by Arnaldo Carvalho de Melo

[permalink] [raw]

Subject: Re: [PATCH v1 5/8] perf topology: Add core_wide

Em Wed, Aug 31, 2022 at 08:58:49AM -0700, Ian Rogers escreveu:
> On Wed, Aug 31, 2022 at 7:40 AM Arnaldo Carvalho de Melo
> <[email protected]> wrote:
> >
> > Em Tue, Aug 30, 2022 at 09:48:43AM -0700, Ian Rogers escreveu:
> > > It is possible to optimize metrics when all SMT threads (CPUs) on a
> > > core are measuring events in system wide mode. For example, TMA
> > > metrics defines CORE_CLKS for Sandybrdige as:
> > >
> > > if SMT is disabled:
> > > CPU_CLK_UNHALTED.THREAD
> > > if SMT is enabled and recording on all SMT threads:
> > > CPU_CLK_UNHALTED.THREAD_ANY / 2
> > > if SMT is enabled and not recording on all SMT threads:
> > > (CPU_CLK_UNHALTED.THREAD/2)*
> > > (1+CPU_CLK_UNHALTED.ONE_THREAD_ACTIVE/CPU_CLK_UNHALTED.REF_XCLK )
> > >
> > > That is two more events are necessary when not gathering counts on all
> > > SMT threads. To distinguish all SMT threads on a core vs system wide
> > > (all CPUs) call the new property core wide. Add a core wide test that
> > > determines the property from user requested CPUs, the topology and
> > > system wide. System wide is required as other processes running on a
> > > SMT thread will change the counts.
> > >
> > > Signed-off-by: Ian Rogers <[email protected]>
> > > ---
> > > tools/perf/util/cputopo.c | 46 +++++++++++++++++++++++++++++++++++++++
> > > tools/perf/util/cputopo.h | 3 +++
> > > tools/perf/util/smt.c | 14 ++++++++++++
> > > tools/perf/util/smt.h | 7 ++++++
> > > 4 files changed, 70 insertions(+)
> > >
> > > diff --git a/tools/perf/util/cputopo.c b/tools/perf/util/cputopo.c
> > > index 511002e52714..1a3ff6449158 100644
> > > --- a/tools/perf/util/cputopo.c
> > > +++ b/tools/perf/util/cputopo.c
> > > @@ -172,6 +172,52 @@ bool cpu_topology__smt_on(const struct cpu_topology *topology)
> > > return false;
> > > }
> > >
> > > +bool cpu_topology__core_wide(const struct cpu_topology *topology,
> > > + const char *user_requested_cpu_list)
> > > +{
> > > + struct perf_cpu_map *user_requested_cpus;
> > > +
> > > + /*
> > > + * If user_requested_cpu_list is empty then all CPUs are recorded and so
> > > + * core_wide is true.
> > > + */
> > > + if (!user_requested_cpu_list)
> > > + return true;
> > > +
> > > + user_requested_cpus = perf_cpu_map__new(user_requested_cpu_list);
> >
> > Don't we need a NULL test here?

> No, NULL is a valid perf_cpu_map, one with no values in it. For
> example, see here:

But in this specific case, if perf_cpu_map__new() returns NULL, its
because it failed to parse user_requested_cpu_list, so checking against
NULL is valid, no?

- Arnaldo

> https://git.kernel.org/pub/scm/linux/kernel/git/acme/linux.git/tree/tools/lib/perf/cpumap.c?h=perf/core#n266
> So there's no way to tell an error (ENOMEM) from an empty map. My
> suggestion is a rewrite of perf_cpu_map, I'd like to do this in a new
> "libperf2" which has a more permissive than GPL license like libbpf.
> That is out-of-scope here.

