Btrfs: improve forever loop when doing balance relocation
[sfrench/cifs-2.6.git] / fs / btrfs / relocation.c
1 /*
2  * Copyright (C) 2009 Oracle.  All rights reserved.
3  *
4  * This program is free software; you can redistribute it and/or
5  * modify it under the terms of the GNU General Public
6  * License v2 as published by the Free Software Foundation.
7  *
8  * This program is distributed in the hope that it will be useful,
9  * but WITHOUT ANY WARRANTY; without even the implied warranty of
10  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
11  * General Public License for more details.
12  *
13  * You should have received a copy of the GNU General Public
14  * License along with this program; if not, write to the
15  * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
16  * Boston, MA 021110-1307, USA.
17  */
18
19 #include <linux/sched.h>
20 #include <linux/pagemap.h>
21 #include <linux/writeback.h>
22 #include <linux/blkdev.h>
23 #include <linux/rbtree.h>
24 #include <linux/slab.h>
25 #include "ctree.h"
26 #include "disk-io.h"
27 #include "transaction.h"
28 #include "volumes.h"
29 #include "locking.h"
30 #include "btrfs_inode.h"
31 #include "async-thread.h"
32 #include "free-space-cache.h"
33 #include "inode-map.h"
34
35 /*
36  * backref_node, mapping_node and tree_block start with this
37  */
38 struct tree_entry {
39         struct rb_node rb_node;
40         u64 bytenr;
41 };
42
43 /*
44  * present a tree block in the backref cache
45  */
46 struct backref_node {
47         struct rb_node rb_node;
48         u64 bytenr;
49
50         u64 new_bytenr;
51         /* objectid of tree block owner, can be not uptodate */
52         u64 owner;
53         /* link to pending, changed or detached list */
54         struct list_head list;
55         /* list of upper level blocks reference this block */
56         struct list_head upper;
57         /* list of child blocks in the cache */
58         struct list_head lower;
59         /* NULL if this node is not tree root */
60         struct btrfs_root *root;
61         /* extent buffer got by COW the block */
62         struct extent_buffer *eb;
63         /* level of tree block */
64         unsigned int level:8;
65         /* is the block in non-reference counted tree */
66         unsigned int cowonly:1;
67         /* 1 if no child node in the cache */
68         unsigned int lowest:1;
69         /* is the extent buffer locked */
70         unsigned int locked:1;
71         /* has the block been processed */
72         unsigned int processed:1;
73         /* have backrefs of this block been checked */
74         unsigned int checked:1;
75         /*
76          * 1 if corresponding block has been cowed but some upper
77          * level block pointers may not point to the new location
78          */
79         unsigned int pending:1;
80         /*
81          * 1 if the backref node isn't connected to any other
82          * backref node.
83          */
84         unsigned int detached:1;
85 };
86
87 /*
88  * present a block pointer in the backref cache
89  */
90 struct backref_edge {
91         struct list_head list[2];
92         struct backref_node *node[2];
93 };
94
95 #define LOWER   0
96 #define UPPER   1
97 #define RELOCATION_RESERVED_NODES       256
98
99 struct backref_cache {
100         /* red black tree of all backref nodes in the cache */
101         struct rb_root rb_root;
102         /* for passing backref nodes to btrfs_reloc_cow_block */
103         struct backref_node *path[BTRFS_MAX_LEVEL];
104         /*
105          * list of blocks that have been cowed but some block
106          * pointers in upper level blocks may not reflect the
107          * new location
108          */
109         struct list_head pending[BTRFS_MAX_LEVEL];
110         /* list of backref nodes with no child node */
111         struct list_head leaves;
112         /* list of blocks that have been cowed in current transaction */
113         struct list_head changed;
114         /* list of detached backref node. */
115         struct list_head detached;
116
117         u64 last_trans;
118
119         int nr_nodes;
120         int nr_edges;
121 };
122
123 /*
124  * map address of tree root to tree
125  */
126 struct mapping_node {
127         struct rb_node rb_node;
128         u64 bytenr;
129         void *data;
130 };
131
132 struct mapping_tree {
133         struct rb_root rb_root;
134         spinlock_t lock;
135 };
136
137 /*
138  * present a tree block to process
139  */
140 struct tree_block {
141         struct rb_node rb_node;
142         u64 bytenr;
143         struct btrfs_key key;
144         unsigned int level:8;
145         unsigned int key_ready:1;
146 };
147
148 #define MAX_EXTENTS 128
149
150 struct file_extent_cluster {
151         u64 start;
152         u64 end;
153         u64 boundary[MAX_EXTENTS];
154         unsigned int nr;
155 };
156
157 struct reloc_control {
158         /* block group to relocate */
159         struct btrfs_block_group_cache *block_group;
160         /* extent tree */
161         struct btrfs_root *extent_root;
162         /* inode for moving data */
163         struct inode *data_inode;
164
165         struct btrfs_block_rsv *block_rsv;
166
167         struct backref_cache backref_cache;
168
169         struct file_extent_cluster cluster;
170         /* tree blocks have been processed */
171         struct extent_io_tree processed_blocks;
172         /* map start of tree root to corresponding reloc tree */
173         struct mapping_tree reloc_root_tree;
174         /* list of reloc trees */
175         struct list_head reloc_roots;
176         /* size of metadata reservation for merging reloc trees */
177         u64 merging_rsv_size;
178         /* size of relocated tree nodes */
179         u64 nodes_relocated;
180         /* reserved size for block group relocation*/
181         u64 reserved_bytes;
182
183         u64 search_start;
184         u64 extents_found;
185
186         unsigned int stage:8;
187         unsigned int create_reloc_tree:1;
188         unsigned int merge_reloc_tree:1;
189         unsigned int found_file_extent:1;
190 };
191
192 /* stages of data relocation */
193 #define MOVE_DATA_EXTENTS       0
194 #define UPDATE_DATA_PTRS        1
195
196 static void remove_backref_node(struct backref_cache *cache,
197                                 struct backref_node *node);
198 static void __mark_block_processed(struct reloc_control *rc,
199                                    struct backref_node *node);
200
201 static void mapping_tree_init(struct mapping_tree *tree)
202 {
203         tree->rb_root = RB_ROOT;
204         spin_lock_init(&tree->lock);
205 }
206
207 static void backref_cache_init(struct backref_cache *cache)
208 {
209         int i;
210         cache->rb_root = RB_ROOT;
211         for (i = 0; i < BTRFS_MAX_LEVEL; i++)
212                 INIT_LIST_HEAD(&cache->pending[i]);
213         INIT_LIST_HEAD(&cache->changed);
214         INIT_LIST_HEAD(&cache->detached);
215         INIT_LIST_HEAD(&cache->leaves);
216 }
217
218 static void backref_cache_cleanup(struct backref_cache *cache)
219 {
220         struct backref_node *node;
221         int i;
222
223         while (!list_empty(&cache->detached)) {
224                 node = list_entry(cache->detached.next,
225                                   struct backref_node, list);
226                 remove_backref_node(cache, node);
227         }
228
229         while (!list_empty(&cache->leaves)) {
230                 node = list_entry(cache->leaves.next,
231                                   struct backref_node, lower);
232                 remove_backref_node(cache, node);
233         }
234
235         cache->last_trans = 0;
236
237         for (i = 0; i < BTRFS_MAX_LEVEL; i++)
238                 BUG_ON(!list_empty(&cache->pending[i]));
239         BUG_ON(!list_empty(&cache->changed));
240         BUG_ON(!list_empty(&cache->detached));
241         BUG_ON(!RB_EMPTY_ROOT(&cache->rb_root));
242         BUG_ON(cache->nr_nodes);
243         BUG_ON(cache->nr_edges);
244 }
245
246 static struct backref_node *alloc_backref_node(struct backref_cache *cache)
247 {
248         struct backref_node *node;
249
250         node = kzalloc(sizeof(*node), GFP_NOFS);
251         if (node) {
252                 INIT_LIST_HEAD(&node->list);
253                 INIT_LIST_HEAD(&node->upper);
254                 INIT_LIST_HEAD(&node->lower);
255                 RB_CLEAR_NODE(&node->rb_node);
256                 cache->nr_nodes++;
257         }
258         return node;
259 }
260
261 static void free_backref_node(struct backref_cache *cache,
262                               struct backref_node *node)
263 {
264         if (node) {
265                 cache->nr_nodes--;
266                 kfree(node);
267         }
268 }
269
270 static struct backref_edge *alloc_backref_edge(struct backref_cache *cache)
271 {
272         struct backref_edge *edge;
273
274         edge = kzalloc(sizeof(*edge), GFP_NOFS);
275         if (edge)
276                 cache->nr_edges++;
277         return edge;
278 }
279
280 static void free_backref_edge(struct backref_cache *cache,
281                               struct backref_edge *edge)
282 {
283         if (edge) {
284                 cache->nr_edges--;
285                 kfree(edge);
286         }
287 }
288
289 static struct rb_node *tree_insert(struct rb_root *root, u64 bytenr,
290                                    struct rb_node *node)
291 {
292         struct rb_node **p = &root->rb_node;
293         struct rb_node *parent = NULL;
294         struct tree_entry *entry;
295
296         while (*p) {
297                 parent = *p;
298                 entry = rb_entry(parent, struct tree_entry, rb_node);
299
300                 if (bytenr < entry->bytenr)
301                         p = &(*p)->rb_left;
302                 else if (bytenr > entry->bytenr)
303                         p = &(*p)->rb_right;
304                 else
305                         return parent;
306         }
307
308         rb_link_node(node, parent, p);
309         rb_insert_color(node, root);
310         return NULL;
311 }
312
313 static struct rb_node *tree_search(struct rb_root *root, u64 bytenr)
314 {
315         struct rb_node *n = root->rb_node;
316         struct tree_entry *entry;
317
318         while (n) {
319                 entry = rb_entry(n, struct tree_entry, rb_node);
320
321                 if (bytenr < entry->bytenr)
322                         n = n->rb_left;
323                 else if (bytenr > entry->bytenr)
324                         n = n->rb_right;
325                 else
326                         return n;
327         }
328         return NULL;
329 }
330
331 static void backref_tree_panic(struct rb_node *rb_node, int errno, u64 bytenr)
332 {
333
334         struct btrfs_fs_info *fs_info = NULL;
335         struct backref_node *bnode = rb_entry(rb_node, struct backref_node,
336                                               rb_node);
337         if (bnode->root)
338                 fs_info = bnode->root->fs_info;
339         btrfs_panic(fs_info, errno, "Inconsistency in backref cache "
340                     "found at offset %llu\n", bytenr);
341 }
342
343 /*
344  * walk up backref nodes until reach node presents tree root
345  */
346 static struct backref_node *walk_up_backref(struct backref_node *node,
347                                             struct backref_edge *edges[],
348                                             int *index)
349 {
350         struct backref_edge *edge;
351         int idx = *index;
352
353         while (!list_empty(&node->upper)) {
354                 edge = list_entry(node->upper.next,
355                                   struct backref_edge, list[LOWER]);
356                 edges[idx++] = edge;
357                 node = edge->node[UPPER];
358         }
359         BUG_ON(node->detached);
360         *index = idx;
361         return node;
362 }
363
364 /*
365  * walk down backref nodes to find start of next reference path
366  */
367 static struct backref_node *walk_down_backref(struct backref_edge *edges[],
368                                               int *index)
369 {
370         struct backref_edge *edge;
371         struct backref_node *lower;
372         int idx = *index;
373
374         while (idx > 0) {
375                 edge = edges[idx - 1];
376                 lower = edge->node[LOWER];
377                 if (list_is_last(&edge->list[LOWER], &lower->upper)) {
378                         idx--;
379                         continue;
380                 }
381                 edge = list_entry(edge->list[LOWER].next,
382                                   struct backref_edge, list[LOWER]);
383                 edges[idx - 1] = edge;
384                 *index = idx;
385                 return edge->node[UPPER];
386         }
387         *index = 0;
388         return NULL;
389 }
390
391 static void unlock_node_buffer(struct backref_node *node)
392 {
393         if (node->locked) {
394                 btrfs_tree_unlock(node->eb);
395                 node->locked = 0;
396         }
397 }
398
399 static void drop_node_buffer(struct backref_node *node)
400 {
401         if (node->eb) {
402                 unlock_node_buffer(node);
403                 free_extent_buffer(node->eb);
404                 node->eb = NULL;
405         }
406 }
407
408 static void drop_backref_node(struct backref_cache *tree,
409                               struct backref_node *node)
410 {
411         BUG_ON(!list_empty(&node->upper));
412
413         drop_node_buffer(node);
414         list_del(&node->list);
415         list_del(&node->lower);
416         if (!RB_EMPTY_NODE(&node->rb_node))
417                 rb_erase(&node->rb_node, &tree->rb_root);
418         free_backref_node(tree, node);
419 }
420
421 /*
422  * remove a backref node from the backref cache
423  */
424 static void remove_backref_node(struct backref_cache *cache,
425                                 struct backref_node *node)
426 {
427         struct backref_node *upper;
428         struct backref_edge *edge;
429
430         if (!node)
431                 return;
432
433         BUG_ON(!node->lowest && !node->detached);
434         while (!list_empty(&node->upper)) {
435                 edge = list_entry(node->upper.next, struct backref_edge,
436                                   list[LOWER]);
437                 upper = edge->node[UPPER];
438                 list_del(&edge->list[LOWER]);
439                 list_del(&edge->list[UPPER]);
440                 free_backref_edge(cache, edge);
441
442                 if (RB_EMPTY_NODE(&upper->rb_node)) {
443                         BUG_ON(!list_empty(&node->upper));
444                         drop_backref_node(cache, node);
445                         node = upper;
446                         node->lowest = 1;
447                         continue;
448                 }
449                 /*
450                  * add the node to leaf node list if no other
451                  * child block cached.
452                  */
453                 if (list_empty(&upper->lower)) {
454                         list_add_tail(&upper->lower, &cache->leaves);
455                         upper->lowest = 1;
456                 }
457         }
458
459         drop_backref_node(cache, node);
460 }
461
462 static void update_backref_node(struct backref_cache *cache,
463                                 struct backref_node *node, u64 bytenr)
464 {
465         struct rb_node *rb_node;
466         rb_erase(&node->rb_node, &cache->rb_root);
467         node->bytenr = bytenr;
468         rb_node = tree_insert(&cache->rb_root, node->bytenr, &node->rb_node);
469         if (rb_node)
470                 backref_tree_panic(rb_node, -EEXIST, bytenr);
471 }
472
473 /*
474  * update backref cache after a transaction commit
475  */
476 static int update_backref_cache(struct btrfs_trans_handle *trans,
477                                 struct backref_cache *cache)
478 {
479         struct backref_node *node;
480         int level = 0;
481
482         if (cache->last_trans == 0) {
483                 cache->last_trans = trans->transid;
484                 return 0;
485         }
486
487         if (cache->last_trans == trans->transid)
488                 return 0;
489
490         /*
491          * detached nodes are used to avoid unnecessary backref
492          * lookup. transaction commit changes the extent tree.
