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#include "raid-stripe-tree.h"
40
41/*
42 * Relocation overview
43 *
44 * [What does relocation do]
45 *
46 * The objective of relocation is to relocate all extents of the target block
47 * group to other block groups.
48 * This is utilized by resize (shrink only), profile converting, compacting
49 * space, or balance routine to spread chunks over devices.
50 *
51 * Before | After
52 * ------------------------------------------------------------------
53 * BG A: 10 data extents | BG A: deleted
54 * BG B: 2 data extents | BG B: 10 data extents (2 old + 8 relocated)
55 * BG C: 1 extents | BG C: 3 data extents (1 old + 2 relocated)
56 *
57 * [How does relocation work]
58 *
59 * 1. Mark the target block group read-only
60 * New extents won't be allocated from the target block group.
61 *
62 * 2.1 Record each extent in the target block group
63 * To build a proper map of extents to be relocated.
64 *
65 * 2.2 Build data reloc tree and reloc trees
66 * Data reloc tree will contain an inode, recording all newly relocated
67 * data extents.
68 * There will be only one data reloc tree for one data block group.
69 *
70 * Reloc tree will be a special snapshot of its source tree, containing
71 * relocated tree blocks.
72 * Each tree referring to a tree block in target block group will get its
73 * reloc tree built.
74 *
75 * 2.3 Swap source tree with its corresponding reloc tree
76 * Each involved tree only refers to new extents after swap.
77 *
78 * 3. Cleanup reloc trees and data reloc tree.
79 * As old extents in the target block group are still referenced by reloc
80 * trees, we need to clean them up before really freeing the target block
81 * group.
82 *
83 * The main complexity is in steps 2.2 and 2.3.
84 *
85 * The entry point of relocation is relocate_block_group() function.
86 */
87
88#define RELOCATION_RESERVED_NODES 256
89/*
90 * map address of tree root to tree
91 */
92struct mapping_node {
93 struct {
94 struct rb_node rb_node;
95 u64 bytenr;
96 }; /* Use rb_simle_node for search/insert */
97 void *data;
98};
99
100struct mapping_tree {
101 struct rb_root rb_root;
102 spinlock_t lock;
103};
104
105/*
106 * present a tree block to process
107 */
108struct tree_block {
109 struct {
110 struct rb_node rb_node;
111 u64 bytenr;
112 }; /* Use rb_simple_node for search/insert */
113 u64 owner;
114 struct btrfs_key key;
115 u8 level;
116 bool key_ready;
117};
118
119#define MAX_EXTENTS 128
120
121struct file_extent_cluster {
122 u64 start;
123 u64 end;
124 u64 boundary[MAX_EXTENTS];
125 unsigned int nr;
126 u64 owning_root;
127};
128
129/* Stages of data relocation. */
130enum reloc_stage {
131 MOVE_DATA_EXTENTS,
132 UPDATE_DATA_PTRS
133};
134
135struct reloc_control {
136 /* block group to relocate */
137 struct btrfs_block_group *block_group;
138 /* extent tree */
139 struct btrfs_root *extent_root;
140 /* inode for moving data */
141 struct inode *data_inode;
142
143 struct btrfs_block_rsv *block_rsv;
144
145 struct btrfs_backref_cache backref_cache;
146
147 struct file_extent_cluster cluster;
148 /* tree blocks have been processed */
149 struct extent_io_tree processed_blocks;
150 /* map start of tree root to corresponding reloc tree */
151 struct mapping_tree reloc_root_tree;
152 /* list of reloc trees */
153 struct list_head reloc_roots;
154 /* list of subvolume trees that get relocated */
155 struct list_head dirty_subvol_roots;
156 /* size of metadata reservation for merging reloc trees */
157 u64 merging_rsv_size;
158 /* size of relocated tree nodes */
159 u64 nodes_relocated;
160 /* reserved size for block group relocation*/
161 u64 reserved_bytes;
162
163 u64 search_start;
164 u64 extents_found;
165
166 enum reloc_stage stage;
167 bool create_reloc_tree;
168 bool merge_reloc_tree;
169 bool found_file_extent;
170};
171
172static void mark_block_processed(struct reloc_control *rc,
173 struct btrfs_backref_node *node)
174{
175 u32 blocksize;
176
177 if (node->level == 0 ||
178 in_range(node->bytenr, rc->block_group->start,
179 rc->block_group->length)) {
180 blocksize = rc->extent_root->fs_info->nodesize;
181 set_extent_bit(&rc->processed_blocks, node->bytenr,
182 node->bytenr + blocksize - 1, EXTENT_DIRTY, NULL);
183 }
184 node->processed = 1;
185}
186
187/*
188 * walk up backref nodes until reach node presents tree root
189 */
190static struct btrfs_backref_node *walk_up_backref(
191 struct btrfs_backref_node *node,
192 struct btrfs_backref_edge *edges[], int *index)
193{
194 struct btrfs_backref_edge *edge;
195 int idx = *index;
196
197 while (!list_empty(&node->upper)) {
198 edge = list_entry(node->upper.next,
199 struct btrfs_backref_edge, list[LOWER]);
200 edges[idx++] = edge;
201 node = edge->node[UPPER];
202 }
203 BUG_ON(node->detached);
204 *index = idx;
205 return node;
206}
207
208/*
209 * walk down backref nodes to find start of next reference path
210 */
211static struct btrfs_backref_node *walk_down_backref(
212 struct btrfs_backref_edge *edges[], int *index)
213{
214 struct btrfs_backref_edge *edge;
215 struct btrfs_backref_node *lower;
216 int idx = *index;
217
218 while (idx > 0) {
219 edge = edges[idx - 1];
220 lower = edge->node[LOWER];
221 if (list_is_last(&edge->list[LOWER], &lower->upper)) {
222 idx--;
223 continue;
224 }
225 edge = list_entry(edge->list[LOWER].next,
226 struct btrfs_backref_edge, list[LOWER]);
227 edges[idx - 1] = edge;
228 *index = idx;
229 return edge->node[UPPER];
230 }
231 *index = 0;
232 return NULL;
233}
234
235static bool reloc_root_is_dead(const struct btrfs_root *root)
236{
237 /*
238 * Pair with set_bit/clear_bit in clean_dirty_subvols and
239 * btrfs_update_reloc_root. We need to see the updated bit before
240 * trying to access reloc_root
241 */
242 smp_rmb();
243 if (test_bit(BTRFS_ROOT_DEAD_RELOC_TREE, &root->state))
244 return true;
245 return false;
246}
247
248/*
249 * Check if this subvolume tree has valid reloc tree.
250 *
251 * Reloc tree after swap is considered dead, thus not considered as valid.
252 * This is enough for most callers, as they don't distinguish dead reloc root
253 * from no reloc root. But btrfs_should_ignore_reloc_root() below is a
254 * special case.
255 */
256static bool have_reloc_root(const struct btrfs_root *root)
257{
258 if (reloc_root_is_dead(root))
259 return false;
260 if (!root->reloc_root)
261 return false;
262 return true;
263}
264
265bool btrfs_should_ignore_reloc_root(const struct btrfs_root *root)
266{
267 struct btrfs_root *reloc_root;
268
269 if (!test_bit(BTRFS_ROOT_SHAREABLE, &root->state))
270 return false;
271
272 /* This root has been merged with its reloc tree, we can ignore it */
273 if (reloc_root_is_dead(root))
274 return true;
275
276 reloc_root = root->reloc_root;
277 if (!reloc_root)
278 return false;
279
280 if (btrfs_header_generation(reloc_root->commit_root) ==
281 root->fs_info->running_transaction->transid)
282 return false;
283 /*
284 * If there is reloc tree and it was created in previous transaction
285 * backref lookup can find the reloc tree, so backref node for the fs
286 * tree root is useless for relocation.
287 */
288 return true;
289}
290
291/*
292 * find reloc tree by address of tree root
293 */
294struct btrfs_root *find_reloc_root(struct btrfs_fs_info *fs_info, u64 bytenr)
295{
296 struct reloc_control *rc = fs_info->reloc_ctl;
297 struct rb_node *rb_node;
298 struct mapping_node *node;
299 struct btrfs_root *root = NULL;
300
301 ASSERT(rc);
302 spin_lock(&rc->reloc_root_tree.lock);
303 rb_node = rb_simple_search(&rc->reloc_root_tree.rb_root, bytenr);
304 if (rb_node) {
305 node = rb_entry(rb_node, struct mapping_node, rb_node);
306 root = node->data;
307 }
308 spin_unlock(&rc->reloc_root_tree.lock);
309 return btrfs_grab_root(root);
310}
311
312/*
313 * For useless nodes, do two major clean ups:
314 *
315 * - Cleanup the children edges and nodes
316 * If child node is also orphan (no parent) during cleanup, then the child
317 * node will also be cleaned up.
318 *
319 * - Freeing up leaves (level 0), keeps nodes detached
320 * For nodes, the node is still cached as "detached"
321 *
322 * Return false if @node is not in the @useless_nodes list.
323 * Return true if @node is in the @useless_nodes list.
324 */
325static bool handle_useless_nodes(struct reloc_control *rc,
326 struct btrfs_backref_node *node)
327{
328 struct btrfs_backref_cache *cache = &rc->backref_cache;
329 struct list_head *useless_node = &cache->useless_node;
330 bool ret = false;
331
332 while (!list_empty(useless_node)) {
333 struct btrfs_backref_node *cur;
334
335 cur = list_first_entry(useless_node, struct btrfs_backref_node,
336 list);
337 list_del_init(&cur->list);
338
339 /* Only tree root nodes can be added to @useless_nodes */
340 ASSERT(list_empty(&cur->upper));
341
342 if (cur == node)
343 ret = true;
344
345 /* The node is the lowest node */
346 if (cur->lowest) {
347 list_del_init(&cur->lower);
348 cur->lowest = 0;
349 }
350
351 /* Cleanup the lower edges */
352 while (!list_empty(&cur->lower)) {
353 struct btrfs_backref_edge *edge;
354 struct btrfs_backref_node *lower;
355
356 edge = list_entry(cur->lower.next,
357 struct btrfs_backref_edge, list[UPPER]);
358 list_del(&edge->list[UPPER]);
359 list_del(&edge->list[LOWER]);
360 lower = edge->node[LOWER];
361 btrfs_backref_free_edge(cache, edge);
362
363 /* Child node is also orphan, queue for cleanup */
364 if (list_empty(&lower->upper))
365 list_add(&lower->list, useless_node);
366 }
367 /* Mark this block processed for relocation */
368 mark_block_processed(rc, cur);
369
370 /*
371 * Backref nodes for tree leaves are deleted from the cache.
372 * Backref nodes for upper level tree blocks are left in the
373 * cache to avoid unnecessary backref lookup.
374 */
375 if (cur->level > 0) {
376 list_add(&cur->list, &cache->detached);
377 cur->detached = 1;
378 } else {
379 rb_erase(&cur->rb_node, &cache->rb_root);
380 btrfs_backref_free_node(cache, cur);
381 }
382 }
383 return ret;
384}
385
386/*
387 * Build backref tree for a given tree block. Root of the backref tree
388 * corresponds the tree block, leaves of the backref tree correspond roots of
389 * b-trees that reference the tree block.
390 *
391 * The basic idea of this function is check backrefs of a given block to find
392 * upper level blocks that reference the block, and then check backrefs of
393 * these upper level blocks recursively. The recursion stops when tree root is
394 * reached or backrefs for the block is cached.
395 *
396 * NOTE: if we find that backrefs for a block are cached, we know backrefs for
397 * all upper level blocks that directly/indirectly reference the block are also
398 * cached.
399 */
400static noinline_for_stack struct btrfs_backref_node *build_backref_tree(
401 struct btrfs_trans_handle *trans,
402 struct reloc_control *rc, struct btrfs_key *node_key,
403 int level, u64 bytenr)
404{
405 struct btrfs_backref_iter *iter;
406 struct btrfs_backref_cache *cache = &rc->backref_cache;
407 /* For searching parent of TREE_BLOCK_REF */
408 struct btrfs_path *path;
409 struct btrfs_backref_node *cur;
410 struct btrfs_backref_node *node = NULL;
411 struct btrfs_backref_edge *edge;
412 int ret;
413
414 iter = btrfs_backref_iter_alloc(rc->extent_root->fs_info);
415 if (!iter)
416 return ERR_PTR(-ENOMEM);
417 path = btrfs_alloc_path();
418 if (!path) {
419 ret = -ENOMEM;
420 goto out;
421 }
422
423 node = btrfs_backref_alloc_node(cache, bytenr, level);
424 if (!node) {
425 ret = -ENOMEM;
426 goto out;
427 }
428
429 node->lowest = 1;
430 cur = node;
431
432 /* Breadth-first search to build backref cache */
433 do {
434 ret = btrfs_backref_add_tree_node(trans, cache, path, iter,
435 node_key, cur);
436 if (ret < 0)
437 goto out;
438
439 edge = list_first_entry_or_null(&cache->pending_edge,
440 struct btrfs_backref_edge, list[UPPER]);
441 /*
442 * The pending list isn't empty, take the first block to
443 * process
444 */
445 if (edge) {
446 list_del_init(&edge->list[UPPER]);
447 cur = edge->node[UPPER];
448 }
449 } while (edge);
450
451 /* Finish the upper linkage of newly added edges/nodes */
452 ret = btrfs_backref_finish_upper_links(cache, node);
453 if (ret < 0)
454 goto out;
455
456 if (handle_useless_nodes(rc, node))
457 node = NULL;
458out:
459 btrfs_free_path(iter->path);
460 kfree(iter);
461 btrfs_free_path(path);
462 if (ret) {
463 btrfs_backref_error_cleanup(cache, node);
464 return ERR_PTR(ret);
465 }
466 ASSERT(!node || !node->detached);
467 ASSERT(list_empty(&cache->useless_node) &&
468 list_empty(&cache->pending_edge));
469 return node;
470}
471
472/*
473 * helper to add backref node for the newly created snapshot.
474 * the backref node is created by cloning backref node that
475 * corresponds to root of source tree
476 */
477static int clone_backref_node(struct btrfs_trans_handle *trans,
478 struct reloc_control *rc,
479 const struct btrfs_root *src,
480 struct btrfs_root *dest)
481{
482 struct btrfs_root *reloc_root = src->reloc_root;
483 struct btrfs_backref_cache *cache = &rc->backref_cache;
484 struct btrfs_backref_node *node = NULL;
485 struct btrfs_backref_node *new_node;
486 struct btrfs_backref_edge *edge;
487 struct btrfs_backref_edge *new_edge;
488 struct rb_node *rb_node;
489
490 rb_node = rb_simple_search(&cache->rb_root, src->commit_root->start);
491 if (rb_node) {
492 node = rb_entry(rb_node, struct btrfs_backref_node, rb_node);
493 if (node->detached)
494 node = NULL;
495 else
496 BUG_ON(node->new_bytenr != reloc_root->node->start);
497 }
498
499 if (!node) {
500 rb_node = rb_simple_search(&cache->rb_root,
501 reloc_root->commit_root->start);
502 if (rb_node) {
503 node = rb_entry(rb_node, struct btrfs_backref_node,
504 rb_node);
505 BUG_ON(node->detached);
506 }
507 }
508
509 if (!node)
510 return 0;
511
512 new_node = btrfs_backref_alloc_node(cache, dest->node->start,
513 node->level);
514 if (!new_node)
515 return -ENOMEM;
516
517 new_node->lowest = node->lowest;
518 new_node->checked = 1;
519 new_node->root = btrfs_grab_root(dest);
520 ASSERT(new_node->root);
521
522 if (!node->lowest) {
523 list_for_each_entry(edge, &node->lower, list[UPPER]) {
524 new_edge = btrfs_backref_alloc_edge(cache);
525 if (!new_edge)
526 goto fail;
527
528 btrfs_backref_link_edge(new_edge, edge->node[LOWER],
529 new_node, LINK_UPPER);
530 }
531 } else {
532 list_add_tail(&new_node->lower, &cache->leaves);
533 }
534
535 rb_node = rb_simple_insert(&cache->rb_root, new_node->bytenr,
536 &new_node->rb_node);
537 if (rb_node)
538 btrfs_backref_panic(trans->fs_info, new_node->bytenr, -EEXIST);
539
540 if (!new_node->lowest) {
541 list_for_each_entry(new_edge, &new_node->lower, list[UPPER]) {
542 list_add_tail(&new_edge->list[LOWER],
543 &new_edge->node[LOWER]->upper);
544 }
545 }
546 return 0;
547fail:
548 while (!list_empty(&new_node->lower)) {
549 new_edge = list_entry(new_node->lower.next,
550 struct btrfs_backref_edge, list[UPPER]);
551 list_del(&new_edge->list[UPPER]);
552 btrfs_backref_free_edge(cache, new_edge);
553 }
554 btrfs_backref_free_node(cache, new_node);
555 return -ENOMEM;
556}
557
558/*
559 * helper to add 'address of tree root -> reloc tree' mapping
560 */
561static int __add_reloc_root(struct btrfs_root *root)
562{
563 struct btrfs_fs_info *fs_info = root->fs_info;
564 struct rb_node *rb_node;
565 struct mapping_node *node;
566 struct reloc_control *rc = fs_info->reloc_ctl;
567
568 node = kmalloc(sizeof(*node), GFP_NOFS);
569 if (!node)
570 return -ENOMEM;
571
572 node->bytenr = root->commit_root->start;
573 node->data = root;
574
575 spin_lock(&rc->reloc_root_tree.lock);
576 rb_node = rb_simple_insert(&rc->reloc_root_tree.rb_root,
577 node->bytenr, &node->rb_node);
578 spin_unlock(&rc->reloc_root_tree.lock);
579 if (rb_node) {
580 btrfs_err(fs_info,
581 "Duplicate root found for start=%llu while inserting into relocation tree",
582 node->bytenr);
583 return -EEXIST;
584 }
585
586 list_add_tail(&root->root_list, &rc->reloc_roots);
587 return 0;
588}
589
590/*
591 * helper to delete the 'address of tree root -> reloc tree'
592 * mapping
593 */
594static void __del_reloc_root(struct btrfs_root *root)
595{
596 struct btrfs_fs_info *fs_info = root->fs_info;
597 struct rb_node *rb_node;
598 struct mapping_node *node = NULL;
599 struct reloc_control *rc = fs_info->reloc_ctl;
600 bool put_ref = false;
601
602 if (rc && root->node) {
603 spin_lock(&rc->reloc_root_tree.lock);
604 rb_node = rb_simple_search(&rc->reloc_root_tree.rb_root,
605 root->commit_root->start);
606 if (rb_node) {
607 node = rb_entry(rb_node, struct mapping_node, rb_node);
608 rb_erase(&node->rb_node, &rc->reloc_root_tree.rb_root);
609 RB_CLEAR_NODE(&node->rb_node);
610 }
611 spin_unlock(&rc->reloc_root_tree.lock);
612 ASSERT(!node || (struct btrfs_root *)node->data == root);
613 }
614
615 /*
616 * We only put the reloc root here if it's on the list. There's a lot
617 * of places where the pattern is to splice the rc->reloc_roots, process
618 * the reloc roots, and then add the reloc root back onto
619 * rc->reloc_roots. If we call __del_reloc_root while it's off of the
620 * list we don't want the reference being dropped, because the guy
621 * messing with the list is in charge of the reference.
622 */
623 spin_lock(&fs_info->trans_lock);
624 if (!list_empty(&root->root_list)) {
625 put_ref = true;
626 list_del_init(&root->root_list);
627 }
628 spin_unlock(&fs_info->trans_lock);
629 if (put_ref)
630 btrfs_put_root(root);
631 kfree(node);
632}
633
634/*
635 * helper to update the 'address of tree root -> reloc tree'
636 * mapping
637 */
638static int __update_reloc_root(struct btrfs_root *root)
639{
640 struct btrfs_fs_info *fs_info = root->fs_info;
641 struct rb_node *rb_node;
642 struct mapping_node *node = NULL;
643 struct reloc_control *rc = fs_info->reloc_ctl;
644
645 spin_lock(&rc->reloc_root_tree.lock);
646 rb_node = rb_simple_search(&rc->reloc_root_tree.rb_root,
647 root->commit_root->start);
648 if (rb_node) {
649 node = rb_entry(rb_node, struct mapping_node, rb_node);
650 rb_erase(&node->rb_node, &rc->reloc_root_tree.rb_root);
651 }
652 spin_unlock(&rc->reloc_root_tree.lock);
653
654 if (!node)
655 return 0;
656 BUG_ON((struct btrfs_root *)node->data != root);
657
658 spin_lock(&rc->reloc_root_tree.lock);
659 node->bytenr = root->node->start;
660 rb_node = rb_simple_insert(&rc->reloc_root_tree.rb_root,
661 node->bytenr, &node->rb_node);
662 spin_unlock(&rc->reloc_root_tree.lock);
663 if (rb_node)
664 btrfs_backref_panic(fs_info, node->bytenr, -EEXIST);
665 return 0;
666}
667
668static struct btrfs_root *create_reloc_root(struct btrfs_trans_handle *trans,
669 struct btrfs_root *root, u64 objectid)
670{
671 struct btrfs_fs_info *fs_info = root->fs_info;
672 struct btrfs_root *reloc_root;
673 struct extent_buffer *eb;
674 struct btrfs_root_item *root_item;
675 struct btrfs_key root_key;
676 int ret = 0;
677 bool must_abort = false;
678
679 root_item = kmalloc(sizeof(*root_item), GFP_NOFS);
680 if (!root_item)
681 return ERR_PTR(-ENOMEM);
682
683 root_key.objectid = BTRFS_TREE_RELOC_OBJECTID;
684 root_key.type = BTRFS_ROOT_ITEM_KEY;
685 root_key.offset = objectid;
686
687 if (btrfs_root_id(root) == objectid) {
688 u64 commit_root_gen;
689
690 /* called by btrfs_init_reloc_root */
691 ret = btrfs_copy_root(trans, root, root->commit_root, &eb,
692 BTRFS_TREE_RELOC_OBJECTID);
693 if (ret)
694 goto fail;
695
696 /*
697 * Set the last_snapshot field to the generation of the commit
698 * root - like this ctree.c:btrfs_block_can_be_shared() behaves
699 * correctly (returns true) when the relocation root is created
700 * either inside the critical section of a transaction commit
701 * (through transaction.c:qgroup_account_snapshot()) and when
702 * it's created before the transaction commit is started.
703 */
704 commit_root_gen = btrfs_header_generation(root->commit_root);
705 btrfs_set_root_last_snapshot(&root->root_item, commit_root_gen);
706 } else {
707 /*
708 * called by btrfs_reloc_post_snapshot_hook.
709 * the source tree is a reloc tree, all tree blocks
710 * modified after it was created have RELOC flag
711 * set in their headers. so it's OK to not update
712 * the 'last_snapshot'.
713 */
714 ret = btrfs_copy_root(trans, root, root->node, &eb,
715 BTRFS_TREE_RELOC_OBJECTID);
716 if (ret)
717 goto fail;
718 }
719
720 /*
721 * We have changed references at this point, we must abort the
722 * transaction if anything fails.
723 */
724 must_abort = true;
725
726 memcpy(root_item, &root->root_item, sizeof(*root_item));
727 btrfs_set_root_bytenr(root_item, eb->start);
728 btrfs_set_root_level(root_item, btrfs_header_level(eb));
729 btrfs_set_root_generation(root_item, trans->transid);
730
731 if (btrfs_root_id(root) == objectid) {
732 btrfs_set_root_refs(root_item, 0);
733 memset(&root_item->drop_progress, 0,
734 sizeof(struct btrfs_disk_key));
735 btrfs_set_root_drop_level(root_item, 0);
736 }
737
738 btrfs_tree_unlock(eb);
739 free_extent_buffer(eb);
740
741 ret = btrfs_insert_root(trans, fs_info->tree_root,
742 &root_key, root_item);
743 if (ret)
744 goto fail;
745
746 kfree(root_item);
747
748 reloc_root = btrfs_read_tree_root(fs_info->tree_root, &root_key);
749 if (IS_ERR(reloc_root)) {
750 ret = PTR_ERR(reloc_root);
751 goto abort;
752 }
753 set_bit(BTRFS_ROOT_SHAREABLE, &reloc_root->state);
754 btrfs_set_root_last_trans(reloc_root, trans->transid);
755 return reloc_root;
756fail:
757 kfree(root_item);
758abort:
759 if (must_abort)
760 btrfs_abort_transaction(trans, ret);
761 return ERR_PTR(ret);
762}
763
764/*
765 * create reloc tree for a given fs tree. reloc tree is just a
766 * snapshot of the fs tree with special root objectid.
767 *
768 * The reloc_root comes out of here with two references, one for
769 * root->reloc_root, and another for being on the rc->reloc_roots list.
770 */
771int btrfs_init_reloc_root(struct btrfs_trans_handle *trans,
772 struct btrfs_root *root)
773{
774 struct btrfs_fs_info *fs_info = root->fs_info;
775 struct btrfs_root *reloc_root;
776 struct reloc_control *rc = fs_info->reloc_ctl;
777 struct btrfs_block_rsv *rsv;
778 int clear_rsv = 0;
779 int ret;
780
781 if (!rc)
782 return 0;
783
784 /*
785 * The subvolume has reloc tree but the swap is finished, no need to
786 * create/update the dead reloc tree
787 */
788 if (reloc_root_is_dead(root))
789 return 0;
790
791 /*
792 * This is subtle but important. We do not do
793 * record_root_in_transaction for reloc roots, instead we record their
794 * corresponding fs root, and then here we update the last trans for the
795 * reloc root. This means that we have to do this for the entire life
796 * of the reloc root, regardless of which stage of the relocation we are
797 * in.
798 */
799 if (root->reloc_root) {
800 reloc_root = root->reloc_root;
801 btrfs_set_root_last_trans(reloc_root, trans->transid);
802 return 0;
803 }
804
805 /*
806 * We are merging reloc roots, we do not need new reloc trees. Also
807 * reloc trees never need their own reloc tree.
808 */
809 if (!rc->create_reloc_tree || btrfs_root_id(root) == BTRFS_TREE_RELOC_OBJECTID)
810 return 0;
811
812 if (!trans->reloc_reserved) {
813 rsv = trans->block_rsv;
814 trans->block_rsv = rc->block_rsv;
815 clear_rsv = 1;
816 }
817 reloc_root = create_reloc_root(trans, root, btrfs_root_id(root));
818 if (clear_rsv)
819 trans->block_rsv = rsv;
820 if (IS_ERR(reloc_root))
821 return PTR_ERR(reloc_root);
822
823 ret = __add_reloc_root(reloc_root);
824 ASSERT(ret != -EEXIST);
825 if (ret) {
826 /* Pairs with create_reloc_root */
827 btrfs_put_root(reloc_root);
828 return ret;
829 }
830 root->reloc_root = btrfs_grab_root(reloc_root);
831 return 0;
832}
833
834/*
835 * update root item of reloc tree
836 */
837int btrfs_update_reloc_root(struct btrfs_trans_handle *trans,
838 struct btrfs_root *root)
839{
840 struct btrfs_fs_info *fs_info = root->fs_info;
841 struct btrfs_root *reloc_root;
842 struct btrfs_root_item *root_item;
843 int ret;
844
845 if (!have_reloc_root(root))
846 return 0;
847
848 reloc_root = root->reloc_root;
849 root_item = &reloc_root->root_item;
850
851 /*
852 * We are probably ok here, but __del_reloc_root() will drop its ref of
853 * the root. We have the ref for root->reloc_root, but just in case
854 * hold it while we update the reloc root.
855 */
856 btrfs_grab_root(reloc_root);
857
858 /* root->reloc_root will stay until current relocation finished */
859 if (fs_info->reloc_ctl && fs_info->reloc_ctl->merge_reloc_tree &&
860 btrfs_root_refs(root_item) == 0) {
861 set_bit(BTRFS_ROOT_DEAD_RELOC_TREE, &root->state);
862 /*
863 * Mark the tree as dead before we change reloc_root so
864 * have_reloc_root will not touch it from now on.
865 */
866 smp_wmb();
867 __del_reloc_root(reloc_root);
868 }
869
870 if (reloc_root->commit_root != reloc_root->node) {
871 __update_reloc_root(reloc_root);
872 btrfs_set_root_node(root_item, reloc_root->node);
873 free_extent_buffer(reloc_root->commit_root);
874 reloc_root->commit_root = btrfs_root_node(reloc_root);
875 }
876
877 ret = btrfs_update_root(trans, fs_info->tree_root,
878 &reloc_root->root_key, root_item);
879 btrfs_put_root(reloc_root);
880 return ret;
881}
882
883/*
884 * get new location of data
885 */
886static int get_new_location(struct inode *reloc_inode, u64 *new_bytenr,
887 u64 bytenr, u64 num_bytes)
888{
889 struct btrfs_root *root = BTRFS_I(reloc_inode)->root;
890 struct btrfs_path *path;
891 struct btrfs_file_extent_item *fi;
892 struct extent_buffer *leaf;
893 int ret;
894
895 path = btrfs_alloc_path();
896 if (!path)
897 return -ENOMEM;
898
899 bytenr -= BTRFS_I(reloc_inode)->reloc_block_group_start;
900 ret = btrfs_lookup_file_extent(NULL, root, path,
901 btrfs_ino(BTRFS_I(reloc_inode)), bytenr, 0);
902 if (ret < 0)
903 goto out;
904 if (ret > 0) {
905 ret = -ENOENT;
906 goto out;
907 }
908
909 leaf = path->nodes[0];
910 fi = btrfs_item_ptr(leaf, path->slots[0],
911 struct btrfs_file_extent_item);
912
913 BUG_ON(btrfs_file_extent_offset(leaf, fi) ||
914 btrfs_file_extent_compression(leaf, fi) ||
915 btrfs_file_extent_encryption(leaf, fi) ||
916 btrfs_file_extent_other_encoding(leaf, fi));
917
918 if (num_bytes != btrfs_file_extent_disk_num_bytes(leaf, fi)) {
919 ret = -EINVAL;
920 goto out;
921 }
922
923 *new_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
924 ret = 0;
925out:
926 btrfs_free_path(path);
927 return ret;
928}
929
930/*
931 * update file extent items in the tree leaf to point to
932 * the new locations.
933 */
934static noinline_for_stack
935int replace_file_extents(struct btrfs_trans_handle *trans,
936 struct reloc_control *rc,
937 struct btrfs_root *root,
938 struct extent_buffer *leaf)
939{
940 struct btrfs_fs_info *fs_info = root->fs_info;
941 struct btrfs_key key;
942 struct btrfs_file_extent_item *fi;
943 struct btrfs_inode *inode = NULL;
944 u64 parent;
945 u64 bytenr;
946 u64 new_bytenr = 0;
947 u64 num_bytes;
948 u64 end;
949 u32 nritems;
950 u32 i;
951 int ret = 0;
952 int first = 1;
953 int dirty = 0;
954
955 if (rc->stage != UPDATE_DATA_PTRS)
956 return 0;
957
958 /* reloc trees always use full backref */
959 if (btrfs_root_id(root) == BTRFS_TREE_RELOC_OBJECTID)
960 parent = leaf->start;
961 else
962 parent = 0;
963
964 nritems = btrfs_header_nritems(leaf);
965 for (i = 0; i < nritems; i++) {
966 struct btrfs_ref ref = { 0 };
967
968 cond_resched();
969 btrfs_item_key_to_cpu(leaf, &key, i);
970 if (key.type != BTRFS_EXTENT_DATA_KEY)
971 continue;
972 fi = btrfs_item_ptr(leaf, i, struct btrfs_file_extent_item);
973 if (btrfs_file_extent_type(leaf, fi) ==
974 BTRFS_FILE_EXTENT_INLINE)
975 continue;
976 bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
977 num_bytes = btrfs_file_extent_disk_num_bytes(leaf, fi);
978 if (bytenr == 0)
979 continue;
980 if (!in_range(bytenr, rc->block_group->start,
981 rc->block_group->length))
982 continue;
983
984 /*
985 * if we are modifying block in fs tree, wait for read_folio
986 * to complete and drop the extent cache
987 */
988 if (btrfs_root_id(root) != BTRFS_TREE_RELOC_OBJECTID) {
989 if (first) {
990 inode = btrfs_find_first_inode(root, key.objectid);
991 first = 0;
992 } else if (inode && btrfs_ino(inode) < key.objectid) {
993 btrfs_add_delayed_iput(inode);
994 inode = btrfs_find_first_inode(root, key.objectid);
995 }
996 if (inode && btrfs_ino(inode) == key.objectid) {
997 struct extent_state *cached_state = NULL;
998
999 end = key.offset +
1000 btrfs_file_extent_num_bytes(leaf, fi);
1001 WARN_ON(!IS_ALIGNED(key.offset,
1002 fs_info->sectorsize));
1003 WARN_ON(!IS_ALIGNED(end, fs_info->sectorsize));
1004 end--;
1005 /* Take mmap lock to serialize with reflinks. */
1006 if (!down_read_trylock(&inode->i_mmap_lock))
1007 continue;
1008 ret = try_lock_extent(&inode->io_tree, key.offset,
1009 end, &cached_state);
1010 if (!ret) {
1011 up_read(&inode->i_mmap_lock);
1012 continue;
1013 }
1014
1015 btrfs_drop_extent_map_range(inode, key.offset, end, true);
1016 unlock_extent(&inode->io_tree, key.offset, end,
1017 &cached_state);
1018 up_read(&inode->i_mmap_lock);
1019 }
1020 }
1021
1022 ret = get_new_location(rc->data_inode, &new_bytenr,
1023 bytenr, num_bytes);
1024 if (ret) {
1025 /*
1026 * Don't have to abort since we've not changed anything
1027 * in the file extent yet.
1028 */
1029 break;
1030 }
1031
1032 btrfs_set_file_extent_disk_bytenr(leaf, fi, new_bytenr);
1033 dirty = 1;
1034
1035 key.offset -= btrfs_file_extent_offset(leaf, fi);
1036 ref.action = BTRFS_ADD_DELAYED_REF;
1037 ref.bytenr = new_bytenr;
1038 ref.num_bytes = num_bytes;
1039 ref.parent = parent;
1040 ref.owning_root = btrfs_root_id(root);
1041 ref.ref_root = btrfs_header_owner(leaf);
1042 btrfs_init_data_ref(&ref, key.objectid, key.offset,
1043 btrfs_root_id(root), false);
1044 ret = btrfs_inc_extent_ref(trans, &ref);
1045 if (ret) {
1046 btrfs_abort_transaction(trans, ret);
1047 break;
1048 }
1049
1050 ref.action = BTRFS_DROP_DELAYED_REF;
1051 ref.bytenr = bytenr;
1052 ref.num_bytes = num_bytes;
1053 ref.parent = parent;
1054 ref.owning_root = btrfs_root_id(root);
1055 ref.ref_root = btrfs_header_owner(leaf);
1056 btrfs_init_data_ref(&ref, key.objectid, key.offset,
1057 btrfs_root_id(root), false);
1058 ret = btrfs_free_extent(trans, &ref);
1059 if (ret) {
1060 btrfs_abort_transaction(trans, ret);
1061 break;
1062 }
1063 }
1064 if (dirty)
1065 btrfs_mark_buffer_dirty(trans, leaf);
1066 if (inode)
1067 btrfs_add_delayed_iput(inode);
1068 return ret;
1069}
1070
1071static noinline_for_stack int memcmp_node_keys(const struct extent_buffer *eb,
1072 int slot, const struct btrfs_path *path,
1073 int level)
1074{
1075 struct btrfs_disk_key key1;
1076 struct btrfs_disk_key key2;
1077 btrfs_node_key(eb, &key1, slot);
1078 btrfs_node_key(path->nodes[level], &key2, path->slots[level]);
1079 return memcmp(&key1, &key2, sizeof(key1));
1080}
1081
1082/*
1083 * try to replace tree blocks in fs tree with the new blocks
1084 * in reloc tree. tree blocks haven't been modified since the
1085 * reloc tree was create can be replaced.
1086 *
1087 * if a block was replaced, level of the block + 1 is returned.
1088 * if no block got replaced, 0 is returned. if there are other
1089 * errors, a negative error number is returned.
1090 */
1091static noinline_for_stack
1092int replace_path(struct btrfs_trans_handle *trans, struct reloc_control *rc,
1093 struct btrfs_root *dest, struct btrfs_root *src,
1094 struct btrfs_path *path, struct btrfs_key *next_key,
1095 int lowest_level, int max_level)
1096{
1097 struct btrfs_fs_info *fs_info = dest->fs_info;
1098 struct extent_buffer *eb;
1099 struct extent_buffer *parent;
1100 struct btrfs_ref ref = { 0 };
1101 struct btrfs_key key;
1102 u64 old_bytenr;
1103 u64 new_bytenr;
1104 u64 old_ptr_gen;
1105 u64 new_ptr_gen;
1106 u64 last_snapshot;
1107 u32 blocksize;
1108 int cow = 0;
1109 int level;
1110 int ret;
1111 int slot;
1112
1113 ASSERT(btrfs_root_id(src) == BTRFS_TREE_RELOC_OBJECTID);
1114 ASSERT(btrfs_root_id(dest) != BTRFS_TREE_RELOC_OBJECTID);
1115
1116 last_snapshot = btrfs_root_last_snapshot(&src->root_item);
1117again:
1118 slot = path->slots[lowest_level];
1119 btrfs_node_key_to_cpu(path->nodes[lowest_level], &key, slot);
1120
1121 eb = btrfs_lock_root_node(dest);
1122 level = btrfs_header_level(eb);
1123
1124 if (level < lowest_level) {
1125 btrfs_tree_unlock(eb);
1126 free_extent_buffer(eb);
1127 return 0;
1128 }
1129
1130 if (cow) {
1131 ret = btrfs_cow_block(trans, dest, eb, NULL, 0, &eb,
1132 BTRFS_NESTING_COW);
1133 if (ret) {
1134 btrfs_tree_unlock(eb);
1135 free_extent_buffer(eb);
1136 return ret;
1137 }
1138 }
1139
1140 if (next_key) {
1141 next_key->objectid = (u64)-1;
1142 next_key->type = (u8)-1;
1143 next_key->offset = (u64)-1;
1144 }
1145
1146 parent = eb;
1147 while (1) {
1148 level = btrfs_header_level(parent);
1149 ASSERT(level >= lowest_level);
1150
1151 ret = btrfs_bin_search(parent, 0, &key, &slot);
1152 if (ret < 0)
1153 break;
1154 if (ret && slot > 0)
1155 slot--;
1156
1157 if (next_key && slot + 1 < btrfs_header_nritems(parent))
1158 btrfs_node_key_to_cpu(parent, next_key, slot + 1);
1159
1160 old_bytenr = btrfs_node_blockptr(parent, slot);
1161 blocksize = fs_info->nodesize;
1162 old_ptr_gen = btrfs_node_ptr_generation(parent, slot);
1163
1164 if (level <= max_level) {
1165 eb = path->nodes[level];
1166 new_bytenr = btrfs_node_blockptr(eb,
1167 path->slots[level]);
1168 new_ptr_gen = btrfs_node_ptr_generation(eb,
1169 path->slots[level]);
1170 } else {
1171 new_bytenr = 0;
1172 new_ptr_gen = 0;
1173 }
1174
1175 if (WARN_ON(new_bytenr > 0 && new_bytenr == old_bytenr)) {
1176 ret = level;
1177 break;
1178 }
1179
1180 if (new_bytenr == 0 || old_ptr_gen > last_snapshot ||
1181 memcmp_node_keys(parent, slot, path, level)) {
1182 if (level <= lowest_level) {
1183 ret = 0;
1184 break;
1185 }
1186
1187 eb = btrfs_read_node_slot(parent, slot);
1188 if (IS_ERR(eb)) {
1189 ret = PTR_ERR(eb);
1190 break;
1191 }
1192 btrfs_tree_lock(eb);
1193 if (cow) {
1194 ret = btrfs_cow_block(trans, dest, eb, parent,
1195 slot, &eb,
1196 BTRFS_NESTING_COW);
1197 if (ret) {
1198 btrfs_tree_unlock(eb);
1199 free_extent_buffer(eb);
1200 break;
1201 }
1202 }
1203
1204 btrfs_tree_unlock(parent);
1205 free_extent_buffer(parent);
1206
1207 parent = eb;
1208 continue;
1209 }
1210
1211 if (!cow) {
1212 btrfs_tree_unlock(parent);
1213 free_extent_buffer(parent);
1214 cow = 1;
1215 goto again;
1216 }
1217
1218 btrfs_node_key_to_cpu(path->nodes[level], &key,
1219 path->slots[level]);
1220 btrfs_release_path(path);
1221
1222 path->lowest_level = level;
1223 set_bit(BTRFS_ROOT_RESET_LOCKDEP_CLASS, &src->state);
1224 ret = btrfs_search_slot(trans, src, &key, path, 0, 1);
1225 clear_bit(BTRFS_ROOT_RESET_LOCKDEP_CLASS, &src->state);
1226 path->lowest_level = 0;
1227 if (ret) {
1228 if (ret > 0)
1229 ret = -ENOENT;
1230 break;
1231 }
1232
1233 /*
1234 * Info qgroup to trace both subtrees.
