[libvirt] Re: kexec/kdump of a kvm guest?

[Mike Snitzer]

On Thu, Jul 24, 2008 at 9:15 AM, Vivek Goyal <vgoyal at redhat.com> wrote:
> On Thu, Jul 24, 2008 at 07:49:59AM -0400, Mike Snitzer wrote:
>> On Thu, Jul 24, 2008 at 4:39 AM, Alexander Graf <agraf at suse.de> wrote:

>> > As you're stating that the host kernel breaks with kvm modules loaded, maybe
>> > someone there could give a hint.
>>
>> OK, I can try using a newer kernel on the host too (e.g. 2.6.25.x) to
>> see how kexec/kdump of the host fairs when kvm modules are loaded.
>>
>> On the guest side of things, as I mentioned in my original post,
>> kexec/kdump wouldn't work within a 2.6.22.19 guest with the host
>> running 2.6.25.4 (with kvm-70).
>>
>
> Hi Mike,
>
> I have never tried kexec/kdump inside a kvm guest. So I don't know if
> historically they have been working or not.

Avi indicated he seems to remember that at least kexec worked last he
tried (didn't provide when/what he tried though).

> Having said that, Why do we need kdump to work inside the guest? In this
> case qemu should be knowing about the memory of guest kernel and should
> be able to capture a kernel crash dump? I am not sure if qemu already does
> that. If not, then probably we should think about it?
>
> To me, kdump is a good solution for baremetal but not for virtualized
> environment where we already have another piece of software running which
> can do the job for us. We will end up wasting memory in every instance
> of guest (memory reserved for kdump kernel in every guest).

I haven't looked into what mechanics qemu provides for collecting the
entire guest memory image; I'll dig deeper at some point.  It seems
the libvirt mid-layer ("virsh dump" - dump the core of a domain to a
file for analysis) doesn't support saving a kvm guest core:
# virsh dump guest10 guest10.dump
libvir: error : this function is not supported by the hypervisor:
virDomainCoreDump
error: Failed to core dump domain guest10 to guest10.dump

Seems that libvirt functionality isn't available yet with kvm (I'm
using libvirt 0.4.2, I'll give libvirt 0.4.4 a try).  cc'ing the
libvirt-list to get their insight.

That aside, having the crash dump collection be multi-phased really
isn't workable (that is if it requires a crashed guest to be manually
saved after the fact).  The host system _could_ be rebooted; whereby
losing the guest's core image.  So automating qemu and/or libvirtd to
trigger a dump would seem worthwhile (maybe its already done?).

So while I agree with you its ideal to not have to waste memory in
each guest for the purposes of kdump; if users want to model a guest
image as closely as possible to what will be deployed on bare metal it
really would be ideal to support a 1:1 functional equivalent with kvm.
 I work with people who refuse to use kvm because of the lack of
kexec/kdump support.

I can do further research but welcome others' insight: do others have
advice on how best to collect a crashed kvm guest's core?

> It will be interesting to look at your results with 2.6.25.x kernels with
> kvm module inserted. Currently I can't think what can possibly be wrong.

If the host's 2.6.25.4 kernel has both the kvm and kvm-intel modules
loaded kexec/kdump does _not_ work (simply hangs the system).  If I
only have the kvm module loaded kexec/kdump works as expected
(likewise if no kvm modules are loaded at all).  So it would appear
that kvm-intel and kexec are definitely mutually exclusive at the
moment (at least on both 2.6.22.x and 2.6.25.x).

Mike