udevd messages lost on exit?

udevd messages lost on exit?

[James Hunt]

Hi All,

I'm investigating a problem in Ubuntu Oneiric where we think udev messages maybe be getting lost due
to the way that we're stopping udevd:

  https://bugs.launchpad.net/ubuntu/+source/udev/+bug/818177

In summary, at the end of our initramfs, we are doing this:

  # Stop udevd, we'll miss a few events while we run init, but we catch up
  udevadm control --exit

  # Move /dev to the real filesystem
  mount -n -o move /dev ${rootmnt}/dev

The problem is that most notably when using lvm2 the "--exit" call seems to be causing udev messages
relating to the *rootfs* to be lost so that the system fails to boot.

Looking at udevd.c, it does the following when requested to exit:

  /* discard queued events and kill workers */
  event_queue_cleanup(udev, EVENT_QUEUED);
  worker_kill(udev, 0);

I'm currently trying to debug what is left in the queue to get a better handle on this, but I am
wondering if "udevadm control --exit" should request that udevd drain the queues rather than
discarding the messages for our case. Is the reason the messages are discarded to avoid slowing down
the boot coupled with the expectation that the caller will use "udevadm trigger" to catch up later?

Kind regards,

James.

Re: udevd messages lost on exit?

[Kay Sievers]

On Fri, Sep 30, 2011 at 10:29, James Hunt <james.hunt@ubuntu.com> wrote:
> I'm investigating a problem in Ubuntu Oneiric where we think udev messages maybe be getting lost due
> to the way that we're stopping udevd:
>
>  https://bugs.launchpad.net/ubuntu/+source/udev/+bug/818177
>
> In summary, at the end of our initramfs, we are doing this:
>
>  # Stop udevd, we'll miss a few events while we run init, but we catch up
>  udevadm control --exit
>
>  # Move /dev to the real filesystem
>  mount -n -o move /dev ${rootmnt}/dev
>
> The problem is that most notably when using lvm2 the "--exit" call seems to be causing udev messages
> relating to the *rootfs* to be lost so that the system fails to boot.
>
> Looking at udevd.c, it does the following when requested to exit:
>
>  /* discard queued events and kill workers */
>  event_queue_cleanup(udev, EVENT_QUEUED);
>  worker_kill(udev, 0);
>
> I'm currently trying to debug what is left in the queue to get a better handle on this, but I am
> wondering if "udevadm control --exit" should request that udevd drain the queues rather than
> discarding the messages for our case. Is the reason the messages are discarded to avoid slowing down
> the boot coupled with the expectation that the caller will use "udevadm trigger" to catch up later?

--exit just does a clean shutdown, it will not check for pending
events, only already running stuff will be properly finished, but
nothing new will be started.

If you want queued stuff, queued in udevd or the kernel, to be all
handled before exiting, add a call to 'settle' before the --exit call.

Kay

Re: udevd messages lost on exit?

[James Hunt]

On 30/09/11 10:03, Kay Sievers wrote:
> --exit just does a clean shutdown, it will not check for pending
> events, only already running stuff will be properly finished, but
> nothing new will be started.
So if events are pending in udevd they will be lost unless trigger is called?

> If you want queued stuff, queued in udevd or the kernel, to be all
> handled before exiting, add a call to 'settle' before the --exit call.
Is that combination atomic?

Kind regards,

James Hunt

Re: udevd messages lost on exit?

[Kay Sievers]

On Fri, Sep 30, 2011 at 11:33, James Hunt <james.hunt@ubuntu.com> wrote:
> On 30/09/11 10:03, Kay Sievers wrote:
>> --exit just does a clean shutdown, it will not check for pending
>> events, only already running stuff will be properly finished, but
>> nothing new will be started.
> So if events are pending in udevd they will be lost unless trigger is called?

There is no idea at all of any continuous handling of events from
initramfs to the real root. These are completely different rule sets
and the stuff from initramfs needs to be replayed anyway.

>> If you want queued stuff, queued in udevd or the kernel, to be all
>> handled before exiting, add a call to 'settle' before the --exit call.
> Is that combination atomic?

See above. You need to trigger anyway. The rootfs might configure the
devices differently than the initramfs. The device database usually
much more stuff to carry.

Kay