I wonder if any of this is caused by the increased task parallelism we have now. Maybe mounting a squashfs file along with another squashfs file, using libmount is just racy and we somehow end up switching the backing store of a loopback device while it is being mounted.