?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
ARC: export "abort" for modules
[linux/fpc-iii.git] / fs / btrfs / relocation.c
blobb0c3a6afe6643a5aec3e104c612e152af7109ee0
1 /*
2 * Copyright (C) 2009 Oracle. All rights reserved.
3 *
4 * This program is free software; you can redistribute it and/or
5 * modify it under the terms of the GNU General Public
6 * License v2 as published by the Free Software Foundation.
7 *
8 * This program is distributed in the hope that it will be useful,
9 * but WITHOUT ANY WARRANTY; without even the implied warranty of
10 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
11 * General Public License for more details.
12 *
13 * You should have received a copy of the GNU General Public
14 * License along with this program; if not, write to the
15 * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
16 * Boston, MA 021110-1307, USA.
17 */
18
19 #include <linux/sched.h>
20 #include <linux/pagemap.h>
21 #include <linux/writeback.h>
22 #include <linux/blkdev.h>
23 #include <linux/rbtree.h>
24 #include <linux/slab.h>
25 #include "ctree.h"
26 #include "disk-io.h"
27 #include "transaction.h"
28 #include "volumes.h"
29 #include "locking.h"
30 #include "btrfs_inode.h"
31 #include "async-thread.h"
32 #include "free-space-cache.h"
33 #include "inode-map.h"
34 #include "qgroup.h"
35
36 /*
37 * backref_node, mapping_node and tree_block start with this
38 */
39 struct tree_entry {
40 struct rb_node rb_node;
41 u64 bytenr;
42 };
43
44 /*
45 * present a tree block in the backref cache
46 */
47 struct backref_node {
48 struct rb_node rb_node;
49 u64 bytenr;
50
51 u64 new_bytenr;
52 /* objectid of tree block owner, can be not uptodate */
53 u64 owner;
54 /* link to pending, changed or detached list */
55 struct list_head list;
56 /* list of upper level blocks reference this block */
57 struct list_head upper;
58 /* list of child blocks in the cache */
59 struct list_head lower;
60 /* NULL if this node is not tree root */
61 struct btrfs_root *root;
62 /* extent buffer got by COW the block */
63 struct extent_buffer *eb;
64 /* level of tree block */
65 unsigned int level:8;
66 /* is the block in non-reference counted tree */
67 unsigned int cowonly:1;
68 /* 1 if no child node in the cache */
69 unsigned int lowest:1;
70 /* is the extent buffer locked */
71 unsigned int locked:1;
72 /* has the block been processed */
73 unsigned int processed:1;
74 /* have backrefs of this block been checked */
75 unsigned int checked:1;
76 /*
77 * 1 if corresponding block has been cowed but some upper
78 * level block pointers may not point to the new location
79 */
80 unsigned int pending:1;
81 /*
82 * 1 if the backref node isn't connected to any other
83 * backref node.
84 */
85 unsigned int detached:1;
86 };
87
88 /*
89 * present a block pointer in the backref cache
90 */
91 struct backref_edge {
92 struct list_head list[2];
93 struct backref_node *node[2];
94 };
95
96 #define LOWER 0
97 #define UPPER 1
98 #define RELOCATION_RESERVED_NODES 256
99
100 struct backref_cache {
101 /* red black tree of all backref nodes in the cache */
102 struct rb_root rb_root;
103 /* for passing backref nodes to btrfs_reloc_cow_block */
104 struct backref_node *path[BTRFS_MAX_LEVEL];
105 /*
106 * list of blocks that have been cowed but some block
107 * pointers in upper level blocks may not reflect the
108 * new location
109 */
110 struct list_head pending[BTRFS_MAX_LEVEL];
111 /* list of backref nodes with no child node */
112 struct list_head leaves;
113 /* list of blocks that have been cowed in current transaction */
114 struct list_head changed;
115 /* list of detached backref node. */
116 struct list_head detached;
117
118 u64 last_trans;
119
120 int nr_nodes;
121 int nr_edges;
122 };
123
124 /*
125 * map address of tree root to tree
126 */
127 struct mapping_node {
128 struct rb_node rb_node;
129 u64 bytenr;
130 void *data;
131 };
132
133 struct mapping_tree {
134 struct rb_root rb_root;
135 spinlock_t lock;
136 };
137
138 /*
139 * present a tree block to process
140 */
141 struct tree_block {
142 struct rb_node rb_node;
143 u64 bytenr;
144 struct btrfs_key key;
145 unsigned int level:8;
146 unsigned int key_ready:1;
147 };
148
149 #define MAX_EXTENTS 128
150
151 struct file_extent_cluster {
152 u64 start;
153 u64 end;
154 u64 boundary[MAX_EXTENTS];
155 unsigned int nr;
156 };
157
158 struct reloc_control {
159 /* block group to relocate */
160 struct btrfs_block_group_cache *block_group;
161 /* extent tree */
162 struct btrfs_root *extent_root;
163 /* inode for moving data */
164 struct inode *data_inode;
165
166 struct btrfs_block_rsv *block_rsv;
167
168 struct backref_cache backref_cache;
169
170 struct file_extent_cluster cluster;
171 /* tree blocks have been processed */
172 struct extent_io_tree processed_blocks;
173 /* map start of tree root to corresponding reloc tree */
174 struct mapping_tree reloc_root_tree;
175 /* list of reloc trees */
176 struct list_head reloc_roots;
177 /* size of metadata reservation for merging reloc trees */
178 u64 merging_rsv_size;
179 /* size of relocated tree nodes */
180 u64 nodes_relocated;
181 /* reserved size for block group relocation*/
182 u64 reserved_bytes;
183
184 u64 search_start;
185 u64 extents_found;
186
187 unsigned int stage:8;
188 unsigned int create_reloc_tree:1;
189 unsigned int merge_reloc_tree:1;
190 unsigned int found_file_extent:1;
191 };
192
193 /* stages of data relocation */
194 #define MOVE_DATA_EXTENTS 0
195 #define UPDATE_DATA_PTRS 1
196
197 static void remove_backref_node(struct backref_cache *cache,
198 struct backref_node *node);
199 static void __mark_block_processed(struct reloc_control *rc,
200 struct backref_node *node);
201
202 static void mapping_tree_init(struct mapping_tree *tree)
203 {
204 tree->rb_root = RB_ROOT;
205 spin_lock_init(&tree->lock);
206 }
207
208 static void backref_cache_init(struct backref_cache *cache)
209 {
210 int i;
211 cache->rb_root = RB_ROOT;
212 for (i = 0; i < BTRFS_MAX_LEVEL; i++)
213 INIT_LIST_HEAD(&cache->pending[i]);
214 INIT_LIST_HEAD(&cache->changed);
215 INIT_LIST_HEAD(&cache->detached);
216 INIT_LIST_HEAD(&cache->leaves);
217 }
218
219 static void backref_cache_cleanup(struct backref_cache *cache)
220 {
221 struct backref_node *node;
222 int i;
223
224 while (!list_empty(&cache->detached)) {
225 node = list_entry(cache->detached.next,
226 struct backref_node, list);
227 remove_backref_node(cache, node);
228 }
229
230 while (!list_empty(&cache->leaves)) {
231 node = list_entry(cache->leaves.next,
232 struct backref_node, lower);
233 remove_backref_node(cache, node);
234 }
235
236 cache->last_trans = 0;
237
238 for (i = 0; i < BTRFS_MAX_LEVEL; i++)
239 ASSERT(list_empty(&cache->pending[i]));
240 ASSERT(list_empty(&cache->changed));
241 ASSERT(list_empty(&cache->detached));
242 ASSERT(RB_EMPTY_ROOT(&cache->rb_root));
243 ASSERT(!cache->nr_nodes);
244 ASSERT(!cache->nr_edges);
245 }
246
247 static struct backref_node *alloc_backref_node(struct backref_cache *cache)
248 {
249 struct backref_node *node;
250
251 node = kzalloc(sizeof(*node), GFP_NOFS);
252 if (node) {
253 INIT_LIST_HEAD(&node->list);
254 INIT_LIST_HEAD(&node->upper);
255 INIT_LIST_HEAD(&node->lower);
256 RB_CLEAR_NODE(&node->rb_node);
257 cache->nr_nodes++;
258 }
259 return node;
260 }
261
262 static void free_backref_node(struct backref_cache *cache,
263 struct backref_node *node)
264 {
265 if (node) {
266 cache->nr_nodes--;
267 kfree(node);
268 }
269 }
270
271 static struct backref_edge *alloc_backref_edge(struct backref_cache *cache)
272 {
273 struct backref_edge *edge;
274
275 edge = kzalloc(sizeof(*edge), GFP_NOFS);
276 if (edge)
277 cache->nr_edges++;
278 return edge;
279 }
280
281 static void free_backref_edge(struct backref_cache *cache,
282 struct backref_edge *edge)
283 {
284 if (edge) {
285 cache->nr_edges--;
286 kfree(edge);
287 }
288 }
289
290 static struct rb_node *tree_insert(struct rb_root *root, u64 bytenr,
291 struct rb_node *node)
292 {
293 struct rb_node **p = &root->rb_node;
294 struct rb_node *parent = NULL;
295 struct tree_entry *entry;
296
297 while (*p) {
298 parent = *p;
299 entry = rb_entry(parent, struct tree_entry, rb_node);
300
301 if (bytenr < entry->bytenr)
302 p = &(*p)->rb_left;
303 else if (bytenr > entry->bytenr)
304 p = &(*p)->rb_right;
305 else
306 return parent;
307 }
308
309 rb_link_node(node, parent, p);
310 rb_insert_color(node, root);
311 return NULL;
312 }
313
314 static struct rb_node *tree_search(struct rb_root *root, u64 bytenr)
315 {
316 struct rb_node *n = root->rb_node;
317 struct tree_entry *entry;
318
319 while (n) {
320 entry = rb_entry(n, struct tree_entry, rb_node);
321
322 if (bytenr < entry->bytenr)
323 n = n->rb_left;
324 else if (bytenr > entry->bytenr)
325 n = n->rb_right;
326 else
327 return n;
328 }
329 return NULL;
330 }
331
332 static void backref_tree_panic(struct rb_node *rb_node, int errno, u64 bytenr)
333 {
334
335 struct btrfs_fs_info *fs_info = NULL;
336 struct backref_node *bnode = rb_entry(rb_node, struct backref_node,
337 rb_node);
338 if (bnode->root)
339 fs_info = bnode->root->fs_info;
340 btrfs_panic(fs_info, errno,
341 "Inconsistency in backref cache found at offset %llu",
342 bytenr);
343 }
344
345 /*
346 * walk up backref nodes until reach node presents tree root
347 */
348 static struct backref_node *walk_up_backref(struct backref_node *node,
349 struct backref_edge *edges[],
350 int *index)
351 {
352 struct backref_edge *edge;
353 int idx = *index;
354
355 while (!list_empty(&node->upper)) {
356 edge = list_entry(node->upper.next,
357 struct backref_edge, list[LOWER]);
358 edges[idx++] = edge;
359 node = edge->node[UPPER];
360 }
361 BUG_ON(node->detached);
362 *index = idx;
363 return node;
364 }
365
366 /*
367 * walk down backref nodes to find start of next reference path
368 */
369 static struct backref_node *walk_down_backref(struct backref_edge *edges[],
370 int *index)
371 {
372 struct backref_edge *edge;
373 struct backref_node *lower;
374 int idx = *index;
375
376 while (idx > 0) {
377 edge = edges[idx - 1];
378 lower = edge->node[LOWER];
379 if (list_is_last(&edge->list[LOWER], &lower->upper)) {
380 idx--;
381 continue;
382 }
383 edge = list_entry(edge->list[LOWER].next,
384 struct backref_edge, list[LOWER]);
385 edges[idx - 1] = edge;
386 *index = idx;
387 return edge->node[UPPER];
388 }
389 *index = 0;
390 return NULL;
391 }
392
393 static void unlock_node_buffer(struct backref_node *node)
394 {
395 if (node->locked) {
396 btrfs_tree_unlock(node->eb);
397 node->locked = 0;
398 }
399 }
400
401 static void drop_node_buffer(struct backref_node *node)
402 {
403 if (node->eb) {
404 unlock_node_buffer(node);
405 free_extent_buffer(node->eb);
406 node->eb = NULL;
407 }
408 }
409
410 static void drop_backref_node(struct backref_cache *tree,
411 struct backref_node *node)
412 {
413 BUG_ON(!list_empty(&node->upper));
414
415 drop_node_buffer(node);
416 list_del(&node->list);
417 list_del(&node->lower);
418 if (!RB_EMPTY_NODE(&node->rb_node))
419 rb_erase(&node->rb_node, &tree->rb_root);
420 free_backref_node(tree, node);
421 }
422
423 /*
424 * remove a backref node from the backref cache
425 */
426 static void remove_backref_node(struct backref_cache *cache,
427 struct backref_node *node)
428 {
429 struct backref_node *upper;
430 struct backref_edge *edge;
431
432 if (!node)
433 return;
434
435 BUG_ON(!node->lowest && !node->detached);
436 while (!list_empty(&node->upper)) {
437 edge = list_entry(node->upper.next, struct backref_edge,
438 list[LOWER]);
439 upper = edge->node[UPPER];
440 list_del(&edge->list[LOWER]);
441 list_del(&edge->list[UPPER]);
442 free_backref_edge(cache, edge);
443
444 if (RB_EMPTY_NODE(&upper->rb_node)) {
445 BUG_ON(!list_empty(&node->upper));
446 drop_backref_node(cache, node);
447 node = upper;
448 node->lowest = 1;
449 continue;
450 }
451 /*
452 * add the node to leaf node list if no other
453 * child block cached.
454 */
455 if (list_empty(&upper->lower)) {
456 list_add_tail(&upper->lower, &cache->leaves);
457 upper->lowest = 1;
458 }
459 }
460
461 drop_backref_node(cache, node);
462 }
463
464 static void update_backref_node(struct backref_cache *cache,
465 struct backref_node *node, u64 bytenr)
466 {
467 struct rb_node *rb_node;
468 rb_erase(&node->rb_node, &cache->rb_root);
469 node->bytenr = bytenr;
470 rb_node = tree_insert(&cache->rb_root, node->bytenr, &node->rb_node);
471 if (rb_node)
472 backref_tree_panic(rb_node, -EEXIST, bytenr);
473 }
474
475 /*
476 * update backref cache after a transaction commit
477 */
478 static int update_backref_cache(struct btrfs_trans_handle *trans,
479 struct backref_cache *cache)
480 {
481 struct backref_node *node;
482 int level = 0;
483
484 if (cache->last_trans == 0) {
485 cache->last_trans = trans->transid;
486 return 0;
487 }
488
489 if (cache->last_trans == trans->transid)
490 return 0;
491
492 /*
493 * detached nodes are used to avoid unnecessary backref
494 * lookup. transaction commit changes the extent tree.
495 * so the detached nodes are no longer useful.
496 */
497 while (!list_empty(&cache->detached)) {
498 node = list_entry(cache->detached.next,
499 struct backref_node, list);
500 remove_backref_node(cache, node);
501 }
502
503 while (!list_empty(&cache->changed)) {
504 node = list_entry(cache->changed.next,
505 struct backref_node, list);
506 list_del_init(&node->list);
507 BUG_ON(node->pending);
508 update_backref_node(cache, node, node->new_bytenr);
509 }
510
511 /*
512 * some nodes can be left in the pending list if there were
513 * errors during processing the pending nodes.
514 */
515 for (level = 0; level < BTRFS_MAX_LEVEL; level++) {
516 list_for_each_entry(node, &cache->pending[level], list) {
517 BUG_ON(!node->pending);
518 if (node->bytenr == node->new_bytenr)
519 continue;
520 update_backref_node(cache, node, node->new_bytenr);
521 }
522 }
523
524 cache->last_trans = 0;
525 return 1;
526 }
527
528
529 static int should_ignore_root(struct btrfs_root *root)
530 {
531 struct btrfs_root *reloc_root;
532
533 if (!test_bit(BTRFS_ROOT_REF_COWS, &root->state))
534 return 0;
535
536 reloc_root = root->reloc_root;
537 if (!reloc_root)
538 return 0;
539
540 if (btrfs_root_last_snapshot(&reloc_root->root_item) ==
541 root->fs_info->running_transaction->transid - 1)
542 return 0;
543 /*
544 * if there is reloc tree and it was created in previous
545 * transaction backref lookup can find the reloc tree,
546 * so backref node for the fs tree root is useless for
547 * relocation.
548 */
549 return 1;
550 }
551 /*
552 * find reloc tree by address of tree root
553 */
554 static struct btrfs_root *find_reloc_root(struct reloc_control *rc,
555 u64 bytenr)
556 {
557 struct rb_node *rb_node;
558 struct mapping_node *node;
559 struct btrfs_root *root = NULL;
560
561 spin_lock(&rc->reloc_root_tree.lock);
562 rb_node = tree_search(&rc->reloc_root_tree.rb_root, bytenr);
563 if (rb_node) {
564 node = rb_entry(rb_node, struct mapping_node, rb_node);
565 root = (struct btrfs_root *)node->data;
566 }
567 spin_unlock(&rc->reloc_root_tree.lock);
568 return root;
569 }
570
571 static int is_cowonly_root(u64 root_objectid)
572 {
573 if (root_objectid == BTRFS_ROOT_TREE_OBJECTID ||
574 root_objectid == BTRFS_EXTENT_TREE_OBJECTID ||
575 root_objectid == BTRFS_CHUNK_TREE_OBJECTID ||
576 root_objectid == BTRFS_DEV_TREE_OBJECTID ||
577 root_objectid == BTRFS_TREE_LOG_OBJECTID ||
578 root_objectid == BTRFS_CSUM_TREE_OBJECTID ||
579 root_objectid == BTRFS_UUID_TREE_OBJECTID ||
580 root_objectid == BTRFS_QUOTA_TREE_OBJECTID ||
581 root_objectid == BTRFS_FREE_SPACE_TREE_OBJECTID)
582 return 1;
583 return 0;
584 }
585
586 static struct btrfs_root *read_fs_root(struct btrfs_fs_info *fs_info,
587 u64 root_objectid)
588 {
589 struct btrfs_key key;
590
591 key.objectid = root_objectid;
592 key.type = BTRFS_ROOT_ITEM_KEY;
593 if (is_cowonly_root(root_objectid))
594 key.offset = 0;
595 else
596 key.offset = (u64)-1;
597
598 return btrfs_get_fs_root(fs_info, &key, false);
599 }
600
601 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
602 static noinline_for_stack
603 struct btrfs_root *find_tree_root(struct reloc_control *rc,
604 struct extent_buffer *leaf,
605 struct btrfs_extent_ref_v0 *ref0)
606 {
607 struct btrfs_root *root;
608 u64 root_objectid = btrfs_ref_root_v0(leaf, ref0);
609 u64 generation = btrfs_ref_generation_v0(leaf, ref0);
610
611 BUG_ON(root_objectid == BTRFS_TREE_RELOC_OBJECTID);
612
613 root = read_fs_root(rc->extent_root->fs_info, root_objectid);
614 BUG_ON(IS_ERR(root));
615
616 if (test_bit(BTRFS_ROOT_REF_COWS, &root->state) &&
617 generation != btrfs_root_generation(&root->root_item))
618 return NULL;
619
620 return root;
621 }
622 #endif
623
624 static noinline_for_stack
625 int find_inline_backref(struct extent_buffer *leaf, int slot,
626 unsigned long *ptr, unsigned long *end)
627 {
628 struct btrfs_key key;
629 struct btrfs_extent_item *ei;
630 struct btrfs_tree_block_info *bi;
631 u32 item_size;
632
633 btrfs_item_key_to_cpu(leaf, &key, slot);
634
635 item_size = btrfs_item_size_nr(leaf, slot);
636 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
637 if (item_size < sizeof(*ei)) {
638 WARN_ON(item_size != sizeof(struct btrfs_extent_item_v0));
639 return 1;
640 }
641 #endif
642 ei = btrfs_item_ptr(leaf, slot, struct btrfs_extent_item);
643 WARN_ON(!(btrfs_extent_flags(leaf, ei) &
644 BTRFS_EXTENT_FLAG_TREE_BLOCK));
645
646 if (key.type == BTRFS_EXTENT_ITEM_KEY &&
647 item_size <= sizeof(*ei) + sizeof(*bi)) {
648 WARN_ON(item_size < sizeof(*ei) + sizeof(*bi));
649 return 1;
650 }
651 if (key.type == BTRFS_METADATA_ITEM_KEY &&
652 item_size <= sizeof(*ei)) {
653 WARN_ON(item_size < sizeof(*ei));
654 return 1;
655 }
656
657 if (key.type == BTRFS_EXTENT_ITEM_KEY) {
658 bi = (struct btrfs_tree_block_info *)(ei + 1);
659 *ptr = (unsigned long)(bi + 1);
660 } else {
661 *ptr = (unsigned long)(ei + 1);
662 }
663 *end = (unsigned long)ei + item_size;
664 return 0;
665 }
666
667 /*
668 * build backref tree for a given tree block. root of the backref tree
669 * corresponds the tree block, leaves of the backref tree correspond
670 * roots of b-trees that reference the tree block.
