GNU Linux-libre 4.9.308-gnu1
[releases.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 #include "qgroup.h"
35
36 /*
37  * backref_node, mapping_node and tree_block start with this
38  */
39 struct tree_entry {
40         struct rb_node rb_node;
41         u64 bytenr;
42 };
43
44 /*
45  * present a tree block in the backref cache
46  */
47 struct backref_node {
48         struct rb_node rb_node;
49         u64 bytenr;
50
51         u64 new_bytenr;
52         /* objectid of tree block owner, can be not uptodate */
53         u64 owner;
54         /* link to pending, changed or detached list */
55         struct list_head list;
56         /* list of upper level blocks reference this block */
57         struct list_head upper;
58         /* list of child blocks in the cache */
59         struct list_head lower;
60         /* NULL if this node is not tree root */
61         struct btrfs_root *root;
62         /* extent buffer got by COW the block */
63         struct extent_buffer *eb;
64         /* level of tree block */
65         unsigned int level:8;
66         /* is the block in non-reference counted tree */
67         unsigned int cowonly:1;
68         /* 1 if no child node in the cache */
69         unsigned int lowest:1;
70         /* is the extent buffer locked */
71         unsigned int locked:1;
72         /* has the block been processed */
73         unsigned int processed:1;
74         /* have backrefs of this block been checked */
75         unsigned int checked:1;
76         /*
77          * 1 if corresponding block has been cowed but some upper
78          * level block pointers may not point to the new location
79          */
80         unsigned int pending:1;
81         /*
82          * 1 if the backref node isn't connected to any other
83          * backref node.
84          */
85         unsigned int detached:1;
86 };
87
88 /*
89  * present a block pointer in the backref cache
90  */
91 struct backref_edge {
92         struct list_head list[2];
93         struct backref_node *node[2];
94 };
95
96 #define LOWER   0
97 #define UPPER   1
98 #define RELOCATION_RESERVED_NODES       256
99
100 struct backref_cache {
101         /* red black tree of all backref nodes in the cache */
102         struct rb_root rb_root;
103         /* for passing backref nodes to btrfs_reloc_cow_block */
104         struct backref_node *path[BTRFS_MAX_LEVEL];
105         /*
106          * list of blocks that have been cowed but some block
107          * pointers in upper level blocks may not reflect the
108          * new location
109          */
110         struct list_head pending[BTRFS_MAX_LEVEL];
111         /* list of backref nodes with no child node */
112         struct list_head leaves;
113         /* list of blocks that have been cowed in current transaction */
114         struct list_head changed;
115         /* list of detached backref node. */
116         struct list_head detached;
117
118         u64 last_trans;
119
120         int nr_nodes;
121         int nr_edges;
122 };
123
124 /*
125  * map address of tree root to tree
126  */
127 struct mapping_node {
128         struct rb_node rb_node;
129         u64 bytenr;
130         void *data;
131 };
132
133 struct mapping_tree {
134         struct rb_root rb_root;
135         spinlock_t lock;
136 };
137
138 /*
139  * present a tree block to process
140  */
141 struct tree_block {
142         struct rb_node rb_node;
143         u64 bytenr;
144         struct btrfs_key key;
145         unsigned int level:8;
146         unsigned int key_ready:1;
147 };
148
149 #define MAX_EXTENTS 128
150
151 struct file_extent_cluster {
152         u64 start;
153         u64 end;
154         u64 boundary[MAX_EXTENTS];
155         unsigned int nr;
156 };
157
158 struct reloc_control {
159         /* block group to relocate */
160         struct btrfs_block_group_cache *block_group;
161         /* extent tree */
162         struct btrfs_root *extent_root;
163         /* inode for moving data */
164         struct inode *data_inode;
165
166         struct btrfs_block_rsv *block_rsv;
167
168         struct backref_cache backref_cache;
169
170         struct file_extent_cluster cluster;
171         /* tree blocks have been processed */
172         struct extent_io_tree processed_blocks;
173         /* map start of tree root to corresponding reloc tree */
174         struct mapping_tree reloc_root_tree;
175         /* list of reloc trees */
176         struct list_head reloc_roots;
177         /* size of metadata reservation for merging reloc trees */
178         u64 merging_rsv_size;
179         /* size of relocated tree nodes */
180         u64 nodes_relocated;
181         /* reserved size for block group relocation*/
182         u64 reserved_bytes;
183
184         u64 search_start;
185         u64 extents_found;
186
187         unsigned int stage:8;
188         unsigned int create_reloc_tree:1;
189         unsigned int merge_reloc_tree:1;
190         unsigned int found_file_extent:1;
191 };
192
193 /* stages of data relocation */
194 #define MOVE_DATA_EXTENTS       0
195 #define UPDATE_DATA_PTRS        1
196
197 static void remove_backref_node(struct backref_cache *cache,
198                                 struct backref_node *node);
199 static void __mark_block_processed(struct reloc_control *rc,
200                                    struct backref_node *node);
201
202 static void mapping_tree_init(struct mapping_tree *tree)
203 {
204         tree->rb_root = RB_ROOT;
205         spin_lock_init(&tree->lock);
206 }
207
208 static void backref_cache_init(struct backref_cache *cache)
209 {
210         int i;
211         cache->rb_root = RB_ROOT;
212         for (i = 0; i < BTRFS_MAX_LEVEL; i++)
213                 INIT_LIST_HEAD(&cache->pending[i]);
214         INIT_LIST_HEAD(&cache->changed);
215         INIT_LIST_HEAD(&cache->detached);
216         INIT_LIST_HEAD(&cache->leaves);
217 }
218
219 static void backref_cache_cleanup(struct backref_cache *cache)
220 {
221         struct backref_node *node;
222         int i;
223
224         while (!list_empty(&cache->detached)) {
225                 node = list_entry(cache->detached.next,
226                                   struct backref_node, list);
227                 remove_backref_node(cache, node);
228         }
229
230         while (!list_empty(&cache->leaves)) {
231                 node = list_entry(cache->leaves.next,
232                                   struct backref_node, lower);
233                 remove_backref_node(cache, node);
234         }
235
236         cache->last_trans = 0;
237
238         for (i = 0; i < BTRFS_MAX_LEVEL; i++)
239                 ASSERT(list_empty(&cache->pending[i]));
240         ASSERT(list_empty(&cache->changed));
241         ASSERT(list_empty(&cache->detached));
242         ASSERT(RB_EMPTY_ROOT(&cache->rb_root));
243         ASSERT(!cache->nr_nodes);
244         ASSERT(!cache->nr_edges);
245 }
246
247 static struct backref_node *alloc_backref_node(struct backref_cache *cache)
248 {
249         struct backref_node *node;
250
251         node = kzalloc(sizeof(*node), GFP_NOFS);
252         if (node) {
253                 INIT_LIST_HEAD(&node->list);
254                 INIT_LIST_HEAD(&node->upper);
255                 INIT_LIST_HEAD(&node->lower);
256                 RB_CLEAR_NODE(&node->rb_node);
257                 cache->nr_nodes++;
258         }
259         return node;
260 }
261
262 static void free_backref_node(struct backref_cache *cache,
263                               struct backref_node *node)
264 {
265         if (node) {
266                 cache->nr_nodes--;
267                 kfree(node);
268         }
269 }
270
271 static struct backref_edge *alloc_backref_edge(struct backref_cache *cache)
272 {
273         struct backref_edge *edge;
274
275         edge = kzalloc(sizeof(*edge), GFP_NOFS);
276         if (edge)
277                 cache->nr_edges++;
278         return edge;
279 }
280
281 static void free_backref_edge(struct backref_cache *cache,
282                               struct backref_edge *edge)
283 {
284         if (edge) {
285                 cache->nr_edges--;
286                 kfree(edge);
287         }
288 }
289
290 static struct rb_node *tree_insert(struct rb_root *root, u64 bytenr,
291                                    struct rb_node *node)
292 {
293         struct rb_node **p = &root->rb_node;
294         struct rb_node *parent = NULL;
295         struct tree_entry *entry;
296
297         while (*p) {
298                 parent = *p;
299                 entry = rb_entry(parent, struct tree_entry, rb_node);
300
301                 if (bytenr < entry->bytenr)
302                         p = &(*p)->rb_left;
303                 else if (bytenr > entry->bytenr)
304                         p = &(*p)->rb_right;
305                 else
306                         return parent;
307         }
308
309         rb_link_node(node, parent, p);
310         rb_insert_color(node, root);
311         return NULL;
312 }
313
314 static struct rb_node *tree_search(struct rb_root *root, u64 bytenr)
315 {
316         struct rb_node *n = root->rb_node;
317         struct tree_entry *entry;
318
319         while (n) {
320                 entry = rb_entry(n, struct tree_entry, rb_node);
321
322                 if (bytenr < entry->bytenr)
323                         n = n->rb_left;
324                 else if (bytenr > entry->bytenr)
325                         n = n->rb_right;
326                 else
327                         return n;
328         }
329         return NULL;
330 }
331
332 static void backref_tree_panic(struct rb_node *rb_node, int errno, u64 bytenr)
333 {
334
335         struct btrfs_fs_info *fs_info = NULL;
336         struct backref_node *bnode = rb_entry(rb_node, struct backref_node,
337                                               rb_node);
338         if (bnode->root)
339                 fs_info = bnode->root->fs_info;
340         btrfs_panic(fs_info, errno,
341                     "Inconsistency in backref cache found at offset %llu",
342                     bytenr);
343 }
344
345 /*
346  * walk up backref nodes until reach node presents tree root
347  */
348 static struct backref_node *walk_up_backref(struct backref_node *node,
349                                             struct backref_edge *edges[],
350                                             int *index)
351 {
352         struct backref_edge *edge;
353         int idx = *index;
354
355         while (!list_empty(&node->upper)) {
356                 edge = list_entry(node->upper.next,
357                                   struct backref_edge, list[LOWER]);
358                 edges[idx++] = edge;
359                 node = edge->node[UPPER];
360         }
361         BUG_ON(node->detached);
362         *index = idx;
363         return node;
364 }
365
366 /*
367  * walk down backref nodes to find start of next reference path
368  */
369 static struct backref_node *walk_down_backref(struct backref_edge *edges[],
370                                               int *index)
371 {
372         struct backref_edge *edge;
373         struct backref_node *lower;
374         int idx = *index;
375
376         while (idx > 0) {
377                 edge = edges[idx - 1];
378                 lower = edge->node[LOWER];
379                 if (list_is_last(&edge->list[LOWER], &lower->upper)) {
380                         idx--;
381                         continue;
382                 }
383                 edge = list_entry(edge->list[LOWER].next,
384                                   struct backref_edge, list[LOWER]);
385                 edges[idx - 1] = edge;
386                 *index = idx;
387                 return edge->node[UPPER];
388         }
389         *index = 0;
390         return NULL;
391 }
392
393 static void unlock_node_buffer(struct backref_node *node)
394 {
395         if (node->locked) {
396                 btrfs_tree_unlock(node->eb);
397                 node->locked = 0;
398         }
399 }
400
401 static void drop_node_buffer(struct backref_node *node)
402 {
403         if (node->eb) {
404                 unlock_node_buffer(node);
405                 free_extent_buffer(node->eb);
406                 node->eb = NULL;
407         }
408 }
409
410 static void drop_backref_node(struct backref_cache *tree,
411                               struct backref_node *node)
412 {
413         BUG_ON(!list_empty(&node->upper));
414
415         drop_node_buffer(node);
416         list_del(&node->list);
417         list_del(&node->lower);
418         if (!RB_EMPTY_NODE(&node->rb_node))
419                 rb_erase(&node->rb_node, &tree->rb_root);
420         free_backref_node(tree, node);
421 }
422
423 /*
424  * remove a backref node from the backref cache
425  */
426 static void remove_backref_node(struct backref_cache *cache,
427                                 struct backref_node *node)
428 {
429         struct backref_node *upper;
430         struct backref_edge *edge;
431
432         if (!node)
433                 return;
434
435         BUG_ON(!node->lowest && !node->detached);
436         while (!list_empty(&node->upper)) {
437                 edge = list_entry(node->upper.next, struct backref_edge,
438                                   list[LOWER]);
439                 upper = edge->node[UPPER];
440                 list_del(&edge->list[LOWER]);
441                 list_del(&edge->list[UPPER]);
442                 free_backref_edge(cache, edge);
443
444                 if (RB_EMPTY_NODE(&upper->rb_node)) {
445                         BUG_ON(!list_empty(&node->upper));
446                         drop_backref_node(cache, node);
447                         node = upper;
448                         node->lowest = 1;
449                         continue;
450                 }
451                 /*
452                  * add the node to leaf node list if no other
453                  * child block cached.
454                  */
455                 if (list_empty(&upper->lower)) {
456                         list_add_tail(&upper->lower, &cache->leaves);
457                         upper->lowest = 1;
458                 }
459         }
460
461         drop_backref_node(cache, node);
462 }
463
464 static void update_backref_node(struct backref_cache *cache,
465                                 struct backref_node *node, u64 bytenr)
466 {
467         struct rb_node *rb_node;
468         rb_erase(&node->rb_node, &cache->rb_root);
469         node->bytenr = bytenr;
470         rb_node = tree_insert(&cache->rb_root, node->bytenr, &node->rb_node);
471         if (rb_node)
472                 backref_tree_panic(rb_node, -EEXIST, bytenr);
473 }
474
475 /*
476  * update backref cache after a transaction commit
477  */
478 static int update_backref_cache(struct btrfs_trans_handle *trans,
479                                 struct backref_cache *cache)
480 {
481         struct backref_node *node;
482         int level = 0;
483
484         if (cache->last_trans == 0) {
485                 cache->last_trans = trans->transid;
486                 return 0;
487         }
488
489         if (cache->last_trans == trans->transid)
490                 return 0;
491
492         /*
493          * detached nodes are used to avoid unnecessary backref
494          * lookup. transaction commit changes the extent tree.
495          * so the detached nodes are no longer useful.
496          */
497         while (!list_empty(&cache->detached)) {
498                 node = list_entry(cache->detached.next,
499                                   struct backref_node, list);
500                 remove_backref_node(cache, node);
501         }
502
503         while (!list_empty(&cache->changed)) {
504                 node = list_entry(cache->changed.next,
505                                   struct backref_node, list);
506                 list_del_init(&node->list);
507                 BUG_ON(node->pending);
508                 update_backref_node(cache, node, node->new_bytenr);
509         }
510
511         /*
512          * some nodes can be left in the pending list if there were
513          * errors during processing the pending nodes.
514          */
515         for (level = 0; level < BTRFS_MAX_LEVEL; level++) {
516                 list_for_each_entry(node, &cache->pending[level], list) {
517                         BUG_ON(!node->pending);
518                         if (node->bytenr == node->new_bytenr)
519                                 continue;
520                         update_backref_node(cache, node, node->new_bytenr);
521                 }
522         }
523
524         cache->last_trans = 0;
525         return 1;
526 }
527
528
529 static int should_ignore_root(struct btrfs_root *root)
530 {
531         struct btrfs_root *reloc_root;
532
533         if (!test_bit(BTRFS_ROOT_REF_COWS, &root->state))
534                 return 0;
535
536         reloc_root = root->reloc_root;
537         if (!reloc_root)
538                 return 0;
539
540         if (btrfs_header_generation(reloc_root->commit_root) ==
541             root->fs_info->running_transaction->transid)
542                 return 0;
543         /*
544          * if there is reloc tree and it was created in previous
545          * transaction backref lookup can find the reloc tree,
546          * so backref node for the fs tree root is useless for
547          * relocation.
548          */
549         return 1;
550 }
551 /*
552  * find reloc tree by address of tree root
553  */
554 static struct btrfs_root *find_reloc_root(struct reloc_control *rc,
555                                           u64 bytenr)
556 {
557         struct rb_node *rb_node;
558         struct mapping_node *node;
559         struct btrfs_root *root = NULL;
560
561         spin_lock(&rc->reloc_root_tree.lock);
562         rb_node = tree_search(&rc->reloc_root_tree.rb_root, bytenr);
563         if (rb_node) {
564                 node = rb_entry(rb_node, struct mapping_node, rb_node);
565                 root = (struct btrfs_root *)node->data;
566         }
567         spin_unlock(&rc->reloc_root_tree.lock);
568         return root;
569 }
570
571 static int is_cowonly_root(u64 root_objectid)
572 {
573         if (root_objectid == BTRFS_ROOT_TREE_OBJECTID ||
574             root_objectid == BTRFS_EXTENT_TREE_OBJECTID ||
575             root_objectid == BTRFS_CHUNK_TREE_OBJECTID ||
576             root_objectid == BTRFS_DEV_TREE_OBJECTID ||
577             root_objectid == BTRFS_TREE_LOG_OBJECTID ||
578             root_objectid == BTRFS_CSUM_TREE_OBJECTID ||
579             root_objectid == BTRFS_UUID_TREE_OBJECTID ||
580             root_objectid == BTRFS_QUOTA_TREE_OBJECTID ||
581             root_objectid == BTRFS_FREE_SPACE_TREE_OBJECTID)
582                 return 1;
583         return 0;
584 }
585
586 static struct btrfs_root *read_fs_root(struct btrfs_fs_info *fs_info,
587                                         u64 root_objectid)
588 {
589         struct btrfs_key key;
590
591         key.objectid = root_objectid;
592         key.type = BTRFS_ROOT_ITEM_KEY;
593         if (is_cowonly_root(root_objectid))
594                 key.offset = 0;
595         else
596                 key.offset = (u64)-1;
597
598         return btrfs_get_fs_root(fs_info, &key, false);
599 }
600
601 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
602 static noinline_for_stack
603 struct btrfs_root *find_tree_root(struct reloc_control *rc,
604                                   struct extent_buffer *leaf,
605                                   struct btrfs_extent_ref_v0 *ref0)
606 {
607         struct btrfs_root *root;
608         u64 root_objectid = btrfs_ref_root_v0(leaf, ref0);
609         u64 generation = btrfs_ref_generation_v0(leaf, ref0);
610
611         BUG_ON(root_objectid == BTRFS_TREE_RELOC_OBJECTID);
612
613         root = read_fs_root(rc->extent_root->fs_info, root_objectid);
614         BUG_ON(IS_ERR(root));
615
616         if (test_bit(BTRFS_ROOT_REF_COWS, &root->state) &&
617             generation != btrfs_root_generation(&root->root_item))
618                 return NULL;
619
620         return root;
621 }
622 #endif
623
624 static noinline_for_stack
625 int find_inline_backref(struct extent_buffer *leaf, int slot,
626                         unsigned long *ptr, unsigned long *end)
627 {
628         struct btrfs_key key;
629         struct btrfs_extent_item *ei;
630         struct btrfs_tree_block_info *bi;
631         u32 item_size;
632
633         btrfs_item_key_to_cpu(leaf, &key, slot);
634
635         item_size = btrfs_item_size_nr(leaf, slot);
636 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
637         if (item_size < sizeof(*ei)) {
638                 WARN_ON(item_size != sizeof(struct btrfs_extent_item_v0));
639                 return 1;
640         }
641 #endif
642         ei = btrfs_item_ptr(leaf, slot, struct btrfs_extent_item);
643         WARN_ON(!(btrfs_extent_flags(leaf, ei) &
644                   BTRFS_EXTENT_FLAG_TREE_BLOCK));
645
646         if (key.type == BTRFS_EXTENT_ITEM_KEY &&
647             item_size <= sizeof(*ei) + sizeof(*bi)) {
648                 WARN_ON(item_size < sizeof(*ei) + sizeof(*bi));
649                 return 1;
650         }
651         if (key.type == BTRFS_METADATA_ITEM_KEY &&
652             item_size <= sizeof(*ei)) {
653                 WARN_ON(item_size < sizeof(*ei));
654                 return 1;
655         }
656
657         if (key.type == BTRFS_EXTENT_ITEM_KEY) {
658                 bi = (struct btrfs_tree_block_info *)(ei + 1);
659                 *ptr = (unsigned long)(bi + 1);
660         } else {
661                 *ptr = (unsigned long)(ei + 1);
662         }
663         *end = (unsigned long)ei + item_size;
664         return 0;
665 }
666
667 /*
668  * build backref tree for a given tree block. root of the backref tree
669  * corresponds the tree block, leaves of the backref tree correspond
670  * roots of b-trees that reference the tree block.
671  *
672  * the basic idea of this function is check backrefs of a given block
673  * to find upper level blocks that reference the block, and then check
674  * backrefs of these upper level blocks recursively. the recursion stop
675  * when tree root is reached or backrefs for the block is cached.
676  *
677  * NOTE: if we find backrefs for a block are cached, we know backrefs
678  * for all upper level blocks that directly/indirectly reference the
679  * block are also cached.
