On 23/02/16 01:46, Shannon Zhao wrote:

On 2016/2/23 1:51, Marc Zyngier wrote:

quoted

On 22/02/16 09:37, Shannon Zhao wrote:

quoted

From: Shannon Zhao <redacted>

We are about to trap and emulate accesses to each PMU register
individually. This adds the context offsets for the AArch64 PMU
registers.

Signed-off-by: Shannon Zhao <redacted>
Reviewed-by: Marc Zyngier <redacted>
Reviewed-by: Andrew Jones <redacted>
---
 arch/arm64/include/asm/kvm_host.h | 15 +++++++++++++++
 1 file changed, 15 insertions(+)

diff --git a/arch/arm64/include/asm/kvm_host.h b/arch/arm64/include/asm/kvm_host.h
index 6f0241f..6bab7fb 100644
--- a/arch/arm64/include/asm/kvm_host.h
+++ b/arch/arm64/include/asm/kvm_host.h

@@ -115,6 +115,21 @@ enum vcpu_sysreg {
 	MDSCR_EL1,	/* Monitor Debug System Control Register */
 	MDCCINT_EL1,	/* Monitor Debug Comms Channel Interrupt Enable Reg */
 
+	/* Performance Monitors Registers */
+	PMCR_EL0,	/* Control Register */
+	PMOVSSET_EL0,	/* Overflow Flag Status Set Register */
+	PMSELR_EL0,	/* Event Counter Selection Register */
+	PMEVCNTR0_EL0,	/* Event Counter Register (0-30) */
+	PMEVCNTR30_EL0 = PMEVCNTR0_EL0 + 30,
+	PMCCNTR_EL0,	/* Cycle Counter Register */
+	PMEVTYPER0_EL0,	/* Event Type Register (0-30) */
+	PMEVTYPER30_EL0 = PMEVTYPER0_EL0 + 30,
+	PMCCFILTR_EL0,	/* Cycle Count Filter Register */
+	PMCNTENSET_EL0,	/* Count Enable Set Register */
+	PMINTENSET_EL1,	/* Interrupt Enable Set Register */
+	PMUSERENR_EL0,	/* User Enable Register */
+	PMSWINC_EL0,	/* Software Increment Register */
+

I've just noticed a rather fundamental issue with this: this makes it
impossible to bisect the whole series.

Ah, sorry. Will fix this.

Thanks.

quoted

I was trying to pinpoint a performance regression with this series, and
started bisecting.

You mean this series introduce a performance regression? Is there any
method to measure that? Then I can have a look too.

I'm not quite sure this is related to this series. What I'm observing is
that hackbench runs faster in a VM spawned with kvmtool than the same VM
run with qemu. As in significantly faster (62 vs 55 seconds - 2 vcpus,
1GB memory).

Given that hackbench doesn't do any IO, I don't really see why we should
see a difference. I'll add support for the PMU to kvmtool today, and
I'll be able to see if that has any impact.

	M.
-- 
Jazz is not dead. It just smells funny...