?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
HID: wacom: Call 'wacom_query_tablet_data' only after 'hid_hw_start'
[linux/fpc-iii.git] / fs / btrfs / relocation.c
blob88cbb5995667951192b94c2627f759a771eda9e3
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 WARN_ON(btrfs_node_blockptr(eb, path2->slots[level]) !=
925 cur->bytenr);
926
927 lower = cur;
928 need_check = true;
929 for (; level < BTRFS_MAX_LEVEL; level++) {
930 if (!path2->nodes[level]) {
931 ASSERT(btrfs_root_bytenr(&root->root_item) ==
932 lower->bytenr);
933 if (should_ignore_root(root))
934 list_add(&lower->list, &useless);
935 else
936 lower->root = root;
937 break;
938 }
939
940 edge = alloc_backref_edge(cache);
941 if (!edge) {
942 err = -ENOMEM;
943 goto out;
944 }
945
946 eb = path2->nodes[level];
947 rb_node = tree_search(&cache->rb_root, eb->start);
948 if (!rb_node) {
949 upper = alloc_backref_node(cache);
950 if (!upper) {
951 free_backref_edge(cache, edge);
952 err = -ENOMEM;
953 goto out;
954 }
955 upper->bytenr = eb->start;
956 upper->owner = btrfs_header_owner(eb);
957 upper->level = lower->level + 1;
958 if (!test_bit(BTRFS_ROOT_REF_COWS,
959 &root->state))
960 upper->cowonly = 1;
961
962 /*
963 * if we know the block isn't shared
964 * we can void checking its backrefs.
965 */
966 if (btrfs_block_can_be_shared(root, eb))
967 upper->checked = 0;
968 else
969 upper->checked = 1;
970
971 /*
972 * add the block to pending list if we
973 * need check its backrefs, we only do this once
974 * while walking up a tree as we will catch
975 * anything else later on.
976 */
977 if (!upper->checked && need_check) {
978 need_check = false;
979 list_add_tail(&edge->list[UPPER],
980 &list);
981 } else {
982 if (upper->checked)
983 need_check = true;
984 INIT_LIST_HEAD(&edge->list[UPPER]);
985 }
986 } else {
987 upper = rb_entry(rb_node, struct backref_node,
988 rb_node);
989 ASSERT(upper->checked);
990 INIT_LIST_HEAD(&edge->list[UPPER]);
991 if (!upper->owner)
992 upper->owner = btrfs_header_owner(eb);
993 }
994 list_add_tail(&edge->list[LOWER], &lower->upper);
995 edge->node[LOWER] = lower;
996 edge->node[UPPER] = upper;
997
998 if (rb_node)
999 break;
1000 lower = upper;
1001 upper = NULL;
1002 }
1003 btrfs_release_path(path2);
1004 next:
1005 if (ptr < end) {
1006 ptr += btrfs_extent_inline_ref_size(key.type);
1007 if (ptr >= end) {
1008 WARN_ON(ptr > end);
1009 ptr = 0;
1010 end = 0;
1011 }
1012 }
1013 if (ptr >= end)
1014 path1->slots[0]++;
1015 }
1016 btrfs_release_path(path1);
1017
1018 cur->checked = 1;
1019 WARN_ON(exist);
1020
1021 /* the pending list isn't empty, take the first block to process */
1022 if (!list_empty(&list)) {
1023 edge = list_entry(list.next, struct backref_edge, list[UPPER]);
1024 list_del_init(&edge->list[UPPER]);
1025 cur = edge->node[UPPER];
1026 goto again;
1027 }
1028
1029 /*
1030 * everything goes well, connect backref nodes and insert backref nodes
1031 * into the cache.
1032 */
1033 ASSERT(node->checked);
1034 cowonly = node->cowonly;
1035 if (!cowonly) {
1036 rb_node = tree_insert(&cache->rb_root, node->bytenr,
1037 &node->rb_node);
1038 if (rb_node)
1039 backref_tree_panic(rb_node, -EEXIST, node->bytenr);
1040 list_add_tail(&node->lower, &cache->leaves);
1041 }
1042
1043 list_for_each_entry(edge, &node->upper, list[LOWER])
1044 list_add_tail(&edge->list[UPPER], &list);
1045
1046 while (!list_empty(&list)) {
1047 edge = list_entry(list.next, struct backref_edge, list[UPPER]);
1048 list_del_init(&edge->list[UPPER]);
1049 upper = edge->node[UPPER];
1050 if (upper->detached) {
1051 list_del(&edge->list[LOWER]);
1052 lower = edge->node[LOWER];
1053 free_backref_edge(cache, edge);
1054 if (list_empty(&lower->upper))
1055 list_add(&lower->list, &useless);
1056 continue;
1057 }
1058
1059 if (!RB_EMPTY_NODE(&upper->rb_node)) {
1060 if (upper->lowest) {
1061 list_del_init(&upper->lower);
1062 upper->lowest = 0;
1063 }
1064
1065 list_add_tail(&edge->list[UPPER], &upper->lower);
1066 continue;
1067 }
1068
1069 if (!upper->checked) {
1070 /*
1071 * Still want to blow up for developers since this is a
1072 * logic bug.
1073 */
1074 ASSERT(0);
1075 err = -EINVAL;
1076 goto out;
1077 }
1078 if (cowonly != upper->cowonly) {
1079 ASSERT(0);
1080 err = -EINVAL;
1081 goto out;
1082 }
1083
1084 if (!cowonly) {
1085 rb_node = tree_insert(&cache->rb_root, upper->bytenr,
1086 &upper->rb_node);
1087 if (rb_node)
1088 backref_tree_panic(rb_node, -EEXIST,
1089 upper->bytenr);
1090 }
1091
1092 list_add_tail(&edge->list[UPPER], &upper->lower);
1093
1094 list_for_each_entry(edge, &upper->upper, list[LOWER])
1095 list_add_tail(&edge->list[UPPER], &list);
1096 }
1097 /*
1098 * process useless backref nodes. backref nodes for tree leaves
1099 * are deleted from the cache. backref nodes for upper level
1100 * tree blocks are left in the cache to avoid unnecessary backref
1101 * lookup.
1102 */
1103 while (!list_empty(&useless)) {
1104 upper = list_entry(useless.next, struct backref_node, list);
1105 list_del_init(&upper->list);
1106 ASSERT(list_empty(&upper->upper));
1107 if (upper == node)
1108 node = NULL;
1109 if (upper->lowest) {
1110 list_del_init(&upper->lower);
1111 upper->lowest = 0;
1112 }
1113 while (!list_empty(&upper->lower)) {
1114 edge = list_entry(upper->lower.next,
1115 struct backref_edge, list[UPPER]);
1116 list_del(&edge->list[UPPER]);
1117 list_del(&edge->list[LOWER]);
1118 lower = edge->node[LOWER];
1119 free_backref_edge(cache, edge);
1120
1121 if (list_empty(&lower->upper))
1122 list_add(&lower->list, &useless);
1123 }
1124 __mark_block_processed(rc, upper);
1125 if (upper->level > 0) {
1126 list_add(&upper->list, &cache->detached);
1127 upper->detached = 1;
1128 } else {
1129 rb_erase(&upper->rb_node, &cache->rb_root);
1130 free_backref_node(cache, upper);
1131 }
1132 }
1133 out:
1134 btrfs_free_path(path1);
1135 btrfs_free_path(path2);
1136 if (err) {
1137 while (!list_empty(&useless)) {
1138 lower = list_entry(useless.next,
1139 struct backref_node, list);
1140 list_del_init(&lower->list);
1141 }
1142 while (!list_empty(&list)) {
1143 edge = list_first_entry(&list, struct backref_edge,
1144 list[UPPER]);
1145 list_del(&edge->list[UPPER]);
1146 list_del(&edge->list[LOWER]);
1147 lower = edge->node[LOWER];
1148 upper = edge->node[UPPER];
1149 free_backref_edge(cache, edge);
1150
1151 /*
1152 * Lower is no longer linked to any upper backref nodes
1153 * and isn't in the cache, we can free it ourselves.
1154 */
1155 if (list_empty(&lower->upper) &&
1156 RB_EMPTY_NODE(&lower->rb_node))
1157 list_add(&lower->list, &useless);
1158
1159 if (!RB_EMPTY_NODE(&upper->rb_node))
1160 continue;
1161
1162 /* Add this guy's upper edges to the list to proces */
1163 list_for_each_entry(edge, &upper->upper, list[LOWER])
1164 list_add_tail(&edge->list[UPPER], &list);
1165 if (list_empty(&upper->upper))
1166 list_add(&upper->list, &useless);
1167 }
1168
1169 while (!list_empty(&useless)) {
1170 lower = list_entry(useless.next,
1171 struct backref_node, list);
1172 list_del_init(&lower->list);
1173 free_backref_node(cache, lower);
1174 }
1175 return ERR_PTR(err);
1176 }
1177 ASSERT(!node || !node->detached);
1178 return node;
1179 }
1180
1181 /*
1182 * helper to add backref node for the newly created snapshot.
1183 * the backref node is created by cloning backref node that
1184 * corresponds to root of source tree
1185 */
1186 static int clone_backref_node(struct btrfs_trans_handle *trans,
1187 struct reloc_control *rc,
1188 struct btrfs_root *src,
1189 struct btrfs_root *dest)
1190 {
1191 struct btrfs_root *reloc_root = src->reloc_root;
1192 struct backref_cache *cache = &rc->backref_cache;
1193 struct backref_node *node = NULL;
1194 struct backref_node *new_node;
1195 struct backref_edge *edge;
1196 struct backref_edge *new_edge;
1197 struct rb_node *rb_node;
1198
1199 if (cache->last_trans > 0)
1200 update_backref_cache(trans, cache);
1201
1202 rb_node = tree_search(&cache->rb_root, src->commit_root->start);
1203 if (rb_node) {
1204 node = rb_entry(rb_node, struct backref_node, rb_node);
1205 if (node->detached)
1206 node = NULL;
1207 else
1208 BUG_ON(node->new_bytenr != reloc_root->node->start);
1209 }
1210
1211 if (!node) {
1212 rb_node = tree_search(&cache->rb_root,
1213 reloc_root->commit_root->start);
1214 if (rb_node) {
1215 node = rb_entry(rb_node, struct backref_node,
1216 rb_node);
1217 BUG_ON(node->detached);
1218 }
1219 }
1220
1221 if (!node)
1222 return 0;
1223
1224 new_node = alloc_backref_node(cache);
1225 if (!new_node)
1226 return -ENOMEM;
1227
1228 new_node->bytenr = dest->node->start;
1229 new_node->level = node->level;
1230 new_node->lowest = node->lowest;
1231 new_node->checked = 1;
1232 new_node->root = dest;
1233
1234 if (!node->lowest) {
1235 list_for_each_entry(edge, &node->lower, list[UPPER]) {
1236 new_edge = alloc_backref_edge(cache);
1237 if (!new_edge)
1238 goto fail;
1239
1240 new_edge->node[UPPER] = new_node;
1241 new_edge->node[LOWER] = edge->node[LOWER];
1242 list_add_tail(&new_edge->list[UPPER],
1243 &new_node->lower);
1244 }
1245 } else {
1246 list_add_tail(&new_node->lower, &cache->leaves);
1247 }
1248
1249 rb_node = tree_insert(&cache->rb_root, new_node->bytenr,
1250 &new_node->rb_node);
1251 if (rb_node)
1252 backref_tree_panic(rb_node, -EEXIST, new_node->bytenr);
1253
1254 if (!new_node->lowest) {
1255 list_for_each_entry(new_edge, &new_node->lower, list[UPPER]) {
1256 list_add_tail(&new_edge->list[LOWER],
1257 &new_edge->node[LOWER]->upper);
1258 }
1259 }
1260 return 0;
1261 fail:
1262 while (!list_empty(&new_node->lower)) {
1263 new_edge = list_entry(new_node->lower.next,
1264 struct backref_edge, list[UPPER]);
1265 list_del(&new_edge->list[UPPER]);
1266 free_backref_edge(cache, new_edge);
1267 }
1268 free_backref_node(cache, new_node);
1269 return -ENOMEM;
1270 }
1271
1272 /*
1273 * helper to add 'address of tree root -> reloc tree' mapping
1274 */
1275 static int __must_check __add_reloc_root(struct btrfs_root *root)
1276 {
1277 struct rb_node *rb_node;
1278 struct mapping_node *node;
1279 struct reloc_control *rc = root->fs_info->reloc_ctl;
1280
1281 node = kmalloc(sizeof(*node), GFP_NOFS);
1282 if (!node)
1283 return -ENOMEM;
1284
1285 node->bytenr = root->node->start;
1286 node->data = root;
1287
1288 spin_lock(&rc->reloc_root_tree.lock);
1289 rb_node = tree_insert(&rc->reloc_root_tree.rb_root,
1290 node->bytenr, &node->rb_node);
1291 spin_unlock(&rc->reloc_root_tree.lock);
1292 if (rb_node) {
1293 btrfs_panic(root->fs_info, -EEXIST, "Duplicate root found "
1294 "for start=%llu while inserting into relocation "
1295 "tree", node->bytenr);
1296 kfree(node);
1297 return -EEXIST;
1298 }
1299
1300 list_add_tail(&root->root_list, &rc->reloc_roots);
1301 return 0;
1302 }
1303
1304 /*
1305 * helper to delete the 'address of tree root -> reloc tree'
1306 * mapping
1307 */
1308 static void __del_reloc_root(struct btrfs_root *root)
1309 {
1310 struct rb_node *rb_node;
1311 struct mapping_node *node = NULL;
1312 struct reloc_control *rc = root->fs_info->reloc_ctl;
1313
1314 spin_lock(&rc->reloc_root_tree.lock);
1315 rb_node = tree_search(&rc->reloc_root_tree.rb_root,
1316 root->node->start);
1317 if (rb_node) {
1318 node = rb_entry(rb_node, struct mapping_node, rb_node);
1319 rb_erase(&node->rb_node, &rc->reloc_root_tree.rb_root);
1320 }
1321 spin_unlock(&rc->reloc_root_tree.lock);
1322
1323 if (!node)
1324 return;
1325 BUG_ON((struct btrfs_root *)node->data != root);
1326
1327 spin_lock(&root->fs_info->trans_lock);
1328 list_del_init(&root->root_list);
1329 spin_unlock(&root->fs_info->trans_lock);
1330 kfree(node);
1331 }
1332
1333 /*
1334 * helper to update the 'address of tree root -> reloc tree'
1335 * mapping
1336 */
1337 static int __update_reloc_root(struct btrfs_root *root, u64 new_bytenr)
1338 {
1339 struct rb_node *rb_node;
1340 struct mapping_node *node = NULL;
1341 struct reloc_control *rc = root->fs_info->reloc_ctl;
1342
1343 spin_lock(&rc->reloc_root_tree.lock);
1344 rb_node = tree_search(&rc->reloc_root_tree.rb_root,
1345 root->node->start);
1346 if (rb_node) {
1347 node = rb_entry(rb_node, struct mapping_node, rb_node);
1348 rb_erase(&node->rb_node, &rc->reloc_root_tree.rb_root);
1349 }
1350 spin_unlock(&rc->reloc_root_tree.lock);
1351
1352 if (!node)
1353 return 0;
1354 BUG_ON((struct btrfs_root *)node->data != root);
1355
1356 spin_lock(&rc->reloc_root_tree.lock);
1357 node->bytenr = new_bytenr;
1358 rb_node = tree_insert(&rc->reloc_root_tree.rb_root,
1359 node->bytenr, &node->rb_node);
1360 spin_unlock(&rc->reloc_root_tree.lock);
1361 if (rb_node)
1362 backref_tree_panic(rb_node, -EEXIST, node->bytenr);
1363 return 0;
1364 }
1365
1366 static struct btrfs_root *create_reloc_root(struct btrfs_trans_handle *trans,
1367 struct btrfs_root *root, u64 objectid)
1368 {
1369 struct btrfs_root *reloc_root;
1370 struct extent_buffer *eb;
1371 struct btrfs_root_item *root_item;
1372 struct btrfs_key root_key;
1373 u64 last_snap = 0;
1374 int ret;
1375
1376 root_item = kmalloc(sizeof(*root_item), GFP_NOFS);
1377 BUG_ON(!root_item);
1378
1379 root_key.objectid = BTRFS_TREE_RELOC_OBJECTID;
1380 root_key.type = BTRFS_ROOT_ITEM_KEY;
1381 root_key.offset = objectid;
1382
1383 if (root->root_key.objectid == objectid) {
1384 /* called by btrfs_init_reloc_root */
1385 ret = btrfs_copy_root(trans, root, root->commit_root, &eb,
1386 BTRFS_TREE_RELOC_OBJECTID);
1387 BUG_ON(ret);
1388
1389 last_snap = btrfs_root_last_snapshot(&root->root_item);
1390 btrfs_set_root_last_snapshot(&root->root_item,
1391 trans->transid - 1);
1392 } else {
1393 /*
1394 * called by btrfs_reloc_post_snapshot_hook.
