Merge tag 'drm-misc-next-fixes-2020-12-22' of git://anongit.freedesktop.org/drm/drm...
[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 reserve_level;
1652         int level;
1653         int max_level;
1654         int replaced = 0;
1655         int ret;
1656         int err = 0;
1657         u32 min_reserved;
1658
1659         path = btrfs_alloc_path();
1660         if (!path)
1661                 return -ENOMEM;
1662         path->reada = READA_FORWARD;
1663
1664         reloc_root = root->reloc_root;
1665         root_item = &reloc_root->root_item;
1666
1667         if (btrfs_disk_key_objectid(&root_item->drop_progress) == 0) {
1668                 level = btrfs_root_level(root_item);
1669                 atomic_inc(&reloc_root->node->refs);
1670                 path->nodes[level] = reloc_root->node;
1671                 path->slots[level] = 0;
1672         } else {
1673                 btrfs_disk_key_to_cpu(&key, &root_item->drop_progress);
1674
1675                 level = root_item->drop_level;
1676                 BUG_ON(level == 0);
1677                 path->lowest_level = level;
1678                 ret = btrfs_search_slot(NULL, reloc_root, &key, path, 0, 0);
1679                 path->lowest_level = 0;
1680                 if (ret < 0) {
1681                         btrfs_free_path(path);
1682                         return ret;
1683                 }
1684
1685                 btrfs_node_key_to_cpu(path->nodes[level], &next_key,
1686                                       path->slots[level]);
1687                 WARN_ON(memcmp(&key, &next_key, sizeof(key)));
1688
1689                 btrfs_unlock_up_safe(path, 0);
1690         }
1691
1692         /*
1693          * In merge_reloc_root(), we modify the upper level pointer to swap the
1694          * tree blocks between reloc tree and subvolume tree.  Thus for tree
1695          * block COW, we COW at most from level 1 to root level for each tree.
1696          *
1697          * Thus the needed metadata size is at most root_level * nodesize,
1698          * and * 2 since we have two trees to COW.
1699          */
1700         reserve_level = max_t(int, 1, btrfs_root_level(root_item));
1701         min_reserved = fs_info->nodesize * reserve_level * 2;
1702         memset(&next_key, 0, sizeof(next_key));
1703
1704         while (1) {
1705                 ret = btrfs_block_rsv_refill(root, rc->block_rsv, min_reserved,
1706                                              BTRFS_RESERVE_FLUSH_LIMIT);
1707                 if (ret) {
1708                         err = ret;
1709                         goto out;
1710                 }
1711                 trans = btrfs_start_transaction(root, 0);
1712                 if (IS_ERR(trans)) {
1713                         err = PTR_ERR(trans);
1714                         trans = NULL;
1715                         goto out;
1716                 }
1717
1718                 /*
1719                  * At this point we no longer have a reloc_control, so we can't
1720                  * depend on btrfs_init_reloc_root to update our last_trans.
1721                  *
1722                  * But that's ok, we started the trans handle on our
1723                  * corresponding fs_root, which means it's been added to the
1724                  * dirty list.  At commit time we'll still call
1725                  * btrfs_update_reloc_root() and update our root item
1726                  * appropriately.
1727                  */
1728                 reloc_root->last_trans = trans->transid;
1729                 trans->block_rsv = rc->block_rsv;
1730
1731                 replaced = 0;
1732                 max_level = level;
1733
1734                 ret = walk_down_reloc_tree(reloc_root, path, &level);
1735                 if (ret < 0) {
1736                         err = ret;
1737                         goto out;
1738                 }
1739                 if (ret > 0)
1740                         break;
1741
1742                 if (!find_next_key(path, level, &key) &&
1743                     btrfs_comp_cpu_keys(&next_key, &key) >= 0) {
1744                         ret = 0;
1745                 } else {
1746                         ret = replace_path(trans, rc, root, reloc_root, path,
1747                                            &next_key, level, max_level);
1748                 }
1749                 if (ret < 0) {
1750                         err = ret;
1751                         goto out;
1752                 }
1753
1754                 if (ret > 0) {
1755                         level = ret;
1756                         btrfs_node_key_to_cpu(path->nodes[level], &key,
1757                                               path->slots[level]);
1758                         replaced = 1;
1759                 }
1760
1761                 ret = walk_up_reloc_tree(reloc_root, path, &level);
1762                 if (ret > 0)
1763                         break;
1764
1765                 BUG_ON(level == 0);
1766                 /*
1767                  * save the merging progress in the drop_progress.
1768                  * this is OK since root refs == 1 in this case.
1769                  */
1770                 btrfs_node_key(path->nodes[level], &root_item->drop_progress,
1771                                path->slots[level]);
1772                 root_item->drop_level = level;
1773
1774                 btrfs_end_transaction_throttle(trans);
1775                 trans = NULL;
1776
1777                 btrfs_btree_balance_dirty(fs_info);
1778
1779                 if (replaced && rc->stage == UPDATE_DATA_PTRS)
1780                         invalidate_extent_cache(root, &key, &next_key);
1781         }
1782
1783         /*
1784          * handle the case only one block in the fs tree need to be
1785          * relocated and the block is tree root.
1786          */
1787         leaf = btrfs_lock_root_node(root);
1788         ret = btrfs_cow_block(trans, root, leaf, NULL, 0, &leaf,
1789                               BTRFS_NESTING_COW);
1790         btrfs_tree_unlock(leaf);
1791         free_extent_buffer(leaf);
1792         if (ret < 0)
1793                 err = ret;
1794 out:
1795         btrfs_free_path(path);
1796
1797         if (err == 0)
1798                 insert_dirty_subvol(trans, rc, root);
1799
1800         if (trans)
1801                 btrfs_end_transaction_throttle(trans);
1802
1803         btrfs_btree_balance_dirty(fs_info);
1804
1805         if (replaced && rc->stage == UPDATE_DATA_PTRS)
1806                 invalidate_extent_cache(root, &key, &next_key);
1807
1808         return err;
1809 }
1810
1811 static noinline_for_stack
1812 int prepare_to_merge(struct reloc_control *rc, int err)
1813 {
1814         struct btrfs_root *root = rc->extent_root;
1815         struct btrfs_fs_info *fs_info = root->fs_info;
1816         struct btrfs_root *reloc_root;
1817         struct btrfs_trans_handle *trans;
1818         LIST_HEAD(reloc_roots);
1819         u64 num_bytes = 0;
1820         int ret;
1821
1822         mutex_lock(&fs_info->reloc_mutex);
1823         rc->merging_rsv_size += fs_info->nodesize * (BTRFS_MAX_LEVEL - 1) * 2;
1824         rc->merging_rsv_size += rc->nodes_relocated * 2;
1825         mutex_unlock(&fs_info->reloc_mutex);
1826
1827 again:
1828         if (!err) {
1829                 num_bytes = rc->merging_rsv_size;
1830                 ret = btrfs_block_rsv_add(root, rc->block_rsv, num_bytes,
1831                                           BTRFS_RESERVE_FLUSH_ALL);
1832                 if (ret)
1833                         err = ret;
1834         }
1835
1836         trans = btrfs_join_transaction(rc->extent_root);
1837         if (IS_ERR(trans)) {
1838                 if (!err)
1839                         btrfs_block_rsv_release(fs_info, rc->block_rsv,
1840                                                 num_bytes, NULL);
1841                 return PTR_ERR(trans);
1842         }
1843
1844         if (!err) {
1845                 if (num_bytes != rc->merging_rsv_size) {
1846                         btrfs_end_transaction(trans);
1847                         btrfs_block_rsv_release(fs_info, rc->block_rsv,
1848                                                 num_bytes, NULL);
1849                         goto again;
1850                 }
1851         }
1852
1853         rc->merge_reloc_tree = 1;
1854
1855         while (!list_empty(&rc->reloc_roots)) {
1856                 reloc_root = list_entry(rc->reloc_roots.next,
1857                                         struct btrfs_root, root_list);
1858                 list_del_init(&reloc_root->root_list);
1859
1860                 root = btrfs_get_fs_root(fs_info, reloc_root->root_key.offset,
1861                                 false);
1862                 BUG_ON(IS_ERR(root));
1863                 BUG_ON(root->reloc_root != reloc_root);
1864
1865                 /*
1866                  * set reference count to 1, so btrfs_recover_relocation
1867                  * knows it should resumes merging
1868                  */
1869                 if (!err)
1870                         btrfs_set_root_refs(&reloc_root->root_item, 1);
1871                 btrfs_update_reloc_root(trans, root);
1872
1873                 list_add(&reloc_root->root_list, &reloc_roots);
1874                 btrfs_put_root(root);
1875         }
1876
1877         list_splice(&reloc_roots, &rc->reloc_roots);
1878
1879         if (!err)
1880                 btrfs_commit_transaction(trans);
1881         else
1882                 btrfs_end_transaction(trans);
1883         return err;
1884 }
1885
1886 static noinline_for_stack
1887 void free_reloc_roots(struct list_head *list)
1888 {
1889         struct btrfs_root *reloc_root, *tmp;
1890
1891         list_for_each_entry_safe(reloc_root, tmp, list, root_list)
1892                 __del_reloc_root(reloc_root);
1893 }
1894
1895 static noinline_for_stack
1896 void merge_reloc_roots(struct reloc_control *rc)
1897 {
1898         struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
1899         struct btrfs_root *root;
1900         struct btrfs_root *reloc_root;
1901         LIST_HEAD(reloc_roots);
1902         int found = 0;
1903         int ret = 0;
1904 again:
1905         root = rc->extent_root;
1906
1907         /*
1908          * this serializes us with btrfs_record_root_in_transaction,
1909          * we have to make sure nobody is in the middle of
1910          * adding their roots to the list while we are
1911          * doing this splice
1912          */
1913         mutex_lock(&fs_info->reloc_mutex);
1914         list_splice_init(&rc->reloc_roots, &reloc_roots);
1915         mutex_unlock(&fs_info->reloc_mutex);
1916
1917         while (!list_empty(&reloc_roots)) {
1918                 found = 1;
1919                 reloc_root = list_entry(reloc_roots.next,
1920                                         struct btrfs_root, root_list);
1921
1922                 root = btrfs_get_fs_root(fs_info, reloc_root->root_key.offset,
1923                                          false);
1924                 if (btrfs_root_refs(&reloc_root->root_item) > 0) {
1925                         BUG_ON(IS_ERR(root));
1926                         BUG_ON(root->reloc_root != reloc_root);
1927                         ret = merge_reloc_root(rc, root);
1928                         btrfs_put_root(root);
1929                         if (ret) {
1930                                 if (list_empty(&reloc_root->root_list))
1931                                         list_add_tail(&reloc_root->root_list,
1932                                                       &reloc_roots);
1933                                 goto out;
1934                         }
1935                 } else {
1936                         if (!IS_ERR(root)) {
1937                                 if (root->reloc_root == reloc_root) {
1938                                         root->reloc_root = NULL;
1939                                         btrfs_put_root(reloc_root);
1940                                 }
1941                                 clear_bit(BTRFS_ROOT_DEAD_RELOC_TREE,
1942                                           &root->state);
1943                                 btrfs_put_root(root);
1944                         }
1945
1946                         list_del_init(&reloc_root->root_list);
1947                         /* Don't forget to queue this reloc root for cleanup */
1948                         list_add_tail(&reloc_root->reloc_dirty_list,
1949                                       &rc->dirty_subvol_roots);
1950                 }
1951         }
1952
1953         if (found) {
1954                 found = 0;
1955                 goto again;
1956         }
1957 out:
1958         if (ret) {
1959                 btrfs_handle_fs_error(fs_info, ret, NULL);
1960                 free_reloc_roots(&reloc_roots);
1961
1962                 /* new reloc root may be added */
1963                 mutex_lock(&fs_info->reloc_mutex);
1964                 list_splice_init(&rc->reloc_roots, &reloc_roots);
1965                 mutex_unlock(&fs_info->reloc_mutex);
1966                 free_reloc_roots(&reloc_roots);
1967         }
1968
1969         /*
1970          * We used to have
1971          *
1972          * BUG_ON(!RB_EMPTY_ROOT(&rc->reloc_root_tree.rb_root));
1973          *
1974          * here, but it's wrong.  If we fail to start the transaction in
1975          * prepare_to_merge() we will have only 0 ref reloc roots, none of which
1976          * have actually been removed from the reloc_root_tree rb tree.  This is
1977          * fine because we're bailing here, and we hold a reference on the root
1978          * for the list that holds it, so these roots will be cleaned up when we
1979          * do the reloc_dirty_list afterwards.  Meanwhile the root->reloc_root
1980          * will be cleaned up on unmount.
