[PATCH v2 00/10] Better fixes for grouping of events

[PATCH v2 00/10] Better fixes for grouping of events

[Ian Rogers]

The rules for grouping events has grown more complex. Topdown events
must be grouped, but flags like --metric-no-group and flags on metrics
don't respect this. Uncore events may be expanded using wild cards for
PMU names, but then the events need reordering so the group members
are adjacent. Rather than fixing metrics, this change fixes the main
event parsing code to first sort and then regroup evsels.

As this is shared functionality changes to it should cause
concern. The change is done with the intent of simplifying and making
more robust the grouping logic, examples are given. If additional
changes are necessary, they are most likely necessary to the
evsel__pmu_name logic as the code avoids breaking groups that are on
the same PMU. The pmu_name is tweaked in the case of software and aux
events, that use groups in a slightly different manner to conventional
events.

The code was manually tested as well as passing perf test on a Intel
tigerlake CPU with intel-pt.

v2. Fix up the commit message on 4/10 (thanks Arnaldo). Drop
    unnecessary v1 5/10 (thanks Kan). evlist->core.nr_groups wasn't
    being correctly maintained after the sort/regrouping and so the
    new patch 10/10 removes that variable and computes it from the
    evlist when necessary, generally just tests.

Ian Rogers (10):
  libperf evlist: Avoid a use of evsel idx
  perf stat: Don't remove all grouped events when CPU maps disagree
  perf record: Early auxtrace initialization before event parsing
  perf stat: Modify the group test
  perf evsel: Allow const evsel for certain accesses
  perf evsel: Add function to compute pmu_name
  perf parse-events: Pass ownership of the group name
  perf parse-events: Sort and group parsed events
  perf evsel: Remove use_uncore_alias
  perf evlist: Remove nr_groups

 tools/lib/perf/evlist.c                  |  31 ++-
 tools/lib/perf/include/internal/evlist.h |   1 -
 tools/lib/perf/include/perf/evlist.h     |   1 +
 tools/perf/arch/x86/util/auxtrace.c      |  17 +-
 tools/perf/arch/x86/util/evlist.c        |  39 ++--
 tools/perf/builtin-record.c              |   8 +-
 tools/perf/builtin-report.c              |   2 +-
 tools/perf/builtin-stat.c                |  24 ++-
 tools/perf/tests/bpf.c                   |   1 -
 tools/perf/tests/parse-events.c          |  22 +-
 tools/perf/tests/pfm.c                   |  12 +-
 tools/perf/util/auxtrace.h               |   2 +
 tools/perf/util/evlist.c                 |   2 +-
 tools/perf/util/evlist.h                 |   8 +-
 tools/perf/util/evsel.c                  |  27 ++-
 tools/perf/util/evsel.h                  |   8 +-
 tools/perf/util/header.c                 |   3 +-
 tools/perf/util/parse-events.c           | 254 +++++++++++------------
 tools/perf/util/parse-events.h           |   7 +-
 tools/perf/util/parse-events.y           |  27 +--
 tools/perf/util/pfm.c                    |   1 -
 tools/perf/util/pmu.c                    |   6 +-
 tools/perf/util/python.c                 |   2 +-
 tools/perf/util/stat-shadow.c            |   2 +-
 24 files changed, 270 insertions(+), 237 deletions(-)

-- 
2.40.0.rc0.216.gc4246ad0f0-goog

[PATCH v2 01/10] libperf evlist: Avoid a use of evsel idx

[Ian Rogers]

Setting the leader iterates the list, so rather than use idx (which
may be changed through list reordering) just count the elements and
set afterwards.

Signed-off-by: Ian Rogers <irogers@google.com>
---
 tools/lib/perf/evlist.c | 13 ++++++-------
 1 file changed, 6 insertions(+), 7 deletions(-)

diff --git a/tools/lib/perf/evlist.c b/tools/lib/perf/evlist.c
index 61b637f29b82..2d6121e89ccb 100644
--- a/tools/lib/perf/evlist.c
+++ b/tools/lib/perf/evlist.c
@@ -687,15 +687,14 @@ perf_evlist__next_mmap(struct perf_evlist *evlist, struct perf_mmap *map,
 
 void __perf_evlist__set_leader(struct list_head *list, struct perf_evsel *leader)
 {
-	struct perf_evsel *first, *last, *evsel;
-
-	first = list_first_entry(list, struct perf_evsel, node);
-	last = list_last_entry(list, struct perf_evsel, node);
-
-	leader->nr_members = last->idx - first->idx + 1;
+	struct perf_evsel *evsel;
+	int n = 0;
 
-	__perf_evlist__for_each_entry(list, evsel)
+	__perf_evlist__for_each_entry(list, evsel) {
 		evsel->leader = leader;
+		n++;
+	}
+	leader->nr_members = n;
 }
 
 void perf_evlist__set_leader(struct perf_evlist *evlist)
-- 
2.40.0.rc0.216.gc4246ad0f0-goog

[PATCH v2 02/10] perf stat: Don't remove all grouped events when CPU maps disagree

[Ian Rogers]

If the events in an evlist's CPU map differ then the entire group is
removed. For example:

```
$ perf stat -e '{imc_free_running/data_read/,imc_free_running/data_write/,cs}' -a sleep 1
WARNING: grouped events cpus do not match, disabling group:
  anon group { imc_free_running/data_read/, imc_free_running/data_write/, cs }
```

Change the behavior so that just the events not matching the leader
are removed. So in the example above, just 'cs' will be removed.

Modify the warning so that it is produced once for each group, rather
than once for the entire evlist. Shrink the scope and size of the
warning text buffer.

Signed-off-by: Ian Rogers <irogers@google.com>
---
 tools/perf/builtin-stat.c | 24 +++++++++++++++---------
 1 file changed, 15 insertions(+), 9 deletions(-)

diff --git a/tools/perf/builtin-stat.c b/tools/perf/builtin-stat.c
index d70b1ec88594..5c12ae5efce5 100644
--- a/tools/perf/builtin-stat.c
+++ b/tools/perf/builtin-stat.c
@@ -181,14 +181,13 @@ static bool cpus_map_matched(struct evsel *a, struct evsel *b)
 
 static void evlist__check_cpu_maps(struct evlist *evlist)
 {
-	struct evsel *evsel, *pos, *leader;
-	char buf[1024];
+	struct evsel *evsel, *warned_leader = NULL;
 
 	if (evlist__has_hybrid(evlist))
 		evlist__warn_hybrid_group(evlist);
 
 	evlist__for_each_entry(evlist, evsel) {
-		leader = evsel__leader(evsel);
+		struct evsel *leader = evsel__leader(evsel);
 
 		/* Check that leader matches cpus with each member. */
 		if (leader == evsel)
@@ -197,19 +196,26 @@ static void evlist__check_cpu_maps(struct evlist *evlist)
 			continue;
 
 		/* If there's mismatch disable the group and warn user. */
-		WARN_ONCE(1, "WARNING: grouped events cpus do not match, disabling group:\n");
-		evsel__group_desc(leader, buf, sizeof(buf));
-		pr_warning("  %s\n", buf);
-
+		if (warned_leader != leader) {
+			char buf[200];
+
+			pr_warning("WARNING: grouped events cpus do not match.\n"
+				"Events with CPUs not matching the leader will "
+				"be removed from the group.\n");
+			evsel__group_desc(leader, buf, sizeof(buf));
+			pr_warning("  %s\n", buf);
+			warned_leader = leader;
+		}
 		if (verbose > 0) {
+			char buf[200];
+
 			cpu_map__snprint(leader->core.cpus, buf, sizeof(buf));
 			pr_warning("     %s: %s\n", leader->name, buf);
 			cpu_map__snprint(evsel->core.cpus, buf, sizeof(buf));
 			pr_warning("     %s: %s\n", evsel->name, buf);
 		}
 
-		for_each_group_evsel(pos, leader)
-			evsel__remove_from_group(pos, leader);
+		evsel__remove_from_group(evsel, leader);
 	}
 }
 
-- 
2.40.0.rc0.216.gc4246ad0f0-goog

Re: [PATCH v2 02/10] perf stat: Don't remove all grouped events when CPU maps disagree

[Liang, Kan]

On 2023-03-02 4:25 p.m., Ian Rogers wrote:
> If the events in an evlist's CPU map differ then the entire group is
> removed. For example:
> 
> ```
> $ perf stat -e '{imc_free_running/data_read/,imc_free_running/data_write/,cs}' -a sleep 1
> WARNING: grouped events cpus do not match, disabling group:
>   anon group { imc_free_running/data_read/, imc_free_running/data_write/, cs }
> ```
> 
> Change the behavior so that just the events not matching the leader
> are removed. So in the example above, just 'cs' will be removed.
> 
> Modify the warning so that it is produced once for each group, rather
> than once for the entire evlist. Shrink the scope and size of the
> warning text buffer.

For the uncore, we usually have to create a group for each uncore PMU.
The number of groups may be big. For example, on ICX, we have 40 CHA
PMUs. For SPR, there should be more CHAs. If we have something like
{cycles,uncore_cha/event=0x1/}, is the warning shown 40 times on ICX?
If so, it should be very annoying.

Maybe it's better to keep the current behavior which only print a
warning once and notify the users that perf will re-group the events.
For the details, they can get it from the -v option.

Thanks,
Kan
> 
> Signed-off-by: Ian Rogers <irogers@google.com>
> ---
>  tools/perf/builtin-stat.c | 24 +++++++++++++++---------
>  1 file changed, 15 insertions(+), 9 deletions(-)
> 
> diff --git a/tools/perf/builtin-stat.c b/tools/perf/builtin-stat.c
> index d70b1ec88594..5c12ae5efce5 100644
> --- a/tools/perf/builtin-stat.c
> +++ b/tools/perf/builtin-stat.c
> @@ -181,14 +181,13 @@ static bool cpus_map_matched(struct evsel *a, struct evsel *b)
>  
>  static void evlist__check_cpu_maps(struct evlist *evlist)
>  {
> -	struct evsel *evsel, *pos, *leader;
> -	char buf[1024];
> +	struct evsel *evsel, *warned_leader = NULL;
>  
>  	if (evlist__has_hybrid(evlist))
>  		evlist__warn_hybrid_group(evlist);
>  
>  	evlist__for_each_entry(evlist, evsel) {
> -		leader = evsel__leader(evsel);
> +		struct evsel *leader = evsel__leader(evsel);
>  
>  		/* Check that leader matches cpus with each member. */
>  		if (leader == evsel)
> @@ -197,19 +196,26 @@ static void evlist__check_cpu_maps(struct evlist *evlist)
>  			continue;
>  
>  		/* If there's mismatch disable the group and warn user. */
> -		WARN_ONCE(1, "WARNING: grouped events cpus do not match, disabling group:\n");
> -		evsel__group_desc(leader, buf, sizeof(buf));
> -		pr_warning("  %s\n", buf);
> -
> +		if (warned_leader != leader) {
> +			char buf[200];
> +
> +			pr_warning("WARNING: grouped events cpus do not match.\n"
> +				"Events with CPUs not matching the leader will "
> +				"be removed from the group.\n");
> +			evsel__group_desc(leader, buf, sizeof(buf));
> +			pr_warning("  %s\n", buf);
> +			warned_leader = leader;
> +		}
>  		if (verbose > 0) {
> +			char buf[200];
> +
>  			cpu_map__snprint(leader->core.cpus, buf, sizeof(buf));
>  			pr_warning("     %s: %s\n", leader->name, buf);
>  			cpu_map__snprint(evsel->core.cpus, buf, sizeof(buf));
>  			pr_warning("     %s: %s\n", evsel->name, buf);
>  		}
>  
> -		for_each_group_evsel(pos, leader)
> -			evsel__remove_from_group(pos, leader);
> +		evsel__remove_from_group(evsel, leader);
>  	}
>  }
>  

Re: [PATCH v2 02/10] perf stat: Don't remove all grouped events when CPU maps disagree

[Ian Rogers]

On Fri, Mar 3, 2023 at 7:50 AM Liang, Kan <kan.liang@linux.intel.com> wrote:
>
>
>
> On 2023-03-02 4:25 p.m., Ian Rogers wrote:
> > If the events in an evlist's CPU map differ then the entire group is
> > removed. For example:
> >
> > ```
> > $ perf stat -e '{imc_free_running/data_read/,imc_free_running/data_write/,cs}' -a sleep 1
> > WARNING: grouped events cpus do not match, disabling group:
> >   anon group { imc_free_running/data_read/, imc_free_running/data_write/, cs }
> > ```
> >
> > Change the behavior so that just the events not matching the leader
> > are removed. So in the example above, just 'cs' will be removed.
> >
> > Modify the warning so that it is produced once for each group, rather
> > than once for the entire evlist. Shrink the scope and size of the
> > warning text buffer.
>
> For the uncore, we usually have to create a group for each uncore PMU.
> The number of groups may be big. For example, on ICX, we have 40 CHA
> PMUs. For SPR, there should be more CHAs. If we have something like
> {cycles,uncore_cha/event=0x1/}, is the warning shown 40 times on ICX?
> If so, it should be very annoying.
>
> Maybe it's better to keep the current behavior which only print a
> warning once and notify the users that perf will re-group the events.
> For the details, they can get it from the -v option.

