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/*
2 * Copyright (C) 2009 Oracle. All rights reserved.
3 *
4 * This program is free software; you can redistribute it and/or
5 * modify it under the terms of the GNU General Public
6 * License v2 as published by the Free Software Foundation.
7 *
8 * This program is distributed in the hope that it will be useful,
9 * but WITHOUT ANY WARRANTY; without even the implied warranty of
10 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
11 * General Public License for more details.
12 *
13 * You should have received a copy of the GNU General Public
14 * License along with this program; if not, write to the
15 * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
16 * Boston, MA 021110-1307, USA.
17 */
18
19#include <linux/sched.h>
20#include <linux/pagemap.h>
21#include <linux/writeback.h>
22#include <linux/blkdev.h>
23#include <linux/rbtree.h>
24#include <linux/slab.h>
25#include "ctree.h"
26#include "disk-io.h"
27#include "transaction.h"
28#include "volumes.h"
29#include "locking.h"
30#include "btrfs_inode.h"
31#include "async-thread.h"
32#include "free-space-cache.h"
33#include "inode-map.h"
34
35/*
36 * backref_node, mapping_node and tree_block start with this
37 */
38struct tree_entry {
39 struct rb_node rb_node;
40 u64 bytenr;
41};
42
43/*
44 * present a tree block in the backref cache
45 */
46struct backref_node {
47 struct rb_node rb_node;
48 u64 bytenr;
49
50 u64 new_bytenr;
51 /* objectid of tree block owner, can be not uptodate */
52 u64 owner;
53 /* link to pending, changed or detached list */
54 struct list_head list;
55 /* list of upper level blocks reference this block */
56 struct list_head upper;
57 /* list of child blocks in the cache */
58 struct list_head lower;
59 /* NULL if this node is not tree root */
60 struct btrfs_root *root;
61 /* extent buffer got by COW the block */
62 struct extent_buffer *eb;
63 /* level of tree block */
64 unsigned int level:8;
65 /* is the block in non-reference counted tree */
66 unsigned int cowonly:1;
67 /* 1 if no child node in the cache */
68 unsigned int lowest:1;
69 /* is the extent buffer locked */
70 unsigned int locked:1;
71 /* has the block been processed */
72 unsigned int processed:1;
73 /* have backrefs of this block been checked */
74 unsigned int checked:1;
75 /*
76 * 1 if corresponding block has been cowed but some upper
77 * level block pointers may not point to the new location
78 */
79 unsigned int pending:1;
80 /*
81 * 1 if the backref node isn't connected to any other
82 * backref node.
83 */
84 unsigned int detached:1;
85};
86
87/*
88 * present a block pointer in the backref cache
89 */
90struct backref_edge {
91 struct list_head list[2];
92 struct backref_node *node[2];
93};
94
95#define LOWER 0
96#define UPPER 1
97#define RELOCATION_RESERVED_NODES 256
98
99struct backref_cache {
100 /* red black tree of all backref nodes in the cache */
101 struct rb_root rb_root;
102 /* for passing backref nodes to btrfs_reloc_cow_block */
103 struct backref_node *path[BTRFS_MAX_LEVEL];
104 /*
105 * list of blocks that have been cowed but some block
106 * pointers in upper level blocks may not reflect the
107 * new location
108 */
109 struct list_head pending[BTRFS_MAX_LEVEL];
110 /* list of backref nodes with no child node */
111 struct list_head leaves;
112 /* list of blocks that have been cowed in current transaction */
113 struct list_head changed;
114 /* list of detached backref node. */
115 struct list_head detached;
116
117 u64 last_trans;
118
119 int nr_nodes;
120 int nr_edges;
121};
122
123/*
124 * map address of tree root to tree
125 */
126struct mapping_node {
127 struct rb_node rb_node;
128 u64 bytenr;
129 void *data;
130};
131
132struct mapping_tree {
133 struct rb_root rb_root;
134 spinlock_t lock;
135};
136
137/*
138 * present a tree block to process
139 */
140struct tree_block {
141 struct rb_node rb_node;
142 u64 bytenr;
143 struct btrfs_key key;
144 unsigned int level:8;
145 unsigned int key_ready:1;
146};
147
148#define MAX_EXTENTS 128
149
150struct file_extent_cluster {
151 u64 start;
152 u64 end;
153 u64 boundary[MAX_EXTENTS];
154 unsigned int nr;
155};
156
157struct reloc_control {
158 /* block group to relocate */
159 struct btrfs_block_group_cache *block_group;
160 /* extent tree */
161 struct btrfs_root *extent_root;
162 /* inode for moving data */
163 struct inode *data_inode;
164
165 struct btrfs_block_rsv *block_rsv;
166
167 struct backref_cache backref_cache;
168
169 struct file_extent_cluster cluster;
170 /* tree blocks have been processed */
171 struct extent_io_tree processed_blocks;
172 /* map start of tree root to corresponding reloc tree */
173 struct mapping_tree reloc_root_tree;
174 /* list of reloc trees */
175 struct list_head reloc_roots;
176 /* size of metadata reservation for merging reloc trees */
177 u64 merging_rsv_size;
178 /* size of relocated tree nodes */
179 u64 nodes_relocated;
180 /* reserved size for block group relocation*/
181 u64 reserved_bytes;
182
183 u64 search_start;
184 u64 extents_found;
185
186 unsigned int stage:8;
187 unsigned int create_reloc_tree:1;
188 unsigned int merge_reloc_tree:1;
189 unsigned int found_file_extent:1;
190};
191
192/* stages of data relocation */
193#define MOVE_DATA_EXTENTS 0
194#define UPDATE_DATA_PTRS 1
195
196static void remove_backref_node(struct backref_cache *cache,
197 struct backref_node *node);
198static void __mark_block_processed(struct reloc_control *rc,
199 struct backref_node *node);
200
201static void mapping_tree_init(struct mapping_tree *tree)
202{
203 tree->rb_root = RB_ROOT;
204 spin_lock_init(&tree->lock);
205}
206
207static void backref_cache_init(struct backref_cache *cache)
208{
209 int i;
210 cache->rb_root = RB_ROOT;
211 for (i = 0; i < BTRFS_MAX_LEVEL; i++)
212 INIT_LIST_HEAD(&cache->pending[i]);
213 INIT_LIST_HEAD(&cache->changed);
214 INIT_LIST_HEAD(&cache->detached);
215 INIT_LIST_HEAD(&cache->leaves);
216}
217
218static void backref_cache_cleanup(struct backref_cache *cache)
219{
220 struct backref_node *node;
221 int i;
222
223 while (!list_empty(&cache->detached)) {
224 node = list_entry(cache->detached.next,
225 struct backref_node, list);
226 remove_backref_node(cache, node);
227 }
228
229 while (!list_empty(&cache->leaves)) {
230 node = list_entry(cache->leaves.next,
231 struct backref_node, lower);
232 remove_backref_node(cache, node);
233 }
234
235 cache->last_trans = 0;
236
237 for (i = 0; i < BTRFS_MAX_LEVEL; i++)
238 BUG_ON(!list_empty(&cache->pending[i]));
239 BUG_ON(!list_empty(&cache->changed));
240 BUG_ON(!list_empty(&cache->detached));
241 BUG_ON(!RB_EMPTY_ROOT(&cache->rb_root));
242 BUG_ON(cache->nr_nodes);
243 BUG_ON(cache->nr_edges);
244}
245
246static struct backref_node *alloc_backref_node(struct backref_cache *cache)
247{
248 struct backref_node *node;
249
250 node = kzalloc(sizeof(*node), GFP_NOFS);
251 if (node) {
252 INIT_LIST_HEAD(&node->list);
253 INIT_LIST_HEAD(&node->upper);
254 INIT_LIST_HEAD(&node->lower);
255 RB_CLEAR_NODE(&node->rb_node);
256 cache->nr_nodes++;
257 }
258 return node;
259}
260
261static void free_backref_node(struct backref_cache *cache,
262 struct backref_node *node)
263{
264 if (node) {
265 cache->nr_nodes--;
266 kfree(node);
267 }
268}
269
270static struct backref_edge *alloc_backref_edge(struct backref_cache *cache)
271{
272 struct backref_edge *edge;
273
274 edge = kzalloc(sizeof(*edge), GFP_NOFS);
275 if (edge)
276 cache->nr_edges++;
277 return edge;
278}
279
280static void free_backref_edge(struct backref_cache *cache,
281 struct backref_edge *edge)
282{
283 if (edge) {
284 cache->nr_edges--;
285 kfree(edge);
286 }
287}
288
289static struct rb_node *tree_insert(struct rb_root *root, u64 bytenr,
290 struct rb_node *node)
291{
292 struct rb_node **p = &root->rb_node;
293 struct rb_node *parent = NULL;
294 struct tree_entry *entry;
295
296 while (*p) {
297 parent = *p;
298 entry = rb_entry(parent, struct tree_entry, rb_node);
299
300 if (bytenr < entry->bytenr)
301 p = &(*p)->rb_left;
302 else if (bytenr > entry->bytenr)
303 p = &(*p)->rb_right;
304 else
305 return parent;
306 }
307
308 rb_link_node(node, parent, p);
309 rb_insert_color(node, root);
310 return NULL;
311}
312
313static struct rb_node *tree_search(struct rb_root *root, u64 bytenr)
314{
315 struct rb_node *n = root->rb_node;
316 struct tree_entry *entry;
317
318 while (n) {
319 entry = rb_entry(n, struct tree_entry, rb_node);
320
321 if (bytenr < entry->bytenr)
322 n = n->rb_left;
323 else if (bytenr > entry->bytenr)
324 n = n->rb_right;
325 else
326 return n;
327 }
328 return NULL;
329}
330
331static void backref_tree_panic(struct rb_node *rb_node, int errno, u64 bytenr)
332{
333
334 struct btrfs_fs_info *fs_info = NULL;
335 struct backref_node *bnode = rb_entry(rb_node, struct backref_node,
336 rb_node);
337 if (bnode->root)
338 fs_info = bnode->root->fs_info;
339 btrfs_panic(fs_info, errno, "Inconsistency in backref cache "
340 "found at offset %llu\n", bytenr);
341}
342
343/*
344 * walk up backref nodes until reach node presents tree root
345 */
346static struct backref_node *walk_up_backref(struct backref_node *node,
347 struct backref_edge *edges[],
348 int *index)
349{
350 struct backref_edge *edge;
351 int idx = *index;
352
353 while (!list_empty(&node->upper)) {
354 edge = list_entry(node->upper.next,
355 struct backref_edge, list[LOWER]);
356 edges[idx++] = edge;
357 node = edge->node[UPPER];
358 }
359 BUG_ON(node->detached);
360 *index = idx;
361 return node;
362}
363
364/*
365 * walk down backref nodes to find start of next reference path
366 */
367static struct backref_node *walk_down_backref(struct backref_edge *edges[],
368 int *index)
369{
370 struct backref_edge *edge;
371 struct backref_node *lower;
372 int idx = *index;
373
374 while (idx > 0) {
375 edge = edges[idx - 1];
376 lower = edge->node[LOWER];
377 if (list_is_last(&edge->list[LOWER], &lower->upper)) {
378 idx--;
379 continue;
380 }
381 edge = list_entry(edge->list[LOWER].next,
382 struct backref_edge, list[LOWER]);
383 edges[idx - 1] = edge;
384 *index = idx;
385 return edge->node[UPPER];
386 }
387 *index = 0;
388 return NULL;
389}
390
391static void unlock_node_buffer(struct backref_node *node)
392{
393 if (node->locked) {
394 btrfs_tree_unlock(node->eb);
395 node->locked = 0;
396 }
397}
398
399static void drop_node_buffer(struct backref_node *node)
400{
401 if (node->eb) {
402 unlock_node_buffer(node);
403 free_extent_buffer(node->eb);
404 node->eb = NULL;
405 }
406}
407
408static void drop_backref_node(struct backref_cache *tree,
409 struct backref_node *node)
410{
411 BUG_ON(!list_empty(&node->upper));
412
413 drop_node_buffer(node);
414 list_del(&node->list);
415 list_del(&node->lower);
416 if (!RB_EMPTY_NODE(&node->rb_node))
417 rb_erase(&node->rb_node, &tree->rb_root);
418 free_backref_node(tree, node);
419}
420
421/*
422 * remove a backref node from the backref cache
423 */
424static void remove_backref_node(struct backref_cache *cache,
425 struct backref_node *node)
426{
427 struct backref_node *upper;
428 struct backref_edge *edge;
429
430 if (!node)
431 return;
432
433 BUG_ON(!node->lowest && !node->detached);
434 while (!list_empty(&node->upper)) {
435 edge = list_entry(node->upper.next, struct backref_edge,
436 list[LOWER]);
437 upper = edge->node[UPPER];
438 list_del(&edge->list[LOWER]);
439 list_del(&edge->list[UPPER]);
440 free_backref_edge(cache, edge);
441
442 if (RB_EMPTY_NODE(&upper->rb_node)) {
443 BUG_ON(!list_empty(&node->upper));
444 drop_backref_node(cache, node);
445 node = upper;
446 node->lowest = 1;
447 continue;
448 }
449 /*
450 * add the node to leaf node list if no other
451 * child block cached.
452 */
453 if (list_empty(&upper->lower)) {
454 list_add_tail(&upper->lower, &cache->leaves);
455 upper->lowest = 1;
456 }
457 }
458
459 drop_backref_node(cache, node);
460}
461
462static void update_backref_node(struct backref_cache *cache,
463 struct backref_node *node, u64 bytenr)
464{
465 struct rb_node *rb_node;
466 rb_erase(&node->rb_node, &cache->rb_root);
467 node->bytenr = bytenr;
468 rb_node = tree_insert(&cache->rb_root, node->bytenr, &node->rb_node);
469 if (rb_node)
470 backref_tree_panic(rb_node, -EEXIST, bytenr);
471}
472
473/*
474 * update backref cache after a transaction commit
475 */
476static int update_backref_cache(struct btrfs_trans_handle *trans,
477 struct backref_cache *cache)
478{
479 struct backref_node *node;
480 int level = 0;
481
482 if (cache->last_trans == 0) {
483 cache->last_trans = trans->transid;
484 return 0;
485 }
486
487 if (cache->last_trans == trans->transid)
488 return 0;
489
490 /*
491 * detached nodes are used to avoid unnecessary backref
492 * lookup. transaction commit changes the extent tree.
493 * so the detached nodes are no longer useful.
494 */
495 while (!list_empty(&cache->detached)) {
496 node = list_entry(cache->detached.next,
497 struct backref_node, list);
498 remove_backref_node(cache, node);
499 }
500
501 while (!list_empty(&cache->changed)) {
502 node = list_entry(cache->changed.next,
503 struct backref_node, list);
504 list_del_init(&node->list);
505 BUG_ON(node->pending);
506 update_backref_node(cache, node, node->new_bytenr);
507 }
508
509 /*
510 * some nodes can be left in the pending list if there were
511 * errors during processing the pending nodes.
512 */
513 for (level = 0; level < BTRFS_MAX_LEVEL; level++) {
514 list_for_each_entry(node, &cache->pending[level], list) {
515 BUG_ON(!node->pending);
516 if (node->bytenr == node->new_bytenr)
517 continue;
518 update_backref_node(cache, node, node->new_bytenr);
519 }
520 }
521
522 cache->last_trans = 0;
523 return 1;
524}
525
526
527static int should_ignore_root(struct btrfs_root *root)
528{
529 struct btrfs_root *reloc_root;
530
531 if (!root->ref_cows)
532 return 0;
533
534 reloc_root = root->reloc_root;
535 if (!reloc_root)
536 return 0;
537
538 if (btrfs_root_last_snapshot(&reloc_root->root_item) ==
539 root->fs_info->running_transaction->transid - 1)
540 return 0;
541 /*
542 * if there is reloc tree and it was created in previous
543 * transaction backref lookup can find the reloc tree,
544 * so backref node for the fs tree root is useless for
545 * relocation.
546 */
547 return 1;
548}
549/*
550 * find reloc tree by address of tree root
551 */
552static struct btrfs_root *find_reloc_root(struct reloc_control *rc,
553 u64 bytenr)
554{
555 struct rb_node *rb_node;
556 struct mapping_node *node;
557 struct btrfs_root *root = NULL;
558
559 spin_lock(&rc->reloc_root_tree.lock);
560 rb_node = tree_search(&rc->reloc_root_tree.rb_root, bytenr);
561 if (rb_node) {
562 node = rb_entry(rb_node, struct mapping_node, rb_node);
563 root = (struct btrfs_root *)node->data;
564 }
565 spin_unlock(&rc->reloc_root_tree.lock);
566 return root;
567}
568
569static int is_cowonly_root(u64 root_objectid)
570{
571 if (root_objectid == BTRFS_ROOT_TREE_OBJECTID ||
572 root_objectid == BTRFS_EXTENT_TREE_OBJECTID ||
573 root_objectid == BTRFS_CHUNK_TREE_OBJECTID ||
574 root_objectid == BTRFS_DEV_TREE_OBJECTID ||
575 root_objectid == BTRFS_TREE_LOG_OBJECTID ||
576 root_objectid == BTRFS_CSUM_TREE_OBJECTID ||
577 root_objectid == BTRFS_UUID_TREE_OBJECTID ||
578 root_objectid == BTRFS_QUOTA_TREE_OBJECTID)
579 return 1;
580 return 0;
581}
582
583static struct btrfs_root *read_fs_root(struct btrfs_fs_info *fs_info,
584 u64 root_objectid)
585{
586 struct btrfs_key key;
587
588 key.objectid = root_objectid;
589 key.type = BTRFS_ROOT_ITEM_KEY;
590 if (is_cowonly_root(root_objectid))
591 key.offset = 0;
592 else
593 key.offset = (u64)-1;
594
595 return btrfs_get_fs_root(fs_info, &key, false);
596}
597
598#ifdef BTRFS_COMPAT_EXTENT_TREE_V0
599static noinline_for_stack
600struct btrfs_root *find_tree_root(struct reloc_control *rc,
601 struct extent_buffer *leaf,
602 struct btrfs_extent_ref_v0 *ref0)
603{
604 struct btrfs_root *root;
605 u64 root_objectid = btrfs_ref_root_v0(leaf, ref0);
606 u64 generation = btrfs_ref_generation_v0(leaf, ref0);
607
608 BUG_ON(root_objectid == BTRFS_TREE_RELOC_OBJECTID);
609
610 root = read_fs_root(rc->extent_root->fs_info, root_objectid);
611 BUG_ON(IS_ERR(root));
612
613 if (root->ref_cows &&
614 generation != btrfs_root_generation(&root->root_item))
615 return NULL;
616
617 return root;
618}
619#endif
620
621static noinline_for_stack
622int find_inline_backref(struct extent_buffer *leaf, int slot,
623 unsigned long *ptr, unsigned long *end)
624{
625 struct btrfs_key key;
626 struct btrfs_extent_item *ei;
627 struct btrfs_tree_block_info *bi;
628 u32 item_size;
629
630 btrfs_item_key_to_cpu(leaf, &key, slot);
631
632 item_size = btrfs_item_size_nr(leaf, slot);
633#ifdef BTRFS_COMPAT_EXTENT_TREE_V0
634 if (item_size < sizeof(*ei)) {
635 WARN_ON(item_size != sizeof(struct btrfs_extent_item_v0));
636 return 1;
637 }
638#endif
639 ei = btrfs_item_ptr(leaf, slot, struct btrfs_extent_item);
640 WARN_ON(!(btrfs_extent_flags(leaf, ei) &
641 BTRFS_EXTENT_FLAG_TREE_BLOCK));
642
643 if (key.type == BTRFS_EXTENT_ITEM_KEY &&
644 item_size <= sizeof(*ei) + sizeof(*bi)) {
645 WARN_ON(item_size < sizeof(*ei) + sizeof(*bi));
646 return 1;
647 }
648 if (key.type == BTRFS_METADATA_ITEM_KEY &&
649 item_size <= sizeof(*ei)) {
650 WARN_ON(item_size < sizeof(*ei));
651 return 1;
652 }
653
654 if (key.type == BTRFS_EXTENT_ITEM_KEY) {
655 bi = (struct btrfs_tree_block_info *)(ei + 1);
656 *ptr = (unsigned long)(bi + 1);
657 } else {
658 *ptr = (unsigned long)(ei + 1);
659 }
660 *end = (unsigned long)ei + item_size;
661 return 0;
662}
663
664/*
665 * build backref tree for a given tree block. root of the backref tree
666 * corresponds the tree block, leaves of the backref tree correspond
667 * roots of b-trees that reference the tree block.
