projects / deliverable / linux.git / blame
? search:
summary | shortlog | log | commit | commitdiff | tree
blob | blame (incremental) | history | HEAD
Btrfs: fix build_backref_tree issue with multiple shared blocks
[deliverable/linux.git] / fs / btrfs / relocation.c
CommitLineData
5d4f98a2
YZ		 1/*
2 * Copyright (C) 2009 Oracle. All rights reserved.
3 *
4 * This program is free software; you can redistribute it and/or
5 * modify it under the terms of the GNU General Public
6 * License v2 as published by the Free Software Foundation.
7 *
8 * This program is distributed in the hope that it will be useful,
9 * but WITHOUT ANY WARRANTY; without even the implied warranty of
10 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
11 * General Public License for more details.
12 *
13 * You should have received a copy of the GNU General Public
14 * License along with this program; if not, write to the
15 * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
16 * Boston, MA 021110-1307, USA.
17 */
18
19#include <linux/sched.h>
20#include <linux/pagemap.h>
21#include <linux/writeback.h>
22#include <linux/blkdev.h>
23#include <linux/rbtree.h>
5a0e3ad6 24#include <linux/slab.h>
5d4f98a2
YZ		 25#include "ctree.h"
26#include "disk-io.h"
27#include "transaction.h"
28#include "volumes.h"
29#include "locking.h"
30#include "btrfs_inode.h"
31#include "async-thread.h"
0af3d00b 32#include "free-space-cache.h"
581bb050 33#include "inode-map.h"
5d4f98a2
YZ		 34
35/*
36 * backref_node, mapping_node and tree_block start with this
37 */
38struct tree_entry {
39 struct rb_node rb_node;
40 u64 bytenr;
41};
42
43/*
44 * present a tree block in the backref cache
45 */
46struct backref_node {
47 struct rb_node rb_node;
48 u64 bytenr;
3fd0a558
YZ		 49
50 u64 new_bytenr;
51 /* objectid of tree block owner, can be not uptodate */
5d4f98a2 52 u64 owner;
3fd0a558
YZ		 53 /* link to pending, changed or detached list */
54 struct list_head list;
5d4f98a2
YZ		 55 /* list of upper level blocks reference this block */
56 struct list_head upper;
57 /* list of child blocks in the cache */
58 struct list_head lower;
59 /* NULL if this node is not tree root */
60 struct btrfs_root *root;
61 /* extent buffer got by COW the block */
62 struct extent_buffer *eb;
63 /* level of tree block */
64 unsigned int level:8;
3fd0a558
YZ		 65 /* is the block in non-reference counted tree */
66 unsigned int cowonly:1;
67 /* 1 if no child node in the cache */
5d4f98a2
YZ		 68 unsigned int lowest:1;
69 /* is the extent buffer locked */
70 unsigned int locked:1;
71 /* has the block been processed */
72 unsigned int processed:1;
73 /* have backrefs of this block been checked */
74 unsigned int checked:1;
3fd0a558
YZ		 75 /*
76 * 1 if corresponding block has been cowed but some upper
77 * level block pointers may not point to the new location
78 */
79 unsigned int pending:1;
80 /*
81 * 1 if the backref node isn't connected to any other
82 * backref node.
83 */
84 unsigned int detached:1;
5d4f98a2
YZ		 85};
86
87/*
88 * present a block pointer in the backref cache
89 */
90struct backref_edge {
91 struct list_head list[2];
92 struct backref_node *node[2];
5d4f98a2
YZ		 93};
94
95#define LOWER 0
96#define UPPER 1
0647bf56 97#define RELOCATION_RESERVED_NODES 256
5d4f98a2
YZ		 98
99struct backref_cache {
100 /* red black tree of all backref nodes in the cache */
101 struct rb_root rb_root;
3fd0a558
YZ		 102 /* for passing backref nodes to btrfs_reloc_cow_block */
103 struct backref_node *path[BTRFS_MAX_LEVEL];
104 /*
105 * list of blocks that have been cowed but some block
106 * pointers in upper level blocks may not reflect the
107 * new location
108 */
5d4f98a2 109 struct list_head pending[BTRFS_MAX_LEVEL];
3fd0a558
YZ		 110 /* list of backref nodes with no child node */
111 struct list_head leaves;
112 /* list of blocks that have been cowed in current transaction */
113 struct list_head changed;
114 /* list of detached backref node. */
115 struct list_head detached;
116
117 u64 last_trans;
118
119 int nr_nodes;
120 int nr_edges;
5d4f98a2
YZ		 121};
122
123/*
124 * map address of tree root to tree
125 */
126struct mapping_node {
127 struct rb_node rb_node;
128 u64 bytenr;
129 void *data;
130};
131
132struct mapping_tree {
133 struct rb_root rb_root;
134 spinlock_t lock;
135};
136
137/*
138 * present a tree block to process
139 */
140struct tree_block {
141 struct rb_node rb_node;
142 u64 bytenr;
143 struct btrfs_key key;
144 unsigned int level:8;
145 unsigned int key_ready:1;
146};
147
0257bb82
YZ		 148#define MAX_EXTENTS 128
149
150struct file_extent_cluster {
151 u64 start;
152 u64 end;
153 u64 boundary[MAX_EXTENTS];
154 unsigned int nr;
155};
156
5d4f98a2
YZ		 157struct reloc_control {
158 /* block group to relocate */
159 struct btrfs_block_group_cache *block_group;
160 /* extent tree */
161 struct btrfs_root *extent_root;
162 /* inode for moving data */
163 struct inode *data_inode;
3fd0a558
YZ		 164
165 struct btrfs_block_rsv *block_rsv;
166
167 struct backref_cache backref_cache;
168
169 struct file_extent_cluster cluster;
5d4f98a2
YZ		 170 /* tree blocks have been processed */
171 struct extent_io_tree processed_blocks;
172 /* map start of tree root to corresponding reloc tree */
173 struct mapping_tree reloc_root_tree;
174 /* list of reloc trees */
175 struct list_head reloc_roots;
3fd0a558
YZ		 176 /* size of metadata reservation for merging reloc trees */
177 u64 merging_rsv_size;
178 /* size of relocated tree nodes */
179 u64 nodes_relocated;
0647bf56
WS		 180 /* reserved size for block group relocation*/
181 u64 reserved_bytes;
3fd0a558 182
5d4f98a2
YZ		 183 u64 search_start;
184 u64 extents_found;
3fd0a558 185
3fd0a558
YZ		 186 unsigned int stage:8;
187 unsigned int create_reloc_tree:1;
188 unsigned int merge_reloc_tree:1;
5d4f98a2 189 unsigned int found_file_extent:1;
5d4f98a2
YZ		 190};
191
192/* stages of data relocation */
193#define MOVE_DATA_EXTENTS 0
194#define UPDATE_DATA_PTRS 1
195
3fd0a558
YZ		 196static void remove_backref_node(struct backref_cache *cache,
197 struct backref_node *node);
198static void __mark_block_processed(struct reloc_control *rc,
199 struct backref_node *node);
5d4f98a2
YZ		 200
201static void mapping_tree_init(struct mapping_tree *tree)
202{
6bef4d31 203 tree->rb_root = RB_ROOT;
5d4f98a2
YZ		 204 spin_lock_init(&tree->lock);
205}
206
207static void backref_cache_init(struct backref_cache *cache)
208{
209 int i;
6bef4d31 210 cache->rb_root = RB_ROOT;
5d4f98a2
YZ		 211 for (i = 0; i < BTRFS_MAX_LEVEL; i++)
212 INIT_LIST_HEAD(&cache->pending[i]);
3fd0a558
YZ		 213 INIT_LIST_HEAD(&cache->changed);
214 INIT_LIST_HEAD(&cache->detached);
215 INIT_LIST_HEAD(&cache->leaves);
216}
217
218static void backref_cache_cleanup(struct backref_cache *cache)
219{
220 struct backref_node *node;
221 int i;
222
223 while (!list_empty(&cache->detached)) {
224 node = list_entry(cache->detached.next,
225 struct backref_node, list);
226 remove_backref_node(cache, node);
227 }
228
229 while (!list_empty(&cache->leaves)) {
230 node = list_entry(cache->leaves.next,
231 struct backref_node, lower);
232 remove_backref_node(cache, node);
233 }
234
235 cache->last_trans = 0;
236
237 for (i = 0; i < BTRFS_MAX_LEVEL; i++)
238 BUG_ON(!list_empty(&cache->pending[i]));
239 BUG_ON(!list_empty(&cache->changed));
240 BUG_ON(!list_empty(&cache->detached));
241 BUG_ON(!RB_EMPTY_ROOT(&cache->rb_root));
242 BUG_ON(cache->nr_nodes);
243 BUG_ON(cache->nr_edges);
244}
245
246static struct backref_node *alloc_backref_node(struct backref_cache *cache)
247{
248 struct backref_node *node;
249
250 node = kzalloc(sizeof(*node), GFP_NOFS);
251 if (node) {
252 INIT_LIST_HEAD(&node->list);
253 INIT_LIST_HEAD(&node->upper);
254 INIT_LIST_HEAD(&node->lower);
255 RB_CLEAR_NODE(&node->rb_node);
256 cache->nr_nodes++;
257 }
258 return node;
259}
260
261static void free_backref_node(struct backref_cache *cache,
262 struct backref_node *node)
263{
264 if (node) {
265 cache->nr_nodes--;
266 kfree(node);
267 }
268}
269
270static struct backref_edge *alloc_backref_edge(struct backref_cache *cache)
271{
272 struct backref_edge *edge;
273
274 edge = kzalloc(sizeof(*edge), GFP_NOFS);
275 if (edge)
276 cache->nr_edges++;
277 return edge;
5d4f98a2
YZ		 278}
279
3fd0a558
YZ		 280static void free_backref_edge(struct backref_cache *cache,
281 struct backref_edge *edge)
5d4f98a2 282{
3fd0a558
YZ		 283 if (edge) {
284 cache->nr_edges--;
285 kfree(edge);
286 }
5d4f98a2
YZ		 287}
288
289static struct rb_node *tree_insert(struct rb_root *root, u64 bytenr,
290 struct rb_node *node)
291{
292 struct rb_node **p = &root->rb_node;
293 struct rb_node *parent = NULL;
294 struct tree_entry *entry;
295
296 while (*p) {
297 parent = *p;
298 entry = rb_entry(parent, struct tree_entry, rb_node);
299
300 if (bytenr < entry->bytenr)
301 p = &(*p)->rb_left;
302 else if (bytenr > entry->bytenr)
303 p = &(*p)->rb_right;
304 else
305 return parent;
306 }
307
308 rb_link_node(node, parent, p);
309 rb_insert_color(node, root);
310 return NULL;
311}
312
313static struct rb_node *tree_search(struct rb_root *root, u64 bytenr)
314{
315 struct rb_node *n = root->rb_node;
316 struct tree_entry *entry;
317
318 while (n) {
319 entry = rb_entry(n, struct tree_entry, rb_node);
320
321 if (bytenr < entry->bytenr)
322 n = n->rb_left;
323 else if (bytenr > entry->bytenr)
324 n = n->rb_right;
325 else
326 return n;
327 }
328 return NULL;
329}
330
48a3b636 331static void backref_tree_panic(struct rb_node *rb_node, int errno, u64 bytenr)
43c04fb1
JM		 332{
333
334 struct btrfs_fs_info *fs_info = NULL;
335 struct backref_node *bnode = rb_entry(rb_node, struct backref_node,
336 rb_node);
337 if (bnode->root)
338 fs_info = bnode->root->fs_info;
339 btrfs_panic(fs_info, errno, "Inconsistency in backref cache "
351fd353 340 "found at offset %llu", bytenr);
43c04fb1
JM		 341}
342
5d4f98a2
YZ		 343/*
344 * walk up backref nodes until reach node presents tree root
345 */
346static struct backref_node *walk_up_backref(struct backref_node *node,
347 struct backref_edge *edges[],
348 int *index)
349{
350 struct backref_edge *edge;
351 int idx = *index;
352
353 while (!list_empty(&node->upper)) {
354 edge = list_entry(node->upper.next,
355 struct backref_edge, list[LOWER]);
356 edges[idx++] = edge;
357 node = edge->node[UPPER];
358 }
3fd0a558 359 BUG_ON(node->detached);
5d4f98a2
YZ		 360 *index = idx;
361 return node;
362}
363
364/*
365 * walk down backref nodes to find start of next reference path
366 */
367static struct backref_node *walk_down_backref(struct backref_edge *edges[],
368 int *index)
369{
370 struct backref_edge *edge;
371 struct backref_node *lower;
372 int idx = *index;
373
374 while (idx > 0) {
375 edge = edges[idx - 1];
376 lower = edge->node[LOWER];
377 if (list_is_last(&edge->list[LOWER], &lower->upper)) {
378 idx--;
379 continue;
380 }
381 edge = list_entry(edge->list[LOWER].next,
382 struct backref_edge, list[LOWER]);
383 edges[idx - 1] = edge;
384 *index = idx;
385 return edge->node[UPPER];
386 }
387 *index = 0;
388 return NULL;
389}
390
3fd0a558
YZ		 391static void unlock_node_buffer(struct backref_node *node)
392{
393 if (node->locked) {
394 btrfs_tree_unlock(node->eb);
395 node->locked = 0;
396 }
397}
398
5d4f98a2
YZ		 399static void drop_node_buffer(struct backref_node *node)
400{
401 if (node->eb) {
3fd0a558 402 unlock_node_buffer(node);
5d4f98a2
YZ		 403 free_extent_buffer(node->eb);
404 node->eb = NULL;
405 }
406}
407
408static void drop_backref_node(struct backref_cache *tree,
409 struct backref_node *node)
410{
5d4f98a2
YZ		 411 BUG_ON(!list_empty(&node->upper));
412
413 drop_node_buffer(node);
3fd0a558 414 list_del(&node->list);
5d4f98a2 415 list_del(&node->lower);
3fd0a558
YZ		 416 if (!RB_EMPTY_NODE(&node->rb_node))
417 rb_erase(&node->rb_node, &tree->rb_root);
418 free_backref_node(tree, node);
5d4f98a2
YZ		 419}
420
421/*
422 * remove a backref node from the backref cache
423 */
424static void remove_backref_node(struct backref_cache *cache,
425 struct backref_node *node)
426{
427 struct backref_node *upper;
428 struct backref_edge *edge;
429
430 if (!node)
431 return;
432
3fd0a558 433 BUG_ON(!node->lowest && !node->detached);
5d4f98a2
YZ		 434 while (!list_empty(&node->upper)) {
435 edge = list_entry(node->upper.next, struct backref_edge,
436 list[LOWER]);
437 upper = edge->node[UPPER];
438 list_del(&edge->list[LOWER]);
439 list_del(&edge->list[UPPER]);
3fd0a558
YZ		 440 free_backref_edge(cache, edge);
441
442 if (RB_EMPTY_NODE(&upper->rb_node)) {
443 BUG_ON(!list_empty(&node->upper));
444 drop_backref_node(cache, node);
445 node = upper;
446 node->lowest = 1;
447 continue;
448 }
5d4f98a2 449 /*
3fd0a558 450 * add the node to leaf node list if no other
5d4f98a2
YZ		 451 * child block cached.
452 */
453 if (list_empty(&upper->lower)) {
3fd0a558 454 list_add_tail(&upper->lower, &cache->leaves);
5d4f98a2
YZ		 455 upper->lowest = 1;
456 }
457 }
3fd0a558 458
5d4f98a2
YZ		 459 drop_backref_node(cache, node);
460}
461
3fd0a558
YZ		 462static void update_backref_node(struct backref_cache *cache,
463 struct backref_node *node, u64 bytenr)
464{
465 struct rb_node *rb_node;
466 rb_erase(&node->rb_node, &cache->rb_root);
467 node->bytenr = bytenr;
468 rb_node = tree_insert(&cache->rb_root, node->bytenr, &node->rb_node);
43c04fb1
JM		 469 if (rb_node)
470 backref_tree_panic(rb_node, -EEXIST, bytenr);
3fd0a558
YZ		 471}
472
473/*
474 * update backref cache after a transaction commit
475 */
476static int update_backref_cache(struct btrfs_trans_handle *trans,
477 struct backref_cache *cache)
478{
479 struct backref_node *node;
480 int level = 0;
481
482 if (cache->last_trans == 0) {
483 cache->last_trans = trans->transid;
484 return 0;
485 }
486
487 if (cache->last_trans == trans->transid)
488 return 0;
489
490 /*
491 * detached nodes are used to avoid unnecessary backref
492 * lookup. transaction commit changes the extent tree.
493 * so the detached nodes are no longer useful.
494 */
495 while (!list_empty(&cache->detached)) {
496 node = list_entry(cache->detached.next,
497 struct backref_node, list);
498 remove_backref_node(cache, node);
499 }
500
501 while (!list_empty(&cache->changed)) {
502 node = list_entry(cache->changed.next,
503 struct backref_node, list);
504 list_del_init(&node->list);
505 BUG_ON(node->pending);
506 update_backref_node(cache, node, node->new_bytenr);
507 }
508
509 /*
510 * some nodes can be left in the pending list if there were
511 * errors during processing the pending nodes.
512 */
513 for (level = 0; level < BTRFS_MAX_LEVEL; level++) {
514 list_for_each_entry(node, &cache->pending[level], list) {
515 BUG_ON(!node->pending);
516 if (node->bytenr == node->new_bytenr)
517 continue;
518 update_backref_node(cache, node, node->new_bytenr);
519 }
520 }
521
522 cache->last_trans = 0;
523 return 1;
524}
525
f2a97a9d 526
3fd0a558
YZ		 527static int should_ignore_root(struct btrfs_root *root)
528{
529 struct btrfs_root *reloc_root;
530
27cdeb70 531 if (!test_bit(BTRFS_ROOT_REF_COWS, &root->state))
3fd0a558
YZ		 532 return 0;
533
534 reloc_root = root->reloc_root;
535 if (!reloc_root)
536 return 0;
537
538 if (btrfs_root_last_snapshot(&reloc_root->root_item) ==
539 root->fs_info->running_transaction->transid - 1)
540 return 0;
541 /*
542 * if there is reloc tree and it was created in previous
543 * transaction backref lookup can find the reloc tree,
544 * so backref node for the fs tree root is useless for
545 * relocation.
546 */
547 return 1;
548}
5d4f98a2
YZ		 549/*
550 * find reloc tree by address of tree root
551 */
552static struct btrfs_root *find_reloc_root(struct reloc_control *rc,
553 u64 bytenr)
554{
555 struct rb_node *rb_node;
556 struct mapping_node *node;
557 struct btrfs_root *root = NULL;
558
559 spin_lock(&rc->reloc_root_tree.lock);
560 rb_node = tree_search(&rc->reloc_root_tree.rb_root, bytenr);
561 if (rb_node) {
562 node = rb_entry(rb_node, struct mapping_node, rb_node);
563 root = (struct btrfs_root *)node->data;
564 }
565 spin_unlock(&rc->reloc_root_tree.lock);
566 return root;
567}
568
569static int is_cowonly_root(u64 root_objectid)
570{
571 if (root_objectid == BTRFS_ROOT_TREE_OBJECTID ||
572 root_objectid == BTRFS_EXTENT_TREE_OBJECTID ||
573 root_objectid == BTRFS_CHUNK_TREE_OBJECTID ||
574 root_objectid == BTRFS_DEV_TREE_OBJECTID ||
575 root_objectid == BTRFS_TREE_LOG_OBJECTID ||
66463748
WS		 576 root_objectid == BTRFS_CSUM_TREE_OBJECTID ||
577 root_objectid == BTRFS_UUID_TREE_OBJECTID ||
578 root_objectid == BTRFS_QUOTA_TREE_OBJECTID)
5d4f98a2
YZ		 579 return 1;
580 return 0;
581}
582
583static struct btrfs_root *read_fs_root(struct btrfs_fs_info *fs_info,
584 u64 root_objectid)
585{
586 struct btrfs_key key;
587
588 key.objectid = root_objectid;
589 key.type = BTRFS_ROOT_ITEM_KEY;
590 if (is_cowonly_root(root_objectid))
591 key.offset = 0;
592 else
593 key.offset = (u64)-1;
594
c00869f1 595 return btrfs_get_fs_root(fs_info, &key, false);
5d4f98a2
YZ		 596}
597
598#ifdef BTRFS_COMPAT_EXTENT_TREE_V0
599static noinline_for_stack
600struct btrfs_root *find_tree_root(struct reloc_control *rc,
601 struct extent_buffer *leaf,
602 struct btrfs_extent_ref_v0 *ref0)
603{
604 struct btrfs_root *root;
605 u64 root_objectid = btrfs_ref_root_v0(leaf, ref0);
606 u64 generation = btrfs_ref_generation_v0(leaf, ref0);
607
608 BUG_ON(root_objectid == BTRFS_TREE_RELOC_OBJECTID);
609
610 root = read_fs_root(rc->extent_root->fs_info, root_objectid);
611 BUG_ON(IS_ERR(root));
612
27cdeb70 613 if (test_bit(BTRFS_ROOT_REF_COWS, &root->state) &&
5d4f98a2
YZ		 614 generation != btrfs_root_generation(&root->root_item))
615 return NULL;
616
617 return root;
618}
619#endif
620
621static noinline_for_stack
622int find_inline_backref(struct extent_buffer *leaf, int slot,
623 unsigned long *ptr, unsigned long *end)
624{
3173a18f 625 struct btrfs_key key;
5d4f98a2
YZ		 626 struct btrfs_extent_item *ei;
627 struct btrfs_tree_block_info *bi;
628 u32 item_size;
629
3173a18f
JB		 630 btrfs_item_key_to_cpu(leaf, &key, slot);
631
5d4f98a2
YZ		 632 item_size = btrfs_item_size_nr(leaf, slot);
633#ifdef BTRFS_COMPAT_EXTENT_TREE_V0
634 if (item_size < sizeof(*ei)) {
635 WARN_ON(item_size != sizeof(struct btrfs_extent_item_v0));
636 return 1;
637 }
638#endif
639 ei = btrfs_item_ptr(leaf, slot, struct btrfs_extent_item);
640 WARN_ON(!(btrfs_extent_flags(leaf, ei) &
641 BTRFS_EXTENT_FLAG_TREE_BLOCK));
642
3173a18f
JB		 643 if (key.type == BTRFS_EXTENT_ITEM_KEY &&
644 item_size <= sizeof(*ei) + sizeof(*bi)) {
5d4f98a2
YZ		 645 WARN_ON(item_size < sizeof(*ei) + sizeof(*bi));
646 return 1;
647 }
d062d13c
JB		 648 if (key.type == BTRFS_METADATA_ITEM_KEY &&
649 item_size <= sizeof(*ei)) {
650 WARN_ON(item_size < sizeof(*ei));
651 return 1;
652 }
5d4f98a2 653
3173a18f
JB		 654 if (key.type == BTRFS_EXTENT_ITEM_KEY) {
655 bi = (struct btrfs_tree_block_info *)(ei + 1);
656 *ptr = (unsigned long)(bi + 1);
657 } else {
658 *ptr = (unsigned long)(ei + 1);
659 }
5d4f98a2
YZ		 660 *end = (unsigned long)ei + item_size;
661 return 0;
662}
663
664/*
665 * build backref tree for a given tree block. root of the backref tree
666 * corresponds the tree block, leaves of the backref tree correspond
667 * roots of b-trees that reference the tree block.