Thanks Kan, I could imagine that but I was also worried about cases
where there are multiple groups like:

```
$ perf stat -e '{imc_free_running/data_read/,cs},{uncore_clock/clockticks/,cs}'
-a sleep 1
WARNING: grouped events cpus do not match.
Events with CPUs not matching the leader will be removed from the group.
 anon group { imc_free_running/data_read/, cs }
WARNING: grouped events cpus do not match.
Events with CPUs not matching the leader will be removed from the group.
 anon group { uncore_clock/clockticks/, cs }

Performance counter stats for 'system wide':

         1,255.75 MiB  imc_free_running/data_read/
            7,571      cs
    1,327,285,527      uncore_clock/clockticks/
            7,571      cs

      1.002772882 seconds time elapsed
```

Knowing that both groups were broken there feels like a value add.
Given that this is a warning, and it can be fixed by moving the event
out of the group or forcing the CPUs, I lean toward being
informative/spammy as the spam is somewhat straightforwardly fixed on
the command line.

Thanks,
Ian

> Thanks,
> Kan
> >
> > Signed-off-by: Ian Rogers <irogers@google.com>
> > ---
> >  tools/perf/builtin-stat.c | 24 +++++++++++++++---------
> >  1 file changed, 15 insertions(+), 9 deletions(-)
> >
> > diff --git a/tools/perf/builtin-stat.c b/tools/perf/builtin-stat.c
> > index d70b1ec88594..5c12ae5efce5 100644
> > --- a/tools/perf/builtin-stat.c
> > +++ b/tools/perf/builtin-stat.c
> > @@ -181,14 +181,13 @@ static bool cpus_map_matched(struct evsel *a, struct evsel *b)
> >
> >  static void evlist__check_cpu_maps(struct evlist *evlist)
> >  {
> > -     struct evsel *evsel, *pos, *leader;
> > -     char buf[1024];
> > +     struct evsel *evsel, *warned_leader = NULL;
> >
> >       if (evlist__has_hybrid(evlist))
> >               evlist__warn_hybrid_group(evlist);
> >
> >       evlist__for_each_entry(evlist, evsel) {
> > -             leader = evsel__leader(evsel);
> > +             struct evsel *leader = evsel__leader(evsel);
> >
> >               /* Check that leader matches cpus with each member. */
> >               if (leader == evsel)
> > @@ -197,19 +196,26 @@ static void evlist__check_cpu_maps(struct evlist *evlist)
> >                       continue;
> >
> >               /* If there's mismatch disable the group and warn user. */
> > -             WARN_ONCE(1, "WARNING: grouped events cpus do not match, disabling group:\n");
> > -             evsel__group_desc(leader, buf, sizeof(buf));
> > -             pr_warning("  %s\n", buf);
> > -
> > +             if (warned_leader != leader) {
> > +                     char buf[200];
> > +
> > +                     pr_warning("WARNING: grouped events cpus do not match.\n"
> > +                             "Events with CPUs not matching the leader will "
> > +                             "be removed from the group.\n");
> > +                     evsel__group_desc(leader, buf, sizeof(buf));
> > +                     pr_warning("  %s\n", buf);
> > +                     warned_leader = leader;
> > +             }
> >               if (verbose > 0) {
> > +                     char buf[200];
> > +
> >                       cpu_map__snprint(leader->core.cpus, buf, sizeof(buf));
> >                       pr_warning("     %s: %s\n", leader->name, buf);
> >                       cpu_map__snprint(evsel->core.cpus, buf, sizeof(buf));
> >                       pr_warning("     %s: %s\n", evsel->name, buf);
> >               }
> >
> > -             for_each_group_evsel(pos, leader)
> > -                     evsel__remove_from_group(pos, leader);
> > +             evsel__remove_from_group(evsel, leader);
> >       }
> >  }
> >

Re: [PATCH v2 02/10] perf stat: Don't remove all grouped events when CPU maps disagree

[Liang, Kan]

On 2023-03-03 11:44 a.m., Ian Rogers wrote:
> On Fri, Mar 3, 2023 at 7:50 AM Liang, Kan <kan.liang@linux.intel.com> wrote:
>>
>>
>>
>> On 2023-03-02 4:25 p.m., Ian Rogers wrote:
>>> If the events in an evlist's CPU map differ then the entire group is
>>> removed. For example:
>>>
>>> ```
>>> $ perf stat -e '{imc_free_running/data_read/,imc_free_running/data_write/,cs}' -a sleep 1
>>> WARNING: grouped events cpus do not match, disabling group:
>>>   anon group { imc_free_running/data_read/, imc_free_running/data_write/, cs }
>>> ```
>>>
>>> Change the behavior so that just the events not matching the leader
>>> are removed. So in the example above, just 'cs' will be removed.
>>>
>>> Modify the warning so that it is produced once for each group, rather
>>> than once for the entire evlist. Shrink the scope and size of the
>>> warning text buffer.
>>
>> For the uncore, we usually have to create a group for each uncore PMU.
>> The number of groups may be big. For example, on ICX, we have 40 CHA
>> PMUs. For SPR, there should be more CHAs. If we have something like
>> {cycles,uncore_cha/event=0x1/}, is the warning shown 40 times on ICX?
>> If so, it should be very annoying.
>>
>> Maybe it's better to keep the current behavior which only print a
>> warning once and notify the users that perf will re-group the events.
>> For the details, they can get it from the -v option.
> 
> Thanks Kan, I could imagine that but I was also worried about cases
> where there are multiple groups like:
> 
> ```
> $ perf stat -e '{imc_free_running/data_read/,cs},{uncore_clock/clockticks/,cs}'
> -a sleep 1
> WARNING: grouped events cpus do not match.
> Events with CPUs not matching the leader will be removed from the group.
>  anon group { imc_free_running/data_read/, cs }
> WARNING: grouped events cpus do not match.
> Events with CPUs not matching the leader will be removed from the group.
>  anon group { uncore_clock/clockticks/, cs }
> 
> Performance counter stats for 'system wide':
> 
>          1,255.75 MiB  imc_free_running/data_read/
>             7,571      cs
>     1,327,285,527      uncore_clock/clockticks/
>             7,571      cs
> 
>       1.002772882 seconds time elapsed
> ```
> 
> Knowing that both groups were broken there feels like a value add.
> Given that this is a warning, and it can be fixed by moving the event
> out of the group or forcing the CPUs, I lean toward being
> informative/spammy as the spam is somewhat straightforwardly fixed on
> the command line.


I did some tests with the patch. The issue I was worried about didn't
occur. The change looks good to me.

But I found another issue. If I specify a CPU set, the group removal
fails. It's not an issue of this patch. It looks like the current perf
compares the user defined CPU set, rather than the PMU's cpumask.

./perf stat -e '{cycles,uncore_cha/event=0x1/}' -C0 sleep 1

 Performance counter stats for 'CPU(s) 0':

     <not counted>      cycles
   <not supported>      uncore_cha/event=0x1/

       1.001783936 seconds time elapsed

Some events weren't counted. Try disabling the NMI watchdog:
        echo 0 > /proc/sys/kernel/nmi_watchdog
        perf stat ...
        echo 1 > /proc/sys/kernel/nmi_watchdog
The events in group usually have to be from the same PMU. Try
reorganizing the group.

Thanks,
Kan

> 
> Thanks,
> Ian
> 
>> Thanks,
>> Kan
>>>
>>> Signed-off-by: Ian Rogers <irogers@google.com>
>>> ---
>>>  tools/perf/builtin-stat.c | 24 +++++++++++++++---------
>>>  1 file changed, 15 insertions(+), 9 deletions(-)
>>>
>>> diff --git a/tools/perf/builtin-stat.c b/tools/perf/builtin-stat.c
>>> index d70b1ec88594..5c12ae5efce5 100644
>>> --- a/tools/perf/builtin-stat.c
>>> +++ b/tools/perf/builtin-stat.c
>>> @@ -181,14 +181,13 @@ static bool cpus_map_matched(struct evsel *a, struct evsel *b)
>>>
>>>  static void evlist__check_cpu_maps(struct evlist *evlist)
>>>  {
>>> -     struct evsel *evsel, *pos, *leader;
>>> -     char buf[1024];
>>> +     struct evsel *evsel, *warned_leader = NULL;
>>>
>>>       if (evlist__has_hybrid(evlist))
>>>               evlist__warn_hybrid_group(evlist);
>>>
>>>       evlist__for_each_entry(evlist, evsel) {
>>> -             leader = evsel__leader(evsel);
>>> +             struct evsel *leader = evsel__leader(evsel);
>>>
>>>               /* Check that leader matches cpus with each member. */
>>>               if (leader == evsel)
>>> @@ -197,19 +196,26 @@ static void evlist__check_cpu_maps(struct evlist *evlist)
>>>                       continue;
>>>
>>>               /* If there's mismatch disable the group and warn user. */
>>> -             WARN_ONCE(1, "WARNING: grouped events cpus do not match, disabling group:\n");
>>> -             evsel__group_desc(leader, buf, sizeof(buf));
>>> -             pr_warning("  %s\n", buf);
>>> -
>>> +             if (warned_leader != leader) {
>>> +                     char buf[200];
>>> +
>>> +                     pr_warning("WARNING: grouped events cpus do not match.\n"
>>> +                             "Events with CPUs not matching the leader will "
>>> +                             "be removed from the group.\n");
>>> +                     evsel__group_desc(leader, buf, sizeof(buf));
>>> +                     pr_warning("  %s\n", buf);
>>> +                     warned_leader = leader;
>>> +             }
>>>               if (verbose > 0) {
>>> +                     char buf[200];
>>> +
>>>                       cpu_map__snprint(leader->core.cpus, buf, sizeof(buf));
>>>                       pr_warning("     %s: %s\n", leader->name, buf);
>>>                       cpu_map__snprint(evsel->core.cpus, buf, sizeof(buf));
>>>                       pr_warning("     %s: %s\n", evsel->name, buf);
>>>               }
>>>
>>> -             for_each_group_evsel(pos, leader)
>>> -                     evsel__remove_from_group(pos, leader);
>>> +             evsel__remove_from_group(evsel, leader);
>>>       }
>>>  }
>>>

[PATCH v2 03/10] perf record: Early auxtrace initialization before event parsing

[Ian Rogers]

This allows event parsing to use the evsel__is_aux_event function,
which is important when determining event grouping.