668 *
669 * the basic idea of this function is check backrefs of a given block
670 * to find upper level blocks that refernece the block, and then check
671 * bakcrefs of these upper level blocks recursively. the recursion stop
672 * when tree root is reached or backrefs for the block is cached.
673 *
674 * NOTE: if we find backrefs for a block are cached, we know backrefs
675 * for all upper level blocks that directly/indirectly reference the
676 * block are also cached.
677 */
678static noinline_for_stack
679struct backref_node *build_backref_tree(struct reloc_control *rc,
680 struct btrfs_key *node_key,
681 int level, u64 bytenr)
682{
683 struct backref_cache *cache = &rc->backref_cache;
684 struct btrfs_path *path1;
685 struct btrfs_path *path2;
686 struct extent_buffer *eb;
687 struct btrfs_root *root;
688 struct backref_node *cur;
689 struct backref_node *upper;
690 struct backref_node *lower;
691 struct backref_node *node = NULL;
692 struct backref_node *exist = NULL;
693 struct backref_edge *edge;
694 struct rb_node *rb_node;
695 struct btrfs_key key;
696 unsigned long end;
697 unsigned long ptr;
698 LIST_HEAD(list);
699 LIST_HEAD(useless);
700 int cowonly;
701 int ret;
702 int err = 0;
703 bool need_check = true;
704
705 path1 = btrfs_alloc_path();
706 path2 = btrfs_alloc_path();
707 if (!path1 || !path2) {
708 err = -ENOMEM;
709 goto out;
710 }
711 path1->reada = 1;
712 path2->reada = 2;
713
714 node = alloc_backref_node(cache);
715 if (!node) {
716 err = -ENOMEM;
717 goto out;
718 }
719
720 node->bytenr = bytenr;
721 node->level = level;
722 node->lowest = 1;
723 cur = node;
724again:
725 end = 0;
726 ptr = 0;
727 key.objectid = cur->bytenr;
728 key.type = BTRFS_METADATA_ITEM_KEY;
729 key.offset = (u64)-1;
730
731 path1->search_commit_root = 1;
732 path1->skip_locking = 1;
733 ret = btrfs_search_slot(NULL, rc->extent_root, &key, path1,
734 0, 0);
735 if (ret < 0) {
736 err = ret;
737 goto out;
738 }
739 BUG_ON(!ret || !path1->slots[0]);
740
741 path1->slots[0]--;
742
743 WARN_ON(cur->checked);
744 if (!list_empty(&cur->upper)) {
745 /*
746 * the backref was added previously when processing
747 * backref of type BTRFS_TREE_BLOCK_REF_KEY
748 */
749 BUG_ON(!list_is_singular(&cur->upper));
750 edge = list_entry(cur->upper.next, struct backref_edge,
751 list[LOWER]);
752 BUG_ON(!list_empty(&edge->list[UPPER]));
753 exist = edge->node[UPPER];
754 /*
755 * add the upper level block to pending list if we need
756 * check its backrefs
757 */
758 if (!exist->checked)
759 list_add_tail(&edge->list[UPPER], &list);
760 } else {
761 exist = NULL;
762 }
763
764 while (1) {
765 cond_resched();
766 eb = path1->nodes[0];
767
768 if (ptr >= end) {
769 if (path1->slots[0] >= btrfs_header_nritems(eb)) {
770 ret = btrfs_next_leaf(rc->extent_root, path1);
771 if (ret < 0) {
772 err = ret;
773 goto out;
774 }
775 if (ret > 0)
776 break;
777 eb = path1->nodes[0];
778 }
779
780 btrfs_item_key_to_cpu(eb, &key, path1->slots[0]);
781 if (key.objectid != cur->bytenr) {
782 WARN_ON(exist);
783 break;
784 }
785
786 if (key.type == BTRFS_EXTENT_ITEM_KEY ||
787 key.type == BTRFS_METADATA_ITEM_KEY) {
788 ret = find_inline_backref(eb, path1->slots[0],
789 &ptr, &end);
790 if (ret)
791 goto next;
792 }
793 }
794
795 if (ptr < end) {
796 /* update key for inline back ref */
797 struct btrfs_extent_inline_ref *iref;
798 iref = (struct btrfs_extent_inline_ref *)ptr;
799 key.type = btrfs_extent_inline_ref_type(eb, iref);
800 key.offset = btrfs_extent_inline_ref_offset(eb, iref);
801 WARN_ON(key.type != BTRFS_TREE_BLOCK_REF_KEY &&
802 key.type != BTRFS_SHARED_BLOCK_REF_KEY);
803 }
804
805 if (exist &&
806 ((key.type == BTRFS_TREE_BLOCK_REF_KEY &&
807 exist->owner == key.offset) ||
808 (key.type == BTRFS_SHARED_BLOCK_REF_KEY &&
809 exist->bytenr == key.offset))) {
810 exist = NULL;
811 goto next;
812 }
813
814#ifdef BTRFS_COMPAT_EXTENT_TREE_V0
815 if (key.type == BTRFS_SHARED_BLOCK_REF_KEY ||
816 key.type == BTRFS_EXTENT_REF_V0_KEY) {
817 if (key.type == BTRFS_EXTENT_REF_V0_KEY) {
818 struct btrfs_extent_ref_v0 *ref0;
819 ref0 = btrfs_item_ptr(eb, path1->slots[0],
820 struct btrfs_extent_ref_v0);
821 if (key.objectid == key.offset) {
822 root = find_tree_root(rc, eb, ref0);
823 if (root && !should_ignore_root(root))
824 cur->root = root;
825 else
826 list_add(&cur->list, &useless);
827 break;
828 }
829 if (is_cowonly_root(btrfs_ref_root_v0(eb,
830 ref0)))
831 cur->cowonly = 1;
832 }
833#else
834 BUG_ON(key.type == BTRFS_EXTENT_REF_V0_KEY);
835 if (key.type == BTRFS_SHARED_BLOCK_REF_KEY) {
836#endif
837 if (key.objectid == key.offset) {
838 /*
839 * only root blocks of reloc trees use
840 * backref of this type.
841 */
842 root = find_reloc_root(rc, cur->bytenr);
843 BUG_ON(!root);
844 cur->root = root;
845 break;
846 }
847
848 edge = alloc_backref_edge(cache);
849 if (!edge) {
850 err = -ENOMEM;
851 goto out;
852 }
853 rb_node = tree_search(&cache->rb_root, key.offset);
854 if (!rb_node) {
855 upper = alloc_backref_node(cache);
856 if (!upper) {
857 free_backref_edge(cache, edge);
858 err = -ENOMEM;
859 goto out;
860 }
861 upper->bytenr = key.offset;
862 upper->level = cur->level + 1;
863 /*
864 * backrefs for the upper level block isn't
865 * cached, add the block to pending list
866 */
867 list_add_tail(&edge->list[UPPER], &list);
868 } else {
869 upper = rb_entry(rb_node, struct backref_node,
870 rb_node);
871 BUG_ON(!upper->checked);
872 INIT_LIST_HEAD(&edge->list[UPPER]);
873 }
874 list_add_tail(&edge->list[LOWER], &cur->upper);
875 edge->node[LOWER] = cur;
876 edge->node[UPPER] = upper;
877
878 goto next;
879 } else if (key.type != BTRFS_TREE_BLOCK_REF_KEY) {
880 goto next;
881 }
882
883 /* key.type == BTRFS_TREE_BLOCK_REF_KEY */
884 root = read_fs_root(rc->extent_root->fs_info, key.offset);
885 if (IS_ERR(root)) {
886 err = PTR_ERR(root);
887 goto out;
888 }
889
890 if (!root->ref_cows)
891 cur->cowonly = 1;
892
893 if (btrfs_root_level(&root->root_item) == cur->level) {
894 /* tree root */
895 BUG_ON(btrfs_root_bytenr(&root->root_item) !=
896 cur->bytenr);
897 if (should_ignore_root(root))
898 list_add(&cur->list, &useless);
899 else
900 cur->root = root;
901 break;
902 }
903
904 level = cur->level + 1;
905
906 /*
907 * searching the tree to find upper level blocks
908 * reference the block.
909 */
910 path2->search_commit_root = 1;
911 path2->skip_locking = 1;
912 path2->lowest_level = level;
913 ret = btrfs_search_slot(NULL, root, node_key, path2, 0, 0);
914 path2->lowest_level = 0;
915 if (ret < 0) {
916 err = ret;
917 goto out;
918 }
919 if (ret > 0 && path2->slots[level] > 0)
920 path2->slots[level]--;
921
922 eb = path2->nodes[level];
923 WARN_ON(btrfs_node_blockptr(eb, path2->slots[level]) !=
924 cur->bytenr);
925
926 lower = cur;
927 need_check = true;
928 for (; level < BTRFS_MAX_LEVEL; level++) {
929 if (!path2->nodes[level]) {
930 BUG_ON(btrfs_root_bytenr(&root->root_item) !=
931 lower->bytenr);
932 if (should_ignore_root(root))
933 list_add(&lower->list, &useless);
934 else
935 lower->root = root;
936 break;
937 }
938
939 edge = alloc_backref_edge(cache);
940 if (!edge) {
941 err = -ENOMEM;
942 goto out;
943 }
944
945 eb = path2->nodes[level];
946 rb_node = tree_search(&cache->rb_root, eb->start);
947 if (!rb_node) {
948 upper = alloc_backref_node(cache);
949 if (!upper) {
950 free_backref_edge(cache, edge);
951 err = -ENOMEM;
952 goto out;
953 }
954 upper->bytenr = eb->start;
955 upper->owner = btrfs_header_owner(eb);
956 upper->level = lower->level + 1;
957 if (!root->ref_cows)
958 upper->cowonly = 1;
959
960 /*
961 * if we know the block isn't shared
962 * we can void checking its backrefs.
963 */
964 if (btrfs_block_can_be_shared(root, eb))
965 upper->checked = 0;
966 else
967 upper->checked = 1;
968
969 /*
970 * add the block to pending list if we
971 * need check its backrefs, we only do this once
972 * while walking up a tree as we will catch
973 * anything else later on.
974 */
975 if (!upper->checked && need_check) {
976 need_check = false;
977 list_add_tail(&edge->list[UPPER],
978 &list);
979 } else
980 INIT_LIST_HEAD(&edge->list[UPPER]);
981 } else {
982 upper = rb_entry(rb_node, struct backref_node,
983 rb_node);
984 BUG_ON(!upper->checked);
985 INIT_LIST_HEAD(&edge->list[UPPER]);
986 if (!upper->owner)
987 upper->owner = btrfs_header_owner(eb);
988 }
989 list_add_tail(&edge->list[LOWER], &lower->upper);
990 edge->node[LOWER] = lower;
991 edge->node[UPPER] = upper;
992
993 if (rb_node)
994 break;
995 lower = upper;
996 upper = NULL;
997 }
998 btrfs_release_path(path2);
999next:
1000 if (ptr < end) {
1001 ptr += btrfs_extent_inline_ref_size(key.type);
1002 if (ptr >= end) {
1003 WARN_ON(ptr > end);
1004 ptr = 0;
1005 end = 0;
1006 }
1007 }
1008 if (ptr >= end)
1009 path1->slots[0]++;
1010 }
1011 btrfs_release_path(path1);
1012
1013 cur->checked = 1;
1014 WARN_ON(exist);
1015
1016 /* the pending list isn't empty, take the first block to process */
1017 if (!list_empty(&list)) {
1018 edge = list_entry(list.next, struct backref_edge, list[UPPER]);
1019 list_del_init(&edge->list[UPPER]);
1020 cur = edge->node[UPPER];
1021 goto again;
1022 }
1023
1024 /*
1025 * everything goes well, connect backref nodes and insert backref nodes
1026 * into the cache.
1027 */
1028 BUG_ON(!node->checked);
1029 cowonly = node->cowonly;
1030 if (!cowonly) {
1031 rb_node = tree_insert(&cache->rb_root, node->bytenr,
1032 &node->rb_node);
1033 if (rb_node)
1034 backref_tree_panic(rb_node, -EEXIST, node->bytenr);
1035 list_add_tail(&node->lower, &cache->leaves);
1036 }
1037
1038 list_for_each_entry(edge, &node->upper, list[LOWER])
1039 list_add_tail(&edge->list[UPPER], &list);
1040
1041 while (!list_empty(&list)) {
1042 edge = list_entry(list.next, struct backref_edge, list[UPPER]);
1043 list_del_init(&edge->list[UPPER]);
1044 upper = edge->node[UPPER];
1045 if (upper->detached) {
1046 list_del(&edge->list[LOWER]);
1047 lower = edge->node[LOWER];
1048 free_backref_edge(cache, edge);
1049 if (list_empty(&lower->upper))
1050 list_add(&lower->list, &useless);
1051 continue;
1052 }
1053
1054 if (!RB_EMPTY_NODE(&upper->rb_node)) {
1055 if (upper->lowest) {
1056 list_del_init(&upper->lower);
1057 upper->lowest = 0;
1058 }
1059
1060 list_add_tail(&edge->list[UPPER], &upper->lower);
1061 continue;
1062 }
1063
1064 BUG_ON(!upper->checked);
1065 BUG_ON(cowonly != upper->cowonly);
1066 if (!cowonly) {
1067 rb_node = tree_insert(&cache->rb_root, upper->bytenr,
1068 &upper->rb_node);
1069 if (rb_node)
1070 backref_tree_panic(rb_node, -EEXIST,
1071 upper->bytenr);
1072 }
1073
1074 list_add_tail(&edge->list[UPPER], &upper->lower);
1075
1076 list_for_each_entry(edge, &upper->upper, list[LOWER])
1077 list_add_tail(&edge->list[UPPER], &list);
1078 }
1079 /*
1080 * process useless backref nodes. backref nodes for tree leaves
1081 * are deleted from the cache. backref nodes for upper level
1082 * tree blocks are left in the cache to avoid unnecessary backref
1083 * lookup.
1084 */
1085 while (!list_empty(&useless)) {
1086 upper = list_entry(useless.next, struct backref_node, list);
1087 list_del_init(&upper->list);
1088 BUG_ON(!list_empty(&upper->upper));
1089 if (upper == node)
1090 node = NULL;
1091 if (upper->lowest) {
1092 list_del_init(&upper->lower);
1093 upper->lowest = 0;
1094 }
1095 while (!list_empty(&upper->lower)) {
1096 edge = list_entry(upper->lower.next,
1097 struct backref_edge, list[UPPER]);
1098 list_del(&edge->list[UPPER]);
1099 list_del(&edge->list[LOWER]);
1100 lower = edge->node[LOWER];
1101 free_backref_edge(cache, edge);
1102
1103 if (list_empty(&lower->upper))
1104 list_add(&lower->list, &useless);
1105 }
1106 __mark_block_processed(rc, upper);
1107 if (upper->level > 0) {
1108 list_add(&upper->list, &cache->detached);
1109 upper->detached = 1;
1110 } else {
1111 rb_erase(&upper->rb_node, &cache->rb_root);
1112 free_backref_node(cache, upper);
1113 }
1114 }
1115out:
1116 btrfs_free_path(path1);
1117 btrfs_free_path(path2);
1118 if (err) {
1119 while (!list_empty(&useless)) {
1120 lower = list_entry(useless.next,
1121 struct backref_node, upper);
1122 list_del_init(&lower->upper);
1123 }
1124 upper = node;
1125 INIT_LIST_HEAD(&list);
1126 while (upper) {
1127 if (RB_EMPTY_NODE(&upper->rb_node)) {
1128 list_splice_tail(&upper->upper, &list);
1129 free_backref_node(cache, upper);
1130 }
1131
1132 if (list_empty(&list))
1133 break;
1134
1135 edge = list_entry(list.next, struct backref_edge,
1136 list[LOWER]);
1137 list_del(&edge->list[LOWER]);
1138 upper = edge->node[UPPER];
1139 free_backref_edge(cache, edge);
1140 }
1141 return ERR_PTR(err);
1142 }
1143 BUG_ON(node && node->detached);
1144 return node;
1145}
1146
1147/*
1148 * helper to add backref node for the newly created snapshot.
