Re: [PATCH v1 05/23] perf pmu: Remove perf_pmu__hybrid_mounted

[Liang, Kan]

On 2023-05-22 10:06 a.m., Ian Rogers wrote:
> On Mon, May 22, 2023 at 4:55 AM Liang, Kan <kan.liang@xxxxxxxxxxxxxxx> wrote:
>>
>>
>>
>> On 2023-05-22 1:21 a.m., Ian Rogers wrote:
>>> On Sun, May 21, 2023 at 12:23 PM Liang, Kan <kan.liang@xxxxxxxxxxxxxxx> wrote:
>>>>
>>>>
>>>>
>>>> On 2023-05-17 10:57 a.m., Ian Rogers wrote:
>>>>> perf_pmu__hybrid_mounted is used to detect whether cpu_core or
>>>>> cpu_atom
>>>>
>>>> Currently, there are only two CPU types for a hybrid machine, core and
>>>> atom. But there may be more CPU types added later. Please see the CPUID
>>>> 1AH EAX enumeration in SDM VOL2. It has several reserved encodings for
>>>> CPU types. It's better not using the hardcode cpu_core/cpu_atom to
>>>> replace the perf_pmu__hybrid_mounted().
>>>>
>>>> Thanks,
>>>> Kan
>>>
>>> This is covered by later patches. Specifically patches like:
>>> patch 7: perf pmu: Add is_core to pmu
>>> patch 20: Split pmus list into core and uncore
>>>
>>> Ultimately in pmus.h we have two scan routines, one for all PMUs and
>>> one for core PMUs. For everything except hybrid (and the name varies a
>>> bit on ARM) the core scan only scans "cpu", on hybrid it scans
>>> "cpu_atom" and "cpu_core". The determination of core vs uncore is
>>> done without using the name, so can support >2 hybrid PMUs. At this
>>> point in the patch series I'm trying to simplify everything so that I
>>> can then build the pmus interface.
>>
>> But if we add a new core type "cpu_whatever" later, we have to hardcode
>> the new name to the perf tool, right? Users have to update the perf tool
>> for the new platforms, otherwise I think the new type will be treated as
>> an uncore PMU.
>>
>> Since the hybrid is Intel only, I think it may be better move the
>> is_pmu_hybrid() to X86 specifc code. For the Intel only code, we already
>> have a naming rule for the hybrid name, "cpu_$". So we don't need to
>> update the tool for every new CPU type.
>>
>> Thanks,
>> Kan
> 
> I don't disagree, but fixing all uses of is_pmu_hybrid and similarly
> perf_pmus__has_hybrid is going to add yet more to a moderately long
> patch series. I think in most cases is_pmu_hybrid can be replaced by a
> core oriented alternative. For example, in pmu.c there is
> perf_pmu__auto_merge_stats that normally returns true that we want to
> merge counts for uncore or non-hybrid PMUs. For hybrid it returns
> false so that cpu_atom and cpu_core counts aren't merged. A core
> oriented alternative would be to return false if the PMU is core and
> the number of core PMUs is >1 - this also avoids any hard coding of
> PMU names and assuming >1 core PMU means they all begin with "cpu_".

I'm OK with the alternative method.

> 
> The scope of fixing the remaining is_pmu_hybrid and perf_pmus__has_hybrid is:
I don't think we need to modify each places listed below.
We just need to update is_pmu_hybrid() and perf_pmus__has_hybrid(). And
probably use is_pmu_hybrid() to replace perf_pmu__skip_empty_cpus().
I don't think we need to change the interface.

It should not be a big patch.

Is there anything I missed?

