Merge branch 'for-linus' of master.kernel.org:/home/rmk/linux-2.6-arm
[pandora-kernel.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
98 struct backref_cache {
99         /* red black tree of all backref nodes in the cache */
100         struct rb_root rb_root;
101         /* for passing backref nodes to btrfs_reloc_cow_block */
102         struct backref_node *path[BTRFS_MAX_LEVEL];
103         /*
104          * list of blocks that have been cowed but some block
105          * pointers in upper level blocks may not reflect the
106          * new location
107          */
108         struct list_head pending[BTRFS_MAX_LEVEL];
109         /* list of backref nodes with no child node */
110         struct list_head leaves;
111         /* list of blocks that have been cowed in current transaction */
112         struct list_head changed;
113         /* list of detached backref node. */
114         struct list_head detached;
115
116         u64 last_trans;
117
118         int nr_nodes;
119         int nr_edges;
120 };
121
122 /*
123  * map address of tree root to tree
124  */
125 struct mapping_node {
126         struct rb_node rb_node;
127         u64 bytenr;
128         void *data;
129 };
130
131 struct mapping_tree {
132         struct rb_root rb_root;
133         spinlock_t lock;
134 };
135
136 /*
137  * present a tree block to process
138  */
139 struct tree_block {
140         struct rb_node rb_node;
141         u64 bytenr;
142         struct btrfs_key key;
143         unsigned int level:8;
144         unsigned int key_ready:1;
145 };
146
147 #define MAX_EXTENTS 128
148
149 struct file_extent_cluster {
150         u64 start;
151         u64 end;
152         u64 boundary[MAX_EXTENTS];
153         unsigned int nr;
154 };
155
156 struct reloc_control {
157         /* block group to relocate */
158         struct btrfs_block_group_cache *block_group;
159         /* extent tree */
160         struct btrfs_root *extent_root;
161         /* inode for moving data */
162         struct inode *data_inode;
163
164         struct btrfs_block_rsv *block_rsv;
165
166         struct backref_cache backref_cache;
167
168         struct file_extent_cluster cluster;
169         /* tree blocks have been processed */
170         struct extent_io_tree processed_blocks;
171         /* map start of tree root to corresponding reloc tree */
172         struct mapping_tree reloc_root_tree;
173         /* list of reloc trees */
174         struct list_head reloc_roots;
175         /* size of metadata reservation for merging reloc trees */
176         u64 merging_rsv_size;
177         /* size of relocated tree nodes */
178         u64 nodes_relocated;
179
180         u64 search_start;
181         u64 extents_found;
182
183         unsigned int stage:8;
184         unsigned int create_reloc_tree:1;
185         unsigned int merge_reloc_tree:1;
186         unsigned int found_file_extent:1;
187         unsigned int commit_transaction:1;
188 };
189
190 /* stages of data relocation */
191 #define MOVE_DATA_EXTENTS       0
192 #define UPDATE_DATA_PTRS        1
193
194 static void remove_backref_node(struct backref_cache *cache,
195                                 struct backref_node *node);
196 static void __mark_block_processed(struct reloc_control *rc,
197                                    struct backref_node *node);
198
199 static void mapping_tree_init(struct mapping_tree *tree)
200 {
201         tree->rb_root = RB_ROOT;
202         spin_lock_init(&tree->lock);
203 }
204
205 static void backref_cache_init(struct backref_cache *cache)
206 {
207         int i;
208         cache->rb_root = RB_ROOT;
209         for (i = 0; i < BTRFS_MAX_LEVEL; i++)
210                 INIT_LIST_HEAD(&cache->pending[i]);
211         INIT_LIST_HEAD(&cache->changed);
212         INIT_LIST_HEAD(&cache->detached);
213         INIT_LIST_HEAD(&cache->leaves);
214 }
215
216 static void backref_cache_cleanup(struct backref_cache *cache)
217 {
218         struct backref_node *node;
219         int i;
220
221         while (!list_empty(&cache->detached)) {
222                 node = list_entry(cache->detached.next,
223                                   struct backref_node, list);
224                 remove_backref_node(cache, node);
225         }
226
227         while (!list_empty(&cache->leaves)) {
228                 node = list_entry(cache->leaves.next,
229                                   struct backref_node, lower);
230                 remove_backref_node(cache, node);
231         }
232
233         cache->last_trans = 0;
234
235         for (i = 0; i < BTRFS_MAX_LEVEL; i++)
236                 BUG_ON(!list_empty(&cache->pending[i]));
237         BUG_ON(!list_empty(&cache->changed));
238         BUG_ON(!list_empty(&cache->detached));
239         BUG_ON(!RB_EMPTY_ROOT(&cache->rb_root));
240         BUG_ON(cache->nr_nodes);
241         BUG_ON(cache->nr_edges);
242 }
243
244 static struct backref_node *alloc_backref_node(struct backref_cache *cache)
245 {
246         struct backref_node *node;
247
248         node = kzalloc(sizeof(*node), GFP_NOFS);
249         if (node) {
250                 INIT_LIST_HEAD(&node->list);
251                 INIT_LIST_HEAD(&node->upper);
252                 INIT_LIST_HEAD(&node->lower);
253                 RB_CLEAR_NODE(&node->rb_node);
254                 cache->nr_nodes++;
255         }
256         return node;
257 }
258
259 static void free_backref_node(struct backref_cache *cache,
260                               struct backref_node *node)
261 {
262         if (node) {
263                 cache->nr_nodes--;
264                 kfree(node);
265         }
266 }
267
268 static struct backref_edge *alloc_backref_edge(struct backref_cache *cache)
269 {
270         struct backref_edge *edge;
271
272         edge = kzalloc(sizeof(*edge), GFP_NOFS);
273         if (edge)
274                 cache->nr_edges++;
275         return edge;
276 }
277
278 static void free_backref_edge(struct backref_cache *cache,
279                               struct backref_edge *edge)
280 {
281         if (edge) {
282                 cache->nr_edges--;
283                 kfree(edge);
284         }
285 }
286
287 static struct rb_node *tree_insert(struct rb_root *root, u64 bytenr,
288                                    struct rb_node *node)
289 {
290         struct rb_node **p = &root->rb_node;
291         struct rb_node *parent = NULL;
292         struct tree_entry *entry;
293
294         while (*p) {
295                 parent = *p;
296                 entry = rb_entry(parent, struct tree_entry, rb_node);
297
298                 if (bytenr < entry->bytenr)
299                         p = &(*p)->rb_left;
300                 else if (bytenr > entry->bytenr)
301                         p = &(*p)->rb_right;
302                 else
303                         return parent;
304         }
305
306         rb_link_node(node, parent, p);
307         rb_insert_color(node, root);
308         return NULL;
309 }
310
311 static struct rb_node *tree_search(struct rb_root *root, u64 bytenr)
312 {
313         struct rb_node *n = root->rb_node;
314         struct tree_entry *entry;
315
316         while (n) {
317                 entry = rb_entry(n, struct tree_entry, rb_node);
318
319                 if (bytenr < entry->bytenr)
320                         n = n->rb_left;
321                 else if (bytenr > entry->bytenr)
322                         n = n->rb_right;
323                 else
324                         return n;
325         }
326         return NULL;
327 }
328
329 /*
330  * walk up backref nodes until reach node presents tree root
331  */
332 static struct backref_node *walk_up_backref(struct backref_node *node,
333                                             struct backref_edge *edges[],
334                                             int *index)
335 {
336         struct backref_edge *edge;
337         int idx = *index;
338
339         while (!list_empty(&node->upper)) {
340                 edge = list_entry(node->upper.next,
341                                   struct backref_edge, list[LOWER]);
342                 edges[idx++] = edge;
343                 node = edge->node[UPPER];
344         }
345         BUG_ON(node->detached);
346         *index = idx;
347         return node;
348 }
349
350 /*
351  * walk down backref nodes to find start of next reference path
352  */
353 static struct backref_node *walk_down_backref(struct backref_edge *edges[],
354                                               int *index)
355 {
356         struct backref_edge *edge;
357         struct backref_node *lower;
358         int idx = *index;
359
360         while (idx > 0) {
361                 edge = edges[idx - 1];
362                 lower = edge->node[LOWER];
363                 if (list_is_last(&edge->list[LOWER], &lower->upper)) {
364                         idx--;
365                         continue;
366                 }
367                 edge = list_entry(edge->list[LOWER].next,
368                                   struct backref_edge, list[LOWER]);
369                 edges[idx - 1] = edge;
370                 *index = idx;
371                 return edge->node[UPPER];
372         }
373         *index = 0;
374         return NULL;
375 }
376
377 static void unlock_node_buffer(struct backref_node *node)
378 {
379         if (node->locked) {
380                 btrfs_tree_unlock(node->eb);
381                 node->locked = 0;
382         }
383 }
384
385 static void drop_node_buffer(struct backref_node *node)
386 {
387         if (node->eb) {
388                 unlock_node_buffer(node);
389                 free_extent_buffer(node->eb);
390                 node->eb = NULL;
391         }
392 }
393
394 static void drop_backref_node(struct backref_cache *tree,
395                               struct backref_node *node)
396 {
397         BUG_ON(!list_empty(&node->upper));
398
399         drop_node_buffer(node);
400         list_del(&node->list);
401         list_del(&node->lower);
402         if (!RB_EMPTY_NODE(&node->rb_node))
403                 rb_erase(&node->rb_node, &tree->rb_root);
404         free_backref_node(tree, node);
405 }
406
407 /*
408  * remove a backref node from the backref cache
409  */
410 static void remove_backref_node(struct backref_cache *cache,
411                                 struct backref_node *node)
412 {
413         struct backref_node *upper;
414         struct backref_edge *edge;
415
416         if (!node)
417                 return;
418
419         BUG_ON(!node->lowest && !node->detached);
420         while (!list_empty(&node->upper)) {
421                 edge = list_entry(node->upper.next, struct backref_edge,
422                                   list[LOWER]);
423                 upper = edge->node[UPPER];
424                 list_del(&edge->list[LOWER]);
425                 list_del(&edge->list[UPPER]);
426                 free_backref_edge(cache, edge);
427
428                 if (RB_EMPTY_NODE(&upper->rb_node)) {
429                         BUG_ON(!list_empty(&node->upper));
430                         drop_backref_node(cache, node);
431                         node = upper;
432                         node->lowest = 1;
433                         continue;
434                 }
435                 /*
436                  * add the node to leaf node list if no other
437                  * child block cached.
438                  */
439                 if (list_empty(&upper->lower)) {
440                         list_add_tail(&upper->lower, &cache->leaves);
441                         upper->lowest = 1;
442                 }
443         }
444
445         drop_backref_node(cache, node);
446 }
447
448 static void update_backref_node(struct backref_cache *cache,
449                                 struct backref_node *node, u64 bytenr)
450 {
451         struct rb_node *rb_node;
452         rb_erase(&node->rb_node, &cache->rb_root);
453         node->bytenr = bytenr;
454         rb_node = tree_insert(&cache->rb_root, node->bytenr, &node->rb_node);
455         BUG_ON(rb_node);
456 }
457
458 /*
459  * update backref cache after a transaction commit
460  */
461 static int update_backref_cache(struct btrfs_trans_handle *trans,
462                                 struct backref_cache *cache)
463 {
464         struct backref_node *node;
465         int level = 0;
466
467         if (cache->last_trans == 0) {
468                 cache->last_trans = trans->transid;
469                 return 0;
470         }
471
472         if (cache->last_trans == trans->transid)
473                 return 0;
474
475         /*
476          * detached nodes are used to avoid unnecessary backref
477          * lookup. transaction commit changes the extent tree.
478          * so the detached nodes are no longer useful.
479          */
480         while (!list_empty(&cache->detached)) {
481                 node = list_entry(cache->detached.next,
482                                   struct backref_node, list);
483                 remove_backref_node(cache, node);
484         }
485
486         while (!list_empty(&cache->changed)) {
487                 node = list_entry(cache->changed.next,
488                                   struct backref_node, list);
489                 list_del_init(&node->list);
490                 BUG_ON(node->pending);
491                 update_backref_node(cache, node, node->new_bytenr);
492         }
493
494         /*
495          * some nodes can be left in the pending list if there were
496          * errors during processing the pending nodes.
497          */
498         for (level = 0; level < BTRFS_MAX_LEVEL; level++) {
499                 list_for_each_entry(node, &cache->pending[level], list) {
500                         BUG_ON(!node->pending);
501                         if (node->bytenr == node->new_bytenr)
502                                 continue;
503                         update_backref_node(cache, node, node->new_bytenr);
504                 }
505         }
506
507         cache->last_trans = 0;
508         return 1;
509 }
510
511
512 static int should_ignore_root(struct btrfs_root *root)
513 {
514         struct btrfs_root *reloc_root;
515
516         if (!root->ref_cows)
517                 return 0;
518
519         reloc_root = root->reloc_root;
520         if (!reloc_root)
521                 return 0;
522
523         if (btrfs_root_last_snapshot(&reloc_root->root_item) ==
524             root->fs_info->running_transaction->transid - 1)
525                 return 0;
526         /*
527          * if there is reloc tree and it was created in previous
528          * transaction backref lookup can find the reloc tree,
529          * so backref node for the fs tree root is useless for
530          * relocation.
531          */
532         return 1;
533 }
534 /*
535  * find reloc tree by address of tree root
536  */
537 static struct btrfs_root *find_reloc_root(struct reloc_control *rc,
538                                           u64 bytenr)
539 {
540         struct rb_node *rb_node;
541         struct mapping_node *node;
542         struct btrfs_root *root = NULL;
543
544         spin_lock(&rc->reloc_root_tree.lock);
545         rb_node = tree_search(&rc->reloc_root_tree.rb_root, bytenr);
546         if (rb_node) {
547                 node = rb_entry(rb_node, struct mapping_node, rb_node);
548                 root = (struct btrfs_root *)node->data;
549         }
550         spin_unlock(&rc->reloc_root_tree.lock);
551         return root;
552 }
553
554 static int is_cowonly_root(u64 root_objectid)
555 {
556         if (root_objectid == BTRFS_ROOT_TREE_OBJECTID ||
557             root_objectid == BTRFS_EXTENT_TREE_OBJECTID ||
558             root_objectid == BTRFS_CHUNK_TREE_OBJECTID ||
559             root_objectid == BTRFS_DEV_TREE_OBJECTID ||
560             root_objectid == BTRFS_TREE_LOG_OBJECTID ||
561             root_objectid == BTRFS_CSUM_TREE_OBJECTID)
562                 return 1;
563         return 0;
564 }
565
566 static struct btrfs_root *read_fs_root(struct btrfs_fs_info *fs_info,
567                                         u64 root_objectid)
568 {
569         struct btrfs_key key;
570
571         key.objectid = root_objectid;
572         key.type = BTRFS_ROOT_ITEM_KEY;
573         if (is_cowonly_root(root_objectid))
574                 key.offset = 0;
575         else
576                 key.offset = (u64)-1;
577
578         return btrfs_read_fs_root_no_name(fs_info, &key);
579 }
580
581 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
582 static noinline_for_stack
583 struct btrfs_root *find_tree_root(struct reloc_control *rc,
584                                   struct extent_buffer *leaf,
585                                   struct btrfs_extent_ref_v0 *ref0)
586 {
587         struct btrfs_root *root;
588         u64 root_objectid = btrfs_ref_root_v0(leaf, ref0);
589         u64 generation = btrfs_ref_generation_v0(leaf, ref0);
590
591         BUG_ON(root_objectid == BTRFS_TREE_RELOC_OBJECTID);
592
593         root = read_fs_root(rc->extent_root->fs_info, root_objectid);
594         BUG_ON(IS_ERR(root));
595
596         if (root->ref_cows &&
597             generation != btrfs_root_generation(&root->root_item))
598                 return NULL;
599
600         return root;
601 }
602 #endif
603
604 static noinline_for_stack
605 int find_inline_backref(struct extent_buffer *leaf, int slot,
606                         unsigned long *ptr, unsigned long *end)
607 {
608         struct btrfs_extent_item *ei;
609         struct btrfs_tree_block_info *bi;
610         u32 item_size;
611
612         item_size = btrfs_item_size_nr(leaf, slot);
613 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
614         if (item_size < sizeof(*ei)) {
615                 WARN_ON(item_size != sizeof(struct btrfs_extent_item_v0));
616                 return 1;
617         }
618 #endif
619         ei = btrfs_item_ptr(leaf, slot, struct btrfs_extent_item);
620         WARN_ON(!(btrfs_extent_flags(leaf, ei) &
621                   BTRFS_EXTENT_FLAG_TREE_BLOCK));
622
623         if (item_size <= sizeof(*ei) + sizeof(*bi)) {
624                 WARN_ON(item_size < sizeof(*ei) + sizeof(*bi));
625                 return 1;
626         }
627
628         bi = (struct btrfs_tree_block_info *)(ei + 1);
629         *ptr = (unsigned long)(bi + 1);
630         *end = (unsigned long)ei + item_size;
631         return 0;
632 }
633
634 /*
635  * build backref tree for a given tree block. root of the backref tree
636  * corresponds the tree block, leaves of the backref tree correspond
637  * roots of b-trees that reference the tree block.
638  *
639  * the basic idea of this function is check backrefs of a given block
640  * to find upper level blocks that refernece the block, and then check
641  * bakcrefs of these upper level blocks recursively. the recursion stop
642  * when tree root is reached or backrefs for the block is cached.
643  *
644  * NOTE: if we find backrefs for a block are cached, we know backrefs
645  * for all upper level blocks that directly/indirectly reference the
646  * block are also cached.
