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"
29 #include "inode-item.h"
30 #include "space-info.h"
32 #include "accessors.h"
33 #include "extent-tree.h"
34 #include "root-tree.h"
35 #include "file-item.h"
36 #include "relocation.h"
38 #include "tree-checker.h"
43 * [What does relocation do]
45 * The objective of relocation is to relocate all extents of the target block
46 * group to other block groups.
47 * This is utilized by resize (shrink only), profile converting, compacting
48 * space, or balance routine to spread chunks over devices.
51 * ------------------------------------------------------------------
52 * BG A: 10 data extents | BG A: deleted
53 * BG B: 2 data extents | BG B: 10 data extents (2 old + 8 relocated)
54 * BG C: 1 extents | BG C: 3 data extents (1 old + 2 relocated)
56 * [How does relocation work]
58 * 1. Mark the target block group read-only
59 * New extents won't be allocated from the target block group.
61 * 2.1 Record each extent in the target block group
62 * To build a proper map of extents to be relocated.
64 * 2.2 Build data reloc tree and reloc trees
65 * Data reloc tree will contain an inode, recording all newly relocated
67 * There will be only one data reloc tree for one data block group.
69 * Reloc tree will be a special snapshot of its source tree, containing
70 * relocated tree blocks.
71 * Each tree referring to a tree block in target block group will get its
74 * 2.3 Swap source tree with its corresponding reloc tree
75 * Each involved tree only refers to new extents after swap.
77 * 3. Cleanup reloc trees and data reloc tree.
78 * As old extents in the target block group are still referenced by reloc
79 * trees, we need to clean them up before really freeing the target block
82 * The main complexity is in steps 2.2 and 2.3.
84 * The entry point of relocation is relocate_block_group() function.
87 #define RELOCATION_RESERVED_NODES 256
89 * map address of tree root to tree
93 struct rb_node rb_node;
95 }; /* Use rb_simle_node for search/insert */
100 struct rb_root rb_root;
105 * present a tree block to process
109 struct rb_node rb_node;
111 }; /* Use rb_simple_node for search/insert */
113 struct btrfs_key key;
114 unsigned int level:8;
115 unsigned int key_ready:1;
118 #define MAX_EXTENTS 128
120 struct file_extent_cluster {
123 u64 boundary[MAX_EXTENTS];
127 struct reloc_control {
128 /* block group to relocate */
129 struct btrfs_block_group *block_group;
131 struct btrfs_root *extent_root;
132 /* inode for moving data */
133 struct inode *data_inode;
135 struct btrfs_block_rsv *block_rsv;
137 struct btrfs_backref_cache backref_cache;
139 struct file_extent_cluster cluster;
140 /* tree blocks have been processed */
141 struct extent_io_tree processed_blocks;
142 /* map start of tree root to corresponding reloc tree */
143 struct mapping_tree reloc_root_tree;
144 /* list of reloc trees */
145 struct list_head reloc_roots;
146 /* list of subvolume trees that get relocated */
147 struct list_head dirty_subvol_roots;
148 /* size of metadata reservation for merging reloc trees */
149 u64 merging_rsv_size;
150 /* size of relocated tree nodes */
152 /* reserved size for block group relocation*/
158 unsigned int stage:8;
159 unsigned int create_reloc_tree:1;
160 unsigned int merge_reloc_tree:1;
161 unsigned int found_file_extent:1;
164 /* stages of data relocation */
165 #define MOVE_DATA_EXTENTS 0
166 #define UPDATE_DATA_PTRS 1
168 static void mark_block_processed(struct reloc_control *rc,
169 struct btrfs_backref_node *node)
173 if (node->level == 0 ||
174 in_range(node->bytenr, rc->block_group->start,
175 rc->block_group->length)) {
176 blocksize = rc->extent_root->fs_info->nodesize;
177 set_extent_bits(&rc->processed_blocks, node->bytenr,
178 node->bytenr + blocksize - 1, EXTENT_DIRTY);
184 static void mapping_tree_init(struct mapping_tree *tree)
186 tree->rb_root = RB_ROOT;
187 spin_lock_init(&tree->lock);
191 * walk up backref nodes until reach node presents tree root
193 static struct btrfs_backref_node *walk_up_backref(
194 struct btrfs_backref_node *node,
195 struct btrfs_backref_edge *edges[], int *index)
197 struct btrfs_backref_edge *edge;
200 while (!list_empty(&node->upper)) {
201 edge = list_entry(node->upper.next,
202 struct btrfs_backref_edge, list[LOWER]);
204 node = edge->node[UPPER];
206 BUG_ON(node->detached);
212 * walk down backref nodes to find start of next reference path
214 static struct btrfs_backref_node *walk_down_backref(
215 struct btrfs_backref_edge *edges[], int *index)
217 struct btrfs_backref_edge *edge;
218 struct btrfs_backref_node *lower;
222 edge = edges[idx - 1];
223 lower = edge->node[LOWER];
224 if (list_is_last(&edge->list[LOWER], &lower->upper)) {
228 edge = list_entry(edge->list[LOWER].next,
229 struct btrfs_backref_edge, list[LOWER]);
230 edges[idx - 1] = edge;
232 return edge->node[UPPER];
238 static void update_backref_node(struct btrfs_backref_cache *cache,
239 struct btrfs_backref_node *node, u64 bytenr)
241 struct rb_node *rb_node;
242 rb_erase(&node->rb_node, &cache->rb_root);
243 node->bytenr = bytenr;
244 rb_node = rb_simple_insert(&cache->rb_root, node->bytenr, &node->rb_node);
246 btrfs_backref_panic(cache->fs_info, bytenr, -EEXIST);
250 * update backref cache after a transaction commit
252 static int update_backref_cache(struct btrfs_trans_handle *trans,
253 struct btrfs_backref_cache *cache)
255 struct btrfs_backref_node *node;
258 if (cache->last_trans == 0) {
259 cache->last_trans = trans->transid;
263 if (cache->last_trans == trans->transid)
267 * detached nodes are used to avoid unnecessary backref
268 * lookup. transaction commit changes the extent tree.
269 * so the detached nodes are no longer useful.
271 while (!list_empty(&cache->detached)) {
272 node = list_entry(cache->detached.next,
273 struct btrfs_backref_node, list);
274 btrfs_backref_cleanup_node(cache, node);
277 while (!list_empty(&cache->changed)) {
278 node = list_entry(cache->changed.next,
279 struct btrfs_backref_node, list);
280 list_del_init(&node->list);
281 BUG_ON(node->pending);
282 update_backref_node(cache, node, node->new_bytenr);
286 * some nodes can be left in the pending list if there were
287 * errors during processing the pending nodes.
289 for (level = 0; level < BTRFS_MAX_LEVEL; level++) {
290 list_for_each_entry(node, &cache->pending[level], list) {
291 BUG_ON(!node->pending);
292 if (node->bytenr == node->new_bytenr)
294 update_backref_node(cache, node, node->new_bytenr);
298 cache->last_trans = 0;
302 static bool reloc_root_is_dead(struct btrfs_root *root)
305 * Pair with set_bit/clear_bit in clean_dirty_subvols and
306 * btrfs_update_reloc_root. We need to see the updated bit before
307 * trying to access reloc_root
310 if (test_bit(BTRFS_ROOT_DEAD_RELOC_TREE, &root->state))
316 * Check if this subvolume tree has valid reloc tree.
318 * Reloc tree after swap is considered dead, thus not considered as valid.
319 * This is enough for most callers, as they don't distinguish dead reloc root
320 * from no reloc root. But btrfs_should_ignore_reloc_root() below is a
323 static bool have_reloc_root(struct btrfs_root *root)
325 if (reloc_root_is_dead(root))
327 if (!root->reloc_root)
332 int btrfs_should_ignore_reloc_root(struct btrfs_root *root)
334 struct btrfs_root *reloc_root;
336 if (!test_bit(BTRFS_ROOT_SHAREABLE, &root->state))
339 /* This root has been merged with its reloc tree, we can ignore it */
340 if (reloc_root_is_dead(root))
343 reloc_root = root->reloc_root;
347 if (btrfs_header_generation(reloc_root->commit_root) ==
348 root->fs_info->running_transaction->transid)
351 * if there is reloc tree and it was created in previous
352 * transaction backref lookup can find the reloc tree,
353 * so backref node for the fs tree root is useless for
360 * find reloc tree by address of tree root
362 struct btrfs_root *find_reloc_root(struct btrfs_fs_info *fs_info, u64 bytenr)
364 struct reloc_control *rc = fs_info->reloc_ctl;
365 struct rb_node *rb_node;
366 struct mapping_node *node;
367 struct btrfs_root *root = NULL;
370 spin_lock(&rc->reloc_root_tree.lock);
371 rb_node = rb_simple_search(&rc->reloc_root_tree.rb_root, bytenr);
373 node = rb_entry(rb_node, struct mapping_node, rb_node);
376 spin_unlock(&rc->reloc_root_tree.lock);
377 return btrfs_grab_root(root);
381 * For useless nodes, do two major clean ups:
383 * - Cleanup the children edges and nodes
384 * If child node is also orphan (no parent) during cleanup, then the child
385 * node will also be cleaned up.
387 * - Freeing up leaves (level 0), keeps nodes detached
388 * For nodes, the node is still cached as "detached"
390 * Return false if @node is not in the @useless_nodes list.
391 * Return true if @node is in the @useless_nodes list.
393 static bool handle_useless_nodes(struct reloc_control *rc,
394 struct btrfs_backref_node *node)
396 struct btrfs_backref_cache *cache = &rc->backref_cache;
397 struct list_head *useless_node = &cache->useless_node;
400 while (!list_empty(useless_node)) {
401 struct btrfs_backref_node *cur;
403 cur = list_first_entry(useless_node, struct btrfs_backref_node,
405 list_del_init(&cur->list);
407 /* Only tree root nodes can be added to @useless_nodes */
408 ASSERT(list_empty(&cur->upper));
413 /* The node is the lowest node */
415 list_del_init(&cur->lower);
419 /* Cleanup the lower edges */
420 while (!list_empty(&cur->lower)) {
421 struct btrfs_backref_edge *edge;
422 struct btrfs_backref_node *lower;
424 edge = list_entry(cur->lower.next,
425 struct btrfs_backref_edge, list[UPPER]);
426 list_del(&edge->list[UPPER]);
427 list_del(&edge->list[LOWER]);
428 lower = edge->node[LOWER];
429 btrfs_backref_free_edge(cache, edge);
431 /* Child node is also orphan, queue for cleanup */
432 if (list_empty(&lower->upper))
433 list_add(&lower->list, useless_node);
435 /* Mark this block processed for relocation */
436 mark_block_processed(rc, cur);
439 * Backref nodes for tree leaves are deleted from the cache.
440 * Backref nodes for upper level tree blocks are left in the
441 * cache to avoid unnecessary backref lookup.
443 if (cur->level > 0) {
444 list_add(&cur->list, &cache->detached);
447 rb_erase(&cur->rb_node, &cache->rb_root);
448 btrfs_backref_free_node(cache, cur);
455 * Build backref tree for a given tree block. Root of the backref tree
456 * corresponds the tree block, leaves of the backref tree correspond roots of
457 * b-trees that reference the tree block.
459 * The basic idea of this function is check backrefs of a given block to find
460 * upper level blocks that reference the block, and then check backrefs of
461 * these upper level blocks recursively. The recursion stops when tree root is
462 * reached or backrefs for the block is cached.
464 * NOTE: if we find that backrefs for a block are cached, we know backrefs for
465 * all upper level blocks that directly/indirectly reference the block are also
468 static noinline_for_stack struct btrfs_backref_node *build_backref_tree(
469 struct reloc_control *rc, struct btrfs_key *node_key,
470 int level, u64 bytenr)
472 struct btrfs_backref_iter *iter;
473 struct btrfs_backref_cache *cache = &rc->backref_cache;
474 /* For searching parent of TREE_BLOCK_REF */
475 struct btrfs_path *path;
476 struct btrfs_backref_node *cur;
477 struct btrfs_backref_node *node = NULL;
478 struct btrfs_backref_edge *edge;
482 iter = btrfs_backref_iter_alloc(rc->extent_root->fs_info);
484 return ERR_PTR(-ENOMEM);
485 path = btrfs_alloc_path();
491 node = btrfs_backref_alloc_node(cache, bytenr, level);
500 /* Breadth-first search to build backref cache */
502 ret = btrfs_backref_add_tree_node(cache, path, iter, node_key,
508 edge = list_first_entry_or_null(&cache->pending_edge,
509 struct btrfs_backref_edge, list[UPPER]);
511 * The pending list isn't empty, take the first block to
515 list_del_init(&edge->list[UPPER]);
516 cur = edge->node[UPPER];
520 /* Finish the upper linkage of newly added edges/nodes */
521 ret = btrfs_backref_finish_upper_links(cache, node);
527 if (handle_useless_nodes(rc, node))
530 btrfs_backref_iter_free(iter);
531 btrfs_free_path(path);
533 btrfs_backref_error_cleanup(cache, node);
536 ASSERT(!node || !node->detached);
537 ASSERT(list_empty(&cache->useless_node) &&
538 list_empty(&cache->pending_edge));
543 * helper to add backref node for the newly created snapshot.
