projects / linux-2.6-block.git / blame
? search:
summary | shortlog | log | commit | commitdiff | tree
blob | blame (incremental) | history | HEAD
wrappers for ->i_mutex access
[linux-2.6-block.git] / fs / btrfs / relocation.c
CommitLineData
5d4f98a2
YZ		 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>
5a0e3ad6 24#include <linux/slab.h>
5d4f98a2
YZ		 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"
0af3d00b 32#include "free-space-cache.h"
581bb050 33#include "inode-map.h"
5d4f98a2
YZ		 34
35/*
36 * backref_node, mapping_node and tree_block start with this
37 */
38struct tree_entry {
39 struct rb_node rb_node;
40 u64 bytenr;
41};
42
43/*
44 * present a tree block in the backref cache
45 */
46struct backref_node {
47 struct rb_node rb_node;
48 u64 bytenr;
3fd0a558
YZ		 49
50 u64 new_bytenr;
51 /* objectid of tree block owner, can be not uptodate */
5d4f98a2 52 u64 owner;
3fd0a558
YZ		 53 /* link to pending, changed or detached list */
54 struct list_head list;
5d4f98a2
YZ		 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;
3fd0a558
YZ		 65 /* is the block in non-reference counted tree */
66 unsigned int cowonly:1;
67 /* 1 if no child node in the cache */
5d4f98a2
YZ		 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;
3fd0a558
YZ		 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;
5d4f98a2
YZ		 85};
86
87/*
88 * present a block pointer in the backref cache
89 */
90struct backref_edge {
91 struct list_head list[2];
92 struct backref_node *node[2];
5d4f98a2
YZ		 93};
94
95#define LOWER 0
96#define UPPER 1
0647bf56 97#define RELOCATION_RESERVED_NODES 256
5d4f98a2
YZ		 98
99struct backref_cache {
100 /* red black tree of all backref nodes in the cache */
101 struct rb_root rb_root;
3fd0a558
YZ		 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 */
5d4f98a2 109 struct list_head pending[BTRFS_MAX_LEVEL];
3fd0a558
YZ		 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;
5d4f98a2
YZ		 121};
122
123/*
124 * map address of tree root to tree
125 */
126struct mapping_node {
127 struct rb_node rb_node;
128 u64 bytenr;
129 void *data;
130};
131
132struct mapping_tree {
133 struct rb_root rb_root;
134 spinlock_t lock;
135};
136
137/*
138 * present a tree block to process
139 */
140struct 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
0257bb82
YZ		 148#define MAX_EXTENTS 128
149
150struct file_extent_cluster {
151 u64 start;
152 u64 end;
153 u64 boundary[MAX_EXTENTS];
154 unsigned int nr;
155};
156
5d4f98a2
YZ		 157struct 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;
3fd0a558
YZ		 164
165 struct btrfs_block_rsv *block_rsv;
166
167 struct backref_cache backref_cache;
168
169 struct file_extent_cluster cluster;
5d4f98a2
YZ		 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;
3fd0a558
YZ		 176 /* size of metadata reservation for merging reloc trees */
177 u64 merging_rsv_size;
178 /* size of relocated tree nodes */
179 u64 nodes_relocated;
0647bf56
WS		 180 /* reserved size for block group relocation*/
181 u64 reserved_bytes;
3fd0a558 182
5d4f98a2
YZ		 183 u64 search_start;
184 u64 extents_found;
3fd0a558 185
3fd0a558
YZ		 186 unsigned int stage:8;
187 unsigned int create_reloc_tree:1;
188 unsigned int merge_reloc_tree:1;
5d4f98a2 189 unsigned int found_file_extent:1;
5d4f98a2
YZ		 190};
191
192/* stages of data relocation */
193#define MOVE_DATA_EXTENTS 0
194#define UPDATE_DATA_PTRS 1
195
3fd0a558
YZ		 196static void remove_backref_node(struct backref_cache *cache,
197 struct backref_node *node);
198static void __mark_block_processed(struct reloc_control *rc,
199 struct backref_node *node);
5d4f98a2
YZ		 200
201static void mapping_tree_init(struct mapping_tree *tree)
202{
6bef4d31 203 tree->rb_root = RB_ROOT;
5d4f98a2
YZ		 204 spin_lock_init(&tree->lock);
205}
206
207static void backref_cache_init(struct backref_cache *cache)
208{
209 int i;
6bef4d31 210 cache->rb_root = RB_ROOT;
5d4f98a2
YZ		 211 for (i = 0; i < BTRFS_MAX_LEVEL; i++)
212 INIT_LIST_HEAD(&cache->pending[i]);
3fd0a558
YZ		 213 INIT_LIST_HEAD(&cache->changed);
214 INIT_LIST_HEAD(&cache->detached);
215 INIT_LIST_HEAD(&cache->leaves);
216}
217
218static 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
246static 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
261static 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
270static 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;
5d4f98a2
YZ		 278}
279
3fd0a558
YZ		 280static void free_backref_edge(struct backref_cache *cache,
281 struct backref_edge *edge)
5d4f98a2 282{
3fd0a558
YZ		 283 if (edge) {
284 cache->nr_edges--;
285 kfree(edge);
286 }
5d4f98a2
YZ		 287}
288
289static 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
313static 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
48a3b636 331static void backref_tree_panic(struct rb_node *rb_node, int errno, u64 bytenr)
43c04fb1
JM		 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 "
351fd353 340 "found at offset %llu", bytenr);
43c04fb1
JM		 341}
342
5d4f98a2
YZ		 343/*
344 * walk up backref nodes until reach node presents tree root
345 */
346static 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 }
3fd0a558 359 BUG_ON(node->detached);
5d4f98a2
YZ		 360 *index = idx;
361 return node;
362}
363
364/*
365 * walk down backref nodes to find start of next reference path
366 */
367static 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
3fd0a558
YZ		 391static 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
5d4f98a2
YZ		 399static void drop_node_buffer(struct backref_node *node)
400{
401 if (node->eb) {
3fd0a558 402 unlock_node_buffer(node);
5d4f98a2
YZ		 403 free_extent_buffer(node->eb);
404 node->eb = NULL;
405 }
406}
407
408static void drop_backref_node(struct backref_cache *tree,
409 struct backref_node *node)
410{
5d4f98a2
YZ		 411 BUG_ON(!list_empty(&node->upper));
412
413 drop_node_buffer(node);
3fd0a558 414 list_del(&node->list);
5d4f98a2 415 list_del(&node->lower);
3fd0a558
YZ		 416 if (!RB_EMPTY_NODE(&node->rb_node))
417 rb_erase(&node->rb_node, &tree->rb_root);
418 free_backref_node(tree, node);
5d4f98a2
YZ		 419}
420
421/*
422 * remove a backref node from the backref cache
423 */
424static 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
3fd0a558 433 BUG_ON(!node->lowest && !node->detached);
5d4f98a2
YZ		 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]);
3fd0a558
YZ		 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 }
5d4f98a2 449 /*
3fd0a558 450 * add the node to leaf node list if no other
5d4f98a2
YZ		 451 * child block cached.
452 */
453 if (list_empty(&upper->lower)) {
3fd0a558 454 list_add_tail(&upper->lower, &cache->leaves);
5d4f98a2
YZ		 455 upper->lowest = 1;
456 }
457 }
3fd0a558 458
5d4f98a2
YZ		 459 drop_backref_node(cache, node);
460}
461
3fd0a558
YZ		 462static 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);
43c04fb1
JM		 469 if (rb_node)
470 backref_tree_panic(rb_node, -EEXIST, bytenr);
3fd0a558
YZ		 471}
472
473/*
474 * update backref cache after a transaction commit
475 */
476static 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
f2a97a9d 526
3fd0a558
YZ		 527static int should_ignore_root(struct btrfs_root *root)
528{
529 struct btrfs_root *reloc_root;
530
27cdeb70 531 if (!test_bit(BTRFS_ROOT_REF_COWS, &root->state))
3fd0a558
YZ		 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}
5d4f98a2
YZ		 549/*
550 * find reloc tree by address of tree root
551 */
552static 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
569static 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 ||
66463748
WS		 576 root_objectid == BTRFS_CSUM_TREE_OBJECTID ||
577 root_objectid == BTRFS_UUID_TREE_OBJECTID ||
578 root_objectid == BTRFS_QUOTA_TREE_OBJECTID)
5d4f98a2
YZ		 579 return 1;
580 return 0;
581}
582
583static 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
c00869f1 595 return btrfs_get_fs_root(fs_info, &key, false);
5d4f98a2
YZ		 596}
597
598#ifdef BTRFS_COMPAT_EXTENT_TREE_V0
599static noinline_for_stack
600struct 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
27cdeb70 613 if (test_bit(BTRFS_ROOT_REF_COWS, &root->state) &&
5d4f98a2
YZ		 614 generation != btrfs_root_generation(&root->root_item))
615 return NULL;
616
617 return root;
618}
619#endif
620
621static noinline_for_stack
622int find_inline_backref(struct extent_buffer *leaf, int slot,
623 unsigned long *ptr, unsigned long *end)
624{
3173a18f 625 struct btrfs_key key;
5d4f98a2
YZ		 626 struct btrfs_extent_item *ei;
627 struct btrfs_tree_block_info *bi;
628 u32 item_size;
629
3173a18f
JB		 630 btrfs_item_key_to_cpu(leaf, &key, slot);
631
5d4f98a2
YZ		 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
3173a18f
JB		 643 if (key.type == BTRFS_EXTENT_ITEM_KEY &&
644 item_size <= sizeof(*ei) + sizeof(*bi)) {
5d4f98a2
YZ		 645 WARN_ON(item_size < sizeof(*ei) + sizeof(*bi));
646 return 1;
647 }
d062d13c
JB		 648 if (key.type == BTRFS_METADATA_ITEM_KEY &&
649 item_size <= sizeof(*ei)) {
650 WARN_ON(item_size < sizeof(*ei));
651 return 1;
652 }
5d4f98a2 653
3173a18f
JB		 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 }
5d4f98a2
YZ		 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 */
3fd0a558
YZ		 678static noinline_for_stack
679struct backref_node *build_backref_tree(struct reloc_control *rc,
680 struct btrfs_key *node_key,
681 int level, u64 bytenr)
5d4f98a2 682{
3fd0a558 683 struct backref_cache *cache = &rc->backref_cache;
5d4f98a2
YZ		 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);
3fd0a558
YZ		 699 LIST_HEAD(useless);
700 int cowonly;
5d4f98a2
YZ		 701 int ret;
702 int err = 0;
b6c60c80 703 bool need_check = true;
5d4f98a2
YZ		 704
705 path1 = btrfs_alloc_path();
706 path2 = btrfs_alloc_path();
707 if (!path1 || !path2) {
708 err = -ENOMEM;
709 goto out;
710 }
e4058b54
DS		 711 path1->reada = READA_FORWARD;
712 path2->reada = READA_FORWARD;
5d4f98a2 713
3fd0a558 714 node = alloc_backref_node(cache);
5d4f98a2
YZ		 715 if (!node) {
716 err = -ENOMEM;
717 goto out;
718 }
719
5d4f98a2 720 node->bytenr = bytenr;
5d4f98a2
YZ		 721 node->level = level;
722 node->lowest = 1;
723 cur = node;
724again:
725 end = 0;
726 ptr = 0;
727 key.objectid = cur->bytenr;
3173a18f 728 key.type = BTRFS_METADATA_ITEM_KEY;
5d4f98a2
YZ		 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 }
75bfb9af
JB		 739 ASSERT(ret);
740 ASSERT(path1->slots[0]);
5d4f98a2
YZ		 741
742 path1->slots[0]--;
743
744 WARN_ON(cur->checked);
745 if (!list_empty(&cur->upper)) {
746 /*
70f23fd6 747 * the backref was added previously when processing
5d4f98a2
YZ		 748 * backref of type BTRFS_TREE_BLOCK_REF_KEY
749 */
75bfb9af 750 ASSERT(list_is_singular(&cur->upper));
5d4f98a2
YZ		 751 edge = list_entry(cur->upper.next, struct backref_edge,
752 list[LOWER]);
75bfb9af 753 ASSERT(list_empty(&edge->list[UPPER]));
5d4f98a2
YZ		 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
3173a18f
JB		 787 if (key.type == BTRFS_EXTENT_ITEM_KEY ||
788 key.type == BTRFS_METADATA_ITEM_KEY) {
5d4f98a2
YZ		 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) {
3fd0a558 818 if (key.type == BTRFS_EXTENT_REF_V0_KEY) {
5d4f98a2
YZ		 819 struct btrfs_extent_ref_v0 *ref0;
820 ref0 = btrfs_item_ptr(eb, path1->slots[0],
821 struct btrfs_extent_ref_v0);
3fd0a558 822 if (key.objectid == key.offset) {
046f264f 823 root = find_tree_root(rc, eb, ref0);
3fd0a558
YZ		 824 if (root && !should_ignore_root(root))
825 cur->root = root;
826 else
827 list_add(&cur->list, &useless);
828 break;
829 }
046f264f
YZ		 830 if (is_cowonly_root(btrfs_ref_root_v0(eb,
831 ref0)))
832 cur->cowonly = 1;
5d4f98a2
YZ		 833 }
834#else
75bfb9af 835 ASSERT(key.type != BTRFS_EXTENT_REF_V0_KEY);
5d4f98a2
YZ		 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);
75bfb9af 844 ASSERT(root);
5d4f98a2
YZ		 845 cur->root = root;
846 break;
847 }
848
3fd0a558 849 edge = alloc_backref_edge(cache);
5d4f98a2
YZ		 850 if (!edge) {
851 err = -ENOMEM;
852 goto out;
853 }
854 rb_node = tree_search(&cache->rb_root, key.offset);
855 if (!rb_node) {
3fd0a558 856 upper = alloc_backref_node(cache);
5d4f98a2 857 if (!upper) {
3fd0a558 858 free_backref_edge(cache, edge);
5d4f98a2
YZ		 859 err = -ENOMEM;
860 goto out;
861 }
5d4f98a2 862 upper->bytenr = key.offset;
5d4f98a2
YZ		 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);
75bfb9af 872 ASSERT(upper->checked);
5d4f98a2
YZ		 873 INIT_LIST_HEAD(&edge->list[UPPER]);
874 }
3fd0a558 875 list_add_tail(&edge->list[LOWER], &cur->upper);
5d4f98a2 876 edge->node[LOWER] = cur;
3fd0a558 877 edge->node[UPPER] = upper;
5d4f98a2
YZ		 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
27cdeb70 891 if (!test_bit(BTRFS_ROOT_REF_COWS, &root->state))
3fd0a558
YZ		 892 cur->cowonly = 1;
893
5d4f98a2
YZ		 894 if (btrfs_root_level(&root->root_item) == cur->level) {
895 /* tree root */
75bfb9af 896 ASSERT(btrfs_root_bytenr(&root->root_item) ==
5d4f98a2 897 cur->bytenr);
3fd0a558
YZ		 898 if (should_ignore_root(root))
899 list_add(&cur->list, &useless);
900 else
901 cur->root = root;
5d4f98a2
YZ		 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 }
33c66f43
YZ		 920 if (ret > 0 && path2->slots[level] > 0)
921 path2->slots[level]--;
5d4f98a2
YZ		 922
923 eb = path2->nodes[level];
924 WARN_ON(btrfs_node_blockptr(eb, path2->slots[level]) !=
925 cur->bytenr);
926
927 lower = cur;
b6c60c80 928 need_check = true;
5d4f98a2
YZ		 929 for (; level < BTRFS_MAX_LEVEL; level++) {
930 if (!path2->nodes[level]) {
75bfb9af 931 ASSERT(btrfs_root_bytenr(&root->root_item) ==
5d4f98a2 932 lower->bytenr);
3fd0a558
YZ		 933 if (should_ignore_root(root))
934 list_add(&lower->list, &useless);
935 else
936 lower->root = root;
5d4f98a2
YZ		 937 break;
938 }
939
3fd0a558 940 edge = alloc_backref_edge(cache);
5d4f98a2
YZ		 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) {
3fd0a558 949 upper = alloc_backref_node(cache);
5d4f98a2 950 if (!upper) {
3fd0a558 951 free_backref_edge(cache, edge);
5d4f98a2
YZ		 952 err = -ENOMEM;
953 goto out;
954 }
5d4f98a2
YZ		 955 upper->bytenr = eb->start;
956 upper->owner = btrfs_header_owner(eb);
957 upper->level = lower->level + 1;
27cdeb70
MX		 958 if (!test_bit(BTRFS_ROOT_REF_COWS,
959 &root->state))
3fd0a558 960 upper->cowonly = 1;
5d4f98a2
YZ		 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
b6c60c80
JB		 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.
