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

[Gleb Natapov]

On Wed, Feb 29, 2012 at 12:08:19PM +0200, Avi Kivity wrote:
> On 02/29/2012 12:05 PM, Gleb Natapov wrote:
> > On Wed, Feb 29, 2012 at 12:00:41PM +0200, Avi Kivity wrote:
> > > On 02/29/2012 11:55 AM, Gleb Natapov wrote:
> > > > > >
> > > > > 
> > > > > How about using a virtio-serial channel for this?  You can transfer 
> > > > > any
> > > > > amount of information (including the dump itself).
> > > > > 
> > > > Isn't it unreliable after the guest panicked? 
> > > 
> > > So is calling hypercalls, or dumping, or writing to the screen.  Of
> > > course calling a hypercall is simpler and so is more reliable.
> > > 
> > Yes, crash can be so severe that it is not even detected by a kernel
> > itself, so not OOPS message even printed. But in most cases if kernel is
> > functional enough to print OOPS it is functional enough to call single
> > hypercall instruction.
> 
> Why not print the oops to virtio-serial?  Or even just a regular serial
> port?  That's what bare metal does.
> 
The more interface is complex the more chances it will fail during
panic. Regular serial is likely more reliable than virtio-serial.
Hypercall is likely more reliable than uart. On serial user can
fake panic notification. Can this be problematic?

> > > > Having special kdump
> > > > kernel that transfers dump to a host via virtio-serial channel though
> > > > sounds interesting. May be that's what you mean.
> > > 
> > > Yes.  The "panic, starting dump" signal should be initiated by the
> > > panicking kernel though, in case the dump fails.
> > > 
> > Then panic hypercall sounds like a reasonable solution.
> 
> It is, but I'm trying to see if we can get away with doing nothing.
> 
Fair enough.

--
                        Gleb.