projects / linux-2.6-block.git / blame
? search:
summary | shortlog | log | commit | commitdiff | tree
blob | blame (incremental) | history | HEAD
Btrfs: convert printk to btrfs_ and fix BTRFS prefix
[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 "
c1c9ff7c 340 "found at offset %llu\n", 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
531 if (!root->ref_cows)
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
613 if (root->ref_cows &&
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 }
026fd317
JB		 711 path1->reada = 1;
712 path2->reada = 2;
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 }
739 BUG_ON(!ret || !path1->slots[0]);
740
741 path1->slots[0]--;
742
743 WARN_ON(cur->checked);
744 if (!list_empty(&cur->upper)) {
745 /*
70f23fd6 746 * the backref was added previously when processing
5d4f98a2
YZ		 747 * backref of type BTRFS_TREE_BLOCK_REF_KEY
748 */
749 BUG_ON(!list_is_singular(&cur->upper));
750 edge = list_entry(cur->upper.next, struct backref_edge,
751 list[LOWER]);
752 BUG_ON(!list_empty(&edge->list[UPPER]));
753 exist = edge->node[UPPER];
754 /*
755 * add the upper level block to pending list if we need
756 * check its backrefs
757 */
758 if (!exist->checked)
759 list_add_tail(&edge->list[UPPER], &list);
760 } else {
761 exist = NULL;
762 }
763
764 while (1) {
765 cond_resched();
766 eb = path1->nodes[0];
767
768 if (ptr >= end) {
769 if (path1->slots[0] >= btrfs_header_nritems(eb)) {
770 ret = btrfs_next_leaf(rc->extent_root, path1);
771 if (ret < 0) {
772 err = ret;
773 goto out;
774 }
775 if (ret > 0)
776 break;
777 eb = path1->nodes[0];
778 }
779
780 btrfs_item_key_to_cpu(eb, &key, path1->slots[0]);
781 if (key.objectid != cur->bytenr) {
782 WARN_ON(exist);
783 break;
784 }
785
3173a18f
JB		 786 if (key.type == BTRFS_EXTENT_ITEM_KEY ||
787 key.type == BTRFS_METADATA_ITEM_KEY) {
5d4f98a2
YZ		 788 ret = find_inline_backref(eb, path1->slots[0],
789 &ptr, &end);
790 if (ret)
791 goto next;
792 }
793 }
794
795 if (ptr < end) {
796 /* update key for inline back ref */
797 struct btrfs_extent_inline_ref *iref;
798 iref = (struct btrfs_extent_inline_ref *)ptr;
799 key.type = btrfs_extent_inline_ref_type(eb, iref);
800 key.offset = btrfs_extent_inline_ref_offset(eb, iref);
801 WARN_ON(key.type != BTRFS_TREE_BLOCK_REF_KEY &&
802 key.type != BTRFS_SHARED_BLOCK_REF_KEY);
803 }
804
805 if (exist &&
806 ((key.type == BTRFS_TREE_BLOCK_REF_KEY &&
807 exist->owner == key.offset) ||
808 (key.type == BTRFS_SHARED_BLOCK_REF_KEY &&
809 exist->bytenr == key.offset))) {
810 exist = NULL;
811 goto next;
812 }
813
814#ifdef BTRFS_COMPAT_EXTENT_TREE_V0
815 if (key.type == BTRFS_SHARED_BLOCK_REF_KEY ||
816 key.type == BTRFS_EXTENT_REF_V0_KEY) {
3fd0a558 817 if (key.type == BTRFS_EXTENT_REF_V0_KEY) {
5d4f98a2
YZ		 818 struct btrfs_extent_ref_v0 *ref0;
819 ref0 = btrfs_item_ptr(eb, path1->slots[0],
820 struct btrfs_extent_ref_v0);
3fd0a558 821 if (key.objectid == key.offset) {
046f264f 822 root = find_tree_root(rc, eb, ref0);
3fd0a558
YZ		 823 if (root && !should_ignore_root(root))
824 cur->root = root;
825 else
826 list_add(&cur->list, &useless);
827 break;
828 }
046f264f
YZ		 829 if (is_cowonly_root(btrfs_ref_root_v0(eb,
830 ref0)))
831 cur->cowonly = 1;
5d4f98a2
YZ		 832 }
833#else
834 BUG_ON(key.type == BTRFS_EXTENT_REF_V0_KEY);
835 if (key.type == BTRFS_SHARED_BLOCK_REF_KEY) {
836#endif
837 if (key.objectid == key.offset) {
838 /*
839 * only root blocks of reloc trees use
840 * backref of this type.
841 */
842 root = find_reloc_root(rc, cur->bytenr);
843 BUG_ON(!root);
844 cur->root = root;
845 break;
846 }
847
3fd0a558 848 edge = alloc_backref_edge(cache);
5d4f98a2
YZ		 849 if (!edge) {
850 err = -ENOMEM;
851 goto out;
852 }
853 rb_node = tree_search(&cache->rb_root, key.offset);
854 if (!rb_node) {
3fd0a558 855 upper = alloc_backref_node(cache);
5d4f98a2 856 if (!upper) {
3fd0a558 857 free_backref_edge(cache, edge);
5d4f98a2
YZ		 858 err = -ENOMEM;
859 goto out;
860 }
5d4f98a2 861 upper->bytenr = key.offset;
5d4f98a2
YZ		 862 upper->level = cur->level + 1;
863 /*
864 * backrefs for the upper level block isn't
865 * cached, add the block to pending list
866 */
867 list_add_tail(&edge->list[UPPER], &list);
868 } else {
869 upper = rb_entry(rb_node, struct backref_node,
870 rb_node);
3fd0a558 871 BUG_ON(!upper->checked);
5d4f98a2
YZ		 872 INIT_LIST_HEAD(&edge->list[UPPER]);
873 }
3fd0a558 874 list_add_tail(&edge->list[LOWER], &cur->upper);
5d4f98a2 875 edge->node[LOWER] = cur;
3fd0a558 876 edge->node[UPPER] = upper;
5d4f98a2
YZ		 877
878 goto next;
879 } else if (key.type != BTRFS_TREE_BLOCK_REF_KEY) {
880 goto next;
881 }
882
883 /* key.type == BTRFS_TREE_BLOCK_REF_KEY */
884 root = read_fs_root(rc->extent_root->fs_info, key.offset);
885 if (IS_ERR(root)) {
886 err = PTR_ERR(root);
887 goto out;
888 }
889
3fd0a558
YZ		 890 if (!root->ref_cows)
891 cur->cowonly = 1;
892
5d4f98a2
YZ		 893 if (btrfs_root_level(&root->root_item) == cur->level) {
894 /* tree root */
895 BUG_ON(btrfs_root_bytenr(&root->root_item) !=
896 cur->bytenr);
3fd0a558
YZ		 897 if (should_ignore_root(root))
898 list_add(&cur->list, &useless);
899 else
900 cur->root = root;
5d4f98a2
YZ		 901 break;
902 }
903
904 level = cur->level + 1;
905
906 /*
907 * searching the tree to find upper level blocks
908 * reference the block.
909 */
910 path2->search_commit_root = 1;
911 path2->skip_locking = 1;
912 path2->lowest_level = level;
913 ret = btrfs_search_slot(NULL, root, node_key, path2, 0, 0);
914 path2->lowest_level = 0;
915 if (ret < 0) {
916 err = ret;
917 goto out;
918 }
33c66f43
YZ		 919 if (ret > 0 && path2->slots[level] > 0)
920 path2->slots[level]--;
5d4f98a2
YZ		 921
922 eb = path2->nodes[level];
923 WARN_ON(btrfs_node_blockptr(eb, path2->slots[level]) !=
924 cur->bytenr);
925
926 lower = cur;
b6c60c80 927 need_check = true;
5d4f98a2
YZ		 928 for (; level < BTRFS_MAX_LEVEL; level++) {
929 if (!path2->nodes[level]) {
930 BUG_ON(btrfs_root_bytenr(&root->root_item) !=
931 lower->bytenr);
3fd0a558
YZ		 932 if (should_ignore_root(root))
933 list_add(&lower->list, &useless);
934 else
935 lower->root = root;
5d4f98a2
YZ		 936 break;
937 }
938
3fd0a558 939 edge = alloc_backref_edge(cache);
5d4f98a2
YZ		 940 if (!edge) {
941 err = -ENOMEM;
942 goto out;
943 }
944
945 eb = path2->nodes[level];
946 rb_node = tree_search(&cache->rb_root, eb->start);
947 if (!rb_node) {
3fd0a558 948 upper = alloc_backref_node(cache);
5d4f98a2 949 if (!upper) {
3fd0a558 950 free_backref_edge(cache, edge);
5d4f98a2
YZ		 951 err = -ENOMEM;
952 goto out;
953 }
5d4f98a2
YZ		 954 upper->bytenr = eb->start;
955 upper->owner = btrfs_header_owner(eb);
956 upper->level = lower->level + 1;
3fd0a558
YZ		 957 if (!root->ref_cows)
958 upper->cowonly = 1;
5d4f98a2
YZ		 959
960 /*
961 * if we know the block isn't shared
962 * we can void checking its backrefs.
963 */
964 if (btrfs_block_can_be_shared(root, eb))
965 upper->checked = 0;
966 else
967 upper->checked = 1;
968
969 /*
970 * add the block to pending list if we
b6c60c80
JB		 971 * need check its backrefs, we only do this once
972 * while walking up a tree as we will catch
973 * anything else later on.
5d4f98a2 974 */
b6c60c80
JB		 975 if (!upper->checked && need_check) {
976 need_check = false;
5d4f98a2
YZ		 977 list_add_tail(&edge->list[UPPER],
978 &list);
979 } else
980 INIT_LIST_HEAD(&edge->list[UPPER]);
981 } else {
982 upper = rb_entry(rb_node, struct backref_node,
983 rb_node);
984 BUG_ON(!upper->checked);
985 INIT_LIST_HEAD(&edge->list[UPPER]);
3fd0a558
YZ		 986 if (!upper->owner)
987 upper->owner = btrfs_header_owner(eb);
5d4f98a2
YZ		 988 }
989 list_add_tail(&edge->list[LOWER], &lower->upper);
5d4f98a2 990 edge->node[LOWER] = lower;
3fd0a558 991 edge->node[UPPER] = upper;
5d4f98a2
YZ		 992
993 if (rb_node)
994 break;
995 lower = upper;
996 upper = NULL;
997 }
b3b4aa74 998 btrfs_release_path(path2);
5d4f98a2
YZ		 999next:
1000 if (ptr < end) {
1001 ptr += btrfs_extent_inline_ref_size(key.type);
1002 if (ptr >= end) {
1003 WARN_ON(ptr > end);
1004 ptr = 0;
1005 end = 0;
1006 }
1007 }
1008 if (ptr >= end)
1009 path1->slots[0]++;
1010 }
b3b4aa74 1011 btrfs_release_path(path1);
5d4f98a2
YZ		 1012
1013 cur->checked = 1;
1014 WARN_ON(exist);
1015
1016 /* the pending list isn't empty, take the first block to process */
1017 if (!list_empty(&list)) {
1018 edge = list_entry(list.next, struct backref_edge, list[UPPER]);
1019 list_del_init(&edge->list[UPPER]);
1020 cur = edge->node[UPPER];
1021 goto again;
1022 }
1023
1024 /*
1025 * everything goes well, connect backref nodes and insert backref nodes
1026 * into the cache.
1027 */
1028 BUG_ON(!node->checked);
3fd0a558
YZ		 1029 cowonly = node->cowonly;
1030 if (!cowonly) {
1031 rb_node = tree_insert(&cache->rb_root, node->bytenr,
1032 &node->rb_node);
43c04fb1
JM		 1033 if (rb_node)
1034 backref_tree_panic(rb_node, -EEXIST, node->bytenr);
3fd0a558
YZ		 1035 list_add_tail(&node->lower, &cache->leaves);
1036 }
5d4f98a2
YZ		 1037
1038 list_for_each_entry(edge, &node->upper, list[LOWER])
1039 list_add_tail(&edge->list[UPPER], &list);
1040
1041 while (!list_empty(&list)) {
1042 edge = list_entry(list.next, struct backref_edge, list[UPPER]);
1043 list_del_init(&edge->list[UPPER]);
1044 upper = edge->node[UPPER];
3fd0a558
YZ		 1045 if (upper->detached) {
1046 list_del(&edge->list[LOWER]);
1047 lower = edge->node[LOWER];
1048 free_backref_edge(cache, edge);
1049 if (list_empty(&lower->upper))
1050 list_add(&lower->list, &useless);
1051 continue;
1052 }
5d4f98a2
YZ		 1053
1054 if (!RB_EMPTY_NODE(&upper->rb_node)) {
1055 if (upper->lowest) {
1056 list_del_init(&upper->lower);
1057 upper->lowest = 0;
1058 }
1059
1060 list_add_tail(&edge->list[UPPER], &upper->lower);
1061 continue;
1062 }
1063
1064 BUG_ON(!upper->checked);
3fd0a558
YZ		 1065 BUG_ON(cowonly != upper->cowonly);
1066 if (!cowonly) {
1067 rb_node = tree_insert(&cache->rb_root, upper->bytenr,
1068 &upper->rb_node);
43c04fb1
JM		 1069 if (rb_node)
1070 backref_tree_panic(rb_node, -EEXIST,
1071 upper->bytenr);
3fd0a558 1072 }
5d4f98a2
YZ		 1073
1074 list_add_tail(&edge->list[UPPER], &upper->lower);
1075
1076 list_for_each_entry(edge, &upper->upper, list[LOWER])
1077 list_add_tail(&edge->list[UPPER], &list);
1078 }
3fd0a558
YZ		 1079 /*
1080 * process useless backref nodes. backref nodes for tree leaves
1081 * are deleted from the cache. backref nodes for upper level
1082 * tree blocks are left in the cache to avoid unnecessary backref
1083 * lookup.
1084 */
1085 while (!list_empty(&useless)) {
1086 upper = list_entry(useless.next, struct backref_node, list);
1087 list_del_init(&upper->list);
1088 BUG_ON(!list_empty(&upper->upper));
1089 if (upper == node)
1090 node = NULL;
1091 if (upper->lowest) {
1092 list_del_init(&upper->lower);
1093 upper->lowest = 0;
1094 }
1095 while (!list_empty(&upper->lower)) {
1096 edge = list_entry(upper->lower.next,
1097 struct backref_edge, list[UPPER]);
1098 list_del(&edge->list[UPPER]);
1099 list_del(&edge->list[LOWER]);
1100 lower = edge->node[LOWER];
1101 free_backref_edge(cache, edge);
1102
1103 if (list_empty(&lower->upper))
1104 list_add(&lower->list, &useless);
1105 }
1106 __mark_block_processed(rc, upper);
1107 if (upper->level > 0) {
1108 list_add(&upper->list, &cache->detached);
1109 upper->detached = 1;
1110 } else {
1111 rb_erase(&upper->rb_node, &cache->rb_root);
1112 free_backref_node(cache, upper);
1113 }
1114 }
5d4f98a2
YZ		 1115out:
1116 btrfs_free_path(path1);
1117 btrfs_free_path(path2);
1118 if (err) {
3fd0a558
YZ		 1119 while (!list_empty(&useless)) {
1120 lower = list_entry(useless.next,
1121 struct backref_node, upper);
1122 list_del_init(&lower->upper);
1123 }
5d4f98a2 1124 upper = node;
3fd0a558 1125 INIT_LIST_HEAD(&list);
5d4f98a2
YZ		 1126 while (upper) {
1127 if (RB_EMPTY_NODE(&upper->rb_node)) {
1128 list_splice_tail(&upper->upper, &list);
3fd0a558 1129 free_backref_node(cache, upper);
5d4f98a2
YZ		 1130 }
1131
1132 if (list_empty(&list))
1133 break;
1134
1135 edge = list_entry(list.next, struct backref_edge,
1136 list[LOWER]);
3fd0a558 1137 list_del(&edge->list[LOWER]);
5d4f98a2 1138 upper = edge->node[UPPER];
3fd0a558 1139 free_backref_edge(cache, edge);
5d4f98a2
YZ		 1140 }
1141 return ERR_PTR(err);
1142 }
3fd0a558 1143 BUG_ON(node && node->detached);
5d4f98a2
YZ		 1144 return node;
1145}
1146
3fd0a558
YZ		 1147/*
1148 * helper to add backref node for the newly created snapshot.
