On 09/28/2011 03:37 PM, Don Zickus wrote:
> On Wed, Sep 28, 2011 at 12:31:40PM +0200, Robert Richter wrote:
> >  On 23.09.11 15:17:13, Don Zickus wrote:
> >  >  @@ -89,6 +89,15 @@ static int notrace __kprobes nmi_handle(unsigned int type, struct pt_regs *regs)
> >  >
> >  >   		handled += a->handler(type, regs);
> >  >
> >  >  +		/*
> >  >  + 		 * Optimization: only loop once if this is not a
> >  >  + 		 * back-to-back NMI.  The idea is nothing is dropped
> >  >  + 		 * on the first NMI, only on the second of a back-to-back
> >  >  + 		 * NMI.  No need to waste cycles going through all the
> >  >  + 		 * handlers.
> >  >  + 		 */
> >  >  +		if (!b2b&&  handled)
> >  >  +			break;
> >
> >  I don't think we can leave this in. As said, there are cases that 2
> >  nmis trigger but the handler is called only once. Only the first would
> >  be handled then, and the second get lost cause there is no 2nd nmi
> >  call.
>
> Right.  Avi, Jeremy what was your objection that needed this optimization
> in the first place?
>

First, iterating over all NMI sources is going to be slow, and a lot 
more so in a guest.  This reduces the performance of perf.

Second, I wanted to use NMIs as a way of waking up a vcpu sleeping with 
interrupts disabled (in the context of Jeremy's paravirt spinlock 
patches).  Looks like we'll have to use paravirtualization for that.

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