1 /*
2 * Copyright (C) 2009 Oracle. All rights reserved.
3 *
4 * This program is free software; you can redistribute it and/or
5 * modify it under the terms of the GNU General Public
6 * License v2 as published by the Free Software Foundation.
7 *
8 * This program is distributed in the hope that it will be useful,
9 * but WITHOUT ANY WARRANTY; without even the implied warranty of
10 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
11 * General Public License for more details.
12 *
13 * You should have received a copy of the GNU General Public
14 * License along with this program; if not, write to the
15 * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
16 * Boston, MA 021110-1307, USA.
17 */
18
19 #include <linux/sched.h>
20 #include <linux/pagemap.h>
21 #include <linux/writeback.h>
22 #include <linux/blkdev.h>
23 #include <linux/rbtree.h>
24 #include <linux/slab.h>
25 #include "ctree.h"
26 #include "disk-io.h"
27 #include "transaction.h"
28 #include "volumes.h"
29 #include "locking.h"
30 #include "btrfs_inode.h"
31 #include "async-thread.h"
32 #include "free-space-cache.h"
33 #include "inode-map.h"
34
35 /*
36 * backref_node, mapping_node and tree_block start with this
37 */
38 struct tree_entry {
39 struct rb_node rb_node;
40 u64 bytenr;
41 };
42
43 /*
44 * present a tree block in the backref cache
45 */
46 struct backref_node {
47 struct rb_node rb_node;
48 u64 bytenr;
49
50 u64 new_bytenr;
51 /* objectid of tree block owner, can be not uptodate */
52 u64 owner;
53 /* link to pending, changed or detached list */
54 struct list_head list;
55 /* list of upper level blocks reference this block */
56 struct list_head upper;
57 /* list of child blocks in the cache */
58 struct list_head lower;
59 /* NULL if this node is not tree root */
60 struct btrfs_root *root;
61 /* extent buffer got by COW the block */
62 struct extent_buffer *eb;
63 /* level of tree block */
64 unsigned int level:8;
65 /* is the block in non-reference counted tree */
66 unsigned int cowonly:1;
67 /* 1 if no child node in the cache */
68 unsigned int lowest:1;
69 /* is the extent buffer locked */
70 unsigned int locked:1;
71 /* has the block been processed */
72 unsigned int processed:1;
73 /* have backrefs of this block been checked */
74 unsigned int checked:1;
75 /*
76 * 1 if corresponding block has been cowed but some upper
77 * level block pointers may not point to the new location
78 */
79 unsigned int pending:1;
80 /*
81 * 1 if the backref node isn't connected to any other
82 * backref node.
83 */
84 unsigned int detached:1;
85 };
86
87 /*
88 * present a block pointer in the backref cache
89 */
90 struct backref_edge {
91 struct list_head list[2];
92 struct backref_node *node[2];
93 };
94
95 #define LOWER 0
96 #define UPPER 1
97
98 struct backref_cache {
99 /* red black tree of all backref nodes in the cache */
100 struct rb_root rb_root;
101 /* for passing backref nodes to btrfs_reloc_cow_block */
102 struct backref_node *path[BTRFS_MAX_LEVEL];
103 /*
104 * list of blocks that have been cowed but some block
105 * pointers in upper level blocks may not reflect the
106 * new location
107 */
108 struct list_head pending[BTRFS_MAX_LEVEL];
109 /* list of backref nodes with no child node */
110 struct list_head leaves;
111 /* list of blocks that have been cowed in current transaction */
112 struct list_head changed;
113 /* list of detached backref node. */
114 struct list_head detached;
115
116 u64 last_trans;
117
118 int nr_nodes;
119 int nr_edges;
120 };
121
122 /*
123 * map address of tree root to tree
124 */
125 struct mapping_node {
126 struct rb_node rb_node;
127 u64 bytenr;
128 void *data;
129 };
130
131 struct mapping_tree {
132 struct rb_root rb_root;
133 spinlock_t lock;
134 };
135
136 /*
137 * present a tree block to process
138 */
139 struct tree_block {
140 struct rb_node rb_node;
141 u64 bytenr;
142 struct btrfs_key key;
143 unsigned int level:8;
144 unsigned int key_ready:1;
145 };
146
147 #define MAX_EXTENTS 128
148
149 struct file_extent_cluster {
150 u64 start;
151 u64 end;
152 u64 boundary[MAX_EXTENTS];
153 unsigned int nr;
154 };
155
156 struct reloc_control {
157 /* block group to relocate */
158 struct btrfs_block_group_cache *block_group;
159 /* extent tree */
160 struct btrfs_root *extent_root;
161 /* inode for moving data */
162 struct inode *data_inode;
163
164 struct btrfs_block_rsv *block_rsv;
165
166 struct backref_cache backref_cache;
167
168 struct file_extent_cluster cluster;
169 /* tree blocks have been processed */
170 struct extent_io_tree processed_blocks;
171 /* map start of tree root to corresponding reloc tree */
172 struct mapping_tree reloc_root_tree;
173 /* list of reloc trees */
174 struct list_head reloc_roots;
175 /* size of metadata reservation for merging reloc trees */
176 u64 merging_rsv_size;
177 /* size of relocated tree nodes */
178 u64 nodes_relocated;
179
180 u64 search_start;
181 u64 extents_found;
182
183 unsigned int stage:8;
184 unsigned int create_reloc_tree:1;
185 unsigned int merge_reloc_tree:1;
186 unsigned int found_file_extent:1;
187 unsigned int commit_transaction:1;
188 };
189
190 /* stages of data relocation */
191 #define MOVE_DATA_EXTENTS 0
192 #define UPDATE_DATA_PTRS 1
193
194 static void remove_backref_node(struct backref_cache *cache,
195 struct backref_node *node);
196 static void __mark_block_processed(struct reloc_control *rc,
197 struct backref_node *node);
198
mapping_tree_init(struct mapping_tree * tree)199 static void mapping_tree_init(struct mapping_tree *tree)
200 {
201 tree->rb_root = RB_ROOT;
202 spin_lock_init(&tree->lock);
203 }
204
backref_cache_init(struct backref_cache * cache)205 static void backref_cache_init(struct backref_cache *cache)
206 {
207 int i;
208 cache->rb_root = RB_ROOT;
209 for (i = 0; i < BTRFS_MAX_LEVEL; i++)
210 INIT_LIST_HEAD(&cache->pending[i]);
211 INIT_LIST_HEAD(&cache->changed);
212 INIT_LIST_HEAD(&cache->detached);
213 INIT_LIST_HEAD(&cache->leaves);
214 }
215
backref_cache_cleanup(struct backref_cache * cache)216 static void backref_cache_cleanup(struct backref_cache *cache)
217 {
218 struct backref_node *node;
219 int i;
220
221 while (!list_empty(&cache->detached)) {
222 node = list_entry(cache->detached.next,
223 struct backref_node, list);
224 remove_backref_node(cache, node);
225 }
226
227 while (!list_empty(&cache->leaves)) {
228 node = list_entry(cache->leaves.next,
229 struct backref_node, lower);
230 remove_backref_node(cache, node);
231 }
232
233 cache->last_trans = 0;
234
235 for (i = 0; i < BTRFS_MAX_LEVEL; i++)
236 BUG_ON(!list_empty(&cache->pending[i]));
237 BUG_ON(!list_empty(&cache->changed));
238 BUG_ON(!list_empty(&cache->detached));
239 BUG_ON(!RB_EMPTY_ROOT(&cache->rb_root));
240 BUG_ON(cache->nr_nodes);
241 BUG_ON(cache->nr_edges);
242 }
243
alloc_backref_node(struct backref_cache * cache)244 static struct backref_node *alloc_backref_node(struct backref_cache *cache)
245 {
246 struct backref_node *node;
247
248 node = kzalloc(sizeof(*node), GFP_NOFS);
249 if (node) {
250 INIT_LIST_HEAD(&node->list);
251 INIT_LIST_HEAD(&node->upper);
252 INIT_LIST_HEAD(&node->lower);
253 RB_CLEAR_NODE(&node->rb_node);
254 cache->nr_nodes++;
255 }
256 return node;
257 }
258
free_backref_node(struct backref_cache * cache,struct backref_node * node)259 static void free_backref_node(struct backref_cache *cache,
260 struct backref_node *node)
261 {
262 if (node) {
263 cache->nr_nodes--;
264 kfree(node);
265 }
266 }
267
alloc_backref_edge(struct backref_cache * cache)268 static struct backref_edge *alloc_backref_edge(struct backref_cache *cache)
269 {
270 struct backref_edge *edge;
271
272 edge = kzalloc(sizeof(*edge), GFP_NOFS);
273 if (edge)
274 cache->nr_edges++;
275 return edge;
276 }
277
free_backref_edge(struct backref_cache * cache,struct backref_edge * edge)278 static void free_backref_edge(struct backref_cache *cache,
279 struct backref_edge *edge)
280 {
281 if (edge) {
282 cache->nr_edges--;
283 kfree(edge);
284 }
285 }
286
tree_insert(struct rb_root * root,u64 bytenr,struct rb_node * node)287 static struct rb_node *tree_insert(struct rb_root *root, u64 bytenr,
288 struct rb_node *node)
289 {
290 struct rb_node **p = &root->rb_node;
291 struct rb_node *parent = NULL;
292 struct tree_entry *entry;
293
294 while (*p) {
295 parent = *p;
296 entry = rb_entry(parent, struct tree_entry, rb_node);
297
298 if (bytenr < entry->bytenr)
299 p = &(*p)->rb_left;
300 else if (bytenr > entry->bytenr)
301 p = &(*p)->rb_right;
302 else
303 return parent;
304 }
305
306 rb_link_node(node, parent, p);
307 rb_insert_color(node, root);
308 return NULL;
309 }
310
tree_search(struct rb_root * root,u64 bytenr)311 static struct rb_node *tree_search(struct rb_root *root, u64 bytenr)
312 {
313 struct rb_node *n = root->rb_node;
314 struct tree_entry *entry;
315
316 while (n) {
317 entry = rb_entry(n, struct tree_entry, rb_node);
318
319 if (bytenr < entry->bytenr)
320 n = n->rb_left;
321 else if (bytenr > entry->bytenr)
322 n = n->rb_right;
323 else
324 return n;
325 }
326 return NULL;
327 }
328
backref_tree_panic(struct rb_node * rb_node,int errno,u64 bytenr)329 static void backref_tree_panic(struct rb_node *rb_node, int errno, u64 bytenr)
330 {
331
332 struct btrfs_fs_info *fs_info = NULL;
333 struct backref_node *bnode = rb_entry(rb_node, struct backref_node,
334 rb_node);
335 if (bnode->root)
336 fs_info = bnode->root->fs_info;
337 btrfs_panic(fs_info, errno, "Inconsistency in backref cache "
338 "found at offset %llu\n", (unsigned long long)bytenr);
339 }
340
341 /*
342 * walk up backref nodes until reach node presents tree root
343 */
walk_up_backref(struct backref_node * node,struct backref_edge * edges[],int * index)344 static struct backref_node *walk_up_backref(struct backref_node *node,
345 struct backref_edge *edges[],
346 int *index)
347 {
348 struct backref_edge *edge;
349 int idx = *index;
350
351 while (!list_empty(&node->upper)) {
352 edge = list_entry(node->upper.next,
353 struct backref_edge, list[LOWER]);
354 edges[idx++] = edge;
355 node = edge->node[UPPER];
356 }
357 BUG_ON(node->detached);
358 *index = idx;
359 return node;
360 }
361
362 /*
363 * walk down backref nodes to find start of next reference path
364 */
walk_down_backref(struct backref_edge * edges[],int * index)365 static struct backref_node *walk_down_backref(struct backref_edge *edges[],
366 int *index)
367 {
368 struct backref_edge *edge;
369 struct backref_node *lower;
370 int idx = *index;
371
372 while (idx > 0) {
373 edge = edges[idx - 1];
374 lower = edge->node[LOWER];
375 if (list_is_last(&edge->list[LOWER], &lower->upper)) {
376 idx--;
377 continue;
378 }
379 edge = list_entry(edge->list[LOWER].next,
380 struct backref_edge, list[LOWER]);
381 edges[idx - 1] = edge;
382 *index = idx;
383 return edge->node[UPPER];
384 }
385 *index = 0;
386 return NULL;
387 }
388
unlock_node_buffer(struct backref_node * node)389 static void unlock_node_buffer(struct backref_node *node)
390 {
391 if (node->locked) {
392 btrfs_tree_unlock(node->eb);
393 node->locked = 0;
394 }
395 }
396
drop_node_buffer(struct backref_node * node)397 static void drop_node_buffer(struct backref_node *node)
398 {
399 if (node->eb) {
400 unlock_node_buffer(node);
401 free_extent_buffer(node->eb);
402 node->eb = NULL;
403 }
404 }
405
drop_backref_node(struct backref_cache * tree,struct backref_node * node)406 static void drop_backref_node(struct backref_cache *tree,
407 struct backref_node *node)
408 {
409 BUG_ON(!list_empty(&node->upper));
410
411 drop_node_buffer(node);
412 list_del(&node->list);
413 list_del(&node->lower);
414 if (!RB_EMPTY_NODE(&node->rb_node))
415 rb_erase(&node->rb_node, &tree->rb_root);
416 free_backref_node(tree, node);
417 }
418
419 /*
420 * remove a backref node from the backref cache
421 */
remove_backref_node(struct backref_cache * cache,struct backref_node * node)422 static void remove_backref_node(struct backref_cache *cache,
423 struct backref_node *node)
424 {
425 struct backref_node *upper;
426 struct backref_edge *edge;
427
428 if (!node)
429 return;
430
431 BUG_ON(!node->lowest && !node->detached);
432 while (!list_empty(&node->upper)) {
433 edge = list_entry(node->upper.next, struct backref_edge,
434 list[LOWER]);
435 upper = edge->node[UPPER];
436 list_del(&edge->list[LOWER]);
437 list_del(&edge->list[UPPER]);
438 free_backref_edge(cache, edge);
439
440 if (RB_EMPTY_NODE(&upper->rb_node)) {
441 BUG_ON(!list_empty(&node->upper));
442 drop_backref_node(cache, node);
443 node = upper;
444 node->lowest = 1;
445 continue;
446 }
447 /*
448 * add the node to leaf node list if no other
449 * child block cached.
450 */
451 if (list_empty(&upper->lower)) {
452 list_add_tail(&upper->lower, &cache->leaves);
453 upper->lowest = 1;
454 }
455 }
456
457 drop_backref_node(cache, node);
458 }
459
update_backref_node(struct backref_cache * cache,struct backref_node * node,u64 bytenr)460 static void update_backref_node(struct backref_cache *cache,
461 struct backref_node *node, u64 bytenr)
462 {
463 struct rb_node *rb_node;
464 rb_erase(&node->rb_node, &cache->rb_root);
465 node->bytenr = bytenr;
466 rb_node = tree_insert(&cache->rb_root, node->bytenr, &node->rb_node);
467 if (rb_node)
468 backref_tree_panic(rb_node, -EEXIST, bytenr);
469 }
470
471 /*
472 * update backref cache after a transaction commit
473 */
update_backref_cache(struct btrfs_trans_handle * trans,struct backref_cache * cache)474 static int update_backref_cache(struct btrfs_trans_handle *trans,
475 struct backref_cache *cache)
476 {
477 struct backref_node *node;
478 int level = 0;
479
480 if (cache->last_trans == 0) {
481 cache->last_trans = trans->transid;
482 return 0;
483 }
484
485 if (cache->last_trans == trans->transid)
486 return 0;
487
488 /*
489 * detached nodes are used to avoid unnecessary backref
490 * lookup. transaction commit changes the extent tree.
491 * so the detached nodes are no longer useful.
492 */
493 while (!list_empty(&cache->detached)) {
494 node = list_entry(cache->detached.next,
495 struct backref_node, list);
496 remove_backref_node(cache, node);
497 }
498
499 while (!list_empty(&cache->changed)) {
500 node = list_entry(cache->changed.next,
501 struct backref_node, list);
502 list_del_init(&node->list);
503 BUG_ON(node->pending);
504 update_backref_node(cache, node, node->new_bytenr);
505 }
506
507 /*
508 * some nodes can be left in the pending list if there were
509 * errors during processing the pending nodes.
510 */
511 for (level = 0; level < BTRFS_MAX_LEVEL; level++) {
512 list_for_each_entry(node, &cache->pending[level], list) {
513 BUG_ON(!node->pending);
514 if (node->bytenr == node->new_bytenr)
515 continue;
516 update_backref_node(cache, node, node->new_bytenr);
517 }
518 }
519
520 cache->last_trans = 0;
521 return 1;
522 }
523
524
should_ignore_root(struct btrfs_root * root)525 static int should_ignore_root(struct btrfs_root *root)
526 {
527 struct btrfs_root *reloc_root;
528
529 if (!root->ref_cows)
530 return 0;
531
532 reloc_root = root->reloc_root;
533 if (!reloc_root)
534 return 0;
535
536 if (btrfs_root_last_snapshot(&reloc_root->root_item) ==
537 root->fs_info->running_transaction->transid - 1)
538 return 0;
539 /*
540 * if there is reloc tree and it was created in previous
541 * transaction backref lookup can find the reloc tree,
542 * so backref node for the fs tree root is useless for
543 * relocation.