1149 * the backref node is created by cloning backref node that
1150 * corresponds to root of source tree
1151 */
1152static int clone_backref_node(struct btrfs_trans_handle *trans,
1153 struct reloc_control *rc,
1154 struct btrfs_root *src,
1155 struct btrfs_root *dest)
1156{
1157 struct btrfs_root *reloc_root = src->reloc_root;
1158 struct backref_cache *cache = &rc->backref_cache;
1159 struct backref_node *node = NULL;
1160 struct backref_node *new_node;
1161 struct backref_edge *edge;
1162 struct backref_edge *new_edge;
1163 struct rb_node *rb_node;
1164
1165 if (cache->last_trans > 0)
1166 update_backref_cache(trans, cache);
1167
1168 rb_node = tree_search(&cache->rb_root, src->commit_root->start);
1169 if (rb_node) {
1170 node = rb_entry(rb_node, struct backref_node, rb_node);
1171 if (node->detached)
1172 node = NULL;
1173 else
1174 BUG_ON(node->new_bytenr != reloc_root->node->start);
1175 }
1176
1177 if (!node) {
1178 rb_node = tree_search(&cache->rb_root,
1179 reloc_root->commit_root->start);
1180 if (rb_node) {
1181 node = rb_entry(rb_node, struct backref_node,
1182 rb_node);
1183 BUG_ON(node->detached);
1184 }
1185 }
1186
1187 if (!node)
1188 return 0;
1189
1190 new_node = alloc_backref_node(cache);
1191 if (!new_node)
1192 return -ENOMEM;
1193
1194 new_node->bytenr = dest->node->start;
1195 new_node->level = node->level;
1196 new_node->lowest = node->lowest;
6848ad64 1197 new_node->checked = 1;
3fd0a558
YZ		 1198 new_node->root = dest;
1199
1200 if (!node->lowest) {
1201 list_for_each_entry(edge, &node->lower, list[UPPER]) {
1202 new_edge = alloc_backref_edge(cache);
1203 if (!new_edge)
1204 goto fail;
1205
1206 new_edge->node[UPPER] = new_node;
1207 new_edge->node[LOWER] = edge->node[LOWER];
1208 list_add_tail(&new_edge->list[UPPER],
1209 &new_node->lower);
1210 }
76b9e23d
MX		 1211 } else {
1212 list_add_tail(&new_node->lower, &cache->leaves);
3fd0a558
YZ		 1213 }
1214
1215 rb_node = tree_insert(&cache->rb_root, new_node->bytenr,
1216 &new_node->rb_node);
43c04fb1
JM		 1217 if (rb_node)
1218 backref_tree_panic(rb_node, -EEXIST, new_node->bytenr);
3fd0a558
YZ		 1219
1220 if (!new_node->lowest) {
1221 list_for_each_entry(new_edge, &new_node->lower, list[UPPER]) {
1222 list_add_tail(&new_edge->list[LOWER],
1223 &new_edge->node[LOWER]->upper);
1224 }
1225 }
1226 return 0;
1227fail:
1228 while (!list_empty(&new_node->lower)) {
1229 new_edge = list_entry(new_node->lower.next,
1230 struct backref_edge, list[UPPER]);
1231 list_del(&new_edge->list[UPPER]);
1232 free_backref_edge(cache, new_edge);
1233 }
1234 free_backref_node(cache, new_node);
1235 return -ENOMEM;
1236}
1237
5d4f98a2
YZ		 1238/*
1239 * helper to add 'address of tree root -> reloc tree' mapping
1240 */
ffd7b339 1241static int __must_check __add_reloc_root(struct btrfs_root *root)
5d4f98a2
YZ		 1242{
1243 struct rb_node *rb_node;
1244 struct mapping_node *node;
1245 struct reloc_control *rc = root->fs_info->reloc_ctl;
1246
1247 node = kmalloc(sizeof(*node), GFP_NOFS);
ffd7b339
JM		 1248 if (!node)
1249 return -ENOMEM;
5d4f98a2
YZ		 1250
1251 node->bytenr = root->node->start;
1252 node->data = root;
1253
1254 spin_lock(&rc->reloc_root_tree.lock);
1255 rb_node = tree_insert(&rc->reloc_root_tree.rb_root,
1256 node->bytenr, &node->rb_node);
1257 spin_unlock(&rc->reloc_root_tree.lock);
ffd7b339 1258 if (rb_node) {
ffd7b339
JM		 1259 btrfs_panic(root->fs_info, -EEXIST, "Duplicate root found "
1260 "for start=%llu while inserting into relocation "
533574c6 1261 "tree\n", node->bytenr);
23291a04
DC		 1262 kfree(node);
1263 return -EEXIST;
ffd7b339 1264 }
5d4f98a2
YZ		 1265
1266 list_add_tail(&root->root_list, &rc->reloc_roots);
1267 return 0;
1268}
1269
1270/*
c974c464 1271 * helper to delete the 'address of tree root -> reloc tree'
5d4f98a2
YZ		 1272 * mapping
1273 */
c974c464 1274static void __del_reloc_root(struct btrfs_root *root)
5d4f98a2
YZ		 1275{
1276 struct rb_node *rb_node;
1277 struct mapping_node *node = NULL;
1278 struct reloc_control *rc = root->fs_info->reloc_ctl;
1279
1280 spin_lock(&rc->reloc_root_tree.lock);
1281 rb_node = tree_search(&rc->reloc_root_tree.rb_root,
c974c464 1282 root->node->start);
5d4f98a2
YZ		 1283 if (rb_node) {
1284 node = rb_entry(rb_node, struct mapping_node, rb_node);
1285 rb_erase(&node->rb_node, &rc->reloc_root_tree.rb_root);
1286 }
1287 spin_unlock(&rc->reloc_root_tree.lock);
1288
8f71f3e0 1289 if (!node)
c974c464 1290 return;
5d4f98a2
YZ		 1291 BUG_ON((struct btrfs_root *)node->data != root);
1292
c974c464
WS		 1293 spin_lock(&root->fs_info->trans_lock);
1294 list_del_init(&root->root_list);
1295 spin_unlock(&root->fs_info->trans_lock);
1296 kfree(node);
1297}
1298
1299/*
1300 * helper to update the 'address of tree root -> reloc tree'
1301 * mapping
1302 */
1303static int __update_reloc_root(struct btrfs_root *root, u64 new_bytenr)
1304{
1305 struct rb_node *rb_node;
1306 struct mapping_node *node = NULL;
1307 struct reloc_control *rc = root->fs_info->reloc_ctl;
1308
1309 spin_lock(&rc->reloc_root_tree.lock);
1310 rb_node = tree_search(&rc->reloc_root_tree.rb_root,
1311 root->node->start);
1312 if (rb_node) {
1313 node = rb_entry(rb_node, struct mapping_node, rb_node);
1314 rb_erase(&node->rb_node, &rc->reloc_root_tree.rb_root);
5d4f98a2 1315 }
c974c464
WS		 1316 spin_unlock(&rc->reloc_root_tree.lock);
1317
1318 if (!node)
1319 return 0;
1320 BUG_ON((struct btrfs_root *)node->data != root);
1321
1322 spin_lock(&rc->reloc_root_tree.lock);
1323 node->bytenr = new_bytenr;
1324 rb_node = tree_insert(&rc->reloc_root_tree.rb_root,
1325 node->bytenr, &node->rb_node);
1326 spin_unlock(&rc->reloc_root_tree.lock);
1327 if (rb_node)
1328 backref_tree_panic(rb_node, -EEXIST, node->bytenr);
5d4f98a2
YZ		 1329 return 0;
1330}
1331
3fd0a558
YZ		 1332static struct btrfs_root *create_reloc_root(struct btrfs_trans_handle *trans,
1333 struct btrfs_root *root, u64 objectid)
5d4f98a2
YZ		 1334{
1335 struct btrfs_root *reloc_root;
1336 struct extent_buffer *eb;
1337 struct btrfs_root_item *root_item;
1338 struct btrfs_key root_key;
5bc7247a 1339 u64 last_snap = 0;
5d4f98a2
YZ		 1340 int ret;
1341
5d4f98a2
YZ		 1342 root_item = kmalloc(sizeof(*root_item), GFP_NOFS);
1343 BUG_ON(!root_item);
1344
1345 root_key.objectid = BTRFS_TREE_RELOC_OBJECTID;
1346 root_key.type = BTRFS_ROOT_ITEM_KEY;
3fd0a558 1347 root_key.offset = objectid;
5d4f98a2 1348
3fd0a558
YZ		 1349 if (root->root_key.objectid == objectid) {
1350 /* called by btrfs_init_reloc_root */
1351 ret = btrfs_copy_root(trans, root, root->commit_root, &eb,
1352 BTRFS_TREE_RELOC_OBJECTID);
1353 BUG_ON(ret);
1354
5bc7247a 1355 last_snap = btrfs_root_last_snapshot(&root->root_item);
3fd0a558
YZ		 1356 btrfs_set_root_last_snapshot(&root->root_item,
1357 trans->transid - 1);
1358 } else {
1359 /*
1360 * called by btrfs_reloc_post_snapshot_hook.
1361 * the source tree is a reloc tree, all tree blocks
1362 * modified after it was created have RELOC flag
1363 * set in their headers. so it's OK to not update
1364 * the 'last_snapshot'.
1365 */
1366 ret = btrfs_copy_root(trans, root, root->node, &eb,
1367 BTRFS_TREE_RELOC_OBJECTID);
1368 BUG_ON(ret);
1369 }
5d4f98a2 1370
5d4f98a2 1371 memcpy(root_item, &root->root_item, sizeof(*root_item));
5d4f98a2
YZ		 1372 btrfs_set_root_bytenr(root_item, eb->start);
1373 btrfs_set_root_level(root_item, btrfs_header_level(eb));
1374 btrfs_set_root_generation(root_item, trans->transid);
3fd0a558
YZ		 1375
1376 if (root->root_key.objectid == objectid) {
1377 btrfs_set_root_refs(root_item, 0);
1378 memset(&root_item->drop_progress, 0,
1379 sizeof(struct btrfs_disk_key));
1380 root_item->drop_level = 0;
5bc7247a
MX		 1381 /*
1382 * abuse rtransid, it is safe because it is impossible to
1383 * receive data into a relocation tree.
1384 */
1385 btrfs_set_root_rtransid(root_item, last_snap);
1386 btrfs_set_root_otransid(root_item, trans->transid);
3fd0a558 1387 }
5d4f98a2
YZ		 1388
1389 btrfs_tree_unlock(eb);
1390 free_extent_buffer(eb);
1391
1392 ret = btrfs_insert_root(trans, root->fs_info->tree_root,
1393 &root_key, root_item);
1394 BUG_ON(ret);
1395 kfree(root_item);
1396
cb517eab 1397 reloc_root = btrfs_read_fs_root(root->fs_info->tree_root, &root_key);
5d4f98a2
YZ		 1398 BUG_ON(IS_ERR(reloc_root));
1399 reloc_root->last_trans = trans->transid;
3fd0a558
YZ		 1400 return reloc_root;
1401}
1402
1403/*
1404 * create reloc tree for a given fs tree. reloc tree is just a
1405 * snapshot of the fs tree with special root objectid.
1406 */
1407int btrfs_init_reloc_root(struct btrfs_trans_handle *trans,
1408 struct btrfs_root *root)
1409{
1410 struct btrfs_root *reloc_root;
1411 struct reloc_control *rc = root->fs_info->reloc_ctl;
20dd2cbf 1412 struct btrfs_block_rsv *rsv;
3fd0a558 1413 int clear_rsv = 0;
ffd7b339 1414 int ret;
3fd0a558
YZ		 1415
1416 if (root->reloc_root) {
1417 reloc_root = root->reloc_root;
1418 reloc_root->last_trans = trans->transid;
1419 return 0;
1420 }
1421
1422 if (!rc || !rc->create_reloc_tree ||
1423 root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID)
1424 return 0;
1425
20dd2cbf
MX		 1426 if (!trans->reloc_reserved) {
1427 rsv = trans->block_rsv;
3fd0a558
YZ		 1428 trans->block_rsv = rc->block_rsv;
1429 clear_rsv = 1;
1430 }
1431 reloc_root = create_reloc_root(trans, root, root->root_key.objectid);
1432 if (clear_rsv)
20dd2cbf 1433 trans->block_rsv = rsv;
5d4f98a2 1434
ffd7b339
JM		 1435 ret = __add_reloc_root(reloc_root);
1436 BUG_ON(ret < 0);
5d4f98a2
YZ		 1437 root->reloc_root = reloc_root;
1438 return 0;
1439}
1440
1441/*
1442 * update root item of reloc tree
1443 */
1444int btrfs_update_reloc_root(struct btrfs_trans_handle *trans,
1445 struct btrfs_root *root)
1446{
1447 struct btrfs_root *reloc_root;
1448 struct btrfs_root_item *root_item;
5d4f98a2
YZ		 1449 int ret;
1450
1451 if (!root->reloc_root)
7585717f 1452 goto out;
5d4f98a2
YZ		 1453
1454 reloc_root = root->reloc_root;
1455 root_item = &reloc_root->root_item;
1456
3fd0a558
YZ		 1457 if (root->fs_info->reloc_ctl->merge_reloc_tree &&
1458 btrfs_root_refs(root_item) == 0) {
5d4f98a2 1459 root->reloc_root = NULL;
c974c464 1460 __del_reloc_root(reloc_root);
5d4f98a2
YZ		 1461 }
1462
5d4f98a2
YZ		 1463 if (reloc_root->commit_root != reloc_root->node) {
1464 btrfs_set_root_node(root_item, reloc_root->node);
1465 free_extent_buffer(reloc_root->commit_root);
1466 reloc_root->commit_root = btrfs_root_node(reloc_root);
1467 }
1468
1469 ret = btrfs_update_root(trans, root->fs_info->tree_root,
1470 &reloc_root->root_key, root_item);
1471 BUG_ON(ret);
7585717f
CM		 1472
1473out:
5d4f98a2
YZ		 1474 return 0;
1475}
1476
1477/*
1478 * helper to find first cached inode with inode number >= objectid
1479 * in a subvolume
1480 */
1481static struct inode *find_next_inode(struct btrfs_root *root, u64 objectid)
1482{
1483 struct rb_node *node;
1484 struct rb_node *prev;
1485 struct btrfs_inode *entry;
1486 struct inode *inode;
1487
1488 spin_lock(&root->inode_lock);
1489again:
1490 node = root->inode_tree.rb_node;
1491 prev = NULL;
1492 while (node) {
1493 prev = node;
1494 entry = rb_entry(node, struct btrfs_inode, rb_node);
1495
33345d01 1496 if (objectid < btrfs_ino(&entry->vfs_inode))
5d4f98a2 1497 node = node->rb_left;
33345d01 1498 else if (objectid > btrfs_ino(&entry->vfs_inode))
5d4f98a2
YZ		 1499 node = node->rb_right;
1500 else
1501 break;
1502 }
1503 if (!node) {
1504 while (prev) {
1505 entry = rb_entry(prev, struct btrfs_inode, rb_node);
33345d01 1506 if (objectid <= btrfs_ino(&entry->vfs_inode)) {
5d4f98a2
YZ		 1507 node = prev;
1508 break;
1509 }
1510 prev = rb_next(prev);
1511 }
1512 }
1513 while (node) {
1514 entry = rb_entry(node, struct btrfs_inode, rb_node);
1515 inode = igrab(&entry->vfs_inode);
1516 if (inode) {
1517 spin_unlock(&root->inode_lock);
1518 return inode;
1519 }
1520
33345d01 1521 objectid = btrfs_ino(&entry->vfs_inode) + 1;
5d4f98a2
YZ		 1522 if (cond_resched_lock(&root->inode_lock))
1523 goto again;
1524
1525 node = rb_next(node);
1526 }
1527 spin_unlock(&root->inode_lock);
1528 return NULL;
1529}
1530
1531static int in_block_group(u64 bytenr,
1532 struct btrfs_block_group_cache *block_group)
1533{
1534 if (bytenr >= block_group->key.objectid &&
1535 bytenr < block_group->key.objectid + block_group->key.offset)
1536 return 1;
1537 return 0;
1538}
1539
1540/*
1541 * get new location of data
1542 */
1543static int get_new_location(struct inode *reloc_inode, u64 *new_bytenr,
1544 u64 bytenr, u64 num_bytes)
1545{
1546 struct btrfs_root *root = BTRFS_I(reloc_inode)->root;
1547 struct btrfs_path *path;
1548 struct btrfs_file_extent_item *fi;
1549 struct extent_buffer *leaf;
1550 int ret;
1551
1552 path = btrfs_alloc_path();
1553 if (!path)
1554 return -ENOMEM;
1555
1556 bytenr -= BTRFS_I(reloc_inode)->index_cnt;
33345d01 1557 ret = btrfs_lookup_file_extent(NULL, root, path, btrfs_ino(reloc_inode),
5d4f98a2
YZ		 1558 bytenr, 0);
1559 if (ret < 0)
1560 goto out;
1561 if (ret > 0) {
1562 ret = -ENOENT;
1563 goto out;
1564 }
1565
1566 leaf = path->nodes[0];
1567 fi = btrfs_item_ptr(leaf, path->slots[0],
1568 struct btrfs_file_extent_item);
1569
1570 BUG_ON(btrfs_file_extent_offset(leaf, fi) ||
1571 btrfs_file_extent_compression(leaf, fi) ||
1572 btrfs_file_extent_encryption(leaf, fi) ||
1573 btrfs_file_extent_other_encoding(leaf, fi));
1574
1575 if (num_bytes != btrfs_file_extent_disk_num_bytes(leaf, fi)) {
83d4cfd4 1576 ret = -EINVAL;
5d4f98a2
YZ		 1577 goto out;
1578 }
1579
3fd0a558 1580 *new_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
5d4f98a2
YZ		 1581 ret = 0;
1582out:
1583 btrfs_free_path(path);
1584 return ret;
1585}
1586
1587/*
1588 * update file extent items in the tree leaf to point to
1589 * the new locations.
1590 */
3fd0a558
YZ		 1591static noinline_for_stack
1592int replace_file_extents(struct btrfs_trans_handle *trans,
1593 struct reloc_control *rc,
1594 struct btrfs_root *root,
1595 struct extent_buffer *leaf)
5d4f98a2
YZ		 1596{
1597 struct btrfs_key key;
1598 struct btrfs_file_extent_item *fi;
1599 struct inode *inode = NULL;
5d4f98a2
YZ		 1600 u64 parent;
1601 u64 bytenr;
3fd0a558 1602 u64 new_bytenr = 0;
5d4f98a2
YZ		 1603 u64 num_bytes;
1604 u64 end;
1605 u32 nritems;
1606 u32 i;
83d4cfd4 1607 int ret = 0;
5d4f98a2
YZ		 1608 int first = 1;
1609 int dirty = 0;
1610
1611 if (rc->stage != UPDATE_DATA_PTRS)
1612 return 0;
1613
1614 /* reloc trees always use full backref */
1615 if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID)
1616 parent = leaf->start;
1617 else
1618 parent = 0;
1619
1620 nritems = btrfs_header_nritems(leaf);
1621 for (i = 0; i < nritems; i++) {
1622 cond_resched();
1623 btrfs_item_key_to_cpu(leaf, &key, i);
1624 if (key.type != BTRFS_EXTENT_DATA_KEY)
1625 continue;
1626 fi = btrfs_item_ptr(leaf, i, struct btrfs_file_extent_item);
1627 if (btrfs_file_extent_type(leaf, fi) ==
1628 BTRFS_FILE_EXTENT_INLINE)
1629 continue;
1630 bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
1631 num_bytes = btrfs_file_extent_disk_num_bytes(leaf, fi);
1632 if (bytenr == 0)
1633 continue;
1634 if (!in_block_group(bytenr, rc->block_group))
1635 continue;
1636
1637 /*
1638 * if we are modifying block in fs tree, wait for readpage
1639 * to complete and drop the extent cache
1640 */
1641 if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) {
5d4f98a2
YZ		 1642 if (first) {
1643 inode = find_next_inode(root, key.objectid);
5d4f98a2 1644 first = 0;
33345d01 1645 } else if (inode && btrfs_ino(inode) < key.objectid) {
3fd0a558 1646 btrfs_add_delayed_iput(inode);
5d4f98a2 1647 inode = find_next_inode(root, key.objectid);
5d4f98a2 1648 }
33345d01 1649 if (inode && btrfs_ino(inode) == key.objectid) {
5d4f98a2
YZ		 1650 end = key.offset +
1651 btrfs_file_extent_num_bytes(leaf, fi);
1652 WARN_ON(!IS_ALIGNED(key.offset,
1653 root->sectorsize));
1654 WARN_ON(!IS_ALIGNED(end, root->sectorsize));
1655 end--;
1656 ret = try_lock_extent(&BTRFS_I(inode)->io_tree,
d0082371 1657 key.offset, end);
5d4f98a2
YZ		 1658 if (!ret)
1659 continue;
1660
1661 btrfs_drop_extent_cache(inode, key.offset, end,
1662 1);
1663 unlock_extent(&BTRFS_I(inode)->io_tree,
d0082371 1664 key.offset, end);
5d4f98a2
YZ		 1665 }
1666 }
1667
1668 ret = get_new_location(rc->data_inode, &new_bytenr,
1669 bytenr, num_bytes);
83d4cfd4
JB		 1670 if (ret) {
1671 /*
1672 * Don't have to abort since we've not changed anything
1673 * in the file extent yet.
1674 */
1675 break;
3fd0a558 1676 }
5d4f98a2
YZ		 1677
1678 btrfs_set_file_extent_disk_bytenr(leaf, fi, new_bytenr);
1679 dirty = 1;
1680
1681 key.offset -= btrfs_file_extent_offset(leaf, fi);
1682 ret = btrfs_inc_extent_ref(trans, root, new_bytenr,
1683 num_bytes, parent,
1684 btrfs_header_owner(leaf),
66d7e7f0 1685 key.objectid, key.offset, 1);
83d4cfd4
JB		 1686 if (ret) {
1687 btrfs_abort_transaction(trans, root, ret);
1688 break;
1689 }
5d4f98a2
YZ		 1690
1691 ret = btrfs_free_extent(trans, root, bytenr, num_bytes,
1692 parent, btrfs_header_owner(leaf),
66d7e7f0 1693 key.objectid, key.offset, 1);
83d4cfd4
JB		 1694 if (ret) {
1695 btrfs_abort_transaction(trans, root, ret);
1696 break;
1697 }
5d4f98a2
YZ		 1698 }
1699 if (dirty)
1700 btrfs_mark_buffer_dirty(leaf);
3fd0a558
YZ		 1701 if (inode)
1702 btrfs_add_delayed_iput(inode);
83d4cfd4 1703 return ret;
5d4f98a2
YZ		 1704}
1705
1706static noinline_for_stack
1707int memcmp_node_keys(struct extent_buffer *eb, int slot,
1708 struct btrfs_path *path, int level)
1709{
1710 struct btrfs_disk_key key1;
1711 struct btrfs_disk_key key2;
1712 btrfs_node_key(eb, &key1, slot);
1713 btrfs_node_key(path->nodes[level], &key2, path->slots[level]);
1714 return memcmp(&key1, &key2, sizeof(key1));
1715}
1716
1717/*
1718 * try to replace tree blocks in fs tree with the new blocks
1719 * in reloc tree. tree blocks haven't been modified since the
1720 * reloc tree was create can be replaced.