671 *
672 * the basic idea of this function is check backrefs of a given block
673 * to find upper level blocks that reference the block, and then check
674 * backrefs of these upper level blocks recursively. the recursion stop
675 * when tree root is reached or backrefs for the block is cached.
676 *
677 * NOTE: if we find backrefs for a block are cached, we know backrefs
678 * for all upper level blocks that directly/indirectly reference the
679 * block are also cached.
680 */
681 static noinline_for_stack
682 struct backref_node *build_backref_tree(struct reloc_control *rc,
683 struct btrfs_key *node_key,
684 int level, u64 bytenr)
685 {
686 struct backref_cache *cache = &rc->backref_cache;
687 struct btrfs_path *path1;
688 struct btrfs_path *path2;
689 struct extent_buffer *eb;
690 struct btrfs_root *root;
691 struct backref_node *cur;
692 struct backref_node *upper;
693 struct backref_node *lower;
694 struct backref_node *node = NULL;
695 struct backref_node *exist = NULL;
696 struct backref_edge *edge;
697 struct rb_node *rb_node;
698 struct btrfs_key key;
699 unsigned long end;
700 unsigned long ptr;
701 LIST_HEAD(list);
702 LIST_HEAD(useless);
703 int cowonly;
704 int ret;
705 int err = 0;
706 bool need_check = true;
707
708 path1 = btrfs_alloc_path();
709 path2 = btrfs_alloc_path();
710 if (!path1 || !path2) {
711 err = -ENOMEM;
712 goto out;
713 }
714 path1->reada = READA_FORWARD;
715 path2->reada = READA_FORWARD;
716
717 node = alloc_backref_node(cache);
718 if (!node) {
719 err = -ENOMEM;
720 goto out;
721 }
722
723 node->bytenr = bytenr;
724 node->level = level;
725 node->lowest = 1;
726 cur = node;
727 again:
728 end = 0;
729 ptr = 0;
730 key.objectid = cur->bytenr;
731 key.type = BTRFS_METADATA_ITEM_KEY;
732 key.offset = (u64)-1;
733
734 path1->search_commit_root = 1;
735 path1->skip_locking = 1;
736 ret = btrfs_search_slot(NULL, rc->extent_root, &key, path1,
737 0, 0);
738 if (ret < 0) {
739 err = ret;
740 goto out;
741 }
742 ASSERT(ret);
743 ASSERT(path1->slots[0]);
744
745 path1->slots[0]--;
746
747 WARN_ON(cur->checked);
748 if (!list_empty(&cur->upper)) {
749 /*
750 * the backref was added previously when processing
751 * backref of type BTRFS_TREE_BLOCK_REF_KEY
752 */
753 ASSERT(list_is_singular(&cur->upper));
754 edge = list_entry(cur->upper.next, struct backref_edge,
755 list[LOWER]);
756 ASSERT(list_empty(&edge->list[UPPER]));
757 exist = edge->node[UPPER];
758 /*
759 * add the upper level block to pending list if we need
760 * check its backrefs
761 */
762 if (!exist->checked)
763 list_add_tail(&edge->list[UPPER], &list);
764 } else {
765 exist = NULL;
766 }
767
768 while (1) {
769 cond_resched();
770 eb = path1->nodes[0];
771
772 if (ptr >= end) {
773 if (path1->slots[0] >= btrfs_header_nritems(eb)) {
774 ret = btrfs_next_leaf(rc->extent_root, path1);
775 if (ret < 0) {
776 err = ret;
777 goto out;
778 }
779 if (ret > 0)
780 break;
781 eb = path1->nodes[0];
782 }
783
784 btrfs_item_key_to_cpu(eb, &key, path1->slots[0]);
785 if (key.objectid != cur->bytenr) {
786 WARN_ON(exist);
787 break;
788 }
789
790 if (key.type == BTRFS_EXTENT_ITEM_KEY ||
791 key.type == BTRFS_METADATA_ITEM_KEY) {
792 ret = find_inline_backref(eb, path1->slots[0],
793 &ptr, &end);
794 if (ret)
795 goto next;
796 }
797 }
798
799 if (ptr < end) {
800 /* update key for inline back ref */
801 struct btrfs_extent_inline_ref *iref;
802 iref = (struct btrfs_extent_inline_ref *)ptr;
803 key.type = btrfs_extent_inline_ref_type(eb, iref);
804 key.offset = btrfs_extent_inline_ref_offset(eb, iref);
805 WARN_ON(key.type != BTRFS_TREE_BLOCK_REF_KEY &&
806 key.type != BTRFS_SHARED_BLOCK_REF_KEY);
807 }
808
809 if (exist &&
810 ((key.type == BTRFS_TREE_BLOCK_REF_KEY &&
811 exist->owner == key.offset) ||
812 (key.type == BTRFS_SHARED_BLOCK_REF_KEY &&
813 exist->bytenr == key.offset))) {
814 exist = NULL;
815 goto next;
816 }
817
818 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
819 if (key.type == BTRFS_SHARED_BLOCK_REF_KEY ||
820 key.type == BTRFS_EXTENT_REF_V0_KEY) {
821 if (key.type == BTRFS_EXTENT_REF_V0_KEY) {
822 struct btrfs_extent_ref_v0 *ref0;
823 ref0 = btrfs_item_ptr(eb, path1->slots[0],
824 struct btrfs_extent_ref_v0);
825 if (key.objectid == key.offset) {
826 root = find_tree_root(rc, eb, ref0);
827 if (root && !should_ignore_root(root))
828 cur->root = root;
829 else
830 list_add(&cur->list, &useless);
831 break;
832 }
833 if (is_cowonly_root(btrfs_ref_root_v0(eb,
834 ref0)))
835 cur->cowonly = 1;
836 }
837 #else
838 ASSERT(key.type != BTRFS_EXTENT_REF_V0_KEY);
839 if (key.type == BTRFS_SHARED_BLOCK_REF_KEY) {
840 #endif
841 if (key.objectid == key.offset) {
842 /*
843 * only root blocks of reloc trees use
844 * backref of this type.
845 */
846 root = find_reloc_root(rc, cur->bytenr);
847 ASSERT(root);
848 cur->root = root;
849 break;
850 }
851
852 edge = alloc_backref_edge(cache);
853 if (!edge) {
854 err = -ENOMEM;
855 goto out;
856 }
857 rb_node = tree_search(&cache->rb_root, key.offset);
858 if (!rb_node) {
859 upper = alloc_backref_node(cache);
860 if (!upper) {
861 free_backref_edge(cache, edge);
862 err = -ENOMEM;
863 goto out;
864 }
865 upper->bytenr = key.offset;
866 upper->level = cur->level + 1;
867 /*
868 * backrefs for the upper level block isn't
869 * cached, add the block to pending list
870 */
871 list_add_tail(&edge->list[UPPER], &list);
872 } else {
873 upper = rb_entry(rb_node, struct backref_node,
874 rb_node);
875 ASSERT(upper->checked);
876 INIT_LIST_HEAD(&edge->list[UPPER]);
877 }
878 list_add_tail(&edge->list[LOWER], &cur->upper);
879 edge->node[LOWER] = cur;
880 edge->node[UPPER] = upper;
881
882 goto next;
883 } else if (key.type != BTRFS_TREE_BLOCK_REF_KEY) {
884 goto next;
885 }
886
887 /* key.type == BTRFS_TREE_BLOCK_REF_KEY */
888 root = read_fs_root(rc->extent_root->fs_info, key.offset);
889 if (IS_ERR(root)) {
890 err = PTR_ERR(root);
891 goto out;
892 }
893
894 if (!test_bit(BTRFS_ROOT_REF_COWS, &root->state))
895 cur->cowonly = 1;
896
897 if (btrfs_root_level(&root->root_item) == cur->level) {
898 /* tree root */
899 ASSERT(btrfs_root_bytenr(&root->root_item) ==
900 cur->bytenr);
901 if (should_ignore_root(root))
902 list_add(&cur->list, &useless);
903 else
904 cur->root = root;
905 break;
906 }
907
908 level = cur->level + 1;
909
910 /*
911 * searching the tree to find upper level blocks
912 * reference the block.
913 */
914 path2->search_commit_root = 1;
915 path2->skip_locking = 1;
916 path2->lowest_level = level;
917 ret = btrfs_search_slot(NULL, root, node_key, path2, 0, 0);
918 path2->lowest_level = 0;
919 if (ret < 0) {
920 err = ret;
921 goto out;
922 }
923 if (ret > 0 && path2->slots[level] > 0)
924 path2->slots[level]--;
925
926 eb = path2->nodes[level];
927 if (btrfs_node_blockptr(eb, path2->slots[level]) !=
928 cur->bytenr) {
929 btrfs_err(root->fs_info,
930 "couldn't find block (%llu) (level %d) in tree (%llu) with key (%llu %u %llu)",
931 cur->bytenr, level - 1, root->objectid,
932 node_key->objectid, node_key->type,
933 node_key->offset);
934 err = -ENOENT;
935 goto out;
936 }
937 lower = cur;
938 need_check = true;
939 for (; level < BTRFS_MAX_LEVEL; level++) {
940 if (!path2->nodes[level]) {
941 ASSERT(btrfs_root_bytenr(&root->root_item) ==
942 lower->bytenr);
943 if (should_ignore_root(root))
944 list_add(&lower->list, &useless);
945 else
946 lower->root = root;
947 break;
948 }
949
950 edge = alloc_backref_edge(cache);
951 if (!edge) {
952 err = -ENOMEM;
953 goto out;
954 }
955
956 eb = path2->nodes[level];
957 rb_node = tree_search(&cache->rb_root, eb->start);
958 if (!rb_node) {
959 upper = alloc_backref_node(cache);
960 if (!upper) {
961 free_backref_edge(cache, edge);
962 err = -ENOMEM;
963 goto out;
964 }
965 upper->bytenr = eb->start;
966 upper->owner = btrfs_header_owner(eb);
967 upper->level = lower->level + 1;
968 if (!test_bit(BTRFS_ROOT_REF_COWS,
969 &root->state))
970 upper->cowonly = 1;
971
972 /*
973 * if we know the block isn't shared
974 * we can void checking its backrefs.
975 */
976 if (btrfs_block_can_be_shared(root, eb))
977 upper->checked = 0;
978 else
979 upper->checked = 1;
980
981 /*
982 * add the block to pending list if we
983 * need check its backrefs, we only do this once
984 * while walking up a tree as we will catch
985 * anything else later on.
986 */
987 if (!upper->checked && need_check) {
988 need_check = false;
989 list_add_tail(&edge->list[UPPER],
990 &list);
991 } else {
992 if (upper->checked)
993 need_check = true;
994 INIT_LIST_HEAD(&edge->list[UPPER]);
995 }
996 } else {
997 upper = rb_entry(rb_node, struct backref_node,
998 rb_node);
999 ASSERT(upper->checked);
1000 INIT_LIST_HEAD(&edge->list[UPPER]);
1001 if (!upper->owner)
1002 upper->owner = btrfs_header_owner(eb);
1003 }
1004 list_add_tail(&edge->list[LOWER], &lower->upper);
1005 edge->node[LOWER] = lower;
1006 edge->node[UPPER] = upper;
1007
1008 if (rb_node)
1009 break;
1010 lower = upper;
1011 upper = NULL;
1012 }
1013 btrfs_release_path(path2);
1014 next:
1015 if (ptr < end) {
1016 ptr += btrfs_extent_inline_ref_size(key.type);
1017 if (ptr >= end) {
1018 WARN_ON(ptr > end);
1019 ptr = 0;
1020 end = 0;
1021 }
1022 }
1023 if (ptr >= end)
1024 path1->slots[0]++;
1025 }
1026 btrfs_release_path(path1);
1027
1028 cur->checked = 1;
1029 WARN_ON(exist);
1030
1031 /* the pending list isn't empty, take the first block to process */
1032 if (!list_empty(&list)) {
1033 edge = list_entry(list.next, struct backref_edge, list[UPPER]);
1034 list_del_init(&edge->list[UPPER]);
1035 cur = edge->node[UPPER];
1036 goto again;
1037 }
1038
1039 /*
1040 * everything goes well, connect backref nodes and insert backref nodes
1041 * into the cache.
1042 */
1043 ASSERT(node->checked);
1044 cowonly = node->cowonly;
1045 if (!cowonly) {
1046 rb_node = tree_insert(&cache->rb_root, node->bytenr,
1047 &node->rb_node);
1048 if (rb_node)
1049 backref_tree_panic(rb_node, -EEXIST, node->bytenr);
1050 list_add_tail(&node->lower, &cache->leaves);
1051 }
1052
1053 list_for_each_entry(edge, &node->upper, list[LOWER])
1054 list_add_tail(&edge->list[UPPER], &list);
1055
1056 while (!list_empty(&list)) {
1057 edge = list_entry(list.next, struct backref_edge, list[UPPER]);
1058 list_del_init(&edge->list[UPPER]);
1059 upper = edge->node[UPPER];
1060 if (upper->detached) {
1061 list_del(&edge->list[LOWER]);
1062 lower = edge->node[LOWER];
1063 free_backref_edge(cache, edge);
1064 if (list_empty(&lower->upper))
1065 list_add(&lower->list, &useless);
1066 continue;
1067 }
1068
1069 if (!RB_EMPTY_NODE(&upper->rb_node)) {
1070 if (upper->lowest) {
1071 list_del_init(&upper->lower);
1072 upper->lowest = 0;
1073 }
1074
1075 list_add_tail(&edge->list[UPPER], &upper->lower);
1076 continue;
1077 }
1078
1079 if (!upper->checked) {
1080 /*
1081 * Still want to blow up for developers since this is a
1082 * logic bug.
1083 */
1084 ASSERT(0);
1085 err = -EINVAL;
1086 goto out;
1087 }
1088 if (cowonly != upper->cowonly) {
1089 ASSERT(0);
1090 err = -EINVAL;
1091 goto out;
1092 }
1093
1094 if (!cowonly) {
1095 rb_node = tree_insert(&cache->rb_root, upper->bytenr,
1096 &upper->rb_node);
1097 if (rb_node)
1098 backref_tree_panic(rb_node, -EEXIST,
1099 upper->bytenr);
1100 }
1101
1102 list_add_tail(&edge->list[UPPER], &upper->lower);
1103
1104 list_for_each_entry(edge, &upper->upper, list[LOWER])
1105 list_add_tail(&edge->list[UPPER], &list);
1106 }
1107 /*
1108 * process useless backref nodes. backref nodes for tree leaves
1109 * are deleted from the cache. backref nodes for upper level
1110 * tree blocks are left in the cache to avoid unnecessary backref
1111 * lookup.
1112 */
1113 while (!list_empty(&useless)) {
1114 upper = list_entry(useless.next, struct backref_node, list);
1115 list_del_init(&upper->list);
1116 ASSERT(list_empty(&upper->upper));
1117 if (upper == node)
1118 node = NULL;
1119 if (upper->lowest) {
1120 list_del_init(&upper->lower);
1121 upper->lowest = 0;
1122 }
1123 while (!list_empty(&upper->lower)) {
1124 edge = list_entry(upper->lower.next,
1125 struct backref_edge, list[UPPER]);
1126 list_del(&edge->list[UPPER]);
1127 list_del(&edge->list[LOWER]);
1128 lower = edge->node[LOWER];
1129 free_backref_edge(cache, edge);
1130
1131 if (list_empty(&lower->upper))
1132 list_add(&lower->list, &useless);
1133 }
1134 __mark_block_processed(rc, upper);
1135 if (upper->level > 0) {
1136 list_add(&upper->list, &cache->detached);
1137 upper->detached = 1;
1138 } else {
1139 rb_erase(&upper->rb_node, &cache->rb_root);
1140 free_backref_node(cache, upper);
1141 }
1142 }
1143 out:
1144 btrfs_free_path(path1);
1145 btrfs_free_path(path2);
1146 if (err) {
1147 while (!list_empty(&useless)) {
1148 lower = list_entry(useless.next,
1149 struct backref_node, list);
1150 list_del_init(&lower->list);
1151 }
1152 while (!list_empty(&list)) {
1153 edge = list_first_entry(&list, struct backref_edge,
1154 list[UPPER]);
1155 list_del(&edge->list[UPPER]);
1156 list_del(&edge->list[LOWER]);
1157 lower = edge->node[LOWER];
1158 upper = edge->node[UPPER];
1159 free_backref_edge(cache, edge);
1160
1161 /*
1162 * Lower is no longer linked to any upper backref nodes
1163 * and isn't in the cache, we can free it ourselves.
1164 */
1165 if (list_empty(&lower->upper) &&
1166 RB_EMPTY_NODE(&lower->rb_node))
1167 list_add(&lower->list, &useless);
1168
1169 if (!RB_EMPTY_NODE(&upper->rb_node))
1170 continue;
1171
1172 /* Add this guy's upper edges to the list to process */
1173 list_for_each_entry(edge, &upper->upper, list[LOWER])
1174 list_add_tail(&edge->list[UPPER], &list);
1175 if (list_empty(&upper->upper))
1176 list_add(&upper->list, &useless);
1177 }
1178
1179 while (!list_empty(&useless)) {
1180 lower = list_entry(useless.next,
1181 struct backref_node, list);
1182 list_del_init(&lower->list);
1183 if (lower == node)
1184 node = NULL;
1185 free_backref_node(cache, lower);
1186 }
1187
1188 free_backref_node(cache, node);
1189 return ERR_PTR(err);
1190 }
1191 ASSERT(!node || !node->detached);
1192 return node;
1193 }
1194
1195 /*
1196 * helper to add backref node for the newly created snapshot.
