Re: [Patch] Enable "sysrq c" handler for domU coredump

[Horms <horms@verge.net.au>]

On Tue, Aug 01, 2006 at 06:06:35PM +0900, Akio Takebe wrote:
> >On Tue, Aug 01, 2006 at 01:39:23PM +0900, Akio Takebe wrote:
> >> Hi, Horms
> >> 
> >> >
> >> >That seems fine to me. Though there was some resistance to a
> >> >patch I sent which adds a panic option to xen-console, which
> >> >is the hypervisor equivalent of sysrq.
> >> >
> >> This patch focus only manually dumping domU's core.
> >> How can I use the panic option of xen-console?
> >> I thought your patch to panic xen.
> >> Can I dump domU's core with the option?
> >
> >Sorry, I was not clear.
> >
> >Yes, my patch is to panic (or more recently cause a crash dump) 
> >in the hypervisor. It does not give any special behaviour to
> >the domains.
> >
> >What I meant was, that the idea of adding a panic to domU's sysrq
> >is similar to my idea of adding panic to the xen console. And I was
> >asked to change this patch to make it trigger a kdump directly,
> >rather than a panic.
> >
> >I think that your sysrq patch and my xen-console panic patches
> >are related, and I wanted to bring that into the discussion.
> >
> Horms, what you said, that I had better call crash_kexec() 
> than panic() on dom0?

Thats not what I was thinking, but now you mention it...

> If so, I agree for dom0 directly calling crash_kexec().
> But because now vmlinux is used as both dom0 and domU,
> we cannot use ifdef for separateing dom0 and domU.
> So I call panic() on both dom0 and domU.
> I also think option like a unknown_nmi_panic is useful 
> and necessary for dump and debug.
> 
> Everyone, how about the below?
> 1. If crash_kexec() is called on dom0, dom0 do hypercall
>    then xen do kexec/kdump
> 2. If crash_kexec() is called on domU, domU do hypercall
>    then xen do panic_domain(), and domU's core is dumped.

I think that is a good idea, because it gives consistent behaviour
for dom0 and domU, even though the mechanics differ. Though if
kexec for domU is merged, things may change at that time.

> So I don't need to modify linux code. :-)

:-)

> Or How about "xm dump"?
> (Probably I use panic_domain() by hypercall.
> This way can probably dump HVM domain.)

That is also a good idea, though it could be done as well,
rather than instead of your crash_kexec() -> hypercall idea.

> I think we need to have the way manually to dump domU, or Xen.
> (Because for the time when domU or Xen don't panic 
> and spinloop and so on) 

That is a good point, which makes your xm dump idea somewhat
more attactive than the crash_kexec() -> hypercall idea.
But I still think they both make sense.

-- 
Horms
  H: http://www.vergenet.net/~horms/
  W: http://www.valinux.co.jp/en/