Re: pv-on-hvm kexec, howto reregister timer interrupt

[Olaf Hering <olaf@aepfle.de>]

On Sat, Jul 23, Ian Campbell wrote:

> On Sat, 2011-07-23 at 07:57 +0100, Keir Fraser wrote:
> > On 22/07/2011 19:49, "Olaf Hering" <olaf@aepfle.de> wrote:
> > 
> > > What is the best way to reregister the timer interrupt in the kexec
> > > kernel in a pv-on-hvm guest? Right now the crash shown below happens
> > > because EVTCHNOP_bind_virq returns -EEXIST. Can the timer on the cpus be
> > > unregistered during shutdown? xen_teardown_timer() does not allow that
> > > for cpu #0.  I added some silly loop which tries to match vcpu/cpu in
> > > bind_virq_to_irq() using EVTCHNOP_status to find the currently used port
> > > number. This helps and booting proceeds, but I feel that cant be the
> > > right approach.
> > 
> > Sounbds like it's simply some Linux-guest issue here to be untangled. All
> > Xen requires you to do is EVTCHNOP_close the old virq-evtchn binding.
> 
> That might be tricky to do in the target kernel, which may not know what
> needs closing. There's EVTCHNOP_reset which looks like it would be a
> sensible thing to call early on in the target but seems like it would
> close to much -- e.g. the xenstore evtchn?

After some more poking around in the reboot code path I have figured it
out, a syscore_ops has to be registerd to shutdown the irqs.
I will post some patches for review.

Olaf