> Thanks,
> Ian
>
> > > + /* Check that every user requested CPU is the complete set of SMT threads on a core. */
> > > + for (u32 i = 0; i < topology->core_cpus_lists; i++) {
> > > + const char *core_cpu_list = topology->core_cpus_list[i];
> > > + struct perf_cpu_map *core_cpus = perf_cpu_map__new(core_cpu_list);
> >
> > Here too, no?
> >
> > > + struct perf_cpu cpu;
> > > + int idx;
> > > + bool has_first, first = true;
> > > +
> > > + perf_cpu_map__for_each_cpu(cpu, idx, core_cpus) {
> > > + if (first) {
> > > + has_first = perf_cpu_map__has(user_requested_cpus, cpu);
> > > + first = false;
> > > + } else {
> > > + /*
> > > + * If the first core CPU is user requested then
> > > + * all subsequent CPUs in the core must be user
> > > + * requested too. If the first CPU isn't user
> > > + * requested then none of the others must be
> > > + * too.
> > > + */
> > > + if (perf_cpu_map__has(user_requested_cpus, cpu) != has_first) {
> > > + perf_cpu_map__put(core_cpus);
> > > + perf_cpu_map__put(user_requested_cpus);
> > > + return false;
> > > + }
> > > + }
> > > + }
> > > + perf_cpu_map__put(core_cpus);
> > > + }
> > > + perf_cpu_map__put(user_requested_cpus);
> > > + return true;
> > > +}
> > > +
> > > static bool has_die_topology(void)
> > > {
> > > char filename[MAXPATHLEN];
> > > diff --git a/tools/perf/util/cputopo.h b/tools/perf/util/cputopo.h
> > > index 469db775a13c..969e5920a00e 100644
> > > --- a/tools/perf/util/cputopo.h
> > > +++ b/tools/perf/util/cputopo.h
> > > @@ -60,6 +60,9 @@ struct cpu_topology *cpu_topology__new(void);
> > > void cpu_topology__delete(struct cpu_topology *tp);
> > > /* Determine from the core list whether SMT was enabled. */
> > > bool cpu_topology__smt_on(const struct cpu_topology *topology);
> > > +/* Are the sets of SMT siblings all enabled or all disabled in user_requested_cpus. */
> > > +bool cpu_topology__core_wide(const struct cpu_topology *topology,
> > > + const char *user_requested_cpu_list);
> > >
> > > struct numa_topology *numa_topology__new(void);
> > > void numa_topology__delete(struct numa_topology *tp);
> > > diff --git a/tools/perf/util/smt.c b/tools/perf/util/smt.c
> > > index ce90c4ee4138..994e9e418227 100644
> > > --- a/tools/perf/util/smt.c
> > > +++ b/tools/perf/util/smt.c
> > > @@ -21,3 +21,17 @@ bool smt_on(const struct cpu_topology *topology)
> > > cached = true;
> > > return cached_result;
> > > }
> > > +
> > > +bool core_wide(bool system_wide, const char *user_requested_cpu_list,
> > > + const struct cpu_topology *topology)
> > > +{
> > > + /* If not everything running on a core is being recorded then we can't use core_wide. */
> > > + if (!system_wide)
> > > + return false;
> > > +
> > > + /* Cheap case that SMT is disabled and therefore we're inherently core_wide. */
> > > + if (!smt_on(topology))
> > > + return true;
> > > +
> > > + return cpu_topology__core_wide(topology, user_requested_cpu_list);
> > > +}
> > > diff --git a/tools/perf/util/smt.h b/tools/perf/util/smt.h
> > > index e26999c6b8d4..ae9095f2c38c 100644
> > > --- a/tools/perf/util/smt.h
> > > +++ b/tools/perf/util/smt.h
> > > @@ -7,4 +7,11 @@ struct cpu_topology;
> > > /* Returns true if SMT (aka hyperthreading) is enabled. */
> > > bool smt_on(const struct cpu_topology *topology);
> > >
> > > +/*
> > > + * Returns true when system wide and all SMT threads for a core are in the
> > > + * user_requested_cpus map.
> > > + */
> > > +bool core_wide(bool system_wide, const char *user_requested_cpu_list,
> > > + const struct cpu_topology *topology);
> > > +
> > > #endif /* __SMT_H */
> > > --
> > > 2.37.2.672.g94769d06f0-goog
> >
> > --
> >
> > - Arnaldo

--

- Arnaldo

2022-08-31 17:21:24

by Ian Rogers

[permalink] [raw]

