On Fri, Nov 11, 2016 at 01:55:49PM -0600, Wei Huang wrote:
>
>
> On 11/11/2016 01:43 AM, Andrew Jones wrote:
> > On Tue, Nov 08, 2016 at 12:17:14PM -0600, Wei Huang wrote:
> >> From: Christopher Covington <c...@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 <c...@codeaurora.org>
> >> Signed-off-by: Wei Huang <w...@redhat.com>
> >> ---
> >> arm/pmu.c | 98
> >> +++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
> >> 1 file changed, 98 insertions(+)
> >>
> >> diff --git a/arm/pmu.c b/arm/pmu.c
> >> index 0b29088..d5e3ac3 100644
> >> --- a/arm/pmu.c
> >> +++ b/arm/pmu.c
> >> @@ -14,6 +14,7 @@
> >> */
> >> #include "libcflat.h"
> >>
> >> +#define PMU_PMCR_E (1 << 0)
> >> #define PMU_PMCR_N_SHIFT 11
> >> #define PMU_PMCR_N_MASK 0x1f
> >> #define PMU_PMCR_ID_SHIFT 16
> >> @@ -21,6 +22,10 @@
> >> #define PMU_PMCR_IMP_SHIFT 24
> >> #define PMU_PMCR_IMP_MASK 0xff
> >>
> >> +#define PMU_CYCLE_IDX 31
> >> +
> >> +#define NR_SAMPLES 10
> >> +
> >> #if defined(__arm__)
> >> static inline uint32_t pmcr_read(void)
> >> {
> >> @@ -29,6 +34,47 @@ static inline uint32_t pmcr_read(void)
> >> asm volatile("mrc p15, 0, %0, c9, c12, 0" : "=r" (ret));
> >> return ret;
> >> }
> >> +
> >> +static inline void pmcr_write(uint32_t value)
> >> +{
> >> + asm volatile("mcr p15, 0, %0, c9, c12, 0" : : "r" (value));
> >> +}
> >> +
> >> +static inline void pmselr_write(uint32_t value)
> >> +{
> >> + asm volatile("mcr p15, 0, %0, c9, c12, 5" : : "r" (value));
> >> +}
> >> +
> >> +static inline void pmxevtyper_write(uint32_t value)
> >> +{
> >> + asm volatile("mcr p15, 0, %0, c9, c13, 1" : : "r" (value));
> >> +}
> >> +
> >> +/*
> >> + * While PMCCNTR can be accessed as a 64 bit coprocessor register,
> >> returning 64
> >> + * bits doesn't seem worth the trouble when differential usage of the
> >> result is
> >> + * expected (with differences that can easily fit in 32 bits). So just
> >> return
> >> + * the lower 32 bits of the cycle count in AArch32.
> >
> > Like I said in the last review, I'd rather we not do this. We should
> > return the full value and then the test case should confirm the upper
> > 32 bits are zero.
> >
>
> Unless I miss something in ARM documentation, ARMv7 PMCCNTR is a 32-bit
> register. We can force it to a more coarse-grained cycle counter with
> PMCR.D bit=1 (see below). But it is still not a 64-bit register. ARMv8
> PMCCNTR_EL0 is a 64-bit register.
>
> "The PMCR.D bit configures whether PMCCNTR increments once every clock
> cycle, or once every 64 clock cycles. "
>
> So I think the comment above in the code is an overstatement, which
> should be deleted or moved down to ARMv8 pmccntr_read() below.
OK, please fix as appropriate, but for the v8 64-bit register, please
don't drop the upper bits until after a unit test has a chance to check
them.
Thanks,
drew
>
> >> + */
> >> +static inline uint32_t pmccntr_read(void)
> >> +{
> >> + uint32_t cycles;
> >> +
> >> + asm volatile("mrc p15, 0, %0, c9, c13, 0" : "=r" (cycles));
> >> + return cycles;
> >> +}
> >> +
> >> +static inline void pmcntenset_write(uint32_t value)
> >> +{
> >> + asm volatile("mcr p15, 0, %0, c9, c12, 1" : : "r" (value));
> >> +}
> >> +
> >> +/* PMCCFILTR is an obsolete name for PMXEVTYPER31 in ARMv7 */
> >> +static inline void pmccfiltr_write(uint32_t value)
> >> +{
> >> + pmselr_write(PMU_CYCLE_IDX);
> >> + pmxevtyper_write(value);
> >> +}
> >> #elif defined(__aarch64__)
> >> static inline uint32_t pmcr_read(void)
> >> {
> >> @@ -37,6 +83,29 @@ static inline uint32_t pmcr_read(void)
> >> asm volatile("mrs %0, pmcr_el0" : "=r" (ret));
> >> return ret;
> >> }
> >> +
> >> +static inline void pmcr_write(uint32_t value)
> >> +{
> >> + asm volatile("msr pmcr_el0, %0" : : "r" (value));
> >> +}
> >> +
> >> +static inline uint32_t pmccntr_read(void)
> >> +{
> >> + uint32_t cycles;
> >> +
> >> + asm volatile("mrs %0, pmccntr_el0" : "=r" (cycles));
> >> + return cycles;
> >> +}
> >> +
> >> +static inline void pmcntenset_write(uint32_t value)
> >> +{
> >> + asm volatile("msr pmcntenset_el0, %0" : : "r" (value));
> >> +}
> >> +
> >> +static inline void pmccfiltr_write(uint32_t value)
> >> +{
> >> + asm volatile("msr pmccfiltr_el0, %0" : : "r" (value));
> >> +}
> >> #endif
> >>
> >> /*
> >> @@ -63,11 +132,40 @@ 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)
> >> +{
> >> + pmcr_write(pmcr_read() | PMU_PMCR_E);
> >> +
> >> + for (int i = 0; i < NR_SAMPLES; i++) {
> >> + unsigned long a, b;
> >> +
> >> + a = pmccntr_read();
> >> + b = pmccntr_read();
> >> +
> >> + if (a >= b) {
> >> + printf("Read %ld then %ld.\n", a, b);
> >> + return false;
> >> + }
> >> + }
> >> +
> >> + pmcr_write(pmcr_read() & ~PMU_PMCR_E);
> >> +
> >> + return true;
> >> +}
> >> +
> >> int main(void)
> >> {
> >> report_prefix_push("pmu");
> >>
> >> + /* init for PMU event access, right now only care about cycle count */
> >> + pmcntenset_write(1 << PMU_CYCLE_IDX);
> >> + pmccfiltr_write(0); /* count cycles in EL0, EL1, but not EL2 */
> >> +
> >> report("Control register", check_pmcr());
> >> + report("Monotonically increasing cycle count", check_cycles_increase());
> >>
> >> return report_summary();
> >> }
> >> --
> >> 1.8.3.1
> >
> > Besides needing to use u64's for registers that return u64's, it
> > looks good to me.
> >
> > drew
> >
>