1149 * the backref node is created by cloning backref node that
1150 * corresponds to root of source tree
1151 */
1152static int clone_backref_node(struct btrfs_trans_handle *trans,
1153 struct reloc_control *rc,
1154 struct btrfs_root *src,
1155 struct btrfs_root *dest)
1156{
1157 struct btrfs_root *reloc_root = src->reloc_root;
1158 struct backref_cache *cache = &rc->backref_cache;
1159 struct backref_node *node = NULL;
1160 struct backref_node *new_node;
1161 struct backref_edge *edge;
1162 struct backref_edge *new_edge;
1163 struct rb_node *rb_node;
1164
1165 if (cache->last_trans > 0)
1166 update_backref_cache(trans, cache);
1167
1168 rb_node = tree_search(&cache->rb_root, src->commit_root->start);
1169 if (rb_node) {
1170 node = rb_entry(rb_node, struct backref_node, rb_node);
1171 if (node->detached)
1172 node = NULL;
1173 else
1174 BUG_ON(node->new_bytenr != reloc_root->node->start);
1175 }
1176
1177 if (!node) {
1178 rb_node = tree_search(&cache->rb_root,
1179 reloc_root->commit_root->start);
1180 if (rb_node) {
1181 node = rb_entry(rb_node, struct backref_node,
1182 rb_node);
1183 BUG_ON(node->detached);
1184 }
1185 }
1186
1187 if (!node)
1188 return 0;
1189
1190 new_node = alloc_backref_node(cache);
1191 if (!new_node)
1192 return -ENOMEM;
1193
1194 new_node->bytenr = dest->node->start;
1195 new_node->level = node->level;
1196 new_node->lowest = node->lowest;
1197 new_node->checked = 1;
1198 new_node->root = dest;
1199
1200 if (!node->lowest) {
1201 list_for_each_entry(edge, &node->lower, list[UPPER]) {
1202 new_edge = alloc_backref_edge(cache);
1203 if (!new_edge)
1204 goto fail;
1205
1206 new_edge->node[UPPER] = new_node;
1207 new_edge->node[LOWER] = edge->node[LOWER];
1208 list_add_tail(&new_edge->list[UPPER],
1209 &new_node->lower);
1210 }
1211 } else {
1212 list_add_tail(&new_node->lower, &cache->leaves);
1213 }
1214
1215 rb_node = tree_insert(&cache->rb_root, new_node->bytenr,
1216 &new_node->rb_node);
1217 if (rb_node)
1218 backref_tree_panic(rb_node, -EEXIST, new_node->bytenr);
1219
1220 if (!new_node->lowest) {
1221 list_for_each_entry(new_edge, &new_node->lower, list[UPPER]) {
1222 list_add_tail(&new_edge->list[LOWER],
1223 &new_edge->node[LOWER]->upper);
1224 }
1225 }
1226 return 0;
1227fail:
1228 while (!list_empty(&new_node->lower)) {
1229 new_edge = list_entry(new_node->lower.next,
1230 struct backref_edge, list[UPPER]);
1231 list_del(&new_edge->list[UPPER]);
1232 free_backref_edge(cache, new_edge);
1233 }
1234 free_backref_node(cache, new_node);
1235 return -ENOMEM;
1236}
1237
1238/*
1239 * helper to add 'address of tree root -> reloc tree' mapping
1240 */
1241static int __must_check __add_reloc_root(struct btrfs_root *root)
1242{
1243 struct rb_node *rb_node;
1244 struct mapping_node *node;
1245 struct reloc_control *rc = root->fs_info->reloc_ctl;
1246
1247 node = kmalloc(sizeof(*node), GFP_NOFS);
1248 if (!node)
1249 return -ENOMEM;
1250
1251 node->bytenr = root->node->start;
1252 node->data = root;
1253
1254 spin_lock(&rc->reloc_root_tree.lock);
1255 rb_node = tree_insert(&rc->reloc_root_tree.rb_root,
1256 node->bytenr, &node->rb_node);
1257 spin_unlock(&rc->reloc_root_tree.lock);
1258 if (rb_node) {
1259 btrfs_panic(root->fs_info, -EEXIST, "Duplicate root found "
1260 "for start=%llu while inserting into relocation "
1261 "tree\n", node->bytenr);
1262 kfree(node);
1263 return -EEXIST;
1264 }
1265
1266 list_add_tail(&root->root_list, &rc->reloc_roots);
1267 return 0;
1268}
1269
1270/*
1271 * helper to delete the 'address of tree root -> reloc tree'
1272 * mapping
1273 */
1274static void __del_reloc_root(struct btrfs_root *root)
1275{
1276 struct rb_node *rb_node;
1277 struct mapping_node *node = NULL;
1278 struct reloc_control *rc = root->fs_info->reloc_ctl;
1279
1280 spin_lock(&rc->reloc_root_tree.lock);
1281 rb_node = tree_search(&rc->reloc_root_tree.rb_root,
1282 root->node->start);
1283 if (rb_node) {
1284 node = rb_entry(rb_node, struct mapping_node, rb_node);
1285 rb_erase(&node->rb_node, &rc->reloc_root_tree.rb_root);
1286 }
1287 spin_unlock(&rc->reloc_root_tree.lock);
1288
1289 if (!node)
1290 return;
1291 BUG_ON((struct btrfs_root *)node->data != root);
1292
1293 spin_lock(&root->fs_info->trans_lock);
1294 list_del_init(&root->root_list);
1295 spin_unlock(&root->fs_info->trans_lock);
1296 kfree(node);
1297}
1298
1299/*
1300 * helper to update the 'address of tree root -> reloc tree'
1301 * mapping
1302 */
1303static int __update_reloc_root(struct btrfs_root *root, u64 new_bytenr)
1304{
1305 struct rb_node *rb_node;
1306 struct mapping_node *node = NULL;
1307 struct reloc_control *rc = root->fs_info->reloc_ctl;
1308
1309 spin_lock(&rc->reloc_root_tree.lock);
1310 rb_node = tree_search(&rc->reloc_root_tree.rb_root,
1311 root->node->start);
1312 if (rb_node) {
1313 node = rb_entry(rb_node, struct mapping_node, rb_node);
1314 rb_erase(&node->rb_node, &rc->reloc_root_tree.rb_root);
1315 }
1316 spin_unlock(&rc->reloc_root_tree.lock);
1317
1318 if (!node)
1319 return 0;
1320 BUG_ON((struct btrfs_root *)node->data != root);
1321
1322 spin_lock(&rc->reloc_root_tree.lock);
1323 node->bytenr = new_bytenr;
1324 rb_node = tree_insert(&rc->reloc_root_tree.rb_root,
1325 node->bytenr, &node->rb_node);
1326 spin_unlock(&rc->reloc_root_tree.lock);
1327 if (rb_node)
1328 backref_tree_panic(rb_node, -EEXIST, node->bytenr);
1329 return 0;
1330}
1331
1332static struct btrfs_root *create_reloc_root(struct btrfs_trans_handle *trans,
1333 struct btrfs_root *root, u64 objectid)
1334{
1335 struct btrfs_root *reloc_root;
1336 struct extent_buffer *eb;
1337 struct btrfs_root_item *root_item;
1338 struct btrfs_key root_key;
1339 u64 last_snap = 0;
1340 int ret;
1341
1342 root_item = kmalloc(sizeof(*root_item), GFP_NOFS);
1343 BUG_ON(!root_item);
1344
1345 root_key.objectid = BTRFS_TREE_RELOC_OBJECTID;
1346 root_key.type = BTRFS_ROOT_ITEM_KEY;
1347 root_key.offset = objectid;
1348
1349 if (root->root_key.objectid == objectid) {
1350 /* called by btrfs_init_reloc_root */
1351 ret = btrfs_copy_root(trans, root, root->commit_root, &eb,
1352 BTRFS_TREE_RELOC_OBJECTID);
1353 BUG_ON(ret);
1354
1355 last_snap = btrfs_root_last_snapshot(&root->root_item);
1356 btrfs_set_root_last_snapshot(&root->root_item,
1357 trans->transid - 1);
1358 } else {
1359 /*
1360 * called by btrfs_reloc_post_snapshot_hook.
1361 * the source tree is a reloc tree, all tree blocks
1362 * modified after it was created have RELOC flag
1363 * set in their headers. so it's OK to not update
1364 * the 'last_snapshot'.
1365 */
1366 ret = btrfs_copy_root(trans, root, root->node, &eb,
1367 BTRFS_TREE_RELOC_OBJECTID);
1368 BUG_ON(ret);
1369 }
1370
1371 memcpy(root_item, &root->root_item, sizeof(*root_item));
1372 btrfs_set_root_bytenr(root_item, eb->start);
1373 btrfs_set_root_level(root_item, btrfs_header_level(eb));
1374 btrfs_set_root_generation(root_item, trans->transid);
1375
1376 if (root->root_key.objectid == objectid) {
1377 btrfs_set_root_refs(root_item, 0);
1378 memset(&root_item->drop_progress, 0,
1379 sizeof(struct btrfs_disk_key));
1380 root_item->drop_level = 0;
1381 /*
1382 * abuse rtransid, it is safe because it is impossible to
1383 * receive data into a relocation tree.
1384 */
1385 btrfs_set_root_rtransid(root_item, last_snap);
1386 btrfs_set_root_otransid(root_item, trans->transid);
1387 }
1388
1389 btrfs_tree_unlock(eb);
1390 free_extent_buffer(eb);
1391
1392 ret = btrfs_insert_root(trans, root->fs_info->tree_root,
1393 &root_key, root_item);
1394 BUG_ON(ret);
1395 kfree(root_item);
1396
1397 reloc_root = btrfs_read_fs_root(root->fs_info->tree_root, &root_key);
1398 BUG_ON(IS_ERR(reloc_root));
1399 reloc_root->last_trans = trans->transid;
1400 return reloc_root;
1401}
1402
1403/*
1404 * create reloc tree for a given fs tree. reloc tree is just a
1405 * snapshot of the fs tree with special root objectid.
1406 */
1407int btrfs_init_reloc_root(struct btrfs_trans_handle *trans,
1408 struct btrfs_root *root)
1409{
1410 struct btrfs_root *reloc_root;
1411 struct reloc_control *rc = root->fs_info->reloc_ctl;
1412 struct btrfs_block_rsv *rsv;
1413 int clear_rsv = 0;
1414 int ret;
1415
1416 if (root->reloc_root) {
1417 reloc_root = root->reloc_root;
1418 reloc_root->last_trans = trans->transid;
1419 return 0;
1420 }
1421
1422 if (!rc || !rc->create_reloc_tree ||
1423 root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID)
1424 return 0;
1425
1426 if (!trans->reloc_reserved) {
1427 rsv = trans->block_rsv;
1428 trans->block_rsv = rc->block_rsv;
1429 clear_rsv = 1;
1430 }
1431 reloc_root = create_reloc_root(trans, root, root->root_key.objectid);
1432 if (clear_rsv)
1433 trans->block_rsv = rsv;
1434
1435 ret = __add_reloc_root(reloc_root);
1436 BUG_ON(ret < 0);
1437 root->reloc_root = reloc_root;
1438 return 0;
1439}
1440
1441/*
1442 * update root item of reloc tree
1443 */
1444int btrfs_update_reloc_root(struct btrfs_trans_handle *trans,
1445 struct btrfs_root *root)
1446{
1447 struct btrfs_root *reloc_root;
1448 struct btrfs_root_item *root_item;
1449 int ret;
1450
1451 if (!root->reloc_root)
1452 goto out;
1453
1454 reloc_root = root->reloc_root;
1455 root_item = &reloc_root->root_item;
1456
1457 if (root->fs_info->reloc_ctl->merge_reloc_tree &&
1458 btrfs_root_refs(root_item) == 0) {
1459 root->reloc_root = NULL;
1460 __del_reloc_root(reloc_root);
1461 }
1462
1463 if (reloc_root->commit_root != reloc_root->node) {
1464 btrfs_set_root_node(root_item, reloc_root->node);
1465 free_extent_buffer(reloc_root->commit_root);
1466 reloc_root->commit_root = btrfs_root_node(reloc_root);
1467 }
1468
1469 ret = btrfs_update_root(trans, root->fs_info->tree_root,
1470 &reloc_root->root_key, root_item);
1471 BUG_ON(ret);
1472
1473out:
1474 return 0;
1475}
1476
1477/*
1478 * helper to find first cached inode with inode number >= objectid
1479 * in a subvolume
1480 */
1481static struct inode *find_next_inode(struct btrfs_root *root, u64 objectid)
1482{
1483 struct rb_node *node;
1484 struct rb_node *prev;
1485 struct btrfs_inode *entry;
1486 struct inode *inode;
1487
1488 spin_lock(&root->inode_lock);
1489again:
1490 node = root->inode_tree.rb_node;
1491 prev = NULL;
1492 while (node) {
1493 prev = node;
1494 entry = rb_entry(node, struct btrfs_inode, rb_node);
1495
1496 if (objectid < btrfs_ino(&entry->vfs_inode))
1497 node = node->rb_left;
1498 else if (objectid > btrfs_ino(&entry->vfs_inode))
1499 node = node->rb_right;
1500 else
1501 break;
1502 }
1503 if (!node) {
1504 while (prev) {
1505 entry = rb_entry(prev, struct btrfs_inode, rb_node);
1506 if (objectid <= btrfs_ino(&entry->vfs_inode)) {
1507 node = prev;
1508 break;
1509 }
1510 prev = rb_next(prev);
1511 }
1512 }
1513 while (node) {
1514 entry = rb_entry(node, struct btrfs_inode, rb_node);
1515 inode = igrab(&entry->vfs_inode);
1516 if (inode) {
1517 spin_unlock(&root->inode_lock);
1518 return inode;
1519 }
1520
1521 objectid = btrfs_ino(&entry->vfs_inode) + 1;
1522 if (cond_resched_lock(&root->inode_lock))
1523 goto again;
1524
1525 node = rb_next(node);
1526 }
1527 spin_unlock(&root->inode_lock);
1528 return NULL;
1529}
1530
1531static int in_block_group(u64 bytenr,
1532 struct btrfs_block_group_cache *block_group)
1533{
1534 if (bytenr >= block_group->key.objectid &&
1535 bytenr < block_group->key.objectid + block_group->key.offset)
1536 return 1;
1537 return 0;
1538}
1539
1540/*
1541 * get new location of data
1542 */
1543static int get_new_location(struct inode *reloc_inode, u64 *new_bytenr,
1544 u64 bytenr, u64 num_bytes)
1545{
1546 struct btrfs_root *root = BTRFS_I(reloc_inode)->root;
1547 struct btrfs_path *path;
1548 struct btrfs_file_extent_item *fi;
1549 struct extent_buffer *leaf;
1550 int ret;
1551
1552 path = btrfs_alloc_path();
1553 if (!path)
1554 return -ENOMEM;
1555
1556 bytenr -= BTRFS_I(reloc_inode)->index_cnt;
1557 ret = btrfs_lookup_file_extent(NULL, root, path, btrfs_ino(reloc_inode),
1558 bytenr, 0);
1559 if (ret < 0)
1560 goto out;
1561 if (ret > 0) {
1562 ret = -ENOENT;
1563 goto out;
1564 }
1565
1566 leaf = path->nodes[0];
1567 fi = btrfs_item_ptr(leaf, path->slots[0],
1568 struct btrfs_file_extent_item);
1569
1570 BUG_ON(btrfs_file_extent_offset(leaf, fi) ||
1571 btrfs_file_extent_compression(leaf, fi) ||
1572 btrfs_file_extent_encryption(leaf, fi) ||
1573 btrfs_file_extent_other_encoding(leaf, fi));
1574
1575 if (num_bytes != btrfs_file_extent_disk_num_bytes(leaf, fi)) {
1576 ret = -EINVAL;
1577 goto out;
1578 }
1579
1580 *new_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
1581 ret = 0;
1582out:
1583 btrfs_free_path(path);
1584 return ret;
1585}
1586
1587/*
1588 * update file extent items in the tree leaf to point to
1589 * the new locations.
1590 */
1591static noinline_for_stack
1592int replace_file_extents(struct btrfs_trans_handle *trans,
1593 struct reloc_control *rc,
1594 struct btrfs_root *root,
1595 struct extent_buffer *leaf)
1596{
1597 struct btrfs_key key;
1598 struct btrfs_file_extent_item *fi;
1599 struct inode *inode = NULL;
1600 u64 parent;
1601 u64 bytenr;
1602 u64 new_bytenr = 0;
1603 u64 num_bytes;
1604 u64 end;
1605 u32 nritems;
1606 u32 i;
1607 int ret = 0;
1608 int first = 1;
1609 int dirty = 0;
1610
1611 if (rc->stage != UPDATE_DATA_PTRS)
1612 return 0;
1613
1614 /* reloc trees always use full backref */
1615 if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID)
1616 parent = leaf->start;
1617 else
1618 parent = 0;
1619
1620 nritems = btrfs_header_nritems(leaf);
1621 for (i = 0; i < nritems; i++) {
1622 cond_resched();
1623 btrfs_item_key_to_cpu(leaf, &key, i);
1624 if (key.type != BTRFS_EXTENT_DATA_KEY)
1625 continue;
1626 fi = btrfs_item_ptr(leaf, i, struct btrfs_file_extent_item);
1627 if (btrfs_file_extent_type(leaf, fi) ==
1628 BTRFS_FILE_EXTENT_INLINE)
1629 continue;
1630 bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
1631 num_bytes = btrfs_file_extent_disk_num_bytes(leaf, fi);
1632 if (bytenr == 0)
1633 continue;
1634 if (!in_block_group(bytenr, rc->block_group))
1635 continue;
1636
1637 /*
1638 * if we are modifying block in fs tree, wait for readpage
1639 * to complete and drop the extent cache
1640 */
1641 if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) {
1642 if (first) {
1643 inode = find_next_inode(root, key.objectid);
1644 first = 0;
1645 } else if (inode && btrfs_ino(inode) < key.objectid) {
1646 btrfs_add_delayed_iput(inode);
1647 inode = find_next_inode(root, key.objectid);
1648 }
1649 if (inode && btrfs_ino(inode) == key.objectid) {
1650 end = key.offset +
1651 btrfs_file_extent_num_bytes(leaf, fi);
1652 WARN_ON(!IS_ALIGNED(key.offset,
1653 root->sectorsize));
1654 WARN_ON(!IS_ALIGNED(end, root->sectorsize));
1655 end--;
1656 ret = try_lock_extent(&BTRFS_I(inode)->io_tree,
1657 key.offset, end);
1658 if (!ret)
1659 continue;
1660
1661 btrfs_drop_extent_cache(inode, key.offset, end,
1662 1);
1663 unlock_extent(&BTRFS_I(inode)->io_tree,
1664 key.offset, end);
1665 }
1666 }
1667
1668 ret = get_new_location(rc->data_inode, &new_bytenr,
1669 bytenr, num_bytes);
1670 if (ret) {
1671 /*
1672 * Don't have to abort since we've not changed anything
1673 * in the file extent yet.
1674 */
1675 break;
1676 }
1677
1678 btrfs_set_file_extent_disk_bytenr(leaf, fi, new_bytenr);
1679 dirty = 1;
1680
1681 key.offset -= btrfs_file_extent_offset(leaf, fi);
1682 ret = btrfs_inc_extent_ref(trans, root, new_bytenr,
1683 num_bytes, parent,
1684 btrfs_header_owner(leaf),
1685 key.objectid, key.offset, 1);
1686 if (ret) {
1687 btrfs_abort_transaction(trans, root, ret);
1688 break;
1689 }
1690
1691 ret = btrfs_free_extent(trans, root, bytenr, num_bytes,
1692 parent, btrfs_header_owner(leaf),
1693 key.objectid, key.offset, 1);
1694 if (ret) {
1695 btrfs_abort_transaction(trans, root, ret);
1696 break;
1697 }
1698 }
1699 if (dirty)
1700 btrfs_mark_buffer_dirty(leaf);
1701 if (inode)
1702 btrfs_add_delayed_iput(inode);
1703 return ret;
1704}
1705
1706static noinline_for_stack
1707int memcmp_node_keys(struct extent_buffer *eb, int slot,
1708 struct btrfs_path *path, int level)
1709{
1710 struct btrfs_disk_key key1;
1711 struct btrfs_disk_key key2;
1712 btrfs_node_key(eb, &key1, slot);
1713 btrfs_node_key(path->nodes[level], &key2, path->slots[level]);
1714 return memcmp(&key1, &key2, sizeof(key1));
1715}
1716
1717/*
1718 * try to replace tree blocks in fs tree with the new blocks
1719 * in reloc tree. tree blocks haven't been modified since the
1720 * reloc tree was create can be replaced.
1721 *
1722 * if a block was replaced, level of the block + 1 is returned.
1723 * if no block got replaced, 0 is returned. if there are other
1724 * errors, a negative error number is returned.