1721 *
1722 * if a block was replaced, level of the block + 1 is returned.
1723 * if no block got replaced, 0 is returned. if there are other
1724 * errors, a negative error number is returned.
1725 */
3fd0a558
YZ		 1726static noinline_for_stack
1727int replace_path(struct btrfs_trans_handle *trans,
1728 struct btrfs_root *dest, struct btrfs_root *src,
1729 struct btrfs_path *path, struct btrfs_key *next_key,
1730 int lowest_level, int max_level)
5d4f98a2
YZ		 1731{
1732 struct extent_buffer *eb;
1733 struct extent_buffer *parent;
1734 struct btrfs_key key;
1735 u64 old_bytenr;
1736 u64 new_bytenr;
1737 u64 old_ptr_gen;
1738 u64 new_ptr_gen;
1739 u64 last_snapshot;
1740 u32 blocksize;
3fd0a558 1741 int cow = 0;
5d4f98a2
YZ		 1742 int level;
1743 int ret;
1744 int slot;
1745
1746 BUG_ON(src->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID);
1747 BUG_ON(dest->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID);
5d4f98a2
YZ		 1748
1749 last_snapshot = btrfs_root_last_snapshot(&src->root_item);
3fd0a558 1750again:
5d4f98a2
YZ		 1751 slot = path->slots[lowest_level];
1752 btrfs_node_key_to_cpu(path->nodes[lowest_level], &key, slot);
1753
1754 eb = btrfs_lock_root_node(dest);
1755 btrfs_set_lock_blocking(eb);
1756 level = btrfs_header_level(eb);
1757
1758 if (level < lowest_level) {
1759 btrfs_tree_unlock(eb);
1760 free_extent_buffer(eb);
1761 return 0;
1762 }
1763
3fd0a558
YZ		 1764 if (cow) {
1765 ret = btrfs_cow_block(trans, dest, eb, NULL, 0, &eb);
1766 BUG_ON(ret);
1767 }
5d4f98a2
YZ		 1768 btrfs_set_lock_blocking(eb);
1769
1770 if (next_key) {
1771 next_key->objectid = (u64)-1;
1772 next_key->type = (u8)-1;
1773 next_key->offset = (u64)-1;
1774 }
1775
1776 parent = eb;
1777 while (1) {
1778 level = btrfs_header_level(parent);
1779 BUG_ON(level < lowest_level);
1780
1781 ret = btrfs_bin_search(parent, &key, level, &slot);
1782 if (ret && slot > 0)
1783 slot--;
1784
1785 if (next_key && slot + 1 < btrfs_header_nritems(parent))
1786 btrfs_node_key_to_cpu(parent, next_key, slot + 1);
1787
1788 old_bytenr = btrfs_node_blockptr(parent, slot);
1789 blocksize = btrfs_level_size(dest, level - 1);
1790 old_ptr_gen = btrfs_node_ptr_generation(parent, slot);
1791
1792 if (level <= max_level) {
1793 eb = path->nodes[level];
1794 new_bytenr = btrfs_node_blockptr(eb,
1795 path->slots[level]);
1796 new_ptr_gen = btrfs_node_ptr_generation(eb,
1797 path->slots[level]);
1798 } else {
1799 new_bytenr = 0;
1800 new_ptr_gen = 0;
1801 }
1802
fae7f21c 1803 if (WARN_ON(new_bytenr > 0 && new_bytenr == old_bytenr)) {
5d4f98a2
YZ		 1804 ret = level;
1805 break;
1806 }
1807
1808 if (new_bytenr == 0 || old_ptr_gen > last_snapshot ||
1809 memcmp_node_keys(parent, slot, path, level)) {
3fd0a558 1810 if (level <= lowest_level) {
5d4f98a2
YZ		 1811 ret = 0;
1812 break;
1813 }
1814
1815 eb = read_tree_block(dest, old_bytenr, blocksize,
1816 old_ptr_gen);
416bc658
JB		 1817 if (!eb || !extent_buffer_uptodate(eb)) {
1818 ret = (!eb) ? -ENOMEM : -EIO;
1819 free_extent_buffer(eb);
379cde74 1820 break;
416bc658 1821 }
5d4f98a2 1822 btrfs_tree_lock(eb);
3fd0a558
YZ		 1823 if (cow) {
1824 ret = btrfs_cow_block(trans, dest, eb, parent,
1825 slot, &eb);
1826 BUG_ON(ret);
5d4f98a2 1827 }
3fd0a558 1828 btrfs_set_lock_blocking(eb);
5d4f98a2
YZ		 1829
1830 btrfs_tree_unlock(parent);
1831 free_extent_buffer(parent);
1832
1833 parent = eb;
1834 continue;
1835 }
1836
3fd0a558
YZ		 1837 if (!cow) {
1838 btrfs_tree_unlock(parent);
1839 free_extent_buffer(parent);
1840 cow = 1;
1841 goto again;
1842 }
1843
5d4f98a2
YZ		 1844 btrfs_node_key_to_cpu(path->nodes[level], &key,
1845 path->slots[level]);
b3b4aa74 1846 btrfs_release_path(path);
5d4f98a2
YZ		 1847
1848 path->lowest_level = level;
1849 ret = btrfs_search_slot(trans, src, &key, path, 0, 1);
1850 path->lowest_level = 0;
1851 BUG_ON(ret);
1852
1853 /*
1854 * swap blocks in fs tree and reloc tree.
1855 */
1856 btrfs_set_node_blockptr(parent, slot, new_bytenr);
1857 btrfs_set_node_ptr_generation(parent, slot, new_ptr_gen);
1858 btrfs_mark_buffer_dirty(parent);
1859
1860 btrfs_set_node_blockptr(path->nodes[level],
1861 path->slots[level], old_bytenr);
1862 btrfs_set_node_ptr_generation(path->nodes[level],
1863 path->slots[level], old_ptr_gen);
1864 btrfs_mark_buffer_dirty(path->nodes[level]);
1865
1866 ret = btrfs_inc_extent_ref(trans, src, old_bytenr, blocksize,
1867 path->nodes[level]->start,
66d7e7f0
AJ		 1868 src->root_key.objectid, level - 1, 0,
1869 1);
5d4f98a2
YZ		 1870 BUG_ON(ret);
1871 ret = btrfs_inc_extent_ref(trans, dest, new_bytenr, blocksize,
1872 0, dest->root_key.objectid, level - 1,
66d7e7f0 1873 0, 1);
5d4f98a2
YZ		 1874 BUG_ON(ret);
1875
1876 ret = btrfs_free_extent(trans, src, new_bytenr, blocksize,
1877 path->nodes[level]->start,
66d7e7f0
AJ		 1878 src->root_key.objectid, level - 1, 0,
1879 1);
5d4f98a2
YZ		 1880 BUG_ON(ret);
1881
1882 ret = btrfs_free_extent(trans, dest, old_bytenr, blocksize,
1883 0, dest->root_key.objectid, level - 1,
66d7e7f0 1884 0, 1);
5d4f98a2
YZ		 1885 BUG_ON(ret);
1886
1887 btrfs_unlock_up_safe(path, 0);
1888
1889 ret = level;
1890 break;
1891 }
1892 btrfs_tree_unlock(parent);
1893 free_extent_buffer(parent);
1894 return ret;
1895}
1896
1897/*
1898 * helper to find next relocated block in reloc tree
1899 */
1900static noinline_for_stack
1901int walk_up_reloc_tree(struct btrfs_root *root, struct btrfs_path *path,
1902 int *level)
1903{
1904 struct extent_buffer *eb;
1905 int i;
1906 u64 last_snapshot;
1907 u32 nritems;
1908
1909 last_snapshot = btrfs_root_last_snapshot(&root->root_item);
1910
1911 for (i = 0; i < *level; i++) {
1912 free_extent_buffer(path->nodes[i]);
1913 path->nodes[i] = NULL;
1914 }
1915
1916 for (i = *level; i < BTRFS_MAX_LEVEL && path->nodes[i]; i++) {
1917 eb = path->nodes[i];
1918 nritems = btrfs_header_nritems(eb);
1919 while (path->slots[i] + 1 < nritems) {
1920 path->slots[i]++;
1921 if (btrfs_node_ptr_generation(eb, path->slots[i]) <=
1922 last_snapshot)
1923 continue;
1924
1925 *level = i;
1926 return 0;
1927 }
1928 free_extent_buffer(path->nodes[i]);
1929 path->nodes[i] = NULL;
1930 }
1931 return 1;
1932}
1933
1934/*
1935 * walk down reloc tree to find relocated block of lowest level
1936 */
1937static noinline_for_stack
1938int walk_down_reloc_tree(struct btrfs_root *root, struct btrfs_path *path,
1939 int *level)
1940{
1941 struct extent_buffer *eb = NULL;
1942 int i;
1943 u64 bytenr;
1944 u64 ptr_gen = 0;
1945 u64 last_snapshot;
1946 u32 blocksize;
1947 u32 nritems;
1948
1949 last_snapshot = btrfs_root_last_snapshot(&root->root_item);
1950
1951 for (i = *level; i > 0; i--) {
1952 eb = path->nodes[i];
1953 nritems = btrfs_header_nritems(eb);
1954 while (path->slots[i] < nritems) {
1955 ptr_gen = btrfs_node_ptr_generation(eb, path->slots[i]);
1956 if (ptr_gen > last_snapshot)
1957 break;
1958 path->slots[i]++;
1959 }
1960 if (path->slots[i] >= nritems) {
1961 if (i == *level)
1962 break;
1963 *level = i + 1;
1964 return 0;
1965 }
1966 if (i == 1) {
1967 *level = i;
1968 return 0;
1969 }
1970
1971 bytenr = btrfs_node_blockptr(eb, path->slots[i]);
1972 blocksize = btrfs_level_size(root, i - 1);
1973 eb = read_tree_block(root, bytenr, blocksize, ptr_gen);
416bc658
JB		 1974 if (!eb || !extent_buffer_uptodate(eb)) {
1975 free_extent_buffer(eb);
1976 return -EIO;
1977 }
5d4f98a2
YZ		 1978 BUG_ON(btrfs_header_level(eb) != i - 1);
1979 path->nodes[i - 1] = eb;
1980 path->slots[i - 1] = 0;
1981 }
1982 return 1;
1983}
1984
1985/*
1986 * invalidate extent cache for file extents whose key in range of
1987 * [min_key, max_key)
1988 */
1989static int invalidate_extent_cache(struct btrfs_root *root,
1990 struct btrfs_key *min_key,
1991 struct btrfs_key *max_key)
1992{
1993 struct inode *inode = NULL;
1994 u64 objectid;
1995 u64 start, end;
33345d01 1996 u64 ino;
5d4f98a2
YZ		 1997
1998 objectid = min_key->objectid;
1999 while (1) {
2000 cond_resched();
2001 iput(inode);
2002
2003 if (objectid > max_key->objectid)
2004 break;
2005
2006 inode = find_next_inode(root, objectid);
2007 if (!inode)
2008 break;
33345d01 2009 ino = btrfs_ino(inode);
5d4f98a2 2010
33345d01 2011 if (ino > max_key->objectid) {
5d4f98a2
YZ		 2012 iput(inode);
2013 break;
2014 }
2015
33345d01 2016 objectid = ino + 1;
5d4f98a2
YZ		 2017 if (!S_ISREG(inode->i_mode))
2018 continue;
2019
33345d01 2020 if (unlikely(min_key->objectid == ino)) {
5d4f98a2
YZ		 2021 if (min_key->type > BTRFS_EXTENT_DATA_KEY)
2022 continue;
2023 if (min_key->type < BTRFS_EXTENT_DATA_KEY)
2024 start = 0;
2025 else {
2026 start = min_key->offset;
2027 WARN_ON(!IS_ALIGNED(start, root->sectorsize));
2028 }
2029 } else {
2030 start = 0;
2031 }
2032
33345d01 2033 if (unlikely(max_key->objectid == ino)) {
5d4f98a2
YZ		 2034 if (max_key->type < BTRFS_EXTENT_DATA_KEY)
2035 continue;
2036 if (max_key->type > BTRFS_EXTENT_DATA_KEY) {
2037 end = (u64)-1;
2038 } else {
2039 if (max_key->offset == 0)
2040 continue;
2041 end = max_key->offset;
2042 WARN_ON(!IS_ALIGNED(end, root->sectorsize));
2043 end--;
2044 }
2045 } else {
2046 end = (u64)-1;
2047 }
2048
2049 /* the lock_extent waits for readpage to complete */
d0082371 2050 lock_extent(&BTRFS_I(inode)->io_tree, start, end);
5d4f98a2 2051 btrfs_drop_extent_cache(inode, start, end, 1);
d0082371 2052 unlock_extent(&BTRFS_I(inode)->io_tree, start, end);
5d4f98a2
YZ		 2053 }
2054 return 0;
2055}
2056
2057static int find_next_key(struct btrfs_path *path, int level,
2058 struct btrfs_key *key)
2059
2060{
2061 while (level < BTRFS_MAX_LEVEL) {
2062 if (!path->nodes[level])
2063 break;
2064 if (path->slots[level] + 1 <
2065 btrfs_header_nritems(path->nodes[level])) {
2066 btrfs_node_key_to_cpu(path->nodes[level], key,
2067 path->slots[level] + 1);
2068 return 0;
2069 }
2070 level++;
2071 }
2072 return 1;
2073}
2074
2075/*
2076 * merge the relocated tree blocks in reloc tree with corresponding
2077 * fs tree.
2078 */
2079static noinline_for_stack int merge_reloc_root(struct reloc_control *rc,
2080 struct btrfs_root *root)
2081{
2082 LIST_HEAD(inode_list);
2083 struct btrfs_key key;
2084 struct btrfs_key next_key;
9e6a0c52 2085 struct btrfs_trans_handle *trans = NULL;
5d4f98a2
YZ		 2086 struct btrfs_root *reloc_root;
2087 struct btrfs_root_item *root_item;
2088 struct btrfs_path *path;
3fd0a558 2089 struct extent_buffer *leaf;
5d4f98a2
YZ		 2090 int level;
2091 int max_level;
2092 int replaced = 0;
2093 int ret;
2094 int err = 0;
3fd0a558 2095 u32 min_reserved;
5d4f98a2
YZ		 2096
2097 path = btrfs_alloc_path();
2098 if (!path)
2099 return -ENOMEM;
026fd317 2100 path->reada = 1;
5d4f98a2
YZ		 2101
2102 reloc_root = root->reloc_root;
2103 root_item = &reloc_root->root_item;
2104
2105 if (btrfs_disk_key_objectid(&root_item->drop_progress) == 0) {
2106 level = btrfs_root_level(root_item);
2107 extent_buffer_get(reloc_root->node);
2108 path->nodes[level] = reloc_root->node;
2109 path->slots[level] = 0;
2110 } else {
2111 btrfs_disk_key_to_cpu(&key, &root_item->drop_progress);
2112
2113 level = root_item->drop_level;
2114 BUG_ON(level == 0);
2115 path->lowest_level = level;
2116 ret = btrfs_search_slot(NULL, reloc_root, &key, path, 0, 0);
33c66f43 2117 path->lowest_level = 0;
5d4f98a2
YZ		 2118 if (ret < 0) {
2119 btrfs_free_path(path);
2120 return ret;
2121 }
2122
2123 btrfs_node_key_to_cpu(path->nodes[level], &next_key,
2124 path->slots[level]);
2125 WARN_ON(memcmp(&key, &next_key, sizeof(key)));
2126
2127 btrfs_unlock_up_safe(path, 0);
2128 }
2129
3fd0a558
YZ		 2130 min_reserved = root->nodesize * (BTRFS_MAX_LEVEL - 1) * 2;
2131 memset(&next_key, 0, sizeof(next_key));
5d4f98a2 2132
3fd0a558 2133 while (1) {
08e007d2
MX		 2134 ret = btrfs_block_rsv_refill(root, rc->block_rsv, min_reserved,
2135 BTRFS_RESERVE_FLUSH_ALL);
3fd0a558 2136 if (ret) {
9e6a0c52
JB		 2137 err = ret;
2138 goto out;
5d4f98a2 2139 }
9e6a0c52
JB		 2140 trans = btrfs_start_transaction(root, 0);
2141 if (IS_ERR(trans)) {
2142 err = PTR_ERR(trans);
2143 trans = NULL;
2144 goto out;
2145 }
2146 trans->block_rsv = rc->block_rsv;
5d4f98a2 2147
5d4f98a2 2148 replaced = 0;
5d4f98a2
YZ		 2149 max_level = level;
2150
2151 ret = walk_down_reloc_tree(reloc_root, path, &level);
2152 if (ret < 0) {
2153 err = ret;
2154 goto out;
2155 }
2156 if (ret > 0)
2157 break;
2158
2159 if (!find_next_key(path, level, &key) &&
2160 btrfs_comp_cpu_keys(&next_key, &key) >= 0) {
2161 ret = 0;
5d4f98a2 2162 } else {
3fd0a558
YZ		 2163 ret = replace_path(trans, root, reloc_root, path,
2164 &next_key, level, max_level);
5d4f98a2
YZ		 2165 }
2166 if (ret < 0) {
2167 err = ret;
2168 goto out;
2169 }
2170
2171 if (ret > 0) {
2172 level = ret;
2173 btrfs_node_key_to_cpu(path->nodes[level], &key,
2174 path->slots[level]);
2175 replaced = 1;
5d4f98a2
YZ		 2176 }
2177
2178 ret = walk_up_reloc_tree(reloc_root, path, &level);
2179 if (ret > 0)
2180 break;
2181
2182 BUG_ON(level == 0);
2183 /*
2184 * save the merging progress in the drop_progress.
2185 * this is OK since root refs == 1 in this case.
2186 */
2187 btrfs_node_key(path->nodes[level], &root_item->drop_progress,
2188 path->slots[level]);
2189 root_item->drop_level = level;
2190
3fd0a558 2191 btrfs_end_transaction_throttle(trans, root);
9e6a0c52 2192 trans = NULL;
5d4f98a2 2193
b53d3f5d 2194 btrfs_btree_balance_dirty(root);
5d4f98a2
YZ		 2195
2196 if (replaced && rc->stage == UPDATE_DATA_PTRS)
2197 invalidate_extent_cache(root, &key, &next_key);
2198 }
2199
2200 /*
2201 * handle the case only one block in the fs tree need to be
2202 * relocated and the block is tree root.
