Re: [PATCH v7 1/9] perf pmu: Add #slots literal support for arm64

[Ian Rogers]

On Sun, Jan 15, 2023 at 6:58 PM Jing Zhang <renyu.zj@xxxxxxxxxxxxxxxxx> wrote:
>
> 在 2023/1/15 上午6:15, Ian Rogers 写道:
> > On Fri, Jan 13, 2023 at 1:22 AM Jing Zhang <renyu.zj@xxxxxxxxxxxxxxxxx> wrote:
> >>
> >> The slots in each architecture may be different, so add #slots literal
> >> to obtain the slots of different architectures, and the #slots can be
> >> applied in the metric. Currently, The #slots just support for arm64,
> >> and other architectures will return NAN.
> >>
> >> On arm64, the value of slots is from the register PMMIR_EL1.SLOT, which
> >> I can read in /sys/bus/event_source/device/armv8_pmuv3_*/caps/slots.
> >> PMMIR_EL1.SLOT might read as zero if the PMU version is lower than
> >> ID_AA64DFR0_EL1_PMUVer_V3P4 or the STALL_SLOT event is not implemented.
> >>
> >> Signed-off-by: Jing Zhang <renyu.zj@xxxxxxxxxxxxxxxxx>
> >> ---
> >>  tools/perf/arch/arm64/util/pmu.c | 34 ++++++++++++++++++++++++++++++++--
> >>  tools/perf/util/expr.c           |  5 +++++
> >>  tools/perf/util/pmu.c            |  6 ++++++
> >>  tools/perf/util/pmu.h            |  1 +
> >>  4 files changed, 44 insertions(+), 2 deletions(-)
> >>
> >> diff --git a/tools/perf/arch/arm64/util/pmu.c b/tools/perf/arch/arm64/util/pmu.c
> >> index 477e513..5f8667b 100644
> >> --- a/tools/perf/arch/arm64/util/pmu.c
> >> +++ b/tools/perf/arch/arm64/util/pmu.c
> >> @@ -3,8 +3,9 @@
> >>  #include <internal/cpumap.h>
> >>  #include "../../../util/cpumap.h"
> >>  #include "../../../util/pmu.h"
> >> +#include <api/fs/fs.h>
> >>
> >> -const struct pmu_events_table *pmu_events_table__find(void)
> >> +static struct perf_pmu *pmu_core__find_same(void)
> >
> > I'm not sure "find_same" is the best name here. I suspect it should be
> > "find_core_pmu" which would agree with is_arm_pmu_core. Unfortunately
> > "core" has become an overloaded term sometimes used interchangeably
> > with CPU, hyperthread or SMT thread, it was a model name for Intel and
> > it is used to distinguish a set of SMT threads running together from a
> > single one. Anyway, for consistency I think perf_pmu__find_core_pmu is
> > the most appropriate name (or pmu__find_core_pmu, I'm not sure why we
> > get the extra perf_ prefix sometimes, in general that indicates the
> > functionality is in libperf).
> >
>
> The reason for using "pmu_core__find_same" before is to indicate that we're
> only dealing with homogeneous cores. And in the tools/perf/util/pmu.c file,
> most of the static functions have "pmu_" prefix, maybe we can use
> "pmu_find_same_core_pmu"? Ian, John, what do you think?

I wouldn't necessarily worry about hybrid given #slots is currently
ARM specific. For hybrid we'd need to know the CPU for the metric. We
do have the list of CPUs (really hyper/SMT threads) that were
requested for the metric:
https://git.kernel.org/pub/scm/linux/kernel/git/acme/linux.git/tree/tools/perf/util/expr.h?h=perf/core#n9
We use it to compute the #core_wide literal. We need the value early
before events are created, hence the string format. You could search
pmus looking for a PMU that isn't uncore and has matching CPUs, but
AMD has multiple core/hardware PMUs (hence Ravi's most recent work)
and so I suspect we'd need to extend PMU to make this work. We've
solved a similar problem to this with the source_count metric
function, that returns a number of aliased events. We could do
something like:
slots(INST_RETIRED.MACRO_FUSED)
and then from the event get the PMU, from the PMU get the CPUs, from a
CPU get the slots. In that case it may just be cleaner to pass the PMU
name to the slots function, so:
slots(cpu_core) or slots(cpu_atom)
But the parser wouldn't understand that cpu_core or cpu_atom were PMU
names and would try to handle them as events or metric references.

Again, I think ignoring the hybrid case is fine in this case. Using
"same" in the function name to imply "not hybrid" I don't think works,
so I think something like pmu__find_core_pmu is best. You could have a
comment and also a:
assert(!perf_pmu__is_hybrid(pmu->name));
Ultimately I'd like to get rid of all notions of hybrid and just pair
events with a PMU. I recently cleaned this up in builtin-list.c.