1981          *
1982          * The remaining nodes will be cleaned up by free_reloc_control.
1983          */
1984 }
1985
1986 static void free_block_list(struct rb_root *blocks)
1987 {
1988         struct tree_block *block;
1989         struct rb_node *rb_node;
1990         while ((rb_node = rb_first(blocks))) {
1991                 block = rb_entry(rb_node, struct tree_block, rb_node);
1992                 rb_erase(rb_node, blocks);
1993                 kfree(block);
1994         }
1995 }
1996
1997 static int record_reloc_root_in_trans(struct btrfs_trans_handle *trans,
1998                                       struct btrfs_root *reloc_root)
1999 {
2000         struct btrfs_fs_info *fs_info = reloc_root->fs_info;
2001         struct btrfs_root *root;
2002         int ret;
2003
2004         if (reloc_root->last_trans == trans->transid)
2005                 return 0;
2006
2007         root = btrfs_get_fs_root(fs_info, reloc_root->root_key.offset, false);
2008         BUG_ON(IS_ERR(root));
2009         BUG_ON(root->reloc_root != reloc_root);
2010         ret = btrfs_record_root_in_trans(trans, root);
2011         btrfs_put_root(root);
2012
2013         return ret;
2014 }
2015
2016 static noinline_for_stack
2017 struct btrfs_root *select_reloc_root(struct btrfs_trans_handle *trans,
2018                                      struct reloc_control *rc,
2019                                      struct btrfs_backref_node *node,
2020                                      struct btrfs_backref_edge *edges[])
2021 {
2022         struct btrfs_backref_node *next;
2023         struct btrfs_root *root;
2024         int index = 0;
2025
2026         next = node;
2027         while (1) {
2028                 cond_resched();
2029                 next = walk_up_backref(next, edges, &index);
2030                 root = next->root;
2031                 BUG_ON(!root);
2032                 BUG_ON(!test_bit(BTRFS_ROOT_SHAREABLE, &root->state));
2033
2034                 if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID) {
2035                         record_reloc_root_in_trans(trans, root);
2036                         break;
2037                 }
2038
2039                 btrfs_record_root_in_trans(trans, root);
2040                 root = root->reloc_root;
2041
2042                 if (next->new_bytenr != root->node->start) {
2043                         BUG_ON(next->new_bytenr);
2044                         BUG_ON(!list_empty(&next->list));
2045                         next->new_bytenr = root->node->start;
2046                         btrfs_put_root(next->root);
2047                         next->root = btrfs_grab_root(root);
2048                         ASSERT(next->root);
2049                         list_add_tail(&next->list,
2050                                       &rc->backref_cache.changed);
2051                         mark_block_processed(rc, next);
2052                         break;
2053                 }
2054
2055                 WARN_ON(1);
2056                 root = NULL;
2057                 next = walk_down_backref(edges, &index);
2058                 if (!next || next->level <= node->level)
2059                         break;
2060         }
2061         if (!root)
2062                 return NULL;
2063
2064         next = node;
2065         /* setup backref node path for btrfs_reloc_cow_block */
2066         while (1) {
2067                 rc->backref_cache.path[next->level] = next;
2068                 if (--index < 0)
2069                         break;
2070                 next = edges[index]->node[UPPER];
2071         }
2072         return root;
2073 }
2074
2075 /*
2076  * Select a tree root for relocation.
2077  *
2078  * Return NULL if the block is not shareable. We should use do_relocation() in
2079  * this case.
2080  *
2081  * Return a tree root pointer if the block is shareable.
2082  * Return -ENOENT if the block is root of reloc tree.
2083  */
2084 static noinline_for_stack
2085 struct btrfs_root *select_one_root(struct btrfs_backref_node *node)
2086 {
2087         struct btrfs_backref_node *next;
2088         struct btrfs_root *root;
2089         struct btrfs_root *fs_root = NULL;
2090         struct btrfs_backref_edge *edges[BTRFS_MAX_LEVEL - 1];
2091         int index = 0;
2092
2093         next = node;
2094         while (1) {
2095                 cond_resched();
2096                 next = walk_up_backref(next, edges, &index);
2097                 root = next->root;
2098                 BUG_ON(!root);
2099
2100                 /* No other choice for non-shareable tree */
2101                 if (!test_bit(BTRFS_ROOT_SHAREABLE, &root->state))
2102                         return root;
2103
2104                 if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID)
2105                         fs_root = root;
2106
2107                 if (next != node)
2108                         return NULL;
2109
2110                 next = walk_down_backref(edges, &index);
2111                 if (!next || next->level <= node->level)
2112                         break;
2113         }
2114
2115         if (!fs_root)
2116                 return ERR_PTR(-ENOENT);
2117         return fs_root;
2118 }
2119
2120 static noinline_for_stack
2121 u64 calcu_metadata_size(struct reloc_control *rc,
2122                         struct btrfs_backref_node *node, int reserve)
2123 {
2124         struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
2125         struct btrfs_backref_node *next = node;
2126         struct btrfs_backref_edge *edge;
2127         struct btrfs_backref_edge *edges[BTRFS_MAX_LEVEL - 1];
2128         u64 num_bytes = 0;
2129         int index = 0;
2130
2131         BUG_ON(reserve && node->processed);
2132
2133         while (next) {
2134                 cond_resched();
2135                 while (1) {
2136                         if (next->processed && (reserve || next != node))
2137                                 break;
2138
2139                         num_bytes += fs_info->nodesize;
2140
2141                         if (list_empty(&next->upper))
2142                                 break;
2143
2144                         edge = list_entry(next->upper.next,
2145                                         struct btrfs_backref_edge, list[LOWER]);
2146                         edges[index++] = edge;
2147                         next = edge->node[UPPER];
2148                 }
2149                 next = walk_down_backref(edges, &index);
2150         }
2151         return num_bytes;
2152 }
2153
2154 static int reserve_metadata_space(struct btrfs_trans_handle *trans,
2155                                   struct reloc_control *rc,
2156                                   struct btrfs_backref_node *node)
2157 {
2158         struct btrfs_root *root = rc->extent_root;
2159         struct btrfs_fs_info *fs_info = root->fs_info;
2160         u64 num_bytes;
2161         int ret;
2162         u64 tmp;
2163
2164         num_bytes = calcu_metadata_size(rc, node, 1) * 2;
2165
2166         trans->block_rsv = rc->block_rsv;
2167         rc->reserved_bytes += num_bytes;
2168
2169         /*
2170          * We are under a transaction here so we can only do limited flushing.
2171          * If we get an enospc just kick back -EAGAIN so we know to drop the
2172          * transaction and try to refill when we can flush all the things.
2173          */
2174         ret = btrfs_block_rsv_refill(root, rc->block_rsv, num_bytes,
2175                                 BTRFS_RESERVE_FLUSH_LIMIT);
2176         if (ret) {
2177                 tmp = fs_info->nodesize * RELOCATION_RESERVED_NODES;
2178                 while (tmp <= rc->reserved_bytes)
2179                         tmp <<= 1;
2180                 /*
2181                  * only one thread can access block_rsv at this point,
2182                  * so we don't need hold lock to protect block_rsv.
2183                  * we expand more reservation size here to allow enough
2184                  * space for relocation and we will return earlier in
2185                  * enospc case.
2186                  */
2187                 rc->block_rsv->size = tmp + fs_info->nodesize *
2188                                       RELOCATION_RESERVED_NODES;
2189                 return -EAGAIN;
2190         }
2191
2192         return 0;
2193 }
2194
2195 /*
2196  * relocate a block tree, and then update pointers in upper level
2197  * blocks that reference the block to point to the new location.
2198  *
2199  * if called by link_to_upper, the block has already been relocated.
2200  * in that case this function just updates pointers.
