Re: [Qemu-devel] [kvm-unit-tests PATCH v13 3/4] arm: pmu: Check cycle count increases

[Andre Przywara <andre.przywara@arm.com>]

Hi,

On 01/12/16 05:16, Wei Huang wrote:
> From: Christopher Covington <cov@codeaurora.org>
> 
> Ensure that reads of the PMCCNTR_EL0 are monotonically increasing,
> even for the smallest delta of two subsequent reads.
> 
> Signed-off-by: Christopher Covington <cov@codeaurora.org>
> Signed-off-by: Wei Huang <wei@redhat.com>
> Reviewed-by: Andrew Jones <drjones@redhat.com>
> ---
>  arm/pmu.c | 94 +++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
>  1 file changed, 94 insertions(+)
> 
> diff --git a/arm/pmu.c b/arm/pmu.c
> index 1fe2b1a..3566a27 100644
> --- a/arm/pmu.c
> +++ b/arm/pmu.c
> @@ -16,6 +16,9 @@
>  #include "asm/barrier.h"
>  #include "asm/processor.h"
>  
> +#define PMU_PMCR_E         (1 << 0)
> +#define PMU_PMCR_C         (1 << 2)
> +#define PMU_PMCR_LC        (1 << 6)
>  #define PMU_PMCR_N_SHIFT   11
>  #define PMU_PMCR_N_MASK    0x1f
>  #define PMU_PMCR_ID_SHIFT  16
> @@ -23,10 +26,57 @@
>  #define PMU_PMCR_IMP_SHIFT 24
>  #define PMU_PMCR_IMP_MASK  0xff
>  
> +#define ID_DFR0_PERFMON_SHIFT 24
> +#define ID_DFR0_PERFMON_MASK  0xf
> +
> +#define PMU_CYCLE_IDX         31
> +
> +#define NR_SAMPLES 10
> +
> +static unsigned int pmu_version;
>  #if defined(__arm__)
>  DEFINE_GET_SYSREG32(pmcr, 0, c9, c12, 0)
> +DEFINE_SET_SYSREG32(pmcr, 0, c9, c12, 0)
> +DEFINE_GET_SYSREG32(id_dfr0, 0, c0, c1, 2)
> +DEFINE_SET_SYSREG32(pmselr, 0, c9, c12, 5)
> +DEFINE_SET_SYSREG32(pmxevtyper, 0, c9, c13, 1)
> +DEFINE_GET_SYSREG32(pmccntr32, 0, c9, c13, 0)
> +DEFINE_SET_SYSREG32(pmccntr32, 0, c9, c13, 0)
> +DEFINE_GET_SYSREG64(pmccntr64, 0, c9)
> +DEFINE_SET_SYSREG64(pmccntr64, 0, c9)
> +DEFINE_SET_SYSREG32(pmcntenset, 0, c9, c12, 1)
> +
> +static inline uint64_t get_pmccntr(void)
> +{
> +	if (pmu_version == 0x3)
> +		return get_pmccntr64();
> +	else
> +		return get_pmccntr32();
> +}
> +
> +static inline void set_pmccntr(uint64_t value)
> +{
> +	if (pmu_version == 0x3)
> +		set_pmccntr64(value);
> +	else
> +		set_pmccntr32(value & 0xffffffff);
> +}
> +
> +/* PMCCFILTR is an obsolete name for PMXEVTYPER31 in ARMv7 */
> +static inline void set_pmccfiltr(uint32_t value)
> +{
> +	set_pmselr(PMU_CYCLE_IDX);
> +	set_pmxevtyper(value);
> +	isb();
> +}
>  #elif defined(__aarch64__)
>  DEFINE_GET_SYSREG32(pmcr, el0)
> +DEFINE_SET_SYSREG32(pmcr, el0)
> +DEFINE_GET_SYSREG32(id_dfr0, el1)
> +DEFINE_GET_SYSREG64(pmccntr, el0);
> +DEFINE_SET_SYSREG64(pmccntr, el0);
> +DEFINE_SET_SYSREG32(pmcntenset, el0);
> +DEFINE_SET_SYSREG32(pmccfiltr, el0);
>  #endif
>  
>  /*
> @@ -52,11 +102,55 @@ static bool check_pmcr(void)
>  	return ((pmcr >> PMU_PMCR_IMP_SHIFT) & PMU_PMCR_IMP_MASK) != 0;
>  }
>  
> +/*
> + * Ensure that the cycle counter progresses between back-to-back reads.
> + */
> +static bool check_cycles_increase(void)
> +{
> +	bool success = true;
> +
> +	/* init before event access, this test only cares about cycle count */
> +	set_pmcntenset(1 << PMU_CYCLE_IDX);
> +	set_pmccfiltr(0); /* count cycles in EL0, EL1, but not EL2 */
> +	set_pmccntr(0);

Why do we need this? Shouldn't PMU_PMCR_C below take care of that?

> +
> +	set_pmcr(get_pmcr() | PMU_PMCR_LC | PMU_PMCR_C | PMU_PMCR_E);
> +
> +	for (int i = 0; i < NR_SAMPLES; i++) {
> +		uint64_t a, b;
> +
> +		a = get_pmccntr();
> +		b = get_pmccntr();
> +
> +		if (a >= b) {
> +			printf("Read %"PRId64" then %"PRId64".\n", a, b);
> +			success = false;
> +			break;
> +		}
> +	}
> +
> +	set_pmcr(get_pmcr() & ~PMU_PMCR_E);
> +
> +	return success;
> +}
> +
> +void pmu_init(void)

Mmh, this function doesn't really initialize anything, does it?
Should it be named pmu_available() or pmu_version() or the like?

And should we bail out early here (or rather at the caller) if this
register reports that no PMU is available? For instance by making it
return a boolean?

> +{
> +	uint32_t dfr0;
> +
> +	/* probe pmu version */
> +	dfr0 = get_id_dfr0();
> +	pmu_version = (dfr0 >> ID_DFR0_PERFMON_SHIFT) & ID_DFR0_PERFMON_MASK;
> +	report_info("PMU version: %d", pmu_version);
> +}
> +
>  int main(void)
>  {
>  	report_prefix_push("pmu");
>  
> +	pmu_init();
>  	report("Control register", check_pmcr());
> +	report("Monotonically increasing cycle count", check_cycles_increase());

I wonder if we should skip this test if check_pmcr() has returned false
before? We let it return a boolean, so it seems quite natural to use
this information here.
This would avoid a lot of false FAILs due to the PMU not being available
(because QEMU is too old, for instance).

Cheers,
Andre.