2203 */
2204 leaf = btrfs_lock_root_node(root);
2205 ret = btrfs_cow_block(trans, root, leaf, NULL, 0, &leaf);
2206 btrfs_tree_unlock(leaf);
2207 free_extent_buffer(leaf);
2208 if (ret < 0)
2209 err = ret;
2210out:
2211 btrfs_free_path(path);
2212
2213 if (err == 0) {
2214 memset(&root_item->drop_progress, 0,
2215 sizeof(root_item->drop_progress));
2216 root_item->drop_level = 0;
2217 btrfs_set_root_refs(root_item, 0);
3fd0a558 2218 btrfs_update_reloc_root(trans, root);
5d4f98a2
YZ		 2219 }
2220
9e6a0c52
JB		 2221 if (trans)
2222 btrfs_end_transaction_throttle(trans, root);
5d4f98a2 2223
b53d3f5d 2224 btrfs_btree_balance_dirty(root);
5d4f98a2 2225
5d4f98a2
YZ		 2226 if (replaced && rc->stage == UPDATE_DATA_PTRS)
2227 invalidate_extent_cache(root, &key, &next_key);
2228
2229 return err;
2230}
2231
3fd0a558
YZ		 2232static noinline_for_stack
2233int prepare_to_merge(struct reloc_control *rc, int err)
5d4f98a2 2234{
3fd0a558 2235 struct btrfs_root *root = rc->extent_root;
5d4f98a2 2236 struct btrfs_root *reloc_root;
3fd0a558
YZ		 2237 struct btrfs_trans_handle *trans;
2238 LIST_HEAD(reloc_roots);
2239 u64 num_bytes = 0;
2240 int ret;
3fd0a558 2241
7585717f 2242 mutex_lock(&root->fs_info->reloc_mutex);
3fd0a558
YZ		 2243 rc->merging_rsv_size += root->nodesize * (BTRFS_MAX_LEVEL - 1) * 2;
2244 rc->merging_rsv_size += rc->nodes_relocated * 2;
7585717f
CM		 2245 mutex_unlock(&root->fs_info->reloc_mutex);
2246
3fd0a558
YZ		 2247again:
2248 if (!err) {
2249 num_bytes = rc->merging_rsv_size;
08e007d2
MX		 2250 ret = btrfs_block_rsv_add(root, rc->block_rsv, num_bytes,
2251 BTRFS_RESERVE_FLUSH_ALL);
3fd0a558
YZ		 2252 if (ret)
2253 err = ret;
2254 }
2255
7a7eaa40 2256 trans = btrfs_join_transaction(rc->extent_root);
3612b495
TI		 2257 if (IS_ERR(trans)) {
2258 if (!err)
2259 btrfs_block_rsv_release(rc->extent_root,
2260 rc->block_rsv, num_bytes);
2261 return PTR_ERR(trans);
2262 }
3fd0a558
YZ		 2263
2264 if (!err) {
2265 if (num_bytes != rc->merging_rsv_size) {
2266 btrfs_end_transaction(trans, rc->extent_root);
2267 btrfs_block_rsv_release(rc->extent_root,
2268 rc->block_rsv, num_bytes);
3fd0a558
YZ		 2269 goto again;
2270 }
2271 }
5d4f98a2 2272
3fd0a558
YZ		 2273 rc->merge_reloc_tree = 1;
2274
2275 while (!list_empty(&rc->reloc_roots)) {
2276 reloc_root = list_entry(rc->reloc_roots.next,
2277 struct btrfs_root, root_list);
2278 list_del_init(&reloc_root->root_list);
5d4f98a2 2279
5d4f98a2
YZ		 2280 root = read_fs_root(reloc_root->fs_info,
2281 reloc_root->root_key.offset);
2282 BUG_ON(IS_ERR(root));
2283 BUG_ON(root->reloc_root != reloc_root);
2284
3fd0a558
YZ		 2285 /*
2286 * set reference count to 1, so btrfs_recover_relocation
2287 * knows it should resumes merging
2288 */
2289 if (!err)
2290 btrfs_set_root_refs(&reloc_root->root_item, 1);
5d4f98a2 2291 btrfs_update_reloc_root(trans, root);
5d4f98a2 2292
3fd0a558
YZ		 2293 list_add(&reloc_root->root_list, &reloc_roots);
2294 }
5d4f98a2 2295
3fd0a558 2296 list_splice(&reloc_roots, &rc->reloc_roots);
5d4f98a2 2297
3fd0a558
YZ		 2298 if (!err)
2299 btrfs_commit_transaction(trans, rc->extent_root);
2300 else
2301 btrfs_end_transaction(trans, rc->extent_root);
2302 return err;
5d4f98a2
YZ		 2303}
2304
aca1bba6
LB		 2305static noinline_for_stack
2306void free_reloc_roots(struct list_head *list)
2307{
2308 struct btrfs_root *reloc_root;
2309
2310 while (!list_empty(list)) {
2311 reloc_root = list_entry(list->next, struct btrfs_root,
2312 root_list);
c974c464 2313 __del_reloc_root(reloc_root);
aca1bba6
LB		 2314 free_extent_buffer(reloc_root->node);
2315 free_extent_buffer(reloc_root->commit_root);
2316 kfree(reloc_root);
2317 }
2318}
2319
3fd0a558
YZ		 2320static noinline_for_stack
2321int merge_reloc_roots(struct reloc_control *rc)
5d4f98a2 2322{
5bc7247a 2323 struct btrfs_trans_handle *trans;
5d4f98a2 2324 struct btrfs_root *root;
3fd0a558 2325 struct btrfs_root *reloc_root;
5bc7247a
MX		 2326 u64 last_snap;
2327 u64 otransid;
2328 u64 objectid;
3fd0a558
YZ		 2329 LIST_HEAD(reloc_roots);
2330 int found = 0;
aca1bba6 2331 int ret = 0;
3fd0a558
YZ		 2332again:
2333 root = rc->extent_root;
7585717f
CM		 2334
2335 /*
2336 * this serializes us with btrfs_record_root_in_transaction,
2337 * we have to make sure nobody is in the middle of
2338 * adding their roots to the list while we are
2339 * doing this splice
2340 */
2341 mutex_lock(&root->fs_info->reloc_mutex);
3fd0a558 2342 list_splice_init(&rc->reloc_roots, &reloc_roots);
7585717f 2343 mutex_unlock(&root->fs_info->reloc_mutex);
5d4f98a2 2344
3fd0a558
YZ		 2345 while (!list_empty(&reloc_roots)) {
2346 found = 1;
2347 reloc_root = list_entry(reloc_roots.next,
2348 struct btrfs_root, root_list);
5d4f98a2 2349
3fd0a558
YZ		 2350 if (btrfs_root_refs(&reloc_root->root_item) > 0) {
2351 root = read_fs_root(reloc_root->fs_info,
2352 reloc_root->root_key.offset);
2353 BUG_ON(IS_ERR(root));
2354 BUG_ON(root->reloc_root != reloc_root);
5d4f98a2 2355
3fd0a558 2356 ret = merge_reloc_root(rc, root);
b37b39cd 2357 if (ret) {
c974c464 2358 __del_reloc_root(reloc_root);
b37b39cd
JB		 2359 free_extent_buffer(reloc_root->node);
2360 free_extent_buffer(reloc_root->commit_root);
2361 kfree(reloc_root);
aca1bba6 2362 goto out;
b37b39cd 2363 }
3fd0a558
YZ		 2364 } else {
2365 list_del_init(&reloc_root->root_list);
2366 }
5bc7247a
MX		 2367
2368 /*
2369 * we keep the old last snapshod transid in rtranid when we
2370 * created the relocation tree.
2371 */
2372 last_snap = btrfs_root_rtransid(&reloc_root->root_item);
2373 otransid = btrfs_root_otransid(&reloc_root->root_item);
2374 objectid = reloc_root->root_key.offset;
2375
2c536799 2376 ret = btrfs_drop_snapshot(reloc_root, rc->block_rsv, 0, 1);
aca1bba6
LB		 2377 if (ret < 0) {
2378 if (list_empty(&reloc_root->root_list))
2379 list_add_tail(&reloc_root->root_list,
2380 &reloc_roots);
2381 goto out;
5bc7247a
MX		 2382 } else if (!ret) {
2383 /*
2384 * recover the last snapshot tranid to avoid
2385 * the space balance break NOCOW.
2386 */
2387 root = read_fs_root(rc->extent_root->fs_info,
2388 objectid);
2389 if (IS_ERR(root))
2390 continue;
2391
5bc7247a
MX		 2392 trans = btrfs_join_transaction(root);
2393 BUG_ON(IS_ERR(trans));
2394
2395 /* Check if the fs/file tree was snapshoted or not. */
2396 if (btrfs_root_last_snapshot(&root->root_item) ==
2397 otransid - 1)
2398 btrfs_set_root_last_snapshot(&root->root_item,
2399 last_snap);
2400
2401 btrfs_end_transaction(trans, root);
aca1bba6 2402 }
5d4f98a2
YZ		 2403 }
2404
3fd0a558
YZ		 2405 if (found) {
2406 found = 0;
2407 goto again;
2408 }
aca1bba6
LB		 2409out:
2410 if (ret) {
2411 btrfs_std_error(root->fs_info, ret);
2412 if (!list_empty(&reloc_roots))
2413 free_reloc_roots(&reloc_roots);
467bb1d2
WS		 2414
2415 /* new reloc root may be added */
2416 mutex_lock(&root->fs_info->reloc_mutex);
2417 list_splice_init(&rc->reloc_roots, &reloc_roots);
2418 mutex_unlock(&root->fs_info->reloc_mutex);
2419 if (!list_empty(&reloc_roots))
2420 free_reloc_roots(&reloc_roots);
aca1bba6
LB		 2421 }
2422
5d4f98a2 2423 BUG_ON(!RB_EMPTY_ROOT(&rc->reloc_root_tree.rb_root));
aca1bba6 2424 return ret;
5d4f98a2
YZ		 2425}
2426
2427static void free_block_list(struct rb_root *blocks)
2428{
2429 struct tree_block *block;
2430 struct rb_node *rb_node;
2431 while ((rb_node = rb_first(blocks))) {
2432 block = rb_entry(rb_node, struct tree_block, rb_node);
2433 rb_erase(rb_node, blocks);
2434 kfree(block);
2435 }
2436}
2437
2438static int record_reloc_root_in_trans(struct btrfs_trans_handle *trans,
2439 struct btrfs_root *reloc_root)
2440{
2441 struct btrfs_root *root;
2442
2443 if (reloc_root->last_trans == trans->transid)
2444 return 0;
2445
2446 root = read_fs_root(reloc_root->fs_info, reloc_root->root_key.offset);
2447 BUG_ON(IS_ERR(root));
2448 BUG_ON(root->reloc_root != reloc_root);
2449
2450 return btrfs_record_root_in_trans(trans, root);
2451}
2452
3fd0a558
YZ		 2453static noinline_for_stack
2454struct btrfs_root *select_reloc_root(struct btrfs_trans_handle *trans,
2455 struct reloc_control *rc,
2456 struct backref_node *node,
2457 struct backref_edge *edges[], int *nr)
5d4f98a2
YZ		 2458{
2459 struct backref_node *next;
2460 struct btrfs_root *root;
3fd0a558
YZ		 2461 int index = 0;
2462
5d4f98a2
YZ		 2463 next = node;
2464 while (1) {
2465 cond_resched();
2466 next = walk_up_backref(next, edges, &index);
2467 root = next->root;
3fd0a558
YZ		 2468 BUG_ON(!root);
2469 BUG_ON(!root->ref_cows);
5d4f98a2
YZ		 2470
2471 if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID) {
2472 record_reloc_root_in_trans(trans, root);
2473 break;
2474 }
2475
3fd0a558
YZ		 2476 btrfs_record_root_in_trans(trans, root);
2477 root = root->reloc_root;
2478
2479 if (next->new_bytenr != root->node->start) {
2480 BUG_ON(next->new_bytenr);
2481 BUG_ON(!list_empty(&next->list));
2482 next->new_bytenr = root->node->start;
2483 next->root = root;
2484 list_add_tail(&next->list,
2485 &rc->backref_cache.changed);
2486 __mark_block_processed(rc, next);
5d4f98a2
YZ		 2487 break;
2488 }
2489
3fd0a558 2490 WARN_ON(1);
5d4f98a2
YZ		 2491 root = NULL;
2492 next = walk_down_backref(edges, &index);
2493 if (!next || next->level <= node->level)
2494 break;
2495 }
3fd0a558
YZ		 2496 if (!root)
2497 return NULL;
5d4f98a2 2498
3fd0a558
YZ		 2499 *nr = index;
2500 next = node;
2501 /* setup backref node path for btrfs_reloc_cow_block */
2502 while (1) {
2503 rc->backref_cache.path[next->level] = next;
2504 if (--index < 0)
2505 break;
2506 next = edges[index]->node[UPPER];
5d4f98a2 2507 }
5d4f98a2
YZ		 2508 return root;
2509}
2510
3fd0a558
YZ		 2511/*
2512 * select a tree root for relocation. return NULL if the block
2513 * is reference counted. we should use do_relocation() in this
2514 * case. return a tree root pointer if the block isn't reference
2515 * counted. return -ENOENT if the block is root of reloc tree.
2516 */
5d4f98a2
YZ		 2517static noinline_for_stack
2518struct btrfs_root *select_one_root(struct btrfs_trans_handle *trans,
2519 struct backref_node *node)
2520{
3fd0a558
YZ		 2521 struct backref_node *next;
2522 struct btrfs_root *root;
2523 struct btrfs_root *fs_root = NULL;
5d4f98a2 2524 struct backref_edge *edges[BTRFS_MAX_LEVEL - 1];
3fd0a558
YZ		 2525 int index = 0;
2526
2527 next = node;
2528 while (1) {
2529 cond_resched();
2530 next = walk_up_backref(next, edges, &index);
2531 root = next->root;
2532 BUG_ON(!root);
2533
25985edc 2534 /* no other choice for non-references counted tree */
3fd0a558
YZ		 2535 if (!root->ref_cows)
2536 return root;
2537
2538 if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID)
2539 fs_root = root;
2540
2541 if (next != node)
2542 return NULL;
2543
2544 next = walk_down_backref(edges, &index);
2545 if (!next || next->level <= node->level)
2546 break;
2547 }
2548
2549 if (!fs_root)
2550 return ERR_PTR(-ENOENT);
2551 return fs_root;
5d4f98a2
YZ		 2552}
2553
2554static noinline_for_stack
3fd0a558
YZ		 2555u64 calcu_metadata_size(struct reloc_control *rc,
2556 struct backref_node *node, int reserve)
5d4f98a2 2557{
3fd0a558
YZ		 2558 struct backref_node *next = node;
2559 struct backref_edge *edge;
2560 struct backref_edge *edges[BTRFS_MAX_LEVEL - 1];
2561 u64 num_bytes = 0;
2562 int index = 0;
2563
2564 BUG_ON(reserve && node->processed);
2565
2566 while (next) {
2567 cond_resched();
2568 while (1) {
2569 if (next->processed && (reserve || next != node))
2570 break;
2571
2572 num_bytes += btrfs_level_size(rc->extent_root,
2573 next->level);
2574
2575 if (list_empty(&next->upper))
2576 break;
2577
2578 edge = list_entry(next->upper.next,
2579 struct backref_edge, list[LOWER]);
2580 edges[index++] = edge;
2581 next = edge->node[UPPER];
2582 }
2583 next = walk_down_backref(edges, &index);
2584 }
2585 return num_bytes;
5d4f98a2
YZ		 2586}
2587
3fd0a558
YZ		 2588static int reserve_metadata_space(struct btrfs_trans_handle *trans,
2589 struct reloc_control *rc,
2590 struct backref_node *node)
5d4f98a2 2591{
3fd0a558
YZ		 2592 struct btrfs_root *root = rc->extent_root;
2593 u64 num_bytes;
2594 int ret;
0647bf56 2595 u64 tmp;
3fd0a558
YZ		 2596
2597 num_bytes = calcu_metadata_size(rc, node, 1) * 2;
5d4f98a2 2598
3fd0a558 2599 trans->block_rsv = rc->block_rsv;
0647bf56
WS		 2600 rc->reserved_bytes += num_bytes;
2601 ret = btrfs_block_rsv_refill(root, rc->block_rsv, num_bytes,
2602 BTRFS_RESERVE_FLUSH_ALL);
3fd0a558 2603 if (ret) {
0647bf56
WS		 2604 if (ret == -EAGAIN) {
2605 tmp = rc->extent_root->nodesize *
2606 RELOCATION_RESERVED_NODES;
2607 while (tmp <= rc->reserved_bytes)
2608 tmp <<= 1;
2609 /*
2610 * only one thread can access block_rsv at this point,
2611 * so we don't need hold lock to protect block_rsv.
2612 * we expand more reservation size here to allow enough
2613 * space for relocation and we will return eailer in
2614 * enospc case.
2615 */
2616 rc->block_rsv->size = tmp + rc->extent_root->nodesize *
2617 RELOCATION_RESERVED_NODES;
2618 }
3fd0a558 2619 return ret;
5d4f98a2 2620 }
3fd0a558 2621
3fd0a558
YZ		 2622 return 0;
2623}
2624
5d4f98a2
YZ		 2625/*
2626 * relocate a block tree, and then update pointers in upper level
2627 * blocks that reference the block to point to the new location.
2628 *
2629 * if called by link_to_upper, the block has already been relocated.
2630 * in that case this function just updates pointers.