668 *
669 * the basic idea of this function is check backrefs of a given block
670 * to find upper level blocks that refernece the block, and then check
671 * bakcrefs of these upper level blocks recursively. the recursion stop
672 * when tree root is reached or backrefs for the block is cached.
673 *
674 * NOTE: if we find backrefs for a block are cached, we know backrefs
675 * for all upper level blocks that directly/indirectly reference the
676 * block are also cached.
677 */
3fd0a558
YZ		 678static noinline_for_stack
679struct backref_node *build_backref_tree(struct reloc_control *rc,
680 struct btrfs_key *node_key,
681 int level, u64 bytenr)
5d4f98a2 682{
3fd0a558 683 struct backref_cache *cache = &rc->backref_cache;
5d4f98a2
YZ		 684 struct btrfs_path *path1;
685 struct btrfs_path *path2;
686 struct extent_buffer *eb;
687 struct btrfs_root *root;
688 struct backref_node *cur;
689 struct backref_node *upper;
690 struct backref_node *lower;
691 struct backref_node *node = NULL;
692 struct backref_node *exist = NULL;
693 struct backref_edge *edge;
694 struct rb_node *rb_node;
695 struct btrfs_key key;
696 unsigned long end;
697 unsigned long ptr;
698 LIST_HEAD(list);
3fd0a558
YZ		 699 LIST_HEAD(useless);
700 int cowonly;
5d4f98a2
YZ		 701 int ret;
702 int err = 0;
b6c60c80 703 bool need_check = true;
5d4f98a2
YZ		 704
705 path1 = btrfs_alloc_path();
706 path2 = btrfs_alloc_path();
707 if (!path1 || !path2) {
708 err = -ENOMEM;
709 goto out;
710 }
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 }
75bfb9af
JB		 739 ASSERT(ret);
740 ASSERT(path1->slots[0]);
5d4f98a2
YZ		 741
742 path1->slots[0]--;
743
744 WARN_ON(cur->checked);
745 if (!list_empty(&cur->upper)) {
746 /*
70f23fd6 747 * the backref was added previously when processing
5d4f98a2
YZ		 748 * backref of type BTRFS_TREE_BLOCK_REF_KEY
749 */
75bfb9af 750 ASSERT(list_is_singular(&cur->upper));
5d4f98a2
YZ		 751 edge = list_entry(cur->upper.next, struct backref_edge,
752 list[LOWER]);
75bfb9af 753 ASSERT(list_empty(&edge->list[UPPER]));
5d4f98a2
YZ		 754 exist = edge->node[UPPER];
755 /*
756 * add the upper level block to pending list if we need
757 * check its backrefs
758 */
759 if (!exist->checked)
760 list_add_tail(&edge->list[UPPER], &list);
761 } else {
762 exist = NULL;
763 }
764
765 while (1) {
766 cond_resched();
767 eb = path1->nodes[0];
768
769 if (ptr >= end) {
770 if (path1->slots[0] >= btrfs_header_nritems(eb)) {
771 ret = btrfs_next_leaf(rc->extent_root, path1);
772 if (ret < 0) {
773 err = ret;
774 goto out;
775 }
776 if (ret > 0)
777 break;
778 eb = path1->nodes[0];
779 }
780
781 btrfs_item_key_to_cpu(eb, &key, path1->slots[0]);
782 if (key.objectid != cur->bytenr) {
783 WARN_ON(exist);
784 break;
785 }
786
3173a18f
JB		 787 if (key.type == BTRFS_EXTENT_ITEM_KEY ||
788 key.type == BTRFS_METADATA_ITEM_KEY) {
5d4f98a2
YZ		 789 ret = find_inline_backref(eb, path1->slots[0],
790 &ptr, &end);
791 if (ret)
792 goto next;
793 }
794 }
795
796 if (ptr < end) {
797 /* update key for inline back ref */
798 struct btrfs_extent_inline_ref *iref;
799 iref = (struct btrfs_extent_inline_ref *)ptr;
800 key.type = btrfs_extent_inline_ref_type(eb, iref);
801 key.offset = btrfs_extent_inline_ref_offset(eb, iref);
802 WARN_ON(key.type != BTRFS_TREE_BLOCK_REF_KEY &&
803 key.type != BTRFS_SHARED_BLOCK_REF_KEY);
804 }
805
806 if (exist &&
807 ((key.type == BTRFS_TREE_BLOCK_REF_KEY &&
808 exist->owner == key.offset) ||
809 (key.type == BTRFS_SHARED_BLOCK_REF_KEY &&
810 exist->bytenr == key.offset))) {
811 exist = NULL;
812 goto next;
813 }
814
815#ifdef BTRFS_COMPAT_EXTENT_TREE_V0
816 if (key.type == BTRFS_SHARED_BLOCK_REF_KEY ||
817 key.type == BTRFS_EXTENT_REF_V0_KEY) {
3fd0a558 818 if (key.type == BTRFS_EXTENT_REF_V0_KEY) {
5d4f98a2
YZ		 819 struct btrfs_extent_ref_v0 *ref0;
820 ref0 = btrfs_item_ptr(eb, path1->slots[0],
821 struct btrfs_extent_ref_v0);
3fd0a558 822 if (key.objectid == key.offset) {
046f264f 823 root = find_tree_root(rc, eb, ref0);
3fd0a558
YZ		 824 if (root && !should_ignore_root(root))
825 cur->root = root;
826 else
827 list_add(&cur->list, &useless);
828 break;
829 }
046f264f
YZ		 830 if (is_cowonly_root(btrfs_ref_root_v0(eb,
831 ref0)))
832 cur->cowonly = 1;
5d4f98a2
YZ		 833 }
834#else
75bfb9af 835 ASSERT(key.type != BTRFS_EXTENT_REF_V0_KEY);
5d4f98a2
YZ		 836 if (key.type == BTRFS_SHARED_BLOCK_REF_KEY) {
837#endif
838 if (key.objectid == key.offset) {
839 /*
840 * only root blocks of reloc trees use
841 * backref of this type.
842 */
843 root = find_reloc_root(rc, cur->bytenr);
75bfb9af 844 ASSERT(root);
5d4f98a2
YZ		 845 cur->root = root;
846 break;
847 }
848
3fd0a558 849 edge = alloc_backref_edge(cache);
5d4f98a2
YZ		 850 if (!edge) {
851 err = -ENOMEM;
852 goto out;
853 }
854 rb_node = tree_search(&cache->rb_root, key.offset);
855 if (!rb_node) {
3fd0a558 856 upper = alloc_backref_node(cache);
5d4f98a2 857 if (!upper) {
3fd0a558 858 free_backref_edge(cache, edge);
5d4f98a2
YZ		 859 err = -ENOMEM;
860 goto out;
861 }
5d4f98a2 862 upper->bytenr = key.offset;
5d4f98a2
YZ		 863 upper->level = cur->level + 1;
864 /*
865 * backrefs for the upper level block isn't
866 * cached, add the block to pending list
867 */
868 list_add_tail(&edge->list[UPPER], &list);
869 } else {
870 upper = rb_entry(rb_node, struct backref_node,
871 rb_node);
75bfb9af 872 ASSERT(upper->checked);
5d4f98a2
YZ		 873 INIT_LIST_HEAD(&edge->list[UPPER]);
874 }
3fd0a558 875 list_add_tail(&edge->list[LOWER], &cur->upper);
5d4f98a2 876 edge->node[LOWER] = cur;
3fd0a558 877 edge->node[UPPER] = upper;
5d4f98a2
YZ		 878
879 goto next;
880 } else if (key.type != BTRFS_TREE_BLOCK_REF_KEY) {
881 goto next;
882 }
883
884 /* key.type == BTRFS_TREE_BLOCK_REF_KEY */
885 root = read_fs_root(rc->extent_root->fs_info, key.offset);
886 if (IS_ERR(root)) {
887 err = PTR_ERR(root);
888 goto out;
889 }
890
27cdeb70 891 if (!test_bit(BTRFS_ROOT_REF_COWS, &root->state))
3fd0a558
YZ		 892 cur->cowonly = 1;
893
5d4f98a2
YZ		 894 if (btrfs_root_level(&root->root_item) == cur->level) {
895 /* tree root */
75bfb9af 896 ASSERT(btrfs_root_bytenr(&root->root_item) ==
5d4f98a2 897 cur->bytenr);
3fd0a558
YZ		 898 if (should_ignore_root(root))
899 list_add(&cur->list, &useless);
900 else
901 cur->root = root;
5d4f98a2
YZ		 902 break;
903 }
904
905 level = cur->level + 1;
906
907 /*
908 * searching the tree to find upper level blocks
909 * reference the block.
910 */
911 path2->search_commit_root = 1;
912 path2->skip_locking = 1;
913 path2->lowest_level = level;
914 ret = btrfs_search_slot(NULL, root, node_key, path2, 0, 0);
915 path2->lowest_level = 0;
916 if (ret < 0) {
917 err = ret;
918 goto out;
919 }
33c66f43
YZ		 920 if (ret > 0 && path2->slots[level] > 0)
921 path2->slots[level]--;
5d4f98a2
YZ		 922
923 eb = path2->nodes[level];
924 WARN_ON(btrfs_node_blockptr(eb, path2->slots[level]) !=
925 cur->bytenr);
926
927 lower = cur;
b6c60c80 928 need_check = true;
5d4f98a2
YZ		 929 for (; level < BTRFS_MAX_LEVEL; level++) {
930 if (!path2->nodes[level]) {
75bfb9af 931 ASSERT(btrfs_root_bytenr(&root->root_item) ==
5d4f98a2 932 lower->bytenr);
3fd0a558
YZ		 933 if (should_ignore_root(root))
934 list_add(&lower->list, &useless);
935 else
936 lower->root = root;
5d4f98a2
YZ		 937 break;
938 }
939
3fd0a558 940 edge = alloc_backref_edge(cache);
5d4f98a2
YZ		 941 if (!edge) {
942 err = -ENOMEM;
943 goto out;
944 }
945
946 eb = path2->nodes[level];
947 rb_node = tree_search(&cache->rb_root, eb->start);
948 if (!rb_node) {
3fd0a558 949 upper = alloc_backref_node(cache);
5d4f98a2 950 if (!upper) {
3fd0a558 951 free_backref_edge(cache, edge);
5d4f98a2
YZ		 952 err = -ENOMEM;
953 goto out;
954 }
5d4f98a2
YZ		 955 upper->bytenr = eb->start;
956 upper->owner = btrfs_header_owner(eb);
957 upper->level = lower->level + 1;
27cdeb70
MX		 958 if (!test_bit(BTRFS_ROOT_REF_COWS,
959 &root->state))
3fd0a558 960 upper->cowonly = 1;
5d4f98a2
YZ		 961
962 /*
963 * if we know the block isn't shared
964 * we can void checking its backrefs.
965 */
966 if (btrfs_block_can_be_shared(root, eb))
967 upper->checked = 0;
968 else
969 upper->checked = 1;
970
971 /*
972 * add the block to pending list if we
b6c60c80
JB		 973 * need check its backrefs, we only do this once
974 * while walking up a tree as we will catch
975 * anything else later on.
5d4f98a2 976 */
b6c60c80
JB		 977 if (!upper->checked && need_check) {
978 need_check = false;
5d4f98a2
YZ		 979 list_add_tail(&edge->list[UPPER],
980 &list);
bbe90514
JB		 981 } else {
982 if (upper->checked)
983 need_check = true;
5d4f98a2 984 INIT_LIST_HEAD(&edge->list[UPPER]);
bbe90514 985 }
5d4f98a2
YZ		 986 } else {
987 upper = rb_entry(rb_node, struct backref_node,
988 rb_node);
75bfb9af 989 ASSERT(upper->checked);
5d4f98a2 990 INIT_LIST_HEAD(&edge->list[UPPER]);
3fd0a558
YZ		 991 if (!upper->owner)
992 upper->owner = btrfs_header_owner(eb);
5d4f98a2
YZ		 993 }
994 list_add_tail(&edge->list[LOWER], &lower->upper);
5d4f98a2 995 edge->node[LOWER] = lower;
3fd0a558 996 edge->node[UPPER] = upper;
5d4f98a2
YZ		 997
998 if (rb_node)
999 break;
1000 lower = upper;
1001 upper = NULL;
1002 }
b3b4aa74 1003 btrfs_release_path(path2);
5d4f98a2
YZ		 1004next:
1005 if (ptr < end) {
1006 ptr += btrfs_extent_inline_ref_size(key.type);
1007 if (ptr >= end) {
1008 WARN_ON(ptr > end);
1009 ptr = 0;
1010 end = 0;
1011 }
1012 }
1013 if (ptr >= end)
1014 path1->slots[0]++;
1015 }
b3b4aa74 1016 btrfs_release_path(path1);
5d4f98a2
YZ		 1017
1018 cur->checked = 1;
1019 WARN_ON(exist);
1020
1021 /* the pending list isn't empty, take the first block to process */
1022 if (!list_empty(&list)) {
1023 edge = list_entry(list.next, struct backref_edge, list[UPPER]);
1024 list_del_init(&edge->list[UPPER]);
1025 cur = edge->node[UPPER];
1026 goto again;
1027 }
1028
1029 /*
1030 * everything goes well, connect backref nodes and insert backref nodes
1031 * into the cache.
1032 */
75bfb9af 1033 ASSERT(node->checked);
3fd0a558
YZ		 1034 cowonly = node->cowonly;
1035 if (!cowonly) {
1036 rb_node = tree_insert(&cache->rb_root, node->bytenr,
1037 &node->rb_node);
43c04fb1
JM		 1038 if (rb_node)
1039 backref_tree_panic(rb_node, -EEXIST, node->bytenr);
3fd0a558
YZ		 1040 list_add_tail(&node->lower, &cache->leaves);
1041 }
5d4f98a2
YZ		 1042
1043 list_for_each_entry(edge, &node->upper, list[LOWER])
1044 list_add_tail(&edge->list[UPPER], &list);
1045
1046 while (!list_empty(&list)) {
1047 edge = list_entry(list.next, struct backref_edge, list[UPPER]);
1048 list_del_init(&edge->list[UPPER]);
1049 upper = edge->node[UPPER];
3fd0a558
YZ		 1050 if (upper->detached) {
1051 list_del(&edge->list[LOWER]);
1052 lower = edge->node[LOWER];
1053 free_backref_edge(cache, edge);
1054 if (list_empty(&lower->upper))
1055 list_add(&lower->list, &useless);
1056 continue;
1057 }
5d4f98a2
YZ		 1058
1059 if (!RB_EMPTY_NODE(&upper->rb_node)) {
1060 if (upper->lowest) {
1061 list_del_init(&upper->lower);
1062 upper->lowest = 0;
1063 }
1064
1065 list_add_tail(&edge->list[UPPER], &upper->lower);
1066 continue;
1067 }
1068
75bfb9af
JB		 1069 if (!upper->checked) {
1070 /*
1071 * Still want to blow up for developers since this is a
1072 * logic bug.
1073 */
1074 ASSERT(0);
1075 err = -EINVAL;
1076 goto out;
1077 }
1078 if (cowonly != upper->cowonly) {
1079 ASSERT(0);
1080 err = -EINVAL;
1081 goto out;
1082 }
1083
3fd0a558
YZ		 1084 if (!cowonly) {
1085 rb_node = tree_insert(&cache->rb_root, upper->bytenr,
1086 &upper->rb_node);
43c04fb1
JM		 1087 if (rb_node)
1088 backref_tree_panic(rb_node, -EEXIST,
1089 upper->bytenr);
3fd0a558 1090 }
5d4f98a2
YZ		 1091
1092 list_add_tail(&edge->list[UPPER], &upper->lower);
1093
1094 list_for_each_entry(edge, &upper->upper, list[LOWER])
1095 list_add_tail(&edge->list[UPPER], &list);
1096 }
3fd0a558
YZ		 1097 /*
1098 * process useless backref nodes. backref nodes for tree leaves
1099 * are deleted from the cache. backref nodes for upper level
1100 * tree blocks are left in the cache to avoid unnecessary backref
1101 * lookup.
1102 */
1103 while (!list_empty(&useless)) {
1104 upper = list_entry(useless.next, struct backref_node, list);
1105 list_del_init(&upper->list);
75bfb9af 1106 ASSERT(list_empty(&upper->upper));
3fd0a558
YZ		 1107 if (upper == node)
1108 node = NULL;
1109 if (upper->lowest) {
1110 list_del_init(&upper->lower);
1111 upper->lowest = 0;
1112 }
1113 while (!list_empty(&upper->lower)) {
1114 edge = list_entry(upper->lower.next,
1115 struct backref_edge, list[UPPER]);
1116 list_del(&edge->list[UPPER]);
1117 list_del(&edge->list[LOWER]);
1118 lower = edge->node[LOWER];
1119 free_backref_edge(cache, edge);
1120
1121 if (list_empty(&lower->upper))
1122 list_add(&lower->list, &useless);
1123 }
1124 __mark_block_processed(rc, upper);
1125 if (upper->level > 0) {
1126 list_add(&upper->list, &cache->detached);
1127 upper->detached = 1;
1128 } else {
1129 rb_erase(&upper->rb_node, &cache->rb_root);
1130 free_backref_node(cache, upper);
1131 }
1132 }
5d4f98a2
YZ		 1133out:
1134 btrfs_free_path(path1);
1135 btrfs_free_path(path2);
1136 if (err) {
3fd0a558
YZ		 1137 while (!list_empty(&useless)) {
1138 lower = list_entry(useless.next,
75bfb9af
JB		 1139 struct backref_node, list);
1140 list_del_init(&lower->list);
3fd0a558 1141 }
75bfb9af
JB		 1142 while (!list_empty(&list)) {
1143 edge = list_first_entry(&list, struct backref_edge,
1144 list[UPPER]);
1145 list_del(&edge->list[UPPER]);
3fd0a558 1146 list_del(&edge->list[LOWER]);
75bfb9af 1147 lower = edge->node[LOWER];
5d4f98a2 1148 upper = edge->node[UPPER];
3fd0a558 1149 free_backref_edge(cache, edge);
75bfb9af
JB		 1150
1151 /*
1152 * Lower is no longer linked to any upper backref nodes
1153 * and isn't in the cache, we can free it ourselves.
1154 */
1155 if (list_empty(&lower->upper) &&
1156 RB_EMPTY_NODE(&lower->rb_node))
1157 list_add(&lower->list, &useless);
1158
1159 if (!RB_EMPTY_NODE(&upper->rb_node))
1160 continue;
1161
1162 /* Add this guy's upper edges to the list to proces */
1163 list_for_each_entry(edge, &upper->upper, list[LOWER])
1164 list_add_tail(&edge->list[UPPER], &list);
1165 if (list_empty(&upper->upper))
1166 list_add(&upper->list, &useless);
1167 }
1168
1169 while (!list_empty(&useless)) {
1170 lower = list_entry(useless.next,
1171 struct backref_node, list);
1172 list_del_init(&lower->list);
1173 free_backref_node(cache, lower);
5d4f98a2
YZ		 1174 }
1175 return ERR_PTR(err);
1176 }
75bfb9af 1177 ASSERT(!node || !node->detached);
5d4f98a2
YZ		 1178 return node;
1179}
1180
3fd0a558
YZ		 1181/*
1182 * helper to add backref node for the newly created snapshot.