Thanks,
Ian

> Thanks,
> Jing
>
> > Aside from that, lgtm. Thanks,
> > Ian
> >
> >>  {
> >>         struct perf_pmu *pmu = NULL;
> >>
> >> @@ -19,8 +20,37 @@ const struct pmu_events_table *pmu_events_table__find(void)
> >>                 if (pmu->cpus->nr != cpu__max_cpu().cpu)
> >>                         return NULL;
> >>
> >> -               return perf_pmu__find_table(pmu);
> >> +               return pmu;
> >>         }
> >>
> >>         return NULL;
> >>  }
> >> +
> >> +const struct pmu_events_table *pmu_events_table__find(void)
> >> +{
> >> +       struct perf_pmu *pmu = pmu_core__find_same();
> >> +
> >> +       if (pmu)
> >> +               return perf_pmu__find_table(pmu);
> >> +
> >> +       return NULL;
> >> +}
> >> +
> >> +double perf_pmu__cpu_slots_per_cycle(void)
> >> +{
> >> +       char path[PATH_MAX];
> >> +       unsigned long long slots = 0;
> >> +       struct perf_pmu *pmu = pmu_core__find_same();
> >> +
> >> +       if (pmu) {
> >> +               scnprintf(path, PATH_MAX,
> >> +                       EVENT_SOURCE_DEVICE_PATH "%s/caps/slots", pmu->name);
> >> +               /*
> >> +                * The value of slots is not greater than 32 bits, but sysfs__read_int
> >> +                * can't read value with 0x prefix, so use sysfs__read_ull instead.
> >> +                */
> >> +               sysfs__read_ull(path, &slots);
> >> +       }
> >> +
> >> +       return (double)slots;
> >> +}
> >> diff --git a/tools/perf/util/expr.c b/tools/perf/util/expr.c
> >> index 00dcde3..9d3076a 100644
> >> --- a/tools/perf/util/expr.c
> >> +++ b/tools/perf/util/expr.c
> >> @@ -19,6 +19,7 @@
> >>  #include <linux/zalloc.h>
> >>  #include <ctype.h>
> >>  #include <math.h>
> >> +#include "pmu.h"
> >>
> >>  #ifdef PARSER_DEBUG
> >>  extern int expr_debug;
> >> @@ -448,6 +449,10 @@ double expr__get_literal(const char *literal, const struct expr_scanner_ctx *ctx
> >>                 result = topology->core_cpus_lists;
> >>                 goto out;
> >>         }
> >> +       if (!strcmp("#slots", literal)) {
> >> +               result = perf_pmu__cpu_slots_per_cycle() ?: NAN;
> >> +               goto out;
> >> +       }
> >>
> >>         pr_err("Unrecognized literal '%s'", literal);
> >>  out:
> >> diff --git a/tools/perf/util/pmu.c b/tools/perf/util/pmu.c
> >> index 2bdeb89..cbb4fbf 100644
> >> --- a/tools/perf/util/pmu.c
> >> +++ b/tools/perf/util/pmu.c
> >> @@ -19,6 +19,7 @@
> >>  #include <regex.h>
> >>  #include <perf/cpumap.h>
> >>  #include <fnmatch.h>
> >> +#include <math.h>
> >>  #include "debug.h"
> >>  #include "evsel.h"
> >>  #include "pmu.h"
> >> @@ -1993,3 +1994,8 @@ int perf_pmu__cpus_match(struct perf_pmu *pmu, struct perf_cpu_map *cpus,
> >>         *ucpus_ptr = unmatched_cpus;
> >>         return 0;
> >>  }
> >> +
> >> +double __weak perf_pmu__cpu_slots_per_cycle(void)
> >> +{
> >> +       return NAN;
> >> +}
> >> diff --git a/tools/perf/util/pmu.h b/tools/perf/util/pmu.h
> >> index 69ca000..fd414ba 100644
> >> --- a/tools/perf/util/pmu.h
> >> +++ b/tools/perf/util/pmu.h
> >> @@ -259,4 +259,5 @@ int perf_pmu__cpus_match(struct perf_pmu *pmu, struct perf_cpu_map *cpus,
> >>
> >>  char *pmu_find_real_name(const char *name);
> >>  char *pmu_find_alias_name(const char *name);
> >> +double perf_pmu__cpu_slots_per_cycle(void);
> >>  #endif /* __PMU_H */
> >> --
> >> 1.8.3.1
> >>