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