544 */
545 return 1;
546 }
547 /*
548 * find reloc tree by address of tree root
549 */
find_reloc_root(struct reloc_control * rc,u64 bytenr)550 static struct btrfs_root *find_reloc_root(struct reloc_control *rc,
551 u64 bytenr)
552 {
553 struct rb_node *rb_node;
554 struct mapping_node *node;
555 struct btrfs_root *root = NULL;
556
557 spin_lock(&rc->reloc_root_tree.lock);
558 rb_node = tree_search(&rc->reloc_root_tree.rb_root, bytenr);
559 if (rb_node) {
560 node = rb_entry(rb_node, struct mapping_node, rb_node);
561 root = (struct btrfs_root *)node->data;
562 }
563 spin_unlock(&rc->reloc_root_tree.lock);
564 return root;
565 }
566
is_cowonly_root(u64 root_objectid)567 static int is_cowonly_root(u64 root_objectid)
568 {
569 if (root_objectid == BTRFS_ROOT_TREE_OBJECTID ||
570 root_objectid == BTRFS_EXTENT_TREE_OBJECTID ||
571 root_objectid == BTRFS_CHUNK_TREE_OBJECTID ||
572 root_objectid == BTRFS_DEV_TREE_OBJECTID ||
573 root_objectid == BTRFS_TREE_LOG_OBJECTID ||
574 root_objectid == BTRFS_CSUM_TREE_OBJECTID)
575 return 1;
576 return 0;
577 }
578
read_fs_root(struct btrfs_fs_info * fs_info,u64 root_objectid)579 static struct btrfs_root *read_fs_root(struct btrfs_fs_info *fs_info,
580 u64 root_objectid)
581 {
582 struct btrfs_key key;
583
584 key.objectid = root_objectid;
585 key.type = BTRFS_ROOT_ITEM_KEY;
586 if (is_cowonly_root(root_objectid))
587 key.offset = 0;
588 else
589 key.offset = (u64)-1;
590
591 return btrfs_read_fs_root_no_name(fs_info, &key);
592 }
593
594 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
595 static noinline_for_stack
find_tree_root(struct reloc_control * rc,struct extent_buffer * leaf,struct btrfs_extent_ref_v0 * ref0)596 struct btrfs_root *find_tree_root(struct reloc_control *rc,
597 struct extent_buffer *leaf,
598 struct btrfs_extent_ref_v0 *ref0)
599 {
600 struct btrfs_root *root;
601 u64 root_objectid = btrfs_ref_root_v0(leaf, ref0);
602 u64 generation = btrfs_ref_generation_v0(leaf, ref0);
603
604 BUG_ON(root_objectid == BTRFS_TREE_RELOC_OBJECTID);
605
606 root = read_fs_root(rc->extent_root->fs_info, root_objectid);
607 BUG_ON(IS_ERR(root));
608
609 if (root->ref_cows &&
610 generation != btrfs_root_generation(&root->root_item))
611 return NULL;
612
613 return root;
614 }
615 #endif
616
617 static noinline_for_stack
find_inline_backref(struct extent_buffer * leaf,int slot,unsigned long * ptr,unsigned long * end)618 int find_inline_backref(struct extent_buffer *leaf, int slot,
619 unsigned long *ptr, unsigned long *end)
620 {
621 struct btrfs_key key;
622 struct btrfs_extent_item *ei;
623 struct btrfs_tree_block_info *bi;
624 u32 item_size;
625
626 btrfs_item_key_to_cpu(leaf, &key, slot);
627
628 item_size = btrfs_item_size_nr(leaf, slot);
629 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
630 if (item_size < sizeof(*ei)) {
631 WARN_ON(item_size != sizeof(struct btrfs_extent_item_v0));
632 return 1;
633 }
634 #endif
635 ei = btrfs_item_ptr(leaf, slot, struct btrfs_extent_item);
636 WARN_ON(!(btrfs_extent_flags(leaf, ei) &
637 BTRFS_EXTENT_FLAG_TREE_BLOCK));
638
639 if (key.type == BTRFS_EXTENT_ITEM_KEY &&
640 item_size <= sizeof(*ei) + sizeof(*bi)) {
641 WARN_ON(item_size < sizeof(*ei) + sizeof(*bi));
642 return 1;
643 }
644
645 if (key.type == BTRFS_EXTENT_ITEM_KEY) {
646 bi = (struct btrfs_tree_block_info *)(ei + 1);
647 *ptr = (unsigned long)(bi + 1);
648 } else {
649 *ptr = (unsigned long)(ei + 1);
650 }
651 *end = (unsigned long)ei + item_size;
652 return 0;
653 }
654
655 /*
656 * build backref tree for a given tree block. root of the backref tree
657 * corresponds the tree block, leaves of the backref tree correspond
658 * roots of b-trees that reference the tree block.
659 *
660 * the basic idea of this function is check backrefs of a given block
661 * to find upper level blocks that refernece the block, and then check
662 * bakcrefs of these upper level blocks recursively. the recursion stop
663 * when tree root is reached or backrefs for the block is cached.
664 *
665 * NOTE: if we find backrefs for a block are cached, we know backrefs
666 * for all upper level blocks that directly/indirectly reference the
667 * block are also cached.
668 */
669 static noinline_for_stack
build_backref_tree(struct reloc_control * rc,struct btrfs_key * node_key,int level,u64 bytenr)670 struct backref_node *build_backref_tree(struct reloc_control *rc,
671 struct btrfs_key *node_key,
672 int level, u64 bytenr)
673 {
674 struct backref_cache *cache = &rc->backref_cache;
675 struct btrfs_path *path1;
676 struct btrfs_path *path2;
677 struct extent_buffer *eb;
678 struct btrfs_root *root;
679 struct backref_node *cur;
680 struct backref_node *upper;
681 struct backref_node *lower;
682 struct backref_node *node = NULL;
683 struct backref_node *exist = NULL;
684 struct backref_edge *edge;
685 struct rb_node *rb_node;
686 struct btrfs_key key;
687 unsigned long end;
688 unsigned long ptr;
689 LIST_HEAD(list);
690 LIST_HEAD(useless);
691 int cowonly;
692 int ret;
693 int err = 0;
694
695 path1 = btrfs_alloc_path();
696 path2 = btrfs_alloc_path();
697 if (!path1 || !path2) {
698 err = -ENOMEM;
699 goto out;
700 }
701 path1->reada = 1;
702 path2->reada = 2;
703
704 node = alloc_backref_node(cache);
705 if (!node) {
706 err = -ENOMEM;
707 goto out;
708 }
709
710 node->bytenr = bytenr;
711 node->level = level;
712 node->lowest = 1;
713 cur = node;
714 again:
715 end = 0;
716 ptr = 0;
717 key.objectid = cur->bytenr;
718 key.type = BTRFS_METADATA_ITEM_KEY;
719 key.offset = (u64)-1;
720
721 path1->search_commit_root = 1;
722 path1->skip_locking = 1;
723 ret = btrfs_search_slot(NULL, rc->extent_root, &key, path1,
724 0, 0);
725 if (ret < 0) {
726 err = ret;
727 goto out;
728 }
729 BUG_ON(!ret || !path1->slots[0]);
730
731 path1->slots[0]--;
732
733 WARN_ON(cur->checked);
734 if (!list_empty(&cur->upper)) {
735 /*
736 * the backref was added previously when processing
737 * backref of type BTRFS_TREE_BLOCK_REF_KEY
738 */
739 BUG_ON(!list_is_singular(&cur->upper));
740 edge = list_entry(cur->upper.next, struct backref_edge,
741 list[LOWER]);
742 BUG_ON(!list_empty(&edge->list[UPPER]));
743 exist = edge->node[UPPER];
744 /*
745 * add the upper level block to pending list if we need
746 * check its backrefs
747 */
748 if (!exist->checked)
749 list_add_tail(&edge->list[UPPER], &list);
750 } else {
751 exist = NULL;
752 }
753
754 while (1) {
755 cond_resched();
756 eb = path1->nodes[0];
757
758 if (ptr >= end) {
759 if (path1->slots[0] >= btrfs_header_nritems(eb)) {
760 ret = btrfs_next_leaf(rc->extent_root, path1);
761 if (ret < 0) {
762 err = ret;
763 goto out;
764 }
765 if (ret > 0)
766 break;
767 eb = path1->nodes[0];
768 }
769
770 btrfs_item_key_to_cpu(eb, &key, path1->slots[0]);
771 if (key.objectid != cur->bytenr) {
772 WARN_ON(exist);
773 break;
774 }
775
776 if (key.type == BTRFS_EXTENT_ITEM_KEY ||
777 key.type == BTRFS_METADATA_ITEM_KEY) {
778 ret = find_inline_backref(eb, path1->slots[0],
779 &ptr, &end);
780 if (ret)
781 goto next;
782 }
783 }
784
785 if (ptr < end) {
786 /* update key for inline back ref */
787 struct btrfs_extent_inline_ref *iref;
788 iref = (struct btrfs_extent_inline_ref *)ptr;
789 key.type = btrfs_extent_inline_ref_type(eb, iref);
790 key.offset = btrfs_extent_inline_ref_offset(eb, iref);
791 WARN_ON(key.type != BTRFS_TREE_BLOCK_REF_KEY &&
792 key.type != BTRFS_SHARED_BLOCK_REF_KEY);
793 }
794
795 if (exist &&
796 ((key.type == BTRFS_TREE_BLOCK_REF_KEY &&
797 exist->owner == key.offset) ||
798 (key.type == BTRFS_SHARED_BLOCK_REF_KEY &&
799 exist->bytenr == key.offset))) {
800 exist = NULL;
801 goto next;
802 }
803
804 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
805 if (key.type == BTRFS_SHARED_BLOCK_REF_KEY ||
806 key.type == BTRFS_EXTENT_REF_V0_KEY) {
807 if (key.type == BTRFS_EXTENT_REF_V0_KEY) {
808 struct btrfs_extent_ref_v0 *ref0;
809 ref0 = btrfs_item_ptr(eb, path1->slots[0],
810 struct btrfs_extent_ref_v0);
811 if (key.objectid == key.offset) {
812 root = find_tree_root(rc, eb, ref0);
813 if (root && !should_ignore_root(root))
814 cur->root = root;
815 else
816 list_add(&cur->list, &useless);
817 break;
818 }
819 if (is_cowonly_root(btrfs_ref_root_v0(eb,
820 ref0)))
821 cur->cowonly = 1;
822 }
823 #else
824 BUG_ON(key.type == BTRFS_EXTENT_REF_V0_KEY);
825 if (key.type == BTRFS_SHARED_BLOCK_REF_KEY) {
826 #endif
827 if (key.objectid == key.offset) {
828 /*
829 * only root blocks of reloc trees use
830 * backref of this type.
831 */
832 root = find_reloc_root(rc, cur->bytenr);
833 BUG_ON(!root);
834 cur->root = root;
835 break;
836 }
837
838 edge = alloc_backref_edge(cache);
839 if (!edge) {
840 err = -ENOMEM;
841 goto out;
842 }
843 rb_node = tree_search(&cache->rb_root, key.offset);
844 if (!rb_node) {
845 upper = alloc_backref_node(cache);
846 if (!upper) {
847 free_backref_edge(cache, edge);
848 err = -ENOMEM;
849 goto out;
850 }
851 upper->bytenr = key.offset;
852 upper->level = cur->level + 1;
853 /*
854 * backrefs for the upper level block isn't
855 * cached, add the block to pending list
856 */
857 list_add_tail(&edge->list[UPPER], &list);
858 } else {
859 upper = rb_entry(rb_node, struct backref_node,
860 rb_node);
861 BUG_ON(!upper->checked);
862 INIT_LIST_HEAD(&edge->list[UPPER]);
863 }
864 list_add_tail(&edge->list[LOWER], &cur->upper);
865 edge->node[LOWER] = cur;
866 edge->node[UPPER] = upper;
867
868 goto next;
869 } else if (key.type != BTRFS_TREE_BLOCK_REF_KEY) {
870 goto next;
871 }
872
873 /* key.type == BTRFS_TREE_BLOCK_REF_KEY */
874 root = read_fs_root(rc->extent_root->fs_info, key.offset);
875 if (IS_ERR(root)) {
876 err = PTR_ERR(root);
877 goto out;
878 }
879
880 if (!root->ref_cows)
881 cur->cowonly = 1;
882
883 if (btrfs_root_level(&root->root_item) == cur->level) {
884 /* tree root */
885 BUG_ON(btrfs_root_bytenr(&root->root_item) !=
886 cur->bytenr);
887 if (should_ignore_root(root))
888 list_add(&cur->list, &useless);
889 else
890 cur->root = root;
891 break;
892 }
893
894 level = cur->level + 1;
895
896 /*
897 * searching the tree to find upper level blocks
898 * reference the block.
899 */
900 path2->search_commit_root = 1;
901 path2->skip_locking = 1;
902 path2->lowest_level = level;
903 ret = btrfs_search_slot(NULL, root, node_key, path2, 0, 0);
904 path2->lowest_level = 0;
905 if (ret < 0) {
906 err = ret;
907 goto out;
908 }
909 if (ret > 0 && path2->slots[level] > 0)
910 path2->slots[level]--;
911
912 eb = path2->nodes[level];
913 WARN_ON(btrfs_node_blockptr(eb, path2->slots[level]) !=
914 cur->bytenr);
915
916 lower = cur;
917 for (; level < BTRFS_MAX_LEVEL; level++) {
918 if (!path2->nodes[level]) {
919 BUG_ON(btrfs_root_bytenr(&root->root_item) !=
920 lower->bytenr);
921 if (should_ignore_root(root))
922 list_add(&lower->list, &useless);
923 else
924 lower->root = root;
925 break;
926 }
927
928 edge = alloc_backref_edge(cache);
929 if (!edge) {
930 err = -ENOMEM;
931 goto out;
932 }
933
934 eb = path2->nodes[level];
935 rb_node = tree_search(&cache->rb_root, eb->start);
936 if (!rb_node) {
937 upper = alloc_backref_node(cache);
938 if (!upper) {
939 free_backref_edge(cache, edge);
940 err = -ENOMEM;
941 goto out;
942 }
943 upper->bytenr = eb->start;
944 upper->owner = btrfs_header_owner(eb);
945 upper->level = lower->level + 1;
946 if (!root->ref_cows)
947 upper->cowonly = 1;
948
949 /*
950 * if we know the block isn't shared
951 * we can void checking its backrefs.
952 */
953 if (btrfs_block_can_be_shared(root, eb))
954 upper->checked = 0;
955 else
956 upper->checked = 1;
957
958 /*
959 * add the block to pending list if we
960 * need check its backrefs. only block
961 * at 'cur->level + 1' is added to the
962 * tail of pending list. this guarantees
963 * we check backrefs from lower level
964 * blocks to upper level blocks.
965 */
966 if (!upper->checked &&
967 level == cur->level + 1) {
968 list_add_tail(&edge->list[UPPER],
969 &list);
970 } else
971 INIT_LIST_HEAD(&edge->list[UPPER]);
972 } else {
973 upper = rb_entry(rb_node, struct backref_node,
974 rb_node);
975 BUG_ON(!upper->checked);
976 INIT_LIST_HEAD(&edge->list[UPPER]);
977 if (!upper->owner)
978 upper->owner = btrfs_header_owner(eb);
979 }
980 list_add_tail(&edge->list[LOWER], &lower->upper);
981 edge->node[LOWER] = lower;
982 edge->node[UPPER] = upper;
983
984 if (rb_node)
985 break;
986 lower = upper;
987 upper = NULL;
988 }
989 btrfs_release_path(path2);
990 next:
991 if (ptr < end) {
992 ptr += btrfs_extent_inline_ref_size(key.type);
993 if (ptr >= end) {
994 WARN_ON(ptr > end);
995 ptr = 0;
996 end = 0;
997 }
998 }
999 if (ptr >= end)
1000 path1->slots[0]++;
1001 }
1002 btrfs_release_path(path1);
1003
1004 cur->checked = 1;
1005 WARN_ON(exist);
1006
1007 /* the pending list isn't empty, take the first block to process */
1008 if (!list_empty(&list)) {
1009 edge = list_entry(list.next, struct backref_edge, list[UPPER]);
1010 list_del_init(&edge->list[UPPER]);
1011 cur = edge->node[UPPER];
1012 goto again;
1013 }
1014
1015 /*
1016 * everything goes well, connect backref nodes and insert backref nodes
1017 * into the cache.
1018 */
1019 BUG_ON(!node->checked);
1020 cowonly = node->cowonly;
1021 if (!cowonly) {
1022 rb_node = tree_insert(&cache->rb_root, node->bytenr,
1023 &node->rb_node);
1024 if (rb_node)
1025 backref_tree_panic(rb_node, -EEXIST, node->bytenr);
1026 list_add_tail(&node->lower, &cache->leaves);
1027 }
1028
1029 list_for_each_entry(edge, &node->upper, list[LOWER])
1030 list_add_tail(&edge->list[UPPER], &list);
1031
1032 while (!list_empty(&list)) {
1033 edge = list_entry(list.next, struct backref_edge, list[UPPER]);
1034 list_del_init(&edge->list[UPPER]);
1035 upper = edge->node[UPPER];
1036 if (upper->detached) {
1037 list_del(&edge->list[LOWER]);
1038 lower = edge->node[LOWER];
1039 free_backref_edge(cache, edge);
1040 if (list_empty(&lower->upper))
1041 list_add(&lower->list, &useless);
1042 continue;
1043 }
1044
1045 if (!RB_EMPTY_NODE(&upper->rb_node)) {
1046 if (upper->lowest) {
1047 list_del_init(&upper->lower);
1048 upper->lowest = 0;
1049 }
1050
1051 list_add_tail(&edge->list[UPPER], &upper->lower);
1052 continue;
1053 }
1054
1055 BUG_ON(!upper->checked);
1056 BUG_ON(cowonly != upper->cowonly);
1057 if (!cowonly) {
1058 rb_node = tree_insert(&cache->rb_root, upper->bytenr,
1059 &upper->rb_node);
1060 if (rb_node)
1061 backref_tree_panic(rb_node, -EEXIST,
1062 upper->bytenr);
1063 }
1064
1065 list_add_tail(&edge->list[UPPER], &upper->lower);
1066
1067 list_for_each_entry(edge, &upper->upper, list[LOWER])
1068 list_add_tail(&edge->list[UPPER], &list);
1069 }
1070 /*
1071 * process useless backref nodes. backref nodes for tree leaves
1072 * are deleted from the cache. backref nodes for upper level
1073 * tree blocks are left in the cache to avoid unnecessary backref
1074 * lookup.
1075 */
1076 while (!list_empty(&useless)) {
1077 upper = list_entry(useless.next, struct backref_node, list);
1078 list_del_init(&upper->list);
1079 BUG_ON(!list_empty(&upper->upper));
1080 if (upper == node)
1081 node = NULL;
1082 if (upper->lowest) {
1083 list_del_init(&upper->lower);
1084 upper->lowest = 0;
1085 }
1086 while (!list_empty(&upper->lower)) {
1087 edge = list_entry(upper->lower.next,
1088 struct backref_edge, list[UPPER]);
1089 list_del(&edge->list[UPPER]);
1090 list_del(&edge->list[LOWER]);
1091 lower = edge->node[LOWER];
1092 free_backref_edge(cache, edge);
1093
1094 if (list_empty(&lower->upper))
1095 list_add(&lower->list, &useless);
1096 }
1097 __mark_block_processed(rc, upper);
1098 if (upper->level > 0) {
1099 list_add(&upper->list, &cache->detached);
1100 upper->detached = 1;
1101 } else {
1102 rb_erase(&upper->rb_node, &cache->rb_root);
1103 free_backref_node(cache, upper);
1104 }
1105 }
1106 out:
1107 btrfs_free_path(path1);
1108 btrfs_free_path(path2);
1109 if (err) {
1110 while (!list_empty(&useless)) {
1111 lower = list_entry(useless.next,
1112 struct backref_node, upper);
1113 list_del_init(&lower->upper);
1114 }
1115 upper = node;
1116 INIT_LIST_HEAD(&list);
1117 while (upper) {
1118 if (RB_EMPTY_NODE(&upper->rb_node)) {
1119 list_splice_tail(&upper->upper, &list);
1120 free_backref_node(cache, upper);
1121 }
1122
1123 if (list_empty(&list))
1124 break;
1125
1126 edge = list_entry(list.next, struct backref_edge,
1127 list[LOWER]);
1128 list_del(&edge->list[LOWER]);
1129 upper = edge->node[UPPER];
1130 free_backref_edge(cache, edge);
1131 }
1132 return ERR_PTR(err);
1133 }
1134 BUG_ON(node && node->detached);
1135 return node;
1136 }
1137
1138 /*
1139 * helper to add backref node for the newly created snapshot.
