1 // SPDX-License-Identifier: GPL-2.0
3 * Copyright (C) 2009 Oracle. All rights reserved.
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>
15 #include "transaction.h"
18 #include "btrfs_inode.h"
19 #include "async-thread.h"
20 #include "free-space-cache.h"
22 #include "print-tree.h"
23 #include "delalloc-space.h"
24 #include "block-group.h"
32 * [What does relocation do]
34 * The objective of relocation is to relocate all extents of the target block
35 * group to other block groups.
36 * This is utilized by resize (shrink only), profile converting, compacting
37 * space, or balance routine to spread chunks over devices.
40 * ------------------------------------------------------------------
41 * BG A: 10 data extents | BG A: deleted
42 * BG B: 2 data extents | BG B: 10 data extents (2 old + 8 relocated)
43 * BG C: 1 extents | BG C: 3 data extents (1 old + 2 relocated)
45 * [How does relocation work]
47 * 1. Mark the target block group read-only
48 * New extents won't be allocated from the target block group.
50 * 2.1 Record each extent in the target block group
51 * To build a proper map of extents to be relocated.
53 * 2.2 Build data reloc tree and reloc trees
54 * Data reloc tree will contain an inode, recording all newly relocated
56 * There will be only one data reloc tree for one data block group.
58 * Reloc tree will be a special snapshot of its source tree, containing
59 * relocated tree blocks.
60 * Each tree referring to a tree block in target block group will get its
63 * 2.3 Swap source tree with its corresponding reloc tree
64 * Each involved tree only refers to new extents after swap.
66 * 3. Cleanup reloc trees and data reloc tree.
67 * As old extents in the target block group are still referenced by reloc
68 * trees, we need to clean them up before really freeing the target block
71 * The main complexity is in steps 2.2 and 2.3.
73 * The entry point of relocation is relocate_block_group() function.
76 #define RELOCATION_RESERVED_NODES 256
78 * map address of tree root to tree
82 struct rb_node rb_node;
84 }; /* Use rb_simle_node for search/insert */
89 struct rb_root rb_root;
94 * present a tree block to process
98 struct rb_node rb_node;
100 }; /* Use rb_simple_node for search/insert */
102 struct btrfs_key key;
103 unsigned int level:8;
104 unsigned int key_ready:1;
107 #define MAX_EXTENTS 128
109 struct file_extent_cluster {
112 u64 boundary[MAX_EXTENTS];
116 struct reloc_control {
117 /* block group to relocate */
118 struct btrfs_block_group *block_group;
120 struct btrfs_root *extent_root;
121 /* inode for moving data */
122 struct inode *data_inode;
124 struct btrfs_block_rsv *block_rsv;
126 struct btrfs_backref_cache backref_cache;
128 struct file_extent_cluster cluster;
129 /* tree blocks have been processed */
130 struct extent_io_tree processed_blocks;
131 /* map start of tree root to corresponding reloc tree */
132 struct mapping_tree reloc_root_tree;
133 /* list of reloc trees */
134 struct list_head reloc_roots;
135 /* list of subvolume trees that get relocated */
136 struct list_head dirty_subvol_roots;
137 /* size of metadata reservation for merging reloc trees */
138 u64 merging_rsv_size;
139 /* size of relocated tree nodes */
141 /* reserved size for block group relocation*/
147 unsigned int stage:8;
148 unsigned int create_reloc_tree:1;
149 unsigned int merge_reloc_tree:1;
150 unsigned int found_file_extent:1;
153 /* stages of data relocation */
154 #define MOVE_DATA_EXTENTS 0
155 #define UPDATE_DATA_PTRS 1
157 static void mark_block_processed(struct reloc_control *rc,
158 struct btrfs_backref_node *node)
162 if (node->level == 0 ||
163 in_range(node->bytenr, rc->block_group->start,
164 rc->block_group->length)) {
165 blocksize = rc->extent_root->fs_info->nodesize;
166 set_extent_bits(&rc->processed_blocks, node->bytenr,
167 node->bytenr + blocksize - 1, EXTENT_DIRTY);
173 static void mapping_tree_init(struct mapping_tree *tree)
175 tree->rb_root = RB_ROOT;
176 spin_lock_init(&tree->lock);
180 * walk up backref nodes until reach node presents tree root
182 static struct btrfs_backref_node *walk_up_backref(
183 struct btrfs_backref_node *node,
184 struct btrfs_backref_edge *edges[], int *index)
186 struct btrfs_backref_edge *edge;
189 while (!list_empty(&node->upper)) {
190 edge = list_entry(node->upper.next,
191 struct btrfs_backref_edge, list[LOWER]);
193 node = edge->node[UPPER];
195 BUG_ON(node->detached);
201 * walk down backref nodes to find start of next reference path
203 static struct btrfs_backref_node *walk_down_backref(
204 struct btrfs_backref_edge *edges[], int *index)
206 struct btrfs_backref_edge *edge;
207 struct btrfs_backref_node *lower;
211 edge = edges[idx - 1];
212 lower = edge->node[LOWER];
213 if (list_is_last(&edge->list[LOWER], &lower->upper)) {
217 edge = list_entry(edge->list[LOWER].next,
218 struct btrfs_backref_edge, list[LOWER]);
219 edges[idx - 1] = edge;
221 return edge->node[UPPER];
227 static void update_backref_node(struct btrfs_backref_cache *cache,
228 struct btrfs_backref_node *node, u64 bytenr)
230 struct rb_node *rb_node;
231 rb_erase(&node->rb_node, &cache->rb_root);
232 node->bytenr = bytenr;
233 rb_node = rb_simple_insert(&cache->rb_root, node->bytenr, &node->rb_node);
235 btrfs_backref_panic(cache->fs_info, bytenr, -EEXIST);
239 * update backref cache after a transaction commit
241 static int update_backref_cache(struct btrfs_trans_handle *trans,
242 struct btrfs_backref_cache *cache)
244 struct btrfs_backref_node *node;
247 if (cache->last_trans == 0) {
248 cache->last_trans = trans->transid;
252 if (cache->last_trans == trans->transid)
256 * detached nodes are used to avoid unnecessary backref
257 * lookup. transaction commit changes the extent tree.
258 * so the detached nodes are no longer useful.
260 while (!list_empty(&cache->detached)) {
261 node = list_entry(cache->detached.next,
262 struct btrfs_backref_node, list);
263 btrfs_backref_cleanup_node(cache, node);
266 while (!list_empty(&cache->changed)) {
267 node = list_entry(cache->changed.next,
268 struct btrfs_backref_node, list);
269 list_del_init(&node->list);
270 BUG_ON(node->pending);
271 update_backref_node(cache, node, node->new_bytenr);
275 * some nodes can be left in the pending list if there were
276 * errors during processing the pending nodes.
278 for (level = 0; level < BTRFS_MAX_LEVEL; level++) {
279 list_for_each_entry(node, &cache->pending[level], list) {
280 BUG_ON(!node->pending);
281 if (node->bytenr == node->new_bytenr)
283 update_backref_node(cache, node, node->new_bytenr);
287 cache->last_trans = 0;
291 static bool reloc_root_is_dead(struct btrfs_root *root)
294 * Pair with set_bit/clear_bit in clean_dirty_subvols and
295 * btrfs_update_reloc_root. We need to see the updated bit before
296 * trying to access reloc_root
299 if (test_bit(BTRFS_ROOT_DEAD_RELOC_TREE, &root->state))
305 * Check if this subvolume tree has valid reloc tree.
307 * Reloc tree after swap is considered dead, thus not considered as valid.
308 * This is enough for most callers, as they don't distinguish dead reloc root
309 * from no reloc root. But btrfs_should_ignore_reloc_root() below is a
312 static bool have_reloc_root(struct btrfs_root *root)
314 if (reloc_root_is_dead(root))
316 if (!root->reloc_root)
321 int btrfs_should_ignore_reloc_root(struct btrfs_root *root)
323 struct btrfs_root *reloc_root;
325 if (!test_bit(BTRFS_ROOT_SHAREABLE, &root->state))
328 /* This root has been merged with its reloc tree, we can ignore it */
329 if (reloc_root_is_dead(root))
332 reloc_root = root->reloc_root;
336 if (btrfs_header_generation(reloc_root->commit_root) ==
337 root->fs_info->running_transaction->transid)
340 * if there is reloc tree and it was created in previous
341 * transaction backref lookup can find the reloc tree,
342 * so backref node for the fs tree root is useless for
349 * find reloc tree by address of tree root
351 struct btrfs_root *find_reloc_root(struct btrfs_fs_info *fs_info, u64 bytenr)
353 struct reloc_control *rc = fs_info->reloc_ctl;
354 struct rb_node *rb_node;
355 struct mapping_node *node;
356 struct btrfs_root *root = NULL;
359 spin_lock(&rc->reloc_root_tree.lock);
360 rb_node = rb_simple_search(&rc->reloc_root_tree.rb_root, bytenr);
362 node = rb_entry(rb_node, struct mapping_node, rb_node);
363 root = (struct btrfs_root *)node->data;
365 spin_unlock(&rc->reloc_root_tree.lock);
366 return btrfs_grab_root(root);
370 * For useless nodes, do two major clean ups:
372 * - Cleanup the children edges and nodes
373 * If child node is also orphan (no parent) during cleanup, then the child
374 * node will also be cleaned up.
376 * - Freeing up leaves (level 0), keeps nodes detached
377 * For nodes, the node is still cached as "detached"
379 * Return false if @node is not in the @useless_nodes list.
380 * Return true if @node is in the @useless_nodes list.
382 static bool handle_useless_nodes(struct reloc_control *rc,
383 struct btrfs_backref_node *node)
385 struct btrfs_backref_cache *cache = &rc->backref_cache;
386 struct list_head *useless_node = &cache->useless_node;
389 while (!list_empty(useless_node)) {
390 struct btrfs_backref_node *cur;
392 cur = list_first_entry(useless_node, struct btrfs_backref_node,
394 list_del_init(&cur->list);
396 /* Only tree root nodes can be added to @useless_nodes */
397 ASSERT(list_empty(&cur->upper));
402 /* The node is the lowest node */
404 list_del_init(&cur->lower);
408 /* Cleanup the lower edges */
409 while (!list_empty(&cur->lower)) {
410 struct btrfs_backref_edge *edge;
411 struct btrfs_backref_node *lower;
413 edge = list_entry(cur->lower.next,
414 struct btrfs_backref_edge, list[UPPER]);
415 list_del(&edge->list[UPPER]);
416 list_del(&edge->list[LOWER]);
417 lower = edge->node[LOWER];
418 btrfs_backref_free_edge(cache, edge);
420 /* Child node is also orphan, queue for cleanup */
421 if (list_empty(&lower->upper))
422 list_add(&lower->list, useless_node);
424 /* Mark this block processed for relocation */
425 mark_block_processed(rc, cur);
428 * Backref nodes for tree leaves are deleted from the cache.
429 * Backref nodes for upper level tree blocks are left in the
430 * cache to avoid unnecessary backref lookup.
432 if (cur->level > 0) {
433 list_add(&cur->list, &cache->detached);
436 rb_erase(&cur->rb_node, &cache->rb_root);
437 btrfs_backref_free_node(cache, cur);
444 * Build backref tree for a given tree block. Root of the backref tree
445 * corresponds the tree block, leaves of the backref tree correspond roots of
446 * b-trees that reference the tree block.
448 * The basic idea of this function is check backrefs of a given block to find
449 * upper level blocks that reference the block, and then check backrefs of
450 * these upper level blocks recursively. The recursion stops when tree root is
451 * reached or backrefs for the block is cached.
453 * NOTE: if we find that backrefs for a block are cached, we know backrefs for
454 * all upper level blocks that directly/indirectly reference the block are also
457 static noinline_for_stack struct btrfs_backref_node *build_backref_tree(
458 struct reloc_control *rc, struct btrfs_key *node_key,
459 int level, u64 bytenr)
461 struct btrfs_backref_iter *iter;
462 struct btrfs_backref_cache *cache = &rc->backref_cache;
463 /* For searching parent of TREE_BLOCK_REF */
464 struct btrfs_path *path;
465 struct btrfs_backref_node *cur;
466 struct btrfs_backref_node *node = NULL;
467 struct btrfs_backref_edge *edge;
471 iter = btrfs_backref_iter_alloc(rc->extent_root->fs_info, GFP_NOFS);
473 return ERR_PTR(-ENOMEM);
474 path = btrfs_alloc_path();
480 node = btrfs_backref_alloc_node(cache, bytenr, level);
489 /* Breadth-first search to build backref cache */
491 ret = btrfs_backref_add_tree_node(cache, path, iter, node_key,
497 edge = list_first_entry_or_null(&cache->pending_edge,
498 struct btrfs_backref_edge, list[UPPER]);
500 * The pending list isn't empty, take the first block to
504 list_del_init(&edge->list[UPPER]);
505 cur = edge->node[UPPER];
509 /* Finish the upper linkage of newly added edges/nodes */
510 ret = btrfs_backref_finish_upper_links(cache, node);
516 if (handle_useless_nodes(rc, node))
519 btrfs_backref_iter_free(iter);
520 btrfs_free_path(path);
522 btrfs_backref_error_cleanup(cache, node);
525 ASSERT(!node || !node->detached);
526 ASSERT(list_empty(&cache->useless_node) &&
527 list_empty(&cache->pending_edge));
532 * helper to add backref node for the newly created snapshot.