1395 * the source tree is a reloc tree, all tree blocks
1396 * modified after it was created have RELOC flag
1397 * set in their headers. so it's OK to not update
1398 * the 'last_snapshot'.
1399 */
1400 ret = btrfs_copy_root(trans, root, root->node, &eb,
1401 BTRFS_TREE_RELOC_OBJECTID);
1402 BUG_ON(ret);
1403 }
1404
1405 memcpy(root_item, &root->root_item, sizeof(*root_item));
1406 btrfs_set_root_bytenr(root_item, eb->start);
1407 btrfs_set_root_level(root_item, btrfs_header_level(eb));
1408 btrfs_set_root_generation(root_item, trans->transid);
1409
1410 if (root->root_key.objectid == objectid) {
1411 btrfs_set_root_refs(root_item, 0);
1412 memset(&root_item->drop_progress, 0,
1413 sizeof(struct btrfs_disk_key));
1414 root_item->drop_level = 0;
1415 /*
1416 * abuse rtransid, it is safe because it is impossible to
1417 * receive data into a relocation tree.
1418 */
1419 btrfs_set_root_rtransid(root_item, last_snap);
1420 btrfs_set_root_otransid(root_item, trans->transid);
1421 }
1422
1423 btrfs_tree_unlock(eb);
1424 free_extent_buffer(eb);
1425
1426 ret = btrfs_insert_root(trans, root->fs_info->tree_root,
1427 &root_key, root_item);
1428 BUG_ON(ret);
1429 kfree(root_item);
1430
1431 reloc_root = btrfs_read_fs_root(root->fs_info->tree_root, &root_key);
1432 BUG_ON(IS_ERR(reloc_root));
1433 reloc_root->last_trans = trans->transid;
1434 return reloc_root;
1435 }
1436
1437 /*
1438 * create reloc tree for a given fs tree. reloc tree is just a
1439 * snapshot of the fs tree with special root objectid.
1440 */
1441 int btrfs_init_reloc_root(struct btrfs_trans_handle *trans,
1442 struct btrfs_root *root)
1443 {
1444 struct btrfs_root *reloc_root;
1445 struct reloc_control *rc = root->fs_info->reloc_ctl;
1446 struct btrfs_block_rsv *rsv;
1447 int clear_rsv = 0;
1448 int ret;
1449
1450 if (root->reloc_root) {
1451 reloc_root = root->reloc_root;
1452 reloc_root->last_trans = trans->transid;
1453 return 0;
1454 }
1455
1456 if (!rc || !rc->create_reloc_tree ||
1457 root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID)
1458 return 0;
1459
1460 if (!trans->reloc_reserved) {
1461 rsv = trans->block_rsv;
1462 trans->block_rsv = rc->block_rsv;
1463 clear_rsv = 1;
1464 }
1465 reloc_root = create_reloc_root(trans, root, root->root_key.objectid);
1466 if (clear_rsv)
1467 trans->block_rsv = rsv;
1468
1469 ret = __add_reloc_root(reloc_root);
1470 BUG_ON(ret < 0);
1471 root->reloc_root = reloc_root;
1472 return 0;
1473 }
1474
1475 /*
1476 * update root item of reloc tree
1477 */
1478 int btrfs_update_reloc_root(struct btrfs_trans_handle *trans,
1479 struct btrfs_root *root)
1480 {
1481 struct btrfs_root *reloc_root;
1482 struct btrfs_root_item *root_item;
1483 int ret;
1484
1485 if (!root->reloc_root)
1486 goto out;
1487
1488 reloc_root = root->reloc_root;
1489 root_item = &reloc_root->root_item;
1490
1491 if (root->fs_info->reloc_ctl->merge_reloc_tree &&
1492 btrfs_root_refs(root_item) == 0) {
1493 root->reloc_root = NULL;
1494 __del_reloc_root(reloc_root);
1495 }
1496
1497 if (reloc_root->commit_root != reloc_root->node) {
1498 btrfs_set_root_node(root_item, reloc_root->node);
1499 free_extent_buffer(reloc_root->commit_root);
1500 reloc_root->commit_root = btrfs_root_node(reloc_root);
1501 }
1502
1503 ret = btrfs_update_root(trans, root->fs_info->tree_root,
1504 &reloc_root->root_key, root_item);
1505 BUG_ON(ret);
1506
1507 out:
1508 return 0;
1509 }
1510
1511 /*
1512 * helper to find first cached inode with inode number >= objectid
1513 * in a subvolume
1514 */
1515 static struct inode *find_next_inode(struct btrfs_root *root, u64 objectid)
1516 {
1517 struct rb_node *node;
1518 struct rb_node *prev;
1519 struct btrfs_inode *entry;
1520 struct inode *inode;
1521
1522 spin_lock(&root->inode_lock);
1523 again:
1524 node = root->inode_tree.rb_node;
1525 prev = NULL;
1526 while (node) {
1527 prev = node;
1528 entry = rb_entry(node, struct btrfs_inode, rb_node);
1529
1530 if (objectid < btrfs_ino(&entry->vfs_inode))
1531 node = node->rb_left;
1532 else if (objectid > btrfs_ino(&entry->vfs_inode))
1533 node = node->rb_right;
1534 else
1535 break;
1536 }
1537 if (!node) {
1538 while (prev) {
1539 entry = rb_entry(prev, struct btrfs_inode, rb_node);
1540 if (objectid <= btrfs_ino(&entry->vfs_inode)) {
1541 node = prev;
1542 break;
1543 }
1544 prev = rb_next(prev);
1545 }
1546 }
1547 while (node) {
1548 entry = rb_entry(node, struct btrfs_inode, rb_node);
1549 inode = igrab(&entry->vfs_inode);
1550 if (inode) {
1551 spin_unlock(&root->inode_lock);
1552 return inode;
1553 }
1554
1555 objectid = btrfs_ino(&entry->vfs_inode) + 1;
1556 if (cond_resched_lock(&root->inode_lock))
1557 goto again;
1558
1559 node = rb_next(node);
1560 }
1561 spin_unlock(&root->inode_lock);
1562 return NULL;
1563 }
1564
1565 static int in_block_group(u64 bytenr,
1566 struct btrfs_block_group_cache *block_group)
1567 {
1568 if (bytenr >= block_group->key.objectid &&
1569 bytenr < block_group->key.objectid + block_group->key.offset)
1570 return 1;
1571 return 0;
1572 }
1573
1574 /*
1575 * get new location of data
1576 */
1577 static int get_new_location(struct inode *reloc_inode, u64 *new_bytenr,
1578 u64 bytenr, u64 num_bytes)
1579 {
1580 struct btrfs_root *root = BTRFS_I(reloc_inode)->root;
1581 struct btrfs_path *path;
1582 struct btrfs_file_extent_item *fi;
1583 struct extent_buffer *leaf;
1584 int ret;
1585
1586 path = btrfs_alloc_path();
1587 if (!path)
1588 return -ENOMEM;
1589
1590 bytenr -= BTRFS_I(reloc_inode)->index_cnt;
1591 ret = btrfs_lookup_file_extent(NULL, root, path, btrfs_ino(reloc_inode),
1592 bytenr, 0);
1593 if (ret < 0)
1594 goto out;
1595 if (ret > 0) {
1596 ret = -ENOENT;
1597 goto out;
1598 }
1599
1600 leaf = path->nodes[0];
1601 fi = btrfs_item_ptr(leaf, path->slots[0],
1602 struct btrfs_file_extent_item);
1603
1604 BUG_ON(btrfs_file_extent_offset(leaf, fi) ||
1605 btrfs_file_extent_compression(leaf, fi) ||
1606 btrfs_file_extent_encryption(leaf, fi) ||
1607 btrfs_file_extent_other_encoding(leaf, fi));
1608
1609 if (num_bytes != btrfs_file_extent_disk_num_bytes(leaf, fi)) {
1610 ret = -EINVAL;
1611 goto out;
1612 }
1613
1614 *new_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
1615 ret = 0;
1616 out:
1617 btrfs_free_path(path);
1618 return ret;
1619 }
1620
1621 /*
1622 * update file extent items in the tree leaf to point to
1623 * the new locations.
1624 */
1625 static noinline_for_stack
1626 int replace_file_extents(struct btrfs_trans_handle *trans,
1627 struct reloc_control *rc,
1628 struct btrfs_root *root,
1629 struct extent_buffer *leaf)
1630 {
1631 struct btrfs_key key;
1632 struct btrfs_file_extent_item *fi;
1633 struct inode *inode = NULL;
1634 u64 parent;
1635 u64 bytenr;
1636 u64 new_bytenr = 0;
1637 u64 num_bytes;
1638 u64 end;
1639 u32 nritems;
1640 u32 i;
1641 int ret = 0;
1642 int first = 1;
1643 int dirty = 0;
1644
1645 if (rc->stage != UPDATE_DATA_PTRS)
1646 return 0;
1647
1648 /* reloc trees always use full backref */
1649 if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID)
1650 parent = leaf->start;
1651 else
1652 parent = 0;
1653
1654 nritems = btrfs_header_nritems(leaf);
1655 for (i = 0; i < nritems; i++) {
1656 cond_resched();
1657 btrfs_item_key_to_cpu(leaf, &key, i);
1658 if (key.type != BTRFS_EXTENT_DATA_KEY)
1659 continue;
1660 fi = btrfs_item_ptr(leaf, i, struct btrfs_file_extent_item);
1661 if (btrfs_file_extent_type(leaf, fi) ==
1662 BTRFS_FILE_EXTENT_INLINE)
1663 continue;
1664 bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
1665 num_bytes = btrfs_file_extent_disk_num_bytes(leaf, fi);
1666 if (bytenr == 0)
1667 continue;
1668 if (!in_block_group(bytenr, rc->block_group))
1669 continue;
1670
1671 /*
1672 * if we are modifying block in fs tree, wait for readpage
1673 * to complete and drop the extent cache
1674 */
1675 if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) {
1676 if (first) {
1677 inode = find_next_inode(root, key.objectid);
1678 first = 0;
1679 } else if (inode && btrfs_ino(inode) < key.objectid) {
1680 btrfs_add_delayed_iput(inode);
1681 inode = find_next_inode(root, key.objectid);
1682 }
1683 if (inode && btrfs_ino(inode) == key.objectid) {
1684 end = key.offset +
1685 btrfs_file_extent_num_bytes(leaf, fi);
1686 WARN_ON(!IS_ALIGNED(key.offset,
1687 root->sectorsize));
1688 WARN_ON(!IS_ALIGNED(end, root->sectorsize));
1689 end--;
1690 ret = try_lock_extent(&BTRFS_I(inode)->io_tree,
1691 key.offset, end);
1692 if (!ret)
1693 continue;
1694
1695 btrfs_drop_extent_cache(inode, key.offset, end,
1696 1);
1697 unlock_extent(&BTRFS_I(inode)->io_tree,
1698 key.offset, end);
1699 }
1700 }
1701
1702 ret = get_new_location(rc->data_inode, &new_bytenr,
1703 bytenr, num_bytes);
1704 if (ret) {
1705 /*
1706 * Don't have to abort since we've not changed anything
1707 * in the file extent yet.
1708 */
1709 break;
1710 }
1711
1712 btrfs_set_file_extent_disk_bytenr(leaf, fi, new_bytenr);
1713 dirty = 1;
1714
1715 key.offset -= btrfs_file_extent_offset(leaf, fi);
1716 ret = btrfs_inc_extent_ref(trans, root, new_bytenr,
1717 num_bytes, parent,
1718 btrfs_header_owner(leaf),
1719 key.objectid, key.offset, 1);
1720 if (ret) {
1721 btrfs_abort_transaction(trans, root, ret);
1722 break;
1723 }
1724
1725 ret = btrfs_free_extent(trans, root, bytenr, num_bytes,
1726 parent, btrfs_header_owner(leaf),
1727 key.objectid, key.offset, 1);
1728 if (ret) {
1729 btrfs_abort_transaction(trans, root, ret);
1730 break;
1731 }
1732 }
1733 if (dirty)
1734 btrfs_mark_buffer_dirty(leaf);
1735 if (inode)
1736 btrfs_add_delayed_iput(inode);
1737 return ret;
1738 }
1739
1740 static noinline_for_stack
1741 int memcmp_node_keys(struct extent_buffer *eb, int slot,
1742 struct btrfs_path *path, int level)
1743 {
1744 struct btrfs_disk_key key1;
1745 struct btrfs_disk_key key2;
1746 btrfs_node_key(eb, &key1, slot);
1747 btrfs_node_key(path->nodes[level], &key2, path->slots[level]);
1748 return memcmp(&key1, &key2, sizeof(key1));
1749 }
1750
1751 /*
1752 * try to replace tree blocks in fs tree with the new blocks
1753 * in reloc tree. tree blocks haven't been modified since the
1754 * reloc tree was create can be replaced.
1755 *
1756 * if a block was replaced, level of the block + 1 is returned.
1757 * if no block got replaced, 0 is returned. if there are other
1758 * errors, a negative error number is returned.
1759 */
1760 static noinline_for_stack
1761 int replace_path(struct btrfs_trans_handle *trans,
1762 struct btrfs_root *dest, struct btrfs_root *src,
1763 struct btrfs_path *path, struct btrfs_key *next_key,
1764 int lowest_level, int max_level)
1765 {
1766 struct extent_buffer *eb;
1767 struct extent_buffer *parent;
1768 struct btrfs_key key;
1769 u64 old_bytenr;
1770 u64 new_bytenr;
1771 u64 old_ptr_gen;
1772 u64 new_ptr_gen;
1773 u64 last_snapshot;
1774 u32 blocksize;
1775 int cow = 0;
1776 int level;
1777 int ret;
1778 int slot;
1779
1780 BUG_ON(src->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID);
1781 BUG_ON(dest->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID);
1782
1783 last_snapshot = btrfs_root_last_snapshot(&src->root_item);
1784 again:
1785 slot = path->slots[lowest_level];
1786 btrfs_node_key_to_cpu(path->nodes[lowest_level], &key, slot);
1787
1788 eb = btrfs_lock_root_node(dest);
1789 btrfs_set_lock_blocking(eb);
1790 level = btrfs_header_level(eb);
1791
1792 if (level < lowest_level) {
1793 btrfs_tree_unlock(eb);
1794 free_extent_buffer(eb);
1795 return 0;
1796 }
1797
1798 if (cow) {
1799 ret = btrfs_cow_block(trans, dest, eb, NULL, 0, &eb);
1800 BUG_ON(ret);
1801 }
1802 btrfs_set_lock_blocking(eb);
1803
1804 if (next_key) {
1805 next_key->objectid = (u64)-1;
1806 next_key->type = (u8)-1;
1807 next_key->offset = (u64)-1;
1808 }
1809
1810 parent = eb;
1811 while (1) {
1812 level = btrfs_header_level(parent);
1813 BUG_ON(level < lowest_level);
1814
1815 ret = btrfs_bin_search(parent, &key, level, &slot);
1816 if (ret && slot > 0)
1817 slot--;
1818
1819 if (next_key && slot + 1 < btrfs_header_nritems(parent))
1820 btrfs_node_key_to_cpu(parent, next_key, slot + 1);
1821
1822 old_bytenr = btrfs_node_blockptr(parent, slot);
1823 blocksize = dest->nodesize;
1824 old_ptr_gen = btrfs_node_ptr_generation(parent, slot);
1825
1826 if (level <= max_level) {
1827 eb = path->nodes[level];
1828 new_bytenr = btrfs_node_blockptr(eb,
1829 path->slots[level]);
1830 new_ptr_gen = btrfs_node_ptr_generation(eb,
1831 path->slots[level]);
1832 } else {
1833 new_bytenr = 0;
1834 new_ptr_gen = 0;
1835 }
1836
1837 if (WARN_ON(new_bytenr > 0 && new_bytenr == old_bytenr)) {
1838 ret = level;
1839 break;
1840 }
1841
1842 if (new_bytenr == 0 || old_ptr_gen > last_snapshot ||
1843 memcmp_node_keys(parent, slot, path, level)) {
1844 if (level <= lowest_level) {
1845 ret = 0;
1846 break;
1847 }
1848
1849 eb = read_tree_block(dest, old_bytenr, old_ptr_gen);
1850 if (IS_ERR(eb)) {
1851 ret = PTR_ERR(eb);
1852 } else if (!extent_buffer_uptodate(eb)) {
1853 ret = -EIO;
1854 free_extent_buffer(eb);
1855 break;
1856 }
1857 btrfs_tree_lock(eb);
1858 if (cow) {
1859 ret = btrfs_cow_block(trans, dest, eb, parent,
1860 slot, &eb);
1861 BUG_ON(ret);
1862 }
1863 btrfs_set_lock_blocking(eb);
1864
1865 btrfs_tree_unlock(parent);
1866 free_extent_buffer(parent);
1867
1868 parent = eb;
1869 continue;
1870 }
1871
1872 if (!cow) {
1873 btrfs_tree_unlock(parent);
1874 free_extent_buffer(parent);
1875 cow = 1;
1876 goto again;
1877 }
1878
1879 btrfs_node_key_to_cpu(path->nodes[level], &key,
1880 path->slots[level]);
1881 btrfs_release_path(path);
1882
1883 path->lowest_level = level;
1884 ret = btrfs_search_slot(trans, src, &key, path, 0, 1);
1885 path->lowest_level = 0;
1886 BUG_ON(ret);
1887
1888 /*
1889 * swap blocks in fs tree and reloc tree.
