bcachefs: Fix snapshotting a subvolume, then renaming it

Subvolume roots and the dirents that point to them are special; they
don't obey the normal snapshot versioning rules because they cross
snapshot boundaries.

We don't keep around older versions of subvolume dirents on rename - we
don't need to, because subvolume dirents are only visible in the parent
subvolume, and we wouldn't be able to match up the different dirent and
inode versions due to crossing the snapshot ID boundary.

That means that when we rename a subvolume, that's been snapshotted, the
older version of the subvolume root will become dangling - it won't have
a dirent that points to it.

That's expected, we just need to tell fsck that this is ok.

Fixes: https://github.com/koverstreet/bcachefs/issues/856
Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>

fs/bcachefs/fsck.c

@@ -321,6 +321,31 @@ static inline bool inode_should_reattach(struct bch_inode_unpacked *inode)
 	    inode->bi_subvol == BCACHEFS_ROOT_SUBVOL)
 		return false;
 
+	/*
+	 * Subvolume roots are special: older versions of subvolume roots may be
+	 * disconnected, it's only the newest version that matters.
+	 *
+	 * We only keep a single dirent pointing to a subvolume root, i.e.
+	 * older versions of snapshots will not have a different dirent pointing
+	 * to the same subvolume root.
+	 *
+	 * This is because dirents that point to subvolumes are only visible in
+	 * the parent subvolume - versioning is not needed - and keeping them
+	 * around would break fsck, because when we're crossing subvolumes we
+	 * don't have a consistent snapshot ID to do check the inode <-> dirent
+	 * relationships.
+	 *
+	 * Thus, a subvolume root that's been renamed after a snapshot will have
+	 * a disconnected older version - that's expected.
+	 *
+	 * Note that taking a snapshot always updates the root inode (to update
+	 * the dirent backpointer), so a subvolume root inode with
+	 * BCH_INODE_has_child_snapshot is never visible.
+	 */
+	if (inode->bi_subvol &&
+	    (inode->bi_flags & BCH_INODE_has_child_snapshot))
+		return false;
+
 	return !inode->bi_dir && !(inode->bi_flags & BCH_INODE_unlinked);
 }
 
@@ -1007,6 +1032,23 @@ static int check_inode_dirent_inode(struct btree_trans *trans,
 	if (ret && !bch2_err_matches(ret, ENOENT))
 		return ret;
 
+	if ((ret || dirent_points_to_inode_nowarn(d, inode)) &&
+	    inode->bi_subvol &&
+	    (inode->bi_flags & BCH_INODE_has_child_snapshot)) {
+		/* Older version of a renamed subvolume root: we won't have a
+		 * correct dirent for it. That's expected, see
+		 * inode_should_reattach().
+		 *
+		 * We don't clear the backpointer field when doing the rename
+		 * because there might be arbitrarily many versions in older
+		 * snapshots.
+		 */
+		inode->bi_dir = 0;
+		inode->bi_dir_offset = 0;
+		*write_inode = true;
+		goto out;
+	}
+
 	if (fsck_err_on(ret,
 			trans, inode_points_to_missing_dirent,
 			"inode points to missing dirent\n%s",
@@ -1027,7 +1069,7 @@ static int check_inode_dirent_inode(struct btree_trans *trans,
 		inode->bi_dir_offset = 0;
 		*write_inode = true;
 	}
-
+out:
 	ret = 0;
 fsck_err:
 	bch2_trans_iter_exit(trans, &dirent_iter);