5d4f98a2 976 */
b6c60c80
JB		 977 if (!upper->checked && need_check) {
978 need_check = false;
5d4f98a2
YZ		 979 list_add_tail(&edge->list[UPPER],
980 &list);
bbe90514
JB		 981 } else {
982 if (upper->checked)
983 need_check = true;
5d4f98a2 984 INIT_LIST_HEAD(&edge->list[UPPER]);
bbe90514 985 }
5d4f98a2
YZ		 986 } else {
987 upper = rb_entry(rb_node, struct backref_node,
988 rb_node);
75bfb9af 989 ASSERT(upper->checked);
5d4f98a2 990 INIT_LIST_HEAD(&edge->list[UPPER]);
3fd0a558
YZ		 991 if (!upper->owner)
992 upper->owner = btrfs_header_owner(eb);
5d4f98a2
YZ		 993 }
994 list_add_tail(&edge->list[LOWER], &lower->upper);
5d4f98a2 995 edge->node[LOWER] = lower;
3fd0a558 996 edge->node[UPPER] = upper;
5d4f98a2
YZ		 997
998 if (rb_node)
999 break;
1000 lower = upper;
1001 upper = NULL;
1002 }
b3b4aa74 1003 btrfs_release_path(path2);
5d4f98a2
YZ		 1004next:
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 }
b3b4aa74 1016 btrfs_release_path(path1);
5d4f98a2
YZ		 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 */
75bfb9af 1033 ASSERT(node->checked);
3fd0a558
YZ		 1034 cowonly = node->cowonly;
1035 if (!cowonly) {
1036 rb_node = tree_insert(&cache->rb_root, node->bytenr,
1037 &node->rb_node);
43c04fb1
JM		 1038 if (rb_node)
1039 backref_tree_panic(rb_node, -EEXIST, node->bytenr);
3fd0a558
YZ		 1040 list_add_tail(&node->lower, &cache->leaves);
1041 }
5d4f98a2
YZ		 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];
3fd0a558
YZ		 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 }
5d4f98a2
YZ		 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
75bfb9af
JB		 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
3fd0a558
YZ		 1084 if (!cowonly) {
1085 rb_node = tree_insert(&cache->rb_root, upper->bytenr,
1086 &upper->rb_node);
43c04fb1
JM		 1087 if (rb_node)
1088 backref_tree_panic(rb_node, -EEXIST,
1089 upper->bytenr);
3fd0a558 1090 }
5d4f98a2
YZ		 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 }
3fd0a558
YZ		 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);
75bfb9af 1106 ASSERT(list_empty(&upper->upper));
3fd0a558
YZ		 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 }
5d4f98a2
YZ		 1133out:
1134 btrfs_free_path(path1);
1135 btrfs_free_path(path2);
1136 if (err) {
3fd0a558
YZ		 1137 while (!list_empty(&useless)) {
1138 lower = list_entry(useless.next,
75bfb9af
JB		 1139 struct backref_node, list);
1140 list_del_init(&lower->list);
3fd0a558 1141 }
75bfb9af
JB		 1142 while (!list_empty(&list)) {
1143 edge = list_first_entry(&list, struct backref_edge,
1144 list[UPPER]);
1145 list_del(&edge->list[UPPER]);
3fd0a558 1146 list_del(&edge->list[LOWER]);
75bfb9af 1147 lower = edge->node[LOWER];
5d4f98a2 1148 upper = edge->node[UPPER];
3fd0a558 1149 free_backref_edge(cache, edge);
75bfb9af
JB		 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);
5d4f98a2
YZ		 1174 }
1175 return ERR_PTR(err);
1176 }
75bfb9af 1177 ASSERT(!node || !node->detached);
5d4f98a2
YZ		 1178 return node;
1179}
1180
3fd0a558
YZ		 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 */
1186static 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;
6848ad64 1231 new_node->checked = 1;
3fd0a558
YZ		 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 }
76b9e23d
MX		 1245 } else {
1246 list_add_tail(&new_node->lower, &cache->leaves);
3fd0a558
YZ		 1247 }
1248
1249 rb_node = tree_insert(&cache->rb_root, new_node->bytenr,
1250 &new_node->rb_node);
43c04fb1
JM		 1251 if (rb_node)
1252 backref_tree_panic(rb_node, -EEXIST, new_node->bytenr);
3fd0a558
YZ		 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;
1261fail:
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
5d4f98a2
YZ		 1272/*
1273 * helper to add 'address of tree root -> reloc tree' mapping
1274 */
ffd7b339 1275static int __must_check __add_reloc_root(struct btrfs_root *root)
5d4f98a2
YZ		 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);
ffd7b339
JM		 1282 if (!node)
1283 return -ENOMEM;
5d4f98a2
YZ		 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);
ffd7b339 1292 if (rb_node) {
ffd7b339
JM		 1293 btrfs_panic(root->fs_info, -EEXIST, "Duplicate root found "
1294 "for start=%llu while inserting into relocation "
351fd353 1295 "tree", node->bytenr);
23291a04
DC		 1296 kfree(node);
1297 return -EEXIST;
ffd7b339 1298 }
5d4f98a2
YZ		 1299
1300 list_add_tail(&root->root_list, &rc->reloc_roots);
1301 return 0;
1302}
1303
1304/*
c974c464 1305 * helper to delete the 'address of tree root -> reloc tree'
5d4f98a2
YZ		 1306 * mapping
1307 */
c974c464 1308static void __del_reloc_root(struct btrfs_root *root)
5d4f98a2
YZ		 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,
c974c464 1316 root->node->start);
5d4f98a2
YZ		 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
8f71f3e0 1323 if (!node)
c974c464 1324 return;
5d4f98a2
YZ		 1325 BUG_ON((struct btrfs_root *)node->data != root);
1326
c974c464
WS		 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 */
1337static 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);
5d4f98a2 1349 }
c974c464
WS		 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);
5d4f98a2
YZ		 1363 return 0;
1364}
1365
3fd0a558
YZ		 1366static struct btrfs_root *create_reloc_root(struct btrfs_trans_handle *trans,
1367 struct btrfs_root *root, u64 objectid)
5d4f98a2
YZ		 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;
5bc7247a 1373 u64 last_snap = 0;
5d4f98a2
YZ		 1374 int ret;
1375
5d4f98a2
YZ		 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;
3fd0a558 1381 root_key.offset = objectid;
5d4f98a2 1382
3fd0a558
YZ		 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
5bc7247a 1389 last_snap = btrfs_root_last_snapshot(&root->root_item);
3fd0a558
YZ		 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 }
5d4f98a2 1404
5d4f98a2 1405 memcpy(root_item, &root->root_item, sizeof(*root_item));
5d4f98a2
YZ		 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);
3fd0a558
YZ		 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;
5bc7247a
MX		 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);
3fd0a558 1421 }
5d4f98a2
YZ		 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
cb517eab 1431 reloc_root = btrfs_read_fs_root(root->fs_info->tree_root, &root_key);
5d4f98a2
YZ		 1432 BUG_ON(IS_ERR(reloc_root));
1433 reloc_root->last_trans = trans->transid;
3fd0a558
YZ		 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 */
1441int 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;
20dd2cbf 1446 struct btrfs_block_rsv *rsv;
3fd0a558 1447 int clear_rsv = 0;
ffd7b339 1448 int ret;
3fd0a558
YZ		 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
20dd2cbf
MX		 1460 if (!trans->reloc_reserved) {
1461 rsv = trans->block_rsv;
3fd0a558
YZ		 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)
20dd2cbf 1467 trans->block_rsv = rsv;
5d4f98a2 1468
ffd7b339
JM		 1469 ret = __add_reloc_root(reloc_root);
1470 BUG_ON(ret < 0);
5d4f98a2
YZ		 1471 root->reloc_root = reloc_root;
1472 return 0;
1473}
1474
1475/*
1476 * update root item of reloc tree
1477 */
1478int 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;
5d4f98a2
YZ		 1483 int ret;
1484
1485 if (!root->reloc_root)
7585717f 1486 goto out;
5d4f98a2
YZ		 1487
1488 reloc_root = root->reloc_root;
1489 root_item = &reloc_root->root_item;
1490
3fd0a558
YZ		 1491 if (root->fs_info->reloc_ctl->merge_reloc_tree &&
1492 btrfs_root_refs(root_item) == 0) {
5d4f98a2 1493 root->reloc_root = NULL;
c974c464 1494 __del_reloc_root(reloc_root);
5d4f98a2
YZ		 1495 }
1496
5d4f98a2
YZ		 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);
7585717f
CM		 1506
1507out:
5d4f98a2
YZ		 1508 return 0;
1509}
1510
1511/*
1512 * helper to find first cached inode with inode number >= objectid
1513 * in a subvolume
1514 */
1515static 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);
1523again:
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
33345d01 1530 if (objectid < btrfs_ino(&entry->vfs_inode))
5d4f98a2 1531 node = node->rb_left;
33345d01 1532 else if (objectid > btrfs_ino(&entry->vfs_inode))
5d4f98a2
YZ		 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);
33345d01 1540 if (objectid <= btrfs_ino(&entry->vfs_inode)) {
5d4f98a2
YZ		 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
33345d01 1555 objectid = btrfs_ino(&entry->vfs_inode) + 1;
5d4f98a2
YZ		 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
1565static 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 */
1577static 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;
33345d01 1591 ret = btrfs_lookup_file_extent(NULL, root, path, btrfs_ino(reloc_inode),
5d4f98a2
YZ		 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)) {
83d4cfd4 1610 ret = -EINVAL;
5d4f98a2
YZ		 1611 goto out;
1612 }
1613
3fd0a558 1614 *new_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
5d4f98a2
YZ		 1615 ret = 0;
1616out:
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 */
3fd0a558
YZ		 1625static noinline_for_stack
1626int replace_file_extents(struct btrfs_trans_handle *trans,
1627 struct reloc_control *rc,
1628 struct btrfs_root *root,
1629 struct extent_buffer *leaf)
5d4f98a2
YZ		 1630{
1631 struct btrfs_key key;
1632 struct btrfs_file_extent_item *fi;
1633 struct inode *inode = NULL;
5d4f98a2
YZ		 1634 u64 parent;
1635 u64 bytenr;
3fd0a558 1636 u64 new_bytenr = 0;
5d4f98a2
YZ		 1637 u64 num_bytes;
1638 u64 end;
1639 u32 nritems;
1640 u32 i;
83d4cfd4 1641 int ret = 0;
5d4f98a2
YZ		 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) {
5d4f98a2
YZ		 1676 if (first) {
1677 inode = find_next_inode(root, key.objectid);
5d4f98a2 1678 first = 0;
33345d01 1679 } else if (inode && btrfs_ino(inode) < key.objectid) {
3fd0a558 1680 btrfs_add_delayed_iput(inode);
5d4f98a2 1681 inode = find_next_inode(root, key.objectid);
5d4f98a2 1682 }
33345d01 1683 if (inode && btrfs_ino(inode) == key.objectid) {
5d4f98a2
YZ		 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,
d0082371 1691 key.offset, end);
5d4f98a2
YZ		 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,
d0082371 1698 key.offset, end);
5d4f98a2
YZ		 1699 }
1700 }
1701
1702 ret = get_new_location(rc->data_inode, &new_bytenr,
1703 bytenr, num_bytes);
83d4cfd4
JB		 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;
3fd0a558 1710 }
5d4f98a2
YZ		 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),
b06c4bf5 1719 key.objectid, key.offset);
83d4cfd4
JB		 1720 if (ret) {
1721 btrfs_abort_transaction(trans, root, ret);
1722 break;
1723 }
5d4f98a2
YZ		 1724
1725 ret = btrfs_free_extent(trans, root, bytenr, num_bytes,
1726 parent, btrfs_header_owner(leaf),
b06c4bf5 1727 key.objectid, key.offset);
83d4cfd4
JB		 1728 if (ret) {
1729 btrfs_abort_transaction(trans, root, ret);
1730 break;
1731 }
5d4f98a2
YZ		 1732 }
1733 if (dirty)
1734 btrfs_mark_buffer_dirty(leaf);
3fd0a558
YZ		 1735 if (inode)
1736 btrfs_add_delayed_iput(inode);
83d4cfd4 1737 return ret;
5d4f98a2
YZ		 1738}
1739
1740static noinline_for_stack
1741int 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 */
3fd0a558
YZ		 1760static noinline_for_stack
1761int 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)
5d4f98a2
YZ		 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;
3fd0a558 1775 int cow = 0;
5d4f98a2
YZ		 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);
5d4f98a2
YZ		 1782
1783 last_snapshot = btrfs_root_last_snapshot(&src->root_item);
3fd0a558 1784again:
5d4f98a2
YZ		 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
3fd0a558
YZ		 1798 if (cow) {
1799 ret = btrfs_cow_block(trans, dest, eb, NULL, 0, &eb);
1800 BUG_ON(ret);
1801 }
5d4f98a2
YZ		 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);
707e8a07 1823 blocksize = dest->nodesize;
5d4f98a2
YZ		 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
fae7f21c 1837 if (WARN_ON(new_bytenr > 0 && new_bytenr == old_bytenr)) {
5d4f98a2
YZ		 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)) {
3fd0a558 1844 if (level <= lowest_level) {
5d4f98a2
YZ		 1845 ret = 0;
1846 break;
1847 }
1848
ce86cd59 1849 eb = read_tree_block(dest, old_bytenr, old_ptr_gen);
64c043de
LB		 1850 if (IS_ERR(eb)) {
1851 ret = PTR_ERR(eb);
1852 } else if (!extent_buffer_uptodate(eb)) {
1853 ret = -EIO;
416bc658 1854 free_extent_buffer(eb);
379cde74 1855 break;
416bc658 1856 }
5d4f98a2 1857 btrfs_tree_lock(eb);
3fd0a558
YZ		 1858 if (cow) {
1859 ret = btrfs_cow_block(trans, dest, eb, parent,
1860 slot, &eb);
1861 BUG_ON(ret);
5d4f98a2 1862 }
3fd0a558 1863 btrfs_set_lock_blocking(eb);
5d4f98a2
YZ		 1864
1865 btrfs_tree_unlock(parent);
1866 free_extent_buffer(parent);
1867
1868 parent = eb;
1869 continue;
1870 }
1871
3fd0a558
YZ		 1872 if (!cow) {
1873 btrfs_tree_unlock(parent);
1874 free_extent_buffer(parent);
1875 cow = 1;
1876 goto again;
1877 }
1878
5d4f98a2
YZ		 1879 btrfs_node_key_to_cpu(path->nodes[level], &key,
1880 path->slots[level]);
b3b4aa74 1881 btrfs_release_path(path);
5d4f98a2
YZ		 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,
b06c4bf5 1903 src->root_key.objectid, level - 1, 0);
5d4f98a2
YZ		 1904 BUG_ON(ret);
1905 ret = btrfs_inc_extent_ref(trans, dest, new_bytenr, blocksize,
1906 0, dest->root_key.objectid, level - 1,
b06c4bf5 1907 0);
5d4f98a2
YZ		 1908 BUG_ON(ret);
1909
1910 ret = btrfs_free_extent(trans, src, new_bytenr, blocksize,
1911 path->nodes[level]->start,
b06c4bf5 1912 src->root_key.objectid, level - 1, 0);
5d4f98a2
YZ		 1913 BUG_ON(ret);
1914
1915 ret = btrfs_free_extent(trans, dest, old_bytenr, blocksize,
1916 0, dest->root_key.objectid, level - 1,
b06c4bf5 1917 0);
5d4f98a2
YZ		 1918 BUG_ON(ret);
1919
1920 btrfs_unlock_up_safe(path, 0);
1921
1922 ret = level;
1923 break;
1924 }
1925 btrfs_tree_unlock(parent);
1926 free_extent_buffer(parent);
1927 return ret;
1928}
1929
1930/*
1931 * helper to find next relocated block in reloc tree
1932 */
1933static noinline_for_stack
1934int walk_up_reloc_tree(struct btrfs_root *root, struct btrfs_path *path,
1935 int *level)
1936{
1937 struct extent_buffer *eb;
1938 int i;
1939 u64 last_snapshot;
1940 u32 nritems;
1941
1942 last_snapshot = btrfs_root_last_snapshot(&root->root_item);
1943
1944 for (i = 0; i < *level; i++) {
1945 free_extent_buffer(path->nodes[i]);
1946 path->nodes[i] = NULL;
1947 }
1948
1949 for (i = *level; i < BTRFS_MAX_LEVEL && path->nodes[i]; i++) {
1950 eb = path->nodes[i];
1951 nritems = btrfs_header_nritems(eb);
1952 while (path->slots[i] + 1 < nritems) {
1953 path->slots[i]++;
1954 if (btrfs_node_ptr_generation(eb, path->slots[i]) <=
1955 last_snapshot)
1956 continue;
1957
1958 *level = i;
1959 return 0;
1960 }
1961 free_extent_buffer(path->nodes[i]);
1962 path->nodes[i] = NULL;
1963 }
1964 return 1;
1965}
1966
1967/*
1968 * walk down reloc tree to find relocated block of lowest level
1969 */
1970static noinline_for_stack
1971int walk_down_reloc_tree(struct btrfs_root *root, struct btrfs_path *path,
1972 int *level)
1973{
1974 struct extent_buffer *eb = NULL;
1975 int i;
1976 u64 bytenr;
1977 u64 ptr_gen = 0;
1978 u64 last_snapshot;
5d4f98a2
YZ		 1979 u32 nritems;
1980
1981 last_snapshot = btrfs_root_last_snapshot(&root->root_item);
1982
1983 for (i = *level; i > 0; i--) {
1984 eb = path->nodes[i];
1985 nritems = btrfs_header_nritems(eb);
1986 while (path->slots[i] < nritems) {
1987 ptr_gen = btrfs_node_ptr_generation(eb, path->slots[i]);
1988 if (ptr_gen > last_snapshot)
1989 break;
1990 path->slots[i]++;
1991 }
1992 if (path->slots[i] >= nritems) {
1993 if (i == *level)
1994 break;
1995 *level = i + 1;
1996 return 0;
1997 }
1998 if (i == 1) {
1999 *level = i;
2000 return 0;
2001 }
2002
2003 bytenr = btrfs_node_blockptr(eb, path->slots[i]);
ce86cd59 2004 eb = read_tree_block(root, bytenr, ptr_gen);
64c043de
LB		 2005 if (IS_ERR(eb)) {
2006 return PTR_ERR(eb);
2007 } else if (!extent_buffer_uptodate(eb)) {
416bc658
JB		 2008 free_extent_buffer(eb);
2009 return -EIO;
2010 }
5d4f98a2
YZ		 2011 BUG_ON(btrfs_header_level(eb) != i - 1);
2012 path->nodes[i - 1] = eb;
2013 path->slots[i - 1] = 0;
2014 }
2015 return 1;
2016}
2017
2018/*
2019 * invalidate extent cache for file extents whose key in range of
2020 * [min_key, max_key)
2021 */
2022static int invalidate_extent_cache(struct btrfs_root *root,
2023 struct btrfs_key *min_key,
2024 struct btrfs_key *max_key)
2025{
2026 struct inode *inode = NULL;
2027 u64 objectid;
2028 u64 start, end;
33345d01 2029 u64 ino;
5d4f98a2
YZ		 2030
2031 objectid = min_key->objectid;
2032 while (1) {
2033 cond_resched();
2034 iput(inode);
2035
2036 if (objectid > max_key->objectid)
2037 break;
2038
2039 inode = find_next_inode(root, objectid);
2040 if (!inode)
2041 break;
33345d01 2042 ino = btrfs_ino(inode);
5d4f98a2 2043
33345d01 2044 if (ino > max_key->objectid) {
5d4f98a2
YZ		 2045 iput(inode);
2046 break;
2047 }
2048
33345d01 2049 objectid = ino + 1;
5d4f98a2
YZ		 2050 if (!S_ISREG(inode->i_mode))
2051 continue;
2052
33345d01 2053 if (unlikely(min_key->objectid == ino)) {
5d4f98a2
YZ		 2054 if (min_key->type > BTRFS_EXTENT_DATA_KEY)
2055 continue;
2056 if (min_key->type < BTRFS_EXTENT_DATA_KEY)
2057 start = 0;
2058 else {
2059 start = min_key->offset;
2060 WARN_ON(!IS_ALIGNED(start, root->sectorsize));
2061 }
2062 } else {
2063 start = 0;
2064 }
2065
33345d01 2066 if (unlikely(max_key->objectid == ino)) {
5d4f98a2
YZ		 2067 if (max_key->type < BTRFS_EXTENT_DATA_KEY)
2068 continue;
2069 if (max_key->type > BTRFS_EXTENT_DATA_KEY) {
2070 end = (u64)-1;
2071 } else {
2072 if (max_key->offset == 0)
2073 continue;
2074 end = max_key->offset;
2075 WARN_ON(!IS_ALIGNED(end, root->sectorsize));
2076 end--;
2077 }
2078 } else {
2079 end = (u64)-1;
2080 }
2081
2082 /* the lock_extent waits for readpage to complete */
d0082371 2083 lock_extent(&BTRFS_I(inode)->io_tree, start, end);
5d4f98a2 2084 btrfs_drop_extent_cache(inode, start, end, 1);
d0082371 2085 unlock_extent(&BTRFS_I(inode)->io_tree, start, end);
5d4f98a2
YZ		 2086 }
2087 return 0;
2088}
2089
2090static int find_next_key(struct btrfs_path *path, int level,
2091 struct btrfs_key *key)
2092
2093{
2094 while (level < BTRFS_MAX_LEVEL) {
2095 if (!path->nodes[level])
2096 break;
2097 if (path->slots[level] + 1 <
2098 btrfs_header_nritems(path->nodes[level])) {
2099 btrfs_node_key_to_cpu(path->nodes[level], key,
2100 path->slots[level] + 1);
2101 return 0;
2102 }
2103 level++;
2104 }
2105 return 1;
2106}
2107
2108/*
2109 * merge the relocated tree blocks in reloc tree with corresponding
2110 * fs tree.