1890 */
1891 btrfs_set_node_blockptr(parent, slot, new_bytenr);
1892 btrfs_set_node_ptr_generation(parent, slot, new_ptr_gen);
1893 btrfs_mark_buffer_dirty(parent);
1894
1895 btrfs_set_node_blockptr(path->nodes[level],
1896 path->slots[level], old_bytenr);
1897 btrfs_set_node_ptr_generation(path->nodes[level],
1898 path->slots[level], old_ptr_gen);
1899 btrfs_mark_buffer_dirty(path->nodes[level]);
1900
1901 ret = btrfs_inc_extent_ref(trans, src, old_bytenr, blocksize,
1902 path->nodes[level]->start,
1903 src->root_key.objectid, level - 1, 0,
1904 1);
1905 BUG_ON(ret);
1906 ret = btrfs_inc_extent_ref(trans, dest, new_bytenr, blocksize,
1907 0, dest->root_key.objectid, level - 1,
1908 0, 1);
1909 BUG_ON(ret);
1910
1911 ret = btrfs_free_extent(trans, src, new_bytenr, blocksize,
1912 path->nodes[level]->start,
1913 src->root_key.objectid, level - 1, 0,
1914 1);
1915 BUG_ON(ret);
1916
1917 ret = btrfs_free_extent(trans, dest, old_bytenr, blocksize,
1918 0, dest->root_key.objectid, level - 1,
1919 0, 1);
1920 BUG_ON(ret);
1921
1922 btrfs_unlock_up_safe(path, 0);
1923
1924 ret = level;
1925 break;
1926 }
1927 btrfs_tree_unlock(parent);
1928 free_extent_buffer(parent);
1929 return ret;
1930 }
1931
1932 /*
1933 * helper to find next relocated block in reloc tree
1934 */
1935 static noinline_for_stack
1936 int walk_up_reloc_tree(struct btrfs_root *root, struct btrfs_path *path,
1937 int *level)
1938 {
1939 struct extent_buffer *eb;
1940 int i;
1941 u64 last_snapshot;
1942 u32 nritems;
1943
1944 last_snapshot = btrfs_root_last_snapshot(&root->root_item);
1945
1946 for (i = 0; i < *level; i++) {
1947 free_extent_buffer(path->nodes[i]);
1948 path->nodes[i] = NULL;
1949 }
1950
1951 for (i = *level; i < BTRFS_MAX_LEVEL && path->nodes[i]; i++) {
1952 eb = path->nodes[i];
1953 nritems = btrfs_header_nritems(eb);
1954 while (path->slots[i] + 1 < nritems) {
1955 path->slots[i]++;
1956 if (btrfs_node_ptr_generation(eb, path->slots[i]) <=
1957 last_snapshot)
1958 continue;
1959
1960 *level = i;
1961 return 0;
1962 }
1963 free_extent_buffer(path->nodes[i]);
1964 path->nodes[i] = NULL;
1965 }
1966 return 1;
1967 }
1968
1969 /*
1970 * walk down reloc tree to find relocated block of lowest level
1971 */
1972 static noinline_for_stack
1973 int walk_down_reloc_tree(struct btrfs_root *root, struct btrfs_path *path,
1974 int *level)
1975 {
1976 struct extent_buffer *eb = NULL;
1977 int i;
1978 u64 bytenr;
1979 u64 ptr_gen = 0;
1980 u64 last_snapshot;
1981 u32 nritems;
1982
1983 last_snapshot = btrfs_root_last_snapshot(&root->root_item);
1984
1985 for (i = *level; i > 0; i--) {
1986 eb = path->nodes[i];
1987 nritems = btrfs_header_nritems(eb);
1988 while (path->slots[i] < nritems) {
1989 ptr_gen = btrfs_node_ptr_generation(eb, path->slots[i]);
1990 if (ptr_gen > last_snapshot)
1991 break;
1992 path->slots[i]++;
1993 }
1994 if (path->slots[i] >= nritems) {
1995 if (i == *level)
1996 break;
1997 *level = i + 1;
1998 return 0;
1999 }
2000 if (i == 1) {
2001 *level = i;
2002 return 0;
2003 }
2004
2005 bytenr = btrfs_node_blockptr(eb, path->slots[i]);
2006 eb = read_tree_block(root, bytenr, ptr_gen);
2007 if (IS_ERR(eb)) {
2008 return PTR_ERR(eb);
2009 } else if (!extent_buffer_uptodate(eb)) {
2010 free_extent_buffer(eb);
2011 return -EIO;
2012 }
2013 BUG_ON(btrfs_header_level(eb) != i - 1);
2014 path->nodes[i - 1] = eb;
2015 path->slots[i - 1] = 0;
2016 }
2017 return 1;
2018 }
2019
2020 /*
2021 * invalidate extent cache for file extents whose key in range of
2022 * [min_key, max_key)
2023 */
2024 static int invalidate_extent_cache(struct btrfs_root *root,
2025 struct btrfs_key *min_key,
2026 struct btrfs_key *max_key)
2027 {
2028 struct inode *inode = NULL;
2029 u64 objectid;
2030 u64 start, end;
2031 u64 ino;
2032
2033 objectid = min_key->objectid;
2034 while (1) {
2035 cond_resched();
2036 iput(inode);
2037
2038 if (objectid > max_key->objectid)
2039 break;
2040
2041 inode = find_next_inode(root, objectid);
2042 if (!inode)
2043 break;
2044 ino = btrfs_ino(inode);
2045
2046 if (ino > max_key->objectid) {
2047 iput(inode);
2048 break;
2049 }
2050
2051 objectid = ino + 1;
2052 if (!S_ISREG(inode->i_mode))
2053 continue;
2054
2055 if (unlikely(min_key->objectid == ino)) {
2056 if (min_key->type > BTRFS_EXTENT_DATA_KEY)
2057 continue;
2058 if (min_key->type < BTRFS_EXTENT_DATA_KEY)
2059 start = 0;
2060 else {
2061 start = min_key->offset;
2062 WARN_ON(!IS_ALIGNED(start, root->sectorsize));
2063 }
2064 } else {
2065 start = 0;
2066 }
2067
2068 if (unlikely(max_key->objectid == ino)) {
2069 if (max_key->type < BTRFS_EXTENT_DATA_KEY)
2070 continue;
2071 if (max_key->type > BTRFS_EXTENT_DATA_KEY) {
2072 end = (u64)-1;
2073 } else {
2074 if (max_key->offset == 0)
2075 continue;
2076 end = max_key->offset;
2077 WARN_ON(!IS_ALIGNED(end, root->sectorsize));
2078 end--;
2079 }
2080 } else {
2081 end = (u64)-1;
2082 }
2083
2084 /* the lock_extent waits for readpage to complete */
2085 lock_extent(&BTRFS_I(inode)->io_tree, start, end);
2086 btrfs_drop_extent_cache(inode, start, end, 1);
2087 unlock_extent(&BTRFS_I(inode)->io_tree, start, end);
2088 }
2089 return 0;
2090 }
2091
2092 static int find_next_key(struct btrfs_path *path, int level,
2093 struct btrfs_key *key)
2094
2095 {
2096 while (level < BTRFS_MAX_LEVEL) {
2097 if (!path->nodes[level])
2098 break;
2099 if (path->slots[level] + 1 <
2100 btrfs_header_nritems(path->nodes[level])) {
2101 btrfs_node_key_to_cpu(path->nodes[level], key,
2102 path->slots[level] + 1);
2103 return 0;
2104 }
2105 level++;
2106 }
2107 return 1;
2108 }
2109
2110 /*
2111 * merge the relocated tree blocks in reloc tree with corresponding
2112 * fs tree.
2113 */
2114 static noinline_for_stack int merge_reloc_root(struct reloc_control *rc,
2115 struct btrfs_root *root)
2116 {
2117 LIST_HEAD(inode_list);
2118 struct btrfs_key key;
2119 struct btrfs_key next_key;
2120 struct btrfs_trans_handle *trans = NULL;
2121 struct btrfs_root *reloc_root;
2122 struct btrfs_root_item *root_item;
2123 struct btrfs_path *path;
2124 struct extent_buffer *leaf;
2125 int level;
2126 int max_level;
2127 int replaced = 0;
2128 int ret;
2129 int err = 0;
2130 u32 min_reserved;
2131
2132 path = btrfs_alloc_path();
2133 if (!path)
2134 return -ENOMEM;
2135 path->reada = 1;
2136
2137 reloc_root = root->reloc_root;
2138 root_item = &reloc_root->root_item;
2139
2140 if (btrfs_disk_key_objectid(&root_item->drop_progress) == 0) {
2141 level = btrfs_root_level(root_item);
2142 extent_buffer_get(reloc_root->node);
2143 path->nodes[level] = reloc_root->node;
2144 path->slots[level] = 0;
2145 } else {
2146 btrfs_disk_key_to_cpu(&key, &root_item->drop_progress);
2147
2148 level = root_item->drop_level;
2149 BUG_ON(level == 0);
2150 path->lowest_level = level;
2151 ret = btrfs_search_slot(NULL, reloc_root, &key, path, 0, 0);
2152 path->lowest_level = 0;
2153 if (ret < 0) {
2154 btrfs_free_path(path);
2155 return ret;
2156 }
2157
2158 btrfs_node_key_to_cpu(path->nodes[level], &next_key,
2159 path->slots[level]);
2160 WARN_ON(memcmp(&key, &next_key, sizeof(key)));
2161
2162 btrfs_unlock_up_safe(path, 0);
2163 }
2164
2165 min_reserved = root->nodesize * (BTRFS_MAX_LEVEL - 1) * 2;
2166 memset(&next_key, 0, sizeof(next_key));
2167
2168 while (1) {
2169 ret = btrfs_block_rsv_refill(root, rc->block_rsv, min_reserved,
2170 BTRFS_RESERVE_FLUSH_ALL);
2171 if (ret) {
2172 err = ret;
2173 goto out;
2174 }
2175 trans = btrfs_start_transaction(root, 0);
2176 if (IS_ERR(trans)) {
2177 err = PTR_ERR(trans);
2178 trans = NULL;
2179 goto out;
2180 }
2181 trans->block_rsv = rc->block_rsv;
2182
2183 replaced = 0;
2184 max_level = level;
2185
2186 ret = walk_down_reloc_tree(reloc_root, path, &level);
2187 if (ret < 0) {
2188 err = ret;
2189 goto out;
2190 }
2191 if (ret > 0)
2192 break;
2193
2194 if (!find_next_key(path, level, &key) &&
2195 btrfs_comp_cpu_keys(&next_key, &key) >= 0) {
2196 ret = 0;
2197 } else {
2198 ret = replace_path(trans, root, reloc_root, path,
2199 &next_key, level, max_level);
2200 }
2201 if (ret < 0) {
2202 err = ret;
2203 goto out;
2204 }
2205
2206 if (ret > 0) {
2207 level = ret;
2208 btrfs_node_key_to_cpu(path->nodes[level], &key,
2209 path->slots[level]);
2210 replaced = 1;
2211 }
2212
2213 ret = walk_up_reloc_tree(reloc_root, path, &level);
2214 if (ret > 0)
2215 break;
2216
2217 BUG_ON(level == 0);
2218 /*
2219 * save the merging progress in the drop_progress.
2220 * this is OK since root refs == 1 in this case.
2221 */
2222 btrfs_node_key(path->nodes[level], &root_item->drop_progress,
2223 path->slots[level]);
2224 root_item->drop_level = level;
2225
2226 btrfs_end_transaction_throttle(trans, root);
2227 trans = NULL;
2228
2229 btrfs_btree_balance_dirty(root);
2230
2231 if (replaced && rc->stage == UPDATE_DATA_PTRS)
2232 invalidate_extent_cache(root, &key, &next_key);
2233 }
2234
2235 /*
2236 * handle the case only one block in the fs tree need to be
2237 * relocated and the block is tree root.