544 * the backref node is created by cloning backref node that
545 * corresponds to root of source tree
547 static int clone_backref_node(struct btrfs_trans_handle *trans,
548 struct reloc_control *rc,
549 struct btrfs_root *src,
550 struct btrfs_root *dest)
552 struct btrfs_root *reloc_root = src->reloc_root;
553 struct btrfs_backref_cache *cache = &rc->backref_cache;
554 struct btrfs_backref_node *node = NULL;
555 struct btrfs_backref_node *new_node;
556 struct btrfs_backref_edge *edge;
557 struct btrfs_backref_edge *new_edge;
558 struct rb_node *rb_node;
560 if (cache->last_trans > 0)
561 update_backref_cache(trans, cache);
563 rb_node = rb_simple_search(&cache->rb_root, src->commit_root->start);
565 node = rb_entry(rb_node, struct btrfs_backref_node, rb_node);
569 BUG_ON(node->new_bytenr != reloc_root->node->start);
573 rb_node = rb_simple_search(&cache->rb_root,
574 reloc_root->commit_root->start);
576 node = rb_entry(rb_node, struct btrfs_backref_node,
578 BUG_ON(node->detached);
585 new_node = btrfs_backref_alloc_node(cache, dest->node->start,
590 new_node->lowest = node->lowest;
591 new_node->checked = 1;
592 new_node->root = btrfs_grab_root(dest);
593 ASSERT(new_node->root);
596 list_for_each_entry(edge, &node->lower, list[UPPER]) {
597 new_edge = btrfs_backref_alloc_edge(cache);
601 btrfs_backref_link_edge(new_edge, edge->node[LOWER],
602 new_node, LINK_UPPER);
605 list_add_tail(&new_node->lower, &cache->leaves);
608 rb_node = rb_simple_insert(&cache->rb_root, new_node->bytenr,
611 btrfs_backref_panic(trans->fs_info, new_node->bytenr, -EEXIST);
613 if (!new_node->lowest) {
614 list_for_each_entry(new_edge, &new_node->lower, list[UPPER]) {
615 list_add_tail(&new_edge->list[LOWER],
616 &new_edge->node[LOWER]->upper);
621 while (!list_empty(&new_node->lower)) {
622 new_edge = list_entry(new_node->lower.next,
623 struct btrfs_backref_edge, list[UPPER]);
624 list_del(&new_edge->list[UPPER]);
625 btrfs_backref_free_edge(cache, new_edge);
627 btrfs_backref_free_node(cache, new_node);
632 * helper to add 'address of tree root -> reloc tree' mapping
634 static int __must_check __add_reloc_root(struct btrfs_root *root)
636 struct btrfs_fs_info *fs_info = root->fs_info;
637 struct rb_node *rb_node;
638 struct mapping_node *node;
639 struct reloc_control *rc = fs_info->reloc_ctl;
641 node = kmalloc(sizeof(*node), GFP_NOFS);
645 node->bytenr = root->commit_root->start;
648 spin_lock(&rc->reloc_root_tree.lock);
649 rb_node = rb_simple_insert(&rc->reloc_root_tree.rb_root,
650 node->bytenr, &node->rb_node);
651 spin_unlock(&rc->reloc_root_tree.lock);
654 "Duplicate root found for start=%llu while inserting into relocation tree",
659 list_add_tail(&root->root_list, &rc->reloc_roots);
664 * helper to delete the 'address of tree root -> reloc tree'
667 static void __del_reloc_root(struct btrfs_root *root)
669 struct btrfs_fs_info *fs_info = root->fs_info;
670 struct rb_node *rb_node;
671 struct mapping_node *node = NULL;
672 struct reloc_control *rc = fs_info->reloc_ctl;
673 bool put_ref = false;
675 if (rc && root->node) {
676 spin_lock(&rc->reloc_root_tree.lock);
677 rb_node = rb_simple_search(&rc->reloc_root_tree.rb_root,
678 root->commit_root->start);
680 node = rb_entry(rb_node, struct mapping_node, rb_node);
681 rb_erase(&node->rb_node, &rc->reloc_root_tree.rb_root);
682 RB_CLEAR_NODE(&node->rb_node);
684 spin_unlock(&rc->reloc_root_tree.lock);
685 ASSERT(!node || (struct btrfs_root *)node->data == root);
689 * We only put the reloc root here if it's on the list. There's a lot
690 * of places where the pattern is to splice the rc->reloc_roots, process
691 * the reloc roots, and then add the reloc root back onto
692 * rc->reloc_roots. If we call __del_reloc_root while it's off of the
693 * list we don't want the reference being dropped, because the guy
694 * messing with the list is in charge of the reference.
696 spin_lock(&fs_info->trans_lock);
697 if (!list_empty(&root->root_list)) {
699 list_del_init(&root->root_list);
701 spin_unlock(&fs_info->trans_lock);
703 btrfs_put_root(root);
708 * helper to update the 'address of tree root -> reloc tree'
711 static int __update_reloc_root(struct btrfs_root *root)
713 struct btrfs_fs_info *fs_info = root->fs_info;
714 struct rb_node *rb_node;
715 struct mapping_node *node = NULL;
716 struct reloc_control *rc = fs_info->reloc_ctl;
718 spin_lock(&rc->reloc_root_tree.lock);
719 rb_node = rb_simple_search(&rc->reloc_root_tree.rb_root,
720 root->commit_root->start);
722 node = rb_entry(rb_node, struct mapping_node, rb_node);
723 rb_erase(&node->rb_node, &rc->reloc_root_tree.rb_root);
725 spin_unlock(&rc->reloc_root_tree.lock);
729 BUG_ON((struct btrfs_root *)node->data != root);
731 spin_lock(&rc->reloc_root_tree.lock);
732 node->bytenr = root->node->start;
733 rb_node = rb_simple_insert(&rc->reloc_root_tree.rb_root,
734 node->bytenr, &node->rb_node);
735 spin_unlock(&rc->reloc_root_tree.lock);
737 btrfs_backref_panic(fs_info, node->bytenr, -EEXIST);
741 static struct btrfs_root *create_reloc_root(struct btrfs_trans_handle *trans,
742 struct btrfs_root *root, u64 objectid)
744 struct btrfs_fs_info *fs_info = root->fs_info;
745 struct btrfs_root *reloc_root;
746 struct extent_buffer *eb;
747 struct btrfs_root_item *root_item;
748 struct btrfs_key root_key;
750 bool must_abort = false;
752 root_item = kmalloc(sizeof(*root_item), GFP_NOFS);
754 return ERR_PTR(-ENOMEM);
756 root_key.objectid = BTRFS_TREE_RELOC_OBJECTID;
757 root_key.type = BTRFS_ROOT_ITEM_KEY;
758 root_key.offset = objectid;
760 if (root->root_key.objectid == objectid) {
763 /* called by btrfs_init_reloc_root */
764 ret = btrfs_copy_root(trans, root, root->commit_root, &eb,
765 BTRFS_TREE_RELOC_OBJECTID);
770 * Set the last_snapshot field to the generation of the commit
771 * root - like this ctree.c:btrfs_block_can_be_shared() behaves
772 * correctly (returns true) when the relocation root is created
773 * either inside the critical section of a transaction commit
774 * (through transaction.c:qgroup_account_snapshot()) and when
775 * it's created before the transaction commit is started.
777 commit_root_gen = btrfs_header_generation(root->commit_root);
778 btrfs_set_root_last_snapshot(&root->root_item, commit_root_gen);
781 * called by btrfs_reloc_post_snapshot_hook.
782 * the source tree is a reloc tree, all tree blocks
783 * modified after it was created have RELOC flag
784 * set in their headers. so it's OK to not update
785 * the 'last_snapshot'.
787 ret = btrfs_copy_root(trans, root, root->node, &eb,
788 BTRFS_TREE_RELOC_OBJECTID);
794 * We have changed references at this point, we must abort the
795 * transaction if anything fails.
799 memcpy(root_item, &root->root_item, sizeof(*root_item));
800 btrfs_set_root_bytenr(root_item, eb->start);
801 btrfs_set_root_level(root_item, btrfs_header_level(eb));
802 btrfs_set_root_generation(root_item, trans->transid);
804 if (root->root_key.objectid == objectid) {
805 btrfs_set_root_refs(root_item, 0);
806 memset(&root_item->drop_progress, 0,
807 sizeof(struct btrfs_disk_key));
808 btrfs_set_root_drop_level(root_item, 0);
811 btrfs_tree_unlock(eb);
812 free_extent_buffer(eb);
814 ret = btrfs_insert_root(trans, fs_info->tree_root,
815 &root_key, root_item);
821 reloc_root = btrfs_read_tree_root(fs_info->tree_root, &root_key);
822 if (IS_ERR(reloc_root)) {
823 ret = PTR_ERR(reloc_root);
826 set_bit(BTRFS_ROOT_SHAREABLE, &reloc_root->state);
827 reloc_root->last_trans = trans->transid;
833 btrfs_abort_transaction(trans, ret);
838 * create reloc tree for a given fs tree. reloc tree is just a
839 * snapshot of the fs tree with special root objectid.
841 * The reloc_root comes out of here with two references, one for
842 * root->reloc_root, and another for being on the rc->reloc_roots list.
844 int btrfs_init_reloc_root(struct btrfs_trans_handle *trans,
845 struct btrfs_root *root)
847 struct btrfs_fs_info *fs_info = root->fs_info;
848 struct btrfs_root *reloc_root;
849 struct reloc_control *rc = fs_info->reloc_ctl;
850 struct btrfs_block_rsv *rsv;
858 * The subvolume has reloc tree but the swap is finished, no need to
859 * create/update the dead reloc tree
861 if (reloc_root_is_dead(root))
865 * This is subtle but important. We do not do
866 * record_root_in_transaction for reloc roots, instead we record their
867 * corresponding fs root, and then here we update the last trans for the
868 * reloc root. This means that we have to do this for the entire life
869 * of the reloc root, regardless of which stage of the relocation we are
872 if (root->reloc_root) {
873 reloc_root = root->reloc_root;
874 reloc_root->last_trans = trans->transid;
879 * We are merging reloc roots, we do not need new reloc trees. Also
880 * reloc trees never need their own reloc tree.
882 if (!rc->create_reloc_tree ||
883 root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID)
886 if (!trans->reloc_reserved) {
887 rsv = trans->block_rsv;
888 trans->block_rsv = rc->block_rsv;
891 reloc_root = create_reloc_root(trans, root, root->root_key.objectid);
893 trans->block_rsv = rsv;
894 if (IS_ERR(reloc_root))
895 return PTR_ERR(reloc_root);
897 ret = __add_reloc_root(reloc_root);
898 ASSERT(ret != -EEXIST);
900 /* Pairs with create_reloc_root */
901 btrfs_put_root(reloc_root);
904 root->reloc_root = btrfs_grab_root(reloc_root);
909 * update root item of reloc tree
911 int btrfs_update_reloc_root(struct btrfs_trans_handle *trans,
912 struct btrfs_root *root)
914 struct btrfs_fs_info *fs_info = root->fs_info;
915 struct btrfs_root *reloc_root;
916 struct btrfs_root_item *root_item;
919 if (!have_reloc_root(root))
922 reloc_root = root->reloc_root;
923 root_item = &reloc_root->root_item;
926 * We are probably ok here, but __del_reloc_root() will drop its ref of
927 * the root. We have the ref for root->reloc_root, but just in case
928 * hold it while we update the reloc root.
930 btrfs_grab_root(reloc_root);
932 /* root->reloc_root will stay until current relocation finished */
933 if (fs_info->reloc_ctl->merge_reloc_tree &&
934 btrfs_root_refs(root_item) == 0) {
935 set_bit(BTRFS_ROOT_DEAD_RELOC_TREE, &root->state);
937 * Mark the tree as dead before we change reloc_root so
938 * have_reloc_root will not touch it from now on.
941 __del_reloc_root(reloc_root);
944 if (reloc_root->commit_root != reloc_root->node) {
945 __update_reloc_root(reloc_root);
946 btrfs_set_root_node(root_item, reloc_root->node);
947 free_extent_buffer(reloc_root->commit_root);
948 reloc_root->commit_root = btrfs_root_node(reloc_root);
951 ret = btrfs_update_root(trans, fs_info->tree_root,
952 &reloc_root->root_key, root_item);
953 btrfs_put_root(reloc_root);
958 * helper to find first cached inode with inode number >= objectid
961 static struct inode *find_next_inode(struct btrfs_root *root, u64 objectid)
963 struct rb_node *node;
964 struct rb_node *prev;
965 struct btrfs_inode *entry;
968 spin_lock(&root->inode_lock);
970 node = root->inode_tree.rb_node;
974 entry = rb_entry(node, struct btrfs_inode, rb_node);
976 if (objectid < btrfs_ino(entry))
977 node = node->rb_left;
978 else if (objectid > btrfs_ino(entry))
979 node = node->rb_right;
985 entry = rb_entry(prev, struct btrfs_inode, rb_node);
986 if (objectid <= btrfs_ino(entry)) {
990 prev = rb_next(prev);
994 entry = rb_entry(node, struct btrfs_inode, rb_node);
995 inode = igrab(&entry->vfs_inode);
997 spin_unlock(&root->inode_lock);
1001 objectid = btrfs_ino(entry) + 1;
1002 if (cond_resched_lock(&root->inode_lock))
1005 node = rb_next(node);
1007 spin_unlock(&root->inode_lock);
1012 * get new location of data
1014 static int get_new_location(struct inode *reloc_inode, u64 *new_bytenr,
1015 u64 bytenr, u64 num_bytes)
1017 struct btrfs_root *root = BTRFS_I(reloc_inode)->root;
1018 struct btrfs_path *path;
1019 struct btrfs_file_extent_item *fi;
1020 struct extent_buffer *leaf;
1023 path = btrfs_alloc_path();
1027 bytenr -= BTRFS_I(reloc_inode)->index_cnt;
1028 ret = btrfs_lookup_file_extent(NULL, root, path,
1029 btrfs_ino(BTRFS_I(reloc_inode)), bytenr, 0);
1037 leaf = path->nodes[0];
1038 fi = btrfs_item_ptr(leaf, path->slots[0],
1039 struct btrfs_file_extent_item);
1041 BUG_ON(btrfs_file_extent_offset(leaf, fi) ||
1042 btrfs_file_extent_compression(leaf, fi) ||
1043 btrfs_file_extent_encryption(leaf, fi) ||
1044 btrfs_file_extent_other_encoding(leaf, fi));
1046 if (num_bytes != btrfs_file_extent_disk_num_bytes(leaf, fi)) {
1051 *new_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
1054 btrfs_free_path(path);
1059 * update file extent items in the tree leaf to point to
1060 * the new locations.
1062 static noinline_for_stack
1063 int replace_file_extents(struct btrfs_trans_handle *trans,
1064 struct reloc_control *rc,
1065 struct btrfs_root *root,
1066 struct extent_buffer *leaf)
1068 struct btrfs_fs_info *fs_info = root->fs_info;
1069 struct btrfs_key key;
1070 struct btrfs_file_extent_item *fi;
1071 struct inode *inode = NULL;
1083 if (rc->stage != UPDATE_DATA_PTRS)
1086 /* reloc trees always use full backref */
1087 if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID)
1088 parent = leaf->start;
1092 nritems = btrfs_header_nritems(leaf);
1093 for (i = 0; i < nritems; i++) {
1094 struct btrfs_ref ref = { 0 };
1097 btrfs_item_key_to_cpu(leaf, &key, i);
1098 if (key.type != BTRFS_EXTENT_DATA_KEY)
1100 fi = btrfs_item_ptr(leaf, i, struct btrfs_file_extent_item);
1101 if (btrfs_file_extent_type(leaf, fi) ==
1102 BTRFS_FILE_EXTENT_INLINE)
1104 bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
1105 num_bytes = btrfs_file_extent_disk_num_bytes(leaf, fi);
1108 if (!in_range(bytenr, rc->block_group->start,
1109 rc->block_group->length))
1113 * if we are modifying block in fs tree, wait for read_folio
1114 * to complete and drop the extent cache
1116 if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) {
1118 inode = find_next_inode(root, key.objectid);
1120 } else if (inode && btrfs_ino(BTRFS_I(inode)) < key.objectid) {
1121 btrfs_add_delayed_iput(BTRFS_I(inode));
1122 inode = find_next_inode(root, key.objectid);
1124 if (inode && btrfs_ino(BTRFS_I(inode)) == key.objectid) {
1125 struct extent_state *cached_state = NULL;
1128 btrfs_file_extent_num_bytes(leaf, fi);
1129 WARN_ON(!IS_ALIGNED(key.offset,
1130 fs_info->sectorsize));
1131 WARN_ON(!IS_ALIGNED(end, fs_info->sectorsize));
1133 ret = try_lock_extent(&BTRFS_I(inode)->io_tree,
1139 btrfs_drop_extent_map_range(BTRFS_I(inode),
1140 key.offset, end, true);
1141 unlock_extent(&BTRFS_I(inode)->io_tree,
1142 key.offset, end, &cached_state);
1146 ret = get_new_location(rc->data_inode, &new_bytenr,
1150 * Don't have to abort since we've not changed anything
1151 * in the file extent yet.