1140 * the backref node is created by cloning backref node that
1141 * corresponds to root of source tree
1142 */
1143 static int clone_backref_node(struct btrfs_trans_handle *trans,
1144 struct reloc_control *rc,
1145 struct btrfs_root *src,
1146 struct btrfs_root *dest)
1147 {
1148 struct btrfs_root *reloc_root = src->reloc_root;
1149 struct backref_cache *cache = &rc->backref_cache;
1150 struct backref_node *node = NULL;
1151 struct backref_node *new_node;
1152 struct backref_edge *edge;
1153 struct backref_edge *new_edge;
1154 struct rb_node *rb_node;
1155
1156 if (cache->last_trans > 0)
1157 update_backref_cache(trans, cache);
1158
1159 rb_node = tree_search(&cache->rb_root, src->commit_root->start);
1160 if (rb_node) {
1161 node = rb_entry(rb_node, struct backref_node, rb_node);
1162 if (node->detached)
1163 node = NULL;
1164 else
1165 BUG_ON(node->new_bytenr != reloc_root->node->start);
1166 }
1167
1168 if (!node) {
1169 rb_node = tree_search(&cache->rb_root,
1170 reloc_root->commit_root->start);
1171 if (rb_node) {
1172 node = rb_entry(rb_node, struct backref_node,
1173 rb_node);
1174 BUG_ON(node->detached);
1175 }
1176 }
1177
1178 if (!node)
1179 return 0;
1180
1181 new_node = alloc_backref_node(cache);
1182 if (!new_node)
1183 return -ENOMEM;
1184
1185 new_node->bytenr = dest->node->start;
1186 new_node->level = node->level;
1187 new_node->lowest = node->lowest;
1188 new_node->checked = 1;
1189 new_node->root = dest;
1190
1191 if (!node->lowest) {
1192 list_for_each_entry(edge, &node->lower, list[UPPER]) {
1193 new_edge = alloc_backref_edge(cache);
1194 if (!new_edge)
1195 goto fail;
1196
1197 new_edge->node[UPPER] = new_node;
1198 new_edge->node[LOWER] = edge->node[LOWER];
1199 list_add_tail(&new_edge->list[UPPER],
1200 &new_node->lower);
1201 }
1202 } else {
1203 list_add_tail(&new_node->lower, &cache->leaves);
1204 }
1205
1206 rb_node = tree_insert(&cache->rb_root, new_node->bytenr,
1207 &new_node->rb_node);
1208 if (rb_node)
1209 backref_tree_panic(rb_node, -EEXIST, new_node->bytenr);
1210
1211 if (!new_node->lowest) {
1212 list_for_each_entry(new_edge, &new_node->lower, list[UPPER]) {
1213 list_add_tail(&new_edge->list[LOWER],
1214 &new_edge->node[LOWER]->upper);
1215 }
1216 }
1217 return 0;
1218 fail:
1219 while (!list_empty(&new_node->lower)) {
1220 new_edge = list_entry(new_node->lower.next,
1221 struct backref_edge, list[UPPER]);
1222 list_del(&new_edge->list[UPPER]);
1223 free_backref_edge(cache, new_edge);
1224 }
1225 free_backref_node(cache, new_node);
1226 return -ENOMEM;
1227 }
1228
1229 /*
1230 * helper to add 'address of tree root -> reloc tree' mapping
1231 */
1232 static int __must_check __add_reloc_root(struct btrfs_root *root)
1233 {
1234 struct rb_node *rb_node;
1235 struct mapping_node *node;
1236 struct reloc_control *rc = root->fs_info->reloc_ctl;
1237
1238 node = kmalloc(sizeof(*node), GFP_NOFS);
1239 if (!node)
1240 return -ENOMEM;
1241
1242 node->bytenr = root->node->start;
1243 node->data = root;
1244
1245 spin_lock(&rc->reloc_root_tree.lock);
1246 rb_node = tree_insert(&rc->reloc_root_tree.rb_root,
1247 node->bytenr, &node->rb_node);
1248 spin_unlock(&rc->reloc_root_tree.lock);
1249 if (rb_node) {
1250 btrfs_panic(root->fs_info, -EEXIST, "Duplicate root found "
1251 "for start=%llu while inserting into relocation "
1252 "tree\n", node->bytenr);
1253 kfree(node);
1254 return -EEXIST;
1255 }
1256
1257 list_add_tail(&root->root_list, &rc->reloc_roots);
1258 return 0;
1259 }
1260
1261 /*
1262 * helper to update/delete the 'address of tree root -> reloc tree'
1263 * mapping
1264 */
1265 static int __update_reloc_root(struct btrfs_root *root, int del)
1266 {
1267 struct rb_node *rb_node;
1268 struct mapping_node *node = NULL;
1269 struct reloc_control *rc = root->fs_info->reloc_ctl;
1270
1271 spin_lock(&rc->reloc_root_tree.lock);
1272 rb_node = tree_search(&rc->reloc_root_tree.rb_root,
1273 root->commit_root->start);
1274 if (rb_node) {
1275 node = rb_entry(rb_node, struct mapping_node, rb_node);
1276 rb_erase(&node->rb_node, &rc->reloc_root_tree.rb_root);
1277 }
1278 spin_unlock(&rc->reloc_root_tree.lock);
1279
1280 if (!node)
1281 return 0;
1282 BUG_ON((struct btrfs_root *)node->data != root);
1283
1284 if (!del) {
1285 spin_lock(&rc->reloc_root_tree.lock);
1286 node->bytenr = root->node->start;
1287 rb_node = tree_insert(&rc->reloc_root_tree.rb_root,
1288 node->bytenr, &node->rb_node);
1289 spin_unlock(&rc->reloc_root_tree.lock);
1290 if (rb_node)
1291 backref_tree_panic(rb_node, -EEXIST, node->bytenr);
1292 } else {
1293 spin_lock(&root->fs_info->trans_lock);
1294 list_del_init(&root->root_list);
1295 spin_unlock(&root->fs_info->trans_lock);
1296 kfree(node);
1297 }
1298 return 0;
1299 }
1300
1301 static struct btrfs_root *create_reloc_root(struct btrfs_trans_handle *trans,
1302 struct btrfs_root *root, u64 objectid)
1303 {
1304 struct btrfs_root *reloc_root;
1305 struct extent_buffer *eb;
1306 struct btrfs_root_item *root_item;
1307 struct btrfs_key root_key;
1308 int ret;
1309
1310 root_item = kmalloc(sizeof(*root_item), GFP_NOFS);
1311 BUG_ON(!root_item);
1312
1313 root_key.objectid = BTRFS_TREE_RELOC_OBJECTID;
1314 root_key.type = BTRFS_ROOT_ITEM_KEY;
1315 root_key.offset = objectid;
1316
1317 if (root->root_key.objectid == objectid) {
1318 /* called by btrfs_init_reloc_root */
1319 ret = btrfs_copy_root(trans, root, root->commit_root, &eb,
1320 BTRFS_TREE_RELOC_OBJECTID);
1321 BUG_ON(ret);
1322
1323 btrfs_set_root_last_snapshot(&root->root_item,
1324 trans->transid - 1);
1325 } else {
1326 /*
1327 * called by btrfs_reloc_post_snapshot_hook.
1328 * the source tree is a reloc tree, all tree blocks
1329 * modified after it was created have RELOC flag
1330 * set in their headers. so it's OK to not update
1331 * the 'last_snapshot'.
1332 */
1333 ret = btrfs_copy_root(trans, root, root->node, &eb,
1334 BTRFS_TREE_RELOC_OBJECTID);
1335 BUG_ON(ret);
1336 }
1337
1338 memcpy(root_item, &root->root_item, sizeof(*root_item));
1339 btrfs_set_root_bytenr(root_item, eb->start);
1340 btrfs_set_root_level(root_item, btrfs_header_level(eb));
1341 btrfs_set_root_generation(root_item, trans->transid);
1342
1343 if (root->root_key.objectid == objectid) {
1344 btrfs_set_root_refs(root_item, 0);
1345 memset(&root_item->drop_progress, 0,
1346 sizeof(struct btrfs_disk_key));
1347 root_item->drop_level = 0;
1348 }
1349
1350 btrfs_tree_unlock(eb);
1351 free_extent_buffer(eb);
1352
1353 ret = btrfs_insert_root(trans, root->fs_info->tree_root,
1354 &root_key, root_item);
1355 BUG_ON(ret);
1356 kfree(root_item);
1357
1358 reloc_root = btrfs_read_fs_root_no_radix(root->fs_info->tree_root,
1359 &root_key);
1360 BUG_ON(IS_ERR(reloc_root));
1361 reloc_root->last_trans = trans->transid;
1362 return reloc_root;
1363 }
1364
1365 /*
1366 * create reloc tree for a given fs tree. reloc tree is just a
1367 * snapshot of the fs tree with special root objectid.
1368 */
1369 int btrfs_init_reloc_root(struct btrfs_trans_handle *trans,
1370 struct btrfs_root *root)
1371 {
1372 struct btrfs_root *reloc_root;
1373 struct reloc_control *rc = root->fs_info->reloc_ctl;
1374 int clear_rsv = 0;
1375 int ret;
1376
1377 if (root->reloc_root) {
1378 reloc_root = root->reloc_root;
1379 reloc_root->last_trans = trans->transid;
1380 return 0;
1381 }
1382
1383 if (!rc || !rc->create_reloc_tree ||
1384 root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID)
1385 return 0;
1386
1387 if (!trans->block_rsv) {
1388 trans->block_rsv = rc->block_rsv;
1389 clear_rsv = 1;
1390 }
1391 reloc_root = create_reloc_root(trans, root, root->root_key.objectid);
1392 if (clear_rsv)
1393 trans->block_rsv = NULL;
1394
1395 ret = __add_reloc_root(reloc_root);
1396 BUG_ON(ret < 0);
1397 root->reloc_root = reloc_root;
1398 return 0;
1399 }
1400
1401 /*
1402 * update root item of reloc tree
1403 */
1404 int btrfs_update_reloc_root(struct btrfs_trans_handle *trans,
1405 struct btrfs_root *root)
1406 {
1407 struct btrfs_root *reloc_root;
1408 struct btrfs_root_item *root_item;
1409 int del = 0;
1410 int ret;
1411
1412 if (!root->reloc_root)
1413 goto out;
1414
1415 reloc_root = root->reloc_root;
1416 root_item = &reloc_root->root_item;
1417
1418 if (root->fs_info->reloc_ctl->merge_reloc_tree &&
1419 btrfs_root_refs(root_item) == 0) {
1420 root->reloc_root = NULL;
1421 del = 1;
1422 }
1423
1424 __update_reloc_root(reloc_root, del);
1425
1426 if (reloc_root->commit_root != reloc_root->node) {
1427 btrfs_set_root_node(root_item, reloc_root->node);
1428 free_extent_buffer(reloc_root->commit_root);
1429 reloc_root->commit_root = btrfs_root_node(reloc_root);
1430 }
1431
1432 ret = btrfs_update_root(trans, root->fs_info->tree_root,
1433 &reloc_root->root_key, root_item);
1434 BUG_ON(ret);
1435
1436 out:
1437 return 0;
1438 }
1439
1440 /*
1441 * helper to find first cached inode with inode number >= objectid
1442 * in a subvolume
1443 */
1444 static struct inode *find_next_inode(struct btrfs_root *root, u64 objectid)
1445 {
1446 struct rb_node *node;
1447 struct rb_node *prev;
1448 struct btrfs_inode *entry;
1449 struct inode *inode;
1450
1451 spin_lock(&root->inode_lock);
1452 again:
1453 node = root->inode_tree.rb_node;
1454 prev = NULL;
1455 while (node) {
1456 prev = node;
1457 entry = rb_entry(node, struct btrfs_inode, rb_node);
1458
1459 if (objectid < btrfs_ino(&entry->vfs_inode))
1460 node = node->rb_left;
1461 else if (objectid > btrfs_ino(&entry->vfs_inode))
1462 node = node->rb_right;
1463 else
1464 break;
1465 }
1466 if (!node) {
1467 while (prev) {
1468 entry = rb_entry(prev, struct btrfs_inode, rb_node);
1469 if (objectid <= btrfs_ino(&entry->vfs_inode)) {
1470 node = prev;
1471 break;
1472 }
1473 prev = rb_next(prev);
1474 }
1475 }
1476 while (node) {
1477 entry = rb_entry(node, struct btrfs_inode, rb_node);
1478 inode = igrab(&entry->vfs_inode);
1479 if (inode) {
1480 spin_unlock(&root->inode_lock);
1481 return inode;
1482 }
1483
1484 objectid = btrfs_ino(&entry->vfs_inode) + 1;
1485 if (cond_resched_lock(&root->inode_lock))
1486 goto again;
1487
1488 node = rb_next(node);
1489 }
1490 spin_unlock(&root->inode_lock);
1491 return NULL;
1492 }
1493
1494 static int in_block_group(u64 bytenr,
1495 struct btrfs_block_group_cache *block_group)
1496 {
1497 if (bytenr >= block_group->key.objectid &&
1498 bytenr < block_group->key.objectid + block_group->key.offset)
1499 return 1;
1500 return 0;
1501 }
1502
1503 /*
1504 * get new location of data
1505 */
1506 static int get_new_location(struct inode *reloc_inode, u64 *new_bytenr,
1507 u64 bytenr, u64 num_bytes)
1508 {
1509 struct btrfs_root *root = BTRFS_I(reloc_inode)->root;
1510 struct btrfs_path *path;
1511 struct btrfs_file_extent_item *fi;
1512 struct extent_buffer *leaf;
1513 int ret;
1514
1515 path = btrfs_alloc_path();
1516 if (!path)
1517 return -ENOMEM;
1518
1519 bytenr -= BTRFS_I(reloc_inode)->index_cnt;
1520 ret = btrfs_lookup_file_extent(NULL, root, path, btrfs_ino(reloc_inode),
1521 bytenr, 0);
1522 if (ret < 0)
1523 goto out;
1524 if (ret > 0) {
1525 ret = -ENOENT;
1526 goto out;
1527 }
1528
1529 leaf = path->nodes[0];
1530 fi = btrfs_item_ptr(leaf, path->slots[0],
1531 struct btrfs_file_extent_item);
1532
1533 BUG_ON(btrfs_file_extent_offset(leaf, fi) ||
1534 btrfs_file_extent_compression(leaf, fi) ||
1535 btrfs_file_extent_encryption(leaf, fi) ||
1536 btrfs_file_extent_other_encoding(leaf, fi));
1537
1538 if (num_bytes != btrfs_file_extent_disk_num_bytes(leaf, fi)) {
1539 ret = 1;
1540 goto out;
1541 }
1542
1543 *new_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
1544 ret = 0;
1545 out:
1546 btrfs_free_path(path);
1547 return ret;
1548 }
1549
1550 /*
1551 * update file extent items in the tree leaf to point to
1552 * the new locations.
1553 */
1554 static noinline_for_stack
1555 int replace_file_extents(struct btrfs_trans_handle *trans,
1556 struct reloc_control *rc,
1557 struct btrfs_root *root,
1558 struct extent_buffer *leaf)
1559 {
1560 struct btrfs_key key;
1561 struct btrfs_file_extent_item *fi;
1562 struct inode *inode = NULL;
1563 u64 parent;
1564 u64 bytenr;
1565 u64 new_bytenr = 0;
1566 u64 num_bytes;
1567 u64 end;
1568 u32 nritems;
1569 u32 i;
1570 int ret;
1571 int first = 1;
1572 int dirty = 0;
1573
1574 if (rc->stage != UPDATE_DATA_PTRS)
1575 return 0;
1576
1577 /* reloc trees always use full backref */
1578 if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID)
1579 parent = leaf->start;
1580 else
1581 parent = 0;
1582
1583 nritems = btrfs_header_nritems(leaf);
1584 for (i = 0; i < nritems; i++) {
1585 cond_resched();
1586 btrfs_item_key_to_cpu(leaf, &key, i);
1587 if (key.type != BTRFS_EXTENT_DATA_KEY)
1588 continue;
1589 fi = btrfs_item_ptr(leaf, i, struct btrfs_file_extent_item);
1590 if (btrfs_file_extent_type(leaf, fi) ==
1591 BTRFS_FILE_EXTENT_INLINE)
1592 continue;
1593 bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
1594 num_bytes = btrfs_file_extent_disk_num_bytes(leaf, fi);
1595 if (bytenr == 0)
1596 continue;
1597 if (!in_block_group(bytenr, rc->block_group))
1598 continue;
1599
1600 /*
1601 * if we are modifying block in fs tree, wait for readpage
1602 * to complete and drop the extent cache
1603 */
1604 if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) {
1605 if (first) {
1606 inode = find_next_inode(root, key.objectid);
1607 first = 0;
1608 } else if (inode && btrfs_ino(inode) < key.objectid) {
1609 btrfs_add_delayed_iput(inode);
1610 inode = find_next_inode(root, key.objectid);
1611 }
1612 if (inode && btrfs_ino(inode) == key.objectid) {
1613 end = key.offset +
1614 btrfs_file_extent_num_bytes(leaf, fi);
1615 WARN_ON(!IS_ALIGNED(key.offset,
1616 root->sectorsize));
1617 WARN_ON(!IS_ALIGNED(end, root->sectorsize));
1618 end--;
1619 ret = try_lock_extent(&BTRFS_I(inode)->io_tree,
1620 key.offset, end);
1621 if (!ret)
1622 continue;
1623
1624 btrfs_drop_extent_cache(inode, key.offset, end,
1625 1);
1626 unlock_extent(&BTRFS_I(inode)->io_tree,
1627 key.offset, end);
1628 }
1629 }
1630
1631 ret = get_new_location(rc->data_inode, &new_bytenr,
1632 bytenr, num_bytes);
1633 if (ret > 0) {
1634 WARN_ON(1);
1635 continue;
1636 }
1637 BUG_ON(ret < 0);
1638
1639 btrfs_set_file_extent_disk_bytenr(leaf, fi, new_bytenr);
1640 dirty = 1;
1641
1642 key.offset -= btrfs_file_extent_offset(leaf, fi);
1643 ret = btrfs_inc_extent_ref(trans, root, new_bytenr,
1644 num_bytes, parent,
1645 btrfs_header_owner(leaf),
1646 key.objectid, key.offset, 1);
1647 BUG_ON(ret);
1648
1649 ret = btrfs_free_extent(trans, root, bytenr, num_bytes,
1650 parent, btrfs_header_owner(leaf),
1651 key.objectid, key.offset, 1);
1652 BUG_ON(ret);
1653 }
1654 if (dirty)
1655 btrfs_mark_buffer_dirty(leaf);
1656 if (inode)
1657 btrfs_add_delayed_iput(inode);
1658 return 0;
1659 }
1660
1661 static noinline_for_stack
1662 int memcmp_node_keys(struct extent_buffer *eb, int slot,
1663 struct btrfs_path *path, int level)
1664 {
1665 struct btrfs_disk_key key1;
1666 struct btrfs_disk_key key2;
1667 btrfs_node_key(eb, &key1, slot);
1668 btrfs_node_key(path->nodes[level], &key2, path->slots[level]);
1669 return memcmp(&key1, &key2, sizeof(key1));
1670 }
1671
1672 /*
1673 * try to replace tree blocks in fs tree with the new blocks
1674 * in reloc tree. tree blocks haven't been modified since the
1675 * reloc tree was create can be replaced.
1676 *
1677 * if a block was replaced, level of the block + 1 is returned.
1678 * if no block got replaced, 0 is returned. if there are other
1679 * errors, a negative error number is returned.
