Re: [PATCH v3 1/6] KVM: Don't use KVM_REQ_PENDING_TIMER

[Avi Kivity <avi@redhat.com>]

On 07/10/2012 11:50 AM, Gleb Natapov wrote:
> On Mon, Jul 09, 2012 at 08:05:40PM +0300, Avi Kivity wrote:
> > It's a write-only bit, set by the timer and cleared by the main loop.
> > Remove it.  Retain the definition since ppc uses it.
> > 
> > Signed-off-by: Avi Kivity <avi@redhat.com>
> > ---
> >  arch/x86/kvm/timer.c | 8 ++------
> >  arch/x86/kvm/x86.c   | 1 -
> >  2 files changed, 2 insertions(+), 7 deletions(-)
> > 
> > diff --git a/arch/x86/kvm/timer.c b/arch/x86/kvm/timer.c
> > index 6b85cc6..c28f838 100644
> > --- a/arch/x86/kvm/timer.c
> > +++ b/arch/x86/kvm/timer.c
> > @@ -27,14 +27,10 @@ enum hrtimer_restart kvm_timer_fn(struct hrtimer *data)
> >  	/*
> >  	 * There is a race window between reading and incrementing, but we do
> >  	 * not care about potentially losing timer events in the !reinject
> > -	 * case anyway. Note: KVM_REQ_PENDING_TIMER is implicitly checked
> > -	 * in vcpu_enter_guest.
> So what about the part that says "KVM_REQ_PENDING_TIMER is implicitly
> checked in vcpu_enter_guest"? This patch drops the check. So now if
> timer triggers while vcpu is in userspace it will enter guest mode
> without injecting timer interrupt on the next ioctl(KVM_RUN). Or do I
> miss something here?

You're right, the bit appears to be write-only, but it isn't.  The check
inside the critical section for vcpu->requests reads it.

I guess we can make the check explicit by doing a 'goto out' if the bit
is set.


-- 
I have a truly marvellous patch that fixes the bug which this
signature is too narrow to contain.