1156 btrfs_set_file_extent_disk_bytenr(leaf, fi, new_bytenr);
1159 key.offset -= btrfs_file_extent_offset(leaf, fi);
1160 btrfs_init_generic_ref(&ref, BTRFS_ADD_DELAYED_REF, new_bytenr,
1162 btrfs_init_data_ref(&ref, btrfs_header_owner(leaf),
1163 key.objectid, key.offset,
1164 root->root_key.objectid, false);
1165 ret = btrfs_inc_extent_ref(trans, &ref);
1167 btrfs_abort_transaction(trans, ret);
1171 btrfs_init_generic_ref(&ref, BTRFS_DROP_DELAYED_REF, bytenr,
1173 btrfs_init_data_ref(&ref, btrfs_header_owner(leaf),
1174 key.objectid, key.offset,
1175 root->root_key.objectid, false);
1176 ret = btrfs_free_extent(trans, &ref);
1178 btrfs_abort_transaction(trans, ret);
1183 btrfs_mark_buffer_dirty(leaf);
1185 btrfs_add_delayed_iput(BTRFS_I(inode));
1189 static noinline_for_stack
1190 int memcmp_node_keys(struct extent_buffer *eb, int slot,
1191 struct btrfs_path *path, int level)
1193 struct btrfs_disk_key key1;
1194 struct btrfs_disk_key key2;
1195 btrfs_node_key(eb, &key1, slot);
1196 btrfs_node_key(path->nodes[level], &key2, path->slots[level]);
1197 return memcmp(&key1, &key2, sizeof(key1));
1201 * try to replace tree blocks in fs tree with the new blocks
1202 * in reloc tree. tree blocks haven't been modified since the
1203 * reloc tree was create can be replaced.
1205 * if a block was replaced, level of the block + 1 is returned.
1206 * if no block got replaced, 0 is returned. if there are other
1207 * errors, a negative error number is returned.
1209 static noinline_for_stack
1210 int replace_path(struct btrfs_trans_handle *trans, struct reloc_control *rc,
1211 struct btrfs_root *dest, struct btrfs_root *src,
1212 struct btrfs_path *path, struct btrfs_key *next_key,
1213 int lowest_level, int max_level)
1215 struct btrfs_fs_info *fs_info = dest->fs_info;
1216 struct extent_buffer *eb;
1217 struct extent_buffer *parent;
1218 struct btrfs_ref ref = { 0 };
1219 struct btrfs_key key;
1231 ASSERT(src->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID);
1232 ASSERT(dest->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID);
1234 last_snapshot = btrfs_root_last_snapshot(&src->root_item);
1236 slot = path->slots[lowest_level];
1237 btrfs_node_key_to_cpu(path->nodes[lowest_level], &key, slot);
1239 eb = btrfs_lock_root_node(dest);
1240 level = btrfs_header_level(eb);
1242 if (level < lowest_level) {
1243 btrfs_tree_unlock(eb);
1244 free_extent_buffer(eb);
1249 ret = btrfs_cow_block(trans, dest, eb, NULL, 0, &eb,
1252 btrfs_tree_unlock(eb);
1253 free_extent_buffer(eb);
1259 next_key->objectid = (u64)-1;
1260 next_key->type = (u8)-1;
1261 next_key->offset = (u64)-1;
1266 level = btrfs_header_level(parent);
1267 ASSERT(level >= lowest_level);
1269 ret = btrfs_bin_search(parent, &key, &slot);
1272 if (ret && slot > 0)
1275 if (next_key && slot + 1 < btrfs_header_nritems(parent))
1276 btrfs_node_key_to_cpu(parent, next_key, slot + 1);
1278 old_bytenr = btrfs_node_blockptr(parent, slot);
1279 blocksize = fs_info->nodesize;
1280 old_ptr_gen = btrfs_node_ptr_generation(parent, slot);
1282 if (level <= max_level) {
1283 eb = path->nodes[level];
1284 new_bytenr = btrfs_node_blockptr(eb,
1285 path->slots[level]);
1286 new_ptr_gen = btrfs_node_ptr_generation(eb,
1287 path->slots[level]);
1293 if (WARN_ON(new_bytenr > 0 && new_bytenr == old_bytenr)) {
1298 if (new_bytenr == 0 || old_ptr_gen > last_snapshot ||
1299 memcmp_node_keys(parent, slot, path, level)) {
1300 if (level <= lowest_level) {
1305 eb = btrfs_read_node_slot(parent, slot);
1310 btrfs_tree_lock(eb);
1312 ret = btrfs_cow_block(trans, dest, eb, parent,
1316 btrfs_tree_unlock(eb);
1317 free_extent_buffer(eb);
1322 btrfs_tree_unlock(parent);
1323 free_extent_buffer(parent);
1330 btrfs_tree_unlock(parent);
1331 free_extent_buffer(parent);
1336 btrfs_node_key_to_cpu(path->nodes[level], &key,
1337 path->slots[level]);
1338 btrfs_release_path(path);
1340 path->lowest_level = level;
1341 set_bit(BTRFS_ROOT_RESET_LOCKDEP_CLASS, &src->state);
1342 ret = btrfs_search_slot(trans, src, &key, path, 0, 1);
1343 clear_bit(BTRFS_ROOT_RESET_LOCKDEP_CLASS, &src->state);
1344 path->lowest_level = 0;
1352 * Info qgroup to trace both subtrees.
1354 * We must trace both trees.
1355 * 1) Tree reloc subtree
1356 * If not traced, we will leak data numbers
1358 * If not traced, we will double count old data
1360 * We don't scan the subtree right now, but only record
1361 * the swapped tree blocks.
1362 * The real subtree rescan is delayed until we have new
1363 * CoW on the subtree root node before transaction commit.
1365 ret = btrfs_qgroup_add_swapped_blocks(trans, dest,
1366 rc->block_group, parent, slot,
1367 path->nodes[level], path->slots[level],
1372 * swap blocks in fs tree and reloc tree.
1374 btrfs_set_node_blockptr(parent, slot, new_bytenr);
1375 btrfs_set_node_ptr_generation(parent, slot, new_ptr_gen);
1376 btrfs_mark_buffer_dirty(parent);
1378 btrfs_set_node_blockptr(path->nodes[level],
1379 path->slots[level], old_bytenr);
1380 btrfs_set_node_ptr_generation(path->nodes[level],
1381 path->slots[level], old_ptr_gen);
1382 btrfs_mark_buffer_dirty(path->nodes[level]);
1384 btrfs_init_generic_ref(&ref, BTRFS_ADD_DELAYED_REF, old_bytenr,
1385 blocksize, path->nodes[level]->start);
1386 btrfs_init_tree_ref(&ref, level - 1, src->root_key.objectid,
1388 ret = btrfs_inc_extent_ref(trans, &ref);
1390 btrfs_abort_transaction(trans, ret);
1393 btrfs_init_generic_ref(&ref, BTRFS_ADD_DELAYED_REF, new_bytenr,
1395 btrfs_init_tree_ref(&ref, level - 1, dest->root_key.objectid, 0,
1397 ret = btrfs_inc_extent_ref(trans, &ref);
1399 btrfs_abort_transaction(trans, ret);
1403 btrfs_init_generic_ref(&ref, BTRFS_DROP_DELAYED_REF, new_bytenr,
1404 blocksize, path->nodes[level]->start);
1405 btrfs_init_tree_ref(&ref, level - 1, src->root_key.objectid,
1407 ret = btrfs_free_extent(trans, &ref);
1409 btrfs_abort_transaction(trans, ret);
1413 btrfs_init_generic_ref(&ref, BTRFS_DROP_DELAYED_REF, old_bytenr,
1415 btrfs_init_tree_ref(&ref, level - 1, dest->root_key.objectid,
1417 ret = btrfs_free_extent(trans, &ref);
1419 btrfs_abort_transaction(trans, ret);
1423 btrfs_unlock_up_safe(path, 0);
1428 btrfs_tree_unlock(parent);
1429 free_extent_buffer(parent);
1434 * helper to find next relocated block in reloc tree
1436 static noinline_for_stack
1437 int walk_up_reloc_tree(struct btrfs_root *root, struct btrfs_path *path,
1440 struct extent_buffer *eb;
1445 last_snapshot = btrfs_root_last_snapshot(&root->root_item);
1447 for (i = 0; i < *level; i++) {
1448 free_extent_buffer(path->nodes[i]);
1449 path->nodes[i] = NULL;
1452 for (i = *level; i < BTRFS_MAX_LEVEL && path->nodes[i]; i++) {
1453 eb = path->nodes[i];
1454 nritems = btrfs_header_nritems(eb);
1455 while (path->slots[i] + 1 < nritems) {
1457 if (btrfs_node_ptr_generation(eb, path->slots[i]) <=
1464 free_extent_buffer(path->nodes[i]);
1465 path->nodes[i] = NULL;
1471 * walk down reloc tree to find relocated block of lowest level
1473 static noinline_for_stack
1474 int walk_down_reloc_tree(struct btrfs_root *root, struct btrfs_path *path,
1477 struct extent_buffer *eb = NULL;
1483 last_snapshot = btrfs_root_last_snapshot(&root->root_item);
1485 for (i = *level; i > 0; i--) {
1486 eb = path->nodes[i];
1487 nritems = btrfs_header_nritems(eb);
1488 while (path->slots[i] < nritems) {
1489 ptr_gen = btrfs_node_ptr_generation(eb, path->slots[i]);
1490 if (ptr_gen > last_snapshot)
1494 if (path->slots[i] >= nritems) {
1505 eb = btrfs_read_node_slot(eb, path->slots[i]);
1508 BUG_ON(btrfs_header_level(eb) != i - 1);
1509 path->nodes[i - 1] = eb;
1510 path->slots[i - 1] = 0;
1516 * invalidate extent cache for file extents whose key in range of
1517 * [min_key, max_key)
1519 static int invalidate_extent_cache(struct btrfs_root *root,
1520 struct btrfs_key *min_key,
1521 struct btrfs_key *max_key)
1523 struct btrfs_fs_info *fs_info = root->fs_info;
1524 struct inode *inode = NULL;
1529 objectid = min_key->objectid;
1531 struct extent_state *cached_state = NULL;
1536 if (objectid > max_key->objectid)
1539 inode = find_next_inode(root, objectid);
1542 ino = btrfs_ino(BTRFS_I(inode));
1544 if (ino > max_key->objectid) {
1550 if (!S_ISREG(inode->i_mode))
1553 if (unlikely(min_key->objectid == ino)) {
1554 if (min_key->type > BTRFS_EXTENT_DATA_KEY)
1556 if (min_key->type < BTRFS_EXTENT_DATA_KEY)
1559 start = min_key->offset;
1560 WARN_ON(!IS_ALIGNED(start, fs_info->sectorsize));
1566 if (unlikely(max_key->objectid == ino)) {
1567 if (max_key->type < BTRFS_EXTENT_DATA_KEY)
1569 if (max_key->type > BTRFS_EXTENT_DATA_KEY) {
1572 if (max_key->offset == 0)
1574 end = max_key->offset;
1575 WARN_ON(!IS_ALIGNED(end, fs_info->sectorsize));
1582 /* the lock_extent waits for read_folio to complete */
1583 lock_extent(&BTRFS_I(inode)->io_tree, start, end, &cached_state);
1584 btrfs_drop_extent_map_range(BTRFS_I(inode), start, end, true);
1585 unlock_extent(&BTRFS_I(inode)->io_tree, start, end, &cached_state);
1590 static int find_next_key(struct btrfs_path *path, int level,
1591 struct btrfs_key *key)
1594 while (level < BTRFS_MAX_LEVEL) {
1595 if (!path->nodes[level])
1597 if (path->slots[level] + 1 <
1598 btrfs_header_nritems(path->nodes[level])) {
1599 btrfs_node_key_to_cpu(path->nodes[level], key,
1600 path->slots[level] + 1);
1609 * Insert current subvolume into reloc_control::dirty_subvol_roots
1611 static int insert_dirty_subvol(struct btrfs_trans_handle *trans,
1612 struct reloc_control *rc,
1613 struct btrfs_root *root)
1615 struct btrfs_root *reloc_root = root->reloc_root;
1616 struct btrfs_root_item *reloc_root_item;
1619 /* @root must be a subvolume tree root with a valid reloc tree */
1620 ASSERT(root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID);
1623 reloc_root_item = &reloc_root->root_item;
1624 memset(&reloc_root_item->drop_progress, 0,
1625 sizeof(reloc_root_item->drop_progress));
1626 btrfs_set_root_drop_level(reloc_root_item, 0);
1627 btrfs_set_root_refs(reloc_root_item, 0);
1628 ret = btrfs_update_reloc_root(trans, root);
1632 if (list_empty(&root->reloc_dirty_list)) {
1633 btrfs_grab_root(root);
1634 list_add_tail(&root->reloc_dirty_list, &rc->dirty_subvol_roots);
1640 static int clean_dirty_subvols(struct reloc_control *rc)
1642 struct btrfs_root *root;
1643 struct btrfs_root *next;
1647 list_for_each_entry_safe(root, next, &rc->dirty_subvol_roots,
1649 if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) {
1650 /* Merged subvolume, cleanup its reloc root */
1651 struct btrfs_root *reloc_root = root->reloc_root;
1653 list_del_init(&root->reloc_dirty_list);
1654 root->reloc_root = NULL;
1656 * Need barrier to ensure clear_bit() only happens after
1657 * root->reloc_root = NULL. Pairs with have_reloc_root.
1660 clear_bit(BTRFS_ROOT_DEAD_RELOC_TREE, &root->state);
1663 * btrfs_drop_snapshot drops our ref we hold for
1664 * ->reloc_root. If it fails however we must
1665 * drop the ref ourselves.
1667 ret2 = btrfs_drop_snapshot(reloc_root, 0, 1);
1669 btrfs_put_root(reloc_root);
1674 btrfs_put_root(root);
1676 /* Orphan reloc tree, just clean it up */
1677 ret2 = btrfs_drop_snapshot(root, 0, 1);
1679 btrfs_put_root(root);
1689 * merge the relocated tree blocks in reloc tree with corresponding
1692 static noinline_for_stack int merge_reloc_root(struct reloc_control *rc,
1693 struct btrfs_root *root)
1695 struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
1696 struct btrfs_key key;
1697 struct btrfs_key next_key;
1698 struct btrfs_trans_handle *trans = NULL;
1699 struct btrfs_root *reloc_root;
1700 struct btrfs_root_item *root_item;
1701 struct btrfs_path *path;
1702 struct extent_buffer *leaf;
1710 path = btrfs_alloc_path();
1713 path->reada = READA_FORWARD;
1715 reloc_root = root->reloc_root;
1716 root_item = &reloc_root->root_item;
1718 if (btrfs_disk_key_objectid(&root_item->drop_progress) == 0) {
1719 level = btrfs_root_level(root_item);
1720 atomic_inc(&reloc_root->node->refs);
1721 path->nodes[level] = reloc_root->node;
1722 path->slots[level] = 0;
1724 btrfs_disk_key_to_cpu(&key, &root_item->drop_progress);
1726 level = btrfs_root_drop_level(root_item);
1728 path->lowest_level = level;
1729 ret = btrfs_search_slot(NULL, reloc_root, &key, path, 0, 0);
1730 path->lowest_level = 0;
1732 btrfs_free_path(path);
1736 btrfs_node_key_to_cpu(path->nodes[level], &next_key,
1737 path->slots[level]);
1738 WARN_ON(memcmp(&key, &next_key, sizeof(key)));
1740 btrfs_unlock_up_safe(path, 0);
1744 * In merge_reloc_root(), we modify the upper level pointer to swap the
1745 * tree blocks between reloc tree and subvolume tree. Thus for tree
1746 * block COW, we COW at most from level 1 to root level for each tree.