1725 */
1726static noinline_for_stack
1727int replace_path(struct btrfs_trans_handle *trans,
1728 struct btrfs_root *dest, struct btrfs_root *src,
1729 struct btrfs_path *path, struct btrfs_key *next_key,
1730 int lowest_level, int max_level)
1731{
1732 struct extent_buffer *eb;
1733 struct extent_buffer *parent;
1734 struct btrfs_key key;
1735 u64 old_bytenr;
1736 u64 new_bytenr;
1737 u64 old_ptr_gen;
1738 u64 new_ptr_gen;
1739 u64 last_snapshot;
1740 u32 blocksize;
1741 int cow = 0;
1742 int level;
1743 int ret;
1744 int slot;
1745
1746 BUG_ON(src->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID);
1747 BUG_ON(dest->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID);
1748
1749 last_snapshot = btrfs_root_last_snapshot(&src->root_item);
1750again:
1751 slot = path->slots[lowest_level];
1752 btrfs_node_key_to_cpu(path->nodes[lowest_level], &key, slot);
1753
1754 eb = btrfs_lock_root_node(dest);
1755 btrfs_set_lock_blocking(eb);
1756 level = btrfs_header_level(eb);
1757
1758 if (level < lowest_level) {
1759 btrfs_tree_unlock(eb);
1760 free_extent_buffer(eb);
1761 return 0;
1762 }
1763
1764 if (cow) {
1765 ret = btrfs_cow_block(trans, dest, eb, NULL, 0, &eb);
1766 BUG_ON(ret);
1767 }
1768 btrfs_set_lock_blocking(eb);
1769
1770 if (next_key) {
1771 next_key->objectid = (u64)-1;
1772 next_key->type = (u8)-1;
1773 next_key->offset = (u64)-1;
1774 }
1775
1776 parent = eb;
1777 while (1) {
1778 level = btrfs_header_level(parent);
1779 BUG_ON(level < lowest_level);
1780
1781 ret = btrfs_bin_search(parent, &key, level, &slot);
1782 if (ret && slot > 0)
1783 slot--;
1784
1785 if (next_key && slot + 1 < btrfs_header_nritems(parent))
1786 btrfs_node_key_to_cpu(parent, next_key, slot + 1);
1787
1788 old_bytenr = btrfs_node_blockptr(parent, slot);
1789 blocksize = btrfs_level_size(dest, level - 1);
1790 old_ptr_gen = btrfs_node_ptr_generation(parent, slot);
1791
1792 if (level <= max_level) {
1793 eb = path->nodes[level];
1794 new_bytenr = btrfs_node_blockptr(eb,
1795 path->slots[level]);
1796 new_ptr_gen = btrfs_node_ptr_generation(eb,
1797 path->slots[level]);
1798 } else {
1799 new_bytenr = 0;
1800 new_ptr_gen = 0;
1801 }
1802
1803 if (WARN_ON(new_bytenr > 0 && new_bytenr == old_bytenr)) {
1804 ret = level;
1805 break;
1806 }
1807
1808 if (new_bytenr == 0 || old_ptr_gen > last_snapshot ||
1809 memcmp_node_keys(parent, slot, path, level)) {
1810 if (level <= lowest_level) {
1811 ret = 0;
1812 break;
1813 }
1814
1815 eb = read_tree_block(dest, old_bytenr, blocksize,
1816 old_ptr_gen);
1817 if (!eb || !extent_buffer_uptodate(eb)) {
1818 ret = (!eb) ? -ENOMEM : -EIO;
1819 free_extent_buffer(eb);
1820 break;
1821 }
1822 btrfs_tree_lock(eb);
1823 if (cow) {
1824 ret = btrfs_cow_block(trans, dest, eb, parent,
1825 slot, &eb);
1826 BUG_ON(ret);
1827 }
1828 btrfs_set_lock_blocking(eb);
1829
1830 btrfs_tree_unlock(parent);
1831 free_extent_buffer(parent);
1832
1833 parent = eb;
1834 continue;
1835 }
1836
1837 if (!cow) {
1838 btrfs_tree_unlock(parent);
1839 free_extent_buffer(parent);
1840 cow = 1;
1841 goto again;
1842 }
1843
1844 btrfs_node_key_to_cpu(path->nodes[level], &key,
1845 path->slots[level]);
1846 btrfs_release_path(path);
1847
1848 path->lowest_level = level;
1849 ret = btrfs_search_slot(trans, src, &key, path, 0, 1);
1850 path->lowest_level = 0;
1851 BUG_ON(ret);
1852
1853 /*
1854 * swap blocks in fs tree and reloc tree.
1855 */
1856 btrfs_set_node_blockptr(parent, slot, new_bytenr);
1857 btrfs_set_node_ptr_generation(parent, slot, new_ptr_gen);
1858 btrfs_mark_buffer_dirty(parent);
1859
1860 btrfs_set_node_blockptr(path->nodes[level],
1861 path->slots[level], old_bytenr);
1862 btrfs_set_node_ptr_generation(path->nodes[level],
1863 path->slots[level], old_ptr_gen);
1864 btrfs_mark_buffer_dirty(path->nodes[level]);
1865
1866 ret = btrfs_inc_extent_ref(trans, src, old_bytenr, blocksize,
1867 path->nodes[level]->start,
1868 src->root_key.objectid, level - 1, 0,
1869 1);
1870 BUG_ON(ret);
1871 ret = btrfs_inc_extent_ref(trans, dest, new_bytenr, blocksize,
1872 0, dest->root_key.objectid, level - 1,
1873 0, 1);
1874 BUG_ON(ret);
1875
1876 ret = btrfs_free_extent(trans, src, new_bytenr, blocksize,
1877 path->nodes[level]->start,
1878 src->root_key.objectid, level - 1, 0,
1879 1);
1880 BUG_ON(ret);
1881
1882 ret = btrfs_free_extent(trans, dest, old_bytenr, blocksize,
1883 0, dest->root_key.objectid, level - 1,
1884 0, 1);
1885 BUG_ON(ret);
1886
1887 btrfs_unlock_up_safe(path, 0);
1888
1889 ret = level;
1890 break;
1891 }
1892 btrfs_tree_unlock(parent);
1893 free_extent_buffer(parent);
1894 return ret;
1895}
1896
1897/*
1898 * helper to find next relocated block in reloc tree
1899 */
1900static noinline_for_stack
1901int walk_up_reloc_tree(struct btrfs_root *root, struct btrfs_path *path,
1902 int *level)
1903{
1904 struct extent_buffer *eb;
1905 int i;
1906 u64 last_snapshot;
1907 u32 nritems;
1908
1909 last_snapshot = btrfs_root_last_snapshot(&root->root_item);
1910
1911 for (i = 0; i < *level; i++) {
1912 free_extent_buffer(path->nodes[i]);
1913 path->nodes[i] = NULL;
1914 }
1915
1916 for (i = *level; i < BTRFS_MAX_LEVEL && path->nodes[i]; i++) {
1917 eb = path->nodes[i];
1918 nritems = btrfs_header_nritems(eb);
1919 while (path->slots[i] + 1 < nritems) {
1920 path->slots[i]++;
1921 if (btrfs_node_ptr_generation(eb, path->slots[i]) <=
1922 last_snapshot)
1923 continue;
1924
1925 *level = i;
1926 return 0;
1927 }
1928 free_extent_buffer(path->nodes[i]);
1929 path->nodes[i] = NULL;
1930 }
1931 return 1;
1932}
1933
1934/*
1935 * walk down reloc tree to find relocated block of lowest level
1936 */
1937static noinline_for_stack
1938int walk_down_reloc_tree(struct btrfs_root *root, struct btrfs_path *path,
1939 int *level)
1940{
1941 struct extent_buffer *eb = NULL;
1942 int i;
1943 u64 bytenr;
1944 u64 ptr_gen = 0;
1945 u64 last_snapshot;
1946 u32 blocksize;
1947 u32 nritems;
1948
1949 last_snapshot = btrfs_root_last_snapshot(&root->root_item);
1950
1951 for (i = *level; i > 0; i--) {
1952 eb = path->nodes[i];
1953 nritems = btrfs_header_nritems(eb);
1954 while (path->slots[i] < nritems) {
1955 ptr_gen = btrfs_node_ptr_generation(eb, path->slots[i]);
1956 if (ptr_gen > last_snapshot)
1957 break;
1958 path->slots[i]++;
1959 }
1960 if (path->slots[i] >= nritems) {
1961 if (i == *level)
1962 break;
1963 *level = i + 1;
1964 return 0;
1965 }
1966 if (i == 1) {
1967 *level = i;
1968 return 0;
1969 }
1970
1971 bytenr = btrfs_node_blockptr(eb, path->slots[i]);
1972 blocksize = btrfs_level_size(root, i - 1);
1973 eb = read_tree_block(root, bytenr, blocksize, ptr_gen);
1974 if (!eb || !extent_buffer_uptodate(eb)) {
1975 free_extent_buffer(eb);
1976 return -EIO;
1977 }
1978 BUG_ON(btrfs_header_level(eb) != i - 1);
1979 path->nodes[i - 1] = eb;
1980 path->slots[i - 1] = 0;
1981 }
1982 return 1;
1983}
1984
1985/*
1986 * invalidate extent cache for file extents whose key in range of
1987 * [min_key, max_key)
1988 */
1989static int invalidate_extent_cache(struct btrfs_root *root,
1990 struct btrfs_key *min_key,
1991 struct btrfs_key *max_key)
1992{
1993 struct inode *inode = NULL;
1994 u64 objectid;
1995 u64 start, end;
1996 u64 ino;
1997
1998 objectid = min_key->objectid;
1999 while (1) {
2000 cond_resched();
2001 iput(inode);
2002
2003 if (objectid > max_key->objectid)
2004 break;
2005
2006 inode = find_next_inode(root, objectid);
2007 if (!inode)
2008 break;
2009 ino = btrfs_ino(inode);
2010
2011 if (ino > max_key->objectid) {
2012 iput(inode);
2013 break;
2014 }
2015
2016 objectid = ino + 1;
2017 if (!S_ISREG(inode->i_mode))
2018 continue;
2019
2020 if (unlikely(min_key->objectid == ino)) {
2021 if (min_key->type > BTRFS_EXTENT_DATA_KEY)
2022 continue;
2023 if (min_key->type < BTRFS_EXTENT_DATA_KEY)
2024 start = 0;
2025 else {
2026 start = min_key->offset;
2027 WARN_ON(!IS_ALIGNED(start, root->sectorsize));
2028 }
2029 } else {
2030 start = 0;
2031 }
2032
2033 if (unlikely(max_key->objectid == ino)) {
2034 if (max_key->type < BTRFS_EXTENT_DATA_KEY)
2035 continue;
2036 if (max_key->type > BTRFS_EXTENT_DATA_KEY) {
2037 end = (u64)-1;
2038 } else {
2039 if (max_key->offset == 0)
2040 continue;
2041 end = max_key->offset;
2042 WARN_ON(!IS_ALIGNED(end, root->sectorsize));
2043 end--;
2044 }
2045 } else {
2046 end = (u64)-1;
2047 }
2048
2049 /* the lock_extent waits for readpage to complete */
2050 lock_extent(&BTRFS_I(inode)->io_tree, start, end);
2051 btrfs_drop_extent_cache(inode, start, end, 1);
2052 unlock_extent(&BTRFS_I(inode)->io_tree, start, end);
2053 }
2054 return 0;
2055}
2056
2057static int find_next_key(struct btrfs_path *path, int level,
2058 struct btrfs_key *key)
2059
2060{
2061 while (level < BTRFS_MAX_LEVEL) {
2062 if (!path->nodes[level])
2063 break;
2064 if (path->slots[level] + 1 <
2065 btrfs_header_nritems(path->nodes[level])) {
2066 btrfs_node_key_to_cpu(path->nodes[level], key,
2067 path->slots[level] + 1);
2068 return 0;
2069 }
2070 level++;
2071 }
2072 return 1;
2073}
2074
2075/*
2076 * merge the relocated tree blocks in reloc tree with corresponding
2077 * fs tree.
2078 */
2079static noinline_for_stack int merge_reloc_root(struct reloc_control *rc,
2080 struct btrfs_root *root)
2081{
2082 LIST_HEAD(inode_list);
2083 struct btrfs_key key;
2084 struct btrfs_key next_key;
2085 struct btrfs_trans_handle *trans = NULL;
2086 struct btrfs_root *reloc_root;
2087 struct btrfs_root_item *root_item;
2088 struct btrfs_path *path;
2089 struct extent_buffer *leaf;
2090 int level;
2091 int max_level;
2092 int replaced = 0;
2093 int ret;
2094 int err = 0;
2095 u32 min_reserved;
2096
2097 path = btrfs_alloc_path();
2098 if (!path)
2099 return -ENOMEM;
2100 path->reada = 1;
2101
2102 reloc_root = root->reloc_root;
2103 root_item = &reloc_root->root_item;
2104
2105 if (btrfs_disk_key_objectid(&root_item->drop_progress) == 0) {
2106 level = btrfs_root_level(root_item);
2107 extent_buffer_get(reloc_root->node);
2108 path->nodes[level] = reloc_root->node;
2109 path->slots[level] = 0;
2110 } else {
2111 btrfs_disk_key_to_cpu(&key, &root_item->drop_progress);
2112
2113 level = root_item->drop_level;
2114 BUG_ON(level == 0);
2115 path->lowest_level = level;
2116 ret = btrfs_search_slot(NULL, reloc_root, &key, path, 0, 0);
2117 path->lowest_level = 0;
2118 if (ret < 0) {
2119 btrfs_free_path(path);
2120 return ret;
2121 }
2122
2123 btrfs_node_key_to_cpu(path->nodes[level], &next_key,
2124 path->slots[level]);
2125 WARN_ON(memcmp(&key, &next_key, sizeof(key)));
2126
2127 btrfs_unlock_up_safe(path, 0);
2128 }
2129
2130 min_reserved = root->nodesize * (BTRFS_MAX_LEVEL - 1) * 2;
2131 memset(&next_key, 0, sizeof(next_key));
2132
2133 while (1) {
2134 ret = btrfs_block_rsv_refill(root, rc->block_rsv, min_reserved,
2135 BTRFS_RESERVE_FLUSH_ALL);
2136 if (ret) {
2137 err = ret;
2138 goto out;
2139 }
2140 trans = btrfs_start_transaction(root, 0);
2141 if (IS_ERR(trans)) {
2142 err = PTR_ERR(trans);
2143 trans = NULL;
2144 goto out;
2145 }
2146 trans->block_rsv = rc->block_rsv;
2147
2148 replaced = 0;
2149 max_level = level;
2150
2151 ret = walk_down_reloc_tree(reloc_root, path, &level);
2152 if (ret < 0) {
2153 err = ret;
2154 goto out;
2155 }
2156 if (ret > 0)
2157 break;
2158
2159 if (!find_next_key(path, level, &key) &&
2160 btrfs_comp_cpu_keys(&next_key, &key) >= 0) {
2161 ret = 0;
2162 } else {
2163 ret = replace_path(trans, root, reloc_root, path,
2164 &next_key, level, max_level);
2165 }
2166 if (ret < 0) {
2167 err = ret;
2168 goto out;
2169 }
2170
2171 if (ret > 0) {
2172 level = ret;
2173 btrfs_node_key_to_cpu(path->nodes[level], &key,
2174 path->slots[level]);
2175 replaced = 1;
2176 }
2177
2178 ret = walk_up_reloc_tree(reloc_root, path, &level);
2179 if (ret > 0)
2180 break;
2181
2182 BUG_ON(level == 0);
2183 /*
2184 * save the merging progress in the drop_progress.
2185 * this is OK since root refs == 1 in this case.
2186 */
2187 btrfs_node_key(path->nodes[level], &root_item->drop_progress,
2188 path->slots[level]);
2189 root_item->drop_level = level;
2190
2191 btrfs_end_transaction_throttle(trans, root);
2192 trans = NULL;
2193
2194 btrfs_btree_balance_dirty(root);
2195
2196 if (replaced && rc->stage == UPDATE_DATA_PTRS)
2197 invalidate_extent_cache(root, &key, &next_key);
2198 }
2199
2200 /*
2201 * handle the case only one block in the fs tree need to be
2202 * relocated and the block is tree root.
2203 */
2204 leaf = btrfs_lock_root_node(root);
2205 ret = btrfs_cow_block(trans, root, leaf, NULL, 0, &leaf);
2206 btrfs_tree_unlock(leaf);
2207 free_extent_buffer(leaf);
2208 if (ret < 0)
2209 err = ret;
2210out:
2211 btrfs_free_path(path);
2212
2213 if (err == 0) {
2214 memset(&root_item->drop_progress, 0,
2215 sizeof(root_item->drop_progress));
2216 root_item->drop_level = 0;
2217 btrfs_set_root_refs(root_item, 0);
2218 btrfs_update_reloc_root(trans, root);
2219 }
2220
2221 if (trans)
2222 btrfs_end_transaction_throttle(trans, root);
2223
2224 btrfs_btree_balance_dirty(root);
2225
2226 if (replaced && rc->stage == UPDATE_DATA_PTRS)
2227 invalidate_extent_cache(root, &key, &next_key);
2228
2229 return err;
2230}
2231
2232static noinline_for_stack
2233int prepare_to_merge(struct reloc_control *rc, int err)
2234{
2235 struct btrfs_root *root = rc->extent_root;
2236 struct btrfs_root *reloc_root;
2237 struct btrfs_trans_handle *trans;
2238 LIST_HEAD(reloc_roots);
2239 u64 num_bytes = 0;
2240 int ret;
2241
2242 mutex_lock(&root->fs_info->reloc_mutex);
2243 rc->merging_rsv_size += root->nodesize * (BTRFS_MAX_LEVEL - 1) * 2;
2244 rc->merging_rsv_size += rc->nodes_relocated * 2;
2245 mutex_unlock(&root->fs_info->reloc_mutex);
2246
2247again:
2248 if (!err) {
2249 num_bytes = rc->merging_rsv_size;
2250 ret = btrfs_block_rsv_add(root, rc->block_rsv, num_bytes,
2251 BTRFS_RESERVE_FLUSH_ALL);
2252 if (ret)
2253 err = ret;
2254 }
2255
2256 trans = btrfs_join_transaction(rc->extent_root);
2257 if (IS_ERR(trans)) {
2258 if (!err)
2259 btrfs_block_rsv_release(rc->extent_root,
2260 rc->block_rsv, num_bytes);
2261 return PTR_ERR(trans);
2262 }
2263
2264 if (!err) {
2265 if (num_bytes != rc->merging_rsv_size) {
2266 btrfs_end_transaction(trans, rc->extent_root);
2267 btrfs_block_rsv_release(rc->extent_root,
2268 rc->block_rsv, num_bytes);
2269 goto again;
2270 }
2271 }
2272
2273 rc->merge_reloc_tree = 1;
2274
2275 while (!list_empty(&rc->reloc_roots)) {
2276 reloc_root = list_entry(rc->reloc_roots.next,
2277 struct btrfs_root, root_list);
2278 list_del_init(&reloc_root->root_list);
2279
2280 root = read_fs_root(reloc_root->fs_info,
2281 reloc_root->root_key.offset);
2282 BUG_ON(IS_ERR(root));
2283 BUG_ON(root->reloc_root != reloc_root);
2284
2285 /*
2286 * set reference count to 1, so btrfs_recover_relocation
2287 * knows it should resumes merging
2288 */
2289 if (!err)
2290 btrfs_set_root_refs(&reloc_root->root_item, 1);
2291 btrfs_update_reloc_root(trans, root);
2292
2293 list_add(&reloc_root->root_list, &reloc_roots);
2294 }
2295
2296 list_splice(&reloc_roots, &rc->reloc_roots);
2297
2298 if (!err)
2299 btrfs_commit_transaction(trans, rc->extent_root);
2300 else
2301 btrfs_end_transaction(trans, rc->extent_root);
2302 return err;
2303}
2304
2305static noinline_for_stack
2306void free_reloc_roots(struct list_head *list)
2307{
2308 struct btrfs_root *reloc_root;
2309
2310 while (!list_empty(list)) {
2311 reloc_root = list_entry(list->next, struct btrfs_root,
2312 root_list);
2313 __del_reloc_root(reloc_root);
2314 }
2315}
2316
2317static noinline_for_stack
2318int merge_reloc_roots(struct reloc_control *rc)
2319{
2320 struct btrfs_root *root;
2321 struct btrfs_root *reloc_root;
2322 u64 last_snap;
2323 u64 otransid;
2324 u64 objectid;
2325 LIST_HEAD(reloc_roots);
2326 int found = 0;
2327 int ret = 0;
2328again:
2329 root = rc->extent_root;
2330
2331 /*
2332 * this serializes us with btrfs_record_root_in_transaction,
2333 * we have to make sure nobody is in the middle of
2334 * adding their roots to the list while we are
2335 * doing this splice
2336 */
2337 mutex_lock(&root->fs_info->reloc_mutex);
2338 list_splice_init(&rc->reloc_roots, &reloc_roots);
2339 mutex_unlock(&root->fs_info->reloc_mutex);
2340
2341 while (!list_empty(&reloc_roots)) {
2342 found = 1;
2343 reloc_root = list_entry(reloc_roots.next,
2344 struct btrfs_root, root_list);
2345
2346 if (btrfs_root_refs(&reloc_root->root_item) > 0) {
2347 root = read_fs_root(reloc_root->fs_info,
2348 reloc_root->root_key.offset);
2349 BUG_ON(IS_ERR(root));
2350 BUG_ON(root->reloc_root != reloc_root);
2351
2352 ret = merge_reloc_root(rc, root);
2353 if (ret) {
2354 if (list_empty(&reloc_root->root_list))
2355 list_add_tail(&reloc_root->root_list,
2356 &reloc_roots);
2357 goto out;
2358 }
2359 } else {
2360 list_del_init(&reloc_root->root_list);
2361 }
2362
2363 /*
2364 * we keep the old last snapshod transid in rtranid when we
2365 * created the relocation tree.