1235 *
1236 * We must trace both trees.
1237 * 1) Tree reloc subtree
1238 * If not traced, we will leak data numbers
1239 * 2) Fs subtree
1240 * If not traced, we will double count old data
1241 *
1242 * We don't scan the subtree right now, but only record
1243 * the swapped tree blocks.
1244 * The real subtree rescan is delayed until we have new
1245 * CoW on the subtree root node before transaction commit.
1246 */
1247 ret = btrfs_qgroup_add_swapped_blocks(dest,
1248 rc->block_group, parent, slot,
1249 path->nodes[level], path->slots[level],
1250 last_snapshot);
1251 if (ret < 0)
1252 break;
1253 /*
1254 * swap blocks in fs tree and reloc tree.
1255 */
1256 btrfs_set_node_blockptr(parent, slot, new_bytenr);
1257 btrfs_set_node_ptr_generation(parent, slot, new_ptr_gen);
1258 btrfs_mark_buffer_dirty(trans, parent);
1259
1260 btrfs_set_node_blockptr(path->nodes[level],
1261 path->slots[level], old_bytenr);
1262 btrfs_set_node_ptr_generation(path->nodes[level],
1263 path->slots[level], old_ptr_gen);
1264 btrfs_mark_buffer_dirty(trans, path->nodes[level]);
1265
1266 ref.action = BTRFS_ADD_DELAYED_REF;
1267 ref.bytenr = old_bytenr;
1268 ref.num_bytes = blocksize;
1269 ref.parent = path->nodes[level]->start;
1270 ref.owning_root = btrfs_root_id(src);
1271 ref.ref_root = btrfs_root_id(src);
1272 btrfs_init_tree_ref(&ref, level - 1, 0, true);
1273 ret = btrfs_inc_extent_ref(trans, &ref);
1274 if (ret) {
1275 btrfs_abort_transaction(trans, ret);
1276 break;
1277 }
1278
1279 ref.action = BTRFS_ADD_DELAYED_REF;
1280 ref.bytenr = new_bytenr;
1281 ref.num_bytes = blocksize;
1282 ref.parent = 0;
1283 ref.owning_root = btrfs_root_id(dest);
1284 ref.ref_root = btrfs_root_id(dest);
1285 btrfs_init_tree_ref(&ref, level - 1, 0, true);
1286 ret = btrfs_inc_extent_ref(trans, &ref);
1287 if (ret) {
1288 btrfs_abort_transaction(trans, ret);
1289 break;
1290 }
1291
1292 /* We don't know the real owning_root, use 0. */
1293 ref.action = BTRFS_DROP_DELAYED_REF;
1294 ref.bytenr = new_bytenr;
1295 ref.num_bytes = blocksize;
1296 ref.parent = path->nodes[level]->start;
1297 ref.owning_root = 0;
1298 ref.ref_root = btrfs_root_id(src);
1299 btrfs_init_tree_ref(&ref, level - 1, 0, true);
1300 ret = btrfs_free_extent(trans, &ref);
1301 if (ret) {
1302 btrfs_abort_transaction(trans, ret);
1303 break;
1304 }
1305
1306 /* We don't know the real owning_root, use 0. */
1307 ref.action = BTRFS_DROP_DELAYED_REF;
1308 ref.bytenr = old_bytenr;
1309 ref.num_bytes = blocksize;
1310 ref.parent = 0;
1311 ref.owning_root = 0;
1312 ref.ref_root = btrfs_root_id(dest);
1313 btrfs_init_tree_ref(&ref, level - 1, 0, true);
1314 ret = btrfs_free_extent(trans, &ref);
1315 if (ret) {
1316 btrfs_abort_transaction(trans, ret);
1317 break;
1318 }
1319
1320 btrfs_unlock_up_safe(path, 0);
1321
1322 ret = level;
1323 break;
1324 }
1325 btrfs_tree_unlock(parent);
1326 free_extent_buffer(parent);
1327 return ret;
1328}
1329
1330/*
1331 * helper to find next relocated block in reloc tree
1332 */
1333static noinline_for_stack
1334int walk_up_reloc_tree(struct btrfs_root *root, struct btrfs_path *path,
1335 int *level)
1336{
1337 struct extent_buffer *eb;
1338 int i;
1339 u64 last_snapshot;
1340 u32 nritems;
1341
1342 last_snapshot = btrfs_root_last_snapshot(&root->root_item);
1343
1344 for (i = 0; i < *level; i++) {
1345 free_extent_buffer(path->nodes[i]);
1346 path->nodes[i] = NULL;
1347 }
1348
1349 for (i = *level; i < BTRFS_MAX_LEVEL && path->nodes[i]; i++) {
1350 eb = path->nodes[i];
1351 nritems = btrfs_header_nritems(eb);
1352 while (path->slots[i] + 1 < nritems) {
1353 path->slots[i]++;
1354 if (btrfs_node_ptr_generation(eb, path->slots[i]) <=
1355 last_snapshot)
1356 continue;
1357
1358 *level = i;
1359 return 0;
1360 }
1361 free_extent_buffer(path->nodes[i]);
1362 path->nodes[i] = NULL;
1363 }
1364 return 1;
1365}
1366
1367/*
1368 * walk down reloc tree to find relocated block of lowest level
1369 */
1370static noinline_for_stack
1371int walk_down_reloc_tree(struct btrfs_root *root, struct btrfs_path *path,
1372 int *level)
1373{
1374 struct extent_buffer *eb = NULL;
1375 int i;
1376 u64 ptr_gen = 0;
1377 u64 last_snapshot;
1378 u32 nritems;
1379
1380 last_snapshot = btrfs_root_last_snapshot(&root->root_item);
1381
1382 for (i = *level; i > 0; i--) {
1383 eb = path->nodes[i];
1384 nritems = btrfs_header_nritems(eb);
1385 while (path->slots[i] < nritems) {
1386 ptr_gen = btrfs_node_ptr_generation(eb, path->slots[i]);
1387 if (ptr_gen > last_snapshot)
1388 break;
1389 path->slots[i]++;
1390 }
1391 if (path->slots[i] >= nritems) {
1392 if (i == *level)
1393 break;
1394 *level = i + 1;
1395 return 0;
1396 }
1397 if (i == 1) {
1398 *level = i;
1399 return 0;
1400 }
1401
1402 eb = btrfs_read_node_slot(eb, path->slots[i]);
1403 if (IS_ERR(eb))
1404 return PTR_ERR(eb);
1405 BUG_ON(btrfs_header_level(eb) != i - 1);
1406 path->nodes[i - 1] = eb;
1407 path->slots[i - 1] = 0;
1408 }
1409 return 1;
1410}
1411
1412/*
1413 * invalidate extent cache for file extents whose key in range of
1414 * [min_key, max_key)
1415 */
1416static int invalidate_extent_cache(struct btrfs_root *root,
1417 const struct btrfs_key *min_key,
1418 const struct btrfs_key *max_key)
1419{
1420 struct btrfs_fs_info *fs_info = root->fs_info;
1421 struct btrfs_inode *inode = NULL;
1422 u64 objectid;
1423 u64 start, end;
1424 u64 ino;
1425
1426 objectid = min_key->objectid;
1427 while (1) {
1428 struct extent_state *cached_state = NULL;
1429
1430 cond_resched();
1431 if (inode)
1432 iput(&inode->vfs_inode);
1433
1434 if (objectid > max_key->objectid)
1435 break;
1436
1437 inode = btrfs_find_first_inode(root, objectid);
1438 if (!inode)
1439 break;
1440 ino = btrfs_ino(inode);
1441
1442 if (ino > max_key->objectid) {
1443 iput(&inode->vfs_inode);
1444 break;
1445 }
1446
1447 objectid = ino + 1;
1448 if (!S_ISREG(inode->vfs_inode.i_mode))
1449 continue;
1450
1451 if (unlikely(min_key->objectid == ino)) {
1452 if (min_key->type > BTRFS_EXTENT_DATA_KEY)
1453 continue;
1454 if (min_key->type < BTRFS_EXTENT_DATA_KEY)
1455 start = 0;
1456 else {
1457 start = min_key->offset;
1458 WARN_ON(!IS_ALIGNED(start, fs_info->sectorsize));
1459 }
1460 } else {
1461 start = 0;
1462 }
1463
1464 if (unlikely(max_key->objectid == ino)) {
1465 if (max_key->type < BTRFS_EXTENT_DATA_KEY)
1466 continue;
1467 if (max_key->type > BTRFS_EXTENT_DATA_KEY) {
1468 end = (u64)-1;
1469 } else {
1470 if (max_key->offset == 0)
1471 continue;
1472 end = max_key->offset;
1473 WARN_ON(!IS_ALIGNED(end, fs_info->sectorsize));
1474 end--;
1475 }
1476 } else {
1477 end = (u64)-1;
1478 }
1479
1480 /* the lock_extent waits for read_folio to complete */
1481 lock_extent(&inode->io_tree, start, end, &cached_state);
1482 btrfs_drop_extent_map_range(inode, start, end, true);
1483 unlock_extent(&inode->io_tree, start, end, &cached_state);
1484 }
1485 return 0;
1486}
1487
1488static int find_next_key(struct btrfs_path *path, int level,
1489 struct btrfs_key *key)
1490
1491{
1492 while (level < BTRFS_MAX_LEVEL) {
1493 if (!path->nodes[level])
1494 break;
1495 if (path->slots[level] + 1 <
1496 btrfs_header_nritems(path->nodes[level])) {
1497 btrfs_node_key_to_cpu(path->nodes[level], key,
1498 path->slots[level] + 1);
1499 return 0;
1500 }
1501 level++;
1502 }
1503 return 1;
1504}
1505
1506/*
1507 * Insert current subvolume into reloc_control::dirty_subvol_roots
1508 */
1509static int insert_dirty_subvol(struct btrfs_trans_handle *trans,
1510 struct reloc_control *rc,
1511 struct btrfs_root *root)
1512{
1513 struct btrfs_root *reloc_root = root->reloc_root;
1514 struct btrfs_root_item *reloc_root_item;
1515 int ret;
1516
1517 /* @root must be a subvolume tree root with a valid reloc tree */
1518 ASSERT(btrfs_root_id(root) != BTRFS_TREE_RELOC_OBJECTID);
1519 ASSERT(reloc_root);
1520
1521 reloc_root_item = &reloc_root->root_item;
1522 memset(&reloc_root_item->drop_progress, 0,
1523 sizeof(reloc_root_item->drop_progress));
1524 btrfs_set_root_drop_level(reloc_root_item, 0);
1525 btrfs_set_root_refs(reloc_root_item, 0);
1526 ret = btrfs_update_reloc_root(trans, root);
1527 if (ret)
1528 return ret;
1529
1530 if (list_empty(&root->reloc_dirty_list)) {
1531 btrfs_grab_root(root);
1532 list_add_tail(&root->reloc_dirty_list, &rc->dirty_subvol_roots);
1533 }
1534
1535 return 0;
1536}
1537
1538static int clean_dirty_subvols(struct reloc_control *rc)
1539{
1540 struct btrfs_root *root;
1541 struct btrfs_root *next;
1542 int ret = 0;
1543 int ret2;
1544
1545 list_for_each_entry_safe(root, next, &rc->dirty_subvol_roots,
1546 reloc_dirty_list) {
1547 if (btrfs_root_id(root) != BTRFS_TREE_RELOC_OBJECTID) {
1548 /* Merged subvolume, cleanup its reloc root */
1549 struct btrfs_root *reloc_root = root->reloc_root;
1550
1551 list_del_init(&root->reloc_dirty_list);
1552 root->reloc_root = NULL;
1553 /*
1554 * Need barrier to ensure clear_bit() only happens after
1555 * root->reloc_root = NULL. Pairs with have_reloc_root.
1556 */
1557 smp_wmb();
1558 clear_bit(BTRFS_ROOT_DEAD_RELOC_TREE, &root->state);
1559 if (reloc_root) {
1560 /*
1561 * btrfs_drop_snapshot drops our ref we hold for
1562 * ->reloc_root. If it fails however we must
1563 * drop the ref ourselves.
1564 */
1565 ret2 = btrfs_drop_snapshot(reloc_root, 0, 1);
1566 if (ret2 < 0) {
1567 btrfs_put_root(reloc_root);
1568 if (!ret)
1569 ret = ret2;
1570 }
1571 }
1572 btrfs_put_root(root);
1573 } else {
1574 /* Orphan reloc tree, just clean it up */
1575 ret2 = btrfs_drop_snapshot(root, 0, 1);
1576 if (ret2 < 0) {
1577 btrfs_put_root(root);
1578 if (!ret)
1579 ret = ret2;
1580 }
1581 }
1582 }
1583 return ret;
1584}
1585
1586/*
1587 * merge the relocated tree blocks in reloc tree with corresponding
1588 * fs tree.
1589 */
1590static noinline_for_stack int merge_reloc_root(struct reloc_control *rc,
1591 struct btrfs_root *root)
1592{
1593 struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
1594 struct btrfs_key key;
1595 struct btrfs_key next_key;
1596 struct btrfs_trans_handle *trans = NULL;
1597 struct btrfs_root *reloc_root;
1598 struct btrfs_root_item *root_item;
1599 struct btrfs_path *path;
1600 struct extent_buffer *leaf;
1601 int reserve_level;
1602 int level;
1603 int max_level;
1604 int replaced = 0;
1605 int ret = 0;
1606 u32 min_reserved;
1607
1608 path = btrfs_alloc_path();
1609 if (!path)
1610 return -ENOMEM;
1611 path->reada = READA_FORWARD;
1612
1613 reloc_root = root->reloc_root;
1614 root_item = &reloc_root->root_item;
1615
1616 if (btrfs_disk_key_objectid(&root_item->drop_progress) == 0) {
1617 level = btrfs_root_level(root_item);
1618 atomic_inc(&reloc_root->node->refs);
1619 path->nodes[level] = reloc_root->node;
1620 path->slots[level] = 0;
1621 } else {
1622 btrfs_disk_key_to_cpu(&key, &root_item->drop_progress);
1623
1624 level = btrfs_root_drop_level(root_item);
1625 BUG_ON(level == 0);
1626 path->lowest_level = level;
1627 ret = btrfs_search_slot(NULL, reloc_root, &key, path, 0, 0);
1628 path->lowest_level = 0;
1629 if (ret < 0) {
1630 btrfs_free_path(path);
1631 return ret;
1632 }
1633
1634 btrfs_node_key_to_cpu(path->nodes[level], &next_key,
1635 path->slots[level]);
1636 WARN_ON(memcmp(&key, &next_key, sizeof(key)));
1637
1638 btrfs_unlock_up_safe(path, 0);
1639 }
1640
1641 /*
1642 * In merge_reloc_root(), we modify the upper level pointer to swap the
1643 * tree blocks between reloc tree and subvolume tree. Thus for tree
1644 * block COW, we COW at most from level 1 to root level for each tree.
1645 *
1646 * Thus the needed metadata size is at most root_level * nodesize,
1647 * and * 2 since we have two trees to COW.
1648 */
1649 reserve_level = max_t(int, 1, btrfs_root_level(root_item));
1650 min_reserved = fs_info->nodesize * reserve_level * 2;
1651 memset(&next_key, 0, sizeof(next_key));
1652
1653 while (1) {
1654 ret = btrfs_block_rsv_refill(fs_info, rc->block_rsv,
1655 min_reserved,
1656 BTRFS_RESERVE_FLUSH_LIMIT);
1657 if (ret)
1658 goto out;
1659 trans = btrfs_start_transaction(root, 0);
1660 if (IS_ERR(trans)) {
1661 ret = PTR_ERR(trans);
1662 trans = NULL;
1663 goto out;
1664 }
1665
1666 /*
1667 * At this point we no longer have a reloc_control, so we can't
1668 * depend on btrfs_init_reloc_root to update our last_trans.
1669 *
1670 * But that's ok, we started the trans handle on our
1671 * corresponding fs_root, which means it's been added to the
1672 * dirty list. At commit time we'll still call
1673 * btrfs_update_reloc_root() and update our root item
1674 * appropriately.
1675 */
1676 btrfs_set_root_last_trans(reloc_root, trans->transid);
1677 trans->block_rsv = rc->block_rsv;
1678
1679 replaced = 0;
1680 max_level = level;
1681
1682 ret = walk_down_reloc_tree(reloc_root, path, &level);
1683 if (ret < 0)
1684 goto out;
1685 if (ret > 0)
1686 break;
1687
1688 if (!find_next_key(path, level, &key) &&
1689 btrfs_comp_cpu_keys(&next_key, &key) >= 0) {
1690 ret = 0;
1691 } else {
1692 ret = replace_path(trans, rc, root, reloc_root, path,
1693 &next_key, level, max_level);
1694 }
1695 if (ret < 0)
1696 goto out;
1697 if (ret > 0) {
1698 level = ret;
1699 btrfs_node_key_to_cpu(path->nodes[level], &key,
1700 path->slots[level]);
1701 replaced = 1;
1702 }
1703
1704 ret = walk_up_reloc_tree(reloc_root, path, &level);
1705 if (ret > 0)
1706 break;
1707
1708 BUG_ON(level == 0);
1709 /*
1710 * save the merging progress in the drop_progress.
1711 * this is OK since root refs == 1 in this case.
1712 */
1713 btrfs_node_key(path->nodes[level], &root_item->drop_progress,
1714 path->slots[level]);
1715 btrfs_set_root_drop_level(root_item, level);
1716
1717 btrfs_end_transaction_throttle(trans);
1718 trans = NULL;
1719
1720 btrfs_btree_balance_dirty(fs_info);
1721
1722 if (replaced && rc->stage == UPDATE_DATA_PTRS)
1723 invalidate_extent_cache(root, &key, &next_key);
1724 }
1725
1726 /*
1727 * handle the case only one block in the fs tree need to be
1728 * relocated and the block is tree root.
1729 */
1730 leaf = btrfs_lock_root_node(root);
1731 ret = btrfs_cow_block(trans, root, leaf, NULL, 0, &leaf,
1732 BTRFS_NESTING_COW);
1733 btrfs_tree_unlock(leaf);
1734 free_extent_buffer(leaf);
1735out:
1736 btrfs_free_path(path);
1737
1738 if (ret == 0) {
1739 ret = insert_dirty_subvol(trans, rc, root);
1740 if (ret)
1741 btrfs_abort_transaction(trans, ret);
1742 }
1743
1744 if (trans)
1745 btrfs_end_transaction_throttle(trans);
1746
1747 btrfs_btree_balance_dirty(fs_info);
1748
1749 if (replaced && rc->stage == UPDATE_DATA_PTRS)
1750 invalidate_extent_cache(root, &key, &next_key);
1751
1752 return ret;
1753}
1754
1755static noinline_for_stack
1756int prepare_to_merge(struct reloc_control *rc, int err)
1757{
1758 struct btrfs_root *root = rc->extent_root;
1759 struct btrfs_fs_info *fs_info = root->fs_info;
1760 struct btrfs_root *reloc_root;
1761 struct btrfs_trans_handle *trans;
1762 LIST_HEAD(reloc_roots);
1763 u64 num_bytes = 0;
1764 int ret;
1765
1766 mutex_lock(&fs_info->reloc_mutex);
1767 rc->merging_rsv_size += fs_info->nodesize * (BTRFS_MAX_LEVEL - 1) * 2;
1768 rc->merging_rsv_size += rc->nodes_relocated * 2;
1769 mutex_unlock(&fs_info->reloc_mutex);
1770
1771again:
1772 if (!err) {
1773 num_bytes = rc->merging_rsv_size;
1774 ret = btrfs_block_rsv_add(fs_info, rc->block_rsv, num_bytes,
1775 BTRFS_RESERVE_FLUSH_ALL);
1776 if (ret)
1777 err = ret;
1778 }
1779
1780 trans = btrfs_join_transaction(rc->extent_root);
1781 if (IS_ERR(trans)) {
1782 if (!err)
1783 btrfs_block_rsv_release(fs_info, rc->block_rsv,
1784 num_bytes, NULL);
1785 return PTR_ERR(trans);
1786 }
1787
1788 if (!err) {
1789 if (num_bytes != rc->merging_rsv_size) {
1790 btrfs_end_transaction(trans);
1791 btrfs_block_rsv_release(fs_info, rc->block_rsv,
1792 num_bytes, NULL);
1793 goto again;
1794 }
1795 }
1796
1797 rc->merge_reloc_tree = true;
1798
1799 while (!list_empty(&rc->reloc_roots)) {
1800 reloc_root = list_entry(rc->reloc_roots.next,
1801 struct btrfs_root, root_list);
1802 list_del_init(&reloc_root->root_list);
1803
1804 root = btrfs_get_fs_root(fs_info, reloc_root->root_key.offset,
1805 false);
1806 if (IS_ERR(root)) {
1807 /*
1808 * Even if we have an error we need this reloc root
1809 * back on our list so we can clean up properly.
1810 */
1811 list_add(&reloc_root->root_list, &reloc_roots);
1812 btrfs_abort_transaction(trans, (int)PTR_ERR(root));
1813 if (!err)
1814 err = PTR_ERR(root);
1815 break;
1816 }
1817
1818 if (unlikely(root->reloc_root != reloc_root)) {
1819 if (root->reloc_root) {
1820 btrfs_err(fs_info,
1821"reloc tree mismatch, root %lld has reloc root key (%lld %u %llu) gen %llu, expect reloc root key (%lld %u %llu) gen %llu",
1822 btrfs_root_id(root),
1823 btrfs_root_id(root->reloc_root),
1824 root->reloc_root->root_key.type,
1825 root->reloc_root->root_key.offset,
1826 btrfs_root_generation(
1827 &root->reloc_root->root_item),
1828 btrfs_root_id(reloc_root),
1829 reloc_root->root_key.type,
1830 reloc_root->root_key.offset,
1831 btrfs_root_generation(
1832 &reloc_root->root_item));
1833 } else {
1834 btrfs_err(fs_info,
1835"reloc tree mismatch, root %lld has no reloc root, expect reloc root key (%lld %u %llu) gen %llu",
1836 btrfs_root_id(root),
1837 btrfs_root_id(reloc_root),
1838 reloc_root->root_key.type,
1839 reloc_root->root_key.offset,
1840 btrfs_root_generation(
1841 &reloc_root->root_item));
1842 }
1843 list_add(&reloc_root->root_list, &reloc_roots);
1844 btrfs_put_root(root);
1845 btrfs_abort_transaction(trans, -EUCLEAN);
1846 if (!err)
1847 err = -EUCLEAN;
1848 break;
1849 }
1850
1851 /*
1852 * set reference count to 1, so btrfs_recover_relocation
1853 * knows it should resumes merging
1854 */
1855 if (!err)
1856 btrfs_set_root_refs(&reloc_root->root_item, 1);
1857 ret = btrfs_update_reloc_root(trans, root);
1858
1859 /*
1860 * Even if we have an error we need this reloc root back on our
1861 * list so we can clean up properly.
1862 */
1863 list_add(&reloc_root->root_list, &reloc_roots);
1864 btrfs_put_root(root);
1865
1866 if (ret) {
1867 btrfs_abort_transaction(trans, ret);
1868 if (!err)
1869 err = ret;
1870 break;
1871 }
1872 }
1873
1874 list_splice(&reloc_roots, &rc->reloc_roots);
1875
1876 if (!err)
1877 err = btrfs_commit_transaction(trans);
1878 else
1879 btrfs_end_transaction(trans);
1880 return err;
1881}
1882
1883static noinline_for_stack
1884void free_reloc_roots(struct list_head *list)
1885{
1886 struct btrfs_root *reloc_root, *tmp;
1887
1888 list_for_each_entry_safe(reloc_root, tmp, list, root_list)
1889 __del_reloc_root(reloc_root);
1890}
1891
1892static noinline_for_stack
1893void merge_reloc_roots(struct reloc_control *rc)
1894{
1895 struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
1896 struct btrfs_root *root;
1897 struct btrfs_root *reloc_root;
1898 LIST_HEAD(reloc_roots);
1899 int found = 0;
1900 int ret = 0;
1901again:
1902 root = rc->extent_root;
1903
1904 /*
1905 * this serializes us with btrfs_record_root_in_transaction,
1906 * we have to make sure nobody is in the middle of
1907 * adding their roots to the list while we are
1908 * doing this splice
1909 */
1910 mutex_lock(&fs_info->reloc_mutex);
1911 list_splice_init(&rc->reloc_roots, &reloc_roots);
1912 mutex_unlock(&fs_info->reloc_mutex);
1913
1914 while (!list_empty(&reloc_roots)) {
1915 found = 1;
1916 reloc_root = list_entry(reloc_roots.next,
1917 struct btrfs_root, root_list);
1918
1919 root = btrfs_get_fs_root(fs_info, reloc_root->root_key.offset,
1920 false);
1921 if (btrfs_root_refs(&reloc_root->root_item) > 0) {
1922 if (WARN_ON(IS_ERR(root))) {
1923 /*
1924 * For recovery we read the fs roots on mount,
1925 * and if we didn't find the root then we marked
1926 * the reloc root as a garbage root. For normal
1927 * relocation obviously the root should exist in
1928 * memory. However there's no reason we can't
1929 * handle the error properly here just in case.
1930 */
1931 ret = PTR_ERR(root);
1932 goto out;
1933 }
1934 if (WARN_ON(root->reloc_root != reloc_root)) {
1935 /*
1936 * This can happen if on-disk metadata has some
1937 * corruption, e.g. bad reloc tree key offset.
1938 */
1939 ret = -EINVAL;
1940 goto out;
1941 }
1942 ret = merge_reloc_root(rc, root);
1943 btrfs_put_root(root);
1944 if (ret) {
1945 if (list_empty(&reloc_root->root_list))
1946 list_add_tail(&reloc_root->root_list,
1947 &reloc_roots);
1948 goto out;
1949 }
1950 } else {
1951 if (!IS_ERR(root)) {
1952 if (root->reloc_root == reloc_root) {
1953 root->reloc_root = NULL;
1954 btrfs_put_root(reloc_root);
1955 }
1956 clear_bit(BTRFS_ROOT_DEAD_RELOC_TREE,
1957 &root->state);
1958 btrfs_put_root(root);
1959 }
1960
1961 list_del_init(&reloc_root->root_list);
1962 /* Don't forget to queue this reloc root for cleanup */
1963 list_add_tail(&reloc_root->reloc_dirty_list,
1964 &rc->dirty_subvol_roots);
1965 }
1966 }
1967
1968 if (found) {
1969 found = 0;
1970 goto again;
1971 }
1972out:
1973 if (ret) {
1974 btrfs_handle_fs_error(fs_info, ret, NULL);
1975 free_reloc_roots(&reloc_roots);
1976
1977 /* new reloc root may be added */
1978 mutex_lock(&fs_info->reloc_mutex);
1979 list_splice_init(&rc->reloc_roots, &reloc_roots);
1980 mutex_unlock(&fs_info->reloc_mutex);
1981 free_reloc_roots(&reloc_roots);
1982 }
1983
1984 /*
1985 * We used to have
1986 *
1987 * BUG_ON(!RB_EMPTY_ROOT(&rc->reloc_root_tree.rb_root));
1988 *
1989 * here, but it's wrong. If we fail to start the transaction in
1990 * prepare_to_merge() we will have only 0 ref reloc roots, none of which
1991 * have actually been removed from the reloc_root_tree rb tree. This is
1992 * fine because we're bailing here, and we hold a reference on the root
1993 * for the list that holds it, so these roots will be cleaned up when we
1994 * do the reloc_dirty_list afterwards. Meanwhile the root->reloc_root
1995 * will be cleaned up on unmount.
1996 *
1997 * The remaining nodes will be cleaned up by free_reloc_control.
1998 */
1999}
2000
2001static void free_block_list(struct rb_root *blocks)
2002{
2003 struct tree_block *block;
2004 struct rb_node *rb_node;
2005 while ((rb_node = rb_first(blocks))) {
2006 block = rb_entry(rb_node, struct tree_block, rb_node);
2007 rb_erase(rb_node, blocks);
2008 kfree(block);
2009 }
2010}
2011
2012static int record_reloc_root_in_trans(struct btrfs_trans_handle *trans,
2013 struct btrfs_root *reloc_root)
2014{
2015 struct btrfs_fs_info *fs_info = reloc_root->fs_info;
2016 struct btrfs_root *root;
2017 int ret;
2018
2019 if (btrfs_get_root_last_trans(reloc_root) == trans->transid)
2020 return 0;
2021
2022 root = btrfs_get_fs_root(fs_info, reloc_root->root_key.offset, false);
2023
2024 /*
2025 * This should succeed, since we can't have a reloc root without having
2026 * already looked up the actual root and created the reloc root for this
2027 * root.
2028 *
2029 * However if there's some sort of corruption where we have a ref to a
2030 * reloc root without a corresponding root this could return ENOENT.
2031 */
2032 if (IS_ERR(root)) {
2033 ASSERT(0);
2034 return PTR_ERR(root);
2035 }
2036 if (root->reloc_root != reloc_root) {
2037 ASSERT(0);
2038 btrfs_err(fs_info,
2039 "root %llu has two reloc roots associated with it",
2040 reloc_root->root_key.offset);
2041 btrfs_put_root(root);
2042 return -EUCLEAN;
2043 }
2044 ret = btrfs_record_root_in_trans(trans, root);
2045 btrfs_put_root(root);
2046
2047 return ret;
2048}
2049
2050static noinline_for_stack
2051struct btrfs_root *select_reloc_root(struct btrfs_trans_handle *trans,
2052 struct reloc_control *rc,
2053 struct btrfs_backref_node *node,
2054 struct btrfs_backref_edge *edges[])
2055{
2056 struct btrfs_backref_node *next;
2057 struct btrfs_root *root;
2058 int index = 0;
2059 int ret;
2060
2061 next = node;
2062 while (1) {
2063 cond_resched();
2064 next = walk_up_backref(next, edges, &index);
2065 root = next->root;
2066
2067 /*
2068 * If there is no root, then our references for this block are
2069 * incomplete, as we should be able to walk all the way up to a
2070 * block that is owned by a root.
2071 *
2072 * This path is only for SHAREABLE roots, so if we come upon a
2073 * non-SHAREABLE root then we have backrefs that resolve
2074 * improperly.
2075 *
2076 * Both of these cases indicate file system corruption, or a bug
2077 * in the backref walking code.
2078 */
2079 if (!root) {
2080 ASSERT(0);
2081 btrfs_err(trans->fs_info,
2082 "bytenr %llu doesn't have a backref path ending in a root",
2083 node->bytenr);
2084 return ERR_PTR(-EUCLEAN);
2085 }
2086 if (!test_bit(BTRFS_ROOT_SHAREABLE, &root->state)) {
2087 ASSERT(0);
2088 btrfs_err(trans->fs_info,
2089 "bytenr %llu has multiple refs with one ending in a non-shareable root",
2090 node->bytenr);
2091 return ERR_PTR(-EUCLEAN);
2092 }
2093
2094 if (btrfs_root_id(root) == BTRFS_TREE_RELOC_OBJECTID) {
2095 ret = record_reloc_root_in_trans(trans, root);
2096 if (ret)
2097 return ERR_PTR(ret);
2098 break;
2099 }
2100
2101 ret = btrfs_record_root_in_trans(trans, root);
2102 if (ret)
2103 return ERR_PTR(ret);
2104 root = root->reloc_root;
2105
2106 /*
2107 * We could have raced with another thread which failed, so
2108 * root->reloc_root may not be set, return ENOENT in this case.
2109 */
2110 if (!root)
2111 return ERR_PTR(-ENOENT);
2112
2113 if (next->new_bytenr != root->node->start) {
2114 /*
2115 * We just created the reloc root, so we shouldn't have
2116 * ->new_bytenr set and this shouldn't be in the changed
2117 * list. If it is then we have multiple roots pointing
2118 * at the same bytenr which indicates corruption, or
2119 * we've made a mistake in the backref walking code.
2120 */
2121 ASSERT(next->new_bytenr == 0);
2122 ASSERT(list_empty(&next->list));
2123 if (next->new_bytenr || !list_empty(&next->list)) {
2124 btrfs_err(trans->fs_info,
2125 "bytenr %llu possibly has multiple roots pointing at the same bytenr %llu",
2126 node->bytenr, next->bytenr);
2127 return ERR_PTR(-EUCLEAN);
2128 }
2129
2130 next->new_bytenr = root->node->start;
2131 btrfs_put_root(next->root);
2132 next->root = btrfs_grab_root(root);
2133 ASSERT(next->root);
2134 list_add_tail(&next->list,
2135 &rc->backref_cache.changed);
2136 mark_block_processed(rc, next);
2137 break;
2138 }
2139
2140 WARN_ON(1);
2141 root = NULL;
2142 next = walk_down_backref(edges, &index);
2143 if (!next || next->level <= node->level)
2144 break;
2145 }
2146 if (!root) {
2147 /*
2148 * This can happen if there's fs corruption or if there's a bug
2149 * in the backref lookup code.
2150 */
2151 ASSERT(0);
2152 return ERR_PTR(-ENOENT);
2153 }
2154
2155 next = node;
2156 /* setup backref node path for btrfs_reloc_cow_block */
2157 while (1) {
2158 rc->backref_cache.path[next->level] = next;
2159 if (--index < 0)
2160 break;
2161 next = edges[index]->node[UPPER];
2162 }
2163 return root;
2164}
2165
2166/*
2167 * Select a tree root for relocation.
2168 *
2169 * Return NULL if the block is not shareable. We should use do_relocation() in
2170 * this case.
2171 *
2172 * Return a tree root pointer if the block is shareable.
2173 * Return -ENOENT if the block is root of reloc tree.
2174 */
2175static noinline_for_stack
2176struct btrfs_root *select_one_root(struct btrfs_backref_node *node)
2177{
2178 struct btrfs_backref_node *next;
2179 struct btrfs_root *root;
2180 struct btrfs_root *fs_root = NULL;
2181 struct btrfs_backref_edge *edges[BTRFS_MAX_LEVEL - 1];
2182 int index = 0;
2183
2184 next = node;
2185 while (1) {
2186 cond_resched();
2187 next = walk_up_backref(next, edges, &index);
2188 root = next->root;
2189
2190 /*
2191 * This can occur if we have incomplete extent refs leading all
2192 * the way up a particular path, in this case return -EUCLEAN.
2193 */
2194 if (!root)
2195 return ERR_PTR(-EUCLEAN);
2196
2197 /* No other choice for non-shareable tree */
2198 if (!test_bit(BTRFS_ROOT_SHAREABLE, &root->state))
2199 return root;
2200
2201 if (btrfs_root_id(root) != BTRFS_TREE_RELOC_OBJECTID)
2202 fs_root = root;
2203
2204 if (next != node)
2205 return NULL;
2206
2207 next = walk_down_backref(edges, &index);
2208 if (!next || next->level <= node->level)
2209 break;
2210 }
2211
2212 if (!fs_root)
2213 return ERR_PTR(-ENOENT);
2214 return fs_root;
2215}
2216
2217static noinline_for_stack u64 calcu_metadata_size(struct reloc_control *rc,
2218 struct btrfs_backref_node *node)
2219{
2220 struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
2221 struct btrfs_backref_node *next = node;
2222 struct btrfs_backref_edge *edge;
2223 struct btrfs_backref_edge *edges[BTRFS_MAX_LEVEL - 1];
2224 u64 num_bytes = 0;
2225 int index = 0;
2226
2227 BUG_ON(node->processed);
2228
2229 while (next) {
2230 cond_resched();
2231 while (1) {
2232 if (next->processed)
2233 break;
2234
2235 num_bytes += fs_info->nodesize;
2236
2237 if (list_empty(&next->upper))
2238 break;
2239
2240 edge = list_entry(next->upper.next,
2241 struct btrfs_backref_edge, list[LOWER]);
2242 edges[index++] = edge;
2243 next = edge->node[UPPER];
2244 }
2245 next = walk_down_backref(edges, &index);
2246 }
2247 return num_bytes;
2248}
2249
2250static int reserve_metadata_space(struct btrfs_trans_handle *trans,
2251 struct reloc_control *rc,
2252 struct btrfs_backref_node *node)
2253{
2254 struct btrfs_root *root = rc->extent_root;
2255 struct btrfs_fs_info *fs_info = root->fs_info;
2256 u64 num_bytes;
2257 int ret;
2258 u64 tmp;
2259
2260 num_bytes = calcu_metadata_size(rc, node) * 2;
2261
2262 trans->block_rsv = rc->block_rsv;
2263 rc->reserved_bytes += num_bytes;
2264
2265 /*
2266 * We are under a transaction here so we can only do limited flushing.
2267 * If we get an enospc just kick back -EAGAIN so we know to drop the
2268 * transaction and try to refill when we can flush all the things.
2269 */
2270 ret = btrfs_block_rsv_refill(fs_info, rc->block_rsv, num_bytes,
2271 BTRFS_RESERVE_FLUSH_LIMIT);
2272 if (ret) {
2273 tmp = fs_info->nodesize * RELOCATION_RESERVED_NODES;
2274 while (tmp <= rc->reserved_bytes)
2275 tmp <<= 1;
2276 /*
2277 * only one thread can access block_rsv at this point,
2278 * so we don't need hold lock to protect block_rsv.