Signed-off-by: Ian Rogers <irogers@google.com>
---
 tools/perf/arch/x86/util/auxtrace.c | 17 +++++++++++++----
 tools/perf/builtin-record.c         |  6 ++++++
 tools/perf/util/auxtrace.h          |  2 ++
 3 files changed, 21 insertions(+), 4 deletions(-)

diff --git a/tools/perf/arch/x86/util/auxtrace.c b/tools/perf/arch/x86/util/auxtrace.c
index 3da506e13f49..de1e4842ea2e 100644
--- a/tools/perf/arch/x86/util/auxtrace.c
+++ b/tools/perf/arch/x86/util/auxtrace.c
@@ -15,6 +15,19 @@
 #include "../../../util/intel-bts.h"
 #include "../../../util/evlist.h"
 
+void auxtrace__early_init(void)
+{
+	struct perf_pmu *intel_pt_pmu;
+	struct perf_pmu *intel_bts_pmu;
+
+	intel_pt_pmu = perf_pmu__find(INTEL_PT_PMU_NAME);
+	if (intel_pt_pmu)
+		intel_pt_pmu->auxtrace = true;
+	intel_bts_pmu = perf_pmu__find(INTEL_BTS_PMU_NAME);
+	if (intel_bts_pmu)
+		intel_bts_pmu->auxtrace = true;
+}
+
 static
 struct auxtrace_record *auxtrace_record__init_intel(struct evlist *evlist,
 						    int *err)
@@ -26,11 +39,7 @@ struct auxtrace_record *auxtrace_record__init_intel(struct evlist *evlist,
 	bool found_bts = false;
 
 	intel_pt_pmu = perf_pmu__find(INTEL_PT_PMU_NAME);
-	if (intel_pt_pmu)
-		intel_pt_pmu->auxtrace = true;
 	intel_bts_pmu = perf_pmu__find(INTEL_BTS_PMU_NAME);
-	if (intel_bts_pmu)
-		intel_bts_pmu->auxtrace = true;
 
 	evlist__for_each_entry(evlist, evsel) {
 		if (intel_pt_pmu && evsel->core.attr.type == intel_pt_pmu->type)
diff --git a/tools/perf/builtin-record.c b/tools/perf/builtin-record.c
index 8374117e66f6..a0870c076dc0 100644
--- a/tools/perf/builtin-record.c
+++ b/tools/perf/builtin-record.c
@@ -3940,6 +3940,10 @@ static int record__init_thread_masks(struct record *rec)
 	return ret;
 }
 
+__weak void auxtrace__early_init(void)
+{
+}
+
 int cmd_record(int argc, const char **argv)
 {
 	int err;
@@ -3985,6 +3989,8 @@ int cmd_record(int argc, const char **argv)
 	if (err)
 		return err;
 
+	auxtrace__early_init();
+
 	argc = parse_options(argc, argv, record_options, record_usage,
 			    PARSE_OPT_STOP_AT_NON_OPTION);
 	if (quiet)
diff --git a/tools/perf/util/auxtrace.h b/tools/perf/util/auxtrace.h
index 29eb82dff574..49a86aa6ac94 100644
--- a/tools/perf/util/auxtrace.h
+++ b/tools/perf/util/auxtrace.h
@@ -457,6 +457,8 @@ struct addr_filters {
 
 struct auxtrace_cache;
 
+void auxtrace__early_init(void);
+
 #ifdef HAVE_AUXTRACE_SUPPORT
 
 u64 compat_auxtrace_mmap__read_head(struct auxtrace_mmap *mm);
-- 
2.40.0.rc0.216.gc4246ad0f0-goog

Re: [PATCH v2 03/10] perf record: Early auxtrace initialization before event parsing

[Liang, Kan]

On 2023-03-02 4:25 p.m., Ian Rogers wrote:
> This allows event parsing to use the evsel__is_aux_event function,
> which is important when determining event grouping.
> 
> Signed-off-by: Ian Rogers <irogers@google.com>
> ---
>  tools/perf/arch/x86/util/auxtrace.c | 17 +++++++++++++----
>  tools/perf/builtin-record.c         |  6 ++++++
>  tools/perf/util/auxtrace.h          |  2 ++
>  3 files changed, 21 insertions(+), 4 deletions(-)
> 
> diff --git a/tools/perf/arch/x86/util/auxtrace.c b/tools/perf/arch/x86/util/auxtrace.c
> index 3da506e13f49..de1e4842ea2e 100644
> --- a/tools/perf/arch/x86/util/auxtrace.c
> +++ b/tools/perf/arch/x86/util/auxtrace.c
> @@ -15,6 +15,19 @@
>  #include "../../../util/intel-bts.h"
>  #include "../../../util/evlist.h"
>  
> +void auxtrace__early_init(void)
> +{
> +	struct perf_pmu *intel_pt_pmu;
> +	struct perf_pmu *intel_bts_pmu;
> +
> +	intel_pt_pmu = perf_pmu__find(INTEL_PT_PMU_NAME);
> +	if (intel_pt_pmu)
> +		intel_pt_pmu->auxtrace = true;
> +	intel_bts_pmu = perf_pmu__find(INTEL_BTS_PMU_NAME);
> +	if (intel_bts_pmu)
> +		intel_bts_pmu->auxtrace = true;
> +}
> +
>  static
>  struct auxtrace_record *auxtrace_record__init_intel(struct evlist *evlist,
>  						    int *err)
> @@ -26,11 +39,7 @@ struct auxtrace_record *auxtrace_record__init_intel(struct evlist *evlist,
>  	bool found_bts = false;
>  
>  	intel_pt_pmu = perf_pmu__find(INTEL_PT_PMU_NAME);
> -	if (intel_pt_pmu)
> -		intel_pt_pmu->auxtrace = true;
>  	intel_bts_pmu = perf_pmu__find(INTEL_BTS_PMU_NAME);
> -	if (intel_bts_pmu)
> -		intel_bts_pmu->auxtrace = true;
>  
>  	evlist__for_each_entry(evlist, evsel) {
>  		if (intel_pt_pmu && evsel->core.attr.type == intel_pt_pmu->type)
> diff --git a/tools/perf/builtin-record.c b/tools/perf/builtin-record.c
> index 8374117e66f6..a0870c076dc0 100644
> --- a/tools/perf/builtin-record.c
> +++ b/tools/perf/builtin-record.c
> @@ -3940,6 +3940,10 @@ static int record__init_thread_masks(struct record *rec)
>  	return ret;
>  }
>  
> +__weak void auxtrace__early_init(void)
> +{
> +}
> +
>  int cmd_record(int argc, const char **argv)
>  {
>  	int err;
> @@ -3985,6 +3989,8 @@ int cmd_record(int argc, const char **argv)
>  	if (err)
>  		return err;
>  
> +	auxtrace__early_init();

So the auxtrace__early_init() will be unconditionally invoked even there
is no PT or BTS events, right?

Maybe we should move the auxtrace__early_init() to evsel__is_aux_event()
and cache the value. The initialization will only be invoked when it's
required.
Something as below (not tested.)

+void auxtrace__init(void)
+{
+	struct perf_pmu *intel_pt_pmu;
+	struct perf_pmu *intel_bts_pmu;
+	static bool cached;
+
+	if (cached)
+		return;
+	intel_pt_pmu = perf_pmu__find(INTEL_PT_PMU_NAME);
+	if (intel_pt_pmu)
+		intel_pt_pmu->auxtrace = true;
+	intel_bts_pmu = perf_pmu__find(INTEL_BTS_PMU_NAME);
+	if (intel_bts_pmu)
+		intel_bts_pmu->auxtrace = true;
+}

bool evsel__is_aux_event(struct evsel *evsel)
{
	struct perf_pmu *pmu = evsel__find_pmu(evsel);
+	auxtrace__init();
	return pmu && pmu->auxtrace;
}



Thanks,
Kan

> +
>  	argc = parse_options(argc, argv, record_options, record_usage,
>  			    PARSE_OPT_STOP_AT_NON_OPTION);
>  	if (quiet)
> diff --git a/tools/perf/util/auxtrace.h b/tools/perf/util/auxtrace.h
> index 29eb82dff574..49a86aa6ac94 100644
> --- a/tools/perf/util/auxtrace.h
> +++ b/tools/perf/util/auxtrace.h
> @@ -457,6 +457,8 @@ struct addr_filters {
>  
>  struct auxtrace_cache;
>  
> +void auxtrace__early_init(void);
> +
>  #ifdef HAVE_AUXTRACE_SUPPORT
>  
>  u64 compat_auxtrace_mmap__read_head(struct auxtrace_mmap *mm);

Re: [PATCH v2 03/10] perf record: Early auxtrace initialization before event parsing

[Adrian Hunter]

On 3/03/23 18:40, Liang, Kan wrote:
> 
> 
> On 2023-03-02 4:25 p.m., Ian Rogers wrote:
>> This allows event parsing to use the evsel__is_aux_event function,
>> which is important when determining event grouping.
>>
>> Signed-off-by: Ian Rogers <irogers@google.com>
>> ---
>>  tools/perf/arch/x86/util/auxtrace.c | 17 +++++++++++++----
>>  tools/perf/builtin-record.c         |  6 ++++++
>>  tools/perf/util/auxtrace.h          |  2 ++
>>  3 files changed, 21 insertions(+), 4 deletions(-)
>>
>> diff --git a/tools/perf/arch/x86/util/auxtrace.c b/tools/perf/arch/x86/util/auxtrace.c
>> index 3da506e13f49..de1e4842ea2e 100644
>> --- a/tools/perf/arch/x86/util/auxtrace.c
>> +++ b/tools/perf/arch/x86/util/auxtrace.c
>> @@ -15,6 +15,19 @@
>>  #include "../../../util/intel-bts.h"
>>  #include "../../../util/evlist.h"
>>  
>> +void auxtrace__early_init(void)
>> +{
>> +	struct perf_pmu *intel_pt_pmu;
>> +	struct perf_pmu *intel_bts_pmu;
>> +
>> +	intel_pt_pmu = perf_pmu__find(INTEL_PT_PMU_NAME);
>> +	if (intel_pt_pmu)
>> +		intel_pt_pmu->auxtrace = true;
>> +	intel_bts_pmu = perf_pmu__find(INTEL_BTS_PMU_NAME);
>> +	if (intel_bts_pmu)
>> +		intel_bts_pmu->auxtrace = true;
>> +}
>> +
>>  static
>>  struct auxtrace_record *auxtrace_record__init_intel(struct evlist *evlist,
>>  						    int *err)
>> @@ -26,11 +39,7 @@ struct auxtrace_record *auxtrace_record__init_intel(struct evlist *evlist,
>>  	bool found_bts = false;
>>  
>>  	intel_pt_pmu = perf_pmu__find(INTEL_PT_PMU_NAME);
>> -	if (intel_pt_pmu)
>> -		intel_pt_pmu->auxtrace = true;
>>  	intel_bts_pmu = perf_pmu__find(INTEL_BTS_PMU_NAME);
>> -	if (intel_bts_pmu)
>> -		intel_bts_pmu->auxtrace = true;
>>  
>>  	evlist__for_each_entry(evlist, evsel) {
>>  		if (intel_pt_pmu && evsel->core.attr.type == intel_pt_pmu->type)
>> diff --git a/tools/perf/builtin-record.c b/tools/perf/builtin-record.c
>> index 8374117e66f6..a0870c076dc0 100644
>> --- a/tools/perf/builtin-record.c
>> +++ b/tools/perf/builtin-record.c
>> @@ -3940,6 +3940,10 @@ static int record__init_thread_masks(struct record *rec)
>>  	return ret;
>>  }
>>  
>> +__weak void auxtrace__early_init(void)
>> +{
>> +}
>> +
>>  int cmd_record(int argc, const char **argv)
>>  {
>>  	int err;
>> @@ -3985,6 +3989,8 @@ int cmd_record(int argc, const char **argv)
>>  	if (err)
>>  		return err;
>>  
>> +	auxtrace__early_init();
> 
> So the auxtrace__early_init() will be unconditionally invoked even there
> is no PT or BTS events, right?
> 
> Maybe we should move the auxtrace__early_init() to evsel__is_aux_event()
> and cache the value. The initialization will only be invoked when it's
> required.

