?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
HID: hiddev: Fix slab-out-of-bounds write in hiddev_ioctl_usage()
[linux/fpc-iii.git] / fs / btrfs / relocation.c
blobdf04309390bba364cf1ba11bdaa752e4c5bdccfa
1 /*
2 * Copyright (C) 2009 Oracle. All rights reserved.
3 *
4 * This program is free software; you can redistribute it and/or
5 * modify it under the terms of the GNU General Public
6 * License v2 as published by the Free Software Foundation.
7 *
8 * This program is distributed in the hope that it will be useful,
9 * but WITHOUT ANY WARRANTY; without even the implied warranty of
10 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
11 * General Public License for more details.
12 *
13 * You should have received a copy of the GNU General Public
14 * License along with this program; if not, write to the
15 * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
16 * Boston, MA 021110-1307, USA.
17 */
18
19 #include <linux/sched.h>
20 #include <linux/pagemap.h>
21 #include <linux/writeback.h>
22 #include <linux/blkdev.h>
23 #include <linux/rbtree.h>
24 #include <linux/slab.h>
25 #include "ctree.h"
26 #include "disk-io.h"
27 #include "transaction.h"
28 #include "volumes.h"
29 #include "locking.h"
30 #include "btrfs_inode.h"
31 #include "async-thread.h"
32 #include "free-space-cache.h"
33 #include "inode-map.h"
34
35 /*
36 * backref_node, mapping_node and tree_block start with this
37 */
38 struct tree_entry {
39 struct rb_node rb_node;
40 u64 bytenr;
41 };
42
43 /*
44 * present a tree block in the backref cache
45 */
46 struct backref_node {
47 struct rb_node rb_node;
48 u64 bytenr;
49
50 u64 new_bytenr;
51 /* objectid of tree block owner, can be not uptodate */
52 u64 owner;
53 /* link to pending, changed or detached list */
54 struct list_head list;
55 /* list of upper level blocks reference this block */
56 struct list_head upper;
57 /* list of child blocks in the cache */
58 struct list_head lower;
59 /* NULL if this node is not tree root */
60 struct btrfs_root *root;
61 /* extent buffer got by COW the block */
62 struct extent_buffer *eb;
63 /* level of tree block */
64 unsigned int level:8;
65 /* is the block in non-reference counted tree */
66 unsigned int cowonly:1;
67 /* 1 if no child node in the cache */
68 unsigned int lowest:1;
69 /* is the extent buffer locked */
70 unsigned int locked:1;
71 /* has the block been processed */
72 unsigned int processed:1;
73 /* have backrefs of this block been checked */
74 unsigned int checked:1;
75 /*
76 * 1 if corresponding block has been cowed but some upper
77 * level block pointers may not point to the new location
78 */
79 unsigned int pending:1;
80 /*
81 * 1 if the backref node isn't connected to any other
82 * backref node.
83 */
84 unsigned int detached:1;
85 };
86
87 /*
88 * present a block pointer in the backref cache
89 */
90 struct backref_edge {
91 struct list_head list[2];
92 struct backref_node *node[2];
93 };
94
95 #define LOWER 0
96 #define UPPER 1
97 #define RELOCATION_RESERVED_NODES 256
98
99 struct backref_cache {
100 /* red black tree of all backref nodes in the cache */
101 struct rb_root rb_root;
102 /* for passing backref nodes to btrfs_reloc_cow_block */
103 struct backref_node *path[BTRFS_MAX_LEVEL];
104 /*
105 * list of blocks that have been cowed but some block
106 * pointers in upper level blocks may not reflect the
107 * new location
108 */
109 struct list_head pending[BTRFS_MAX_LEVEL];
110 /* list of backref nodes with no child node */
111 struct list_head leaves;
112 /* list of blocks that have been cowed in current transaction */
113 struct list_head changed;
114 /* list of detached backref node. */
115 struct list_head detached;
116
117 u64 last_trans;
118
119 int nr_nodes;
120 int nr_edges;
121 };
122
123 /*
124 * map address of tree root to tree
125 */
126 struct mapping_node {
127 struct rb_node rb_node;
128 u64 bytenr;
129 void *data;
130 };
131
132 struct mapping_tree {
133 struct rb_root rb_root;
134 spinlock_t lock;
135 };
136
137 /*
138 * present a tree block to process
139 */
140 struct tree_block {
141 struct rb_node rb_node;
142 u64 bytenr;
143 struct btrfs_key key;
144 unsigned int level:8;
145 unsigned int key_ready:1;
146 };
147
148 #define MAX_EXTENTS 128
149
150 struct file_extent_cluster {
151 u64 start;
152 u64 end;
153 u64 boundary[MAX_EXTENTS];
154 unsigned int nr;
155 };
156
157 struct reloc_control {
158 /* block group to relocate */
159 struct btrfs_block_group_cache *block_group;
160 /* extent tree */
161 struct btrfs_root *extent_root;
162 /* inode for moving data */
163 struct inode *data_inode;
164
165 struct btrfs_block_rsv *block_rsv;
166
167 struct backref_cache backref_cache;
168
169 struct file_extent_cluster cluster;
170 /* tree blocks have been processed */
171 struct extent_io_tree processed_blocks;
172 /* map start of tree root to corresponding reloc tree */
173 struct mapping_tree reloc_root_tree;
174 /* list of reloc trees */
175 struct list_head reloc_roots;
176 /* size of metadata reservation for merging reloc trees */
177 u64 merging_rsv_size;
178 /* size of relocated tree nodes */
179 u64 nodes_relocated;
180 /* reserved size for block group relocation*/
181 u64 reserved_bytes;
182
183 u64 search_start;
184 u64 extents_found;
185
186 unsigned int stage:8;
187 unsigned int create_reloc_tree:1;
188 unsigned int merge_reloc_tree:1;
189 unsigned int found_file_extent:1;
190 };
191
192 /* stages of data relocation */
193 #define MOVE_DATA_EXTENTS 0
194 #define UPDATE_DATA_PTRS 1
195
196 static void remove_backref_node(struct backref_cache *cache,
197 struct backref_node *node);
198 static void __mark_block_processed(struct reloc_control *rc,
199 struct backref_node *node);
200
201 static void mapping_tree_init(struct mapping_tree *tree)
202 {
203 tree->rb_root = RB_ROOT;
204 spin_lock_init(&tree->lock);
205 }
206
207 static void backref_cache_init(struct backref_cache *cache)
208 {
209 int i;
210 cache->rb_root = RB_ROOT;
211 for (i = 0; i < BTRFS_MAX_LEVEL; i++)
212 INIT_LIST_HEAD(&cache->pending[i]);
213 INIT_LIST_HEAD(&cache->changed);
214 INIT_LIST_HEAD(&cache->detached);
215 INIT_LIST_HEAD(&cache->leaves);
216 }
217
218 static void backref_cache_cleanup(struct backref_cache *cache)
219 {
220 struct backref_node *node;
221 int i;
222
223 while (!list_empty(&cache->detached)) {
224 node = list_entry(cache->detached.next,
225 struct backref_node, list);
226 remove_backref_node(cache, node);
227 }
228
229 while (!list_empty(&cache->leaves)) {
230 node = list_entry(cache->leaves.next,
231 struct backref_node, lower);
232 remove_backref_node(cache, node);
233 }
234
235 cache->last_trans = 0;
236
237 for (i = 0; i < BTRFS_MAX_LEVEL; i++)
238 BUG_ON(!list_empty(&cache->pending[i]));
239 BUG_ON(!list_empty(&cache->changed));
240 BUG_ON(!list_empty(&cache->detached));
241 BUG_ON(!RB_EMPTY_ROOT(&cache->rb_root));
242 BUG_ON(cache->nr_nodes);
243 BUG_ON(cache->nr_edges);
244 }
245
246 static struct backref_node *alloc_backref_node(struct backref_cache *cache)
247 {
248 struct backref_node *node;
249
250 node = kzalloc(sizeof(*node), GFP_NOFS);
251 if (node) {
252 INIT_LIST_HEAD(&node->list);
253 INIT_LIST_HEAD(&node->upper);
254 INIT_LIST_HEAD(&node->lower);
255 RB_CLEAR_NODE(&node->rb_node);
256 cache->nr_nodes++;
257 }
258 return node;
259 }
260
261 static void free_backref_node(struct backref_cache *cache,
262 struct backref_node *node)
263 {
264 if (node) {
265 cache->nr_nodes--;
266 kfree(node);
267 }
268 }
269
270 static struct backref_edge *alloc_backref_edge(struct backref_cache *cache)
271 {
272 struct backref_edge *edge;
273
274 edge = kzalloc(sizeof(*edge), GFP_NOFS);
275 if (edge)
276 cache->nr_edges++;
277 return edge;
278 }
279
280 static void free_backref_edge(struct backref_cache *cache,
281 struct backref_edge *edge)
282 {
283 if (edge) {
284 cache->nr_edges--;
285 kfree(edge);
286 }
287 }
288
289 static struct rb_node *tree_insert(struct rb_root *root, u64 bytenr,
290 struct rb_node *node)
291 {
292 struct rb_node **p = &root->rb_node;
293 struct rb_node *parent = NULL;
294 struct tree_entry *entry;
295
296 while (*p) {
297 parent = *p;
298 entry = rb_entry(parent, struct tree_entry, rb_node);
299
300 if (bytenr < entry->bytenr)
301 p = &(*p)->rb_left;
302 else if (bytenr > entry->bytenr)
303 p = &(*p)->rb_right;
304 else
305 return parent;
306 }
307
308 rb_link_node(node, parent, p);
309 rb_insert_color(node, root);
310 return NULL;
311 }
312
313 static struct rb_node *tree_search(struct rb_root *root, u64 bytenr)
314 {
315 struct rb_node *n = root->rb_node;
316 struct tree_entry *entry;
317
318 while (n) {
319 entry = rb_entry(n, struct tree_entry, rb_node);
320
321 if (bytenr < entry->bytenr)
322 n = n->rb_left;
323 else if (bytenr > entry->bytenr)
324 n = n->rb_right;
325 else
326 return n;
327 }
328 return NULL;
329 }
330
331 static void backref_tree_panic(struct rb_node *rb_node, int errno, u64 bytenr)
332 {
333
334 struct btrfs_fs_info *fs_info = NULL;
335 struct backref_node *bnode = rb_entry(rb_node, struct backref_node,
336 rb_node);
337 if (bnode->root)
338 fs_info = bnode->root->fs_info;
339 btrfs_panic(fs_info, errno, "Inconsistency in backref cache "
340 "found at offset %llu", bytenr);
341 }
342
343 /*
344 * walk up backref nodes until reach node presents tree root
345 */
346 static struct backref_node *walk_up_backref(struct backref_node *node,
347 struct backref_edge *edges[],
348 int *index)
349 {
350 struct backref_edge *edge;
351 int idx = *index;
352
353 while (!list_empty(&node->upper)) {
354 edge = list_entry(node->upper.next,
355 struct backref_edge, list[LOWER]);
356 edges[idx++] = edge;
357 node = edge->node[UPPER];
358 }
359 BUG_ON(node->detached);
360 *index = idx;
361 return node;
362 }
363
364 /*
365 * walk down backref nodes to find start of next reference path
366 */
367 static struct backref_node *walk_down_backref(struct backref_edge *edges[],
368 int *index)
369 {
370 struct backref_edge *edge;
371 struct backref_node *lower;
372 int idx = *index;
373
374 while (idx > 0) {
375 edge = edges[idx - 1];
376 lower = edge->node[LOWER];
377 if (list_is_last(&edge->list[LOWER], &lower->upper)) {
378 idx--;
379 continue;
380 }
381 edge = list_entry(edge->list[LOWER].next,
382 struct backref_edge, list[LOWER]);
383 edges[idx - 1] = edge;
384 *index = idx;
385 return edge->node[UPPER];
386 }
387 *index = 0;
388 return NULL;
389 }
390
391 static void unlock_node_buffer(struct backref_node *node)
392 {
393 if (node->locked) {
394 btrfs_tree_unlock(node->eb);
395 node->locked = 0;
396 }
397 }
398
399 static void drop_node_buffer(struct backref_node *node)
400 {
401 if (node->eb) {
402 unlock_node_buffer(node);
403 free_extent_buffer(node->eb);
404 node->eb = NULL;
405 }
406 }
407
408 static void drop_backref_node(struct backref_cache *tree,
409 struct backref_node *node)
410 {
411 BUG_ON(!list_empty(&node->upper));
412
413 drop_node_buffer(node);
414 list_del(&node->list);
415 list_del(&node->lower);
416 if (!RB_EMPTY_NODE(&node->rb_node))
417 rb_erase(&node->rb_node, &tree->rb_root);
418 free_backref_node(tree, node);
419 }
420
421 /*
422 * remove a backref node from the backref cache
423 */
424 static void remove_backref_node(struct backref_cache *cache,
425 struct backref_node *node)
426 {
427 struct backref_node *upper;
428 struct backref_edge *edge;
429
430 if (!node)
431 return;
432
433 BUG_ON(!node->lowest && !node->detached);
434 while (!list_empty(&node->upper)) {
435 edge = list_entry(node->upper.next, struct backref_edge,
436 list[LOWER]);
437 upper = edge->node[UPPER];
438 list_del(&edge->list[LOWER]);
439 list_del(&edge->list[UPPER]);
440 free_backref_edge(cache, edge);
441
442 if (RB_EMPTY_NODE(&upper->rb_node)) {
443 BUG_ON(!list_empty(&node->upper));
444 drop_backref_node(cache, node);
445 node = upper;
446 node->lowest = 1;
447 continue;
448 }
449 /*
450 * add the node to leaf node list if no other
451 * child block cached.
452 */
453 if (list_empty(&upper->lower)) {
454 list_add_tail(&upper->lower, &cache->leaves);
455 upper->lowest = 1;
456 }
457 }
458
459 drop_backref_node(cache, node);
460 }
461
462 static void update_backref_node(struct backref_cache *cache,
463 struct backref_node *node, u64 bytenr)
464 {
465 struct rb_node *rb_node;
466 rb_erase(&node->rb_node, &cache->rb_root);
467 node->bytenr = bytenr;
468 rb_node = tree_insert(&cache->rb_root, node->bytenr, &node->rb_node);
469 if (rb_node)
470 backref_tree_panic(rb_node, -EEXIST, bytenr);
471 }
472
473 /*
474 * update backref cache after a transaction commit
475 */
476 static int update_backref_cache(struct btrfs_trans_handle *trans,
477 struct backref_cache *cache)
478 {
479 struct backref_node *node;
480 int level = 0;
481
482 if (cache->last_trans == 0) {
483 cache->last_trans = trans->transid;
484 return 0;
485 }
486
487 if (cache->last_trans == trans->transid)
488 return 0;
489
490 /*
491 * detached nodes are used to avoid unnecessary backref
492 * lookup. transaction commit changes the extent tree.
493 * so the detached nodes are no longer useful.
494 */
495 while (!list_empty(&cache->detached)) {
496 node = list_entry(cache->detached.next,
497 struct backref_node, list);
498 remove_backref_node(cache, node);
499 }
500
501 while (!list_empty(&cache->changed)) {
502 node = list_entry(cache->changed.next,
503 struct backref_node, list);
504 list_del_init(&node->list);
505 BUG_ON(node->pending);
506 update_backref_node(cache, node, node->new_bytenr);
507 }
508
509 /*
510 * some nodes can be left in the pending list if there were
511 * errors during processing the pending nodes.
512 */
513 for (level = 0; level < BTRFS_MAX_LEVEL; level++) {
514 list_for_each_entry(node, &cache->pending[level], list) {
515 BUG_ON(!node->pending);
516 if (node->bytenr == node->new_bytenr)
517 continue;
518 update_backref_node(cache, node, node->new_bytenr);
519 }
520 }
521
522 cache->last_trans = 0;
523 return 1;
524 }
525
526
527 static int should_ignore_root(struct btrfs_root *root)
528 {
529 struct btrfs_root *reloc_root;
530
531 if (!test_bit(BTRFS_ROOT_REF_COWS, &root->state))
532 return 0;
533
534 reloc_root = root->reloc_root;
535 if (!reloc_root)
536 return 0;
537
538 if (btrfs_root_last_snapshot(&reloc_root->root_item) ==
539 root->fs_info->running_transaction->transid - 1)
540 return 0;
541 /*
542 * if there is reloc tree and it was created in previous
543 * transaction backref lookup can find the reloc tree,
544 * so backref node for the fs tree root is useless for
545 * relocation.
546 */
547 return 1;
548 }
549 /*
550 * find reloc tree by address of tree root
551 */
552 static struct btrfs_root *find_reloc_root(struct reloc_control *rc,
553 u64 bytenr)
554 {
555 struct rb_node *rb_node;
556 struct mapping_node *node;
557 struct btrfs_root *root = NULL;
558
559 spin_lock(&rc->reloc_root_tree.lock);
560 rb_node = tree_search(&rc->reloc_root_tree.rb_root, bytenr);
561 if (rb_node) {
562 node = rb_entry(rb_node, struct mapping_node, rb_node);
563 root = (struct btrfs_root *)node->data;
564 }
565 spin_unlock(&rc->reloc_root_tree.lock);
566 return root;
567 }
568
569 static int is_cowonly_root(u64 root_objectid)
570 {
571 if (root_objectid == BTRFS_ROOT_TREE_OBJECTID ||
572 root_objectid == BTRFS_EXTENT_TREE_OBJECTID ||
573 root_objectid == BTRFS_CHUNK_TREE_OBJECTID ||
574 root_objectid == BTRFS_DEV_TREE_OBJECTID ||
575 root_objectid == BTRFS_TREE_LOG_OBJECTID ||
576 root_objectid == BTRFS_CSUM_TREE_OBJECTID ||
577 root_objectid == BTRFS_UUID_TREE_OBJECTID ||
578 root_objectid == BTRFS_QUOTA_TREE_OBJECTID)
579 return 1;
580 return 0;
581 }
582
583 static struct btrfs_root *read_fs_root(struct btrfs_fs_info *fs_info,
584 u64 root_objectid)
585 {
586 struct btrfs_key key;
587
588 key.objectid = root_objectid;
589 key.type = BTRFS_ROOT_ITEM_KEY;
590 if (is_cowonly_root(root_objectid))
591 key.offset = 0;
592 else
593 key.offset = (u64)-1;
594
595 return btrfs_get_fs_root(fs_info, &key, false);
596 }
597
598 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
599 static noinline_for_stack
600 struct btrfs_root *find_tree_root(struct reloc_control *rc,
601 struct extent_buffer *leaf,
602 struct btrfs_extent_ref_v0 *ref0)
603 {
604 struct btrfs_root *root;
605 u64 root_objectid = btrfs_ref_root_v0(leaf, ref0);
606 u64 generation = btrfs_ref_generation_v0(leaf, ref0);
607
608 BUG_ON(root_objectid == BTRFS_TREE_RELOC_OBJECTID);
609
610 root = read_fs_root(rc->extent_root->fs_info, root_objectid);
611 BUG_ON(IS_ERR(root));
612
613 if (test_bit(BTRFS_ROOT_REF_COWS, &root->state) &&
614 generation != btrfs_root_generation(&root->root_item))
615 return NULL;
616
617 return root;
618 }
619 #endif
620
621 static noinline_for_stack
622 int find_inline_backref(struct extent_buffer *leaf, int slot,
623 unsigned long *ptr, unsigned long *end)
624 {
625 struct btrfs_key key;
626 struct btrfs_extent_item *ei;
627 struct btrfs_tree_block_info *bi;
628 u32 item_size;
629
630 btrfs_item_key_to_cpu(leaf, &key, slot);
631
632 item_size = btrfs_item_size_nr(leaf, slot);
633 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
634 if (item_size < sizeof(*ei)) {
635 WARN_ON(item_size != sizeof(struct btrfs_extent_item_v0));
636 return 1;
637 }
638 #endif
639 ei = btrfs_item_ptr(leaf, slot, struct btrfs_extent_item);
640 WARN_ON(!(btrfs_extent_flags(leaf, ei) &
641 BTRFS_EXTENT_FLAG_TREE_BLOCK));
642
643 if (key.type == BTRFS_EXTENT_ITEM_KEY &&
644 item_size <= sizeof(*ei) + sizeof(*bi)) {
645 WARN_ON(item_size < sizeof(*ei) + sizeof(*bi));
646 return 1;
647 }
648 if (key.type == BTRFS_METADATA_ITEM_KEY &&
649 item_size <= sizeof(*ei)) {
650 WARN_ON(item_size < sizeof(*ei));
651 return 1;
652 }
653
654 if (key.type == BTRFS_EXTENT_ITEM_KEY) {
655 bi = (struct btrfs_tree_block_info *)(ei + 1);
656 *ptr = (unsigned long)(bi + 1);
657 } else {
658 *ptr = (unsigned long)(ei + 1);
659 }
660 *end = (unsigned long)ei + item_size;
661 return 0;
662 }
663
664 /*
665 * build backref tree for a given tree block. root of the backref tree
666 * corresponds the tree block, leaves of the backref tree correspond
667 * roots of b-trees that reference the tree block.
