jfs: nlink overflow in jfs_rename

[ Upstream commit 9218dc2 ]

If nlink is maximal for a directory (-1) and inside that directory you
perform a rename for some child directory (not moving from the parent),
then the nlink of the first directory is first incremented and later
decremented. Normally this is fine, but when nlink = -1 this causes a
wrap around to 0, and then drop_nlink issues a warning.

After applying the patch syzbot no longer issues any warnings. I also
ran some basic fs tests to look for any regressions.

Signed-off-by: Jori Koolstra <jkoolstra@xs4all.nl>
Reported-by: syzbot+9131ddfd7870623b719f@syzkaller.appspotmail.com
Closes: https://syzbot.org/bug?extid=9131ddfd7870623b719f
Signed-off-by: Dave Kleikamp <dave.kleikamp@oracle.com>
Signed-off-by: Sasha Levin <sashal@kernel.org>

Author: LATENTBOUNCE

fs/jfs/namei.c

@@ -1228,7 +1228,7 @@ static int jfs_rename(struct mnt_idmap *idmap, struct inode *old_dir,
 				jfs_err("jfs_rename: dtInsert returned -EIO");
 			goto out_tx;
 		}
-		if (S_ISDIR(old_ip->i_mode))
+		if (S_ISDIR(old_ip->i_mode) && old_dir != new_dir)
 			inc_nlink(new_dir);
 	}
 	/*
@@ -1244,7 +1244,9 @@ static int jfs_rename(struct mnt_idmap *idmap, struct inode *old_dir,
 		goto out_tx;
 	}
 	if (S_ISDIR(old_ip->i_mode)) {
-		drop_nlink(old_dir);
+		if (new_ip || old_dir != new_dir)
+			drop_nlink(old_dir);
+
 		if (old_dir != new_dir) {
 			/*
 			 * Change inode number of parent for moved directory