btrfs: send: fix duplicated rmdir operations when using extrefs

Commit 29d6d30f5c8a ("Btrfs: send, don't send rmdir for same target
multiple times") has fixed an issue that a send stream contained a rmdir
operation for the same directory multiple times. After that fix we keep
track of the last directory for which we sent a rmdir operation and
compare with it before sending a rmdir for the parent inode of a deleted
hardlink we are processing. But there is still a corner case that in
between rmdir dir operations for the same inode we find deleted hardlinks
for other parent inodes, so tracking just the last inode for which we sent
a rmdir operation is not enough.

Hardlinks of a file in the same directory are stored in the same INODE_REF
item, but if the number of hardlinks is too large and can not fit in a
leaf, we use INODE_EXTREF items to store them. The key of an INODE_EXTREF
item is (inode_id, INODE_EXTREF, hash[name, parent ino]), so between two
hardlinks for the same parent directory, we can find others for other
parent directories. For example for the reproducer below we get the
following (from a btrfs inspect-internal dump-tree output):

item 0 key (259 INODE_EXTREF 2309449) itemoff 16257 itemsize 26
index 6925 parent 257 namelen 8 name: foo.6923
item 1 key (259 INODE_EXTREF 2311350) itemoff 16231 itemsize 26
index 6588 parent 258 namelen 8 name: foo.6587
item 2 key (259 INODE_EXTREF 2457395) itemoff 16205 itemsize 26
index 6611 parent 257 namelen 8 name: foo.6609
(...)

So tracking the last directory's inode number does not work in this case
since we process a link for parent inode 257, then for 258 and then back
again for 257, and that second time we process a deleted link for 257 we
think we have not yet sent a rmdir operation.

Fix this by using a rbtree to keep track of all the directories for which
we have already sent rmdir operations, and add those directories to the
'check_dirs' ref list in process_recorded_refs() only if the directory is
not yet in the rbtree, otherwise skip it since it means we have already
sent a rmdir operation for that directory.

The following test script reproduces the problem:

$ cat test.sh
#!/bin/bash

DEV=/dev/sdi
MNT=/mnt/sdi

mkfs.btrfs -f $DEV
mount $DEV $MNT

mkdir $MNT/a $MNT/b

echo 123 > $MNT/a/foo
for ((i = 1; i <= 1000; i++)); do
ln $MNT/a/foo $MNT/a/foo.$i
ln $MNT/a/foo $MNT/b/foo.$i
done

btrfs subvolume snapshot -r $MNT $MNT/snap1
btrfs send $MNT/snap1 -f /tmp/base.send

rm -r $MNT/a $MNT/b

btrfs subvolume snapshot -r $MNT $MNT/snap2
btrfs send -p $MNT/snap1 $MNT/snap2 -f /tmp/incremental.send

umount $MNT
mkfs.btrfs -f $DEV
mount $DEV $MNT

btrfs receive $MNT -f /tmp/base.send
btrfs receive $MNT -f /tmp/incremental.send

rm -f /tmp/base.send /tmp/incremental.send

umount $MNT

When running it, it fails like this:

$ ./test.sh
(...)
At subvol snap1
At snapshot snap2
ERROR: rmdir o257-9-0 failed: No such file or directory

CC: <[e-mail]>
Reviewed-by: Filipe Manana <[e-mail]>
Signed-off-by: Ting-Chang Hou <[e-mail]>
[ Updated changelog ]
Signed-off-by: Filipe Manana <[e-mail]>
Signed-off-by: David Sterba <[e-mail]>

fs/btrfs/send.c

···
 	return ret;
 }
 
+static int rbtree_check_dir_ref_comp(const void *k, const struct rb_node *node)
+{
+	const struct recorded_ref *data = k;
+	const struct recorded_ref *ref = rb_entry(node, struct recorded_ref, node);
+
+	if (data->dir > ref->dir)
+		return 1;
+	if (data->dir < ref->dir)
+		return -1;
+	if (data->dir_gen > ref->dir_gen)
+		return 1;
+	if (data->dir_gen < ref->dir_gen)
+		return -1;
+	return 0;
+}
+
+static bool rbtree_check_dir_ref_less(struct rb_node *node, const struct rb_node *parent)
+{
+	const struct recorded_ref *entry = rb_entry(node, struct recorded_ref, node);
+
+	return rbtree_check_dir_ref_comp(entry, parent) < 0;
+}
+
+static int record_check_dir_ref_in_tree(struct rb_root *root,
+			struct recorded_ref *ref, struct list_head *list)
+{
+	struct recorded_ref *tmp_ref;
+	int ret;
+
+	if (rb_find(ref, root, rbtree_check_dir_ref_comp))
+		return 0;
+
+	ret = dup_ref(ref, list);
+	if (ret < 0)
+		return ret;
+
+	tmp_ref = list_last_entry(list, struct recorded_ref, list);
+	rb_add(&tmp_ref->node, root, rbtree_check_dir_ref_less);
+	tmp_ref->root = root;
+	return 0;
+}
+
 static int rename_current_inode(struct send_ctx *sctx,
 				struct fs_path *current_path,
 				struct fs_path *new_path)
···
 	struct recorded_ref *cur;
 	struct recorded_ref *cur2;
 	LIST_HEAD(check_dirs);
+	struct rb_root rbtree_check_dirs = RB_ROOT;
 	struct fs_path *valid_path = NULL;
 	u64 ow_inode = 0;
 	u64 ow_gen;
 	u64 ow_mode;
-	u64 last_dir_ino_rm = 0;
 	bool did_overwrite = false;
 	bool is_orphan = false;
 	bool can_rename = true;
···
 					goto out;
 			}
 		}
-		ret = dup_ref(cur, &check_dirs);
+		ret = record_check_dir_ref_in_tree(&rbtree_check_dirs, cur, &check_dirs);
 		if (ret < 0)
 			goto out;
 	}
···
 		}
 
 		list_for_each_entry(cur, &sctx->deleted_refs, list) {
-			ret = dup_ref(cur, &check_dirs);
+			ret = record_check_dir_ref_in_tree(&rbtree_check_dirs, cur, &check_dirs);
 			if (ret < 0)
 				goto out;
 		}
···
 		 * We have a moved dir. Add the old parent to check_dirs
 		 */
 		cur = list_first_entry(&sctx->deleted_refs, struct recorded_ref, list);
-		ret = dup_ref(cur, &check_dirs);
+		ret = record_check_dir_ref_in_tree(&rbtree_check_dirs, cur, &check_dirs);
 		if (ret < 0)
 			goto out;
 	} else if (!S_ISDIR(sctx->cur_inode_mode)) {
···
 				if (is_current_inode_path(sctx, cur->full_path))
 					fs_path_reset(&sctx->cur_inode_path);
 			}
-			ret = dup_ref(cur, &check_dirs);
+			ret = record_check_dir_ref_in_tree(&rbtree_check_dirs, cur, &check_dirs);
 			if (ret < 0)
 				goto out;
 		}
···
 			ret = cache_dir_utimes(sctx, cur->dir, cur->dir_gen);
 			if (ret < 0)
 				goto out;
-		} else if (ret == inode_state_did_delete &&
-			   cur->dir != last_dir_ino_rm) {
+		} else if (ret == inode_state_did_delete) {
 			ret = can_rmdir(sctx, cur->dir, cur->dir_gen);
 			if (ret < 0)
 				goto out;
···
 				ret = send_rmdir(sctx, valid_path);
 				if (ret < 0)
 					goto out;
-				last_dir_ino_rm = cur->dir;
 			}
 		}
 	}