493          * so the detached nodes are no longer useful.
494          */
495         while (!list_empty(&cache->detached)) {
496                 node = list_entry(cache->detached.next,
497                                   struct backref_node, list);
498                 remove_backref_node(cache, node);
499         }
500
501         while (!list_empty(&cache->changed)) {
502                 node = list_entry(cache->changed.next,
503                                   struct backref_node, list);
504                 list_del_init(&node->list);
505                 BUG_ON(node->pending);
506                 update_backref_node(cache, node, node->new_bytenr);
507         }
508
509         /*
510          * some nodes can be left in the pending list if there were
511          * errors during processing the pending nodes.
512          */
513         for (level = 0; level < BTRFS_MAX_LEVEL; level++) {
514                 list_for_each_entry(node, &cache->pending[level], list) {
515                         BUG_ON(!node->pending);
516                         if (node->bytenr == node->new_bytenr)
517                                 continue;
518                         update_backref_node(cache, node, node->new_bytenr);
519                 }
520         }
521
522         cache->last_trans = 0;
523         return 1;
524 }
525
526
527 static int should_ignore_root(struct btrfs_root *root)
528 {
529         struct btrfs_root *reloc_root;
530
531         if (!root->ref_cows)
532                 return 0;
533
534         reloc_root = root->reloc_root;
535         if (!reloc_root)
536                 return 0;
537
538         if (btrfs_root_last_snapshot(&reloc_root->root_item) ==
539             root->fs_info->running_transaction->transid - 1)
540                 return 0;
541         /*
542          * if there is reloc tree and it was created in previous
543          * transaction backref lookup can find the reloc tree,
544          * so backref node for the fs tree root is useless for
545          * relocation.
546          */
547         return 1;
548 }
549 /*
550  * find reloc tree by address of tree root
551  */
552 static struct btrfs_root *find_reloc_root(struct reloc_control *rc,
553                                           u64 bytenr)
554 {
555         struct rb_node *rb_node;
556         struct mapping_node *node;
557         struct btrfs_root *root = NULL;
558
559         spin_lock(&rc->reloc_root_tree.lock);
560         rb_node = tree_search(&rc->reloc_root_tree.rb_root, bytenr);
561         if (rb_node) {
562                 node = rb_entry(rb_node, struct mapping_node, rb_node);
563                 root = (struct btrfs_root *)node->data;
564         }
565         spin_unlock(&rc->reloc_root_tree.lock);
566         return root;
567 }
568
569 static int is_cowonly_root(u64 root_objectid)
570 {
571         if (root_objectid == BTRFS_ROOT_TREE_OBJECTID ||
572             root_objectid == BTRFS_EXTENT_TREE_OBJECTID ||
573             root_objectid == BTRFS_CHUNK_TREE_OBJECTID ||
574             root_objectid == BTRFS_DEV_TREE_OBJECTID ||
575             root_objectid == BTRFS_TREE_LOG_OBJECTID ||
576             root_objectid == BTRFS_CSUM_TREE_OBJECTID ||
577             root_objectid == BTRFS_UUID_TREE_OBJECTID ||
578             root_objectid == BTRFS_QUOTA_TREE_OBJECTID)
579                 return 1;
580         return 0;
581 }
582
583 static struct btrfs_root *read_fs_root(struct btrfs_fs_info *fs_info,
584                                         u64 root_objectid)
585 {
586         struct btrfs_key key;
587
588         key.objectid = root_objectid;
589         key.type = BTRFS_ROOT_ITEM_KEY;
590         if (is_cowonly_root(root_objectid))
591                 key.offset = 0;
592         else
593                 key.offset = (u64)-1;
594
595         return btrfs_get_fs_root(fs_info, &key, false);
596 }
597
598 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
599 static noinline_for_stack
600 struct btrfs_root *find_tree_root(struct reloc_control *rc,
601                                   struct extent_buffer *leaf,
602                                   struct btrfs_extent_ref_v0 *ref0)
603 {
604         struct btrfs_root *root;
605         u64 root_objectid = btrfs_ref_root_v0(leaf, ref0);
606         u64 generation = btrfs_ref_generation_v0(leaf, ref0);
607
608         BUG_ON(root_objectid == BTRFS_TREE_RELOC_OBJECTID);
609
610         root = read_fs_root(rc->extent_root->fs_info, root_objectid);
611         BUG_ON(IS_ERR(root));
612
613         if (root->ref_cows &&
614             generation != btrfs_root_generation(&root->root_item))
615                 return NULL;
616
617         return root;
618 }
619 #endif
620
621 static noinline_for_stack
622 int find_inline_backref(struct extent_buffer *leaf, int slot,
623                         unsigned long *ptr, unsigned long *end)
624 {
625         struct btrfs_key key;
626         struct btrfs_extent_item *ei;
627         struct btrfs_tree_block_info *bi;
628         u32 item_size;
629
630         btrfs_item_key_to_cpu(leaf, &key, slot);
631
632         item_size = btrfs_item_size_nr(leaf, slot);
633 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
634         if (item_size < sizeof(*ei)) {
635                 WARN_ON(item_size != sizeof(struct btrfs_extent_item_v0));
636                 return 1;
637         }
638 #endif
639         ei = btrfs_item_ptr(leaf, slot, struct btrfs_extent_item);
640         WARN_ON(!(btrfs_extent_flags(leaf, ei) &
641                   BTRFS_EXTENT_FLAG_TREE_BLOCK));
642
643         if (key.type == BTRFS_EXTENT_ITEM_KEY &&
644             item_size <= sizeof(*ei) + sizeof(*bi)) {
645                 WARN_ON(item_size < sizeof(*ei) + sizeof(*bi));
646                 return 1;
647         }
648         if (key.type == BTRFS_METADATA_ITEM_KEY &&
649             item_size <= sizeof(*ei)) {
650                 WARN_ON(item_size < sizeof(*ei));
651                 return 1;
652         }
653
654         if (key.type == BTRFS_EXTENT_ITEM_KEY) {
655                 bi = (struct btrfs_tree_block_info *)(ei + 1);
656                 *ptr = (unsigned long)(bi + 1);
657         } else {
658                 *ptr = (unsigned long)(ei + 1);
659         }
660         *end = (unsigned long)ei + item_size;
661         return 0;
662 }
663
664 /*
665  * build backref tree for a given tree block. root of the backref tree
666  * corresponds the tree block, leaves of the backref tree correspond
667  * roots of b-trees that reference the tree block.
668  *
669  * the basic idea of this function is check backrefs of a given block
670  * to find upper level blocks that refernece the block, and then check
671  * bakcrefs of these upper level blocks recursively. the recursion stop
672  * when tree root is reached or backrefs for the block is cached.
673  *
674  * NOTE: if we find backrefs for a block are cached, we know backrefs
675  * for all upper level blocks that directly/indirectly reference the
676  * block are also cached.
677  */
678 static noinline_for_stack
679 struct backref_node *build_backref_tree(struct reloc_control *rc,
680                                         struct btrfs_key *node_key,
681                                         int level, u64 bytenr)
682 {
683         struct backref_cache *cache = &rc->backref_cache;
684         struct btrfs_path *path1;
685         struct btrfs_path *path2;
686         struct extent_buffer *eb;
687         struct btrfs_root *root;
688         struct backref_node *cur;
689         struct backref_node *upper;
690         struct backref_node *lower;
691         struct backref_node *node = NULL;
692         struct backref_node *exist = NULL;
693         struct backref_edge *edge;
694         struct rb_node *rb_node;
695         struct btrfs_key key;
696         unsigned long end;
697         unsigned long ptr;
698         LIST_HEAD(list);
699         LIST_HEAD(useless);
700         int cowonly;
701         int ret;
702         int err = 0;
703         bool need_check = true;
704
705         path1 = btrfs_alloc_path();
706         path2 = btrfs_alloc_path();
707         if (!path1 || !path2) {
708                 err = -ENOMEM;
709                 goto out;
710         }
711         path1->reada = 1;
712         path2->reada = 2;
713
714         node = alloc_backref_node(cache);
715         if (!node) {
716                 err = -ENOMEM;
717                 goto out;
718         }
719
720         node->bytenr = bytenr;
721         node->level = level;
722         node->lowest = 1;
723         cur = node;
724 again:
725         end = 0;
726         ptr = 0;
727         key.objectid = cur->bytenr;
728         key.type = BTRFS_METADATA_ITEM_KEY;
729         key.offset = (u64)-1;
730
731         path1->search_commit_root = 1;
732         path1->skip_locking = 1;
733         ret = btrfs_search_slot(NULL, rc->extent_root, &key, path1,
734                                 0, 0);
735         if (ret < 0) {
736                 err = ret;
737                 goto out;
738         }
739         BUG_ON(!ret || !path1->slots[0]);
740
741         path1->slots[0]--;
742
743         WARN_ON(cur->checked);
744         if (!list_empty(&cur->upper)) {
745                 /*
746                  * the backref was added previously when processing
747                  * backref of type BTRFS_TREE_BLOCK_REF_KEY
748                  */
749                 BUG_ON(!list_is_singular(&cur->upper));
750                 edge = list_entry(cur->upper.next, struct backref_edge,
751                                   list[LOWER]);
752                 BUG_ON(!list_empty(&edge->list[UPPER]));
753                 exist = edge->node[UPPER];
754                 /*
755                  * add the upper level block to pending list if we need
756                  * check its backrefs
757                  */
758                 if (!exist->checked)
759                         list_add_tail(&edge->list[UPPER], &list);
760         } else {
761                 exist = NULL;
762         }
763
764         while (1) {
765                 cond_resched();
766                 eb = path1->nodes[0];
767
768                 if (ptr >= end) {
769                         if (path1->slots[0] >= btrfs_header_nritems(eb)) {
770                                 ret = btrfs_next_leaf(rc->extent_root, path1);
771                                 if (ret < 0) {
772                                         err = ret;
773                                         goto out;
774                                 }
775                                 if (ret > 0)
776                                         break;
777                                 eb = path1->nodes[0];
778                         }
779
780                         btrfs_item_key_to_cpu(eb, &key, path1->slots[0]);
781                         if (key.objectid != cur->bytenr) {
782                                 WARN_ON(exist);
783                                 break;
784                         }
785
786                         if (key.type == BTRFS_EXTENT_ITEM_KEY ||
787                             key.type == BTRFS_METADATA_ITEM_KEY) {
788                                 ret = find_inline_backref(eb, path1->slots[0],
789                                                           &ptr, &end);
790                                 if (ret)
791                                         goto next;
792                         }
793                 }
794
795                 if (ptr < end) {
796                         /* update key for inline back ref */
797                         struct btrfs_extent_inline_ref *iref;
798                         iref = (struct btrfs_extent_inline_ref *)ptr;
799                         key.type = btrfs_extent_inline_ref_type(eb, iref);
800                         key.offset = btrfs_extent_inline_ref_offset(eb, iref);
801                         WARN_ON(key.type != BTRFS_TREE_BLOCK_REF_KEY &&
802                                 key.type != BTRFS_SHARED_BLOCK_REF_KEY);
803                 }
804
805                 if (exist &&
806                     ((key.type == BTRFS_TREE_BLOCK_REF_KEY &&
807                       exist->owner == key.offset) ||
808                      (key.type == BTRFS_SHARED_BLOCK_REF_KEY &&
809                       exist->bytenr == key.offset))) {
810                         exist = NULL;
811                         goto next;
812                 }
813
814 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
815                 if (key.type == BTRFS_SHARED_BLOCK_REF_KEY ||
816                     key.type == BTRFS_EXTENT_REF_V0_KEY) {
817                         if (key.type == BTRFS_EXTENT_REF_V0_KEY) {
818                                 struct btrfs_extent_ref_v0 *ref0;
819                                 ref0 = btrfs_item_ptr(eb, path1->slots[0],
820                                                 struct btrfs_extent_ref_v0);
821                                 if (key.objectid == key.offset) {
822                                         root = find_tree_root(rc, eb, ref0);
823                                         if (root && !should_ignore_root(root))
824                                                 cur->root = root;
825                                         else
826                                                 list_add(&cur->list, &useless);
827                                         break;
828                                 }
829                                 if (is_cowonly_root(btrfs_ref_root_v0(eb,
830                                                                       ref0)))
831                                         cur->cowonly = 1;
832                         }
833 #else
834                 BUG_ON(key.type == BTRFS_EXTENT_REF_V0_KEY);
835                 if (key.type == BTRFS_SHARED_BLOCK_REF_KEY) {
836 #endif
837                         if (key.objectid == key.offset) {
838                                 /*
839                                  * only root blocks of reloc trees use
840                                  * backref of this type.
841                                  */
842                                 root = find_reloc_root(rc, cur->bytenr);
843                                 BUG_ON(!root);
844                                 cur->root = root;
845                                 break;
846                         }
847
848                         edge = alloc_backref_edge(cache);
849                         if (!edge) {
850                                 err = -ENOMEM;
851                                 goto out;
852                         }
853                         rb_node = tree_search(&cache->rb_root, key.offset);
854                         if (!rb_node) {
855                                 upper = alloc_backref_node(cache);
856                                 if (!upper) {
857                                         free_backref_edge(cache, edge);
858                                         err = -ENOMEM;
859                                         goto out;
860                                 }
861                                 upper->bytenr = key.offset;
862                                 upper->level = cur->level + 1;
863                                 /*
864                                  *  backrefs for the upper level block isn't
865                                  *  cached, add the block to pending list
866                                  */
867                                 list_add_tail(&edge->list[UPPER], &list);
868                         } else {
869                                 upper = rb_entry(rb_node, struct backref_node,
870                                                  rb_node);
871                                 BUG_ON(!upper->checked);
872                                 INIT_LIST_HEAD(&edge->list[UPPER]);
873                         }
874                         list_add_tail(&edge->list[LOWER], &cur->upper);
875                         edge->node[LOWER] = cur;
876                         edge->node[UPPER] = upper;
877
878                         goto next;
879                 } else if (key.type != BTRFS_TREE_BLOCK_REF_KEY) {
880                         goto next;
881                 }
882
883                 /* key.type == BTRFS_TREE_BLOCK_REF_KEY */
884                 root = read_fs_root(rc->extent_root->fs_info, key.offset);
885                 if (IS_ERR(root)) {
886                         err = PTR_ERR(root);
887                         goto out;
888                 }
889
890                 if (!root->ref_cows)
891                         cur->cowonly = 1;
892
893                 if (btrfs_root_level(&root->root_item) == cur->level) {
894                         /* tree root */
895                         BUG_ON(btrfs_root_bytenr(&root->root_item) !=
896                                cur->bytenr);
897                         if (should_ignore_root(root))
898                                 list_add(&cur->list, &useless);
899                         else
900                                 cur->root = root;
901                         break;
902                 }
903
904                 level = cur->level + 1;
905
906                 /*
907                  * searching the tree to find upper level blocks
908                  * reference the block.