2631 */
2632static int do_relocation(struct btrfs_trans_handle *trans,
3fd0a558 2633 struct reloc_control *rc,
5d4f98a2
YZ		 2634 struct backref_node *node,
2635 struct btrfs_key *key,
2636 struct btrfs_path *path, int lowest)
2637{
2638 struct backref_node *upper;
2639 struct backref_edge *edge;
2640 struct backref_edge *edges[BTRFS_MAX_LEVEL - 1];
2641 struct btrfs_root *root;
2642 struct extent_buffer *eb;
2643 u32 blocksize;
2644 u64 bytenr;
2645 u64 generation;
2646 int nr;
2647 int slot;
2648 int ret;
2649 int err = 0;
2650
2651 BUG_ON(lowest && node->eb);
2652
2653 path->lowest_level = node->level + 1;
3fd0a558 2654 rc->backref_cache.path[node->level] = node;
5d4f98a2
YZ		 2655 list_for_each_entry(edge, &node->upper, list[LOWER]) {
2656 cond_resched();
5d4f98a2
YZ		 2657
2658 upper = edge->node[UPPER];
3fd0a558
YZ		 2659 root = select_reloc_root(trans, rc, upper, edges, &nr);
2660 BUG_ON(!root);
2661
2662 if (upper->eb && !upper->locked) {
2663 if (!lowest) {
2664 ret = btrfs_bin_search(upper->eb, key,
2665 upper->level, &slot);
2666 BUG_ON(ret);
2667 bytenr = btrfs_node_blockptr(upper->eb, slot);
2668 if (node->eb->start == bytenr)
2669 goto next;
2670 }
5d4f98a2 2671 drop_node_buffer(upper);
3fd0a558 2672 }
5d4f98a2
YZ		 2673
2674 if (!upper->eb) {
2675 ret = btrfs_search_slot(trans, root, key, path, 0, 1);
2676 if (ret < 0) {
2677 err = ret;
2678 break;
2679 }
2680 BUG_ON(ret > 0);
2681
3fd0a558
YZ		 2682 if (!upper->eb) {
2683 upper->eb = path->nodes[upper->level];
2684 path->nodes[upper->level] = NULL;
2685 } else {
2686 BUG_ON(upper->eb != path->nodes[upper->level]);
2687 }
5d4f98a2 2688
3fd0a558
YZ		 2689 upper->locked = 1;
2690 path->locks[upper->level] = 0;
5d4f98a2 2691
3fd0a558 2692 slot = path->slots[upper->level];
b3b4aa74 2693 btrfs_release_path(path);
5d4f98a2
YZ		 2694 } else {
2695 ret = btrfs_bin_search(upper->eb, key, upper->level,
2696 &slot);
2697 BUG_ON(ret);
2698 }
2699
2700 bytenr = btrfs_node_blockptr(upper->eb, slot);
3fd0a558
YZ		 2701 if (lowest) {
2702 BUG_ON(bytenr != node->bytenr);
5d4f98a2 2703 } else {
3fd0a558
YZ		 2704 if (node->eb->start == bytenr)
2705 goto next;
5d4f98a2
YZ		 2706 }
2707
2708 blocksize = btrfs_level_size(root, node->level);
2709 generation = btrfs_node_ptr_generation(upper->eb, slot);
2710 eb = read_tree_block(root, bytenr, blocksize, generation);
416bc658
JB		 2711 if (!eb || !extent_buffer_uptodate(eb)) {
2712 free_extent_buffer(eb);
97d9a8a4
TI		 2713 err = -EIO;
2714 goto next;
2715 }
5d4f98a2
YZ		 2716 btrfs_tree_lock(eb);
2717 btrfs_set_lock_blocking(eb);
2718
2719 if (!node->eb) {
2720 ret = btrfs_cow_block(trans, root, eb, upper->eb,
2721 slot, &eb);
3fd0a558
YZ		 2722 btrfs_tree_unlock(eb);
2723 free_extent_buffer(eb);
5d4f98a2
YZ		 2724 if (ret < 0) {
2725 err = ret;
3fd0a558 2726 goto next;
5d4f98a2 2727 }
3fd0a558 2728 BUG_ON(node->eb != eb);
5d4f98a2
YZ		 2729 } else {
2730 btrfs_set_node_blockptr(upper->eb, slot,
2731 node->eb->start);
2732 btrfs_set_node_ptr_generation(upper->eb, slot,
2733 trans->transid);
2734 btrfs_mark_buffer_dirty(upper->eb);
2735
2736 ret = btrfs_inc_extent_ref(trans, root,
2737 node->eb->start, blocksize,
2738 upper->eb->start,
2739 btrfs_header_owner(upper->eb),
66d7e7f0 2740 node->level, 0, 1);
5d4f98a2
YZ		 2741 BUG_ON(ret);
2742
2743 ret = btrfs_drop_subtree(trans, root, eb, upper->eb);
2744 BUG_ON(ret);
5d4f98a2 2745 }
3fd0a558
YZ		 2746next:
2747 if (!upper->pending)
2748 drop_node_buffer(upper);
2749 else
2750 unlock_node_buffer(upper);
2751 if (err)
2752 break;
5d4f98a2 2753 }
3fd0a558
YZ		 2754
2755 if (!err && node->pending) {
2756 drop_node_buffer(node);
2757 list_move_tail(&node->list, &rc->backref_cache.changed);
2758 node->pending = 0;
2759 }
2760
5d4f98a2 2761 path->lowest_level = 0;
3fd0a558 2762 BUG_ON(err == -ENOSPC);
5d4f98a2
YZ		 2763 return err;
2764}
2765
2766static int link_to_upper(struct btrfs_trans_handle *trans,
3fd0a558 2767 struct reloc_control *rc,
5d4f98a2
YZ		 2768 struct backref_node *node,
2769 struct btrfs_path *path)
2770{
2771 struct btrfs_key key;
5d4f98a2
YZ		 2772
2773 btrfs_node_key_to_cpu(node->eb, &key, 0);
3fd0a558 2774 return do_relocation(trans, rc, node, &key, path, 0);
5d4f98a2
YZ		 2775}
2776
2777static int finish_pending_nodes(struct btrfs_trans_handle *trans,
3fd0a558
YZ		 2778 struct reloc_control *rc,
2779 struct btrfs_path *path, int err)
5d4f98a2 2780{
3fd0a558
YZ		 2781 LIST_HEAD(list);
2782 struct backref_cache *cache = &rc->backref_cache;
5d4f98a2
YZ		 2783 struct backref_node *node;
2784 int level;
2785 int ret;
5d4f98a2
YZ		 2786
2787 for (level = 0; level < BTRFS_MAX_LEVEL; level++) {
2788 while (!list_empty(&cache->pending[level])) {
2789 node = list_entry(cache->pending[level].next,
3fd0a558
YZ		 2790 struct backref_node, list);
2791 list_move_tail(&node->list, &list);
2792 BUG_ON(!node->pending);
5d4f98a2 2793
3fd0a558
YZ		 2794 if (!err) {
2795 ret = link_to_upper(trans, rc, node, path);
2796 if (ret < 0)
2797 err = ret;
2798 }
5d4f98a2 2799 }
3fd0a558 2800 list_splice_init(&list, &cache->pending[level]);
5d4f98a2 2801 }
5d4f98a2
YZ		 2802 return err;
2803}
2804
2805static void mark_block_processed(struct reloc_control *rc,
3fd0a558
YZ		 2806 u64 bytenr, u32 blocksize)
2807{
2808 set_extent_bits(&rc->processed_blocks, bytenr, bytenr + blocksize - 1,
2809 EXTENT_DIRTY, GFP_NOFS);
2810}
2811
2812static void __mark_block_processed(struct reloc_control *rc,
2813 struct backref_node *node)
5d4f98a2
YZ		 2814{
2815 u32 blocksize;
2816 if (node->level == 0 ||
2817 in_block_group(node->bytenr, rc->block_group)) {
2818 blocksize = btrfs_level_size(rc->extent_root, node->level);
3fd0a558 2819 mark_block_processed(rc, node->bytenr, blocksize);
5d4f98a2
YZ		 2820 }
2821 node->processed = 1;
2822}
2823
2824/*
2825 * mark a block and all blocks directly/indirectly reference the block
2826 * as processed.
2827 */
2828static void update_processed_blocks(struct reloc_control *rc,
2829 struct backref_node *node)
2830{
2831 struct backref_node *next = node;
2832 struct backref_edge *edge;
2833 struct backref_edge *edges[BTRFS_MAX_LEVEL - 1];
2834 int index = 0;
2835
2836 while (next) {
2837 cond_resched();
2838 while (1) {
2839 if (next->processed)
2840 break;
2841
3fd0a558 2842 __mark_block_processed(rc, next);
5d4f98a2
YZ		 2843
2844 if (list_empty(&next->upper))
2845 break;
2846
2847 edge = list_entry(next->upper.next,
2848 struct backref_edge, list[LOWER]);
2849 edges[index++] = edge;
2850 next = edge->node[UPPER];
2851 }
2852 next = walk_down_backref(edges, &index);
2853 }
2854}
2855
3fd0a558
YZ		 2856static int tree_block_processed(u64 bytenr, u32 blocksize,
2857 struct reloc_control *rc)
2858{
2859 if (test_range_bit(&rc->processed_blocks, bytenr,
2860 bytenr + blocksize - 1, EXTENT_DIRTY, 1, NULL))
2861 return 1;
2862 return 0;
5d4f98a2
YZ		 2863}
2864
2865static int get_tree_block_key(struct reloc_control *rc,
2866 struct tree_block *block)
2867{
2868 struct extent_buffer *eb;
2869
2870 BUG_ON(block->key_ready);
2871 eb = read_tree_block(rc->extent_root, block->bytenr,
2872 block->key.objectid, block->key.offset);
416bc658
JB		 2873 if (!eb || !extent_buffer_uptodate(eb)) {
2874 free_extent_buffer(eb);
2875 return -EIO;
2876 }
5d4f98a2
YZ		 2877 WARN_ON(btrfs_header_level(eb) != block->level);
2878 if (block->level == 0)
2879 btrfs_item_key_to_cpu(eb, &block->key, 0);
2880 else
2881 btrfs_node_key_to_cpu(eb, &block->key, 0);
2882 free_extent_buffer(eb);
2883 block->key_ready = 1;
2884 return 0;
2885}
2886
2887static int reada_tree_block(struct reloc_control *rc,
2888 struct tree_block *block)
2889{
2890 BUG_ON(block->key_ready);
3173a18f
JB		 2891 if (block->key.type == BTRFS_METADATA_ITEM_KEY)
2892 readahead_tree_block(rc->extent_root, block->bytenr,
2893 block->key.objectid,
2894 rc->extent_root->leafsize);
2895 else
2896 readahead_tree_block(rc->extent_root, block->bytenr,
2897 block->key.objectid, block->key.offset);
5d4f98a2
YZ		 2898 return 0;
2899}
2900
2901/*
2902 * helper function to relocate a tree block
2903 */
2904static int relocate_tree_block(struct btrfs_trans_handle *trans,
2905 struct reloc_control *rc,
2906 struct backref_node *node,
2907 struct btrfs_key *key,
2908 struct btrfs_path *path)
2909{
2910 struct btrfs_root *root;
3fd0a558
YZ		 2911 int ret = 0;
2912
2913 if (!node)
2914 return 0;
5d4f98a2 2915
3fd0a558 2916 BUG_ON(node->processed);
5d4f98a2 2917 root = select_one_root(trans, node);
3fd0a558 2918 if (root == ERR_PTR(-ENOENT)) {
5d4f98a2 2919 update_processed_blocks(rc, node);
3fd0a558 2920 goto out;
5d4f98a2
YZ		 2921 }
2922
3fd0a558
YZ		 2923 if (!root || root->ref_cows) {
2924 ret = reserve_metadata_space(trans, rc, node);
2925 if (ret)
5d4f98a2 2926 goto out;
5d4f98a2
YZ		 2927 }
2928
3fd0a558
YZ		 2929 if (root) {
2930 if (root->ref_cows) {
2931 BUG_ON(node->new_bytenr);
2932 BUG_ON(!list_empty(&node->list));
2933 btrfs_record_root_in_trans(trans, root);
2934 root = root->reloc_root;
2935 node->new_bytenr = root->node->start;
2936 node->root = root;
2937 list_add_tail(&node->list, &rc->backref_cache.changed);
2938 } else {
2939 path->lowest_level = node->level;
2940 ret = btrfs_search_slot(trans, root, key, path, 0, 1);
b3b4aa74 2941 btrfs_release_path(path);
3fd0a558
YZ		 2942 if (ret > 0)
2943 ret = 0;
2944 }
2945 if (!ret)
2946 update_processed_blocks(rc, node);
2947 } else {
2948 ret = do_relocation(trans, rc, node, key, path, 1);
2949 }
5d4f98a2 2950out:
0647bf56 2951 if (ret || node->level == 0 || node->cowonly)
3fd0a558 2952 remove_backref_node(&rc->backref_cache, node);
5d4f98a2
YZ		 2953 return ret;
2954}
2955
2956/*
2957 * relocate a list of blocks
2958 */
2959static noinline_for_stack
2960int relocate_tree_blocks(struct btrfs_trans_handle *trans,
2961 struct reloc_control *rc, struct rb_root *blocks)
2962{
5d4f98a2
YZ		 2963 struct backref_node *node;
2964 struct btrfs_path *path;
2965 struct tree_block *block;
2966 struct rb_node *rb_node;
5d4f98a2
YZ		 2967 int ret;
2968 int err = 0;
2969
2970 path = btrfs_alloc_path();
e1a12670
LB		 2971 if (!path) {
2972 err = -ENOMEM;
34c2b290 2973 goto out_free_blocks;
e1a12670 2974 }
5d4f98a2 2975
5d4f98a2
YZ		 2976 rb_node = rb_first(blocks);
2977 while (rb_node) {
2978 block = rb_entry(rb_node, struct tree_block, rb_node);
5d4f98a2
YZ		 2979 if (!block->key_ready)
2980 reada_tree_block(rc, block);
2981 rb_node = rb_next(rb_node);
2982 }
2983
2984 rb_node = rb_first(blocks);
2985 while (rb_node) {
2986 block = rb_entry(rb_node, struct tree_block, rb_node);
34c2b290
DS		 2987 if (!block->key_ready) {
2988 err = get_tree_block_key(rc, block);
2989 if (err)
2990 goto out_free_path;
2991 }
5d4f98a2
YZ		 2992 rb_node = rb_next(rb_node);
2993 }
2994
2995 rb_node = rb_first(blocks);
2996 while (rb_node) {
2997 block = rb_entry(rb_node, struct tree_block, rb_node);
2998
3fd0a558 2999 node = build_backref_tree(rc, &block->key,
5d4f98a2
YZ		 3000 block->level, block->bytenr);
3001 if (IS_ERR(node)) {
3002 err = PTR_ERR(node);
3003 goto out;
3004 }
3005
3006 ret = relocate_tree_block(trans, rc, node, &block->key,
3007 path);
3008 if (ret < 0) {
3fd0a558
YZ		 3009 if (ret != -EAGAIN || rb_node == rb_first(blocks))
3010 err = ret;
5d4f98a2
YZ		 3011 goto out;
3012 }
5d4f98a2
YZ		 3013 rb_node = rb_next(rb_node);
3014 }
5d4f98a2 3015out:
3fd0a558 3016 err = finish_pending_nodes(trans, rc, path, err);
5d4f98a2 3017
34c2b290 3018out_free_path:
5d4f98a2 3019 btrfs_free_path(path);
34c2b290 3020out_free_blocks:
e1a12670 3021 free_block_list(blocks);
5d4f98a2
YZ		 3022 return err;
3023}
3024
efa56464
YZ		 3025static noinline_for_stack
3026int prealloc_file_extent_cluster(struct inode *inode,
3027 struct file_extent_cluster *cluster)
3028{
3029 u64 alloc_hint = 0;
3030 u64 start;
3031 u64 end;
3032 u64 offset = BTRFS_I(inode)->index_cnt;
3033 u64 num_bytes;
3034 int nr = 0;
3035 int ret = 0;
3036
3037 BUG_ON(cluster->start != cluster->boundary[0]);
3038 mutex_lock(&inode->i_mutex);
3039
3040 ret = btrfs_check_data_free_space(inode, cluster->end +
3041 1 - cluster->start);
3042 if (ret)
3043 goto out;
3044
3045 while (nr < cluster->nr) {
3046 start = cluster->boundary[nr] - offset;
3047 if (nr + 1 < cluster->nr)
3048 end = cluster->boundary[nr + 1] - 1 - offset;
3049 else
3050 end = cluster->end - offset;
3051
d0082371 3052 lock_extent(&BTRFS_I(inode)->io_tree, start, end);
efa56464
YZ		 3053 num_bytes = end + 1 - start;
3054 ret = btrfs_prealloc_file_range(inode, 0, start,
3055 num_bytes, num_bytes,
3056 end + 1, &alloc_hint);
d0082371 3057 unlock_extent(&BTRFS_I(inode)->io_tree, start, end);
efa56464
YZ		 3058 if (ret)
3059 break;
3060 nr++;
3061 }
3062 btrfs_free_reserved_data_space(inode, cluster->end +
3063 1 - cluster->start);
3064out:
3065 mutex_unlock(&inode->i_mutex);
3066 return ret;
3067}
3068
5d4f98a2 3069static noinline_for_stack
0257bb82
YZ		 3070int setup_extent_mapping(struct inode *inode, u64 start, u64 end,
3071 u64 block_start)
3072{
3073 struct btrfs_root *root = BTRFS_I(inode)->root;
3074 struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
3075 struct extent_map *em;
3076 int ret = 0;
3077
172ddd60 3078 em = alloc_extent_map();
0257bb82
YZ		 3079 if (!em)
3080 return -ENOMEM;
3081
3082 em->start = start;
3083 em->len = end + 1 - start;
3084 em->block_len = em->len;
3085 em->block_start = block_start;
3086 em->bdev = root->fs_info->fs_devices->latest_bdev;
3087 set_bit(EXTENT_FLAG_PINNED, &em->flags);
3088
d0082371 3089 lock_extent(&BTRFS_I(inode)->io_tree, start, end);
0257bb82
YZ		 3090 while (1) {
3091 write_lock(&em_tree->lock);
09a2a8f9 3092 ret = add_extent_mapping(em_tree, em, 0);
0257bb82
YZ		 3093 write_unlock(&em_tree->lock);
3094 if (ret != -EEXIST) {
3095 free_extent_map(em);
3096 break;
3097 }
3098 btrfs_drop_extent_cache(inode, start, end, 0);
3099 }
d0082371 3100 unlock_extent(&BTRFS_I(inode)->io_tree, start, end);
0257bb82
YZ		 3101 return ret;
3102}
3103
3104static int relocate_file_extent_cluster(struct inode *inode,
3105 struct file_extent_cluster *cluster)
5d4f98a2
YZ		 3106{
3107 u64 page_start;
3108 u64 page_end;
0257bb82
YZ		 3109 u64 offset = BTRFS_I(inode)->index_cnt;
3110 unsigned long index;
5d4f98a2 3111 unsigned long last_index;
5d4f98a2
YZ		 3112 struct page *page;
3113 struct file_ra_state *ra;
3b16a4e3 3114 gfp_t mask = btrfs_alloc_write_mask(inode->i_mapping);
0257bb82 3115 int nr = 0;
5d4f98a2
YZ		 3116 int ret = 0;
3117
0257bb82
YZ		 3118 if (!cluster->nr)
3119 return 0;
3120
5d4f98a2
YZ		 3121 ra = kzalloc(sizeof(*ra), GFP_NOFS);
3122 if (!ra)
3123 return -ENOMEM;
3124
efa56464
YZ		 3125 ret = prealloc_file_extent_cluster(inode, cluster);
3126 if (ret)
3127 goto out;
0257bb82 3128
efa56464 3129 file_ra_state_init(ra, inode->i_mapping);
5d4f98a2 3130
0257bb82
YZ		 3131 ret = setup_extent_mapping(inode, cluster->start - offset,
3132 cluster->end - offset, cluster->start);
5d4f98a2 3133 if (ret)
efa56464 3134 goto out;
5d4f98a2 3135
efa56464
YZ		 3136 index = (cluster->start - offset) >> PAGE_CACHE_SHIFT;
3137 last_index = (cluster->end - offset) >> PAGE_CACHE_SHIFT;
0257bb82 3138 while (index <= last_index) {
efa56464
YZ		 3139 ret = btrfs_delalloc_reserve_metadata(inode, PAGE_CACHE_SIZE);
3140 if (ret)
3141 goto out;
3142
0257bb82 3143 page = find_lock_page(inode->i_mapping, index);
5d4f98a2 3144 if (!page) {
0257bb82
YZ		 3145 page_cache_sync_readahead(inode->i_mapping,
3146 ra, NULL, index,
3147 last_index + 1 - index);
a94733d0 3148 page = find_or_create_page(inode->i_mapping, index,
3b16a4e3 3149 mask);
0257bb82 3150 if (!page) {
efa56464
YZ		 3151 btrfs_delalloc_release_metadata(inode,
3152 PAGE_CACHE_SIZE);
0257bb82 3153 ret = -ENOMEM;
efa56464 3154 goto out;
0257bb82 3155 }
5d4f98a2 3156 }
0257bb82
YZ		 3157
3158 if (PageReadahead(page)) {
3159 page_cache_async_readahead(inode->i_mapping,
3160 ra, NULL, page, index,
3161 last_index + 1 - index);
3162 }
3163
5d4f98a2
YZ		 3164 if (!PageUptodate(page)) {
3165 btrfs_readpage(NULL, page);
3166 lock_page(page);
3167 if (!PageUptodate(page)) {
3168 unlock_page(page);
3169 page_cache_release(page);
efa56464
YZ		 3170 btrfs_delalloc_release_metadata(inode,
3171 PAGE_CACHE_SIZE);
5d4f98a2 3172 ret = -EIO;
efa56464 3173 goto out;
5d4f98a2
YZ		 3174 }
3175 }
5d4f98a2 3176
4eee4fa4 3177 page_start = page_offset(page);
5d4f98a2 3178 page_end = page_start + PAGE_CACHE_SIZE - 1;
0257bb82 3179
d0082371 3180 lock_extent(&BTRFS_I(inode)->io_tree, page_start, page_end);
0257bb82 3181
5d4f98a2
YZ		 3182 set_page_extent_mapped(page);
3183
0257bb82
YZ		 3184 if (nr < cluster->nr &&
3185 page_start + offset == cluster->boundary[nr]) {
3186 set_extent_bits(&BTRFS_I(inode)->io_tree,
3187 page_start, page_end,
5d4f98a2 3188 EXTENT_BOUNDARY, GFP_NOFS);
0257bb82
YZ		 3189 nr++;
3190 }
5d4f98a2 3191
efa56464 3192 btrfs_set_extent_delalloc(inode, page_start, page_end, NULL);
5d4f98a2 3193 set_page_dirty(page);
5d4f98a2 3194
0257bb82 3195 unlock_extent(&BTRFS_I(inode)->io_tree,
d0082371 3196 page_start, page_end);
5d4f98a2
YZ		 3197 unlock_page(page);
3198 page_cache_release(page);
0257bb82
YZ		 3199
3200 index++;
efa56464
YZ		 3201 balance_dirty_pages_ratelimited(inode->i_mapping);
3202 btrfs_throttle(BTRFS_I(inode)->root);
5d4f98a2 3203 }
0257bb82 3204 WARN_ON(nr != cluster->nr);
efa56464 3205out:
5d4f98a2 3206 kfree(ra);
5d4f98a2
YZ		 3207 return ret;
3208}
3209
3210static noinline_for_stack
0257bb82
YZ		 3211int relocate_data_extent(struct inode *inode, struct btrfs_key *extent_key,
3212 struct file_extent_cluster *cluster)
5d4f98a2 3213{
0257bb82 3214 int ret;
5d4f98a2 3215
0257bb82
YZ		 3216 if (cluster->nr > 0 && extent_key->objectid != cluster->end + 1) {
3217 ret = relocate_file_extent_cluster(inode, cluster);
3218 if (ret)
3219 return ret;
3220 cluster->nr = 0;
5d4f98a2 3221 }
5d4f98a2 3222
0257bb82
YZ		 3223 if (!cluster->nr)
3224 cluster->start = extent_key->objectid;
3225 else
3226 BUG_ON(cluster->nr >= MAX_EXTENTS);
3227 cluster->end = extent_key->objectid + extent_key->offset - 1;
3228 cluster->boundary[cluster->nr] = extent_key->objectid;
3229 cluster->nr++;
3230
3231 if (cluster->nr >= MAX_EXTENTS) {
3232 ret = relocate_file_extent_cluster(inode, cluster);
3233 if (ret)
3234 return ret;
3235 cluster->nr = 0;
3236 }
3237 return 0;
5d4f98a2
YZ		 3238}
3239
3240#ifdef BTRFS_COMPAT_EXTENT_TREE_V0
3241static int get_ref_objectid_v0(struct reloc_control *rc,
3242 struct btrfs_path *path,
3243 struct btrfs_key *extent_key,
3244 u64 *ref_objectid, int *path_change)
3245{
3246 struct btrfs_key key;
3247 struct extent_buffer *leaf;
3248 struct btrfs_extent_ref_v0 *ref0;
3249 int ret;
3250 int slot;
3251
3252 leaf = path->nodes[0];
3253 slot = path->slots[0];
3254 while (1) {
3255 if (slot >= btrfs_header_nritems(leaf)) {
3256 ret = btrfs_next_leaf(rc->extent_root, path);
3257 if (ret < 0)
3258 return ret;
3259 BUG_ON(ret > 0);
3260 leaf = path->nodes[0];
3261 slot = path->slots[0];
3262 if (path_change)
3263 *path_change = 1;
3264 }
3265 btrfs_item_key_to_cpu(leaf, &key, slot);
3266 if (key.objectid != extent_key->objectid)
3267 return -ENOENT;
3268
3269 if (key.type != BTRFS_EXTENT_REF_V0_KEY) {
3270 slot++;
3271 continue;
3272 }
3273 ref0 = btrfs_item_ptr(leaf, slot,
3274 struct btrfs_extent_ref_v0);
3275 *ref_objectid = btrfs_ref_objectid_v0(leaf, ref0);
3276 break;
3277 }
3278 return 0;
3279}
3280#endif
3281
3282/*
3283 * helper to add a tree block to the list.
