On Wed, Apr 30 2008, Paul E. McKenney wrote:
quoted text
> On Wed, Apr 30, 2008 at 01:35:42PM +0200, Jens Axboe wrote:
> > On Tue, Apr 29 2008, Jeremy Fitzhardinge wrote:
> > > Jens Axboe wrote:
> > > >-int xen_smp_call_function_mask(cpumask_t mask, void (*func)(void *),
> > > >-			       void *info, int wait)
> > > >  
> > > [...]
> > > >-	/* Send a message to other CPUs and wait for them to respond */
> > > >-	xen_send_IPI_mask(mask, XEN_CALL_FUNCTION_VECTOR);
> > > >-
> > > >-	/* Make sure other vcpus get a chance to run if they need to. */
> > > >-	yield = false;
> > > >-	for_each_cpu_mask(cpu, mask)
> > > >-		if (xen_vcpu_stolen(cpu))
> > > >-			yield = true;
> > > >-
> > > >-	if (yield)
> > > >-		HYPERVISOR_sched_op(SCHEDOP_yield, 0);
> > > >  
> > > 
> > > I added this to deal with the case where you're sending an IPI to 
> > > another VCPU which isn't currently running on a real cpu.  In this case 
> > > you could end up spinning while the other VCPU is waiting for a real CPU 
> > > to run on.  (Basically the same problem that spinlocks have in a virtual 
> > > environment.)
> > > 
> > > However, this is at best a partial solution to the problem, and I never 
> > > benchmarked if it really makes a difference.  Since any other virtual 
> > > environment would have the same problem, its best if we can solve it 
> > > generically.  (Of course a synchronous single-target cross-cpu call is a 
> > > simple cross-cpu rpc, which could be implemented very efficiently in the 
> > > host/hypervisor by simply doing a vcpu context switch...)
> > 
> > So, what would your advice be? Seems safe enough to ignore for now and
> > attack it if it becomes a real problem.
> 
> How about an arch-specific function/macro invoked in the spin loop?
> The generic implementation would do nothing, but things like Xen
> could implement as above.

Xen could just stuff that bit into its arch_send_call_function_ipi(),
something like the below should be fine. My question to Jeremy was more
of the order of whether it should be kept or not, I guess it's safer to
just keep it and retain the existing behaviour (and let Jeremy/others
evaluate it at will later on). Note that I got rid of the yield bool and
break when we called the hypervisor.

Jeremy, shall I add this?

diff --git a/arch/x86/xen/smp.c b/arch/x86/xen/smp.c
index 2dfe093..064e6dc 100644
--- a/arch/x86/xen/smp.c
+++ b/arch/x86/xen/smp.c
@@ -352,7 +352,17 @@ static void xen_send_IPI_mask(cpumask_t mask, enum ipi_vector vector)
 
 void xen_smp_send_call_function_ipi(cpumask_t mask)
 {
+	int cpu;
+
 	xen_send_IPI_mask(mask, XEN_CALL_FUNCTION_VECTOR);
+
+	/* Make sure other vcpus get a chance to run if they need to. */
+	for_each_cpu_mask(cpu, mask) {
+		if (xen_vcpu_stolen(cpu)) {
+			HYPERVISOR_sched_op(SCHEDOP_yield, 0);
+			break;
+		}
+	}
 }
 
 void xen_smp_send_call_function_single_ipi(int cpu)

-- 
Jens Axboe

--
unsubscribe notice
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/