533 * the backref node is created by cloning backref node that
534 * corresponds to root of source tree
536 static int clone_backref_node(struct btrfs_trans_handle *trans,
537 struct reloc_control *rc,
538 struct btrfs_root *src,
539 struct btrfs_root *dest)
541 struct btrfs_root *reloc_root = src->reloc_root;
542 struct btrfs_backref_cache *cache = &rc->backref_cache;
543 struct btrfs_backref_node *node = NULL;
544 struct btrfs_backref_node *new_node;
545 struct btrfs_backref_edge *edge;
546 struct btrfs_backref_edge *new_edge;
547 struct rb_node *rb_node;
549 if (cache->last_trans > 0)
550 update_backref_cache(trans, cache);
552 rb_node = rb_simple_search(&cache->rb_root, src->commit_root->start);
554 node = rb_entry(rb_node, struct btrfs_backref_node, rb_node);
558 BUG_ON(node->new_bytenr != reloc_root->node->start);
562 rb_node = rb_simple_search(&cache->rb_root,
563 reloc_root->commit_root->start);
565 node = rb_entry(rb_node, struct btrfs_backref_node,
567 BUG_ON(node->detached);
574 new_node = btrfs_backref_alloc_node(cache, dest->node->start,
579 new_node->lowest = node->lowest;
580 new_node->checked = 1;
581 new_node->root = btrfs_grab_root(dest);
582 ASSERT(new_node->root);
585 list_for_each_entry(edge, &node->lower, list[UPPER]) {
586 new_edge = btrfs_backref_alloc_edge(cache);
590 btrfs_backref_link_edge(new_edge, edge->node[LOWER],
591 new_node, LINK_UPPER);
594 list_add_tail(&new_node->lower, &cache->leaves);
597 rb_node = rb_simple_insert(&cache->rb_root, new_node->bytenr,
600 btrfs_backref_panic(trans->fs_info, new_node->bytenr, -EEXIST);
602 if (!new_node->lowest) {
603 list_for_each_entry(new_edge, &new_node->lower, list[UPPER]) {
604 list_add_tail(&new_edge->list[LOWER],
605 &new_edge->node[LOWER]->upper);
610 while (!list_empty(&new_node->lower)) {
611 new_edge = list_entry(new_node->lower.next,
612 struct btrfs_backref_edge, list[UPPER]);
613 list_del(&new_edge->list[UPPER]);
614 btrfs_backref_free_edge(cache, new_edge);
616 btrfs_backref_free_node(cache, new_node);
621 * helper to add 'address of tree root -> reloc tree' mapping
623 static int __must_check __add_reloc_root(struct btrfs_root *root)
625 struct btrfs_fs_info *fs_info = root->fs_info;
626 struct rb_node *rb_node;
627 struct mapping_node *node;
628 struct reloc_control *rc = fs_info->reloc_ctl;
630 node = kmalloc(sizeof(*node), GFP_NOFS);
634 node->bytenr = root->commit_root->start;
637 spin_lock(&rc->reloc_root_tree.lock);
638 rb_node = rb_simple_insert(&rc->reloc_root_tree.rb_root,
639 node->bytenr, &node->rb_node);
640 spin_unlock(&rc->reloc_root_tree.lock);
643 "Duplicate root found for start=%llu while inserting into relocation tree",
648 list_add_tail(&root->root_list, &rc->reloc_roots);
653 * helper to delete the 'address of tree root -> reloc tree'
656 static void __del_reloc_root(struct btrfs_root *root)
658 struct btrfs_fs_info *fs_info = root->fs_info;
659 struct rb_node *rb_node;
660 struct mapping_node *node = NULL;
661 struct reloc_control *rc = fs_info->reloc_ctl;
662 bool put_ref = false;
664 if (rc && root->node) {
665 spin_lock(&rc->reloc_root_tree.lock);
666 rb_node = rb_simple_search(&rc->reloc_root_tree.rb_root,
667 root->commit_root->start);
669 node = rb_entry(rb_node, struct mapping_node, rb_node);
670 rb_erase(&node->rb_node, &rc->reloc_root_tree.rb_root);
671 RB_CLEAR_NODE(&node->rb_node);
673 spin_unlock(&rc->reloc_root_tree.lock);
674 ASSERT(!node || (struct btrfs_root *)node->data == root);
678 * We only put the reloc root here if it's on the list. There's a lot
679 * of places where the pattern is to splice the rc->reloc_roots, process
680 * the reloc roots, and then add the reloc root back onto
681 * rc->reloc_roots. If we call __del_reloc_root while it's off of the
682 * list we don't want the reference being dropped, because the guy
683 * messing with the list is in charge of the reference.
685 spin_lock(&fs_info->trans_lock);
686 if (!list_empty(&root->root_list)) {
688 list_del_init(&root->root_list);
690 spin_unlock(&fs_info->trans_lock);
692 btrfs_put_root(root);
697 * helper to update the 'address of tree root -> reloc tree'
700 static int __update_reloc_root(struct btrfs_root *root)
702 struct btrfs_fs_info *fs_info = root->fs_info;
703 struct rb_node *rb_node;
704 struct mapping_node *node = NULL;
705 struct reloc_control *rc = fs_info->reloc_ctl;
707 spin_lock(&rc->reloc_root_tree.lock);
708 rb_node = rb_simple_search(&rc->reloc_root_tree.rb_root,
709 root->commit_root->start);
711 node = rb_entry(rb_node, struct mapping_node, rb_node);
712 rb_erase(&node->rb_node, &rc->reloc_root_tree.rb_root);
714 spin_unlock(&rc->reloc_root_tree.lock);
718 BUG_ON((struct btrfs_root *)node->data != root);
720 spin_lock(&rc->reloc_root_tree.lock);
721 node->bytenr = root->node->start;
722 rb_node = rb_simple_insert(&rc->reloc_root_tree.rb_root,
723 node->bytenr, &node->rb_node);
724 spin_unlock(&rc->reloc_root_tree.lock);
726 btrfs_backref_panic(fs_info, node->bytenr, -EEXIST);
730 static struct btrfs_root *create_reloc_root(struct btrfs_trans_handle *trans,
731 struct btrfs_root *root, u64 objectid)
733 struct btrfs_fs_info *fs_info = root->fs_info;
734 struct btrfs_root *reloc_root;
735 struct extent_buffer *eb;
736 struct btrfs_root_item *root_item;
737 struct btrfs_key root_key;
739 bool must_abort = false;
741 root_item = kmalloc(sizeof(*root_item), GFP_NOFS);
743 return ERR_PTR(-ENOMEM);
745 root_key.objectid = BTRFS_TREE_RELOC_OBJECTID;
746 root_key.type = BTRFS_ROOT_ITEM_KEY;
747 root_key.offset = objectid;
749 if (root->root_key.objectid == objectid) {
752 /* called by btrfs_init_reloc_root */
753 ret = btrfs_copy_root(trans, root, root->commit_root, &eb,
754 BTRFS_TREE_RELOC_OBJECTID);
759 * Set the last_snapshot field to the generation of the commit
760 * root - like this ctree.c:btrfs_block_can_be_shared() behaves
761 * correctly (returns true) when the relocation root is created
762 * either inside the critical section of a transaction commit
763 * (through transaction.c:qgroup_account_snapshot()) and when
764 * it's created before the transaction commit is started.
766 commit_root_gen = btrfs_header_generation(root->commit_root);
767 btrfs_set_root_last_snapshot(&root->root_item, commit_root_gen);
770 * called by btrfs_reloc_post_snapshot_hook.
771 * the source tree is a reloc tree, all tree blocks
772 * modified after it was created have RELOC flag
773 * set in their headers. so it's OK to not update
774 * the 'last_snapshot'.
776 ret = btrfs_copy_root(trans, root, root->node, &eb,
777 BTRFS_TREE_RELOC_OBJECTID);
783 * We have changed references at this point, we must abort the
784 * transaction if anything fails.
788 memcpy(root_item, &root->root_item, sizeof(*root_item));
789 btrfs_set_root_bytenr(root_item, eb->start);
790 btrfs_set_root_level(root_item, btrfs_header_level(eb));
791 btrfs_set_root_generation(root_item, trans->transid);
793 if (root->root_key.objectid == objectid) {
794 btrfs_set_root_refs(root_item, 0);
795 memset(&root_item->drop_progress, 0,
796 sizeof(struct btrfs_disk_key));
797 btrfs_set_root_drop_level(root_item, 0);
800 btrfs_tree_unlock(eb);
801 free_extent_buffer(eb);
803 ret = btrfs_insert_root(trans, fs_info->tree_root,
804 &root_key, root_item);
810 reloc_root = btrfs_read_tree_root(fs_info->tree_root, &root_key);
811 if (IS_ERR(reloc_root)) {
812 ret = PTR_ERR(reloc_root);
815 set_bit(BTRFS_ROOT_SHAREABLE, &reloc_root->state);
816 reloc_root->last_trans = trans->transid;
822 btrfs_abort_transaction(trans, ret);
827 * create reloc tree for a given fs tree. reloc tree is just a
828 * snapshot of the fs tree with special root objectid.
830 * The reloc_root comes out of here with two references, one for
831 * root->reloc_root, and another for being on the rc->reloc_roots list.
833 int btrfs_init_reloc_root(struct btrfs_trans_handle *trans,
834 struct btrfs_root *root)
836 struct btrfs_fs_info *fs_info = root->fs_info;
837 struct btrfs_root *reloc_root;
838 struct reloc_control *rc = fs_info->reloc_ctl;
839 struct btrfs_block_rsv *rsv;
847 * The subvolume has reloc tree but the swap is finished, no need to
848 * create/update the dead reloc tree
850 if (reloc_root_is_dead(root))
854 * This is subtle but important. We do not do
855 * record_root_in_transaction for reloc roots, instead we record their
856 * corresponding fs root, and then here we update the last trans for the
857 * reloc root. This means that we have to do this for the entire life
858 * of the reloc root, regardless of which stage of the relocation we are
861 if (root->reloc_root) {
862 reloc_root = root->reloc_root;
863 reloc_root->last_trans = trans->transid;
868 * We are merging reloc roots, we do not need new reloc trees. Also
869 * reloc trees never need their own reloc tree.
871 if (!rc->create_reloc_tree ||
872 root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID)
875 if (!trans->reloc_reserved) {
876 rsv = trans->block_rsv;
877 trans->block_rsv = rc->block_rsv;
880 reloc_root = create_reloc_root(trans, root, root->root_key.objectid);
882 trans->block_rsv = rsv;
883 if (IS_ERR(reloc_root))
884 return PTR_ERR(reloc_root);
886 ret = __add_reloc_root(reloc_root);
887 ASSERT(ret != -EEXIST);
889 /* Pairs with create_reloc_root */
890 btrfs_put_root(reloc_root);
893 root->reloc_root = btrfs_grab_root(reloc_root);
898 * update root item of reloc tree
900 int btrfs_update_reloc_root(struct btrfs_trans_handle *trans,
901 struct btrfs_root *root)
903 struct btrfs_fs_info *fs_info = root->fs_info;
904 struct btrfs_root *reloc_root;
905 struct btrfs_root_item *root_item;
908 if (!have_reloc_root(root))
911 reloc_root = root->reloc_root;
912 root_item = &reloc_root->root_item;
915 * We are probably ok here, but __del_reloc_root() will drop its ref of
916 * the root. We have the ref for root->reloc_root, but just in case
917 * hold it while we update the reloc root.
919 btrfs_grab_root(reloc_root);
921 /* root->reloc_root will stay until current relocation finished */
922 if (fs_info->reloc_ctl->merge_reloc_tree &&
923 btrfs_root_refs(root_item) == 0) {
924 set_bit(BTRFS_ROOT_DEAD_RELOC_TREE, &root->state);
926 * Mark the tree as dead before we change reloc_root so
927 * have_reloc_root will not touch it from now on.
930 __del_reloc_root(reloc_root);
933 if (reloc_root->commit_root != reloc_root->node) {
934 __update_reloc_root(reloc_root);
935 btrfs_set_root_node(root_item, reloc_root->node);
936 free_extent_buffer(reloc_root->commit_root);
937 reloc_root->commit_root = btrfs_root_node(reloc_root);
940 ret = btrfs_update_root(trans, fs_info->tree_root,
941 &reloc_root->root_key, root_item);
942 btrfs_put_root(reloc_root);
947 * helper to find first cached inode with inode number >= objectid
950 static struct inode *find_next_inode(struct btrfs_root *root, u64 objectid)
952 struct rb_node *node;
953 struct rb_node *prev;
954 struct btrfs_inode *entry;
957 spin_lock(&root->inode_lock);
959 node = root->inode_tree.rb_node;
963 entry = rb_entry(node, struct btrfs_inode, rb_node);
965 if (objectid < btrfs_ino(entry))
966 node = node->rb_left;
967 else if (objectid > btrfs_ino(entry))
968 node = node->rb_right;
974 entry = rb_entry(prev, struct btrfs_inode, rb_node);
975 if (objectid <= btrfs_ino(entry)) {
979 prev = rb_next(prev);
983 entry = rb_entry(node, struct btrfs_inode, rb_node);
984 inode = igrab(&entry->vfs_inode);
986 spin_unlock(&root->inode_lock);
990 objectid = btrfs_ino(entry) + 1;
991 if (cond_resched_lock(&root->inode_lock))
994 node = rb_next(node);
996 spin_unlock(&root->inode_lock);
1001 * get new location of data
1003 static int get_new_location(struct inode *reloc_inode, u64 *new_bytenr,
1004 u64 bytenr, u64 num_bytes)
1006 struct btrfs_root *root = BTRFS_I(reloc_inode)->root;
1007 struct btrfs_path *path;
1008 struct btrfs_file_extent_item *fi;
1009 struct extent_buffer *leaf;
1012 path = btrfs_alloc_path();
1016 bytenr -= BTRFS_I(reloc_inode)->index_cnt;
1017 ret = btrfs_lookup_file_extent(NULL, root, path,
1018 btrfs_ino(BTRFS_I(reloc_inode)), bytenr, 0);
1026 leaf = path->nodes[0];
1027 fi = btrfs_item_ptr(leaf, path->slots[0],
1028 struct btrfs_file_extent_item);
1030 BUG_ON(btrfs_file_extent_offset(leaf, fi) ||
1031 btrfs_file_extent_compression(leaf, fi) ||
1032 btrfs_file_extent_encryption(leaf, fi) ||
1033 btrfs_file_extent_other_encoding(leaf, fi));
1035 if (num_bytes != btrfs_file_extent_disk_num_bytes(leaf, fi)) {
1040 *new_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
1043 btrfs_free_path(path);
1048 * update file extent items in the tree leaf to point to
1049 * the new locations.
1051 static noinline_for_stack
1052 int replace_file_extents(struct btrfs_trans_handle *trans,
1053 struct reloc_control *rc,
1054 struct btrfs_root *root,
1055 struct extent_buffer *leaf)
1057 struct btrfs_fs_info *fs_info = root->fs_info;
1058 struct btrfs_key key;
1059 struct btrfs_file_extent_item *fi;
1060 struct inode *inode = NULL;
1072 if (rc->stage != UPDATE_DATA_PTRS)
1075 /* reloc trees always use full backref */
1076 if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID)
1077 parent = leaf->start;
1081 nritems = btrfs_header_nritems(leaf);
1082 for (i = 0; i < nritems; i++) {
1083 struct btrfs_ref ref = { 0 };
1086 btrfs_item_key_to_cpu(leaf, &key, i);
1087 if (key.type != BTRFS_EXTENT_DATA_KEY)
1089 fi = btrfs_item_ptr(leaf, i, struct btrfs_file_extent_item);
1090 if (btrfs_file_extent_type(leaf, fi) ==
1091 BTRFS_FILE_EXTENT_INLINE)
1093 bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
1094 num_bytes = btrfs_file_extent_disk_num_bytes(leaf, fi);
1097 if (!in_range(bytenr, rc->block_group->start,
1098 rc->block_group->length))
1102 * if we are modifying block in fs tree, wait for readpage
1103 * to complete and drop the extent cache
1105 if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) {
1107 inode = find_next_inode(root, key.objectid);
1109 } else if (inode && btrfs_ino(BTRFS_I(inode)) < key.objectid) {
1110 btrfs_add_delayed_iput(inode);
1111 inode = find_next_inode(root, key.objectid);
1113 if (inode && btrfs_ino(BTRFS_I(inode)) == key.objectid) {
1115 btrfs_file_extent_num_bytes(leaf, fi);
1116 WARN_ON(!IS_ALIGNED(key.offset,
1117 fs_info->sectorsize));
1118 WARN_ON(!IS_ALIGNED(end, fs_info->sectorsize));
1120 ret = try_lock_extent(&BTRFS_I(inode)->io_tree,
1125 btrfs_drop_extent_cache(BTRFS_I(inode),
1126 key.offset, end, 1);
1127 unlock_extent(&BTRFS_I(inode)->io_tree,
1132 ret = get_new_location(rc->data_inode, &new_bytenr,
1136 * Don't have to abort since we've not changed anything
1137 * in the file extent yet.
