Re: [PATCH 4/4] KVM: Rework of guest debug state writing

[Marcelo Tosatti <mtosatti@redhat.com>]

On Thu, Feb 04, 2010 at 04:41:44PM +0100, Jan Kiszka wrote:
> Jan Kiszka wrote:
> > Marcelo Tosatti wrote:
> >> On Thu, Feb 04, 2010 at 01:33:50AM +0100, Jan Kiszka wrote:
> >>> Marcelo Tosatti wrote:
> >>>> On Wed, Feb 03, 2010 at 10:29:45PM +0100, Jan Kiszka wrote:
> >>>>> So far we synchronized any dirty VCPU state back into the kernel before
> >>>>> updating the guest debug state. This was a tribute to a deficit in x86
> >>>>> kernels before 2.6.33. But as this is an arch-dependent issue, it is
> >>>>> better handle in the x86 part of KVM and remove the writeback point for
> >>>>> generic code.
> >>>> Jan,
> >>>>
> >>>> This patch breaks migration.
> >>> Can you elaborate what you did? I can't reproduce, and I do not see any
> >>> conceptual issue (given that guest debugging conflicts with migration
> >>> anyway).
> >> kvm-autotest fails (migration only, install is ok, both Linux and Win
> >> guests). Not sure why, perhaps the unconditional KVM_SET_GUEST_DEBUG
> >> corrupts state somehow? 
> >>
> >> Tested with io thread enabled.
> > 
> > That's this default-off thing, so... OK, confirmed, investigating.
> > 
> 
> Heisenbug: It first also popped up (in form of a frozen migration
> target) after removing this patch, but now it's totally unreproducible,
> whatever patch I apply or revert from my series. Base is current master.
>
> I tend to think there is a hidden issue of iothread vs. migration,
> unrelated to this patch.

Probably many :)

Do you have c5f32c99c6855d466737daf1cd262e7e92062f87 (from qemu-kvm.git
uq/master) in?

With kvm-autotest the failure is not sporadic (and the above commit
applied): with KVM_SET_GUEST_DEBUG in arch_put_regs all migration 
tests fail, without, all of them succeed. 

So env->kvm_guest_debug has been zeroed by cpu_x86_init, which means
the writeback via KVM_SET_GUEST_DEBUG does almost nothing. It does
get_rflags and set_rflags in the kernel.

Test box is off, but the synchronous writeback via qemu_system_reset
in main, after machine and vcpu thread initialization, might be
problematic. But it would be nice to understand this.

Unrelated to this problem, won't put_vcpu_events, which is executed 
after KVM_SET_GUEST_DEBUG, overwrite any queued debug exceptions?