668 *
669 * the basic idea of this function is check backrefs of a given block
670 * to find upper level blocks that refernece the block, and then check
671 * bakcrefs of these upper level blocks recursively. the recursion stop
672 * when tree root is reached or backrefs for the block is cached.
673 *
674 * NOTE: if we find backrefs for a block are cached, we know backrefs
675 * for all upper level blocks that directly/indirectly reference the
676 * block are also cached.
677 */
678 static noinline_for_stack
679 struct backref_node *build_backref_tree(struct reloc_control *rc,
680 struct btrfs_key *node_key,
681 int level, u64 bytenr)
682 {
683 struct backref_cache *cache = &rc->backref_cache;
684 struct btrfs_path *path1;
685 struct btrfs_path *path2;
686 struct extent_buffer *eb;
687 struct btrfs_root *root;
688 struct backref_node *cur;
689 struct backref_node *upper;
690 struct backref_node *lower;
691 struct backref_node *node = NULL;
692 struct backref_node *exist = NULL;
693 struct backref_edge *edge;
694 struct rb_node *rb_node;
695 struct btrfs_key key;
696 unsigned long end;
697 unsigned long ptr;
698 LIST_HEAD(list);
699 LIST_HEAD(useless);
700 int cowonly;
701 int ret;
702 int err = 0;
703 bool need_check = true;
704
705 path1 = btrfs_alloc_path();
706 path2 = btrfs_alloc_path();
707 if (!path1 || !path2) {
708 err = -ENOMEM;
709 goto out;
710 }
711 path1->reada = 1;
712 path2->reada = 2;
713
714 node = alloc_backref_node(cache);
715 if (!node) {
716 err = -ENOMEM;
717 goto out;
718 }
719
720 node->bytenr = bytenr;
721 node->level = level;
722 node->lowest = 1;
723 cur = node;
724 again:
725 end = 0;
726 ptr = 0;
727 key.objectid = cur->bytenr;
728 key.type = BTRFS_METADATA_ITEM_KEY;
729 key.offset = (u64)-1;
730
731 path1->search_commit_root = 1;
732 path1->skip_locking = 1;
733 ret = btrfs_search_slot(NULL, rc->extent_root, &key, path1,
734 0, 0);
735 if (ret < 0) {
736 err = ret;
737 goto out;
738 }
739 ASSERT(ret);
740 ASSERT(path1->slots[0]);
741
742 path1->slots[0]--;
743
744 WARN_ON(cur->checked);
745 if (!list_empty(&cur->upper)) {
746 /*
747 * the backref was added previously when processing
748 * backref of type BTRFS_TREE_BLOCK_REF_KEY
749 */
750 ASSERT(list_is_singular(&cur->upper));
751 edge = list_entry(cur->upper.next, struct backref_edge,
752 list[LOWER]);
753 ASSERT(list_empty(&edge->list[UPPER]));
754 exist = edge->node[UPPER];
755 /*
756 * add the upper level block to pending list if we need
757 * check its backrefs
758 */
759 if (!exist->checked)
760 list_add_tail(&edge->list[UPPER], &list);
761 } else {
762 exist = NULL;
763 }
764
765 while (1) {
766 cond_resched();
767 eb = path1->nodes[0];
768
769 if (ptr >= end) {
770 if (path1->slots[0] >= btrfs_header_nritems(eb)) {
771 ret = btrfs_next_leaf(rc->extent_root, path1);
772 if (ret < 0) {
773 err = ret;
774 goto out;
775 }
776 if (ret > 0)
777 break;
778 eb = path1->nodes[0];
779 }
780
781 btrfs_item_key_to_cpu(eb, &key, path1->slots[0]);
782 if (key.objectid != cur->bytenr) {
783 WARN_ON(exist);
784 break;
785 }
786
787 if (key.type == BTRFS_EXTENT_ITEM_KEY ||
788 key.type == BTRFS_METADATA_ITEM_KEY) {
789 ret = find_inline_backref(eb, path1->slots[0],
790 &ptr, &end);
791 if (ret)
792 goto next;
793 }
794 }
795
796 if (ptr < end) {
797 /* update key for inline back ref */
798 struct btrfs_extent_inline_ref *iref;
799 iref = (struct btrfs_extent_inline_ref *)ptr;
800 key.type = btrfs_extent_inline_ref_type(eb, iref);
801 key.offset = btrfs_extent_inline_ref_offset(eb, iref);
802 WARN_ON(key.type != BTRFS_TREE_BLOCK_REF_KEY &&
803 key.type != BTRFS_SHARED_BLOCK_REF_KEY);
804 }
805
806 if (exist &&
807 ((key.type == BTRFS_TREE_BLOCK_REF_KEY &&
808 exist->owner == key.offset) ||
809 (key.type == BTRFS_SHARED_BLOCK_REF_KEY &&
810 exist->bytenr == key.offset))) {
811 exist = NULL;
812 goto next;
813 }
814
815 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
816 if (key.type == BTRFS_SHARED_BLOCK_REF_KEY ||
817 key.type == BTRFS_EXTENT_REF_V0_KEY) {
818 if (key.type == BTRFS_EXTENT_REF_V0_KEY) {
819 struct btrfs_extent_ref_v0 *ref0;
820 ref0 = btrfs_item_ptr(eb, path1->slots[0],
821 struct btrfs_extent_ref_v0);
822 if (key.objectid == key.offset) {
823 root = find_tree_root(rc, eb, ref0);
824 if (root && !should_ignore_root(root))
825 cur->root = root;
826 else
827 list_add(&cur->list, &useless);
828 break;
829 }
830 if (is_cowonly_root(btrfs_ref_root_v0(eb,
831 ref0)))
832 cur->cowonly = 1;
833 }
834 #else
835 ASSERT(key.type != BTRFS_EXTENT_REF_V0_KEY);
836 if (key.type == BTRFS_SHARED_BLOCK_REF_KEY) {
837 #endif
838 if (key.objectid == key.offset) {
839 /*
840 * only root blocks of reloc trees use
841 * backref of this type.
842 */
843 root = find_reloc_root(rc, cur->bytenr);
844 ASSERT(root);
845 cur->root = root;
846 break;
847 }
848
849 edge = alloc_backref_edge(cache);
850 if (!edge) {
851 err = -ENOMEM;
852 goto out;
853 }
854 rb_node = tree_search(&cache->rb_root, key.offset);
855 if (!rb_node) {
856 upper = alloc_backref_node(cache);
857 if (!upper) {
858 free_backref_edge(cache, edge);
859 err = -ENOMEM;
860 goto out;
861 }
862 upper->bytenr = key.offset;
863 upper->level = cur->level + 1;
864 /*
865 * backrefs for the upper level block isn't
866 * cached, add the block to pending list
867 */
868 list_add_tail(&edge->list[UPPER], &list);
869 } else {
870 upper = rb_entry(rb_node, struct backref_node,
871 rb_node);
872 ASSERT(upper->checked);
873 INIT_LIST_HEAD(&edge->list[UPPER]);
874 }
875 list_add_tail(&edge->list[LOWER], &cur->upper);
876 edge->node[LOWER] = cur;
877 edge->node[UPPER] = upper;
878
879 goto next;
880 } else if (key.type != BTRFS_TREE_BLOCK_REF_KEY) {
881 goto next;
882 }
883
884 /* key.type == BTRFS_TREE_BLOCK_REF_KEY */
885 root = read_fs_root(rc->extent_root->fs_info, key.offset);
886 if (IS_ERR(root)) {
887 err = PTR_ERR(root);
888 goto out;
889 }
890
891 if (!test_bit(BTRFS_ROOT_REF_COWS, &root->state))
892 cur->cowonly = 1;
893
894 if (btrfs_root_level(&root->root_item) == cur->level) {
895 /* tree root */
896 ASSERT(btrfs_root_bytenr(&root->root_item) ==
897 cur->bytenr);
898 if (should_ignore_root(root))
899 list_add(&cur->list, &useless);
900 else
901 cur->root = root;
902 break;
903 }
904
905 level = cur->level + 1;
906
907 /*
908 * searching the tree to find upper level blocks
909 * reference the block.
910 */
911 path2->search_commit_root = 1;
912 path2->skip_locking = 1;
913 path2->lowest_level = level;
914 ret = btrfs_search_slot(NULL, root, node_key, path2, 0, 0);
915 path2->lowest_level = 0;
916 if (ret < 0) {
917 err = ret;
918 goto out;
919 }
920 if (ret > 0 && path2->slots[level] > 0)
921 path2->slots[level]--;
922
923 eb = path2->nodes[level];
924 if (btrfs_node_blockptr(eb, path2->slots[level]) !=
925 cur->bytenr) {
926 btrfs_err(root->fs_info,
927 "couldn't find block (%llu) (level %d) in tree (%llu) with key (%llu %u %llu)",
928 cur->bytenr, level - 1, root->objectid,
929 node_key->objectid, node_key->type,
930 node_key->offset);
931 err = -ENOENT;
932 goto out;
933 }
934 lower = cur;
935 need_check = true;
936 for (; level < BTRFS_MAX_LEVEL; level++) {
937 if (!path2->nodes[level]) {
938 ASSERT(btrfs_root_bytenr(&root->root_item) ==
939 lower->bytenr);
940 if (should_ignore_root(root))
941 list_add(&lower->list, &useless);
942 else
943 lower->root = root;
944 break;
945 }
946
947 edge = alloc_backref_edge(cache);
948 if (!edge) {
949 err = -ENOMEM;
950 goto out;
951 }
952
953 eb = path2->nodes[level];
954 rb_node = tree_search(&cache->rb_root, eb->start);
955 if (!rb_node) {
956 upper = alloc_backref_node(cache);
957 if (!upper) {
958 free_backref_edge(cache, edge);
959 err = -ENOMEM;
960 goto out;
961 }
962 upper->bytenr = eb->start;
963 upper->owner = btrfs_header_owner(eb);
964 upper->level = lower->level + 1;
965 if (!test_bit(BTRFS_ROOT_REF_COWS,
966 &root->state))
967 upper->cowonly = 1;
968
969 /*
970 * if we know the block isn't shared
971 * we can void checking its backrefs.
972 */
973 if (btrfs_block_can_be_shared(root, eb))
974 upper->checked = 0;
975 else
976 upper->checked = 1;
977
978 /*
979 * add the block to pending list if we
980 * need check its backrefs, we only do this once
981 * while walking up a tree as we will catch
982 * anything else later on.
983 */
984 if (!upper->checked && need_check) {
985 need_check = false;
986 list_add_tail(&edge->list[UPPER],
987 &list);
988 } else {
989 if (upper->checked)
990 need_check = true;
991 INIT_LIST_HEAD(&edge->list[UPPER]);
992 }
993 } else {
994 upper = rb_entry(rb_node, struct backref_node,
995 rb_node);
996 ASSERT(upper->checked);
997 INIT_LIST_HEAD(&edge->list[UPPER]);
998 if (!upper->owner)
999 upper->owner = btrfs_header_owner(eb);
1000 }
1001 list_add_tail(&edge->list[LOWER], &lower->upper);
1002 edge->node[LOWER] = lower;
1003 edge->node[UPPER] = upper;
1004
1005 if (rb_node)
1006 break;
1007 lower = upper;
1008 upper = NULL;
1009 }
1010 btrfs_release_path(path2);
1011 next:
1012 if (ptr < end) {
1013 ptr += btrfs_extent_inline_ref_size(key.type);
1014 if (ptr >= end) {
1015 WARN_ON(ptr > end);
1016 ptr = 0;
1017 end = 0;
1018 }
1019 }
1020 if (ptr >= end)
1021 path1->slots[0]++;
1022 }
1023 btrfs_release_path(path1);
1024
1025 cur->checked = 1;
1026 WARN_ON(exist);
1027
1028 /* the pending list isn't empty, take the first block to process */
1029 if (!list_empty(&list)) {
1030 edge = list_entry(list.next, struct backref_edge, list[UPPER]);
1031 list_del_init(&edge->list[UPPER]);
1032 cur = edge->node[UPPER];
1033 goto again;
1034 }
1035
1036 /*
1037 * everything goes well, connect backref nodes and insert backref nodes
1038 * into the cache.
1039 */
1040 ASSERT(node->checked);
1041 cowonly = node->cowonly;
1042 if (!cowonly) {
1043 rb_node = tree_insert(&cache->rb_root, node->bytenr,
1044 &node->rb_node);
1045 if (rb_node)
1046 backref_tree_panic(rb_node, -EEXIST, node->bytenr);
1047 list_add_tail(&node->lower, &cache->leaves);
1048 }
1049
1050 list_for_each_entry(edge, &node->upper, list[LOWER])
1051 list_add_tail(&edge->list[UPPER], &list);
1052
1053 while (!list_empty(&list)) {
1054 edge = list_entry(list.next, struct backref_edge, list[UPPER]);
1055 list_del_init(&edge->list[UPPER]);
1056 upper = edge->node[UPPER];
1057 if (upper->detached) {
1058 list_del(&edge->list[LOWER]);
1059 lower = edge->node[LOWER];
1060 free_backref_edge(cache, edge);
1061 if (list_empty(&lower->upper))
1062 list_add(&lower->list, &useless);
1063 continue;
1064 }
1065
1066 if (!RB_EMPTY_NODE(&upper->rb_node)) {
1067 if (upper->lowest) {
1068 list_del_init(&upper->lower);
1069 upper->lowest = 0;
1070 }
1071
1072 list_add_tail(&edge->list[UPPER], &upper->lower);
1073 continue;
1074 }
1075
1076 if (!upper->checked) {
1077 /*
1078 * Still want to blow up for developers since this is a
1079 * logic bug.
1080 */
1081 ASSERT(0);
1082 err = -EINVAL;
1083 goto out;
1084 }
1085 if (cowonly != upper->cowonly) {
1086 ASSERT(0);
1087 err = -EINVAL;
1088 goto out;
1089 }
1090
1091 if (!cowonly) {
1092 rb_node = tree_insert(&cache->rb_root, upper->bytenr,
1093 &upper->rb_node);
1094 if (rb_node)
1095 backref_tree_panic(rb_node, -EEXIST,
1096 upper->bytenr);
1097 }
1098
1099 list_add_tail(&edge->list[UPPER], &upper->lower);
1100
1101 list_for_each_entry(edge, &upper->upper, list[LOWER])
1102 list_add_tail(&edge->list[UPPER], &list);
1103 }
1104 /*
1105 * process useless backref nodes. backref nodes for tree leaves
1106 * are deleted from the cache. backref nodes for upper level
1107 * tree blocks are left in the cache to avoid unnecessary backref
1108 * lookup.
1109 */
1110 while (!list_empty(&useless)) {
1111 upper = list_entry(useless.next, struct backref_node, list);
1112 list_del_init(&upper->list);
1113 ASSERT(list_empty(&upper->upper));
1114 if (upper == node)
1115 node = NULL;
1116 if (upper->lowest) {
1117 list_del_init(&upper->lower);
1118 upper->lowest = 0;
1119 }
1120 while (!list_empty(&upper->lower)) {
1121 edge = list_entry(upper->lower.next,
1122 struct backref_edge, list[UPPER]);
1123 list_del(&edge->list[UPPER]);
1124 list_del(&edge->list[LOWER]);
1125 lower = edge->node[LOWER];
1126 free_backref_edge(cache, edge);
1127
1128 if (list_empty(&lower->upper))
1129 list_add(&lower->list, &useless);
1130 }
1131 __mark_block_processed(rc, upper);
1132 if (upper->level > 0) {
1133 list_add(&upper->list, &cache->detached);
1134 upper->detached = 1;
1135 } else {
1136 rb_erase(&upper->rb_node, &cache->rb_root);
1137 free_backref_node(cache, upper);
1138 }
1139 }
1140 out:
1141 btrfs_free_path(path1);
1142 btrfs_free_path(path2);
1143 if (err) {
1144 while (!list_empty(&useless)) {
1145 lower = list_entry(useless.next,
1146 struct backref_node, list);
1147 list_del_init(&lower->list);
1148 }
1149 while (!list_empty(&list)) {
1150 edge = list_first_entry(&list, struct backref_edge,
1151 list[UPPER]);
1152 list_del(&edge->list[UPPER]);
1153 list_del(&edge->list[LOWER]);
1154 lower = edge->node[LOWER];
1155 upper = edge->node[UPPER];
1156 free_backref_edge(cache, edge);
1157
1158 /*
1159 * Lower is no longer linked to any upper backref nodes
1160 * and isn't in the cache, we can free it ourselves.
1161 */
1162 if (list_empty(&lower->upper) &&
1163 RB_EMPTY_NODE(&lower->rb_node))
1164 list_add(&lower->list, &useless);
1165
1166 if (!RB_EMPTY_NODE(&upper->rb_node))
1167 continue;
1168
1169 /* Add this guy's upper edges to the list to proces */
1170 list_for_each_entry(edge, &upper->upper, list[LOWER])
1171 list_add_tail(&edge->list[UPPER], &list);
1172 if (list_empty(&upper->upper))
1173 list_add(&upper->list, &useless);
1174 }
1175
1176 while (!list_empty(&useless)) {
1177 lower = list_entry(useless.next,
1178 struct backref_node, list);
1179 list_del_init(&lower->list);
1180 free_backref_node(cache, lower);
1181 }
1182 return ERR_PTR(err);
1183 }
1184 ASSERT(!node || !node->detached);
1185 return node;
1186 }
1187
1188 /*
1189 * helper to add backref node for the newly created snapshot.