1197 * the backref node is created by cloning backref node that
1198 * corresponds to root of source tree
1199 */
1200 static int clone_backref_node(struct btrfs_trans_handle *trans,
1201 struct reloc_control *rc,
1202 struct btrfs_root *src,
1203 struct btrfs_root *dest)
1204 {
1205 struct btrfs_root *reloc_root = src->reloc_root;
1206 struct backref_cache *cache = &rc->backref_cache;
1207 struct backref_node *node = NULL;
1208 struct backref_node *new_node;
1209 struct backref_edge *edge;
1210 struct backref_edge *new_edge;
1211 struct rb_node *rb_node;
1212
1213 if (cache->last_trans > 0)
1214 update_backref_cache(trans, cache);
1215
1216 rb_node = tree_search(&cache->rb_root, src->commit_root->start);
1217 if (rb_node) {
1218 node = rb_entry(rb_node, struct backref_node, rb_node);
1219 if (node->detached)
1220 node = NULL;
1221 else
1222 BUG_ON(node->new_bytenr != reloc_root->node->start);
1223 }
1224
1225 if (!node) {
1226 rb_node = tree_search(&cache->rb_root,
1227 reloc_root->commit_root->start);
1228 if (rb_node) {
1229 node = rb_entry(rb_node, struct backref_node,
1230 rb_node);
1231 BUG_ON(node->detached);
1232 }
1233 }
1234
1235 if (!node)
1236 return 0;
1237
1238 new_node = alloc_backref_node(cache);
1239 if (!new_node)
1240 return -ENOMEM;
1241
1242 new_node->bytenr = dest->node->start;
1243 new_node->level = node->level;
1244 new_node->lowest = node->lowest;
1245 new_node->checked = 1;
1246 new_node->root = dest;
1247
1248 if (!node->lowest) {
1249 list_for_each_entry(edge, &node->lower, list[UPPER]) {
1250 new_edge = alloc_backref_edge(cache);
1251 if (!new_edge)
1252 goto fail;
1253
1254 new_edge->node[UPPER] = new_node;
1255 new_edge->node[LOWER] = edge->node[LOWER];
1256 list_add_tail(&new_edge->list[UPPER],
1257 &new_node->lower);
1258 }
1259 } else {
1260 list_add_tail(&new_node->lower, &cache->leaves);
1261 }
1262
1263 rb_node = tree_insert(&cache->rb_root, new_node->bytenr,
1264 &new_node->rb_node);
1265 if (rb_node)
1266 backref_tree_panic(rb_node, -EEXIST, new_node->bytenr);
1267
1268 if (!new_node->lowest) {
1269 list_for_each_entry(new_edge, &new_node->lower, list[UPPER]) {
1270 list_add_tail(&new_edge->list[LOWER],
1271 &new_edge->node[LOWER]->upper);
1272 }
1273 }
1274 return 0;
1275 fail:
1276 while (!list_empty(&new_node->lower)) {
1277 new_edge = list_entry(new_node->lower.next,
1278 struct backref_edge, list[UPPER]);
1279 list_del(&new_edge->list[UPPER]);
1280 free_backref_edge(cache, new_edge);
1281 }
1282 free_backref_node(cache, new_node);
1283 return -ENOMEM;
1284 }
1285
1286 /*
1287 * helper to add 'address of tree root -> reloc tree' mapping
1288 */
1289 static int __must_check __add_reloc_root(struct btrfs_root *root)
1290 {
1291 struct rb_node *rb_node;
1292 struct mapping_node *node;
1293 struct reloc_control *rc = root->fs_info->reloc_ctl;
1294
1295 node = kmalloc(sizeof(*node), GFP_NOFS);
1296 if (!node)
1297 return -ENOMEM;
1298
1299 node->bytenr = root->node->start;
1300 node->data = root;
1301
1302 spin_lock(&rc->reloc_root_tree.lock);
1303 rb_node = tree_insert(&rc->reloc_root_tree.rb_root,
1304 node->bytenr, &node->rb_node);
1305 spin_unlock(&rc->reloc_root_tree.lock);
1306 if (rb_node) {
1307 btrfs_panic(root->fs_info, -EEXIST,
1308 "Duplicate root found for start=%llu while inserting into relocation tree",
1309 node->bytenr);
1310 kfree(node);
1311 return -EEXIST;
1312 }
1313
1314 list_add_tail(&root->root_list, &rc->reloc_roots);
1315 return 0;
1316 }
1317
1318 /*
1319 * helper to delete the 'address of tree root -> reloc tree'
1320 * mapping
1321 */
1322 static void __del_reloc_root(struct btrfs_root *root)
1323 {
1324 struct rb_node *rb_node;
1325 struct mapping_node *node = NULL;
1326 struct reloc_control *rc = root->fs_info->reloc_ctl;
1327
1328 if (rc && root->node) {
1329 spin_lock(&rc->reloc_root_tree.lock);
1330 rb_node = tree_search(&rc->reloc_root_tree.rb_root,
1331 root->node->start);
1332 if (rb_node) {
1333 node = rb_entry(rb_node, struct mapping_node, rb_node);
1334 rb_erase(&node->rb_node, &rc->reloc_root_tree.rb_root);
1335 }
1336 spin_unlock(&rc->reloc_root_tree.lock);
1337 if (!node)
1338 return;
1339 BUG_ON((struct btrfs_root *)node->data != root);
1340 }
1341
1342 spin_lock(&root->fs_info->trans_lock);
1343 list_del_init(&root->root_list);
1344 spin_unlock(&root->fs_info->trans_lock);
1345 kfree(node);
1346 }
1347
1348 /*
1349 * helper to update the 'address of tree root -> reloc tree'
1350 * mapping
1351 */
1352 static int __update_reloc_root(struct btrfs_root *root, u64 new_bytenr)
1353 {
1354 struct rb_node *rb_node;
1355 struct mapping_node *node = NULL;
1356 struct reloc_control *rc = root->fs_info->reloc_ctl;
1357
1358 spin_lock(&rc->reloc_root_tree.lock);
1359 rb_node = tree_search(&rc->reloc_root_tree.rb_root,
1360 root->node->start);
1361 if (rb_node) {
1362 node = rb_entry(rb_node, struct mapping_node, rb_node);
1363 rb_erase(&node->rb_node, &rc->reloc_root_tree.rb_root);
1364 }
1365 spin_unlock(&rc->reloc_root_tree.lock);
1366
1367 if (!node)
1368 return 0;
1369 BUG_ON((struct btrfs_root *)node->data != root);
1370
1371 spin_lock(&rc->reloc_root_tree.lock);
1372 node->bytenr = new_bytenr;
1373 rb_node = tree_insert(&rc->reloc_root_tree.rb_root,
1374 node->bytenr, &node->rb_node);
1375 spin_unlock(&rc->reloc_root_tree.lock);
1376 if (rb_node)
1377 backref_tree_panic(rb_node, -EEXIST, node->bytenr);
1378 return 0;
1379 }
1380
1381 static struct btrfs_root *create_reloc_root(struct btrfs_trans_handle *trans,
1382 struct btrfs_root *root, u64 objectid)
1383 {
1384 struct btrfs_root *reloc_root;
1385 struct extent_buffer *eb;
1386 struct btrfs_root_item *root_item;
1387 struct btrfs_key root_key;
1388 u64 last_snap = 0;
1389 int ret;
1390
1391 root_item = kmalloc(sizeof(*root_item), GFP_NOFS);
1392 BUG_ON(!root_item);
1393
1394 root_key.objectid = BTRFS_TREE_RELOC_OBJECTID;
1395 root_key.type = BTRFS_ROOT_ITEM_KEY;
1396 root_key.offset = objectid;
1397
1398 if (root->root_key.objectid == objectid) {
1399 u64 commit_root_gen;
1400
1401 /* called by btrfs_init_reloc_root */
1402 ret = btrfs_copy_root(trans, root, root->commit_root, &eb,
1403 BTRFS_TREE_RELOC_OBJECTID);
1404 BUG_ON(ret);
1405 last_snap = btrfs_root_last_snapshot(&root->root_item);
1406 /*
1407 * Set the last_snapshot field to the generation of the commit
1408 * root - like this ctree.c:btrfs_block_can_be_shared() behaves
1409 * correctly (returns true) when the relocation root is created
1410 * either inside the critical section of a transaction commit
1411 * (through transaction.c:qgroup_account_snapshot()) and when
1412 * it's created before the transaction commit is started.
1413 */
1414 commit_root_gen = btrfs_header_generation(root->commit_root);
1415 btrfs_set_root_last_snapshot(&root->root_item, commit_root_gen);
1416 } else {
1417 /*
1418 * called by btrfs_reloc_post_snapshot_hook.
1419 * the source tree is a reloc tree, all tree blocks
1420 * modified after it was created have RELOC flag
1421 * set in their headers. so it's OK to not update
1422 * the 'last_snapshot'.
1423 */
1424 ret = btrfs_copy_root(trans, root, root->node, &eb,
1425 BTRFS_TREE_RELOC_OBJECTID);
1426 BUG_ON(ret);
1427 }
1428
1429 memcpy(root_item, &root->root_item, sizeof(*root_item));
1430 btrfs_set_root_bytenr(root_item, eb->start);
1431 btrfs_set_root_level(root_item, btrfs_header_level(eb));
1432 btrfs_set_root_generation(root_item, trans->transid);
1433
1434 if (root->root_key.objectid == objectid) {
1435 btrfs_set_root_refs(root_item, 0);
1436 memset(&root_item->drop_progress, 0,
1437 sizeof(struct btrfs_disk_key));
1438 root_item->drop_level = 0;
1439 /*
1440 * abuse rtransid, it is safe because it is impossible to
1441 * receive data into a relocation tree.
1442 */
1443 btrfs_set_root_rtransid(root_item, last_snap);
1444 btrfs_set_root_otransid(root_item, trans->transid);
1445 }
1446
1447 btrfs_tree_unlock(eb);
1448 free_extent_buffer(eb);
1449
1450 ret = btrfs_insert_root(trans, root->fs_info->tree_root,
1451 &root_key, root_item);
1452 BUG_ON(ret);
1453 kfree(root_item);
1454
1455 reloc_root = btrfs_read_fs_root(root->fs_info->tree_root, &root_key);
1456 BUG_ON(IS_ERR(reloc_root));
1457 reloc_root->last_trans = trans->transid;
1458 return reloc_root;
1459 }
1460
1461 /*
1462 * create reloc tree for a given fs tree. reloc tree is just a
1463 * snapshot of the fs tree with special root objectid.
1464 */
1465 int btrfs_init_reloc_root(struct btrfs_trans_handle *trans,
1466 struct btrfs_root *root)
1467 {
1468 struct btrfs_root *reloc_root;
1469 struct reloc_control *rc = root->fs_info->reloc_ctl;
1470 struct btrfs_block_rsv *rsv;
1471 int clear_rsv = 0;
1472 int ret;
1473
1474 if (root->reloc_root) {
1475 reloc_root = root->reloc_root;
1476 reloc_root->last_trans = trans->transid;
1477 return 0;
1478 }
1479
1480 if (!rc || !rc->create_reloc_tree ||
1481 root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID)
1482 return 0;
1483
1484 if (!trans->reloc_reserved) {
1485 rsv = trans->block_rsv;
1486 trans->block_rsv = rc->block_rsv;
1487 clear_rsv = 1;
1488 }
1489 reloc_root = create_reloc_root(trans, root, root->root_key.objectid);
1490 if (clear_rsv)
1491 trans->block_rsv = rsv;
1492
1493 ret = __add_reloc_root(reloc_root);
1494 BUG_ON(ret < 0);
1495 root->reloc_root = reloc_root;
1496 return 0;
1497 }
1498
1499 /*
1500 * update root item of reloc tree
1501 */
1502 int btrfs_update_reloc_root(struct btrfs_trans_handle *trans,
1503 struct btrfs_root *root)
1504 {
1505 struct btrfs_root *reloc_root;
1506 struct btrfs_root_item *root_item;
1507 int ret;
1508
1509 if (!root->reloc_root)
1510 goto out;
1511
1512 reloc_root = root->reloc_root;
1513 root_item = &reloc_root->root_item;
1514
1515 if (root->fs_info->reloc_ctl->merge_reloc_tree &&
1516 btrfs_root_refs(root_item) == 0) {
1517 root->reloc_root = NULL;
1518 __del_reloc_root(reloc_root);
1519 }
1520
1521 if (reloc_root->commit_root != reloc_root->node) {
1522 btrfs_set_root_node(root_item, reloc_root->node);
1523 free_extent_buffer(reloc_root->commit_root);
1524 reloc_root->commit_root = btrfs_root_node(reloc_root);
1525 }
1526
1527 ret = btrfs_update_root(trans, root->fs_info->tree_root,
1528 &reloc_root->root_key, root_item);
1529 BUG_ON(ret);
1530
1531 out:
1532 return 0;
1533 }
1534
1535 /*
1536 * helper to find first cached inode with inode number >= objectid
1537 * in a subvolume
1538 */
1539 static struct inode *find_next_inode(struct btrfs_root *root, u64 objectid)
1540 {
1541 struct rb_node *node;
1542 struct rb_node *prev;
1543 struct btrfs_inode *entry;
1544 struct inode *inode;
1545
1546 spin_lock(&root->inode_lock);
1547 again:
1548 node = root->inode_tree.rb_node;
1549 prev = NULL;
1550 while (node) {
1551 prev = node;
1552 entry = rb_entry(node, struct btrfs_inode, rb_node);
1553
1554 if (objectid < btrfs_ino(&entry->vfs_inode))
1555 node = node->rb_left;
1556 else if (objectid > btrfs_ino(&entry->vfs_inode))
1557 node = node->rb_right;
1558 else
1559 break;
1560 }
1561 if (!node) {
1562 while (prev) {
1563 entry = rb_entry(prev, struct btrfs_inode, rb_node);
1564 if (objectid <= btrfs_ino(&entry->vfs_inode)) {
1565 node = prev;
1566 break;
1567 }
1568 prev = rb_next(prev);
1569 }
1570 }
1571 while (node) {
1572 entry = rb_entry(node, struct btrfs_inode, rb_node);
1573 inode = igrab(&entry->vfs_inode);
1574 if (inode) {
1575 spin_unlock(&root->inode_lock);
1576 return inode;
1577 }
1578
1579 objectid = btrfs_ino(&entry->vfs_inode) + 1;
1580 if (cond_resched_lock(&root->inode_lock))
1581 goto again;
1582
1583 node = rb_next(node);
1584 }
1585 spin_unlock(&root->inode_lock);
1586 return NULL;
1587 }
1588
1589 static int in_block_group(u64 bytenr,
1590 struct btrfs_block_group_cache *block_group)
1591 {
1592 if (bytenr >= block_group->key.objectid &&
1593 bytenr < block_group->key.objectid + block_group->key.offset)
1594 return 1;
1595 return 0;
1596 }
1597
1598 /*
1599 * get new location of data
1600 */
1601 static int get_new_location(struct inode *reloc_inode, u64 *new_bytenr,
1602 u64 bytenr, u64 num_bytes)
1603 {
1604 struct btrfs_root *root = BTRFS_I(reloc_inode)->root;
1605 struct btrfs_path *path;
1606 struct btrfs_file_extent_item *fi;
1607 struct extent_buffer *leaf;
1608 int ret;
1609
1610 path = btrfs_alloc_path();
1611 if (!path)
1612 return -ENOMEM;
1613
1614 bytenr -= BTRFS_I(reloc_inode)->index_cnt;
1615 ret = btrfs_lookup_file_extent(NULL, root, path, btrfs_ino(reloc_inode),
1616 bytenr, 0);
1617 if (ret < 0)
1618 goto out;
1619 if (ret > 0) {
1620 ret = -ENOENT;
1621 goto out;
1622 }
1623
1624 leaf = path->nodes[0];
1625 fi = btrfs_item_ptr(leaf, path->slots[0],
1626 struct btrfs_file_extent_item);
1627
1628 BUG_ON(btrfs_file_extent_offset(leaf, fi) ||
1629 btrfs_file_extent_compression(leaf, fi) ||
1630 btrfs_file_extent_encryption(leaf, fi) ||
1631 btrfs_file_extent_other_encoding(leaf, fi));
1632
1633 if (num_bytes != btrfs_file_extent_disk_num_bytes(leaf, fi)) {
1634 ret = -EINVAL;
1635 goto out;
1636 }
1637
1638 *new_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
1639 ret = 0;
1640 out:
1641 btrfs_free_path(path);
1642 return ret;
1643 }
1644
1645 /*
1646 * update file extent items in the tree leaf to point to
1647 * the new locations.
1648 */
1649 static noinline_for_stack
1650 int replace_file_extents(struct btrfs_trans_handle *trans,
1651 struct reloc_control *rc,
1652 struct btrfs_root *root,
1653 struct extent_buffer *leaf)
1654 {
1655 struct btrfs_key key;
1656 struct btrfs_file_extent_item *fi;
1657 struct inode *inode = NULL;
1658 u64 parent;
1659 u64 bytenr;
1660 u64 new_bytenr = 0;
1661 u64 num_bytes;
1662 u64 end;
1663 u32 nritems;
1664 u32 i;
1665 int ret = 0;
1666 int first = 1;
1667 int dirty = 0;
1668
1669 if (rc->stage != UPDATE_DATA_PTRS)
1670 return 0;
1671
1672 /* reloc trees always use full backref */
1673 if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID)
1674 parent = leaf->start;
1675 else
1676 parent = 0;
1677
1678 nritems = btrfs_header_nritems(leaf);
1679 for (i = 0; i < nritems; i++) {
1680 cond_resched();
1681 btrfs_item_key_to_cpu(leaf, &key, i);
1682 if (key.type != BTRFS_EXTENT_DATA_KEY)
1683 continue;
1684 fi = btrfs_item_ptr(leaf, i, struct btrfs_file_extent_item);
1685 if (btrfs_file_extent_type(leaf, fi) ==
1686 BTRFS_FILE_EXTENT_INLINE)
1687 continue;
1688 bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
1689 num_bytes = btrfs_file_extent_disk_num_bytes(leaf, fi);
1690 if (bytenr == 0)
1691 continue;
1692 if (!in_block_group(bytenr, rc->block_group))
1693 continue;
1694
1695 /*
1696 * if we are modifying block in fs tree, wait for readpage
1697 * to complete and drop the extent cache
1698 */
1699 if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) {
1700 if (first) {
1701 inode = find_next_inode(root, key.objectid);
1702 first = 0;
1703 } else if (inode && btrfs_ino(inode) < key.objectid) {
1704 btrfs_add_delayed_iput(inode);
1705 inode = find_next_inode(root, key.objectid);
1706 }
1707 if (inode && btrfs_ino(inode) == key.objectid) {
1708 end = key.offset +
1709 btrfs_file_extent_num_bytes(leaf, fi);
1710 WARN_ON(!IS_ALIGNED(key.offset,
1711 root->sectorsize));
1712 WARN_ON(!IS_ALIGNED(end, root->sectorsize));
1713 end--;
1714 ret = try_lock_extent(&BTRFS_I(inode)->io_tree,
1715 key.offset, end);
1716 if (!ret)
1717 continue;
1718
1719 btrfs_drop_extent_cache(inode, key.offset, end,
1720 1);
1721 unlock_extent(&BTRFS_I(inode)->io_tree,
1722 key.offset, end);
1723 }
1724 }
1725
1726 ret = get_new_location(rc->data_inode, &new_bytenr,
1727 bytenr, num_bytes);
1728 if (ret) {
1729 /*
1730 * Don't have to abort since we've not changed anything
1731 * in the file extent yet.
1732 */
1733 break;
1734 }
1735
1736 btrfs_set_file_extent_disk_bytenr(leaf, fi, new_bytenr);
1737 dirty = 1;
1738
1739 key.offset -= btrfs_file_extent_offset(leaf, fi);
1740 ret = btrfs_inc_extent_ref(trans, root, new_bytenr,
1741 num_bytes, parent,
1742 btrfs_header_owner(leaf),
1743 key.objectid, key.offset);
1744 if (ret) {
1745 btrfs_abort_transaction(trans, ret);
1746 break;
1747 }
1748
1749 ret = btrfs_free_extent(trans, root, bytenr, num_bytes,
1750 parent, btrfs_header_owner(leaf),
1751 key.objectid, key.offset);
1752 if (ret) {
1753 btrfs_abort_transaction(trans, ret);
1754 break;
1755 }
1756 }
1757 if (dirty)
1758 btrfs_mark_buffer_dirty(leaf);
1759 if (inode)
1760 btrfs_add_delayed_iput(inode);
1761 return ret;
1762 }
1763
1764 static noinline_for_stack
1765 int memcmp_node_keys(struct extent_buffer *eb, int slot,
1766 struct btrfs_path *path, int level)
1767 {
1768 struct btrfs_disk_key key1;
1769 struct btrfs_disk_key key2;
1770 btrfs_node_key(eb, &key1, slot);
1771 btrfs_node_key(path->nodes[level], &key2, path->slots[level]);
1772 return memcmp(&key1, &key2, sizeof(key1));
1773 }
1774
1775 /*
1776 * try to replace tree blocks in fs tree with the new blocks
1777 * in reloc tree. tree blocks haven't been modified since the
1778 * reloc tree was create can be replaced.
1779 *
1780 * if a block was replaced, level of the block + 1 is returned.
1781 * if no block got replaced, 0 is returned. if there are other
1782 * errors, a negative error number is returned.