2279 * we expand more reservation size here to allow enough
2280 * space for relocation and we will return earlier in
2281 * enospc case.
2282 */
2283 rc->block_rsv->size = tmp + fs_info->nodesize *
2284 RELOCATION_RESERVED_NODES;
2285 return -EAGAIN;
2286 }
2287
2288 return 0;
2289}
2290
2291/*
2292 * relocate a block tree, and then update pointers in upper level
2293 * blocks that reference the block to point to the new location.
2294 *
2295 * if called by link_to_upper, the block has already been relocated.
2296 * in that case this function just updates pointers.
2297 */
2298static int do_relocation(struct btrfs_trans_handle *trans,
2299 struct reloc_control *rc,
2300 struct btrfs_backref_node *node,
2301 struct btrfs_key *key,
2302 struct btrfs_path *path, int lowest)
2303{
2304 struct btrfs_backref_node *upper;
2305 struct btrfs_backref_edge *edge;
2306 struct btrfs_backref_edge *edges[BTRFS_MAX_LEVEL - 1];
2307 struct btrfs_root *root;
2308 struct extent_buffer *eb;
2309 u32 blocksize;
2310 u64 bytenr;
2311 int slot;
2312 int ret = 0;
2313
2314 /*
2315 * If we are lowest then this is the first time we're processing this
2316 * block, and thus shouldn't have an eb associated with it yet.
2317 */
2318 ASSERT(!lowest || !node->eb);
2319
2320 path->lowest_level = node->level + 1;
2321 rc->backref_cache.path[node->level] = node;
2322 list_for_each_entry(edge, &node->upper, list[LOWER]) {
2323 cond_resched();
2324
2325 upper = edge->node[UPPER];
2326 root = select_reloc_root(trans, rc, upper, edges);
2327 if (IS_ERR(root)) {
2328 ret = PTR_ERR(root);
2329 goto next;
2330 }
2331
2332 if (upper->eb && !upper->locked) {
2333 if (!lowest) {
2334 ret = btrfs_bin_search(upper->eb, 0, key, &slot);
2335 if (ret < 0)
2336 goto next;
2337 BUG_ON(ret);
2338 bytenr = btrfs_node_blockptr(upper->eb, slot);
2339 if (node->eb->start == bytenr)
2340 goto next;
2341 }
2342 btrfs_backref_drop_node_buffer(upper);
2343 }
2344
2345 if (!upper->eb) {
2346 ret = btrfs_search_slot(trans, root, key, path, 0, 1);
2347 if (ret) {
2348 if (ret > 0)
2349 ret = -ENOENT;
2350
2351 btrfs_release_path(path);
2352 break;
2353 }
2354
2355 if (!upper->eb) {
2356 upper->eb = path->nodes[upper->level];
2357 path->nodes[upper->level] = NULL;
2358 } else {
2359 BUG_ON(upper->eb != path->nodes[upper->level]);
2360 }
2361
2362 upper->locked = 1;
2363 path->locks[upper->level] = 0;
2364
2365 slot = path->slots[upper->level];
2366 btrfs_release_path(path);
2367 } else {
2368 ret = btrfs_bin_search(upper->eb, 0, key, &slot);
2369 if (ret < 0)
2370 goto next;
2371 BUG_ON(ret);
2372 }
2373
2374 bytenr = btrfs_node_blockptr(upper->eb, slot);
2375 if (lowest) {
2376 if (bytenr != node->bytenr) {
2377 btrfs_err(root->fs_info,
2378 "lowest leaf/node mismatch: bytenr %llu node->bytenr %llu slot %d upper %llu",
2379 bytenr, node->bytenr, slot,
2380 upper->eb->start);
2381 ret = -EIO;
2382 goto next;
2383 }
2384 } else {
2385 if (node->eb->start == bytenr)
2386 goto next;
2387 }
2388
2389 blocksize = root->fs_info->nodesize;
2390 eb = btrfs_read_node_slot(upper->eb, slot);
2391 if (IS_ERR(eb)) {
2392 ret = PTR_ERR(eb);
2393 goto next;
2394 }
2395 btrfs_tree_lock(eb);
2396
2397 if (!node->eb) {
2398 ret = btrfs_cow_block(trans, root, eb, upper->eb,
2399 slot, &eb, BTRFS_NESTING_COW);
2400 btrfs_tree_unlock(eb);
2401 free_extent_buffer(eb);
2402 if (ret < 0)
2403 goto next;
2404 /*
2405 * We've just COWed this block, it should have updated
2406 * the correct backref node entry.
2407 */
2408 ASSERT(node->eb == eb);
2409 } else {
2410 struct btrfs_ref ref = {
2411 .action = BTRFS_ADD_DELAYED_REF,
2412 .bytenr = node->eb->start,
2413 .num_bytes = blocksize,
2414 .parent = upper->eb->start,
2415 .owning_root = btrfs_header_owner(upper->eb),
2416 .ref_root = btrfs_header_owner(upper->eb),
2417 };
2418
2419 btrfs_set_node_blockptr(upper->eb, slot,
2420 node->eb->start);
2421 btrfs_set_node_ptr_generation(upper->eb, slot,
2422 trans->transid);
2423 btrfs_mark_buffer_dirty(trans, upper->eb);
2424
2425 btrfs_init_tree_ref(&ref, node->level,
2426 btrfs_root_id(root), false);
2427 ret = btrfs_inc_extent_ref(trans, &ref);
2428 if (!ret)
2429 ret = btrfs_drop_subtree(trans, root, eb,
2430 upper->eb);
2431 if (ret)
2432 btrfs_abort_transaction(trans, ret);
2433 }
2434next:
2435 if (!upper->pending)
2436 btrfs_backref_drop_node_buffer(upper);
2437 else
2438 btrfs_backref_unlock_node_buffer(upper);
2439 if (ret)
2440 break;
2441 }
2442
2443 if (!ret && node->pending) {
2444 btrfs_backref_drop_node_buffer(node);
2445 list_move_tail(&node->list, &rc->backref_cache.changed);
2446 node->pending = 0;
2447 }
2448
2449 path->lowest_level = 0;
2450
2451 /*
2452 * We should have allocated all of our space in the block rsv and thus
2453 * shouldn't ENOSPC.
2454 */
2455 ASSERT(ret != -ENOSPC);
2456 return ret;
2457}
2458
2459static int link_to_upper(struct btrfs_trans_handle *trans,
2460 struct reloc_control *rc,
2461 struct btrfs_backref_node *node,
2462 struct btrfs_path *path)
2463{
2464 struct btrfs_key key;
2465
2466 btrfs_node_key_to_cpu(node->eb, &key, 0);
2467 return do_relocation(trans, rc, node, &key, path, 0);
2468}
2469
2470static int finish_pending_nodes(struct btrfs_trans_handle *trans,
2471 struct reloc_control *rc,
2472 struct btrfs_path *path, int err)
2473{
2474 LIST_HEAD(list);
2475 struct btrfs_backref_cache *cache = &rc->backref_cache;
2476 struct btrfs_backref_node *node;
2477 int level;
2478 int ret;
2479
2480 for (level = 0; level < BTRFS_MAX_LEVEL; level++) {
2481 while (!list_empty(&cache->pending[level])) {
2482 node = list_entry(cache->pending[level].next,
2483 struct btrfs_backref_node, list);
2484 list_move_tail(&node->list, &list);
2485 BUG_ON(!node->pending);
2486
2487 if (!err) {
2488 ret = link_to_upper(trans, rc, node, path);
2489 if (ret < 0)
2490 err = ret;
2491 }
2492 }
2493 list_splice_init(&list, &cache->pending[level]);
2494 }
2495 return err;
2496}
2497
2498/*
2499 * mark a block and all blocks directly/indirectly reference the block
2500 * as processed.
2501 */
2502static void update_processed_blocks(struct reloc_control *rc,
2503 struct btrfs_backref_node *node)
2504{
2505 struct btrfs_backref_node *next = node;
2506 struct btrfs_backref_edge *edge;
2507 struct btrfs_backref_edge *edges[BTRFS_MAX_LEVEL - 1];
2508 int index = 0;
2509
2510 while (next) {
2511 cond_resched();
2512 while (1) {
2513 if (next->processed)
2514 break;
2515
2516 mark_block_processed(rc, next);
2517
2518 if (list_empty(&next->upper))
2519 break;
2520
2521 edge = list_entry(next->upper.next,
2522 struct btrfs_backref_edge, list[LOWER]);
2523 edges[index++] = edge;
2524 next = edge->node[UPPER];
2525 }
2526 next = walk_down_backref(edges, &index);
2527 }
2528}
2529
2530static int tree_block_processed(u64 bytenr, struct reloc_control *rc)
2531{
2532 u32 blocksize = rc->extent_root->fs_info->nodesize;
2533
2534 if (test_range_bit(&rc->processed_blocks, bytenr,
2535 bytenr + blocksize - 1, EXTENT_DIRTY, NULL))
2536 return 1;
2537 return 0;
2538}
2539
2540static int get_tree_block_key(struct btrfs_fs_info *fs_info,
2541 struct tree_block *block)
2542{
2543 struct btrfs_tree_parent_check check = {
2544 .level = block->level,
2545 .owner_root = block->owner,
2546 .transid = block->key.offset
2547 };
2548 struct extent_buffer *eb;
2549
2550 eb = read_tree_block(fs_info, block->bytenr, &check);
2551 if (IS_ERR(eb))
2552 return PTR_ERR(eb);
2553 if (!extent_buffer_uptodate(eb)) {
2554 free_extent_buffer(eb);
2555 return -EIO;
2556 }
2557 if (block->level == 0)
2558 btrfs_item_key_to_cpu(eb, &block->key, 0);
2559 else
2560 btrfs_node_key_to_cpu(eb, &block->key, 0);
2561 free_extent_buffer(eb);
2562 block->key_ready = true;
2563 return 0;
2564}
2565
2566/*
2567 * helper function to relocate a tree block
2568 */
2569static int relocate_tree_block(struct btrfs_trans_handle *trans,
2570 struct reloc_control *rc,
2571 struct btrfs_backref_node *node,
2572 struct btrfs_key *key,
2573 struct btrfs_path *path)
2574{
2575 struct btrfs_root *root;
2576 int ret = 0;
2577
2578 if (!node)
2579 return 0;
2580
2581 /*
2582 * If we fail here we want to drop our backref_node because we are going
2583 * to start over and regenerate the tree for it.
2584 */
2585 ret = reserve_metadata_space(trans, rc, node);
2586 if (ret)
2587 goto out;
2588
2589 BUG_ON(node->processed);
2590 root = select_one_root(node);
2591 if (IS_ERR(root)) {
2592 ret = PTR_ERR(root);
2593
2594 /* See explanation in select_one_root for the -EUCLEAN case. */
2595 ASSERT(ret == -ENOENT);
2596 if (ret == -ENOENT) {
2597 ret = 0;
2598 update_processed_blocks(rc, node);
2599 }
2600 goto out;
2601 }
2602
2603 if (root) {
2604 if (test_bit(BTRFS_ROOT_SHAREABLE, &root->state)) {
2605 /*
2606 * This block was the root block of a root, and this is
2607 * the first time we're processing the block and thus it
2608 * should not have had the ->new_bytenr modified and
2609 * should have not been included on the changed list.
2610 *
2611 * However in the case of corruption we could have
2612 * multiple refs pointing to the same block improperly,
2613 * and thus we would trip over these checks. ASSERT()
2614 * for the developer case, because it could indicate a
2615 * bug in the backref code, however error out for a
2616 * normal user in the case of corruption.
2617 */
2618 ASSERT(node->new_bytenr == 0);
2619 ASSERT(list_empty(&node->list));
2620 if (node->new_bytenr || !list_empty(&node->list)) {
2621 btrfs_err(root->fs_info,
2622 "bytenr %llu has improper references to it",
2623 node->bytenr);
2624 ret = -EUCLEAN;
2625 goto out;
2626 }
2627 ret = btrfs_record_root_in_trans(trans, root);
2628 if (ret)
2629 goto out;
2630 /*
2631 * Another thread could have failed, need to check if we
2632 * have reloc_root actually set.
2633 */
2634 if (!root->reloc_root) {
2635 ret = -ENOENT;
2636 goto out;
2637 }
2638 root = root->reloc_root;
2639 node->new_bytenr = root->node->start;
2640 btrfs_put_root(node->root);
2641 node->root = btrfs_grab_root(root);
2642 ASSERT(node->root);
2643 list_add_tail(&node->list, &rc->backref_cache.changed);
2644 } else {
2645 path->lowest_level = node->level;
2646 if (root == root->fs_info->chunk_root)
2647 btrfs_reserve_chunk_metadata(trans, false);
2648 ret = btrfs_search_slot(trans, root, key, path, 0, 1);
2649 btrfs_release_path(path);
2650 if (root == root->fs_info->chunk_root)
2651 btrfs_trans_release_chunk_metadata(trans);
2652 if (ret > 0)
2653 ret = 0;
2654 }
2655 if (!ret)
2656 update_processed_blocks(rc, node);
2657 } else {
2658 ret = do_relocation(trans, rc, node, key, path, 1);
2659 }
2660out:
2661 if (ret || node->level == 0 || node->cowonly)
2662 btrfs_backref_cleanup_node(&rc->backref_cache, node);
2663 return ret;
2664}
2665
2666/*
2667 * relocate a list of blocks
2668 */
2669static noinline_for_stack
2670int relocate_tree_blocks(struct btrfs_trans_handle *trans,
2671 struct reloc_control *rc, struct rb_root *blocks)
2672{
2673 struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
2674 struct btrfs_backref_node *node;
2675 struct btrfs_path *path;
2676 struct tree_block *block;
2677 struct tree_block *next;
2678 int ret = 0;
2679
2680 path = btrfs_alloc_path();
2681 if (!path) {
2682 ret = -ENOMEM;
2683 goto out_free_blocks;
2684 }
2685
2686 /* Kick in readahead for tree blocks with missing keys */
2687 rbtree_postorder_for_each_entry_safe(block, next, blocks, rb_node) {
2688 if (!block->key_ready)
2689 btrfs_readahead_tree_block(fs_info, block->bytenr,
2690 block->owner, 0,
2691 block->level);
2692 }
2693
2694 /* Get first keys */
2695 rbtree_postorder_for_each_entry_safe(block, next, blocks, rb_node) {
2696 if (!block->key_ready) {
2697 ret = get_tree_block_key(fs_info, block);
2698 if (ret)
2699 goto out_free_path;
2700 }
2701 }
2702
2703 /* Do tree relocation */
2704 rbtree_postorder_for_each_entry_safe(block, next, blocks, rb_node) {
2705 node = build_backref_tree(trans, rc, &block->key,
2706 block->level, block->bytenr);
2707 if (IS_ERR(node)) {
2708 ret = PTR_ERR(node);
2709 goto out;
2710 }
2711
2712 ret = relocate_tree_block(trans, rc, node, &block->key,
2713 path);
2714 if (ret < 0)
2715 break;
2716 }
2717out:
2718 ret = finish_pending_nodes(trans, rc, path, ret);
2719
2720out_free_path:
2721 btrfs_free_path(path);
2722out_free_blocks:
2723 free_block_list(blocks);
2724 return ret;
2725}
2726
2727static noinline_for_stack int prealloc_file_extent_cluster(struct reloc_control *rc)
2728{
2729 const struct file_extent_cluster *cluster = &rc->cluster;
2730 struct btrfs_inode *inode = BTRFS_I(rc->data_inode);
2731 u64 alloc_hint = 0;
2732 u64 start;
2733 u64 end;
2734 u64 offset = inode->reloc_block_group_start;
2735 u64 num_bytes;
2736 int nr;
2737 int ret = 0;
2738 u64 i_size = i_size_read(&inode->vfs_inode);
2739 u64 prealloc_start = cluster->start - offset;
2740 u64 prealloc_end = cluster->end - offset;
2741 u64 cur_offset = prealloc_start;
2742
2743 /*
2744 * For subpage case, previous i_size may not be aligned to PAGE_SIZE.
2745 * This means the range [i_size, PAGE_END + 1) is filled with zeros by
2746 * btrfs_do_readpage() call of previously relocated file cluster.
2747 *
2748 * If the current cluster starts in the above range, btrfs_do_readpage()
2749 * will skip the read, and relocate_one_folio() will later writeback
2750 * the padding zeros as new data, causing data corruption.
2751 *
2752 * Here we have to manually invalidate the range (i_size, PAGE_END + 1).
2753 */
2754 if (!PAGE_ALIGNED(i_size)) {
2755 struct address_space *mapping = inode->vfs_inode.i_mapping;
2756 struct btrfs_fs_info *fs_info = inode->root->fs_info;
2757 const u32 sectorsize = fs_info->sectorsize;
2758 struct folio *folio;
2759
2760 ASSERT(sectorsize < PAGE_SIZE);
2761 ASSERT(IS_ALIGNED(i_size, sectorsize));
2762
2763 /*
2764 * Subpage can't handle page with DIRTY but without UPTODATE
2765 * bit as it can lead to the following deadlock:
2766 *
2767 * btrfs_read_folio()
2768 * | Page already *locked*
2769 * |- btrfs_lock_and_flush_ordered_range()
2770 * |- btrfs_start_ordered_extent()
2771 * |- extent_write_cache_pages()
2772 * |- lock_page()
2773 * We try to lock the page we already hold.
2774 *
2775 * Here we just writeback the whole data reloc inode, so that
2776 * we will be ensured to have no dirty range in the page, and
2777 * are safe to clear the uptodate bits.
2778 *
2779 * This shouldn't cause too much overhead, as we need to write
2780 * the data back anyway.
2781 */
2782 ret = filemap_write_and_wait(mapping);
2783 if (ret < 0)
2784 return ret;
2785
2786 clear_extent_bits(&inode->io_tree, i_size,
2787 round_up(i_size, PAGE_SIZE) - 1,
2788 EXTENT_UPTODATE);
2789 folio = filemap_lock_folio(mapping, i_size >> PAGE_SHIFT);
2790 /*
2791 * If page is freed we don't need to do anything then, as we
2792 * will re-read the whole page anyway.
2793 */
2794 if (!IS_ERR(folio)) {
2795 btrfs_subpage_clear_uptodate(fs_info, folio, i_size,
2796 round_up(i_size, PAGE_SIZE) - i_size);
2797 folio_unlock(folio);
2798 folio_put(folio);
2799 }
2800 }
2801
2802 BUG_ON(cluster->start != cluster->boundary[0]);
2803 ret = btrfs_alloc_data_chunk_ondemand(inode,
2804 prealloc_end + 1 - prealloc_start);
2805 if (ret)
2806 return ret;
2807
2808 btrfs_inode_lock(inode, 0);
2809 for (nr = 0; nr < cluster->nr; nr++) {
2810 struct extent_state *cached_state = NULL;
2811
2812 start = cluster->boundary[nr] - offset;
2813 if (nr + 1 < cluster->nr)
2814 end = cluster->boundary[nr + 1] - 1 - offset;
2815 else
2816 end = cluster->end - offset;
2817
2818 lock_extent(&inode->io_tree, start, end, &cached_state);
2819 num_bytes = end + 1 - start;
2820 ret = btrfs_prealloc_file_range(&inode->vfs_inode, 0, start,
2821 num_bytes, num_bytes,
2822 end + 1, &alloc_hint);
2823 cur_offset = end + 1;
2824 unlock_extent(&inode->io_tree, start, end, &cached_state);
2825 if (ret)
2826 break;
2827 }
2828 btrfs_inode_unlock(inode, 0);
2829
2830 if (cur_offset < prealloc_end)
2831 btrfs_free_reserved_data_space_noquota(inode->root->fs_info,
2832 prealloc_end + 1 - cur_offset);
2833 return ret;
2834}
2835
2836static noinline_for_stack int setup_relocation_extent_mapping(struct reloc_control *rc)
2837{
2838 struct btrfs_inode *inode = BTRFS_I(rc->data_inode);
2839 struct extent_map *em;
2840 struct extent_state *cached_state = NULL;
2841 u64 offset = inode->reloc_block_group_start;
2842 u64 start = rc->cluster.start - offset;
2843 u64 end = rc->cluster.end - offset;
2844 int ret = 0;
2845
2846 em = alloc_extent_map();
2847 if (!em)
2848 return -ENOMEM;
2849
2850 em->start = start;
2851 em->len = end + 1 - start;
2852 em->disk_bytenr = rc->cluster.start;
2853 em->disk_num_bytes = em->len;
2854 em->ram_bytes = em->len;
2855 em->flags |= EXTENT_FLAG_PINNED;
2856
2857 lock_extent(&inode->io_tree, start, end, &cached_state);
2858 ret = btrfs_replace_extent_map_range(inode, em, false);
2859 unlock_extent(&inode->io_tree, start, end, &cached_state);
2860 free_extent_map(em);
2861
2862 return ret;
2863}
2864
2865/*
2866 * Allow error injection to test balance/relocation cancellation
2867 */
2868noinline int btrfs_should_cancel_balance(const struct btrfs_fs_info *fs_info)
2869{
2870 return atomic_read(&fs_info->balance_cancel_req) ||
2871 atomic_read(&fs_info->reloc_cancel_req) ||
2872 fatal_signal_pending(current);
2873}
2874ALLOW_ERROR_INJECTION(btrfs_should_cancel_balance, TRUE);
2875
2876static u64 get_cluster_boundary_end(const struct file_extent_cluster *cluster,
2877 int cluster_nr)
2878{
2879 /* Last extent, use cluster end directly */
2880 if (cluster_nr >= cluster->nr - 1)
2881 return cluster->end;
2882
2883 /* Use next boundary start*/
2884 return cluster->boundary[cluster_nr + 1] - 1;
2885}
2886
2887static int relocate_one_folio(struct reloc_control *rc,
2888 struct file_ra_state *ra,
2889 int *cluster_nr, unsigned long index)
2890{
2891 const struct file_extent_cluster *cluster = &rc->cluster;
2892 struct inode *inode = rc->data_inode;
2893 struct btrfs_fs_info *fs_info = inode_to_fs_info(inode);
2894 u64 offset = BTRFS_I(inode)->reloc_block_group_start;
2895 const unsigned long last_index = (cluster->end - offset) >> PAGE_SHIFT;
2896 gfp_t mask = btrfs_alloc_write_mask(inode->i_mapping);
2897 struct folio *folio;
2898 u64 folio_start;
2899 u64 folio_end;
2900 u64 cur;
2901 int ret;
2902 const bool use_rst = btrfs_need_stripe_tree_update(fs_info, rc->block_group->flags);
2903
2904 ASSERT(index <= last_index);
2905again:
2906 folio = filemap_lock_folio(inode->i_mapping, index);
2907 if (IS_ERR(folio)) {
2908
2909 /*
2910 * On relocation we're doing readahead on the relocation inode,
2911 * but if the filesystem is backed by a RAID stripe tree we can
2912 * get ENOENT (e.g. due to preallocated extents not being
2913 * mapped in the RST) from the lookup.
2914 *
2915 * But readahead doesn't handle the error and submits invalid
2916 * reads to the device, causing a assertion failures.
2917 */
2918 if (!use_rst)
2919 page_cache_sync_readahead(inode->i_mapping, ra, NULL,
2920 index, last_index + 1 - index);
2921 folio = __filemap_get_folio(inode->i_mapping, index,
2922 FGP_LOCK | FGP_ACCESSED | FGP_CREAT,
2923 mask);
2924 if (IS_ERR(folio))
2925 return PTR_ERR(folio);
2926 }
2927
2928 WARN_ON(folio_order(folio));
2929
2930 if (folio_test_readahead(folio) && !use_rst)
2931 page_cache_async_readahead(inode->i_mapping, ra, NULL,
2932 folio, last_index + 1 - index);
2933
2934 if (!folio_test_uptodate(folio)) {
2935 btrfs_read_folio(NULL, folio);
2936 folio_lock(folio);
2937 if (!folio_test_uptodate(folio)) {
2938 ret = -EIO;
2939 goto release_folio;
2940 }
2941 if (folio->mapping != inode->i_mapping) {
2942 folio_unlock(folio);
2943 folio_put(folio);
2944 goto again;
2945 }
2946 }
2947
2948 /*
2949 * We could have lost folio private when we dropped the lock to read the
2950 * folio above, make sure we set_page_extent_mapped here so we have any
2951 * of the subpage blocksize stuff we need in place.
2952 */
2953 ret = set_folio_extent_mapped(folio);
2954 if (ret < 0)
2955 goto release_folio;
2956
2957 folio_start = folio_pos(folio);
2958 folio_end = folio_start + PAGE_SIZE - 1;
2959
2960 /*
2961 * Start from the cluster, as for subpage case, the cluster can start
2962 * inside the folio.
2963 */
2964 cur = max(folio_start, cluster->boundary[*cluster_nr] - offset);
2965 while (cur <= folio_end) {
2966 struct extent_state *cached_state = NULL;
2967 u64 extent_start = cluster->boundary[*cluster_nr] - offset;
2968 u64 extent_end = get_cluster_boundary_end(cluster,
2969 *cluster_nr) - offset;
2970 u64 clamped_start = max(folio_start, extent_start);
2971 u64 clamped_end = min(folio_end, extent_end);
2972 u32 clamped_len = clamped_end + 1 - clamped_start;
2973
2974 /* Reserve metadata for this range */
2975 ret = btrfs_delalloc_reserve_metadata(BTRFS_I(inode),
2976 clamped_len, clamped_len,
2977 false);
2978 if (ret)
2979 goto release_folio;
2980
2981 /* Mark the range delalloc and dirty for later writeback */
2982 lock_extent(&BTRFS_I(inode)->io_tree, clamped_start, clamped_end,
2983 &cached_state);
2984 ret = btrfs_set_extent_delalloc(BTRFS_I(inode), clamped_start,
2985 clamped_end, 0, &cached_state);
2986 if (ret) {
2987 clear_extent_bit(&BTRFS_I(inode)->io_tree,
2988 clamped_start, clamped_end,
2989 EXTENT_LOCKED | EXTENT_BOUNDARY,
2990 &cached_state);
2991 btrfs_delalloc_release_metadata(BTRFS_I(inode),
2992 clamped_len, true);
2993 btrfs_delalloc_release_extents(BTRFS_I(inode),
2994 clamped_len);
2995 goto release_folio;
2996 }
2997 btrfs_folio_set_dirty(fs_info, folio, clamped_start, clamped_len);
2998
2999 /*
3000 * Set the boundary if it's inside the folio.
3001 * Data relocation requires the destination extents to have the
3002 * same size as the source.
3003 * EXTENT_BOUNDARY bit prevents current extent from being merged
3004 * with previous extent.
3005 */
3006 if (in_range(cluster->boundary[*cluster_nr] - offset, folio_start, PAGE_SIZE)) {
3007 u64 boundary_start = cluster->boundary[*cluster_nr] -
3008 offset;
3009 u64 boundary_end = boundary_start +
3010 fs_info->sectorsize - 1;
3011
3012 set_extent_bit(&BTRFS_I(inode)->io_tree,
3013 boundary_start, boundary_end,
3014 EXTENT_BOUNDARY, NULL);
3015 }
3016 unlock_extent(&BTRFS_I(inode)->io_tree, clamped_start, clamped_end,
3017 &cached_state);
3018 btrfs_delalloc_release_extents(BTRFS_I(inode), clamped_len);
3019 cur += clamped_len;
3020
3021 /* Crossed extent end, go to next extent */
3022 if (cur >= extent_end) {
3023 (*cluster_nr)++;
3024 /* Just finished the last extent of the cluster, exit. */
3025 if (*cluster_nr >= cluster->nr)
3026 break;
3027 }
3028 }
3029 folio_unlock(folio);
3030 folio_put(folio);
3031
3032 balance_dirty_pages_ratelimited(inode->i_mapping);
3033 btrfs_throttle(fs_info);
3034 if (btrfs_should_cancel_balance(fs_info))
3035 ret = -ECANCELED;
3036 return ret;
3037
3038release_folio:
3039 folio_unlock(folio);
3040 folio_put(folio);
3041 return ret;
3042}
3043
3044static int relocate_file_extent_cluster(struct reloc_control *rc)
3045{
3046 struct inode *inode = rc->data_inode;
3047 const struct file_extent_cluster *cluster = &rc->cluster;
3048 u64 offset = BTRFS_I(inode)->reloc_block_group_start;
3049 unsigned long index;
3050 unsigned long last_index;
3051 struct file_ra_state *ra;
3052 int cluster_nr = 0;
3053 int ret = 0;
3054
3055 if (!cluster->nr)
3056 return 0;
3057
3058 ra = kzalloc(sizeof(*ra), GFP_NOFS);
3059 if (!ra)
3060 return -ENOMEM;
3061
3062 ret = prealloc_file_extent_cluster(rc);
3063 if (ret)
3064 goto out;
3065
3066 file_ra_state_init(ra, inode->i_mapping);
3067
3068 ret = setup_relocation_extent_mapping(rc);
3069 if (ret)
3070 goto out;
3071
3072 last_index = (cluster->end - offset) >> PAGE_SHIFT;
3073 for (index = (cluster->start - offset) >> PAGE_SHIFT;
3074 index <= last_index && !ret; index++)
3075 ret = relocate_one_folio(rc, ra, &cluster_nr, index);
3076 if (ret == 0)
3077 WARN_ON(cluster_nr != cluster->nr);
3078out:
3079 kfree(ra);
3080 return ret;
3081}
3082
3083static noinline_for_stack int relocate_data_extent(struct reloc_control *rc,
3084 const struct btrfs_key *extent_key)
3085{
3086 struct inode *inode = rc->data_inode;
3087 struct file_extent_cluster *cluster = &rc->cluster;
3088 int ret;
3089 struct btrfs_root *root = BTRFS_I(inode)->root;
3090
3091 if (cluster->nr > 0 && extent_key->objectid != cluster->end + 1) {
3092 ret = relocate_file_extent_cluster(rc);
3093 if (ret)
3094 return ret;
3095 cluster->nr = 0;
3096 }
3097
3098 /*
3099 * Under simple quotas, we set root->relocation_src_root when we find
3100 * the extent. If adjacent extents have different owners, we can't merge
3101 * them while relocating. Handle this by storing the owning root that
3102 * started a cluster and if we see an extent from a different root break
3103 * cluster formation (just like the above case of non-adjacent extents).
3104 *
3105 * Without simple quotas, relocation_src_root is always 0, so we should
3106 * never see a mismatch, and it should have no effect on relocation
3107 * clusters.
3108 */
3109 if (cluster->nr > 0 && cluster->owning_root != root->relocation_src_root) {
3110 u64 tmp = root->relocation_src_root;
3111
3112 /*
3113 * root->relocation_src_root is the state that actually affects
3114 * the preallocation we do here, so set it to the root owning
3115 * the cluster we need to relocate.
3116 */
3117 root->relocation_src_root = cluster->owning_root;
3118 ret = relocate_file_extent_cluster(rc);
3119 if (ret)
3120 return ret;
3121 cluster->nr = 0;
3122 /* And reset it back for the current extent's owning root. */
3123 root->relocation_src_root = tmp;
3124 }
3125
3126 if (!cluster->nr) {
3127 cluster->start = extent_key->objectid;
3128 cluster->owning_root = root->relocation_src_root;
3129 }
3130 else
3131 BUG_ON(cluster->nr >= MAX_EXTENTS);
3132 cluster->end = extent_key->objectid + extent_key->offset - 1;
3133 cluster->boundary[cluster->nr] = extent_key->objectid;
3134 cluster->nr++;
3135
3136 if (cluster->nr >= MAX_EXTENTS) {
3137 ret = relocate_file_extent_cluster(rc);
3138 if (ret)
3139 return ret;
3140 cluster->nr = 0;
3141 }
3142 return 0;
3143}
3144
3145/*
3146 * helper to add a tree block to the list.
3147 * the major work is getting the generation and level of the block
3148 */
3149static int add_tree_block(struct reloc_control *rc,
3150 const struct btrfs_key *extent_key,
3151 struct btrfs_path *path,
3152 struct rb_root *blocks)
3153{
3154 struct extent_buffer *eb;
3155 struct btrfs_extent_item *ei;
3156 struct btrfs_tree_block_info *bi;
3157 struct tree_block *block;
3158 struct rb_node *rb_node;
3159 u32 item_size;
3160 int level = -1;
3161 u64 generation;
3162 u64 owner = 0;
3163
3164 eb = path->nodes[0];
3165 item_size = btrfs_item_size(eb, path->slots[0]);
3166
3167 if (extent_key->type == BTRFS_METADATA_ITEM_KEY ||
3168 item_size >= sizeof(*ei) + sizeof(*bi)) {
3169 unsigned long ptr = 0, end;
3170
3171 ei = btrfs_item_ptr(eb, path->slots[0],
3172 struct btrfs_extent_item);
3173 end = (unsigned long)ei + item_size;
3174 if (extent_key->type == BTRFS_EXTENT_ITEM_KEY) {
3175 bi = (struct btrfs_tree_block_info *)(ei + 1);
3176 level = btrfs_tree_block_level(eb, bi);
3177 ptr = (unsigned long)(bi + 1);
3178 } else {
3179 level = (int)extent_key->offset;
3180 ptr = (unsigned long)(ei + 1);
3181 }
3182 generation = btrfs_extent_generation(eb, ei);
3183
3184 /*
3185 * We're reading random blocks without knowing their owner ahead
3186 * of time. This is ok most of the time, as all reloc roots and
3187 * fs roots have the same lock type. However normal trees do
3188 * not, and the only way to know ahead of time is to read the
3189 * inline ref offset. We know it's an fs root if
3190 *
3191 * 1. There's more than one ref.
3192 * 2. There's a SHARED_DATA_REF_KEY set.
3193 * 3. FULL_BACKREF is set on the flags.
3194 *
3195 * Otherwise it's safe to assume that the ref offset == the
3196 * owner of this block, so we can use that when calling
3197 * read_tree_block.
3198 */
3199 if (btrfs_extent_refs(eb, ei) == 1 &&
3200 !(btrfs_extent_flags(eb, ei) &
3201 BTRFS_BLOCK_FLAG_FULL_BACKREF) &&
3202 ptr < end) {
3203 struct btrfs_extent_inline_ref *iref;
3204 int type;
3205
3206 iref = (struct btrfs_extent_inline_ref *)ptr;
3207 type = btrfs_get_extent_inline_ref_type(eb, iref,
3208 BTRFS_REF_TYPE_BLOCK);
3209 if (type == BTRFS_REF_TYPE_INVALID)
3210 return -EINVAL;
3211 if (type == BTRFS_TREE_BLOCK_REF_KEY)
3212 owner = btrfs_extent_inline_ref_offset(eb, iref);
3213 }
3214 } else {
3215 btrfs_print_leaf(eb);
3216 btrfs_err(rc->block_group->fs_info,
3217 "unrecognized tree backref at tree block %llu slot %u",
3218 eb->start, path->slots[0]);
3219 btrfs_release_path(path);
3220 return -EUCLEAN;
3221 }
3222
3223 btrfs_release_path(path);
3224
3225 BUG_ON(level == -1);
3226
3227 block = kmalloc(sizeof(*block), GFP_NOFS);
3228 if (!block)
3229 return -ENOMEM;
3230
3231 block->bytenr = extent_key->objectid;
3232 block->key.objectid = rc->extent_root->fs_info->nodesize;
3233 block->key.offset = generation;
3234 block->level = level;
3235 block->key_ready = false;
3236 block->owner = owner;
3237
3238 rb_node = rb_simple_insert(blocks, block->bytenr, &block->rb_node);
3239 if (rb_node)
3240 btrfs_backref_panic(rc->extent_root->fs_info, block->bytenr,
3241 -EEXIST);
3242
3243 return 0;
3244}
3245
3246/*
3247 * helper to add tree blocks for backref of type BTRFS_SHARED_DATA_REF_KEY
3248 */
3249static int __add_tree_block(struct reloc_control *rc,
3250 u64 bytenr, u32 blocksize,
3251 struct rb_root *blocks)
3252{
3253 struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
3254 struct btrfs_path *path;
3255 struct btrfs_key key;
3256 int ret;
3257 bool skinny = btrfs_fs_incompat(fs_info, SKINNY_METADATA);
3258
3259 if (tree_block_processed(bytenr, rc))
3260 return 0;
3261
3262 if (rb_simple_search(blocks, bytenr))
3263 return 0;
3264
3265 path = btrfs_alloc_path();
3266 if (!path)
3267 return -ENOMEM;
3268again:
3269 key.objectid = bytenr;
3270 if (skinny) {
3271 key.type = BTRFS_METADATA_ITEM_KEY;
3272 key.offset = (u64)-1;
3273 } else {
3274 key.type = BTRFS_EXTENT_ITEM_KEY;
3275 key.offset = blocksize;
3276 }
3277
3278 path->search_commit_root = 1;
3279 path->skip_locking = 1;
3280 ret = btrfs_search_slot(NULL, rc->extent_root, &key, path, 0, 0);
3281 if (ret < 0)
3282 goto out;
3283
3284 if (ret > 0 && skinny) {
3285 if (path->slots[0]) {
3286 path->slots[0]--;
3287 btrfs_item_key_to_cpu(path->nodes[0], &key,
3288 path->slots[0]);
3289 if (key.objectid == bytenr &&
3290 (key.type == BTRFS_METADATA_ITEM_KEY ||
3291 (key.type == BTRFS_EXTENT_ITEM_KEY &&
3292 key.offset == blocksize)))
3293 ret = 0;
3294 }
3295
3296 if (ret) {
3297 skinny = false;
3298 btrfs_release_path(path);
3299 goto again;
3300 }
3301 }
3302 if (ret) {
3303 ASSERT(ret == 1);
3304 btrfs_print_leaf(path->nodes[0]);
3305 btrfs_err(fs_info,
3306 "tree block extent item (%llu) is not found in extent tree",
3307 bytenr);
3308 WARN_ON(1);
3309 ret = -EINVAL;
3310 goto out;
3311 }
3312
3313 ret = add_tree_block(rc, &key, path, blocks);
3314out:
3315 btrfs_free_path(path);
3316 return ret;
3317}
3318
3319static int delete_block_group_cache(struct btrfs_fs_info *fs_info,
3320 struct btrfs_block_group *block_group,
3321 struct inode *inode,
3322 u64 ino)
3323{
3324 struct btrfs_root *root = fs_info->tree_root;
3325 struct btrfs_trans_handle *trans;
3326 int ret = 0;
3327
3328 if (inode)
3329 goto truncate;
3330
3331 inode = btrfs_iget(ino, root);
3332 if (IS_ERR(inode))
3333 return -ENOENT;
3334
3335truncate:
3336 ret = btrfs_check_trunc_cache_free_space(fs_info,
3337 &fs_info->global_block_rsv);
3338 if (ret)
3339 goto out;
3340
3341 trans = btrfs_join_transaction(root);
3342 if (IS_ERR(trans)) {
3343 ret = PTR_ERR(trans);
3344 goto out;
3345 }
3346
3347 ret = btrfs_truncate_free_space_cache(trans, block_group, inode);
3348
3349 btrfs_end_transaction(trans);
3350 btrfs_btree_balance_dirty(fs_info);
3351out:
3352 iput(inode);
3353 return ret;
3354}
3355
3356/*
3357 * Locate the free space cache EXTENT_DATA in root tree leaf and delete the
3358 * cache inode, to avoid free space cache data extent blocking data relocation.