2111 */
2112static noinline_for_stack int merge_reloc_root(struct reloc_control *rc,
2113 struct btrfs_root *root)
2114{
2115 LIST_HEAD(inode_list);
2116 struct btrfs_key key;
2117 struct btrfs_key next_key;
9e6a0c52 2118 struct btrfs_trans_handle *trans = NULL;
5d4f98a2
YZ		 2119 struct btrfs_root *reloc_root;
2120 struct btrfs_root_item *root_item;
2121 struct btrfs_path *path;
3fd0a558 2122 struct extent_buffer *leaf;
5d4f98a2
YZ		 2123 int level;
2124 int max_level;
2125 int replaced = 0;
2126 int ret;
2127 int err = 0;
3fd0a558 2128 u32 min_reserved;
5d4f98a2
YZ		 2129
2130 path = btrfs_alloc_path();
2131 if (!path)
2132 return -ENOMEM;
e4058b54 2133 path->reada = READA_FORWARD;
5d4f98a2
YZ		 2134
2135 reloc_root = root->reloc_root;
2136 root_item = &reloc_root->root_item;
2137
2138 if (btrfs_disk_key_objectid(&root_item->drop_progress) == 0) {
2139 level = btrfs_root_level(root_item);
2140 extent_buffer_get(reloc_root->node);
2141 path->nodes[level] = reloc_root->node;
2142 path->slots[level] = 0;
2143 } else {
2144 btrfs_disk_key_to_cpu(&key, &root_item->drop_progress);
2145
2146 level = root_item->drop_level;
2147 BUG_ON(level == 0);
2148 path->lowest_level = level;
2149 ret = btrfs_search_slot(NULL, reloc_root, &key, path, 0, 0);
33c66f43 2150 path->lowest_level = 0;
5d4f98a2
YZ		 2151 if (ret < 0) {
2152 btrfs_free_path(path);
2153 return ret;
2154 }
2155
2156 btrfs_node_key_to_cpu(path->nodes[level], &next_key,
2157 path->slots[level]);
2158 WARN_ON(memcmp(&key, &next_key, sizeof(key)));
2159
2160 btrfs_unlock_up_safe(path, 0);
2161 }
2162
3fd0a558
YZ		 2163 min_reserved = root->nodesize * (BTRFS_MAX_LEVEL - 1) * 2;
2164 memset(&next_key, 0, sizeof(next_key));
5d4f98a2 2165
3fd0a558 2166 while (1) {
08e007d2
MX		 2167 ret = btrfs_block_rsv_refill(root, rc->block_rsv, min_reserved,
2168 BTRFS_RESERVE_FLUSH_ALL);
3fd0a558 2169 if (ret) {
9e6a0c52
JB		 2170 err = ret;
2171 goto out;
5d4f98a2 2172 }
9e6a0c52
JB		 2173 trans = btrfs_start_transaction(root, 0);
2174 if (IS_ERR(trans)) {
2175 err = PTR_ERR(trans);
2176 trans = NULL;
2177 goto out;
2178 }
2179 trans->block_rsv = rc->block_rsv;
5d4f98a2 2180
5d4f98a2 2181 replaced = 0;
5d4f98a2
YZ		 2182 max_level = level;
2183
2184 ret = walk_down_reloc_tree(reloc_root, path, &level);
2185 if (ret < 0) {
2186 err = ret;
2187 goto out;
2188 }
2189 if (ret > 0)
2190 break;
2191
2192 if (!find_next_key(path, level, &key) &&
2193 btrfs_comp_cpu_keys(&next_key, &key) >= 0) {
2194 ret = 0;
5d4f98a2 2195 } else {
3fd0a558
YZ		 2196 ret = replace_path(trans, root, reloc_root, path,
2197 &next_key, level, max_level);
5d4f98a2
YZ		 2198 }
2199 if (ret < 0) {
2200 err = ret;
2201 goto out;
2202 }
2203
2204 if (ret > 0) {
2205 level = ret;
2206 btrfs_node_key_to_cpu(path->nodes[level], &key,
2207 path->slots[level]);
2208 replaced = 1;
5d4f98a2
YZ		 2209 }
2210
2211 ret = walk_up_reloc_tree(reloc_root, path, &level);
2212 if (ret > 0)
2213 break;
2214
2215 BUG_ON(level == 0);
2216 /*
2217 * save the merging progress in the drop_progress.
2218 * this is OK since root refs == 1 in this case.
2219 */
2220 btrfs_node_key(path->nodes[level], &root_item->drop_progress,
2221 path->slots[level]);
2222 root_item->drop_level = level;
2223
3fd0a558 2224 btrfs_end_transaction_throttle(trans, root);
9e6a0c52 2225 trans = NULL;
5d4f98a2 2226
b53d3f5d 2227 btrfs_btree_balance_dirty(root);
5d4f98a2
YZ		 2228
2229 if (replaced && rc->stage == UPDATE_DATA_PTRS)
2230 invalidate_extent_cache(root, &key, &next_key);
2231 }
2232
2233 /*
2234 * handle the case only one block in the fs tree need to be
2235 * relocated and the block is tree root.
2236 */
2237 leaf = btrfs_lock_root_node(root);
2238 ret = btrfs_cow_block(trans, root, leaf, NULL, 0, &leaf);
2239 btrfs_tree_unlock(leaf);
2240 free_extent_buffer(leaf);
2241 if (ret < 0)
2242 err = ret;
2243out:
2244 btrfs_free_path(path);
2245
2246 if (err == 0) {
2247 memset(&root_item->drop_progress, 0,
2248 sizeof(root_item->drop_progress));
2249 root_item->drop_level = 0;
2250 btrfs_set_root_refs(root_item, 0);
3fd0a558 2251 btrfs_update_reloc_root(trans, root);
5d4f98a2
YZ		 2252 }
2253
9e6a0c52
JB		 2254 if (trans)
2255 btrfs_end_transaction_throttle(trans, root);
5d4f98a2 2256
b53d3f5d 2257 btrfs_btree_balance_dirty(root);
5d4f98a2 2258
5d4f98a2
YZ		 2259 if (replaced && rc->stage == UPDATE_DATA_PTRS)
2260 invalidate_extent_cache(root, &key, &next_key);
2261
2262 return err;
2263}
2264
3fd0a558
YZ		 2265static noinline_for_stack
2266int prepare_to_merge(struct reloc_control *rc, int err)
5d4f98a2 2267{
3fd0a558 2268 struct btrfs_root *root = rc->extent_root;
5d4f98a2 2269 struct btrfs_root *reloc_root;
3fd0a558
YZ		 2270 struct btrfs_trans_handle *trans;
2271 LIST_HEAD(reloc_roots);
2272 u64 num_bytes = 0;
2273 int ret;
3fd0a558 2274
7585717f 2275 mutex_lock(&root->fs_info->reloc_mutex);
3fd0a558
YZ		 2276 rc->merging_rsv_size += root->nodesize * (BTRFS_MAX_LEVEL - 1) * 2;
2277 rc->merging_rsv_size += rc->nodes_relocated * 2;
7585717f
CM		 2278 mutex_unlock(&root->fs_info->reloc_mutex);
2279
3fd0a558
YZ		 2280again:
2281 if (!err) {
2282 num_bytes = rc->merging_rsv_size;
08e007d2
MX		 2283 ret = btrfs_block_rsv_add(root, rc->block_rsv, num_bytes,
2284 BTRFS_RESERVE_FLUSH_ALL);
3fd0a558
YZ		 2285 if (ret)
2286 err = ret;
2287 }
2288
7a7eaa40 2289 trans = btrfs_join_transaction(rc->extent_root);
3612b495
TI		 2290 if (IS_ERR(trans)) {
2291 if (!err)
2292 btrfs_block_rsv_release(rc->extent_root,
2293 rc->block_rsv, num_bytes);
2294 return PTR_ERR(trans);
2295 }
3fd0a558
YZ		 2296
2297 if (!err) {
2298 if (num_bytes != rc->merging_rsv_size) {
2299 btrfs_end_transaction(trans, rc->extent_root);
2300 btrfs_block_rsv_release(rc->extent_root,
2301 rc->block_rsv, num_bytes);
3fd0a558
YZ		 2302 goto again;
2303 }
2304 }
5d4f98a2 2305
3fd0a558
YZ		 2306 rc->merge_reloc_tree = 1;
2307
2308 while (!list_empty(&rc->reloc_roots)) {
2309 reloc_root = list_entry(rc->reloc_roots.next,
2310 struct btrfs_root, root_list);
2311 list_del_init(&reloc_root->root_list);
5d4f98a2 2312
5d4f98a2
YZ		 2313 root = read_fs_root(reloc_root->fs_info,
2314 reloc_root->root_key.offset);
2315 BUG_ON(IS_ERR(root));
2316 BUG_ON(root->reloc_root != reloc_root);
2317
3fd0a558
YZ		 2318 /*
2319 * set reference count to 1, so btrfs_recover_relocation
2320 * knows it should resumes merging
2321 */
2322 if (!err)
2323 btrfs_set_root_refs(&reloc_root->root_item, 1);
5d4f98a2 2324 btrfs_update_reloc_root(trans, root);
5d4f98a2 2325
3fd0a558
YZ		 2326 list_add(&reloc_root->root_list, &reloc_roots);
2327 }
5d4f98a2 2328
3fd0a558 2329 list_splice(&reloc_roots, &rc->reloc_roots);
5d4f98a2 2330
3fd0a558
YZ		 2331 if (!err)
2332 btrfs_commit_transaction(trans, rc->extent_root);
2333 else
2334 btrfs_end_transaction(trans, rc->extent_root);
2335 return err;
5d4f98a2
YZ		 2336}
2337
aca1bba6
LB		 2338static noinline_for_stack
2339void free_reloc_roots(struct list_head *list)
2340{
2341 struct btrfs_root *reloc_root;
2342
2343 while (!list_empty(list)) {
2344 reloc_root = list_entry(list->next, struct btrfs_root,
2345 root_list);
c974c464 2346 __del_reloc_root(reloc_root);
aca1bba6
LB		 2347 }
2348}
2349
3fd0a558 2350static noinline_for_stack
94404e82 2351void merge_reloc_roots(struct reloc_control *rc)
5d4f98a2 2352{
5d4f98a2 2353 struct btrfs_root *root;
3fd0a558 2354 struct btrfs_root *reloc_root;
5bc7247a
MX		 2355 u64 last_snap;
2356 u64 otransid;
2357 u64 objectid;
3fd0a558
YZ		 2358 LIST_HEAD(reloc_roots);
2359 int found = 0;
aca1bba6 2360 int ret = 0;
3fd0a558
YZ		 2361again:
2362 root = rc->extent_root;
7585717f
CM		 2363
2364 /*
2365 * this serializes us with btrfs_record_root_in_transaction,
2366 * we have to make sure nobody is in the middle of
2367 * adding their roots to the list while we are
2368 * doing this splice
2369 */
2370 mutex_lock(&root->fs_info->reloc_mutex);
3fd0a558 2371 list_splice_init(&rc->reloc_roots, &reloc_roots);
7585717f 2372 mutex_unlock(&root->fs_info->reloc_mutex);
5d4f98a2 2373
3fd0a558
YZ		 2374 while (!list_empty(&reloc_roots)) {
2375 found = 1;
2376 reloc_root = list_entry(reloc_roots.next,
2377 struct btrfs_root, root_list);
5d4f98a2 2378
3fd0a558
YZ		 2379 if (btrfs_root_refs(&reloc_root->root_item) > 0) {
2380 root = read_fs_root(reloc_root->fs_info,
2381 reloc_root->root_key.offset);
2382 BUG_ON(IS_ERR(root));
2383 BUG_ON(root->reloc_root != reloc_root);
5d4f98a2 2384
3fd0a558 2385 ret = merge_reloc_root(rc, root);
b37b39cd 2386 if (ret) {
25e293c2
WS		 2387 if (list_empty(&reloc_root->root_list))
2388 list_add_tail(&reloc_root->root_list,
2389 &reloc_roots);
aca1bba6 2390 goto out;
b37b39cd 2391 }
3fd0a558
YZ		 2392 } else {
2393 list_del_init(&reloc_root->root_list);
2394 }
5bc7247a
MX		 2395
2396 /*
2397 * we keep the old last snapshod transid in rtranid when we
2398 * created the relocation tree.
