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

[Gleb Natapov]

On Wed, Mar 21, 2012 at 05:35:49PM +0800, Wen Congyang wrote:
> At 03/21/2012 05:11 PM, Gleb Natapov Wrote:
> > On Wed, Mar 21, 2012 at 08:56:03AM +0800, Wen Congyang wrote:
> >> At 03/20/2012 11:45 PM, Gleb Natapov Wrote:
> >>> On Tue, Mar 20, 2012 at 05:59:16PM +0800, Wen Congyang wrote:
> >>>> At 03/19/2012 03:33 PM, Wen Congyang Wrote:
> >>>>> At 03/08/2012 03:57 PM, Wen Congyang Wrote:
> >>>>>> We can know the guest is paniced when the guest runs on xen.
> >>>>>> But we do not have such feature on kvm.
> >>>>>>
> >>>>>> Another purpose of this feature is: management app(for example:
> >>>>>> libvirt) can do auto dump when the guest is crashed. If management
> >>>>>> app does not do auto dump, the guest's user can do dump by hand if
> >>>>>> he sees the guest is paniced.
> >>>>>>
> >>>>>> I touch the hypervisor instead of using virtio-serial, because
> >>>>>> 1. it is simple
> >>>>>> 2. the virtio-serial is an optional device, and the guest may
> >>>>>>    not have such device.
> >>>>>>
> >>>>>> Changes from v2 to v3:
> >>>>>> 1. correct spelling
> >>>>>>
> >>>>>> Changes from v1 to v2:
> >>>>>> 1. split up host and guest-side changes
> >>>>>> 2. introduce new request flag to avoid changing return values.
> >>>>>> --
> >>>>>> To unsubscribe from this list: send the line "unsubscribe 
> >>>>>> linux-kernel" in
> >>>>>> the body of a message to address@hidden
> >>>>>> More majordomo info at  http://vger.kernel.org/majordomo-info.html
> >>>>>> Please read the FAQ at  http://www.tux.org/lkml/
> >>>>>>
> >>>>>
> >>>>>
> >>>>> Hi all:
> >>>>>
> >>>>> we neet this feature, but we don't decide how to implement it.
> >>>>> We have two solution:
> >>>>> 1. use vmcall
> >>>>> 2. use virtio-serial.
> >>>>>
> >>>>> I will not change this patch set before we decide how to do it.
> >>>>> Can we make a decision recent days?
> >>>>
> >>>> Anybody can decide which solution to use?
> >>>>
> >>> To make an informed decision we need to have at least raw idea how
> >>> virtio-serial variant will look.
> >>
> >> Hmm, I think we can do this:
> >> 1. reset the virtio-serial device or reset the port we use to notice
> >>    the host that guest is panicked.
> >> 2. write some specific messages to the port
> >>
> >> So the port should have fixed name. If this port is opened by the userspace
> >> before the guest is paniced, I am not sure whether we can use it(because a
> >> port only can be opened once at the same time).
> > Yes, IMO we should dedicate one virtio-serial port for panic
> > notifications. Just like we dedicate one for a console.
> > 
> >> We cannot call any function in the module, so we may need to write a simple
> >> driver for virtio-serial(like diskdump's disk driver).
> >>
> > netconsole uses standard NIC drivers in polling mode to send OOPSes
> > over the network and it mostly works. So I think using virtio-serial
> > driver is not out of question, but with IRQ disabled of course.
> 
> The code for netconsole is in which file?
drivers/net/netconsole.c naturally.

> Another question: we cannot call the function in the module directly in the
> kernel.
True I think. netconsole and actual NIC driver have a layer of abstraction
between them. Modules can call functions from other modules. Your
module can register to panic_notifier_list (like IPMI does) and call
functions from virtio-serial. Or panic handling can be added directly
to virtio-serial since it will have to be modified anyway. Something
like netpoll API will have to be added to it.

> > 
> >> I donot know how to implement it now. But I guess that it may be 
> >> complicated.
> >>
> > 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.
> 
> If we implement it by virtio-serial, I agree with passing more useful message
> to host.
> 

--
                        Gleb.