1 // SPDX-License-Identifier: GPL-2.0
3 * Copyright (C) 2014 Facebook. All rights reserved.
6 #include <linux/sched.h>
7 #include <linux/stacktrace.h>
12 #include "delayed-ref.h"
13 #include "ref-verify.h"
15 #include "accessors.h"
18 * Used to keep track the roots and number of refs each root has for a given
19 * bytenr. This just tracks the number of direct references, no shared
29 * These are meant to represent what should exist in the extent tree, these can
30 * be used to verify the extent tree is consistent as these should all match
31 * what the extent tree says.
45 * Whenever we add/remove a reference we record the action. The action maps
46 * back to the delayed ref action. We hold the ref we are changing in the
47 * action so we can account for the history properly, and we record the root we
48 * were called with since it could be different from ref_root. We also store
49 * stack traces because that's how I roll.
55 struct list_head list;
56 unsigned long trace[MAX_TRACE];
57 unsigned int trace_len;
61 * One of these for every block we reference, it holds the roots and references
62 * to it as well as all of the ref actions that have occurred to it. We never
63 * free it until we unmount the file system in order to make sure re-allocations
64 * are happening properly.
75 struct list_head actions;
78 static struct block_entry *insert_block_entry(struct rb_root *root,
79 struct block_entry *be)
81 struct rb_node **p = &root->rb_node;
82 struct rb_node *parent_node = NULL;
83 struct block_entry *entry;
87 entry = rb_entry(parent_node, struct block_entry, node);
88 if (entry->bytenr > be->bytenr)
90 else if (entry->bytenr < be->bytenr)
96 rb_link_node(&be->node, parent_node, p);
97 rb_insert_color(&be->node, root);
101 static struct block_entry *lookup_block_entry(struct rb_root *root, u64 bytenr)
104 struct block_entry *entry = NULL;
108 entry = rb_entry(n, struct block_entry, node);
109 if (entry->bytenr < bytenr)
111 else if (entry->bytenr > bytenr)
119 static struct root_entry *insert_root_entry(struct rb_root *root,
120 struct root_entry *re)
122 struct rb_node **p = &root->rb_node;
123 struct rb_node *parent_node = NULL;
124 struct root_entry *entry;
128 entry = rb_entry(parent_node, struct root_entry, node);
129 if (entry->root_objectid > re->root_objectid)
131 else if (entry->root_objectid < re->root_objectid)
137 rb_link_node(&re->node, parent_node, p);
138 rb_insert_color(&re->node, root);
143 static int comp_refs(struct ref_entry *ref1, struct ref_entry *ref2)
145 if (ref1->root_objectid < ref2->root_objectid)
147 if (ref1->root_objectid > ref2->root_objectid)
149 if (ref1->parent < ref2->parent)
151 if (ref1->parent > ref2->parent)
153 if (ref1->owner < ref2->owner)
155 if (ref1->owner > ref2->owner)
157 if (ref1->offset < ref2->offset)
159 if (ref1->offset > ref2->offset)
164 static struct ref_entry *insert_ref_entry(struct rb_root *root,
165 struct ref_entry *ref)
167 struct rb_node **p = &root->rb_node;
168 struct rb_node *parent_node = NULL;
169 struct ref_entry *entry;
174 entry = rb_entry(parent_node, struct ref_entry, node);
175 cmp = comp_refs(entry, ref);
184 rb_link_node(&ref->node, parent_node, p);
185 rb_insert_color(&ref->node, root);
190 static struct root_entry *lookup_root_entry(struct rb_root *root, u64 objectid)