1680 */
1681 static noinline_for_stack
1682 int replace_path(struct btrfs_trans_handle *trans,
1683 struct btrfs_root *dest, struct btrfs_root *src,
1684 struct btrfs_path *path, struct btrfs_key *next_key,
1685 int lowest_level, int max_level)
1686 {
1687 struct extent_buffer *eb;
1688 struct extent_buffer *parent;
1689 struct btrfs_key key;
1690 u64 old_bytenr;
1691 u64 new_bytenr;
1692 u64 old_ptr_gen;
1693 u64 new_ptr_gen;
1694 u64 last_snapshot;
1695 u32 blocksize;
1696 int cow = 0;
1697 int level;
1698 int ret;
1699 int slot;
1700
1701 BUG_ON(src->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID);
1702 BUG_ON(dest->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID);
1703
1704 last_snapshot = btrfs_root_last_snapshot(&src->root_item);
1705 again:
1706 slot = path->slots[lowest_level];
1707 btrfs_node_key_to_cpu(path->nodes[lowest_level], &key, slot);
1708
1709 eb = btrfs_lock_root_node(dest);
1710 btrfs_set_lock_blocking(eb);
1711 level = btrfs_header_level(eb);
1712
1713 if (level < lowest_level) {
1714 btrfs_tree_unlock(eb);
1715 free_extent_buffer(eb);
1716 return 0;
1717 }
1718
1719 if (cow) {
1720 ret = btrfs_cow_block(trans, dest, eb, NULL, 0, &eb);
1721 BUG_ON(ret);
1722 }
1723 btrfs_set_lock_blocking(eb);
1724
1725 if (next_key) {
1726 next_key->objectid = (u64)-1;
1727 next_key->type = (u8)-1;
1728 next_key->offset = (u64)-1;
1729 }
1730
1731 parent = eb;
1732 while (1) {
1733 level = btrfs_header_level(parent);
1734 BUG_ON(level < lowest_level);
1735
1736 ret = btrfs_bin_search(parent, &key, level, &slot);
1737 if (ret && slot > 0)
1738 slot--;
1739
1740 if (next_key && slot + 1 < btrfs_header_nritems(parent))
1741 btrfs_node_key_to_cpu(parent, next_key, slot + 1);
1742
1743 old_bytenr = btrfs_node_blockptr(parent, slot);
1744 blocksize = btrfs_level_size(dest, level - 1);
1745 old_ptr_gen = btrfs_node_ptr_generation(parent, slot);
1746
1747 if (level <= max_level) {
1748 eb = path->nodes[level];
1749 new_bytenr = btrfs_node_blockptr(eb,
1750 path->slots[level]);
1751 new_ptr_gen = btrfs_node_ptr_generation(eb,
1752 path->slots[level]);
1753 } else {
1754 new_bytenr = 0;
1755 new_ptr_gen = 0;
1756 }
1757
1758 if (new_bytenr > 0 && new_bytenr == old_bytenr) {
1759 WARN_ON(1);
1760 ret = level;
1761 break;
1762 }
1763
1764 if (new_bytenr == 0 || old_ptr_gen > last_snapshot ||
1765 memcmp_node_keys(parent, slot, path, level)) {
1766 if (level <= lowest_level) {
1767 ret = 0;
1768 break;
1769 }
1770
1771 eb = read_tree_block(dest, old_bytenr, blocksize,
1772 old_ptr_gen);
1773 if (!eb || !extent_buffer_uptodate(eb)) {
1774 ret = (!eb) ? -ENOMEM : -EIO;
1775 free_extent_buffer(eb);
1776 break;
1777 }
1778 btrfs_tree_lock(eb);
1779 if (cow) {
1780 ret = btrfs_cow_block(trans, dest, eb, parent,
1781 slot, &eb);
1782 BUG_ON(ret);
1783 }
1784 btrfs_set_lock_blocking(eb);
1785
1786 btrfs_tree_unlock(parent);
1787 free_extent_buffer(parent);
1788
1789 parent = eb;
1790 continue;
1791 }
1792
1793 if (!cow) {
1794 btrfs_tree_unlock(parent);
1795 free_extent_buffer(parent);
1796 cow = 1;
1797 goto again;
1798 }
1799
1800 btrfs_node_key_to_cpu(path->nodes[level], &key,
1801 path->slots[level]);
1802 btrfs_release_path(path);
1803
1804 path->lowest_level = level;
1805 ret = btrfs_search_slot(trans, src, &key, path, 0, 1);
1806 path->lowest_level = 0;
1807 BUG_ON(ret);
1808
1809 /*
1810 * swap blocks in fs tree and reloc tree.
1811 */
1812 btrfs_set_node_blockptr(parent, slot, new_bytenr);
1813 btrfs_set_node_ptr_generation(parent, slot, new_ptr_gen);
1814 btrfs_mark_buffer_dirty(parent);
1815
1816 btrfs_set_node_blockptr(path->nodes[level],
1817 path->slots[level], old_bytenr);
1818 btrfs_set_node_ptr_generation(path->nodes[level],
1819 path->slots[level], old_ptr_gen);
1820 btrfs_mark_buffer_dirty(path->nodes[level]);
1821
1822 ret = btrfs_inc_extent_ref(trans, src, old_bytenr, blocksize,
1823 path->nodes[level]->start,
1824 src->root_key.objectid, level - 1, 0,
1825 1);
1826 BUG_ON(ret);
1827 ret = btrfs_inc_extent_ref(trans, dest, new_bytenr, blocksize,
1828 0, dest->root_key.objectid, level - 1,
1829 0, 1);
1830 BUG_ON(ret);
1831
1832 ret = btrfs_free_extent(trans, src, new_bytenr, blocksize,
1833 path->nodes[level]->start,
1834 src->root_key.objectid, level - 1, 0,
1835 1);
1836 BUG_ON(ret);
1837
1838 ret = btrfs_free_extent(trans, dest, old_bytenr, blocksize,
1839 0, dest->root_key.objectid, level - 1,
1840 0, 1);
1841 BUG_ON(ret);
1842
1843 btrfs_unlock_up_safe(path, 0);
1844
1845 ret = level;
1846 break;
1847 }
1848 btrfs_tree_unlock(parent);
1849 free_extent_buffer(parent);
1850 return ret;
1851 }
1852
1853 /*
1854 * helper to find next relocated block in reloc tree
1855 */
1856 static noinline_for_stack
1857 int walk_up_reloc_tree(struct btrfs_root *root, struct btrfs_path *path,
1858 int *level)
1859 {
1860 struct extent_buffer *eb;
1861 int i;
1862 u64 last_snapshot;
1863 u32 nritems;
1864
1865 last_snapshot = btrfs_root_last_snapshot(&root->root_item);
1866
1867 for (i = 0; i < *level; i++) {
1868 free_extent_buffer(path->nodes[i]);
1869 path->nodes[i] = NULL;
1870 }
1871
1872 for (i = *level; i < BTRFS_MAX_LEVEL && path->nodes[i]; i++) {
1873 eb = path->nodes[i];
1874 nritems = btrfs_header_nritems(eb);
1875 while (path->slots[i] + 1 < nritems) {
1876 path->slots[i]++;
1877 if (btrfs_node_ptr_generation(eb, path->slots[i]) <=
1878 last_snapshot)
1879 continue;
1880
1881 *level = i;
1882 return 0;
1883 }
1884 free_extent_buffer(path->nodes[i]);
1885 path->nodes[i] = NULL;
1886 }
1887 return 1;
1888 }
1889
1890 /*
1891 * walk down reloc tree to find relocated block of lowest level
1892 */
1893 static noinline_for_stack
1894 int walk_down_reloc_tree(struct btrfs_root *root, struct btrfs_path *path,
1895 int *level)
1896 {
1897 struct extent_buffer *eb = NULL;
1898 int i;
1899 u64 bytenr;
1900 u64 ptr_gen = 0;
1901 u64 last_snapshot;
1902 u32 blocksize;
1903 u32 nritems;
1904
1905 last_snapshot = btrfs_root_last_snapshot(&root->root_item);
1906
1907 for (i = *level; i > 0; i--) {
1908 eb = path->nodes[i];
1909 nritems = btrfs_header_nritems(eb);
1910 while (path->slots[i] < nritems) {
1911 ptr_gen = btrfs_node_ptr_generation(eb, path->slots[i]);
1912 if (ptr_gen > last_snapshot)
1913 break;
1914 path->slots[i]++;
1915 }
1916 if (path->slots[i] >= nritems) {
1917 if (i == *level)
1918 break;
1919 *level = i + 1;
1920 return 0;
1921 }
1922 if (i == 1) {
1923 *level = i;
1924 return 0;
1925 }
1926
1927 bytenr = btrfs_node_blockptr(eb, path->slots[i]);
1928 blocksize = btrfs_level_size(root, i - 1);
1929 eb = read_tree_block(root, bytenr, blocksize, ptr_gen);
1930 if (!eb || !extent_buffer_uptodate(eb)) {
1931 free_extent_buffer(eb);
1932 return -EIO;
1933 }
1934 BUG_ON(btrfs_header_level(eb) != i - 1);
1935 path->nodes[i - 1] = eb;
1936 path->slots[i - 1] = 0;
1937 }
1938 return 1;
1939 }
1940
1941 /*
1942 * invalidate extent cache for file extents whose key in range of
1943 * [min_key, max_key)
1944 */
1945 static int invalidate_extent_cache(struct btrfs_root *root,
1946 struct btrfs_key *min_key,
1947 struct btrfs_key *max_key)
1948 {
1949 struct inode *inode = NULL;
1950 u64 objectid;
1951 u64 start, end;
1952 u64 ino;
1953
1954 objectid = min_key->objectid;
1955 while (1) {
1956 cond_resched();
1957 iput(inode);
1958
1959 if (objectid > max_key->objectid)
1960 break;
1961
1962 inode = find_next_inode(root, objectid);
1963 if (!inode)
1964 break;
1965 ino = btrfs_ino(inode);
1966
1967 if (ino > max_key->objectid) {
1968 iput(inode);
1969 break;
1970 }
1971
1972 objectid = ino + 1;
1973 if (!S_ISREG(inode->i_mode))
1974 continue;
1975
1976 if (unlikely(min_key->objectid == ino)) {
1977 if (min_key->type > BTRFS_EXTENT_DATA_KEY)
1978 continue;
1979 if (min_key->type < BTRFS_EXTENT_DATA_KEY)
1980 start = 0;
1981 else {
1982 start = min_key->offset;
1983 WARN_ON(!IS_ALIGNED(start, root->sectorsize));
1984 }
1985 } else {
1986 start = 0;
1987 }
1988
1989 if (unlikely(max_key->objectid == ino)) {
1990 if (max_key->type < BTRFS_EXTENT_DATA_KEY)
1991 continue;
1992 if (max_key->type > BTRFS_EXTENT_DATA_KEY) {
1993 end = (u64)-1;
1994 } else {
1995 if (max_key->offset == 0)
1996 continue;
1997 end = max_key->offset;
1998 WARN_ON(!IS_ALIGNED(end, root->sectorsize));
1999 end--;
2000 }
2001 } else {
2002 end = (u64)-1;
2003 }
2004
2005 /* the lock_extent waits for readpage to complete */
2006 lock_extent(&BTRFS_I(inode)->io_tree, start, end);
2007 btrfs_drop_extent_cache(inode, start, end, 1);
2008 unlock_extent(&BTRFS_I(inode)->io_tree, start, end);
2009 }
2010 return 0;
2011 }
2012
2013 static int find_next_key(struct btrfs_path *path, int level,
2014 struct btrfs_key *key)
2015
2016 {
2017 while (level < BTRFS_MAX_LEVEL) {
2018 if (!path->nodes[level])
2019 break;
2020 if (path->slots[level] + 1 <
2021 btrfs_header_nritems(path->nodes[level])) {
2022 btrfs_node_key_to_cpu(path->nodes[level], key,
2023 path->slots[level] + 1);
2024 return 0;
2025 }
2026 level++;
2027 }
2028 return 1;
2029 }
2030
2031 /*
2032 * merge the relocated tree blocks in reloc tree with corresponding
2033 * fs tree.
2034 */
2035 static noinline_for_stack int merge_reloc_root(struct reloc_control *rc,
2036 struct btrfs_root *root)
2037 {
2038 LIST_HEAD(inode_list);
2039 struct btrfs_key key;
2040 struct btrfs_key next_key;
2041 struct btrfs_trans_handle *trans;
2042 struct btrfs_root *reloc_root;
2043 struct btrfs_root_item *root_item;
2044 struct btrfs_path *path;
2045 struct extent_buffer *leaf;
2046 int level;
2047 int max_level;
2048 int replaced = 0;
2049 int ret;
2050 int err = 0;
2051 u32 min_reserved;
2052
2053 path = btrfs_alloc_path();
2054 if (!path)
2055 return -ENOMEM;
2056 path->reada = 1;
2057
2058 reloc_root = root->reloc_root;
2059 root_item = &reloc_root->root_item;
2060
2061 if (btrfs_disk_key_objectid(&root_item->drop_progress) == 0) {
2062 level = btrfs_root_level(root_item);
2063 extent_buffer_get(reloc_root->node);
2064 path->nodes[level] = reloc_root->node;
2065 path->slots[level] = 0;
2066 } else {
2067 btrfs_disk_key_to_cpu(&key, &root_item->drop_progress);
2068
2069 level = root_item->drop_level;
2070 BUG_ON(level == 0);
2071 path->lowest_level = level;
2072 ret = btrfs_search_slot(NULL, reloc_root, &key, path, 0, 0);
2073 path->lowest_level = 0;
2074 if (ret < 0) {
2075 btrfs_free_path(path);
2076 return ret;
2077 }
2078
2079 btrfs_node_key_to_cpu(path->nodes[level], &next_key,
2080 path->slots[level]);
2081 WARN_ON(memcmp(&key, &next_key, sizeof(key)));
2082
2083 btrfs_unlock_up_safe(path, 0);
2084 }
2085
2086 min_reserved = root->nodesize * (BTRFS_MAX_LEVEL - 1) * 2;
2087 memset(&next_key, 0, sizeof(next_key));
2088
2089 while (1) {
2090 trans = btrfs_start_transaction(root, 0);
2091 BUG_ON(IS_ERR(trans));
2092 trans->block_rsv = rc->block_rsv;
2093
2094 ret = btrfs_block_rsv_refill(root, rc->block_rsv, min_reserved,
2095 BTRFS_RESERVE_FLUSH_ALL);
2096 if (ret) {
2097 BUG_ON(ret != -EAGAIN);
2098 ret = btrfs_commit_transaction(trans, root);
2099 BUG_ON(ret);
2100 continue;
2101 }
2102
2103 replaced = 0;
2104 max_level = level;
2105
2106 ret = walk_down_reloc_tree(reloc_root, path, &level);
2107 if (ret < 0) {
2108 err = ret;
2109 goto out;
2110 }
2111 if (ret > 0)
2112 break;
2113
2114 if (!find_next_key(path, level, &key) &&
2115 btrfs_comp_cpu_keys(&next_key, &key) >= 0) {
2116 ret = 0;
2117 } else {
2118 ret = replace_path(trans, root, reloc_root, path,
2119 &next_key, level, max_level);
2120 }
2121 if (ret < 0) {
2122 err = ret;
2123 goto out;
2124 }
2125
2126 if (ret > 0) {
2127 level = ret;
2128 btrfs_node_key_to_cpu(path->nodes[level], &key,
2129 path->slots[level]);
2130 replaced = 1;
2131 }
2132
2133 ret = walk_up_reloc_tree(reloc_root, path, &level);
2134 if (ret > 0)
2135 break;
2136
2137 BUG_ON(level == 0);
2138 /*
2139 * save the merging progress in the drop_progress.
2140 * this is OK since root refs == 1 in this case.
2141 */
2142 btrfs_node_key(path->nodes[level], &root_item->drop_progress,
2143 path->slots[level]);
2144 root_item->drop_level = level;
2145
2146 btrfs_end_transaction_throttle(trans, root);
2147
2148 btrfs_btree_balance_dirty(root);
2149
2150 if (replaced && rc->stage == UPDATE_DATA_PTRS)
2151 invalidate_extent_cache(root, &key, &next_key);
2152 }
2153
2154 /*
2155 * handle the case only one block in the fs tree need to be
2156 * relocated and the block is tree root.
