Re: KVM Test result, kernel 873c05f.., userspace d102d75..

[Marcelo Tosatti <mtosatti@redhat.com>]

On Thu, Apr 24, 2008 at 04:44:27PM +0300, Avi Kivity wrote:
> Chris Lalancette wrote:
> >Avi Kivity wrote:
> >  
> >>Ok.  __pit_timer_fn() is called from an interrupt, which then calls 
> >>smp_call_function_single(), which calls spin_lock().  If we've already 
> >>taken the lock, we hang.
> >>
> >>    
> >
> >Ah.  Just adding a "me too"; I didn't get a chance to debug it yesterday, 
> >but I
> >was seeing similar problems.  If I disabled in-kernel pit with 
> >-no-kvm-pit, all
> >was well.
> >  
> 
> How to fix it, though?  the only idea that comes to mind is to affine 
> the hrtimer with vcpu0 (like the local apic timers) which would mean we 
> only need to unwait the waitqueue, and never need to send the IPI.  
> Would slightly improve performance as well.

Yes, agree. 

For now I think just revert

--- a/arch/x86/kvm/i8254.c
+++ b/arch/x86/kvm/i8254.c
@@ -200,10 +200,8 @@ int __pit_timer_fn(struct kvm_kpit_state *ps)

        atomic_inc(&pt->pending);
        smp_mb__after_atomic_inc();
-       if (vcpu0 && waitqueue_active(&vcpu0->wq)) {
-               vcpu0->arch.mp_state = KVM_MP_STATE_RUNNABLE;
-               wake_up_interruptible(&vcpu0->wq);
-       }
+       if (vcpu0)
+               kvm_vcpu_kick(vcpu0);

And add a big fat FIXME.

-------------------------------------------------------------------------
This SF.net email is sponsored by the 2008 JavaOne(SM) Conference 
Don't miss this year's exciting event. There's still time to save $100. 
Use priority code J8TL2D2. 
http://ad.doubleclick.net/clk;198757673;13503038;p?http://java.sun.com/javaone