680  */
681 static noinline_for_stack
682 struct backref_node *build_backref_tree(struct reloc_control *rc,
683                                         struct btrfs_key *node_key,
684                                         int level, u64 bytenr)
685 {
686         struct backref_cache *cache = &rc->backref_cache;
687         struct btrfs_path *path1;
688         struct btrfs_path *path2;
689         struct extent_buffer *eb;
690         struct btrfs_root *root;
691         struct backref_node *cur;
692         struct backref_node *upper;
693         struct backref_node *lower;
694         struct backref_node *node = NULL;
695         struct backref_node *exist = NULL;
696         struct backref_edge *edge;
697         struct rb_node *rb_node;
698         struct btrfs_key key;
699         unsigned long end;
700         unsigned long ptr;
701         LIST_HEAD(list);
702         LIST_HEAD(useless);
703         int cowonly;
704         int ret;
705         int err = 0;
706         bool need_check = true;
707
708         path1 = btrfs_alloc_path();
709         path2 = btrfs_alloc_path();
710         if (!path1 || !path2) {
711                 err = -ENOMEM;
712                 goto out;
713         }
714         path1->reada = READA_FORWARD;
715         path2->reada = READA_FORWARD;
716
717         node = alloc_backref_node(cache);
718         if (!node) {
719                 err = -ENOMEM;
720                 goto out;
721         }
722
723         node->bytenr = bytenr;
724         node->level = level;
725         node->lowest = 1;
726         cur = node;
727 again:
728         end = 0;
729         ptr = 0;
730         key.objectid = cur->bytenr;
731         key.type = BTRFS_METADATA_ITEM_KEY;
732         key.offset = (u64)-1;
733
734         path1->search_commit_root = 1;
735         path1->skip_locking = 1;
736         ret = btrfs_search_slot(NULL, rc->extent_root, &key, path1,
737                                 0, 0);
738         if (ret < 0) {
739                 err = ret;
740                 goto out;
741         }
742         ASSERT(ret);
743         ASSERT(path1->slots[0]);
744
745         path1->slots[0]--;
746
747         WARN_ON(cur->checked);
748         if (!list_empty(&cur->upper)) {
749                 /*
750                  * the backref was added previously when processing
751                  * backref of type BTRFS_TREE_BLOCK_REF_KEY
752                  */
753                 ASSERT(list_is_singular(&cur->upper));
754                 edge = list_entry(cur->upper.next, struct backref_edge,
755                                   list[LOWER]);
756                 ASSERT(list_empty(&edge->list[UPPER]));
757                 exist = edge->node[UPPER];
758                 /*
759                  * add the upper level block to pending list if we need
760                  * check its backrefs
761                  */
762                 if (!exist->checked)
763                         list_add_tail(&edge->list[UPPER], &list);
764         } else {
765                 exist = NULL;
766         }
767
768         while (1) {
769                 cond_resched();
770                 eb = path1->nodes[0];
771
772                 if (ptr >= end) {
773                         if (path1->slots[0] >= btrfs_header_nritems(eb)) {
774                                 ret = btrfs_next_leaf(rc->extent_root, path1);
775                                 if (ret < 0) {
776                                         err = ret;
777                                         goto out;
778                                 }
779                                 if (ret > 0)
780                                         break;
781                                 eb = path1->nodes[0];
782                         }
783
784                         btrfs_item_key_to_cpu(eb, &key, path1->slots[0]);
785                         if (key.objectid != cur->bytenr) {
786                                 WARN_ON(exist);
787                                 break;
788                         }
789
790                         if (key.type == BTRFS_EXTENT_ITEM_KEY ||
791                             key.type == BTRFS_METADATA_ITEM_KEY) {
792                                 ret = find_inline_backref(eb, path1->slots[0],
793                                                           &ptr, &end);
794                                 if (ret)
795                                         goto next;
796                         }
797                 }
798
799                 if (ptr < end) {
800                         /* update key for inline back ref */
801                         struct btrfs_extent_inline_ref *iref;
802                         iref = (struct btrfs_extent_inline_ref *)ptr;
803                         key.type = btrfs_extent_inline_ref_type(eb, iref);
804                         key.offset = btrfs_extent_inline_ref_offset(eb, iref);
805                         WARN_ON(key.type != BTRFS_TREE_BLOCK_REF_KEY &&
806                                 key.type != BTRFS_SHARED_BLOCK_REF_KEY);
807                 }
808
809                 if (exist &&
810                     ((key.type == BTRFS_TREE_BLOCK_REF_KEY &&
811                       exist->owner == key.offset) ||
812                      (key.type == BTRFS_SHARED_BLOCK_REF_KEY &&
813                       exist->bytenr == key.offset))) {
814                         exist = NULL;
815                         goto next;
816                 }
817
818 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
819                 if (key.type == BTRFS_SHARED_BLOCK_REF_KEY ||
820                     key.type == BTRFS_EXTENT_REF_V0_KEY) {
821                         if (key.type == BTRFS_EXTENT_REF_V0_KEY) {
822                                 struct btrfs_extent_ref_v0 *ref0;
823                                 ref0 = btrfs_item_ptr(eb, path1->slots[0],
824                                                 struct btrfs_extent_ref_v0);
825                                 if (key.objectid == key.offset) {
826                                         root = find_tree_root(rc, eb, ref0);
827                                         if (root && !should_ignore_root(root))
828                                                 cur->root = root;
829                                         else
830                                                 list_add(&cur->list, &useless);
831                                         break;
832                                 }
833                                 if (is_cowonly_root(btrfs_ref_root_v0(eb,
834                                                                       ref0)))
835                                         cur->cowonly = 1;
836                         }
837 #else
838                 ASSERT(key.type != BTRFS_EXTENT_REF_V0_KEY);
839                 if (key.type == BTRFS_SHARED_BLOCK_REF_KEY) {
840 #endif
841                         if (key.objectid == key.offset) {
842                                 /*
843                                  * only root blocks of reloc trees use
844                                  * backref of this type.
845                                  */
846                                 root = find_reloc_root(rc, cur->bytenr);
847                                 ASSERT(root);
848                                 cur->root = root;
849                                 break;
850                         }
851
852                         edge = alloc_backref_edge(cache);
853                         if (!edge) {
854                                 err = -ENOMEM;
855                                 goto out;
856                         }
857                         rb_node = tree_search(&cache->rb_root, key.offset);
858                         if (!rb_node) {
859                                 upper = alloc_backref_node(cache);
860                                 if (!upper) {
861                                         free_backref_edge(cache, edge);
862                                         err = -ENOMEM;
863                                         goto out;
864                                 }
865                                 upper->bytenr = key.offset;
866                                 upper->level = cur->level + 1;
867                                 /*
868                                  *  backrefs for the upper level block isn't
869                                  *  cached, add the block to pending list
870                                  */
871                                 list_add_tail(&edge->list[UPPER], &list);
872                         } else {
873                                 upper = rb_entry(rb_node, struct backref_node,
874                                                  rb_node);
875                                 ASSERT(upper->checked);
876                                 INIT_LIST_HEAD(&edge->list[UPPER]);
877                         }
878                         list_add_tail(&edge->list[LOWER], &cur->upper);
879                         edge->node[LOWER] = cur;
880                         edge->node[UPPER] = upper;
881
882                         goto next;
883                 } else if (key.type != BTRFS_TREE_BLOCK_REF_KEY) {
884                         goto next;
885                 }
886
887                 /* key.type == BTRFS_TREE_BLOCK_REF_KEY */
888                 root = read_fs_root(rc->extent_root->fs_info, key.offset);
889                 if (IS_ERR(root)) {
890                         err = PTR_ERR(root);
891                         goto out;
892                 }
893
894                 if (!test_bit(BTRFS_ROOT_REF_COWS, &root->state))
895                         cur->cowonly = 1;
896
897                 if (btrfs_root_level(&root->root_item) == cur->level) {
898                         /* tree root */
899                         ASSERT(btrfs_root_bytenr(&root->root_item) ==
900                                cur->bytenr);
901                         if (should_ignore_root(root))
902                                 list_add(&cur->list, &useless);
903                         else
904                                 cur->root = root;
905                         break;
906                 }
907
908                 level = cur->level + 1;
909
910                 /*
911                  * searching the tree to find upper level blocks
912                  * reference the block.
913                  */
914                 path2->search_commit_root = 1;
915                 path2->skip_locking = 1;
916                 path2->lowest_level = level;
917                 ret = btrfs_search_slot(NULL, root, node_key, path2, 0, 0);
918                 path2->lowest_level = 0;
919                 if (ret < 0) {
920                         err = ret;
921                         goto out;
922                 }
923                 if (ret > 0 && path2->slots[level] > 0)
924                         path2->slots[level]--;
925
926                 eb = path2->nodes[level];
927                 if (btrfs_node_blockptr(eb, path2->slots[level]) !=
928                     cur->bytenr) {
929                         btrfs_err(root->fs_info,
930         "couldn't find block (%llu) (level %d) in tree (%llu) with key (%llu %u %llu)",
931                                   cur->bytenr, level - 1, root->objectid,
932                                   node_key->objectid, node_key->type,
933                                   node_key->offset);
934                         err = -ENOENT;
935                         goto out;
936                 }
937                 lower = cur;
938                 need_check = true;
939                 for (; level < BTRFS_MAX_LEVEL; level++) {
940                         if (!path2->nodes[level]) {
941                                 ASSERT(btrfs_root_bytenr(&root->root_item) ==
942                                        lower->bytenr);
943                                 if (should_ignore_root(root))
944                                         list_add(&lower->list, &useless);
945                                 else
946                                         lower->root = root;
947                                 break;
948                         }
949
950                         edge = alloc_backref_edge(cache);
951                         if (!edge) {
952                                 err = -ENOMEM;
953                                 goto out;
954                         }
955
956                         eb = path2->nodes[level];
957                         rb_node = tree_search(&cache->rb_root, eb->start);
958                         if (!rb_node) {
959                                 upper = alloc_backref_node(cache);
960                                 if (!upper) {
961                                         free_backref_edge(cache, edge);
962                                         err = -ENOMEM;
963                                         goto out;
964                                 }
965                                 upper->bytenr = eb->start;
966                                 upper->owner = btrfs_header_owner(eb);
967                                 upper->level = lower->level + 1;
968                                 if (!test_bit(BTRFS_ROOT_REF_COWS,
969                                               &root->state))
970                                         upper->cowonly = 1;
971
972                                 /*
973                                  * if we know the block isn't shared
974                                  * we can void checking its backrefs.
975                                  */
976                                 if (btrfs_block_can_be_shared(root, eb))
977                                         upper->checked = 0;
978                                 else
979                                         upper->checked = 1;
980
981                                 /*
982                                  * add the block to pending list if we
983                                  * need check its backrefs, we only do this once
984                                  * while walking up a tree as we will catch
985                                  * anything else later on.
986                                  */
987                                 if (!upper->checked && need_check) {
988                                         need_check = false;
989                                         list_add_tail(&edge->list[UPPER],
990                                                       &list);
991                                 } else {
992                                         if (upper->checked)
993                                                 need_check = true;
994                                         INIT_LIST_HEAD(&edge->list[UPPER]);
995                                 }
996                         } else {
997                                 upper = rb_entry(rb_node, struct backref_node,
998                                                  rb_node);
999                                 ASSERT(upper->checked);
1000                                 INIT_LIST_HEAD(&edge->list[UPPER]);
1001                                 if (!upper->owner)
1002                                         upper->owner = btrfs_header_owner(eb);
1003                         }
1004                         list_add_tail(&edge->list[LOWER], &lower->upper);
1005                         edge->node[LOWER] = lower;
1006                         edge->node[UPPER] = upper;
1007
1008                         if (rb_node)
1009                                 break;
1010                         lower = upper;
1011                         upper = NULL;
1012                 }
1013                 btrfs_release_path(path2);
1014 next:
1015                 if (ptr < end) {
1016                         ptr += btrfs_extent_inline_ref_size(key.type);
1017                         if (ptr >= end) {
1018                                 WARN_ON(ptr > end);
1019                                 ptr = 0;
1020                                 end = 0;
1021                         }
1022                 }
1023                 if (ptr >= end)
1024                         path1->slots[0]++;
1025         }
1026         btrfs_release_path(path1);
1027
1028         cur->checked = 1;
1029         WARN_ON(exist);
1030
1031         /* the pending list isn't empty, take the first block to process */
1032         if (!list_empty(&list)) {
1033                 edge = list_entry(list.next, struct backref_edge, list[UPPER]);
1034                 list_del_init(&edge->list[UPPER]);
1035                 cur = edge->node[UPPER];
1036                 goto again;
1037         }
1038
1039         /*
1040          * everything goes well, connect backref nodes and insert backref nodes
1041          * into the cache.
1042          */
1043         ASSERT(node->checked);
1044         cowonly = node->cowonly;
1045         if (!cowonly) {
1046                 rb_node = tree_insert(&cache->rb_root, node->bytenr,
1047                                       &node->rb_node);
1048                 if (rb_node)
1049                         backref_tree_panic(rb_node, -EEXIST, node->bytenr);
1050                 list_add_tail(&node->lower, &cache->leaves);
1051         }
1052
1053         list_for_each_entry(edge, &node->upper, list[LOWER])
1054                 list_add_tail(&edge->list[UPPER], &list);
1055
1056         while (!list_empty(&list)) {
1057                 edge = list_entry(list.next, struct backref_edge, list[UPPER]);
1058                 list_del_init(&edge->list[UPPER]);
1059                 upper = edge->node[UPPER];
1060                 if (upper->detached) {
1061                         list_del(&edge->list[LOWER]);
1062                         lower = edge->node[LOWER];
1063                         free_backref_edge(cache, edge);
1064                         if (list_empty(&lower->upper))
1065                                 list_add(&lower->list, &useless);
1066                         continue;
1067                 }
1068
1069                 if (!RB_EMPTY_NODE(&upper->rb_node)) {
1070                         if (upper->lowest) {
1071                                 list_del_init(&upper->lower);
1072                                 upper->lowest = 0;
1073                         }
1074
1075                         list_add_tail(&edge->list[UPPER], &upper->lower);
1076                         continue;
1077                 }
1078
1079                 if (!upper->checked) {
1080                         /*
1081                          * Still want to blow up for developers since this is a
1082                          * logic bug.
1083                          */
1084                         ASSERT(0);
1085                         err = -EINVAL;
1086                         goto out;
1087                 }
1088                 if (cowonly != upper->cowonly) {
1089                         ASSERT(0);
1090                         err = -EINVAL;
1091                         goto out;
1092                 }
1093
1094                 if (!cowonly) {
1095                         rb_node = tree_insert(&cache->rb_root, upper->bytenr,
1096                                               &upper->rb_node);
1097                         if (rb_node)
1098                                 backref_tree_panic(rb_node, -EEXIST,
1099                                                    upper->bytenr);
1100                 }
1101
1102                 list_add_tail(&edge->list[UPPER], &upper->lower);
1103
1104                 list_for_each_entry(edge, &upper->upper, list[LOWER])
1105                         list_add_tail(&edge->list[UPPER], &list);
1106         }
1107         /*
1108          * process useless backref nodes. backref nodes for tree leaves
1109          * are deleted from the cache. backref nodes for upper level
1110          * tree blocks are left in the cache to avoid unnecessary backref
1111          * lookup.
1112          */
1113         while (!list_empty(&useless)) {
1114                 upper = list_entry(useless.next, struct backref_node, list);
1115                 list_del_init(&upper->list);
1116                 ASSERT(list_empty(&upper->upper));
1117                 if (upper == node)
1118                         node = NULL;
1119                 if (upper->lowest) {
1120                         list_del_init(&upper->lower);
1121                         upper->lowest = 0;
1122                 }
1123                 while (!list_empty(&upper->lower)) {
1124                         edge = list_entry(upper->lower.next,
1125                                           struct backref_edge, list[UPPER]);
1126                         list_del(&edge->list[UPPER]);
1127                         list_del(&edge->list[LOWER]);
1128                         lower = edge->node[LOWER];
1129                         free_backref_edge(cache, edge);
1130
1131                         if (list_empty(&lower->upper))
1132                                 list_add(&lower->list, &useless);
1133                 }
1134                 __mark_block_processed(rc, upper);
1135                 if (upper->level > 0) {
1136                         list_add(&upper->list, &cache->detached);
1137                         upper->detached = 1;
1138                 } else {
1139                         rb_erase(&upper->rb_node, &cache->rb_root);
1140                         free_backref_node(cache, upper);
1141                 }
1142         }
1143 out:
1144         btrfs_free_path(path1);
1145         btrfs_free_path(path2);
1146         if (err) {
1147                 while (!list_empty(&useless)) {
1148                         lower = list_entry(useless.next,
1149                                            struct backref_node, list);
1150                         list_del_init(&lower->list);
1151                 }
1152                 while (!list_empty(&list)) {
1153                         edge = list_first_entry(&list, struct backref_edge,
1154                                                 list[UPPER]);
1155                         list_del(&edge->list[UPPER]);
1156                         list_del(&edge->list[LOWER]);
1157                         lower = edge->node[LOWER];
1158                         upper = edge->node[UPPER];
1159                         free_backref_edge(cache, edge);
1160
1161                         /*
1162                          * Lower is no longer linked to any upper backref nodes
1163                          * and isn't in the cache, we can free it ourselves.
1164                          */
1165                         if (list_empty(&lower->upper) &&
1166                             RB_EMPTY_NODE(&lower->rb_node))
1167                                 list_add(&lower->list, &useless);
1168
1169                         if (!RB_EMPTY_NODE(&upper->rb_node))
1170                                 continue;
1171
1172                         /* Add this guy's upper edges to the list to process */
1173                         list_for_each_entry(edge, &upper->upper, list[LOWER])
1174                                 list_add_tail(&edge->list[UPPER], &list);
1175                         if (list_empty(&upper->upper))
1176                                 list_add(&upper->list, &useless);
1177                 }
1178
1179                 while (!list_empty(&useless)) {
1180                         lower = list_entry(useless.next,
1181                                            struct backref_node, list);
1182                         list_del_init(&lower->list);
1183                         if (lower == node)
1184                                 node = NULL;
1185                         free_backref_node(cache, lower);
1186                 }
1187
1188                 remove_backref_node(cache, node);
1189                 return ERR_PTR(err);
1190         }
1191         ASSERT(!node || !node->detached);
1192         return node;
1193 }
1194
1195 /*
1196  * helper to add backref node for the newly created snapshot.