Thanks,
Kan
> ```
> $ grep -rn perf_pmus__has_hybrid tools/perf
> tools/perf/util/header.c:1592:  if (perf_pmus__has_hybrid()) {
> tools/perf/util/mem-events.c:132:               if (!perf_pmus__has_hybrid()) {
> tools/perf/util/mem-events.c:199:               if (!perf_pmus__has_hybrid()) {
> tools/perf/util/evsel.c:3139:   if (!perf_pmus__has_hybrid())
> tools/perf/util/pmus.h:21:bool perf_pmus__has_hybrid(void);
> tools/perf/util/stat-display.c:684:     if (!perf_pmus__has_hybrid())
> tools/perf/util/metricgroup.c:277:      bool all_pmus = !strcmp(pmu,
> "all") || !perf_pmus__has_hybrid() || !is_pmu_hybrid(pmu);
> tools/perf/util/pmus.c:474:bool perf_pmus__has_hybrid(void)
> tools/perf/util/cputopo.c:477:  if (!perf_pmus__has_hybrid())
> tools/perf/tests/attr.c:188:    if (perf_pmus__has_hybrid())
> tools/perf/tests/topology.c:44: if (!perf_pmus__has_hybrid()) {
> tools/perf/tests/parse-metric.c:306:    if (!perf_pmus__has_hybrid()) {
> tools/perf/tests/switch-tracking.c:378: if (perf_pmus__has_hybrid()) {
> tools/perf/builtin-record.c:1297:           perf_pmus__has_hybrid()) {
> tools/perf/builtin-record.c:2196:       if (!perf_pmus__has_hybrid())
> tools/perf/builtin-record.c:4196:       rec->opts.target.hybrid =
> perf_pmus__has_hybrid();
> tools/perf/builtin-stat.c:2463: target.hybrid = perf_pmus__has_hybrid();
> tools/perf/arch/x86/util/perf_regs.c:295:       if (perf_pmus__has_hybrid()) {
> tools/perf/arch/x86/util/evlist.c:21:   if (!perf_pmus__has_hybrid())
> tools/perf/arch/x86/tests/hybrid.c:284: if (!perf_pmus__has_hybrid())
> 
> $ grep -rn is_pmu_hybrid tools/perf
> tools/perf/util/pmu.c:1433:bool is_pmu_hybrid(const char *name)
> tools/perf/util/pmu.c:1445:     return !is_pmu_hybrid(pmu->name);
> tools/perf/util/pmu.h:224:bool is_pmu_hybrid(const char *name);
> tools/perf/util/metricgroup.c:277:      bool all_pmus = !strcmp(pmu,
> "all") || !perf_pmus__
> has_hybrid() || !is_pmu_hybrid(pmu);
> tools/perf/util/pmus.c:482:                     if (is_pmu_hybrid(pmu->name)) {
> ```
> 
> So, I think it makes sense to do it as a follow up.
> 
> Thanks,
> Ian
> 
>>>
>>> Thanks,
>>> Ian
>>>
>>>>> is mounted with a non-empty cpus file by
>>>>> pmu_lookup. pmu_lookup will attempt to read the cpus file too and so
>>>>> the check can be folded into this.
>>>>>
>>>>> Checking hybrid_mounted in pmu_is_uncore is redundant as the next
>>>>> cpumask read will fail returning false.
>>>>>
>>>>> Reduce the scope of perf_pmu__find_hybrid_pmu by making it static.
>>>>>
>>>>> Signed-off-by: Ian Rogers <irogers@xxxxxxxxxx>
>>>>> ---
>>>>>  tools/perf/util/pmu-hybrid.c | 15 +--------------
>>>>>  tools/perf/util/pmu-hybrid.h |  3 ---
>>>>>  tools/perf/util/pmu.c        | 26 ++++++++++++++------------
>>>>>  3 files changed, 15 insertions(+), 29 deletions(-)
>>>>>
>>>>> diff --git a/tools/perf/util/pmu-hybrid.c b/tools/perf/util/pmu-hybrid.c
>>>>> index bc4cb0738c35..7fe943dd3217 100644
>>>>> --- a/tools/perf/util/pmu-hybrid.c
>>>>> +++ b/tools/perf/util/pmu-hybrid.c
>>>>> @@ -18,20 +18,7 @@
>>>>>
>>>>>  LIST_HEAD(perf_pmu__hybrid_pmus);
>>>>>
>>>>> -bool perf_pmu__hybrid_mounted(const char *name)
>>>>> -{
>>>>> -     int cpu;
>>>>> -     char pmu_name[PATH_MAX];
>>>>> -     struct perf_pmu pmu = {.name = pmu_name};
>>>>> -
>>>>> -     if (strncmp(name, "cpu_", 4))
>>>>> -             return false;
>>>>> -
>>>>> -     strlcpy(pmu_name, name, sizeof(pmu_name));
>>>>> -     return perf_pmu__scan_file(&pmu, "cpus", "%u", &cpu) > 0;
>>>>> -}
>>>>> -