Although perf_pmu__find() will be called unconditionally via
record__auxtrace_init() anyway.

> Something as below (not tested.)
> 
> +void auxtrace__init(void)
> +{
> +	struct perf_pmu *intel_pt_pmu;
> +	struct perf_pmu *intel_bts_pmu;
> +	static bool cached;
> +
> +	if (cached)
> +		return;
> +	intel_pt_pmu = perf_pmu__find(INTEL_PT_PMU_NAME);
> +	if (intel_pt_pmu)
> +		intel_pt_pmu->auxtrace = true;
> +	intel_bts_pmu = perf_pmu__find(INTEL_BTS_PMU_NAME);
> +	if (intel_bts_pmu)
> +		intel_bts_pmu->auxtrace = true;
> +}
> 
> bool evsel__is_aux_event(struct evsel *evsel)
> {
> 	struct perf_pmu *pmu = evsel__find_pmu(evsel);
> +	auxtrace__init();
> 	return pmu && pmu->auxtrace;
> }
> 
> 
> 
> Thanks,
> Kan
> 
>> +
>>  	argc = parse_options(argc, argv, record_options, record_usage,
>>  			    PARSE_OPT_STOP_AT_NON_OPTION);
>>  	if (quiet)
>> diff --git a/tools/perf/util/auxtrace.h b/tools/perf/util/auxtrace.h
>> index 29eb82dff574..49a86aa6ac94 100644
>> --- a/tools/perf/util/auxtrace.h
>> +++ b/tools/perf/util/auxtrace.h
>> @@ -457,6 +457,8 @@ struct addr_filters {
>>  
>>  struct auxtrace_cache;
>>  
>> +void auxtrace__early_init(void);
>> +
>>  #ifdef HAVE_AUXTRACE_SUPPORT
>>  
>>  u64 compat_auxtrace_mmap__read_head(struct auxtrace_mmap *mm);

Re: [PATCH v2 03/10] perf record: Early auxtrace initialization before event parsing

[Adrian Hunter]

On 5/03/23 10:32, Adrian Hunter wrote:
> On 3/03/23 18:40, Liang, Kan wrote:
>>
>>
>> On 2023-03-02 4:25 p.m., Ian Rogers wrote:
>>> This allows event parsing to use the evsel__is_aux_event function,
>>> which is important when determining event grouping.
>>>
>>> Signed-off-by: Ian Rogers <irogers@google.com>
>>> ---
>>>  tools/perf/arch/x86/util/auxtrace.c | 17 +++++++++++++----
>>>  tools/perf/builtin-record.c         |  6 ++++++
>>>  tools/perf/util/auxtrace.h          |  2 ++
>>>  3 files changed, 21 insertions(+), 4 deletions(-)
>>>
>>> diff --git a/tools/perf/arch/x86/util/auxtrace.c b/tools/perf/arch/x86/util/auxtrace.c
>>> index 3da506e13f49..de1e4842ea2e 100644
>>> --- a/tools/perf/arch/x86/util/auxtrace.c
>>> +++ b/tools/perf/arch/x86/util/auxtrace.c
>>> @@ -15,6 +15,19 @@
>>>  #include "../../../util/intel-bts.h"
>>>  #include "../../../util/evlist.h"
>>>  
>>> +void auxtrace__early_init(void)
>>> +{
>>> +	struct perf_pmu *intel_pt_pmu;
>>> +	struct perf_pmu *intel_bts_pmu;
>>> +
>>> +	intel_pt_pmu = perf_pmu__find(INTEL_PT_PMU_NAME);
>>> +	if (intel_pt_pmu)
>>> +		intel_pt_pmu->auxtrace = true;
>>> +	intel_bts_pmu = perf_pmu__find(INTEL_BTS_PMU_NAME);
>>> +	if (intel_bts_pmu)
>>> +		intel_bts_pmu->auxtrace = true;
>>> +}
>>> +
>>>  static
>>>  struct auxtrace_record *auxtrace_record__init_intel(struct evlist *evlist,
>>>  						    int *err)
>>> @@ -26,11 +39,7 @@ struct auxtrace_record *auxtrace_record__init_intel(struct evlist *evlist,
>>>  	bool found_bts = false;
>>>  
>>>  	intel_pt_pmu = perf_pmu__find(INTEL_PT_PMU_NAME);
>>> -	if (intel_pt_pmu)
>>> -		intel_pt_pmu->auxtrace = true;
>>>  	intel_bts_pmu = perf_pmu__find(INTEL_BTS_PMU_NAME);
>>> -	if (intel_bts_pmu)
>>> -		intel_bts_pmu->auxtrace = true;
>>>  
>>>  	evlist__for_each_entry(evlist, evsel) {
>>>  		if (intel_pt_pmu && evsel->core.attr.type == intel_pt_pmu->type)
>>> diff --git a/tools/perf/builtin-record.c b/tools/perf/builtin-record.c
>>> index 8374117e66f6..a0870c076dc0 100644
>>> --- a/tools/perf/builtin-record.c
>>> +++ b/tools/perf/builtin-record.c
>>> @@ -3940,6 +3940,10 @@ static int record__init_thread_masks(struct record *rec)
>>>  	return ret;
>>>  }
>>>  
>>> +__weak void auxtrace__early_init(void)
>>> +{
>>> +}
>>> +
>>>  int cmd_record(int argc, const char **argv)
>>>  {
>>>  	int err;
>>> @@ -3985,6 +3989,8 @@ int cmd_record(int argc, const char **argv)
>>>  	if (err)
>>>  		return err;
>>>  
>>> +	auxtrace__early_init();
>>
>> So the auxtrace__early_init() will be unconditionally invoked even there
>> is no PT or BTS events, right?
>>
>> Maybe we should move the auxtrace__early_init() to evsel__is_aux_event()
>> and cache the value. The initialization will only be invoked when it's
>> required.
> 
> Although perf_pmu__find() will be called unconditionally via
> record__auxtrace_init() anyway.

However auxtrace__early_init() is before parsing 'verbose' so
debug prints don't work anymore.

How about this instead:

diff --git a/tools/perf/arch/x86/util/auxtrace.c
b/tools/perf/arch/x86/util/auxtrace.c
index 3da506e13f49d..330d03216b0e6 100644
--- a/tools/perf/arch/x86/util/auxtrace.c
+++ b/tools/perf/arch/x86/util/auxtrace.c
@@ -26,11 +26,7 @@ struct auxtrace_record
*auxtrace_record__init_intel(struct evlist *evlist,
 	bool found_bts = false;

 	intel_pt_pmu = perf_pmu__find(INTEL_PT_PMU_NAME);
-	if (intel_pt_pmu)
-		intel_pt_pmu->auxtrace = true;
 	intel_bts_pmu = perf_pmu__find(INTEL_BTS_PMU_NAME);
-	if (intel_bts_pmu)
-		intel_bts_pmu->auxtrace = true;

 	evlist__for_each_entry(evlist, evsel) {
 		if (intel_pt_pmu && evsel->core.attr.type == intel_pt_pmu->type)
diff --git a/tools/perf/arch/x86/util/pmu.c b/tools/perf/arch/x86/util/pmu.c
index 358340b342431..f73b80dcd8bdc 100644
--- a/tools/perf/arch/x86/util/pmu.c
+++ b/tools/perf/arch/x86/util/pmu.c
@@ -27,10 +27,14 @@ static bool cached_list;
 struct perf_event_attr *perf_pmu__get_default_config(struct perf_pmu
*pmu __maybe_unused)
 {
 #ifdef HAVE_AUXTRACE_SUPPORT
-	if (!strcmp(pmu->name, INTEL_PT_PMU_NAME))
+	if (!strcmp(pmu->name, INTEL_PT_PMU_NAME)) {
+		pmu->auxtrace = true;
 		return intel_pt_pmu_default_config(pmu);
-	if (!strcmp(pmu->name, INTEL_BTS_PMU_NAME))
+	}
+	if (!strcmp(pmu->name, INTEL_BTS_PMU_NAME)) {
+		pmu->auxtrace = true;
 		pmu->selectable = true;
+	}
 #endif
 	return NULL;
 }



> 
>> Something as below (not tested.)
>>
>> +void auxtrace__init(void)
>> +{
>> +	struct perf_pmu *intel_pt_pmu;
>> +	struct perf_pmu *intel_bts_pmu;
>> +	static bool cached;
>> +
>> +	if (cached)
>> +		return;
>> +	intel_pt_pmu = perf_pmu__find(INTEL_PT_PMU_NAME);
>> +	if (intel_pt_pmu)
>> +		intel_pt_pmu->auxtrace = true;
>> +	intel_bts_pmu = perf_pmu__find(INTEL_BTS_PMU_NAME);
>> +	if (intel_bts_pmu)
>> +		intel_bts_pmu->auxtrace = true;
>> +}
>>
>> bool evsel__is_aux_event(struct evsel *evsel)
>> {
>> 	struct perf_pmu *pmu = evsel__find_pmu(evsel);
>> +	auxtrace__init();
>> 	return pmu && pmu->auxtrace;
>> }
>>
>>
>>
>> Thanks,
>> Kan
>>
>>> +
>>>  	argc = parse_options(argc, argv, record_options, record_usage,
>>>  			    PARSE_OPT_STOP_AT_NON_OPTION);
>>>  	if (quiet)
>>> diff --git a/tools/perf/util/auxtrace.h b/tools/perf/util/auxtrace.h
>>> index 29eb82dff574..49a86aa6ac94 100644
>>> --- a/tools/perf/util/auxtrace.h
>>> +++ b/tools/perf/util/auxtrace.h
>>> @@ -457,6 +457,8 @@ struct addr_filters {
>>>  
>>>  struct auxtrace_cache;
>>>  
>>> +void auxtrace__early_init(void);
>>> +
>>>  #ifdef HAVE_AUXTRACE_SUPPORT
>>>  
>>>  u64 compat_auxtrace_mmap__read_head(struct auxtrace_mmap *mm);
> 

Re: [PATCH v2 03/10] perf record: Early auxtrace initialization before event parsing

[Liang, Kan]