2157 */
2158 leaf = btrfs_lock_root_node(root);
2159 ret = btrfs_cow_block(trans, root, leaf, NULL, 0, &leaf);
2160 btrfs_tree_unlock(leaf);
2161 free_extent_buffer(leaf);
2162 if (ret < 0)
2163 err = ret;
2164 out:
2165 btrfs_free_path(path);
2166
2167 if (err == 0) {
2168 memset(&root_item->drop_progress, 0,
2169 sizeof(root_item->drop_progress));
2170 root_item->drop_level = 0;
2171 btrfs_set_root_refs(root_item, 0);
2172 btrfs_update_reloc_root(trans, root);
2173 }
2174
2175 btrfs_end_transaction_throttle(trans, root);
2176
2177 btrfs_btree_balance_dirty(root);
2178
2179 if (replaced && rc->stage == UPDATE_DATA_PTRS)
2180 invalidate_extent_cache(root, &key, &next_key);
2181
2182 return err;
2183 }
2184
2185 static noinline_for_stack
2186 int prepare_to_merge(struct reloc_control *rc, int err)
2187 {
2188 struct btrfs_root *root = rc->extent_root;
2189 struct btrfs_root *reloc_root;
2190 struct btrfs_trans_handle *trans;
2191 LIST_HEAD(reloc_roots);
2192 u64 num_bytes = 0;
2193 int ret;
2194
2195 mutex_lock(&root->fs_info->reloc_mutex);
2196 rc->merging_rsv_size += root->nodesize * (BTRFS_MAX_LEVEL - 1) * 2;
2197 rc->merging_rsv_size += rc->nodes_relocated * 2;
2198 mutex_unlock(&root->fs_info->reloc_mutex);
2199
2200 again:
2201 if (!err) {
2202 num_bytes = rc->merging_rsv_size;
2203 ret = btrfs_block_rsv_add(root, rc->block_rsv, num_bytes,
2204 BTRFS_RESERVE_FLUSH_ALL);
2205 if (ret)
2206 err = ret;
2207 }
2208
2209 trans = btrfs_join_transaction(rc->extent_root);
2210 if (IS_ERR(trans)) {
2211 if (!err)
2212 btrfs_block_rsv_release(rc->extent_root,
2213 rc->block_rsv, num_bytes);
2214 return PTR_ERR(trans);
2215 }
2216
2217 if (!err) {
2218 if (num_bytes != rc->merging_rsv_size) {
2219 btrfs_end_transaction(trans, rc->extent_root);
2220 btrfs_block_rsv_release(rc->extent_root,
2221 rc->block_rsv, num_bytes);
2222 goto again;
2223 }
2224 }
2225
2226 rc->merge_reloc_tree = 1;
2227
2228 while (!list_empty(&rc->reloc_roots)) {
2229 reloc_root = list_entry(rc->reloc_roots.next,
2230 struct btrfs_root, root_list);
2231 list_del_init(&reloc_root->root_list);
2232
2233 root = read_fs_root(reloc_root->fs_info,
2234 reloc_root->root_key.offset);
2235 BUG_ON(IS_ERR(root));
2236 BUG_ON(root->reloc_root != reloc_root);
2237
2238 /*
2239 * set reference count to 1, so btrfs_recover_relocation
2240 * knows it should resumes merging
2241 */
2242 if (!err)
2243 btrfs_set_root_refs(&reloc_root->root_item, 1);
2244 btrfs_update_reloc_root(trans, root);
2245
2246 list_add(&reloc_root->root_list, &reloc_roots);
2247 }
2248
2249 list_splice(&reloc_roots, &rc->reloc_roots);
2250
2251 if (!err)
2252 btrfs_commit_transaction(trans, rc->extent_root);
2253 else
2254 btrfs_end_transaction(trans, rc->extent_root);
2255 return err;
2256 }
2257
2258 static noinline_for_stack
2259 void free_reloc_roots(struct list_head *list)
2260 {
2261 struct btrfs_root *reloc_root;
2262
2263 while (!list_empty(list)) {
2264 reloc_root = list_entry(list->next, struct btrfs_root,
2265 root_list);
2266 __update_reloc_root(reloc_root, 1);
2267 free_extent_buffer(reloc_root->node);
2268 free_extent_buffer(reloc_root->commit_root);
2269 kfree(reloc_root);
2270 }
2271 }
2272
2273 static noinline_for_stack
2274 int merge_reloc_roots(struct reloc_control *rc)
2275 {
2276 struct btrfs_root *root;
2277 struct btrfs_root *reloc_root;
2278 LIST_HEAD(reloc_roots);
2279 int found = 0;
2280 int ret = 0;
2281 again:
2282 root = rc->extent_root;
2283
2284 /*
2285 * this serializes us with btrfs_record_root_in_transaction,
2286 * we have to make sure nobody is in the middle of
2287 * adding their roots to the list while we are
2288 * doing this splice
2289 */
2290 mutex_lock(&root->fs_info->reloc_mutex);
2291 list_splice_init(&rc->reloc_roots, &reloc_roots);
2292 mutex_unlock(&root->fs_info->reloc_mutex);
2293
2294 while (!list_empty(&reloc_roots)) {
2295 found = 1;
2296 reloc_root = list_entry(reloc_roots.next,
2297 struct btrfs_root, root_list);
2298
2299 if (btrfs_root_refs(&reloc_root->root_item) > 0) {
2300 root = read_fs_root(reloc_root->fs_info,
2301 reloc_root->root_key.offset);
2302 BUG_ON(IS_ERR(root));
2303 BUG_ON(root->reloc_root != reloc_root);
2304
2305 ret = merge_reloc_root(rc, root);
2306 if (ret)
2307 goto out;
2308 } else {
2309 list_del_init(&reloc_root->root_list);
2310 }
2311 ret = btrfs_drop_snapshot(reloc_root, rc->block_rsv, 0, 1);
2312 if (ret < 0) {
2313 if (list_empty(&reloc_root->root_list))
2314 list_add_tail(&reloc_root->root_list,
2315 &reloc_roots);
2316 goto out;
2317 }
2318 }
2319
2320 if (found) {
2321 found = 0;
2322 goto again;
2323 }
2324 out:
2325 if (ret) {
2326 btrfs_std_error(root->fs_info, ret);
2327 if (!list_empty(&reloc_roots))
2328 free_reloc_roots(&reloc_roots);
2329 }
2330
2331 BUG_ON(!RB_EMPTY_ROOT(&rc->reloc_root_tree.rb_root));
2332 return ret;
2333 }
2334
2335 static void free_block_list(struct rb_root *blocks)
2336 {
2337 struct tree_block *block;
2338 struct rb_node *rb_node;
2339 while ((rb_node = rb_first(blocks))) {
2340 block = rb_entry(rb_node, struct tree_block, rb_node);
2341 rb_erase(rb_node, blocks);
2342 kfree(block);
2343 }
2344 }
2345
2346 static int record_reloc_root_in_trans(struct btrfs_trans_handle *trans,
2347 struct btrfs_root *reloc_root)
2348 {
2349 struct btrfs_root *root;
2350
2351 if (reloc_root->last_trans == trans->transid)
2352 return 0;
2353
2354 root = read_fs_root(reloc_root->fs_info, reloc_root->root_key.offset);
2355 BUG_ON(IS_ERR(root));
2356 BUG_ON(root->reloc_root != reloc_root);
2357
2358 return btrfs_record_root_in_trans(trans, root);
2359 }
2360
2361 static noinline_for_stack
2362 struct btrfs_root *select_reloc_root(struct btrfs_trans_handle *trans,
2363 struct reloc_control *rc,
2364 struct backref_node *node,
2365 struct backref_edge *edges[], int *nr)
2366 {
2367 struct backref_node *next;
2368 struct btrfs_root *root;
2369 int index = 0;
2370
2371 next = node;
2372 while (1) {
2373 cond_resched();
2374 next = walk_up_backref(next, edges, &index);
2375 root = next->root;
2376 BUG_ON(!root);
2377 BUG_ON(!root->ref_cows);
2378
2379 if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID) {
2380 record_reloc_root_in_trans(trans, root);
2381 break;
2382 }
2383
2384 btrfs_record_root_in_trans(trans, root);
2385 root = root->reloc_root;
2386
2387 if (next->new_bytenr != root->node->start) {
2388 BUG_ON(next->new_bytenr);
2389 BUG_ON(!list_empty(&next->list));
2390 next->new_bytenr = root->node->start;
2391 next->root = root;
2392 list_add_tail(&next->list,
2393 &rc->backref_cache.changed);
2394 __mark_block_processed(rc, next);
2395 break;
2396 }
2397
2398 WARN_ON(1);
2399 root = NULL;
2400 next = walk_down_backref(edges, &index);
2401 if (!next || next->level <= node->level)
2402 break;
2403 }
2404 if (!root)
2405 return NULL;
2406
2407 *nr = index;
2408 next = node;
2409 /* setup backref node path for btrfs_reloc_cow_block */
2410 while (1) {
2411 rc->backref_cache.path[next->level] = next;
2412 if (--index < 0)
2413 break;
2414 next = edges[index]->node[UPPER];
2415 }
2416 return root;
2417 }
2418
2419 /*
2420 * select a tree root for relocation. return NULL if the block
2421 * is reference counted. we should use do_relocation() in this
2422 * case. return a tree root pointer if the block isn't reference
2423 * counted. return -ENOENT if the block is root of reloc tree.
2424 */
2425 static noinline_for_stack
2426 struct btrfs_root *select_one_root(struct btrfs_trans_handle *trans,
2427 struct backref_node *node)
2428 {
2429 struct backref_node *next;
2430 struct btrfs_root *root;
2431 struct btrfs_root *fs_root = NULL;
2432 struct backref_edge *edges[BTRFS_MAX_LEVEL - 1];
2433 int index = 0;
2434
2435 next = node;
2436 while (1) {
2437 cond_resched();
2438 next = walk_up_backref(next, edges, &index);
2439 root = next->root;
2440 BUG_ON(!root);
2441
2442 /* no other choice for non-references counted tree */
2443 if (!root->ref_cows)
2444 return root;
2445
2446 if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID)
2447 fs_root = root;
2448
2449 if (next != node)
2450 return NULL;
2451
2452 next = walk_down_backref(edges, &index);
2453 if (!next || next->level <= node->level)
2454 break;
2455 }
2456
2457 if (!fs_root)
2458 return ERR_PTR(-ENOENT);
2459 return fs_root;
2460 }
2461
2462 static noinline_for_stack
2463 u64 calcu_metadata_size(struct reloc_control *rc,
2464 struct backref_node *node, int reserve)
2465 {
2466 struct backref_node *next = node;
2467 struct backref_edge *edge;
2468 struct backref_edge *edges[BTRFS_MAX_LEVEL - 1];
2469 u64 num_bytes = 0;
2470 int index = 0;
2471
2472 BUG_ON(reserve && node->processed);
2473
2474 while (next) {
2475 cond_resched();
2476 while (1) {
2477 if (next->processed && (reserve || next != node))
2478 break;
2479
2480 num_bytes += btrfs_level_size(rc->extent_root,
2481 next->level);
2482
2483 if (list_empty(&next->upper))
2484 break;
2485
2486 edge = list_entry(next->upper.next,
2487 struct backref_edge, list[LOWER]);
2488 edges[index++] = edge;
2489 next = edge->node[UPPER];
2490 }
2491 next = walk_down_backref(edges, &index);
2492 }
2493 return num_bytes;
2494 }
2495
2496 static int reserve_metadata_space(struct btrfs_trans_handle *trans,
2497 struct reloc_control *rc,
2498 struct backref_node *node)
2499 {
2500 struct btrfs_root *root = rc->extent_root;
2501 u64 num_bytes;
2502 int ret;
2503
2504 num_bytes = calcu_metadata_size(rc, node, 1) * 2;
2505
2506 trans->block_rsv = rc->block_rsv;
2507 ret = btrfs_block_rsv_add(root, rc->block_rsv, num_bytes,
2508 BTRFS_RESERVE_FLUSH_ALL);
2509 if (ret) {
2510 if (ret == -EAGAIN)
2511 rc->commit_transaction = 1;
2512 return ret;
2513 }
2514
2515 return 0;
2516 }
2517
2518 static void release_metadata_space(struct reloc_control *rc,
2519 struct backref_node *node)
2520 {
2521 u64 num_bytes = calcu_metadata_size(rc, node, 0) * 2;
2522 btrfs_block_rsv_release(rc->extent_root, rc->block_rsv, num_bytes);
2523 }
2524
2525 /*
2526 * relocate a block tree, and then update pointers in upper level
2527 * blocks that reference the block to point to the new location.
2528 *
2529 * if called by link_to_upper, the block has already been relocated.
2530 * in that case this function just updates pointers.
2531 */
2532 static int do_relocation(struct btrfs_trans_handle *trans,
2533 struct reloc_control *rc,
2534 struct backref_node *node,
2535 struct btrfs_key *key,
2536 struct btrfs_path *path, int lowest)
2537 {
2538 struct backref_node *upper;
2539 struct backref_edge *edge;
2540 struct backref_edge *edges[BTRFS_MAX_LEVEL - 1];
2541 struct btrfs_root *root;
2542 struct extent_buffer *eb;
2543 u32 blocksize;
2544 u64 bytenr;
2545 u64 generation;
2546 int nr;
2547 int slot;
2548 int ret;
2549 int err = 0;
2550
2551 BUG_ON(lowest && node->eb);
2552
2553 path->lowest_level = node->level + 1;
2554 rc->backref_cache.path[node->level] = node;
2555 list_for_each_entry(edge, &node->upper, list[LOWER]) {
2556 cond_resched();
2557
2558 upper = edge->node[UPPER];
2559 root = select_reloc_root(trans, rc, upper, edges, &nr);
2560 BUG_ON(!root);
2561
2562 if (upper->eb && !upper->locked) {
2563 if (!lowest) {
2564 ret = btrfs_bin_search(upper->eb, key,
2565 upper->level, &slot);
2566 BUG_ON(ret);
2567 bytenr = btrfs_node_blockptr(upper->eb, slot);
2568 if (node->eb->start == bytenr)
2569 goto next;
2570 }
2571 drop_node_buffer(upper);
2572 }
2573
2574 if (!upper->eb) {
2575 ret = btrfs_search_slot(trans, root, key, path, 0, 1);
2576 if (ret < 0) {
2577 err = ret;
2578 break;
2579 }
2580 BUG_ON(ret > 0);
2581
2582 if (!upper->eb) {
2583 upper->eb = path->nodes[upper->level];
2584 path->nodes[upper->level] = NULL;
2585 } else {
2586 BUG_ON(upper->eb != path->nodes[upper->level]);
2587 }
2588
2589 upper->locked = 1;
2590 path->locks[upper->level] = 0;
2591
2592 slot = path->slots[upper->level];
2593 btrfs_release_path(path);
2594 } else {
2595 ret = btrfs_bin_search(upper->eb, key, upper->level,
2596 &slot);
2597 BUG_ON(ret);
2598 }
2599
2600 bytenr = btrfs_node_blockptr(upper->eb, slot);
2601 if (lowest) {
2602 BUG_ON(bytenr != node->bytenr);
2603 } else {
2604 if (node->eb->start == bytenr)
2605 goto next;
2606 }
2607
2608 blocksize = btrfs_level_size(root, node->level);
2609 generation = btrfs_node_ptr_generation(upper->eb, slot);
2610 eb = read_tree_block(root, bytenr, blocksize, generation);
2611 if (!eb || !extent_buffer_uptodate(eb)) {
2612 free_extent_buffer(eb);
2613 err = -EIO;
2614 goto next;
2615 }
2616 btrfs_tree_lock(eb);
2617 btrfs_set_lock_blocking(eb);
2618
2619 if (!node->eb) {
2620 ret = btrfs_cow_block(trans, root, eb, upper->eb,
2621 slot, &eb);
2622 btrfs_tree_unlock(eb);
2623 free_extent_buffer(eb);
2624 if (ret < 0) {
2625 err = ret;
2626 goto next;
2627 }
2628 BUG_ON(node->eb != eb);
2629 } else {
2630 btrfs_set_node_blockptr(upper->eb, slot,
2631 node->eb->start);
2632 btrfs_set_node_ptr_generation(upper->eb, slot,
2633 trans->transid);
2634 btrfs_mark_buffer_dirty(upper->eb);
2635
2636 ret = btrfs_inc_extent_ref(trans, root,
2637 node->eb->start, blocksize,
2638 upper->eb->start,
2639 btrfs_header_owner(upper->eb),
2640 node->level, 0, 1);
2641 BUG_ON(ret);
2642
2643 ret = btrfs_drop_subtree(trans, root, eb, upper->eb);
2644 BUG_ON(ret);
2645 }
2646 next:
2647 if (!upper->pending)
2648 drop_node_buffer(upper);
2649 else
2650 unlock_node_buffer(upper);
2651 if (err)
2652 break;
2653 }
2654
2655 if (!err && node->pending) {
2656 drop_node_buffer(node);
2657 list_move_tail(&node->list, &rc->backref_cache.changed);
2658 node->pending = 0;
2659 }
2660
2661 path->lowest_level = 0;
2662 BUG_ON(err == -ENOSPC);
2663 return err;
2664 }
2665
2666 static int link_to_upper(struct btrfs_trans_handle *trans,
2667 struct reloc_control *rc,
2668 struct backref_node *node,
2669 struct btrfs_path *path)
2670 {
2671 struct btrfs_key key;
2672
2673 btrfs_node_key_to_cpu(node->eb, &key, 0);
2674 return do_relocation(trans, rc, node, &key, path, 0);
2675 }
2676
2677 static int finish_pending_nodes(struct btrfs_trans_handle *trans,
2678 struct reloc_control *rc,
2679 struct btrfs_path *path, int err)
2680 {
2681 LIST_HEAD(list);
2682 struct backref_cache *cache = &rc->backref_cache;
2683 struct backref_node *node;
2684 int level;
2685 int ret;
2686
2687 for (level = 0; level < BTRFS_MAX_LEVEL; level++) {
2688 while (!list_empty(&cache->pending[level])) {
2689 node = list_entry(cache->pending[level].next,
2690 struct backref_node, list);
2691 list_move_tail(&node->list, &list);
2692 BUG_ON(!node->pending);
2693
2694 if (!err) {
2695 ret = link_to_upper(trans, rc, node, path);
2696 if (ret < 0)
2697 err = ret;
2698 }
2699 }
2700 list_splice_init(&list, &cache->pending[level]);
2701 }
2702 return err;
2703 }
2704
2705 static void mark_block_processed(struct reloc_control *rc,
2706 u64 bytenr, u32 blocksize)
2707 {
2708 set_extent_bits(&rc->processed_blocks, bytenr, bytenr + blocksize - 1,
2709 EXTENT_DIRTY, GFP_NOFS);
2710 }
2711
2712 static void __mark_block_processed(struct reloc_control *rc,
2713 struct backref_node *node)
2714 {
2715 u32 blocksize;
2716 if (node->level == 0 ||
2717 in_block_group(node->bytenr, rc->block_group)) {
2718 blocksize = btrfs_level_size(rc->extent_root, node->level);
2719 mark_block_processed(rc, node->bytenr, blocksize);
2720 }
2721 node->processed = 1;
2722 }
2723
2724 /*
2725 * mark a block and all blocks directly/indirectly reference the block
2726 * as processed.