2366 */
2367 last_snap = btrfs_root_rtransid(&reloc_root->root_item);
2368 otransid = btrfs_root_otransid(&reloc_root->root_item);
2369 objectid = reloc_root->root_key.offset;
2370
2371 ret = btrfs_drop_snapshot(reloc_root, rc->block_rsv, 0, 1);
2372 if (ret < 0) {
2373 if (list_empty(&reloc_root->root_list))
2374 list_add_tail(&reloc_root->root_list,
2375 &reloc_roots);
2376 goto out;
2377 }
2378 }
2379
2380 if (found) {
2381 found = 0;
2382 goto again;
2383 }
2384out:
2385 if (ret) {
2386 btrfs_std_error(root->fs_info, ret);
2387 if (!list_empty(&reloc_roots))
2388 free_reloc_roots(&reloc_roots);
2389
2390 /* new reloc root may be added */
2391 mutex_lock(&root->fs_info->reloc_mutex);
2392 list_splice_init(&rc->reloc_roots, &reloc_roots);
2393 mutex_unlock(&root->fs_info->reloc_mutex);
2394 if (!list_empty(&reloc_roots))
2395 free_reloc_roots(&reloc_roots);
2396 }
2397
2398 BUG_ON(!RB_EMPTY_ROOT(&rc->reloc_root_tree.rb_root));
2399 return ret;
2400}
2401
2402static void free_block_list(struct rb_root *blocks)
2403{
2404 struct tree_block *block;
2405 struct rb_node *rb_node;
2406 while ((rb_node = rb_first(blocks))) {
2407 block = rb_entry(rb_node, struct tree_block, rb_node);
2408 rb_erase(rb_node, blocks);
2409 kfree(block);
2410 }
2411}
2412
2413static int record_reloc_root_in_trans(struct btrfs_trans_handle *trans,
2414 struct btrfs_root *reloc_root)
2415{
2416 struct btrfs_root *root;
2417
2418 if (reloc_root->last_trans == trans->transid)
2419 return 0;
2420
2421 root = read_fs_root(reloc_root->fs_info, reloc_root->root_key.offset);
2422 BUG_ON(IS_ERR(root));
2423 BUG_ON(root->reloc_root != reloc_root);
2424
2425 return btrfs_record_root_in_trans(trans, root);
2426}
2427
2428static noinline_for_stack
2429struct btrfs_root *select_reloc_root(struct btrfs_trans_handle *trans,
2430 struct reloc_control *rc,
2431 struct backref_node *node,
2432 struct backref_edge *edges[])
2433{
2434 struct backref_node *next;
2435 struct btrfs_root *root;
2436 int index = 0;
2437
2438 next = node;
2439 while (1) {
2440 cond_resched();
2441 next = walk_up_backref(next, edges, &index);
2442 root = next->root;
2443 BUG_ON(!root);
2444 BUG_ON(!root->ref_cows);
2445
2446 if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID) {
2447 record_reloc_root_in_trans(trans, root);
2448 break;
2449 }
2450
2451 btrfs_record_root_in_trans(trans, root);
2452 root = root->reloc_root;
2453
2454 if (next->new_bytenr != root->node->start) {
2455 BUG_ON(next->new_bytenr);
2456 BUG_ON(!list_empty(&next->list));
2457 next->new_bytenr = root->node->start;
2458 next->root = root;
2459 list_add_tail(&next->list,
2460 &rc->backref_cache.changed);
2461 __mark_block_processed(rc, next);
2462 break;
2463 }
2464
2465 WARN_ON(1);
2466 root = NULL;
2467 next = walk_down_backref(edges, &index);
2468 if (!next || next->level <= node->level)
2469 break;
2470 }
2471 if (!root)
2472 return NULL;
2473
2474 next = node;
2475 /* setup backref node path for btrfs_reloc_cow_block */
2476 while (1) {
2477 rc->backref_cache.path[next->level] = next;
2478 if (--index < 0)
2479 break;
2480 next = edges[index]->node[UPPER];
2481 }
2482 return root;
2483}
2484
2485/*
2486 * select a tree root for relocation. return NULL if the block
2487 * is reference counted. we should use do_relocation() in this
2488 * case. return a tree root pointer if the block isn't reference
2489 * counted. return -ENOENT if the block is root of reloc tree.
2490 */
2491static noinline_for_stack
2492struct btrfs_root *select_one_root(struct btrfs_trans_handle *trans,
2493 struct backref_node *node)
2494{
2495 struct backref_node *next;
2496 struct btrfs_root *root;
2497 struct btrfs_root *fs_root = NULL;
2498 struct backref_edge *edges[BTRFS_MAX_LEVEL - 1];
2499 int index = 0;
2500
2501 next = node;
2502 while (1) {
2503 cond_resched();
2504 next = walk_up_backref(next, edges, &index);
2505 root = next->root;
2506 BUG_ON(!root);
2507
2508 /* no other choice for non-references counted tree */
2509 if (!root->ref_cows)
2510 return root;
2511
2512 if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID)
2513 fs_root = root;
2514
2515 if (next != node)
2516 return NULL;
2517
2518 next = walk_down_backref(edges, &index);
2519 if (!next || next->level <= node->level)
2520 break;
2521 }
2522
2523 if (!fs_root)
2524 return ERR_PTR(-ENOENT);
2525 return fs_root;
2526}
2527
2528static noinline_for_stack
2529u64 calcu_metadata_size(struct reloc_control *rc,
2530 struct backref_node *node, int reserve)
2531{
2532 struct backref_node *next = node;
2533 struct backref_edge *edge;
2534 struct backref_edge *edges[BTRFS_MAX_LEVEL - 1];
2535 u64 num_bytes = 0;
2536 int index = 0;
2537
2538 BUG_ON(reserve && node->processed);
2539
2540 while (next) {
2541 cond_resched();
2542 while (1) {
2543 if (next->processed && (reserve || next != node))
2544 break;
2545
2546 num_bytes += btrfs_level_size(rc->extent_root,
2547 next->level);
2548
2549 if (list_empty(&next->upper))
2550 break;
2551
2552 edge = list_entry(next->upper.next,
2553 struct backref_edge, list[LOWER]);
2554 edges[index++] = edge;
2555 next = edge->node[UPPER];
2556 }
2557 next = walk_down_backref(edges, &index);
2558 }
2559 return num_bytes;
2560}
2561
2562static int reserve_metadata_space(struct btrfs_trans_handle *trans,
2563 struct reloc_control *rc,
2564 struct backref_node *node)
2565{
2566 struct btrfs_root *root = rc->extent_root;
2567 u64 num_bytes;
2568 int ret;
2569 u64 tmp;
2570
2571 num_bytes = calcu_metadata_size(rc, node, 1) * 2;
2572
2573 trans->block_rsv = rc->block_rsv;
2574 rc->reserved_bytes += num_bytes;
2575 ret = btrfs_block_rsv_refill(root, rc->block_rsv, num_bytes,
2576 BTRFS_RESERVE_FLUSH_ALL);
2577 if (ret) {
2578 if (ret == -EAGAIN) {
2579 tmp = rc->extent_root->nodesize *
2580 RELOCATION_RESERVED_NODES;
2581 while (tmp <= rc->reserved_bytes)
2582 tmp <<= 1;
2583 /*
2584 * only one thread can access block_rsv at this point,
2585 * so we don't need hold lock to protect block_rsv.
2586 * we expand more reservation size here to allow enough
2587 * space for relocation and we will return eailer in
2588 * enospc case.
2589 */
2590 rc->block_rsv->size = tmp + rc->extent_root->nodesize *
2591 RELOCATION_RESERVED_NODES;
2592 }
2593 return ret;
2594 }
2595
2596 return 0;
2597}
2598
2599/*
2600 * relocate a block tree, and then update pointers in upper level
2601 * blocks that reference the block to point to the new location.
2602 *
2603 * if called by link_to_upper, the block has already been relocated.
2604 * in that case this function just updates pointers.
2605 */
2606static int do_relocation(struct btrfs_trans_handle *trans,
2607 struct reloc_control *rc,
2608 struct backref_node *node,
2609 struct btrfs_key *key,
2610 struct btrfs_path *path, int lowest)
2611{
2612 struct backref_node *upper;
2613 struct backref_edge *edge;
2614 struct backref_edge *edges[BTRFS_MAX_LEVEL - 1];
2615 struct btrfs_root *root;
2616 struct extent_buffer *eb;
2617 u32 blocksize;
2618 u64 bytenr;
2619 u64 generation;
2620 int slot;
2621 int ret;
2622 int err = 0;
2623
2624 BUG_ON(lowest && node->eb);
2625
2626 path->lowest_level = node->level + 1;
2627 rc->backref_cache.path[node->level] = node;
2628 list_for_each_entry(edge, &node->upper, list[LOWER]) {
2629 cond_resched();
2630
2631 upper = edge->node[UPPER];
2632 root = select_reloc_root(trans, rc, upper, edges);
2633 BUG_ON(!root);
2634
2635 if (upper->eb && !upper->locked) {
2636 if (!lowest) {
2637 ret = btrfs_bin_search(upper->eb, key,
2638 upper->level, &slot);
2639 BUG_ON(ret);
2640 bytenr = btrfs_node_blockptr(upper->eb, slot);
2641 if (node->eb->start == bytenr)
2642 goto next;
2643 }
2644 drop_node_buffer(upper);
2645 }
2646
2647 if (!upper->eb) {
2648 ret = btrfs_search_slot(trans, root, key, path, 0, 1);
2649 if (ret < 0) {
2650 err = ret;
2651 break;
2652 }
2653 BUG_ON(ret > 0);
2654
2655 if (!upper->eb) {
2656 upper->eb = path->nodes[upper->level];
2657 path->nodes[upper->level] = NULL;
2658 } else {
2659 BUG_ON(upper->eb != path->nodes[upper->level]);
2660 }
2661
2662 upper->locked = 1;
2663 path->locks[upper->level] = 0;
2664
2665 slot = path->slots[upper->level];
2666 btrfs_release_path(path);
2667 } else {
2668 ret = btrfs_bin_search(upper->eb, key, upper->level,
2669 &slot);
2670 BUG_ON(ret);
2671 }
2672
2673 bytenr = btrfs_node_blockptr(upper->eb, slot);
2674 if (lowest) {
2675 BUG_ON(bytenr != node->bytenr);
2676 } else {
2677 if (node->eb->start == bytenr)
2678 goto next;
2679 }
2680
2681 blocksize = btrfs_level_size(root, node->level);
2682 generation = btrfs_node_ptr_generation(upper->eb, slot);
2683 eb = read_tree_block(root, bytenr, blocksize, generation);
2684 if (!eb || !extent_buffer_uptodate(eb)) {
2685 free_extent_buffer(eb);
2686 err = -EIO;
2687 goto next;
2688 }
2689 btrfs_tree_lock(eb);
2690 btrfs_set_lock_blocking(eb);
2691
2692 if (!node->eb) {
2693 ret = btrfs_cow_block(trans, root, eb, upper->eb,
2694 slot, &eb);
2695 btrfs_tree_unlock(eb);
2696 free_extent_buffer(eb);
2697 if (ret < 0) {
2698 err = ret;
2699 goto next;
2700 }
2701 BUG_ON(node->eb != eb);
2702 } else {
2703 btrfs_set_node_blockptr(upper->eb, slot,
2704 node->eb->start);
2705 btrfs_set_node_ptr_generation(upper->eb, slot,
2706 trans->transid);
2707 btrfs_mark_buffer_dirty(upper->eb);
2708
2709 ret = btrfs_inc_extent_ref(trans, root,
2710 node->eb->start, blocksize,
2711 upper->eb->start,
2712 btrfs_header_owner(upper->eb),
2713 node->level, 0, 1);
2714 BUG_ON(ret);
2715
2716 ret = btrfs_drop_subtree(trans, root, eb, upper->eb);
2717 BUG_ON(ret);
2718 }
2719next:
2720 if (!upper->pending)
2721 drop_node_buffer(upper);
2722 else
2723 unlock_node_buffer(upper);
2724 if (err)
2725 break;
2726 }
2727
2728 if (!err && node->pending) {
2729 drop_node_buffer(node);
2730 list_move_tail(&node->list, &rc->backref_cache.changed);
2731 node->pending = 0;
2732 }
2733
2734 path->lowest_level = 0;
2735 BUG_ON(err == -ENOSPC);
2736 return err;
2737}
2738
2739static int link_to_upper(struct btrfs_trans_handle *trans,
2740 struct reloc_control *rc,
2741 struct backref_node *node,
2742 struct btrfs_path *path)
2743{
2744 struct btrfs_key key;
2745
2746 btrfs_node_key_to_cpu(node->eb, &key, 0);
2747 return do_relocation(trans, rc, node, &key, path, 0);
2748}
2749
2750static int finish_pending_nodes(struct btrfs_trans_handle *trans,
2751 struct reloc_control *rc,
2752 struct btrfs_path *path, int err)
2753{
2754 LIST_HEAD(list);
2755 struct backref_cache *cache = &rc->backref_cache;
2756 struct backref_node *node;
2757 int level;
2758 int ret;
2759
2760 for (level = 0; level < BTRFS_MAX_LEVEL; level++) {
2761 while (!list_empty(&cache->pending[level])) {
2762 node = list_entry(cache->pending[level].next,
2763 struct backref_node, list);
2764 list_move_tail(&node->list, &list);
2765 BUG_ON(!node->pending);
2766
2767 if (!err) {
2768 ret = link_to_upper(trans, rc, node, path);
2769 if (ret < 0)
2770 err = ret;
2771 }
2772 }
2773 list_splice_init(&list, &cache->pending[level]);
2774 }
2775 return err;
2776}
2777
2778static void mark_block_processed(struct reloc_control *rc,
2779 u64 bytenr, u32 blocksize)
2780{
2781 set_extent_bits(&rc->processed_blocks, bytenr, bytenr + blocksize - 1,
2782 EXTENT_DIRTY, GFP_NOFS);
2783}
2784
2785static void __mark_block_processed(struct reloc_control *rc,
2786 struct backref_node *node)
2787{
2788 u32 blocksize;
2789 if (node->level == 0 ||
2790 in_block_group(node->bytenr, rc->block_group)) {
2791 blocksize = btrfs_level_size(rc->extent_root, node->level);
2792 mark_block_processed(rc, node->bytenr, blocksize);
2793 }
2794 node->processed = 1;
2795}
2796
2797/*
2798 * mark a block and all blocks directly/indirectly reference the block
2799 * as processed.