1183 * the backref node is created by cloning backref node that
1184 * corresponds to root of source tree
1185 */
1186static int clone_backref_node(struct btrfs_trans_handle *trans,
1187 struct reloc_control *rc,
1188 struct btrfs_root *src,
1189 struct btrfs_root *dest)
1190{
1191 struct btrfs_root *reloc_root = src->reloc_root;
1192 struct backref_cache *cache = &rc->backref_cache;
1193 struct backref_node *node = NULL;
1194 struct backref_node *new_node;
1195 struct backref_edge *edge;
1196 struct backref_edge *new_edge;
1197 struct rb_node *rb_node;
1198
1199 if (cache->last_trans > 0)
1200 update_backref_cache(trans, cache);
1201
1202 rb_node = tree_search(&cache->rb_root, src->commit_root->start);
1203 if (rb_node) {
1204 node = rb_entry(rb_node, struct backref_node, rb_node);
1205 if (node->detached)
1206 node = NULL;
1207 else
1208 BUG_ON(node->new_bytenr != reloc_root->node->start);
1209 }
1210
1211 if (!node) {
1212 rb_node = tree_search(&cache->rb_root,
1213 reloc_root->commit_root->start);
1214 if (rb_node) {
1215 node = rb_entry(rb_node, struct backref_node,
1216 rb_node);
1217 BUG_ON(node->detached);
1218 }
1219 }
1220
1221 if (!node)
1222 return 0;
1223
1224 new_node = alloc_backref_node(cache);
1225 if (!new_node)
1226 return -ENOMEM;
1227
1228 new_node->bytenr = dest->node->start;
1229 new_node->level = node->level;
1230 new_node->lowest = node->lowest;
6848ad64 1231 new_node->checked = 1;
3fd0a558
YZ		 1232 new_node->root = dest;
1233
1234 if (!node->lowest) {
1235 list_for_each_entry(edge, &node->lower, list[UPPER]) {
1236 new_edge = alloc_backref_edge(cache);
1237 if (!new_edge)
1238 goto fail;
1239
1240 new_edge->node[UPPER] = new_node;
1241 new_edge->node[LOWER] = edge->node[LOWER];
1242 list_add_tail(&new_edge->list[UPPER],
1243 &new_node->lower);
1244 }
76b9e23d
MX		 1245 } else {
1246 list_add_tail(&new_node->lower, &cache->leaves);
3fd0a558
YZ		 1247 }
1248
1249 rb_node = tree_insert(&cache->rb_root, new_node->bytenr,
1250 &new_node->rb_node);
43c04fb1
JM		 1251 if (rb_node)
1252 backref_tree_panic(rb_node, -EEXIST, new_node->bytenr);
3fd0a558
YZ		 1253
1254 if (!new_node->lowest) {
1255 list_for_each_entry(new_edge, &new_node->lower, list[UPPER]) {
1256 list_add_tail(&new_edge->list[LOWER],
1257 &new_edge->node[LOWER]->upper);
1258 }
1259 }
1260 return 0;
1261fail:
1262 while (!list_empty(&new_node->lower)) {
1263 new_edge = list_entry(new_node->lower.next,
1264 struct backref_edge, list[UPPER]);
1265 list_del(&new_edge->list[UPPER]);
1266 free_backref_edge(cache, new_edge);
1267 }
1268 free_backref_node(cache, new_node);
1269 return -ENOMEM;
1270}
1271
5d4f98a2
YZ		 1272/*
1273 * helper to add 'address of tree root -> reloc tree' mapping
1274 */
ffd7b339 1275static int __must_check __add_reloc_root(struct btrfs_root *root)
5d4f98a2
YZ		 1276{
1277 struct rb_node *rb_node;
1278 struct mapping_node *node;
1279 struct reloc_control *rc = root->fs_info->reloc_ctl;
1280
1281 node = kmalloc(sizeof(*node), GFP_NOFS);
ffd7b339
JM		 1282 if (!node)
1283 return -ENOMEM;
5d4f98a2
YZ		 1284
1285 node->bytenr = root->node->start;
1286 node->data = root;
1287
1288 spin_lock(&rc->reloc_root_tree.lock);
1289 rb_node = tree_insert(&rc->reloc_root_tree.rb_root,
1290 node->bytenr, &node->rb_node);
1291 spin_unlock(&rc->reloc_root_tree.lock);
ffd7b339 1292 if (rb_node) {
ffd7b339
JM		 1293 btrfs_panic(root->fs_info, -EEXIST, "Duplicate root found "
1294 "for start=%llu while inserting into relocation "
351fd353 1295 "tree", node->bytenr);
23291a04
DC		 1296 kfree(node);
1297 return -EEXIST;
ffd7b339 1298 }
5d4f98a2
YZ		 1299
1300 list_add_tail(&root->root_list, &rc->reloc_roots);
1301 return 0;
1302}
1303
1304/*
c974c464 1305 * helper to delete the 'address of tree root -> reloc tree'
5d4f98a2
YZ		 1306 * mapping
1307 */
c974c464 1308static void __del_reloc_root(struct btrfs_root *root)
5d4f98a2
YZ		 1309{
1310 struct rb_node *rb_node;
1311 struct mapping_node *node = NULL;
1312 struct reloc_control *rc = root->fs_info->reloc_ctl;
1313
1314 spin_lock(&rc->reloc_root_tree.lock);
1315 rb_node = tree_search(&rc->reloc_root_tree.rb_root,
c974c464 1316 root->node->start);
5d4f98a2
YZ		 1317 if (rb_node) {
1318 node = rb_entry(rb_node, struct mapping_node, rb_node);
1319 rb_erase(&node->rb_node, &rc->reloc_root_tree.rb_root);
1320 }
1321 spin_unlock(&rc->reloc_root_tree.lock);
1322
8f71f3e0 1323 if (!node)
c974c464 1324 return;
5d4f98a2
YZ		 1325 BUG_ON((struct btrfs_root *)node->data != root);
1326
c974c464
WS		 1327 spin_lock(&root->fs_info->trans_lock);
1328 list_del_init(&root->root_list);
1329 spin_unlock(&root->fs_info->trans_lock);
1330 kfree(node);
1331}
1332
1333/*
1334 * helper to update the 'address of tree root -> reloc tree'
1335 * mapping
1336 */
1337static int __update_reloc_root(struct btrfs_root *root, u64 new_bytenr)
1338{
1339 struct rb_node *rb_node;
1340 struct mapping_node *node = NULL;
1341 struct reloc_control *rc = root->fs_info->reloc_ctl;
1342
1343 spin_lock(&rc->reloc_root_tree.lock);
1344 rb_node = tree_search(&rc->reloc_root_tree.rb_root,
1345 root->node->start);
1346 if (rb_node) {
1347 node = rb_entry(rb_node, struct mapping_node, rb_node);
1348 rb_erase(&node->rb_node, &rc->reloc_root_tree.rb_root);
5d4f98a2 1349 }
c974c464
WS		 1350 spin_unlock(&rc->reloc_root_tree.lock);
1351
1352 if (!node)
1353 return 0;
1354 BUG_ON((struct btrfs_root *)node->data != root);
1355
1356 spin_lock(&rc->reloc_root_tree.lock);
1357 node->bytenr = new_bytenr;
1358 rb_node = tree_insert(&rc->reloc_root_tree.rb_root,
1359 node->bytenr, &node->rb_node);
1360 spin_unlock(&rc->reloc_root_tree.lock);
1361 if (rb_node)
1362 backref_tree_panic(rb_node, -EEXIST, node->bytenr);
5d4f98a2
YZ		 1363 return 0;
1364}
1365
3fd0a558
YZ		 1366static struct btrfs_root *create_reloc_root(struct btrfs_trans_handle *trans,
1367 struct btrfs_root *root, u64 objectid)
5d4f98a2
YZ		 1368{
1369 struct btrfs_root *reloc_root;
1370 struct extent_buffer *eb;
1371 struct btrfs_root_item *root_item;
1372 struct btrfs_key root_key;
5bc7247a 1373 u64 last_snap = 0;
5d4f98a2
YZ		 1374 int ret;
1375
5d4f98a2
YZ		 1376 root_item = kmalloc(sizeof(*root_item), GFP_NOFS);
1377 BUG_ON(!root_item);
1378
1379 root_key.objectid = BTRFS_TREE_RELOC_OBJECTID;
1380 root_key.type = BTRFS_ROOT_ITEM_KEY;
3fd0a558 1381 root_key.offset = objectid;
5d4f98a2 1382
3fd0a558
YZ		 1383 if (root->root_key.objectid == objectid) {
1384 /* called by btrfs_init_reloc_root */
1385 ret = btrfs_copy_root(trans, root, root->commit_root, &eb,
1386 BTRFS_TREE_RELOC_OBJECTID);
1387 BUG_ON(ret);
1388
5bc7247a 1389 last_snap = btrfs_root_last_snapshot(&root->root_item);
3fd0a558
YZ		 1390 btrfs_set_root_last_snapshot(&root->root_item,
1391 trans->transid - 1);
1392 } else {
1393 /*
1394 * called by btrfs_reloc_post_snapshot_hook.
1395 * the source tree is a reloc tree, all tree blocks
1396 * modified after it was created have RELOC flag
1397 * set in their headers. so it's OK to not update
1398 * the 'last_snapshot'.
1399 */
1400 ret = btrfs_copy_root(trans, root, root->node, &eb,
1401 BTRFS_TREE_RELOC_OBJECTID);
1402 BUG_ON(ret);
1403 }
5d4f98a2 1404
5d4f98a2 1405 memcpy(root_item, &root->root_item, sizeof(*root_item));
5d4f98a2
YZ		 1406 btrfs_set_root_bytenr(root_item, eb->start);
1407 btrfs_set_root_level(root_item, btrfs_header_level(eb));
1408 btrfs_set_root_generation(root_item, trans->transid);
3fd0a558
YZ		 1409
1410 if (root->root_key.objectid == objectid) {
1411 btrfs_set_root_refs(root_item, 0);
1412 memset(&root_item->drop_progress, 0,
1413 sizeof(struct btrfs_disk_key));
1414 root_item->drop_level = 0;
5bc7247a
MX		 1415 /*
1416 * abuse rtransid, it is safe because it is impossible to
1417 * receive data into a relocation tree.
1418 */
1419 btrfs_set_root_rtransid(root_item, last_snap);
1420 btrfs_set_root_otransid(root_item, trans->transid);
3fd0a558 1421 }
5d4f98a2
YZ		 1422
1423 btrfs_tree_unlock(eb);
1424 free_extent_buffer(eb);
1425
1426 ret = btrfs_insert_root(trans, root->fs_info->tree_root,
1427 &root_key, root_item);
1428 BUG_ON(ret);
1429 kfree(root_item);
1430
cb517eab 1431 reloc_root = btrfs_read_fs_root(root->fs_info->tree_root, &root_key);
5d4f98a2
YZ		 1432 BUG_ON(IS_ERR(reloc_root));
1433 reloc_root->last_trans = trans->transid;
3fd0a558
YZ		 1434 return reloc_root;
1435}
1436
1437/*
1438 * create reloc tree for a given fs tree. reloc tree is just a
1439 * snapshot of the fs tree with special root objectid.
1440 */
1441int btrfs_init_reloc_root(struct btrfs_trans_handle *trans,
1442 struct btrfs_root *root)
1443{
1444 struct btrfs_root *reloc_root;
1445 struct reloc_control *rc = root->fs_info->reloc_ctl;
20dd2cbf 1446 struct btrfs_block_rsv *rsv;
3fd0a558 1447 int clear_rsv = 0;
ffd7b339 1448 int ret;
3fd0a558
YZ		 1449
1450 if (root->reloc_root) {
1451 reloc_root = root->reloc_root;
1452 reloc_root->last_trans = trans->transid;
1453 return 0;
1454 }
1455
1456 if (!rc || !rc->create_reloc_tree ||
1457 root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID)
1458 return 0;
1459
20dd2cbf
MX		 1460 if (!trans->reloc_reserved) {
1461 rsv = trans->block_rsv;
3fd0a558
YZ		 1462 trans->block_rsv = rc->block_rsv;
1463 clear_rsv = 1;
1464 }
1465 reloc_root = create_reloc_root(trans, root, root->root_key.objectid);
1466 if (clear_rsv)
20dd2cbf 1467 trans->block_rsv = rsv;
5d4f98a2 1468
ffd7b339
JM		 1469 ret = __add_reloc_root(reloc_root);
1470 BUG_ON(ret < 0);
5d4f98a2
YZ		 1471 root->reloc_root = reloc_root;
1472 return 0;
1473}
1474
1475/*
1476 * update root item of reloc tree
1477 */
1478int btrfs_update_reloc_root(struct btrfs_trans_handle *trans,
1479 struct btrfs_root *root)
1480{
1481 struct btrfs_root *reloc_root;
1482 struct btrfs_root_item *root_item;
5d4f98a2
YZ		 1483 int ret;
1484
1485 if (!root->reloc_root)
7585717f 1486 goto out;
5d4f98a2
YZ		 1487
1488 reloc_root = root->reloc_root;
1489 root_item = &reloc_root->root_item;
1490
3fd0a558
YZ		 1491 if (root->fs_info->reloc_ctl->merge_reloc_tree &&
1492 btrfs_root_refs(root_item) == 0) {
5d4f98a2 1493 root->reloc_root = NULL;
c974c464 1494 __del_reloc_root(reloc_root);
5d4f98a2
YZ		 1495 }
1496
5d4f98a2
YZ		 1497 if (reloc_root->commit_root != reloc_root->node) {
1498 btrfs_set_root_node(root_item, reloc_root->node);
1499 free_extent_buffer(reloc_root->commit_root);
1500 reloc_root->commit_root = btrfs_root_node(reloc_root);
1501 }
1502
1503 ret = btrfs_update_root(trans, root->fs_info->tree_root,
1504 &reloc_root->root_key, root_item);
1505 BUG_ON(ret);
7585717f
CM		 1506
1507out:
5d4f98a2
YZ		 1508 return 0;
1509}
1510
1511/*
1512 * helper to find first cached inode with inode number >= objectid
1513 * in a subvolume
1514 */
1515static struct inode *find_next_inode(struct btrfs_root *root, u64 objectid)
1516{
1517 struct rb_node *node;
1518 struct rb_node *prev;
1519 struct btrfs_inode *entry;
1520 struct inode *inode;
1521
1522 spin_lock(&root->inode_lock);
1523again:
1524 node = root->inode_tree.rb_node;
1525 prev = NULL;
1526 while (node) {
1527 prev = node;
1528 entry = rb_entry(node, struct btrfs_inode, rb_node);
1529
33345d01 1530 if (objectid < btrfs_ino(&entry->vfs_inode))
5d4f98a2 1531 node = node->rb_left;
33345d01 1532 else if (objectid > btrfs_ino(&entry->vfs_inode))
5d4f98a2
YZ		 1533 node = node->rb_right;
1534 else
1535 break;
1536 }
1537 if (!node) {
1538 while (prev) {
1539 entry = rb_entry(prev, struct btrfs_inode, rb_node);
33345d01 1540 if (objectid <= btrfs_ino(&entry->vfs_inode)) {
5d4f98a2
YZ		 1541 node = prev;
1542 break;
1543 }
1544 prev = rb_next(prev);
1545 }
1546 }
1547 while (node) {
1548 entry = rb_entry(node, struct btrfs_inode, rb_node);
1549 inode = igrab(&entry->vfs_inode);
1550 if (inode) {
1551 spin_unlock(&root->inode_lock);
1552 return inode;
1553 }
1554
33345d01 1555 objectid = btrfs_ino(&entry->vfs_inode) + 1;
5d4f98a2
YZ		 1556 if (cond_resched_lock(&root->inode_lock))
1557 goto again;
1558
1559 node = rb_next(node);
1560 }
1561 spin_unlock(&root->inode_lock);
1562 return NULL;
1563}
1564
1565static int in_block_group(u64 bytenr,
1566 struct btrfs_block_group_cache *block_group)
1567{
1568 if (bytenr >= block_group->key.objectid &&
1569 bytenr < block_group->key.objectid + block_group->key.offset)
1570 return 1;
1571 return 0;
1572}
1573
1574/*
1575 * get new location of data
1576 */
1577static int get_new_location(struct inode *reloc_inode, u64 *new_bytenr,
1578 u64 bytenr, u64 num_bytes)
1579{
1580 struct btrfs_root *root = BTRFS_I(reloc_inode)->root;
1581 struct btrfs_path *path;
1582 struct btrfs_file_extent_item *fi;
1583 struct extent_buffer *leaf;
1584 int ret;
1585
1586 path = btrfs_alloc_path();
1587 if (!path)
1588 return -ENOMEM;
1589
1590 bytenr -= BTRFS_I(reloc_inode)->index_cnt;
33345d01 1591 ret = btrfs_lookup_file_extent(NULL, root, path, btrfs_ino(reloc_inode),
5d4f98a2
YZ		 1592 bytenr, 0);
1593 if (ret < 0)
1594 goto out;
1595 if (ret > 0) {
1596 ret = -ENOENT;
1597 goto out;
1598 }
1599
1600 leaf = path->nodes[0];
1601 fi = btrfs_item_ptr(leaf, path->slots[0],
1602 struct btrfs_file_extent_item);
1603
1604 BUG_ON(btrfs_file_extent_offset(leaf, fi) ||
1605 btrfs_file_extent_compression(leaf, fi) ||
1606 btrfs_file_extent_encryption(leaf, fi) ||
1607 btrfs_file_extent_other_encoding(leaf, fi));
1608
1609 if (num_bytes != btrfs_file_extent_disk_num_bytes(leaf, fi)) {
83d4cfd4 1610 ret = -EINVAL;
5d4f98a2
YZ		 1611 goto out;
1612 }
1613
3fd0a558 1614 *new_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
5d4f98a2
YZ		 1615 ret = 0;
1616out:
1617 btrfs_free_path(path);
1618 return ret;
1619}
1620
1621/*
1622 * update file extent items in the tree leaf to point to
1623 * the new locations.
1624 */
3fd0a558
YZ		 1625static noinline_for_stack
1626int replace_file_extents(struct btrfs_trans_handle *trans,
1627 struct reloc_control *rc,
1628 struct btrfs_root *root,
1629 struct extent_buffer *leaf)
5d4f98a2
YZ		 1630{
1631 struct btrfs_key key;
1632 struct btrfs_file_extent_item *fi;
1633 struct inode *inode = NULL;
5d4f98a2
YZ		 1634 u64 parent;
1635 u64 bytenr;
3fd0a558 1636 u64 new_bytenr = 0;
5d4f98a2
YZ		 1637 u64 num_bytes;
1638 u64 end;
1639 u32 nritems;
1640 u32 i;
83d4cfd4 1641 int ret = 0;
5d4f98a2
YZ		 1642 int first = 1;
1643 int dirty = 0;
1644
1645 if (rc->stage != UPDATE_DATA_PTRS)
1646 return 0;
1647
1648 /* reloc trees always use full backref */
1649 if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID)
1650 parent = leaf->start;
1651 else
1652 parent = 0;
1653
1654 nritems = btrfs_header_nritems(leaf);
1655 for (i = 0; i < nritems; i++) {
1656 cond_resched();
1657 btrfs_item_key_to_cpu(leaf, &key, i);
1658 if (key.type != BTRFS_EXTENT_DATA_KEY)
1659 continue;
1660 fi = btrfs_item_ptr(leaf, i, struct btrfs_file_extent_item);
1661 if (btrfs_file_extent_type(leaf, fi) ==
1662 BTRFS_FILE_EXTENT_INLINE)
1663 continue;
1664 bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
1665 num_bytes = btrfs_file_extent_disk_num_bytes(leaf, fi);
1666 if (bytenr == 0)
1667 continue;
1668 if (!in_block_group(bytenr, rc->block_group))
1669 continue;
1670
1671 /*
1672 * if we are modifying block in fs tree, wait for readpage
1673 * to complete and drop the extent cache
1674 */
1675 if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) {
5d4f98a2
YZ		 1676 if (first) {
1677 inode = find_next_inode(root, key.objectid);
5d4f98a2 1678 first = 0;
33345d01 1679 } else if (inode && btrfs_ino(inode) < key.objectid) {
3fd0a558 1680 btrfs_add_delayed_iput(inode);
5d4f98a2 1681 inode = find_next_inode(root, key.objectid);
5d4f98a2 1682 }
33345d01 1683 if (inode && btrfs_ino(inode) == key.objectid) {
5d4f98a2
YZ		 1684 end = key.offset +
1685 btrfs_file_extent_num_bytes(leaf, fi);
1686 WARN_ON(!IS_ALIGNED(key.offset,
1687 root->sectorsize));
1688 WARN_ON(!IS_ALIGNED(end, root->sectorsize));
1689 end--;
1690 ret = try_lock_extent(&BTRFS_I(inode)->io_tree,
d0082371 1691 key.offset, end);
5d4f98a2
YZ		 1692 if (!ret)
1693 continue;
1694
1695 btrfs_drop_extent_cache(inode, key.offset, end,
1696 1);
1697 unlock_extent(&BTRFS_I(inode)->io_tree,
d0082371 1698 key.offset, end);
5d4f98a2
YZ		 1699 }
1700 }
1701
1702 ret = get_new_location(rc->data_inode, &new_bytenr,
1703 bytenr, num_bytes);
83d4cfd4
JB		 1704 if (ret) {
1705 /*
1706 * Don't have to abort since we've not changed anything
1707 * in the file extent yet.
1708 */
1709 break;
3fd0a558 1710 }
5d4f98a2
YZ		 1711
1712 btrfs_set_file_extent_disk_bytenr(leaf, fi, new_bytenr);
1713 dirty = 1;
1714
1715 key.offset -= btrfs_file_extent_offset(leaf, fi);
1716 ret = btrfs_inc_extent_ref(trans, root, new_bytenr,
1717 num_bytes, parent,
1718 btrfs_header_owner(leaf),
66d7e7f0 1719 key.objectid, key.offset, 1);
83d4cfd4
JB		 1720 if (ret) {
1721 btrfs_abort_transaction(trans, root, ret);
1722 break;
1723 }
5d4f98a2
YZ		 1724
1725 ret = btrfs_free_extent(trans, root, bytenr, num_bytes,
1726 parent, btrfs_header_owner(leaf),
66d7e7f0 1727 key.objectid, key.offset, 1);
83d4cfd4
JB		 1728 if (ret) {
1729 btrfs_abort_transaction(trans, root, ret);
1730 break;
1731 }
5d4f98a2
YZ		 1732 }
1733 if (dirty)
1734 btrfs_mark_buffer_dirty(leaf);
3fd0a558
YZ		 1735 if (inode)
1736 btrfs_add_delayed_iput(inode);
83d4cfd4 1737 return ret;
5d4f98a2
YZ		 1738}
1739
1740static noinline_for_stack
1741int memcmp_node_keys(struct extent_buffer *eb, int slot,
1742 struct btrfs_path *path, int level)
1743{
1744 struct btrfs_disk_key key1;
1745 struct btrfs_disk_key key2;
1746 btrfs_node_key(eb, &key1, slot);
1747 btrfs_node_key(path->nodes[level], &key2, path->slots[level]);
1748 return memcmp(&key1, &key2, sizeof(key1));
1749}
1750
1751/*
1752 * try to replace tree blocks in fs tree with the new blocks
1753 * in reloc tree. tree blocks haven't been modified since the
1754 * reloc tree was create can be replaced.
1755 *
1756 * if a block was replaced, level of the block + 1 is returned.
1757 * if no block got replaced, 0 is returned. if there are other
1758 * errors, a negative error number is returned.