909                  */
910                 path2->search_commit_root = 1;
911                 path2->skip_locking = 1;
912                 path2->lowest_level = level;
913                 ret = btrfs_search_slot(NULL, root, node_key, path2, 0, 0);
914                 path2->lowest_level = 0;
915                 if (ret < 0) {
916                         err = ret;
917                         goto out;
918                 }
919                 if (ret > 0 && path2->slots[level] > 0)
920                         path2->slots[level]--;
921
922                 eb = path2->nodes[level];
923                 WARN_ON(btrfs_node_blockptr(eb, path2->slots[level]) !=
924                         cur->bytenr);
925
926                 lower = cur;
927                 need_check = true;
928                 for (; level < BTRFS_MAX_LEVEL; level++) {
929                         if (!path2->nodes[level]) {
930                                 BUG_ON(btrfs_root_bytenr(&root->root_item) !=
931                                        lower->bytenr);
932                                 if (should_ignore_root(root))
933                                         list_add(&lower->list, &useless);
934                                 else
935                                         lower->root = root;
936                                 break;
937                         }
938
939                         edge = alloc_backref_edge(cache);
940                         if (!edge) {
941                                 err = -ENOMEM;
942                                 goto out;
943                         }
944
945                         eb = path2->nodes[level];
946                         rb_node = tree_search(&cache->rb_root, eb->start);
947                         if (!rb_node) {
948                                 upper = alloc_backref_node(cache);
949                                 if (!upper) {
950                                         free_backref_edge(cache, edge);
951                                         err = -ENOMEM;
952                                         goto out;
953                                 }
954                                 upper->bytenr = eb->start;
955                                 upper->owner = btrfs_header_owner(eb);
956                                 upper->level = lower->level + 1;
957                                 if (!root->ref_cows)
958                                         upper->cowonly = 1;
959
960                                 /*
961                                  * if we know the block isn't shared
962                                  * we can void checking its backrefs.
963                                  */
964                                 if (btrfs_block_can_be_shared(root, eb))
965                                         upper->checked = 0;
966                                 else
967                                         upper->checked = 1;
968
969                                 /*
970                                  * add the block to pending list if we
971                                  * need check its backrefs, we only do this once
972                                  * while walking up a tree as we will catch
973                                  * anything else later on.
974                                  */
975                                 if (!upper->checked && need_check) {
976                                         need_check = false;
977                                         list_add_tail(&edge->list[UPPER],
978                                                       &list);
979                                 } else
980                                         INIT_LIST_HEAD(&edge->list[UPPER]);
981                         } else {
982                                 upper = rb_entry(rb_node, struct backref_node,
983                                                  rb_node);
984                                 BUG_ON(!upper->checked);
985                                 INIT_LIST_HEAD(&edge->list[UPPER]);
986                                 if (!upper->owner)
987                                         upper->owner = btrfs_header_owner(eb);
988                         }
989                         list_add_tail(&edge->list[LOWER], &lower->upper);
990                         edge->node[LOWER] = lower;
991                         edge->node[UPPER] = upper;
992
993                         if (rb_node)
994                                 break;
995                         lower = upper;
996                         upper = NULL;
997                 }
998                 btrfs_release_path(path2);
999 next:
1000                 if (ptr < end) {
1001                         ptr += btrfs_extent_inline_ref_size(key.type);
1002                         if (ptr >= end) {
1003                                 WARN_ON(ptr > end);
1004                                 ptr = 0;
1005                                 end = 0;
1006                         }
1007                 }
1008                 if (ptr >= end)
1009                         path1->slots[0]++;
1010         }
1011         btrfs_release_path(path1);
1012
1013         cur->checked = 1;
1014         WARN_ON(exist);
1015
1016         /* the pending list isn't empty, take the first block to process */
1017         if (!list_empty(&list)) {
1018                 edge = list_entry(list.next, struct backref_edge, list[UPPER]);
1019                 list_del_init(&edge->list[UPPER]);
1020                 cur = edge->node[UPPER];
1021                 goto again;
1022         }
1023
1024         /*
1025          * everything goes well, connect backref nodes and insert backref nodes
1026          * into the cache.
1027          */
1028         BUG_ON(!node->checked);
1029         cowonly = node->cowonly;
1030         if (!cowonly) {
1031                 rb_node = tree_insert(&cache->rb_root, node->bytenr,
1032                                       &node->rb_node);
1033                 if (rb_node)
1034                         backref_tree_panic(rb_node, -EEXIST, node->bytenr);
1035                 list_add_tail(&node->lower, &cache->leaves);
1036         }
1037
1038         list_for_each_entry(edge, &node->upper, list[LOWER])
1039                 list_add_tail(&edge->list[UPPER], &list);
1040
1041         while (!list_empty(&list)) {
1042                 edge = list_entry(list.next, struct backref_edge, list[UPPER]);
1043                 list_del_init(&edge->list[UPPER]);
1044                 upper = edge->node[UPPER];
1045                 if (upper->detached) {
1046                         list_del(&edge->list[LOWER]);
1047                         lower = edge->node[LOWER];
1048                         free_backref_edge(cache, edge);
1049                         if (list_empty(&lower->upper))
1050                                 list_add(&lower->list, &useless);
1051                         continue;
1052                 }
1053
1054                 if (!RB_EMPTY_NODE(&upper->rb_node)) {
1055                         if (upper->lowest) {
1056                                 list_del_init(&upper->lower);
1057                                 upper->lowest = 0;
1058                         }
1059
1060                         list_add_tail(&edge->list[UPPER], &upper->lower);
1061                         continue;
1062                 }
1063
1064                 BUG_ON(!upper->checked);
1065                 BUG_ON(cowonly != upper->cowonly);
1066                 if (!cowonly) {
1067                         rb_node = tree_insert(&cache->rb_root, upper->bytenr,
1068                                               &upper->rb_node);
1069                         if (rb_node)
1070                                 backref_tree_panic(rb_node, -EEXIST,
1071                                                    upper->bytenr);
1072                 }
1073
1074                 list_add_tail(&edge->list[UPPER], &upper->lower);
1075
1076                 list_for_each_entry(edge, &upper->upper, list[LOWER])
1077                         list_add_tail(&edge->list[UPPER], &list);
1078         }
1079         /*
1080          * process useless backref nodes. backref nodes for tree leaves
1081          * are deleted from the cache. backref nodes for upper level
1082          * tree blocks are left in the cache to avoid unnecessary backref
1083          * lookup.
1084          */
1085         while (!list_empty(&useless)) {
1086                 upper = list_entry(useless.next, struct backref_node, list);
1087                 list_del_init(&upper->list);
1088                 BUG_ON(!list_empty(&upper->upper));
1089                 if (upper == node)
1090                         node = NULL;
1091                 if (upper->lowest) {
1092                         list_del_init(&upper->lower);
1093                         upper->lowest = 0;
1094                 }
1095                 while (!list_empty(&upper->lower)) {
1096                         edge = list_entry(upper->lower.next,
1097                                           struct backref_edge, list[UPPER]);
1098                         list_del(&edge->list[UPPER]);
1099                         list_del(&edge->list[LOWER]);
1100                         lower = edge->node[LOWER];
1101                         free_backref_edge(cache, edge);
1102
1103                         if (list_empty(&lower->upper))
1104                                 list_add(&lower->list, &useless);
1105                 }
1106                 __mark_block_processed(rc, upper);
1107                 if (upper->level > 0) {
1108                         list_add(&upper->list, &cache->detached);
1109                         upper->detached = 1;
1110                 } else {
1111                         rb_erase(&upper->rb_node, &cache->rb_root);
1112                         free_backref_node(cache, upper);
1113                 }
1114         }
1115 out:
1116         btrfs_free_path(path1);
1117         btrfs_free_path(path2);
1118         if (err) {
1119                 while (!list_empty(&useless)) {
1120                         lower = list_entry(useless.next,
1121                                            struct backref_node, upper);
1122                         list_del_init(&lower->upper);
1123                 }
1124                 upper = node;
1125                 INIT_LIST_HEAD(&list);
1126                 while (upper) {
1127                         if (RB_EMPTY_NODE(&upper->rb_node)) {
1128                                 list_splice_tail(&upper->upper, &list);
1129                                 free_backref_node(cache, upper);
1130                         }
1131
1132                         if (list_empty(&list))
1133                                 break;
1134
1135                         edge = list_entry(list.next, struct backref_edge,
1136                                           list[LOWER]);
1137                         list_del(&edge->list[LOWER]);
1138                         upper = edge->node[UPPER];
1139                         free_backref_edge(cache, edge);
1140                 }
1141                 return ERR_PTR(err);
1142         }
1143         BUG_ON(node && node->detached);
1144         return node;
1145 }
1146
1147 /*
1148  * helper to add backref node for the newly created snapshot.
1149  * the backref node is created by cloning backref node that
1150  * corresponds to root of source tree
1151  */
1152 static int clone_backref_node(struct btrfs_trans_handle *trans,
1153                               struct reloc_control *rc,
1154                               struct btrfs_root *src,
1155                               struct btrfs_root *dest)
1156 {
1157         struct btrfs_root *reloc_root = src->reloc_root;
1158         struct backref_cache *cache = &rc->backref_cache;
1159         struct backref_node *node = NULL;
1160         struct backref_node *new_node;
1161         struct backref_edge *edge;
1162         struct backref_edge *new_edge;
1163         struct rb_node *rb_node;
1164
1165         if (cache->last_trans > 0)
1166                 update_backref_cache(trans, cache);
1167
1168         rb_node = tree_search(&cache->rb_root, src->commit_root->start);
1169         if (rb_node) {
1170                 node = rb_entry(rb_node, struct backref_node, rb_node);
1171                 if (node->detached)
1172                         node = NULL;
1173                 else
1174                         BUG_ON(node->new_bytenr != reloc_root->node->start);
1175         }
1176
1177         if (!node) {
1178                 rb_node = tree_search(&cache->rb_root,
1179                                       reloc_root->commit_root->start);
1180                 if (rb_node) {
1181                         node = rb_entry(rb_node, struct backref_node,
1182                                         rb_node);
1183                         BUG_ON(node->detached);
1184                 }
1185         }
1186
1187         if (!node)
1188                 return 0;
1189
1190         new_node = alloc_backref_node(cache);
1191         if (!new_node)
1192                 return -ENOMEM;
1193
1194         new_node->bytenr = dest->node->start;
1195         new_node->level = node->level;
1196         new_node->lowest = node->lowest;
1197         new_node->checked = 1;
1198         new_node->root = dest;
1199
1200         if (!node->lowest) {
1201                 list_for_each_entry(edge, &node->lower, list[UPPER]) {
1202                         new_edge = alloc_backref_edge(cache);
1203                         if (!new_edge)
1204                                 goto fail;
1205
1206                         new_edge->node[UPPER] = new_node;
1207                         new_edge->node[LOWER] = edge->node[LOWER];
1208                         list_add_tail(&new_edge->list[UPPER],
1209                                       &new_node->lower);
1210                 }
1211         } else {
1212                 list_add_tail(&new_node->lower, &cache->leaves);
1213         }
1214
1215         rb_node = tree_insert(&cache->rb_root, new_node->bytenr,
1216                               &new_node->rb_node);
1217         if (rb_node)
1218                 backref_tree_panic(rb_node, -EEXIST, new_node->bytenr);
1219
1220         if (!new_node->lowest) {
1221                 list_for_each_entry(new_edge, &new_node->lower, list[UPPER]) {
1222                         list_add_tail(&new_edge->list[LOWER],
1223                                       &new_edge->node[LOWER]->upper);
1224                 }
1225         }
1226         return 0;
1227 fail:
1228         while (!list_empty(&new_node->lower)) {
1229                 new_edge = list_entry(new_node->lower.next,
1230                                       struct backref_edge, list[UPPER]);
1231                 list_del(&new_edge->list[UPPER]);
1232                 free_backref_edge(cache, new_edge);
1233         }
1234         free_backref_node(cache, new_node);
1235         return -ENOMEM;
1236 }
1237
1238 /*
1239  * helper to add 'address of tree root -> reloc tree' mapping
1240  */
1241 static int __must_check __add_reloc_root(struct btrfs_root *root)
1242 {
1243         struct rb_node *rb_node;
1244         struct mapping_node *node;
1245         struct reloc_control *rc = root->fs_info->reloc_ctl;
1246
1247         node = kmalloc(sizeof(*node), GFP_NOFS);
1248         if (!node)
1249                 return -ENOMEM;
1250
1251         node->bytenr = root->node->start;
1252         node->data = root;
1253
1254         spin_lock(&rc->reloc_root_tree.lock);
1255         rb_node = tree_insert(&rc->reloc_root_tree.rb_root,
1256                               node->bytenr, &node->rb_node);
1257         spin_unlock(&rc->reloc_root_tree.lock);
1258         if (rb_node) {
1259                 btrfs_panic(root->fs_info, -EEXIST, "Duplicate root found "
1260                             "for start=%llu while inserting into relocation "
1261                             "tree\n", node->bytenr);
1262                 kfree(node);
1263                 return -EEXIST;
1264         }
1265
1266         list_add_tail(&root->root_list, &rc->reloc_roots);
1267         return 0;
1268 }
1269
1270 /*
1271  * helper to delete the 'address of tree root -> reloc tree'
1272  * mapping
1273  */
1274 static void __del_reloc_root(struct btrfs_root *root)
1275 {
1276         struct rb_node *rb_node;
1277         struct mapping_node *node = NULL;
1278         struct reloc_control *rc = root->fs_info->reloc_ctl;
1279
1280         spin_lock(&rc->reloc_root_tree.lock);
1281         rb_node = tree_search(&rc->reloc_root_tree.rb_root,
1282                               root->node->start);
1283         if (rb_node) {
1284                 node = rb_entry(rb_node, struct mapping_node, rb_node);
1285                 rb_erase(&node->rb_node, &rc->reloc_root_tree.rb_root);
1286         }
1287         spin_unlock(&rc->reloc_root_tree.lock);
1288
1289         if (!node)
1290                 return;
1291         BUG_ON((struct btrfs_root *)node->data != root);
1292
1293         spin_lock(&root->fs_info->trans_lock);
1294         list_del_init(&root->root_list);
1295         spin_unlock(&root->fs_info->trans_lock);
1296         kfree(node);
1297 }
1298
1299 /*
1300  * helper to update the 'address of tree root -> reloc tree'
1301  * mapping
1302  */
1303 static int __update_reloc_root(struct btrfs_root *root, u64 new_bytenr)
1304 {
1305         struct rb_node *rb_node;
1306         struct mapping_node *node = NULL;
1307         struct reloc_control *rc = root->fs_info->reloc_ctl;
1308
1309         spin_lock(&rc->reloc_root_tree.lock);
1310         rb_node = tree_search(&rc->reloc_root_tree.rb_root,
1311                               root->node->start);
1312         if (rb_node) {
1313                 node = rb_entry(rb_node, struct mapping_node, rb_node);
1314                 rb_erase(&node->rb_node, &rc->reloc_root_tree.rb_root);
1315         }
1316         spin_unlock(&rc->reloc_root_tree.lock);
1317
1318         if (!node)
1319                 return 0;
1320         BUG_ON((struct btrfs_root *)node->data != root);
1321
1322         spin_lock(&rc->reloc_root_tree.lock);
1323         node->bytenr = new_bytenr;
1324         rb_node = tree_insert(&rc->reloc_root_tree.rb_root,
1325                               node->bytenr, &node->rb_node);
1326         spin_unlock(&rc->reloc_root_tree.lock);
1327         if (rb_node)
1328                 backref_tree_panic(rb_node, -EEXIST, node->bytenr);
1329         return 0;
1330 }
1331
1332 static struct btrfs_root *create_reloc_root(struct btrfs_trans_handle *trans,
1333                                         struct btrfs_root *root, u64 objectid)
1334 {
1335         struct btrfs_root *reloc_root;
1336         struct extent_buffer *eb;
1337         struct btrfs_root_item *root_item;
1338         struct btrfs_key root_key;
1339         u64 last_snap = 0;
1340         int ret;
1341
1342         root_item = kmalloc(sizeof(*root_item), GFP_NOFS);
1343         BUG_ON(!root_item);
1344
1345         root_key.objectid = BTRFS_TREE_RELOC_OBJECTID;
1346         root_key.type = BTRFS_ROOT_ITEM_KEY;
1347         root_key.offset = objectid;
1348
1349         if (root->root_key.objectid == objectid) {
1350                 /* called by btrfs_init_reloc_root */
1351                 ret = btrfs_copy_root(trans, root, root->commit_root, &eb,
1352                                       BTRFS_TREE_RELOC_OBJECTID);
1353                 BUG_ON(ret);
1354
1355                 last_snap = btrfs_root_last_snapshot(&root->root_item);
1356                 btrfs_set_root_last_snapshot(&root->root_item,
1357                                              trans->transid - 1);
1358         } else {
1359                 /*
1360                  * called by btrfs_reloc_post_snapshot_hook.
