Loading...
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * Copyright (C) 2009 Oracle. All rights reserved.
4 */
5
6#include <linux/sched.h>
7#include <linux/pagemap.h>
8#include <linux/writeback.h>
9#include <linux/blkdev.h>
10#include <linux/rbtree.h>
11#include <linux/slab.h>
12#include <linux/error-injection.h>
13#include "ctree.h"
14#include "disk-io.h"
15#include "transaction.h"
16#include "volumes.h"
17#include "locking.h"
18#include "btrfs_inode.h"
19#include "async-thread.h"
20#include "free-space-cache.h"
21#include "qgroup.h"
22#include "print-tree.h"
23#include "delalloc-space.h"
24#include "block-group.h"
25#include "backref.h"
26#include "misc.h"
27#include "subpage.h"
28#include "zoned.h"
29#include "inode-item.h"
30#include "space-info.h"
31#include "fs.h"
32#include "accessors.h"
33#include "extent-tree.h"
34#include "root-tree.h"
35#include "file-item.h"
36#include "relocation.h"
37#include "super.h"
38#include "tree-checker.h"
39
40/*
41 * Relocation overview
42 *
43 * [What does relocation do]
44 *
45 * The objective of relocation is to relocate all extents of the target block
46 * group to other block groups.
47 * This is utilized by resize (shrink only), profile converting, compacting
48 * space, or balance routine to spread chunks over devices.
49 *
50 * Before | After
51 * ------------------------------------------------------------------
52 * BG A: 10 data extents | BG A: deleted
53 * BG B: 2 data extents | BG B: 10 data extents (2 old + 8 relocated)
54 * BG C: 1 extents | BG C: 3 data extents (1 old + 2 relocated)
55 *
56 * [How does relocation work]
57 *
58 * 1. Mark the target block group read-only
59 * New extents won't be allocated from the target block group.
60 *
61 * 2.1 Record each extent in the target block group
62 * To build a proper map of extents to be relocated.
63 *
64 * 2.2 Build data reloc tree and reloc trees
65 * Data reloc tree will contain an inode, recording all newly relocated
66 * data extents.
67 * There will be only one data reloc tree for one data block group.
68 *
69 * Reloc tree will be a special snapshot of its source tree, containing
70 * relocated tree blocks.
71 * Each tree referring to a tree block in target block group will get its
72 * reloc tree built.
73 *
74 * 2.3 Swap source tree with its corresponding reloc tree
75 * Each involved tree only refers to new extents after swap.
76 *
77 * 3. Cleanup reloc trees and data reloc tree.
78 * As old extents in the target block group are still referenced by reloc
79 * trees, we need to clean them up before really freeing the target block
80 * group.
81 *
82 * The main complexity is in steps 2.2 and 2.3.
83 *
84 * The entry point of relocation is relocate_block_group() function.
85 */
86
87#define RELOCATION_RESERVED_NODES 256
88/*
89 * map address of tree root to tree
90 */
91struct mapping_node {
92 struct {
93 struct rb_node rb_node;
94 u64 bytenr;
95 }; /* Use rb_simle_node for search/insert */
96 void *data;
97};
98
99struct mapping_tree {
100 struct rb_root rb_root;
101 spinlock_t lock;
102};
103
104/*
105 * present a tree block to process
106 */
107struct tree_block {
108 struct {
109 struct rb_node rb_node;
110 u64 bytenr;
111 }; /* Use rb_simple_node for search/insert */
112 u64 owner;
113 struct btrfs_key key;
114 unsigned int level:8;
115 unsigned int key_ready:1;
116};
117
118#define MAX_EXTENTS 128
119
120struct file_extent_cluster {
121 u64 start;
122 u64 end;
123 u64 boundary[MAX_EXTENTS];
124 unsigned int nr;
125};
126
127struct reloc_control {
128 /* block group to relocate */
129 struct btrfs_block_group *block_group;
130 /* extent tree */
131 struct btrfs_root *extent_root;
132 /* inode for moving data */
133 struct inode *data_inode;
134
135 struct btrfs_block_rsv *block_rsv;
136
137 struct btrfs_backref_cache backref_cache;
138
139 struct file_extent_cluster cluster;
140 /* tree blocks have been processed */
141 struct extent_io_tree processed_blocks;
142 /* map start of tree root to corresponding reloc tree */
143 struct mapping_tree reloc_root_tree;
144 /* list of reloc trees */
145 struct list_head reloc_roots;
146 /* list of subvolume trees that get relocated */
147 struct list_head dirty_subvol_roots;
148 /* size of metadata reservation for merging reloc trees */
149 u64 merging_rsv_size;
150 /* size of relocated tree nodes */
151 u64 nodes_relocated;
152 /* reserved size for block group relocation*/
153 u64 reserved_bytes;
154
155 u64 search_start;
156 u64 extents_found;
157
158 unsigned int stage:8;
159 unsigned int create_reloc_tree:1;
160 unsigned int merge_reloc_tree:1;
161 unsigned int found_file_extent:1;
162};
163
164/* stages of data relocation */
165#define MOVE_DATA_EXTENTS 0
166#define UPDATE_DATA_PTRS 1
167
168static void mark_block_processed(struct reloc_control *rc,
169 struct btrfs_backref_node *node)
170{
171 u32 blocksize;
172
173 if (node->level == 0 ||
174 in_range(node->bytenr, rc->block_group->start,
175 rc->block_group->length)) {
176 blocksize = rc->extent_root->fs_info->nodesize;
177 set_extent_bits(&rc->processed_blocks, node->bytenr,
178 node->bytenr + blocksize - 1, EXTENT_DIRTY);
179 }
180 node->processed = 1;
181}
182
183
184static void mapping_tree_init(struct mapping_tree *tree)
185{
186 tree->rb_root = RB_ROOT;
187 spin_lock_init(&tree->lock);
188}
189
190/*
191 * walk up backref nodes until reach node presents tree root
192 */
193static struct btrfs_backref_node *walk_up_backref(
194 struct btrfs_backref_node *node,
195 struct btrfs_backref_edge *edges[], int *index)
196{
197 struct btrfs_backref_edge *edge;
198 int idx = *index;
199
200 while (!list_empty(&node->upper)) {
201 edge = list_entry(node->upper.next,
202 struct btrfs_backref_edge, list[LOWER]);
203 edges[idx++] = edge;
204 node = edge->node[UPPER];
205 }
206 BUG_ON(node->detached);
207 *index = idx;
208 return node;
209}
210
211/*
212 * walk down backref nodes to find start of next reference path
213 */
214static struct btrfs_backref_node *walk_down_backref(
215 struct btrfs_backref_edge *edges[], int *index)
216{
217 struct btrfs_backref_edge *edge;
218 struct btrfs_backref_node *lower;
219 int idx = *index;
220
221 while (idx > 0) {
222 edge = edges[idx - 1];
223 lower = edge->node[LOWER];
224 if (list_is_last(&edge->list[LOWER], &lower->upper)) {
225 idx--;
226 continue;
227 }
228 edge = list_entry(edge->list[LOWER].next,
229 struct btrfs_backref_edge, list[LOWER]);
230 edges[idx - 1] = edge;
231 *index = idx;
232 return edge->node[UPPER];
233 }
234 *index = 0;
235 return NULL;
236}
237
238static void update_backref_node(struct btrfs_backref_cache *cache,
239 struct btrfs_backref_node *node, u64 bytenr)
240{
241 struct rb_node *rb_node;
242 rb_erase(&node->rb_node, &cache->rb_root);
243 node->bytenr = bytenr;
244 rb_node = rb_simple_insert(&cache->rb_root, node->bytenr, &node->rb_node);
245 if (rb_node)
246 btrfs_backref_panic(cache->fs_info, bytenr, -EEXIST);
247}
248
249/*
250 * update backref cache after a transaction commit
251 */
252static int update_backref_cache(struct btrfs_trans_handle *trans,
253 struct btrfs_backref_cache *cache)
254{
255 struct btrfs_backref_node *node;
256 int level = 0;
257
258 if (cache->last_trans == 0) {
259 cache->last_trans = trans->transid;
260 return 0;
261 }
262
263 if (cache->last_trans == trans->transid)
264 return 0;
265
266 /*
267 * detached nodes are used to avoid unnecessary backref
268 * lookup. transaction commit changes the extent tree.
269 * so the detached nodes are no longer useful.
270 */
271 while (!list_empty(&cache->detached)) {
272 node = list_entry(cache->detached.next,
273 struct btrfs_backref_node, list);
274 btrfs_backref_cleanup_node(cache, node);
275 }
276
277 while (!list_empty(&cache->changed)) {
278 node = list_entry(cache->changed.next,
279 struct btrfs_backref_node, list);
280 list_del_init(&node->list);
281 BUG_ON(node->pending);
282 update_backref_node(cache, node, node->new_bytenr);
283 }
284
285 /*
286 * some nodes can be left in the pending list if there were
287 * errors during processing the pending nodes.
288 */
289 for (level = 0; level < BTRFS_MAX_LEVEL; level++) {
290 list_for_each_entry(node, &cache->pending[level], list) {
291 BUG_ON(!node->pending);
292 if (node->bytenr == node->new_bytenr)
293 continue;
294 update_backref_node(cache, node, node->new_bytenr);
295 }
296 }
297
298 cache->last_trans = 0;
299 return 1;
300}
301
302static bool reloc_root_is_dead(struct btrfs_root *root)
303{
304 /*
305 * Pair with set_bit/clear_bit in clean_dirty_subvols and
306 * btrfs_update_reloc_root. We need to see the updated bit before
307 * trying to access reloc_root
308 */
309 smp_rmb();
310 if (test_bit(BTRFS_ROOT_DEAD_RELOC_TREE, &root->state))
311 return true;
312 return false;
313}
314
315/*
316 * Check if this subvolume tree has valid reloc tree.
317 *
318 * Reloc tree after swap is considered dead, thus not considered as valid.
319 * This is enough for most callers, as they don't distinguish dead reloc root
320 * from no reloc root. But btrfs_should_ignore_reloc_root() below is a
321 * special case.
322 */
323static bool have_reloc_root(struct btrfs_root *root)
324{
325 if (reloc_root_is_dead(root))
326 return false;
327 if (!root->reloc_root)
328 return false;
329 return true;
330}
331
332int btrfs_should_ignore_reloc_root(struct btrfs_root *root)
333{
334 struct btrfs_root *reloc_root;
335
336 if (!test_bit(BTRFS_ROOT_SHAREABLE, &root->state))
337 return 0;
338
339 /* This root has been merged with its reloc tree, we can ignore it */
340 if (reloc_root_is_dead(root))
341 return 1;
342
343 reloc_root = root->reloc_root;
344 if (!reloc_root)
345 return 0;
346
347 if (btrfs_header_generation(reloc_root->commit_root) ==
348 root->fs_info->running_transaction->transid)
349 return 0;
350 /*
351 * if there is reloc tree and it was created in previous
352 * transaction backref lookup can find the reloc tree,
353 * so backref node for the fs tree root is useless for
354 * relocation.
355 */
356 return 1;
357}
358
359/*
360 * find reloc tree by address of tree root
361 */
362struct btrfs_root *find_reloc_root(struct btrfs_fs_info *fs_info, u64 bytenr)
363{
364 struct reloc_control *rc = fs_info->reloc_ctl;
365 struct rb_node *rb_node;
366 struct mapping_node *node;
367 struct btrfs_root *root = NULL;
368
369 ASSERT(rc);
370 spin_lock(&rc->reloc_root_tree.lock);
371 rb_node = rb_simple_search(&rc->reloc_root_tree.rb_root, bytenr);
372 if (rb_node) {
373 node = rb_entry(rb_node, struct mapping_node, rb_node);
374 root = node->data;
375 }
376 spin_unlock(&rc->reloc_root_tree.lock);
377 return btrfs_grab_root(root);
378}
379
380/*
381 * For useless nodes, do two major clean ups:
382 *
383 * - Cleanup the children edges and nodes
384 * If child node is also orphan (no parent) during cleanup, then the child
385 * node will also be cleaned up.
386 *
387 * - Freeing up leaves (level 0), keeps nodes detached
388 * For nodes, the node is still cached as "detached"
389 *
390 * Return false if @node is not in the @useless_nodes list.
391 * Return true if @node is in the @useless_nodes list.
392 */
393static bool handle_useless_nodes(struct reloc_control *rc,
394 struct btrfs_backref_node *node)
395{
396 struct btrfs_backref_cache *cache = &rc->backref_cache;
397 struct list_head *useless_node = &cache->useless_node;
398 bool ret = false;
399
400 while (!list_empty(useless_node)) {
401 struct btrfs_backref_node *cur;
402
403 cur = list_first_entry(useless_node, struct btrfs_backref_node,
404 list);
405 list_del_init(&cur->list);
406
407 /* Only tree root nodes can be added to @useless_nodes */
408 ASSERT(list_empty(&cur->upper));
409
410 if (cur == node)
411 ret = true;
412
413 /* The node is the lowest node */
414 if (cur->lowest) {
415 list_del_init(&cur->lower);
416 cur->lowest = 0;
417 }
418
419 /* Cleanup the lower edges */
420 while (!list_empty(&cur->lower)) {
421 struct btrfs_backref_edge *edge;
422 struct btrfs_backref_node *lower;
423
424 edge = list_entry(cur->lower.next,
425 struct btrfs_backref_edge, list[UPPER]);
426 list_del(&edge->list[UPPER]);
427 list_del(&edge->list[LOWER]);
428 lower = edge->node[LOWER];
429 btrfs_backref_free_edge(cache, edge);
430
431 /* Child node is also orphan, queue for cleanup */
432 if (list_empty(&lower->upper))
433 list_add(&lower->list, useless_node);
434 }
435 /* Mark this block processed for relocation */
436 mark_block_processed(rc, cur);
437
438 /*
439 * Backref nodes for tree leaves are deleted from the cache.
440 * Backref nodes for upper level tree blocks are left in the
441 * cache to avoid unnecessary backref lookup.
442 */
443 if (cur->level > 0) {
444 list_add(&cur->list, &cache->detached);
445 cur->detached = 1;
446 } else {
447 rb_erase(&cur->rb_node, &cache->rb_root);
448 btrfs_backref_free_node(cache, cur);
449 }
450 }
451 return ret;
452}
453
454/*
455 * Build backref tree for a given tree block. Root of the backref tree
456 * corresponds the tree block, leaves of the backref tree correspond roots of
457 * b-trees that reference the tree block.
458 *
459 * The basic idea of this function is check backrefs of a given block to find
460 * upper level blocks that reference the block, and then check backrefs of
461 * these upper level blocks recursively. The recursion stops when tree root is
462 * reached or backrefs for the block is cached.
463 *
464 * NOTE: if we find that backrefs for a block are cached, we know backrefs for
465 * all upper level blocks that directly/indirectly reference the block are also
466 * cached.
467 */
468static noinline_for_stack struct btrfs_backref_node *build_backref_tree(
469 struct reloc_control *rc, struct btrfs_key *node_key,
470 int level, u64 bytenr)
471{
472 struct btrfs_backref_iter *iter;
473 struct btrfs_backref_cache *cache = &rc->backref_cache;
474 /* For searching parent of TREE_BLOCK_REF */
475 struct btrfs_path *path;
476 struct btrfs_backref_node *cur;
477 struct btrfs_backref_node *node = NULL;
478 struct btrfs_backref_edge *edge;
479 int ret;
480 int err = 0;
481
482 iter = btrfs_backref_iter_alloc(rc->extent_root->fs_info);
483 if (!iter)
484 return ERR_PTR(-ENOMEM);
485 path = btrfs_alloc_path();
486 if (!path) {
487 err = -ENOMEM;
488 goto out;
489 }
490
491 node = btrfs_backref_alloc_node(cache, bytenr, level);
492 if (!node) {
493 err = -ENOMEM;
494 goto out;
495 }
496
497 node->lowest = 1;
498 cur = node;
499
500 /* Breadth-first search to build backref cache */
501 do {
502 ret = btrfs_backref_add_tree_node(cache, path, iter, node_key,
503 cur);
504 if (ret < 0) {
505 err = ret;
506 goto out;
507 }
508 edge = list_first_entry_or_null(&cache->pending_edge,
509 struct btrfs_backref_edge, list[UPPER]);
510 /*
511 * The pending list isn't empty, take the first block to
512 * process
513 */
514 if (edge) {
515 list_del_init(&edge->list[UPPER]);
516 cur = edge->node[UPPER];
517 }
518 } while (edge);
519
520 /* Finish the upper linkage of newly added edges/nodes */
521 ret = btrfs_backref_finish_upper_links(cache, node);
522 if (ret < 0) {
523 err = ret;
524 goto out;
525 }
526
527 if (handle_useless_nodes(rc, node))
528 node = NULL;
529out:
530 btrfs_backref_iter_free(iter);
531 btrfs_free_path(path);
532 if (err) {
533 btrfs_backref_error_cleanup(cache, node);
534 return ERR_PTR(err);
535 }
536 ASSERT(!node || !node->detached);
537 ASSERT(list_empty(&cache->useless_node) &&
538 list_empty(&cache->pending_edge));
539 return node;
540}
541
542/*
543 * helper to add backref node for the newly created snapshot.
544 * the backref node is created by cloning backref node that
545 * corresponds to root of source tree
546 */
547static int clone_backref_node(struct btrfs_trans_handle *trans,
548 struct reloc_control *rc,
549 struct btrfs_root *src,
550 struct btrfs_root *dest)
551{
552 struct btrfs_root *reloc_root = src->reloc_root;
553 struct btrfs_backref_cache *cache = &rc->backref_cache;
554 struct btrfs_backref_node *node = NULL;
555 struct btrfs_backref_node *new_node;
556 struct btrfs_backref_edge *edge;
557 struct btrfs_backref_edge *new_edge;
558 struct rb_node *rb_node;
559
560 if (cache->last_trans > 0)
561 update_backref_cache(trans, cache);
562
563 rb_node = rb_simple_search(&cache->rb_root, src->commit_root->start);
564 if (rb_node) {
565 node = rb_entry(rb_node, struct btrfs_backref_node, rb_node);
566 if (node->detached)
567 node = NULL;
568 else
569 BUG_ON(node->new_bytenr != reloc_root->node->start);
570 }
571
572 if (!node) {
573 rb_node = rb_simple_search(&cache->rb_root,
574 reloc_root->commit_root->start);
575 if (rb_node) {
576 node = rb_entry(rb_node, struct btrfs_backref_node,
577 rb_node);
578 BUG_ON(node->detached);
579 }
580 }
581
582 if (!node)
583 return 0;
584
585 new_node = btrfs_backref_alloc_node(cache, dest->node->start,
586 node->level);
587 if (!new_node)
588 return -ENOMEM;
589
590 new_node->lowest = node->lowest;
591 new_node->checked = 1;
592 new_node->root = btrfs_grab_root(dest);
593 ASSERT(new_node->root);
594
595 if (!node->lowest) {
596 list_for_each_entry(edge, &node->lower, list[UPPER]) {
597 new_edge = btrfs_backref_alloc_edge(cache);
598 if (!new_edge)
599 goto fail;
600
601 btrfs_backref_link_edge(new_edge, edge->node[LOWER],
602 new_node, LINK_UPPER);
603 }
604 } else {
605 list_add_tail(&new_node->lower, &cache->leaves);
606 }
607
608 rb_node = rb_simple_insert(&cache->rb_root, new_node->bytenr,
609 &new_node->rb_node);
610 if (rb_node)
611 btrfs_backref_panic(trans->fs_info, new_node->bytenr, -EEXIST);
612
613 if (!new_node->lowest) {
614 list_for_each_entry(new_edge, &new_node->lower, list[UPPER]) {
615 list_add_tail(&new_edge->list[LOWER],
616 &new_edge->node[LOWER]->upper);
617 }
618 }
619 return 0;
620fail:
621 while (!list_empty(&new_node->lower)) {
622 new_edge = list_entry(new_node->lower.next,
623 struct btrfs_backref_edge, list[UPPER]);
624 list_del(&new_edge->list[UPPER]);
625 btrfs_backref_free_edge(cache, new_edge);
626 }
627 btrfs_backref_free_node(cache, new_node);
628 return -ENOMEM;
629}
630
631/*
632 * helper to add 'address of tree root -> reloc tree' mapping
633 */
634static int __must_check __add_reloc_root(struct btrfs_root *root)
635{
636 struct btrfs_fs_info *fs_info = root->fs_info;
637 struct rb_node *rb_node;
638 struct mapping_node *node;
639 struct reloc_control *rc = fs_info->reloc_ctl;
640
641 node = kmalloc(sizeof(*node), GFP_NOFS);
642 if (!node)
643 return -ENOMEM;
644
645 node->bytenr = root->commit_root->start;
646 node->data = root;
647
648 spin_lock(&rc->reloc_root_tree.lock);
649 rb_node = rb_simple_insert(&rc->reloc_root_tree.rb_root,
650 node->bytenr, &node->rb_node);
651 spin_unlock(&rc->reloc_root_tree.lock);
652 if (rb_node) {
653 btrfs_err(fs_info,
654 "Duplicate root found for start=%llu while inserting into relocation tree",
655 node->bytenr);
656 return -EEXIST;
657 }
658
659 list_add_tail(&root->root_list, &rc->reloc_roots);
660 return 0;
661}
662
663/*
664 * helper to delete the 'address of tree root -> reloc tree'
665 * mapping
666 */
667static void __del_reloc_root(struct btrfs_root *root)
668{
669 struct btrfs_fs_info *fs_info = root->fs_info;
670 struct rb_node *rb_node;
671 struct mapping_node *node = NULL;
672 struct reloc_control *rc = fs_info->reloc_ctl;
673 bool put_ref = false;
674
675 if (rc && root->node) {
676 spin_lock(&rc->reloc_root_tree.lock);
677 rb_node = rb_simple_search(&rc->reloc_root_tree.rb_root,
678 root->commit_root->start);
679 if (rb_node) {
680 node = rb_entry(rb_node, struct mapping_node, rb_node);
681 rb_erase(&node->rb_node, &rc->reloc_root_tree.rb_root);
682 RB_CLEAR_NODE(&node->rb_node);
683 }
684 spin_unlock(&rc->reloc_root_tree.lock);
685 ASSERT(!node || (struct btrfs_root *)node->data == root);
686 }
687
688 /*
689 * We only put the reloc root here if it's on the list. There's a lot
690 * of places where the pattern is to splice the rc->reloc_roots, process
691 * the reloc roots, and then add the reloc root back onto
692 * rc->reloc_roots. If we call __del_reloc_root while it's off of the
693 * list we don't want the reference being dropped, because the guy
694 * messing with the list is in charge of the reference.
695 */
696 spin_lock(&fs_info->trans_lock);
697 if (!list_empty(&root->root_list)) {
698 put_ref = true;
699 list_del_init(&root->root_list);
700 }
701 spin_unlock(&fs_info->trans_lock);
702 if (put_ref)
703 btrfs_put_root(root);
704 kfree(node);
705}
706
707/*
708 * helper to update the 'address of tree root -> reloc tree'
709 * mapping
710 */
711static int __update_reloc_root(struct btrfs_root *root)
712{
713 struct btrfs_fs_info *fs_info = root->fs_info;
714 struct rb_node *rb_node;
715 struct mapping_node *node = NULL;
716 struct reloc_control *rc = fs_info->reloc_ctl;
717
718 spin_lock(&rc->reloc_root_tree.lock);
719 rb_node = rb_simple_search(&rc->reloc_root_tree.rb_root,
720 root->commit_root->start);
721 if (rb_node) {
722 node = rb_entry(rb_node, struct mapping_node, rb_node);
723 rb_erase(&node->rb_node, &rc->reloc_root_tree.rb_root);
724 }
725 spin_unlock(&rc->reloc_root_tree.lock);
726
727 if (!node)
728 return 0;
729 BUG_ON((struct btrfs_root *)node->data != root);
730
731 spin_lock(&rc->reloc_root_tree.lock);
732 node->bytenr = root->node->start;
733 rb_node = rb_simple_insert(&rc->reloc_root_tree.rb_root,
734 node->bytenr, &node->rb_node);
735 spin_unlock(&rc->reloc_root_tree.lock);
736 if (rb_node)
737 btrfs_backref_panic(fs_info, node->bytenr, -EEXIST);
738 return 0;
739}
740
741static struct btrfs_root *create_reloc_root(struct btrfs_trans_handle *trans,
742 struct btrfs_root *root, u64 objectid)
743{
744 struct btrfs_fs_info *fs_info = root->fs_info;
745 struct btrfs_root *reloc_root;
746 struct extent_buffer *eb;
747 struct btrfs_root_item *root_item;
748 struct btrfs_key root_key;
749 int ret = 0;
750 bool must_abort = false;
751
752 root_item = kmalloc(sizeof(*root_item), GFP_NOFS);
753 if (!root_item)
754 return ERR_PTR(-ENOMEM);
755
756 root_key.objectid = BTRFS_TREE_RELOC_OBJECTID;
757 root_key.type = BTRFS_ROOT_ITEM_KEY;
758 root_key.offset = objectid;
759
760 if (root->root_key.objectid == objectid) {
761 u64 commit_root_gen;
762
763 /* called by btrfs_init_reloc_root */
764 ret = btrfs_copy_root(trans, root, root->commit_root, &eb,
765 BTRFS_TREE_RELOC_OBJECTID);
766 if (ret)
767 goto fail;
768
769 /*
770 * Set the last_snapshot field to the generation of the commit
771 * root - like this ctree.c:btrfs_block_can_be_shared() behaves
772 * correctly (returns true) when the relocation root is created
773 * either inside the critical section of a transaction commit
774 * (through transaction.c:qgroup_account_snapshot()) and when
775 * it's created before the transaction commit is started.
776 */
777 commit_root_gen = btrfs_header_generation(root->commit_root);
778 btrfs_set_root_last_snapshot(&root->root_item, commit_root_gen);
779 } else {
780 /*
781 * called by btrfs_reloc_post_snapshot_hook.
782 * the source tree is a reloc tree, all tree blocks
783 * modified after it was created have RELOC flag
784 * set in their headers. so it's OK to not update
785 * the 'last_snapshot'.
786 */
787 ret = btrfs_copy_root(trans, root, root->node, &eb,
788 BTRFS_TREE_RELOC_OBJECTID);
789 if (ret)
790 goto fail;
791 }
792
793 /*
794 * We have changed references at this point, we must abort the
795 * transaction if anything fails.
796 */
797 must_abort = true;
798
799 memcpy(root_item, &root->root_item, sizeof(*root_item));
800 btrfs_set_root_bytenr(root_item, eb->start);
801 btrfs_set_root_level(root_item, btrfs_header_level(eb));
802 btrfs_set_root_generation(root_item, trans->transid);
803
804 if (root->root_key.objectid == objectid) {
805 btrfs_set_root_refs(root_item, 0);
806 memset(&root_item->drop_progress, 0,
807 sizeof(struct btrfs_disk_key));
808 btrfs_set_root_drop_level(root_item, 0);
809 }
810
811 btrfs_tree_unlock(eb);
812 free_extent_buffer(eb);
813
814 ret = btrfs_insert_root(trans, fs_info->tree_root,
815 &root_key, root_item);
816 if (ret)
817 goto fail;
818
819 kfree(root_item);
820
821 reloc_root = btrfs_read_tree_root(fs_info->tree_root, &root_key);
822 if (IS_ERR(reloc_root)) {
823 ret = PTR_ERR(reloc_root);
824 goto abort;
825 }
826 set_bit(BTRFS_ROOT_SHAREABLE, &reloc_root->state);
827 reloc_root->last_trans = trans->transid;
828 return reloc_root;
829fail:
830 kfree(root_item);
831abort:
832 if (must_abort)
833 btrfs_abort_transaction(trans, ret);
834 return ERR_PTR(ret);
835}
836
837/*
838 * create reloc tree for a given fs tree. reloc tree is just a
839 * snapshot of the fs tree with special root objectid.
840 *
841 * The reloc_root comes out of here with two references, one for
842 * root->reloc_root, and another for being on the rc->reloc_roots list.
843 */
844int btrfs_init_reloc_root(struct btrfs_trans_handle *trans,
845 struct btrfs_root *root)
846{
847 struct btrfs_fs_info *fs_info = root->fs_info;
848 struct btrfs_root *reloc_root;
849 struct reloc_control *rc = fs_info->reloc_ctl;
850 struct btrfs_block_rsv *rsv;
851 int clear_rsv = 0;
852 int ret;
853
854 if (!rc)
855 return 0;
856
857 /*
858 * The subvolume has reloc tree but the swap is finished, no need to
859 * create/update the dead reloc tree
860 */
861 if (reloc_root_is_dead(root))
862 return 0;
863
864 /*
865 * This is subtle but important. We do not do
866 * record_root_in_transaction for reloc roots, instead we record their
867 * corresponding fs root, and then here we update the last trans for the
868 * reloc root. This means that we have to do this for the entire life
869 * of the reloc root, regardless of which stage of the relocation we are
870 * in.
871 */
872 if (root->reloc_root) {
873 reloc_root = root->reloc_root;
874 reloc_root->last_trans = trans->transid;
875 return 0;
876 }
877
878 /*
879 * We are merging reloc roots, we do not need new reloc trees. Also
880 * reloc trees never need their own reloc tree.
881 */
882 if (!rc->create_reloc_tree ||
883 root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID)
884 return 0;
885
886 if (!trans->reloc_reserved) {
887 rsv = trans->block_rsv;
888 trans->block_rsv = rc->block_rsv;
889 clear_rsv = 1;
890 }
891 reloc_root = create_reloc_root(trans, root, root->root_key.objectid);
892 if (clear_rsv)
893 trans->block_rsv = rsv;
894 if (IS_ERR(reloc_root))
895 return PTR_ERR(reloc_root);
896
897 ret = __add_reloc_root(reloc_root);
898 ASSERT(ret != -EEXIST);
899 if (ret) {
900 /* Pairs with create_reloc_root */
901 btrfs_put_root(reloc_root);
902 return ret;
903 }
904 root->reloc_root = btrfs_grab_root(reloc_root);
905 return 0;
906}
907
908/*
909 * update root item of reloc tree
910 */
911int btrfs_update_reloc_root(struct btrfs_trans_handle *trans,
912 struct btrfs_root *root)
913{
914 struct btrfs_fs_info *fs_info = root->fs_info;
915 struct btrfs_root *reloc_root;
916 struct btrfs_root_item *root_item;
917 int ret;
918
919 if (!have_reloc_root(root))
920 return 0;
921
922 reloc_root = root->reloc_root;
923 root_item = &reloc_root->root_item;
924
925 /*
926 * We are probably ok here, but __del_reloc_root() will drop its ref of
927 * the root. We have the ref for root->reloc_root, but just in case
928 * hold it while we update the reloc root.
929 */
930 btrfs_grab_root(reloc_root);
931
932 /* root->reloc_root will stay until current relocation finished */
933 if (fs_info->reloc_ctl->merge_reloc_tree &&
934 btrfs_root_refs(root_item) == 0) {
935 set_bit(BTRFS_ROOT_DEAD_RELOC_TREE, &root->state);
936 /*
937 * Mark the tree as dead before we change reloc_root so
938 * have_reloc_root will not touch it from now on.
939 */
940 smp_wmb();
941 __del_reloc_root(reloc_root);
942 }
943
944 if (reloc_root->commit_root != reloc_root->node) {
945 __update_reloc_root(reloc_root);
946 btrfs_set_root_node(root_item, reloc_root->node);
947 free_extent_buffer(reloc_root->commit_root);
948 reloc_root->commit_root = btrfs_root_node(reloc_root);
949 }
950
951 ret = btrfs_update_root(trans, fs_info->tree_root,
952 &reloc_root->root_key, root_item);
953 btrfs_put_root(reloc_root);
954 return ret;
955}
956
957/*
958 * helper to find first cached inode with inode number >= objectid
959 * in a subvolume
960 */
961static struct inode *find_next_inode(struct btrfs_root *root, u64 objectid)
962{
963 struct rb_node *node;
964 struct rb_node *prev;
965 struct btrfs_inode *entry;
966 struct inode *inode;
967
968 spin_lock(&root->inode_lock);
969again:
970 node = root->inode_tree.rb_node;
971 prev = NULL;
972 while (node) {
973 prev = node;
974 entry = rb_entry(node, struct btrfs_inode, rb_node);
975
976 if (objectid < btrfs_ino(entry))
977 node = node->rb_left;
978 else if (objectid > btrfs_ino(entry))
979 node = node->rb_right;
980 else
981 break;
982 }
983 if (!node) {
984 while (prev) {
985 entry = rb_entry(prev, struct btrfs_inode, rb_node);
986 if (objectid <= btrfs_ino(entry)) {
987 node = prev;
988 break;
989 }
990 prev = rb_next(prev);
991 }
992 }
993 while (node) {
994 entry = rb_entry(node, struct btrfs_inode, rb_node);
995 inode = igrab(&entry->vfs_inode);
996 if (inode) {
997 spin_unlock(&root->inode_lock);
998 return inode;
999 }
1000
1001 objectid = btrfs_ino(entry) + 1;
1002 if (cond_resched_lock(&root->inode_lock))
1003 goto again;
1004
1005 node = rb_next(node);
1006 }
1007 spin_unlock(&root->inode_lock);
1008 return NULL;
1009}
1010
1011/*
1012 * get new location of data
1013 */
1014static int get_new_location(struct inode *reloc_inode, u64 *new_bytenr,
1015 u64 bytenr, u64 num_bytes)
1016{
1017 struct btrfs_root *root = BTRFS_I(reloc_inode)->root;
1018 struct btrfs_path *path;
1019 struct btrfs_file_extent_item *fi;
1020 struct extent_buffer *leaf;
1021 int ret;
1022
1023 path = btrfs_alloc_path();
1024 if (!path)
1025 return -ENOMEM;
1026
1027 bytenr -= BTRFS_I(reloc_inode)->index_cnt;
1028 ret = btrfs_lookup_file_extent(NULL, root, path,
1029 btrfs_ino(BTRFS_I(reloc_inode)), bytenr, 0);
1030 if (ret < 0)
1031 goto out;
1032 if (ret > 0) {
1033 ret = -ENOENT;
1034 goto out;
1035 }
1036
1037 leaf = path->nodes[0];
1038 fi = btrfs_item_ptr(leaf, path->slots[0],
1039 struct btrfs_file_extent_item);
1040
1041 BUG_ON(btrfs_file_extent_offset(leaf, fi) ||
1042 btrfs_file_extent_compression(leaf, fi) ||
1043 btrfs_file_extent_encryption(leaf, fi) ||
1044 btrfs_file_extent_other_encoding(leaf, fi));
1045
1046 if (num_bytes != btrfs_file_extent_disk_num_bytes(leaf, fi)) {
1047 ret = -EINVAL;
1048 goto out;
1049 }
1050
1051 *new_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
1052 ret = 0;
1053out:
1054 btrfs_free_path(path);
1055 return ret;
1056}
1057
1058/*
1059 * update file extent items in the tree leaf to point to
1060 * the new locations.
1061 */
1062static noinline_for_stack
1063int replace_file_extents(struct btrfs_trans_handle *trans,
1064 struct reloc_control *rc,
1065 struct btrfs_root *root,
1066 struct extent_buffer *leaf)
1067{
1068 struct btrfs_fs_info *fs_info = root->fs_info;
1069 struct btrfs_key key;
1070 struct btrfs_file_extent_item *fi;
1071 struct inode *inode = NULL;
1072 u64 parent;
1073 u64 bytenr;
1074 u64 new_bytenr = 0;
1075 u64 num_bytes;
1076 u64 end;
1077 u32 nritems;
1078 u32 i;
1079 int ret = 0;
1080 int first = 1;
1081 int dirty = 0;
1082
1083 if (rc->stage != UPDATE_DATA_PTRS)
1084 return 0;
1085
1086 /* reloc trees always use full backref */
1087 if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID)
1088 parent = leaf->start;
1089 else
1090 parent = 0;
1091
1092 nritems = btrfs_header_nritems(leaf);
1093 for (i = 0; i < nritems; i++) {
1094 struct btrfs_ref ref = { 0 };
1095
1096 cond_resched();
1097 btrfs_item_key_to_cpu(leaf, &key, i);
1098 if (key.type != BTRFS_EXTENT_DATA_KEY)
1099 continue;
1100 fi = btrfs_item_ptr(leaf, i, struct btrfs_file_extent_item);
1101 if (btrfs_file_extent_type(leaf, fi) ==
1102 BTRFS_FILE_EXTENT_INLINE)
1103 continue;
1104 bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
1105 num_bytes = btrfs_file_extent_disk_num_bytes(leaf, fi);
1106 if (bytenr == 0)
1107 continue;
1108 if (!in_range(bytenr, rc->block_group->start,
1109 rc->block_group->length))
1110 continue;
1111
1112 /*
1113 * if we are modifying block in fs tree, wait for read_folio
1114 * to complete and drop the extent cache
1115 */
1116 if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) {
1117 if (first) {
1118 inode = find_next_inode(root, key.objectid);
1119 first = 0;
1120 } else if (inode && btrfs_ino(BTRFS_I(inode)) < key.objectid) {
1121 btrfs_add_delayed_iput(BTRFS_I(inode));
1122 inode = find_next_inode(root, key.objectid);
1123 }
1124 if (inode && btrfs_ino(BTRFS_I(inode)) == key.objectid) {
1125 struct extent_state *cached_state = NULL;
1126
1127 end = key.offset +
1128 btrfs_file_extent_num_bytes(leaf, fi);
1129 WARN_ON(!IS_ALIGNED(key.offset,
1130 fs_info->sectorsize));
1131 WARN_ON(!IS_ALIGNED(end, fs_info->sectorsize));
1132 end--;
1133 ret = try_lock_extent(&BTRFS_I(inode)->io_tree,
1134 key.offset, end,
1135 &cached_state);
1136 if (!ret)
1137 continue;
1138
1139 btrfs_drop_extent_map_range(BTRFS_I(inode),
1140 key.offset, end, true);
1141 unlock_extent(&BTRFS_I(inode)->io_tree,
1142 key.offset, end, &cached_state);
1143 }
1144 }
1145
1146 ret = get_new_location(rc->data_inode, &new_bytenr,
1147 bytenr, num_bytes);
1148 if (ret) {
1149 /*
1150 * Don't have to abort since we've not changed anything
1151 * in the file extent yet.
