?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
Merge branch 'nfs-for-2.6.37' of git://git.linux-nfs.org/projects/trondmy/nfs-2.6
[linux-2.6.git] / fs / btrfs / relocation.c
blobb37d723b9d4a8a137b05256a85788074b5f03c5e
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
33 /*
34 * backref_node, mapping_node and tree_block start with this
35 */
36 struct tree_entry {
37 struct rb_node rb_node;
38 u64 bytenr;
39 };
40
41 /*
42 * present a tree block in the backref cache
43 */
44 struct backref_node {
45 struct rb_node rb_node;
46 u64 bytenr;
47
48 u64 new_bytenr;
49 /* objectid of tree block owner, can be not uptodate */
50 u64 owner;
51 /* link to pending, changed or detached list */
52 struct list_head list;
53 /* list of upper level blocks reference this block */
54 struct list_head upper;
55 /* list of child blocks in the cache */
56 struct list_head lower;
57 /* NULL if this node is not tree root */
58 struct btrfs_root *root;
59 /* extent buffer got by COW the block */
60 struct extent_buffer *eb;
61 /* level of tree block */
62 unsigned int level:8;
63 /* is the block in non-reference counted tree */
64 unsigned int cowonly:1;
65 /* 1 if no child node in the cache */
66 unsigned int lowest:1;
67 /* is the extent buffer locked */
68 unsigned int locked:1;
69 /* has the block been processed */
70 unsigned int processed:1;
71 /* have backrefs of this block been checked */
72 unsigned int checked:1;
73 /*
74 * 1 if corresponding block has been cowed but some upper
75 * level block pointers may not point to the new location
76 */
77 unsigned int pending:1;
78 /*
79 * 1 if the backref node isn't connected to any other
80 * backref node.
81 */
82 unsigned int detached:1;
83 };
84
85 /*
86 * present a block pointer in the backref cache
87 */
88 struct backref_edge {
89 struct list_head list[2];
90 struct backref_node *node[2];
91 };
92
93 #define LOWER 0
94 #define UPPER 1
95
96 struct backref_cache {
97 /* red black tree of all backref nodes in the cache */
98 struct rb_root rb_root;
99 /* for passing backref nodes to btrfs_reloc_cow_block */
100 struct backref_node *path[BTRFS_MAX_LEVEL];
101 /*
102 * list of blocks that have been cowed but some block
103 * pointers in upper level blocks may not reflect the
104 * new location
105 */
106 struct list_head pending[BTRFS_MAX_LEVEL];
107 /* list of backref nodes with no child node */
108 struct list_head leaves;
109 /* list of blocks that have been cowed in current transaction */
110 struct list_head changed;
111 /* list of detached backref node. */
112 struct list_head detached;
113
114 u64 last_trans;
115
116 int nr_nodes;
117 int nr_edges;
118 };
119
120 /*
121 * map address of tree root to tree
122 */
123 struct mapping_node {
124 struct rb_node rb_node;
125 u64 bytenr;
126 void *data;
127 };
128
129 struct mapping_tree {
130 struct rb_root rb_root;
131 spinlock_t lock;
132 };
133
134 /*
135 * present a tree block to process
136 */
137 struct tree_block {
138 struct rb_node rb_node;
139 u64 bytenr;
140 struct btrfs_key key;
141 unsigned int level:8;
142 unsigned int key_ready:1;
143 };
144
145 #define MAX_EXTENTS 128
146
147 struct file_extent_cluster {
148 u64 start;
149 u64 end;
150 u64 boundary[MAX_EXTENTS];
151 unsigned int nr;
152 };
153
154 struct reloc_control {
155 /* block group to relocate */
156 struct btrfs_block_group_cache *block_group;
157 /* extent tree */
158 struct btrfs_root *extent_root;
159 /* inode for moving data */
160 struct inode *data_inode;
161
162 struct btrfs_block_rsv *block_rsv;
163
164 struct backref_cache backref_cache;
165
166 struct file_extent_cluster cluster;
167 /* tree blocks have been processed */
168 struct extent_io_tree processed_blocks;
169 /* map start of tree root to corresponding reloc tree */
170 struct mapping_tree reloc_root_tree;
171 /* list of reloc trees */
172 struct list_head reloc_roots;
173 /* size of metadata reservation for merging reloc trees */
174 u64 merging_rsv_size;
175 /* size of relocated tree nodes */
176 u64 nodes_relocated;
177
178 u64 search_start;
179 u64 extents_found;
180
181 int block_rsv_retries;
182
183 unsigned int stage:8;
184 unsigned int create_reloc_tree:1;
185 unsigned int merge_reloc_tree:1;
186 unsigned int found_file_extent:1;
187 unsigned int commit_transaction:1;
188 };
189
190 /* stages of data relocation */
191 #define MOVE_DATA_EXTENTS 0
192 #define UPDATE_DATA_PTRS 1
193
194 static void remove_backref_node(struct backref_cache *cache,
195 struct backref_node *node);
196 static void __mark_block_processed(struct reloc_control *rc,
197 struct backref_node *node);
198
199 static void mapping_tree_init(struct mapping_tree *tree)
200 {
201 tree->rb_root = RB_ROOT;
202 spin_lock_init(&tree->lock);
203 }
204
205 static void backref_cache_init(struct backref_cache *cache)
206 {
207 int i;
208 cache->rb_root = RB_ROOT;
209 for (i = 0; i < BTRFS_MAX_LEVEL; i++)
210 INIT_LIST_HEAD(&cache->pending[i]);
211 INIT_LIST_HEAD(&cache->changed);
212 INIT_LIST_HEAD(&cache->detached);
213 INIT_LIST_HEAD(&cache->leaves);
214 }
215
216 static void backref_cache_cleanup(struct backref_cache *cache)
217 {
218 struct backref_node *node;
219 int i;
220
221 while (!list_empty(&cache->detached)) {
222 node = list_entry(cache->detached.next,
223 struct backref_node, list);
224 remove_backref_node(cache, node);
225 }
226
227 while (!list_empty(&cache->leaves)) {
228 node = list_entry(cache->leaves.next,
229 struct backref_node, lower);
230 remove_backref_node(cache, node);
231 }
232
233 cache->last_trans = 0;
234
235 for (i = 0; i < BTRFS_MAX_LEVEL; i++)
236 BUG_ON(!list_empty(&cache->pending[i]));
237 BUG_ON(!list_empty(&cache->changed));
238 BUG_ON(!list_empty(&cache->detached));
239 BUG_ON(!RB_EMPTY_ROOT(&cache->rb_root));
240 BUG_ON(cache->nr_nodes);
241 BUG_ON(cache->nr_edges);
242 }
243
244 static struct backref_node *alloc_backref_node(struct backref_cache *cache)
245 {
246 struct backref_node *node;
247
248 node = kzalloc(sizeof(*node), GFP_NOFS);
249 if (node) {
250 INIT_LIST_HEAD(&node->list);
251 INIT_LIST_HEAD(&node->upper);
252 INIT_LIST_HEAD(&node->lower);
253 RB_CLEAR_NODE(&node->rb_node);
254 cache->nr_nodes++;
255 }
256 return node;
257 }
258
259 static void free_backref_node(struct backref_cache *cache,
260 struct backref_node *node)
261 {
262 if (node) {
263 cache->nr_nodes--;
264 kfree(node);
265 }
266 }
267
268 static struct backref_edge *alloc_backref_edge(struct backref_cache *cache)
269 {
270 struct backref_edge *edge;
271
272 edge = kzalloc(sizeof(*edge), GFP_NOFS);
273 if (edge)
274 cache->nr_edges++;
275 return edge;
276 }
277
278 static void free_backref_edge(struct backref_cache *cache,
279 struct backref_edge *edge)
280 {
281 if (edge) {
282 cache->nr_edges--;
283 kfree(edge);
284 }
285 }
286
287 static struct rb_node *tree_insert(struct rb_root *root, u64 bytenr,
288 struct rb_node *node)
289 {
290 struct rb_node **p = &root->rb_node;
291 struct rb_node *parent = NULL;
292 struct tree_entry *entry;
293
294 while (*p) {
295 parent = *p;
296 entry = rb_entry(parent, struct tree_entry, rb_node);
297
298 if (bytenr < entry->bytenr)
299 p = &(*p)->rb_left;
300 else if (bytenr > entry->bytenr)
301 p = &(*p)->rb_right;
302 else
303 return parent;
304 }
305
306 rb_link_node(node, parent, p);
307 rb_insert_color(node, root);
308 return NULL;
309 }
310
311 static struct rb_node *tree_search(struct rb_root *root, u64 bytenr)
312 {
313 struct rb_node *n = root->rb_node;
314 struct tree_entry *entry;
315
316 while (n) {
317 entry = rb_entry(n, struct tree_entry, rb_node);
318
319 if (bytenr < entry->bytenr)
320 n = n->rb_left;
321 else if (bytenr > entry->bytenr)
322 n = n->rb_right;
323 else
324 return n;
325 }
326 return NULL;
327 }
328
329 /*
330 * walk up backref nodes until reach node presents tree root
331 */
332 static struct backref_node *walk_up_backref(struct backref_node *node,
333 struct backref_edge *edges[],
334 int *index)
335 {
336 struct backref_edge *edge;
337 int idx = *index;
338
339 while (!list_empty(&node->upper)) {
340 edge = list_entry(node->upper.next,
341 struct backref_edge, list[LOWER]);
342 edges[idx++] = edge;
343 node = edge->node[UPPER];
344 }
345 BUG_ON(node->detached);
346 *index = idx;
347 return node;
348 }
349
350 /*
351 * walk down backref nodes to find start of next reference path
352 */
353 static struct backref_node *walk_down_backref(struct backref_edge *edges[],
354 int *index)
355 {
356 struct backref_edge *edge;
357 struct backref_node *lower;
358 int idx = *index;
359
360 while (idx > 0) {
361 edge = edges[idx - 1];
362 lower = edge->node[LOWER];
363 if (list_is_last(&edge->list[LOWER], &lower->upper)) {
364 idx--;
365 continue;
366 }
367 edge = list_entry(edge->list[LOWER].next,
368 struct backref_edge, list[LOWER]);
369 edges[idx - 1] = edge;
370 *index = idx;
371 return edge->node[UPPER];
372 }
373 *index = 0;
374 return NULL;
375 }
376
377 static void unlock_node_buffer(struct backref_node *node)
378 {
379 if (node->locked) {
380 btrfs_tree_unlock(node->eb);
381 node->locked = 0;
382 }
383 }
384
385 static void drop_node_buffer(struct backref_node *node)
386 {
387 if (node->eb) {
388 unlock_node_buffer(node);
389 free_extent_buffer(node->eb);
390 node->eb = NULL;
391 }
392 }
393
394 static void drop_backref_node(struct backref_cache *tree,
395 struct backref_node *node)
396 {
397 BUG_ON(!list_empty(&node->upper));
398
399 drop_node_buffer(node);
400 list_del(&node->list);
401 list_del(&node->lower);
402 if (!RB_EMPTY_NODE(&node->rb_node))
403 rb_erase(&node->rb_node, &tree->rb_root);
404 free_backref_node(tree, node);
405 }
406
407 /*
408 * remove a backref node from the backref cache
409 */
410 static void remove_backref_node(struct backref_cache *cache,
411 struct backref_node *node)
412 {
413 struct backref_node *upper;
414 struct backref_edge *edge;
415
416 if (!node)
417 return;
418
419 BUG_ON(!node->lowest && !node->detached);
420 while (!list_empty(&node->upper)) {
421 edge = list_entry(node->upper.next, struct backref_edge,
422 list[LOWER]);
423 upper = edge->node[UPPER];
424 list_del(&edge->list[LOWER]);
425 list_del(&edge->list[UPPER]);
426 free_backref_edge(cache, edge);
427
428 if (RB_EMPTY_NODE(&upper->rb_node)) {
429 BUG_ON(!list_empty(&node->upper));
430 drop_backref_node(cache, node);
431 node = upper;
432 node->lowest = 1;
433 continue;
434 }
435 /*
436 * add the node to leaf node list if no other
437 * child block cached.
438 */
439 if (list_empty(&upper->lower)) {
440 list_add_tail(&upper->lower, &cache->leaves);
441 upper->lowest = 1;
442 }
443 }
444
445 drop_backref_node(cache, node);
446 }
447
448 static void update_backref_node(struct backref_cache *cache,
449 struct backref_node *node, u64 bytenr)
450 {
451 struct rb_node *rb_node;
452 rb_erase(&node->rb_node, &cache->rb_root);
453 node->bytenr = bytenr;
454 rb_node = tree_insert(&cache->rb_root, node->bytenr, &node->rb_node);
455 BUG_ON(rb_node);
456 }
457
458 /*
459 * update backref cache after a transaction commit
460 */
461 static int update_backref_cache(struct btrfs_trans_handle *trans,
462 struct backref_cache *cache)
463 {
464 struct backref_node *node;
465 int level = 0;
466
467 if (cache->last_trans == 0) {
468 cache->last_trans = trans->transid;
469 return 0;
470 }
471
472 if (cache->last_trans == trans->transid)
473 return 0;
474
475 /*
476 * detached nodes are used to avoid unnecessary backref
477 * lookup. transaction commit changes the extent tree.
478 * so the detached nodes are no longer useful.
479 */
480 while (!list_empty(&cache->detached)) {
481 node = list_entry(cache->detached.next,
482 struct backref_node, list);
483 remove_backref_node(cache, node);
484 }
485
486 while (!list_empty(&cache->changed)) {
487 node = list_entry(cache->changed.next,
488 struct backref_node, list);
489 list_del_init(&node->list);
490 BUG_ON(node->pending);
491 update_backref_node(cache, node, node->new_bytenr);
492 }
493
494 /*
495 * some nodes can be left in the pending list if there were
496 * errors during processing the pending nodes.
497 */
498 for (level = 0; level < BTRFS_MAX_LEVEL; level++) {
499 list_for_each_entry(node, &cache->pending[level], list) {
500 BUG_ON(!node->pending);
501 if (node->bytenr == node->new_bytenr)
502 continue;
503 update_backref_node(cache, node, node->new_bytenr);
504 }
505 }
506
507 cache->last_trans = 0;
508 return 1;
509 }
510
511 static int should_ignore_root(struct btrfs_root *root)
512 {
513 struct btrfs_root *reloc_root;
514
515 if (!root->ref_cows)
516 return 0;
517
518 reloc_root = root->reloc_root;
519 if (!reloc_root)
520 return 0;
521
522 if (btrfs_root_last_snapshot(&reloc_root->root_item) ==
523 root->fs_info->running_transaction->transid - 1)
524 return 0;
525 /*
526 * if there is reloc tree and it was created in previous
527 * transaction backref lookup can find the reloc tree,
528 * so backref node for the fs tree root is useless for
529 * relocation.
530 */
531 return 1;
532 }
533
534 /*
535 * find reloc tree by address of tree root
536 */
537 static struct btrfs_root *find_reloc_root(struct reloc_control *rc,
538 u64 bytenr)
539 {
540 struct rb_node *rb_node;
541 struct mapping_node *node;
542 struct btrfs_root *root = NULL;
543
544 spin_lock(&rc->reloc_root_tree.lock);
545 rb_node = tree_search(&rc->reloc_root_tree.rb_root, bytenr);
546 if (rb_node) {
547 node = rb_entry(rb_node, struct mapping_node, rb_node);
548 root = (struct btrfs_root *)node->data;
549 }
550 spin_unlock(&rc->reloc_root_tree.lock);
551 return root;
552 }
553
554 static int is_cowonly_root(u64 root_objectid)
555 {
556 if (root_objectid == BTRFS_ROOT_TREE_OBJECTID ||
557 root_objectid == BTRFS_EXTENT_TREE_OBJECTID ||
558 root_objectid == BTRFS_CHUNK_TREE_OBJECTID ||
559 root_objectid == BTRFS_DEV_TREE_OBJECTID ||
560 root_objectid == BTRFS_TREE_LOG_OBJECTID ||
561 root_objectid == BTRFS_CSUM_TREE_OBJECTID)
562 return 1;
563 return 0;
564 }
565
566 static struct btrfs_root *read_fs_root(struct btrfs_fs_info *fs_info,
567 u64 root_objectid)
568 {
569 struct btrfs_key key;
570
571 key.objectid = root_objectid;
572 key.type = BTRFS_ROOT_ITEM_KEY;
573 if (is_cowonly_root(root_objectid))
574 key.offset = 0;
575 else
576 key.offset = (u64)-1;
577
578 return btrfs_read_fs_root_no_name(fs_info, &key);
579 }
580
581 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
582 static noinline_for_stack
583 struct btrfs_root *find_tree_root(struct reloc_control *rc,
584 struct extent_buffer *leaf,
585 struct btrfs_extent_ref_v0 *ref0)
586 {
587 struct btrfs_root *root;
588 u64 root_objectid = btrfs_ref_root_v0(leaf, ref0);
589 u64 generation = btrfs_ref_generation_v0(leaf, ref0);
590
591 BUG_ON(root_objectid == BTRFS_TREE_RELOC_OBJECTID);
592
593 root = read_fs_root(rc->extent_root->fs_info, root_objectid);
594 BUG_ON(IS_ERR(root));
595
596 if (root->ref_cows &&
597 generation != btrfs_root_generation(&root->root_item))
598 return NULL;
599
600 return root;
601 }
602 #endif
603
604 static noinline_for_stack
605 int find_inline_backref(struct extent_buffer *leaf, int slot,
606 unsigned long *ptr, unsigned long *end)
607 {
608 struct btrfs_extent_item *ei;
609 struct btrfs_tree_block_info *bi;
610 u32 item_size;
611
612 item_size = btrfs_item_size_nr(leaf, slot);
613 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
614 if (item_size < sizeof(*ei)) {
615 WARN_ON(item_size != sizeof(struct btrfs_extent_item_v0));
616 return 1;
617 }
618 #endif
619 ei = btrfs_item_ptr(leaf, slot, struct btrfs_extent_item);
620 WARN_ON(!(btrfs_extent_flags(leaf, ei) &
621 BTRFS_EXTENT_FLAG_TREE_BLOCK));
622
623 if (item_size <= sizeof(*ei) + sizeof(*bi)) {
624 WARN_ON(item_size < sizeof(*ei) + sizeof(*bi));
625 return 1;
626 }
627
628 bi = (struct btrfs_tree_block_info *)(ei + 1);
629 *ptr = (unsigned long)(bi + 1);
630 *end = (unsigned long)ei + item_size;
631 return 0;
632 }
633
634 /*
635 * build backref tree for a given tree block. root of the backref tree
636 * corresponds the tree block, leaves of the backref tree correspond
637 * roots of b-trees that reference the tree block.