2399 */
2400 last_snap = btrfs_root_rtransid(&reloc_root->root_item);
2401 otransid = btrfs_root_otransid(&reloc_root->root_item);
2402 objectid = reloc_root->root_key.offset;
2403
2c536799 2404 ret = btrfs_drop_snapshot(reloc_root, rc->block_rsv, 0, 1);
aca1bba6
LB		 2405 if (ret < 0) {
2406 if (list_empty(&reloc_root->root_list))
2407 list_add_tail(&reloc_root->root_list,
2408 &reloc_roots);
2409 goto out;
2410 }
5d4f98a2
YZ		 2411 }
2412
3fd0a558
YZ		 2413 if (found) {
2414 found = 0;
2415 goto again;
2416 }
aca1bba6
LB		 2417out:
2418 if (ret) {
a4553fef 2419 btrfs_std_error(root->fs_info, ret, NULL);
aca1bba6
LB		 2420 if (!list_empty(&reloc_roots))
2421 free_reloc_roots(&reloc_roots);
467bb1d2
WS		 2422
2423 /* new reloc root may be added */
2424 mutex_lock(&root->fs_info->reloc_mutex);
2425 list_splice_init(&rc->reloc_roots, &reloc_roots);
2426 mutex_unlock(&root->fs_info->reloc_mutex);
2427 if (!list_empty(&reloc_roots))
2428 free_reloc_roots(&reloc_roots);
aca1bba6
LB		 2429 }
2430
5d4f98a2 2431 BUG_ON(!RB_EMPTY_ROOT(&rc->reloc_root_tree.rb_root));
5d4f98a2
YZ		 2432}
2433
2434static void free_block_list(struct rb_root *blocks)
2435{
2436 struct tree_block *block;
2437 struct rb_node *rb_node;
2438 while ((rb_node = rb_first(blocks))) {
2439 block = rb_entry(rb_node, struct tree_block, rb_node);
2440 rb_erase(rb_node, blocks);
2441 kfree(block);
2442 }
2443}
2444
2445static int record_reloc_root_in_trans(struct btrfs_trans_handle *trans,
2446 struct btrfs_root *reloc_root)
2447{
2448 struct btrfs_root *root;
2449
2450 if (reloc_root->last_trans == trans->transid)
2451 return 0;
2452
2453 root = read_fs_root(reloc_root->fs_info, reloc_root->root_key.offset);
2454 BUG_ON(IS_ERR(root));
2455 BUG_ON(root->reloc_root != reloc_root);
2456
2457 return btrfs_record_root_in_trans(trans, root);
2458}
2459
3fd0a558
YZ		 2460static noinline_for_stack
2461struct btrfs_root *select_reloc_root(struct btrfs_trans_handle *trans,
2462 struct reloc_control *rc,
2463 struct backref_node *node,
dc4103f9 2464 struct backref_edge *edges[])
5d4f98a2
YZ		 2465{
2466 struct backref_node *next;
2467 struct btrfs_root *root;
3fd0a558
YZ		 2468 int index = 0;
2469
5d4f98a2
YZ		 2470 next = node;
2471 while (1) {
2472 cond_resched();
2473 next = walk_up_backref(next, edges, &index);
2474 root = next->root;
3fd0a558 2475 BUG_ON(!root);
27cdeb70 2476 BUG_ON(!test_bit(BTRFS_ROOT_REF_COWS, &root->state));
5d4f98a2
YZ		 2477
2478 if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID) {
2479 record_reloc_root_in_trans(trans, root);
2480 break;
2481 }
2482
3fd0a558
YZ		 2483 btrfs_record_root_in_trans(trans, root);
2484 root = root->reloc_root;
2485
2486 if (next->new_bytenr != root->node->start) {
2487 BUG_ON(next->new_bytenr);
2488 BUG_ON(!list_empty(&next->list));
2489 next->new_bytenr = root->node->start;
2490 next->root = root;
2491 list_add_tail(&next->list,
2492 &rc->backref_cache.changed);
2493 __mark_block_processed(rc, next);
5d4f98a2
YZ		 2494 break;
2495 }
2496
3fd0a558 2497 WARN_ON(1);
5d4f98a2
YZ		 2498 root = NULL;
2499 next = walk_down_backref(edges, &index);
2500 if (!next || next->level <= node->level)
2501 break;
2502 }
3fd0a558
YZ		 2503 if (!root)
2504 return NULL;
5d4f98a2 2505
3fd0a558
YZ		 2506 next = node;
2507 /* setup backref node path for btrfs_reloc_cow_block */
2508 while (1) {
2509 rc->backref_cache.path[next->level] = next;
2510 if (--index < 0)
2511 break;
2512 next = edges[index]->node[UPPER];
5d4f98a2 2513 }
5d4f98a2
YZ		 2514 return root;
2515}
2516
3fd0a558
YZ		 2517/*
2518 * select a tree root for relocation. return NULL if the block
2519 * is reference counted. we should use do_relocation() in this
2520 * case. return a tree root pointer if the block isn't reference
2521 * counted. return -ENOENT if the block is root of reloc tree.
2522 */
5d4f98a2 2523static noinline_for_stack
147d256e 2524struct btrfs_root *select_one_root(struct backref_node *node)
5d4f98a2 2525{
3fd0a558
YZ		 2526 struct backref_node *next;
2527 struct btrfs_root *root;
2528 struct btrfs_root *fs_root = NULL;
5d4f98a2 2529 struct backref_edge *edges[BTRFS_MAX_LEVEL - 1];
3fd0a558
YZ		 2530 int index = 0;
2531
2532 next = node;
2533 while (1) {
2534 cond_resched();
2535 next = walk_up_backref(next, edges, &index);
2536 root = next->root;
2537 BUG_ON(!root);
2538
25985edc 2539 /* no other choice for non-references counted tree */
27cdeb70 2540 if (!test_bit(BTRFS_ROOT_REF_COWS, &root->state))
3fd0a558
YZ		 2541 return root;
2542
2543 if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID)
2544 fs_root = root;
2545
2546 if (next != node)
2547 return NULL;
2548
2549 next = walk_down_backref(edges, &index);
2550 if (!next || next->level <= node->level)
2551 break;
2552 }
2553
2554 if (!fs_root)
2555 return ERR_PTR(-ENOENT);
2556 return fs_root;
5d4f98a2
YZ		 2557}
2558
2559static noinline_for_stack
3fd0a558
YZ		 2560u64 calcu_metadata_size(struct reloc_control *rc,
2561 struct backref_node *node, int reserve)
5d4f98a2 2562{
3fd0a558
YZ		 2563 struct backref_node *next = node;
2564 struct backref_edge *edge;
2565 struct backref_edge *edges[BTRFS_MAX_LEVEL - 1];
2566 u64 num_bytes = 0;
2567 int index = 0;
2568
2569 BUG_ON(reserve && node->processed);
2570
2571 while (next) {
2572 cond_resched();
2573 while (1) {
2574 if (next->processed && (reserve || next != node))
2575 break;
2576
707e8a07 2577 num_bytes += rc->extent_root->nodesize;
3fd0a558
YZ		 2578
2579 if (list_empty(&next->upper))
2580 break;
2581
2582 edge = list_entry(next->upper.next,
2583 struct backref_edge, list[LOWER]);
2584 edges[index++] = edge;
2585 next = edge->node[UPPER];
2586 }
2587 next = walk_down_backref(edges, &index);
2588 }
2589 return num_bytes;
5d4f98a2
YZ		 2590}
2591
3fd0a558
YZ		 2592static int reserve_metadata_space(struct btrfs_trans_handle *trans,
2593 struct reloc_control *rc,
2594 struct backref_node *node)
5d4f98a2 2595{
3fd0a558
YZ		 2596 struct btrfs_root *root = rc->extent_root;
2597 u64 num_bytes;
2598 int ret;
0647bf56 2599 u64 tmp;
3fd0a558
YZ		 2600
2601 num_bytes = calcu_metadata_size(rc, node, 1) * 2;
5d4f98a2 2602
3fd0a558 2603 trans->block_rsv = rc->block_rsv;
0647bf56
WS		 2604 rc->reserved_bytes += num_bytes;
2605 ret = btrfs_block_rsv_refill(root, rc->block_rsv, num_bytes,
2606 BTRFS_RESERVE_FLUSH_ALL);
3fd0a558 2607 if (ret) {
0647bf56
WS		 2608 if (ret == -EAGAIN) {
2609 tmp = rc->extent_root->nodesize *
2610 RELOCATION_RESERVED_NODES;
2611 while (tmp <= rc->reserved_bytes)
2612 tmp <<= 1;
2613 /*
2614 * only one thread can access block_rsv at this point,
2615 * so we don't need hold lock to protect block_rsv.
2616 * we expand more reservation size here to allow enough
2617 * space for relocation and we will return eailer in
2618 * enospc case.
2619 */
2620 rc->block_rsv->size = tmp + rc->extent_root->nodesize *
2621 RELOCATION_RESERVED_NODES;
2622 }
3fd0a558 2623 return ret;
5d4f98a2 2624 }
3fd0a558 2625
3fd0a558
YZ		 2626 return 0;
2627}
2628
5d4f98a2
YZ		 2629/*
2630 * relocate a block tree, and then update pointers in upper level
2631 * blocks that reference the block to point to the new location.
2632 *
2633 * if called by link_to_upper, the block has already been relocated.
2634 * in that case this function just updates pointers.
2635 */
2636static int do_relocation(struct btrfs_trans_handle *trans,
3fd0a558 2637 struct reloc_control *rc,
5d4f98a2
YZ		 2638 struct backref_node *node,
2639 struct btrfs_key *key,
2640 struct btrfs_path *path, int lowest)
2641{
2642 struct backref_node *upper;
2643 struct backref_edge *edge;
2644 struct backref_edge *edges[BTRFS_MAX_LEVEL - 1];
2645 struct btrfs_root *root;
2646 struct extent_buffer *eb;
2647 u32 blocksize;
2648 u64 bytenr;
2649 u64 generation;
5d4f98a2
YZ		 2650 int slot;
2651 int ret;
2652 int err = 0;
2653
2654 BUG_ON(lowest && node->eb);
2655
2656 path->lowest_level = node->level + 1;
3fd0a558 2657 rc->backref_cache.path[node->level] = node;
5d4f98a2
YZ		 2658 list_for_each_entry(edge, &node->upper, list[LOWER]) {
2659 cond_resched();
5d4f98a2
YZ		 2660
2661 upper = edge->node[UPPER];
dc4103f9 2662 root = select_reloc_root(trans, rc, upper, edges);
3fd0a558
YZ		 2663 BUG_ON(!root);
2664
2665 if (upper->eb && !upper->locked) {
2666 if (!lowest) {
2667 ret = btrfs_bin_search(upper->eb, key,
2668 upper->level, &slot);
2669 BUG_ON(ret);
2670 bytenr = btrfs_node_blockptr(upper->eb, slot);
2671 if (node->eb->start == bytenr)
2672 goto next;
2673 }
5d4f98a2 2674 drop_node_buffer(upper);
3fd0a558 2675 }
5d4f98a2
YZ		 2676
2677 if (!upper->eb) {
2678 ret = btrfs_search_slot(trans, root, key, path, 0, 1);
2679 if (ret < 0) {
2680 err = ret;
2681 break;
2682 }
2683 BUG_ON(ret > 0);
2684
3fd0a558
YZ		 2685 if (!upper->eb) {
2686 upper->eb = path->nodes[upper->level];
2687 path->nodes[upper->level] = NULL;
2688 } else {
2689 BUG_ON(upper->eb != path->nodes[upper->level]);
2690 }
5d4f98a2 2691
3fd0a558
YZ		 2692 upper->locked = 1;
2693 path->locks[upper->level] = 0;
5d4f98a2 2694
3fd0a558 2695 slot = path->slots[upper->level];
b3b4aa74 2696 btrfs_release_path(path);
5d4f98a2
YZ		 2697 } else {
2698 ret = btrfs_bin_search(upper->eb, key, upper->level,
2699 &slot);
2700 BUG_ON(ret);
2701 }
2702
2703 bytenr = btrfs_node_blockptr(upper->eb, slot);
3fd0a558
YZ		 2704 if (lowest) {
2705 BUG_ON(bytenr != node->bytenr);
5d4f98a2 2706 } else {
3fd0a558
YZ		 2707 if (node->eb->start == bytenr)
2708 goto next;
5d4f98a2
YZ		 2709 }
2710
707e8a07 2711 blocksize = root->nodesize;
5d4f98a2 2712 generation = btrfs_node_ptr_generation(upper->eb, slot);
ce86cd59 2713 eb = read_tree_block(root, bytenr, generation);
64c043de
LB		 2714 if (IS_ERR(eb)) {
2715 err = PTR_ERR(eb);
2716 goto next;
2717 } else if (!extent_buffer_uptodate(eb)) {
416bc658 2718 free_extent_buffer(eb);
97d9a8a4
TI		 2719 err = -EIO;
2720 goto next;
2721 }
5d4f98a2
YZ		 2722 btrfs_tree_lock(eb);
2723 btrfs_set_lock_blocking(eb);
2724
2725 if (!node->eb) {
2726 ret = btrfs_cow_block(trans, root, eb, upper->eb,
2727 slot, &eb);
3fd0a558
YZ		 2728 btrfs_tree_unlock(eb);
2729 free_extent_buffer(eb);
5d4f98a2
YZ		 2730 if (ret < 0) {
2731 err = ret;
3fd0a558 2732 goto next;
5d4f98a2 2733 }
3fd0a558 2734 BUG_ON(node->eb != eb);
5d4f98a2
YZ		 2735 } else {
2736 btrfs_set_node_blockptr(upper->eb, slot,
2737 node->eb->start);
2738 btrfs_set_node_ptr_generation(upper->eb, slot,
2739 trans->transid);
2740 btrfs_mark_buffer_dirty(upper->eb);
2741
2742 ret = btrfs_inc_extent_ref(trans, root,
2743 node->eb->start, blocksize,
2744 upper->eb->start,
2745 btrfs_header_owner(upper->eb),
b06c4bf5 2746 node->level, 0);
5d4f98a2
YZ		 2747 BUG_ON(ret);
2748
2749 ret = btrfs_drop_subtree(trans, root, eb, upper->eb);
2750 BUG_ON(ret);
5d4f98a2 2751 }
3fd0a558
YZ		 2752next:
2753 if (!upper->pending)
2754 drop_node_buffer(upper);
2755 else
2756 unlock_node_buffer(upper);
2757 if (err)
2758 break;
5d4f98a2 2759 }
3fd0a558
YZ		 2760
2761 if (!err && node->pending) {
2762 drop_node_buffer(node);
2763 list_move_tail(&node->list, &rc->backref_cache.changed);
2764 node->pending = 0;
2765 }
2766
5d4f98a2 2767 path->lowest_level = 0;
3fd0a558 2768 BUG_ON(err == -ENOSPC);
5d4f98a2
YZ		 2769 return err;
2770}
2771
2772static int link_to_upper(struct btrfs_trans_handle *trans,
3fd0a558 2773 struct reloc_control *rc,
5d4f98a2
YZ		 2774 struct backref_node *node,
2775 struct btrfs_path *path)
2776{
2777 struct btrfs_key key;
5d4f98a2
YZ		 2778
2779 btrfs_node_key_to_cpu(node->eb, &key, 0);
3fd0a558 2780 return do_relocation(trans, rc, node, &key, path, 0);
5d4f98a2
YZ		 2781}
2782
2783static int finish_pending_nodes(struct btrfs_trans_handle *trans,
3fd0a558
YZ		 2784 struct reloc_control *rc,
2785 struct btrfs_path *path, int err)
5d4f98a2 2786{
3fd0a558
YZ		 2787 LIST_HEAD(list);
2788 struct backref_cache *cache = &rc->backref_cache;
5d4f98a2
YZ		 2789 struct backref_node *node;
2790 int level;
2791 int ret;
5d4f98a2
YZ		 2792
2793 for (level = 0; level < BTRFS_MAX_LEVEL; level++) {
2794 while (!list_empty(&cache->pending[level])) {
2795 node = list_entry(cache->pending[level].next,
3fd0a558
YZ		 2796 struct backref_node, list);
2797 list_move_tail(&node->list, &list);
2798 BUG_ON(!node->pending);
5d4f98a2 2799
3fd0a558
YZ		 2800 if (!err) {
2801 ret = link_to_upper(trans, rc, node, path);
2802 if (ret < 0)
2803 err = ret;
2804 }
5d4f98a2 2805 }
3fd0a558 2806 list_splice_init(&list, &cache->pending[level]);
5d4f98a2 2807 }
5d4f98a2
YZ		 2808 return err;
2809}
2810
2811static void mark_block_processed(struct reloc_control *rc,
3fd0a558
YZ		 2812 u64 bytenr, u32 blocksize)
2813{
2814 set_extent_bits(&rc->processed_blocks, bytenr, bytenr + blocksize - 1,
2815 EXTENT_DIRTY, GFP_NOFS);
2816}
2817
2818static void __mark_block_processed(struct reloc_control *rc,
2819 struct backref_node *node)
5d4f98a2
YZ		 2820{
2821 u32 blocksize;
2822 if (node->level == 0 ||
2823 in_block_group(node->bytenr, rc->block_group)) {
707e8a07 2824 blocksize = rc->extent_root->nodesize;
3fd0a558 2825 mark_block_processed(rc, node->bytenr, blocksize);
5d4f98a2
YZ		 2826 }
2827 node->processed = 1;
2828}
2829
2830/*
2831 * mark a block and all blocks directly/indirectly reference the block
2832 * as processed.