1152 */
1153 break;
1154 }
1155
1156 btrfs_set_file_extent_disk_bytenr(leaf, fi, new_bytenr);
1157 dirty = 1;
1158
1159 key.offset -= btrfs_file_extent_offset(leaf, fi);
1160 btrfs_init_generic_ref(&ref, BTRFS_ADD_DELAYED_REF, new_bytenr,
1161 num_bytes, parent);
1162 btrfs_init_data_ref(&ref, btrfs_header_owner(leaf),
1163 key.objectid, key.offset,
1164 root->root_key.objectid, false);
1165 ret = btrfs_inc_extent_ref(trans, &ref);
1166 if (ret) {
1167 btrfs_abort_transaction(trans, ret);
1168 break;
1169 }
1170
1171 btrfs_init_generic_ref(&ref, BTRFS_DROP_DELAYED_REF, bytenr,
1172 num_bytes, parent);
1173 btrfs_init_data_ref(&ref, btrfs_header_owner(leaf),
1174 key.objectid, key.offset,
1175 root->root_key.objectid, false);
1176 ret = btrfs_free_extent(trans, &ref);
1177 if (ret) {
1178 btrfs_abort_transaction(trans, ret);
1179 break;
1180 }
1181 }
1182 if (dirty)
1183 btrfs_mark_buffer_dirty(leaf);
1184 if (inode)
1185 btrfs_add_delayed_iput(BTRFS_I(inode));
1186 return ret;
1187}
1188
1189static noinline_for_stack
1190int memcmp_node_keys(struct extent_buffer *eb, int slot,
1191 struct btrfs_path *path, int level)
1192{
1193 struct btrfs_disk_key key1;
1194 struct btrfs_disk_key key2;
1195 btrfs_node_key(eb, &key1, slot);
1196 btrfs_node_key(path->nodes[level], &key2, path->slots[level]);
1197 return memcmp(&key1, &key2, sizeof(key1));
1198}
1199
1200/*
1201 * try to replace tree blocks in fs tree with the new blocks
1202 * in reloc tree. tree blocks haven't been modified since the
1203 * reloc tree was create can be replaced.
1204 *
1205 * if a block was replaced, level of the block + 1 is returned.
1206 * if no block got replaced, 0 is returned. if there are other
1207 * errors, a negative error number is returned.
1208 */
1209static noinline_for_stack
1210int replace_path(struct btrfs_trans_handle *trans, struct reloc_control *rc,
1211 struct btrfs_root *dest, struct btrfs_root *src,
1212 struct btrfs_path *path, struct btrfs_key *next_key,
1213 int lowest_level, int max_level)
1214{
1215 struct btrfs_fs_info *fs_info = dest->fs_info;
1216 struct extent_buffer *eb;
1217 struct extent_buffer *parent;
1218 struct btrfs_ref ref = { 0 };
1219 struct btrfs_key key;
1220 u64 old_bytenr;
1221 u64 new_bytenr;
1222 u64 old_ptr_gen;
1223 u64 new_ptr_gen;
1224 u64 last_snapshot;
1225 u32 blocksize;
1226 int cow = 0;
1227 int level;
1228 int ret;
1229 int slot;
1230
1231 ASSERT(src->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID);
1232 ASSERT(dest->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID);
1233
1234 last_snapshot = btrfs_root_last_snapshot(&src->root_item);
1235again:
1236 slot = path->slots[lowest_level];
1237 btrfs_node_key_to_cpu(path->nodes[lowest_level], &key, slot);
1238
1239 eb = btrfs_lock_root_node(dest);
1240 level = btrfs_header_level(eb);
1241
1242 if (level < lowest_level) {
1243 btrfs_tree_unlock(eb);
1244 free_extent_buffer(eb);
1245 return 0;
1246 }
1247
1248 if (cow) {
1249 ret = btrfs_cow_block(trans, dest, eb, NULL, 0, &eb,
1250 BTRFS_NESTING_COW);
1251 if (ret) {
1252 btrfs_tree_unlock(eb);
1253 free_extent_buffer(eb);
1254 return ret;
1255 }
1256 }
1257
1258 if (next_key) {
1259 next_key->objectid = (u64)-1;
1260 next_key->type = (u8)-1;
1261 next_key->offset = (u64)-1;
1262 }
1263
1264 parent = eb;
1265 while (1) {
1266 level = btrfs_header_level(parent);
1267 ASSERT(level >= lowest_level);
1268
1269 ret = btrfs_bin_search(parent, &key, &slot);
1270 if (ret < 0)
1271 break;
1272 if (ret && slot > 0)
1273 slot--;
1274
1275 if (next_key && slot + 1 < btrfs_header_nritems(parent))
1276 btrfs_node_key_to_cpu(parent, next_key, slot + 1);
1277
1278 old_bytenr = btrfs_node_blockptr(parent, slot);
1279 blocksize = fs_info->nodesize;
1280 old_ptr_gen = btrfs_node_ptr_generation(parent, slot);
1281
1282 if (level <= max_level) {
1283 eb = path->nodes[level];
1284 new_bytenr = btrfs_node_blockptr(eb,
1285 path->slots[level]);
1286 new_ptr_gen = btrfs_node_ptr_generation(eb,
1287 path->slots[level]);
1288 } else {
1289 new_bytenr = 0;
1290 new_ptr_gen = 0;
1291 }
1292
1293 if (WARN_ON(new_bytenr > 0 && new_bytenr == old_bytenr)) {
1294 ret = level;
1295 break;
1296 }
1297
1298 if (new_bytenr == 0 || old_ptr_gen > last_snapshot ||
1299 memcmp_node_keys(parent, slot, path, level)) {
1300 if (level <= lowest_level) {
1301 ret = 0;
1302 break;
1303 }
1304
1305 eb = btrfs_read_node_slot(parent, slot);
1306 if (IS_ERR(eb)) {
1307 ret = PTR_ERR(eb);
1308 break;
1309 }
1310 btrfs_tree_lock(eb);
1311 if (cow) {
1312 ret = btrfs_cow_block(trans, dest, eb, parent,
1313 slot, &eb,
1314 BTRFS_NESTING_COW);
1315 if (ret) {
1316 btrfs_tree_unlock(eb);
1317 free_extent_buffer(eb);
1318 break;
1319 }
1320 }
1321
1322 btrfs_tree_unlock(parent);
1323 free_extent_buffer(parent);
1324
1325 parent = eb;
1326 continue;
1327 }
1328
1329 if (!cow) {
1330 btrfs_tree_unlock(parent);
1331 free_extent_buffer(parent);
1332 cow = 1;
1333 goto again;
1334 }
1335
1336 btrfs_node_key_to_cpu(path->nodes[level], &key,
1337 path->slots[level]);
1338 btrfs_release_path(path);
1339
1340 path->lowest_level = level;
1341 set_bit(BTRFS_ROOT_RESET_LOCKDEP_CLASS, &src->state);
1342 ret = btrfs_search_slot(trans, src, &key, path, 0, 1);
1343 clear_bit(BTRFS_ROOT_RESET_LOCKDEP_CLASS, &src->state);
1344 path->lowest_level = 0;
1345 if (ret) {
1346 if (ret > 0)
1347 ret = -ENOENT;
1348 break;
1349 }
1350
1351 /*
1352 * Info qgroup to trace both subtrees.
1353 *
1354 * We must trace both trees.
1355 * 1) Tree reloc subtree
1356 * If not traced, we will leak data numbers
1357 * 2) Fs subtree
1358 * If not traced, we will double count old data
1359 *
1360 * We don't scan the subtree right now, but only record
1361 * the swapped tree blocks.
1362 * The real subtree rescan is delayed until we have new
1363 * CoW on the subtree root node before transaction commit.
1364 */
1365 ret = btrfs_qgroup_add_swapped_blocks(trans, dest,
1366 rc->block_group, parent, slot,
1367 path->nodes[level], path->slots[level],
1368 last_snapshot);
1369 if (ret < 0)
1370 break;
1371 /*
1372 * swap blocks in fs tree and reloc tree.
1373 */
1374 btrfs_set_node_blockptr(parent, slot, new_bytenr);
1375 btrfs_set_node_ptr_generation(parent, slot, new_ptr_gen);
1376 btrfs_mark_buffer_dirty(parent);
1377
1378 btrfs_set_node_blockptr(path->nodes[level],
1379 path->slots[level], old_bytenr);
1380 btrfs_set_node_ptr_generation(path->nodes[level],
1381 path->slots[level], old_ptr_gen);
1382 btrfs_mark_buffer_dirty(path->nodes[level]);
1383
1384 btrfs_init_generic_ref(&ref, BTRFS_ADD_DELAYED_REF, old_bytenr,
1385 blocksize, path->nodes[level]->start);
1386 btrfs_init_tree_ref(&ref, level - 1, src->root_key.objectid,
1387 0, true);
1388 ret = btrfs_inc_extent_ref(trans, &ref);
1389 if (ret) {
1390 btrfs_abort_transaction(trans, ret);
1391 break;
1392 }
1393 btrfs_init_generic_ref(&ref, BTRFS_ADD_DELAYED_REF, new_bytenr,
1394 blocksize, 0);
1395 btrfs_init_tree_ref(&ref, level - 1, dest->root_key.objectid, 0,
1396 true);
1397 ret = btrfs_inc_extent_ref(trans, &ref);
1398 if (ret) {
1399 btrfs_abort_transaction(trans, ret);
1400 break;
1401 }
1402
1403 btrfs_init_generic_ref(&ref, BTRFS_DROP_DELAYED_REF, new_bytenr,
1404 blocksize, path->nodes[level]->start);
1405 btrfs_init_tree_ref(&ref, level - 1, src->root_key.objectid,
1406 0, true);
1407 ret = btrfs_free_extent(trans, &ref);
1408 if (ret) {
1409 btrfs_abort_transaction(trans, ret);
1410 break;
1411 }
1412
1413 btrfs_init_generic_ref(&ref, BTRFS_DROP_DELAYED_REF, old_bytenr,
1414 blocksize, 0);
1415 btrfs_init_tree_ref(&ref, level - 1, dest->root_key.objectid,
1416 0, true);
1417 ret = btrfs_free_extent(trans, &ref);
1418 if (ret) {
1419 btrfs_abort_transaction(trans, ret);
1420 break;
1421 }
1422
1423 btrfs_unlock_up_safe(path, 0);
1424
1425 ret = level;
1426 break;
1427 }
1428 btrfs_tree_unlock(parent);
1429 free_extent_buffer(parent);
1430 return ret;
1431}
1432
1433/*
1434 * helper to find next relocated block in reloc tree
1435 */
1436static noinline_for_stack
1437int walk_up_reloc_tree(struct btrfs_root *root, struct btrfs_path *path,
1438 int *level)
1439{
1440 struct extent_buffer *eb;
1441 int i;
1442 u64 last_snapshot;
1443 u32 nritems;
1444
1445 last_snapshot = btrfs_root_last_snapshot(&root->root_item);
1446
1447 for (i = 0; i < *level; i++) {
1448 free_extent_buffer(path->nodes[i]);
1449 path->nodes[i] = NULL;
1450 }
1451
1452 for (i = *level; i < BTRFS_MAX_LEVEL && path->nodes[i]; i++) {
1453 eb = path->nodes[i];
1454 nritems = btrfs_header_nritems(eb);
1455 while (path->slots[i] + 1 < nritems) {
1456 path->slots[i]++;
1457 if (btrfs_node_ptr_generation(eb, path->slots[i]) <=
1458 last_snapshot)
1459 continue;
1460
1461 *level = i;
1462 return 0;
1463 }
1464 free_extent_buffer(path->nodes[i]);
1465 path->nodes[i] = NULL;
1466 }
1467 return 1;
1468}
1469
1470/*
1471 * walk down reloc tree to find relocated block of lowest level
1472 */
1473static noinline_for_stack
1474int walk_down_reloc_tree(struct btrfs_root *root, struct btrfs_path *path,
1475 int *level)
1476{
1477 struct extent_buffer *eb = NULL;
1478 int i;
1479 u64 ptr_gen = 0;
1480 u64 last_snapshot;
1481 u32 nritems;
1482
1483 last_snapshot = btrfs_root_last_snapshot(&root->root_item);
1484
1485 for (i = *level; i > 0; i--) {
1486 eb = path->nodes[i];
1487 nritems = btrfs_header_nritems(eb);
1488 while (path->slots[i] < nritems) {
1489 ptr_gen = btrfs_node_ptr_generation(eb, path->slots[i]);
1490 if (ptr_gen > last_snapshot)
1491 break;
1492 path->slots[i]++;
1493 }
1494 if (path->slots[i] >= nritems) {
1495 if (i == *level)
1496 break;
1497 *level = i + 1;
1498 return 0;
1499 }
1500 if (i == 1) {
1501 *level = i;
1502 return 0;
1503 }
1504
1505 eb = btrfs_read_node_slot(eb, path->slots[i]);
1506 if (IS_ERR(eb))
1507 return PTR_ERR(eb);
1508 BUG_ON(btrfs_header_level(eb) != i - 1);
1509 path->nodes[i - 1] = eb;
1510 path->slots[i - 1] = 0;
1511 }
1512 return 1;
1513}
1514
1515/*
1516 * invalidate extent cache for file extents whose key in range of
1517 * [min_key, max_key)
1518 */
1519static int invalidate_extent_cache(struct btrfs_root *root,
1520 struct btrfs_key *min_key,
1521 struct btrfs_key *max_key)
1522{
1523 struct btrfs_fs_info *fs_info = root->fs_info;
1524 struct inode *inode = NULL;
1525 u64 objectid;
1526 u64 start, end;
1527 u64 ino;
1528
1529 objectid = min_key->objectid;
1530 while (1) {
1531 struct extent_state *cached_state = NULL;
1532
1533 cond_resched();
1534 iput(inode);
1535
1536 if (objectid > max_key->objectid)
1537 break;
1538
1539 inode = find_next_inode(root, objectid);
1540 if (!inode)
1541 break;
1542 ino = btrfs_ino(BTRFS_I(inode));
1543
1544 if (ino > max_key->objectid) {
1545 iput(inode);
1546 break;
1547 }
1548
1549 objectid = ino + 1;
1550 if (!S_ISREG(inode->i_mode))
1551 continue;
1552
1553 if (unlikely(min_key->objectid == ino)) {
1554 if (min_key->type > BTRFS_EXTENT_DATA_KEY)
1555 continue;
1556 if (min_key->type < BTRFS_EXTENT_DATA_KEY)
1557 start = 0;
1558 else {
1559 start = min_key->offset;
1560 WARN_ON(!IS_ALIGNED(start, fs_info->sectorsize));
1561 }
1562 } else {
1563 start = 0;
1564 }
1565
1566 if (unlikely(max_key->objectid == ino)) {
1567 if (max_key->type < BTRFS_EXTENT_DATA_KEY)
1568 continue;
1569 if (max_key->type > BTRFS_EXTENT_DATA_KEY) {
1570 end = (u64)-1;
1571 } else {
1572 if (max_key->offset == 0)
1573 continue;
1574 end = max_key->offset;
1575 WARN_ON(!IS_ALIGNED(end, fs_info->sectorsize));
1576 end--;
1577 }
1578 } else {
1579 end = (u64)-1;
1580 }
1581
1582 /* the lock_extent waits for read_folio to complete */
1583 lock_extent(&BTRFS_I(inode)->io_tree, start, end, &cached_state);
1584 btrfs_drop_extent_map_range(BTRFS_I(inode), start, end, true);
1585 unlock_extent(&BTRFS_I(inode)->io_tree, start, end, &cached_state);
1586 }
1587 return 0;
1588}
1589
1590static int find_next_key(struct btrfs_path *path, int level,
1591 struct btrfs_key *key)
1592
1593{
1594 while (level < BTRFS_MAX_LEVEL) {
1595 if (!path->nodes[level])
1596 break;
1597 if (path->slots[level] + 1 <
1598 btrfs_header_nritems(path->nodes[level])) {
1599 btrfs_node_key_to_cpu(path->nodes[level], key,
1600 path->slots[level] + 1);
1601 return 0;
1602 }
1603 level++;
1604 }
1605 return 1;
1606}
1607
1608/*
1609 * Insert current subvolume into reloc_control::dirty_subvol_roots
1610 */
1611static int insert_dirty_subvol(struct btrfs_trans_handle *trans,
1612 struct reloc_control *rc,
1613 struct btrfs_root *root)
1614{
1615 struct btrfs_root *reloc_root = root->reloc_root;
1616 struct btrfs_root_item *reloc_root_item;
1617 int ret;
1618
1619 /* @root must be a subvolume tree root with a valid reloc tree */
1620 ASSERT(root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID);
1621 ASSERT(reloc_root);
1622
1623 reloc_root_item = &reloc_root->root_item;
1624 memset(&reloc_root_item->drop_progress, 0,
1625 sizeof(reloc_root_item->drop_progress));
1626 btrfs_set_root_drop_level(reloc_root_item, 0);
1627 btrfs_set_root_refs(reloc_root_item, 0);
1628 ret = btrfs_update_reloc_root(trans, root);
1629 if (ret)
1630 return ret;
1631
1632 if (list_empty(&root->reloc_dirty_list)) {
1633 btrfs_grab_root(root);
1634 list_add_tail(&root->reloc_dirty_list, &rc->dirty_subvol_roots);
1635 }
1636
1637 return 0;
1638}
1639
1640static int clean_dirty_subvols(struct reloc_control *rc)
1641{
1642 struct btrfs_root *root;
1643 struct btrfs_root *next;
1644 int ret = 0;
1645 int ret2;
1646
1647 list_for_each_entry_safe(root, next, &rc->dirty_subvol_roots,
1648 reloc_dirty_list) {
1649 if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) {
1650 /* Merged subvolume, cleanup its reloc root */
1651 struct btrfs_root *reloc_root = root->reloc_root;
1652
1653 list_del_init(&root->reloc_dirty_list);
1654 root->reloc_root = NULL;
1655 /*
1656 * Need barrier to ensure clear_bit() only happens after
1657 * root->reloc_root = NULL. Pairs with have_reloc_root.
1658 */
1659 smp_wmb();
1660 clear_bit(BTRFS_ROOT_DEAD_RELOC_TREE, &root->state);
1661 if (reloc_root) {
1662 /*
1663 * btrfs_drop_snapshot drops our ref we hold for
1664 * ->reloc_root. If it fails however we must
1665 * drop the ref ourselves.
1666 */
1667 ret2 = btrfs_drop_snapshot(reloc_root, 0, 1);
1668 if (ret2 < 0) {
1669 btrfs_put_root(reloc_root);
1670 if (!ret)
1671 ret = ret2;
1672 }
1673 }
1674 btrfs_put_root(root);
1675 } else {
1676 /* Orphan reloc tree, just clean it up */
1677 ret2 = btrfs_drop_snapshot(root, 0, 1);
1678 if (ret2 < 0) {
1679 btrfs_put_root(root);
1680 if (!ret)
1681 ret = ret2;
1682 }
1683 }
1684 }
1685 return ret;
1686}
1687
1688/*
1689 * merge the relocated tree blocks in reloc tree with corresponding
1690 * fs tree.
1691 */
1692static noinline_for_stack int merge_reloc_root(struct reloc_control *rc,
1693 struct btrfs_root *root)
1694{
1695 struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
1696 struct btrfs_key key;
1697 struct btrfs_key next_key;
1698 struct btrfs_trans_handle *trans = NULL;
1699 struct btrfs_root *reloc_root;
1700 struct btrfs_root_item *root_item;
1701 struct btrfs_path *path;
1702 struct extent_buffer *leaf;
1703 int reserve_level;
1704 int level;
1705 int max_level;
1706 int replaced = 0;
1707 int ret = 0;
1708 u32 min_reserved;
1709
1710 path = btrfs_alloc_path();
1711 if (!path)
1712 return -ENOMEM;
1713 path->reada = READA_FORWARD;
1714
1715 reloc_root = root->reloc_root;
1716 root_item = &reloc_root->root_item;
1717
1718 if (btrfs_disk_key_objectid(&root_item->drop_progress) == 0) {
1719 level = btrfs_root_level(root_item);
1720 atomic_inc(&reloc_root->node->refs);
1721 path->nodes[level] = reloc_root->node;
1722 path->slots[level] = 0;
1723 } else {
1724 btrfs_disk_key_to_cpu(&key, &root_item->drop_progress);
1725
1726 level = btrfs_root_drop_level(root_item);
1727 BUG_ON(level == 0);
1728 path->lowest_level = level;
1729 ret = btrfs_search_slot(NULL, reloc_root, &key, path, 0, 0);
1730 path->lowest_level = 0;
1731 if (ret < 0) {
1732 btrfs_free_path(path);
1733 return ret;
1734 }
1735
1736 btrfs_node_key_to_cpu(path->nodes[level], &next_key,
1737 path->slots[level]);
1738 WARN_ON(memcmp(&key, &next_key, sizeof(key)));
1739
1740 btrfs_unlock_up_safe(path, 0);
1741 }
1742
1743 /*
1744 * In merge_reloc_root(), we modify the upper level pointer to swap the
1745 * tree blocks between reloc tree and subvolume tree. Thus for tree
1746 * block COW, we COW at most from level 1 to root level for each tree.
1747 *
1748 * Thus the needed metadata size is at most root_level * nodesize,
1749 * and * 2 since we have two trees to COW.
1750 */
1751 reserve_level = max_t(int, 1, btrfs_root_level(root_item));
1752 min_reserved = fs_info->nodesize * reserve_level * 2;
1753 memset(&next_key, 0, sizeof(next_key));
1754
1755 while (1) {
1756 ret = btrfs_block_rsv_refill(fs_info, rc->block_rsv,
1757 min_reserved,
1758 BTRFS_RESERVE_FLUSH_LIMIT);
1759 if (ret)
1760 goto out;
1761 trans = btrfs_start_transaction(root, 0);
1762 if (IS_ERR(trans)) {
1763 ret = PTR_ERR(trans);
1764 trans = NULL;
1765 goto out;
1766 }
1767
1768 /*
1769 * At this point we no longer have a reloc_control, so we can't
1770 * depend on btrfs_init_reloc_root to update our last_trans.
1771 *
1772 * But that's ok, we started the trans handle on our
1773 * corresponding fs_root, which means it's been added to the
1774 * dirty list. At commit time we'll still call
1775 * btrfs_update_reloc_root() and update our root item
1776 * appropriately.
1777 */
1778 reloc_root->last_trans = trans->transid;
1779 trans->block_rsv = rc->block_rsv;
1780
1781 replaced = 0;
1782 max_level = level;
1783
1784 ret = walk_down_reloc_tree(reloc_root, path, &level);
1785 if (ret < 0)
1786 goto out;
1787 if (ret > 0)
1788 break;
1789
1790 if (!find_next_key(path, level, &key) &&
1791 btrfs_comp_cpu_keys(&next_key, &key) >= 0) {
1792 ret = 0;
1793 } else {
1794 ret = replace_path(trans, rc, root, reloc_root, path,
1795 &next_key, level, max_level);
1796 }
1797 if (ret < 0)
1798 goto out;
1799 if (ret > 0) {
1800 level = ret;
1801 btrfs_node_key_to_cpu(path->nodes[level], &key,
1802 path->slots[level]);
1803 replaced = 1;
1804 }
1805
1806 ret = walk_up_reloc_tree(reloc_root, path, &level);
1807 if (ret > 0)
1808 break;
1809
1810 BUG_ON(level == 0);
1811 /*
1812 * save the merging progress in the drop_progress.
1813 * this is OK since root refs == 1 in this case.
1814 */
1815 btrfs_node_key(path->nodes[level], &root_item->drop_progress,
1816 path->slots[level]);
1817 btrfs_set_root_drop_level(root_item, level);
1818
1819 btrfs_end_transaction_throttle(trans);
1820 trans = NULL;
1821
1822 btrfs_btree_balance_dirty(fs_info);
1823
1824 if (replaced && rc->stage == UPDATE_DATA_PTRS)
1825 invalidate_extent_cache(root, &key, &next_key);
1826 }
1827
1828 /*
1829 * handle the case only one block in the fs tree need to be
1830 * relocated and the block is tree root.
1831 */
1832 leaf = btrfs_lock_root_node(root);
1833 ret = btrfs_cow_block(trans, root, leaf, NULL, 0, &leaf,
1834 BTRFS_NESTING_COW);
1835 btrfs_tree_unlock(leaf);
1836 free_extent_buffer(leaf);
1837out:
1838 btrfs_free_path(path);
1839
1840 if (ret == 0) {
1841 ret = insert_dirty_subvol(trans, rc, root);
1842 if (ret)
1843 btrfs_abort_transaction(trans, ret);
1844 }
1845
1846 if (trans)
1847 btrfs_end_transaction_throttle(trans);
1848
1849 btrfs_btree_balance_dirty(fs_info);
1850
1851 if (replaced && rc->stage == UPDATE_DATA_PTRS)
1852 invalidate_extent_cache(root, &key, &next_key);
1853
1854 return ret;
1855}
1856
1857static noinline_for_stack
1858int prepare_to_merge(struct reloc_control *rc, int err)
1859{
1860 struct btrfs_root *root = rc->extent_root;
1861 struct btrfs_fs_info *fs_info = root->fs_info;
1862 struct btrfs_root *reloc_root;
1863 struct btrfs_trans_handle *trans;
1864 LIST_HEAD(reloc_roots);
1865 u64 num_bytes = 0;
1866 int ret;
1867
1868 mutex_lock(&fs_info->reloc_mutex);
1869 rc->merging_rsv_size += fs_info->nodesize * (BTRFS_MAX_LEVEL - 1) * 2;
1870 rc->merging_rsv_size += rc->nodes_relocated * 2;
1871 mutex_unlock(&fs_info->reloc_mutex);
1872
1873again:
1874 if (!err) {
1875 num_bytes = rc->merging_rsv_size;
1876 ret = btrfs_block_rsv_add(fs_info, rc->block_rsv, num_bytes,
1877 BTRFS_RESERVE_FLUSH_ALL);
1878 if (ret)
1879 err = ret;
1880 }
1881
1882 trans = btrfs_join_transaction(rc->extent_root);
1883 if (IS_ERR(trans)) {
1884 if (!err)
1885 btrfs_block_rsv_release(fs_info, rc->block_rsv,
1886 num_bytes, NULL);
1887 return PTR_ERR(trans);
1888 }
1889
1890 if (!err) {
1891 if (num_bytes != rc->merging_rsv_size) {
1892 btrfs_end_transaction(trans);
1893 btrfs_block_rsv_release(fs_info, rc->block_rsv,
1894 num_bytes, NULL);
1895 goto again;
1896 }
1897 }
1898
1899 rc->merge_reloc_tree = 1;
1900
1901 while (!list_empty(&rc->reloc_roots)) {
1902 reloc_root = list_entry(rc->reloc_roots.next,
1903 struct btrfs_root, root_list);
1904 list_del_init(&reloc_root->root_list);
1905
1906 root = btrfs_get_fs_root(fs_info, reloc_root->root_key.offset,
1907 false);
1908 if (IS_ERR(root)) {
1909 /*
1910 * Even if we have an error we need this reloc root
1911 * back on our list so we can clean up properly.
1912 */
1913 list_add(&reloc_root->root_list, &reloc_roots);
1914 btrfs_abort_transaction(trans, (int)PTR_ERR(root));
1915 if (!err)
1916 err = PTR_ERR(root);
1917 break;
1918 }
1919 ASSERT(root->reloc_root == reloc_root);
1920
1921 /*
1922 * set reference count to 1, so btrfs_recover_relocation
1923 * knows it should resumes merging
1924 */
1925 if (!err)
1926 btrfs_set_root_refs(&reloc_root->root_item, 1);
1927 ret = btrfs_update_reloc_root(trans, root);
1928
1929 /*
1930 * Even if we have an error we need this reloc root back on our
1931 * list so we can clean up properly.
1932 */
1933 list_add(&reloc_root->root_list, &reloc_roots);
1934 btrfs_put_root(root);
1935
1936 if (ret) {
1937 btrfs_abort_transaction(trans, ret);
1938 if (!err)
1939 err = ret;
1940 break;
1941 }
1942 }
1943
1944 list_splice(&reloc_roots, &rc->reloc_roots);
1945
1946 if (!err)
1947 err = btrfs_commit_transaction(trans);
1948 else
1949 btrfs_end_transaction(trans);
1950 return err;
1951}
1952
1953static noinline_for_stack
1954void free_reloc_roots(struct list_head *list)
1955{
1956 struct btrfs_root *reloc_root, *tmp;
1957
1958 list_for_each_entry_safe(reloc_root, tmp, list, root_list)
1959 __del_reloc_root(reloc_root);
1960}
1961
1962static noinline_for_stack
1963void merge_reloc_roots(struct reloc_control *rc)
1964{
1965 struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
1966 struct btrfs_root *root;
1967 struct btrfs_root *reloc_root;
1968 LIST_HEAD(reloc_roots);
1969 int found = 0;
1970 int ret = 0;
1971again:
1972 root = rc->extent_root;
1973
1974 /*
1975 * this serializes us with btrfs_record_root_in_transaction,
1976 * we have to make sure nobody is in the middle of
1977 * adding their roots to the list while we are
1978 * doing this splice
1979 */
1980 mutex_lock(&fs_info->reloc_mutex);
1981 list_splice_init(&rc->reloc_roots, &reloc_roots);
1982 mutex_unlock(&fs_info->reloc_mutex);
1983
1984 while (!list_empty(&reloc_roots)) {
1985 found = 1;
1986 reloc_root = list_entry(reloc_roots.next,
1987 struct btrfs_root, root_list);
1988
1989 root = btrfs_get_fs_root(fs_info, reloc_root->root_key.offset,
1990 false);
1991 if (btrfs_root_refs(&reloc_root->root_item) > 0) {
1992 if (IS_ERR(root)) {
1993 /*
1994 * For recovery we read the fs roots on mount,
1995 * and if we didn't find the root then we marked
1996 * the reloc root as a garbage root. For normal
1997 * relocation obviously the root should exist in
1998 * memory. However there's no reason we can't
1999 * handle the error properly here just in case.
2000 */
2001 ASSERT(0);
2002 ret = PTR_ERR(root);
2003 goto out;
2004 }
2005 if (root->reloc_root != reloc_root) {
2006 /*
2007 * This is actually impossible without something
2008 * going really wrong (like weird race condition
2009 * or cosmic rays).
2010 */
2011 ASSERT(0);
2012 ret = -EINVAL;
2013 goto out;
2014 }
2015 ret = merge_reloc_root(rc, root);
2016 btrfs_put_root(root);
2017 if (ret) {
2018 if (list_empty(&reloc_root->root_list))
2019 list_add_tail(&reloc_root->root_list,
2020 &reloc_roots);
2021 goto out;
2022 }
2023 } else {
2024 if (!IS_ERR(root)) {
2025 if (root->reloc_root == reloc_root) {
2026 root->reloc_root = NULL;
2027 btrfs_put_root(reloc_root);
2028 }
2029 clear_bit(BTRFS_ROOT_DEAD_RELOC_TREE,
2030 &root->state);
2031 btrfs_put_root(root);
2032 }
2033
2034 list_del_init(&reloc_root->root_list);
2035 /* Don't forget to queue this reloc root for cleanup */
2036 list_add_tail(&reloc_root->reloc_dirty_list,
2037 &rc->dirty_subvol_roots);
2038 }
2039 }
2040
2041 if (found) {
2042 found = 0;
2043 goto again;
2044 }
2045out:
2046 if (ret) {
2047 btrfs_handle_fs_error(fs_info, ret, NULL);
2048 free_reloc_roots(&reloc_roots);
2049
2050 /* new reloc root may be added */
2051 mutex_lock(&fs_info->reloc_mutex);
2052 list_splice_init(&rc->reloc_roots, &reloc_roots);
2053 mutex_unlock(&fs_info->reloc_mutex);
2054 free_reloc_roots(&reloc_roots);
2055 }
2056
2057 /*
2058 * We used to have
2059 *
2060 * BUG_ON(!RB_EMPTY_ROOT(&rc->reloc_root_tree.rb_root));
2061 *
2062 * here, but it's wrong. If we fail to start the transaction in
2063 * prepare_to_merge() we will have only 0 ref reloc roots, none of which
2064 * have actually been removed from the reloc_root_tree rb tree. This is
2065 * fine because we're bailing here, and we hold a reference on the root
2066 * for the list that holds it, so these roots will be cleaned up when we
2067 * do the reloc_dirty_list afterwards. Meanwhile the root->reloc_root
2068 * will be cleaned up on unmount.
2069 *
2070 * The remaining nodes will be cleaned up by free_reloc_control.
2071 */
2072}
2073
2074static void free_block_list(struct rb_root *blocks)
2075{
2076 struct tree_block *block;
2077 struct rb_node *rb_node;
2078 while ((rb_node = rb_first(blocks))) {
2079 block = rb_entry(rb_node, struct tree_block, rb_node);
2080 rb_erase(rb_node, blocks);
2081 kfree(block);
2082 }
2083}
2084
2085static int record_reloc_root_in_trans(struct btrfs_trans_handle *trans,
2086 struct btrfs_root *reloc_root)
2087{
2088 struct btrfs_fs_info *fs_info = reloc_root->fs_info;
2089 struct btrfs_root *root;
2090 int ret;
2091
2092 if (reloc_root->last_trans == trans->transid)
2093 return 0;
2094
2095 root = btrfs_get_fs_root(fs_info, reloc_root->root_key.offset, false);
2096
2097 /*
2098 * This should succeed, since we can't have a reloc root without having
2099 * already looked up the actual root and created the reloc root for this
2100 * root.
2101 *
2102 * However if there's some sort of corruption where we have a ref to a
2103 * reloc root without a corresponding root this could return ENOENT.
2104 */
2105 if (IS_ERR(root)) {
2106 ASSERT(0);
2107 return PTR_ERR(root);
2108 }
2109 if (root->reloc_root != reloc_root) {
2110 ASSERT(0);
2111 btrfs_err(fs_info,
2112 "root %llu has two reloc roots associated with it",
2113 reloc_root->root_key.offset);
2114 btrfs_put_root(root);
2115 return -EUCLEAN;
2116 }
2117 ret = btrfs_record_root_in_trans(trans, root);
2118 btrfs_put_root(root);
2119
2120 return ret;
2121}
2122
2123static noinline_for_stack
2124struct btrfs_root *select_reloc_root(struct btrfs_trans_handle *trans,
2125 struct reloc_control *rc,
2126 struct btrfs_backref_node *node,
2127 struct btrfs_backref_edge *edges[])
2128{
2129 struct btrfs_backref_node *next;
2130 struct btrfs_root *root;
2131 int index = 0;
2132 int ret;
2133
2134 next = node;
2135 while (1) {
2136 cond_resched();
2137 next = walk_up_backref(next, edges, &index);
2138 root = next->root;
2139
2140 /*
2141 * If there is no root, then our references for this block are
2142 * incomplete, as we should be able to walk all the way up to a
2143 * block that is owned by a root.
2144 *
2145 * This path is only for SHAREABLE roots, so if we come upon a
2146 * non-SHAREABLE root then we have backrefs that resolve
2147 * improperly.
2148 *
2149 * Both of these cases indicate file system corruption, or a bug
2150 * in the backref walking code.
2151 */
2152 if (!root) {
2153 ASSERT(0);
2154 btrfs_err(trans->fs_info,
2155 "bytenr %llu doesn't have a backref path ending in a root",
2156 node->bytenr);
2157 return ERR_PTR(-EUCLEAN);
2158 }
2159 if (!test_bit(BTRFS_ROOT_SHAREABLE, &root->state)) {
2160 ASSERT(0);
2161 btrfs_err(trans->fs_info,
2162 "bytenr %llu has multiple refs with one ending in a non-shareable root",
2163 node->bytenr);
2164 return ERR_PTR(-EUCLEAN);
2165 }
2166
2167 if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID) {
2168 ret = record_reloc_root_in_trans(trans, root);
2169 if (ret)
2170 return ERR_PTR(ret);
2171 break;
2172 }
2173
2174 ret = btrfs_record_root_in_trans(trans, root);
2175 if (ret)
2176 return ERR_PTR(ret);
2177 root = root->reloc_root;
2178
2179 /*
2180 * We could have raced with another thread which failed, so
2181 * root->reloc_root may not be set, return ENOENT in this case.