638 *
639 * the basic idea of this function is check backrefs of a given block
640 * to find upper level blocks that refernece the block, and then check
641 * bakcrefs of these upper level blocks recursively. the recursion stop
642 * when tree root is reached or backrefs for the block is cached.
643 *
644 * NOTE: if we find backrefs for a block are cached, we know backrefs
645 * for all upper level blocks that directly/indirectly reference the
646 * block are also cached.
647 */
648 static noinline_for_stack
649 struct backref_node *build_backref_tree(struct reloc_control *rc,
650 struct btrfs_key *node_key,
651 int level, u64 bytenr)
652 {
653 struct backref_cache *cache = &rc->backref_cache;
654 struct btrfs_path *path1;
655 struct btrfs_path *path2;
656 struct extent_buffer *eb;
657 struct btrfs_root *root;
658 struct backref_node *cur;
659 struct backref_node *upper;
660 struct backref_node *lower;
661 struct backref_node *node = NULL;
662 struct backref_node *exist = NULL;
663 struct backref_edge *edge;
664 struct rb_node *rb_node;
665 struct btrfs_key key;
666 unsigned long end;
667 unsigned long ptr;
668 LIST_HEAD(list);
669 LIST_HEAD(useless);
670 int cowonly;
671 int ret;
672 int err = 0;
673
674 path1 = btrfs_alloc_path();
675 path2 = btrfs_alloc_path();
676 if (!path1 || !path2) {
677 err = -ENOMEM;
678 goto out;
679 }
680
681 node = alloc_backref_node(cache);
682 if (!node) {
683 err = -ENOMEM;
684 goto out;
685 }
686
687 node->bytenr = bytenr;
688 node->level = level;
689 node->lowest = 1;
690 cur = node;
691 again:
692 end = 0;
693 ptr = 0;
694 key.objectid = cur->bytenr;
695 key.type = BTRFS_EXTENT_ITEM_KEY;
696 key.offset = (u64)-1;
697
698 path1->search_commit_root = 1;
699 path1->skip_locking = 1;
700 ret = btrfs_search_slot(NULL, rc->extent_root, &key, path1,
701 0, 0);
702 if (ret < 0) {
703 err = ret;
704 goto out;
705 }
706 BUG_ON(!ret || !path1->slots[0]);
707
708 path1->slots[0]--;
709
710 WARN_ON(cur->checked);
711 if (!list_empty(&cur->upper)) {
712 /*
713 * the backref was added previously when processsing
714 * backref of type BTRFS_TREE_BLOCK_REF_KEY
715 */
716 BUG_ON(!list_is_singular(&cur->upper));
717 edge = list_entry(cur->upper.next, struct backref_edge,
718 list[LOWER]);
719 BUG_ON(!list_empty(&edge->list[UPPER]));
720 exist = edge->node[UPPER];
721 /*
722 * add the upper level block to pending list if we need
723 * check its backrefs
724 */
725 if (!exist->checked)
726 list_add_tail(&edge->list[UPPER], &list);
727 } else {
728 exist = NULL;
729 }
730
731 while (1) {
732 cond_resched();
733 eb = path1->nodes[0];
734
735 if (ptr >= end) {
736 if (path1->slots[0] >= btrfs_header_nritems(eb)) {
737 ret = btrfs_next_leaf(rc->extent_root, path1);
738 if (ret < 0) {
739 err = ret;
740 goto out;
741 }
742 if (ret > 0)
743 break;
744 eb = path1->nodes[0];
745 }
746
747 btrfs_item_key_to_cpu(eb, &key, path1->slots[0]);
748 if (key.objectid != cur->bytenr) {
749 WARN_ON(exist);
750 break;
751 }
752
753 if (key.type == BTRFS_EXTENT_ITEM_KEY) {
754 ret = find_inline_backref(eb, path1->slots[0],
755 &ptr, &end);
756 if (ret)
757 goto next;
758 }
759 }
760
761 if (ptr < end) {
762 /* update key for inline back ref */
763 struct btrfs_extent_inline_ref *iref;
764 iref = (struct btrfs_extent_inline_ref *)ptr;
765 key.type = btrfs_extent_inline_ref_type(eb, iref);
766 key.offset = btrfs_extent_inline_ref_offset(eb, iref);
767 WARN_ON(key.type != BTRFS_TREE_BLOCK_REF_KEY &&
768 key.type != BTRFS_SHARED_BLOCK_REF_KEY);
769 }
770
771 if (exist &&
772 ((key.type == BTRFS_TREE_BLOCK_REF_KEY &&
773 exist->owner == key.offset) ||
774 (key.type == BTRFS_SHARED_BLOCK_REF_KEY &&
775 exist->bytenr == key.offset))) {
776 exist = NULL;
777 goto next;
778 }
779
780 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
781 if (key.type == BTRFS_SHARED_BLOCK_REF_KEY ||
782 key.type == BTRFS_EXTENT_REF_V0_KEY) {
783 if (key.type == BTRFS_EXTENT_REF_V0_KEY) {
784 struct btrfs_extent_ref_v0 *ref0;
785 ref0 = btrfs_item_ptr(eb, path1->slots[0],
786 struct btrfs_extent_ref_v0);
787 if (key.objectid == key.offset) {
788 root = find_tree_root(rc, eb, ref0);
789 if (root && !should_ignore_root(root))
790 cur->root = root;
791 else
792 list_add(&cur->list, &useless);
793 break;
794 }
795 if (is_cowonly_root(btrfs_ref_root_v0(eb,
796 ref0)))
797 cur->cowonly = 1;
798 }
799 #else
800 BUG_ON(key.type == BTRFS_EXTENT_REF_V0_KEY);
801 if (key.type == BTRFS_SHARED_BLOCK_REF_KEY) {
802 #endif
803 if (key.objectid == key.offset) {
804 /*
805 * only root blocks of reloc trees use
806 * backref of this type.
807 */
808 root = find_reloc_root(rc, cur->bytenr);
809 BUG_ON(!root);
810 cur->root = root;
811 break;
812 }
813
814 edge = alloc_backref_edge(cache);
815 if (!edge) {
816 err = -ENOMEM;
817 goto out;
818 }
819 rb_node = tree_search(&cache->rb_root, key.offset);
820 if (!rb_node) {
821 upper = alloc_backref_node(cache);
822 if (!upper) {
823 free_backref_edge(cache, edge);
824 err = -ENOMEM;
825 goto out;
826 }
827 upper->bytenr = key.offset;
828 upper->level = cur->level + 1;
829 /*
830 * backrefs for the upper level block isn't
831 * cached, add the block to pending list
832 */
833 list_add_tail(&edge->list[UPPER], &list);
834 } else {
835 upper = rb_entry(rb_node, struct backref_node,
836 rb_node);
837 BUG_ON(!upper->checked);
838 INIT_LIST_HEAD(&edge->list[UPPER]);
839 }
840 list_add_tail(&edge->list[LOWER], &cur->upper);
841 edge->node[LOWER] = cur;
842 edge->node[UPPER] = upper;
843
844 goto next;
845 } else if (key.type != BTRFS_TREE_BLOCK_REF_KEY) {
846 goto next;
847 }
848
849 /* key.type == BTRFS_TREE_BLOCK_REF_KEY */
850 root = read_fs_root(rc->extent_root->fs_info, key.offset);
851 if (IS_ERR(root)) {
852 err = PTR_ERR(root);
853 goto out;
854 }
855
856 if (!root->ref_cows)
857 cur->cowonly = 1;
858
859 if (btrfs_root_level(&root->root_item) == cur->level) {
860 /* tree root */
861 BUG_ON(btrfs_root_bytenr(&root->root_item) !=
862 cur->bytenr);
863 if (should_ignore_root(root))
864 list_add(&cur->list, &useless);
865 else
866 cur->root = root;
867 break;
868 }
869
870 level = cur->level + 1;
871
872 /*
873 * searching the tree to find upper level blocks
874 * reference the block.
875 */
876 path2->search_commit_root = 1;
877 path2->skip_locking = 1;
878 path2->lowest_level = level;
879 ret = btrfs_search_slot(NULL, root, node_key, path2, 0, 0);
880 path2->lowest_level = 0;
881 if (ret < 0) {
882 err = ret;
883 goto out;
884 }
885 if (ret > 0 && path2->slots[level] > 0)
886 path2->slots[level]--;
887
888 eb = path2->nodes[level];
889 WARN_ON(btrfs_node_blockptr(eb, path2->slots[level]) !=
890 cur->bytenr);
891
892 lower = cur;
893 for (; level < BTRFS_MAX_LEVEL; level++) {
894 if (!path2->nodes[level]) {
895 BUG_ON(btrfs_root_bytenr(&root->root_item) !=
896 lower->bytenr);
897 if (should_ignore_root(root))
898 list_add(&lower->list, &useless);
899 else
900 lower->root = root;
901 break;
902 }
903
904 edge = alloc_backref_edge(cache);
905 if (!edge) {
906 err = -ENOMEM;
907 goto out;
908 }
909
910 eb = path2->nodes[level];
911 rb_node = tree_search(&cache->rb_root, eb->start);
912 if (!rb_node) {
913 upper = alloc_backref_node(cache);
914 if (!upper) {
915 free_backref_edge(cache, edge);
916 err = -ENOMEM;
917 goto out;
918 }
919 upper->bytenr = eb->start;
920 upper->owner = btrfs_header_owner(eb);
921 upper->level = lower->level + 1;
922 if (!root->ref_cows)
923 upper->cowonly = 1;
924
925 /*
926 * if we know the block isn't shared
927 * we can void checking its backrefs.
928 */
929 if (btrfs_block_can_be_shared(root, eb))
930 upper->checked = 0;
931 else
932 upper->checked = 1;
933
934 /*
935 * add the block to pending list if we
936 * need check its backrefs. only block
937 * at 'cur->level + 1' is added to the
938 * tail of pending list. this guarantees
939 * we check backrefs from lower level
940 * blocks to upper level blocks.
941 */
942 if (!upper->checked &&
943 level == cur->level + 1) {
944 list_add_tail(&edge->list[UPPER],
945 &list);
946 } else
947 INIT_LIST_HEAD(&edge->list[UPPER]);
948 } else {
949 upper = rb_entry(rb_node, struct backref_node,
950 rb_node);
951 BUG_ON(!upper->checked);
952 INIT_LIST_HEAD(&edge->list[UPPER]);
953 if (!upper->owner)
954 upper->owner = btrfs_header_owner(eb);
955 }
956 list_add_tail(&edge->list[LOWER], &lower->upper);
957 edge->node[LOWER] = lower;
958 edge->node[UPPER] = upper;
959
960 if (rb_node)
961 break;
962 lower = upper;
963 upper = NULL;
964 }
965 btrfs_release_path(root, path2);
966 next:
967 if (ptr < end) {
968 ptr += btrfs_extent_inline_ref_size(key.type);
969 if (ptr >= end) {
970 WARN_ON(ptr > end);
971 ptr = 0;
972 end = 0;
973 }
974 }
975 if (ptr >= end)
976 path1->slots[0]++;
977 }
978 btrfs_release_path(rc->extent_root, path1);
979
980 cur->checked = 1;
981 WARN_ON(exist);
982
983 /* the pending list isn't empty, take the first block to process */
984 if (!list_empty(&list)) {
985 edge = list_entry(list.next, struct backref_edge, list[UPPER]);
986 list_del_init(&edge->list[UPPER]);
987 cur = edge->node[UPPER];
988 goto again;
989 }
990
991 /*
992 * everything goes well, connect backref nodes and insert backref nodes
993 * into the cache.
994 */
995 BUG_ON(!node->checked);
996 cowonly = node->cowonly;
997 if (!cowonly) {
998 rb_node = tree_insert(&cache->rb_root, node->bytenr,
999 &node->rb_node);
1000 BUG_ON(rb_node);
1001 list_add_tail(&node->lower, &cache->leaves);
1002 }
1003
1004 list_for_each_entry(edge, &node->upper, list[LOWER])
1005 list_add_tail(&edge->list[UPPER], &list);
1006
1007 while (!list_empty(&list)) {
1008 edge = list_entry(list.next, struct backref_edge, list[UPPER]);
1009 list_del_init(&edge->list[UPPER]);
1010 upper = edge->node[UPPER];
1011 if (upper->detached) {
1012 list_del(&edge->list[LOWER]);
1013 lower = edge->node[LOWER];
1014 free_backref_edge(cache, edge);
1015 if (list_empty(&lower->upper))
1016 list_add(&lower->list, &useless);
1017 continue;
1018 }
1019
1020 if (!RB_EMPTY_NODE(&upper->rb_node)) {
1021 if (upper->lowest) {
1022 list_del_init(&upper->lower);
1023 upper->lowest = 0;
1024 }
1025
1026 list_add_tail(&edge->list[UPPER], &upper->lower);
1027 continue;
1028 }
1029
1030 BUG_ON(!upper->checked);
1031 BUG_ON(cowonly != upper->cowonly);
1032 if (!cowonly) {
1033 rb_node = tree_insert(&cache->rb_root, upper->bytenr,
1034 &upper->rb_node);
1035 BUG_ON(rb_node);
1036 }
1037
1038 list_add_tail(&edge->list[UPPER], &upper->lower);
1039
1040 list_for_each_entry(edge, &upper->upper, list[LOWER])
1041 list_add_tail(&edge->list[UPPER], &list);
1042 }
1043 /*
1044 * process useless backref nodes. backref nodes for tree leaves
1045 * are deleted from the cache. backref nodes for upper level
1046 * tree blocks are left in the cache to avoid unnecessary backref
1047 * lookup.
1048 */
1049 while (!list_empty(&useless)) {
1050 upper = list_entry(useless.next, struct backref_node, list);
1051 list_del_init(&upper->list);
1052 BUG_ON(!list_empty(&upper->upper));
1053 if (upper == node)
1054 node = NULL;
1055 if (upper->lowest) {
1056 list_del_init(&upper->lower);
1057 upper->lowest = 0;
1058 }
1059 while (!list_empty(&upper->lower)) {
1060 edge = list_entry(upper->lower.next,
1061 struct backref_edge, list[UPPER]);
1062 list_del(&edge->list[UPPER]);
1063 list_del(&edge->list[LOWER]);
1064 lower = edge->node[LOWER];
1065 free_backref_edge(cache, edge);
1066
1067 if (list_empty(&lower->upper))
1068 list_add(&lower->list, &useless);
1069 }
1070 __mark_block_processed(rc, upper);
1071 if (upper->level > 0) {
1072 list_add(&upper->list, &cache->detached);
1073 upper->detached = 1;
1074 } else {
1075 rb_erase(&upper->rb_node, &cache->rb_root);
1076 free_backref_node(cache, upper);
1077 }
1078 }
1079 out:
1080 btrfs_free_path(path1);
1081 btrfs_free_path(path2);
1082 if (err) {
1083 while (!list_empty(&useless)) {
1084 lower = list_entry(useless.next,
1085 struct backref_node, upper);
1086 list_del_init(&lower->upper);
1087 }
1088 upper = node;
1089 INIT_LIST_HEAD(&list);
1090 while (upper) {
1091 if (RB_EMPTY_NODE(&upper->rb_node)) {
1092 list_splice_tail(&upper->upper, &list);
1093 free_backref_node(cache, upper);
1094 }
1095
1096 if (list_empty(&list))
1097 break;
1098
1099 edge = list_entry(list.next, struct backref_edge,
1100 list[LOWER]);
1101 list_del(&edge->list[LOWER]);
1102 upper = edge->node[UPPER];
1103 free_backref_edge(cache, edge);
1104 }
1105 return ERR_PTR(err);
1106 }
1107 BUG_ON(node && node->detached);
1108 return node;
1109 }
1110
1111 /*
1112 * helper to add backref node for the newly created snapshot.