2201  */
2202 static int do_relocation(struct btrfs_trans_handle *trans,
2203                          struct reloc_control *rc,
2204                          struct btrfs_backref_node *node,
2205                          struct btrfs_key *key,
2206                          struct btrfs_path *path, int lowest)
2207 {
2208         struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
2209         struct btrfs_backref_node *upper;
2210         struct btrfs_backref_edge *edge;
2211         struct btrfs_backref_edge *edges[BTRFS_MAX_LEVEL - 1];
2212         struct btrfs_root *root;
2213         struct extent_buffer *eb;
2214         u32 blocksize;
2215         u64 bytenr;
2216         u64 generation;
2217         int slot;
2218         int ret;
2219         int err = 0;
2220
2221         BUG_ON(lowest && node->eb);
2222
2223         path->lowest_level = node->level + 1;
2224         rc->backref_cache.path[node->level] = node;
2225         list_for_each_entry(edge, &node->upper, list[LOWER]) {
2226                 struct btrfs_key first_key;
2227                 struct btrfs_ref ref = { 0 };
2228
2229                 cond_resched();
2230
2231                 upper = edge->node[UPPER];
2232                 root = select_reloc_root(trans, rc, upper, edges);
2233                 BUG_ON(!root);
2234
2235                 if (upper->eb && !upper->locked) {
2236                         if (!lowest) {
2237                                 ret = btrfs_bin_search(upper->eb, key, &slot);
2238                                 if (ret < 0) {
2239                                         err = ret;
2240                                         goto next;
2241                                 }
2242                                 BUG_ON(ret);
2243                                 bytenr = btrfs_node_blockptr(upper->eb, slot);
2244                                 if (node->eb->start == bytenr)
2245                                         goto next;
2246                         }
2247                         btrfs_backref_drop_node_buffer(upper);
2248                 }
2249
2250                 if (!upper->eb) {
2251                         ret = btrfs_search_slot(trans, root, key, path, 0, 1);
2252                         if (ret) {
2253                                 if (ret < 0)
2254                                         err = ret;
2255                                 else
2256                                         err = -ENOENT;
2257
2258                                 btrfs_release_path(path);
2259                                 break;
2260                         }
2261
2262                         if (!upper->eb) {
2263                                 upper->eb = path->nodes[upper->level];
2264                                 path->nodes[upper->level] = NULL;
2265                         } else {
2266                                 BUG_ON(upper->eb != path->nodes[upper->level]);
2267                         }
2268
2269                         upper->locked = 1;
2270                         path->locks[upper->level] = 0;
2271
2272                         slot = path->slots[upper->level];
2273                         btrfs_release_path(path);
2274                 } else {
2275                         ret = btrfs_bin_search(upper->eb, key, &slot);
2276                         if (ret < 0) {
2277                                 err = ret;
2278                                 goto next;
2279                         }
2280                         BUG_ON(ret);
2281                 }
2282
2283                 bytenr = btrfs_node_blockptr(upper->eb, slot);
2284                 if (lowest) {
2285                         if (bytenr != node->bytenr) {
2286                                 btrfs_err(root->fs_info,
2287                 "lowest leaf/node mismatch: bytenr %llu node->bytenr %llu slot %d upper %llu",
2288                                           bytenr, node->bytenr, slot,
2289                                           upper->eb->start);
2290                                 err = -EIO;
2291                                 goto next;
2292                         }
2293                 } else {
2294                         if (node->eb->start == bytenr)
2295                                 goto next;
2296                 }
2297
2298                 blocksize = root->fs_info->nodesize;
2299                 generation = btrfs_node_ptr_generation(upper->eb, slot);
2300                 btrfs_node_key_to_cpu(upper->eb, &first_key, slot);
2301                 eb = read_tree_block(fs_info, bytenr, generation,
2302                                      upper->level - 1, &first_key);
2303                 if (IS_ERR(eb)) {
2304                         err = PTR_ERR(eb);
2305                         goto next;
2306                 } else if (!extent_buffer_uptodate(eb)) {
2307                         free_extent_buffer(eb);
2308                         err = -EIO;
2309                         goto next;
2310                 }
2311                 btrfs_tree_lock(eb);
2312                 btrfs_set_lock_blocking_write(eb);
2313
2314                 if (!node->eb) {
2315                         ret = btrfs_cow_block(trans, root, eb, upper->eb,
2316                                               slot, &eb, BTRFS_NESTING_COW);
2317                         btrfs_tree_unlock(eb);
2318                         free_extent_buffer(eb);
2319                         if (ret < 0) {
2320                                 err = ret;
2321                                 goto next;
2322                         }
2323                         BUG_ON(node->eb != eb);
2324                 } else {
2325                         btrfs_set_node_blockptr(upper->eb, slot,
2326                                                 node->eb->start);
2327                         btrfs_set_node_ptr_generation(upper->eb, slot,
2328                                                       trans->transid);
2329                         btrfs_mark_buffer_dirty(upper->eb);
2330
2331                         btrfs_init_generic_ref(&ref, BTRFS_ADD_DELAYED_REF,
2332                                                node->eb->start, blocksize,
2333                                                upper->eb->start);
2334                         ref.real_root = root->root_key.objectid;
2335                         btrfs_init_tree_ref(&ref, node->level,
2336                                             btrfs_header_owner(upper->eb));
2337                         ret = btrfs_inc_extent_ref(trans, &ref);
2338                         BUG_ON(ret);
2339
2340                         ret = btrfs_drop_subtree(trans, root, eb, upper->eb);
2341                         BUG_ON(ret);
2342                 }
2343 next:
2344                 if (!upper->pending)
2345                         btrfs_backref_drop_node_buffer(upper);
2346                 else
2347                         btrfs_backref_unlock_node_buffer(upper);
2348                 if (err)
2349                         break;
2350         }
2351
2352         if (!err && node->pending) {
2353                 btrfs_backref_drop_node_buffer(node);
2354                 list_move_tail(&node->list, &rc->backref_cache.changed);
2355                 node->pending = 0;
2356         }
2357
2358         path->lowest_level = 0;
2359         BUG_ON(err == -ENOSPC);
2360         return err;
2361 }
2362
2363 static int link_to_upper(struct btrfs_trans_handle *trans,
2364                          struct reloc_control *rc,
2365                          struct btrfs_backref_node *node,
2366                          struct btrfs_path *path)
2367 {
2368         struct btrfs_key key;
2369
2370         btrfs_node_key_to_cpu(node->eb, &key, 0);
2371         return do_relocation(trans, rc, node, &key, path, 0);
2372 }
2373
2374 static int finish_pending_nodes(struct btrfs_trans_handle *trans,
2375                                 struct reloc_control *rc,
2376                                 struct btrfs_path *path, int err)
2377 {
2378         LIST_HEAD(list);
2379         struct btrfs_backref_cache *cache = &rc->backref_cache;
2380         struct btrfs_backref_node *node;
2381         int level;
2382         int ret;
2383
2384         for (level = 0; level < BTRFS_MAX_LEVEL; level++) {
2385                 while (!list_empty(&cache->pending[level])) {
2386                         node = list_entry(cache->pending[level].next,
2387                                           struct btrfs_backref_node, list);
2388                         list_move_tail(&node->list, &list);
2389                         BUG_ON(!node->pending);
2390
2391                         if (!err) {
2392                                 ret = link_to_upper(trans, rc, node, path);
2393                                 if (ret < 0)
2394                                         err = ret;
2395                         }
2396                 }
2397                 list_splice_init(&list, &cache->pending[level]);
2398         }
2399         return err;
2400 }
2401
2402 /*
2403  * mark a block and all blocks directly/indirectly reference the block
2404  * as processed.
2405  */
2406 static void update_processed_blocks(struct reloc_control *rc,
2407                                     struct btrfs_backref_node *node)
2408 {
2409         struct btrfs_backref_node *next = node;
2410         struct btrfs_backref_edge *edge;
2411         struct btrfs_backref_edge *edges[BTRFS_MAX_LEVEL - 1];
2412         int index = 0;
2413
2414         while (next) {
2415                 cond_resched();
2416                 while (1) {
2417                         if (next->processed)
2418                                 break;
2419
2420                         mark_block_processed(rc, next);
2421
2422                         if (list_empty(&next->upper))
2423                                 break;
2424
2425                         edge = list_entry(next->upper.next,
2426                                         struct btrfs_backref_edge, list[LOWER]);
2427                         edges[index++] = edge;
2428                         next = edge->node[UPPER];
2429                 }
2430                 next = walk_down_backref(edges, &index);
2431         }
2432 }
2433
2434 static int tree_block_processed(u64 bytenr, struct reloc_control *rc)
2435 {
2436         u32 blocksize = rc->extent_root->fs_info->nodesize;
2437
2438         if (test_range_bit(&rc->processed_blocks, bytenr,
2439                            bytenr + blocksize - 1, EXTENT_DIRTY, 1, NULL))
2440                 return 1;
2441         return 0;
2442 }
2443
2444 static int get_tree_block_key(struct btrfs_fs_info *fs_info,
2445                               struct tree_block *block)
2446 {
2447         struct extent_buffer *eb;
2448
2449         eb = read_tree_block(fs_info, block->bytenr, block->key.offset,
2450                              block->level, NULL);
2451         if (IS_ERR(eb)) {
2452                 return PTR_ERR(eb);
2453         } else if (!extent_buffer_uptodate(eb)) {
2454                 free_extent_buffer(eb);
2455                 return -EIO;
2456         }
2457         if (block->level == 0)
2458                 btrfs_item_key_to_cpu(eb, &block->key, 0);
2459         else
2460                 btrfs_node_key_to_cpu(eb, &block->key, 0);
2461         free_extent_buffer(eb);
2462         block->key_ready = 1;
2463         return 0;
2464 }
2465
2466 /*
2467  * helper function to relocate a tree block
2468  */
2469 static int relocate_tree_block(struct btrfs_trans_handle *trans,
2470                                 struct reloc_control *rc,
2471                                 struct btrfs_backref_node *node,
2472                                 struct btrfs_key *key,
2473                                 struct btrfs_path *path)
2474 {
2475         struct btrfs_root *root;
2476         int ret = 0;
2477
2478         if (!node)
2479                 return 0;
2480
2481         /*
2482          * If we fail here we want to drop our backref_node because we are going
2483          * to start over and regenerate the tree for it.