2182 */
2183 if (!root)
2184 return ERR_PTR(-ENOENT);
2185
2186 if (next->new_bytenr != root->node->start) {
2187 /*
2188 * We just created the reloc root, so we shouldn't have
2189 * ->new_bytenr set and this shouldn't be in the changed
2190 * list. If it is then we have multiple roots pointing
2191 * at the same bytenr which indicates corruption, or
2192 * we've made a mistake in the backref walking code.
2193 */
2194 ASSERT(next->new_bytenr == 0);
2195 ASSERT(list_empty(&next->list));
2196 if (next->new_bytenr || !list_empty(&next->list)) {
2197 btrfs_err(trans->fs_info,
2198 "bytenr %llu possibly has multiple roots pointing at the same bytenr %llu",
2199 node->bytenr, next->bytenr);
2200 return ERR_PTR(-EUCLEAN);
2201 }
2202
2203 next->new_bytenr = root->node->start;
2204 btrfs_put_root(next->root);
2205 next->root = btrfs_grab_root(root);
2206 ASSERT(next->root);
2207 list_add_tail(&next->list,
2208 &rc->backref_cache.changed);
2209 mark_block_processed(rc, next);
2210 break;
2211 }
2212
2213 WARN_ON(1);
2214 root = NULL;
2215 next = walk_down_backref(edges, &index);
2216 if (!next || next->level <= node->level)
2217 break;
2218 }
2219 if (!root) {
2220 /*
2221 * This can happen if there's fs corruption or if there's a bug
2222 * in the backref lookup code.
2223 */
2224 ASSERT(0);
2225 return ERR_PTR(-ENOENT);
2226 }
2227
2228 next = node;
2229 /* setup backref node path for btrfs_reloc_cow_block */
2230 while (1) {
2231 rc->backref_cache.path[next->level] = next;
2232 if (--index < 0)
2233 break;
2234 next = edges[index]->node[UPPER];
2235 }
2236 return root;
2237}
2238
2239/*
2240 * Select a tree root for relocation.
2241 *
2242 * Return NULL if the block is not shareable. We should use do_relocation() in
2243 * this case.
2244 *
2245 * Return a tree root pointer if the block is shareable.
2246 * Return -ENOENT if the block is root of reloc tree.
2247 */
2248static noinline_for_stack
2249struct btrfs_root *select_one_root(struct btrfs_backref_node *node)
2250{
2251 struct btrfs_backref_node *next;
2252 struct btrfs_root *root;
2253 struct btrfs_root *fs_root = NULL;
2254 struct btrfs_backref_edge *edges[BTRFS_MAX_LEVEL - 1];
2255 int index = 0;
2256
2257 next = node;
2258 while (1) {
2259 cond_resched();
2260 next = walk_up_backref(next, edges, &index);
2261 root = next->root;
2262
2263 /*
2264 * This can occur if we have incomplete extent refs leading all
2265 * the way up a particular path, in this case return -EUCLEAN.
2266 */
2267 if (!root)
2268 return ERR_PTR(-EUCLEAN);
2269
2270 /* No other choice for non-shareable tree */
2271 if (!test_bit(BTRFS_ROOT_SHAREABLE, &root->state))
2272 return root;
2273
2274 if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID)
2275 fs_root = root;
2276
2277 if (next != node)
2278 return NULL;
2279
2280 next = walk_down_backref(edges, &index);
2281 if (!next || next->level <= node->level)
2282 break;
2283 }
2284
2285 if (!fs_root)
2286 return ERR_PTR(-ENOENT);
2287 return fs_root;
2288}
2289
2290static noinline_for_stack
2291u64 calcu_metadata_size(struct reloc_control *rc,
2292 struct btrfs_backref_node *node, int reserve)
2293{
2294 struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
2295 struct btrfs_backref_node *next = node;
2296 struct btrfs_backref_edge *edge;
2297 struct btrfs_backref_edge *edges[BTRFS_MAX_LEVEL - 1];
2298 u64 num_bytes = 0;
2299 int index = 0;
2300
2301 BUG_ON(reserve && node->processed);
2302
2303 while (next) {
2304 cond_resched();
2305 while (1) {
2306 if (next->processed && (reserve || next != node))
2307 break;
2308
2309 num_bytes += fs_info->nodesize;
2310
2311 if (list_empty(&next->upper))
2312 break;
2313
2314 edge = list_entry(next->upper.next,
2315 struct btrfs_backref_edge, list[LOWER]);
2316 edges[index++] = edge;
2317 next = edge->node[UPPER];
2318 }
2319 next = walk_down_backref(edges, &index);
2320 }
2321 return num_bytes;
2322}
2323
2324static int reserve_metadata_space(struct btrfs_trans_handle *trans,
2325 struct reloc_control *rc,
2326 struct btrfs_backref_node *node)
2327{
2328 struct btrfs_root *root = rc->extent_root;
2329 struct btrfs_fs_info *fs_info = root->fs_info;
2330 u64 num_bytes;
2331 int ret;
2332 u64 tmp;
2333
2334 num_bytes = calcu_metadata_size(rc, node, 1) * 2;
2335
2336 trans->block_rsv = rc->block_rsv;
2337 rc->reserved_bytes += num_bytes;
2338
2339 /*
2340 * We are under a transaction here so we can only do limited flushing.
2341 * If we get an enospc just kick back -EAGAIN so we know to drop the
2342 * transaction and try to refill when we can flush all the things.
2343 */
2344 ret = btrfs_block_rsv_refill(fs_info, rc->block_rsv, num_bytes,
2345 BTRFS_RESERVE_FLUSH_LIMIT);
2346 if (ret) {
2347 tmp = fs_info->nodesize * RELOCATION_RESERVED_NODES;
2348 while (tmp <= rc->reserved_bytes)
2349 tmp <<= 1;
2350 /*
2351 * only one thread can access block_rsv at this point,
2352 * so we don't need hold lock to protect block_rsv.
2353 * we expand more reservation size here to allow enough
2354 * space for relocation and we will return earlier in
2355 * enospc case.
2356 */
2357 rc->block_rsv->size = tmp + fs_info->nodesize *
2358 RELOCATION_RESERVED_NODES;
2359 return -EAGAIN;
2360 }
2361
2362 return 0;
2363}
2364
2365/*
2366 * relocate a block tree, and then update pointers in upper level
2367 * blocks that reference the block to point to the new location.
2368 *
2369 * if called by link_to_upper, the block has already been relocated.
2370 * in that case this function just updates pointers.
2371 */
2372static int do_relocation(struct btrfs_trans_handle *trans,
2373 struct reloc_control *rc,
2374 struct btrfs_backref_node *node,
2375 struct btrfs_key *key,
2376 struct btrfs_path *path, int lowest)
2377{
2378 struct btrfs_backref_node *upper;
2379 struct btrfs_backref_edge *edge;
2380 struct btrfs_backref_edge *edges[BTRFS_MAX_LEVEL - 1];
2381 struct btrfs_root *root;
2382 struct extent_buffer *eb;
2383 u32 blocksize;
2384 u64 bytenr;
2385 int slot;
2386 int ret = 0;
2387
2388 /*
2389 * If we are lowest then this is the first time we're processing this
2390 * block, and thus shouldn't have an eb associated with it yet.
2391 */
2392 ASSERT(!lowest || !node->eb);
2393
2394 path->lowest_level = node->level + 1;
2395 rc->backref_cache.path[node->level] = node;
2396 list_for_each_entry(edge, &node->upper, list[LOWER]) {
2397 struct btrfs_ref ref = { 0 };
2398
2399 cond_resched();
2400
2401 upper = edge->node[UPPER];
2402 root = select_reloc_root(trans, rc, upper, edges);
2403 if (IS_ERR(root)) {
2404 ret = PTR_ERR(root);
2405 goto next;
2406 }
2407
2408 if (upper->eb && !upper->locked) {
2409 if (!lowest) {
2410 ret = btrfs_bin_search(upper->eb, key, &slot);
2411 if (ret < 0)
2412 goto next;
2413 BUG_ON(ret);
2414 bytenr = btrfs_node_blockptr(upper->eb, slot);
2415 if (node->eb->start == bytenr)
2416 goto next;
2417 }
2418 btrfs_backref_drop_node_buffer(upper);
2419 }
2420
2421 if (!upper->eb) {
2422 ret = btrfs_search_slot(trans, root, key, path, 0, 1);
2423 if (ret) {
2424 if (ret > 0)
2425 ret = -ENOENT;
2426
2427 btrfs_release_path(path);
2428 break;
2429 }
2430
2431 if (!upper->eb) {
2432 upper->eb = path->nodes[upper->level];
2433 path->nodes[upper->level] = NULL;
2434 } else {
2435 BUG_ON(upper->eb != path->nodes[upper->level]);
2436 }
2437
2438 upper->locked = 1;
2439 path->locks[upper->level] = 0;
2440
2441 slot = path->slots[upper->level];
2442 btrfs_release_path(path);
2443 } else {
2444 ret = btrfs_bin_search(upper->eb, key, &slot);
2445 if (ret < 0)
2446 goto next;
2447 BUG_ON(ret);
2448 }
2449
2450 bytenr = btrfs_node_blockptr(upper->eb, slot);
2451 if (lowest) {
2452 if (bytenr != node->bytenr) {
2453 btrfs_err(root->fs_info,
2454 "lowest leaf/node mismatch: bytenr %llu node->bytenr %llu slot %d upper %llu",
2455 bytenr, node->bytenr, slot,
2456 upper->eb->start);
2457 ret = -EIO;
2458 goto next;
2459 }
2460 } else {
2461 if (node->eb->start == bytenr)
2462 goto next;
2463 }
2464
2465 blocksize = root->fs_info->nodesize;
2466 eb = btrfs_read_node_slot(upper->eb, slot);
2467 if (IS_ERR(eb)) {
2468 ret = PTR_ERR(eb);
2469 goto next;
2470 }
2471 btrfs_tree_lock(eb);
2472
2473 if (!node->eb) {
2474 ret = btrfs_cow_block(trans, root, eb, upper->eb,
2475 slot, &eb, BTRFS_NESTING_COW);
2476 btrfs_tree_unlock(eb);
2477 free_extent_buffer(eb);
2478 if (ret < 0)
2479 goto next;
2480 /*
2481 * We've just COWed this block, it should have updated
2482 * the correct backref node entry.
2483 */
2484 ASSERT(node->eb == eb);
2485 } else {
2486 btrfs_set_node_blockptr(upper->eb, slot,
2487 node->eb->start);
2488 btrfs_set_node_ptr_generation(upper->eb, slot,
2489 trans->transid);
2490 btrfs_mark_buffer_dirty(upper->eb);
2491
2492 btrfs_init_generic_ref(&ref, BTRFS_ADD_DELAYED_REF,
2493 node->eb->start, blocksize,
2494 upper->eb->start);
2495 btrfs_init_tree_ref(&ref, node->level,
2496 btrfs_header_owner(upper->eb),
2497 root->root_key.objectid, false);
2498 ret = btrfs_inc_extent_ref(trans, &ref);
2499 if (!ret)
2500 ret = btrfs_drop_subtree(trans, root, eb,
2501 upper->eb);
2502 if (ret)
2503 btrfs_abort_transaction(trans, ret);
2504 }
2505next:
2506 if (!upper->pending)
2507 btrfs_backref_drop_node_buffer(upper);
2508 else
2509 btrfs_backref_unlock_node_buffer(upper);
2510 if (ret)
2511 break;
2512 }
2513
2514 if (!ret && node->pending) {
2515 btrfs_backref_drop_node_buffer(node);
2516 list_move_tail(&node->list, &rc->backref_cache.changed);
2517 node->pending = 0;
2518 }
2519
2520 path->lowest_level = 0;
2521
2522 /*
2523 * We should have allocated all of our space in the block rsv and thus
2524 * shouldn't ENOSPC.
2525 */
2526 ASSERT(ret != -ENOSPC);
2527 return ret;
2528}
2529
2530static int link_to_upper(struct btrfs_trans_handle *trans,
2531 struct reloc_control *rc,
2532 struct btrfs_backref_node *node,
2533 struct btrfs_path *path)
2534{
2535 struct btrfs_key key;
2536
2537 btrfs_node_key_to_cpu(node->eb, &key, 0);
2538 return do_relocation(trans, rc, node, &key, path, 0);
2539}
2540
2541static int finish_pending_nodes(struct btrfs_trans_handle *trans,
2542 struct reloc_control *rc,
2543 struct btrfs_path *path, int err)
2544{
2545 LIST_HEAD(list);
2546 struct btrfs_backref_cache *cache = &rc->backref_cache;
2547 struct btrfs_backref_node *node;
2548 int level;
2549 int ret;
2550
2551 for (level = 0; level < BTRFS_MAX_LEVEL; level++) {
2552 while (!list_empty(&cache->pending[level])) {
2553 node = list_entry(cache->pending[level].next,
2554 struct btrfs_backref_node, list);
2555 list_move_tail(&node->list, &list);
2556 BUG_ON(!node->pending);
2557
2558 if (!err) {
2559 ret = link_to_upper(trans, rc, node, path);
2560 if (ret < 0)
2561 err = ret;
2562 }
2563 }
2564 list_splice_init(&list, &cache->pending[level]);
2565 }
2566 return err;
2567}
2568
2569/*
2570 * mark a block and all blocks directly/indirectly reference the block
2571 * as processed.
2572 */
2573static void update_processed_blocks(struct reloc_control *rc,
2574 struct btrfs_backref_node *node)
2575{
2576 struct btrfs_backref_node *next = node;
2577 struct btrfs_backref_edge *edge;
2578 struct btrfs_backref_edge *edges[BTRFS_MAX_LEVEL - 1];
2579 int index = 0;
2580
2581 while (next) {
2582 cond_resched();
2583 while (1) {
2584 if (next->processed)
2585 break;
2586
2587 mark_block_processed(rc, next);
2588
2589 if (list_empty(&next->upper))
2590 break;
2591
2592 edge = list_entry(next->upper.next,
2593 struct btrfs_backref_edge, list[LOWER]);
2594 edges[index++] = edge;
2595 next = edge->node[UPPER];
2596 }
2597 next = walk_down_backref(edges, &index);
2598 }
2599}
2600
2601static int tree_block_processed(u64 bytenr, struct reloc_control *rc)
2602{
2603 u32 blocksize = rc->extent_root->fs_info->nodesize;
2604
2605 if (test_range_bit(&rc->processed_blocks, bytenr,
2606 bytenr + blocksize - 1, EXTENT_DIRTY, 1, NULL))
2607 return 1;
2608 return 0;
2609}
2610
2611static int get_tree_block_key(struct btrfs_fs_info *fs_info,
2612 struct tree_block *block)
2613{
2614 struct btrfs_tree_parent_check check = {
2615 .level = block->level,
2616 .owner_root = block->owner,
2617 .transid = block->key.offset
2618 };
2619 struct extent_buffer *eb;
2620
2621 eb = read_tree_block(fs_info, block->bytenr, &check);
2622 if (IS_ERR(eb))
2623 return PTR_ERR(eb);
2624 if (!extent_buffer_uptodate(eb)) {
2625 free_extent_buffer(eb);
2626 return -EIO;
2627 }
2628 if (block->level == 0)
2629 btrfs_item_key_to_cpu(eb, &block->key, 0);
2630 else
2631 btrfs_node_key_to_cpu(eb, &block->key, 0);
2632 free_extent_buffer(eb);
2633 block->key_ready = 1;
2634 return 0;
2635}
2636
2637/*
2638 * helper function to relocate a tree block
2639 */
2640static int relocate_tree_block(struct btrfs_trans_handle *trans,
2641 struct reloc_control *rc,
2642 struct btrfs_backref_node *node,
2643 struct btrfs_key *key,
2644 struct btrfs_path *path)
2645{
2646 struct btrfs_root *root;
2647 int ret = 0;
2648
2649 if (!node)
2650 return 0;
2651
2652 /*
2653 * If we fail here we want to drop our backref_node because we are going
2654 * to start over and regenerate the tree for it.
2655 */
2656 ret = reserve_metadata_space(trans, rc, node);
2657 if (ret)
2658 goto out;
2659
2660 BUG_ON(node->processed);
2661 root = select_one_root(node);
2662 if (IS_ERR(root)) {
2663 ret = PTR_ERR(root);
2664
2665 /* See explanation in select_one_root for the -EUCLEAN case. */
2666 ASSERT(ret == -ENOENT);
2667 if (ret == -ENOENT) {
2668 ret = 0;
2669 update_processed_blocks(rc, node);
2670 }
2671 goto out;
2672 }
2673
2674 if (root) {
2675 if (test_bit(BTRFS_ROOT_SHAREABLE, &root->state)) {
2676 /*
2677 * This block was the root block of a root, and this is
2678 * the first time we're processing the block and thus it
2679 * should not have had the ->new_bytenr modified and
2680 * should have not been included on the changed list.
2681 *
2682 * However in the case of corruption we could have
2683 * multiple refs pointing to the same block improperly,
2684 * and thus we would trip over these checks. ASSERT()
2685 * for the developer case, because it could indicate a
2686 * bug in the backref code, however error out for a
2687 * normal user in the case of corruption.
2688 */
2689 ASSERT(node->new_bytenr == 0);
2690 ASSERT(list_empty(&node->list));
2691 if (node->new_bytenr || !list_empty(&node->list)) {
2692 btrfs_err(root->fs_info,
2693 "bytenr %llu has improper references to it",
2694 node->bytenr);
2695 ret = -EUCLEAN;
2696 goto out;
2697 }
2698 ret = btrfs_record_root_in_trans(trans, root);
2699 if (ret)
2700 goto out;
2701 /*
2702 * Another thread could have failed, need to check if we
2703 * have reloc_root actually set.
2704 */
2705 if (!root->reloc_root) {
2706 ret = -ENOENT;
2707 goto out;
2708 }
2709 root = root->reloc_root;
2710 node->new_bytenr = root->node->start;
2711 btrfs_put_root(node->root);
2712 node->root = btrfs_grab_root(root);
2713 ASSERT(node->root);
2714 list_add_tail(&node->list, &rc->backref_cache.changed);
2715 } else {
2716 path->lowest_level = node->level;
2717 if (root == root->fs_info->chunk_root)
2718 btrfs_reserve_chunk_metadata(trans, false);
2719 ret = btrfs_search_slot(trans, root, key, path, 0, 1);
2720 btrfs_release_path(path);
2721 if (root == root->fs_info->chunk_root)
2722 btrfs_trans_release_chunk_metadata(trans);
2723 if (ret > 0)
2724 ret = 0;
2725 }
2726 if (!ret)
2727 update_processed_blocks(rc, node);
2728 } else {
2729 ret = do_relocation(trans, rc, node, key, path, 1);
2730 }
2731out:
2732 if (ret || node->level == 0 || node->cowonly)
2733 btrfs_backref_cleanup_node(&rc->backref_cache, node);
2734 return ret;
2735}
2736
2737/*
2738 * relocate a list of blocks
2739 */
2740static noinline_for_stack
2741int relocate_tree_blocks(struct btrfs_trans_handle *trans,
2742 struct reloc_control *rc, struct rb_root *blocks)
2743{
2744 struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
2745 struct btrfs_backref_node *node;
2746 struct btrfs_path *path;
2747 struct tree_block *block;
2748 struct tree_block *next;
2749 int ret;
2750 int err = 0;
2751
2752 path = btrfs_alloc_path();
2753 if (!path) {
2754 err = -ENOMEM;
2755 goto out_free_blocks;
2756 }
2757
2758 /* Kick in readahead for tree blocks with missing keys */
2759 rbtree_postorder_for_each_entry_safe(block, next, blocks, rb_node) {
2760 if (!block->key_ready)
2761 btrfs_readahead_tree_block(fs_info, block->bytenr,
2762 block->owner, 0,
2763 block->level);
2764 }
2765
2766 /* Get first keys */
2767 rbtree_postorder_for_each_entry_safe(block, next, blocks, rb_node) {
2768 if (!block->key_ready) {
2769 err = get_tree_block_key(fs_info, block);
2770 if (err)
2771 goto out_free_path;
2772 }
2773 }
2774
2775 /* Do tree relocation */
2776 rbtree_postorder_for_each_entry_safe(block, next, blocks, rb_node) {
2777 node = build_backref_tree(rc, &block->key,
2778 block->level, block->bytenr);
2779 if (IS_ERR(node)) {
2780 err = PTR_ERR(node);
2781 goto out;
2782 }
2783
2784 ret = relocate_tree_block(trans, rc, node, &block->key,
2785 path);
2786 if (ret < 0) {
2787 err = ret;
2788 break;
2789 }
2790 }
2791out:
2792 err = finish_pending_nodes(trans, rc, path, err);
2793
2794out_free_path:
2795 btrfs_free_path(path);
2796out_free_blocks:
2797 free_block_list(blocks);
2798 return err;
2799}
2800
2801static noinline_for_stack int prealloc_file_extent_cluster(
2802 struct btrfs_inode *inode,
2803 struct file_extent_cluster *cluster)
2804{
2805 u64 alloc_hint = 0;
2806 u64 start;
2807 u64 end;
2808 u64 offset = inode->index_cnt;
2809 u64 num_bytes;
2810 int nr;
2811 int ret = 0;
2812 u64 i_size = i_size_read(&inode->vfs_inode);
2813 u64 prealloc_start = cluster->start - offset;
2814 u64 prealloc_end = cluster->end - offset;
2815 u64 cur_offset = prealloc_start;
2816
2817 /*
2818 * For subpage case, previous i_size may not be aligned to PAGE_SIZE.
2819 * This means the range [i_size, PAGE_END + 1) is filled with zeros by
2820 * btrfs_do_readpage() call of previously relocated file cluster.
2821 *
2822 * If the current cluster starts in the above range, btrfs_do_readpage()
2823 * will skip the read, and relocate_one_page() will later writeback
2824 * the padding zeros as new data, causing data corruption.
2825 *
2826 * Here we have to manually invalidate the range (i_size, PAGE_END + 1).
2827 */
2828 if (!IS_ALIGNED(i_size, PAGE_SIZE)) {
2829 struct address_space *mapping = inode->vfs_inode.i_mapping;
2830 struct btrfs_fs_info *fs_info = inode->root->fs_info;
2831 const u32 sectorsize = fs_info->sectorsize;
2832 struct page *page;
2833
2834 ASSERT(sectorsize < PAGE_SIZE);
2835 ASSERT(IS_ALIGNED(i_size, sectorsize));
2836
2837 /*
2838 * Subpage can't handle page with DIRTY but without UPTODATE
2839 * bit as it can lead to the following deadlock:
2840 *
2841 * btrfs_read_folio()
2842 * | Page already *locked*
2843 * |- btrfs_lock_and_flush_ordered_range()
2844 * |- btrfs_start_ordered_extent()
2845 * |- extent_write_cache_pages()
2846 * |- lock_page()
2847 * We try to lock the page we already hold.
2848 *
2849 * Here we just writeback the whole data reloc inode, so that
2850 * we will be ensured to have no dirty range in the page, and
2851 * are safe to clear the uptodate bits.
2852 *
2853 * This shouldn't cause too much overhead, as we need to write
2854 * the data back anyway.
2855 */
2856 ret = filemap_write_and_wait(mapping);
2857 if (ret < 0)
2858 return ret;
2859
2860 clear_extent_bits(&inode->io_tree, i_size,
2861 round_up(i_size, PAGE_SIZE) - 1,
2862 EXTENT_UPTODATE);
2863 page = find_lock_page(mapping, i_size >> PAGE_SHIFT);
2864 /*
2865 * If page is freed we don't need to do anything then, as we
2866 * will re-read the whole page anyway.
2867 */
2868 if (page) {
2869 btrfs_subpage_clear_uptodate(fs_info, page, i_size,
2870 round_up(i_size, PAGE_SIZE) - i_size);
2871 unlock_page(page);
2872 put_page(page);
2873 }
2874 }
2875
2876 BUG_ON(cluster->start != cluster->boundary[0]);
2877 ret = btrfs_alloc_data_chunk_ondemand(inode,
2878 prealloc_end + 1 - prealloc_start);
2879 if (ret)
2880 return ret;
2881
2882 btrfs_inode_lock(inode, 0);
2883 for (nr = 0; nr < cluster->nr; nr++) {
2884 struct extent_state *cached_state = NULL;
2885
2886 start = cluster->boundary[nr] - offset;
2887 if (nr + 1 < cluster->nr)
2888 end = cluster->boundary[nr + 1] - 1 - offset;
2889 else
2890 end = cluster->end - offset;
2891
2892 lock_extent(&inode->io_tree, start, end, &cached_state);
2893 num_bytes = end + 1 - start;
2894 ret = btrfs_prealloc_file_range(&inode->vfs_inode, 0, start,
2895 num_bytes, num_bytes,
2896 end + 1, &alloc_hint);
2897 cur_offset = end + 1;
2898 unlock_extent(&inode->io_tree, start, end, &cached_state);
2899 if (ret)
2900 break;
2901 }
2902 btrfs_inode_unlock(inode, 0);
2903
2904 if (cur_offset < prealloc_end)
2905 btrfs_free_reserved_data_space_noquota(inode->root->fs_info,
2906 prealloc_end + 1 - cur_offset);
2907 return ret;
2908}
2909
2910static noinline_for_stack int setup_relocation_extent_mapping(struct inode *inode,
2911 u64 start, u64 end, u64 block_start)
2912{
2913 struct extent_map *em;
2914 struct extent_state *cached_state = NULL;
2915 int ret = 0;
2916
2917 em = alloc_extent_map();
2918 if (!em)
2919 return -ENOMEM;
2920
2921 em->start = start;
2922 em->len = end + 1 - start;
2923 em->block_len = em->len;
2924 em->block_start = block_start;
2925 set_bit(EXTENT_FLAG_PINNED, &em->flags);
2926
2927 lock_extent(&BTRFS_I(inode)->io_tree, start, end, &cached_state);
2928 ret = btrfs_replace_extent_map_range(BTRFS_I(inode), em, false);
2929 unlock_extent(&BTRFS_I(inode)->io_tree, start, end, &cached_state);
2930 free_extent_map(em);
2931
2932 return ret;
2933}
2934
2935/*
2936 * Allow error injection to test balance/relocation cancellation
2937 */
2938noinline int btrfs_should_cancel_balance(struct btrfs_fs_info *fs_info)
2939{
2940 return atomic_read(&fs_info->balance_cancel_req) ||
2941 atomic_read(&fs_info->reloc_cancel_req) ||
2942 fatal_signal_pending(current);
2943}
2944ALLOW_ERROR_INJECTION(btrfs_should_cancel_balance, TRUE);
2945
2946static u64 get_cluster_boundary_end(struct file_extent_cluster *cluster,
2947 int cluster_nr)
2948{
2949 /* Last extent, use cluster end directly */
2950 if (cluster_nr >= cluster->nr - 1)
2951 return cluster->end;
2952
2953 /* Use next boundary start*/
2954 return cluster->boundary[cluster_nr + 1] - 1;
2955}
2956
2957static int relocate_one_page(struct inode *inode, struct file_ra_state *ra,
2958 struct file_extent_cluster *cluster,
2959 int *cluster_nr, unsigned long page_index)
2960{
2961 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2962 u64 offset = BTRFS_I(inode)->index_cnt;
2963 const unsigned long last_index = (cluster->end - offset) >> PAGE_SHIFT;
2964 gfp_t mask = btrfs_alloc_write_mask(inode->i_mapping);
2965 struct page *page;
2966 u64 page_start;
2967 u64 page_end;
2968 u64 cur;
2969 int ret;
2970
2971 ASSERT(page_index <= last_index);
2972 page = find_lock_page(inode->i_mapping, page_index);
2973 if (!page) {
2974 page_cache_sync_readahead(inode->i_mapping, ra, NULL,
2975 page_index, last_index + 1 - page_index);
2976 page = find_or_create_page(inode->i_mapping, page_index, mask);
2977 if (!page)
2978 return -ENOMEM;
2979 }
2980 ret = set_page_extent_mapped(page);
2981 if (ret < 0)
2982 goto release_page;
2983
2984 if (PageReadahead(page))
2985 page_cache_async_readahead(inode->i_mapping, ra, NULL,
2986 page_folio(page), page_index,
2987 last_index + 1 - page_index);
2988
2989 if (!PageUptodate(page)) {
2990 btrfs_read_folio(NULL, page_folio(page));
2991 lock_page(page);
2992 if (!PageUptodate(page)) {
2993 ret = -EIO;
2994 goto release_page;
2995 }
2996 }
2997
2998 page_start = page_offset(page);
2999 page_end = page_start + PAGE_SIZE - 1;
3000
3001 /*
3002 * Start from the cluster, as for subpage case, the cluster can start
3003 * inside the page.
3004 */
3005 cur = max(page_start, cluster->boundary[*cluster_nr] - offset);
3006 while (cur <= page_end) {
3007 struct extent_state *cached_state = NULL;
3008 u64 extent_start = cluster->boundary[*cluster_nr] - offset;
3009 u64 extent_end = get_cluster_boundary_end(cluster,
3010 *cluster_nr) - offset;
3011 u64 clamped_start = max(page_start, extent_start);
3012 u64 clamped_end = min(page_end, extent_end);
3013 u32 clamped_len = clamped_end + 1 - clamped_start;
3014
3015 /* Reserve metadata for this range */
3016 ret = btrfs_delalloc_reserve_metadata(BTRFS_I(inode),
3017 clamped_len, clamped_len,
3018 false);
3019 if (ret)
3020 goto release_page;
3021
3022 /* Mark the range delalloc and dirty for later writeback */
3023 lock_extent(&BTRFS_I(inode)->io_tree, clamped_start, clamped_end,
3024 &cached_state);
3025 ret = btrfs_set_extent_delalloc(BTRFS_I(inode), clamped_start,
3026 clamped_end, 0, &cached_state);
3027 if (ret) {
3028 clear_extent_bit(&BTRFS_I(inode)->io_tree,
3029 clamped_start, clamped_end,
3030 EXTENT_LOCKED | EXTENT_BOUNDARY,
3031 &cached_state);
3032 btrfs_delalloc_release_metadata(BTRFS_I(inode),
3033 clamped_len, true);
3034 btrfs_delalloc_release_extents(BTRFS_I(inode),
3035 clamped_len);
3036 goto release_page;
3037 }
3038 btrfs_page_set_dirty(fs_info, page, clamped_start, clamped_len);
3039
3040 /*
3041 * Set the boundary if it's inside the page.
3042 * Data relocation requires the destination extents to have the
3043 * same size as the source.
3044 * EXTENT_BOUNDARY bit prevents current extent from being merged
3045 * with previous extent.
3046 */
3047 if (in_range(cluster->boundary[*cluster_nr] - offset,
3048 page_start, PAGE_SIZE)) {
3049 u64 boundary_start = cluster->boundary[*cluster_nr] -
3050 offset;
3051 u64 boundary_end = boundary_start +
3052 fs_info->sectorsize - 1;
3053
3054 set_extent_bits(&BTRFS_I(inode)->io_tree,
3055 boundary_start, boundary_end,
3056 EXTENT_BOUNDARY);
3057 }
3058 unlock_extent(&BTRFS_I(inode)->io_tree, clamped_start, clamped_end,
3059 &cached_state);
3060 btrfs_delalloc_release_extents(BTRFS_I(inode), clamped_len);
3061 cur += clamped_len;
3062
3063 /* Crossed extent end, go to next extent */
3064 if (cur >= extent_end) {
3065 (*cluster_nr)++;
3066 /* Just finished the last extent of the cluster, exit. */
3067 if (*cluster_nr >= cluster->nr)
3068 break;
3069 }
3070 }
3071 unlock_page(page);
3072 put_page(page);
3073
3074 balance_dirty_pages_ratelimited(inode->i_mapping);
3075 btrfs_throttle(fs_info);
3076 if (btrfs_should_cancel_balance(fs_info))
3077 ret = -ECANCELED;
3078 return ret;
3079
3080release_page:
3081 unlock_page(page);
3082 put_page(page);
3083 return ret;
3084}
3085
3086static int relocate_file_extent_cluster(struct inode *inode,
3087 struct file_extent_cluster *cluster)
3088{
3089 u64 offset = BTRFS_I(inode)->index_cnt;
3090 unsigned long index;
3091 unsigned long last_index;
3092 struct file_ra_state *ra;
3093 int cluster_nr = 0;
3094 int ret = 0;
3095
3096 if (!cluster->nr)
3097 return 0;
3098
3099 ra = kzalloc(sizeof(*ra), GFP_NOFS);
3100 if (!ra)
3101 return -ENOMEM;
3102
3103 ret = prealloc_file_extent_cluster(BTRFS_I(inode), cluster);
3104 if (ret)
3105 goto out;
3106
3107 file_ra_state_init(ra, inode->i_mapping);
3108
3109 ret = setup_relocation_extent_mapping(inode, cluster->start - offset,
3110 cluster->end - offset, cluster->start);
3111 if (ret)
3112 goto out;
3113
3114 last_index = (cluster->end - offset) >> PAGE_SHIFT;
3115 for (index = (cluster->start - offset) >> PAGE_SHIFT;
3116 index <= last_index && !ret; index++)
3117 ret = relocate_one_page(inode, ra, cluster, &cluster_nr, index);
3118 if (ret == 0)
3119 WARN_ON(cluster_nr != cluster->nr);
3120out:
3121 kfree(ra);
3122 return ret;
3123}
3124
3125static noinline_for_stack
3126int relocate_data_extent(struct inode *inode, struct btrfs_key *extent_key,
3127 struct file_extent_cluster *cluster)
3128{
3129 int ret;
3130
3131 if (cluster->nr > 0 && extent_key->objectid != cluster->end + 1) {
3132 ret = relocate_file_extent_cluster(inode, cluster);
3133 if (ret)
3134 return ret;
3135 cluster->nr = 0;
3136 }
3137
3138 if (!cluster->nr)
3139 cluster->start = extent_key->objectid;
3140 else
3141 BUG_ON(cluster->nr >= MAX_EXTENTS);
3142 cluster->end = extent_key->objectid + extent_key->offset - 1;
3143 cluster->boundary[cluster->nr] = extent_key->objectid;
3144 cluster->nr++;
3145
3146 if (cluster->nr >= MAX_EXTENTS) {
3147 ret = relocate_file_extent_cluster(inode, cluster);
3148 if (ret)
3149 return ret;
3150 cluster->nr = 0;
3151 }
3152 return 0;
3153}
3154
3155/*
3156 * helper to add a tree block to the list.
3157 * the major work is getting the generation and level of the block
3158 */
3159static int add_tree_block(struct reloc_control *rc,
3160 struct btrfs_key *extent_key,
3161 struct btrfs_path *path,
3162 struct rb_root *blocks)
3163{
3164 struct extent_buffer *eb;
3165 struct btrfs_extent_item *ei;
3166 struct btrfs_tree_block_info *bi;
3167 struct tree_block *block;
3168 struct rb_node *rb_node;
3169 u32 item_size;
3170 int level = -1;
3171 u64 generation;
3172 u64 owner = 0;
3173
3174 eb = path->nodes[0];
3175 item_size = btrfs_item_size(eb, path->slots[0]);
3176
3177 if (extent_key->type == BTRFS_METADATA_ITEM_KEY ||
3178 item_size >= sizeof(*ei) + sizeof(*bi)) {
3179 unsigned long ptr = 0, end;
3180
3181 ei = btrfs_item_ptr(eb, path->slots[0],
3182 struct btrfs_extent_item);
3183 end = (unsigned long)ei + item_size;
3184 if (extent_key->type == BTRFS_EXTENT_ITEM_KEY) {
3185 bi = (struct btrfs_tree_block_info *)(ei + 1);
3186 level = btrfs_tree_block_level(eb, bi);
3187 ptr = (unsigned long)(bi + 1);
3188 } else {
3189 level = (int)extent_key->offset;
3190 ptr = (unsigned long)(ei + 1);
3191 }
3192 generation = btrfs_extent_generation(eb, ei);
3193
3194 /*
3195 * We're reading random blocks without knowing their owner ahead
3196 * of time. This is ok most of the time, as all reloc roots and
3197 * fs roots have the same lock type. However normal trees do
3198 * not, and the only way to know ahead of time is to read the
3199 * inline ref offset. We know it's an fs root if
3200 *
3201 * 1. There's more than one ref.
3202 * 2. There's a SHARED_DATA_REF_KEY set.
3203 * 3. FULL_BACKREF is set on the flags.
3204 *
3205 * Otherwise it's safe to assume that the ref offset == the
3206 * owner of this block, so we can use that when calling
3207 * read_tree_block.