2727 */
2728 static void update_processed_blocks(struct reloc_control *rc,
2729 struct backref_node *node)
2730 {
2731 struct backref_node *next = node;
2732 struct backref_edge *edge;
2733 struct backref_edge *edges[BTRFS_MAX_LEVEL - 1];
2734 int index = 0;
2735
2736 while (next) {
2737 cond_resched();
2738 while (1) {
2739 if (next->processed)
2740 break;
2741
2742 __mark_block_processed(rc, next);
2743
2744 if (list_empty(&next->upper))
2745 break;
2746
2747 edge = list_entry(next->upper.next,
2748 struct backref_edge, list[LOWER]);
2749 edges[index++] = edge;
2750 next = edge->node[UPPER];
2751 }
2752 next = walk_down_backref(edges, &index);
2753 }
2754 }
2755
2756 static int tree_block_processed(u64 bytenr, u32 blocksize,
2757 struct reloc_control *rc)
2758 {
2759 if (test_range_bit(&rc->processed_blocks, bytenr,
2760 bytenr + blocksize - 1, EXTENT_DIRTY, 1, NULL))
2761 return 1;
2762 return 0;
2763 }
2764
2765 static int get_tree_block_key(struct reloc_control *rc,
2766 struct tree_block *block)
2767 {
2768 struct extent_buffer *eb;
2769
2770 BUG_ON(block->key_ready);
2771 eb = read_tree_block(rc->extent_root, block->bytenr,
2772 block->key.objectid, block->key.offset);
2773 if (!eb || !extent_buffer_uptodate(eb)) {
2774 free_extent_buffer(eb);
2775 return -EIO;
2776 }
2777 WARN_ON(btrfs_header_level(eb) != block->level);
2778 if (block->level == 0)
2779 btrfs_item_key_to_cpu(eb, &block->key, 0);
2780 else
2781 btrfs_node_key_to_cpu(eb, &block->key, 0);
2782 free_extent_buffer(eb);
2783 block->key_ready = 1;
2784 return 0;
2785 }
2786
2787 static int reada_tree_block(struct reloc_control *rc,
2788 struct tree_block *block)
2789 {
2790 BUG_ON(block->key_ready);
2791 if (block->key.type == BTRFS_METADATA_ITEM_KEY)
2792 readahead_tree_block(rc->extent_root, block->bytenr,
2793 block->key.objectid,
2794 rc->extent_root->leafsize);
2795 else
2796 readahead_tree_block(rc->extent_root, block->bytenr,
2797 block->key.objectid, block->key.offset);
2798 return 0;
2799 }
2800
2801 /*
2802 * helper function to relocate a tree block
2803 */
2804 static int relocate_tree_block(struct btrfs_trans_handle *trans,
2805 struct reloc_control *rc,
2806 struct backref_node *node,
2807 struct btrfs_key *key,
2808 struct btrfs_path *path)
2809 {
2810 struct btrfs_root *root;
2811 int release = 0;
2812 int ret = 0;
2813
2814 if (!node)
2815 return 0;
2816
2817 BUG_ON(node->processed);
2818 root = select_one_root(trans, node);
2819 if (root == ERR_PTR(-ENOENT)) {
2820 update_processed_blocks(rc, node);
2821 goto out;
2822 }
2823
2824 if (!root || root->ref_cows) {
2825 ret = reserve_metadata_space(trans, rc, node);
2826 if (ret)
2827 goto out;
2828 release = 1;
2829 }
2830
2831 if (root) {
2832 if (root->ref_cows) {
2833 BUG_ON(node->new_bytenr);
2834 BUG_ON(!list_empty(&node->list));
2835 btrfs_record_root_in_trans(trans, root);
2836 root = root->reloc_root;
2837 node->new_bytenr = root->node->start;
2838 node->root = root;
2839 list_add_tail(&node->list, &rc->backref_cache.changed);
2840 } else {
2841 path->lowest_level = node->level;
2842 ret = btrfs_search_slot(trans, root, key, path, 0, 1);
2843 btrfs_release_path(path);
2844 if (ret > 0)
2845 ret = 0;
2846 }
2847 if (!ret)
2848 update_processed_blocks(rc, node);
2849 } else {
2850 ret = do_relocation(trans, rc, node, key, path, 1);
2851 }
2852 out:
2853 if (ret || node->level == 0 || node->cowonly) {
2854 if (release)
2855 release_metadata_space(rc, node);
2856 remove_backref_node(&rc->backref_cache, node);
2857 }
2858 return ret;
2859 }
2860
2861 /*
2862 * relocate a list of blocks
2863 */
2864 static noinline_for_stack
2865 int relocate_tree_blocks(struct btrfs_trans_handle *trans,
2866 struct reloc_control *rc, struct rb_root *blocks)
2867 {
2868 struct backref_node *node;
2869 struct btrfs_path *path;
2870 struct tree_block *block;
2871 struct rb_node *rb_node;
2872 int ret;
2873 int err = 0;
2874
2875 path = btrfs_alloc_path();
2876 if (!path) {
2877 err = -ENOMEM;
2878 goto out_free_blocks;
2879 }
2880
2881 rb_node = rb_first(blocks);
2882 while (rb_node) {
2883 block = rb_entry(rb_node, struct tree_block, rb_node);
2884 if (!block->key_ready)
2885 reada_tree_block(rc, block);
2886 rb_node = rb_next(rb_node);
2887 }
2888
2889 rb_node = rb_first(blocks);
2890 while (rb_node) {
2891 block = rb_entry(rb_node, struct tree_block, rb_node);
2892 if (!block->key_ready) {
2893 err = get_tree_block_key(rc, block);
2894 if (err)
2895 goto out_free_path;
2896 }
2897 rb_node = rb_next(rb_node);
2898 }
2899
2900 rb_node = rb_first(blocks);
2901 while (rb_node) {
2902 block = rb_entry(rb_node, struct tree_block, rb_node);
2903
2904 node = build_backref_tree(rc, &block->key,
2905 block->level, block->bytenr);
2906 if (IS_ERR(node)) {
2907 err = PTR_ERR(node);
2908 goto out;
2909 }
2910
2911 ret = relocate_tree_block(trans, rc, node, &block->key,
2912 path);
2913 if (ret < 0) {
2914 if (ret != -EAGAIN || rb_node == rb_first(blocks))
2915 err = ret;
2916 goto out;
2917 }
2918 rb_node = rb_next(rb_node);
2919 }
2920 out:
2921 err = finish_pending_nodes(trans, rc, path, err);
2922
2923 out_free_path:
2924 btrfs_free_path(path);
2925 out_free_blocks:
2926 free_block_list(blocks);
2927 return err;
2928 }
2929
2930 static noinline_for_stack
2931 int prealloc_file_extent_cluster(struct inode *inode,
2932 struct file_extent_cluster *cluster)
2933 {
2934 u64 alloc_hint = 0;
2935 u64 start;
2936 u64 end;
2937 u64 offset = BTRFS_I(inode)->index_cnt;
2938 u64 num_bytes;
2939 int nr = 0;
2940 int ret = 0;
2941
2942 BUG_ON(cluster->start != cluster->boundary[0]);
2943 mutex_lock(&inode->i_mutex);
2944
2945 ret = btrfs_check_data_free_space(inode, cluster->end +
2946 1 - cluster->start);
2947 if (ret)
2948 goto out;
2949
2950 while (nr < cluster->nr) {
2951 start = cluster->boundary[nr] - offset;
2952 if (nr + 1 < cluster->nr)
2953 end = cluster->boundary[nr + 1] - 1 - offset;
2954 else
2955 end = cluster->end - offset;
2956
2957 lock_extent(&BTRFS_I(inode)->io_tree, start, end);
2958 num_bytes = end + 1 - start;
2959 ret = btrfs_prealloc_file_range(inode, 0, start,
2960 num_bytes, num_bytes,
2961 end + 1, &alloc_hint);
2962 unlock_extent(&BTRFS_I(inode)->io_tree, start, end);
2963 if (ret)
2964 break;
2965 nr++;
2966 }
2967 btrfs_free_reserved_data_space(inode, cluster->end +
2968 1 - cluster->start);
2969 out:
2970 mutex_unlock(&inode->i_mutex);
2971 return ret;
2972 }
2973
2974 static noinline_for_stack
2975 int setup_extent_mapping(struct inode *inode, u64 start, u64 end,
2976 u64 block_start)
2977 {
2978 struct btrfs_root *root = BTRFS_I(inode)->root;
2979 struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
2980 struct extent_map *em;
2981 int ret = 0;
2982
2983 em = alloc_extent_map();
2984 if (!em)
2985 return -ENOMEM;
2986
2987 em->start = start;
2988 em->len = end + 1 - start;
2989 em->block_len = em->len;
2990 em->block_start = block_start;
2991 em->bdev = root->fs_info->fs_devices->latest_bdev;
2992 set_bit(EXTENT_FLAG_PINNED, &em->flags);
2993
2994 lock_extent(&BTRFS_I(inode)->io_tree, start, end);
2995 while (1) {
2996 write_lock(&em_tree->lock);
2997 ret = add_extent_mapping(em_tree, em, 0);
2998 write_unlock(&em_tree->lock);
2999 if (ret != -EEXIST) {
3000 free_extent_map(em);
3001 break;
3002 }
3003 btrfs_drop_extent_cache(inode, start, end, 0);
3004 }
3005 unlock_extent(&BTRFS_I(inode)->io_tree, start, end);
3006 return ret;
3007 }
3008
3009 static int relocate_file_extent_cluster(struct inode *inode,
3010 struct file_extent_cluster *cluster)
3011 {
3012 u64 page_start;
3013 u64 page_end;
3014 u64 offset = BTRFS_I(inode)->index_cnt;
3015 unsigned long index;
3016 unsigned long last_index;
3017 struct page *page;
3018 struct file_ra_state *ra;
3019 gfp_t mask = btrfs_alloc_write_mask(inode->i_mapping);
3020 int nr = 0;
3021 int ret = 0;
3022
3023 if (!cluster->nr)
3024 return 0;
3025
3026 ra = kzalloc(sizeof(*ra), GFP_NOFS);
3027 if (!ra)
3028 return -ENOMEM;
3029
3030 ret = prealloc_file_extent_cluster(inode, cluster);
3031 if (ret)
3032 goto out;
3033
3034 file_ra_state_init(ra, inode->i_mapping);
3035
3036 ret = setup_extent_mapping(inode, cluster->start - offset,
3037 cluster->end - offset, cluster->start);
3038 if (ret)
3039 goto out;
3040
3041 index = (cluster->start - offset) >> PAGE_CACHE_SHIFT;
3042 last_index = (cluster->end - offset) >> PAGE_CACHE_SHIFT;
3043 while (index <= last_index) {
3044 ret = btrfs_delalloc_reserve_metadata(inode, PAGE_CACHE_SIZE);
3045 if (ret)
3046 goto out;
3047
3048 page = find_lock_page(inode->i_mapping, index);
3049 if (!page) {
3050 page_cache_sync_readahead(inode->i_mapping,
3051 ra, NULL, index,
3052 last_index + 1 - index);
3053 page = find_or_create_page(inode->i_mapping, index,
3054 mask);
3055 if (!page) {
3056 btrfs_delalloc_release_metadata(inode,
3057 PAGE_CACHE_SIZE);
3058 ret = -ENOMEM;
3059 goto out;
3060 }
3061 }
3062
3063 if (PageReadahead(page)) {
3064 page_cache_async_readahead(inode->i_mapping,
3065 ra, NULL, page, index,
3066 last_index + 1 - index);
3067 }
3068
3069 if (!PageUptodate(page)) {
3070 btrfs_readpage(NULL, page);
3071 lock_page(page);
3072 if (!PageUptodate(page)) {
3073 unlock_page(page);
3074 page_cache_release(page);
3075 btrfs_delalloc_release_metadata(inode,
3076 PAGE_CACHE_SIZE);
3077 ret = -EIO;
3078 goto out;
3079 }
3080 }
3081
3082 page_start = page_offset(page);
3083 page_end = page_start + PAGE_CACHE_SIZE - 1;
3084
3085 lock_extent(&BTRFS_I(inode)->io_tree, page_start, page_end);
3086
3087 set_page_extent_mapped(page);
3088
3089 if (nr < cluster->nr &&
3090 page_start + offset == cluster->boundary[nr]) {
3091 set_extent_bits(&BTRFS_I(inode)->io_tree,
3092 page_start, page_end,
3093 EXTENT_BOUNDARY, GFP_NOFS);
3094 nr++;
3095 }
3096
3097 btrfs_set_extent_delalloc(inode, page_start, page_end, NULL);
3098 set_page_dirty(page);
3099
3100 unlock_extent(&BTRFS_I(inode)->io_tree,
3101 page_start, page_end);
3102 unlock_page(page);
3103 page_cache_release(page);
3104
3105 index++;
3106 balance_dirty_pages_ratelimited(inode->i_mapping);
3107 btrfs_throttle(BTRFS_I(inode)->root);
3108 }
3109 WARN_ON(nr != cluster->nr);
3110 out:
3111 kfree(ra);
3112 return ret;
3113 }
3114
3115 static noinline_for_stack
3116 int relocate_data_extent(struct inode *inode, struct btrfs_key *extent_key,
3117 struct file_extent_cluster *cluster)
3118 {
3119 int ret;
3120
3121 if (cluster->nr > 0 && extent_key->objectid != cluster->end + 1) {
3122 ret = relocate_file_extent_cluster(inode, cluster);
3123 if (ret)
3124 return ret;
3125 cluster->nr = 0;
3126 }
3127
3128 if (!cluster->nr)
3129 cluster->start = extent_key->objectid;
3130 else
3131 BUG_ON(cluster->nr >= MAX_EXTENTS);
3132 cluster->end = extent_key->objectid + extent_key->offset - 1;
3133 cluster->boundary[cluster->nr] = extent_key->objectid;
3134 cluster->nr++;
3135
3136 if (cluster->nr >= MAX_EXTENTS) {
3137 ret = relocate_file_extent_cluster(inode, cluster);
3138 if (ret)
3139 return ret;
3140 cluster->nr = 0;
3141 }
3142 return 0;
3143 }
3144
3145 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
3146 static int get_ref_objectid_v0(struct reloc_control *rc,
3147 struct btrfs_path *path,
3148 struct btrfs_key *extent_key,
3149 u64 *ref_objectid, int *path_change)
3150 {
3151 struct btrfs_key key;
3152 struct extent_buffer *leaf;
3153 struct btrfs_extent_ref_v0 *ref0;
3154 int ret;
3155 int slot;
3156
3157 leaf = path->nodes[0];
3158 slot = path->slots[0];
3159 while (1) {
3160 if (slot >= btrfs_header_nritems(leaf)) {
3161 ret = btrfs_next_leaf(rc->extent_root, path);
3162 if (ret < 0)
3163 return ret;
3164 BUG_ON(ret > 0);
3165 leaf = path->nodes[0];
3166 slot = path->slots[0];
3167 if (path_change)
3168 *path_change = 1;
3169 }
3170 btrfs_item_key_to_cpu(leaf, &key, slot);
3171 if (key.objectid != extent_key->objectid)
3172 return -ENOENT;
3173
3174 if (key.type != BTRFS_EXTENT_REF_V0_KEY) {
3175 slot++;
3176 continue;
3177 }
3178 ref0 = btrfs_item_ptr(leaf, slot,
3179 struct btrfs_extent_ref_v0);
3180 *ref_objectid = btrfs_ref_objectid_v0(leaf, ref0);
3181 break;
3182 }
3183 return 0;
3184 }
3185 #endif
3186
3187 /*
3188 * helper to add a tree block to the list.
3189 * the major work is getting the generation and level of the block
3190 */
3191 static int add_tree_block(struct reloc_control *rc,
3192 struct btrfs_key *extent_key,
3193 struct btrfs_path *path,
3194 struct rb_root *blocks)
3195 {
3196 struct extent_buffer *eb;
3197 struct btrfs_extent_item *ei;
3198 struct btrfs_tree_block_info *bi;
3199 struct tree_block *block;
3200 struct rb_node *rb_node;
3201 u32 item_size;
3202 int level = -1;
3203 int generation;
3204
3205 eb = path->nodes[0];
3206 item_size = btrfs_item_size_nr(eb, path->slots[0]);
3207
3208 if (extent_key->type == BTRFS_METADATA_ITEM_KEY ||
3209 item_size >= sizeof(*ei) + sizeof(*bi)) {
3210 ei = btrfs_item_ptr(eb, path->slots[0],
3211 struct btrfs_extent_item);
3212 if (extent_key->type == BTRFS_EXTENT_ITEM_KEY) {
3213 bi = (struct btrfs_tree_block_info *)(ei + 1);
3214 level = btrfs_tree_block_level(eb, bi);
3215 } else {
3216 level = (int)extent_key->offset;
3217 }
3218 generation = btrfs_extent_generation(eb, ei);
3219 } else {
3220 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
3221 u64 ref_owner;
3222 int ret;
3223
3224 BUG_ON(item_size != sizeof(struct btrfs_extent_item_v0));
3225 ret = get_ref_objectid_v0(rc, path, extent_key,
3226 &ref_owner, NULL);
3227 if (ret < 0)
3228 return ret;
3229 BUG_ON(ref_owner >= BTRFS_MAX_LEVEL);
3230 level = (int)ref_owner;
3231 /* FIXME: get real generation */
3232 generation = 0;
3233 #else
3234 BUG();
3235 #endif
3236 }
3237
3238 btrfs_release_path(path);
3239
3240 BUG_ON(level == -1);
3241
3242 block = kmalloc(sizeof(*block), GFP_NOFS);
3243 if (!block)
3244 return -ENOMEM;
3245
3246 block->bytenr = extent_key->objectid;
3247 block->key.objectid = rc->extent_root->leafsize;
3248 block->key.offset = generation;
3249 block->level = level;
3250 block->key_ready = 0;
3251
3252 rb_node = tree_insert(blocks, block->bytenr, &block->rb_node);
3253 if (rb_node)
3254 backref_tree_panic(rb_node, -EEXIST, block->bytenr);
3255
3256 return 0;
3257 }
3258
3259 /*
3260 * helper to add tree blocks for backref of type BTRFS_SHARED_DATA_REF_KEY
3261 */
3262 static int __add_tree_block(struct reloc_control *rc,
3263 u64 bytenr, u32 blocksize,
3264 struct rb_root *blocks)
3265 {
3266 struct btrfs_path *path;
3267 struct btrfs_key key;
3268 int ret;
3269
3270 if (tree_block_processed(bytenr, blocksize, rc))
3271 return 0;
3272
3273 if (tree_search(blocks, bytenr))
3274 return 0;
3275
3276 path = btrfs_alloc_path();
3277 if (!path)
3278 return -ENOMEM;
3279
3280 key.objectid = bytenr;
3281 key.type = BTRFS_EXTENT_ITEM_KEY;
3282 key.offset = blocksize;
3283
3284 path->search_commit_root = 1;
3285 path->skip_locking = 1;
3286 ret = btrfs_search_slot(NULL, rc->extent_root, &key, path, 0, 0);
3287 if (ret < 0)
3288 goto out;
3289
3290 btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0]);
3291 if (ret > 0) {
3292 if (key.objectid == bytenr &&
3293 key.type == BTRFS_METADATA_ITEM_KEY)
3294 ret = 0;
3295 }
3296 BUG_ON(ret);
3297
3298 ret = add_tree_block(rc, &key, path, blocks);
3299 out:
3300 btrfs_free_path(path);
3301 return ret;
3302 }
3303
3304 /*
3305 * helper to check if the block use full backrefs for pointers in it
3306 */
3307 static int block_use_full_backref(struct reloc_control *rc,
3308 struct extent_buffer *eb)
3309 {
3310 u64 flags;
3311 int ret;
3312
3313 if (btrfs_header_flag(eb, BTRFS_HEADER_FLAG_RELOC) ||
3314 btrfs_header_backref_rev(eb) < BTRFS_MIXED_BACKREF_REV)
3315 return 1;
3316
3317 ret = btrfs_lookup_extent_info(NULL, rc->extent_root,
3318 eb->start, btrfs_header_level(eb), 1,
3319 NULL, &flags);
3320 BUG_ON(ret);
3321
3322 if (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF)
3323 ret = 1;
3324 else
3325 ret = 0;
3326 return ret;
3327 }
3328
3329 static int delete_block_group_cache(struct btrfs_fs_info *fs_info,
3330 struct inode *inode, u64 ino)
3331 {
3332 struct btrfs_key key;
3333 struct btrfs_path *path;
3334 struct btrfs_root *root = fs_info->tree_root;
3335 struct btrfs_trans_handle *trans;
3336 int ret = 0;
3337
3338 if (inode)
3339 goto truncate;
3340
3341 key.objectid = ino;
3342 key.type = BTRFS_INODE_ITEM_KEY;
3343 key.offset = 0;
3344
3345 inode = btrfs_iget(fs_info->sb, &key, root, NULL);
3346 if (IS_ERR(inode) || is_bad_inode(inode)) {
3347 if (!IS_ERR(inode))
3348 iput(inode);
3349 return -ENOENT;
3350 }
3351
3352 truncate:
3353 ret = btrfs_check_trunc_cache_free_space(root,
3354 &fs_info->global_block_rsv);
3355 if (ret)
3356 goto out;
3357
3358 path = btrfs_alloc_path();
3359 if (!path) {
3360 ret = -ENOMEM;
3361 goto out;
3362 }
3363
3364 trans = btrfs_join_transaction(root);
3365 if (IS_ERR(trans)) {
3366 btrfs_free_path(path);
3367 ret = PTR_ERR(trans);
3368 goto out;
3369 }
3370
3371 ret = btrfs_truncate_free_space_cache(root, trans, path, inode);
3372
3373 btrfs_free_path(path);
3374 btrfs_end_transaction(trans, root);
3375 btrfs_btree_balance_dirty(root);
3376 out:
3377 iput(inode);
3378 return ret;
3379 }
3380
3381 /*
3382 * helper to add tree blocks for backref of type BTRFS_EXTENT_DATA_REF_KEY
3383 * this function scans fs tree to find blocks reference the data extent
3384 */
3385 static int find_data_references(struct reloc_control *rc,
3386 struct btrfs_key *extent_key,
3387 struct extent_buffer *leaf,
3388 struct btrfs_extent_data_ref *ref,
3389 struct rb_root *blocks)
3390 {
3391 struct btrfs_path *path;
3392 struct tree_block *block;
3393 struct btrfs_root *root;
3394 struct btrfs_file_extent_item *fi;
3395 struct rb_node *rb_node;
3396 struct btrfs_key key;
3397 u64 ref_root;
3398 u64 ref_objectid;
3399 u64 ref_offset;
3400 u32 ref_count;
3401 u32 nritems;
3402 int err = 0;
3403 int added = 0;
3404 int counted;
3405 int ret;
3406
3407 ref_root = btrfs_extent_data_ref_root(leaf, ref);
3408 ref_objectid = btrfs_extent_data_ref_objectid(leaf, ref);
3409 ref_offset = btrfs_extent_data_ref_offset(leaf, ref);
3410 ref_count = btrfs_extent_data_ref_count(leaf, ref);
3411
3412 /*
3413 * This is an extent belonging to the free space cache, lets just delete
3414 * it and redo the search.