3359 */
3360static int delete_v1_space_cache(struct extent_buffer *leaf,
3361 struct btrfs_block_group *block_group,
3362 u64 data_bytenr)
3363{
3364 u64 space_cache_ino;
3365 struct btrfs_file_extent_item *ei;
3366 struct btrfs_key key;
3367 bool found = false;
3368 int i;
3369 int ret;
3370
3371 if (btrfs_header_owner(leaf) != BTRFS_ROOT_TREE_OBJECTID)
3372 return 0;
3373
3374 for (i = 0; i < btrfs_header_nritems(leaf); i++) {
3375 u8 type;
3376
3377 btrfs_item_key_to_cpu(leaf, &key, i);
3378 if (key.type != BTRFS_EXTENT_DATA_KEY)
3379 continue;
3380 ei = btrfs_item_ptr(leaf, i, struct btrfs_file_extent_item);
3381 type = btrfs_file_extent_type(leaf, ei);
3382
3383 if ((type == BTRFS_FILE_EXTENT_REG ||
3384 type == BTRFS_FILE_EXTENT_PREALLOC) &&
3385 btrfs_file_extent_disk_bytenr(leaf, ei) == data_bytenr) {
3386 found = true;
3387 space_cache_ino = key.objectid;
3388 break;
3389 }
3390 }
3391 if (!found)
3392 return -ENOENT;
3393 ret = delete_block_group_cache(leaf->fs_info, block_group, NULL,
3394 space_cache_ino);
3395 return ret;
3396}
3397
3398/*
3399 * helper to find all tree blocks that reference a given data extent
3400 */
3401static noinline_for_stack int add_data_references(struct reloc_control *rc,
3402 const struct btrfs_key *extent_key,
3403 struct btrfs_path *path,
3404 struct rb_root *blocks)
3405{
3406 struct btrfs_backref_walk_ctx ctx = { 0 };
3407 struct ulist_iterator leaf_uiter;
3408 struct ulist_node *ref_node = NULL;
3409 const u32 blocksize = rc->extent_root->fs_info->nodesize;
3410 int ret = 0;
3411
3412 btrfs_release_path(path);
3413
3414 ctx.bytenr = extent_key->objectid;
3415 ctx.skip_inode_ref_list = true;
3416 ctx.fs_info = rc->extent_root->fs_info;
3417
3418 ret = btrfs_find_all_leafs(&ctx);
3419 if (ret < 0)
3420 return ret;
3421
3422 ULIST_ITER_INIT(&leaf_uiter);
3423 while ((ref_node = ulist_next(ctx.refs, &leaf_uiter))) {
3424 struct btrfs_tree_parent_check check = { 0 };
3425 struct extent_buffer *eb;
3426
3427 eb = read_tree_block(ctx.fs_info, ref_node->val, &check);
3428 if (IS_ERR(eb)) {
3429 ret = PTR_ERR(eb);
3430 break;
3431 }
3432 ret = delete_v1_space_cache(eb, rc->block_group,
3433 extent_key->objectid);
3434 free_extent_buffer(eb);
3435 if (ret < 0)
3436 break;
3437 ret = __add_tree_block(rc, ref_node->val, blocksize, blocks);
3438 if (ret < 0)
3439 break;
3440 }
3441 if (ret < 0)
3442 free_block_list(blocks);
3443 ulist_free(ctx.refs);
3444 return ret;
3445}
3446
3447/*
3448 * helper to find next unprocessed extent
3449 */
3450static noinline_for_stack
3451int find_next_extent(struct reloc_control *rc, struct btrfs_path *path,
3452 struct btrfs_key *extent_key)
3453{
3454 struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
3455 struct btrfs_key key;
3456 struct extent_buffer *leaf;
3457 u64 start, end, last;
3458 int ret;
3459
3460 last = rc->block_group->start + rc->block_group->length;
3461 while (1) {
3462 bool block_found;
3463
3464 cond_resched();
3465 if (rc->search_start >= last) {
3466 ret = 1;
3467 break;
3468 }
3469
3470 key.objectid = rc->search_start;
3471 key.type = BTRFS_EXTENT_ITEM_KEY;
3472 key.offset = 0;
3473
3474 path->search_commit_root = 1;
3475 path->skip_locking = 1;
3476 ret = btrfs_search_slot(NULL, rc->extent_root, &key, path,
3477 0, 0);
3478 if (ret < 0)
3479 break;
3480next:
3481 leaf = path->nodes[0];
3482 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
3483 ret = btrfs_next_leaf(rc->extent_root, path);
3484 if (ret != 0)
3485 break;
3486 leaf = path->nodes[0];
3487 }
3488
3489 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
3490 if (key.objectid >= last) {
3491 ret = 1;
3492 break;
3493 }
3494
3495 if (key.type != BTRFS_EXTENT_ITEM_KEY &&
3496 key.type != BTRFS_METADATA_ITEM_KEY) {
3497 path->slots[0]++;
3498 goto next;
3499 }
3500
3501 if (key.type == BTRFS_EXTENT_ITEM_KEY &&
3502 key.objectid + key.offset <= rc->search_start) {
3503 path->slots[0]++;
3504 goto next;
3505 }
3506
3507 if (key.type == BTRFS_METADATA_ITEM_KEY &&
3508 key.objectid + fs_info->nodesize <=
3509 rc->search_start) {
3510 path->slots[0]++;
3511 goto next;
3512 }
3513
3514 block_found = find_first_extent_bit(&rc->processed_blocks,
3515 key.objectid, &start, &end,
3516 EXTENT_DIRTY, NULL);
3517
3518 if (block_found && start <= key.objectid) {
3519 btrfs_release_path(path);
3520 rc->search_start = end + 1;
3521 } else {
3522 if (key.type == BTRFS_EXTENT_ITEM_KEY)
3523 rc->search_start = key.objectid + key.offset;
3524 else
3525 rc->search_start = key.objectid +
3526 fs_info->nodesize;
3527 memcpy(extent_key, &key, sizeof(key));
3528 return 0;
3529 }
3530 }
3531 btrfs_release_path(path);
3532 return ret;
3533}
3534
3535static void set_reloc_control(struct reloc_control *rc)
3536{
3537 struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
3538
3539 mutex_lock(&fs_info->reloc_mutex);
3540 fs_info->reloc_ctl = rc;
3541 mutex_unlock(&fs_info->reloc_mutex);
3542}
3543
3544static void unset_reloc_control(struct reloc_control *rc)
3545{
3546 struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
3547
3548 mutex_lock(&fs_info->reloc_mutex);
3549 fs_info->reloc_ctl = NULL;
3550 mutex_unlock(&fs_info->reloc_mutex);
3551}
3552
3553static noinline_for_stack
3554int prepare_to_relocate(struct reloc_control *rc)
3555{
3556 struct btrfs_trans_handle *trans;
3557 int ret;
3558
3559 rc->block_rsv = btrfs_alloc_block_rsv(rc->extent_root->fs_info,
3560 BTRFS_BLOCK_RSV_TEMP);
3561 if (!rc->block_rsv)
3562 return -ENOMEM;
3563
3564 memset(&rc->cluster, 0, sizeof(rc->cluster));
3565 rc->search_start = rc->block_group->start;
3566 rc->extents_found = 0;
3567 rc->nodes_relocated = 0;
3568 rc->merging_rsv_size = 0;
3569 rc->reserved_bytes = 0;
3570 rc->block_rsv->size = rc->extent_root->fs_info->nodesize *
3571 RELOCATION_RESERVED_NODES;
3572 ret = btrfs_block_rsv_refill(rc->extent_root->fs_info,
3573 rc->block_rsv, rc->block_rsv->size,
3574 BTRFS_RESERVE_FLUSH_ALL);
3575 if (ret)
3576 return ret;
3577
3578 rc->create_reloc_tree = true;
3579 set_reloc_control(rc);
3580
3581 trans = btrfs_join_transaction(rc->extent_root);
3582 if (IS_ERR(trans)) {
3583 unset_reloc_control(rc);
3584 /*
3585 * extent tree is not a ref_cow tree and has no reloc_root to
3586 * cleanup. And callers are responsible to free the above
3587 * block rsv.
3588 */
3589 return PTR_ERR(trans);
3590 }
3591
3592 ret = btrfs_commit_transaction(trans);
3593 if (ret)
3594 unset_reloc_control(rc);
3595
3596 return ret;
3597}
3598
3599static noinline_for_stack int relocate_block_group(struct reloc_control *rc)
3600{
3601 struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
3602 struct rb_root blocks = RB_ROOT;
3603 struct btrfs_key key;
3604 struct btrfs_trans_handle *trans = NULL;
3605 struct btrfs_path *path;
3606 struct btrfs_extent_item *ei;
3607 u64 flags;
3608 int ret;
3609 int err = 0;
3610 int progress = 0;
3611
3612 path = btrfs_alloc_path();
3613 if (!path)
3614 return -ENOMEM;
3615 path->reada = READA_FORWARD;
3616
3617 ret = prepare_to_relocate(rc);
3618 if (ret) {
3619 err = ret;
3620 goto out_free;
3621 }
3622
3623 while (1) {
3624 rc->reserved_bytes = 0;
3625 ret = btrfs_block_rsv_refill(fs_info, rc->block_rsv,
3626 rc->block_rsv->size,
3627 BTRFS_RESERVE_FLUSH_ALL);
3628 if (ret) {
3629 err = ret;
3630 break;
3631 }
3632 progress++;
3633 trans = btrfs_start_transaction(rc->extent_root, 0);
3634 if (IS_ERR(trans)) {
3635 err = PTR_ERR(trans);
3636 trans = NULL;
3637 break;
3638 }
3639restart:
3640 if (rc->backref_cache.last_trans != trans->transid)
3641 btrfs_backref_release_cache(&rc->backref_cache);
3642 rc->backref_cache.last_trans = trans->transid;
3643
3644 ret = find_next_extent(rc, path, &key);
3645 if (ret < 0)
3646 err = ret;
3647 if (ret != 0)
3648 break;
3649
3650 rc->extents_found++;
3651
3652 ei = btrfs_item_ptr(path->nodes[0], path->slots[0],
3653 struct btrfs_extent_item);
3654 flags = btrfs_extent_flags(path->nodes[0], ei);
3655
3656 /*
3657 * If we are relocating a simple quota owned extent item, we
3658 * need to note the owner on the reloc data root so that when
3659 * we allocate the replacement item, we can attribute it to the
3660 * correct eventual owner (rather than the reloc data root).
3661 */
3662 if (btrfs_qgroup_mode(fs_info) == BTRFS_QGROUP_MODE_SIMPLE) {
3663 struct btrfs_root *root = BTRFS_I(rc->data_inode)->root;
3664 u64 owning_root_id = btrfs_get_extent_owner_root(fs_info,
3665 path->nodes[0],
3666 path->slots[0]);
3667
3668 root->relocation_src_root = owning_root_id;
3669 }
3670
3671 if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
3672 ret = add_tree_block(rc, &key, path, &blocks);
3673 } else if (rc->stage == UPDATE_DATA_PTRS &&
3674 (flags & BTRFS_EXTENT_FLAG_DATA)) {
3675 ret = add_data_references(rc, &key, path, &blocks);
3676 } else {
3677 btrfs_release_path(path);
3678 ret = 0;
3679 }
3680 if (ret < 0) {
3681 err = ret;
3682 break;
3683 }
3684
3685 if (!RB_EMPTY_ROOT(&blocks)) {
3686 ret = relocate_tree_blocks(trans, rc, &blocks);
3687 if (ret < 0) {
3688 if (ret != -EAGAIN) {
3689 err = ret;
3690 break;
3691 }
3692 rc->extents_found--;
3693 rc->search_start = key.objectid;
3694 }
3695 }
3696
3697 btrfs_end_transaction_throttle(trans);
3698 btrfs_btree_balance_dirty(fs_info);
3699 trans = NULL;
3700
3701 if (rc->stage == MOVE_DATA_EXTENTS &&
3702 (flags & BTRFS_EXTENT_FLAG_DATA)) {
3703 rc->found_file_extent = true;
3704 ret = relocate_data_extent(rc, &key);
3705 if (ret < 0) {
3706 err = ret;
3707 break;
3708 }
3709 }
3710 if (btrfs_should_cancel_balance(fs_info)) {
3711 err = -ECANCELED;
3712 break;
3713 }
3714 }
3715 if (trans && progress && err == -ENOSPC) {
3716 ret = btrfs_force_chunk_alloc(trans, rc->block_group->flags);
3717 if (ret == 1) {
3718 err = 0;
3719 progress = 0;
3720 goto restart;
3721 }
3722 }
3723
3724 btrfs_release_path(path);
3725 clear_extent_bits(&rc->processed_blocks, 0, (u64)-1, EXTENT_DIRTY);
3726
3727 if (trans) {
3728 btrfs_end_transaction_throttle(trans);
3729 btrfs_btree_balance_dirty(fs_info);
3730 }
3731
3732 if (!err) {
3733 ret = relocate_file_extent_cluster(rc);
3734 if (ret < 0)
3735 err = ret;
3736 }
3737
3738 rc->create_reloc_tree = false;
3739 set_reloc_control(rc);
3740
3741 btrfs_backref_release_cache(&rc->backref_cache);
3742 btrfs_block_rsv_release(fs_info, rc->block_rsv, (u64)-1, NULL);
3743
3744 /*
3745 * Even in the case when the relocation is cancelled, we should all go
3746 * through prepare_to_merge() and merge_reloc_roots().
3747 *
3748 * For error (including cancelled balance), prepare_to_merge() will
3749 * mark all reloc trees orphan, then queue them for cleanup in
3750 * merge_reloc_roots()
3751 */
3752 err = prepare_to_merge(rc, err);
3753
3754 merge_reloc_roots(rc);
3755
3756 rc->merge_reloc_tree = false;
3757 unset_reloc_control(rc);
3758 btrfs_block_rsv_release(fs_info, rc->block_rsv, (u64)-1, NULL);
3759
3760 /* get rid of pinned extents */
3761 trans = btrfs_join_transaction(rc->extent_root);
3762 if (IS_ERR(trans)) {
3763 err = PTR_ERR(trans);
3764 goto out_free;
3765 }
3766 ret = btrfs_commit_transaction(trans);
3767 if (ret && !err)
3768 err = ret;
3769out_free:
3770 ret = clean_dirty_subvols(rc);
3771 if (ret < 0 && !err)
3772 err = ret;
3773 btrfs_free_block_rsv(fs_info, rc->block_rsv);
3774 btrfs_free_path(path);
3775 return err;
3776}
3777
3778static int __insert_orphan_inode(struct btrfs_trans_handle *trans,
3779 struct btrfs_root *root, u64 objectid)
3780{
3781 struct btrfs_path *path;
3782 struct btrfs_inode_item *item;
3783 struct extent_buffer *leaf;
3784 int ret;
3785
3786 path = btrfs_alloc_path();
3787 if (!path)
3788 return -ENOMEM;
3789
3790 ret = btrfs_insert_empty_inode(trans, root, path, objectid);
3791 if (ret)
3792 goto out;
3793
3794 leaf = path->nodes[0];
3795 item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_inode_item);
3796 memzero_extent_buffer(leaf, (unsigned long)item, sizeof(*item));
3797 btrfs_set_inode_generation(leaf, item, 1);
3798 btrfs_set_inode_size(leaf, item, 0);
3799 btrfs_set_inode_mode(leaf, item, S_IFREG | 0600);
3800 btrfs_set_inode_flags(leaf, item, BTRFS_INODE_NOCOMPRESS |
3801 BTRFS_INODE_PREALLOC);
3802 btrfs_mark_buffer_dirty(trans, leaf);
3803out:
3804 btrfs_free_path(path);
3805 return ret;
3806}
3807
3808static void delete_orphan_inode(struct btrfs_trans_handle *trans,
3809 struct btrfs_root *root, u64 objectid)
3810{
3811 struct btrfs_path *path;
3812 struct btrfs_key key;
3813 int ret = 0;
3814
3815 path = btrfs_alloc_path();
3816 if (!path) {
3817 ret = -ENOMEM;
3818 goto out;
3819 }
3820
3821 key.objectid = objectid;
3822 key.type = BTRFS_INODE_ITEM_KEY;
3823 key.offset = 0;
3824 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
3825 if (ret) {
3826 if (ret > 0)
3827 ret = -ENOENT;
3828 goto out;
3829 }
3830 ret = btrfs_del_item(trans, root, path);
3831out:
3832 if (ret)
3833 btrfs_abort_transaction(trans, ret);
3834 btrfs_free_path(path);
3835}
3836
3837/*
3838 * helper to create inode for data relocation.
3839 * the inode is in data relocation tree and its link count is 0
3840 */
3841static noinline_for_stack struct inode *create_reloc_inode(
3842 struct btrfs_fs_info *fs_info,
3843 const struct btrfs_block_group *group)
3844{
3845 struct inode *inode = NULL;
3846 struct btrfs_trans_handle *trans;
3847 struct btrfs_root *root;
3848 u64 objectid;
3849 int ret = 0;
3850
3851 root = btrfs_grab_root(fs_info->data_reloc_root);
3852 trans = btrfs_start_transaction(root, 6);
3853 if (IS_ERR(trans)) {
3854 btrfs_put_root(root);
3855 return ERR_CAST(trans);
3856 }
3857
3858 ret = btrfs_get_free_objectid(root, &objectid);
3859 if (ret)
3860 goto out;
3861
3862 ret = __insert_orphan_inode(trans, root, objectid);
3863 if (ret)
3864 goto out;
3865
3866 inode = btrfs_iget(objectid, root);
3867 if (IS_ERR(inode)) {
3868 delete_orphan_inode(trans, root, objectid);
3869 ret = PTR_ERR(inode);
3870 inode = NULL;
3871 goto out;
3872 }
3873 BTRFS_I(inode)->reloc_block_group_start = group->start;
3874
3875 ret = btrfs_orphan_add(trans, BTRFS_I(inode));
3876out:
3877 btrfs_put_root(root);
3878 btrfs_end_transaction(trans);
3879 btrfs_btree_balance_dirty(fs_info);
3880 if (ret) {
3881 iput(inode);
3882 inode = ERR_PTR(ret);
3883 }
3884 return inode;
3885}
3886
3887/*
3888 * Mark start of chunk relocation that is cancellable. Check if the cancellation
3889 * has been requested meanwhile and don't start in that case.
3890 *
3891 * Return:
3892 * 0 success
3893 * -EINPROGRESS operation is already in progress, that's probably a bug
3894 * -ECANCELED cancellation request was set before the operation started
3895 */
3896static int reloc_chunk_start(struct btrfs_fs_info *fs_info)
3897{
3898 if (test_and_set_bit(BTRFS_FS_RELOC_RUNNING, &fs_info->flags)) {
3899 /* This should not happen */
3900 btrfs_err(fs_info, "reloc already running, cannot start");
3901 return -EINPROGRESS;
3902 }
3903
3904 if (atomic_read(&fs_info->reloc_cancel_req) > 0) {
3905 btrfs_info(fs_info, "chunk relocation canceled on start");
3906 /*
3907 * On cancel, clear all requests but let the caller mark
3908 * the end after cleanup operations.
3909 */
3910 atomic_set(&fs_info->reloc_cancel_req, 0);
3911 return -ECANCELED;
3912 }
3913 return 0;
3914}
3915
3916/*
3917 * Mark end of chunk relocation that is cancellable and wake any waiters.
3918 */
3919static void reloc_chunk_end(struct btrfs_fs_info *fs_info)
3920{
3921 /* Requested after start, clear bit first so any waiters can continue */
3922 if (atomic_read(&fs_info->reloc_cancel_req) > 0)
3923 btrfs_info(fs_info, "chunk relocation canceled during operation");
3924 clear_and_wake_up_bit(BTRFS_FS_RELOC_RUNNING, &fs_info->flags);
3925 atomic_set(&fs_info->reloc_cancel_req, 0);
3926}
3927
3928static struct reloc_control *alloc_reloc_control(struct btrfs_fs_info *fs_info)
3929{
3930 struct reloc_control *rc;
3931
3932 rc = kzalloc(sizeof(*rc), GFP_NOFS);
3933 if (!rc)
3934 return NULL;
3935
3936 INIT_LIST_HEAD(&rc->reloc_roots);
3937 INIT_LIST_HEAD(&rc->dirty_subvol_roots);
3938 btrfs_backref_init_cache(fs_info, &rc->backref_cache, true);
3939 rc->reloc_root_tree.rb_root = RB_ROOT;
3940 spin_lock_init(&rc->reloc_root_tree.lock);
3941 extent_io_tree_init(fs_info, &rc->processed_blocks, IO_TREE_RELOC_BLOCKS);
3942 return rc;
3943}
3944
3945static void free_reloc_control(struct reloc_control *rc)
3946{
3947 struct mapping_node *node, *tmp;
3948
3949 free_reloc_roots(&rc->reloc_roots);
3950 rbtree_postorder_for_each_entry_safe(node, tmp,
3951 &rc->reloc_root_tree.rb_root, rb_node)
3952 kfree(node);
3953
3954 kfree(rc);
3955}
3956
3957/*
3958 * Print the block group being relocated
3959 */
3960static void describe_relocation(struct btrfs_block_group *block_group)
3961{
3962 char buf[128] = {'\0'};
3963
3964 btrfs_describe_block_groups(block_group->flags, buf, sizeof(buf));
3965
3966 btrfs_info(block_group->fs_info, "relocating block group %llu flags %s",
3967 block_group->start, buf);
3968}
3969
3970static const char *stage_to_string(enum reloc_stage stage)
3971{
3972 if (stage == MOVE_DATA_EXTENTS)
3973 return "move data extents";
3974 if (stage == UPDATE_DATA_PTRS)
3975 return "update data pointers";
3976 return "unknown";
3977}
3978
3979/*
3980 * function to relocate all extents in a block group.
3981 */
3982int btrfs_relocate_block_group(struct btrfs_fs_info *fs_info, u64 group_start)
3983{
3984 struct btrfs_block_group *bg;
3985 struct btrfs_root *extent_root = btrfs_extent_root(fs_info, group_start);
3986 struct reloc_control *rc;
3987 struct inode *inode;
3988 struct btrfs_path *path;
3989 int ret;
3990 int rw = 0;
3991 int err = 0;
3992
3993 /*
3994 * This only gets set if we had a half-deleted snapshot on mount. We
3995 * cannot allow relocation to start while we're still trying to clean up
3996 * these pending deletions.
3997 */
3998 ret = wait_on_bit(&fs_info->flags, BTRFS_FS_UNFINISHED_DROPS, TASK_INTERRUPTIBLE);
3999 if (ret)
4000 return ret;
4001
4002 /* We may have been woken up by close_ctree, so bail if we're closing. */
4003 if (btrfs_fs_closing(fs_info))
4004 return -EINTR;
4005
4006 bg = btrfs_lookup_block_group(fs_info, group_start);
4007 if (!bg)
4008 return -ENOENT;
4009
4010 /*
4011 * Relocation of a data block group creates ordered extents. Without
4012 * sb_start_write(), we can freeze the filesystem while unfinished
4013 * ordered extents are left. Such ordered extents can cause a deadlock
4014 * e.g. when syncfs() is waiting for their completion but they can't
4015 * finish because they block when joining a transaction, due to the
4016 * fact that the freeze locks are being held in write mode.
4017 */
4018 if (bg->flags & BTRFS_BLOCK_GROUP_DATA)
4019 ASSERT(sb_write_started(fs_info->sb));
4020
4021 if (btrfs_pinned_by_swapfile(fs_info, bg)) {
4022 btrfs_put_block_group(bg);
4023 return -ETXTBSY;
4024 }
4025
4026 rc = alloc_reloc_control(fs_info);
4027 if (!rc) {
4028 btrfs_put_block_group(bg);
4029 return -ENOMEM;
4030 }
4031
4032 ret = reloc_chunk_start(fs_info);
4033 if (ret < 0) {
4034 err = ret;
4035 goto out_put_bg;
4036 }
4037
4038 rc->extent_root = extent_root;
4039 rc->block_group = bg;
4040
4041 ret = btrfs_inc_block_group_ro(rc->block_group, true);
4042 if (ret) {
4043 err = ret;
4044 goto out;
4045 }
4046 rw = 1;
4047
4048 path = btrfs_alloc_path();
4049 if (!path) {
4050 err = -ENOMEM;
4051 goto out;
4052 }
4053
4054 inode = lookup_free_space_inode(rc->block_group, path);
4055 btrfs_free_path(path);
4056
4057 if (!IS_ERR(inode))
4058 ret = delete_block_group_cache(fs_info, rc->block_group, inode, 0);
4059 else
4060 ret = PTR_ERR(inode);
4061
4062 if (ret && ret != -ENOENT) {
4063 err = ret;
4064 goto out;
4065 }
4066
4067 rc->data_inode = create_reloc_inode(fs_info, rc->block_group);
4068 if (IS_ERR(rc->data_inode)) {
4069 err = PTR_ERR(rc->data_inode);
4070 rc->data_inode = NULL;
4071 goto out;
4072 }
4073
4074 describe_relocation(rc->block_group);
4075
4076 btrfs_wait_block_group_reservations(rc->block_group);
4077 btrfs_wait_nocow_writers(rc->block_group);
4078 btrfs_wait_ordered_roots(fs_info, U64_MAX, rc->block_group);
4079
4080 ret = btrfs_zone_finish(rc->block_group);
4081 WARN_ON(ret && ret != -EAGAIN);
4082
4083 while (1) {
4084 enum reloc_stage 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(BTRFS_I(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 ret2;
4176 int ret = 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 goto out;
4192 if (ret > 0) {
4193 if (path->slots[0] == 0)
4194 break;
4195 path->slots[0]--;
4196 }
4197 ret = 0;
4198 leaf = path->nodes[0];
4199 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
4200 btrfs_release_path(path);
4201
4202 if (key.objectid != BTRFS_TREE_RELOC_OBJECTID ||
4203 key.type != BTRFS_ROOT_ITEM_KEY)
4204 break;
4205
4206 reloc_root = btrfs_read_tree_root(fs_info->tree_root, &key);
4207 if (IS_ERR(reloc_root)) {
4208 ret = PTR_ERR(reloc_root);
4209 goto out;
4210 }
4211
4212 set_bit(BTRFS_ROOT_SHAREABLE, &reloc_root->state);
4213 list_add(&reloc_root->root_list, &reloc_roots);
4214
4215 if (btrfs_root_refs(&reloc_root->root_item) > 0) {
4216 fs_root = btrfs_get_fs_root(fs_info,
4217 reloc_root->root_key.offset, false);
4218 if (IS_ERR(fs_root)) {
4219 ret = PTR_ERR(fs_root);
4220 if (ret != -ENOENT)
4221 goto out;
4222 ret = mark_garbage_root(reloc_root);
4223 if (ret < 0)
4224 goto out;
4225 ret = 0;
4226 } else {
4227 btrfs_put_root(fs_root);
4228 }
4229 }
4230
4231 if (key.offset == 0)
4232 break;
4233
4234 key.offset--;
4235 }
4236 btrfs_release_path(path);
4237
4238 if (list_empty(&reloc_roots))
4239 goto out;
4240
4241 rc = alloc_reloc_control(fs_info);
4242 if (!rc) {
4243 ret = -ENOMEM;
4244 goto out;
4245 }
4246
4247 ret = reloc_chunk_start(fs_info);
4248 if (ret < 0)
4249 goto out_end;
4250
4251 rc->extent_root = btrfs_extent_root(fs_info, 0);
4252
4253 set_reloc_control(rc);
4254
4255 trans = btrfs_join_transaction(rc->extent_root);
4256 if (IS_ERR(trans)) {
4257 ret = PTR_ERR(trans);
4258 goto out_unset;
4259 }
4260
4261 rc->merge_reloc_tree = true;
4262
4263 while (!list_empty(&reloc_roots)) {
4264 reloc_root = list_entry(reloc_roots.next,
4265 struct btrfs_root, root_list);
4266 list_del(&reloc_root->root_list);
4267
4268 if (btrfs_root_refs(&reloc_root->root_item) == 0) {
4269 list_add_tail(&reloc_root->root_list,
4270 &rc->reloc_roots);
4271 continue;
4272 }
4273
4274 fs_root = btrfs_get_fs_root(fs_info, reloc_root->root_key.offset,
4275 false);
4276 if (IS_ERR(fs_root)) {
4277 ret = PTR_ERR(fs_root);
4278 list_add_tail(&reloc_root->root_list, &reloc_roots);
4279 btrfs_end_transaction(trans);
4280 goto out_unset;
4281 }
4282
4283 ret = __add_reloc_root(reloc_root);
4284 ASSERT(ret != -EEXIST);
4285 if (ret) {
4286 list_add_tail(&reloc_root->root_list, &reloc_roots);
4287 btrfs_put_root(fs_root);
4288 btrfs_end_transaction(trans);
4289 goto out_unset;
4290 }
4291 fs_root->reloc_root = btrfs_grab_root(reloc_root);
4292 btrfs_put_root(fs_root);
4293 }
4294
4295 ret = btrfs_commit_transaction(trans);
4296 if (ret)
4297 goto out_unset;
4298
4299 merge_reloc_roots(rc);
4300
4301 unset_reloc_control(rc);
4302
4303 trans = btrfs_join_transaction(rc->extent_root);
4304 if (IS_ERR(trans)) {
4305 ret = PTR_ERR(trans);
4306 goto out_clean;
4307 }
4308 ret = btrfs_commit_transaction(trans);
4309out_clean:
4310 ret2 = clean_dirty_subvols(rc);
4311 if (ret2 < 0 && !ret)
4312 ret = ret2;
4313out_unset:
4314 unset_reloc_control(rc);
4315out_end:
4316 reloc_chunk_end(fs_info);
4317 free_reloc_control(rc);
4318out:
4319 free_reloc_roots(&reloc_roots);
4320
4321 btrfs_free_path(path);
4322
4323 if (ret == 0) {
4324 /* cleanup orphan inode in data relocation tree */
4325 fs_root = btrfs_grab_root(fs_info->data_reloc_root);
4326 ASSERT(fs_root);
4327 ret = btrfs_orphan_cleanup(fs_root);
4328 btrfs_put_root(fs_root);
4329 }
4330 return ret;
4331}
4332
4333/*
4334 * helper to add ordered checksum for data relocation.
4335 *
4336 * cloning checksum properly handles the nodatasum extents.
4337 * it also saves CPU time to re-calculate the checksum.
4338 */
4339int btrfs_reloc_clone_csums(struct btrfs_ordered_extent *ordered)
4340{
4341 struct btrfs_inode *inode = ordered->inode;
4342 struct btrfs_fs_info *fs_info = inode->root->fs_info;
4343 u64 disk_bytenr = ordered->file_offset + inode->reloc_block_group_start;
4344 struct btrfs_root *csum_root = btrfs_csum_root(fs_info, disk_bytenr);
4345 LIST_HEAD(list);
4346 int ret;
4347
4348 ret = btrfs_lookup_csums_list(csum_root, disk_bytenr,
4349 disk_bytenr + ordered->num_bytes - 1,
4350 &list, false);
4351 if (ret < 0) {
4352 btrfs_mark_ordered_extent_error(ordered);
4353 return ret;
4354 }
4355
4356 while (!list_empty(&list)) {
4357 struct btrfs_ordered_sum *sums =
4358 list_entry(list.next, struct btrfs_ordered_sum, list);
4359
4360 list_del_init(&sums->list);
4361
4362 /*
4363 * We need to offset the new_bytenr based on where the csum is.
4364 * We need to do this because we will read in entire prealloc
4365 * extents but we may have written to say the middle of the
4366 * prealloc extent, so we need to make sure the csum goes with
4367 * the right disk offset.
4368 *
4369 * We can do this because the data reloc inode refers strictly
4370 * to the on disk bytes, so we don't have to worry about
4371 * disk_len vs real len like with real inodes since it's all
4372 * disk length.
4373 */
4374 sums->logical = ordered->disk_bytenr + sums->logical - disk_bytenr;
4375 btrfs_add_ordered_sum(ordered, sums);
4376 }
4377
4378 return 0;
4379}
4380
4381int btrfs_reloc_cow_block(struct btrfs_trans_handle *trans,
4382 struct btrfs_root *root,
4383 const struct extent_buffer *buf,
4384 struct extent_buffer *cow)
4385{
4386 struct btrfs_fs_info *fs_info = root->fs_info;
4387 struct reloc_control *rc;
4388 struct btrfs_backref_node *node;
4389 int first_cow = 0;
4390 int level;
4391 int ret = 0;
4392
4393 rc = fs_info->reloc_ctl;
4394 if (!rc)
4395 return 0;
4396
4397 BUG_ON(rc->stage == UPDATE_DATA_PTRS && btrfs_is_data_reloc_root(root));
4398
4399 level = btrfs_header_level(buf);
4400 if (btrfs_header_generation(buf) <=
4401 btrfs_root_last_snapshot(&root->root_item))
4402 first_cow = 1;
4403
4404 if (btrfs_root_id(root) == BTRFS_TREE_RELOC_OBJECTID && rc->create_reloc_tree) {
4405 WARN_ON(!first_cow && level == 0);
4406
4407 node = rc->backref_cache.path[level];
4408
4409 /*
4410 * If node->bytenr != buf->start and node->new_bytenr !=
4411 * buf->start then we've got the wrong backref node for what we
4412 * expected to see here and the cache is incorrect.
4413 */
4414 if (unlikely(node->bytenr != buf->start && node->new_bytenr != buf->start)) {
4415 btrfs_err(fs_info,
4416"bytenr %llu was found but our backref cache was expecting %llu or %llu",
4417 buf->start, node->bytenr, node->new_bytenr);
4418 return -EUCLEAN;
4419 }
4420
4421 btrfs_backref_drop_node_buffer(node);
4422 atomic_inc(&cow->refs);
4423 node->eb = cow;
4424 node->new_bytenr = cow->start;
4425
4426 if (!node->pending) {
4427 list_move_tail(&node->list,
4428 &rc->backref_cache.pending[level]);
4429 node->pending = 1;
4430 }
4431
4432 if (first_cow)
4433 mark_block_processed(rc, node);
4434
4435 if (first_cow && level > 0)
4436 rc->nodes_relocated += buf->len;
4437 }
4438
4439 if (level == 0 && first_cow && rc->stage == UPDATE_DATA_PTRS)
4440 ret = replace_file_extents(trans, rc, root, cow);
4441 return ret;
4442}
4443
4444/*
4445 * called before creating snapshot. it calculates metadata reservation
4446 * required for relocating tree blocks in the snapshot
4447 */
4448void btrfs_reloc_pre_snapshot(struct btrfs_pending_snapshot *pending,
4449 u64 *bytes_to_reserve)
4450{
4451 struct btrfs_root *root = pending->root;
4452 struct reloc_control *rc = root->fs_info->reloc_ctl;
4453
4454 if (!rc || !have_reloc_root(root))
4455 return;
4456
4457 if (!rc->merge_reloc_tree)
4458 return;
4459
4460 root = root->reloc_root;
4461 BUG_ON(btrfs_root_refs(&root->root_item) == 0);
4462 /*
4463 * relocation is in the stage of merging trees. the space
4464 * used by merging a reloc tree is twice the size of
4465 * relocated tree nodes in the worst case. half for cowing
4466 * the reloc tree, half for cowing the fs tree. the space
4467 * used by cowing the reloc tree will be freed after the
4468 * tree is dropped. if we create snapshot, cowing the fs
4469 * tree may use more space than it frees. so we need
4470 * reserve extra space.
4471 */
4472 *bytes_to_reserve += rc->nodes_relocated;
4473}
4474
4475/*
4476 * called after snapshot is created. migrate block reservation
4477 * and create reloc root for the newly created snapshot
4478 *
4479 * This is similar to btrfs_init_reloc_root(), we come out of here with two
4480 * references held on the reloc_root, one for root->reloc_root and one for
4481 * rc->reloc_roots.
4482 */
4483int btrfs_reloc_post_snapshot(struct btrfs_trans_handle *trans,
4484 struct btrfs_pending_snapshot *pending)
4485{
4486 struct btrfs_root *root = pending->root;
4487 struct btrfs_root *reloc_root;
4488 struct btrfs_root *new_root;
4489 struct reloc_control *rc = root->fs_info->reloc_ctl;
4490 int ret;
4491
4492 if (!rc || !have_reloc_root(root))
4493 return 0;
4494
4495 rc = root->fs_info->reloc_ctl;
4496 rc->merging_rsv_size += rc->nodes_relocated;
4497
4498 if (rc->merge_reloc_tree) {
4499 ret = btrfs_block_rsv_migrate(&pending->block_rsv,
4500 rc->block_rsv,
4501 rc->nodes_relocated, true);
4502 if (ret)
4503 return ret;
4504 }
4505
4506 new_root = pending->snap;
4507 reloc_root = create_reloc_root(trans, root->reloc_root, btrfs_root_id(new_root));
4508 if (IS_ERR(reloc_root))
4509 return PTR_ERR(reloc_root);
4510
4511 ret = __add_reloc_root(reloc_root);
4512 ASSERT(ret != -EEXIST);
4513 if (ret) {
4514 /* Pairs with create_reloc_root */
4515 btrfs_put_root(reloc_root);
4516 return ret;
4517 }
4518 new_root->reloc_root = btrfs_grab_root(reloc_root);
4519
4520 if (rc->create_reloc_tree)
4521 ret = clone_backref_node(trans, rc, root, reloc_root);
4522 return ret;
4523}
4524
4525/*
4526 * Get the current bytenr for the block group which is being relocated.
4527 *
4528 * Return U64_MAX if no running relocation.