647  */
648 static noinline_for_stack
649 struct backref_node *build_backref_tree(struct reloc_control *rc,
650                                         struct btrfs_key *node_key,
651                                         int level, u64 bytenr)
652 {
653         struct backref_cache *cache = &rc->backref_cache;
654         struct btrfs_path *path1;
655         struct btrfs_path *path2;
656         struct extent_buffer *eb;
657         struct btrfs_root *root;
658         struct backref_node *cur;
659         struct backref_node *upper;
660         struct backref_node *lower;
661         struct backref_node *node = NULL;
662         struct backref_node *exist = NULL;
663         struct backref_edge *edge;
664         struct rb_node *rb_node;
665         struct btrfs_key key;
666         unsigned long end;
667         unsigned long ptr;
668         LIST_HEAD(list);
669         LIST_HEAD(useless);
670         int cowonly;
671         int ret;
672         int err = 0;
673
674         path1 = btrfs_alloc_path();
675         path2 = btrfs_alloc_path();
676         if (!path1 || !path2) {
677                 err = -ENOMEM;
678                 goto out;
679         }
680
681         node = alloc_backref_node(cache);
682         if (!node) {
683                 err = -ENOMEM;
684                 goto out;
685         }
686
687         node->bytenr = bytenr;
688         node->level = level;
689         node->lowest = 1;
690         cur = node;
691 again:
692         end = 0;
693         ptr = 0;
694         key.objectid = cur->bytenr;
695         key.type = BTRFS_EXTENT_ITEM_KEY;
696         key.offset = (u64)-1;
697
698         path1->search_commit_root = 1;
699         path1->skip_locking = 1;
700         ret = btrfs_search_slot(NULL, rc->extent_root, &key, path1,
701                                 0, 0);
702         if (ret < 0) {
703                 err = ret;
704                 goto out;
705         }
706         BUG_ON(!ret || !path1->slots[0]);
707
708         path1->slots[0]--;
709
710         WARN_ON(cur->checked);
711         if (!list_empty(&cur->upper)) {
712                 /*
713                  * the backref was added previously when processing
714                  * backref of type BTRFS_TREE_BLOCK_REF_KEY
715                  */
716                 BUG_ON(!list_is_singular(&cur->upper));
717                 edge = list_entry(cur->upper.next, struct backref_edge,
718                                   list[LOWER]);
719                 BUG_ON(!list_empty(&edge->list[UPPER]));
720                 exist = edge->node[UPPER];
721                 /*
722                  * add the upper level block to pending list if we need
723                  * check its backrefs
724                  */
725                 if (!exist->checked)
726                         list_add_tail(&edge->list[UPPER], &list);
727         } else {
728                 exist = NULL;
729         }
730
731         while (1) {
732                 cond_resched();
733                 eb = path1->nodes[0];
734
735                 if (ptr >= end) {
736                         if (path1->slots[0] >= btrfs_header_nritems(eb)) {
737                                 ret = btrfs_next_leaf(rc->extent_root, path1);
738                                 if (ret < 0) {
739                                         err = ret;
740                                         goto out;
741                                 }
742                                 if (ret > 0)
743                                         break;
744                                 eb = path1->nodes[0];
745                         }
746
747                         btrfs_item_key_to_cpu(eb, &key, path1->slots[0]);
748                         if (key.objectid != cur->bytenr) {
749                                 WARN_ON(exist);
750                                 break;
751                         }
752
753                         if (key.type == BTRFS_EXTENT_ITEM_KEY) {
754                                 ret = find_inline_backref(eb, path1->slots[0],
755                                                           &ptr, &end);
756                                 if (ret)
757                                         goto next;
758                         }
759                 }
760
761                 if (ptr < end) {
762                         /* update key for inline back ref */
763                         struct btrfs_extent_inline_ref *iref;
764                         iref = (struct btrfs_extent_inline_ref *)ptr;
765                         key.type = btrfs_extent_inline_ref_type(eb, iref);
766                         key.offset = btrfs_extent_inline_ref_offset(eb, iref);
767                         WARN_ON(key.type != BTRFS_TREE_BLOCK_REF_KEY &&
768                                 key.type != BTRFS_SHARED_BLOCK_REF_KEY);
769                 }
770
771                 if (exist &&
772                     ((key.type == BTRFS_TREE_BLOCK_REF_KEY &&
773                       exist->owner == key.offset) ||
774                      (key.type == BTRFS_SHARED_BLOCK_REF_KEY &&
775                       exist->bytenr == key.offset))) {
776                         exist = NULL;
777                         goto next;
778                 }
779
780 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
781                 if (key.type == BTRFS_SHARED_BLOCK_REF_KEY ||
782                     key.type == BTRFS_EXTENT_REF_V0_KEY) {
783                         if (key.type == BTRFS_EXTENT_REF_V0_KEY) {
784                                 struct btrfs_extent_ref_v0 *ref0;
785                                 ref0 = btrfs_item_ptr(eb, path1->slots[0],
786                                                 struct btrfs_extent_ref_v0);
787                                 if (key.objectid == key.offset) {
788                                         root = find_tree_root(rc, eb, ref0);
789                                         if (root && !should_ignore_root(root))
790                                                 cur->root = root;
791                                         else
792                                                 list_add(&cur->list, &useless);
793                                         break;
794                                 }
795                                 if (is_cowonly_root(btrfs_ref_root_v0(eb,
796                                                                       ref0)))
797                                         cur->cowonly = 1;
798                         }
799 #else
800                 BUG_ON(key.type == BTRFS_EXTENT_REF_V0_KEY);
801                 if (key.type == BTRFS_SHARED_BLOCK_REF_KEY) {
802 #endif
803                         if (key.objectid == key.offset) {
804                                 /*
805                                  * only root blocks of reloc trees use
806                                  * backref of this type.
807                                  */
808                                 root = find_reloc_root(rc, cur->bytenr);
809                                 BUG_ON(!root);
810                                 cur->root = root;
811                                 break;
812                         }
813
814                         edge = alloc_backref_edge(cache);
815                         if (!edge) {
816                                 err = -ENOMEM;
817                                 goto out;
818                         }
819                         rb_node = tree_search(&cache->rb_root, key.offset);
820                         if (!rb_node) {
821                                 upper = alloc_backref_node(cache);
822                                 if (!upper) {
823                                         free_backref_edge(cache, edge);
824                                         err = -ENOMEM;
825                                         goto out;
826                                 }
827                                 upper->bytenr = key.offset;
828                                 upper->level = cur->level + 1;
829                                 /*
830                                  *  backrefs for the upper level block isn't
831                                  *  cached, add the block to pending list
832                                  */
833                                 list_add_tail(&edge->list[UPPER], &list);
834                         } else {
835                                 upper = rb_entry(rb_node, struct backref_node,
836                                                  rb_node);
837                                 BUG_ON(!upper->checked);
838                                 INIT_LIST_HEAD(&edge->list[UPPER]);
839                         }
840                         list_add_tail(&edge->list[LOWER], &cur->upper);
841                         edge->node[LOWER] = cur;
842                         edge->node[UPPER] = upper;
843
844                         goto next;
845                 } else if (key.type != BTRFS_TREE_BLOCK_REF_KEY) {
846                         goto next;
847                 }
848
849                 /* key.type == BTRFS_TREE_BLOCK_REF_KEY */
850                 root = read_fs_root(rc->extent_root->fs_info, key.offset);
851                 if (IS_ERR(root)) {
852                         err = PTR_ERR(root);
853                         goto out;
854                 }
855
856                 if (!root->ref_cows)
857                         cur->cowonly = 1;
858
859                 if (btrfs_root_level(&root->root_item) == cur->level) {
860                         /* tree root */
861                         BUG_ON(btrfs_root_bytenr(&root->root_item) !=
862                                cur->bytenr);
863                         if (should_ignore_root(root))
864                                 list_add(&cur->list, &useless);
865                         else
866                                 cur->root = root;
867                         break;
868                 }
869
870                 level = cur->level + 1;
871
872                 /*
873                  * searching the tree to find upper level blocks
874                  * reference the block.
875                  */
876                 path2->search_commit_root = 1;
877                 path2->skip_locking = 1;
878                 path2->lowest_level = level;
879                 ret = btrfs_search_slot(NULL, root, node_key, path2, 0, 0);
880                 path2->lowest_level = 0;
881                 if (ret < 0) {
882                         err = ret;
883                         goto out;
884                 }
885                 if (ret > 0 && path2->slots[level] > 0)
886                         path2->slots[level]--;
887
888                 eb = path2->nodes[level];
889                 WARN_ON(btrfs_node_blockptr(eb, path2->slots[level]) !=
890                         cur->bytenr);
891
892                 lower = cur;
893                 for (; level < BTRFS_MAX_LEVEL; level++) {
894                         if (!path2->nodes[level]) {
895                                 BUG_ON(btrfs_root_bytenr(&root->root_item) !=
896                                        lower->bytenr);
897                                 if (should_ignore_root(root))
898                                         list_add(&lower->list, &useless);
899                                 else
900                                         lower->root = root;
901                                 break;
902                         }
903
904                         edge = alloc_backref_edge(cache);
905                         if (!edge) {
906                                 err = -ENOMEM;
907                                 goto out;
908                         }
909
910                         eb = path2->nodes[level];
911                         rb_node = tree_search(&cache->rb_root, eb->start);
912                         if (!rb_node) {
913                                 upper = alloc_backref_node(cache);
914                                 if (!upper) {
915                                         free_backref_edge(cache, edge);
916                                         err = -ENOMEM;
917                                         goto out;
918                                 }
919                                 upper->bytenr = eb->start;
920                                 upper->owner = btrfs_header_owner(eb);
921                                 upper->level = lower->level + 1;
922                                 if (!root->ref_cows)
923                                         upper->cowonly = 1;
924
925                                 /*
926                                  * if we know the block isn't shared
927                                  * we can void checking its backrefs.
928                                  */
929                                 if (btrfs_block_can_be_shared(root, eb))
930                                         upper->checked = 0;
931                                 else
932                                         upper->checked = 1;
933
934                                 /*
935                                  * add the block to pending list if we
936                                  * need check its backrefs. only block
937                                  * at 'cur->level + 1' is added to the
938                                  * tail of pending list. this guarantees
939                                  * we check backrefs from lower level
940                                  * blocks to upper level blocks.
941                                  */
942                                 if (!upper->checked &&
943                                     level == cur->level + 1) {
944                                         list_add_tail(&edge->list[UPPER],
945                                                       &list);
946                                 } else
947                                         INIT_LIST_HEAD(&edge->list[UPPER]);
948                         } else {
949                                 upper = rb_entry(rb_node, struct backref_node,
950                                                  rb_node);
951                                 BUG_ON(!upper->checked);
952                                 INIT_LIST_HEAD(&edge->list[UPPER]);
953                                 if (!upper->owner)
954                                         upper->owner = btrfs_header_owner(eb);
955                         }
956                         list_add_tail(&edge->list[LOWER], &lower->upper);
957                         edge->node[LOWER] = lower;
958                         edge->node[UPPER] = upper;
959
960                         if (rb_node)
961                                 break;
962                         lower = upper;
963                         upper = NULL;
964                 }
965                 btrfs_release_path(path2);
966 next:
967                 if (ptr < end) {
968                         ptr += btrfs_extent_inline_ref_size(key.type);
969                         if (ptr >= end) {
970                                 WARN_ON(ptr > end);
971                                 ptr = 0;
972                                 end = 0;
973                         }
974                 }
975                 if (ptr >= end)
976                         path1->slots[0]++;
977         }
978         btrfs_release_path(path1);
979
980         cur->checked = 1;
981         WARN_ON(exist);
982
983         /* the pending list isn't empty, take the first block to process */
984         if (!list_empty(&list)) {
985                 edge = list_entry(list.next, struct backref_edge, list[UPPER]);
986                 list_del_init(&edge->list[UPPER]);
987                 cur = edge->node[UPPER];
988                 goto again;
989         }
990
991         /*
992          * everything goes well, connect backref nodes and insert backref nodes
993          * into the cache.
994          */
995         BUG_ON(!node->checked);
996         cowonly = node->cowonly;
997         if (!cowonly) {
998                 rb_node = tree_insert(&cache->rb_root, node->bytenr,
999                                       &node->rb_node);
1000                 BUG_ON(rb_node);
1001                 list_add_tail(&node->lower, &cache->leaves);
1002         }
1003
1004         list_for_each_entry(edge, &node->upper, list[LOWER])
1005                 list_add_tail(&edge->list[UPPER], &list);
1006
1007         while (!list_empty(&list)) {
1008                 edge = list_entry(list.next, struct backref_edge, list[UPPER]);
1009                 list_del_init(&edge->list[UPPER]);
1010                 upper = edge->node[UPPER];
1011                 if (upper->detached) {
1012                         list_del(&edge->list[LOWER]);
1013                         lower = edge->node[LOWER];
1014                         free_backref_edge(cache, edge);
1015                         if (list_empty(&lower->upper))
1016                                 list_add(&lower->list, &useless);
1017                         continue;
1018                 }
1019
1020                 if (!RB_EMPTY_NODE(&upper->rb_node)) {
1021                         if (upper->lowest) {
1022                                 list_del_init(&upper->lower);
1023                                 upper->lowest = 0;
1024                         }
1025
1026                         list_add_tail(&edge->list[UPPER], &upper->lower);
1027                         continue;
1028                 }
1029
1030                 BUG_ON(!upper->checked);
1031                 BUG_ON(cowonly != upper->cowonly);
1032                 if (!cowonly) {
1033                         rb_node = tree_insert(&cache->rb_root, upper->bytenr,
1034                                               &upper->rb_node);
1035                         BUG_ON(rb_node);
1036                 }
1037
1038                 list_add_tail(&edge->list[UPPER], &upper->lower);
1039
1040                 list_for_each_entry(edge, &upper->upper, list[LOWER])
1041                         list_add_tail(&edge->list[UPPER], &list);
1042         }
1043         /*
1044          * process useless backref nodes. backref nodes for tree leaves
1045          * are deleted from the cache. backref nodes for upper level
1046          * tree blocks are left in the cache to avoid unnecessary backref
1047          * lookup.
1048          */
1049         while (!list_empty(&useless)) {
1050                 upper = list_entry(useless.next, struct backref_node, list);
1051                 list_del_init(&upper->list);
1052                 BUG_ON(!list_empty(&upper->upper));
1053                 if (upper == node)
1054                         node = NULL;
1055                 if (upper->lowest) {
1056                         list_del_init(&upper->lower);
1057                         upper->lowest = 0;
1058                 }
1059                 while (!list_empty(&upper->lower)) {
1060                         edge = list_entry(upper->lower.next,
1061                                           struct backref_edge, list[UPPER]);
1062                         list_del(&edge->list[UPPER]);
1063                         list_del(&edge->list[LOWER]);
1064                         lower = edge->node[LOWER];
1065                         free_backref_edge(cache, edge);
1066
1067                         if (list_empty(&lower->upper))
1068                                 list_add(&lower->list, &useless);
1069                 }
1070                 __mark_block_processed(rc, upper);
1071                 if (upper->level > 0) {
1072                         list_add(&upper->list, &cache->detached);
1073                         upper->detached = 1;
1074                 } else {
1075                         rb_erase(&upper->rb_node, &cache->rb_root);
1076                         free_backref_node(cache, upper);
1077                 }
1078         }
1079 out:
1080         btrfs_free_path(path1);
1081         btrfs_free_path(path2);
1082         if (err) {
1083                 while (!list_empty(&useless)) {
1084                         lower = list_entry(useless.next,
1085                                            struct backref_node, upper);
1086                         list_del_init(&lower->upper);
1087                 }
1088                 upper = node;
1089                 INIT_LIST_HEAD(&list);
1090                 while (upper) {
1091                         if (RB_EMPTY_NODE(&upper->rb_node)) {
1092                                 list_splice_tail(&upper->upper, &list);
1093                                 free_backref_node(cache, upper);
1094                         }
1095
1096                         if (list_empty(&list))
1097                                 break;
1098
1099                         edge = list_entry(list.next, struct backref_edge,
1100                                           list[LOWER]);
1101                         list_del(&edge->list[LOWER]);
1102                         upper = edge->node[UPPER];
1103                         free_backref_edge(cache, edge);
1104                 }
1105                 return ERR_PTR(err);
1106         }
1107         BUG_ON(node && node->detached);
1108         return node;
1109 }
1110
1111 /*
1112  * helper to add backref node for the newly created snapshot.
1113  * the backref node is created by cloning backref node that
1114  * corresponds to root of source tree
1115  */
1116 static int clone_backref_node(struct btrfs_trans_handle *trans,
1117                               struct reloc_control *rc,
1118                               struct btrfs_root *src,
1119                               struct btrfs_root *dest)
1120 {
1121         struct btrfs_root *reloc_root = src->reloc_root;
1122         struct backref_cache *cache = &rc->backref_cache;
1123         struct backref_node *node = NULL;
1124         struct backref_node *new_node;
1125         struct backref_edge *edge;
1126         struct backref_edge *new_edge;
1127         struct rb_node *rb_node;
1128
1129         if (cache->last_trans > 0)
1130                 update_backref_cache(trans, cache);
1131
1132         rb_node = tree_search(&cache->rb_root, src->commit_root->start);
1133         if (rb_node) {
1134                 node = rb_entry(rb_node, struct backref_node, rb_node);
1135                 if (node->detached)
1136                         node = NULL;
1137                 else
1138                         BUG_ON(node->new_bytenr != reloc_root->node->start);
1139         }
1140
1141         if (!node) {
1142                 rb_node = tree_search(&cache->rb_root,
1143                                       reloc_root->commit_root->start);
1144                 if (rb_node) {
1145                         node = rb_entry(rb_node, struct backref_node,
1146                                         rb_node);
1147                         BUG_ON(node->detached);
1148                 }
1149         }
1150
1151         if (!node)
1152                 return 0;
1153
1154         new_node = alloc_backref_node(cache);
1155         if (!new_node)
1156                 return -ENOMEM;
1157
1158         new_node->bytenr = dest->node->start;
1159         new_node->level = node->level;
1160         new_node->lowest = node->lowest;
1161         new_node->checked = 1;
1162         new_node->root = dest;
1163
1164         if (!node->lowest) {
1165                 list_for_each_entry(edge, &node->lower, list[UPPER]) {
1166                         new_edge = alloc_backref_edge(cache);
1167                         if (!new_edge)
1168                                 goto fail;
1169
1170                         new_edge->node[UPPER] = new_node;
1171                         new_edge->node[LOWER] = edge->node[LOWER];
1172                         list_add_tail(&new_edge->list[UPPER],
1173                                       &new_node->lower);
1174                 }
1175         }
1176
1177         rb_node = tree_insert(&cache->rb_root, new_node->bytenr,
1178                               &new_node->rb_node);
1179         BUG_ON(rb_node);
1180
1181         if (!new_node->lowest) {
1182                 list_for_each_entry(new_edge, &new_node->lower, list[UPPER]) {
1183                         list_add_tail(&new_edge->list[LOWER],
1184                                       &new_edge->node[LOWER]->upper);
1185                 }
1186         }
1187         return 0;
1188 fail:
1189         while (!list_empty(&new_node->lower)) {
1190                 new_edge = list_entry(new_node->lower.next,
1191                                       struct backref_edge, list[UPPER]);
1192                 list_del(&new_edge->list[UPPER]);
1193                 free_backref_edge(cache, new_edge);
1194         }
1195         free_backref_node(cache, new_node);
1196         return -ENOMEM;
1197 }
1198
1199 /*
1200  * helper to add 'address of tree root -> reloc tree' mapping
1201  */
1202 static int __add_reloc_root(struct btrfs_root *root)
1203 {
1204         struct rb_node *rb_node;
1205         struct mapping_node *node;
1206         struct reloc_control *rc = root->fs_info->reloc_ctl;
1207
1208         node = kmalloc(sizeof(*node), GFP_NOFS);
1209         BUG_ON(!node);
1210
1211         node->bytenr = root->node->start;
1212         node->data = root;
1213
1214         spin_lock(&rc->reloc_root_tree.lock);
1215         rb_node = tree_insert(&rc->reloc_root_tree.rb_root,
1216                               node->bytenr, &node->rb_node);
1217         spin_unlock(&rc->reloc_root_tree.lock);
1218         BUG_ON(rb_node);
1219
1220         list_add_tail(&root->root_list, &rc->reloc_roots);
1221         return 0;
1222 }
1223
1224 /*
1225  * helper to update/delete the 'address of tree root -> reloc tree'
1226  * mapping
1227  */
1228 static int __update_reloc_root(struct btrfs_root *root, int del)
1229 {
1230         struct rb_node *rb_node;
1231         struct mapping_node *node = NULL;
1232         struct reloc_control *rc = root->fs_info->reloc_ctl;
1233
1234         spin_lock(&rc->reloc_root_tree.lock);
1235         rb_node = tree_search(&rc->reloc_root_tree.rb_root,
1236                               root->commit_root->start);
1237         if (rb_node) {
1238                 node = rb_entry(rb_node, struct mapping_node, rb_node);
1239                 rb_erase(&node->rb_node, &rc->reloc_root_tree.rb_root);
1240         }
1241         spin_unlock(&rc->reloc_root_tree.lock);
1242
1243         BUG_ON((struct btrfs_root *)node->data != root);
1244
1245         if (!del) {
1246                 spin_lock(&rc->reloc_root_tree.lock);
1247                 node->bytenr = root->node->start;
1248                 rb_node = tree_insert(&rc->reloc_root_tree.rb_root,
1249                                       node->bytenr, &node->rb_node);
1250                 spin_unlock(&rc->reloc_root_tree.lock);
1251                 BUG_ON(rb_node);
1252         } else {
1253                 list_del_init(&root->root_list);
1254                 kfree(node);
1255         }
1256         return 0;
1257 }
1258
1259 static struct btrfs_root *create_reloc_root(struct btrfs_trans_handle *trans,
1260                                         struct btrfs_root *root, u64 objectid)
1261 {
1262         struct btrfs_root *reloc_root;
1263         struct extent_buffer *eb;
1264         struct btrfs_root_item *root_item;
1265         struct btrfs_key root_key;
1266         int ret;
1267
1268         root_item = kmalloc(sizeof(*root_item), GFP_NOFS);
1269         BUG_ON(!root_item);
1270
1271         root_key.objectid = BTRFS_TREE_RELOC_OBJECTID;
1272         root_key.type = BTRFS_ROOT_ITEM_KEY;
1273         root_key.offset = objectid;
1274
1275         if (root->root_key.objectid == objectid) {
1276                 /* called by btrfs_init_reloc_root */
1277                 ret = btrfs_copy_root(trans, root, root->commit_root, &eb,
1278                                       BTRFS_TREE_RELOC_OBJECTID);
1279                 BUG_ON(ret);
1280
1281                 btrfs_set_root_last_snapshot(&root->root_item,
1282                                              trans->transid - 1);
1283         } else {
1284                 /*
1285                  * called by btrfs_reloc_post_snapshot_hook.