1759 */
3fd0a558
YZ		 1760static noinline_for_stack
1761int replace_path(struct btrfs_trans_handle *trans,
1762 struct btrfs_root *dest, struct btrfs_root *src,
1763 struct btrfs_path *path, struct btrfs_key *next_key,
1764 int lowest_level, int max_level)
5d4f98a2
YZ		 1765{
1766 struct extent_buffer *eb;
1767 struct extent_buffer *parent;
1768 struct btrfs_key key;
1769 u64 old_bytenr;
1770 u64 new_bytenr;
1771 u64 old_ptr_gen;
1772 u64 new_ptr_gen;
1773 u64 last_snapshot;
1774 u32 blocksize;
3fd0a558 1775 int cow = 0;
5d4f98a2
YZ		 1776 int level;
1777 int ret;
1778 int slot;
1779
1780 BUG_ON(src->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID);
1781 BUG_ON(dest->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID);
5d4f98a2
YZ		 1782
1783 last_snapshot = btrfs_root_last_snapshot(&src->root_item);
3fd0a558 1784again:
5d4f98a2
YZ		 1785 slot = path->slots[lowest_level];
1786 btrfs_node_key_to_cpu(path->nodes[lowest_level], &key, slot);
1787
1788 eb = btrfs_lock_root_node(dest);
1789 btrfs_set_lock_blocking(eb);
1790 level = btrfs_header_level(eb);
1791
1792 if (level < lowest_level) {
1793 btrfs_tree_unlock(eb);
1794 free_extent_buffer(eb);
1795 return 0;
1796 }
1797
3fd0a558
YZ		 1798 if (cow) {
1799 ret = btrfs_cow_block(trans, dest, eb, NULL, 0, &eb);
1800 BUG_ON(ret);
1801 }
5d4f98a2
YZ		 1802 btrfs_set_lock_blocking(eb);
1803
1804 if (next_key) {
1805 next_key->objectid = (u64)-1;
1806 next_key->type = (u8)-1;
1807 next_key->offset = (u64)-1;
1808 }
1809
1810 parent = eb;
1811 while (1) {
1812 level = btrfs_header_level(parent);
1813 BUG_ON(level < lowest_level);
1814
1815 ret = btrfs_bin_search(parent, &key, level, &slot);
1816 if (ret && slot > 0)
1817 slot--;
1818
1819 if (next_key && slot + 1 < btrfs_header_nritems(parent))
1820 btrfs_node_key_to_cpu(parent, next_key, slot + 1);
1821
1822 old_bytenr = btrfs_node_blockptr(parent, slot);
707e8a07 1823 blocksize = dest->nodesize;
5d4f98a2
YZ		 1824 old_ptr_gen = btrfs_node_ptr_generation(parent, slot);
1825
1826 if (level <= max_level) {
1827 eb = path->nodes[level];
1828 new_bytenr = btrfs_node_blockptr(eb,
1829 path->slots[level]);
1830 new_ptr_gen = btrfs_node_ptr_generation(eb,
1831 path->slots[level]);
1832 } else {
1833 new_bytenr = 0;
1834 new_ptr_gen = 0;
1835 }
1836
fae7f21c 1837 if (WARN_ON(new_bytenr > 0 && new_bytenr == old_bytenr)) {
5d4f98a2
YZ		 1838 ret = level;
1839 break;
1840 }
1841
1842 if (new_bytenr == 0 || old_ptr_gen > last_snapshot ||
1843 memcmp_node_keys(parent, slot, path, level)) {
3fd0a558 1844 if (level <= lowest_level) {
5d4f98a2
YZ		 1845 ret = 0;
1846 break;
1847 }
1848
1849 eb = read_tree_block(dest, old_bytenr, blocksize,
1850 old_ptr_gen);
416bc658
JB		 1851 if (!eb || !extent_buffer_uptodate(eb)) {
1852 ret = (!eb) ? -ENOMEM : -EIO;
1853 free_extent_buffer(eb);
379cde74 1854 break;
416bc658 1855 }
5d4f98a2 1856 btrfs_tree_lock(eb);
3fd0a558
YZ		 1857 if (cow) {
1858 ret = btrfs_cow_block(trans, dest, eb, parent,
1859 slot, &eb);
1860 BUG_ON(ret);
5d4f98a2 1861 }
3fd0a558 1862 btrfs_set_lock_blocking(eb);
5d4f98a2
YZ		 1863
1864 btrfs_tree_unlock(parent);
1865 free_extent_buffer(parent);
1866
1867 parent = eb;
1868 continue;
1869 }
1870
3fd0a558
YZ		 1871 if (!cow) {
1872 btrfs_tree_unlock(parent);
1873 free_extent_buffer(parent);
1874 cow = 1;
1875 goto again;
1876 }
1877
5d4f98a2
YZ		 1878 btrfs_node_key_to_cpu(path->nodes[level], &key,
1879 path->slots[level]);
b3b4aa74 1880 btrfs_release_path(path);
5d4f98a2
YZ		 1881
1882 path->lowest_level = level;
1883 ret = btrfs_search_slot(trans, src, &key, path, 0, 1);
1884 path->lowest_level = 0;
1885 BUG_ON(ret);
1886
1887 /*
1888 * swap blocks in fs tree and reloc tree.
1889 */
1890 btrfs_set_node_blockptr(parent, slot, new_bytenr);
1891 btrfs_set_node_ptr_generation(parent, slot, new_ptr_gen);
1892 btrfs_mark_buffer_dirty(parent);
1893
1894 btrfs_set_node_blockptr(path->nodes[level],
1895 path->slots[level], old_bytenr);
1896 btrfs_set_node_ptr_generation(path->nodes[level],
1897 path->slots[level], old_ptr_gen);
1898 btrfs_mark_buffer_dirty(path->nodes[level]);
1899
1900 ret = btrfs_inc_extent_ref(trans, src, old_bytenr, blocksize,
1901 path->nodes[level]->start,
66d7e7f0
AJ		 1902 src->root_key.objectid, level - 1, 0,
1903 1);
5d4f98a2
YZ		 1904 BUG_ON(ret);
1905 ret = btrfs_inc_extent_ref(trans, dest, new_bytenr, blocksize,
1906 0, dest->root_key.objectid, level - 1,
66d7e7f0 1907 0, 1);
5d4f98a2
YZ		 1908 BUG_ON(ret);
1909
1910 ret = btrfs_free_extent(trans, src, new_bytenr, blocksize,
1911 path->nodes[level]->start,
66d7e7f0
AJ		 1912 src->root_key.objectid, level - 1, 0,
1913 1);
5d4f98a2
YZ		 1914 BUG_ON(ret);
1915
1916 ret = btrfs_free_extent(trans, dest, old_bytenr, blocksize,
1917 0, dest->root_key.objectid, level - 1,
66d7e7f0 1918 0, 1);
5d4f98a2
YZ		 1919 BUG_ON(ret);
1920
1921 btrfs_unlock_up_safe(path, 0);
1922
1923 ret = level;
1924 break;
1925 }
1926 btrfs_tree_unlock(parent);
1927 free_extent_buffer(parent);
1928 return ret;
1929}
1930
1931/*
1932 * helper to find next relocated block in reloc tree
1933 */
1934static noinline_for_stack
1935int walk_up_reloc_tree(struct btrfs_root *root, struct btrfs_path *path,
1936 int *level)
1937{
1938 struct extent_buffer *eb;
1939 int i;
1940 u64 last_snapshot;
1941 u32 nritems;
1942
1943 last_snapshot = btrfs_root_last_snapshot(&root->root_item);
1944
1945 for (i = 0; i < *level; i++) {
1946 free_extent_buffer(path->nodes[i]);
1947 path->nodes[i] = NULL;
1948 }
1949
1950 for (i = *level; i < BTRFS_MAX_LEVEL && path->nodes[i]; i++) {
1951 eb = path->nodes[i];
1952 nritems = btrfs_header_nritems(eb);
1953 while (path->slots[i] + 1 < nritems) {
1954 path->slots[i]++;
1955 if (btrfs_node_ptr_generation(eb, path->slots[i]) <=
1956 last_snapshot)
1957 continue;
1958
1959 *level = i;
1960 return 0;
1961 }
1962 free_extent_buffer(path->nodes[i]);
1963 path->nodes[i] = NULL;
1964 }
1965 return 1;
1966}
1967
1968/*
1969 * walk down reloc tree to find relocated block of lowest level
1970 */
1971static noinline_for_stack
1972int walk_down_reloc_tree(struct btrfs_root *root, struct btrfs_path *path,
1973 int *level)
1974{
1975 struct extent_buffer *eb = NULL;
1976 int i;
1977 u64 bytenr;
1978 u64 ptr_gen = 0;
1979 u64 last_snapshot;
1980 u32 blocksize;
1981 u32 nritems;
1982
1983 last_snapshot = btrfs_root_last_snapshot(&root->root_item);
1984
1985 for (i = *level; i > 0; i--) {
1986 eb = path->nodes[i];
1987 nritems = btrfs_header_nritems(eb);
1988 while (path->slots[i] < nritems) {
1989 ptr_gen = btrfs_node_ptr_generation(eb, path->slots[i]);
1990 if (ptr_gen > last_snapshot)
1991 break;
1992 path->slots[i]++;
1993 }
1994 if (path->slots[i] >= nritems) {
1995 if (i == *level)
1996 break;
1997 *level = i + 1;
1998 return 0;
1999 }
2000 if (i == 1) {
2001 *level = i;
2002 return 0;
2003 }
2004
2005 bytenr = btrfs_node_blockptr(eb, path->slots[i]);
707e8a07 2006 blocksize = root->nodesize;
5d4f98a2 2007 eb = read_tree_block(root, bytenr, blocksize, ptr_gen);
416bc658
JB		 2008 if (!eb || !extent_buffer_uptodate(eb)) {
2009 free_extent_buffer(eb);
2010 return -EIO;
2011 }
5d4f98a2
YZ		 2012 BUG_ON(btrfs_header_level(eb) != i - 1);
2013 path->nodes[i - 1] = eb;
2014 path->slots[i - 1] = 0;
2015 }
2016 return 1;
2017}
2018
2019/*
2020 * invalidate extent cache for file extents whose key in range of
2021 * [min_key, max_key)
2022 */
2023static int invalidate_extent_cache(struct btrfs_root *root,
2024 struct btrfs_key *min_key,
2025 struct btrfs_key *max_key)
2026{
2027 struct inode *inode = NULL;
2028 u64 objectid;
2029 u64 start, end;
33345d01 2030 u64 ino;
5d4f98a2
YZ		 2031
2032 objectid = min_key->objectid;
2033 while (1) {
2034 cond_resched();
2035 iput(inode);
2036
2037 if (objectid > max_key->objectid)
2038 break;
2039
2040 inode = find_next_inode(root, objectid);
2041 if (!inode)
2042 break;
33345d01 2043 ino = btrfs_ino(inode);
5d4f98a2 2044
33345d01 2045 if (ino > max_key->objectid) {
5d4f98a2
YZ		 2046 iput(inode);
2047 break;
2048 }
2049
33345d01 2050 objectid = ino + 1;
5d4f98a2
YZ		 2051 if (!S_ISREG(inode->i_mode))
2052 continue;
2053
33345d01 2054 if (unlikely(min_key->objectid == ino)) {
5d4f98a2
YZ		 2055 if (min_key->type > BTRFS_EXTENT_DATA_KEY)
2056 continue;
2057 if (min_key->type < BTRFS_EXTENT_DATA_KEY)
2058 start = 0;
2059 else {
2060 start = min_key->offset;
2061 WARN_ON(!IS_ALIGNED(start, root->sectorsize));
2062 }
2063 } else {
2064 start = 0;
2065 }
2066
33345d01 2067 if (unlikely(max_key->objectid == ino)) {
5d4f98a2
YZ		 2068 if (max_key->type < BTRFS_EXTENT_DATA_KEY)
2069 continue;
2070 if (max_key->type > BTRFS_EXTENT_DATA_KEY) {
2071 end = (u64)-1;
2072 } else {
2073 if (max_key->offset == 0)
2074 continue;
2075 end = max_key->offset;
2076 WARN_ON(!IS_ALIGNED(end, root->sectorsize));
2077 end--;
2078 }
2079 } else {
2080 end = (u64)-1;
2081 }
2082
2083 /* the lock_extent waits for readpage to complete */
d0082371 2084 lock_extent(&BTRFS_I(inode)->io_tree, start, end);
5d4f98a2 2085 btrfs_drop_extent_cache(inode, start, end, 1);
d0082371 2086 unlock_extent(&BTRFS_I(inode)->io_tree, start, end);
5d4f98a2
YZ		 2087 }
2088 return 0;
2089}
2090
2091static int find_next_key(struct btrfs_path *path, int level,
2092 struct btrfs_key *key)
2093
2094{
2095 while (level < BTRFS_MAX_LEVEL) {
2096 if (!path->nodes[level])
2097 break;
2098 if (path->slots[level] + 1 <
2099 btrfs_header_nritems(path->nodes[level])) {
2100 btrfs_node_key_to_cpu(path->nodes[level], key,
2101 path->slots[level] + 1);
2102 return 0;
2103 }
2104 level++;
2105 }
2106 return 1;
2107}
2108
2109/*
2110 * merge the relocated tree blocks in reloc tree with corresponding
2111 * fs tree.
2112 */
2113static noinline_for_stack int merge_reloc_root(struct reloc_control *rc,
2114 struct btrfs_root *root)
2115{
2116 LIST_HEAD(inode_list);
2117 struct btrfs_key key;
2118 struct btrfs_key next_key;
9e6a0c52 2119 struct btrfs_trans_handle *trans = NULL;
5d4f98a2
YZ		 2120 struct btrfs_root *reloc_root;
2121 struct btrfs_root_item *root_item;
2122 struct btrfs_path *path;
3fd0a558 2123 struct extent_buffer *leaf;
5d4f98a2
YZ		 2124 int level;
2125 int max_level;
2126 int replaced = 0;
2127 int ret;
2128 int err = 0;
3fd0a558 2129 u32 min_reserved;
5d4f98a2
YZ		 2130
2131 path = btrfs_alloc_path();
2132 if (!path)
2133 return -ENOMEM;
026fd317 2134 path->reada = 1;
5d4f98a2
YZ		 2135
2136 reloc_root = root->reloc_root;
2137 root_item = &reloc_root->root_item;
2138
2139 if (btrfs_disk_key_objectid(&root_item->drop_progress) == 0) {
2140 level = btrfs_root_level(root_item);
2141 extent_buffer_get(reloc_root->node);
2142 path->nodes[level] = reloc_root->node;
2143 path->slots[level] = 0;
2144 } else {
2145 btrfs_disk_key_to_cpu(&key, &root_item->drop_progress);
2146
2147 level = root_item->drop_level;
2148 BUG_ON(level == 0);
2149 path->lowest_level = level;
2150 ret = btrfs_search_slot(NULL, reloc_root, &key, path, 0, 0);
33c66f43 2151 path->lowest_level = 0;
5d4f98a2
YZ		 2152 if (ret < 0) {
2153 btrfs_free_path(path);
2154 return ret;
2155 }
2156
2157 btrfs_node_key_to_cpu(path->nodes[level], &next_key,
2158 path->slots[level]);
2159 WARN_ON(memcmp(&key, &next_key, sizeof(key)));
2160
2161 btrfs_unlock_up_safe(path, 0);
2162 }
2163
3fd0a558
YZ		 2164 min_reserved = root->nodesize * (BTRFS_MAX_LEVEL - 1) * 2;
2165 memset(&next_key, 0, sizeof(next_key));
5d4f98a2 2166
3fd0a558 2167 while (1) {
08e007d2
MX		 2168 ret = btrfs_block_rsv_refill(root, rc->block_rsv, min_reserved,
2169 BTRFS_RESERVE_FLUSH_ALL);
3fd0a558 2170 if (ret) {
9e6a0c52
JB		 2171 err = ret;
2172 goto out;
5d4f98a2 2173 }
9e6a0c52
JB		 2174 trans = btrfs_start_transaction(root, 0);
2175 if (IS_ERR(trans)) {
2176 err = PTR_ERR(trans);
2177 trans = NULL;
2178 goto out;
2179 }
2180 trans->block_rsv = rc->block_rsv;
5d4f98a2 2181
5d4f98a2 2182 replaced = 0;
5d4f98a2
YZ		 2183 max_level = level;
2184
2185 ret = walk_down_reloc_tree(reloc_root, path, &level);
2186 if (ret < 0) {
2187 err = ret;
2188 goto out;
2189 }
2190 if (ret > 0)
2191 break;
2192
2193 if (!find_next_key(path, level, &key) &&
2194 btrfs_comp_cpu_keys(&next_key, &key) >= 0) {
2195 ret = 0;
5d4f98a2 2196 } else {
3fd0a558
YZ		 2197 ret = replace_path(trans, root, reloc_root, path,
2198 &next_key, level, max_level);
5d4f98a2
YZ		 2199 }
2200 if (ret < 0) {
2201 err = ret;
2202 goto out;
2203 }
2204
2205 if (ret > 0) {
2206 level = ret;
2207 btrfs_node_key_to_cpu(path->nodes[level], &key,
2208 path->slots[level]);
2209 replaced = 1;
5d4f98a2
YZ		 2210 }
2211
2212 ret = walk_up_reloc_tree(reloc_root, path, &level);
2213 if (ret > 0)
2214 break;
2215
2216 BUG_ON(level == 0);
2217 /*
2218 * save the merging progress in the drop_progress.
2219 * this is OK since root refs == 1 in this case.
2220 */
2221 btrfs_node_key(path->nodes[level], &root_item->drop_progress,
2222 path->slots[level]);
2223 root_item->drop_level = level;
2224
3fd0a558 2225 btrfs_end_transaction_throttle(trans, root);
9e6a0c52 2226 trans = NULL;
5d4f98a2 2227
b53d3f5d 2228 btrfs_btree_balance_dirty(root);
5d4f98a2
YZ		 2229
2230 if (replaced && rc->stage == UPDATE_DATA_PTRS)
2231 invalidate_extent_cache(root, &key, &next_key);
2232 }
2233
2234 /*
2235 * handle the case only one block in the fs tree need to be
2236 * relocated and the block is tree root.
2237 */
2238 leaf = btrfs_lock_root_node(root);
2239 ret = btrfs_cow_block(trans, root, leaf, NULL, 0, &leaf);
2240 btrfs_tree_unlock(leaf);
2241 free_extent_buffer(leaf);
2242 if (ret < 0)
2243 err = ret;
2244out:
2245 btrfs_free_path(path);
2246
2247 if (err == 0) {
2248 memset(&root_item->drop_progress, 0,
2249 sizeof(root_item->drop_progress));
2250 root_item->drop_level = 0;
2251 btrfs_set_root_refs(root_item, 0);
3fd0a558 2252 btrfs_update_reloc_root(trans, root);
5d4f98a2
YZ		 2253 }
2254
9e6a0c52
JB		 2255 if (trans)
2256 btrfs_end_transaction_throttle(trans, root);
5d4f98a2 2257
b53d3f5d 2258 btrfs_btree_balance_dirty(root);
5d4f98a2 2259
5d4f98a2
YZ		 2260 if (replaced && rc->stage == UPDATE_DATA_PTRS)
2261 invalidate_extent_cache(root, &key, &next_key);
2262
2263 return err;
2264}
2265
3fd0a558
YZ		 2266static noinline_for_stack
2267int prepare_to_merge(struct reloc_control *rc, int err)
5d4f98a2 2268{
3fd0a558 2269 struct btrfs_root *root = rc->extent_root;
5d4f98a2 2270 struct btrfs_root *reloc_root;
3fd0a558
YZ		 2271 struct btrfs_trans_handle *trans;
2272 LIST_HEAD(reloc_roots);
2273 u64 num_bytes = 0;
2274 int ret;
3fd0a558 2275
7585717f 2276 mutex_lock(&root->fs_info->reloc_mutex);
3fd0a558
YZ		 2277 rc->merging_rsv_size += root->nodesize * (BTRFS_MAX_LEVEL - 1) * 2;
2278 rc->merging_rsv_size += rc->nodes_relocated * 2;
7585717f
CM		 2279 mutex_unlock(&root->fs_info->reloc_mutex);
2280
3fd0a558
YZ		 2281again:
2282 if (!err) {
2283 num_bytes = rc->merging_rsv_size;
08e007d2
MX		 2284 ret = btrfs_block_rsv_add(root, rc->block_rsv, num_bytes,
2285 BTRFS_RESERVE_FLUSH_ALL);
3fd0a558
YZ		 2286 if (ret)
2287 err = ret;
2288 }
2289
7a7eaa40 2290 trans = btrfs_join_transaction(rc->extent_root);
3612b495
TI		 2291 if (IS_ERR(trans)) {
2292 if (!err)
2293 btrfs_block_rsv_release(rc->extent_root,
2294 rc->block_rsv, num_bytes);
2295 return PTR_ERR(trans);
2296 }
3fd0a558
YZ		 2297
2298 if (!err) {
2299 if (num_bytes != rc->merging_rsv_size) {
2300 btrfs_end_transaction(trans, rc->extent_root);
2301 btrfs_block_rsv_release(rc->extent_root,
2302 rc->block_rsv, num_bytes);
3fd0a558
YZ		 2303 goto again;
2304 }
2305 }
5d4f98a2 2306
3fd0a558
YZ		 2307 rc->merge_reloc_tree = 1;
2308
2309 while (!list_empty(&rc->reloc_roots)) {
2310 reloc_root = list_entry(rc->reloc_roots.next,
2311 struct btrfs_root, root_list);
2312 list_del_init(&reloc_root->root_list);
5d4f98a2 2313
5d4f98a2
YZ		 2314 root = read_fs_root(reloc_root->fs_info,
2315 reloc_root->root_key.offset);
2316 BUG_ON(IS_ERR(root));
2317 BUG_ON(root->reloc_root != reloc_root);
2318
3fd0a558
YZ		 2319 /*
2320 * set reference count to 1, so btrfs_recover_relocation
2321 * knows it should resumes merging
2322 */
2323 if (!err)
2324 btrfs_set_root_refs(&reloc_root->root_item, 1);
5d4f98a2 2325 btrfs_update_reloc_root(trans, root);
5d4f98a2 2326
3fd0a558
YZ		 2327 list_add(&reloc_root->root_list, &reloc_roots);
2328 }
5d4f98a2 2329
3fd0a558 2330 list_splice(&reloc_roots, &rc->reloc_roots);
5d4f98a2 2331
3fd0a558
YZ		 2332 if (!err)
2333 btrfs_commit_transaction(trans, rc->extent_root);
2334 else
2335 btrfs_end_transaction(trans, rc->extent_root);
2336 return err;
5d4f98a2
YZ		 2337}
2338
aca1bba6
LB		 2339static noinline_for_stack
2340void free_reloc_roots(struct list_head *list)
2341{
2342 struct btrfs_root *reloc_root;
2343
2344 while (!list_empty(list)) {
2345 reloc_root = list_entry(list->next, struct btrfs_root,
2346 root_list);
c974c464 2347 __del_reloc_root(reloc_root);
aca1bba6
LB		 2348 }
2349}
2350
3fd0a558
YZ		 2351static noinline_for_stack
2352int merge_reloc_roots(struct reloc_control *rc)
5d4f98a2 2353{
5d4f98a2 2354 struct btrfs_root *root;
3fd0a558 2355 struct btrfs_root *reloc_root;
5bc7247a
MX		 2356 u64 last_snap;
2357 u64 otransid;
2358 u64 objectid;
3fd0a558
YZ		 2359 LIST_HEAD(reloc_roots);
2360 int found = 0;
aca1bba6 2361 int ret = 0;
3fd0a558
YZ		 2362again:
2363 root = rc->extent_root;
7585717f
CM		 2364
2365 /*
2366 * this serializes us with btrfs_record_root_in_transaction,
2367 * we have to make sure nobody is in the middle of
2368 * adding their roots to the list while we are
2369 * doing this splice
2370 */
2371 mutex_lock(&root->fs_info->reloc_mutex);
3fd0a558 2372 list_splice_init(&rc->reloc_roots, &reloc_roots);
7585717f 2373 mutex_unlock(&root->fs_info->reloc_mutex);
5d4f98a2 2374
3fd0a558
YZ		 2375 while (!list_empty(&reloc_roots)) {
2376 found = 1;
2377 reloc_root = list_entry(reloc_roots.next,
2378 struct btrfs_root, root_list);
5d4f98a2 2379
3fd0a558
YZ		 2380 if (btrfs_root_refs(&reloc_root->root_item) > 0) {
2381 root = read_fs_root(reloc_root->fs_info,
2382 reloc_root->root_key.offset);
2383 BUG_ON(IS_ERR(root));
2384 BUG_ON(root->reloc_root != reloc_root);
5d4f98a2 2385
3fd0a558 2386 ret = merge_reloc_root(rc, root);
b37b39cd 2387 if (ret) {
25e293c2
WS		 2388 if (list_empty(&reloc_root->root_list))
2389 list_add_tail(&reloc_root->root_list,
2390 &reloc_roots);
aca1bba6 2391 goto out;
b37b39cd 2392 }
3fd0a558
YZ		 2393 } else {
2394 list_del_init(&reloc_root->root_list);
2395 }
5bc7247a
MX		 2396
2397 /*
2398 * we keep the old last snapshod transid in rtranid when we
2399 * created the relocation tree.