4529 */
4530u64 btrfs_get_reloc_bg_bytenr(const struct btrfs_fs_info *fs_info)
4531{
4532 u64 logical = U64_MAX;
4533
4534 lockdep_assert_held(&fs_info->reloc_mutex);
4535
4536 if (fs_info->reloc_ctl && fs_info->reloc_ctl->block_group)
4537 logical = fs_info->reloc_ctl->block_group->start;
4538 return logical;
4539}
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 u8 level;
115 bool key_ready;
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 u64 owning_root;
126};
127
128/* Stages of data relocation. */
129enum reloc_stage {
130 MOVE_DATA_EXTENTS,
131 UPDATE_DATA_PTRS
132};
133
134struct reloc_control {
135 /* block group to relocate */
136 struct btrfs_block_group *block_group;
137 /* extent tree */
138 struct btrfs_root *extent_root;
139 /* inode for moving data */
140 struct inode *data_inode;
141
142 struct btrfs_block_rsv *block_rsv;
143
144 struct btrfs_backref_cache backref_cache;
145
146 struct file_extent_cluster cluster;
147 /* tree blocks have been processed */
148 struct extent_io_tree processed_blocks;
149 /* map start of tree root to corresponding reloc tree */
150 struct mapping_tree reloc_root_tree;
151 /* list of reloc trees */
152 struct list_head reloc_roots;
153 /* list of subvolume trees that get relocated */
154 struct list_head dirty_subvol_roots;
155 /* size of metadata reservation for merging reloc trees */
156 u64 merging_rsv_size;
157 /* size of relocated tree nodes */
158 u64 nodes_relocated;
159 /* reserved size for block group relocation*/
160 u64 reserved_bytes;
161
162 u64 search_start;
163 u64 extents_found;
164
165 enum reloc_stage stage;
166 bool create_reloc_tree;
167 bool merge_reloc_tree;
168 bool found_file_extent;
169};
170
171static void mark_block_processed(struct reloc_control *rc,
172 struct btrfs_backref_node *node)
173{
174 u32 blocksize;
175
176 if (node->level == 0 ||
177 in_range(node->bytenr, rc->block_group->start,
178 rc->block_group->length)) {
179 blocksize = rc->extent_root->fs_info->nodesize;
180 set_extent_bit(&rc->processed_blocks, node->bytenr,
181 node->bytenr + blocksize - 1, EXTENT_DIRTY, NULL);
182 }
183 node->processed = 1;
184}
185
186/*
187 * walk up backref nodes until reach node presents tree root
188 */
189static struct btrfs_backref_node *walk_up_backref(
190 struct btrfs_backref_node *node,
191 struct btrfs_backref_edge *edges[], int *index)
192{
193 struct btrfs_backref_edge *edge;
194 int idx = *index;
195
196 while (!list_empty(&node->upper)) {
197 edge = list_entry(node->upper.next,
198 struct btrfs_backref_edge, list[LOWER]);
199 edges[idx++] = edge;
200 node = edge->node[UPPER];
201 }
202 BUG_ON(node->detached);
203 *index = idx;
204 return node;
205}
206
207/*
208 * walk down backref nodes to find start of next reference path
209 */
210static struct btrfs_backref_node *walk_down_backref(
211 struct btrfs_backref_edge *edges[], int *index)
212{
213 struct btrfs_backref_edge *edge;
214 struct btrfs_backref_node *lower;
215 int idx = *index;
216
217 while (idx > 0) {
218 edge = edges[idx - 1];
219 lower = edge->node[LOWER];
220 if (list_is_last(&edge->list[LOWER], &lower->upper)) {
221 idx--;
222 continue;
223 }
224 edge = list_entry(edge->list[LOWER].next,
225 struct btrfs_backref_edge, list[LOWER]);
226 edges[idx - 1] = edge;
227 *index = idx;
228 return edge->node[UPPER];
229 }
230 *index = 0;
231 return NULL;
232}
233
234static void update_backref_node(struct btrfs_backref_cache *cache,
235 struct btrfs_backref_node *node, u64 bytenr)
236{
237 struct rb_node *rb_node;
238 rb_erase(&node->rb_node, &cache->rb_root);
239 node->bytenr = bytenr;
240 rb_node = rb_simple_insert(&cache->rb_root, node->bytenr, &node->rb_node);
241 if (rb_node)
242 btrfs_backref_panic(cache->fs_info, bytenr, -EEXIST);
243}
244
245/*
246 * update backref cache after a transaction commit
247 */
248static int update_backref_cache(struct btrfs_trans_handle *trans,
249 struct btrfs_backref_cache *cache)
250{
251 struct btrfs_backref_node *node;
252 int level = 0;
253
254 if (cache->last_trans == 0) {
255 cache->last_trans = trans->transid;
256 return 0;
257 }
258
259 if (cache->last_trans == trans->transid)
260 return 0;
261
262 /*
263 * detached nodes are used to avoid unnecessary backref
264 * lookup. transaction commit changes the extent tree.
265 * so the detached nodes are no longer useful.
266 */
267 while (!list_empty(&cache->detached)) {
268 node = list_entry(cache->detached.next,
269 struct btrfs_backref_node, list);
270 btrfs_backref_cleanup_node(cache, node);
271 }
272
273 while (!list_empty(&cache->changed)) {
274 node = list_entry(cache->changed.next,
275 struct btrfs_backref_node, list);
276 list_del_init(&node->list);
277 BUG_ON(node->pending);
278 update_backref_node(cache, node, node->new_bytenr);
279 }
280
281 /*
282 * some nodes can be left in the pending list if there were
283 * errors during processing the pending nodes.
284 */
285 for (level = 0; level < BTRFS_MAX_LEVEL; level++) {
286 list_for_each_entry(node, &cache->pending[level], list) {
287 BUG_ON(!node->pending);
288 if (node->bytenr == node->new_bytenr)
289 continue;
290 update_backref_node(cache, node, node->new_bytenr);
291 }
292 }
293
294 cache->last_trans = 0;
295 return 1;
296}
297
298static bool reloc_root_is_dead(const struct btrfs_root *root)
299{
300 /*
301 * Pair with set_bit/clear_bit in clean_dirty_subvols and
302 * btrfs_update_reloc_root. We need to see the updated bit before
303 * trying to access reloc_root
304 */
305 smp_rmb();
306 if (test_bit(BTRFS_ROOT_DEAD_RELOC_TREE, &root->state))
307 return true;
308 return false;
309}
310
311/*
312 * Check if this subvolume tree has valid reloc tree.
313 *
314 * Reloc tree after swap is considered dead, thus not considered as valid.
315 * This is enough for most callers, as they don't distinguish dead reloc root
316 * from no reloc root. But btrfs_should_ignore_reloc_root() below is a
317 * special case.
318 */
319static bool have_reloc_root(const struct btrfs_root *root)
320{
321 if (reloc_root_is_dead(root))
322 return false;
323 if (!root->reloc_root)
324 return false;
325 return true;
326}
327
328bool btrfs_should_ignore_reloc_root(const struct btrfs_root *root)
329{
330 struct btrfs_root *reloc_root;
331
332 if (!test_bit(BTRFS_ROOT_SHAREABLE, &root->state))
333 return false;
334
335 /* This root has been merged with its reloc tree, we can ignore it */
336 if (reloc_root_is_dead(root))
337 return true;
338
339 reloc_root = root->reloc_root;
340 if (!reloc_root)
341 return false;
342
343 if (btrfs_header_generation(reloc_root->commit_root) ==
344 root->fs_info->running_transaction->transid)
345 return false;
346 /*
347 * If there is reloc tree and it was created in previous transaction
348 * backref lookup can find the reloc tree, so backref node for the fs
349 * tree root is useless for relocation.
350 */
351 return true;
352}
353
354/*
355 * find reloc tree by address of tree root
356 */
357struct btrfs_root *find_reloc_root(struct btrfs_fs_info *fs_info, u64 bytenr)
358{
359 struct reloc_control *rc = fs_info->reloc_ctl;
360 struct rb_node *rb_node;
361 struct mapping_node *node;
362 struct btrfs_root *root = NULL;
363
364 ASSERT(rc);
365 spin_lock(&rc->reloc_root_tree.lock);
366 rb_node = rb_simple_search(&rc->reloc_root_tree.rb_root, bytenr);
367 if (rb_node) {
368 node = rb_entry(rb_node, struct mapping_node, rb_node);
369 root = node->data;
370 }
371 spin_unlock(&rc->reloc_root_tree.lock);
372 return btrfs_grab_root(root);
373}
374
375/*
376 * For useless nodes, do two major clean ups:
377 *
378 * - Cleanup the children edges and nodes
379 * If child node is also orphan (no parent) during cleanup, then the child
380 * node will also be cleaned up.
381 *
382 * - Freeing up leaves (level 0), keeps nodes detached
383 * For nodes, the node is still cached as "detached"
384 *
385 * Return false if @node is not in the @useless_nodes list.
386 * Return true if @node is in the @useless_nodes list.
387 */
388static bool handle_useless_nodes(struct reloc_control *rc,
389 struct btrfs_backref_node *node)
390{
391 struct btrfs_backref_cache *cache = &rc->backref_cache;
392 struct list_head *useless_node = &cache->useless_node;
393 bool ret = false;
394
395 while (!list_empty(useless_node)) {
396 struct btrfs_backref_node *cur;
397
398 cur = list_first_entry(useless_node, struct btrfs_backref_node,
399 list);
400 list_del_init(&cur->list);
401
402 /* Only tree root nodes can be added to @useless_nodes */
403 ASSERT(list_empty(&cur->upper));
404
405 if (cur == node)
406 ret = true;
407
408 /* The node is the lowest node */
409 if (cur->lowest) {
410 list_del_init(&cur->lower);
411 cur->lowest = 0;
412 }
413
414 /* Cleanup the lower edges */
415 while (!list_empty(&cur->lower)) {
416 struct btrfs_backref_edge *edge;
417 struct btrfs_backref_node *lower;
418
419 edge = list_entry(cur->lower.next,
420 struct btrfs_backref_edge, list[UPPER]);
421 list_del(&edge->list[UPPER]);
422 list_del(&edge->list[LOWER]);
423 lower = edge->node[LOWER];
424 btrfs_backref_free_edge(cache, edge);
425
426 /* Child node is also orphan, queue for cleanup */
427 if (list_empty(&lower->upper))
428 list_add(&lower->list, useless_node);
429 }
430 /* Mark this block processed for relocation */
431 mark_block_processed(rc, cur);
432
433 /*
434 * Backref nodes for tree leaves are deleted from the cache.
435 * Backref nodes for upper level tree blocks are left in the
436 * cache to avoid unnecessary backref lookup.
437 */
438 if (cur->level > 0) {
439 list_add(&cur->list, &cache->detached);
440 cur->detached = 1;
441 } else {
442 rb_erase(&cur->rb_node, &cache->rb_root);
443 btrfs_backref_free_node(cache, cur);
444 }
445 }
446 return ret;
447}
448
449/*
450 * Build backref tree for a given tree block. Root of the backref tree
451 * corresponds the tree block, leaves of the backref tree correspond roots of
452 * b-trees that reference the tree block.
453 *
454 * The basic idea of this function is check backrefs of a given block to find
455 * upper level blocks that reference the block, and then check backrefs of
456 * these upper level blocks recursively. The recursion stops when tree root is
457 * reached or backrefs for the block is cached.
458 *
459 * NOTE: if we find that backrefs for a block are cached, we know backrefs for
460 * all upper level blocks that directly/indirectly reference the block are also
461 * cached.
462 */
463static noinline_for_stack struct btrfs_backref_node *build_backref_tree(
464 struct btrfs_trans_handle *trans,
465 struct reloc_control *rc, struct btrfs_key *node_key,
466 int level, u64 bytenr)
467{
468 struct btrfs_backref_iter *iter;
469 struct btrfs_backref_cache *cache = &rc->backref_cache;
470 /* For searching parent of TREE_BLOCK_REF */
471 struct btrfs_path *path;
472 struct btrfs_backref_node *cur;
473 struct btrfs_backref_node *node = NULL;
474 struct btrfs_backref_edge *edge;
475 int ret;
476 int err = 0;
477
478 iter = btrfs_backref_iter_alloc(rc->extent_root->fs_info);
479 if (!iter)
480 return ERR_PTR(-ENOMEM);
481 path = btrfs_alloc_path();
482 if (!path) {
483 err = -ENOMEM;
484 goto out;
485 }
486
487 node = btrfs_backref_alloc_node(cache, bytenr, level);
488 if (!node) {
489 err = -ENOMEM;
490 goto out;
491 }
492
493 node->lowest = 1;
494 cur = node;
495
496 /* Breadth-first search to build backref cache */
497 do {
498 ret = btrfs_backref_add_tree_node(trans, cache, path, iter,
499 node_key, cur);
500 if (ret < 0) {
501 err = ret;
502 goto out;
503 }
504 edge = list_first_entry_or_null(&cache->pending_edge,
505 struct btrfs_backref_edge, list[UPPER]);
506 /*
507 * The pending list isn't empty, take the first block to
508 * process
509 */
510 if (edge) {
511 list_del_init(&edge->list[UPPER]);
512 cur = edge->node[UPPER];
513 }
514 } while (edge);
515
516 /* Finish the upper linkage of newly added edges/nodes */
517 ret = btrfs_backref_finish_upper_links(cache, node);
518 if (ret < 0) {
519 err = ret;
520 goto out;
521 }
522
523 if (handle_useless_nodes(rc, node))
524 node = NULL;
525out:
526 btrfs_backref_iter_free(iter);
527 btrfs_free_path(path);
528 if (err) {
529 btrfs_backref_error_cleanup(cache, node);
530 return ERR_PTR(err);
531 }
532 ASSERT(!node || !node->detached);
533 ASSERT(list_empty(&cache->useless_node) &&
534 list_empty(&cache->pending_edge));
535 return node;
536}
537
538/*
539 * helper to add backref node for the newly created snapshot.
540 * the backref node is created by cloning backref node that
541 * corresponds to root of source tree
542 */
543static int clone_backref_node(struct btrfs_trans_handle *trans,
544 struct reloc_control *rc,
545 const struct btrfs_root *src,
546 struct btrfs_root *dest)
547{
548 struct btrfs_root *reloc_root = src->reloc_root;
549 struct btrfs_backref_cache *cache = &rc->backref_cache;
550 struct btrfs_backref_node *node = NULL;
551 struct btrfs_backref_node *new_node;
552 struct btrfs_backref_edge *edge;
553 struct btrfs_backref_edge *new_edge;
554 struct rb_node *rb_node;
555
556 if (cache->last_trans > 0)
557 update_backref_cache(trans, cache);
558
559 rb_node = rb_simple_search(&cache->rb_root, src->commit_root->start);
560 if (rb_node) {
561 node = rb_entry(rb_node, struct btrfs_backref_node, rb_node);
562 if (node->detached)
563 node = NULL;
564 else
565 BUG_ON(node->new_bytenr != reloc_root->node->start);
566 }
567
568 if (!node) {
569 rb_node = rb_simple_search(&cache->rb_root,
570 reloc_root->commit_root->start);
571 if (rb_node) {
572 node = rb_entry(rb_node, struct btrfs_backref_node,
573 rb_node);
574 BUG_ON(node->detached);
575 }
576 }
577
578 if (!node)
579 return 0;
580
581 new_node = btrfs_backref_alloc_node(cache, dest->node->start,
582 node->level);
583 if (!new_node)
584 return -ENOMEM;
585
586 new_node->lowest = node->lowest;
587 new_node->checked = 1;
588 new_node->root = btrfs_grab_root(dest);
589 ASSERT(new_node->root);
590
591 if (!node->lowest) {
592 list_for_each_entry(edge, &node->lower, list[UPPER]) {
593 new_edge = btrfs_backref_alloc_edge(cache);
594 if (!new_edge)
595 goto fail;
596
597 btrfs_backref_link_edge(new_edge, edge->node[LOWER],
598 new_node, LINK_UPPER);
599 }
600 } else {
601 list_add_tail(&new_node->lower, &cache->leaves);
602 }
603
604 rb_node = rb_simple_insert(&cache->rb_root, new_node->bytenr,
605 &new_node->rb_node);
606 if (rb_node)
607 btrfs_backref_panic(trans->fs_info, new_node->bytenr, -EEXIST);
608
609 if (!new_node->lowest) {
610 list_for_each_entry(new_edge, &new_node->lower, list[UPPER]) {
611 list_add_tail(&new_edge->list[LOWER],
612 &new_edge->node[LOWER]->upper);
613 }
614 }
615 return 0;
616fail:
617 while (!list_empty(&new_node->lower)) {
618 new_edge = list_entry(new_node->lower.next,
619 struct btrfs_backref_edge, list[UPPER]);
620 list_del(&new_edge->list[UPPER]);
621 btrfs_backref_free_edge(cache, new_edge);
622 }
623 btrfs_backref_free_node(cache, new_node);
624 return -ENOMEM;
625}
626
627/*
628 * helper to add 'address of tree root -> reloc tree' mapping
629 */
630static int __add_reloc_root(struct btrfs_root *root)
631{
632 struct btrfs_fs_info *fs_info = root->fs_info;
633 struct rb_node *rb_node;
634 struct mapping_node *node;
635 struct reloc_control *rc = fs_info->reloc_ctl;
636
637 node = kmalloc(sizeof(*node), GFP_NOFS);
638 if (!node)
639 return -ENOMEM;
640
641 node->bytenr = root->commit_root->start;
642 node->data = root;
643
644 spin_lock(&rc->reloc_root_tree.lock);
645 rb_node = rb_simple_insert(&rc->reloc_root_tree.rb_root,
646 node->bytenr, &node->rb_node);
647 spin_unlock(&rc->reloc_root_tree.lock);
648 if (rb_node) {
649 btrfs_err(fs_info,
650 "Duplicate root found for start=%llu while inserting into relocation tree",
651 node->bytenr);
652 return -EEXIST;
653 }
654
655 list_add_tail(&root->root_list, &rc->reloc_roots);
656 return 0;
657}
658
659/*
660 * helper to delete the 'address of tree root -> reloc tree'
661 * mapping
662 */
663static void __del_reloc_root(struct btrfs_root *root)
664{
665 struct btrfs_fs_info *fs_info = root->fs_info;
666 struct rb_node *rb_node;
667 struct mapping_node *node = NULL;
668 struct reloc_control *rc = fs_info->reloc_ctl;
669 bool put_ref = false;
670
671 if (rc && root->node) {
672 spin_lock(&rc->reloc_root_tree.lock);
673 rb_node = rb_simple_search(&rc->reloc_root_tree.rb_root,
674 root->commit_root->start);
675 if (rb_node) {
676 node = rb_entry(rb_node, struct mapping_node, rb_node);
677 rb_erase(&node->rb_node, &rc->reloc_root_tree.rb_root);
678 RB_CLEAR_NODE(&node->rb_node);
679 }
680 spin_unlock(&rc->reloc_root_tree.lock);
681 ASSERT(!node || (struct btrfs_root *)node->data == root);
682 }
683
684 /*
685 * We only put the reloc root here if it's on the list. There's a lot
686 * of places where the pattern is to splice the rc->reloc_roots, process
687 * the reloc roots, and then add the reloc root back onto
688 * rc->reloc_roots. If we call __del_reloc_root while it's off of the
689 * list we don't want the reference being dropped, because the guy
690 * messing with the list is in charge of the reference.
691 */
692 spin_lock(&fs_info->trans_lock);
693 if (!list_empty(&root->root_list)) {
694 put_ref = true;
695 list_del_init(&root->root_list);
696 }
697 spin_unlock(&fs_info->trans_lock);
698 if (put_ref)
699 btrfs_put_root(root);
700 kfree(node);
701}
702
703/*
704 * helper to update the 'address of tree root -> reloc tree'
705 * mapping
706 */
707static int __update_reloc_root(struct btrfs_root *root)
708{
709 struct btrfs_fs_info *fs_info = root->fs_info;
710 struct rb_node *rb_node;
711 struct mapping_node *node = NULL;
712 struct reloc_control *rc = fs_info->reloc_ctl;
713
714 spin_lock(&rc->reloc_root_tree.lock);
715 rb_node = rb_simple_search(&rc->reloc_root_tree.rb_root,
716 root->commit_root->start);
717 if (rb_node) {
718 node = rb_entry(rb_node, struct mapping_node, rb_node);
719 rb_erase(&node->rb_node, &rc->reloc_root_tree.rb_root);
720 }
721 spin_unlock(&rc->reloc_root_tree.lock);
722
723 if (!node)
724 return 0;
725 BUG_ON((struct btrfs_root *)node->data != root);
726
727 spin_lock(&rc->reloc_root_tree.lock);
728 node->bytenr = root->node->start;
729 rb_node = rb_simple_insert(&rc->reloc_root_tree.rb_root,
730 node->bytenr, &node->rb_node);
731 spin_unlock(&rc->reloc_root_tree.lock);
732 if (rb_node)
733 btrfs_backref_panic(fs_info, node->bytenr, -EEXIST);
734 return 0;
735}
736
737static struct btrfs_root *create_reloc_root(struct btrfs_trans_handle *trans,
738 struct btrfs_root *root, u64 objectid)
739{
740 struct btrfs_fs_info *fs_info = root->fs_info;
741 struct btrfs_root *reloc_root;
742 struct extent_buffer *eb;
743 struct btrfs_root_item *root_item;
744 struct btrfs_key root_key;
745 int ret = 0;
746 bool must_abort = false;
747
748 root_item = kmalloc(sizeof(*root_item), GFP_NOFS);
749 if (!root_item)
750 return ERR_PTR(-ENOMEM);
751
752 root_key.objectid = BTRFS_TREE_RELOC_OBJECTID;
753 root_key.type = BTRFS_ROOT_ITEM_KEY;
754 root_key.offset = objectid;
755
756 if (root->root_key.objectid == objectid) {
757 u64 commit_root_gen;
758
759 /* called by btrfs_init_reloc_root */
760 ret = btrfs_copy_root(trans, root, root->commit_root, &eb,
761 BTRFS_TREE_RELOC_OBJECTID);
762 if (ret)
763 goto fail;
764
765 /*
766 * Set the last_snapshot field to the generation of the commit
767 * root - like this ctree.c:btrfs_block_can_be_shared() behaves
768 * correctly (returns true) when the relocation root is created
769 * either inside the critical section of a transaction commit
770 * (through transaction.c:qgroup_account_snapshot()) and when
771 * it's created before the transaction commit is started.
772 */
773 commit_root_gen = btrfs_header_generation(root->commit_root);
774 btrfs_set_root_last_snapshot(&root->root_item, commit_root_gen);
775 } else {
776 /*
777 * called by btrfs_reloc_post_snapshot_hook.
778 * the source tree is a reloc tree, all tree blocks
779 * modified after it was created have RELOC flag
780 * set in their headers. so it's OK to not update
781 * the 'last_snapshot'.
782 */
783 ret = btrfs_copy_root(trans, root, root->node, &eb,
784 BTRFS_TREE_RELOC_OBJECTID);
785 if (ret)
786 goto fail;
787 }
788
789 /*
790 * We have changed references at this point, we must abort the
791 * transaction if anything fails.
792 */
793 must_abort = true;
794
795 memcpy(root_item, &root->root_item, sizeof(*root_item));
796 btrfs_set_root_bytenr(root_item, eb->start);
797 btrfs_set_root_level(root_item, btrfs_header_level(eb));
798 btrfs_set_root_generation(root_item, trans->transid);
799
800 if (root->root_key.objectid == objectid) {
801 btrfs_set_root_refs(root_item, 0);
802 memset(&root_item->drop_progress, 0,
803 sizeof(struct btrfs_disk_key));
804 btrfs_set_root_drop_level(root_item, 0);
805 }
806
807 btrfs_tree_unlock(eb);
808 free_extent_buffer(eb);
809
810 ret = btrfs_insert_root(trans, fs_info->tree_root,
811 &root_key, root_item);
812 if (ret)
813 goto fail;
814
815 kfree(root_item);
816
817 reloc_root = btrfs_read_tree_root(fs_info->tree_root, &root_key);
818 if (IS_ERR(reloc_root)) {
819 ret = PTR_ERR(reloc_root);
820 goto abort;
821 }
822 set_bit(BTRFS_ROOT_SHAREABLE, &reloc_root->state);
823 reloc_root->last_trans = trans->transid;
824 return reloc_root;
825fail:
826 kfree(root_item);
827abort:
828 if (must_abort)
829 btrfs_abort_transaction(trans, ret);
830 return ERR_PTR(ret);
831}
832
833/*
834 * create reloc tree for a given fs tree. reloc tree is just a
835 * snapshot of the fs tree with special root objectid.
836 *
837 * The reloc_root comes out of here with two references, one for
838 * root->reloc_root, and another for being on the rc->reloc_roots list.
839 */
840int btrfs_init_reloc_root(struct btrfs_trans_handle *trans,
841 struct btrfs_root *root)
842{
843 struct btrfs_fs_info *fs_info = root->fs_info;
844 struct btrfs_root *reloc_root;
845 struct reloc_control *rc = fs_info->reloc_ctl;
846 struct btrfs_block_rsv *rsv;
847 int clear_rsv = 0;
848 int ret;
849
850 if (!rc)
851 return 0;
852
853 /*
854 * The subvolume has reloc tree but the swap is finished, no need to
855 * create/update the dead reloc tree
856 */
857 if (reloc_root_is_dead(root))
858 return 0;
859
860 /*
861 * This is subtle but important. We do not do
862 * record_root_in_transaction for reloc roots, instead we record their
863 * corresponding fs root, and then here we update the last trans for the
864 * reloc root. This means that we have to do this for the entire life
865 * of the reloc root, regardless of which stage of the relocation we are
866 * in.
867 */
868 if (root->reloc_root) {
869 reloc_root = root->reloc_root;
870 reloc_root->last_trans = trans->transid;
871 return 0;
872 }
873
874 /*
875 * We are merging reloc roots, we do not need new reloc trees. Also
876 * reloc trees never need their own reloc tree.
877 */
878 if (!rc->create_reloc_tree ||
879 root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID)
880 return 0;
881
882 if (!trans->reloc_reserved) {
883 rsv = trans->block_rsv;
884 trans->block_rsv = rc->block_rsv;
885 clear_rsv = 1;
886 }
887 reloc_root = create_reloc_root(trans, root, root->root_key.objectid);
888 if (clear_rsv)
889 trans->block_rsv = rsv;
890 if (IS_ERR(reloc_root))
891 return PTR_ERR(reloc_root);
892
893 ret = __add_reloc_root(reloc_root);
894 ASSERT(ret != -EEXIST);
895 if (ret) {
896 /* Pairs with create_reloc_root */
897 btrfs_put_root(reloc_root);
898 return ret;
899 }
900 root->reloc_root = btrfs_grab_root(reloc_root);
901 return 0;
902}
903
904/*
905 * update root item of reloc tree
906 */
907int btrfs_update_reloc_root(struct btrfs_trans_handle *trans,
908 struct btrfs_root *root)
909{
910 struct btrfs_fs_info *fs_info = root->fs_info;
911 struct btrfs_root *reloc_root;
912 struct btrfs_root_item *root_item;
913 int ret;
914
915 if (!have_reloc_root(root))
916 return 0;
917
918 reloc_root = root->reloc_root;
919 root_item = &reloc_root->root_item;
920
921 /*
922 * We are probably ok here, but __del_reloc_root() will drop its ref of
923 * the root. We have the ref for root->reloc_root, but just in case
924 * hold it while we update the reloc root.
925 */
926 btrfs_grab_root(reloc_root);
927
928 /* root->reloc_root will stay until current relocation finished */
929 if (fs_info->reloc_ctl->merge_reloc_tree &&
930 btrfs_root_refs(root_item) == 0) {
931 set_bit(BTRFS_ROOT_DEAD_RELOC_TREE, &root->state);
932 /*
933 * Mark the tree as dead before we change reloc_root so
934 * have_reloc_root will not touch it from now on.
935 */
936 smp_wmb();
937 __del_reloc_root(reloc_root);
938 }
939
940 if (reloc_root->commit_root != reloc_root->node) {
941 __update_reloc_root(reloc_root);
942 btrfs_set_root_node(root_item, reloc_root->node);
943 free_extent_buffer(reloc_root->commit_root);
944 reloc_root->commit_root = btrfs_root_node(reloc_root);
945 }
946
947 ret = btrfs_update_root(trans, fs_info->tree_root,
948 &reloc_root->root_key, root_item);
949 btrfs_put_root(reloc_root);
950 return ret;
951}
952
953/*
954 * helper to find first cached inode with inode number >= objectid
955 * in a subvolume
956 */
957static struct inode *find_next_inode(struct btrfs_root *root, u64 objectid)
958{
959 struct rb_node *node;
960 struct rb_node *prev;
961 struct btrfs_inode *entry;
962 struct inode *inode;
963
964 spin_lock(&root->inode_lock);
965again:
966 node = root->inode_tree.rb_node;
967 prev = NULL;
968 while (node) {
969 prev = node;
970 entry = rb_entry(node, struct btrfs_inode, rb_node);
971
972 if (objectid < btrfs_ino(entry))
973 node = node->rb_left;
974 else if (objectid > btrfs_ino(entry))
975 node = node->rb_right;
976 else
977 break;
978 }
979 if (!node) {
980 while (prev) {
981 entry = rb_entry(prev, struct btrfs_inode, rb_node);
982 if (objectid <= btrfs_ino(entry)) {
983 node = prev;
984 break;
985 }
986 prev = rb_next(prev);
987 }
988 }
989 while (node) {
990 entry = rb_entry(node, struct btrfs_inode, rb_node);
991 inode = igrab(&entry->vfs_inode);
992 if (inode) {
993 spin_unlock(&root->inode_lock);
994 return inode;
995 }
996
997 objectid = btrfs_ino(entry) + 1;
998 if (cond_resched_lock(&root->inode_lock))
999 goto again;
1000
1001 node = rb_next(node);
1002 }
1003 spin_unlock(&root->inode_lock);
1004 return NULL;
1005}
1006
1007/*
1008 * get new location of data
1009 */
1010static int get_new_location(struct inode *reloc_inode, u64 *new_bytenr,
1011 u64 bytenr, u64 num_bytes)
1012{
1013 struct btrfs_root *root = BTRFS_I(reloc_inode)->root;
1014 struct btrfs_path *path;
1015 struct btrfs_file_extent_item *fi;
1016 struct extent_buffer *leaf;
1017 int ret;
1018
1019 path = btrfs_alloc_path();
1020 if (!path)
1021 return -ENOMEM;
1022
1023 bytenr -= BTRFS_I(reloc_inode)->index_cnt;
1024 ret = btrfs_lookup_file_extent(NULL, root, path,
1025 btrfs_ino(BTRFS_I(reloc_inode)), bytenr, 0);
1026 if (ret < 0)
1027 goto out;
1028 if (ret > 0) {
1029 ret = -ENOENT;
1030 goto out;
1031 }
1032
1033 leaf = path->nodes[0];
1034 fi = btrfs_item_ptr(leaf, path->slots[0],
1035 struct btrfs_file_extent_item);
1036
1037 BUG_ON(btrfs_file_extent_offset(leaf, fi) ||
1038 btrfs_file_extent_compression(leaf, fi) ||
1039 btrfs_file_extent_encryption(leaf, fi) ||
1040 btrfs_file_extent_other_encoding(leaf, fi));
1041
1042 if (num_bytes != btrfs_file_extent_disk_num_bytes(leaf, fi)) {
1043 ret = -EINVAL;
1044 goto out;
1045 }
1046
1047 *new_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
1048 ret = 0;
1049out:
1050 btrfs_free_path(path);
1051 return ret;
1052}
1053
1054/*
1055 * update file extent items in the tree leaf to point to
1056 * the new locations.
1057 */
1058static noinline_for_stack
1059int replace_file_extents(struct btrfs_trans_handle *trans,
1060 struct reloc_control *rc,
1061 struct btrfs_root *root,
1062 struct extent_buffer *leaf)
1063{
1064 struct btrfs_fs_info *fs_info = root->fs_info;
1065 struct btrfs_key key;
1066 struct btrfs_file_extent_item *fi;
1067 struct inode *inode = NULL;
1068 u64 parent;
1069 u64 bytenr;
1070 u64 new_bytenr = 0;
1071 u64 num_bytes;
1072 u64 end;
1073 u32 nritems;
1074 u32 i;
1075 int ret = 0;
1076 int first = 1;
1077 int dirty = 0;
1078
1079 if (rc->stage != UPDATE_DATA_PTRS)
1080 return 0;
1081
1082 /* reloc trees always use full backref */
1083 if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID)
1084 parent = leaf->start;
1085 else
1086 parent = 0;
1087
1088 nritems = btrfs_header_nritems(leaf);
1089 for (i = 0; i < nritems; i++) {
1090 struct btrfs_ref ref = { 0 };
1091
1092 cond_resched();
1093 btrfs_item_key_to_cpu(leaf, &key, i);
1094 if (key.type != BTRFS_EXTENT_DATA_KEY)
1095 continue;
1096 fi = btrfs_item_ptr(leaf, i, struct btrfs_file_extent_item);
1097 if (btrfs_file_extent_type(leaf, fi) ==
1098 BTRFS_FILE_EXTENT_INLINE)
1099 continue;
1100 bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
1101 num_bytes = btrfs_file_extent_disk_num_bytes(leaf, fi);
1102 if (bytenr == 0)
1103 continue;
1104 if (!in_range(bytenr, rc->block_group->start,
1105 rc->block_group->length))
1106 continue;
1107
1108 /*
1109 * if we are modifying block in fs tree, wait for read_folio
1110 * to complete and drop the extent cache
1111 */
1112 if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) {
1113 if (first) {
1114 inode = find_next_inode(root, key.objectid);
1115 first = 0;
1116 } else if (inode && btrfs_ino(BTRFS_I(inode)) < key.objectid) {
1117 btrfs_add_delayed_iput(BTRFS_I(inode));
1118 inode = find_next_inode(root, key.objectid);
1119 }
1120 if (inode && btrfs_ino(BTRFS_I(inode)) == key.objectid) {
1121 struct extent_state *cached_state = NULL;
1122
1123 end = key.offset +
1124 btrfs_file_extent_num_bytes(leaf, fi);
1125 WARN_ON(!IS_ALIGNED(key.offset,
1126 fs_info->sectorsize));
1127 WARN_ON(!IS_ALIGNED(end, fs_info->sectorsize));
1128 end--;
1129 ret = try_lock_extent(&BTRFS_I(inode)->io_tree,
1130 key.offset, end,
1131 &cached_state);
1132 if (!ret)
1133 continue;
1134
1135 btrfs_drop_extent_map_range(BTRFS_I(inode),
1136 key.offset, end, true);
1137 unlock_extent(&BTRFS_I(inode)->io_tree,
1138 key.offset, end, &cached_state);
1139 }
1140 }
1141
1142 ret = get_new_location(rc->data_inode, &new_bytenr,
1143 bytenr, num_bytes);
1144 if (ret) {
1145 /*
1146 * Don't have to abort since we've not changed anything
1147 * in the file extent yet.
1148 */
1149 break;
1150 }
1151
1152 btrfs_set_file_extent_disk_bytenr(leaf, fi, new_bytenr);
1153 dirty = 1;
1154
1155 key.offset -= btrfs_file_extent_offset(leaf, fi);
1156 btrfs_init_generic_ref(&ref, BTRFS_ADD_DELAYED_REF, new_bytenr,
1157 num_bytes, parent, root->root_key.objectid);
1158 btrfs_init_data_ref(&ref, btrfs_header_owner(leaf),
1159 key.objectid, key.offset,
1160 root->root_key.objectid, false);
1161 ret = btrfs_inc_extent_ref(trans, &ref);
1162 if (ret) {
1163 btrfs_abort_transaction(trans, ret);
1164 break;
1165 }
1166
1167 btrfs_init_generic_ref(&ref, BTRFS_DROP_DELAYED_REF, bytenr,
1168 num_bytes, parent, root->root_key.objectid);
1169 btrfs_init_data_ref(&ref, btrfs_header_owner(leaf),
1170 key.objectid, key.offset,
1171 root->root_key.objectid, false);
1172 ret = btrfs_free_extent(trans, &ref);
1173 if (ret) {
1174 btrfs_abort_transaction(trans, ret);
1175 break;
1176 }
1177 }
1178 if (dirty)
1179 btrfs_mark_buffer_dirty(trans, leaf);
1180 if (inode)
1181 btrfs_add_delayed_iput(BTRFS_I(inode));
1182 return ret;
1183}
1184
1185static noinline_for_stack int memcmp_node_keys(const struct extent_buffer *eb,
1186 int slot, const struct btrfs_path *path,
1187 int level)
1188{
1189 struct btrfs_disk_key key1;
1190 struct btrfs_disk_key key2;
1191 btrfs_node_key(eb, &key1, slot);
1192 btrfs_node_key(path->nodes[level], &key2, path->slots[level]);
1193 return memcmp(&key1, &key2, sizeof(key1));
1194}
1195
1196/*
1197 * try to replace tree blocks in fs tree with the new blocks
1198 * in reloc tree. tree blocks haven't been modified since the
1199 * reloc tree was create can be replaced.
1200 *
1201 * if a block was replaced, level of the block + 1 is returned.
1202 * if no block got replaced, 0 is returned. if there are other
1203 * errors, a negative error number is returned.
1204 */
1205static noinline_for_stack
1206int replace_path(struct btrfs_trans_handle *trans, struct reloc_control *rc,
1207 struct btrfs_root *dest, struct btrfs_root *src,
1208 struct btrfs_path *path, struct btrfs_key *next_key,
1209 int lowest_level, int max_level)
1210{
1211 struct btrfs_fs_info *fs_info = dest->fs_info;
1212 struct extent_buffer *eb;
1213 struct extent_buffer *parent;
1214 struct btrfs_ref ref = { 0 };
1215 struct btrfs_key key;
1216 u64 old_bytenr;
1217 u64 new_bytenr;
1218 u64 old_ptr_gen;
1219 u64 new_ptr_gen;
1220 u64 last_snapshot;
1221 u32 blocksize;
1222 int cow = 0;
1223 int level;
1224 int ret;
1225 int slot;
1226
1227 ASSERT(src->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID);
1228 ASSERT(dest->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID);
1229
1230 last_snapshot = btrfs_root_last_snapshot(&src->root_item);
1231again:
1232 slot = path->slots[lowest_level];
1233 btrfs_node_key_to_cpu(path->nodes[lowest_level], &key, slot);
1234
1235 eb = btrfs_lock_root_node(dest);
1236 level = btrfs_header_level(eb);
1237
1238 if (level < lowest_level) {
1239 btrfs_tree_unlock(eb);
1240 free_extent_buffer(eb);
1241 return 0;
1242 }
1243
1244 if (cow) {
1245 ret = btrfs_cow_block(trans, dest, eb, NULL, 0, &eb,
1246 BTRFS_NESTING_COW);
1247 if (ret) {
1248 btrfs_tree_unlock(eb);
1249 free_extent_buffer(eb);
1250 return ret;
1251 }
1252 }
1253
1254 if (next_key) {
1255 next_key->objectid = (u64)-1;
1256 next_key->type = (u8)-1;
1257 next_key->offset = (u64)-1;
1258 }
1259
1260 parent = eb;
1261 while (1) {
1262 level = btrfs_header_level(parent);
1263 ASSERT(level >= lowest_level);
1264
1265 ret = btrfs_bin_search(parent, 0, &key, &slot);
1266 if (ret < 0)
1267 break;
1268 if (ret && slot > 0)
1269 slot--;
1270
1271 if (next_key && slot + 1 < btrfs_header_nritems(parent))
1272 btrfs_node_key_to_cpu(parent, next_key, slot + 1);
1273
1274 old_bytenr = btrfs_node_blockptr(parent, slot);
1275 blocksize = fs_info->nodesize;
1276 old_ptr_gen = btrfs_node_ptr_generation(parent, slot);
1277
1278 if (level <= max_level) {
1279 eb = path->nodes[level];
1280 new_bytenr = btrfs_node_blockptr(eb,
1281 path->slots[level]);
1282 new_ptr_gen = btrfs_node_ptr_generation(eb,
1283 path->slots[level]);
1284 } else {
1285 new_bytenr = 0;
1286 new_ptr_gen = 0;
1287 }
1288
1289 if (WARN_ON(new_bytenr > 0 && new_bytenr == old_bytenr)) {
1290 ret = level;
1291 break;
1292 }
1293
1294 if (new_bytenr == 0 || old_ptr_gen > last_snapshot ||
1295 memcmp_node_keys(parent, slot, path, level)) {
1296 if (level <= lowest_level) {
1297 ret = 0;
1298 break;
1299 }
1300
1301 eb = btrfs_read_node_slot(parent, slot);
1302 if (IS_ERR(eb)) {
1303 ret = PTR_ERR(eb);
1304 break;
1305 }
1306 btrfs_tree_lock(eb);
1307 if (cow) {
1308 ret = btrfs_cow_block(trans, dest, eb, parent,
1309 slot, &eb,
1310 BTRFS_NESTING_COW);
1311 if (ret) {
1312 btrfs_tree_unlock(eb);
1313 free_extent_buffer(eb);
1314 break;
1315 }
1316 }
1317
1318 btrfs_tree_unlock(parent);
1319 free_extent_buffer(parent);
1320
1321 parent = eb;
1322 continue;
1323 }
1324
1325 if (!cow) {
1326 btrfs_tree_unlock(parent);
1327 free_extent_buffer(parent);
1328 cow = 1;
1329 goto again;
1330 }
1331
1332 btrfs_node_key_to_cpu(path->nodes[level], &key,
1333 path->slots[level]);
1334 btrfs_release_path(path);
1335
1336 path->lowest_level = level;
1337 set_bit(BTRFS_ROOT_RESET_LOCKDEP_CLASS, &src->state);
1338 ret = btrfs_search_slot(trans, src, &key, path, 0, 1);
1339 clear_bit(BTRFS_ROOT_RESET_LOCKDEP_CLASS, &src->state);
1340 path->lowest_level = 0;
1341 if (ret) {
1342 if (ret > 0)
1343 ret = -ENOENT;
1344 break;
1345 }
1346
1347 /*
1348 * Info qgroup to trace both subtrees.