1286                  * the source tree is a reloc tree, all tree blocks
1287                  * modified after it was created have RELOC flag
1288                  * set in their headers. so it's OK to not update
1289                  * the 'last_snapshot'.
1290                  */
1291                 ret = btrfs_copy_root(trans, root, root->node, &eb,
1292                                       BTRFS_TREE_RELOC_OBJECTID);
1293                 BUG_ON(ret);
1294         }
1295
1296         memcpy(root_item, &root->root_item, sizeof(*root_item));
1297         btrfs_set_root_bytenr(root_item, eb->start);
1298         btrfs_set_root_level(root_item, btrfs_header_level(eb));
1299         btrfs_set_root_generation(root_item, trans->transid);
1300
1301         if (root->root_key.objectid == objectid) {
1302                 btrfs_set_root_refs(root_item, 0);
1303                 memset(&root_item->drop_progress, 0,
1304                        sizeof(struct btrfs_disk_key));
1305                 root_item->drop_level = 0;
1306         }
1307
1308         btrfs_tree_unlock(eb);
1309         free_extent_buffer(eb);
1310
1311         ret = btrfs_insert_root(trans, root->fs_info->tree_root,
1312                                 &root_key, root_item);
1313         BUG_ON(ret);
1314         kfree(root_item);
1315
1316         reloc_root = btrfs_read_fs_root_no_radix(root->fs_info->tree_root,
1317                                                  &root_key);
1318         BUG_ON(IS_ERR(reloc_root));
1319         reloc_root->last_trans = trans->transid;
1320         return reloc_root;
1321 }
1322
1323 /*
1324  * create reloc tree for a given fs tree. reloc tree is just a
1325  * snapshot of the fs tree with special root objectid.
1326  */
1327 int btrfs_init_reloc_root(struct btrfs_trans_handle *trans,
1328                           struct btrfs_root *root)
1329 {
1330         struct btrfs_root *reloc_root;
1331         struct reloc_control *rc = root->fs_info->reloc_ctl;
1332         int clear_rsv = 0;
1333
1334         if (root->reloc_root) {
1335                 reloc_root = root->reloc_root;
1336                 reloc_root->last_trans = trans->transid;
1337                 return 0;
1338         }
1339
1340         if (!rc || !rc->create_reloc_tree ||
1341             root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID)
1342                 return 0;
1343
1344         if (!trans->block_rsv) {
1345                 trans->block_rsv = rc->block_rsv;
1346                 clear_rsv = 1;
1347         }
1348         reloc_root = create_reloc_root(trans, root, root->root_key.objectid);
1349         if (clear_rsv)
1350                 trans->block_rsv = NULL;
1351
1352         __add_reloc_root(reloc_root);
1353         root->reloc_root = reloc_root;
1354         return 0;
1355 }
1356
1357 /*
1358  * update root item of reloc tree
1359  */
1360 int btrfs_update_reloc_root(struct btrfs_trans_handle *trans,
1361                             struct btrfs_root *root)
1362 {
1363         struct btrfs_root *reloc_root;
1364         struct btrfs_root_item *root_item;
1365         int del = 0;
1366         int ret;
1367
1368         if (!root->reloc_root)
1369                 return 0;
1370
1371         reloc_root = root->reloc_root;
1372         root_item = &reloc_root->root_item;
1373
1374         if (root->fs_info->reloc_ctl->merge_reloc_tree &&
1375             btrfs_root_refs(root_item) == 0) {
1376                 root->reloc_root = NULL;
1377                 del = 1;
1378         }
1379
1380         __update_reloc_root(reloc_root, del);
1381
1382         if (reloc_root->commit_root != reloc_root->node) {
1383                 btrfs_set_root_node(root_item, reloc_root->node);
1384                 free_extent_buffer(reloc_root->commit_root);
1385                 reloc_root->commit_root = btrfs_root_node(reloc_root);
1386         }
1387
1388         ret = btrfs_update_root(trans, root->fs_info->tree_root,
1389                                 &reloc_root->root_key, root_item);
1390         BUG_ON(ret);
1391         return 0;
1392 }
1393
1394 /*
1395  * helper to find first cached inode with inode number >= objectid
1396  * in a subvolume
1397  */
1398 static struct inode *find_next_inode(struct btrfs_root *root, u64 objectid)
1399 {
1400         struct rb_node *node;
1401         struct rb_node *prev;
1402         struct btrfs_inode *entry;
1403         struct inode *inode;
1404
1405         spin_lock(&root->inode_lock);
1406 again:
1407         node = root->inode_tree.rb_node;
1408         prev = NULL;
1409         while (node) {
1410                 prev = node;
1411                 entry = rb_entry(node, struct btrfs_inode, rb_node);
1412
1413                 if (objectid < btrfs_ino(&entry->vfs_inode))
1414                         node = node->rb_left;
1415                 else if (objectid > btrfs_ino(&entry->vfs_inode))
1416                         node = node->rb_right;
1417                 else
1418                         break;
1419         }
1420         if (!node) {
1421                 while (prev) {
1422                         entry = rb_entry(prev, struct btrfs_inode, rb_node);
1423                         if (objectid <= btrfs_ino(&entry->vfs_inode)) {
1424                                 node = prev;
1425                                 break;
1426                         }
1427                         prev = rb_next(prev);
1428                 }
1429         }
1430         while (node) {
1431                 entry = rb_entry(node, struct btrfs_inode, rb_node);
1432                 inode = igrab(&entry->vfs_inode);
1433                 if (inode) {
1434                         spin_unlock(&root->inode_lock);
1435                         return inode;
1436                 }
1437
1438                 objectid = btrfs_ino(&entry->vfs_inode) + 1;
1439                 if (cond_resched_lock(&root->inode_lock))
1440                         goto again;
1441
1442                 node = rb_next(node);
1443         }
1444         spin_unlock(&root->inode_lock);
1445         return NULL;
1446 }
1447
1448 static int in_block_group(u64 bytenr,
1449                           struct btrfs_block_group_cache *block_group)
1450 {
1451         if (bytenr >= block_group->key.objectid &&
1452             bytenr < block_group->key.objectid + block_group->key.offset)
1453                 return 1;
1454         return 0;
1455 }
1456
1457 /*
1458  * get new location of data
1459  */
1460 static int get_new_location(struct inode *reloc_inode, u64 *new_bytenr,
1461                             u64 bytenr, u64 num_bytes)
1462 {
1463         struct btrfs_root *root = BTRFS_I(reloc_inode)->root;
1464         struct btrfs_path *path;
1465         struct btrfs_file_extent_item *fi;
1466         struct extent_buffer *leaf;
1467         int ret;
1468
1469         path = btrfs_alloc_path();
1470         if (!path)
1471                 return -ENOMEM;
1472
1473         bytenr -= BTRFS_I(reloc_inode)->index_cnt;
1474         ret = btrfs_lookup_file_extent(NULL, root, path, btrfs_ino(reloc_inode),
1475                                        bytenr, 0);
1476         if (ret < 0)
1477                 goto out;
1478         if (ret > 0) {
1479                 ret = -ENOENT;
1480                 goto out;
1481         }
1482
1483         leaf = path->nodes[0];
1484         fi = btrfs_item_ptr(leaf, path->slots[0],
1485                             struct btrfs_file_extent_item);
1486
1487         BUG_ON(btrfs_file_extent_offset(leaf, fi) ||
1488                btrfs_file_extent_compression(leaf, fi) ||
1489                btrfs_file_extent_encryption(leaf, fi) ||
1490                btrfs_file_extent_other_encoding(leaf, fi));
1491
1492         if (num_bytes != btrfs_file_extent_disk_num_bytes(leaf, fi)) {
1493                 ret = 1;
1494                 goto out;
1495         }
1496
1497         *new_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
1498         ret = 0;
1499 out:
1500         btrfs_free_path(path);
1501         return ret;
1502 }
1503
1504 /*
1505  * update file extent items in the tree leaf to point to
1506  * the new locations.
1507  */
1508 static noinline_for_stack
1509 int replace_file_extents(struct btrfs_trans_handle *trans,
1510                          struct reloc_control *rc,
1511                          struct btrfs_root *root,
1512                          struct extent_buffer *leaf)
1513 {
1514         struct btrfs_key key;
1515         struct btrfs_file_extent_item *fi;
1516         struct inode *inode = NULL;
1517         u64 parent;
1518         u64 bytenr;
1519         u64 new_bytenr = 0;
1520         u64 num_bytes;
1521         u64 end;
1522         u32 nritems;
1523         u32 i;
1524         int ret;
1525         int first = 1;
1526         int dirty = 0;
1527
1528         if (rc->stage != UPDATE_DATA_PTRS)
1529                 return 0;
1530
1531         /* reloc trees always use full backref */
1532         if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID)
1533                 parent = leaf->start;
1534         else
1535                 parent = 0;
1536
1537         nritems = btrfs_header_nritems(leaf);
1538         for (i = 0; i < nritems; i++) {
1539                 cond_resched();
1540                 btrfs_item_key_to_cpu(leaf, &key, i);
1541                 if (key.type != BTRFS_EXTENT_DATA_KEY)
1542                         continue;
1543                 fi = btrfs_item_ptr(leaf, i, struct btrfs_file_extent_item);
1544                 if (btrfs_file_extent_type(leaf, fi) ==
1545                     BTRFS_FILE_EXTENT_INLINE)
1546                         continue;
1547                 bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
1548                 num_bytes = btrfs_file_extent_disk_num_bytes(leaf, fi);
1549                 if (bytenr == 0)
1550                         continue;
1551                 if (!in_block_group(bytenr, rc->block_group))
1552                         continue;
1553
1554                 /*
1555                  * if we are modifying block in fs tree, wait for readpage
1556                  * to complete and drop the extent cache
1557                  */
1558                 if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) {
1559                         if (first) {
1560                                 inode = find_next_inode(root, key.objectid);
1561                                 first = 0;
1562                         } else if (inode && btrfs_ino(inode) < key.objectid) {
1563                                 btrfs_add_delayed_iput(inode);
1564                                 inode = find_next_inode(root, key.objectid);
1565                         }
1566                         if (inode && btrfs_ino(inode) == key.objectid) {
1567                                 end = key.offset +
1568                                       btrfs_file_extent_num_bytes(leaf, fi);
1569                                 WARN_ON(!IS_ALIGNED(key.offset,
1570                                                     root->sectorsize));
1571                                 WARN_ON(!IS_ALIGNED(end, root->sectorsize));
1572                                 end--;
1573                                 ret = try_lock_extent(&BTRFS_I(inode)->io_tree,
1574                                                       key.offset, end,
1575                                                       GFP_NOFS);
1576                                 if (!ret)
1577                                         continue;
1578
1579                                 btrfs_drop_extent_cache(inode, key.offset, end,
1580                                                         1);
1581                                 unlock_extent(&BTRFS_I(inode)->io_tree,
1582                                               key.offset, end, GFP_NOFS);
1583                         }
1584                 }
1585
1586                 ret = get_new_location(rc->data_inode, &new_bytenr,
1587                                        bytenr, num_bytes);
1588                 if (ret > 0) {
1589                         WARN_ON(1);
1590                         continue;
1591                 }
1592                 BUG_ON(ret < 0);
1593
1594                 btrfs_set_file_extent_disk_bytenr(leaf, fi, new_bytenr);
1595                 dirty = 1;
1596
1597                 key.offset -= btrfs_file_extent_offset(leaf, fi);
1598                 ret = btrfs_inc_extent_ref(trans, root, new_bytenr,
1599                                            num_bytes, parent,
1600                                            btrfs_header_owner(leaf),
1601                                            key.objectid, key.offset);
1602                 BUG_ON(ret);
1603
1604                 ret = btrfs_free_extent(trans, root, bytenr, num_bytes,
1605                                         parent, btrfs_header_owner(leaf),
1606                                         key.objectid, key.offset);
1607                 BUG_ON(ret);
1608         }
1609         if (dirty)
1610                 btrfs_mark_buffer_dirty(leaf);
1611         if (inode)
1612                 btrfs_add_delayed_iput(inode);
1613         return 0;
1614 }
1615
1616 static noinline_for_stack
1617 int memcmp_node_keys(struct extent_buffer *eb, int slot,
1618                      struct btrfs_path *path, int level)
1619 {
1620         struct btrfs_disk_key key1;
1621         struct btrfs_disk_key key2;
1622         btrfs_node_key(eb, &key1, slot);
1623         btrfs_node_key(path->nodes[level], &key2, path->slots[level]);
1624         return memcmp(&key1, &key2, sizeof(key1));
1625 }
1626
1627 /*
1628  * try to replace tree blocks in fs tree with the new blocks
1629  * in reloc tree. tree blocks haven't been modified since the
1630  * reloc tree was create can be replaced.
1631  *
1632  * if a block was replaced, level of the block + 1 is returned.
1633  * if no block got replaced, 0 is returned. if there are other
1634  * errors, a negative error number is returned.
1635  */
1636 static noinline_for_stack
1637 int replace_path(struct btrfs_trans_handle *trans,
1638                  struct btrfs_root *dest, struct btrfs_root *src,
1639                  struct btrfs_path *path, struct btrfs_key *next_key,
1640                  int lowest_level, int max_level)
1641 {
1642         struct extent_buffer *eb;
1643         struct extent_buffer *parent;
1644         struct btrfs_key key;
1645         u64 old_bytenr;
1646         u64 new_bytenr;
1647         u64 old_ptr_gen;
1648         u64 new_ptr_gen;
1649         u64 last_snapshot;
1650         u32 blocksize;
1651         int cow = 0;
1652         int level;
1653         int ret;
1654         int slot;
1655
1656         BUG_ON(src->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID);
1657         BUG_ON(dest->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID);
1658
1659         last_snapshot = btrfs_root_last_snapshot(&src->root_item);
1660 again:
1661         slot = path->slots[lowest_level];
1662         btrfs_node_key_to_cpu(path->nodes[lowest_level], &key, slot);
1663
1664         eb = btrfs_lock_root_node(dest);
1665         btrfs_set_lock_blocking(eb);
1666         level = btrfs_header_level(eb);
1667
1668         if (level < lowest_level) {
1669                 btrfs_tree_unlock(eb);
1670                 free_extent_buffer(eb);
1671                 return 0;
1672         }
1673
1674         if (cow) {
1675                 ret = btrfs_cow_block(trans, dest, eb, NULL, 0, &eb);
1676                 BUG_ON(ret);
1677         }
1678         btrfs_set_lock_blocking(eb);
1679
1680         if (next_key) {
1681                 next_key->objectid = (u64)-1;
1682                 next_key->type = (u8)-1;
1683                 next_key->offset = (u64)-1;
1684         }
1685
1686         parent = eb;
1687         while (1) {
1688                 level = btrfs_header_level(parent);
1689                 BUG_ON(level < lowest_level);
1690
1691                 ret = btrfs_bin_search(parent, &key, level, &slot);
1692                 if (ret && slot > 0)
1693                         slot--;
1694
1695                 if (next_key && slot + 1 < btrfs_header_nritems(parent))
1696                         btrfs_node_key_to_cpu(parent, next_key, slot + 1);
1697
1698                 old_bytenr = btrfs_node_blockptr(parent, slot);
1699                 blocksize = btrfs_level_size(dest, level - 1);
1700                 old_ptr_gen = btrfs_node_ptr_generation(parent, slot);
1701
1702                 if (level <= max_level) {
1703                         eb = path->nodes[level];
1704                         new_bytenr = btrfs_node_blockptr(eb,
1705                                                         path->slots[level]);
1706                         new_ptr_gen = btrfs_node_ptr_generation(eb,
1707                                                         path->slots[level]);
1708                 } else {
1709                         new_bytenr = 0;
1710                         new_ptr_gen = 0;
1711                 }
1712
1713                 if (new_bytenr > 0 && new_bytenr == old_bytenr) {
1714                         WARN_ON(1);
1715                         ret = level;
1716                         break;
1717                 }
1718
1719                 if (new_bytenr == 0 || old_ptr_gen > last_snapshot ||
1720                     memcmp_node_keys(parent, slot, path, level)) {
1721                         if (level <= lowest_level) {
1722                                 ret = 0;
1723                                 break;
1724                         }
1725
1726                         eb = read_tree_block(dest, old_bytenr, blocksize,
1727                                              old_ptr_gen);
1728                         BUG_ON(!eb);
1729                         btrfs_tree_lock(eb);
1730                         if (cow) {
1731                                 ret = btrfs_cow_block(trans, dest, eb, parent,
1732                                                       slot, &eb);
1733                                 BUG_ON(ret);
1734                         }
1735                         btrfs_set_lock_blocking(eb);
1736
1737                         btrfs_tree_unlock(parent);
1738                         free_extent_buffer(parent);
1739
1740                         parent = eb;
1741                         continue;
1742                 }
1743
1744                 if (!cow) {
1745                         btrfs_tree_unlock(parent);
1746                         free_extent_buffer(parent);
1747                         cow = 1;
1748                         goto again;
1749                 }
1750
1751                 btrfs_node_key_to_cpu(path->nodes[level], &key,
1752                                       path->slots[level]);
1753                 btrfs_release_path(path);
1754
1755                 path->lowest_level = level;
1756                 ret = btrfs_search_slot(trans, src, &key, path, 0, 1);
1757                 path->lowest_level = 0;
1758                 BUG_ON(ret);
1759
1760                 /*
1761                  * swap blocks in fs tree and reloc tree.