3208 */
3209 if (btrfs_extent_refs(eb, ei) == 1 &&
3210 !(btrfs_extent_flags(eb, ei) &
3211 BTRFS_BLOCK_FLAG_FULL_BACKREF) &&
3212 ptr < end) {
3213 struct btrfs_extent_inline_ref *iref;
3214 int type;
3215
3216 iref = (struct btrfs_extent_inline_ref *)ptr;
3217 type = btrfs_get_extent_inline_ref_type(eb, iref,
3218 BTRFS_REF_TYPE_BLOCK);
3219 if (type == BTRFS_REF_TYPE_INVALID)
3220 return -EINVAL;
3221 if (type == BTRFS_TREE_BLOCK_REF_KEY)
3222 owner = btrfs_extent_inline_ref_offset(eb, iref);
3223 }
3224 } else if (unlikely(item_size == sizeof(struct btrfs_extent_item_v0))) {
3225 btrfs_print_v0_err(eb->fs_info);
3226 btrfs_handle_fs_error(eb->fs_info, -EINVAL, NULL);
3227 return -EINVAL;
3228 } else {
3229 BUG();
3230 }
3231
3232 btrfs_release_path(path);
3233
3234 BUG_ON(level == -1);
3235
3236 block = kmalloc(sizeof(*block), GFP_NOFS);
3237 if (!block)
3238 return -ENOMEM;
3239
3240 block->bytenr = extent_key->objectid;
3241 block->key.objectid = rc->extent_root->fs_info->nodesize;
3242 block->key.offset = generation;
3243 block->level = level;
3244 block->key_ready = 0;
3245 block->owner = owner;
3246
3247 rb_node = rb_simple_insert(blocks, block->bytenr, &block->rb_node);
3248 if (rb_node)
3249 btrfs_backref_panic(rc->extent_root->fs_info, block->bytenr,
3250 -EEXIST);
3251
3252 return 0;
3253}
3254
3255/*
3256 * helper to add tree blocks for backref of type BTRFS_SHARED_DATA_REF_KEY
3257 */
3258static int __add_tree_block(struct reloc_control *rc,
3259 u64 bytenr, u32 blocksize,
3260 struct rb_root *blocks)
3261{
3262 struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
3263 struct btrfs_path *path;
3264 struct btrfs_key key;
3265 int ret;
3266 bool skinny = btrfs_fs_incompat(fs_info, SKINNY_METADATA);
3267
3268 if (tree_block_processed(bytenr, rc))
3269 return 0;
3270
3271 if (rb_simple_search(blocks, bytenr))
3272 return 0;
3273
3274 path = btrfs_alloc_path();
3275 if (!path)
3276 return -ENOMEM;
3277again:
3278 key.objectid = bytenr;
3279 if (skinny) {
3280 key.type = BTRFS_METADATA_ITEM_KEY;
3281 key.offset = (u64)-1;
3282 } else {
3283 key.type = BTRFS_EXTENT_ITEM_KEY;
3284 key.offset = blocksize;
3285 }
3286
3287 path->search_commit_root = 1;
3288 path->skip_locking = 1;
3289 ret = btrfs_search_slot(NULL, rc->extent_root, &key, path, 0, 0);
3290 if (ret < 0)
3291 goto out;
3292
3293 if (ret > 0 && skinny) {
3294 if (path->slots[0]) {
3295 path->slots[0]--;
3296 btrfs_item_key_to_cpu(path->nodes[0], &key,
3297 path->slots[0]);
3298 if (key.objectid == bytenr &&
3299 (key.type == BTRFS_METADATA_ITEM_KEY ||
3300 (key.type == BTRFS_EXTENT_ITEM_KEY &&
3301 key.offset == blocksize)))
3302 ret = 0;
3303 }
3304
3305 if (ret) {
3306 skinny = false;
3307 btrfs_release_path(path);
3308 goto again;
3309 }
3310 }
3311 if (ret) {
3312 ASSERT(ret == 1);
3313 btrfs_print_leaf(path->nodes[0]);
3314 btrfs_err(fs_info,
3315 "tree block extent item (%llu) is not found in extent tree",
3316 bytenr);
3317 WARN_ON(1);
3318 ret = -EINVAL;
3319 goto out;
3320 }
3321
3322 ret = add_tree_block(rc, &key, path, blocks);
3323out:
3324 btrfs_free_path(path);
3325 return ret;
3326}
3327
3328static int delete_block_group_cache(struct btrfs_fs_info *fs_info,
3329 struct btrfs_block_group *block_group,
3330 struct inode *inode,
3331 u64 ino)
3332{
3333 struct btrfs_root *root = fs_info->tree_root;
3334 struct btrfs_trans_handle *trans;
3335 int ret = 0;
3336
3337 if (inode)
3338 goto truncate;
3339
3340 inode = btrfs_iget(fs_info->sb, ino, root);
3341 if (IS_ERR(inode))
3342 return -ENOENT;
3343
3344truncate:
3345 ret = btrfs_check_trunc_cache_free_space(fs_info,
3346 &fs_info->global_block_rsv);
3347 if (ret)
3348 goto out;
3349
3350 trans = btrfs_join_transaction(root);
3351 if (IS_ERR(trans)) {
3352 ret = PTR_ERR(trans);
3353 goto out;
3354 }
3355
3356 ret = btrfs_truncate_free_space_cache(trans, block_group, inode);
3357
3358 btrfs_end_transaction(trans);
3359 btrfs_btree_balance_dirty(fs_info);
3360out:
3361 iput(inode);
3362 return ret;
3363}
3364
3365/*
3366 * Locate the free space cache EXTENT_DATA in root tree leaf and delete the
3367 * cache inode, to avoid free space cache data extent blocking data relocation.
3368 */
3369static int delete_v1_space_cache(struct extent_buffer *leaf,
3370 struct btrfs_block_group *block_group,
3371 u64 data_bytenr)
3372{
3373 u64 space_cache_ino;
3374 struct btrfs_file_extent_item *ei;
3375 struct btrfs_key key;
3376 bool found = false;
3377 int i;
3378 int ret;
3379
3380 if (btrfs_header_owner(leaf) != BTRFS_ROOT_TREE_OBJECTID)
3381 return 0;
3382
3383 for (i = 0; i < btrfs_header_nritems(leaf); i++) {
3384 u8 type;
3385
3386 btrfs_item_key_to_cpu(leaf, &key, i);
3387 if (key.type != BTRFS_EXTENT_DATA_KEY)
3388 continue;
3389 ei = btrfs_item_ptr(leaf, i, struct btrfs_file_extent_item);
3390 type = btrfs_file_extent_type(leaf, ei);
3391
3392 if ((type == BTRFS_FILE_EXTENT_REG ||
3393 type == BTRFS_FILE_EXTENT_PREALLOC) &&
3394 btrfs_file_extent_disk_bytenr(leaf, ei) == data_bytenr) {
3395 found = true;
3396 space_cache_ino = key.objectid;
3397 break;
3398 }
3399 }
3400 if (!found)
3401 return -ENOENT;
3402 ret = delete_block_group_cache(leaf->fs_info, block_group, NULL,
3403 space_cache_ino);
3404 return ret;
3405}
3406
3407/*
3408 * helper to find all tree blocks that reference a given data extent
3409 */
3410static noinline_for_stack
3411int add_data_references(struct reloc_control *rc,
3412 struct btrfs_key *extent_key,
3413 struct btrfs_path *path,
3414 struct rb_root *blocks)
3415{
3416 struct btrfs_backref_walk_ctx ctx = { 0 };
3417 struct ulist_iterator leaf_uiter;
3418 struct ulist_node *ref_node = NULL;
3419 const u32 blocksize = rc->extent_root->fs_info->nodesize;
3420 int ret = 0;
3421
3422 btrfs_release_path(path);
3423
3424 ctx.bytenr = extent_key->objectid;
3425 ctx.ignore_extent_item_pos = true;
3426 ctx.fs_info = rc->extent_root->fs_info;
3427
3428 ret = btrfs_find_all_leafs(&ctx);
3429 if (ret < 0)
3430 return ret;
3431
3432 ULIST_ITER_INIT(&leaf_uiter);
3433 while ((ref_node = ulist_next(ctx.refs, &leaf_uiter))) {
3434 struct btrfs_tree_parent_check check = { 0 };
3435 struct extent_buffer *eb;
3436
3437 eb = read_tree_block(ctx.fs_info, ref_node->val, &check);
3438 if (IS_ERR(eb)) {
3439 ret = PTR_ERR(eb);
3440 break;
3441 }
3442 ret = delete_v1_space_cache(eb, rc->block_group,
3443 extent_key->objectid);
3444 free_extent_buffer(eb);
3445 if (ret < 0)
3446 break;
3447 ret = __add_tree_block(rc, ref_node->val, blocksize, blocks);
3448 if (ret < 0)
3449 break;
3450 }
3451 if (ret < 0)
3452 free_block_list(blocks);
3453 ulist_free(ctx.refs);
3454 return ret;
3455}
3456
3457/*
3458 * helper to find next unprocessed extent
3459 */
3460static noinline_for_stack
3461int find_next_extent(struct reloc_control *rc, struct btrfs_path *path,
3462 struct btrfs_key *extent_key)
3463{
3464 struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
3465 struct btrfs_key key;
3466 struct extent_buffer *leaf;
3467 u64 start, end, last;
3468 int ret;
3469
3470 last = rc->block_group->start + rc->block_group->length;
3471 while (1) {
3472 cond_resched();
3473 if (rc->search_start >= last) {
3474 ret = 1;
3475 break;
3476 }
3477
3478 key.objectid = rc->search_start;
3479 key.type = BTRFS_EXTENT_ITEM_KEY;
3480 key.offset = 0;
3481
3482 path->search_commit_root = 1;
3483 path->skip_locking = 1;
3484 ret = btrfs_search_slot(NULL, rc->extent_root, &key, path,
3485 0, 0);
3486 if (ret < 0)
3487 break;
3488next:
3489 leaf = path->nodes[0];
3490 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
3491 ret = btrfs_next_leaf(rc->extent_root, path);
3492 if (ret != 0)
3493 break;
3494 leaf = path->nodes[0];
3495 }
3496
3497 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
3498 if (key.objectid >= last) {
3499 ret = 1;
3500 break;
3501 }
3502
3503 if (key.type != BTRFS_EXTENT_ITEM_KEY &&
3504 key.type != BTRFS_METADATA_ITEM_KEY) {
3505 path->slots[0]++;
3506 goto next;
3507 }
3508
3509 if (key.type == BTRFS_EXTENT_ITEM_KEY &&
3510 key.objectid + key.offset <= rc->search_start) {
3511 path->slots[0]++;
3512 goto next;
3513 }
3514
3515 if (key.type == BTRFS_METADATA_ITEM_KEY &&
3516 key.objectid + fs_info->nodesize <=
3517 rc->search_start) {
3518 path->slots[0]++;
3519 goto next;
3520 }
3521
3522 ret = find_first_extent_bit(&rc->processed_blocks,
3523 key.objectid, &start, &end,
3524 EXTENT_DIRTY, NULL);
3525
3526 if (ret == 0 && start <= key.objectid) {
3527 btrfs_release_path(path);
3528 rc->search_start = end + 1;
3529 } else {
3530 if (key.type == BTRFS_EXTENT_ITEM_KEY)
3531 rc->search_start = key.objectid + key.offset;
3532 else
3533 rc->search_start = key.objectid +
3534 fs_info->nodesize;
3535 memcpy(extent_key, &key, sizeof(key));
3536 return 0;
3537 }
3538 }
3539 btrfs_release_path(path);
3540 return ret;
3541}
3542
3543static void set_reloc_control(struct reloc_control *rc)
3544{
3545 struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
3546
3547 mutex_lock(&fs_info->reloc_mutex);
3548 fs_info->reloc_ctl = rc;
3549 mutex_unlock(&fs_info->reloc_mutex);
3550}
3551
3552static void unset_reloc_control(struct reloc_control *rc)
3553{
3554 struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
3555
3556 mutex_lock(&fs_info->reloc_mutex);
3557 fs_info->reloc_ctl = NULL;
3558 mutex_unlock(&fs_info->reloc_mutex);
3559}
3560
3561static noinline_for_stack
3562int prepare_to_relocate(struct reloc_control *rc)
3563{
3564 struct btrfs_trans_handle *trans;
3565 int ret;
3566
3567 rc->block_rsv = btrfs_alloc_block_rsv(rc->extent_root->fs_info,
3568 BTRFS_BLOCK_RSV_TEMP);
3569 if (!rc->block_rsv)
3570 return -ENOMEM;
3571
3572 memset(&rc->cluster, 0, sizeof(rc->cluster));
3573 rc->search_start = rc->block_group->start;
3574 rc->extents_found = 0;
3575 rc->nodes_relocated = 0;
3576 rc->merging_rsv_size = 0;
3577 rc->reserved_bytes = 0;
3578 rc->block_rsv->size = rc->extent_root->fs_info->nodesize *
3579 RELOCATION_RESERVED_NODES;
3580 ret = btrfs_block_rsv_refill(rc->extent_root->fs_info,
3581 rc->block_rsv, rc->block_rsv->size,
3582 BTRFS_RESERVE_FLUSH_ALL);
3583 if (ret)
3584 return ret;
3585
3586 rc->create_reloc_tree = 1;
3587 set_reloc_control(rc);
3588
3589 trans = btrfs_join_transaction(rc->extent_root);
3590 if (IS_ERR(trans)) {
3591 unset_reloc_control(rc);
3592 /*
3593 * extent tree is not a ref_cow tree and has no reloc_root to
3594 * cleanup. And callers are responsible to free the above
3595 * block rsv.
3596 */
3597 return PTR_ERR(trans);
3598 }
3599
3600 ret = btrfs_commit_transaction(trans);
3601 if (ret)
3602 unset_reloc_control(rc);
3603
3604 return ret;
3605}
3606
3607static noinline_for_stack int relocate_block_group(struct reloc_control *rc)
3608{
3609 struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
3610 struct rb_root blocks = RB_ROOT;
3611 struct btrfs_key key;
3612 struct btrfs_trans_handle *trans = NULL;
3613 struct btrfs_path *path;
3614 struct btrfs_extent_item *ei;
3615 u64 flags;
3616 int ret;
3617 int err = 0;
3618 int progress = 0;
3619
3620 path = btrfs_alloc_path();
3621 if (!path)
3622 return -ENOMEM;
3623 path->reada = READA_FORWARD;
3624
3625 ret = prepare_to_relocate(rc);
3626 if (ret) {
3627 err = ret;
3628 goto out_free;
3629 }
3630
3631 while (1) {
3632 rc->reserved_bytes = 0;
3633 ret = btrfs_block_rsv_refill(fs_info, rc->block_rsv,
3634 rc->block_rsv->size,
3635 BTRFS_RESERVE_FLUSH_ALL);
3636 if (ret) {
3637 err = ret;
3638 break;
3639 }
3640 progress++;
3641 trans = btrfs_start_transaction(rc->extent_root, 0);
3642 if (IS_ERR(trans)) {
3643 err = PTR_ERR(trans);
3644 trans = NULL;
3645 break;
3646 }
3647restart:
3648 if (update_backref_cache(trans, &rc->backref_cache)) {
3649 btrfs_end_transaction(trans);
3650 trans = NULL;
3651 continue;
3652 }
3653
3654 ret = find_next_extent(rc, path, &key);
3655 if (ret < 0)
3656 err = ret;
3657 if (ret != 0)
3658 break;
3659
3660 rc->extents_found++;
3661
3662 ei = btrfs_item_ptr(path->nodes[0], path->slots[0],
3663 struct btrfs_extent_item);
3664 flags = btrfs_extent_flags(path->nodes[0], ei);
3665
3666 if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
3667 ret = add_tree_block(rc, &key, path, &blocks);
3668 } else if (rc->stage == UPDATE_DATA_PTRS &&
3669 (flags & BTRFS_EXTENT_FLAG_DATA)) {
3670 ret = add_data_references(rc, &key, path, &blocks);
3671 } else {
3672 btrfs_release_path(path);
3673 ret = 0;
3674 }
3675 if (ret < 0) {
3676 err = ret;
3677 break;
3678 }
3679
3680 if (!RB_EMPTY_ROOT(&blocks)) {
3681 ret = relocate_tree_blocks(trans, rc, &blocks);
3682 if (ret < 0) {
3683 if (ret != -EAGAIN) {
3684 err = ret;
3685 break;
3686 }
3687 rc->extents_found--;
3688 rc->search_start = key.objectid;
3689 }
3690 }
3691
3692 btrfs_end_transaction_throttle(trans);
3693 btrfs_btree_balance_dirty(fs_info);
3694 trans = NULL;
3695
3696 if (rc->stage == MOVE_DATA_EXTENTS &&
3697 (flags & BTRFS_EXTENT_FLAG_DATA)) {
3698 rc->found_file_extent = 1;
3699 ret = relocate_data_extent(rc->data_inode,
3700 &key, &rc->cluster);
3701 if (ret < 0) {
3702 err = ret;
3703 break;
3704 }
3705 }
3706 if (btrfs_should_cancel_balance(fs_info)) {
3707 err = -ECANCELED;
3708 break;
3709 }
3710 }
3711 if (trans && progress && err == -ENOSPC) {
3712 ret = btrfs_force_chunk_alloc(trans, rc->block_group->flags);
3713 if (ret == 1) {
3714 err = 0;
3715 progress = 0;
3716 goto restart;
3717 }
3718 }
3719
3720 btrfs_release_path(path);
3721 clear_extent_bits(&rc->processed_blocks, 0, (u64)-1, EXTENT_DIRTY);
3722
3723 if (trans) {
3724 btrfs_end_transaction_throttle(trans);
3725 btrfs_btree_balance_dirty(fs_info);
3726 }
3727
3728 if (!err) {
3729 ret = relocate_file_extent_cluster(rc->data_inode,
3730 &rc->cluster);
3731 if (ret < 0)
3732 err = ret;
3733 }
3734
3735 rc->create_reloc_tree = 0;
3736 set_reloc_control(rc);
3737
3738 btrfs_backref_release_cache(&rc->backref_cache);
3739 btrfs_block_rsv_release(fs_info, rc->block_rsv, (u64)-1, NULL);
3740
3741 /*
3742 * Even in the case when the relocation is cancelled, we should all go
3743 * through prepare_to_merge() and merge_reloc_roots().
3744 *
3745 * For error (including cancelled balance), prepare_to_merge() will
3746 * mark all reloc trees orphan, then queue them for cleanup in
3747 * merge_reloc_roots()
3748 */
3749 err = prepare_to_merge(rc, err);
3750
3751 merge_reloc_roots(rc);
3752
3753 rc->merge_reloc_tree = 0;
3754 unset_reloc_control(rc);
3755 btrfs_block_rsv_release(fs_info, rc->block_rsv, (u64)-1, NULL);
3756
3757 /* get rid of pinned extents */
3758 trans = btrfs_join_transaction(rc->extent_root);
3759 if (IS_ERR(trans)) {
3760 err = PTR_ERR(trans);
3761 goto out_free;
3762 }
3763 ret = btrfs_commit_transaction(trans);
3764 if (ret && !err)
3765 err = ret;
3766out_free:
3767 ret = clean_dirty_subvols(rc);
3768 if (ret < 0 && !err)
3769 err = ret;
3770 btrfs_free_block_rsv(fs_info, rc->block_rsv);
3771 btrfs_free_path(path);
3772 return err;
3773}
3774
3775static int __insert_orphan_inode(struct btrfs_trans_handle *trans,
3776 struct btrfs_root *root, u64 objectid)
3777{
3778 struct btrfs_path *path;
3779 struct btrfs_inode_item *item;
3780 struct extent_buffer *leaf;
3781 int ret;
3782
3783 path = btrfs_alloc_path();
3784 if (!path)
3785 return -ENOMEM;
3786
3787 ret = btrfs_insert_empty_inode(trans, root, path, objectid);
3788 if (ret)
3789 goto out;
3790
3791 leaf = path->nodes[0];
3792 item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_inode_item);
3793 memzero_extent_buffer(leaf, (unsigned long)item, sizeof(*item));
3794 btrfs_set_inode_generation(leaf, item, 1);
3795 btrfs_set_inode_size(leaf, item, 0);
3796 btrfs_set_inode_mode(leaf, item, S_IFREG | 0600);
3797 btrfs_set_inode_flags(leaf, item, BTRFS_INODE_NOCOMPRESS |
3798 BTRFS_INODE_PREALLOC);
3799 btrfs_mark_buffer_dirty(leaf);
3800out:
3801 btrfs_free_path(path);
3802 return ret;
3803}
3804
3805static void delete_orphan_inode(struct btrfs_trans_handle *trans,
3806 struct btrfs_root *root, u64 objectid)
3807{
3808 struct btrfs_path *path;
3809 struct btrfs_key key;
3810 int ret = 0;
3811
3812 path = btrfs_alloc_path();
3813 if (!path) {
3814 ret = -ENOMEM;
3815 goto out;
3816 }
3817
3818 key.objectid = objectid;
3819 key.type = BTRFS_INODE_ITEM_KEY;
3820 key.offset = 0;
3821 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
3822 if (ret) {
3823 if (ret > 0)
3824 ret = -ENOENT;
3825 goto out;
3826 }
3827 ret = btrfs_del_item(trans, root, path);
3828out:
3829 if (ret)
3830 btrfs_abort_transaction(trans, ret);
3831 btrfs_free_path(path);
3832}
3833
3834/*
3835 * helper to create inode for data relocation.
3836 * the inode is in data relocation tree and its link count is 0
3837 */
3838static noinline_for_stack
3839struct inode *create_reloc_inode(struct btrfs_fs_info *fs_info,
3840 struct btrfs_block_group *group)
3841{
3842 struct inode *inode = NULL;
3843 struct btrfs_trans_handle *trans;
3844 struct btrfs_root *root;
3845 u64 objectid;
3846 int err = 0;
3847
3848 root = btrfs_grab_root(fs_info->data_reloc_root);
3849 trans = btrfs_start_transaction(root, 6);
3850 if (IS_ERR(trans)) {
3851 btrfs_put_root(root);
3852 return ERR_CAST(trans);
3853 }
3854
3855 err = btrfs_get_free_objectid(root, &objectid);
3856 if (err)
3857 goto out;
3858
3859 err = __insert_orphan_inode(trans, root, objectid);
3860 if (err)
3861 goto out;
3862
3863 inode = btrfs_iget(fs_info->sb, objectid, root);
3864 if (IS_ERR(inode)) {
3865 delete_orphan_inode(trans, root, objectid);
3866 err = PTR_ERR(inode);
3867 inode = NULL;
3868 goto out;
3869 }
3870 BTRFS_I(inode)->index_cnt = group->start;
3871
3872 err = btrfs_orphan_add(trans, BTRFS_I(inode));
3873out:
3874 btrfs_put_root(root);
3875 btrfs_end_transaction(trans);
3876 btrfs_btree_balance_dirty(fs_info);
3877 if (err) {
3878 iput(inode);
3879 inode = ERR_PTR(err);
3880 }
3881 return inode;
3882}
3883
3884/*
3885 * Mark start of chunk relocation that is cancellable. Check if the cancellation
3886 * has been requested meanwhile and don't start in that case.
3887 *
3888 * Return:
3889 * 0 success
3890 * -EINPROGRESS operation is already in progress, that's probably a bug
3891 * -ECANCELED cancellation request was set before the operation started
3892 */
3893static int reloc_chunk_start(struct btrfs_fs_info *fs_info)
3894{
3895 if (test_and_set_bit(BTRFS_FS_RELOC_RUNNING, &fs_info->flags)) {
3896 /* This should not happen */
3897 btrfs_err(fs_info, "reloc already running, cannot start");
3898 return -EINPROGRESS;
3899 }
3900
3901 if (atomic_read(&fs_info->reloc_cancel_req) > 0) {
3902 btrfs_info(fs_info, "chunk relocation canceled on start");
3903 /*
3904 * On cancel, clear all requests but let the caller mark
3905 * the end after cleanup operations.
3906 */
3907 atomic_set(&fs_info->reloc_cancel_req, 0);
3908 return -ECANCELED;
3909 }
3910 return 0;
3911}
3912
3913/*
3914 * Mark end of chunk relocation that is cancellable and wake any waiters.
3915 */
3916static void reloc_chunk_end(struct btrfs_fs_info *fs_info)
3917{
3918 /* Requested after start, clear bit first so any waiters can continue */
3919 if (atomic_read(&fs_info->reloc_cancel_req) > 0)
3920 btrfs_info(fs_info, "chunk relocation canceled during operation");
3921 clear_and_wake_up_bit(BTRFS_FS_RELOC_RUNNING, &fs_info->flags);
3922 atomic_set(&fs_info->reloc_cancel_req, 0);
3923}
3924
3925static struct reloc_control *alloc_reloc_control(struct btrfs_fs_info *fs_info)
3926{
3927 struct reloc_control *rc;
3928
3929 rc = kzalloc(sizeof(*rc), GFP_NOFS);
3930 if (!rc)
3931 return NULL;
3932
3933 INIT_LIST_HEAD(&rc->reloc_roots);
3934 INIT_LIST_HEAD(&rc->dirty_subvol_roots);
3935 btrfs_backref_init_cache(fs_info, &rc->backref_cache, 1);
3936 mapping_tree_init(&rc->reloc_root_tree);
3937 extent_io_tree_init(fs_info, &rc->processed_blocks, IO_TREE_RELOC_BLOCKS);
3938 return rc;
3939}
3940
3941static void free_reloc_control(struct reloc_control *rc)
3942{
3943 struct mapping_node *node, *tmp;
3944
3945 free_reloc_roots(&rc->reloc_roots);
3946 rbtree_postorder_for_each_entry_safe(node, tmp,
3947 &rc->reloc_root_tree.rb_root, rb_node)
3948 kfree(node);
3949
3950 kfree(rc);
3951}
3952
3953/*
3954 * Print the block group being relocated
3955 */
3956static void describe_relocation(struct btrfs_fs_info *fs_info,
3957 struct btrfs_block_group *block_group)
3958{
3959 char buf[128] = {'\0'};
3960
3961 btrfs_describe_block_groups(block_group->flags, buf, sizeof(buf));
3962
3963 btrfs_info(fs_info,
3964 "relocating block group %llu flags %s",
3965 block_group->start, buf);
3966}
3967
3968static const char *stage_to_string(int stage)
3969{
3970 if (stage == MOVE_DATA_EXTENTS)
3971 return "move data extents";
3972 if (stage == UPDATE_DATA_PTRS)
3973 return "update data pointers";
3974 return "unknown";
3975}
3976
3977/*
3978 * function to relocate all extents in a block group.
3979 */
3980int btrfs_relocate_block_group(struct btrfs_fs_info *fs_info, u64 group_start)
3981{
3982 struct btrfs_block_group *bg;
3983 struct btrfs_root *extent_root = btrfs_extent_root(fs_info, group_start);
3984 struct reloc_control *rc;
3985 struct inode *inode;
3986 struct btrfs_path *path;
3987 int ret;
3988 int rw = 0;
3989 int err = 0;
3990
3991 /*
3992 * This only gets set if we had a half-deleted snapshot on mount. We
3993 * cannot allow relocation to start while we're still trying to clean up
3994 * these pending deletions.
3995 */
3996 ret = wait_on_bit(&fs_info->flags, BTRFS_FS_UNFINISHED_DROPS, TASK_INTERRUPTIBLE);
3997 if (ret)
3998 return ret;
3999
4000 /* We may have been woken up by close_ctree, so bail if we're closing. */
4001 if (btrfs_fs_closing(fs_info))
4002 return -EINTR;
4003
4004 bg = btrfs_lookup_block_group(fs_info, group_start);
4005 if (!bg)
4006 return -ENOENT;
4007
4008 /*
4009 * Relocation of a data block group creates ordered extents. Without
4010 * sb_start_write(), we can freeze the filesystem while unfinished
4011 * ordered extents are left. Such ordered extents can cause a deadlock
4012 * e.g. when syncfs() is waiting for their completion but they can't
4013 * finish because they block when joining a transaction, due to the
4014 * fact that the freeze locks are being held in write mode.
4015 */
4016 if (bg->flags & BTRFS_BLOCK_GROUP_DATA)
4017 ASSERT(sb_write_started(fs_info->sb));
4018
4019 if (btrfs_pinned_by_swapfile(fs_info, bg)) {
4020 btrfs_put_block_group(bg);
4021 return -ETXTBSY;
4022 }
4023
4024 rc = alloc_reloc_control(fs_info);
4025 if (!rc) {
4026 btrfs_put_block_group(bg);
4027 return -ENOMEM;
4028 }
4029
4030 ret = reloc_chunk_start(fs_info);
4031 if (ret < 0) {
4032 err = ret;
4033 goto out_put_bg;
4034 }
4035
4036 rc->extent_root = extent_root;
4037 rc->block_group = bg;
4038
4039 ret = btrfs_inc_block_group_ro(rc->block_group, true);
4040 if (ret) {
4041 err = ret;
4042 goto out;
4043 }
4044 rw = 1;
4045
4046 path = btrfs_alloc_path();
4047 if (!path) {
4048 err = -ENOMEM;
4049 goto out;
4050 }
4051
4052 inode = lookup_free_space_inode(rc->block_group, path);
4053 btrfs_free_path(path);
4054
4055 if (!IS_ERR(inode))
4056 ret = delete_block_group_cache(fs_info, rc->block_group, inode, 0);
4057 else
4058 ret = PTR_ERR(inode);
4059
4060 if (ret && ret != -ENOENT) {
4061 err = ret;
4062 goto out;
4063 }
4064
4065 rc->data_inode = create_reloc_inode(fs_info, rc->block_group);
4066 if (IS_ERR(rc->data_inode)) {
4067 err = PTR_ERR(rc->data_inode);
4068 rc->data_inode = NULL;
4069 goto out;
4070 }
4071
4072 describe_relocation(fs_info, rc->block_group);
4073
4074 btrfs_wait_block_group_reservations(rc->block_group);
4075 btrfs_wait_nocow_writers(rc->block_group);
4076 btrfs_wait_ordered_roots(fs_info, U64_MAX,
4077 rc->block_group->start,
4078 rc->block_group->length);
4079
4080 ret = btrfs_zone_finish(rc->block_group);
4081 WARN_ON(ret && ret != -EAGAIN);
4082
4083 while (1) {
4084 int finishes_stage;
4085
4086 mutex_lock(&fs_info->cleaner_mutex);
4087 ret = relocate_block_group(rc);
4088 mutex_unlock(&fs_info->cleaner_mutex);
4089 if (ret < 0)
4090 err = ret;
4091
4092 finishes_stage = rc->stage;
4093 /*
4094 * We may have gotten ENOSPC after we already dirtied some
4095 * extents. If writeout happens while we're relocating a
4096 * different block group we could end up hitting the
4097 * BUG_ON(rc->stage == UPDATE_DATA_PTRS) in
4098 * btrfs_reloc_cow_block. Make sure we write everything out
4099 * properly so we don't trip over this problem, and then break
4100 * out of the loop if we hit an error.
4101 */
4102 if (rc->stage == MOVE_DATA_EXTENTS && rc->found_file_extent) {
4103 ret = btrfs_wait_ordered_range(rc->data_inode, 0,
4104 (u64)-1);
4105 if (ret)
4106 err = ret;
4107 invalidate_mapping_pages(rc->data_inode->i_mapping,
4108 0, -1);
4109 rc->stage = UPDATE_DATA_PTRS;
4110 }
4111
4112 if (err < 0)
4113 goto out;
4114
4115 if (rc->extents_found == 0)
4116 break;
4117
4118 btrfs_info(fs_info, "found %llu extents, stage: %s",
4119 rc->extents_found, stage_to_string(finishes_stage));
4120 }
4121
4122 WARN_ON(rc->block_group->pinned > 0);
4123 WARN_ON(rc->block_group->reserved > 0);
4124 WARN_ON(rc->block_group->used > 0);
4125out:
4126 if (err && rw)
4127 btrfs_dec_block_group_ro(rc->block_group);
4128 iput(rc->data_inode);
4129out_put_bg:
4130 btrfs_put_block_group(bg);
4131 reloc_chunk_end(fs_info);
4132 free_reloc_control(rc);
4133 return err;
4134}
4135
4136static noinline_for_stack int mark_garbage_root(struct btrfs_root *root)
4137{
4138 struct btrfs_fs_info *fs_info = root->fs_info;
4139 struct btrfs_trans_handle *trans;
4140 int ret, err;
4141
4142 trans = btrfs_start_transaction(fs_info->tree_root, 0);
4143 if (IS_ERR(trans))
4144 return PTR_ERR(trans);
4145
4146 memset(&root->root_item.drop_progress, 0,
4147 sizeof(root->root_item.drop_progress));
4148 btrfs_set_root_drop_level(&root->root_item, 0);
4149 btrfs_set_root_refs(&root->root_item, 0);
4150 ret = btrfs_update_root(trans, fs_info->tree_root,
4151 &root->root_key, &root->root_item);
4152
4153 err = btrfs_end_transaction(trans);
4154 if (err)
4155 return err;
4156 return ret;
4157}
4158
4159/*
4160 * recover relocation interrupted by system crash.
4161 *
4162 * this function resumes merging reloc trees with corresponding fs trees.
4163 * this is important for keeping the sharing of tree blocks
4164 */
4165int btrfs_recover_relocation(struct btrfs_fs_info *fs_info)
4166{
4167 LIST_HEAD(reloc_roots);
4168 struct btrfs_key key;
4169 struct btrfs_root *fs_root;
4170 struct btrfs_root *reloc_root;
4171 struct btrfs_path *path;
4172 struct extent_buffer *leaf;
4173 struct reloc_control *rc = NULL;
4174 struct btrfs_trans_handle *trans;
4175 int ret;
4176 int err = 0;
4177
4178 path = btrfs_alloc_path();
4179 if (!path)
4180 return -ENOMEM;
4181 path->reada = READA_BACK;
4182
4183 key.objectid = BTRFS_TREE_RELOC_OBJECTID;
4184 key.type = BTRFS_ROOT_ITEM_KEY;
4185 key.offset = (u64)-1;
4186
4187 while (1) {
4188 ret = btrfs_search_slot(NULL, fs_info->tree_root, &key,
4189 path, 0, 0);
4190 if (ret < 0) {
4191 err = ret;
4192 goto out;
4193 }
4194 if (ret > 0) {
4195 if (path->slots[0] == 0)
4196 break;
4197 path->slots[0]--;
4198 }
4199 leaf = path->nodes[0];
4200 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
4201 btrfs_release_path(path);
4202
4203 if (key.objectid != BTRFS_TREE_RELOC_OBJECTID ||
4204 key.type != BTRFS_ROOT_ITEM_KEY)
4205 break;
4206
4207 reloc_root = btrfs_read_tree_root(fs_info->tree_root, &key);
4208 if (IS_ERR(reloc_root)) {
4209 err = PTR_ERR(reloc_root);
4210 goto out;
4211 }
4212
4213 set_bit(BTRFS_ROOT_SHAREABLE, &reloc_root->state);
4214 list_add(&reloc_root->root_list, &reloc_roots);
4215
4216 if (btrfs_root_refs(&reloc_root->root_item) > 0) {
4217 fs_root = btrfs_get_fs_root(fs_info,
4218 reloc_root->root_key.offset, false);
4219 if (IS_ERR(fs_root)) {
4220 ret = PTR_ERR(fs_root);
4221 if (ret != -ENOENT) {
4222 err = ret;
4223 goto out;
4224 }
4225 ret = mark_garbage_root(reloc_root);
4226 if (ret < 0) {
4227 err = ret;
4228 goto out;
4229 }
4230 } else {
4231 btrfs_put_root(fs_root);
4232 }
4233 }
4234
4235 if (key.offset == 0)
4236 break;
4237
4238 key.offset--;
4239 }
4240 btrfs_release_path(path);
4241
4242 if (list_empty(&reloc_roots))
4243 goto out;
4244
4245 rc = alloc_reloc_control(fs_info);
4246 if (!rc) {
4247 err = -ENOMEM;
4248 goto out;
4249 }
4250
4251 ret = reloc_chunk_start(fs_info);
4252 if (ret < 0) {
4253 err = ret;
4254 goto out_end;
4255 }
4256
4257 rc->extent_root = btrfs_extent_root(fs_info, 0);
4258
4259 set_reloc_control(rc);
4260
4261 trans = btrfs_join_transaction(rc->extent_root);
4262 if (IS_ERR(trans)) {
4263 err = PTR_ERR(trans);
4264 goto out_unset;
4265 }
4266
4267 rc->merge_reloc_tree = 1;
4268
4269 while (!list_empty(&reloc_roots)) {
4270 reloc_root = list_entry(reloc_roots.next,
4271 struct btrfs_root, root_list);
4272 list_del(&reloc_root->root_list);
4273
4274 if (btrfs_root_refs(&reloc_root->root_item) == 0) {
4275 list_add_tail(&reloc_root->root_list,
4276 &rc->reloc_roots);
4277 continue;
4278 }
4279
4280 fs_root = btrfs_get_fs_root(fs_info, reloc_root->root_key.offset,
4281 false);
4282 if (IS_ERR(fs_root)) {
4283 err = PTR_ERR(fs_root);
4284 list_add_tail(&reloc_root->root_list, &reloc_roots);
4285 btrfs_end_transaction(trans);
4286 goto out_unset;
4287 }
4288
4289 err = __add_reloc_root(reloc_root);
4290 ASSERT(err != -EEXIST);
4291 if (err) {
4292 list_add_tail(&reloc_root->root_list, &reloc_roots);
4293 btrfs_put_root(fs_root);
4294 btrfs_end_transaction(trans);
4295 goto out_unset;
4296 }
4297 fs_root->reloc_root = btrfs_grab_root(reloc_root);
4298 btrfs_put_root(fs_root);
4299 }
4300
4301 err = btrfs_commit_transaction(trans);
4302 if (err)
4303 goto out_unset;
4304
4305 merge_reloc_roots(rc);
4306
4307 unset_reloc_control(rc);
4308
4309 trans = btrfs_join_transaction(rc->extent_root);
4310 if (IS_ERR(trans)) {
4311 err = PTR_ERR(trans);
4312 goto out_clean;
4313 }
4314 err = btrfs_commit_transaction(trans);
4315out_clean:
4316 ret = clean_dirty_subvols(rc);
4317 if (ret < 0 && !err)
4318 err = ret;
4319out_unset:
4320 unset_reloc_control(rc);
4321out_end:
4322 reloc_chunk_end(fs_info);
4323 free_reloc_control(rc);
4324out:
4325 free_reloc_roots(&reloc_roots);
4326
4327 btrfs_free_path(path);
4328
4329 if (err == 0) {
4330 /* cleanup orphan inode in data relocation tree */
4331 fs_root = btrfs_grab_root(fs_info->data_reloc_root);
4332 ASSERT(fs_root);
4333 err = btrfs_orphan_cleanup(fs_root);
4334 btrfs_put_root(fs_root);
4335 }
4336 return err;
4337}
4338
4339/*
4340 * helper to add ordered checksum for data relocation.