1197  * the backref node is created by cloning backref node that
1198  * corresponds to root of source tree
1199  */
1200 static int clone_backref_node(struct btrfs_trans_handle *trans,
1201                               struct reloc_control *rc,
1202                               struct btrfs_root *src,
1203                               struct btrfs_root *dest)
1204 {
1205         struct btrfs_root *reloc_root = src->reloc_root;
1206         struct backref_cache *cache = &rc->backref_cache;
1207         struct backref_node *node = NULL;
1208         struct backref_node *new_node;
1209         struct backref_edge *edge;
1210         struct backref_edge *new_edge;
1211         struct rb_node *rb_node;
1212
1213         if (cache->last_trans > 0)
1214                 update_backref_cache(trans, cache);
1215
1216         rb_node = tree_search(&cache->rb_root, src->commit_root->start);
1217         if (rb_node) {
1218                 node = rb_entry(rb_node, struct backref_node, rb_node);
1219                 if (node->detached)
1220                         node = NULL;
1221                 else
1222                         BUG_ON(node->new_bytenr != reloc_root->node->start);
1223         }
1224
1225         if (!node) {
1226                 rb_node = tree_search(&cache->rb_root,
1227                                       reloc_root->commit_root->start);
1228                 if (rb_node) {
1229                         node = rb_entry(rb_node, struct backref_node,
1230                                         rb_node);
1231                         BUG_ON(node->detached);
1232                 }
1233         }
1234
1235         if (!node)
1236                 return 0;
1237
1238         new_node = alloc_backref_node(cache);
1239         if (!new_node)
1240                 return -ENOMEM;
1241
1242         new_node->bytenr = dest->node->start;
1243         new_node->level = node->level;
1244         new_node->lowest = node->lowest;
1245         new_node->checked = 1;
1246         new_node->root = dest;
1247
1248         if (!node->lowest) {
1249                 list_for_each_entry(edge, &node->lower, list[UPPER]) {
1250                         new_edge = alloc_backref_edge(cache);
1251                         if (!new_edge)
1252                                 goto fail;
1253
1254                         new_edge->node[UPPER] = new_node;
1255                         new_edge->node[LOWER] = edge->node[LOWER];
1256                         list_add_tail(&new_edge->list[UPPER],
1257                                       &new_node->lower);
1258                 }
1259         } else {
1260                 list_add_tail(&new_node->lower, &cache->leaves);
1261         }
1262
1263         rb_node = tree_insert(&cache->rb_root, new_node->bytenr,
1264                               &new_node->rb_node);
1265         if (rb_node)
1266                 backref_tree_panic(rb_node, -EEXIST, new_node->bytenr);
1267
1268         if (!new_node->lowest) {
1269                 list_for_each_entry(new_edge, &new_node->lower, list[UPPER]) {
1270                         list_add_tail(&new_edge->list[LOWER],
1271                                       &new_edge->node[LOWER]->upper);
1272                 }
1273         }
1274         return 0;
1275 fail:
1276         while (!list_empty(&new_node->lower)) {
1277                 new_edge = list_entry(new_node->lower.next,
1278                                       struct backref_edge, list[UPPER]);
1279                 list_del(&new_edge->list[UPPER]);
1280                 free_backref_edge(cache, new_edge);
1281         }
1282         free_backref_node(cache, new_node);
1283         return -ENOMEM;
1284 }
1285
1286 /*
1287  * helper to add 'address of tree root -> reloc tree' mapping
1288  */
1289 static int __must_check __add_reloc_root(struct btrfs_root *root)
1290 {
1291         struct rb_node *rb_node;
1292         struct mapping_node *node;
1293         struct reloc_control *rc = root->fs_info->reloc_ctl;
1294
1295         node = kmalloc(sizeof(*node), GFP_NOFS);
1296         if (!node)
1297                 return -ENOMEM;
1298
1299         node->bytenr = root->commit_root->start;
1300         node->data = root;
1301
1302         spin_lock(&rc->reloc_root_tree.lock);
1303         rb_node = tree_insert(&rc->reloc_root_tree.rb_root,
1304                               node->bytenr, &node->rb_node);
1305         spin_unlock(&rc->reloc_root_tree.lock);
1306         if (rb_node) {
1307                 btrfs_panic(root->fs_info, -EEXIST,
1308                             "Duplicate root found for start=%llu while inserting into relocation tree",
1309                             node->bytenr);
1310                 kfree(node);
1311                 return -EEXIST;
1312         }
1313
1314         list_add_tail(&root->root_list, &rc->reloc_roots);
1315         return 0;
1316 }
1317
1318 /*
1319  * helper to delete the 'address of tree root -> reloc tree'
1320  * mapping
1321  */
1322 static void __del_reloc_root(struct btrfs_root *root)
1323 {
1324         struct rb_node *rb_node;
1325         struct mapping_node *node = NULL;
1326         struct reloc_control *rc = root->fs_info->reloc_ctl;
1327
1328         if (rc && root->node) {
1329                 spin_lock(&rc->reloc_root_tree.lock);
1330                 rb_node = tree_search(&rc->reloc_root_tree.rb_root,
1331                                       root->commit_root->start);
1332                 if (rb_node) {
1333                         node = rb_entry(rb_node, struct mapping_node, rb_node);
1334                         rb_erase(&node->rb_node, &rc->reloc_root_tree.rb_root);
1335                         RB_CLEAR_NODE(&node->rb_node);
1336                 }
1337                 spin_unlock(&rc->reloc_root_tree.lock);
1338                 ASSERT(!node || (struct btrfs_root *)node->data == root);
1339         }
1340
1341         spin_lock(&root->fs_info->trans_lock);
1342         list_del_init(&root->root_list);
1343         spin_unlock(&root->fs_info->trans_lock);
1344         kfree(node);
1345 }
1346
1347 /*
1348  * helper to update the 'address of tree root -> reloc tree'
1349  * mapping
1350  */
1351 static int __update_reloc_root(struct btrfs_root *root)
1352 {
1353         struct rb_node *rb_node;
1354         struct mapping_node *node = NULL;
1355         struct reloc_control *rc = root->fs_info->reloc_ctl;
1356
1357         spin_lock(&rc->reloc_root_tree.lock);
1358         rb_node = tree_search(&rc->reloc_root_tree.rb_root,
1359                               root->commit_root->start);
1360         if (rb_node) {
1361                 node = rb_entry(rb_node, struct mapping_node, rb_node);
1362                 rb_erase(&node->rb_node, &rc->reloc_root_tree.rb_root);
1363         }
1364         spin_unlock(&rc->reloc_root_tree.lock);
1365
1366         if (!node)
1367                 return 0;
1368         BUG_ON((struct btrfs_root *)node->data != root);
1369
1370         spin_lock(&rc->reloc_root_tree.lock);
1371         node->bytenr = root->node->start;
1372         rb_node = tree_insert(&rc->reloc_root_tree.rb_root,
1373                               node->bytenr, &node->rb_node);
1374         spin_unlock(&rc->reloc_root_tree.lock);
1375         if (rb_node)
1376                 backref_tree_panic(rb_node, -EEXIST, node->bytenr);
1377         return 0;
1378 }
1379
1380 static struct btrfs_root *create_reloc_root(struct btrfs_trans_handle *trans,
1381                                         struct btrfs_root *root, u64 objectid)
1382 {
1383         struct btrfs_root *reloc_root;
1384         struct extent_buffer *eb;
1385         struct btrfs_root_item *root_item;
1386         struct btrfs_key root_key;
1387         u64 last_snap = 0;
1388         int ret;
1389
1390         root_item = kmalloc(sizeof(*root_item), GFP_NOFS);
1391         BUG_ON(!root_item);
1392
1393         root_key.objectid = BTRFS_TREE_RELOC_OBJECTID;
1394         root_key.type = BTRFS_ROOT_ITEM_KEY;
1395         root_key.offset = objectid;
1396
1397         if (root->root_key.objectid == objectid) {
1398                 u64 commit_root_gen;
1399
1400                 /* called by btrfs_init_reloc_root */
1401                 ret = btrfs_copy_root(trans, root, root->commit_root, &eb,
1402                                       BTRFS_TREE_RELOC_OBJECTID);
1403                 BUG_ON(ret);
1404                 last_snap = btrfs_root_last_snapshot(&root->root_item);
1405                 /*
1406                  * Set the last_snapshot field to the generation of the commit
1407                  * root - like this ctree.c:btrfs_block_can_be_shared() behaves
1408                  * correctly (returns true) when the relocation root is created
1409                  * either inside the critical section of a transaction commit
1410                  * (through transaction.c:qgroup_account_snapshot()) and when
1411                  * it's created before the transaction commit is started.
1412                  */
1413                 commit_root_gen = btrfs_header_generation(root->commit_root);
1414                 btrfs_set_root_last_snapshot(&root->root_item, commit_root_gen);
1415         } else {
1416                 /*
1417                  * called by btrfs_reloc_post_snapshot_hook.
1418                  * the source tree is a reloc tree, all tree blocks
1419                  * modified after it was created have RELOC flag
1420                  * set in their headers. so it's OK to not update
1421                  * the 'last_snapshot'.
1422                  */
1423                 ret = btrfs_copy_root(trans, root, root->node, &eb,
1424                                       BTRFS_TREE_RELOC_OBJECTID);
1425                 BUG_ON(ret);
1426         }
1427
1428         memcpy(root_item, &root->root_item, sizeof(*root_item));
1429         btrfs_set_root_bytenr(root_item, eb->start);
1430         btrfs_set_root_level(root_item, btrfs_header_level(eb));
1431         btrfs_set_root_generation(root_item, trans->transid);
1432
1433         if (root->root_key.objectid == objectid) {
1434                 btrfs_set_root_refs(root_item, 0);
1435                 memset(&root_item->drop_progress, 0,
1436                        sizeof(struct btrfs_disk_key));
1437                 root_item->drop_level = 0;
1438                 /*
1439                  * abuse rtransid, it is safe because it is impossible to
1440                  * receive data into a relocation tree.
1441                  */
1442                 btrfs_set_root_rtransid(root_item, last_snap);
1443                 btrfs_set_root_otransid(root_item, trans->transid);
1444         }
1445
1446         btrfs_tree_unlock(eb);
1447         free_extent_buffer(eb);
1448
1449         ret = btrfs_insert_root(trans, root->fs_info->tree_root,
1450                                 &root_key, root_item);
1451         BUG_ON(ret);
1452         kfree(root_item);
1453
1454         reloc_root = btrfs_read_fs_root(root->fs_info->tree_root, &root_key);
1455         BUG_ON(IS_ERR(reloc_root));
1456         reloc_root->last_trans = trans->transid;
1457         return reloc_root;
1458 }
1459
1460 /*
1461  * create reloc tree for a given fs tree. reloc tree is just a
1462  * snapshot of the fs tree with special root objectid.
1463  */
1464 int btrfs_init_reloc_root(struct btrfs_trans_handle *trans,
1465                           struct btrfs_root *root)
1466 {
1467         struct btrfs_root *reloc_root;
1468         struct reloc_control *rc = root->fs_info->reloc_ctl;
1469         struct btrfs_block_rsv *rsv;
1470         int clear_rsv = 0;
1471         int ret;
1472
1473         if (root->reloc_root) {
1474                 reloc_root = root->reloc_root;
1475                 reloc_root->last_trans = trans->transid;
1476                 return 0;
1477         }
1478
1479         if (!rc || !rc->create_reloc_tree ||
1480             root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID)
1481                 return 0;
1482
1483         if (!trans->reloc_reserved) {
1484                 rsv = trans->block_rsv;
1485                 trans->block_rsv = rc->block_rsv;
1486                 clear_rsv = 1;
1487         }
1488         reloc_root = create_reloc_root(trans, root, root->root_key.objectid);
1489         if (clear_rsv)
1490                 trans->block_rsv = rsv;
1491
1492         ret = __add_reloc_root(reloc_root);
1493         BUG_ON(ret < 0);
1494         root->reloc_root = reloc_root;
1495         return 0;
1496 }
1497
1498 /*
1499  * update root item of reloc tree
1500  */
1501 int btrfs_update_reloc_root(struct btrfs_trans_handle *trans,
1502                             struct btrfs_root *root)
1503 {
1504         struct btrfs_root *reloc_root;
1505         struct btrfs_root_item *root_item;
1506         int ret;
1507
1508         if (!root->reloc_root)
1509                 goto out;
1510
1511         reloc_root = root->reloc_root;
1512         root_item = &reloc_root->root_item;
1513
1514         if (root->fs_info->reloc_ctl->merge_reloc_tree &&
1515             btrfs_root_refs(root_item) == 0) {
1516                 root->reloc_root = NULL;
1517                 __del_reloc_root(reloc_root);
1518         }
1519
1520         if (reloc_root->commit_root != reloc_root->node) {
1521                 __update_reloc_root(reloc_root);
1522                 btrfs_set_root_node(root_item, reloc_root->node);
1523                 free_extent_buffer(reloc_root->commit_root);
1524                 reloc_root->commit_root = btrfs_root_node(reloc_root);
1525         }
1526
1527         ret = btrfs_update_root(trans, root->fs_info->tree_root,
1528                                 &reloc_root->root_key, root_item);
1529         BUG_ON(ret);
1530
1531 out:
1532         return 0;
1533 }
1534
1535 /*
1536  * helper to find first cached inode with inode number >= objectid
1537  * in a subvolume
1538  */
1539 static struct inode *find_next_inode(struct btrfs_root *root, u64 objectid)
1540 {
1541         struct rb_node *node;
1542         struct rb_node *prev;
1543         struct btrfs_inode *entry;
1544         struct inode *inode;
1545
1546         spin_lock(&root->inode_lock);
1547 again:
1548         node = root->inode_tree.rb_node;
1549         prev = NULL;
1550         while (node) {
1551                 prev = node;
1552                 entry = rb_entry(node, struct btrfs_inode, rb_node);
1553
1554                 if (objectid < btrfs_ino(&entry->vfs_inode))
1555                         node = node->rb_left;
1556                 else if (objectid > btrfs_ino(&entry->vfs_inode))
1557                         node = node->rb_right;
1558                 else
1559                         break;
1560         }
1561         if (!node) {
1562                 while (prev) {
1563                         entry = rb_entry(prev, struct btrfs_inode, rb_node);
1564                         if (objectid <= btrfs_ino(&entry->vfs_inode)) {
1565                                 node = prev;
1566                                 break;
1567                         }
1568                         prev = rb_next(prev);
1569                 }
1570         }
1571         while (node) {
1572                 entry = rb_entry(node, struct btrfs_inode, rb_node);
1573                 inode = igrab(&entry->vfs_inode);
1574                 if (inode) {
1575                         spin_unlock(&root->inode_lock);
1576                         return inode;
1577                 }
1578
1579                 objectid = btrfs_ino(&entry->vfs_inode) + 1;
1580                 if (cond_resched_lock(&root->inode_lock))
1581                         goto again;
1582
1583                 node = rb_next(node);
1584         }
1585         spin_unlock(&root->inode_lock);
1586         return NULL;
1587 }
1588
1589 static int in_block_group(u64 bytenr,
1590                           struct btrfs_block_group_cache *block_group)
1591 {
1592         if (bytenr >= block_group->key.objectid &&
1593             bytenr < block_group->key.objectid + block_group->key.offset)
1594                 return 1;
1595         return 0;
1596 }
1597
1598 /*
1599  * get new location of data
1600  */
1601 static int get_new_location(struct inode *reloc_inode, u64 *new_bytenr,
1602                             u64 bytenr, u64 num_bytes)
1603 {
1604         struct btrfs_root *root = BTRFS_I(reloc_inode)->root;
1605         struct btrfs_path *path;
1606         struct btrfs_file_extent_item *fi;
1607         struct extent_buffer *leaf;
1608         int ret;
1609
1610         path = btrfs_alloc_path();
1611         if (!path)
1612                 return -ENOMEM;
1613
1614         bytenr -= BTRFS_I(reloc_inode)->index_cnt;
1615         ret = btrfs_lookup_file_extent(NULL, root, path, btrfs_ino(reloc_inode),
1616                                        bytenr, 0);
1617         if (ret < 0)
1618                 goto out;
1619         if (ret > 0) {
1620                 ret = -ENOENT;
1621                 goto out;
1622         }
1623
1624         leaf = path->nodes[0];
1625         fi = btrfs_item_ptr(leaf, path->slots[0],
1626                             struct btrfs_file_extent_item);
1627
1628         BUG_ON(btrfs_file_extent_offset(leaf, fi) ||
1629                btrfs_file_extent_compression(leaf, fi) ||
1630                btrfs_file_extent_encryption(leaf, fi) ||
1631                btrfs_file_extent_other_encoding(leaf, fi));
1632
1633         if (num_bytes != btrfs_file_extent_disk_num_bytes(leaf, fi)) {
1634                 ret = -EINVAL;
1635                 goto out;
1636         }
1637
1638         *new_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
1639         ret = 0;
1640 out:
1641         btrfs_free_path(path);
1642         return ret;
1643 }
1644
1645 /*
1646  * update file extent items in the tree leaf to point to
1647  * the new locations.
1648  */
1649 static noinline_for_stack
1650 int replace_file_extents(struct btrfs_trans_handle *trans,
1651                          struct reloc_control *rc,
1652                          struct btrfs_root *root,
1653                          struct extent_buffer *leaf)
1654 {
1655         struct btrfs_key key;
1656         struct btrfs_file_extent_item *fi;
1657         struct inode *inode = NULL;
1658         u64 parent;
1659         u64 bytenr;
1660         u64 new_bytenr = 0;
1661         u64 num_bytes;
1662         u64 end;
1663         u32 nritems;
1664         u32 i;
1665         int ret = 0;
1666         int first = 1;
1667         int dirty = 0;
1668
1669         if (rc->stage != UPDATE_DATA_PTRS)
1670                 return 0;
1671
1672         /* reloc trees always use full backref */
1673         if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID)
1674                 parent = leaf->start;
1675         else
1676                 parent = 0;
1677
1678         nritems = btrfs_header_nritems(leaf);
1679         for (i = 0; i < nritems; i++) {
1680                 cond_resched();
1681                 btrfs_item_key_to_cpu(leaf, &key, i);
1682                 if (key.type != BTRFS_EXTENT_DATA_KEY)
1683                         continue;
1684                 fi = btrfs_item_ptr(leaf, i, struct btrfs_file_extent_item);
1685                 if (btrfs_file_extent_type(leaf, fi) ==
1686                     BTRFS_FILE_EXTENT_INLINE)
1687                         continue;
1688                 bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
1689                 num_bytes = btrfs_file_extent_disk_num_bytes(leaf, fi);
1690                 if (bytenr == 0)
1691                         continue;
1692                 if (!in_block_group(bytenr, rc->block_group))
1693                         continue;
1694
1695                 /*
1696                  * if we are modifying block in fs tree, wait for readpage
1697                  * to complete and drop the extent cache
1698                  */
1699                 if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) {
1700                         if (first) {
1701                                 inode = find_next_inode(root, key.objectid);
1702                                 first = 0;
1703                         } else if (inode && btrfs_ino(inode) < key.objectid) {
1704                                 btrfs_add_delayed_iput(inode);
1705                                 inode = find_next_inode(root, key.objectid);
1706                         }
1707                         if (inode && btrfs_ino(inode) == key.objectid) {
1708                                 end = key.offset +
1709                                       btrfs_file_extent_num_bytes(leaf, fi);
1710                                 WARN_ON(!IS_ALIGNED(key.offset,
1711                                                     root->sectorsize));
1712                                 WARN_ON(!IS_ALIGNED(end, root->sectorsize));
1713                                 end--;
1714                                 ret = try_lock_extent(&BTRFS_I(inode)->io_tree,
1715                                                       key.offset, end);
1716                                 if (!ret)
1717                                         continue;
1718
1719                                 btrfs_drop_extent_cache(inode, key.offset, end,
1720                                                         1);
1721                                 unlock_extent(&BTRFS_I(inode)->io_tree,
1722                                               key.offset, end);
1723                         }
1724                 }
1725
1726                 ret = get_new_location(rc->data_inode, &new_bytenr,
1727                                        bytenr, num_bytes);
1728                 if (ret) {
1729                         /*
1730                          * Don't have to abort since we've not changed anything
1731                          * in the file extent yet.
1732                          */
1733                         break;
1734                 }
1735
1736                 btrfs_set_file_extent_disk_bytenr(leaf, fi, new_bytenr);
1737                 dirty = 1;
1738
1739                 key.offset -= btrfs_file_extent_offset(leaf, fi);
1740                 ret = btrfs_inc_extent_ref(trans, root, new_bytenr,
1741                                            num_bytes, parent,
1742                                            btrfs_header_owner(leaf),
1743                                            key.objectid, key.offset);
1744                 if (ret) {
1745                         btrfs_abort_transaction(trans, ret);
1746                         break;
1747                 }
1748
1749                 ret = btrfs_free_extent(trans, root, bytenr, num_bytes,
1750                                         parent, btrfs_header_owner(leaf),
1751                                         key.objectid, key.offset);
1752                 if (ret) {
1753                         btrfs_abort_transaction(trans, ret);
1754                         break;
1755                 }
1756         }
1757         if (dirty)
1758                 btrfs_mark_buffer_dirty(leaf);
1759         if (inode)
1760                 btrfs_add_delayed_iput(inode);
1761         return ret;
1762 }
1763
1764 static noinline_for_stack
1765 int memcmp_node_keys(struct extent_buffer *eb, int slot,
1766                      struct btrfs_path *path, int level)
1767 {
1768         struct btrfs_disk_key key1;
1769         struct btrfs_disk_key key2;
1770         btrfs_node_key(eb, &key1, slot);
1771         btrfs_node_key(path->nodes[level], &key2, path->slots[level]);
1772         return memcmp(&key1, &key2, sizeof(key1));
1773 }
1774
1775 /*
1776  * try to replace tree blocks in fs tree with the new blocks
1777  * in reloc tree. tree blocks haven't been modified since the
1778  * reloc tree was create can be replaced.
1779  *
1780  * if a block was replaced, level of the block + 1 is returned.
1781  * if no block got replaced, 0 is returned. if there are other
1782  * errors, a negative error number is returned.
1783  */
1784 static noinline_for_stack
1785 int replace_path(struct btrfs_trans_handle *trans,
1786                  struct btrfs_root *dest, struct btrfs_root *src,
1787                  struct btrfs_path *path, struct btrfs_key *next_key,
1788                  int lowest_level, int max_level)
1789 {
1790         struct extent_buffer *eb;
1791         struct extent_buffer *parent;
1792         struct btrfs_key key;
1793         u64 old_bytenr;
1794         u64 new_bytenr;
1795         u64 old_ptr_gen;
1796         u64 new_ptr_gen;
1797         u64 last_snapshot;
1798         u32 blocksize;
1799         int cow = 0;
1800         int level;
1801         int ret;
1802         int slot;
1803
1804         ASSERT(src->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID);
1805         ASSERT(dest->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID);
1806
1807         last_snapshot = btrfs_root_last_snapshot(&src->root_item);
1808 again:
1809         slot = path->slots[lowest_level];
1810         btrfs_node_key_to_cpu(path->nodes[lowest_level], &key, slot);
1811
1812         eb = btrfs_lock_root_node(dest);
1813         btrfs_set_lock_blocking(eb);
1814         level = btrfs_header_level(eb);
1815
1816         if (level < lowest_level) {
1817                 btrfs_tree_unlock(eb);
1818                 free_extent_buffer(eb);
1819                 return 0;
1820         }
1821
1822         if (cow) {
1823                 ret = btrfs_cow_block(trans, dest, eb, NULL, 0, &eb);
1824                 BUG_ON(ret);
1825         }
1826         btrfs_set_lock_blocking(eb);
1827
1828         if (next_key) {
1829                 next_key->objectid = (u64)-1;
1830                 next_key->type = (u8)-1;
1831                 next_key->offset = (u64)-1;
1832         }
1833
1834         parent = eb;
1835         while (1) {
1836                 level = btrfs_header_level(parent);
1837                 ASSERT(level >= lowest_level);
1838
1839                 ret = btrfs_bin_search(parent, &key, level, &slot);
1840                 if (ret && slot > 0)
1841                         slot--;
1842
1843                 if (next_key && slot + 1 < btrfs_header_nritems(parent))
1844                         btrfs_node_key_to_cpu(parent, next_key, slot + 1);
1845
1846                 old_bytenr = btrfs_node_blockptr(parent, slot);
1847                 blocksize = dest->nodesize;
1848                 old_ptr_gen = btrfs_node_ptr_generation(parent, slot);
1849
1850                 if (level <= max_level) {
1851                         eb = path->nodes[level];
1852                         new_bytenr = btrfs_node_blockptr(eb,
1853                                                         path->slots[level]);
1854                         new_ptr_gen = btrfs_node_ptr_generation(eb,
1855                                                         path->slots[level]);
1856                 } else {
1857                         new_bytenr = 0;
1858                         new_ptr_gen = 0;
1859                 }
1860
1861                 if (WARN_ON(new_bytenr > 0 && new_bytenr == old_bytenr)) {
1862                         ret = level;
1863                         break;
1864                 }
1865
1866                 if (new_bytenr == 0 || old_ptr_gen > last_snapshot ||
1867                     memcmp_node_keys(parent, slot, path, level)) {
1868                         if (level <= lowest_level) {
1869                                 ret = 0;
1870                                 break;
1871                         }
1872
1873                         eb = read_tree_block(dest, old_bytenr, old_ptr_gen);
1874                         if (IS_ERR(eb)) {
1875                                 ret = PTR_ERR(eb);
1876                                 break;
1877                         } else if (!extent_buffer_uptodate(eb)) {
1878                                 ret = -EIO;
1879                                 free_extent_buffer(eb);
1880                                 break;
1881                         }
1882                         btrfs_tree_lock(eb);
1883                         if (cow) {
1884                                 ret = btrfs_cow_block(trans, dest, eb, parent,
1885                                                       slot, &eb);
1886                                 BUG_ON(ret);
1887                         }
1888                         btrfs_set_lock_blocking(eb);
1889
1890                         btrfs_tree_unlock(parent);
1891                         free_extent_buffer(parent);
1892
1893                         parent = eb;
1894                         continue;
1895                 }
1896
1897                 if (!cow) {
1898                         btrfs_tree_unlock(parent);
1899                         free_extent_buffer(parent);
1900                         cow = 1;
1901                         goto again;
1902                 }
1903
1904                 btrfs_node_key_to_cpu(path->nodes[level], &key,
1905                                       path->slots[level]);
1906                 btrfs_release_path(path);
1907
1908                 path->lowest_level = level;
1909                 ret = btrfs_search_slot(trans, src, &key, path, 0, 1);
1910                 path->lowest_level = 0;
1911                 BUG_ON(ret);
1912
1913                 /*
1914                  * swap blocks in fs tree and reloc tree.