>>>>> -struct perf_pmu *perf_pmu__find_hybrid_pmu(const char *name)
>>>>> +static struct perf_pmu *perf_pmu__find_hybrid_pmu(const char *name)
>>>>>  {
>>>>>       struct perf_pmu *pmu;
>>>>>
>>>>> diff --git a/tools/perf/util/pmu-hybrid.h b/tools/perf/util/pmu-hybrid.h
>>>>> index 206b94931531..8dbcae935020 100644
>>>>> --- a/tools/perf/util/pmu-hybrid.h
>>>>> +++ b/tools/perf/util/pmu-hybrid.h
>>>>> @@ -13,9 +13,6 @@ extern struct list_head perf_pmu__hybrid_pmus;
>>>>>  #define perf_pmu__for_each_hybrid_pmu(pmu)   \
>>>>>       list_for_each_entry(pmu, &perf_pmu__hybrid_pmus, hybrid_list)
>>>>>
>>>>> -bool perf_pmu__hybrid_mounted(const char *name);
>>>>> -
>>>>> -struct perf_pmu *perf_pmu__find_hybrid_pmu(const char *name);
>>>>>  bool perf_pmu__is_hybrid(const char *name);
>>>>>
>>>>>  static inline int perf_pmu__hybrid_pmu_num(void)
>>>>> diff --git a/tools/perf/util/pmu.c b/tools/perf/util/pmu.c
>>>>> index 1e0be23d4dd7..729b1f166f80 100644
>>>>> --- a/tools/perf/util/pmu.c
>>>>> +++ b/tools/perf/util/pmu.c
>>>>> @@ -617,9 +617,6 @@ static bool pmu_is_uncore(int dirfd, const char *name)
>>>>>  {
>>>>>       int fd;
>>>>>
>>>>> -     if (perf_pmu__hybrid_mounted(name))
>>>>> -             return false;
>>>>> -
>>>>>       fd = perf_pmu__pathname_fd(dirfd, name, "cpumask", O_PATH);
>>>>>       if (fd < 0)
>>>>>               return false;
>>>>> @@ -898,6 +895,16 @@ static int pmu_max_precise(int dirfd, struct perf_pmu *pmu)
>>>>>       return max_precise;
>>>>>  }
>>>>>
>>>>> +/**
>>>>> + * perf_pmu__skip_empty_cpus() - should pmu_lookup skip the named PMU if the
>>>>> + *      cpus or cpumask file isn't present?
>>>>> + * @name: Name of PMU.
>>>>> + */
>>>>> +static bool perf_pmu__skip_empty_cpus(const char *name)
>>>>> +{
>>>>> +     return !strcmp(name, "cpu_core") || !strcmp(name, "cpu_atom");
>>>>> +}
>>>>> +
>>>>>  static struct perf_pmu *pmu_lookup(int dirfd, const char *lookup_name)
>>>>>  {
>>>>>       struct perf_pmu *pmu;
>>>>> @@ -905,15 +912,8 @@ static struct perf_pmu *pmu_lookup(int dirfd, const char *lookup_name)
>>>>>       LIST_HEAD(aliases);
>>>>>       __u32 type;
>>>>>       char *name = pmu_find_real_name(lookup_name);
>>>>> -     bool is_hybrid = perf_pmu__hybrid_mounted(name);
>>>>>       char *alias_name;
>>>>>
>>>>> -     /*
>>>>> -      * Check pmu name for hybrid and the pmu may be invalid in sysfs
>>>>> -      */
>>>>> -     if (!strncmp(name, "cpu_", 4) && !is_hybrid)
>>>>> -             return NULL;
>>>>> -
>>>>>       /*
>>>>>        * The pmu data we store & need consists of the pmu
>>>>>        * type value and format definitions. Load both right
>>>>> @@ -933,8 +933,10 @@ static struct perf_pmu *pmu_lookup(int dirfd, const char *lookup_name)
>>>>>               return NULL;
>>>>>
>>>>>       pmu->cpus = pmu_cpumask(dirfd, name);
>>>>> -     pmu->name = strdup(name);
>>>>> +     if (!pmu->cpus && perf_pmu__skip_empty_cpus(name))
>>>>> +             goto err;
>>>>>
>>>>> +     pmu->name = strdup(name);
>>>>>       if (!pmu->name)
>>>>>               goto err;
>>>>>
>>>>> @@ -964,7 +966,7 @@ static struct perf_pmu *pmu_lookup(int dirfd, const char *lookup_name)
>>>>>       list_splice(&aliases, &pmu->aliases);
>>>>>       list_add_tail(&pmu->list, &pmus);
>>>>>
>>>>> -     if (is_hybrid)
>>>>> +     if (!strcmp(name, "cpu_core") || !strcmp(name, "cpu_atom"))
>>>>>               list_add_tail(&pmu->hybrid_list, &perf_pmu__hybrid_pmus);
>>>>>       else
>>>>>               INIT_LIST_HEAD(&pmu->hybrid_list);