2833 */
2834static void update_processed_blocks(struct reloc_control *rc,
2835 struct backref_node *node)
2836{
2837 struct backref_node *next = node;
2838 struct backref_edge *edge;
2839 struct backref_edge *edges[BTRFS_MAX_LEVEL - 1];
2840 int index = 0;
2841
2842 while (next) {
2843 cond_resched();
2844 while (1) {
2845 if (next->processed)
2846 break;
2847
3fd0a558 2848 __mark_block_processed(rc, next);
5d4f98a2
YZ		 2849
2850 if (list_empty(&next->upper))
2851 break;
2852
2853 edge = list_entry(next->upper.next,
2854 struct backref_edge, list[LOWER]);
2855 edges[index++] = edge;
2856 next = edge->node[UPPER];
2857 }
2858 next = walk_down_backref(edges, &index);
2859 }
2860}
2861
7476dfda 2862static int tree_block_processed(u64 bytenr, struct reloc_control *rc)
3fd0a558 2863{
7476dfda
DS		 2864 u32 blocksize = rc->extent_root->nodesize;
2865
3fd0a558
YZ		 2866 if (test_range_bit(&rc->processed_blocks, bytenr,
2867 bytenr + blocksize - 1, EXTENT_DIRTY, 1, NULL))
2868 return 1;
2869 return 0;
5d4f98a2
YZ		 2870}
2871
2872static int get_tree_block_key(struct reloc_control *rc,
2873 struct tree_block *block)
2874{
2875 struct extent_buffer *eb;
2876
2877 BUG_ON(block->key_ready);
2878 eb = read_tree_block(rc->extent_root, block->bytenr,
ce86cd59 2879 block->key.offset);
64c043de
LB		 2880 if (IS_ERR(eb)) {
2881 return PTR_ERR(eb);
2882 } else if (!extent_buffer_uptodate(eb)) {
416bc658
JB		 2883 free_extent_buffer(eb);
2884 return -EIO;
2885 }
5d4f98a2
YZ		 2886 WARN_ON(btrfs_header_level(eb) != block->level);
2887 if (block->level == 0)
2888 btrfs_item_key_to_cpu(eb, &block->key, 0);
2889 else
2890 btrfs_node_key_to_cpu(eb, &block->key, 0);
2891 free_extent_buffer(eb);
2892 block->key_ready = 1;
2893 return 0;
2894}
2895
5d4f98a2
YZ		 2896/*
2897 * helper function to relocate a tree block
2898 */
2899static int relocate_tree_block(struct btrfs_trans_handle *trans,
2900 struct reloc_control *rc,
2901 struct backref_node *node,
2902 struct btrfs_key *key,
2903 struct btrfs_path *path)
2904{
2905 struct btrfs_root *root;
3fd0a558
YZ		 2906 int ret = 0;
2907
2908 if (!node)
2909 return 0;
5d4f98a2 2910
3fd0a558 2911 BUG_ON(node->processed);
147d256e 2912 root = select_one_root(node);
3fd0a558 2913 if (root == ERR_PTR(-ENOENT)) {
5d4f98a2 2914 update_processed_blocks(rc, node);
3fd0a558 2915 goto out;
5d4f98a2
YZ		 2916 }
2917
27cdeb70 2918 if (!root || test_bit(BTRFS_ROOT_REF_COWS, &root->state)) {
3fd0a558
YZ		 2919 ret = reserve_metadata_space(trans, rc, node);
2920 if (ret)
5d4f98a2 2921 goto out;
5d4f98a2
YZ		 2922 }
2923
3fd0a558 2924 if (root) {
27cdeb70 2925 if (test_bit(BTRFS_ROOT_REF_COWS, &root->state)) {
3fd0a558
YZ		 2926 BUG_ON(node->new_bytenr);
2927 BUG_ON(!list_empty(&node->list));
2928 btrfs_record_root_in_trans(trans, root);
2929 root = root->reloc_root;
2930 node->new_bytenr = root->node->start;
2931 node->root = root;
2932 list_add_tail(&node->list, &rc->backref_cache.changed);
2933 } else {
2934 path->lowest_level = node->level;
2935 ret = btrfs_search_slot(trans, root, key, path, 0, 1);
b3b4aa74 2936 btrfs_release_path(path);
3fd0a558
YZ		 2937 if (ret > 0)
2938 ret = 0;
2939 }
2940 if (!ret)
2941 update_processed_blocks(rc, node);
2942 } else {
2943 ret = do_relocation(trans, rc, node, key, path, 1);
2944 }
5d4f98a2 2945out:
0647bf56 2946 if (ret || node->level == 0 || node->cowonly)
3fd0a558 2947 remove_backref_node(&rc->backref_cache, node);
5d4f98a2
YZ		 2948 return ret;
2949}
2950
2951/*
2952 * relocate a list of blocks
2953 */
2954static noinline_for_stack
2955int relocate_tree_blocks(struct btrfs_trans_handle *trans,
2956 struct reloc_control *rc, struct rb_root *blocks)
2957{
5d4f98a2
YZ		 2958 struct backref_node *node;
2959 struct btrfs_path *path;
2960 struct tree_block *block;
2961 struct rb_node *rb_node;
5d4f98a2
YZ		 2962 int ret;
2963 int err = 0;
2964
2965 path = btrfs_alloc_path();
e1a12670
LB		 2966 if (!path) {
2967 err = -ENOMEM;
34c2b290 2968 goto out_free_blocks;
e1a12670 2969 }
5d4f98a2 2970
5d4f98a2
YZ		 2971 rb_node = rb_first(blocks);
2972 while (rb_node) {
2973 block = rb_entry(rb_node, struct tree_block, rb_node);
5d4f98a2 2974 if (!block->key_ready)
d3e46fea 2975 readahead_tree_block(rc->extent_root, block->bytenr);
5d4f98a2
YZ		 2976 rb_node = rb_next(rb_node);
2977 }
2978
2979 rb_node = rb_first(blocks);
2980 while (rb_node) {
2981 block = rb_entry(rb_node, struct tree_block, rb_node);
34c2b290
DS		 2982 if (!block->key_ready) {
2983 err = get_tree_block_key(rc, block);
2984 if (err)
2985 goto out_free_path;
2986 }
5d4f98a2
YZ		 2987 rb_node = rb_next(rb_node);
2988 }
2989
2990 rb_node = rb_first(blocks);
2991 while (rb_node) {
2992 block = rb_entry(rb_node, struct tree_block, rb_node);
2993
3fd0a558 2994 node = build_backref_tree(rc, &block->key,
5d4f98a2
YZ		 2995 block->level, block->bytenr);
2996 if (IS_ERR(node)) {
2997 err = PTR_ERR(node);
2998 goto out;
2999 }
3000
3001 ret = relocate_tree_block(trans, rc, node, &block->key,
3002 path);
3003 if (ret < 0) {
3fd0a558
YZ		 3004 if (ret != -EAGAIN || rb_node == rb_first(blocks))
3005 err = ret;
5d4f98a2
YZ		 3006 goto out;
3007 }
5d4f98a2
YZ		 3008 rb_node = rb_next(rb_node);
3009 }
5d4f98a2 3010out:
3fd0a558 3011 err = finish_pending_nodes(trans, rc, path, err);
5d4f98a2 3012
34c2b290 3013out_free_path:
5d4f98a2 3014 btrfs_free_path(path);
34c2b290 3015out_free_blocks:
e1a12670 3016 free_block_list(blocks);
5d4f98a2
YZ		 3017 return err;
3018}
3019
efa56464
YZ		 3020static noinline_for_stack
3021int prealloc_file_extent_cluster(struct inode *inode,
3022 struct file_extent_cluster *cluster)
3023{
3024 u64 alloc_hint = 0;
3025 u64 start;
3026 u64 end;
3027 u64 offset = BTRFS_I(inode)->index_cnt;
3028 u64 num_bytes;
3029 int nr = 0;
3030 int ret = 0;
3031
3032 BUG_ON(cluster->start != cluster->boundary[0]);
5955102c 3033 inode_lock(inode);
efa56464 3034
7cf5b976
QW		 3035 ret = btrfs_check_data_free_space(inode, cluster->start,
3036 cluster->end + 1 - cluster->start);
efa56464
YZ		 3037 if (ret)
3038 goto out;
3039
3040 while (nr < cluster->nr) {
3041 start = cluster->boundary[nr] - offset;
3042 if (nr + 1 < cluster->nr)
3043 end = cluster->boundary[nr + 1] - 1 - offset;
3044 else
3045 end = cluster->end - offset;
3046
d0082371 3047 lock_extent(&BTRFS_I(inode)->io_tree, start, end);
efa56464
YZ		 3048 num_bytes = end + 1 - start;
3049 ret = btrfs_prealloc_file_range(inode, 0, start,
3050 num_bytes, num_bytes,
3051 end + 1, &alloc_hint);
d0082371 3052 unlock_extent(&BTRFS_I(inode)->io_tree, start, end);
efa56464
YZ		 3053 if (ret)
3054 break;
3055 nr++;
3056 }
7cf5b976
QW		 3057 btrfs_free_reserved_data_space(inode, cluster->start,
3058 cluster->end + 1 - cluster->start);
efa56464 3059out:
5955102c 3060 inode_unlock(inode);
efa56464
YZ		 3061 return ret;
3062}
3063
5d4f98a2 3064static noinline_for_stack
0257bb82
YZ		 3065int setup_extent_mapping(struct inode *inode, u64 start, u64 end,
3066 u64 block_start)
3067{
3068 struct btrfs_root *root = BTRFS_I(inode)->root;
3069 struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
3070 struct extent_map *em;
3071 int ret = 0;
3072
172ddd60 3073 em = alloc_extent_map();
0257bb82
YZ		 3074 if (!em)
3075 return -ENOMEM;
3076
3077 em->start = start;
3078 em->len = end + 1 - start;
3079 em->block_len = em->len;
3080 em->block_start = block_start;
3081 em->bdev = root->fs_info->fs_devices->latest_bdev;
3082 set_bit(EXTENT_FLAG_PINNED, &em->flags);
3083
d0082371 3084 lock_extent(&BTRFS_I(inode)->io_tree, start, end);
0257bb82
YZ		 3085 while (1) {
3086 write_lock(&em_tree->lock);
09a2a8f9 3087 ret = add_extent_mapping(em_tree, em, 0);
0257bb82
YZ		 3088 write_unlock(&em_tree->lock);
3089 if (ret != -EEXIST) {
3090 free_extent_map(em);
3091 break;
3092 }
3093 btrfs_drop_extent_cache(inode, start, end, 0);
3094 }
d0082371 3095 unlock_extent(&BTRFS_I(inode)->io_tree, start, end);
0257bb82
YZ		 3096 return ret;
3097}
3098
3099static int relocate_file_extent_cluster(struct inode *inode,
3100 struct file_extent_cluster *cluster)
5d4f98a2
YZ		 3101{
3102 u64 page_start;
3103 u64 page_end;
0257bb82
YZ		 3104 u64 offset = BTRFS_I(inode)->index_cnt;
3105 unsigned long index;
5d4f98a2 3106 unsigned long last_index;
5d4f98a2
YZ		 3107 struct page *page;
3108 struct file_ra_state *ra;
3b16a4e3 3109 gfp_t mask = btrfs_alloc_write_mask(inode->i_mapping);
0257bb82 3110 int nr = 0;
5d4f98a2
YZ		 3111 int ret = 0;
3112
0257bb82
YZ		 3113 if (!cluster->nr)
3114 return 0;
3115
5d4f98a2
YZ		 3116 ra = kzalloc(sizeof(*ra), GFP_NOFS);
3117 if (!ra)
3118 return -ENOMEM;
3119
efa56464
YZ		 3120 ret = prealloc_file_extent_cluster(inode, cluster);
3121 if (ret)
3122 goto out;
0257bb82 3123
efa56464 3124 file_ra_state_init(ra, inode->i_mapping);
5d4f98a2 3125
0257bb82
YZ		 3126 ret = setup_extent_mapping(inode, cluster->start - offset,
3127 cluster->end - offset, cluster->start);
5d4f98a2 3128 if (ret)
efa56464 3129 goto out;
5d4f98a2 3130
efa56464
YZ		 3131 index = (cluster->start - offset) >> PAGE_CACHE_SHIFT;
3132 last_index = (cluster->end - offset) >> PAGE_CACHE_SHIFT;
0257bb82 3133 while (index <= last_index) {
efa56464
YZ		 3134 ret = btrfs_delalloc_reserve_metadata(inode, PAGE_CACHE_SIZE);
3135 if (ret)
3136 goto out;
3137
0257bb82 3138 page = find_lock_page(inode->i_mapping, index);
5d4f98a2 3139 if (!page) {
0257bb82
YZ		 3140 page_cache_sync_readahead(inode->i_mapping,
3141 ra, NULL, index,
3142 last_index + 1 - index);
a94733d0 3143 page = find_or_create_page(inode->i_mapping, index,
3b16a4e3 3144 mask);
0257bb82 3145 if (!page) {
efa56464
YZ		 3146 btrfs_delalloc_release_metadata(inode,
3147 PAGE_CACHE_SIZE);
0257bb82 3148 ret = -ENOMEM;
efa56464 3149 goto out;
0257bb82 3150 }
5d4f98a2 3151 }
0257bb82
YZ		 3152
3153 if (PageReadahead(page)) {
3154 page_cache_async_readahead(inode->i_mapping,
3155 ra, NULL, page, index,
3156 last_index + 1 - index);
3157 }
3158
5d4f98a2
YZ		 3159 if (!PageUptodate(page)) {
3160 btrfs_readpage(NULL, page);
3161 lock_page(page);
3162 if (!PageUptodate(page)) {
3163 unlock_page(page);
3164 page_cache_release(page);
efa56464
YZ		 3165 btrfs_delalloc_release_metadata(inode,
3166 PAGE_CACHE_SIZE);
5d4f98a2 3167 ret = -EIO;
efa56464 3168 goto out;
5d4f98a2
YZ		 3169 }
3170 }
5d4f98a2 3171
4eee4fa4 3172 page_start = page_offset(page);
5d4f98a2 3173 page_end = page_start + PAGE_CACHE_SIZE - 1;
0257bb82 3174
d0082371 3175 lock_extent(&BTRFS_I(inode)->io_tree, page_start, page_end);
0257bb82 3176
5d4f98a2
YZ		 3177 set_page_extent_mapped(page);
3178
0257bb82
YZ		 3179 if (nr < cluster->nr &&
3180 page_start + offset == cluster->boundary[nr]) {
3181 set_extent_bits(&BTRFS_I(inode)->io_tree,
3182 page_start, page_end,
5d4f98a2 3183 EXTENT_BOUNDARY, GFP_NOFS);
0257bb82
YZ		 3184 nr++;
3185 }
5d4f98a2 3186
efa56464 3187 btrfs_set_extent_delalloc(inode, page_start, page_end, NULL);
5d4f98a2 3188 set_page_dirty(page);
5d4f98a2 3189
0257bb82 3190 unlock_extent(&BTRFS_I(inode)->io_tree,
d0082371 3191 page_start, page_end);
5d4f98a2
YZ		 3192 unlock_page(page);
3193 page_cache_release(page);
0257bb82
YZ		 3194
3195 index++;
efa56464
YZ		 3196 balance_dirty_pages_ratelimited(inode->i_mapping);
3197 btrfs_throttle(BTRFS_I(inode)->root);
5d4f98a2 3198 }
0257bb82 3199 WARN_ON(nr != cluster->nr);
efa56464 3200out:
5d4f98a2 3201 kfree(ra);
5d4f98a2
YZ		 3202 return ret;
3203}
3204
3205static noinline_for_stack
0257bb82
YZ		 3206int relocate_data_extent(struct inode *inode, struct btrfs_key *extent_key,
3207 struct file_extent_cluster *cluster)
5d4f98a2 3208{
0257bb82 3209 int ret;
5d4f98a2 3210
0257bb82
YZ		 3211 if (cluster->nr > 0 && extent_key->objectid != cluster->end + 1) {
3212 ret = relocate_file_extent_cluster(inode, cluster);
3213 if (ret)
3214 return ret;
3215 cluster->nr = 0;
5d4f98a2 3216 }
5d4f98a2 3217
0257bb82
YZ		 3218 if (!cluster->nr)
3219 cluster->start = extent_key->objectid;
3220 else
3221 BUG_ON(cluster->nr >= MAX_EXTENTS);
3222 cluster->end = extent_key->objectid + extent_key->offset - 1;
3223 cluster->boundary[cluster->nr] = extent_key->objectid;
3224 cluster->nr++;
3225
3226 if (cluster->nr >= MAX_EXTENTS) {
3227 ret = relocate_file_extent_cluster(inode, cluster);
3228 if (ret)
3229 return ret;
3230 cluster->nr = 0;
3231 }
3232 return 0;
5d4f98a2
YZ		 3233}
3234
3235#ifdef BTRFS_COMPAT_EXTENT_TREE_V0
3236static int get_ref_objectid_v0(struct reloc_control *rc,
3237 struct btrfs_path *path,
3238 struct btrfs_key *extent_key,
3239 u64 *ref_objectid, int *path_change)
3240{
3241 struct btrfs_key key;
3242 struct extent_buffer *leaf;
3243 struct btrfs_extent_ref_v0 *ref0;
3244 int ret;
3245 int slot;
3246
3247 leaf = path->nodes[0];
3248 slot = path->slots[0];
3249 while (1) {
3250 if (slot >= btrfs_header_nritems(leaf)) {
3251 ret = btrfs_next_leaf(rc->extent_root, path);
3252 if (ret < 0)
3253 return ret;
3254 BUG_ON(ret > 0);
3255 leaf = path->nodes[0];
3256 slot = path->slots[0];
3257 if (path_change)
3258 *path_change = 1;
3259 }
3260 btrfs_item_key_to_cpu(leaf, &key, slot);
3261 if (key.objectid != extent_key->objectid)
3262 return -ENOENT;
3263
3264 if (key.type != BTRFS_EXTENT_REF_V0_KEY) {
3265 slot++;
3266 continue;
3267 }
3268 ref0 = btrfs_item_ptr(leaf, slot,
3269 struct btrfs_extent_ref_v0);
3270 *ref_objectid = btrfs_ref_objectid_v0(leaf, ref0);
3271 break;
3272 }
3273 return 0;
3274}
3275#endif
3276
3277/*
3278 * helper to add a tree block to the list.