1915                  */
1916                 btrfs_set_node_blockptr(parent, slot, new_bytenr);
1917                 btrfs_set_node_ptr_generation(parent, slot, new_ptr_gen);
1918                 btrfs_mark_buffer_dirty(parent);
1919
1920                 btrfs_set_node_blockptr(path->nodes[level],
1921                                         path->slots[level], old_bytenr);
1922                 btrfs_set_node_ptr_generation(path->nodes[level],
1923                                               path->slots[level], old_ptr_gen);
1924                 btrfs_mark_buffer_dirty(path->nodes[level]);
1925
1926                 ret = btrfs_inc_extent_ref(trans, src, old_bytenr, blocksize,
1927                                         path->nodes[level]->start,
1928                                         src->root_key.objectid, level - 1, 0);
1929                 BUG_ON(ret);
1930                 ret = btrfs_inc_extent_ref(trans, dest, new_bytenr, blocksize,
1931                                         0, dest->root_key.objectid, level - 1,
1932                                         0);
1933                 BUG_ON(ret);
1934
1935                 ret = btrfs_free_extent(trans, src, new_bytenr, blocksize,
1936                                         path->nodes[level]->start,
1937                                         src->root_key.objectid, level - 1, 0);
1938                 BUG_ON(ret);
1939
1940                 ret = btrfs_free_extent(trans, dest, old_bytenr, blocksize,
1941                                         0, dest->root_key.objectid, level - 1,
1942                                         0);
1943                 BUG_ON(ret);
1944
1945                 btrfs_unlock_up_safe(path, 0);
1946
1947                 ret = level;
1948                 break;
1949         }
1950         btrfs_tree_unlock(parent);
1951         free_extent_buffer(parent);
1952         return ret;
1953 }
1954
1955 /*
1956  * helper to find next relocated block in reloc tree
1957  */
1958 static noinline_for_stack
1959 int walk_up_reloc_tree(struct btrfs_root *root, struct btrfs_path *path,
1960                        int *level)
1961 {
1962         struct extent_buffer *eb;
1963         int i;
1964         u64 last_snapshot;
1965         u32 nritems;
1966
1967         last_snapshot = btrfs_root_last_snapshot(&root->root_item);
1968
1969         for (i = 0; i < *level; i++) {
1970                 free_extent_buffer(path->nodes[i]);
1971                 path->nodes[i] = NULL;
1972         }
1973
1974         for (i = *level; i < BTRFS_MAX_LEVEL && path->nodes[i]; i++) {
1975                 eb = path->nodes[i];
1976                 nritems = btrfs_header_nritems(eb);
1977                 while (path->slots[i] + 1 < nritems) {
1978                         path->slots[i]++;
1979                         if (btrfs_node_ptr_generation(eb, path->slots[i]) <=
1980                             last_snapshot)
1981                                 continue;
1982
1983                         *level = i;
1984                         return 0;
1985                 }
1986                 free_extent_buffer(path->nodes[i]);
1987                 path->nodes[i] = NULL;
1988         }
1989         return 1;
1990 }
1991
1992 /*
1993  * walk down reloc tree to find relocated block of lowest level
1994  */
1995 static noinline_for_stack
1996 int walk_down_reloc_tree(struct btrfs_root *root, struct btrfs_path *path,
1997                          int *level)
1998 {
1999         struct extent_buffer *eb = NULL;
2000         int i;
2001         u64 bytenr;
2002         u64 ptr_gen = 0;
2003         u64 last_snapshot;
2004         u32 nritems;
2005
2006         last_snapshot = btrfs_root_last_snapshot(&root->root_item);
2007
2008         for (i = *level; i > 0; i--) {
2009                 eb = path->nodes[i];
2010                 nritems = btrfs_header_nritems(eb);
2011                 while (path->slots[i] < nritems) {
2012                         ptr_gen = btrfs_node_ptr_generation(eb, path->slots[i]);
2013                         if (ptr_gen > last_snapshot)
2014                                 break;
2015                         path->slots[i]++;
2016                 }
2017                 if (path->slots[i] >= nritems) {
2018                         if (i == *level)
2019                                 break;
2020                         *level = i + 1;
2021                         return 0;
2022                 }
2023                 if (i == 1) {
2024                         *level = i;
2025                         return 0;
2026                 }
2027
2028                 bytenr = btrfs_node_blockptr(eb, path->slots[i]);
2029                 eb = read_tree_block(root, bytenr, ptr_gen);
2030                 if (IS_ERR(eb)) {
2031                         return PTR_ERR(eb);
2032                 } else if (!extent_buffer_uptodate(eb)) {
2033                         free_extent_buffer(eb);
2034                         return -EIO;
2035                 }
2036                 BUG_ON(btrfs_header_level(eb) != i - 1);
2037                 path->nodes[i - 1] = eb;
2038                 path->slots[i - 1] = 0;
2039         }
2040         return 1;
2041 }
2042
2043 /*
2044  * invalidate extent cache for file extents whose key in range of
2045  * [min_key, max_key)
2046  */
2047 static int invalidate_extent_cache(struct btrfs_root *root,
2048                                    struct btrfs_key *min_key,
2049                                    struct btrfs_key *max_key)
2050 {
2051         struct inode *inode = NULL;
2052         u64 objectid;
2053         u64 start, end;
2054         u64 ino;
2055
2056         objectid = min_key->objectid;
2057         while (1) {
2058                 cond_resched();
2059                 iput(inode);
2060
2061                 if (objectid > max_key->objectid)
2062                         break;
2063
2064                 inode = find_next_inode(root, objectid);
2065                 if (!inode)
2066                         break;
2067                 ino = btrfs_ino(inode);
2068
2069                 if (ino > max_key->objectid) {
2070                         iput(inode);
2071                         break;
2072                 }
2073
2074                 objectid = ino + 1;
2075                 if (!S_ISREG(inode->i_mode))
2076                         continue;
2077
2078                 if (unlikely(min_key->objectid == ino)) {
2079                         if (min_key->type > BTRFS_EXTENT_DATA_KEY)
2080                                 continue;
2081                         if (min_key->type < BTRFS_EXTENT_DATA_KEY)
2082                                 start = 0;
2083                         else {
2084                                 start = min_key->offset;
2085                                 WARN_ON(!IS_ALIGNED(start, root->sectorsize));
2086                         }
2087                 } else {
2088                         start = 0;
2089                 }
2090
2091                 if (unlikely(max_key->objectid == ino)) {
2092                         if (max_key->type < BTRFS_EXTENT_DATA_KEY)
2093                                 continue;
2094                         if (max_key->type > BTRFS_EXTENT_DATA_KEY) {
2095                                 end = (u64)-1;
2096                         } else {
2097                                 if (max_key->offset == 0)
2098                                         continue;
2099                                 end = max_key->offset;
2100                                 WARN_ON(!IS_ALIGNED(end, root->sectorsize));
2101                                 end--;
2102                         }
2103                 } else {
2104                         end = (u64)-1;
2105                 }
2106
2107                 /* the lock_extent waits for readpage to complete */
2108                 lock_extent(&BTRFS_I(inode)->io_tree, start, end);
2109                 btrfs_drop_extent_cache(inode, start, end, 1);
2110                 unlock_extent(&BTRFS_I(inode)->io_tree, start, end);
2111         }
2112         return 0;
2113 }
2114
2115 static int find_next_key(struct btrfs_path *path, int level,
2116                          struct btrfs_key *key)
2117
2118 {
2119         while (level < BTRFS_MAX_LEVEL) {
2120                 if (!path->nodes[level])
2121                         break;
2122                 if (path->slots[level] + 1 <
2123                     btrfs_header_nritems(path->nodes[level])) {
2124                         btrfs_node_key_to_cpu(path->nodes[level], key,
2125                                               path->slots[level] + 1);
2126                         return 0;
2127                 }
2128                 level++;
2129         }
2130         return 1;
2131 }
2132
2133 /*
2134  * merge the relocated tree blocks in reloc tree with corresponding
2135  * fs tree.
2136  */
2137 static noinline_for_stack int merge_reloc_root(struct reloc_control *rc,
2138                                                struct btrfs_root *root)
2139 {
2140         LIST_HEAD(inode_list);
2141         struct btrfs_key key;
2142         struct btrfs_key next_key;
2143         struct btrfs_trans_handle *trans = NULL;
2144         struct btrfs_root *reloc_root;
2145         struct btrfs_root_item *root_item;
2146         struct btrfs_path *path;
2147         struct extent_buffer *leaf;
2148         int level;
2149         int max_level;
2150         int replaced = 0;
2151         int ret;
2152         int err = 0;
2153         u32 min_reserved;
2154
2155         path = btrfs_alloc_path();
2156         if (!path)
2157                 return -ENOMEM;
2158         path->reada = READA_FORWARD;
2159
2160         reloc_root = root->reloc_root;
2161         root_item = &reloc_root->root_item;
2162
2163         if (btrfs_disk_key_objectid(&root_item->drop_progress) == 0) {
2164                 level = btrfs_root_level(root_item);
2165                 extent_buffer_get(reloc_root->node);
2166                 path->nodes[level] = reloc_root->node;
2167                 path->slots[level] = 0;
2168         } else {
2169                 btrfs_disk_key_to_cpu(&key, &root_item->drop_progress);
2170
2171                 level = root_item->drop_level;
2172                 BUG_ON(level == 0);
2173                 path->lowest_level = level;
2174                 ret = btrfs_search_slot(NULL, reloc_root, &key, path, 0, 0);
2175                 path->lowest_level = 0;
2176                 if (ret < 0) {
2177                         btrfs_free_path(path);
2178                         return ret;
2179                 }
2180
2181                 btrfs_node_key_to_cpu(path->nodes[level], &next_key,
2182                                       path->slots[level]);
2183                 WARN_ON(memcmp(&key, &next_key, sizeof(key)));
2184
2185                 btrfs_unlock_up_safe(path, 0);
2186         }
2187
2188         min_reserved = root->nodesize * (BTRFS_MAX_LEVEL - 1) * 2;
2189         memset(&next_key, 0, sizeof(next_key));
2190
2191         while (1) {
2192                 ret = btrfs_block_rsv_refill(root, rc->block_rsv, min_reserved,
2193                                              BTRFS_RESERVE_FLUSH_ALL);
2194                 if (ret) {
2195                         err = ret;
2196                         goto out;
2197                 }
2198                 trans = btrfs_start_transaction(root, 0);
2199                 if (IS_ERR(trans)) {
2200                         err = PTR_ERR(trans);
2201                         trans = NULL;
2202                         goto out;
2203                 }
2204                 trans->block_rsv = rc->block_rsv;
2205
2206                 replaced = 0;
2207                 max_level = level;
2208
2209                 ret = walk_down_reloc_tree(reloc_root, path, &level);
2210                 if (ret < 0) {
2211                         err = ret;
2212                         goto out;
2213                 }
2214                 if (ret > 0)
2215                         break;
2216
2217                 if (!find_next_key(path, level, &key) &&
2218                     btrfs_comp_cpu_keys(&next_key, &key) >= 0) {
2219                         ret = 0;
2220                 } else {
2221                         ret = replace_path(trans, root, reloc_root, path,
2222                                            &next_key, level, max_level);
2223                 }
2224                 if (ret < 0) {
2225                         err = ret;
2226                         goto out;
2227                 }
2228
2229                 if (ret > 0) {
2230                         level = ret;
2231                         btrfs_node_key_to_cpu(path->nodes[level], &key,
2232                                               path->slots[level]);
2233                         replaced = 1;
2234                 }
2235
2236                 ret = walk_up_reloc_tree(reloc_root, path, &level);
2237                 if (ret > 0)
2238                         break;
2239
2240                 BUG_ON(level == 0);
2241                 /*
2242                  * save the merging progress in the drop_progress.
2243                  * this is OK since root refs == 1 in this case.
2244                  */
2245                 btrfs_node_key(path->nodes[level], &root_item->drop_progress,
2246                                path->slots[level]);
2247                 root_item->drop_level = level;
2248
2249                 btrfs_end_transaction_throttle(trans, root);
2250                 trans = NULL;
2251
2252                 btrfs_btree_balance_dirty(root);
2253
2254                 if (replaced && rc->stage == UPDATE_DATA_PTRS)
2255                         invalidate_extent_cache(root, &key, &next_key);
2256         }
2257
2258         /*
2259          * handle the case only one block in the fs tree need to be
2260          * relocated and the block is tree root.
2261          */
2262         leaf = btrfs_lock_root_node(root);
2263         ret = btrfs_cow_block(trans, root, leaf, NULL, 0, &leaf);
2264         btrfs_tree_unlock(leaf);
2265         free_extent_buffer(leaf);
2266         if (ret < 0)
2267                 err = ret;
2268 out:
2269         btrfs_free_path(path);
2270
2271         if (err == 0) {
2272                 memset(&root_item->drop_progress, 0,
2273                        sizeof(root_item->drop_progress));
2274                 root_item->drop_level = 0;
2275                 btrfs_set_root_refs(root_item, 0);
2276                 btrfs_update_reloc_root(trans, root);
2277         }
2278
2279         if (trans)
2280                 btrfs_end_transaction_throttle(trans, root);
2281
2282         btrfs_btree_balance_dirty(root);
2283
2284         if (replaced && rc->stage == UPDATE_DATA_PTRS)
2285                 invalidate_extent_cache(root, &key, &next_key);
2286
2287         return err;
2288 }
2289
2290 static noinline_for_stack
2291 int prepare_to_merge(struct reloc_control *rc, int err)
2292 {
2293         struct btrfs_root *root = rc->extent_root;
2294         struct btrfs_root *reloc_root;
2295         struct btrfs_trans_handle *trans;
2296         LIST_HEAD(reloc_roots);
2297         u64 num_bytes = 0;
2298         int ret;
2299
2300         mutex_lock(&root->fs_info->reloc_mutex);
2301         rc->merging_rsv_size += root->nodesize * (BTRFS_MAX_LEVEL - 1) * 2;
2302         rc->merging_rsv_size += rc->nodes_relocated * 2;
2303         mutex_unlock(&root->fs_info->reloc_mutex);
2304
2305 again:
2306         if (!err) {
2307                 num_bytes = rc->merging_rsv_size;
2308                 ret = btrfs_block_rsv_add(root, rc->block_rsv, num_bytes,
2309                                           BTRFS_RESERVE_FLUSH_ALL);
2310                 if (ret)
2311                         err = ret;
2312         }
2313
2314         trans = btrfs_join_transaction(rc->extent_root);
2315         if (IS_ERR(trans)) {
2316                 if (!err)
2317                         btrfs_block_rsv_release(rc->extent_root,
2318                                                 rc->block_rsv, num_bytes);
2319                 return PTR_ERR(trans);
2320         }
2321
2322         if (!err) {
2323                 if (num_bytes != rc->merging_rsv_size) {
2324                         btrfs_end_transaction(trans, rc->extent_root);
2325                         btrfs_block_rsv_release(rc->extent_root,
2326                                                 rc->block_rsv, num_bytes);
2327                         goto again;
2328                 }
2329         }
2330
2331         rc->merge_reloc_tree = 1;
2332
2333         while (!list_empty(&rc->reloc_roots)) {
2334                 reloc_root = list_entry(rc->reloc_roots.next,
2335                                         struct btrfs_root, root_list);
2336                 list_del_init(&reloc_root->root_list);
2337
2338                 root = read_fs_root(reloc_root->fs_info,
2339                                     reloc_root->root_key.offset);
2340                 BUG_ON(IS_ERR(root));
2341                 BUG_ON(root->reloc_root != reloc_root);
2342
2343                 /*
2344                  * set reference count to 1, so btrfs_recover_relocation
2345                  * knows it should resumes merging
2346                  */
2347                 if (!err)
2348                         btrfs_set_root_refs(&reloc_root->root_item, 1);
2349                 btrfs_update_reloc_root(trans, root);
2350
2351                 list_add(&reloc_root->root_list, &reloc_roots);
2352         }
2353
2354         list_splice(&reloc_roots, &rc->reloc_roots);
2355
2356         if (!err)
2357                 btrfs_commit_transaction(trans, rc->extent_root);
2358         else
2359                 btrfs_end_transaction(trans, rc->extent_root);
2360         return err;
2361 }
2362
2363 static noinline_for_stack
2364 void free_reloc_roots(struct list_head *list)
2365 {
2366         struct btrfs_root *reloc_root;
2367
2368         while (!list_empty(list)) {
2369                 reloc_root = list_entry(list->next, struct btrfs_root,
2370                                         root_list);
2371                 __del_reloc_root(reloc_root);
2372                 free_extent_buffer(reloc_root->node);
2373                 free_extent_buffer(reloc_root->commit_root);
2374                 reloc_root->node = NULL;
2375                 reloc_root->commit_root = NULL;
2376         }
2377 }
2378
2379 static noinline_for_stack
2380 void merge_reloc_roots(struct reloc_control *rc)
2381 {
2382         struct btrfs_root *root;
2383         struct btrfs_root *reloc_root;
2384         u64 last_snap;
2385         u64 otransid;
2386         u64 objectid;
2387         LIST_HEAD(reloc_roots);
2388         int found = 0;
2389         int ret = 0;
2390 again:
2391         root = rc->extent_root;
2392
2393         /*
2394          * this serializes us with btrfs_record_root_in_transaction,
2395          * we have to make sure nobody is in the middle of
2396          * adding their roots to the list while we are
2397          * doing this splice
2398          */
2399         mutex_lock(&root->fs_info->reloc_mutex);
2400         list_splice_init(&rc->reloc_roots, &reloc_roots);
2401         mutex_unlock(&root->fs_info->reloc_mutex);
2402
2403         while (!list_empty(&reloc_roots)) {
2404                 found = 1;
2405                 reloc_root = list_entry(reloc_roots.next,
2406                                         struct btrfs_root, root_list);
2407
2408                 if (btrfs_root_refs(&reloc_root->root_item) > 0) {
2409                         root = read_fs_root(reloc_root->fs_info,
2410                                             reloc_root->root_key.offset);
2411                         BUG_ON(IS_ERR(root));
2412                         BUG_ON(root->reloc_root != reloc_root);
2413
2414                         ret = merge_reloc_root(rc, root);
2415                         if (ret) {
2416                                 if (list_empty(&reloc_root->root_list))
2417                                         list_add_tail(&reloc_root->root_list,
2418                                                       &reloc_roots);
2419                                 goto out;
2420                         }
2421                 } else {
2422                         list_del_init(&reloc_root->root_list);
2423                 }
2424
2425                 /*
2426                  * we keep the old last snapshot transid in rtranid when we
2427                  * created the relocation tree.