1361                  * the source tree is a reloc tree, all tree blocks
1362                  * modified after it was created have RELOC flag
1363                  * set in their headers. so it's OK to not update
1364                  * the 'last_snapshot'.
1365                  */
1366                 ret = btrfs_copy_root(trans, root, root->node, &eb,
1367                                       BTRFS_TREE_RELOC_OBJECTID);
1368                 BUG_ON(ret);
1369         }
1370
1371         memcpy(root_item, &root->root_item, sizeof(*root_item));
1372         btrfs_set_root_bytenr(root_item, eb->start);
1373         btrfs_set_root_level(root_item, btrfs_header_level(eb));
1374         btrfs_set_root_generation(root_item, trans->transid);
1375
1376         if (root->root_key.objectid == objectid) {
1377                 btrfs_set_root_refs(root_item, 0);
1378                 memset(&root_item->drop_progress, 0,
1379                        sizeof(struct btrfs_disk_key));
1380                 root_item->drop_level = 0;
1381                 /*
1382                  * abuse rtransid, it is safe because it is impossible to
1383                  * receive data into a relocation tree.
1384                  */
1385                 btrfs_set_root_rtransid(root_item, last_snap);
1386                 btrfs_set_root_otransid(root_item, trans->transid);
1387         }
1388
1389         btrfs_tree_unlock(eb);
1390         free_extent_buffer(eb);
1391
1392         ret = btrfs_insert_root(trans, root->fs_info->tree_root,
1393                                 &root_key, root_item);
1394         BUG_ON(ret);
1395         kfree(root_item);
1396
1397         reloc_root = btrfs_read_fs_root(root->fs_info->tree_root, &root_key);
1398         BUG_ON(IS_ERR(reloc_root));
1399         reloc_root->last_trans = trans->transid;
1400         return reloc_root;
1401 }
1402
1403 /*
1404  * create reloc tree for a given fs tree. reloc tree is just a
1405  * snapshot of the fs tree with special root objectid.
1406  */
1407 int btrfs_init_reloc_root(struct btrfs_trans_handle *trans,
1408                           struct btrfs_root *root)
1409 {
1410         struct btrfs_root *reloc_root;
1411         struct reloc_control *rc = root->fs_info->reloc_ctl;
1412         struct btrfs_block_rsv *rsv;
1413         int clear_rsv = 0;
1414         int ret;
1415
1416         if (root->reloc_root) {
1417                 reloc_root = root->reloc_root;
1418                 reloc_root->last_trans = trans->transid;
1419                 return 0;
1420         }
1421
1422         if (!rc || !rc->create_reloc_tree ||
1423             root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID)
1424                 return 0;
1425
1426         if (!trans->reloc_reserved) {
1427                 rsv = trans->block_rsv;
1428                 trans->block_rsv = rc->block_rsv;
1429                 clear_rsv = 1;
1430         }
1431         reloc_root = create_reloc_root(trans, root, root->root_key.objectid);
1432         if (clear_rsv)
1433                 trans->block_rsv = rsv;
1434
1435         ret = __add_reloc_root(reloc_root);
1436         BUG_ON(ret < 0);
1437         root->reloc_root = reloc_root;
1438         return 0;
1439 }
1440
1441 /*
1442  * update root item of reloc tree
1443  */
1444 int btrfs_update_reloc_root(struct btrfs_trans_handle *trans,
1445                             struct btrfs_root *root)
1446 {
1447         struct btrfs_root *reloc_root;
1448         struct btrfs_root_item *root_item;
1449         int ret;
1450
1451         if (!root->reloc_root)
1452                 goto out;
1453
1454         reloc_root = root->reloc_root;
1455         root_item = &reloc_root->root_item;
1456
1457         if (root->fs_info->reloc_ctl->merge_reloc_tree &&
1458             btrfs_root_refs(root_item) == 0) {
1459                 root->reloc_root = NULL;
1460                 __del_reloc_root(reloc_root);
1461         }
1462
1463         if (reloc_root->commit_root != reloc_root->node) {
1464                 btrfs_set_root_node(root_item, reloc_root->node);
1465                 free_extent_buffer(reloc_root->commit_root);
1466                 reloc_root->commit_root = btrfs_root_node(reloc_root);
1467         }
1468
1469         ret = btrfs_update_root(trans, root->fs_info->tree_root,
1470                                 &reloc_root->root_key, root_item);
1471         BUG_ON(ret);
1472
1473 out:
1474         return 0;
1475 }
1476
1477 /*
1478  * helper to find first cached inode with inode number >= objectid
1479  * in a subvolume
1480  */
1481 static struct inode *find_next_inode(struct btrfs_root *root, u64 objectid)
1482 {
1483         struct rb_node *node;
1484         struct rb_node *prev;
1485         struct btrfs_inode *entry;
1486         struct inode *inode;
1487
1488         spin_lock(&root->inode_lock);
1489 again:
1490         node = root->inode_tree.rb_node;
1491         prev = NULL;
1492         while (node) {
1493                 prev = node;
1494                 entry = rb_entry(node, struct btrfs_inode, rb_node);
1495
1496                 if (objectid < btrfs_ino(&entry->vfs_inode))
1497                         node = node->rb_left;
1498                 else if (objectid > btrfs_ino(&entry->vfs_inode))
1499                         node = node->rb_right;
1500                 else
1501                         break;
1502         }
1503         if (!node) {
1504                 while (prev) {
1505                         entry = rb_entry(prev, struct btrfs_inode, rb_node);
1506                         if (objectid <= btrfs_ino(&entry->vfs_inode)) {
1507                                 node = prev;
1508                                 break;
1509                         }
1510                         prev = rb_next(prev);
1511                 }
1512         }
1513         while (node) {
1514                 entry = rb_entry(node, struct btrfs_inode, rb_node);
1515                 inode = igrab(&entry->vfs_inode);
1516                 if (inode) {
1517                         spin_unlock(&root->inode_lock);
1518                         return inode;
1519                 }
1520
1521                 objectid = btrfs_ino(&entry->vfs_inode) + 1;
1522                 if (cond_resched_lock(&root->inode_lock))
1523                         goto again;
1524
1525                 node = rb_next(node);
1526         }
1527         spin_unlock(&root->inode_lock);
1528         return NULL;
1529 }
1530
1531 static int in_block_group(u64 bytenr,
1532                           struct btrfs_block_group_cache *block_group)
1533 {
1534         if (bytenr >= block_group->key.objectid &&
1535             bytenr < block_group->key.objectid + block_group->key.offset)
1536                 return 1;
1537         return 0;
1538 }
1539
1540 /*
1541  * get new location of data
1542  */
1543 static int get_new_location(struct inode *reloc_inode, u64 *new_bytenr,
1544                             u64 bytenr, u64 num_bytes)
1545 {
1546         struct btrfs_root *root = BTRFS_I(reloc_inode)->root;
1547         struct btrfs_path *path;
1548         struct btrfs_file_extent_item *fi;
1549         struct extent_buffer *leaf;
1550         int ret;
1551
1552         path = btrfs_alloc_path();
1553         if (!path)
1554                 return -ENOMEM;
1555
1556         bytenr -= BTRFS_I(reloc_inode)->index_cnt;
1557         ret = btrfs_lookup_file_extent(NULL, root, path, btrfs_ino(reloc_inode),
1558                                        bytenr, 0);
1559         if (ret < 0)
1560                 goto out;
1561         if (ret > 0) {
1562                 ret = -ENOENT;
1563                 goto out;
1564         }
1565
1566         leaf = path->nodes[0];
1567         fi = btrfs_item_ptr(leaf, path->slots[0],
1568                             struct btrfs_file_extent_item);
1569
1570         BUG_ON(btrfs_file_extent_offset(leaf, fi) ||
1571                btrfs_file_extent_compression(leaf, fi) ||
1572                btrfs_file_extent_encryption(leaf, fi) ||
1573                btrfs_file_extent_other_encoding(leaf, fi));
1574
1575         if (num_bytes != btrfs_file_extent_disk_num_bytes(leaf, fi)) {
1576                 ret = -EINVAL;
1577                 goto out;
1578         }
1579
1580         *new_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
1581         ret = 0;
1582 out:
1583         btrfs_free_path(path);
1584         return ret;
1585 }
1586
1587 /*
1588  * update file extent items in the tree leaf to point to
1589  * the new locations.
1590  */
1591 static noinline_for_stack
1592 int replace_file_extents(struct btrfs_trans_handle *trans,
1593                          struct reloc_control *rc,
1594                          struct btrfs_root *root,
1595                          struct extent_buffer *leaf)
1596 {
1597         struct btrfs_key key;
1598         struct btrfs_file_extent_item *fi;
1599         struct inode *inode = NULL;
1600         u64 parent;
1601         u64 bytenr;
1602         u64 new_bytenr = 0;
1603         u64 num_bytes;
1604         u64 end;
1605         u32 nritems;
1606         u32 i;
1607         int ret = 0;
1608         int first = 1;
1609         int dirty = 0;
1610
1611         if (rc->stage != UPDATE_DATA_PTRS)
1612                 return 0;
1613
1614         /* reloc trees always use full backref */
1615         if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID)
1616                 parent = leaf->start;
1617         else
1618                 parent = 0;
1619
1620         nritems = btrfs_header_nritems(leaf);
1621         for (i = 0; i < nritems; i++) {
1622                 cond_resched();
1623                 btrfs_item_key_to_cpu(leaf, &key, i);
1624                 if (key.type != BTRFS_EXTENT_DATA_KEY)
1625                         continue;
1626                 fi = btrfs_item_ptr(leaf, i, struct btrfs_file_extent_item);
1627                 if (btrfs_file_extent_type(leaf, fi) ==
1628                     BTRFS_FILE_EXTENT_INLINE)
1629                         continue;
1630                 bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
1631                 num_bytes = btrfs_file_extent_disk_num_bytes(leaf, fi);
1632                 if (bytenr == 0)
1633                         continue;
1634                 if (!in_block_group(bytenr, rc->block_group))
1635                         continue;
1636
1637                 /*
1638                  * if we are modifying block in fs tree, wait for readpage
1639                  * to complete and drop the extent cache
1640                  */
1641                 if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) {
1642                         if (first) {
1643                                 inode = find_next_inode(root, key.objectid);
1644                                 first = 0;
1645                         } else if (inode && btrfs_ino(inode) < key.objectid) {
1646                                 btrfs_add_delayed_iput(inode);
1647                                 inode = find_next_inode(root, key.objectid);
1648                         }
1649                         if (inode && btrfs_ino(inode) == key.objectid) {
1650                                 end = key.offset +
1651                                       btrfs_file_extent_num_bytes(leaf, fi);
1652                                 WARN_ON(!IS_ALIGNED(key.offset,
1653                                                     root->sectorsize));
1654                                 WARN_ON(!IS_ALIGNED(end, root->sectorsize));
1655                                 end--;
1656                                 ret = try_lock_extent(&BTRFS_I(inode)->io_tree,
1657                                                       key.offset, end);
1658                                 if (!ret)
1659                                         continue;
1660
1661                                 btrfs_drop_extent_cache(inode, key.offset, end,
1662                                                         1);
1663                                 unlock_extent(&BTRFS_I(inode)->io_tree,
1664                                               key.offset, end);
1665                         }
1666                 }
1667
1668                 ret = get_new_location(rc->data_inode, &new_bytenr,
1669                                        bytenr, num_bytes);
1670                 if (ret) {
1671                         /*
1672                          * Don't have to abort since we've not changed anything
1673                          * in the file extent yet.
1674                          */
1675                         break;
1676                 }
1677
1678                 btrfs_set_file_extent_disk_bytenr(leaf, fi, new_bytenr);
1679                 dirty = 1;
1680
1681                 key.offset -= btrfs_file_extent_offset(leaf, fi);
1682                 ret = btrfs_inc_extent_ref(trans, root, new_bytenr,
1683                                            num_bytes, parent,
1684                                            btrfs_header_owner(leaf),
1685                                            key.objectid, key.offset, 1);
1686                 if (ret) {
1687                         btrfs_abort_transaction(trans, root, ret);
1688                         break;
1689                 }
1690
1691                 ret = btrfs_free_extent(trans, root, bytenr, num_bytes,
1692                                         parent, btrfs_header_owner(leaf),
1693                                         key.objectid, key.offset, 1);
1694                 if (ret) {
1695                         btrfs_abort_transaction(trans, root, ret);
1696                         break;
1697                 }
1698         }
1699         if (dirty)
1700                 btrfs_mark_buffer_dirty(leaf);
1701         if (inode)
1702                 btrfs_add_delayed_iput(inode);
1703         return ret;
1704 }
1705
1706 static noinline_for_stack
1707 int memcmp_node_keys(struct extent_buffer *eb, int slot,
1708                      struct btrfs_path *path, int level)
1709 {
1710         struct btrfs_disk_key key1;
1711         struct btrfs_disk_key key2;
1712         btrfs_node_key(eb, &key1, slot);
1713         btrfs_node_key(path->nodes[level], &key2, path->slots[level]);
1714         return memcmp(&key1, &key2, sizeof(key1));
1715 }
1716
1717 /*
1718  * try to replace tree blocks in fs tree with the new blocks
1719  * in reloc tree. tree blocks haven't been modified since the
1720  * reloc tree was create can be replaced.