193 struct root_entry *entry = NULL;
197 entry = rb_entry(n, struct root_entry, node);
198 if (entry->root_objectid < objectid)
200 else if (entry->root_objectid > objectid)
208 #ifdef CONFIG_STACKTRACE
209 static void __save_stack_trace(struct ref_action *ra)
211 ra->trace_len = stack_trace_save(ra->trace, MAX_TRACE, 2);
214 static void __print_stack_trace(struct btrfs_fs_info *fs_info,
215 struct ref_action *ra)
217 if (ra->trace_len == 0) {
218 btrfs_err(fs_info, " ref-verify: no stacktrace");
221 stack_trace_print(ra->trace, ra->trace_len, 2);
224 static inline void __save_stack_trace(struct ref_action *ra)
228 static inline void __print_stack_trace(struct btrfs_fs_info *fs_info,
229 struct ref_action *ra)
231 btrfs_err(fs_info, " ref-verify: no stacktrace support");
235 static void free_block_entry(struct block_entry *be)
237 struct root_entry *re;
238 struct ref_entry *ref;
239 struct ref_action *ra;
242 while ((n = rb_first(&be->roots))) {
243 re = rb_entry(n, struct root_entry, node);
244 rb_erase(&re->node, &be->roots);
248 while((n = rb_first(&be->refs))) {
249 ref = rb_entry(n, struct ref_entry, node);
250 rb_erase(&ref->node, &be->refs);
254 while (!list_empty(&be->actions)) {
255 ra = list_first_entry(&be->actions, struct ref_action,
263 static struct block_entry *add_block_entry(struct btrfs_fs_info *fs_info,
267 struct block_entry *be = NULL, *exist;
268 struct root_entry *re = NULL;
270 re = kzalloc(sizeof(struct root_entry), GFP_NOFS);
271 be = kzalloc(sizeof(struct block_entry), GFP_NOFS);
275 return ERR_PTR(-ENOMEM);
280 re->root_objectid = root_objectid;
283 spin_lock(&fs_info->ref_verify_lock);
284 exist = insert_block_entry(&fs_info->block_tree, be);
287 struct root_entry *exist_re;
289 exist_re = insert_root_entry(&exist->roots, re);
304 INIT_LIST_HEAD(&be->actions);
306 insert_root_entry(&be->roots, re);
312 static int add_tree_block(struct btrfs_fs_info *fs_info, u64 ref_root,
313 u64 parent, u64 bytenr, int level)
315 struct block_entry *be;
316 struct root_entry *re;
317 struct ref_entry *ref = NULL, *exist;
319 ref = kmalloc(sizeof(struct ref_entry), GFP_NOFS);
324 ref->root_objectid = 0;
326 ref->root_objectid = ref_root;
327 ref->parent = parent;
332 be = add_block_entry(fs_info, bytenr, fs_info->nodesize, ref_root);
343 re = lookup_root_entry(&be->roots, ref_root);
347 exist = insert_ref_entry(&be->refs, ref);
352 spin_unlock(&fs_info->ref_verify_lock);
357 static int add_shared_data_ref(struct btrfs_fs_info *fs_info,
358 u64 parent, u32 num_refs, u64 bytenr,
361 struct block_entry *be;
362 struct ref_entry *ref;
364 ref = kzalloc(sizeof(struct ref_entry), GFP_NOFS);
367 be = add_block_entry(fs_info, bytenr, num_bytes, 0);
372 be->num_refs += num_refs;
374 ref->parent = parent;
375 ref->num_refs = num_refs;
376 if (insert_ref_entry(&be->refs, ref)) {
377 spin_unlock(&fs_info->ref_verify_lock);
378 btrfs_err(fs_info, "existing shared ref when reading from disk?");
382 spin_unlock(&fs_info->ref_verify_lock);
386 static int add_extent_data_ref(struct btrfs_fs_info *fs_info,
387 struct extent_buffer *leaf,
388 struct btrfs_extent_data_ref *dref,
389 u64 bytenr, u64 num_bytes)
391 struct block_entry *be;
392 struct ref_entry *ref;
393 struct root_entry *re;
394 u64 ref_root = btrfs_extent_data_ref_root(leaf, dref);