1142 btrfs_set_file_extent_disk_bytenr(leaf, fi, new_bytenr);
1145 key.offset -= btrfs_file_extent_offset(leaf, fi);
1146 btrfs_init_generic_ref(&ref, BTRFS_ADD_DELAYED_REF, new_bytenr,
1148 ref.real_root = root->root_key.objectid;
1149 btrfs_init_data_ref(&ref, btrfs_header_owner(leaf),
1150 key.objectid, key.offset);
1151 ret = btrfs_inc_extent_ref(trans, &ref);
1153 btrfs_abort_transaction(trans, ret);
1157 btrfs_init_generic_ref(&ref, BTRFS_DROP_DELAYED_REF, bytenr,
1159 ref.real_root = root->root_key.objectid;
1160 btrfs_init_data_ref(&ref, btrfs_header_owner(leaf),
1161 key.objectid, key.offset);
1162 ret = btrfs_free_extent(trans, &ref);
1164 btrfs_abort_transaction(trans, ret);
1169 btrfs_mark_buffer_dirty(leaf);
1171 btrfs_add_delayed_iput(inode);
1175 static noinline_for_stack
1176 int memcmp_node_keys(struct extent_buffer *eb, int slot,
1177 struct btrfs_path *path, int level)
1179 struct btrfs_disk_key key1;
1180 struct btrfs_disk_key key2;
1181 btrfs_node_key(eb, &key1, slot);
1182 btrfs_node_key(path->nodes[level], &key2, path->slots[level]);
1183 return memcmp(&key1, &key2, sizeof(key1));
1187 * try to replace tree blocks in fs tree with the new blocks
1188 * in reloc tree. tree blocks haven't been modified since the
1189 * reloc tree was create can be replaced.
1191 * if a block was replaced, level of the block + 1 is returned.
1192 * if no block got replaced, 0 is returned. if there are other
1193 * errors, a negative error number is returned.
1195 static noinline_for_stack
1196 int replace_path(struct btrfs_trans_handle *trans, struct reloc_control *rc,
1197 struct btrfs_root *dest, struct btrfs_root *src,
1198 struct btrfs_path *path, struct btrfs_key *next_key,
1199 int lowest_level, int max_level)
1201 struct btrfs_fs_info *fs_info = dest->fs_info;
1202 struct extent_buffer *eb;
1203 struct extent_buffer *parent;
1204 struct btrfs_ref ref = { 0 };
1205 struct btrfs_key key;
1217 ASSERT(src->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID);
1218 ASSERT(dest->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID);
1220 last_snapshot = btrfs_root_last_snapshot(&src->root_item);
1222 slot = path->slots[lowest_level];
1223 btrfs_node_key_to_cpu(path->nodes[lowest_level], &key, slot);
1225 eb = btrfs_lock_root_node(dest);
1226 level = btrfs_header_level(eb);
1228 if (level < lowest_level) {
1229 btrfs_tree_unlock(eb);
1230 free_extent_buffer(eb);
1235 ret = btrfs_cow_block(trans, dest, eb, NULL, 0, &eb,
1238 btrfs_tree_unlock(eb);
1239 free_extent_buffer(eb);
1245 next_key->objectid = (u64)-1;
1246 next_key->type = (u8)-1;
1247 next_key->offset = (u64)-1;
1252 level = btrfs_header_level(parent);
1253 ASSERT(level >= lowest_level);
1255 ret = btrfs_bin_search(parent, &key, &slot);
1258 if (ret && slot > 0)
1261 if (next_key && slot + 1 < btrfs_header_nritems(parent))
1262 btrfs_node_key_to_cpu(parent, next_key, slot + 1);
1264 old_bytenr = btrfs_node_blockptr(parent, slot);
1265 blocksize = fs_info->nodesize;
1266 old_ptr_gen = btrfs_node_ptr_generation(parent, slot);
1268 if (level <= max_level) {
1269 eb = path->nodes[level];
1270 new_bytenr = btrfs_node_blockptr(eb,
1271 path->slots[level]);
1272 new_ptr_gen = btrfs_node_ptr_generation(eb,
1273 path->slots[level]);
1279 if (WARN_ON(new_bytenr > 0 && new_bytenr == old_bytenr)) {
1284 if (new_bytenr == 0 || old_ptr_gen > last_snapshot ||
1285 memcmp_node_keys(parent, slot, path, level)) {
1286 if (level <= lowest_level) {
1291 eb = btrfs_read_node_slot(parent, slot);
1296 btrfs_tree_lock(eb);
1298 ret = btrfs_cow_block(trans, dest, eb, parent,
1302 btrfs_tree_unlock(eb);
1303 free_extent_buffer(eb);
1308 btrfs_tree_unlock(parent);
1309 free_extent_buffer(parent);
1316 btrfs_tree_unlock(parent);
1317 free_extent_buffer(parent);
1322 btrfs_node_key_to_cpu(path->nodes[level], &key,
1323 path->slots[level]);
1324 btrfs_release_path(path);
1326 path->lowest_level = level;
1327 ret = btrfs_search_slot(trans, src, &key, path, 0, 1);
1328 path->lowest_level = 0;
1336 * Info qgroup to trace both subtrees.
1338 * We must trace both trees.
1339 * 1) Tree reloc subtree
1340 * If not traced, we will leak data numbers
1342 * If not traced, we will double count old data
1344 * We don't scan the subtree right now, but only record
1345 * the swapped tree blocks.
1346 * The real subtree rescan is delayed until we have new
1347 * CoW on the subtree root node before transaction commit.
1349 ret = btrfs_qgroup_add_swapped_blocks(trans, dest,
1350 rc->block_group, parent, slot,
1351 path->nodes[level], path->slots[level],
1356 * swap blocks in fs tree and reloc tree.
1358 btrfs_set_node_blockptr(parent, slot, new_bytenr);
1359 btrfs_set_node_ptr_generation(parent, slot, new_ptr_gen);
1360 btrfs_mark_buffer_dirty(parent);
1362 btrfs_set_node_blockptr(path->nodes[level],
1363 path->slots[level], old_bytenr);
1364 btrfs_set_node_ptr_generation(path->nodes[level],
1365 path->slots[level], old_ptr_gen);
1366 btrfs_mark_buffer_dirty(path->nodes[level]);
1368 btrfs_init_generic_ref(&ref, BTRFS_ADD_DELAYED_REF, old_bytenr,
1369 blocksize, path->nodes[level]->start);
1370 ref.skip_qgroup = true;
1371 btrfs_init_tree_ref(&ref, level - 1, src->root_key.objectid);
1372 ret = btrfs_inc_extent_ref(trans, &ref);
1374 btrfs_abort_transaction(trans, ret);
1377 btrfs_init_generic_ref(&ref, BTRFS_ADD_DELAYED_REF, new_bytenr,
1379 ref.skip_qgroup = true;
1380 btrfs_init_tree_ref(&ref, level - 1, dest->root_key.objectid);
1381 ret = btrfs_inc_extent_ref(trans, &ref);
1383 btrfs_abort_transaction(trans, ret);
1387 btrfs_init_generic_ref(&ref, BTRFS_DROP_DELAYED_REF, new_bytenr,
1388 blocksize, path->nodes[level]->start);
1389 btrfs_init_tree_ref(&ref, level - 1, src->root_key.objectid);
1390 ref.skip_qgroup = true;
1391 ret = btrfs_free_extent(trans, &ref);
1393 btrfs_abort_transaction(trans, ret);
1397 btrfs_init_generic_ref(&ref, BTRFS_DROP_DELAYED_REF, old_bytenr,
1399 btrfs_init_tree_ref(&ref, level - 1, dest->root_key.objectid);
1400 ref.skip_qgroup = true;
1401 ret = btrfs_free_extent(trans, &ref);
1403 btrfs_abort_transaction(trans, ret);
1407 btrfs_unlock_up_safe(path, 0);
1412 btrfs_tree_unlock(parent);
1413 free_extent_buffer(parent);
1418 * helper to find next relocated block in reloc tree
1420 static noinline_for_stack
1421 int walk_up_reloc_tree(struct btrfs_root *root, struct btrfs_path *path,
1424 struct extent_buffer *eb;
1429 last_snapshot = btrfs_root_last_snapshot(&root->root_item);
1431 for (i = 0; i < *level; i++) {
1432 free_extent_buffer(path->nodes[i]);
1433 path->nodes[i] = NULL;
1436 for (i = *level; i < BTRFS_MAX_LEVEL && path->nodes[i]; i++) {
1437 eb = path->nodes[i];
1438 nritems = btrfs_header_nritems(eb);
1439 while (path->slots[i] + 1 < nritems) {
1441 if (btrfs_node_ptr_generation(eb, path->slots[i]) <=
1448 free_extent_buffer(path->nodes[i]);
1449 path->nodes[i] = NULL;
1455 * walk down reloc tree to find relocated block of lowest level
1457 static noinline_for_stack
1458 int walk_down_reloc_tree(struct btrfs_root *root, struct btrfs_path *path,
1461 struct extent_buffer *eb = NULL;
1467 last_snapshot = btrfs_root_last_snapshot(&root->root_item);
1469 for (i = *level; i > 0; i--) {
1470 eb = path->nodes[i];
1471 nritems = btrfs_header_nritems(eb);
1472 while (path->slots[i] < nritems) {
1473 ptr_gen = btrfs_node_ptr_generation(eb, path->slots[i]);
1474 if (ptr_gen > last_snapshot)
1478 if (path->slots[i] >= nritems) {
1489 eb = btrfs_read_node_slot(eb, path->slots[i]);
1492 BUG_ON(btrfs_header_level(eb) != i - 1);
1493 path->nodes[i - 1] = eb;
1494 path->slots[i - 1] = 0;
1500 * invalidate extent cache for file extents whose key in range of
1501 * [min_key, max_key)
1503 static int invalidate_extent_cache(struct btrfs_root *root,
1504 struct btrfs_key *min_key,
1505 struct btrfs_key *max_key)
1507 struct btrfs_fs_info *fs_info = root->fs_info;
1508 struct inode *inode = NULL;
1513 objectid = min_key->objectid;
1518 if (objectid > max_key->objectid)
1521 inode = find_next_inode(root, objectid);
1524 ino = btrfs_ino(BTRFS_I(inode));
1526 if (ino > max_key->objectid) {
1532 if (!S_ISREG(inode->i_mode))
1535 if (unlikely(min_key->objectid == ino)) {
1536 if (min_key->type > BTRFS_EXTENT_DATA_KEY)
1538 if (min_key->type < BTRFS_EXTENT_DATA_KEY)
1541 start = min_key->offset;
1542 WARN_ON(!IS_ALIGNED(start, fs_info->sectorsize));
1548 if (unlikely(max_key->objectid == ino)) {
1549 if (max_key->type < BTRFS_EXTENT_DATA_KEY)
1551 if (max_key->type > BTRFS_EXTENT_DATA_KEY) {
1554 if (max_key->offset == 0)
1556 end = max_key->offset;
1557 WARN_ON(!IS_ALIGNED(end, fs_info->sectorsize));
1564 /* the lock_extent waits for readpage to complete */
1565 lock_extent(&BTRFS_I(inode)->io_tree, start, end);
1566 btrfs_drop_extent_cache(BTRFS_I(inode), start, end, 1);
1567 unlock_extent(&BTRFS_I(inode)->io_tree, start, end);
1572 static int find_next_key(struct btrfs_path *path, int level,
1573 struct btrfs_key *key)
1576 while (level < BTRFS_MAX_LEVEL) {
1577 if (!path->nodes[level])
1579 if (path->slots[level] + 1 <
1580 btrfs_header_nritems(path->nodes[level])) {
1581 btrfs_node_key_to_cpu(path->nodes[level], key,
1582 path->slots[level] + 1);
1591 * Insert current subvolume into reloc_control::dirty_subvol_roots
1593 static int insert_dirty_subvol(struct btrfs_trans_handle *trans,
1594 struct reloc_control *rc,
1595 struct btrfs_root *root)
1597 struct btrfs_root *reloc_root = root->reloc_root;
1598 struct btrfs_root_item *reloc_root_item;
1601 /* @root must be a subvolume tree root with a valid reloc tree */
1602 ASSERT(root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID);
1605 reloc_root_item = &reloc_root->root_item;
1606 memset(&reloc_root_item->drop_progress, 0,
1607 sizeof(reloc_root_item->drop_progress));
1608 btrfs_set_root_drop_level(reloc_root_item, 0);
1609 btrfs_set_root_refs(reloc_root_item, 0);
1610 ret = btrfs_update_reloc_root(trans, root);
1614 if (list_empty(&root->reloc_dirty_list)) {
1615 btrfs_grab_root(root);
1616 list_add_tail(&root->reloc_dirty_list, &rc->dirty_subvol_roots);
1622 static int clean_dirty_subvols(struct reloc_control *rc)
1624 struct btrfs_root *root;
1625 struct btrfs_root *next;
1629 list_for_each_entry_safe(root, next, &rc->dirty_subvol_roots,
1631 if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) {
1632 /* Merged subvolume, cleanup its reloc root */
1633 struct btrfs_root *reloc_root = root->reloc_root;
1635 list_del_init(&root->reloc_dirty_list);
1636 root->reloc_root = NULL;
1638 * Need barrier to ensure clear_bit() only happens after
1639 * root->reloc_root = NULL. Pairs with have_reloc_root.
1642 clear_bit(BTRFS_ROOT_DEAD_RELOC_TREE, &root->state);
1645 * btrfs_drop_snapshot drops our ref we hold for
1646 * ->reloc_root. If it fails however we must
1647 * drop the ref ourselves.
1649 ret2 = btrfs_drop_snapshot(reloc_root, 0, 1);
1651 btrfs_put_root(reloc_root);
1656 btrfs_put_root(root);
1658 /* Orphan reloc tree, just clean it up */
1659 ret2 = btrfs_drop_snapshot(root, 0, 1);
1661 btrfs_put_root(root);
1671 * merge the relocated tree blocks in reloc tree with corresponding
1674 static noinline_for_stack int merge_reloc_root(struct reloc_control *rc,
1675 struct btrfs_root *root)
1677 struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
1678 struct btrfs_key key;
1679 struct btrfs_key next_key;
1680 struct btrfs_trans_handle *trans = NULL;
1681 struct btrfs_root *reloc_root;
1682 struct btrfs_root_item *root_item;
1683 struct btrfs_path *path;
1684 struct extent_buffer *leaf;
1692 path = btrfs_alloc_path();
1695 path->reada = READA_FORWARD;
1697 reloc_root = root->reloc_root;
1698 root_item = &reloc_root->root_item;
1700 if (btrfs_disk_key_objectid(&root_item->drop_progress) == 0) {
1701 level = btrfs_root_level(root_item);
1702 atomic_inc(&reloc_root->node->refs);
1703 path->nodes[level] = reloc_root->node;
1704 path->slots[level] = 0;
1706 btrfs_disk_key_to_cpu(&key, &root_item->drop_progress);
1708 level = btrfs_root_drop_level(root_item);
1710 path->lowest_level = level;
1711 ret = btrfs_search_slot(NULL, reloc_root, &key, path, 0, 0);
1712 path->lowest_level = 0;
1714 btrfs_free_path(path);
1718 btrfs_node_key_to_cpu(path->nodes[level], &next_key,
1719 path->slots[level]);
1720 WARN_ON(memcmp(&key, &next_key, sizeof(key)));
1722 btrfs_unlock_up_safe(path, 0);
1726 * In merge_reloc_root(), we modify the upper level pointer to swap the
1727 * tree blocks between reloc tree and subvolume tree. Thus for tree
1728 * block COW, we COW at most from level 1 to root level for each tree.