1721  *
1722  * if a block was replaced, level of the block + 1 is returned.
1723  * if no block got replaced, 0 is returned. if there are other
1724  * errors, a negative error number is returned.
1725  */
1726 static noinline_for_stack
1727 int replace_path(struct btrfs_trans_handle *trans,
1728                  struct btrfs_root *dest, struct btrfs_root *src,
1729                  struct btrfs_path *path, struct btrfs_key *next_key,
1730                  int lowest_level, int max_level)
1731 {
1732         struct extent_buffer *eb;
1733         struct extent_buffer *parent;
1734         struct btrfs_key key;
1735         u64 old_bytenr;
1736         u64 new_bytenr;
1737         u64 old_ptr_gen;
1738         u64 new_ptr_gen;
1739         u64 last_snapshot;
1740         u32 blocksize;
1741         int cow = 0;
1742         int level;
1743         int ret;
1744         int slot;
1745
1746         BUG_ON(src->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID);
1747         BUG_ON(dest->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID);
1748
1749         last_snapshot = btrfs_root_last_snapshot(&src->root_item);
1750 again:
1751         slot = path->slots[lowest_level];
1752         btrfs_node_key_to_cpu(path->nodes[lowest_level], &key, slot);
1753
1754         eb = btrfs_lock_root_node(dest);
1755         btrfs_set_lock_blocking(eb);
1756         level = btrfs_header_level(eb);
1757
1758         if (level < lowest_level) {
1759                 btrfs_tree_unlock(eb);
1760                 free_extent_buffer(eb);
1761                 return 0;
1762         }
1763
1764         if (cow) {
1765                 ret = btrfs_cow_block(trans, dest, eb, NULL, 0, &eb);
1766                 BUG_ON(ret);
1767         }
1768         btrfs_set_lock_blocking(eb);
1769
1770         if (next_key) {
1771                 next_key->objectid = (u64)-1;
1772                 next_key->type = (u8)-1;
1773                 next_key->offset = (u64)-1;
1774         }
1775
1776         parent = eb;
1777         while (1) {
1778                 level = btrfs_header_level(parent);
1779                 BUG_ON(level < lowest_level);
1780
1781                 ret = btrfs_bin_search(parent, &key, level, &slot);
1782                 if (ret && slot > 0)
1783                         slot--;
1784
1785                 if (next_key && slot + 1 < btrfs_header_nritems(parent))
1786                         btrfs_node_key_to_cpu(parent, next_key, slot + 1);
1787
1788                 old_bytenr = btrfs_node_blockptr(parent, slot);
1789                 blocksize = btrfs_level_size(dest, level - 1);
1790                 old_ptr_gen = btrfs_node_ptr_generation(parent, slot);
1791
1792                 if (level <= max_level) {
1793                         eb = path->nodes[level];
1794                         new_bytenr = btrfs_node_blockptr(eb,
1795                                                         path->slots[level]);
1796                         new_ptr_gen = btrfs_node_ptr_generation(eb,
1797                                                         path->slots[level]);
1798                 } else {
1799                         new_bytenr = 0;
1800                         new_ptr_gen = 0;
1801                 }
1802
1803                 if (WARN_ON(new_bytenr > 0 && new_bytenr == old_bytenr)) {
1804                         ret = level;
1805                         break;
1806                 }
1807
1808                 if (new_bytenr == 0 || old_ptr_gen > last_snapshot ||
1809                     memcmp_node_keys(parent, slot, path, level)) {
1810                         if (level <= lowest_level) {
1811                                 ret = 0;
1812                                 break;
1813                         }
1814
1815                         eb = read_tree_block(dest, old_bytenr, blocksize,
1816                                              old_ptr_gen);
1817                         if (!eb || !extent_buffer_uptodate(eb)) {
1818                                 ret = (!eb) ? -ENOMEM : -EIO;
1819                                 free_extent_buffer(eb);
1820                                 break;
1821                         }
1822                         btrfs_tree_lock(eb);
1823                         if (cow) {
1824                                 ret = btrfs_cow_block(trans, dest, eb, parent,
1825                                                       slot, &eb);
1826                                 BUG_ON(ret);
1827                         }
1828                         btrfs_set_lock_blocking(eb);
1829
1830                         btrfs_tree_unlock(parent);
1831                         free_extent_buffer(parent);
1832
1833                         parent = eb;
1834                         continue;
1835                 }
1836
1837                 if (!cow) {
1838                         btrfs_tree_unlock(parent);
1839                         free_extent_buffer(parent);
1840                         cow = 1;
1841                         goto again;
1842                 }
1843
1844                 btrfs_node_key_to_cpu(path->nodes[level], &key,
1845                                       path->slots[level]);
1846                 btrfs_release_path(path);
1847
1848                 path->lowest_level = level;
1849                 ret = btrfs_search_slot(trans, src, &key, path, 0, 1);
1850                 path->lowest_level = 0;
1851                 BUG_ON(ret);
1852
1853                 /*
1854                  * swap blocks in fs tree and reloc tree.
1855                  */
1856                 btrfs_set_node_blockptr(parent, slot, new_bytenr);
1857                 btrfs_set_node_ptr_generation(parent, slot, new_ptr_gen);
1858                 btrfs_mark_buffer_dirty(parent);
1859
1860                 btrfs_set_node_blockptr(path->nodes[level],
1861                                         path->slots[level], old_bytenr);
1862                 btrfs_set_node_ptr_generation(path->nodes[level],
1863                                               path->slots[level], old_ptr_gen);
1864                 btrfs_mark_buffer_dirty(path->nodes[level]);
1865
1866                 ret = btrfs_inc_extent_ref(trans, src, old_bytenr, blocksize,
1867                                         path->nodes[level]->start,
1868                                         src->root_key.objectid, level - 1, 0,
1869                                         1);
1870                 BUG_ON(ret);
1871                 ret = btrfs_inc_extent_ref(trans, dest, new_bytenr, blocksize,
1872                                         0, dest->root_key.objectid, level - 1,
1873                                         0, 1);
1874                 BUG_ON(ret);
1875
1876                 ret = btrfs_free_extent(trans, src, new_bytenr, blocksize,
1877                                         path->nodes[level]->start,
1878                                         src->root_key.objectid, level - 1, 0,
1879                                         1);
1880                 BUG_ON(ret);
1881
1882                 ret = btrfs_free_extent(trans, dest, old_bytenr, blocksize,
1883                                         0, dest->root_key.objectid, level - 1,
1884                                         0, 1);
1885                 BUG_ON(ret);
1886
1887                 btrfs_unlock_up_safe(path, 0);
1888
1889                 ret = level;
1890                 break;
1891         }
1892         btrfs_tree_unlock(parent);
1893         free_extent_buffer(parent);
1894         return ret;
1895 }
1896
1897 /*
1898  * helper to find next relocated block in reloc tree
1899  */
1900 static noinline_for_stack
1901 int walk_up_reloc_tree(struct btrfs_root *root, struct btrfs_path *path,
1902                        int *level)
1903 {
1904         struct extent_buffer *eb;
1905         int i;
1906         u64 last_snapshot;
1907         u32 nritems;
1908
1909         last_snapshot = btrfs_root_last_snapshot(&root->root_item);
1910
1911         for (i = 0; i < *level; i++) {
1912                 free_extent_buffer(path->nodes[i]);
1913                 path->nodes[i] = NULL;
1914         }
1915
1916         for (i = *level; i < BTRFS_MAX_LEVEL && path->nodes[i]; i++) {
1917                 eb = path->nodes[i];
1918                 nritems = btrfs_header_nritems(eb);
1919                 while (path->slots[i] + 1 < nritems) {
1920                         path->slots[i]++;
1921                         if (btrfs_node_ptr_generation(eb, path->slots[i]) <=
1922                             last_snapshot)
1923                                 continue;
1924
1925                         *level = i;
1926                         return 0;
1927                 }
1928                 free_extent_buffer(path->nodes[i]);
1929                 path->nodes[i] = NULL;
1930         }
1931         return 1;
1932 }
1933
1934 /*
1935  * walk down reloc tree to find relocated block of lowest level
1936  */
1937 static noinline_for_stack
1938 int walk_down_reloc_tree(struct btrfs_root *root, struct btrfs_path *path,
1939                          int *level)
1940 {
1941         struct extent_buffer *eb = NULL;
1942         int i;
1943         u64 bytenr;
1944         u64 ptr_gen = 0;
1945         u64 last_snapshot;
1946         u32 blocksize;
1947         u32 nritems;
1948
1949         last_snapshot = btrfs_root_last_snapshot(&root->root_item);
1950
1951         for (i = *level; i > 0; i--) {
1952                 eb = path->nodes[i];
1953                 nritems = btrfs_header_nritems(eb);
1954                 while (path->slots[i] < nritems) {
1955                         ptr_gen = btrfs_node_ptr_generation(eb, path->slots[i]);
1956                         if (ptr_gen > last_snapshot)
1957                                 break;
1958                         path->slots[i]++;
1959                 }
1960                 if (path->slots[i] >= nritems) {
1961                         if (i == *level)
1962                                 break;
1963                         *level = i + 1;
1964                         return 0;
1965                 }
1966                 if (i == 1) {
1967                         *level = i;
1968                         return 0;
1969                 }
1970
1971                 bytenr = btrfs_node_blockptr(eb, path->slots[i]);
1972                 blocksize = btrfs_level_size(root, i - 1);
1973                 eb = read_tree_block(root, bytenr, blocksize, ptr_gen);
1974                 if (!eb || !extent_buffer_uptodate(eb)) {
1975                         free_extent_buffer(eb);
1976                         return -EIO;
1977                 }
1978                 BUG_ON(btrfs_header_level(eb) != i - 1);
1979                 path->nodes[i - 1] = eb;
1980                 path->slots[i - 1] = 0;
1981         }
1982         return 1;
1983 }
1984
1985 /*
1986  * invalidate extent cache for file extents whose key in range of
1987  * [min_key, max_key)
1988  */
1989 static int invalidate_extent_cache(struct btrfs_root *root,
1990                                    struct btrfs_key *min_key,
1991                                    struct btrfs_key *max_key)
1992 {
1993         struct inode *inode = NULL;
1994         u64 objectid;
1995         u64 start, end;
1996         u64 ino;
1997
1998         objectid = min_key->objectid;
1999         while (1) {
2000                 cond_resched();
2001                 iput(inode);
2002
2003                 if (objectid > max_key->objectid)
2004                         break;
2005
2006                 inode = find_next_inode(root, objectid);
2007                 if (!inode)
2008                         break;
2009                 ino = btrfs_ino(inode);
2010
2011                 if (ino > max_key->objectid) {
2012                         iput(inode);
2013                         break;
2014                 }
2015
2016                 objectid = ino + 1;
2017                 if (!S_ISREG(inode->i_mode))
2018                         continue;
2019
2020                 if (unlikely(min_key->objectid == ino)) {
2021                         if (min_key->type > BTRFS_EXTENT_DATA_KEY)
2022                                 continue;
2023                         if (min_key->type < BTRFS_EXTENT_DATA_KEY)
2024                                 start = 0;
2025                         else {
2026                                 start = min_key->offset;
2027                                 WARN_ON(!IS_ALIGNED(start, root->sectorsize));
2028                         }
2029                 } else {
2030                         start = 0;
2031                 }
2032
2033                 if (unlikely(max_key->objectid == ino)) {
2034                         if (max_key->type < BTRFS_EXTENT_DATA_KEY)
2035                                 continue;
2036                         if (max_key->type > BTRFS_EXTENT_DATA_KEY) {
2037                                 end = (u64)-1;
2038                         } else {
2039                                 if (max_key->offset == 0)
2040                                         continue;
2041                                 end = max_key->offset;
2042                                 WARN_ON(!IS_ALIGNED(end, root->sectorsize));
2043                                 end--;
2044                         }
2045                 } else {
2046                         end = (u64)-1;
2047                 }
2048
2049                 /* the lock_extent waits for readpage to complete */
2050                 lock_extent(&BTRFS_I(inode)->io_tree, start, end);
2051                 btrfs_drop_extent_cache(inode, start, end, 1);
2052                 unlock_extent(&BTRFS_I(inode)->io_tree, start, end);
2053         }
2054         return 0;
2055 }
2056
2057 static int find_next_key(struct btrfs_path *path, int level,
2058                          struct btrfs_key *key)
2059
2060 {
2061         while (level < BTRFS_MAX_LEVEL) {
2062                 if (!path->nodes[level])
2063                         break;
2064                 if (path->slots[level] + 1 <
2065                     btrfs_header_nritems(path->nodes[level])) {
2066                         btrfs_node_key_to_cpu(path->nodes[level], key,
2067                                               path->slots[level] + 1);
2068                         return 0;
2069                 }
2070                 level++;
2071         }
2072         return 1;
2073 }
2074
2075 /*
2076  * merge the relocated tree blocks in reloc tree with corresponding
2077  * fs tree.
2078  */
2079 static noinline_for_stack int merge_reloc_root(struct reloc_control *rc,
2080                                                struct btrfs_root *root)
2081 {
2082         LIST_HEAD(inode_list);
2083         struct btrfs_key key;
2084         struct btrfs_key next_key;
2085         struct btrfs_trans_handle *trans = NULL;
2086         struct btrfs_root *reloc_root;
2087         struct btrfs_root_item *root_item;
2088         struct btrfs_path *path;
2089         struct extent_buffer *leaf;
2090         int level;
2091         int max_level;
2092         int replaced = 0;
2093         int ret;
2094         int err = 0;
2095         u32 min_reserved;
2096
2097         path = btrfs_alloc_path();
2098         if (!path)
2099                 return -ENOMEM;
2100         path->reada = 1;
2101
2102         reloc_root = root->reloc_root;
2103         root_item = &reloc_root->root_item;
2104
2105         if (btrfs_disk_key_objectid(&root_item->drop_progress) == 0) {
2106                 level = btrfs_root_level(root_item);
2107                 extent_buffer_get(reloc_root->node);
2108                 path->nodes[level] = reloc_root->node;
2109                 path->slots[level] = 0;
2110         } else {
2111                 btrfs_disk_key_to_cpu(&key, &root_item->drop_progress);
2112
2113                 level = root_item->drop_level;
2114                 BUG_ON(level == 0);
2115                 path->lowest_level = level;
2116                 ret = btrfs_search_slot(NULL, reloc_root, &key, path, 0, 0);
2117                 path->lowest_level = 0;
2118                 if (ret < 0) {
2119                         btrfs_free_path(path);
2120                         return ret;
2121                 }
2122
2123                 btrfs_node_key_to_cpu(path->nodes[level], &next_key,
2124                                       path->slots[level]);
2125                 WARN_ON(memcmp(&key, &next_key, sizeof(key)));
2126
2127                 btrfs_unlock_up_safe(path, 0);
2128         }
2129
2130         min_reserved = root->nodesize * (BTRFS_MAX_LEVEL - 1) * 2;
2131         memset(&next_key, 0, sizeof(next_key));
2132
2133         while (1) {
2134                 ret = btrfs_block_rsv_refill(root, rc->block_rsv, min_reserved,
2135                                              BTRFS_RESERVE_FLUSH_ALL);
2136                 if (ret) {
2137                         err = ret;
2138                         goto out;
2139                 }
2140                 trans = btrfs_start_transaction(root, 0);
2141                 if (IS_ERR(trans)) {
2142                         err = PTR_ERR(trans);
2143                         trans = NULL;
2144                         goto out;
2145                 }
2146                 trans->block_rsv = rc->block_rsv;
2147
2148                 replaced = 0;
2149                 max_level = level;
2150
2151                 ret = walk_down_reloc_tree(reloc_root, path, &level);
2152                 if (ret < 0) {
2153                         err = ret;
2154                         goto out;
2155                 }
2156                 if (ret > 0)
2157                         break;
2158
2159                 if (!find_next_key(path, level, &key) &&
2160                     btrfs_comp_cpu_keys(&next_key, &key) >= 0) {
2161                         ret = 0;
2162                 } else {
2163                         ret = replace_path(trans, root, reloc_root, path,
2164                                            &next_key, level, max_level);
2165                 }
2166                 if (ret < 0) {
2167                         err = ret;
2168                         goto out;
2169                 }
2170
2171                 if (ret > 0) {
2172                         level = ret;
2173                         btrfs_node_key_to_cpu(path->nodes[level], &key,
2174                                               path->slots[level]);
2175                         replaced = 1;
2176                 }
2177
2178                 ret = walk_up_reloc_tree(reloc_root, path, &level);
2179                 if (ret > 0)
2180                         break;
2181
2182                 BUG_ON(level == 0);
2183                 /*
2184                  * save the merging progress in the drop_progress.