3415 */
3416 if (ref_root == BTRFS_ROOT_TREE_OBJECTID) {
3417 ret = delete_block_group_cache(rc->extent_root->fs_info,
3418 NULL, ref_objectid);
3419 if (ret != -ENOENT)
3420 return ret;
3421 ret = 0;
3422 }
3423
3424 path = btrfs_alloc_path();
3425 if (!path)
3426 return -ENOMEM;
3427 path->reada = 1;
3428
3429 root = read_fs_root(rc->extent_root->fs_info, ref_root);
3430 if (IS_ERR(root)) {
3431 err = PTR_ERR(root);
3432 goto out;
3433 }
3434
3435 key.objectid = ref_objectid;
3436 key.type = BTRFS_EXTENT_DATA_KEY;
3437 if (ref_offset > ((u64)-1 << 32))
3438 key.offset = 0;
3439 else
3440 key.offset = ref_offset;
3441
3442 path->search_commit_root = 1;
3443 path->skip_locking = 1;
3444 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
3445 if (ret < 0) {
3446 err = ret;
3447 goto out;
3448 }
3449
3450 leaf = path->nodes[0];
3451 nritems = btrfs_header_nritems(leaf);
3452 /*
3453 * the references in tree blocks that use full backrefs
3454 * are not counted in
3455 */
3456 if (block_use_full_backref(rc, leaf))
3457 counted = 0;
3458 else
3459 counted = 1;
3460 rb_node = tree_search(blocks, leaf->start);
3461 if (rb_node) {
3462 if (counted)
3463 added = 1;
3464 else
3465 path->slots[0] = nritems;
3466 }
3467
3468 while (ref_count > 0) {
3469 while (path->slots[0] >= nritems) {
3470 ret = btrfs_next_leaf(root, path);
3471 if (ret < 0) {
3472 err = ret;
3473 goto out;
3474 }
3475 if (ret > 0) {
3476 WARN_ON(1);
3477 goto out;
3478 }
3479
3480 leaf = path->nodes[0];
3481 nritems = btrfs_header_nritems(leaf);
3482 added = 0;
3483
3484 if (block_use_full_backref(rc, leaf))
3485 counted = 0;
3486 else
3487 counted = 1;
3488 rb_node = tree_search(blocks, leaf->start);
3489 if (rb_node) {
3490 if (counted)
3491 added = 1;
3492 else
3493 path->slots[0] = nritems;
3494 }
3495 }
3496
3497 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
3498 if (key.objectid != ref_objectid ||
3499 key.type != BTRFS_EXTENT_DATA_KEY) {
3500 WARN_ON(1);
3501 break;
3502 }
3503
3504 fi = btrfs_item_ptr(leaf, path->slots[0],
3505 struct btrfs_file_extent_item);
3506
3507 if (btrfs_file_extent_type(leaf, fi) ==
3508 BTRFS_FILE_EXTENT_INLINE)
3509 goto next;
3510
3511 if (btrfs_file_extent_disk_bytenr(leaf, fi) !=
3512 extent_key->objectid)
3513 goto next;
3514
3515 key.offset -= btrfs_file_extent_offset(leaf, fi);
3516 if (key.offset != ref_offset)
3517 goto next;
3518
3519 if (counted)
3520 ref_count--;
3521 if (added)
3522 goto next;
3523
3524 if (!tree_block_processed(leaf->start, leaf->len, rc)) {
3525 block = kmalloc(sizeof(*block), GFP_NOFS);
3526 if (!block) {
3527 err = -ENOMEM;
3528 break;
3529 }
3530 block->bytenr = leaf->start;
3531 btrfs_item_key_to_cpu(leaf, &block->key, 0);
3532 block->level = 0;
3533 block->key_ready = 1;
3534 rb_node = tree_insert(blocks, block->bytenr,
3535 &block->rb_node);
3536 if (rb_node)
3537 backref_tree_panic(rb_node, -EEXIST,
3538 block->bytenr);
3539 }
3540 if (counted)
3541 added = 1;
3542 else
3543 path->slots[0] = nritems;
3544 next:
3545 path->slots[0]++;
3546
3547 }
3548 out:
3549 btrfs_free_path(path);
3550 return err;
3551 }
3552
3553 /*
3554 * helper to find all tree blocks that reference a given data extent
3555 */
3556 static noinline_for_stack
3557 int add_data_references(struct reloc_control *rc,
3558 struct btrfs_key *extent_key,
3559 struct btrfs_path *path,
3560 struct rb_root *blocks)
3561 {
3562 struct btrfs_key key;
3563 struct extent_buffer *eb;
3564 struct btrfs_extent_data_ref *dref;
3565 struct btrfs_extent_inline_ref *iref;
3566 unsigned long ptr;
3567 unsigned long end;
3568 u32 blocksize = btrfs_level_size(rc->extent_root, 0);
3569 int ret;
3570 int err = 0;
3571
3572 eb = path->nodes[0];
3573 ptr = btrfs_item_ptr_offset(eb, path->slots[0]);
3574 end = ptr + btrfs_item_size_nr(eb, path->slots[0]);
3575 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
3576 if (ptr + sizeof(struct btrfs_extent_item_v0) == end)
3577 ptr = end;
3578 else
3579 #endif
3580 ptr += sizeof(struct btrfs_extent_item);
3581
3582 while (ptr < end) {
3583 iref = (struct btrfs_extent_inline_ref *)ptr;
3584 key.type = btrfs_extent_inline_ref_type(eb, iref);
3585 if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
3586 key.offset = btrfs_extent_inline_ref_offset(eb, iref);
3587 ret = __add_tree_block(rc, key.offset, blocksize,
3588 blocks);
3589 } else if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
3590 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
3591 ret = find_data_references(rc, extent_key,
3592 eb, dref, blocks);
3593 } else {
3594 BUG();
3595 }
3596 ptr += btrfs_extent_inline_ref_size(key.type);
3597 }
3598 WARN_ON(ptr > end);
3599
3600 while (1) {
3601 cond_resched();
3602 eb = path->nodes[0];
3603 if (path->slots[0] >= btrfs_header_nritems(eb)) {
3604 ret = btrfs_next_leaf(rc->extent_root, path);
3605 if (ret < 0) {
3606 err = ret;
3607 break;
3608 }
3609 if (ret > 0)
3610 break;
3611 eb = path->nodes[0];
3612 }
3613
3614 btrfs_item_key_to_cpu(eb, &key, path->slots[0]);
3615 if (key.objectid != extent_key->objectid)
3616 break;
3617
3618 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
3619 if (key.type == BTRFS_SHARED_DATA_REF_KEY ||
3620 key.type == BTRFS_EXTENT_REF_V0_KEY) {
3621 #else
3622 BUG_ON(key.type == BTRFS_EXTENT_REF_V0_KEY);
3623 if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
3624 #endif
3625 ret = __add_tree_block(rc, key.offset, blocksize,
3626 blocks);
3627 } else if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
3628 dref = btrfs_item_ptr(eb, path->slots[0],
3629 struct btrfs_extent_data_ref);
3630 ret = find_data_references(rc, extent_key,
3631 eb, dref, blocks);
3632 } else {
3633 ret = 0;
3634 }
3635 if (ret) {
3636 err = ret;
3637 break;
3638 }
3639 path->slots[0]++;
3640 }
3641 btrfs_release_path(path);
3642 if (err)
3643 free_block_list(blocks);
3644 return err;
3645 }
3646
3647 /*
3648 * helper to find next unprocessed extent
3649 */
3650 static noinline_for_stack
3651 int find_next_extent(struct btrfs_trans_handle *trans,
3652 struct reloc_control *rc, struct btrfs_path *path,
3653 struct btrfs_key *extent_key)
3654 {
3655 struct btrfs_key key;
3656 struct extent_buffer *leaf;
3657 u64 start, end, last;
3658 int ret;
3659
3660 last = rc->block_group->key.objectid + rc->block_group->key.offset;
3661 while (1) {
3662 cond_resched();
3663 if (rc->search_start >= last) {
3664 ret = 1;
3665 break;
3666 }
3667
3668 key.objectid = rc->search_start;
3669 key.type = BTRFS_EXTENT_ITEM_KEY;
3670 key.offset = 0;
3671
3672 path->search_commit_root = 1;
3673 path->skip_locking = 1;
3674 ret = btrfs_search_slot(NULL, rc->extent_root, &key, path,
3675 0, 0);
3676 if (ret < 0)
3677 break;
3678 next:
3679 leaf = path->nodes[0];
3680 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
3681 ret = btrfs_next_leaf(rc->extent_root, path);
3682 if (ret != 0)
3683 break;
3684 leaf = path->nodes[0];
3685 }
3686
3687 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
3688 if (key.objectid >= last) {
3689 ret = 1;
3690 break;
3691 }
3692
3693 if (key.type != BTRFS_EXTENT_ITEM_KEY &&
3694 key.type != BTRFS_METADATA_ITEM_KEY) {
3695 path->slots[0]++;
3696 goto next;
3697 }
3698
3699 if (key.type == BTRFS_EXTENT_ITEM_KEY &&
3700 key.objectid + key.offset <= rc->search_start) {
3701 path->slots[0]++;
3702 goto next;
3703 }
3704
3705 if (key.type == BTRFS_METADATA_ITEM_KEY &&
3706 key.objectid + rc->extent_root->leafsize <=
3707 rc->search_start) {
3708 path->slots[0]++;
3709 goto next;
3710 }
3711
3712 ret = find_first_extent_bit(&rc->processed_blocks,
3713 key.objectid, &start, &end,
3714 EXTENT_DIRTY, NULL);
3715
3716 if (ret == 0 && start <= key.objectid) {
3717 btrfs_release_path(path);
3718 rc->search_start = end + 1;
3719 } else {
3720 if (key.type == BTRFS_EXTENT_ITEM_KEY)
3721 rc->search_start = key.objectid + key.offset;
3722 else
3723 rc->search_start = key.objectid +
3724 rc->extent_root->leafsize;
3725 memcpy(extent_key, &key, sizeof(key));
3726 return 0;
3727 }
3728 }
3729 btrfs_release_path(path);
3730 return ret;
3731 }
3732
3733 static void set_reloc_control(struct reloc_control *rc)
3734 {
3735 struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
3736
3737 mutex_lock(&fs_info->reloc_mutex);
3738 fs_info->reloc_ctl = rc;
3739 mutex_unlock(&fs_info->reloc_mutex);
3740 }
3741
3742 static void unset_reloc_control(struct reloc_control *rc)
3743 {
3744 struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
3745
3746 mutex_lock(&fs_info->reloc_mutex);
3747 fs_info->reloc_ctl = NULL;
3748 mutex_unlock(&fs_info->reloc_mutex);
3749 }
3750
3751 static int check_extent_flags(u64 flags)
3752 {
3753 if ((flags & BTRFS_EXTENT_FLAG_DATA) &&
3754 (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK))
3755 return 1;
3756 if (!(flags & BTRFS_EXTENT_FLAG_DATA) &&
3757 !(flags & BTRFS_EXTENT_FLAG_TREE_BLOCK))
3758 return 1;
3759 if ((flags & BTRFS_EXTENT_FLAG_DATA) &&
3760 (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
3761 return 1;
3762 return 0;
3763 }
3764
3765 static noinline_for_stack
3766 int prepare_to_relocate(struct reloc_control *rc)
3767 {
3768 struct btrfs_trans_handle *trans;
3769 int ret;
3770
3771 rc->block_rsv = btrfs_alloc_block_rsv(rc->extent_root,
3772 BTRFS_BLOCK_RSV_TEMP);
3773 if (!rc->block_rsv)
3774 return -ENOMEM;
3775
3776 /*
3777 * reserve some space for creating reloc trees.
3778 * btrfs_init_reloc_root will use them when there
3779 * is no reservation in transaction handle.
3780 */
3781 ret = btrfs_block_rsv_add(rc->extent_root, rc->block_rsv,
3782 rc->extent_root->nodesize * 256,
3783 BTRFS_RESERVE_FLUSH_ALL);
3784 if (ret)
3785 return ret;
3786
3787 memset(&rc->cluster, 0, sizeof(rc->cluster));
3788 rc->search_start = rc->block_group->key.objectid;
3789 rc->extents_found = 0;
3790 rc->nodes_relocated = 0;
3791 rc->merging_rsv_size = 0;
3792
3793 rc->create_reloc_tree = 1;
3794 set_reloc_control(rc);
3795
3796 trans = btrfs_join_transaction(rc->extent_root);
3797 if (IS_ERR(trans)) {
3798 unset_reloc_control(rc);
3799 /*
3800 * extent tree is not a ref_cow tree and has no reloc_root to
3801 * cleanup. And callers are responsible to free the above
3802 * block rsv.
3803 */
3804 return PTR_ERR(trans);
3805 }
3806 btrfs_commit_transaction(trans, rc->extent_root);
3807 return 0;
3808 }
3809
3810 static noinline_for_stack int relocate_block_group(struct reloc_control *rc)
3811 {
3812 struct rb_root blocks = RB_ROOT;
3813 struct btrfs_key key;
3814 struct btrfs_trans_handle *trans = NULL;
3815 struct btrfs_path *path;
3816 struct btrfs_extent_item *ei;
3817 u64 flags;
3818 u32 item_size;
3819 int ret;
3820 int err = 0;
3821 int progress = 0;
3822
3823 path = btrfs_alloc_path();
3824 if (!path)
3825 return -ENOMEM;
3826 path->reada = 1;
3827
3828 ret = prepare_to_relocate(rc);
3829 if (ret) {
3830 err = ret;
3831 goto out_free;
3832 }
3833
3834 while (1) {
3835 progress++;
3836 trans = btrfs_start_transaction(rc->extent_root, 0);
3837 if (IS_ERR(trans)) {
3838 err = PTR_ERR(trans);
3839 trans = NULL;
3840 break;
3841 }
3842 restart:
3843 if (update_backref_cache(trans, &rc->backref_cache)) {
3844 btrfs_end_transaction(trans, rc->extent_root);
3845 continue;
3846 }
3847
3848 ret = find_next_extent(trans, rc, path, &key);
3849 if (ret < 0)
3850 err = ret;
3851 if (ret != 0)
3852 break;
3853
3854 rc->extents_found++;
3855
3856 ei = btrfs_item_ptr(path->nodes[0], path->slots[0],
3857 struct btrfs_extent_item);
3858 item_size = btrfs_item_size_nr(path->nodes[0], path->slots[0]);
3859 if (item_size >= sizeof(*ei)) {
3860 flags = btrfs_extent_flags(path->nodes[0], ei);
3861 ret = check_extent_flags(flags);
3862 BUG_ON(ret);
3863
3864 } else {
3865 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
3866 u64 ref_owner;
3867 int path_change = 0;
3868
3869 BUG_ON(item_size !=
3870 sizeof(struct btrfs_extent_item_v0));
3871 ret = get_ref_objectid_v0(rc, path, &key, &ref_owner,
3872 &path_change);
3873 if (ref_owner < BTRFS_FIRST_FREE_OBJECTID)
3874 flags = BTRFS_EXTENT_FLAG_TREE_BLOCK;
3875 else
3876 flags = BTRFS_EXTENT_FLAG_DATA;
3877
3878 if (path_change) {
3879 btrfs_release_path(path);
3880
3881 path->search_commit_root = 1;
3882 path->skip_locking = 1;
3883 ret = btrfs_search_slot(NULL, rc->extent_root,
3884 &key, path, 0, 0);
3885 if (ret < 0) {
3886 err = ret;
3887 break;
3888 }
3889 BUG_ON(ret > 0);
3890 }
3891 #else
3892 BUG();
3893 #endif
3894 }
3895
3896 if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
3897 ret = add_tree_block(rc, &key, path, &blocks);
3898 } else if (rc->stage == UPDATE_DATA_PTRS &&
3899 (flags & BTRFS_EXTENT_FLAG_DATA)) {
3900 ret = add_data_references(rc, &key, path, &blocks);
3901 } else {
3902 btrfs_release_path(path);
3903 ret = 0;
3904 }
3905 if (ret < 0) {
3906 err = ret;
3907 break;
3908 }
3909
3910 if (!RB_EMPTY_ROOT(&blocks)) {
3911 ret = relocate_tree_blocks(trans, rc, &blocks);
3912 if (ret < 0) {
3913 if (ret != -EAGAIN) {
3914 err = ret;
3915 break;
3916 }
3917 rc->extents_found--;
3918 rc->search_start = key.objectid;
3919 }
3920 }
3921
3922 ret = btrfs_block_rsv_check(rc->extent_root, rc->block_rsv, 5);
3923 if (ret < 0) {
3924 if (ret != -ENOSPC) {
3925 err = ret;
3926 WARN_ON(1);
3927 break;
3928 }
3929 rc->commit_transaction = 1;
3930 }
3931
3932 if (rc->commit_transaction) {
3933 rc->commit_transaction = 0;
3934 ret = btrfs_commit_transaction(trans, rc->extent_root);
3935 BUG_ON(ret);
3936 } else {
3937 btrfs_end_transaction_throttle(trans, rc->extent_root);
3938 btrfs_btree_balance_dirty(rc->extent_root);
3939 }
3940 trans = NULL;
3941
3942 if (rc->stage == MOVE_DATA_EXTENTS &&
3943 (flags & BTRFS_EXTENT_FLAG_DATA)) {
3944 rc->found_file_extent = 1;
3945 ret = relocate_data_extent(rc->data_inode,
3946 &key, &rc->cluster);
3947 if (ret < 0) {
3948 err = ret;
3949 break;
3950 }
3951 }
3952 }
3953 if (trans && progress && err == -ENOSPC) {
3954 ret = btrfs_force_chunk_alloc(trans, rc->extent_root,
3955 rc->block_group->flags);
3956 if (ret == 0) {
3957 err = 0;
3958 progress = 0;
3959 goto restart;
3960 }
3961 }
3962
3963 btrfs_release_path(path);
3964 clear_extent_bits(&rc->processed_blocks, 0, (u64)-1, EXTENT_DIRTY,
3965 GFP_NOFS);
3966
3967 if (trans) {
3968 btrfs_end_transaction_throttle(trans, rc->extent_root);
3969 btrfs_btree_balance_dirty(rc->extent_root);
3970 }
3971
3972 if (!err) {
3973 ret = relocate_file_extent_cluster(rc->data_inode,
3974 &rc->cluster);
3975 if (ret < 0)
3976 err = ret;
3977 }
3978
3979 rc->create_reloc_tree = 0;
3980 set_reloc_control(rc);
3981
3982 backref_cache_cleanup(&rc->backref_cache);
3983 btrfs_block_rsv_release(rc->extent_root, rc->block_rsv, (u64)-1);
3984
3985 err = prepare_to_merge(rc, err);
3986
3987 merge_reloc_roots(rc);
3988
3989 rc->merge_reloc_tree = 0;
3990 unset_reloc_control(rc);
3991 btrfs_block_rsv_release(rc->extent_root, rc->block_rsv, (u64)-1);
3992
3993 /* get rid of pinned extents */
3994 trans = btrfs_join_transaction(rc->extent_root);
3995 if (IS_ERR(trans))
3996 err = PTR_ERR(trans);
3997 else
3998 btrfs_commit_transaction(trans, rc->extent_root);
3999 out_free:
4000 btrfs_free_block_rsv(rc->extent_root, rc->block_rsv);
4001 btrfs_free_path(path);
4002 return err;
4003 }
4004
4005 static int __insert_orphan_inode(struct btrfs_trans_handle *trans,
4006 struct btrfs_root *root, u64 objectid)
4007 {
4008 struct btrfs_path *path;
4009 struct btrfs_inode_item *item;
4010 struct extent_buffer *leaf;
4011 int ret;
4012
4013 path = btrfs_alloc_path();
4014 if (!path)
4015 return -ENOMEM;
4016
4017 ret = btrfs_insert_empty_inode(trans, root, path, objectid);
4018 if (ret)
4019 goto out;
4020
4021 leaf = path->nodes[0];
4022 item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_inode_item);
4023 memset_extent_buffer(leaf, 0, (unsigned long)item, sizeof(*item));
4024 btrfs_set_inode_generation(leaf, item, 1);
4025 btrfs_set_inode_size(leaf, item, 0);
4026 btrfs_set_inode_mode(leaf, item, S_IFREG | 0600);
4027 btrfs_set_inode_flags(leaf, item, BTRFS_INODE_NOCOMPRESS |
4028 BTRFS_INODE_PREALLOC);
4029 btrfs_mark_buffer_dirty(leaf);
4030 btrfs_release_path(path);
4031 out:
4032 btrfs_free_path(path);
4033 return ret;
4034 }
4035
4036 /*
4037 * helper to create inode for data relocation.