4341 *
4342 * cloning checksum properly handles the nodatasum extents.
4343 * it also saves CPU time to re-calculate the checksum.
4344 */
4345int btrfs_reloc_clone_csums(struct btrfs_inode *inode, u64 file_pos, u64 len)
4346{
4347 struct btrfs_fs_info *fs_info = inode->root->fs_info;
4348 struct btrfs_root *csum_root;
4349 struct btrfs_ordered_sum *sums;
4350 struct btrfs_ordered_extent *ordered;
4351 int ret;
4352 u64 disk_bytenr;
4353 u64 new_bytenr;
4354 LIST_HEAD(list);
4355
4356 ordered = btrfs_lookup_ordered_extent(inode, file_pos);
4357 BUG_ON(ordered->file_offset != file_pos || ordered->num_bytes != len);
4358
4359 disk_bytenr = file_pos + inode->index_cnt;
4360 csum_root = btrfs_csum_root(fs_info, disk_bytenr);
4361 ret = btrfs_lookup_csums_list(csum_root, disk_bytenr,
4362 disk_bytenr + len - 1, &list, 0, false);
4363 if (ret)
4364 goto out;
4365
4366 while (!list_empty(&list)) {
4367 sums = list_entry(list.next, struct btrfs_ordered_sum, list);
4368 list_del_init(&sums->list);
4369
4370 /*
4371 * We need to offset the new_bytenr based on where the csum is.
4372 * We need to do this because we will read in entire prealloc
4373 * extents but we may have written to say the middle of the
4374 * prealloc extent, so we need to make sure the csum goes with
4375 * the right disk offset.
4376 *
4377 * We can do this because the data reloc inode refers strictly
4378 * to the on disk bytes, so we don't have to worry about
4379 * disk_len vs real len like with real inodes since it's all
4380 * disk length.
4381 */
4382 new_bytenr = ordered->disk_bytenr + sums->bytenr - disk_bytenr;
4383 sums->bytenr = new_bytenr;
4384
4385 btrfs_add_ordered_sum(ordered, sums);
4386 }
4387out:
4388 btrfs_put_ordered_extent(ordered);
4389 return ret;
4390}
4391
4392int btrfs_reloc_cow_block(struct btrfs_trans_handle *trans,
4393 struct btrfs_root *root, struct extent_buffer *buf,
4394 struct extent_buffer *cow)
4395{
4396 struct btrfs_fs_info *fs_info = root->fs_info;
4397 struct reloc_control *rc;
4398 struct btrfs_backref_node *node;
4399 int first_cow = 0;
4400 int level;
4401 int ret = 0;
4402
4403 rc = fs_info->reloc_ctl;
4404 if (!rc)
4405 return 0;
4406
4407 BUG_ON(rc->stage == UPDATE_DATA_PTRS && btrfs_is_data_reloc_root(root));
4408
4409 level = btrfs_header_level(buf);
4410 if (btrfs_header_generation(buf) <=
4411 btrfs_root_last_snapshot(&root->root_item))
4412 first_cow = 1;
4413
4414 if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID &&
4415 rc->create_reloc_tree) {
4416 WARN_ON(!first_cow && level == 0);
4417
4418 node = rc->backref_cache.path[level];
4419 BUG_ON(node->bytenr != buf->start &&
4420 node->new_bytenr != buf->start);
4421
4422 btrfs_backref_drop_node_buffer(node);
4423 atomic_inc(&cow->refs);
4424 node->eb = cow;
4425 node->new_bytenr = cow->start;
4426
4427 if (!node->pending) {
4428 list_move_tail(&node->list,
4429 &rc->backref_cache.pending[level]);
4430 node->pending = 1;
4431 }
4432
4433 if (first_cow)
4434 mark_block_processed(rc, node);
4435
4436 if (first_cow && level > 0)
4437 rc->nodes_relocated += buf->len;
4438 }
4439
4440 if (level == 0 && first_cow && rc->stage == UPDATE_DATA_PTRS)
4441 ret = replace_file_extents(trans, rc, root, cow);
4442 return ret;
4443}
4444
4445/*
4446 * called before creating snapshot. it calculates metadata reservation
4447 * required for relocating tree blocks in the snapshot
4448 */
4449void btrfs_reloc_pre_snapshot(struct btrfs_pending_snapshot *pending,
4450 u64 *bytes_to_reserve)
4451{
4452 struct btrfs_root *root = pending->root;
4453 struct reloc_control *rc = root->fs_info->reloc_ctl;
4454
4455 if (!rc || !have_reloc_root(root))
4456 return;
4457
4458 if (!rc->merge_reloc_tree)
4459 return;
4460
4461 root = root->reloc_root;
4462 BUG_ON(btrfs_root_refs(&root->root_item) == 0);
4463 /*
4464 * relocation is in the stage of merging trees. the space
4465 * used by merging a reloc tree is twice the size of
4466 * relocated tree nodes in the worst case. half for cowing
4467 * the reloc tree, half for cowing the fs tree. the space
4468 * used by cowing the reloc tree will be freed after the
4469 * tree is dropped. if we create snapshot, cowing the fs
4470 * tree may use more space than it frees. so we need
4471 * reserve extra space.
4472 */
4473 *bytes_to_reserve += rc->nodes_relocated;
4474}
4475
4476/*
4477 * called after snapshot is created. migrate block reservation
4478 * and create reloc root for the newly created snapshot
4479 *
4480 * This is similar to btrfs_init_reloc_root(), we come out of here with two
4481 * references held on the reloc_root, one for root->reloc_root and one for
4482 * rc->reloc_roots.
4483 */
4484int btrfs_reloc_post_snapshot(struct btrfs_trans_handle *trans,
4485 struct btrfs_pending_snapshot *pending)
4486{
4487 struct btrfs_root *root = pending->root;
4488 struct btrfs_root *reloc_root;
4489 struct btrfs_root *new_root;
4490 struct reloc_control *rc = root->fs_info->reloc_ctl;
4491 int ret;
4492
4493 if (!rc || !have_reloc_root(root))
4494 return 0;
4495
4496 rc = root->fs_info->reloc_ctl;
4497 rc->merging_rsv_size += rc->nodes_relocated;
4498
4499 if (rc->merge_reloc_tree) {
4500 ret = btrfs_block_rsv_migrate(&pending->block_rsv,
4501 rc->block_rsv,
4502 rc->nodes_relocated, true);
4503 if (ret)
4504 return ret;
4505 }
4506
4507 new_root = pending->snap;
4508 reloc_root = create_reloc_root(trans, root->reloc_root,
4509 new_root->root_key.objectid);
4510 if (IS_ERR(reloc_root))
4511 return PTR_ERR(reloc_root);
4512
4513 ret = __add_reloc_root(reloc_root);
4514 ASSERT(ret != -EEXIST);
4515 if (ret) {
4516 /* Pairs with create_reloc_root */
4517 btrfs_put_root(reloc_root);
4518 return ret;
4519 }
4520 new_root->reloc_root = btrfs_grab_root(reloc_root);
4521
4522 if (rc->create_reloc_tree)
4523 ret = clone_backref_node(trans, rc, root, reloc_root);
4524 return ret;
4525}
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * Copyright (C) 2009 Oracle. All rights reserved.
4 */
5
6#include <linux/sched.h>
7#include <linux/pagemap.h>
8#include <linux/writeback.h>
9#include <linux/blkdev.h>
10#include <linux/rbtree.h>
11#include <linux/slab.h>
12#include <linux/error-injection.h>
13#include "ctree.h"
14#include "disk-io.h"
15#include "transaction.h"
16#include "volumes.h"
17#include "locking.h"
18#include "btrfs_inode.h"
19#include "async-thread.h"
20#include "free-space-cache.h"
21#include "inode-map.h"
22#include "qgroup.h"
23#include "print-tree.h"
24#include "delalloc-space.h"
25#include "block-group.h"
26#include "backref.h"
27#include "misc.h"
28
29/*
30 * Relocation overview
31 *
32 * [What does relocation do]
33 *
34 * The objective of relocation is to relocate all extents of the target block
35 * group to other block groups.
36 * This is utilized by resize (shrink only), profile converting, compacting
37 * space, or balance routine to spread chunks over devices.
38 *
39 * Before | After
40 * ------------------------------------------------------------------
41 * BG A: 10 data extents | BG A: deleted
42 * BG B: 2 data extents | BG B: 10 data extents (2 old + 8 relocated)
43 * BG C: 1 extents | BG C: 3 data extents (1 old + 2 relocated)
44 *
45 * [How does relocation work]
46 *
47 * 1. Mark the target block group read-only
48 * New extents won't be allocated from the target block group.
49 *
50 * 2.1 Record each extent in the target block group
51 * To build a proper map of extents to be relocated.
52 *
53 * 2.2 Build data reloc tree and reloc trees
54 * Data reloc tree will contain an inode, recording all newly relocated
55 * data extents.
56 * There will be only one data reloc tree for one data block group.
57 *
58 * Reloc tree will be a special snapshot of its source tree, containing
59 * relocated tree blocks.
60 * Each tree referring to a tree block in target block group will get its
61 * reloc tree built.
62 *
63 * 2.3 Swap source tree with its corresponding reloc tree
64 * Each involved tree only refers to new extents after swap.
65 *
66 * 3. Cleanup reloc trees and data reloc tree.
67 * As old extents in the target block group are still referenced by reloc
68 * trees, we need to clean them up before really freeing the target block
69 * group.
70 *
71 * The main complexity is in steps 2.2 and 2.3.
72 *
73 * The entry point of relocation is relocate_block_group() function.
74 */
75
76#define RELOCATION_RESERVED_NODES 256
77/*
78 * map address of tree root to tree
79 */
80struct mapping_node {
81 struct {
82 struct rb_node rb_node;
83 u64 bytenr;
84 }; /* Use rb_simle_node for search/insert */
85 void *data;
86};
87
88struct mapping_tree {
89 struct rb_root rb_root;
90 spinlock_t lock;
91};
92
93/*
94 * present a tree block to process
95 */
96struct tree_block {
97 struct {
98 struct rb_node rb_node;
99 u64 bytenr;
100 }; /* Use rb_simple_node for search/insert */
101 struct btrfs_key key;
102 unsigned int level:8;
103 unsigned int key_ready:1;
104};
105
106#define MAX_EXTENTS 128
107
108struct file_extent_cluster {
109 u64 start;
110 u64 end;
111 u64 boundary[MAX_EXTENTS];
112 unsigned int nr;
113};
114
115struct reloc_control {
116 /* block group to relocate */
117 struct btrfs_block_group *block_group;
118 /* extent tree */
119 struct btrfs_root *extent_root;
120 /* inode for moving data */
121 struct inode *data_inode;
122
123 struct btrfs_block_rsv *block_rsv;
124
125 struct btrfs_backref_cache backref_cache;
126
127 struct file_extent_cluster cluster;
128 /* tree blocks have been processed */
129 struct extent_io_tree processed_blocks;
130 /* map start of tree root to corresponding reloc tree */
131 struct mapping_tree reloc_root_tree;
132 /* list of reloc trees */
133 struct list_head reloc_roots;
134 /* list of subvolume trees that get relocated */
135 struct list_head dirty_subvol_roots;
136 /* size of metadata reservation for merging reloc trees */
137 u64 merging_rsv_size;
138 /* size of relocated tree nodes */
139 u64 nodes_relocated;
140 /* reserved size for block group relocation*/
141 u64 reserved_bytes;
142
143 u64 search_start;
144 u64 extents_found;
145
146 unsigned int stage:8;
147 unsigned int create_reloc_tree:1;
148 unsigned int merge_reloc_tree:1;
149 unsigned int found_file_extent:1;
150};
151
152/* stages of data relocation */
153#define MOVE_DATA_EXTENTS 0
154#define UPDATE_DATA_PTRS 1
155
156static void mark_block_processed(struct reloc_control *rc,
157 struct btrfs_backref_node *node)
158{
159 u32 blocksize;
160
161 if (node->level == 0 ||
162 in_range(node->bytenr, rc->block_group->start,
163 rc->block_group->length)) {
164 blocksize = rc->extent_root->fs_info->nodesize;
165 set_extent_bits(&rc->processed_blocks, node->bytenr,
166 node->bytenr + blocksize - 1, EXTENT_DIRTY);
167 }
168 node->processed = 1;
169}
170
171
172static void mapping_tree_init(struct mapping_tree *tree)
173{
174 tree->rb_root = RB_ROOT;
175 spin_lock_init(&tree->lock);
176}
177
178/*
179 * walk up backref nodes until reach node presents tree root
180 */
181static struct btrfs_backref_node *walk_up_backref(
182 struct btrfs_backref_node *node,
183 struct btrfs_backref_edge *edges[], int *index)
184{
185 struct btrfs_backref_edge *edge;
186 int idx = *index;
187
188 while (!list_empty(&node->upper)) {
189 edge = list_entry(node->upper.next,
190 struct btrfs_backref_edge, list[LOWER]);
191 edges[idx++] = edge;
192 node = edge->node[UPPER];
193 }
194 BUG_ON(node->detached);
195 *index = idx;
196 return node;
197}
198
199/*
200 * walk down backref nodes to find start of next reference path
201 */
202static struct btrfs_backref_node *walk_down_backref(
203 struct btrfs_backref_edge *edges[], int *index)
204{
205 struct btrfs_backref_edge *edge;
206 struct btrfs_backref_node *lower;
207 int idx = *index;
208
209 while (idx > 0) {
210 edge = edges[idx - 1];
211 lower = edge->node[LOWER];
212 if (list_is_last(&edge->list[LOWER], &lower->upper)) {
213 idx--;
214 continue;
215 }
216 edge = list_entry(edge->list[LOWER].next,
217 struct btrfs_backref_edge, list[LOWER]);
218 edges[idx - 1] = edge;
219 *index = idx;
220 return edge->node[UPPER];
221 }
222 *index = 0;
223 return NULL;
224}
225
226static void update_backref_node(struct btrfs_backref_cache *cache,
227 struct btrfs_backref_node *node, u64 bytenr)
228{
229 struct rb_node *rb_node;
230 rb_erase(&node->rb_node, &cache->rb_root);
231 node->bytenr = bytenr;
232 rb_node = rb_simple_insert(&cache->rb_root, node->bytenr, &node->rb_node);
233 if (rb_node)
234 btrfs_backref_panic(cache->fs_info, bytenr, -EEXIST);
235}
236
237/*
238 * update backref cache after a transaction commit
239 */
240static int update_backref_cache(struct btrfs_trans_handle *trans,
241 struct btrfs_backref_cache *cache)
242{
243 struct btrfs_backref_node *node;
244 int level = 0;
245
246 if (cache->last_trans == 0) {
247 cache->last_trans = trans->transid;
248 return 0;
249 }
250
251 if (cache->last_trans == trans->transid)
252 return 0;
253
254 /*
255 * detached nodes are used to avoid unnecessary backref
256 * lookup. transaction commit changes the extent tree.
257 * so the detached nodes are no longer useful.
258 */
259 while (!list_empty(&cache->detached)) {
260 node = list_entry(cache->detached.next,
261 struct btrfs_backref_node, list);
262 btrfs_backref_cleanup_node(cache, node);
263 }
264
265 while (!list_empty(&cache->changed)) {
266 node = list_entry(cache->changed.next,
267 struct btrfs_backref_node, list);
268 list_del_init(&node->list);
269 BUG_ON(node->pending);
270 update_backref_node(cache, node, node->new_bytenr);
271 }
272
273 /*
274 * some nodes can be left in the pending list if there were
275 * errors during processing the pending nodes.
276 */
277 for (level = 0; level < BTRFS_MAX_LEVEL; level++) {
278 list_for_each_entry(node, &cache->pending[level], list) {
279 BUG_ON(!node->pending);
280 if (node->bytenr == node->new_bytenr)
281 continue;
282 update_backref_node(cache, node, node->new_bytenr);
283 }
284 }
285
286 cache->last_trans = 0;
287 return 1;
288}
289
290static bool reloc_root_is_dead(struct btrfs_root *root)
291{
292 /*
293 * Pair with set_bit/clear_bit in clean_dirty_subvols and
294 * btrfs_update_reloc_root. We need to see the updated bit before
295 * trying to access reloc_root
296 */
297 smp_rmb();
298 if (test_bit(BTRFS_ROOT_DEAD_RELOC_TREE, &root->state))
299 return true;
300 return false;
301}
302
303/*
304 * Check if this subvolume tree has valid reloc tree.
305 *
306 * Reloc tree after swap is considered dead, thus not considered as valid.
307 * This is enough for most callers, as they don't distinguish dead reloc root
308 * from no reloc root. But btrfs_should_ignore_reloc_root() below is a
309 * special case.
310 */
311static bool have_reloc_root(struct btrfs_root *root)
312{
313 if (reloc_root_is_dead(root))
314 return false;
315 if (!root->reloc_root)
316 return false;
317 return true;
318}
319
320int btrfs_should_ignore_reloc_root(struct btrfs_root *root)
321{
322 struct btrfs_root *reloc_root;
323
324 if (!test_bit(BTRFS_ROOT_SHAREABLE, &root->state))
325 return 0;
326
327 /* This root has been merged with its reloc tree, we can ignore it */
328 if (reloc_root_is_dead(root))
329 return 1;
330
331 reloc_root = root->reloc_root;
332 if (!reloc_root)
333 return 0;
334
335 if (btrfs_header_generation(reloc_root->commit_root) ==
336 root->fs_info->running_transaction->transid)
337 return 0;
338 /*
339 * if there is reloc tree and it was created in previous
340 * transaction backref lookup can find the reloc tree,
341 * so backref node for the fs tree root is useless for
342 * relocation.
343 */
344 return 1;
345}
346
347/*
348 * find reloc tree by address of tree root
349 */
350struct btrfs_root *find_reloc_root(struct btrfs_fs_info *fs_info, u64 bytenr)
351{
352 struct reloc_control *rc = fs_info->reloc_ctl;
353 struct rb_node *rb_node;
354 struct mapping_node *node;
355 struct btrfs_root *root = NULL;
356
357 ASSERT(rc);
358 spin_lock(&rc->reloc_root_tree.lock);
359 rb_node = rb_simple_search(&rc->reloc_root_tree.rb_root, bytenr);
360 if (rb_node) {
361 node = rb_entry(rb_node, struct mapping_node, rb_node);
362 root = (struct btrfs_root *)node->data;
363 }
364 spin_unlock(&rc->reloc_root_tree.lock);
365 return btrfs_grab_root(root);
366}
367
368/*
369 * For useless nodes, do two major clean ups:
370 *
371 * - Cleanup the children edges and nodes
372 * If child node is also orphan (no parent) during cleanup, then the child
373 * node will also be cleaned up.
374 *
375 * - Freeing up leaves (level 0), keeps nodes detached
376 * For nodes, the node is still cached as "detached"
377 *
378 * Return false if @node is not in the @useless_nodes list.
379 * Return true if @node is in the @useless_nodes list.
380 */
381static bool handle_useless_nodes(struct reloc_control *rc,
382 struct btrfs_backref_node *node)
383{
384 struct btrfs_backref_cache *cache = &rc->backref_cache;
385 struct list_head *useless_node = &cache->useless_node;
386 bool ret = false;
387
388 while (!list_empty(useless_node)) {
389 struct btrfs_backref_node *cur;
390
391 cur = list_first_entry(useless_node, struct btrfs_backref_node,
392 list);
393 list_del_init(&cur->list);
394
395 /* Only tree root nodes can be added to @useless_nodes */
396 ASSERT(list_empty(&cur->upper));
397
398 if (cur == node)
399 ret = true;
400
401 /* The node is the lowest node */
402 if (cur->lowest) {
403 list_del_init(&cur->lower);
404 cur->lowest = 0;
405 }
406
407 /* Cleanup the lower edges */
408 while (!list_empty(&cur->lower)) {
409 struct btrfs_backref_edge *edge;
410 struct btrfs_backref_node *lower;
411
412 edge = list_entry(cur->lower.next,
413 struct btrfs_backref_edge, list[UPPER]);
414 list_del(&edge->list[UPPER]);
415 list_del(&edge->list[LOWER]);
416 lower = edge->node[LOWER];
417 btrfs_backref_free_edge(cache, edge);
418
419 /* Child node is also orphan, queue for cleanup */
420 if (list_empty(&lower->upper))
421 list_add(&lower->list, useless_node);
422 }
423 /* Mark this block processed for relocation */
424 mark_block_processed(rc, cur);
425
426 /*
427 * Backref nodes for tree leaves are deleted from the cache.
428 * Backref nodes for upper level tree blocks are left in the
429 * cache to avoid unnecessary backref lookup.
430 */
431 if (cur->level > 0) {
432 list_add(&cur->list, &cache->detached);
433 cur->detached = 1;
434 } else {
435 rb_erase(&cur->rb_node, &cache->rb_root);
436 btrfs_backref_free_node(cache, cur);
437 }
438 }
439 return ret;
440}
441
442/*
443 * Build backref tree for a given tree block. Root of the backref tree
444 * corresponds the tree block, leaves of the backref tree correspond roots of
445 * b-trees that reference the tree block.
446 *
447 * The basic idea of this function is check backrefs of a given block to find
448 * upper level blocks that reference the block, and then check backrefs of
449 * these upper level blocks recursively. The recursion stops when tree root is
450 * reached or backrefs for the block is cached.
451 *
452 * NOTE: if we find that backrefs for a block are cached, we know backrefs for
453 * all upper level blocks that directly/indirectly reference the block are also
454 * cached.
455 */
456static noinline_for_stack struct btrfs_backref_node *build_backref_tree(
457 struct reloc_control *rc, struct btrfs_key *node_key,
458 int level, u64 bytenr)
459{
460 struct btrfs_backref_iter *iter;
461 struct btrfs_backref_cache *cache = &rc->backref_cache;
462 /* For searching parent of TREE_BLOCK_REF */
463 struct btrfs_path *path;
464 struct btrfs_backref_node *cur;
465 struct btrfs_backref_node *node = NULL;
466 struct btrfs_backref_edge *edge;
467 int ret;
468 int err = 0;
469
470 iter = btrfs_backref_iter_alloc(rc->extent_root->fs_info, GFP_NOFS);
471 if (!iter)
472 return ERR_PTR(-ENOMEM);
473 path = btrfs_alloc_path();
474 if (!path) {
475 err = -ENOMEM;
476 goto out;
477 }
478
479 node = btrfs_backref_alloc_node(cache, bytenr, level);
480 if (!node) {
481 err = -ENOMEM;
482 goto out;
483 }
484
485 node->lowest = 1;
486 cur = node;
487
488 /* Breadth-first search to build backref cache */
489 do {
490 ret = btrfs_backref_add_tree_node(cache, path, iter, node_key,
491 cur);
492 if (ret < 0) {
493 err = ret;
494 goto out;
495 }
496 edge = list_first_entry_or_null(&cache->pending_edge,
497 struct btrfs_backref_edge, list[UPPER]);
498 /*
499 * The pending list isn't empty, take the first block to
500 * process
501 */
502 if (edge) {
503 list_del_init(&edge->list[UPPER]);
504 cur = edge->node[UPPER];
505 }
506 } while (edge);
507
508 /* Finish the upper linkage of newly added edges/nodes */
509 ret = btrfs_backref_finish_upper_links(cache, node);
510 if (ret < 0) {
511 err = ret;
512 goto out;
513 }
514
515 if (handle_useless_nodes(rc, node))
516 node = NULL;
517out:
518 btrfs_backref_iter_free(iter);
519 btrfs_free_path(path);
520 if (err) {
521 btrfs_backref_error_cleanup(cache, node);
522 return ERR_PTR(err);
523 }
524 ASSERT(!node || !node->detached);
525 ASSERT(list_empty(&cache->useless_node) &&
526 list_empty(&cache->pending_edge));
527 return node;
528}
529
530/*
531 * helper to add backref node for the newly created snapshot.
532 * the backref node is created by cloning backref node that
533 * corresponds to root of source tree
534 */
535static int clone_backref_node(struct btrfs_trans_handle *trans,
536 struct reloc_control *rc,
537 struct btrfs_root *src,
538 struct btrfs_root *dest)
539{
540 struct btrfs_root *reloc_root = src->reloc_root;
541 struct btrfs_backref_cache *cache = &rc->backref_cache;
542 struct btrfs_backref_node *node = NULL;
543 struct btrfs_backref_node *new_node;
544 struct btrfs_backref_edge *edge;
545 struct btrfs_backref_edge *new_edge;
546 struct rb_node *rb_node;
547
548 if (cache->last_trans > 0)
549 update_backref_cache(trans, cache);
550
551 rb_node = rb_simple_search(&cache->rb_root, src->commit_root->start);
552 if (rb_node) {
553 node = rb_entry(rb_node, struct btrfs_backref_node, rb_node);
554 if (node->detached)
555 node = NULL;
556 else
557 BUG_ON(node->new_bytenr != reloc_root->node->start);
558 }
559
560 if (!node) {
561 rb_node = rb_simple_search(&cache->rb_root,
562 reloc_root->commit_root->start);
563 if (rb_node) {
564 node = rb_entry(rb_node, struct btrfs_backref_node,
565 rb_node);
566 BUG_ON(node->detached);
567 }
568 }
569
570 if (!node)
571 return 0;
572
573 new_node = btrfs_backref_alloc_node(cache, dest->node->start,
574 node->level);
575 if (!new_node)
576 return -ENOMEM;
577
578 new_node->lowest = node->lowest;
579 new_node->checked = 1;
580 new_node->root = btrfs_grab_root(dest);
581 ASSERT(new_node->root);
582
583 if (!node->lowest) {
584 list_for_each_entry(edge, &node->lower, list[UPPER]) {
585 new_edge = btrfs_backref_alloc_edge(cache);
586 if (!new_edge)
587 goto fail;
588
589 btrfs_backref_link_edge(new_edge, edge->node[LOWER],
590 new_node, LINK_UPPER);
591 }
592 } else {
593 list_add_tail(&new_node->lower, &cache->leaves);
594 }
595
596 rb_node = rb_simple_insert(&cache->rb_root, new_node->bytenr,
597 &new_node->rb_node);
598 if (rb_node)
599 btrfs_backref_panic(trans->fs_info, new_node->bytenr, -EEXIST);
600
601 if (!new_node->lowest) {
602 list_for_each_entry(new_edge, &new_node->lower, list[UPPER]) {
603 list_add_tail(&new_edge->list[LOWER],
604 &new_edge->node[LOWER]->upper);
605 }
606 }
607 return 0;
608fail:
609 while (!list_empty(&new_node->lower)) {
610 new_edge = list_entry(new_node->lower.next,
611 struct btrfs_backref_edge, list[UPPER]);
612 list_del(&new_edge->list[UPPER]);
613 btrfs_backref_free_edge(cache, new_edge);
614 }
615 btrfs_backref_free_node(cache, new_node);
616 return -ENOMEM;
617}
618
619/*
620 * helper to add 'address of tree root -> reloc tree' mapping
621 */
622static int __must_check __add_reloc_root(struct btrfs_root *root)
623{
624 struct btrfs_fs_info *fs_info = root->fs_info;
625 struct rb_node *rb_node;
626 struct mapping_node *node;
627 struct reloc_control *rc = fs_info->reloc_ctl;
628
629 node = kmalloc(sizeof(*node), GFP_NOFS);
630 if (!node)
631 return -ENOMEM;
632
633 node->bytenr = root->commit_root->start;
634 node->data = root;
635
636 spin_lock(&rc->reloc_root_tree.lock);
637 rb_node = rb_simple_insert(&rc->reloc_root_tree.rb_root,
638 node->bytenr, &node->rb_node);
639 spin_unlock(&rc->reloc_root_tree.lock);
640 if (rb_node) {
641 btrfs_panic(fs_info, -EEXIST,
642 "Duplicate root found for start=%llu while inserting into relocation tree",
643 node->bytenr);
644 }
645
646 list_add_tail(&root->root_list, &rc->reloc_roots);
647 return 0;
648}
649
650/*
651 * helper to delete the 'address of tree root -> reloc tree'
652 * mapping
653 */
654static void __del_reloc_root(struct btrfs_root *root)
655{
656 struct btrfs_fs_info *fs_info = root->fs_info;
657 struct rb_node *rb_node;
658 struct mapping_node *node = NULL;
659 struct reloc_control *rc = fs_info->reloc_ctl;
660 bool put_ref = false;
661
662 if (rc && root->node) {
663 spin_lock(&rc->reloc_root_tree.lock);
664 rb_node = rb_simple_search(&rc->reloc_root_tree.rb_root,
665 root->commit_root->start);
666 if (rb_node) {
667 node = rb_entry(rb_node, struct mapping_node, rb_node);
668 rb_erase(&node->rb_node, &rc->reloc_root_tree.rb_root);
669 RB_CLEAR_NODE(&node->rb_node);
670 }
671 spin_unlock(&rc->reloc_root_tree.lock);
672 if (!node)
673 return;
674 BUG_ON((struct btrfs_root *)node->data != root);
675 }
676
677 /*
678 * We only put the reloc root here if it's on the list. There's a lot
679 * of places where the pattern is to splice the rc->reloc_roots, process
680 * the reloc roots, and then add the reloc root back onto
681 * rc->reloc_roots. If we call __del_reloc_root while it's off of the
682 * list we don't want the reference being dropped, because the guy
683 * messing with the list is in charge of the reference.
684 */
685 spin_lock(&fs_info->trans_lock);
686 if (!list_empty(&root->root_list)) {
687 put_ref = true;
688 list_del_init(&root->root_list);
689 }
690 spin_unlock(&fs_info->trans_lock);
691 if (put_ref)
692 btrfs_put_root(root);
693 kfree(node);
694}
695
696/*
697 * helper to update the 'address of tree root -> reloc tree'
698 * mapping
699 */
700static int __update_reloc_root(struct btrfs_root *root)
701{
702 struct btrfs_fs_info *fs_info = root->fs_info;
703 struct rb_node *rb_node;
704 struct mapping_node *node = NULL;
705 struct reloc_control *rc = fs_info->reloc_ctl;
706
707 spin_lock(&rc->reloc_root_tree.lock);
708 rb_node = rb_simple_search(&rc->reloc_root_tree.rb_root,
709 root->commit_root->start);
710 if (rb_node) {
711 node = rb_entry(rb_node, struct mapping_node, rb_node);
712 rb_erase(&node->rb_node, &rc->reloc_root_tree.rb_root);
713 }
714 spin_unlock(&rc->reloc_root_tree.lock);
715
716 if (!node)
717 return 0;
718 BUG_ON((struct btrfs_root *)node->data != root);
719
720 spin_lock(&rc->reloc_root_tree.lock);
721 node->bytenr = root->node->start;
722 rb_node = rb_simple_insert(&rc->reloc_root_tree.rb_root,
723 node->bytenr, &node->rb_node);
724 spin_unlock(&rc->reloc_root_tree.lock);
725 if (rb_node)
726 btrfs_backref_panic(fs_info, node->bytenr, -EEXIST);
727 return 0;
728}
729
730static struct btrfs_root *create_reloc_root(struct btrfs_trans_handle *trans,
731 struct btrfs_root *root, u64 objectid)
732{
733 struct btrfs_fs_info *fs_info = root->fs_info;
734 struct btrfs_root *reloc_root;
735 struct extent_buffer *eb;
736 struct btrfs_root_item *root_item;
737 struct btrfs_key root_key;
738 int ret;
739
740 root_item = kmalloc(sizeof(*root_item), GFP_NOFS);
741 BUG_ON(!root_item);
742
743 root_key.objectid = BTRFS_TREE_RELOC_OBJECTID;
744 root_key.type = BTRFS_ROOT_ITEM_KEY;
745 root_key.offset = objectid;
746
747 if (root->root_key.objectid == objectid) {
748 u64 commit_root_gen;
749
750 /* called by btrfs_init_reloc_root */
751 ret = btrfs_copy_root(trans, root, root->commit_root, &eb,
752 BTRFS_TREE_RELOC_OBJECTID);
753 BUG_ON(ret);
754 /*
755 * Set the last_snapshot field to the generation of the commit
756 * root - like this ctree.c:btrfs_block_can_be_shared() behaves
757 * correctly (returns true) when the relocation root is created
758 * either inside the critical section of a transaction commit
759 * (through transaction.c:qgroup_account_snapshot()) and when
760 * it's created before the transaction commit is started.
761 */
762 commit_root_gen = btrfs_header_generation(root->commit_root);
763 btrfs_set_root_last_snapshot(&root->root_item, commit_root_gen);
764 } else {
765 /*
766 * called by btrfs_reloc_post_snapshot_hook.
767 * the source tree is a reloc tree, all tree blocks
768 * modified after it was created have RELOC flag
769 * set in their headers. so it's OK to not update
770 * the 'last_snapshot'.
771 */
772 ret = btrfs_copy_root(trans, root, root->node, &eb,
773 BTRFS_TREE_RELOC_OBJECTID);
774 BUG_ON(ret);
775 }
776
777 memcpy(root_item, &root->root_item, sizeof(*root_item));
778 btrfs_set_root_bytenr(root_item, eb->start);
779 btrfs_set_root_level(root_item, btrfs_header_level(eb));
780 btrfs_set_root_generation(root_item, trans->transid);
781
782 if (root->root_key.objectid == objectid) {
783 btrfs_set_root_refs(root_item, 0);
784 memset(&root_item->drop_progress, 0,
785 sizeof(struct btrfs_disk_key));
786 root_item->drop_level = 0;
787 }
788
789 btrfs_tree_unlock(eb);
790 free_extent_buffer(eb);
791
792 ret = btrfs_insert_root(trans, fs_info->tree_root,
793 &root_key, root_item);
794 BUG_ON(ret);
795 kfree(root_item);
796
797 reloc_root = btrfs_read_tree_root(fs_info->tree_root, &root_key);
798 BUG_ON(IS_ERR(reloc_root));
799 set_bit(BTRFS_ROOT_SHAREABLE, &reloc_root->state);
800 reloc_root->last_trans = trans->transid;
801 return reloc_root;
802}
803
804/*
805 * create reloc tree for a given fs tree. reloc tree is just a
806 * snapshot of the fs tree with special root objectid.
807 *
808 * The reloc_root comes out of here with two references, one for
809 * root->reloc_root, and another for being on the rc->reloc_roots list.
810 */
811int btrfs_init_reloc_root(struct btrfs_trans_handle *trans,
812 struct btrfs_root *root)
813{
814 struct btrfs_fs_info *fs_info = root->fs_info;
815 struct btrfs_root *reloc_root;
816 struct reloc_control *rc = fs_info->reloc_ctl;
817 struct btrfs_block_rsv *rsv;
818 int clear_rsv = 0;
819 int ret;
820
821 if (!rc)
822 return 0;
823
824 /*
825 * The subvolume has reloc tree but the swap is finished, no need to
826 * create/update the dead reloc tree
827 */
828 if (reloc_root_is_dead(root))
829 return 0;
830
831 /*
832 * This is subtle but important. We do not do
833 * record_root_in_transaction for reloc roots, instead we record their
834 * corresponding fs root, and then here we update the last trans for the
835 * reloc root. This means that we have to do this for the entire life
836 * of the reloc root, regardless of which stage of the relocation we are
837 * in.
838 */
839 if (root->reloc_root) {
840 reloc_root = root->reloc_root;
841 reloc_root->last_trans = trans->transid;
842 return 0;
843 }
844
845 /*
846 * We are merging reloc roots, we do not need new reloc trees. Also
847 * reloc trees never need their own reloc tree.