2185                  * this is OK since root refs == 1 in this case.
2186                  */
2187                 btrfs_node_key(path->nodes[level], &root_item->drop_progress,
2188                                path->slots[level]);
2189                 root_item->drop_level = level;
2190
2191                 btrfs_end_transaction_throttle(trans, root);
2192                 trans = NULL;
2193
2194                 btrfs_btree_balance_dirty(root);
2195
2196                 if (replaced && rc->stage == UPDATE_DATA_PTRS)
2197                         invalidate_extent_cache(root, &key, &next_key);
2198         }
2199
2200         /*
2201          * handle the case only one block in the fs tree need to be
2202          * relocated and the block is tree root.
2203          */
2204         leaf = btrfs_lock_root_node(root);
2205         ret = btrfs_cow_block(trans, root, leaf, NULL, 0, &leaf);
2206         btrfs_tree_unlock(leaf);
2207         free_extent_buffer(leaf);
2208         if (ret < 0)
2209                 err = ret;
2210 out:
2211         btrfs_free_path(path);
2212
2213         if (err == 0) {
2214                 memset(&root_item->drop_progress, 0,
2215                        sizeof(root_item->drop_progress));
2216                 root_item->drop_level = 0;
2217                 btrfs_set_root_refs(root_item, 0);
2218                 btrfs_update_reloc_root(trans, root);
2219         }
2220
2221         if (trans)
2222                 btrfs_end_transaction_throttle(trans, root);
2223
2224         btrfs_btree_balance_dirty(root);
2225
2226         if (replaced && rc->stage == UPDATE_DATA_PTRS)
2227                 invalidate_extent_cache(root, &key, &next_key);
2228
2229         return err;
2230 }
2231
2232 static noinline_for_stack
2233 int prepare_to_merge(struct reloc_control *rc, int err)
2234 {
2235         struct btrfs_root *root = rc->extent_root;
2236         struct btrfs_root *reloc_root;
2237         struct btrfs_trans_handle *trans;
2238         LIST_HEAD(reloc_roots);
2239         u64 num_bytes = 0;
2240         int ret;
2241
2242         mutex_lock(&root->fs_info->reloc_mutex);
2243         rc->merging_rsv_size += root->nodesize * (BTRFS_MAX_LEVEL - 1) * 2;
2244         rc->merging_rsv_size += rc->nodes_relocated * 2;
2245         mutex_unlock(&root->fs_info->reloc_mutex);
2246
2247 again:
2248         if (!err) {
2249                 num_bytes = rc->merging_rsv_size;
2250                 ret = btrfs_block_rsv_add(root, rc->block_rsv, num_bytes,
2251                                           BTRFS_RESERVE_FLUSH_ALL);
2252                 if (ret)
2253                         err = ret;
2254         }
2255
2256         trans = btrfs_join_transaction(rc->extent_root);
2257         if (IS_ERR(trans)) {
2258                 if (!err)
2259                         btrfs_block_rsv_release(rc->extent_root,
2260                                                 rc->block_rsv, num_bytes);
2261                 return PTR_ERR(trans);
2262         }
2263
2264         if (!err) {
2265                 if (num_bytes != rc->merging_rsv_size) {
2266                         btrfs_end_transaction(trans, rc->extent_root);
2267                         btrfs_block_rsv_release(rc->extent_root,
2268                                                 rc->block_rsv, num_bytes);
2269                         goto again;
2270                 }
2271         }
2272
2273         rc->merge_reloc_tree = 1;
2274
2275         while (!list_empty(&rc->reloc_roots)) {
2276                 reloc_root = list_entry(rc->reloc_roots.next,
2277                                         struct btrfs_root, root_list);
2278                 list_del_init(&reloc_root->root_list);
2279
2280                 root = read_fs_root(reloc_root->fs_info,
2281                                     reloc_root->root_key.offset);
2282                 BUG_ON(IS_ERR(root));
2283                 BUG_ON(root->reloc_root != reloc_root);
2284
2285                 /*
2286                  * set reference count to 1, so btrfs_recover_relocation
2287                  * knows it should resumes merging
2288                  */
2289                 if (!err)
2290                         btrfs_set_root_refs(&reloc_root->root_item, 1);
2291                 btrfs_update_reloc_root(trans, root);
2292
2293                 list_add(&reloc_root->root_list, &reloc_roots);
2294         }
2295
2296         list_splice(&reloc_roots, &rc->reloc_roots);
2297
2298         if (!err)
2299                 btrfs_commit_transaction(trans, rc->extent_root);
2300         else
2301                 btrfs_end_transaction(trans, rc->extent_root);
2302         return err;
2303 }
2304
2305 static noinline_for_stack
2306 void free_reloc_roots(struct list_head *list)
2307 {
2308         struct btrfs_root *reloc_root;
2309
2310         while (!list_empty(list)) {
2311                 reloc_root = list_entry(list->next, struct btrfs_root,
2312                                         root_list);
2313                 __del_reloc_root(reloc_root);
2314                 free_extent_buffer(reloc_root->node);
2315                 free_extent_buffer(reloc_root->commit_root);
2316                 kfree(reloc_root);
2317         }
2318 }
2319
2320 static noinline_for_stack
2321 int merge_reloc_roots(struct reloc_control *rc)
2322 {
2323         struct btrfs_trans_handle *trans;
2324         struct btrfs_root *root;
2325         struct btrfs_root *reloc_root;
2326         u64 last_snap;
2327         u64 otransid;
2328         u64 objectid;
2329         LIST_HEAD(reloc_roots);
2330         int found = 0;
2331         int ret = 0;
2332 again:
2333         root = rc->extent_root;
2334
2335         /*
2336          * this serializes us with btrfs_record_root_in_transaction,
2337          * we have to make sure nobody is in the middle of
2338          * adding their roots to the list while we are
2339          * doing this splice
2340          */
2341         mutex_lock(&root->fs_info->reloc_mutex);
2342         list_splice_init(&rc->reloc_roots, &reloc_roots);
2343         mutex_unlock(&root->fs_info->reloc_mutex);
2344
2345         while (!list_empty(&reloc_roots)) {
2346                 found = 1;
2347                 reloc_root = list_entry(reloc_roots.next,
2348                                         struct btrfs_root, root_list);
2349
2350                 if (btrfs_root_refs(&reloc_root->root_item) > 0) {
2351                         root = read_fs_root(reloc_root->fs_info,
2352                                             reloc_root->root_key.offset);
2353                         BUG_ON(IS_ERR(root));
2354                         BUG_ON(root->reloc_root != reloc_root);
2355
2356                         ret = merge_reloc_root(rc, root);
2357                         if (ret) {
2358                                 __del_reloc_root(reloc_root);
2359                                 free_extent_buffer(reloc_root->node);
2360                                 free_extent_buffer(reloc_root->commit_root);
2361                                 kfree(reloc_root);
2362                                 goto out;
2363                         }
2364                 } else {
2365                         list_del_init(&reloc_root->root_list);
2366                 }
2367
2368                 /*
2369                  * we keep the old last snapshod transid in rtranid when we
2370                  * created the relocation tree.
2371                  */
2372                 last_snap = btrfs_root_rtransid(&reloc_root->root_item);
2373                 otransid = btrfs_root_otransid(&reloc_root->root_item);
2374                 objectid = reloc_root->root_key.offset;
2375
2376                 ret = btrfs_drop_snapshot(reloc_root, rc->block_rsv, 0, 1);
2377                 if (ret < 0) {
2378                         if (list_empty(&reloc_root->root_list))
2379                                 list_add_tail(&reloc_root->root_list,
2380                                               &reloc_roots);
2381                         goto out;
2382                 } else if (!ret) {
2383                         /*
2384                          * recover the last snapshot tranid to avoid
2385                          * the space balance break NOCOW.
2386                          */
2387                         root = read_fs_root(rc->extent_root->fs_info,
2388                                             objectid);
2389                         if (IS_ERR(root))
2390                                 continue;
2391
2392                         trans = btrfs_join_transaction(root);
2393                         BUG_ON(IS_ERR(trans));
2394
2395                         /* Check if the fs/file tree was snapshoted or not. */
2396                         if (btrfs_root_last_snapshot(&root->root_item) ==
2397                             otransid - 1)
2398                                 btrfs_set_root_last_snapshot(&root->root_item,
2399                                                              last_snap);
2400                                 
2401                         btrfs_end_transaction(trans, root);
2402                 }
2403         }
2404
2405         if (found) {
2406                 found = 0;
2407                 goto again;
2408         }
2409 out:
2410         if (ret) {
2411                 btrfs_std_error(root->fs_info, ret);
2412                 if (!list_empty(&reloc_roots))
2413                         free_reloc_roots(&reloc_roots);
2414
2415                 /* new reloc root may be added */
2416                 mutex_lock(&root->fs_info->reloc_mutex);
2417                 list_splice_init(&rc->reloc_roots, &reloc_roots);
2418                 mutex_unlock(&root->fs_info->reloc_mutex);
2419                 if (!list_empty(&reloc_roots))
2420                         free_reloc_roots(&reloc_roots);
2421         }
2422
2423         BUG_ON(!RB_EMPTY_ROOT(&rc->reloc_root_tree.rb_root));
2424         return ret;
2425 }
2426
2427 static void free_block_list(struct rb_root *blocks)
2428 {
2429         struct tree_block *block;
2430         struct rb_node *rb_node;
2431         while ((rb_node = rb_first(blocks))) {
2432                 block = rb_entry(rb_node, struct tree_block, rb_node);
2433                 rb_erase(rb_node, blocks);
2434                 kfree(block);
2435         }
2436 }
2437
2438 static int record_reloc_root_in_trans(struct btrfs_trans_handle *trans,
2439                                       struct btrfs_root *reloc_root)
2440 {
2441         struct btrfs_root *root;
2442
2443         if (reloc_root->last_trans == trans->transid)
2444                 return 0;
2445
2446         root = read_fs_root(reloc_root->fs_info, reloc_root->root_key.offset);
2447         BUG_ON(IS_ERR(root));
2448         BUG_ON(root->reloc_root != reloc_root);
2449
2450         return btrfs_record_root_in_trans(trans, root);
2451 }
2452
2453 static noinline_for_stack
2454 struct btrfs_root *select_reloc_root(struct btrfs_trans_handle *trans,
2455                                      struct reloc_control *rc,
2456                                      struct backref_node *node,
2457                                      struct backref_edge *edges[], int *nr)
2458 {
2459         struct backref_node *next;
2460         struct btrfs_root *root;
2461         int index = 0;
2462
2463         next = node;
2464         while (1) {
2465                 cond_resched();
2466                 next = walk_up_backref(next, edges, &index);
2467                 root = next->root;
2468                 BUG_ON(!root);
2469                 BUG_ON(!root->ref_cows);
2470
2471                 if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID) {
2472                         record_reloc_root_in_trans(trans, root);
2473                         break;
2474                 }
2475
2476                 btrfs_record_root_in_trans(trans, root);
2477                 root = root->reloc_root;
2478
2479                 if (next->new_bytenr != root->node->start) {
2480                         BUG_ON(next->new_bytenr);
2481                         BUG_ON(!list_empty(&next->list));
2482                         next->new_bytenr = root->node->start;
2483                         next->root = root;
2484                         list_add_tail(&next->list,
2485                                       &rc->backref_cache.changed);
2486                         __mark_block_processed(rc, next);
2487                         break;
2488                 }
2489
2490                 WARN_ON(1);
2491                 root = NULL;
2492                 next = walk_down_backref(edges, &index);
2493                 if (!next || next->level <= node->level)
2494                         break;
2495         }
2496         if (!root)
2497                 return NULL;
2498
2499         *nr = index;
2500         next = node;
2501         /* setup backref node path for btrfs_reloc_cow_block */
2502         while (1) {
2503                 rc->backref_cache.path[next->level] = next;
2504                 if (--index < 0)
2505                         break;
2506                 next = edges[index]->node[UPPER];
2507         }
2508         return root;
2509 }
2510
2511 /*
2512  * select a tree root for relocation. return NULL if the block
2513  * is reference counted. we should use do_relocation() in this
2514  * case. return a tree root pointer if the block isn't reference
2515  * counted. return -ENOENT if the block is root of reloc tree.