2238 */
2239 leaf = btrfs_lock_root_node(root);
2240 ret = btrfs_cow_block(trans, root, leaf, NULL, 0, &leaf);
2241 btrfs_tree_unlock(leaf);
2242 free_extent_buffer(leaf);
2243 if (ret < 0)
2244 err = ret;
2245 out:
2246 btrfs_free_path(path);
2247
2248 if (err == 0) {
2249 memset(&root_item->drop_progress, 0,
2250 sizeof(root_item->drop_progress));
2251 root_item->drop_level = 0;
2252 btrfs_set_root_refs(root_item, 0);
2253 btrfs_update_reloc_root(trans, root);
2254 }
2255
2256 if (trans)
2257 btrfs_end_transaction_throttle(trans, root);
2258
2259 btrfs_btree_balance_dirty(root);
2260
2261 if (replaced && rc->stage == UPDATE_DATA_PTRS)
2262 invalidate_extent_cache(root, &key, &next_key);
2263
2264 return err;
2265 }
2266
2267 static noinline_for_stack
2268 int prepare_to_merge(struct reloc_control *rc, int err)
2269 {
2270 struct btrfs_root *root = rc->extent_root;
2271 struct btrfs_root *reloc_root;
2272 struct btrfs_trans_handle *trans;
2273 LIST_HEAD(reloc_roots);
2274 u64 num_bytes = 0;
2275 int ret;
2276
2277 mutex_lock(&root->fs_info->reloc_mutex);
2278 rc->merging_rsv_size += root->nodesize * (BTRFS_MAX_LEVEL - 1) * 2;
2279 rc->merging_rsv_size += rc->nodes_relocated * 2;
2280 mutex_unlock(&root->fs_info->reloc_mutex);
2281
2282 again:
2283 if (!err) {
2284 num_bytes = rc->merging_rsv_size;
2285 ret = btrfs_block_rsv_add(root, rc->block_rsv, num_bytes,
2286 BTRFS_RESERVE_FLUSH_ALL);
2287 if (ret)
2288 err = ret;
2289 }
2290
2291 trans = btrfs_join_transaction(rc->extent_root);
2292 if (IS_ERR(trans)) {
2293 if (!err)
2294 btrfs_block_rsv_release(rc->extent_root,
2295 rc->block_rsv, num_bytes);
2296 return PTR_ERR(trans);
2297 }
2298
2299 if (!err) {
2300 if (num_bytes != rc->merging_rsv_size) {
2301 btrfs_end_transaction(trans, rc->extent_root);
2302 btrfs_block_rsv_release(rc->extent_root,
2303 rc->block_rsv, num_bytes);
2304 goto again;
2305 }
2306 }
2307
2308 rc->merge_reloc_tree = 1;
2309
2310 while (!list_empty(&rc->reloc_roots)) {
2311 reloc_root = list_entry(rc->reloc_roots.next,
2312 struct btrfs_root, root_list);
2313 list_del_init(&reloc_root->root_list);
2314
2315 root = read_fs_root(reloc_root->fs_info,
2316 reloc_root->root_key.offset);
2317 BUG_ON(IS_ERR(root));
2318 BUG_ON(root->reloc_root != reloc_root);
2319
2320 /*
2321 * set reference count to 1, so btrfs_recover_relocation
2322 * knows it should resumes merging
2323 */
2324 if (!err)
2325 btrfs_set_root_refs(&reloc_root->root_item, 1);
2326 btrfs_update_reloc_root(trans, root);
2327
2328 list_add(&reloc_root->root_list, &reloc_roots);
2329 }
2330
2331 list_splice(&reloc_roots, &rc->reloc_roots);
2332
2333 if (!err)
2334 btrfs_commit_transaction(trans, rc->extent_root);
2335 else
2336 btrfs_end_transaction(trans, rc->extent_root);
2337 return err;
2338 }
2339
2340 static noinline_for_stack
2341 void free_reloc_roots(struct list_head *list)
2342 {
2343 struct btrfs_root *reloc_root;
2344
2345 while (!list_empty(list)) {
2346 reloc_root = list_entry(list->next, struct btrfs_root,
2347 root_list);
2348 __del_reloc_root(reloc_root);
2349 }
2350 }
2351
2352 static noinline_for_stack
2353 void merge_reloc_roots(struct reloc_control *rc)
2354 {
2355 struct btrfs_root *root;
2356 struct btrfs_root *reloc_root;
2357 u64 last_snap;
2358 u64 otransid;
2359 u64 objectid;
2360 LIST_HEAD(reloc_roots);
2361 int found = 0;
2362 int ret = 0;
2363 again:
2364 root = rc->extent_root;
2365
2366 /*
2367 * this serializes us with btrfs_record_root_in_transaction,
2368 * we have to make sure nobody is in the middle of
2369 * adding their roots to the list while we are
2370 * doing this splice
2371 */
2372 mutex_lock(&root->fs_info->reloc_mutex);
2373 list_splice_init(&rc->reloc_roots, &reloc_roots);
2374 mutex_unlock(&root->fs_info->reloc_mutex);
2375
2376 while (!list_empty(&reloc_roots)) {
2377 found = 1;
2378 reloc_root = list_entry(reloc_roots.next,
2379 struct btrfs_root, root_list);
2380
2381 if (btrfs_root_refs(&reloc_root->root_item) > 0) {
2382 root = read_fs_root(reloc_root->fs_info,
2383 reloc_root->root_key.offset);
2384 BUG_ON(IS_ERR(root));
2385 BUG_ON(root->reloc_root != reloc_root);
2386
2387 ret = merge_reloc_root(rc, root);
2388 if (ret) {
2389 if (list_empty(&reloc_root->root_list))
2390 list_add_tail(&reloc_root->root_list,
2391 &reloc_roots);
2392 goto out;
2393 }
2394 } else {
2395 list_del_init(&reloc_root->root_list);
2396 }
2397
2398 /*
2399 * we keep the old last snapshod transid in rtranid when we
2400 * created the relocation tree.
2401 */
2402 last_snap = btrfs_root_rtransid(&reloc_root->root_item);
2403 otransid = btrfs_root_otransid(&reloc_root->root_item);
2404 objectid = reloc_root->root_key.offset;
2405
2406 ret = btrfs_drop_snapshot(reloc_root, rc->block_rsv, 0, 1);
2407 if (ret < 0) {
2408 if (list_empty(&reloc_root->root_list))
2409 list_add_tail(&reloc_root->root_list,
2410 &reloc_roots);
2411 goto out;
2412 }
2413 }
2414
2415 if (found) {
2416 found = 0;
2417 goto again;
2418 }
2419 out:
2420 if (ret) {
2421 btrfs_std_error(root->fs_info, ret);
2422 if (!list_empty(&reloc_roots))
2423 free_reloc_roots(&reloc_roots);
2424
2425 /* new reloc root may be added */
2426 mutex_lock(&root->fs_info->reloc_mutex);
2427 list_splice_init(&rc->reloc_roots, &reloc_roots);
2428 mutex_unlock(&root->fs_info->reloc_mutex);
2429 if (!list_empty(&reloc_roots))
2430 free_reloc_roots(&reloc_roots);
2431 }
2432
2433 BUG_ON(!RB_EMPTY_ROOT(&rc->reloc_root_tree.rb_root));
2434 }
2435
2436 static void free_block_list(struct rb_root *blocks)
2437 {
2438 struct tree_block *block;
2439 struct rb_node *rb_node;
2440 while ((rb_node = rb_first(blocks))) {
2441 block = rb_entry(rb_node, struct tree_block, rb_node);
2442 rb_erase(rb_node, blocks);
2443 kfree(block);
2444 }
2445 }
2446
2447 static int record_reloc_root_in_trans(struct btrfs_trans_handle *trans,
2448 struct btrfs_root *reloc_root)
2449 {
2450 struct btrfs_root *root;
2451
2452 if (reloc_root->last_trans == trans->transid)
2453 return 0;
2454
2455 root = read_fs_root(reloc_root->fs_info, reloc_root->root_key.offset);
2456 BUG_ON(IS_ERR(root));
2457 BUG_ON(root->reloc_root != reloc_root);
2458
2459 return btrfs_record_root_in_trans(trans, root);
2460 }
2461
2462 static noinline_for_stack
2463 struct btrfs_root *select_reloc_root(struct btrfs_trans_handle *trans,
2464 struct reloc_control *rc,
2465 struct backref_node *node,
2466 struct backref_edge *edges[])
2467 {
2468 struct backref_node *next;
2469 struct btrfs_root *root;
2470 int index = 0;
2471
2472 next = node;
2473 while (1) {
2474 cond_resched();
2475 next = walk_up_backref(next, edges, &index);
2476 root = next->root;
2477 BUG_ON(!root);
2478 BUG_ON(!test_bit(BTRFS_ROOT_REF_COWS, &root->state));
2479
2480 if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID) {
2481 record_reloc_root_in_trans(trans, root);
2482 break;
2483 }
2484
2485 btrfs_record_root_in_trans(trans, root);
2486 root = root->reloc_root;
2487
2488 if (next->new_bytenr != root->node->start) {
2489 BUG_ON(next->new_bytenr);
2490 BUG_ON(!list_empty(&next->list));
2491 next->new_bytenr = root->node->start;
2492 next->root = root;
2493 list_add_tail(&next->list,
2494 &rc->backref_cache.changed);
2495 __mark_block_processed(rc, next);
2496 break;
2497 }
2498
2499 WARN_ON(1);
2500 root = NULL;
2501 next = walk_down_backref(edges, &index);
2502 if (!next || next->level <= node->level)
2503 break;
2504 }
2505 if (!root)
2506 return NULL;
2507
2508 next = node;
2509 /* setup backref node path for btrfs_reloc_cow_block */
2510 while (1) {
2511 rc->backref_cache.path[next->level] = next;
2512 if (--index < 0)
2513 break;
2514 next = edges[index]->node[UPPER];
2515 }
2516 return root;
2517 }
2518
2519 /*
2520 * select a tree root for relocation. return NULL if the block
2521 * is reference counted. we should use do_relocation() in this
2522 * case. return a tree root pointer if the block isn't reference
2523 * counted. return -ENOENT if the block is root of reloc tree.
2524 */
2525 static noinline_for_stack
2526 struct btrfs_root *select_one_root(struct btrfs_trans_handle *trans,
2527 struct backref_node *node)
2528 {
2529 struct backref_node *next;
2530 struct btrfs_root *root;
2531 struct btrfs_root *fs_root = NULL;
2532 struct backref_edge *edges[BTRFS_MAX_LEVEL - 1];
2533 int index = 0;
2534
2535 next = node;
2536 while (1) {
2537 cond_resched();
2538 next = walk_up_backref(next, edges, &index);
2539 root = next->root;
2540 BUG_ON(!root);
2541
2542 /* no other choice for non-references counted tree */
2543 if (!test_bit(BTRFS_ROOT_REF_COWS, &root->state))
2544 return root;
2545
2546 if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID)
2547 fs_root = root;
2548
2549 if (next != node)
2550 return NULL;
2551
2552 next = walk_down_backref(edges, &index);
2553 if (!next || next->level <= node->level)
2554 break;
2555 }
2556
2557 if (!fs_root)
2558 return ERR_PTR(-ENOENT);
2559 return fs_root;
2560 }
2561
2562 static noinline_for_stack
2563 u64 calcu_metadata_size(struct reloc_control *rc,
2564 struct backref_node *node, int reserve)
2565 {
2566 struct backref_node *next = node;
2567 struct backref_edge *edge;
2568 struct backref_edge *edges[BTRFS_MAX_LEVEL - 1];
2569 u64 num_bytes = 0;
2570 int index = 0;
2571
2572 BUG_ON(reserve && node->processed);
2573
2574 while (next) {
2575 cond_resched();
2576 while (1) {
2577 if (next->processed && (reserve || next != node))
2578 break;
2579
2580 num_bytes += rc->extent_root->nodesize;
2581
2582 if (list_empty(&next->upper))
2583 break;
2584
2585 edge = list_entry(next->upper.next,
2586 struct backref_edge, list[LOWER]);
2587 edges[index++] = edge;
2588 next = edge->node[UPPER];
2589 }
2590 next = walk_down_backref(edges, &index);
2591 }
2592 return num_bytes;
2593 }
2594
2595 static int reserve_metadata_space(struct btrfs_trans_handle *trans,
2596 struct reloc_control *rc,
2597 struct backref_node *node)
2598 {
2599 struct btrfs_root *root = rc->extent_root;
2600 u64 num_bytes;
2601 int ret;
2602 u64 tmp;
2603
2604 num_bytes = calcu_metadata_size(rc, node, 1) * 2;
2605
2606 trans->block_rsv = rc->block_rsv;
2607 rc->reserved_bytes += num_bytes;
2608 ret = btrfs_block_rsv_refill(root, rc->block_rsv, num_bytes,
2609 BTRFS_RESERVE_FLUSH_ALL);
2610 if (ret) {
2611 if (ret == -EAGAIN) {
2612 tmp = rc->extent_root->nodesize *
2613 RELOCATION_RESERVED_NODES;
2614 while (tmp <= rc->reserved_bytes)
2615 tmp <<= 1;
2616 /*
2617 * only one thread can access block_rsv at this point,
2618 * so we don't need hold lock to protect block_rsv.
2619 * we expand more reservation size here to allow enough
2620 * space for relocation and we will return eailer in
2621 * enospc case.
2622 */
2623 rc->block_rsv->size = tmp + rc->extent_root->nodesize *
2624 RELOCATION_RESERVED_NODES;
2625 }
2626 return ret;
2627 }
2628
2629 return 0;
2630 }
2631
2632 /*
2633 * relocate a block tree, and then update pointers in upper level
2634 * blocks that reference the block to point to the new location.
2635 *
2636 * if called by link_to_upper, the block has already been relocated.
2637 * in that case this function just updates pointers.
2638 */
2639 static int do_relocation(struct btrfs_trans_handle *trans,
2640 struct reloc_control *rc,
2641 struct backref_node *node,
2642 struct btrfs_key *key,
2643 struct btrfs_path *path, int lowest)
2644 {
2645 struct backref_node *upper;
2646 struct backref_edge *edge;
2647 struct backref_edge *edges[BTRFS_MAX_LEVEL - 1];
2648 struct btrfs_root *root;
2649 struct extent_buffer *eb;
2650 u32 blocksize;
2651 u64 bytenr;
2652 u64 generation;
2653 int slot;
2654 int ret;
2655 int err = 0;
2656
2657 BUG_ON(lowest && node->eb);
2658
2659 path->lowest_level = node->level + 1;
2660 rc->backref_cache.path[node->level] = node;
2661 list_for_each_entry(edge, &node->upper, list[LOWER]) {
2662 cond_resched();
2663
2664 upper = edge->node[UPPER];
2665 root = select_reloc_root(trans, rc, upper, edges);
2666 BUG_ON(!root);
2667
2668 if (upper->eb && !upper->locked) {
2669 if (!lowest) {
2670 ret = btrfs_bin_search(upper->eb, key,
2671 upper->level, &slot);
2672 BUG_ON(ret);
2673 bytenr = btrfs_node_blockptr(upper->eb, slot);
2674 if (node->eb->start == bytenr)
2675 goto next;
2676 }
2677 drop_node_buffer(upper);
2678 }
2679
2680 if (!upper->eb) {
2681 ret = btrfs_search_slot(trans, root, key, path, 0, 1);
2682 if (ret < 0) {
2683 err = ret;
2684 break;
2685 }
2686 BUG_ON(ret > 0);
2687
2688 if (!upper->eb) {
2689 upper->eb = path->nodes[upper->level];
2690 path->nodes[upper->level] = NULL;
2691 } else {
2692 BUG_ON(upper->eb != path->nodes[upper->level]);
2693 }
2694
2695 upper->locked = 1;
2696 path->locks[upper->level] = 0;
2697
2698 slot = path->slots[upper->level];
2699 btrfs_release_path(path);
2700 } else {
2701 ret = btrfs_bin_search(upper->eb, key, upper->level,
2702 &slot);
2703 BUG_ON(ret);
2704 }
2705
2706 bytenr = btrfs_node_blockptr(upper->eb, slot);
2707 if (lowest) {
2708 BUG_ON(bytenr != node->bytenr);
2709 } else {
2710 if (node->eb->start == bytenr)
2711 goto next;
2712 }
2713
2714 blocksize = root->nodesize;
2715 generation = btrfs_node_ptr_generation(upper->eb, slot);
2716 eb = read_tree_block(root, bytenr, generation);
2717 if (IS_ERR(eb)) {
2718 err = PTR_ERR(eb);
2719 goto next;
2720 } else if (!extent_buffer_uptodate(eb)) {
2721 free_extent_buffer(eb);
2722 err = -EIO;
2723 goto next;
2724 }
2725 btrfs_tree_lock(eb);
2726 btrfs_set_lock_blocking(eb);
2727
2728 if (!node->eb) {
2729 ret = btrfs_cow_block(trans, root, eb, upper->eb,
2730 slot, &eb);
2731 btrfs_tree_unlock(eb);
2732 free_extent_buffer(eb);
2733 if (ret < 0) {
2734 err = ret;
2735 goto next;
2736 }
2737 BUG_ON(node->eb != eb);
2738 } else {
2739 btrfs_set_node_blockptr(upper->eb, slot,
2740 node->eb->start);
2741 btrfs_set_node_ptr_generation(upper->eb, slot,
2742 trans->transid);
2743 btrfs_mark_buffer_dirty(upper->eb);
2744
2745 ret = btrfs_inc_extent_ref(trans, root,
2746 node->eb->start, blocksize,
2747 upper->eb->start,
2748 btrfs_header_owner(upper->eb),
2749 node->level, 0, 1);
2750 BUG_ON(ret);
2751
2752 ret = btrfs_drop_subtree(trans, root, eb, upper->eb);
2753 BUG_ON(ret);
2754 }
2755 next:
2756 if (!upper->pending)
2757 drop_node_buffer(upper);
2758 else
2759 unlock_node_buffer(upper);
2760 if (err)
2761 break;
2762 }
2763
2764 if (!err && node->pending) {
2765 drop_node_buffer(node);
2766 list_move_tail(&node->list, &rc->backref_cache.changed);
2767 node->pending = 0;
2768 }
2769
2770 path->lowest_level = 0;
2771 BUG_ON(err == -ENOSPC);
2772 return err;
2773 }
2774
2775 static int link_to_upper(struct btrfs_trans_handle *trans,
2776 struct reloc_control *rc,
2777 struct backref_node *node,
2778 struct btrfs_path *path)
2779 {
2780 struct btrfs_key key;
2781
2782 btrfs_node_key_to_cpu(node->eb, &key, 0);
2783 return do_relocation(trans, rc, node, &key, path, 0);
2784 }
2785
2786 static int finish_pending_nodes(struct btrfs_trans_handle *trans,
2787 struct reloc_control *rc,
2788 struct btrfs_path *path, int err)
2789 {
2790 LIST_HEAD(list);
2791 struct backref_cache *cache = &rc->backref_cache;
2792 struct backref_node *node;
2793 int level;
2794 int ret;
2795
2796 for (level = 0; level < BTRFS_MAX_LEVEL; level++) {
2797 while (!list_empty(&cache->pending[level])) {
2798 node = list_entry(cache->pending[level].next,
2799 struct backref_node, list);
2800 list_move_tail(&node->list, &list);
2801 BUG_ON(!node->pending);
2802
2803 if (!err) {
2804 ret = link_to_upper(trans, rc, node, path);
2805 if (ret < 0)
2806 err = ret;
2807 }
2808 }
2809 list_splice_init(&list, &cache->pending[level]);
2810 }
2811 return err;
2812 }
2813
2814 static void mark_block_processed(struct reloc_control *rc,
2815 u64 bytenr, u32 blocksize)
2816 {
2817 set_extent_bits(&rc->processed_blocks, bytenr, bytenr + blocksize - 1,
2818 EXTENT_DIRTY, GFP_NOFS);
2819 }
2820
2821 static void __mark_block_processed(struct reloc_control *rc,
2822 struct backref_node *node)
2823 {
2824 u32 blocksize;
2825 if (node->level == 0 ||
2826 in_block_group(node->bytenr, rc->block_group)) {
2827 blocksize = rc->extent_root->nodesize;
2828 mark_block_processed(rc, node->bytenr, blocksize);
2829 }
2830 node->processed = 1;
2831 }
2832
2833 /*
2834 * mark a block and all blocks directly/indirectly reference the block
2835 * as processed.