3284 * the major work is getting the generation and level of the block
3285 */
3286static int add_tree_block(struct reloc_control *rc,
3287 struct btrfs_key *extent_key,
3288 struct btrfs_path *path,
3289 struct rb_root *blocks)
3290{
3291 struct extent_buffer *eb;
3292 struct btrfs_extent_item *ei;
3293 struct btrfs_tree_block_info *bi;
3294 struct tree_block *block;
3295 struct rb_node *rb_node;
3296 u32 item_size;
3297 int level = -1;
7fdf4b60 3298 u64 generation;
5d4f98a2
YZ		 3299
3300 eb = path->nodes[0];
3301 item_size = btrfs_item_size_nr(eb, path->slots[0]);
3302
3173a18f
JB		 3303 if (extent_key->type == BTRFS_METADATA_ITEM_KEY ||
3304 item_size >= sizeof(*ei) + sizeof(*bi)) {
5d4f98a2
YZ		 3305 ei = btrfs_item_ptr(eb, path->slots[0],
3306 struct btrfs_extent_item);
3173a18f
JB		 3307 if (extent_key->type == BTRFS_EXTENT_ITEM_KEY) {
3308 bi = (struct btrfs_tree_block_info *)(ei + 1);
3309 level = btrfs_tree_block_level(eb, bi);
3310 } else {
3311 level = (int)extent_key->offset;
3312 }
5d4f98a2 3313 generation = btrfs_extent_generation(eb, ei);
5d4f98a2
YZ		 3314 } else {
3315#ifdef BTRFS_COMPAT_EXTENT_TREE_V0
3316 u64 ref_owner;
3317 int ret;
3318
3319 BUG_ON(item_size != sizeof(struct btrfs_extent_item_v0));
3320 ret = get_ref_objectid_v0(rc, path, extent_key,
3321 &ref_owner, NULL);
411fc6bc
AK		 3322 if (ret < 0)
3323 return ret;
5d4f98a2
YZ		 3324 BUG_ON(ref_owner >= BTRFS_MAX_LEVEL);
3325 level = (int)ref_owner;
3326 /* FIXME: get real generation */
3327 generation = 0;
3328#else
3329 BUG();
3330#endif
3331 }
3332
b3b4aa74 3333 btrfs_release_path(path);
5d4f98a2
YZ		 3334
3335 BUG_ON(level == -1);
3336
3337 block = kmalloc(sizeof(*block), GFP_NOFS);
3338 if (!block)
3339 return -ENOMEM;
3340
3341 block->bytenr = extent_key->objectid;
3173a18f 3342 block->key.objectid = rc->extent_root->leafsize;
5d4f98a2
YZ		 3343 block->key.offset = generation;
3344 block->level = level;
3345 block->key_ready = 0;
3346
3347 rb_node = tree_insert(blocks, block->bytenr, &block->rb_node);
43c04fb1
JM		 3348 if (rb_node)
3349 backref_tree_panic(rb_node, -EEXIST, block->bytenr);
5d4f98a2
YZ		 3350
3351 return 0;
3352}
3353
3354/*
3355 * helper to add tree blocks for backref of type BTRFS_SHARED_DATA_REF_KEY
3356 */
3357static int __add_tree_block(struct reloc_control *rc,
3358 u64 bytenr, u32 blocksize,
3359 struct rb_root *blocks)
3360{
3361 struct btrfs_path *path;
3362 struct btrfs_key key;
3363 int ret;
aee68ee5
JB		 3364 bool skinny = btrfs_fs_incompat(rc->extent_root->fs_info,
3365 SKINNY_METADATA);
5d4f98a2
YZ		 3366
3367 if (tree_block_processed(bytenr, blocksize, rc))
3368 return 0;
3369
3370 if (tree_search(blocks, bytenr))
3371 return 0;
3372
3373 path = btrfs_alloc_path();
3374 if (!path)
3375 return -ENOMEM;
aee68ee5 3376again:
5d4f98a2 3377 key.objectid = bytenr;
aee68ee5
JB		 3378 if (skinny) {
3379 key.type = BTRFS_METADATA_ITEM_KEY;
3380 key.offset = (u64)-1;
3381 } else {
3382 key.type = BTRFS_EXTENT_ITEM_KEY;
3383 key.offset = blocksize;
3384 }
5d4f98a2
YZ		 3385
3386 path->search_commit_root = 1;
3387 path->skip_locking = 1;
3388 ret = btrfs_search_slot(NULL, rc->extent_root, &key, path, 0, 0);
3389 if (ret < 0)
3390 goto out;
5d4f98a2 3391
aee68ee5
JB		 3392 if (ret > 0 && skinny) {
3393 if (path->slots[0]) {
3394 path->slots[0]--;
3395 btrfs_item_key_to_cpu(path->nodes[0], &key,
3396 path->slots[0]);
3397 if (key.objectid == bytenr &&
3398 (key.type == BTRFS_METADATA_ITEM_KEY ||
3399 (key.type == BTRFS_EXTENT_ITEM_KEY &&
3400 key.offset == blocksize)))
3401 ret = 0;
3402 }
3403
3404 if (ret) {
3405 skinny = false;
3406 btrfs_release_path(path);
3407 goto again;
3408 }
3173a18f
JB		 3409 }
3410 BUG_ON(ret);
3411
5d4f98a2
YZ		 3412 ret = add_tree_block(rc, &key, path, blocks);
3413out:
3414 btrfs_free_path(path);
3415 return ret;
3416}
3417
3418/*
3419 * helper to check if the block use full backrefs for pointers in it
3420 */
3421static int block_use_full_backref(struct reloc_control *rc,
3422 struct extent_buffer *eb)
3423{
5d4f98a2
YZ		 3424 u64 flags;
3425 int ret;
3426
3427 if (btrfs_header_flag(eb, BTRFS_HEADER_FLAG_RELOC) ||
3428 btrfs_header_backref_rev(eb) < BTRFS_MIXED_BACKREF_REV)
3429 return 1;
3430
3fd0a558 3431 ret = btrfs_lookup_extent_info(NULL, rc->extent_root,
3173a18f
JB		 3432 eb->start, btrfs_header_level(eb), 1,
3433 NULL, &flags);
5d4f98a2
YZ		 3434 BUG_ON(ret);
3435
5d4f98a2
YZ		 3436 if (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF)
3437 ret = 1;
3438 else
3439 ret = 0;
5d4f98a2
YZ		 3440 return ret;
3441}
3442
0af3d00b
JB		 3443static int delete_block_group_cache(struct btrfs_fs_info *fs_info,
3444 struct inode *inode, u64 ino)
3445{
3446 struct btrfs_key key;
0af3d00b
JB		 3447 struct btrfs_root *root = fs_info->tree_root;
3448 struct btrfs_trans_handle *trans;
0af3d00b
JB		 3449 int ret = 0;
3450
3451 if (inode)
3452 goto truncate;
3453
3454 key.objectid = ino;
3455 key.type = BTRFS_INODE_ITEM_KEY;
3456 key.offset = 0;
3457
3458 inode = btrfs_iget(fs_info->sb, &key, root, NULL);
f54fb859
TI		 3459 if (IS_ERR(inode) || is_bad_inode(inode)) {
3460 if (!IS_ERR(inode))
0af3d00b
JB		 3461 iput(inode);
3462 return -ENOENT;
3463 }
3464
3465truncate:
7b61cd92
MX		 3466 ret = btrfs_check_trunc_cache_free_space(root,
3467 &fs_info->global_block_rsv);
3468 if (ret)
3469 goto out;
3470
7a7eaa40 3471 trans = btrfs_join_transaction(root);
0af3d00b 3472 if (IS_ERR(trans)) {
3612b495 3473 ret = PTR_ERR(trans);
0af3d00b
JB		 3474 goto out;
3475 }
3476
74514323 3477 ret = btrfs_truncate_free_space_cache(root, trans, inode);
0af3d00b 3478
0af3d00b 3479 btrfs_end_transaction(trans, root);
b53d3f5d 3480 btrfs_btree_balance_dirty(root);
0af3d00b
JB		 3481out:
3482 iput(inode);
3483 return ret;
3484}
3485
5d4f98a2
YZ		 3486/*
3487 * helper to add tree blocks for backref of type BTRFS_EXTENT_DATA_REF_KEY
3488 * this function scans fs tree to find blocks reference the data extent
3489 */
3490static int find_data_references(struct reloc_control *rc,
3491 struct btrfs_key *extent_key,
3492 struct extent_buffer *leaf,
3493 struct btrfs_extent_data_ref *ref,
3494 struct rb_root *blocks)
3495{
3496 struct btrfs_path *path;
3497 struct tree_block *block;
3498 struct btrfs_root *root;
3499 struct btrfs_file_extent_item *fi;
3500 struct rb_node *rb_node;
3501 struct btrfs_key key;
3502 u64 ref_root;
3503 u64 ref_objectid;
3504 u64 ref_offset;
3505 u32 ref_count;
3506 u32 nritems;
3507 int err = 0;
3508 int added = 0;
3509 int counted;
3510 int ret;
3511
5d4f98a2
YZ		 3512 ref_root = btrfs_extent_data_ref_root(leaf, ref);
3513 ref_objectid = btrfs_extent_data_ref_objectid(leaf, ref);
3514 ref_offset = btrfs_extent_data_ref_offset(leaf, ref);
3515 ref_count = btrfs_extent_data_ref_count(leaf, ref);
3516
0af3d00b
JB		 3517 /*
3518 * This is an extent belonging to the free space cache, lets just delete
3519 * it and redo the search.
3520 */
3521 if (ref_root == BTRFS_ROOT_TREE_OBJECTID) {
3522 ret = delete_block_group_cache(rc->extent_root->fs_info,
3523 NULL, ref_objectid);
3524 if (ret != -ENOENT)
3525 return ret;
3526 ret = 0;
3527 }
3528
3529 path = btrfs_alloc_path();
3530 if (!path)
3531 return -ENOMEM;
026fd317 3532 path->reada = 1;
0af3d00b 3533
5d4f98a2
YZ		 3534 root = read_fs_root(rc->extent_root->fs_info, ref_root);
3535 if (IS_ERR(root)) {
3536 err = PTR_ERR(root);
3537 goto out;
3538 }
3539
3540 key.objectid = ref_objectid;
5d4f98a2 3541 key.type = BTRFS_EXTENT_DATA_KEY;
84850e8d
YZ		 3542 if (ref_offset > ((u64)-1 << 32))
3543 key.offset = 0;
3544 else
3545 key.offset = ref_offset;
5d4f98a2
YZ		 3546
3547 path->search_commit_root = 1;
3548 path->skip_locking = 1;
3549 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
3550 if (ret < 0) {
3551 err = ret;
3552 goto out;
3553 }
3554
3555 leaf = path->nodes[0];
3556 nritems = btrfs_header_nritems(leaf);
3557 /*
3558 * the references in tree blocks that use full backrefs
3559 * are not counted in
3560 */
3561 if (block_use_full_backref(rc, leaf))
3562 counted = 0;
3563 else
3564 counted = 1;
3565 rb_node = tree_search(blocks, leaf->start);
3566 if (rb_node) {
3567 if (counted)
3568 added = 1;
3569 else
3570 path->slots[0] = nritems;
3571 }
3572
3573 while (ref_count > 0) {
3574 while (path->slots[0] >= nritems) {
3575 ret = btrfs_next_leaf(root, path);
3576 if (ret < 0) {
3577 err = ret;
3578 goto out;
3579 }
fae7f21c 3580 if (WARN_ON(ret > 0))
5d4f98a2 3581 goto out;
5d4f98a2
YZ		 3582
3583 leaf = path->nodes[0];
3584 nritems = btrfs_header_nritems(leaf);
3585 added = 0;
3586
3587 if (block_use_full_backref(rc, leaf))
3588 counted = 0;
3589 else
3590 counted = 1;
3591 rb_node = tree_search(blocks, leaf->start);
3592 if (rb_node) {
3593 if (counted)
3594 added = 1;
3595 else
3596 path->slots[0] = nritems;
3597 }
3598 }
3599
3600 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
fae7f21c
DG		 3601 if (WARN_ON(key.objectid != ref_objectid ||
3602 key.type != BTRFS_EXTENT_DATA_KEY))
5d4f98a2 3603 break;
5d4f98a2
YZ		 3604
3605 fi = btrfs_item_ptr(leaf, path->slots[0],
3606 struct btrfs_file_extent_item);
3607
3608 if (btrfs_file_extent_type(leaf, fi) ==
3609 BTRFS_FILE_EXTENT_INLINE)
3610 goto next;
3611
3612 if (btrfs_file_extent_disk_bytenr(leaf, fi) !=
3613 extent_key->objectid)
3614 goto next;
3615
3616 key.offset -= btrfs_file_extent_offset(leaf, fi);
3617 if (key.offset != ref_offset)
3618 goto next;
3619
3620 if (counted)
3621 ref_count--;
3622 if (added)
3623 goto next;
3624
3625 if (!tree_block_processed(leaf->start, leaf->len, rc)) {
3626 block = kmalloc(sizeof(*block), GFP_NOFS);
3627 if (!block) {
3628 err = -ENOMEM;
3629 break;
3630 }
3631 block->bytenr = leaf->start;
3632 btrfs_item_key_to_cpu(leaf, &block->key, 0);
3633 block->level = 0;
3634 block->key_ready = 1;
3635 rb_node = tree_insert(blocks, block->bytenr,
3636 &block->rb_node);
43c04fb1
JM		 3637 if (rb_node)
3638 backref_tree_panic(rb_node, -EEXIST,
3639 block->bytenr);
5d4f98a2
YZ		 3640 }
3641 if (counted)
3642 added = 1;
3643 else
3644 path->slots[0] = nritems;
3645next:
3646 path->slots[0]++;
3647
3648 }
3649out:
3650 btrfs_free_path(path);
3651 return err;
3652}
3653
3654/*
2c016dc2 3655 * helper to find all tree blocks that reference a given data extent
5d4f98a2
YZ		 3656 */
3657static noinline_for_stack
3658int add_data_references(struct reloc_control *rc,
3659 struct btrfs_key *extent_key,
3660 struct btrfs_path *path,
3661 struct rb_root *blocks)
3662{
3663 struct btrfs_key key;
3664 struct extent_buffer *eb;
3665 struct btrfs_extent_data_ref *dref;
3666 struct btrfs_extent_inline_ref *iref;
3667 unsigned long ptr;
3668 unsigned long end;
3fd0a558 3669 u32 blocksize = btrfs_level_size(rc->extent_root, 0);
647f63bd 3670 int ret = 0;
5d4f98a2
YZ		 3671 int err = 0;
3672
5d4f98a2
YZ		 3673 eb = path->nodes[0];
3674 ptr = btrfs_item_ptr_offset(eb, path->slots[0]);
3675 end = ptr + btrfs_item_size_nr(eb, path->slots[0]);
3676#ifdef BTRFS_COMPAT_EXTENT_TREE_V0
3677 if (ptr + sizeof(struct btrfs_extent_item_v0) == end)
3678 ptr = end;
3679 else
3680#endif
3681 ptr += sizeof(struct btrfs_extent_item);
3682
3683 while (ptr < end) {
3684 iref = (struct btrfs_extent_inline_ref *)ptr;
3685 key.type = btrfs_extent_inline_ref_type(eb, iref);
3686 if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
3687 key.offset = btrfs_extent_inline_ref_offset(eb, iref);
3688 ret = __add_tree_block(rc, key.offset, blocksize,
3689 blocks);
3690 } else if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
3691 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
3692 ret = find_data_references(rc, extent_key,
3693 eb, dref, blocks);
3694 } else {
3695 BUG();
3696 }
647f63bd
FDBM		 3697 if (ret) {
3698 err = ret;
3699 goto out;
3700 }
5d4f98a2
YZ		 3701 ptr += btrfs_extent_inline_ref_size(key.type);
3702 }
3703 WARN_ON(ptr > end);
3704
3705 while (1) {
3706 cond_resched();
3707 eb = path->nodes[0];
3708 if (path->slots[0] >= btrfs_header_nritems(eb)) {
3709 ret = btrfs_next_leaf(rc->extent_root, path);
3710 if (ret < 0) {
3711 err = ret;
3712 break;
3713 }
3714 if (ret > 0)
3715 break;
3716 eb = path->nodes[0];
3717 }
3718
3719 btrfs_item_key_to_cpu(eb, &key, path->slots[0]);
3720 if (key.objectid != extent_key->objectid)
3721 break;
3722
3723#ifdef BTRFS_COMPAT_EXTENT_TREE_V0
3724 if (key.type == BTRFS_SHARED_DATA_REF_KEY ||
3725 key.type == BTRFS_EXTENT_REF_V0_KEY) {
3726#else
3727 BUG_ON(key.type == BTRFS_EXTENT_REF_V0_KEY);
3728 if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
3729#endif
3730 ret = __add_tree_block(rc, key.offset, blocksize,
3731 blocks);
3732 } else if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
3733 dref = btrfs_item_ptr(eb, path->slots[0],
3734 struct btrfs_extent_data_ref);
3735 ret = find_data_references(rc, extent_key,
3736 eb, dref, blocks);
3737 } else {
3738 ret = 0;
3739 }
3740 if (ret) {
3741 err = ret;
3742 break;
3743 }
3744 path->slots[0]++;
3745 }
647f63bd 3746out:
b3b4aa74 3747 btrfs_release_path(path);
5d4f98a2
YZ		 3748 if (err)
3749 free_block_list(blocks);
3750 return err;
3751}