2516  */
2517 static noinline_for_stack
2518 struct btrfs_root *select_one_root(struct btrfs_trans_handle *trans,
2519                                    struct backref_node *node)
2520 {
2521         struct backref_node *next;
2522         struct btrfs_root *root;
2523         struct btrfs_root *fs_root = NULL;
2524         struct backref_edge *edges[BTRFS_MAX_LEVEL - 1];
2525         int index = 0;
2526
2527         next = node;
2528         while (1) {
2529                 cond_resched();
2530                 next = walk_up_backref(next, edges, &index);
2531                 root = next->root;
2532                 BUG_ON(!root);
2533
2534                 /* no other choice for non-references counted tree */
2535                 if (!root->ref_cows)
2536                         return root;
2537
2538                 if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID)
2539                         fs_root = root;
2540
2541                 if (next != node)
2542                         return NULL;
2543
2544                 next = walk_down_backref(edges, &index);
2545                 if (!next || next->level <= node->level)
2546                         break;
2547         }
2548
2549         if (!fs_root)
2550                 return ERR_PTR(-ENOENT);
2551         return fs_root;
2552 }
2553
2554 static noinline_for_stack
2555 u64 calcu_metadata_size(struct reloc_control *rc,
2556                         struct backref_node *node, int reserve)
2557 {
2558         struct backref_node *next = node;
2559         struct backref_edge *edge;
2560         struct backref_edge *edges[BTRFS_MAX_LEVEL - 1];
2561         u64 num_bytes = 0;
2562         int index = 0;
2563
2564         BUG_ON(reserve && node->processed);
2565
2566         while (next) {
2567                 cond_resched();
2568                 while (1) {
2569                         if (next->processed && (reserve || next != node))
2570                                 break;
2571
2572                         num_bytes += btrfs_level_size(rc->extent_root,
2573                                                       next->level);
2574
2575                         if (list_empty(&next->upper))
2576                                 break;
2577
2578                         edge = list_entry(next->upper.next,
2579                                           struct backref_edge, list[LOWER]);
2580                         edges[index++] = edge;
2581                         next = edge->node[UPPER];
2582                 }
2583                 next = walk_down_backref(edges, &index);
2584         }
2585         return num_bytes;
2586 }
2587
2588 static int reserve_metadata_space(struct btrfs_trans_handle *trans,
2589                                   struct reloc_control *rc,
2590                                   struct backref_node *node)
2591 {
2592         struct btrfs_root *root = rc->extent_root;
2593         u64 num_bytes;
2594         int ret;
2595         u64 tmp;
2596
2597         num_bytes = calcu_metadata_size(rc, node, 1) * 2;
2598
2599         trans->block_rsv = rc->block_rsv;
2600         rc->reserved_bytes += num_bytes;
2601         ret = btrfs_block_rsv_refill(root, rc->block_rsv, num_bytes,
2602                                 BTRFS_RESERVE_FLUSH_ALL);
2603         if (ret) {
2604                 if (ret == -EAGAIN) {
2605                         tmp = rc->extent_root->nodesize *
2606                                 RELOCATION_RESERVED_NODES;
2607                         while (tmp <= rc->reserved_bytes)
2608                                 tmp <<= 1;
2609                         /*
2610                          * only one thread can access block_rsv at this point,
2611                          * so we don't need hold lock to protect block_rsv.
2612                          * we expand more reservation size here to allow enough
2613                          * space for relocation and we will return eailer in
2614                          * enospc case.
2615                          */
2616                         rc->block_rsv->size = tmp + rc->extent_root->nodesize *
2617                                               RELOCATION_RESERVED_NODES;
2618                 }
2619                 return ret;
2620         }
2621
2622         return 0;
2623 }
2624
2625 /*
2626  * relocate a block tree, and then update pointers in upper level
2627  * blocks that reference the block to point to the new location.
2628  *
2629  * if called by link_to_upper, the block has already been relocated.
2630  * in that case this function just updates pointers.
2631  */
2632 static int do_relocation(struct btrfs_trans_handle *trans,
2633                          struct reloc_control *rc,
2634                          struct backref_node *node,
2635                          struct btrfs_key *key,
2636                          struct btrfs_path *path, int lowest)
2637 {
2638         struct backref_node *upper;
2639         struct backref_edge *edge;
2640         struct backref_edge *edges[BTRFS_MAX_LEVEL - 1];
2641         struct btrfs_root *root;
2642         struct extent_buffer *eb;
2643         u32 blocksize;
2644         u64 bytenr;
2645         u64 generation;
2646         int nr;
2647         int slot;
2648         int ret;
2649         int err = 0;
2650
2651         BUG_ON(lowest && node->eb);
2652
2653         path->lowest_level = node->level + 1;
2654         rc->backref_cache.path[node->level] = node;
2655         list_for_each_entry(edge, &node->upper, list[LOWER]) {
2656                 cond_resched();
2657
2658                 upper = edge->node[UPPER];
2659                 root = select_reloc_root(trans, rc, upper, edges, &nr);
2660                 BUG_ON(!root);
2661
2662                 if (upper->eb && !upper->locked) {
2663                         if (!lowest) {
2664                                 ret = btrfs_bin_search(upper->eb, key,
2665                                                        upper->level, &slot);
2666                                 BUG_ON(ret);
2667                                 bytenr = btrfs_node_blockptr(upper->eb, slot);
2668                                 if (node->eb->start == bytenr)
2669                                         goto next;
2670                         }
2671                         drop_node_buffer(upper);
2672                 }
2673
2674                 if (!upper->eb) {
2675                         ret = btrfs_search_slot(trans, root, key, path, 0, 1);
2676                         if (ret < 0) {
2677                                 err = ret;
2678                                 break;
2679                         }
2680                         BUG_ON(ret > 0);
2681
2682                         if (!upper->eb) {
2683                                 upper->eb = path->nodes[upper->level];
2684                                 path->nodes[upper->level] = NULL;
2685                         } else {
2686                                 BUG_ON(upper->eb != path->nodes[upper->level]);
2687                         }
2688
2689                         upper->locked = 1;
2690                         path->locks[upper->level] = 0;
2691
2692                         slot = path->slots[upper->level];
2693                         btrfs_release_path(path);
2694                 } else {
2695                         ret = btrfs_bin_search(upper->eb, key, upper->level,
2696                                                &slot);
2697                         BUG_ON(ret);
2698                 }
2699
2700                 bytenr = btrfs_node_blockptr(upper->eb, slot);
2701                 if (lowest) {
2702                         BUG_ON(bytenr != node->bytenr);
2703                 } else {
2704                         if (node->eb->start == bytenr)
2705                                 goto next;
2706                 }
2707
2708                 blocksize = btrfs_level_size(root, node->level);
2709                 generation = btrfs_node_ptr_generation(upper->eb, slot);
2710                 eb = read_tree_block(root, bytenr, blocksize, generation);
2711                 if (!eb || !extent_buffer_uptodate(eb)) {
2712                         free_extent_buffer(eb);
2713                         err = -EIO;
2714                         goto next;
2715                 }
2716                 btrfs_tree_lock(eb);
2717                 btrfs_set_lock_blocking(eb);
2718
2719                 if (!node->eb) {
2720                         ret = btrfs_cow_block(trans, root, eb, upper->eb,
2721                                               slot, &eb);
2722                         btrfs_tree_unlock(eb);
2723                         free_extent_buffer(eb);
2724                         if (ret < 0) {
2725                                 err = ret;
2726                                 goto next;
2727                         }
2728                         BUG_ON(node->eb != eb);
2729                 } else {
2730                         btrfs_set_node_blockptr(upper->eb, slot,
2731                                                 node->eb->start);
2732                         btrfs_set_node_ptr_generation(upper->eb, slot,
2733                                                       trans->transid);
2734                         btrfs_mark_buffer_dirty(upper->eb);
2735
2736                         ret = btrfs_inc_extent_ref(trans, root,
2737                                                 node->eb->start, blocksize,
2738                                                 upper->eb->start,
2739                                                 btrfs_header_owner(upper->eb),
2740                                                 node->level, 0, 1);
2741                         BUG_ON(ret);
2742
2743                         ret = btrfs_drop_subtree(trans, root, eb, upper->eb);
2744                         BUG_ON(ret);
2745                 }
2746 next:
2747                 if (!upper->pending)
2748                         drop_node_buffer(upper);
2749                 else
2750                         unlock_node_buffer(upper);
2751                 if (err)
2752                         break;
2753         }
2754
2755         if (!err && node->pending) {
2756                 drop_node_buffer(node);
2757                 list_move_tail(&node->list, &rc->backref_cache.changed);
2758                 node->pending = 0;
2759         }
2760
2761         path->lowest_level = 0;
2762         BUG_ON(err == -ENOSPC);
2763         return err;
2764 }
2765
2766 static int link_to_upper(struct btrfs_trans_handle *trans,
2767                          struct reloc_control *rc,
2768                          struct backref_node *node,
2769                          struct btrfs_path *path)
2770 {
2771         struct btrfs_key key;
2772
2773         btrfs_node_key_to_cpu(node->eb, &key, 0);
2774         return do_relocation(trans, rc, node, &key, path, 0);
2775 }
2776
2777 static int finish_pending_nodes(struct btrfs_trans_handle *trans,
2778                                 struct reloc_control *rc,
2779                                 struct btrfs_path *path, int err)
2780 {
2781         LIST_HEAD(list);
2782         struct backref_cache *cache = &rc->backref_cache;
2783         struct backref_node *node;
2784         int level;
2785         int ret;
2786
2787         for (level = 0; level < BTRFS_MAX_LEVEL; level++) {
2788                 while (!list_empty(&cache->pending[level])) {
2789                         node = list_entry(cache->pending[level].next,
2790                                           struct backref_node, list);
2791                         list_move_tail(&node->list, &list);
2792                         BUG_ON(!node->pending);
2793
2794                         if (!err) {
2795                                 ret = link_to_upper(trans, rc, node, path);
2796                                 if (ret < 0)
2797                                         err = ret;
2798                         }
2799                 }
2800                 list_splice_init(&list, &cache->pending[level]);
2801         }
2802         return err;
2803 }
2804
2805 static void mark_block_processed(struct reloc_control *rc,
2806                                  u64 bytenr, u32 blocksize)
2807 {
2808         set_extent_bits(&rc->processed_blocks, bytenr, bytenr + blocksize - 1,
2809                         EXTENT_DIRTY, GFP_NOFS);
2810 }
2811
2812 static void __mark_block_processed(struct reloc_control *rc,
2813                                    struct backref_node *node)
2814 {
2815         u32 blocksize;
2816         if (node->level == 0 ||
2817             in_block_group(node->bytenr, rc->block_group)) {
2818                 blocksize = btrfs_level_size(rc->extent_root, node->level);
2819                 mark_block_processed(rc, node->bytenr, blocksize);
2820         }
2821         node->processed = 1;
2822 }
2823
2824 /*
2825  * mark a block and all blocks directly/indirectly reference the block
2826  * as processed.
2827  */
2828 static void update_processed_blocks(struct reloc_control *rc,
2829                                     struct backref_node *node)
2830 {
2831         struct backref_node *next = node;
2832         struct backref_edge *edge;
2833         struct backref_edge *edges[BTRFS_MAX_LEVEL - 1];
2834         int index = 0;
2835
2836         while (next) {
2837                 cond_resched();
2838                 while (1) {
2839                         if (next->processed)
2840                                 break;
2841
2842                         __mark_block_processed(rc, next);
2843
2844                         if (list_empty(&next->upper))
2845                                 break;
2846
2847                         edge = list_entry(next->upper.next,
2848                                           struct backref_edge, list[LOWER]);
2849                         edges[index++] = edge;
2850                         next = edge->node[UPPER];
2851                 }
2852                 next = walk_down_backref(edges, &index);
2853         }
2854 }
2855
2856 static int tree_block_processed(u64 bytenr, u32 blocksize,
2857                                 struct reloc_control *rc)
2858 {
2859         if (test_range_bit(&rc->processed_blocks, bytenr,
2860                            bytenr + blocksize - 1, EXTENT_DIRTY, 1, NULL))
2861                 return 1;
2862         return 0;
2863 }
2864
2865 static int get_tree_block_key(struct reloc_control *rc,
2866                               struct tree_block *block)
2867 {
2868         struct extent_buffer *eb;
2869
2870         BUG_ON(block->key_ready);
2871         eb = read_tree_block(rc->extent_root, block->bytenr,
2872                              block->key.objectid, block->key.offset);
2873         if (!eb || !extent_buffer_uptodate(eb)) {
2874                 free_extent_buffer(eb);
2875                 return -EIO;
2876         }
2877         WARN_ON(btrfs_header_level(eb) != block->level);
2878         if (block->level == 0)
2879                 btrfs_item_key_to_cpu(eb, &block->key, 0);
2880         else
2881                 btrfs_node_key_to_cpu(eb, &block->key, 0);
2882         free_extent_buffer(eb);
2883         block->key_ready = 1;
2884         return 0;
2885 }
2886
2887 static int reada_tree_block(struct reloc_control *rc,
2888                             struct tree_block *block)
2889 {
2890         BUG_ON(block->key_ready);
2891         if (block->key.type == BTRFS_METADATA_ITEM_KEY)
2892                 readahead_tree_block(rc->extent_root, block->bytenr,
2893                                      block->key.objectid,
2894                                      rc->extent_root->leafsize);
2895         else
2896                 readahead_tree_block(rc->extent_root, block->bytenr,
2897                                      block->key.objectid, block->key.offset);
2898         return 0;
2899 }
2900
2901 /*
2902  * helper function to relocate a tree block
2903  */
2904 static int relocate_tree_block(struct btrfs_trans_handle *trans,
2905                                 struct reloc_control *rc,
2906                                 struct backref_node *node,
2907                                 struct btrfs_key *key,
2908                                 struct btrfs_path *path)
2909 {
2910         struct btrfs_root *root;
2911         int ret = 0;
2912
2913         if (!node)
2914                 return 0;
2915
2916         BUG_ON(node->processed);
2917         root = select_one_root(trans, node);
2918         if (root == ERR_PTR(-ENOENT)) {
2919                 update_processed_blocks(rc, node);
2920                 goto out;
2921         }
2922
2923         if (!root || root->ref_cows) {
2924                 ret = reserve_metadata_space(trans, rc, node);
2925                 if (ret)
2926                         goto out;
2927         }
2928
2929         if (root) {
2930                 if (root->ref_cows) {
2931                         BUG_ON(node->new_bytenr);
2932                         BUG_ON(!list_empty(&node->list));
2933                         btrfs_record_root_in_trans(trans, root);
2934                         root = root->reloc_root;
2935                         node->new_bytenr = root->node->start;
2936                         node->root = root;
2937                         list_add_tail(&node->list, &rc->backref_cache.changed);
2938                 } else {
2939                         path->lowest_level = node->level;
2940                         ret = btrfs_search_slot(trans, root, key, path, 0, 1);
2941                         btrfs_release_path(path);
2942                         if (ret > 0)
2943                                 ret = 0;