2400 */
2401 last_snap = btrfs_root_rtransid(&reloc_root->root_item);
2402 otransid = btrfs_root_otransid(&reloc_root->root_item);
2403 objectid = reloc_root->root_key.offset;
2404
2c536799 2405 ret = btrfs_drop_snapshot(reloc_root, rc->block_rsv, 0, 1);
aca1bba6
LB		 2406 if (ret < 0) {
2407 if (list_empty(&reloc_root->root_list))
2408 list_add_tail(&reloc_root->root_list,
2409 &reloc_roots);
2410 goto out;
2411 }
5d4f98a2
YZ		 2412 }
2413
3fd0a558
YZ		 2414 if (found) {
2415 found = 0;
2416 goto again;
2417 }
aca1bba6
LB		 2418out:
2419 if (ret) {
2420 btrfs_std_error(root->fs_info, ret);
2421 if (!list_empty(&reloc_roots))
2422 free_reloc_roots(&reloc_roots);
467bb1d2
WS		 2423
2424 /* new reloc root may be added */
2425 mutex_lock(&root->fs_info->reloc_mutex);
2426 list_splice_init(&rc->reloc_roots, &reloc_roots);
2427 mutex_unlock(&root->fs_info->reloc_mutex);
2428 if (!list_empty(&reloc_roots))
2429 free_reloc_roots(&reloc_roots);
aca1bba6
LB		 2430 }
2431
5d4f98a2 2432 BUG_ON(!RB_EMPTY_ROOT(&rc->reloc_root_tree.rb_root));
aca1bba6 2433 return ret;
5d4f98a2
YZ		 2434}
2435
2436static void free_block_list(struct rb_root *blocks)
2437{
2438 struct tree_block *block;
2439 struct rb_node *rb_node;
2440 while ((rb_node = rb_first(blocks))) {
2441 block = rb_entry(rb_node, struct tree_block, rb_node);
2442 rb_erase(rb_node, blocks);
2443 kfree(block);
2444 }
2445}
2446
2447static int record_reloc_root_in_trans(struct btrfs_trans_handle *trans,
2448 struct btrfs_root *reloc_root)
2449{
2450 struct btrfs_root *root;
2451
2452 if (reloc_root->last_trans == trans->transid)
2453 return 0;
2454
2455 root = read_fs_root(reloc_root->fs_info, reloc_root->root_key.offset);
2456 BUG_ON(IS_ERR(root));
2457 BUG_ON(root->reloc_root != reloc_root);
2458
2459 return btrfs_record_root_in_trans(trans, root);
2460}
2461
3fd0a558
YZ		 2462static noinline_for_stack
2463struct btrfs_root *select_reloc_root(struct btrfs_trans_handle *trans,
2464 struct reloc_control *rc,
2465 struct backref_node *node,
dc4103f9 2466 struct backref_edge *edges[])
5d4f98a2
YZ		 2467{
2468 struct backref_node *next;
2469 struct btrfs_root *root;
3fd0a558
YZ		 2470 int index = 0;
2471
5d4f98a2
YZ		 2472 next = node;
2473 while (1) {
2474 cond_resched();
2475 next = walk_up_backref(next, edges, &index);
2476 root = next->root;
3fd0a558 2477 BUG_ON(!root);
27cdeb70 2478 BUG_ON(!test_bit(BTRFS_ROOT_REF_COWS, &root->state));
5d4f98a2
YZ		 2479
2480 if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID) {
2481 record_reloc_root_in_trans(trans, root);
2482 break;
2483 }
2484
3fd0a558
YZ		 2485 btrfs_record_root_in_trans(trans, root);
2486 root = root->reloc_root;
2487
2488 if (next->new_bytenr != root->node->start) {
2489 BUG_ON(next->new_bytenr);
2490 BUG_ON(!list_empty(&next->list));
2491 next->new_bytenr = root->node->start;
2492 next->root = root;
2493 list_add_tail(&next->list,
2494 &rc->backref_cache.changed);
2495 __mark_block_processed(rc, next);
5d4f98a2
YZ		 2496 break;
2497 }
2498
3fd0a558 2499 WARN_ON(1);
5d4f98a2
YZ		 2500 root = NULL;
2501 next = walk_down_backref(edges, &index);
2502 if (!next || next->level <= node->level)
2503 break;
2504 }
3fd0a558
YZ		 2505 if (!root)
2506 return NULL;
5d4f98a2 2507
3fd0a558
YZ		 2508 next = node;
2509 /* setup backref node path for btrfs_reloc_cow_block */
2510 while (1) {
2511 rc->backref_cache.path[next->level] = next;
2512 if (--index < 0)
2513 break;
2514 next = edges[index]->node[UPPER];
5d4f98a2 2515 }
5d4f98a2
YZ		 2516 return root;
2517}
2518
3fd0a558
YZ		 2519/*
2520 * select a tree root for relocation. return NULL if the block
2521 * is reference counted. we should use do_relocation() in this
2522 * case. return a tree root pointer if the block isn't reference
2523 * counted. return -ENOENT if the block is root of reloc tree.
2524 */
5d4f98a2
YZ		 2525static noinline_for_stack
2526struct btrfs_root *select_one_root(struct btrfs_trans_handle *trans,
2527 struct backref_node *node)
2528{
3fd0a558
YZ		 2529 struct backref_node *next;
2530 struct btrfs_root *root;
2531 struct btrfs_root *fs_root = NULL;
5d4f98a2 2532 struct backref_edge *edges[BTRFS_MAX_LEVEL - 1];
3fd0a558
YZ		 2533 int index = 0;
2534
2535 next = node;
2536 while (1) {
2537 cond_resched();
2538 next = walk_up_backref(next, edges, &index);
2539 root = next->root;
2540 BUG_ON(!root);
2541
25985edc 2542 /* no other choice for non-references counted tree */
27cdeb70 2543 if (!test_bit(BTRFS_ROOT_REF_COWS, &root->state))
3fd0a558
YZ		 2544 return root;
2545
2546 if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID)
2547 fs_root = root;
2548
2549 if (next != node)
2550 return NULL;
2551
2552 next = walk_down_backref(edges, &index);
2553 if (!next || next->level <= node->level)
2554 break;
2555 }
2556
2557 if (!fs_root)
2558 return ERR_PTR(-ENOENT);
2559 return fs_root;
5d4f98a2
YZ		 2560}
2561
2562static noinline_for_stack
3fd0a558
YZ		 2563u64 calcu_metadata_size(struct reloc_control *rc,
2564 struct backref_node *node, int reserve)
5d4f98a2 2565{
3fd0a558
YZ		 2566 struct backref_node *next = node;
2567 struct backref_edge *edge;
2568 struct backref_edge *edges[BTRFS_MAX_LEVEL - 1];
2569 u64 num_bytes = 0;
2570 int index = 0;
2571
2572 BUG_ON(reserve && node->processed);
2573
2574 while (next) {
2575 cond_resched();
2576 while (1) {
2577 if (next->processed && (reserve || next != node))
2578 break;
2579
707e8a07 2580 num_bytes += rc->extent_root->nodesize;
3fd0a558
YZ		 2581
2582 if (list_empty(&next->upper))
2583 break;
2584
2585 edge = list_entry(next->upper.next,
2586 struct backref_edge, list[LOWER]);
2587 edges[index++] = edge;
2588 next = edge->node[UPPER];
2589 }
2590 next = walk_down_backref(edges, &index);
2591 }
2592 return num_bytes;
5d4f98a2
YZ		 2593}
2594
3fd0a558
YZ		 2595static int reserve_metadata_space(struct btrfs_trans_handle *trans,
2596 struct reloc_control *rc,
2597 struct backref_node *node)
5d4f98a2 2598{
3fd0a558
YZ		 2599 struct btrfs_root *root = rc->extent_root;
2600 u64 num_bytes;
2601 int ret;
0647bf56 2602 u64 tmp;
3fd0a558
YZ		 2603
2604 num_bytes = calcu_metadata_size(rc, node, 1) * 2;
5d4f98a2 2605
3fd0a558 2606 trans->block_rsv = rc->block_rsv;
0647bf56
WS		 2607 rc->reserved_bytes += num_bytes;
2608 ret = btrfs_block_rsv_refill(root, rc->block_rsv, num_bytes,
2609 BTRFS_RESERVE_FLUSH_ALL);
3fd0a558 2610 if (ret) {
0647bf56
WS		 2611 if (ret == -EAGAIN) {
2612 tmp = rc->extent_root->nodesize *
2613 RELOCATION_RESERVED_NODES;
2614 while (tmp <= rc->reserved_bytes)
2615 tmp <<= 1;
2616 /*
2617 * only one thread can access block_rsv at this point,
2618 * so we don't need hold lock to protect block_rsv.
2619 * we expand more reservation size here to allow enough
2620 * space for relocation and we will return eailer in
2621 * enospc case.
2622 */
2623 rc->block_rsv->size = tmp + rc->extent_root->nodesize *
2624 RELOCATION_RESERVED_NODES;
2625 }
3fd0a558 2626 return ret;
5d4f98a2 2627 }
3fd0a558 2628
3fd0a558
YZ		 2629 return 0;
2630}
2631
5d4f98a2
YZ		 2632/*
2633 * relocate a block tree, and then update pointers in upper level
2634 * blocks that reference the block to point to the new location.
2635 *
2636 * if called by link_to_upper, the block has already been relocated.
2637 * in that case this function just updates pointers.
2638 */
2639static int do_relocation(struct btrfs_trans_handle *trans,
3fd0a558 2640 struct reloc_control *rc,
5d4f98a2
YZ		 2641 struct backref_node *node,
2642 struct btrfs_key *key,
2643 struct btrfs_path *path, int lowest)
2644{
2645 struct backref_node *upper;
2646 struct backref_edge *edge;
2647 struct backref_edge *edges[BTRFS_MAX_LEVEL - 1];
2648 struct btrfs_root *root;
2649 struct extent_buffer *eb;
2650 u32 blocksize;
2651 u64 bytenr;
2652 u64 generation;
5d4f98a2
YZ		 2653 int slot;
2654 int ret;
2655 int err = 0;
2656
2657 BUG_ON(lowest && node->eb);
2658
2659 path->lowest_level = node->level + 1;
3fd0a558 2660 rc->backref_cache.path[node->level] = node;
5d4f98a2
YZ		 2661 list_for_each_entry(edge, &node->upper, list[LOWER]) {
2662 cond_resched();
5d4f98a2
YZ		 2663
2664 upper = edge->node[UPPER];
dc4103f9 2665 root = select_reloc_root(trans, rc, upper, edges);
3fd0a558
YZ		 2666 BUG_ON(!root);
2667
2668 if (upper->eb && !upper->locked) {
2669 if (!lowest) {
2670 ret = btrfs_bin_search(upper->eb, key,
2671 upper->level, &slot);
2672 BUG_ON(ret);
2673 bytenr = btrfs_node_blockptr(upper->eb, slot);
2674 if (node->eb->start == bytenr)
2675 goto next;
2676 }
5d4f98a2 2677 drop_node_buffer(upper);
3fd0a558 2678 }
5d4f98a2
YZ		 2679
2680 if (!upper->eb) {
2681 ret = btrfs_search_slot(trans, root, key, path, 0, 1);
2682 if (ret < 0) {
2683 err = ret;
2684 break;
2685 }
2686 BUG_ON(ret > 0);
2687
3fd0a558
YZ		 2688 if (!upper->eb) {
2689 upper->eb = path->nodes[upper->level];
2690 path->nodes[upper->level] = NULL;
2691 } else {
2692 BUG_ON(upper->eb != path->nodes[upper->level]);
2693 }
5d4f98a2 2694
3fd0a558
YZ		 2695 upper->locked = 1;
2696 path->locks[upper->level] = 0;
5d4f98a2 2697
3fd0a558 2698 slot = path->slots[upper->level];
b3b4aa74 2699 btrfs_release_path(path);
5d4f98a2
YZ		 2700 } else {
2701 ret = btrfs_bin_search(upper->eb, key, upper->level,
2702 &slot);
2703 BUG_ON(ret);
2704 }
2705
2706 bytenr = btrfs_node_blockptr(upper->eb, slot);
3fd0a558
YZ		 2707 if (lowest) {
2708 BUG_ON(bytenr != node->bytenr);
5d4f98a2 2709 } else {
3fd0a558
YZ		 2710 if (node->eb->start == bytenr)
2711 goto next;
5d4f98a2
YZ		 2712 }
2713
707e8a07 2714 blocksize = root->nodesize;
5d4f98a2
YZ		 2715 generation = btrfs_node_ptr_generation(upper->eb, slot);
2716 eb = read_tree_block(root, bytenr, blocksize, generation);
416bc658
JB		 2717 if (!eb || !extent_buffer_uptodate(eb)) {
2718 free_extent_buffer(eb);
97d9a8a4
TI		 2719 err = -EIO;
2720 goto next;
2721 }
5d4f98a2
YZ		 2722 btrfs_tree_lock(eb);
2723 btrfs_set_lock_blocking(eb);
2724
2725 if (!node->eb) {
2726 ret = btrfs_cow_block(trans, root, eb, upper->eb,
2727 slot, &eb);
3fd0a558
YZ		 2728 btrfs_tree_unlock(eb);
2729 free_extent_buffer(eb);
5d4f98a2
YZ		 2730 if (ret < 0) {
2731 err = ret;
3fd0a558 2732 goto next;
5d4f98a2 2733 }
3fd0a558 2734 BUG_ON(node->eb != eb);
5d4f98a2
YZ		 2735 } else {
2736 btrfs_set_node_blockptr(upper->eb, slot,
2737 node->eb->start);
2738 btrfs_set_node_ptr_generation(upper->eb, slot,
2739 trans->transid);
2740 btrfs_mark_buffer_dirty(upper->eb);
2741
2742 ret = btrfs_inc_extent_ref(trans, root,
2743 node->eb->start, blocksize,
2744 upper->eb->start,
2745 btrfs_header_owner(upper->eb),
66d7e7f0 2746 node->level, 0, 1);
5d4f98a2
YZ		 2747 BUG_ON(ret);
2748
2749 ret = btrfs_drop_subtree(trans, root, eb, upper->eb);
2750 BUG_ON(ret);
5d4f98a2 2751 }
3fd0a558
YZ		 2752next:
2753 if (!upper->pending)
2754 drop_node_buffer(upper);
2755 else
2756 unlock_node_buffer(upper);
2757 if (err)
2758 break;
5d4f98a2 2759 }
3fd0a558
YZ		 2760
2761 if (!err && node->pending) {
2762 drop_node_buffer(node);
2763 list_move_tail(&node->list, &rc->backref_cache.changed);
2764 node->pending = 0;
2765 }
2766
5d4f98a2 2767 path->lowest_level = 0;
3fd0a558 2768 BUG_ON(err == -ENOSPC);
5d4f98a2
YZ		 2769 return err;
2770}
2771
2772static int link_to_upper(struct btrfs_trans_handle *trans,
3fd0a558 2773 struct reloc_control *rc,
5d4f98a2
YZ		 2774 struct backref_node *node,
2775 struct btrfs_path *path)
2776{
2777 struct btrfs_key key;
5d4f98a2
YZ		 2778
2779 btrfs_node_key_to_cpu(node->eb, &key, 0);
3fd0a558 2780 return do_relocation(trans, rc, node, &key, path, 0);
5d4f98a2
YZ		 2781}
2782
2783static int finish_pending_nodes(struct btrfs_trans_handle *trans,
3fd0a558
YZ		 2784 struct reloc_control *rc,
2785 struct btrfs_path *path, int err)
5d4f98a2 2786{
3fd0a558
YZ		 2787 LIST_HEAD(list);
2788 struct backref_cache *cache = &rc->backref_cache;
5d4f98a2
YZ		 2789 struct backref_node *node;
2790 int level;
2791 int ret;
5d4f98a2
YZ		 2792
2793 for (level = 0; level < BTRFS_MAX_LEVEL; level++) {
2794 while (!list_empty(&cache->pending[level])) {
2795 node = list_entry(cache->pending[level].next,
3fd0a558
YZ		 2796 struct backref_node, list);
2797 list_move_tail(&node->list, &list);
2798 BUG_ON(!node->pending);
5d4f98a2 2799
3fd0a558
YZ		 2800 if (!err) {
2801 ret = link_to_upper(trans, rc, node, path);
2802 if (ret < 0)
2803 err = ret;
2804 }
5d4f98a2 2805 }
3fd0a558 2806 list_splice_init(&list, &cache->pending[level]);
5d4f98a2 2807 }
5d4f98a2
YZ		 2808 return err;
2809}
2810
2811static void mark_block_processed(struct reloc_control *rc,
3fd0a558
YZ		 2812 u64 bytenr, u32 blocksize)
2813{
2814 set_extent_bits(&rc->processed_blocks, bytenr, bytenr + blocksize - 1,
2815 EXTENT_DIRTY, GFP_NOFS);
2816}
2817
2818static void __mark_block_processed(struct reloc_control *rc,
2819 struct backref_node *node)
5d4f98a2
YZ		 2820{
2821 u32 blocksize;
2822 if (node->level == 0 ||
2823 in_block_group(node->bytenr, rc->block_group)) {
707e8a07 2824 blocksize = rc->extent_root->nodesize;
3fd0a558 2825 mark_block_processed(rc, node->bytenr, blocksize);
5d4f98a2
YZ		 2826 }
2827 node->processed = 1;
2828}
2829
2830/*
2831 * mark a block and all blocks directly/indirectly reference the block
2832 * as processed.