1730 * Thus the needed metadata size is at most root_level * nodesize,
1731 * and * 2 since we have two trees to COW.
1733 reserve_level = max_t(int, 1, btrfs_root_level(root_item));
1734 min_reserved = fs_info->nodesize * reserve_level * 2;
1735 memset(&next_key, 0, sizeof(next_key));
1738 ret = btrfs_block_rsv_refill(root, rc->block_rsv, min_reserved,
1739 BTRFS_RESERVE_FLUSH_LIMIT);
1742 trans = btrfs_start_transaction(root, 0);
1743 if (IS_ERR(trans)) {
1744 ret = PTR_ERR(trans);
1750 * At this point we no longer have a reloc_control, so we can't
1751 * depend on btrfs_init_reloc_root to update our last_trans.
1753 * But that's ok, we started the trans handle on our
1754 * corresponding fs_root, which means it's been added to the
1755 * dirty list. At commit time we'll still call
1756 * btrfs_update_reloc_root() and update our root item
1759 reloc_root->last_trans = trans->transid;
1760 trans->block_rsv = rc->block_rsv;
1765 ret = walk_down_reloc_tree(reloc_root, path, &level);
1771 if (!find_next_key(path, level, &key) &&
1772 btrfs_comp_cpu_keys(&next_key, &key) >= 0) {
1775 ret = replace_path(trans, rc, root, reloc_root, path,
1776 &next_key, level, max_level);
1782 btrfs_node_key_to_cpu(path->nodes[level], &key,
1783 path->slots[level]);
1787 ret = walk_up_reloc_tree(reloc_root, path, &level);
1793 * save the merging progress in the drop_progress.
1794 * this is OK since root refs == 1 in this case.
1796 btrfs_node_key(path->nodes[level], &root_item->drop_progress,
1797 path->slots[level]);
1798 btrfs_set_root_drop_level(root_item, level);
1800 btrfs_end_transaction_throttle(trans);
1803 btrfs_btree_balance_dirty(fs_info);
1805 if (replaced && rc->stage == UPDATE_DATA_PTRS)
1806 invalidate_extent_cache(root, &key, &next_key);
1810 * handle the case only one block in the fs tree need to be
1811 * relocated and the block is tree root.
1813 leaf = btrfs_lock_root_node(root);
1814 ret = btrfs_cow_block(trans, root, leaf, NULL, 0, &leaf,
1816 btrfs_tree_unlock(leaf);
1817 free_extent_buffer(leaf);
1819 btrfs_free_path(path);
1822 ret = insert_dirty_subvol(trans, rc, root);
1824 btrfs_abort_transaction(trans, ret);
1828 btrfs_end_transaction_throttle(trans);
1830 btrfs_btree_balance_dirty(fs_info);
1832 if (replaced && rc->stage == UPDATE_DATA_PTRS)
1833 invalidate_extent_cache(root, &key, &next_key);
1838 static noinline_for_stack
1839 int prepare_to_merge(struct reloc_control *rc, int err)
1841 struct btrfs_root *root = rc->extent_root;
1842 struct btrfs_fs_info *fs_info = root->fs_info;
1843 struct btrfs_root *reloc_root;
1844 struct btrfs_trans_handle *trans;
1845 LIST_HEAD(reloc_roots);
1849 mutex_lock(&fs_info->reloc_mutex);
1850 rc->merging_rsv_size += fs_info->nodesize * (BTRFS_MAX_LEVEL - 1) * 2;
1851 rc->merging_rsv_size += rc->nodes_relocated * 2;
1852 mutex_unlock(&fs_info->reloc_mutex);
1856 num_bytes = rc->merging_rsv_size;
1857 ret = btrfs_block_rsv_add(root, rc->block_rsv, num_bytes,
1858 BTRFS_RESERVE_FLUSH_ALL);
1863 trans = btrfs_join_transaction(rc->extent_root);
1864 if (IS_ERR(trans)) {
1866 btrfs_block_rsv_release(fs_info, rc->block_rsv,
1868 return PTR_ERR(trans);
1872 if (num_bytes != rc->merging_rsv_size) {
1873 btrfs_end_transaction(trans);
1874 btrfs_block_rsv_release(fs_info, rc->block_rsv,
1880 rc->merge_reloc_tree = 1;
1882 while (!list_empty(&rc->reloc_roots)) {
1883 reloc_root = list_entry(rc->reloc_roots.next,
1884 struct btrfs_root, root_list);
1885 list_del_init(&reloc_root->root_list);
1887 root = btrfs_get_fs_root(fs_info, reloc_root->root_key.offset,
1891 * Even if we have an error we need this reloc root
1892 * back on our list so we can clean up properly.
1894 list_add(&reloc_root->root_list, &reloc_roots);
1895 btrfs_abort_transaction(trans, (int)PTR_ERR(root));
1897 err = PTR_ERR(root);
1900 ASSERT(root->reloc_root == reloc_root);
1903 * set reference count to 1, so btrfs_recover_relocation
1904 * knows it should resumes merging
1907 btrfs_set_root_refs(&reloc_root->root_item, 1);
1908 ret = btrfs_update_reloc_root(trans, root);
1911 * Even if we have an error we need this reloc root back on our
1912 * list so we can clean up properly.
1914 list_add(&reloc_root->root_list, &reloc_roots);
1915 btrfs_put_root(root);
1918 btrfs_abort_transaction(trans, ret);
1925 list_splice(&reloc_roots, &rc->reloc_roots);
1928 err = btrfs_commit_transaction(trans);
1930 btrfs_end_transaction(trans);
1934 static noinline_for_stack
1935 void free_reloc_roots(struct list_head *list)
1937 struct btrfs_root *reloc_root, *tmp;
1939 list_for_each_entry_safe(reloc_root, tmp, list, root_list)
1940 __del_reloc_root(reloc_root);
1943 static noinline_for_stack
1944 void merge_reloc_roots(struct reloc_control *rc)
1946 struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
1947 struct btrfs_root *root;
1948 struct btrfs_root *reloc_root;
1949 LIST_HEAD(reloc_roots);
1953 root = rc->extent_root;
1956 * this serializes us with btrfs_record_root_in_transaction,
1957 * we have to make sure nobody is in the middle of
1958 * adding their roots to the list while we are
1961 mutex_lock(&fs_info->reloc_mutex);
1962 list_splice_init(&rc->reloc_roots, &reloc_roots);
1963 mutex_unlock(&fs_info->reloc_mutex);
1965 while (!list_empty(&reloc_roots)) {
1967 reloc_root = list_entry(reloc_roots.next,
1968 struct btrfs_root, root_list);
1970 root = btrfs_get_fs_root(fs_info, reloc_root->root_key.offset,
1972 if (btrfs_root_refs(&reloc_root->root_item) > 0) {
1975 * For recovery we read the fs roots on mount,
1976 * and if we didn't find the root then we marked
1977 * the reloc root as a garbage root. For normal
1978 * relocation obviously the root should exist in
1979 * memory. However there's no reason we can't
1980 * handle the error properly here just in case.
1983 ret = PTR_ERR(root);
1986 if (root->reloc_root != reloc_root) {
1988 * This is actually impossible without something
1989 * going really wrong (like weird race condition
1996 ret = merge_reloc_root(rc, root);
1997 btrfs_put_root(root);
1999 if (list_empty(&reloc_root->root_list))
2000 list_add_tail(&reloc_root->root_list,
2005 if (!IS_ERR(root)) {
2006 if (root->reloc_root == reloc_root) {
2007 root->reloc_root = NULL;
2008 btrfs_put_root(reloc_root);
2010 clear_bit(BTRFS_ROOT_DEAD_RELOC_TREE,
2012 btrfs_put_root(root);
2015 list_del_init(&reloc_root->root_list);
2016 /* Don't forget to queue this reloc root for cleanup */
2017 list_add_tail(&reloc_root->reloc_dirty_list,
2018 &rc->dirty_subvol_roots);
2028 btrfs_handle_fs_error(fs_info, ret, NULL);
2029 free_reloc_roots(&reloc_roots);
2031 /* new reloc root may be added */
2032 mutex_lock(&fs_info->reloc_mutex);
2033 list_splice_init(&rc->reloc_roots, &reloc_roots);
2034 mutex_unlock(&fs_info->reloc_mutex);
2035 free_reloc_roots(&reloc_roots);
2041 * BUG_ON(!RB_EMPTY_ROOT(&rc->reloc_root_tree.rb_root));
2043 * here, but it's wrong. If we fail to start the transaction in
2044 * prepare_to_merge() we will have only 0 ref reloc roots, none of which
2045 * have actually been removed from the reloc_root_tree rb tree. This is
2046 * fine because we're bailing here, and we hold a reference on the root
2047 * for the list that holds it, so these roots will be cleaned up when we
2048 * do the reloc_dirty_list afterwards. Meanwhile the root->reloc_root
2049 * will be cleaned up on unmount.
2051 * The remaining nodes will be cleaned up by free_reloc_control.
2055 static void free_block_list(struct rb_root *blocks)
2057 struct tree_block *block;
2058 struct rb_node *rb_node;
2059 while ((rb_node = rb_first(blocks))) {
2060 block = rb_entry(rb_node, struct tree_block, rb_node);
2061 rb_erase(rb_node, blocks);
2066 static int record_reloc_root_in_trans(struct btrfs_trans_handle *trans,
2067 struct btrfs_root *reloc_root)
2069 struct btrfs_fs_info *fs_info = reloc_root->fs_info;
2070 struct btrfs_root *root;
2073 if (reloc_root->last_trans == trans->transid)
2076 root = btrfs_get_fs_root(fs_info, reloc_root->root_key.offset, false);
2079 * This should succeed, since we can't have a reloc root without having
2080 * already looked up the actual root and created the reloc root for this
2083 * However if there's some sort of corruption where we have a ref to a
2084 * reloc root without a corresponding root this could return ENOENT.
2088 return PTR_ERR(root);
2090 if (root->reloc_root != reloc_root) {
2093 "root %llu has two reloc roots associated with it",
2094 reloc_root->root_key.offset);
2095 btrfs_put_root(root);
2098 ret = btrfs_record_root_in_trans(trans, root);
2099 btrfs_put_root(root);
2104 static noinline_for_stack
2105 struct btrfs_root *select_reloc_root(struct btrfs_trans_handle *trans,
2106 struct reloc_control *rc,
2107 struct btrfs_backref_node *node,
2108 struct btrfs_backref_edge *edges[])
2110 struct btrfs_backref_node *next;
2111 struct btrfs_root *root;
2118 next = walk_up_backref(next, edges, &index);
2122 * If there is no root, then our references for this block are
2123 * incomplete, as we should be able to walk all the way up to a
2124 * block that is owned by a root.
2126 * This path is only for SHAREABLE roots, so if we come upon a
2127 * non-SHAREABLE root then we have backrefs that resolve
2130 * Both of these cases indicate file system corruption, or a bug
2131 * in the backref walking code.
2135 btrfs_err(trans->fs_info,
2136 "bytenr %llu doesn't have a backref path ending in a root",
2138 return ERR_PTR(-EUCLEAN);
2140 if (!test_bit(BTRFS_ROOT_SHAREABLE, &root->state)) {
2142 btrfs_err(trans->fs_info,
2143 "bytenr %llu has multiple refs with one ending in a non-shareable root",
2145 return ERR_PTR(-EUCLEAN);
2148 if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID) {
2149 ret = record_reloc_root_in_trans(trans, root);
2151 return ERR_PTR(ret);
2155 ret = btrfs_record_root_in_trans(trans, root);
2157 return ERR_PTR(ret);
2158 root = root->reloc_root;
2161 * We could have raced with another thread which failed, so
2162 * root->reloc_root may not be set, return ENOENT in this case.
2165 return ERR_PTR(-ENOENT);
2167 if (next->new_bytenr != root->node->start) {
2169 * We just created the reloc root, so we shouldn't have
2170 * ->new_bytenr set and this shouldn't be in the changed
2171 * list. If it is then we have multiple roots pointing
2172 * at the same bytenr which indicates corruption, or
2173 * we've made a mistake in the backref walking code.
2175 ASSERT(next->new_bytenr == 0);
2176 ASSERT(list_empty(&next->list));
2177 if (next->new_bytenr || !list_empty(&next->list)) {
2178 btrfs_err(trans->fs_info,
2179 "bytenr %llu possibly has multiple roots pointing at the same bytenr %llu",
2180 node->bytenr, next->bytenr);
2181 return ERR_PTR(-EUCLEAN);
2184 next->new_bytenr = root->node->start;
2185 btrfs_put_root(next->root);
2186 next->root = btrfs_grab_root(root);
2188 list_add_tail(&next->list,
2189 &rc->backref_cache.changed);
2190 mark_block_processed(rc, next);
2196 next = walk_down_backref(edges, &index);
2197 if (!next || next->level <= node->level)
2202 * This can happen if there's fs corruption or if there's a bug
2203 * in the backref lookup code.
2206 return ERR_PTR(-ENOENT);
2210 /* setup backref node path for btrfs_reloc_cow_block */
2212 rc->backref_cache.path[next->level] = next;
2215 next = edges[index]->node[UPPER];
2221 * Select a tree root for relocation.
2223 * Return NULL if the block is not shareable. We should use do_relocation() in
2226 * Return a tree root pointer if the block is shareable.
2227 * Return -ENOENT if the block is root of reloc tree.
2229 static noinline_for_stack
2230 struct btrfs_root *select_one_root(struct btrfs_backref_node *node)
2232 struct btrfs_backref_node *next;
2233 struct btrfs_root *root;
2234 struct btrfs_root *fs_root = NULL;
2235 struct btrfs_backref_edge *edges[BTRFS_MAX_LEVEL - 1];
2241 next = walk_up_backref(next, edges, &index);
2245 * This can occur if we have incomplete extent refs leading all
2246 * the way up a particular path, in this case return -EUCLEAN.