2428                  */
2429                 last_snap = btrfs_root_rtransid(&reloc_root->root_item);
2430                 otransid = btrfs_root_otransid(&reloc_root->root_item);
2431                 objectid = reloc_root->root_key.offset;
2432
2433                 ret = btrfs_drop_snapshot(reloc_root, rc->block_rsv, 0, 1);
2434                 if (ret < 0) {
2435                         if (list_empty(&reloc_root->root_list))
2436                                 list_add_tail(&reloc_root->root_list,
2437                                               &reloc_roots);
2438                         goto out;
2439                 }
2440         }
2441
2442         if (found) {
2443                 found = 0;
2444                 goto again;
2445         }
2446 out:
2447         if (ret) {
2448                 btrfs_handle_fs_error(root->fs_info, ret, NULL);
2449                 if (!list_empty(&reloc_roots))
2450                         free_reloc_roots(&reloc_roots);
2451
2452                 /* new reloc root may be added */
2453                 mutex_lock(&root->fs_info->reloc_mutex);
2454                 list_splice_init(&rc->reloc_roots, &reloc_roots);
2455                 mutex_unlock(&root->fs_info->reloc_mutex);
2456                 if (!list_empty(&reloc_roots))
2457                         free_reloc_roots(&reloc_roots);
2458         }
2459
2460         /*
2461          * We used to have
2462          *
2463          * BUG_ON(!RB_EMPTY_ROOT(&rc->reloc_root_tree.rb_root));
2464          *
2465          * here, but it's wrong.  If we fail to start the transaction in
2466          * prepare_to_merge() we will have only 0 ref reloc roots, none of which
2467          * have actually been removed from the reloc_root_tree rb tree.  This is
2468          * fine because we're bailing here, and we hold a reference on the root
2469          * for the list that holds it, so these roots will be cleaned up when we
2470          * do the reloc_dirty_list afterwards.  Meanwhile the root->reloc_root
2471          * will be cleaned up on unmount.
2472          *
2473          * The remaining nodes will be cleaned up by free_reloc_control.
2474          */
2475 }
2476
2477 static void free_block_list(struct rb_root *blocks)
2478 {
2479         struct tree_block *block;
2480         struct rb_node *rb_node;
2481         while ((rb_node = rb_first(blocks))) {
2482                 block = rb_entry(rb_node, struct tree_block, rb_node);
2483                 rb_erase(rb_node, blocks);
2484                 kfree(block);
2485         }
2486 }
2487
2488 static int record_reloc_root_in_trans(struct btrfs_trans_handle *trans,
2489                                       struct btrfs_root *reloc_root)
2490 {
2491         struct btrfs_root *root;
2492
2493         if (reloc_root->last_trans == trans->transid)
2494                 return 0;
2495
2496         root = read_fs_root(reloc_root->fs_info, reloc_root->root_key.offset);
2497         BUG_ON(IS_ERR(root));
2498         BUG_ON(root->reloc_root != reloc_root);
2499
2500         return btrfs_record_root_in_trans(trans, root);
2501 }
2502
2503 static noinline_for_stack
2504 struct btrfs_root *select_reloc_root(struct btrfs_trans_handle *trans,
2505                                      struct reloc_control *rc,
2506                                      struct backref_node *node,
2507                                      struct backref_edge *edges[])
2508 {
2509         struct backref_node *next;
2510         struct btrfs_root *root;
2511         int index = 0;
2512
2513         next = node;
2514         while (1) {
2515                 cond_resched();
2516                 next = walk_up_backref(next, edges, &index);
2517                 root = next->root;
2518                 BUG_ON(!root);
2519                 BUG_ON(!test_bit(BTRFS_ROOT_REF_COWS, &root->state));
2520
2521                 if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID) {
2522                         record_reloc_root_in_trans(trans, root);
2523                         break;
2524                 }
2525
2526                 btrfs_record_root_in_trans(trans, root);
2527                 root = root->reloc_root;
2528
2529                 if (next->new_bytenr != root->node->start) {
2530                         BUG_ON(next->new_bytenr);
2531                         BUG_ON(!list_empty(&next->list));
2532                         next->new_bytenr = root->node->start;
2533                         next->root = root;
2534                         list_add_tail(&next->list,
2535                                       &rc->backref_cache.changed);
2536                         __mark_block_processed(rc, next);
2537                         break;
2538                 }
2539
2540                 WARN_ON(1);
2541                 root = NULL;
2542                 next = walk_down_backref(edges, &index);
2543                 if (!next || next->level <= node->level)
2544                         break;
2545         }
2546         if (!root)
2547                 return NULL;
2548
2549         next = node;
2550         /* setup backref node path for btrfs_reloc_cow_block */
2551         while (1) {
2552                 rc->backref_cache.path[next->level] = next;
2553                 if (--index < 0)
2554                         break;
2555                 next = edges[index]->node[UPPER];
2556         }
2557         return root;
2558 }
2559
2560 /*
2561  * select a tree root for relocation. return NULL if the block
2562  * is reference counted. we should use do_relocation() in this
2563  * case. return a tree root pointer if the block isn't reference
2564  * counted. return -ENOENT if the block is root of reloc tree.
2565  */
2566 static noinline_for_stack
2567 struct btrfs_root *select_one_root(struct backref_node *node)
2568 {
2569         struct backref_node *next;
2570         struct btrfs_root *root;
2571         struct btrfs_root *fs_root = NULL;
2572         struct backref_edge *edges[BTRFS_MAX_LEVEL - 1];
2573         int index = 0;
2574
2575         next = node;
2576         while (1) {
2577                 cond_resched();
2578                 next = walk_up_backref(next, edges, &index);
2579                 root = next->root;
2580                 BUG_ON(!root);
2581
2582                 /* no other choice for non-references counted tree */
2583                 if (!test_bit(BTRFS_ROOT_REF_COWS, &root->state))
2584                         return root;
2585
2586                 if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID)
2587                         fs_root = root;
2588
2589                 if (next != node)
2590                         return NULL;
2591
2592                 next = walk_down_backref(edges, &index);
2593                 if (!next || next->level <= node->level)
2594                         break;
2595         }
2596
2597         if (!fs_root)
2598                 return ERR_PTR(-ENOENT);
2599         return fs_root;
2600 }
2601
2602 static noinline_for_stack
2603 u64 calcu_metadata_size(struct reloc_control *rc,
2604                         struct backref_node *node, int reserve)
2605 {
2606         struct backref_node *next = node;
2607         struct backref_edge *edge;
2608         struct backref_edge *edges[BTRFS_MAX_LEVEL - 1];
2609         u64 num_bytes = 0;
2610         int index = 0;
2611
2612         BUG_ON(reserve && node->processed);
2613
2614         while (next) {
2615                 cond_resched();
2616                 while (1) {
2617                         if (next->processed && (reserve || next != node))
2618                                 break;
2619
2620                         num_bytes += rc->extent_root->nodesize;
2621
2622                         if (list_empty(&next->upper))
2623                                 break;
2624
2625                         edge = list_entry(next->upper.next,
2626                                           struct backref_edge, list[LOWER]);
2627                         edges[index++] = edge;
2628                         next = edge->node[UPPER];
2629                 }
2630                 next = walk_down_backref(edges, &index);
2631         }
2632         return num_bytes;
2633 }
2634
2635 static int reserve_metadata_space(struct btrfs_trans_handle *trans,
2636                                   struct reloc_control *rc,
2637                                   struct backref_node *node)
2638 {
2639         struct btrfs_root *root = rc->extent_root;
2640         u64 num_bytes;
2641         int ret;
2642         u64 tmp;
2643
2644         num_bytes = calcu_metadata_size(rc, node, 1) * 2;
2645
2646         trans->block_rsv = rc->block_rsv;
2647         rc->reserved_bytes += num_bytes;
2648
2649         /*
2650          * We are under a transaction here so we can only do limited flushing.
2651          * If we get an enospc just kick back -EAGAIN so we know to drop the
2652          * transaction and try to refill when we can flush all the things.
2653          */
2654         ret = btrfs_block_rsv_refill(root, rc->block_rsv, num_bytes,
2655                                 BTRFS_RESERVE_FLUSH_LIMIT);
2656         if (ret) {
2657                 tmp = rc->extent_root->nodesize * RELOCATION_RESERVED_NODES;
2658                 while (tmp <= rc->reserved_bytes)
2659                         tmp <<= 1;
2660                 /*
2661                  * only one thread can access block_rsv at this point,
2662                  * so we don't need hold lock to protect block_rsv.
2663                  * we expand more reservation size here to allow enough
2664                  * space for relocation and we will return eailer in
2665                  * enospc case.
2666                  */
2667                 rc->block_rsv->size = tmp + rc->extent_root->nodesize *
2668                         RELOCATION_RESERVED_NODES;
2669                 return -EAGAIN;
2670         }
2671
2672         return 0;
2673 }
2674
2675 /*
2676  * relocate a block tree, and then update pointers in upper level
2677  * blocks that reference the block to point to the new location.
2678  *
2679  * if called by link_to_upper, the block has already been relocated.
2680  * in that case this function just updates pointers.
2681  */
2682 static int do_relocation(struct btrfs_trans_handle *trans,
2683                          struct reloc_control *rc,
2684                          struct backref_node *node,
2685                          struct btrfs_key *key,
2686                          struct btrfs_path *path, int lowest)
2687 {
2688         struct backref_node *upper;
2689         struct backref_edge *edge;
2690         struct backref_edge *edges[BTRFS_MAX_LEVEL - 1];
2691         struct btrfs_root *root;
2692         struct extent_buffer *eb;
2693         u32 blocksize;
2694         u64 bytenr;
2695         u64 generation;
2696         int slot;
2697         int ret;
2698         int err = 0;
2699
2700         BUG_ON(lowest && node->eb);
2701
2702         path->lowest_level = node->level + 1;
2703         rc->backref_cache.path[node->level] = node;
2704         list_for_each_entry(edge, &node->upper, list[LOWER]) {
2705                 cond_resched();
2706
2707                 upper = edge->node[UPPER];
2708                 root = select_reloc_root(trans, rc, upper, edges);
2709                 BUG_ON(!root);
2710
2711                 if (upper->eb && !upper->locked) {
2712                         if (!lowest) {
2713                                 ret = btrfs_bin_search(upper->eb, key,
2714                                                        upper->level, &slot);
2715                                 BUG_ON(ret);
2716                                 bytenr = btrfs_node_blockptr(upper->eb, slot);
2717                                 if (node->eb->start == bytenr)
2718                                         goto next;
2719                         }
2720                         drop_node_buffer(upper);
2721                 }
2722
2723                 if (!upper->eb) {
2724                         ret = btrfs_search_slot(trans, root, key, path, 0, 1);
2725                         if (ret) {
2726                                 if (ret < 0)
2727                                         err = ret;
2728                                 else
2729                                         err = -ENOENT;
2730
2731                                 btrfs_release_path(path);
2732                                 break;
2733                         }
2734
2735                         if (!upper->eb) {
2736                                 upper->eb = path->nodes[upper->level];
2737                                 path->nodes[upper->level] = NULL;
2738                         } else {
2739                                 BUG_ON(upper->eb != path->nodes[upper->level]);
2740                         }
2741
2742                         upper->locked = 1;
2743                         path->locks[upper->level] = 0;
2744
2745                         slot = path->slots[upper->level];
2746                         btrfs_release_path(path);
2747                 } else {
2748                         ret = btrfs_bin_search(upper->eb, key, upper->level,
2749                                                &slot);
2750                         BUG_ON(ret);
2751                 }
2752
2753                 bytenr = btrfs_node_blockptr(upper->eb, slot);
2754                 if (lowest) {
2755                         if (bytenr != node->bytenr) {
2756                                 btrfs_err(root->fs_info,
2757                 "lowest leaf/node mismatch: bytenr %llu node->bytenr %llu slot %d upper %llu",
2758                                           bytenr, node->bytenr, slot,
2759                                           upper->eb->start);
2760                                 err = -EIO;
2761                                 goto next;
2762                         }
2763                 } else {
2764                         if (node->eb->start == bytenr)
2765                                 goto next;
2766                 }
2767
2768                 blocksize = root->nodesize;
2769                 generation = btrfs_node_ptr_generation(upper->eb, slot);
2770                 eb = read_tree_block(root, bytenr, generation);
2771                 if (IS_ERR(eb)) {
2772                         err = PTR_ERR(eb);
2773                         goto next;
2774                 } else if (!extent_buffer_uptodate(eb)) {
2775                         free_extent_buffer(eb);
2776                         err = -EIO;
2777                         goto next;
2778                 }
2779                 btrfs_tree_lock(eb);
2780                 btrfs_set_lock_blocking(eb);
2781
2782                 if (!node->eb) {
2783                         ret = btrfs_cow_block(trans, root, eb, upper->eb,
2784                                               slot, &eb);
2785                         btrfs_tree_unlock(eb);
2786                         free_extent_buffer(eb);
2787                         if (ret < 0) {
2788                                 err = ret;
2789                                 goto next;
2790                         }
2791                         BUG_ON(node->eb != eb);
2792                 } else {
2793                         btrfs_set_node_blockptr(upper->eb, slot,
2794                                                 node->eb->start);
2795                         btrfs_set_node_ptr_generation(upper->eb, slot,
2796                                                       trans->transid);
2797                         btrfs_mark_buffer_dirty(upper->eb);
2798
2799                         ret = btrfs_inc_extent_ref(trans, root,
2800                                                 node->eb->start, blocksize,
2801                                                 upper->eb->start,
2802                                                 btrfs_header_owner(upper->eb),
2803                                                 node->level, 0);
2804                         BUG_ON(ret);
2805
2806                         ret = btrfs_drop_subtree(trans, root, eb, upper->eb);
2807                         BUG_ON(ret);
2808                 }
2809 next:
2810                 if (!upper->pending)
2811                         drop_node_buffer(upper);
2812                 else
2813                         unlock_node_buffer(upper);
2814                 if (err)
2815                         break;
2816         }
2817
2818         if (!err && node->pending) {
2819                 drop_node_buffer(node);
2820                 list_move_tail(&node->list, &rc->backref_cache.changed);
2821                 node->pending = 0;
2822         }
2823
2824         path->lowest_level = 0;
2825         BUG_ON(err == -ENOSPC);
2826         return err;
2827 }
2828
2829 static int link_to_upper(struct btrfs_trans_handle *trans,
2830                          struct reloc_control *rc,
2831                          struct backref_node *node,
2832                          struct btrfs_path *path)
2833 {
2834         struct btrfs_key key;
2835
2836         btrfs_node_key_to_cpu(node->eb, &key, 0);
2837         return do_relocation(trans, rc, node, &key, path, 0);
2838 }
2839
2840 static int finish_pending_nodes(struct btrfs_trans_handle *trans,
2841                                 struct reloc_control *rc,
2842                                 struct btrfs_path *path, int err)
2843 {
2844         LIST_HEAD(list);
2845         struct backref_cache *cache = &rc->backref_cache;
2846         struct backref_node *node;
2847         int level;
2848         int ret;
2849
2850         for (level = 0; level < BTRFS_MAX_LEVEL; level++) {
2851                 while (!list_empty(&cache->pending[level])) {
2852                         node = list_entry(cache->pending[level].next,
2853                                           struct backref_node, list);
2854                         list_move_tail(&node->list, &list);
2855                         BUG_ON(!node->pending);
2856
2857                         if (!err) {
2858                                 ret = link_to_upper(trans, rc, node, path);
2859                                 if (ret < 0)
2860                                         err = ret;
2861                         }
2862                 }
2863                 list_splice_init(&list, &cache->pending[level]);
2864         }
2865         return err;
2866 }
2867
2868 static void mark_block_processed(struct reloc_control *rc,
2869                                  u64 bytenr, u32 blocksize)
2870 {
2871         set_extent_bits(&rc->processed_blocks, bytenr, bytenr + blocksize - 1,
2872                         EXTENT_DIRTY);
2873 }
2874
2875 static void __mark_block_processed(struct reloc_control *rc,
2876                                    struct backref_node *node)
2877 {
2878         u32 blocksize;
2879         if (node->level == 0 ||
2880             in_block_group(node->bytenr, rc->block_group)) {
2881                 blocksize = rc->extent_root->nodesize;
2882                 mark_block_processed(rc, node->bytenr, blocksize);
2883         }
2884         node->processed = 1;
2885 }
2886
2887 /*
2888  * mark a block and all blocks directly/indirectly reference the block
2889  * as processed.