1190 * the backref node is created by cloning backref node that
1191 * corresponds to root of source tree
1192 */
1193 static int clone_backref_node(struct btrfs_trans_handle *trans,
1194 struct reloc_control *rc,
1195 struct btrfs_root *src,
1196 struct btrfs_root *dest)
1197 {
1198 struct btrfs_root *reloc_root = src->reloc_root;
1199 struct backref_cache *cache = &rc->backref_cache;
1200 struct backref_node *node = NULL;
1201 struct backref_node *new_node;
1202 struct backref_edge *edge;
1203 struct backref_edge *new_edge;
1204 struct rb_node *rb_node;
1205
1206 if (cache->last_trans > 0)
1207 update_backref_cache(trans, cache);
1208
1209 rb_node = tree_search(&cache->rb_root, src->commit_root->start);
1210 if (rb_node) {
1211 node = rb_entry(rb_node, struct backref_node, rb_node);
1212 if (node->detached)
1213 node = NULL;
1214 else
1215 BUG_ON(node->new_bytenr != reloc_root->node->start);
1216 }
1217
1218 if (!node) {
1219 rb_node = tree_search(&cache->rb_root,
1220 reloc_root->commit_root->start);
1221 if (rb_node) {
1222 node = rb_entry(rb_node, struct backref_node,
1223 rb_node);
1224 BUG_ON(node->detached);
1225 }
1226 }
1227
1228 if (!node)
1229 return 0;
1230
1231 new_node = alloc_backref_node(cache);
1232 if (!new_node)
1233 return -ENOMEM;
1234
1235 new_node->bytenr = dest->node->start;
1236 new_node->level = node->level;
1237 new_node->lowest = node->lowest;
1238 new_node->checked = 1;
1239 new_node->root = dest;
1240
1241 if (!node->lowest) {
1242 list_for_each_entry(edge, &node->lower, list[UPPER]) {
1243 new_edge = alloc_backref_edge(cache);
1244 if (!new_edge)
1245 goto fail;
1246
1247 new_edge->node[UPPER] = new_node;
1248 new_edge->node[LOWER] = edge->node[LOWER];
1249 list_add_tail(&new_edge->list[UPPER],
1250 &new_node->lower);
1251 }
1252 } else {
1253 list_add_tail(&new_node->lower, &cache->leaves);
1254 }
1255
1256 rb_node = tree_insert(&cache->rb_root, new_node->bytenr,
1257 &new_node->rb_node);
1258 if (rb_node)
1259 backref_tree_panic(rb_node, -EEXIST, new_node->bytenr);
1260
1261 if (!new_node->lowest) {
1262 list_for_each_entry(new_edge, &new_node->lower, list[UPPER]) {
1263 list_add_tail(&new_edge->list[LOWER],
1264 &new_edge->node[LOWER]->upper);
1265 }
1266 }
1267 return 0;
1268 fail:
1269 while (!list_empty(&new_node->lower)) {
1270 new_edge = list_entry(new_node->lower.next,
1271 struct backref_edge, list[UPPER]);
1272 list_del(&new_edge->list[UPPER]);
1273 free_backref_edge(cache, new_edge);
1274 }
1275 free_backref_node(cache, new_node);
1276 return -ENOMEM;
1277 }
1278
1279 /*
1280 * helper to add 'address of tree root -> reloc tree' mapping
1281 */
1282 static int __must_check __add_reloc_root(struct btrfs_root *root)
1283 {
1284 struct rb_node *rb_node;
1285 struct mapping_node *node;
1286 struct reloc_control *rc = root->fs_info->reloc_ctl;
1287
1288 node = kmalloc(sizeof(*node), GFP_NOFS);
1289 if (!node)
1290 return -ENOMEM;
1291
1292 node->bytenr = root->commit_root->start;
1293 node->data = root;
1294
1295 spin_lock(&rc->reloc_root_tree.lock);
1296 rb_node = tree_insert(&rc->reloc_root_tree.rb_root,
1297 node->bytenr, &node->rb_node);
1298 spin_unlock(&rc->reloc_root_tree.lock);
1299 if (rb_node) {
1300 btrfs_panic(root->fs_info, -EEXIST, "Duplicate root found "
1301 "for start=%llu while inserting into relocation "
1302 "tree", node->bytenr);
1303 kfree(node);
1304 return -EEXIST;
1305 }
1306
1307 list_add_tail(&root->root_list, &rc->reloc_roots);
1308 return 0;
1309 }
1310
1311 /*
1312 * helper to delete the 'address of tree root -> reloc tree'
1313 * mapping
1314 */
1315 static void __del_reloc_root(struct btrfs_root *root)
1316 {
1317 struct rb_node *rb_node;
1318 struct mapping_node *node = NULL;
1319 struct reloc_control *rc = root->fs_info->reloc_ctl;
1320
1321 if (rc && root->node) {
1322 spin_lock(&rc->reloc_root_tree.lock);
1323 rb_node = tree_search(&rc->reloc_root_tree.rb_root,
1324 root->commit_root->start);
1325 if (rb_node) {
1326 node = rb_entry(rb_node, struct mapping_node, rb_node);
1327 rb_erase(&node->rb_node, &rc->reloc_root_tree.rb_root);
1328 RB_CLEAR_NODE(&node->rb_node);
1329 }
1330 spin_unlock(&rc->reloc_root_tree.lock);
1331 if (!node)
1332 return;
1333 BUG_ON((struct btrfs_root *)node->data != root);
1334 }
1335
1336 spin_lock(&root->fs_info->trans_lock);
1337 list_del_init(&root->root_list);
1338 spin_unlock(&root->fs_info->trans_lock);
1339 kfree(node);
1340 }
1341
1342 /*
1343 * helper to update the 'address of tree root -> reloc tree'
1344 * mapping
1345 */
1346 static int __update_reloc_root(struct btrfs_root *root)
1347 {
1348 struct rb_node *rb_node;
1349 struct mapping_node *node = NULL;
1350 struct reloc_control *rc = root->fs_info->reloc_ctl;
1351
1352 spin_lock(&rc->reloc_root_tree.lock);
1353 rb_node = tree_search(&rc->reloc_root_tree.rb_root,
1354 root->commit_root->start);
1355 if (rb_node) {
1356 node = rb_entry(rb_node, struct mapping_node, rb_node);
1357 rb_erase(&node->rb_node, &rc->reloc_root_tree.rb_root);
1358 }
1359 spin_unlock(&rc->reloc_root_tree.lock);
1360
1361 if (!node)
1362 return 0;
1363 BUG_ON((struct btrfs_root *)node->data != root);
1364
1365 spin_lock(&rc->reloc_root_tree.lock);
1366 node->bytenr = root->node->start;
1367 rb_node = tree_insert(&rc->reloc_root_tree.rb_root,
1368 node->bytenr, &node->rb_node);
1369 spin_unlock(&rc->reloc_root_tree.lock);
1370 if (rb_node)
1371 backref_tree_panic(rb_node, -EEXIST, node->bytenr);
1372 return 0;
1373 }
1374
1375 static struct btrfs_root *create_reloc_root(struct btrfs_trans_handle *trans,
1376 struct btrfs_root *root, u64 objectid)
1377 {
1378 struct btrfs_root *reloc_root;
1379 struct extent_buffer *eb;
1380 struct btrfs_root_item *root_item;
1381 struct btrfs_key root_key;
1382 u64 last_snap = 0;
1383 int ret;
1384
1385 root_item = kmalloc(sizeof(*root_item), GFP_NOFS);
1386 BUG_ON(!root_item);
1387
1388 root_key.objectid = BTRFS_TREE_RELOC_OBJECTID;
1389 root_key.type = BTRFS_ROOT_ITEM_KEY;
1390 root_key.offset = objectid;
1391
1392 if (root->root_key.objectid == objectid) {
1393 /* called by btrfs_init_reloc_root */
1394 ret = btrfs_copy_root(trans, root, root->commit_root, &eb,
1395 BTRFS_TREE_RELOC_OBJECTID);
1396 BUG_ON(ret);
1397
1398 last_snap = btrfs_root_last_snapshot(&root->root_item);
1399 btrfs_set_root_last_snapshot(&root->root_item,
1400 trans->transid - 1);
1401 } else {
1402 /*
1403 * called by btrfs_reloc_post_snapshot_hook.
1404 * the source tree is a reloc tree, all tree blocks
1405 * modified after it was created have RELOC flag
1406 * set in their headers. so it's OK to not update
1407 * the 'last_snapshot'.
1408 */
1409 ret = btrfs_copy_root(trans, root, root->node, &eb,
1410 BTRFS_TREE_RELOC_OBJECTID);
1411 BUG_ON(ret);
1412 }
1413
1414 memcpy(root_item, &root->root_item, sizeof(*root_item));
1415 btrfs_set_root_bytenr(root_item, eb->start);
1416 btrfs_set_root_level(root_item, btrfs_header_level(eb));
1417 btrfs_set_root_generation(root_item, trans->transid);
1418
1419 if (root->root_key.objectid == objectid) {
1420 btrfs_set_root_refs(root_item, 0);
1421 memset(&root_item->drop_progress, 0,
1422 sizeof(struct btrfs_disk_key));
1423 root_item->drop_level = 0;
1424 /*
1425 * abuse rtransid, it is safe because it is impossible to
1426 * receive data into a relocation tree.
1427 */
1428 btrfs_set_root_rtransid(root_item, last_snap);
1429 btrfs_set_root_otransid(root_item, trans->transid);
1430 }
1431
1432 btrfs_tree_unlock(eb);
1433 free_extent_buffer(eb);
1434
1435 ret = btrfs_insert_root(trans, root->fs_info->tree_root,
1436 &root_key, root_item);
1437 BUG_ON(ret);
1438 kfree(root_item);
1439
1440 reloc_root = btrfs_read_fs_root(root->fs_info->tree_root, &root_key);
1441 BUG_ON(IS_ERR(reloc_root));
1442 reloc_root->last_trans = trans->transid;
1443 return reloc_root;
1444 }
1445
1446 /*
1447 * create reloc tree for a given fs tree. reloc tree is just a
1448 * snapshot of the fs tree with special root objectid.
1449 */
1450 int btrfs_init_reloc_root(struct btrfs_trans_handle *trans,
1451 struct btrfs_root *root)
1452 {
1453 struct btrfs_root *reloc_root;
1454 struct reloc_control *rc = root->fs_info->reloc_ctl;
1455 struct btrfs_block_rsv *rsv;
1456 int clear_rsv = 0;
1457 int ret;
1458
1459 if (root->reloc_root) {
1460 reloc_root = root->reloc_root;
1461 reloc_root->last_trans = trans->transid;
1462 return 0;
1463 }
1464
1465 if (!rc || !rc->create_reloc_tree ||
1466 root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID)
1467 return 0;
1468
1469 if (!trans->reloc_reserved) {
1470 rsv = trans->block_rsv;
1471 trans->block_rsv = rc->block_rsv;
1472 clear_rsv = 1;
1473 }
1474 reloc_root = create_reloc_root(trans, root, root->root_key.objectid);
1475 if (clear_rsv)
1476 trans->block_rsv = rsv;
1477
1478 ret = __add_reloc_root(reloc_root);
1479 BUG_ON(ret < 0);
1480 root->reloc_root = reloc_root;
1481 return 0;
1482 }
1483
1484 /*
1485 * update root item of reloc tree
1486 */
1487 int btrfs_update_reloc_root(struct btrfs_trans_handle *trans,
1488 struct btrfs_root *root)
1489 {
1490 struct btrfs_root *reloc_root;
1491 struct btrfs_root_item *root_item;
1492 int ret;
1493
1494 if (!root->reloc_root)
1495 goto out;
1496
1497 reloc_root = root->reloc_root;
1498 root_item = &reloc_root->root_item;
1499
1500 if (root->fs_info->reloc_ctl->merge_reloc_tree &&
1501 btrfs_root_refs(root_item) == 0) {
1502 root->reloc_root = NULL;
1503 __del_reloc_root(reloc_root);
1504 }
1505
1506 if (reloc_root->commit_root != reloc_root->node) {
1507 __update_reloc_root(reloc_root);
1508 btrfs_set_root_node(root_item, reloc_root->node);
1509 free_extent_buffer(reloc_root->commit_root);
1510 reloc_root->commit_root = btrfs_root_node(reloc_root);
1511 }
1512
1513 ret = btrfs_update_root(trans, root->fs_info->tree_root,
1514 &reloc_root->root_key, root_item);
1515 BUG_ON(ret);
1516
1517 out:
1518 return 0;
1519 }
1520
1521 /*
1522 * helper to find first cached inode with inode number >= objectid
1523 * in a subvolume
1524 */
1525 static struct inode *find_next_inode(struct btrfs_root *root, u64 objectid)
1526 {
1527 struct rb_node *node;
1528 struct rb_node *prev;
1529 struct btrfs_inode *entry;
1530 struct inode *inode;
1531
1532 spin_lock(&root->inode_lock);
1533 again:
1534 node = root->inode_tree.rb_node;
1535 prev = NULL;
1536 while (node) {
1537 prev = node;
1538 entry = rb_entry(node, struct btrfs_inode, rb_node);
1539
1540 if (objectid < btrfs_ino(&entry->vfs_inode))
1541 node = node->rb_left;
1542 else if (objectid > btrfs_ino(&entry->vfs_inode))
1543 node = node->rb_right;
1544 else
1545 break;
1546 }
1547 if (!node) {
1548 while (prev) {
1549 entry = rb_entry(prev, struct btrfs_inode, rb_node);
1550 if (objectid <= btrfs_ino(&entry->vfs_inode)) {
1551 node = prev;
1552 break;
1553 }
1554 prev = rb_next(prev);
1555 }
1556 }
1557 while (node) {
1558 entry = rb_entry(node, struct btrfs_inode, rb_node);
1559 inode = igrab(&entry->vfs_inode);
1560 if (inode) {
1561 spin_unlock(&root->inode_lock);
1562 return inode;
1563 }
1564
1565 objectid = btrfs_ino(&entry->vfs_inode) + 1;
1566 if (cond_resched_lock(&root->inode_lock))
1567 goto again;
1568
1569 node = rb_next(node);
1570 }
1571 spin_unlock(&root->inode_lock);
1572 return NULL;
1573 }
1574
1575 static int in_block_group(u64 bytenr,
1576 struct btrfs_block_group_cache *block_group)
1577 {
1578 if (bytenr >= block_group->key.objectid &&
1579 bytenr < block_group->key.objectid + block_group->key.offset)
1580 return 1;
1581 return 0;
1582 }
1583
1584 /*
1585 * get new location of data
1586 */
1587 static int get_new_location(struct inode *reloc_inode, u64 *new_bytenr,
1588 u64 bytenr, u64 num_bytes)
1589 {
1590 struct btrfs_root *root = BTRFS_I(reloc_inode)->root;
1591 struct btrfs_path *path;
1592 struct btrfs_file_extent_item *fi;
1593 struct extent_buffer *leaf;
1594 int ret;
1595
1596 path = btrfs_alloc_path();
1597 if (!path)
1598 return -ENOMEM;
1599
1600 bytenr -= BTRFS_I(reloc_inode)->index_cnt;
1601 ret = btrfs_lookup_file_extent(NULL, root, path, btrfs_ino(reloc_inode),
1602 bytenr, 0);
1603 if (ret < 0)
1604 goto out;
1605 if (ret > 0) {
1606 ret = -ENOENT;
1607 goto out;
1608 }
1609
1610 leaf = path->nodes[0];
1611 fi = btrfs_item_ptr(leaf, path->slots[0],
1612 struct btrfs_file_extent_item);
1613
1614 BUG_ON(btrfs_file_extent_offset(leaf, fi) ||
1615 btrfs_file_extent_compression(leaf, fi) ||
1616 btrfs_file_extent_encryption(leaf, fi) ||
1617 btrfs_file_extent_other_encoding(leaf, fi));
1618
1619 if (num_bytes != btrfs_file_extent_disk_num_bytes(leaf, fi)) {
1620 ret = -EINVAL;
1621 goto out;
1622 }
1623
1624 *new_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
1625 ret = 0;
1626 out:
1627 btrfs_free_path(path);
1628 return ret;
1629 }
1630
1631 /*
1632 * update file extent items in the tree leaf to point to
1633 * the new locations.
1634 */
1635 static noinline_for_stack
1636 int replace_file_extents(struct btrfs_trans_handle *trans,
1637 struct reloc_control *rc,
1638 struct btrfs_root *root,
1639 struct extent_buffer *leaf)
1640 {
1641 struct btrfs_key key;
1642 struct btrfs_file_extent_item *fi;
1643 struct inode *inode = NULL;
1644 u64 parent;
1645 u64 bytenr;
1646 u64 new_bytenr = 0;
1647 u64 num_bytes;
1648 u64 end;
1649 u32 nritems;
1650 u32 i;
1651 int ret = 0;
1652 int first = 1;
1653 int dirty = 0;
1654
1655 if (rc->stage != UPDATE_DATA_PTRS)
1656 return 0;
1657
1658 /* reloc trees always use full backref */
1659 if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID)
1660 parent = leaf->start;
1661 else
1662 parent = 0;
1663
1664 nritems = btrfs_header_nritems(leaf);
1665 for (i = 0; i < nritems; i++) {
1666 cond_resched();
1667 btrfs_item_key_to_cpu(leaf, &key, i);
1668 if (key.type != BTRFS_EXTENT_DATA_KEY)
1669 continue;
1670 fi = btrfs_item_ptr(leaf, i, struct btrfs_file_extent_item);
1671 if (btrfs_file_extent_type(leaf, fi) ==
1672 BTRFS_FILE_EXTENT_INLINE)
1673 continue;
1674 bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
1675 num_bytes = btrfs_file_extent_disk_num_bytes(leaf, fi);
1676 if (bytenr == 0)
1677 continue;
1678 if (!in_block_group(bytenr, rc->block_group))
1679 continue;
1680
1681 /*
1682 * if we are modifying block in fs tree, wait for readpage
1683 * to complete and drop the extent cache
1684 */
1685 if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) {
1686 if (first) {
1687 inode = find_next_inode(root, key.objectid);
1688 first = 0;
1689 } else if (inode && btrfs_ino(inode) < key.objectid) {
1690 btrfs_add_delayed_iput(inode);
1691 inode = find_next_inode(root, key.objectid);
1692 }
1693 if (inode && btrfs_ino(inode) == key.objectid) {
1694 end = key.offset +
1695 btrfs_file_extent_num_bytes(leaf, fi);
1696 WARN_ON(!IS_ALIGNED(key.offset,
1697 root->sectorsize));
1698 WARN_ON(!IS_ALIGNED(end, root->sectorsize));
1699 end--;
1700 ret = try_lock_extent(&BTRFS_I(inode)->io_tree,
1701 key.offset, end);
1702 if (!ret)
1703 continue;
1704
1705 btrfs_drop_extent_cache(inode, key.offset, end,
1706 1);
1707 unlock_extent(&BTRFS_I(inode)->io_tree,
1708 key.offset, end);
1709 }
1710 }
1711
1712 ret = get_new_location(rc->data_inode, &new_bytenr,
1713 bytenr, num_bytes);
1714 if (ret) {
1715 /*
1716 * Don't have to abort since we've not changed anything
1717 * in the file extent yet.
1718 */
1719 break;
1720 }
1721
1722 btrfs_set_file_extent_disk_bytenr(leaf, fi, new_bytenr);
1723 dirty = 1;
1724
1725 key.offset -= btrfs_file_extent_offset(leaf, fi);
1726 ret = btrfs_inc_extent_ref(trans, root, new_bytenr,
1727 num_bytes, parent,
1728 btrfs_header_owner(leaf),
1729 key.objectid, key.offset);
1730 if (ret) {
1731 btrfs_abort_transaction(trans, root, ret);
1732 break;
1733 }
1734
1735 ret = btrfs_free_extent(trans, root, bytenr, num_bytes,
1736 parent, btrfs_header_owner(leaf),
1737 key.objectid, key.offset);
1738 if (ret) {
1739 btrfs_abort_transaction(trans, root, ret);
1740 break;
1741 }
1742 }
1743 if (dirty)
1744 btrfs_mark_buffer_dirty(leaf);
1745 if (inode)
1746 btrfs_add_delayed_iput(inode);
1747 return ret;
1748 }
1749
1750 static noinline_for_stack
1751 int memcmp_node_keys(struct extent_buffer *eb, int slot,
1752 struct btrfs_path *path, int level)
1753 {
1754 struct btrfs_disk_key key1;
1755 struct btrfs_disk_key key2;
1756 btrfs_node_key(eb, &key1, slot);
1757 btrfs_node_key(path->nodes[level], &key2, path->slots[level]);
1758 return memcmp(&key1, &key2, sizeof(key1));
1759 }
1760
1761 /*
1762 * try to replace tree blocks in fs tree with the new blocks
1763 * in reloc tree. tree blocks haven't been modified since the
1764 * reloc tree was create can be replaced.
1765 *
1766 * if a block was replaced, level of the block + 1 is returned.
1767 * if no block got replaced, 0 is returned. if there are other
1768 * errors, a negative error number is returned.