2249 return ERR_PTR(-EUCLEAN);
2251 /* No other choice for non-shareable tree */
2252 if (!test_bit(BTRFS_ROOT_SHAREABLE, &root->state))
2255 if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID)
2261 next = walk_down_backref(edges, &index);
2262 if (!next || next->level <= node->level)
2267 return ERR_PTR(-ENOENT);
2271 static noinline_for_stack
2272 u64 calcu_metadata_size(struct reloc_control *rc,
2273 struct btrfs_backref_node *node, int reserve)
2275 struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
2276 struct btrfs_backref_node *next = node;
2277 struct btrfs_backref_edge *edge;
2278 struct btrfs_backref_edge *edges[BTRFS_MAX_LEVEL - 1];
2282 BUG_ON(reserve && node->processed);
2287 if (next->processed && (reserve || next != node))
2290 num_bytes += fs_info->nodesize;
2292 if (list_empty(&next->upper))
2295 edge = list_entry(next->upper.next,
2296 struct btrfs_backref_edge, list[LOWER]);
2297 edges[index++] = edge;
2298 next = edge->node[UPPER];
2300 next = walk_down_backref(edges, &index);
2305 static int reserve_metadata_space(struct btrfs_trans_handle *trans,
2306 struct reloc_control *rc,
2307 struct btrfs_backref_node *node)
2309 struct btrfs_root *root = rc->extent_root;
2310 struct btrfs_fs_info *fs_info = root->fs_info;
2315 num_bytes = calcu_metadata_size(rc, node, 1) * 2;
2317 trans->block_rsv = rc->block_rsv;
2318 rc->reserved_bytes += num_bytes;
2321 * We are under a transaction here so we can only do limited flushing.
2322 * If we get an enospc just kick back -EAGAIN so we know to drop the
2323 * transaction and try to refill when we can flush all the things.
2325 ret = btrfs_block_rsv_refill(root, rc->block_rsv, num_bytes,
2326 BTRFS_RESERVE_FLUSH_LIMIT);
2328 tmp = fs_info->nodesize * RELOCATION_RESERVED_NODES;
2329 while (tmp <= rc->reserved_bytes)
2332 * only one thread can access block_rsv at this point,
2333 * so we don't need hold lock to protect block_rsv.
2334 * we expand more reservation size here to allow enough
2335 * space for relocation and we will return earlier in
2338 rc->block_rsv->size = tmp + fs_info->nodesize *
2339 RELOCATION_RESERVED_NODES;
2347 * relocate a block tree, and then update pointers in upper level
2348 * blocks that reference the block to point to the new location.
2350 * if called by link_to_upper, the block has already been relocated.
2351 * in that case this function just updates pointers.
2353 static int do_relocation(struct btrfs_trans_handle *trans,
2354 struct reloc_control *rc,
2355 struct btrfs_backref_node *node,
2356 struct btrfs_key *key,
2357 struct btrfs_path *path, int lowest)
2359 struct btrfs_backref_node *upper;
2360 struct btrfs_backref_edge *edge;
2361 struct btrfs_backref_edge *edges[BTRFS_MAX_LEVEL - 1];
2362 struct btrfs_root *root;
2363 struct extent_buffer *eb;
2370 * If we are lowest then this is the first time we're processing this
2371 * block, and thus shouldn't have an eb associated with it yet.
2373 ASSERT(!lowest || !node->eb);
2375 path->lowest_level = node->level + 1;
2376 rc->backref_cache.path[node->level] = node;
2377 list_for_each_entry(edge, &node->upper, list[LOWER]) {
2378 struct btrfs_ref ref = { 0 };
2382 upper = edge->node[UPPER];
2383 root = select_reloc_root(trans, rc, upper, edges);
2385 ret = PTR_ERR(root);
2389 if (upper->eb && !upper->locked) {
2391 ret = btrfs_bin_search(upper->eb, key, &slot);
2395 bytenr = btrfs_node_blockptr(upper->eb, slot);
2396 if (node->eb->start == bytenr)
2399 btrfs_backref_drop_node_buffer(upper);
2403 ret = btrfs_search_slot(trans, root, key, path, 0, 1);
2408 btrfs_release_path(path);
2413 upper->eb = path->nodes[upper->level];
2414 path->nodes[upper->level] = NULL;
2416 BUG_ON(upper->eb != path->nodes[upper->level]);
2420 path->locks[upper->level] = 0;
2422 slot = path->slots[upper->level];
2423 btrfs_release_path(path);
2425 ret = btrfs_bin_search(upper->eb, key, &slot);
2431 bytenr = btrfs_node_blockptr(upper->eb, slot);
2433 if (bytenr != node->bytenr) {
2434 btrfs_err(root->fs_info,
2435 "lowest leaf/node mismatch: bytenr %llu node->bytenr %llu slot %d upper %llu",
2436 bytenr, node->bytenr, slot,
2442 if (node->eb->start == bytenr)
2446 blocksize = root->fs_info->nodesize;
2447 eb = btrfs_read_node_slot(upper->eb, slot);
2452 btrfs_tree_lock(eb);
2455 ret = btrfs_cow_block(trans, root, eb, upper->eb,
2456 slot, &eb, BTRFS_NESTING_COW);
2457 btrfs_tree_unlock(eb);
2458 free_extent_buffer(eb);
2462 * We've just COWed this block, it should have updated
2463 * the correct backref node entry.
2465 ASSERT(node->eb == eb);
2467 btrfs_set_node_blockptr(upper->eb, slot,
2469 btrfs_set_node_ptr_generation(upper->eb, slot,
2471 btrfs_mark_buffer_dirty(upper->eb);
2473 btrfs_init_generic_ref(&ref, BTRFS_ADD_DELAYED_REF,
2474 node->eb->start, blocksize,
2476 ref.real_root = root->root_key.objectid;
2477 btrfs_init_tree_ref(&ref, node->level,
2478 btrfs_header_owner(upper->eb));
2479 ret = btrfs_inc_extent_ref(trans, &ref);
2481 ret = btrfs_drop_subtree(trans, root, eb,
2484 btrfs_abort_transaction(trans, ret);
2487 if (!upper->pending)
2488 btrfs_backref_drop_node_buffer(upper);
2490 btrfs_backref_unlock_node_buffer(upper);
2495 if (!ret && node->pending) {
2496 btrfs_backref_drop_node_buffer(node);
2497 list_move_tail(&node->list, &rc->backref_cache.changed);
2501 path->lowest_level = 0;
2504 * We should have allocated all of our space in the block rsv and thus
2507 ASSERT(ret != -ENOSPC);
2511 static int link_to_upper(struct btrfs_trans_handle *trans,
2512 struct reloc_control *rc,
2513 struct btrfs_backref_node *node,
2514 struct btrfs_path *path)
2516 struct btrfs_key key;
2518 btrfs_node_key_to_cpu(node->eb, &key, 0);
2519 return do_relocation(trans, rc, node, &key, path, 0);
2522 static int finish_pending_nodes(struct btrfs_trans_handle *trans,
2523 struct reloc_control *rc,
2524 struct btrfs_path *path, int err)
2527 struct btrfs_backref_cache *cache = &rc->backref_cache;
2528 struct btrfs_backref_node *node;
2532 for (level = 0; level < BTRFS_MAX_LEVEL; level++) {
2533 while (!list_empty(&cache->pending[level])) {
2534 node = list_entry(cache->pending[level].next,
2535 struct btrfs_backref_node, list);
2536 list_move_tail(&node->list, &list);
2537 BUG_ON(!node->pending);
2540 ret = link_to_upper(trans, rc, node, path);
2545 list_splice_init(&list, &cache->pending[level]);
2551 * mark a block and all blocks directly/indirectly reference the block
2554 static void update_processed_blocks(struct reloc_control *rc,
2555 struct btrfs_backref_node *node)
2557 struct btrfs_backref_node *next = node;
2558 struct btrfs_backref_edge *edge;
2559 struct btrfs_backref_edge *edges[BTRFS_MAX_LEVEL - 1];
2565 if (next->processed)
2568 mark_block_processed(rc, next);
2570 if (list_empty(&next->upper))
2573 edge = list_entry(next->upper.next,
2574 struct btrfs_backref_edge, list[LOWER]);
2575 edges[index++] = edge;
2576 next = edge->node[UPPER];
2578 next = walk_down_backref(edges, &index);
2582 static int tree_block_processed(u64 bytenr, struct reloc_control *rc)
2584 u32 blocksize = rc->extent_root->fs_info->nodesize;
2586 if (test_range_bit(&rc->processed_blocks, bytenr,
2587 bytenr + blocksize - 1, EXTENT_DIRTY, 1, NULL))
2592 static int get_tree_block_key(struct btrfs_fs_info *fs_info,
2593 struct tree_block *block)
2595 struct extent_buffer *eb;
2597 eb = read_tree_block(fs_info, block->bytenr, block->owner,
2598 block->key.offset, block->level, NULL);
2601 } else if (!extent_buffer_uptodate(eb)) {
2602 free_extent_buffer(eb);
2605 if (block->level == 0)
2606 btrfs_item_key_to_cpu(eb, &block->key, 0);
2608 btrfs_node_key_to_cpu(eb, &block->key, 0);
2609 free_extent_buffer(eb);
2610 block->key_ready = 1;
2615 * helper function to relocate a tree block
2617 static int relocate_tree_block(struct btrfs_trans_handle *trans,
2618 struct reloc_control *rc,
2619 struct btrfs_backref_node *node,
2620 struct btrfs_key *key,
2621 struct btrfs_path *path)
2623 struct btrfs_root *root;
2630 * If we fail here we want to drop our backref_node because we are going
2631 * to start over and regenerate the tree for it.
2633 ret = reserve_metadata_space(trans, rc, node);
2637 BUG_ON(node->processed);
2638 root = select_one_root(node);
2640 ret = PTR_ERR(root);
2642 /* See explanation in select_one_root for the -EUCLEAN case. */
2643 ASSERT(ret == -ENOENT);
2644 if (ret == -ENOENT) {
2646 update_processed_blocks(rc, node);
2652 if (test_bit(BTRFS_ROOT_SHAREABLE, &root->state)) {
2654 * This block was the root block of a root, and this is
2655 * the first time we're processing the block and thus it
2656 * should not have had the ->new_bytenr modified and
2657 * should have not been included on the changed list.
2659 * However in the case of corruption we could have
2660 * multiple refs pointing to the same block improperly,
2661 * and thus we would trip over these checks. ASSERT()
2662 * for the developer case, because it could indicate a
2663 * bug in the backref code, however error out for a
2664 * normal user in the case of corruption.
2666 ASSERT(node->new_bytenr == 0);
2667 ASSERT(list_empty(&node->list));
2668 if (node->new_bytenr || !list_empty(&node->list)) {
2669 btrfs_err(root->fs_info,
2670 "bytenr %llu has improper references to it",
2675 ret = btrfs_record_root_in_trans(trans, root);
2679 * Another thread could have failed, need to check if we
2680 * have reloc_root actually set.
2682 if (!root->reloc_root) {
2686 root = root->reloc_root;
2687 node->new_bytenr = root->node->start;
2688 btrfs_put_root(node->root);
2689 node->root = btrfs_grab_root(root);
2691 list_add_tail(&node->list, &rc->backref_cache.changed);
2693 path->lowest_level = node->level;
2694 ret = btrfs_search_slot(trans, root, key, path, 0, 1);
2695 btrfs_release_path(path);
2700 update_processed_blocks(rc, node);
2702 ret = do_relocation(trans, rc, node, key, path, 1);
2705 if (ret || node->level == 0 || node->cowonly)
2706 btrfs_backref_cleanup_node(&rc->backref_cache, node);
2711 * relocate a list of blocks
2713 static noinline_for_stack
2714 int relocate_tree_blocks(struct btrfs_trans_handle *trans,
2715 struct reloc_control *rc, struct rb_root *blocks)
2717 struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
2718 struct btrfs_backref_node *node;
2719 struct btrfs_path *path;
2720 struct tree_block *block;
2721 struct tree_block *next;
2725 path = btrfs_alloc_path();
2728 goto out_free_blocks;
2731 /* Kick in readahead for tree blocks with missing keys */
2732 rbtree_postorder_for_each_entry_safe(block, next, blocks, rb_node) {
2733 if (!block->key_ready)
2734 btrfs_readahead_tree_block(fs_info, block->bytenr,
2739 /* Get first keys */
2740 rbtree_postorder_for_each_entry_safe(block, next, blocks, rb_node) {
2741 if (!block->key_ready) {
2742 err = get_tree_block_key(fs_info, block);
2748 /* Do tree relocation */
2749 rbtree_postorder_for_each_entry_safe(block, next, blocks, rb_node) {
2750 node = build_backref_tree(rc, &block->key,
2751 block->level, block->bytenr);
2753 err = PTR_ERR(node);
2757 ret = relocate_tree_block(trans, rc, node, &block->key,
2765 err = finish_pending_nodes(trans, rc, path, err);
2768 btrfs_free_path(path);
2770 free_block_list(blocks);
2774 static noinline_for_stack int prealloc_file_extent_cluster(
2775 struct btrfs_inode *inode,
2776 struct file_extent_cluster *cluster)
2781 u64 offset = inode->index_cnt;
2785 u64 i_size = i_size_read(&inode->vfs_inode);
2786 u64 prealloc_start = cluster->start - offset;
2787 u64 prealloc_end = cluster->end - offset;
2788 u64 cur_offset = prealloc_start;
2791 * For subpage case, previous i_size may not be aligned to PAGE_SIZE.
2792 * This means the range [i_size, PAGE_END + 1) is filled with zeros by
2793 * btrfs_do_readpage() call of previously relocated file cluster.
2795 * If the current cluster starts in the above range, btrfs_do_readpage()
2796 * will skip the read, and relocate_one_page() will later writeback
2797 * the padding zeros as new data, causing data corruption.
2799 * Here we have to manually invalidate the range (i_size, PAGE_END + 1).
2801 if (!IS_ALIGNED(i_size, PAGE_SIZE)) {
2802 struct address_space *mapping = inode->vfs_inode.i_mapping;
2803 struct btrfs_fs_info *fs_info = inode->root->fs_info;
2804 const u32 sectorsize = fs_info->sectorsize;
2807 ASSERT(sectorsize < PAGE_SIZE);
2808 ASSERT(IS_ALIGNED(i_size, sectorsize));
2811 * Subpage can't handle page with DIRTY but without UPTODATE
2812 * bit as it can lead to the following deadlock:
2815 * | Page already *locked*
2816 * |- btrfs_lock_and_flush_ordered_range()
2817 * |- btrfs_start_ordered_extent()
2818 * |- extent_write_cache_pages()
2820 * We try to lock the page we already hold.
2822 * Here we just writeback the whole data reloc inode, so that
2823 * we will be ensured to have no dirty range in the page, and
2824 * are safe to clear the uptodate bits.
2826 * This shouldn't cause too much overhead, as we need to write
2827 * the data back anyway.
2829 ret = filemap_write_and_wait(mapping);
2833 clear_extent_bits(&inode->io_tree, i_size,
2834 round_up(i_size, PAGE_SIZE) - 1,
2836 page = find_lock_page(mapping, i_size >> PAGE_SHIFT);
2838 * If page is freed we don't need to do anything then, as we
2839 * will re-read the whole page anyway.