2484          */
2485         ret = reserve_metadata_space(trans, rc, node);
2486         if (ret)
2487                 goto out;
2488
2489         BUG_ON(node->processed);
2490         root = select_one_root(node);
2491         if (root == ERR_PTR(-ENOENT)) {
2492                 update_processed_blocks(rc, node);
2493                 goto out;
2494         }
2495
2496         if (root) {
2497                 if (test_bit(BTRFS_ROOT_SHAREABLE, &root->state)) {
2498                         BUG_ON(node->new_bytenr);
2499                         BUG_ON(!list_empty(&node->list));
2500                         btrfs_record_root_in_trans(trans, root);
2501                         root = root->reloc_root;
2502                         node->new_bytenr = root->node->start;
2503                         btrfs_put_root(node->root);
2504                         node->root = btrfs_grab_root(root);
2505                         ASSERT(node->root);
2506                         list_add_tail(&node->list, &rc->backref_cache.changed);
2507                 } else {
2508                         path->lowest_level = node->level;
2509                         ret = btrfs_search_slot(trans, root, key, path, 0, 1);
2510                         btrfs_release_path(path);
2511                         if (ret > 0)
2512                                 ret = 0;
2513                 }
2514                 if (!ret)
2515                         update_processed_blocks(rc, node);
2516         } else {
2517                 ret = do_relocation(trans, rc, node, key, path, 1);
2518         }
2519 out:
2520         if (ret || node->level == 0 || node->cowonly)
2521                 btrfs_backref_cleanup_node(&rc->backref_cache, node);
2522         return ret;
2523 }
2524
2525 /*
2526  * relocate a list of blocks
2527  */
2528 static noinline_for_stack
2529 int relocate_tree_blocks(struct btrfs_trans_handle *trans,
2530                          struct reloc_control *rc, struct rb_root *blocks)
2531 {
2532         struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
2533         struct btrfs_backref_node *node;
2534         struct btrfs_path *path;
2535         struct tree_block *block;
2536         struct tree_block *next;
2537         int ret;
2538         int err = 0;
2539
2540         path = btrfs_alloc_path();
2541         if (!path) {
2542                 err = -ENOMEM;
2543                 goto out_free_blocks;
2544         }
2545
2546         /* Kick in readahead for tree blocks with missing keys */
2547         rbtree_postorder_for_each_entry_safe(block, next, blocks, rb_node) {
2548                 if (!block->key_ready)
2549                         readahead_tree_block(fs_info, block->bytenr);
2550         }
2551
2552         /* Get first keys */
2553         rbtree_postorder_for_each_entry_safe(block, next, blocks, rb_node) {
2554                 if (!block->key_ready) {
2555                         err = get_tree_block_key(fs_info, block);
2556                         if (err)
2557                                 goto out_free_path;
2558                 }
2559         }
2560
2561         /* Do tree relocation */
2562         rbtree_postorder_for_each_entry_safe(block, next, blocks, rb_node) {
2563                 node = build_backref_tree(rc, &block->key,
2564                                           block->level, block->bytenr);
2565                 if (IS_ERR(node)) {
2566                         err = PTR_ERR(node);
2567                         goto out;
2568                 }
2569
2570                 ret = relocate_tree_block(trans, rc, node, &block->key,
2571                                           path);
2572                 if (ret < 0) {
2573                         err = ret;
2574                         break;
2575                 }
2576         }
2577 out:
2578         err = finish_pending_nodes(trans, rc, path, err);
2579
2580 out_free_path:
2581         btrfs_free_path(path);
2582 out_free_blocks:
2583         free_block_list(blocks);
2584         return err;
2585 }
2586
2587 static noinline_for_stack int prealloc_file_extent_cluster(
2588                                 struct btrfs_inode *inode,
2589                                 struct file_extent_cluster *cluster)
2590 {
2591         u64 alloc_hint = 0;
2592         u64 start;
2593         u64 end;
2594         u64 offset = inode->index_cnt;
2595         u64 num_bytes;
2596         int nr;
2597         int ret = 0;
2598         u64 prealloc_start = cluster->start - offset;
2599         u64 prealloc_end = cluster->end - offset;
2600         u64 cur_offset = prealloc_start;
2601
2602         BUG_ON(cluster->start != cluster->boundary[0]);
2603         ret = btrfs_alloc_data_chunk_ondemand(inode,
2604                                               prealloc_end + 1 - prealloc_start);
2605         if (ret)
2606                 return ret;
2607
2608         inode_lock(&inode->vfs_inode);
2609         for (nr = 0; nr < cluster->nr; nr++) {
2610                 start = cluster->boundary[nr] - offset;
2611                 if (nr + 1 < cluster->nr)
2612                         end = cluster->boundary[nr + 1] - 1 - offset;
2613                 else
2614                         end = cluster->end - offset;
2615
2616                 lock_extent(&inode->io_tree, start, end);
2617                 num_bytes = end + 1 - start;
2618                 ret = btrfs_prealloc_file_range(&inode->vfs_inode, 0, start,
2619                                                 num_bytes, num_bytes,
2620                                                 end + 1, &alloc_hint);
2621                 cur_offset = end + 1;
2622                 unlock_extent(&inode->io_tree, start, end);
2623                 if (ret)
2624                         break;
2625         }
2626         inode_unlock(&inode->vfs_inode);
2627
2628         if (cur_offset < prealloc_end)
2629                 btrfs_free_reserved_data_space_noquota(inode->root->fs_info,
2630                                                prealloc_end + 1 - cur_offset);
2631         return ret;
2632 }
2633
2634 static noinline_for_stack
2635 int setup_extent_mapping(struct inode *inode, u64 start, u64 end,
2636                          u64 block_start)
2637 {
2638         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
2639         struct extent_map *em;
2640         int ret = 0;
2641
2642         em = alloc_extent_map();
2643         if (!em)
2644                 return -ENOMEM;
2645
2646         em->start = start;
2647         em->len = end + 1 - start;
2648         em->block_len = em->len;
2649         em->block_start = block_start;
2650         set_bit(EXTENT_FLAG_PINNED, &em->flags);
2651
2652         lock_extent(&BTRFS_I(inode)->io_tree, start, end);
2653         while (1) {
2654                 write_lock(&em_tree->lock);
2655                 ret = add_extent_mapping(em_tree, em, 0);
2656                 write_unlock(&em_tree->lock);
2657                 if (ret != -EEXIST) {
2658                         free_extent_map(em);
2659                         break;
2660                 }
2661                 btrfs_drop_extent_cache(BTRFS_I(inode), start, end, 0);
2662         }
2663         unlock_extent(&BTRFS_I(inode)->io_tree, start, end);
2664         return ret;
2665 }
2666
2667 /*
2668  * Allow error injection to test balance cancellation
2669  */
2670 int btrfs_should_cancel_balance(struct btrfs_fs_info *fs_info)
2671 {
2672         return atomic_read(&fs_info->balance_cancel_req) ||
2673                 fatal_signal_pending(current);
2674 }
2675 ALLOW_ERROR_INJECTION(btrfs_should_cancel_balance, TRUE);
2676
2677 static int relocate_file_extent_cluster(struct inode *inode,
2678                                         struct file_extent_cluster *cluster)
2679 {
2680         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2681         u64 page_start;
2682         u64 page_end;
2683         u64 offset = BTRFS_I(inode)->index_cnt;
2684         unsigned long index;
2685         unsigned long last_index;
2686         struct page *page;
2687         struct file_ra_state *ra;
2688         gfp_t mask = btrfs_alloc_write_mask(inode->i_mapping);
2689         int nr = 0;
2690         int ret = 0;
2691
2692         if (!cluster->nr)
2693                 return 0;
2694
2695         ra = kzalloc(sizeof(*ra), GFP_NOFS);
2696         if (!ra)
2697                 return -ENOMEM;
2698
2699         ret = prealloc_file_extent_cluster(BTRFS_I(inode), cluster);
2700         if (ret)
2701                 goto out;
2702
2703         file_ra_state_init(ra, inode->i_mapping);
2704
2705         ret = setup_extent_mapping(inode, cluster->start - offset,
2706                                    cluster->end - offset, cluster->start);
2707         if (ret)
2708                 goto out;
2709
2710         index = (cluster->start - offset) >> PAGE_SHIFT;
2711         last_index = (cluster->end - offset) >> PAGE_SHIFT;
2712         while (index <= last_index) {
2713                 ret = btrfs_delalloc_reserve_metadata(BTRFS_I(inode),
2714                                 PAGE_SIZE);
2715                 if (ret)
2716                         goto out;
2717
2718                 page = find_lock_page(inode->i_mapping, index);
2719                 if (!page) {
2720                         page_cache_sync_readahead(inode->i_mapping,
2721                                                   ra, NULL, index,
2722                                                   last_index + 1 - index);
2723                         page = find_or_create_page(inode->i_mapping, index,
2724                                                    mask);
2725                         if (!page) {
2726                                 btrfs_delalloc_release_metadata(BTRFS_I(inode),
2727                                                         PAGE_SIZE, true);
2728                                 btrfs_delalloc_release_extents(BTRFS_I(inode),
2729                                                         PAGE_SIZE);
2730                                 ret = -ENOMEM;
2731                                 goto out;
2732                         }
2733                 }
2734
2735                 if (PageReadahead(page)) {
2736                         page_cache_async_readahead(inode->i_mapping,
2737                                                    ra, NULL, page, index,
2738                                                    last_index + 1 - index);
2739                 }
2740
2741                 if (!PageUptodate(page)) {
2742                         btrfs_readpage(NULL, page);
2743                         lock_page(page);
2744                         if (!PageUptodate(page)) {
2745                                 unlock_page(page);
2746                                 put_page(page);
2747                                 btrfs_delalloc_release_metadata(BTRFS_I(inode),
2748                                                         PAGE_SIZE, true);
2749                                 btrfs_delalloc_release_extents(BTRFS_I(inode),
2750                                                                PAGE_SIZE);
2751                                 ret = -EIO;
2752                                 goto out;
2753                         }
2754                 }
2755
2756                 page_start = page_offset(page);
2757                 page_end = page_start + PAGE_SIZE - 1;
2758
2759                 lock_extent(&BTRFS_I(inode)->io_tree, page_start, page_end);
2760
2761                 set_page_extent_mapped(page);
2762
2763                 if (nr < cluster->nr &&
2764                     page_start + offset == cluster->boundary[nr]) {
2765                         set_extent_bits(&BTRFS_I(inode)->io_tree,
2766                                         page_start, page_end,
2767                                         EXTENT_BOUNDARY);
2768                         nr++;
2769                 }
2770
2771                 ret = btrfs_set_extent_delalloc(BTRFS_I(inode), page_start,
2772                                                 page_end, 0, NULL);
2773                 if (ret) {
2774                         unlock_page(page);
2775                         put_page(page);
2776                         btrfs_delalloc_release_metadata(BTRFS_I(inode),
2777                                                          PAGE_SIZE, true);
2778                         btrfs_delalloc_release_extents(BTRFS_I(inode),
2779                                                        PAGE_SIZE);
2780
2781                         clear_extent_bits(&BTRFS_I(inode)->io_tree,
2782                                           page_start, page_end,
2783                                           EXTENT_LOCKED | EXTENT_BOUNDARY);
2784                         goto out;
2785
2786                 }
2787                 set_page_dirty(page);
2788
2789                 unlock_extent(&BTRFS_I(inode)->io_tree,
2790                               page_start, page_end);
2791                 unlock_page(page);
2792                 put_page(page);
2793
2794                 index++;
2795                 btrfs_delalloc_release_extents(BTRFS_I(inode), PAGE_SIZE);
2796                 balance_dirty_pages_ratelimited(inode->i_mapping);
2797                 btrfs_throttle(fs_info);
2798                 if (btrfs_should_cancel_balance(fs_info)) {
2799                         ret = -ECANCELED;
2800                         goto out;
2801                 }
2802         }
2803         WARN_ON(nr != cluster->nr);
2804 out:
2805         kfree(ra);
2806         return ret;
2807 }
2808
2809 static noinline_for_stack
2810 int relocate_data_extent(struct inode *inode, struct btrfs_key *extent_key,
2811                          struct file_extent_cluster *cluster)
2812 {
2813         int ret;
2814
2815         if (cluster->nr > 0 && extent_key->objectid != cluster->end + 1) {
2816                 ret = relocate_file_extent_cluster(inode, cluster);
2817                 if (ret)
2818                         return ret;
2819                 cluster->nr = 0;
2820         }
2821
2822         if (!cluster->nr)
2823                 cluster->start = extent_key->objectid;
2824         else
2825                 BUG_ON(cluster->nr >= MAX_EXTENTS);
2826         cluster->end = extent_key->objectid + extent_key->offset - 1;
2827         cluster->boundary[cluster->nr] = extent_key->objectid;
2828         cluster->nr++;
2829
2830         if (cluster->nr >= MAX_EXTENTS) {
2831                 ret = relocate_file_extent_cluster(inode, cluster);
2832                 if (ret)
2833                         return ret;
2834                 cluster->nr = 0;
2835         }
2836         return 0;
2837 }
2838
2839 /*
2840  * helper to add a tree block to the list.