2800 */
2801static void update_processed_blocks(struct reloc_control *rc,
2802 struct backref_node *node)
2803{
2804 struct backref_node *next = node;
2805 struct backref_edge *edge;
2806 struct backref_edge *edges[BTRFS_MAX_LEVEL - 1];
2807 int index = 0;
2808
2809 while (next) {
2810 cond_resched();
2811 while (1) {
2812 if (next->processed)
2813 break;
2814
2815 __mark_block_processed(rc, next);
2816
2817 if (list_empty(&next->upper))
2818 break;
2819
2820 edge = list_entry(next->upper.next,
2821 struct backref_edge, list[LOWER]);
2822 edges[index++] = edge;
2823 next = edge->node[UPPER];
2824 }
2825 next = walk_down_backref(edges, &index);
2826 }
2827}
2828
2829static int tree_block_processed(u64 bytenr, u32 blocksize,
2830 struct reloc_control *rc)
2831{
2832 if (test_range_bit(&rc->processed_blocks, bytenr,
2833 bytenr + blocksize - 1, EXTENT_DIRTY, 1, NULL))
2834 return 1;
2835 return 0;
2836}
2837
2838static int get_tree_block_key(struct reloc_control *rc,
2839 struct tree_block *block)
2840{
2841 struct extent_buffer *eb;
2842
2843 BUG_ON(block->key_ready);
2844 eb = read_tree_block(rc->extent_root, block->bytenr,
2845 block->key.objectid, block->key.offset);
2846 if (!eb || !extent_buffer_uptodate(eb)) {
2847 free_extent_buffer(eb);
2848 return -EIO;
2849 }
2850 WARN_ON(btrfs_header_level(eb) != block->level);
2851 if (block->level == 0)
2852 btrfs_item_key_to_cpu(eb, &block->key, 0);
2853 else
2854 btrfs_node_key_to_cpu(eb, &block->key, 0);
2855 free_extent_buffer(eb);
2856 block->key_ready = 1;
2857 return 0;
2858}
2859
2860static int reada_tree_block(struct reloc_control *rc,
2861 struct tree_block *block)
2862{
2863 BUG_ON(block->key_ready);
2864 if (block->key.type == BTRFS_METADATA_ITEM_KEY)
2865 readahead_tree_block(rc->extent_root, block->bytenr,
2866 block->key.objectid,
2867 rc->extent_root->leafsize);
2868 else
2869 readahead_tree_block(rc->extent_root, block->bytenr,
2870 block->key.objectid, block->key.offset);
2871 return 0;
2872}
2873
2874/*
2875 * helper function to relocate a tree block
2876 */
2877static int relocate_tree_block(struct btrfs_trans_handle *trans,
2878 struct reloc_control *rc,
2879 struct backref_node *node,
2880 struct btrfs_key *key,
2881 struct btrfs_path *path)
2882{
2883 struct btrfs_root *root;
2884 int ret = 0;
2885
2886 if (!node)
2887 return 0;
2888
2889 BUG_ON(node->processed);
2890 root = select_one_root(trans, node);
2891 if (root == ERR_PTR(-ENOENT)) {
2892 update_processed_blocks(rc, node);
2893 goto out;
2894 }
2895
2896 if (!root || root->ref_cows) {
2897 ret = reserve_metadata_space(trans, rc, node);
2898 if (ret)
2899 goto out;
2900 }
2901
2902 if (root) {
2903 if (root->ref_cows) {
2904 BUG_ON(node->new_bytenr);
2905 BUG_ON(!list_empty(&node->list));
2906 btrfs_record_root_in_trans(trans, root);
2907 root = root->reloc_root;
2908 node->new_bytenr = root->node->start;
2909 node->root = root;
2910 list_add_tail(&node->list, &rc->backref_cache.changed);
2911 } else {
2912 path->lowest_level = node->level;
2913 ret = btrfs_search_slot(trans, root, key, path, 0, 1);
2914 btrfs_release_path(path);
2915 if (ret > 0)
2916 ret = 0;
2917 }
2918 if (!ret)
2919 update_processed_blocks(rc, node);
2920 } else {
2921 ret = do_relocation(trans, rc, node, key, path, 1);
2922 }
2923out:
2924 if (ret || node->level == 0 || node->cowonly)
2925 remove_backref_node(&rc->backref_cache, node);
2926 return ret;
2927}
2928
2929/*
2930 * relocate a list of blocks
2931 */
2932static noinline_for_stack
2933int relocate_tree_blocks(struct btrfs_trans_handle *trans,
2934 struct reloc_control *rc, struct rb_root *blocks)
2935{
2936 struct backref_node *node;
2937 struct btrfs_path *path;
2938 struct tree_block *block;
2939 struct rb_node *rb_node;
2940 int ret;
2941 int err = 0;
2942
2943 path = btrfs_alloc_path();
2944 if (!path) {
2945 err = -ENOMEM;
2946 goto out_free_blocks;
2947 }
2948
2949 rb_node = rb_first(blocks);
2950 while (rb_node) {
2951 block = rb_entry(rb_node, struct tree_block, rb_node);
2952 if (!block->key_ready)
2953 reada_tree_block(rc, block);
2954 rb_node = rb_next(rb_node);
2955 }
2956
2957 rb_node = rb_first(blocks);
2958 while (rb_node) {
2959 block = rb_entry(rb_node, struct tree_block, rb_node);
2960 if (!block->key_ready) {
2961 err = get_tree_block_key(rc, block);
2962 if (err)
2963 goto out_free_path;
2964 }
2965 rb_node = rb_next(rb_node);
2966 }
2967
2968 rb_node = rb_first(blocks);
2969 while (rb_node) {
2970 block = rb_entry(rb_node, struct tree_block, rb_node);
2971
2972 node = build_backref_tree(rc, &block->key,
2973 block->level, block->bytenr);
2974 if (IS_ERR(node)) {
2975 err = PTR_ERR(node);
2976 goto out;
2977 }
2978
2979 ret = relocate_tree_block(trans, rc, node, &block->key,
2980 path);
2981 if (ret < 0) {
2982 if (ret != -EAGAIN || rb_node == rb_first(blocks))
2983 err = ret;
2984 goto out;
2985 }
2986 rb_node = rb_next(rb_node);
2987 }
2988out:
2989 err = finish_pending_nodes(trans, rc, path, err);
2990
2991out_free_path:
2992 btrfs_free_path(path);
2993out_free_blocks:
2994 free_block_list(blocks);
2995 return err;
2996}
2997
2998static noinline_for_stack
2999int prealloc_file_extent_cluster(struct inode *inode,
3000 struct file_extent_cluster *cluster)
3001{
3002 u64 alloc_hint = 0;
3003 u64 start;
3004 u64 end;
3005 u64 offset = BTRFS_I(inode)->index_cnt;
3006 u64 num_bytes;
3007 int nr = 0;
3008 int ret = 0;
3009
3010 BUG_ON(cluster->start != cluster->boundary[0]);
3011 mutex_lock(&inode->i_mutex);
3012
3013 ret = btrfs_check_data_free_space(inode, cluster->end +
3014 1 - cluster->start);
3015 if (ret)
3016 goto out;
3017
3018 while (nr < cluster->nr) {
3019 start = cluster->boundary[nr] - offset;
3020 if (nr + 1 < cluster->nr)
3021 end = cluster->boundary[nr + 1] - 1 - offset;
3022 else
3023 end = cluster->end - offset;
3024
3025 lock_extent(&BTRFS_I(inode)->io_tree, start, end);
3026 num_bytes = end + 1 - start;
3027 ret = btrfs_prealloc_file_range(inode, 0, start,
3028 num_bytes, num_bytes,
3029 end + 1, &alloc_hint);
3030 unlock_extent(&BTRFS_I(inode)->io_tree, start, end);
3031 if (ret)
3032 break;
3033 nr++;
3034 }
3035 btrfs_free_reserved_data_space(inode, cluster->end +
3036 1 - cluster->start);
3037out:
3038 mutex_unlock(&inode->i_mutex);
3039 return ret;
3040}
3041
3042static noinline_for_stack
3043int setup_extent_mapping(struct inode *inode, u64 start, u64 end,
3044 u64 block_start)
3045{
3046 struct btrfs_root *root = BTRFS_I(inode)->root;
3047 struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
3048 struct extent_map *em;
3049 int ret = 0;
3050
3051 em = alloc_extent_map();
3052 if (!em)
3053 return -ENOMEM;
3054
3055 em->start = start;
3056 em->len = end + 1 - start;
3057 em->block_len = em->len;
3058 em->block_start = block_start;
3059 em->bdev = root->fs_info->fs_devices->latest_bdev;
3060 set_bit(EXTENT_FLAG_PINNED, &em->flags);
3061
3062 lock_extent(&BTRFS_I(inode)->io_tree, start, end);
3063 while (1) {
3064 write_lock(&em_tree->lock);
3065 ret = add_extent_mapping(em_tree, em, 0);
3066 write_unlock(&em_tree->lock);
3067 if (ret != -EEXIST) {
3068 free_extent_map(em);
3069 break;
3070 }
3071 btrfs_drop_extent_cache(inode, start, end, 0);
3072 }
3073 unlock_extent(&BTRFS_I(inode)->io_tree, start, end);
3074 return ret;
3075}
3076
3077static int relocate_file_extent_cluster(struct inode *inode,
3078 struct file_extent_cluster *cluster)
3079{
3080 u64 page_start;
3081 u64 page_end;
3082 u64 offset = BTRFS_I(inode)->index_cnt;
3083 unsigned long index;
3084 unsigned long last_index;
3085 struct page *page;
3086 struct file_ra_state *ra;
3087 gfp_t mask = btrfs_alloc_write_mask(inode->i_mapping);
3088 int nr = 0;
3089 int ret = 0;
3090
3091 if (!cluster->nr)
3092 return 0;
3093
3094 ra = kzalloc(sizeof(*ra), GFP_NOFS);
3095 if (!ra)
3096 return -ENOMEM;
3097
3098 ret = prealloc_file_extent_cluster(inode, cluster);
3099 if (ret)
3100 goto out;
3101
3102 file_ra_state_init(ra, inode->i_mapping);
3103
3104 ret = setup_extent_mapping(inode, cluster->start - offset,
3105 cluster->end - offset, cluster->start);
3106 if (ret)
3107 goto out;
3108
3109 index = (cluster->start - offset) >> PAGE_CACHE_SHIFT;
3110 last_index = (cluster->end - offset) >> PAGE_CACHE_SHIFT;
3111 while (index <= last_index) {
3112 ret = btrfs_delalloc_reserve_metadata(inode, PAGE_CACHE_SIZE);
3113 if (ret)
3114 goto out;
3115
3116 page = find_lock_page(inode->i_mapping, index);
3117 if (!page) {
3118 page_cache_sync_readahead(inode->i_mapping,
3119 ra, NULL, index,
3120 last_index + 1 - index);
3121 page = find_or_create_page(inode->i_mapping, index,
3122 mask);
3123 if (!page) {
3124 btrfs_delalloc_release_metadata(inode,
3125 PAGE_CACHE_SIZE);
3126 ret = -ENOMEM;
3127 goto out;
3128 }
3129 }
3130
3131 if (PageReadahead(page)) {
3132 page_cache_async_readahead(inode->i_mapping,
3133 ra, NULL, page, index,
3134 last_index + 1 - index);
3135 }
3136
3137 if (!PageUptodate(page)) {
3138 btrfs_readpage(NULL, page);
3139 lock_page(page);
3140 if (!PageUptodate(page)) {
3141 unlock_page(page);
3142 page_cache_release(page);
3143 btrfs_delalloc_release_metadata(inode,
3144 PAGE_CACHE_SIZE);
3145 ret = -EIO;
3146 goto out;
3147 }
3148 }
3149
3150 page_start = page_offset(page);
3151 page_end = page_start + PAGE_CACHE_SIZE - 1;
3152
3153 lock_extent(&BTRFS_I(inode)->io_tree, page_start, page_end);
3154
3155 set_page_extent_mapped(page);
3156
3157 if (nr < cluster->nr &&
3158 page_start + offset == cluster->boundary[nr]) {
3159 set_extent_bits(&BTRFS_I(inode)->io_tree,
3160 page_start, page_end,
3161 EXTENT_BOUNDARY, GFP_NOFS);
3162 nr++;
3163 }
3164
3165 btrfs_set_extent_delalloc(inode, page_start, page_end, NULL);
3166 set_page_dirty(page);
3167
3168 unlock_extent(&BTRFS_I(inode)->io_tree,
3169 page_start, page_end);
3170 unlock_page(page);
3171 page_cache_release(page);
3172
3173 index++;
3174 balance_dirty_pages_ratelimited(inode->i_mapping);
3175 btrfs_throttle(BTRFS_I(inode)->root);
3176 }
3177 WARN_ON(nr != cluster->nr);
3178out:
3179 kfree(ra);
3180 return ret;
3181}
3182
3183static noinline_for_stack
3184int relocate_data_extent(struct inode *inode, struct btrfs_key *extent_key,
3185 struct file_extent_cluster *cluster)
3186{
3187 int ret;
3188
3189 if (cluster->nr > 0 && extent_key->objectid != cluster->end + 1) {
3190 ret = relocate_file_extent_cluster(inode, cluster);
3191 if (ret)
3192 return ret;
3193 cluster->nr = 0;
3194 }
3195
3196 if (!cluster->nr)
3197 cluster->start = extent_key->objectid;
3198 else
3199 BUG_ON(cluster->nr >= MAX_EXTENTS);
3200 cluster->end = extent_key->objectid + extent_key->offset - 1;
3201 cluster->boundary[cluster->nr] = extent_key->objectid;
3202 cluster->nr++;
3203
3204 if (cluster->nr >= MAX_EXTENTS) {
3205 ret = relocate_file_extent_cluster(inode, cluster);
3206 if (ret)
3207 return ret;
3208 cluster->nr = 0;
3209 }
3210 return 0;
3211}
3212
3213#ifdef BTRFS_COMPAT_EXTENT_TREE_V0
3214static int get_ref_objectid_v0(struct reloc_control *rc,
3215 struct btrfs_path *path,
3216 struct btrfs_key *extent_key,
3217 u64 *ref_objectid, int *path_change)
3218{
3219 struct btrfs_key key;
3220 struct extent_buffer *leaf;
3221 struct btrfs_extent_ref_v0 *ref0;
3222 int ret;
3223 int slot;
3224
3225 leaf = path->nodes[0];
3226 slot = path->slots[0];
3227 while (1) {
3228 if (slot >= btrfs_header_nritems(leaf)) {
3229 ret = btrfs_next_leaf(rc->extent_root, path);
3230 if (ret < 0)
3231 return ret;
3232 BUG_ON(ret > 0);
3233 leaf = path->nodes[0];
3234 slot = path->slots[0];
3235 if (path_change)
3236 *path_change = 1;
3237 }
3238 btrfs_item_key_to_cpu(leaf, &key, slot);
3239 if (key.objectid != extent_key->objectid)
3240 return -ENOENT;
3241
3242 if (key.type != BTRFS_EXTENT_REF_V0_KEY) {
3243 slot++;
3244 continue;
3245 }
3246 ref0 = btrfs_item_ptr(leaf, slot,
3247 struct btrfs_extent_ref_v0);
3248 *ref_objectid = btrfs_ref_objectid_v0(leaf, ref0);
3249 break;
3250 }
3251 return 0;
3252}
3253#endif
3254
3255/*
3256 * helper to add a tree block to the list.
3257 * the major work is getting the generation and level of the block
3258 */
3259static int add_tree_block(struct reloc_control *rc,
3260 struct btrfs_key *extent_key,
3261 struct btrfs_path *path,
3262 struct rb_root *blocks)
3263{
3264 struct extent_buffer *eb;
3265 struct btrfs_extent_item *ei;
3266 struct btrfs_tree_block_info *bi;
3267 struct tree_block *block;
3268 struct rb_node *rb_node;
3269 u32 item_size;
3270 int level = -1;
3271 u64 generation;
3272
3273 eb = path->nodes[0];
3274 item_size = btrfs_item_size_nr(eb, path->slots[0]);
3275
3276 if (extent_key->type == BTRFS_METADATA_ITEM_KEY ||
3277 item_size >= sizeof(*ei) + sizeof(*bi)) {
3278 ei = btrfs_item_ptr(eb, path->slots[0],
3279 struct btrfs_extent_item);
3280 if (extent_key->type == BTRFS_EXTENT_ITEM_KEY) {
3281 bi = (struct btrfs_tree_block_info *)(ei + 1);
3282 level = btrfs_tree_block_level(eb, bi);
3283 } else {
3284 level = (int)extent_key->offset;
3285 }
3286 generation = btrfs_extent_generation(eb, ei);
3287 } else {
3288#ifdef BTRFS_COMPAT_EXTENT_TREE_V0
3289 u64 ref_owner;
3290 int ret;
3291
3292 BUG_ON(item_size != sizeof(struct btrfs_extent_item_v0));
3293 ret = get_ref_objectid_v0(rc, path, extent_key,
3294 &ref_owner, NULL);
3295 if (ret < 0)
3296 return ret;
3297 BUG_ON(ref_owner >= BTRFS_MAX_LEVEL);
3298 level = (int)ref_owner;
3299 /* FIXME: get real generation */
3300 generation = 0;
3301#else
3302 BUG();
3303#endif
3304 }
3305
3306 btrfs_release_path(path);
3307
3308 BUG_ON(level == -1);
3309
3310 block = kmalloc(sizeof(*block), GFP_NOFS);
3311 if (!block)
3312 return -ENOMEM;
3313
3314 block->bytenr = extent_key->objectid;
3315 block->key.objectid = rc->extent_root->leafsize;
3316 block->key.offset = generation;
3317 block->level = level;
3318 block->key_ready = 0;
3319
3320 rb_node = tree_insert(blocks, block->bytenr, &block->rb_node);
3321 if (rb_node)
3322 backref_tree_panic(rb_node, -EEXIST, block->bytenr);
3323
3324 return 0;
3325}
3326
3327/*
3328 * helper to add tree blocks for backref of type BTRFS_SHARED_DATA_REF_KEY
3329 */
3330static int __add_tree_block(struct reloc_control *rc,
3331 u64 bytenr, u32 blocksize,
3332 struct rb_root *blocks)
3333{
3334 struct btrfs_path *path;
3335 struct btrfs_key key;
3336 int ret;
3337 bool skinny = btrfs_fs_incompat(rc->extent_root->fs_info,
3338 SKINNY_METADATA);
3339
3340 if (tree_block_processed(bytenr, blocksize, rc))
3341 return 0;
3342
3343 if (tree_search(blocks, bytenr))
3344 return 0;
3345
3346 path = btrfs_alloc_path();
3347 if (!path)
3348 return -ENOMEM;
3349again:
3350 key.objectid = bytenr;
3351 if (skinny) {
3352 key.type = BTRFS_METADATA_ITEM_KEY;
3353 key.offset = (u64)-1;
3354 } else {
3355 key.type = BTRFS_EXTENT_ITEM_KEY;
3356 key.offset = blocksize;
3357 }
3358
3359 path->search_commit_root = 1;
3360 path->skip_locking = 1;
3361 ret = btrfs_search_slot(NULL, rc->extent_root, &key, path, 0, 0);
3362 if (ret < 0)
3363 goto out;
3364
3365 if (ret > 0 && skinny) {
3366 if (path->slots[0]) {
3367 path->slots[0]--;
3368 btrfs_item_key_to_cpu(path->nodes[0], &key,
3369 path->slots[0]);
3370 if (key.objectid == bytenr &&
3371 (key.type == BTRFS_METADATA_ITEM_KEY ||
3372 (key.type == BTRFS_EXTENT_ITEM_KEY &&
3373 key.offset == blocksize)))
3374 ret = 0;
3375 }
3376
3377 if (ret) {
3378 skinny = false;
3379 btrfs_release_path(path);
3380 goto again;
3381 }
3382 }
3383 BUG_ON(ret);
3384
3385 ret = add_tree_block(rc, &key, path, blocks);
3386out:
3387 btrfs_free_path(path);
3388 return ret;
3389}
3390
3391/*
3392 * helper to check if the block use full backrefs for pointers in it
3393 */
3394static int block_use_full_backref(struct reloc_control *rc,
3395 struct extent_buffer *eb)
3396{
3397 u64 flags;
3398 int ret;
3399
3400 if (btrfs_header_flag(eb, BTRFS_HEADER_FLAG_RELOC) ||
3401 btrfs_header_backref_rev(eb) < BTRFS_MIXED_BACKREF_REV)
3402 return 1;
3403
3404 ret = btrfs_lookup_extent_info(NULL, rc->extent_root,
3405 eb->start, btrfs_header_level(eb), 1,
3406 NULL, &flags);
3407 BUG_ON(ret);
3408
3409 if (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF)
3410 ret = 1;
3411 else
3412 ret = 0;
3413 return ret;
3414}
3415
3416static int delete_block_group_cache(struct btrfs_fs_info *fs_info,
3417 struct inode *inode, u64 ino)
3418{
3419 struct btrfs_key key;
3420 struct btrfs_root *root = fs_info->tree_root;
3421 struct btrfs_trans_handle *trans;
3422 int ret = 0;
3423
3424 if (inode)
3425 goto truncate;
3426
3427 key.objectid = ino;
3428 key.type = BTRFS_INODE_ITEM_KEY;
3429 key.offset = 0;
3430
3431 inode = btrfs_iget(fs_info->sb, &key, root, NULL);
3432 if (IS_ERR(inode) || is_bad_inode(inode)) {
3433 if (!IS_ERR(inode))
3434 iput(inode);
3435 return -ENOENT;
3436 }
3437
3438truncate:
3439 ret = btrfs_check_trunc_cache_free_space(root,
3440 &fs_info->global_block_rsv);
3441 if (ret)
3442 goto out;
3443
3444 trans = btrfs_join_transaction(root);
3445 if (IS_ERR(trans)) {
3446 ret = PTR_ERR(trans);
3447 goto out;
3448 }
3449
3450 ret = btrfs_truncate_free_space_cache(root, trans, inode);
3451
3452 btrfs_end_transaction(trans, root);
3453 btrfs_btree_balance_dirty(root);
3454out:
3455 iput(inode);
3456 return ret;
3457}
3458
3459/*
3460 * helper to add tree blocks for backref of type BTRFS_EXTENT_DATA_REF_KEY
3461 * this function scans fs tree to find blocks reference the data extent
3462 */
3463static int find_data_references(struct reloc_control *rc,
3464 struct btrfs_key *extent_key,
3465 struct extent_buffer *leaf,
3466 struct btrfs_extent_data_ref *ref,
3467 struct rb_root *blocks)
3468{
3469 struct btrfs_path *path;
3470 struct tree_block *block;
3471 struct btrfs_root *root;
3472 struct btrfs_file_extent_item *fi;
3473 struct rb_node *rb_node;
3474 struct btrfs_key key;
3475 u64 ref_root;
3476 u64 ref_objectid;
3477 u64 ref_offset;
3478 u32 ref_count;
3479 u32 nritems;
3480 int err = 0;
3481 int added = 0;
3482 int counted;
3483 int ret;
3484
3485 ref_root = btrfs_extent_data_ref_root(leaf, ref);
3486 ref_objectid = btrfs_extent_data_ref_objectid(leaf, ref);
3487 ref_offset = btrfs_extent_data_ref_offset(leaf, ref);
3488 ref_count = btrfs_extent_data_ref_count(leaf, ref);
3489
3490 /*
3491 * This is an extent belonging to the free space cache, lets just delete
3492 * it and redo the search.