1113 * the backref node is created by cloning backref node that
1114 * corresponds to root of source tree
1115 */
1116 static int clone_backref_node(struct btrfs_trans_handle *trans,
1117 struct reloc_control *rc,
1118 struct btrfs_root *src,
1119 struct btrfs_root *dest)
1120 {
1121 struct btrfs_root *reloc_root = src->reloc_root;
1122 struct backref_cache *cache = &rc->backref_cache;
1123 struct backref_node *node = NULL;
1124 struct backref_node *new_node;
1125 struct backref_edge *edge;
1126 struct backref_edge *new_edge;
1127 struct rb_node *rb_node;
1128
1129 if (cache->last_trans > 0)
1130 update_backref_cache(trans, cache);
1131
1132 rb_node = tree_search(&cache->rb_root, src->commit_root->start);
1133 if (rb_node) {
1134 node = rb_entry(rb_node, struct backref_node, rb_node);
1135 if (node->detached)
1136 node = NULL;
1137 else
1138 BUG_ON(node->new_bytenr != reloc_root->node->start);
1139 }
1140
1141 if (!node) {
1142 rb_node = tree_search(&cache->rb_root,
1143 reloc_root->commit_root->start);
1144 if (rb_node) {
1145 node = rb_entry(rb_node, struct backref_node,
1146 rb_node);
1147 BUG_ON(node->detached);
1148 }
1149 }
1150
1151 if (!node)
1152 return 0;
1153
1154 new_node = alloc_backref_node(cache);
1155 if (!new_node)
1156 return -ENOMEM;
1157
1158 new_node->bytenr = dest->node->start;
1159 new_node->level = node->level;
1160 new_node->lowest = node->lowest;
1161 new_node->root = dest;
1162
1163 if (!node->lowest) {
1164 list_for_each_entry(edge, &node->lower, list[UPPER]) {
1165 new_edge = alloc_backref_edge(cache);
1166 if (!new_edge)
1167 goto fail;
1168
1169 new_edge->node[UPPER] = new_node;
1170 new_edge->node[LOWER] = edge->node[LOWER];
1171 list_add_tail(&new_edge->list[UPPER],
1172 &new_node->lower);
1173 }
1174 }
1175
1176 rb_node = tree_insert(&cache->rb_root, new_node->bytenr,
1177 &new_node->rb_node);
1178 BUG_ON(rb_node);
1179
1180 if (!new_node->lowest) {
1181 list_for_each_entry(new_edge, &new_node->lower, list[UPPER]) {
1182 list_add_tail(&new_edge->list[LOWER],
1183 &new_edge->node[LOWER]->upper);
1184 }
1185 }
1186 return 0;
1187 fail:
1188 while (!list_empty(&new_node->lower)) {
1189 new_edge = list_entry(new_node->lower.next,
1190 struct backref_edge, list[UPPER]);
1191 list_del(&new_edge->list[UPPER]);
1192 free_backref_edge(cache, new_edge);
1193 }
1194 free_backref_node(cache, new_node);
1195 return -ENOMEM;
1196 }
1197
1198 /*
1199 * helper to add 'address of tree root -> reloc tree' mapping
1200 */
1201 static int __add_reloc_root(struct btrfs_root *root)
1202 {
1203 struct rb_node *rb_node;
1204 struct mapping_node *node;
1205 struct reloc_control *rc = root->fs_info->reloc_ctl;
1206
1207 node = kmalloc(sizeof(*node), GFP_NOFS);
1208 BUG_ON(!node);
1209
1210 node->bytenr = root->node->start;
1211 node->data = root;
1212
1213 spin_lock(&rc->reloc_root_tree.lock);
1214 rb_node = tree_insert(&rc->reloc_root_tree.rb_root,
1215 node->bytenr, &node->rb_node);
1216 spin_unlock(&rc->reloc_root_tree.lock);
1217 BUG_ON(rb_node);
1218
1219 list_add_tail(&root->root_list, &rc->reloc_roots);
1220 return 0;
1221 }
1222
1223 /*
1224 * helper to update/delete the 'address of tree root -> reloc tree'
1225 * mapping
1226 */
1227 static int __update_reloc_root(struct btrfs_root *root, int del)
1228 {
1229 struct rb_node *rb_node;
1230 struct mapping_node *node = NULL;
1231 struct reloc_control *rc = root->fs_info->reloc_ctl;
1232
1233 spin_lock(&rc->reloc_root_tree.lock);
1234 rb_node = tree_search(&rc->reloc_root_tree.rb_root,
1235 root->commit_root->start);
1236 if (rb_node) {
1237 node = rb_entry(rb_node, struct mapping_node, rb_node);
1238 rb_erase(&node->rb_node, &rc->reloc_root_tree.rb_root);
1239 }
1240 spin_unlock(&rc->reloc_root_tree.lock);
1241
1242 BUG_ON((struct btrfs_root *)node->data != root);
1243
1244 if (!del) {
1245 spin_lock(&rc->reloc_root_tree.lock);
1246 node->bytenr = root->node->start;
1247 rb_node = tree_insert(&rc->reloc_root_tree.rb_root,
1248 node->bytenr, &node->rb_node);
1249 spin_unlock(&rc->reloc_root_tree.lock);
1250 BUG_ON(rb_node);
1251 } else {
1252 list_del_init(&root->root_list);
1253 kfree(node);
1254 }
1255 return 0;
1256 }
1257
1258 static struct btrfs_root *create_reloc_root(struct btrfs_trans_handle *trans,
1259 struct btrfs_root *root, u64 objectid)
1260 {
1261 struct btrfs_root *reloc_root;
1262 struct extent_buffer *eb;
1263 struct btrfs_root_item *root_item;
1264 struct btrfs_key root_key;
1265 int ret;
1266
1267 root_item = kmalloc(sizeof(*root_item), GFP_NOFS);
1268 BUG_ON(!root_item);
1269
1270 root_key.objectid = BTRFS_TREE_RELOC_OBJECTID;
1271 root_key.type = BTRFS_ROOT_ITEM_KEY;
1272 root_key.offset = objectid;
1273
1274 if (root->root_key.objectid == objectid) {
1275 /* called by btrfs_init_reloc_root */
1276 ret = btrfs_copy_root(trans, root, root->commit_root, &eb,
1277 BTRFS_TREE_RELOC_OBJECTID);
1278 BUG_ON(ret);
1279
1280 btrfs_set_root_last_snapshot(&root->root_item,
1281 trans->transid - 1);
1282 } else {
1283 /*
1284 * called by btrfs_reloc_post_snapshot_hook.
1285 * the source tree is a reloc tree, all tree blocks
1286 * modified after it was created have RELOC flag
1287 * set in their headers. so it's OK to not update
1288 * the 'last_snapshot'.
1289 */
1290 ret = btrfs_copy_root(trans, root, root->node, &eb,
1291 BTRFS_TREE_RELOC_OBJECTID);
1292 BUG_ON(ret);
1293 }
1294
1295 memcpy(root_item, &root->root_item, sizeof(*root_item));
1296 btrfs_set_root_bytenr(root_item, eb->start);
1297 btrfs_set_root_level(root_item, btrfs_header_level(eb));
1298 btrfs_set_root_generation(root_item, trans->transid);
1299
1300 if (root->root_key.objectid == objectid) {
1301 btrfs_set_root_refs(root_item, 0);
1302 memset(&root_item->drop_progress, 0,
1303 sizeof(struct btrfs_disk_key));
1304 root_item->drop_level = 0;
1305 }
1306
1307 btrfs_tree_unlock(eb);
1308 free_extent_buffer(eb);
1309
1310 ret = btrfs_insert_root(trans, root->fs_info->tree_root,
1311 &root_key, root_item);
1312 BUG_ON(ret);
1313 kfree(root_item);
1314
1315 reloc_root = btrfs_read_fs_root_no_radix(root->fs_info->tree_root,
1316 &root_key);
1317 BUG_ON(IS_ERR(reloc_root));
1318 reloc_root->last_trans = trans->transid;
1319 return reloc_root;
1320 }
1321
1322 /*
1323 * create reloc tree for a given fs tree. reloc tree is just a
1324 * snapshot of the fs tree with special root objectid.
1325 */
1326 int btrfs_init_reloc_root(struct btrfs_trans_handle *trans,
1327 struct btrfs_root *root)
1328 {
1329 struct btrfs_root *reloc_root;
1330 struct reloc_control *rc = root->fs_info->reloc_ctl;
1331 int clear_rsv = 0;
1332
1333 if (root->reloc_root) {
1334 reloc_root = root->reloc_root;
1335 reloc_root->last_trans = trans->transid;
1336 return 0;
1337 }
1338
1339 if (!rc || !rc->create_reloc_tree ||
1340 root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID)
1341 return 0;
1342
1343 if (!trans->block_rsv) {
1344 trans->block_rsv = rc->block_rsv;
1345 clear_rsv = 1;
1346 }
1347 reloc_root = create_reloc_root(trans, root, root->root_key.objectid);
1348 if (clear_rsv)
1349 trans->block_rsv = NULL;
1350
1351 __add_reloc_root(reloc_root);
1352 root->reloc_root = reloc_root;
1353 return 0;
1354 }
1355
1356 /*
1357 * update root item of reloc tree
1358 */
1359 int btrfs_update_reloc_root(struct btrfs_trans_handle *trans,
1360 struct btrfs_root *root)
1361 {
1362 struct btrfs_root *reloc_root;
1363 struct btrfs_root_item *root_item;
1364 int del = 0;
1365 int ret;
1366
1367 if (!root->reloc_root)
1368 return 0;
1369
1370 reloc_root = root->reloc_root;
1371 root_item = &reloc_root->root_item;
1372
1373 if (root->fs_info->reloc_ctl->merge_reloc_tree &&
1374 btrfs_root_refs(root_item) == 0) {
1375 root->reloc_root = NULL;
1376 del = 1;
1377 }
1378
1379 __update_reloc_root(reloc_root, del);
1380
1381 if (reloc_root->commit_root != reloc_root->node) {
1382 btrfs_set_root_node(root_item, reloc_root->node);
1383 free_extent_buffer(reloc_root->commit_root);
1384 reloc_root->commit_root = btrfs_root_node(reloc_root);
1385 }
1386
1387 ret = btrfs_update_root(trans, root->fs_info->tree_root,
1388 &reloc_root->root_key, root_item);
1389 BUG_ON(ret);
1390 return 0;
1391 }
1392
1393 /*
1394 * helper to find first cached inode with inode number >= objectid
1395 * in a subvolume
1396 */
1397 static struct inode *find_next_inode(struct btrfs_root *root, u64 objectid)
1398 {
1399 struct rb_node *node;
1400 struct rb_node *prev;
1401 struct btrfs_inode *entry;
1402 struct inode *inode;
1403
1404 spin_lock(&root->inode_lock);
1405 again:
1406 node = root->inode_tree.rb_node;
1407 prev = NULL;
1408 while (node) {
1409 prev = node;
1410 entry = rb_entry(node, struct btrfs_inode, rb_node);
1411
1412 if (objectid < entry->vfs_inode.i_ino)
1413 node = node->rb_left;
1414 else if (objectid > entry->vfs_inode.i_ino)
1415 node = node->rb_right;
1416 else
1417 break;
1418 }
1419 if (!node) {
1420 while (prev) {
1421 entry = rb_entry(prev, struct btrfs_inode, rb_node);
1422 if (objectid <= entry->vfs_inode.i_ino) {
1423 node = prev;
1424 break;
1425 }
1426 prev = rb_next(prev);
1427 }
1428 }
1429 while (node) {
1430 entry = rb_entry(node, struct btrfs_inode, rb_node);
1431 inode = igrab(&entry->vfs_inode);
1432 if (inode) {
1433 spin_unlock(&root->inode_lock);
1434 return inode;
1435 }
1436
1437 objectid = entry->vfs_inode.i_ino + 1;
1438 if (cond_resched_lock(&root->inode_lock))
1439 goto again;
1440
1441 node = rb_next(node);
1442 }
1443 spin_unlock(&root->inode_lock);
1444 return NULL;
1445 }
1446
1447 static int in_block_group(u64 bytenr,
1448 struct btrfs_block_group_cache *block_group)
1449 {
1450 if (bytenr >= block_group->key.objectid &&
1451 bytenr < block_group->key.objectid + block_group->key.offset)
1452 return 1;
1453 return 0;
1454 }
1455
1456 /*
1457 * get new location of data
1458 */
1459 static int get_new_location(struct inode *reloc_inode, u64 *new_bytenr,
1460 u64 bytenr, u64 num_bytes)
1461 {
1462 struct btrfs_root *root = BTRFS_I(reloc_inode)->root;
1463 struct btrfs_path *path;
1464 struct btrfs_file_extent_item *fi;
1465 struct extent_buffer *leaf;
1466 int ret;
1467
1468 path = btrfs_alloc_path();
1469 if (!path)
1470 return -ENOMEM;
1471
1472 bytenr -= BTRFS_I(reloc_inode)->index_cnt;
1473 ret = btrfs_lookup_file_extent(NULL, root, path, reloc_inode->i_ino,
1474 bytenr, 0);
1475 if (ret < 0)
1476 goto out;
1477 if (ret > 0) {
1478 ret = -ENOENT;
1479 goto out;
1480 }
1481
1482 leaf = path->nodes[0];
1483 fi = btrfs_item_ptr(leaf, path->slots[0],
1484 struct btrfs_file_extent_item);
1485
1486 BUG_ON(btrfs_file_extent_offset(leaf, fi) ||
1487 btrfs_file_extent_compression(leaf, fi) ||
1488 btrfs_file_extent_encryption(leaf, fi) ||
1489 btrfs_file_extent_other_encoding(leaf, fi));
1490
1491 if (num_bytes != btrfs_file_extent_disk_num_bytes(leaf, fi)) {
1492 ret = 1;
1493 goto out;
1494 }
1495
1496 *new_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
1497 ret = 0;
1498 out:
1499 btrfs_free_path(path);
1500 return ret;
1501 }
1502
1503 /*
1504 * update file extent items in the tree leaf to point to
1505 * the new locations.
1506 */
1507 static noinline_for_stack
1508 int replace_file_extents(struct btrfs_trans_handle *trans,
1509 struct reloc_control *rc,
1510 struct btrfs_root *root,
1511 struct extent_buffer *leaf)
1512 {
1513 struct btrfs_key key;
1514 struct btrfs_file_extent_item *fi;
1515 struct inode *inode = NULL;
1516 u64 parent;
1517 u64 bytenr;
1518 u64 new_bytenr = 0;
1519 u64 num_bytes;
1520 u64 end;
1521 u32 nritems;
1522 u32 i;
1523 int ret;
1524 int first = 1;
1525 int dirty = 0;
1526
1527 if (rc->stage != UPDATE_DATA_PTRS)
1528 return 0;
1529
1530 /* reloc trees always use full backref */
1531 if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID)
1532 parent = leaf->start;
1533 else
1534 parent = 0;
1535
1536 nritems = btrfs_header_nritems(leaf);
1537 for (i = 0; i < nritems; i++) {
1538 cond_resched();
1539 btrfs_item_key_to_cpu(leaf, &key, i);
1540 if (key.type != BTRFS_EXTENT_DATA_KEY)
1541 continue;
1542 fi = btrfs_item_ptr(leaf, i, struct btrfs_file_extent_item);
1543 if (btrfs_file_extent_type(leaf, fi) ==
1544 BTRFS_FILE_EXTENT_INLINE)
1545 continue;
1546 bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
1547 num_bytes = btrfs_file_extent_disk_num_bytes(leaf, fi);
1548 if (bytenr == 0)
1549 continue;
1550 if (!in_block_group(bytenr, rc->block_group))
1551 continue;
1552
1553 /*
1554 * if we are modifying block in fs tree, wait for readpage
1555 * to complete and drop the extent cache
1556 */
1557 if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) {
1558 if (first) {
1559 inode = find_next_inode(root, key.objectid);
1560 first = 0;
1561 } else if (inode && inode->i_ino < key.objectid) {
1562 btrfs_add_delayed_iput(inode);
1563 inode = find_next_inode(root, key.objectid);
1564 }
1565 if (inode && inode->i_ino == key.objectid) {
1566 end = key.offset +
1567 btrfs_file_extent_num_bytes(leaf, fi);
1568 WARN_ON(!IS_ALIGNED(key.offset,
1569 root->sectorsize));
1570 WARN_ON(!IS_ALIGNED(end, root->sectorsize));
1571 end--;
1572 ret = try_lock_extent(&BTRFS_I(inode)->io_tree,
1573 key.offset, end,
1574 GFP_NOFS);
1575 if (!ret)
1576 continue;
1577
1578 btrfs_drop_extent_cache(inode, key.offset, end,
1579 1);
1580 unlock_extent(&BTRFS_I(inode)->io_tree,
1581 key.offset, end, GFP_NOFS);
1582 }
1583 }
1584
1585 ret = get_new_location(rc->data_inode, &new_bytenr,
1586 bytenr, num_bytes);
1587 if (ret > 0) {
1588 WARN_ON(1);
1589 continue;
1590 }
1591 BUG_ON(ret < 0);
1592
1593 btrfs_set_file_extent_disk_bytenr(leaf, fi, new_bytenr);
1594 dirty = 1;
1595
1596 key.offset -= btrfs_file_extent_offset(leaf, fi);
1597 ret = btrfs_inc_extent_ref(trans, root, new_bytenr,
1598 num_bytes, parent,
1599 btrfs_header_owner(leaf),
1600 key.objectid, key.offset);
1601 BUG_ON(ret);
1602
1603 ret = btrfs_free_extent(trans, root, bytenr, num_bytes,
1604 parent, btrfs_header_owner(leaf),
1605 key.objectid, key.offset);
1606 BUG_ON(ret);
1607 }
1608 if (dirty)
1609 btrfs_mark_buffer_dirty(leaf);
1610 if (inode)
1611 btrfs_add_delayed_iput(inode);
1612 return 0;
1613 }
1614
1615 static noinline_for_stack
1616 int memcmp_node_keys(struct extent_buffer *eb, int slot,
1617 struct btrfs_path *path, int level)
1618 {
1619 struct btrfs_disk_key key1;
1620 struct btrfs_disk_key key2;
1621 btrfs_node_key(eb, &key1, slot);
1622 btrfs_node_key(path->nodes[level], &key2, path->slots[level]);
1623 return memcmp(&key1, &key2, sizeof(key1));
1624 }
1625
1626 /*
1627 * try to replace tree blocks in fs tree with the new blocks
1628 * in reloc tree. tree blocks haven't been modified since the
1629 * reloc tree was create can be replaced.
1630 *
1631 * if a block was replaced, level of the block + 1 is returned.
1632 * if no block got replaced, 0 is returned. if there are other
1633 * errors, a negative error number is returned.