Subject: Re: [PATCH v1 5/8] perf topology: Add core_wide

On Wed, Aug 31, 2022 at 9:20 AM Arnaldo Carvalho de Melo
<[email protected]> wrote:
>
> Em Wed, Aug 31, 2022 at 08:58:49AM -0700, Ian Rogers escreveu:
> > On Wed, Aug 31, 2022 at 7:40 AM Arnaldo Carvalho de Melo
> > <[email protected]> wrote:
> > >
> > > Em Tue, Aug 30, 2022 at 09:48:43AM -0700, Ian Rogers escreveu:
> > > > It is possible to optimize metrics when all SMT threads (CPUs) on a
> > > > core are measuring events in system wide mode. For example, TMA
> > > > metrics defines CORE_CLKS for Sandybrdige as:
> > > >
> > > > if SMT is disabled:
> > > > CPU_CLK_UNHALTED.THREAD
> > > > if SMT is enabled and recording on all SMT threads:
> > > > CPU_CLK_UNHALTED.THREAD_ANY / 2
> > > > if SMT is enabled and not recording on all SMT threads:
> > > > (CPU_CLK_UNHALTED.THREAD/2)*
> > > > (1+CPU_CLK_UNHALTED.ONE_THREAD_ACTIVE/CPU_CLK_UNHALTED.REF_XCLK )
> > > >
> > > > That is two more events are necessary when not gathering counts on all
> > > > SMT threads. To distinguish all SMT threads on a core vs system wide
> > > > (all CPUs) call the new property core wide. Add a core wide test that
> > > > determines the property from user requested CPUs, the topology and
> > > > system wide. System wide is required as other processes running on a
> > > > SMT thread will change the counts.
> > > >
> > > > Signed-off-by: Ian Rogers <[email protected]>
> > > > ---
> > > > tools/perf/util/cputopo.c | 46 +++++++++++++++++++++++++++++++++++++++
> > > > tools/perf/util/cputopo.h | 3 +++
> > > > tools/perf/util/smt.c | 14 ++++++++++++
> > > > tools/perf/util/smt.h | 7 ++++++
> > > > 4 files changed, 70 insertions(+)
> > > >
> > > > diff --git a/tools/perf/util/cputopo.c b/tools/perf/util/cputopo.c
> > > > index 511002e52714..1a3ff6449158 100644
> > > > --- a/tools/perf/util/cputopo.c
> > > > +++ b/tools/perf/util/cputopo.c
> > > > @@ -172,6 +172,52 @@ bool cpu_topology__smt_on(const struct cpu_topology *topology)
> > > > return false;
> > > > }
> > > >
> > > > +bool cpu_topology__core_wide(const struct cpu_topology *topology,
> > > > + const char *user_requested_cpu_list)
> > > > +{
> > > > + struct perf_cpu_map *user_requested_cpus;
> > > > +
> > > > + /*
> > > > + * If user_requested_cpu_list is empty then all CPUs are recorded and so
> > > > + * core_wide is true.
> > > > + */
> > > > + if (!user_requested_cpu_list)
> > > > + return true;
> > > > +
> > > > + user_requested_cpus = perf_cpu_map__new(user_requested_cpu_list);
> > >
> > > Don't we need a NULL test here?
>
> > No, NULL is a valid perf_cpu_map, one with no values in it. For
> > example, see here:
>
> But in this specific case, if perf_cpu_map__new() returns NULL, its
> because it failed to parse user_requested_cpu_list, so checking against
> NULL is valid, no?

What about the empty string?