2842 btrfs_subpage_clear_uptodate(fs_info, page, i_size,
2843 round_up(i_size, PAGE_SIZE) - i_size);
2849 BUG_ON(cluster->start != cluster->boundary[0]);
2850 ret = btrfs_alloc_data_chunk_ondemand(inode,
2851 prealloc_end + 1 - prealloc_start);
2855 btrfs_inode_lock(&inode->vfs_inode, 0);
2856 for (nr = 0; nr < cluster->nr; nr++) {
2857 start = cluster->boundary[nr] - offset;
2858 if (nr + 1 < cluster->nr)
2859 end = cluster->boundary[nr + 1] - 1 - offset;
2861 end = cluster->end - offset;
2863 lock_extent(&inode->io_tree, start, end);
2864 num_bytes = end + 1 - start;
2865 ret = btrfs_prealloc_file_range(&inode->vfs_inode, 0, start,
2866 num_bytes, num_bytes,
2867 end + 1, &alloc_hint);
2868 cur_offset = end + 1;
2869 unlock_extent(&inode->io_tree, start, end);
2873 btrfs_inode_unlock(&inode->vfs_inode, 0);
2875 if (cur_offset < prealloc_end)
2876 btrfs_free_reserved_data_space_noquota(inode->root->fs_info,
2877 prealloc_end + 1 - cur_offset);
2881 static noinline_for_stack
2882 int setup_extent_mapping(struct inode *inode, u64 start, u64 end,
2885 struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
2886 struct extent_map *em;
2889 em = alloc_extent_map();
2894 em->len = end + 1 - start;
2895 em->block_len = em->len;
2896 em->block_start = block_start;
2897 set_bit(EXTENT_FLAG_PINNED, &em->flags);
2899 lock_extent(&BTRFS_I(inode)->io_tree, start, end);
2901 write_lock(&em_tree->lock);
2902 ret = add_extent_mapping(em_tree, em, 0);
2903 write_unlock(&em_tree->lock);
2904 if (ret != -EEXIST) {
2905 free_extent_map(em);
2908 btrfs_drop_extent_cache(BTRFS_I(inode), start, end, 0);
2910 unlock_extent(&BTRFS_I(inode)->io_tree, start, end);
2915 * Allow error injection to test balance/relocation cancellation
2917 noinline int btrfs_should_cancel_balance(struct btrfs_fs_info *fs_info)
2919 return atomic_read(&fs_info->balance_cancel_req) ||
2920 atomic_read(&fs_info->reloc_cancel_req) ||
2921 fatal_signal_pending(current);
2923 ALLOW_ERROR_INJECTION(btrfs_should_cancel_balance, TRUE);
2925 static u64 get_cluster_boundary_end(struct file_extent_cluster *cluster,
2928 /* Last extent, use cluster end directly */
2929 if (cluster_nr >= cluster->nr - 1)
2930 return cluster->end;
2932 /* Use next boundary start*/
2933 return cluster->boundary[cluster_nr + 1] - 1;
2936 static int relocate_one_page(struct inode *inode, struct file_ra_state *ra,
2937 struct file_extent_cluster *cluster,
2938 int *cluster_nr, unsigned long page_index)
2940 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2941 u64 offset = BTRFS_I(inode)->index_cnt;
2942 const unsigned long last_index = (cluster->end - offset) >> PAGE_SHIFT;
2943 gfp_t mask = btrfs_alloc_write_mask(inode->i_mapping);
2950 ASSERT(page_index <= last_index);
2951 page = find_lock_page(inode->i_mapping, page_index);
2953 page_cache_sync_readahead(inode->i_mapping, ra, NULL,
2954 page_index, last_index + 1 - page_index);
2955 page = find_or_create_page(inode->i_mapping, page_index, mask);
2959 ret = set_page_extent_mapped(page);
2963 if (PageReadahead(page))
2964 page_cache_async_readahead(inode->i_mapping, ra, NULL, page,
2965 page_index, last_index + 1 - page_index);
2967 if (!PageUptodate(page)) {
2968 btrfs_readpage(NULL, page);
2970 if (!PageUptodate(page)) {
2976 page_start = page_offset(page);
2977 page_end = page_start + PAGE_SIZE - 1;
2980 * Start from the cluster, as for subpage case, the cluster can start
2983 cur = max(page_start, cluster->boundary[*cluster_nr] - offset);
2984 while (cur <= page_end) {
2985 u64 extent_start = cluster->boundary[*cluster_nr] - offset;
2986 u64 extent_end = get_cluster_boundary_end(cluster,
2987 *cluster_nr) - offset;
2988 u64 clamped_start = max(page_start, extent_start);
2989 u64 clamped_end = min(page_end, extent_end);
2990 u32 clamped_len = clamped_end + 1 - clamped_start;
2992 /* Reserve metadata for this range */
2993 ret = btrfs_delalloc_reserve_metadata(BTRFS_I(inode),
2998 /* Mark the range delalloc and dirty for later writeback */
2999 lock_extent(&BTRFS_I(inode)->io_tree, clamped_start, clamped_end);
3000 ret = btrfs_set_extent_delalloc(BTRFS_I(inode), clamped_start,
3001 clamped_end, 0, NULL);
3003 clear_extent_bits(&BTRFS_I(inode)->io_tree,
3004 clamped_start, clamped_end,
3005 EXTENT_LOCKED | EXTENT_BOUNDARY);
3006 btrfs_delalloc_release_metadata(BTRFS_I(inode),
3008 btrfs_delalloc_release_extents(BTRFS_I(inode),
3012 btrfs_page_set_dirty(fs_info, page, clamped_start, clamped_len);
3015 * Set the boundary if it's inside the page.
3016 * Data relocation requires the destination extents to have the
3017 * same size as the source.
3018 * EXTENT_BOUNDARY bit prevents current extent from being merged
3019 * with previous extent.
3021 if (in_range(cluster->boundary[*cluster_nr] - offset,
3022 page_start, PAGE_SIZE)) {
3023 u64 boundary_start = cluster->boundary[*cluster_nr] -
3025 u64 boundary_end = boundary_start +
3026 fs_info->sectorsize - 1;
3028 set_extent_bits(&BTRFS_I(inode)->io_tree,
3029 boundary_start, boundary_end,
3032 unlock_extent(&BTRFS_I(inode)->io_tree, clamped_start, clamped_end);
3033 btrfs_delalloc_release_extents(BTRFS_I(inode), clamped_len);
3036 /* Crossed extent end, go to next extent */
3037 if (cur >= extent_end) {
3039 /* Just finished the last extent of the cluster, exit. */
3040 if (*cluster_nr >= cluster->nr)
3047 balance_dirty_pages_ratelimited(inode->i_mapping);
3048 btrfs_throttle(fs_info);
3049 if (btrfs_should_cancel_balance(fs_info))
3059 static int relocate_file_extent_cluster(struct inode *inode,
3060 struct file_extent_cluster *cluster)
3062 u64 offset = BTRFS_I(inode)->index_cnt;
3063 unsigned long index;
3064 unsigned long last_index;
3065 struct file_ra_state *ra;
3072 ra = kzalloc(sizeof(*ra), GFP_NOFS);
3076 ret = prealloc_file_extent_cluster(BTRFS_I(inode), cluster);
3080 file_ra_state_init(ra, inode->i_mapping);
3082 ret = setup_extent_mapping(inode, cluster->start - offset,
3083 cluster->end - offset, cluster->start);
3087 last_index = (cluster->end - offset) >> PAGE_SHIFT;
3088 for (index = (cluster->start - offset) >> PAGE_SHIFT;
3089 index <= last_index && !ret; index++)
3090 ret = relocate_one_page(inode, ra, cluster, &cluster_nr, index);
3092 WARN_ON(cluster_nr != cluster->nr);
3098 static noinline_for_stack
3099 int relocate_data_extent(struct inode *inode, struct btrfs_key *extent_key,
3100 struct file_extent_cluster *cluster)
3104 if (cluster->nr > 0 && extent_key->objectid != cluster->end + 1) {
3105 ret = relocate_file_extent_cluster(inode, cluster);
3112 cluster->start = extent_key->objectid;
3114 BUG_ON(cluster->nr >= MAX_EXTENTS);
3115 cluster->end = extent_key->objectid + extent_key->offset - 1;
3116 cluster->boundary[cluster->nr] = extent_key->objectid;
3119 if (cluster->nr >= MAX_EXTENTS) {
3120 ret = relocate_file_extent_cluster(inode, cluster);
3129 * helper to add a tree block to the list.
3130 * the major work is getting the generation and level of the block
3132 static int add_tree_block(struct reloc_control *rc,
3133 struct btrfs_key *extent_key,
3134 struct btrfs_path *path,
3135 struct rb_root *blocks)
3137 struct extent_buffer *eb;
3138 struct btrfs_extent_item *ei;
3139 struct btrfs_tree_block_info *bi;
3140 struct tree_block *block;
3141 struct rb_node *rb_node;
3147 eb = path->nodes[0];
3148 item_size = btrfs_item_size_nr(eb, path->slots[0]);
3150 if (extent_key->type == BTRFS_METADATA_ITEM_KEY ||
3151 item_size >= sizeof(*ei) + sizeof(*bi)) {
3152 unsigned long ptr = 0, end;
3154 ei = btrfs_item_ptr(eb, path->slots[0],
3155 struct btrfs_extent_item);
3156 end = (unsigned long)ei + item_size;
3157 if (extent_key->type == BTRFS_EXTENT_ITEM_KEY) {
3158 bi = (struct btrfs_tree_block_info *)(ei + 1);
3159 level = btrfs_tree_block_level(eb, bi);
3160 ptr = (unsigned long)(bi + 1);
3162 level = (int)extent_key->offset;
3163 ptr = (unsigned long)(ei + 1);
3165 generation = btrfs_extent_generation(eb, ei);
3168 * We're reading random blocks without knowing their owner ahead
3169 * of time. This is ok most of the time, as all reloc roots and
3170 * fs roots have the same lock type. However normal trees do
3171 * not, and the only way to know ahead of time is to read the
3172 * inline ref offset. We know it's an fs root if
3174 * 1. There's more than one ref.
3175 * 2. There's a SHARED_DATA_REF_KEY set.
3176 * 3. FULL_BACKREF is set on the flags.
3178 * Otherwise it's safe to assume that the ref offset == the
3179 * owner of this block, so we can use that when calling
3182 if (btrfs_extent_refs(eb, ei) == 1 &&
3183 !(btrfs_extent_flags(eb, ei) &
3184 BTRFS_BLOCK_FLAG_FULL_BACKREF) &&
3186 struct btrfs_extent_inline_ref *iref;
3189 iref = (struct btrfs_extent_inline_ref *)ptr;
3190 type = btrfs_get_extent_inline_ref_type(eb, iref,
3191 BTRFS_REF_TYPE_BLOCK);
3192 if (type == BTRFS_REF_TYPE_INVALID)
3194 if (type == BTRFS_TREE_BLOCK_REF_KEY)
3195 owner = btrfs_extent_inline_ref_offset(eb, iref);
3197 } else if (unlikely(item_size == sizeof(struct btrfs_extent_item_v0))) {
3198 btrfs_print_v0_err(eb->fs_info);
3199 btrfs_handle_fs_error(eb->fs_info, -EINVAL, NULL);
3205 btrfs_release_path(path);
3207 BUG_ON(level == -1);
3209 block = kmalloc(sizeof(*block), GFP_NOFS);
3213 block->bytenr = extent_key->objectid;
3214 block->key.objectid = rc->extent_root->fs_info->nodesize;
3215 block->key.offset = generation;
3216 block->level = level;
3217 block->key_ready = 0;
3218 block->owner = owner;
3220 rb_node = rb_simple_insert(blocks, block->bytenr, &block->rb_node);
3222 btrfs_backref_panic(rc->extent_root->fs_info, block->bytenr,
3229 * helper to add tree blocks for backref of type BTRFS_SHARED_DATA_REF_KEY
3231 static int __add_tree_block(struct reloc_control *rc,
3232 u64 bytenr, u32 blocksize,
3233 struct rb_root *blocks)
3235 struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
3236 struct btrfs_path *path;
3237 struct btrfs_key key;
3239 bool skinny = btrfs_fs_incompat(fs_info, SKINNY_METADATA);
3241 if (tree_block_processed(bytenr, rc))
3244 if (rb_simple_search(blocks, bytenr))
3247 path = btrfs_alloc_path();
3251 key.objectid = bytenr;
3253 key.type = BTRFS_METADATA_ITEM_KEY;
3254 key.offset = (u64)-1;
3256 key.type = BTRFS_EXTENT_ITEM_KEY;
3257 key.offset = blocksize;
3260 path->search_commit_root = 1;
3261 path->skip_locking = 1;
3262 ret = btrfs_search_slot(NULL, rc->extent_root, &key, path, 0, 0);
3266 if (ret > 0 && skinny) {
3267 if (path->slots[0]) {
3269 btrfs_item_key_to_cpu(path->nodes[0], &key,
3271 if (key.objectid == bytenr &&
3272 (key.type == BTRFS_METADATA_ITEM_KEY ||
3273 (key.type == BTRFS_EXTENT_ITEM_KEY &&
3274 key.offset == blocksize)))
3280 btrfs_release_path(path);
3286 btrfs_print_leaf(path->nodes[0]);
3288 "tree block extent item (%llu) is not found in extent tree",
3295 ret = add_tree_block(rc, &key, path, blocks);
3297 btrfs_free_path(path);
3301 static int delete_block_group_cache(struct btrfs_fs_info *fs_info,
3302 struct btrfs_block_group *block_group,
3303 struct inode *inode,
3306 struct btrfs_root *root = fs_info->tree_root;
3307 struct btrfs_trans_handle *trans;
3313 inode = btrfs_iget(fs_info->sb, ino, root);
3318 ret = btrfs_check_trunc_cache_free_space(fs_info,
3319 &fs_info->global_block_rsv);
3323 trans = btrfs_join_transaction(root);
3324 if (IS_ERR(trans)) {
3325 ret = PTR_ERR(trans);
3329 ret = btrfs_truncate_free_space_cache(trans, block_group, inode);
3331 btrfs_end_transaction(trans);
3332 btrfs_btree_balance_dirty(fs_info);
3339 * Locate the free space cache EXTENT_DATA in root tree leaf and delete the
3340 * cache inode, to avoid free space cache data extent blocking data relocation.