848 */
849 if (!rc->create_reloc_tree ||
850 root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID)
851 return 0;
852
853 if (!trans->reloc_reserved) {
854 rsv = trans->block_rsv;
855 trans->block_rsv = rc->block_rsv;
856 clear_rsv = 1;
857 }
858 reloc_root = create_reloc_root(trans, root, root->root_key.objectid);
859 if (clear_rsv)
860 trans->block_rsv = rsv;
861
862 ret = __add_reloc_root(reloc_root);
863 BUG_ON(ret < 0);
864 root->reloc_root = btrfs_grab_root(reloc_root);
865 return 0;
866}
867
868/*
869 * update root item of reloc tree
870 */
871int btrfs_update_reloc_root(struct btrfs_trans_handle *trans,
872 struct btrfs_root *root)
873{
874 struct btrfs_fs_info *fs_info = root->fs_info;
875 struct btrfs_root *reloc_root;
876 struct btrfs_root_item *root_item;
877 int ret;
878
879 if (!have_reloc_root(root))
880 goto out;
881
882 reloc_root = root->reloc_root;
883 root_item = &reloc_root->root_item;
884
885 /*
886 * We are probably ok here, but __del_reloc_root() will drop its ref of
887 * the root. We have the ref for root->reloc_root, but just in case
888 * hold it while we update the reloc root.
889 */
890 btrfs_grab_root(reloc_root);
891
892 /* root->reloc_root will stay until current relocation finished */
893 if (fs_info->reloc_ctl->merge_reloc_tree &&
894 btrfs_root_refs(root_item) == 0) {
895 set_bit(BTRFS_ROOT_DEAD_RELOC_TREE, &root->state);
896 /*
897 * Mark the tree as dead before we change reloc_root so
898 * have_reloc_root will not touch it from now on.
899 */
900 smp_wmb();
901 __del_reloc_root(reloc_root);
902 }
903
904 if (reloc_root->commit_root != reloc_root->node) {
905 __update_reloc_root(reloc_root);
906 btrfs_set_root_node(root_item, reloc_root->node);
907 free_extent_buffer(reloc_root->commit_root);
908 reloc_root->commit_root = btrfs_root_node(reloc_root);
909 }
910
911 ret = btrfs_update_root(trans, fs_info->tree_root,
912 &reloc_root->root_key, root_item);
913 BUG_ON(ret);
914 btrfs_put_root(reloc_root);
915out:
916 return 0;
917}
918
919/*
920 * helper to find first cached inode with inode number >= objectid
921 * in a subvolume
922 */
923static struct inode *find_next_inode(struct btrfs_root *root, u64 objectid)
924{
925 struct rb_node *node;
926 struct rb_node *prev;
927 struct btrfs_inode *entry;
928 struct inode *inode;
929
930 spin_lock(&root->inode_lock);
931again:
932 node = root->inode_tree.rb_node;
933 prev = NULL;
934 while (node) {
935 prev = node;
936 entry = rb_entry(node, struct btrfs_inode, rb_node);
937
938 if (objectid < btrfs_ino(entry))
939 node = node->rb_left;
940 else if (objectid > btrfs_ino(entry))
941 node = node->rb_right;
942 else
943 break;
944 }
945 if (!node) {
946 while (prev) {
947 entry = rb_entry(prev, struct btrfs_inode, rb_node);
948 if (objectid <= btrfs_ino(entry)) {
949 node = prev;
950 break;
951 }
952 prev = rb_next(prev);
953 }
954 }
955 while (node) {
956 entry = rb_entry(node, struct btrfs_inode, rb_node);
957 inode = igrab(&entry->vfs_inode);
958 if (inode) {
959 spin_unlock(&root->inode_lock);
960 return inode;
961 }
962
963 objectid = btrfs_ino(entry) + 1;
964 if (cond_resched_lock(&root->inode_lock))
965 goto again;
966
967 node = rb_next(node);
968 }
969 spin_unlock(&root->inode_lock);
970 return NULL;
971}
972
973/*
974 * get new location of data
975 */
976static int get_new_location(struct inode *reloc_inode, u64 *new_bytenr,
977 u64 bytenr, u64 num_bytes)
978{
979 struct btrfs_root *root = BTRFS_I(reloc_inode)->root;
980 struct btrfs_path *path;
981 struct btrfs_file_extent_item *fi;
982 struct extent_buffer *leaf;
983 int ret;
984
985 path = btrfs_alloc_path();
986 if (!path)
987 return -ENOMEM;
988
989 bytenr -= BTRFS_I(reloc_inode)->index_cnt;
990 ret = btrfs_lookup_file_extent(NULL, root, path,
991 btrfs_ino(BTRFS_I(reloc_inode)), bytenr, 0);
992 if (ret < 0)
993 goto out;
994 if (ret > 0) {
995 ret = -ENOENT;
996 goto out;
997 }
998
999 leaf = path->nodes[0];
1000 fi = btrfs_item_ptr(leaf, path->slots[0],
1001 struct btrfs_file_extent_item);
1002
1003 BUG_ON(btrfs_file_extent_offset(leaf, fi) ||
1004 btrfs_file_extent_compression(leaf, fi) ||
1005 btrfs_file_extent_encryption(leaf, fi) ||
1006 btrfs_file_extent_other_encoding(leaf, fi));
1007
1008 if (num_bytes != btrfs_file_extent_disk_num_bytes(leaf, fi)) {
1009 ret = -EINVAL;
1010 goto out;
1011 }
1012
1013 *new_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
1014 ret = 0;
1015out:
1016 btrfs_free_path(path);
1017 return ret;
1018}
1019
1020/*
1021 * update file extent items in the tree leaf to point to
1022 * the new locations.
1023 */
1024static noinline_for_stack
1025int replace_file_extents(struct btrfs_trans_handle *trans,
1026 struct reloc_control *rc,
1027 struct btrfs_root *root,
1028 struct extent_buffer *leaf)
1029{
1030 struct btrfs_fs_info *fs_info = root->fs_info;
1031 struct btrfs_key key;
1032 struct btrfs_file_extent_item *fi;
1033 struct inode *inode = NULL;
1034 u64 parent;
1035 u64 bytenr;
1036 u64 new_bytenr = 0;
1037 u64 num_bytes;
1038 u64 end;
1039 u32 nritems;
1040 u32 i;
1041 int ret = 0;
1042 int first = 1;
1043 int dirty = 0;
1044
1045 if (rc->stage != UPDATE_DATA_PTRS)
1046 return 0;
1047
1048 /* reloc trees always use full backref */
1049 if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID)
1050 parent = leaf->start;
1051 else
1052 parent = 0;
1053
1054 nritems = btrfs_header_nritems(leaf);
1055 for (i = 0; i < nritems; i++) {
1056 struct btrfs_ref ref = { 0 };
1057
1058 cond_resched();
1059 btrfs_item_key_to_cpu(leaf, &key, i);
1060 if (key.type != BTRFS_EXTENT_DATA_KEY)
1061 continue;
1062 fi = btrfs_item_ptr(leaf, i, struct btrfs_file_extent_item);
1063 if (btrfs_file_extent_type(leaf, fi) ==
1064 BTRFS_FILE_EXTENT_INLINE)
1065 continue;
1066 bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
1067 num_bytes = btrfs_file_extent_disk_num_bytes(leaf, fi);
1068 if (bytenr == 0)
1069 continue;
1070 if (!in_range(bytenr, rc->block_group->start,
1071 rc->block_group->length))
1072 continue;
1073
1074 /*
1075 * if we are modifying block in fs tree, wait for readpage
1076 * to complete and drop the extent cache
1077 */
1078 if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) {
1079 if (first) {
1080 inode = find_next_inode(root, key.objectid);
1081 first = 0;
1082 } else if (inode && btrfs_ino(BTRFS_I(inode)) < key.objectid) {
1083 btrfs_add_delayed_iput(inode);
1084 inode = find_next_inode(root, key.objectid);
1085 }
1086 if (inode && btrfs_ino(BTRFS_I(inode)) == key.objectid) {
1087 end = key.offset +
1088 btrfs_file_extent_num_bytes(leaf, fi);
1089 WARN_ON(!IS_ALIGNED(key.offset,
1090 fs_info->sectorsize));
1091 WARN_ON(!IS_ALIGNED(end, fs_info->sectorsize));
1092 end--;
1093 ret = try_lock_extent(&BTRFS_I(inode)->io_tree,
1094 key.offset, end);
1095 if (!ret)
1096 continue;
1097
1098 btrfs_drop_extent_cache(BTRFS_I(inode),
1099 key.offset, end, 1);
1100 unlock_extent(&BTRFS_I(inode)->io_tree,
1101 key.offset, end);
1102 }
1103 }
1104
1105 ret = get_new_location(rc->data_inode, &new_bytenr,
1106 bytenr, num_bytes);
1107 if (ret) {
1108 /*
1109 * Don't have to abort since we've not changed anything
1110 * in the file extent yet.
1111 */
1112 break;
1113 }
1114
1115 btrfs_set_file_extent_disk_bytenr(leaf, fi, new_bytenr);
1116 dirty = 1;
1117
1118 key.offset -= btrfs_file_extent_offset(leaf, fi);
1119 btrfs_init_generic_ref(&ref, BTRFS_ADD_DELAYED_REF, new_bytenr,
1120 num_bytes, parent);
1121 ref.real_root = root->root_key.objectid;
1122 btrfs_init_data_ref(&ref, btrfs_header_owner(leaf),
1123 key.objectid, key.offset);
1124 ret = btrfs_inc_extent_ref(trans, &ref);
1125 if (ret) {
1126 btrfs_abort_transaction(trans, ret);
1127 break;
1128 }
1129
1130 btrfs_init_generic_ref(&ref, BTRFS_DROP_DELAYED_REF, bytenr,
1131 num_bytes, parent);
1132 ref.real_root = root->root_key.objectid;
1133 btrfs_init_data_ref(&ref, btrfs_header_owner(leaf),
1134 key.objectid, key.offset);
1135 ret = btrfs_free_extent(trans, &ref);
1136 if (ret) {
1137 btrfs_abort_transaction(trans, ret);
1138 break;
1139 }
1140 }
1141 if (dirty)
1142 btrfs_mark_buffer_dirty(leaf);
1143 if (inode)
1144 btrfs_add_delayed_iput(inode);
1145 return ret;
1146}
1147
1148static noinline_for_stack
1149int memcmp_node_keys(struct extent_buffer *eb, int slot,
1150 struct btrfs_path *path, int level)
1151{
1152 struct btrfs_disk_key key1;
1153 struct btrfs_disk_key key2;
1154 btrfs_node_key(eb, &key1, slot);
1155 btrfs_node_key(path->nodes[level], &key2, path->slots[level]);
1156 return memcmp(&key1, &key2, sizeof(key1));
1157}
1158
1159/*
1160 * try to replace tree blocks in fs tree with the new blocks
1161 * in reloc tree. tree blocks haven't been modified since the
1162 * reloc tree was create can be replaced.
1163 *
1164 * if a block was replaced, level of the block + 1 is returned.
1165 * if no block got replaced, 0 is returned. if there are other
1166 * errors, a negative error number is returned.
1167 */
1168static noinline_for_stack
1169int replace_path(struct btrfs_trans_handle *trans, struct reloc_control *rc,
1170 struct btrfs_root *dest, struct btrfs_root *src,
1171 struct btrfs_path *path, struct btrfs_key *next_key,
1172 int lowest_level, int max_level)
1173{
1174 struct btrfs_fs_info *fs_info = dest->fs_info;
1175 struct extent_buffer *eb;
1176 struct extent_buffer *parent;
1177 struct btrfs_ref ref = { 0 };
1178 struct btrfs_key key;
1179 u64 old_bytenr;
1180 u64 new_bytenr;
1181 u64 old_ptr_gen;
1182 u64 new_ptr_gen;
1183 u64 last_snapshot;
1184 u32 blocksize;
1185 int cow = 0;
1186 int level;
1187 int ret;
1188 int slot;
1189
1190 BUG_ON(src->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID);
1191 BUG_ON(dest->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID);
1192
1193 last_snapshot = btrfs_root_last_snapshot(&src->root_item);
1194again:
1195 slot = path->slots[lowest_level];
1196 btrfs_node_key_to_cpu(path->nodes[lowest_level], &key, slot);
1197
1198 eb = btrfs_lock_root_node(dest);
1199 btrfs_set_lock_blocking_write(eb);
1200 level = btrfs_header_level(eb);
1201
1202 if (level < lowest_level) {
1203 btrfs_tree_unlock(eb);
1204 free_extent_buffer(eb);
1205 return 0;
1206 }
1207
1208 if (cow) {
1209 ret = btrfs_cow_block(trans, dest, eb, NULL, 0, &eb);
1210 BUG_ON(ret);
1211 }
1212 btrfs_set_lock_blocking_write(eb);
1213
1214 if (next_key) {
1215 next_key->objectid = (u64)-1;
1216 next_key->type = (u8)-1;
1217 next_key->offset = (u64)-1;
1218 }
1219
1220 parent = eb;
1221 while (1) {
1222 struct btrfs_key first_key;
1223
1224 level = btrfs_header_level(parent);
1225 BUG_ON(level < lowest_level);
1226
1227 ret = btrfs_bin_search(parent, &key, &slot);
1228 if (ret < 0)
1229 break;
1230 if (ret && slot > 0)
1231 slot--;
1232
1233 if (next_key && slot + 1 < btrfs_header_nritems(parent))
1234 btrfs_node_key_to_cpu(parent, next_key, slot + 1);
1235
1236 old_bytenr = btrfs_node_blockptr(parent, slot);
1237 blocksize = fs_info->nodesize;
1238 old_ptr_gen = btrfs_node_ptr_generation(parent, slot);
1239 btrfs_node_key_to_cpu(parent, &first_key, slot);
1240
1241 if (level <= max_level) {
1242 eb = path->nodes[level];
1243 new_bytenr = btrfs_node_blockptr(eb,
1244 path->slots[level]);
1245 new_ptr_gen = btrfs_node_ptr_generation(eb,
1246 path->slots[level]);
1247 } else {
1248 new_bytenr = 0;
1249 new_ptr_gen = 0;
1250 }
1251
1252 if (WARN_ON(new_bytenr > 0 && new_bytenr == old_bytenr)) {
1253 ret = level;
1254 break;
1255 }
1256
1257 if (new_bytenr == 0 || old_ptr_gen > last_snapshot ||
1258 memcmp_node_keys(parent, slot, path, level)) {
1259 if (level <= lowest_level) {
1260 ret = 0;
1261 break;
1262 }
1263
1264 eb = read_tree_block(fs_info, old_bytenr, old_ptr_gen,
1265 level - 1, &first_key);
1266 if (IS_ERR(eb)) {
1267 ret = PTR_ERR(eb);
1268 break;
1269 } else if (!extent_buffer_uptodate(eb)) {
1270 ret = -EIO;
1271 free_extent_buffer(eb);
1272 break;
1273 }
1274 btrfs_tree_lock(eb);
1275 if (cow) {
1276 ret = btrfs_cow_block(trans, dest, eb, parent,
1277 slot, &eb);
1278 BUG_ON(ret);
1279 }
1280 btrfs_set_lock_blocking_write(eb);
1281
1282 btrfs_tree_unlock(parent);
1283 free_extent_buffer(parent);
1284
1285 parent = eb;
1286 continue;
1287 }
1288
1289 if (!cow) {
1290 btrfs_tree_unlock(parent);
1291 free_extent_buffer(parent);
1292 cow = 1;
1293 goto again;
1294 }
1295
1296 btrfs_node_key_to_cpu(path->nodes[level], &key,
1297 path->slots[level]);
1298 btrfs_release_path(path);
1299
1300 path->lowest_level = level;
1301 ret = btrfs_search_slot(trans, src, &key, path, 0, 1);
1302 path->lowest_level = 0;
1303 BUG_ON(ret);
1304
1305 /*
1306 * Info qgroup to trace both subtrees.
1307 *
1308 * We must trace both trees.
1309 * 1) Tree reloc subtree
1310 * If not traced, we will leak data numbers
1311 * 2) Fs subtree
1312 * If not traced, we will double count old data
1313 *
1314 * We don't scan the subtree right now, but only record
1315 * the swapped tree blocks.
1316 * The real subtree rescan is delayed until we have new
1317 * CoW on the subtree root node before transaction commit.
1318 */
1319 ret = btrfs_qgroup_add_swapped_blocks(trans, dest,
1320 rc->block_group, parent, slot,
1321 path->nodes[level], path->slots[level],
1322 last_snapshot);
1323 if (ret < 0)
1324 break;
1325 /*
1326 * swap blocks in fs tree and reloc tree.
1327 */
1328 btrfs_set_node_blockptr(parent, slot, new_bytenr);
1329 btrfs_set_node_ptr_generation(parent, slot, new_ptr_gen);
1330 btrfs_mark_buffer_dirty(parent);
1331
1332 btrfs_set_node_blockptr(path->nodes[level],
1333 path->slots[level], old_bytenr);
1334 btrfs_set_node_ptr_generation(path->nodes[level],
1335 path->slots[level], old_ptr_gen);
1336 btrfs_mark_buffer_dirty(path->nodes[level]);
1337
1338 btrfs_init_generic_ref(&ref, BTRFS_ADD_DELAYED_REF, old_bytenr,
1339 blocksize, path->nodes[level]->start);
1340 ref.skip_qgroup = true;
1341 btrfs_init_tree_ref(&ref, level - 1, src->root_key.objectid);
1342 ret = btrfs_inc_extent_ref(trans, &ref);
1343 BUG_ON(ret);
1344 btrfs_init_generic_ref(&ref, BTRFS_ADD_DELAYED_REF, new_bytenr,
1345 blocksize, 0);
1346 ref.skip_qgroup = true;
1347 btrfs_init_tree_ref(&ref, level - 1, dest->root_key.objectid);
1348 ret = btrfs_inc_extent_ref(trans, &ref);
1349 BUG_ON(ret);
1350
1351 btrfs_init_generic_ref(&ref, BTRFS_DROP_DELAYED_REF, new_bytenr,
1352 blocksize, path->nodes[level]->start);
1353 btrfs_init_tree_ref(&ref, level - 1, src->root_key.objectid);
1354 ref.skip_qgroup = true;
1355 ret = btrfs_free_extent(trans, &ref);
1356 BUG_ON(ret);
1357
1358 btrfs_init_generic_ref(&ref, BTRFS_DROP_DELAYED_REF, old_bytenr,
1359 blocksize, 0);
1360 btrfs_init_tree_ref(&ref, level - 1, dest->root_key.objectid);
1361 ref.skip_qgroup = true;
1362 ret = btrfs_free_extent(trans, &ref);
1363 BUG_ON(ret);
1364
1365 btrfs_unlock_up_safe(path, 0);
1366
1367 ret = level;
1368 break;
1369 }
1370 btrfs_tree_unlock(parent);
1371 free_extent_buffer(parent);
1372 return ret;
1373}
1374
1375/*
1376 * helper to find next relocated block in reloc tree
1377 */
1378static noinline_for_stack
1379int walk_up_reloc_tree(struct btrfs_root *root, struct btrfs_path *path,
1380 int *level)
1381{
1382 struct extent_buffer *eb;
1383 int i;
1384 u64 last_snapshot;
1385 u32 nritems;
1386
1387 last_snapshot = btrfs_root_last_snapshot(&root->root_item);
1388
1389 for (i = 0; i < *level; i++) {
1390 free_extent_buffer(path->nodes[i]);
1391 path->nodes[i] = NULL;
1392 }
1393
1394 for (i = *level; i < BTRFS_MAX_LEVEL && path->nodes[i]; i++) {
1395 eb = path->nodes[i];
1396 nritems = btrfs_header_nritems(eb);
1397 while (path->slots[i] + 1 < nritems) {
1398 path->slots[i]++;
1399 if (btrfs_node_ptr_generation(eb, path->slots[i]) <=
1400 last_snapshot)
1401 continue;
1402
1403 *level = i;
1404 return 0;
1405 }
1406 free_extent_buffer(path->nodes[i]);
1407 path->nodes[i] = NULL;
1408 }
1409 return 1;
1410}
1411
1412/*
1413 * walk down reloc tree to find relocated block of lowest level
1414 */
1415static noinline_for_stack
1416int walk_down_reloc_tree(struct btrfs_root *root, struct btrfs_path *path,
1417 int *level)
1418{
1419 struct btrfs_fs_info *fs_info = root->fs_info;
1420 struct extent_buffer *eb = NULL;
1421 int i;
1422 u64 bytenr;
1423 u64 ptr_gen = 0;
1424 u64 last_snapshot;
1425 u32 nritems;
1426
1427 last_snapshot = btrfs_root_last_snapshot(&root->root_item);
1428
1429 for (i = *level; i > 0; i--) {
1430 struct btrfs_key first_key;
1431
1432 eb = path->nodes[i];
1433 nritems = btrfs_header_nritems(eb);
1434 while (path->slots[i] < nritems) {
1435 ptr_gen = btrfs_node_ptr_generation(eb, path->slots[i]);
1436 if (ptr_gen > last_snapshot)
1437 break;
1438 path->slots[i]++;
1439 }
1440 if (path->slots[i] >= nritems) {
1441 if (i == *level)
1442 break;
1443 *level = i + 1;
1444 return 0;
1445 }
1446 if (i == 1) {
1447 *level = i;
1448 return 0;
1449 }
1450
1451 bytenr = btrfs_node_blockptr(eb, path->slots[i]);
1452 btrfs_node_key_to_cpu(eb, &first_key, path->slots[i]);
1453 eb = read_tree_block(fs_info, bytenr, ptr_gen, i - 1,
1454 &first_key);
1455 if (IS_ERR(eb)) {
1456 return PTR_ERR(eb);
1457 } else if (!extent_buffer_uptodate(eb)) {
1458 free_extent_buffer(eb);
1459 return -EIO;
1460 }
1461 BUG_ON(btrfs_header_level(eb) != i - 1);
1462 path->nodes[i - 1] = eb;
1463 path->slots[i - 1] = 0;
1464 }
1465 return 1;
1466}
1467
1468/*
1469 * invalidate extent cache for file extents whose key in range of
1470 * [min_key, max_key)
1471 */
1472static int invalidate_extent_cache(struct btrfs_root *root,
1473 struct btrfs_key *min_key,
1474 struct btrfs_key *max_key)
1475{
1476 struct btrfs_fs_info *fs_info = root->fs_info;
1477 struct inode *inode = NULL;
1478 u64 objectid;
1479 u64 start, end;
1480 u64 ino;
1481
1482 objectid = min_key->objectid;
1483 while (1) {
1484 cond_resched();
1485 iput(inode);
1486
1487 if (objectid > max_key->objectid)
1488 break;
1489
1490 inode = find_next_inode(root, objectid);
1491 if (!inode)
1492 break;
1493 ino = btrfs_ino(BTRFS_I(inode));
1494
1495 if (ino > max_key->objectid) {
1496 iput(inode);
1497 break;
1498 }
1499
1500 objectid = ino + 1;
1501 if (!S_ISREG(inode->i_mode))
1502 continue;
1503
1504 if (unlikely(min_key->objectid == ino)) {
1505 if (min_key->type > BTRFS_EXTENT_DATA_KEY)
1506 continue;
1507 if (min_key->type < BTRFS_EXTENT_DATA_KEY)
1508 start = 0;
1509 else {
1510 start = min_key->offset;
1511 WARN_ON(!IS_ALIGNED(start, fs_info->sectorsize));
1512 }
1513 } else {
1514 start = 0;
1515 }
1516
1517 if (unlikely(max_key->objectid == ino)) {
1518 if (max_key->type < BTRFS_EXTENT_DATA_KEY)
1519 continue;
1520 if (max_key->type > BTRFS_EXTENT_DATA_KEY) {
1521 end = (u64)-1;
1522 } else {
1523 if (max_key->offset == 0)
1524 continue;
1525 end = max_key->offset;
1526 WARN_ON(!IS_ALIGNED(end, fs_info->sectorsize));
1527 end--;
1528 }
1529 } else {
1530 end = (u64)-1;
1531 }
1532
1533 /* the lock_extent waits for readpage to complete */
1534 lock_extent(&BTRFS_I(inode)->io_tree, start, end);
1535 btrfs_drop_extent_cache(BTRFS_I(inode), start, end, 1);
1536 unlock_extent(&BTRFS_I(inode)->io_tree, start, end);
1537 }
1538 return 0;
1539}
1540
1541static int find_next_key(struct btrfs_path *path, int level,
1542 struct btrfs_key *key)
1543
1544{
1545 while (level < BTRFS_MAX_LEVEL) {
1546 if (!path->nodes[level])
1547 break;
1548 if (path->slots[level] + 1 <
1549 btrfs_header_nritems(path->nodes[level])) {
1550 btrfs_node_key_to_cpu(path->nodes[level], key,
1551 path->slots[level] + 1);
1552 return 0;
1553 }
1554 level++;
1555 }
1556 return 1;
1557}
1558
1559/*
1560 * Insert current subvolume into reloc_control::dirty_subvol_roots
1561 */
1562static void insert_dirty_subvol(struct btrfs_trans_handle *trans,
1563 struct reloc_control *rc,
1564 struct btrfs_root *root)
1565{
1566 struct btrfs_root *reloc_root = root->reloc_root;
1567 struct btrfs_root_item *reloc_root_item;
1568
1569 /* @root must be a subvolume tree root with a valid reloc tree */
1570 ASSERT(root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID);
1571 ASSERT(reloc_root);
1572
1573 reloc_root_item = &reloc_root->root_item;
1574 memset(&reloc_root_item->drop_progress, 0,
1575 sizeof(reloc_root_item->drop_progress));
1576 reloc_root_item->drop_level = 0;
1577 btrfs_set_root_refs(reloc_root_item, 0);
1578 btrfs_update_reloc_root(trans, root);
1579
1580 if (list_empty(&root->reloc_dirty_list)) {
1581 btrfs_grab_root(root);
1582 list_add_tail(&root->reloc_dirty_list, &rc->dirty_subvol_roots);
1583 }
1584}
1585
1586static int clean_dirty_subvols(struct reloc_control *rc)
1587{
1588 struct btrfs_root *root;
1589 struct btrfs_root *next;
1590 int ret = 0;
1591 int ret2;
1592
1593 list_for_each_entry_safe(root, next, &rc->dirty_subvol_roots,
1594 reloc_dirty_list) {
1595 if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) {
1596 /* Merged subvolume, cleanup its reloc root */
1597 struct btrfs_root *reloc_root = root->reloc_root;
1598
1599 list_del_init(&root->reloc_dirty_list);
1600 root->reloc_root = NULL;
1601 /*
1602 * Need barrier to ensure clear_bit() only happens after
1603 * root->reloc_root = NULL. Pairs with have_reloc_root.
1604 */
1605 smp_wmb();
1606 clear_bit(BTRFS_ROOT_DEAD_RELOC_TREE, &root->state);
1607 if (reloc_root) {
1608 /*
1609 * btrfs_drop_snapshot drops our ref we hold for
1610 * ->reloc_root. If it fails however we must
1611 * drop the ref ourselves.
1612 */
1613 ret2 = btrfs_drop_snapshot(reloc_root, 0, 1);
1614 if (ret2 < 0) {
1615 btrfs_put_root(reloc_root);
1616 if (!ret)
1617 ret = ret2;
1618 }
1619 }
1620 btrfs_put_root(root);
1621 } else {
1622 /* Orphan reloc tree, just clean it up */
1623 ret2 = btrfs_drop_snapshot(root, 0, 1);
1624 if (ret2 < 0) {
1625 btrfs_put_root(root);
1626 if (!ret)
1627 ret = ret2;
1628 }
1629 }
1630 }
1631 return ret;
1632}
1633
1634/*
1635 * merge the relocated tree blocks in reloc tree with corresponding
1636 * fs tree.
1637 */
1638static noinline_for_stack int merge_reloc_root(struct reloc_control *rc,
1639 struct btrfs_root *root)
1640{
1641 struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
1642 struct btrfs_key key;
1643 struct btrfs_key next_key;
1644 struct btrfs_trans_handle *trans = NULL;
1645 struct btrfs_root *reloc_root;
1646 struct btrfs_root_item *root_item;
1647 struct btrfs_path *path;
1648 struct extent_buffer *leaf;
1649 int level;
1650 int max_level;
1651 int replaced = 0;
1652 int ret;
1653 int err = 0;
1654 u32 min_reserved;
1655
1656 path = btrfs_alloc_path();
1657 if (!path)
1658 return -ENOMEM;
1659 path->reada = READA_FORWARD;
1660
1661 reloc_root = root->reloc_root;
1662 root_item = &reloc_root->root_item;
1663
1664 if (btrfs_disk_key_objectid(&root_item->drop_progress) == 0) {
1665 level = btrfs_root_level(root_item);
1666 atomic_inc(&reloc_root->node->refs);
1667 path->nodes[level] = reloc_root->node;
1668 path->slots[level] = 0;
1669 } else {
1670 btrfs_disk_key_to_cpu(&key, &root_item->drop_progress);
1671
1672 level = root_item->drop_level;
1673 BUG_ON(level == 0);
1674 path->lowest_level = level;
1675 ret = btrfs_search_slot(NULL, reloc_root, &key, path, 0, 0);
1676 path->lowest_level = 0;
1677 if (ret < 0) {
1678 btrfs_free_path(path);
1679 return ret;
1680 }
1681
1682 btrfs_node_key_to_cpu(path->nodes[level], &next_key,
1683 path->slots[level]);
1684 WARN_ON(memcmp(&key, &next_key, sizeof(key)));
1685
1686 btrfs_unlock_up_safe(path, 0);
1687 }
1688
1689 /*
1690 * In merge_reloc_root(), we modify the upper level pointer to swap the
1691 * tree blocks between reloc tree and subvolume tree. Thus for tree
1692 * block COW, we COW at most from level 1 to root level for each tree.
1693 *
1694 * Thus the needed metadata size is at most root_level * nodesize,
1695 * and * 2 since we have two trees to COW.
1696 */
1697 min_reserved = fs_info->nodesize * btrfs_root_level(root_item) * 2;
1698 memset(&next_key, 0, sizeof(next_key));
1699
1700 while (1) {
1701 ret = btrfs_block_rsv_refill(root, rc->block_rsv, min_reserved,
1702 BTRFS_RESERVE_FLUSH_LIMIT);
1703 if (ret) {
1704 err = ret;
1705 goto out;
1706 }
1707 trans = btrfs_start_transaction(root, 0);
1708 if (IS_ERR(trans)) {
1709 err = PTR_ERR(trans);
1710 trans = NULL;
1711 goto out;
1712 }
1713
1714 /*
1715 * At this point we no longer have a reloc_control, so we can't
1716 * depend on btrfs_init_reloc_root to update our last_trans.
1717 *
1718 * But that's ok, we started the trans handle on our
1719 * corresponding fs_root, which means it's been added to the
1720 * dirty list. At commit time we'll still call
1721 * btrfs_update_reloc_root() and update our root item
1722 * appropriately.
1723 */
1724 reloc_root->last_trans = trans->transid;
1725 trans->block_rsv = rc->block_rsv;
1726
1727 replaced = 0;
1728 max_level = level;
1729
1730 ret = walk_down_reloc_tree(reloc_root, path, &level);
1731 if (ret < 0) {
1732 err = ret;
1733 goto out;
1734 }
1735 if (ret > 0)
1736 break;
1737
1738 if (!find_next_key(path, level, &key) &&
1739 btrfs_comp_cpu_keys(&next_key, &key) >= 0) {
1740 ret = 0;
1741 } else {
1742 ret = replace_path(trans, rc, root, reloc_root, path,
1743 &next_key, level, max_level);
1744 }
1745 if (ret < 0) {
1746 err = ret;
1747 goto out;
1748 }
1749
1750 if (ret > 0) {
1751 level = ret;
1752 btrfs_node_key_to_cpu(path->nodes[level], &key,
1753 path->slots[level]);
1754 replaced = 1;
1755 }
1756
1757 ret = walk_up_reloc_tree(reloc_root, path, &level);
1758 if (ret > 0)
1759 break;
1760
1761 BUG_ON(level == 0);
1762 /*
1763 * save the merging progress in the drop_progress.
1764 * this is OK since root refs == 1 in this case.
1765 */
1766 btrfs_node_key(path->nodes[level], &root_item->drop_progress,
1767 path->slots[level]);
1768 root_item->drop_level = level;
1769
1770 btrfs_end_transaction_throttle(trans);
1771 trans = NULL;
1772
1773 btrfs_btree_balance_dirty(fs_info);
1774
1775 if (replaced && rc->stage == UPDATE_DATA_PTRS)
1776 invalidate_extent_cache(root, &key, &next_key);
1777 }
1778
1779 /*
1780 * handle the case only one block in the fs tree need to be
1781 * relocated and the block is tree root.
1782 */
1783 leaf = btrfs_lock_root_node(root);
1784 ret = btrfs_cow_block(trans, root, leaf, NULL, 0, &leaf);
1785 btrfs_tree_unlock(leaf);
1786 free_extent_buffer(leaf);
1787 if (ret < 0)
1788 err = ret;
1789out:
1790 btrfs_free_path(path);
1791
1792 if (err == 0)
1793 insert_dirty_subvol(trans, rc, root);
1794
1795 if (trans)
1796 btrfs_end_transaction_throttle(trans);
1797
1798 btrfs_btree_balance_dirty(fs_info);
1799
1800 if (replaced && rc->stage == UPDATE_DATA_PTRS)
1801 invalidate_extent_cache(root, &key, &next_key);
1802
1803 return err;
1804}
1805
1806static noinline_for_stack
1807int prepare_to_merge(struct reloc_control *rc, int err)
1808{
1809 struct btrfs_root *root = rc->extent_root;
1810 struct btrfs_fs_info *fs_info = root->fs_info;
1811 struct btrfs_root *reloc_root;
1812 struct btrfs_trans_handle *trans;
1813 LIST_HEAD(reloc_roots);
1814 u64 num_bytes = 0;
1815 int ret;
1816
1817 mutex_lock(&fs_info->reloc_mutex);
1818 rc->merging_rsv_size += fs_info->nodesize * (BTRFS_MAX_LEVEL - 1) * 2;
1819 rc->merging_rsv_size += rc->nodes_relocated * 2;
1820 mutex_unlock(&fs_info->reloc_mutex);
1821
1822again:
1823 if (!err) {
1824 num_bytes = rc->merging_rsv_size;
1825 ret = btrfs_block_rsv_add(root, rc->block_rsv, num_bytes,
1826 BTRFS_RESERVE_FLUSH_ALL);
1827 if (ret)
1828 err = ret;
1829 }
1830
1831 trans = btrfs_join_transaction(rc->extent_root);
1832 if (IS_ERR(trans)) {
1833 if (!err)
1834 btrfs_block_rsv_release(fs_info, rc->block_rsv,
1835 num_bytes, NULL);
1836 return PTR_ERR(trans);
1837 }
1838
1839 if (!err) {
1840 if (num_bytes != rc->merging_rsv_size) {
1841 btrfs_end_transaction(trans);
1842 btrfs_block_rsv_release(fs_info, rc->block_rsv,
1843 num_bytes, NULL);
1844 goto again;
1845 }
1846 }
1847
1848 rc->merge_reloc_tree = 1;
1849
1850 while (!list_empty(&rc->reloc_roots)) {
1851 reloc_root = list_entry(rc->reloc_roots.next,
1852 struct btrfs_root, root_list);
1853 list_del_init(&reloc_root->root_list);
1854
1855 root = btrfs_get_fs_root(fs_info, reloc_root->root_key.offset,
1856 false);
1857 BUG_ON(IS_ERR(root));
1858 BUG_ON(root->reloc_root != reloc_root);
1859
1860 /*
1861 * set reference count to 1, so btrfs_recover_relocation
1862 * knows it should resumes merging
1863 */
1864 if (!err)
1865 btrfs_set_root_refs(&reloc_root->root_item, 1);
1866 btrfs_update_reloc_root(trans, root);
1867
1868 list_add(&reloc_root->root_list, &reloc_roots);
1869 btrfs_put_root(root);
1870 }
1871
1872 list_splice(&reloc_roots, &rc->reloc_roots);
1873
1874 if (!err)
1875 btrfs_commit_transaction(trans);
1876 else
1877 btrfs_end_transaction(trans);
1878 return err;
1879}
1880
1881static noinline_for_stack
1882void free_reloc_roots(struct list_head *list)
1883{
1884 struct btrfs_root *reloc_root, *tmp;
1885
1886 list_for_each_entry_safe(reloc_root, tmp, list, root_list)
1887 __del_reloc_root(reloc_root);
1888}
1889
1890static noinline_for_stack
1891void merge_reloc_roots(struct reloc_control *rc)
1892{
1893 struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
1894 struct btrfs_root *root;
1895 struct btrfs_root *reloc_root;
1896 LIST_HEAD(reloc_roots);
1897 int found = 0;
1898 int ret = 0;
1899again:
1900 root = rc->extent_root;
1901
1902 /*
1903 * this serializes us with btrfs_record_root_in_transaction,
1904 * we have to make sure nobody is in the middle of
1905 * adding their roots to the list while we are
1906 * doing this splice
1907 */
1908 mutex_lock(&fs_info->reloc_mutex);
1909 list_splice_init(&rc->reloc_roots, &reloc_roots);
1910 mutex_unlock(&fs_info->reloc_mutex);
1911
1912 while (!list_empty(&reloc_roots)) {
1913 found = 1;
1914 reloc_root = list_entry(reloc_roots.next,
1915 struct btrfs_root, root_list);
1916
1917 root = btrfs_get_fs_root(fs_info, reloc_root->root_key.offset,
1918 false);
1919 if (btrfs_root_refs(&reloc_root->root_item) > 0) {
1920 BUG_ON(IS_ERR(root));
1921 BUG_ON(root->reloc_root != reloc_root);
1922 ret = merge_reloc_root(rc, root);
1923 btrfs_put_root(root);
1924 if (ret) {
1925 if (list_empty(&reloc_root->root_list))
1926 list_add_tail(&reloc_root->root_list,
1927 &reloc_roots);
1928 goto out;
1929 }
1930 } else {
1931 if (!IS_ERR(root)) {
1932 if (root->reloc_root == reloc_root) {
1933 root->reloc_root = NULL;
1934 btrfs_put_root(reloc_root);
1935 }
1936 clear_bit(BTRFS_ROOT_DEAD_RELOC_TREE,
1937 &root->state);
1938 btrfs_put_root(root);
1939 }
1940
1941 list_del_init(&reloc_root->root_list);
1942 /* Don't forget to queue this reloc root for cleanup */
1943 list_add_tail(&reloc_root->reloc_dirty_list,
1944 &rc->dirty_subvol_roots);
1945 }
1946 }
1947
1948 if (found) {
1949 found = 0;
1950 goto again;
1951 }
1952out:
1953 if (ret) {
1954 btrfs_handle_fs_error(fs_info, ret, NULL);
1955 free_reloc_roots(&reloc_roots);
1956
1957 /* new reloc root may be added */
1958 mutex_lock(&fs_info->reloc_mutex);
1959 list_splice_init(&rc->reloc_roots, &reloc_roots);
1960 mutex_unlock(&fs_info->reloc_mutex);
1961 free_reloc_roots(&reloc_roots);
1962 }
1963
1964 /*
1965 * We used to have
1966 *
1967 * BUG_ON(!RB_EMPTY_ROOT(&rc->reloc_root_tree.rb_root));
1968 *
1969 * here, but it's wrong. If we fail to start the transaction in
1970 * prepare_to_merge() we will have only 0 ref reloc roots, none of which
1971 * have actually been removed from the reloc_root_tree rb tree. This is
1972 * fine because we're bailing here, and we hold a reference on the root
1973 * for the list that holds it, so these roots will be cleaned up when we
1974 * do the reloc_dirty_list afterwards. Meanwhile the root->reloc_root
1975 * will be cleaned up on unmount.