1634 */
1635 static noinline_for_stack
1636 int replace_path(struct btrfs_trans_handle *trans,
1637 struct btrfs_root *dest, struct btrfs_root *src,
1638 struct btrfs_path *path, struct btrfs_key *next_key,
1639 int lowest_level, int max_level)
1640 {
1641 struct extent_buffer *eb;
1642 struct extent_buffer *parent;
1643 struct btrfs_key key;
1644 u64 old_bytenr;
1645 u64 new_bytenr;
1646 u64 old_ptr_gen;
1647 u64 new_ptr_gen;
1648 u64 last_snapshot;
1649 u32 blocksize;
1650 int cow = 0;
1651 int level;
1652 int ret;
1653 int slot;
1654
1655 BUG_ON(src->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID);
1656 BUG_ON(dest->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID);
1657
1658 last_snapshot = btrfs_root_last_snapshot(&src->root_item);
1659 again:
1660 slot = path->slots[lowest_level];
1661 btrfs_node_key_to_cpu(path->nodes[lowest_level], &key, slot);
1662
1663 eb = btrfs_lock_root_node(dest);
1664 btrfs_set_lock_blocking(eb);
1665 level = btrfs_header_level(eb);
1666
1667 if (level < lowest_level) {
1668 btrfs_tree_unlock(eb);
1669 free_extent_buffer(eb);
1670 return 0;
1671 }
1672
1673 if (cow) {
1674 ret = btrfs_cow_block(trans, dest, eb, NULL, 0, &eb);
1675 BUG_ON(ret);
1676 }
1677 btrfs_set_lock_blocking(eb);
1678
1679 if (next_key) {
1680 next_key->objectid = (u64)-1;
1681 next_key->type = (u8)-1;
1682 next_key->offset = (u64)-1;
1683 }
1684
1685 parent = eb;
1686 while (1) {
1687 level = btrfs_header_level(parent);
1688 BUG_ON(level < lowest_level);
1689
1690 ret = btrfs_bin_search(parent, &key, level, &slot);
1691 if (ret && slot > 0)
1692 slot--;
1693
1694 if (next_key && slot + 1 < btrfs_header_nritems(parent))
1695 btrfs_node_key_to_cpu(parent, next_key, slot + 1);
1696
1697 old_bytenr = btrfs_node_blockptr(parent, slot);
1698 blocksize = btrfs_level_size(dest, level - 1);
1699 old_ptr_gen = btrfs_node_ptr_generation(parent, slot);
1700
1701 if (level <= max_level) {
1702 eb = path->nodes[level];
1703 new_bytenr = btrfs_node_blockptr(eb,
1704 path->slots[level]);
1705 new_ptr_gen = btrfs_node_ptr_generation(eb,
1706 path->slots[level]);
1707 } else {
1708 new_bytenr = 0;
1709 new_ptr_gen = 0;
1710 }
1711
1712 if (new_bytenr > 0 && new_bytenr == old_bytenr) {
1713 WARN_ON(1);
1714 ret = level;
1715 break;
1716 }
1717
1718 if (new_bytenr == 0 || old_ptr_gen > last_snapshot ||
1719 memcmp_node_keys(parent, slot, path, level)) {
1720 if (level <= lowest_level) {
1721 ret = 0;
1722 break;
1723 }
1724
1725 eb = read_tree_block(dest, old_bytenr, blocksize,
1726 old_ptr_gen);
1727 btrfs_tree_lock(eb);
1728 if (cow) {
1729 ret = btrfs_cow_block(trans, dest, eb, parent,
1730 slot, &eb);
1731 BUG_ON(ret);
1732 }
1733 btrfs_set_lock_blocking(eb);
1734
1735 btrfs_tree_unlock(parent);
1736 free_extent_buffer(parent);
1737
1738 parent = eb;
1739 continue;
1740 }
1741
1742 if (!cow) {
1743 btrfs_tree_unlock(parent);
1744 free_extent_buffer(parent);
1745 cow = 1;
1746 goto again;
1747 }
1748
1749 btrfs_node_key_to_cpu(path->nodes[level], &key,
1750 path->slots[level]);
1751 btrfs_release_path(src, path);
1752
1753 path->lowest_level = level;
1754 ret = btrfs_search_slot(trans, src, &key, path, 0, 1);
1755 path->lowest_level = 0;
1756 BUG_ON(ret);
1757
1758 /*
1759 * swap blocks in fs tree and reloc tree.
1760 */
1761 btrfs_set_node_blockptr(parent, slot, new_bytenr);
1762 btrfs_set_node_ptr_generation(parent, slot, new_ptr_gen);
1763 btrfs_mark_buffer_dirty(parent);
1764
1765 btrfs_set_node_blockptr(path->nodes[level],
1766 path->slots[level], old_bytenr);
1767 btrfs_set_node_ptr_generation(path->nodes[level],
1768 path->slots[level], old_ptr_gen);
1769 btrfs_mark_buffer_dirty(path->nodes[level]);
1770
1771 ret = btrfs_inc_extent_ref(trans, src, old_bytenr, blocksize,
1772 path->nodes[level]->start,
1773 src->root_key.objectid, level - 1, 0);
1774 BUG_ON(ret);
1775 ret = btrfs_inc_extent_ref(trans, dest, new_bytenr, blocksize,
1776 0, dest->root_key.objectid, level - 1,
1777 0);
1778 BUG_ON(ret);
1779
1780 ret = btrfs_free_extent(trans, src, new_bytenr, blocksize,
1781 path->nodes[level]->start,
1782 src->root_key.objectid, level - 1, 0);
1783 BUG_ON(ret);
1784
1785 ret = btrfs_free_extent(trans, dest, old_bytenr, blocksize,
1786 0, dest->root_key.objectid, level - 1,
1787 0);
1788 BUG_ON(ret);
1789
1790 btrfs_unlock_up_safe(path, 0);
1791
1792 ret = level;
1793 break;
1794 }
1795 btrfs_tree_unlock(parent);
1796 free_extent_buffer(parent);
1797 return ret;
1798 }
1799
1800 /*
1801 * helper to find next relocated block in reloc tree
1802 */
1803 static noinline_for_stack
1804 int walk_up_reloc_tree(struct btrfs_root *root, struct btrfs_path *path,
1805 int *level)
1806 {
1807 struct extent_buffer *eb;
1808 int i;
1809 u64 last_snapshot;
1810 u32 nritems;
1811
1812 last_snapshot = btrfs_root_last_snapshot(&root->root_item);
1813
1814 for (i = 0; i < *level; i++) {
1815 free_extent_buffer(path->nodes[i]);
1816 path->nodes[i] = NULL;
1817 }
1818
1819 for (i = *level; i < BTRFS_MAX_LEVEL && path->nodes[i]; i++) {
1820 eb = path->nodes[i];
1821 nritems = btrfs_header_nritems(eb);
1822 while (path->slots[i] + 1 < nritems) {
1823 path->slots[i]++;
1824 if (btrfs_node_ptr_generation(eb, path->slots[i]) <=
1825 last_snapshot)
1826 continue;
1827
1828 *level = i;
1829 return 0;
1830 }
1831 free_extent_buffer(path->nodes[i]);
1832 path->nodes[i] = NULL;
1833 }
1834 return 1;
1835 }
1836
1837 /*
1838 * walk down reloc tree to find relocated block of lowest level
1839 */
1840 static noinline_for_stack
1841 int walk_down_reloc_tree(struct btrfs_root *root, struct btrfs_path *path,
1842 int *level)
1843 {
1844 struct extent_buffer *eb = NULL;
1845 int i;
1846 u64 bytenr;
1847 u64 ptr_gen = 0;
1848 u64 last_snapshot;
1849 u32 blocksize;
1850 u32 nritems;
1851
1852 last_snapshot = btrfs_root_last_snapshot(&root->root_item);
1853
1854 for (i = *level; i > 0; i--) {
1855 eb = path->nodes[i];
1856 nritems = btrfs_header_nritems(eb);
1857 while (path->slots[i] < nritems) {
1858 ptr_gen = btrfs_node_ptr_generation(eb, path->slots[i]);
1859 if (ptr_gen > last_snapshot)
1860 break;
1861 path->slots[i]++;
1862 }
1863 if (path->slots[i] >= nritems) {
1864 if (i == *level)
1865 break;
1866 *level = i + 1;
1867 return 0;
1868 }
1869 if (i == 1) {
1870 *level = i;
1871 return 0;
1872 }
1873
1874 bytenr = btrfs_node_blockptr(eb, path->slots[i]);
1875 blocksize = btrfs_level_size(root, i - 1);
1876 eb = read_tree_block(root, bytenr, blocksize, ptr_gen);
1877 BUG_ON(btrfs_header_level(eb) != i - 1);
1878 path->nodes[i - 1] = eb;
1879 path->slots[i - 1] = 0;
1880 }
1881 return 1;
1882 }
1883
1884 /*
1885 * invalidate extent cache for file extents whose key in range of
1886 * [min_key, max_key)
1887 */
1888 static int invalidate_extent_cache(struct btrfs_root *root,
1889 struct btrfs_key *min_key,
1890 struct btrfs_key *max_key)
1891 {
1892 struct inode *inode = NULL;
1893 u64 objectid;
1894 u64 start, end;
1895
1896 objectid = min_key->objectid;
1897 while (1) {
1898 cond_resched();
1899 iput(inode);
1900
1901 if (objectid > max_key->objectid)
1902 break;
1903
1904 inode = find_next_inode(root, objectid);
1905 if (!inode)
1906 break;
1907
1908 if (inode->i_ino > max_key->objectid) {
1909 iput(inode);
1910 break;
1911 }
1912
1913 objectid = inode->i_ino + 1;
1914 if (!S_ISREG(inode->i_mode))
1915 continue;
1916
1917 if (unlikely(min_key->objectid == inode->i_ino)) {
1918 if (min_key->type > BTRFS_EXTENT_DATA_KEY)
1919 continue;
1920 if (min_key->type < BTRFS_EXTENT_DATA_KEY)
1921 start = 0;
1922 else {
1923 start = min_key->offset;
1924 WARN_ON(!IS_ALIGNED(start, root->sectorsize));
1925 }
1926 } else {
1927 start = 0;
1928 }
1929
1930 if (unlikely(max_key->objectid == inode->i_ino)) {
1931 if (max_key->type < BTRFS_EXTENT_DATA_KEY)
1932 continue;
1933 if (max_key->type > BTRFS_EXTENT_DATA_KEY) {
1934 end = (u64)-1;
1935 } else {
1936 if (max_key->offset == 0)
1937 continue;
1938 end = max_key->offset;
1939 WARN_ON(!IS_ALIGNED(end, root->sectorsize));
1940 end--;
1941 }
1942 } else {
1943 end = (u64)-1;
1944 }
1945
1946 /* the lock_extent waits for readpage to complete */
1947 lock_extent(&BTRFS_I(inode)->io_tree, start, end, GFP_NOFS);
1948 btrfs_drop_extent_cache(inode, start, end, 1);
1949 unlock_extent(&BTRFS_I(inode)->io_tree, start, end, GFP_NOFS);
1950 }
1951 return 0;
1952 }
1953
1954 static int find_next_key(struct btrfs_path *path, int level,
1955 struct btrfs_key *key)
1956
1957 {
1958 while (level < BTRFS_MAX_LEVEL) {
1959 if (!path->nodes[level])
1960 break;
1961 if (path->slots[level] + 1 <
1962 btrfs_header_nritems(path->nodes[level])) {
1963 btrfs_node_key_to_cpu(path->nodes[level], key,
1964 path->slots[level] + 1);
1965 return 0;
1966 }
1967 level++;
1968 }
1969 return 1;
1970 }
1971
1972 /*
1973 * merge the relocated tree blocks in reloc tree with corresponding
1974 * fs tree.
1975 */
1976 static noinline_for_stack int merge_reloc_root(struct reloc_control *rc,
1977 struct btrfs_root *root)
1978 {
1979 LIST_HEAD(inode_list);
1980 struct btrfs_key key;
1981 struct btrfs_key next_key;
1982 struct btrfs_trans_handle *trans;
1983 struct btrfs_root *reloc_root;
1984 struct btrfs_root_item *root_item;
1985 struct btrfs_path *path;
1986 struct extent_buffer *leaf;
1987 unsigned long nr;
1988 int level;
1989 int max_level;
1990 int replaced = 0;
1991 int ret;
1992 int err = 0;
1993 u32 min_reserved;
1994
1995 path = btrfs_alloc_path();
1996 if (!path)
1997 return -ENOMEM;
1998
1999 reloc_root = root->reloc_root;
2000 root_item = &reloc_root->root_item;
2001
2002 if (btrfs_disk_key_objectid(&root_item->drop_progress) == 0) {
2003 level = btrfs_root_level(root_item);
2004 extent_buffer_get(reloc_root->node);
2005 path->nodes[level] = reloc_root->node;
2006 path->slots[level] = 0;
2007 } else {
2008 btrfs_disk_key_to_cpu(&key, &root_item->drop_progress);
2009
2010 level = root_item->drop_level;
2011 BUG_ON(level == 0);
2012 path->lowest_level = level;
2013 ret = btrfs_search_slot(NULL, reloc_root, &key, path, 0, 0);
2014 path->lowest_level = 0;
2015 if (ret < 0) {
2016 btrfs_free_path(path);
2017 return ret;
2018 }
2019
2020 btrfs_node_key_to_cpu(path->nodes[level], &next_key,
2021 path->slots[level]);
2022 WARN_ON(memcmp(&key, &next_key, sizeof(key)));
2023
2024 btrfs_unlock_up_safe(path, 0);
2025 }
2026
2027 min_reserved = root->nodesize * (BTRFS_MAX_LEVEL - 1) * 2;
2028 memset(&next_key, 0, sizeof(next_key));
2029
2030 while (1) {
2031 trans = btrfs_start_transaction(root, 0);
2032 trans->block_rsv = rc->block_rsv;
2033
2034 ret = btrfs_block_rsv_check(trans, root, rc->block_rsv,
2035 min_reserved, 0);
2036 if (ret) {
2037 BUG_ON(ret != -EAGAIN);
2038 ret = btrfs_commit_transaction(trans, root);
2039 BUG_ON(ret);
2040 continue;
2041 }
2042
2043 replaced = 0;
2044 max_level = level;
2045
2046 ret = walk_down_reloc_tree(reloc_root, path, &level);
2047 if (ret < 0) {
2048 err = ret;
2049 goto out;
2050 }
2051 if (ret > 0)
2052 break;
2053
2054 if (!find_next_key(path, level, &key) &&
2055 btrfs_comp_cpu_keys(&next_key, &key) >= 0) {
2056 ret = 0;
2057 } else {
2058 ret = replace_path(trans, root, reloc_root, path,
2059 &next_key, level, max_level);
2060 }
2061 if (ret < 0) {
2062 err = ret;
2063 goto out;
2064 }
2065
2066 if (ret > 0) {
2067 level = ret;
2068 btrfs_node_key_to_cpu(path->nodes[level], &key,
2069 path->slots[level]);
2070 replaced = 1;
2071 }
2072
2073 ret = walk_up_reloc_tree(reloc_root, path, &level);
2074 if (ret > 0)
2075 break;
2076
2077 BUG_ON(level == 0);
2078 /*
2079 * save the merging progress in the drop_progress.
2080 * this is OK since root refs == 1 in this case.
2081 */
2082 btrfs_node_key(path->nodes[level], &root_item->drop_progress,
2083 path->slots[level]);
2084 root_item->drop_level = level;
2085
2086 nr = trans->blocks_used;
2087 btrfs_end_transaction_throttle(trans, root);
2088
2089 btrfs_btree_balance_dirty(root, nr);
2090
2091 if (replaced && rc->stage == UPDATE_DATA_PTRS)
2092 invalidate_extent_cache(root, &key, &next_key);
2093 }
2094
2095 /*
2096 * handle the case only one block in the fs tree need to be
2097 * relocated and the block is tree root.