1762                  */
1763                 btrfs_set_node_blockptr(parent, slot, new_bytenr);
1764                 btrfs_set_node_ptr_generation(parent, slot, new_ptr_gen);
1765                 btrfs_mark_buffer_dirty(parent);
1766
1767                 btrfs_set_node_blockptr(path->nodes[level],
1768                                         path->slots[level], old_bytenr);
1769                 btrfs_set_node_ptr_generation(path->nodes[level],
1770                                               path->slots[level], old_ptr_gen);
1771                 btrfs_mark_buffer_dirty(path->nodes[level]);
1772
1773                 ret = btrfs_inc_extent_ref(trans, src, old_bytenr, blocksize,
1774                                         path->nodes[level]->start,
1775                                         src->root_key.objectid, level - 1, 0);
1776                 BUG_ON(ret);
1777                 ret = btrfs_inc_extent_ref(trans, dest, new_bytenr, blocksize,
1778                                         0, dest->root_key.objectid, level - 1,
1779                                         0);
1780                 BUG_ON(ret);
1781
1782                 ret = btrfs_free_extent(trans, src, new_bytenr, blocksize,
1783                                         path->nodes[level]->start,
1784                                         src->root_key.objectid, level - 1, 0);
1785                 BUG_ON(ret);
1786
1787                 ret = btrfs_free_extent(trans, dest, old_bytenr, blocksize,
1788                                         0, dest->root_key.objectid, level - 1,
1789                                         0);
1790                 BUG_ON(ret);
1791
1792                 btrfs_unlock_up_safe(path, 0);
1793
1794                 ret = level;
1795                 break;
1796         }
1797         btrfs_tree_unlock(parent);
1798         free_extent_buffer(parent);
1799         return ret;
1800 }
1801
1802 /*
1803  * helper to find next relocated block in reloc tree
1804  */
1805 static noinline_for_stack
1806 int walk_up_reloc_tree(struct btrfs_root *root, struct btrfs_path *path,
1807                        int *level)
1808 {
1809         struct extent_buffer *eb;
1810         int i;
1811         u64 last_snapshot;
1812         u32 nritems;
1813
1814         last_snapshot = btrfs_root_last_snapshot(&root->root_item);
1815
1816         for (i = 0; i < *level; i++) {
1817                 free_extent_buffer(path->nodes[i]);
1818                 path->nodes[i] = NULL;
1819         }
1820
1821         for (i = *level; i < BTRFS_MAX_LEVEL && path->nodes[i]; i++) {
1822                 eb = path->nodes[i];
1823                 nritems = btrfs_header_nritems(eb);
1824                 while (path->slots[i] + 1 < nritems) {
1825                         path->slots[i]++;
1826                         if (btrfs_node_ptr_generation(eb, path->slots[i]) <=
1827                             last_snapshot)
1828                                 continue;
1829
1830                         *level = i;
1831                         return 0;
1832                 }
1833                 free_extent_buffer(path->nodes[i]);
1834                 path->nodes[i] = NULL;
1835         }
1836         return 1;
1837 }
1838
1839 /*
1840  * walk down reloc tree to find relocated block of lowest level
1841  */
1842 static noinline_for_stack
1843 int walk_down_reloc_tree(struct btrfs_root *root, struct btrfs_path *path,
1844                          int *level)
1845 {
1846         struct extent_buffer *eb = NULL;
1847         int i;
1848         u64 bytenr;
1849         u64 ptr_gen = 0;
1850         u64 last_snapshot;
1851         u32 blocksize;
1852         u32 nritems;
1853
1854         last_snapshot = btrfs_root_last_snapshot(&root->root_item);
1855
1856         for (i = *level; i > 0; i--) {
1857                 eb = path->nodes[i];
1858                 nritems = btrfs_header_nritems(eb);
1859                 while (path->slots[i] < nritems) {
1860                         ptr_gen = btrfs_node_ptr_generation(eb, path->slots[i]);
1861                         if (ptr_gen > last_snapshot)
1862                                 break;
1863                         path->slots[i]++;
1864                 }
1865                 if (path->slots[i] >= nritems) {
1866                         if (i == *level)
1867                                 break;
1868                         *level = i + 1;
1869                         return 0;
1870                 }
1871                 if (i == 1) {
1872                         *level = i;
1873                         return 0;
1874                 }
1875
1876                 bytenr = btrfs_node_blockptr(eb, path->slots[i]);
1877                 blocksize = btrfs_level_size(root, i - 1);
1878                 eb = read_tree_block(root, bytenr, blocksize, ptr_gen);
1879                 BUG_ON(btrfs_header_level(eb) != i - 1);
1880                 path->nodes[i - 1] = eb;
1881                 path->slots[i - 1] = 0;
1882         }
1883         return 1;
1884 }
1885
1886 /*
1887  * invalidate extent cache for file extents whose key in range of
1888  * [min_key, max_key)
1889  */
1890 static int invalidate_extent_cache(struct btrfs_root *root,
1891                                    struct btrfs_key *min_key,
1892                                    struct btrfs_key *max_key)
1893 {
1894         struct inode *inode = NULL;
1895         u64 objectid;
1896         u64 start, end;
1897         u64 ino;
1898
1899         objectid = min_key->objectid;
1900         while (1) {
1901                 cond_resched();
1902                 iput(inode);
1903
1904                 if (objectid > max_key->objectid)
1905                         break;
1906
1907                 inode = find_next_inode(root, objectid);
1908                 if (!inode)
1909                         break;
1910                 ino = btrfs_ino(inode);
1911
1912                 if (ino > max_key->objectid) {
1913                         iput(inode);
1914                         break;
1915                 }
1916
1917                 objectid = ino + 1;
1918                 if (!S_ISREG(inode->i_mode))
1919                         continue;
1920
1921                 if (unlikely(min_key->objectid == ino)) {
1922                         if (min_key->type > BTRFS_EXTENT_DATA_KEY)
1923                                 continue;
1924                         if (min_key->type < BTRFS_EXTENT_DATA_KEY)
1925                                 start = 0;
1926                         else {
1927                                 start = min_key->offset;
1928                                 WARN_ON(!IS_ALIGNED(start, root->sectorsize));
1929                         }
1930                 } else {
1931                         start = 0;
1932                 }
1933
1934                 if (unlikely(max_key->objectid == ino)) {
1935                         if (max_key->type < BTRFS_EXTENT_DATA_KEY)
1936                                 continue;
1937                         if (max_key->type > BTRFS_EXTENT_DATA_KEY) {
1938                                 end = (u64)-1;
1939                         } else {
1940                                 if (max_key->offset == 0)
1941                                         continue;
1942                                 end = max_key->offset;
1943                                 WARN_ON(!IS_ALIGNED(end, root->sectorsize));
1944                                 end--;
1945                         }
1946                 } else {
1947                         end = (u64)-1;
1948                 }
1949
1950                 /* the lock_extent waits for readpage to complete */
1951                 lock_extent(&BTRFS_I(inode)->io_tree, start, end, GFP_NOFS);
1952                 btrfs_drop_extent_cache(inode, start, end, 1);
1953                 unlock_extent(&BTRFS_I(inode)->io_tree, start, end, GFP_NOFS);
1954         }
1955         return 0;
1956 }
1957
1958 static int find_next_key(struct btrfs_path *path, int level,
1959                          struct btrfs_key *key)
1960
1961 {
1962         while (level < BTRFS_MAX_LEVEL) {
1963                 if (!path->nodes[level])
1964                         break;
1965                 if (path->slots[level] + 1 <
1966                     btrfs_header_nritems(path->nodes[level])) {
1967                         btrfs_node_key_to_cpu(path->nodes[level], key,
1968                                               path->slots[level] + 1);
1969                         return 0;
1970                 }
1971                 level++;
1972         }
1973         return 1;
1974 }
1975
1976 /*
1977  * merge the relocated tree blocks in reloc tree with corresponding
1978  * fs tree.
1979  */
1980 static noinline_for_stack int merge_reloc_root(struct reloc_control *rc,
1981                                                struct btrfs_root *root)
1982 {
1983         LIST_HEAD(inode_list);
1984         struct btrfs_key key;
1985         struct btrfs_key next_key;
1986         struct btrfs_trans_handle *trans;
1987         struct btrfs_root *reloc_root;
1988         struct btrfs_root_item *root_item;
1989         struct btrfs_path *path;
1990         struct extent_buffer *leaf;
1991         unsigned long nr;
1992         int level;
1993         int max_level;
1994         int replaced = 0;
1995         int ret;
1996         int err = 0;
1997         u32 min_reserved;
1998
1999         path = btrfs_alloc_path();
2000         if (!path)
2001                 return -ENOMEM;
2002
2003         reloc_root = root->reloc_root;
2004         root_item = &reloc_root->root_item;
2005
2006         if (btrfs_disk_key_objectid(&root_item->drop_progress) == 0) {
2007                 level = btrfs_root_level(root_item);
2008                 extent_buffer_get(reloc_root->node);
2009                 path->nodes[level] = reloc_root->node;
2010                 path->slots[level] = 0;
2011         } else {
2012                 btrfs_disk_key_to_cpu(&key, &root_item->drop_progress);
2013
2014                 level = root_item->drop_level;
2015                 BUG_ON(level == 0);
2016                 path->lowest_level = level;
2017                 ret = btrfs_search_slot(NULL, reloc_root, &key, path, 0, 0);
2018                 path->lowest_level = 0;
2019                 if (ret < 0) {
2020                         btrfs_free_path(path);
2021                         return ret;
2022                 }
2023
2024                 btrfs_node_key_to_cpu(path->nodes[level], &next_key,
2025                                       path->slots[level]);
2026                 WARN_ON(memcmp(&key, &next_key, sizeof(key)));
2027
2028                 btrfs_unlock_up_safe(path, 0);
2029         }
2030
2031         min_reserved = root->nodesize * (BTRFS_MAX_LEVEL - 1) * 2;
2032         memset(&next_key, 0, sizeof(next_key));
2033
2034         while (1) {
2035                 trans = btrfs_start_transaction(root, 0);
2036                 BUG_ON(IS_ERR(trans));
2037                 trans->block_rsv = rc->block_rsv;
2038
2039                 ret = btrfs_block_rsv_check(trans, root, rc->block_rsv,
2040                                             min_reserved, 0);
2041                 if (ret) {
2042                         BUG_ON(ret != -EAGAIN);
2043                         ret = btrfs_commit_transaction(trans, root);
2044                         BUG_ON(ret);
2045                         continue;
2046                 }
2047
2048                 replaced = 0;
2049                 max_level = level;
2050
2051                 ret = walk_down_reloc_tree(reloc_root, path, &level);
2052                 if (ret < 0) {
2053                         err = ret;
2054                         goto out;
2055                 }
2056                 if (ret > 0)
2057                         break;
2058
2059                 if (!find_next_key(path, level, &key) &&
2060                     btrfs_comp_cpu_keys(&next_key, &key) >= 0) {
2061                         ret = 0;
2062                 } else {
2063                         ret = replace_path(trans, root, reloc_root, path,
2064                                            &next_key, level, max_level);
2065                 }
2066                 if (ret < 0) {
2067                         err = ret;
2068                         goto out;
2069                 }
2070
2071                 if (ret > 0) {
2072                         level = ret;
2073                         btrfs_node_key_to_cpu(path->nodes[level], &key,
2074                                               path->slots[level]);
2075                         replaced = 1;
2076                 }
2077
2078                 ret = walk_up_reloc_tree(reloc_root, path, &level);
2079                 if (ret > 0)
2080                         break;
2081
2082                 BUG_ON(level == 0);
2083                 /*
2084                  * save the merging progress in the drop_progress.
2085                  * this is OK since root refs == 1 in this case.
2086                  */
2087                 btrfs_node_key(path->nodes[level], &root_item->drop_progress,
2088                                path->slots[level]);
2089                 root_item->drop_level = level;
2090
2091                 nr = trans->blocks_used;
2092                 btrfs_end_transaction_throttle(trans, root);
2093
2094                 btrfs_btree_balance_dirty(root, nr);
2095
2096                 if (replaced && rc->stage == UPDATE_DATA_PTRS)
2097                         invalidate_extent_cache(root, &key, &next_key);
2098         }
2099
2100         /*
2101          * handle the case only one block in the fs tree need to be
2102          * relocated and the block is tree root.
2103          */
2104         leaf = btrfs_lock_root_node(root);
2105         ret = btrfs_cow_block(trans, root, leaf, NULL, 0, &leaf);
2106         btrfs_tree_unlock(leaf);
2107         free_extent_buffer(leaf);
2108         if (ret < 0)
2109                 err = ret;
2110 out:
2111         btrfs_free_path(path);
2112
2113         if (err == 0) {
2114                 memset(&root_item->drop_progress, 0,
2115                        sizeof(root_item->drop_progress));
2116                 root_item->drop_level = 0;
2117                 btrfs_set_root_refs(root_item, 0);
2118                 btrfs_update_reloc_root(trans, root);
2119         }
2120
2121         nr = trans->blocks_used;
2122         btrfs_end_transaction_throttle(trans, root);
2123
2124         btrfs_btree_balance_dirty(root, nr);
2125
2126         if (replaced && rc->stage == UPDATE_DATA_PTRS)
2127                 invalidate_extent_cache(root, &key, &next_key);
2128
2129         return err;
2130 }
2131
2132 static noinline_for_stack
2133 int prepare_to_merge(struct reloc_control *rc, int err)
2134 {
2135         struct btrfs_root *root = rc->extent_root;
2136         struct btrfs_root *reloc_root;
2137         struct btrfs_trans_handle *trans;
2138         LIST_HEAD(reloc_roots);
2139         u64 num_bytes = 0;
2140         int ret;
2141
2142         mutex_lock(&root->fs_info->trans_mutex);
2143         rc->merging_rsv_size += root->nodesize * (BTRFS_MAX_LEVEL - 1) * 2;
2144         rc->merging_rsv_size += rc->nodes_relocated * 2;
2145         mutex_unlock(&root->fs_info->trans_mutex);
2146 again:
2147         if (!err) {
2148                 num_bytes = rc->merging_rsv_size;
2149                 ret = btrfs_block_rsv_add(NULL, root, rc->block_rsv,
2150                                           num_bytes);
2151                 if (ret)
2152                         err = ret;
2153         }
2154
2155         trans = btrfs_join_transaction(rc->extent_root, 1);
2156         if (IS_ERR(trans)) {
2157                 if (!err)
2158                         btrfs_block_rsv_release(rc->extent_root,
2159                                                 rc->block_rsv, num_bytes);
2160                 return PTR_ERR(trans);
2161         }
2162
2163         if (!err) {
2164                 if (num_bytes != rc->merging_rsv_size) {
2165                         btrfs_end_transaction(trans, rc->extent_root);
2166                         btrfs_block_rsv_release(rc->extent_root,
2167                                                 rc->block_rsv, num_bytes);
2168                         goto again;
2169                 }
2170         }
2171
2172         rc->merge_reloc_tree = 1;
2173
2174         while (!list_empty(&rc->reloc_roots)) {
2175                 reloc_root = list_entry(rc->reloc_roots.next,
2176                                         struct btrfs_root, root_list);
2177                 list_del_init(&reloc_root->root_list);
2178
2179                 root = read_fs_root(reloc_root->fs_info,
2180                                     reloc_root->root_key.offset);
2181                 BUG_ON(IS_ERR(root));
2182                 BUG_ON(root->reloc_root != reloc_root);
2183
2184                 /*
2185                  * set reference count to 1, so btrfs_recover_relocation
2186                  * knows it should resumes merging
2187                  */
2188                 if (!err)
2189                         btrfs_set_root_refs(&reloc_root->root_item, 1);
2190                 btrfs_update_reloc_root(trans, root);
2191
2192                 list_add(&reloc_root->root_list, &reloc_roots);
2193         }
2194
2195         list_splice(&reloc_roots, &rc->reloc_roots);
2196
2197         if (!err)
2198                 btrfs_commit_transaction(trans, rc->extent_root);
2199         else
2200                 btrfs_end_transaction(trans, rc->extent_root);
2201         return err;
2202 }
2203
2204 static noinline_for_stack
2205 int merge_reloc_roots(struct reloc_control *rc)
2206 {
2207         struct btrfs_root *root;
2208         struct btrfs_root *reloc_root;
2209         LIST_HEAD(reloc_roots);
2210         int found = 0;
2211         int ret;
2212 again:
2213         root = rc->extent_root;
2214         mutex_lock(&root->fs_info->trans_mutex);
2215         list_splice_init(&rc->reloc_roots, &reloc_roots);
2216         mutex_unlock(&root->fs_info->trans_mutex);
2217
2218         while (!list_empty(&reloc_roots)) {
2219                 found = 1;
2220                 reloc_root = list_entry(reloc_roots.next,
2221                                         struct btrfs_root, root_list);
2222
2223                 if (btrfs_root_refs(&reloc_root->root_item) > 0) {
2224                         root = read_fs_root(reloc_root->fs_info,
2225                                             reloc_root->root_key.offset);
2226                         BUG_ON(IS_ERR(root));
2227                         BUG_ON(root->reloc_root != reloc_root);
2228
2229                         ret = merge_reloc_root(rc, root);
2230                         BUG_ON(ret);
2231                 } else {
2232                         list_del_init(&reloc_root->root_list);
2233                 }
2234                 btrfs_drop_snapshot(reloc_root, rc->block_rsv, 0);
2235         }
2236
2237         if (found) {
2238                 found = 0;
2239                 goto again;
2240         }
2241         BUG_ON(!RB_EMPTY_ROOT(&rc->reloc_root_tree.rb_root));
2242         return 0;
2243 }
2244
2245 static void free_block_list(struct rb_root *blocks)
2246 {
2247         struct tree_block *block;
2248         struct rb_node *rb_node;
2249         while ((rb_node = rb_first(blocks))) {
2250                 block = rb_entry(rb_node, struct tree_block, rb_node);
2251                 rb_erase(rb_node, blocks);
2252                 kfree(block);
2253         }
2254 }
2255
2256 static int record_reloc_root_in_trans(struct btrfs_trans_handle *trans,
2257                                       struct btrfs_root *reloc_root)
2258 {
2259         struct btrfs_root *root;
2260
2261         if (reloc_root->last_trans == trans->transid)
2262                 return 0;
2263
2264         root = read_fs_root(reloc_root->fs_info, reloc_root->root_key.offset);
2265         BUG_ON(IS_ERR(root));
2266         BUG_ON(root->reloc_root != reloc_root);
2267
2268         return btrfs_record_root_in_trans(trans, root);
2269 }
2270
2271 static noinline_for_stack
2272 struct btrfs_root *select_reloc_root(struct btrfs_trans_handle *trans,
2273                                      struct reloc_control *rc,
2274                                      struct backref_node *node,
2275                                      struct backref_edge *edges[], int *nr)
2276 {
2277         struct backref_node *next;
2278         struct btrfs_root *root;
2279         int index = 0;
2280
2281         next = node;
2282         while (1) {
2283                 cond_resched();
2284                 next = walk_up_backref(next, edges, &index);
2285                 root = next->root;
2286                 BUG_ON(!root);
2287                 BUG_ON(!root->ref_cows);
2288
2289                 if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID) {
2290                         record_reloc_root_in_trans(trans, root);
2291                         break;
2292                 }
2293
2294                 btrfs_record_root_in_trans(trans, root);
2295                 root = root->reloc_root;
2296
2297                 if (next->new_bytenr != root->node->start) {
2298                         BUG_ON(next->new_bytenr);
2299                         BUG_ON(!list_empty(&next->list));
2300                         next->new_bytenr = root->node->start;
2301                         next->root = root;
2302                         list_add_tail(&next->list,
2303                                       &rc->backref_cache.changed);
2304                         __mark_block_processed(rc, next);
2305                         break;
2306                 }
2307
2308                 WARN_ON(1);
2309                 root = NULL;
2310                 next = walk_down_backref(edges, &index);
2311                 if (!next || next->level <= node->level)
2312                         break;
2313         }
2314         if (!root)
2315                 return NULL;
2316
2317         *nr = index;
2318         next = node;
2319         /* setup backref node path for btrfs_reloc_cow_block */
2320         while (1) {
2321                 rc->backref_cache.path[next->level] = next;
2322                 if (--index < 0)
2323                         break;
2324                 next = edges[index]->node[UPPER];
2325         }
2326         return root;
2327 }
2328
2329 /*
2330  * select a tree root for relocation. return NULL if the block
2331  * is reference counted. we should use do_relocation() in this
2332  * case. return a tree root pointer if the block isn't reference
2333  * counted. return -ENOENT if the block is root of reloc tree.