1748 * Thus the needed metadata size is at most root_level * nodesize,
1749 * and * 2 since we have two trees to COW.
1751 reserve_level = max_t(int, 1, btrfs_root_level(root_item));
1752 min_reserved = fs_info->nodesize * reserve_level * 2;
1753 memset(&next_key, 0, sizeof(next_key));
1756 ret = btrfs_block_rsv_refill(fs_info, rc->block_rsv,
1758 BTRFS_RESERVE_FLUSH_LIMIT);
1761 trans = btrfs_start_transaction(root, 0);
1762 if (IS_ERR(trans)) {
1763 ret = PTR_ERR(trans);
1769 * At this point we no longer have a reloc_control, so we can't
1770 * depend on btrfs_init_reloc_root to update our last_trans.
1772 * But that's ok, we started the trans handle on our
1773 * corresponding fs_root, which means it's been added to the
1774 * dirty list. At commit time we'll still call
1775 * btrfs_update_reloc_root() and update our root item
1778 reloc_root->last_trans = trans->transid;
1779 trans->block_rsv = rc->block_rsv;
1784 ret = walk_down_reloc_tree(reloc_root, path, &level);
1790 if (!find_next_key(path, level, &key) &&
1791 btrfs_comp_cpu_keys(&next_key, &key) >= 0) {
1794 ret = replace_path(trans, rc, root, reloc_root, path,
1795 &next_key, level, max_level);
1801 btrfs_node_key_to_cpu(path->nodes[level], &key,
1802 path->slots[level]);
1806 ret = walk_up_reloc_tree(reloc_root, path, &level);
1812 * save the merging progress in the drop_progress.
1813 * this is OK since root refs == 1 in this case.
1815 btrfs_node_key(path->nodes[level], &root_item->drop_progress,
1816 path->slots[level]);
1817 btrfs_set_root_drop_level(root_item, level);
1819 btrfs_end_transaction_throttle(trans);
1822 btrfs_btree_balance_dirty(fs_info);
1824 if (replaced && rc->stage == UPDATE_DATA_PTRS)
1825 invalidate_extent_cache(root, &key, &next_key);
1829 * handle the case only one block in the fs tree need to be
1830 * relocated and the block is tree root.
1832 leaf = btrfs_lock_root_node(root);
1833 ret = btrfs_cow_block(trans, root, leaf, NULL, 0, &leaf,
1835 btrfs_tree_unlock(leaf);
1836 free_extent_buffer(leaf);
1838 btrfs_free_path(path);
1841 ret = insert_dirty_subvol(trans, rc, root);
1843 btrfs_abort_transaction(trans, ret);
1847 btrfs_end_transaction_throttle(trans);
1849 btrfs_btree_balance_dirty(fs_info);
1851 if (replaced && rc->stage == UPDATE_DATA_PTRS)
1852 invalidate_extent_cache(root, &key, &next_key);
1857 static noinline_for_stack
1858 int prepare_to_merge(struct reloc_control *rc, int err)
1860 struct btrfs_root *root = rc->extent_root;
1861 struct btrfs_fs_info *fs_info = root->fs_info;
1862 struct btrfs_root *reloc_root;
1863 struct btrfs_trans_handle *trans;
1864 LIST_HEAD(reloc_roots);
1868 mutex_lock(&fs_info->reloc_mutex);
1869 rc->merging_rsv_size += fs_info->nodesize * (BTRFS_MAX_LEVEL - 1) * 2;
1870 rc->merging_rsv_size += rc->nodes_relocated * 2;
1871 mutex_unlock(&fs_info->reloc_mutex);
1875 num_bytes = rc->merging_rsv_size;
1876 ret = btrfs_block_rsv_add(fs_info, rc->block_rsv, num_bytes,
1877 BTRFS_RESERVE_FLUSH_ALL);
1882 trans = btrfs_join_transaction(rc->extent_root);
1883 if (IS_ERR(trans)) {
1885 btrfs_block_rsv_release(fs_info, rc->block_rsv,
1887 return PTR_ERR(trans);
1891 if (num_bytes != rc->merging_rsv_size) {
1892 btrfs_end_transaction(trans);
1893 btrfs_block_rsv_release(fs_info, rc->block_rsv,
1899 rc->merge_reloc_tree = 1;
1901 while (!list_empty(&rc->reloc_roots)) {
1902 reloc_root = list_entry(rc->reloc_roots.next,
1903 struct btrfs_root, root_list);
1904 list_del_init(&reloc_root->root_list);
1906 root = btrfs_get_fs_root(fs_info, reloc_root->root_key.offset,
1910 * Even if we have an error we need this reloc root
1911 * back on our list so we can clean up properly.
1913 list_add(&reloc_root->root_list, &reloc_roots);
1914 btrfs_abort_transaction(trans, (int)PTR_ERR(root));
1916 err = PTR_ERR(root);
1919 ASSERT(root->reloc_root == reloc_root);
1922 * set reference count to 1, so btrfs_recover_relocation
1923 * knows it should resumes merging
1926 btrfs_set_root_refs(&reloc_root->root_item, 1);
1927 ret = btrfs_update_reloc_root(trans, root);
1930 * Even if we have an error we need this reloc root back on our
1931 * list so we can clean up properly.
1933 list_add(&reloc_root->root_list, &reloc_roots);
1934 btrfs_put_root(root);
1937 btrfs_abort_transaction(trans, ret);
1944 list_splice(&reloc_roots, &rc->reloc_roots);
1947 err = btrfs_commit_transaction(trans);
1949 btrfs_end_transaction(trans);
1953 static noinline_for_stack
1954 void free_reloc_roots(struct list_head *list)
1956 struct btrfs_root *reloc_root, *tmp;
1958 list_for_each_entry_safe(reloc_root, tmp, list, root_list)
1959 __del_reloc_root(reloc_root);
1962 static noinline_for_stack
1963 void merge_reloc_roots(struct reloc_control *rc)
1965 struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
1966 struct btrfs_root *root;
1967 struct btrfs_root *reloc_root;
1968 LIST_HEAD(reloc_roots);
1972 root = rc->extent_root;
1975 * this serializes us with btrfs_record_root_in_transaction,
1976 * we have to make sure nobody is in the middle of
1977 * adding their roots to the list while we are
1980 mutex_lock(&fs_info->reloc_mutex);
1981 list_splice_init(&rc->reloc_roots, &reloc_roots);
1982 mutex_unlock(&fs_info->reloc_mutex);
1984 while (!list_empty(&reloc_roots)) {
1986 reloc_root = list_entry(reloc_roots.next,
1987 struct btrfs_root, root_list);
1989 root = btrfs_get_fs_root(fs_info, reloc_root->root_key.offset,
1991 if (btrfs_root_refs(&reloc_root->root_item) > 0) {
1994 * For recovery we read the fs roots on mount,
1995 * and if we didn't find the root then we marked
1996 * the reloc root as a garbage root. For normal
1997 * relocation obviously the root should exist in
1998 * memory. However there's no reason we can't
1999 * handle the error properly here just in case.
2002 ret = PTR_ERR(root);
2005 if (root->reloc_root != reloc_root) {
2007 * This is actually impossible without something
2008 * going really wrong (like weird race condition
2015 ret = merge_reloc_root(rc, root);
2016 btrfs_put_root(root);
2018 if (list_empty(&reloc_root->root_list))
2019 list_add_tail(&reloc_root->root_list,
2024 if (!IS_ERR(root)) {
2025 if (root->reloc_root == reloc_root) {
2026 root->reloc_root = NULL;
2027 btrfs_put_root(reloc_root);
2029 clear_bit(BTRFS_ROOT_DEAD_RELOC_TREE,
2031 btrfs_put_root(root);
2034 list_del_init(&reloc_root->root_list);
2035 /* Don't forget to queue this reloc root for cleanup */
2036 list_add_tail(&reloc_root->reloc_dirty_list,
2037 &rc->dirty_subvol_roots);
2047 btrfs_handle_fs_error(fs_info, ret, NULL);
2048 free_reloc_roots(&reloc_roots);
2050 /* new reloc root may be added */
2051 mutex_lock(&fs_info->reloc_mutex);
2052 list_splice_init(&rc->reloc_roots, &reloc_roots);
2053 mutex_unlock(&fs_info->reloc_mutex);
2054 free_reloc_roots(&reloc_roots);
2060 * BUG_ON(!RB_EMPTY_ROOT(&rc->reloc_root_tree.rb_root));
2062 * here, but it's wrong. If we fail to start the transaction in
2063 * prepare_to_merge() we will have only 0 ref reloc roots, none of which
2064 * have actually been removed from the reloc_root_tree rb tree. This is
2065 * fine because we're bailing here, and we hold a reference on the root
2066 * for the list that holds it, so these roots will be cleaned up when we
2067 * do the reloc_dirty_list afterwards. Meanwhile the root->reloc_root
2068 * will be cleaned up on unmount.
2070 * The remaining nodes will be cleaned up by free_reloc_control.
2074 static void free_block_list(struct rb_root *blocks)
2076 struct tree_block *block;
2077 struct rb_node *rb_node;
2078 while ((rb_node = rb_first(blocks))) {
2079 block = rb_entry(rb_node, struct tree_block, rb_node);
2080 rb_erase(rb_node, blocks);
2085 static int record_reloc_root_in_trans(struct btrfs_trans_handle *trans,
2086 struct btrfs_root *reloc_root)
2088 struct btrfs_fs_info *fs_info = reloc_root->fs_info;
2089 struct btrfs_root *root;
2092 if (reloc_root->last_trans == trans->transid)
2095 root = btrfs_get_fs_root(fs_info, reloc_root->root_key.offset, false);
2098 * This should succeed, since we can't have a reloc root without having
2099 * already looked up the actual root and created the reloc root for this
2102 * However if there's some sort of corruption where we have a ref to a
2103 * reloc root without a corresponding root this could return ENOENT.
2107 return PTR_ERR(root);
2109 if (root->reloc_root != reloc_root) {
2112 "root %llu has two reloc roots associated with it",
2113 reloc_root->root_key.offset);
2114 btrfs_put_root(root);
2117 ret = btrfs_record_root_in_trans(trans, root);
2118 btrfs_put_root(root);
2123 static noinline_for_stack
2124 struct btrfs_root *select_reloc_root(struct btrfs_trans_handle *trans,
2125 struct reloc_control *rc,
2126 struct btrfs_backref_node *node,
2127 struct btrfs_backref_edge *edges[])
2129 struct btrfs_backref_node *next;
2130 struct btrfs_root *root;
2137 next = walk_up_backref(next, edges, &index);
2141 * If there is no root, then our references for this block are
2142 * incomplete, as we should be able to walk all the way up to a
2143 * block that is owned by a root.
2145 * This path is only for SHAREABLE roots, so if we come upon a
2146 * non-SHAREABLE root then we have backrefs that resolve
2149 * Both of these cases indicate file system corruption, or a bug
2150 * in the backref walking code.
2154 btrfs_err(trans->fs_info,
2155 "bytenr %llu doesn't have a backref path ending in a root",
2157 return ERR_PTR(-EUCLEAN);
2159 if (!test_bit(BTRFS_ROOT_SHAREABLE, &root->state)) {
2161 btrfs_err(trans->fs_info,
2162 "bytenr %llu has multiple refs with one ending in a non-shareable root",
2164 return ERR_PTR(-EUCLEAN);
2167 if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID) {
2168 ret = record_reloc_root_in_trans(trans, root);
2170 return ERR_PTR(ret);
2174 ret = btrfs_record_root_in_trans(trans, root);
2176 return ERR_PTR(ret);
2177 root = root->reloc_root;
2180 * We could have raced with another thread which failed, so
2181 * root->reloc_root may not be set, return ENOENT in this case.
2184 return ERR_PTR(-ENOENT);
2186 if (next->new_bytenr != root->node->start) {
2188 * We just created the reloc root, so we shouldn't have
2189 * ->new_bytenr set and this shouldn't be in the changed
2190 * list. If it is then we have multiple roots pointing
2191 * at the same bytenr which indicates corruption, or
2192 * we've made a mistake in the backref walking code.
2194 ASSERT(next->new_bytenr == 0);
2195 ASSERT(list_empty(&next->list));
2196 if (next->new_bytenr || !list_empty(&next->list)) {
2197 btrfs_err(trans->fs_info,
2198 "bytenr %llu possibly has multiple roots pointing at the same bytenr %llu",
2199 node->bytenr, next->bytenr);
2200 return ERR_PTR(-EUCLEAN);
2203 next->new_bytenr = root->node->start;
2204 btrfs_put_root(next->root);
2205 next->root = btrfs_grab_root(root);
2207 list_add_tail(&next->list,
2208 &rc->backref_cache.changed);
2209 mark_block_processed(rc, next);
2215 next = walk_down_backref(edges, &index);
2216 if (!next || next->level <= node->level)
2221 * This can happen if there's fs corruption or if there's a bug
2222 * in the backref lookup code.
2225 return ERR_PTR(-ENOENT);
2229 /* setup backref node path for btrfs_reloc_cow_block */
2231 rc->backref_cache.path[next->level] = next;
2234 next = edges[index]->node[UPPER];
2240 * Select a tree root for relocation.
2242 * Return NULL if the block is not shareable. We should use do_relocation() in
2245 * Return a tree root pointer if the block is shareable.
2246 * Return -ENOENT if the block is root of reloc tree.
2248 static noinline_for_stack
2249 struct btrfs_root *select_one_root(struct btrfs_backref_node *node)
2251 struct btrfs_backref_node *next;
2252 struct btrfs_root *root;
2253 struct btrfs_root *fs_root = NULL;
2254 struct btrfs_backref_edge *edges[BTRFS_MAX_LEVEL - 1];
2260 next = walk_up_backref(next, edges, &index);
2264 * This can occur if we have incomplete extent refs leading all
2265 * the way up a particular path, in this case return -EUCLEAN.