2841  * the major work is getting the generation and level of the block
2842  */
2843 static int add_tree_block(struct reloc_control *rc,
2844                           struct btrfs_key *extent_key,
2845                           struct btrfs_path *path,
2846                           struct rb_root *blocks)
2847 {
2848         struct extent_buffer *eb;
2849         struct btrfs_extent_item *ei;
2850         struct btrfs_tree_block_info *bi;
2851         struct tree_block *block;
2852         struct rb_node *rb_node;
2853         u32 item_size;
2854         int level = -1;
2855         u64 generation;
2856
2857         eb =  path->nodes[0];
2858         item_size = btrfs_item_size_nr(eb, path->slots[0]);
2859
2860         if (extent_key->type == BTRFS_METADATA_ITEM_KEY ||
2861             item_size >= sizeof(*ei) + sizeof(*bi)) {
2862                 ei = btrfs_item_ptr(eb, path->slots[0],
2863                                 struct btrfs_extent_item);
2864                 if (extent_key->type == BTRFS_EXTENT_ITEM_KEY) {
2865                         bi = (struct btrfs_tree_block_info *)(ei + 1);
2866                         level = btrfs_tree_block_level(eb, bi);
2867                 } else {
2868                         level = (int)extent_key->offset;
2869                 }
2870                 generation = btrfs_extent_generation(eb, ei);
2871         } else if (unlikely(item_size == sizeof(struct btrfs_extent_item_v0))) {
2872                 btrfs_print_v0_err(eb->fs_info);
2873                 btrfs_handle_fs_error(eb->fs_info, -EINVAL, NULL);
2874                 return -EINVAL;
2875         } else {
2876                 BUG();
2877         }
2878
2879         btrfs_release_path(path);
2880
2881         BUG_ON(level == -1);
2882
2883         block = kmalloc(sizeof(*block), GFP_NOFS);
2884         if (!block)
2885                 return -ENOMEM;
2886
2887         block->bytenr = extent_key->objectid;
2888         block->key.objectid = rc->extent_root->fs_info->nodesize;
2889         block->key.offset = generation;
2890         block->level = level;
2891         block->key_ready = 0;
2892
2893         rb_node = rb_simple_insert(blocks, block->bytenr, &block->rb_node);
2894         if (rb_node)
2895                 btrfs_backref_panic(rc->extent_root->fs_info, block->bytenr,
2896                                     -EEXIST);
2897
2898         return 0;
2899 }
2900
2901 /*
2902  * helper to add tree blocks for backref of type BTRFS_SHARED_DATA_REF_KEY
2903  */
2904 static int __add_tree_block(struct reloc_control *rc,
2905                             u64 bytenr, u32 blocksize,
2906                             struct rb_root *blocks)
2907 {
2908         struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
2909         struct btrfs_path *path;
2910         struct btrfs_key key;
2911         int ret;
2912         bool skinny = btrfs_fs_incompat(fs_info, SKINNY_METADATA);
2913
2914         if (tree_block_processed(bytenr, rc))
2915                 return 0;
2916
2917         if (rb_simple_search(blocks, bytenr))
2918                 return 0;
2919
2920         path = btrfs_alloc_path();
2921         if (!path)
2922                 return -ENOMEM;
2923 again:
2924         key.objectid = bytenr;
2925         if (skinny) {
2926                 key.type = BTRFS_METADATA_ITEM_KEY;
2927                 key.offset = (u64)-1;
2928         } else {
2929                 key.type = BTRFS_EXTENT_ITEM_KEY;
2930                 key.offset = blocksize;
2931         }
2932
2933         path->search_commit_root = 1;
2934         path->skip_locking = 1;
2935         ret = btrfs_search_slot(NULL, rc->extent_root, &key, path, 0, 0);
2936         if (ret < 0)
2937                 goto out;
2938
2939         if (ret > 0 && skinny) {
2940                 if (path->slots[0]) {
2941                         path->slots[0]--;
2942                         btrfs_item_key_to_cpu(path->nodes[0], &key,
2943                                               path->slots[0]);
2944                         if (key.objectid == bytenr &&
2945                             (key.type == BTRFS_METADATA_ITEM_KEY ||
2946                              (key.type == BTRFS_EXTENT_ITEM_KEY &&
2947                               key.offset == blocksize)))
2948                                 ret = 0;
2949                 }
2950
2951                 if (ret) {
2952                         skinny = false;
2953                         btrfs_release_path(path);
2954                         goto again;
2955                 }
2956         }
2957         if (ret) {
2958                 ASSERT(ret == 1);
2959                 btrfs_print_leaf(path->nodes[0]);
2960                 btrfs_err(fs_info,
2961              "tree block extent item (%llu) is not found in extent tree",
2962                      bytenr);
2963                 WARN_ON(1);
2964                 ret = -EINVAL;
2965                 goto out;
2966         }
2967
2968         ret = add_tree_block(rc, &key, path, blocks);
2969 out:
2970         btrfs_free_path(path);
2971         return ret;
2972 }
2973
2974 static int delete_block_group_cache(struct btrfs_fs_info *fs_info,
2975                                     struct btrfs_block_group *block_group,
2976                                     struct inode *inode,
2977                                     u64 ino)
2978 {
2979         struct btrfs_root *root = fs_info->tree_root;
2980         struct btrfs_trans_handle *trans;
2981         int ret = 0;
2982
2983         if (inode)
2984                 goto truncate;
2985
2986         inode = btrfs_iget(fs_info->sb, ino, root);
2987         if (IS_ERR(inode))
2988                 return -ENOENT;
2989
2990 truncate:
2991         ret = btrfs_check_trunc_cache_free_space(fs_info,
2992                                                  &fs_info->global_block_rsv);
2993         if (ret)
2994                 goto out;
2995
2996         trans = btrfs_join_transaction(root);
2997         if (IS_ERR(trans)) {
2998                 ret = PTR_ERR(trans);
2999                 goto out;
3000         }
3001
3002         ret = btrfs_truncate_free_space_cache(trans, block_group, inode);
3003
3004         btrfs_end_transaction(trans);
3005         btrfs_btree_balance_dirty(fs_info);
3006 out:
3007         iput(inode);
3008         return ret;
3009 }
3010
3011 /*
3012  * Locate the free space cache EXTENT_DATA in root tree leaf and delete the
3013  * cache inode, to avoid free space cache data extent blocking data relocation.
3014  */
3015 static int delete_v1_space_cache(struct extent_buffer *leaf,
3016                                  struct btrfs_block_group *block_group,
3017                                  u64 data_bytenr)
3018 {
3019         u64 space_cache_ino;
3020         struct btrfs_file_extent_item *ei;
3021         struct btrfs_key key;
3022         bool found = false;
3023         int i;
3024         int ret;
3025
3026         if (btrfs_header_owner(leaf) != BTRFS_ROOT_TREE_OBJECTID)
3027                 return 0;
3028
3029         for (i = 0; i < btrfs_header_nritems(leaf); i++) {
3030                 btrfs_item_key_to_cpu(leaf, &key, i);
3031                 if (key.type != BTRFS_EXTENT_DATA_KEY)
3032                         continue;
3033                 ei = btrfs_item_ptr(leaf, i, struct btrfs_file_extent_item);
3034                 if (btrfs_file_extent_type(leaf, ei) == BTRFS_FILE_EXTENT_REG &&
3035                     btrfs_file_extent_disk_bytenr(leaf, ei) == data_bytenr) {
3036                         found = true;
3037                         space_cache_ino = key.objectid;
3038                         break;
3039                 }
3040         }
3041         if (!found)
3042                 return -ENOENT;
3043         ret = delete_block_group_cache(leaf->fs_info, block_group, NULL,
3044                                         space_cache_ino);
3045         return ret;
3046 }
3047
3048 /*
3049  * helper to find all tree blocks that reference a given data extent
3050  */
3051 static noinline_for_stack
3052 int add_data_references(struct reloc_control *rc,
3053                         struct btrfs_key *extent_key,
3054                         struct btrfs_path *path,
3055                         struct rb_root *blocks)
3056 {
3057         struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
3058         struct ulist *leaves = NULL;
3059         struct ulist_iterator leaf_uiter;
3060         struct ulist_node *ref_node = NULL;
3061         const u32 blocksize = fs_info->nodesize;
3062         int ret = 0;
3063
3064         btrfs_release_path(path);
3065         ret = btrfs_find_all_leafs(NULL, fs_info, extent_key->objectid,
3066                                    0, &leaves, NULL, true);
3067         if (ret < 0)
3068                 return ret;
3069
3070         ULIST_ITER_INIT(&leaf_uiter);
3071         while ((ref_node = ulist_next(leaves, &leaf_uiter))) {
3072                 struct extent_buffer *eb;
3073
3074                 eb = read_tree_block(fs_info, ref_node->val, 0, 0, NULL);
3075                 if (IS_ERR(eb)) {
3076                         ret = PTR_ERR(eb);
3077                         break;
3078                 }
3079                 ret = delete_v1_space_cache(eb, rc->block_group,
3080                                             extent_key->objectid);
3081                 free_extent_buffer(eb);
3082                 if (ret < 0)
3083                         break;
3084                 ret = __add_tree_block(rc, ref_node->val, blocksize, blocks);
3085                 if (ret < 0)
3086                         break;
3087         }
3088         if (ret < 0)
3089                 free_block_list(blocks);
3090         ulist_free(leaves);
3091         return ret;
3092 }
3093
3094 /*
3095  * helper to find next unprocessed extent
3096  */
3097 static noinline_for_stack
3098 int find_next_extent(struct reloc_control *rc, struct btrfs_path *path,
3099                      struct btrfs_key *extent_key)
3100 {
3101         struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
3102         struct btrfs_key key;
3103         struct extent_buffer *leaf;
3104         u64 start, end, last;
3105         int ret;
3106
3107         last = rc->block_group->start + rc->block_group->length;
3108         while (1) {
3109                 cond_resched();
3110                 if (rc->search_start >= last) {
3111                         ret = 1;
3112                         break;
3113                 }
3114
3115                 key.objectid = rc->search_start;
3116                 key.type = BTRFS_EXTENT_ITEM_KEY;
3117                 key.offset = 0;
3118
3119                 path->search_commit_root = 1;
3120                 path->skip_locking = 1;
3121                 ret = btrfs_search_slot(NULL, rc->extent_root, &key, path,
3122                                         0, 0);
3123                 if (ret < 0)
3124                         break;
3125 next:
3126                 leaf = path->nodes[0];
3127                 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
3128                         ret = btrfs_next_leaf(rc->extent_root, path);
3129                         if (ret != 0)
3130                                 break;
3131                         leaf = path->nodes[0];
3132                 }
3133
3134                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
3135                 if (key.objectid >= last) {
3136                         ret = 1;
3137                         break;
3138                 }
3139
3140                 if (key.type != BTRFS_EXTENT_ITEM_KEY &&
3141                     key.type != BTRFS_METADATA_ITEM_KEY) {
3142                         path->slots[0]++;
3143                         goto next;
3144                 }
3145
3146                 if (key.type == BTRFS_EXTENT_ITEM_KEY &&
3147                     key.objectid + key.offset <= rc->search_start) {
3148                         path->slots[0]++;
3149                         goto next;
3150                 }
3151
3152                 if (key.type == BTRFS_METADATA_ITEM_KEY &&
3153                     key.objectid + fs_info->nodesize <=
3154                     rc->search_start) {
3155                         path->slots[0]++;
3156                         goto next;
3157                 }