2334  */
2335 static noinline_for_stack
2336 struct btrfs_root *select_one_root(struct btrfs_trans_handle *trans,
2337                                    struct backref_node *node)
2338 {
2339         struct backref_node *next;
2340         struct btrfs_root *root;
2341         struct btrfs_root *fs_root = NULL;
2342         struct backref_edge *edges[BTRFS_MAX_LEVEL - 1];
2343         int index = 0;
2344
2345         next = node;
2346         while (1) {
2347                 cond_resched();
2348                 next = walk_up_backref(next, edges, &index);
2349                 root = next->root;
2350                 BUG_ON(!root);
2351
2352                 /* no other choice for non-references counted tree */
2353                 if (!root->ref_cows)
2354                         return root;
2355
2356                 if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID)
2357                         fs_root = root;
2358
2359                 if (next != node)
2360                         return NULL;
2361
2362                 next = walk_down_backref(edges, &index);
2363                 if (!next || next->level <= node->level)
2364                         break;
2365         }
2366
2367         if (!fs_root)
2368                 return ERR_PTR(-ENOENT);
2369         return fs_root;
2370 }
2371
2372 static noinline_for_stack
2373 u64 calcu_metadata_size(struct reloc_control *rc,
2374                         struct backref_node *node, int reserve)
2375 {
2376         struct backref_node *next = node;
2377         struct backref_edge *edge;
2378         struct backref_edge *edges[BTRFS_MAX_LEVEL - 1];
2379         u64 num_bytes = 0;
2380         int index = 0;
2381
2382         BUG_ON(reserve && node->processed);
2383
2384         while (next) {
2385                 cond_resched();
2386                 while (1) {
2387                         if (next->processed && (reserve || next != node))
2388                                 break;
2389
2390                         num_bytes += btrfs_level_size(rc->extent_root,
2391                                                       next->level);
2392
2393                         if (list_empty(&next->upper))
2394                                 break;
2395
2396                         edge = list_entry(next->upper.next,
2397                                           struct backref_edge, list[LOWER]);
2398                         edges[index++] = edge;
2399                         next = edge->node[UPPER];
2400                 }
2401                 next = walk_down_backref(edges, &index);
2402         }
2403         return num_bytes;
2404 }
2405
2406 static int reserve_metadata_space(struct btrfs_trans_handle *trans,
2407                                   struct reloc_control *rc,
2408                                   struct backref_node *node)
2409 {
2410         struct btrfs_root *root = rc->extent_root;
2411         u64 num_bytes;
2412         int ret;
2413
2414         num_bytes = calcu_metadata_size(rc, node, 1) * 2;
2415
2416         trans->block_rsv = rc->block_rsv;
2417         ret = btrfs_block_rsv_add(trans, root, rc->block_rsv, num_bytes);
2418         if (ret) {
2419                 if (ret == -EAGAIN)
2420                         rc->commit_transaction = 1;
2421                 return ret;
2422         }
2423
2424         return 0;
2425 }
2426
2427 static void release_metadata_space(struct reloc_control *rc,
2428                                    struct backref_node *node)
2429 {
2430         u64 num_bytes = calcu_metadata_size(rc, node, 0) * 2;
2431         btrfs_block_rsv_release(rc->extent_root, rc->block_rsv, num_bytes);
2432 }
2433
2434 /*
2435  * relocate a block tree, and then update pointers in upper level
2436  * blocks that reference the block to point to the new location.
2437  *
2438  * if called by link_to_upper, the block has already been relocated.
2439  * in that case this function just updates pointers.
2440  */
2441 static int do_relocation(struct btrfs_trans_handle *trans,
2442                          struct reloc_control *rc,
2443                          struct backref_node *node,
2444                          struct btrfs_key *key,
2445                          struct btrfs_path *path, int lowest)
2446 {
2447         struct backref_node *upper;
2448         struct backref_edge *edge;
2449         struct backref_edge *edges[BTRFS_MAX_LEVEL - 1];
2450         struct btrfs_root *root;
2451         struct extent_buffer *eb;
2452         u32 blocksize;
2453         u64 bytenr;
2454         u64 generation;
2455         int nr;
2456         int slot;
2457         int ret;
2458         int err = 0;
2459
2460         BUG_ON(lowest && node->eb);
2461
2462         path->lowest_level = node->level + 1;
2463         rc->backref_cache.path[node->level] = node;
2464         list_for_each_entry(edge, &node->upper, list[LOWER]) {
2465                 cond_resched();
2466
2467                 upper = edge->node[UPPER];
2468                 root = select_reloc_root(trans, rc, upper, edges, &nr);
2469                 BUG_ON(!root);
2470
2471                 if (upper->eb && !upper->locked) {
2472                         if (!lowest) {
2473                                 ret = btrfs_bin_search(upper->eb, key,
2474                                                        upper->level, &slot);
2475                                 BUG_ON(ret);
2476                                 bytenr = btrfs_node_blockptr(upper->eb, slot);
2477                                 if (node->eb->start == bytenr)
2478                                         goto next;
2479                         }
2480                         drop_node_buffer(upper);
2481                 }
2482
2483                 if (!upper->eb) {
2484                         ret = btrfs_search_slot(trans, root, key, path, 0, 1);
2485                         if (ret < 0) {
2486                                 err = ret;
2487                                 break;
2488                         }
2489                         BUG_ON(ret > 0);
2490
2491                         if (!upper->eb) {
2492                                 upper->eb = path->nodes[upper->level];
2493                                 path->nodes[upper->level] = NULL;
2494                         } else {
2495                                 BUG_ON(upper->eb != path->nodes[upper->level]);
2496                         }
2497
2498                         upper->locked = 1;
2499                         path->locks[upper->level] = 0;
2500
2501                         slot = path->slots[upper->level];
2502                         btrfs_release_path(path);
2503                 } else {
2504                         ret = btrfs_bin_search(upper->eb, key, upper->level,
2505                                                &slot);
2506                         BUG_ON(ret);
2507                 }
2508
2509                 bytenr = btrfs_node_blockptr(upper->eb, slot);
2510                 if (lowest) {
2511                         BUG_ON(bytenr != node->bytenr);
2512                 } else {
2513                         if (node->eb->start == bytenr)
2514                                 goto next;
2515                 }
2516
2517                 blocksize = btrfs_level_size(root, node->level);
2518                 generation = btrfs_node_ptr_generation(upper->eb, slot);
2519                 eb = read_tree_block(root, bytenr, blocksize, generation);
2520                 if (!eb) {
2521                         err = -EIO;
2522                         goto next;
2523                 }
2524                 btrfs_tree_lock(eb);
2525                 btrfs_set_lock_blocking(eb);
2526
2527                 if (!node->eb) {
2528                         ret = btrfs_cow_block(trans, root, eb, upper->eb,
2529                                               slot, &eb);
2530                         btrfs_tree_unlock(eb);
2531                         free_extent_buffer(eb);
2532                         if (ret < 0) {
2533                                 err = ret;
2534                                 goto next;
2535                         }
2536                         BUG_ON(node->eb != eb);
2537                 } else {
2538                         btrfs_set_node_blockptr(upper->eb, slot,
2539                                                 node->eb->start);
2540                         btrfs_set_node_ptr_generation(upper->eb, slot,
2541                                                       trans->transid);
2542                         btrfs_mark_buffer_dirty(upper->eb);
2543
2544                         ret = btrfs_inc_extent_ref(trans, root,
2545                                                 node->eb->start, blocksize,
2546                                                 upper->eb->start,
2547                                                 btrfs_header_owner(upper->eb),
2548                                                 node->level, 0);
2549                         BUG_ON(ret);
2550
2551                         ret = btrfs_drop_subtree(trans, root, eb, upper->eb);
2552                         BUG_ON(ret);
2553                 }
2554 next:
2555                 if (!upper->pending)
2556                         drop_node_buffer(upper);
2557                 else
2558                         unlock_node_buffer(upper);
2559                 if (err)
2560                         break;
2561         }
2562
2563         if (!err && node->pending) {
2564                 drop_node_buffer(node);
2565                 list_move_tail(&node->list, &rc->backref_cache.changed);
2566                 node->pending = 0;
2567         }
2568
2569         path->lowest_level = 0;
2570         BUG_ON(err == -ENOSPC);
2571         return err;
2572 }
2573
2574 static int link_to_upper(struct btrfs_trans_handle *trans,
2575                          struct reloc_control *rc,
2576                          struct backref_node *node,
2577                          struct btrfs_path *path)
2578 {
2579         struct btrfs_key key;
2580
2581         btrfs_node_key_to_cpu(node->eb, &key, 0);
2582         return do_relocation(trans, rc, node, &key, path, 0);
2583 }
2584
2585 static int finish_pending_nodes(struct btrfs_trans_handle *trans,
2586                                 struct reloc_control *rc,
2587                                 struct btrfs_path *path, int err)
2588 {
2589         LIST_HEAD(list);
2590         struct backref_cache *cache = &rc->backref_cache;
2591         struct backref_node *node;
2592         int level;
2593         int ret;
2594
2595         for (level = 0; level < BTRFS_MAX_LEVEL; level++) {
2596                 while (!list_empty(&cache->pending[level])) {
2597                         node = list_entry(cache->pending[level].next,
2598                                           struct backref_node, list);
2599                         list_move_tail(&node->list, &list);
2600                         BUG_ON(!node->pending);
2601
2602                         if (!err) {
2603                                 ret = link_to_upper(trans, rc, node, path);
2604                                 if (ret < 0)
2605                                         err = ret;
2606                         }
2607                 }
2608                 list_splice_init(&list, &cache->pending[level]);
2609         }
2610         return err;
2611 }
2612
2613 static void mark_block_processed(struct reloc_control *rc,
2614                                  u64 bytenr, u32 blocksize)
2615 {
2616         set_extent_bits(&rc->processed_blocks, bytenr, bytenr + blocksize - 1,
2617                         EXTENT_DIRTY, GFP_NOFS);
2618 }
2619
2620 static void __mark_block_processed(struct reloc_control *rc,
2621                                    struct backref_node *node)
2622 {
2623         u32 blocksize;
2624         if (node->level == 0 ||
2625             in_block_group(node->bytenr, rc->block_group)) {
2626                 blocksize = btrfs_level_size(rc->extent_root, node->level);
2627                 mark_block_processed(rc, node->bytenr, blocksize);
2628         }
2629         node->processed = 1;
2630 }
2631
2632 /*
2633  * mark a block and all blocks directly/indirectly reference the block
2634  * as processed.
