[libvirt] [RFC PATCH] udev: Remove udev handle from main loop when udev thread stops

[Marc Hartmayer]

On Thu, Feb 07, 2019 at 05:08 PM +0100, Marc Hartmayer <mhartmay at linux.ibm.com> wrote:
> Commit "nodedev: Move device enumumeration out of nodeStateInitialize"
> (9f0ae0b18e3e620) has moved the heavy task of device enumeration into
> a separate thread. The problem with this commit is that there is a
> functionality change in the cleanup when udevEnumerateDevices
> fails. Before commit 9f0ae0b18e3e620, the entire driver was cleaned up
> by calling nodeStateCleanup (defined in node_device_udev.c) which
> resulted in libvirtd stopping with the error message
> 'daemonRunStateInit:800 : Driver state initialization failed'. With
> the commit 9f0ae0b18e3e620 there is only a signal to the udev thread
> that it must stop. This means that for example the watch handle isn't
> removed from the main loop and this can result in the main loop
> consuming 100% CPU time as soon as a new udev event occurs.
>
> This patch proposes a simple solution to the described problem. In
> case the udev thread stops the watch handle is removed from the main
> loop.

Another option would be to stop libvirtd if udevEnumerateDevices fails
in nodeStateInitializeEnumerate.

[…snip]