2836 */
2837 static void update_processed_blocks(struct reloc_control *rc,
2838 struct backref_node *node)
2839 {
2840 struct backref_node *next = node;
2841 struct backref_edge *edge;
2842 struct backref_edge *edges[BTRFS_MAX_LEVEL - 1];
2843 int index = 0;
2844
2845 while (next) {
2846 cond_resched();
2847 while (1) {
2848 if (next->processed)
2849 break;
2850
2851 __mark_block_processed(rc, next);
2852
2853 if (list_empty(&next->upper))
2854 break;
2855
2856 edge = list_entry(next->upper.next,
2857 struct backref_edge, list[LOWER]);
2858 edges[index++] = edge;
2859 next = edge->node[UPPER];
2860 }
2861 next = walk_down_backref(edges, &index);
2862 }
2863 }
2864
2865 static int tree_block_processed(u64 bytenr, struct reloc_control *rc)
2866 {
2867 u32 blocksize = rc->extent_root->nodesize;
2868
2869 if (test_range_bit(&rc->processed_blocks, bytenr,
2870 bytenr + blocksize - 1, EXTENT_DIRTY, 1, NULL))
2871 return 1;
2872 return 0;
2873 }
2874
2875 static int get_tree_block_key(struct reloc_control *rc,
2876 struct tree_block *block)
2877 {
2878 struct extent_buffer *eb;
2879
2880 BUG_ON(block->key_ready);
2881 eb = read_tree_block(rc->extent_root, block->bytenr,
2882 block->key.offset);
2883 if (IS_ERR(eb)) {
2884 return PTR_ERR(eb);
2885 } else if (!extent_buffer_uptodate(eb)) {
2886 free_extent_buffer(eb);
2887 return -EIO;
2888 }
2889 WARN_ON(btrfs_header_level(eb) != block->level);
2890 if (block->level == 0)
2891 btrfs_item_key_to_cpu(eb, &block->key, 0);
2892 else
2893 btrfs_node_key_to_cpu(eb, &block->key, 0);
2894 free_extent_buffer(eb);
2895 block->key_ready = 1;
2896 return 0;
2897 }
2898
2899 /*
2900 * helper function to relocate a tree block
2901 */
2902 static int relocate_tree_block(struct btrfs_trans_handle *trans,
2903 struct reloc_control *rc,
2904 struct backref_node *node,
2905 struct btrfs_key *key,
2906 struct btrfs_path *path)
2907 {
2908 struct btrfs_root *root;
2909 int ret = 0;
2910
2911 if (!node)
2912 return 0;
2913
2914 BUG_ON(node->processed);
2915 root = select_one_root(trans, node);
2916 if (root == ERR_PTR(-ENOENT)) {
2917 update_processed_blocks(rc, node);
2918 goto out;
2919 }
2920
2921 if (!root || test_bit(BTRFS_ROOT_REF_COWS, &root->state)) {
2922 ret = reserve_metadata_space(trans, rc, node);
2923 if (ret)
2924 goto out;
2925 }
2926
2927 if (root) {
2928 if (test_bit(BTRFS_ROOT_REF_COWS, &root->state)) {
2929 BUG_ON(node->new_bytenr);
2930 BUG_ON(!list_empty(&node->list));
2931 btrfs_record_root_in_trans(trans, root);
2932 root = root->reloc_root;
2933 node->new_bytenr = root->node->start;
2934 node->root = root;
2935 list_add_tail(&node->list, &rc->backref_cache.changed);
2936 } else {
2937 path->lowest_level = node->level;
2938 ret = btrfs_search_slot(trans, root, key, path, 0, 1);
2939 btrfs_release_path(path);
2940 if (ret > 0)
2941 ret = 0;
2942 }
2943 if (!ret)
2944 update_processed_blocks(rc, node);
2945 } else {
2946 ret = do_relocation(trans, rc, node, key, path, 1);
2947 }
2948 out:
2949 if (ret || node->level == 0 || node->cowonly)
2950 remove_backref_node(&rc->backref_cache, node);
2951 return ret;
2952 }
2953
2954 /*
2955 * relocate a list of blocks
2956 */
2957 static noinline_for_stack
2958 int relocate_tree_blocks(struct btrfs_trans_handle *trans,
2959 struct reloc_control *rc, struct rb_root *blocks)
2960 {
2961 struct backref_node *node;
2962 struct btrfs_path *path;
2963 struct tree_block *block;
2964 struct rb_node *rb_node;
2965 int ret;
2966 int err = 0;
2967
2968 path = btrfs_alloc_path();
2969 if (!path) {
2970 err = -ENOMEM;
2971 goto out_free_blocks;
2972 }
2973
2974 rb_node = rb_first(blocks);
2975 while (rb_node) {
2976 block = rb_entry(rb_node, struct tree_block, rb_node);
2977 if (!block->key_ready)
2978 readahead_tree_block(rc->extent_root, block->bytenr);
2979 rb_node = rb_next(rb_node);
2980 }
2981
2982 rb_node = rb_first(blocks);
2983 while (rb_node) {
2984 block = rb_entry(rb_node, struct tree_block, rb_node);
2985 if (!block->key_ready) {
2986 err = get_tree_block_key(rc, block);
2987 if (err)
2988 goto out_free_path;
2989 }
2990 rb_node = rb_next(rb_node);
2991 }
2992
2993 rb_node = rb_first(blocks);
2994 while (rb_node) {
2995 block = rb_entry(rb_node, struct tree_block, rb_node);
2996
2997 node = build_backref_tree(rc, &block->key,
2998 block->level, block->bytenr);
2999 if (IS_ERR(node)) {
3000 err = PTR_ERR(node);
3001 goto out;
3002 }
3003
3004 ret = relocate_tree_block(trans, rc, node, &block->key,
3005 path);
3006 if (ret < 0) {
3007 if (ret != -EAGAIN || rb_node == rb_first(blocks))
3008 err = ret;
3009 goto out;
3010 }
3011 rb_node = rb_next(rb_node);
3012 }
3013 out:
3014 err = finish_pending_nodes(trans, rc, path, err);
3015
3016 out_free_path:
3017 btrfs_free_path(path);
3018 out_free_blocks:
3019 free_block_list(blocks);
3020 return err;
3021 }
3022
3023 static noinline_for_stack
3024 int prealloc_file_extent_cluster(struct inode *inode,
3025 struct file_extent_cluster *cluster)
3026 {
3027 u64 alloc_hint = 0;
3028 u64 start;
3029 u64 end;
3030 u64 offset = BTRFS_I(inode)->index_cnt;
3031 u64 num_bytes;
3032 int nr = 0;
3033 int ret = 0;
3034
3035 BUG_ON(cluster->start != cluster->boundary[0]);
3036 mutex_lock(&inode->i_mutex);
3037
3038 ret = btrfs_check_data_free_space(inode, cluster->end +
3039 1 - cluster->start, 0);
3040 if (ret)
3041 goto out;
3042
3043 while (nr < cluster->nr) {
3044 start = cluster->boundary[nr] - offset;
3045 if (nr + 1 < cluster->nr)
3046 end = cluster->boundary[nr + 1] - 1 - offset;
3047 else
3048 end = cluster->end - offset;
3049
3050 lock_extent(&BTRFS_I(inode)->io_tree, start, end);
3051 num_bytes = end + 1 - start;
3052 ret = btrfs_prealloc_file_range(inode, 0, start,
3053 num_bytes, num_bytes,
3054 end + 1, &alloc_hint);
3055 unlock_extent(&BTRFS_I(inode)->io_tree, start, end);
3056 if (ret)
3057 break;
3058 nr++;
3059 }
3060 btrfs_free_reserved_data_space(inode, cluster->end +
3061 1 - cluster->start);
3062 out:
3063 mutex_unlock(&inode->i_mutex);
3064 return ret;
3065 }
3066
3067 static noinline_for_stack
3068 int setup_extent_mapping(struct inode *inode, u64 start, u64 end,
3069 u64 block_start)
3070 {
3071 struct btrfs_root *root = BTRFS_I(inode)->root;
3072 struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
3073 struct extent_map *em;
3074 int ret = 0;
3075
3076 em = alloc_extent_map();
3077 if (!em)
3078 return -ENOMEM;
3079
3080 em->start = start;
3081 em->len = end + 1 - start;
3082 em->block_len = em->len;
3083 em->block_start = block_start;
3084 em->bdev = root->fs_info->fs_devices->latest_bdev;
3085 set_bit(EXTENT_FLAG_PINNED, &em->flags);
3086
3087 lock_extent(&BTRFS_I(inode)->io_tree, start, end);
3088 while (1) {
3089 write_lock(&em_tree->lock);
3090 ret = add_extent_mapping(em_tree, em, 0);
3091 write_unlock(&em_tree->lock);
3092 if (ret != -EEXIST) {
3093 free_extent_map(em);
3094 break;
3095 }
3096 btrfs_drop_extent_cache(inode, start, end, 0);
3097 }
3098 unlock_extent(&BTRFS_I(inode)->io_tree, start, end);
3099 return ret;
3100 }
3101
3102 static int relocate_file_extent_cluster(struct inode *inode,
3103 struct file_extent_cluster *cluster)
3104 {
3105 u64 page_start;
3106 u64 page_end;
3107 u64 offset = BTRFS_I(inode)->index_cnt;
3108 unsigned long index;
3109 unsigned long last_index;
3110 struct page *page;
3111 struct file_ra_state *ra;
3112 gfp_t mask = btrfs_alloc_write_mask(inode->i_mapping);
3113 int nr = 0;
3114 int ret = 0;
3115
3116 if (!cluster->nr)
3117 return 0;
3118
3119 ra = kzalloc(sizeof(*ra), GFP_NOFS);
3120 if (!ra)
3121 return -ENOMEM;
3122
3123 ret = prealloc_file_extent_cluster(inode, cluster);
3124 if (ret)
3125 goto out;
3126
3127 file_ra_state_init(ra, inode->i_mapping);
3128
3129 ret = setup_extent_mapping(inode, cluster->start - offset,
3130 cluster->end - offset, cluster->start);
3131 if (ret)
3132 goto out;
3133
3134 index = (cluster->start - offset) >> PAGE_CACHE_SHIFT;
3135 last_index = (cluster->end - offset) >> PAGE_CACHE_SHIFT;
3136 while (index <= last_index) {
3137 ret = btrfs_delalloc_reserve_metadata(inode, PAGE_CACHE_SIZE);
3138 if (ret)
3139 goto out;
3140
3141 page = find_lock_page(inode->i_mapping, index);
3142 if (!page) {
3143 page_cache_sync_readahead(inode->i_mapping,
3144 ra, NULL, index,
3145 last_index + 1 - index);
3146 page = find_or_create_page(inode->i_mapping, index,
3147 mask);
3148 if (!page) {
3149 btrfs_delalloc_release_metadata(inode,
3150 PAGE_CACHE_SIZE);
3151 ret = -ENOMEM;
3152 goto out;
3153 }
3154 }
3155
3156 if (PageReadahead(page)) {
3157 page_cache_async_readahead(inode->i_mapping,
3158 ra, NULL, page, index,
3159 last_index + 1 - index);
3160 }
3161
3162 if (!PageUptodate(page)) {
3163 btrfs_readpage(NULL, page);
3164 lock_page(page);
3165 if (!PageUptodate(page)) {
3166 unlock_page(page);
3167 page_cache_release(page);
3168 btrfs_delalloc_release_metadata(inode,
3169 PAGE_CACHE_SIZE);
3170 ret = -EIO;
3171 goto out;
3172 }
3173 }
3174
3175 page_start = page_offset(page);
3176 page_end = page_start + PAGE_CACHE_SIZE - 1;
3177
3178 lock_extent(&BTRFS_I(inode)->io_tree, page_start, page_end);
3179
3180 set_page_extent_mapped(page);
3181
3182 if (nr < cluster->nr &&
3183 page_start + offset == cluster->boundary[nr]) {
3184 set_extent_bits(&BTRFS_I(inode)->io_tree,
3185 page_start, page_end,
3186 EXTENT_BOUNDARY, GFP_NOFS);
3187 nr++;
3188 }
3189
3190 btrfs_set_extent_delalloc(inode, page_start, page_end, NULL);
3191 set_page_dirty(page);
3192
3193 unlock_extent(&BTRFS_I(inode)->io_tree,
3194 page_start, page_end);
3195 unlock_page(page);
3196 page_cache_release(page);
3197
3198 index++;
3199 balance_dirty_pages_ratelimited(inode->i_mapping);
3200 btrfs_throttle(BTRFS_I(inode)->root);
3201 }
3202 WARN_ON(nr != cluster->nr);
3203 out:
3204 kfree(ra);
3205 return ret;
3206 }
3207
3208 static noinline_for_stack
3209 int relocate_data_extent(struct inode *inode, struct btrfs_key *extent_key,
3210 struct file_extent_cluster *cluster)
3211 {
3212 int ret;
3213
3214 if (cluster->nr > 0 && extent_key->objectid != cluster->end + 1) {
3215 ret = relocate_file_extent_cluster(inode, cluster);
3216 if (ret)
3217 return ret;
3218 cluster->nr = 0;
3219 }
3220
3221 if (!cluster->nr)
3222 cluster->start = extent_key->objectid;
3223 else
3224 BUG_ON(cluster->nr >= MAX_EXTENTS);
3225 cluster->end = extent_key->objectid + extent_key->offset - 1;
3226 cluster->boundary[cluster->nr] = extent_key->objectid;
3227 cluster->nr++;
3228
3229 if (cluster->nr >= MAX_EXTENTS) {
3230 ret = relocate_file_extent_cluster(inode, cluster);
3231 if (ret)
3232 return ret;
3233 cluster->nr = 0;
3234 }
3235 return 0;
3236 }
3237
3238 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
3239 static int get_ref_objectid_v0(struct reloc_control *rc,
3240 struct btrfs_path *path,
3241 struct btrfs_key *extent_key,
3242 u64 *ref_objectid, int *path_change)
3243 {
3244 struct btrfs_key key;
3245 struct extent_buffer *leaf;
3246 struct btrfs_extent_ref_v0 *ref0;
3247 int ret;
3248 int slot;
3249
3250 leaf = path->nodes[0];
3251 slot = path->slots[0];
3252 while (1) {
3253 if (slot >= btrfs_header_nritems(leaf)) {
3254 ret = btrfs_next_leaf(rc->extent_root, path);
3255 if (ret < 0)
3256 return ret;
3257 BUG_ON(ret > 0);
3258 leaf = path->nodes[0];
3259 slot = path->slots[0];
3260 if (path_change)
3261 *path_change = 1;
3262 }
3263 btrfs_item_key_to_cpu(leaf, &key, slot);
3264 if (key.objectid != extent_key->objectid)
3265 return -ENOENT;
3266
3267 if (key.type != BTRFS_EXTENT_REF_V0_KEY) {
3268 slot++;
3269 continue;
3270 }
3271 ref0 = btrfs_item_ptr(leaf, slot,
3272 struct btrfs_extent_ref_v0);
3273 *ref_objectid = btrfs_ref_objectid_v0(leaf, ref0);
3274 break;
3275 }
3276 return 0;
3277 }
3278 #endif
3279
3280 /*
3281 * helper to add a tree block to the list.