3752
3753/*
2c016dc2 3754 * helper to find next unprocessed extent
5d4f98a2
YZ		 3755 */
3756static noinline_for_stack
3757int find_next_extent(struct btrfs_trans_handle *trans,
3fd0a558
YZ		 3758 struct reloc_control *rc, struct btrfs_path *path,
3759 struct btrfs_key *extent_key)
5d4f98a2
YZ		 3760{
3761 struct btrfs_key key;
3762 struct extent_buffer *leaf;
3763 u64 start, end, last;
3764 int ret;
3765
3766 last = rc->block_group->key.objectid + rc->block_group->key.offset;
3767 while (1) {
3768 cond_resched();
3769 if (rc->search_start >= last) {
3770 ret = 1;
3771 break;
3772 }
3773
3774 key.objectid = rc->search_start;
3775 key.type = BTRFS_EXTENT_ITEM_KEY;
3776 key.offset = 0;
3777
3778 path->search_commit_root = 1;
3779 path->skip_locking = 1;
3780 ret = btrfs_search_slot(NULL, rc->extent_root, &key, path,
3781 0, 0);
3782 if (ret < 0)
3783 break;
3784next:
3785 leaf = path->nodes[0];
3786 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
3787 ret = btrfs_next_leaf(rc->extent_root, path);
3788 if (ret != 0)
3789 break;
3790 leaf = path->nodes[0];
3791 }
3792
3793 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
3794 if (key.objectid >= last) {
3795 ret = 1;
3796 break;
3797 }
3798
3173a18f
JB		 3799 if (key.type != BTRFS_EXTENT_ITEM_KEY &&
3800 key.type != BTRFS_METADATA_ITEM_KEY) {
3801 path->slots[0]++;
3802 goto next;
3803 }
3804
3805 if (key.type == BTRFS_EXTENT_ITEM_KEY &&
5d4f98a2
YZ		 3806 key.objectid + key.offset <= rc->search_start) {
3807 path->slots[0]++;
3808 goto next;
3809 }
3810
3173a18f
JB		 3811 if (key.type == BTRFS_METADATA_ITEM_KEY &&
3812 key.objectid + rc->extent_root->leafsize <=
3813 rc->search_start) {
3814 path->slots[0]++;
3815 goto next;
3816 }
3817
5d4f98a2
YZ		 3818 ret = find_first_extent_bit(&rc->processed_blocks,
3819 key.objectid, &start, &end,
e6138876 3820 EXTENT_DIRTY, NULL);
5d4f98a2
YZ		 3821
3822 if (ret == 0 && start <= key.objectid) {
b3b4aa74 3823 btrfs_release_path(path);
5d4f98a2
YZ		 3824 rc->search_start = end + 1;
3825 } else {
3173a18f
JB		 3826 if (key.type == BTRFS_EXTENT_ITEM_KEY)
3827 rc->search_start = key.objectid + key.offset;
3828 else
3829 rc->search_start = key.objectid +
3830 rc->extent_root->leafsize;
3fd0a558 3831 memcpy(extent_key, &key, sizeof(key));
5d4f98a2
YZ		 3832 return 0;
3833 }
3834 }
b3b4aa74 3835 btrfs_release_path(path);
5d4f98a2
YZ		 3836 return ret;
3837}
3838
3839static void set_reloc_control(struct reloc_control *rc)
3840{
3841 struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
7585717f
CM		 3842
3843 mutex_lock(&fs_info->reloc_mutex);
5d4f98a2 3844 fs_info->reloc_ctl = rc;
7585717f 3845 mutex_unlock(&fs_info->reloc_mutex);
5d4f98a2
YZ		 3846}
3847
3848static void unset_reloc_control(struct reloc_control *rc)
3849{
3850 struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
7585717f
CM		 3851
3852 mutex_lock(&fs_info->reloc_mutex);
5d4f98a2 3853 fs_info->reloc_ctl = NULL;
7585717f 3854 mutex_unlock(&fs_info->reloc_mutex);
5d4f98a2
YZ		 3855}
3856
3857static int check_extent_flags(u64 flags)
3858{
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_EXTENT_FLAG_TREE_BLOCK))
3864 return 1;
3865 if ((flags & BTRFS_EXTENT_FLAG_DATA) &&
3866 (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
3867 return 1;
3868 return 0;
3869}
3870
3fd0a558
YZ		 3871static noinline_for_stack
3872int prepare_to_relocate(struct reloc_control *rc)
3873{
3874 struct btrfs_trans_handle *trans;
3fd0a558 3875
66d8f3dd
MX		 3876 rc->block_rsv = btrfs_alloc_block_rsv(rc->extent_root,
3877 BTRFS_BLOCK_RSV_TEMP);
3fd0a558
YZ		 3878 if (!rc->block_rsv)
3879 return -ENOMEM;
3880
3fd0a558
YZ		 3881 memset(&rc->cluster, 0, sizeof(rc->cluster));
3882 rc->search_start = rc->block_group->key.objectid;
3883 rc->extents_found = 0;
3884 rc->nodes_relocated = 0;
3885 rc->merging_rsv_size = 0;
0647bf56
WS		 3886 rc->reserved_bytes = 0;
3887 rc->block_rsv->size = rc->extent_root->nodesize *
3888 RELOCATION_RESERVED_NODES;
3fd0a558
YZ		 3889
3890 rc->create_reloc_tree = 1;
3891 set_reloc_control(rc);
3892
7a7eaa40 3893 trans = btrfs_join_transaction(rc->extent_root);
28818947
LB		 3894 if (IS_ERR(trans)) {
3895 unset_reloc_control(rc);
3896 /*
3897 * extent tree is not a ref_cow tree and has no reloc_root to
3898 * cleanup. And callers are responsible to free the above
3899 * block rsv.
3900 */
3901 return PTR_ERR(trans);
3902 }
3fd0a558
YZ		 3903 btrfs_commit_transaction(trans, rc->extent_root);
3904 return 0;
3905}
76dda93c 3906
5d4f98a2
YZ		 3907static noinline_for_stack int relocate_block_group(struct reloc_control *rc)
3908{
3909 struct rb_root blocks = RB_ROOT;
3910 struct btrfs_key key;
3911 struct btrfs_trans_handle *trans = NULL;
3912 struct btrfs_path *path;
3913 struct btrfs_extent_item *ei;
5d4f98a2
YZ		 3914 u64 flags;
3915 u32 item_size;
3916 int ret;
3917 int err = 0;
c87f08ca 3918 int progress = 0;
5d4f98a2
YZ		 3919
3920 path = btrfs_alloc_path();
3fd0a558 3921 if (!path)
5d4f98a2 3922 return -ENOMEM;
026fd317 3923 path->reada = 1;
5d4f98a2 3924
3fd0a558
YZ		 3925 ret = prepare_to_relocate(rc);
3926 if (ret) {
3927 err = ret;
3928 goto out_free;
3929 }
5d4f98a2
YZ		 3930
3931 while (1) {
0647bf56
WS		 3932 rc->reserved_bytes = 0;
3933 ret = btrfs_block_rsv_refill(rc->extent_root,
3934 rc->block_rsv, rc->block_rsv->size,
3935 BTRFS_RESERVE_FLUSH_ALL);
3936 if (ret) {
3937 err = ret;
3938 break;
3939 }
c87f08ca 3940 progress++;
a22285a6 3941 trans = btrfs_start_transaction(rc->extent_root, 0);
0f788c58
LB		 3942 if (IS_ERR(trans)) {
3943 err = PTR_ERR(trans);
3944 trans = NULL;
3945 break;
3946 }
c87f08ca 3947restart:
3fd0a558
YZ		 3948 if (update_backref_cache(trans, &rc->backref_cache)) {
3949 btrfs_end_transaction(trans, rc->extent_root);
3950 continue;
3951 }
3952
3953 ret = find_next_extent(trans, rc, path, &key);
5d4f98a2
YZ		 3954 if (ret < 0)
3955 err = ret;
3956 if (ret != 0)
3957 break;
3958
3959 rc->extents_found++;
3960
3961 ei = btrfs_item_ptr(path->nodes[0], path->slots[0],
3962 struct btrfs_extent_item);
3fd0a558 3963 item_size = btrfs_item_size_nr(path->nodes[0], path->slots[0]);
5d4f98a2
YZ		 3964 if (item_size >= sizeof(*ei)) {
3965 flags = btrfs_extent_flags(path->nodes[0], ei);
3966 ret = check_extent_flags(flags);
3967 BUG_ON(ret);
3968
3969 } else {
3970#ifdef BTRFS_COMPAT_EXTENT_TREE_V0
3971 u64 ref_owner;
3972 int path_change = 0;
3973
3974 BUG_ON(item_size !=
3975 sizeof(struct btrfs_extent_item_v0));
3976 ret = get_ref_objectid_v0(rc, path, &key, &ref_owner,
3977 &path_change);
3978 if (ref_owner < BTRFS_FIRST_FREE_OBJECTID)
3979 flags = BTRFS_EXTENT_FLAG_TREE_BLOCK;
3980 else
3981 flags = BTRFS_EXTENT_FLAG_DATA;
3982
3983 if (path_change) {
b3b4aa74 3984 btrfs_release_path(path);
5d4f98a2
YZ		 3985
3986 path->search_commit_root = 1;
3987 path->skip_locking = 1;
3988 ret = btrfs_search_slot(NULL, rc->extent_root,
3989 &key, path, 0, 0);
3990 if (ret < 0) {
3991 err = ret;
3992 break;
3993 }
3994 BUG_ON(ret > 0);
3995 }
3996#else
3997 BUG();
3998#endif
3999 }
4000
4001 if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
4002 ret = add_tree_block(rc, &key, path, &blocks);
4003 } else if (rc->stage == UPDATE_DATA_PTRS &&
3fd0a558 4004 (flags & BTRFS_EXTENT_FLAG_DATA)) {
5d4f98a2
YZ		 4005 ret = add_data_references(rc, &key, path, &blocks);
4006 } else {
b3b4aa74 4007 btrfs_release_path(path);
5d4f98a2
YZ		 4008 ret = 0;
4009 }
4010 if (ret < 0) {
3fd0a558 4011 err = ret;
5d4f98a2
YZ		 4012 break;
4013 }
4014
4015 if (!RB_EMPTY_ROOT(&blocks)) {
4016 ret = relocate_tree_blocks(trans, rc, &blocks);
4017 if (ret < 0) {
3fd0a558
YZ		 4018 if (ret != -EAGAIN) {
4019 err = ret;
4020 break;
4021 }
4022 rc->extents_found--;
4023 rc->search_start = key.objectid;
4024 }
4025 }
4026
0647bf56
WS		 4027 btrfs_end_transaction_throttle(trans, rc->extent_root);
4028 btrfs_btree_balance_dirty(rc->extent_root);
5d4f98a2 4029 trans = NULL;
5d4f98a2
YZ		 4030
4031 if (rc->stage == MOVE_DATA_EXTENTS &&
4032 (flags & BTRFS_EXTENT_FLAG_DATA)) {
4033 rc->found_file_extent = 1;
0257bb82 4034 ret = relocate_data_extent(rc->data_inode,
3fd0a558 4035 &key, &rc->cluster);
5d4f98a2
YZ		 4036 if (ret < 0) {
4037 err = ret;
4038 break;
4039 }
4040 }
4041 }
c87f08ca
CM		 4042 if (trans && progress && err == -ENOSPC) {
4043 ret = btrfs_force_chunk_alloc(trans, rc->extent_root,
4044 rc->block_group->flags);
4045 if (ret == 0) {
4046 err = 0;
4047 progress = 0;
4048 goto restart;
4049 }
4050 }
3fd0a558 4051
b3b4aa74 4052 btrfs_release_path(path);
3fd0a558
YZ		 4053 clear_extent_bits(&rc->processed_blocks, 0, (u64)-1, EXTENT_DIRTY,
4054 GFP_NOFS);
5d4f98a2
YZ		 4055
4056 if (trans) {
3fd0a558 4057 btrfs_end_transaction_throttle(trans, rc->extent_root);
b53d3f5d 4058 btrfs_btree_balance_dirty(rc->extent_root);
5d4f98a2
YZ		 4059 }
4060
0257bb82 4061 if (!err) {
3fd0a558
YZ		 4062 ret = relocate_file_extent_cluster(rc->data_inode,
4063 &rc->cluster);
0257bb82
YZ		 4064 if (ret < 0)
4065 err = ret;
4066 }
4067
3fd0a558
YZ		 4068 rc->create_reloc_tree = 0;
4069 set_reloc_control(rc);
0257bb82 4070
3fd0a558
YZ		 4071 backref_cache_cleanup(&rc->backref_cache);
4072 btrfs_block_rsv_release(rc->extent_root, rc->block_rsv, (u64)-1);
5d4f98a2 4073
3fd0a558 4074 err = prepare_to_merge(rc, err);
5d4f98a2
YZ		 4075
4076 merge_reloc_roots(rc);
4077
3fd0a558 4078 rc->merge_reloc_tree = 0;
5d4f98a2 4079 unset_reloc_control(rc);
3fd0a558 4080 btrfs_block_rsv_release(rc->extent_root, rc->block_rsv, (u64)-1);
5d4f98a2
YZ		 4081
4082 /* get rid of pinned extents */
7a7eaa40 4083 trans = btrfs_join_transaction(rc->extent_root);
3612b495
TI		 4084 if (IS_ERR(trans))
4085 err = PTR_ERR(trans);
4086 else
4087 btrfs_commit_transaction(trans, rc->extent_root);
3fd0a558
YZ		 4088out_free:
4089 btrfs_free_block_rsv(rc->extent_root, rc->block_rsv);
4090 btrfs_free_path(path);
5d4f98a2
YZ		 4091 return err;
4092}
4093
4094static int __insert_orphan_inode(struct btrfs_trans_handle *trans,
0257bb82 4095 struct btrfs_root *root, u64 objectid)
5d4f98a2
YZ		 4096{
4097 struct btrfs_path *path;
4098 struct btrfs_inode_item *item;
4099 struct extent_buffer *leaf;
4100 int ret;
4101
4102 path = btrfs_alloc_path();
4103 if (!path)
4104 return -ENOMEM;
4105
4106 ret = btrfs_insert_empty_inode(trans, root, path, objectid);
4107 if (ret)
4108 goto out;
4109
4110 leaf = path->nodes[0];
4111 item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_inode_item);
4112 memset_extent_buffer(leaf, 0, (unsigned long)item, sizeof(*item));
4113 btrfs_set_inode_generation(leaf, item, 1);
0257bb82 4114 btrfs_set_inode_size(leaf, item, 0);
5d4f98a2 4115 btrfs_set_inode_mode(leaf, item, S_IFREG | 0600);
3fd0a558
YZ		 4116 btrfs_set_inode_flags(leaf, item, BTRFS_INODE_NOCOMPRESS |
4117 BTRFS_INODE_PREALLOC);
5d4f98a2 4118 btrfs_mark_buffer_dirty(leaf);
b3b4aa74 4119 btrfs_release_path(path);
5d4f98a2
YZ		 4120out:
4121 btrfs_free_path(path);
4122 return ret;
4123}
4124
4125/*
4126 * helper to create inode for data relocation.
4127 * the inode is in data relocation tree and its link count is 0
4128 */
3fd0a558
YZ		 4129static noinline_for_stack
4130struct inode *create_reloc_inode(struct btrfs_fs_info *fs_info,
4131 struct btrfs_block_group_cache *group)
5d4f98a2
YZ		 4132{
4133 struct inode *inode = NULL;
4134 struct btrfs_trans_handle *trans;
4135 struct btrfs_root *root;
4136 struct btrfs_key key;
5d4f98a2
YZ		 4137 u64 objectid = BTRFS_FIRST_FREE_OBJECTID;
4138 int err = 0;
4139
4140 root = read_fs_root(fs_info, BTRFS_DATA_RELOC_TREE_OBJECTID);
4141 if (IS_ERR(root))
4142 return ERR_CAST(root);
4143
a22285a6 4144 trans = btrfs_start_transaction(root, 6);
3fd0a558
YZ		 4145 if (IS_ERR(trans))
4146 return ERR_CAST(trans);
5d4f98a2 4147
581bb050 4148 err = btrfs_find_free_objectid(root, &objectid);
5d4f98a2
YZ		 4149 if (err)
4150 goto out;
4151
0257bb82 4152 err = __insert_orphan_inode(trans, root, objectid);
5d4f98a2
YZ		 4153 BUG_ON(err);
4154
4155 key.objectid = objectid;
4156 key.type = BTRFS_INODE_ITEM_KEY;
4157 key.offset = 0;
73f73415 4158 inode = btrfs_iget(root->fs_info->sb, &key, root, NULL);
5d4f98a2
YZ		 4159 BUG_ON(IS_ERR(inode) || is_bad_inode(inode));
4160 BTRFS_I(inode)->index_cnt = group->key.objectid;
4161
4162 err = btrfs_orphan_add(trans, inode);
4163out:
5d4f98a2 4164 btrfs_end_transaction(trans, root);
b53d3f5d 4165 btrfs_btree_balance_dirty(root);
5d4f98a2
YZ		 4166 if (err) {
4167 if (inode)
4168 iput(inode);
4169 inode = ERR_PTR(err);
4170 }
4171 return inode;
4172}
4173
a9995eec 4174static struct reloc_control *alloc_reloc_control(struct btrfs_fs_info *fs_info)
3fd0a558
YZ		 4175{
4176 struct reloc_control *rc;
4177
4178 rc = kzalloc(sizeof(*rc), GFP_NOFS);
4179 if (!rc)
4180 return NULL;
4181
4182 INIT_LIST_HEAD(&rc->reloc_roots);
4183 backref_cache_init(&rc->backref_cache);
4184 mapping_tree_init(&rc->reloc_root_tree);
a9995eec
JB		 4185 extent_io_tree_init(&rc->processed_blocks,
4186 fs_info->btree_inode->i_mapping);
3fd0a558
YZ		 4187 return rc;
4188}
4189
5d4f98a2
YZ		 4190/*
4191 * function to relocate all extents in a block group.