On 2023-03-06 4:31 a.m., Adrian Hunter wrote:
> On 5/03/23 10:32, Adrian Hunter wrote:
>> On 3/03/23 18:40, Liang, Kan wrote:
>>>
>>>
>>> On 2023-03-02 4:25 p.m., Ian Rogers wrote:
>>>> This allows event parsing to use the evsel__is_aux_event function,
>>>> which is important when determining event grouping.
>>>>
>>>> Signed-off-by: Ian Rogers <irogers@google.com>
>>>> ---
>>>>  tools/perf/arch/x86/util/auxtrace.c | 17 +++++++++++++----
>>>>  tools/perf/builtin-record.c         |  6 ++++++
>>>>  tools/perf/util/auxtrace.h          |  2 ++
>>>>  3 files changed, 21 insertions(+), 4 deletions(-)
>>>>
>>>> diff --git a/tools/perf/arch/x86/util/auxtrace.c b/tools/perf/arch/x86/util/auxtrace.c
>>>> index 3da506e13f49..de1e4842ea2e 100644
>>>> --- a/tools/perf/arch/x86/util/auxtrace.c
>>>> +++ b/tools/perf/arch/x86/util/auxtrace.c
>>>> @@ -15,6 +15,19 @@
>>>>  #include "../../../util/intel-bts.h"
>>>>  #include "../../../util/evlist.h"
>>>>  
>>>> +void auxtrace__early_init(void)
>>>> +{
>>>> +	struct perf_pmu *intel_pt_pmu;
>>>> +	struct perf_pmu *intel_bts_pmu;
>>>> +
>>>> +	intel_pt_pmu = perf_pmu__find(INTEL_PT_PMU_NAME);
>>>> +	if (intel_pt_pmu)
>>>> +		intel_pt_pmu->auxtrace = true;
>>>> +	intel_bts_pmu = perf_pmu__find(INTEL_BTS_PMU_NAME);
>>>> +	if (intel_bts_pmu)
>>>> +		intel_bts_pmu->auxtrace = true;
>>>> +}
>>>> +
>>>>  static
>>>>  struct auxtrace_record *auxtrace_record__init_intel(struct evlist *evlist,
>>>>  						    int *err)
>>>> @@ -26,11 +39,7 @@ struct auxtrace_record *auxtrace_record__init_intel(struct evlist *evlist,
>>>>  	bool found_bts = false;
>>>>  
>>>>  	intel_pt_pmu = perf_pmu__find(INTEL_PT_PMU_NAME);
>>>> -	if (intel_pt_pmu)
>>>> -		intel_pt_pmu->auxtrace = true;
>>>>  	intel_bts_pmu = perf_pmu__find(INTEL_BTS_PMU_NAME);
>>>> -	if (intel_bts_pmu)
>>>> -		intel_bts_pmu->auxtrace = true;
>>>>  
>>>>  	evlist__for_each_entry(evlist, evsel) {
>>>>  		if (intel_pt_pmu && evsel->core.attr.type == intel_pt_pmu->type)
>>>> diff --git a/tools/perf/builtin-record.c b/tools/perf/builtin-record.c
>>>> index 8374117e66f6..a0870c076dc0 100644
>>>> --- a/tools/perf/builtin-record.c
>>>> +++ b/tools/perf/builtin-record.c
>>>> @@ -3940,6 +3940,10 @@ static int record__init_thread_masks(struct record *rec)
>>>>  	return ret;
>>>>  }
>>>>  
>>>> +__weak void auxtrace__early_init(void)
>>>> +{
>>>> +}
>>>> +
>>>>  int cmd_record(int argc, const char **argv)
>>>>  {
>>>>  	int err;
>>>> @@ -3985,6 +3989,8 @@ int cmd_record(int argc, const char **argv)
>>>>  	if (err)
>>>>  		return err;
>>>>  
>>>> +	auxtrace__early_init();
>>>
>>> So the auxtrace__early_init() will be unconditionally invoked even there
>>> is no PT or BTS events, right?
>>>
>>> Maybe we should move the auxtrace__early_init() to evsel__is_aux_event()
>>> and cache the value. The initialization will only be invoked when it's
>>> required.
>>
>> Although perf_pmu__find() will be called unconditionally via
>> record__auxtrace_init() anyway.
> 
> However auxtrace__early_init() is before parsing 'verbose' so
> debug prints don't work anymore.
> 
> How about this instead:

Yes, I think it should be a better place to initialize them.

Thanks,
Kan
> 
> diff --git a/tools/perf/arch/x86/util/auxtrace.c
> b/tools/perf/arch/x86/util/auxtrace.c
> index 3da506e13f49d..330d03216b0e6 100644
> --- a/tools/perf/arch/x86/util/auxtrace.c
> +++ b/tools/perf/arch/x86/util/auxtrace.c
> @@ -26,11 +26,7 @@ struct auxtrace_record
> *auxtrace_record__init_intel(struct evlist *evlist,
>  	bool found_bts = false;
> 
>  	intel_pt_pmu = perf_pmu__find(INTEL_PT_PMU_NAME);
> -	if (intel_pt_pmu)
> -		intel_pt_pmu->auxtrace = true;
>  	intel_bts_pmu = perf_pmu__find(INTEL_BTS_PMU_NAME);
> -	if (intel_bts_pmu)
> -		intel_bts_pmu->auxtrace = true;
> 
>  	evlist__for_each_entry(evlist, evsel) {
>  		if (intel_pt_pmu && evsel->core.attr.type == intel_pt_pmu->type)
> diff --git a/tools/perf/arch/x86/util/pmu.c b/tools/perf/arch/x86/util/pmu.c
> index 358340b342431..f73b80dcd8bdc 100644
> --- a/tools/perf/arch/x86/util/pmu.c
> +++ b/tools/perf/arch/x86/util/pmu.c
> @@ -27,10 +27,14 @@ static bool cached_list;
>  struct perf_event_attr *perf_pmu__get_default_config(struct perf_pmu
> *pmu __maybe_unused)
>  {
>  #ifdef HAVE_AUXTRACE_SUPPORT
> -	if (!strcmp(pmu->name, INTEL_PT_PMU_NAME))
> +	if (!strcmp(pmu->name, INTEL_PT_PMU_NAME)) {
> +		pmu->auxtrace = true;
>  		return intel_pt_pmu_default_config(pmu);
> -	if (!strcmp(pmu->name, INTEL_BTS_PMU_NAME))
> +	}
> +	if (!strcmp(pmu->name, INTEL_BTS_PMU_NAME)) {
> +		pmu->auxtrace = true;
>  		pmu->selectable = true;
> +	}
>  #endif
>  	return NULL;
>  }
> 
> 
> 
>>
>>> Something as below (not tested.)
>>>
>>> +void auxtrace__init(void)
>>> +{
>>> +	struct perf_pmu *intel_pt_pmu;
>>> +	struct perf_pmu *intel_bts_pmu;
>>> +	static bool cached;
>>> +
>>> +	if (cached)
>>> +		return;
>>> +	intel_pt_pmu = perf_pmu__find(INTEL_PT_PMU_NAME);
>>> +	if (intel_pt_pmu)
>>> +		intel_pt_pmu->auxtrace = true;
>>> +	intel_bts_pmu = perf_pmu__find(INTEL_BTS_PMU_NAME);
>>> +	if (intel_bts_pmu)
>>> +		intel_bts_pmu->auxtrace = true;
>>> +}
>>>
>>> bool evsel__is_aux_event(struct evsel *evsel)
>>> {
>>> 	struct perf_pmu *pmu = evsel__find_pmu(evsel);
>>> +	auxtrace__init();
>>> 	return pmu && pmu->auxtrace;
>>> }
>>>
>>>
>>>
>>> Thanks,
>>> Kan
>>>
>>>> +
>>>>  	argc = parse_options(argc, argv, record_options, record_usage,
>>>>  			    PARSE_OPT_STOP_AT_NON_OPTION);
>>>>  	if (quiet)
>>>> diff --git a/tools/perf/util/auxtrace.h b/tools/perf/util/auxtrace.h
>>>> index 29eb82dff574..49a86aa6ac94 100644
>>>> --- a/tools/perf/util/auxtrace.h
>>>> +++ b/tools/perf/util/auxtrace.h
>>>> @@ -457,6 +457,8 @@ struct addr_filters {
>>>>  
>>>>  struct auxtrace_cache;
>>>>  
>>>> +void auxtrace__early_init(void);
>>>> +
>>>>  #ifdef HAVE_AUXTRACE_SUPPORT
>>>>  
>>>>  u64 compat_auxtrace_mmap__read_head(struct auxtrace_mmap *mm);
>>
> 

[PATCH v2 04/10] perf stat: Modify the group test

[Ian Rogers]

Currently nr_members is 0 for an event with no group, however, they
are always a leader of their own group. A later change will make that
count 1 because the event is its own leader. Make the find_stat logic
consistent with this, an improvement suggested by Namhyung Kim.

Suggested-by: Namhyung Kim <namhyung@kernel.org>
Signed-off-by: Ian Rogers <irogers@google.com>
---
 tools/perf/util/stat-shadow.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/tools/perf/util/stat-shadow.c b/tools/perf/util/stat-shadow.c
index ef85f1ae1ab2..eeccab6751d7 100644
--- a/tools/perf/util/stat-shadow.c
+++ b/tools/perf/util/stat-shadow.c
@@ -163,7 +163,7 @@ static double find_stat(const struct evsel *evsel, int aggr_idx, enum stat_type
 			continue;
 
 		/* Ignore evsels that are part of different groups. */
-		if (evsel->core.leader->nr_members &&
+		if (evsel->core.leader->nr_members > 1 &&
 		    evsel->core.leader != cur->core.leader)
 			continue;
 		/* Ignore evsels with mismatched modifiers. */
-- 
2.40.0.rc0.216.gc4246ad0f0-goog

[PATCH v2 05/10] perf evsel: Allow const evsel for certain accesses

[Ian Rogers]

List sorting, added later to evlist, passes const elements requiring
helper functions to also be const. Make the argument to
evsel__find_pmu, evsel__is_aux_event and evsel__leader const.

Signed-off-by: Ian Rogers <irogers@google.com>
---
 tools/perf/util/evsel.c  | 2 +-
 tools/perf/util/evsel.h  | 6 +++---
 tools/perf/util/pmu.c    | 6 +++---
 tools/perf/util/python.c | 2 +-
 4 files changed, 8 insertions(+), 8 deletions(-)

diff --git a/tools/perf/util/evsel.c b/tools/perf/util/evsel.c
index 51e8ce6edddc..2dc2c24252bb 100644
--- a/tools/perf/util/evsel.c
+++ b/tools/perf/util/evsel.c
@@ -3139,7 +3139,7 @@ bool evsel__is_hybrid(const struct evsel *evsel)
 	return evsel->pmu_name && perf_pmu__is_hybrid(evsel->pmu_name);
 }
 
-struct evsel *evsel__leader(struct evsel *evsel)
+struct evsel *evsel__leader(const struct evsel *evsel)
 {
 	return container_of(evsel->core.leader, struct evsel, core);
 }
diff --git a/tools/perf/util/evsel.h b/tools/perf/util/evsel.h
index 814a49ebb7e3..676c499323e9 100644
--- a/tools/perf/util/evsel.h
+++ b/tools/perf/util/evsel.h
@@ -212,8 +212,8 @@ int evsel__object_config(size_t object_size,
 			 int (*init)(struct evsel *evsel),
 			 void (*fini)(struct evsel *evsel));
 
-struct perf_pmu *evsel__find_pmu(struct evsel *evsel);
-bool evsel__is_aux_event(struct evsel *evsel);
+struct perf_pmu *evsel__find_pmu(const struct evsel *evsel);
+bool evsel__is_aux_event(const struct evsel *evsel);
 
 struct evsel *evsel__new_idx(struct perf_event_attr *attr, int idx);
 
@@ -505,7 +505,7 @@ int evsel__store_ids(struct evsel *evsel, struct evlist *evlist);
 
 void evsel__zero_per_pkg(struct evsel *evsel);
 bool evsel__is_hybrid(const struct evsel *evsel);
-struct evsel *evsel__leader(struct evsel *evsel);
+struct evsel *evsel__leader(const struct evsel *evsel);
 bool evsel__has_leader(struct evsel *evsel, struct evsel *leader);
 bool evsel__is_leader(struct evsel *evsel);
 void evsel__set_leader(struct evsel *evsel, struct evsel *leader);
diff --git a/tools/perf/util/pmu.c b/tools/perf/util/pmu.c
index 43b6182d96b7..45d9b8e28e16 100644
--- a/tools/perf/util/pmu.c
+++ b/tools/perf/util/pmu.c
@@ -988,7 +988,7 @@ struct perf_pmu *perf_pmu__scan(struct perf_pmu *pmu)
 	return NULL;
 }
 
-struct perf_pmu *evsel__find_pmu(struct evsel *evsel)
+struct perf_pmu *evsel__find_pmu(const struct evsel *evsel)
 {
 	struct perf_pmu *pmu = NULL;
 
@@ -1000,11 +1000,11 @@ struct perf_pmu *evsel__find_pmu(struct evsel *evsel)
 			break;
 	}
 
-	evsel->pmu = pmu;
+	((struct evsel *)evsel)->pmu = pmu;
 	return pmu;
 }
 
-bool evsel__is_aux_event(struct evsel *evsel)
+bool evsel__is_aux_event(const struct evsel *evsel)
 {
 	struct perf_pmu *pmu = evsel__find_pmu(evsel);
 
diff --git a/tools/perf/util/python.c b/tools/perf/util/python.c
index 42e8b813d010..ab48ffbb6448 100644
--- a/tools/perf/util/python.c
+++ b/tools/perf/util/python.c
@@ -83,7 +83,7 @@ const char *perf_env__arch(struct perf_env *env __maybe_unused)
  * far, for the perf python binding known usecases, revisit if this become
  * necessary.
  */
-struct perf_pmu *evsel__find_pmu(struct evsel *evsel __maybe_unused)
+struct perf_pmu *evsel__find_pmu(const struct evsel *evsel __maybe_unused)
 {
 	return NULL;
 }
-- 
2.40.0.rc0.216.gc4246ad0f0-goog