3493 */
3494 if (ref_root == BTRFS_ROOT_TREE_OBJECTID) {
3495 ret = delete_block_group_cache(rc->extent_root->fs_info,
3496 NULL, ref_objectid);
3497 if (ret != -ENOENT)
3498 return ret;
3499 ret = 0;
3500 }
3501
3502 path = btrfs_alloc_path();
3503 if (!path)
3504 return -ENOMEM;
3505 path->reada = 1;
3506
3507 root = read_fs_root(rc->extent_root->fs_info, ref_root);
3508 if (IS_ERR(root)) {
3509 err = PTR_ERR(root);
3510 goto out;
3511 }
3512
3513 key.objectid = ref_objectid;
3514 key.type = BTRFS_EXTENT_DATA_KEY;
3515 if (ref_offset > ((u64)-1 << 32))
3516 key.offset = 0;
3517 else
3518 key.offset = ref_offset;
3519
3520 path->search_commit_root = 1;
3521 path->skip_locking = 1;
3522 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
3523 if (ret < 0) {
3524 err = ret;
3525 goto out;
3526 }
3527
3528 leaf = path->nodes[0];
3529 nritems = btrfs_header_nritems(leaf);
3530 /*
3531 * the references in tree blocks that use full backrefs
3532 * are not counted in
3533 */
3534 if (block_use_full_backref(rc, leaf))
3535 counted = 0;
3536 else
3537 counted = 1;
3538 rb_node = tree_search(blocks, leaf->start);
3539 if (rb_node) {
3540 if (counted)
3541 added = 1;
3542 else
3543 path->slots[0] = nritems;
3544 }
3545
3546 while (ref_count > 0) {
3547 while (path->slots[0] >= nritems) {
3548 ret = btrfs_next_leaf(root, path);
3549 if (ret < 0) {
3550 err = ret;
3551 goto out;
3552 }
3553 if (WARN_ON(ret > 0))
3554 goto out;
3555
3556 leaf = path->nodes[0];
3557 nritems = btrfs_header_nritems(leaf);
3558 added = 0;
3559
3560 if (block_use_full_backref(rc, leaf))
3561 counted = 0;
3562 else
3563 counted = 1;
3564 rb_node = tree_search(blocks, leaf->start);
3565 if (rb_node) {
3566 if (counted)
3567 added = 1;
3568 else
3569 path->slots[0] = nritems;
3570 }
3571 }
3572
3573 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
3574 if (WARN_ON(key.objectid != ref_objectid ||
3575 key.type != BTRFS_EXTENT_DATA_KEY))
3576 break;
3577
3578 fi = btrfs_item_ptr(leaf, path->slots[0],
3579 struct btrfs_file_extent_item);
3580
3581 if (btrfs_file_extent_type(leaf, fi) ==
3582 BTRFS_FILE_EXTENT_INLINE)
3583 goto next;
3584
3585 if (btrfs_file_extent_disk_bytenr(leaf, fi) !=
3586 extent_key->objectid)
3587 goto next;
3588
3589 key.offset -= btrfs_file_extent_offset(leaf, fi);
3590 if (key.offset != ref_offset)
3591 goto next;
3592
3593 if (counted)
3594 ref_count--;
3595 if (added)
3596 goto next;
3597
3598 if (!tree_block_processed(leaf->start, leaf->len, rc)) {
3599 block = kmalloc(sizeof(*block), GFP_NOFS);
3600 if (!block) {
3601 err = -ENOMEM;
3602 break;
3603 }
3604 block->bytenr = leaf->start;
3605 btrfs_item_key_to_cpu(leaf, &block->key, 0);
3606 block->level = 0;
3607 block->key_ready = 1;
3608 rb_node = tree_insert(blocks, block->bytenr,
3609 &block->rb_node);
3610 if (rb_node)
3611 backref_tree_panic(rb_node, -EEXIST,
3612 block->bytenr);
3613 }
3614 if (counted)
3615 added = 1;
3616 else
3617 path->slots[0] = nritems;
3618next:
3619 path->slots[0]++;
3620
3621 }
3622out:
3623 btrfs_free_path(path);
3624 return err;
3625}
3626
3627/*
3628 * helper to find all tree blocks that reference a given data extent
3629 */
3630static noinline_for_stack
3631int add_data_references(struct reloc_control *rc,
3632 struct btrfs_key *extent_key,
3633 struct btrfs_path *path,
3634 struct rb_root *blocks)
3635{
3636 struct btrfs_key key;
3637 struct extent_buffer *eb;
3638 struct btrfs_extent_data_ref *dref;
3639 struct btrfs_extent_inline_ref *iref;
3640 unsigned long ptr;
3641 unsigned long end;
3642 u32 blocksize = btrfs_level_size(rc->extent_root, 0);
3643 int ret = 0;
3644 int err = 0;
3645
3646 eb = path->nodes[0];
3647 ptr = btrfs_item_ptr_offset(eb, path->slots[0]);
3648 end = ptr + btrfs_item_size_nr(eb, path->slots[0]);
3649#ifdef BTRFS_COMPAT_EXTENT_TREE_V0
3650 if (ptr + sizeof(struct btrfs_extent_item_v0) == end)
3651 ptr = end;
3652 else
3653#endif
3654 ptr += sizeof(struct btrfs_extent_item);
3655
3656 while (ptr < end) {
3657 iref = (struct btrfs_extent_inline_ref *)ptr;
3658 key.type = btrfs_extent_inline_ref_type(eb, iref);
3659 if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
3660 key.offset = btrfs_extent_inline_ref_offset(eb, iref);
3661 ret = __add_tree_block(rc, key.offset, blocksize,
3662 blocks);
3663 } else if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
3664 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
3665 ret = find_data_references(rc, extent_key,
3666 eb, dref, blocks);
3667 } else {
3668 BUG();
3669 }
3670 if (ret) {
3671 err = ret;
3672 goto out;
3673 }
3674 ptr += btrfs_extent_inline_ref_size(key.type);
3675 }
3676 WARN_ON(ptr > end);
3677
3678 while (1) {
3679 cond_resched();
3680 eb = path->nodes[0];
3681 if (path->slots[0] >= btrfs_header_nritems(eb)) {
3682 ret = btrfs_next_leaf(rc->extent_root, path);
3683 if (ret < 0) {
3684 err = ret;
3685 break;
3686 }
3687 if (ret > 0)
3688 break;
3689 eb = path->nodes[0];
3690 }
3691
3692 btrfs_item_key_to_cpu(eb, &key, path->slots[0]);
3693 if (key.objectid != extent_key->objectid)
3694 break;
3695
3696#ifdef BTRFS_COMPAT_EXTENT_TREE_V0
3697 if (key.type == BTRFS_SHARED_DATA_REF_KEY ||
3698 key.type == BTRFS_EXTENT_REF_V0_KEY) {
3699#else
3700 BUG_ON(key.type == BTRFS_EXTENT_REF_V0_KEY);
3701 if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
3702#endif
3703 ret = __add_tree_block(rc, key.offset, blocksize,
3704 blocks);
3705 } else if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
3706 dref = btrfs_item_ptr(eb, path->slots[0],
3707 struct btrfs_extent_data_ref);
3708 ret = find_data_references(rc, extent_key,
3709 eb, dref, blocks);
3710 } else {
3711 ret = 0;
3712 }
3713 if (ret) {
3714 err = ret;
3715 break;
3716 }
3717 path->slots[0]++;
3718 }
3719out:
3720 btrfs_release_path(path);
3721 if (err)
3722 free_block_list(blocks);
3723 return err;
3724}
3725
3726/*
3727 * helper to find next unprocessed extent
3728 */
3729static noinline_for_stack
3730int find_next_extent(struct btrfs_trans_handle *trans,
3731 struct reloc_control *rc, struct btrfs_path *path,
3732 struct btrfs_key *extent_key)
3733{
3734 struct btrfs_key key;
3735 struct extent_buffer *leaf;
3736 u64 start, end, last;
3737 int ret;
3738
3739 last = rc->block_group->key.objectid + rc->block_group->key.offset;
3740 while (1) {
3741 cond_resched();
3742 if (rc->search_start >= last) {
3743 ret = 1;
3744 break;
3745 }
3746
3747 key.objectid = rc->search_start;
3748 key.type = BTRFS_EXTENT_ITEM_KEY;
3749 key.offset = 0;
3750
3751 path->search_commit_root = 1;
3752 path->skip_locking = 1;
3753 ret = btrfs_search_slot(NULL, rc->extent_root, &key, path,
3754 0, 0);
3755 if (ret < 0)
3756 break;
3757next:
3758 leaf = path->nodes[0];
3759 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
3760 ret = btrfs_next_leaf(rc->extent_root, path);
3761 if (ret != 0)
3762 break;
3763 leaf = path->nodes[0];
3764 }
3765
3766 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
3767 if (key.objectid >= last) {
3768 ret = 1;
3769 break;
3770 }
3771
3772 if (key.type != BTRFS_EXTENT_ITEM_KEY &&
3773 key.type != BTRFS_METADATA_ITEM_KEY) {
3774 path->slots[0]++;
3775 goto next;
3776 }
3777
3778 if (key.type == BTRFS_EXTENT_ITEM_KEY &&
3779 key.objectid + key.offset <= rc->search_start) {
3780 path->slots[0]++;
3781 goto next;
3782 }
3783
3784 if (key.type == BTRFS_METADATA_ITEM_KEY &&
3785 key.objectid + rc->extent_root->leafsize <=
3786 rc->search_start) {
3787 path->slots[0]++;
3788 goto next;
3789 }
3790
3791 ret = find_first_extent_bit(&rc->processed_blocks,
3792 key.objectid, &start, &end,
3793 EXTENT_DIRTY, NULL);
3794
3795 if (ret == 0 && start <= key.objectid) {
3796 btrfs_release_path(path);
3797 rc->search_start = end + 1;
3798 } else {
3799 if (key.type == BTRFS_EXTENT_ITEM_KEY)
3800 rc->search_start = key.objectid + key.offset;
3801 else
3802 rc->search_start = key.objectid +
3803 rc->extent_root->leafsize;
3804 memcpy(extent_key, &key, sizeof(key));
3805 return 0;
3806 }
3807 }
3808 btrfs_release_path(path);
3809 return ret;
3810}
3811
3812static void set_reloc_control(struct reloc_control *rc)
3813{
3814 struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
3815
3816 mutex_lock(&fs_info->reloc_mutex);
3817 fs_info->reloc_ctl = rc;
3818 mutex_unlock(&fs_info->reloc_mutex);
3819}
3820
3821static void unset_reloc_control(struct reloc_control *rc)
3822{
3823 struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
3824
3825 mutex_lock(&fs_info->reloc_mutex);
3826 fs_info->reloc_ctl = NULL;
3827 mutex_unlock(&fs_info->reloc_mutex);
3828}
3829
3830static int check_extent_flags(u64 flags)
3831{
3832 if ((flags & BTRFS_EXTENT_FLAG_DATA) &&
3833 (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK))
3834 return 1;
3835 if (!(flags & BTRFS_EXTENT_FLAG_DATA) &&
3836 !(flags & BTRFS_EXTENT_FLAG_TREE_BLOCK))
3837 return 1;
3838 if ((flags & BTRFS_EXTENT_FLAG_DATA) &&
3839 (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
3840 return 1;
3841 return 0;
3842}
3843
3844static noinline_for_stack
3845int prepare_to_relocate(struct reloc_control *rc)
3846{
3847 struct btrfs_trans_handle *trans;
3848
3849 rc->block_rsv = btrfs_alloc_block_rsv(rc->extent_root,
3850 BTRFS_BLOCK_RSV_TEMP);
3851 if (!rc->block_rsv)
3852 return -ENOMEM;
3853
3854 memset(&rc->cluster, 0, sizeof(rc->cluster));
3855 rc->search_start = rc->block_group->key.objectid;
3856 rc->extents_found = 0;
3857 rc->nodes_relocated = 0;
3858 rc->merging_rsv_size = 0;
3859 rc->reserved_bytes = 0;
3860 rc->block_rsv->size = rc->extent_root->nodesize *
3861 RELOCATION_RESERVED_NODES;
3862
3863 rc->create_reloc_tree = 1;
3864 set_reloc_control(rc);
3865
3866 trans = btrfs_join_transaction(rc->extent_root);
3867 if (IS_ERR(trans)) {
3868 unset_reloc_control(rc);
3869 /*
3870 * extent tree is not a ref_cow tree and has no reloc_root to
3871 * cleanup. And callers are responsible to free the above
3872 * block rsv.
3873 */
3874 return PTR_ERR(trans);
3875 }
3876 btrfs_commit_transaction(trans, rc->extent_root);
3877 return 0;
3878}
3879
3880static noinline_for_stack int relocate_block_group(struct reloc_control *rc)
3881{
3882 struct rb_root blocks = RB_ROOT;
3883 struct btrfs_key key;
3884 struct btrfs_trans_handle *trans = NULL;
3885 struct btrfs_path *path;
3886 struct btrfs_extent_item *ei;
3887 u64 flags;
3888 u32 item_size;
3889 int ret;
3890 int err = 0;
3891 int progress = 0;
3892
3893 path = btrfs_alloc_path();
3894 if (!path)
3895 return -ENOMEM;
3896 path->reada = 1;
3897
3898 ret = prepare_to_relocate(rc);
3899 if (ret) {
3900 err = ret;
3901 goto out_free;
3902 }
3903
3904 while (1) {
3905 rc->reserved_bytes = 0;
3906 ret = btrfs_block_rsv_refill(rc->extent_root,
3907 rc->block_rsv, rc->block_rsv->size,
3908 BTRFS_RESERVE_FLUSH_ALL);
3909 if (ret) {
3910 err = ret;
3911 break;
3912 }
3913 progress++;
3914 trans = btrfs_start_transaction(rc->extent_root, 0);
3915 if (IS_ERR(trans)) {
3916 err = PTR_ERR(trans);
3917 trans = NULL;
3918 break;
3919 }
3920restart:
3921 if (update_backref_cache(trans, &rc->backref_cache)) {
3922 btrfs_end_transaction(trans, rc->extent_root);
3923 continue;
3924 }
3925
3926 ret = find_next_extent(trans, rc, path, &key);
3927 if (ret < 0)
3928 err = ret;
3929 if (ret != 0)
3930 break;
3931
3932 rc->extents_found++;
3933
3934 ei = btrfs_item_ptr(path->nodes[0], path->slots[0],
3935 struct btrfs_extent_item);
3936 item_size = btrfs_item_size_nr(path->nodes[0], path->slots[0]);
3937 if (item_size >= sizeof(*ei)) {
3938 flags = btrfs_extent_flags(path->nodes[0], ei);
3939 ret = check_extent_flags(flags);
3940 BUG_ON(ret);
3941
3942 } else {
3943#ifdef BTRFS_COMPAT_EXTENT_TREE_V0
3944 u64 ref_owner;
3945 int path_change = 0;
3946
3947 BUG_ON(item_size !=
3948 sizeof(struct btrfs_extent_item_v0));
3949 ret = get_ref_objectid_v0(rc, path, &key, &ref_owner,
3950 &path_change);
3951 if (ref_owner < BTRFS_FIRST_FREE_OBJECTID)
3952 flags = BTRFS_EXTENT_FLAG_TREE_BLOCK;
3953 else
3954 flags = BTRFS_EXTENT_FLAG_DATA;
3955
3956 if (path_change) {
3957 btrfs_release_path(path);
3958
3959 path->search_commit_root = 1;
3960 path->skip_locking = 1;
3961 ret = btrfs_search_slot(NULL, rc->extent_root,
3962 &key, path, 0, 0);
3963 if (ret < 0) {
3964 err = ret;
3965 break;
3966 }
3967 BUG_ON(ret > 0);
3968 }
3969#else
3970 BUG();
3971#endif
3972 }
3973
3974 if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
3975 ret = add_tree_block(rc, &key, path, &blocks);
3976 } else if (rc->stage == UPDATE_DATA_PTRS &&
3977 (flags & BTRFS_EXTENT_FLAG_DATA)) {
3978 ret = add_data_references(rc, &key, path, &blocks);
3979 } else {
3980 btrfs_release_path(path);
3981 ret = 0;
3982 }
3983 if (ret < 0) {
3984 err = ret;
3985 break;
3986 }
3987
3988 if (!RB_EMPTY_ROOT(&blocks)) {
3989 ret = relocate_tree_blocks(trans, rc, &blocks);
3990 if (ret < 0) {
3991 /*
3992 * if we fail to relocate tree blocks, force to update
3993 * backref cache when committing transaction.
3994 */
3995 rc->backref_cache.last_trans = trans->transid - 1;
3996
3997 if (ret != -EAGAIN) {
3998 err = ret;
3999 break;
4000 }
4001 rc->extents_found--;
4002 rc->search_start = key.objectid;
4003 }
4004 }
4005
4006 btrfs_end_transaction_throttle(trans, rc->extent_root);
4007 btrfs_btree_balance_dirty(rc->extent_root);
4008 trans = NULL;
4009
4010 if (rc->stage == MOVE_DATA_EXTENTS &&
4011 (flags & BTRFS_EXTENT_FLAG_DATA)) {
4012 rc->found_file_extent = 1;
4013 ret = relocate_data_extent(rc->data_inode,
4014 &key, &rc->cluster);
4015 if (ret < 0) {
4016 err = ret;
4017 break;
4018 }
4019 }
4020 }
4021 if (trans && progress && err == -ENOSPC) {
4022 ret = btrfs_force_chunk_alloc(trans, rc->extent_root,
4023 rc->block_group->flags);
4024 if (ret == 0) {
4025 err = 0;
4026 progress = 0;
4027 goto restart;
4028 }
4029 }
4030
4031 btrfs_release_path(path);
4032 clear_extent_bits(&rc->processed_blocks, 0, (u64)-1, EXTENT_DIRTY,
4033 GFP_NOFS);
4034
4035 if (trans) {
4036 btrfs_end_transaction_throttle(trans, rc->extent_root);
4037 btrfs_btree_balance_dirty(rc->extent_root);
4038 }
4039
4040 if (!err) {
4041 ret = relocate_file_extent_cluster(rc->data_inode,
4042 &rc->cluster);
4043 if (ret < 0)
4044 err = ret;
4045 }
4046
4047 rc->create_reloc_tree = 0;
4048 set_reloc_control(rc);
4049
4050 backref_cache_cleanup(&rc->backref_cache);
4051 btrfs_block_rsv_release(rc->extent_root, rc->block_rsv, (u64)-1);
4052
4053 err = prepare_to_merge(rc, err);
4054
4055 merge_reloc_roots(rc);
4056
4057 rc->merge_reloc_tree = 0;
4058 unset_reloc_control(rc);
4059 btrfs_block_rsv_release(rc->extent_root, rc->block_rsv, (u64)-1);
4060
4061 /* get rid of pinned extents */
4062 trans = btrfs_join_transaction(rc->extent_root);
4063 if (IS_ERR(trans))
4064 err = PTR_ERR(trans);
4065 else
4066 btrfs_commit_transaction(trans, rc->extent_root);
4067out_free:
4068 btrfs_free_block_rsv(rc->extent_root, rc->block_rsv);
4069 btrfs_free_path(path);
4070 return err;
4071}
4072
4073static int __insert_orphan_inode(struct btrfs_trans_handle *trans,
4074 struct btrfs_root *root, u64 objectid)
4075{
4076 struct btrfs_path *path;
4077 struct btrfs_inode_item *item;
4078 struct extent_buffer *leaf;
4079 int ret;
4080
4081 path = btrfs_alloc_path();
4082 if (!path)
4083 return -ENOMEM;
4084
4085 ret = btrfs_insert_empty_inode(trans, root, path, objectid);
4086 if (ret)
4087 goto out;
4088
4089 leaf = path->nodes[0];
4090 item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_inode_item);
4091 memset_extent_buffer(leaf, 0, (unsigned long)item, sizeof(*item));
4092 btrfs_set_inode_generation(leaf, item, 1);
4093 btrfs_set_inode_size(leaf, item, 0);
4094 btrfs_set_inode_mode(leaf, item, S_IFREG | 0600);
4095 btrfs_set_inode_flags(leaf, item, BTRFS_INODE_NOCOMPRESS |
4096 BTRFS_INODE_PREALLOC);
4097 btrfs_mark_buffer_dirty(leaf);
4098 btrfs_release_path(path);
4099out:
4100 btrfs_free_path(path);
4101 return ret;
4102}
4103
4104/*
4105 * helper to create inode for data relocation.