3158
3159                 ret = find_first_extent_bit(&rc->processed_blocks,
3160                                             key.objectid, &start, &end,
3161                                             EXTENT_DIRTY, NULL);
3162
3163                 if (ret == 0 && start <= key.objectid) {
3164                         btrfs_release_path(path);
3165                         rc->search_start = end + 1;
3166                 } else {
3167                         if (key.type == BTRFS_EXTENT_ITEM_KEY)
3168                                 rc->search_start = key.objectid + key.offset;
3169                         else
3170                                 rc->search_start = key.objectid +
3171                                         fs_info->nodesize;
3172                         memcpy(extent_key, &key, sizeof(key));
3173                         return 0;
3174                 }
3175         }
3176         btrfs_release_path(path);
3177         return ret;
3178 }
3179
3180 static void set_reloc_control(struct reloc_control *rc)
3181 {
3182         struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
3183
3184         mutex_lock(&fs_info->reloc_mutex);
3185         fs_info->reloc_ctl = rc;
3186         mutex_unlock(&fs_info->reloc_mutex);
3187 }
3188
3189 static void unset_reloc_control(struct reloc_control *rc)
3190 {
3191         struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
3192
3193         mutex_lock(&fs_info->reloc_mutex);
3194         fs_info->reloc_ctl = NULL;
3195         mutex_unlock(&fs_info->reloc_mutex);
3196 }
3197
3198 static int check_extent_flags(u64 flags)
3199 {
3200         if ((flags & BTRFS_EXTENT_FLAG_DATA) &&
3201             (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK))
3202                 return 1;
3203         if (!(flags & BTRFS_EXTENT_FLAG_DATA) &&
3204             !(flags & BTRFS_EXTENT_FLAG_TREE_BLOCK))
3205                 return 1;
3206         if ((flags & BTRFS_EXTENT_FLAG_DATA) &&
3207             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
3208                 return 1;
3209         return 0;
3210 }
3211
3212 static noinline_for_stack
3213 int prepare_to_relocate(struct reloc_control *rc)
3214 {
3215         struct btrfs_trans_handle *trans;
3216         int ret;
3217
3218         rc->block_rsv = btrfs_alloc_block_rsv(rc->extent_root->fs_info,
3219                                               BTRFS_BLOCK_RSV_TEMP);
3220         if (!rc->block_rsv)
3221                 return -ENOMEM;
3222
3223         memset(&rc->cluster, 0, sizeof(rc->cluster));
3224         rc->search_start = rc->block_group->start;
3225         rc->extents_found = 0;
3226         rc->nodes_relocated = 0;
3227         rc->merging_rsv_size = 0;
3228         rc->reserved_bytes = 0;
3229         rc->block_rsv->size = rc->extent_root->fs_info->nodesize *
3230                               RELOCATION_RESERVED_NODES;
3231         ret = btrfs_block_rsv_refill(rc->extent_root,
3232                                      rc->block_rsv, rc->block_rsv->size,
3233                                      BTRFS_RESERVE_FLUSH_ALL);
3234         if (ret)
3235                 return ret;
3236
3237         rc->create_reloc_tree = 1;
3238         set_reloc_control(rc);
3239
3240         trans = btrfs_join_transaction(rc->extent_root);
3241         if (IS_ERR(trans)) {
3242                 unset_reloc_control(rc);
3243                 /*
3244                  * extent tree is not a ref_cow tree and has no reloc_root to
3245                  * cleanup.  And callers are responsible to free the above
3246                  * block rsv.
3247                  */
3248                 return PTR_ERR(trans);
3249         }
3250         btrfs_commit_transaction(trans);
3251         return 0;
3252 }
3253
3254 static noinline_for_stack int relocate_block_group(struct reloc_control *rc)
3255 {
3256         struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
3257         struct rb_root blocks = RB_ROOT;
3258         struct btrfs_key key;
3259         struct btrfs_trans_handle *trans = NULL;
3260         struct btrfs_path *path;
3261         struct btrfs_extent_item *ei;
3262         u64 flags;
3263         u32 item_size;
3264         int ret;
3265         int err = 0;
3266         int progress = 0;
3267
3268         path = btrfs_alloc_path();
3269         if (!path)
3270                 return -ENOMEM;
3271         path->reada = READA_FORWARD;
3272
3273         ret = prepare_to_relocate(rc);
3274         if (ret) {
3275                 err = ret;
3276                 goto out_free;
3277         }
3278
3279         while (1) {
3280                 rc->reserved_bytes = 0;
3281                 ret = btrfs_block_rsv_refill(rc->extent_root,
3282                                         rc->block_rsv, rc->block_rsv->size,
3283                                         BTRFS_RESERVE_FLUSH_ALL);
3284                 if (ret) {
3285                         err = ret;
3286                         break;
3287                 }
3288                 progress++;
3289                 trans = btrfs_start_transaction(rc->extent_root, 0);
3290                 if (IS_ERR(trans)) {
3291                         err = PTR_ERR(trans);
3292                         trans = NULL;
3293                         break;
3294                 }
3295 restart:
3296                 if (update_backref_cache(trans, &rc->backref_cache)) {
3297                         btrfs_end_transaction(trans);
3298                         trans = NULL;
3299                         continue;
3300                 }
3301
3302                 ret = find_next_extent(rc, path, &key);
3303                 if (ret < 0)
3304                         err = ret;
3305                 if (ret != 0)
3306                         break;
3307
3308                 rc->extents_found++;
3309
3310                 ei = btrfs_item_ptr(path->nodes[0], path->slots[0],
3311                                     struct btrfs_extent_item);
3312                 item_size = btrfs_item_size_nr(path->nodes[0], path->slots[0]);
3313                 if (item_size >= sizeof(*ei)) {
3314                         flags = btrfs_extent_flags(path->nodes[0], ei);
3315                         ret = check_extent_flags(flags);
3316                         BUG_ON(ret);
3317                 } else if (unlikely(item_size == sizeof(struct btrfs_extent_item_v0))) {
3318                         err = -EINVAL;
3319                         btrfs_print_v0_err(trans->fs_info);
3320                         btrfs_abort_transaction(trans, err);
3321                         break;
3322                 } else {
3323                         BUG();
3324                 }
3325
3326                 if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
3327                         ret = add_tree_block(rc, &key, path, &blocks);
3328                 } else if (rc->stage == UPDATE_DATA_PTRS &&
3329                            (flags & BTRFS_EXTENT_FLAG_DATA)) {
3330                         ret = add_data_references(rc, &key, path, &blocks);
3331                 } else {
3332                         btrfs_release_path(path);
3333                         ret = 0;
3334                 }
3335                 if (ret < 0) {
3336                         err = ret;
3337                         break;
3338                 }
3339
3340                 if (!RB_EMPTY_ROOT(&blocks)) {
3341                         ret = relocate_tree_blocks(trans, rc, &blocks);
3342                         if (ret < 0) {
3343                                 if (ret != -EAGAIN) {
3344                                         err = ret;
3345                                         break;
3346                                 }
3347                                 rc->extents_found--;
3348                                 rc->search_start = key.objectid;
3349                         }
3350                 }
3351
3352                 btrfs_end_transaction_throttle(trans);
3353                 btrfs_btree_balance_dirty(fs_info);
3354                 trans = NULL;
3355
3356                 if (rc->stage == MOVE_DATA_EXTENTS &&
3357                     (flags & BTRFS_EXTENT_FLAG_DATA)) {
3358                         rc->found_file_extent = 1;
3359                         ret = relocate_data_extent(rc->data_inode,
3360                                                    &key, &rc->cluster);
3361                         if (ret < 0) {
3362                                 err = ret;
3363                                 break;
3364                         }
3365                 }
3366                 if (btrfs_should_cancel_balance(fs_info)) {
3367                         err = -ECANCELED;
3368                         break;
3369                 }
3370         }
3371         if (trans && progress && err == -ENOSPC) {
3372                 ret = btrfs_force_chunk_alloc(trans, rc->block_group->flags);
3373                 if (ret == 1) {
3374                         err = 0;
3375                         progress = 0;
3376                         goto restart;
3377                 }
3378         }
3379
3380         btrfs_release_path(path);
3381         clear_extent_bits(&rc->processed_blocks, 0, (u64)-1, EXTENT_DIRTY);
3382
3383         if (trans) {
3384                 btrfs_end_transaction_throttle(trans);
3385                 btrfs_btree_balance_dirty(fs_info);
3386         }
3387
3388         if (!err) {
3389                 ret = relocate_file_extent_cluster(rc->data_inode,
3390                                                    &rc->cluster);
3391                 if (ret < 0)
3392                         err = ret;
3393         }
3394
3395         rc->create_reloc_tree = 0;
3396         set_reloc_control(rc);
3397
3398         btrfs_backref_release_cache(&rc->backref_cache);
3399         btrfs_block_rsv_release(fs_info, rc->block_rsv, (u64)-1, NULL);
3400
3401         /*
3402          * Even in the case when the relocation is cancelled, we should all go
3403          * through prepare_to_merge() and merge_reloc_roots().
3404          *
3405          * For error (including cancelled balance), prepare_to_merge() will
3406          * mark all reloc trees orphan, then queue them for cleanup in
3407          * merge_reloc_roots()
3408          */
3409         err = prepare_to_merge(rc, err);
3410
3411         merge_reloc_roots(rc);
3412
3413         rc->merge_reloc_tree = 0;
3414         unset_reloc_control(rc);
3415         btrfs_block_rsv_release(fs_info, rc->block_rsv, (u64)-1, NULL);
3416
3417         /* get rid of pinned extents */
3418         trans = btrfs_join_transaction(rc->extent_root);
3419         if (IS_ERR(trans)) {
3420                 err = PTR_ERR(trans);
3421                 goto out_free;
3422         }
3423         btrfs_commit_transaction(trans);
3424 out_free:
3425         ret = clean_dirty_subvols(rc);
3426         if (ret < 0 && !err)
3427                 err = ret;
3428         btrfs_free_block_rsv(fs_info, rc->block_rsv);
3429         btrfs_free_path(path);
3430         return err;
3431 }
3432
3433 static int __insert_orphan_inode(struct btrfs_trans_handle *trans,
3434                                  struct btrfs_root *root, u64 objectid)
3435 {
3436         struct btrfs_path *path;
3437         struct btrfs_inode_item *item;
3438         struct extent_buffer *leaf;
3439         int ret;
3440
3441         path = btrfs_alloc_path();
3442         if (!path)
3443                 return -ENOMEM;
3444
3445         ret = btrfs_insert_empty_inode(trans, root, path, objectid);
3446         if (ret)
3447                 goto out;
3448
3449         leaf = path->nodes[0];
3450         item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_inode_item);
3451         memzero_extent_buffer(leaf, (unsigned long)item, sizeof(*item));
3452         btrfs_set_inode_generation(leaf, item, 1);
3453         btrfs_set_inode_size(leaf, item, 0);
3454         btrfs_set_inode_mode(leaf, item, S_IFREG | 0600);
3455         btrfs_set_inode_flags(leaf, item, BTRFS_INODE_NOCOMPRESS |
3456                                           BTRFS_INODE_PREALLOC);
3457         btrfs_mark_buffer_dirty(leaf);
3458 out:
3459         btrfs_free_path(path);
3460         return ret;
3461 }
3462
3463 /*
3464  * helper to create inode for data relocation.