Re: [PATCH v2 05/10] perf evsel: Allow const evsel for certain accesses

[Namhyung Kim]

Hi Ian,

On Thu, Mar 2, 2023 at 1:26 PM Ian Rogers <irogers@google.com> wrote:
>
> List sorting, added later to evlist, passes const elements requiring
> helper functions to also be const. Make the argument to
> evsel__find_pmu, evsel__is_aux_event and evsel__leader const.
>
> Signed-off-by: Ian Rogers <irogers@google.com>
> ---
>  tools/perf/util/evsel.c  | 2 +-
>  tools/perf/util/evsel.h  | 6 +++---
>  tools/perf/util/pmu.c    | 6 +++---
>  tools/perf/util/python.c | 2 +-
>  4 files changed, 8 insertions(+), 8 deletions(-)
>
> diff --git a/tools/perf/util/evsel.c b/tools/perf/util/evsel.c
> index 51e8ce6edddc..2dc2c24252bb 100644
> --- a/tools/perf/util/evsel.c
> +++ b/tools/perf/util/evsel.c
> @@ -3139,7 +3139,7 @@ bool evsel__is_hybrid(const struct evsel *evsel)
>         return evsel->pmu_name && perf_pmu__is_hybrid(evsel->pmu_name);
>  }
>
> -struct evsel *evsel__leader(struct evsel *evsel)
> +struct evsel *evsel__leader(const struct evsel *evsel)

It seems like it should return const struct evsel *, right?

>  {
>         return container_of(evsel->core.leader, struct evsel, core);
>  }
> diff --git a/tools/perf/util/evsel.h b/tools/perf/util/evsel.h
> index 814a49ebb7e3..676c499323e9 100644
> --- a/tools/perf/util/evsel.h
> +++ b/tools/perf/util/evsel.h
> @@ -212,8 +212,8 @@ int evsel__object_config(size_t object_size,
>                          int (*init)(struct evsel *evsel),
>                          void (*fini)(struct evsel *evsel));
>
> -struct perf_pmu *evsel__find_pmu(struct evsel *evsel);
> -bool evsel__is_aux_event(struct evsel *evsel);
> +struct perf_pmu *evsel__find_pmu(const struct evsel *evsel);
> +bool evsel__is_aux_event(const struct evsel *evsel);
>
>  struct evsel *evsel__new_idx(struct perf_event_attr *attr, int idx);
>
> @@ -505,7 +505,7 @@ int evsel__store_ids(struct evsel *evsel, struct evlist *evlist);
>
>  void evsel__zero_per_pkg(struct evsel *evsel);
>  bool evsel__is_hybrid(const struct evsel *evsel);
> -struct evsel *evsel__leader(struct evsel *evsel);
> +struct evsel *evsel__leader(const struct evsel *evsel);
>  bool evsel__has_leader(struct evsel *evsel, struct evsel *leader);
>  bool evsel__is_leader(struct evsel *evsel);
>  void evsel__set_leader(struct evsel *evsel, struct evsel *leader);
> diff --git a/tools/perf/util/pmu.c b/tools/perf/util/pmu.c
> index 43b6182d96b7..45d9b8e28e16 100644
> --- a/tools/perf/util/pmu.c
> +++ b/tools/perf/util/pmu.c
> @@ -988,7 +988,7 @@ struct perf_pmu *perf_pmu__scan(struct perf_pmu *pmu)
>         return NULL;
>  }
>
> -struct perf_pmu *evsel__find_pmu(struct evsel *evsel)
> +struct perf_pmu *evsel__find_pmu(const struct evsel *evsel)
>  {
>         struct perf_pmu *pmu = NULL;
>
> @@ -1000,11 +1000,11 @@ struct perf_pmu *evsel__find_pmu(struct evsel *evsel)
>                         break;
>         }
>
> -       evsel->pmu = pmu;
> +       ((struct evsel *)evsel)->pmu = pmu;

Maybe we can move it out of the function and make callers set it.

Thanks,
Namhyung


>         return pmu;
>  }
>
> -bool evsel__is_aux_event(struct evsel *evsel)
> +bool evsel__is_aux_event(const struct evsel *evsel)
>  {
>         struct perf_pmu *pmu = evsel__find_pmu(evsel);
>
> diff --git a/tools/perf/util/python.c b/tools/perf/util/python.c
> index 42e8b813d010..ab48ffbb6448 100644
> --- a/tools/perf/util/python.c
> +++ b/tools/perf/util/python.c
> @@ -83,7 +83,7 @@ const char *perf_env__arch(struct perf_env *env __maybe_unused)
>   * far, for the perf python binding known usecases, revisit if this become
>   * necessary.
>   */
> -struct perf_pmu *evsel__find_pmu(struct evsel *evsel __maybe_unused)
> +struct perf_pmu *evsel__find_pmu(const struct evsel *evsel __maybe_unused)
>  {
>         return NULL;
>  }
> --
> 2.40.0.rc0.216.gc4246ad0f0-goog
>

Re: [PATCH v2 05/10] perf evsel: Allow const evsel for certain accesses

[Ian Rogers]

On Thu, Mar 2, 2023 at 4:14 PM Namhyung Kim <namhyung@kernel.org> wrote:
>
> Hi Ian,
>
> On Thu, Mar 2, 2023 at 1:26 PM Ian Rogers <irogers@google.com> wrote:
> >
> > List sorting, added later to evlist, passes const elements requiring
> > helper functions to also be const. Make the argument to
> > evsel__find_pmu, evsel__is_aux_event and evsel__leader const.
> >
> > Signed-off-by: Ian Rogers <irogers@google.com>
> > ---
> >  tools/perf/util/evsel.c  | 2 +-
> >  tools/perf/util/evsel.h  | 6 +++---
> >  tools/perf/util/pmu.c    | 6 +++---
> >  tools/perf/util/python.c | 2 +-
> >  4 files changed, 8 insertions(+), 8 deletions(-)
> >
> > diff --git a/tools/perf/util/evsel.c b/tools/perf/util/evsel.c
> > index 51e8ce6edddc..2dc2c24252bb 100644
> > --- a/tools/perf/util/evsel.c
> > +++ b/tools/perf/util/evsel.c
> > @@ -3139,7 +3139,7 @@ bool evsel__is_hybrid(const struct evsel *evsel)
> >         return evsel->pmu_name && perf_pmu__is_hybrid(evsel->pmu_name);
> >  }
> >
> > -struct evsel *evsel__leader(struct evsel *evsel)
> > +struct evsel *evsel__leader(const struct evsel *evsel)
>
> It seems like it should return const struct evsel *, right?

Yes, but that breaks things :-)

> >  {
> >         return container_of(evsel->core.leader, struct evsel, core);
> >  }
> > diff --git a/tools/perf/util/evsel.h b/tools/perf/util/evsel.h
> > index 814a49ebb7e3..676c499323e9 100644
> > --- a/tools/perf/util/evsel.h
> > +++ b/tools/perf/util/evsel.h
> > @@ -212,8 +212,8 @@ int evsel__object_config(size_t object_size,
> >                          int (*init)(struct evsel *evsel),
> >                          void (*fini)(struct evsel *evsel));
> >
> > -struct perf_pmu *evsel__find_pmu(struct evsel *evsel);
> > -bool evsel__is_aux_event(struct evsel *evsel);
> > +struct perf_pmu *evsel__find_pmu(const struct evsel *evsel);
> > +bool evsel__is_aux_event(const struct evsel *evsel);
> >
> >  struct evsel *evsel__new_idx(struct perf_event_attr *attr, int idx);
> >
> > @@ -505,7 +505,7 @@ int evsel__store_ids(struct evsel *evsel, struct evlist *evlist);
> >
> >  void evsel__zero_per_pkg(struct evsel *evsel);
> >  bool evsel__is_hybrid(const struct evsel *evsel);
> > -struct evsel *evsel__leader(struct evsel *evsel);
> > +struct evsel *evsel__leader(const struct evsel *evsel);
> >  bool evsel__has_leader(struct evsel *evsel, struct evsel *leader);
> >  bool evsel__is_leader(struct evsel *evsel);
> >  void evsel__set_leader(struct evsel *evsel, struct evsel *leader);
> > diff --git a/tools/perf/util/pmu.c b/tools/perf/util/pmu.c
> > index 43b6182d96b7..45d9b8e28e16 100644
> > --- a/tools/perf/util/pmu.c
> > +++ b/tools/perf/util/pmu.c
> > @@ -988,7 +988,7 @@ struct perf_pmu *perf_pmu__scan(struct perf_pmu *pmu)
> >         return NULL;
> >  }
> >
> > -struct perf_pmu *evsel__find_pmu(struct evsel *evsel)
> > +struct perf_pmu *evsel__find_pmu(const struct evsel *evsel)
> >  {
> >         struct perf_pmu *pmu = NULL;
> >
> > @@ -1000,11 +1000,11 @@ struct perf_pmu *evsel__find_pmu(struct evsel *evsel)
> >                         break;
> >         }
> >
> > -       evsel->pmu = pmu;
> > +       ((struct evsel *)evsel)->pmu = pmu;
>
> Maybe we can move it out of the function and make callers set it.

Right, I considered this a case where in C++ we'd use mutable so
casting away the const wasn't the worst offense.

Thanks,
Ian

> Thanks,
> Namhyung
>
>
> >         return pmu;
> >  }
> >
> > -bool evsel__is_aux_event(struct evsel *evsel)
> > +bool evsel__is_aux_event(const struct evsel *evsel)
> >  {
> >         struct perf_pmu *pmu = evsel__find_pmu(evsel);
> >
> > diff --git a/tools/perf/util/python.c b/tools/perf/util/python.c
> > index 42e8b813d010..ab48ffbb6448 100644
> > --- a/tools/perf/util/python.c
> > +++ b/tools/perf/util/python.c
> > @@ -83,7 +83,7 @@ const char *perf_env__arch(struct perf_env *env __maybe_unused)
> >   * far, for the perf python binding known usecases, revisit if this become
> >   * necessary.
> >   */
> > -struct perf_pmu *evsel__find_pmu(struct evsel *evsel __maybe_unused)
> > +struct perf_pmu *evsel__find_pmu(const struct evsel *evsel __maybe_unused)
> >  {
> >         return NULL;
> >  }
> > --
> > 2.40.0.rc0.216.gc4246ad0f0-goog
> >

[PATCH v2 06/10] perf evsel: Add function to compute pmu_name

[Ian Rogers]

The computed pmu_name respects software events and aux event groups,
such that the pmu_name is changed to be that of the aux event leader
or group leader for software events. This is done as a later change
will split events that are in different PMUs into different groups.