3342 static int delete_v1_space_cache(struct extent_buffer *leaf,
3343 struct btrfs_block_group *block_group,
3346 u64 space_cache_ino;
3347 struct btrfs_file_extent_item *ei;
3348 struct btrfs_key key;
3353 if (btrfs_header_owner(leaf) != BTRFS_ROOT_TREE_OBJECTID)
3356 for (i = 0; i < btrfs_header_nritems(leaf); i++) {
3359 btrfs_item_key_to_cpu(leaf, &key, i);
3360 if (key.type != BTRFS_EXTENT_DATA_KEY)
3362 ei = btrfs_item_ptr(leaf, i, struct btrfs_file_extent_item);
3363 type = btrfs_file_extent_type(leaf, ei);
3365 if ((type == BTRFS_FILE_EXTENT_REG ||
3366 type == BTRFS_FILE_EXTENT_PREALLOC) &&
3367 btrfs_file_extent_disk_bytenr(leaf, ei) == data_bytenr) {
3369 space_cache_ino = key.objectid;
3375 ret = delete_block_group_cache(leaf->fs_info, block_group, NULL,
3381 * helper to find all tree blocks that reference a given data extent
3383 static noinline_for_stack
3384 int add_data_references(struct reloc_control *rc,
3385 struct btrfs_key *extent_key,
3386 struct btrfs_path *path,
3387 struct rb_root *blocks)
3389 struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
3390 struct ulist *leaves = NULL;
3391 struct ulist_iterator leaf_uiter;
3392 struct ulist_node *ref_node = NULL;
3393 const u32 blocksize = fs_info->nodesize;
3396 btrfs_release_path(path);
3397 ret = btrfs_find_all_leafs(NULL, fs_info, extent_key->objectid,
3398 0, &leaves, NULL, true);
3402 ULIST_ITER_INIT(&leaf_uiter);
3403 while ((ref_node = ulist_next(leaves, &leaf_uiter))) {
3404 struct extent_buffer *eb;
3406 eb = read_tree_block(fs_info, ref_node->val, 0, 0, 0, NULL);
3411 ret = delete_v1_space_cache(eb, rc->block_group,
3412 extent_key->objectid);
3413 free_extent_buffer(eb);
3416 ret = __add_tree_block(rc, ref_node->val, blocksize, blocks);
3421 free_block_list(blocks);
3427 * helper to find next unprocessed extent
3429 static noinline_for_stack
3430 int find_next_extent(struct reloc_control *rc, struct btrfs_path *path,
3431 struct btrfs_key *extent_key)
3433 struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
3434 struct btrfs_key key;
3435 struct extent_buffer *leaf;
3436 u64 start, end, last;
3439 last = rc->block_group->start + rc->block_group->length;
3442 if (rc->search_start >= last) {
3447 key.objectid = rc->search_start;
3448 key.type = BTRFS_EXTENT_ITEM_KEY;
3451 path->search_commit_root = 1;
3452 path->skip_locking = 1;
3453 ret = btrfs_search_slot(NULL, rc->extent_root, &key, path,
3458 leaf = path->nodes[0];
3459 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
3460 ret = btrfs_next_leaf(rc->extent_root, path);
3463 leaf = path->nodes[0];
3466 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
3467 if (key.objectid >= last) {
3472 if (key.type != BTRFS_EXTENT_ITEM_KEY &&
3473 key.type != BTRFS_METADATA_ITEM_KEY) {
3478 if (key.type == BTRFS_EXTENT_ITEM_KEY &&
3479 key.objectid + key.offset <= rc->search_start) {
3484 if (key.type == BTRFS_METADATA_ITEM_KEY &&
3485 key.objectid + fs_info->nodesize <=
3491 ret = find_first_extent_bit(&rc->processed_blocks,
3492 key.objectid, &start, &end,
3493 EXTENT_DIRTY, NULL);
3495 if (ret == 0 && start <= key.objectid) {
3496 btrfs_release_path(path);
3497 rc->search_start = end + 1;
3499 if (key.type == BTRFS_EXTENT_ITEM_KEY)
3500 rc->search_start = key.objectid + key.offset;
3502 rc->search_start = key.objectid +
3504 memcpy(extent_key, &key, sizeof(key));
3508 btrfs_release_path(path);
3512 static void set_reloc_control(struct reloc_control *rc)
3514 struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
3516 mutex_lock(&fs_info->reloc_mutex);
3517 fs_info->reloc_ctl = rc;
3518 mutex_unlock(&fs_info->reloc_mutex);
3521 static void unset_reloc_control(struct reloc_control *rc)
3523 struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
3525 mutex_lock(&fs_info->reloc_mutex);
3526 fs_info->reloc_ctl = NULL;
3527 mutex_unlock(&fs_info->reloc_mutex);
3530 static noinline_for_stack
3531 int prepare_to_relocate(struct reloc_control *rc)
3533 struct btrfs_trans_handle *trans;
3536 rc->block_rsv = btrfs_alloc_block_rsv(rc->extent_root->fs_info,
3537 BTRFS_BLOCK_RSV_TEMP);
3541 memset(&rc->cluster, 0, sizeof(rc->cluster));
3542 rc->search_start = rc->block_group->start;
3543 rc->extents_found = 0;
3544 rc->nodes_relocated = 0;
3545 rc->merging_rsv_size = 0;
3546 rc->reserved_bytes = 0;
3547 rc->block_rsv->size = rc->extent_root->fs_info->nodesize *
3548 RELOCATION_RESERVED_NODES;
3549 ret = btrfs_block_rsv_refill(rc->extent_root,
3550 rc->block_rsv, rc->block_rsv->size,
3551 BTRFS_RESERVE_FLUSH_ALL);
3555 rc->create_reloc_tree = 1;
3556 set_reloc_control(rc);
3558 trans = btrfs_join_transaction(rc->extent_root);
3559 if (IS_ERR(trans)) {
3560 unset_reloc_control(rc);
3562 * extent tree is not a ref_cow tree and has no reloc_root to
3563 * cleanup. And callers are responsible to free the above
3566 return PTR_ERR(trans);
3568 return btrfs_commit_transaction(trans);
3571 static noinline_for_stack int relocate_block_group(struct reloc_control *rc)
3573 struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
3574 struct rb_root blocks = RB_ROOT;
3575 struct btrfs_key key;
3576 struct btrfs_trans_handle *trans = NULL;
3577 struct btrfs_path *path;
3578 struct btrfs_extent_item *ei;
3584 path = btrfs_alloc_path();
3587 path->reada = READA_FORWARD;
3589 ret = prepare_to_relocate(rc);
3596 rc->reserved_bytes = 0;
3597 ret = btrfs_block_rsv_refill(rc->extent_root,
3598 rc->block_rsv, rc->block_rsv->size,
3599 BTRFS_RESERVE_FLUSH_ALL);
3605 trans = btrfs_start_transaction(rc->extent_root, 0);
3606 if (IS_ERR(trans)) {
3607 err = PTR_ERR(trans);
3612 if (update_backref_cache(trans, &rc->backref_cache)) {
3613 btrfs_end_transaction(trans);
3618 ret = find_next_extent(rc, path, &key);
3624 rc->extents_found++;
3626 ei = btrfs_item_ptr(path->nodes[0], path->slots[0],
3627 struct btrfs_extent_item);
3628 flags = btrfs_extent_flags(path->nodes[0], ei);
3630 if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
3631 ret = add_tree_block(rc, &key, path, &blocks);
3632 } else if (rc->stage == UPDATE_DATA_PTRS &&
3633 (flags & BTRFS_EXTENT_FLAG_DATA)) {
3634 ret = add_data_references(rc, &key, path, &blocks);
3636 btrfs_release_path(path);
3644 if (!RB_EMPTY_ROOT(&blocks)) {
3645 ret = relocate_tree_blocks(trans, rc, &blocks);
3647 if (ret != -EAGAIN) {
3651 rc->extents_found--;
3652 rc->search_start = key.objectid;
3656 btrfs_end_transaction_throttle(trans);
3657 btrfs_btree_balance_dirty(fs_info);
3660 if (rc->stage == MOVE_DATA_EXTENTS &&
3661 (flags & BTRFS_EXTENT_FLAG_DATA)) {
3662 rc->found_file_extent = 1;
3663 ret = relocate_data_extent(rc->data_inode,
3664 &key, &rc->cluster);
3670 if (btrfs_should_cancel_balance(fs_info)) {
3675 if (trans && progress && err == -ENOSPC) {
3676 ret = btrfs_force_chunk_alloc(trans, rc->block_group->flags);
3684 btrfs_release_path(path);
3685 clear_extent_bits(&rc->processed_blocks, 0, (u64)-1, EXTENT_DIRTY);
3688 btrfs_end_transaction_throttle(trans);
3689 btrfs_btree_balance_dirty(fs_info);
3693 ret = relocate_file_extent_cluster(rc->data_inode,
3699 rc->create_reloc_tree = 0;
3700 set_reloc_control(rc);
3702 btrfs_backref_release_cache(&rc->backref_cache);
3703 btrfs_block_rsv_release(fs_info, rc->block_rsv, (u64)-1, NULL);
3706 * Even in the case when the relocation is cancelled, we should all go
3707 * through prepare_to_merge() and merge_reloc_roots().
3709 * For error (including cancelled balance), prepare_to_merge() will
3710 * mark all reloc trees orphan, then queue them for cleanup in
3711 * merge_reloc_roots()
3713 err = prepare_to_merge(rc, err);
3715 merge_reloc_roots(rc);
3717 rc->merge_reloc_tree = 0;
3718 unset_reloc_control(rc);
3719 btrfs_block_rsv_release(fs_info, rc->block_rsv, (u64)-1, NULL);
3721 /* get rid of pinned extents */
3722 trans = btrfs_join_transaction(rc->extent_root);
3723 if (IS_ERR(trans)) {
3724 err = PTR_ERR(trans);
3727 ret = btrfs_commit_transaction(trans);
3731 ret = clean_dirty_subvols(rc);
3732 if (ret < 0 && !err)
3734 btrfs_free_block_rsv(fs_info, rc->block_rsv);
3735 btrfs_free_path(path);
3739 static int __insert_orphan_inode(struct btrfs_trans_handle *trans,
3740 struct btrfs_root *root, u64 objectid)
3742 struct btrfs_path *path;
3743 struct btrfs_inode_item *item;
3744 struct extent_buffer *leaf;
3747 path = btrfs_alloc_path();
3751 ret = btrfs_insert_empty_inode(trans, root, path, objectid);
3755 leaf = path->nodes[0];
3756 item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_inode_item);
3757 memzero_extent_buffer(leaf, (unsigned long)item, sizeof(*item));
3758 btrfs_set_inode_generation(leaf, item, 1);
3759 btrfs_set_inode_size(leaf, item, 0);
3760 btrfs_set_inode_mode(leaf, item, S_IFREG | 0600);
3761 btrfs_set_inode_flags(leaf, item, BTRFS_INODE_NOCOMPRESS |
3762 BTRFS_INODE_PREALLOC);
3763 btrfs_mark_buffer_dirty(leaf);
3765 btrfs_free_path(path);
3769 static void delete_orphan_inode(struct btrfs_trans_handle *trans,
3770 struct btrfs_root *root, u64 objectid)
3772 struct btrfs_path *path;
3773 struct btrfs_key key;
3776 path = btrfs_alloc_path();
3782 key.objectid = objectid;
3783 key.type = BTRFS_INODE_ITEM_KEY;
3785 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
3791 ret = btrfs_del_item(trans, root, path);
3794 btrfs_abort_transaction(trans, ret);
3795 btrfs_free_path(path);
3799 * helper to create inode for data relocation.
3800 * the inode is in data relocation tree and its link count is 0
3802 static noinline_for_stack
3803 struct inode *create_reloc_inode(struct btrfs_fs_info *fs_info,
3804 struct btrfs_block_group *group)
3806 struct inode *inode = NULL;
3807 struct btrfs_trans_handle *trans;
3808 struct btrfs_root *root;
3812 root = btrfs_grab_root(fs_info->data_reloc_root);
3813 trans = btrfs_start_transaction(root, 6);
3814 if (IS_ERR(trans)) {
3815 btrfs_put_root(root);
3816 return ERR_CAST(trans);
3819 err = btrfs_get_free_objectid(root, &objectid);
3823 err = __insert_orphan_inode(trans, root, objectid);
3827 inode = btrfs_iget(fs_info->sb, objectid, root);
3828 if (IS_ERR(inode)) {
3829 delete_orphan_inode(trans, root, objectid);
3830 err = PTR_ERR(inode);
3834 BTRFS_I(inode)->index_cnt = group->start;
3836 err = btrfs_orphan_add(trans, BTRFS_I(inode));
3838 btrfs_put_root(root);
3839 btrfs_end_transaction(trans);
3840 btrfs_btree_balance_dirty(fs_info);
3844 inode = ERR_PTR(err);
3850 * Mark start of chunk relocation that is cancellable. Check if the cancellation
3851 * has been requested meanwhile and don't start in that case.
3855 * -EINPROGRESS operation is already in progress, that's probably a bug
3856 * -ECANCELED cancellation request was set before the operation started
3857 * -EAGAIN can not start because there are ongoing send operations
3859 static int reloc_chunk_start(struct btrfs_fs_info *fs_info)
3861 spin_lock(&fs_info->send_reloc_lock);
3862 if (fs_info->send_in_progress) {
3863 btrfs_warn_rl(fs_info,
3864 "cannot run relocation while send operations are in progress (%d in progress)",
3865 fs_info->send_in_progress);
3866 spin_unlock(&fs_info->send_reloc_lock);
3869 if (test_and_set_bit(BTRFS_FS_RELOC_RUNNING, &fs_info->flags)) {
3870 /* This should not happen */
3871 spin_unlock(&fs_info->send_reloc_lock);
3872 btrfs_err(fs_info, "reloc already running, cannot start");
3873 return -EINPROGRESS;
3875 spin_unlock(&fs_info->send_reloc_lock);
3877 if (atomic_read(&fs_info->reloc_cancel_req) > 0) {
3878 btrfs_info(fs_info, "chunk relocation canceled on start");
3880 * On cancel, clear all requests but let the caller mark
3881 * the end after cleanup operations.
3883 atomic_set(&fs_info->reloc_cancel_req, 0);
3890 * Mark end of chunk relocation that is cancellable and wake any waiters.
3892 static void reloc_chunk_end(struct btrfs_fs_info *fs_info)
3894 /* Requested after start, clear bit first so any waiters can continue */
3895 if (atomic_read(&fs_info->reloc_cancel_req) > 0)
3896 btrfs_info(fs_info, "chunk relocation canceled during operation");
3897 spin_lock(&fs_info->send_reloc_lock);
3898 clear_and_wake_up_bit(BTRFS_FS_RELOC_RUNNING, &fs_info->flags);
3899 spin_unlock(&fs_info->send_reloc_lock);
3900 atomic_set(&fs_info->reloc_cancel_req, 0);
3903 static struct reloc_control *alloc_reloc_control(struct btrfs_fs_info *fs_info)
3905 struct reloc_control *rc;
3907 rc = kzalloc(sizeof(*rc), GFP_NOFS);
3911 INIT_LIST_HEAD(&rc->reloc_roots);
3912 INIT_LIST_HEAD(&rc->dirty_subvol_roots);
3913 btrfs_backref_init_cache(fs_info, &rc->backref_cache, 1);
3914 mapping_tree_init(&rc->reloc_root_tree);
3915 extent_io_tree_init(fs_info, &rc->processed_blocks,
3916 IO_TREE_RELOC_BLOCKS, NULL);
3920 static void free_reloc_control(struct reloc_control *rc)
3922 struct mapping_node *node, *tmp;
3924 free_reloc_roots(&rc->reloc_roots);
3925 rbtree_postorder_for_each_entry_safe(node, tmp,
3926 &rc->reloc_root_tree.rb_root, rb_node)
3933 * Print the block group being relocated
3935 static void describe_relocation(struct btrfs_fs_info *fs_info,
3936 struct btrfs_block_group *block_group)
3938 char buf[128] = {'\0'};
3940 btrfs_describe_block_groups(block_group->flags, buf, sizeof(buf));
3943 "relocating block group %llu flags %s",
3944 block_group->start, buf);
3947 static const char *stage_to_string(int stage)
3949 if (stage == MOVE_DATA_EXTENTS)
3950 return "move data extents";
3951 if (stage == UPDATE_DATA_PTRS)
3952 return "update data pointers";
3957 * function to relocate all extents in a block group.
