Re: [RFC] KVM: Fix simultaneous NMIs

[Avi Kivity <avi@redhat.com>]

On 09/19/2011 06:22 PM, Marcelo Tosatti wrote:
> On Mon, Sep 19, 2011 at 06:09:39PM +0300, Avi Kivity wrote:
> >  On 09/19/2011 05:54 PM, Marcelo Tosatti wrote:
> >  >On Mon, Sep 19, 2011 at 05:30:27PM +0300, Avi Kivity wrote:
> >  >>   On 09/19/2011 04:54 PM, Marcelo Tosatti wrote:
> >  >>   >>    >>
> >  >>   >>    >>     Yes, due to NMI-blocked-by-STI.  A really touchy area.
> >  >>   >>    >And we don't need the window exit notification then? I don't understand
> >  >>   >>    >what nmi_in_progress is supposed to do here.
> >  >>   >>
> >  >>   >>    We need the window notification in both cases.  If we're recovering
> >  >>   >>    from STI, then we don't need to collapse NMIs.  If we're completing
> >  >>   >>    an NMI handler, then we do need to collapse NMIs (since the queue
> >  >>   >>    length is two, and we just completed one).
> >  >>   >
> >  >>   >I don't understand what is the point with nmi_in_progress, and the above
> >  >>   >hunk, either. Can't inject_nmi do:
> >  >>   >
> >  >>   >if (nmi_injected + atomic_read(nmi_pending)<    2)
> >  >>   >       atomic_inc(nmi_pending)
> >  >>   >
> >  >>   >Instead of collapsing somewhere else?
> >  >>
> >  >>   We could.  It's not atomic though - two threads executing in
> >  >>   parallel could raise the value to three.  Could do a cmpxchg loop
> >  >>   does an increment bounded to two.  I guess this is a lot clearer,
> >  >>   thanks.
> >  >>
> >  >>   >You'd also have to change
> >  >>   >nmi_injected handling in arch code so its value is not "hidden", in
> >  >>   >complete_interrupts().
> >  >>
> >  >>   Or maybe make raising nmi_injected not decrement nmi_pending.  So:
> >  >>
> >  >>     nmi_pending: total number of interrupts in queue
> >  >>     nmi_injected: of these, how many are currently being injected
> >  >>
> >  >>   yes?
> >  >
> >  >Yes, at the expense of decrementing on subarch code (which is fine,
> >  >apparently).
> >  >
> >
> >  Hm, we have no place to decrement.
>
> Decrement when setting nmi_injected = false, increment when setting
> nmi_injected = true, in vmx/svm.c.

That gives a queue length of 3: one running nmi and nmi_pending = 2.

> >   We need to do that when IRET
> >  executes, but we don't want to request an NMI window exit in the
> >  common case of nmi_pending = 1.
>
> Do not enable nmi window if nmi_injected = true?
>

We have to, since we need a back-to-back nmi if the queue length > 1 
(including the running nmi).

-- 
error compiling committee.c: too many arguments to function