2635  */
2636 static void update_processed_blocks(struct reloc_control *rc,
2637                                     struct backref_node *node)
2638 {
2639         struct backref_node *next = node;
2640         struct backref_edge *edge;
2641         struct backref_edge *edges[BTRFS_MAX_LEVEL - 1];
2642         int index = 0;
2643
2644         while (next) {
2645                 cond_resched();
2646                 while (1) {
2647                         if (next->processed)
2648                                 break;
2649
2650                         __mark_block_processed(rc, next);
2651
2652                         if (list_empty(&next->upper))
2653                                 break;
2654
2655                         edge = list_entry(next->upper.next,
2656                                           struct backref_edge, list[LOWER]);
2657                         edges[index++] = edge;
2658                         next = edge->node[UPPER];
2659                 }
2660                 next = walk_down_backref(edges, &index);
2661         }
2662 }
2663
2664 static int tree_block_processed(u64 bytenr, u32 blocksize,
2665                                 struct reloc_control *rc)
2666 {
2667         if (test_range_bit(&rc->processed_blocks, bytenr,
2668                            bytenr + blocksize - 1, EXTENT_DIRTY, 1, NULL))
2669                 return 1;
2670         return 0;
2671 }
2672
2673 static int get_tree_block_key(struct reloc_control *rc,
2674                               struct tree_block *block)
2675 {
2676         struct extent_buffer *eb;
2677
2678         BUG_ON(block->key_ready);
2679         eb = read_tree_block(rc->extent_root, block->bytenr,
2680                              block->key.objectid, block->key.offset);
2681         BUG_ON(!eb);
2682         WARN_ON(btrfs_header_level(eb) != block->level);
2683         if (block->level == 0)
2684                 btrfs_item_key_to_cpu(eb, &block->key, 0);
2685         else
2686                 btrfs_node_key_to_cpu(eb, &block->key, 0);
2687         free_extent_buffer(eb);
2688         block->key_ready = 1;
2689         return 0;
2690 }
2691
2692 static int reada_tree_block(struct reloc_control *rc,
2693                             struct tree_block *block)
2694 {
2695         BUG_ON(block->key_ready);
2696         readahead_tree_block(rc->extent_root, block->bytenr,
2697                              block->key.objectid, block->key.offset);
2698         return 0;
2699 }
2700
2701 /*
2702  * helper function to relocate a tree block
2703  */
2704 static int relocate_tree_block(struct btrfs_trans_handle *trans,
2705                                 struct reloc_control *rc,
2706                                 struct backref_node *node,
2707                                 struct btrfs_key *key,
2708                                 struct btrfs_path *path)
2709 {
2710         struct btrfs_root *root;
2711         int release = 0;
2712         int ret = 0;
2713
2714         if (!node)
2715                 return 0;
2716
2717         BUG_ON(node->processed);
2718         root = select_one_root(trans, node);
2719         if (root == ERR_PTR(-ENOENT)) {
2720                 update_processed_blocks(rc, node);
2721                 goto out;
2722         }
2723
2724         if (!root || root->ref_cows) {
2725                 ret = reserve_metadata_space(trans, rc, node);
2726                 if (ret)
2727                         goto out;
2728                 release = 1;
2729         }
2730
2731         if (root) {
2732                 if (root->ref_cows) {
2733                         BUG_ON(node->new_bytenr);
2734                         BUG_ON(!list_empty(&node->list));
2735                         btrfs_record_root_in_trans(trans, root);
2736                         root = root->reloc_root;
2737                         node->new_bytenr = root->node->start;
2738                         node->root = root;
2739                         list_add_tail(&node->list, &rc->backref_cache.changed);
2740                 } else {
2741                         path->lowest_level = node->level;
2742                         ret = btrfs_search_slot(trans, root, key, path, 0, 1);
2743                         btrfs_release_path(path);
2744                         if (ret > 0)
2745                                 ret = 0;
2746                 }
2747                 if (!ret)
2748                         update_processed_blocks(rc, node);
2749         } else {
2750                 ret = do_relocation(trans, rc, node, key, path, 1);
2751         }
2752 out:
2753         if (ret || node->level == 0 || node->cowonly) {
2754                 if (release)
2755                         release_metadata_space(rc, node);
2756                 remove_backref_node(&rc->backref_cache, node);
2757         }
2758         return ret;
2759 }
2760
2761 /*
2762  * relocate a list of blocks
2763  */
2764 static noinline_for_stack
2765 int relocate_tree_blocks(struct btrfs_trans_handle *trans,
2766                          struct reloc_control *rc, struct rb_root *blocks)
2767 {
2768         struct backref_node *node;
2769         struct btrfs_path *path;
2770         struct tree_block *block;
2771         struct rb_node *rb_node;
2772         int ret;
2773         int err = 0;
2774
2775         path = btrfs_alloc_path();
2776         if (!path)
2777                 return -ENOMEM;
2778
2779         rb_node = rb_first(blocks);
2780         while (rb_node) {
2781                 block = rb_entry(rb_node, struct tree_block, rb_node);
2782                 if (!block->key_ready)
2783                         reada_tree_block(rc, block);
2784                 rb_node = rb_next(rb_node);
2785         }
2786
2787         rb_node = rb_first(blocks);
2788         while (rb_node) {
2789                 block = rb_entry(rb_node, struct tree_block, rb_node);
2790                 if (!block->key_ready)
2791                         get_tree_block_key(rc, block);
2792                 rb_node = rb_next(rb_node);
2793         }
2794
2795         rb_node = rb_first(blocks);
2796         while (rb_node) {
2797                 block = rb_entry(rb_node, struct tree_block, rb_node);
2798
2799                 node = build_backref_tree(rc, &block->key,
2800                                           block->level, block->bytenr);
2801                 if (IS_ERR(node)) {
2802                         err = PTR_ERR(node);
2803                         goto out;
2804                 }
2805
2806                 ret = relocate_tree_block(trans, rc, node, &block->key,
2807                                           path);
2808                 if (ret < 0) {
2809                         if (ret != -EAGAIN || rb_node == rb_first(blocks))
2810                                 err = ret;
2811                         goto out;
2812                 }
2813                 rb_node = rb_next(rb_node);
2814         }
2815 out:
2816         free_block_list(blocks);
2817         err = finish_pending_nodes(trans, rc, path, err);
2818
2819         btrfs_free_path(path);
2820         return err;
2821 }
2822
2823 static noinline_for_stack
2824 int prealloc_file_extent_cluster(struct inode *inode,
2825                                  struct file_extent_cluster *cluster)
2826 {
2827         u64 alloc_hint = 0;
2828         u64 start;
2829         u64 end;
2830         u64 offset = BTRFS_I(inode)->index_cnt;
2831         u64 num_bytes;
2832         int nr = 0;
2833         int ret = 0;
2834
2835         BUG_ON(cluster->start != cluster->boundary[0]);
2836         mutex_lock(&inode->i_mutex);
2837
2838         ret = btrfs_check_data_free_space(inode, cluster->end +
2839                                           1 - cluster->start);
2840         if (ret)
2841                 goto out;
2842
2843         while (nr < cluster->nr) {
2844                 start = cluster->boundary[nr] - offset;
2845                 if (nr + 1 < cluster->nr)
2846                         end = cluster->boundary[nr + 1] - 1 - offset;
2847                 else
2848                         end = cluster->end - offset;
2849
2850                 lock_extent(&BTRFS_I(inode)->io_tree, start, end, GFP_NOFS);
2851                 num_bytes = end + 1 - start;
2852                 ret = btrfs_prealloc_file_range(inode, 0, start,
2853                                                 num_bytes, num_bytes,
2854                                                 end + 1, &alloc_hint);
2855                 unlock_extent(&BTRFS_I(inode)->io_tree, start, end, GFP_NOFS);
2856                 if (ret)
2857                         break;
2858                 nr++;
2859         }
2860         btrfs_free_reserved_data_space(inode, cluster->end +
2861                                        1 - cluster->start);
2862 out:
2863         mutex_unlock(&inode->i_mutex);
2864         return ret;
2865 }
2866
2867 static noinline_for_stack
2868 int setup_extent_mapping(struct inode *inode, u64 start, u64 end,
2869                          u64 block_start)
2870 {
2871         struct btrfs_root *root = BTRFS_I(inode)->root;
2872         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
2873         struct extent_map *em;
2874         int ret = 0;
2875
2876         em = alloc_extent_map();
2877         if (!em)
2878                 return -ENOMEM;
2879
2880         em->start = start;
2881         em->len = end + 1 - start;
2882         em->block_len = em->len;
2883         em->block_start = block_start;
2884         em->bdev = root->fs_info->fs_devices->latest_bdev;
2885         set_bit(EXTENT_FLAG_PINNED, &em->flags);
2886
2887         lock_extent(&BTRFS_I(inode)->io_tree, start, end, GFP_NOFS);
2888         while (1) {
2889                 write_lock(&em_tree->lock);
2890                 ret = add_extent_mapping(em_tree, em);
2891                 write_unlock(&em_tree->lock);
2892                 if (ret != -EEXIST) {
2893                         free_extent_map(em);
2894                         break;
2895                 }
2896                 btrfs_drop_extent_cache(inode, start, end, 0);
2897         }
2898         unlock_extent(&BTRFS_I(inode)->io_tree, start, end, GFP_NOFS);
2899         return ret;
2900 }
2901
2902 static int relocate_file_extent_cluster(struct inode *inode,
2903                                         struct file_extent_cluster *cluster)
2904 {
2905         u64 page_start;
2906         u64 page_end;
2907         u64 offset = BTRFS_I(inode)->index_cnt;
2908         unsigned long index;
2909         unsigned long last_index;
2910         struct page *page;
2911         struct file_ra_state *ra;
2912         int nr = 0;
2913         int ret = 0;
2914
2915         if (!cluster->nr)
2916                 return 0;
2917
2918         ra = kzalloc(sizeof(*ra), GFP_NOFS);
2919         if (!ra)
2920                 return -ENOMEM;
2921
2922         ret = prealloc_file_extent_cluster(inode, cluster);
2923         if (ret)
2924                 goto out;
2925
2926         file_ra_state_init(ra, inode->i_mapping);
2927
2928         ret = setup_extent_mapping(inode, cluster->start - offset,
2929                                    cluster->end - offset, cluster->start);
2930         if (ret)
2931                 goto out;
2932
2933         index = (cluster->start - offset) >> PAGE_CACHE_SHIFT;
2934         last_index = (cluster->end - offset) >> PAGE_CACHE_SHIFT;
2935         while (index <= last_index) {
2936                 ret = btrfs_delalloc_reserve_metadata(inode, PAGE_CACHE_SIZE);
2937                 if (ret)
2938                         goto out;
2939
2940                 page = find_lock_page(inode->i_mapping, index);
2941                 if (!page) {
2942                         page_cache_sync_readahead(inode->i_mapping,
2943                                                   ra, NULL, index,
2944                                                   last_index + 1 - index);
2945                         page = grab_cache_page(inode->i_mapping, index);
2946                         if (!page) {
2947                                 btrfs_delalloc_release_metadata(inode,
2948                                                         PAGE_CACHE_SIZE);
2949                                 ret = -ENOMEM;
2950                                 goto out;
2951                         }
2952                 }
2953
2954                 if (PageReadahead(page)) {
2955                         page_cache_async_readahead(inode->i_mapping,
2956                                                    ra, NULL, page, index,
2957                                                    last_index + 1 - index);
2958                 }
2959
2960                 if (!PageUptodate(page)) {
2961                         btrfs_readpage(NULL, page);
2962                         lock_page(page);
2963                         if (!PageUptodate(page)) {
2964                                 unlock_page(page);
2965                                 page_cache_release(page);
2966                                 btrfs_delalloc_release_metadata(inode,
2967                                                         PAGE_CACHE_SIZE);
2968                                 ret = -EIO;
2969                                 goto out;
2970                         }
2971                 }
2972
2973                 page_start = (u64)page->index << PAGE_CACHE_SHIFT;
2974                 page_end = page_start + PAGE_CACHE_SIZE - 1;
2975
2976                 lock_extent(&BTRFS_I(inode)->io_tree,
2977                             page_start, page_end, GFP_NOFS);
2978
2979                 set_page_extent_mapped(page);
2980
2981                 if (nr < cluster->nr &&
2982                     page_start + offset == cluster->boundary[nr]) {
2983                         set_extent_bits(&BTRFS_I(inode)->io_tree,
2984                                         page_start, page_end,
2985                                         EXTENT_BOUNDARY, GFP_NOFS);
2986                         nr++;
2987                 }
2988
2989                 btrfs_set_extent_delalloc(inode, page_start, page_end, NULL);
2990                 set_page_dirty(page);
2991
2992                 unlock_extent(&BTRFS_I(inode)->io_tree,
2993                               page_start, page_end, GFP_NOFS);
2994                 unlock_page(page);
2995                 page_cache_release(page);
2996
2997                 index++;
2998                 balance_dirty_pages_ratelimited(inode->i_mapping);
2999                 btrfs_throttle(BTRFS_I(inode)->root);
3000         }
3001         WARN_ON(nr != cluster->nr);
3002 out:
3003         kfree(ra);
3004         return ret;
3005 }
3006
3007 static noinline_for_stack
3008 int relocate_data_extent(struct inode *inode, struct btrfs_key *extent_key,
3009                          struct file_extent_cluster *cluster)
3010 {
3011         int ret;
3012
3013         if (cluster->nr > 0 && extent_key->objectid != cluster->end + 1) {
3014                 ret = relocate_file_extent_cluster(inode, cluster);
3015                 if (ret)
3016                         return ret;
3017                 cluster->nr = 0;
3018         }
3019
3020         if (!cluster->nr)
3021                 cluster->start = extent_key->objectid;
3022         else
3023                 BUG_ON(cluster->nr >= MAX_EXTENTS);
3024         cluster->end = extent_key->objectid + extent_key->offset - 1;
3025         cluster->boundary[cluster->nr] = extent_key->objectid;
3026         cluster->nr++;
3027
3028         if (cluster->nr >= MAX_EXTENTS) {
3029                 ret = relocate_file_extent_cluster(inode, cluster);
3030                 if (ret)
3031                         return ret;
3032                 cluster->nr = 0;
3033         }
3034         return 0;
3035 }
3036
3037 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
3038 static int get_ref_objectid_v0(struct reloc_control *rc,
3039                                struct btrfs_path *path,
3040                                struct btrfs_key *extent_key,
3041                                u64 *ref_objectid, int *path_change)
3042 {
3043         struct btrfs_key key;
3044         struct extent_buffer *leaf;
3045         struct btrfs_extent_ref_v0 *ref0;
3046         int ret;
3047         int slot;
3048
3049         leaf = path->nodes[0];
3050         slot = path->slots[0];
3051         while (1) {
3052                 if (slot >= btrfs_header_nritems(leaf)) {
3053                         ret = btrfs_next_leaf(rc->extent_root, path);
3054                         if (ret < 0)
3055                                 return ret;
3056                         BUG_ON(ret > 0);
3057                         leaf = path->nodes[0];
3058                         slot = path->slots[0];
3059                         if (path_change)
3060                                 *path_change = 1;
3061                 }
3062                 btrfs_item_key_to_cpu(leaf, &key, slot);
3063                 if (key.objectid != extent_key->objectid)
3064                         return -ENOENT;
3065
3066                 if (key.type != BTRFS_EXTENT_REF_V0_KEY) {
3067                         slot++;
3068                         continue;
3069                 }
3070                 ref0 = btrfs_item_ptr(leaf, slot,
3071                                 struct btrfs_extent_ref_v0);
3072                 *ref_objectid = btrfs_ref_objectid_v0(leaf, ref0);
3073                 break;
3074         }
3075         return 0;
3076 }
3077 #endif
3078
3079 /*
3080  * helper to add a tree block to the list.
3081  * the major work is getting the generation and level of the block
3082  */
3083 static int add_tree_block(struct reloc_control *rc,
3084                           struct btrfs_key *extent_key,
3085                           struct btrfs_path *path,
3086                           struct rb_root *blocks)
3087 {
3088         struct extent_buffer *eb;
3089         struct btrfs_extent_item *ei;
3090         struct btrfs_tree_block_info *bi;
3091         struct tree_block *block;
3092         struct rb_node *rb_node;
3093         u32 item_size;
3094         int level = -1;
3095         int generation;
3096
3097         eb =  path->nodes[0];
3098         item_size = btrfs_item_size_nr(eb, path->slots[0]);
3099
3100         if (item_size >= sizeof(*ei) + sizeof(*bi)) {
3101                 ei = btrfs_item_ptr(eb, path->slots[0],
3102                                 struct btrfs_extent_item);
3103                 bi = (struct btrfs_tree_block_info *)(ei + 1);
3104                 generation = btrfs_extent_generation(eb, ei);
3105                 level = btrfs_tree_block_level(eb, bi);
3106         } else {
3107 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
3108                 u64 ref_owner;
3109                 int ret;
3110
3111                 BUG_ON(item_size != sizeof(struct btrfs_extent_item_v0));
3112                 ret = get_ref_objectid_v0(rc, path, extent_key,
3113                                           &ref_owner, NULL);
3114                 if (ret < 0)
3115                         return ret;
3116                 BUG_ON(ref_owner >= BTRFS_MAX_LEVEL);
3117                 level = (int)ref_owner;
3118                 /* FIXME: get real generation */
3119                 generation = 0;
3120 #else
3121                 BUG();
3122 #endif
3123         }
3124
3125         btrfs_release_path(path);
3126
3127         BUG_ON(level == -1);
3128
3129         block = kmalloc(sizeof(*block), GFP_NOFS);
3130         if (!block)
3131                 return -ENOMEM;
3132
3133         block->bytenr = extent_key->objectid;
3134         block->key.objectid = extent_key->offset;
3135         block->key.offset = generation;
3136         block->level = level;
3137         block->key_ready = 0;
3138
3139         rb_node = tree_insert(blocks, block->bytenr, &block->rb_node);
3140         BUG_ON(rb_node);
3141
3142         return 0;
3143 }
3144
3145 /*
3146  * helper to add tree blocks for backref of type BTRFS_SHARED_DATA_REF_KEY
3147  */
3148 static int __add_tree_block(struct reloc_control *rc,
3149                             u64 bytenr, u32 blocksize,
3150                             struct rb_root *blocks)
3151 {
3152         struct btrfs_path *path;
3153         struct btrfs_key key;
3154         int ret;
3155
3156         if (tree_block_processed(bytenr, blocksize, rc))
3157                 return 0;
3158
3159         if (tree_search(blocks, bytenr))
3160                 return 0;
3161
3162         path = btrfs_alloc_path();
3163         if (!path)
3164                 return -ENOMEM;
3165
3166         key.objectid = bytenr;
3167         key.type = BTRFS_EXTENT_ITEM_KEY;
3168         key.offset = blocksize;
3169
3170         path->search_commit_root = 1;
3171         path->skip_locking = 1;
3172         ret = btrfs_search_slot(NULL, rc->extent_root, &key, path, 0, 0);
3173         if (ret < 0)
3174                 goto out;
3175         BUG_ON(ret);
3176
3177         btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0]);
3178         ret = add_tree_block(rc, &key, path, blocks);
3179 out:
3180         btrfs_free_path(path);
3181         return ret;
3182 }
3183
3184 /*
3185  * helper to check if the block use full backrefs for pointers in it
3186  */
3187 static int block_use_full_backref(struct reloc_control *rc,
3188                                   struct extent_buffer *eb)
3189 {
3190         u64 flags;
3191         int ret;
3192
3193         if (btrfs_header_flag(eb, BTRFS_HEADER_FLAG_RELOC) ||
3194             btrfs_header_backref_rev(eb) < BTRFS_MIXED_BACKREF_REV)
3195                 return 1;
3196
3197         ret = btrfs_lookup_extent_info(NULL, rc->extent_root,
3198                                        eb->start, eb->len, NULL, &flags);
3199         BUG_ON(ret);
3200
3201         if (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF)
3202                 ret = 1;
3203         else
3204                 ret = 0;
3205         return ret;
3206 }
3207
3208 static int delete_block_group_cache(struct btrfs_fs_info *fs_info,
3209                                     struct inode *inode, u64 ino)
3210 {
3211         struct btrfs_key key;
3212         struct btrfs_path *path;
3213         struct btrfs_root *root = fs_info->tree_root;
3214         struct btrfs_trans_handle *trans;
3215         unsigned long nr;
3216         int ret = 0;
3217
3218         if (inode)
3219                 goto truncate;
3220
3221         key.objectid = ino;
3222         key.type = BTRFS_INODE_ITEM_KEY;
3223         key.offset = 0;
3224
3225         inode = btrfs_iget(fs_info->sb, &key, root, NULL);
3226         if (IS_ERR_OR_NULL(inode) || is_bad_inode(inode)) {
3227                 if (inode && !IS_ERR(inode))
3228                         iput(inode);
3229                 return -ENOENT;
3230         }
3231
3232 truncate:
3233         path = btrfs_alloc_path();
3234         if (!path) {
3235                 ret = -ENOMEM;
3236                 goto out;
3237         }
3238
3239         trans = btrfs_join_transaction(root, 0);
3240         if (IS_ERR(trans)) {
3241                 btrfs_free_path(path);
3242                 ret = PTR_ERR(trans);
3243                 goto out;
3244         }
3245
3246         ret = btrfs_truncate_free_space_cache(root, trans, path, inode);
3247
3248         btrfs_free_path(path);
3249         nr = trans->blocks_used;
3250         btrfs_end_transaction(trans, root);
3251         btrfs_btree_balance_dirty(root, nr);
3252 out:
3253         iput(inode);
3254         return ret;
3255 }
3256
3257 /*
3258  * helper to add tree blocks for backref of type BTRFS_EXTENT_DATA_REF_KEY
3259  * this function scans fs tree to find blocks reference the data extent
3260  */
3261 static int find_data_references(struct reloc_control *rc,
3262                                 struct btrfs_key *extent_key,
3263                                 struct extent_buffer *leaf,
3264                                 struct btrfs_extent_data_ref *ref,
3265                                 struct rb_root *blocks)
3266 {
3267         struct btrfs_path *path;
3268         struct tree_block *block;
3269         struct btrfs_root *root;
3270         struct btrfs_file_extent_item *fi;
3271         struct rb_node *rb_node;
3272         struct btrfs_key key;
3273         u64 ref_root;
3274         u64 ref_objectid;
3275         u64 ref_offset;
3276         u32 ref_count;
3277         u32 nritems;
3278         int err = 0;
3279         int added = 0;
3280         int counted;
3281         int ret;
3282
3283         ref_root = btrfs_extent_data_ref_root(leaf, ref);
3284         ref_objectid = btrfs_extent_data_ref_objectid(leaf, ref);
3285         ref_offset = btrfs_extent_data_ref_offset(leaf, ref);
3286         ref_count = btrfs_extent_data_ref_count(leaf, ref);
3287
3288         /*
3289          * This is an extent belonging to the free space cache, lets just delete
3290          * it and redo the search.