3959 int btrfs_relocate_block_group(struct btrfs_fs_info *fs_info, u64 group_start)
3961 struct btrfs_block_group *bg;
3962 struct btrfs_root *extent_root = fs_info->extent_root;
3963 struct reloc_control *rc;
3964 struct inode *inode;
3965 struct btrfs_path *path;
3970 bg = btrfs_lookup_block_group(fs_info, group_start);
3974 if (btrfs_pinned_by_swapfile(fs_info, bg)) {
3975 btrfs_put_block_group(bg);
3979 rc = alloc_reloc_control(fs_info);
3981 btrfs_put_block_group(bg);
3985 ret = reloc_chunk_start(fs_info);
3991 rc->extent_root = extent_root;
3992 rc->block_group = bg;
3994 ret = btrfs_inc_block_group_ro(rc->block_group, true);
4001 path = btrfs_alloc_path();
4007 inode = lookup_free_space_inode(rc->block_group, path);
4008 btrfs_free_path(path);
4011 ret = delete_block_group_cache(fs_info, rc->block_group, inode, 0);
4013 ret = PTR_ERR(inode);
4015 if (ret && ret != -ENOENT) {
4020 rc->data_inode = create_reloc_inode(fs_info, rc->block_group);
4021 if (IS_ERR(rc->data_inode)) {
4022 err = PTR_ERR(rc->data_inode);
4023 rc->data_inode = NULL;
4027 describe_relocation(fs_info, rc->block_group);
4029 btrfs_wait_block_group_reservations(rc->block_group);
4030 btrfs_wait_nocow_writers(rc->block_group);
4031 btrfs_wait_ordered_roots(fs_info, U64_MAX,
4032 rc->block_group->start,
4033 rc->block_group->length);
4038 mutex_lock(&fs_info->cleaner_mutex);
4039 ret = relocate_block_group(rc);
4040 mutex_unlock(&fs_info->cleaner_mutex);
4044 finishes_stage = rc->stage;
4046 * We may have gotten ENOSPC after we already dirtied some
4047 * extents. If writeout happens while we're relocating a
4048 * different block group we could end up hitting the
4049 * BUG_ON(rc->stage == UPDATE_DATA_PTRS) in
4050 * btrfs_reloc_cow_block. Make sure we write everything out
4051 * properly so we don't trip over this problem, and then break
4052 * out of the loop if we hit an error.
4054 if (rc->stage == MOVE_DATA_EXTENTS && rc->found_file_extent) {
4055 ret = btrfs_wait_ordered_range(rc->data_inode, 0,
4059 invalidate_mapping_pages(rc->data_inode->i_mapping,
4061 rc->stage = UPDATE_DATA_PTRS;
4067 if (rc->extents_found == 0)
4070 btrfs_info(fs_info, "found %llu extents, stage: %s",
4071 rc->extents_found, stage_to_string(finishes_stage));
4074 WARN_ON(rc->block_group->pinned > 0);
4075 WARN_ON(rc->block_group->reserved > 0);
4076 WARN_ON(rc->block_group->used > 0);
4079 btrfs_dec_block_group_ro(rc->block_group);
4080 iput(rc->data_inode);
4082 btrfs_put_block_group(bg);
4083 reloc_chunk_end(fs_info);
4084 free_reloc_control(rc);
4088 static noinline_for_stack int mark_garbage_root(struct btrfs_root *root)
4090 struct btrfs_fs_info *fs_info = root->fs_info;
4091 struct btrfs_trans_handle *trans;
4094 trans = btrfs_start_transaction(fs_info->tree_root, 0);
4096 return PTR_ERR(trans);
4098 memset(&root->root_item.drop_progress, 0,
4099 sizeof(root->root_item.drop_progress));
4100 btrfs_set_root_drop_level(&root->root_item, 0);
4101 btrfs_set_root_refs(&root->root_item, 0);
4102 ret = btrfs_update_root(trans, fs_info->tree_root,
4103 &root->root_key, &root->root_item);
4105 err = btrfs_end_transaction(trans);
4112 * recover relocation interrupted by system crash.
4114 * this function resumes merging reloc trees with corresponding fs trees.
4115 * this is important for keeping the sharing of tree blocks
4117 int btrfs_recover_relocation(struct btrfs_root *root)
4119 struct btrfs_fs_info *fs_info = root->fs_info;
4120 LIST_HEAD(reloc_roots);
4121 struct btrfs_key key;
4122 struct btrfs_root *fs_root;
4123 struct btrfs_root *reloc_root;
4124 struct btrfs_path *path;
4125 struct extent_buffer *leaf;
4126 struct reloc_control *rc = NULL;
4127 struct btrfs_trans_handle *trans;
4131 path = btrfs_alloc_path();
4134 path->reada = READA_BACK;
4136 key.objectid = BTRFS_TREE_RELOC_OBJECTID;
4137 key.type = BTRFS_ROOT_ITEM_KEY;
4138 key.offset = (u64)-1;
4141 ret = btrfs_search_slot(NULL, fs_info->tree_root, &key,
4148 if (path->slots[0] == 0)
4152 leaf = path->nodes[0];
4153 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
4154 btrfs_release_path(path);
4156 if (key.objectid != BTRFS_TREE_RELOC_OBJECTID ||
4157 key.type != BTRFS_ROOT_ITEM_KEY)
4160 reloc_root = btrfs_read_tree_root(root, &key);
4161 if (IS_ERR(reloc_root)) {
4162 err = PTR_ERR(reloc_root);
4166 set_bit(BTRFS_ROOT_SHAREABLE, &reloc_root->state);
4167 list_add(&reloc_root->root_list, &reloc_roots);
4169 if (btrfs_root_refs(&reloc_root->root_item) > 0) {
4170 fs_root = btrfs_get_fs_root(fs_info,
4171 reloc_root->root_key.offset, false);
4172 if (IS_ERR(fs_root)) {
4173 ret = PTR_ERR(fs_root);
4174 if (ret != -ENOENT) {
4178 ret = mark_garbage_root(reloc_root);
4184 btrfs_put_root(fs_root);
4188 if (key.offset == 0)
4193 btrfs_release_path(path);
4195 if (list_empty(&reloc_roots))
4198 rc = alloc_reloc_control(fs_info);
4204 ret = reloc_chunk_start(fs_info);
4210 rc->extent_root = fs_info->extent_root;
4212 set_reloc_control(rc);
4214 trans = btrfs_join_transaction(rc->extent_root);
4215 if (IS_ERR(trans)) {
4216 err = PTR_ERR(trans);
4220 rc->merge_reloc_tree = 1;
4222 while (!list_empty(&reloc_roots)) {
4223 reloc_root = list_entry(reloc_roots.next,
4224 struct btrfs_root, root_list);
4225 list_del(&reloc_root->root_list);
4227 if (btrfs_root_refs(&reloc_root->root_item) == 0) {
4228 list_add_tail(&reloc_root->root_list,
4233 fs_root = btrfs_get_fs_root(fs_info, reloc_root->root_key.offset,
4235 if (IS_ERR(fs_root)) {
4236 err = PTR_ERR(fs_root);
4237 list_add_tail(&reloc_root->root_list, &reloc_roots);
4238 btrfs_end_transaction(trans);
4242 err = __add_reloc_root(reloc_root);
4243 ASSERT(err != -EEXIST);
4245 list_add_tail(&reloc_root->root_list, &reloc_roots);
4246 btrfs_put_root(fs_root);
4247 btrfs_end_transaction(trans);
4250 fs_root->reloc_root = btrfs_grab_root(reloc_root);
4251 btrfs_put_root(fs_root);
4254 err = btrfs_commit_transaction(trans);
4258 merge_reloc_roots(rc);
4260 unset_reloc_control(rc);
4262 trans = btrfs_join_transaction(rc->extent_root);
4263 if (IS_ERR(trans)) {
4264 err = PTR_ERR(trans);
4267 err = btrfs_commit_transaction(trans);
4269 ret = clean_dirty_subvols(rc);
4270 if (ret < 0 && !err)
4273 unset_reloc_control(rc);
4275 reloc_chunk_end(fs_info);
4276 free_reloc_control(rc);
4278 free_reloc_roots(&reloc_roots);
4280 btrfs_free_path(path);
4283 /* cleanup orphan inode in data relocation tree */
4284 fs_root = btrfs_grab_root(fs_info->data_reloc_root);
4286 err = btrfs_orphan_cleanup(fs_root);
4287 btrfs_put_root(fs_root);
4293 * helper to add ordered checksum for data relocation.
4295 * cloning checksum properly handles the nodatasum extents.
4296 * it also saves CPU time to re-calculate the checksum.
4298 int btrfs_reloc_clone_csums(struct btrfs_inode *inode, u64 file_pos, u64 len)
4300 struct btrfs_fs_info *fs_info = inode->root->fs_info;
4301 struct btrfs_ordered_sum *sums;
4302 struct btrfs_ordered_extent *ordered;
4308 ordered = btrfs_lookup_ordered_extent(inode, file_pos);
4309 BUG_ON(ordered->file_offset != file_pos || ordered->num_bytes != len);
4311 disk_bytenr = file_pos + inode->index_cnt;
4312 ret = btrfs_lookup_csums_range(fs_info->csum_root, disk_bytenr,
4313 disk_bytenr + len - 1, &list, 0);
4317 while (!list_empty(&list)) {
4318 sums = list_entry(list.next, struct btrfs_ordered_sum, list);
4319 list_del_init(&sums->list);
4322 * We need to offset the new_bytenr based on where the csum is.
4323 * We need to do this because we will read in entire prealloc
4324 * extents but we may have written to say the middle of the
4325 * prealloc extent, so we need to make sure the csum goes with
4326 * the right disk offset.
4328 * We can do this because the data reloc inode refers strictly
4329 * to the on disk bytes, so we don't have to worry about
4330 * disk_len vs real len like with real inodes since it's all
4333 new_bytenr = ordered->disk_bytenr + sums->bytenr - disk_bytenr;
4334 sums->bytenr = new_bytenr;
4336 btrfs_add_ordered_sum(ordered, sums);
4339 btrfs_put_ordered_extent(ordered);
4343 int btrfs_reloc_cow_block(struct btrfs_trans_handle *trans,
4344 struct btrfs_root *root, struct extent_buffer *buf,
4345 struct extent_buffer *cow)
4347 struct btrfs_fs_info *fs_info = root->fs_info;
4348 struct reloc_control *rc;
4349 struct btrfs_backref_node *node;
4354 rc = fs_info->reloc_ctl;
4358 BUG_ON(rc->stage == UPDATE_DATA_PTRS && btrfs_is_data_reloc_root(root));
4360 level = btrfs_header_level(buf);
4361 if (btrfs_header_generation(buf) <=
4362 btrfs_root_last_snapshot(&root->root_item))
4365 if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID &&
4366 rc->create_reloc_tree) {
4367 WARN_ON(!first_cow && level == 0);
4369 node = rc->backref_cache.path[level];
4370 BUG_ON(node->bytenr != buf->start &&
4371 node->new_bytenr != buf->start);
4373 btrfs_backref_drop_node_buffer(node);
4374 atomic_inc(&cow->refs);
4376 node->new_bytenr = cow->start;
4378 if (!node->pending) {
4379 list_move_tail(&node->list,
4380 &rc->backref_cache.pending[level]);
4385 mark_block_processed(rc, node);
4387 if (first_cow && level > 0)
4388 rc->nodes_relocated += buf->len;
4391 if (level == 0 && first_cow && rc->stage == UPDATE_DATA_PTRS)
4392 ret = replace_file_extents(trans, rc, root, cow);
4397 * called before creating snapshot. it calculates metadata reservation
4398 * required for relocating tree blocks in the snapshot
4400 void btrfs_reloc_pre_snapshot(struct btrfs_pending_snapshot *pending,
4401 u64 *bytes_to_reserve)
4403 struct btrfs_root *root = pending->root;
4404 struct reloc_control *rc = root->fs_info->reloc_ctl;
4406 if (!rc || !have_reloc_root(root))
4409 if (!rc->merge_reloc_tree)
4412 root = root->reloc_root;
4413 BUG_ON(btrfs_root_refs(&root->root_item) == 0);
4415 * relocation is in the stage of merging trees. the space
4416 * used by merging a reloc tree is twice the size of
4417 * relocated tree nodes in the worst case. half for cowing
4418 * the reloc tree, half for cowing the fs tree. the space
4419 * used by cowing the reloc tree will be freed after the
4420 * tree is dropped. if we create snapshot, cowing the fs
4421 * tree may use more space than it frees. so we need
4422 * reserve extra space.
4424 *bytes_to_reserve += rc->nodes_relocated;
4428 * called after snapshot is created. migrate block reservation
4429 * and create reloc root for the newly created snapshot
4431 * This is similar to btrfs_init_reloc_root(), we come out of here with two
4432 * references held on the reloc_root, one for root->reloc_root and one for
4435 int btrfs_reloc_post_snapshot(struct btrfs_trans_handle *trans,
4436 struct btrfs_pending_snapshot *pending)
4438 struct btrfs_root *root = pending->root;
4439 struct btrfs_root *reloc_root;
4440 struct btrfs_root *new_root;
4441 struct reloc_control *rc = root->fs_info->reloc_ctl;
4444 if (!rc || !have_reloc_root(root))
4447 rc = root->fs_info->reloc_ctl;
4448 rc->merging_rsv_size += rc->nodes_relocated;
4450 if (rc->merge_reloc_tree) {
4451 ret = btrfs_block_rsv_migrate(&pending->block_rsv,
4453 rc->nodes_relocated, true);
4458 new_root = pending->snap;
4459 reloc_root = create_reloc_root(trans, root->reloc_root,
4460 new_root->root_key.objectid);
4461 if (IS_ERR(reloc_root))
4462 return PTR_ERR(reloc_root);
4464 ret = __add_reloc_root(reloc_root);
4465 ASSERT(ret != -EEXIST);
4467 /* Pairs with create_reloc_root */
4468 btrfs_put_root(reloc_root);
4471 new_root->reloc_root = btrfs_grab_root(reloc_root);
4473 if (rc->create_reloc_tree)
4474 ret = clone_backref_node(trans, rc, root, reloc_root);