On Sat, 2011-01-29 at 09:56 +0100, Jan Kiszka wrote:
> On 2011-01-29 03:07, Glauber Costa wrote:
> > On Fri, 2011-01-28 at 22:09 +0100, Jan Kiszka wrote:
> >> On 2011-01-28 20:48, Glauber Costa wrote:
> >>> Up to know, we were relying on guest cooperation to turn off kvmclock.
> >>> I just realized that even though this is fine and nice, a more robust
> >>> method is to (also) turn it off on vcpu_reset on the hypervisor side.
> >>> This will protect us against reboots, and we don't expect the guest
> >>> to reset its cpu during normal operation anyway.
> >>>
> >>> Signed-off-by: Glauber Costa <[email protected]>
> >>> ---
> >>> arch/x86/kvm/x86.c | 5 +++++
> >>> 1 files changed, 5 insertions(+), 0 deletions(-)
> >>>
> >>> diff --git a/arch/x86/kvm/x86.c b/arch/x86/kvm/x86.c
> >>> index bcc0efc..38b55b3 100644
> >>> --- a/arch/x86/kvm/x86.c
> >>> +++ b/arch/x86/kvm/x86.c
> >>> @@ -5878,6 +5878,11 @@ int kvm_arch_vcpu_reset(struct kvm_vcpu *vcpu)
> >>> kvm_make_request(KVM_REQ_EVENT, vcpu);
> >>> vcpu->arch.apf.msr_val = 0;
> >>>
> >>> + if (vcpu->arch.time_page) {
> >>> + kvm_release_page_dirty(vcpu->arch.time_page);
> >>> + vcpu->arch.time_page = NULL;
> >>> + }
> >>> +
> >>
> >> kvm_arch_vcpu_reset is only called on vcpu setup and when it receives a
> >> sipi (provided in-kernel irqchip is in use). If you want this page to be
> >> consistently reset on guest reboot, you have to trigger this from user
> >> space. But I thought we are doing this already in qemu, don't we?
> >
> > Humm, you might as well be right regarding reboots.
> > But in the end, it doesn't affect correctness here. If we're resetting
> > the vcpu, we should not let that kind of data live.
> >
>
> Right, just checked that we reset other states like nmi_pending or
> async_pf here as well. So doing the same for the time_page looks
> appropriate.
>
> But I think you should encapsulate the pattern above in a function and
> substitute other occurrences at this chance. Also, the changelog should
> clarify in which cases the code matters.
Fair. I will respin this today.
> Jan
>