2268 return ERR_PTR(-EUCLEAN);
2270 /* No other choice for non-shareable tree */
2271 if (!test_bit(BTRFS_ROOT_SHAREABLE, &root->state))
2274 if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID)
2280 next = walk_down_backref(edges, &index);
2281 if (!next || next->level <= node->level)
2286 return ERR_PTR(-ENOENT);
2290 static noinline_for_stack
2291 u64 calcu_metadata_size(struct reloc_control *rc,
2292 struct btrfs_backref_node *node, int reserve)
2294 struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
2295 struct btrfs_backref_node *next = node;
2296 struct btrfs_backref_edge *edge;
2297 struct btrfs_backref_edge *edges[BTRFS_MAX_LEVEL - 1];
2301 BUG_ON(reserve && node->processed);
2306 if (next->processed && (reserve || next != node))
2309 num_bytes += fs_info->nodesize;
2311 if (list_empty(&next->upper))
2314 edge = list_entry(next->upper.next,
2315 struct btrfs_backref_edge, list[LOWER]);
2316 edges[index++] = edge;
2317 next = edge->node[UPPER];
2319 next = walk_down_backref(edges, &index);
2324 static int reserve_metadata_space(struct btrfs_trans_handle *trans,
2325 struct reloc_control *rc,
2326 struct btrfs_backref_node *node)
2328 struct btrfs_root *root = rc->extent_root;
2329 struct btrfs_fs_info *fs_info = root->fs_info;
2334 num_bytes = calcu_metadata_size(rc, node, 1) * 2;
2336 trans->block_rsv = rc->block_rsv;
2337 rc->reserved_bytes += num_bytes;
2340 * We are under a transaction here so we can only do limited flushing.
2341 * If we get an enospc just kick back -EAGAIN so we know to drop the
2342 * transaction and try to refill when we can flush all the things.
2344 ret = btrfs_block_rsv_refill(fs_info, rc->block_rsv, num_bytes,
2345 BTRFS_RESERVE_FLUSH_LIMIT);
2347 tmp = fs_info->nodesize * RELOCATION_RESERVED_NODES;
2348 while (tmp <= rc->reserved_bytes)
2351 * only one thread can access block_rsv at this point,
2352 * so we don't need hold lock to protect block_rsv.
2353 * we expand more reservation size here to allow enough
2354 * space for relocation and we will return earlier in
2357 rc->block_rsv->size = tmp + fs_info->nodesize *
2358 RELOCATION_RESERVED_NODES;
2366 * relocate a block tree, and then update pointers in upper level
2367 * blocks that reference the block to point to the new location.
2369 * if called by link_to_upper, the block has already been relocated.
2370 * in that case this function just updates pointers.
2372 static int do_relocation(struct btrfs_trans_handle *trans,
2373 struct reloc_control *rc,
2374 struct btrfs_backref_node *node,
2375 struct btrfs_key *key,
2376 struct btrfs_path *path, int lowest)
2378 struct btrfs_backref_node *upper;
2379 struct btrfs_backref_edge *edge;
2380 struct btrfs_backref_edge *edges[BTRFS_MAX_LEVEL - 1];
2381 struct btrfs_root *root;
2382 struct extent_buffer *eb;
2389 * If we are lowest then this is the first time we're processing this
2390 * block, and thus shouldn't have an eb associated with it yet.
2392 ASSERT(!lowest || !node->eb);
2394 path->lowest_level = node->level + 1;
2395 rc->backref_cache.path[node->level] = node;
2396 list_for_each_entry(edge, &node->upper, list[LOWER]) {
2397 struct btrfs_ref ref = { 0 };
2401 upper = edge->node[UPPER];
2402 root = select_reloc_root(trans, rc, upper, edges);
2404 ret = PTR_ERR(root);
2408 if (upper->eb && !upper->locked) {
2410 ret = btrfs_bin_search(upper->eb, key, &slot);
2414 bytenr = btrfs_node_blockptr(upper->eb, slot);
2415 if (node->eb->start == bytenr)
2418 btrfs_backref_drop_node_buffer(upper);
2422 ret = btrfs_search_slot(trans, root, key, path, 0, 1);
2427 btrfs_release_path(path);
2432 upper->eb = path->nodes[upper->level];
2433 path->nodes[upper->level] = NULL;
2435 BUG_ON(upper->eb != path->nodes[upper->level]);
2439 path->locks[upper->level] = 0;
2441 slot = path->slots[upper->level];
2442 btrfs_release_path(path);
2444 ret = btrfs_bin_search(upper->eb, key, &slot);
2450 bytenr = btrfs_node_blockptr(upper->eb, slot);
2452 if (bytenr != node->bytenr) {
2453 btrfs_err(root->fs_info,
2454 "lowest leaf/node mismatch: bytenr %llu node->bytenr %llu slot %d upper %llu",
2455 bytenr, node->bytenr, slot,
2461 if (node->eb->start == bytenr)
2465 blocksize = root->fs_info->nodesize;
2466 eb = btrfs_read_node_slot(upper->eb, slot);
2471 btrfs_tree_lock(eb);
2474 ret = btrfs_cow_block(trans, root, eb, upper->eb,
2475 slot, &eb, BTRFS_NESTING_COW);
2476 btrfs_tree_unlock(eb);
2477 free_extent_buffer(eb);
2481 * We've just COWed this block, it should have updated
2482 * the correct backref node entry.
2484 ASSERT(node->eb == eb);
2486 btrfs_set_node_blockptr(upper->eb, slot,
2488 btrfs_set_node_ptr_generation(upper->eb, slot,
2490 btrfs_mark_buffer_dirty(upper->eb);
2492 btrfs_init_generic_ref(&ref, BTRFS_ADD_DELAYED_REF,
2493 node->eb->start, blocksize,
2495 btrfs_init_tree_ref(&ref, node->level,
2496 btrfs_header_owner(upper->eb),
2497 root->root_key.objectid, false);
2498 ret = btrfs_inc_extent_ref(trans, &ref);
2500 ret = btrfs_drop_subtree(trans, root, eb,
2503 btrfs_abort_transaction(trans, ret);
2506 if (!upper->pending)
2507 btrfs_backref_drop_node_buffer(upper);
2509 btrfs_backref_unlock_node_buffer(upper);
2514 if (!ret && node->pending) {
2515 btrfs_backref_drop_node_buffer(node);
2516 list_move_tail(&node->list, &rc->backref_cache.changed);
2520 path->lowest_level = 0;
2523 * We should have allocated all of our space in the block rsv and thus
2526 ASSERT(ret != -ENOSPC);
2530 static int link_to_upper(struct btrfs_trans_handle *trans,
2531 struct reloc_control *rc,
2532 struct btrfs_backref_node *node,
2533 struct btrfs_path *path)
2535 struct btrfs_key key;
2537 btrfs_node_key_to_cpu(node->eb, &key, 0);
2538 return do_relocation(trans, rc, node, &key, path, 0);
2541 static int finish_pending_nodes(struct btrfs_trans_handle *trans,
2542 struct reloc_control *rc,
2543 struct btrfs_path *path, int err)
2546 struct btrfs_backref_cache *cache = &rc->backref_cache;
2547 struct btrfs_backref_node *node;
2551 for (level = 0; level < BTRFS_MAX_LEVEL; level++) {
2552 while (!list_empty(&cache->pending[level])) {
2553 node = list_entry(cache->pending[level].next,
2554 struct btrfs_backref_node, list);
2555 list_move_tail(&node->list, &list);
2556 BUG_ON(!node->pending);
2559 ret = link_to_upper(trans, rc, node, path);
2564 list_splice_init(&list, &cache->pending[level]);
2570 * mark a block and all blocks directly/indirectly reference the block
2573 static void update_processed_blocks(struct reloc_control *rc,
2574 struct btrfs_backref_node *node)
2576 struct btrfs_backref_node *next = node;
2577 struct btrfs_backref_edge *edge;
2578 struct btrfs_backref_edge *edges[BTRFS_MAX_LEVEL - 1];
2584 if (next->processed)
2587 mark_block_processed(rc, next);
2589 if (list_empty(&next->upper))
2592 edge = list_entry(next->upper.next,
2593 struct btrfs_backref_edge, list[LOWER]);
2594 edges[index++] = edge;
2595 next = edge->node[UPPER];
2597 next = walk_down_backref(edges, &index);
2601 static int tree_block_processed(u64 bytenr, struct reloc_control *rc)
2603 u32 blocksize = rc->extent_root->fs_info->nodesize;
2605 if (test_range_bit(&rc->processed_blocks, bytenr,
2606 bytenr + blocksize - 1, EXTENT_DIRTY, 1, NULL))
2611 static int get_tree_block_key(struct btrfs_fs_info *fs_info,
2612 struct tree_block *block)
2614 struct btrfs_tree_parent_check check = {
2615 .level = block->level,
2616 .owner_root = block->owner,
2617 .transid = block->key.offset
2619 struct extent_buffer *eb;
2621 eb = read_tree_block(fs_info, block->bytenr, &check);
2624 if (!extent_buffer_uptodate(eb)) {
2625 free_extent_buffer(eb);
2628 if (block->level == 0)
2629 btrfs_item_key_to_cpu(eb, &block->key, 0);
2631 btrfs_node_key_to_cpu(eb, &block->key, 0);
2632 free_extent_buffer(eb);
2633 block->key_ready = 1;
2638 * helper function to relocate a tree block
2640 static int relocate_tree_block(struct btrfs_trans_handle *trans,
2641 struct reloc_control *rc,
2642 struct btrfs_backref_node *node,
2643 struct btrfs_key *key,
2644 struct btrfs_path *path)
2646 struct btrfs_root *root;
2653 * If we fail here we want to drop our backref_node because we are going
2654 * to start over and regenerate the tree for it.
2656 ret = reserve_metadata_space(trans, rc, node);
2660 BUG_ON(node->processed);
2661 root = select_one_root(node);
2663 ret = PTR_ERR(root);
2665 /* See explanation in select_one_root for the -EUCLEAN case. */
2666 ASSERT(ret == -ENOENT);
2667 if (ret == -ENOENT) {
2669 update_processed_blocks(rc, node);
2675 if (test_bit(BTRFS_ROOT_SHAREABLE, &root->state)) {
2677 * This block was the root block of a root, and this is
2678 * the first time we're processing the block and thus it
2679 * should not have had the ->new_bytenr modified and
2680 * should have not been included on the changed list.
2682 * However in the case of corruption we could have
2683 * multiple refs pointing to the same block improperly,
2684 * and thus we would trip over these checks. ASSERT()
2685 * for the developer case, because it could indicate a
2686 * bug in the backref code, however error out for a
2687 * normal user in the case of corruption.
2689 ASSERT(node->new_bytenr == 0);
2690 ASSERT(list_empty(&node->list));
2691 if (node->new_bytenr || !list_empty(&node->list)) {
2692 btrfs_err(root->fs_info,
2693 "bytenr %llu has improper references to it",
2698 ret = btrfs_record_root_in_trans(trans, root);
2702 * Another thread could have failed, need to check if we
2703 * have reloc_root actually set.
2705 if (!root->reloc_root) {
2709 root = root->reloc_root;
2710 node->new_bytenr = root->node->start;
2711 btrfs_put_root(node->root);
2712 node->root = btrfs_grab_root(root);
2714 list_add_tail(&node->list, &rc->backref_cache.changed);
2716 path->lowest_level = node->level;
2717 if (root == root->fs_info->chunk_root)
2718 btrfs_reserve_chunk_metadata(trans, false);
2719 ret = btrfs_search_slot(trans, root, key, path, 0, 1);
2720 btrfs_release_path(path);
2721 if (root == root->fs_info->chunk_root)
2722 btrfs_trans_release_chunk_metadata(trans);
2727 update_processed_blocks(rc, node);
2729 ret = do_relocation(trans, rc, node, key, path, 1);
2732 if (ret || node->level == 0 || node->cowonly)
2733 btrfs_backref_cleanup_node(&rc->backref_cache, node);
2738 * relocate a list of blocks
2740 static noinline_for_stack
2741 int relocate_tree_blocks(struct btrfs_trans_handle *trans,
2742 struct reloc_control *rc, struct rb_root *blocks)
2744 struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
2745 struct btrfs_backref_node *node;
2746 struct btrfs_path *path;
2747 struct tree_block *block;
2748 struct tree_block *next;
2752 path = btrfs_alloc_path();
2755 goto out_free_blocks;
2758 /* Kick in readahead for tree blocks with missing keys */
2759 rbtree_postorder_for_each_entry_safe(block, next, blocks, rb_node) {
2760 if (!block->key_ready)
2761 btrfs_readahead_tree_block(fs_info, block->bytenr,
2766 /* Get first keys */
2767 rbtree_postorder_for_each_entry_safe(block, next, blocks, rb_node) {
2768 if (!block->key_ready) {
2769 err = get_tree_block_key(fs_info, block);
2775 /* Do tree relocation */
2776 rbtree_postorder_for_each_entry_safe(block, next, blocks, rb_node) {
2777 node = build_backref_tree(rc, &block->key,
2778 block->level, block->bytenr);
2780 err = PTR_ERR(node);
2784 ret = relocate_tree_block(trans, rc, node, &block->key,
2792 err = finish_pending_nodes(trans, rc, path, err);
2795 btrfs_free_path(path);
2797 free_block_list(blocks);
2801 static noinline_for_stack int prealloc_file_extent_cluster(
2802 struct btrfs_inode *inode,
2803 struct file_extent_cluster *cluster)
2808 u64 offset = inode->index_cnt;
2812 u64 i_size = i_size_read(&inode->vfs_inode);
2813 u64 prealloc_start = cluster->start - offset;
2814 u64 prealloc_end = cluster->end - offset;
2815 u64 cur_offset = prealloc_start;
2818 * For subpage case, previous i_size may not be aligned to PAGE_SIZE.
2819 * This means the range [i_size, PAGE_END + 1) is filled with zeros by
2820 * btrfs_do_readpage() call of previously relocated file cluster.
2822 * If the current cluster starts in the above range, btrfs_do_readpage()
2823 * will skip the read, and relocate_one_page() will later writeback
2824 * the padding zeros as new data, causing data corruption.
2826 * Here we have to manually invalidate the range (i_size, PAGE_END + 1).
2828 if (!PAGE_ALIGNED(i_size)) {
2829 struct address_space *mapping = inode->vfs_inode.i_mapping;
2830 struct btrfs_fs_info *fs_info = inode->root->fs_info;
2831 const u32 sectorsize = fs_info->sectorsize;
2834 ASSERT(sectorsize < PAGE_SIZE);
2835 ASSERT(IS_ALIGNED(i_size, sectorsize));
2838 * Subpage can't handle page with DIRTY but without UPTODATE
2839 * bit as it can lead to the following deadlock:
2841 * btrfs_read_folio()
2842 * | Page already *locked*
2843 * |- btrfs_lock_and_flush_ordered_range()
2844 * |- btrfs_start_ordered_extent()
2845 * |- extent_write_cache_pages()
2847 * We try to lock the page we already hold.
2849 * Here we just writeback the whole data reloc inode, so that
2850 * we will be ensured to have no dirty range in the page, and
2851 * are safe to clear the uptodate bits.
2853 * This shouldn't cause too much overhead, as we need to write
2854 * the data back anyway.
2856 ret = filemap_write_and_wait(mapping);
2860 clear_extent_bits(&inode->io_tree, i_size,
2861 round_up(i_size, PAGE_SIZE) - 1,
2863 page = find_lock_page(mapping, i_size >> PAGE_SHIFT);
2865 * If page is freed we don't need to do anything then, as we
2866 * will re-read the whole page anyway.