2098 */
2099 leaf = btrfs_lock_root_node(root);
2100 ret = btrfs_cow_block(trans, root, leaf, NULL, 0, &leaf);
2101 btrfs_tree_unlock(leaf);
2102 free_extent_buffer(leaf);
2103 if (ret < 0)
2104 err = ret;
2105 out:
2106 btrfs_free_path(path);
2107
2108 if (err == 0) {
2109 memset(&root_item->drop_progress, 0,
2110 sizeof(root_item->drop_progress));
2111 root_item->drop_level = 0;
2112 btrfs_set_root_refs(root_item, 0);
2113 btrfs_update_reloc_root(trans, root);
2114 }
2115
2116 nr = trans->blocks_used;
2117 btrfs_end_transaction_throttle(trans, root);
2118
2119 btrfs_btree_balance_dirty(root, nr);
2120
2121 if (replaced && rc->stage == UPDATE_DATA_PTRS)
2122 invalidate_extent_cache(root, &key, &next_key);
2123
2124 return err;
2125 }
2126
2127 static noinline_for_stack
2128 int prepare_to_merge(struct reloc_control *rc, int err)
2129 {
2130 struct btrfs_root *root = rc->extent_root;
2131 struct btrfs_root *reloc_root;
2132 struct btrfs_trans_handle *trans;
2133 LIST_HEAD(reloc_roots);
2134 u64 num_bytes = 0;
2135 int ret;
2136 int retries = 0;
2137
2138 mutex_lock(&root->fs_info->trans_mutex);
2139 rc->merging_rsv_size += root->nodesize * (BTRFS_MAX_LEVEL - 1) * 2;
2140 rc->merging_rsv_size += rc->nodes_relocated * 2;
2141 mutex_unlock(&root->fs_info->trans_mutex);
2142 again:
2143 if (!err) {
2144 num_bytes = rc->merging_rsv_size;
2145 ret = btrfs_block_rsv_add(NULL, root, rc->block_rsv,
2146 num_bytes, &retries);
2147 if (ret)
2148 err = ret;
2149 }
2150
2151 trans = btrfs_join_transaction(rc->extent_root, 1);
2152
2153 if (!err) {
2154 if (num_bytes != rc->merging_rsv_size) {
2155 btrfs_end_transaction(trans, rc->extent_root);
2156 btrfs_block_rsv_release(rc->extent_root,
2157 rc->block_rsv, num_bytes);
2158 retries = 0;
2159 goto again;
2160 }
2161 }
2162
2163 rc->merge_reloc_tree = 1;
2164
2165 while (!list_empty(&rc->reloc_roots)) {
2166 reloc_root = list_entry(rc->reloc_roots.next,
2167 struct btrfs_root, root_list);
2168 list_del_init(&reloc_root->root_list);
2169
2170 root = read_fs_root(reloc_root->fs_info,
2171 reloc_root->root_key.offset);
2172 BUG_ON(IS_ERR(root));
2173 BUG_ON(root->reloc_root != reloc_root);
2174
2175 /*
2176 * set reference count to 1, so btrfs_recover_relocation
2177 * knows it should resumes merging
2178 */
2179 if (!err)
2180 btrfs_set_root_refs(&reloc_root->root_item, 1);
2181 btrfs_update_reloc_root(trans, root);
2182
2183 list_add(&reloc_root->root_list, &reloc_roots);
2184 }
2185
2186 list_splice(&reloc_roots, &rc->reloc_roots);
2187
2188 if (!err)
2189 btrfs_commit_transaction(trans, rc->extent_root);
2190 else
2191 btrfs_end_transaction(trans, rc->extent_root);
2192 return err;
2193 }
2194
2195 static noinline_for_stack
2196 int merge_reloc_roots(struct reloc_control *rc)
2197 {
2198 struct btrfs_root *root;
2199 struct btrfs_root *reloc_root;
2200 LIST_HEAD(reloc_roots);
2201 int found = 0;
2202 int ret;
2203 again:
2204 root = rc->extent_root;
2205 mutex_lock(&root->fs_info->trans_mutex);
2206 list_splice_init(&rc->reloc_roots, &reloc_roots);
2207 mutex_unlock(&root->fs_info->trans_mutex);
2208
2209 while (!list_empty(&reloc_roots)) {
2210 found = 1;
2211 reloc_root = list_entry(reloc_roots.next,
2212 struct btrfs_root, root_list);
2213
2214 if (btrfs_root_refs(&reloc_root->root_item) > 0) {
2215 root = read_fs_root(reloc_root->fs_info,
2216 reloc_root->root_key.offset);
2217 BUG_ON(IS_ERR(root));
2218 BUG_ON(root->reloc_root != reloc_root);
2219
2220 ret = merge_reloc_root(rc, root);
2221 BUG_ON(ret);
2222 } else {
2223 list_del_init(&reloc_root->root_list);
2224 }
2225 btrfs_drop_snapshot(reloc_root, rc->block_rsv, 0);
2226 }
2227
2228 if (found) {
2229 found = 0;
2230 goto again;
2231 }
2232 BUG_ON(!RB_EMPTY_ROOT(&rc->reloc_root_tree.rb_root));
2233 return 0;
2234 }
2235
2236 static void free_block_list(struct rb_root *blocks)
2237 {
2238 struct tree_block *block;
2239 struct rb_node *rb_node;
2240 while ((rb_node = rb_first(blocks))) {
2241 block = rb_entry(rb_node, struct tree_block, rb_node);
2242 rb_erase(rb_node, blocks);
2243 kfree(block);
2244 }
2245 }
2246
2247 static int record_reloc_root_in_trans(struct btrfs_trans_handle *trans,
2248 struct btrfs_root *reloc_root)
2249 {
2250 struct btrfs_root *root;
2251
2252 if (reloc_root->last_trans == trans->transid)
2253 return 0;
2254
2255 root = read_fs_root(reloc_root->fs_info, reloc_root->root_key.offset);
2256 BUG_ON(IS_ERR(root));
2257 BUG_ON(root->reloc_root != reloc_root);
2258
2259 return btrfs_record_root_in_trans(trans, root);
2260 }
2261
2262 static noinline_for_stack
2263 struct btrfs_root *select_reloc_root(struct btrfs_trans_handle *trans,
2264 struct reloc_control *rc,
2265 struct backref_node *node,
2266 struct backref_edge *edges[], int *nr)
2267 {
2268 struct backref_node *next;
2269 struct btrfs_root *root;
2270 int index = 0;
2271
2272 next = node;
2273 while (1) {
2274 cond_resched();
2275 next = walk_up_backref(next, edges, &index);
2276 root = next->root;
2277 BUG_ON(!root);
2278 BUG_ON(!root->ref_cows);
2279
2280 if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID) {
2281 record_reloc_root_in_trans(trans, root);
2282 break;
2283 }
2284
2285 btrfs_record_root_in_trans(trans, root);
2286 root = root->reloc_root;
2287
2288 if (next->new_bytenr != root->node->start) {
2289 BUG_ON(next->new_bytenr);
2290 BUG_ON(!list_empty(&next->list));
2291 next->new_bytenr = root->node->start;
2292 next->root = root;
2293 list_add_tail(&next->list,
2294 &rc->backref_cache.changed);
2295 __mark_block_processed(rc, next);
2296 break;
2297 }
2298
2299 WARN_ON(1);
2300 root = NULL;
2301 next = walk_down_backref(edges, &index);
2302 if (!next || next->level <= node->level)
2303 break;
2304 }
2305 if (!root)
2306 return NULL;
2307
2308 *nr = index;
2309 next = node;
2310 /* setup backref node path for btrfs_reloc_cow_block */
2311 while (1) {
2312 rc->backref_cache.path[next->level] = next;
2313 if (--index < 0)
2314 break;
2315 next = edges[index]->node[UPPER];
2316 }
2317 return root;
2318 }
2319
2320 /*
2321 * select a tree root for relocation. return NULL if the block
2322 * is reference counted. we should use do_relocation() in this
2323 * case. return a tree root pointer if the block isn't reference
2324 * counted. return -ENOENT if the block is root of reloc tree.
2325 */
2326 static noinline_for_stack
2327 struct btrfs_root *select_one_root(struct btrfs_trans_handle *trans,
2328 struct backref_node *node)
2329 {
2330 struct backref_node *next;
2331 struct btrfs_root *root;
2332 struct btrfs_root *fs_root = NULL;
2333 struct backref_edge *edges[BTRFS_MAX_LEVEL - 1];
2334 int index = 0;
2335
2336 next = node;
2337 while (1) {
2338 cond_resched();
2339 next = walk_up_backref(next, edges, &index);
2340 root = next->root;
2341 BUG_ON(!root);
2342
2343 /* no other choice for non-refernce counted tree */
2344 if (!root->ref_cows)
2345 return root;
2346
2347 if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID)
2348 fs_root = root;
2349
2350 if (next != node)
2351 return NULL;
2352
2353 next = walk_down_backref(edges, &index);
2354 if (!next || next->level <= node->level)
2355 break;
2356 }
2357
2358 if (!fs_root)
2359 return ERR_PTR(-ENOENT);
2360 return fs_root;
2361 }
2362
2363 static noinline_for_stack
2364 u64 calcu_metadata_size(struct reloc_control *rc,
2365 struct backref_node *node, int reserve)
2366 {
2367 struct backref_node *next = node;
2368 struct backref_edge *edge;
2369 struct backref_edge *edges[BTRFS_MAX_LEVEL - 1];
2370 u64 num_bytes = 0;
2371 int index = 0;
2372
2373 BUG_ON(reserve && node->processed);
2374
2375 while (next) {
2376 cond_resched();
2377 while (1) {
2378 if (next->processed && (reserve || next != node))
2379 break;
2380
2381 num_bytes += btrfs_level_size(rc->extent_root,
2382 next->level);
2383
2384 if (list_empty(&next->upper))
2385 break;
2386
2387 edge = list_entry(next->upper.next,
2388 struct backref_edge, list[LOWER]);
2389 edges[index++] = edge;
2390 next = edge->node[UPPER];
2391 }
2392 next = walk_down_backref(edges, &index);
2393 }
2394 return num_bytes;
2395 }
2396
2397 static int reserve_metadata_space(struct btrfs_trans_handle *trans,
2398 struct reloc_control *rc,
2399 struct backref_node *node)
2400 {
2401 struct btrfs_root *root = rc->extent_root;
2402 u64 num_bytes;
2403 int ret;
2404
2405 num_bytes = calcu_metadata_size(rc, node, 1) * 2;
2406
2407 trans->block_rsv = rc->block_rsv;
2408 ret = btrfs_block_rsv_add(trans, root, rc->block_rsv, num_bytes,
2409 &rc->block_rsv_retries);
2410 if (ret) {
2411 if (ret == -EAGAIN)
2412 rc->commit_transaction = 1;
2413 return ret;
2414 }
2415
2416 rc->block_rsv_retries = 0;
2417 return 0;
2418 }
2419
2420 static void release_metadata_space(struct reloc_control *rc,
2421 struct backref_node *node)
2422 {
2423 u64 num_bytes = calcu_metadata_size(rc, node, 0) * 2;
2424 btrfs_block_rsv_release(rc->extent_root, rc->block_rsv, num_bytes);
2425 }
2426
2427 /*
2428 * relocate a block tree, and then update pointers in upper level
2429 * blocks that reference the block to point to the new location.
2430 *
2431 * if called by link_to_upper, the block has already been relocated.
2432 * in that case this function just updates pointers.
2433 */
2434 static int do_relocation(struct btrfs_trans_handle *trans,
2435 struct reloc_control *rc,
2436 struct backref_node *node,
2437 struct btrfs_key *key,
2438 struct btrfs_path *path, int lowest)
2439 {
2440 struct backref_node *upper;
2441 struct backref_edge *edge;
2442 struct backref_edge *edges[BTRFS_MAX_LEVEL - 1];
2443 struct btrfs_root *root;
2444 struct extent_buffer *eb;
2445 u32 blocksize;
2446 u64 bytenr;
2447 u64 generation;
2448 int nr;
2449 int slot;
2450 int ret;
2451 int err = 0;
2452
2453 BUG_ON(lowest && node->eb);
2454
2455 path->lowest_level = node->level + 1;
2456 rc->backref_cache.path[node->level] = node;
2457 list_for_each_entry(edge, &node->upper, list[LOWER]) {
2458 cond_resched();
2459
2460 upper = edge->node[UPPER];
2461 root = select_reloc_root(trans, rc, upper, edges, &nr);
2462 BUG_ON(!root);
2463
2464 if (upper->eb && !upper->locked) {
2465 if (!lowest) {
2466 ret = btrfs_bin_search(upper->eb, key,
2467 upper->level, &slot);
2468 BUG_ON(ret);
2469 bytenr = btrfs_node_blockptr(upper->eb, slot);
2470 if (node->eb->start == bytenr)
2471 goto next;
2472 }
2473 drop_node_buffer(upper);
2474 }
2475
2476 if (!upper->eb) {
2477 ret = btrfs_search_slot(trans, root, key, path, 0, 1);
2478 if (ret < 0) {
2479 err = ret;
2480 break;
2481 }
2482 BUG_ON(ret > 0);
2483
2484 if (!upper->eb) {
2485 upper->eb = path->nodes[upper->level];
2486 path->nodes[upper->level] = NULL;
2487 } else {
2488 BUG_ON(upper->eb != path->nodes[upper->level]);
2489 }
2490
2491 upper->locked = 1;
2492 path->locks[upper->level] = 0;
2493
2494 slot = path->slots[upper->level];
2495 btrfs_release_path(NULL, path);
2496 } else {
2497 ret = btrfs_bin_search(upper->eb, key, upper->level,
2498 &slot);
2499 BUG_ON(ret);
2500 }
2501
2502 bytenr = btrfs_node_blockptr(upper->eb, slot);
2503 if (lowest) {
2504 BUG_ON(bytenr != node->bytenr);
2505 } else {
2506 if (node->eb->start == bytenr)
2507 goto next;
2508 }
2509
2510 blocksize = btrfs_level_size(root, node->level);
2511 generation = btrfs_node_ptr_generation(upper->eb, slot);
2512 eb = read_tree_block(root, bytenr, blocksize, generation);
2513 btrfs_tree_lock(eb);
2514 btrfs_set_lock_blocking(eb);
2515
2516 if (!node->eb) {
2517 ret = btrfs_cow_block(trans, root, eb, upper->eb,
2518 slot, &eb);
2519 btrfs_tree_unlock(eb);
2520 free_extent_buffer(eb);
2521 if (ret < 0) {
2522 err = ret;
2523 goto next;
2524 }
2525 BUG_ON(node->eb != eb);
2526 } else {
2527 btrfs_set_node_blockptr(upper->eb, slot,
2528 node->eb->start);
2529 btrfs_set_node_ptr_generation(upper->eb, slot,
2530 trans->transid);
2531 btrfs_mark_buffer_dirty(upper->eb);
2532
2533 ret = btrfs_inc_extent_ref(trans, root,
2534 node->eb->start, blocksize,
2535 upper->eb->start,
2536 btrfs_header_owner(upper->eb),
2537 node->level, 0);
2538 BUG_ON(ret);
2539
2540 ret = btrfs_drop_subtree(trans, root, eb, upper->eb);
2541 BUG_ON(ret);
2542 }
2543 next:
2544 if (!upper->pending)
2545 drop_node_buffer(upper);
2546 else
2547 unlock_node_buffer(upper);
2548 if (err)
2549 break;
2550 }
2551
2552 if (!err && node->pending) {
2553 drop_node_buffer(node);
2554 list_move_tail(&node->list, &rc->backref_cache.changed);
2555 node->pending = 0;
2556 }
2557
2558 path->lowest_level = 0;
2559 BUG_ON(err == -ENOSPC);
2560 return err;
2561 }
2562
2563 static int link_to_upper(struct btrfs_trans_handle *trans,
2564 struct reloc_control *rc,
2565 struct backref_node *node,
2566 struct btrfs_path *path)
2567 {
2568 struct btrfs_key key;
2569
2570 btrfs_node_key_to_cpu(node->eb, &key, 0);
2571 return do_relocation(trans, rc, node, &key, path, 0);
2572 }
2573
2574 static int finish_pending_nodes(struct btrfs_trans_handle *trans,
2575 struct reloc_control *rc,
2576 struct btrfs_path *path, int err)
2577 {
2578 LIST_HEAD(list);
2579 struct backref_cache *cache = &rc->backref_cache;
2580 struct backref_node *node;
2581 int level;
2582 int ret;
2583
2584 for (level = 0; level < BTRFS_MAX_LEVEL; level++) {
2585 while (!list_empty(&cache->pending[level])) {
2586 node = list_entry(cache->pending[level].next,
2587 struct backref_node, list);
2588 list_move_tail(&node->list, &list);
2589 BUG_ON(!node->pending);
2590
2591 if (!err) {
2592 ret = link_to_upper(trans, rc, node, path);
2593 if (ret < 0)
2594 err = ret;
2595 }
2596 }
2597 list_splice_init(&list, &cache->pending[level]);
2598 }
2599 return err;
2600 }
2601
2602 static void mark_block_processed(struct reloc_control *rc,
2603 u64 bytenr, u32 blocksize)
2604 {
2605 set_extent_bits(&rc->processed_blocks, bytenr, bytenr + blocksize - 1,
2606 EXTENT_DIRTY, GFP_NOFS);
2607 }
2608
2609 static void __mark_block_processed(struct reloc_control *rc,
2610 struct backref_node *node)
2611 {
2612 u32 blocksize;
2613 if (node->level == 0 ||
2614 in_block_group(node->bytenr, rc->block_group)) {
2615 blocksize = btrfs_level_size(rc->extent_root, node->level);
2616 mark_block_processed(rc, node->bytenr, blocksize);
2617 }
2618 node->processed = 1;
2619 }
2620
2621 /*
2622 * mark a block and all blocks directly/indirectly reference the block
2623 * as processed.