1783 */
1784 static noinline_for_stack
1785 int replace_path(struct btrfs_trans_handle *trans,
1786 struct btrfs_root *dest, struct btrfs_root *src,
1787 struct btrfs_path *path, struct btrfs_key *next_key,
1788 int lowest_level, int max_level)
1789 {
1790 struct extent_buffer *eb;
1791 struct extent_buffer *parent;
1792 struct btrfs_key key;
1793 u64 old_bytenr;
1794 u64 new_bytenr;
1795 u64 old_ptr_gen;
1796 u64 new_ptr_gen;
1797 u64 last_snapshot;
1798 u32 blocksize;
1799 int cow = 0;
1800 int level;
1801 int ret;
1802 int slot;
1803
1804 BUG_ON(src->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID);
1805 BUG_ON(dest->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID);
1806
1807 last_snapshot = btrfs_root_last_snapshot(&src->root_item);
1808 again:
1809 slot = path->slots[lowest_level];
1810 btrfs_node_key_to_cpu(path->nodes[lowest_level], &key, slot);
1811
1812 eb = btrfs_lock_root_node(dest);
1813 btrfs_set_lock_blocking(eb);
1814 level = btrfs_header_level(eb);
1815
1816 if (level < lowest_level) {
1817 btrfs_tree_unlock(eb);
1818 free_extent_buffer(eb);
1819 return 0;
1820 }
1821
1822 if (cow) {
1823 ret = btrfs_cow_block(trans, dest, eb, NULL, 0, &eb);
1824 BUG_ON(ret);
1825 }
1826 btrfs_set_lock_blocking(eb);
1827
1828 if (next_key) {
1829 next_key->objectid = (u64)-1;
1830 next_key->type = (u8)-1;
1831 next_key->offset = (u64)-1;
1832 }
1833
1834 parent = eb;
1835 while (1) {
1836 level = btrfs_header_level(parent);
1837 BUG_ON(level < lowest_level);
1838
1839 ret = btrfs_bin_search(parent, &key, level, &slot);
1840 if (ret && slot > 0)
1841 slot--;
1842
1843 if (next_key && slot + 1 < btrfs_header_nritems(parent))
1844 btrfs_node_key_to_cpu(parent, next_key, slot + 1);
1845
1846 old_bytenr = btrfs_node_blockptr(parent, slot);
1847 blocksize = dest->nodesize;
1848 old_ptr_gen = btrfs_node_ptr_generation(parent, slot);
1849
1850 if (level <= max_level) {
1851 eb = path->nodes[level];
1852 new_bytenr = btrfs_node_blockptr(eb,
1853 path->slots[level]);
1854 new_ptr_gen = btrfs_node_ptr_generation(eb,
1855 path->slots[level]);
1856 } else {
1857 new_bytenr = 0;
1858 new_ptr_gen = 0;
1859 }
1860
1861 if (WARN_ON(new_bytenr > 0 && new_bytenr == old_bytenr)) {
1862 ret = level;
1863 break;
1864 }
1865
1866 if (new_bytenr == 0 || old_ptr_gen > last_snapshot ||
1867 memcmp_node_keys(parent, slot, path, level)) {
1868 if (level <= lowest_level) {
1869 ret = 0;
1870 break;
1871 }
1872
1873 eb = read_tree_block(dest, old_bytenr, old_ptr_gen);
1874 if (IS_ERR(eb)) {
1875 ret = PTR_ERR(eb);
1876 break;
1877 } else if (!extent_buffer_uptodate(eb)) {
1878 ret = -EIO;
1879 free_extent_buffer(eb);
1880 break;
1881 }
1882 btrfs_tree_lock(eb);
1883 if (cow) {
1884 ret = btrfs_cow_block(trans, dest, eb, parent,
1885 slot, &eb);
1886 BUG_ON(ret);
1887 }
1888 btrfs_set_lock_blocking(eb);
1889
1890 btrfs_tree_unlock(parent);
1891 free_extent_buffer(parent);
1892
1893 parent = eb;
1894 continue;
1895 }
1896
1897 if (!cow) {
1898 btrfs_tree_unlock(parent);
1899 free_extent_buffer(parent);
1900 cow = 1;
1901 goto again;
1902 }
1903
1904 btrfs_node_key_to_cpu(path->nodes[level], &key,
1905 path->slots[level]);
1906 btrfs_release_path(path);
1907
1908 path->lowest_level = level;
1909 ret = btrfs_search_slot(trans, src, &key, path, 0, 1);
1910 path->lowest_level = 0;
1911 BUG_ON(ret);
1912
1913 /*
1914 * swap blocks in fs tree and reloc tree.
1915 */
1916 btrfs_set_node_blockptr(parent, slot, new_bytenr);
1917 btrfs_set_node_ptr_generation(parent, slot, new_ptr_gen);
1918 btrfs_mark_buffer_dirty(parent);
1919
1920 btrfs_set_node_blockptr(path->nodes[level],
1921 path->slots[level], old_bytenr);
1922 btrfs_set_node_ptr_generation(path->nodes[level],
1923 path->slots[level], old_ptr_gen);
1924 btrfs_mark_buffer_dirty(path->nodes[level]);
1925
1926 ret = btrfs_inc_extent_ref(trans, src, old_bytenr, blocksize,
1927 path->nodes[level]->start,
1928 src->root_key.objectid, level - 1, 0);
1929 BUG_ON(ret);
1930 ret = btrfs_inc_extent_ref(trans, dest, new_bytenr, blocksize,
1931 0, dest->root_key.objectid, level - 1,
1932 0);
1933 BUG_ON(ret);
1934
1935 ret = btrfs_free_extent(trans, src, new_bytenr, blocksize,
1936 path->nodes[level]->start,
1937 src->root_key.objectid, level - 1, 0);
1938 BUG_ON(ret);
1939
1940 ret = btrfs_free_extent(trans, dest, old_bytenr, blocksize,
1941 0, dest->root_key.objectid, level - 1,
1942 0);
1943 BUG_ON(ret);
1944
1945 btrfs_unlock_up_safe(path, 0);
1946
1947 ret = level;
1948 break;
1949 }
1950 btrfs_tree_unlock(parent);
1951 free_extent_buffer(parent);
1952 return ret;
1953 }
1954
1955 /*
1956 * helper to find next relocated block in reloc tree
1957 */
1958 static noinline_for_stack
1959 int walk_up_reloc_tree(struct btrfs_root *root, struct btrfs_path *path,
1960 int *level)
1961 {
1962 struct extent_buffer *eb;
1963 int i;
1964 u64 last_snapshot;
1965 u32 nritems;
1966
1967 last_snapshot = btrfs_root_last_snapshot(&root->root_item);
1968
1969 for (i = 0; i < *level; i++) {
1970 free_extent_buffer(path->nodes[i]);
1971 path->nodes[i] = NULL;
1972 }
1973
1974 for (i = *level; i < BTRFS_MAX_LEVEL && path->nodes[i]; i++) {
1975 eb = path->nodes[i];
1976 nritems = btrfs_header_nritems(eb);
1977 while (path->slots[i] + 1 < nritems) {
1978 path->slots[i]++;
1979 if (btrfs_node_ptr_generation(eb, path->slots[i]) <=
1980 last_snapshot)
1981 continue;
1982
1983 *level = i;
1984 return 0;
1985 }
1986 free_extent_buffer(path->nodes[i]);
1987 path->nodes[i] = NULL;
1988 }
1989 return 1;
1990 }
1991
1992 /*
1993 * walk down reloc tree to find relocated block of lowest level
1994 */
1995 static noinline_for_stack
1996 int walk_down_reloc_tree(struct btrfs_root *root, struct btrfs_path *path,
1997 int *level)
1998 {
1999 struct extent_buffer *eb = NULL;
2000 int i;
2001 u64 bytenr;
2002 u64 ptr_gen = 0;
2003 u64 last_snapshot;
2004 u32 nritems;
2005
2006 last_snapshot = btrfs_root_last_snapshot(&root->root_item);
2007
2008 for (i = *level; i > 0; i--) {
2009 eb = path->nodes[i];
2010 nritems = btrfs_header_nritems(eb);
2011 while (path->slots[i] < nritems) {
2012 ptr_gen = btrfs_node_ptr_generation(eb, path->slots[i]);
2013 if (ptr_gen > last_snapshot)
2014 break;
2015 path->slots[i]++;
2016 }
2017 if (path->slots[i] >= nritems) {
2018 if (i == *level)
2019 break;
2020 *level = i + 1;
2021 return 0;
2022 }
2023 if (i == 1) {
2024 *level = i;
2025 return 0;
2026 }
2027
2028 bytenr = btrfs_node_blockptr(eb, path->slots[i]);
2029 eb = read_tree_block(root, bytenr, ptr_gen);
2030 if (IS_ERR(eb)) {
2031 return PTR_ERR(eb);
2032 } else if (!extent_buffer_uptodate(eb)) {
2033 free_extent_buffer(eb);
2034 return -EIO;
2035 }
2036 BUG_ON(btrfs_header_level(eb) != i - 1);
2037 path->nodes[i - 1] = eb;
2038 path->slots[i - 1] = 0;
2039 }
2040 return 1;
2041 }
2042
2043 /*
2044 * invalidate extent cache for file extents whose key in range of
2045 * [min_key, max_key)
2046 */
2047 static int invalidate_extent_cache(struct btrfs_root *root,
2048 struct btrfs_key *min_key,
2049 struct btrfs_key *max_key)
2050 {
2051 struct inode *inode = NULL;
2052 u64 objectid;
2053 u64 start, end;
2054 u64 ino;
2055
2056 objectid = min_key->objectid;
2057 while (1) {
2058 cond_resched();
2059 iput(inode);
2060
2061 if (objectid > max_key->objectid)
2062 break;
2063
2064 inode = find_next_inode(root, objectid);
2065 if (!inode)
2066 break;
2067 ino = btrfs_ino(inode);
2068
2069 if (ino > max_key->objectid) {
2070 iput(inode);
2071 break;
2072 }
2073
2074 objectid = ino + 1;
2075 if (!S_ISREG(inode->i_mode))
2076 continue;
2077
2078 if (unlikely(min_key->objectid == ino)) {
2079 if (min_key->type > BTRFS_EXTENT_DATA_KEY)
2080 continue;
2081 if (min_key->type < BTRFS_EXTENT_DATA_KEY)
2082 start = 0;
2083 else {
2084 start = min_key->offset;
2085 WARN_ON(!IS_ALIGNED(start, root->sectorsize));
2086 }
2087 } else {
2088 start = 0;
2089 }
2090
2091 if (unlikely(max_key->objectid == ino)) {
2092 if (max_key->type < BTRFS_EXTENT_DATA_KEY)
2093 continue;
2094 if (max_key->type > BTRFS_EXTENT_DATA_KEY) {
2095 end = (u64)-1;
2096 } else {
2097 if (max_key->offset == 0)
2098 continue;
2099 end = max_key->offset;
2100 WARN_ON(!IS_ALIGNED(end, root->sectorsize));
2101 end--;
2102 }
2103 } else {
2104 end = (u64)-1;
2105 }
2106
2107 /* the lock_extent waits for readpage to complete */
2108 lock_extent(&BTRFS_I(inode)->io_tree, start, end);
2109 btrfs_drop_extent_cache(inode, start, end, 1);
2110 unlock_extent(&BTRFS_I(inode)->io_tree, start, end);
2111 }
2112 return 0;
2113 }
2114
2115 static int find_next_key(struct btrfs_path *path, int level,
2116 struct btrfs_key *key)
2117
2118 {
2119 while (level < BTRFS_MAX_LEVEL) {
2120 if (!path->nodes[level])
2121 break;
2122 if (path->slots[level] + 1 <
2123 btrfs_header_nritems(path->nodes[level])) {
2124 btrfs_node_key_to_cpu(path->nodes[level], key,
2125 path->slots[level] + 1);
2126 return 0;
2127 }
2128 level++;
2129 }
2130 return 1;
2131 }
2132
2133 /*
2134 * merge the relocated tree blocks in reloc tree with corresponding
2135 * fs tree.
2136 */
2137 static noinline_for_stack int merge_reloc_root(struct reloc_control *rc,
2138 struct btrfs_root *root)
2139 {
2140 LIST_HEAD(inode_list);
2141 struct btrfs_key key;
2142 struct btrfs_key next_key;
2143 struct btrfs_trans_handle *trans = NULL;
2144 struct btrfs_root *reloc_root;
2145 struct btrfs_root_item *root_item;
2146 struct btrfs_path *path;
2147 struct extent_buffer *leaf;
2148 int level;
2149 int max_level;
2150 int replaced = 0;
2151 int ret;
2152 int err = 0;
2153 u32 min_reserved;
2154
2155 path = btrfs_alloc_path();
2156 if (!path)
2157 return -ENOMEM;
2158 path->reada = READA_FORWARD;
2159
2160 reloc_root = root->reloc_root;
2161 root_item = &reloc_root->root_item;
2162
2163 if (btrfs_disk_key_objectid(&root_item->drop_progress) == 0) {
2164 level = btrfs_root_level(root_item);
2165 extent_buffer_get(reloc_root->node);
2166 path->nodes[level] = reloc_root->node;
2167 path->slots[level] = 0;
2168 } else {
2169 btrfs_disk_key_to_cpu(&key, &root_item->drop_progress);
2170
2171 level = root_item->drop_level;
2172 BUG_ON(level == 0);
2173 path->lowest_level = level;
2174 ret = btrfs_search_slot(NULL, reloc_root, &key, path, 0, 0);
2175 path->lowest_level = 0;
2176 if (ret < 0) {
2177 btrfs_free_path(path);
2178 return ret;
2179 }
2180
2181 btrfs_node_key_to_cpu(path->nodes[level], &next_key,
2182 path->slots[level]);
2183 WARN_ON(memcmp(&key, &next_key, sizeof(key)));
2184
2185 btrfs_unlock_up_safe(path, 0);
2186 }
2187
2188 min_reserved = root->nodesize * (BTRFS_MAX_LEVEL - 1) * 2;
2189 memset(&next_key, 0, sizeof(next_key));
2190
2191 while (1) {
2192 ret = btrfs_block_rsv_refill(root, rc->block_rsv, min_reserved,
2193 BTRFS_RESERVE_FLUSH_ALL);
2194 if (ret) {
2195 err = ret;
2196 goto out;
2197 }
2198 trans = btrfs_start_transaction(root, 0);
2199 if (IS_ERR(trans)) {
2200 err = PTR_ERR(trans);
2201 trans = NULL;
2202 goto out;
2203 }
2204 trans->block_rsv = rc->block_rsv;
2205
2206 replaced = 0;
2207 max_level = level;
2208
2209 ret = walk_down_reloc_tree(reloc_root, path, &level);
2210 if (ret < 0) {
2211 err = ret;
2212 goto out;
2213 }
2214 if (ret > 0)
2215 break;
2216
2217 if (!find_next_key(path, level, &key) &&
2218 btrfs_comp_cpu_keys(&next_key, &key) >= 0) {
2219 ret = 0;
2220 } else {
2221 ret = replace_path(trans, root, reloc_root, path,
2222 &next_key, level, max_level);
2223 }
2224 if (ret < 0) {
2225 err = ret;
2226 goto out;
2227 }
2228
2229 if (ret > 0) {
2230 level = ret;
2231 btrfs_node_key_to_cpu(path->nodes[level], &key,
2232 path->slots[level]);
2233 replaced = 1;
2234 }
2235
2236 ret = walk_up_reloc_tree(reloc_root, path, &level);
2237 if (ret > 0)
2238 break;
2239
2240 BUG_ON(level == 0);
2241 /*
2242 * save the merging progress in the drop_progress.
2243 * this is OK since root refs == 1 in this case.
2244 */
2245 btrfs_node_key(path->nodes[level], &root_item->drop_progress,
2246 path->slots[level]);
2247 root_item->drop_level = level;
2248
2249 btrfs_end_transaction_throttle(trans, root);
2250 trans = NULL;
2251
2252 btrfs_btree_balance_dirty(root);
2253
2254 if (replaced && rc->stage == UPDATE_DATA_PTRS)
2255 invalidate_extent_cache(root, &key, &next_key);
2256 }
2257
2258 /*
2259 * handle the case only one block in the fs tree need to be
2260 * relocated and the block is tree root.
2261 */
2262 leaf = btrfs_lock_root_node(root);
2263 ret = btrfs_cow_block(trans, root, leaf, NULL, 0, &leaf);
2264 btrfs_tree_unlock(leaf);
2265 free_extent_buffer(leaf);
2266 if (ret < 0)
2267 err = ret;
2268 out:
2269 btrfs_free_path(path);
2270
2271 if (err == 0) {
2272 memset(&root_item->drop_progress, 0,
2273 sizeof(root_item->drop_progress));
2274 root_item->drop_level = 0;
2275 btrfs_set_root_refs(root_item, 0);
2276 btrfs_update_reloc_root(trans, root);
2277 }
2278
2279 if (trans)
2280 btrfs_end_transaction_throttle(trans, root);
2281
2282 btrfs_btree_balance_dirty(root);
2283
2284 if (replaced && rc->stage == UPDATE_DATA_PTRS)
2285 invalidate_extent_cache(root, &key, &next_key);
2286
2287 return err;
2288 }
2289
2290 static noinline_for_stack
2291 int prepare_to_merge(struct reloc_control *rc, int err)
2292 {
2293 struct btrfs_root *root = rc->extent_root;
2294 struct btrfs_root *reloc_root;
2295 struct btrfs_trans_handle *trans;
2296 LIST_HEAD(reloc_roots);
2297 u64 num_bytes = 0;
2298 int ret;
2299
2300 mutex_lock(&root->fs_info->reloc_mutex);
2301 rc->merging_rsv_size += root->nodesize * (BTRFS_MAX_LEVEL - 1) * 2;
2302 rc->merging_rsv_size += rc->nodes_relocated * 2;
2303 mutex_unlock(&root->fs_info->reloc_mutex);
2304
2305 again:
2306 if (!err) {
2307 num_bytes = rc->merging_rsv_size;
2308 ret = btrfs_block_rsv_add(root, rc->block_rsv, num_bytes,
2309 BTRFS_RESERVE_FLUSH_ALL);
2310 if (ret)
2311 err = ret;
2312 }
2313
2314 trans = btrfs_join_transaction(rc->extent_root);
2315 if (IS_ERR(trans)) {
2316 if (!err)
2317 btrfs_block_rsv_release(rc->extent_root,
2318 rc->block_rsv, num_bytes);
2319 return PTR_ERR(trans);
2320 }
2321
2322 if (!err) {
2323 if (num_bytes != rc->merging_rsv_size) {
2324 btrfs_end_transaction(trans, rc->extent_root);
2325 btrfs_block_rsv_release(rc->extent_root,
2326 rc->block_rsv, num_bytes);
2327 goto again;
2328 }
2329 }
2330
2331 rc->merge_reloc_tree = 1;
2332
2333 while (!list_empty(&rc->reloc_roots)) {
2334 reloc_root = list_entry(rc->reloc_roots.next,
2335 struct btrfs_root, root_list);
2336 list_del_init(&reloc_root->root_list);
2337
2338 root = read_fs_root(reloc_root->fs_info,
2339 reloc_root->root_key.offset);
2340 BUG_ON(IS_ERR(root));
2341 BUG_ON(root->reloc_root != reloc_root);
2342
2343 /*
2344 * set reference count to 1, so btrfs_recover_relocation
2345 * knows it should resumes merging
2346 */
2347 if (!err)
2348 btrfs_set_root_refs(&reloc_root->root_item, 1);
2349 btrfs_update_reloc_root(trans, root);
2350
2351 list_add(&reloc_root->root_list, &reloc_roots);
2352 }
2353
2354 list_splice(&reloc_roots, &rc->reloc_roots);
2355
2356 if (!err)
2357 btrfs_commit_transaction(trans, rc->extent_root);
2358 else
2359 btrfs_end_transaction(trans, rc->extent_root);
2360 return err;
2361 }
2362
2363 static noinline_for_stack
2364 void free_reloc_roots(struct list_head *list)
2365 {
2366 struct btrfs_root *reloc_root;
2367
2368 while (!list_empty(list)) {
2369 reloc_root = list_entry(list->next, struct btrfs_root,
2370 root_list);
2371 __del_reloc_root(reloc_root);
2372 free_extent_buffer(reloc_root->node);
2373 free_extent_buffer(reloc_root->commit_root);
2374 reloc_root->node = NULL;
2375 reloc_root->commit_root = NULL;
2376 }
2377 }
2378
2379 static noinline_for_stack
2380 void merge_reloc_roots(struct reloc_control *rc)
2381 {
2382 struct btrfs_root *root;
2383 struct btrfs_root *reloc_root;
2384 u64 last_snap;
2385 u64 otransid;
2386 u64 objectid;
2387 LIST_HEAD(reloc_roots);
2388 int found = 0;
2389 int ret = 0;
2390 again:
2391 root = rc->extent_root;
2392
2393 /*
2394 * this serializes us with btrfs_record_root_in_transaction,
2395 * we have to make sure nobody is in the middle of
2396 * adding their roots to the list while we are
2397 * doing this splice
2398 */
2399 mutex_lock(&root->fs_info->reloc_mutex);
2400 list_splice_init(&rc->reloc_roots, &reloc_roots);
2401 mutex_unlock(&root->fs_info->reloc_mutex);
2402
2403 while (!list_empty(&reloc_roots)) {
2404 found = 1;
2405 reloc_root = list_entry(reloc_roots.next,
2406 struct btrfs_root, root_list);
2407
2408 if (btrfs_root_refs(&reloc_root->root_item) > 0) {
2409 root = read_fs_root(reloc_root->fs_info,
2410 reloc_root->root_key.offset);
2411 BUG_ON(IS_ERR(root));
2412 BUG_ON(root->reloc_root != reloc_root);
2413
2414 ret = merge_reloc_root(rc, root);
2415 if (ret) {
2416 if (list_empty(&reloc_root->root_list))
2417 list_add_tail(&reloc_root->root_list,
2418 &reloc_roots);
2419 goto out;
2420 }
2421 } else {
2422 list_del_init(&reloc_root->root_list);
2423 }
2424
2425 /*
2426 * we keep the old last snapshot transid in rtranid when we
2427 * created the relocation tree.