1769 */
1770 static noinline_for_stack
1771 int replace_path(struct btrfs_trans_handle *trans,
1772 struct btrfs_root *dest, struct btrfs_root *src,
1773 struct btrfs_path *path, struct btrfs_key *next_key,
1774 int lowest_level, int max_level)
1775 {
1776 struct extent_buffer *eb;
1777 struct extent_buffer *parent;
1778 struct btrfs_key key;
1779 u64 old_bytenr;
1780 u64 new_bytenr;
1781 u64 old_ptr_gen;
1782 u64 new_ptr_gen;
1783 u64 last_snapshot;
1784 u32 blocksize;
1785 int cow = 0;
1786 int level;
1787 int ret;
1788 int slot;
1789
1790 BUG_ON(src->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID);
1791 BUG_ON(dest->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID);
1792
1793 last_snapshot = btrfs_root_last_snapshot(&src->root_item);
1794 again:
1795 slot = path->slots[lowest_level];
1796 btrfs_node_key_to_cpu(path->nodes[lowest_level], &key, slot);
1797
1798 eb = btrfs_lock_root_node(dest);
1799 btrfs_set_lock_blocking(eb);
1800 level = btrfs_header_level(eb);
1801
1802 if (level < lowest_level) {
1803 btrfs_tree_unlock(eb);
1804 free_extent_buffer(eb);
1805 return 0;
1806 }
1807
1808 if (cow) {
1809 ret = btrfs_cow_block(trans, dest, eb, NULL, 0, &eb);
1810 BUG_ON(ret);
1811 }
1812 btrfs_set_lock_blocking(eb);
1813
1814 if (next_key) {
1815 next_key->objectid = (u64)-1;
1816 next_key->type = (u8)-1;
1817 next_key->offset = (u64)-1;
1818 }
1819
1820 parent = eb;
1821 while (1) {
1822 level = btrfs_header_level(parent);
1823 BUG_ON(level < lowest_level);
1824
1825 ret = btrfs_bin_search(parent, &key, level, &slot);
1826 if (ret && slot > 0)
1827 slot--;
1828
1829 if (next_key && slot + 1 < btrfs_header_nritems(parent))
1830 btrfs_node_key_to_cpu(parent, next_key, slot + 1);
1831
1832 old_bytenr = btrfs_node_blockptr(parent, slot);
1833 blocksize = dest->nodesize;
1834 old_ptr_gen = btrfs_node_ptr_generation(parent, slot);
1835
1836 if (level <= max_level) {
1837 eb = path->nodes[level];
1838 new_bytenr = btrfs_node_blockptr(eb,
1839 path->slots[level]);
1840 new_ptr_gen = btrfs_node_ptr_generation(eb,
1841 path->slots[level]);
1842 } else {
1843 new_bytenr = 0;
1844 new_ptr_gen = 0;
1845 }
1846
1847 if (WARN_ON(new_bytenr > 0 && new_bytenr == old_bytenr)) {
1848 ret = level;
1849 break;
1850 }
1851
1852 if (new_bytenr == 0 || old_ptr_gen > last_snapshot ||
1853 memcmp_node_keys(parent, slot, path, level)) {
1854 if (level <= lowest_level) {
1855 ret = 0;
1856 break;
1857 }
1858
1859 eb = read_tree_block(dest, old_bytenr, old_ptr_gen);
1860 if (IS_ERR(eb)) {
1861 ret = PTR_ERR(eb);
1862 } else if (!extent_buffer_uptodate(eb)) {
1863 ret = -EIO;
1864 free_extent_buffer(eb);
1865 break;
1866 }
1867 btrfs_tree_lock(eb);
1868 if (cow) {
1869 ret = btrfs_cow_block(trans, dest, eb, parent,
1870 slot, &eb);
1871 BUG_ON(ret);
1872 }
1873 btrfs_set_lock_blocking(eb);
1874
1875 btrfs_tree_unlock(parent);
1876 free_extent_buffer(parent);
1877
1878 parent = eb;
1879 continue;
1880 }
1881
1882 if (!cow) {
1883 btrfs_tree_unlock(parent);
1884 free_extent_buffer(parent);
1885 cow = 1;
1886 goto again;
1887 }
1888
1889 btrfs_node_key_to_cpu(path->nodes[level], &key,
1890 path->slots[level]);
1891 btrfs_release_path(path);
1892
1893 path->lowest_level = level;
1894 ret = btrfs_search_slot(trans, src, &key, path, 0, 1);
1895 path->lowest_level = 0;
1896 BUG_ON(ret);
1897
1898 /*
1899 * swap blocks in fs tree and reloc tree.
1900 */
1901 btrfs_set_node_blockptr(parent, slot, new_bytenr);
1902 btrfs_set_node_ptr_generation(parent, slot, new_ptr_gen);
1903 btrfs_mark_buffer_dirty(parent);
1904
1905 btrfs_set_node_blockptr(path->nodes[level],
1906 path->slots[level], old_bytenr);
1907 btrfs_set_node_ptr_generation(path->nodes[level],
1908 path->slots[level], old_ptr_gen);
1909 btrfs_mark_buffer_dirty(path->nodes[level]);
1910
1911 ret = btrfs_inc_extent_ref(trans, src, old_bytenr, blocksize,
1912 path->nodes[level]->start,
1913 src->root_key.objectid, level - 1, 0);
1914 BUG_ON(ret);
1915 ret = btrfs_inc_extent_ref(trans, dest, new_bytenr, blocksize,
1916 0, dest->root_key.objectid, level - 1,
1917 0);
1918 BUG_ON(ret);
1919
1920 ret = btrfs_free_extent(trans, src, new_bytenr, blocksize,
1921 path->nodes[level]->start,
1922 src->root_key.objectid, level - 1, 0);
1923 BUG_ON(ret);
1924
1925 ret = btrfs_free_extent(trans, dest, old_bytenr, blocksize,
1926 0, dest->root_key.objectid, level - 1,
1927 0);
1928 BUG_ON(ret);
1929
1930 btrfs_unlock_up_safe(path, 0);
1931
1932 ret = level;
1933 break;
1934 }
1935 btrfs_tree_unlock(parent);
1936 free_extent_buffer(parent);
1937 return ret;
1938 }
1939
1940 /*
1941 * helper to find next relocated block in reloc tree
1942 */
1943 static noinline_for_stack
1944 int walk_up_reloc_tree(struct btrfs_root *root, struct btrfs_path *path,
1945 int *level)
1946 {
1947 struct extent_buffer *eb;
1948 int i;
1949 u64 last_snapshot;
1950 u32 nritems;
1951
1952 last_snapshot = btrfs_root_last_snapshot(&root->root_item);
1953
1954 for (i = 0; i < *level; i++) {
1955 free_extent_buffer(path->nodes[i]);
1956 path->nodes[i] = NULL;
1957 }
1958
1959 for (i = *level; i < BTRFS_MAX_LEVEL && path->nodes[i]; i++) {
1960 eb = path->nodes[i];
1961 nritems = btrfs_header_nritems(eb);
1962 while (path->slots[i] + 1 < nritems) {
1963 path->slots[i]++;
1964 if (btrfs_node_ptr_generation(eb, path->slots[i]) <=
1965 last_snapshot)
1966 continue;
1967
1968 *level = i;
1969 return 0;
1970 }
1971 free_extent_buffer(path->nodes[i]);
1972 path->nodes[i] = NULL;
1973 }
1974 return 1;
1975 }
1976
1977 /*
1978 * walk down reloc tree to find relocated block of lowest level
1979 */
1980 static noinline_for_stack
1981 int walk_down_reloc_tree(struct btrfs_root *root, struct btrfs_path *path,
1982 int *level)
1983 {
1984 struct extent_buffer *eb = NULL;
1985 int i;
1986 u64 bytenr;
1987 u64 ptr_gen = 0;
1988 u64 last_snapshot;
1989 u32 nritems;
1990
1991 last_snapshot = btrfs_root_last_snapshot(&root->root_item);
1992
1993 for (i = *level; i > 0; i--) {
1994 eb = path->nodes[i];
1995 nritems = btrfs_header_nritems(eb);
1996 while (path->slots[i] < nritems) {
1997 ptr_gen = btrfs_node_ptr_generation(eb, path->slots[i]);
1998 if (ptr_gen > last_snapshot)
1999 break;
2000 path->slots[i]++;
2001 }
2002 if (path->slots[i] >= nritems) {
2003 if (i == *level)
2004 break;
2005 *level = i + 1;
2006 return 0;
2007 }
2008 if (i == 1) {
2009 *level = i;
2010 return 0;
2011 }
2012
2013 bytenr = btrfs_node_blockptr(eb, path->slots[i]);
2014 eb = read_tree_block(root, bytenr, ptr_gen);
2015 if (IS_ERR(eb)) {
2016 return PTR_ERR(eb);
2017 } else if (!extent_buffer_uptodate(eb)) {
2018 free_extent_buffer(eb);
2019 return -EIO;
2020 }
2021 BUG_ON(btrfs_header_level(eb) != i - 1);
2022 path->nodes[i - 1] = eb;
2023 path->slots[i - 1] = 0;
2024 }
2025 return 1;
2026 }
2027
2028 /*
2029 * invalidate extent cache for file extents whose key in range of
2030 * [min_key, max_key)
2031 */
2032 static int invalidate_extent_cache(struct btrfs_root *root,
2033 struct btrfs_key *min_key,
2034 struct btrfs_key *max_key)
2035 {
2036 struct inode *inode = NULL;
2037 u64 objectid;
2038 u64 start, end;
2039 u64 ino;
2040
2041 objectid = min_key->objectid;
2042 while (1) {
2043 cond_resched();
2044 iput(inode);
2045
2046 if (objectid > max_key->objectid)
2047 break;
2048
2049 inode = find_next_inode(root, objectid);
2050 if (!inode)
2051 break;
2052 ino = btrfs_ino(inode);
2053
2054 if (ino > max_key->objectid) {
2055 iput(inode);
2056 break;
2057 }
2058
2059 objectid = ino + 1;
2060 if (!S_ISREG(inode->i_mode))
2061 continue;
2062
2063 if (unlikely(min_key->objectid == ino)) {
2064 if (min_key->type > BTRFS_EXTENT_DATA_KEY)
2065 continue;
2066 if (min_key->type < BTRFS_EXTENT_DATA_KEY)
2067 start = 0;
2068 else {
2069 start = min_key->offset;
2070 WARN_ON(!IS_ALIGNED(start, root->sectorsize));
2071 }
2072 } else {
2073 start = 0;
2074 }
2075
2076 if (unlikely(max_key->objectid == ino)) {
2077 if (max_key->type < BTRFS_EXTENT_DATA_KEY)
2078 continue;
2079 if (max_key->type > BTRFS_EXTENT_DATA_KEY) {
2080 end = (u64)-1;
2081 } else {
2082 if (max_key->offset == 0)
2083 continue;
2084 end = max_key->offset;
2085 WARN_ON(!IS_ALIGNED(end, root->sectorsize));
2086 end--;
2087 }
2088 } else {
2089 end = (u64)-1;
2090 }
2091
2092 /* the lock_extent waits for readpage to complete */
2093 lock_extent(&BTRFS_I(inode)->io_tree, start, end);
2094 btrfs_drop_extent_cache(inode, start, end, 1);
2095 unlock_extent(&BTRFS_I(inode)->io_tree, start, end);
2096 }
2097 return 0;
2098 }
2099
2100 static int find_next_key(struct btrfs_path *path, int level,
2101 struct btrfs_key *key)
2102
2103 {
2104 while (level < BTRFS_MAX_LEVEL) {
2105 if (!path->nodes[level])
2106 break;
2107 if (path->slots[level] + 1 <
2108 btrfs_header_nritems(path->nodes[level])) {
2109 btrfs_node_key_to_cpu(path->nodes[level], key,
2110 path->slots[level] + 1);
2111 return 0;
2112 }
2113 level++;
2114 }
2115 return 1;
2116 }
2117
2118 /*
2119 * merge the relocated tree blocks in reloc tree with corresponding
2120 * fs tree.
2121 */
2122 static noinline_for_stack int merge_reloc_root(struct reloc_control *rc,
2123 struct btrfs_root *root)
2124 {
2125 LIST_HEAD(inode_list);
2126 struct btrfs_key key;
2127 struct btrfs_key next_key;
2128 struct btrfs_trans_handle *trans = NULL;
2129 struct btrfs_root *reloc_root;
2130 struct btrfs_root_item *root_item;
2131 struct btrfs_path *path;
2132 struct extent_buffer *leaf;
2133 int level;
2134 int max_level;
2135 int replaced = 0;
2136 int ret;
2137 int err = 0;
2138 u32 min_reserved;
2139
2140 path = btrfs_alloc_path();
2141 if (!path)
2142 return -ENOMEM;
2143 path->reada = 1;
2144
2145 reloc_root = root->reloc_root;
2146 root_item = &reloc_root->root_item;
2147
2148 if (btrfs_disk_key_objectid(&root_item->drop_progress) == 0) {
2149 level = btrfs_root_level(root_item);
2150 extent_buffer_get(reloc_root->node);
2151 path->nodes[level] = reloc_root->node;
2152 path->slots[level] = 0;
2153 } else {
2154 btrfs_disk_key_to_cpu(&key, &root_item->drop_progress);
2155
2156 level = root_item->drop_level;
2157 BUG_ON(level == 0);
2158 path->lowest_level = level;
2159 ret = btrfs_search_slot(NULL, reloc_root, &key, path, 0, 0);
2160 path->lowest_level = 0;
2161 if (ret < 0) {
2162 btrfs_free_path(path);
2163 return ret;
2164 }
2165
2166 btrfs_node_key_to_cpu(path->nodes[level], &next_key,
2167 path->slots[level]);
2168 WARN_ON(memcmp(&key, &next_key, sizeof(key)));
2169
2170 btrfs_unlock_up_safe(path, 0);
2171 }
2172
2173 min_reserved = root->nodesize * (BTRFS_MAX_LEVEL - 1) * 2;
2174 memset(&next_key, 0, sizeof(next_key));
2175
2176 while (1) {
2177 ret = btrfs_block_rsv_refill(root, rc->block_rsv, min_reserved,
2178 BTRFS_RESERVE_FLUSH_ALL);
2179 if (ret) {
2180 err = ret;
2181 goto out;
2182 }
2183 trans = btrfs_start_transaction(root, 0);
2184 if (IS_ERR(trans)) {
2185 err = PTR_ERR(trans);
2186 trans = NULL;
2187 goto out;
2188 }
2189 trans->block_rsv = rc->block_rsv;
2190
2191 replaced = 0;
2192 max_level = level;
2193
2194 ret = walk_down_reloc_tree(reloc_root, path, &level);
2195 if (ret < 0) {
2196 err = ret;
2197 goto out;
2198 }
2199 if (ret > 0)
2200 break;
2201
2202 if (!find_next_key(path, level, &key) &&
2203 btrfs_comp_cpu_keys(&next_key, &key) >= 0) {
2204 ret = 0;
2205 } else {
2206 ret = replace_path(trans, root, reloc_root, path,
2207 &next_key, level, max_level);
2208 }
2209 if (ret < 0) {
2210 err = ret;
2211 goto out;
2212 }
2213
2214 if (ret > 0) {
2215 level = ret;
2216 btrfs_node_key_to_cpu(path->nodes[level], &key,
2217 path->slots[level]);
2218 replaced = 1;
2219 }
2220
2221 ret = walk_up_reloc_tree(reloc_root, path, &level);
2222 if (ret > 0)
2223 break;
2224
2225 BUG_ON(level == 0);
2226 /*
2227 * save the merging progress in the drop_progress.
2228 * this is OK since root refs == 1 in this case.
2229 */
2230 btrfs_node_key(path->nodes[level], &root_item->drop_progress,
2231 path->slots[level]);
2232 root_item->drop_level = level;
2233
2234 btrfs_end_transaction_throttle(trans, root);
2235 trans = NULL;
2236
2237 btrfs_btree_balance_dirty(root);
2238
2239 if (replaced && rc->stage == UPDATE_DATA_PTRS)
2240 invalidate_extent_cache(root, &key, &next_key);
2241 }
2242
2243 /*
2244 * handle the case only one block in the fs tree need to be
2245 * relocated and the block is tree root.
2246 */
2247 leaf = btrfs_lock_root_node(root);
2248 ret = btrfs_cow_block(trans, root, leaf, NULL, 0, &leaf);
2249 btrfs_tree_unlock(leaf);
2250 free_extent_buffer(leaf);
2251 if (ret < 0)
2252 err = ret;
2253 out:
2254 btrfs_free_path(path);
2255
2256 if (err == 0) {
2257 memset(&root_item->drop_progress, 0,
2258 sizeof(root_item->drop_progress));
2259 root_item->drop_level = 0;
2260 btrfs_set_root_refs(root_item, 0);
2261 btrfs_update_reloc_root(trans, root);
2262 }
2263
2264 if (trans)
2265 btrfs_end_transaction_throttle(trans, root);
2266
2267 btrfs_btree_balance_dirty(root);
2268
2269 if (replaced && rc->stage == UPDATE_DATA_PTRS)
2270 invalidate_extent_cache(root, &key, &next_key);
2271
2272 return err;
2273 }
2274
2275 static noinline_for_stack
2276 int prepare_to_merge(struct reloc_control *rc, int err)
2277 {
2278 struct btrfs_root *root = rc->extent_root;
2279 struct btrfs_root *reloc_root;
2280 struct btrfs_trans_handle *trans;
2281 LIST_HEAD(reloc_roots);
2282 u64 num_bytes = 0;
2283 int ret;
2284
2285 mutex_lock(&root->fs_info->reloc_mutex);
2286 rc->merging_rsv_size += root->nodesize * (BTRFS_MAX_LEVEL - 1) * 2;
2287 rc->merging_rsv_size += rc->nodes_relocated * 2;
2288 mutex_unlock(&root->fs_info->reloc_mutex);
2289
2290 again:
2291 if (!err) {
2292 num_bytes = rc->merging_rsv_size;
2293 ret = btrfs_block_rsv_add(root, rc->block_rsv, num_bytes,
2294 BTRFS_RESERVE_FLUSH_ALL);
2295 if (ret)
2296 err = ret;
2297 }
2298
2299 trans = btrfs_join_transaction(rc->extent_root);
2300 if (IS_ERR(trans)) {
2301 if (!err)
2302 btrfs_block_rsv_release(rc->extent_root,
2303 rc->block_rsv, num_bytes);
2304 return PTR_ERR(trans);
2305 }
2306
2307 if (!err) {
2308 if (num_bytes != rc->merging_rsv_size) {
2309 btrfs_end_transaction(trans, rc->extent_root);
2310 btrfs_block_rsv_release(rc->extent_root,
2311 rc->block_rsv, num_bytes);
2312 goto again;
2313 }
2314 }
2315
2316 rc->merge_reloc_tree = 1;
2317
2318 while (!list_empty(&rc->reloc_roots)) {
2319 reloc_root = list_entry(rc->reloc_roots.next,
2320 struct btrfs_root, root_list);
2321 list_del_init(&reloc_root->root_list);
2322
2323 root = read_fs_root(reloc_root->fs_info,
2324 reloc_root->root_key.offset);
2325 BUG_ON(IS_ERR(root));
2326 BUG_ON(root->reloc_root != reloc_root);
2327
2328 /*
2329 * set reference count to 1, so btrfs_recover_relocation
2330 * knows it should resumes merging
2331 */
2332 if (!err)
2333 btrfs_set_root_refs(&reloc_root->root_item, 1);
2334 btrfs_update_reloc_root(trans, root);
2335
2336 list_add(&reloc_root->root_list, &reloc_roots);
2337 }
2338
2339 list_splice(&reloc_roots, &rc->reloc_roots);
2340
2341 if (!err)
2342 btrfs_commit_transaction(trans, rc->extent_root);
2343 else
2344 btrfs_end_transaction(trans, rc->extent_root);
2345 return err;
2346 }
2347
2348 static noinline_for_stack
2349 void free_reloc_roots(struct list_head *list)
2350 {
2351 struct btrfs_root *reloc_root;
2352
2353 while (!list_empty(list)) {
2354 reloc_root = list_entry(list->next, struct btrfs_root,
2355 root_list);
2356 __del_reloc_root(reloc_root);
2357 free_extent_buffer(reloc_root->node);
2358 free_extent_buffer(reloc_root->commit_root);
2359 reloc_root->node = NULL;
2360 reloc_root->commit_root = NULL;
2361 }
2362 }
2363
2364 static noinline_for_stack
2365 void merge_reloc_roots(struct reloc_control *rc)
2366 {
2367 struct btrfs_root *root;
2368 struct btrfs_root *reloc_root;
2369 u64 last_snap;
2370 u64 otransid;
2371 u64 objectid;
2372 LIST_HEAD(reloc_roots);
2373 int found = 0;
2374 int ret = 0;
2375 again:
2376 root = rc->extent_root;
2377
2378 /*
2379 * this serializes us with btrfs_record_root_in_transaction,
2380 * we have to make sure nobody is in the middle of
2381 * adding their roots to the list while we are
2382 * doing this splice
2383 */
2384 mutex_lock(&root->fs_info->reloc_mutex);
2385 list_splice_init(&rc->reloc_roots, &reloc_roots);
2386 mutex_unlock(&root->fs_info->reloc_mutex);
2387
2388 while (!list_empty(&reloc_roots)) {
2389 found = 1;
2390 reloc_root = list_entry(reloc_roots.next,
2391 struct btrfs_root, root_list);
2392
2393 if (btrfs_root_refs(&reloc_root->root_item) > 0) {
2394 root = read_fs_root(reloc_root->fs_info,
2395 reloc_root->root_key.offset);
2396 BUG_ON(IS_ERR(root));
2397 BUG_ON(root->reloc_root != reloc_root);
2398
2399 ret = merge_reloc_root(rc, root);
2400 if (ret) {
2401 if (list_empty(&reloc_root->root_list))
2402 list_add_tail(&reloc_root->root_list,
2403 &reloc_roots);
2404 goto out;
2405 }
2406 } else {
2407 list_del_init(&reloc_root->root_list);
2408 }
2409
2410 /*
2411 * we keep the old last snapshod transid in rtranid when we
2412 * created the relocation tree.