1976 *
1977 * The remaining nodes will be cleaned up by free_reloc_control.
1978 */
1979}
1980
1981static void free_block_list(struct rb_root *blocks)
1982{
1983 struct tree_block *block;
1984 struct rb_node *rb_node;
1985 while ((rb_node = rb_first(blocks))) {
1986 block = rb_entry(rb_node, struct tree_block, rb_node);
1987 rb_erase(rb_node, blocks);
1988 kfree(block);
1989 }
1990}
1991
1992static int record_reloc_root_in_trans(struct btrfs_trans_handle *trans,
1993 struct btrfs_root *reloc_root)
1994{
1995 struct btrfs_fs_info *fs_info = reloc_root->fs_info;
1996 struct btrfs_root *root;
1997 int ret;
1998
1999 if (reloc_root->last_trans == trans->transid)
2000 return 0;
2001
2002 root = btrfs_get_fs_root(fs_info, reloc_root->root_key.offset, false);
2003 BUG_ON(IS_ERR(root));
2004 BUG_ON(root->reloc_root != reloc_root);
2005 ret = btrfs_record_root_in_trans(trans, root);
2006 btrfs_put_root(root);
2007
2008 return ret;
2009}
2010
2011static noinline_for_stack
2012struct btrfs_root *select_reloc_root(struct btrfs_trans_handle *trans,
2013 struct reloc_control *rc,
2014 struct btrfs_backref_node *node,
2015 struct btrfs_backref_edge *edges[])
2016{
2017 struct btrfs_backref_node *next;
2018 struct btrfs_root *root;
2019 int index = 0;
2020
2021 next = node;
2022 while (1) {
2023 cond_resched();
2024 next = walk_up_backref(next, edges, &index);
2025 root = next->root;
2026 BUG_ON(!root);
2027 BUG_ON(!test_bit(BTRFS_ROOT_SHAREABLE, &root->state));
2028
2029 if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID) {
2030 record_reloc_root_in_trans(trans, root);
2031 break;
2032 }
2033
2034 btrfs_record_root_in_trans(trans, root);
2035 root = root->reloc_root;
2036
2037 if (next->new_bytenr != root->node->start) {
2038 BUG_ON(next->new_bytenr);
2039 BUG_ON(!list_empty(&next->list));
2040 next->new_bytenr = root->node->start;
2041 btrfs_put_root(next->root);
2042 next->root = btrfs_grab_root(root);
2043 ASSERT(next->root);
2044 list_add_tail(&next->list,
2045 &rc->backref_cache.changed);
2046 mark_block_processed(rc, next);
2047 break;
2048 }
2049
2050 WARN_ON(1);
2051 root = NULL;
2052 next = walk_down_backref(edges, &index);
2053 if (!next || next->level <= node->level)
2054 break;
2055 }
2056 if (!root)
2057 return NULL;
2058
2059 next = node;
2060 /* setup backref node path for btrfs_reloc_cow_block */
2061 while (1) {
2062 rc->backref_cache.path[next->level] = next;
2063 if (--index < 0)
2064 break;
2065 next = edges[index]->node[UPPER];
2066 }
2067 return root;
2068}
2069
2070/*
2071 * Select a tree root for relocation.
2072 *
2073 * Return NULL if the block is not shareable. We should use do_relocation() in
2074 * this case.
2075 *
2076 * Return a tree root pointer if the block is shareable.
2077 * Return -ENOENT if the block is root of reloc tree.
2078 */
2079static noinline_for_stack
2080struct btrfs_root *select_one_root(struct btrfs_backref_node *node)
2081{
2082 struct btrfs_backref_node *next;
2083 struct btrfs_root *root;
2084 struct btrfs_root *fs_root = NULL;
2085 struct btrfs_backref_edge *edges[BTRFS_MAX_LEVEL - 1];
2086 int index = 0;
2087
2088 next = node;
2089 while (1) {
2090 cond_resched();
2091 next = walk_up_backref(next, edges, &index);
2092 root = next->root;
2093 BUG_ON(!root);
2094
2095 /* No other choice for non-shareable tree */
2096 if (!test_bit(BTRFS_ROOT_SHAREABLE, &root->state))
2097 return root;
2098
2099 if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID)
2100 fs_root = root;
2101
2102 if (next != node)
2103 return NULL;
2104
2105 next = walk_down_backref(edges, &index);
2106 if (!next || next->level <= node->level)
2107 break;
2108 }
2109
2110 if (!fs_root)
2111 return ERR_PTR(-ENOENT);
2112 return fs_root;
2113}
2114
2115static noinline_for_stack
2116u64 calcu_metadata_size(struct reloc_control *rc,
2117 struct btrfs_backref_node *node, int reserve)
2118{
2119 struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
2120 struct btrfs_backref_node *next = node;
2121 struct btrfs_backref_edge *edge;
2122 struct btrfs_backref_edge *edges[BTRFS_MAX_LEVEL - 1];
2123 u64 num_bytes = 0;
2124 int index = 0;
2125
2126 BUG_ON(reserve && node->processed);
2127
2128 while (next) {
2129 cond_resched();
2130 while (1) {
2131 if (next->processed && (reserve || next != node))
2132 break;
2133
2134 num_bytes += fs_info->nodesize;
2135
2136 if (list_empty(&next->upper))
2137 break;
2138
2139 edge = list_entry(next->upper.next,
2140 struct btrfs_backref_edge, list[LOWER]);
2141 edges[index++] = edge;
2142 next = edge->node[UPPER];
2143 }
2144 next = walk_down_backref(edges, &index);
2145 }
2146 return num_bytes;
2147}
2148
2149static int reserve_metadata_space(struct btrfs_trans_handle *trans,
2150 struct reloc_control *rc,
2151 struct btrfs_backref_node *node)
2152{
2153 struct btrfs_root *root = rc->extent_root;
2154 struct btrfs_fs_info *fs_info = root->fs_info;
2155 u64 num_bytes;
2156 int ret;
2157 u64 tmp;
2158
2159 num_bytes = calcu_metadata_size(rc, node, 1) * 2;
2160
2161 trans->block_rsv = rc->block_rsv;
2162 rc->reserved_bytes += num_bytes;
2163
2164 /*
2165 * We are under a transaction here so we can only do limited flushing.
2166 * If we get an enospc just kick back -EAGAIN so we know to drop the
2167 * transaction and try to refill when we can flush all the things.
2168 */
2169 ret = btrfs_block_rsv_refill(root, rc->block_rsv, num_bytes,
2170 BTRFS_RESERVE_FLUSH_LIMIT);
2171 if (ret) {
2172 tmp = fs_info->nodesize * RELOCATION_RESERVED_NODES;
2173 while (tmp <= rc->reserved_bytes)
2174 tmp <<= 1;
2175 /*
2176 * only one thread can access block_rsv at this point,
2177 * so we don't need hold lock to protect block_rsv.
2178 * we expand more reservation size here to allow enough
2179 * space for relocation and we will return earlier in
2180 * enospc case.
2181 */
2182 rc->block_rsv->size = tmp + fs_info->nodesize *
2183 RELOCATION_RESERVED_NODES;
2184 return -EAGAIN;
2185 }
2186
2187 return 0;
2188}
2189
2190/*
2191 * relocate a block tree, and then update pointers in upper level
2192 * blocks that reference the block to point to the new location.
2193 *
2194 * if called by link_to_upper, the block has already been relocated.
2195 * in that case this function just updates pointers.
2196 */
2197static int do_relocation(struct btrfs_trans_handle *trans,
2198 struct reloc_control *rc,
2199 struct btrfs_backref_node *node,
2200 struct btrfs_key *key,
2201 struct btrfs_path *path, int lowest)
2202{
2203 struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
2204 struct btrfs_backref_node *upper;
2205 struct btrfs_backref_edge *edge;
2206 struct btrfs_backref_edge *edges[BTRFS_MAX_LEVEL - 1];
2207 struct btrfs_root *root;
2208 struct extent_buffer *eb;
2209 u32 blocksize;
2210 u64 bytenr;
2211 u64 generation;
2212 int slot;
2213 int ret;
2214 int err = 0;
2215
2216 BUG_ON(lowest && node->eb);
2217
2218 path->lowest_level = node->level + 1;
2219 rc->backref_cache.path[node->level] = node;
2220 list_for_each_entry(edge, &node->upper, list[LOWER]) {
2221 struct btrfs_key first_key;
2222 struct btrfs_ref ref = { 0 };
2223
2224 cond_resched();
2225
2226 upper = edge->node[UPPER];
2227 root = select_reloc_root(trans, rc, upper, edges);
2228 BUG_ON(!root);
2229
2230 if (upper->eb && !upper->locked) {
2231 if (!lowest) {
2232 ret = btrfs_bin_search(upper->eb, key, &slot);
2233 if (ret < 0) {
2234 err = ret;
2235 goto next;
2236 }
2237 BUG_ON(ret);
2238 bytenr = btrfs_node_blockptr(upper->eb, slot);
2239 if (node->eb->start == bytenr)
2240 goto next;
2241 }
2242 btrfs_backref_drop_node_buffer(upper);
2243 }
2244
2245 if (!upper->eb) {
2246 ret = btrfs_search_slot(trans, root, key, path, 0, 1);
2247 if (ret) {
2248 if (ret < 0)
2249 err = ret;
2250 else
2251 err = -ENOENT;
2252
2253 btrfs_release_path(path);
2254 break;
2255 }
2256
2257 if (!upper->eb) {
2258 upper->eb = path->nodes[upper->level];
2259 path->nodes[upper->level] = NULL;
2260 } else {
2261 BUG_ON(upper->eb != path->nodes[upper->level]);
2262 }
2263
2264 upper->locked = 1;
2265 path->locks[upper->level] = 0;
2266
2267 slot = path->slots[upper->level];
2268 btrfs_release_path(path);
2269 } else {
2270 ret = btrfs_bin_search(upper->eb, key, &slot);
2271 if (ret < 0) {
2272 err = ret;
2273 goto next;
2274 }
2275 BUG_ON(ret);
2276 }
2277
2278 bytenr = btrfs_node_blockptr(upper->eb, slot);
2279 if (lowest) {
2280 if (bytenr != node->bytenr) {
2281 btrfs_err(root->fs_info,
2282 "lowest leaf/node mismatch: bytenr %llu node->bytenr %llu slot %d upper %llu",
2283 bytenr, node->bytenr, slot,
2284 upper->eb->start);
2285 err = -EIO;
2286 goto next;
2287 }
2288 } else {
2289 if (node->eb->start == bytenr)
2290 goto next;
2291 }
2292
2293 blocksize = root->fs_info->nodesize;
2294 generation = btrfs_node_ptr_generation(upper->eb, slot);
2295 btrfs_node_key_to_cpu(upper->eb, &first_key, slot);
2296 eb = read_tree_block(fs_info, bytenr, generation,
2297 upper->level - 1, &first_key);
2298 if (IS_ERR(eb)) {
2299 err = PTR_ERR(eb);
2300 goto next;
2301 } else if (!extent_buffer_uptodate(eb)) {
2302 free_extent_buffer(eb);
2303 err = -EIO;
2304 goto next;
2305 }
2306 btrfs_tree_lock(eb);
2307 btrfs_set_lock_blocking_write(eb);
2308
2309 if (!node->eb) {
2310 ret = btrfs_cow_block(trans, root, eb, upper->eb,
2311 slot, &eb);
2312 btrfs_tree_unlock(eb);
2313 free_extent_buffer(eb);
2314 if (ret < 0) {
2315 err = ret;
2316 goto next;
2317 }
2318 BUG_ON(node->eb != eb);
2319 } else {
2320 btrfs_set_node_blockptr(upper->eb, slot,
2321 node->eb->start);
2322 btrfs_set_node_ptr_generation(upper->eb, slot,
2323 trans->transid);
2324 btrfs_mark_buffer_dirty(upper->eb);
2325
2326 btrfs_init_generic_ref(&ref, BTRFS_ADD_DELAYED_REF,
2327 node->eb->start, blocksize,
2328 upper->eb->start);
2329 ref.real_root = root->root_key.objectid;
2330 btrfs_init_tree_ref(&ref, node->level,
2331 btrfs_header_owner(upper->eb));
2332 ret = btrfs_inc_extent_ref(trans, &ref);
2333 BUG_ON(ret);
2334
2335 ret = btrfs_drop_subtree(trans, root, eb, upper->eb);
2336 BUG_ON(ret);
2337 }
2338next:
2339 if (!upper->pending)
2340 btrfs_backref_drop_node_buffer(upper);
2341 else
2342 btrfs_backref_unlock_node_buffer(upper);
2343 if (err)
2344 break;
2345 }
2346
2347 if (!err && node->pending) {
2348 btrfs_backref_drop_node_buffer(node);
2349 list_move_tail(&node->list, &rc->backref_cache.changed);
2350 node->pending = 0;
2351 }
2352
2353 path->lowest_level = 0;
2354 BUG_ON(err == -ENOSPC);
2355 return err;
2356}
2357
2358static int link_to_upper(struct btrfs_trans_handle *trans,
2359 struct reloc_control *rc,
2360 struct btrfs_backref_node *node,
2361 struct btrfs_path *path)
2362{
2363 struct btrfs_key key;
2364
2365 btrfs_node_key_to_cpu(node->eb, &key, 0);
2366 return do_relocation(trans, rc, node, &key, path, 0);
2367}
2368
2369static int finish_pending_nodes(struct btrfs_trans_handle *trans,
2370 struct reloc_control *rc,
2371 struct btrfs_path *path, int err)
2372{
2373 LIST_HEAD(list);
2374 struct btrfs_backref_cache *cache = &rc->backref_cache;
2375 struct btrfs_backref_node *node;
2376 int level;
2377 int ret;
2378
2379 for (level = 0; level < BTRFS_MAX_LEVEL; level++) {
2380 while (!list_empty(&cache->pending[level])) {
2381 node = list_entry(cache->pending[level].next,
2382 struct btrfs_backref_node, list);
2383 list_move_tail(&node->list, &list);
2384 BUG_ON(!node->pending);
2385
2386 if (!err) {
2387 ret = link_to_upper(trans, rc, node, path);
2388 if (ret < 0)
2389 err = ret;
2390 }
2391 }
2392 list_splice_init(&list, &cache->pending[level]);
2393 }
2394 return err;
2395}
2396
2397/*
2398 * mark a block and all blocks directly/indirectly reference the block
2399 * as processed.
2400 */
2401static void update_processed_blocks(struct reloc_control *rc,
2402 struct btrfs_backref_node *node)
2403{
2404 struct btrfs_backref_node *next = node;
2405 struct btrfs_backref_edge *edge;
2406 struct btrfs_backref_edge *edges[BTRFS_MAX_LEVEL - 1];
2407 int index = 0;
2408
2409 while (next) {
2410 cond_resched();
2411 while (1) {
2412 if (next->processed)
2413 break;
2414
2415 mark_block_processed(rc, next);
2416
2417 if (list_empty(&next->upper))
2418 break;
2419
2420 edge = list_entry(next->upper.next,
2421 struct btrfs_backref_edge, list[LOWER]);
2422 edges[index++] = edge;
2423 next = edge->node[UPPER];
2424 }
2425 next = walk_down_backref(edges, &index);
2426 }
2427}
2428
2429static int tree_block_processed(u64 bytenr, struct reloc_control *rc)
2430{
2431 u32 blocksize = rc->extent_root->fs_info->nodesize;
2432
2433 if (test_range_bit(&rc->processed_blocks, bytenr,
2434 bytenr + blocksize - 1, EXTENT_DIRTY, 1, NULL))
2435 return 1;
2436 return 0;
2437}
2438
2439static int get_tree_block_key(struct btrfs_fs_info *fs_info,
2440 struct tree_block *block)
2441{
2442 struct extent_buffer *eb;
2443
2444 eb = read_tree_block(fs_info, block->bytenr, block->key.offset,
2445 block->level, NULL);
2446 if (IS_ERR(eb)) {
2447 return PTR_ERR(eb);
2448 } else if (!extent_buffer_uptodate(eb)) {
2449 free_extent_buffer(eb);
2450 return -EIO;
2451 }
2452 if (block->level == 0)
2453 btrfs_item_key_to_cpu(eb, &block->key, 0);
2454 else
2455 btrfs_node_key_to_cpu(eb, &block->key, 0);
2456 free_extent_buffer(eb);
2457 block->key_ready = 1;
2458 return 0;
2459}
2460
2461/*
2462 * helper function to relocate a tree block
2463 */
2464static int relocate_tree_block(struct btrfs_trans_handle *trans,
2465 struct reloc_control *rc,
2466 struct btrfs_backref_node *node,
2467 struct btrfs_key *key,
2468 struct btrfs_path *path)
2469{
2470 struct btrfs_root *root;
2471 int ret = 0;
2472
2473 if (!node)
2474 return 0;
2475
2476 /*
2477 * If we fail here we want to drop our backref_node because we are going
2478 * to start over and regenerate the tree for it.
2479 */
2480 ret = reserve_metadata_space(trans, rc, node);
2481 if (ret)
2482 goto out;
2483
2484 BUG_ON(node->processed);
2485 root = select_one_root(node);
2486 if (root == ERR_PTR(-ENOENT)) {
2487 update_processed_blocks(rc, node);
2488 goto out;
2489 }
2490
2491 if (root) {
2492 if (test_bit(BTRFS_ROOT_SHAREABLE, &root->state)) {
2493 BUG_ON(node->new_bytenr);
2494 BUG_ON(!list_empty(&node->list));
2495 btrfs_record_root_in_trans(trans, root);
2496 root = root->reloc_root;
2497 node->new_bytenr = root->node->start;
2498 btrfs_put_root(node->root);
2499 node->root = btrfs_grab_root(root);
2500 ASSERT(node->root);
2501 list_add_tail(&node->list, &rc->backref_cache.changed);
2502 } else {
2503 path->lowest_level = node->level;
2504 ret = btrfs_search_slot(trans, root, key, path, 0, 1);
2505 btrfs_release_path(path);
2506 if (ret > 0)
2507 ret = 0;
2508 }
2509 if (!ret)
2510 update_processed_blocks(rc, node);
2511 } else {
2512 ret = do_relocation(trans, rc, node, key, path, 1);
2513 }
2514out:
2515 if (ret || node->level == 0 || node->cowonly)
2516 btrfs_backref_cleanup_node(&rc->backref_cache, node);
2517 return ret;
2518}
2519
2520/*
2521 * relocate a list of blocks
2522 */
2523static noinline_for_stack
2524int relocate_tree_blocks(struct btrfs_trans_handle *trans,
2525 struct reloc_control *rc, struct rb_root *blocks)
2526{
2527 struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
2528 struct btrfs_backref_node *node;
2529 struct btrfs_path *path;
2530 struct tree_block *block;
2531 struct tree_block *next;
2532 int ret;
2533 int err = 0;
2534
2535 path = btrfs_alloc_path();
2536 if (!path) {
2537 err = -ENOMEM;
2538 goto out_free_blocks;
2539 }
2540
2541 /* Kick in readahead for tree blocks with missing keys */
2542 rbtree_postorder_for_each_entry_safe(block, next, blocks, rb_node) {
2543 if (!block->key_ready)
2544 readahead_tree_block(fs_info, block->bytenr);
2545 }
2546
2547 /* Get first keys */
2548 rbtree_postorder_for_each_entry_safe(block, next, blocks, rb_node) {
2549 if (!block->key_ready) {
2550 err = get_tree_block_key(fs_info, block);
2551 if (err)
2552 goto out_free_path;
2553 }
2554 }
2555
2556 /* Do tree relocation */
2557 rbtree_postorder_for_each_entry_safe(block, next, blocks, rb_node) {
2558 node = build_backref_tree(rc, &block->key,
2559 block->level, block->bytenr);
2560 if (IS_ERR(node)) {
2561 err = PTR_ERR(node);
2562 goto out;
2563 }
2564
2565 ret = relocate_tree_block(trans, rc, node, &block->key,
2566 path);
2567 if (ret < 0) {
2568 err = ret;
2569 break;
2570 }
2571 }
2572out:
2573 err = finish_pending_nodes(trans, rc, path, err);
2574
2575out_free_path:
2576 btrfs_free_path(path);
2577out_free_blocks:
2578 free_block_list(blocks);
2579 return err;
2580}
2581
2582static noinline_for_stack int prealloc_file_extent_cluster(
2583 struct btrfs_inode *inode,
2584 struct file_extent_cluster *cluster)
2585{
2586 u64 alloc_hint = 0;
2587 u64 start;
2588 u64 end;
2589 u64 offset = inode->index_cnt;
2590 u64 num_bytes;
2591 int nr;
2592 int ret = 0;
2593 u64 prealloc_start = cluster->start - offset;
2594 u64 prealloc_end = cluster->end - offset;
2595 u64 cur_offset = prealloc_start;
2596
2597 BUG_ON(cluster->start != cluster->boundary[0]);
2598 ret = btrfs_alloc_data_chunk_ondemand(inode,
2599 prealloc_end + 1 - prealloc_start);
2600 if (ret)
2601 return ret;
2602
2603 inode_lock(&inode->vfs_inode);
2604 for (nr = 0; nr < cluster->nr; nr++) {
2605 start = cluster->boundary[nr] - offset;
2606 if (nr + 1 < cluster->nr)
2607 end = cluster->boundary[nr + 1] - 1 - offset;
2608 else
2609 end = cluster->end - offset;
2610
2611 lock_extent(&inode->io_tree, start, end);
2612 num_bytes = end + 1 - start;
2613 ret = btrfs_prealloc_file_range(&inode->vfs_inode, 0, start,
2614 num_bytes, num_bytes,
2615 end + 1, &alloc_hint);
2616 cur_offset = end + 1;
2617 unlock_extent(&inode->io_tree, start, end);
2618 if (ret)
2619 break;
2620 }
2621 inode_unlock(&inode->vfs_inode);
2622
2623 if (cur_offset < prealloc_end)
2624 btrfs_free_reserved_data_space_noquota(inode->root->fs_info,
2625 prealloc_end + 1 - cur_offset);
2626 return ret;
2627}
2628
2629static noinline_for_stack
2630int setup_extent_mapping(struct inode *inode, u64 start, u64 end,
2631 u64 block_start)
2632{
2633 struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
2634 struct extent_map *em;
2635 int ret = 0;
2636
2637 em = alloc_extent_map();
2638 if (!em)
2639 return -ENOMEM;
2640
2641 em->start = start;
2642 em->len = end + 1 - start;
2643 em->block_len = em->len;
2644 em->block_start = block_start;
2645 set_bit(EXTENT_FLAG_PINNED, &em->flags);
2646
2647 lock_extent(&BTRFS_I(inode)->io_tree, start, end);
2648 while (1) {
2649 write_lock(&em_tree->lock);
2650 ret = add_extent_mapping(em_tree, em, 0);
2651 write_unlock(&em_tree->lock);
2652 if (ret != -EEXIST) {
2653 free_extent_map(em);
2654 break;
2655 }
2656 btrfs_drop_extent_cache(BTRFS_I(inode), start, end, 0);
2657 }
2658 unlock_extent(&BTRFS_I(inode)->io_tree, start, end);
2659 return ret;
2660}
2661
2662/*
2663 * Allow error injection to test balance cancellation
2664 */
2665int btrfs_should_cancel_balance(struct btrfs_fs_info *fs_info)
2666{
2667 return atomic_read(&fs_info->balance_cancel_req) ||
2668 fatal_signal_pending(current);
2669}
2670ALLOW_ERROR_INJECTION(btrfs_should_cancel_balance, TRUE);
2671
2672static int relocate_file_extent_cluster(struct inode *inode,
2673 struct file_extent_cluster *cluster)
2674{
2675 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2676 u64 page_start;
2677 u64 page_end;
2678 u64 offset = BTRFS_I(inode)->index_cnt;
2679 unsigned long index;
2680 unsigned long last_index;
2681 struct page *page;
2682 struct file_ra_state *ra;
2683 gfp_t mask = btrfs_alloc_write_mask(inode->i_mapping);
2684 int nr = 0;
2685 int ret = 0;
2686
2687 if (!cluster->nr)
2688 return 0;
2689
2690 ra = kzalloc(sizeof(*ra), GFP_NOFS);
2691 if (!ra)
2692 return -ENOMEM;
2693
2694 ret = prealloc_file_extent_cluster(BTRFS_I(inode), cluster);
2695 if (ret)
2696 goto out;
2697
2698 file_ra_state_init(ra, inode->i_mapping);
2699
2700 ret = setup_extent_mapping(inode, cluster->start - offset,
2701 cluster->end - offset, cluster->start);
2702 if (ret)
2703 goto out;
2704
2705 index = (cluster->start - offset) >> PAGE_SHIFT;
2706 last_index = (cluster->end - offset) >> PAGE_SHIFT;
2707 while (index <= last_index) {
2708 ret = btrfs_delalloc_reserve_metadata(BTRFS_I(inode),
2709 PAGE_SIZE);
2710 if (ret)
2711 goto out;
2712
2713 page = find_lock_page(inode->i_mapping, index);
2714 if (!page) {
2715 page_cache_sync_readahead(inode->i_mapping,
2716 ra, NULL, index,
2717 last_index + 1 - index);
2718 page = find_or_create_page(inode->i_mapping, index,
2719 mask);
2720 if (!page) {
2721 btrfs_delalloc_release_metadata(BTRFS_I(inode),
2722 PAGE_SIZE, true);
2723 btrfs_delalloc_release_extents(BTRFS_I(inode),
2724 PAGE_SIZE);
2725 ret = -ENOMEM;
2726 goto out;
2727 }
2728 }
2729
2730 if (PageReadahead(page)) {
2731 page_cache_async_readahead(inode->i_mapping,
2732 ra, NULL, page, index,
2733 last_index + 1 - index);
2734 }
2735
2736 if (!PageUptodate(page)) {
2737 btrfs_readpage(NULL, page);
2738 lock_page(page);
2739 if (!PageUptodate(page)) {
2740 unlock_page(page);
2741 put_page(page);
2742 btrfs_delalloc_release_metadata(BTRFS_I(inode),
2743 PAGE_SIZE, true);
2744 btrfs_delalloc_release_extents(BTRFS_I(inode),
2745 PAGE_SIZE);
2746 ret = -EIO;
2747 goto out;
2748 }
2749 }
2750
2751 page_start = page_offset(page);
2752 page_end = page_start + PAGE_SIZE - 1;
2753
2754 lock_extent(&BTRFS_I(inode)->io_tree, page_start, page_end);
2755
2756 set_page_extent_mapped(page);
2757
2758 if (nr < cluster->nr &&
2759 page_start + offset == cluster->boundary[nr]) {
2760 set_extent_bits(&BTRFS_I(inode)->io_tree,
2761 page_start, page_end,
2762 EXTENT_BOUNDARY);
2763 nr++;
2764 }
2765
2766 ret = btrfs_set_extent_delalloc(BTRFS_I(inode), page_start,
2767 page_end, 0, NULL);
2768 if (ret) {
2769 unlock_page(page);
2770 put_page(page);
2771 btrfs_delalloc_release_metadata(BTRFS_I(inode),
2772 PAGE_SIZE, true);
2773 btrfs_delalloc_release_extents(BTRFS_I(inode),
2774 PAGE_SIZE);
2775
2776 clear_extent_bits(&BTRFS_I(inode)->io_tree,
2777 page_start, page_end,
2778 EXTENT_LOCKED | EXTENT_BOUNDARY);
2779 goto out;
2780
2781 }
2782 set_page_dirty(page);
2783
2784 unlock_extent(&BTRFS_I(inode)->io_tree,
2785 page_start, page_end);
2786 unlock_page(page);
2787 put_page(page);
2788
2789 index++;
2790 btrfs_delalloc_release_extents(BTRFS_I(inode), PAGE_SIZE);
2791 balance_dirty_pages_ratelimited(inode->i_mapping);
2792 btrfs_throttle(fs_info);
2793 if (btrfs_should_cancel_balance(fs_info)) {
2794 ret = -ECANCELED;
2795 goto out;
2796 }
2797 }
2798 WARN_ON(nr != cluster->nr);
2799out:
2800 kfree(ra);
2801 return ret;
2802}
2803
2804static noinline_for_stack
2805int relocate_data_extent(struct inode *inode, struct btrfs_key *extent_key,
2806 struct file_extent_cluster *cluster)
2807{
2808 int ret;
2809
2810 if (cluster->nr > 0 && extent_key->objectid != cluster->end + 1) {
2811 ret = relocate_file_extent_cluster(inode, cluster);
2812 if (ret)
2813 return ret;
2814 cluster->nr = 0;
2815 }
2816
2817 if (!cluster->nr)
2818 cluster->start = extent_key->objectid;
2819 else
2820 BUG_ON(cluster->nr >= MAX_EXTENTS);
2821 cluster->end = extent_key->objectid + extent_key->offset - 1;
2822 cluster->boundary[cluster->nr] = extent_key->objectid;
2823 cluster->nr++;
2824
2825 if (cluster->nr >= MAX_EXTENTS) {
2826 ret = relocate_file_extent_cluster(inode, cluster);
2827 if (ret)
2828 return ret;
2829 cluster->nr = 0;
2830 }
2831 return 0;
2832}
2833
2834/*
2835 * helper to add a tree block to the list.
2836 * the major work is getting the generation and level of the block
2837 */
2838static int add_tree_block(struct reloc_control *rc,
2839 struct btrfs_key *extent_key,
2840 struct btrfs_path *path,
2841 struct rb_root *blocks)
2842{
2843 struct extent_buffer *eb;
2844 struct btrfs_extent_item *ei;
2845 struct btrfs_tree_block_info *bi;
2846 struct tree_block *block;
2847 struct rb_node *rb_node;
2848 u32 item_size;
2849 int level = -1;
2850 u64 generation;
2851
2852 eb = path->nodes[0];
2853 item_size = btrfs_item_size_nr(eb, path->slots[0]);
2854
2855 if (extent_key->type == BTRFS_METADATA_ITEM_KEY ||
2856 item_size >= sizeof(*ei) + sizeof(*bi)) {
2857 ei = btrfs_item_ptr(eb, path->slots[0],
2858 struct btrfs_extent_item);
2859 if (extent_key->type == BTRFS_EXTENT_ITEM_KEY) {
2860 bi = (struct btrfs_tree_block_info *)(ei + 1);
2861 level = btrfs_tree_block_level(eb, bi);
2862 } else {
2863 level = (int)extent_key->offset;
2864 }
2865 generation = btrfs_extent_generation(eb, ei);
2866 } else if (unlikely(item_size == sizeof(struct btrfs_extent_item_v0))) {
2867 btrfs_print_v0_err(eb->fs_info);
2868 btrfs_handle_fs_error(eb->fs_info, -EINVAL, NULL);
2869 return -EINVAL;
2870 } else {
2871 BUG();
2872 }
2873
2874 btrfs_release_path(path);
2875
2876 BUG_ON(level == -1);
2877
2878 block = kmalloc(sizeof(*block), GFP_NOFS);
2879 if (!block)
2880 return -ENOMEM;
2881
2882 block->bytenr = extent_key->objectid;
2883 block->key.objectid = rc->extent_root->fs_info->nodesize;
2884 block->key.offset = generation;
2885 block->level = level;
2886 block->key_ready = 0;
2887
2888 rb_node = rb_simple_insert(blocks, block->bytenr, &block->rb_node);
2889 if (rb_node)
2890 btrfs_backref_panic(rc->extent_root->fs_info, block->bytenr,
2891 -EEXIST);
2892
2893 return 0;
2894}
2895
2896/*
2897 * helper to add tree blocks for backref of type BTRFS_SHARED_DATA_REF_KEY
2898 */
2899static int __add_tree_block(struct reloc_control *rc,
2900 u64 bytenr, u32 blocksize,
2901 struct rb_root *blocks)
2902{
2903 struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
2904 struct btrfs_path *path;
2905 struct btrfs_key key;
2906 int ret;
2907 bool skinny = btrfs_fs_incompat(fs_info, SKINNY_METADATA);
2908
2909 if (tree_block_processed(bytenr, rc))
2910 return 0;
2911
2912 if (rb_simple_search(blocks, bytenr))
2913 return 0;
2914
2915 path = btrfs_alloc_path();
2916 if (!path)
2917 return -ENOMEM;
2918again:
2919 key.objectid = bytenr;
2920 if (skinny) {
2921 key.type = BTRFS_METADATA_ITEM_KEY;
2922 key.offset = (u64)-1;
2923 } else {
2924 key.type = BTRFS_EXTENT_ITEM_KEY;
2925 key.offset = blocksize;
2926 }
2927
2928 path->search_commit_root = 1;
2929 path->skip_locking = 1;
2930 ret = btrfs_search_slot(NULL, rc->extent_root, &key, path, 0, 0);
2931 if (ret < 0)
2932 goto out;
2933
2934 if (ret > 0 && skinny) {
2935 if (path->slots[0]) {
2936 path->slots[0]--;
2937 btrfs_item_key_to_cpu(path->nodes[0], &key,
2938 path->slots[0]);
2939 if (key.objectid == bytenr &&
2940 (key.type == BTRFS_METADATA_ITEM_KEY ||
2941 (key.type == BTRFS_EXTENT_ITEM_KEY &&
2942 key.offset == blocksize)))
2943 ret = 0;
2944 }
2945
2946 if (ret) {
2947 skinny = false;
2948 btrfs_release_path(path);
2949 goto again;
2950 }
2951 }
2952 if (ret) {
2953 ASSERT(ret == 1);
2954 btrfs_print_leaf(path->nodes[0]);
2955 btrfs_err(fs_info,
2956 "tree block extent item (%llu) is not found in extent tree",
2957 bytenr);
2958 WARN_ON(1);
2959 ret = -EINVAL;
2960 goto out;
2961 }
2962
2963 ret = add_tree_block(rc, &key, path, blocks);
2964out:
2965 btrfs_free_path(path);
2966 return ret;
2967}
2968
2969static int delete_block_group_cache(struct btrfs_fs_info *fs_info,
2970 struct btrfs_block_group *block_group,
2971 struct inode *inode,
2972 u64 ino)
2973{
2974 struct btrfs_root *root = fs_info->tree_root;
2975 struct btrfs_trans_handle *trans;
2976 int ret = 0;
2977
2978 if (inode)
2979 goto truncate;
2980
2981 inode = btrfs_iget(fs_info->sb, ino, root);
2982 if (IS_ERR(inode))
2983 return -ENOENT;
2984
2985truncate:
2986 ret = btrfs_check_trunc_cache_free_space(fs_info,
2987 &fs_info->global_block_rsv);
2988 if (ret)
2989 goto out;
2990
2991 trans = btrfs_join_transaction(root);
2992 if (IS_ERR(trans)) {
2993 ret = PTR_ERR(trans);
2994 goto out;
2995 }
2996
2997 ret = btrfs_truncate_free_space_cache(trans, block_group, inode);
2998
2999 btrfs_end_transaction(trans);
3000 btrfs_btree_balance_dirty(fs_info);
3001out:
3002 iput(inode);
3003 return ret;
3004}
3005
3006/*
3007 * Locate the free space cache EXTENT_DATA in root tree leaf and delete the
3008 * cache inode, to avoid free space cache data extent blocking data relocation.