2869 btrfs_subpage_clear_uptodate(fs_info, page, i_size,
2870 round_up(i_size, PAGE_SIZE) - i_size);
2876 BUG_ON(cluster->start != cluster->boundary[0]);
2877 ret = btrfs_alloc_data_chunk_ondemand(inode,
2878 prealloc_end + 1 - prealloc_start);
2882 btrfs_inode_lock(inode, 0);
2883 for (nr = 0; nr < cluster->nr; nr++) {
2884 struct extent_state *cached_state = NULL;
2886 start = cluster->boundary[nr] - offset;
2887 if (nr + 1 < cluster->nr)
2888 end = cluster->boundary[nr + 1] - 1 - offset;
2890 end = cluster->end - offset;
2892 lock_extent(&inode->io_tree, start, end, &cached_state);
2893 num_bytes = end + 1 - start;
2894 ret = btrfs_prealloc_file_range(&inode->vfs_inode, 0, start,
2895 num_bytes, num_bytes,
2896 end + 1, &alloc_hint);
2897 cur_offset = end + 1;
2898 unlock_extent(&inode->io_tree, start, end, &cached_state);
2902 btrfs_inode_unlock(inode, 0);
2904 if (cur_offset < prealloc_end)
2905 btrfs_free_reserved_data_space_noquota(inode->root->fs_info,
2906 prealloc_end + 1 - cur_offset);
2910 static noinline_for_stack int setup_relocation_extent_mapping(struct inode *inode,
2911 u64 start, u64 end, u64 block_start)
2913 struct extent_map *em;
2914 struct extent_state *cached_state = NULL;
2917 em = alloc_extent_map();
2922 em->len = end + 1 - start;
2923 em->block_len = em->len;
2924 em->block_start = block_start;
2925 set_bit(EXTENT_FLAG_PINNED, &em->flags);
2927 lock_extent(&BTRFS_I(inode)->io_tree, start, end, &cached_state);
2928 ret = btrfs_replace_extent_map_range(BTRFS_I(inode), em, false);
2929 unlock_extent(&BTRFS_I(inode)->io_tree, start, end, &cached_state);
2930 free_extent_map(em);
2936 * Allow error injection to test balance/relocation cancellation
2938 noinline int btrfs_should_cancel_balance(struct btrfs_fs_info *fs_info)
2940 return atomic_read(&fs_info->balance_cancel_req) ||
2941 atomic_read(&fs_info->reloc_cancel_req) ||
2942 fatal_signal_pending(current);
2944 ALLOW_ERROR_INJECTION(btrfs_should_cancel_balance, TRUE);
2946 static u64 get_cluster_boundary_end(struct file_extent_cluster *cluster,
2949 /* Last extent, use cluster end directly */
2950 if (cluster_nr >= cluster->nr - 1)
2951 return cluster->end;
2953 /* Use next boundary start*/
2954 return cluster->boundary[cluster_nr + 1] - 1;
2957 static int relocate_one_page(struct inode *inode, struct file_ra_state *ra,
2958 struct file_extent_cluster *cluster,
2959 int *cluster_nr, unsigned long page_index)
2961 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2962 u64 offset = BTRFS_I(inode)->index_cnt;
2963 const unsigned long last_index = (cluster->end - offset) >> PAGE_SHIFT;
2964 gfp_t mask = btrfs_alloc_write_mask(inode->i_mapping);
2971 ASSERT(page_index <= last_index);
2972 page = find_lock_page(inode->i_mapping, page_index);
2974 page_cache_sync_readahead(inode->i_mapping, ra, NULL,
2975 page_index, last_index + 1 - page_index);
2976 page = find_or_create_page(inode->i_mapping, page_index, mask);
2980 ret = set_page_extent_mapped(page);
2984 if (PageReadahead(page))
2985 page_cache_async_readahead(inode->i_mapping, ra, NULL,
2986 page_folio(page), page_index,
2987 last_index + 1 - page_index);
2989 if (!PageUptodate(page)) {
2990 btrfs_read_folio(NULL, page_folio(page));
2992 if (!PageUptodate(page)) {
2998 page_start = page_offset(page);
2999 page_end = page_start + PAGE_SIZE - 1;
3002 * Start from the cluster, as for subpage case, the cluster can start
3005 cur = max(page_start, cluster->boundary[*cluster_nr] - offset);
3006 while (cur <= page_end) {
3007 struct extent_state *cached_state = NULL;
3008 u64 extent_start = cluster->boundary[*cluster_nr] - offset;
3009 u64 extent_end = get_cluster_boundary_end(cluster,
3010 *cluster_nr) - offset;
3011 u64 clamped_start = max(page_start, extent_start);
3012 u64 clamped_end = min(page_end, extent_end);
3013 u32 clamped_len = clamped_end + 1 - clamped_start;
3015 /* Reserve metadata for this range */
3016 ret = btrfs_delalloc_reserve_metadata(BTRFS_I(inode),
3017 clamped_len, clamped_len,
3022 /* Mark the range delalloc and dirty for later writeback */
3023 lock_extent(&BTRFS_I(inode)->io_tree, clamped_start, clamped_end,
3025 ret = btrfs_set_extent_delalloc(BTRFS_I(inode), clamped_start,
3026 clamped_end, 0, &cached_state);
3028 clear_extent_bit(&BTRFS_I(inode)->io_tree,
3029 clamped_start, clamped_end,
3030 EXTENT_LOCKED | EXTENT_BOUNDARY,
3032 btrfs_delalloc_release_metadata(BTRFS_I(inode),
3034 btrfs_delalloc_release_extents(BTRFS_I(inode),
3038 btrfs_page_set_dirty(fs_info, page, clamped_start, clamped_len);
3041 * Set the boundary if it's inside the page.
3042 * Data relocation requires the destination extents to have the
3043 * same size as the source.
3044 * EXTENT_BOUNDARY bit prevents current extent from being merged
3045 * with previous extent.
3047 if (in_range(cluster->boundary[*cluster_nr] - offset,
3048 page_start, PAGE_SIZE)) {
3049 u64 boundary_start = cluster->boundary[*cluster_nr] -
3051 u64 boundary_end = boundary_start +
3052 fs_info->sectorsize - 1;
3054 set_extent_bits(&BTRFS_I(inode)->io_tree,
3055 boundary_start, boundary_end,
3058 unlock_extent(&BTRFS_I(inode)->io_tree, clamped_start, clamped_end,
3060 btrfs_delalloc_release_extents(BTRFS_I(inode), clamped_len);
3063 /* Crossed extent end, go to next extent */
3064 if (cur >= extent_end) {
3066 /* Just finished the last extent of the cluster, exit. */
3067 if (*cluster_nr >= cluster->nr)
3074 balance_dirty_pages_ratelimited(inode->i_mapping);
3075 btrfs_throttle(fs_info);
3076 if (btrfs_should_cancel_balance(fs_info))
3086 static int relocate_file_extent_cluster(struct inode *inode,
3087 struct file_extent_cluster *cluster)
3089 u64 offset = BTRFS_I(inode)->index_cnt;
3090 unsigned long index;
3091 unsigned long last_index;
3092 struct file_ra_state *ra;
3099 ra = kzalloc(sizeof(*ra), GFP_NOFS);
3103 ret = prealloc_file_extent_cluster(BTRFS_I(inode), cluster);
3107 file_ra_state_init(ra, inode->i_mapping);
3109 ret = setup_relocation_extent_mapping(inode, cluster->start - offset,
3110 cluster->end - offset, cluster->start);
3114 last_index = (cluster->end - offset) >> PAGE_SHIFT;
3115 for (index = (cluster->start - offset) >> PAGE_SHIFT;
3116 index <= last_index && !ret; index++)
3117 ret = relocate_one_page(inode, ra, cluster, &cluster_nr, index);
3119 WARN_ON(cluster_nr != cluster->nr);
3125 static noinline_for_stack
3126 int relocate_data_extent(struct inode *inode, struct btrfs_key *extent_key,
3127 struct file_extent_cluster *cluster)
3131 if (cluster->nr > 0 && extent_key->objectid != cluster->end + 1) {
3132 ret = relocate_file_extent_cluster(inode, cluster);
3139 cluster->start = extent_key->objectid;
3141 BUG_ON(cluster->nr >= MAX_EXTENTS);
3142 cluster->end = extent_key->objectid + extent_key->offset - 1;
3143 cluster->boundary[cluster->nr] = extent_key->objectid;
3146 if (cluster->nr >= MAX_EXTENTS) {
3147 ret = relocate_file_extent_cluster(inode, cluster);
3156 * helper to add a tree block to the list.
3157 * the major work is getting the generation and level of the block
3159 static int add_tree_block(struct reloc_control *rc,
3160 struct btrfs_key *extent_key,
3161 struct btrfs_path *path,
3162 struct rb_root *blocks)
3164 struct extent_buffer *eb;
3165 struct btrfs_extent_item *ei;
3166 struct btrfs_tree_block_info *bi;
3167 struct tree_block *block;
3168 struct rb_node *rb_node;
3174 eb = path->nodes[0];
3175 item_size = btrfs_item_size(eb, path->slots[0]);
3177 if (extent_key->type == BTRFS_METADATA_ITEM_KEY ||
3178 item_size >= sizeof(*ei) + sizeof(*bi)) {
3179 unsigned long ptr = 0, end;
3181 ei = btrfs_item_ptr(eb, path->slots[0],
3182 struct btrfs_extent_item);
3183 end = (unsigned long)ei + item_size;
3184 if (extent_key->type == BTRFS_EXTENT_ITEM_KEY) {
3185 bi = (struct btrfs_tree_block_info *)(ei + 1);
3186 level = btrfs_tree_block_level(eb, bi);
3187 ptr = (unsigned long)(bi + 1);
3189 level = (int)extent_key->offset;
3190 ptr = (unsigned long)(ei + 1);
3192 generation = btrfs_extent_generation(eb, ei);
3195 * We're reading random blocks without knowing their owner ahead
3196 * of time. This is ok most of the time, as all reloc roots and
3197 * fs roots have the same lock type. However normal trees do
3198 * not, and the only way to know ahead of time is to read the
3199 * inline ref offset. We know it's an fs root if
3201 * 1. There's more than one ref.
3202 * 2. There's a SHARED_DATA_REF_KEY set.
3203 * 3. FULL_BACKREF is set on the flags.
3205 * Otherwise it's safe to assume that the ref offset == the
3206 * owner of this block, so we can use that when calling
3209 if (btrfs_extent_refs(eb, ei) == 1 &&
3210 !(btrfs_extent_flags(eb, ei) &
3211 BTRFS_BLOCK_FLAG_FULL_BACKREF) &&
3213 struct btrfs_extent_inline_ref *iref;
3216 iref = (struct btrfs_extent_inline_ref *)ptr;
3217 type = btrfs_get_extent_inline_ref_type(eb, iref,
3218 BTRFS_REF_TYPE_BLOCK);
3219 if (type == BTRFS_REF_TYPE_INVALID)
3221 if (type == BTRFS_TREE_BLOCK_REF_KEY)
3222 owner = btrfs_extent_inline_ref_offset(eb, iref);
3224 } else if (unlikely(item_size == sizeof(struct btrfs_extent_item_v0))) {
3225 btrfs_print_v0_err(eb->fs_info);
3226 btrfs_handle_fs_error(eb->fs_info, -EINVAL, NULL);
3232 btrfs_release_path(path);
3234 BUG_ON(level == -1);
3236 block = kmalloc(sizeof(*block), GFP_NOFS);
3240 block->bytenr = extent_key->objectid;
3241 block->key.objectid = rc->extent_root->fs_info->nodesize;
3242 block->key.offset = generation;
3243 block->level = level;
3244 block->key_ready = 0;
3245 block->owner = owner;
3247 rb_node = rb_simple_insert(blocks, block->bytenr, &block->rb_node);
3249 btrfs_backref_panic(rc->extent_root->fs_info, block->bytenr,
3256 * helper to add tree blocks for backref of type BTRFS_SHARED_DATA_REF_KEY
3258 static int __add_tree_block(struct reloc_control *rc,
3259 u64 bytenr, u32 blocksize,
3260 struct rb_root *blocks)
3262 struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
3263 struct btrfs_path *path;
3264 struct btrfs_key key;
3266 bool skinny = btrfs_fs_incompat(fs_info, SKINNY_METADATA);
3268 if (tree_block_processed(bytenr, rc))
3271 if (rb_simple_search(blocks, bytenr))
3274 path = btrfs_alloc_path();
3278 key.objectid = bytenr;
3280 key.type = BTRFS_METADATA_ITEM_KEY;
3281 key.offset = (u64)-1;
3283 key.type = BTRFS_EXTENT_ITEM_KEY;
3284 key.offset = blocksize;
3287 path->search_commit_root = 1;
3288 path->skip_locking = 1;
3289 ret = btrfs_search_slot(NULL, rc->extent_root, &key, path, 0, 0);
3293 if (ret > 0 && skinny) {
3294 if (path->slots[0]) {
3296 btrfs_item_key_to_cpu(path->nodes[0], &key,
3298 if (key.objectid == bytenr &&
3299 (key.type == BTRFS_METADATA_ITEM_KEY ||
3300 (key.type == BTRFS_EXTENT_ITEM_KEY &&
3301 key.offset == blocksize)))
3307 btrfs_release_path(path);
3313 btrfs_print_leaf(path->nodes[0]);
3315 "tree block extent item (%llu) is not found in extent tree",
3322 ret = add_tree_block(rc, &key, path, blocks);
3324 btrfs_free_path(path);
3328 static int delete_block_group_cache(struct btrfs_fs_info *fs_info,
3329 struct btrfs_block_group *block_group,
3330 struct inode *inode,
3333 struct btrfs_root *root = fs_info->tree_root;
3334 struct btrfs_trans_handle *trans;
3340 inode = btrfs_iget(fs_info->sb, ino, root);
3345 ret = btrfs_check_trunc_cache_free_space(fs_info,
3346 &fs_info->global_block_rsv);
3350 trans = btrfs_join_transaction(root);
3351 if (IS_ERR(trans)) {
3352 ret = PTR_ERR(trans);
3356 ret = btrfs_truncate_free_space_cache(trans, block_group, inode);
3358 btrfs_end_transaction(trans);
3359 btrfs_btree_balance_dirty(fs_info);
3366 * Locate the free space cache EXTENT_DATA in root tree leaf and delete the
3367 * cache inode, to avoid free space cache data extent blocking data relocation.