2833 */
2834static void update_processed_blocks(struct reloc_control *rc,
2835 struct backref_node *node)
2836{
2837 struct backref_node *next = node;
2838 struct backref_edge *edge;
2839 struct backref_edge *edges[BTRFS_MAX_LEVEL - 1];
2840 int index = 0;
2841
2842 while (next) {
2843 cond_resched();
2844 while (1) {
2845 if (next->processed)
2846 break;
2847
3fd0a558 2848 __mark_block_processed(rc, next);
5d4f98a2
YZ		 2849
2850 if (list_empty(&next->upper))
2851 break;
2852
2853 edge = list_entry(next->upper.next,
2854 struct backref_edge, list[LOWER]);
2855 edges[index++] = edge;
2856 next = edge->node[UPPER];
2857 }
2858 next = walk_down_backref(edges, &index);
2859 }
2860}
2861
3fd0a558
YZ		 2862static int tree_block_processed(u64 bytenr, u32 blocksize,
2863 struct reloc_control *rc)
2864{
2865 if (test_range_bit(&rc->processed_blocks, bytenr,
2866 bytenr + blocksize - 1, EXTENT_DIRTY, 1, NULL))
2867 return 1;
2868 return 0;
5d4f98a2
YZ		 2869}
2870
2871static int get_tree_block_key(struct reloc_control *rc,
2872 struct tree_block *block)
2873{
2874 struct extent_buffer *eb;
2875
2876 BUG_ON(block->key_ready);
2877 eb = read_tree_block(rc->extent_root, block->bytenr,
2878 block->key.objectid, block->key.offset);
416bc658
JB		 2879 if (!eb || !extent_buffer_uptodate(eb)) {
2880 free_extent_buffer(eb);
2881 return -EIO;
2882 }
5d4f98a2
YZ		 2883 WARN_ON(btrfs_header_level(eb) != block->level);
2884 if (block->level == 0)
2885 btrfs_item_key_to_cpu(eb, &block->key, 0);
2886 else
2887 btrfs_node_key_to_cpu(eb, &block->key, 0);
2888 free_extent_buffer(eb);
2889 block->key_ready = 1;
2890 return 0;
2891}
2892
2893static int reada_tree_block(struct reloc_control *rc,
2894 struct tree_block *block)
2895{
2896 BUG_ON(block->key_ready);
3173a18f
JB		 2897 if (block->key.type == BTRFS_METADATA_ITEM_KEY)
2898 readahead_tree_block(rc->extent_root, block->bytenr,
2899 block->key.objectid,
707e8a07 2900 rc->extent_root->nodesize);
3173a18f
JB		 2901 else
2902 readahead_tree_block(rc->extent_root, block->bytenr,
2903 block->key.objectid, block->key.offset);
5d4f98a2
YZ		 2904 return 0;
2905}
2906
2907/*
2908 * helper function to relocate a tree block
2909 */
2910static int relocate_tree_block(struct btrfs_trans_handle *trans,
2911 struct reloc_control *rc,
2912 struct backref_node *node,
2913 struct btrfs_key *key,
2914 struct btrfs_path *path)
2915{
2916 struct btrfs_root *root;
3fd0a558
YZ		 2917 int ret = 0;
2918
2919 if (!node)
2920 return 0;
5d4f98a2 2921
3fd0a558 2922 BUG_ON(node->processed);
5d4f98a2 2923 root = select_one_root(trans, node);
3fd0a558 2924 if (root == ERR_PTR(-ENOENT)) {
5d4f98a2 2925 update_processed_blocks(rc, node);
3fd0a558 2926 goto out;
5d4f98a2
YZ		 2927 }
2928
27cdeb70 2929 if (!root || test_bit(BTRFS_ROOT_REF_COWS, &root->state)) {
3fd0a558
YZ		 2930 ret = reserve_metadata_space(trans, rc, node);
2931 if (ret)
5d4f98a2 2932 goto out;
5d4f98a2
YZ		 2933 }
2934
3fd0a558 2935 if (root) {
27cdeb70 2936 if (test_bit(BTRFS_ROOT_REF_COWS, &root->state)) {
3fd0a558
YZ		 2937 BUG_ON(node->new_bytenr);
2938 BUG_ON(!list_empty(&node->list));
2939 btrfs_record_root_in_trans(trans, root);
2940 root = root->reloc_root;
2941 node->new_bytenr = root->node->start;
2942 node->root = root;
2943 list_add_tail(&node->list, &rc->backref_cache.changed);
2944 } else {
2945 path->lowest_level = node->level;
2946 ret = btrfs_search_slot(trans, root, key, path, 0, 1);
b3b4aa74 2947 btrfs_release_path(path);
3fd0a558
YZ		 2948 if (ret > 0)
2949 ret = 0;
2950 }
2951 if (!ret)
2952 update_processed_blocks(rc, node);
2953 } else {
2954 ret = do_relocation(trans, rc, node, key, path, 1);
2955 }
5d4f98a2 2956out:
0647bf56 2957 if (ret || node->level == 0 || node->cowonly)
3fd0a558 2958 remove_backref_node(&rc->backref_cache, node);
5d4f98a2
YZ		 2959 return ret;
2960}
2961
2962/*
2963 * relocate a list of blocks
2964 */
2965static noinline_for_stack
2966int relocate_tree_blocks(struct btrfs_trans_handle *trans,
2967 struct reloc_control *rc, struct rb_root *blocks)
2968{
5d4f98a2
YZ		 2969 struct backref_node *node;
2970 struct btrfs_path *path;
2971 struct tree_block *block;
2972 struct rb_node *rb_node;
5d4f98a2
YZ		 2973 int ret;
2974 int err = 0;
2975
2976 path = btrfs_alloc_path();
e1a12670
LB		 2977 if (!path) {
2978 err = -ENOMEM;
34c2b290 2979 goto out_free_blocks;
e1a12670 2980 }
5d4f98a2 2981
5d4f98a2
YZ		 2982 rb_node = rb_first(blocks);
2983 while (rb_node) {
2984 block = rb_entry(rb_node, struct tree_block, rb_node);
5d4f98a2
YZ		 2985 if (!block->key_ready)
2986 reada_tree_block(rc, block);
2987 rb_node = rb_next(rb_node);
2988 }
2989
2990 rb_node = rb_first(blocks);
2991 while (rb_node) {
2992 block = rb_entry(rb_node, struct tree_block, rb_node);
34c2b290
DS		 2993 if (!block->key_ready) {
2994 err = get_tree_block_key(rc, block);
2995 if (err)
2996 goto out_free_path;
2997 }
5d4f98a2
YZ		 2998 rb_node = rb_next(rb_node);
2999 }
3000
3001 rb_node = rb_first(blocks);
3002 while (rb_node) {
3003 block = rb_entry(rb_node, struct tree_block, rb_node);
3004
3fd0a558 3005 node = build_backref_tree(rc, &block->key,
5d4f98a2
YZ		 3006 block->level, block->bytenr);
3007 if (IS_ERR(node)) {
3008 err = PTR_ERR(node);
3009 goto out;
3010 }
3011
3012 ret = relocate_tree_block(trans, rc, node, &block->key,
3013 path);
3014 if (ret < 0) {
3fd0a558
YZ		 3015 if (ret != -EAGAIN || rb_node == rb_first(blocks))
3016 err = ret;
5d4f98a2
YZ		 3017 goto out;
3018 }
5d4f98a2
YZ		 3019 rb_node = rb_next(rb_node);
3020 }
5d4f98a2 3021out:
3fd0a558 3022 err = finish_pending_nodes(trans, rc, path, err);
5d4f98a2 3023
34c2b290 3024out_free_path:
5d4f98a2 3025 btrfs_free_path(path);
34c2b290 3026out_free_blocks:
e1a12670 3027 free_block_list(blocks);
5d4f98a2
YZ		 3028 return err;
3029}
3030
efa56464
YZ		 3031static noinline_for_stack
3032int prealloc_file_extent_cluster(struct inode *inode,
3033 struct file_extent_cluster *cluster)
3034{
3035 u64 alloc_hint = 0;
3036 u64 start;
3037 u64 end;
3038 u64 offset = BTRFS_I(inode)->index_cnt;
3039 u64 num_bytes;
3040 int nr = 0;
3041 int ret = 0;
3042
3043 BUG_ON(cluster->start != cluster->boundary[0]);
3044 mutex_lock(&inode->i_mutex);
3045
3046 ret = btrfs_check_data_free_space(inode, cluster->end +
3047 1 - cluster->start);
3048 if (ret)
3049 goto out;
3050
3051 while (nr < cluster->nr) {
3052 start = cluster->boundary[nr] - offset;
3053 if (nr + 1 < cluster->nr)
3054 end = cluster->boundary[nr + 1] - 1 - offset;
3055 else
3056 end = cluster->end - offset;
3057
d0082371 3058 lock_extent(&BTRFS_I(inode)->io_tree, start, end);
efa56464
YZ		 3059 num_bytes = end + 1 - start;
3060 ret = btrfs_prealloc_file_range(inode, 0, start,
3061 num_bytes, num_bytes,
3062 end + 1, &alloc_hint);
d0082371 3063 unlock_extent(&BTRFS_I(inode)->io_tree, start, end);
efa56464
YZ		 3064 if (ret)
3065 break;
3066 nr++;
3067 }
3068 btrfs_free_reserved_data_space(inode, cluster->end +
3069 1 - cluster->start);
3070out:
3071 mutex_unlock(&inode->i_mutex);
3072 return ret;
3073}
3074
5d4f98a2 3075static noinline_for_stack
0257bb82
YZ		 3076int setup_extent_mapping(struct inode *inode, u64 start, u64 end,
3077 u64 block_start)
3078{
3079 struct btrfs_root *root = BTRFS_I(inode)->root;
3080 struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
3081 struct extent_map *em;
3082 int ret = 0;
3083
172ddd60 3084 em = alloc_extent_map();
0257bb82
YZ		 3085 if (!em)
3086 return -ENOMEM;
3087
3088 em->start = start;
3089 em->len = end + 1 - start;
3090 em->block_len = em->len;
3091 em->block_start = block_start;
3092 em->bdev = root->fs_info->fs_devices->latest_bdev;
3093 set_bit(EXTENT_FLAG_PINNED, &em->flags);
3094
d0082371 3095 lock_extent(&BTRFS_I(inode)->io_tree, start, end);
0257bb82
YZ		 3096 while (1) {
3097 write_lock(&em_tree->lock);
09a2a8f9 3098 ret = add_extent_mapping(em_tree, em, 0);
0257bb82
YZ		 3099 write_unlock(&em_tree->lock);
3100 if (ret != -EEXIST) {
3101 free_extent_map(em);
3102 break;
3103 }
3104 btrfs_drop_extent_cache(inode, start, end, 0);
3105 }
d0082371 3106 unlock_extent(&BTRFS_I(inode)->io_tree, start, end);
0257bb82
YZ		 3107 return ret;
3108}
3109
3110static int relocate_file_extent_cluster(struct inode *inode,
3111 struct file_extent_cluster *cluster)
5d4f98a2
YZ		 3112{
3113 u64 page_start;
3114 u64 page_end;
0257bb82
YZ		 3115 u64 offset = BTRFS_I(inode)->index_cnt;
3116 unsigned long index;
5d4f98a2 3117 unsigned long last_index;
5d4f98a2
YZ		 3118 struct page *page;
3119 struct file_ra_state *ra;
3b16a4e3 3120 gfp_t mask = btrfs_alloc_write_mask(inode->i_mapping);
0257bb82 3121 int nr = 0;
5d4f98a2
YZ		 3122 int ret = 0;
3123
0257bb82
YZ		 3124 if (!cluster->nr)
3125 return 0;
3126
5d4f98a2
YZ		 3127 ra = kzalloc(sizeof(*ra), GFP_NOFS);
3128 if (!ra)
3129 return -ENOMEM;
3130
efa56464
YZ		 3131 ret = prealloc_file_extent_cluster(inode, cluster);
3132 if (ret)
3133 goto out;
0257bb82 3134
efa56464 3135 file_ra_state_init(ra, inode->i_mapping);
5d4f98a2 3136
0257bb82
YZ		 3137 ret = setup_extent_mapping(inode, cluster->start - offset,
3138 cluster->end - offset, cluster->start);
5d4f98a2 3139 if (ret)
efa56464 3140 goto out;
5d4f98a2 3141
efa56464
YZ		 3142 index = (cluster->start - offset) >> PAGE_CACHE_SHIFT;
3143 last_index = (cluster->end - offset) >> PAGE_CACHE_SHIFT;
0257bb82 3144 while (index <= last_index) {
efa56464
YZ		 3145 ret = btrfs_delalloc_reserve_metadata(inode, PAGE_CACHE_SIZE);
3146 if (ret)
3147 goto out;
3148
0257bb82 3149 page = find_lock_page(inode->i_mapping, index);
5d4f98a2 3150 if (!page) {
0257bb82
YZ		 3151 page_cache_sync_readahead(inode->i_mapping,
3152 ra, NULL, index,
3153 last_index + 1 - index);
a94733d0 3154 page = find_or_create_page(inode->i_mapping, index,
3b16a4e3 3155 mask);
0257bb82 3156 if (!page) {
efa56464
YZ		 3157 btrfs_delalloc_release_metadata(inode,
3158 PAGE_CACHE_SIZE);
0257bb82 3159 ret = -ENOMEM;
efa56464 3160 goto out;
0257bb82 3161 }
5d4f98a2 3162 }
0257bb82
YZ		 3163
3164 if (PageReadahead(page)) {
3165 page_cache_async_readahead(inode->i_mapping,
3166 ra, NULL, page, index,
3167 last_index + 1 - index);
3168 }
3169
5d4f98a2
YZ		 3170 if (!PageUptodate(page)) {
3171 btrfs_readpage(NULL, page);
3172 lock_page(page);
3173 if (!PageUptodate(page)) {
3174 unlock_page(page);
3175 page_cache_release(page);
efa56464
YZ		 3176 btrfs_delalloc_release_metadata(inode,
3177 PAGE_CACHE_SIZE);
5d4f98a2 3178 ret = -EIO;
efa56464 3179 goto out;
5d4f98a2
YZ		 3180 }
3181 }
5d4f98a2 3182
4eee4fa4 3183 page_start = page_offset(page);
5d4f98a2 3184 page_end = page_start + PAGE_CACHE_SIZE - 1;
0257bb82 3185
d0082371 3186 lock_extent(&BTRFS_I(inode)->io_tree, page_start, page_end);
0257bb82 3187
5d4f98a2
YZ		 3188 set_page_extent_mapped(page);
3189
0257bb82
YZ		 3190 if (nr < cluster->nr &&
3191 page_start + offset == cluster->boundary[nr]) {
3192 set_extent_bits(&BTRFS_I(inode)->io_tree,
3193 page_start, page_end,
5d4f98a2 3194 EXTENT_BOUNDARY, GFP_NOFS);
0257bb82
YZ		 3195 nr++;
3196 }
5d4f98a2 3197
efa56464 3198 btrfs_set_extent_delalloc(inode, page_start, page_end, NULL);
5d4f98a2 3199 set_page_dirty(page);
5d4f98a2 3200
0257bb82 3201 unlock_extent(&BTRFS_I(inode)->io_tree,
d0082371 3202 page_start, page_end);
5d4f98a2
YZ		 3203 unlock_page(page);
3204 page_cache_release(page);
0257bb82
YZ		 3205
3206 index++;
efa56464
YZ		 3207 balance_dirty_pages_ratelimited(inode->i_mapping);
3208 btrfs_throttle(BTRFS_I(inode)->root);
5d4f98a2 3209 }
0257bb82 3210 WARN_ON(nr != cluster->nr);
efa56464 3211out:
5d4f98a2 3212 kfree(ra);
5d4f98a2
YZ		 3213 return ret;
3214}
3215
3216static noinline_for_stack
0257bb82
YZ		 3217int relocate_data_extent(struct inode *inode, struct btrfs_key *extent_key,
3218 struct file_extent_cluster *cluster)
5d4f98a2 3219{
0257bb82 3220 int ret;
5d4f98a2 3221
0257bb82
YZ		 3222 if (cluster->nr > 0 && extent_key->objectid != cluster->end + 1) {
3223 ret = relocate_file_extent_cluster(inode, cluster);
3224 if (ret)
3225 return ret;
3226 cluster->nr = 0;
5d4f98a2 3227 }
5d4f98a2 3228
0257bb82
YZ		 3229 if (!cluster->nr)
3230 cluster->start = extent_key->objectid;
3231 else
3232 BUG_ON(cluster->nr >= MAX_EXTENTS);
3233 cluster->end = extent_key->objectid + extent_key->offset - 1;
3234 cluster->boundary[cluster->nr] = extent_key->objectid;
3235 cluster->nr++;
3236
3237 if (cluster->nr >= MAX_EXTENTS) {
3238 ret = relocate_file_extent_cluster(inode, cluster);
3239 if (ret)
3240 return ret;
3241 cluster->nr = 0;
3242 }
3243 return 0;
5d4f98a2
YZ		 3244}
3245
3246#ifdef BTRFS_COMPAT_EXTENT_TREE_V0
3247static int get_ref_objectid_v0(struct reloc_control *rc,
3248 struct btrfs_path *path,
3249 struct btrfs_key *extent_key,
3250 u64 *ref_objectid, int *path_change)
3251{
3252 struct btrfs_key key;
3253 struct extent_buffer *leaf;
3254 struct btrfs_extent_ref_v0 *ref0;
3255 int ret;
3256 int slot;
3257
3258 leaf = path->nodes[0];
3259 slot = path->slots[0];
3260 while (1) {
3261 if (slot >= btrfs_header_nritems(leaf)) {
3262 ret = btrfs_next_leaf(rc->extent_root, path);
3263 if (ret < 0)
3264 return ret;
3265 BUG_ON(ret > 0);
3266 leaf = path->nodes[0];
3267 slot = path->slots[0];
3268 if (path_change)
3269 *path_change = 1;
3270 }
3271 btrfs_item_key_to_cpu(leaf, &key, slot);
3272 if (key.objectid != extent_key->objectid)
3273 return -ENOENT;
3274
3275 if (key.type != BTRFS_EXTENT_REF_V0_KEY) {
3276 slot++;
3277 continue;
3278 }
3279 ref0 = btrfs_item_ptr(leaf, slot,
3280 struct btrfs_extent_ref_v0);
3281 *ref_objectid = btrfs_ref_objectid_v0(leaf, ref0);
3282 break;
3283 }
3284 return 0;
3285}
3286#endif
3287
3288/*
3289 * helper to add a tree block to the list.
3290 * the major work is getting the generation and level of the block
3291 */
3292static int add_tree_block(struct reloc_control *rc,
3293 struct btrfs_key *extent_key,
3294 struct btrfs_path *path,
3295 struct rb_root *blocks)
3296{
3297 struct extent_buffer *eb;
3298 struct btrfs_extent_item *ei;
3299 struct btrfs_tree_block_info *bi;
3300 struct tree_block *block;
3301 struct rb_node *rb_node;
3302 u32 item_size;
3303 int level = -1;
7fdf4b60 3304 u64 generation;
5d4f98a2
YZ		 3305
3306 eb = path->nodes[0];
3307 item_size = btrfs_item_size_nr(eb, path->slots[0]);
3308
3173a18f
JB		 3309 if (extent_key->type == BTRFS_METADATA_ITEM_KEY ||
3310 item_size >= sizeof(*ei) + sizeof(*bi)) {
5d4f98a2
YZ		 3311 ei = btrfs_item_ptr(eb, path->slots[0],
3312 struct btrfs_extent_item);
3173a18f
JB		 3313 if (extent_key->type == BTRFS_EXTENT_ITEM_KEY) {
3314 bi = (struct btrfs_tree_block_info *)(ei + 1);
3315 level = btrfs_tree_block_level(eb, bi);
3316 } else {
3317 level = (int)extent_key->offset;
3318 }
5d4f98a2 3319 generation = btrfs_extent_generation(eb, ei);
5d4f98a2
YZ		 3320 } else {
3321#ifdef BTRFS_COMPAT_EXTENT_TREE_V0
3322 u64 ref_owner;
3323 int ret;
3324
3325 BUG_ON(item_size != sizeof(struct btrfs_extent_item_v0));
3326 ret = get_ref_objectid_v0(rc, path, extent_key,
3327 &ref_owner, NULL);
411fc6bc
AK		 3328 if (ret < 0)
3329 return ret;
5d4f98a2
YZ		 3330 BUG_ON(ref_owner >= BTRFS_MAX_LEVEL);
3331 level = (int)ref_owner;
3332 /* FIXME: get real generation */
3333 generation = 0;
3334#else
3335 BUG();
3336#endif
3337 }
3338
b3b4aa74 3339 btrfs_release_path(path);
5d4f98a2
YZ		 3340
3341 BUG_ON(level == -1);
3342
3343 block = kmalloc(sizeof(*block), GFP_NOFS);
3344 if (!block)
3345 return -ENOMEM;
3346
3347 block->bytenr = extent_key->objectid;
707e8a07 3348 block->key.objectid = rc->extent_root->nodesize;
5d4f98a2
YZ		 3349 block->key.offset = generation;
3350 block->level = level;
3351 block->key_ready = 0;
3352
3353 rb_node = tree_insert(blocks, block->bytenr, &block->rb_node);
43c04fb1
JM		 3354 if (rb_node)
3355 backref_tree_panic(rb_node, -EEXIST, block->bytenr);
5d4f98a2
YZ		 3356
3357 return 0;
3358}
3359
3360/*
3361 * helper to add tree blocks for backref of type BTRFS_SHARED_DATA_REF_KEY
3362 */
3363static int __add_tree_block(struct reloc_control *rc,
3364 u64 bytenr, u32 blocksize,
3365 struct rb_root *blocks)
3366{
3367 struct btrfs_path *path;
3368 struct btrfs_key key;
3369 int ret;
aee68ee5
JB		 3370 bool skinny = btrfs_fs_incompat(rc->extent_root->fs_info,
3371 SKINNY_METADATA);
5d4f98a2
YZ		 3372
3373 if (tree_block_processed(bytenr, blocksize, rc))
3374 return 0;
3375
3376 if (tree_search(blocks, bytenr))
3377 return 0;
3378
3379 path = btrfs_alloc_path();
3380 if (!path)
3381 return -ENOMEM;
aee68ee5 3382again:
5d4f98a2 3383 key.objectid = bytenr;
aee68ee5
JB		 3384 if (skinny) {
3385 key.type = BTRFS_METADATA_ITEM_KEY;
3386 key.offset = (u64)-1;
3387 } else {
3388 key.type = BTRFS_EXTENT_ITEM_KEY;
3389 key.offset = blocksize;
3390 }
5d4f98a2
YZ		 3391
3392 path->search_commit_root = 1;
3393 path->skip_locking = 1;
3394 ret = btrfs_search_slot(NULL, rc->extent_root, &key, path, 0, 0);
3395 if (ret < 0)
3396 goto out;
5d4f98a2 3397
aee68ee5
JB		 3398 if (ret > 0 && skinny) {
3399 if (path->slots[0]) {
3400 path->slots[0]--;
3401 btrfs_item_key_to_cpu(path->nodes[0], &key,
3402 path->slots[0]);
3403 if (key.objectid == bytenr &&
3404 (key.type == BTRFS_METADATA_ITEM_KEY ||
3405 (key.type == BTRFS_EXTENT_ITEM_KEY &&
3406 key.offset == blocksize)))
3407 ret = 0;
3408 }
3409
3410 if (ret) {
3411 skinny = false;
3412 btrfs_release_path(path);
3413 goto again;
3414 }
3173a18f
JB		 3415 }
3416 BUG_ON(ret);
3417
5d4f98a2
YZ		 3418 ret = add_tree_block(rc, &key, path, blocks);
3419out:
3420 btrfs_free_path(path);
3421 return ret;
3422}
3423
3424/*
3425 * helper to check if the block use full backrefs for pointers in it
3426 */
3427static int block_use_full_backref(struct reloc_control *rc,
3428 struct extent_buffer *eb)
3429{
5d4f98a2
YZ		 3430 u64 flags;
3431 int ret;
3432
3433 if (btrfs_header_flag(eb, BTRFS_HEADER_FLAG_RELOC) ||
3434 btrfs_header_backref_rev(eb) < BTRFS_MIXED_BACKREF_REV)
3435 return 1;
3436
3fd0a558 3437 ret = btrfs_lookup_extent_info(NULL, rc->extent_root,
3173a18f
JB		 3438 eb->start, btrfs_header_level(eb), 1,
3439 NULL, &flags);
5d4f98a2
YZ		 3440 BUG_ON(ret);
3441
5d4f98a2
YZ		 3442 if (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF)
3443 ret = 1;
3444 else
3445 ret = 0;
5d4f98a2
YZ		 3446 return ret;
3447}
3448
0af3d00b
JB		 3449static int delete_block_group_cache(struct btrfs_fs_info *fs_info,
3450 struct inode *inode, u64 ino)
3451{
3452 struct btrfs_key key;
0af3d00b
JB		 3453 struct btrfs_root *root = fs_info->tree_root;
3454 struct btrfs_trans_handle *trans;
0af3d00b
JB		 3455 int ret = 0;
3456
3457 if (inode)
3458 goto truncate;
3459
3460 key.objectid = ino;
3461 key.type = BTRFS_INODE_ITEM_KEY;
3462 key.offset = 0;
3463
3464 inode = btrfs_iget(fs_info->sb, &key, root, NULL);
f54fb859
TI		 3465 if (IS_ERR(inode) || is_bad_inode(inode)) {
3466 if (!IS_ERR(inode))
0af3d00b
JB		 3467 iput(inode);
3468 return -ENOENT;
3469 }
3470
3471truncate:
7b61cd92
MX		 3472 ret = btrfs_check_trunc_cache_free_space(root,
3473 &fs_info->global_block_rsv);
3474 if (ret)
3475 goto out;
3476
7a7eaa40 3477 trans = btrfs_join_transaction(root);
0af3d00b 3478 if (IS_ERR(trans)) {
3612b495 3479 ret = PTR_ERR(trans);
0af3d00b
JB		 3480 goto out;
3481 }
3482
74514323 3483 ret = btrfs_truncate_free_space_cache(root, trans, inode);
0af3d00b 3484
0af3d00b 3485 btrfs_end_transaction(trans, root);
b53d3f5d 3486 btrfs_btree_balance_dirty(root);
0af3d00b
JB		 3487out:
3488 iput(inode);
3489 return ret;
3490}
3491
5d4f98a2
YZ		 3492/*
3493 * helper to add tree blocks for backref of type BTRFS_EXTENT_DATA_REF_KEY
3494 * this function scans fs tree to find blocks reference the data extent
3495 */
3496static int find_data_references(struct reloc_control *rc,
3497 struct btrfs_key *extent_key,
3498 struct extent_buffer *leaf,
3499 struct btrfs_extent_data_ref *ref,
3500 struct rb_root *blocks)
3501{
3502 struct btrfs_path *path;
3503 struct tree_block *block;
3504 struct btrfs_root *root;
3505 struct btrfs_file_extent_item *fi;
3506 struct rb_node *rb_node;
3507 struct btrfs_key key;
3508 u64 ref_root;
3509 u64 ref_objectid;
3510 u64 ref_offset;
3511 u32 ref_count;
3512 u32 nritems;
3513 int err = 0;
3514 int added = 0;
3515 int counted;
3516 int ret;
3517
5d4f98a2
YZ		 3518 ref_root = btrfs_extent_data_ref_root(leaf, ref);
3519 ref_objectid = btrfs_extent_data_ref_objectid(leaf, ref);
3520 ref_offset = btrfs_extent_data_ref_offset(leaf, ref);
3521 ref_count = btrfs_extent_data_ref_count(leaf, ref);
3522
0af3d00b
JB		 3523 /*
3524 * This is an extent belonging to the free space cache, lets just delete
3525 * it and redo the search.