Signed-off-by: Ian Rogers <irogers@google.com>
---
 tools/perf/util/evsel.c | 24 ++++++++++++++++++++++++
 tools/perf/util/evsel.h |  1 +
 2 files changed, 25 insertions(+)

diff --git a/tools/perf/util/evsel.c b/tools/perf/util/evsel.c
index 2dc2c24252bb..9c6b486f8bd4 100644
--- a/tools/perf/util/evsel.c
+++ b/tools/perf/util/evsel.c
@@ -821,6 +821,30 @@ const char *evsel__name(struct evsel *evsel)
 	return "unknown";
 }
 
+const char *evsel__pmu_name(const struct evsel *evsel)
+{
+	const struct evsel *leader;
+
+	/* If the pmu_name is set use it. pmu_name isn't set for CPU and software events. */
+	if (evsel->pmu_name)
+		return evsel->pmu_name;
+	/*
+	 * Software events may be in a group with other uncore PMU events. Use
+	 * the pmu_name of the group leader to avoid breaking the software event
+	 * out of the group.
+	 *
+	 * Aux event leaders, like intel_pt, expect a group with events from
+	 * other PMUs, so substitute the AUX event's PMU in this case.
+	 */
+	leader  = evsel__leader(evsel);
+	if ((evsel->core.attr.type == PERF_TYPE_SOFTWARE || evsel__is_aux_event(leader)) &&
+	    leader->pmu_name) {
+		return leader->pmu_name;
+	}
+
+	return "cpu";
+}
+
 const char *evsel__metric_id(const struct evsel *evsel)
 {
 	if (evsel->metric_id)
diff --git a/tools/perf/util/evsel.h b/tools/perf/util/evsel.h
index 676c499323e9..72121194d3b1 100644
--- a/tools/perf/util/evsel.h
+++ b/tools/perf/util/evsel.h
@@ -280,6 +280,7 @@ int arch_evsel__hw_name(struct evsel *evsel, char *bf, size_t size);
 
 int __evsel__hw_cache_type_op_res_name(u8 type, u8 op, u8 result, char *bf, size_t size);
 const char *evsel__name(struct evsel *evsel);
+const char *evsel__pmu_name(const struct evsel *evsel);
 const char *evsel__metric_id(const struct evsel *evsel);
 
 static inline bool evsel__is_tool(const struct evsel *evsel)
-- 
2.40.0.rc0.216.gc4246ad0f0-goog

Re: [PATCH v2 06/10] perf evsel: Add function to compute pmu_name

[Namhyung Kim]

On Thu, Mar 2, 2023 at 1:26 PM Ian Rogers <irogers@google.com> wrote:
>
> The computed pmu_name respects software events and aux event groups,
> such that the pmu_name is changed to be that of the aux event leader
> or group leader for software events. This is done as a later change
> will split events that are in different PMUs into different groups.
>
> Signed-off-by: Ian Rogers <irogers@google.com>
> ---
>  tools/perf/util/evsel.c | 24 ++++++++++++++++++++++++
>  tools/perf/util/evsel.h |  1 +
>  2 files changed, 25 insertions(+)
>
> diff --git a/tools/perf/util/evsel.c b/tools/perf/util/evsel.c
> index 2dc2c24252bb..9c6b486f8bd4 100644
> --- a/tools/perf/util/evsel.c
> +++ b/tools/perf/util/evsel.c
> @@ -821,6 +821,30 @@ const char *evsel__name(struct evsel *evsel)
>         return "unknown";
>  }
>
> +const char *evsel__pmu_name(const struct evsel *evsel)
> +{
> +       const struct evsel *leader;
> +
> +       /* If the pmu_name is set use it. pmu_name isn't set for CPU and software events. */
> +       if (evsel->pmu_name)
> +               return evsel->pmu_name;

I'm wondering if we can just use evsel->pmu->name if it's set.

> +       /*
> +        * Software events may be in a group with other uncore PMU events. Use
> +        * the pmu_name of the group leader to avoid breaking the software event
> +        * out of the group.
> +        *
> +        * Aux event leaders, like intel_pt, expect a group with events from
> +        * other PMUs, so substitute the AUX event's PMU in this case.
> +        */

Looks like we need to rename this function as it doesn't simply return
the pmu name in the above cases.  Maybe evsel__group_pmu_name() ?

Thanks,
Namhyung


> +       leader  = evsel__leader(evsel);
> +       if ((evsel->core.attr.type == PERF_TYPE_SOFTWARE || evsel__is_aux_event(leader)) &&
> +           leader->pmu_name) {
> +               return leader->pmu_name;
> +       }
> +
> +       return "cpu";
> +}
> +
>  const char *evsel__metric_id(const struct evsel *evsel)
>  {
>         if (evsel->metric_id)
> diff --git a/tools/perf/util/evsel.h b/tools/perf/util/evsel.h
> index 676c499323e9..72121194d3b1 100644
> --- a/tools/perf/util/evsel.h
> +++ b/tools/perf/util/evsel.h
> @@ -280,6 +280,7 @@ int arch_evsel__hw_name(struct evsel *evsel, char *bf, size_t size);
>
>  int __evsel__hw_cache_type_op_res_name(u8 type, u8 op, u8 result, char *bf, size_t size);
>  const char *evsel__name(struct evsel *evsel);
> +const char *evsel__pmu_name(const struct evsel *evsel);
>  const char *evsel__metric_id(const struct evsel *evsel);
>
>  static inline bool evsel__is_tool(const struct evsel *evsel)
> --
> 2.40.0.rc0.216.gc4246ad0f0-goog
>

Re: [PATCH v2 06/10] perf evsel: Add function to compute pmu_name

[Ian Rogers]

On Thu, Mar 2, 2023 at 4:19 PM Namhyung Kim <namhyung@kernel.org> wrote:
>
> On Thu, Mar 2, 2023 at 1:26 PM Ian Rogers <irogers@google.com> wrote:
> >
> > The computed pmu_name respects software events and aux event groups,
> > such that the pmu_name is changed to be that of the aux event leader
> > or group leader for software events. This is done as a later change
> > will split events that are in different PMUs into different groups.
> >
> > Signed-off-by: Ian Rogers <irogers@google.com>
> > ---
> >  tools/perf/util/evsel.c | 24 ++++++++++++++++++++++++
> >  tools/perf/util/evsel.h |  1 +
> >  2 files changed, 25 insertions(+)
> >
> > diff --git a/tools/perf/util/evsel.c b/tools/perf/util/evsel.c
> > index 2dc2c24252bb..9c6b486f8bd4 100644
> > --- a/tools/perf/util/evsel.c
> > +++ b/tools/perf/util/evsel.c
> > @@ -821,6 +821,30 @@ const char *evsel__name(struct evsel *evsel)
> >         return "unknown";
> >  }
> >
> > +const char *evsel__pmu_name(const struct evsel *evsel)
> > +{
> > +       const struct evsel *leader;
> > +
> > +       /* If the pmu_name is set use it. pmu_name isn't set for CPU and software events. */
> > +       if (evsel->pmu_name)
> > +               return evsel->pmu_name;
>
> I'm wondering if we can just use evsel->pmu->name if it's set.

Agreed. I think that can be looked into in follow up work.

> > +       /*
> > +        * Software events may be in a group with other uncore PMU events. Use
> > +        * the pmu_name of the group leader to avoid breaking the software event
> > +        * out of the group.
> > +        *
> > +        * Aux event leaders, like intel_pt, expect a group with events from
> > +        * other PMUs, so substitute the AUX event's PMU in this case.
> > +        */
>
> Looks like we need to rename this function as it doesn't simply return
> the pmu name in the above cases.  Maybe evsel__group_pmu_name() ?

Agreed, that is a better name. I'll add it in v3.

Thanks,
Ian

> Thanks,
> Namhyung
>
>
> > +       leader  = evsel__leader(evsel);
> > +       if ((evsel->core.attr.type == PERF_TYPE_SOFTWARE || evsel__is_aux_event(leader)) &&
> > +           leader->pmu_name) {
> > +               return leader->pmu_name;
> > +       }
> > +
> > +       return "cpu";
> > +}
> > +
> >  const char *evsel__metric_id(const struct evsel *evsel)
> >  {
> >         if (evsel->metric_id)
> > diff --git a/tools/perf/util/evsel.h b/tools/perf/util/evsel.h
> > index 676c499323e9..72121194d3b1 100644
> > --- a/tools/perf/util/evsel.h
> > +++ b/tools/perf/util/evsel.h
> > @@ -280,6 +280,7 @@ int arch_evsel__hw_name(struct evsel *evsel, char *bf, size_t size);
> >
> >  int __evsel__hw_cache_type_op_res_name(u8 type, u8 op, u8 result, char *bf, size_t size);
> >  const char *evsel__name(struct evsel *evsel);
> > +const char *evsel__pmu_name(const struct evsel *evsel);
> >  const char *evsel__metric_id(const struct evsel *evsel);
> >
> >  static inline bool evsel__is_tool(const struct evsel *evsel)
> > --
> > 2.40.0.rc0.216.gc4246ad0f0-goog
> >

[PATCH v2 07/10] perf parse-events: Pass ownership of the group name

[Ian Rogers]

Pass ownership of the group name rather than copying and freeing the
original. This saves a memory allocation and copy.

Signed-off-by: Ian Rogers <irogers@google.com>
---
 tools/perf/util/parse-events.c | 3 ++-
 tools/perf/util/parse-events.y | 2 +-
 2 files changed, 3 insertions(+), 2 deletions(-)

diff --git a/tools/perf/util/parse-events.c b/tools/perf/util/parse-events.c
index 0336ff27c15f..1be454697d57 100644
--- a/tools/perf/util/parse-events.c
+++ b/tools/perf/util/parse-events.c
@@ -1761,6 +1761,7 @@ parse_events__set_leader_for_uncore_aliase(char *name, struct list_head *list,
 
 handled:
 	ret = 1;
+	free(name);
 out:
 	free(leaders);
 	return ret;
@@ -1786,7 +1787,7 @@ void parse_events__set_leader(char *name, struct list_head *list,
 
 	leader = arch_evlist__leader(list);
 	__perf_evlist__set_leader(list, &leader->core);
-	leader->group_name = name ? strdup(name) : NULL;
+	leader->group_name = name;
 	list_move(&leader->core.node, list);
 }
 
diff --git a/tools/perf/util/parse-events.y b/tools/perf/util/parse-events.y
index be8c51770051..541b8dde2063 100644
--- a/tools/perf/util/parse-events.y
+++ b/tools/perf/util/parse-events.y
@@ -202,8 +202,8 @@ PE_NAME '{' events '}'
 	struct list_head *list = $3;
 
 	inc_group_count(list, _parse_state);
+	/* Takes ownership of $1. */
 	parse_events__set_leader($1, list, _parse_state);
-	free($1);
 	$$ = list;
 }
 |
-- 
2.40.0.rc0.216.gc4246ad0f0-goog

[PATCH v2 08/10] perf parse-events: Sort and group parsed events

[Ian Rogers]

This change is intended to be a no-op for most current cases, the
default sort order is the order the events were parsed. Where it
varies is in how groups are handled. Previously an uncore and core
event that are grouped would most often cause the group to be removed:

```
$ perf stat -e '{instructions,uncore_imc_free_running_0/data_total/}' -a sleep 1
WARNING: grouped events cpus do not match, disabling group:
  anon group { instructions, uncore_imc_free_running_0/data_total/ }
...
```

However, when wildcards are used the events should be re-sorted and
re-grouped in parse_events__set_leader, but this currently fails for
simple examples:

```
$ perf stat -e '{uncore_imc_free_running/data_read/,uncore_imc_free_running/data_write/}' -a sleep 1

 Performance counter stats for 'system wide':

     <not counted> MiB  uncore_imc_free_running/data_read/
     <not counted> MiB  uncore_imc_free_running/data_write/

       1.000996992 seconds time elapsed
```

A futher failure mode, fixed in this patch, is to force topdown events
into a group.

This change moves sorting the evsels in the evlist after parsing. It
requires parsing to set up groups. First the evsels are sorted
respecting the existing groupings and parse order, but also reordering
to ensure evsels of the same PMU and group appear together. So that
software and aux events respect groups, their pmu_name is taken from
the group leader. The sorting is done with list_sort removing a memory
allocation.

After sorting a pass is done to correct the group leaders and for
topdown events ensuring they have a group leader.