1349 *
1350 * We must trace both trees.
1351 * 1) Tree reloc subtree
1352 * If not traced, we will leak data numbers
1353 * 2) Fs subtree
1354 * If not traced, we will double count old data
1355 *
1356 * We don't scan the subtree right now, but only record
1357 * the swapped tree blocks.
1358 * The real subtree rescan is delayed until we have new
1359 * CoW on the subtree root node before transaction commit.
1360 */
1361 ret = btrfs_qgroup_add_swapped_blocks(trans, dest,
1362 rc->block_group, parent, slot,
1363 path->nodes[level], path->slots[level],
1364 last_snapshot);
1365 if (ret < 0)
1366 break;
1367 /*
1368 * swap blocks in fs tree and reloc tree.
1369 */
1370 btrfs_set_node_blockptr(parent, slot, new_bytenr);
1371 btrfs_set_node_ptr_generation(parent, slot, new_ptr_gen);
1372 btrfs_mark_buffer_dirty(trans, parent);
1373
1374 btrfs_set_node_blockptr(path->nodes[level],
1375 path->slots[level], old_bytenr);
1376 btrfs_set_node_ptr_generation(path->nodes[level],
1377 path->slots[level], old_ptr_gen);
1378 btrfs_mark_buffer_dirty(trans, path->nodes[level]);
1379
1380 btrfs_init_generic_ref(&ref, BTRFS_ADD_DELAYED_REF, old_bytenr,
1381 blocksize, path->nodes[level]->start,
1382 src->root_key.objectid);
1383 btrfs_init_tree_ref(&ref, level - 1, src->root_key.objectid,
1384 0, true);
1385 ret = btrfs_inc_extent_ref(trans, &ref);
1386 if (ret) {
1387 btrfs_abort_transaction(trans, ret);
1388 break;
1389 }
1390 btrfs_init_generic_ref(&ref, BTRFS_ADD_DELAYED_REF, new_bytenr,
1391 blocksize, 0, dest->root_key.objectid);
1392 btrfs_init_tree_ref(&ref, level - 1, dest->root_key.objectid, 0,
1393 true);
1394 ret = btrfs_inc_extent_ref(trans, &ref);
1395 if (ret) {
1396 btrfs_abort_transaction(trans, ret);
1397 break;
1398 }
1399
1400 /* We don't know the real owning_root, use 0. */
1401 btrfs_init_generic_ref(&ref, BTRFS_DROP_DELAYED_REF, new_bytenr,
1402 blocksize, path->nodes[level]->start, 0);
1403 btrfs_init_tree_ref(&ref, level - 1, src->root_key.objectid,
1404 0, true);
1405 ret = btrfs_free_extent(trans, &ref);
1406 if (ret) {
1407 btrfs_abort_transaction(trans, ret);
1408 break;
1409 }
1410
1411 /* We don't know the real owning_root, use 0. */
1412 btrfs_init_generic_ref(&ref, BTRFS_DROP_DELAYED_REF, old_bytenr,
1413 blocksize, 0, 0);
1414 btrfs_init_tree_ref(&ref, level - 1, dest->root_key.objectid,
1415 0, true);
1416 ret = btrfs_free_extent(trans, &ref);
1417 if (ret) {
1418 btrfs_abort_transaction(trans, ret);
1419 break;
1420 }
1421
1422 btrfs_unlock_up_safe(path, 0);
1423
1424 ret = level;
1425 break;
1426 }
1427 btrfs_tree_unlock(parent);
1428 free_extent_buffer(parent);
1429 return ret;
1430}
1431
1432/*
1433 * helper to find next relocated block in reloc tree
1434 */
1435static noinline_for_stack
1436int walk_up_reloc_tree(struct btrfs_root *root, struct btrfs_path *path,
1437 int *level)
1438{
1439 struct extent_buffer *eb;
1440 int i;
1441 u64 last_snapshot;
1442 u32 nritems;
1443
1444 last_snapshot = btrfs_root_last_snapshot(&root->root_item);
1445
1446 for (i = 0; i < *level; i++) {
1447 free_extent_buffer(path->nodes[i]);
1448 path->nodes[i] = NULL;
1449 }
1450
1451 for (i = *level; i < BTRFS_MAX_LEVEL && path->nodes[i]; i++) {
1452 eb = path->nodes[i];
1453 nritems = btrfs_header_nritems(eb);
1454 while (path->slots[i] + 1 < nritems) {
1455 path->slots[i]++;
1456 if (btrfs_node_ptr_generation(eb, path->slots[i]) <=
1457 last_snapshot)
1458 continue;
1459
1460 *level = i;
1461 return 0;
1462 }
1463 free_extent_buffer(path->nodes[i]);
1464 path->nodes[i] = NULL;
1465 }
1466 return 1;
1467}
1468
1469/*
1470 * walk down reloc tree to find relocated block of lowest level
1471 */
1472static noinline_for_stack
1473int walk_down_reloc_tree(struct btrfs_root *root, struct btrfs_path *path,
1474 int *level)
1475{
1476 struct extent_buffer *eb = NULL;
1477 int i;
1478 u64 ptr_gen = 0;
1479 u64 last_snapshot;
1480 u32 nritems;
1481
1482 last_snapshot = btrfs_root_last_snapshot(&root->root_item);
1483
1484 for (i = *level; i > 0; i--) {
1485 eb = path->nodes[i];
1486 nritems = btrfs_header_nritems(eb);
1487 while (path->slots[i] < nritems) {
1488 ptr_gen = btrfs_node_ptr_generation(eb, path->slots[i]);
1489 if (ptr_gen > last_snapshot)
1490 break;
1491 path->slots[i]++;
1492 }
1493 if (path->slots[i] >= nritems) {
1494 if (i == *level)
1495 break;
1496 *level = i + 1;
1497 return 0;
1498 }
1499 if (i == 1) {
1500 *level = i;
1501 return 0;
1502 }
1503
1504 eb = btrfs_read_node_slot(eb, path->slots[i]);
1505 if (IS_ERR(eb))
1506 return PTR_ERR(eb);
1507 BUG_ON(btrfs_header_level(eb) != i - 1);
1508 path->nodes[i - 1] = eb;
1509 path->slots[i - 1] = 0;
1510 }
1511 return 1;
1512}
1513
1514/*
1515 * invalidate extent cache for file extents whose key in range of
1516 * [min_key, max_key)
1517 */
1518static int invalidate_extent_cache(struct btrfs_root *root,
1519 const struct btrfs_key *min_key,
1520 const struct btrfs_key *max_key)
1521{
1522 struct btrfs_fs_info *fs_info = root->fs_info;
1523 struct inode *inode = NULL;
1524 u64 objectid;
1525 u64 start, end;
1526 u64 ino;
1527
1528 objectid = min_key->objectid;
1529 while (1) {
1530 struct extent_state *cached_state = NULL;
1531
1532 cond_resched();
1533 iput(inode);
1534
1535 if (objectid > max_key->objectid)
1536 break;
1537
1538 inode = find_next_inode(root, objectid);
1539 if (!inode)
1540 break;
1541 ino = btrfs_ino(BTRFS_I(inode));
1542
1543 if (ino > max_key->objectid) {
1544 iput(inode);
1545 break;
1546 }
1547
1548 objectid = ino + 1;
1549 if (!S_ISREG(inode->i_mode))
1550 continue;
1551
1552 if (unlikely(min_key->objectid == ino)) {
1553 if (min_key->type > BTRFS_EXTENT_DATA_KEY)
1554 continue;
1555 if (min_key->type < BTRFS_EXTENT_DATA_KEY)
1556 start = 0;
1557 else {
1558 start = min_key->offset;
1559 WARN_ON(!IS_ALIGNED(start, fs_info->sectorsize));
1560 }
1561 } else {
1562 start = 0;
1563 }
1564
1565 if (unlikely(max_key->objectid == ino)) {
1566 if (max_key->type < BTRFS_EXTENT_DATA_KEY)
1567 continue;
1568 if (max_key->type > BTRFS_EXTENT_DATA_KEY) {
1569 end = (u64)-1;
1570 } else {
1571 if (max_key->offset == 0)
1572 continue;
1573 end = max_key->offset;
1574 WARN_ON(!IS_ALIGNED(end, fs_info->sectorsize));
1575 end--;
1576 }
1577 } else {
1578 end = (u64)-1;
1579 }
1580
1581 /* the lock_extent waits for read_folio to complete */
1582 lock_extent(&BTRFS_I(inode)->io_tree, start, end, &cached_state);
1583 btrfs_drop_extent_map_range(BTRFS_I(inode), start, end, true);
1584 unlock_extent(&BTRFS_I(inode)->io_tree, start, end, &cached_state);
1585 }
1586 return 0;
1587}
1588
1589static int find_next_key(struct btrfs_path *path, int level,
1590 struct btrfs_key *key)
1591
1592{
1593 while (level < BTRFS_MAX_LEVEL) {
1594 if (!path->nodes[level])
1595 break;
1596 if (path->slots[level] + 1 <
1597 btrfs_header_nritems(path->nodes[level])) {
1598 btrfs_node_key_to_cpu(path->nodes[level], key,
1599 path->slots[level] + 1);
1600 return 0;
1601 }
1602 level++;
1603 }
1604 return 1;
1605}
1606
1607/*
1608 * Insert current subvolume into reloc_control::dirty_subvol_roots
1609 */
1610static int insert_dirty_subvol(struct btrfs_trans_handle *trans,
1611 struct reloc_control *rc,
1612 struct btrfs_root *root)
1613{
1614 struct btrfs_root *reloc_root = root->reloc_root;
1615 struct btrfs_root_item *reloc_root_item;
1616 int ret;
1617
1618 /* @root must be a subvolume tree root with a valid reloc tree */
1619 ASSERT(root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID);
1620 ASSERT(reloc_root);
1621
1622 reloc_root_item = &reloc_root->root_item;
1623 memset(&reloc_root_item->drop_progress, 0,
1624 sizeof(reloc_root_item->drop_progress));
1625 btrfs_set_root_drop_level(reloc_root_item, 0);
1626 btrfs_set_root_refs(reloc_root_item, 0);
1627 ret = btrfs_update_reloc_root(trans, root);
1628 if (ret)
1629 return ret;
1630
1631 if (list_empty(&root->reloc_dirty_list)) {
1632 btrfs_grab_root(root);
1633 list_add_tail(&root->reloc_dirty_list, &rc->dirty_subvol_roots);
1634 }
1635
1636 return 0;
1637}
1638
1639static int clean_dirty_subvols(struct reloc_control *rc)
1640{
1641 struct btrfs_root *root;
1642 struct btrfs_root *next;
1643 int ret = 0;
1644 int ret2;
1645
1646 list_for_each_entry_safe(root, next, &rc->dirty_subvol_roots,
1647 reloc_dirty_list) {
1648 if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) {
1649 /* Merged subvolume, cleanup its reloc root */
1650 struct btrfs_root *reloc_root = root->reloc_root;
1651
1652 list_del_init(&root->reloc_dirty_list);
1653 root->reloc_root = NULL;
1654 /*
1655 * Need barrier to ensure clear_bit() only happens after
1656 * root->reloc_root = NULL. Pairs with have_reloc_root.
1657 */
1658 smp_wmb();
1659 clear_bit(BTRFS_ROOT_DEAD_RELOC_TREE, &root->state);
1660 if (reloc_root) {
1661 /*
1662 * btrfs_drop_snapshot drops our ref we hold for
1663 * ->reloc_root. If it fails however we must
1664 * drop the ref ourselves.
1665 */
1666 ret2 = btrfs_drop_snapshot(reloc_root, 0, 1);
1667 if (ret2 < 0) {
1668 btrfs_put_root(reloc_root);
1669 if (!ret)
1670 ret = ret2;
1671 }
1672 }
1673 btrfs_put_root(root);
1674 } else {
1675 /* Orphan reloc tree, just clean it up */
1676 ret2 = btrfs_drop_snapshot(root, 0, 1);
1677 if (ret2 < 0) {
1678 btrfs_put_root(root);
1679 if (!ret)
1680 ret = ret2;
1681 }
1682 }
1683 }
1684 return ret;
1685}
1686
1687/*
1688 * merge the relocated tree blocks in reloc tree with corresponding
1689 * fs tree.
1690 */
1691static noinline_for_stack int merge_reloc_root(struct reloc_control *rc,
1692 struct btrfs_root *root)
1693{
1694 struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
1695 struct btrfs_key key;
1696 struct btrfs_key next_key;
1697 struct btrfs_trans_handle *trans = NULL;
1698 struct btrfs_root *reloc_root;
1699 struct btrfs_root_item *root_item;
1700 struct btrfs_path *path;
1701 struct extent_buffer *leaf;
1702 int reserve_level;
1703 int level;
1704 int max_level;
1705 int replaced = 0;
1706 int ret = 0;
1707 u32 min_reserved;
1708
1709 path = btrfs_alloc_path();
1710 if (!path)
1711 return -ENOMEM;
1712 path->reada = READA_FORWARD;
1713
1714 reloc_root = root->reloc_root;
1715 root_item = &reloc_root->root_item;
1716
1717 if (btrfs_disk_key_objectid(&root_item->drop_progress) == 0) {
1718 level = btrfs_root_level(root_item);
1719 atomic_inc(&reloc_root->node->refs);
1720 path->nodes[level] = reloc_root->node;
1721 path->slots[level] = 0;
1722 } else {
1723 btrfs_disk_key_to_cpu(&key, &root_item->drop_progress);
1724
1725 level = btrfs_root_drop_level(root_item);
1726 BUG_ON(level == 0);
1727 path->lowest_level = level;
1728 ret = btrfs_search_slot(NULL, reloc_root, &key, path, 0, 0);
1729 path->lowest_level = 0;
1730 if (ret < 0) {
1731 btrfs_free_path(path);
1732 return ret;
1733 }
1734
1735 btrfs_node_key_to_cpu(path->nodes[level], &next_key,
1736 path->slots[level]);
1737 WARN_ON(memcmp(&key, &next_key, sizeof(key)));
1738
1739 btrfs_unlock_up_safe(path, 0);
1740 }
1741
1742 /*
1743 * In merge_reloc_root(), we modify the upper level pointer to swap the
1744 * tree blocks between reloc tree and subvolume tree. Thus for tree
1745 * block COW, we COW at most from level 1 to root level for each tree.
1746 *
1747 * Thus the needed metadata size is at most root_level * nodesize,
1748 * and * 2 since we have two trees to COW.
1749 */
1750 reserve_level = max_t(int, 1, btrfs_root_level(root_item));
1751 min_reserved = fs_info->nodesize * reserve_level * 2;
1752 memset(&next_key, 0, sizeof(next_key));
1753
1754 while (1) {
1755 ret = btrfs_block_rsv_refill(fs_info, rc->block_rsv,
1756 min_reserved,
1757 BTRFS_RESERVE_FLUSH_LIMIT);
1758 if (ret)
1759 goto out;
1760 trans = btrfs_start_transaction(root, 0);
1761 if (IS_ERR(trans)) {
1762 ret = PTR_ERR(trans);
1763 trans = NULL;
1764 goto out;
1765 }
1766
1767 /*
1768 * At this point we no longer have a reloc_control, so we can't
1769 * depend on btrfs_init_reloc_root to update our last_trans.
1770 *
1771 * But that's ok, we started the trans handle on our
1772 * corresponding fs_root, which means it's been added to the
1773 * dirty list. At commit time we'll still call
1774 * btrfs_update_reloc_root() and update our root item
1775 * appropriately.
1776 */
1777 reloc_root->last_trans = trans->transid;
1778 trans->block_rsv = rc->block_rsv;
1779
1780 replaced = 0;
1781 max_level = level;
1782
1783 ret = walk_down_reloc_tree(reloc_root, path, &level);
1784 if (ret < 0)
1785 goto out;
1786 if (ret > 0)
1787 break;
1788
1789 if (!find_next_key(path, level, &key) &&
1790 btrfs_comp_cpu_keys(&next_key, &key) >= 0) {
1791 ret = 0;
1792 } else {
1793 ret = replace_path(trans, rc, root, reloc_root, path,
1794 &next_key, level, max_level);
1795 }
1796 if (ret < 0)
1797 goto out;
1798 if (ret > 0) {
1799 level = ret;
1800 btrfs_node_key_to_cpu(path->nodes[level], &key,
1801 path->slots[level]);
1802 replaced = 1;
1803 }
1804
1805 ret = walk_up_reloc_tree(reloc_root, path, &level);
1806 if (ret > 0)
1807 break;
1808
1809 BUG_ON(level == 0);
1810 /*
1811 * save the merging progress in the drop_progress.
1812 * this is OK since root refs == 1 in this case.
1813 */
1814 btrfs_node_key(path->nodes[level], &root_item->drop_progress,
1815 path->slots[level]);
1816 btrfs_set_root_drop_level(root_item, level);
1817
1818 btrfs_end_transaction_throttle(trans);
1819 trans = NULL;
1820
1821 btrfs_btree_balance_dirty(fs_info);
1822
1823 if (replaced && rc->stage == UPDATE_DATA_PTRS)
1824 invalidate_extent_cache(root, &key, &next_key);
1825 }
1826
1827 /*
1828 * handle the case only one block in the fs tree need to be
1829 * relocated and the block is tree root.
1830 */
1831 leaf = btrfs_lock_root_node(root);
1832 ret = btrfs_cow_block(trans, root, leaf, NULL, 0, &leaf,
1833 BTRFS_NESTING_COW);
1834 btrfs_tree_unlock(leaf);
1835 free_extent_buffer(leaf);
1836out:
1837 btrfs_free_path(path);
1838
1839 if (ret == 0) {
1840 ret = insert_dirty_subvol(trans, rc, root);
1841 if (ret)
1842 btrfs_abort_transaction(trans, ret);
1843 }
1844
1845 if (trans)
1846 btrfs_end_transaction_throttle(trans);
1847
1848 btrfs_btree_balance_dirty(fs_info);
1849
1850 if (replaced && rc->stage == UPDATE_DATA_PTRS)
1851 invalidate_extent_cache(root, &key, &next_key);
1852
1853 return ret;
1854}
1855
1856static noinline_for_stack
1857int prepare_to_merge(struct reloc_control *rc, int err)
1858{
1859 struct btrfs_root *root = rc->extent_root;
1860 struct btrfs_fs_info *fs_info = root->fs_info;
1861 struct btrfs_root *reloc_root;
1862 struct btrfs_trans_handle *trans;
1863 LIST_HEAD(reloc_roots);
1864 u64 num_bytes = 0;
1865 int ret;
1866
1867 mutex_lock(&fs_info->reloc_mutex);
1868 rc->merging_rsv_size += fs_info->nodesize * (BTRFS_MAX_LEVEL - 1) * 2;
1869 rc->merging_rsv_size += rc->nodes_relocated * 2;
1870 mutex_unlock(&fs_info->reloc_mutex);
1871
1872again:
1873 if (!err) {
1874 num_bytes = rc->merging_rsv_size;
1875 ret = btrfs_block_rsv_add(fs_info, rc->block_rsv, num_bytes,
1876 BTRFS_RESERVE_FLUSH_ALL);
1877 if (ret)
1878 err = ret;
1879 }
1880
1881 trans = btrfs_join_transaction(rc->extent_root);
1882 if (IS_ERR(trans)) {
1883 if (!err)
1884 btrfs_block_rsv_release(fs_info, rc->block_rsv,
1885 num_bytes, NULL);
1886 return PTR_ERR(trans);
1887 }
1888
1889 if (!err) {
1890 if (num_bytes != rc->merging_rsv_size) {
1891 btrfs_end_transaction(trans);
1892 btrfs_block_rsv_release(fs_info, rc->block_rsv,
1893 num_bytes, NULL);
1894 goto again;
1895 }
1896 }
1897
1898 rc->merge_reloc_tree = true;
1899
1900 while (!list_empty(&rc->reloc_roots)) {
1901 reloc_root = list_entry(rc->reloc_roots.next,
1902 struct btrfs_root, root_list);
1903 list_del_init(&reloc_root->root_list);
1904
1905 root = btrfs_get_fs_root(fs_info, reloc_root->root_key.offset,
1906 false);
1907 if (IS_ERR(root)) {
1908 /*
1909 * Even if we have an error we need this reloc root
1910 * back on our list so we can clean up properly.
1911 */
1912 list_add(&reloc_root->root_list, &reloc_roots);
1913 btrfs_abort_transaction(trans, (int)PTR_ERR(root));
1914 if (!err)
1915 err = PTR_ERR(root);
1916 break;
1917 }
1918
1919 if (unlikely(root->reloc_root != reloc_root)) {
1920 if (root->reloc_root) {
1921 btrfs_err(fs_info,
1922"reloc tree mismatch, root %lld has reloc root key (%lld %u %llu) gen %llu, expect reloc root key (%lld %u %llu) gen %llu",
1923 root->root_key.objectid,
1924 root->reloc_root->root_key.objectid,
1925 root->reloc_root->root_key.type,
1926 root->reloc_root->root_key.offset,
1927 btrfs_root_generation(
1928 &root->reloc_root->root_item),
1929 reloc_root->root_key.objectid,
1930 reloc_root->root_key.type,
1931 reloc_root->root_key.offset,
1932 btrfs_root_generation(
1933 &reloc_root->root_item));
1934 } else {
1935 btrfs_err(fs_info,
1936"reloc tree mismatch, root %lld has no reloc root, expect reloc root key (%lld %u %llu) gen %llu",
1937 root->root_key.objectid,
1938 reloc_root->root_key.objectid,
1939 reloc_root->root_key.type,
1940 reloc_root->root_key.offset,
1941 btrfs_root_generation(
1942 &reloc_root->root_item));
1943 }
1944 list_add(&reloc_root->root_list, &reloc_roots);
1945 btrfs_put_root(root);
1946 btrfs_abort_transaction(trans, -EUCLEAN);
1947 if (!err)
1948 err = -EUCLEAN;
1949 break;
1950 }
1951
1952 /*
1953 * set reference count to 1, so btrfs_recover_relocation
1954 * knows it should resumes merging
1955 */
1956 if (!err)
1957 btrfs_set_root_refs(&reloc_root->root_item, 1);
1958 ret = btrfs_update_reloc_root(trans, root);
1959
1960 /*
1961 * Even if we have an error we need this reloc root back on our
1962 * list so we can clean up properly.
1963 */
1964 list_add(&reloc_root->root_list, &reloc_roots);
1965 btrfs_put_root(root);
1966
1967 if (ret) {
1968 btrfs_abort_transaction(trans, ret);
1969 if (!err)
1970 err = ret;
1971 break;
1972 }
1973 }
1974
1975 list_splice(&reloc_roots, &rc->reloc_roots);
1976
1977 if (!err)
1978 err = btrfs_commit_transaction(trans);
1979 else
1980 btrfs_end_transaction(trans);
1981 return err;
1982}
1983
1984static noinline_for_stack
1985void free_reloc_roots(struct list_head *list)
1986{
1987 struct btrfs_root *reloc_root, *tmp;
1988
1989 list_for_each_entry_safe(reloc_root, tmp, list, root_list)
1990 __del_reloc_root(reloc_root);
1991}
1992
1993static noinline_for_stack
1994void merge_reloc_roots(struct reloc_control *rc)
1995{
1996 struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
1997 struct btrfs_root *root;
1998 struct btrfs_root *reloc_root;
1999 LIST_HEAD(reloc_roots);
2000 int found = 0;
2001 int ret = 0;
2002again:
2003 root = rc->extent_root;
2004
2005 /*
2006 * this serializes us with btrfs_record_root_in_transaction,
2007 * we have to make sure nobody is in the middle of
2008 * adding their roots to the list while we are
2009 * doing this splice
2010 */
2011 mutex_lock(&fs_info->reloc_mutex);
2012 list_splice_init(&rc->reloc_roots, &reloc_roots);
2013 mutex_unlock(&fs_info->reloc_mutex);
2014
2015 while (!list_empty(&reloc_roots)) {
2016 found = 1;
2017 reloc_root = list_entry(reloc_roots.next,
2018 struct btrfs_root, root_list);
2019
2020 root = btrfs_get_fs_root(fs_info, reloc_root->root_key.offset,
2021 false);
2022 if (btrfs_root_refs(&reloc_root->root_item) > 0) {
2023 if (WARN_ON(IS_ERR(root))) {
2024 /*
2025 * For recovery we read the fs roots on mount,
2026 * and if we didn't find the root then we marked
2027 * the reloc root as a garbage root. For normal
2028 * relocation obviously the root should exist in
2029 * memory. However there's no reason we can't
2030 * handle the error properly here just in case.
2031 */
2032 ret = PTR_ERR(root);
2033 goto out;
2034 }
2035 if (WARN_ON(root->reloc_root != reloc_root)) {
2036 /*
2037 * This can happen if on-disk metadata has some
2038 * corruption, e.g. bad reloc tree key offset.
2039 */
2040 ret = -EINVAL;
2041 goto out;
2042 }
2043 ret = merge_reloc_root(rc, root);
2044 btrfs_put_root(root);
2045 if (ret) {
2046 if (list_empty(&reloc_root->root_list))
2047 list_add_tail(&reloc_root->root_list,
2048 &reloc_roots);
2049 goto out;
2050 }
2051 } else {
2052 if (!IS_ERR(root)) {
2053 if (root->reloc_root == reloc_root) {
2054 root->reloc_root = NULL;
2055 btrfs_put_root(reloc_root);
2056 }
2057 clear_bit(BTRFS_ROOT_DEAD_RELOC_TREE,
2058 &root->state);
2059 btrfs_put_root(root);
2060 }
2061
2062 list_del_init(&reloc_root->root_list);
2063 /* Don't forget to queue this reloc root for cleanup */
2064 list_add_tail(&reloc_root->reloc_dirty_list,
2065 &rc->dirty_subvol_roots);
2066 }
2067 }
2068
2069 if (found) {
2070 found = 0;
2071 goto again;
2072 }
2073out:
2074 if (ret) {
2075 btrfs_handle_fs_error(fs_info, ret, NULL);
2076 free_reloc_roots(&reloc_roots);
2077
2078 /* new reloc root may be added */
2079 mutex_lock(&fs_info->reloc_mutex);
2080 list_splice_init(&rc->reloc_roots, &reloc_roots);
2081 mutex_unlock(&fs_info->reloc_mutex);
2082 free_reloc_roots(&reloc_roots);
2083 }
2084
2085 /*
2086 * We used to have
2087 *
2088 * BUG_ON(!RB_EMPTY_ROOT(&rc->reloc_root_tree.rb_root));
2089 *
2090 * here, but it's wrong. If we fail to start the transaction in
2091 * prepare_to_merge() we will have only 0 ref reloc roots, none of which
2092 * have actually been removed from the reloc_root_tree rb tree. This is
2093 * fine because we're bailing here, and we hold a reference on the root
2094 * for the list that holds it, so these roots will be cleaned up when we
2095 * do the reloc_dirty_list afterwards. Meanwhile the root->reloc_root
2096 * will be cleaned up on unmount.
2097 *
2098 * The remaining nodes will be cleaned up by free_reloc_control.
2099 */
2100}
2101
2102static void free_block_list(struct rb_root *blocks)
2103{
2104 struct tree_block *block;
2105 struct rb_node *rb_node;
2106 while ((rb_node = rb_first(blocks))) {
2107 block = rb_entry(rb_node, struct tree_block, rb_node);
2108 rb_erase(rb_node, blocks);
2109 kfree(block);
2110 }
2111}
2112
2113static int record_reloc_root_in_trans(struct btrfs_trans_handle *trans,
2114 struct btrfs_root *reloc_root)
2115{
2116 struct btrfs_fs_info *fs_info = reloc_root->fs_info;
2117 struct btrfs_root *root;
2118 int ret;
2119
2120 if (reloc_root->last_trans == trans->transid)
2121 return 0;
2122
2123 root = btrfs_get_fs_root(fs_info, reloc_root->root_key.offset, false);
2124
2125 /*
2126 * This should succeed, since we can't have a reloc root without having
2127 * already looked up the actual root and created the reloc root for this
2128 * root.
2129 *
2130 * However if there's some sort of corruption where we have a ref to a
2131 * reloc root without a corresponding root this could return ENOENT.
2132 */
2133 if (IS_ERR(root)) {
2134 ASSERT(0);
2135 return PTR_ERR(root);
2136 }
2137 if (root->reloc_root != reloc_root) {
2138 ASSERT(0);
2139 btrfs_err(fs_info,
2140 "root %llu has two reloc roots associated with it",
2141 reloc_root->root_key.offset);
2142 btrfs_put_root(root);
2143 return -EUCLEAN;
2144 }
2145 ret = btrfs_record_root_in_trans(trans, root);
2146 btrfs_put_root(root);
2147
2148 return ret;
2149}
2150
2151static noinline_for_stack
2152struct btrfs_root *select_reloc_root(struct btrfs_trans_handle *trans,
2153 struct reloc_control *rc,
2154 struct btrfs_backref_node *node,
2155 struct btrfs_backref_edge *edges[])
2156{
2157 struct btrfs_backref_node *next;
2158 struct btrfs_root *root;
2159 int index = 0;
2160 int ret;
2161
2162 next = node;
2163 while (1) {
2164 cond_resched();
2165 next = walk_up_backref(next, edges, &index);
2166 root = next->root;
2167
2168 /*
2169 * If there is no root, then our references for this block are
2170 * incomplete, as we should be able to walk all the way up to a
2171 * block that is owned by a root.
2172 *
2173 * This path is only for SHAREABLE roots, so if we come upon a
2174 * non-SHAREABLE root then we have backrefs that resolve
2175 * improperly.
2176 *
2177 * Both of these cases indicate file system corruption, or a bug
2178 * in the backref walking code.
2179 */
2180 if (!root) {
2181 ASSERT(0);
2182 btrfs_err(trans->fs_info,
2183 "bytenr %llu doesn't have a backref path ending in a root",
2184 node->bytenr);
2185 return ERR_PTR(-EUCLEAN);
2186 }
2187 if (!test_bit(BTRFS_ROOT_SHAREABLE, &root->state)) {
2188 ASSERT(0);
2189 btrfs_err(trans->fs_info,
2190 "bytenr %llu has multiple refs with one ending in a non-shareable root",
2191 node->bytenr);
2192 return ERR_PTR(-EUCLEAN);
2193 }
2194
2195 if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID) {
2196 ret = record_reloc_root_in_trans(trans, root);
2197 if (ret)
2198 return ERR_PTR(ret);
2199 break;
2200 }
2201
2202 ret = btrfs_record_root_in_trans(trans, root);
2203 if (ret)
2204 return ERR_PTR(ret);
2205 root = root->reloc_root;
2206
2207 /*
2208 * We could have raced with another thread which failed, so
2209 * root->reloc_root may not be set, return ENOENT in this case.
2210 */
2211 if (!root)
2212 return ERR_PTR(-ENOENT);
2213
2214 if (next->new_bytenr != root->node->start) {
2215 /*
2216 * We just created the reloc root, so we shouldn't have
2217 * ->new_bytenr set and this shouldn't be in the changed
2218 * list. If it is then we have multiple roots pointing
2219 * at the same bytenr which indicates corruption, or
2220 * we've made a mistake in the backref walking code.
2221 */
2222 ASSERT(next->new_bytenr == 0);
2223 ASSERT(list_empty(&next->list));
2224 if (next->new_bytenr || !list_empty(&next->list)) {
2225 btrfs_err(trans->fs_info,
2226 "bytenr %llu possibly has multiple roots pointing at the same bytenr %llu",
2227 node->bytenr, next->bytenr);
2228 return ERR_PTR(-EUCLEAN);
2229 }
2230
2231 next->new_bytenr = root->node->start;
2232 btrfs_put_root(next->root);
2233 next->root = btrfs_grab_root(root);
2234 ASSERT(next->root);
2235 list_add_tail(&next->list,
2236 &rc->backref_cache.changed);
2237 mark_block_processed(rc, next);
2238 break;
2239 }
2240
2241 WARN_ON(1);
2242 root = NULL;
2243 next = walk_down_backref(edges, &index);
2244 if (!next || next->level <= node->level)
2245 break;
2246 }
2247 if (!root) {
2248 /*
2249 * This can happen if there's fs corruption or if there's a bug
2250 * in the backref lookup code.
2251 */
2252 ASSERT(0);
2253 return ERR_PTR(-ENOENT);
2254 }
2255
2256 next = node;
2257 /* setup backref node path for btrfs_reloc_cow_block */
2258 while (1) {
2259 rc->backref_cache.path[next->level] = next;
2260 if (--index < 0)
2261 break;
2262 next = edges[index]->node[UPPER];
2263 }
2264 return root;
2265}
2266
2267/*
2268 * Select a tree root for relocation.
2269 *
2270 * Return NULL if the block is not shareable. We should use do_relocation() in
2271 * this case.
2272 *
2273 * Return a tree root pointer if the block is shareable.
2274 * Return -ENOENT if the block is root of reloc tree.
2275 */
2276static noinline_for_stack
2277struct btrfs_root *select_one_root(struct btrfs_backref_node *node)
2278{
2279 struct btrfs_backref_node *next;
2280 struct btrfs_root *root;
2281 struct btrfs_root *fs_root = NULL;
2282 struct btrfs_backref_edge *edges[BTRFS_MAX_LEVEL - 1];
2283 int index = 0;
2284
2285 next = node;
2286 while (1) {
2287 cond_resched();
2288 next = walk_up_backref(next, edges, &index);
2289 root = next->root;
2290
2291 /*
2292 * This can occur if we have incomplete extent refs leading all
2293 * the way up a particular path, in this case return -EUCLEAN.
2294 */
2295 if (!root)
2296 return ERR_PTR(-EUCLEAN);
2297
2298 /* No other choice for non-shareable tree */
2299 if (!test_bit(BTRFS_ROOT_SHAREABLE, &root->state))
2300 return root;
2301
2302 if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID)
2303 fs_root = root;
2304
2305 if (next != node)
2306 return NULL;
2307
2308 next = walk_down_backref(edges, &index);
2309 if (!next || next->level <= node->level)
2310 break;
2311 }
2312
2313 if (!fs_root)
2314 return ERR_PTR(-ENOENT);
2315 return fs_root;
2316}
2317
2318static noinline_for_stack
2319u64 calcu_metadata_size(struct reloc_control *rc,
2320 struct btrfs_backref_node *node, int reserve)
2321{
2322 struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
2323 struct btrfs_backref_node *next = node;
2324 struct btrfs_backref_edge *edge;
2325 struct btrfs_backref_edge *edges[BTRFS_MAX_LEVEL - 1];
2326 u64 num_bytes = 0;
2327 int index = 0;
2328
2329 BUG_ON(reserve && node->processed);
2330
2331 while (next) {
2332 cond_resched();
2333 while (1) {
2334 if (next->processed && (reserve || next != node))
2335 break;
2336
2337 num_bytes += fs_info->nodesize;
2338
2339 if (list_empty(&next->upper))
2340 break;
2341
2342 edge = list_entry(next->upper.next,
2343 struct btrfs_backref_edge, list[LOWER]);
2344 edges[index++] = edge;
2345 next = edge->node[UPPER];
2346 }
2347 next = walk_down_backref(edges, &index);
2348 }
2349 return num_bytes;
2350}
2351
2352static int reserve_metadata_space(struct btrfs_trans_handle *trans,
2353 struct reloc_control *rc,
2354 struct btrfs_backref_node *node)
2355{
2356 struct btrfs_root *root = rc->extent_root;
2357 struct btrfs_fs_info *fs_info = root->fs_info;
2358 u64 num_bytes;
2359 int ret;
2360 u64 tmp;
2361
2362 num_bytes = calcu_metadata_size(rc, node, 1) * 2;
2363
2364 trans->block_rsv = rc->block_rsv;
2365 rc->reserved_bytes += num_bytes;
2366
2367 /*
2368 * We are under a transaction here so we can only do limited flushing.
2369 * If we get an enospc just kick back -EAGAIN so we know to drop the
2370 * transaction and try to refill when we can flush all the things.
2371 */
2372 ret = btrfs_block_rsv_refill(fs_info, rc->block_rsv, num_bytes,
2373 BTRFS_RESERVE_FLUSH_LIMIT);
2374 if (ret) {
2375 tmp = fs_info->nodesize * RELOCATION_RESERVED_NODES;
2376 while (tmp <= rc->reserved_bytes)
2377 tmp <<= 1;
2378 /*
2379 * only one thread can access block_rsv at this point,
2380 * so we don't need hold lock to protect block_rsv.
2381 * we expand more reservation size here to allow enough
2382 * space for relocation and we will return earlier in
2383 * enospc case.
2384 */
2385 rc->block_rsv->size = tmp + fs_info->nodesize *
2386 RELOCATION_RESERVED_NODES;
2387 return -EAGAIN;
2388 }
2389
2390 return 0;
2391}
2392
2393/*
2394 * relocate a block tree, and then update pointers in upper level
2395 * blocks that reference the block to point to the new location.
2396 *
2397 * if called by link_to_upper, the block has already been relocated.
2398 * in that case this function just updates pointers.
2399 */
2400static int do_relocation(struct btrfs_trans_handle *trans,
2401 struct reloc_control *rc,
2402 struct btrfs_backref_node *node,
2403 struct btrfs_key *key,
2404 struct btrfs_path *path, int lowest)
2405{
2406 struct btrfs_backref_node *upper;
2407 struct btrfs_backref_edge *edge;
2408 struct btrfs_backref_edge *edges[BTRFS_MAX_LEVEL - 1];
2409 struct btrfs_root *root;
2410 struct extent_buffer *eb;
2411 u32 blocksize;
2412 u64 bytenr;
2413 int slot;
2414 int ret = 0;
2415
2416 /*
2417 * If we are lowest then this is the first time we're processing this
2418 * block, and thus shouldn't have an eb associated with it yet.