3279 * the major work is getting the generation and level of the block
3280 */
3281static int add_tree_block(struct reloc_control *rc,
3282 struct btrfs_key *extent_key,
3283 struct btrfs_path *path,
3284 struct rb_root *blocks)
3285{
3286 struct extent_buffer *eb;
3287 struct btrfs_extent_item *ei;
3288 struct btrfs_tree_block_info *bi;
3289 struct tree_block *block;
3290 struct rb_node *rb_node;
3291 u32 item_size;
3292 int level = -1;
7fdf4b60 3293 u64 generation;
5d4f98a2
YZ		 3294
3295 eb = path->nodes[0];
3296 item_size = btrfs_item_size_nr(eb, path->slots[0]);
3297
3173a18f
JB		 3298 if (extent_key->type == BTRFS_METADATA_ITEM_KEY ||
3299 item_size >= sizeof(*ei) + sizeof(*bi)) {
5d4f98a2
YZ		 3300 ei = btrfs_item_ptr(eb, path->slots[0],
3301 struct btrfs_extent_item);
3173a18f
JB		 3302 if (extent_key->type == BTRFS_EXTENT_ITEM_KEY) {
3303 bi = (struct btrfs_tree_block_info *)(ei + 1);
3304 level = btrfs_tree_block_level(eb, bi);
3305 } else {
3306 level = (int)extent_key->offset;
3307 }
5d4f98a2 3308 generation = btrfs_extent_generation(eb, ei);
5d4f98a2
YZ		 3309 } else {
3310#ifdef BTRFS_COMPAT_EXTENT_TREE_V0
3311 u64 ref_owner;
3312 int ret;
3313
3314 BUG_ON(item_size != sizeof(struct btrfs_extent_item_v0));
3315 ret = get_ref_objectid_v0(rc, path, extent_key,
3316 &ref_owner, NULL);
411fc6bc
AK		 3317 if (ret < 0)
3318 return ret;
5d4f98a2
YZ		 3319 BUG_ON(ref_owner >= BTRFS_MAX_LEVEL);
3320 level = (int)ref_owner;
3321 /* FIXME: get real generation */
3322 generation = 0;
3323#else
3324 BUG();
3325#endif
3326 }
3327
b3b4aa74 3328 btrfs_release_path(path);
5d4f98a2
YZ		 3329
3330 BUG_ON(level == -1);
3331
3332 block = kmalloc(sizeof(*block), GFP_NOFS);
3333 if (!block)
3334 return -ENOMEM;
3335
3336 block->bytenr = extent_key->objectid;
707e8a07 3337 block->key.objectid = rc->extent_root->nodesize;
5d4f98a2
YZ		 3338 block->key.offset = generation;
3339 block->level = level;
3340 block->key_ready = 0;
3341
3342 rb_node = tree_insert(blocks, block->bytenr, &block->rb_node);
43c04fb1
JM		 3343 if (rb_node)
3344 backref_tree_panic(rb_node, -EEXIST, block->bytenr);
5d4f98a2
YZ		 3345
3346 return 0;
3347}
3348
3349/*
3350 * helper to add tree blocks for backref of type BTRFS_SHARED_DATA_REF_KEY
3351 */
3352static int __add_tree_block(struct reloc_control *rc,
3353 u64 bytenr, u32 blocksize,
3354 struct rb_root *blocks)
3355{
3356 struct btrfs_path *path;
3357 struct btrfs_key key;
3358 int ret;
aee68ee5
JB		 3359 bool skinny = btrfs_fs_incompat(rc->extent_root->fs_info,
3360 SKINNY_METADATA);
5d4f98a2 3361
7476dfda 3362 if (tree_block_processed(bytenr, rc))
5d4f98a2
YZ		 3363 return 0;
3364
3365 if (tree_search(blocks, bytenr))
3366 return 0;
3367
3368 path = btrfs_alloc_path();
3369 if (!path)
3370 return -ENOMEM;
aee68ee5 3371again:
5d4f98a2 3372 key.objectid = bytenr;
aee68ee5
JB		 3373 if (skinny) {
3374 key.type = BTRFS_METADATA_ITEM_KEY;
3375 key.offset = (u64)-1;
3376 } else {
3377 key.type = BTRFS_EXTENT_ITEM_KEY;
3378 key.offset = blocksize;
3379 }
5d4f98a2
YZ		 3380
3381 path->search_commit_root = 1;
3382 path->skip_locking = 1;
3383 ret = btrfs_search_slot(NULL, rc->extent_root, &key, path, 0, 0);
3384 if (ret < 0)
3385 goto out;
5d4f98a2 3386
aee68ee5
JB		 3387 if (ret > 0 && skinny) {
3388 if (path->slots[0]) {
3389 path->slots[0]--;
3390 btrfs_item_key_to_cpu(path->nodes[0], &key,
3391 path->slots[0]);
3392 if (key.objectid == bytenr &&
3393 (key.type == BTRFS_METADATA_ITEM_KEY ||
3394 (key.type == BTRFS_EXTENT_ITEM_KEY &&
3395 key.offset == blocksize)))
3396 ret = 0;
3397 }
3398
3399 if (ret) {
3400 skinny = false;
3401 btrfs_release_path(path);
3402 goto again;
3403 }
3173a18f
JB		 3404 }
3405 BUG_ON(ret);
3406
5d4f98a2
YZ		 3407 ret = add_tree_block(rc, &key, path, blocks);
3408out:
3409 btrfs_free_path(path);
3410 return ret;
3411}
3412
3413/*
3414 * helper to check if the block use full backrefs for pointers in it
3415 */
3416static int block_use_full_backref(struct reloc_control *rc,
3417 struct extent_buffer *eb)
3418{
5d4f98a2
YZ		 3419 u64 flags;
3420 int ret;
3421
3422 if (btrfs_header_flag(eb, BTRFS_HEADER_FLAG_RELOC) ||
3423 btrfs_header_backref_rev(eb) < BTRFS_MIXED_BACKREF_REV)
3424 return 1;
3425
3fd0a558 3426 ret = btrfs_lookup_extent_info(NULL, rc->extent_root,
3173a18f
JB		 3427 eb->start, btrfs_header_level(eb), 1,
3428 NULL, &flags);
5d4f98a2
YZ		 3429 BUG_ON(ret);
3430
5d4f98a2
YZ		 3431 if (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF)
3432 ret = 1;
3433 else
3434 ret = 0;
5d4f98a2
YZ		 3435 return ret;
3436}
3437
0af3d00b 3438static int delete_block_group_cache(struct btrfs_fs_info *fs_info,
1bbc621e
CM		 3439 struct btrfs_block_group_cache *block_group,
3440 struct inode *inode,
3441 u64 ino)
0af3d00b
JB		 3442{
3443 struct btrfs_key key;
0af3d00b
JB		 3444 struct btrfs_root *root = fs_info->tree_root;
3445 struct btrfs_trans_handle *trans;
0af3d00b
JB		 3446 int ret = 0;
3447
3448 if (inode)
3449 goto truncate;
3450
3451 key.objectid = ino;
3452 key.type = BTRFS_INODE_ITEM_KEY;
3453 key.offset = 0;
3454
3455 inode = btrfs_iget(fs_info->sb, &key, root, NULL);
f54fb859
TI		 3456 if (IS_ERR(inode) || is_bad_inode(inode)) {
3457 if (!IS_ERR(inode))
0af3d00b
JB		 3458 iput(inode);
3459 return -ENOENT;
3460 }
3461
3462truncate:
7b61cd92
MX		 3463 ret = btrfs_check_trunc_cache_free_space(root,
3464 &fs_info->global_block_rsv);
3465 if (ret)
3466 goto out;
3467
7a7eaa40 3468 trans = btrfs_join_transaction(root);
0af3d00b 3469 if (IS_ERR(trans)) {
3612b495 3470 ret = PTR_ERR(trans);
0af3d00b
JB		 3471 goto out;
3472 }
3473
1bbc621e 3474 ret = btrfs_truncate_free_space_cache(root, trans, block_group, inode);
0af3d00b 3475
0af3d00b 3476 btrfs_end_transaction(trans, root);
b53d3f5d 3477 btrfs_btree_balance_dirty(root);
0af3d00b
JB		 3478out:
3479 iput(inode);
3480 return ret;
3481}
3482
5d4f98a2
YZ		 3483/*
3484 * helper to add tree blocks for backref of type BTRFS_EXTENT_DATA_REF_KEY
3485 * this function scans fs tree to find blocks reference the data extent
3486 */
3487static int find_data_references(struct reloc_control *rc,
3488 struct btrfs_key *extent_key,
3489 struct extent_buffer *leaf,
3490 struct btrfs_extent_data_ref *ref,
3491 struct rb_root *blocks)
3492{
3493 struct btrfs_path *path;
3494 struct tree_block *block;
3495 struct btrfs_root *root;
3496 struct btrfs_file_extent_item *fi;
3497 struct rb_node *rb_node;
3498 struct btrfs_key key;
3499 u64 ref_root;
3500 u64 ref_objectid;
3501 u64 ref_offset;
3502 u32 ref_count;
3503 u32 nritems;
3504 int err = 0;
3505 int added = 0;
3506 int counted;
3507 int ret;
3508
5d4f98a2
YZ		 3509 ref_root = btrfs_extent_data_ref_root(leaf, ref);
3510 ref_objectid = btrfs_extent_data_ref_objectid(leaf, ref);
3511 ref_offset = btrfs_extent_data_ref_offset(leaf, ref);
3512 ref_count = btrfs_extent_data_ref_count(leaf, ref);
3513
0af3d00b
JB		 3514 /*
3515 * This is an extent belonging to the free space cache, lets just delete
3516 * it and redo the search.
3517 */
3518 if (ref_root == BTRFS_ROOT_TREE_OBJECTID) {
3519 ret = delete_block_group_cache(rc->extent_root->fs_info,
1bbc621e 3520 rc->block_group,
0af3d00b
JB		 3521 NULL, ref_objectid);
3522 if (ret != -ENOENT)
3523 return ret;
3524 ret = 0;
3525 }
3526
3527 path = btrfs_alloc_path();
3528 if (!path)
3529 return -ENOMEM;
e4058b54 3530 path->reada = READA_FORWARD;
0af3d00b 3531
5d4f98a2
YZ		 3532 root = read_fs_root(rc->extent_root->fs_info, ref_root);
3533 if (IS_ERR(root)) {
3534 err = PTR_ERR(root);
3535 goto out;
3536 }
3537
3538 key.objectid = ref_objectid;
5d4f98a2 3539 key.type = BTRFS_EXTENT_DATA_KEY;
84850e8d
YZ		 3540 if (ref_offset > ((u64)-1 << 32))
3541 key.offset = 0;
3542 else
3543 key.offset = ref_offset;
5d4f98a2
YZ		 3544
3545 path->search_commit_root = 1;
3546 path->skip_locking = 1;
3547 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
3548 if (ret < 0) {
3549 err = ret;
3550 goto out;
3551 }
3552
3553 leaf = path->nodes[0];
3554 nritems = btrfs_header_nritems(leaf);
3555 /*
3556 * the references in tree blocks that use full backrefs
3557 * are not counted in
3558 */
3559 if (block_use_full_backref(rc, leaf))
3560 counted = 0;
3561 else
3562 counted = 1;
3563 rb_node = tree_search(blocks, leaf->start);
3564 if (rb_node) {
3565 if (counted)
3566 added = 1;
3567 else
3568 path->slots[0] = nritems;
3569 }
3570
3571 while (ref_count > 0) {
3572 while (path->slots[0] >= nritems) {
3573 ret = btrfs_next_leaf(root, path);
3574 if (ret < 0) {
3575 err = ret;
3576 goto out;
3577 }
fae7f21c 3578 if (WARN_ON(ret > 0))
5d4f98a2 3579 goto out;
5d4f98a2
YZ		 3580
3581 leaf = path->nodes[0];
3582 nritems = btrfs_header_nritems(leaf);
3583 added = 0;
3584
3585 if (block_use_full_backref(rc, leaf))
3586 counted = 0;
3587 else
3588 counted = 1;
3589 rb_node = tree_search(blocks, leaf->start);
3590 if (rb_node) {
3591 if (counted)
3592 added = 1;
3593 else
3594 path->slots[0] = nritems;
3595 }
3596 }
3597
3598 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
fae7f21c
DG		 3599 if (WARN_ON(key.objectid != ref_objectid ||
3600 key.type != BTRFS_EXTENT_DATA_KEY))
5d4f98a2 3601 break;
5d4f98a2
YZ		 3602
3603 fi = btrfs_item_ptr(leaf, path->slots[0],
3604 struct btrfs_file_extent_item);
3605
3606 if (btrfs_file_extent_type(leaf, fi) ==
3607 BTRFS_FILE_EXTENT_INLINE)
3608 goto next;
3609
3610 if (btrfs_file_extent_disk_bytenr(leaf, fi) !=
3611 extent_key->objectid)
3612 goto next;
3613
3614 key.offset -= btrfs_file_extent_offset(leaf, fi);
3615 if (key.offset != ref_offset)
3616 goto next;
3617
3618 if (counted)
3619 ref_count--;
3620 if (added)
3621 goto next;
3622
7476dfda 3623 if (!tree_block_processed(leaf->start, rc)) {
5d4f98a2
YZ		 3624 block = kmalloc(sizeof(*block), GFP_NOFS);
3625 if (!block) {
3626 err = -ENOMEM;
3627 break;
3628 }
3629 block->bytenr = leaf->start;
3630 btrfs_item_key_to_cpu(leaf, &block->key, 0);
3631 block->level = 0;
3632 block->key_ready = 1;
3633 rb_node = tree_insert(blocks, block->bytenr,
3634 &block->rb_node);
43c04fb1
JM		 3635 if (rb_node)
3636 backref_tree_panic(rb_node, -EEXIST,
3637 block->bytenr);
5d4f98a2
YZ		 3638 }
3639 if (counted)
3640 added = 1;
3641 else
3642 path->slots[0] = nritems;
3643next:
3644 path->slots[0]++;
3645
3646 }
3647out:
3648 btrfs_free_path(path);
3649 return err;
3650}
3651
3652/*
2c016dc2 3653 * helper to find all tree blocks that reference a given data extent
5d4f98a2
YZ		 3654 */
3655static noinline_for_stack
3656int add_data_references(struct reloc_control *rc,
3657 struct btrfs_key *extent_key,
3658 struct btrfs_path *path,
3659 struct rb_root *blocks)
3660{
3661 struct btrfs_key key;
3662 struct extent_buffer *eb;
3663 struct btrfs_extent_data_ref *dref;
3664 struct btrfs_extent_inline_ref *iref;
3665 unsigned long ptr;
3666 unsigned long end;
707e8a07 3667 u32 blocksize = rc->extent_root->nodesize;
647f63bd 3668 int ret = 0;
5d4f98a2
YZ		 3669 int err = 0;
3670
5d4f98a2
YZ		 3671 eb = path->nodes[0];
3672 ptr = btrfs_item_ptr_offset(eb, path->slots[0]);
3673 end = ptr + btrfs_item_size_nr(eb, path->slots[0]);
3674#ifdef BTRFS_COMPAT_EXTENT_TREE_V0
3675 if (ptr + sizeof(struct btrfs_extent_item_v0) == end)
3676 ptr = end;
3677 else
3678#endif
3679 ptr += sizeof(struct btrfs_extent_item);
3680
3681 while (ptr < end) {
3682 iref = (struct btrfs_extent_inline_ref *)ptr;
3683 key.type = btrfs_extent_inline_ref_type(eb, iref);
3684 if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
3685 key.offset = btrfs_extent_inline_ref_offset(eb, iref);
3686 ret = __add_tree_block(rc, key.offset, blocksize,
3687 blocks);
3688 } else if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
3689 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
3690 ret = find_data_references(rc, extent_key,
3691 eb, dref, blocks);
3692 } else {
3693 BUG();
3694 }
647f63bd
FDBM		 3695 if (ret) {
3696 err = ret;
3697 goto out;
3698 }
5d4f98a2
YZ		 3699 ptr += btrfs_extent_inline_ref_size(key.type);
3700 }
3701 WARN_ON(ptr > end);
3702
3703 while (1) {
3704 cond_resched();
3705 eb = path->nodes[0];
3706 if (path->slots[0] >= btrfs_header_nritems(eb)) {
3707 ret = btrfs_next_leaf(rc->extent_root, path);
3708 if (ret < 0) {
3709 err = ret;
3710 break;
3711 }
3712 if (ret > 0)
3713 break;
3714 eb = path->nodes[0];
3715 }
3716
3717 btrfs_item_key_to_cpu(eb, &key, path->slots[0]);
3718 if (key.objectid != extent_key->objectid)
3719 break;
3720
3721#ifdef BTRFS_COMPAT_EXTENT_TREE_V0
3722 if (key.type == BTRFS_SHARED_DATA_REF_KEY ||
3723 key.type == BTRFS_EXTENT_REF_V0_KEY) {
