So I finally figured out the cause of the random test failures I've been
seeing. It seems udev has a race condition. While the test suite is
busy adding and removing partitions, udev can get behind a bit and while
processing the removal event of the partition, it can decide to unlink
the dev node of the by now recreated partition, so the partition device
vanishes out from under the test script momentarily, until udev
processes the re-add event and puts it back.
Hopefully this will be solved whenever udev finally drops support for
systems without devtmpfs and gets rid of all calls to mknod and unlink
for the dev node.
In the mean time, I wonder if we could add a udevadm settle between each
test and if that would help? Hrm... guess I'll go try it.