2890  */
2891 static void update_processed_blocks(struct reloc_control *rc,
2892                                     struct backref_node *node)
2893 {
2894         struct backref_node *next = node;
2895         struct backref_edge *edge;
2896         struct backref_edge *edges[BTRFS_MAX_LEVEL - 1];
2897         int index = 0;
2898
2899         while (next) {
2900                 cond_resched();
2901                 while (1) {
2902                         if (next->processed)
2903                                 break;
2904
2905                         __mark_block_processed(rc, next);
2906
2907                         if (list_empty(&next->upper))
2908                                 break;
2909
2910                         edge = list_entry(next->upper.next,
2911                                           struct backref_edge, list[LOWER]);
2912                         edges[index++] = edge;
2913                         next = edge->node[UPPER];
2914                 }
2915                 next = walk_down_backref(edges, &index);
2916         }
2917 }
2918
2919 static int tree_block_processed(u64 bytenr, struct reloc_control *rc)
2920 {
2921         u32 blocksize = rc->extent_root->nodesize;
2922
2923         if (test_range_bit(&rc->processed_blocks, bytenr,
2924                            bytenr + blocksize - 1, EXTENT_DIRTY, 1, NULL))
2925                 return 1;
2926         return 0;
2927 }
2928
2929 static int get_tree_block_key(struct reloc_control *rc,
2930                               struct tree_block *block)
2931 {
2932         struct extent_buffer *eb;
2933
2934         BUG_ON(block->key_ready);
2935         eb = read_tree_block(rc->extent_root, block->bytenr,
2936                              block->key.offset);
2937         if (IS_ERR(eb)) {
2938                 return PTR_ERR(eb);
2939         } else if (!extent_buffer_uptodate(eb)) {
2940                 free_extent_buffer(eb);
2941                 return -EIO;
2942         }
2943         WARN_ON(btrfs_header_level(eb) != block->level);
2944         if (block->level == 0)
2945                 btrfs_item_key_to_cpu(eb, &block->key, 0);
2946         else
2947                 btrfs_node_key_to_cpu(eb, &block->key, 0);
2948         free_extent_buffer(eb);
2949         block->key_ready = 1;
2950         return 0;
2951 }
2952
2953 /*
2954  * helper function to relocate a tree block
2955  */
2956 static int relocate_tree_block(struct btrfs_trans_handle *trans,
2957                                 struct reloc_control *rc,
2958                                 struct backref_node *node,
2959                                 struct btrfs_key *key,
2960                                 struct btrfs_path *path)
2961 {
2962         struct btrfs_root *root;
2963         int ret = 0;
2964
2965         if (!node)
2966                 return 0;
2967
2968         BUG_ON(node->processed);
2969         root = select_one_root(node);
2970         if (root == ERR_PTR(-ENOENT)) {
2971                 update_processed_blocks(rc, node);
2972                 goto out;
2973         }
2974
2975         if (!root || test_bit(BTRFS_ROOT_REF_COWS, &root->state)) {
2976                 ret = reserve_metadata_space(trans, rc, node);
2977                 if (ret)
2978                         goto out;
2979         }
2980
2981         if (root) {
2982                 if (test_bit(BTRFS_ROOT_REF_COWS, &root->state)) {
2983                         BUG_ON(node->new_bytenr);
2984                         BUG_ON(!list_empty(&node->list));
2985                         btrfs_record_root_in_trans(trans, root);
2986                         root = root->reloc_root;
2987                         node->new_bytenr = root->node->start;
2988                         node->root = root;
2989                         list_add_tail(&node->list, &rc->backref_cache.changed);
2990                 } else {
2991                         path->lowest_level = node->level;
2992                         ret = btrfs_search_slot(trans, root, key, path, 0, 1);
2993                         btrfs_release_path(path);
2994                         if (ret > 0)
2995                                 ret = 0;
2996                 }
2997                 if (!ret)
2998                         update_processed_blocks(rc, node);
2999         } else {
3000                 ret = do_relocation(trans, rc, node, key, path, 1);
3001         }
3002 out:
3003         if (ret || node->level == 0 || node->cowonly)
3004                 remove_backref_node(&rc->backref_cache, node);
3005         return ret;
3006 }
3007
3008 /*
3009  * relocate a list of blocks
3010  */
3011 static noinline_for_stack
3012 int relocate_tree_blocks(struct btrfs_trans_handle *trans,
3013                          struct reloc_control *rc, struct rb_root *blocks)
3014 {
3015         struct backref_node *node;
3016         struct btrfs_path *path;
3017         struct tree_block *block;
3018         struct rb_node *rb_node;
3019         int ret;
3020         int err = 0;
3021
3022         path = btrfs_alloc_path();
3023         if (!path) {
3024                 err = -ENOMEM;
3025                 goto out_free_blocks;
3026         }
3027
3028         rb_node = rb_first(blocks);
3029         while (rb_node) {
3030                 block = rb_entry(rb_node, struct tree_block, rb_node);
3031                 if (!block->key_ready)
3032                         readahead_tree_block(rc->extent_root, block->bytenr);
3033                 rb_node = rb_next(rb_node);
3034         }
3035
3036         rb_node = rb_first(blocks);
3037         while (rb_node) {
3038                 block = rb_entry(rb_node, struct tree_block, rb_node);
3039                 if (!block->key_ready) {
3040                         err = get_tree_block_key(rc, block);
3041                         if (err)
3042                                 goto out_free_path;
3043                 }
3044                 rb_node = rb_next(rb_node);
3045         }
3046
3047         rb_node = rb_first(blocks);
3048         while (rb_node) {
3049                 block = rb_entry(rb_node, struct tree_block, rb_node);
3050
3051                 node = build_backref_tree(rc, &block->key,
3052                                           block->level, block->bytenr);
3053                 if (IS_ERR(node)) {
3054                         err = PTR_ERR(node);
3055                         goto out;
3056                 }
3057
3058                 ret = relocate_tree_block(trans, rc, node, &block->key,
3059                                           path);
3060                 if (ret < 0) {
3061                         if (ret != -EAGAIN || rb_node == rb_first(blocks))
3062                                 err = ret;
3063                         goto out;
3064                 }
3065                 rb_node = rb_next(rb_node);
3066         }
3067 out:
3068         err = finish_pending_nodes(trans, rc, path, err);
3069
3070 out_free_path:
3071         btrfs_free_path(path);
3072 out_free_blocks:
3073         free_block_list(blocks);
3074         return err;
3075 }
3076
3077 static noinline_for_stack
3078 int prealloc_file_extent_cluster(struct inode *inode,
3079                                  struct file_extent_cluster *cluster)
3080 {
3081         u64 alloc_hint = 0;
3082         u64 start;
3083         u64 end;
3084         u64 offset = BTRFS_I(inode)->index_cnt;
3085         u64 num_bytes;
3086         int nr = 0;
3087         int ret = 0;
3088         u64 prealloc_start = cluster->start - offset;
3089         u64 prealloc_end = cluster->end - offset;
3090         u64 cur_offset;
3091
3092         BUG_ON(cluster->start != cluster->boundary[0]);
3093         inode_lock(inode);
3094
3095         ret = btrfs_check_data_free_space(inode, prealloc_start,
3096                                           prealloc_end + 1 - prealloc_start);
3097         if (ret)
3098                 goto out;
3099
3100         cur_offset = prealloc_start;
3101         while (nr < cluster->nr) {
3102                 start = cluster->boundary[nr] - offset;
3103                 if (nr + 1 < cluster->nr)
3104                         end = cluster->boundary[nr + 1] - 1 - offset;
3105                 else
3106                         end = cluster->end - offset;
3107
3108                 lock_extent(&BTRFS_I(inode)->io_tree, start, end);
3109                 num_bytes = end + 1 - start;
3110                 if (cur_offset < start)
3111                         btrfs_free_reserved_data_space(inode, cur_offset,
3112                                         start - cur_offset);
3113                 ret = btrfs_prealloc_file_range(inode, 0, start,
3114                                                 num_bytes, num_bytes,
3115                                                 end + 1, &alloc_hint);
3116                 cur_offset = end + 1;
3117                 unlock_extent(&BTRFS_I(inode)->io_tree, start, end);
3118                 if (ret)
3119                         break;
3120                 nr++;
3121         }
3122         if (cur_offset < prealloc_end)
3123                 btrfs_free_reserved_data_space(inode, cur_offset,
3124                                        prealloc_end + 1 - cur_offset);
3125 out:
3126         inode_unlock(inode);
3127         return ret;
3128 }
3129
3130 static noinline_for_stack
3131 int setup_extent_mapping(struct inode *inode, u64 start, u64 end,
3132                          u64 block_start)
3133 {
3134         struct btrfs_root *root = BTRFS_I(inode)->root;
3135         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
3136         struct extent_map *em;
3137         int ret = 0;
3138
3139         em = alloc_extent_map();
3140         if (!em)
3141                 return -ENOMEM;
3142
3143         em->start = start;
3144         em->len = end + 1 - start;
3145         em->block_len = em->len;
3146         em->block_start = block_start;
3147         em->bdev = root->fs_info->fs_devices->latest_bdev;
3148         set_bit(EXTENT_FLAG_PINNED, &em->flags);
3149
3150         lock_extent(&BTRFS_I(inode)->io_tree, start, end);
3151         while (1) {
3152                 write_lock(&em_tree->lock);
3153                 ret = add_extent_mapping(em_tree, em, 0);
3154                 write_unlock(&em_tree->lock);
3155                 if (ret != -EEXIST) {
3156                         free_extent_map(em);
3157                         break;
3158                 }
3159                 btrfs_drop_extent_cache(inode, start, end, 0);
3160         }
3161         unlock_extent(&BTRFS_I(inode)->io_tree, start, end);
3162         return ret;
3163 }
3164
3165 static int relocate_file_extent_cluster(struct inode *inode,
3166                                         struct file_extent_cluster *cluster)
3167 {
3168         u64 page_start;
3169         u64 page_end;
3170         u64 offset = BTRFS_I(inode)->index_cnt;
3171         unsigned long index;
3172         unsigned long last_index;
3173         struct page *page;
3174         struct file_ra_state *ra;
3175         gfp_t mask = btrfs_alloc_write_mask(inode->i_mapping);
3176         int nr = 0;
3177         int ret = 0;
3178
3179         if (!cluster->nr)
3180                 return 0;
3181
3182         ra = kzalloc(sizeof(*ra), GFP_NOFS);
3183         if (!ra)
3184                 return -ENOMEM;
3185
3186         ret = prealloc_file_extent_cluster(inode, cluster);
3187         if (ret)
3188                 goto out;
3189
3190         file_ra_state_init(ra, inode->i_mapping);
3191
3192         ret = setup_extent_mapping(inode, cluster->start - offset,
3193                                    cluster->end - offset, cluster->start);
3194         if (ret)
3195                 goto out;
3196
3197         index = (cluster->start - offset) >> PAGE_SHIFT;
3198         last_index = (cluster->end - offset) >> PAGE_SHIFT;
3199         while (index <= last_index) {
3200                 ret = btrfs_delalloc_reserve_metadata(inode, PAGE_SIZE);
3201                 if (ret)
3202                         goto out;
3203
3204                 page = find_lock_page(inode->i_mapping, index);
3205                 if (!page) {
3206                         page_cache_sync_readahead(inode->i_mapping,
3207                                                   ra, NULL, index,
3208                                                   last_index + 1 - index);
3209                         page = find_or_create_page(inode->i_mapping, index,
3210                                                    mask);
3211                         if (!page) {
3212                                 btrfs_delalloc_release_metadata(inode,
3213                                                         PAGE_SIZE);
3214                                 ret = -ENOMEM;
3215                                 goto out;
3216                         }
3217                 }
3218
3219                 if (PageReadahead(page)) {
3220                         page_cache_async_readahead(inode->i_mapping,
3221                                                    ra, NULL, page, index,
3222                                                    last_index + 1 - index);
3223                 }
3224
3225                 if (!PageUptodate(page)) {
3226                         btrfs_readpage(NULL, page);
3227                         lock_page(page);
3228                         if (!PageUptodate(page)) {
3229                                 unlock_page(page);
3230                                 put_page(page);
3231                                 btrfs_delalloc_release_metadata(inode,
3232                                                         PAGE_SIZE);
3233                                 ret = -EIO;
3234                                 goto out;
3235                         }
3236                 }
3237
3238                 page_start = page_offset(page);
3239                 page_end = page_start + PAGE_SIZE - 1;
3240
3241                 lock_extent(&BTRFS_I(inode)->io_tree, page_start, page_end);
3242
3243                 set_page_extent_mapped(page);
3244
3245                 if (nr < cluster->nr &&
3246                     page_start + offset == cluster->boundary[nr]) {
3247                         set_extent_bits(&BTRFS_I(inode)->io_tree,
3248                                         page_start, page_end,
3249                                         EXTENT_BOUNDARY);
3250                         nr++;
3251                 }
3252
3253                 btrfs_set_extent_delalloc(inode, page_start, page_end, NULL, 0);
3254                 set_page_dirty(page);
3255
3256                 unlock_extent(&BTRFS_I(inode)->io_tree,
3257                               page_start, page_end);
3258                 unlock_page(page);
3259                 put_page(page);
3260
3261                 index++;
3262                 balance_dirty_pages_ratelimited(inode->i_mapping);
3263                 btrfs_throttle(BTRFS_I(inode)->root);
3264         }
3265         WARN_ON(nr != cluster->nr);
3266 out:
3267         kfree(ra);
3268         return ret;
3269 }
3270
3271 static noinline_for_stack
3272 int relocate_data_extent(struct inode *inode, struct btrfs_key *extent_key,
3273                          struct file_extent_cluster *cluster)
3274 {
3275         int ret;
3276
3277         if (cluster->nr > 0 && extent_key->objectid != cluster->end + 1) {
3278                 ret = relocate_file_extent_cluster(inode, cluster);
3279                 if (ret)
3280                         return ret;
3281                 cluster->nr = 0;
3282         }
3283
3284         if (!cluster->nr)
3285                 cluster->start = extent_key->objectid;
3286         else
3287                 BUG_ON(cluster->nr >= MAX_EXTENTS);
3288         cluster->end = extent_key->objectid + extent_key->offset - 1;
3289         cluster->boundary[cluster->nr] = extent_key->objectid;
3290         cluster->nr++;
3291
3292         if (cluster->nr >= MAX_EXTENTS) {
3293                 ret = relocate_file_extent_cluster(inode, cluster);
3294                 if (ret)
3295                         return ret;
3296                 cluster->nr = 0;
3297         }
3298         return 0;
3299 }
3300
3301 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
3302 static int get_ref_objectid_v0(struct reloc_control *rc,
3303                                struct btrfs_path *path,
3304                                struct btrfs_key *extent_key,
3305                                u64 *ref_objectid, int *path_change)
3306 {
3307         struct btrfs_key key;
3308         struct extent_buffer *leaf;
3309         struct btrfs_extent_ref_v0 *ref0;
3310         int ret;
3311         int slot;
3312
3313         leaf = path->nodes[0];
3314         slot = path->slots[0];
3315         while (1) {
3316                 if (slot >= btrfs_header_nritems(leaf)) {
3317                         ret = btrfs_next_leaf(rc->extent_root, path);
3318                         if (ret < 0)
3319                                 return ret;
3320                         BUG_ON(ret > 0);
3321                         leaf = path->nodes[0];
3322                         slot = path->slots[0];
3323                         if (path_change)
3324                                 *path_change = 1;
3325                 }
3326                 btrfs_item_key_to_cpu(leaf, &key, slot);
3327                 if (key.objectid != extent_key->objectid)
3328                         return -ENOENT;
3329
3330                 if (key.type != BTRFS_EXTENT_REF_V0_KEY) {
3331                         slot++;
3332                         continue;
3333                 }
3334                 ref0 = btrfs_item_ptr(leaf, slot,
3335                                 struct btrfs_extent_ref_v0);
3336                 *ref_objectid = btrfs_ref_objectid_v0(leaf, ref0);
3337                 break;
3338         }
3339         return 0;
3340 }
3341 #endif
3342
3343 /*
3344  * helper to add a tree block to the list.
3345  * the major work is getting the generation and level of the block
3346  */
3347 static int add_tree_block(struct reloc_control *rc,
3348                           struct btrfs_key *extent_key,
3349                           struct btrfs_path *path,
3350                           struct rb_root *blocks)
3351 {
3352         struct extent_buffer *eb;
3353         struct btrfs_extent_item *ei;
3354         struct btrfs_tree_block_info *bi;
3355         struct tree_block *block;
3356         struct rb_node *rb_node;
3357         u32 item_size;
3358         int level = -1;
3359         u64 generation;
3360
3361         eb =  path->nodes[0];
3362         item_size = btrfs_item_size_nr(eb, path->slots[0]);
3363
3364         if (extent_key->type == BTRFS_METADATA_ITEM_KEY ||
3365             item_size >= sizeof(*ei) + sizeof(*bi)) {
3366                 ei = btrfs_item_ptr(eb, path->slots[0],
3367                                 struct btrfs_extent_item);
3368                 if (extent_key->type == BTRFS_EXTENT_ITEM_KEY) {
3369                         bi = (struct btrfs_tree_block_info *)(ei + 1);
3370                         level = btrfs_tree_block_level(eb, bi);
3371                 } else {
3372                         level = (int)extent_key->offset;
3373                 }
3374                 generation = btrfs_extent_generation(eb, ei);
3375         } else {
3376 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
3377                 u64 ref_owner;
3378                 int ret;
3379
3380                 BUG_ON(item_size != sizeof(struct btrfs_extent_item_v0));
3381                 ret = get_ref_objectid_v0(rc, path, extent_key,
3382                                           &ref_owner, NULL);
3383                 if (ret < 0)
3384                         return ret;
3385                 BUG_ON(ref_owner >= BTRFS_MAX_LEVEL);
3386                 level = (int)ref_owner;
3387                 /* FIXME: get real generation */
3388                 generation = 0;
3389 #else
3390                 BUG();
3391 #endif
3392         }
3393
3394         btrfs_release_path(path);
3395
3396         BUG_ON(level == -1);
3397
3398         block = kmalloc(sizeof(*block), GFP_NOFS);
3399         if (!block)
3400                 return -ENOMEM;
3401
3402         block->bytenr = extent_key->objectid;
3403         block->key.objectid = rc->extent_root->nodesize;
3404         block->key.offset = generation;
3405         block->level = level;
3406         block->key_ready = 0;
3407
3408         rb_node = tree_insert(blocks, block->bytenr, &block->rb_node);
3409         if (rb_node)
3410                 backref_tree_panic(rb_node, -EEXIST, block->bytenr);
3411
3412         return 0;
3413 }
3414
3415 /*
3416  * helper to add tree blocks for backref of type BTRFS_SHARED_DATA_REF_KEY
3417  */
3418 static int __add_tree_block(struct reloc_control *rc,
3419                             u64 bytenr, u32 blocksize,
3420                             struct rb_root *blocks)
3421 {
3422         struct btrfs_path *path;
3423         struct btrfs_key key;
3424         int ret;
3425         bool skinny = btrfs_fs_incompat(rc->extent_root->fs_info,
3426                                         SKINNY_METADATA);
3427
3428         if (tree_block_processed(bytenr, rc))
3429                 return 0;
3430
3431         if (tree_search(blocks, bytenr))
3432                 return 0;
3433
3434         path = btrfs_alloc_path();
3435         if (!path)
3436                 return -ENOMEM;
3437 again:
3438         key.objectid = bytenr;
3439         if (skinny) {
3440                 key.type = BTRFS_METADATA_ITEM_KEY;
3441                 key.offset = (u64)-1;
3442         } else {
3443                 key.type = BTRFS_EXTENT_ITEM_KEY;
3444                 key.offset = blocksize;
3445         }
3446
3447         path->search_commit_root = 1;
3448         path->skip_locking = 1;
3449         ret = btrfs_search_slot(NULL, rc->extent_root, &key, path, 0, 0);
3450         if (ret < 0)
3451                 goto out;
3452
3453         if (ret > 0 && skinny) {
3454                 if (path->slots[0]) {
3455                         path->slots[0]--;
3456                         btrfs_item_key_to_cpu(path->nodes[0], &key,
3457                                               path->slots[0]);
3458                         if (key.objectid == bytenr &&
3459                             (key.type == BTRFS_METADATA_ITEM_KEY ||
3460                              (key.type == BTRFS_EXTENT_ITEM_KEY &&
3461                               key.offset == blocksize)))
3462                                 ret = 0;
3463                 }
3464
3465                 if (ret) {
3466                         skinny = false;
3467                         btrfs_release_path(path);
3468                         goto again;
3469                 }
3470         }
3471         BUG_ON(ret);
3472
3473         ret = add_tree_block(rc, &key, path, blocks);
3474 out:
3475         btrfs_free_path(path);
3476         return ret;
3477 }
3478
3479 /*
3480  * helper to check if the block use full backrefs for pointers in it
3481  */
3482 static int block_use_full_backref(struct reloc_control *rc,
3483                                   struct extent_buffer *eb)
3484 {
3485         u64 flags;
3486         int ret;
3487
3488         if (btrfs_header_flag(eb, BTRFS_HEADER_FLAG_RELOC) ||
3489             btrfs_header_backref_rev(eb) < BTRFS_MIXED_BACKREF_REV)
3490                 return 1;
3491
3492         ret = btrfs_lookup_extent_info(NULL, rc->extent_root,
3493                                        eb->start, btrfs_header_level(eb), 1,
3494                                        NULL, &flags);
3495         BUG_ON(ret);
3496
3497         if (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF)
3498                 ret = 1;
3499         else
3500                 ret = 0;
3501         return ret;
3502 }
3503
3504 static int delete_block_group_cache(struct btrfs_fs_info *fs_info,
3505                                     struct btrfs_block_group_cache *block_group,
3506                                     struct inode *inode,
3507                                     u64 ino)
3508 {
3509         struct btrfs_key key;
3510         struct btrfs_root *root = fs_info->tree_root;
3511         struct btrfs_trans_handle *trans;
3512         int ret = 0;
3513
3514         if (inode)
3515                 goto truncate;
3516
3517         key.objectid = ino;
3518         key.type = BTRFS_INODE_ITEM_KEY;
3519         key.offset = 0;
3520
3521         inode = btrfs_iget(fs_info->sb, &key, root, NULL);
3522         if (IS_ERR(inode) || is_bad_inode(inode)) {
3523                 if (!IS_ERR(inode))
3524                         iput(inode);
3525                 return -ENOENT;
3526         }
3527
3528 truncate:
3529         ret = btrfs_check_trunc_cache_free_space(root,
3530                                                  &fs_info->global_block_rsv);
3531         if (ret)
3532                 goto out;
3533
3534         trans = btrfs_join_transaction(root);
3535         if (IS_ERR(trans)) {
3536                 ret = PTR_ERR(trans);
3537                 goto out;
3538         }
3539
3540         ret = btrfs_truncate_free_space_cache(root, trans, block_group, inode);
3541
3542         btrfs_end_transaction(trans, root);
3543         btrfs_btree_balance_dirty(root);
3544 out:
3545         iput(inode);
3546         return ret;
3547 }
3548
3549 /*
3550  * helper to add tree blocks for backref of type BTRFS_EXTENT_DATA_REF_KEY
3551  * this function scans fs tree to find blocks reference the data extent
3552  */
3553 static int find_data_references(struct reloc_control *rc,
3554                                 struct btrfs_key *extent_key,
3555                                 struct extent_buffer *leaf,
3556                                 struct btrfs_extent_data_ref *ref,
3557                                 struct rb_root *blocks)
3558 {
3559         struct btrfs_path *path;
3560         struct tree_block *block;
3561         struct btrfs_root *root;
3562         struct btrfs_file_extent_item *fi;
3563         struct rb_node *rb_node;
3564         struct btrfs_key key;
3565         u64 ref_root;
3566         u64 ref_objectid;
3567         u64 ref_offset;
3568         u32 ref_count;
3569         u32 nritems;
3570         int err = 0;
3571         int added = 0;
3572         int counted;
3573         int ret;
3574
3575         ref_root = btrfs_extent_data_ref_root(leaf, ref);
3576         ref_objectid = btrfs_extent_data_ref_objectid(leaf, ref);
3577         ref_offset = btrfs_extent_data_ref_offset(leaf, ref);
3578         ref_count = btrfs_extent_data_ref_count(leaf, ref);
3579
3580         /*
3581          * This is an extent belonging to the free space cache, lets just delete
3582          * it and redo the search.
