On Tue, Jul 21, 2015 at 09:16:57AM +0800, Shannon Zhao wrote:

On 2015/7/17 18:21, Christoffer Dall wrote:

quoted

On Fri, Jul 17, 2015 at 04:45:44PM +0800, Shannon Zhao wrote:

quoted

On 2015/7/17 3:55, Christoffer Dall wrote:

quoted

On Mon, Jul 06, 2015 at 10:17:34AM +0800, shannon.zhao at linaro.org wrote:

quoted

From: Shannon Zhao <redacted>

Add reset handler which gets host value of PMCR_EL0 and make writable
bits architecturally UNKNOWN. Add access handler which emulates
writing and reading PMCR_EL0 register.

Signed-off-by: Shannon Zhao <redacted>
---
 arch/arm64/kvm/sys_regs.c | 41 ++++++++++++++++++++++++++++++++++++++++-
 1 file changed, 40 insertions(+), 1 deletion(-)

diff --git a/arch/arm64/kvm/sys_regs.c b/arch/arm64/kvm/sys_regs.c
index c370b40..152ee17 100644
--- a/arch/arm64/kvm/sys_regs.c
+++ b/arch/arm64/kvm/sys_regs.c

@@ -33,6 +33,7 @@
 #include <asm/kvm_emulate.h>
 #include <asm/kvm_host.h>
 #include <asm/kvm_mmu.h>
+#include <asm/pmu.h>
 
 #include <trace/events/kvm.h>
 

@@ -236,6 +237,44 @@ static void reset_mpidr(struct kvm_vcpu *vcpu, const struct sys_reg_desc *r)
 	vcpu_sys_reg(vcpu, MPIDR_EL1) = (1ULL << 31) | mpidr;
 }
 
+static void reset_pmcr_el0(struct kvm_vcpu *vcpu, const struct sys_reg_desc *r)
+{
+	u32 pmcr;
+
+	asm volatile("mrs %0, pmcr_el0\n" : "=r" (pmcr));
+	vcpu_sys_reg(vcpu, PMCR_EL0) = (pmcr & ~ARMV8_PMCR_MASK)
+				       | (ARMV8_PMCR_MASK & 0xdecafbad);

You could add a comment that this resets to UNKNOWN as to not make
people confused about the pseudo-random hex value.

Ok.

quoted

Have we thought about whether we want to tell the guest that it has the
same PMU as available on the real hardware, or does the virtualization
layer suggest to us that we should adjust this somehow?

I guess here the number of PMU counters is what we can adjust, right?
Are we worried about that the host will run out of counters when guest
and host register lots of events?

that's what I wonder; if perf itself reserves a counter for example,
then we'll at best be able to measure with N-1 counters for the guest (N
being the number of counters on the physical CPU), so why tell the guest
we have N counters?

I'm not sure whether perf itself reserves one counter. Here I just pass
the hardware information to guest.

quoted

Of course, we can never even be guaranteed to have N-1 counters
avaialable either, so maybe the sane choice is just to tell the guest
what kind of hardware we have, and then fiddle the best we can with the
remaining counters?  After all, correct functionality of the guest
doesn't depend on this, it's a best-effort kind of thing...

Thoughts?

quoted

The PMU of cortex-a57 has 6 counters. IIUC, if one of the guest vcpu
process registers 6 events and some host process register 6 events too,
these events will be monitored in real hardware PMU counter when the
related process runs on the cpu. And when other processes are scheduled
to run, it will switch the contexts of PMU counters.

That depends on the way the counters are used by perf I think.  What if
you have system wide events?  What if the KVM (vcpu) process itself is
being monitored for some events?

Not sure what will happen when the number of monitored events is greater
than counters. I guess the perf layer may balance the events?

I'm not sure either, but we need to have thought about these things, I
suppose.  You should probably look at the code or try it out.

Thanks,
-Christoffer