3465  * the inode is in data relocation tree and its link count is 0
3466  */
3467 static noinline_for_stack
3468 struct inode *create_reloc_inode(struct btrfs_fs_info *fs_info,
3469                                  struct btrfs_block_group *group)
3470 {
3471         struct inode *inode = NULL;
3472         struct btrfs_trans_handle *trans;
3473         struct btrfs_root *root;
3474         u64 objectid;
3475         int err = 0;
3476
3477         root = btrfs_grab_root(fs_info->data_reloc_root);
3478         trans = btrfs_start_transaction(root, 6);
3479         if (IS_ERR(trans)) {
3480                 btrfs_put_root(root);
3481                 return ERR_CAST(trans);
3482         }
3483
3484         err = btrfs_find_free_objectid(root, &objectid);
3485         if (err)
3486                 goto out;
3487
3488         err = __insert_orphan_inode(trans, root, objectid);
3489         BUG_ON(err);
3490
3491         inode = btrfs_iget(fs_info->sb, objectid, root);
3492         BUG_ON(IS_ERR(inode));
3493         BTRFS_I(inode)->index_cnt = group->start;
3494
3495         err = btrfs_orphan_add(trans, BTRFS_I(inode));
3496 out:
3497         btrfs_put_root(root);
3498         btrfs_end_transaction(trans);
3499         btrfs_btree_balance_dirty(fs_info);
3500         if (err) {
3501                 if (inode)
3502                         iput(inode);
3503                 inode = ERR_PTR(err);
3504         }
3505         return inode;
3506 }
3507
3508 static struct reloc_control *alloc_reloc_control(struct btrfs_fs_info *fs_info)
3509 {
3510         struct reloc_control *rc;
3511
3512         rc = kzalloc(sizeof(*rc), GFP_NOFS);
3513         if (!rc)
3514                 return NULL;
3515
3516         INIT_LIST_HEAD(&rc->reloc_roots);
3517         INIT_LIST_HEAD(&rc->dirty_subvol_roots);
3518         btrfs_backref_init_cache(fs_info, &rc->backref_cache, 1);
3519         mapping_tree_init(&rc->reloc_root_tree);
3520         extent_io_tree_init(fs_info, &rc->processed_blocks,
3521                             IO_TREE_RELOC_BLOCKS, NULL);
3522         return rc;
3523 }
3524
3525 static void free_reloc_control(struct reloc_control *rc)
3526 {
3527         struct mapping_node *node, *tmp;
3528
3529         free_reloc_roots(&rc->reloc_roots);
3530         rbtree_postorder_for_each_entry_safe(node, tmp,
3531                         &rc->reloc_root_tree.rb_root, rb_node)
3532                 kfree(node);
3533
3534         kfree(rc);
3535 }
3536
3537 /*
3538  * Print the block group being relocated
3539  */
3540 static void describe_relocation(struct btrfs_fs_info *fs_info,
3541                                 struct btrfs_block_group *block_group)
3542 {
3543         char buf[128] = {'\0'};
3544
3545         btrfs_describe_block_groups(block_group->flags, buf, sizeof(buf));
3546
3547         btrfs_info(fs_info,
3548                    "relocating block group %llu flags %s",
3549                    block_group->start, buf);
3550 }
3551
3552 static const char *stage_to_string(int stage)
3553 {
3554         if (stage == MOVE_DATA_EXTENTS)
3555                 return "move data extents";
3556         if (stage == UPDATE_DATA_PTRS)
3557                 return "update data pointers";
3558         return "unknown";
3559 }
3560
3561 /*
3562  * function to relocate all extents in a block group.
3563  */
3564 int btrfs_relocate_block_group(struct btrfs_fs_info *fs_info, u64 group_start)
3565 {
3566         struct btrfs_block_group *bg;
3567         struct btrfs_root *extent_root = fs_info->extent_root;
3568         struct reloc_control *rc;
3569         struct inode *inode;
3570         struct btrfs_path *path;
3571         int ret;
3572         int rw = 0;
3573         int err = 0;
3574
3575         bg = btrfs_lookup_block_group(fs_info, group_start);
3576         if (!bg)
3577                 return -ENOENT;
3578
3579         if (btrfs_pinned_by_swapfile(fs_info, bg)) {
3580                 btrfs_put_block_group(bg);
3581                 return -ETXTBSY;
3582         }
3583
3584         rc = alloc_reloc_control(fs_info);
3585         if (!rc) {
3586                 btrfs_put_block_group(bg);
3587                 return -ENOMEM;
3588         }
3589
3590         rc->extent_root = extent_root;
3591         rc->block_group = bg;
3592
3593         ret = btrfs_inc_block_group_ro(rc->block_group, true);
3594         if (ret) {
3595                 err = ret;
3596                 goto out;
3597         }
3598         rw = 1;
3599
3600         path = btrfs_alloc_path();
3601         if (!path) {
3602                 err = -ENOMEM;
3603                 goto out;
3604         }
3605
3606         inode = lookup_free_space_inode(rc->block_group, path);
3607         btrfs_free_path(path);
3608
3609         if (!IS_ERR(inode))
3610                 ret = delete_block_group_cache(fs_info, rc->block_group, inode, 0);
3611         else
3612                 ret = PTR_ERR(inode);
3613
3614         if (ret && ret != -ENOENT) {
3615                 err = ret;
3616                 goto out;
3617         }
3618
3619         rc->data_inode = create_reloc_inode(fs_info, rc->block_group);
3620         if (IS_ERR(rc->data_inode)) {
3621                 err = PTR_ERR(rc->data_inode);
3622                 rc->data_inode = NULL;
3623                 goto out;
3624         }
3625
3626         describe_relocation(fs_info, rc->block_group);
3627
3628         btrfs_wait_block_group_reservations(rc->block_group);
3629         btrfs_wait_nocow_writers(rc->block_group);
3630         btrfs_wait_ordered_roots(fs_info, U64_MAX,
3631                                  rc->block_group->start,
3632                                  rc->block_group->length);
3633
3634         while (1) {
3635                 int finishes_stage;
3636
3637                 mutex_lock(&fs_info->cleaner_mutex);
3638                 ret = relocate_block_group(rc);
3639                 mutex_unlock(&fs_info->cleaner_mutex);
3640                 if (ret < 0)
3641                         err = ret;
3642
3643                 finishes_stage = rc->stage;
3644                 /*
3645                  * We may have gotten ENOSPC after we already dirtied some
3646                  * extents.  If writeout happens while we're relocating a
3647                  * different block group we could end up hitting the
3648                  * BUG_ON(rc->stage == UPDATE_DATA_PTRS) in
3649                  * btrfs_reloc_cow_block.  Make sure we write everything out
3650                  * properly so we don't trip over this problem, and then break
3651                  * out of the loop if we hit an error.
3652                  */
3653                 if (rc->stage == MOVE_DATA_EXTENTS && rc->found_file_extent) {
3654                         ret = btrfs_wait_ordered_range(rc->data_inode, 0,
3655                                                        (u64)-1);
3656                         if (ret)
3657                                 err = ret;
3658                         invalidate_mapping_pages(rc->data_inode->i_mapping,
3659                                                  0, -1);
3660                         rc->stage = UPDATE_DATA_PTRS;
3661                 }
3662
3663                 if (err < 0)
3664                         goto out;
3665
3666                 if (rc->extents_found == 0)
3667                         break;
3668
3669                 btrfs_info(fs_info, "found %llu extents, stage: %s",
3670                            rc->extents_found, stage_to_string(finishes_stage));
3671         }
3672
3673         WARN_ON(rc->block_group->pinned > 0);
3674         WARN_ON(rc->block_group->reserved > 0);
3675         WARN_ON(rc->block_group->used > 0);
3676 out:
3677         if (err && rw)
3678                 btrfs_dec_block_group_ro(rc->block_group);
3679         iput(rc->data_inode);
3680         btrfs_put_block_group(rc->block_group);
3681         free_reloc_control(rc);
3682         return err;
3683 }
3684
3685 static noinline_for_stack int mark_garbage_root(struct btrfs_root *root)
3686 {
3687         struct btrfs_fs_info *fs_info = root->fs_info;
3688         struct btrfs_trans_handle *trans;
3689         int ret, err;
3690
3691         trans = btrfs_start_transaction(fs_info->tree_root, 0);
3692         if (IS_ERR(trans))
3693                 return PTR_ERR(trans);
3694
3695         memset(&root->root_item.drop_progress, 0,
3696                 sizeof(root->root_item.drop_progress));
3697         root->root_item.drop_level = 0;
3698         btrfs_set_root_refs(&root->root_item, 0);
3699         ret = btrfs_update_root(trans, fs_info->tree_root,
3700                                 &root->root_key, &root->root_item);
3701
3702         err = btrfs_end_transaction(trans);
3703         if (err)
3704                 return err;
3705         return ret;
3706 }
3707
3708 /*
3709  * recover relocation interrupted by system crash.
3710  *
3711  * this function resumes merging reloc trees with corresponding fs trees.