3009 */
3010static int delete_v1_space_cache(struct extent_buffer *leaf,
3011 struct btrfs_block_group *block_group,
3012 u64 data_bytenr)
3013{
3014 u64 space_cache_ino;
3015 struct btrfs_file_extent_item *ei;
3016 struct btrfs_key key;
3017 bool found = false;
3018 int i;
3019 int ret;
3020
3021 if (btrfs_header_owner(leaf) != BTRFS_ROOT_TREE_OBJECTID)
3022 return 0;
3023
3024 for (i = 0; i < btrfs_header_nritems(leaf); i++) {
3025 btrfs_item_key_to_cpu(leaf, &key, i);
3026 if (key.type != BTRFS_EXTENT_DATA_KEY)
3027 continue;
3028 ei = btrfs_item_ptr(leaf, i, struct btrfs_file_extent_item);
3029 if (btrfs_file_extent_type(leaf, ei) == BTRFS_FILE_EXTENT_REG &&
3030 btrfs_file_extent_disk_bytenr(leaf, ei) == data_bytenr) {
3031 found = true;
3032 space_cache_ino = key.objectid;
3033 break;
3034 }
3035 }
3036 if (!found)
3037 return -ENOENT;
3038 ret = delete_block_group_cache(leaf->fs_info, block_group, NULL,
3039 space_cache_ino);
3040 return ret;
3041}
3042
3043/*
3044 * helper to find all tree blocks that reference a given data extent
3045 */
3046static noinline_for_stack
3047int add_data_references(struct reloc_control *rc,
3048 struct btrfs_key *extent_key,
3049 struct btrfs_path *path,
3050 struct rb_root *blocks)
3051{
3052 struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
3053 struct ulist *leaves = NULL;
3054 struct ulist_iterator leaf_uiter;
3055 struct ulist_node *ref_node = NULL;
3056 const u32 blocksize = fs_info->nodesize;
3057 int ret = 0;
3058
3059 btrfs_release_path(path);
3060 ret = btrfs_find_all_leafs(NULL, fs_info, extent_key->objectid,
3061 0, &leaves, NULL, true);
3062 if (ret < 0)
3063 return ret;
3064
3065 ULIST_ITER_INIT(&leaf_uiter);
3066 while ((ref_node = ulist_next(leaves, &leaf_uiter))) {
3067 struct extent_buffer *eb;
3068
3069 eb = read_tree_block(fs_info, ref_node->val, 0, 0, NULL);
3070 if (IS_ERR(eb)) {
3071 ret = PTR_ERR(eb);
3072 break;
3073 }
3074 ret = delete_v1_space_cache(eb, rc->block_group,
3075 extent_key->objectid);
3076 free_extent_buffer(eb);
3077 if (ret < 0)
3078 break;
3079 ret = __add_tree_block(rc, ref_node->val, blocksize, blocks);
3080 if (ret < 0)
3081 break;
3082 }
3083 if (ret < 0)
3084 free_block_list(blocks);
3085 ulist_free(leaves);
3086 return ret;
3087}
3088
3089/*
3090 * helper to find next unprocessed extent
3091 */
3092static noinline_for_stack
3093int find_next_extent(struct reloc_control *rc, struct btrfs_path *path,
3094 struct btrfs_key *extent_key)
3095{
3096 struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
3097 struct btrfs_key key;
3098 struct extent_buffer *leaf;
3099 u64 start, end, last;
3100 int ret;
3101
3102 last = rc->block_group->start + rc->block_group->length;
3103 while (1) {
3104 cond_resched();
3105 if (rc->search_start >= last) {
3106 ret = 1;
3107 break;
3108 }
3109
3110 key.objectid = rc->search_start;
3111 key.type = BTRFS_EXTENT_ITEM_KEY;
3112 key.offset = 0;
3113
3114 path->search_commit_root = 1;
3115 path->skip_locking = 1;
3116 ret = btrfs_search_slot(NULL, rc->extent_root, &key, path,
3117 0, 0);
3118 if (ret < 0)
3119 break;
3120next:
3121 leaf = path->nodes[0];
3122 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
3123 ret = btrfs_next_leaf(rc->extent_root, path);
3124 if (ret != 0)
3125 break;
3126 leaf = path->nodes[0];
3127 }
3128
3129 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
3130 if (key.objectid >= last) {
3131 ret = 1;
3132 break;
3133 }
3134
3135 if (key.type != BTRFS_EXTENT_ITEM_KEY &&
3136 key.type != BTRFS_METADATA_ITEM_KEY) {
3137 path->slots[0]++;
3138 goto next;
3139 }
3140
3141 if (key.type == BTRFS_EXTENT_ITEM_KEY &&
3142 key.objectid + key.offset <= rc->search_start) {
3143 path->slots[0]++;
3144 goto next;
3145 }
3146
3147 if (key.type == BTRFS_METADATA_ITEM_KEY &&
3148 key.objectid + fs_info->nodesize <=
3149 rc->search_start) {
3150 path->slots[0]++;
3151 goto next;
3152 }
3153
3154 ret = find_first_extent_bit(&rc->processed_blocks,
3155 key.objectid, &start, &end,
3156 EXTENT_DIRTY, NULL);
3157
3158 if (ret == 0 && start <= key.objectid) {
3159 btrfs_release_path(path);
3160 rc->search_start = end + 1;
3161 } else {
3162 if (key.type == BTRFS_EXTENT_ITEM_KEY)
3163 rc->search_start = key.objectid + key.offset;
3164 else
3165 rc->search_start = key.objectid +
3166 fs_info->nodesize;
3167 memcpy(extent_key, &key, sizeof(key));
3168 return 0;
3169 }
3170 }
3171 btrfs_release_path(path);
3172 return ret;
3173}
3174
3175static void set_reloc_control(struct reloc_control *rc)
3176{
3177 struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
3178
3179 mutex_lock(&fs_info->reloc_mutex);
3180 fs_info->reloc_ctl = rc;
3181 mutex_unlock(&fs_info->reloc_mutex);
3182}
3183
3184static void unset_reloc_control(struct reloc_control *rc)
3185{
3186 struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
3187
3188 mutex_lock(&fs_info->reloc_mutex);
3189 fs_info->reloc_ctl = NULL;
3190 mutex_unlock(&fs_info->reloc_mutex);
3191}
3192
3193static int check_extent_flags(u64 flags)
3194{
3195 if ((flags & BTRFS_EXTENT_FLAG_DATA) &&
3196 (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK))
3197 return 1;
3198 if (!(flags & BTRFS_EXTENT_FLAG_DATA) &&
3199 !(flags & BTRFS_EXTENT_FLAG_TREE_BLOCK))
3200 return 1;
3201 if ((flags & BTRFS_EXTENT_FLAG_DATA) &&
3202 (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
3203 return 1;
3204 return 0;
3205}
3206
3207static noinline_for_stack
3208int prepare_to_relocate(struct reloc_control *rc)
3209{
3210 struct btrfs_trans_handle *trans;
3211 int ret;
3212
3213 rc->block_rsv = btrfs_alloc_block_rsv(rc->extent_root->fs_info,
3214 BTRFS_BLOCK_RSV_TEMP);
3215 if (!rc->block_rsv)
3216 return -ENOMEM;
3217
3218 memset(&rc->cluster, 0, sizeof(rc->cluster));
3219 rc->search_start = rc->block_group->start;
3220 rc->extents_found = 0;
3221 rc->nodes_relocated = 0;
3222 rc->merging_rsv_size = 0;
3223 rc->reserved_bytes = 0;
3224 rc->block_rsv->size = rc->extent_root->fs_info->nodesize *
3225 RELOCATION_RESERVED_NODES;
3226 ret = btrfs_block_rsv_refill(rc->extent_root,
3227 rc->block_rsv, rc->block_rsv->size,
3228 BTRFS_RESERVE_FLUSH_ALL);
3229 if (ret)
3230 return ret;
3231
3232 rc->create_reloc_tree = 1;
3233 set_reloc_control(rc);
3234
3235 trans = btrfs_join_transaction(rc->extent_root);
3236 if (IS_ERR(trans)) {
3237 unset_reloc_control(rc);
3238 /*
3239 * extent tree is not a ref_cow tree and has no reloc_root to
3240 * cleanup. And callers are responsible to free the above
3241 * block rsv.
3242 */
3243 return PTR_ERR(trans);
3244 }
3245 btrfs_commit_transaction(trans);
3246 return 0;
3247}
3248
3249static noinline_for_stack int relocate_block_group(struct reloc_control *rc)
3250{
3251 struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
3252 struct rb_root blocks = RB_ROOT;
3253 struct btrfs_key key;
3254 struct btrfs_trans_handle *trans = NULL;
3255 struct btrfs_path *path;
3256 struct btrfs_extent_item *ei;
3257 u64 flags;
3258 u32 item_size;
3259 int ret;
3260 int err = 0;
3261 int progress = 0;
3262
3263 path = btrfs_alloc_path();
3264 if (!path)
3265 return -ENOMEM;
3266 path->reada = READA_FORWARD;
3267
3268 ret = prepare_to_relocate(rc);
3269 if (ret) {
3270 err = ret;
3271 goto out_free;
3272 }
3273
3274 while (1) {
3275 rc->reserved_bytes = 0;
3276 ret = btrfs_block_rsv_refill(rc->extent_root,
3277 rc->block_rsv, rc->block_rsv->size,
3278 BTRFS_RESERVE_FLUSH_ALL);
3279 if (ret) {
3280 err = ret;
3281 break;
3282 }
3283 progress++;
3284 trans = btrfs_start_transaction(rc->extent_root, 0);
3285 if (IS_ERR(trans)) {
3286 err = PTR_ERR(trans);
3287 trans = NULL;
3288 break;
3289 }
3290restart:
3291 if (update_backref_cache(trans, &rc->backref_cache)) {
3292 btrfs_end_transaction(trans);
3293 trans = NULL;
3294 continue;
3295 }
3296
3297 ret = find_next_extent(rc, path, &key);
3298 if (ret < 0)
3299 err = ret;
3300 if (ret != 0)
3301 break;
3302
3303 rc->extents_found++;
3304
3305 ei = btrfs_item_ptr(path->nodes[0], path->slots[0],
3306 struct btrfs_extent_item);
3307 item_size = btrfs_item_size_nr(path->nodes[0], path->slots[0]);
3308 if (item_size >= sizeof(*ei)) {
3309 flags = btrfs_extent_flags(path->nodes[0], ei);
3310 ret = check_extent_flags(flags);
3311 BUG_ON(ret);
3312 } else if (unlikely(item_size == sizeof(struct btrfs_extent_item_v0))) {
3313 err = -EINVAL;
3314 btrfs_print_v0_err(trans->fs_info);
3315 btrfs_abort_transaction(trans, err);
3316 break;
3317 } else {
3318 BUG();
3319 }
3320
3321 if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
3322 ret = add_tree_block(rc, &key, path, &blocks);
3323 } else if (rc->stage == UPDATE_DATA_PTRS &&
3324 (flags & BTRFS_EXTENT_FLAG_DATA)) {
3325 ret = add_data_references(rc, &key, path, &blocks);
3326 } else {
3327 btrfs_release_path(path);
3328 ret = 0;
3329 }
3330 if (ret < 0) {
3331 err = ret;
3332 break;
3333 }
3334
3335 if (!RB_EMPTY_ROOT(&blocks)) {
3336 ret = relocate_tree_blocks(trans, rc, &blocks);
3337 if (ret < 0) {
3338 if (ret != -EAGAIN) {
3339 err = ret;
3340 break;
3341 }
3342 rc->extents_found--;
3343 rc->search_start = key.objectid;
3344 }
3345 }
3346
3347 btrfs_end_transaction_throttle(trans);
3348 btrfs_btree_balance_dirty(fs_info);
3349 trans = NULL;
3350
3351 if (rc->stage == MOVE_DATA_EXTENTS &&
3352 (flags & BTRFS_EXTENT_FLAG_DATA)) {
3353 rc->found_file_extent = 1;
3354 ret = relocate_data_extent(rc->data_inode,
3355 &key, &rc->cluster);
3356 if (ret < 0) {
3357 err = ret;
3358 break;
3359 }
3360 }
3361 if (btrfs_should_cancel_balance(fs_info)) {
3362 err = -ECANCELED;
3363 break;
3364 }
3365 }
3366 if (trans && progress && err == -ENOSPC) {
3367 ret = btrfs_force_chunk_alloc(trans, rc->block_group->flags);
3368 if (ret == 1) {
3369 err = 0;
3370 progress = 0;
3371 goto restart;
3372 }
3373 }
3374
3375 btrfs_release_path(path);
3376 clear_extent_bits(&rc->processed_blocks, 0, (u64)-1, EXTENT_DIRTY);
3377
3378 if (trans) {
3379 btrfs_end_transaction_throttle(trans);
3380 btrfs_btree_balance_dirty(fs_info);
3381 }
3382
3383 if (!err) {
3384 ret = relocate_file_extent_cluster(rc->data_inode,
3385 &rc->cluster);
3386 if (ret < 0)
3387 err = ret;
3388 }
3389
3390 rc->create_reloc_tree = 0;
3391 set_reloc_control(rc);
3392
3393 btrfs_backref_release_cache(&rc->backref_cache);
3394 btrfs_block_rsv_release(fs_info, rc->block_rsv, (u64)-1, NULL);
3395
3396 /*
3397 * Even in the case when the relocation is cancelled, we should all go
3398 * through prepare_to_merge() and merge_reloc_roots().
3399 *
3400 * For error (including cancelled balance), prepare_to_merge() will
3401 * mark all reloc trees orphan, then queue them for cleanup in
3402 * merge_reloc_roots()
3403 */
3404 err = prepare_to_merge(rc, err);
3405
3406 merge_reloc_roots(rc);
3407
3408 rc->merge_reloc_tree = 0;
3409 unset_reloc_control(rc);
3410 btrfs_block_rsv_release(fs_info, rc->block_rsv, (u64)-1, NULL);
3411
3412 /* get rid of pinned extents */
3413 trans = btrfs_join_transaction(rc->extent_root);
3414 if (IS_ERR(trans)) {
3415 err = PTR_ERR(trans);
3416 goto out_free;
3417 }
3418 btrfs_commit_transaction(trans);
3419out_free:
3420 ret = clean_dirty_subvols(rc);
3421 if (ret < 0 && !err)
3422 err = ret;
3423 btrfs_free_block_rsv(fs_info, rc->block_rsv);
3424 btrfs_free_path(path);
3425 return err;
3426}
3427
3428static int __insert_orphan_inode(struct btrfs_trans_handle *trans,
3429 struct btrfs_root *root, u64 objectid)
3430{
3431 struct btrfs_path *path;
3432 struct btrfs_inode_item *item;
3433 struct extent_buffer *leaf;
3434 int ret;
3435
3436 path = btrfs_alloc_path();
3437 if (!path)
3438 return -ENOMEM;
3439
3440 ret = btrfs_insert_empty_inode(trans, root, path, objectid);
3441 if (ret)
3442 goto out;
3443
3444 leaf = path->nodes[0];
3445 item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_inode_item);
3446 memzero_extent_buffer(leaf, (unsigned long)item, sizeof(*item));
3447 btrfs_set_inode_generation(leaf, item, 1);
3448 btrfs_set_inode_size(leaf, item, 0);
3449 btrfs_set_inode_mode(leaf, item, S_IFREG | 0600);
3450 btrfs_set_inode_flags(leaf, item, BTRFS_INODE_NOCOMPRESS |
3451 BTRFS_INODE_PREALLOC);
3452 btrfs_mark_buffer_dirty(leaf);
3453out:
3454 btrfs_free_path(path);
3455 return ret;
3456}
3457
3458/*
3459 * helper to create inode for data relocation.
3460 * the inode is in data relocation tree and its link count is 0
3461 */
3462static noinline_for_stack
3463struct inode *create_reloc_inode(struct btrfs_fs_info *fs_info,
3464 struct btrfs_block_group *group)
3465{
3466 struct inode *inode = NULL;
3467 struct btrfs_trans_handle *trans;
3468 struct btrfs_root *root;
3469 u64 objectid;
3470 int err = 0;
3471
3472 root = btrfs_grab_root(fs_info->data_reloc_root);
3473 trans = btrfs_start_transaction(root, 6);
3474 if (IS_ERR(trans)) {
3475 btrfs_put_root(root);
3476 return ERR_CAST(trans);
3477 }
3478
3479 err = btrfs_find_free_objectid(root, &objectid);
3480 if (err)
3481 goto out;
3482
3483 err = __insert_orphan_inode(trans, root, objectid);
3484 BUG_ON(err);
3485
3486 inode = btrfs_iget(fs_info->sb, objectid, root);
3487 BUG_ON(IS_ERR(inode));
3488 BTRFS_I(inode)->index_cnt = group->start;
3489
3490 err = btrfs_orphan_add(trans, BTRFS_I(inode));
3491out:
3492 btrfs_put_root(root);
3493 btrfs_end_transaction(trans);
3494 btrfs_btree_balance_dirty(fs_info);
3495 if (err) {
3496 if (inode)
3497 iput(inode);
3498 inode = ERR_PTR(err);
3499 }
3500 return inode;
3501}
3502
3503static struct reloc_control *alloc_reloc_control(struct btrfs_fs_info *fs_info)
3504{
3505 struct reloc_control *rc;
3506
3507 rc = kzalloc(sizeof(*rc), GFP_NOFS);
3508 if (!rc)
3509 return NULL;
3510
3511 INIT_LIST_HEAD(&rc->reloc_roots);
3512 INIT_LIST_HEAD(&rc->dirty_subvol_roots);
3513 btrfs_backref_init_cache(fs_info, &rc->backref_cache, 1);
3514 mapping_tree_init(&rc->reloc_root_tree);
3515 extent_io_tree_init(fs_info, &rc->processed_blocks,
3516 IO_TREE_RELOC_BLOCKS, NULL);
3517 return rc;
3518}
3519
3520static void free_reloc_control(struct reloc_control *rc)
3521{
3522 struct mapping_node *node, *tmp;
3523
3524 free_reloc_roots(&rc->reloc_roots);
3525 rbtree_postorder_for_each_entry_safe(node, tmp,
3526 &rc->reloc_root_tree.rb_root, rb_node)
3527 kfree(node);
3528
3529 kfree(rc);
3530}
3531
3532/*
3533 * Print the block group being relocated
3534 */
3535static void describe_relocation(struct btrfs_fs_info *fs_info,
3536 struct btrfs_block_group *block_group)
3537{
3538 char buf[128] = {'\0'};
3539
3540 btrfs_describe_block_groups(block_group->flags, buf, sizeof(buf));
3541
3542 btrfs_info(fs_info,
3543 "relocating block group %llu flags %s",
3544 block_group->start, buf);
3545}
3546
3547static const char *stage_to_string(int stage)
3548{
3549 if (stage == MOVE_DATA_EXTENTS)
3550 return "move data extents";
3551 if (stage == UPDATE_DATA_PTRS)
3552 return "update data pointers";
3553 return "unknown";
3554}
3555
3556/*
3557 * function to relocate all extents in a block group.
3558 */
3559int btrfs_relocate_block_group(struct btrfs_fs_info *fs_info, u64 group_start)
3560{
3561 struct btrfs_block_group *bg;
3562 struct btrfs_root *extent_root = fs_info->extent_root;
3563 struct reloc_control *rc;
3564 struct inode *inode;
3565 struct btrfs_path *path;
3566 int ret;
3567 int rw = 0;
3568 int err = 0;
3569
3570 bg = btrfs_lookup_block_group(fs_info, group_start);
3571 if (!bg)
3572 return -ENOENT;
3573
3574 if (btrfs_pinned_by_swapfile(fs_info, bg)) {
3575 btrfs_put_block_group(bg);
3576 return -ETXTBSY;
3577 }
3578
3579 rc = alloc_reloc_control(fs_info);
3580 if (!rc) {
3581 btrfs_put_block_group(bg);
3582 return -ENOMEM;
3583 }
3584
3585 rc->extent_root = extent_root;
3586 rc->block_group = bg;
3587
3588 ret = btrfs_inc_block_group_ro(rc->block_group, true);
3589 if (ret) {
3590 err = ret;
3591 goto out;
3592 }
3593 rw = 1;
3594
3595 path = btrfs_alloc_path();
3596 if (!path) {
3597 err = -ENOMEM;
3598 goto out;
3599 }
3600
3601 inode = lookup_free_space_inode(rc->block_group, path);
3602 btrfs_free_path(path);
3603
3604 if (!IS_ERR(inode))
3605 ret = delete_block_group_cache(fs_info, rc->block_group, inode, 0);
3606 else
3607 ret = PTR_ERR(inode);
3608
3609 if (ret && ret != -ENOENT) {
3610 err = ret;
3611 goto out;
3612 }
3613
3614 rc->data_inode = create_reloc_inode(fs_info, rc->block_group);
3615 if (IS_ERR(rc->data_inode)) {
3616 err = PTR_ERR(rc->data_inode);
3617 rc->data_inode = NULL;
3618 goto out;
3619 }
3620
3621 describe_relocation(fs_info, rc->block_group);
3622
3623 btrfs_wait_block_group_reservations(rc->block_group);
3624 btrfs_wait_nocow_writers(rc->block_group);
3625 btrfs_wait_ordered_roots(fs_info, U64_MAX,
3626 rc->block_group->start,
3627 rc->block_group->length);
3628
3629 while (1) {
3630 int finishes_stage;
3631
3632 mutex_lock(&fs_info->cleaner_mutex);
3633 ret = relocate_block_group(rc);
3634 mutex_unlock(&fs_info->cleaner_mutex);
3635 if (ret < 0)
3636 err = ret;
3637
3638 finishes_stage = rc->stage;
3639 /*
3640 * We may have gotten ENOSPC after we already dirtied some
3641 * extents. If writeout happens while we're relocating a
3642 * different block group we could end up hitting the
3643 * BUG_ON(rc->stage == UPDATE_DATA_PTRS) in
3644 * btrfs_reloc_cow_block. Make sure we write everything out
3645 * properly so we don't trip over this problem, and then break
3646 * out of the loop if we hit an error.
3647 */
3648 if (rc->stage == MOVE_DATA_EXTENTS && rc->found_file_extent) {
3649 ret = btrfs_wait_ordered_range(rc->data_inode, 0,
3650 (u64)-1);
3651 if (ret)
3652 err = ret;
3653 invalidate_mapping_pages(rc->data_inode->i_mapping,
3654 0, -1);
3655 rc->stage = UPDATE_DATA_PTRS;
3656 }
3657
3658 if (err < 0)
3659 goto out;
3660
3661 if (rc->extents_found == 0)
3662 break;
3663
3664 btrfs_info(fs_info, "found %llu extents, stage: %s",
3665 rc->extents_found, stage_to_string(finishes_stage));
3666 }
3667
3668 WARN_ON(rc->block_group->pinned > 0);
3669 WARN_ON(rc->block_group->reserved > 0);
3670 WARN_ON(rc->block_group->used > 0);
3671out:
3672 if (err && rw)
3673 btrfs_dec_block_group_ro(rc->block_group);
3674 iput(rc->data_inode);
3675 btrfs_put_block_group(rc->block_group);
3676 free_reloc_control(rc);
3677 return err;
3678}
3679
3680static noinline_for_stack int mark_garbage_root(struct btrfs_root *root)
3681{
3682 struct btrfs_fs_info *fs_info = root->fs_info;
3683 struct btrfs_trans_handle *trans;
3684 int ret, err;
3685
3686 trans = btrfs_start_transaction(fs_info->tree_root, 0);
3687 if (IS_ERR(trans))
3688 return PTR_ERR(trans);
3689
3690 memset(&root->root_item.drop_progress, 0,
3691 sizeof(root->root_item.drop_progress));
3692 root->root_item.drop_level = 0;
3693 btrfs_set_root_refs(&root->root_item, 0);
3694 ret = btrfs_update_root(trans, fs_info->tree_root,
3695 &root->root_key, &root->root_item);
3696
3697 err = btrfs_end_transaction(trans);
3698 if (err)
3699 return err;
3700 return ret;
3701}
3702
3703/*
3704 * recover relocation interrupted by system crash.
3705 *
3706 * this function resumes merging reloc trees with corresponding fs trees.
3707 * this is important for keeping the sharing of tree blocks
3708 */
3709int btrfs_recover_relocation(struct btrfs_root *root)
3710{
3711 struct btrfs_fs_info *fs_info = root->fs_info;
3712 LIST_HEAD(reloc_roots);
3713 struct btrfs_key key;
3714 struct btrfs_root *fs_root;
3715 struct btrfs_root *reloc_root;
3716 struct btrfs_path *path;
3717 struct extent_buffer *leaf;
3718 struct reloc_control *rc = NULL;
3719 struct btrfs_trans_handle *trans;
3720 int ret;
3721 int err = 0;
3722
3723 path = btrfs_alloc_path();
3724 if (!path)
3725 return -ENOMEM;
3726 path->reada = READA_BACK;
3727
3728 key.objectid = BTRFS_TREE_RELOC_OBJECTID;
3729 key.type = BTRFS_ROOT_ITEM_KEY;
3730 key.offset = (u64)-1;
3731
3732 while (1) {
3733 ret = btrfs_search_slot(NULL, fs_info->tree_root, &key,
3734 path, 0, 0);
3735 if (ret < 0) {
3736 err = ret;
3737 goto out;
3738 }
3739 if (ret > 0) {
3740 if (path->slots[0] == 0)
3741 break;
3742 path->slots[0]--;
3743 }
3744 leaf = path->nodes[0];
3745 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
3746 btrfs_release_path(path);
3747
3748 if (key.objectid != BTRFS_TREE_RELOC_OBJECTID ||
3749 key.type != BTRFS_ROOT_ITEM_KEY)
3750 break;
3751
3752 reloc_root = btrfs_read_tree_root(root, &key);
3753 if (IS_ERR(reloc_root)) {
3754 err = PTR_ERR(reloc_root);
3755 goto out;
3756 }
3757
3758 set_bit(BTRFS_ROOT_SHAREABLE, &reloc_root->state);
3759 list_add(&reloc_root->root_list, &reloc_roots);
3760
3761 if (btrfs_root_refs(&reloc_root->root_item) > 0) {
3762 fs_root = btrfs_get_fs_root(fs_info,
3763 reloc_root->root_key.offset, false);
3764 if (IS_ERR(fs_root)) {
3765 ret = PTR_ERR(fs_root);
3766 if (ret != -ENOENT) {
3767 err = ret;
3768 goto out;
3769 }
3770 ret = mark_garbage_root(reloc_root);
3771 if (ret < 0) {
3772 err = ret;
3773 goto out;
3774 }
3775 } else {
3776 btrfs_put_root(fs_root);
3777 }
3778 }
3779
3780 if (key.offset == 0)
3781 break;
3782
3783 key.offset--;
3784 }
3785 btrfs_release_path(path);
3786
3787 if (list_empty(&reloc_roots))
3788 goto out;
3789
3790 rc = alloc_reloc_control(fs_info);
3791 if (!rc) {
3792 err = -ENOMEM;
3793 goto out;
3794 }
3795
3796 rc->extent_root = fs_info->extent_root;
3797
3798 set_reloc_control(rc);
3799
3800 trans = btrfs_join_transaction(rc->extent_root);
3801 if (IS_ERR(trans)) {
3802 err = PTR_ERR(trans);
3803 goto out_unset;
3804 }
3805
3806 rc->merge_reloc_tree = 1;
3807
3808 while (!list_empty(&reloc_roots)) {
3809 reloc_root = list_entry(reloc_roots.next,
3810 struct btrfs_root, root_list);
3811 list_del(&reloc_root->root_list);
3812
3813 if (btrfs_root_refs(&reloc_root->root_item) == 0) {
3814 list_add_tail(&reloc_root->root_list,
3815 &rc->reloc_roots);
3816 continue;
3817 }
3818
3819 fs_root = btrfs_get_fs_root(fs_info, reloc_root->root_key.offset,
3820 false);
3821 if (IS_ERR(fs_root)) {
3822 err = PTR_ERR(fs_root);
3823 list_add_tail(&reloc_root->root_list, &reloc_roots);
3824 btrfs_end_transaction(trans);
3825 goto out_unset;
3826 }
3827
3828 err = __add_reloc_root(reloc_root);
3829 BUG_ON(err < 0); /* -ENOMEM or logic error */
3830 fs_root->reloc_root = btrfs_grab_root(reloc_root);
3831 btrfs_put_root(fs_root);
3832 }
3833
3834 err = btrfs_commit_transaction(trans);
3835 if (err)
3836 goto out_unset;
3837
3838 merge_reloc_roots(rc);
3839
3840 unset_reloc_control(rc);
3841
3842 trans = btrfs_join_transaction(rc->extent_root);
3843 if (IS_ERR(trans)) {
3844 err = PTR_ERR(trans);
3845 goto out_clean;
3846 }
3847 err = btrfs_commit_transaction(trans);
3848out_clean:
3849 ret = clean_dirty_subvols(rc);
3850 if (ret < 0 && !err)
3851 err = ret;
3852out_unset:
3853 unset_reloc_control(rc);
3854 free_reloc_control(rc);
3855out:
3856 free_reloc_roots(&reloc_roots);
3857
3858 btrfs_free_path(path);
3859
3860 if (err == 0) {
3861 /* cleanup orphan inode in data relocation tree */
3862 fs_root = btrfs_grab_root(fs_info->data_reloc_root);
3863 ASSERT(fs_root);
3864 err = btrfs_orphan_cleanup(fs_root);
3865 btrfs_put_root(fs_root);
3866 }
3867 return err;
3868}
3869
3870/*
3871 * helper to add ordered checksum for data relocation.
3872 *
3873 * cloning checksum properly handles the nodatasum extents.
3874 * it also saves CPU time to re-calculate the checksum.
3875 */
3876int btrfs_reloc_clone_csums(struct btrfs_inode *inode, u64 file_pos, u64 len)
3877{
3878 struct btrfs_fs_info *fs_info = inode->root->fs_info;
3879 struct btrfs_ordered_sum *sums;
3880 struct btrfs_ordered_extent *ordered;
3881 int ret;
3882 u64 disk_bytenr;
3883 u64 new_bytenr;
3884 LIST_HEAD(list);
3885
3886 ordered = btrfs_lookup_ordered_extent(inode, file_pos);
3887 BUG_ON(ordered->file_offset != file_pos || ordered->num_bytes != len);
3888
3889 disk_bytenr = file_pos + inode->index_cnt;
3890 ret = btrfs_lookup_csums_range(fs_info->csum_root, disk_bytenr,
3891 disk_bytenr + len - 1, &list, 0);
3892 if (ret)
3893 goto out;
3894
3895 while (!list_empty(&list)) {
3896 sums = list_entry(list.next, struct btrfs_ordered_sum, list);
3897 list_del_init(&sums->list);
3898
3899 /*
3900 * We need to offset the new_bytenr based on where the csum is.
3901 * We need to do this because we will read in entire prealloc
3902 * extents but we may have written to say the middle of the
3903 * prealloc extent, so we need to make sure the csum goes with
3904 * the right disk offset.
3905 *
3906 * We can do this because the data reloc inode refers strictly
3907 * to the on disk bytes, so we don't have to worry about
3908 * disk_len vs real len like with real inodes since it's all
3909 * disk length.
3910 */
3911 new_bytenr = ordered->disk_bytenr + sums->bytenr - disk_bytenr;
3912 sums->bytenr = new_bytenr;
3913
3914 btrfs_add_ordered_sum(ordered, sums);
3915 }
3916out:
3917 btrfs_put_ordered_extent(ordered);
3918 return ret;
3919}
3920
3921int btrfs_reloc_cow_block(struct btrfs_trans_handle *trans,
3922 struct btrfs_root *root, struct extent_buffer *buf,
3923 struct extent_buffer *cow)
3924{
3925 struct btrfs_fs_info *fs_info = root->fs_info;
3926 struct reloc_control *rc;
3927 struct btrfs_backref_node *node;
3928 int first_cow = 0;
3929 int level;
3930 int ret = 0;
3931
3932 rc = fs_info->reloc_ctl;
3933 if (!rc)
3934 return 0;
3935
3936 BUG_ON(rc->stage == UPDATE_DATA_PTRS &&
3937 root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID);
3938
3939 level = btrfs_header_level(buf);
3940 if (btrfs_header_generation(buf) <=
3941 btrfs_root_last_snapshot(&root->root_item))
3942 first_cow = 1;
3943
3944 if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID &&
3945 rc->create_reloc_tree) {
3946 WARN_ON(!first_cow && level == 0);
3947
3948 node = rc->backref_cache.path[level];
3949 BUG_ON(node->bytenr != buf->start &&
3950 node->new_bytenr != buf->start);
3951
3952 btrfs_backref_drop_node_buffer(node);
3953 atomic_inc(&cow->refs);
3954 node->eb = cow;
3955 node->new_bytenr = cow->start;
3956
3957 if (!node->pending) {
3958 list_move_tail(&node->list,
3959 &rc->backref_cache.pending[level]);
3960 node->pending = 1;
3961 }
3962
3963 if (first_cow)
3964 mark_block_processed(rc, node);
3965
3966 if (first_cow && level > 0)
3967 rc->nodes_relocated += buf->len;
3968 }
3969
3970 if (level == 0 && first_cow && rc->stage == UPDATE_DATA_PTRS)
3971 ret = replace_file_extents(trans, rc, root, cow);
3972 return ret;
3973}
3974
3975/*
3976 * called before creating snapshot. it calculates metadata reservation
3977 * required for relocating tree blocks in the snapshot
3978 */
3979void btrfs_reloc_pre_snapshot(struct btrfs_pending_snapshot *pending,
3980 u64 *bytes_to_reserve)
3981{
3982 struct btrfs_root *root = pending->root;
3983 struct reloc_control *rc = root->fs_info->reloc_ctl;
3984
3985 if (!rc || !have_reloc_root(root))
3986 return;
3987
3988 if (!rc->merge_reloc_tree)
3989 return;
3990
3991 root = root->reloc_root;
3992 BUG_ON(btrfs_root_refs(&root->root_item) == 0);
3993 /*
3994 * relocation is in the stage of merging trees. the space
3995 * used by merging a reloc tree is twice the size of
3996 * relocated tree nodes in the worst case. half for cowing
3997 * the reloc tree, half for cowing the fs tree. the space
3998 * used by cowing the reloc tree will be freed after the
3999 * tree is dropped. if we create snapshot, cowing the fs
4000 * tree may use more space than it frees. so we need
4001 * reserve extra space.
4002 */
4003 *bytes_to_reserve += rc->nodes_relocated;
4004}
4005
4006/*
4007 * called after snapshot is created. migrate block reservation
4008 * and create reloc root for the newly created snapshot
4009 *
4010 * This is similar to btrfs_init_reloc_root(), we come out of here with two
4011 * references held on the reloc_root, one for root->reloc_root and one for
4012 * rc->reloc_roots.
4013 */
4014int btrfs_reloc_post_snapshot(struct btrfs_trans_handle *trans,
4015 struct btrfs_pending_snapshot *pending)
4016{
4017 struct btrfs_root *root = pending->root;
4018 struct btrfs_root *reloc_root;
4019 struct btrfs_root *new_root;
4020 struct reloc_control *rc = root->fs_info->reloc_ctl;
4021 int ret;
4022
4023 if (!rc || !have_reloc_root(root))
4024 return 0;
4025
4026 rc = root->fs_info->reloc_ctl;
4027 rc->merging_rsv_size += rc->nodes_relocated;
4028
4029 if (rc->merge_reloc_tree) {
4030 ret = btrfs_block_rsv_migrate(&pending->block_rsv,
4031 rc->block_rsv,
4032 rc->nodes_relocated, true);
4033 if (ret)
4034 return ret;
4035 }
4036
4037 new_root = pending->snap;
4038 reloc_root = create_reloc_root(trans, root->reloc_root,
4039 new_root->root_key.objectid);
4040 if (IS_ERR(reloc_root))
4041 return PTR_ERR(reloc_root);
4042
4043 ret = __add_reloc_root(reloc_root);
4044 BUG_ON(ret < 0);
4045 new_root->reloc_root = btrfs_grab_root(reloc_root);
4046
4047 if (rc->create_reloc_tree)
4048 ret = clone_backref_node(trans, rc, root, reloc_root);
4049 return ret;
4050}