3282 * the major work is getting the generation and level of the block
3283 */
3284 static int add_tree_block(struct reloc_control *rc,
3285 struct btrfs_key *extent_key,
3286 struct btrfs_path *path,
3287 struct rb_root *blocks)
3288 {
3289 struct extent_buffer *eb;
3290 struct btrfs_extent_item *ei;
3291 struct btrfs_tree_block_info *bi;
3292 struct tree_block *block;
3293 struct rb_node *rb_node;
3294 u32 item_size;
3295 int level = -1;
3296 u64 generation;
3297
3298 eb = path->nodes[0];
3299 item_size = btrfs_item_size_nr(eb, path->slots[0]);
3300
3301 if (extent_key->type == BTRFS_METADATA_ITEM_KEY ||
3302 item_size >= sizeof(*ei) + sizeof(*bi)) {
3303 ei = btrfs_item_ptr(eb, path->slots[0],
3304 struct btrfs_extent_item);
3305 if (extent_key->type == BTRFS_EXTENT_ITEM_KEY) {
3306 bi = (struct btrfs_tree_block_info *)(ei + 1);
3307 level = btrfs_tree_block_level(eb, bi);
3308 } else {
3309 level = (int)extent_key->offset;
3310 }
3311 generation = btrfs_extent_generation(eb, ei);
3312 } else {
3313 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
3314 u64 ref_owner;
3315 int ret;
3316
3317 BUG_ON(item_size != sizeof(struct btrfs_extent_item_v0));
3318 ret = get_ref_objectid_v0(rc, path, extent_key,
3319 &ref_owner, NULL);
3320 if (ret < 0)
3321 return ret;
3322 BUG_ON(ref_owner >= BTRFS_MAX_LEVEL);
3323 level = (int)ref_owner;
3324 /* FIXME: get real generation */
3325 generation = 0;
3326 #else
3327 BUG();
3328 #endif
3329 }
3330
3331 btrfs_release_path(path);
3332
3333 BUG_ON(level == -1);
3334
3335 block = kmalloc(sizeof(*block), GFP_NOFS);
3336 if (!block)
3337 return -ENOMEM;
3338
3339 block->bytenr = extent_key->objectid;
3340 block->key.objectid = rc->extent_root->nodesize;
3341 block->key.offset = generation;
3342 block->level = level;
3343 block->key_ready = 0;
3344
3345 rb_node = tree_insert(blocks, block->bytenr, &block->rb_node);
3346 if (rb_node)
3347 backref_tree_panic(rb_node, -EEXIST, block->bytenr);
3348
3349 return 0;
3350 }
3351
3352 /*
3353 * helper to add tree blocks for backref of type BTRFS_SHARED_DATA_REF_KEY
3354 */
3355 static int __add_tree_block(struct reloc_control *rc,
3356 u64 bytenr, u32 blocksize,
3357 struct rb_root *blocks)
3358 {
3359 struct btrfs_path *path;
3360 struct btrfs_key key;
3361 int ret;
3362 bool skinny = btrfs_fs_incompat(rc->extent_root->fs_info,
3363 SKINNY_METADATA);
3364
3365 if (tree_block_processed(bytenr, rc))
3366 return 0;
3367
3368 if (tree_search(blocks, bytenr))
3369 return 0;
3370
3371 path = btrfs_alloc_path();
3372 if (!path)
3373 return -ENOMEM;
3374 again:
3375 key.objectid = bytenr;
3376 if (skinny) {
3377 key.type = BTRFS_METADATA_ITEM_KEY;
3378 key.offset = (u64)-1;
3379 } else {
3380 key.type = BTRFS_EXTENT_ITEM_KEY;
3381 key.offset = blocksize;
3382 }
3383
3384 path->search_commit_root = 1;
3385 path->skip_locking = 1;
3386 ret = btrfs_search_slot(NULL, rc->extent_root, &key, path, 0, 0);
3387 if (ret < 0)
3388 goto out;
3389
3390 if (ret > 0 && skinny) {
3391 if (path->slots[0]) {
3392 path->slots[0]--;
3393 btrfs_item_key_to_cpu(path->nodes[0], &key,
3394 path->slots[0]);
3395 if (key.objectid == bytenr &&
3396 (key.type == BTRFS_METADATA_ITEM_KEY ||
3397 (key.type == BTRFS_EXTENT_ITEM_KEY &&
3398 key.offset == blocksize)))
3399 ret = 0;
3400 }
3401
3402 if (ret) {
3403 skinny = false;
3404 btrfs_release_path(path);
3405 goto again;
3406 }
3407 }
3408 BUG_ON(ret);
3409
3410 ret = add_tree_block(rc, &key, path, blocks);
3411 out:
3412 btrfs_free_path(path);
3413 return ret;
3414 }
3415
3416 /*
3417 * helper to check if the block use full backrefs for pointers in it
3418 */
3419 static int block_use_full_backref(struct reloc_control *rc,
3420 struct extent_buffer *eb)
3421 {
3422 u64 flags;
3423 int ret;
3424
3425 if (btrfs_header_flag(eb, BTRFS_HEADER_FLAG_RELOC) ||
3426 btrfs_header_backref_rev(eb) < BTRFS_MIXED_BACKREF_REV)
3427 return 1;
3428
3429 ret = btrfs_lookup_extent_info(NULL, rc->extent_root,
3430 eb->start, btrfs_header_level(eb), 1,
3431 NULL, &flags);
3432 BUG_ON(ret);
3433
3434 if (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF)
3435 ret = 1;
3436 else
3437 ret = 0;
3438 return ret;
3439 }
3440
3441 static int delete_block_group_cache(struct btrfs_fs_info *fs_info,
3442 struct btrfs_block_group_cache *block_group,
3443 struct inode *inode,
3444 u64 ino)
3445 {
3446 struct btrfs_key key;
3447 struct btrfs_root *root = fs_info->tree_root;
3448 struct btrfs_trans_handle *trans;
3449 int ret = 0;
3450
3451 if (inode)
3452 goto truncate;
3453
3454 key.objectid = ino;
3455 key.type = BTRFS_INODE_ITEM_KEY;
3456 key.offset = 0;
3457
3458 inode = btrfs_iget(fs_info->sb, &key, root, NULL);
3459 if (IS_ERR(inode) || is_bad_inode(inode)) {
3460 if (!IS_ERR(inode))
3461 iput(inode);
3462 return -ENOENT;
3463 }
3464
3465 truncate:
3466 ret = btrfs_check_trunc_cache_free_space(root,
3467 &fs_info->global_block_rsv);
3468 if (ret)
3469 goto out;
3470
3471 trans = btrfs_join_transaction(root);
3472 if (IS_ERR(trans)) {
3473 ret = PTR_ERR(trans);
3474 goto out;
3475 }
3476
3477 ret = btrfs_truncate_free_space_cache(root, trans, block_group, inode);
3478
3479 btrfs_end_transaction(trans, root);
3480 btrfs_btree_balance_dirty(root);
3481 out:
3482 iput(inode);
3483 return ret;
3484 }
3485
3486 /*
3487 * helper to add tree blocks for backref of type BTRFS_EXTENT_DATA_REF_KEY
3488 * this function scans fs tree to find blocks reference the data extent
3489 */
3490 static int find_data_references(struct reloc_control *rc,
3491 struct btrfs_key *extent_key,
3492 struct extent_buffer *leaf,
3493 struct btrfs_extent_data_ref *ref,
3494 struct rb_root *blocks)
3495 {
3496 struct btrfs_path *path;
3497 struct tree_block *block;
3498 struct btrfs_root *root;
3499 struct btrfs_file_extent_item *fi;
3500 struct rb_node *rb_node;
3501 struct btrfs_key key;
3502 u64 ref_root;
3503 u64 ref_objectid;
3504 u64 ref_offset;
3505 u32 ref_count;
3506 u32 nritems;
3507 int err = 0;
3508 int added = 0;
3509 int counted;
3510 int ret;
3511
3512 ref_root = btrfs_extent_data_ref_root(leaf, ref);
3513 ref_objectid = btrfs_extent_data_ref_objectid(leaf, ref);
3514 ref_offset = btrfs_extent_data_ref_offset(leaf, ref);
3515 ref_count = btrfs_extent_data_ref_count(leaf, ref);
3516
3517 /*
3518 * This is an extent belonging to the free space cache, lets just delete
3519 * it and redo the search.
3520 */
3521 if (ref_root == BTRFS_ROOT_TREE_OBJECTID) {
3522 ret = delete_block_group_cache(rc->extent_root->fs_info,
3523 rc->block_group,
3524 NULL, ref_objectid);
3525 if (ret != -ENOENT)
3526 return ret;
3527 ret = 0;
3528 }
3529
3530 path = btrfs_alloc_path();
3531 if (!path)
3532 return -ENOMEM;
3533 path->reada = 1;
3534
3535 root = read_fs_root(rc->extent_root->fs_info, ref_root);
3536 if (IS_ERR(root)) {
3537 err = PTR_ERR(root);
3538 goto out;
3539 }
3540
3541 key.objectid = ref_objectid;
3542 key.type = BTRFS_EXTENT_DATA_KEY;
3543 if (ref_offset > ((u64)-1 << 32))
3544 key.offset = 0;
3545 else
3546 key.offset = ref_offset;
3547
3548 path->search_commit_root = 1;
3549 path->skip_locking = 1;
3550 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
3551 if (ret < 0) {
3552 err = ret;
3553 goto out;
3554 }
3555
3556 leaf = path->nodes[0];
3557 nritems = btrfs_header_nritems(leaf);
3558 /*
3559 * the references in tree blocks that use full backrefs
3560 * are not counted in
3561 */
3562 if (block_use_full_backref(rc, leaf))
3563 counted = 0;
3564 else
3565 counted = 1;
3566 rb_node = tree_search(blocks, leaf->start);
3567 if (rb_node) {
3568 if (counted)
3569 added = 1;
3570 else
3571 path->slots[0] = nritems;
3572 }
3573
3574 while (ref_count > 0) {
3575 while (path->slots[0] >= nritems) {
3576 ret = btrfs_next_leaf(root, path);
3577 if (ret < 0) {
3578 err = ret;
3579 goto out;
3580 }
3581 if (WARN_ON(ret > 0))
3582 goto out;
3583
3584 leaf = path->nodes[0];
3585 nritems = btrfs_header_nritems(leaf);
3586 added = 0;
3587
3588 if (block_use_full_backref(rc, leaf))
3589 counted = 0;
3590 else
3591 counted = 1;
3592 rb_node = tree_search(blocks, leaf->start);
3593 if (rb_node) {
3594 if (counted)
3595 added = 1;
3596 else
3597 path->slots[0] = nritems;
3598 }
3599 }
3600
3601 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
3602 if (WARN_ON(key.objectid != ref_objectid ||
3603 key.type != BTRFS_EXTENT_DATA_KEY))
3604 break;
3605
3606 fi = btrfs_item_ptr(leaf, path->slots[0],
3607 struct btrfs_file_extent_item);
3608
3609 if (btrfs_file_extent_type(leaf, fi) ==
3610 BTRFS_FILE_EXTENT_INLINE)
3611 goto next;
3612
3613 if (btrfs_file_extent_disk_bytenr(leaf, fi) !=
3614 extent_key->objectid)
3615 goto next;
3616
3617 key.offset -= btrfs_file_extent_offset(leaf, fi);
3618 if (key.offset != ref_offset)
3619 goto next;
3620
3621 if (counted)
3622 ref_count--;
3623 if (added)
3624 goto next;
3625
3626 if (!tree_block_processed(leaf->start, rc)) {
3627 block = kmalloc(sizeof(*block), GFP_NOFS);
3628 if (!block) {
3629 err = -ENOMEM;
3630 break;
3631 }
3632 block->bytenr = leaf->start;
3633 btrfs_item_key_to_cpu(leaf, &block->key, 0);
3634 block->level = 0;
3635 block->key_ready = 1;
3636 rb_node = tree_insert(blocks, block->bytenr,
3637 &block->rb_node);
3638 if (rb_node)
3639 backref_tree_panic(rb_node, -EEXIST,
3640 block->bytenr);
3641 }
3642 if (counted)
3643 added = 1;
3644 else
3645 path->slots[0] = nritems;
3646 next:
3647 path->slots[0]++;
3648
3649 }
3650 out:
3651 btrfs_free_path(path);
3652 return err;
3653 }
3654
3655 /*
3656 * helper to find all tree blocks that reference a given data extent
3657 */
3658 static noinline_for_stack
3659 int add_data_references(struct reloc_control *rc,
3660 struct btrfs_key *extent_key,
3661 struct btrfs_path *path,
3662 struct rb_root *blocks)
3663 {
3664 struct btrfs_key key;
3665 struct extent_buffer *eb;
3666 struct btrfs_extent_data_ref *dref;
3667 struct btrfs_extent_inline_ref *iref;
3668 unsigned long ptr;
3669 unsigned long end;
3670 u32 blocksize = rc->extent_root->nodesize;
3671 int ret = 0;
3672 int err = 0;
3673
3674 eb = path->nodes[0];
3675 ptr = btrfs_item_ptr_offset(eb, path->slots[0]);
3676 end = ptr + btrfs_item_size_nr(eb, path->slots[0]);
3677 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
3678 if (ptr + sizeof(struct btrfs_extent_item_v0) == end)
3679 ptr = end;
3680 else
3681 #endif
3682 ptr += sizeof(struct btrfs_extent_item);
3683
3684 while (ptr < end) {
3685 iref = (struct btrfs_extent_inline_ref *)ptr;
3686 key.type = btrfs_extent_inline_ref_type(eb, iref);
3687 if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
3688 key.offset = btrfs_extent_inline_ref_offset(eb, iref);
3689 ret = __add_tree_block(rc, key.offset, blocksize,
3690 blocks);
3691 } else if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
3692 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
3693 ret = find_data_references(rc, extent_key,
3694 eb, dref, blocks);
3695 } else {
3696 BUG();
3697 }
3698 if (ret) {
3699 err = ret;
3700 goto out;
3701 }
3702 ptr += btrfs_extent_inline_ref_size(key.type);
3703 }
3704 WARN_ON(ptr > end);
3705
3706 while (1) {
3707 cond_resched();
3708 eb = path->nodes[0];
3709 if (path->slots[0] >= btrfs_header_nritems(eb)) {
3710 ret = btrfs_next_leaf(rc->extent_root, path);
3711 if (ret < 0) {
3712 err = ret;
3713 break;
3714 }
3715 if (ret > 0)
3716 break;
3717 eb = path->nodes[0];
3718 }
3719
3720 btrfs_item_key_to_cpu(eb, &key, path->slots[0]);
3721 if (key.objectid != extent_key->objectid)
3722 break;
3723
3724 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
3725 if (key.type == BTRFS_SHARED_DATA_REF_KEY ||
3726 key.type == BTRFS_EXTENT_REF_V0_KEY) {
3727 #else
3728 BUG_ON(key.type == BTRFS_EXTENT_REF_V0_KEY);
3729 if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
3730 #endif
3731 ret = __add_tree_block(rc, key.offset, blocksize,
3732 blocks);
3733 } else if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
3734 dref = btrfs_item_ptr(eb, path->slots[0],
3735 struct btrfs_extent_data_ref);
3736 ret = find_data_references(rc, extent_key,
3737 eb, dref, blocks);
3738 } else {
3739 ret = 0;
3740 }
3741 if (ret) {