3712  * this is important for keeping the sharing of tree blocks
3713  */
3714 int btrfs_recover_relocation(struct btrfs_root *root)
3715 {
3716         struct btrfs_fs_info *fs_info = root->fs_info;
3717         LIST_HEAD(reloc_roots);
3718         struct btrfs_key key;
3719         struct btrfs_root *fs_root;
3720         struct btrfs_root *reloc_root;
3721         struct btrfs_path *path;
3722         struct extent_buffer *leaf;
3723         struct reloc_control *rc = NULL;
3724         struct btrfs_trans_handle *trans;
3725         int ret;
3726         int err = 0;
3727
3728         path = btrfs_alloc_path();
3729         if (!path)
3730                 return -ENOMEM;
3731         path->reada = READA_BACK;
3732
3733         key.objectid = BTRFS_TREE_RELOC_OBJECTID;
3734         key.type = BTRFS_ROOT_ITEM_KEY;
3735         key.offset = (u64)-1;
3736
3737         while (1) {
3738                 ret = btrfs_search_slot(NULL, fs_info->tree_root, &key,
3739                                         path, 0, 0);
3740                 if (ret < 0) {
3741                         err = ret;
3742                         goto out;
3743                 }
3744                 if (ret > 0) {
3745                         if (path->slots[0] == 0)
3746                                 break;
3747                         path->slots[0]--;
3748                 }
3749                 leaf = path->nodes[0];
3750                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
3751                 btrfs_release_path(path);
3752
3753                 if (key.objectid != BTRFS_TREE_RELOC_OBJECTID ||
3754                     key.type != BTRFS_ROOT_ITEM_KEY)
3755                         break;
3756
3757                 reloc_root = btrfs_read_tree_root(root, &key);
3758                 if (IS_ERR(reloc_root)) {
3759                         err = PTR_ERR(reloc_root);
3760                         goto out;
3761                 }
3762
3763                 set_bit(BTRFS_ROOT_SHAREABLE, &reloc_root->state);
3764                 list_add(&reloc_root->root_list, &reloc_roots);
3765
3766                 if (btrfs_root_refs(&reloc_root->root_item) > 0) {
3767                         fs_root = btrfs_get_fs_root(fs_info,
3768                                         reloc_root->root_key.offset, false);
3769                         if (IS_ERR(fs_root)) {
3770                                 ret = PTR_ERR(fs_root);
3771                                 if (ret != -ENOENT) {
3772                                         err = ret;
3773                                         goto out;
3774                                 }
3775                                 ret = mark_garbage_root(reloc_root);
3776                                 if (ret < 0) {
3777                                         err = ret;
3778                                         goto out;
3779                                 }
3780                         } else {
3781                                 btrfs_put_root(fs_root);
3782                         }
3783                 }
3784
3785                 if (key.offset == 0)
3786                         break;
3787
3788                 key.offset--;
3789         }
3790         btrfs_release_path(path);
3791
3792         if (list_empty(&reloc_roots))
3793                 goto out;
3794
3795         rc = alloc_reloc_control(fs_info);
3796         if (!rc) {
3797                 err = -ENOMEM;
3798                 goto out;
3799         }
3800
3801         rc->extent_root = fs_info->extent_root;
3802
3803         set_reloc_control(rc);
3804
3805         trans = btrfs_join_transaction(rc->extent_root);
3806         if (IS_ERR(trans)) {
3807                 err = PTR_ERR(trans);
3808                 goto out_unset;
3809         }
3810
3811         rc->merge_reloc_tree = 1;
3812
3813         while (!list_empty(&reloc_roots)) {
3814                 reloc_root = list_entry(reloc_roots.next,
3815                                         struct btrfs_root, root_list);
3816                 list_del(&reloc_root->root_list);
3817
3818                 if (btrfs_root_refs(&reloc_root->root_item) == 0) {
3819                         list_add_tail(&reloc_root->root_list,
3820                                       &rc->reloc_roots);
3821                         continue;
3822                 }
3823
3824                 fs_root = btrfs_get_fs_root(fs_info, reloc_root->root_key.offset,
3825                                             false);
3826                 if (IS_ERR(fs_root)) {
3827                         err = PTR_ERR(fs_root);
3828                         list_add_tail(&reloc_root->root_list, &reloc_roots);
3829                         btrfs_end_transaction(trans);
3830                         goto out_unset;
3831                 }
3832
3833                 err = __add_reloc_root(reloc_root);
3834                 BUG_ON(err < 0); /* -ENOMEM or logic error */
3835                 fs_root->reloc_root = btrfs_grab_root(reloc_root);
3836                 btrfs_put_root(fs_root);
3837         }
3838
3839         err = btrfs_commit_transaction(trans);
3840         if (err)
3841                 goto out_unset;
3842
3843         merge_reloc_roots(rc);
3844
3845         unset_reloc_control(rc);
3846
3847         trans = btrfs_join_transaction(rc->extent_root);
3848         if (IS_ERR(trans)) {
3849                 err = PTR_ERR(trans);
3850                 goto out_clean;
3851         }
3852         err = btrfs_commit_transaction(trans);
3853 out_clean:
3854         ret = clean_dirty_subvols(rc);
3855         if (ret < 0 && !err)
3856                 err = ret;
3857 out_unset:
3858         unset_reloc_control(rc);
3859         free_reloc_control(rc);
3860 out:
3861         free_reloc_roots(&reloc_roots);
3862
3863         btrfs_free_path(path);
3864
3865         if (err == 0) {
3866                 /* cleanup orphan inode in data relocation tree */
3867                 fs_root = btrfs_grab_root(fs_info->data_reloc_root);
3868                 ASSERT(fs_root);
3869                 err = btrfs_orphan_cleanup(fs_root);
3870                 btrfs_put_root(fs_root);
3871         }
3872         return err;
3873 }
3874
3875 /*
3876  * helper to add ordered checksum for data relocation.
3877  *
3878  * cloning checksum properly handles the nodatasum extents.
3879  * it also saves CPU time to re-calculate the checksum.
3880  */
3881 int btrfs_reloc_clone_csums(struct btrfs_inode *inode, u64 file_pos, u64 len)
3882 {
3883         struct btrfs_fs_info *fs_info = inode->root->fs_info;
3884         struct btrfs_ordered_sum *sums;
3885         struct btrfs_ordered_extent *ordered;
3886         int ret;
3887         u64 disk_bytenr;
3888         u64 new_bytenr;
3889         LIST_HEAD(list);
3890
3891         ordered = btrfs_lookup_ordered_extent(inode, file_pos);
3892         BUG_ON(ordered->file_offset != file_pos || ordered->num_bytes != len);
3893
3894         disk_bytenr = file_pos + inode->index_cnt;
3895         ret = btrfs_lookup_csums_range(fs_info->csum_root, disk_bytenr,
3896                                        disk_bytenr + len - 1, &list, 0);
3897         if (ret)
3898                 goto out;
3899
3900         while (!list_empty(&list)) {
3901                 sums = list_entry(list.next, struct btrfs_ordered_sum, list);
3902                 list_del_init(&sums->list);
3903
3904                 /*
3905                  * We need to offset the new_bytenr based on where the csum is.
3906                  * We need to do this because we will read in entire prealloc
3907                  * extents but we may have written to say the middle of the
3908                  * prealloc extent, so we need to make sure the csum goes with
3909                  * the right disk offset.
3910                  *
3911                  * We can do this because the data reloc inode refers strictly
3912                  * to the on disk bytes, so we don't have to worry about
3913                  * disk_len vs real len like with real inodes since it's all
3914                  * disk length.
3915                  */
3916                 new_bytenr = ordered->disk_bytenr + sums->bytenr - disk_bytenr;
3917                 sums->bytenr = new_bytenr;
3918
3919                 btrfs_add_ordered_sum(ordered, sums);
3920         }
3921 out:
3922         btrfs_put_ordered_extent(ordered);
3923         return ret;
3924 }
3925
3926 int btrfs_reloc_cow_block(struct btrfs_trans_handle *trans,
3927                           struct btrfs_root *root, struct extent_buffer *buf,
3928                           struct extent_buffer *cow)
3929 {
3930         struct btrfs_fs_info *fs_info = root->fs_info;
3931         struct reloc_control *rc;
3932         struct btrfs_backref_node *node;
3933         int first_cow = 0;
3934         int level;
3935         int ret = 0;
3936
3937         rc = fs_info->reloc_ctl;
3938         if (!rc)
3939                 return 0;
3940
3941         BUG_ON(rc->stage == UPDATE_DATA_PTRS &&
3942                root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID);
3943
3944         level = btrfs_header_level(buf);
3945         if (btrfs_header_generation(buf) <=
3946             btrfs_root_last_snapshot(&root->root_item))
3947                 first_cow = 1;
3948
3949         if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID &&
3950             rc->create_reloc_tree) {
3951                 WARN_ON(!first_cow && level == 0);
3952
3953                 node = rc->backref_cache.path[level];
3954                 BUG_ON(node->bytenr != buf->start &&
3955                        node->new_bytenr != buf->start);
3956
3957                 btrfs_backref_drop_node_buffer(node);
3958                 atomic_inc(&cow->refs);
3959                 node->eb = cow;
3960                 node->new_bytenr = cow->start;
3961
3962                 if (!node->pending) {
3963                         list_move_tail(&node->list,
3964                                        &rc->backref_cache.pending[level]);
3965                         node->pending = 1;
3966                 }
3967
3968                 if (first_cow)
3969                         mark_block_processed(rc, node);
3970
3971                 if (first_cow && level > 0)
3972                         rc->nodes_relocated += buf->len;
3973         }
3974
3975         if (level == 0 && first_cow && rc->stage == UPDATE_DATA_PTRS)
3976                 ret = replace_file_extents(trans, rc, root, cow);
3977         return ret;
3978 }
3979
3980 /*
3981  * called before creating snapshot. it calculates metadata reservation
3982  * required for relocating tree blocks in the snapshot
3983  */
3984 void btrfs_reloc_pre_snapshot(struct btrfs_pending_snapshot *pending,
3985                               u64 *bytes_to_reserve)
3986 {
3987         struct btrfs_root *root = pending->root;
3988         struct reloc_control *rc = root->fs_info->reloc_ctl;
3989
3990         if (!rc || !have_reloc_root(root))
3991                 return;
3992
3993         if (!rc->merge_reloc_tree)
3994                 return;
3995
3996         root = root->reloc_root;
3997         BUG_ON(btrfs_root_refs(&root->root_item) == 0);
3998         /*
3999          * relocation is in the stage of merging trees. the space
4000          * used by merging a reloc tree is twice the size of
4001          * relocated tree nodes in the worst case. half for cowing
4002          * the reloc tree, half for cowing the fs tree. the space
4003          * used by cowing the reloc tree will be freed after the
4004          * tree is dropped. if we create snapshot, cowing the fs
4005          * tree may use more space than it frees. so we need
4006          * reserve extra space.
4007          */
4008         *bytes_to_reserve += rc->nodes_relocated;
4009 }
4010
4011 /*
4012  * called after snapshot is created. migrate block reservation
4013  * and create reloc root for the newly created snapshot
4014  *
4015  * This is similar to btrfs_init_reloc_root(), we come out of here with two
4016  * references held on the reloc_root, one for root->reloc_root and one for
4017  * rc->reloc_roots.
4018  */
4019 int btrfs_reloc_post_snapshot(struct btrfs_trans_handle *trans,
4020                                struct btrfs_pending_snapshot *pending)
4021 {
4022         struct btrfs_root *root = pending->root;
4023         struct btrfs_root *reloc_root;
4024         struct btrfs_root *new_root;
4025         struct reloc_control *rc = root->fs_info->reloc_ctl;
4026         int ret;
4027
4028         if (!rc || !have_reloc_root(root))
4029                 return 0;
4030
4031         rc = root->fs_info->reloc_ctl;
4032         rc->merging_rsv_size += rc->nodes_relocated;
4033
4034         if (rc->merge_reloc_tree) {
4035                 ret = btrfs_block_rsv_migrate(&pending->block_rsv,
4036                                               rc->block_rsv,
4037                                               rc->nodes_relocated, true);
4038                 if (ret)
4039                         return ret;
4040         }
4041
4042         new_root = pending->snap;
4043         reloc_root = create_reloc_root(trans, root->reloc_root,
4044                                        new_root->root_key.objectid);
4045         if (IS_ERR(reloc_root))
4046                 return PTR_ERR(reloc_root);
4047
4048         ret = __add_reloc_root(reloc_root);
4049         BUG_ON(ret < 0);
4050         new_root->reloc_root = btrfs_grab_root(reloc_root);
4051
4052         if (rc->create_reloc_tree)
4053                 ret = clone_backref_node(trans, rc, root, reloc_root);
4054         return ret;
4055 }