4038 * the inode is in data relocation tree and its link count is 0
4039 */
4040 static noinline_for_stack
4041 struct inode *create_reloc_inode(struct btrfs_fs_info *fs_info,
4042 struct btrfs_block_group_cache *group)
4043 {
4044 struct inode *inode = NULL;
4045 struct btrfs_trans_handle *trans;
4046 struct btrfs_root *root;
4047 struct btrfs_key key;
4048 u64 objectid = BTRFS_FIRST_FREE_OBJECTID;
4049 int err = 0;
4050
4051 root = read_fs_root(fs_info, BTRFS_DATA_RELOC_TREE_OBJECTID);
4052 if (IS_ERR(root))
4053 return ERR_CAST(root);
4054
4055 trans = btrfs_start_transaction(root, 6);
4056 if (IS_ERR(trans))
4057 return ERR_CAST(trans);
4058
4059 err = btrfs_find_free_objectid(root, &objectid);
4060 if (err)
4061 goto out;
4062
4063 err = __insert_orphan_inode(trans, root, objectid);
4064 BUG_ON(err);
4065
4066 key.objectid = objectid;
4067 key.type = BTRFS_INODE_ITEM_KEY;
4068 key.offset = 0;
4069 inode = btrfs_iget(root->fs_info->sb, &key, root, NULL);
4070 BUG_ON(IS_ERR(inode) || is_bad_inode(inode));
4071 BTRFS_I(inode)->index_cnt = group->key.objectid;
4072
4073 err = btrfs_orphan_add(trans, inode);
4074 out:
4075 btrfs_end_transaction(trans, root);
4076 btrfs_btree_balance_dirty(root);
4077 if (err) {
4078 if (inode)
4079 iput(inode);
4080 inode = ERR_PTR(err);
4081 }
4082 return inode;
4083 }
4084
4085 static struct reloc_control *alloc_reloc_control(struct btrfs_fs_info *fs_info)
4086 {
4087 struct reloc_control *rc;
4088
4089 rc = kzalloc(sizeof(*rc), GFP_NOFS);
4090 if (!rc)
4091 return NULL;
4092
4093 INIT_LIST_HEAD(&rc->reloc_roots);
4094 backref_cache_init(&rc->backref_cache);
4095 mapping_tree_init(&rc->reloc_root_tree);
4096 extent_io_tree_init(&rc->processed_blocks,
4097 fs_info->btree_inode->i_mapping);
4098 return rc;
4099 }
4100
4101 /*
4102 * function to relocate all extents in a block group.
4103 */
4104 int btrfs_relocate_block_group(struct btrfs_root *extent_root, u64 group_start)
4105 {
4106 struct btrfs_fs_info *fs_info = extent_root->fs_info;
4107 struct reloc_control *rc;
4108 struct inode *inode;
4109 struct btrfs_path *path;
4110 int ret;
4111 int rw = 0;
4112 int err = 0;
4113
4114 rc = alloc_reloc_control(fs_info);
4115 if (!rc)
4116 return -ENOMEM;
4117
4118 rc->extent_root = extent_root;
4119
4120 rc->block_group = btrfs_lookup_block_group(fs_info, group_start);
4121 BUG_ON(!rc->block_group);
4122
4123 if (!rc->block_group->ro) {
4124 ret = btrfs_set_block_group_ro(extent_root, rc->block_group);
4125 if (ret) {
4126 err = ret;
4127 goto out;
4128 }
4129 rw = 1;
4130 }
4131
4132 path = btrfs_alloc_path();
4133 if (!path) {
4134 err = -ENOMEM;
4135 goto out;
4136 }
4137
4138 inode = lookup_free_space_inode(fs_info->tree_root, rc->block_group,
4139 path);
4140 btrfs_free_path(path);
4141
4142 if (!IS_ERR(inode))
4143 ret = delete_block_group_cache(fs_info, inode, 0);
4144 else
4145 ret = PTR_ERR(inode);
4146
4147 if (ret && ret != -ENOENT) {
4148 err = ret;
4149 goto out;
4150 }
4151
4152 rc->data_inode = create_reloc_inode(fs_info, rc->block_group);
4153 if (IS_ERR(rc->data_inode)) {
4154 err = PTR_ERR(rc->data_inode);
4155 rc->data_inode = NULL;
4156 goto out;
4157 }
4158
4159 printk(KERN_INFO "btrfs: relocating block group %llu flags %llu\n",
4160 (unsigned long long)rc->block_group->key.objectid,
4161 (unsigned long long)rc->block_group->flags);
4162
4163 ret = btrfs_start_delalloc_inodes(fs_info->tree_root, 0);
4164 if (ret < 0) {
4165 err = ret;
4166 goto out;
4167 }
4168 btrfs_wait_ordered_extents(fs_info->tree_root, 0);
4169
4170 while (1) {
4171 mutex_lock(&fs_info->cleaner_mutex);
4172 ret = relocate_block_group(rc);
4173 mutex_unlock(&fs_info->cleaner_mutex);
4174 if (ret < 0) {
4175 err = ret;
4176 goto out;
4177 }
4178
4179 if (rc->extents_found == 0)
4180 break;
4181
4182 printk(KERN_INFO "btrfs: found %llu extents\n",
4183 (unsigned long long)rc->extents_found);
4184
4185 if (rc->stage == MOVE_DATA_EXTENTS && rc->found_file_extent) {
4186 btrfs_wait_ordered_range(rc->data_inode, 0, (u64)-1);
4187 invalidate_mapping_pages(rc->data_inode->i_mapping,
4188 0, -1);
4189 rc->stage = UPDATE_DATA_PTRS;
4190 }
4191 }
4192
4193 filemap_write_and_wait_range(fs_info->btree_inode->i_mapping,
4194 rc->block_group->key.objectid,
4195 rc->block_group->key.objectid +
4196 rc->block_group->key.offset - 1);
4197
4198 WARN_ON(rc->block_group->pinned > 0);
4199 WARN_ON(rc->block_group->reserved > 0);
4200 WARN_ON(btrfs_block_group_used(&rc->block_group->item) > 0);
4201 out:
4202 if (err && rw)
4203 btrfs_set_block_group_rw(extent_root, rc->block_group);
4204 iput(rc->data_inode);
4205 btrfs_put_block_group(rc->block_group);
4206 kfree(rc);
4207 return err;
4208 }
4209
4210 static noinline_for_stack int mark_garbage_root(struct btrfs_root *root)
4211 {
4212 struct btrfs_trans_handle *trans;
4213 int ret, err;
4214
4215 trans = btrfs_start_transaction(root->fs_info->tree_root, 0);
4216 if (IS_ERR(trans))
4217 return PTR_ERR(trans);
4218
4219 memset(&root->root_item.drop_progress, 0,
4220 sizeof(root->root_item.drop_progress));
4221 root->root_item.drop_level = 0;
4222 btrfs_set_root_refs(&root->root_item, 0);
4223 ret = btrfs_update_root(trans, root->fs_info->tree_root,
4224 &root->root_key, &root->root_item);
4225
4226 err = btrfs_end_transaction(trans, root->fs_info->tree_root);
4227 if (err)
4228 return err;
4229 return ret;
4230 }
4231
4232 /*
4233 * recover relocation interrupted by system crash.
4234 *
4235 * this function resumes merging reloc trees with corresponding fs trees.
4236 * this is important for keeping the sharing of tree blocks
4237 */
4238 int btrfs_recover_relocation(struct btrfs_root *root)
4239 {
4240 LIST_HEAD(reloc_roots);
4241 struct btrfs_key key;
4242 struct btrfs_root *fs_root;
4243 struct btrfs_root *reloc_root;
4244 struct btrfs_path *path;
4245 struct extent_buffer *leaf;
4246 struct reloc_control *rc = NULL;
4247 struct btrfs_trans_handle *trans;
4248 int ret;
4249 int err = 0;
4250
4251 path = btrfs_alloc_path();
4252 if (!path)
4253 return -ENOMEM;
4254 path->reada = -1;
4255
4256 key.objectid = BTRFS_TREE_RELOC_OBJECTID;
4257 key.type = BTRFS_ROOT_ITEM_KEY;
4258 key.offset = (u64)-1;
4259
4260 while (1) {
4261 ret = btrfs_search_slot(NULL, root->fs_info->tree_root, &key,
4262 path, 0, 0);
4263 if (ret < 0) {
4264 err = ret;
4265 goto out;
4266 }
4267 if (ret > 0) {
4268 if (path->slots[0] == 0)
4269 break;
4270 path->slots[0]--;
4271 }
4272 leaf = path->nodes[0];
4273 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
4274 btrfs_release_path(path);
4275
4276 if (key.objectid != BTRFS_TREE_RELOC_OBJECTID ||
4277 key.type != BTRFS_ROOT_ITEM_KEY)
4278 break;
4279
4280 reloc_root = btrfs_read_fs_root_no_radix(root, &key);
4281 if (IS_ERR(reloc_root)) {
4282 err = PTR_ERR(reloc_root);
4283 goto out;
4284 }
4285
4286 list_add(&reloc_root->root_list, &reloc_roots);
4287
4288 if (btrfs_root_refs(&reloc_root->root_item) > 0) {
4289 fs_root = read_fs_root(root->fs_info,
4290 reloc_root->root_key.offset);
4291 if (IS_ERR(fs_root)) {
4292 ret = PTR_ERR(fs_root);
4293 if (ret != -ENOENT) {
4294 err = ret;
4295 goto out;
4296 }
4297 ret = mark_garbage_root(reloc_root);
4298 if (ret < 0) {
4299 err = ret;
4300 goto out;
4301 }
4302 }
4303 }
4304
4305 if (key.offset == 0)
4306 break;
4307
4308 key.offset--;
4309 }
4310 btrfs_release_path(path);
4311
4312 if (list_empty(&reloc_roots))
4313 goto out;
4314
4315 rc = alloc_reloc_control(root->fs_info);
4316 if (!rc) {
4317 err = -ENOMEM;
4318 goto out;
4319 }
4320
4321 rc->extent_root = root->fs_info->extent_root;
4322
4323 set_reloc_control(rc);
4324
4325 trans = btrfs_join_transaction(rc->extent_root);
4326 if (IS_ERR(trans)) {
4327 unset_reloc_control(rc);
4328 err = PTR_ERR(trans);
4329 goto out_free;
4330 }
4331
4332 rc->merge_reloc_tree = 1;
4333
4334 while (!list_empty(&reloc_roots)) {
4335 reloc_root = list_entry(reloc_roots.next,
4336 struct btrfs_root, root_list);
4337 list_del(&reloc_root->root_list);
4338
4339 if (btrfs_root_refs(&reloc_root->root_item) == 0) {
4340 list_add_tail(&reloc_root->root_list,
4341 &rc->reloc_roots);
4342 continue;
4343 }
4344
4345 fs_root = read_fs_root(root->fs_info,
4346 reloc_root->root_key.offset);
4347 if (IS_ERR(fs_root)) {
4348 err = PTR_ERR(fs_root);
4349 goto out_free;
4350 }
4351
4352 err = __add_reloc_root(reloc_root);
4353 BUG_ON(err < 0); /* -ENOMEM or logic error */
4354 fs_root->reloc_root = reloc_root;
4355 }
4356
4357 err = btrfs_commit_transaction(trans, rc->extent_root);
4358 if (err)
4359 goto out_free;
4360
4361 merge_reloc_roots(rc);
4362
4363 unset_reloc_control(rc);
4364
4365 trans = btrfs_join_transaction(rc->extent_root);
4366 if (IS_ERR(trans))
4367 err = PTR_ERR(trans);
4368 else
4369 err = btrfs_commit_transaction(trans, rc->extent_root);
4370 out_free:
4371 kfree(rc);
4372 out:
4373 if (!list_empty(&reloc_roots))
4374 free_reloc_roots(&reloc_roots);
4375
4376 btrfs_free_path(path);
4377
4378 if (err == 0) {
4379 /* cleanup orphan inode in data relocation tree */
4380 fs_root = read_fs_root(root->fs_info,
4381 BTRFS_DATA_RELOC_TREE_OBJECTID);
4382 if (IS_ERR(fs_root))
4383 err = PTR_ERR(fs_root);
4384 else
4385 err = btrfs_orphan_cleanup(fs_root);
4386 }
4387 return err;
4388 }
4389
4390 /*
4391 * helper to add ordered checksum for data relocation.
4392 *
4393 * cloning checksum properly handles the nodatasum extents.
4394 * it also saves CPU time to re-calculate the checksum.
4395 */
4396 int btrfs_reloc_clone_csums(struct inode *inode, u64 file_pos, u64 len)
4397 {
4398 struct btrfs_ordered_sum *sums;
4399 struct btrfs_sector_sum *sector_sum;
4400 struct btrfs_ordered_extent *ordered;
4401 struct btrfs_root *root = BTRFS_I(inode)->root;
4402 size_t offset;
4403 int ret;
4404 u64 disk_bytenr;
4405 LIST_HEAD(list);
4406
4407 ordered = btrfs_lookup_ordered_extent(inode, file_pos);
4408 BUG_ON(ordered->file_offset != file_pos || ordered->len != len);
4409
4410 disk_bytenr = file_pos + BTRFS_I(inode)->index_cnt;
4411 ret = btrfs_lookup_csums_range(root->fs_info->csum_root, disk_bytenr,
4412 disk_bytenr + len - 1, &list, 0);
4413 if (ret)
4414 goto out;
4415
4416 while (!list_empty(&list)) {
4417 sums = list_entry(list.next, struct btrfs_ordered_sum, list);
4418 list_del_init(&sums->list);
4419
4420 sector_sum = sums->sums;
4421 sums->bytenr = ordered->start;
4422
4423 offset = 0;
4424 while (offset < sums->len) {
4425 sector_sum->bytenr += ordered->start - disk_bytenr;
4426 sector_sum++;
4427 offset += root->sectorsize;
4428 }
4429
4430 btrfs_add_ordered_sum(inode, ordered, sums);
4431 }
4432 out:
4433 btrfs_put_ordered_extent(ordered);
4434 return ret;
4435 }
4436
4437 void btrfs_reloc_cow_block(struct btrfs_trans_handle *trans,
4438 struct btrfs_root *root, struct extent_buffer *buf,
4439 struct extent_buffer *cow)
4440 {
4441 struct reloc_control *rc;
4442 struct backref_node *node;
4443 int first_cow = 0;
4444 int level;
4445 int ret;
4446
4447 rc = root->fs_info->reloc_ctl;
4448 if (!rc)
4449 return;
4450
4451 BUG_ON(rc->stage == UPDATE_DATA_PTRS &&
4452 root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID);
4453
4454 level = btrfs_header_level(buf);
4455 if (btrfs_header_generation(buf) <=
4456 btrfs_root_last_snapshot(&root->root_item))
4457 first_cow = 1;
4458
4459 if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID &&
4460 rc->create_reloc_tree) {
4461 WARN_ON(!first_cow && level == 0);
4462
4463 node = rc->backref_cache.path[level];
4464 BUG_ON(node->bytenr != buf->start &&
4465 node->new_bytenr != buf->start);
4466
4467 drop_node_buffer(node);
4468 extent_buffer_get(cow);
4469 node->eb = cow;
4470 node->new_bytenr = cow->start;
4471
4472 if (!node->pending) {
4473 list_move_tail(&node->list,
4474 &rc->backref_cache.pending[level]);
4475 node->pending = 1;
4476 }
4477
4478 if (first_cow)
4479 __mark_block_processed(rc, node);
4480
4481 if (first_cow && level > 0)
4482 rc->nodes_relocated += buf->len;
4483 }
4484
4485 if (level == 0 && first_cow && rc->stage == UPDATE_DATA_PTRS) {
4486 ret = replace_file_extents(trans, rc, root, cow);
4487 BUG_ON(ret);
4488 }
4489 }
4490
4491 /*
4492 * called before creating snapshot. it calculates metadata reservation
4493 * requried for relocating tree blocks in the snapshot
4494 */
4495 void btrfs_reloc_pre_snapshot(struct btrfs_trans_handle *trans,
4496 struct btrfs_pending_snapshot *pending,
4497 u64 *bytes_to_reserve)
4498 {
4499 struct btrfs_root *root;
4500 struct reloc_control *rc;
4501
4502 root = pending->root;
4503 if (!root->reloc_root)
4504 return;
4505
4506 rc = root->fs_info->reloc_ctl;
4507 if (!rc->merge_reloc_tree)
4508 return;
4509
4510 root = root->reloc_root;
4511 BUG_ON(btrfs_root_refs(&root->root_item) == 0);
4512 /*
4513 * relocation is in the stage of merging trees. the space
4514 * used by merging a reloc tree is twice the size of
4515 * relocated tree nodes in the worst case. half for cowing
4516 * the reloc tree, half for cowing the fs tree. the space
4517 * used by cowing the reloc tree will be freed after the
4518 * tree is dropped. if we create snapshot, cowing the fs
4519 * tree may use more space than it frees. so we need
4520 * reserve extra space.
4521 */
4522 *bytes_to_reserve += rc->nodes_relocated;
4523 }
4524
4525 /*
4526 * called after snapshot is created. migrate block reservation
4527 * and create reloc root for the newly created snapshot
4528 */
4529 int btrfs_reloc_post_snapshot(struct btrfs_trans_handle *trans,
4530 struct btrfs_pending_snapshot *pending)
4531 {
4532 struct btrfs_root *root = pending->root;
4533 struct btrfs_root *reloc_root;
4534 struct btrfs_root *new_root;
4535 struct reloc_control *rc;
4536 int ret;
4537
4538 if (!root->reloc_root)
4539 return 0;
4540
4541 rc = root->fs_info->reloc_ctl;
4542 rc->merging_rsv_size += rc->nodes_relocated;
4543
4544 if (rc->merge_reloc_tree) {
4545 ret = btrfs_block_rsv_migrate(&pending->block_rsv,
4546 rc->block_rsv,
4547 rc->nodes_relocated);
4548 if (ret)
4549 return ret;
4550 }
4551
4552 new_root = pending->snap;
4553 reloc_root = create_reloc_root(trans, root->reloc_root,
4554 new_root->root_key.objectid);
4555 if (IS_ERR(reloc_root))
4556 return PTR_ERR(reloc_root);
4557
4558 ret = __add_reloc_root(reloc_root);
4559 BUG_ON(ret < 0);
4560 new_root->reloc_root = reloc_root;
4561
4562 if (rc->create_reloc_tree)
4563 ret = clone_backref_node(trans, rc, root, reloc_root);
4564 return ret;
4565 }
4566