3583          */
3584         if (ref_root == BTRFS_ROOT_TREE_OBJECTID) {
3585                 ret = delete_block_group_cache(rc->extent_root->fs_info,
3586                                                rc->block_group,
3587                                                NULL, ref_objectid);
3588                 if (ret != -ENOENT)
3589                         return ret;
3590                 ret = 0;
3591         }
3592
3593         path = btrfs_alloc_path();
3594         if (!path)
3595                 return -ENOMEM;
3596         path->reada = READA_FORWARD;
3597
3598         root = read_fs_root(rc->extent_root->fs_info, ref_root);
3599         if (IS_ERR(root)) {
3600                 err = PTR_ERR(root);
3601                 goto out;
3602         }
3603
3604         key.objectid = ref_objectid;
3605         key.type = BTRFS_EXTENT_DATA_KEY;
3606         if (ref_offset > ((u64)-1 << 32))
3607                 key.offset = 0;
3608         else
3609                 key.offset = ref_offset;
3610
3611         path->search_commit_root = 1;
3612         path->skip_locking = 1;
3613         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
3614         if (ret < 0) {
3615                 err = ret;
3616                 goto out;
3617         }
3618
3619         leaf = path->nodes[0];
3620         nritems = btrfs_header_nritems(leaf);
3621         /*
3622          * the references in tree blocks that use full backrefs
3623          * are not counted in
3624          */
3625         if (block_use_full_backref(rc, leaf))
3626                 counted = 0;
3627         else
3628                 counted = 1;
3629         rb_node = tree_search(blocks, leaf->start);
3630         if (rb_node) {
3631                 if (counted)
3632                         added = 1;
3633                 else
3634                         path->slots[0] = nritems;
3635         }
3636
3637         while (ref_count > 0) {
3638                 while (path->slots[0] >= nritems) {
3639                         ret = btrfs_next_leaf(root, path);
3640                         if (ret < 0) {
3641                                 err = ret;
3642                                 goto out;
3643                         }
3644                         if (WARN_ON(ret > 0))
3645                                 goto out;
3646
3647                         leaf = path->nodes[0];
3648                         nritems = btrfs_header_nritems(leaf);
3649                         added = 0;
3650
3651                         if (block_use_full_backref(rc, leaf))
3652                                 counted = 0;
3653                         else
3654                                 counted = 1;
3655                         rb_node = tree_search(blocks, leaf->start);
3656                         if (rb_node) {
3657                                 if (counted)
3658                                         added = 1;
3659                                 else
3660                                         path->slots[0] = nritems;
3661                         }
3662                 }
3663
3664                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
3665                 if (WARN_ON(key.objectid != ref_objectid ||
3666                     key.type != BTRFS_EXTENT_DATA_KEY))
3667                         break;
3668
3669                 fi = btrfs_item_ptr(leaf, path->slots[0],
3670                                     struct btrfs_file_extent_item);
3671
3672                 if (btrfs_file_extent_type(leaf, fi) ==
3673                     BTRFS_FILE_EXTENT_INLINE)
3674                         goto next;
3675
3676                 if (btrfs_file_extent_disk_bytenr(leaf, fi) !=
3677                     extent_key->objectid)
3678                         goto next;
3679
3680                 key.offset -= btrfs_file_extent_offset(leaf, fi);
3681                 if (key.offset != ref_offset)
3682                         goto next;
3683
3684                 if (counted)
3685                         ref_count--;
3686                 if (added)
3687                         goto next;
3688
3689                 if (!tree_block_processed(leaf->start, rc)) {
3690                         block = kmalloc(sizeof(*block), GFP_NOFS);
3691                         if (!block) {
3692                                 err = -ENOMEM;
3693                                 break;
3694                         }
3695                         block->bytenr = leaf->start;
3696                         btrfs_item_key_to_cpu(leaf, &block->key, 0);
3697                         block->level = 0;
3698                         block->key_ready = 1;
3699                         rb_node = tree_insert(blocks, block->bytenr,
3700                                               &block->rb_node);
3701                         if (rb_node)
3702                                 backref_tree_panic(rb_node, -EEXIST,
3703                                                    block->bytenr);
3704                 }
3705                 if (counted)
3706                         added = 1;
3707                 else
3708                         path->slots[0] = nritems;
3709 next:
3710                 path->slots[0]++;
3711
3712         }
3713 out:
3714         btrfs_free_path(path);
3715         return err;
3716 }
3717
3718 /*
3719  * helper to find all tree blocks that reference a given data extent
3720  */
3721 static noinline_for_stack
3722 int add_data_references(struct reloc_control *rc,
3723                         struct btrfs_key *extent_key,
3724                         struct btrfs_path *path,
3725                         struct rb_root *blocks)
3726 {
3727         struct btrfs_key key;
3728         struct extent_buffer *eb;
3729         struct btrfs_extent_data_ref *dref;
3730         struct btrfs_extent_inline_ref *iref;
3731         unsigned long ptr;
3732         unsigned long end;
3733         u32 blocksize = rc->extent_root->nodesize;
3734         int ret = 0;
3735         int err = 0;
3736
3737         eb = path->nodes[0];
3738         ptr = btrfs_item_ptr_offset(eb, path->slots[0]);
3739         end = ptr + btrfs_item_size_nr(eb, path->slots[0]);
3740 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
3741         if (ptr + sizeof(struct btrfs_extent_item_v0) == end)
3742                 ptr = end;
3743         else
3744 #endif
3745                 ptr += sizeof(struct btrfs_extent_item);
3746
3747         while (ptr < end) {
3748                 iref = (struct btrfs_extent_inline_ref *)ptr;
3749                 key.type = btrfs_extent_inline_ref_type(eb, iref);
3750                 if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
3751                         key.offset = btrfs_extent_inline_ref_offset(eb, iref);
3752                         ret = __add_tree_block(rc, key.offset, blocksize,
3753                                                blocks);
3754                 } else if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
3755                         dref = (struct btrfs_extent_data_ref *)(&iref->offset);
3756                         ret = find_data_references(rc, extent_key,
3757                                                    eb, dref, blocks);
3758                 } else {
3759                         BUG();
3760                 }
3761                 if (ret) {
3762                         err = ret;
3763                         goto out;
3764                 }
3765                 ptr += btrfs_extent_inline_ref_size(key.type);
3766         }
3767         WARN_ON(ptr > end);
3768
3769         while (1) {
3770                 cond_resched();
3771                 eb = path->nodes[0];
3772                 if (path->slots[0] >= btrfs_header_nritems(eb)) {
3773                         ret = btrfs_next_leaf(rc->extent_root, path);
3774                         if (ret < 0) {
3775                                 err = ret;
3776                                 break;
3777                         }
3778                         if (ret > 0)
3779                                 break;
3780                         eb = path->nodes[0];
3781                 }
3782
3783                 btrfs_item_key_to_cpu(eb, &key, path->slots[0]);
3784                 if (key.objectid != extent_key->objectid)
3785                         break;
3786
3787 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
3788                 if (key.type == BTRFS_SHARED_DATA_REF_KEY ||
3789                     key.type == BTRFS_EXTENT_REF_V0_KEY) {
3790 #else
3791                 BUG_ON(key.type == BTRFS_EXTENT_REF_V0_KEY);
3792                 if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
3793 #endif
3794                         ret = __add_tree_block(rc, key.offset, blocksize,
3795                                                blocks);
3796                 } else if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
3797                         dref = btrfs_item_ptr(eb, path->slots[0],
3798                                               struct btrfs_extent_data_ref);
3799                         ret = find_data_references(rc, extent_key,
3800                                                    eb, dref, blocks);
3801                 } else {
3802                         ret = 0;
3803                 }
3804                 if (ret) {
3805                         err = ret;
3806                         break;
3807                 }
3808                 path->slots[0]++;
3809         }
3810 out:
3811         btrfs_release_path(path);
3812         if (err)
3813                 free_block_list(blocks);
3814         return err;
3815 }
3816
3817 /*
3818  * helper to find next unprocessed extent
3819  */
3820 static noinline_for_stack
3821 int find_next_extent(struct reloc_control *rc, struct btrfs_path *path,
3822                      struct btrfs_key *extent_key)
3823 {
3824         struct btrfs_key key;
3825         struct extent_buffer *leaf;
3826         u64 start, end, last;
3827         int ret;
3828
3829         last = rc->block_group->key.objectid + rc->block_group->key.offset;
3830         while (1) {
3831                 cond_resched();
3832                 if (rc->search_start >= last) {
3833                         ret = 1;
3834                         break;
3835                 }
3836
3837                 key.objectid = rc->search_start;
3838                 key.type = BTRFS_EXTENT_ITEM_KEY;
3839                 key.offset = 0;
3840
3841                 path->search_commit_root = 1;
3842                 path->skip_locking = 1;
3843                 ret = btrfs_search_slot(NULL, rc->extent_root, &key, path,
3844                                         0, 0);
3845                 if (ret < 0)
3846                         break;
3847 next:
3848                 leaf = path->nodes[0];
3849                 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
3850                         ret = btrfs_next_leaf(rc->extent_root, path);
3851                         if (ret != 0)
3852                                 break;
3853                         leaf = path->nodes[0];
3854                 }
3855
3856                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
3857                 if (key.objectid >= last) {
3858                         ret = 1;
3859                         break;
3860                 }
3861
3862                 if (key.type != BTRFS_EXTENT_ITEM_KEY &&
3863                     key.type != BTRFS_METADATA_ITEM_KEY) {
3864                         path->slots[0]++;
3865                         goto next;
3866                 }
3867
3868                 if (key.type == BTRFS_EXTENT_ITEM_KEY &&
3869                     key.objectid + key.offset <= rc->search_start) {
3870                         path->slots[0]++;
3871                         goto next;
3872                 }
3873
3874                 if (key.type == BTRFS_METADATA_ITEM_KEY &&
3875                     key.objectid + rc->extent_root->nodesize <=
3876                     rc->search_start) {
3877                         path->slots[0]++;
3878                         goto next;
3879                 }
3880
3881                 ret = find_first_extent_bit(&rc->processed_blocks,
3882                                             key.objectid, &start, &end,
3883                                             EXTENT_DIRTY, NULL);
3884
3885                 if (ret == 0 && start <= key.objectid) {
3886                         btrfs_release_path(path);
3887                         rc->search_start = end + 1;
3888                 } else {
3889                         if (key.type == BTRFS_EXTENT_ITEM_KEY)
3890                                 rc->search_start = key.objectid + key.offset;
3891                         else
3892                                 rc->search_start = key.objectid +
3893                                         rc->extent_root->nodesize;
3894                         memcpy(extent_key, &key, sizeof(key));
3895                         return 0;
3896                 }
3897         }
3898         btrfs_release_path(path);
3899         return ret;
3900 }
3901
3902 static void set_reloc_control(struct reloc_control *rc)
3903 {
3904         struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
3905
3906         mutex_lock(&fs_info->reloc_mutex);
3907         fs_info->reloc_ctl = rc;
3908         mutex_unlock(&fs_info->reloc_mutex);
3909 }
3910
3911 static void unset_reloc_control(struct reloc_control *rc)
3912 {
3913         struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
3914
3915         mutex_lock(&fs_info->reloc_mutex);
3916         fs_info->reloc_ctl = NULL;
3917         mutex_unlock(&fs_info->reloc_mutex);
3918 }
3919
3920 static int check_extent_flags(u64 flags)
3921 {
3922         if ((flags & BTRFS_EXTENT_FLAG_DATA) &&
3923             (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK))
3924                 return 1;
3925         if (!(flags & BTRFS_EXTENT_FLAG_DATA) &&
3926             !(flags & BTRFS_EXTENT_FLAG_TREE_BLOCK))
3927                 return 1;
3928         if ((flags & BTRFS_EXTENT_FLAG_DATA) &&
3929             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
3930                 return 1;
3931         return 0;
3932 }
3933
3934 static noinline_for_stack
3935 int prepare_to_relocate(struct reloc_control *rc)
3936 {
3937         struct btrfs_trans_handle *trans;
3938         int ret;
3939
3940         rc->block_rsv = btrfs_alloc_block_rsv(rc->extent_root,
3941                                               BTRFS_BLOCK_RSV_TEMP);
3942         if (!rc->block_rsv)
3943                 return -ENOMEM;
3944
3945         memset(&rc->cluster, 0, sizeof(rc->cluster));
3946         rc->search_start = rc->block_group->key.objectid;
3947         rc->extents_found = 0;
3948         rc->nodes_relocated = 0;
3949         rc->merging_rsv_size = 0;
3950         rc->reserved_bytes = 0;
3951         rc->block_rsv->size = rc->extent_root->nodesize *
3952                               RELOCATION_RESERVED_NODES;
3953         ret = btrfs_block_rsv_refill(rc->extent_root,
3954                                      rc->block_rsv, rc->block_rsv->size,
3955                                      BTRFS_RESERVE_FLUSH_ALL);
3956         if (ret)
3957                 return ret;
3958
3959         rc->create_reloc_tree = 1;
3960         set_reloc_control(rc);
3961
3962         trans = btrfs_join_transaction(rc->extent_root);
3963         if (IS_ERR(trans)) {
3964                 unset_reloc_control(rc);
3965                 /*
3966                  * extent tree is not a ref_cow tree and has no reloc_root to
3967                  * cleanup.  And callers are responsible to free the above
3968                  * block rsv.
3969                  */
3970                 return PTR_ERR(trans);
3971         }
3972         btrfs_commit_transaction(trans, rc->extent_root);
3973         return 0;
3974 }
3975
3976 /*
3977  * Qgroup fixer for data chunk relocation.
3978  * The data relocation is done in the following steps
3979  * 1) Copy data extents into data reloc tree
3980  * 2) Create tree reloc tree(special snapshot) for related subvolumes
3981  * 3) Modify file extents in tree reloc tree
3982  * 4) Merge tree reloc tree with original fs tree, by swapping tree blocks
3983  *
3984  * The problem is, data and tree reloc tree are not accounted to qgroup,
3985  * and 4) will only info qgroup to track tree blocks change, not file extents
3986  * in the tree blocks.
3987  *
3988  * The good news is, related data extents are all in data reloc tree, so we
3989  * only need to info qgroup to track all file extents in data reloc tree
3990  * before commit trans.
3991  */
3992 static int qgroup_fix_relocated_data_extents(struct btrfs_trans_handle *trans,
3993                                              struct reloc_control *rc)
3994 {
3995         struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
3996         struct inode *inode = rc->data_inode;
3997         struct btrfs_root *data_reloc_root = BTRFS_I(inode)->root;
3998         struct btrfs_path *path;
3999         struct btrfs_key key;
4000         int ret = 0;
4001
4002         if (!test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags))
4003                 return 0;
4004
4005         /*
4006          * Only for stage where we update data pointers the qgroup fix is
4007          * valid.
4008          * For MOVING_DATA stage, we will miss the timing of swapping tree
4009          * blocks, and won't fix it.
4010          */
4011         if (!(rc->stage == UPDATE_DATA_PTRS && rc->extents_found))
4012                 return 0;
4013
4014         path = btrfs_alloc_path();
4015         if (!path)
4016                 return -ENOMEM;
4017         key.objectid = btrfs_ino(inode);
4018         key.type = BTRFS_EXTENT_DATA_KEY;
4019         key.offset = 0;
4020
4021         ret = btrfs_search_slot(NULL, data_reloc_root, &key, path, 0, 0);
4022         if (ret < 0)
4023                 goto out;
4024
4025         lock_extent(&BTRFS_I(inode)->io_tree, 0, (u64)-1);
4026         while (1) {
4027                 struct btrfs_file_extent_item *fi;
4028
4029                 btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0]);
4030                 if (key.objectid > btrfs_ino(inode))
4031                         break;
4032                 if (key.type != BTRFS_EXTENT_DATA_KEY)
4033                         goto next;
4034                 fi = btrfs_item_ptr(path->nodes[0], path->slots[0],
4035                                     struct btrfs_file_extent_item);
4036                 if (btrfs_file_extent_type(path->nodes[0], fi) !=
4037                                 BTRFS_FILE_EXTENT_REG)
4038                         goto next;
4039                 ret = btrfs_qgroup_insert_dirty_extent(trans, fs_info,
4040                         btrfs_file_extent_disk_bytenr(path->nodes[0], fi),
4041                         btrfs_file_extent_disk_num_bytes(path->nodes[0], fi),
4042                         GFP_NOFS);
4043                 if (ret < 0)
4044                         break;
4045 next:
4046                 ret = btrfs_next_item(data_reloc_root, path);
4047                 if (ret < 0)
4048                         break;
4049                 if (ret > 0) {
4050                         ret = 0;
4051                         break;
4052                 }
4053         }
4054         unlock_extent(&BTRFS_I(inode)->io_tree, 0 , (u64)-1);
4055 out:
4056         btrfs_free_path(path);
4057         return ret;
4058 }
4059
4060 static noinline_for_stack int relocate_block_group(struct reloc_control *rc)
4061 {
4062         struct rb_root blocks = RB_ROOT;
4063         struct btrfs_key key;
4064         struct btrfs_trans_handle *trans = NULL;
4065         struct btrfs_path *path;
4066         struct btrfs_extent_item *ei;
4067         u64 flags;
4068         u32 item_size;
4069         int ret;
4070         int err = 0;
4071         int progress = 0;
4072
4073         path = btrfs_alloc_path();
4074         if (!path)
4075                 return -ENOMEM;
4076         path->reada = READA_FORWARD;
4077
4078         ret = prepare_to_relocate(rc);
4079         if (ret) {
4080                 err = ret;
4081                 goto out_free;
4082         }
4083
4084         while (1) {
4085                 rc->reserved_bytes = 0;
4086                 ret = btrfs_block_rsv_refill(rc->extent_root,
4087                                         rc->block_rsv, rc->block_rsv->size,
4088                                         BTRFS_RESERVE_FLUSH_ALL);
4089                 if (ret) {
4090                         err = ret;
4091                         break;
4092                 }
4093                 progress++;
4094                 trans = btrfs_start_transaction(rc->extent_root, 0);
4095                 if (IS_ERR(trans)) {
4096                         err = PTR_ERR(trans);
4097                         trans = NULL;
4098                         break;
4099                 }
4100 restart:
4101                 if (update_backref_cache(trans, &rc->backref_cache)) {
4102                         btrfs_end_transaction(trans, rc->extent_root);
4103                         continue;
4104                 }
4105
4106                 ret = find_next_extent(rc, path, &key);
4107                 if (ret < 0)
4108                         err = ret;
4109                 if (ret != 0)
4110                         break;
4111
4112                 rc->extents_found++;
4113
4114                 ei = btrfs_item_ptr(path->nodes[0], path->slots[0],
4115                                     struct btrfs_extent_item);
4116                 item_size = btrfs_item_size_nr(path->nodes[0], path->slots[0]);
4117                 if (item_size >= sizeof(*ei)) {
4118                         flags = btrfs_extent_flags(path->nodes[0], ei);
4119                         ret = check_extent_flags(flags);
4120                         BUG_ON(ret);
4121
4122                 } else {
4123 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
4124                         u64 ref_owner;
4125                         int path_change = 0;
4126
4127                         BUG_ON(item_size !=
4128                                sizeof(struct btrfs_extent_item_v0));
4129                         ret = get_ref_objectid_v0(rc, path, &key, &ref_owner,
4130                                                   &path_change);
4131                         if (ret < 0) {
4132                                 err = ret;
4133                                 break;
4134                         }
4135                         if (ref_owner < BTRFS_FIRST_FREE_OBJECTID)
4136                                 flags = BTRFS_EXTENT_FLAG_TREE_BLOCK;
4137                         else
4138                                 flags = BTRFS_EXTENT_FLAG_DATA;
4139
4140                         if (path_change) {
4141                                 btrfs_release_path(path);
4142
4143                                 path->search_commit_root = 1;
4144                                 path->skip_locking = 1;
4145                                 ret = btrfs_search_slot(NULL, rc->extent_root,
4146                                                         &key, path, 0, 0);
4147                                 if (ret < 0) {
4148                                         err = ret;
4149                                         break;
4150                                 }
4151                                 BUG_ON(ret > 0);
4152                         }
4153 #else
4154                         BUG();
4155 #endif
4156                 }
4157
4158                 if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
4159                         ret = add_tree_block(rc, &key, path, &blocks);
4160                 } else if (rc->stage == UPDATE_DATA_PTRS &&
4161                            (flags & BTRFS_EXTENT_FLAG_DATA)) {
4162                         ret = add_data_references(rc, &key, path, &blocks);
4163                 } else {
4164                         btrfs_release_path(path);
4165                         ret = 0;
4166                 }
4167                 if (ret < 0) {
4168                         err = ret;
4169                         break;
4170                 }
4171
4172                 if (!RB_EMPTY_ROOT(&blocks)) {
4173                         ret = relocate_tree_blocks(trans, rc, &blocks);
4174                         if (ret < 0) {
4175                                 /*
4176                                  * if we fail to relocate tree blocks, force to update
4177                                  * backref cache when committing transaction.
4178                                  */
4179                                 rc->backref_cache.last_trans = trans->transid - 1;
4180
4181                                 if (ret != -EAGAIN) {
4182                                         err = ret;
4183                                         break;
4184                                 }
4185                                 rc->extents_found--;
4186                                 rc->search_start = key.objectid;
4187                         }
4188                 }
4189
4190                 btrfs_end_transaction_throttle(trans, rc->extent_root);
4191                 btrfs_btree_balance_dirty(rc->extent_root);
4192                 trans = NULL;
4193
4194                 if (rc->stage == MOVE_DATA_EXTENTS &&
4195                     (flags & BTRFS_EXTENT_FLAG_DATA)) {
4196                         rc->found_file_extent = 1;
4197                         ret = relocate_data_extent(rc->data_inode,
4198                                                    &key, &rc->cluster);
4199                         if (ret < 0) {
4200                                 err = ret;
4201                                 break;
4202                         }
4203                 }
4204         }
4205         if (trans && progress && err == -ENOSPC) {
4206                 ret = btrfs_force_chunk_alloc(trans, rc->extent_root,
4207                                               rc->block_group->flags);
4208                 if (ret == 1) {
4209                         err = 0;
4210                         progress = 0;
4211                         goto restart;
4212                 }
4213         }
4214
4215         btrfs_release_path(path);
4216         clear_extent_bits(&rc->processed_blocks, 0, (u64)-1, EXTENT_DIRTY);
4217
4218         if (trans) {
4219                 btrfs_end_transaction_throttle(trans, rc->extent_root);
4220                 btrfs_btree_balance_dirty(rc->extent_root);
4221         }
4222
4223         if (!err) {
4224                 ret = relocate_file_extent_cluster(rc->data_inode,
4225                                                    &rc->cluster);
4226                 if (ret < 0)
4227                         err = ret;
4228         }
4229
4230         rc->create_reloc_tree = 0;
4231         set_reloc_control(rc);
4232
4233         backref_cache_cleanup(&rc->backref_cache);
4234         btrfs_block_rsv_release(rc->extent_root, rc->block_rsv, (u64)-1);
4235
4236         err = prepare_to_merge(rc, err);
4237
4238         merge_reloc_roots(rc);
4239
4240         rc->merge_reloc_tree = 0;
4241         unset_reloc_control(rc);
4242         btrfs_block_rsv_release(rc->extent_root, rc->block_rsv, (u64)-1);
4243
4244         /* get rid of pinned extents */
4245         trans = btrfs_join_transaction(rc->extent_root);
4246         if (IS_ERR(trans)) {
4247                 err = PTR_ERR(trans);
4248                 goto out_free;
4249         }
4250         ret = qgroup_fix_relocated_data_extents(trans, rc);
4251         if (ret < 0) {
4252                 btrfs_abort_transaction(trans, ret);
4253                 if (!err)
4254                         err = ret;
4255                 goto out_free;
4256         }
4257         btrfs_commit_transaction(trans, rc->extent_root);
4258 out_free:
4259         btrfs_free_block_rsv(rc->extent_root, rc->block_rsv);
4260         btrfs_free_path(path);
4261         return err;
4262 }
4263
4264 static int __insert_orphan_inode(struct btrfs_trans_handle *trans,
4265                                  struct btrfs_root *root, u64 objectid)
4266 {
4267         struct btrfs_path *path;
4268         struct btrfs_inode_item *item;
4269         struct extent_buffer *leaf;
4270         int ret;
4271
4272         path = btrfs_alloc_path();
4273         if (!path)
4274                 return -ENOMEM;
4275
4276         ret = btrfs_insert_empty_inode(trans, root, path, objectid);
4277         if (ret)
4278                 goto out;
4279
4280         leaf = path->nodes[0];
4281         item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_inode_item);
4282         memset_extent_buffer(leaf, 0, (unsigned long)item, sizeof(*item));
4283         btrfs_set_inode_generation(leaf, item, 1);
4284         btrfs_set_inode_size(leaf, item, 0);
4285         btrfs_set_inode_mode(leaf, item, S_IFREG | 0600);
4286         btrfs_set_inode_flags(leaf, item, BTRFS_INODE_NOCOMPRESS |
4287                                           BTRFS_INODE_PREALLOC);
4288         btrfs_mark_buffer_dirty(leaf);
4289 out:
4290         btrfs_free_path(path);
4291         return ret;
4292 }
4293
4294 /*
4295  * helper to create inode for data relocation.
