Re: [Qemu-devel] [PATCH] kvm: notify host when guest paniced

[Avi Kivity <avi@redhat.com>]

On 02/29/2012 12:19 PM, Daniel P. Berrange wrote:
> On Wed, Feb 29, 2012 at 12:05:32PM +0200, Avi Kivity wrote:
> > On 02/29/2012 11:58 AM, Daniel P. Berrange wrote:
> > > > 
> > > > How about using a virtio-serial channel for this?  You can transfer any
> > > > amount of information (including the dump itself).
> > >
> > > When the guest OS has crashed, any dumps will be done from the host
> > > OS using libvirt's core dump mechanism. The guest OS isn't involved
> > > and is likely too dead to be of any use anyway. Likewise it is
> > > quite probably too dead to work a virtio-serial channel or any
> > > similarly complex device. We're really just after the simplest
> > > possible notification that the guest kernel has paniced.
> > 
> > If it's alive enough to panic, it's alive enough to kexec its kdump
> > kernel.  After that it can do anything.
> > 
> > Guest-internal dumps are more useful IMO that host-initiated dumps.  In
> > a cloud, the host-initiated dump is left on the host, outside the reach
> > of the guest admin, outside the guest image where all the symbols are,
> > and sometimes not even on the same host if a live migration occurred. 
> > It's more useful in small setups, or if the problem is in the
> > hypervisor, not the guest.
>
> I don't think guest vs host dumps should be considered mutually exclusive,
> they both have pluses+minuses.

True.

> Configuring kexec+kdump requires non-negligable guest admin configuration
> work before it's usable, and this work is guest OS specific, if it is possible
> at all. 

I think it's on by default on Windows and requires installing a package
on Linux, which may be part of the default configuration on many distros.

> A permanent panic notifier that's built in the kernel by default
> requires zero guest admin config, and can allow host admin to automate
> collection of dumps across all their hosts/guests. The KVM hypercall
> notification is fairly trivially ported to any OS kernel, by comparison
> with a full virtio + virtio-serial impl.

That's the path of least resistance.  But it's not necessarily the best
path.  We end up with a wide set of disconnected ABIs instead of a
narrow set that is more flexible.

-- 
error compiling committee.c: too many arguments to function