3724#else
3725 BUG_ON(key.type == BTRFS_EXTENT_REF_V0_KEY);
3726 if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
3727#endif
3728 ret = __add_tree_block(rc, key.offset, blocksize,
3729 blocks);
3730 } else if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
3731 dref = btrfs_item_ptr(eb, path->slots[0],
3732 struct btrfs_extent_data_ref);
3733 ret = find_data_references(rc, extent_key,
3734 eb, dref, blocks);
3735 } else {
3736 ret = 0;
3737 }
3738 if (ret) {
3739 err = ret;
3740 break;
3741 }
3742 path->slots[0]++;
3743 }
647f63bd 3744out:
b3b4aa74 3745 btrfs_release_path(path);
5d4f98a2
YZ		 3746 if (err)
3747 free_block_list(blocks);
3748 return err;
3749}
3750
3751/*
2c016dc2 3752 * helper to find next unprocessed extent
5d4f98a2
YZ		 3753 */
3754static noinline_for_stack
147d256e 3755int find_next_extent(struct reloc_control *rc, struct btrfs_path *path,
3fd0a558 3756 struct btrfs_key *extent_key)
5d4f98a2
YZ		 3757{
3758 struct btrfs_key key;
3759 struct extent_buffer *leaf;
3760 u64 start, end, last;
3761 int ret;
3762
3763 last = rc->block_group->key.objectid + rc->block_group->key.offset;
3764 while (1) {
3765 cond_resched();
3766 if (rc->search_start >= last) {
3767 ret = 1;
3768 break;
3769 }
3770
3771 key.objectid = rc->search_start;
3772 key.type = BTRFS_EXTENT_ITEM_KEY;
3773 key.offset = 0;
3774
3775 path->search_commit_root = 1;
3776 path->skip_locking = 1;
3777 ret = btrfs_search_slot(NULL, rc->extent_root, &key, path,
3778 0, 0);
3779 if (ret < 0)
3780 break;
3781next:
3782 leaf = path->nodes[0];
3783 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
3784 ret = btrfs_next_leaf(rc->extent_root, path);
3785 if (ret != 0)
3786 break;
3787 leaf = path->nodes[0];
3788 }
3789
3790 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
3791 if (key.objectid >= last) {
3792 ret = 1;
3793 break;
3794 }
3795
3173a18f
JB		 3796 if (key.type != BTRFS_EXTENT_ITEM_KEY &&
3797 key.type != BTRFS_METADATA_ITEM_KEY) {
3798 path->slots[0]++;
3799 goto next;
3800 }
3801
3802 if (key.type == BTRFS_EXTENT_ITEM_KEY &&
5d4f98a2
YZ		 3803 key.objectid + key.offset <= rc->search_start) {
3804 path->slots[0]++;
3805 goto next;
3806 }
3807
3173a18f 3808 if (key.type == BTRFS_METADATA_ITEM_KEY &&
707e8a07 3809 key.objectid + rc->extent_root->nodesize <=
3173a18f
JB		 3810 rc->search_start) {
3811 path->slots[0]++;
3812 goto next;
3813 }
3814
5d4f98a2
YZ		 3815 ret = find_first_extent_bit(&rc->processed_blocks,
3816 key.objectid, &start, &end,
e6138876 3817 EXTENT_DIRTY, NULL);
5d4f98a2
YZ		 3818
3819 if (ret == 0 && start <= key.objectid) {
b3b4aa74 3820 btrfs_release_path(path);
5d4f98a2
YZ		 3821 rc->search_start = end + 1;
3822 } else {
3173a18f
JB		 3823 if (key.type == BTRFS_EXTENT_ITEM_KEY)
3824 rc->search_start = key.objectid + key.offset;
3825 else
3826 rc->search_start = key.objectid +
707e8a07 3827 rc->extent_root->nodesize;
3fd0a558 3828 memcpy(extent_key, &key, sizeof(key));
5d4f98a2
YZ		 3829 return 0;
3830 }
3831 }
b3b4aa74 3832 btrfs_release_path(path);
5d4f98a2
YZ		 3833 return ret;
3834}
3835
3836static void set_reloc_control(struct reloc_control *rc)
3837{
3838 struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
7585717f
CM		 3839
3840 mutex_lock(&fs_info->reloc_mutex);
5d4f98a2 3841 fs_info->reloc_ctl = rc;
7585717f 3842 mutex_unlock(&fs_info->reloc_mutex);
5d4f98a2
YZ		 3843}
3844
3845static void unset_reloc_control(struct reloc_control *rc)
3846{
3847 struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
7585717f
CM		 3848
3849 mutex_lock(&fs_info->reloc_mutex);
5d4f98a2 3850 fs_info->reloc_ctl = NULL;
7585717f 3851 mutex_unlock(&fs_info->reloc_mutex);
5d4f98a2
YZ		 3852}
3853
3854static int check_extent_flags(u64 flags)
3855{
3856 if ((flags & BTRFS_EXTENT_FLAG_DATA) &&
3857 (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK))
3858 return 1;
3859 if (!(flags & BTRFS_EXTENT_FLAG_DATA) &&
3860 !(flags & BTRFS_EXTENT_FLAG_TREE_BLOCK))
3861 return 1;
3862 if ((flags & BTRFS_EXTENT_FLAG_DATA) &&
3863 (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
3864 return 1;
3865 return 0;
3866}
3867
3fd0a558
YZ		 3868static noinline_for_stack
3869int prepare_to_relocate(struct reloc_control *rc)
3870{
3871 struct btrfs_trans_handle *trans;
3fd0a558 3872
66d8f3dd
MX		 3873 rc->block_rsv = btrfs_alloc_block_rsv(rc->extent_root,
3874 BTRFS_BLOCK_RSV_TEMP);
3fd0a558
YZ		 3875 if (!rc->block_rsv)
3876 return -ENOMEM;
3877
3fd0a558
YZ		 3878 memset(&rc->cluster, 0, sizeof(rc->cluster));
3879 rc->search_start = rc->block_group->key.objectid;
3880 rc->extents_found = 0;
3881 rc->nodes_relocated = 0;
3882 rc->merging_rsv_size = 0;
0647bf56
WS		 3883 rc->reserved_bytes = 0;
3884 rc->block_rsv->size = rc->extent_root->nodesize *
3885 RELOCATION_RESERVED_NODES;
3fd0a558
YZ		 3886
3887 rc->create_reloc_tree = 1;
3888 set_reloc_control(rc);
3889
7a7eaa40 3890 trans = btrfs_join_transaction(rc->extent_root);
28818947
LB		 3891 if (IS_ERR(trans)) {
3892 unset_reloc_control(rc);
3893 /*
3894 * extent tree is not a ref_cow tree and has no reloc_root to
3895 * cleanup. And callers are responsible to free the above
3896 * block rsv.
3897 */
3898 return PTR_ERR(trans);
3899 }
3fd0a558
YZ		 3900 btrfs_commit_transaction(trans, rc->extent_root);
3901 return 0;
3902}
76dda93c 3903
5d4f98a2
YZ		 3904static noinline_for_stack int relocate_block_group(struct reloc_control *rc)
3905{
3906 struct rb_root blocks = RB_ROOT;
3907 struct btrfs_key key;
3908 struct btrfs_trans_handle *trans = NULL;
3909 struct btrfs_path *path;
3910 struct btrfs_extent_item *ei;
5d4f98a2
YZ		 3911 u64 flags;
3912 u32 item_size;
3913 int ret;
3914 int err = 0;
c87f08ca 3915 int progress = 0;
5d4f98a2
YZ		 3916
3917 path = btrfs_alloc_path();
3fd0a558 3918 if (!path)
5d4f98a2 3919 return -ENOMEM;
e4058b54 3920 path->reada = READA_FORWARD;
5d4f98a2 3921
3fd0a558
YZ		 3922 ret = prepare_to_relocate(rc);
3923 if (ret) {
3924 err = ret;
3925 goto out_free;
3926 }
5d4f98a2
YZ		 3927
3928 while (1) {
0647bf56
WS		 3929 rc->reserved_bytes = 0;
3930 ret = btrfs_block_rsv_refill(rc->extent_root,
3931 rc->block_rsv, rc->block_rsv->size,
3932 BTRFS_RESERVE_FLUSH_ALL);
3933 if (ret) {
3934 err = ret;
3935 break;
3936 }
c87f08ca 3937 progress++;
a22285a6 3938 trans = btrfs_start_transaction(rc->extent_root, 0);
0f788c58
LB		 3939 if (IS_ERR(trans)) {
3940 err = PTR_ERR(trans);
3941 trans = NULL;
3942 break;
3943 }
c87f08ca 3944restart:
3fd0a558
YZ		 3945 if (update_backref_cache(trans, &rc->backref_cache)) {
3946 btrfs_end_transaction(trans, rc->extent_root);
3947 continue;
3948 }
3949
147d256e 3950 ret = find_next_extent(rc, path, &key);
5d4f98a2
YZ		 3951 if (ret < 0)
3952 err = ret;
3953 if (ret != 0)
3954 break;
3955
3956 rc->extents_found++;
3957
3958 ei = btrfs_item_ptr(path->nodes[0], path->slots[0],
3959 struct btrfs_extent_item);
3fd0a558 3960 item_size = btrfs_item_size_nr(path->nodes[0], path->slots[0]);
5d4f98a2
YZ		 3961 if (item_size >= sizeof(*ei)) {
3962 flags = btrfs_extent_flags(path->nodes[0], ei);
3963 ret = check_extent_flags(flags);
3964 BUG_ON(ret);
3965
3966 } else {
3967#ifdef BTRFS_COMPAT_EXTENT_TREE_V0
3968 u64 ref_owner;
3969 int path_change = 0;
3970
3971 BUG_ON(item_size !=
3972 sizeof(struct btrfs_extent_item_v0));
3973 ret = get_ref_objectid_v0(rc, path, &key, &ref_owner,
3974 &path_change);
4b3576e4
Z		 3975 if (ret < 0) {
3976 err = ret;
3977 break;
3978 }
5d4f98a2
YZ		 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) {
b3b4aa74 3985 btrfs_release_path(path);
5d4f98a2
YZ		 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 &&
3fd0a558 4005 (flags & BTRFS_EXTENT_FLAG_DATA)) {
5d4f98a2
YZ		 4006 ret = add_data_references(rc, &key, path, &blocks);
4007 } else {
b3b4aa74 4008 btrfs_release_path(path);
5d4f98a2
YZ		 4009 ret = 0;
4010 }
4011 if (ret < 0) {
3fd0a558 4012 err = ret;
5d4f98a2
YZ		 4013 break;
4014 }
4015
4016 if (!RB_EMPTY_ROOT(&blocks)) {
4017 ret = relocate_tree_blocks(trans, rc, &blocks);
4018 if (ret < 0) {
1708cc57
WS		 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
3fd0a558
YZ		 4025 if (ret != -EAGAIN) {
4026 err = ret;
4027 break;
4028 }
4029 rc->extents_found--;
4030 rc->search_start = key.objectid;
4031 }
4032 }
4033
0647bf56
WS		 4034 btrfs_end_transaction_throttle(trans, rc->extent_root);
4035 btrfs_btree_balance_dirty(rc->extent_root);
5d4f98a2 4036 trans = NULL;
5d4f98a2
YZ		 4037
4038 if (rc->stage == MOVE_DATA_EXTENTS &&
4039 (flags & BTRFS_EXTENT_FLAG_DATA)) {
4040 rc->found_file_extent = 1;
0257bb82 4041 ret = relocate_data_extent(rc->data_inode,
3fd0a558 4042 &key, &rc->cluster);
5d4f98a2
YZ		 4043 if (ret < 0) {
4044 err = ret;
4045 break;
4046 }
4047 }
4048 }
c87f08ca
CM		 4049 if (trans && progress && err == -ENOSPC) {
4050 ret = btrfs_force_chunk_alloc(trans, rc->extent_root,
4051 rc->block_group->flags);
9689457b 4052 if (ret == 1) {
c87f08ca
CM		 4053 err = 0;
4054 progress = 0;
4055 goto restart;
4056 }
4057 }
3fd0a558 4058
b3b4aa74 4059 btrfs_release_path(path);
3fd0a558
YZ		 4060 clear_extent_bits(&rc->processed_blocks, 0, (u64)-1, EXTENT_DIRTY,
4061 GFP_NOFS);
5d4f98a2
YZ		 4062
4063 if (trans) {
3fd0a558 4064 btrfs_end_transaction_throttle(trans, rc->extent_root);
b53d3f5d 4065 btrfs_btree_balance_dirty(rc->extent_root);
5d4f98a2
YZ		 4066 }
4067
0257bb82 4068 if (!err) {
3fd0a558
YZ		 4069 ret = relocate_file_extent_cluster(rc->data_inode,
4070 &rc->cluster);
0257bb82
YZ		 4071 if (ret < 0)
4072 err = ret;
4073 }
4074
3fd0a558
YZ		 4075 rc->create_reloc_tree = 0;
4076 set_reloc_control(rc);
0257bb82 4077
3fd0a558
YZ		 4078 backref_cache_cleanup(&rc->backref_cache);
4079 btrfs_block_rsv_release(rc->extent_root, rc->block_rsv, (u64)-1);
5d4f98a2 4080
3fd0a558 4081 err = prepare_to_merge(rc, err);
5d4f98a2
YZ		 4082
4083 merge_reloc_roots(rc);
4084
3fd0a558 4085 rc->merge_reloc_tree = 0;
5d4f98a2 4086 unset_reloc_control(rc);
3fd0a558 4087 btrfs_block_rsv_release(rc->extent_root, rc->block_rsv, (u64)-1);
5d4f98a2
YZ		 4088
4089 /* get rid of pinned extents */
7a7eaa40 4090 trans = btrfs_join_transaction(rc->extent_root);
3612b495
TI		 4091 if (IS_ERR(trans))
4092 err = PTR_ERR(trans);
4093 else
4094 btrfs_commit_transaction(trans, rc->extent_root);
3fd0a558
YZ		 4095out_free:
4096 btrfs_free_block_rsv(rc->extent_root, rc->block_rsv);
4097 btrfs_free_path(path);
5d4f98a2
YZ		 4098 return err;
4099}
4100
4101static int __insert_orphan_inode(struct btrfs_trans_handle *trans,
0257bb82 4102 struct btrfs_root *root, u64 objectid)
5d4f98a2
YZ		 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);
0257bb82 4121 btrfs_set_inode_size(leaf, item, 0);
5d4f98a2 4122 btrfs_set_inode_mode(leaf, item, S_IFREG | 0600);
3fd0a558
YZ		 4123 btrfs_set_inode_flags(leaf, item, BTRFS_INODE_NOCOMPRESS |
4124 BTRFS_INODE_PREALLOC);
5d4f98a2 4125 btrfs_mark_buffer_dirty(leaf);
5d4f98a2
YZ		 4126out:
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 */
3fd0a558
YZ		 4135static noinline_for_stack
4136struct inode *create_reloc_inode(struct btrfs_fs_info *fs_info,
4137 struct btrfs_block_group_cache *group)
5d4f98a2
YZ		 4138{
4139 struct inode *inode = NULL;
4140 struct btrfs_trans_handle *trans;
4141 struct btrfs_root *root;
4142 struct btrfs_key key;
4624900d 4143 u64 objectid;
5d4f98a2
YZ		 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
a22285a6 4150 trans = btrfs_start_transaction(root, 6);
3fd0a558
YZ		 4151 if (IS_ERR(trans))
4152 return ERR_CAST(trans);
5d4f98a2 4153
581bb050 4154 err = btrfs_find_free_objectid(root, &objectid);
5d4f98a2
YZ		 4155 if (err)
4156 goto out;
4157
0257bb82 4158 err = __insert_orphan_inode(trans, root, objectid);
5d4f98a2
YZ		 4159 BUG_ON(err);
4160
4161 key.objectid = objectid;
4162 key.type = BTRFS_INODE_ITEM_KEY;
4163 key.offset = 0;
73f73415 4164 inode = btrfs_iget(root->fs_info->sb, &key, root, NULL);
5d4f98a2
YZ		 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);
4169out:
5d4f98a2 4170 btrfs_end_transaction(trans, root);
b53d3f5d 4171 btrfs_btree_balance_dirty(root);
5d4f98a2
YZ		 4172 if (err) {
4173 if (inode)
4174 iput(inode);
4175 inode = ERR_PTR(err);
4176 }
4177 return inode;
4178}
4179
a9995eec 4180static struct reloc_control *alloc_reloc_control(struct btrfs_fs_info *fs_info)
3fd0a558
YZ		 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);
a9995eec
JB		 4191 extent_io_tree_init(&rc->processed_blocks,
4192 fs_info->btree_inode->i_mapping);
3fd0a558
YZ		 4193 return rc;
4194}
4195
5d4f98a2
YZ		 4196/*
4197 * function to relocate all extents in a block group.