3291          */
3292         if (ref_root == BTRFS_ROOT_TREE_OBJECTID) {
3293                 ret = delete_block_group_cache(rc->extent_root->fs_info,
3294                                                NULL, ref_objectid);
3295                 if (ret != -ENOENT)
3296                         return ret;
3297                 ret = 0;
3298         }
3299
3300         path = btrfs_alloc_path();
3301         if (!path)
3302                 return -ENOMEM;
3303
3304         root = read_fs_root(rc->extent_root->fs_info, ref_root);
3305         if (IS_ERR(root)) {
3306                 err = PTR_ERR(root);
3307                 goto out;
3308         }
3309
3310         key.objectid = ref_objectid;
3311         key.offset = ref_offset;
3312         key.type = BTRFS_EXTENT_DATA_KEY;
3313
3314         path->search_commit_root = 1;
3315         path->skip_locking = 1;
3316         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
3317         if (ret < 0) {
3318                 err = ret;
3319                 goto out;
3320         }
3321
3322         leaf = path->nodes[0];
3323         nritems = btrfs_header_nritems(leaf);
3324         /*
3325          * the references in tree blocks that use full backrefs
3326          * are not counted in
3327          */
3328         if (block_use_full_backref(rc, leaf))
3329                 counted = 0;
3330         else
3331                 counted = 1;
3332         rb_node = tree_search(blocks, leaf->start);
3333         if (rb_node) {
3334                 if (counted)
3335                         added = 1;
3336                 else
3337                         path->slots[0] = nritems;
3338         }
3339
3340         while (ref_count > 0) {
3341                 while (path->slots[0] >= nritems) {
3342                         ret = btrfs_next_leaf(root, path);
3343                         if (ret < 0) {
3344                                 err = ret;
3345                                 goto out;
3346                         }
3347                         if (ret > 0) {
3348                                 WARN_ON(1);
3349                                 goto out;
3350                         }
3351
3352                         leaf = path->nodes[0];
3353                         nritems = btrfs_header_nritems(leaf);
3354                         added = 0;
3355
3356                         if (block_use_full_backref(rc, leaf))
3357                                 counted = 0;
3358                         else
3359                                 counted = 1;
3360                         rb_node = tree_search(blocks, leaf->start);
3361                         if (rb_node) {
3362                                 if (counted)
3363                                         added = 1;
3364                                 else
3365                                         path->slots[0] = nritems;
3366                         }
3367                 }
3368
3369                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
3370                 if (key.objectid != ref_objectid ||
3371                     key.type != BTRFS_EXTENT_DATA_KEY) {
3372                         WARN_ON(1);
3373                         break;
3374                 }
3375
3376                 fi = btrfs_item_ptr(leaf, path->slots[0],
3377                                     struct btrfs_file_extent_item);
3378
3379                 if (btrfs_file_extent_type(leaf, fi) ==
3380                     BTRFS_FILE_EXTENT_INLINE)
3381                         goto next;
3382
3383                 if (btrfs_file_extent_disk_bytenr(leaf, fi) !=
3384                     extent_key->objectid)
3385                         goto next;
3386
3387                 key.offset -= btrfs_file_extent_offset(leaf, fi);
3388                 if (key.offset != ref_offset)
3389                         goto next;
3390
3391                 if (counted)
3392                         ref_count--;
3393                 if (added)
3394                         goto next;
3395
3396                 if (!tree_block_processed(leaf->start, leaf->len, rc)) {
3397                         block = kmalloc(sizeof(*block), GFP_NOFS);
3398                         if (!block) {
3399                                 err = -ENOMEM;
3400                                 break;
3401                         }
3402                         block->bytenr = leaf->start;
3403                         btrfs_item_key_to_cpu(leaf, &block->key, 0);
3404                         block->level = 0;
3405                         block->key_ready = 1;
3406                         rb_node = tree_insert(blocks, block->bytenr,
3407                                               &block->rb_node);
3408                         BUG_ON(rb_node);
3409                 }
3410                 if (counted)
3411                         added = 1;
3412                 else
3413                         path->slots[0] = nritems;
3414 next:
3415                 path->slots[0]++;
3416
3417         }
3418 out:
3419         btrfs_free_path(path);
3420         return err;
3421 }
3422
3423 /*
3424  * hepler to find all tree blocks that reference a given data extent
3425  */
3426 static noinline_for_stack
3427 int add_data_references(struct reloc_control *rc,
3428                         struct btrfs_key *extent_key,
3429                         struct btrfs_path *path,
3430                         struct rb_root *blocks)
3431 {
3432         struct btrfs_key key;
3433         struct extent_buffer *eb;
3434         struct btrfs_extent_data_ref *dref;
3435         struct btrfs_extent_inline_ref *iref;
3436         unsigned long ptr;
3437         unsigned long end;
3438         u32 blocksize = btrfs_level_size(rc->extent_root, 0);
3439         int ret;
3440         int err = 0;
3441
3442         eb = path->nodes[0];
3443         ptr = btrfs_item_ptr_offset(eb, path->slots[0]);
3444         end = ptr + btrfs_item_size_nr(eb, path->slots[0]);
3445 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
3446         if (ptr + sizeof(struct btrfs_extent_item_v0) == end)
3447                 ptr = end;
3448         else
3449 #endif
3450                 ptr += sizeof(struct btrfs_extent_item);
3451
3452         while (ptr < end) {
3453                 iref = (struct btrfs_extent_inline_ref *)ptr;
3454                 key.type = btrfs_extent_inline_ref_type(eb, iref);
3455                 if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
3456                         key.offset = btrfs_extent_inline_ref_offset(eb, iref);
3457                         ret = __add_tree_block(rc, key.offset, blocksize,
3458                                                blocks);
3459                 } else if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
3460                         dref = (struct btrfs_extent_data_ref *)(&iref->offset);
3461                         ret = find_data_references(rc, extent_key,
3462                                                    eb, dref, blocks);
3463                 } else {
3464                         BUG();
3465                 }
3466                 ptr += btrfs_extent_inline_ref_size(key.type);
3467         }
3468         WARN_ON(ptr > end);
3469
3470         while (1) {
3471                 cond_resched();
3472                 eb = path->nodes[0];
3473                 if (path->slots[0] >= btrfs_header_nritems(eb)) {
3474                         ret = btrfs_next_leaf(rc->extent_root, path);
3475                         if (ret < 0) {
3476                                 err = ret;
3477                                 break;
3478                         }
3479                         if (ret > 0)
3480                                 break;
3481                         eb = path->nodes[0];
3482                 }
3483
3484                 btrfs_item_key_to_cpu(eb, &key, path->slots[0]);
3485                 if (key.objectid != extent_key->objectid)
3486                         break;
3487
3488 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
3489                 if (key.type == BTRFS_SHARED_DATA_REF_KEY ||
3490                     key.type == BTRFS_EXTENT_REF_V0_KEY) {
3491 #else
3492                 BUG_ON(key.type == BTRFS_EXTENT_REF_V0_KEY);
3493                 if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
3494 #endif
3495                         ret = __add_tree_block(rc, key.offset, blocksize,
3496                                                blocks);
3497                 } else if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
3498                         dref = btrfs_item_ptr(eb, path->slots[0],
3499                                               struct btrfs_extent_data_ref);
3500                         ret = find_data_references(rc, extent_key,
3501                                                    eb, dref, blocks);
3502                 } else {
3503                         ret = 0;
3504                 }
3505                 if (ret) {
3506                         err = ret;
3507                         break;
3508                 }
3509                 path->slots[0]++;
3510         }
3511         btrfs_release_path(path);
3512         if (err)
3513                 free_block_list(blocks);
3514         return err;
3515 }
3516
3517 /*
3518  * hepler to find next unprocessed extent
3519  */
3520 static noinline_for_stack
3521 int find_next_extent(struct btrfs_trans_handle *trans,
3522                      struct reloc_control *rc, struct btrfs_path *path,
3523                      struct btrfs_key *extent_key)
3524 {
3525         struct btrfs_key key;
3526         struct extent_buffer *leaf;
3527         u64 start, end, last;
3528         int ret;
3529
3530         last = rc->block_group->key.objectid + rc->block_group->key.offset;
3531         while (1) {
3532                 cond_resched();
3533                 if (rc->search_start >= last) {
3534                         ret = 1;
3535                         break;
3536                 }
3537
3538                 key.objectid = rc->search_start;
3539                 key.type = BTRFS_EXTENT_ITEM_KEY;
3540                 key.offset = 0;
3541
3542                 path->search_commit_root = 1;
3543                 path->skip_locking = 1;
3544                 ret = btrfs_search_slot(NULL, rc->extent_root, &key, path,
3545                                         0, 0);
3546                 if (ret < 0)
3547                         break;
3548 next:
3549                 leaf = path->nodes[0];
3550                 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
3551                         ret = btrfs_next_leaf(rc->extent_root, path);
3552                         if (ret != 0)
3553                                 break;
3554                         leaf = path->nodes[0];
3555                 }
3556
3557                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
3558                 if (key.objectid >= last) {
3559                         ret = 1;
3560                         break;
3561                 }
3562
3563                 if (key.type != BTRFS_EXTENT_ITEM_KEY ||
3564                     key.objectid + key.offset <= rc->search_start) {
3565                         path->slots[0]++;
3566                         goto next;
3567                 }
3568
3569                 ret = find_first_extent_bit(&rc->processed_blocks,
3570                                             key.objectid, &start, &end,
3571                                             EXTENT_DIRTY);
3572
3573                 if (ret == 0 && start <= key.objectid) {
3574                         btrfs_release_path(path);
3575                         rc->search_start = end + 1;
3576                 } else {
3577                         rc->search_start = key.objectid + key.offset;
3578                         memcpy(extent_key, &key, sizeof(key));
3579                         return 0;
3580                 }
3581         }
3582         btrfs_release_path(path);
3583         return ret;
3584 }
3585
3586 static void set_reloc_control(struct reloc_control *rc)
3587 {
3588         struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
3589         mutex_lock(&fs_info->trans_mutex);
3590         fs_info->reloc_ctl = rc;
3591         mutex_unlock(&fs_info->trans_mutex);
3592 }
3593
3594 static void unset_reloc_control(struct reloc_control *rc)
3595 {
3596         struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
3597         mutex_lock(&fs_info->trans_mutex);
3598         fs_info->reloc_ctl = NULL;
3599         mutex_unlock(&fs_info->trans_mutex);
3600 }
3601
3602 static int check_extent_flags(u64 flags)
3603 {
3604         if ((flags & BTRFS_EXTENT_FLAG_DATA) &&
3605             (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK))
3606                 return 1;
3607         if (!(flags & BTRFS_EXTENT_FLAG_DATA) &&
3608             !(flags & BTRFS_EXTENT_FLAG_TREE_BLOCK))
3609                 return 1;
3610         if ((flags & BTRFS_EXTENT_FLAG_DATA) &&
3611             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
3612                 return 1;
3613         return 0;
3614 }
3615
3616 static noinline_for_stack
3617 int prepare_to_relocate(struct reloc_control *rc)
3618 {
3619         struct btrfs_trans_handle *trans;
3620         int ret;
3621
3622         rc->block_rsv = btrfs_alloc_block_rsv(rc->extent_root);
3623         if (!rc->block_rsv)
3624                 return -ENOMEM;
3625
3626         /*
3627          * reserve some space for creating reloc trees.
3628          * btrfs_init_reloc_root will use them when there
3629          * is no reservation in transaction handle.
3630          */
3631         ret = btrfs_block_rsv_add(NULL, rc->extent_root, rc->block_rsv,
3632                                   rc->extent_root->nodesize * 256);
3633         if (ret)
3634                 return ret;
3635
3636         rc->block_rsv->refill_used = 1;
3637         btrfs_add_durable_block_rsv(rc->extent_root->fs_info, rc->block_rsv);
3638
3639         memset(&rc->cluster, 0, sizeof(rc->cluster));
3640         rc->search_start = rc->block_group->key.objectid;
3641         rc->extents_found = 0;
3642         rc->nodes_relocated = 0;
3643         rc->merging_rsv_size = 0;
3644
3645         rc->create_reloc_tree = 1;
3646         set_reloc_control(rc);
3647
3648         trans = btrfs_join_transaction(rc->extent_root, 1);
3649         BUG_ON(IS_ERR(trans));
3650         btrfs_commit_transaction(trans, rc->extent_root);
3651         return 0;
3652 }
3653
3654 static noinline_for_stack int relocate_block_group(struct reloc_control *rc)
3655 {
3656         struct rb_root blocks = RB_ROOT;
3657         struct btrfs_key key;
3658         struct btrfs_trans_handle *trans = NULL;
3659         struct btrfs_path *path;
3660         struct btrfs_extent_item *ei;
3661         unsigned long nr;
3662         u64 flags;
3663         u32 item_size;
3664         int ret;
3665         int err = 0;
3666         int progress = 0;
3667
3668         path = btrfs_alloc_path();
3669         if (!path)
3670                 return -ENOMEM;
3671
3672         ret = prepare_to_relocate(rc);
3673         if (ret) {
3674                 err = ret;
3675                 goto out_free;
3676         }
3677
3678         while (1) {
3679                 progress++;
3680                 trans = btrfs_start_transaction(rc->extent_root, 0);
3681                 BUG_ON(IS_ERR(trans));
3682 restart:
3683                 if (update_backref_cache(trans, &rc->backref_cache)) {
3684                         btrfs_end_transaction(trans, rc->extent_root);
3685                         continue;
3686                 }
3687
3688                 ret = find_next_extent(trans, rc, path, &key);
3689                 if (ret < 0)
3690                         err = ret;
3691                 if (ret != 0)
3692                         break;
3693
3694                 rc->extents_found++;
3695
3696                 ei = btrfs_item_ptr(path->nodes[0], path->slots[0],
3697                                     struct btrfs_extent_item);
3698                 item_size = btrfs_item_size_nr(path->nodes[0], path->slots[0]);
3699                 if (item_size >= sizeof(*ei)) {
3700                         flags = btrfs_extent_flags(path->nodes[0], ei);
3701                         ret = check_extent_flags(flags);
3702                         BUG_ON(ret);
3703
3704                 } else {
3705 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
3706                         u64 ref_owner;
3707                         int path_change = 0;
3708
3709                         BUG_ON(item_size !=
3710                                sizeof(struct btrfs_extent_item_v0));
3711                         ret = get_ref_objectid_v0(rc, path, &key, &ref_owner,
3712                                                   &path_change);
3713                         if (ref_owner < BTRFS_FIRST_FREE_OBJECTID)
3714                                 flags = BTRFS_EXTENT_FLAG_TREE_BLOCK;
3715                         else
3716                                 flags = BTRFS_EXTENT_FLAG_DATA;
3717
3718                         if (path_change) {
3719                                 btrfs_release_path(path);
3720
3721                                 path->search_commit_root = 1;
3722                                 path->skip_locking = 1;
3723                                 ret = btrfs_search_slot(NULL, rc->extent_root,
3724                                                         &key, path, 0, 0);
3725                                 if (ret < 0) {
3726                                         err = ret;
3727                                         break;
3728                                 }
3729                                 BUG_ON(ret > 0);
3730                         }
3731 #else
3732                         BUG();
3733 #endif
3734                 }
3735
3736                 if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
3737                         ret = add_tree_block(rc, &key, path, &blocks);
3738                 } else if (rc->stage == UPDATE_DATA_PTRS &&
3739                            (flags & BTRFS_EXTENT_FLAG_DATA)) {
3740                         ret = add_data_references(rc, &key, path, &blocks);
3741                 } else {
3742                         btrfs_release_path(path);
3743                         ret = 0;
3744                 }
3745                 if (ret < 0) {
3746                         err = ret;
3747                         break;
3748                 }
3749
3750                 if (!RB_EMPTY_ROOT(&blocks)) {
3751                         ret = relocate_tree_blocks(trans, rc, &blocks);
3752                         if (ret < 0) {
3753                                 if (ret != -EAGAIN) {
3754                                         err = ret;
3755                                         break;
3756                                 }
3757                                 rc->extents_found--;
3758                                 rc->search_start = key.objectid;
3759                         }
3760                 }
3761
3762                 ret = btrfs_block_rsv_check(trans, rc->extent_root,
3763                                             rc->block_rsv, 0, 5);
3764                 if (ret < 0) {
3765                         if (ret != -EAGAIN) {
3766                                 err = ret;
3767                                 WARN_ON(1);
3768                                 break;
3769                         }
3770                         rc->commit_transaction = 1;
3771                 }
3772
3773                 if (rc->commit_transaction) {
3774                         rc->commit_transaction = 0;
3775                         ret = btrfs_commit_transaction(trans, rc->extent_root);
3776                         BUG_ON(ret);
3777                 } else {
3778                         nr = trans->blocks_used;
3779                         btrfs_end_transaction_throttle(trans, rc->extent_root);
3780                         btrfs_btree_balance_dirty(rc->extent_root, nr);
3781                 }
3782                 trans = NULL;
3783
3784                 if (rc->stage == MOVE_DATA_EXTENTS &&
3785                     (flags & BTRFS_EXTENT_FLAG_DATA)) {
3786                         rc->found_file_extent = 1;
3787                         ret = relocate_data_extent(rc->data_inode,
3788                                                    &key, &rc->cluster);
3789                         if (ret < 0) {
3790                                 err = ret;
3791                                 break;
3792                         }
3793                 }
3794         }
3795         if (trans && progress && err == -ENOSPC) {
3796                 ret = btrfs_force_chunk_alloc(trans, rc->extent_root,
3797                                               rc->block_group->flags);
3798                 if (ret == 0) {
3799                         err = 0;
3800                         progress = 0;
3801                         goto restart;
3802                 }
3803         }
3804
3805         btrfs_release_path(path);
3806         clear_extent_bits(&rc->processed_blocks, 0, (u64)-1, EXTENT_DIRTY,
3807                           GFP_NOFS);
3808
3809         if (trans) {
3810                 nr = trans->blocks_used;
3811                 btrfs_end_transaction_throttle(trans, rc->extent_root);
3812                 btrfs_btree_balance_dirty(rc->extent_root, nr);
3813         }
3814
3815         if (!err) {
3816                 ret = relocate_file_extent_cluster(rc->data_inode,
3817                                                    &rc->cluster);
3818                 if (ret < 0)
3819                         err = ret;
3820         }
3821
3822         rc->create_reloc_tree = 0;
3823         set_reloc_control(rc);
3824
3825         backref_cache_cleanup(&rc->backref_cache);
3826         btrfs_block_rsv_release(rc->extent_root, rc->block_rsv, (u64)-1);
3827
3828         err = prepare_to_merge(rc, err);
3829
3830         merge_reloc_roots(rc);
3831
3832         rc->merge_reloc_tree = 0;
3833         unset_reloc_control(rc);
3834         btrfs_block_rsv_release(rc->extent_root, rc->block_rsv, (u64)-1);
3835
3836         /* get rid of pinned extents */
3837         trans = btrfs_join_transaction(rc->extent_root, 1);
3838         if (IS_ERR(trans))
3839                 err = PTR_ERR(trans);
3840         else
3841                 btrfs_commit_transaction(trans, rc->extent_root);
3842 out_free:
3843         btrfs_free_block_rsv(rc->extent_root, rc->block_rsv);
3844         btrfs_free_path(path);
3845         return err;
3846 }
3847
3848 static int __insert_orphan_inode(struct btrfs_trans_handle *trans,
3849                                  struct btrfs_root *root, u64 objectid)
3850 {
3851         struct btrfs_path *path;
3852         struct btrfs_inode_item *item;
3853         struct extent_buffer *leaf;
3854         int ret;
3855
3856         path = btrfs_alloc_path();
3857         if (!path)
3858                 return -ENOMEM;
3859
3860         ret = btrfs_insert_empty_inode(trans, root, path, objectid);
3861         if (ret)
3862                 goto out;
3863
3864         leaf = path->nodes[0];
3865         item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_inode_item);
3866         memset_extent_buffer(leaf, 0, (unsigned long)item, sizeof(*item));
3867         btrfs_set_inode_generation(leaf, item, 1);
3868         btrfs_set_inode_size(leaf, item, 0);
3869         btrfs_set_inode_mode(leaf, item, S_IFREG | 0600);
3870         btrfs_set_inode_flags(leaf, item, BTRFS_INODE_NOCOMPRESS |
3871                                           BTRFS_INODE_PREALLOC);
3872         btrfs_mark_buffer_dirty(leaf);
3873         btrfs_release_path(path);
3874 out:
3875         btrfs_free_path(path);
3876         return ret;
3877 }
3878
3879 /*
3880  * helper to create inode for data relocation.
