Re: [PATCH 0/2] Introduce panic hypercall

[Gleb Natapov <gleb@redhat.com>]

On Tue, Jun 21, 2011 at 11:56:26AM +0300, Avi Kivity wrote:
> On 06/21/2011 11:41 AM, Gleb Natapov wrote:
> >On Tue, Jun 21, 2011 at 11:09:21AM +0300, Avi Kivity wrote:
> >>  On 06/21/2011 09:02 AM, Gleb Natapov wrote:
> >>  >On Mon, Jun 20, 2011 at 07:34:57PM +0300, Avi Kivity wrote:
> >>  >>   >The only two things which came to my mind are:
> >>  >>   >
> >>  >>   >    * NMI (aka. ipmitool diag) - already available in qemu/kvm - but requires
> >>  >>   >      in-guest kexec/kdump
> >>  >>   >    * Hardware-Watchdog (also available in qemu/libvirt)
> >>  >>
> >>  >>   A watchdog has the advantage that is also detects lockups.
> >>  >>
> >>  >And has disadvantage that all time base heuristics are bite us in the
> >>  >end.
> >>
> >>  The perf-based watchdog counts clocks-not-halted, not time, so it is
> >>  safe from time issues.
> >So it counts only instruction that guest actually executed? That's
> >perfect then. How much overhead it has in a guest though?
> 
> Should be pretty low, especially if the guest is idle.  There were
> some negative reports about the pmu From David Ahern, so it needs to
> be verified.
But it will be running not only when the guest is idle.

> 
> 
> >>                          We could make the hardware watchdog cheat in
> >>  the same way.
> >>
> >Something like steal time, but for watchdog. But this become complicated fast.
> >Watchdog emulation will have to move into kernel for starter.
> 
> Why?  You can use a performance counter from userspace.
> 
Heh, haven't thought about such way of implementing watchdog device.
But the same question again: what impact on performance constantly running
guest under perf is? Doesn't running guest under perf involve a lot of
NMIs (and hence vmexists)?

--
			Gleb.