3526 */
3527 if (ref_root == BTRFS_ROOT_TREE_OBJECTID) {
3528 ret = delete_block_group_cache(rc->extent_root->fs_info,
3529 NULL, ref_objectid);
3530 if (ret != -ENOENT)
3531 return ret;
3532 ret = 0;
3533 }
3534
3535 path = btrfs_alloc_path();
3536 if (!path)
3537 return -ENOMEM;
026fd317 3538 path->reada = 1;
0af3d00b 3539
5d4f98a2
YZ		 3540 root = read_fs_root(rc->extent_root->fs_info, ref_root);
3541 if (IS_ERR(root)) {
3542 err = PTR_ERR(root);
3543 goto out;
3544 }
3545
3546 key.objectid = ref_objectid;
5d4f98a2 3547 key.type = BTRFS_EXTENT_DATA_KEY;
84850e8d
YZ		 3548 if (ref_offset > ((u64)-1 << 32))
3549 key.offset = 0;
3550 else
3551 key.offset = ref_offset;
5d4f98a2
YZ		 3552
3553 path->search_commit_root = 1;
3554 path->skip_locking = 1;
3555 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
3556 if (ret < 0) {
3557 err = ret;
3558 goto out;
3559 }
3560
3561 leaf = path->nodes[0];
3562 nritems = btrfs_header_nritems(leaf);
3563 /*
3564 * the references in tree blocks that use full backrefs
3565 * are not counted in
3566 */
3567 if (block_use_full_backref(rc, leaf))
3568 counted = 0;
3569 else
3570 counted = 1;
3571 rb_node = tree_search(blocks, leaf->start);
3572 if (rb_node) {
3573 if (counted)
3574 added = 1;
3575 else
3576 path->slots[0] = nritems;
3577 }
3578
3579 while (ref_count > 0) {
3580 while (path->slots[0] >= nritems) {
3581 ret = btrfs_next_leaf(root, path);
3582 if (ret < 0) {
3583 err = ret;
3584 goto out;
3585 }
fae7f21c 3586 if (WARN_ON(ret > 0))
5d4f98a2 3587 goto out;
5d4f98a2
YZ		 3588
3589 leaf = path->nodes[0];
3590 nritems = btrfs_header_nritems(leaf);
3591 added = 0;
3592
3593 if (block_use_full_backref(rc, leaf))
3594 counted = 0;
3595 else
3596 counted = 1;
3597 rb_node = tree_search(blocks, leaf->start);
3598 if (rb_node) {
3599 if (counted)
3600 added = 1;
3601 else
3602 path->slots[0] = nritems;
3603 }
3604 }
3605
3606 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
fae7f21c
DG		 3607 if (WARN_ON(key.objectid != ref_objectid ||
3608 key.type != BTRFS_EXTENT_DATA_KEY))
5d4f98a2 3609 break;
5d4f98a2
YZ		 3610
3611 fi = btrfs_item_ptr(leaf, path->slots[0],
3612 struct btrfs_file_extent_item);
3613
3614 if (btrfs_file_extent_type(leaf, fi) ==
3615 BTRFS_FILE_EXTENT_INLINE)
3616 goto next;
3617
3618 if (btrfs_file_extent_disk_bytenr(leaf, fi) !=
3619 extent_key->objectid)
3620 goto next;
3621
3622 key.offset -= btrfs_file_extent_offset(leaf, fi);
3623 if (key.offset != ref_offset)
3624 goto next;
3625
3626 if (counted)
3627 ref_count--;
3628 if (added)
3629 goto next;
3630
3631 if (!tree_block_processed(leaf->start, leaf->len, rc)) {
3632 block = kmalloc(sizeof(*block), GFP_NOFS);
3633 if (!block) {
3634 err = -ENOMEM;
3635 break;
3636 }
3637 block->bytenr = leaf->start;
3638 btrfs_item_key_to_cpu(leaf, &block->key, 0);
3639 block->level = 0;
3640 block->key_ready = 1;
3641 rb_node = tree_insert(blocks, block->bytenr,
3642 &block->rb_node);
43c04fb1
JM		 3643 if (rb_node)
3644 backref_tree_panic(rb_node, -EEXIST,
3645 block->bytenr);
5d4f98a2
YZ		 3646 }
3647 if (counted)
3648 added = 1;
3649 else
3650 path->slots[0] = nritems;
3651next:
3652 path->slots[0]++;
3653
3654 }
3655out:
3656 btrfs_free_path(path);
3657 return err;
3658}
3659
3660/*
2c016dc2 3661 * helper to find all tree blocks that reference a given data extent
5d4f98a2
YZ		 3662 */
3663static noinline_for_stack
3664int add_data_references(struct reloc_control *rc,
3665 struct btrfs_key *extent_key,
3666 struct btrfs_path *path,
3667 struct rb_root *blocks)
3668{
3669 struct btrfs_key key;
3670 struct extent_buffer *eb;
3671 struct btrfs_extent_data_ref *dref;
3672 struct btrfs_extent_inline_ref *iref;
3673 unsigned long ptr;
3674 unsigned long end;
707e8a07 3675 u32 blocksize = rc->extent_root->nodesize;
647f63bd 3676 int ret = 0;
5d4f98a2
YZ		 3677 int err = 0;
3678
5d4f98a2
YZ		 3679 eb = path->nodes[0];
3680 ptr = btrfs_item_ptr_offset(eb, path->slots[0]);
3681 end = ptr + btrfs_item_size_nr(eb, path->slots[0]);
3682#ifdef BTRFS_COMPAT_EXTENT_TREE_V0
3683 if (ptr + sizeof(struct btrfs_extent_item_v0) == end)
3684 ptr = end;
3685 else
3686#endif
3687 ptr += sizeof(struct btrfs_extent_item);
3688
3689 while (ptr < end) {
3690 iref = (struct btrfs_extent_inline_ref *)ptr;
3691 key.type = btrfs_extent_inline_ref_type(eb, iref);
3692 if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
3693 key.offset = btrfs_extent_inline_ref_offset(eb, iref);
3694 ret = __add_tree_block(rc, key.offset, blocksize,
3695 blocks);
3696 } else if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
3697 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
3698 ret = find_data_references(rc, extent_key,
3699 eb, dref, blocks);
3700 } else {
3701 BUG();
3702 }
647f63bd
FDBM		 3703 if (ret) {
3704 err = ret;
3705 goto out;
3706 }
5d4f98a2
YZ		 3707 ptr += btrfs_extent_inline_ref_size(key.type);
3708 }
3709 WARN_ON(ptr > end);
3710
3711 while (1) {
3712 cond_resched();
3713 eb = path->nodes[0];
3714 if (path->slots[0] >= btrfs_header_nritems(eb)) {
3715 ret = btrfs_next_leaf(rc->extent_root, path);
3716 if (ret < 0) {
3717 err = ret;
3718 break;
3719 }
3720 if (ret > 0)
3721 break;
3722 eb = path->nodes[0];
3723 }
3724
3725 btrfs_item_key_to_cpu(eb, &key, path->slots[0]);
3726 if (key.objectid != extent_key->objectid)
3727 break;
3728
3729#ifdef BTRFS_COMPAT_EXTENT_TREE_V0
3730 if (key.type == BTRFS_SHARED_DATA_REF_KEY ||
3731 key.type == BTRFS_EXTENT_REF_V0_KEY) {
3732#else
3733 BUG_ON(key.type == BTRFS_EXTENT_REF_V0_KEY);
3734 if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
3735#endif
3736 ret = __add_tree_block(rc, key.offset, blocksize,
3737 blocks);
3738 } else if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
3739 dref = btrfs_item_ptr(eb, path->slots[0],
3740 struct btrfs_extent_data_ref);
3741 ret = find_data_references(rc, extent_key,
3742 eb, dref, blocks);
3743 } else {
3744 ret = 0;
3745 }
3746 if (ret) {
3747 err = ret;
3748 break;
3749 }
3750 path->slots[0]++;
3751 }
647f63bd 3752out:
b3b4aa74 3753 btrfs_release_path(path);
5d4f98a2
YZ		 3754 if (err)
3755 free_block_list(blocks);
3756 return err;
3757}
3758
3759/*
2c016dc2 3760 * helper to find next unprocessed extent
5d4f98a2
YZ		 3761 */
3762static noinline_for_stack
3763int find_next_extent(struct btrfs_trans_handle *trans,
3fd0a558
YZ		 3764 struct reloc_control *rc, struct btrfs_path *path,
3765 struct btrfs_key *extent_key)
5d4f98a2
YZ		 3766{
3767 struct btrfs_key key;
3768 struct extent_buffer *leaf;
3769 u64 start, end, last;
3770 int ret;
3771
3772 last = rc->block_group->key.objectid + rc->block_group->key.offset;
3773 while (1) {
3774 cond_resched();
3775 if (rc->search_start >= last) {
3776 ret = 1;
3777 break;
3778 }
3779
3780 key.objectid = rc->search_start;
3781 key.type = BTRFS_EXTENT_ITEM_KEY;
3782 key.offset = 0;
3783
3784 path->search_commit_root = 1;
3785 path->skip_locking = 1;
3786 ret = btrfs_search_slot(NULL, rc->extent_root, &key, path,
3787 0, 0);
3788 if (ret < 0)
3789 break;
3790next:
3791 leaf = path->nodes[0];
3792 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
3793 ret = btrfs_next_leaf(rc->extent_root, path);
3794 if (ret != 0)
3795 break;
3796 leaf = path->nodes[0];
3797 }
3798
3799 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
3800 if (key.objectid >= last) {
3801 ret = 1;
3802 break;
3803 }
3804
3173a18f
JB		 3805 if (key.type != BTRFS_EXTENT_ITEM_KEY &&
3806 key.type != BTRFS_METADATA_ITEM_KEY) {
3807 path->slots[0]++;
3808 goto next;
3809 }
3810
3811 if (key.type == BTRFS_EXTENT_ITEM_KEY &&
5d4f98a2
YZ		 3812 key.objectid + key.offset <= rc->search_start) {
3813 path->slots[0]++;
3814 goto next;
3815 }
3816
3173a18f 3817 if (key.type == BTRFS_METADATA_ITEM_KEY &&
707e8a07 3818 key.objectid + rc->extent_root->nodesize <=
3173a18f
JB		 3819 rc->search_start) {
3820 path->slots[0]++;
3821 goto next;
3822 }
3823
5d4f98a2
YZ		 3824 ret = find_first_extent_bit(&rc->processed_blocks,
3825 key.objectid, &start, &end,
e6138876 3826 EXTENT_DIRTY, NULL);
5d4f98a2
YZ		 3827
3828 if (ret == 0 && start <= key.objectid) {
b3b4aa74 3829 btrfs_release_path(path);
5d4f98a2
YZ		 3830 rc->search_start = end + 1;
3831 } else {
3173a18f
JB		 3832 if (key.type == BTRFS_EXTENT_ITEM_KEY)
3833 rc->search_start = key.objectid + key.offset;
3834 else
3835 rc->search_start = key.objectid +
707e8a07 3836 rc->extent_root->nodesize;
3fd0a558 3837 memcpy(extent_key, &key, sizeof(key));
5d4f98a2
YZ		 3838 return 0;
3839 }
3840 }
b3b4aa74 3841 btrfs_release_path(path);
5d4f98a2
YZ		 3842 return ret;
3843}
3844
3845static void set_reloc_control(struct reloc_control *rc)
3846{
3847 struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
7585717f
CM		 3848
3849 mutex_lock(&fs_info->reloc_mutex);
5d4f98a2 3850 fs_info->reloc_ctl = rc;
7585717f 3851 mutex_unlock(&fs_info->reloc_mutex);
5d4f98a2
YZ		 3852}
3853
3854static void unset_reloc_control(struct reloc_control *rc)
3855{
3856 struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
7585717f
CM		 3857
3858 mutex_lock(&fs_info->reloc_mutex);
5d4f98a2 3859 fs_info->reloc_ctl = NULL;
7585717f 3860 mutex_unlock(&fs_info->reloc_mutex);
5d4f98a2
YZ		 3861}
3862
3863static int check_extent_flags(u64 flags)
3864{
3865 if ((flags & BTRFS_EXTENT_FLAG_DATA) &&
3866 (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK))
3867 return 1;
3868 if (!(flags & BTRFS_EXTENT_FLAG_DATA) &&
3869 !(flags & BTRFS_EXTENT_FLAG_TREE_BLOCK))
3870 return 1;
3871 if ((flags & BTRFS_EXTENT_FLAG_DATA) &&
3872 (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
3873 return 1;
3874 return 0;
3875}
3876
3fd0a558
YZ		 3877static noinline_for_stack
3878int prepare_to_relocate(struct reloc_control *rc)
3879{
3880 struct btrfs_trans_handle *trans;
3fd0a558 3881
66d8f3dd
MX		 3882 rc->block_rsv = btrfs_alloc_block_rsv(rc->extent_root,
3883 BTRFS_BLOCK_RSV_TEMP);
3fd0a558
YZ		 3884 if (!rc->block_rsv)
3885 return -ENOMEM;
3886
3fd0a558
YZ		 3887 memset(&rc->cluster, 0, sizeof(rc->cluster));
3888 rc->search_start = rc->block_group->key.objectid;
3889 rc->extents_found = 0;
3890 rc->nodes_relocated = 0;
3891 rc->merging_rsv_size = 0;
0647bf56
WS		 3892 rc->reserved_bytes = 0;
3893 rc->block_rsv->size = rc->extent_root->nodesize *
3894 RELOCATION_RESERVED_NODES;
3fd0a558
YZ		 3895
3896 rc->create_reloc_tree = 1;
3897 set_reloc_control(rc);
3898
7a7eaa40 3899 trans = btrfs_join_transaction(rc->extent_root);
28818947
LB		 3900 if (IS_ERR(trans)) {
3901 unset_reloc_control(rc);
3902 /*
3903 * extent tree is not a ref_cow tree and has no reloc_root to
3904 * cleanup. And callers are responsible to free the above
3905 * block rsv.
3906 */
3907 return PTR_ERR(trans);
3908 }
3fd0a558
YZ		 3909 btrfs_commit_transaction(trans, rc->extent_root);
3910 return 0;
3911}
76dda93c 3912
5d4f98a2
YZ		 3913static noinline_for_stack int relocate_block_group(struct reloc_control *rc)
3914{
3915 struct rb_root blocks = RB_ROOT;
3916 struct btrfs_key key;
3917 struct btrfs_trans_handle *trans = NULL;
3918 struct btrfs_path *path;
3919 struct btrfs_extent_item *ei;
5d4f98a2
YZ		 3920 u64 flags;
3921 u32 item_size;
3922 int ret;
3923 int err = 0;
c87f08ca 3924 int progress = 0;
5d4f98a2
YZ		 3925
3926 path = btrfs_alloc_path();
3fd0a558 3927 if (!path)
5d4f98a2 3928 return -ENOMEM;
026fd317 3929 path->reada = 1;
5d4f98a2 3930
3fd0a558
YZ		 3931 ret = prepare_to_relocate(rc);
3932 if (ret) {
3933 err = ret;
3934 goto out_free;
3935 }
5d4f98a2
YZ		 3936
3937 while (1) {
0647bf56
WS		 3938 rc->reserved_bytes = 0;
3939 ret = btrfs_block_rsv_refill(rc->extent_root,
3940 rc->block_rsv, rc->block_rsv->size,
3941 BTRFS_RESERVE_FLUSH_ALL);
3942 if (ret) {
3943 err = ret;
3944 break;
3945 }
c87f08ca 3946 progress++;
a22285a6 3947 trans = btrfs_start_transaction(rc->extent_root, 0);
0f788c58
LB		 3948 if (IS_ERR(trans)) {
3949 err = PTR_ERR(trans);
3950 trans = NULL;
3951 break;
3952 }
c87f08ca 3953restart:
3fd0a558
YZ		 3954 if (update_backref_cache(trans, &rc->backref_cache)) {
3955 btrfs_end_transaction(trans, rc->extent_root);
3956 continue;
3957 }
3958
3959 ret = find_next_extent(trans, rc, path, &key);
5d4f98a2
YZ		 3960 if (ret < 0)
3961 err = ret;
3962 if (ret != 0)
3963 break;
3964
3965 rc->extents_found++;
3966
3967 ei = btrfs_item_ptr(path->nodes[0], path->slots[0],
3968 struct btrfs_extent_item);
3fd0a558 3969 item_size = btrfs_item_size_nr(path->nodes[0], path->slots[0]);
5d4f98a2
YZ		 3970 if (item_size >= sizeof(*ei)) {
3971 flags = btrfs_extent_flags(path->nodes[0], ei);
3972 ret = check_extent_flags(flags);
3973 BUG_ON(ret);
3974
3975 } else {
3976#ifdef BTRFS_COMPAT_EXTENT_TREE_V0
3977 u64 ref_owner;
3978 int path_change = 0;
3979
3980 BUG_ON(item_size !=
3981 sizeof(struct btrfs_extent_item_v0));
3982 ret = get_ref_objectid_v0(rc, path, &key, &ref_owner,
3983 &path_change);
3984 if (ref_owner < BTRFS_FIRST_FREE_OBJECTID)
3985 flags = BTRFS_EXTENT_FLAG_TREE_BLOCK;
3986 else
3987 flags = BTRFS_EXTENT_FLAG_DATA;
3988
3989 if (path_change) {
b3b4aa74 3990 btrfs_release_path(path);
5d4f98a2
YZ		 3991
3992 path->search_commit_root = 1;
3993 path->skip_locking = 1;
3994 ret = btrfs_search_slot(NULL, rc->extent_root,
3995 &key, path, 0, 0);
3996 if (ret < 0) {
3997 err = ret;
3998 break;
3999 }
4000 BUG_ON(ret > 0);
4001 }
4002#else
4003 BUG();
4004#endif
4005 }
4006
4007 if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
4008 ret = add_tree_block(rc, &key, path, &blocks);
4009 } else if (rc->stage == UPDATE_DATA_PTRS &&
3fd0a558 4010 (flags & BTRFS_EXTENT_FLAG_DATA)) {
5d4f98a2
YZ		 4011 ret = add_data_references(rc, &key, path, &blocks);
4012 } else {
b3b4aa74 4013 btrfs_release_path(path);
5d4f98a2
YZ		 4014 ret = 0;
4015 }
4016 if (ret < 0) {
3fd0a558 4017 err = ret;
5d4f98a2
YZ		 4018 break;
4019 }
4020
4021 if (!RB_EMPTY_ROOT(&blocks)) {
4022 ret = relocate_tree_blocks(trans, rc, &blocks);
4023 if (ret < 0) {
1708cc57
WS		 4024 /*
4025 * if we fail to relocate tree blocks, force to update
4026 * backref cache when committing transaction.
4027 */
4028 rc->backref_cache.last_trans = trans->transid - 1;
4029
3fd0a558
YZ		 4030 if (ret != -EAGAIN) {
4031 err = ret;
4032 break;
4033 }
4034 rc->extents_found--;
4035 rc->search_start = key.objectid;
4036 }
4037 }
4038
0647bf56
WS		 4039 btrfs_end_transaction_throttle(trans, rc->extent_root);
4040 btrfs_btree_balance_dirty(rc->extent_root);
5d4f98a2 4041 trans = NULL;
5d4f98a2
YZ		 4042
4043 if (rc->stage == MOVE_DATA_EXTENTS &&
4044 (flags & BTRFS_EXTENT_FLAG_DATA)) {
4045 rc->found_file_extent = 1;
0257bb82 4046 ret = relocate_data_extent(rc->data_inode,
3fd0a558 4047 &key, &rc->cluster);
5d4f98a2
YZ		 4048 if (ret < 0) {
4049 err = ret;
4050 break;
4051 }
4052 }
4053 }
c87f08ca
CM		 4054 if (trans && progress && err == -ENOSPC) {
4055 ret = btrfs_force_chunk_alloc(trans, rc->extent_root,
4056 rc->block_group->flags);
4057 if (ret == 0) {
4058 err = 0;
4059 progress = 0;
4060 goto restart;
4061 }
4062 }
3fd0a558 4063
b3b4aa74 4064 btrfs_release_path(path);
3fd0a558
YZ		 4065 clear_extent_bits(&rc->processed_blocks, 0, (u64)-1, EXTENT_DIRTY,
4066 GFP_NOFS);
5d4f98a2
YZ		 4067
4068 if (trans) {
3fd0a558 4069 btrfs_end_transaction_throttle(trans, rc->extent_root);
b53d3f5d 4070 btrfs_btree_balance_dirty(rc->extent_root);
5d4f98a2
YZ		 4071 }
4072
0257bb82 4073 if (!err) {
3fd0a558
YZ		 4074 ret = relocate_file_extent_cluster(rc->data_inode,
4075 &rc->cluster);
0257bb82
YZ		 4076 if (ret < 0)
4077 err = ret;
4078 }
4079
3fd0a558
YZ		 4080 rc->create_reloc_tree = 0;
4081 set_reloc_control(rc);
0257bb82 4082
3fd0a558
YZ		 4083 backref_cache_cleanup(&rc->backref_cache);
4084 btrfs_block_rsv_release(rc->extent_root, rc->block_rsv, (u64)-1);
5d4f98a2 4085
3fd0a558 4086 err = prepare_to_merge(rc, err);
5d4f98a2
YZ		 4087
4088 merge_reloc_roots(rc);
4089
3fd0a558 4090 rc->merge_reloc_tree = 0;
5d4f98a2 4091 unset_reloc_control(rc);
3fd0a558 4092 btrfs_block_rsv_release(rc->extent_root, rc->block_rsv, (u64)-1);
5d4f98a2
YZ		 4093
4094 /* get rid of pinned extents */
7a7eaa40 4095 trans = btrfs_join_transaction(rc->extent_root);
3612b495
TI		 4096 if (IS_ERR(trans))
4097 err = PTR_ERR(trans);
4098 else
4099 btrfs_commit_transaction(trans, rc->extent_root);
3fd0a558
YZ		 4100out_free:
4101 btrfs_free_block_rsv(rc->extent_root, rc->block_rsv);
4102 btrfs_free_path(path);
5d4f98a2
YZ		 4103 return err;
4104}
4105
4106static int __insert_orphan_inode(struct btrfs_trans_handle *trans,
0257bb82 4107 struct btrfs_root *root, u64 objectid)
5d4f98a2
YZ		 4108{
4109 struct btrfs_path *path;
4110 struct btrfs_inode_item *item;
4111 struct extent_buffer *leaf;
4112 int ret;
4113
4114 path = btrfs_alloc_path();
4115 if (!path)
4116 return -ENOMEM;
4117
4118 ret = btrfs_insert_empty_inode(trans, root, path, objectid);
4119 if (ret)
4120 goto out;
4121
4122 leaf = path->nodes[0];
4123 item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_inode_item);
4124 memset_extent_buffer(leaf, 0, (unsigned long)item, sizeof(*item));
4125 btrfs_set_inode_generation(leaf, item, 1);
0257bb82 4126 btrfs_set_inode_size(leaf, item, 0);
5d4f98a2 4127 btrfs_set_inode_mode(leaf, item, S_IFREG | 0600);
3fd0a558
YZ		 4128 btrfs_set_inode_flags(leaf, item, BTRFS_INODE_NOCOMPRESS |
4129 BTRFS_INODE_PREALLOC);
5d4f98a2 4130 btrfs_mark_buffer_dirty(leaf);
5d4f98a2
YZ		 4131out:
4132 btrfs_free_path(path);
4133 return ret;
4134}
4135
4136/*
4137 * helper to create inode for data relocation.