4106 * the inode is in data relocation tree and its link count is 0
4107 */
4108static noinline_for_stack
4109struct inode *create_reloc_inode(struct btrfs_fs_info *fs_info,
4110 struct btrfs_block_group_cache *group)
4111{
4112 struct inode *inode = NULL;
4113 struct btrfs_trans_handle *trans;
4114 struct btrfs_root *root;
4115 struct btrfs_key key;
4116 u64 objectid = BTRFS_FIRST_FREE_OBJECTID;
4117 int err = 0;
4118
4119 root = read_fs_root(fs_info, BTRFS_DATA_RELOC_TREE_OBJECTID);
4120 if (IS_ERR(root))
4121 return ERR_CAST(root);
4122
4123 trans = btrfs_start_transaction(root, 6);
4124 if (IS_ERR(trans))
4125 return ERR_CAST(trans);
4126
4127 err = btrfs_find_free_objectid(root, &objectid);
4128 if (err)
4129 goto out;
4130
4131 err = __insert_orphan_inode(trans, root, objectid);
4132 BUG_ON(err);
4133
4134 key.objectid = objectid;
4135 key.type = BTRFS_INODE_ITEM_KEY;
4136 key.offset = 0;
4137 inode = btrfs_iget(root->fs_info->sb, &key, root, NULL);
4138 BUG_ON(IS_ERR(inode) || is_bad_inode(inode));
4139 BTRFS_I(inode)->index_cnt = group->key.objectid;
4140
4141 err = btrfs_orphan_add(trans, inode);
4142out:
4143 btrfs_end_transaction(trans, root);
4144 btrfs_btree_balance_dirty(root);
4145 if (err) {
4146 if (inode)
4147 iput(inode);
4148 inode = ERR_PTR(err);
4149 }
4150 return inode;
4151}
4152
4153static struct reloc_control *alloc_reloc_control(struct btrfs_fs_info *fs_info)
4154{
4155 struct reloc_control *rc;
4156
4157 rc = kzalloc(sizeof(*rc), GFP_NOFS);
4158 if (!rc)
4159 return NULL;
4160
4161 INIT_LIST_HEAD(&rc->reloc_roots);
4162 backref_cache_init(&rc->backref_cache);
4163 mapping_tree_init(&rc->reloc_root_tree);
4164 extent_io_tree_init(&rc->processed_blocks,
4165 fs_info->btree_inode->i_mapping);
4166 return rc;
4167}
4168
4169/*
4170 * function to relocate all extents in a block group.
4171 */
4172int btrfs_relocate_block_group(struct btrfs_root *extent_root, u64 group_start)
4173{
4174 struct btrfs_fs_info *fs_info = extent_root->fs_info;
4175 struct reloc_control *rc;
4176 struct inode *inode;
4177 struct btrfs_path *path;
4178 int ret;
4179 int rw = 0;
4180 int err = 0;
4181
4182 rc = alloc_reloc_control(fs_info);
4183 if (!rc)
4184 return -ENOMEM;
4185
4186 rc->extent_root = extent_root;
4187
4188 rc->block_group = btrfs_lookup_block_group(fs_info, group_start);
4189 BUG_ON(!rc->block_group);
4190
4191 if (!rc->block_group->ro) {
4192 ret = btrfs_set_block_group_ro(extent_root, rc->block_group);
4193 if (ret) {
4194 err = ret;
4195 goto out;
4196 }
4197 rw = 1;
4198 }
4199
4200 path = btrfs_alloc_path();
4201 if (!path) {
4202 err = -ENOMEM;
4203 goto out;
4204 }
4205
4206 inode = lookup_free_space_inode(fs_info->tree_root, rc->block_group,
4207 path);
4208 btrfs_free_path(path);
4209
4210 if (!IS_ERR(inode))
4211 ret = delete_block_group_cache(fs_info, inode, 0);
4212 else
4213 ret = PTR_ERR(inode);
4214
4215 if (ret && ret != -ENOENT) {
4216 err = ret;
4217 goto out;
4218 }
4219
4220 rc->data_inode = create_reloc_inode(fs_info, rc->block_group);
4221 if (IS_ERR(rc->data_inode)) {
4222 err = PTR_ERR(rc->data_inode);
4223 rc->data_inode = NULL;
4224 goto out;
4225 }
4226
4227 btrfs_info(extent_root->fs_info, "relocating block group %llu flags %llu",
4228 rc->block_group->key.objectid, rc->block_group->flags);
4229
4230 ret = btrfs_start_delalloc_roots(fs_info, 0, -1);
4231 if (ret < 0) {
4232 err = ret;
4233 goto out;
4234 }
4235 btrfs_wait_ordered_roots(fs_info, -1);
4236
4237 while (1) {
4238 mutex_lock(&fs_info->cleaner_mutex);
4239 ret = relocate_block_group(rc);
4240 mutex_unlock(&fs_info->cleaner_mutex);
4241 if (ret < 0) {
4242 err = ret;
4243 goto out;
4244 }
4245
4246 if (rc->extents_found == 0)
4247 break;
4248
4249 btrfs_info(extent_root->fs_info, "found %llu extents",
4250 rc->extents_found);
4251
4252 if (rc->stage == MOVE_DATA_EXTENTS && rc->found_file_extent) {
4253 ret = btrfs_wait_ordered_range(rc->data_inode, 0,
4254 (u64)-1);
4255 if (ret) {
4256 err = ret;
4257 goto out;
4258 }
4259 invalidate_mapping_pages(rc->data_inode->i_mapping,
4260 0, -1);
4261 rc->stage = UPDATE_DATA_PTRS;
4262 }
4263 }
4264
4265 WARN_ON(rc->block_group->pinned > 0);
4266 WARN_ON(rc->block_group->reserved > 0);
4267 WARN_ON(btrfs_block_group_used(&rc->block_group->item) > 0);
4268out:
4269 if (err && rw)
4270 btrfs_set_block_group_rw(extent_root, rc->block_group);
4271 iput(rc->data_inode);
4272 btrfs_put_block_group(rc->block_group);
4273 kfree(rc);
4274 return err;
4275}
4276
4277static noinline_for_stack int mark_garbage_root(struct btrfs_root *root)
4278{
4279 struct btrfs_trans_handle *trans;
4280 int ret, err;
4281
4282 trans = btrfs_start_transaction(root->fs_info->tree_root, 0);
4283 if (IS_ERR(trans))
4284 return PTR_ERR(trans);
4285
4286 memset(&root->root_item.drop_progress, 0,
4287 sizeof(root->root_item.drop_progress));
4288 root->root_item.drop_level = 0;
4289 btrfs_set_root_refs(&root->root_item, 0);
4290 ret = btrfs_update_root(trans, root->fs_info->tree_root,
4291 &root->root_key, &root->root_item);
4292
4293 err = btrfs_end_transaction(trans, root->fs_info->tree_root);
4294 if (err)
4295 return err;
4296 return ret;
4297}
4298
4299/*
4300 * recover relocation interrupted by system crash.
4301 *
4302 * this function resumes merging reloc trees with corresponding fs trees.
4303 * this is important for keeping the sharing of tree blocks
4304 */
4305int btrfs_recover_relocation(struct btrfs_root *root)
4306{
4307 LIST_HEAD(reloc_roots);
4308 struct btrfs_key key;
4309 struct btrfs_root *fs_root;
4310 struct btrfs_root *reloc_root;
4311 struct btrfs_path *path;
4312 struct extent_buffer *leaf;
4313 struct reloc_control *rc = NULL;
4314 struct btrfs_trans_handle *trans;
4315 int ret;
4316 int err = 0;
4317
4318 path = btrfs_alloc_path();
4319 if (!path)
4320 return -ENOMEM;
4321 path->reada = -1;
4322
4323 key.objectid = BTRFS_TREE_RELOC_OBJECTID;
4324 key.type = BTRFS_ROOT_ITEM_KEY;
4325 key.offset = (u64)-1;
4326
4327 while (1) {
4328 ret = btrfs_search_slot(NULL, root->fs_info->tree_root, &key,
4329 path, 0, 0);
4330 if (ret < 0) {
4331 err = ret;
4332 goto out;
4333 }
4334 if (ret > 0) {
4335 if (path->slots[0] == 0)
4336 break;
4337 path->slots[0]--;
4338 }
4339 leaf = path->nodes[0];
4340 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
4341 btrfs_release_path(path);
4342
4343 if (key.objectid != BTRFS_TREE_RELOC_OBJECTID ||
4344 key.type != BTRFS_ROOT_ITEM_KEY)
4345 break;
4346
4347 reloc_root = btrfs_read_fs_root(root, &key);
4348 if (IS_ERR(reloc_root)) {
4349 err = PTR_ERR(reloc_root);
4350 goto out;
4351 }
4352
4353 list_add(&reloc_root->root_list, &reloc_roots);
4354
4355 if (btrfs_root_refs(&reloc_root->root_item) > 0) {
4356 fs_root = read_fs_root(root->fs_info,
4357 reloc_root->root_key.offset);
4358 if (IS_ERR(fs_root)) {
4359 ret = PTR_ERR(fs_root);
4360 if (ret != -ENOENT) {
4361 err = ret;
4362 goto out;
4363 }
4364 ret = mark_garbage_root(reloc_root);
4365 if (ret < 0) {
4366 err = ret;
4367 goto out;
4368 }
4369 }
4370 }
4371
4372 if (key.offset == 0)
4373 break;
4374
4375 key.offset--;
4376 }
4377 btrfs_release_path(path);
4378
4379 if (list_empty(&reloc_roots))
4380 goto out;
4381
4382 rc = alloc_reloc_control(root->fs_info);
4383 if (!rc) {
4384 err = -ENOMEM;
4385 goto out;
4386 }
4387
4388 rc->extent_root = root->fs_info->extent_root;
4389
4390 set_reloc_control(rc);
4391
4392 trans = btrfs_join_transaction(rc->extent_root);
4393 if (IS_ERR(trans)) {
4394 unset_reloc_control(rc);
4395 err = PTR_ERR(trans);
4396 goto out_free;
4397 }
4398
4399 rc->merge_reloc_tree = 1;
4400
4401 while (!list_empty(&reloc_roots)) {
4402 reloc_root = list_entry(reloc_roots.next,
4403 struct btrfs_root, root_list);
4404 list_del(&reloc_root->root_list);
4405
4406 if (btrfs_root_refs(&reloc_root->root_item) == 0) {
4407 list_add_tail(&reloc_root->root_list,
4408 &rc->reloc_roots);
4409 continue;
4410 }
4411
4412 fs_root = read_fs_root(root->fs_info,
4413 reloc_root->root_key.offset);
4414 if (IS_ERR(fs_root)) {
4415 err = PTR_ERR(fs_root);
4416 goto out_free;
4417 }
4418
4419 err = __add_reloc_root(reloc_root);
4420 BUG_ON(err < 0); /* -ENOMEM or logic error */
4421 fs_root->reloc_root = reloc_root;
4422 }
4423
4424 err = btrfs_commit_transaction(trans, rc->extent_root);
4425 if (err)
4426 goto out_free;
4427
4428 merge_reloc_roots(rc);
4429
4430 unset_reloc_control(rc);
4431
4432 trans = btrfs_join_transaction(rc->extent_root);
4433 if (IS_ERR(trans))
4434 err = PTR_ERR(trans);
4435 else
4436 err = btrfs_commit_transaction(trans, rc->extent_root);
4437out_free:
4438 kfree(rc);
4439out:
4440 if (!list_empty(&reloc_roots))
4441 free_reloc_roots(&reloc_roots);
4442
4443 btrfs_free_path(path);
4444
4445 if (err == 0) {
4446 /* cleanup orphan inode in data relocation tree */
4447 fs_root = read_fs_root(root->fs_info,
4448 BTRFS_DATA_RELOC_TREE_OBJECTID);
4449 if (IS_ERR(fs_root))
4450 err = PTR_ERR(fs_root);
4451 else
4452 err = btrfs_orphan_cleanup(fs_root);
4453 }
4454 return err;
4455}
4456
4457/*
4458 * helper to add ordered checksum for data relocation.
4459 *
4460 * cloning checksum properly handles the nodatasum extents.
4461 * it also saves CPU time to re-calculate the checksum.
4462 */
4463int btrfs_reloc_clone_csums(struct inode *inode, u64 file_pos, u64 len)
4464{
4465 struct btrfs_ordered_sum *sums;
4466 struct btrfs_ordered_extent *ordered;
4467 struct btrfs_root *root = BTRFS_I(inode)->root;
4468 int ret;
4469 u64 disk_bytenr;
4470 u64 new_bytenr;
4471 LIST_HEAD(list);
4472
4473 ordered = btrfs_lookup_ordered_extent(inode, file_pos);
4474 BUG_ON(ordered->file_offset != file_pos || ordered->len != len);
4475
4476 disk_bytenr = file_pos + BTRFS_I(inode)->index_cnt;
4477 ret = btrfs_lookup_csums_range(root->fs_info->csum_root, disk_bytenr,
4478 disk_bytenr + len - 1, &list, 0);
4479 if (ret)
4480 goto out;
4481
4482 while (!list_empty(&list)) {
4483 sums = list_entry(list.next, struct btrfs_ordered_sum, list);
4484 list_del_init(&sums->list);
4485
4486 /*
4487 * We need to offset the new_bytenr based on where the csum is.
4488 * We need to do this because we will read in entire prealloc
4489 * extents but we may have written to say the middle of the
4490 * prealloc extent, so we need to make sure the csum goes with
4491 * the right disk offset.
4492 *
4493 * We can do this because the data reloc inode refers strictly
4494 * to the on disk bytes, so we don't have to worry about
4495 * disk_len vs real len like with real inodes since it's all
4496 * disk length.
4497 */
4498 new_bytenr = ordered->start + (sums->bytenr - disk_bytenr);
4499 sums->bytenr = new_bytenr;
4500
4501 btrfs_add_ordered_sum(inode, ordered, sums);
4502 }
4503out:
4504 btrfs_put_ordered_extent(ordered);
4505 return ret;
4506}
4507
4508int btrfs_reloc_cow_block(struct btrfs_trans_handle *trans,
4509 struct btrfs_root *root, struct extent_buffer *buf,
4510 struct extent_buffer *cow)
4511{
4512 struct reloc_control *rc;
4513 struct backref_node *node;
4514 int first_cow = 0;
4515 int level;
4516 int ret = 0;
4517
4518 rc = root->fs_info->reloc_ctl;
4519 if (!rc)
4520 return 0;
4521
4522 BUG_ON(rc->stage == UPDATE_DATA_PTRS &&
4523 root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID);
4524
4525 if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID) {
4526 if (buf == root->node)
4527 __update_reloc_root(root, cow->start);
4528 }
4529
4530 level = btrfs_header_level(buf);
4531 if (btrfs_header_generation(buf) <=
4532 btrfs_root_last_snapshot(&root->root_item))
4533 first_cow = 1;
4534
4535 if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID &&
4536 rc->create_reloc_tree) {
4537 WARN_ON(!first_cow && level == 0);
4538
4539 node = rc->backref_cache.path[level];
4540 BUG_ON(node->bytenr != buf->start &&
4541 node->new_bytenr != buf->start);
4542
4543 drop_node_buffer(node);
4544 extent_buffer_get(cow);
4545 node->eb = cow;
4546 node->new_bytenr = cow->start;
4547
4548 if (!node->pending) {
4549 list_move_tail(&node->list,
4550 &rc->backref_cache.pending[level]);
4551 node->pending = 1;
4552 }
4553
4554 if (first_cow)
4555 __mark_block_processed(rc, node);
4556
4557 if (first_cow && level > 0)
4558 rc->nodes_relocated += buf->len;
4559 }
4560
4561 if (level == 0 && first_cow && rc->stage == UPDATE_DATA_PTRS)
4562 ret = replace_file_extents(trans, rc, root, cow);
4563 return ret;
4564}
4565
4566/*
4567 * called before creating snapshot. it calculates metadata reservation
4568 * requried for relocating tree blocks in the snapshot
4569 */
4570void btrfs_reloc_pre_snapshot(struct btrfs_trans_handle *trans,
4571 struct btrfs_pending_snapshot *pending,
4572 u64 *bytes_to_reserve)
4573{
4574 struct btrfs_root *root;
4575 struct reloc_control *rc;
4576
4577 root = pending->root;
4578 if (!root->reloc_root)
4579 return;
4580
4581 rc = root->fs_info->reloc_ctl;
4582 if (!rc->merge_reloc_tree)
4583 return;
4584
4585 root = root->reloc_root;
4586 BUG_ON(btrfs_root_refs(&root->root_item) == 0);
4587 /*
4588 * relocation is in the stage of merging trees. the space
4589 * used by merging a reloc tree is twice the size of
4590 * relocated tree nodes in the worst case. half for cowing
4591 * the reloc tree, half for cowing the fs tree. the space
4592 * used by cowing the reloc tree will be freed after the
4593 * tree is dropped. if we create snapshot, cowing the fs
4594 * tree may use more space than it frees. so we need
4595 * reserve extra space.
4596 */
4597 *bytes_to_reserve += rc->nodes_relocated;
4598}
4599
4600/*
4601 * called after snapshot is created. migrate block reservation
4602 * and create reloc root for the newly created snapshot
4603 */
4604int btrfs_reloc_post_snapshot(struct btrfs_trans_handle *trans,
4605 struct btrfs_pending_snapshot *pending)
4606{
4607 struct btrfs_root *root = pending->root;
4608 struct btrfs_root *reloc_root;
4609 struct btrfs_root *new_root;
4610 struct reloc_control *rc;
4611 int ret;
4612
4613 if (!root->reloc_root)
4614 return 0;
4615
4616 rc = root->fs_info->reloc_ctl;
4617 rc->merging_rsv_size += rc->nodes_relocated;
4618
4619 if (rc->merge_reloc_tree) {
4620 ret = btrfs_block_rsv_migrate(&pending->block_rsv,
4621 rc->block_rsv,
4622 rc->nodes_relocated);
4623 if (ret)
4624 return ret;
4625 }
4626
4627 new_root = pending->snap;
4628 reloc_root = create_reloc_root(trans, root->reloc_root,
4629 new_root->root_key.objectid);
4630 if (IS_ERR(reloc_root))
4631 return PTR_ERR(reloc_root);
4632
4633 ret = __add_reloc_root(reloc_root);
4634 BUG_ON(ret < 0);
4635 new_root->reloc_root = reloc_root;
4636
4637 if (rc->create_reloc_tree)
4638 ret = clone_backref_node(trans, rc, root, reloc_root);
4639 return ret;
4640}