2944                 }
2945                 if (!ret)
2946                         update_processed_blocks(rc, node);
2947         } else {
2948                 ret = do_relocation(trans, rc, node, key, path, 1);
2949         }
2950 out:
2951         if (ret || node->level == 0 || node->cowonly)
2952                 remove_backref_node(&rc->backref_cache, node);
2953         return ret;
2954 }
2955
2956 /*
2957  * relocate a list of blocks
2958  */
2959 static noinline_for_stack
2960 int relocate_tree_blocks(struct btrfs_trans_handle *trans,
2961                          struct reloc_control *rc, struct rb_root *blocks)
2962 {
2963         struct backref_node *node;
2964         struct btrfs_path *path;
2965         struct tree_block *block;
2966         struct rb_node *rb_node;
2967         int ret;
2968         int err = 0;
2969
2970         path = btrfs_alloc_path();
2971         if (!path) {
2972                 err = -ENOMEM;
2973                 goto out_free_blocks;
2974         }
2975
2976         rb_node = rb_first(blocks);
2977         while (rb_node) {
2978                 block = rb_entry(rb_node, struct tree_block, rb_node);
2979                 if (!block->key_ready)
2980                         reada_tree_block(rc, block);
2981                 rb_node = rb_next(rb_node);
2982         }
2983
2984         rb_node = rb_first(blocks);
2985         while (rb_node) {
2986                 block = rb_entry(rb_node, struct tree_block, rb_node);
2987                 if (!block->key_ready) {
2988                         err = get_tree_block_key(rc, block);
2989                         if (err)
2990                                 goto out_free_path;
2991                 }
2992                 rb_node = rb_next(rb_node);
2993         }
2994
2995         rb_node = rb_first(blocks);
2996         while (rb_node) {
2997                 block = rb_entry(rb_node, struct tree_block, rb_node);
2998
2999                 node = build_backref_tree(rc, &block->key,
3000                                           block->level, block->bytenr);
3001                 if (IS_ERR(node)) {
3002                         err = PTR_ERR(node);
3003                         goto out;
3004                 }
3005
3006                 ret = relocate_tree_block(trans, rc, node, &block->key,
3007                                           path);
3008                 if (ret < 0) {
3009                         if (ret != -EAGAIN || rb_node == rb_first(blocks))
3010                                 err = ret;
3011                         goto out;
3012                 }
3013                 rb_node = rb_next(rb_node);
3014         }
3015 out:
3016         err = finish_pending_nodes(trans, rc, path, err);
3017
3018 out_free_path:
3019         btrfs_free_path(path);
3020 out_free_blocks:
3021         free_block_list(blocks);
3022         return err;
3023 }
3024
3025 static noinline_for_stack
3026 int prealloc_file_extent_cluster(struct inode *inode,
3027                                  struct file_extent_cluster *cluster)
3028 {
3029         u64 alloc_hint = 0;
3030         u64 start;
3031         u64 end;
3032         u64 offset = BTRFS_I(inode)->index_cnt;
3033         u64 num_bytes;
3034         int nr = 0;
3035         int ret = 0;
3036
3037         BUG_ON(cluster->start != cluster->boundary[0]);
3038         mutex_lock(&inode->i_mutex);
3039
3040         ret = btrfs_check_data_free_space(inode, cluster->end +
3041                                           1 - cluster->start);
3042         if (ret)
3043                 goto out;
3044
3045         while (nr < cluster->nr) {
3046                 start = cluster->boundary[nr] - offset;
3047                 if (nr + 1 < cluster->nr)
3048                         end = cluster->boundary[nr + 1] - 1 - offset;
3049                 else
3050                         end = cluster->end - offset;
3051
3052                 lock_extent(&BTRFS_I(inode)->io_tree, start, end);
3053                 num_bytes = end + 1 - start;
3054                 ret = btrfs_prealloc_file_range(inode, 0, start,
3055                                                 num_bytes, num_bytes,
3056                                                 end + 1, &alloc_hint);
3057                 unlock_extent(&BTRFS_I(inode)->io_tree, start, end);
3058                 if (ret)
3059                         break;
3060                 nr++;
3061         }
3062         btrfs_free_reserved_data_space(inode, cluster->end +
3063                                        1 - cluster->start);
3064 out:
3065         mutex_unlock(&inode->i_mutex);
3066         return ret;
3067 }
3068
3069 static noinline_for_stack
3070 int setup_extent_mapping(struct inode *inode, u64 start, u64 end,
3071                          u64 block_start)
3072 {
3073         struct btrfs_root *root = BTRFS_I(inode)->root;
3074         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
3075         struct extent_map *em;
3076         int ret = 0;
3077
3078         em = alloc_extent_map();
3079         if (!em)
3080                 return -ENOMEM;
3081
3082         em->start = start;
3083         em->len = end + 1 - start;
3084         em->block_len = em->len;
3085         em->block_start = block_start;
3086         em->bdev = root->fs_info->fs_devices->latest_bdev;
3087         set_bit(EXTENT_FLAG_PINNED, &em->flags);
3088
3089         lock_extent(&BTRFS_I(inode)->io_tree, start, end);
3090         while (1) {
3091                 write_lock(&em_tree->lock);
3092                 ret = add_extent_mapping(em_tree, em, 0);
3093                 write_unlock(&em_tree->lock);
3094                 if (ret != -EEXIST) {
3095                         free_extent_map(em);
3096                         break;
3097                 }
3098                 btrfs_drop_extent_cache(inode, start, end, 0);
3099         }
3100         unlock_extent(&BTRFS_I(inode)->io_tree, start, end);
3101         return ret;
3102 }
3103
3104 static int relocate_file_extent_cluster(struct inode *inode,
3105                                         struct file_extent_cluster *cluster)
3106 {
3107         u64 page_start;
3108         u64 page_end;
3109         u64 offset = BTRFS_I(inode)->index_cnt;
3110         unsigned long index;
3111         unsigned long last_index;
3112         struct page *page;
3113         struct file_ra_state *ra;
3114         gfp_t mask = btrfs_alloc_write_mask(inode->i_mapping);
3115         int nr = 0;
3116         int ret = 0;
3117
3118         if (!cluster->nr)
3119                 return 0;
3120
3121         ra = kzalloc(sizeof(*ra), GFP_NOFS);
3122         if (!ra)
3123                 return -ENOMEM;
3124
3125         ret = prealloc_file_extent_cluster(inode, cluster);
3126         if (ret)
3127                 goto out;
3128
3129         file_ra_state_init(ra, inode->i_mapping);
3130
3131         ret = setup_extent_mapping(inode, cluster->start - offset,
3132                                    cluster->end - offset, cluster->start);
3133         if (ret)
3134                 goto out;
3135
3136         index = (cluster->start - offset) >> PAGE_CACHE_SHIFT;
3137         last_index = (cluster->end - offset) >> PAGE_CACHE_SHIFT;
3138         while (index <= last_index) {
3139                 ret = btrfs_delalloc_reserve_metadata(inode, PAGE_CACHE_SIZE);
3140                 if (ret)
3141                         goto out;
3142
3143                 page = find_lock_page(inode->i_mapping, index);
3144                 if (!page) {
3145                         page_cache_sync_readahead(inode->i_mapping,
3146                                                   ra, NULL, index,
3147                                                   last_index + 1 - index);
3148                         page = find_or_create_page(inode->i_mapping, index,
3149                                                    mask);
3150                         if (!page) {
3151                                 btrfs_delalloc_release_metadata(inode,
3152                                                         PAGE_CACHE_SIZE);
3153                                 ret = -ENOMEM;
3154                                 goto out;
3155                         }
3156                 }
3157
3158                 if (PageReadahead(page)) {
3159                         page_cache_async_readahead(inode->i_mapping,
3160                                                    ra, NULL, page, index,
3161                                                    last_index + 1 - index);
3162                 }
3163
3164                 if (!PageUptodate(page)) {
3165                         btrfs_readpage(NULL, page);
3166                         lock_page(page);
3167                         if (!PageUptodate(page)) {
3168                                 unlock_page(page);
3169                                 page_cache_release(page);
3170                                 btrfs_delalloc_release_metadata(inode,
3171                                                         PAGE_CACHE_SIZE);
3172                                 ret = -EIO;
3173                                 goto out;
3174                         }
3175                 }
3176
3177                 page_start = page_offset(page);
3178                 page_end = page_start + PAGE_CACHE_SIZE - 1;
3179
3180                 lock_extent(&BTRFS_I(inode)->io_tree, page_start, page_end);
3181
3182                 set_page_extent_mapped(page);
3183
3184                 if (nr < cluster->nr &&
3185                     page_start + offset == cluster->boundary[nr]) {
3186                         set_extent_bits(&BTRFS_I(inode)->io_tree,
3187                                         page_start, page_end,
3188                                         EXTENT_BOUNDARY, GFP_NOFS);
3189                         nr++;
3190                 }
3191
3192                 btrfs_set_extent_delalloc(inode, page_start, page_end, NULL);
3193                 set_page_dirty(page);
3194
3195                 unlock_extent(&BTRFS_I(inode)->io_tree,
3196                               page_start, page_end);
3197                 unlock_page(page);
3198                 page_cache_release(page);
3199
3200                 index++;
3201                 balance_dirty_pages_ratelimited(inode->i_mapping);
3202                 btrfs_throttle(BTRFS_I(inode)->root);
3203         }
3204         WARN_ON(nr != cluster->nr);
3205 out:
3206         kfree(ra);
3207         return ret;
3208 }
3209
3210 static noinline_for_stack
3211 int relocate_data_extent(struct inode *inode, struct btrfs_key *extent_key,
3212                          struct file_extent_cluster *cluster)
3213 {
3214         int ret;
3215
3216         if (cluster->nr > 0 && extent_key->objectid != cluster->end + 1) {
3217                 ret = relocate_file_extent_cluster(inode, cluster);
3218                 if (ret)
3219                         return ret;
3220                 cluster->nr = 0;
3221         }
3222
3223         if (!cluster->nr)
3224                 cluster->start = extent_key->objectid;
3225         else
3226                 BUG_ON(cluster->nr >= MAX_EXTENTS);
3227         cluster->end = extent_key->objectid + extent_key->offset - 1;
3228         cluster->boundary[cluster->nr] = extent_key->objectid;
3229         cluster->nr++;
3230
3231         if (cluster->nr >= MAX_EXTENTS) {
3232                 ret = relocate_file_extent_cluster(inode, cluster);
3233                 if (ret)
3234                         return ret;
3235                 cluster->nr = 0;
3236         }
3237         return 0;
3238 }
3239
3240 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
3241 static int get_ref_objectid_v0(struct reloc_control *rc,
3242                                struct btrfs_path *path,
3243                                struct btrfs_key *extent_key,
3244                                u64 *ref_objectid, int *path_change)
3245 {
3246         struct btrfs_key key;
3247         struct extent_buffer *leaf;
3248         struct btrfs_extent_ref_v0 *ref0;
3249         int ret;
3250         int slot;
3251
3252         leaf = path->nodes[0];
3253         slot = path->slots[0];
3254         while (1) {
3255                 if (slot >= btrfs_header_nritems(leaf)) {
3256                         ret = btrfs_next_leaf(rc->extent_root, path);
3257                         if (ret < 0)
3258                                 return ret;
3259                         BUG_ON(ret > 0);
3260                         leaf = path->nodes[0];
3261                         slot = path->slots[0];
3262                         if (path_change)
3263                                 *path_change = 1;
3264                 }
3265                 btrfs_item_key_to_cpu(leaf, &key, slot);
3266                 if (key.objectid != extent_key->objectid)
3267                         return -ENOENT;
3268
3269                 if (key.type != BTRFS_EXTENT_REF_V0_KEY) {
3270                         slot++;
3271                         continue;
3272                 }
3273                 ref0 = btrfs_item_ptr(leaf, slot,
3274                                 struct btrfs_extent_ref_v0);
3275                 *ref_objectid = btrfs_ref_objectid_v0(leaf, ref0);
3276                 break;
3277         }
3278         return 0;
3279 }
3280 #endif
3281
3282 /*
3283  * helper to add a tree block to the list.
3284  * the major work is getting the generation and level of the block
3285  */
3286 static int add_tree_block(struct reloc_control *rc,
3287                           struct btrfs_key *extent_key,
3288                           struct btrfs_path *path,
3289                           struct rb_root *blocks)
3290 {
3291         struct extent_buffer *eb;
3292         struct btrfs_extent_item *ei;
3293         struct btrfs_tree_block_info *bi;
3294         struct tree_block *block;
3295         struct rb_node *rb_node;
3296         u32 item_size;
3297         int level = -1;
3298         u64 generation;
3299
3300         eb =  path->nodes[0];
3301         item_size = btrfs_item_size_nr(eb, path->slots[0]);
3302
3303         if (extent_key->type == BTRFS_METADATA_ITEM_KEY ||
3304             item_size >= sizeof(*ei) + sizeof(*bi)) {
3305                 ei = btrfs_item_ptr(eb, path->slots[0],
3306                                 struct btrfs_extent_item);
3307                 if (extent_key->type == BTRFS_EXTENT_ITEM_KEY) {
3308                         bi = (struct btrfs_tree_block_info *)(ei + 1);
3309                         level = btrfs_tree_block_level(eb, bi);
3310                 } else {
3311                         level = (int)extent_key->offset;
3312                 }
3313                 generation = btrfs_extent_generation(eb, ei);
3314         } else {
3315 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
3316                 u64 ref_owner;
3317                 int ret;
3318
3319                 BUG_ON(item_size != sizeof(struct btrfs_extent_item_v0));
3320                 ret = get_ref_objectid_v0(rc, path, extent_key,
3321                                           &ref_owner, NULL);
3322                 if (ret < 0)
3323                         return ret;
3324                 BUG_ON(ref_owner >= BTRFS_MAX_LEVEL);
3325                 level = (int)ref_owner;
3326                 /* FIXME: get real generation */
3327                 generation = 0;
3328 #else
3329                 BUG();
3330 #endif
3331         }
3332
3333         btrfs_release_path(path);
3334
3335         BUG_ON(level == -1);
3336
3337         block = kmalloc(sizeof(*block), GFP_NOFS);
3338         if (!block)
3339                 return -ENOMEM;
3340
3341         block->bytenr = extent_key->objectid;
3342         block->key.objectid = rc->extent_root->leafsize;
3343         block->key.offset = generation;
3344         block->level = level;
3345         block->key_ready = 0;
3346
3347         rb_node = tree_insert(blocks, block->bytenr, &block->rb_node);
3348         if (rb_node)
3349                 backref_tree_panic(rb_node, -EEXIST, block->bytenr);
3350
3351         return 0;
3352 }
3353
3354 /*
3355  * helper to add tree blocks for backref of type BTRFS_SHARED_DATA_REF_KEY
3356  */
3357 static int __add_tree_block(struct reloc_control *rc,
3358                             u64 bytenr, u32 blocksize,
3359                             struct rb_root *blocks)
3360 {
3361         struct btrfs_path *path;
3362         struct btrfs_key key;
3363         int ret;
3364         bool skinny = btrfs_fs_incompat(rc->extent_root->fs_info,
3365                                         SKINNY_METADATA);
3366
3367         if (tree_block_processed(bytenr, blocksize, rc))
3368                 return 0;
3369
3370         if (tree_search(blocks, bytenr))
3371                 return 0;
3372
3373         path = btrfs_alloc_path();
3374         if (!path)
3375                 return -ENOMEM;
3376 again:
3377         key.objectid = bytenr;
3378         if (skinny) {
3379                 key.type = BTRFS_METADATA_ITEM_KEY;
3380                 key.offset = (u64)-1;
3381         } else {
3382                 key.type = BTRFS_EXTENT_ITEM_KEY;
3383                 key.offset = blocksize;
3384         }
3385
3386         path->search_commit_root = 1;
3387         path->skip_locking = 1;
3388         ret = btrfs_search_slot(NULL, rc->extent_root, &key, path, 0, 0);
3389         if (ret < 0)
3390                 goto out;
3391
3392         if (ret > 0 && skinny) {
3393                 if (path->slots[0]) {
3394                         path->slots[0]--;
3395                         btrfs_item_key_to_cpu(path->nodes[0], &key,
3396                               &