2413 */
2414 last_snap = btrfs_root_rtransid(&reloc_root->root_item);
2415 otransid = btrfs_root_otransid(&reloc_root->root_item);
2416 objectid = reloc_root->root_key.offset;
2417
2418 ret = btrfs_drop_snapshot(reloc_root, rc->block_rsv, 0, 1);
2419 if (ret < 0) {
2420 if (list_empty(&reloc_root->root_list))
2421 list_add_tail(&reloc_root->root_list,
2422 &reloc_roots);
2423 goto out;
2424 }
2425 }
2426
2427 if (found) {
2428 found = 0;
2429 goto again;
2430 }
2431 out:
2432 if (ret) {
2433 btrfs_std_error(root->fs_info, ret, NULL);
2434 if (!list_empty(&reloc_roots))
2435 free_reloc_roots(&reloc_roots);
2436
2437 /* new reloc root may be added */
2438 mutex_lock(&root->fs_info->reloc_mutex);
2439 list_splice_init(&rc->reloc_roots, &reloc_roots);
2440 mutex_unlock(&root->fs_info->reloc_mutex);
2441 if (!list_empty(&reloc_roots))
2442 free_reloc_roots(&reloc_roots);
2443 }
2444
2445 /*
2446 * We used to have
2447 *
2448 * BUG_ON(!RB_EMPTY_ROOT(&rc->reloc_root_tree.rb_root));
2449 *
2450 * here, but it's wrong. If we fail to start the transaction in
2451 * prepare_to_merge() we will have only 0 ref reloc roots, none of which
2452 * have actually been removed from the reloc_root_tree rb tree. This is
2453 * fine because we're bailing here, and we hold a reference on the root
2454 * for the list that holds it, so these roots will be cleaned up when we
2455 * do the reloc_dirty_list afterwards. Meanwhile the root->reloc_root
2456 * will be cleaned up on unmount.
2457 *
2458 * The remaining nodes will be cleaned up by free_reloc_control.
2459 */
2460 }
2461
2462 static void free_block_list(struct rb_root *blocks)
2463 {
2464 struct tree_block *block;
2465 struct rb_node *rb_node;
2466 while ((rb_node = rb_first(blocks))) {
2467 block = rb_entry(rb_node, struct tree_block, rb_node);
2468 rb_erase(rb_node, blocks);
2469 kfree(block);
2470 }
2471 }
2472
2473 static int record_reloc_root_in_trans(struct btrfs_trans_handle *trans,
2474 struct btrfs_root *reloc_root)
2475 {
2476 struct btrfs_root *root;
2477
2478 if (reloc_root->last_trans == trans->transid)
2479 return 0;
2480
2481 root = read_fs_root(reloc_root->fs_info, reloc_root->root_key.offset);
2482 BUG_ON(IS_ERR(root));
2483 BUG_ON(root->reloc_root != reloc_root);
2484
2485 return btrfs_record_root_in_trans(trans, root);
2486 }
2487
2488 static noinline_for_stack
2489 struct btrfs_root *select_reloc_root(struct btrfs_trans_handle *trans,
2490 struct reloc_control *rc,
2491 struct backref_node *node,
2492 struct backref_edge *edges[])
2493 {
2494 struct backref_node *next;
2495 struct btrfs_root *root;
2496 int index = 0;
2497
2498 next = node;
2499 while (1) {
2500 cond_resched();
2501 next = walk_up_backref(next, edges, &index);
2502 root = next->root;
2503 BUG_ON(!root);
2504 BUG_ON(!test_bit(BTRFS_ROOT_REF_COWS, &root->state));
2505
2506 if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID) {
2507 record_reloc_root_in_trans(trans, root);
2508 break;
2509 }
2510
2511 btrfs_record_root_in_trans(trans, root);
2512 root = root->reloc_root;
2513
2514 if (next->new_bytenr != root->node->start) {
2515 BUG_ON(next->new_bytenr);
2516 BUG_ON(!list_empty(&next->list));
2517 next->new_bytenr = root->node->start;
2518 next->root = root;
2519 list_add_tail(&next->list,
2520 &rc->backref_cache.changed);
2521 __mark_block_processed(rc, next);
2522 break;
2523 }
2524
2525 WARN_ON(1);
2526 root = NULL;
2527 next = walk_down_backref(edges, &index);
2528 if (!next || next->level <= node->level)
2529 break;
2530 }
2531 if (!root)
2532 return NULL;
2533
2534 next = node;
2535 /* setup backref node path for btrfs_reloc_cow_block */
2536 while (1) {
2537 rc->backref_cache.path[next->level] = next;
2538 if (--index < 0)
2539 break;
2540 next = edges[index]->node[UPPER];
2541 }
2542 return root;
2543 }
2544
2545 /*
2546 * select a tree root for relocation. return NULL if the block
2547 * is reference counted. we should use do_relocation() in this
2548 * case. return a tree root pointer if the block isn't reference
2549 * counted. return -ENOENT if the block is root of reloc tree.
2550 */
2551 static noinline_for_stack
2552 struct btrfs_root *select_one_root(struct backref_node *node)
2553 {
2554 struct backref_node *next;
2555 struct btrfs_root *root;
2556 struct btrfs_root *fs_root = NULL;
2557 struct backref_edge *edges[BTRFS_MAX_LEVEL - 1];
2558 int index = 0;
2559
2560 next = node;
2561 while (1) {
2562 cond_resched();
2563 next = walk_up_backref(next, edges, &index);
2564 root = next->root;
2565 BUG_ON(!root);
2566
2567 /* no other choice for non-references counted tree */
2568 if (!test_bit(BTRFS_ROOT_REF_COWS, &root->state))
2569 return root;
2570
2571 if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID)
2572 fs_root = root;
2573
2574 if (next != node)
2575 return NULL;
2576
2577 next = walk_down_backref(edges, &index);
2578 if (!next || next->level <= node->level)
2579 break;
2580 }
2581
2582 if (!fs_root)
2583 return ERR_PTR(-ENOENT);
2584 return fs_root;
2585 }
2586
2587 static noinline_for_stack
2588 u64 calcu_metadata_size(struct reloc_control *rc,
2589 struct backref_node *node, int reserve)
2590 {
2591 struct backref_node *next = node;
2592 struct backref_edge *edge;
2593 struct backref_edge *edges[BTRFS_MAX_LEVEL - 1];
2594 u64 num_bytes = 0;
2595 int index = 0;
2596
2597 BUG_ON(reserve && node->processed);
2598
2599 while (next) {
2600 cond_resched();
2601 while (1) {
2602 if (next->processed && (reserve || next != node))
2603 break;
2604
2605 num_bytes += rc->extent_root->nodesize;
2606
2607 if (list_empty(&next->upper))
2608 break;
2609
2610 edge = list_entry(next->upper.next,
2611 struct backref_edge, list[LOWER]);
2612 edges[index++] = edge;
2613 next = edge->node[UPPER];
2614 }
2615 next = walk_down_backref(edges, &index);
2616 }
2617 return num_bytes;
2618 }
2619
2620 static int reserve_metadata_space(struct btrfs_trans_handle *trans,
2621 struct reloc_control *rc,
2622 struct backref_node *node)
2623 {
2624 struct btrfs_root *root = rc->extent_root;
2625 u64 num_bytes;
2626 int ret;
2627 u64 tmp;
2628
2629 num_bytes = calcu_metadata_size(rc, node, 1) * 2;
2630
2631 trans->block_rsv = rc->block_rsv;
2632 rc->reserved_bytes += num_bytes;
2633 ret = btrfs_block_rsv_refill(root, rc->block_rsv, num_bytes,
2634 BTRFS_RESERVE_FLUSH_ALL);
2635 if (ret) {
2636 if (ret == -EAGAIN) {
2637 tmp = rc->extent_root->nodesize *
2638 RELOCATION_RESERVED_NODES;
2639 while (tmp <= rc->reserved_bytes)
2640 tmp <<= 1;
2641 /*
2642 * only one thread can access block_rsv at this point,
2643 * so we don't need hold lock to protect block_rsv.
2644 * we expand more reservation size here to allow enough
2645 * space for relocation and we will return eailer in
2646 * enospc case.
2647 */
2648 rc->block_rsv->size = tmp + rc->extent_root->nodesize *
2649 RELOCATION_RESERVED_NODES;
2650 }
2651 return ret;
2652 }
2653
2654 return 0;
2655 }
2656
2657 /*
2658 * relocate a block tree, and then update pointers in upper level
2659 * blocks that reference the block to point to the new location.
2660 *
2661 * if called by link_to_upper, the block has already been relocated.
2662 * in that case this function just updates pointers.
2663 */
2664 static int do_relocation(struct btrfs_trans_handle *trans,
2665 struct reloc_control *rc,
2666 struct backref_node *node,
2667 struct btrfs_key *key,
2668 struct btrfs_path *path, int lowest)
2669 {
2670 struct backref_node *upper;
2671 struct backref_edge *edge;
2672 struct backref_edge *edges[BTRFS_MAX_LEVEL - 1];
2673 struct btrfs_root *root;
2674 struct extent_buffer *eb;
2675 u32 blocksize;
2676 u64 bytenr;
2677 u64 generation;
2678 int slot;
2679 int ret;
2680 int err = 0;
2681
2682 BUG_ON(lowest && node->eb);
2683
2684 path->lowest_level = node->level + 1;
2685 rc->backref_cache.path[node->level] = node;
2686 list_for_each_entry(edge, &node->upper, list[LOWER]) {
2687 cond_resched();
2688
2689 upper = edge->node[UPPER];
2690 root = select_reloc_root(trans, rc, upper, edges);
2691 BUG_ON(!root);
2692
2693 if (upper->eb && !upper->locked) {
2694 if (!lowest) {
2695 ret = btrfs_bin_search(upper->eb, key,
2696 upper->level, &slot);
2697 BUG_ON(ret);
2698 bytenr = btrfs_node_blockptr(upper->eb, slot);
2699 if (node->eb->start == bytenr)
2700 goto next;
2701 }
2702 drop_node_buffer(upper);
2703 }
2704
2705 if (!upper->eb) {
2706 ret = btrfs_search_slot(trans, root, key, path, 0, 1);
2707 if (ret) {
2708 if (ret < 0)
2709 err = ret;
2710 else
2711 err = -ENOENT;
2712
2713 btrfs_release_path(path);
2714 break;
2715 }
2716
2717 if (!upper->eb) {
2718 upper->eb = path->nodes[upper->level];
2719 path->nodes[upper->level] = NULL;
2720 } else {
2721 BUG_ON(upper->eb != path->nodes[upper->level]);
2722 }
2723
2724 upper->locked = 1;
2725 path->locks[upper->level] = 0;
2726
2727 slot = path->slots[upper->level];
2728 btrfs_release_path(path);
2729 } else {
2730 ret = btrfs_bin_search(upper->eb, key, upper->level,
2731 &slot);
2732 BUG_ON(ret);
2733 }
2734
2735 bytenr = btrfs_node_blockptr(upper->eb, slot);
2736 if (lowest) {
2737 BUG_ON(bytenr != node->bytenr);
2738 } else {
2739 if (node->eb->start == bytenr)
2740 goto next;
2741 }
2742
2743 blocksize = root->nodesize;
2744 generation = btrfs_node_ptr_generation(upper->eb, slot);
2745 eb = read_tree_block(root, bytenr, generation);
2746 if (IS_ERR(eb)) {
2747 err = PTR_ERR(eb);
2748 goto next;
2749 } else if (!extent_buffer_uptodate(eb)) {
2750 free_extent_buffer(eb);
2751 err = -EIO;
2752 goto next;
2753 }
2754 btrfs_tree_lock(eb);
2755 btrfs_set_lock_blocking(eb);
2756
2757 if (!node->eb) {
2758 ret = btrfs_cow_block(trans, root, eb, upper->eb,
2759 slot, &eb);
2760 btrfs_tree_unlock(eb);
2761 free_extent_buffer(eb);
2762 if (ret < 0) {
2763 err = ret;
2764 goto next;
2765 }
2766 BUG_ON(node->eb != eb);
2767 } else {
2768 btrfs_set_node_blockptr(upper->eb, slot,
2769 node->eb->start);
2770 btrfs_set_node_ptr_generation(upper->eb, slot,
2771 trans->transid);
2772 btrfs_mark_buffer_dirty(upper->eb);
2773
2774 ret = btrfs_inc_extent_ref(trans, root,
2775 node->eb->start, blocksize,
2776 upper->eb->start,
2777 btrfs_header_owner(upper->eb),
2778 node->level, 0);
2779 BUG_ON(ret);
2780
2781 ret = btrfs_drop_subtree(trans, root, eb, upper->eb);
2782 BUG_ON(ret);
2783 }
2784 next:
2785 if (!upper->pending)
2786 drop_node_buffer(upper);
2787 else
2788 unlock_node_buffer(upper);
2789 if (err)
2790 break;
2791 }
2792
2793 if (!err && node->pending) {
2794 drop_node_buffer(node);
2795 list_move_tail(&node->list, &rc->backref_cache.changed);
2796 node->pending = 0;
2797 }
2798
2799 path->lowest_level = 0;
2800 BUG_ON(err == -ENOSPC);
2801 return err;
2802 }
2803
2804 static int link_to_upper(struct btrfs_trans_handle *trans,
2805 struct reloc_control *rc,
2806 struct backref_node *node,
2807 struct btrfs_path *path)
2808 {
2809 struct btrfs_key key;
2810
2811 btrfs_node_key_to_cpu(node->eb, &key, 0);
2812 return do_relocation(trans, rc, node, &key, path, 0);
2813 }
2814
2815 static int finish_pending_nodes(struct btrfs_trans_handle *trans,
2816 struct reloc_control *rc,
2817 struct btrfs_path *path, int err)
2818 {
2819 LIST_HEAD(list);
2820 struct backref_cache *cache = &rc->backref_cache;
2821 struct backref_node *node;
2822 int level;
2823 int ret;
2824
2825 for (level = 0; level < BTRFS_MAX_LEVEL; level++) {
2826 while (!list_empty(&cache->pending[level])) {
2827 node = list_entry(cache->pending[level].next,
2828 struct backref_node, list);
2829 list_move_tail(&node->list, &list);
2830 BUG_ON(!node->pending);
2831
2832 if (!err) {
2833 ret = link_to_upper(trans, rc, node, path);
2834 if (ret < 0)
2835 err = ret;
2836 }
2837 }
2838 list_splice_init(&list, &cache->pending[level]);
2839 }
2840 return err;
2841 }
2842
2843 static void mark_block_processed(struct reloc_control *rc,
2844 u64 bytenr, u32 blocksize)
2845 {
2846 set_extent_bits(&rc->processed_blocks, bytenr, bytenr + blocksize - 1,
2847 EXTENT_DIRTY, GFP_NOFS);
2848 }
2849
2850 static void __mark_block_processed(struct reloc_control *rc,
2851 struct backref_node *node)
2852 {
2853 u32 blocksize;
2854 if (node->level == 0 ||
2855 in_block_group(node->bytenr, rc->block_group)) {
2856 blocksize = rc->extent_root->nodesize;
2857 mark_block_processed(rc, node->bytenr, blocksize);
2858 }
2859 node->processed = 1;
2860 }
2861
2862 /*
2863 * mark a block and all blocks directly/indirectly reference the block
2864 * as processed.