2624 */
2625 static void update_processed_blocks(struct reloc_control *rc,
2626 struct backref_node *node)
2627 {
2628 struct backref_node *next = node;
2629 struct backref_edge *edge;
2630 struct backref_edge *edges[BTRFS_MAX_LEVEL - 1];
2631 int index = 0;
2632
2633 while (next) {
2634 cond_resched();
2635 while (1) {
2636 if (next->processed)
2637 break;
2638
2639 __mark_block_processed(rc, next);
2640
2641 if (list_empty(&next->upper))
2642 break;
2643
2644 edge = list_entry(next->upper.next,
2645 struct backref_edge, list[LOWER]);
2646 edges[index++] = edge;
2647 next = edge->node[UPPER];
2648 }
2649 next = walk_down_backref(edges, &index);
2650 }
2651 }
2652
2653 static int tree_block_processed(u64 bytenr, u32 blocksize,
2654 struct reloc_control *rc)
2655 {
2656 if (test_range_bit(&rc->processed_blocks, bytenr,
2657 bytenr + blocksize - 1, EXTENT_DIRTY, 1, NULL))
2658 return 1;
2659 return 0;
2660 }
2661
2662 static int get_tree_block_key(struct reloc_control *rc,
2663 struct tree_block *block)
2664 {
2665 struct extent_buffer *eb;
2666
2667 BUG_ON(block->key_ready);
2668 eb = read_tree_block(rc->extent_root, block->bytenr,
2669 block->key.objectid, block->key.offset);
2670 WARN_ON(btrfs_header_level(eb) != block->level);
2671 if (block->level == 0)
2672 btrfs_item_key_to_cpu(eb, &block->key, 0);
2673 else
2674 btrfs_node_key_to_cpu(eb, &block->key, 0);
2675 free_extent_buffer(eb);
2676 block->key_ready = 1;
2677 return 0;
2678 }
2679
2680 static int reada_tree_block(struct reloc_control *rc,
2681 struct tree_block *block)
2682 {
2683 BUG_ON(block->key_ready);
2684 readahead_tree_block(rc->extent_root, block->bytenr,
2685 block->key.objectid, block->key.offset);
2686 return 0;
2687 }
2688
2689 /*
2690 * helper function to relocate a tree block
2691 */
2692 static int relocate_tree_block(struct btrfs_trans_handle *trans,
2693 struct reloc_control *rc,
2694 struct backref_node *node,
2695 struct btrfs_key *key,
2696 struct btrfs_path *path)
2697 {
2698 struct btrfs_root *root;
2699 int release = 0;
2700 int ret = 0;
2701
2702 if (!node)
2703 return 0;
2704
2705 BUG_ON(node->processed);
2706 root = select_one_root(trans, node);
2707 if (root == ERR_PTR(-ENOENT)) {
2708 update_processed_blocks(rc, node);
2709 goto out;
2710 }
2711
2712 if (!root || root->ref_cows) {
2713 ret = reserve_metadata_space(trans, rc, node);
2714 if (ret)
2715 goto out;
2716 release = 1;
2717 }
2718
2719 if (root) {
2720 if (root->ref_cows) {
2721 BUG_ON(node->new_bytenr);
2722 BUG_ON(!list_empty(&node->list));
2723 btrfs_record_root_in_trans(trans, root);
2724 root = root->reloc_root;
2725 node->new_bytenr = root->node->start;
2726 node->root = root;
2727 list_add_tail(&node->list, &rc->backref_cache.changed);
2728 } else {
2729 path->lowest_level = node->level;
2730 ret = btrfs_search_slot(trans, root, key, path, 0, 1);
2731 btrfs_release_path(root, path);
2732 if (ret > 0)
2733 ret = 0;
2734 }
2735 if (!ret)
2736 update_processed_blocks(rc, node);
2737 } else {
2738 ret = do_relocation(trans, rc, node, key, path, 1);
2739 }
2740 out:
2741 if (ret || node->level == 0 || node->cowonly) {
2742 if (release)
2743 release_metadata_space(rc, node);
2744 remove_backref_node(&rc->backref_cache, node);
2745 }
2746 return ret;
2747 }
2748
2749 /*
2750 * relocate a list of blocks
2751 */
2752 static noinline_for_stack
2753 int relocate_tree_blocks(struct btrfs_trans_handle *trans,
2754 struct reloc_control *rc, struct rb_root *blocks)
2755 {
2756 struct backref_node *node;
2757 struct btrfs_path *path;
2758 struct tree_block *block;
2759 struct rb_node *rb_node;
2760 int ret;
2761 int err = 0;
2762
2763 path = btrfs_alloc_path();
2764 if (!path)
2765 return -ENOMEM;
2766
2767 rb_node = rb_first(blocks);
2768 while (rb_node) {
2769 block = rb_entry(rb_node, struct tree_block, rb_node);
2770 if (!block->key_ready)
2771 reada_tree_block(rc, block);
2772 rb_node = rb_next(rb_node);
2773 }
2774
2775 rb_node = rb_first(blocks);
2776 while (rb_node) {
2777 block = rb_entry(rb_node, struct tree_block, rb_node);
2778 if (!block->key_ready)
2779 get_tree_block_key(rc, block);
2780 rb_node = rb_next(rb_node);
2781 }
2782
2783 rb_node = rb_first(blocks);
2784 while (rb_node) {
2785 block = rb_entry(rb_node, struct tree_block, rb_node);
2786
2787 node = build_backref_tree(rc, &block->key,
2788 block->level, block->bytenr);
2789 if (IS_ERR(node)) {
2790 err = PTR_ERR(node);
2791 goto out;
2792 }
2793
2794 ret = relocate_tree_block(trans, rc, node, &block->key,
2795 path);
2796 if (ret < 0) {
2797 if (ret != -EAGAIN || rb_node == rb_first(blocks))
2798 err = ret;
2799 goto out;
2800 }
2801 rb_node = rb_next(rb_node);
2802 }
2803 out:
2804 free_block_list(blocks);
2805 err = finish_pending_nodes(trans, rc, path, err);
2806
2807 btrfs_free_path(path);
2808 return err;
2809 }
2810
2811 static noinline_for_stack
2812 int prealloc_file_extent_cluster(struct inode *inode,
2813 struct file_extent_cluster *cluster)
2814 {
2815 u64 alloc_hint = 0;
2816 u64 start;
2817 u64 end;
2818 u64 offset = BTRFS_I(inode)->index_cnt;
2819 u64 num_bytes;
2820 int nr = 0;
2821 int ret = 0;
2822
2823 BUG_ON(cluster->start != cluster->boundary[0]);
2824 mutex_lock(&inode->i_mutex);
2825
2826 ret = btrfs_check_data_free_space(inode, cluster->end +
2827 1 - cluster->start);
2828 if (ret)
2829 goto out;
2830
2831 while (nr < cluster->nr) {
2832 start = cluster->boundary[nr] - offset;
2833 if (nr + 1 < cluster->nr)
2834 end = cluster->boundary[nr + 1] - 1 - offset;
2835 else
2836 end = cluster->end - offset;
2837
2838 lock_extent(&BTRFS_I(inode)->io_tree, start, end, GFP_NOFS);
2839 num_bytes = end + 1 - start;
2840 ret = btrfs_prealloc_file_range(inode, 0, start,
2841 num_bytes, num_bytes,
2842 end + 1, &alloc_hint);
2843 unlock_extent(&BTRFS_I(inode)->io_tree, start, end, GFP_NOFS);
2844 if (ret)
2845 break;
2846 nr++;
2847 }
2848 btrfs_free_reserved_data_space(inode, cluster->end +
2849 1 - cluster->start);
2850 out:
2851 mutex_unlock(&inode->i_mutex);
2852 return ret;
2853 }
2854
2855 static noinline_for_stack
2856 int setup_extent_mapping(struct inode *inode, u64 start, u64 end,
2857 u64 block_start)
2858 {
2859 struct btrfs_root *root = BTRFS_I(inode)->root;
2860 struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
2861 struct extent_map *em;
2862 int ret = 0;
2863
2864 em = alloc_extent_map(GFP_NOFS);
2865 if (!em)
2866 return -ENOMEM;
2867
2868 em->start = start;
2869 em->len = end + 1 - start;
2870 em->block_len = em->len;
2871 em->block_start = block_start;
2872 em->bdev = root->fs_info->fs_devices->latest_bdev;
2873 set_bit(EXTENT_FLAG_PINNED, &em->flags);
2874
2875 lock_extent(&BTRFS_I(inode)->io_tree, start, end, GFP_NOFS);
2876 while (1) {
2877 write_lock(&em_tree->lock);
2878 ret = add_extent_mapping(em_tree, em);
2879 write_unlock(&em_tree->lock);
2880 if (ret != -EEXIST) {
2881 free_extent_map(em);
2882 break;
2883 }
2884 btrfs_drop_extent_cache(inode, start, end, 0);
2885 }
2886 unlock_extent(&BTRFS_I(inode)->io_tree, start, end, GFP_NOFS);
2887 return ret;
2888 }
2889
2890 static int relocate_file_extent_cluster(struct inode *inode,
2891 struct file_extent_cluster *cluster)
2892 {
2893 u64 page_start;
2894 u64 page_end;
2895 u64 offset = BTRFS_I(inode)->index_cnt;
2896 unsigned long index;
2897 unsigned long last_index;
2898 struct page *page;
2899 struct file_ra_state *ra;
2900 int nr = 0;
2901 int ret = 0;
2902
2903 if (!cluster->nr)
2904 return 0;
2905
2906 ra = kzalloc(sizeof(*ra), GFP_NOFS);
2907 if (!ra)
2908 return -ENOMEM;
2909
2910 ret = prealloc_file_extent_cluster(inode, cluster);
2911 if (ret)
2912 goto out;
2913
2914 file_ra_state_init(ra, inode->i_mapping);
2915
2916 ret = setup_extent_mapping(inode, cluster->start - offset,
2917 cluster->end - offset, cluster->start);
2918 if (ret)
2919 goto out;
2920
2921 index = (cluster->start - offset) >> PAGE_CACHE_SHIFT;
2922 last_index = (cluster->end - offset) >> PAGE_CACHE_SHIFT;
2923 while (index <= last_index) {
2924 ret = btrfs_delalloc_reserve_metadata(inode, PAGE_CACHE_SIZE);
2925 if (ret)
2926 goto out;
2927
2928 page = find_lock_page(inode->i_mapping, index);
2929 if (!page) {
2930 page_cache_sync_readahead(inode->i_mapping,
2931 ra, NULL, index,
2932 last_index + 1 - index);
2933 page = grab_cache_page(inode->i_mapping, index);
2934 if (!page) {
2935 btrfs_delalloc_release_metadata(inode,
2936 PAGE_CACHE_SIZE);
2937 ret = -ENOMEM;
2938 goto out;
2939 }
2940 }
2941
2942 if (PageReadahead(page)) {
2943 page_cache_async_readahead(inode->i_mapping,
2944 ra, NULL, page, index,
2945 last_index + 1 - index);
2946 }
2947
2948 if (!PageUptodate(page)) {
2949 btrfs_readpage(NULL, page);
2950 lock_page(page);
2951 if (!PageUptodate(page)) {
2952 unlock_page(page);
2953 page_cache_release(page);
2954 btrfs_delalloc_release_metadata(inode,
2955 PAGE_CACHE_SIZE);
2956 ret = -EIO;
2957 goto out;
2958 }
2959 }
2960
2961 page_start = (u64)page->index << PAGE_CACHE_SHIFT;
2962 page_end = page_start + PAGE_CACHE_SIZE - 1;
2963
2964 lock_extent(&BTRFS_I(inode)->io_tree,
2965 page_start, page_end, GFP_NOFS);
2966
2967 set_page_extent_mapped(page);
2968
2969 if (nr < cluster->nr &&
2970 page_start + offset == cluster->boundary[nr]) {
2971 set_extent_bits(&BTRFS_I(inode)->io_tree,
2972 page_start, page_end,
2973 EXTENT_BOUNDARY, GFP_NOFS);
2974 nr++;
2975 }
2976
2977 btrfs_set_extent_delalloc(inode, page_start, page_end, NULL);
2978 set_page_dirty(page);
2979
2980 unlock_extent(&BTRFS_I(inode)->io_tree,
2981 page_start, page_end, GFP_NOFS);
2982 unlock_page(page);
2983 page_cache_release(page);
2984
2985 index++;
2986 balance_dirty_pages_ratelimited(inode->i_mapping);
2987 btrfs_throttle(BTRFS_I(inode)->root);
2988 }
2989 WARN_ON(nr != cluster->nr);
2990 out:
2991 kfree(ra);
2992 return ret;
2993 }
2994
2995 static noinline_for_stack
2996 int relocate_data_extent(struct inode *inode, struct btrfs_key *extent_key,
2997 struct file_extent_cluster *cluster)
2998 {
2999 int ret;
3000
3001 if (cluster->nr > 0 && extent_key->objectid != cluster->end + 1) {
3002 ret = relocate_file_extent_cluster(inode, cluster);
3003 if (ret)
3004 return ret;
3005 cluster->nr = 0;
3006 }
3007
3008 if (!cluster->nr)
3009 cluster->start = extent_key->objectid;
3010 else
3011 BUG_ON(cluster->nr >= MAX_EXTENTS);
3012 cluster->end = extent_key->objectid + extent_key->offset - 1;
3013 cluster->boundary[cluster->nr] = extent_key->objectid;
3014 cluster->nr++;
3015
3016 if (cluster->nr >= MAX_EXTENTS) {
3017 ret = relocate_file_extent_cluster(inode, cluster);
3018 if (ret)
3019 return ret;
3020 cluster->nr = 0;
3021 }
3022 return 0;
3023 }
3024
3025 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
3026 static int get_ref_objectid_v0(struct reloc_control *rc,
3027 struct btrfs_path *path,
3028 struct btrfs_key *extent_key,
3029 u64 *ref_objectid, int *path_change)
3030 {
3031 struct btrfs_key key;
3032 struct extent_buffer *leaf;
3033 struct btrfs_extent_ref_v0 *ref0;
3034 int ret;
3035 int slot;
3036
3037 leaf = path->nodes[0];
3038 slot = path->slots[0];
3039 while (1) {
3040 if (slot >= btrfs_header_nritems(leaf)) {
3041 ret = btrfs_next_leaf(rc->extent_root, path);
3042 if (ret < 0)
3043 return ret;
3044 BUG_ON(ret > 0);
3045 leaf = path->nodes[0];
3046 slot = path->slots[0];
3047 if (path_change)
3048 *path_change = 1;
3049 }
3050 btrfs_item_key_to_cpu(leaf, &key, slot);
3051 if (key.objectid != extent_key->objectid)
3052 return -ENOENT;
3053
3054 if (key.type != BTRFS_EXTENT_REF_V0_KEY) {
3055 slot++;
3056 continue;
3057 }
3058 ref0 = btrfs_item_ptr(leaf, slot,
3059 struct btrfs_extent_ref_v0);
3060 *ref_objectid = btrfs_ref_objectid_v0(leaf, ref0);
3061 break;
3062 }
3063 return 0;
3064 }
3065 #endif
3066
3067 /*
3068 * helper to add a tree block to the list.
3069 * the major work is getting the generation and level of the block
3070 */
3071 static int add_tree_block(struct reloc_control *rc,
3072 struct btrfs_key *extent_key,
3073 struct btrfs_path *path,
3074 struct rb_root *blocks)
3075 {
3076 struct extent_buffer *eb;
3077 struct btrfs_extent_item *ei;
3078 struct btrfs_tree_block_info *bi;
3079 struct tree_block *block;
3080 struct rb_node *rb_node;
3081 u32 item_size;
3082 int level = -1;
3083 int generation;
3084
3085 eb = path->nodes[0];
3086 item_size = btrfs_item_size_nr(eb, path->slots[0]);
3087
3088 if (item_size >= sizeof(*ei) + sizeof(*bi)) {
3089 ei = btrfs_item_ptr(eb, path->slots[0],
3090 struct btrfs_extent_item);
3091 bi = (struct btrfs_tree_block_info *)(ei + 1);
3092 generation = btrfs_extent_generation(eb, ei);
3093 level = btrfs_tree_block_level(eb, bi);
3094 } else {
3095 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
3096 u64 ref_owner;
3097 int ret;
3098
3099 BUG_ON(item_size != sizeof(struct btrfs_extent_item_v0));
3100 ret = get_ref_objectid_v0(rc, path, extent_key,
3101 &ref_owner, NULL);
3102 BUG_ON(ref_owner >= BTRFS_MAX_LEVEL);
3103 level = (int)ref_owner;
3104 /* FIXME: get real generation */
3105 generation = 0;
3106 #else
3107 BUG();
3108 #endif
3109 }
3110
3111 btrfs_release_path(rc->extent_root, path);
3112
3113 BUG_ON(level == -1);
3114
3115 block = kmalloc(sizeof(*block), GFP_NOFS);
3116 if (!block)
3117 return -ENOMEM;
3118
3119 block->bytenr = extent_key->objectid;
3120 block->key.objectid = extent_key->offset;
3121 block->key.offset = generation;
3122 block->level = level;
3123 block->key_ready = 0;
3124
3125 rb_node = tree_insert(blocks, block->bytenr, &block->rb_node);
3126 BUG_ON(rb_node);
3127
3128 return 0;
3129 }
3130
3131 /*
3132 * helper to add tree blocks for backref of type BTRFS_SHARED_DATA_REF_KEY
3133 */
3134 static int __add_tree_block(struct reloc_control *rc,
3135 u64 bytenr, u32 blocksize,
3136 struct rb_root *blocks)
3137 {
3138 struct btrfs_path *path;
3139 struct btrfs_key key;
3140 int ret;
3141
3142 if (tree_block_processed(bytenr, blocksize, rc))
3143 return 0;
3144
3145 if (tree_search(blocks, bytenr))
3146 return 0;
3147
3148 path = btrfs_alloc_path();
3149 if (!path)
3150 return -ENOMEM;
3151
3152 key.objectid = bytenr;
3153 key.type = BTRFS_EXTENT_ITEM_KEY;
3154 key.offset = blocksize;
3155
3156 path->search_commit_root = 1;
3157 path->skip_locking = 1;
3158 ret = btrfs_search_slot(NULL, rc->extent_root, &key, path, 0, 0);
3159 if (ret < 0)
3160 goto out;
3161 BUG_ON(ret);
3162
3163 btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0]);
3164 ret = add_tree_block(rc, &key, path, blocks);
3165 out:
3166 btrfs_free_path(path);
3167 return ret;
3168 }
3169
3170 /*
3171 * helper to check if the block use full backrefs for pointers in it
3172 */
3173 static int block_use_full_backref(struct reloc_control *rc,
3174 struct extent_buffer *eb)
3175 {
3176 u64 flags;
3177 int ret;
3178
3179 if (btrfs_header_flag(eb, BTRFS_HEADER_FLAG_RELOC) ||
3180 btrfs_header_backref_rev(eb) < BTRFS_MIXED_BACKREF_REV)
3181 return 1;
3182
3183 ret = btrfs_lookup_extent_info(NULL, rc->extent_root,
3184 eb->start, eb->len, NULL, &flags);
3185 BUG_ON(ret);
3186
3187 if (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF)
3188 ret = 1;
3189 else
3190 ret = 0;
3191 return ret;
3192 }
3193
3194 /*
3195 * helper to add tree blocks for backref of type BTRFS_EXTENT_DATA_REF_KEY
3196 * this function scans fs tree to find blocks reference the data extent
3197 */
3198 static int find_data_references(struct reloc_control *rc,
3199 struct btrfs_key *extent_key,
3200 struct extent_buffer *leaf,
3201 struct btrfs_extent_data_ref *ref,
3202 struct rb_root *blocks)
3203 {
3204 struct btrfs_path *path;
3205 struct tree_block *block;