2419 */
2420 ASSERT(!lowest || !node->eb);
2421
2422 path->lowest_level = node->level + 1;
2423 rc->backref_cache.path[node->level] = node;
2424 list_for_each_entry(edge, &node->upper, list[LOWER]) {
2425 struct btrfs_ref ref = { 0 };
2426
2427 cond_resched();
2428
2429 upper = edge->node[UPPER];
2430 root = select_reloc_root(trans, rc, upper, edges);
2431 if (IS_ERR(root)) {
2432 ret = PTR_ERR(root);
2433 goto next;
2434 }
2435
2436 if (upper->eb && !upper->locked) {
2437 if (!lowest) {
2438 ret = btrfs_bin_search(upper->eb, 0, key, &slot);
2439 if (ret < 0)
2440 goto next;
2441 BUG_ON(ret);
2442 bytenr = btrfs_node_blockptr(upper->eb, slot);
2443 if (node->eb->start == bytenr)
2444 goto next;
2445 }
2446 btrfs_backref_drop_node_buffer(upper);
2447 }
2448
2449 if (!upper->eb) {
2450 ret = btrfs_search_slot(trans, root, key, path, 0, 1);
2451 if (ret) {
2452 if (ret > 0)
2453 ret = -ENOENT;
2454
2455 btrfs_release_path(path);
2456 break;
2457 }
2458
2459 if (!upper->eb) {
2460 upper->eb = path->nodes[upper->level];
2461 path->nodes[upper->level] = NULL;
2462 } else {
2463 BUG_ON(upper->eb != path->nodes[upper->level]);
2464 }
2465
2466 upper->locked = 1;
2467 path->locks[upper->level] = 0;
2468
2469 slot = path->slots[upper->level];
2470 btrfs_release_path(path);
2471 } else {
2472 ret = btrfs_bin_search(upper->eb, 0, key, &slot);
2473 if (ret < 0)
2474 goto next;
2475 BUG_ON(ret);
2476 }
2477
2478 bytenr = btrfs_node_blockptr(upper->eb, slot);
2479 if (lowest) {
2480 if (bytenr != node->bytenr) {
2481 btrfs_err(root->fs_info,
2482 "lowest leaf/node mismatch: bytenr %llu node->bytenr %llu slot %d upper %llu",
2483 bytenr, node->bytenr, slot,
2484 upper->eb->start);
2485 ret = -EIO;
2486 goto next;
2487 }
2488 } else {
2489 if (node->eb->start == bytenr)
2490 goto next;
2491 }
2492
2493 blocksize = root->fs_info->nodesize;
2494 eb = btrfs_read_node_slot(upper->eb, slot);
2495 if (IS_ERR(eb)) {
2496 ret = PTR_ERR(eb);
2497 goto next;
2498 }
2499 btrfs_tree_lock(eb);
2500
2501 if (!node->eb) {
2502 ret = btrfs_cow_block(trans, root, eb, upper->eb,
2503 slot, &eb, BTRFS_NESTING_COW);
2504 btrfs_tree_unlock(eb);
2505 free_extent_buffer(eb);
2506 if (ret < 0)
2507 goto next;
2508 /*
2509 * We've just COWed this block, it should have updated
2510 * the correct backref node entry.
2511 */
2512 ASSERT(node->eb == eb);
2513 } else {
2514 btrfs_set_node_blockptr(upper->eb, slot,
2515 node->eb->start);
2516 btrfs_set_node_ptr_generation(upper->eb, slot,
2517 trans->transid);
2518 btrfs_mark_buffer_dirty(trans, upper->eb);
2519
2520 btrfs_init_generic_ref(&ref, BTRFS_ADD_DELAYED_REF,
2521 node->eb->start, blocksize,
2522 upper->eb->start,
2523 btrfs_header_owner(upper->eb));
2524 btrfs_init_tree_ref(&ref, node->level,
2525 btrfs_header_owner(upper->eb),
2526 root->root_key.objectid, false);
2527 ret = btrfs_inc_extent_ref(trans, &ref);
2528 if (!ret)
2529 ret = btrfs_drop_subtree(trans, root, eb,
2530 upper->eb);
2531 if (ret)
2532 btrfs_abort_transaction(trans, ret);
2533 }
2534next:
2535 if (!upper->pending)
2536 btrfs_backref_drop_node_buffer(upper);
2537 else
2538 btrfs_backref_unlock_node_buffer(upper);
2539 if (ret)
2540 break;
2541 }
2542
2543 if (!ret && node->pending) {
2544 btrfs_backref_drop_node_buffer(node);
2545 list_move_tail(&node->list, &rc->backref_cache.changed);
2546 node->pending = 0;
2547 }
2548
2549 path->lowest_level = 0;
2550
2551 /*
2552 * We should have allocated all of our space in the block rsv and thus
2553 * shouldn't ENOSPC.
2554 */
2555 ASSERT(ret != -ENOSPC);
2556 return ret;
2557}
2558
2559static int link_to_upper(struct btrfs_trans_handle *trans,
2560 struct reloc_control *rc,
2561 struct btrfs_backref_node *node,
2562 struct btrfs_path *path)
2563{
2564 struct btrfs_key key;
2565
2566 btrfs_node_key_to_cpu(node->eb, &key, 0);
2567 return do_relocation(trans, rc, node, &key, path, 0);
2568}
2569
2570static int finish_pending_nodes(struct btrfs_trans_handle *trans,
2571 struct reloc_control *rc,
2572 struct btrfs_path *path, int err)
2573{
2574 LIST_HEAD(list);
2575 struct btrfs_backref_cache *cache = &rc->backref_cache;
2576 struct btrfs_backref_node *node;
2577 int level;
2578 int ret;
2579
2580 for (level = 0; level < BTRFS_MAX_LEVEL; level++) {
2581 while (!list_empty(&cache->pending[level])) {
2582 node = list_entry(cache->pending[level].next,
2583 struct btrfs_backref_node, list);
2584 list_move_tail(&node->list, &list);
2585 BUG_ON(!node->pending);
2586
2587 if (!err) {
2588 ret = link_to_upper(trans, rc, node, path);
2589 if (ret < 0)
2590 err = ret;
2591 }
2592 }
2593 list_splice_init(&list, &cache->pending[level]);
2594 }
2595 return err;
2596}
2597
2598/*
2599 * mark a block and all blocks directly/indirectly reference the block
2600 * as processed.
2601 */
2602static void update_processed_blocks(struct reloc_control *rc,
2603 struct btrfs_backref_node *node)
2604{
2605 struct btrfs_backref_node *next = node;
2606 struct btrfs_backref_edge *edge;
2607 struct btrfs_backref_edge *edges[BTRFS_MAX_LEVEL - 1];
2608 int index = 0;
2609
2610 while (next) {
2611 cond_resched();
2612 while (1) {
2613 if (next->processed)
2614 break;
2615
2616 mark_block_processed(rc, next);
2617
2618 if (list_empty(&next->upper))
2619 break;
2620
2621 edge = list_entry(next->upper.next,
2622 struct btrfs_backref_edge, list[LOWER]);
2623 edges[index++] = edge;
2624 next = edge->node[UPPER];
2625 }
2626 next = walk_down_backref(edges, &index);
2627 }
2628}
2629
2630static int tree_block_processed(u64 bytenr, struct reloc_control *rc)
2631{
2632 u32 blocksize = rc->extent_root->fs_info->nodesize;
2633
2634 if (test_range_bit(&rc->processed_blocks, bytenr,
2635 bytenr + blocksize - 1, EXTENT_DIRTY, NULL))
2636 return 1;
2637 return 0;
2638}
2639
2640static int get_tree_block_key(struct btrfs_fs_info *fs_info,
2641 struct tree_block *block)
2642{
2643 struct btrfs_tree_parent_check check = {
2644 .level = block->level,
2645 .owner_root = block->owner,
2646 .transid = block->key.offset
2647 };
2648 struct extent_buffer *eb;
2649
2650 eb = read_tree_block(fs_info, block->bytenr, &check);
2651 if (IS_ERR(eb))
2652 return PTR_ERR(eb);
2653 if (!extent_buffer_uptodate(eb)) {
2654 free_extent_buffer(eb);
2655 return -EIO;
2656 }
2657 if (block->level == 0)
2658 btrfs_item_key_to_cpu(eb, &block->key, 0);
2659 else
2660 btrfs_node_key_to_cpu(eb, &block->key, 0);
2661 free_extent_buffer(eb);
2662 block->key_ready = true;
2663 return 0;
2664}
2665
2666/*
2667 * helper function to relocate a tree block
2668 */
2669static int relocate_tree_block(struct btrfs_trans_handle *trans,
2670 struct reloc_control *rc,
2671 struct btrfs_backref_node *node,
2672 struct btrfs_key *key,
2673 struct btrfs_path *path)
2674{
2675 struct btrfs_root *root;
2676 int ret = 0;
2677
2678 if (!node)
2679 return 0;
2680
2681 /*
2682 * If we fail here we want to drop our backref_node because we are going
2683 * to start over and regenerate the tree for it.
2684 */
2685 ret = reserve_metadata_space(trans, rc, node);
2686 if (ret)
2687 goto out;
2688
2689 BUG_ON(node->processed);
2690 root = select_one_root(node);
2691 if (IS_ERR(root)) {
2692 ret = PTR_ERR(root);
2693
2694 /* See explanation in select_one_root for the -EUCLEAN case. */
2695 ASSERT(ret == -ENOENT);
2696 if (ret == -ENOENT) {
2697 ret = 0;
2698 update_processed_blocks(rc, node);
2699 }
2700 goto out;
2701 }
2702
2703 if (root) {
2704 if (test_bit(BTRFS_ROOT_SHAREABLE, &root->state)) {
2705 /*
2706 * This block was the root block of a root, and this is
2707 * the first time we're processing the block and thus it
2708 * should not have had the ->new_bytenr modified and
2709 * should have not been included on the changed list.
2710 *
2711 * However in the case of corruption we could have
2712 * multiple refs pointing to the same block improperly,
2713 * and thus we would trip over these checks. ASSERT()
2714 * for the developer case, because it could indicate a
2715 * bug in the backref code, however error out for a
2716 * normal user in the case of corruption.
2717 */
2718 ASSERT(node->new_bytenr == 0);
2719 ASSERT(list_empty(&node->list));
2720 if (node->new_bytenr || !list_empty(&node->list)) {
2721 btrfs_err(root->fs_info,
2722 "bytenr %llu has improper references to it",
2723 node->bytenr);
2724 ret = -EUCLEAN;
2725 goto out;
2726 }
2727 ret = btrfs_record_root_in_trans(trans, root);
2728 if (ret)
2729 goto out;
2730 /*
2731 * Another thread could have failed, need to check if we
2732 * have reloc_root actually set.
2733 */
2734 if (!root->reloc_root) {
2735 ret = -ENOENT;
2736 goto out;
2737 }
2738 root = root->reloc_root;
2739 node->new_bytenr = root->node->start;
2740 btrfs_put_root(node->root);
2741 node->root = btrfs_grab_root(root);
2742 ASSERT(node->root);
2743 list_add_tail(&node->list, &rc->backref_cache.changed);
2744 } else {
2745 path->lowest_level = node->level;
2746 if (root == root->fs_info->chunk_root)
2747 btrfs_reserve_chunk_metadata(trans, false);
2748 ret = btrfs_search_slot(trans, root, key, path, 0, 1);
2749 btrfs_release_path(path);
2750 if (root == root->fs_info->chunk_root)
2751 btrfs_trans_release_chunk_metadata(trans);
2752 if (ret > 0)
2753 ret = 0;
2754 }
2755 if (!ret)
2756 update_processed_blocks(rc, node);
2757 } else {
2758 ret = do_relocation(trans, rc, node, key, path, 1);
2759 }
2760out:
2761 if (ret || node->level == 0 || node->cowonly)
2762 btrfs_backref_cleanup_node(&rc->backref_cache, node);
2763 return ret;
2764}
2765
2766/*
2767 * relocate a list of blocks
2768 */
2769static noinline_for_stack
2770int relocate_tree_blocks(struct btrfs_trans_handle *trans,
2771 struct reloc_control *rc, struct rb_root *blocks)
2772{
2773 struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
2774 struct btrfs_backref_node *node;
2775 struct btrfs_path *path;
2776 struct tree_block *block;
2777 struct tree_block *next;
2778 int ret;
2779 int err = 0;
2780
2781 path = btrfs_alloc_path();
2782 if (!path) {
2783 err = -ENOMEM;
2784 goto out_free_blocks;
2785 }
2786
2787 /* Kick in readahead for tree blocks with missing keys */
2788 rbtree_postorder_for_each_entry_safe(block, next, blocks, rb_node) {
2789 if (!block->key_ready)
2790 btrfs_readahead_tree_block(fs_info, block->bytenr,
2791 block->owner, 0,
2792 block->level);
2793 }
2794
2795 /* Get first keys */
2796 rbtree_postorder_for_each_entry_safe(block, next, blocks, rb_node) {
2797 if (!block->key_ready) {
2798 err = get_tree_block_key(fs_info, block);
2799 if (err)
2800 goto out_free_path;
2801 }
2802 }
2803
2804 /* Do tree relocation */
2805 rbtree_postorder_for_each_entry_safe(block, next, blocks, rb_node) {
2806 node = build_backref_tree(trans, rc, &block->key,
2807 block->level, block->bytenr);
2808 if (IS_ERR(node)) {
2809 err = PTR_ERR(node);
2810 goto out;
2811 }
2812
2813 ret = relocate_tree_block(trans, rc, node, &block->key,
2814 path);
2815 if (ret < 0) {
2816 err = ret;
2817 break;
2818 }
2819 }
2820out:
2821 err = finish_pending_nodes(trans, rc, path, err);
2822
2823out_free_path:
2824 btrfs_free_path(path);
2825out_free_blocks:
2826 free_block_list(blocks);
2827 return err;
2828}
2829
2830static noinline_for_stack int prealloc_file_extent_cluster(
2831 struct btrfs_inode *inode,
2832 const struct file_extent_cluster *cluster)
2833{
2834 u64 alloc_hint = 0;
2835 u64 start;
2836 u64 end;
2837 u64 offset = inode->index_cnt;
2838 u64 num_bytes;
2839 int nr;
2840 int ret = 0;
2841 u64 i_size = i_size_read(&inode->vfs_inode);
2842 u64 prealloc_start = cluster->start - offset;
2843 u64 prealloc_end = cluster->end - offset;
2844 u64 cur_offset = prealloc_start;
2845
2846 /*
2847 * For subpage case, previous i_size may not be aligned to PAGE_SIZE.
2848 * This means the range [i_size, PAGE_END + 1) is filled with zeros by
2849 * btrfs_do_readpage() call of previously relocated file cluster.
2850 *
2851 * If the current cluster starts in the above range, btrfs_do_readpage()
2852 * will skip the read, and relocate_one_page() will later writeback
2853 * the padding zeros as new data, causing data corruption.
2854 *
2855 * Here we have to manually invalidate the range (i_size, PAGE_END + 1).
2856 */
2857 if (!PAGE_ALIGNED(i_size)) {
2858 struct address_space *mapping = inode->vfs_inode.i_mapping;
2859 struct btrfs_fs_info *fs_info = inode->root->fs_info;
2860 const u32 sectorsize = fs_info->sectorsize;
2861 struct page *page;
2862
2863 ASSERT(sectorsize < PAGE_SIZE);
2864 ASSERT(IS_ALIGNED(i_size, sectorsize));
2865
2866 /*
2867 * Subpage can't handle page with DIRTY but without UPTODATE
2868 * bit as it can lead to the following deadlock:
2869 *
2870 * btrfs_read_folio()
2871 * | Page already *locked*
2872 * |- btrfs_lock_and_flush_ordered_range()
2873 * |- btrfs_start_ordered_extent()
2874 * |- extent_write_cache_pages()
2875 * |- lock_page()
2876 * We try to lock the page we already hold.
2877 *
2878 * Here we just writeback the whole data reloc inode, so that
2879 * we will be ensured to have no dirty range in the page, and
2880 * are safe to clear the uptodate bits.
2881 *
2882 * This shouldn't cause too much overhead, as we need to write
2883 * the data back anyway.
2884 */
2885 ret = filemap_write_and_wait(mapping);
2886 if (ret < 0)
2887 return ret;
2888
2889 clear_extent_bits(&inode->io_tree, i_size,
2890 round_up(i_size, PAGE_SIZE) - 1,
2891 EXTENT_UPTODATE);
2892 page = find_lock_page(mapping, i_size >> PAGE_SHIFT);
2893 /*
2894 * If page is freed we don't need to do anything then, as we
2895 * will re-read the whole page anyway.
2896 */
2897 if (page) {
2898 btrfs_subpage_clear_uptodate(fs_info, page_folio(page), i_size,
2899 round_up(i_size, PAGE_SIZE) - i_size);
2900 unlock_page(page);
2901 put_page(page);
2902 }
2903 }
2904
2905 BUG_ON(cluster->start != cluster->boundary[0]);
2906 ret = btrfs_alloc_data_chunk_ondemand(inode,
2907 prealloc_end + 1 - prealloc_start);
2908 if (ret)
2909 return ret;
2910
2911 btrfs_inode_lock(inode, 0);
2912 for (nr = 0; nr < cluster->nr; nr++) {
2913 struct extent_state *cached_state = NULL;
2914
2915 start = cluster->boundary[nr] - offset;
2916 if (nr + 1 < cluster->nr)
2917 end = cluster->boundary[nr + 1] - 1 - offset;
2918 else
2919 end = cluster->end - offset;
2920
2921 lock_extent(&inode->io_tree, start, end, &cached_state);
2922 num_bytes = end + 1 - start;
2923 ret = btrfs_prealloc_file_range(&inode->vfs_inode, 0, start,
2924 num_bytes, num_bytes,
2925 end + 1, &alloc_hint);
2926 cur_offset = end + 1;
2927 unlock_extent(&inode->io_tree, start, end, &cached_state);
2928 if (ret)
2929 break;
2930 }
2931 btrfs_inode_unlock(inode, 0);
2932
2933 if (cur_offset < prealloc_end)
2934 btrfs_free_reserved_data_space_noquota(inode->root->fs_info,
2935 prealloc_end + 1 - cur_offset);
2936 return ret;
2937}
2938
2939static noinline_for_stack int setup_relocation_extent_mapping(struct inode *inode,
2940 u64 start, u64 end, u64 block_start)
2941{
2942 struct extent_map *em;
2943 struct extent_state *cached_state = NULL;
2944 int ret = 0;
2945
2946 em = alloc_extent_map();
2947 if (!em)
2948 return -ENOMEM;
2949
2950 em->start = start;
2951 em->len = end + 1 - start;
2952 em->block_len = em->len;
2953 em->block_start = block_start;
2954 em->flags |= EXTENT_FLAG_PINNED;
2955
2956 lock_extent(&BTRFS_I(inode)->io_tree, start, end, &cached_state);
2957 ret = btrfs_replace_extent_map_range(BTRFS_I(inode), em, false);
2958 unlock_extent(&BTRFS_I(inode)->io_tree, start, end, &cached_state);
2959 free_extent_map(em);
2960
2961 return ret;
2962}
2963
2964/*
2965 * Allow error injection to test balance/relocation cancellation
2966 */
2967noinline int btrfs_should_cancel_balance(const struct btrfs_fs_info *fs_info)
2968{
2969 return atomic_read(&fs_info->balance_cancel_req) ||
2970 atomic_read(&fs_info->reloc_cancel_req) ||
2971 fatal_signal_pending(current);
2972}
2973ALLOW_ERROR_INJECTION(btrfs_should_cancel_balance, TRUE);
2974
2975static u64 get_cluster_boundary_end(const struct file_extent_cluster *cluster,
2976 int cluster_nr)
2977{
2978 /* Last extent, use cluster end directly */
2979 if (cluster_nr >= cluster->nr - 1)
2980 return cluster->end;
2981
2982 /* Use next boundary start*/
2983 return cluster->boundary[cluster_nr + 1] - 1;
2984}
2985
2986static int relocate_one_page(struct inode *inode, struct file_ra_state *ra,
2987 const struct file_extent_cluster *cluster,
2988 int *cluster_nr, unsigned long page_index)
2989{
2990 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2991 u64 offset = BTRFS_I(inode)->index_cnt;
2992 const unsigned long last_index = (cluster->end - offset) >> PAGE_SHIFT;
2993 gfp_t mask = btrfs_alloc_write_mask(inode->i_mapping);
2994 struct page *page;
2995 u64 page_start;
2996 u64 page_end;
2997 u64 cur;
2998 int ret;
2999
3000 ASSERT(page_index <= last_index);
3001 page = find_lock_page(inode->i_mapping, page_index);
3002 if (!page) {
3003 page_cache_sync_readahead(inode->i_mapping, ra, NULL,
3004 page_index, last_index + 1 - page_index);
3005 page = find_or_create_page(inode->i_mapping, page_index, mask);
3006 if (!page)
3007 return -ENOMEM;
3008 }
3009
3010 if (PageReadahead(page))
3011 page_cache_async_readahead(inode->i_mapping, ra, NULL,
3012 page_folio(page), page_index,
3013 last_index + 1 - page_index);
3014
3015 if (!PageUptodate(page)) {
3016 btrfs_read_folio(NULL, page_folio(page));
3017 lock_page(page);
3018 if (!PageUptodate(page)) {
3019 ret = -EIO;
3020 goto release_page;
3021 }
3022 }
3023
3024 /*
3025 * We could have lost page private when we dropped the lock to read the
3026 * page above, make sure we set_page_extent_mapped here so we have any
3027 * of the subpage blocksize stuff we need in place.
3028 */
3029 ret = set_page_extent_mapped(page);
3030 if (ret < 0)
3031 goto release_page;
3032
3033 page_start = page_offset(page);
3034 page_end = page_start + PAGE_SIZE - 1;
3035
3036 /*
3037 * Start from the cluster, as for subpage case, the cluster can start
3038 * inside the page.
3039 */
3040 cur = max(page_start, cluster->boundary[*cluster_nr] - offset);
3041 while (cur <= page_end) {
3042 struct extent_state *cached_state = NULL;
3043 u64 extent_start = cluster->boundary[*cluster_nr] - offset;
3044 u64 extent_end = get_cluster_boundary_end(cluster,
3045 *cluster_nr) - offset;
3046 u64 clamped_start = max(page_start, extent_start);
3047 u64 clamped_end = min(page_end, extent_end);
3048 u32 clamped_len = clamped_end + 1 - clamped_start;
3049
3050 /* Reserve metadata for this range */
3051 ret = btrfs_delalloc_reserve_metadata(BTRFS_I(inode),
3052 clamped_len, clamped_len,
3053 false);
3054 if (ret)
3055 goto release_page;
3056
3057 /* Mark the range delalloc and dirty for later writeback */
3058 lock_extent(&BTRFS_I(inode)->io_tree, clamped_start, clamped_end,
3059 &cached_state);
3060 ret = btrfs_set_extent_delalloc(BTRFS_I(inode), clamped_start,
3061 clamped_end, 0, &cached_state);
3062 if (ret) {
3063 clear_extent_bit(&BTRFS_I(inode)->io_tree,
3064 clamped_start, clamped_end,
3065 EXTENT_LOCKED | EXTENT_BOUNDARY,
3066 &cached_state);
3067 btrfs_delalloc_release_metadata(BTRFS_I(inode),
3068 clamped_len, true);
3069 btrfs_delalloc_release_extents(BTRFS_I(inode),
3070 clamped_len);
3071 goto release_page;
3072 }
3073 btrfs_folio_set_dirty(fs_info, page_folio(page),
3074 clamped_start, clamped_len);
3075
3076 /*
3077 * Set the boundary if it's inside the page.
3078 * Data relocation requires the destination extents to have the
3079 * same size as the source.
3080 * EXTENT_BOUNDARY bit prevents current extent from being merged
3081 * with previous extent.
3082 */
3083 if (in_range(cluster->boundary[*cluster_nr] - offset,
3084 page_start, PAGE_SIZE)) {
3085 u64 boundary_start = cluster->boundary[*cluster_nr] -
3086 offset;
3087 u64 boundary_end = boundary_start +
3088 fs_info->sectorsize - 1;
3089
3090 set_extent_bit(&BTRFS_I(inode)->io_tree,
3091 boundary_start, boundary_end,
3092 EXTENT_BOUNDARY, NULL);
3093 }
3094 unlock_extent(&BTRFS_I(inode)->io_tree, clamped_start, clamped_end,
3095 &cached_state);
3096 btrfs_delalloc_release_extents(BTRFS_I(inode), clamped_len);
3097 cur += clamped_len;
3098
3099 /* Crossed extent end, go to next extent */
3100 if (cur >= extent_end) {
3101 (*cluster_nr)++;
3102 /* Just finished the last extent of the cluster, exit. */
3103 if (*cluster_nr >= cluster->nr)
3104 break;
3105 }
3106 }
3107 unlock_page(page);
3108 put_page(page);
3109
3110 balance_dirty_pages_ratelimited(inode->i_mapping);
3111 btrfs_throttle(fs_info);
3112 if (btrfs_should_cancel_balance(fs_info))
3113 ret = -ECANCELED;
3114 return ret;
3115
3116release_page:
3117 unlock_page(page);
3118 put_page(page);
3119 return ret;
3120}
3121
3122static int relocate_file_extent_cluster(struct inode *inode,
3123 const struct file_extent_cluster *cluster)
3124{
3125 u64 offset = BTRFS_I(inode)->index_cnt;
3126 unsigned long index;
3127 unsigned long last_index;
3128 struct file_ra_state *ra;
3129 int cluster_nr = 0;
3130 int ret = 0;
3131
3132 if (!cluster->nr)
3133 return 0;
3134
3135 ra = kzalloc(sizeof(*ra), GFP_NOFS);
3136 if (!ra)
3137 return -ENOMEM;
3138
3139 ret = prealloc_file_extent_cluster(BTRFS_I(inode), cluster);
3140 if (ret)
3141 goto out;
3142
3143 file_ra_state_init(ra, inode->i_mapping);
3144
3145 ret = setup_relocation_extent_mapping(inode, cluster->start - offset,
3146 cluster->end - offset, cluster->start);
3147 if (ret)
3148 goto out;
3149
3150 last_index = (cluster->end - offset) >> PAGE_SHIFT;
3151 for (index = (cluster->start - offset) >> PAGE_SHIFT;
3152 index <= last_index && !ret; index++)
3153 ret = relocate_one_page(inode, ra, cluster, &cluster_nr, index);
3154 if (ret == 0)
3155 WARN_ON(cluster_nr != cluster->nr);
3156out:
3157 kfree(ra);
3158 return ret;
3159}
3160
3161static noinline_for_stack int relocate_data_extent(struct inode *inode,
3162 const struct btrfs_key *extent_key,
3163 struct file_extent_cluster *cluster)
3164{
3165 int ret;
3166 struct btrfs_root *root = BTRFS_I(inode)->root;
3167
3168 if (cluster->nr > 0 && extent_key->objectid != cluster->end + 1) {
3169 ret = relocate_file_extent_cluster(inode, cluster);
3170 if (ret)
3171 return ret;
3172 cluster->nr = 0;
3173 }
3174
3175 /*
3176 * Under simple quotas, we set root->relocation_src_root when we find
3177 * the extent. If adjacent extents have different owners, we can't merge
3178 * them while relocating. Handle this by storing the owning root that
3179 * started a cluster and if we see an extent from a different root break
3180 * cluster formation (just like the above case of non-adjacent extents).
3181 *
3182 * Without simple quotas, relocation_src_root is always 0, so we should
3183 * never see a mismatch, and it should have no effect on relocation
3184 * clusters.
3185 */
3186 if (cluster->nr > 0 && cluster->owning_root != root->relocation_src_root) {
3187 u64 tmp = root->relocation_src_root;
3188
3189 /*
3190 * root->relocation_src_root is the state that actually affects
3191 * the preallocation we do here, so set it to the root owning
3192 * the cluster we need to relocate.
3193 */
3194 root->relocation_src_root = cluster->owning_root;
3195 ret = relocate_file_extent_cluster(inode, cluster);
3196 if (ret)
3197 return ret;
3198 cluster->nr = 0;
3199 /* And reset it back for the current extent's owning root. */
3200 root->relocation_src_root = tmp;
3201 }
3202
3203 if (!cluster->nr) {
3204 cluster->start = extent_key->objectid;
3205 cluster->owning_root = root->relocation_src_root;
3206 }
3207 else
3208 BUG_ON(cluster->nr >= MAX_EXTENTS);
3209 cluster->end = extent_key->objectid + extent_key->offset - 1;
3210 cluster->boundary[cluster->nr] = extent_key->objectid;
3211 cluster->nr++;
3212
3213 if (cluster->nr >= MAX_EXTENTS) {
3214 ret = relocate_file_extent_cluster(inode, cluster);
3215 if (ret)
3216 return ret;
3217 cluster->nr = 0;
3218 }
3219 return 0;
3220}
3221
3222/*
3223 * helper to add a tree block to the list.
3224 * the major work is getting the generation and level of the block
3225 */
3226static int add_tree_block(struct reloc_control *rc,
3227 const struct btrfs_key *extent_key,
3228 struct btrfs_path *path,
3229 struct rb_root *blocks)
3230{
3231 struct extent_buffer *eb;
3232 struct btrfs_extent_item *ei;
3233 struct btrfs_tree_block_info *bi;
3234 struct tree_block *block;
3235 struct rb_node *rb_node;
3236 u32 item_size;
3237 int level = -1;
3238 u64 generation;
3239 u64 owner = 0;
3240
3241 eb = path->nodes[0];
3242 item_size = btrfs_item_size(eb, path->slots[0]);
3243
3244 if (extent_key->type == BTRFS_METADATA_ITEM_KEY ||
3245 item_size >= sizeof(*ei) + sizeof(*bi)) {
3246 unsigned long ptr = 0, end;
3247
3248 ei = btrfs_item_ptr(eb, path->slots[0],
3249 struct btrfs_extent_item);
3250 end = (unsigned long)ei + item_size;
3251 if (extent_key->type == BTRFS_EXTENT_ITEM_KEY) {
3252 bi = (struct btrfs_tree_block_info *)(ei + 1);
3253 level = btrfs_tree_block_level(eb, bi);
3254 ptr = (unsigned long)(bi + 1);
3255 } else {
3256 level = (int)extent_key->offset;
3257 ptr = (unsigned long)(ei + 1);
3258 }
3259 generation = btrfs_extent_generation(eb, ei);
3260
3261 /*
3262 * We're reading random blocks without knowing their owner ahead
3263 * of time. This is ok most of the time, as all reloc roots and
3264 * fs roots have the same lock type. However normal trees do
3265 * not, and the only way to know ahead of time is to read the
3266 * inline ref offset. We know it's an fs root if
3267 *
3268 * 1. There's more than one ref.
3269 * 2. There's a SHARED_DATA_REF_KEY set.
3270 * 3. FULL_BACKREF is set on the flags.
3271 *
3272 * Otherwise it's safe to assume that the ref offset == the
3273 * owner of this block, so we can use that when calling
3274 * read_tree_block.
3275 */
3276 if (btrfs_extent_refs(eb, ei) == 1 &&
3277 !(btrfs_extent_flags(eb, ei) &
3278 BTRFS_BLOCK_FLAG_FULL_BACKREF) &&
3279 ptr < end) {
3280 struct btrfs_extent_inline_ref *iref;
3281 int type;
3282
3283 iref = (struct btrfs_extent_inline_ref *)ptr;
3284 type = btrfs_get_extent_inline_ref_type(eb, iref,
3285 BTRFS_REF_TYPE_BLOCK);
3286 if (type == BTRFS_REF_TYPE_INVALID)
3287 return -EINVAL;
3288 if (type == BTRFS_TREE_BLOCK_REF_KEY)
3289 owner = btrfs_extent_inline_ref_offset(eb, iref);
3290 }
3291 } else {
3292 btrfs_print_leaf(eb);
3293 btrfs_err(rc->block_group->fs_info,
3294 "unrecognized tree backref at tree block %llu slot %u",
3295 eb->start, path->slots[0]);
3296 btrfs_release_path(path);
3297 return -EUCLEAN;
3298 }
3299
3300 btrfs_release_path(path);
3301
3302 BUG_ON(level == -1);
3303
3304 block = kmalloc(sizeof(*block), GFP_NOFS);
3305 if (!block)
3306 return -ENOMEM;
3307
3308 block->bytenr = extent_key->objectid;
3309 block->key.objectid = rc->extent_root->fs_info->nodesize;
3310 block->key.offset = generation;
3311 block->level = level;
3312 block->key_ready = false;
3313 block->owner = owner;
3314
3315 rb_node = rb_simple_insert(blocks, block->bytenr, &block->rb_node);
3316 if (rb_node)
3317 btrfs_backref_panic(rc->extent_root->fs_info, block->bytenr,
3318 -EEXIST);
3319
3320 return 0;
3321}
3322
3323/*
3324 * helper to add tree blocks for backref of type BTRFS_SHARED_DATA_REF_KEY
3325 */
3326static int __add_tree_block(struct reloc_control *rc,
3327 u64 bytenr, u32 blocksize,
3328 struct rb_root *blocks)
3329{
3330 struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
3331 struct btrfs_path *path;
3332 struct btrfs_key key;
3333 int ret;
3334 bool skinny = btrfs_fs_incompat(fs_info, SKINNY_METADATA);
3335
3336 if (tree_block_processed(bytenr, rc))
3337 return 0;
3338
3339 if (rb_simple_search(blocks, bytenr))
3340 return 0;
3341
3342 path = btrfs_alloc_path();
3343 if (!path)
3344 return -ENOMEM;
3345again:
3346 key.objectid = bytenr;
3347 if (skinny) {
3348 key.type = BTRFS_METADATA_ITEM_KEY;
3349 key.offset = (u64)-1;
3350 } else {
3351 key.type = BTRFS_EXTENT_ITEM_KEY;
3352 key.offset = blocksize;
3353 }
3354
3355 path->search_commit_root = 1;
3356 path->skip_locking = 1;
3357 ret = btrfs_search_slot(NULL, rc->extent_root, &key, path, 0, 0);
3358 if (ret < 0)
3359 goto out;
3360
3361 if (ret > 0 && skinny) {
3362 if (path->slots[0]) {
3363 path->slots[0]--;
3364 btrfs_item_key_to_cpu(path->nodes[0], &key,
3365 path->slots[0]);
3366 if (key.objectid == bytenr &&
3367 (key.type == BTRFS_METADATA_ITEM_KEY ||
3368 (key.type == BTRFS_EXTENT_ITEM_KEY &&
3369 key.offset == blocksize)))
3370 ret = 0;
3371 }
3372
3373 if (ret) {
3374 skinny = false;
3375 btrfs_release_path(path);
3376 goto again;
3377 }
3378 }
3379 if (ret) {
3380 ASSERT(ret == 1);
3381 btrfs_print_leaf(path->nodes[0]);
3382 btrfs_err(fs_info,
3383 "tree block extent item (%llu) is not found in extent tree",
3384 bytenr);
3385 WARN_ON(1);
3386 ret = -EINVAL;
3387 goto out;
3388 }
3389
3390 ret = add_tree_block(rc, &key, path, blocks);
3391out:
3392 btrfs_free_path(path);
3393 return ret;
3394}
3395
3396static int delete_block_group_cache(struct btrfs_fs_info *fs_info,
3397 struct btrfs_block_group *block_group,
3398 struct inode *inode,
3399 u64 ino)
3400{
3401 struct btrfs_root *root = fs_info->tree_root;
3402 struct btrfs_trans_handle *trans;
3403 int ret = 0;
3404
3405 if (inode)
3406 goto truncate;
3407
3408 inode = btrfs_iget(fs_info->sb, ino, root);
3409 if (IS_ERR(inode))
3410 return -ENOENT;
3411
3412truncate:
3413 ret = btrfs_check_trunc_cache_free_space(fs_info,
3414 &fs_info->global_block_rsv);
3415 if (ret)
3416 goto out;
3417
3418 trans = btrfs_join_transaction(root);
3419 if (IS_ERR(trans)) {
3420 ret = PTR_ERR(trans);
3421 goto out;
3422 }
3423
3424 ret = btrfs_truncate_free_space_cache(trans, block_group, inode);
3425
3426 btrfs_end_transaction(trans);
3427 btrfs_btree_balance_dirty(fs_info);
3428out:
3429 iput(inode);
3430 return ret;
3431}
3432
3433/*
3434 * Locate the free space cache EXTENT_DATA in root tree leaf and delete the
3435 * cache inode, to avoid free space cache data extent blocking data relocation.