4138 * the inode is in data relocation tree and its link count is 0
4139 */
3fd0a558
YZ		 4140static noinline_for_stack
4141struct inode *create_reloc_inode(struct btrfs_fs_info *fs_info,
4142 struct btrfs_block_group_cache *group)
5d4f98a2
YZ		 4143{
4144 struct inode *inode = NULL;
4145 struct btrfs_trans_handle *trans;
4146 struct btrfs_root *root;
4147 struct btrfs_key key;
5d4f98a2
YZ		 4148 u64 objectid = BTRFS_FIRST_FREE_OBJECTID;
4149 int err = 0;
4150
4151 root = read_fs_root(fs_info, BTRFS_DATA_RELOC_TREE_OBJECTID);
4152 if (IS_ERR(root))
4153 return ERR_CAST(root);
4154
a22285a6 4155 trans = btrfs_start_transaction(root, 6);
3fd0a558
YZ		 4156 if (IS_ERR(trans))
4157 return ERR_CAST(trans);
5d4f98a2 4158
581bb050 4159 err = btrfs_find_free_objectid(root, &objectid);
5d4f98a2
YZ		 4160 if (err)
4161 goto out;
4162
0257bb82 4163 err = __insert_orphan_inode(trans, root, objectid);
5d4f98a2
YZ		 4164 BUG_ON(err);
4165
4166 key.objectid = objectid;
4167 key.type = BTRFS_INODE_ITEM_KEY;
4168 key.offset = 0;
73f73415 4169 inode = btrfs_iget(root->fs_info->sb, &key, root, NULL);
5d4f98a2
YZ		 4170 BUG_ON(IS_ERR(inode) || is_bad_inode(inode));
4171 BTRFS_I(inode)->index_cnt = group->key.objectid;
4172
4173 err = btrfs_orphan_add(trans, inode);
4174out:
5d4f98a2 4175 btrfs_end_transaction(trans, root);
b53d3f5d 4176 btrfs_btree_balance_dirty(root);
5d4f98a2
YZ		 4177 if (err) {
4178 if (inode)
4179 iput(inode);
4180 inode = ERR_PTR(err);
4181 }
4182 return inode;
4183}
4184
a9995eec 4185static struct reloc_control *alloc_reloc_control(struct btrfs_fs_info *fs_info)
3fd0a558
YZ		 4186{
4187 struct reloc_control *rc;
4188
4189 rc = kzalloc(sizeof(*rc), GFP_NOFS);
4190 if (!rc)
4191 return NULL;
4192
4193 INIT_LIST_HEAD(&rc->reloc_roots);
4194 backref_cache_init(&rc->backref_cache);
4195 mapping_tree_init(&rc->reloc_root_tree);
a9995eec
JB		 4196 extent_io_tree_init(&rc->processed_blocks,
4197 fs_info->btree_inode->i_mapping);
3fd0a558
YZ		 4198 return rc;
4199}
4200
5d4f98a2
YZ		 4201/*
4202 * function to relocate all extents in a block group.
4203 */
4204int btrfs_relocate_block_group(struct btrfs_root *extent_root, u64 group_start)
4205{
4206 struct btrfs_fs_info *fs_info = extent_root->fs_info;
4207 struct reloc_control *rc;
0af3d00b
JB		 4208 struct inode *inode;
4209 struct btrfs_path *path;
5d4f98a2 4210 int ret;
f0486c68 4211 int rw = 0;
5d4f98a2
YZ		 4212 int err = 0;
4213
a9995eec 4214 rc = alloc_reloc_control(fs_info);
5d4f98a2
YZ		 4215 if (!rc)
4216 return -ENOMEM;
4217
f0486c68 4218 rc->extent_root = extent_root;
3fd0a558 4219
5d4f98a2
YZ		 4220 rc->block_group = btrfs_lookup_block_group(fs_info, group_start);
4221 BUG_ON(!rc->block_group);
4222
f0486c68
YZ		 4223 if (!rc->block_group->ro) {
4224 ret = btrfs_set_block_group_ro(extent_root, rc->block_group);
4225 if (ret) {
4226 err = ret;
4227 goto out;
4228 }
4229 rw = 1;
4230 }
4231
0af3d00b
JB		 4232 path = btrfs_alloc_path();
4233 if (!path) {
4234 err = -ENOMEM;
4235 goto out;
4236 }
4237
4238 inode = lookup_free_space_inode(fs_info->tree_root, rc->block_group,
4239 path);
4240 btrfs_free_path(path);
4241
4242 if (!IS_ERR(inode))
4243 ret = delete_block_group_cache(fs_info, inode, 0);
4244 else
4245 ret = PTR_ERR(inode);
4246
4247 if (ret && ret != -ENOENT) {
4248 err = ret;
4249 goto out;
4250 }
4251
5d4f98a2
YZ		 4252 rc->data_inode = create_reloc_inode(fs_info, rc->block_group);
4253 if (IS_ERR(rc->data_inode)) {
4254 err = PTR_ERR(rc->data_inode);
4255 rc->data_inode = NULL;
4256 goto out;
4257 }
4258
efe120a0 4259 btrfs_info(extent_root->fs_info, "relocating block group %llu flags %llu",
c1c9ff7c 4260 rc->block_group->key.objectid, rc->block_group->flags);
5d4f98a2 4261
6c255e67 4262 ret = btrfs_start_delalloc_roots(fs_info, 0, -1);
8ccf6f19
MX		 4263 if (ret < 0) {
4264 err = ret;
4265 goto out;
4266 }
b0244199 4267 btrfs_wait_ordered_roots(fs_info, -1);
5d4f98a2
YZ		 4268
4269 while (1) {
76dda93c 4270 mutex_lock(&fs_info->cleaner_mutex);
5d4f98a2 4271 ret = relocate_block_group(rc);
76dda93c 4272 mutex_unlock(&fs_info->cleaner_mutex);
5d4f98a2
YZ		 4273 if (ret < 0) {
4274 err = ret;
3fd0a558 4275 goto out;
5d4f98a2
YZ		 4276 }
4277
4278 if (rc->extents_found == 0)
4279 break;
4280
efe120a0 4281 btrfs_info(extent_root->fs_info, "found %llu extents",
c1c9ff7c 4282 rc->extents_found);
5d4f98a2
YZ		 4283
4284 if (rc->stage == MOVE_DATA_EXTENTS && rc->found_file_extent) {
0ef8b726
JB		 4285 ret = btrfs_wait_ordered_range(rc->data_inode, 0,
4286 (u64)-1);
4287 if (ret) {
4288 err = ret;
4289 goto out;
4290 }
5d4f98a2
YZ		 4291 invalidate_mapping_pages(rc->data_inode->i_mapping,
4292 0, -1);
4293 rc->stage = UPDATE_DATA_PTRS;
5d4f98a2
YZ		 4294 }
4295 }
4296
5d4f98a2
YZ		 4297 WARN_ON(rc->block_group->pinned > 0);
4298 WARN_ON(rc->block_group->reserved > 0);
4299 WARN_ON(btrfs_block_group_used(&rc->block_group->item) > 0);
4300out:
f0486c68
YZ		 4301 if (err && rw)
4302 btrfs_set_block_group_rw(extent_root, rc->block_group);
5d4f98a2 4303 iput(rc->data_inode);
5d4f98a2
YZ		 4304 btrfs_put_block_group(rc->block_group);
4305 kfree(rc);
4306 return err;
4307}
4308
76dda93c
YZ		 4309static noinline_for_stack int mark_garbage_root(struct btrfs_root *root)
4310{
4311 struct btrfs_trans_handle *trans;
79787eaa 4312 int ret, err;
76dda93c 4313
a22285a6 4314 trans = btrfs_start_transaction(root->fs_info->tree_root, 0);
79787eaa
JM		 4315 if (IS_ERR(trans))
4316 return PTR_ERR(trans);
76dda93c
YZ		 4317
4318 memset(&root->root_item.drop_progress, 0,
4319 sizeof(root->root_item.drop_progress));
4320 root->root_item.drop_level = 0;
4321 btrfs_set_root_refs(&root->root_item, 0);
4322 ret = btrfs_update_root(trans, root->fs_info->tree_root,
4323 &root->root_key, &root->root_item);
76dda93c 4324
79787eaa
JM		 4325 err = btrfs_end_transaction(trans, root->fs_info->tree_root);
4326 if (err)
4327 return err;
4328 return ret;
76dda93c
YZ		 4329}
4330
5d4f98a2
YZ		 4331/*
4332 * recover relocation interrupted by system crash.
4333 *
4334 * this function resumes merging reloc trees with corresponding fs trees.
4335 * this is important for keeping the sharing of tree blocks
4336 */
4337int btrfs_recover_relocation(struct btrfs_root *root)
4338{
4339 LIST_HEAD(reloc_roots);
4340 struct btrfs_key key;
4341 struct btrfs_root *fs_root;
4342 struct btrfs_root *reloc_root;
4343 struct btrfs_path *path;
4344 struct extent_buffer *leaf;
4345 struct reloc_control *rc = NULL;
4346 struct btrfs_trans_handle *trans;
4347 int ret;
4348 int err = 0;
4349
4350 path = btrfs_alloc_path();
4351 if (!path)
4352 return -ENOMEM;
026fd317 4353 path->reada = -1;
5d4f98a2
YZ		 4354
4355 key.objectid = BTRFS_TREE_RELOC_OBJECTID;
4356 key.type = BTRFS_ROOT_ITEM_KEY;
4357 key.offset = (u64)-1;
4358
4359 while (1) {
4360 ret = btrfs_search_slot(NULL, root->fs_info->tree_root, &key,
4361 path, 0, 0);
4362 if (ret < 0) {
4363 err = ret;
4364 goto out;
4365 }
4366 if (ret > 0) {
4367 if (path->slots[0] == 0)
4368 break;
4369 path->slots[0]--;
4370 }
4371 leaf = path->nodes[0];
4372 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
b3b4aa74 4373 btrfs_release_path(path);
5d4f98a2
YZ		 4374
4375 if (key.objectid != BTRFS_TREE_RELOC_OBJECTID ||
4376 key.type != BTRFS_ROOT_ITEM_KEY)
4377 break;
4378
cb517eab 4379 reloc_root = btrfs_read_fs_root(root, &key);
5d4f98a2
YZ		 4380 if (IS_ERR(reloc_root)) {
4381 err = PTR_ERR(reloc_root);
4382 goto out;
4383 }
4384
4385 list_add(&reloc_root->root_list, &reloc_roots);
4386
4387 if (btrfs_root_refs(&reloc_root->root_item) > 0) {
4388 fs_root = read_fs_root(root->fs_info,
4389 reloc_root->root_key.offset);
4390 if (IS_ERR(fs_root)) {
76dda93c
YZ		 4391 ret = PTR_ERR(fs_root);
4392 if (ret != -ENOENT) {
4393 err = ret;
4394 goto out;
4395 }
79787eaa
JM		 4396 ret = mark_garbage_root(reloc_root);
4397 if (ret < 0) {
4398 err = ret;
4399 goto out;
4400 }
5d4f98a2
YZ		 4401 }
4402 }
4403
4404 if (key.offset == 0)
4405 break;
4406
4407 key.offset--;
4408 }
b3b4aa74 4409 btrfs_release_path(path);
5d4f98a2
YZ		 4410
4411 if (list_empty(&reloc_roots))
4412 goto out;
4413
a9995eec 4414 rc = alloc_reloc_control(root->fs_info);
5d4f98a2
YZ		 4415 if (!rc) {
4416 err = -ENOMEM;
4417 goto out;
4418 }
4419
5d4f98a2
YZ		 4420 rc->extent_root = root->fs_info->extent_root;
4421
4422 set_reloc_control(rc);
4423
7a7eaa40 4424 trans = btrfs_join_transaction(rc->extent_root);
3612b495
TI		 4425 if (IS_ERR(trans)) {
4426 unset_reloc_control(rc);
4427 err = PTR_ERR(trans);
4428 goto out_free;
4429 }
3fd0a558
YZ		 4430
4431 rc->merge_reloc_tree = 1;
4432
5d4f98a2
YZ		 4433 while (!list_empty(&reloc_roots)) {
4434 reloc_root = list_entry(reloc_roots.next,
4435 struct btrfs_root, root_list);
4436 list_del(&reloc_root->root_list);
4437
4438 if (btrfs_root_refs(&reloc_root->root_item) == 0) {
4439 list_add_tail(&reloc_root->root_list,
4440 &rc->reloc_roots);
4441 continue;
4442 }
4443
4444 fs_root = read_fs_root(root->fs_info,
4445 reloc_root->root_key.offset);
79787eaa
JM		 4446 if (IS_ERR(fs_root)) {
4447 err = PTR_ERR(fs_root);
4448 goto out_free;
4449 }
5d4f98a2 4450
ffd7b339 4451 err = __add_reloc_root(reloc_root);
79787eaa 4452 BUG_ON(err < 0); /* -ENOMEM or logic error */
5d4f98a2
YZ		 4453 fs_root->reloc_root = reloc_root;
4454 }
4455
79787eaa
JM		 4456 err = btrfs_commit_transaction(trans, rc->extent_root);
4457 if (err)
4458 goto out_free;
5d4f98a2
YZ		 4459
4460 merge_reloc_roots(rc);
4461
4462 unset_reloc_control(rc);
4463
7a7eaa40 4464 trans = btrfs_join_transaction(rc->extent_root);
3612b495
TI		 4465 if (IS_ERR(trans))
4466 err = PTR_ERR(trans);
4467 else
79787eaa 4468 err = btrfs_commit_transaction(trans, rc->extent_root);
3612b495 4469out_free:
3fd0a558 4470 kfree(rc);
3612b495 4471out:
aca1bba6
LB		 4472 if (!list_empty(&reloc_roots))
4473 free_reloc_roots(&reloc_roots);
4474
5d4f98a2
YZ		 4475 btrfs_free_path(path);
4476
4477 if (err == 0) {
4478 /* cleanup orphan inode in data relocation tree */
4479 fs_root = read_fs_root(root->fs_info,
4480 BTRFS_DATA_RELOC_TREE_OBJECTID);
4481 if (IS_ERR(fs_root))
4482 err = PTR_ERR(fs_root);
d7ce5843 4483 else
66b4ffd1 4484 err = btrfs_orphan_cleanup(fs_root);
5d4f98a2
YZ		 4485 }
4486 return err;
4487}
4488
4489/*
4490 * helper to add ordered checksum for data relocation.
4491 *
4492 * cloning checksum properly handles the nodatasum extents.
4493 * it also saves CPU time to re-calculate the checksum.
4494 */
4495int btrfs_reloc_clone_csums(struct inode *inode, u64 file_pos, u64 len)
4496{
4497 struct btrfs_ordered_sum *sums;
5d4f98a2
YZ		 4498 struct btrfs_ordered_extent *ordered;
4499 struct btrfs_root *root = BTRFS_I(inode)->root;
5d4f98a2
YZ		 4500 int ret;
4501 u64 disk_bytenr;
4577b014 4502 u64 new_bytenr;
5d4f98a2
YZ		 4503 LIST_HEAD(list);
4504
4505 ordered = btrfs_lookup_ordered_extent(inode, file_pos);
4506 BUG_ON(ordered->file_offset != file_pos || ordered->len != len);
4507
4508 disk_bytenr = file_pos + BTRFS_I(inode)->index_cnt;
4509 ret = btrfs_lookup_csums_range(root->fs_info->csum_root, disk_bytenr,
a2de733c 4510 disk_bytenr + len - 1, &list, 0);
79787eaa
JM		 4511 if (ret)
4512 goto out;
5d4f98a2
YZ		 4513
4514 while (!list_empty(&list)) {
4515 sums = list_entry(list.next, struct btrfs_ordered_sum, list);
4516 list_del_init(&sums->list);
4517
4577b014
JB		 4518 /*
4519 * We need to offset the new_bytenr based on where the csum is.
4520 * We need to do this because we will read in entire prealloc
4521 * extents but we may have written to say the middle of the
4522 * prealloc extent, so we need to make sure the csum goes with
4523 * the right disk offset.
4524 *
4525 * We can do this because the data reloc inode refers strictly
4526 * to the on disk bytes, so we don't have to worry about
4527 * disk_len vs real len like with real inodes since it's all
4528 * disk length.
4529 */
4530 new_bytenr = ordered->start + (sums->bytenr - disk_bytenr);
4531 sums->bytenr = new_bytenr;
5d4f98a2
YZ		 4532
4533 btrfs_add_ordered_sum(inode, ordered, sums);
4534 }
79787eaa 4535out:
5d4f98a2 4536 btrfs_put_ordered_extent(ordered);
411fc6bc 4537 return ret;
5d4f98a2 4538}
3fd0a558 4539
83d4cfd4
JB		 4540int btrfs_reloc_cow_block(struct btrfs_trans_handle *trans,
4541 struct btrfs_root *root, struct extent_buffer *buf,
4542 struct extent_buffer *cow)
3fd0a558
YZ		 4543{
4544 struct reloc_control *rc;
4545 struct backref_node *node;
4546 int first_cow = 0;
4547 int level;
83d4cfd4 4548 int ret = 0;
3fd0a558
YZ		 4549
4550 rc = root->fs_info->reloc_ctl;
4551 if (!rc)
83d4cfd4 4552 return 0;
3fd0a558
YZ		 4553
4554 BUG_ON(rc->stage == UPDATE_DATA_PTRS &&
4555 root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID);
4556
c974c464
WS		 4557 if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID) {
4558 if (buf == root->node)
4559 __update_reloc_root(root, cow->start);
4560 }
4561
3fd0a558
YZ		 4562 level = btrfs_header_level(buf);
4563 if (btrfs_header_generation(buf) <=
4564 btrfs_root_last_snapshot(&root->root_item))
4565 first_cow = 1;
4566
4567 if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID &&
4568 rc->create_reloc_tree) {
4569 WARN_ON(!first_cow && level == 0);
4570
4571 node = rc->backref_cache.path[level];
4572 BUG_ON(node->bytenr != buf->start &&
4573 node->new_bytenr != buf->start);
4574
4575 drop_node_buffer(node);
4576 extent_buffer_get(cow);
4577 node->eb = cow;
4578 node->new_bytenr = cow->start;
4579
4580 if (!node->pending) {
4581 list_move_tail(&node->list,
4582 &rc->backref_cache.pending[level]);
4583 node->pending = 1;
4584 }
4585
4586 if (first_cow)
4587 __mark_block_processed(rc, node);
4588
4589 if (first_cow && level > 0)
4590 rc->nodes_relocated += buf->len;
4591 }
4592
83d4cfd4 4593 if (level == 0 && first_cow && rc->stage == UPDATE_DATA_PTRS)
3fd0a558 4594 ret = replace_file_extents(trans, rc, root, cow);
83d4cfd4 4595 return ret;
3fd0a558
YZ		 4596}
4597
4598/*
4599 * called before creating snapshot. it calculates metadata reservation
4600 * requried for relocating tree blocks in the snapshot
4601 */
4602void btrfs_reloc_pre_snapshot(struct btrfs_trans_handle *trans,
4603 struct btrfs_pending_snapshot *pending,
4604 u64 *bytes_to_reserve)
4605{
4606 struct btrfs_root *root;
4607 struct reloc_control *rc;
4608
4609 root = pending->root;
4610 if (!root->reloc_root)
4611 return;
4612
4613 rc = root->fs_info->reloc_ctl;
4614 if (!rc->merge_reloc_tree)
4615 return;
4616
4617 root = root->reloc_root;
4618 BUG_ON(btrfs_root_refs(&root->root_item) == 0);
4619 /*
4620 * relocation is in the stage of merging trees. the space
4621 * used by merging a reloc tree is twice the size of
4622 * relocated tree nodes in the worst case. half for cowing
4623 * the reloc tree, half for cowing the fs tree. the space
4624 * used by cowing the reloc tree will be freed after the
4625 * tree is dropped. if we create snapshot, cowing the fs
4626 * tree may use more space than it frees. so we need
4627 * reserve extra space.
4628 */
4629 *bytes_to_reserve += rc->nodes_relocated;
4630}
4631
4632/*
4633 * called after snapshot is created. migrate block reservation
4634 * and create reloc root for the newly created snapshot
4635 */
49b25e05 4636int btrfs_reloc_post_snapshot(struct btrfs_trans_handle *trans,
3fd0a558
YZ		 4637 struct btrfs_pending_snapshot *pending)
4638{
4639 struct btrfs_root *root = pending->root;
4640 struct btrfs_root *reloc_root;
4641 struct btrfs_root *new_root;
4642 struct reloc_control *rc;
4643 int ret;
4644
4645 if (!root->reloc_root)
49b25e05 4646 return 0;
3fd0a558
YZ		 4647
4648 rc = root->fs_info->reloc_ctl;
4649 rc->merging_rsv_size += rc->nodes_relocated;
4650
4651 if (rc->merge_reloc_tree) {
4652 ret = btrfs_block_rsv_migrate(&pending->block_rsv,
4653 rc->block_rsv,
4654 rc->nodes_relocated);
49b25e05
JM		 4655 if (ret)
4656 return ret;
3fd0a558
YZ		 4657 }
4658
4659 new_root = pending->snap;
4660 reloc_root = create_reloc_root(trans, root->reloc_root,
4661 new_root->root_key.objectid);
49b25e05
JM		 4662 if (IS_ERR(reloc_root))
4663 return PTR_ERR(reloc_root);
3fd0a558 4664
ffd7b339
JM		 4665 ret = __add_reloc_root(reloc_root);
4666 BUG_ON(ret < 0);
3fd0a558
YZ		 4667 new_root->reloc_root = reloc_root;
4668
49b25e05 4669 if (rc->create_reloc_tree)
3fd0a558 4670 ret = clone_backref_node(trans, rc, root, reloc_root);
49b25e05 4671 return ret;
3fd0a558 4672}
Linux kernel - Deliverables
This page took 0.466692 seconds and 5 git commands to generate.