3742 err = ret;
3743 break;
3744 }
3745 path->slots[0]++;
3746 }
3747 out:
3748 btrfs_release_path(path);
3749 if (err)
3750 free_block_list(blocks);
3751 return err;
3752 }
3753
3754 /*
3755 * helper to find next unprocessed extent
3756 */
3757 static noinline_for_stack
3758 int find_next_extent(struct btrfs_trans_handle *trans,
3759 struct reloc_control *rc, struct btrfs_path *path,
3760 struct btrfs_key *extent_key)
3761 {
3762 struct btrfs_key key;
3763 struct extent_buffer *leaf;
3764 u64 start, end, last;
3765 int ret;
3766
3767 last = rc->block_group->key.objectid + rc->block_group->key.offset;
3768 while (1) {
3769 cond_resched();
3770 if (rc->search_start >= last) {
3771 ret = 1;
3772 break;
3773 }
3774
3775 key.objectid = rc->search_start;
3776 key.type = BTRFS_EXTENT_ITEM_KEY;
3777 key.offset = 0;
3778
3779 path->search_commit_root = 1;
3780 path->skip_locking = 1;
3781 ret = btrfs_search_slot(NULL, rc->extent_root, &key, path,
3782 0, 0);
3783 if (ret < 0)
3784 break;
3785 next:
3786 leaf = path->nodes[0];
3787 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
3788 ret = btrfs_next_leaf(rc->extent_root, path);
3789 if (ret != 0)
3790 break;
3791 leaf = path->nodes[0];
3792 }
3793
3794 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
3795 if (key.objectid >= last) {
3796 ret = 1;
3797 break;
3798 }
3799
3800 if (key.type != BTRFS_EXTENT_ITEM_KEY &&
3801 key.type != BTRFS_METADATA_ITEM_KEY) {
3802 path->slots[0]++;
3803 goto next;
3804 }
3805
3806 if (key.type == BTRFS_EXTENT_ITEM_KEY &&
3807 key.objectid + key.offset <= rc->search_start) {
3808 path->slots[0]++;
3809 goto next;
3810 }
3811
3812 if (key.type == BTRFS_METADATA_ITEM_KEY &&
3813 key.objectid + rc->extent_root->nodesize <=
3814 rc->search_start) {
3815 path->slots[0]++;
3816 goto next;
3817 }
3818
3819 ret = find_first_extent_bit(&rc->processed_blocks,
3820 key.objectid, &start, &end,
3821 EXTENT_DIRTY, NULL);
3822
3823 if (ret == 0 && start <= key.objectid) {
3824 btrfs_release_path(path);
3825 rc->search_start = end + 1;
3826 } else {
3827 if (key.type == BTRFS_EXTENT_ITEM_KEY)
3828 rc->search_start = key.objectid + key.offset;
3829 else
3830 rc->search_start = key.objectid +
3831 rc->extent_root->nodesize;
3832 memcpy(extent_key, &key, sizeof(key));
3833 return 0;
3834 }
3835 }
3836 btrfs_release_path(path);
3837 return ret;
3838 }
3839
3840 static void set_reloc_control(struct reloc_control *rc)
3841 {
3842 struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
3843
3844 mutex_lock(&fs_info->reloc_mutex);
3845 fs_info->reloc_ctl = rc;
3846 mutex_unlock(&fs_info->reloc_mutex);
3847 }
3848
3849 static void unset_reloc_control(struct reloc_control *rc)
3850 {
3851 struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
3852
3853 mutex_lock(&fs_info->reloc_mutex);
3854 fs_info->reloc_ctl = NULL;
3855 mutex_unlock(&fs_info->reloc_mutex);
3856 }
3857
3858 static int check_extent_flags(u64 flags)
3859 {
3860 if ((flags & BTRFS_EXTENT_FLAG_DATA) &&
3861 (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK))
3862 return 1;
3863 if (!(flags & BTRFS_EXTENT_FLAG_DATA) &&
3864 !(flags & BTRFS_EXTENT_FLAG_TREE_BLOCK))
3865 return 1;
3866 if ((flags & BTRFS_EXTENT_FLAG_DATA) &&
3867 (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
3868 return 1;
3869 return 0;
3870 }
3871
3872 static noinline_for_stack
3873 int prepare_to_relocate(struct reloc_control *rc)
3874 {
3875 struct btrfs_trans_handle *trans;
3876
3877 rc->block_rsv = btrfs_alloc_block_rsv(rc->extent_root,
3878 BTRFS_BLOCK_RSV_TEMP);
3879 if (!rc->block_rsv)
3880 return -ENOMEM;
3881
3882 memset(&rc->cluster, 0, sizeof(rc->cluster));
3883 rc->search_start = rc->block_group->key.objectid;
3884 rc->extents_found = 0;
3885 rc->nodes_relocated = 0;
3886 rc->merging_rsv_size = 0;
3887 rc->reserved_bytes = 0;
3888 rc->block_rsv->size = rc->extent_root->nodesize *
3889 RELOCATION_RESERVED_NODES;
3890
3891 rc->create_reloc_tree = 1;
3892 set_reloc_control(rc);
3893
3894 trans = btrfs_join_transaction(rc->extent_root);
3895 if (IS_ERR(trans)) {
3896 unset_reloc_control(rc);
3897 /*
3898 * extent tree is not a ref_cow tree and has no reloc_root to
3899 * cleanup. And callers are responsible to free the above
3900 * block rsv.
3901 */
3902 return PTR_ERR(trans);
3903 }
3904 btrfs_commit_transaction(trans, rc->extent_root);
3905 return 0;
3906 }
3907
3908 static noinline_for_stack int relocate_block_group(struct reloc_control *rc)
3909 {
3910 struct rb_root blocks = RB_ROOT;
3911 struct btrfs_key key;
3912 struct btrfs_trans_handle *trans = NULL;
3913 struct btrfs_path *path;
3914 struct btrfs_extent_item *ei;
3915 u64 flags;
3916 u32 item_size;
3917 int ret;
3918 int err = 0;
3919 int progress = 0;
3920
3921 path = btrfs_alloc_path();
3922 if (!path)
3923 return -ENOMEM;
3924 path->reada = 1;
3925
3926 ret = prepare_to_relocate(rc);
3927 if (ret) {
3928 err = ret;
3929 goto out_free;
3930 }
3931
3932 while (1) {
3933 rc->reserved_bytes = 0;
3934 ret = btrfs_block_rsv_refill(rc->extent_root,
3935 rc->block_rsv, rc->block_rsv->size,
3936 BTRFS_RESERVE_FLUSH_ALL);
3937 if (ret) {
3938 err = ret;
3939 break;
3940 }
3941 progress++;
3942 trans = btrfs_start_transaction(rc->extent_root, 0);
3943 if (IS_ERR(trans)) {
3944 err = PTR_ERR(trans);
3945 trans = NULL;
3946 break;
3947 }
3948 restart:
3949 if (update_backref_cache(trans, &rc->backref_cache)) {
3950 btrfs_end_transaction(trans, rc->extent_root);
3951 continue;
3952 }
3953
3954 ret = find_next_extent(trans, rc, path, &key);
3955 if (ret < 0)
3956 err = ret;
3957 if (ret != 0)
3958 break;
3959
3960 rc->extents_found++;
3961
3962 ei = btrfs_item_ptr(path->nodes[0], path->slots[0],
3963 struct btrfs_extent_item);
3964 item_size = btrfs_item_size_nr(path->nodes[0], path->slots[0]);
3965 if (item_size >= sizeof(*ei)) {
3966 flags = btrfs_extent_flags(path->nodes[0], ei);
3967 ret = check_extent_flags(flags);
3968 BUG_ON(ret);
3969
3970 } else {
3971 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
3972 u64 ref_owner;
3973 int path_change = 0;
3974
3975 BUG_ON(item_size !=
3976 sizeof(struct btrfs_extent_item_v0));
3977 ret = get_ref_objectid_v0(rc, path, &key, &ref_owner,
3978 &path_change);
3979 if (ref_owner < BTRFS_FIRST_FREE_OBJECTID)
3980 flags = BTRFS_EXTENT_FLAG_TREE_BLOCK;
3981 else
3982 flags = BTRFS_EXTENT_FLAG_DATA;
3983
3984 if (path_change) {
3985 btrfs_release_path(path);
3986
3987 path->search_commit_root = 1;
3988 path->skip_locking = 1;
3989 ret = btrfs_search_slot(NULL, rc->extent_root,
3990 &key, path, 0, 0);
3991 if (ret < 0) {
3992 err = ret;
3993 break;
3994 }
3995 BUG_ON(ret > 0);
3996 }
3997 #else
3998 BUG();
3999 #endif
4000 }
4001
4002 if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
4003 ret = add_tree_block(rc, &key, path, &blocks);
4004 } else if (rc->stage == UPDATE_DATA_PTRS &&
4005 (flags & BTRFS_EXTENT_FLAG_DATA)) {
4006 ret = add_data_references(rc, &key, path, &blocks);
4007 } else {
4008 btrfs_release_path(path);
4009 ret = 0;
4010 }
4011 if (ret < 0) {
4012 err = ret;
4013 break;
4014 }
4015
4016 if (!RB_EMPTY_ROOT(&blocks)) {
4017 ret = relocate_tree_blocks(trans, rc, &blocks);
4018 if (ret < 0) {
4019 /*
4020 * if we fail to relocate tree blocks, force to update
4021 * backref cache when committing transaction.
4022 */
4023 rc->backref_cache.last_trans = trans->transid - 1;
4024
4025 if (ret != -EAGAIN) {
4026 err = ret;
4027 break;
4028 }
4029 rc->extents_found--;
4030 rc->search_start = key.objectid;
4031 }
4032 }
4033
4034 btrfs_end_transaction_throttle(trans, rc->extent_root);
4035 btrfs_btree_balance_dirty(rc->extent_root);
4036 trans = NULL;
4037
4038 if (rc->stage == MOVE_DATA_EXTENTS &&
4039 (flags & BTRFS_EXTENT_FLAG_DATA)) {
4040 rc->found_file_extent = 1;
4041 ret = relocate_data_extent(rc->data_inode,
4042 &key, &rc->cluster);
4043 if (ret < 0) {
4044 err = ret;
4045 break;
4046 }
4047 }
4048 }
4049 if (trans && progress && err == -ENOSPC) {
4050 ret = btrfs_force_chunk_alloc(trans, rc->extent_root,
4051 rc->block_group->flags);
4052 if (ret == 1) {
4053 err = 0;
4054 progress = 0;
4055 goto restart;
4056 }
4057 }
4058
4059 btrfs_release_path(path);
4060 clear_extent_bits(&rc->processed_blocks, 0, (u64)-1, EXTENT_DIRTY,
4061 GFP_NOFS);
4062
4063 if (trans) {
4064 btrfs_end_transaction_throttle(trans, rc->extent_root);
4065 btrfs_btree_balance_dirty(rc->extent_root);
4066 }
4067
4068 if (!err) {
4069 ret = relocate_file_extent_cluster(rc->data_inode,
4070 &rc->cluster);
4071 if (ret < 0)
4072 err = ret;
4073 }
4074
4075 rc->create_reloc_tree = 0;
4076 set_reloc_control(rc);
4077
4078 backref_cache_cleanup(&rc->backref_cache);
4079 btrfs_block_rsv_release(rc->extent_root, rc->block_rsv, (u64)-1);
4080
4081 err = prepare_to_merge(rc, err);
4082
4083 merge_reloc_roots(rc);
4084
4085 rc->merge_reloc_tree = 0;
4086 unset_reloc_control(rc);
4087 btrfs_block_rsv_release(rc->extent_root, rc->block_rsv, (u64)-1);
4088
4089 /* get rid of pinned extents */
4090 trans = btrfs_join_transaction(rc->extent_root);
4091 if (IS_ERR(trans))
4092 err = PTR_ERR(trans);
4093 else
4094 btrfs_commit_transaction(trans, rc->extent_root);
4095 out_free:
4096 btrfs_free_block_rsv(rc->extent_root, rc->block_rsv);
4097 btrfs_free_path(path);
4098 return err;
4099 }
4100
4101 static int __insert_orphan_inode(struct btrfs_trans_handle *trans,
4102 struct btrfs_root *root, u64 objectid)
4103 {
4104 struct btrfs_path *path;
4105 struct btrfs_inode_item *item;
4106 struct extent_buffer *leaf;
4107 int ret;
4108
4109 path = btrfs_alloc_path();
4110 if (!path)
4111 return -ENOMEM;
4112
4113 ret = btrfs_insert_empty_inode(trans, root, path, objectid);
4114 if (ret)
4115 goto out;
4116
4117 leaf = path->nodes[0];
4118 item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_inode_item);
4119 memset_extent_buffer(leaf, 0, (unsigned long)item, sizeof(*item));
4120 btrfs_set_inode_generation(leaf, item, 1);
4121 btrfs_set_inode_size(leaf, item, 0);
4122 btrfs_set_inode_mode(leaf, item, S_IFREG | 0600);
4123 btrfs_set_inode_flags(leaf, item, BTRFS_INODE_NOCOMPRESS |
4124 BTRFS_INODE_PREALLOC);
4125 btrfs_mark_buffer_dirty(leaf);
4126 out:
4127 btrfs_free_path(path);
4128 return ret;
4129 }
4130
4131 /*
4132 * helper to create inode for data relocation.
4133 * the inode is in data relocation tree and its link count is 0
4134 */
4135 static noinline_for_stack
4136 struct inode *create_reloc_inode(struct btrfs_fs_info *fs_info,
4137 struct btrfs_block_group_cache *group)
4138 {
4139 struct inode *inode = NULL;
4140 struct btrfs_trans_handle *trans;
4141 struct btrfs_root *root;
4142 struct btrfs_key key;
4143 u64 objectid = BTRFS_FIRST_FREE_OBJECTID;
4144 int err = 0;
4145
4146 root = read_fs_root(fs_info, BTRFS_DATA_RELOC_TREE_OBJECTID);
4147 if (IS_ERR(root))
4148 return ERR_CAST(root);
4149
4150 trans = btrfs_start_transaction(root, 6);
4151 if (IS_ERR(trans))
4152 return ERR_CAST(trans);
4153
4154 err = btrfs_find_free_objectid(root, &objectid);
4155 if (err)
4156 goto out;
4157
4158 err = __insert_orphan_inode(trans, root, objectid);
4159 BUG_ON(err);
4160
4161 key.objectid = objectid;
4162 key.type = BTRFS_INODE_ITEM_KEY;
4163 key.offset = 0;
4164 inode = btrfs_iget(root->fs_info->sb, &key, root, NULL);
4165 BUG_ON(IS_ERR(inode) || is_bad_inode(inode));
4166 BTRFS_I(inode)->index_cnt = group->key.objectid;
4167
4168 err = btrfs_orphan_add(trans, inode);
4169 out:
4170 btrfs_end_transaction(trans, root);
4171 btrfs_btree_balance_dirty(root);
4172 if (err) {
4173 if (inode)
4174 iput(inode);
4175 inode = ERR_PTR(err);
4176 }
4177 return inode;
4178 }
4179
4180 static struct reloc_control *alloc_reloc_control(struct btrfs_fs_info *fs_info)
4181 {
4182 struct reloc_control *rc;
4183
4184 rc = kzalloc(sizeof(*rc), GFP_NOFS);
4185 if (!rc)
4186 return NULL;
4187
4188 INIT_LIST_HEAD(&rc->reloc_roots);
4189 backref_cache_init(&rc->backref_cache);
4190 mapping_tree_init(&rc->reloc_root_tree);
4191 extent_io_tree_init(&rc->processed_blocks,
4192 fs_info->btree_inode->i_mapping);
4193 return rc;
4194 }
4195
4196 /*
4197 * function to relocate all extents in a block group.