3436 */
3437static int delete_v1_space_cache(struct extent_buffer *leaf,
3438 struct btrfs_block_group *block_group,
3439 u64 data_bytenr)
3440{
3441 u64 space_cache_ino;
3442 struct btrfs_file_extent_item *ei;
3443 struct btrfs_key key;
3444 bool found = false;
3445 int i;
3446 int ret;
3447
3448 if (btrfs_header_owner(leaf) != BTRFS_ROOT_TREE_OBJECTID)
3449 return 0;
3450
3451 for (i = 0; i < btrfs_header_nritems(leaf); i++) {
3452 u8 type;
3453
3454 btrfs_item_key_to_cpu(leaf, &key, i);
3455 if (key.type != BTRFS_EXTENT_DATA_KEY)
3456 continue;
3457 ei = btrfs_item_ptr(leaf, i, struct btrfs_file_extent_item);
3458 type = btrfs_file_extent_type(leaf, ei);
3459
3460 if ((type == BTRFS_FILE_EXTENT_REG ||
3461 type == BTRFS_FILE_EXTENT_PREALLOC) &&
3462 btrfs_file_extent_disk_bytenr(leaf, ei) == data_bytenr) {
3463 found = true;
3464 space_cache_ino = key.objectid;
3465 break;
3466 }
3467 }
3468 if (!found)
3469 return -ENOENT;
3470 ret = delete_block_group_cache(leaf->fs_info, block_group, NULL,
3471 space_cache_ino);
3472 return ret;
3473}
3474
3475/*
3476 * helper to find all tree blocks that reference a given data extent
3477 */
3478static noinline_for_stack int add_data_references(struct reloc_control *rc,
3479 const struct btrfs_key *extent_key,
3480 struct btrfs_path *path,
3481 struct rb_root *blocks)
3482{
3483 struct btrfs_backref_walk_ctx ctx = { 0 };
3484 struct ulist_iterator leaf_uiter;
3485 struct ulist_node *ref_node = NULL;
3486 const u32 blocksize = rc->extent_root->fs_info->nodesize;
3487 int ret = 0;
3488
3489 btrfs_release_path(path);
3490
3491 ctx.bytenr = extent_key->objectid;
3492 ctx.skip_inode_ref_list = true;
3493 ctx.fs_info = rc->extent_root->fs_info;
3494
3495 ret = btrfs_find_all_leafs(&ctx);
3496 if (ret < 0)
3497 return ret;
3498
3499 ULIST_ITER_INIT(&leaf_uiter);
3500 while ((ref_node = ulist_next(ctx.refs, &leaf_uiter))) {
3501 struct btrfs_tree_parent_check check = { 0 };
3502 struct extent_buffer *eb;
3503
3504 eb = read_tree_block(ctx.fs_info, ref_node->val, &check);
3505 if (IS_ERR(eb)) {
3506 ret = PTR_ERR(eb);
3507 break;
3508 }
3509 ret = delete_v1_space_cache(eb, rc->block_group,
3510 extent_key->objectid);
3511 free_extent_buffer(eb);
3512 if (ret < 0)
3513 break;
3514 ret = __add_tree_block(rc, ref_node->val, blocksize, blocks);
3515 if (ret < 0)
3516 break;
3517 }
3518 if (ret < 0)
3519 free_block_list(blocks);
3520 ulist_free(ctx.refs);
3521 return ret;
3522}
3523
3524/*
3525 * helper to find next unprocessed extent
3526 */
3527static noinline_for_stack
3528int find_next_extent(struct reloc_control *rc, struct btrfs_path *path,
3529 struct btrfs_key *extent_key)
3530{
3531 struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
3532 struct btrfs_key key;
3533 struct extent_buffer *leaf;
3534 u64 start, end, last;
3535 int ret;
3536
3537 last = rc->block_group->start + rc->block_group->length;
3538 while (1) {
3539 bool block_found;
3540
3541 cond_resched();
3542 if (rc->search_start >= last) {
3543 ret = 1;
3544 break;
3545 }
3546
3547 key.objectid = rc->search_start;
3548 key.type = BTRFS_EXTENT_ITEM_KEY;
3549 key.offset = 0;
3550
3551 path->search_commit_root = 1;
3552 path->skip_locking = 1;
3553 ret = btrfs_search_slot(NULL, rc->extent_root, &key, path,
3554 0, 0);
3555 if (ret < 0)
3556 break;
3557next:
3558 leaf = path->nodes[0];
3559 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
3560 ret = btrfs_next_leaf(rc->extent_root, path);
3561 if (ret != 0)
3562 break;
3563 leaf = path->nodes[0];
3564 }
3565
3566 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
3567 if (key.objectid >= last) {
3568 ret = 1;
3569 break;
3570 }
3571
3572 if (key.type != BTRFS_EXTENT_ITEM_KEY &&
3573 key.type != BTRFS_METADATA_ITEM_KEY) {
3574 path->slots[0]++;
3575 goto next;
3576 }
3577
3578 if (key.type == BTRFS_EXTENT_ITEM_KEY &&
3579 key.objectid + key.offset <= rc->search_start) {
3580 path->slots[0]++;
3581 goto next;
3582 }
3583
3584 if (key.type == BTRFS_METADATA_ITEM_KEY &&
3585 key.objectid + fs_info->nodesize <=
3586 rc->search_start) {
3587 path->slots[0]++;
3588 goto next;
3589 }
3590
3591 block_found = find_first_extent_bit(&rc->processed_blocks,
3592 key.objectid, &start, &end,
3593 EXTENT_DIRTY, NULL);
3594
3595 if (block_found && start <= key.objectid) {
3596 btrfs_release_path(path);
3597 rc->search_start = end + 1;
3598 } else {
3599 if (key.type == BTRFS_EXTENT_ITEM_KEY)
3600 rc->search_start = key.objectid + key.offset;
3601 else
3602 rc->search_start = key.objectid +
3603 fs_info->nodesize;
3604 memcpy(extent_key, &key, sizeof(key));
3605 return 0;
3606 }
3607 }
3608 btrfs_release_path(path);
3609 return ret;
3610}
3611
3612static void set_reloc_control(struct reloc_control *rc)
3613{
3614 struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
3615
3616 mutex_lock(&fs_info->reloc_mutex);
3617 fs_info->reloc_ctl = rc;
3618 mutex_unlock(&fs_info->reloc_mutex);
3619}
3620
3621static void unset_reloc_control(struct reloc_control *rc)
3622{
3623 struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
3624
3625 mutex_lock(&fs_info->reloc_mutex);
3626 fs_info->reloc_ctl = NULL;
3627 mutex_unlock(&fs_info->reloc_mutex);
3628}
3629
3630static noinline_for_stack
3631int prepare_to_relocate(struct reloc_control *rc)
3632{
3633 struct btrfs_trans_handle *trans;
3634 int ret;
3635
3636 rc->block_rsv = btrfs_alloc_block_rsv(rc->extent_root->fs_info,
3637 BTRFS_BLOCK_RSV_TEMP);
3638 if (!rc->block_rsv)
3639 return -ENOMEM;
3640
3641 memset(&rc->cluster, 0, sizeof(rc->cluster));
3642 rc->search_start = rc->block_group->start;
3643 rc->extents_found = 0;
3644 rc->nodes_relocated = 0;
3645 rc->merging_rsv_size = 0;
3646 rc->reserved_bytes = 0;
3647 rc->block_rsv->size = rc->extent_root->fs_info->nodesize *
3648 RELOCATION_RESERVED_NODES;
3649 ret = btrfs_block_rsv_refill(rc->extent_root->fs_info,
3650 rc->block_rsv, rc->block_rsv->size,
3651 BTRFS_RESERVE_FLUSH_ALL);
3652 if (ret)
3653 return ret;
3654
3655 rc->create_reloc_tree = true;
3656 set_reloc_control(rc);
3657
3658 trans = btrfs_join_transaction(rc->extent_root);
3659 if (IS_ERR(trans)) {
3660 unset_reloc_control(rc);
3661 /*
3662 * extent tree is not a ref_cow tree and has no reloc_root to
3663 * cleanup. And callers are responsible to free the above
3664 * block rsv.
3665 */
3666 return PTR_ERR(trans);
3667 }
3668
3669 ret = btrfs_commit_transaction(trans);
3670 if (ret)
3671 unset_reloc_control(rc);
3672
3673 return ret;
3674}
3675
3676static noinline_for_stack int relocate_block_group(struct reloc_control *rc)
3677{
3678 struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
3679 struct rb_root blocks = RB_ROOT;
3680 struct btrfs_key key;
3681 struct btrfs_trans_handle *trans = NULL;
3682 struct btrfs_path *path;
3683 struct btrfs_extent_item *ei;
3684 u64 flags;
3685 int ret;
3686 int err = 0;
3687 int progress = 0;
3688
3689 path = btrfs_alloc_path();
3690 if (!path)
3691 return -ENOMEM;
3692 path->reada = READA_FORWARD;
3693
3694 ret = prepare_to_relocate(rc);
3695 if (ret) {
3696 err = ret;
3697 goto out_free;
3698 }
3699
3700 while (1) {
3701 rc->reserved_bytes = 0;
3702 ret = btrfs_block_rsv_refill(fs_info, rc->block_rsv,
3703 rc->block_rsv->size,
3704 BTRFS_RESERVE_FLUSH_ALL);
3705 if (ret) {
3706 err = ret;
3707 break;
3708 }
3709 progress++;
3710 trans = btrfs_start_transaction(rc->extent_root, 0);
3711 if (IS_ERR(trans)) {
3712 err = PTR_ERR(trans);
3713 trans = NULL;
3714 break;
3715 }
3716restart:
3717 if (update_backref_cache(trans, &rc->backref_cache)) {
3718 btrfs_end_transaction(trans);
3719 trans = NULL;
3720 continue;
3721 }
3722
3723 ret = find_next_extent(rc, path, &key);
3724 if (ret < 0)
3725 err = ret;
3726 if (ret != 0)
3727 break;
3728
3729 rc->extents_found++;
3730
3731 ei = btrfs_item_ptr(path->nodes[0], path->slots[0],
3732 struct btrfs_extent_item);
3733 flags = btrfs_extent_flags(path->nodes[0], ei);
3734
3735 /*
3736 * If we are relocating a simple quota owned extent item, we
3737 * need to note the owner on the reloc data root so that when
3738 * we allocate the replacement item, we can attribute it to the
3739 * correct eventual owner (rather than the reloc data root).
3740 */
3741 if (btrfs_qgroup_mode(fs_info) == BTRFS_QGROUP_MODE_SIMPLE) {
3742 struct btrfs_root *root = BTRFS_I(rc->data_inode)->root;
3743 u64 owning_root_id = btrfs_get_extent_owner_root(fs_info,
3744 path->nodes[0],
3745 path->slots[0]);
3746
3747 root->relocation_src_root = owning_root_id;
3748 }
3749
3750 if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
3751 ret = add_tree_block(rc, &key, path, &blocks);
3752 } else if (rc->stage == UPDATE_DATA_PTRS &&
3753 (flags & BTRFS_EXTENT_FLAG_DATA)) {
3754 ret = add_data_references(rc, &key, path, &blocks);
3755 } else {
3756 btrfs_release_path(path);
3757 ret = 0;
3758 }
3759 if (ret < 0) {
3760 err = ret;
3761 break;
3762 }
3763
3764 if (!RB_EMPTY_ROOT(&blocks)) {
3765 ret = relocate_tree_blocks(trans, rc, &blocks);
3766 if (ret < 0) {
3767 if (ret != -EAGAIN) {
3768 err = ret;
3769 break;
3770 }
3771 rc->extents_found--;
3772 rc->search_start = key.objectid;
3773 }
3774 }
3775
3776 btrfs_end_transaction_throttle(trans);
3777 btrfs_btree_balance_dirty(fs_info);
3778 trans = NULL;
3779
3780 if (rc->stage == MOVE_DATA_EXTENTS &&
3781 (flags & BTRFS_EXTENT_FLAG_DATA)) {
3782 rc->found_file_extent = true;
3783 ret = relocate_data_extent(rc->data_inode,
3784 &key, &rc->cluster);
3785 if (ret < 0) {
3786 err = ret;
3787 break;
3788 }
3789 }
3790 if (btrfs_should_cancel_balance(fs_info)) {
3791 err = -ECANCELED;
3792 break;
3793 }
3794 }
3795 if (trans && progress && err == -ENOSPC) {
3796 ret = btrfs_force_chunk_alloc(trans, rc->block_group->flags);
3797 if (ret == 1) {
3798 err = 0;
3799 progress = 0;
3800 goto restart;
3801 }
3802 }
3803
3804 btrfs_release_path(path);
3805 clear_extent_bits(&rc->processed_blocks, 0, (u64)-1, EXTENT_DIRTY);
3806
3807 if (trans) {
3808 btrfs_end_transaction_throttle(trans);
3809 btrfs_btree_balance_dirty(fs_info);
3810 }
3811
3812 if (!err) {
3813 ret = relocate_file_extent_cluster(rc->data_inode,
3814 &rc->cluster);
3815 if (ret < 0)
3816 err = ret;
3817 }
3818
3819 rc->create_reloc_tree = false;
3820 set_reloc_control(rc);
3821
3822 btrfs_backref_release_cache(&rc->backref_cache);
3823 btrfs_block_rsv_release(fs_info, rc->block_rsv, (u64)-1, NULL);
3824
3825 /*
3826 * Even in the case when the relocation is cancelled, we should all go
3827 * through prepare_to_merge() and merge_reloc_roots().
3828 *
3829 * For error (including cancelled balance), prepare_to_merge() will
3830 * mark all reloc trees orphan, then queue them for cleanup in
3831 * merge_reloc_roots()
3832 */
3833 err = prepare_to_merge(rc, err);
3834
3835 merge_reloc_roots(rc);
3836
3837 rc->merge_reloc_tree = false;
3838 unset_reloc_control(rc);
3839 btrfs_block_rsv_release(fs_info, rc->block_rsv, (u64)-1, NULL);
3840
3841 /* get rid of pinned extents */
3842 trans = btrfs_join_transaction(rc->extent_root);
3843 if (IS_ERR(trans)) {
3844 err = PTR_ERR(trans);
3845 goto out_free;
3846 }
3847 ret = btrfs_commit_transaction(trans);
3848 if (ret && !err)
3849 err = ret;
3850out_free:
3851 ret = clean_dirty_subvols(rc);
3852 if (ret < 0 && !err)
3853 err = ret;
3854 btrfs_free_block_rsv(fs_info, rc->block_rsv);
3855 btrfs_free_path(path);
3856 return err;
3857}
3858
3859static int __insert_orphan_inode(struct btrfs_trans_handle *trans,
3860 struct btrfs_root *root, u64 objectid)
3861{
3862 struct btrfs_path *path;
3863 struct btrfs_inode_item *item;
3864 struct extent_buffer *leaf;
3865 int ret;
3866
3867 path = btrfs_alloc_path();
3868 if (!path)
3869 return -ENOMEM;
3870
3871 ret = btrfs_insert_empty_inode(trans, root, path, objectid);
3872 if (ret)
3873 goto out;
3874
3875 leaf = path->nodes[0];
3876 item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_inode_item);
3877 memzero_extent_buffer(leaf, (unsigned long)item, sizeof(*item));
3878 btrfs_set_inode_generation(leaf, item, 1);
3879 btrfs_set_inode_size(leaf, item, 0);
3880 btrfs_set_inode_mode(leaf, item, S_IFREG | 0600);
3881 btrfs_set_inode_flags(leaf, item, BTRFS_INODE_NOCOMPRESS |
3882 BTRFS_INODE_PREALLOC);
3883 btrfs_mark_buffer_dirty(trans, leaf);
3884out:
3885 btrfs_free_path(path);
3886 return ret;
3887}
3888
3889static void delete_orphan_inode(struct btrfs_trans_handle *trans,
3890 struct btrfs_root *root, u64 objectid)
3891{
3892 struct btrfs_path *path;
3893 struct btrfs_key key;
3894 int ret = 0;
3895
3896 path = btrfs_alloc_path();
3897 if (!path) {
3898 ret = -ENOMEM;
3899 goto out;
3900 }
3901
3902 key.objectid = objectid;
3903 key.type = BTRFS_INODE_ITEM_KEY;
3904 key.offset = 0;
3905 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
3906 if (ret) {
3907 if (ret > 0)
3908 ret = -ENOENT;
3909 goto out;
3910 }
3911 ret = btrfs_del_item(trans, root, path);
3912out:
3913 if (ret)
3914 btrfs_abort_transaction(trans, ret);
3915 btrfs_free_path(path);
3916}
3917
3918/*
3919 * helper to create inode for data relocation.
3920 * the inode is in data relocation tree and its link count is 0
3921 */
3922static noinline_for_stack struct inode *create_reloc_inode(
3923 struct btrfs_fs_info *fs_info,
3924 const struct btrfs_block_group *group)
3925{
3926 struct inode *inode = NULL;
3927 struct btrfs_trans_handle *trans;
3928 struct btrfs_root *root;
3929 u64 objectid;
3930 int err = 0;
3931
3932 root = btrfs_grab_root(fs_info->data_reloc_root);
3933 trans = btrfs_start_transaction(root, 6);
3934 if (IS_ERR(trans)) {
3935 btrfs_put_root(root);
3936 return ERR_CAST(trans);
3937 }
3938
3939 err = btrfs_get_free_objectid(root, &objectid);
3940 if (err)
3941 goto out;
3942
3943 err = __insert_orphan_inode(trans, root, objectid);
3944 if (err)
3945 goto out;
3946
3947 inode = btrfs_iget(fs_info->sb, objectid, root);
3948 if (IS_ERR(inode)) {
3949 delete_orphan_inode(trans, root, objectid);
3950 err = PTR_ERR(inode);
3951 inode = NULL;
3952 goto out;
3953 }
3954 BTRFS_I(inode)->index_cnt = group->start;
3955
3956 err = btrfs_orphan_add(trans, BTRFS_I(inode));
3957out:
3958 btrfs_put_root(root);
3959 btrfs_end_transaction(trans);
3960 btrfs_btree_balance_dirty(fs_info);
3961 if (err) {
3962 iput(inode);
3963 inode = ERR_PTR(err);
3964 }
3965 return inode;
3966}
3967
3968/*
3969 * Mark start of chunk relocation that is cancellable. Check if the cancellation
3970 * has been requested meanwhile and don't start in that case.
3971 *
3972 * Return:
3973 * 0 success
3974 * -EINPROGRESS operation is already in progress, that's probably a bug
3975 * -ECANCELED cancellation request was set before the operation started
3976 */
3977static int reloc_chunk_start(struct btrfs_fs_info *fs_info)
3978{
3979 if (test_and_set_bit(BTRFS_FS_RELOC_RUNNING, &fs_info->flags)) {
3980 /* This should not happen */
3981 btrfs_err(fs_info, "reloc already running, cannot start");
3982 return -EINPROGRESS;
3983 }
3984
3985 if (atomic_read(&fs_info->reloc_cancel_req) > 0) {
3986 btrfs_info(fs_info, "chunk relocation canceled on start");
3987 /*
3988 * On cancel, clear all requests but let the caller mark
3989 * the end after cleanup operations.
3990 */
3991 atomic_set(&fs_info->reloc_cancel_req, 0);
3992 return -ECANCELED;
3993 }
3994 return 0;
3995}
3996
3997/*
3998 * Mark end of chunk relocation that is cancellable and wake any waiters.
3999 */
4000static void reloc_chunk_end(struct btrfs_fs_info *fs_info)
4001{
4002 /* Requested after start, clear bit first so any waiters can continue */
4003 if (atomic_read(&fs_info->reloc_cancel_req) > 0)
4004 btrfs_info(fs_info, "chunk relocation canceled during operation");
4005 clear_and_wake_up_bit(BTRFS_FS_RELOC_RUNNING, &fs_info->flags);
4006 atomic_set(&fs_info->reloc_cancel_req, 0);
4007}
4008
4009static struct reloc_control *alloc_reloc_control(struct btrfs_fs_info *fs_info)
4010{
4011 struct reloc_control *rc;
4012
4013 rc = kzalloc(sizeof(*rc), GFP_NOFS);
4014 if (!rc)
4015 return NULL;
4016
4017 INIT_LIST_HEAD(&rc->reloc_roots);
4018 INIT_LIST_HEAD(&rc->dirty_subvol_roots);
4019 btrfs_backref_init_cache(fs_info, &rc->backref_cache, true);
4020 rc->reloc_root_tree.rb_root = RB_ROOT;
4021 spin_lock_init(&rc->reloc_root_tree.lock);
4022 extent_io_tree_init(fs_info, &rc->processed_blocks, IO_TREE_RELOC_BLOCKS);
4023 return rc;
4024}
4025
4026static void free_reloc_control(struct reloc_control *rc)
4027{
4028 struct mapping_node *node, *tmp;
4029
4030 free_reloc_roots(&rc->reloc_roots);
4031 rbtree_postorder_for_each_entry_safe(node, tmp,
4032 &rc->reloc_root_tree.rb_root, rb_node)
4033 kfree(node);
4034
4035 kfree(rc);
4036}
4037
4038/*
4039 * Print the block group being relocated
4040 */
4041static void describe_relocation(struct btrfs_fs_info *fs_info,
4042 struct btrfs_block_group *block_group)
4043{
4044 char buf[128] = {'\0'};
4045
4046 btrfs_describe_block_groups(block_group->flags, buf, sizeof(buf));
4047
4048 btrfs_info(fs_info,
4049 "relocating block group %llu flags %s",
4050 block_group->start, buf);
4051}
4052
4053static const char *stage_to_string(enum reloc_stage stage)
4054{
4055 if (stage == MOVE_DATA_EXTENTS)
4056 return "move data extents";
4057 if (stage == UPDATE_DATA_PTRS)
4058 return "update data pointers";
4059 return "unknown";
4060}
4061
4062/*
4063 * function to relocate all extents in a block group.
4064 */
4065int btrfs_relocate_block_group(struct btrfs_fs_info *fs_info, u64 group_start)
4066{
4067 struct btrfs_block_group *bg;
4068 struct btrfs_root *extent_root = btrfs_extent_root(fs_info, group_start);
4069 struct reloc_control *rc;
4070 struct inode *inode;
4071 struct btrfs_path *path;
4072 int ret;
4073 int rw = 0;
4074 int err = 0;
4075
4076 /*
4077 * This only gets set if we had a half-deleted snapshot on mount. We
4078 * cannot allow relocation to start while we're still trying to clean up
4079 * these pending deletions.
4080 */
4081 ret = wait_on_bit(&fs_info->flags, BTRFS_FS_UNFINISHED_DROPS, TASK_INTERRUPTIBLE);
4082 if (ret)
4083 return ret;
4084
4085 /* We may have been woken up by close_ctree, so bail if we're closing. */
4086 if (btrfs_fs_closing(fs_info))
4087 return -EINTR;
4088
4089 bg = btrfs_lookup_block_group(fs_info, group_start);
4090 if (!bg)
4091 return -ENOENT;
4092
4093 /*
4094 * Relocation of a data block group creates ordered extents. Without
4095 * sb_start_write(), we can freeze the filesystem while unfinished
4096 * ordered extents are left. Such ordered extents can cause a deadlock
4097 * e.g. when syncfs() is waiting for their completion but they can't
4098 * finish because they block when joining a transaction, due to the
4099 * fact that the freeze locks are being held in write mode.
4100 */
4101 if (bg->flags & BTRFS_BLOCK_GROUP_DATA)
4102 ASSERT(sb_write_started(fs_info->sb));
4103
4104 if (btrfs_pinned_by_swapfile(fs_info, bg)) {
4105 btrfs_put_block_group(bg);
4106 return -ETXTBSY;
4107 }
4108
4109 rc = alloc_reloc_control(fs_info);
4110 if (!rc) {
4111 btrfs_put_block_group(bg);
4112 return -ENOMEM;
4113 }
4114
4115 ret = reloc_chunk_start(fs_info);
4116 if (ret < 0) {
4117 err = ret;
4118 goto out_put_bg;
4119 }
4120
4121 rc->extent_root = extent_root;
4122 rc->block_group = bg;
4123
4124 ret = btrfs_inc_block_group_ro(rc->block_group, true);
4125 if (ret) {
4126 err = ret;
4127 goto out;
4128 }
4129 rw = 1;
4130
4131 path = btrfs_alloc_path();
4132 if (!path) {
4133 err = -ENOMEM;
4134 goto out;
4135 }
4136
4137 inode = lookup_free_space_inode(rc->block_group, path);
4138 btrfs_free_path(path);
4139
4140 if (!IS_ERR(inode))
4141 ret = delete_block_group_cache(fs_info, rc->block_group, inode, 0);
4142 else
4143 ret = PTR_ERR(inode);
4144
4145 if (ret && ret != -ENOENT) {
4146 err = ret;
4147 goto out;
4148 }
4149
4150 rc->data_inode = create_reloc_inode(fs_info, rc->block_group);
4151 if (IS_ERR(rc->data_inode)) {
4152 err = PTR_ERR(rc->data_inode);
4153 rc->data_inode = NULL;
4154 goto out;
4155 }
4156
4157 describe_relocation(fs_info, rc->block_group);
4158
4159 btrfs_wait_block_group_reservations(rc->block_group);
4160 btrfs_wait_nocow_writers(rc->block_group);
4161 btrfs_wait_ordered_roots(fs_info, U64_MAX,
4162 rc->block_group->start,
4163 rc->block_group->length);
4164
4165 ret = btrfs_zone_finish(rc->block_group);
4166 WARN_ON(ret && ret != -EAGAIN);
4167
4168 while (1) {
4169 enum reloc_stage finishes_stage;
4170
4171 mutex_lock(&fs_info->cleaner_mutex);
4172 ret = relocate_block_group(rc);
4173 mutex_unlock(&fs_info->cleaner_mutex);
4174 if (ret < 0)
4175 err = ret;
4176
4177 finishes_stage = rc->stage;
4178 /*
4179 * We may have gotten ENOSPC after we already dirtied some
4180 * extents. If writeout happens while we're relocating a
4181 * different block group we could end up hitting the
4182 * BUG_ON(rc->stage == UPDATE_DATA_PTRS) in
4183 * btrfs_reloc_cow_block. Make sure we write everything out
4184 * properly so we don't trip over this problem, and then break
4185 * out of the loop if we hit an error.
4186 */
4187 if (rc->stage == MOVE_DATA_EXTENTS && rc->found_file_extent) {
4188 ret = btrfs_wait_ordered_range(rc->data_inode, 0,
4189 (u64)-1);
4190 if (ret)
4191 err = ret;
4192 invalidate_mapping_pages(rc->data_inode->i_mapping,
4193 0, -1);
4194 rc->stage = UPDATE_DATA_PTRS;
4195 }
4196
4197 if (err < 0)
4198 goto out;
4199
4200 if (rc->extents_found == 0)
4201 break;
4202
4203 btrfs_info(fs_info, "found %llu extents, stage: %s",
4204 rc->extents_found, stage_to_string(finishes_stage));
4205 }
4206
4207 WARN_ON(rc->block_group->pinned > 0);
4208 WARN_ON(rc->block_group->reserved > 0);
4209 WARN_ON(rc->block_group->used > 0);
4210out:
4211 if (err && rw)
4212 btrfs_dec_block_group_ro(rc->block_group);
4213 iput(rc->data_inode);
4214out_put_bg:
4215 btrfs_put_block_group(bg);
4216 reloc_chunk_end(fs_info);
4217 free_reloc_control(rc);
4218 return err;
4219}
4220
4221static noinline_for_stack int mark_garbage_root(struct btrfs_root *root)
4222{
4223 struct btrfs_fs_info *fs_info = root->fs_info;
4224 struct btrfs_trans_handle *trans;
4225 int ret, err;
4226
4227 trans = btrfs_start_transaction(fs_info->tree_root, 0);
4228 if (IS_ERR(trans))
4229 return PTR_ERR(trans);
4230
4231 memset(&root->root_item.drop_progress, 0,
4232 sizeof(root->root_item.drop_progress));
4233 btrfs_set_root_drop_level(&root->root_item, 0);
4234 btrfs_set_root_refs(&root->root_item, 0);
4235 ret = btrfs_update_root(trans, fs_info->tree_root,
4236 &root->root_key, &root->root_item);
4237
4238 err = btrfs_end_transaction(trans);
4239 if (err)
4240 return err;
4241 return ret;
4242}
4243
4244/*
4245 * recover relocation interrupted by system crash.
4246 *
4247 * this function resumes merging reloc trees with corresponding fs trees.
4248 * this is important for keeping the sharing of tree blocks
4249 */
4250int btrfs_recover_relocation(struct btrfs_fs_info *fs_info)
4251{
4252 LIST_HEAD(reloc_roots);
4253 struct btrfs_key key;
4254 struct btrfs_root *fs_root;
4255 struct btrfs_root *reloc_root;
4256 struct btrfs_path *path;
4257 struct extent_buffer *leaf;
4258 struct reloc_control *rc = NULL;
4259 struct btrfs_trans_handle *trans;
4260 int ret;
4261 int err = 0;
4262
4263 path = btrfs_alloc_path();
4264 if (!path)
4265 return -ENOMEM;
4266 path->reada = READA_BACK;
4267
4268 key.objectid = BTRFS_TREE_RELOC_OBJECTID;
4269 key.type = BTRFS_ROOT_ITEM_KEY;
4270 key.offset = (u64)-1;
4271
4272 while (1) {
4273 ret = btrfs_search_slot(NULL, fs_info->tree_root, &key,
4274 path, 0, 0);
4275 if (ret < 0) {
4276 err = ret;
4277 goto out;
4278 }
4279 if (ret > 0) {
4280 if (path->slots[0] == 0)
4281 break;
4282 path->slots[0]--;
4283 }
4284 leaf = path->nodes[0];
4285 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
4286 btrfs_release_path(path);
4287
4288 if (key.objectid != BTRFS_TREE_RELOC_OBJECTID ||
4289 key.type != BTRFS_ROOT_ITEM_KEY)
4290 break;
4291
4292 reloc_root = btrfs_read_tree_root(fs_info->tree_root, &key);
4293 if (IS_ERR(reloc_root)) {
4294 err = PTR_ERR(reloc_root);
4295 goto out;
4296 }
4297
4298 set_bit(BTRFS_ROOT_SHAREABLE, &reloc_root->state);
4299 list_add(&reloc_root->root_list, &reloc_roots);
4300
4301 if (btrfs_root_refs(&reloc_root->root_item) > 0) {
4302 fs_root = btrfs_get_fs_root(fs_info,
4303 reloc_root->root_key.offset, false);
4304 if (IS_ERR(fs_root)) {
4305 ret = PTR_ERR(fs_root);
4306 if (ret != -ENOENT) {
4307 err = ret;
4308 goto out;
4309 }
4310 ret = mark_garbage_root(reloc_root);
4311 if (ret < 0) {
4312 err = ret;
4313 goto out;
4314 }
4315 } else {
4316 btrfs_put_root(fs_root);
4317 }
4318 }
4319
4320 if (key.offset == 0)
4321 break;
4322
4323 key.offset--;
4324 }
4325 btrfs_release_path(path);
4326
4327 if (list_empty(&reloc_roots))
4328 goto out;
4329
4330 rc = alloc_reloc_control(fs_info);
4331 if (!rc) {
4332 err = -ENOMEM;
4333 goto out;
4334 }
4335
4336 ret = reloc_chunk_start(fs_info);
4337 if (ret < 0) {
4338 err = ret;
4339 goto out_end;
4340 }
4341
4342 rc->extent_root = btrfs_extent_root(fs_info, 0);
4343
4344 set_reloc_control(rc);
4345
4346 trans = btrfs_join_transaction(rc->extent_root);
4347 if (IS_ERR(trans)) {
4348 err = PTR_ERR(trans);
4349 goto out_unset;
4350 }
4351
4352 rc->merge_reloc_tree = true;
4353
4354 while (!list_empty(&reloc_roots)) {
4355 reloc_root = list_entry(reloc_roots.next,
4356 struct btrfs_root, root_list);
4357 list_del(&reloc_root->root_list);
4358
4359 if (btrfs_root_refs(&reloc_root->root_item) == 0) {
4360 list_add_tail(&reloc_root->root_list,
4361 &rc->reloc_roots);
4362 continue;
4363 }
4364
4365 fs_root = btrfs_get_fs_root(fs_info, reloc_root->root_key.offset,
4366 false);
4367 if (IS_ERR(fs_root)) {
4368 err = PTR_ERR(fs_root);
4369 list_add_tail(&reloc_root->root_list, &reloc_roots);
4370 btrfs_end_transaction(trans);
4371 goto out_unset;
4372 }
4373
4374 err = __add_reloc_root(reloc_root);
4375 ASSERT(err != -EEXIST);
4376 if (err) {
4377 list_add_tail(&reloc_root->root_list, &reloc_roots);
4378 btrfs_put_root(fs_root);
4379 btrfs_end_transaction(trans);
4380 goto out_unset;
4381 }
4382 fs_root->reloc_root = btrfs_grab_root(reloc_root);
4383 btrfs_put_root(fs_root);
4384 }
4385
4386 err = btrfs_commit_transaction(trans);
4387 if (err)
4388 goto out_unset;
4389
4390 merge_reloc_roots(rc);
4391
4392 unset_reloc_control(rc);
4393
4394 trans = btrfs_join_transaction(rc->extent_root);
4395 if (IS_ERR(trans)) {
4396 err = PTR_ERR(trans);
4397 goto out_clean;
4398 }
4399 err = btrfs_commit_transaction(trans);
4400out_clean:
4401 ret = clean_dirty_subvols(rc);
4402 if (ret < 0 && !err)
4403 err = ret;
4404out_unset:
4405 unset_reloc_control(rc);
4406out_end:
4407 reloc_chunk_end(fs_info);
4408 free_reloc_control(rc);
4409out:
4410 free_reloc_roots(&reloc_roots);
4411
4412 btrfs_free_path(path);
4413
4414 if (err == 0) {
4415 /* cleanup orphan inode in data relocation tree */
4416 fs_root = btrfs_grab_root(fs_info->data_reloc_root);
4417 ASSERT(fs_root);
4418 err = btrfs_orphan_cleanup(fs_root);
4419 btrfs_put_root(fs_root);
4420 }
4421 return err;
4422}
4423
4424/*
4425 * helper to add ordered checksum for data relocation.
4426 *
4427 * cloning checksum properly handles the nodatasum extents.
4428 * it also saves CPU time to re-calculate the checksum.
4429 */
4430int btrfs_reloc_clone_csums(struct btrfs_ordered_extent *ordered)
4431{
4432 struct btrfs_inode *inode = BTRFS_I(ordered->inode);
4433 struct btrfs_fs_info *fs_info = inode->root->fs_info;
4434 u64 disk_bytenr = ordered->file_offset + inode->index_cnt;
4435 struct btrfs_root *csum_root = btrfs_csum_root(fs_info, disk_bytenr);
4436 LIST_HEAD(list);
4437 int ret;
4438
4439 ret = btrfs_lookup_csums_list(csum_root, disk_bytenr,
4440 disk_bytenr + ordered->num_bytes - 1,
4441 &list, 0, false);
4442 if (ret)
4443 return ret;
4444
4445 while (!list_empty(&list)) {
4446 struct btrfs_ordered_sum *sums =
4447 list_entry(list.next, struct btrfs_ordered_sum, list);
4448
4449 list_del_init(&sums->list);
4450
4451 /*
4452 * We need to offset the new_bytenr based on where the csum is.
4453 * We need to do this because we will read in entire prealloc
4454 * extents but we may have written to say the middle of the
4455 * prealloc extent, so we need to make sure the csum goes with
4456 * the right disk offset.
4457 *
4458 * We can do this because the data reloc inode refers strictly
4459 * to the on disk bytes, so we don't have to worry about
4460 * disk_len vs real len like with real inodes since it's all
4461 * disk length.
4462 */
4463 sums->logical = ordered->disk_bytenr + sums->logical - disk_bytenr;
4464 btrfs_add_ordered_sum(ordered, sums);
4465 }
4466
4467 return 0;
4468}
4469
4470int btrfs_reloc_cow_block(struct btrfs_trans_handle *trans,
4471 struct btrfs_root *root,
4472 const struct extent_buffer *buf,
4473 struct extent_buffer *cow)
4474{
4475 struct btrfs_fs_info *fs_info = root->fs_info;
4476 struct reloc_control *rc;
4477 struct btrfs_backref_node *node;
4478 int first_cow = 0;
4479 int level;
4480 int ret = 0;
4481
4482 rc = fs_info->reloc_ctl;
4483 if (!rc)
4484 return 0;
4485
4486 BUG_ON(rc->stage == UPDATE_DATA_PTRS && btrfs_is_data_reloc_root(root));
4487
4488 level = btrfs_header_level(buf);
4489 if (btrfs_header_generation(buf) <=
4490 btrfs_root_last_snapshot(&root->root_item))
4491 first_cow = 1;
4492
4493 if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID &&
4494 rc->create_reloc_tree) {
4495 WARN_ON(!first_cow && level == 0);
4496
4497 node = rc->backref_cache.path[level];
4498 BUG_ON(node->bytenr != buf->start &&
4499 node->new_bytenr != buf->start);
4500
4501 btrfs_backref_drop_node_buffer(node);
4502 atomic_inc(&cow->refs);
4503 node->eb = cow;
4504 node->new_bytenr = cow->start;
4505
4506 if (!node->pending) {
4507 list_move_tail(&node->list,
4508 &rc->backref_cache.pending[level]);
4509 node->pending = 1;
4510 }
4511
4512 if (first_cow)
4513 mark_block_processed(rc, node);
4514
4515 if (first_cow && level > 0)
4516 rc->nodes_relocated += buf->len;
4517 }
4518
4519 if (level == 0 && first_cow && rc->stage == UPDATE_DATA_PTRS)
4520 ret = replace_file_extents(trans, rc, root, cow);
4521 return ret;
4522}
4523
4524/*
4525 * called before creating snapshot. it calculates metadata reservation
4526 * required for relocating tree blocks in the snapshot
4527 */
4528void btrfs_reloc_pre_snapshot(struct btrfs_pending_snapshot *pending,
4529 u64 *bytes_to_reserve)
4530{
4531 struct btrfs_root *root = pending->root;
4532 struct reloc_control *rc = root->fs_info->reloc_ctl;
4533
4534 if (!rc || !have_reloc_root(root))
4535 return;
4536
4537 if (!rc->merge_reloc_tree)
4538 return;
4539
4540 root = root->reloc_root;
4541 BUG_ON(btrfs_root_refs(&root->root_item) == 0);
4542 /*
4543 * relocation is in the stage of merging trees. the space
4544 * used by merging a reloc tree is twice the size of
4545 * relocated tree nodes in the worst case. half for cowing
4546 * the reloc tree, half for cowing the fs tree. the space
4547 * used by cowing the reloc tree will be freed after the
4548 * tree is dropped. if we create snapshot, cowing the fs
4549 * tree may use more space than it frees. so we need
4550 * reserve extra space.
4551 */
4552 *bytes_to_reserve += rc->nodes_relocated;
4553}
4554
4555/*
4556 * called after snapshot is created. migrate block reservation
4557 * and create reloc root for the newly created snapshot
4558 *
4559 * This is similar to btrfs_init_reloc_root(), we come out of here with two
4560 * references held on the reloc_root, one for root->reloc_root and one for
4561 * rc->reloc_roots.
4562 */
4563int btrfs_reloc_post_snapshot(struct btrfs_trans_handle *trans,
4564 struct btrfs_pending_snapshot *pending)
4565{
4566 struct btrfs_root *root = pending->root;
4567 struct btrfs_root *reloc_root;
4568 struct btrfs_root *new_root;
4569 struct reloc_control *rc = root->fs_info->reloc_ctl;
4570 int ret;
4571
4572 if (!rc || !have_reloc_root(root))
4573 return 0;
4574
4575 rc = root->fs_info->reloc_ctl;
4576 rc->merging_rsv_size += rc->nodes_relocated;
4577
4578 if (rc->merge_reloc_tree) {
4579 ret = btrfs_block_rsv_migrate(&pending->block_rsv,
4580 rc->block_rsv,
4581 rc->nodes_relocated, true);
4582 if (ret)
4583 return ret;
4584 }
4585
4586 new_root = pending->snap;
4587 reloc_root = create_reloc_root(trans, root->reloc_root,
4588 new_root->root_key.objectid);
4589 if (IS_ERR(reloc_root))
4590 return PTR_ERR(reloc_root);
4591
4592 ret = __add_reloc_root(reloc_root);
4593 ASSERT(ret != -EEXIST);
4594 if (ret) {
4595 /* Pairs with create_reloc_root */
4596 btrfs_put_root(reloc_root);
4597 return ret;
4598 }
4599 new_root->reloc_root = btrfs_grab_root(reloc_root);
4600
4601 if (rc->create_reloc_tree)
4602 ret = clone_backref_node(trans, rc, root, reloc_root);
4603 return ret;
4604}
4605
4606/*
4607 * Get the current bytenr for the block group which is being relocated.
4608 *
4609 * Return U64_MAX if no running relocation.
4610 */
4611u64 btrfs_get_reloc_bg_bytenr(const struct btrfs_fs_info *fs_info)
4612{
4613 u64 logical = U64_MAX;
4614
4615 lockdep_assert_held(&fs_info->reloc_mutex);
4616
4617 if (fs_info->reloc_ctl && fs_info->reloc_ctl->block_group)
4618 logical = fs_info->reloc_ctl->block_group->start;
4619 return logical;
4620}