3881  * the inode is in data relocation tree and its link count is 0
3882  */
3883 static noinline_for_stack
3884 struct inode *create_reloc_inode(struct btrfs_fs_info *fs_info,
3885                                  struct btrfs_block_group_cache *group)
3886 {
3887         struct inode *inode = NULL;
3888         struct btrfs_trans_handle *trans;
3889         struct btrfs_root *root;
3890         struct btrfs_key key;
3891         unsigned long nr;
3892         u64 objectid = BTRFS_FIRST_FREE_OBJECTID;
3893         int err = 0;
3894
3895         root = read_fs_root(fs_info, BTRFS_DATA_RELOC_TREE_OBJECTID);
3896         if (IS_ERR(root))
3897                 return ERR_CAST(root);
3898
3899         trans = btrfs_start_transaction(root, 6);
3900         if (IS_ERR(trans))
3901                 return ERR_CAST(trans);
3902
3903         err = btrfs_find_free_objectid(root, &objectid);
3904         if (err)
3905                 goto out;
3906
3907         err = __insert_orphan_inode(trans, root, objectid);
3908         BUG_ON(err);
3909
3910         key.objectid = objectid;
3911         key.type = BTRFS_INODE_ITEM_KEY;
3912         key.offset = 0;
3913         inode = btrfs_iget(root->fs_info->sb, &key, root, NULL);
3914         BUG_ON(IS_ERR(inode) || is_bad_inode(inode));
3915         BTRFS_I(inode)->index_cnt = group->key.objectid;
3916
3917         err = btrfs_orphan_add(trans, inode);
3918 out:
3919         nr = trans->blocks_used;
3920         btrfs_end_transaction(trans, root);
3921         btrfs_btree_balance_dirty(root, nr);
3922         if (err) {
3923                 if (inode)
3924                         iput(inode);
3925                 inode = ERR_PTR(err);
3926         }
3927         return inode;
3928 }
3929
3930 static struct reloc_control *alloc_reloc_control(void)
3931 {
3932         struct reloc_control *rc;
3933
3934         rc = kzalloc(sizeof(*rc), GFP_NOFS);
3935         if (!rc)
3936                 return NULL;
3937
3938         INIT_LIST_HEAD(&rc->reloc_roots);
3939         backref_cache_init(&rc->backref_cache);
3940         mapping_tree_init(&rc->reloc_root_tree);
3941         extent_io_tree_init(&rc->processed_blocks, NULL);
3942         return rc;
3943 }
3944
3945 /*
3946  * function to relocate all extents in a block group.
3947  */
3948 int btrfs_relocate_block_group(struct btrfs_root *extent_root, u64 group_start)
3949 {
3950         struct btrfs_fs_info *fs_info = extent_root->fs_info;
3951         struct reloc_control *rc;
3952         struct inode *inode;
3953         struct btrfs_path *path;
3954         int ret;
3955         int rw = 0;
3956         int err = 0;
3957
3958         rc = alloc_reloc_control();
3959         if (!rc)
3960                 return -ENOMEM;
3961
3962         rc->extent_root = extent_root;
3963
3964         rc->block_group = btrfs_lookup_block_group(fs_info, group_start);
3965         BUG_ON(!rc->block_group);
3966
3967         if (!rc->block_group->ro) {
3968                 ret = btrfs_set_block_group_ro(extent_root, rc->block_group);
3969                 if (ret) {
3970                         err = ret;
3971                         goto out;
3972                 }
3973                 rw = 1;
3974         }
3975
3976         path = btrfs_alloc_path();
3977         if (!path) {
3978                 err = -ENOMEM;
3979                 goto out;
3980         }
3981
3982         inode = lookup_free_space_inode(fs_info->tree_root, rc->block_group,
3983                                         path);
3984         btrfs_free_path(path);
3985
3986         if (!IS_ERR(inode))
3987                 ret = delete_block_group_cache(fs_info, inode, 0);
3988         else
3989                 ret = PTR_ERR(inode);
3990
3991         if (ret && ret != -ENOENT) {
3992                 err = ret;
3993                 goto out;
3994         }
3995
3996         rc->data_inode = create_reloc_inode(fs_info, rc->block_group);
3997         if (IS_ERR(rc->data_inode)) {
3998                 err = PTR_ERR(rc->data_inode);
3999                 rc->data_inode = NULL;
4000                 goto out;
4001         }
4002
4003         printk(KERN_INFO "btrfs: relocating block group %llu flags %llu\n",
4004                (unsigned long long)rc->block_group->key.objectid,
4005                (unsigned long long)rc->block_group->flags);
4006
4007         btrfs_start_delalloc_inodes(fs_info->tree_root, 0);
4008         btrfs_wait_ordered_extents(fs_info->tree_root, 0, 0);
4009
4010         while (1) {
4011                 mutex_lock(&fs_info->cleaner_mutex);
4012
4013                 btrfs_clean_old_snapshots(fs_info->tree_root);
4014                 ret = relocate_block_group(rc);
4015
4016                 mutex_unlock(&fs_info->cleaner_mutex);
4017                 if (ret < 0) {
4018                         err = ret;
4019                         goto out;
4020                 }
4021
4022                 if (rc->extents_found == 0)
4023                         break;
4024
4025                 printk(KERN_INFO "btrfs: found %llu extents\n",
4026                         (unsigned long long)rc->extents_found);
4027
4028                 if (rc->stage == MOVE_DATA_EXTENTS && rc->found_file_extent) {
4029                         btrfs_wait_ordered_range(rc->data_inode, 0, (u64)-1);
4030                         invalidate_mapping_pages(rc->data_inode->i_mapping,
4031                                                  0, -1);
4032                         rc->stage = UPDATE_DATA_PTRS;
4033                 }
4034         }
4035
4036         filemap_write_and_wait_range(fs_info->btree_inode->i_mapping,
4037                                      rc->block_group->key.objectid,
4038                                      rc->block_group->key.objectid +
4039                                      rc->block_group->key.offset - 1);
4040
4041         WARN_ON(rc->block_group->pinned > 0);
4042         WARN_ON(rc->block_group->reserved > 0);
4043         WARN_ON(btrfs_block_group_used(&rc->block_group->item) > 0);
4044 out:
4045         if (err && rw)
4046                 btrfs_set_block_group_rw(extent_root, rc->block_group);
4047         iput(rc->data_inode);
4048         btrfs_put_block_group(rc->block_group);
4049         kfree(rc);
4050         return err;
4051 }
4052
4053 static noinline_for_stack int mark_garbage_root(struct btrfs_root *root)
4054 {
4055         struct btrfs_trans_handle *trans;
4056         int ret;
4057
4058         trans = btrfs_start_transaction(root->fs_info->tree_root, 0);
4059         BUG_ON(IS_ERR(trans));
4060
4061         memset(&root->root_item.drop_progress, 0,
4062                 sizeof(root->root_item.drop_progress));
4063         root->root_item.drop_level = 0;
4064         btrfs_set_root_refs(&root->root_item, 0);
4065         ret = btrfs_update_root(trans, root->fs_info->tree_root,
4066                                 &root->root_key, &root->root_item);
4067         BUG_ON(ret);
4068
4069         ret = btrfs_end_transaction(trans, root->fs_info->tree_root);
4070         BUG_ON(ret);
4071         return 0;
4072 }
4073
4074 /*
4075  * recover relocation interrupted by system crash.
4076  *
4077  * this function resumes merging reloc trees with corresponding fs trees.
4078  * this is important for keeping the sharing of tree blocks
4079  */
4080 int btrfs_recover_relocation(struct btrfs_root *root)
4081 {
4082         LIST_HEAD(reloc_roots);
4083         struct btrfs_key key;
4084         struct btrfs_root *fs_root;
4085         struct btrfs_root *reloc_root;
4086         struct btrfs_path *path;
4087         struct extent_buffer *leaf;
4088         struct reloc_control *rc = NULL;
4089         struct btrfs_trans_handle *trans;
4090         int ret;
4091         int err = 0;
4092
4093         path = btrfs_alloc_path();
4094         if (!path)
4095                 return -ENOMEM;
4096
4097         key.objectid = BTRFS_TREE_RELOC_OBJECTID;
4098         key.type = BTRFS_ROOT_ITEM_KEY;
4099         key.offset = (u64)-1;
4100
4101         while (1) {
4102                 ret = btrfs_search_slot(NULL, root->fs_info->tree_root, &key,
4103                                         path, 0, 0);
4104                 if (ret < 0) {
4105                         err = ret;
4106                         goto out;
4107                 }
4108                 if (ret > 0) {
4109                         if (path->slots[0] == 0)
4110                                 break;
4111                         path->slots[0]--;
4112                 }
4113                 leaf = path->nodes[0];
4114                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
4115                 btrfs_release_path(path);
4116
4117                 if (key.objectid != BTRFS_TREE_RELOC_OBJECTID ||
4118                     key.type != BTRFS_ROOT_ITEM_KEY)
4119                         break;
4120
4121                 reloc_root = btrfs_read_fs_root_no_radix(root, &key);
4122                 if (IS_ERR(reloc_root)) {
4123                         err = PTR_ERR(reloc_root);
4124                         goto out;
4125                 }
4126
4127                 list_add(&reloc_root->root_list, &reloc_roots);
4128
4129                 if (btrfs_root_refs(&reloc_root->root_item) > 0) {
4130                         fs_root = read_fs_root(root->fs_info,
4131                                                reloc_root->root_key.offset);
4132                         if (IS_ERR(fs_root)) {
4133                                 ret = PTR_ERR(fs_root);
4134                                 if (ret != -ENOENT) {
4135                                         err = ret;
4136                                         goto out;
4137                                 }
4138                                 mark_garbage_root(reloc_root);
4139                         }
4140                 }
4141
4142                 if (key.offset == 0)
4143                         break;
4144
4145                 key.offset--;
4146         }
4147         btrfs_release_path(path);
4148
4149         if (list_empty(&reloc_roots))
4150                 goto out;
4151
4152         rc = alloc_reloc_control();
4153         if (!rc) {
4154                 err = -ENOMEM;
4155                 goto out;
4156         }
4157
4158         rc->extent_root = root->fs_info->extent_root;
4159
4160         set_reloc_control(rc);
4161
4162         trans = btrfs_join_transaction(rc->extent_root, 1);
4163         if (IS_ERR(trans)) {
4164                 unset_reloc_control(rc);
4165                 err = PTR_ERR(trans);
4166                 goto out_free;
4167         }
4168
4169         rc->merge_reloc_tree = 1;
4170
4171         while (!list_empty(&reloc_roots)) {
4172                 reloc_root = list_entry(reloc_roots.next,
4173                                         struct btrfs_root, root_list);
4174                 list_del(&reloc_root->root_list);
4175
4176                 if (btrfs_root_refs(&reloc_root->root_item) == 0) {
4177                         list_add_tail(&reloc_root->root_list,
4178                                       &rc->reloc_roots);
4179                         continue;
4180                 }
4181
4182                 fs_root = read_fs_root(root->fs_info,
4183                                        reloc_root->root_key.offset);
4184                 BUG_ON(IS_ERR(fs_root));
4185
4186                 __add_reloc_root(reloc_root);
4187                 fs_root->reloc_root = reloc_root;
4188         }
4189
4190         btrfs_commit_transaction(trans, rc->extent_root);
4191
4192         merge_reloc_roots(rc);
4193
4194         unset_reloc_control(rc);
4195
4196         trans = btrfs_join_transaction(rc->extent_root, 1);
4197         if (IS_ERR(trans))
4198                 err = PTR_ERR(trans);
4199         else
4200                 btrfs_commit_transaction(trans, rc->extent_root);
4201 out_free:
4202         kfree(rc);
4203 out:
4204         while (!list_empty(&reloc_roots)) {
4205                 reloc_root = list_entry(reloc_roots.next,
4206                                         struct btrfs_root, root_list);
4207                 list_del(&reloc_root->root_list);
4208                 free_extent_buffer(reloc_root->node);
4209                 free_extent_buffer(reloc_root->commit_root);
4210                 kfree(reloc_root);
4211         }
4212         btrfs_free_path(path);
4213
4214         if (err == 0) {
4215                 /* cleanup orphan inode in data relocation tree */
4216                 fs_root = read_fs_root(root->fs_info,
4217                                        BTRFS_DATA_RELOC_TREE_OBJECTID);
4218                 if (IS_ERR(fs_root))
4219                         err = PTR_ERR(fs_root);
4220                 else
4221                         err = btrfs_orphan_cleanup(fs_root);
4222         }
4223         return err;
4224 }
4225
4226 /*
4227  * helper to add ordered checksum for data relocation.
4228  *
4229  * cloning checksum properly handles the nodatasum extents.
4230  * it also saves CPU time to re-calculate the checksum.
4231  */
4232 int btrfs_reloc_clone_csums(struct inode *inode, u64 file_pos, u64 len)
4233 {
4234         struct btrfs_ordered_sum *sums;
4235         struct btrfs_sector_sum *sector_sum;
4236         struct btrfs_ordered_extent *ordered;
4237         struct btrfs_root *root = BTRFS_I(inode)->root;
4238         size_t offset;
4239         int ret;
4240         u64 disk_bytenr;
4241         LIST_HEAD(list);
4242
4243         ordered = btrfs_lookup_ordered_extent(inode, file_pos);
4244         BUG_ON(ordered->file_offset != file_pos || ordered->len != len);
4245
4246         disk_bytenr = file_pos + BTRFS_I(inode)->index_cnt;
4247         ret = btrfs_lookup_csums_range(root->fs_info->csum_root, disk_bytenr,
4248                                        disk_bytenr + len - 1, &list, 0);
4249
4250         while (!list_empty(&list)) {
4251                 sums = list_entry(list.next, struct btrfs_ordered_sum, list);
4252                 list_del_init(&sums->list);
4253
4254                 sector_sum = sums->sums;
4255                 sums->bytenr = ordered->start;
4256
4257                 offset = 0;
4258                 while (offset < sums->len) {
4259                         sector_sum->bytenr += ordered->start - disk_bytenr;
4260                         sector_sum++;
4261                         offset += root->sectorsize;
4262                 }
4263
4264                 btrfs_add_ordered_sum(inode, ordered, sums);
4265         }
4266         btrfs_put_ordered_extent(ordered);
4267         return ret;
4268 }
4269
4270 void btrfs_reloc_cow_block(struct btrfs_trans_handle *trans,
4271                            struct btrfs_root *root, struct extent_buffer *buf,
4272                            struct extent_buffer *cow)
4273 {
4274         struct reloc_control *rc;
4275         struct backref_node *node;
4276         int first_cow = 0;
4277         int level;
4278         int ret;
4279
4280         rc = root->fs_info->reloc_ctl;
4281         if (!rc)
4282                 return;
4283
4284         BUG_ON(rc->stage == UPDATE_DATA_PTRS &&
4285                root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID);
4286
4287         level = btrfs_header_level(buf);
4288         if (btrfs_header_generation(buf) <=
4289             btrfs_root_last_snapshot(&root->root_item))
4290                 first_cow = 1;
4291
4292         if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID &&
4293             rc->create_reloc_tree) {
4294                 WARN_ON(!first_cow && level == 0);
4295
4296                 node = rc->backref_cache.path[level];
4297                 BUG_ON(node->bytenr != buf->start &&
4298                        node->new_bytenr != buf->start);
4299
4300                 drop_node_buffer(node);
4301                 extent_buffer_get(cow);
4302                 node->eb = cow;
4303                 node->new_bytenr = cow->start;
4304
4305                 if (!node->pending) {
4306                         list_move_tail(&node->list,
4307                                        &rc->backref_cache.pending[level]);
4308                         node->pending = 1;
4309                 }
4310
4311                 if (first_cow)
4312                         __mark_block_processed(rc, node);
4313
4314                 if (first_cow && level > 0)
4315                         rc->nodes_relocated += buf->len;
4316         }
4317
4318         if (level == 0 && first_cow && rc->stage == UPDATE_DATA_PTRS) {
4319                 ret = replace_file_extents(trans, rc, root, cow);
4320                 BUG_ON(ret);
4321         }
4322 }
4323
4324 /*
4325  * called before creating snapshot. it calculates metadata reservation
4326  * requried for relocating tree blocks in the snapshot
4327  */
4328 void btrfs_reloc_pre_snapshot(struct btrfs_trans_handle *trans,
4329                               struct btrfs_pending_snapshot *pending,
4330                               u64 *bytes_to_reserve)
4331 {
4332         struct btrfs_root *root;
4333         struct reloc_control *rc;
4334
4335         root = pending->root;
4336         if (!root->reloc_root)
4337                 return;
4338
4339         rc = root->fs_info->reloc_ctl;
4340         if (!rc->merge_reloc_tree)
4341                 return;
4342
4343         root = root->reloc_root;
4344         BUG_ON(btrfs_root_refs(&root->root_item) == 0);
4345         /*
4346          * relocation is in the stage of merging trees. the space
4347          * used by merging a reloc tree is twice the size of
4348          * relocated tree nodes in the worst case. half for cowing
4349          * the reloc tree, half for cowing the fs tree. the space
4350          * used by cowing the reloc tree will be freed after the
4351          * tree is dropped. if we create snapshot, cowing the fs
4352          * tree may use more space than it frees. so we need
4353          * reserve extra space.
4354          */
4355         *bytes_to_reserve += rc->nodes_relocated;
4356 }
4357
4358 /*
4359  * called after snapshot is created. migrate block reservation
4360  * and create reloc root for the newly created snapshot
4361  */
4362 void btrfs_reloc_post_snapshot(struct btrfs_trans_handle *trans,
4363                                struct btrfs_pending_snapshot *pending)
4364 {
4365         struct btrfs_root *root = pending->root;
4366         struct btrfs_root *reloc_root;
4367         struct btrfs_root *new_root;
4368         struct reloc_control *rc;
4369         int ret;
4370
4371         if (!root->reloc_root)
4372                 return;
4373
4374         rc = root->fs_info->reloc_ctl;
4375         rc->merging_rsv_size += rc->nodes_relocated;
4376
4377         if (rc->merge_reloc_tree) {
4378                 ret = btrfs_block_rsv_migrate(&pending->block_rsv,
4379                                               rc->block_rsv,
4380                                               rc->nodes_relocated);
4381                 BUG_ON(ret);
4382         }
4383
4384         new_root = pending->snap;
4385         reloc_root = create_reloc_root(trans, root->reloc_root,
4386                                        new_root->root_key.objectid);
4387
4388         __add_reloc_root(reloc_root);
4389         new_root->reloc_root = reloc_root;
4390
4391         if (rc->create_reloc_tree) {
4392                 ret = clone_backref_node(trans, rc, root, reloc_root);
4393                 BUG_ON(ret);
4394         }
4395 }