395 u64 owner = btrfs_extent_data_ref_objectid(leaf, dref);
396 u64 offset = btrfs_extent_data_ref_offset(leaf, dref);
397 u32 num_refs = btrfs_extent_data_ref_count(leaf, dref);
399 ref = kzalloc(sizeof(struct ref_entry), GFP_NOFS);
402 be = add_block_entry(fs_info, bytenr, num_bytes, ref_root);
407 be->num_refs += num_refs;
411 ref->root_objectid = ref_root;
412 ref->offset = offset;
413 ref->num_refs = num_refs;
414 if (insert_ref_entry(&be->refs, ref)) {
415 spin_unlock(&fs_info->ref_verify_lock);
416 btrfs_err(fs_info, "existing ref when reading from disk?");
421 re = lookup_root_entry(&be->roots, ref_root);
423 spin_unlock(&fs_info->ref_verify_lock);
424 btrfs_err(fs_info, "missing root in new block entry?");
427 re->num_refs += num_refs;
428 spin_unlock(&fs_info->ref_verify_lock);
432 static int process_extent_item(struct btrfs_fs_info *fs_info,
433 struct btrfs_path *path, struct btrfs_key *key,
434 int slot, int *tree_block_level)
436 struct btrfs_extent_item *ei;
437 struct btrfs_extent_inline_ref *iref;
438 struct btrfs_extent_data_ref *dref;
439 struct btrfs_shared_data_ref *sref;
440 struct extent_buffer *leaf = path->nodes[0];
441 u32 item_size = btrfs_item_size(leaf, slot);
442 unsigned long end, ptr;
443 u64 offset, flags, count;
446 ei = btrfs_item_ptr(leaf, slot, struct btrfs_extent_item);
447 flags = btrfs_extent_flags(leaf, ei);
449 if ((key->type == BTRFS_EXTENT_ITEM_KEY) &&
450 flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
451 struct btrfs_tree_block_info *info;
453 info = (struct btrfs_tree_block_info *)(ei + 1);
454 *tree_block_level = btrfs_tree_block_level(leaf, info);
455 iref = (struct btrfs_extent_inline_ref *)(info + 1);
457 if (key->type == BTRFS_METADATA_ITEM_KEY)
458 *tree_block_level = key->offset;
459 iref = (struct btrfs_extent_inline_ref *)(ei + 1);
462 ptr = (unsigned long)iref;
463 end = (unsigned long)ei + item_size;
465 iref = (struct btrfs_extent_inline_ref *)ptr;
466 type = btrfs_extent_inline_ref_type(leaf, iref);
467 offset = btrfs_extent_inline_ref_offset(leaf, iref);
469 case BTRFS_TREE_BLOCK_REF_KEY:
470 ret = add_tree_block(fs_info, offset, 0, key->objectid,
473 case BTRFS_SHARED_BLOCK_REF_KEY:
474 ret = add_tree_block(fs_info, 0, offset, key->objectid,
477 case BTRFS_EXTENT_DATA_REF_KEY:
478 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
479 ret = add_extent_data_ref(fs_info, leaf, dref,
480 key->objectid, key->offset);
482 case BTRFS_SHARED_DATA_REF_KEY:
483 sref = (struct btrfs_shared_data_ref *)(iref + 1);
484 count = btrfs_shared_data_ref_count(leaf, sref);
485 ret = add_shared_data_ref(fs_info, offset, count,
486 key->objectid, key->offset);
488 case BTRFS_EXTENT_OWNER_REF_KEY:
489 WARN_ON(!btrfs_fs_incompat(fs_info, SIMPLE_QUOTA));
492 btrfs_err(fs_info, "invalid key type in iref");
498 ptr += btrfs_extent_inline_ref_size(type);
503 static int process_leaf(struct btrfs_root *root,
504 struct btrfs_path *path, u64 *bytenr, u64 *num_bytes,
505 int *tree_block_level)
507 struct btrfs_fs_info *fs_info = root->fs_info;
508 struct extent_buffer *leaf = path->nodes[0];
509 struct btrfs_extent_data_ref *dref;
510 struct btrfs_shared_data_ref *sref;
513 struct btrfs_key key;
514 int nritems = btrfs_header_nritems(leaf);
516 for (i = 0; i < nritems; i++) {
