Comment 101 for bug 332270

There are two parts to this bug.

The actual bug is being caused by "lvm vgscan" opening all block devices on the system writable and then closing them again, which triggers the inotify event. There seems to be no reason they cannot be opened readonly, since the only thing it does is call fstat() and read() on them.

Since the devices include any LVM PVs (which is what it's looking for), this gives you a recursive loop where LVM PVs appear to have been changed, so udev calls "lvm vgscan" again to update its tables.

Part #1 of the fix will be a patch to LVM to only open block devices with O_RDONLY

Part #2 of the fix will be a patch to udev to not add the inotify watch until all RUN rules are processed; this will prevent the initial onslaught on boot, and hopefully guard against anything it runs changing the device resulting in reprocessing (which is better handled with udev rules ordering anyway).

It will obviously not protect against a "udevadm trigger" once booted, so it's possible that we'll trip over this with another program later and will have to fix that one in the same manner as LVM.

It's not yet clear why this does not affect _everyone_, there appears to be a race involved as slowing udev down with various log levels - or using only one core - is enough to hide the bug for some people. I need to work out why that is before uploading the fixes