3369 static int delete_v1_space_cache(struct extent_buffer *leaf,
3370 struct btrfs_block_group *block_group,
3373 u64 space_cache_ino;
3374 struct btrfs_file_extent_item *ei;
3375 struct btrfs_key key;
3380 if (btrfs_header_owner(leaf) != BTRFS_ROOT_TREE_OBJECTID)
3383 for (i = 0; i < btrfs_header_nritems(leaf); i++) {
3386 btrfs_item_key_to_cpu(leaf, &key, i);
3387 if (key.type != BTRFS_EXTENT_DATA_KEY)
3389 ei = btrfs_item_ptr(leaf, i, struct btrfs_file_extent_item);
3390 type = btrfs_file_extent_type(leaf, ei);
3392 if ((type == BTRFS_FILE_EXTENT_REG ||
3393 type == BTRFS_FILE_EXTENT_PREALLOC) &&
3394 btrfs_file_extent_disk_bytenr(leaf, ei) == data_bytenr) {
3396 space_cache_ino = key.objectid;
3402 ret = delete_block_group_cache(leaf->fs_info, block_group, NULL,
3408 * helper to find all tree blocks that reference a given data extent
3410 static noinline_for_stack
3411 int add_data_references(struct reloc_control *rc,
3412 struct btrfs_key *extent_key,
3413 struct btrfs_path *path,
3414 struct rb_root *blocks)
3416 struct btrfs_backref_walk_ctx ctx = { 0 };
3417 struct ulist_iterator leaf_uiter;
3418 struct ulist_node *ref_node = NULL;
3419 const u32 blocksize = rc->extent_root->fs_info->nodesize;
3422 btrfs_release_path(path);
3424 ctx.bytenr = extent_key->objectid;
3425 ctx.ignore_extent_item_pos = true;
3426 ctx.fs_info = rc->extent_root->fs_info;
3428 ret = btrfs_find_all_leafs(&ctx);
3432 ULIST_ITER_INIT(&leaf_uiter);
3433 while ((ref_node = ulist_next(ctx.refs, &leaf_uiter))) {
3434 struct btrfs_tree_parent_check check = { 0 };
3435 struct extent_buffer *eb;
3437 eb = read_tree_block(ctx.fs_info, ref_node->val, &check);
3442 ret = delete_v1_space_cache(eb, rc->block_group,
3443 extent_key->objectid);
3444 free_extent_buffer(eb);
3447 ret = __add_tree_block(rc, ref_node->val, blocksize, blocks);
3452 free_block_list(blocks);
3453 ulist_free(ctx.refs);
3458 * helper to find next unprocessed extent
3460 static noinline_for_stack
3461 int find_next_extent(struct reloc_control *rc, struct btrfs_path *path,
3462 struct btrfs_key *extent_key)
3464 struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
3465 struct btrfs_key key;
3466 struct extent_buffer *leaf;
3467 u64 start, end, last;
3470 last = rc->block_group->start + rc->block_group->length;
3473 if (rc->search_start >= last) {
3478 key.objectid = rc->search_start;
3479 key.type = BTRFS_EXTENT_ITEM_KEY;
3482 path->search_commit_root = 1;
3483 path->skip_locking = 1;
3484 ret = btrfs_search_slot(NULL, rc->extent_root, &key, path,
3489 leaf = path->nodes[0];
3490 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
3491 ret = btrfs_next_leaf(rc->extent_root, path);
3494 leaf = path->nodes[0];
3497 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
3498 if (key.objectid >= last) {
3503 if (key.type != BTRFS_EXTENT_ITEM_KEY &&
3504 key.type != BTRFS_METADATA_ITEM_KEY) {
3509 if (key.type == BTRFS_EXTENT_ITEM_KEY &&
3510 key.objectid + key.offset <= rc->search_start) {
3515 if (key.type == BTRFS_METADATA_ITEM_KEY &&
3516 key.objectid + fs_info->nodesize <=
3522 ret = find_first_extent_bit(&rc->processed_blocks,
3523 key.objectid, &start, &end,
3524 EXTENT_DIRTY, NULL);
3526 if (ret == 0 && start <= key.objectid) {
3527 btrfs_release_path(path);
3528 rc->search_start = end + 1;
3530 if (key.type == BTRFS_EXTENT_ITEM_KEY)
3531 rc->search_start = key.objectid + key.offset;
3533 rc->search_start = key.objectid +
3535 memcpy(extent_key, &key, sizeof(key));
3539 btrfs_release_path(path);
3543 static void set_reloc_control(struct reloc_control *rc)
3545 struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
3547 mutex_lock(&fs_info->reloc_mutex);
3548 fs_info->reloc_ctl = rc;
3549 mutex_unlock(&fs_info->reloc_mutex);
3552 static void unset_reloc_control(struct reloc_control *rc)
3554 struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
3556 mutex_lock(&fs_info->reloc_mutex);
3557 fs_info->reloc_ctl = NULL;
3558 mutex_unlock(&fs_info->reloc_mutex);
3561 static noinline_for_stack
3562 int prepare_to_relocate(struct reloc_control *rc)
3564 struct btrfs_trans_handle *trans;
3567 rc->block_rsv = btrfs_alloc_block_rsv(rc->extent_root->fs_info,
3568 BTRFS_BLOCK_RSV_TEMP);
3572 memset(&rc->cluster, 0, sizeof(rc->cluster));
3573 rc->search_start = rc->block_group->start;
3574 rc->extents_found = 0;
3575 rc->nodes_relocated = 0;
3576 rc->merging_rsv_size = 0;
3577 rc->reserved_bytes = 0;
3578 rc->block_rsv->size = rc->extent_root->fs_info->nodesize *
3579 RELOCATION_RESERVED_NODES;
3580 ret = btrfs_block_rsv_refill(rc->extent_root->fs_info,
3581 rc->block_rsv, rc->block_rsv->size,
3582 BTRFS_RESERVE_FLUSH_ALL);
3586 rc->create_reloc_tree = 1;
3587 set_reloc_control(rc);
3589 trans = btrfs_join_transaction(rc->extent_root);
3590 if (IS_ERR(trans)) {
3591 unset_reloc_control(rc);
3593 * extent tree is not a ref_cow tree and has no reloc_root to
3594 * cleanup. And callers are responsible to free the above
3597 return PTR_ERR(trans);
3600 ret = btrfs_commit_transaction(trans);
3602 unset_reloc_control(rc);
3607 static noinline_for_stack int relocate_block_group(struct reloc_control *rc)
3609 struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
3610 struct rb_root blocks = RB_ROOT;
3611 struct btrfs_key key;
3612 struct btrfs_trans_handle *trans = NULL;
3613 struct btrfs_path *path;
3614 struct btrfs_extent_item *ei;
3620 path = btrfs_alloc_path();
3623 path->reada = READA_FORWARD;
3625 ret = prepare_to_relocate(rc);
3632 rc->reserved_bytes = 0;
3633 ret = btrfs_block_rsv_refill(fs_info, rc->block_rsv,
3634 rc->block_rsv->size,
3635 BTRFS_RESERVE_FLUSH_ALL);
3641 trans = btrfs_start_transaction(rc->extent_root, 0);
3642 if (IS_ERR(trans)) {
3643 err = PTR_ERR(trans);
3648 if (update_backref_cache(trans, &rc->backref_cache)) {
3649 btrfs_end_transaction(trans);
3654 ret = find_next_extent(rc, path, &key);
3660 rc->extents_found++;
3662 ei = btrfs_item_ptr(path->nodes[0], path->slots[0],
3663 struct btrfs_extent_item);
3664 flags = btrfs_extent_flags(path->nodes[0], ei);
3666 if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
3667 ret = add_tree_block(rc, &key, path, &blocks);
3668 } else if (rc->stage == UPDATE_DATA_PTRS &&
3669 (flags & BTRFS_EXTENT_FLAG_DATA)) {
3670 ret = add_data_references(rc, &key, path, &blocks);
3672 btrfs_release_path(path);
3680 if (!RB_EMPTY_ROOT(&blocks)) {
3681 ret = relocate_tree_blocks(trans, rc, &blocks);
3683 if (ret != -EAGAIN) {
3687 rc->extents_found--;
3688 rc->search_start = key.objectid;
3692 btrfs_end_transaction_throttle(trans);
3693 btrfs_btree_balance_dirty(fs_info);
3696 if (rc->stage == MOVE_DATA_EXTENTS &&
3697 (flags & BTRFS_EXTENT_FLAG_DATA)) {
3698 rc->found_file_extent = 1;
3699 ret = relocate_data_extent(rc->data_inode,
3700 &key, &rc->cluster);
3706 if (btrfs_should_cancel_balance(fs_info)) {
3711 if (trans && progress && err == -ENOSPC) {
3712 ret = btrfs_force_chunk_alloc(trans, rc->block_group->flags);
3720 btrfs_release_path(path);
3721 clear_extent_bits(&rc->processed_blocks, 0, (u64)-1, EXTENT_DIRTY);
3724 btrfs_end_transaction_throttle(trans);
3725 btrfs_btree_balance_dirty(fs_info);
3729 ret = relocate_file_extent_cluster(rc->data_inode,
3735 rc->create_reloc_tree = 0;
3736 set_reloc_control(rc);
3738 btrfs_backref_release_cache(&rc->backref_cache);
3739 btrfs_block_rsv_release(fs_info, rc->block_rsv, (u64)-1, NULL);
3742 * Even in the case when the relocation is cancelled, we should all go
3743 * through prepare_to_merge() and merge_reloc_roots().
3745 * For error (including cancelled balance), prepare_to_merge() will
3746 * mark all reloc trees orphan, then queue them for cleanup in
3747 * merge_reloc_roots()
3749 err = prepare_to_merge(rc, err);
3751 merge_reloc_roots(rc);
3753 rc->merge_reloc_tree = 0;
3754 unset_reloc_control(rc);
3755 btrfs_block_rsv_release(fs_info, rc->block_rsv, (u64)-1, NULL);
3757 /* get rid of pinned extents */
3758 trans = btrfs_join_transaction(rc->extent_root);
3759 if (IS_ERR(trans)) {
3760 err = PTR_ERR(trans);
3763 ret = btrfs_commit_transaction(trans);
3767 ret = clean_dirty_subvols(rc);
3768 if (ret < 0 && !err)
3770 btrfs_free_block_rsv(fs_info, rc->block_rsv);
3771 btrfs_free_path(path);
3775 static int __insert_orphan_inode(struct btrfs_trans_handle *trans,
3776 struct btrfs_root *root, u64 objectid)
3778 struct btrfs_path *path;
3779 struct btrfs_inode_item *item;
3780 struct extent_buffer *leaf;
3783 path = btrfs_alloc_path();
3787 ret = btrfs_insert_empty_inode(trans, root, path, objectid);
3791 leaf = path->nodes[0];
3792 item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_inode_item);
3793 memzero_extent_buffer(leaf, (unsigned long)item, sizeof(*item));
3794 btrfs_set_inode_generation(leaf, item, 1);
3795 btrfs_set_inode_size(leaf, item, 0);
3796 btrfs_set_inode_mode(leaf, item, S_IFREG | 0600);
3797 btrfs_set_inode_flags(leaf, item, BTRFS_INODE_NOCOMPRESS |
3798 BTRFS_INODE_PREALLOC);
3799 btrfs_mark_buffer_dirty(leaf);
3801 btrfs_free_path(path);
3805 static void delete_orphan_inode(struct btrfs_trans_handle *trans,
3806 struct btrfs_root *root, u64 objectid)
3808 struct btrfs_path *path;
3809 struct btrfs_key key;
3812 path = btrfs_alloc_path();
3818 key.objectid = objectid;
3819 key.type = BTRFS_INODE_ITEM_KEY;
3821 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
3827 ret = btrfs_del_item(trans, root, path);
3830 btrfs_abort_transaction(trans, ret);
3831 btrfs_free_path(path);
3835 * helper to create inode for data relocation.
3836 * the inode is in data relocation tree and its link count is 0
3838 static noinline_for_stack
3839 struct inode *create_reloc_inode(struct btrfs_fs_info *fs_info,
3840 struct btrfs_block_group *group)
3842 struct inode *inode = NULL;
3843 struct btrfs_trans_handle *trans;
3844 struct btrfs_root *root;
3848 root = btrfs_grab_root(fs_info->data_reloc_root);
3849 trans = btrfs_start_transaction(root, 6);
3850 if (IS_ERR(trans)) {
3851 btrfs_put_root(root);
3852 return ERR_CAST(trans);
3855 err = btrfs_get_free_objectid(root, &objectid);
3859 err = __insert_orphan_inode(trans, root, objectid);
3863 inode = btrfs_iget(fs_info->sb, objectid, root);
3864 if (IS_ERR(inode)) {
3865 delete_orphan_inode(trans, root, objectid);
3866 err = PTR_ERR(inode);
3870 BTRFS_I(inode)->index_cnt = group->start;
3872 err = btrfs_orphan_add(trans, BTRFS_I(inode));
3874 btrfs_put_root(root);
3875 btrfs_end_transaction(trans);
3876 btrfs_btree_balance_dirty(fs_info);
3879 inode = ERR_PTR(err);
3885 * Mark start of chunk relocation that is cancellable. Check if the cancellation
3886 * has been requested meanwhile and don't start in that case.
3890 * -EINPROGRESS operation is already in progress, that's probably a bug
3891 * -ECANCELED cancellation request was set before the operation started
3893 static int reloc_chunk_start(struct btrfs_fs_info *fs_info)
3895 if (test_and_set_bit(BTRFS_FS_RELOC_RUNNING, &fs_info->flags)) {
3896 /* This should not happen */
3897 btrfs_err(fs_info, "reloc already running, cannot start");
3898 return -EINPROGRESS;
3901 if (atomic_read(&fs_info->reloc_cancel_req) > 0) {
3902 btrfs_info(fs_info, "chunk relocation canceled on start");
3904 * On cancel, clear all requests but let the caller mark
3905 * the end after cleanup operations.
3907 atomic_set(&fs_info->reloc_cancel_req, 0);
3914 * Mark end of chunk relocation that is cancellable and wake any waiters.
3916 static void reloc_chunk_end(struct btrfs_fs_info *fs_info)
3918 /* Requested after start, clear bit first so any waiters can continue */
3919 if (atomic_read(&fs_info->reloc_cancel_req) > 0)
3920 btrfs_info(fs_info, "chunk relocation canceled during operation");
3921 clear_and_wake_up_bit(BTRFS_FS_RELOC_RUNNING, &fs_info->flags);
3922 atomic_set(&fs_info->reloc_cancel_req, 0);
3925 static struct reloc_control *alloc_reloc_control(struct btrfs_fs_info *fs_info)
3927 struct reloc_control *rc;
3929 rc = kzalloc(sizeof(*rc), GFP_NOFS);
3933 INIT_LIST_HEAD(&rc->reloc_roots);
3934 INIT_LIST_HEAD(&rc->dirty_subvol_roots);
3935 btrfs_backref_init_cache(fs_info, &rc->backref_cache, 1);
3936 mapping_tree_init(&rc->reloc_root_tree);
3937 extent_io_tree_init(fs_info, &rc->processed_blocks, IO_TREE_RELOC_BLOCKS);
3941 static void free_reloc_control(struct reloc_control *rc)
3943 struct mapping_node *node, *tmp;
3945 free_reloc_roots(&rc->reloc_roots);
3946 rbtree_postorder_for_each_entry_safe(node, tmp,
3947 &rc->reloc_root_tree.rb_root, rb_node)
3954 * Print the block group being relocated
3956 static void describe_relocation(struct btrfs_fs_info *fs_info,
3957 struct btrfs_block_group *block_group)
3959 char buf[128] = {'\0'};
3961 btrfs_describe_block_groups(block_group->flags, buf, sizeof(buf));
3964 "relocating block group %llu flags %s",
3965 block_group->start, buf);
3968 static const char *stage_to_string(int stage)
3970 if (stage == MOVE_DATA_EXTENTS)
3971 return "move data extents";
3972 if (stage == UPDATE_DATA_PTRS)
3973 return "update data pointers";
3978 * function to relocate all extents in a block group.
