Re: [Qemu-devel] [PATCH 0/2 v3] kvm: notify host when guest panicked

[Gleb Natapov <gleb@redhat.com>]

On Thu, Mar 22, 2012 at 03:44:45PM +0800, Wen Congyang wrote:
> At 03/22/2012 03:31 PM, Gleb Natapov Wrote:
> > On Thu, Mar 22, 2012 at 09:05:12AM +0800, Wen Congyang wrote:
> >> At 03/22/2012 03:19 AM, Anthony Liguori Wrote:
> >>> On 03/21/2012 11:25 AM, Avi Kivity wrote:
> >>>> On 03/21/2012 06:18 PM, Corey Minyard wrote:
> >>>>>
> >>>>>> Look at drivers/char/ipmi/ipmi_msghandler.c. It has code to send panic
> >>>>>> event over IMPI. The code is pretty complex. Of course if we a going to
> >>>>>> implement something more complex than simple hypercall for panic
> >>>>>> notification we better do something more interesting with it than just
> >>>>>> saying "panic happened", like sending stack traces on all cpus for
> >>>>>> instance.
> >>>>>
> >>>>> I doubt that's the best example, unfortunately.  The IPMI event log
> >>>>> has limited space and it has to be send a little piece at a time since
> >>>>> each log entry is 14 bytes.  It just prints the panic string, nothing
> >>>>> else.  Not that it isn't useful, it has saved my butt before.
> >>>>>
> >>>>> You have lots of interesting options with paravirtualization.  You
> >>>>> could, for instance, create a console driver that delivered all
> >>>>> console output efficiently through a hypercall.  That would be really
> >>>>> easy.  Or, as you mention, a custom way to deliver panic information.
> >>>>> Collecting information like stack traces would be harder to
> >>>>> accomplish, as I don't think there is currently a way to get it except
> >>>>> by sending it to printk.
> >>>>
> >>>> That already exists; virtio-console (or serial console emulation) can do
> >>>> the job.
> >>>
> >>> I think the use case here is pretty straight forward: if the guest finds
> >>> itself in bad place, it wants to indicate that to the host.
> >>>
> >>> We shouldn't rely on any device drivers or complex code.  It should be
> >>> as close to a single instruction as possible that can run even if
> >>> interrupts are disabled.
> >>>
> >>> An out instruction fits this very well.  I think a simple protocol like:
> >>
> >> This solution is more simple than using virtio-serial.
> >>
> >>>
> >>> inl PORT -> returns a magic number indicating the presence of qemucalls
> >>
> >> I donot understantd this instruction's purpose.
> >>
> >>> inl PORT+1 -> returns a bitmap of supported features
> >>
> >> Hmm, we can execute this instruction when guest starts. If the userspace
> >> does not process panicked event, there is no need to notify it.
> >>
> >>>
> >>> outl PORT+1 -> data reg1
> >>> outl PORT+2 -> data reg2
> >>> outl PORT+N -> data regN
> >>
> >> We can get the register value from vmcs. So there is no need to tell
> >> the register value to the host.
> >>
> > No device should examine register value. Ideally QEMU would read
> > registers only during migration.
> 
> I mean: if the qemu(or other app) want to know the register value, it can
> get it from vmcs. So there is no need to pass register value from guest
> to host.
> 
I understand what you mean and I am saying that reading register values
shouldn't be part of the protocol. Examining cpu state postmortem is OK
of course.

> Another question: each outl will cause vmexit?
> 
Yes, IIRC you can pass through io port to a guest but KVM doesn't do it.

> Thanks
> Wen Congyang
> 
> > 
> >> If we decide to avoid touching hypervisor, I agree with this solution.
> >>
> >> Thanks
> >> Wen Congyang
> >>>
> >>> outl PORT -> qemucall of index value with arguments 1..N
> >>>
> >>> Regards,
> >>>
> >>> Anthony Liguori
> >>>
> >>>>
> >>>> In fact the feature can be implemented 100% host side by searching for a
> >>>> panic string signature in the console logs.
> >>>>
> >>>
> >>> -- 
> >>> To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
> >>> the body of a message to majordomo@vger.kernel.org
> >>> More majordomo info at  http://vger.kernel.org/majordomo-info.html
> >>> Please read the FAQ at  http://www.tux.org/lkml/
> >>>
> > 
> > --
> > 			Gleb.
> > --
> > To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
> > the body of a message to majordomo@vger.kernel.org
> > More majordomo info at  http://vger.kernel.org/majordomo-info.html
> > Please read the FAQ at  http://www.tux.org/lkml/
> > 

--
			Gleb.