517 btrfs_item_key_to_cpu(leaf, &key, i);
519 case BTRFS_EXTENT_ITEM_KEY:
520 *num_bytes = key.offset;
522 case BTRFS_METADATA_ITEM_KEY:
523 *bytenr = key.objectid;
524 ret = process_extent_item(fs_info, path, &key, i,
527 case BTRFS_TREE_BLOCK_REF_KEY:
528 ret = add_tree_block(fs_info, key.offset, 0,
529 key.objectid, *tree_block_level);
531 case BTRFS_SHARED_BLOCK_REF_KEY:
532 ret = add_tree_block(fs_info, 0, key.offset,
533 key.objectid, *tree_block_level);
535 case BTRFS_EXTENT_DATA_REF_KEY:
536 dref = btrfs_item_ptr(leaf, i,
537 struct btrfs_extent_data_ref);
538 ret = add_extent_data_ref(fs_info, leaf, dref, *bytenr,
541 case BTRFS_SHARED_DATA_REF_KEY:
542 sref = btrfs_item_ptr(leaf, i,
543 struct btrfs_shared_data_ref);
544 count = btrfs_shared_data_ref_count(leaf, sref);
545 ret = add_shared_data_ref(fs_info, key.offset, count,
546 *bytenr, *num_bytes);
557 /* Walk down to the leaf from the given level */
558 static int walk_down_tree(struct btrfs_root *root, struct btrfs_path *path,
559 int level, u64 *bytenr, u64 *num_bytes,
560 int *tree_block_level)
562 struct extent_buffer *eb;
567 eb = btrfs_read_node_slot(path->nodes[level],
571 btrfs_tree_read_lock(eb);
572 path->nodes[level-1] = eb;
573 path->slots[level-1] = 0;
574 path->locks[level-1] = BTRFS_READ_LOCK;
576 ret = process_leaf(root, path, bytenr, num_bytes,
586 /* Walk up to the next node that needs to be processed */
587 static int walk_up_tree(struct btrfs_path *path, int *level)
591 for (l = 0; l < BTRFS_MAX_LEVEL; l++) {
597 btrfs_header_nritems(path->nodes[l])) {
602 btrfs_tree_unlock_rw(path->nodes[l], path->locks[l]);
603 free_extent_buffer(path->nodes[l]);
604 path->nodes[l] = NULL;
612 static void dump_ref_action(struct btrfs_fs_info *fs_info,
613 struct ref_action *ra)
616 " Ref action %d, root %llu, ref_root %llu, parent %llu, owner %llu, offset %llu, num_refs %llu",
617 ra->action, ra->root, ra->ref.root_objectid, ra->ref.parent,
618 ra->ref.owner, ra->ref.offset, ra->ref.num_refs);
619 __print_stack_trace(fs_info, ra);
623 * Dumps all the information from the block entry to printk, it's going to be
626 static void dump_block_entry(struct btrfs_fs_info *fs_info,
627 struct block_entry *be)
629 struct ref_entry *ref;
630 struct root_entry *re;
631 struct ref_action *ra;
635 "dumping block entry [%llu %llu], num_refs %llu, metadata %d, from disk %d",
636 be->bytenr, be->len, be->num_refs, be->metadata,
639 for (n = rb_first(&be->refs); n; n = rb_next(n)) {
640 ref = rb_entry(n, struct ref_entry, node);
642 " ref root %llu, parent %llu, owner %llu, offset %llu, num_refs %llu",
643 ref->root_objectid, ref->parent, ref->owner,
644 ref->offset, ref->num_refs);
647 for (n = rb_first(&be->roots); n; n = rb_next(n)) {
648 re = rb_entry(n, struct root_entry, node);
649 btrfs_err(fs_info, " root entry %llu, num_refs %llu",
650 re->root_objectid, re->num_refs);
653 list_for_each_entry(ra, &be->actions, list)
654 dump_ref_action(fs_info, ra);
658 * Called when we modify a ref for a bytenr.
660 * This will add an action item to the given bytenr and do sanity checks to make
661 * sure we haven't messed something up. If we are making a new allocation and
662 * this block entry has history we will delete all previous actions as long as
663 * our sanity checks pass as they are no longer needed.