3206 struct btrfs_root *root;
3207 struct btrfs_file_extent_item *fi;
3208 struct rb_node *rb_node;
3209 struct btrfs_key key;
3210 u64 ref_root;
3211 u64 ref_objectid;
3212 u64 ref_offset;
3213 u32 ref_count;
3214 u32 nritems;
3215 int err = 0;
3216 int added = 0;
3217 int counted;
3218 int ret;
3219
3220 path = btrfs_alloc_path();
3221 if (!path)
3222 return -ENOMEM;
3223
3224 ref_root = btrfs_extent_data_ref_root(leaf, ref);
3225 ref_objectid = btrfs_extent_data_ref_objectid(leaf, ref);
3226 ref_offset = btrfs_extent_data_ref_offset(leaf, ref);
3227 ref_count = btrfs_extent_data_ref_count(leaf, ref);
3228
3229 root = read_fs_root(rc->extent_root->fs_info, ref_root);
3230 if (IS_ERR(root)) {
3231 err = PTR_ERR(root);
3232 goto out;
3233 }
3234
3235 key.objectid = ref_objectid;
3236 key.offset = ref_offset;
3237 key.type = BTRFS_EXTENT_DATA_KEY;
3238
3239 path->search_commit_root = 1;
3240 path->skip_locking = 1;
3241 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
3242 if (ret < 0) {
3243 err = ret;
3244 goto out;
3245 }
3246
3247 leaf = path->nodes[0];
3248 nritems = btrfs_header_nritems(leaf);
3249 /*
3250 * the references in tree blocks that use full backrefs
3251 * are not counted in
3252 */
3253 if (block_use_full_backref(rc, leaf))
3254 counted = 0;
3255 else
3256 counted = 1;
3257 rb_node = tree_search(blocks, leaf->start);
3258 if (rb_node) {
3259 if (counted)
3260 added = 1;
3261 else
3262 path->slots[0] = nritems;
3263 }
3264
3265 while (ref_count > 0) {
3266 while (path->slots[0] >= nritems) {
3267 ret = btrfs_next_leaf(root, path);
3268 if (ret < 0) {
3269 err = ret;
3270 goto out;
3271 }
3272 if (ret > 0) {
3273 WARN_ON(1);
3274 goto out;
3275 }
3276
3277 leaf = path->nodes[0];
3278 nritems = btrfs_header_nritems(leaf);
3279 added = 0;
3280
3281 if (block_use_full_backref(rc, leaf))
3282 counted = 0;
3283 else
3284 counted = 1;
3285 rb_node = tree_search(blocks, leaf->start);
3286 if (rb_node) {
3287 if (counted)
3288 added = 1;
3289 else
3290 path->slots[0] = nritems;
3291 }
3292 }
3293
3294 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
3295 if (key.objectid != ref_objectid ||
3296 key.type != BTRFS_EXTENT_DATA_KEY) {
3297 WARN_ON(1);
3298 break;
3299 }
3300
3301 fi = btrfs_item_ptr(leaf, path->slots[0],
3302 struct btrfs_file_extent_item);
3303
3304 if (btrfs_file_extent_type(leaf, fi) ==
3305 BTRFS_FILE_EXTENT_INLINE)
3306 goto next;
3307
3308 if (btrfs_file_extent_disk_bytenr(leaf, fi) !=
3309 extent_key->objectid)
3310 goto next;
3311
3312 key.offset -= btrfs_file_extent_offset(leaf, fi);
3313 if (key.offset != ref_offset)
3314 goto next;
3315
3316 if (counted)
3317 ref_count--;
3318 if (added)
3319 goto next;
3320
3321 if (!tree_block_processed(leaf->start, leaf->len, rc)) {
3322 block = kmalloc(sizeof(*block), GFP_NOFS);
3323 if (!block) {
3324 err = -ENOMEM;
3325 break;
3326 }
3327 block->bytenr = leaf->start;
3328 btrfs_item_key_to_cpu(leaf, &block->key, 0);
3329 block->level = 0;
3330 block->key_ready = 1;
3331 rb_node = tree_insert(blocks, block->bytenr,
3332 &block->rb_node);
3333 BUG_ON(rb_node);
3334 }
3335 if (counted)
3336 added = 1;
3337 else
3338 path->slots[0] = nritems;
3339 next:
3340 path->slots[0]++;
3341
3342 }
3343 out:
3344 btrfs_free_path(path);
3345 return err;
3346 }
3347
3348 /*
3349 * hepler to find all tree blocks that reference a given data extent
3350 */
3351 static noinline_for_stack
3352 int add_data_references(struct reloc_control *rc,
3353 struct btrfs_key *extent_key,
3354 struct btrfs_path *path,
3355 struct rb_root *blocks)
3356 {
3357 struct btrfs_key key;
3358 struct extent_buffer *eb;
3359 struct btrfs_extent_data_ref *dref;
3360 struct btrfs_extent_inline_ref *iref;
3361 unsigned long ptr;
3362 unsigned long end;
3363 u32 blocksize = btrfs_level_size(rc->extent_root, 0);
3364 int ret;
3365 int err = 0;
3366
3367 eb = path->nodes[0];
3368 ptr = btrfs_item_ptr_offset(eb, path->slots[0]);
3369 end = ptr + btrfs_item_size_nr(eb, path->slots[0]);
3370 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
3371 if (ptr + sizeof(struct btrfs_extent_item_v0) == end)
3372 ptr = end;
3373 else
3374 #endif
3375 ptr += sizeof(struct btrfs_extent_item);
3376
3377 while (ptr < end) {
3378 iref = (struct btrfs_extent_inline_ref *)ptr;
3379 key.type = btrfs_extent_inline_ref_type(eb, iref);
3380 if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
3381 key.offset = btrfs_extent_inline_ref_offset(eb, iref);
3382 ret = __add_tree_block(rc, key.offset, blocksize,
3383 blocks);
3384 } else if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
3385 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
3386 ret = find_data_references(rc, extent_key,
3387 eb, dref, blocks);
3388 } else {
3389 BUG();
3390 }
3391 ptr += btrfs_extent_inline_ref_size(key.type);
3392 }
3393 WARN_ON(ptr > end);
3394
3395 while (1) {
3396 cond_resched();
3397 eb = path->nodes[0];
3398 if (path->slots[0] >= btrfs_header_nritems(eb)) {
3399 ret = btrfs_next_leaf(rc->extent_root, path);
3400 if (ret < 0) {
3401 err = ret;
3402 break;
3403 }
3404 if (ret > 0)
3405 break;
3406 eb = path->nodes[0];
3407 }
3408
3409 btrfs_item_key_to_cpu(eb, &key, path->slots[0]);
3410 if (key.objectid != extent_key->objectid)
3411 break;
3412
3413 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
3414 if (key.type == BTRFS_SHARED_DATA_REF_KEY ||
3415 key.type == BTRFS_EXTENT_REF_V0_KEY) {
3416 #else
3417 BUG_ON(key.type == BTRFS_EXTENT_REF_V0_KEY);
3418 if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
3419 #endif
3420 ret = __add_tree_block(rc, key.offset, blocksize,
3421 blocks);
3422 } else if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
3423 dref = btrfs_item_ptr(eb, path->slots[0],
3424 struct btrfs_extent_data_ref);
3425 ret = find_data_references(rc, extent_key,
3426 eb, dref, blocks);
3427 } else {
3428 ret = 0;
3429 }
3430 if (ret) {
3431 err = ret;
3432 break;
3433 }
3434 path->slots[0]++;
3435 }
3436 btrfs_release_path(rc->extent_root, path);
3437 if (err)
3438 free_block_list(blocks);
3439 return err;
3440 }
3441
3442 /*
3443 * hepler to find next unprocessed extent
3444 */
3445 static noinline_for_stack
3446 int find_next_extent(struct btrfs_trans_handle *trans,
3447 struct reloc_control *rc, struct btrfs_path *path,
3448 struct btrfs_key *extent_key)
3449 {
3450 struct btrfs_key key;
3451 struct extent_buffer *leaf;
3452 u64 start, end, last;
3453 int ret;
3454
3455 last = rc->block_group->key.objectid + rc->block_group->key.offset;
3456 while (1) {
3457 cond_resched();
3458 if (rc->search_start >= last) {
3459 ret = 1;
3460 break;
3461 }
3462
3463 key.objectid = rc->search_start;
3464 key.type = BTRFS_EXTENT_ITEM_KEY;
3465 key.offset = 0;
3466
3467 path->search_commit_root = 1;
3468 path->skip_locking = 1;
3469 ret = btrfs_search_slot(NULL, rc->extent_root, &key, path,
3470 0, 0);
3471 if (ret < 0)
3472 break;
3473 next:
3474 leaf = path->nodes[0];
3475 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
3476 ret = btrfs_next_leaf(rc->extent_root, path);
3477 if (ret != 0)
3478 break;
3479 leaf = path->nodes[0];
3480 }
3481
3482 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
3483 if (key.objectid >= last) {
3484 ret = 1;
3485 break;
3486 }
3487
3488 if (key.type != BTRFS_EXTENT_ITEM_KEY ||
3489 key.objectid + key.offset <= rc->search_start) {
3490 path->slots[0]++;
3491 goto next;
3492 }
3493
3494 ret = find_first_extent_bit(&rc->processed_blocks,
3495 key.objectid, &start, &end,
3496 EXTENT_DIRTY);
3497
3498 if (ret == 0 && start <= key.objectid) {
3499 btrfs_release_path(rc->extent_root, path);
3500 rc->search_start = end + 1;
3501 } else {
3502 rc->search_start = key.objectid + key.offset;
3503 memcpy(extent_key, &key, sizeof(key));
3504 return 0;
3505 }
3506 }
3507 btrfs_release_path(rc->extent_root, path);
3508 return ret;
3509 }
3510
3511 static void set_reloc_control(struct reloc_control *rc)
3512 {
3513 struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
3514 mutex_lock(&fs_info->trans_mutex);
3515 fs_info->reloc_ctl = rc;
3516 mutex_unlock(&fs_info->trans_mutex);
3517 }
3518
3519 static void unset_reloc_control(struct reloc_control *rc)
3520 {
3521 struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
3522 mutex_lock(&fs_info->trans_mutex);
3523 fs_info->reloc_ctl = NULL;
3524 mutex_unlock(&fs_info->trans_mutex);
3525 }
3526
3527 static int check_extent_flags(u64 flags)
3528 {
3529 if ((flags & BTRFS_EXTENT_FLAG_DATA) &&
3530 (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK))
3531 return 1;
3532 if (!(flags & BTRFS_EXTENT_FLAG_DATA) &&
3533 !(flags & BTRFS_EXTENT_FLAG_TREE_BLOCK))
3534 return 1;
3535 if ((flags & BTRFS_EXTENT_FLAG_DATA) &&
3536 (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
3537 return 1;
3538 return 0;
3539 }
3540
3541 static noinline_for_stack
3542 int prepare_to_relocate(struct reloc_control *rc)
3543 {
3544 struct btrfs_trans_handle *trans;
3545 int ret;
3546
3547 rc->block_rsv = btrfs_alloc_block_rsv(rc->extent_root);
3548 if (!rc->block_rsv)
3549 return -ENOMEM;
3550
3551 /*
3552 * reserve some space for creating reloc trees.
3553 * btrfs_init_reloc_root will use them when there
3554 * is no reservation in transaction handle.
3555 */
3556 ret = btrfs_block_rsv_add(NULL, rc->extent_root, rc->block_rsv,
3557 rc->extent_root->nodesize * 256,
3558 &rc->block_rsv_retries);
3559 if (ret)
3560 return ret;
3561
3562 rc->block_rsv->refill_used = 1;
3563 btrfs_add_durable_block_rsv(rc->extent_root->fs_info, rc->block_rsv);
3564
3565 memset(&rc->cluster, 0, sizeof(rc->cluster));
3566 rc->search_start = rc->block_group->key.objectid;
3567 rc->extents_found = 0;
3568 rc->nodes_relocated = 0;
3569 rc->merging_rsv_size = 0;
3570 rc->block_rsv_retries = 0;
3571
3572 rc->create_reloc_tree = 1;
3573 set_reloc_control(rc);
3574
3575 trans = btrfs_join_transaction(rc->extent_root, 1);
3576 btrfs_commit_transaction(trans, rc->extent_root);
3577 return 0;
3578 }
3579
3580 static noinline_for_stack int relocate_block_group(struct reloc_control *rc)
3581 {
3582 struct rb_root blocks = RB_ROOT;
3583 struct btrfs_key key;
3584 struct btrfs_trans_handle *trans = NULL;
3585 struct btrfs_path *path;
3586 struct btrfs_extent_item *ei;
3587 unsigned long nr;
3588 u64 flags;
3589 u32 item_size;
3590 int ret;
3591 int err = 0;
3592
3593 path = btrfs_alloc_path();
3594 if (!path)
3595 return -ENOMEM;
3596
3597 ret = prepare_to_relocate(rc);
3598 if (ret) {
3599 err = ret;
3600 goto out_free;
3601 }
3602
3603 while (1) {
3604 trans = btrfs_start_transaction(rc->extent_root, 0);
3605
3606 if (update_backref_cache(trans, &rc->backref_cache)) {
3607 btrfs_end_transaction(trans, rc->extent_root);
3608 continue;
3609 }
3610
3611 ret = find_next_extent(trans, rc, path, &key);
3612 if (ret < 0)
3613 err = ret;
3614 if (ret != 0)
3615 break;
3616
3617 rc->extents_found++;
3618
3619 ei = btrfs_item_ptr(path->nodes[0], path->slots[0],
3620 struct btrfs_extent_item);
3621 item_size = btrfs_item_size_nr(path->nodes[0], path->slots[0]);
3622 if (item_size >= sizeof(*ei)) {
3623 flags = btrfs_extent_flags(path->nodes[0], ei);
3624 ret = check_extent_flags(flags);
3625 BUG_ON(ret);
3626
3627 } else {
3628 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
3629 u64 ref_owner;
3630 int path_change = 0;
3631
3632 BUG_ON(item_size !=
3633 sizeof(struct btrfs_extent_item_v0));
3634 ret = get_ref_objectid_v0(rc, path, &key, &ref_owner,
3635 &path_change);
3636 if (ref_owner < BTRFS_FIRST_FREE_OBJECTID)
3637 flags = BTRFS_EXTENT_FLAG_TREE_BLOCK;
3638 else
3639 flags = BTRFS_EXTENT_FLAG_DATA;
3640
3641 if (path_change) {
3642 btrfs_release_path(rc->extent_root, path);
3643
3644 path->search_commit_root = 1;
3645 path->skip_locking = 1;
3646 ret = btrfs_search_slot(NULL, rc->extent_root,
3647 &key, path, 0, 0);
3648 if (ret < 0) {
3649 err = ret;
3650 break;
3651 }
3652 BUG_ON(ret > 0);
3653 }
3654 #else
3655 BUG();
3656 #endif
3657 }
3658
3659 if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
3660 ret = add_tree_block(rc, &key, path, &blocks);
3661 } else if (rc->stage == UPDATE_DATA_PTRS &&
3662 (flags & BTRFS_EXTENT_FLAG_DATA)) {
3663 ret = add_data_references(rc, &key, path, &blocks);
3664 } else {
3665 btrfs_release_path(rc->extent_root, path);
3666 ret = 0;
3667 }
3668 if (ret < 0) {
3669 err = ret;
3670 break;
3671 }
3672
3673 if (!RB_EMPTY_ROOT(&blocks)) {
3674 ret = relocate_tree_blocks(trans, rc, &blocks);
3675 if (ret < 0) {
3676 if (ret != -EAGAIN) {
3677 err = ret;
3678 break;
3679 }
3680 rc->extents_found--;
3681 rc->search_start = key.objectid;
3682 }
3683 }
3684
3685 ret = btrfs_block_rsv_check(trans, rc->extent_root,
3686 rc->block_rsv, 0, 5);
3687 if (ret < 0) {
3688 if (ret != -EAGAIN) {
3689 err = ret;
3690 WARN_ON(1);
3691 break;
3692 }
3693 rc->commit_transaction = 1;
3694 }
3695
3696 if (rc->commit_transaction) {
3697 rc->commit_transaction = 0;
3698 ret = btrfs_commit_transaction(trans, rc->extent_root);
3699 BUG_ON(ret);
3700 } else {
3701 nr = trans->blocks_used;
3702 btrfs_end_transaction_throttle(trans, rc->extent_root);
3703 btrfs_btree_balance_dirty(rc->extent_root, nr);
3704 }
3705 trans = NULL;
3706
3707 if (rc->stage == MOVE_DATA_EXTENTS &&
3708 (flags & BTRFS_EXTENT_FLAG_DATA)) {
3709 rc->found_file_extent = 1;
3710 ret = relocate_data_extent(rc->data_inode,
3711 &key, &rc->cluster);
3712 if (ret < 0) {
3713 err = ret;
3714 break;
3715 }
3716 }
3717 }
3718
3719 btrfs_release_path(rc->extent_root, path);
3720 clear_extent_bits(&rc->processed_blocks, 0, (u64)-1, EXTENT_DIRTY,
3721 GFP_NOFS);
3722
3723 if (trans) {
3724 nr = trans->blocks_used;
3725 btrfs_end_transaction_throttle(trans, rc->extent_root);
3726 btrfs_btree_balance_dirty(rc->extent_root, nr);
3727 }
3728
3729 if (!err) {
3730 ret = relocate_file_extent_cluster(rc->data_inode,
3731 &rc->cluster);
3732 if (ret < 0)
3733 err = ret;
3734 }
3735
3736 rc->create_reloc_tree = 0;
3737 set_reloc_control(rc);
3738
3739 backref_cache_cleanup(&rc->backref_cache);
3740 btrfs_block_rsv_release(rc->extent_root, rc->block_rsv, (u64)-1);
3741
3742 err = prepare_to_merge(rc, err);
3743
3744 merge_reloc_roots(rc);
3745
3746 rc->merge_reloc_tree = 0;
3747 unset_reloc_control(rc);
3748 btrfs_block_rsv_release(rc->extent_root, rc->block_rsv, (u64)-1);
3749
3750 /* get rid of pinned extents */
3751 trans = btrfs_join_transaction(rc->extent_root, 1);
3752 btrfs_commit_transaction(trans, rc->extent_root);
3753 out_free:
3754 btrfs_free_block_rsv(rc->extent_root, rc->block_rsv);
3755 btrfs_free_path(path);
3756 return err;
3757 }
3758
3759 static int __insert_orphan_inode(struct btrfs_trans_handle *trans,
3760 struct btrfs_root *root, u64 objectid)
3761 {
3762 struct btrfs_path *path;
3763 struct btrfs_inode_item *item;
3764 struct extent_buffer *leaf;
3765 int ret;
3766
3767 path = btrfs_alloc_path();
3768 if (!path)
3769 return -ENOMEM;
3770
3771 ret = btrfs_insert_empty_inode(trans, root, path, objectid);
3772 if (ret)
3773 goto out;
3774
3775 leaf = path->nodes[0];
3776 item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_inode_item);
3777 memset_extent_buffer(leaf, 0, (unsigned long)item, sizeof(*item));
3778 btrfs_set_inode_generation(leaf, item, 1);
3779 btrfs_set_inode_size(leaf, item, 0);
3780 btrfs_set_inode_mode(leaf, item, S_IFREG | 0600);
3781 btrfs_set_inode_flags(leaf, item, BTRFS_INODE_NOCOMPRESS |
3782 BTRFS_INODE_PREALLOC);
3783 btrfs_mark_buffer_dirty(leaf);
3784 btrfs_release_path(root, path);
3785 out:
3786 btrfs_free_path(path);
3787 return ret;
3788 }
3789
3790 /*
3791 * helper to create inode for data relocation.