4198 */
4199 int btrfs_relocate_block_group(struct btrfs_root *extent_root, u64 group_start)
4200 {
4201 struct btrfs_fs_info *fs_info = extent_root->fs_info;
4202 struct reloc_control *rc;
4203 struct inode *inode;
4204 struct btrfs_path *path;
4205 int ret;
4206 int rw = 0;
4207 int err = 0;
4208
4209 rc = alloc_reloc_control(fs_info);
4210 if (!rc)
4211 return -ENOMEM;
4212
4213 rc->extent_root = extent_root;
4214
4215 rc->block_group = btrfs_lookup_block_group(fs_info, group_start);
4216 BUG_ON(!rc->block_group);
4217
4218 if (!rc->block_group->ro) {
4219 ret = btrfs_set_block_group_ro(extent_root, rc->block_group);
4220 if (ret) {
4221 err = ret;
4222 goto out;
4223 }
4224 rw = 1;
4225 }
4226
4227 path = btrfs_alloc_path();
4228 if (!path) {
4229 err = -ENOMEM;
4230 goto out;
4231 }
4232
4233 inode = lookup_free_space_inode(fs_info->tree_root, rc->block_group,
4234 path);
4235 btrfs_free_path(path);
4236
4237 if (!IS_ERR(inode))
4238 ret = delete_block_group_cache(fs_info, rc->block_group, inode, 0);
4239 else
4240 ret = PTR_ERR(inode);
4241
4242 if (ret && ret != -ENOENT) {
4243 err = ret;
4244 goto out;
4245 }
4246
4247 rc->data_inode = create_reloc_inode(fs_info, rc->block_group);
4248 if (IS_ERR(rc->data_inode)) {
4249 err = PTR_ERR(rc->data_inode);
4250 rc->data_inode = NULL;
4251 goto out;
4252 }
4253
4254 btrfs_info(extent_root->fs_info, "relocating block group %llu flags %llu",
4255 rc->block_group->key.objectid, rc->block_group->flags);
4256
4257 ret = btrfs_start_delalloc_roots(fs_info, 0, -1);
4258 if (ret < 0) {
4259 err = ret;
4260 goto out;
4261 }
4262 btrfs_wait_ordered_roots(fs_info, -1);
4263
4264 while (1) {
4265 mutex_lock(&fs_info->cleaner_mutex);
4266 ret = relocate_block_group(rc);
4267 mutex_unlock(&fs_info->cleaner_mutex);
4268 if (ret < 0) {
4269 err = ret;
4270 goto out;
4271 }
4272
4273 if (rc->extents_found == 0)
4274 break;
4275
4276 btrfs_info(extent_root->fs_info, "found %llu extents",
4277 rc->extents_found);
4278
4279 if (rc->stage == MOVE_DATA_EXTENTS && rc->found_file_extent) {
4280 ret = btrfs_wait_ordered_range(rc->data_inode, 0,
4281 (u64)-1);
4282 if (ret) {
4283 err = ret;
4284 goto out;
4285 }
4286 invalidate_mapping_pages(rc->data_inode->i_mapping,
4287 0, -1);
4288 rc->stage = UPDATE_DATA_PTRS;
4289 }
4290 }
4291
4292 WARN_ON(rc->block_group->pinned > 0);
4293 WARN_ON(rc->block_group->reserved > 0);
4294 WARN_ON(btrfs_block_group_used(&rc->block_group->item) > 0);
4295 out:
4296 if (err && rw)
4297 btrfs_set_block_group_rw(extent_root, rc->block_group);
4298 iput(rc->data_inode);
4299 btrfs_put_block_group(rc->block_group);
4300 kfree(rc);
4301 return err;
4302 }
4303
4304 static noinline_for_stack int mark_garbage_root(struct btrfs_root *root)
4305 {
4306 struct btrfs_trans_handle *trans;
4307 int ret, err;
4308
4309 trans = btrfs_start_transaction(root->fs_info->tree_root, 0);
4310 if (IS_ERR(trans))
4311 return PTR_ERR(trans);
4312
4313 memset(&root->root_item.drop_progress, 0,
4314 sizeof(root->root_item.drop_progress));
4315 root->root_item.drop_level = 0;
4316 btrfs_set_root_refs(&root->root_item, 0);
4317 ret = btrfs_update_root(trans, root->fs_info->tree_root,
4318 &root->root_key, &root->root_item);
4319
4320 err = btrfs_end_transaction(trans, root->fs_info->tree_root);
4321 if (err)
4322 return err;
4323 return ret;
4324 }
4325
4326 /*
4327 * recover relocation interrupted by system crash.
4328 *
4329 * this function resumes merging reloc trees with corresponding fs trees.
4330 * this is important for keeping the sharing of tree blocks
4331 */
4332 int btrfs_recover_relocation(struct btrfs_root *root)
4333 {
4334 LIST_HEAD(reloc_roots);
4335 struct btrfs_key key;
4336 struct btrfs_root *fs_root;
4337 struct btrfs_root *reloc_root;
4338 struct btrfs_path *path;
4339 struct extent_buffer *leaf;
4340 struct reloc_control *rc = NULL;
4341 struct btrfs_trans_handle *trans;
4342 int ret;
4343 int err = 0;
4344
4345 path = btrfs_alloc_path();
4346 if (!path)
4347 return -ENOMEM;
4348 path->reada = -1;
4349
4350 key.objectid = BTRFS_TREE_RELOC_OBJECTID;
4351 key.type = BTRFS_ROOT_ITEM_KEY;
4352 key.offset = (u64)-1;
4353
4354 while (1) {
4355 ret = btrfs_search_slot(NULL, root->fs_info->tree_root, &key,
4356 path, 0, 0);
4357 if (ret < 0) {
4358 err = ret;
4359 goto out;
4360 }
4361 if (ret > 0) {
4362 if (path->slots[0] == 0)
4363 break;
4364 path->slots[0]--;
4365 }
4366 leaf = path->nodes[0];
4367 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
4368 btrfs_release_path(path);
4369
4370 if (key.objectid != BTRFS_TREE_RELOC_OBJECTID ||
4371 key.type != BTRFS_ROOT_ITEM_KEY)
4372 break;
4373
4374 reloc_root = btrfs_read_fs_root(root, &key);
4375 if (IS_ERR(reloc_root)) {
4376 err = PTR_ERR(reloc_root);
4377 goto out;
4378 }
4379
4380 list_add(&reloc_root->root_list, &reloc_roots);
4381
4382 if (btrfs_root_refs(&reloc_root->root_item) > 0) {
4383 fs_root = read_fs_root(root->fs_info,
4384 reloc_root->root_key.offset);
4385 if (IS_ERR(fs_root)) {
4386 ret = PTR_ERR(fs_root);
4387 if (ret != -ENOENT) {
4388 err = ret;
4389 goto out;
4390 }
4391 ret = mark_garbage_root(reloc_root);
4392 if (ret < 0) {
4393 err = ret;
4394 goto out;
4395 }
4396 }
4397 }
4398
4399 if (key.offset == 0)
4400 break;
4401
4402 key.offset--;
4403 }
4404 btrfs_release_path(path);
4405
4406 if (list_empty(&reloc_roots))
4407 goto out;
4408
4409 rc = alloc_reloc_control(root->fs_info);
4410 if (!rc) {
4411 err = -ENOMEM;
4412 goto out;
4413 }
4414
4415 rc->extent_root = root->fs_info->extent_root;
4416
4417 set_reloc_control(rc);
4418
4419 trans = btrfs_join_transaction(rc->extent_root);
4420 if (IS_ERR(trans)) {
4421 unset_reloc_control(rc);
4422 err = PTR_ERR(trans);
4423 goto out_free;
4424 }
4425
4426 rc->merge_reloc_tree = 1;
4427
4428 while (!list_empty(&reloc_roots)) {
4429 reloc_root = list_entry(reloc_roots.next,
4430 struct btrfs_root, root_list);
4431 list_del(&reloc_root->root_list);
4432
4433 if (btrfs_root_refs(&reloc_root->root_item) == 0) {
4434 list_add_tail(&reloc_root->root_list,
4435 &rc->reloc_roots);
4436 continue;
4437 }
4438
4439 fs_root = read_fs_root(root->fs_info,
4440 reloc_root->root_key.offset);
4441 if (IS_ERR(fs_root)) {
4442 err = PTR_ERR(fs_root);
4443 goto out_free;
4444 }
4445
4446 err = __add_reloc_root(reloc_root);
4447 BUG_ON(err < 0); /* -ENOMEM or logic error */
4448 fs_root->reloc_root = reloc_root;
4449 }
4450
4451 err = btrfs_commit_transaction(trans, rc->extent_root);
4452 if (err)
4453 goto out_free;
4454
4455 merge_reloc_roots(rc);
4456
4457 unset_reloc_control(rc);
4458
4459 trans = btrfs_join_transaction(rc->extent_root);
4460 if (IS_ERR(trans))
4461 err = PTR_ERR(trans);
4462 else
4463 err = btrfs_commit_transaction(trans, rc->extent_root);
4464 out_free:
4465 kfree(rc);
4466 out:
4467 if (!list_empty(&reloc_roots))
4468 free_reloc_roots(&reloc_roots);
4469
4470 btrfs_free_path(path);
4471
4472 if (err == 0) {
4473 /* cleanup orphan inode in data relocation tree */
4474 fs_root = read_fs_root(root->fs_info,
4475 BTRFS_DATA_RELOC_TREE_OBJECTID);
4476 if (IS_ERR(fs_root))
4477 err = PTR_ERR(fs_root);
4478 else
4479 err = btrfs_orphan_cleanup(fs_root);
4480 }
4481 return err;
4482 }
4483
4484 /*
4485 * helper to add ordered checksum for data relocation.
4486 *
4487 * cloning checksum properly handles the nodatasum extents.
4488 * it also saves CPU time to re-calculate the checksum.
4489 */
4490 int btrfs_reloc_clone_csums(struct inode *inode, u64 file_pos, u64 len)
4491 {
4492 struct btrfs_ordered_sum *sums;
4493 struct btrfs_ordered_extent *ordered;
4494 struct btrfs_root *root = BTRFS_I(inode)->root;
4495 int ret;
4496 u64 disk_bytenr;
4497 u64 new_bytenr;
4498 LIST_HEAD(list);
4499
4500 ordered = btrfs_lookup_ordered_extent(inode, file_pos);
4501 BUG_ON(ordered->file_offset != file_pos || ordered->len != len);
4502
4503 disk_bytenr = file_pos + BTRFS_I(inode)->index_cnt;
4504 ret = btrfs_lookup_csums_range(root->fs_info->csum_root, disk_bytenr,
4505 disk_bytenr + len - 1, &list, 0);
4506 if (ret)
4507 goto out;
4508
4509 while (!list_empty(&list)) {
4510 sums = list_entry(list.next, struct btrfs_ordered_sum, list);
4511 list_del_init(&sums->list);
4512
4513 /*
4514 * We need to offset the new_bytenr based on where the csum is.
4515 * We need to do this because we will read in entire prealloc
4516 * extents but we may have written to say the middle of the
4517 * prealloc extent, so we need to make sure the csum goes with
4518 * the right disk offset.
4519 *
4520 * We can do this because the data reloc inode refers strictly
4521 * to the on disk bytes, so we don't have to worry about
4522 * disk_len vs real len like with real inodes since it's all
4523 * disk length.
4524 */
4525 new_bytenr = ordered->start + (sums->bytenr - disk_bytenr);
4526 sums->bytenr = new_bytenr;
4527
4528 btrfs_add_ordered_sum(inode, ordered, sums);
4529 }
4530 out:
4531 btrfs_put_ordered_extent(ordered);
4532 return ret;
4533 }
4534
4535 int btrfs_reloc_cow_block(struct btrfs_trans_handle *trans,
4536 struct btrfs_root *root, struct extent_buffer *buf,
4537 struct extent_buffer *cow)
4538 {
4539 struct reloc_control *rc;
4540 struct backref_node *node;
4541 int first_cow = 0;
4542 int level;
4543 int ret = 0;
4544
4545 rc = root->fs_info->reloc_ctl;
4546 if (!rc)
4547 return 0;
4548
4549 BUG_ON(rc->stage == UPDATE_DATA_PTRS &&
4550 root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID);
4551
4552 if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID) {
4553 if (buf == root->node)
4554 __update_reloc_root(root, cow->start);
4555 }
4556
4557 level = btrfs_header_level(buf);
4558 if (btrfs_header_generation(buf) <=
4559 btrfs_root_last_snapshot(&root->root_item))
4560 first_cow = 1;
4561
4562 if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID &&
4563 rc->create_reloc_tree) {
4564 WARN_ON(!first_cow && level == 0);
4565
4566 node = rc->backref_cache.path[level];
4567 BUG_ON(node->bytenr != buf->start &&
4568 node->new_bytenr != buf->start);
4569
4570 drop_node_buffer(node);
4571 extent_buffer_get(cow);
4572 node->eb = cow;
4573 node->new_bytenr = cow->start;
4574
4575 if (!node->pending) {
4576 list_move_tail(&node->list,
4577 &rc->backref_cache.pending[level]);
4578 node->pending = 1;
4579 }
4580
4581 if (first_cow)
4582 __mark_block_processed(rc, node);
4583
4584 if (first_cow && level > 0)
4585 rc->nodes_relocated += buf->len;
4586 }
4587
4588 if (level == 0 && first_cow && rc->stage == UPDATE_DATA_PTRS)
4589 ret = replace_file_extents(trans, rc, root, cow);
4590 return ret;
4591 }
4592
4593 /*
4594 * called before creating snapshot. it calculates metadata reservation
4595 * requried for relocating tree blocks in the snapshot
4596 */
4597 void btrfs_reloc_pre_snapshot(struct btrfs_trans_handle *trans,
4598 struct btrfs_pending_snapshot *pending,
4599 u64 *bytes_to_reserve)
4600 {
4601 struct btrfs_root *root;
4602 struct reloc_control *rc;
4603
4604 root = pending->root;
4605 if (!root->reloc_root)
4606 return;
4607
4608 rc = root->fs_info->reloc_ctl;
4609 if (!rc->merge_reloc_tree)
4610 return;
4611
4612 root = root->reloc_root;
4613 BUG_ON(btrfs_root_refs(&root->root_item) == 0);
4614 /*
4615 * relocation is in the stage of merging trees. the space
4616 * used by merging a reloc tree is twice the size of
4617 * relocated tree nodes in the worst case. half for cowing
4618 * the reloc tree, half for cowing the fs tree. the space
4619 * used by cowing the reloc tree will be freed after the
4620 * tree is dropped. if we create snapshot, cowing the fs
4621 * tree may use more space than it frees. so we need
4622 * reserve extra space.
4623 */
4624 *bytes_to_reserve += rc->nodes_relocated;
4625 }
4626
4627 /*
4628 * called after snapshot is created. migrate block reservation
4629 * and create reloc root for the newly created snapshot
4630 */
4631 int btrfs_reloc_post_snapshot(struct btrfs_trans_handle *trans,
4632 struct btrfs_pending_snapshot *pending)
4633 {
4634 struct btrfs_root *root = pending->root;
4635 struct btrfs_root *reloc_root;
4636 struct btrfs_root *new_root;
4637 struct reloc_control *rc;
4638 int ret;
4639
4640 if (!root->reloc_root)
4641 return 0;
4642
4643 rc = root->fs_info->reloc_ctl;
4644 rc->merging_rsv_size += rc->nodes_relocated;
4645
4646 if (rc->merge_reloc_tree) {
4647 ret = btrfs_block_rsv_migrate(&pending->block_rsv,
4648 rc->block_rsv,
4649 rc->nodes_relocated);
4650 if (ret)
4651 return ret;
4652 }
4653
4654 new_root = pending->snap;
4655 reloc_root = create_reloc_root(trans, root->reloc_root,
4656 new_root->root_key.objectid);
4657 if (IS_ERR(reloc_root))
4658 return PTR_ERR(reloc_root);
4659
4660 ret = __add_reloc_root(reloc_root);
4661 BUG_ON(ret < 0);
4662 new_root->reloc_root = reloc_root;
4663
4664 if (rc->create_reloc_tree)
4665 ret = clone_backref_node(trans, rc, root, reloc_root);
4666 return ret;
4667 }
Linux with added FPC-III support