4198 */
4199int 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;
0af3d00b
JB		 4203 struct inode *inode;
4204 struct btrfs_path *path;
5d4f98a2 4205 int ret;
f0486c68 4206 int rw = 0;
5d4f98a2
YZ		 4207 int err = 0;
4208
a9995eec 4209 rc = alloc_reloc_control(fs_info);
5d4f98a2
YZ		 4210 if (!rc)
4211 return -ENOMEM;
4212
f0486c68 4213 rc->extent_root = extent_root;
3fd0a558 4214
5d4f98a2
YZ		 4215 rc->block_group = btrfs_lookup_block_group(fs_info, group_start);
4216 BUG_ON(!rc->block_group);
4217
868f401a
Z		 4218 ret = btrfs_inc_block_group_ro(extent_root, rc->block_group);
4219 if (ret) {
4220 err = ret;
4221 goto out;
f0486c68 4222 }
868f401a 4223 rw = 1;
f0486c68 4224
0af3d00b
JB		 4225 path = btrfs_alloc_path();
4226 if (!path) {
4227 err = -ENOMEM;
4228 goto out;
4229 }
4230
4231 inode = lookup_free_space_inode(fs_info->tree_root, rc->block_group,
4232 path);
4233 btrfs_free_path(path);
4234
4235 if (!IS_ERR(inode))
1bbc621e 4236 ret = delete_block_group_cache(fs_info, rc->block_group, inode, 0);
0af3d00b
JB		 4237 else
4238 ret = PTR_ERR(inode);
4239
4240 if (ret && ret != -ENOENT) {
4241 err = ret;
4242 goto out;
4243 }
4244
5d4f98a2
YZ		 4245 rc->data_inode = create_reloc_inode(fs_info, rc->block_group);
4246 if (IS_ERR(rc->data_inode)) {
4247 err = PTR_ERR(rc->data_inode);
4248 rc->data_inode = NULL;
4249 goto out;
4250 }
4251
efe120a0 4252 btrfs_info(extent_root->fs_info, "relocating block group %llu flags %llu",
c1c9ff7c 4253 rc->block_group->key.objectid, rc->block_group->flags);
5d4f98a2 4254
6c255e67 4255 ret = btrfs_start_delalloc_roots(fs_info, 0, -1);
8ccf6f19
MX		 4256 if (ret < 0) {
4257 err = ret;
4258 goto out;
4259 }
b0244199 4260 btrfs_wait_ordered_roots(fs_info, -1);
5d4f98a2
YZ		 4261
4262 while (1) {
76dda93c 4263 mutex_lock(&fs_info->cleaner_mutex);
5d4f98a2 4264 ret = relocate_block_group(rc);
76dda93c 4265 mutex_unlock(&fs_info->cleaner_mutex);
5d4f98a2
YZ		 4266 if (ret < 0) {
4267 err = ret;
3fd0a558 4268 goto out;
5d4f98a2
YZ		 4269 }
4270
4271 if (rc->extents_found == 0)
4272 break;
4273
efe120a0 4274 btrfs_info(extent_root->fs_info, "found %llu extents",
c1c9ff7c 4275 rc->extents_found);
5d4f98a2
YZ		 4276
4277 if (rc->stage == MOVE_DATA_EXTENTS && rc->found_file_extent) {
0ef8b726
JB		 4278 ret = btrfs_wait_ordered_range(rc->data_inode, 0,
4279 (u64)-1);
4280 if (ret) {
4281 err = ret;
4282 goto out;
4283 }
5d4f98a2
YZ		 4284 invalidate_mapping_pages(rc->data_inode->i_mapping,
4285 0, -1);
4286 rc->stage = UPDATE_DATA_PTRS;
5d4f98a2
YZ		 4287 }
4288 }
4289
5d4f98a2
YZ		 4290 WARN_ON(rc->block_group->pinned > 0);
4291 WARN_ON(rc->block_group->reserved > 0);
4292 WARN_ON(btrfs_block_group_used(&rc->block_group->item) > 0);
4293out:
f0486c68 4294 if (err && rw)
868f401a 4295 btrfs_dec_block_group_ro(extent_root, rc->block_group);
5d4f98a2 4296 iput(rc->data_inode);
5d4f98a2
YZ		 4297 btrfs_put_block_group(rc->block_group);
4298 kfree(rc);
4299 return err;
4300}
4301
76dda93c
YZ		 4302static noinline_for_stack int mark_garbage_root(struct btrfs_root *root)
4303{
4304 struct btrfs_trans_handle *trans;
79787eaa 4305 int ret, err;
76dda93c 4306
a22285a6 4307 trans = btrfs_start_transaction(root->fs_info->tree_root, 0);
79787eaa
JM		 4308 if (IS_ERR(trans))
4309 return PTR_ERR(trans);
76dda93c
YZ		 4310
4311 memset(&root->root_item.drop_progress, 0,
4312 sizeof(root->root_item.drop_progress));
4313 root->root_item.drop_level = 0;
4314 btrfs_set_root_refs(&root->root_item, 0);
4315 ret = btrfs_update_root(trans, root->fs_info->tree_root,
4316 &root->root_key, &root->root_item);
76dda93c 4317
79787eaa
JM		 4318 err = btrfs_end_transaction(trans, root->fs_info->tree_root);
4319 if (err)
4320 return err;
4321 return ret;
76dda93c
YZ		 4322}
4323
5d4f98a2
YZ		 4324/*
4325 * recover relocation interrupted by system crash.
4326 *
4327 * this function resumes merging reloc trees with corresponding fs trees.
4328 * this is important for keeping the sharing of tree blocks
4329 */
4330int btrfs_recover_relocation(struct btrfs_root *root)
4331{
4332 LIST_HEAD(reloc_roots);
4333 struct btrfs_key key;
4334 struct btrfs_root *fs_root;
4335 struct btrfs_root *reloc_root;
4336 struct btrfs_path *path;
4337 struct extent_buffer *leaf;
4338 struct reloc_control *rc = NULL;
4339 struct btrfs_trans_handle *trans;
4340 int ret;
4341 int err = 0;
4342
4343 path = btrfs_alloc_path();
4344 if (!path)
4345 return -ENOMEM;
e4058b54 4346 path->reada = READA_BACK;
5d4f98a2
YZ		 4347
4348 key.objectid = BTRFS_TREE_RELOC_OBJECTID;
4349 key.type = BTRFS_ROOT_ITEM_KEY;
4350 key.offset = (u64)-1;
4351
4352 while (1) {
4353 ret = btrfs_search_slot(NULL, root->fs_info->tree_root, &key,
4354 path, 0, 0);
4355 if (ret < 0) {
4356 err = ret;
4357 goto out;
4358 }
4359 if (ret > 0) {
4360 if (path->slots[0] == 0)
4361 break;
4362 path->slots[0]--;
4363 }
4364 leaf = path->nodes[0];
4365 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
b3b4aa74 4366 btrfs_release_path(path);
5d4f98a2
YZ		 4367
4368 if (key.objectid != BTRFS_TREE_RELOC_OBJECTID ||
4369 key.type != BTRFS_ROOT_ITEM_KEY)
4370 break;
4371
cb517eab 4372 reloc_root = btrfs_read_fs_root(root, &key);
5d4f98a2
YZ		 4373 if (IS_ERR(reloc_root)) {
4374 err = PTR_ERR(reloc_root);
4375 goto out;
4376 }
4377
4378 list_add(&reloc_root->root_list, &reloc_roots);
4379
4380 if (btrfs_root_refs(&reloc_root->root_item) > 0) {
4381 fs_root = read_fs_root(root->fs_info,
4382 reloc_root->root_key.offset);
4383 if (IS_ERR(fs_root)) {
76dda93c
YZ		 4384 ret = PTR_ERR(fs_root);
4385 if (ret != -ENOENT) {
4386 err = ret;
4387 goto out;
4388 }
79787eaa
JM		 4389 ret = mark_garbage_root(reloc_root);
4390 if (ret < 0) {
4391 err = ret;
4392 goto out;
4393 }
5d4f98a2
YZ		 4394 }
4395 }
4396
4397 if (key.offset == 0)
4398 break;
4399
4400 key.offset--;
4401 }
b3b4aa74 4402 btrfs_release_path(path);
5d4f98a2
YZ		 4403
4404 if (list_empty(&reloc_roots))
4405 goto out;
4406
a9995eec 4407 rc = alloc_reloc_control(root->fs_info);
5d4f98a2
YZ		 4408 if (!rc) {
4409 err = -ENOMEM;
4410 goto out;
4411 }
4412
5d4f98a2
YZ		 4413 rc->extent_root = root->fs_info->extent_root;
4414
4415 set_reloc_control(rc);
4416
7a7eaa40 4417 trans = btrfs_join_transaction(rc->extent_root);
3612b495
TI		 4418 if (IS_ERR(trans)) {
4419 unset_reloc_control(rc);
4420 err = PTR_ERR(trans);
4421 goto out_free;
4422 }
3fd0a558
YZ		 4423
4424 rc->merge_reloc_tree = 1;
4425
5d4f98a2
YZ		 4426 while (!list_empty(&reloc_roots)) {
4427 reloc_root = list_entry(reloc_roots.next,
4428 struct btrfs_root, root_list);
4429 list_del(&reloc_root->root_list);
4430
4431 if (btrfs_root_refs(&reloc_root->root_item) == 0) {
4432 list_add_tail(&reloc_root->root_list,
4433 &rc->reloc_roots);
4434 continue;
4435 }
4436
4437 fs_root = read_fs_root(root->fs_info,
4438 reloc_root->root_key.offset);
79787eaa
JM		 4439 if (IS_ERR(fs_root)) {
4440 err = PTR_ERR(fs_root);
4441 goto out_free;
4442 }
5d4f98a2 4443
ffd7b339 4444 err = __add_reloc_root(reloc_root);
79787eaa 4445 BUG_ON(err < 0); /* -ENOMEM or logic error */
5d4f98a2
YZ		 4446 fs_root->reloc_root = reloc_root;
4447 }
4448
79787eaa
JM		 4449 err = btrfs_commit_transaction(trans, rc->extent_root);
4450 if (err)
4451 goto out_free;
5d4f98a2
YZ		 4452
4453 merge_reloc_roots(rc);
4454
4455 unset_reloc_control(rc);
4456
7a7eaa40 4457 trans = btrfs_join_transaction(rc->extent_root);
3612b495
TI		 4458 if (IS_ERR(trans))
4459 err = PTR_ERR(trans);
4460 else
79787eaa 4461 err = btrfs_commit_transaction(trans, rc->extent_root);
3612b495 4462out_free:
3fd0a558 4463 kfree(rc);
3612b495 4464out:
aca1bba6
LB		 4465 if (!list_empty(&reloc_roots))
4466 free_reloc_roots(&reloc_roots);
4467
5d4f98a2
YZ		 4468 btrfs_free_path(path);
4469
4470 if (err == 0) {
4471 /* cleanup orphan inode in data relocation tree */
4472 fs_root = read_fs_root(root->fs_info,
4473 BTRFS_DATA_RELOC_TREE_OBJECTID);
4474 if (IS_ERR(fs_root))
4475 err = PTR_ERR(fs_root);
d7ce5843 4476 else
66b4ffd1 4477 err = btrfs_orphan_cleanup(fs_root);
5d4f98a2
YZ		 4478 }
4479 return err;
4480}
4481
4482/*
4483 * helper to add ordered checksum for data relocation.
4484 *
4485 * cloning checksum properly handles the nodatasum extents.
4486 * it also saves CPU time to re-calculate the checksum.
4487 */
4488int btrfs_reloc_clone_csums(struct inode *inode, u64 file_pos, u64 len)
4489{
4490 struct btrfs_ordered_sum *sums;
5d4f98a2
YZ		 4491 struct btrfs_ordered_extent *ordered;
4492 struct btrfs_root *root = BTRFS_I(inode)->root;
5d4f98a2
YZ		 4493 int ret;
4494 u64 disk_bytenr;
4577b014 4495 u64 new_bytenr;
5d4f98a2
YZ		 4496 LIST_HEAD(list);
4497
4498 ordered = btrfs_lookup_ordered_extent(inode, file_pos);
4499 BUG_ON(ordered->file_offset != file_pos || ordered->len != len);
4500
4501 disk_bytenr = file_pos + BTRFS_I(inode)->index_cnt;
4502 ret = btrfs_lookup_csums_range(root->fs_info->csum_root, disk_bytenr,
a2de733c 4503 disk_bytenr + len - 1, &list, 0);
79787eaa
JM		 4504 if (ret)
4505 goto out;
5d4f98a2
YZ		 4506
4507 while (!list_empty(&list)) {
4508 sums = list_entry(list.next, struct btrfs_ordered_sum, list);
4509 list_del_init(&sums->list);
4510
4577b014
JB		 4511 /*
4512 * We need to offset the new_bytenr based on where the csum is.
4513 * We need to do this because we will read in entire prealloc
4514 * extents but we may have written to say the middle of the
4515 * prealloc extent, so we need to make sure the csum goes with
4516 * the right disk offset.
4517 *
4518 * We can do this because the data reloc inode refers strictly
4519 * to the on disk bytes, so we don't have to worry about
4520 * disk_len vs real len like with real inodes since it's all
4521 * disk length.
4522 */
4523 new_bytenr = ordered->start + (sums->bytenr - disk_bytenr);
4524 sums->bytenr = new_bytenr;
5d4f98a2
YZ		 4525
4526 btrfs_add_ordered_sum(inode, ordered, sums);
4527 }
79787eaa 4528out:
5d4f98a2 4529 btrfs_put_ordered_extent(ordered);
411fc6bc 4530 return ret;
5d4f98a2 4531}
3fd0a558 4532
83d4cfd4
JB		 4533int btrfs_reloc_cow_block(struct btrfs_trans_handle *trans,
4534 struct btrfs_root *root, struct extent_buffer *buf,
4535 struct extent_buffer *cow)
3fd0a558
YZ		 4536{
4537 struct reloc_control *rc;
4538 struct backref_node *node;
4539 int first_cow = 0;
4540 int level;
83d4cfd4 4541 int ret = 0;
3fd0a558
YZ		 4542
4543 rc = root->fs_info->reloc_ctl;
4544 if (!rc)
83d4cfd4 4545 return 0;
3fd0a558
YZ		 4546
4547 BUG_ON(rc->stage == UPDATE_DATA_PTRS &&
4548 root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID);
4549
c974c464
WS		 4550 if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID) {
4551 if (buf == root->node)
4552 __update_reloc_root(root, cow->start);
4553 }
4554
3fd0a558
YZ		 4555 level = btrfs_header_level(buf);
4556 if (btrfs_header_generation(buf) <=
4557 btrfs_root_last_snapshot(&root->root_item))
4558 first_cow = 1;
4559
4560 if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID &&
4561 rc->create_reloc_tree) {
4562 WARN_ON(!first_cow && level == 0);
4563
4564 node = rc->backref_cache.path[level];
4565 BUG_ON(node->bytenr != buf->start &&
4566 node->new_bytenr != buf->start);
4567
4568 drop_node_buffer(node);
4569 extent_buffer_get(cow);
4570 node->eb = cow;
4571 node->new_bytenr = cow->start;
4572
4573 if (!node->pending) {
4574 list_move_tail(&node->list,
4575 &rc->backref_cache.pending[level]);
4576 node->pending = 1;
4577 }
4578
4579 if (first_cow)
4580 __mark_block_processed(rc, node);
4581
4582 if (first_cow && level > 0)
4583 rc->nodes_relocated += buf->len;
4584 }
4585
83d4cfd4 4586 if (level == 0 && first_cow && rc->stage == UPDATE_DATA_PTRS)
3fd0a558 4587 ret = replace_file_extents(trans, rc, root, cow);
83d4cfd4 4588 return ret;
3fd0a558
YZ		 4589}
4590
4591/*
4592 * called before creating snapshot. it calculates metadata reservation
4593 * requried for relocating tree blocks in the snapshot
4594 */
147d256e 4595void btrfs_reloc_pre_snapshot(struct btrfs_pending_snapshot *pending,
3fd0a558
YZ		 4596 u64 *bytes_to_reserve)
4597{
4598 struct btrfs_root *root;
4599 struct reloc_control *rc;
4600
4601 root = pending->root;
4602 if (!root->reloc_root)
4603 return;
4604
4605 rc = root->fs_info->reloc_ctl;
4606 if (!rc->merge_reloc_tree)
4607 return;
4608
4609 root = root->reloc_root;
4610 BUG_ON(btrfs_root_refs(&root->root_item) == 0);
4611 /*
4612 * relocation is in the stage of merging trees. the space
4613 * used by merging a reloc tree is twice the size of
4614 * relocated tree nodes in the worst case. half for cowing
4615 * the reloc tree, half for cowing the fs tree. the space
4616 * used by cowing the reloc tree will be freed after the
4617 * tree is dropped. if we create snapshot, cowing the fs
4618 * tree may use more space than it frees. so we need
4619 * reserve extra space.
4620 */
4621 *bytes_to_reserve += rc->nodes_relocated;
4622}
4623
4624/*
4625 * called after snapshot is created. migrate block reservation
4626 * and create reloc root for the newly created snapshot
4627 */
49b25e05 4628int btrfs_reloc_post_snapshot(struct btrfs_trans_handle *trans,
3fd0a558
YZ		 4629 struct btrfs_pending_snapshot *pending)
4630{
4631 struct btrfs_root *root = pending->root;
4632 struct btrfs_root *reloc_root;
4633 struct btrfs_root *new_root;
4634 struct reloc_control *rc;
4635 int ret;
4636
4637 if (!root->reloc_root)
49b25e05 4638 return 0;
3fd0a558
YZ		 4639
4640 rc = root->fs_info->reloc_ctl;
4641 rc->merging_rsv_size += rc->nodes_relocated;
4642
4643 if (rc->merge_reloc_tree) {
4644 ret = btrfs_block_rsv_migrate(&pending->block_rsv,
4645 rc->block_rsv,
4646 rc->nodes_relocated);
49b25e05
JM		 4647 if (ret)
4648 return ret;
3fd0a558
YZ		 4649 }
4650
4651 new_root = pending->snap;
4652 reloc_root = create_reloc_root(trans, root->reloc_root,
4653 new_root->root_key.objectid);
49b25e05
JM		 4654 if (IS_ERR(reloc_root))
4655 return PTR_ERR(reloc_root);
3fd0a558 4656
ffd7b339
JM		 4657 ret = __add_reloc_root(reloc_root);
4658 BUG_ON(ret < 0);
3fd0a558
YZ		 4659 new_root->reloc_root = reloc_root;
4660
49b25e05 4661 if (rc->create_reloc_tree)
3fd0a558 4662 ret = clone_backref_node(trans, rc, root, reloc_root);
49b25e05 4663 return ret;
3fd0a558 4664}
Linux 6.x block layer and io_uring tree(s)