2428 */
2429 last_snap = btrfs_root_rtransid(&reloc_root->root_item);
2430 otransid = btrfs_root_otransid(&reloc_root->root_item);
2431 objectid = reloc_root->root_key.offset;
2432
2433 ret = btrfs_drop_snapshot(reloc_root, rc->block_rsv, 0, 1);
2434 if (ret < 0) {
2435 if (list_empty(&reloc_root->root_list))
2436 list_add_tail(&reloc_root->root_list,
2437 &reloc_roots);
2438 goto out;
2439 }
2440 }
2441
2442 if (found) {
2443 found = 0;
2444 goto again;
2445 }
2446 out:
2447 if (ret) {
2448 btrfs_handle_fs_error(root->fs_info, ret, NULL);
2449 if (!list_empty(&reloc_roots))
2450 free_reloc_roots(&reloc_roots);
2451
2452 /* new reloc root may be added */
2453 mutex_lock(&root->fs_info->reloc_mutex);
2454 list_splice_init(&rc->reloc_roots, &reloc_roots);
2455 mutex_unlock(&root->fs_info->reloc_mutex);
2456 if (!list_empty(&reloc_roots))
2457 free_reloc_roots(&reloc_roots);
2458 }
2459
2460 BUG_ON(!RB_EMPTY_ROOT(&rc->reloc_root_tree.rb_root));
2461 }
2462
2463 static void free_block_list(struct rb_root *blocks)
2464 {
2465 struct tree_block *block;
2466 struct rb_node *rb_node;
2467 while ((rb_node = rb_first(blocks))) {
2468 block = rb_entry(rb_node, struct tree_block, rb_node);
2469 rb_erase(rb_node, blocks);
2470 kfree(block);
2471 }
2472 }
2473
2474 static int record_reloc_root_in_trans(struct btrfs_trans_handle *trans,
2475 struct btrfs_root *reloc_root)
2476 {
2477 struct btrfs_root *root;
2478
2479 if (reloc_root->last_trans == trans->transid)
2480 return 0;
2481
2482 root = read_fs_root(reloc_root->fs_info, reloc_root->root_key.offset);
2483 BUG_ON(IS_ERR(root));
2484 BUG_ON(root->reloc_root != reloc_root);
2485
2486 return btrfs_record_root_in_trans(trans, root);
2487 }
2488
2489 static noinline_for_stack
2490 struct btrfs_root *select_reloc_root(struct btrfs_trans_handle *trans,
2491 struct reloc_control *rc,
2492 struct backref_node *node,
2493 struct backref_edge *edges[])
2494 {
2495 struct backref_node *next;
2496 struct btrfs_root *root;
2497 int index = 0;
2498
2499 next = node;
2500 while (1) {
2501 cond_resched();
2502 next = walk_up_backref(next, edges, &index);
2503 root = next->root;
2504 BUG_ON(!root);
2505 BUG_ON(!test_bit(BTRFS_ROOT_REF_COWS, &root->state));
2506
2507 if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID) {
2508 record_reloc_root_in_trans(trans, root);
2509 break;
2510 }
2511
2512 btrfs_record_root_in_trans(trans, root);
2513 root = root->reloc_root;
2514
2515 if (next->new_bytenr != root->node->start) {
2516 BUG_ON(next->new_bytenr);
2517 BUG_ON(!list_empty(&next->list));
2518 next->new_bytenr = root->node->start;
2519 next->root = root;
2520 list_add_tail(&next->list,
2521 &rc->backref_cache.changed);
2522 __mark_block_processed(rc, next);
2523 break;
2524 }
2525
2526 WARN_ON(1);
2527 root = NULL;
2528 next = walk_down_backref(edges, &index);
2529 if (!next || next->level <= node->level)
2530 break;
2531 }
2532 if (!root)
2533 return NULL;
2534
2535 next = node;
2536 /* setup backref node path for btrfs_reloc_cow_block */
2537 while (1) {
2538 rc->backref_cache.path[next->level] = next;
2539 if (--index < 0)
2540 break;
2541 next = edges[index]->node[UPPER];
2542 }
2543 return root;
2544 }
2545
2546 /*
2547 * select a tree root for relocation. return NULL if the block
2548 * is reference counted. we should use do_relocation() in this
2549 * case. return a tree root pointer if the block isn't reference
2550 * counted. return -ENOENT if the block is root of reloc tree.
2551 */
2552 static noinline_for_stack
2553 struct btrfs_root *select_one_root(struct backref_node *node)
2554 {
2555 struct backref_node *next;
2556 struct btrfs_root *root;
2557 struct btrfs_root *fs_root = NULL;
2558 struct backref_edge *edges[BTRFS_MAX_LEVEL - 1];
2559 int index = 0;
2560
2561 next = node;
2562 while (1) {
2563 cond_resched();
2564 next = walk_up_backref(next, edges, &index);
2565 root = next->root;
2566 BUG_ON(!root);
2567
2568 /* no other choice for non-references counted tree */
2569 if (!test_bit(BTRFS_ROOT_REF_COWS, &root->state))
2570 return root;
2571
2572 if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID)
2573 fs_root = root;
2574
2575 if (next != node)
2576 return NULL;
2577
2578 next = walk_down_backref(edges, &index);
2579 if (!next || next->level <= node->level)
2580 break;
2581 }
2582
2583 if (!fs_root)
2584 return ERR_PTR(-ENOENT);
2585 return fs_root;
2586 }
2587
2588 static noinline_for_stack
2589 u64 calcu_metadata_size(struct reloc_control *rc,
2590 struct backref_node *node, int reserve)
2591 {
2592 struct backref_node *next = node;
2593 struct backref_edge *edge;
2594 struct backref_edge *edges[BTRFS_MAX_LEVEL - 1];
2595 u64 num_bytes = 0;
2596 int index = 0;
2597
2598 BUG_ON(reserve && node->processed);
2599
2600 while (next) {
2601 cond_resched();
2602 while (1) {
2603 if (next->processed && (reserve || next != node))
2604 break;
2605
2606 num_bytes += rc->extent_root->nodesize;
2607
2608 if (list_empty(&next->upper))
2609 break;
2610
2611 edge = list_entry(next->upper.next,
2612 struct backref_edge, list[LOWER]);
2613 edges[index++] = edge;
2614 next = edge->node[UPPER];
2615 }
2616 next = walk_down_backref(edges, &index);
2617 }
2618 return num_bytes;
2619 }
2620
2621 static int reserve_metadata_space(struct btrfs_trans_handle *trans,
2622 struct reloc_control *rc,
2623 struct backref_node *node)
2624 {
2625 struct btrfs_root *root = rc->extent_root;
2626 u64 num_bytes;
2627 int ret;
2628 u64 tmp;
2629
2630 num_bytes = calcu_metadata_size(rc, node, 1) * 2;
2631
2632 trans->block_rsv = rc->block_rsv;
2633 rc->reserved_bytes += num_bytes;
2634
2635 /*
2636 * We are under a transaction here so we can only do limited flushing.
2637 * If we get an enospc just kick back -EAGAIN so we know to drop the
2638 * transaction and try to refill when we can flush all the things.
2639 */
2640 ret = btrfs_block_rsv_refill(root, rc->block_rsv, num_bytes,
2641 BTRFS_RESERVE_FLUSH_LIMIT);
2642 if (ret) {
2643 tmp = rc->extent_root->nodesize * RELOCATION_RESERVED_NODES;
2644 while (tmp <= rc->reserved_bytes)
2645 tmp <<= 1;
2646 /*
2647 * only one thread can access block_rsv at this point,
2648 * so we don't need hold lock to protect block_rsv.
2649 * we expand more reservation size here to allow enough
2650 * space for relocation and we will return eailer in
2651 * enospc case.
2652 */
2653 rc->block_rsv->size = tmp + rc->extent_root->nodesize *
2654 RELOCATION_RESERVED_NODES;
2655 return -EAGAIN;
2656 }
2657
2658 return 0;
2659 }
2660
2661 /*
2662 * relocate a block tree, and then update pointers in upper level
2663 * blocks that reference the block to point to the new location.
2664 *
2665 * if called by link_to_upper, the block has already been relocated.
2666 * in that case this function just updates pointers.
2667 */
2668 static int do_relocation(struct btrfs_trans_handle *trans,
2669 struct reloc_control *rc,
2670 struct backref_node *node,
2671 struct btrfs_key *key,
2672 struct btrfs_path *path, int lowest)
2673 {
2674 struct backref_node *upper;
2675 struct backref_edge *edge;
2676 struct backref_edge *edges[BTRFS_MAX_LEVEL - 1];
2677 struct btrfs_root *root;
2678 struct extent_buffer *eb;
2679 u32 blocksize;
2680 u64 bytenr;
2681 u64 generation;
2682 int slot;
2683 int ret;
2684 int err = 0;
2685
2686 BUG_ON(lowest && node->eb);
2687
2688 path->lowest_level = node->level + 1;
2689 rc->backref_cache.path[node->level] = node;
2690 list_for_each_entry(edge, &node->upper, list[LOWER]) {
2691 cond_resched();
2692
2693 upper = edge->node[UPPER];
2694 root = select_reloc_root(trans, rc, upper, edges);
2695 BUG_ON(!root);
2696
2697 if (upper->eb && !upper->locked) {
2698 if (!lowest) {
2699 ret = btrfs_bin_search(upper->eb, key,
2700 upper->level, &slot);
2701 BUG_ON(ret);
2702 bytenr = btrfs_node_blockptr(upper->eb, slot);
2703 if (node->eb->start == bytenr)
2704 goto next;
2705 }
2706 drop_node_buffer(upper);
2707 }
2708
2709 if (!upper->eb) {
2710 ret = btrfs_search_slot(trans, root, key, path, 0, 1);
2711 if (ret) {
2712 if (ret < 0)
2713 err = ret;
2714 else
2715 err = -ENOENT;
2716
2717 btrfs_release_path(path);
2718 break;
2719 }
2720
2721 if (!upper->eb) {
2722 upper->eb = path->nodes[upper->level];
2723 path->nodes[upper->level] = NULL;
2724 } else {
2725 BUG_ON(upper->eb != path->nodes[upper->level]);
2726 }
2727
2728 upper->locked = 1;
2729 path->locks[upper->level] = 0;
2730
2731 slot = path->slots[upper->level];
2732 btrfs_release_path(path);
2733 } else {
2734 ret = btrfs_bin_search(upper->eb, key, upper->level,
2735 &slot);
2736 BUG_ON(ret);
2737 }
2738
2739 bytenr = btrfs_node_blockptr(upper->eb, slot);
2740 if (lowest) {
2741 if (bytenr != node->bytenr) {
2742 btrfs_err(root->fs_info,
2743 "lowest leaf/node mismatch: bytenr %llu node->bytenr %llu slot %d upper %llu",
2744 bytenr, node->bytenr, slot,
2745 upper->eb->start);
2746 err = -EIO;
2747 goto next;
2748 }
2749 } else {
2750 if (node->eb->start == bytenr)
2751 goto next;
2752 }
2753
2754 blocksize = root->nodesize;
2755 generation = btrfs_node_ptr_generation(upper->eb, slot);
2756 eb = read_tree_block(root, bytenr, generation);
2757 if (IS_ERR(eb)) {
2758 err = PTR_ERR(eb);
2759 goto next;
2760 } else if (!extent_buffer_uptodate(eb)) {
2761 free_extent_buffer(eb);
2762 err = -EIO;
2763 goto next;
2764 }
2765 btrfs_tree_lock(eb);
2766 btrfs_set_lock_blocking(eb);
2767
2768 if (!node->eb) {
2769 ret = btrfs_cow_block(trans, root, eb, upper->eb,
2770 slot, &eb);
2771 btrfs_tree_unlock(eb);
2772 free_extent_buffer(eb);
2773 if (ret < 0) {
2774 err = ret;
2775 goto next;
2776 }
2777 BUG_ON(node->eb != eb);
2778 } else {
2779 btrfs_set_node_blockptr(upper->eb, slot,
2780 node->eb->start);
2781 btrfs_set_node_ptr_generation(upper->eb, slot,
2782 trans->transid);
2783 btrfs_mark_buffer_dirty(upper->eb);
2784
2785 ret = btrfs_inc_extent_ref(trans, root,
2786 node->eb->start, blocksize,
2787 upper->eb->start,
2788 btrfs_header_owner(upper->eb),
2789 node->level, 0);
2790 BUG_ON(ret);
2791
2792 ret = btrfs_drop_subtree(trans, root, eb, upper->eb);
2793 BUG_ON(ret);
2794 }
2795 next:
2796 if (!upper->pending)
2797 drop_node_buffer(upper);
2798 else
2799 unlock_node_buffer(upper);
2800 if (err)
2801 break;
2802 }
2803
2804 if (!err && node->pending) {
2805 drop_node_buffer(node);
2806 list_move_tail(&node->list, &rc->backref_cache.changed);
2807 node->pending = 0;
2808 }
2809
2810 path->lowest_level = 0;
2811 BUG_ON(err == -ENOSPC);
2812 return err;
2813 }
2814
2815 static int link_to_upper(struct btrfs_trans_handle *trans,
2816 struct reloc_control *rc,
2817 struct backref_node *node,
2818 struct btrfs_path *path)
2819 {
2820 struct btrfs_key key;
2821
2822 btrfs_node_key_to_cpu(node->eb, &key, 0);
2823 return do_relocation(trans, rc, node, &key, path, 0);
2824 }
2825
2826 static int finish_pending_nodes(struct btrfs_trans_handle *trans,
2827 struct reloc_control *rc,
2828 struct btrfs_path *path, int err)
2829 {
2830 LIST_HEAD(list);
2831 struct backref_cache *cache = &rc->backref_cache;
2832 struct backref_node *node;
2833 int level;
2834 int ret;
2835
2836 for (level = 0; level < BTRFS_MAX_LEVEL; level++) {
2837 while (!list_empty(&cache->pending[level])) {
2838 node = list_entry(cache->pending[level].next,
2839 struct backref_node, list);
2840 list_move_tail(&node->list, &list);
2841 BUG_ON(!node->pending);
2842
2843 if (!err) {
2844 ret = link_to_upper(trans, rc, node, path);
2845 if (ret < 0)
2846 err = ret;
2847 }
2848 }
2849 list_splice_init(&list, &cache->pending[level]);
2850 }
2851 return err;
2852 }
2853
2854 static void mark_block_processed(struct reloc_control *rc,
2855 u64 bytenr, u32 blocksize)
2856 {
2857 set_extent_bits(&rc->processed_blocks, bytenr, bytenr + blocksize - 1,
2858 EXTENT_DIRTY);
2859 }
2860
2861 static void __mark_block_processed(struct reloc_control *rc,
2862 struct backref_node *node)
2863 {
2864 u32 blocksize;
2865 if (node->level == 0 ||
2866 in_block_group(node->bytenr, rc->block_group)) {
2867 blocksize = rc->extent_root->nodesize;
2868 mark_block_processed(rc, node->bytenr, blocksize);
2869 }
2870 node->processed = 1;
2871 }
2872
2873 /*
2874 * mark a block and all blocks directly/indirectly reference the block
2875 * as processed.