Thanks,
Ian

> - Arnaldo
>
> > https://git.kernel.org/pub/scm/linux/kernel/git/acme/linux.git/tree/tools/lib/perf/cpumap.c?h=perf/core#n266
> > So there's no way to tell an error (ENOMEM) from an empty map. My
> > suggestion is a rewrite of perf_cpu_map, I'd like to do this in a new
> > "libperf2" which has a more permissive than GPL license like libbpf.
> > That is out-of-scope here.
>
> > Thanks,
> > Ian
> >
> > > > + /* Check that every user requested CPU is the complete set of SMT threads on a core. */
> > > > + for (u32 i = 0; i < topology->core_cpus_lists; i++) {
> > > > + const char *core_cpu_list = topology->core_cpus_list[i];
> > > > + struct perf_cpu_map *core_cpus = perf_cpu_map__new(core_cpu_list);
> > >
> > > Here too, no?
> > >
> > > > + struct perf_cpu cpu;
> > > > + int idx;
> > > > + bool has_first, first = true;
> > > > +
> > > > + perf_cpu_map__for_each_cpu(cpu, idx, core_cpus) {
> > > > + if (first) {
> > > > + has_first = perf_cpu_map__has(user_requested_cpus, cpu);
> > > > + first = false;
> > > > + } else {
> > > > + /*
> > > > + * If the first core CPU is user requested then
> > > > + * all subsequent CPUs in the core must be user
> > > > + * requested too. If the first CPU isn't user
> > > > + * requested then none of the others must be
> > > > + * too.
> > > > + */
> > > > + if (perf_cpu_map__has(user_requested_cpus, cpu) != has_first) {
> > > > + perf_cpu_map__put(core_cpus);
> > > > + perf_cpu_map__put(user_requested_cpus);
> > > > + return false;
> > > > + }
> > > > + }
> > > > + }
> > > > + perf_cpu_map__put(core_cpus);
> > > > + }
> > > > + perf_cpu_map__put(user_requested_cpus);
> > > > + return true;
> > > > +}
> > > > +
> > > > static bool has_die_topology(void)
> > > > {
> > > > char filename[MAXPATHLEN];
> > > > diff --git a/tools/perf/util/cputopo.h b/tools/perf/util/cputopo.h
> > > > index 469db775a13c..969e5920a00e 100644
> > > > --- a/tools/perf/util/cputopo.h
> > > > +++ b/tools/perf/util/cputopo.h
> > > > @@ -60,6 +60,9 @@ struct cpu_topology *cpu_topology__new(void);
> > > > void cpu_topology__delete(struct cpu_topology *tp);
> > > > /* Determine from the core list whether SMT was enabled. */
> > > > bool cpu_topology__smt_on(const struct cpu_topology *topology);
> > > > +/* Are the sets of SMT siblings all enabled or all disabled in user_requested_cpus. */
> > > > +bool cpu_topology__core_wide(const struct cpu_topology *topology,
> > > > + const char *user_requested_cpu_list);
> > > >
> > > > struct numa_topology *numa_topology__new(void);
> > > > void numa_topology__delete(struct numa_topology *tp);
> > > > diff --git a/tools/perf/util/smt.c b/tools/perf/util/smt.c
> > > > index ce90c4ee4138..994e9e418227 100644
> > > > --- a/tools/perf/util/smt.c
> > > > +++ b/tools/perf/util/smt.c
> > > > @@ -21,3 +21,17 @@ bool smt_on(const struct cpu_topology *topology)
> > > > cached = true;
> > > > return cached_result;
> > > > }
> > > > +
> > > > +bool core_wide(bool system_wide, const char *user_requested_cpu_list,
> > > > + const struct cpu_topology *topology)
> > > > +{
> > > > + /* If not everything running on a core is being recorded then we can't use core_wide. */
> > > > + if (!system_wide)
> > > > + return false;
> > > > +
> > > > + /* Cheap case that SMT is disabled and therefore we're inherently core_wide. */
> > > > + if (!smt_on(topology))
> > > > + return true;
> > > > +
> > > > + return cpu_topology__core_wide(topology, user_requested_cpu_list);
> > > > +}
> > > > diff --git a/tools/perf/util/smt.h b/tools/perf/util/smt.h
> > > > index e26999c6b8d4..ae9095f2c38c 100644
> > > > --- a/tools/perf/util/smt.h
> > > > +++ b/tools/perf/util/smt.h
> > > > @@ -7,4 +7,11 @@ struct cpu_topology;
> > > > /* Returns true if SMT (aka hyperthreading) is enabled. */
> > > > bool smt_on(const struct cpu_topology *topology);
> > > >
> > > > +/*
> > > > + * Returns true when system wide and all SMT threads for a core are in the
> > > > + * user_requested_cpus map.
> > > > + */
> > > > +bool core_wide(bool system_wide, const char *user_requested_cpu_list,
> > > > + const struct cpu_topology *topology);
> > > > +
> > > > #endif /* __SMT_H */
> > > > --
> > > > 2.37.2.672.g94769d06f0-goog
> > >
> > > --
> > >
> > > - Arnaldo
>
> --
>
> - Arnaldo