665 int btrfs_ref_tree_mod(struct btrfs_fs_info *fs_info,
666 struct btrfs_ref *generic_ref)
668 struct ref_entry *ref = NULL, *exist;
669 struct ref_action *ra = NULL;
670 struct block_entry *be = NULL;
671 struct root_entry *re = NULL;
672 int action = generic_ref->action;
675 u64 bytenr = generic_ref->bytenr;
676 u64 num_bytes = generic_ref->len;
677 u64 parent = generic_ref->parent;
682 if (!btrfs_test_opt(fs_info, REF_VERIFY))
685 if (generic_ref->type == BTRFS_REF_METADATA) {
687 ref_root = generic_ref->tree_ref.ref_root;
688 owner = generic_ref->tree_ref.level;
689 } else if (!parent) {
690 ref_root = generic_ref->data_ref.ref_root;
691 owner = generic_ref->data_ref.ino;
692 offset = generic_ref->data_ref.offset;
694 metadata = owner < BTRFS_FIRST_FREE_OBJECTID;
696 ref = kzalloc(sizeof(struct ref_entry), GFP_NOFS);
697 ra = kmalloc(sizeof(struct ref_action), GFP_NOFS);
705 ref->parent = parent;
707 ref->root_objectid = ref_root;
708 ref->offset = offset;
709 ref->num_refs = (action == BTRFS_DROP_DELAYED_REF) ? -1 : 1;
711 memcpy(&ra->ref, ref, sizeof(struct ref_entry));
713 * Save the extra info from the delayed ref in the ref action to make it
714 * easier to figure out what is happening. The real ref's we add to the
715 * ref tree need to reflect what we save on disk so it matches any
716 * on-disk refs we pre-loaded.
718 ra->ref.owner = owner;
719 ra->ref.offset = offset;
720 ra->ref.root_objectid = ref_root;
721 __save_stack_trace(ra);
723 INIT_LIST_HEAD(&ra->list);
725 ra->root = generic_ref->real_root;
728 * This is an allocation, preallocate the block_entry in case we haven't
732 if (action == BTRFS_ADD_DELAYED_EXTENT) {
734 * For subvol_create we'll just pass in whatever the parent root
735 * is and the new root objectid, so let's not treat the passed
736 * in root as if it really has a ref for this bytenr.
738 be = add_block_entry(fs_info, bytenr, num_bytes, ref_root);
749 if (be->num_refs != 1) {
751 "re-allocated a block that still has references to it!");
752 dump_block_entry(fs_info, be);
753 dump_ref_action(fs_info, ra);
759 while (!list_empty(&be->actions)) {
760 struct ref_action *tmp;
762 tmp = list_first_entry(&be->actions, struct ref_action,
764 list_del(&tmp->list);
768 struct root_entry *tmp;
771 re = kmalloc(sizeof(struct root_entry), GFP_NOFS);
779 * This is the root that is modifying us, so it's the
780 * one we want to lookup below when we modify the
783 ref_root = generic_ref->real_root;
784 re->root_objectid = generic_ref->real_root;
788 spin_lock(&fs_info->ref_verify_lock);
789 be = lookup_block_entry(&fs_info->block_tree, bytenr);
792 "trying to do action %d to bytenr %llu num_bytes %llu but there is no existing entry!",
793 action, bytenr, num_bytes);
794 dump_ref_action(fs_info, ra);
798 } else if (be->num_refs == 0) {
800 "trying to do action %d for a bytenr that has 0 total references",
802 dump_block_entry(fs_info, be);
803 dump_ref_action(fs_info, ra);
810 tmp = insert_root_entry(&be->roots, re);
818 exist = insert_ref_entry(&be->refs, ref);
820 if (action == BTRFS_DROP_DELAYED_REF) {
821 if (exist->num_refs == 0) {
823 "dropping a ref for a existing root that doesn't have a ref on the block");
824 dump_block_entry(fs_info, be);
825 dump_ref_action(fs_info, ra);
831 if (exist->num_refs == 0) {
832 rb_erase(&exist->node, &be->refs);
835 } else if (!be->metadata) {
839 "attempting to add another ref for an existing ref on a tree block");
840 dump_block_entry(fs_info, be);
841 dump_ref_action(fs_info, ra);
848 if (action == BTRFS_DROP_DELAYED_REF) {
850 "dropping a ref for a root that doesn't have a ref on the block");
851 dump_block_entry(fs_info, be);
852 dump_ref_action(fs_info, ra);
859 if (!parent && !re) {
860 re = lookup_root_entry(&be->roots, ref_root);
863 * This shouldn't happen because we will add our re
864 * above when we lookup the be with !parent, but just in
865 * case catch this case so we don't panic because I
866 * didn't think of some other corner case.