3980 int btrfs_relocate_block_group(struct btrfs_fs_info *fs_info, u64 group_start)
3982 struct btrfs_block_group *bg;
3983 struct btrfs_root *extent_root = btrfs_extent_root(fs_info, group_start);
3984 struct reloc_control *rc;
3985 struct inode *inode;
3986 struct btrfs_path *path;
3992 * This only gets set if we had a half-deleted snapshot on mount. We
3993 * cannot allow relocation to start while we're still trying to clean up
3994 * these pending deletions.
3996 ret = wait_on_bit(&fs_info->flags, BTRFS_FS_UNFINISHED_DROPS, TASK_INTERRUPTIBLE);
4000 /* We may have been woken up by close_ctree, so bail if we're closing. */
4001 if (btrfs_fs_closing(fs_info))
4004 bg = btrfs_lookup_block_group(fs_info, group_start);
4009 * Relocation of a data block group creates ordered extents. Without
4010 * sb_start_write(), we can freeze the filesystem while unfinished
4011 * ordered extents are left. Such ordered extents can cause a deadlock
4012 * e.g. when syncfs() is waiting for their completion but they can't
4013 * finish because they block when joining a transaction, due to the
4014 * fact that the freeze locks are being held in write mode.
4016 if (bg->flags & BTRFS_BLOCK_GROUP_DATA)
4017 ASSERT(sb_write_started(fs_info->sb));
4019 if (btrfs_pinned_by_swapfile(fs_info, bg)) {
4020 btrfs_put_block_group(bg);
4024 rc = alloc_reloc_control(fs_info);
4026 btrfs_put_block_group(bg);
4030 ret = reloc_chunk_start(fs_info);
4036 rc->extent_root = extent_root;
4037 rc->block_group = bg;
4039 ret = btrfs_inc_block_group_ro(rc->block_group, true);
4046 path = btrfs_alloc_path();
4052 inode = lookup_free_space_inode(rc->block_group, path);
4053 btrfs_free_path(path);
4056 ret = delete_block_group_cache(fs_info, rc->block_group, inode, 0);
4058 ret = PTR_ERR(inode);
4060 if (ret && ret != -ENOENT) {
4065 rc->data_inode = create_reloc_inode(fs_info, rc->block_group);
4066 if (IS_ERR(rc->data_inode)) {
4067 err = PTR_ERR(rc->data_inode);
4068 rc->data_inode = NULL;
4072 describe_relocation(fs_info, rc->block_group);
4074 btrfs_wait_block_group_reservations(rc->block_group);
4075 btrfs_wait_nocow_writers(rc->block_group);
4076 btrfs_wait_ordered_roots(fs_info, U64_MAX,
4077 rc->block_group->start,
4078 rc->block_group->length);
4080 ret = btrfs_zone_finish(rc->block_group);
4081 WARN_ON(ret && ret != -EAGAIN);
4086 mutex_lock(&fs_info->cleaner_mutex);
4087 ret = relocate_block_group(rc);
4088 mutex_unlock(&fs_info->cleaner_mutex);
4092 finishes_stage = rc->stage;
4094 * We may have gotten ENOSPC after we already dirtied some
4095 * extents. If writeout happens while we're relocating a
4096 * different block group we could end up hitting the
4097 * BUG_ON(rc->stage == UPDATE_DATA_PTRS) in
4098 * btrfs_reloc_cow_block. Make sure we write everything out
4099 * properly so we don't trip over this problem, and then break
4100 * out of the loop if we hit an error.
4102 if (rc->stage == MOVE_DATA_EXTENTS && rc->found_file_extent) {
4103 ret = btrfs_wait_ordered_range(rc->data_inode, 0,
4107 invalidate_mapping_pages(rc->data_inode->i_mapping,
4109 rc->stage = UPDATE_DATA_PTRS;
4115 if (rc->extents_found == 0)
4118 btrfs_info(fs_info, "found %llu extents, stage: %s",
4119 rc->extents_found, stage_to_string(finishes_stage));
4122 WARN_ON(rc->block_group->pinned > 0);
4123 WARN_ON(rc->block_group->reserved > 0);
4124 WARN_ON(rc->block_group->used > 0);
4127 btrfs_dec_block_group_ro(rc->block_group);
4128 iput(rc->data_inode);
4130 btrfs_put_block_group(bg);
4131 reloc_chunk_end(fs_info);
4132 free_reloc_control(rc);
4136 static noinline_for_stack int mark_garbage_root(struct btrfs_root *root)
4138 struct btrfs_fs_info *fs_info = root->fs_info;
4139 struct btrfs_trans_handle *trans;
4142 trans = btrfs_start_transaction(fs_info->tree_root, 0);
4144 return PTR_ERR(trans);
4146 memset(&root->root_item.drop_progress, 0,
4147 sizeof(root->root_item.drop_progress));
4148 btrfs_set_root_drop_level(&root->root_item, 0);
4149 btrfs_set_root_refs(&root->root_item, 0);
4150 ret = btrfs_update_root(trans, fs_info->tree_root,
4151 &root->root_key, &root->root_item);
4153 err = btrfs_end_transaction(trans);
4160 * recover relocation interrupted by system crash.
4162 * this function resumes merging reloc trees with corresponding fs trees.
4163 * this is important for keeping the sharing of tree blocks
4165 int btrfs_recover_relocation(struct btrfs_fs_info *fs_info)
4167 LIST_HEAD(reloc_roots);
4168 struct btrfs_key key;
4169 struct btrfs_root *fs_root;
4170 struct btrfs_root *reloc_root;
4171 struct btrfs_path *path;
4172 struct extent_buffer *leaf;
4173 struct reloc_control *rc = NULL;
4174 struct btrfs_trans_handle *trans;
4178 path = btrfs_alloc_path();
4181 path->reada = READA_BACK;
4183 key.objectid = BTRFS_TREE_RELOC_OBJECTID;
4184 key.type = BTRFS_ROOT_ITEM_KEY;
4185 key.offset = (u64)-1;
4188 ret = btrfs_search_slot(NULL, fs_info->tree_root, &key,
4195 if (path->slots[0] == 0)
4199 leaf = path->nodes[0];
4200 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
4201 btrfs_release_path(path);
4203 if (key.objectid != BTRFS_TREE_RELOC_OBJECTID ||
4204 key.type != BTRFS_ROOT_ITEM_KEY)
4207 reloc_root = btrfs_read_tree_root(fs_info->tree_root, &key);
4208 if (IS_ERR(reloc_root)) {
4209 err = PTR_ERR(reloc_root);
4213 set_bit(BTRFS_ROOT_SHAREABLE, &reloc_root->state);
4214 list_add(&reloc_root->root_list, &reloc_roots);
4216 if (btrfs_root_refs(&reloc_root->root_item) > 0) {
4217 fs_root = btrfs_get_fs_root(fs_info,
4218 reloc_root->root_key.offset, false);
4219 if (IS_ERR(fs_root)) {
4220 ret = PTR_ERR(fs_root);
4221 if (ret != -ENOENT) {
4225 ret = mark_garbage_root(reloc_root);
4231 btrfs_put_root(fs_root);
4235 if (key.offset == 0)
4240 btrfs_release_path(path);
4242 if (list_empty(&reloc_roots))
4245 rc = alloc_reloc_control(fs_info);
4251 ret = reloc_chunk_start(fs_info);
4257 rc->extent_root = btrfs_extent_root(fs_info, 0);
4259 set_reloc_control(rc);
4261 trans = btrfs_join_transaction(rc->extent_root);
4262 if (IS_ERR(trans)) {
4263 err = PTR_ERR(trans);
4267 rc->merge_reloc_tree = 1;
4269 while (!list_empty(&reloc_roots)) {
4270 reloc_root = list_entry(reloc_roots.next,
4271 struct btrfs_root, root_list);
4272 list_del(&reloc_root->root_list);
4274 if (btrfs_root_refs(&reloc_root->root_item) == 0) {
4275 list_add_tail(&reloc_root->root_list,
4280 fs_root = btrfs_get_fs_root(fs_info, reloc_root->root_key.offset,
4282 if (IS_ERR(fs_root)) {
4283 err = PTR_ERR(fs_root);
4284 list_add_tail(&reloc_root->root_list, &reloc_roots);
4285 btrfs_end_transaction(trans);
4289 err = __add_reloc_root(reloc_root);
4290 ASSERT(err != -EEXIST);
4292 list_add_tail(&reloc_root->root_list, &reloc_roots);
4293 btrfs_put_root(fs_root);
4294 btrfs_end_transaction(trans);
4297 fs_root->reloc_root = btrfs_grab_root(reloc_root);
4298 btrfs_put_root(fs_root);
4301 err = btrfs_commit_transaction(trans);
4305 merge_reloc_roots(rc);
4307 unset_reloc_control(rc);
4309 trans = btrfs_join_transaction(rc->extent_root);
4310 if (IS_ERR(trans)) {
4311 err = PTR_ERR(trans);
4314 err = btrfs_commit_transaction(trans);
4316 ret = clean_dirty_subvols(rc);
4317 if (ret < 0 && !err)
4320 unset_reloc_control(rc);
4322 reloc_chunk_end(fs_info);
4323 free_reloc_control(rc);
4325 free_reloc_roots(&reloc_roots);
4327 btrfs_free_path(path);
4330 /* cleanup orphan inode in data relocation tree */
4331 fs_root = btrfs_grab_root(fs_info->data_reloc_root);
4333 err = btrfs_orphan_cleanup(fs_root);
4334 btrfs_put_root(fs_root);
4340 * helper to add ordered checksum for data relocation.
4342 * cloning checksum properly handles the nodatasum extents.
4343 * it also saves CPU time to re-calculate the checksum.
4345 int btrfs_reloc_clone_csums(struct btrfs_inode *inode, u64 file_pos, u64 len)
4347 struct btrfs_fs_info *fs_info = inode->root->fs_info;
4348 struct btrfs_root *csum_root;
4349 struct btrfs_ordered_sum *sums;
4350 struct btrfs_ordered_extent *ordered;
4356 ordered = btrfs_lookup_ordered_extent(inode, file_pos);
4357 BUG_ON(ordered->file_offset != file_pos || ordered->num_bytes != len);
4359 disk_bytenr = file_pos + inode->index_cnt;
4360 csum_root = btrfs_csum_root(fs_info, disk_bytenr);
4361 ret = btrfs_lookup_csums_list(csum_root, disk_bytenr,
4362 disk_bytenr + len - 1, &list, 0, false);
4366 while (!list_empty(&list)) {
4367 sums = list_entry(list.next, struct btrfs_ordered_sum, list);
4368 list_del_init(&sums->list);
4371 * We need to offset the new_bytenr based on where the csum is.
4372 * We need to do this because we will read in entire prealloc
4373 * extents but we may have written to say the middle of the
4374 * prealloc extent, so we need to make sure the csum goes with
4375 * the right disk offset.
4377 * We can do this because the data reloc inode refers strictly
4378 * to the on disk bytes, so we don't have to worry about
4379 * disk_len vs real len like with real inodes since it's all
4382 new_bytenr = ordered->disk_bytenr + sums->bytenr - disk_bytenr;
4383 sums->bytenr = new_bytenr;
4385 btrfs_add_ordered_sum(ordered, sums);
4388 btrfs_put_ordered_extent(ordered);
4392 int btrfs_reloc_cow_block(struct btrfs_trans_handle *trans,
4393 struct btrfs_root *root, struct extent_buffer *buf,
4394 struct extent_buffer *cow)
4396 struct btrfs_fs_info *fs_info = root->fs_info;
4397 struct reloc_control *rc;
4398 struct btrfs_backref_node *node;
4403 rc = fs_info->reloc_ctl;
4407 BUG_ON(rc->stage == UPDATE_DATA_PTRS && btrfs_is_data_reloc_root(root));
4409 level = btrfs_header_level(buf);
4410 if (btrfs_header_generation(buf) <=
4411 btrfs_root_last_snapshot(&root->root_item))
4414 if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID &&
4415 rc->create_reloc_tree) {
4416 WARN_ON(!first_cow && level == 0);
4418 node = rc->backref_cache.path[level];
4419 BUG_ON(node->bytenr != buf->start &&
4420 node->new_bytenr != buf->start);
4422 btrfs_backref_drop_node_buffer(node);
4423 atomic_inc(&cow->refs);
4425 node->new_bytenr = cow->start;
4427 if (!node->pending) {
4428 list_move_tail(&node->list,
4429 &rc->backref_cache.pending[level]);
4434 mark_block_processed(rc, node);
4436 if (first_cow && level > 0)
4437 rc->nodes_relocated += buf->len;
4440 if (level == 0 && first_cow && rc->stage == UPDATE_DATA_PTRS)
4441 ret = replace_file_extents(trans, rc, root, cow);
4446 * called before creating snapshot. it calculates metadata reservation
4447 * required for relocating tree blocks in the snapshot
4449 void btrfs_reloc_pre_snapshot(struct btrfs_pending_snapshot *pending,
4450 u64 *bytes_to_reserve)
4452 struct btrfs_root *root = pending->root;
4453 struct reloc_control *rc = root->fs_info->reloc_ctl;
4455 if (!rc || !have_reloc_root(root))
4458 if (!rc->merge_reloc_tree)
4461 root = root->reloc_root;
4462 BUG_ON(btrfs_root_refs(&root->root_item) == 0);
4464 * relocation is in the stage of merging trees. the space
4465 * used by merging a reloc tree is twice the size of
4466 * relocated tree nodes in the worst case. half for cowing
4467 * the reloc tree, half for cowing the fs tree. the space
4468 * used by cowing the reloc tree will be freed after the
4469 * tree is dropped. if we create snapshot, cowing the fs
4470 * tree may use more space than it frees. so we need
4471 * reserve extra space.
4473 *bytes_to_reserve += rc->nodes_relocated;
4477 * called after snapshot is created. migrate block reservation
4478 * and create reloc root for the newly created snapshot
4480 * This is similar to btrfs_init_reloc_root(), we come out of here with two
4481 * references held on the reloc_root, one for root->reloc_root and one for
4484 int btrfs_reloc_post_snapshot(struct btrfs_trans_handle *trans,
4485 struct btrfs_pending_snapshot *pending)
4487 struct btrfs_root *root = pending->root;
4488 struct btrfs_root *reloc_root;
4489 struct btrfs_root *new_root;
4490 struct reloc_control *rc = root->fs_info->reloc_ctl;
4493 if (!rc || !have_reloc_root(root))
4496 rc = root->fs_info->reloc_ctl;
4497 rc->merging_rsv_size += rc->nodes_relocated;
4499 if (rc->merge_reloc_tree) {
4500 ret = btrfs_block_rsv_migrate(&pending->block_rsv,
4502 rc->nodes_relocated, true);
4507 new_root = pending->snap;
4508 reloc_root = create_reloc_root(trans, root->reloc_root,
4509 new_root->root_key.objectid);
4510 if (IS_ERR(reloc_root))
4511 return PTR_ERR(reloc_root);
4513 ret = __add_reloc_root(reloc_root);
4514 ASSERT(ret != -EEXIST);
4516 /* Pairs with create_reloc_root */
4517 btrfs_put_root(reloc_root);
4520 new_root->reloc_root = btrfs_grab_root(reloc_root);
4522 if (rc->create_reloc_tree)
4523 ret = clone_backref_node(trans, rc, root, reloc_root);