2865 */
2866 static void update_processed_blocks(struct reloc_control *rc,
2867 struct backref_node *node)
2868 {
2869 struct backref_node *next = node;
2870 struct backref_edge *edge;
2871 struct backref_edge *edges[BTRFS_MAX_LEVEL - 1];
2872 int index = 0;
2873
2874 while (next) {
2875 cond_resched();
2876 while (1) {
2877 if (next->processed)
2878 break;
2879
2880 __mark_block_processed(rc, next);
2881
2882 if (list_empty(&next->upper))
2883 break;
2884
2885 edge = list_entry(next->upper.next,
2886 struct backref_edge, list[LOWER]);
2887 edges[index++] = edge;
2888 next = edge->node[UPPER];
2889 }
2890 next = walk_down_backref(edges, &index);
2891 }
2892 }
2893
2894 static int tree_block_processed(u64 bytenr, struct reloc_control *rc)
2895 {
2896 u32 blocksize = rc->extent_root->nodesize;
2897
2898 if (test_range_bit(&rc->processed_blocks, bytenr,
2899 bytenr + blocksize - 1, EXTENT_DIRTY, 1, NULL))
2900 return 1;
2901 return 0;
2902 }
2903
2904 static int get_tree_block_key(struct reloc_control *rc,
2905 struct tree_block *block)
2906 {
2907 struct extent_buffer *eb;
2908
2909 BUG_ON(block->key_ready);
2910 eb = read_tree_block(rc->extent_root, block->bytenr,
2911 block->key.offset);
2912 if (IS_ERR(eb)) {
2913 return PTR_ERR(eb);
2914 } else if (!extent_buffer_uptodate(eb)) {
2915 free_extent_buffer(eb);
2916 return -EIO;
2917 }
2918 WARN_ON(btrfs_header_level(eb) != block->level);
2919 if (block->level == 0)
2920 btrfs_item_key_to_cpu(eb, &block->key, 0);
2921 else
2922 btrfs_node_key_to_cpu(eb, &block->key, 0);
2923 free_extent_buffer(eb);
2924 block->key_ready = 1;
2925 return 0;
2926 }
2927
2928 /*
2929 * helper function to relocate a tree block
2930 */
2931 static int relocate_tree_block(struct btrfs_trans_handle *trans,
2932 struct reloc_control *rc,
2933 struct backref_node *node,
2934 struct btrfs_key *key,
2935 struct btrfs_path *path)
2936 {
2937 struct btrfs_root *root;
2938 int ret = 0;
2939
2940 if (!node)
2941 return 0;
2942
2943 BUG_ON(node->processed);
2944 root = select_one_root(node);
2945 if (root == ERR_PTR(-ENOENT)) {
2946 update_processed_blocks(rc, node);
2947 goto out;
2948 }
2949
2950 if (!root || test_bit(BTRFS_ROOT_REF_COWS, &root->state)) {
2951 ret = reserve_metadata_space(trans, rc, node);
2952 if (ret)
2953 goto out;
2954 }
2955
2956 if (root) {
2957 if (test_bit(BTRFS_ROOT_REF_COWS, &root->state)) {
2958 BUG_ON(node->new_bytenr);
2959 BUG_ON(!list_empty(&node->list));
2960 btrfs_record_root_in_trans(trans, root);
2961 root = root->reloc_root;
2962 node->new_bytenr = root->node->start;
2963 node->root = root;
2964 list_add_tail(&node->list, &rc->backref_cache.changed);
2965 } else {
2966 path->lowest_level = node->level;
2967 ret = btrfs_search_slot(trans, root, key, path, 0, 1);
2968 btrfs_release_path(path);
2969 if (ret > 0)
2970 ret = 0;
2971 }
2972 if (!ret)
2973 update_processed_blocks(rc, node);
2974 } else {
2975 ret = do_relocation(trans, rc, node, key, path, 1);
2976 }
2977 out:
2978 if (ret || node->level == 0 || node->cowonly)
2979 remove_backref_node(&rc->backref_cache, node);
2980 return ret;
2981 }
2982
2983 /*
2984 * relocate a list of blocks
2985 */
2986 static noinline_for_stack
2987 int relocate_tree_blocks(struct btrfs_trans_handle *trans,
2988 struct reloc_control *rc, struct rb_root *blocks)
2989 {
2990 struct backref_node *node;
2991 struct btrfs_path *path;
2992 struct tree_block *block;
2993 struct rb_node *rb_node;
2994 int ret;
2995 int err = 0;
2996
2997 path = btrfs_alloc_path();
2998 if (!path) {
2999 err = -ENOMEM;
3000 goto out_free_blocks;
3001 }
3002
3003 rb_node = rb_first(blocks);
3004 while (rb_node) {
3005 block = rb_entry(rb_node, struct tree_block, rb_node);
3006 if (!block->key_ready)
3007 readahead_tree_block(rc->extent_root, block->bytenr);
3008 rb_node = rb_next(rb_node);
3009 }
3010
3011 rb_node = rb_first(blocks);
3012 while (rb_node) {
3013 block = rb_entry(rb_node, struct tree_block, rb_node);
3014 if (!block->key_ready) {
3015 err = get_tree_block_key(rc, block);
3016 if (err)
3017 goto out_free_path;
3018 }
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
3026 node = build_backref_tree(rc, &block->key,
3027 block->level, block->bytenr);
3028 if (IS_ERR(node)) {
3029 err = PTR_ERR(node);
3030 goto out;
3031 }
3032
3033 ret = relocate_tree_block(trans, rc, node, &block->key,
3034 path);
3035 if (ret < 0) {
3036 if (ret != -EAGAIN || rb_node == rb_first(blocks))
3037 err = ret;
3038 goto out;
3039 }
3040 rb_node = rb_next(rb_node);
3041 }
3042 out:
3043 err = finish_pending_nodes(trans, rc, path, err);
3044
3045 out_free_path:
3046 btrfs_free_path(path);
3047 out_free_blocks:
3048 free_block_list(blocks);
3049 return err;
3050 }
3051
3052 static noinline_for_stack
3053 int prealloc_file_extent_cluster(struct inode *inode,
3054 struct file_extent_cluster *cluster)
3055 {
3056 u64 alloc_hint = 0;
3057 u64 start;
3058 u64 end;
3059 u64 offset = BTRFS_I(inode)->index_cnt;
3060 u64 num_bytes;
3061 int nr = 0;
3062 int ret = 0;
3063
3064 BUG_ON(cluster->start != cluster->boundary[0]);
3065 mutex_lock(&inode->i_mutex);
3066
3067 ret = btrfs_check_data_free_space(inode, cluster->start,
3068 cluster->end + 1 - cluster->start);
3069 if (ret)
3070 goto out;
3071
3072 while (nr < cluster->nr) {
3073 start = cluster->boundary[nr] - offset;
3074 if (nr + 1 < cluster->nr)
3075 end = cluster->boundary[nr + 1] - 1 - offset;
3076 else
3077 end = cluster->end - offset;
3078
3079 lock_extent(&BTRFS_I(inode)->io_tree, start, end);
3080 num_bytes = end + 1 - start;
3081 ret = btrfs_prealloc_file_range(inode, 0, start,
3082 num_bytes, num_bytes,
3083 end + 1, &alloc_hint);
3084 unlock_extent(&BTRFS_I(inode)->io_tree, start, end);
3085 if (ret)
3086 break;
3087 nr++;
3088 }
3089 btrfs_free_reserved_data_space(inode, cluster->start,
3090 cluster->end + 1 - cluster->start);
3091 out:
3092 mutex_unlock(&inode->i_mutex);
3093 return ret;
3094 }
3095
3096 static noinline_for_stack
3097 int setup_extent_mapping(struct inode *inode, u64 start, u64 end,
3098 u64 block_start)
3099 {
3100 struct btrfs_root *root = BTRFS_I(inode)->root;
3101 struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
3102 struct extent_map *em;
3103 int ret = 0;
3104
3105 em = alloc_extent_map();
3106 if (!em)
3107 return -ENOMEM;
3108
3109 em->start = start;
3110 em->len = end + 1 - start;
3111 em->block_len = em->len;
3112 em->block_start = block_start;
3113 em->bdev = root->fs_info->fs_devices->latest_bdev;
3114 set_bit(EXTENT_FLAG_PINNED, &em->flags);
3115
3116 lock_extent(&BTRFS_I(inode)->io_tree, start, end);
3117 while (1) {
3118 write_lock(&em_tree->lock);
3119 ret = add_extent_mapping(em_tree, em, 0);
3120 write_unlock(&em_tree->lock);
3121 if (ret != -EEXIST) {
3122 free_extent_map(em);
3123 break;
3124 }
3125 btrfs_drop_extent_cache(inode, start, end, 0);
3126 }
3127 unlock_extent(&BTRFS_I(inode)->io_tree, start, end);
3128 return ret;
3129 }
3130
3131 static int relocate_file_extent_cluster(struct inode *inode,
3132 struct file_extent_cluster *cluster)
3133 {
3134 u64 page_start;
3135 u64 page_end;
3136 u64 offset = BTRFS_I(inode)->index_cnt;
3137 unsigned long index;
3138 unsigned long last_index;
3139 struct page *page;
3140 struct file_ra_state *ra;
3141 gfp_t mask = btrfs_alloc_write_mask(inode->i_mapping);
3142 int nr = 0;
3143 int ret = 0;
3144
3145 if (!cluster->nr)
3146 return 0;
3147
3148 ra = kzalloc(sizeof(*ra), GFP_NOFS);
3149 if (!ra)
3150 return -ENOMEM;
3151
3152 ret = prealloc_file_extent_cluster(inode, cluster);
3153 if (ret)
3154 goto out;
3155
3156 file_ra_state_init(ra, inode->i_mapping);
3157
3158 ret = setup_extent_mapping(inode, cluster->start - offset,
3159 cluster->end - offset, cluster->start);
3160 if (ret)
3161 goto out;
3162
3163 index = (cluster->start - offset) >> PAGE_CACHE_SHIFT;
3164 last_index = (cluster->end - offset) >> PAGE_CACHE_SHIFT;
3165 while (index <= last_index) {
3166 ret = btrfs_delalloc_reserve_metadata(inode, PAGE_CACHE_SIZE);
3167 if (ret)
3168 goto out;
3169
3170 page = find_lock_page(inode->i_mapping, index);
3171 if (!page) {
3172 page_cache_sync_readahead(inode->i_mapping,
3173 ra, NULL, index,
3174 last_index + 1 - index);
3175 page = find_or_create_page(inode->i_mapping, index,
3176 mask);
3177 if (!page) {
3178 btrfs_delalloc_release_metadata(inode,
3179 PAGE_CACHE_SIZE);
3180 ret = -ENOMEM;
3181 goto out;
3182 }
3183 }
3184
3185 if (PageReadahead(page)) {
3186 page_cache_async_readahead(inode->i_mapping,
3187 ra, NULL, page, index,
3188 last_index + 1 - index);
3189 }
3190
3191 if (!PageUptodate(page)) {
3192 btrfs_readpage(NULL, page);
3193 lock_page(page);
3194 if (!PageUptodate(page)) {
3195 unlock_page(page);
3196 page_cache_release(page);
3197 btrfs_delalloc_release_metadata(inode,
3198 PAGE_CACHE_SIZE);
3199 ret = -EIO;
3200 goto out;
3201 }
3202 }
3203
3204 page_start = page_offset(page);
3205 page_end = page_start + PAGE_CACHE_SIZE - 1;
3206
3207 lock_extent(&BTRFS_I(inode)->io_tree, page_start, page_end);
3208
3209 set_page_extent_mapped(page);
3210
3211 if (nr < cluster->nr &&
3212 page_start + offset == cluster->boundary[nr]) {
3213 set_extent_bits(&BTRFS_I(inode)->io_tree,
3214 page_start, page_end,
3215 EXTENT_BOUNDARY, GFP_NOFS);
3216 nr++;
3217 }
3218
3219 btrfs_set_extent_delalloc(inode, page_start, page_end, NULL);
3220 set_page_dirty(page);
3221
3222 unlock_extent(&BTRFS_I(inode)->io_tree,
3223 page_start, page_end);
3224 unlock_page(page);
3225 page_cache_release(page);
3226
3227 index++;
3228 balance_dirty_pages_ratelimited(inode->i_mapping);
3229 btrfs_throttle(BTRFS_I(inode)->root);
3230 }
3231 WARN_ON(nr != cluster->nr);
3232 out:
3233 kfree(ra);
3234 return ret;
3235 }
3236
3237 static noinline_for_stack
3238 int relocate_data_extent(struct inode *inode, struct btrfs_key *extent_key,
3239 struct file_extent_cluster *cluster)
3240 {
3241 int ret;
3242
3243 if (cluster->nr > 0 && extent_key->objectid != cluster->end + 1) {
3244 ret = relocate_file_extent_cluster(inode, cluster);
3245 if (ret)
3246 return ret;
3247 cluster->nr = 0;
3248 }
3249
3250 if (!cluster->nr)
3251 cluster->start = extent_key->objectid;
3252 else
3253 BUG_ON(cluster->nr >= MAX_EXTENTS);
3254 cluster->end = extent_key->objectid + extent_key->offset - 1;
3255 cluster->boundary[cluster->nr] = extent_key->objectid;
3256 cluster->nr++;
3257
3258 if (cluster->nr >= MAX_EXTENTS) {
3259 ret = relocate_file_extent_cluster(inode, cluster);
3260 if (ret)
3261 return ret;
3262 cluster->nr = 0;
3263 }
3264 return 0;
3265 }
3266
3267 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
3268 static int get_ref_objectid_v0(struct reloc_control *rc,
3269 struct btrfs_path *path,
3270 struct btrfs_key *extent_key,
3271 u64 *ref_objectid, int *path_change)
3272 {
3273 struct btrfs_key key;
3274 struct extent_buffer *leaf;
3275 struct btrfs_extent_ref_v0 *ref0;
3276 int ret;
3277 int slot;
3278
3279 leaf = path->nodes[0];
3280 slot = path->slots[0];
3281 while (1) {
3282 if (slot >= btrfs_header_nritems(leaf)) {
3283 ret = btrfs_next_leaf(rc->extent_root, path);
3284 if (ret < 0)
3285 return ret;
3286 BUG_ON(ret > 0);
3287 leaf = path->nodes[0];
3288 slot = path->slots[0];
3289 if (path_change)
3290 *path_change = 1;
3291 }
3292 btrfs_item_key_to_cpu(leaf, &key, slot);
3293 if (key.objectid != extent_key->objectid)
3294 return -ENOENT;
3295
3296 if (key.type != BTRFS_EXTENT_REF_V0_KEY) {
3297 slot++;
3298 continue;
3299 }
3300 ref0 = btrfs_item_ptr(leaf, slot,
3301 struct btrfs_extent_ref_v0);
3302 *ref_objectid = btrfs_ref_objectid_v0(leaf, ref0);
3303 break;
3304 }
3305 return 0;
3306 }
3307 #endif
3308
3309 /*
3310 * helper to add a tree block to the list.
3311 * the major work is getting the generation and level of the block
3312 */
3313 static int add_tree_block(struct reloc_control *rc,
3314 struct btrfs_key *extent_key,
3315 struct btrfs_path *path,
3316 struct rb_root *blocks)
3317 {
3318 struct extent_buffer *eb;
3319 struct btrfs_extent_item *ei;
3320 struct btrfs_tree_block_info *bi;
3321 struct tree_block *block;
3322 struct rb_node *rb_node;
3323 u32 item_size;
3324 int level = -1;
3325 u64 generation;
3326
3327 eb = path->nodes[0];
3328 item_size = btrfs_item_size_nr(eb, path->slots[0]);
3329
3330 if (extent_key->type == BTRFS_METADATA_ITEM_KEY ||
3331 item_size >= sizeof(*ei) + sizeof(*bi)) {
3332 ei = btrfs_item_ptr(eb, path->slots[0],
3333 struct btrfs_extent_item);
3334 if (extent_key->type == BTRFS_EXTENT_ITEM_KEY) {
3335 bi = (struct btrfs_tree_block_info *)(ei + 1);
3336 level = btrfs_tree_block_level(eb, bi);
3337 } else {
3338 level = (int)extent_key->offset;
3339 }
3340 generation = btrfs_extent_generation(eb, ei);
3341 } else {
3342 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
3343 u64 ref_owner;
3344 int ret;
3345
3346 BUG_ON(item_size != sizeof(struct btrfs_extent_item_v0));
3347 ret = get_ref_objectid_v0(rc, path, extent_key,
3348 &ref_owner, NULL);
3349 if (ret < 0)
3350 return ret;
3351 BUG_ON(ref_owner >= BTRFS_MAX_LEVEL);
3352 level = (int)ref_owner;
3353 /* FIXME: get real generation */
3354 generation = 0;
3355 #else
3356 BUG();
3357 #endif
3358 }
3359
3360 btrfs_release_path(path);
3361
3362 BUG_ON(level == -1);
3363
3364 block = kmalloc(sizeof(*block), GFP_NOFS);
3365 if (!block)
3366 return -ENOMEM;
3367
3368 block->bytenr = extent_key->objectid;
3369 block->key.objectid = rc->extent_root->nodesize;
3370 block->key.offset = generation;
3371 block->level = level;
3372 block->key_ready = 0;
3373
3374 rb_node = tree_insert(blocks, block->bytenr, &block->rb_node);
3375 if (rb_node)
3376 backref_tree_panic(rb_node, -EEXIST, block->bytenr);
3377
3378 return 0;
3379 }
3380
3381 /*
3382 * helper to add tree blocks for backref of type BTRFS_SHARED_DATA_REF_KEY
3383 */
3384 static int __add_tree_block(struct reloc_control *rc,
3385 u64 bytenr, u32 blocksize,
3386 struct rb_root *blocks)
3387 {
3388 struct btrfs_path *path;
3389 struct btrfs_key key;
3390 int ret;
3391 bool skinny = btrfs_fs_incompat(rc->extent_root->fs_info,
3392 SKINNY_METADATA);
3393
3394 if (tree_block_processed(bytenr, rc))
3395 return 0;
3396
3397 if (tree_search(blocks, bytenr))
3398 return 0;
3399
3400 path = btrfs_alloc_path();
3401 if (!path)
3402 return -ENOMEM;
3403 again:
3404 key.objectid = bytenr;
3405 if (skinny) {
3406 key.type = BTRFS_METADATA_ITEM_KEY;
3407 key.offset = (u64)-1;
3408 } else {
3409 key.type = BTRFS_EXTENT_ITEM_KEY;
3410 key.offset = blocksize;
3411 }
3412
3413 path->search_commit_root = 1;
3414 path->skip_locking = 1;
3415 ret = btrfs_search_slot(NULL, rc->extent_root, &key, path, 0, 0);
3416 if (ret < 0)
3417 goto out;
3418
3419 if (ret > 0 && skinny) {
3420 if (path->slots[0]) {
3421 path->slots[0]--;
3422 btrfs_item_key_to_cpu(path->nodes[0], &key,
3423 path->slots[0]);
3424 if (key.objectid == bytenr &&
3425 (key.type == BTRFS_METADATA_ITEM_KEY ||
3426 (key.type == BTRFS_EXTENT_ITEM_KEY &&
3427 key.offset == blocksize)))
3428 ret = 0;
3429 }
3430
3431 if (ret) {
3432 skinny = false;
3433 btrfs_release_path(path);
3434 goto again;
3435 }
3436 }
3437 BUG_ON(ret);
3438
3439 ret = add_tree_block(rc, &key, path, blocks);
3440 out:
3441 btrfs_free_path(path);
3442 return ret;
3443 }
3444
3445 /*
3446 * helper to check if the block use full backrefs for pointers in it
3447 */
3448 static int block_use_full_backref(struct reloc_control *rc,
3449 struct extent_buffer *eb)
3450 {
3451 u64 flags;
3452 int ret;
3453
3454 if (btrfs_header_flag(eb, BTRFS_HEADER_FLAG_RELOC) ||
3455 btrfs_header_backref_rev(eb) < BTRFS_MIXED_BACKREF_REV)
3456 return 1;
3457
3458 ret = btrfs_lookup_extent_info(NULL, rc->extent_root,
3459 eb->start, btrfs_header_level(eb), 1,
3460 NULL, &flags);
3461 BUG_ON(ret);
3462
3463 if (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF)
3464 ret = 1;
3465 else
3466 ret = 0;
3467 return ret;
3468 }
3469
3470 static int delete_block_group_cache(struct btrfs_fs_info *fs_info,
3471 struct btrfs_block_group_cache *block_group,
3472 struct inode *inode,
3473 u64 ino)
3474 {
3475 struct btrfs_key key;
3476 struct btrfs_root *root = fs_info->tree_root;
3477 struct btrfs_trans_handle *trans;
3478 int ret = 0;
3479
3480 if (inode)
3481 goto truncate;
3482
3483 key.objectid = ino;
3484 key.type = BTRFS_INODE_ITEM_KEY;
3485 key.offset = 0;
3486
3487 inode = btrfs_iget(fs_info->sb, &key, root, NULL);
3488 if (IS_ERR(inode) || is_bad_inode(inode)) {
3489 if (!IS_ERR(inode))
3490 iput(inode);
3491 return -ENOENT;
3492 }
3493
3494 truncate:
3495 ret = btrfs_check_trunc_cache_free_space(root,
3496 &fs_info->global_block_rsv);
3497 if (ret)
3498 goto out;
3499
3500 trans = btrfs_join_transaction(root);
3501 if (IS_ERR(trans)) {
3502 ret = PTR_ERR(trans);
3503 goto out;
3504 }
3505
3506 ret = btrfs_truncate_free_space_cache(root, trans, block_group, inode);
3507
3508 btrfs_end_transaction(trans, root);
3509 btrfs_btree_balance_dirty(root);
3510 out:
3511 iput(inode);
3512 return ret;
3513 }
3514
3515 /*
3516 * helper to add tree blocks for backref of type BTRFS_EXTENT_DATA_REF_KEY
3517 * this function scans fs tree to find blocks reference the data extent
3518 */
3519 static int find_data_references(struct reloc_control *rc,
3520 struct btrfs_key *extent_key,
3521 struct extent_buffer *leaf,
3522 struct btrfs_extent_data_ref *ref,
3523 struct rb_root *blocks)
3524 {
3525 struct btrfs_path *path;
3526 struct tree_block *block;
3527 struct btrfs_root *root;
3528 struct btrfs_file_extent_item *fi;
3529 struct rb_node *rb_node;
3530 struct btrfs_key key;
3531 u64 ref_root;
3532 u64 ref_objectid;
3533 u64 ref_offset;
3534 u32 ref_count;
3535 u32 nritems;
3536 int err = 0;
3537 int added = 0;
3538 int counted;
3539 int ret;
3540
3541 ref_root = btrfs_extent_data_ref_root(leaf, ref);
3542 ref_objectid = btrfs_extent_data_ref_objectid(leaf, ref);
3543 ref_offset = btrfs_extent_data_ref_offset(leaf, ref);
3544 ref_count = btrfs_extent_data_ref_count(leaf, ref);
3545
3546 /*
3547 * This is an extent belonging to the free space cache, lets just delete
3548 * it and redo the search.
