domain pause vs. watchdog timer

domain pause vs. watchdog timer

[Laszlo Ersek]

Hi,

if a domain is paused (= not scheduled), its process that keeps 
otherwise kicking /dev/watchdog and (thusly the hypervisor) "freezes" 
too. Should pausing a domain suspend (or "inactivate") its watchdog_timers?

(I believe this must have been considered and the answer is "you 
shouldn't do that", but a confirmation/refutation would be nice.)

Thanks
L.

Re: domain pause vs. watchdog timer

[Jan Beulich]

>>> On 13.10.11 at 15:41, Laszlo Ersek <lersek@redhat.com> wrote:
> if a domain is paused (= not scheduled), its process that keeps 
> otherwise kicking /dev/watchdog and (thusly the hypervisor) "freezes" 
> too. Should pausing a domain suspend (or "inactivate") its watchdog_timers?

I would say yes it should. And you say it doesn't?

> (I believe this must have been considered and the answer is "you 
> shouldn't do that", but a confirmation/refutation would be nice.)

If a domain gets paused from the tools stack, it's not very reasonable
to expect the operator there to know whether a particular domain
has a watchdog timer activated.

Jan

Re: domain pause vs. watchdog timer

[Laszlo Ersek]

On 10/13/11 16:21, Jan Beulich wrote:
>>>> On 13.10.11 at 15:41, Laszlo Ersek<lersek@redhat.com>  wrote:
>> if a domain is paused (= not scheduled), its process that keeps
>> otherwise kicking /dev/watchdog and (thusly the hypervisor) "freezes"
>> too. Should pausing a domain suspend (or "inactivate") its watchdog_timers?
>
> I would say yes it should. And you say it doesn't?

I must have missed a changeset (or a hunk?) related to pausing:

https://bugzilla.redhat.com/show_bug.cgi?id=742880#c1

I backported 21529 from xen-unstable to the RHEL-5 hypervisor version, 
and Igor tested pausing a Fedora 16 HVM guest. The domU was destroyed 
while it was paused.

Thanks!
L.

Re: domain pause vs. watchdog timer

[Keir Fraser]

On 13/10/2011 15:38, "Laszlo Ersek" <lersek@redhat.com> wrote:

> On 10/13/11 16:21, Jan Beulich wrote:
>>>>> On 13.10.11 at 15:41, Laszlo Ersek<lersek@redhat.com>  wrote:
>>> if a domain is paused (= not scheduled), its process that keeps
>>> otherwise kicking /dev/watchdog and (thusly the hypervisor) "freezes"
>>> too. Should pausing a domain suspend (or "inactivate") its watchdog_timers?
>> 
>> I would say yes it should. And you say it doesn't?
> 
> I must have missed a changeset (or a hunk?) related to pausing:
> 
> https://bugzilla.redhat.com/show_bug.cgi?id=742880#c1
> 
> I backported 21529 from xen-unstable to the RHEL-5 hypervisor version,
> and Igor tested pausing a Fedora 16 HVM guest. The domU was destroyed
> while it was paused.

You are just as likely to upset a watchdog implemented entirely internally
to the guest, and which we have no chance to disable/suspend when a domain
is paused. Domain pausing from the dom0 command line really doesn't have a
production use case afaik -- just don't do it is my helpful suggestion.

 -- Keir

Re: domain pause vs. watchdog timer

[Laszlo Ersek]

On 10/13/11 16:56, Keir Fraser wrote:
> On 13/10/2011 15:38, "Laszlo Ersek"<lersek@redhat.com>  wrote:
>
>> On 10/13/11 16:21, Jan Beulich wrote:
>>>>>> On 13.10.11 at 15:41, Laszlo Ersek<lersek@redhat.com>   wrote:
>>>> if a domain is paused (= not scheduled), its process that keeps
>>>> otherwise kicking /dev/watchdog and (thusly the hypervisor) "freezes"
>>>> too. Should pausing a domain suspend (or "inactivate") its watchdog_timers?
>>>
>>> I would say yes it should. And you say it doesn't?
>>
>> I must have missed a changeset (or a hunk?) related to pausing:
>>
>> https://bugzilla.redhat.com/show_bug.cgi?id=742880#c1
>>
>> I backported 21529 from xen-unstable to the RHEL-5 hypervisor version,
>> and Igor tested pausing a Fedora 16 HVM guest. The domU was destroyed
>> while it was paused.
>
> You are just as likely to upset a watchdog implemented entirely internally
> to the guest, and which we have no chance to disable/suspend when a domain
> is paused. Domain pausing from the dom0 command line really doesn't have a
> production use case afaik -- just don't do it is my helpful suggestion.

Thank you both for the answers.
L.

Re: domain pause vs. watchdog timer

[Keir Fraser]

On 13/10/2011 14:41, "Laszlo Ersek" <lersek@redhat.com> wrote:

> Hi,
> 
> if a domain is paused (= not scheduled), its process that keeps
> otherwise kicking /dev/watchdog and (thusly the hypervisor) "freezes"
> too. Should pausing a domain suspend (or "inactivate") its watchdog_timers?
> 
> (I believe this must have been considered and the answer is "you
> shouldn't do that", but a confirmation/refutation would be nice.)

Yes, the ability to pause a domain from the toolstack command line was added
way back in the beginning largely because... Well, because it could be done,
and it was easy! It's not actually a very good idea. You won't actually find
the toolstack itself automatically pausing a domain in very many places at
all -- most places where it does are debug/diagnostic code, or where a
domain is about to be destroyed anyway.

So, yes, in short: don't do that! :-)

 -- Keir

> Thanks
> L.
> 
> _______________________________________________
> Xen-devel mailing list
> Xen-devel@lists.xensource.com
> http://lists.xensource.com/xen-devel