4192 */
4193int btrfs_relocate_block_group(struct btrfs_root *extent_root, u64 group_start)
4194{
4195 struct btrfs_fs_info *fs_info = extent_root->fs_info;
4196 struct reloc_control *rc;
0af3d00b
JB		 4197 struct inode *inode;
4198 struct btrfs_path *path;
5d4f98a2 4199 int ret;
f0486c68 4200 int rw = 0;
5d4f98a2
YZ		 4201 int err = 0;
4202
a9995eec 4203 rc = alloc_reloc_control(fs_info);
5d4f98a2
YZ		 4204 if (!rc)
4205 return -ENOMEM;
4206
f0486c68 4207 rc->extent_root = extent_root;
3fd0a558 4208
5d4f98a2
YZ		 4209 rc->block_group = btrfs_lookup_block_group(fs_info, group_start);
4210 BUG_ON(!rc->block_group);
4211
f0486c68
YZ		 4212 if (!rc->block_group->ro) {
4213 ret = btrfs_set_block_group_ro(extent_root, rc->block_group);
4214 if (ret) {
4215 err = ret;
4216 goto out;
4217 }
4218 rw = 1;
4219 }
4220
0af3d00b
JB		 4221 path = btrfs_alloc_path();
4222 if (!path) {
4223 err = -ENOMEM;
4224 goto out;
4225 }
4226
4227 inode = lookup_free_space_inode(fs_info->tree_root, rc->block_group,
4228 path);
4229 btrfs_free_path(path);
4230
4231 if (!IS_ERR(inode))
4232 ret = delete_block_group_cache(fs_info, inode, 0);
4233 else
4234 ret = PTR_ERR(inode);
4235
4236 if (ret && ret != -ENOENT) {
4237 err = ret;
4238 goto out;
4239 }
4240
5d4f98a2
YZ		 4241 rc->data_inode = create_reloc_inode(fs_info, rc->block_group);
4242 if (IS_ERR(rc->data_inode)) {
4243 err = PTR_ERR(rc->data_inode);
4244 rc->data_inode = NULL;
4245 goto out;
4246 }
4247
efe120a0 4248 btrfs_info(extent_root->fs_info, "relocating block group %llu flags %llu",
c1c9ff7c 4249 rc->block_group->key.objectid, rc->block_group->flags);
5d4f98a2 4250
91aef86f 4251 ret = btrfs_start_delalloc_roots(fs_info, 0);
8ccf6f19
MX		 4252 if (ret < 0) {
4253 err = ret;
4254 goto out;
4255 }
b0244199 4256 btrfs_wait_ordered_roots(fs_info, -1);
5d4f98a2
YZ		 4257
4258 while (1) {
76dda93c 4259 mutex_lock(&fs_info->cleaner_mutex);
5d4f98a2 4260 ret = relocate_block_group(rc);
76dda93c 4261 mutex_unlock(&fs_info->cleaner_mutex);
5d4f98a2
YZ		 4262 if (ret < 0) {
4263 err = ret;
3fd0a558 4264 goto out;
5d4f98a2
YZ		 4265 }
4266
4267 if (rc->extents_found == 0)
4268 break;
4269
efe120a0 4270 btrfs_info(extent_root->fs_info, "found %llu extents",
c1c9ff7c 4271 rc->extents_found);
5d4f98a2
YZ		 4272
4273 if (rc->stage == MOVE_DATA_EXTENTS && rc->found_file_extent) {
0ef8b726
JB		 4274 ret = btrfs_wait_ordered_range(rc->data_inode, 0,
4275 (u64)-1);
4276 if (ret) {
4277 err = ret;
4278 goto out;
4279 }
5d4f98a2
YZ		 4280 invalidate_mapping_pages(rc->data_inode->i_mapping,
4281 0, -1);
4282 rc->stage = UPDATE_DATA_PTRS;
5d4f98a2
YZ		 4283 }
4284 }
4285
5d4f98a2
YZ		 4286 WARN_ON(rc->block_group->pinned > 0);
4287 WARN_ON(rc->block_group->reserved > 0);
4288 WARN_ON(btrfs_block_group_used(&rc->block_group->item) > 0);
4289out:
f0486c68
YZ		 4290 if (err && rw)
4291 btrfs_set_block_group_rw(extent_root, rc->block_group);
5d4f98a2 4292 iput(rc->data_inode);
5d4f98a2
YZ		 4293 btrfs_put_block_group(rc->block_group);
4294 kfree(rc);
4295 return err;
4296}
4297
76dda93c
YZ		 4298static noinline_for_stack int mark_garbage_root(struct btrfs_root *root)
4299{
4300 struct btrfs_trans_handle *trans;
79787eaa 4301 int ret, err;
76dda93c 4302
a22285a6 4303 trans = btrfs_start_transaction(root->fs_info->tree_root, 0);
79787eaa
JM		 4304 if (IS_ERR(trans))
4305 return PTR_ERR(trans);
76dda93c
YZ		 4306
4307 memset(&root->root_item.drop_progress, 0,
4308 sizeof(root->root_item.drop_progress));
4309 root->root_item.drop_level = 0;
4310 btrfs_set_root_refs(&root->root_item, 0);
4311 ret = btrfs_update_root(trans, root->fs_info->tree_root,
4312 &root->root_key, &root->root_item);
76dda93c 4313
79787eaa
JM		 4314 err = btrfs_end_transaction(trans, root->fs_info->tree_root);
4315 if (err)
4316 return err;
4317 return ret;
76dda93c
YZ		 4318}
4319
5d4f98a2
YZ		 4320/*
4321 * recover relocation interrupted by system crash.
4322 *
4323 * this function resumes merging reloc trees with corresponding fs trees.
4324 * this is important for keeping the sharing of tree blocks
4325 */
4326int btrfs_recover_relocation(struct btrfs_root *root)
4327{
4328 LIST_HEAD(reloc_roots);
4329 struct btrfs_key key;
4330 struct btrfs_root *fs_root;
4331 struct btrfs_root *reloc_root;
4332 struct btrfs_path *path;
4333 struct extent_buffer *leaf;
4334 struct reloc_control *rc = NULL;
4335 struct btrfs_trans_handle *trans;
4336 int ret;
4337 int err = 0;
4338
4339 path = btrfs_alloc_path();
4340 if (!path)
4341 return -ENOMEM;
026fd317 4342 path->reada = -1;
5d4f98a2
YZ		 4343
4344 key.objectid = BTRFS_TREE_RELOC_OBJECTID;
4345 key.type = BTRFS_ROOT_ITEM_KEY;
4346 key.offset = (u64)-1;
4347
4348 while (1) {
4349 ret = btrfs_search_slot(NULL, root->fs_info->tree_root, &key,
4350 path, 0, 0);
4351 if (ret < 0) {
4352 err = ret;
4353 goto out;
4354 }
4355 if (ret > 0) {
4356 if (path->slots[0] == 0)
4357 break;
4358 path->slots[0]--;
4359 }
4360 leaf = path->nodes[0];
4361 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
b3b4aa74 4362 btrfs_release_path(path);
5d4f98a2
YZ		 4363
4364 if (key.objectid != BTRFS_TREE_RELOC_OBJECTID ||
4365 key.type != BTRFS_ROOT_ITEM_KEY)
4366 break;
4367
cb517eab 4368 reloc_root = btrfs_read_fs_root(root, &key);
5d4f98a2
YZ		 4369 if (IS_ERR(reloc_root)) {
4370 err = PTR_ERR(reloc_root);
4371 goto out;
4372 }
4373
4374 list_add(&reloc_root->root_list, &reloc_roots);
4375
4376 if (btrfs_root_refs(&reloc_root->root_item) > 0) {
4377 fs_root = read_fs_root(root->fs_info,
4378 reloc_root->root_key.offset);
4379 if (IS_ERR(fs_root)) {
76dda93c
YZ		 4380 ret = PTR_ERR(fs_root);
4381 if (ret != -ENOENT) {
4382 err = ret;
4383 goto out;
4384 }
79787eaa
JM		 4385 ret = mark_garbage_root(reloc_root);
4386 if (ret < 0) {
4387 err = ret;
4388 goto out;
4389 }
5d4f98a2
YZ		 4390 }
4391 }
4392
4393 if (key.offset == 0)
4394 break;
4395
4396 key.offset--;
4397 }
b3b4aa74 4398 btrfs_release_path(path);
5d4f98a2
YZ		 4399
4400 if (list_empty(&reloc_roots))
4401 goto out;
4402
a9995eec 4403 rc = alloc_reloc_control(root->fs_info);
5d4f98a2
YZ		 4404 if (!rc) {
4405 err = -ENOMEM;
4406 goto out;
4407 }
4408
5d4f98a2
YZ		 4409 rc->extent_root = root->fs_info->extent_root;
4410
4411 set_reloc_control(rc);
4412
7a7eaa40 4413 trans = btrfs_join_transaction(rc->extent_root);
3612b495
TI		 4414 if (IS_ERR(trans)) {
4415 unset_reloc_control(rc);
4416 err = PTR_ERR(trans);
4417 goto out_free;
4418 }
3fd0a558
YZ		 4419
4420 rc->merge_reloc_tree = 1;
4421
5d4f98a2
YZ		 4422 while (!list_empty(&reloc_roots)) {
4423 reloc_root = list_entry(reloc_roots.next,
4424 struct btrfs_root, root_list);
4425 list_del(&reloc_root->root_list);
4426
4427 if (btrfs_root_refs(&reloc_root->root_item) == 0) {
4428 list_add_tail(&reloc_root->root_list,
4429 &rc->reloc_roots);
4430 continue;
4431 }
4432
4433 fs_root = read_fs_root(root->fs_info,
4434 reloc_root->root_key.offset);
79787eaa
JM		 4435 if (IS_ERR(fs_root)) {
4436 err = PTR_ERR(fs_root);
4437 goto out_free;
4438 }
5d4f98a2 4439
ffd7b339 4440 err = __add_reloc_root(reloc_root);
79787eaa 4441 BUG_ON(err < 0); /* -ENOMEM or logic error */
5d4f98a2
YZ		 4442 fs_root->reloc_root = reloc_root;
4443 }
4444
79787eaa
JM		 4445 err = btrfs_commit_transaction(trans, rc->extent_root);
4446 if (err)
4447 goto out_free;
5d4f98a2
YZ		 4448
4449 merge_reloc_roots(rc);
4450
4451 unset_reloc_control(rc);
4452
7a7eaa40 4453 trans = btrfs_join_transaction(rc->extent_root);
3612b495
TI		 4454 if (IS_ERR(trans))
4455 err = PTR_ERR(trans);
4456 else
79787eaa 4457 err = btrfs_commit_transaction(trans, rc->extent_root);
3612b495 4458out_free:
3fd0a558 4459 kfree(rc);
3612b495 4460out:
aca1bba6
LB		 4461 if (!list_empty(&reloc_roots))
4462 free_reloc_roots(&reloc_roots);
4463
5d4f98a2
YZ		 4464 btrfs_free_path(path);
4465
4466 if (err == 0) {
4467 /* cleanup orphan inode in data relocation tree */
4468 fs_root = read_fs_root(root->fs_info,
4469 BTRFS_DATA_RELOC_TREE_OBJECTID);
4470 if (IS_ERR(fs_root))
4471 err = PTR_ERR(fs_root);
d7ce5843 4472 else
66b4ffd1 4473 err = btrfs_orphan_cleanup(fs_root);
5d4f98a2
YZ		 4474 }
4475 return err;
4476}
4477
4478/*
4479 * helper to add ordered checksum for data relocation.
4480 *
4481 * cloning checksum properly handles the nodatasum extents.
4482 * it also saves CPU time to re-calculate the checksum.
4483 */
4484int btrfs_reloc_clone_csums(struct inode *inode, u64 file_pos, u64 len)
4485{
4486 struct btrfs_ordered_sum *sums;
5d4f98a2
YZ		 4487 struct btrfs_ordered_extent *ordered;
4488 struct btrfs_root *root = BTRFS_I(inode)->root;
5d4f98a2
YZ		 4489 int ret;
4490 u64 disk_bytenr;
4577b014 4491 u64 new_bytenr;
5d4f98a2
YZ		 4492 LIST_HEAD(list);
4493
4494 ordered = btrfs_lookup_ordered_extent(inode, file_pos);
4495 BUG_ON(ordered->file_offset != file_pos || ordered->len != len);
4496
4497 disk_bytenr = file_pos + BTRFS_I(inode)->index_cnt;
4498 ret = btrfs_lookup_csums_range(root->fs_info->csum_root, disk_bytenr,
a2de733c 4499 disk_bytenr + len - 1, &list, 0);
79787eaa
JM		 4500 if (ret)
4501 goto out;
5d4f98a2
YZ		 4502
4503 while (!list_empty(&list)) {
4504 sums = list_entry(list.next, struct btrfs_ordered_sum, list);
4505 list_del_init(&sums->list);
4506
4577b014
JB		 4507 /*
4508 * We need to offset the new_bytenr based on where the csum is.
4509 * We need to do this because we will read in entire prealloc
4510 * extents but we may have written to say the middle of the
4511 * prealloc extent, so we need to make sure the csum goes with
4512 * the right disk offset.
4513 *
4514 * We can do this because the data reloc inode refers strictly
4515 * to the on disk bytes, so we don't have to worry about
4516 * disk_len vs real len like with real inodes since it's all
4517 * disk length.
4518 */
4519 new_bytenr = ordered->start + (sums->bytenr - disk_bytenr);
4520 sums->bytenr = new_bytenr;
5d4f98a2
YZ		 4521
4522 btrfs_add_ordered_sum(inode, ordered, sums);
4523 }
79787eaa 4524out:
5d4f98a2 4525 btrfs_put_ordered_extent(ordered);
411fc6bc 4526 return ret;
5d4f98a2 4527}
3fd0a558 4528
83d4cfd4
JB		 4529int btrfs_reloc_cow_block(struct btrfs_trans_handle *trans,
4530 struct btrfs_root *root, struct extent_buffer *buf,
4531 struct extent_buffer *cow)
3fd0a558
YZ		 4532{
4533 struct reloc_control *rc;
4534 struct backref_node *node;
4535 int first_cow = 0;
4536 int level;
83d4cfd4 4537 int ret = 0;
3fd0a558
YZ		 4538
4539 rc = root->fs_info->reloc_ctl;
4540 if (!rc)
83d4cfd4 4541 return 0;
3fd0a558
YZ		 4542
4543 BUG_ON(rc->stage == UPDATE_DATA_PTRS &&
4544 root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID);
4545
c974c464
WS		 4546 if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID) {
4547 if (buf == root->node)
4548 __update_reloc_root(root, cow->start);
4549 }
4550
3fd0a558
YZ		 4551 level = btrfs_header_level(buf);
4552 if (btrfs_header_generation(buf) <=
4553 btrfs_root_last_snapshot(&root->root_item))
4554 first_cow = 1;
4555
4556 if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID &&
4557 rc->create_reloc_tree) {
4558 WARN_ON(!first_cow && level == 0);
4559
4560 node = rc->backref_cache.path[level];
4561 BUG_ON(node->bytenr != buf->start &&
4562 node->new_bytenr != buf->start);
4563
4564 drop_node_buffer(node);
4565 extent_buffer_get(cow);
4566 node->eb = cow;
4567 node->new_bytenr = cow->start;
4568
4569 if (!node->pending) {
4570 list_move_tail(&node->list,
4571 &rc->backref_cache.pending[level]);
4572 node->pending = 1;
4573 }
4574
4575 if (first_cow)
4576 __mark_block_processed(rc, node);
4577
4578 if (first_cow && level > 0)
4579 rc->nodes_relocated += buf->len;
4580 }
4581
83d4cfd4 4582 if (level == 0 && first_cow && rc->stage == UPDATE_DATA_PTRS)
3fd0a558 4583 ret = replace_file_extents(trans, rc, root, cow);
83d4cfd4 4584 return ret;
3fd0a558
YZ		 4585}
4586
4587/*
4588 * called before creating snapshot. it calculates metadata reservation
4589 * requried for relocating tree blocks in the snapshot
4590 */
4591void btrfs_reloc_pre_snapshot(struct btrfs_trans_handle *trans,
4592 struct btrfs_pending_snapshot *pending,
4593 u64 *bytes_to_reserve)
4594{
4595 struct btrfs_root *root;
4596 struct reloc_control *rc;
4597
4598 root = pending->root;
4599 if (!root->reloc_root)
4600 return;
4601
4602 rc = root->fs_info->reloc_ctl;
4603 if (!rc->merge_reloc_tree)
4604 return;
4605
4606 root = root->reloc_root;
4607 BUG_ON(btrfs_root_refs(&root->root_item) == 0);
4608 /*
4609 * relocation is in the stage of merging trees. the space
4610 * used by merging a reloc tree is twice the size of
4611 * relocated tree nodes in the worst case. half for cowing
4612 * the reloc tree, half for cowing the fs tree. the space
4613 * used by cowing the reloc tree will be freed after the
4614 * tree is dropped. if we create snapshot, cowing the fs
4615 * tree may use more space than it frees. so we need
4616 * reserve extra space.
4617 */
4618 *bytes_to_reserve += rc->nodes_relocated;
4619}
4620
4621/*
4622 * called after snapshot is created. migrate block reservation
4623 * and create reloc root for the newly created snapshot
4624 */
49b25e05 4625int btrfs_reloc_post_snapshot(struct btrfs_trans_handle *trans,
3fd0a558
YZ		 4626 struct btrfs_pending_snapshot *pending)
4627{
4628 struct btrfs_root *root = pending->root;
4629 struct btrfs_root *reloc_root;
4630 struct btrfs_root *new_root;
4631 struct reloc_control *rc;
4632 int ret;
4633
4634 if (!root->reloc_root)
49b25e05 4635 return 0;
3fd0a558
YZ		 4636
4637 rc = root->fs_info->reloc_ctl;
4638 rc->merging_rsv_size += rc->nodes_relocated;
4639
4640 if (rc->merge_reloc_tree) {
4641 ret = btrfs_block_rsv_migrate(&pending->block_rsv,
4642 rc->block_rsv,
4643 rc->nodes_relocated);
49b25e05
JM		 4644 if (ret)
4645 return ret;
3fd0a558
YZ		 4646 }
4647
4648 new_root = pending->snap;
4649 reloc_root = create_reloc_root(trans, root->reloc_root,
4650 new_root->root_key.objectid);
49b25e05
JM		 4651 if (IS_ERR(reloc_root))
4652 return PTR_ERR(reloc_root);
3fd0a558 4653
ffd7b339
JM		 4654 ret = __add_reloc_root(reloc_root);
4655 BUG_ON(ret < 0);
3fd0a558
YZ		 4656 new_root->reloc_root = reloc_root;
4657
49b25e05 4658 if (rc->create_reloc_tree)
3fd0a558 4659 ret = clone_backref_node(trans, rc, root, reloc_root);
49b25e05 4660 return ret;
3fd0a558 4661}
Linux 6.x block layer and io_uring tree(s)