3792 * the inode is in data relocation tree and its link count is 0
3793 */
3794 static noinline_for_stack
3795 struct inode *create_reloc_inode(struct btrfs_fs_info *fs_info,
3796 struct btrfs_block_group_cache *group)
3797 {
3798 struct inode *inode = NULL;
3799 struct btrfs_trans_handle *trans;
3800 struct btrfs_root *root;
3801 struct btrfs_key key;
3802 unsigned long nr;
3803 u64 objectid = BTRFS_FIRST_FREE_OBJECTID;
3804 int err = 0;
3805
3806 root = read_fs_root(fs_info, BTRFS_DATA_RELOC_TREE_OBJECTID);
3807 if (IS_ERR(root))
3808 return ERR_CAST(root);
3809
3810 trans = btrfs_start_transaction(root, 6);
3811 if (IS_ERR(trans))
3812 return ERR_CAST(trans);
3813
3814 err = btrfs_find_free_objectid(trans, root, objectid, &objectid);
3815 if (err)
3816 goto out;
3817
3818 err = __insert_orphan_inode(trans, root, objectid);
3819 BUG_ON(err);
3820
3821 key.objectid = objectid;
3822 key.type = BTRFS_INODE_ITEM_KEY;
3823 key.offset = 0;
3824 inode = btrfs_iget(root->fs_info->sb, &key, root, NULL);
3825 BUG_ON(IS_ERR(inode) || is_bad_inode(inode));
3826 BTRFS_I(inode)->index_cnt = group->key.objectid;
3827
3828 err = btrfs_orphan_add(trans, inode);
3829 out:
3830 nr = trans->blocks_used;
3831 btrfs_end_transaction(trans, root);
3832 btrfs_btree_balance_dirty(root, nr);
3833 if (err) {
3834 if (inode)
3835 iput(inode);
3836 inode = ERR_PTR(err);
3837 }
3838 return inode;
3839 }
3840
3841 static struct reloc_control *alloc_reloc_control(void)
3842 {
3843 struct reloc_control *rc;
3844
3845 rc = kzalloc(sizeof(*rc), GFP_NOFS);
3846 if (!rc)
3847 return NULL;
3848
3849 INIT_LIST_HEAD(&rc->reloc_roots);
3850 backref_cache_init(&rc->backref_cache);
3851 mapping_tree_init(&rc->reloc_root_tree);
3852 extent_io_tree_init(&rc->processed_blocks, NULL, GFP_NOFS);
3853 return rc;
3854 }
3855
3856 /*
3857 * function to relocate all extents in a block group.
3858 */
3859 int btrfs_relocate_block_group(struct btrfs_root *extent_root, u64 group_start)
3860 {
3861 struct btrfs_fs_info *fs_info = extent_root->fs_info;
3862 struct reloc_control *rc;
3863 int ret;
3864 int rw = 0;
3865 int err = 0;
3866
3867 rc = alloc_reloc_control();
3868 if (!rc)
3869 return -ENOMEM;
3870
3871 rc->extent_root = extent_root;
3872
3873 rc->block_group = btrfs_lookup_block_group(fs_info, group_start);
3874 BUG_ON(!rc->block_group);
3875
3876 if (!rc->block_group->ro) {
3877 ret = btrfs_set_block_group_ro(extent_root, rc->block_group);
3878 if (ret) {
3879 err = ret;
3880 goto out;
3881 }
3882 rw = 1;
3883 }
3884
3885 rc->data_inode = create_reloc_inode(fs_info, rc->block_group);
3886 if (IS_ERR(rc->data_inode)) {
3887 err = PTR_ERR(rc->data_inode);
3888 rc->data_inode = NULL;
3889 goto out;
3890 }
3891
3892 printk(KERN_INFO "btrfs: relocating block group %llu flags %llu\n",
3893 (unsigned long long)rc->block_group->key.objectid,
3894 (unsigned long long)rc->block_group->flags);
3895
3896 btrfs_start_delalloc_inodes(fs_info->tree_root, 0);
3897 btrfs_wait_ordered_extents(fs_info->tree_root, 0, 0);
3898
3899 while (1) {
3900 mutex_lock(&fs_info->cleaner_mutex);
3901
3902 btrfs_clean_old_snapshots(fs_info->tree_root);
3903 ret = relocate_block_group(rc);
3904
3905 mutex_unlock(&fs_info->cleaner_mutex);
3906 if (ret < 0) {
3907 err = ret;
3908 goto out;
3909 }
3910
3911 if (rc->extents_found == 0)
3912 break;
3913
3914 printk(KERN_INFO "btrfs: found %llu extents\n",
3915 (unsigned long long)rc->extents_found);
3916
3917 if (rc->stage == MOVE_DATA_EXTENTS && rc->found_file_extent) {
3918 btrfs_wait_ordered_range(rc->data_inode, 0, (u64)-1);
3919 invalidate_mapping_pages(rc->data_inode->i_mapping,
3920 0, -1);
3921 rc->stage = UPDATE_DATA_PTRS;
3922 }
3923 }
3924
3925 filemap_write_and_wait_range(fs_info->btree_inode->i_mapping,
3926 rc->block_group->key.objectid,
3927 rc->block_group->key.objectid +
3928 rc->block_group->key.offset - 1);
3929
3930 WARN_ON(rc->block_group->pinned > 0);
3931 WARN_ON(rc->block_group->reserved > 0);
3932 WARN_ON(btrfs_block_group_used(&rc->block_group->item) > 0);
3933 out:
3934 if (err && rw)
3935 btrfs_set_block_group_rw(extent_root, rc->block_group);
3936 iput(rc->data_inode);
3937 btrfs_put_block_group(rc->block_group);
3938 kfree(rc);
3939 return err;
3940 }
3941
3942 static noinline_for_stack int mark_garbage_root(struct btrfs_root *root)
3943 {
3944 struct btrfs_trans_handle *trans;
3945 int ret;
3946
3947 trans = btrfs_start_transaction(root->fs_info->tree_root, 0);
3948
3949 memset(&root->root_item.drop_progress, 0,
3950 sizeof(root->root_item.drop_progress));
3951 root->root_item.drop_level = 0;
3952 btrfs_set_root_refs(&root->root_item, 0);
3953 ret = btrfs_update_root(trans, root->fs_info->tree_root,
3954 &root->root_key, &root->root_item);
3955 BUG_ON(ret);
3956
3957 ret = btrfs_end_transaction(trans, root->fs_info->tree_root);
3958 BUG_ON(ret);
3959 return 0;
3960 }
3961
3962 /*
3963 * recover relocation interrupted by system crash.
3964 *
3965 * this function resumes merging reloc trees with corresponding fs trees.
3966 * this is important for keeping the sharing of tree blocks
3967 */
3968 int btrfs_recover_relocation(struct btrfs_root *root)
3969 {
3970 LIST_HEAD(reloc_roots);
3971 struct btrfs_key key;
3972 struct btrfs_root *fs_root;
3973 struct btrfs_root *reloc_root;
3974 struct btrfs_path *path;
3975 struct extent_buffer *leaf;
3976 struct reloc_control *rc = NULL;
3977 struct btrfs_trans_handle *trans;
3978 int ret;
3979 int err = 0;
3980
3981 path = btrfs_alloc_path();
3982 if (!path)
3983 return -ENOMEM;
3984
3985 key.objectid = BTRFS_TREE_RELOC_OBJECTID;
3986 key.type = BTRFS_ROOT_ITEM_KEY;
3987 key.offset = (u64)-1;
3988
3989 while (1) {
3990 ret = btrfs_search_slot(NULL, root->fs_info->tree_root, &key,
3991 path, 0, 0);
3992 if (ret < 0) {
3993 err = ret;
3994 goto out;
3995 }
3996 if (ret > 0) {
3997 if (path->slots[0] == 0)
3998 break;
3999 path->slots[0]--;
4000 }
4001 leaf = path->nodes[0];
4002 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
4003 btrfs_release_path(root->fs_info->tree_root, path);
4004
4005 if (key.objectid != BTRFS_TREE_RELOC_OBJECTID ||
4006 key.type != BTRFS_ROOT_ITEM_KEY)
4007 break;
4008
4009 reloc_root = btrfs_read_fs_root_no_radix(root, &key);
4010 if (IS_ERR(reloc_root)) {
4011 err = PTR_ERR(reloc_root);
4012 goto out;
4013 }
4014
4015 list_add(&reloc_root->root_list, &reloc_roots);
4016
4017 if (btrfs_root_refs(&reloc_root->root_item) > 0) {
4018 fs_root = read_fs_root(root->fs_info,
4019 reloc_root->root_key.offset);
4020 if (IS_ERR(fs_root)) {
4021 ret = PTR_ERR(fs_root);
4022 if (ret != -ENOENT) {
4023 err = ret;
4024 goto out;
4025 }
4026 mark_garbage_root(reloc_root);
4027 }
4028 }
4029
4030 if (key.offset == 0)
4031 break;
4032
4033 key.offset--;
4034 }
4035 btrfs_release_path(root->fs_info->tree_root, path);
4036
4037 if (list_empty(&reloc_roots))
4038 goto out;
4039
4040 rc = alloc_reloc_control();
4041 if (!rc) {
4042 err = -ENOMEM;
4043 goto out;
4044 }
4045
4046 rc->extent_root = root->fs_info->extent_root;
4047
4048 set_reloc_control(rc);
4049
4050 trans = btrfs_join_transaction(rc->extent_root, 1);
4051
4052 rc->merge_reloc_tree = 1;
4053
4054 while (!list_empty(&reloc_roots)) {
4055 reloc_root = list_entry(reloc_roots.next,
4056 struct btrfs_root, root_list);
4057 list_del(&reloc_root->root_list);
4058
4059 if (btrfs_root_refs(&reloc_root->root_item) == 0) {
4060 list_add_tail(&reloc_root->root_list,
4061 &rc->reloc_roots);
4062 continue;
4063 }
4064
4065 fs_root = read_fs_root(root->fs_info,
4066 reloc_root->root_key.offset);
4067 BUG_ON(IS_ERR(fs_root));
4068
4069 __add_reloc_root(reloc_root);
4070 fs_root->reloc_root = reloc_root;
4071 }
4072
4073 btrfs_commit_transaction(trans, rc->extent_root);
4074
4075 merge_reloc_roots(rc);
4076
4077 unset_reloc_control(rc);
4078
4079 trans = btrfs_join_transaction(rc->extent_root, 1);
4080 btrfs_commit_transaction(trans, rc->extent_root);
4081 out:
4082 kfree(rc);
4083 while (!list_empty(&reloc_roots)) {
4084 reloc_root = list_entry(reloc_roots.next,
4085 struct btrfs_root, root_list);
4086 list_del(&reloc_root->root_list);
4087 free_extent_buffer(reloc_root->node);
4088 free_extent_buffer(reloc_root->commit_root);
4089 kfree(reloc_root);
4090 }
4091 btrfs_free_path(path);
4092
4093 if (err == 0) {
4094 /* cleanup orphan inode in data relocation tree */
4095 fs_root = read_fs_root(root->fs_info,
4096 BTRFS_DATA_RELOC_TREE_OBJECTID);
4097 if (IS_ERR(fs_root))
4098 err = PTR_ERR(fs_root);
4099 else
4100 btrfs_orphan_cleanup(fs_root);
4101 }
4102 return err;
4103 }
4104
4105 /*
4106 * helper to add ordered checksum for data relocation.
4107 *
4108 * cloning checksum properly handles the nodatasum extents.
4109 * it also saves CPU time to re-calculate the checksum.
4110 */
4111 int btrfs_reloc_clone_csums(struct inode *inode, u64 file_pos, u64 len)
4112 {
4113 struct btrfs_ordered_sum *sums;
4114 struct btrfs_sector_sum *sector_sum;
4115 struct btrfs_ordered_extent *ordered;
4116 struct btrfs_root *root = BTRFS_I(inode)->root;
4117 size_t offset;
4118 int ret;
4119 u64 disk_bytenr;
4120 LIST_HEAD(list);
4121
4122 ordered = btrfs_lookup_ordered_extent(inode, file_pos);
4123 BUG_ON(ordered->file_offset != file_pos || ordered->len != len);
4124
4125 disk_bytenr = file_pos + BTRFS_I(inode)->index_cnt;
4126 ret = btrfs_lookup_csums_range(root->fs_info->csum_root, disk_bytenr,
4127 disk_bytenr + len - 1, &list);
4128
4129 while (!list_empty(&list)) {
4130 sums = list_entry(list.next, struct btrfs_ordered_sum, list);
4131 list_del_init(&sums->list);
4132
4133 sector_sum = sums->sums;
4134 sums->bytenr = ordered->start;
4135
4136 offset = 0;
4137 while (offset < sums->len) {
4138 sector_sum->bytenr += ordered->start - disk_bytenr;
4139 sector_sum++;
4140 offset += root->sectorsize;
4141 }
4142
4143 btrfs_add_ordered_sum(inode, ordered, sums);
4144 }
4145 btrfs_put_ordered_extent(ordered);
4146 return 0;
4147 }
4148
4149 void btrfs_reloc_cow_block(struct btrfs_trans_handle *trans,
4150 struct btrfs_root *root, struct extent_buffer *buf,
4151 struct extent_buffer *cow)
4152 {
4153 struct reloc_control *rc;
4154 struct backref_node *node;
4155 int first_cow = 0;
4156 int level;
4157 int ret;
4158
4159 rc = root->fs_info->reloc_ctl;
4160 if (!rc)
4161 return;
4162
4163 BUG_ON(rc->stage == UPDATE_DATA_PTRS &&
4164 root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID);
4165
4166 level = btrfs_header_level(buf);
4167 if (btrfs_header_generation(buf) <=
4168 btrfs_root_last_snapshot(&root->root_item))
4169 first_cow = 1;
4170
4171 if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID &&
4172 rc->create_reloc_tree) {
4173 WARN_ON(!first_cow && level == 0);
4174
4175 node = rc->backref_cache.path[level];
4176 BUG_ON(node->bytenr != buf->start &&
4177 node->new_bytenr != buf->start);
4178
4179 drop_node_buffer(node);
4180 extent_buffer_get(cow);
4181 node->eb = cow;
4182 node->new_bytenr = cow->start;
4183
4184 if (!node->pending) {
4185 list_move_tail(&node->list,
4186 &rc->backref_cache.pending[level]);
4187 node->pending = 1;
4188 }
4189
4190 if (first_cow)
4191 __mark_block_processed(rc, node);
4192
4193 if (first_cow && level > 0)
4194 rc->nodes_relocated += buf->len;
4195 }
4196
4197 if (level == 0 && first_cow && rc->stage == UPDATE_DATA_PTRS) {
4198 ret = replace_file_extents(trans, rc, root, cow);
4199 BUG_ON(ret);
4200 }
4201 }
4202
4203 /*
4204 * called before creating snapshot. it calculates metadata reservation
4205 * requried for relocating tree blocks in the snapshot
4206 */
4207 void btrfs_reloc_pre_snapshot(struct btrfs_trans_handle *trans,
4208 struct btrfs_pending_snapshot *pending,
4209 u64 *bytes_to_reserve)
4210 {
4211 struct btrfs_root *root;
4212 struct reloc_control *rc;
4213
4214 root = pending->root;
4215 if (!root->reloc_root)
4216 return;
4217
4218 rc = root->fs_info->reloc_ctl;
4219 if (!rc->merge_reloc_tree)
4220 return;
4221
4222 root = root->reloc_root;
4223 BUG_ON(btrfs_root_refs(&root->root_item) == 0);
4224 /*
4225 * relocation is in the stage of merging trees. the space
4226 * used by merging a reloc tree is twice the size of
4227 * relocated tree nodes in the worst case. half for cowing
4228 * the reloc tree, half for cowing the fs tree. the space
4229 * used by cowing the reloc tree will be freed after the
4230 * tree is dropped. if we create snapshot, cowing the fs
4231 * tree may use more space than it frees. so we need
4232 * reserve extra space.
4233 */
4234 *bytes_to_reserve += rc->nodes_relocated;
4235 }
4236
4237 /*
4238 * called after snapshot is created. migrate block reservation
4239 * and create reloc root for the newly created snapshot
4240 */
4241 void btrfs_reloc_post_snapshot(struct btrfs_trans_handle *trans,
4242 struct btrfs_pending_snapshot *pending)
4243 {
4244 struct btrfs_root *root = pending->root;
4245 struct btrfs_root *reloc_root;
4246 struct btrfs_root *new_root;
4247 struct reloc_control *rc;
4248 int ret;
4249
4250 if (!root->reloc_root)
4251 return;
4252
4253 rc = root->fs_info->reloc_ctl;
4254 rc->merging_rsv_size += rc->nodes_relocated;
4255
4256 if (rc->merge_reloc_tree) {
4257 ret = btrfs_block_rsv_migrate(&pending->block_rsv,
4258 rc->block_rsv,
4259 rc->nodes_relocated);
4260 BUG_ON(ret);
4261 }
4262
4263 new_root = pending->snap;
4264 reloc_root = create_reloc_root(trans, root->reloc_root,
4265 new_root->root_key.objectid);
4266
4267 __add_reloc_root(reloc_root);
4268 new_root->reloc_root = reloc_root;
4269
4270 if (rc->create_reloc_tree) {
4271 ret = clone_backref_node(trans, rc, root, reloc_root);
4272 BUG_ON(ret);
4273 }
4274 }
Linus' kernel tree