2876 */
2877 static void update_processed_blocks(struct reloc_control *rc,
2878 struct backref_node *node)
2879 {
2880 struct backref_node *next = node;
2881 struct backref_edge *edge;
2882 struct backref_edge *edges[BTRFS_MAX_LEVEL - 1];
2883 int index = 0;
2884
2885 while (next) {
2886 cond_resched();
2887 while (1) {
2888 if (next->processed)
2889 break;
2890
2891 __mark_block_processed(rc, next);
2892
2893 if (list_empty(&next->upper))
2894 break;
2895
2896 edge = list_entry(next->upper.next,
2897 struct backref_edge, list[LOWER]);
2898 edges[index++] = edge;
2899 next = edge->node[UPPER];
2900 }
2901 next = walk_down_backref(edges, &index);
2902 }
2903 }
2904
2905 static int tree_block_processed(u64 bytenr, struct reloc_control *rc)
2906 {
2907 u32 blocksize = rc->extent_root->nodesize;
2908
2909 if (test_range_bit(&rc->processed_blocks, bytenr,
2910 bytenr + blocksize - 1, EXTENT_DIRTY, 1, NULL))
2911 return 1;
2912 return 0;
2913 }
2914
2915 static int get_tree_block_key(struct reloc_control *rc,
2916 struct tree_block *block)
2917 {
2918 struct extent_buffer *eb;
2919
2920 BUG_ON(block->key_ready);
2921 eb = read_tree_block(rc->extent_root, block->bytenr,
2922 block->key.offset);
2923 if (IS_ERR(eb)) {
2924 return PTR_ERR(eb);
2925 } else if (!extent_buffer_uptodate(eb)) {
2926 free_extent_buffer(eb);
2927 return -EIO;
2928 }
2929 WARN_ON(btrfs_header_level(eb) != block->level);
2930 if (block->level == 0)
2931 btrfs_item_key_to_cpu(eb, &block->key, 0);
2932 else
2933 btrfs_node_key_to_cpu(eb, &block->key, 0);
2934 free_extent_buffer(eb);
2935 block->key_ready = 1;
2936 return 0;
2937 }
2938
2939 /*
2940 * helper function to relocate a tree block
2941 */
2942 static int relocate_tree_block(struct btrfs_trans_handle *trans,
2943 struct reloc_control *rc,
2944 struct backref_node *node,
2945 struct btrfs_key *key,
2946 struct btrfs_path *path)
2947 {
2948 struct btrfs_root *root;
2949 int ret = 0;
2950
2951 if (!node)
2952 return 0;
2953
2954 BUG_ON(node->processed);
2955 root = select_one_root(node);
2956 if (root == ERR_PTR(-ENOENT)) {
2957 update_processed_blocks(rc, node);
2958 goto out;
2959 }
2960
2961 if (!root || test_bit(BTRFS_ROOT_REF_COWS, &root->state)) {
2962 ret = reserve_metadata_space(trans, rc, node);
2963 if (ret)
2964 goto out;
2965 }
2966
2967 if (root) {
2968 if (test_bit(BTRFS_ROOT_REF_COWS, &root->state)) {
2969 BUG_ON(node->new_bytenr);
2970 BUG_ON(!list_empty(&node->list));
2971 btrfs_record_root_in_trans(trans, root);
2972 root = root->reloc_root;
2973 node->new_bytenr = root->node->start;
2974 node->root = root;
2975 list_add_tail(&node->list, &rc->backref_cache.changed);
2976 } else {
2977 path->lowest_level = node->level;
2978 ret = btrfs_search_slot(trans, root, key, path, 0, 1);
2979 btrfs_release_path(path);
2980 if (ret > 0)
2981 ret = 0;
2982 }
2983 if (!ret)
2984 update_processed_blocks(rc, node);
2985 } else {
2986 ret = do_relocation(trans, rc, node, key, path, 1);
2987 }
2988 out:
2989 if (ret || node->level == 0 || node->cowonly)
2990 remove_backref_node(&rc->backref_cache, node);
2991 return ret;
2992 }
2993
2994 /*
2995 * relocate a list of blocks
2996 */
2997 static noinline_for_stack
2998 int relocate_tree_blocks(struct btrfs_trans_handle *trans,
2999 struct reloc_control *rc, struct rb_root *blocks)
3000 {
3001 struct backref_node *node;
3002 struct btrfs_path *path;
3003 struct tree_block *block;
3004 struct rb_node *rb_node;
3005 int ret;
3006 int err = 0;
3007
3008 path = btrfs_alloc_path();
3009 if (!path) {
3010 err = -ENOMEM;
3011 goto out_free_blocks;
3012 }
3013
3014 rb_node = rb_first(blocks);
3015 while (rb_node) {
3016 block = rb_entry(rb_node, struct tree_block, rb_node);
3017 if (!block->key_ready)
3018 readahead_tree_block(rc->extent_root, block->bytenr);
3019 rb_node = rb_next(rb_node);
3020 }
3021
3022 rb_node = rb_first(blocks);
3023 while (rb_node) {
3024 block = rb_entry(rb_node, struct tree_block, rb_node);
3025 if (!block->key_ready) {
3026 err = get_tree_block_key(rc, block);
3027 if (err)
3028 goto out_free_path;
3029 }
3030 rb_node = rb_next(rb_node);
3031 }
3032
3033 rb_node = rb_first(blocks);
3034 while (rb_node) {
3035 block = rb_entry(rb_node, struct tree_block, rb_node);
3036
3037 node = build_backref_tree(rc, &block->key,
3038 block->level, block->bytenr);
3039 if (IS_ERR(node)) {
3040 err = PTR_ERR(node);
3041 goto out;
3042 }
3043
3044 ret = relocate_tree_block(trans, rc, node, &block->key,
3045 path);
3046 if (ret < 0) {
3047 if (ret != -EAGAIN || rb_node == rb_first(blocks))
3048 err = ret;
3049 goto out;
3050 }
3051 rb_node = rb_next(rb_node);
3052 }
3053 out:
3054 err = finish_pending_nodes(trans, rc, path, err);
3055
3056 out_free_path:
3057 btrfs_free_path(path);
3058 out_free_blocks:
3059 free_block_list(blocks);
3060 return err;
3061 }
3062
3063 static noinline_for_stack
3064 int prealloc_file_extent_cluster(struct inode *inode,
3065 struct file_extent_cluster *cluster)
3066 {
3067 u64 alloc_hint = 0;
3068 u64 start;
3069 u64 end;
3070 u64 offset = BTRFS_I(inode)->index_cnt;
3071 u64 num_bytes;
3072 int nr = 0;
3073 int ret = 0;
3074 u64 prealloc_start = cluster->start - offset;
3075 u64 prealloc_end = cluster->end - offset;
3076 u64 cur_offset;
3077
3078 BUG_ON(cluster->start != cluster->boundary[0]);
3079 inode_lock(inode);
3080
3081 ret = btrfs_check_data_free_space(inode, prealloc_start,
3082 prealloc_end + 1 - prealloc_start);
3083 if (ret)
3084 goto out;
3085
3086 cur_offset = prealloc_start;
3087 while (nr < cluster->nr) {
3088 start = cluster->boundary[nr] - offset;
3089 if (nr + 1 < cluster->nr)
3090 end = cluster->boundary[nr + 1] - 1 - offset;
3091 else
3092 end = cluster->end - offset;
3093
3094 lock_extent(&BTRFS_I(inode)->io_tree, start, end);
3095 num_bytes = end + 1 - start;
3096 if (cur_offset < start)
3097 btrfs_free_reserved_data_space(inode, cur_offset,
3098 start - cur_offset);
3099 ret = btrfs_prealloc_file_range(inode, 0, start,
3100 num_bytes, num_bytes,
3101 end + 1, &alloc_hint);
3102 cur_offset = end + 1;
3103 unlock_extent(&BTRFS_I(inode)->io_tree, start, end);
3104 if (ret)
3105 break;
3106 nr++;
3107 }
3108 if (cur_offset < prealloc_end)
3109 btrfs_free_reserved_data_space(inode, cur_offset,
3110 prealloc_end + 1 - cur_offset);
3111 out:
3112 inode_unlock(inode);
3113 return ret;
3114 }
3115
3116 static noinline_for_stack
3117 int setup_extent_mapping(struct inode *inode, u64 start, u64 end,
3118 u64 block_start)
3119 {
3120 struct btrfs_root *root = BTRFS_I(inode)->root;
3121 struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
3122 struct extent_map *em;
3123 int ret = 0;
3124
3125 em = alloc_extent_map();
3126 if (!em)
3127 return -ENOMEM;
3128
3129 em->start = start;
3130 em->len = end + 1 - start;
3131 em->block_len = em->len;
3132 em->block_start = block_start;
3133 em->bdev = root->fs_info->fs_devices->latest_bdev;
3134 set_bit(EXTENT_FLAG_PINNED, &em->flags);
3135
3136 lock_extent(&BTRFS_I(inode)->io_tree, start, end);
3137 while (1) {
3138 write_lock(&em_tree->lock);
3139 ret = add_extent_mapping(em_tree, em, 0);
3140 write_unlock(&em_tree->lock);
3141 if (ret != -EEXIST) {
3142 free_extent_map(em);
3143 break;
3144 }
3145 btrfs_drop_extent_cache(inode, start, end, 0);
3146 }
3147 unlock_extent(&BTRFS_I(inode)->io_tree, start, end);
3148 return ret;
3149 }
3150
3151 static int relocate_file_extent_cluster(struct inode *inode,
3152 struct file_extent_cluster *cluster)
3153 {
3154 u64 page_start;
3155 u64 page_end;
3156 u64 offset = BTRFS_I(inode)->index_cnt;
3157 unsigned long index;
3158 unsigned long last_index;
3159 struct page *page;
3160 struct file_ra_state *ra;
3161 gfp_t mask = btrfs_alloc_write_mask(inode->i_mapping);
3162 int nr = 0;
3163 int ret = 0;
3164
3165 if (!cluster->nr)
3166 return 0;
3167
3168 ra = kzalloc(sizeof(*ra), GFP_NOFS);
3169 if (!ra)
3170 return -ENOMEM;
3171
3172 ret = prealloc_file_extent_cluster(inode, cluster);
3173 if (ret)
3174 goto out;
3175
3176 file_ra_state_init(ra, inode->i_mapping);
3177
3178 ret = setup_extent_mapping(inode, cluster->start - offset,
3179 cluster->end - offset, cluster->start);
3180 if (ret)
3181 goto out;
3182
3183 index = (cluster->start - offset) >> PAGE_SHIFT;
3184 last_index = (cluster->end - offset) >> PAGE_SHIFT;
3185 while (index <= last_index) {
3186 ret = btrfs_delalloc_reserve_metadata(inode, PAGE_SIZE);
3187 if (ret)
3188 goto out;
3189
3190 page = find_lock_page(inode->i_mapping, index);
3191 if (!page) {
3192 page_cache_sync_readahead(inode->i_mapping,
3193 ra, NULL, index,
3194 last_index + 1 - index);
3195 page = find_or_create_page(inode->i_mapping, index,
3196 mask);
3197 if (!page) {
3198 btrfs_delalloc_release_metadata(inode,
3199 PAGE_SIZE);
3200 ret = -ENOMEM;
3201 goto out;
3202 }
3203 }
3204
3205 if (PageReadahead(page)) {
3206 page_cache_async_readahead(inode->i_mapping,
3207 ra, NULL, page, index,
3208 last_index + 1 - index);
3209 }
3210
3211 if (!PageUptodate(page)) {
3212 btrfs_readpage(NULL, page);
3213 lock_page(page);
3214 if (!PageUptodate(page)) {
3215 unlock_page(page);
3216 put_page(page);
3217 btrfs_delalloc_release_metadata(inode,
3218 PAGE_SIZE);
3219 ret = -EIO;
3220 goto out;
3221 }
3222 }
3223
3224 page_start = page_offset(page);
3225 page_end = page_start + PAGE_SIZE - 1;
3226
3227 lock_extent(&BTRFS_I(inode)->io_tree, page_start, page_end);
3228
3229 set_page_extent_mapped(page);
3230
3231 if (nr < cluster->nr &&
3232 page_start + offset == cluster->boundary[nr]) {
3233 set_extent_bits(&BTRFS_I(inode)->io_tree,
3234 page_start, page_end,
3235 EXTENT_BOUNDARY);
3236 nr++;
3237 }
3238
3239 btrfs_set_extent_delalloc(inode, page_start, page_end, NULL, 0);
3240 set_page_dirty(page);
3241
3242 unlock_extent(&BTRFS_I(inode)->io_tree,
3243 page_start, page_end);
3244 unlock_page(page);
3245 put_page(page);
3246
3247 index++;
3248 balance_dirty_pages_ratelimited(inode->i_mapping);
3249 btrfs_throttle(BTRFS_I(inode)->root);
3250 }
3251 WARN_ON(nr != cluster->nr);
3252 out:
3253 kfree(ra);
3254 return ret;
3255 }
3256
3257 static noinline_for_stack
3258 int relocate_data_extent(struct inode *inode, struct btrfs_key *extent_key,
3259 struct file_extent_cluster *cluster)
3260 {
3261 int ret;
3262
3263 if (cluster->nr > 0 && extent_key->objectid != cluster->end + 1) {
3264 ret = relocate_file_extent_cluster(inode, cluster);
3265 if (ret)
3266 return ret;
3267 cluster->nr = 0;
3268 }
3269
3270 if (!cluster->nr)
3271 cluster->start = extent_key->objectid;
3272 else
3273 BUG_ON(cluster->nr >= MAX_EXTENTS);
3274 cluster->end = extent_key->objectid + extent_key->offset - 1;
3275 cluster->boundary[cluster->nr] = extent_key->objectid;
3276 cluster->nr++;
3277
3278 if (cluster->nr >= MAX_EXTENTS) {
3279 ret = relocate_file_extent_cluster(inode, cluster);
3280 if (ret)
3281 return ret;
3282 cluster->nr = 0;
3283 }
3284 return 0;
3285 }
3286
3287 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
3288 static int get_ref_objectid_v0(struct reloc_control *rc,
3289 struct btrfs_path *path,
3290 struct btrfs_key *extent_key,
3291 u64 *ref_objectid, int *path_change)
3292 {
3293 struct btrfs_key key;
3294 struct extent_buffer *leaf;
3295 struct btrfs_extent_ref_v0 *ref0;
3296 int ret;
3297 int slot;
3298
3299 leaf = path->nodes[0];
3300 slot = path->slots[0];
3301 while (1) {
3302 if (slot >= btrfs_header_nritems(leaf)) {
3303 ret = btrfs_next_leaf(rc->extent_root, path);
3304 if (ret < 0)
3305 return ret;
3306 BUG_ON(ret > 0);
3307 leaf = path->nodes[0];
3308 slot = path->slots[0];
3309 if (path_change)
3310 *path_change = 1;
3311 }
3312 btrfs_item_key_to_cpu(leaf, &key, slot);
3313 if (key.objectid != extent_key->objectid)
3314 return -ENOENT;
3315
3316 if (key.type != BTRFS_EXTENT_REF_V0_KEY) {
3317 slot++;
3318 continue;
3319 }
3320 ref0 = btrfs_item_ptr(leaf, slot,
3321 struct btrfs_extent_ref_v0);
3322 *ref_objectid = btrfs_ref_objectid_v0(leaf, ref0);
3323 break;
3324 }
3325 return 0;
3326 }
3327 #endif
3328
3329 /*
3330 * helper to add a tree block to the list.
3331 * the major work is getting the generation and level of the block
3332 */
3333 static int add_tree_block(struct reloc_control *rc,
3334 struct btrfs_key *extent_key,
3335 struct btrfs_path *path,
3336 struct rb_root *blocks)
3337 {
3338 struct extent_buffer *eb;
3339 struct btrfs_extent_item *ei;
3340 struct btrfs_tree_block_info *bi;
3341 struct tree_block *block;
3342 struct rb_node *rb_node;
3343 u32 item_size;
3344 int level = -1;
3345 u64 generation;
3346
3347 eb = path->nodes[0];
3348 item_size = btrfs_item_size_nr(eb, path->slots[0]);
3349
3350 if (extent_key->type == BTRFS_METADATA_ITEM_KEY ||
3351 item_size >= sizeof(*ei) + sizeof(*bi)) {
3352 ei = btrfs_item_ptr(eb, path->slots[0],
3353 struct btrfs_extent_item);
3354 if (extent_key->type == BTRFS_EXTENT_ITEM_KEY) {
3355 bi = (struct btrfs_tree_block_info *)(ei + 1);
3356 level = btrfs_tree_block_level(eb, bi);
3357 } else {
3358 level = (int)extent_key->offset;
3359 }
3360 generation = btrfs_extent_generation(eb, ei);
3361 } else {
3362 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
3363 u64 ref_owner;
3364 int ret;
3365
3366 BUG_ON(item_size != sizeof(struct btrfs_extent_item_v0));
3367 ret = get_ref_objectid_v0(rc, path, extent_key,
3368 &ref_owner, NULL);
3369 if (ret < 0)
3370 return ret;
3371 BUG_ON(ref_owner >= BTRFS_MAX_LEVEL);
3372 level = (int)ref_owner;
3373 /* FIXME: get real generation */
3374 generation = 0;
3375 #else
3376 BUG();
3377 #endif
3378 }
3379
3380 btrfs_release_path(path);
3381
3382 BUG_ON(level == -1);
3383
3384 block = kmalloc(sizeof(*block), GFP_NOFS);
3385 if (!block)
3386 return -ENOMEM;
3387
3388 block->bytenr = extent_key->objectid;
3389 block->key.objectid = rc->extent_root->nodesize;
3390 block->key.offset = generation;
3391 block->level = level;
3392 block->key_ready = 0;
3393
3394 rb_node = tree_insert(blocks, block->bytenr, &block->rb_node);
3395 if (rb_node)
3396 backref_tree_panic(rb_node, -EEXIST, block->bytenr);
3397
3398 return 0;
3399 }
3400
3401 /*
3402 * helper to add tree blocks for backref of type BTRFS_SHARED_DATA_REF_KEY
3403 */
3404 static int __add_tree_block(struct reloc_control *rc,
3405 u64 bytenr, u32 blocksize,
3406 struct rb_root *blocks)
3407 {
3408 struct btrfs_path *path;
3409 struct btrfs_key key;
3410 int ret;
3411 bool skinny = btrfs_fs_incompat(rc->extent_root->fs_info,
3412 SKINNY_METADATA);
3413
3414 if (tree_block_processed(bytenr, rc))
3415 return 0;
3416
3417 if (tree_search(blocks, bytenr))
3418 return 0;
3419
3420 path = btrfs_alloc_path();
3421 if (!path)
3422 return -ENOMEM;
3423 again:
3424 key.objectid = bytenr;
3425 if (skinny) {
3426 key.type = BTRFS_METADATA_ITEM_KEY;
3427 key.offset = (u64)-1;
3428 } else {
3429 key.type = BTRFS_EXTENT_ITEM_KEY;
3430 key.offset = blocksize;
3431 }
3432
3433 path->search_commit_root = 1;
3434 path->skip_locking = 1;
3435 ret = btrfs_search_slot(NULL, rc->extent_root, &key, path, 0, 0);
3436 if (ret < 0)
3437 goto out;
3438
3439 if (ret > 0 && skinny) {
3440 if (path->slots[0]) {
3441 path->slots[0]--;
3442 btrfs_item_key_to_cpu(path->nodes[0], &key,
3443 path->slots[0]);
3444 if (key.objectid == bytenr &&
3445 (key.type == BTRFS_METADATA_ITEM_KEY ||
3446 (key.type == BTRFS_EXTENT_ITEM_KEY &&
3447 key.offset == blocksize)))
3448 ret = 0;
3449 }
3450
3451 if (ret) {
3452 skinny = false;
3453 btrfs_release_path(path);
3454 goto again;
3455 }
3456 }
3457 BUG_ON(ret);
3458
3459 ret = add_tree_block(rc, &key, path, blocks);
3460 out:
3461 btrfs_free_path(path);
3462 return ret;
3463 }
3464
3465 /*
3466 * helper to check if the block use full backrefs for pointers in it
3467 */
3468 static int block_use_full_backref(struct reloc_control *rc,
3469 struct extent_buffer *eb)
3470 {
3471 u64 flags;
3472 int ret;
3473
3474 if (btrfs_header_flag(eb, BTRFS_HEADER_FLAG_RELOC) ||
3475 btrfs_header_backref_rev(eb) < BTRFS_MIXED_BACKREF_REV)
3476 return 1;
3477
3478 ret = btrfs_lookup_extent_info(NULL, rc->extent_root,
3479 eb->start, btrfs_header_level(eb), 1,
3480 NULL, &flags);
3481 BUG_ON(ret);
3482
3483 if (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF)
3484 ret = 1;
3485 else
3486 ret = 0;
3487 return ret;
3488 }
3489
3490 static int delete_block_group_cache(struct btrfs_fs_info *fs_info,
3491 struct btrfs_block_group_cache *block_group,
3492 struct inode *inode,
3493 u64 ino)
3494 {
3495 struct btrfs_key key;
3496 struct btrfs_root *root = fs_info->tree_root;
3497 struct btrfs_trans_handle *trans;
3498 int ret = 0;
3499
3500 if (inode)
3501 goto truncate;
3502
3503 key.objectid = ino;
3504 key.type = BTRFS_INODE_ITEM_KEY;
3505 key.offset = 0;
3506
3507 inode = btrfs_iget(fs_info->sb, &key, root, NULL);
3508 if (IS_ERR(inode) || is_bad_inode(inode)) {
3509 if (!IS_ERR(inode))
3510 iput(inode);
3511 return -ENOENT;
3512 }
3513
3514 truncate:
3515 ret = btrfs_check_trunc_cache_free_space(root,
3516 &fs_info->global_block_rsv);
3517 if (ret)
3518 goto out;
3519
3520 trans = btrfs_join_transaction(root);
3521 if (IS_ERR(trans)) {
3522 ret = PTR_ERR(trans);
3523 goto out;
3524 }
3525
3526 ret = btrfs_truncate_free_space_cache(root, trans, block_group, inode);
3527
3528 btrfs_end_transaction(trans, root);
3529 btrfs_btree_balance_dirty(root);
3530 out:
3531 iput(inode);
3532 return ret;
3533 }
3534
3535 /*
3536 * helper to add tree blocks for backref of type BTRFS_EXTENT_DATA_REF_KEY
3537 * this function scans fs tree to find blocks reference the data extent
3538 */
3539 static int find_data_references(struct reloc_control *rc,
3540 struct btrfs_key *extent_key,
3541 struct extent_buffer *leaf,
3542 struct btrfs_extent_data_ref *ref,
3543 struct rb_root *blocks)
3544 {
3545 struct btrfs_path *path;
3546 struct tree_block *block;
3547 struct btrfs_root *root;
3548 struct btrfs_file_extent_item *fi;
3549 struct rb_node *rb_node;
3550 struct btrfs_key key;
3551 u64 ref_root;
3552 u64 ref_objectid;
3553 u64 ref_offset;
3554 u32 ref_count;
3555 u32 nritems;
3556 int err = 0;
3557 int added = 0;
3558 int counted;
3559 int ret;
3560
3561 ref_root = btrfs_extent_data_ref_root(leaf, ref);
3562 ref_objectid = btrfs_extent_data_ref_objectid(leaf, ref);
3563 ref_offset = btrfs_extent_data_ref_offset(leaf, ref);
3564 ref_count = btrfs_extent_data_ref_count(leaf, ref);
3565
3566 /*
3567 * This is an extent belonging to the free space cache, lets just delete
3568 * it and redo the search.