3549 */
3550 if (ref_root == BTRFS_ROOT_TREE_OBJECTID) {
3551 ret = delete_block_group_cache(rc->extent_root->fs_info,
3552 rc->block_group,
3553 NULL, ref_objectid);
3554 if (ret != -ENOENT)
3555 return ret;
3556 ret = 0;
3557 }
3558
3559 path = btrfs_alloc_path();
3560 if (!path)
3561 return -ENOMEM;
3562 path->reada = 1;
3563
3564 root = read_fs_root(rc->extent_root->fs_info, ref_root);
3565 if (IS_ERR(root)) {
3566 err = PTR_ERR(root);
3567 goto out;
3568 }
3569
3570 key.objectid = ref_objectid;
3571 key.type = BTRFS_EXTENT_DATA_KEY;
3572 if (ref_offset > ((u64)-1 << 32))
3573 key.offset = 0;
3574 else
3575 key.offset = ref_offset;
3576
3577 path->search_commit_root = 1;
3578 path->skip_locking = 1;
3579 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
3580 if (ret < 0) {
3581 err = ret;
3582 goto out;
3583 }
3584
3585 leaf = path->nodes[0];
3586 nritems = btrfs_header_nritems(leaf);
3587 /*
3588 * the references in tree blocks that use full backrefs
3589 * are not counted in
3590 */
3591 if (block_use_full_backref(rc, leaf))
3592 counted = 0;
3593 else
3594 counted = 1;
3595 rb_node = tree_search(blocks, leaf->start);
3596 if (rb_node) {
3597 if (counted)
3598 added = 1;
3599 else
3600 path->slots[0] = nritems;
3601 }
3602
3603 while (ref_count > 0) {
3604 while (path->slots[0] >= nritems) {
3605 ret = btrfs_next_leaf(root, path);
3606 if (ret < 0) {
3607 err = ret;
3608 goto out;
3609 }
3610 if (WARN_ON(ret > 0))
3611 goto out;
3612
3613 leaf = path->nodes[0];
3614 nritems = btrfs_header_nritems(leaf);
3615 added = 0;
3616
3617 if (block_use_full_backref(rc, leaf))
3618 counted = 0;
3619 else
3620 counted = 1;
3621 rb_node = tree_search(blocks, leaf->start);
3622 if (rb_node) {
3623 if (counted)
3624 added = 1;
3625 else
3626 path->slots[0] = nritems;
3627 }
3628 }
3629
3630 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
3631 if (WARN_ON(key.objectid != ref_objectid ||
3632 key.type != BTRFS_EXTENT_DATA_KEY))
3633 break;
3634
3635 fi = btrfs_item_ptr(leaf, path->slots[0],
3636 struct btrfs_file_extent_item);
3637
3638 if (btrfs_file_extent_type(leaf, fi) ==
3639 BTRFS_FILE_EXTENT_INLINE)
3640 goto next;
3641
3642 if (btrfs_file_extent_disk_bytenr(leaf, fi) !=
3643 extent_key->objectid)
3644 goto next;
3645
3646 key.offset -= btrfs_file_extent_offset(leaf, fi);
3647 if (key.offset != ref_offset)
3648 goto next;
3649
3650 if (counted)
3651 ref_count--;
3652 if (added)
3653 goto next;
3654
3655 if (!tree_block_processed(leaf->start, rc)) {
3656 block = kmalloc(sizeof(*block), GFP_NOFS);
3657 if (!block) {
3658 err = -ENOMEM;
3659 break;
3660 }
3661 block->bytenr = leaf->start;
3662 btrfs_item_key_to_cpu(leaf, &block->key, 0);
3663 block->level = 0;
3664 block->key_ready = 1;
3665 rb_node = tree_insert(blocks, block->bytenr,
3666 &block->rb_node);
3667 if (rb_node)
3668 backref_tree_panic(rb_node, -EEXIST,
3669 block->bytenr);
3670 }
3671 if (counted)
3672 added = 1;
3673 else
3674 path->slots[0] = nritems;
3675 next:
3676 path->slots[0]++;
3677
3678 }
3679 out:
3680 btrfs_free_path(path);
3681 return err;
3682 }
3683
3684 /*
3685 * helper to find all tree blocks that reference a given data extent
3686 */
3687 static noinline_for_stack
3688 int add_data_references(struct reloc_control *rc,
3689 struct btrfs_key *extent_key,
3690 struct btrfs_path *path,
3691 struct rb_root *blocks)
3692 {
3693 struct btrfs_key key;
3694 struct extent_buffer *eb;
3695 struct btrfs_extent_data_ref *dref;
3696 struct btrfs_extent_inline_ref *iref;
3697 unsigned long ptr;
3698 unsigned long end;
3699 u32 blocksize = rc->extent_root->nodesize;
3700 int ret = 0;
3701 int err = 0;
3702
3703 eb = path->nodes[0];
3704 ptr = btrfs_item_ptr_offset(eb, path->slots[0]);
3705 end = ptr + btrfs_item_size_nr(eb, path->slots[0]);
3706 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
3707 if (ptr + sizeof(struct btrfs_extent_item_v0) == end)
3708 ptr = end;
3709 else
3710 #endif
3711 ptr += sizeof(struct btrfs_extent_item);
3712
3713 while (ptr < end) {
3714 iref = (struct btrfs_extent_inline_ref *)ptr;
3715 key.type = btrfs_extent_inline_ref_type(eb, iref);
3716 if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
3717 key.offset = btrfs_extent_inline_ref_offset(eb, iref);
3718 ret = __add_tree_block(rc, key.offset, blocksize,
3719 blocks);
3720 } else if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
3721 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
3722 ret = find_data_references(rc, extent_key,
3723 eb, dref, blocks);
3724 } else {
3725 BUG();
3726 }
3727 if (ret) {
3728 err = ret;
3729 goto out;
3730 }
3731 ptr += btrfs_extent_inline_ref_size(key.type);
3732 }
3733 WARN_ON(ptr > end);
3734
3735 while (1) {
3736 cond_resched();
3737 eb = path->nodes[0];
3738 if (path->slots[0] >= btrfs_header_nritems(eb)) {
3739 ret = btrfs_next_leaf(rc->extent_root, path);
3740 if (ret < 0) {
3741 err = ret;
3742 break;
3743 }
3744 if (ret > 0)
3745 break;
3746 eb = path->nodes[0];
3747 }
3748
3749 btrfs_item_key_to_cpu(eb, &key, path->slots[0]);
3750 if (key.objectid != extent_key->objectid)
3751 break;
3752
3753 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
3754 if (key.type == BTRFS_SHARED_DATA_REF_KEY ||
3755 key.type == BTRFS_EXTENT_REF_V0_KEY) {
3756 #else
3757 BUG_ON(key.type == BTRFS_EXTENT_REF_V0_KEY);
3758 if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
3759 #endif
3760 ret = __add_tree_block(rc, key.offset, blocksize,
3761 blocks);
3762 } else if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
3763 dref = btrfs_item_ptr(eb, path->slots[0],
3764 struct btrfs_extent_data_ref);
3765 ret = find_data_references(rc, extent_key,
3766 eb, dref, blocks);
3767 } else {
3768 ret = 0;
3769 }
3770 if (ret) {
3771 err = ret;
3772 break;
3773 }
3774 path->slots[0]++;
3775 }
3776 out:
3777 btrfs_release_path(path);
3778 if (err)
3779 free_block_list(blocks);
3780 return err;
3781 }
3782
3783 /*
3784 * helper to find next unprocessed extent
3785 */
3786 static noinline_for_stack
3787 int find_next_extent(struct reloc_control *rc, struct btrfs_path *path,
3788 struct btrfs_key *extent_key)
3789 {
3790 struct btrfs_key key;
3791 struct extent_buffer *leaf;
3792 u64 start, end, last;
3793 int ret;
3794
3795 last = rc->block_group->key.objectid + rc->block_group->key.offset;
3796 while (1) {
3797 cond_resched();
3798 if (rc->search_start >= last) {
3799 ret = 1;
3800 break;
3801 }
3802
3803 key.objectid = rc->search_start;
3804 key.type = BTRFS_EXTENT_ITEM_KEY;
3805 key.offset = 0;
3806
3807 path->search_commit_root = 1;
3808 path->skip_locking = 1;
3809 ret = btrfs_search_slot(NULL, rc->extent_root, &key, path,
3810 0, 0);
3811 if (ret < 0)
3812 break;
3813 next:
3814 leaf = path->nodes[0];
3815 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
3816 ret = btrfs_next_leaf(rc->extent_root, path);
3817 if (ret != 0)
3818 break;
3819 leaf = path->nodes[0];
3820 }
3821
3822 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
3823 if (key.objectid >= last) {
3824 ret = 1;
3825 break;
3826 }
3827
3828 if (key.type != BTRFS_EXTENT_ITEM_KEY &&
3829 key.type != BTRFS_METADATA_ITEM_KEY) {
3830 path->slots[0]++;
3831 goto next;
3832 }
3833
3834 if (key.type == BTRFS_EXTENT_ITEM_KEY &&
3835 key.objectid + key.offset <= rc->search_start) {
3836 path->slots[0]++;
3837 goto next;
3838 }
3839
3840 if (key.type == BTRFS_METADATA_ITEM_KEY &&
3841 key.objectid + rc->extent_root->nodesize <=
3842 rc->search_start) {
3843 path->slots[0]++;
3844 goto next;
3845 }
3846
3847 ret = find_first_extent_bit(&rc->processed_blocks,
3848 key.objectid, &start, &end,
3849 EXTENT_DIRTY, NULL);
3850
3851 if (ret == 0 && start <= key.objectid) {
3852 btrfs_release_path(path);
3853 rc->search_start = end + 1;
3854 } else {
3855 if (key.type == BTRFS_EXTENT_ITEM_KEY)
3856 rc->search_start = key.objectid + key.offset;
3857 else
3858 rc->search_start = key.objectid +
3859 rc->extent_root->nodesize;
3860 memcpy(extent_key, &key, sizeof(key));
3861 return 0;
3862 }
3863 }
3864 btrfs_release_path(path);
3865 return ret;
3866 }
3867
3868 static void set_reloc_control(struct reloc_control *rc)
3869 {
3870 struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
3871
3872 mutex_lock(&fs_info->reloc_mutex);
3873 fs_info->reloc_ctl = rc;
3874 mutex_unlock(&fs_info->reloc_mutex);
3875 }
3876
3877 static void unset_reloc_control(struct reloc_control *rc)
3878 {
3879 struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
3880
3881 mutex_lock(&fs_info->reloc_mutex);
3882 fs_info->reloc_ctl = NULL;
3883 mutex_unlock(&fs_info->reloc_mutex);
3884 }
3885
3886 static int check_extent_flags(u64 flags)
3887 {
3888 if ((flags & BTRFS_EXTENT_FLAG_DATA) &&
3889 (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK))
3890 return 1;
3891 if (!(flags & BTRFS_EXTENT_FLAG_DATA) &&
3892 !(flags & BTRFS_EXTENT_FLAG_TREE_BLOCK))
3893 return 1;
3894 if ((flags & BTRFS_EXTENT_FLAG_DATA) &&
3895 (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
3896 return 1;
3897 return 0;
3898 }
3899
3900 static noinline_for_stack
3901 int prepare_to_relocate(struct reloc_control *rc)
3902 {
3903 struct btrfs_trans_handle *trans;
3904
3905 rc->block_rsv = btrfs_alloc_block_rsv(rc->extent_root,
3906 BTRFS_BLOCK_RSV_TEMP);
3907 if (!rc->block_rsv)
3908 return -ENOMEM;
3909
3910 memset(&rc->cluster, 0, sizeof(rc->cluster));
3911 rc->search_start = rc->block_group->key.objectid;
3912 rc->extents_found = 0;
3913 rc->nodes_relocated = 0;
3914 rc->merging_rsv_size = 0;
3915 rc->reserved_bytes = 0;
3916 rc->block_rsv->size = rc->extent_root->nodesize *
3917 RELOCATION_RESERVED_NODES;
3918
3919 rc->create_reloc_tree = 1;
3920 set_reloc_control(rc);
3921
3922 trans = btrfs_join_transaction(rc->extent_root);
3923 if (IS_ERR(trans)) {
3924 unset_reloc_control(rc);
3925 /*
3926 * extent tree is not a ref_cow tree and has no reloc_root to
3927 * cleanup. And callers are responsible to free the above
3928 * block rsv.
3929 */
3930 return PTR_ERR(trans);
3931 }
3932 btrfs_commit_transaction(trans, rc->extent_root);
3933 return 0;
3934 }
3935
3936 static noinline_for_stack int relocate_block_group(struct reloc_control *rc)
3937 {
3938 struct rb_root blocks = RB_ROOT;
3939 struct btrfs_key key;
3940 struct btrfs_trans_handle *trans = NULL;
3941 struct btrfs_path *path;
3942 struct btrfs_extent_item *ei;
3943 u64 flags;
3944 u32 item_size;
3945 int ret;
3946 int err = 0;
3947 int progress = 0;
3948
3949 path = btrfs_alloc_path();
3950 if (!path)
3951 return -ENOMEM;
3952 path->reada = 1;
3953
3954 ret = prepare_to_relocate(rc);
3955 if (ret) {
3956 err = ret;
3957 goto out_free;
3958 }
3959
3960 while (1) {
3961 rc->reserved_bytes = 0;
3962 ret = btrfs_block_rsv_refill(rc->extent_root,
3963 rc->block_rsv, rc->block_rsv->size,
3964 BTRFS_RESERVE_FLUSH_ALL);
3965 if (ret) {
3966 err = ret;
3967 break;
3968 }
3969 progress++;
3970 trans = btrfs_start_transaction(rc->extent_root, 0);
3971 if (IS_ERR(trans)) {
3972 err = PTR_ERR(trans);
3973 trans = NULL;
3974 break;
3975 }
3976 restart:
3977 if (update_backref_cache(trans, &rc->backref_cache)) {
3978 btrfs_end_transaction(trans, rc->extent_root);
3979 continue;
3980 }
3981
3982 ret = find_next_extent(rc, path, &key);
3983 if (ret < 0)
3984 err = ret;
3985 if (ret != 0)
3986 break;
3987
3988 rc->extents_found++;
3989
3990 ei = btrfs_item_ptr(path->nodes[0], path->slots[0],
3991 struct btrfs_extent_item);
3992 item_size = btrfs_item_size_nr(path->nodes[0], path->slots[0]);
3993 if (item_size >= sizeof(*ei)) {
3994 flags = btrfs_extent_flags(path->nodes[0], ei);
3995 ret = check_extent_flags(flags);
3996 BUG_ON(ret);
3997
3998 } else {
3999 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
4000 u64 ref_owner;
4001 int path_change = 0;
4002
4003 BUG_ON(item_size !=
4004 sizeof(struct btrfs_extent_item_v0));
4005 ret = get_ref_objectid_v0(rc, path, &key, &ref_owner,
4006 &path_change);
4007 if (ret < 0) {
4008 err = ret;
4009 break;
4010 }
4011 if (ref_owner < BTRFS_FIRST_FREE_OBJECTID)
4012 flags = BTRFS_EXTENT_FLAG_TREE_BLOCK;
4013 else
4014 flags = BTRFS_EXTENT_FLAG_DATA;
4015
4016 if (path_change) {
4017 btrfs_release_path(path);
4018
4019 path->search_commit_root = 1;
4020 path->skip_locking = 1;
4021 ret = btrfs_search_slot(NULL, rc->extent_root,
4022 &key, path, 0, 0);
4023 if (ret < 0) {
4024 err = ret;
4025 break;
4026 }
4027 BUG_ON(ret > 0);
4028 }
4029 #else
4030 BUG();
4031 #endif
4032 }
4033
4034 if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
4035 ret = add_tree_block(rc, &key, path, &blocks);
4036 } else if (rc->stage == UPDATE_DATA_PTRS &&
4037 (flags & BTRFS_EXTENT_FLAG_DATA)) {
4038 ret = add_data_references(rc, &key, path, &blocks);
4039 } else {
4040 btrfs_release_path(path);
4041 ret = 0;
4042 }
4043 if (ret < 0) {
4044 err = ret;
4045 break;
4046 }
4047
4048 if (!RB_EMPTY_ROOT(&blocks)) {
4049 ret = relocate_tree_blocks(trans, rc, &blocks);
4050 if (ret < 0) {
4051 /*
4052 * if we fail to relocate tree blocks, force to update
4053 * backref cache when committing transaction.
4054 */
4055 rc->backref_cache.last_trans = trans->transid - 1;
4056
4057 if (ret != -EAGAIN) {
4058 err = ret;
4059 break;
4060 }
4061 rc->extents_found--;
4062 rc->search_start = key.objectid;
4063 }
4064 }
4065
4066 btrfs_end_transaction_throttle(trans, rc->extent_root);
4067 btrfs_btree_balance_dirty(rc->extent_root);
4068 trans = NULL;
4069
4070 if (rc->stage == MOVE_DATA_EXTENTS &&
4071 (flags & BTRFS_EXTENT_FLAG_DATA)) {
4072 rc->found_file_extent = 1;
4073 ret = relocate_data_extent(rc->data_inode,
4074 &key, &rc->cluster);
4075 if (ret < 0) {
4076 err = ret;
4077 break;
4078 }
4079 }
4080 }
4081 if (trans && progress && err == -ENOSPC) {
4082 ret = btrfs_force_chunk_alloc(trans, rc->extent_root,
4083 rc->block_group->flags);
4084 if (ret == 1) {
4085 err = 0;
4086 progress = 0;
4087 goto restart;
4088 }
4089 }
4090
4091 btrfs_release_path(path);
4092 clear_extent_bits(&rc->processed_blocks, 0, (u64)-1, EXTENT_DIRTY,
4093 GFP_NOFS);
4094
4095 if (trans) {
4096 btrfs_end_transaction_throttle(trans, rc->extent_root);
4097 btrfs_btree_balance_dirty(rc->extent_root);
4098 }
4099
4100 if (!err) {
4101 ret = relocate_file_extent_cluster(rc->data_inode,
4102 &rc->cluster);
4103 if (ret < 0)
4104 err = ret;
4105 }
4106
4107 rc->create_reloc_tree = 0;
4108 set_reloc_control(rc);
4109
4110 backref_cache_cleanup(&rc->backref_cache);
4111 btrfs_block_rsv_release(rc->extent_root, rc->block_rsv, (u64)-1);
4112
4113 err = prepare_to_merge(rc, err);
4114
4115 merge_reloc_roots(rc);
4116
4117 rc->merge_reloc_tree = 0;
4118 unset_reloc_control(rc);
4119 btrfs_block_rsv_release(rc->extent_root, rc->block_rsv, (u64)-1);
4120
4121 /* get rid of pinned extents */
4122 trans = btrfs_join_transaction(rc->extent_root);
4123 if (IS_ERR(trans))
4124 err = PTR_ERR(trans);
4125 else
4126 btrfs_commit_transaction(trans, rc->extent_root);
4127 out_free:
4128 btrfs_free_block_rsv(rc->extent_root, rc->block_rsv);
4129 btrfs_free_path(path);
4130 return err;
4131 }
4132
4133 static int __insert_orphan_inode(struct btrfs_trans_handle *trans,
4134 struct btrfs_root *root, u64 objectid)
4135 {
4136 struct btrfs_path *path;
4137 struct btrfs_inode_item *item;
4138 struct extent_buffer *leaf;
4139 int ret;
4140
4141 path = btrfs_alloc_path();
4142 if (!path)
4143 return -ENOMEM;
4144
4145 ret = btrfs_insert_empty_inode(trans, root, path, objectid);
4146 if (ret)
4147 goto out;
4148
4149 leaf = path->nodes[0];
4150 item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_inode_item);
4151 memset_extent_buffer(leaf, 0, (unsigned long)item, sizeof(*item));
4152 btrfs_set_inode_generation(leaf, item, 1);
4153 btrfs_set_inode_size(leaf, item, 0);
4154 btrfs_set_inode_mode(leaf, item, S_IFREG | 0600);
4155 btrfs_set_inode_flags(leaf, item, BTRFS_INODE_NOCOMPRESS |
4156 BTRFS_INODE_PREALLOC);
4157 btrfs_mark_buffer_dirty(leaf);
4158 out:
4159 btrfs_free_path(path);
4160 return ret;
4161 }
4162
4163 /*
4164 * helper to create inode for data relocation.