868 btrfs_err(fs_info, "failed to find root %llu for %llu",
869 generic_ref->real_root, be->bytenr);
870 dump_block_entry(fs_info, be);
871 dump_ref_action(fs_info, ra);
876 if (action == BTRFS_DROP_DELAYED_REF) {
880 } else if (action == BTRFS_ADD_DELAYED_REF) {
885 list_add_tail(&ra->list, &be->actions);
888 spin_unlock(&fs_info->ref_verify_lock);
891 btrfs_clear_opt(fs_info->mount_opt, REF_VERIFY);
895 /* Free up the ref cache */
896 void btrfs_free_ref_cache(struct btrfs_fs_info *fs_info)
898 struct block_entry *be;
901 if (!btrfs_test_opt(fs_info, REF_VERIFY))
904 spin_lock(&fs_info->ref_verify_lock);
905 while ((n = rb_first(&fs_info->block_tree))) {
906 be = rb_entry(n, struct block_entry, node);
907 rb_erase(&be->node, &fs_info->block_tree);
908 free_block_entry(be);
909 cond_resched_lock(&fs_info->ref_verify_lock);
911 spin_unlock(&fs_info->ref_verify_lock);
914 void btrfs_free_ref_tree_range(struct btrfs_fs_info *fs_info, u64 start,
917 struct block_entry *be = NULL, *entry;
920 if (!btrfs_test_opt(fs_info, REF_VERIFY))
923 spin_lock(&fs_info->ref_verify_lock);
924 n = fs_info->block_tree.rb_node;
926 entry = rb_entry(n, struct block_entry, node);
927 if (entry->bytenr < start) {
929 } else if (entry->bytenr > start) {
935 /* We want to get as close to start as possible */
937 (entry->bytenr < start && be->bytenr > start) ||
938 (entry->bytenr < start && entry->bytenr > be->bytenr))
943 * Could have an empty block group, maybe have something to check for
944 * this case to verify we were actually empty?
947 spin_unlock(&fs_info->ref_verify_lock);
953 be = rb_entry(n, struct block_entry, node);
955 if (be->bytenr < start && be->bytenr + be->len > start) {
957 "block entry overlaps a block group [%llu,%llu]!",
959 dump_block_entry(fs_info, be);
962 if (be->bytenr < start)
964 if (be->bytenr >= start + len)
966 if (be->bytenr + be->len > start + len) {
968 "block entry overlaps a block group [%llu,%llu]!",
970 dump_block_entry(fs_info, be);
972 rb_erase(&be->node, &fs_info->block_tree);
973 free_block_entry(be);
975 spin_unlock(&fs_info->ref_verify_lock);
978 /* Walk down all roots and build the ref tree, meant to be called at mount */
979 int btrfs_build_ref_tree(struct btrfs_fs_info *fs_info)
981 struct btrfs_root *extent_root;
982 struct btrfs_path *path;
983 struct extent_buffer *eb;
984 int tree_block_level = 0;
985 u64 bytenr = 0, num_bytes = 0;
988 if (!btrfs_test_opt(fs_info, REF_VERIFY))
991 path = btrfs_alloc_path();
995 extent_root = btrfs_extent_root(fs_info, 0);
996 eb = btrfs_read_lock_root_node(extent_root);
997 level = btrfs_header_level(eb);
998 path->nodes[level] = eb;
999 path->slots[level] = 0;
1000 path->locks[level] = BTRFS_READ_LOCK;
1004 * We have to keep track of the bytenr/num_bytes we last hit
1005 * because we could have run out of space for an inline ref, and
1006 * would have had to added a ref key item which may appear on a
1007 * different leaf from the original extent item.
1009 ret = walk_down_tree(extent_root, path, level,
1010 &bytenr, &num_bytes, &tree_block_level);
1013 ret = walk_up_tree(path, &level);
1022 btrfs_clear_opt(fs_info->mount_opt, REF_VERIFY);
1023 btrfs_free_ref_cache(fs_info);
1025 btrfs_free_path(path);