3569 */
3570 if (ref_root == BTRFS_ROOT_TREE_OBJECTID) {
3571 ret = delete_block_group_cache(rc->extent_root->fs_info,
3572 rc->block_group,
3573 NULL, ref_objectid);
3574 if (ret != -ENOENT)
3575 return ret;
3576 ret = 0;
3577 }
3578
3579 path = btrfs_alloc_path();
3580 if (!path)
3581 return -ENOMEM;
3582 path->reada = READA_FORWARD;
3583
3584 root = read_fs_root(rc->extent_root->fs_info, ref_root);
3585 if (IS_ERR(root)) {
3586 err = PTR_ERR(root);
3587 goto out;
3588 }
3589
3590 key.objectid = ref_objectid;
3591 key.type = BTRFS_EXTENT_DATA_KEY;
3592 if (ref_offset > ((u64)-1 << 32))
3593 key.offset = 0;
3594 else
3595 key.offset = ref_offset;
3596
3597 path->search_commit_root = 1;
3598 path->skip_locking = 1;
3599 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
3600 if (ret < 0) {
3601 err = ret;
3602 goto out;
3603 }
3604
3605 leaf = path->nodes[0];
3606 nritems = btrfs_header_nritems(leaf);
3607 /*
3608 * the references in tree blocks that use full backrefs
3609 * are not counted in
3610 */
3611 if (block_use_full_backref(rc, leaf))
3612 counted = 0;
3613 else
3614 counted = 1;
3615 rb_node = tree_search(blocks, leaf->start);
3616 if (rb_node) {
3617 if (counted)
3618 added = 1;
3619 else
3620 path->slots[0] = nritems;
3621 }
3622
3623 while (ref_count > 0) {
3624 while (path->slots[0] >= nritems) {
3625 ret = btrfs_next_leaf(root, path);
3626 if (ret < 0) {
3627 err = ret;
3628 goto out;
3629 }
3630 if (WARN_ON(ret > 0))
3631 goto out;
3632
3633 leaf = path->nodes[0];
3634 nritems = btrfs_header_nritems(leaf);
3635 added = 0;
3636
3637 if (block_use_full_backref(rc, leaf))
3638 counted = 0;
3639 else
3640 counted = 1;
3641 rb_node = tree_search(blocks, leaf->start);
3642 if (rb_node) {
3643 if (counted)
3644 added = 1;
3645 else
3646 path->slots[0] = nritems;
3647 }
3648 }
3649
3650 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
3651 if (WARN_ON(key.objectid != ref_objectid ||
3652 key.type != BTRFS_EXTENT_DATA_KEY))
3653 break;
3654
3655 fi = btrfs_item_ptr(leaf, path->slots[0],
3656 struct btrfs_file_extent_item);
3657
3658 if (btrfs_file_extent_type(leaf, fi) ==
3659 BTRFS_FILE_EXTENT_INLINE)
3660 goto next;
3661
3662 if (btrfs_file_extent_disk_bytenr(leaf, fi) !=
3663 extent_key->objectid)
3664 goto next;
3665
3666 key.offset -= btrfs_file_extent_offset(leaf, fi);
3667 if (key.offset != ref_offset)
3668 goto next;
3669
3670 if (counted)
3671 ref_count--;
3672 if (added)
3673 goto next;
3674
3675 if (!tree_block_processed(leaf->start, rc)) {
3676 block = kmalloc(sizeof(*block), GFP_NOFS);
3677 if (!block) {
3678 err = -ENOMEM;
3679 break;
3680 }
3681 block->bytenr = leaf->start;
3682 btrfs_item_key_to_cpu(leaf, &block->key, 0);
3683 block->level = 0;
3684 block->key_ready = 1;
3685 rb_node = tree_insert(blocks, block->bytenr,
3686 &block->rb_node);
3687 if (rb_node)
3688 backref_tree_panic(rb_node, -EEXIST,
3689 block->bytenr);
3690 }
3691 if (counted)
3692 added = 1;
3693 else
3694 path->slots[0] = nritems;
3695 next:
3696 path->slots[0]++;
3697
3698 }
3699 out:
3700 btrfs_free_path(path);
3701 return err;
3702 }
3703
3704 /*
3705 * helper to find all tree blocks that reference a given data extent
3706 */
3707 static noinline_for_stack
3708 int add_data_references(struct reloc_control *rc,
3709 struct btrfs_key *extent_key,
3710 struct btrfs_path *path,
3711 struct rb_root *blocks)
3712 {
3713 struct btrfs_key key;
3714 struct extent_buffer *eb;
3715 struct btrfs_extent_data_ref *dref;
3716 struct btrfs_extent_inline_ref *iref;
3717 unsigned long ptr;
3718 unsigned long end;
3719 u32 blocksize = rc->extent_root->nodesize;
3720 int ret = 0;
3721 int err = 0;
3722
3723 eb = path->nodes[0];
3724 ptr = btrfs_item_ptr_offset(eb, path->slots[0]);
3725 end = ptr + btrfs_item_size_nr(eb, path->slots[0]);
3726 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
3727 if (ptr + sizeof(struct btrfs_extent_item_v0) == end)
3728 ptr = end;
3729 else
3730 #endif
3731 ptr += sizeof(struct btrfs_extent_item);
3732
3733 while (ptr < end) {
3734 iref = (struct btrfs_extent_inline_ref *)ptr;
3735 key.type = btrfs_extent_inline_ref_type(eb, iref);
3736 if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
3737 key.offset = btrfs_extent_inline_ref_offset(eb, iref);
3738 ret = __add_tree_block(rc, key.offset, blocksize,
3739 blocks);
3740 } else if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
3741 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
3742 ret = find_data_references(rc, extent_key,
3743 eb, dref, blocks);
3744 } else {
3745 BUG();
3746 }
3747 if (ret) {
3748 err = ret;
3749 goto out;
3750 }
3751 ptr += btrfs_extent_inline_ref_size(key.type);
3752 }
3753 WARN_ON(ptr > end);
3754
3755 while (1) {
3756 cond_resched();
3757 eb = path->nodes[0];
3758 if (path->slots[0] >= btrfs_header_nritems(eb)) {
3759 ret = btrfs_next_leaf(rc->extent_root, path);
3760 if (ret < 0) {
3761 err = ret;
3762 break;
3763 }
3764 if (ret > 0)
3765 break;
3766 eb = path->nodes[0];
3767 }
3768
3769 btrfs_item_key_to_cpu(eb, &key, path->slots[0]);
3770 if (key.objectid != extent_key->objectid)
3771 break;
3772
3773 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
3774 if (key.type == BTRFS_SHARED_DATA_REF_KEY ||
3775 key.type == BTRFS_EXTENT_REF_V0_KEY) {
3776 #else
3777 BUG_ON(key.type == BTRFS_EXTENT_REF_V0_KEY);
3778 if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
3779 #endif
3780 ret = __add_tree_block(rc, key.offset, blocksize,
3781 blocks);
3782 } else if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
3783 dref = btrfs_item_ptr(eb, path->slots[0],
3784 struct btrfs_extent_data_ref);
3785 ret = find_data_references(rc, extent_key,
3786 eb, dref, blocks);
3787 } else {
3788 ret = 0;
3789 }
3790 if (ret) {
3791 err = ret;
3792 break;
3793 }
3794 path->slots[0]++;
3795 }
3796 out:
3797 btrfs_release_path(path);
3798 if (err)
3799 free_block_list(blocks);
3800 return err;
3801 }
3802
3803 /*
3804 * helper to find next unprocessed extent
3805 */
3806 static noinline_for_stack
3807 int find_next_extent(struct reloc_control *rc, struct btrfs_path *path,
3808 struct btrfs_key *extent_key)
3809 {
3810 struct btrfs_key key;
3811 struct extent_buffer *leaf;
3812 u64 start, end, last;
3813 int ret;
3814
3815 last = rc->block_group->key.objectid + rc->block_group->key.offset;
3816 while (1) {
3817 cond_resched();
3818 if (rc->search_start >= last) {
3819 ret = 1;
3820 break;
3821 }
3822
3823 key.objectid = rc->search_start;
3824 key.type = BTRFS_EXTENT_ITEM_KEY;
3825 key.offset = 0;
3826
3827 path->search_commit_root = 1;
3828 path->skip_locking = 1;
3829 ret = btrfs_search_slot(NULL, rc->extent_root, &key, path,
3830 0, 0);
3831 if (ret < 0)
3832 break;
3833 next:
3834 leaf = path->nodes[0];
3835 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
3836 ret = btrfs_next_leaf(rc->extent_root, path);
3837 if (ret != 0)
3838 break;
3839 leaf = path->nodes[0];
3840 }
3841
3842 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
3843 if (key.objectid >= last) {
3844 ret = 1;
3845 break;
3846 }
3847
3848 if (key.type != BTRFS_EXTENT_ITEM_KEY &&
3849 key.type != BTRFS_METADATA_ITEM_KEY) {
3850 path->slots[0]++;
3851 goto next;
3852 }
3853
3854 if (key.type == BTRFS_EXTENT_ITEM_KEY &&
3855 key.objectid + key.offset <= rc->search_start) {
3856 path->slots[0]++;
3857 goto next;
3858 }
3859
3860 if (key.type == BTRFS_METADATA_ITEM_KEY &&
3861 key.objectid + rc->extent_root->nodesize <=
3862 rc->search_start) {
3863 path->slots[0]++;
3864 goto next;
3865 }
3866
3867 ret = find_first_extent_bit(&rc->processed_blocks,
3868 key.objectid, &start, &end,
3869 EXTENT_DIRTY, NULL);
3870
3871 if (ret == 0 && start <= key.objectid) {
3872 btrfs_release_path(path);
3873 rc->search_start = end + 1;
3874 } else {
3875 if (key.type == BTRFS_EXTENT_ITEM_KEY)
3876 rc->search_start = key.objectid + key.offset;
3877 else
3878 rc->search_start = key.objectid +
3879 rc->extent_root->nodesize;
3880 memcpy(extent_key, &key, sizeof(key));
3881 return 0;
3882 }
3883 }
3884 btrfs_release_path(path);
3885 return ret;
3886 }
3887
3888 static void set_reloc_control(struct reloc_control *rc)
3889 {
3890 struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
3891
3892 mutex_lock(&fs_info->reloc_mutex);
3893 fs_info->reloc_ctl = rc;
3894 mutex_unlock(&fs_info->reloc_mutex);
3895 }
3896
3897 static void unset_reloc_control(struct reloc_control *rc)
3898 {
3899 struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
3900
3901 mutex_lock(&fs_info->reloc_mutex);
3902 fs_info->reloc_ctl = NULL;
3903 mutex_unlock(&fs_info->reloc_mutex);
3904 }
3905
3906 static int check_extent_flags(u64 flags)
3907 {
3908 if ((flags & BTRFS_EXTENT_FLAG_DATA) &&
3909 (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK))
3910 return 1;
3911 if (!(flags & BTRFS_EXTENT_FLAG_DATA) &&
3912 !(flags & BTRFS_EXTENT_FLAG_TREE_BLOCK))
3913 return 1;
3914 if ((flags & BTRFS_EXTENT_FLAG_DATA) &&
3915 (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
3916 return 1;
3917 return 0;
3918 }
3919
3920 static noinline_for_stack
3921 int prepare_to_relocate(struct reloc_control *rc)
3922 {
3923 struct btrfs_trans_handle *trans;
3924 int ret;
3925
3926 rc->block_rsv = btrfs_alloc_block_rsv(rc->extent_root,
3927 BTRFS_BLOCK_RSV_TEMP);
3928 if (!rc->block_rsv)
3929 return -ENOMEM;
3930
3931 memset(&rc->cluster, 0, sizeof(rc->cluster));
3932 rc->search_start = rc->block_group->key.objectid;
3933 rc->extents_found = 0;
3934 rc->nodes_relocated = 0;
3935 rc->merging_rsv_size = 0;
3936 rc->reserved_bytes = 0;
3937 rc->block_rsv->size = rc->extent_root->nodesize *
3938 RELOCATION_RESERVED_NODES;
3939 ret = btrfs_block_rsv_refill(rc->extent_root,
3940 rc->block_rsv, rc->block_rsv->size,
3941 BTRFS_RESERVE_FLUSH_ALL);
3942 if (ret)
3943 return ret;
3944
3945 rc->create_reloc_tree = 1;
3946 set_reloc_control(rc);
3947
3948 trans = btrfs_join_transaction(rc->extent_root);
3949 if (IS_ERR(trans)) {
3950 unset_reloc_control(rc);
3951 /*
3952 * extent tree is not a ref_cow tree and has no reloc_root to
3953 * cleanup. And callers are responsible to free the above
3954 * block rsv.
3955 */
3956 return PTR_ERR(trans);
3957 }
3958 btrfs_commit_transaction(trans, rc->extent_root);
3959 return 0;
3960 }
3961
3962 /*
3963 * Qgroup fixer for data chunk relocation.
3964 * The data relocation is done in the following steps
3965 * 1) Copy data extents into data reloc tree
3966 * 2) Create tree reloc tree(special snapshot) for related subvolumes
3967 * 3) Modify file extents in tree reloc tree
3968 * 4) Merge tree reloc tree with original fs tree, by swapping tree blocks
3969 *
3970 * The problem is, data and tree reloc tree are not accounted to qgroup,
3971 * and 4) will only info qgroup to track tree blocks change, not file extents
3972 * in the tree blocks.
3973 *
3974 * The good news is, related data extents are all in data reloc tree, so we
3975 * only need to info qgroup to track all file extents in data reloc tree
3976 * before commit trans.
3977 */
3978 static int qgroup_fix_relocated_data_extents(struct btrfs_trans_handle *trans,
3979 struct reloc_control *rc)
3980 {
3981 struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
3982 struct inode *inode = rc->data_inode;
3983 struct btrfs_root *data_reloc_root = BTRFS_I(inode)->root;
3984 struct btrfs_path *path;
3985 struct btrfs_key key;
3986 int ret = 0;
3987
3988 if (!test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags))
3989 return 0;
3990
3991 /*
3992 * Only for stage where we update data pointers the qgroup fix is
3993 * valid.
3994 * For MOVING_DATA stage, we will miss the timing of swapping tree
3995 * blocks, and won't fix it.
3996 */
3997 if (!(rc->stage == UPDATE_DATA_PTRS && rc->extents_found))
3998 return 0;
3999
4000 path = btrfs_alloc_path();
4001 if (!path)
4002 return -ENOMEM;
4003 key.objectid = btrfs_ino(inode);
4004 key.type = BTRFS_EXTENT_DATA_KEY;
4005 key.offset = 0;
4006
4007 ret = btrfs_search_slot(NULL, data_reloc_root, &key, path, 0, 0);
4008 if (ret < 0)
4009 goto out;
4010
4011 lock_extent(&BTRFS_I(inode)->io_tree, 0, (u64)-1);
4012 while (1) {
4013 struct btrfs_file_extent_item *fi;
4014
4015 btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0]);
4016 if (key.objectid > btrfs_ino(inode))
4017 break;
4018 if (key.type != BTRFS_EXTENT_DATA_KEY)
4019 goto next;
4020 fi = btrfs_item_ptr(path->nodes[0], path->slots[0],
4021 struct btrfs_file_extent_item);
4022 if (btrfs_file_extent_type(path->nodes[0], fi) !=
4023 BTRFS_FILE_EXTENT_REG)
4024 goto next;
4025 ret = btrfs_qgroup_insert_dirty_extent(trans, fs_info,
4026 btrfs_file_extent_disk_bytenr(path->nodes[0], fi),
4027 btrfs_file_extent_disk_num_bytes(path->nodes[0], fi),
4028 GFP_NOFS);
4029 if (ret < 0)
4030 break;
4031 next:
4032 ret = btrfs_next_item(data_reloc_root, path);
4033 if (ret < 0)
4034 break;
4035 if (ret > 0) {
4036 ret = 0;
4037 break;
4038 }
4039 }
4040 unlock_extent(&BTRFS_I(inode)->io_tree, 0 , (u64)-1);
4041 out:
4042 btrfs_free_path(path);
4043 return ret;
4044 }
4045
4046 static noinline_for_stack int relocate_block_group(struct reloc_control *rc)
4047 {
4048 struct rb_root blocks = RB_ROOT;
4049 struct btrfs_key key;
4050 struct btrfs_trans_handle *trans = NULL;
4051 struct btrfs_path *path;
4052 struct btrfs_extent_item *ei;
4053 u64 flags;
4054 u32 item_size;
4055 int ret;
4056 int err = 0;
4057 int progress = 0;
4058
4059 path = btrfs_alloc_path();
4060 if (!path)
4061 return -ENOMEM;
4062 path->reada = READA_FORWARD;
4063
4064 ret = prepare_to_relocate(rc);
4065 if (ret) {
4066 err = ret;
4067 goto out_free;
4068 }
4069
4070 while (1) {
4071 rc->reserved_bytes = 0;
4072 ret = btrfs_block_rsv_refill(rc->extent_root,
4073 rc->block_rsv, rc->block_rsv->size,
4074 BTRFS_RESERVE_FLUSH_ALL);
4075 if (ret) {
4076 err = ret;
4077 break;
4078 }
4079 progress++;
4080 trans = btrfs_start_transaction(rc->extent_root, 0);
4081 if (IS_ERR(trans)) {
4082 err = PTR_ERR(trans);
4083 trans = NULL;
4084 break;
4085 }
4086 restart:
4087 if (update_backref_cache(trans, &rc->backref_cache)) {
4088 btrfs_end_transaction(trans, rc->extent_root);
4089 continue;
4090 }
4091
4092 ret = find_next_extent(rc, path, &key);
4093 if (ret < 0)
4094 err = ret;
4095 if (ret != 0)
4096 break;
4097
4098 rc->extents_found++;
4099
4100 ei = btrfs_item_ptr(path->nodes[0], path->slots[0],
4101 struct btrfs_extent_item);
4102 item_size = btrfs_item_size_nr(path->nodes[0], path->slots[0]);
4103 if (item_size >= sizeof(*ei)) {
4104 flags = btrfs_extent_flags(path->nodes[0], ei);
4105 ret = check_extent_flags(flags);
4106 BUG_ON(ret);
4107
4108 } else {
4109 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
4110 u64 ref_owner;
4111 int path_change = 0;
4112
4113 BUG_ON(item_size !=
4114 sizeof(struct btrfs_extent_item_v0));
4115 ret = get_ref_objectid_v0(rc, path, &key, &ref_owner,
4116 &path_change);
4117 if (ret < 0) {
4118 err = ret;
4119 break;
4120 }
4121 if (ref_owner < BTRFS_FIRST_FREE_OBJECTID)
4122 flags = BTRFS_EXTENT_FLAG_TREE_BLOCK;
4123 else
4124 flags = BTRFS_EXTENT_FLAG_DATA;
4125
4126 if (path_change) {
4127 btrfs_release_path(path);
4128
4129 path->search_commit_root = 1;
4130 path->skip_locking = 1;
4131 ret = btrfs_search_slot(NULL, rc->extent_root,
4132 &key, path, 0, 0);
4133 if (ret < 0) {
4134 err = ret;
4135 break;
4136 }
4137 BUG_ON(ret > 0);
4138 }
4139 #else
4140 BUG();
4141 #endif
4142 }
4143
4144 if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
4145 ret = add_tree_block(rc, &key, path, &blocks);
4146 } else if (rc->stage == UPDATE_DATA_PTRS &&
4147 (flags & BTRFS_EXTENT_FLAG_DATA)) {
4148 ret = add_data_references(rc, &key, path, &blocks);
4149 } else {
4150 btrfs_release_path(path);
4151 ret = 0;
4152 }
4153 if (ret < 0) {
4154 err = ret;
4155 break;
4156 }
4157
4158 if (!RB_EMPTY_ROOT(&blocks)) {
4159 ret = relocate_tree_blocks(trans, rc, &blocks);
4160 if (ret < 0) {
4161 /*
4162 * if we fail to relocate tree blocks, force to update
4163 * backref cache when committing transaction.
4164 */
4165 rc->backref_cache.last_trans = trans->transid - 1;
4166
4167 if (ret != -EAGAIN) {
4168 err = ret;
4169 break;
4170 }
4171 rc->extents_found--;
4172 rc->search_start = key.objectid;
4173 }
4174 }
4175
4176 btrfs_end_transaction_throttle(trans, rc->extent_root);
4177 btrfs_btree_balance_dirty(rc->extent_root);
4178 trans = NULL;
4179
4180 if (rc->stage == MOVE_DATA_EXTENTS &&
4181 (flags & BTRFS_EXTENT_FLAG_DATA)) {
4182 rc->found_file_extent = 1;
4183 ret = relocate_data_extent(rc->data_inode,
4184 &key, &rc->cluster);
4185 if (ret < 0) {
4186 err = ret;
4187 break;
4188 }
4189 }
4190 }
4191 if (trans && progress && err == -ENOSPC) {
4192 ret = btrfs_force_chunk_alloc(trans, rc->extent_root,
4193 rc->block_group->flags);
4194 if (ret == 1) {
4195 err = 0;
4196 progress = 0;
4197 goto restart;
4198 }
4199 }
4200
4201 btrfs_release_path(path);
4202 clear_extent_bits(&rc->processed_blocks, 0, (u64)-1, EXTENT_DIRTY);
4203
4204 if (trans) {
4205 btrfs_end_transaction_throttle(trans, rc->extent_root);
4206 btrfs_btree_balance_dirty(rc->extent_root);
4207 }
4208
4209 if (!err) {
4210 ret = relocate_file_extent_cluster(rc->data_inode,
4211 &rc->cluster);
4212 if (ret < 0)
4213 err = ret;
4214 }
4215
4216 rc->create_reloc_tree = 0;
4217 set_reloc_control(rc);
4218
4219 backref_cache_cleanup(&rc->backref_cache);
4220 btrfs_block_rsv_release(rc->extent_root, rc->block_rsv, (u64)-1);
4221
4222 err = prepare_to_merge(rc, err);
4223
4224 merge_reloc_roots(rc);
4225
4226 rc->merge_reloc_tree = 0;
4227 unset_reloc_control(rc);
4228 btrfs_block_rsv_release(rc->extent_root, rc->block_rsv, (u64)-1);
4229
4230 /* get rid of pinned extents */
4231 trans = btrfs_join_transaction(rc->extent_root);
4232 if (IS_ERR(trans)) {
4233 err = PTR_ERR(trans);
4234 goto out_free;
4235 }
4236 ret = qgroup_fix_relocated_data_extents(trans, rc);
4237 if (ret < 0) {
4238 btrfs_abort_transaction(trans, ret);
4239 if (!err)
4240 err = ret;
4241 goto out_free;
4242 }
4243 btrfs_commit_transaction(trans, rc->extent_root);
4244 out_free:
4245 btrfs_free_block_rsv(rc->extent_root, rc->block_rsv);
4246 btrfs_free_path(path);
4247 return err;
4248 }
4249
4250 static int __insert_orphan_inode(struct btrfs_trans_handle *trans,
4251 struct btrfs_root *root, u64 objectid)
4252 {
4253 struct btrfs_path *path;
4254 struct btrfs_inode_item *item;
4255 struct extent_buffer *leaf;
4256 int ret;
4257
4258 path = btrfs_alloc_path();
4259 if (!path)
4260 return -ENOMEM;
4261
4262 ret = btrfs_insert_empty_inode(trans, root, path, objectid);
4263 if (ret)
4264 goto out;
4265
4266 leaf = path->nodes[0];
4267 item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_inode_item);
4268 memset_extent_buffer(leaf, 0, (unsigned long)item, sizeof(*item));
4269 btrfs_set_inode_generation(leaf, item, 1);
4270 btrfs_set_inode_size(leaf, item, 0);
4271 btrfs_set_inode_mode(leaf, item, S_IFREG | 0600);
4272 btrfs_set_inode_flags(leaf, item, BTRFS_INODE_NOCOMPRESS |
4273 BTRFS_INODE_PREALLOC);
4274 btrfs_mark_buffer_dirty(leaf);
4275 out:
4276 btrfs_free_path(path);
4277 return ret;
4278 }
4279
4280 /*
4281 * helper to create inode for data relocation.