4165 * the inode is in data relocation tree and its link count is 0
4166 */
4167 static noinline_for_stack
4168 struct inode *create_reloc_inode(struct btrfs_fs_info *fs_info,
4169 struct btrfs_block_group_cache *group)
4170 {
4171 struct inode *inode = NULL;
4172 struct btrfs_trans_handle *trans;
4173 struct btrfs_root *root;
4174 struct btrfs_key key;
4175 u64 objectid;
4176 int err = 0;
4177
4178 root = read_fs_root(fs_info, BTRFS_DATA_RELOC_TREE_OBJECTID);
4179 if (IS_ERR(root))
4180 return ERR_CAST(root);
4181
4182 trans = btrfs_start_transaction(root, 6);
4183 if (IS_ERR(trans))
4184 return ERR_CAST(trans);
4185
4186 err = btrfs_find_free_objectid(root, &objectid);
4187 if (err)
4188 goto out;
4189
4190 err = __insert_orphan_inode(trans, root, objectid);
4191 BUG_ON(err);
4192
4193 key.objectid = objectid;
4194 key.type = BTRFS_INODE_ITEM_KEY;
4195 key.offset = 0;
4196 inode = btrfs_iget(root->fs_info->sb, &key, root, NULL);
4197 BUG_ON(IS_ERR(inode) || is_bad_inode(inode));
4198 BTRFS_I(inode)->index_cnt = group->key.objectid;
4199
4200 err = btrfs_orphan_add(trans, inode);
4201 out:
4202 btrfs_end_transaction(trans, root);
4203 btrfs_btree_balance_dirty(root);
4204 if (err) {
4205 if (inode)
4206 iput(inode);
4207 inode = ERR_PTR(err);
4208 }
4209 return inode;
4210 }
4211
4212 static struct reloc_control *alloc_reloc_control(struct btrfs_fs_info *fs_info)
4213 {
4214 struct reloc_control *rc;
4215
4216 rc = kzalloc(sizeof(*rc), GFP_NOFS);
4217 if (!rc)
4218 return NULL;
4219
4220 INIT_LIST_HEAD(&rc->reloc_roots);
4221 backref_cache_init(&rc->backref_cache);
4222 mapping_tree_init(&rc->reloc_root_tree);
4223 extent_io_tree_init(&rc->processed_blocks,
4224 fs_info->btree_inode->i_mapping);
4225 return rc;
4226 }
4227
4228 /*
4229 * function to relocate all extents in a block group.
4230 */
4231 int btrfs_relocate_block_group(struct btrfs_root *extent_root, u64 group_start)
4232 {
4233 struct btrfs_fs_info *fs_info = extent_root->fs_info;
4234 struct reloc_control *rc;
4235 struct inode *inode;
4236 struct btrfs_path *path;
4237 int ret;
4238 int rw = 0;
4239 int err = 0;
4240
4241 rc = alloc_reloc_control(fs_info);
4242 if (!rc)
4243 return -ENOMEM;
4244
4245 rc->extent_root = extent_root;
4246
4247 rc->block_group = btrfs_lookup_block_group(fs_info, group_start);
4248 BUG_ON(!rc->block_group);
4249
4250 ret = btrfs_inc_block_group_ro(extent_root, rc->block_group);
4251 if (ret) {
4252 err = ret;
4253 goto out;
4254 }
4255 rw = 1;
4256
4257 path = btrfs_alloc_path();
4258 if (!path) {
4259 err = -ENOMEM;
4260 goto out;
4261 }
4262
4263 inode = lookup_free_space_inode(fs_info->tree_root, rc->block_group,
4264 path);
4265 btrfs_free_path(path);
4266
4267 if (!IS_ERR(inode))
4268 ret = delete_block_group_cache(fs_info, rc->block_group, inode, 0);
4269 else
4270 ret = PTR_ERR(inode);
4271
4272 if (ret && ret != -ENOENT) {
4273 err = ret;
4274 goto out;
4275 }
4276
4277 rc->data_inode = create_reloc_inode(fs_info, rc->block_group);
4278 if (IS_ERR(rc->data_inode)) {
4279 err = PTR_ERR(rc->data_inode);
4280 rc->data_inode = NULL;
4281 goto out;
4282 }
4283
4284 btrfs_info(extent_root->fs_info, "relocating block group %llu flags %llu",
4285 rc->block_group->key.objectid, rc->block_group->flags);
4286
4287 ret = btrfs_start_delalloc_roots(fs_info, 0, -1);
4288 if (ret < 0) {
4289 err = ret;
4290 goto out;
4291 }
4292 btrfs_wait_ordered_roots(fs_info, -1);
4293
4294 while (1) {
4295 mutex_lock(&fs_info->cleaner_mutex);
4296 ret = relocate_block_group(rc);
4297 mutex_unlock(&fs_info->cleaner_mutex);
4298 if (ret < 0) {
4299 err = ret;
4300 goto out;
4301 }
4302
4303 if (rc->extents_found == 0)
4304 break;
4305
4306 btrfs_info(extent_root->fs_info, "found %llu extents",
4307 rc->extents_found);
4308
4309 if (rc->stage == MOVE_DATA_EXTENTS && rc->found_file_extent) {
4310 ret = btrfs_wait_ordered_range(rc->data_inode, 0,
4311 (u64)-1);
4312 if (ret) {
4313 err = ret;
4314 goto out;
4315 }
4316 invalidate_mapping_pages(rc->data_inode->i_mapping,
4317 0, -1);
4318 rc->stage = UPDATE_DATA_PTRS;
4319 }
4320 }
4321
4322 WARN_ON(rc->block_group->pinned > 0);
4323 WARN_ON(rc->block_group->reserved > 0);
4324 WARN_ON(btrfs_block_group_used(&rc->block_group->item) > 0);
4325 out:
4326 if (err && rw)
4327 btrfs_dec_block_group_ro(extent_root, rc->block_group);
4328 iput(rc->data_inode);
4329 btrfs_put_block_group(rc->block_group);
4330 kfree(rc);
4331 return err;
4332 }
4333
4334 static noinline_for_stack int mark_garbage_root(struct btrfs_root *root)
4335 {
4336 struct btrfs_trans_handle *trans;
4337 int ret, err;
4338
4339 trans = btrfs_start_transaction(root->fs_info->tree_root, 0);
4340 if (IS_ERR(trans))
4341 return PTR_ERR(trans);
4342
4343 memset(&root->root_item.drop_progress, 0,
4344 sizeof(root->root_item.drop_progress));
4345 root->root_item.drop_level = 0;
4346 btrfs_set_root_refs(&root->root_item, 0);
4347 ret = btrfs_update_root(trans, root->fs_info->tree_root,
4348 &root->root_key, &root->root_item);
4349
4350 err = btrfs_end_transaction(trans, root->fs_info->tree_root);
4351 if (err)
4352 return err;
4353 return ret;
4354 }
4355
4356 /*
4357 * recover relocation interrupted by system crash.
4358 *
4359 * this function resumes merging reloc trees with corresponding fs trees.
4360 * this is important for keeping the sharing of tree blocks
4361 */
4362 int btrfs_recover_relocation(struct btrfs_root *root)
4363 {
4364 LIST_HEAD(reloc_roots);
4365 struct btrfs_key key;
4366 struct btrfs_root *fs_root;
4367 struct btrfs_root *reloc_root;
4368 struct btrfs_path *path;
4369 struct extent_buffer *leaf;
4370 struct reloc_control *rc = NULL;
4371 struct btrfs_trans_handle *trans;
4372 int ret;
4373 int err = 0;
4374
4375 path = btrfs_alloc_path();
4376 if (!path)
4377 return -ENOMEM;
4378 path->reada = -1;
4379
4380 key.objectid = BTRFS_TREE_RELOC_OBJECTID;
4381 key.type = BTRFS_ROOT_ITEM_KEY;
4382 key.offset = (u64)-1;
4383
4384 while (1) {
4385 ret = btrfs_search_slot(NULL, root->fs_info->tree_root, &key,
4386 path, 0, 0);
4387 if (ret < 0) {
4388 err = ret;
4389 goto out;
4390 }
4391 if (ret > 0) {
4392 if (path->slots[0] == 0)
4393 break;
4394 path->slots[0]--;
4395 }
4396 leaf = path->nodes[0];
4397 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
4398 btrfs_release_path(path);
4399
4400 if (key.objectid != BTRFS_TREE_RELOC_OBJECTID ||
4401 key.type != BTRFS_ROOT_ITEM_KEY)
4402 break;
4403
4404 reloc_root = btrfs_read_fs_root(root, &key);
4405 if (IS_ERR(reloc_root)) {
4406 err = PTR_ERR(reloc_root);
4407 goto out;
4408 }
4409
4410 list_add(&reloc_root->root_list, &reloc_roots);
4411
4412 if (btrfs_root_refs(&reloc_root->root_item) > 0) {
4413 fs_root = read_fs_root(root->fs_info,
4414 reloc_root->root_key.offset);
4415 if (IS_ERR(fs_root)) {
4416 ret = PTR_ERR(fs_root);
4417 if (ret != -ENOENT) {
4418 err = ret;
4419 goto out;
4420 }
4421 ret = mark_garbage_root(reloc_root);
4422 if (ret < 0) {
4423 err = ret;
4424 goto out;
4425 }
4426 }
4427 }
4428
4429 if (key.offset == 0)
4430 break;
4431
4432 key.offset--;
4433 }
4434 btrfs_release_path(path);
4435
4436 if (list_empty(&reloc_roots))
4437 goto out;
4438
4439 rc = alloc_reloc_control(root->fs_info);
4440 if (!rc) {
4441 err = -ENOMEM;
4442 goto out;
4443 }
4444
4445 rc->extent_root = root->fs_info->extent_root;
4446
4447 set_reloc_control(rc);
4448
4449 trans = btrfs_join_transaction(rc->extent_root);
4450 if (IS_ERR(trans)) {
4451 unset_reloc_control(rc);
4452 err = PTR_ERR(trans);
4453 goto out_free;
4454 }
4455
4456 rc->merge_reloc_tree = 1;
4457
4458 while (!list_empty(&reloc_roots)) {
4459 reloc_root = list_entry(reloc_roots.next,
4460 struct btrfs_root, root_list);
4461 list_del(&reloc_root->root_list);
4462
4463 if (btrfs_root_refs(&reloc_root->root_item) == 0) {
4464 list_add_tail(&reloc_root->root_list,
4465 &rc->reloc_roots);
4466 continue;
4467 }
4468
4469 fs_root = read_fs_root(root->fs_info,
4470 reloc_root->root_key.offset);
4471 if (IS_ERR(fs_root)) {
4472 err = PTR_ERR(fs_root);
4473 list_add_tail(&reloc_root->root_list, &reloc_roots);
4474 goto out_free;
4475 }
4476
4477 err = __add_reloc_root(reloc_root);
4478 BUG_ON(err < 0); /* -ENOMEM or logic error */
4479 fs_root->reloc_root = reloc_root;
4480 }
4481
4482 err = btrfs_commit_transaction(trans, rc->extent_root);
4483 if (err)
4484 goto out_free;
4485
4486 merge_reloc_roots(rc);
4487
4488 unset_reloc_control(rc);
4489
4490 trans = btrfs_join_transaction(rc->extent_root);
4491 if (IS_ERR(trans))
4492 err = PTR_ERR(trans);
4493 else
4494 err = btrfs_commit_transaction(trans, rc->extent_root);
4495 out_free:
4496 kfree(rc);
4497 out:
4498 if (!list_empty(&reloc_roots))
4499 free_reloc_roots(&reloc_roots);
4500
4501 btrfs_free_path(path);
4502
4503 if (err == 0) {
4504 /* cleanup orphan inode in data relocation tree */
4505 fs_root = read_fs_root(root->fs_info,
4506 BTRFS_DATA_RELOC_TREE_OBJECTID);
4507 if (IS_ERR(fs_root))
4508 err = PTR_ERR(fs_root);
4509 else
4510 err = btrfs_orphan_cleanup(fs_root);
4511 }
4512 return err;
4513 }
4514
4515 /*
4516 * helper to add ordered checksum for data relocation.
4517 *
4518 * cloning checksum properly handles the nodatasum extents.
4519 * it also saves CPU time to re-calculate the checksum.
4520 */
4521 int btrfs_reloc_clone_csums(struct inode *inode, u64 file_pos, u64 len)
4522 {
4523 struct btrfs_ordered_sum *sums;
4524 struct btrfs_ordered_extent *ordered;
4525 struct btrfs_root *root = BTRFS_I(inode)->root;
4526 int ret;
4527 u64 disk_bytenr;
4528 u64 new_bytenr;
4529 LIST_HEAD(list);
4530
4531 ordered = btrfs_lookup_ordered_extent(inode, file_pos);
4532 BUG_ON(ordered->file_offset != file_pos || ordered->len != len);
4533
4534 disk_bytenr = file_pos + BTRFS_I(inode)->index_cnt;
4535 ret = btrfs_lookup_csums_range(root->fs_info->csum_root, disk_bytenr,
4536 disk_bytenr + len - 1, &list, 0);
4537 if (ret)
4538 goto out;
4539
4540 while (!list_empty(&list)) {
4541 sums = list_entry(list.next, struct btrfs_ordered_sum, list);
4542 list_del_init(&sums->list);
4543
4544 /*
4545 * We need to offset the new_bytenr based on where the csum is.
4546 * We need to do this because we will read in entire prealloc
4547 * extents but we may have written to say the middle of the
4548 * prealloc extent, so we need to make sure the csum goes with
4549 * the right disk offset.
4550 *
4551 * We can do this because the data reloc inode refers strictly
4552 * to the on disk bytes, so we don't have to worry about
4553 * disk_len vs real len like with real inodes since it's all
4554 * disk length.
4555 */
4556 new_bytenr = ordered->start + (sums->bytenr - disk_bytenr);
4557 sums->bytenr = new_bytenr;
4558
4559 btrfs_add_ordered_sum(inode, ordered, sums);
4560 }
4561 out:
4562 btrfs_put_ordered_extent(ordered);
4563 return ret;
4564 }
4565
4566 int btrfs_reloc_cow_block(struct btrfs_trans_handle *trans,
4567 struct btrfs_root *root, struct extent_buffer *buf,
4568 struct extent_buffer *cow)
4569 {
4570 struct reloc_control *rc;
4571 struct backref_node *node;
4572 int first_cow = 0;
4573 int level;
4574 int ret = 0;
4575
4576 rc = root->fs_info->reloc_ctl;
4577 if (!rc)
4578 return 0;
4579
4580 BUG_ON(rc->stage == UPDATE_DATA_PTRS &&
4581 root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID);
4582
4583 level = btrfs_header_level(buf);
4584 if (btrfs_header_generation(buf) <=
4585 btrfs_root_last_snapshot(&root->root_item))
4586 first_cow = 1;
4587
4588 if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID &&
4589 rc->create_reloc_tree) {
4590 WARN_ON(!first_cow && level == 0);
4591
4592 node = rc->backref_cache.path[level];
4593 BUG_ON(node->bytenr != buf->start &&
4594 node->new_bytenr != buf->start);
4595
4596 drop_node_buffer(node);
4597 extent_buffer_get(cow);
4598 node->eb = cow;
4599 node->new_bytenr = cow->start;
4600
4601 if (!node->pending) {
4602 list_move_tail(&node->list,
4603 &rc->backref_cache.pending[level]);
4604 node->pending = 1;
4605 }
4606
4607 if (first_cow)
4608 __mark_block_processed(rc, node);
4609
4610 if (first_cow && level > 0)
4611 rc->nodes_relocated += buf->len;
4612 }
4613
4614 if (level == 0 && first_cow && rc->stage == UPDATE_DATA_PTRS)
4615 ret = replace_file_extents(trans, rc, root, cow);
4616 return ret;
4617 }
4618
4619 /*
4620 * called before creating snapshot. it calculates metadata reservation
4621 * requried for relocating tree blocks in the snapshot
4622 */
4623 void btrfs_reloc_pre_snapshot(struct btrfs_pending_snapshot *pending,
4624 u64 *bytes_to_reserve)
4625 {
4626 struct btrfs_root *root;
4627 struct reloc_control *rc;
4628
4629 root = pending->root;
4630 if (!root->reloc_root)
4631 return;
4632
4633 rc = root->fs_info->reloc_ctl;
4634 if (!rc->merge_reloc_tree)
4635 return;
4636
4637 root = root->reloc_root;
4638 BUG_ON(btrfs_root_refs(&root->root_item) == 0);
4639 /*
4640 * relocation is in the stage of merging trees. the space
4641 * used by merging a reloc tree is twice the size of
4642 * relocated tree nodes in the worst case. half for cowing
4643 * the reloc tree, half for cowing the fs tree. the space
4644 * used by cowing the reloc tree will be freed after the
4645 * tree is dropped. if we create snapshot, cowing the fs
4646 * tree may use more space than it frees. so we need
4647 * reserve extra space.
4648 */
4649 *bytes_to_reserve += rc->nodes_relocated;
4650 }
4651
4652 /*
4653 * called after snapshot is created. migrate block reservation
4654 * and create reloc root for the newly created snapshot
4655 */
4656 int btrfs_reloc_post_snapshot(struct btrfs_trans_handle *trans,
4657 struct btrfs_pending_snapshot *pending)
4658 {
4659 struct btrfs_root *root = pending->root;
4660 struct btrfs_root *reloc_root;
4661 struct btrfs_root *new_root;
4662 struct reloc_control *rc;
4663 int ret;
4664
4665 if (!root->reloc_root)
4666 return 0;
4667
4668 rc = root->fs_info->reloc_ctl;
4669 rc->merging_rsv_size += rc->nodes_relocated;
4670
4671 if (rc->merge_reloc_tree) {
4672 ret = btrfs_block_rsv_migrate(&pending->block_rsv,
4673 rc->block_rsv,
4674 rc->nodes_relocated);
4675 if (ret)
4676 return ret;
4677 }
4678
4679 new_root = pending->snap;
4680 reloc_root = create_reloc_root(trans, root->reloc_root,
4681 new_root->root_key.objectid);
4682 if (IS_ERR(reloc_root))
4683 return PTR_ERR(reloc_root);
4684
4685 ret = __add_reloc_root(reloc_root);
4686 BUG_ON(ret < 0);
4687 new_root->reloc_root = reloc_root;
4688
4689 if (rc->create_reloc_tree)
4690 ret = clone_backref_node(trans, rc, root, reloc_root);
4691 return ret;
4692 }
Linux with added FPC-III support