On Wed, 2018-01-17 at 18:26 +0100, David Woodhouse wrote:
>
> > In both switching to idle, and back to the vCPU, we should hit this
> > case and not the 'else' case that does the IBPB:
> >
> > 1710 if ( (per_cpu(curr_vcpu, cpu) == next) ||
> > 1711 (is_idle_domain(nextd) && cpu_online(cpu)) )
> > 1712 {
> > 1713 local_irq_enable();
> > 1714 }
>
> I tested that; it doesn't seem to be the case. We end up here with prev
> being the idle thread, next being the actual vCPU, and
> per_cpu(curr_vcpu, cpu) being the idle thread too. So we still do the
> IBPB even when we've just switch from a given vCPU to idle and back
> again.
>
> That's not actually tested on Xen master, but the code here looks very
> much the same as what I actually did test.
This appears to make the excessive IBPBs go away. There might be better
approaches.
diff --git a/xen/arch/x86/domain.c b/xen/arch/x86/domain.c
index 04e9902..b8a9d54 100644
--- a/xen/arch/x86/domain.c
+++ b/xen/arch/x86/domain.c
@@ -68,6 +68,7 @@
#include <asm/pv/mm.h>
#include <asm/spec_ctrl.h>
+DEFINE_PER_CPU(struct vcpu *, last_vcpu); /* Last non-idle vCPU */
DEFINE_PER_CPU(struct vcpu *, curr_vcpu);
static void default_idle(void);
@@ -1745,8 +1746,14 @@ void context_switch(struct vcpu *prev, struct vcpu *next)
ctxt_switch_levelling(next);
- if ( opt_ibpb )
+ /* IBPB on switching to a non-idle vCPU, if that vCPU was not
+ * the last one to be scheduled on this pCPU */
+ if ( opt_ibpb && !is_idle_cpu(next) &&
+ per_cpu(last_vcpu, cpu) != next )
+ {
+ per_cpu(last_vcpu, cpu) = next;
wrmsrl(MSR_PRED_CMD, PRED_CMD_IBPB);
+ }
}
context_saved(prev);
smime.p7s
Description: S/MIME cryptographic signature
_______________________________________________ Xen-devel mailing list Xen-devel@lists.xenproject.org https://lists.xenproject.org/mailman/listinfo/xen-devel