4282 * the inode is in data relocation tree and its link count is 0
4283 */
4284 static noinline_for_stack
4285 struct inode *create_reloc_inode(struct btrfs_fs_info *fs_info,
4286 struct btrfs_block_group_cache *group)
4287 {
4288 struct inode *inode = NULL;
4289 struct btrfs_trans_handle *trans;
4290 struct btrfs_root *root;
4291 struct btrfs_key key;
4292 u64 objectid;
4293 int err = 0;
4294
4295 root = read_fs_root(fs_info, BTRFS_DATA_RELOC_TREE_OBJECTID);
4296 if (IS_ERR(root))
4297 return ERR_CAST(root);
4298
4299 trans = btrfs_start_transaction(root, 6);
4300 if (IS_ERR(trans))
4301 return ERR_CAST(trans);
4302
4303 err = btrfs_find_free_objectid(root, &objectid);
4304 if (err)
4305 goto out;
4306
4307 err = __insert_orphan_inode(trans, root, objectid);
4308 BUG_ON(err);
4309
4310 key.objectid = objectid;
4311 key.type = BTRFS_INODE_ITEM_KEY;
4312 key.offset = 0;
4313 inode = btrfs_iget(root->fs_info->sb, &key, root, NULL);
4314 BUG_ON(IS_ERR(inode) || is_bad_inode(inode));
4315 BTRFS_I(inode)->index_cnt = group->key.objectid;
4316
4317 err = btrfs_orphan_add(trans, inode);
4318 out:
4319 btrfs_end_transaction(trans, root);
4320 btrfs_btree_balance_dirty(root);
4321 if (err) {
4322 if (inode)
4323 iput(inode);
4324 inode = ERR_PTR(err);
4325 }
4326 return inode;
4327 }
4328
4329 static struct reloc_control *alloc_reloc_control(struct btrfs_fs_info *fs_info)
4330 {
4331 struct reloc_control *rc;
4332
4333 rc = kzalloc(sizeof(*rc), GFP_NOFS);
4334 if (!rc)
4335 return NULL;
4336
4337 INIT_LIST_HEAD(&rc->reloc_roots);
4338 backref_cache_init(&rc->backref_cache);
4339 mapping_tree_init(&rc->reloc_root_tree);
4340 extent_io_tree_init(&rc->processed_blocks,
4341 fs_info->btree_inode->i_mapping);
4342 return rc;
4343 }
4344
4345 /*
4346 * function to relocate all extents in a block group.
4347 */
4348 int btrfs_relocate_block_group(struct btrfs_root *extent_root, u64 group_start)
4349 {
4350 struct btrfs_fs_info *fs_info = extent_root->fs_info;
4351 struct reloc_control *rc;
4352 struct inode *inode;
4353 struct btrfs_path *path;
4354 int ret;
4355 int rw = 0;
4356 int err = 0;
4357
4358 rc = alloc_reloc_control(fs_info);
4359 if (!rc)
4360 return -ENOMEM;
4361
4362 rc->extent_root = extent_root;
4363
4364 rc->block_group = btrfs_lookup_block_group(fs_info, group_start);
4365 BUG_ON(!rc->block_group);
4366
4367 ret = btrfs_inc_block_group_ro(extent_root, rc->block_group);
4368 if (ret) {
4369 err = ret;
4370 goto out;
4371 }
4372 rw = 1;
4373
4374 path = btrfs_alloc_path();
4375 if (!path) {
4376 err = -ENOMEM;
4377 goto out;
4378 }
4379
4380 inode = lookup_free_space_inode(fs_info->tree_root, rc->block_group,
4381 path);
4382 btrfs_free_path(path);
4383
4384 if (!IS_ERR(inode))
4385 ret = delete_block_group_cache(fs_info, rc->block_group, inode, 0);
4386 else
4387 ret = PTR_ERR(inode);
4388
4389 if (ret && ret != -ENOENT) {
4390 err = ret;
4391 goto out;
4392 }
4393
4394 rc->data_inode = create_reloc_inode(fs_info, rc->block_group);
4395 if (IS_ERR(rc->data_inode)) {
4396 err = PTR_ERR(rc->data_inode);
4397 rc->data_inode = NULL;
4398 goto out;
4399 }
4400
4401 btrfs_info(extent_root->fs_info,
4402 "relocating block group %llu flags %llu",
4403 rc->block_group->key.objectid, rc->block_group->flags);
4404
4405 btrfs_wait_block_group_reservations(rc->block_group);
4406 btrfs_wait_nocow_writers(rc->block_group);
4407 btrfs_wait_ordered_roots(fs_info, -1,
4408 rc->block_group->key.objectid,
4409 rc->block_group->key.offset);
4410
4411 while (1) {
4412 mutex_lock(&fs_info->cleaner_mutex);
4413 ret = relocate_block_group(rc);
4414 mutex_unlock(&fs_info->cleaner_mutex);
4415 if (ret < 0) {
4416 err = ret;
4417 goto out;
4418 }
4419
4420 if (rc->extents_found == 0)
4421 break;
4422
4423 btrfs_info(extent_root->fs_info, "found %llu extents",
4424 rc->extents_found);
4425
4426 if (rc->stage == MOVE_DATA_EXTENTS && rc->found_file_extent) {
4427 ret = btrfs_wait_ordered_range(rc->data_inode, 0,
4428 (u64)-1);
4429 if (ret) {
4430 err = ret;
4431 goto out;
4432 }
4433 invalidate_mapping_pages(rc->data_inode->i_mapping,
4434 0, -1);
4435 rc->stage = UPDATE_DATA_PTRS;
4436 }
4437 }
4438
4439 WARN_ON(rc->block_group->pinned > 0);
4440 WARN_ON(rc->block_group->reserved > 0);
4441 WARN_ON(btrfs_block_group_used(&rc->block_group->item) > 0);
4442 out:
4443 if (err && rw)
4444 btrfs_dec_block_group_ro(extent_root, rc->block_group);
4445 iput(rc->data_inode);
4446 btrfs_put_block_group(rc->block_group);
4447 kfree(rc);
4448 return err;
4449 }
4450
4451 static noinline_for_stack int mark_garbage_root(struct btrfs_root *root)
4452 {
4453 struct btrfs_trans_handle *trans;
4454 int ret, err;
4455
4456 trans = btrfs_start_transaction(root->fs_info->tree_root, 0);
4457 if (IS_ERR(trans))
4458 return PTR_ERR(trans);
4459
4460 memset(&root->root_item.drop_progress, 0,
4461 sizeof(root->root_item.drop_progress));
4462 root->root_item.drop_level = 0;
4463 btrfs_set_root_refs(&root->root_item, 0);
4464 ret = btrfs_update_root(trans, root->fs_info->tree_root,
4465 &root->root_key, &root->root_item);
4466
4467 err = btrfs_end_transaction(trans, root->fs_info->tree_root);
4468 if (err)
4469 return err;
4470 return ret;
4471 }
4472
4473 /*
4474 * recover relocation interrupted by system crash.
4475 *
4476 * this function resumes merging reloc trees with corresponding fs trees.
4477 * this is important for keeping the sharing of tree blocks
4478 */
4479 int btrfs_recover_relocation(struct btrfs_root *root)
4480 {
4481 LIST_HEAD(reloc_roots);
4482 struct btrfs_key key;
4483 struct btrfs_root *fs_root;
4484 struct btrfs_root *reloc_root;
4485 struct btrfs_path *path;
4486 struct extent_buffer *leaf;
4487 struct reloc_control *rc = NULL;
4488 struct btrfs_trans_handle *trans;
4489 int ret;
4490 int err = 0;
4491
4492 path = btrfs_alloc_path();
4493 if (!path)
4494 return -ENOMEM;
4495 path->reada = READA_BACK;
4496
4497 key.objectid = BTRFS_TREE_RELOC_OBJECTID;
4498 key.type = BTRFS_ROOT_ITEM_KEY;
4499 key.offset = (u64)-1;
4500
4501 while (1) {
4502 ret = btrfs_search_slot(NULL, root->fs_info->tree_root, &key,
4503 path, 0, 0);
4504 if (ret < 0) {
4505 err = ret;
4506 goto out;
4507 }
4508 if (ret > 0) {
4509 if (path->slots[0] == 0)
4510 break;
4511 path->slots[0]--;
4512 }
4513 leaf = path->nodes[0];
4514 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
4515 btrfs_release_path(path);
4516
4517 if (key.objectid != BTRFS_TREE_RELOC_OBJECTID ||
4518 key.type != BTRFS_ROOT_ITEM_KEY)
4519 break;
4520
4521 reloc_root = btrfs_read_fs_root(root, &key);
4522 if (IS_ERR(reloc_root)) {
4523 err = PTR_ERR(reloc_root);
4524 goto out;
4525 }
4526
4527 list_add(&reloc_root->root_list, &reloc_roots);
4528
4529 if (btrfs_root_refs(&reloc_root->root_item) > 0) {
4530 fs_root = read_fs_root(root->fs_info,
4531 reloc_root->root_key.offset);
4532 if (IS_ERR(fs_root)) {
4533 ret = PTR_ERR(fs_root);
4534 if (ret != -ENOENT) {
4535 err = ret;
4536 goto out;
4537 }
4538 ret = mark_garbage_root(reloc_root);
4539 if (ret < 0) {
4540 err = ret;
4541 goto out;
4542 }
4543 }
4544 }
4545
4546 if (key.offset == 0)
4547 break;
4548
4549 key.offset--;
4550 }
4551 btrfs_release_path(path);
4552
4553 if (list_empty(&reloc_roots))
4554 goto out;
4555
4556 rc = alloc_reloc_control(root->fs_info);
4557 if (!rc) {
4558 err = -ENOMEM;
4559 goto out;
4560 }
4561
4562 rc->extent_root = root->fs_info->extent_root;
4563
4564 set_reloc_control(rc);
4565
4566 trans = btrfs_join_transaction(rc->extent_root);
4567 if (IS_ERR(trans)) {
4568 unset_reloc_control(rc);
4569 err = PTR_ERR(trans);
4570 goto out_free;
4571 }
4572
4573 rc->merge_reloc_tree = 1;
4574
4575 while (!list_empty(&reloc_roots)) {
4576 reloc_root = list_entry(reloc_roots.next,
4577 struct btrfs_root, root_list);
4578 list_del(&reloc_root->root_list);
4579
4580 if (btrfs_root_refs(&reloc_root->root_item) == 0) {
4581 list_add_tail(&reloc_root->root_list,
4582 &rc->reloc_roots);
4583 continue;
4584 }
4585
4586 fs_root = read_fs_root(root->fs_info,
4587 reloc_root->root_key.offset);
4588 if (IS_ERR(fs_root)) {
4589 err = PTR_ERR(fs_root);
4590 goto out_free;
4591 }
4592
4593 err = __add_reloc_root(reloc_root);
4594 BUG_ON(err < 0); /* -ENOMEM or logic error */
4595 fs_root->reloc_root = reloc_root;
4596 }
4597
4598 err = btrfs_commit_transaction(trans, rc->extent_root);
4599 if (err)
4600 goto out_free;
4601
4602 merge_reloc_roots(rc);
4603
4604 unset_reloc_control(rc);
4605
4606 trans = btrfs_join_transaction(rc->extent_root);
4607 if (IS_ERR(trans)) {
4608 err = PTR_ERR(trans);
4609 goto out_free;
4610 }
4611 err = qgroup_fix_relocated_data_extents(trans, rc);
4612 if (err < 0) {
4613 btrfs_abort_transaction(trans, err);
4614 goto out_free;
4615 }
4616 err = btrfs_commit_transaction(trans, rc->extent_root);
4617 out_free:
4618 kfree(rc);
4619 out:
4620 if (!list_empty(&reloc_roots))
4621 free_reloc_roots(&reloc_roots);
4622
4623 btrfs_free_path(path);
4624
4625 if (err == 0) {
4626 /* cleanup orphan inode in data relocation tree */
4627 fs_root = read_fs_root(root->fs_info,
4628 BTRFS_DATA_RELOC_TREE_OBJECTID);
4629 if (IS_ERR(fs_root))
4630 err = PTR_ERR(fs_root);
4631 else
4632 err = btrfs_orphan_cleanup(fs_root);
4633 }
4634 return err;
4635 }
4636
4637 /*
4638 * helper to add ordered checksum for data relocation.
4639 *
4640 * cloning checksum properly handles the nodatasum extents.
4641 * it also saves CPU time to re-calculate the checksum.
4642 */
4643 int btrfs_reloc_clone_csums(struct inode *inode, u64 file_pos, u64 len)
4644 {
4645 struct btrfs_ordered_sum *sums;
4646 struct btrfs_ordered_extent *ordered;
4647 struct btrfs_root *root = BTRFS_I(inode)->root;
4648 int ret;
4649 u64 disk_bytenr;
4650 u64 new_bytenr;
4651 LIST_HEAD(list);
4652
4653 ordered = btrfs_lookup_ordered_extent(inode, file_pos);
4654 BUG_ON(ordered->file_offset != file_pos || ordered->len != len);
4655
4656 disk_bytenr = file_pos + BTRFS_I(inode)->index_cnt;
4657 ret = btrfs_lookup_csums_range(root->fs_info->csum_root, disk_bytenr,
4658 disk_bytenr + len - 1, &list, 0);
4659 if (ret)
4660 goto out;
4661
4662 while (!list_empty(&list)) {
4663 sums = list_entry(list.next, struct btrfs_ordered_sum, list);
4664 list_del_init(&sums->list);
4665
4666 /*
4667 * We need to offset the new_bytenr based on where the csum is.
4668 * We need to do this because we will read in entire prealloc
4669 * extents but we may have written to say the middle of the
4670 * prealloc extent, so we need to make sure the csum goes with
4671 * the right disk offset.
4672 *
4673 * We can do this because the data reloc inode refers strictly
4674 * to the on disk bytes, so we don't have to worry about
4675 * disk_len vs real len like with real inodes since it's all
4676 * disk length.
4677 */
4678 new_bytenr = ordered->start + (sums->bytenr - disk_bytenr);
4679 sums->bytenr = new_bytenr;
4680
4681 btrfs_add_ordered_sum(inode, ordered, sums);
4682 }
4683 out:
4684 btrfs_put_ordered_extent(ordered);
4685 return ret;
4686 }
4687
4688 int btrfs_reloc_cow_block(struct btrfs_trans_handle *trans,
4689 struct btrfs_root *root, struct extent_buffer *buf,
4690 struct extent_buffer *cow)
4691 {
4692 struct reloc_control *rc;
4693 struct backref_node *node;
4694 int first_cow = 0;
4695 int level;
4696 int ret = 0;
4697
4698 rc = root->fs_info->reloc_ctl;
4699 if (!rc)
4700 return 0;
4701
4702 BUG_ON(rc->stage == UPDATE_DATA_PTRS &&
4703 root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID);
4704
4705 if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID) {
4706 if (buf == root->node)
4707 __update_reloc_root(root, cow->start);
4708 }
4709
4710 level = btrfs_header_level(buf);
4711 if (btrfs_header_generation(buf) <=
4712 btrfs_root_last_snapshot(&root->root_item))
4713 first_cow = 1;
4714
4715 if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID &&
4716 rc->create_reloc_tree) {
4717 WARN_ON(!first_cow && level == 0);
4718
4719 node = rc->backref_cache.path[level];
4720 BUG_ON(node->bytenr != buf->start &&
4721 node->new_bytenr != buf->start);
4722
4723 drop_node_buffer(node);
4724 extent_buffer_get(cow);
4725 node->eb = cow;
4726 node->new_bytenr = cow->start;
4727
4728 if (!node->pending) {
4729 list_move_tail(&node->list,
4730 &rc->backref_cache.pending[level]);
4731 node->pending = 1;
4732 }
4733
4734 if (first_cow)
4735 __mark_block_processed(rc, node);
4736
4737 if (first_cow && level > 0)
4738 rc->nodes_relocated += buf->len;
4739 }
4740
4741 if (level == 0 && first_cow && rc->stage == UPDATE_DATA_PTRS)
4742 ret = replace_file_extents(trans, rc, root, cow);
4743 return ret;
4744 }
4745
4746 /*
4747 * called before creating snapshot. it calculates metadata reservation
4748 * required for relocating tree blocks in the snapshot
4749 */
4750 void btrfs_reloc_pre_snapshot(struct btrfs_pending_snapshot *pending,
4751 u64 *bytes_to_reserve)
4752 {
4753 struct btrfs_root *root;
4754 struct reloc_control *rc;
4755
4756 root = pending->root;
4757 if (!root->reloc_root)
4758 return;
4759
4760 rc = root->fs_info->reloc_ctl;
4761 if (!rc->merge_reloc_tree)
4762 return;
4763
4764 root = root->reloc_root;
4765 BUG_ON(btrfs_root_refs(&root->root_item) == 0);
4766 /*
4767 * relocation is in the stage of merging trees. the space
4768 * used by merging a reloc tree is twice the size of
4769 * relocated tree nodes in the worst case. half for cowing
4770 * the reloc tree, half for cowing the fs tree. the space
4771 * used by cowing the reloc tree will be freed after the
4772 * tree is dropped. if we create snapshot, cowing the fs
4773 * tree may use more space than it frees. so we need
4774 * reserve extra space.
4775 */
4776 *bytes_to_reserve += rc->nodes_relocated;
4777 }
4778
4779 /*
4780 * called after snapshot is created. migrate block reservation
4781 * and create reloc root for the newly created snapshot
4782 */
4783 int btrfs_reloc_post_snapshot(struct btrfs_trans_handle *trans,
4784 struct btrfs_pending_snapshot *pending)
4785 {
4786 struct btrfs_root *root = pending->root;
4787 struct btrfs_root *reloc_root;
4788 struct btrfs_root *new_root;
4789 struct reloc_control *rc;
4790 int ret;
4791
4792 if (!root->reloc_root)
4793 return 0;
4794
4795 rc = root->fs_info->reloc_ctl;
4796 rc->merging_rsv_size += rc->nodes_relocated;
4797
4798 if (rc->merge_reloc_tree) {
4799 ret = btrfs_block_rsv_migrate(&pending->block_rsv,
4800 rc->block_rsv,
4801 rc->nodes_relocated, 1);
4802 if (ret)
4803 return ret;
4804 }
4805
4806 new_root = pending->snap;
4807 reloc_root = create_reloc_root(trans, root->reloc_root,
4808 new_root->root_key.objectid);
4809 if (IS_ERR(reloc_root))
4810 return PTR_ERR(reloc_root);
4811
4812 ret = __add_reloc_root(reloc_root);
4813 BUG_ON(ret < 0);
4814 new_root->reloc_root = reloc_root;
4815
4816 if (rc->create_reloc_tree)
4817 ret = clone_backref_node(trans, rc, root, reloc_root);
4818 return ret;
4819 }
Linux with added FPC-III support