4296  * the inode is in data relocation tree and its link count is 0
4297  */
4298 static noinline_for_stack
4299 struct inode *create_reloc_inode(struct btrfs_fs_info *fs_info,
4300                                  struct btrfs_block_group_cache *group)
4301 {
4302         struct inode *inode = NULL;
4303         struct btrfs_trans_handle *trans;
4304         struct btrfs_root *root;
4305         struct btrfs_key key;
4306         u64 objectid;
4307         int err = 0;
4308
4309         root = read_fs_root(fs_info, BTRFS_DATA_RELOC_TREE_OBJECTID);
4310         if (IS_ERR(root))
4311                 return ERR_CAST(root);
4312
4313         trans = btrfs_start_transaction(root, 6);
4314         if (IS_ERR(trans))
4315                 return ERR_CAST(trans);
4316
4317         err = btrfs_find_free_objectid(root, &objectid);
4318         if (err)
4319                 goto out;
4320
4321         err = __insert_orphan_inode(trans, root, objectid);
4322         BUG_ON(err);
4323
4324         key.objectid = objectid;
4325         key.type = BTRFS_INODE_ITEM_KEY;
4326         key.offset = 0;
4327         inode = btrfs_iget(root->fs_info->sb, &key, root, NULL);
4328         BUG_ON(IS_ERR(inode) || is_bad_inode(inode));
4329         BTRFS_I(inode)->index_cnt = group->key.objectid;
4330
4331         err = btrfs_orphan_add(trans, inode);
4332 out:
4333         btrfs_end_transaction(trans, root);
4334         btrfs_btree_balance_dirty(root);
4335         if (err) {
4336                 if (inode)
4337                         iput(inode);
4338                 inode = ERR_PTR(err);
4339         }
4340         return inode;
4341 }
4342
4343 static struct reloc_control *alloc_reloc_control(struct btrfs_fs_info *fs_info)
4344 {
4345         struct reloc_control *rc;
4346
4347         rc = kzalloc(sizeof(*rc), GFP_NOFS);
4348         if (!rc)
4349                 return NULL;
4350
4351         INIT_LIST_HEAD(&rc->reloc_roots);
4352         backref_cache_init(&rc->backref_cache);
4353         mapping_tree_init(&rc->reloc_root_tree);
4354         extent_io_tree_init(&rc->processed_blocks,
4355                             fs_info->btree_inode->i_mapping);
4356         return rc;
4357 }
4358
4359 /*
4360  * function to relocate all extents in a block group.
4361  */
4362 int btrfs_relocate_block_group(struct btrfs_root *extent_root, u64 group_start)
4363 {
4364         struct btrfs_fs_info *fs_info = extent_root->fs_info;
4365         struct reloc_control *rc;
4366         struct inode *inode;
4367         struct btrfs_path *path;
4368         int ret;
4369         int rw = 0;
4370         int err = 0;
4371
4372         rc = alloc_reloc_control(fs_info);
4373         if (!rc)
4374                 return -ENOMEM;
4375
4376         rc->extent_root = extent_root;
4377
4378         rc->block_group = btrfs_lookup_block_group(fs_info, group_start);
4379         BUG_ON(!rc->block_group);
4380
4381         ret = btrfs_inc_block_group_ro(extent_root, rc->block_group);
4382         if (ret) {
4383                 err = ret;
4384                 goto out;
4385         }
4386         rw = 1;
4387
4388         path = btrfs_alloc_path();
4389         if (!path) {
4390                 err = -ENOMEM;
4391                 goto out;
4392         }
4393
4394         inode = lookup_free_space_inode(fs_info->tree_root, rc->block_group,
4395                                         path);
4396         btrfs_free_path(path);
4397
4398         if (!IS_ERR(inode))
4399                 ret = delete_block_group_cache(fs_info, rc->block_group, inode, 0);
4400         else
4401                 ret = PTR_ERR(inode);
4402
4403         if (ret && ret != -ENOENT) {
4404                 err = ret;
4405                 goto out;
4406         }
4407
4408         rc->data_inode = create_reloc_inode(fs_info, rc->block_group);
4409         if (IS_ERR(rc->data_inode)) {
4410                 err = PTR_ERR(rc->data_inode);
4411                 rc->data_inode = NULL;
4412                 goto out;
4413         }
4414
4415         btrfs_info(extent_root->fs_info,
4416                    "relocating block group %llu flags %llu",
4417                    rc->block_group->key.objectid, rc->block_group->flags);
4418
4419         btrfs_wait_block_group_reservations(rc->block_group);
4420         btrfs_wait_nocow_writers(rc->block_group);
4421         btrfs_wait_ordered_roots(fs_info, -1,
4422                                  rc->block_group->key.objectid,
4423                                  rc->block_group->key.offset);
4424
4425         while (1) {
4426                 mutex_lock(&fs_info->cleaner_mutex);
4427                 ret = relocate_block_group(rc);
4428                 mutex_unlock(&fs_info->cleaner_mutex);
4429                 if (ret < 0) {
4430                         err = ret;
4431                         goto out;
4432                 }
4433
4434                 if (rc->extents_found == 0)
4435                         break;
4436
4437                 btrfs_info(extent_root->fs_info, "found %llu extents",
4438                         rc->extents_found);
4439
4440                 if (rc->stage == MOVE_DATA_EXTENTS && rc->found_file_extent) {
4441                         ret = btrfs_wait_ordered_range(rc->data_inode, 0,
4442                                                        (u64)-1);
4443                         if (ret) {
4444                                 err = ret;
4445                                 goto out;
4446                         }
4447                         invalidate_mapping_pages(rc->data_inode->i_mapping,
4448                                                  0, -1);
4449                         rc->stage = UPDATE_DATA_PTRS;
4450                 }
4451         }
4452
4453         WARN_ON(rc->block_group->pinned > 0);
4454         WARN_ON(rc->block_group->reserved > 0);
4455         WARN_ON(btrfs_block_group_used(&rc->block_group->item) > 0);
4456 out:
4457         if (err && rw)
4458                 btrfs_dec_block_group_ro(extent_root, rc->block_group);
4459         iput(rc->data_inode);
4460         btrfs_put_block_group(rc->block_group);
4461         kfree(rc);
4462         return err;
4463 }
4464
4465 static noinline_for_stack int mark_garbage_root(struct btrfs_root *root)
4466 {
4467         struct btrfs_trans_handle *trans;
4468         int ret, err;
4469
4470         trans = btrfs_start_transaction(root->fs_info->tree_root, 0);
4471         if (IS_ERR(trans))
4472                 return PTR_ERR(trans);
4473
4474         memset(&root->root_item.drop_progress, 0,
4475                 sizeof(root->root_item.drop_progress));
4476         root->root_item.drop_level = 0;
4477         btrfs_set_root_refs(&root->root_item, 0);
4478         ret = btrfs_update_root(trans, root->fs_info->tree_root,
4479                                 &root->root_key, &root->root_item);
4480
4481         err = btrfs_end_transaction(trans, root->fs_info->tree_root);
4482         if (err)
4483                 return err;
4484         return ret;
4485 }
4486
4487 /*
4488  * recover relocation interrupted by system crash.
4489  *
4490  * this function resumes merging reloc trees with corresponding fs trees.
4491  * this is important for keeping the sharing of tree blocks
4492  */
4493 int btrfs_recover_relocation(struct btrfs_root *root)
4494 {
4495         LIST_HEAD(reloc_roots);
4496         struct btrfs_key key;
4497         struct btrfs_root *fs_root;
4498         struct btrfs_root *reloc_root;
4499         struct btrfs_path *path;
4500         struct extent_buffer *leaf;
4501         struct reloc_control *rc = NULL;
4502         struct btrfs_trans_handle *trans;
4503         int ret;
4504         int err = 0;
4505
4506         path = btrfs_alloc_path();
4507         if (!path)
4508                 return -ENOMEM;
4509         path->reada = READA_BACK;
4510
4511         key.objectid = BTRFS_TREE_RELOC_OBJECTID;
4512         key.type = BTRFS_ROOT_ITEM_KEY;
4513         key.offset = (u64)-1;
4514
4515         while (1) {
4516                 ret = btrfs_search_slot(NULL, root->fs_info->tree_root, &key,
4517                                         path, 0, 0);
4518                 if (ret < 0) {
4519                         err = ret;
4520                         goto out;
4521                 }
4522                 if (ret > 0) {
4523                         if (path->slots[0] == 0)
4524                                 break;
4525                         path->slots[0]--;
4526                 }
4527                 leaf = path->nodes[0];
4528                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
4529                 btrfs_release_path(path);
4530
4531                 if (key.objectid != BTRFS_TREE_RELOC_OBJECTID ||
4532                     key.type != BTRFS_ROOT_ITEM_KEY)
4533                         break;
4534
4535                 reloc_root = btrfs_read_fs_root(root, &key);
4536                 if (IS_ERR(reloc_root)) {
4537                         err = PTR_ERR(reloc_root);
4538                         goto out;
4539                 }
4540
4541                 list_add(&reloc_root->root_list, &reloc_roots);
4542
4543                 if (btrfs_root_refs(&reloc_root->root_item) > 0) {
4544                         fs_root = read_fs_root(root->fs_info,
4545                                                reloc_root->root_key.offset);
4546                         if (IS_ERR(fs_root)) {
4547                                 ret = PTR_ERR(fs_root);
4548                                 if (ret != -ENOENT) {
4549                                         err = ret;
4550                                         goto out;
4551                                 }
4552                                 ret = mark_garbage_root(reloc_root);
4553                                 if (ret < 0) {
4554                                         err = ret;
4555                                         goto out;
4556                                 }
4557                         }
4558                 }
4559
4560                 if (key.offset == 0)
4561                         break;
4562
4563                 key.offset--;
4564         }
4565         btrfs_release_path(path);
4566
4567         if (list_empty(&reloc_roots))
4568                 goto out;
4569
4570         rc = alloc_reloc_control(root->fs_info);
4571         if (!rc) {
4572                 err = -ENOMEM;
4573                 goto out;
4574         }
4575
4576         rc->extent_root = root->fs_info->extent_root;
4577
4578         set_reloc_control(rc);
4579
4580         trans = btrfs_join_transaction(rc->extent_root);
4581         if (IS_ERR(trans)) {
4582                 unset_reloc_control(rc);
4583                 err = PTR_ERR(trans);
4584                 goto out_free;
4585         }
4586
4587         rc->merge_reloc_tree = 1;
4588
4589         while (!list_empty(&reloc_roots)) {
4590                 reloc_root = list_entry(reloc_roots.next,
4591                                         struct btrfs_root, root_list);
4592                 list_del(&reloc_root->root_list);
4593
4594                 if (btrfs_root_refs(&reloc_root->root_item) == 0) {
4595                         list_add_tail(&reloc_root->root_list,
4596                                       &rc->reloc_roots);
4597                         continue;
4598                 }
4599
4600                 fs_root = read_fs_root(root->fs_info,
4601                                        reloc_root->root_key.offset);
4602                 if (IS_ERR(fs_root)) {
4603                         err = PTR_ERR(fs_root);
4604                         list_add_tail(&reloc_root->root_list, &reloc_roots);
4605                         goto out_free;
4606                 }
4607
4608                 err = __add_reloc_root(reloc_root);
4609                 BUG_ON(err < 0); /* -ENOMEM or logic error */
4610                 fs_root->reloc_root = reloc_root;
4611         }
4612
4613         err = btrfs_commit_transaction(trans, rc->extent_root);
4614         if (err)
4615                 goto out_free;
4616
4617         merge_reloc_roots(rc);
4618
4619         unset_reloc_control(rc);
4620
4621         trans = btrfs_join_transaction(rc->extent_root);
4622         if (IS_ERR(trans)) {
4623                 err = PTR_ERR(trans);
4624                 goto out_free;
4625         }
4626         err = qgroup_fix_relocated_data_extents(trans, rc);
4627         if (err < 0) {
4628                 btrfs_abort_transaction(trans, err);
4629                 goto out_free;
4630         }
4631         err = btrfs_commit_transaction(trans, rc->extent_root);
4632 out_free:
4633         kfree(rc);
4634 out:
4635         if (!list_empty(&reloc_roots))
4636                 free_reloc_roots(&reloc_roots);
4637
4638         btrfs_free_path(path);
4639
4640         if (err == 0) {
4641                 /* cleanup orphan inode in data relocation tree */
4642                 fs_root = read_fs_root(root->fs_info,
4643                                        BTRFS_DATA_RELOC_TREE_OBJECTID);
4644                 if (IS_ERR(fs_root))
4645                         err = PTR_ERR(fs_root);
4646                 else
4647                         err = btrfs_orphan_cleanup(fs_root);
4648         }
4649         return err;
4650 }
4651
4652 /*
4653  * helper to add ordered checksum for data relocation.
4654  *
4655  * cloning checksum properly handles the nodatasum extents.
4656  * it also saves CPU time to re-calculate the checksum.
4657  */
4658 int btrfs_reloc_clone_csums(struct inode *inode, u64 file_pos, u64 len)
4659 {
4660         struct btrfs_ordered_sum *sums;
4661         struct btrfs_ordered_extent *ordered;
4662         struct btrfs_root *root = BTRFS_I(inode)->root;
4663         int ret;
4664         u64 disk_bytenr;
4665         u64 new_bytenr;
4666         LIST_HEAD(list);
4667
4668         ordered = btrfs_lookup_ordered_extent(inode, file_pos);
4669         BUG_ON(ordered->file_offset != file_pos || ordered->len != len);
4670
4671         disk_bytenr = file_pos + BTRFS_I(inode)->index_cnt;
4672         ret = btrfs_lookup_csums_range(root->fs_info->csum_root, disk_bytenr,
4673                                        disk_bytenr + len - 1, &list, 0);
4674         if (ret)
4675                 goto out;
4676
4677         while (!list_empty(&list)) {
4678                 sums = list_entry(list.next, struct btrfs_ordered_sum, list);
4679                 list_del_init(&sums->list);
4680
4681                 /*
4682                  * We need to offset the new_bytenr based on where the csum is.
4683                  * We need to do this because we will read in entire prealloc
4684                  * extents but we may have written to say the middle of the
4685                  * prealloc extent, so we need to make sure the csum goes with
4686                  * the right disk offset.
4687                  *
4688                  * We can do this because the data reloc inode refers strictly
4689                  * to the on disk bytes, so we don't have to worry about
4690                  * disk_len vs real len like with real inodes since it's all
4691                  * disk length.
4692                  */
4693                 new_bytenr = ordered->start + (sums->bytenr - disk_bytenr);
4694                 sums->bytenr = new_bytenr;
4695
4696                 btrfs_add_ordered_sum(inode, ordered, sums);
4697         }
4698 out:
4699         btrfs_put_ordered_extent(ordered);
4700         return ret;
4701 }
4702
4703 int btrfs_reloc_cow_block(struct btrfs_trans_handle *trans,
4704                           struct btrfs_root *root, struct extent_buffer *buf,
4705                           struct extent_buffer *cow)
4706 {
4707         struct reloc_control *rc;
4708         struct backref_node *node;
4709         int first_cow = 0;
4710         int level;
4711         int ret = 0;
4712
4713         rc = root->fs_info->reloc_ctl;
4714         if (!rc)
4715                 return 0;
4716
4717         BUG_ON(rc->stage == UPDATE_DATA_PTRS &&
4718                root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID);
4719
4720         level = btrfs_header_level(buf);
4721         if (btrfs_header_generation(buf) <=
4722             btrfs_root_last_snapshot(&root->root_item))
4723                 first_cow = 1;
4724
4725         if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID &&
4726             rc->create_reloc_tree) {
4727                 WARN_ON(!first_cow && level == 0);
4728
4729                 node = rc->backref_cache.path[level];
4730                 BUG_ON(node->bytenr != buf->start &&
4731                        node->new_bytenr != buf->start);
4732
4733                 drop_node_buffer(node);
4734                 extent_buffer_get(cow);
4735                 node->eb = cow;
4736                 node->new_bytenr = cow->start;
4737
4738                 if (!node->pending) {
4739                         list_move_tail(&node->list,
4740                                        &rc->backref_cache.pending[level]);
4741                         node->pending = 1;
4742                 }
4743
4744                 if (first_cow)
4745                         __mark_block_processed(rc, node);
4746
4747                 if (first_cow && level > 0)
4748                         rc->nodes_relocated += buf->len;
4749         }
4750
4751         if (level == 0 && first_cow && rc->stage == UPDATE_DATA_PTRS)
4752                 ret = replace_file_extents(trans, rc, root, cow);
4753         return ret;
4754 }
4755
4756 /*
4757  * called before creating snapshot. it calculates metadata reservation
4758  * required for relocating tree blocks in the snapshot
4759  */
4760 void btrfs_reloc_pre_snapshot(struct btrfs_pending_snapshot *pending,
4761                               u64 *bytes_to_reserve)
4762 {
4763         struct btrfs_root *root;
4764         struct reloc_control *rc;
4765
4766         root = pending->root;
4767         if (!root->reloc_root)
4768                 return;
4769
4770         rc = root->fs_info->reloc_ctl;
4771         if (!rc->merge_reloc_tree)
4772                 return;
4773
4774         root = root->reloc_root;
4775         BUG_ON(btrfs_root_refs(&root->root_item) == 0);
4776         /*
4777          * relocation is in the stage of merging trees. the space
4778          * used by merging a reloc tree is twice the size of
4779          * relocated tree nodes in the worst case. half for cowing
4780          * the reloc tree, half for cowing the fs tree. the space
4781          * used by cowing the reloc tree will be freed after the
4782          * tree is dropped. if we create snapshot, cowing the fs
4783          * tree may use more space than it frees. so we need
4784          * reserve extra space.
4785          */
4786         *bytes_to_reserve += rc->nodes_relocated;
4787 }
4788
4789 /*
4790  * called after snapshot is created. migrate block reservation
4791  * and create reloc root for the newly created snapshot
4792  */
4793 int btrfs_reloc_post_snapshot(struct btrfs_trans_handle *trans,
4794                                struct btrfs_pending_snapshot *pending)
4795 {
4796         struct btrfs_root *root = pending->root;
4797         struct btrfs_root *reloc_root;
4798         struct btrfs_root *new_root;
4799         struct reloc_control *rc;
4800         int ret;
4801
4802         if (!root->reloc_root)
4803                 return 0;
4804
4805         rc = root->fs_info->reloc_ctl;
4806         rc->merging_rsv_size += rc->nodes_relocated;
4807
4808         if (rc->merge_reloc_tree) {
4809                 ret = btrfs_block_rsv_migrate(&pending->block_rsv,
4810                                               rc->block_rsv,
4811                                               rc->nodes_relocated, 1);
4812                 if (ret)
4813                         return ret;
4814         }
4815
4816         new_root = pending->snap;
4817         reloc_root = create_reloc_root(trans, root->reloc_root,
4818                                        new_root->root_key.objectid);
4819         if (IS_ERR(reloc_root))
4820                 return PTR_ERR(reloc_root);
4821
4822         ret = __add_reloc_root(reloc_root);
4823         BUG_ON(ret < 0);
4824         new_root->reloc_root = reloc_root;
4825
4826         if (rc->create_reloc_tree)
4827                 ret = clone_backref_node(trans, rc, root, reloc_root);
4828         return ret;
4829 }