This fixes the problems seen before:

```
$ perf stat -e '{uncore_imc_free_running/data_read/,uncore_imc_free_running/data_write/}' -a sleep 1

 Performance counter stats for 'system wide':

            727.42 MiB  uncore_imc_free_running/data_read/
             81.84 MiB  uncore_imc_free_running/data_write/

       1.000948615 seconds time elapsed
```

As well as making groups not fail for cases like:

```
$ perf stat -e '{imc_free_running_0/data_total/,imc_free_running_1/data_total/}' -a sleep 1

 Performance counter stats for 'system wide':

            256.47 MiB  imc_free_running_0/data_total/
            256.48 MiB  imc_free_running_1/data_total/

       1.001165442 seconds time elapsed
```

Signed-off-by: Ian Rogers <irogers@google.com>
---
 tools/perf/arch/x86/util/evlist.c |  39 ++---
 tools/perf/util/evlist.h          |   2 +-
 tools/perf/util/parse-events.c    | 240 +++++++++++++++---------------
 tools/perf/util/parse-events.h    |   3 +-
 tools/perf/util/parse-events.y    |   4 +-
 5 files changed, 136 insertions(+), 152 deletions(-)

diff --git a/tools/perf/arch/x86/util/evlist.c b/tools/perf/arch/x86/util/evlist.c
index 8a7ae4162563..d4193479a364 100644
--- a/tools/perf/arch/x86/util/evlist.c
+++ b/tools/perf/arch/x86/util/evlist.c
@@ -65,29 +65,22 @@ int arch_evlist__add_default_attrs(struct evlist *evlist,
 	return ___evlist__add_default_attrs(evlist, attrs, nr_attrs);
 }
 
-struct evsel *arch_evlist__leader(struct list_head *list)
+int arch_evlist__cmp(const struct evsel *lhs, const struct evsel *rhs)
 {
-	struct evsel *evsel, *first, *slots = NULL;
-	bool has_topdown = false;
-
-	first = list_first_entry(list, struct evsel, core.node);
-
-	if (!topdown_sys_has_perf_metrics())
-		return first;
-
-	/* If there is a slots event and a topdown event then the slots event comes first. */
-	__evlist__for_each_entry(list, evsel) {
-		if (evsel->pmu_name && !strncmp(evsel->pmu_name, "cpu", 3) && evsel->name) {
-			if (strcasestr(evsel->name, "slots")) {
-				slots = evsel;
-				if (slots == first)
-					return first;
-			}
-			if (strcasestr(evsel->name, "topdown"))
-				has_topdown = true;
-			if (slots && has_topdown)
-				return slots;
-		}
+	if (topdown_sys_has_perf_metrics() &&
+	    (!lhs->pmu_name || !strncmp(lhs->pmu_name, "cpu", 3))) {
+		/* Ensure the topdown slots comes first. */
+		if (strcasestr(lhs->name, "slots"))
+			return -1;
+		if (strcasestr(rhs->name, "slots"))
+			return 1;
+		/* Followed by topdown events. */
+		if (strcasestr(lhs->name, "topdown") && !strcasestr(rhs->name, "topdown"))
+			return -1;
+		if (!strcasestr(lhs->name, "topdown") && strcasestr(rhs->name, "topdown"))
+			return 1;
 	}
-	return first;
+
+	/* Default ordering by insertion index. */
+	return lhs->core.idx - rhs->core.idx;
 }
diff --git a/tools/perf/util/evlist.h b/tools/perf/util/evlist.h
index 01fa9d592c5a..d89d8f92802b 100644
--- a/tools/perf/util/evlist.h
+++ b/tools/perf/util/evlist.h
@@ -119,7 +119,7 @@ int arch_evlist__add_default_attrs(struct evlist *evlist,
 #define evlist__add_default_attrs(evlist, array) \
 	arch_evlist__add_default_attrs(evlist, array, ARRAY_SIZE(array))
 
-struct evsel *arch_evlist__leader(struct list_head *list);
+int arch_evlist__cmp(const struct evsel *lhs, const struct evsel *rhs);
 
 int evlist__add_dummy(struct evlist *evlist);
 struct evsel *evlist__add_aux_dummy(struct evlist *evlist, bool system_wide);
diff --git a/tools/perf/util/parse-events.c b/tools/perf/util/parse-events.c
index 1be454697d57..d6eb0a54dd2d 100644
--- a/tools/perf/util/parse-events.c
+++ b/tools/perf/util/parse-events.c
@@ -1,6 +1,7 @@
 // SPDX-License-Identifier: GPL-2.0
 #include <linux/hw_breakpoint.h>
 #include <linux/err.h>
+#include <linux/list_sort.h>
 #include <linux/zalloc.h>
 #include <dirent.h>
 #include <errno.h>
@@ -1655,125 +1656,7 @@ int parse_events__modifier_group(struct list_head *list,
 	return parse_events__modifier_event(list, event_mod, true);
 }
 
-/*
- * Check if the two uncore PMUs are from the same uncore block
- * The format of the uncore PMU name is uncore_#blockname_#pmuidx
- */
-static bool is_same_uncore_block(const char *pmu_name_a, const char *pmu_name_b)
-{
-	char *end_a, *end_b;
-
-	end_a = strrchr(pmu_name_a, '_');
-	end_b = strrchr(pmu_name_b, '_');
-
-	if (!end_a || !end_b)
-		return false;
-
-	if ((end_a - pmu_name_a) != (end_b - pmu_name_b))
-		return false;
-
-	return (strncmp(pmu_name_a, pmu_name_b, end_a - pmu_name_a) == 0);
-}
-
-static int
-parse_events__set_leader_for_uncore_aliase(char *name, struct list_head *list,
-					   struct parse_events_state *parse_state)
-{
-	struct evsel *evsel, *leader;
-	uintptr_t *leaders;
-	bool is_leader = true;
-	int i, nr_pmu = 0, total_members, ret = 0;
-
-	leader = list_first_entry(list, struct evsel, core.node);
-	evsel = list_last_entry(list, struct evsel, core.node);
-	total_members = evsel->core.idx - leader->core.idx + 1;
-
-	leaders = calloc(total_members, sizeof(uintptr_t));
-	if (WARN_ON(!leaders))
-		return 0;
-
-	/*
-	 * Going through the whole group and doing sanity check.
-	 * All members must use alias, and be from the same uncore block.
-	 * Also, storing the leader events in an array.
-	 */
-	__evlist__for_each_entry(list, evsel) {
-
-		/* Only split the uncore group which members use alias */
-		if (!evsel->use_uncore_alias)
-			goto out;
-
-		/* The events must be from the same uncore block */
-		if (!is_same_uncore_block(leader->pmu_name, evsel->pmu_name))
-			goto out;
-
-		if (!is_leader)
-			continue;
-		/*
-		 * If the event's PMU name starts to repeat, it must be a new
-		 * event. That can be used to distinguish the leader from
-		 * other members, even they have the same event name.
-		 */
-		if ((leader != evsel) &&
-		    !strcmp(leader->pmu_name, evsel->pmu_name)) {
-			is_leader = false;
-			continue;
-		}
-
-		/* Store the leader event for each PMU */
-		leaders[nr_pmu++] = (uintptr_t) evsel;
-	}
-
-	/* only one event alias */
-	if (nr_pmu == total_members) {
-		parse_state->nr_groups--;
-		goto handled;
-	}
-
-	/*
-	 * An uncore event alias is a joint name which means the same event
-	 * runs on all PMUs of a block.
-	 * Perf doesn't support mixed events from different PMUs in the same
-	 * group. The big group has to be split into multiple small groups
-	 * which only include the events from the same PMU.
-	 *
-	 * Here the uncore event aliases must be from the same uncore block.
-	 * The number of PMUs must be same for each alias. The number of new
-	 * small groups equals to the number of PMUs.
-	 * Setting the leader event for corresponding members in each group.
-	 */
-	i = 0;
-	__evlist__for_each_entry(list, evsel) {
-		if (i >= nr_pmu)
-			i = 0;
-		evsel__set_leader(evsel, (struct evsel *) leaders[i++]);
-	}
-
-	/* The number of members and group name are same for each group */
-	for (i = 0; i < nr_pmu; i++) {
-		evsel = (struct evsel *) leaders[i];
-		evsel->core.nr_members = total_members / nr_pmu;
-		evsel->group_name = name ? strdup(name) : NULL;
-	}
-
-	/* Take the new small groups into account */
-	parse_state->nr_groups += nr_pmu - 1;
-
-handled:
-	ret = 1;
-	free(name);
-out:
-	free(leaders);
-	return ret;
-}
-
-__weak struct evsel *arch_evlist__leader(struct list_head *list)
-{
-	return list_first_entry(list, struct evsel, core.node);
-}
-
-void parse_events__set_leader(char *name, struct list_head *list,
-			      struct parse_events_state *parse_state)
+void parse_events__set_leader(char *name, struct list_head *list)
 {
 	struct evsel *leader;
 
@@ -1782,13 +1665,9 @@ void parse_events__set_leader(char *name, struct list_head *list,
 		return;
 	}
 
-	if (parse_events__set_leader_for_uncore_aliase(name, list, parse_state))
-		return;
-
-	leader = arch_evlist__leader(list);
+	leader = list_first_entry(list, struct evsel, core.node);
 	__perf_evlist__set_leader(list, &leader->core);
 	leader->group_name = name;
-	list_move(&leader->core.node, list);
 }
 
 /* list_event is assumed to point to malloc'ed memory */
@@ -2245,6 +2124,117 @@ static int parse_events__with_hybrid_pmu(struct parse_events_state *parse_state,
 	return ret;
 }
 
+__weak int arch_evlist__cmp(const struct evsel *lhs, const struct evsel *rhs)
+{
+	/* Order by insertion index. */
+	return lhs->core.idx - rhs->core.idx;
+}
+
+static int evlist__cmp(void *state, const struct list_head *l, const struct list_head *r)
+{
+	const struct perf_evsel *lhs_core = container_of(l, struct perf_evsel, node);
+	const struct evsel *lhs = container_of(lhs_core, struct evsel, core);
+	const struct perf_evsel *rhs_core = container_of(r, struct perf_evsel, node);
+	const struct evsel *rhs = container_of(rhs_core, struct evsel, core);
+	int *leader_idx = state;
+	int lhs_leader_idx = *leader_idx, rhs_leader_idx = *leader_idx, ret;
+	const char *lhs_pmu_name, *rhs_pmu_name;
+
+	/*
+	 * First sort by grouping/leader. Read the leader idx only if the evsel
+	 * is part of a group, as -1 indicates no group.
+	 */
+	if (lhs_core->leader != lhs_core || lhs_core->nr_members > 1)
+		lhs_leader_idx = lhs_core->leader->idx;
+	if (rhs_core->leader != rhs_core || rhs_core->nr_members > 1)
+		rhs_leader_idx = rhs_core->leader->idx;
+
+	if (lhs_leader_idx != rhs_leader_idx)
+		return lhs_leader_idx - rhs_leader_idx;
+
+	/* Group by PMU. Groups can't span PMUs. */
+	lhs_pmu_name = evsel__pmu_name(lhs);
+	rhs_pmu_name = evsel__pmu_name(rhs);
+	ret = strcmp(lhs_pmu_name, rhs_pmu_name);
+	if (ret)
+		return ret;
+
+	/* Architecture specific sorting. */
+	return arch_evlist__cmp(lhs, rhs);
+}
+
+static void parse_events__sort_events_and_fix_groups(struct list_head *list)
+{
+	int idx = -1;
+	struct evsel *pos, *cur_leader = NULL;
+	struct perf_evsel *cur_leaders_grp = NULL;
+
+	/*
+	 * Compute index to insert ungrouped events at. Place them where the
+	 * first ungrouped event appears.
+	 */
+	list_for_each_entry(pos, list, core.node) {
+		const struct evsel *pos_leader = evsel__leader(pos);
+
+		if (pos != pos_leader || pos->core.nr_members > 1)
+			continue;
+
+		idx = pos->core.idx;
+		break;
+	}
+
+	/* Sort events. */
+	list_sort(&idx, list, evlist__cmp);
+
+	/*
+	 * Recompute groups, splitting for PMUs and adding groups for events
+	 * that require them.
+	 */
+	idx = 0;
+	list_for_each_entry(pos, list, core.node) {
+		const struct evsel *pos_leader = evsel__leader(pos);
+		const char *pos_pmu_name = evsel__pmu_name(pos);
+		const char *cur_leader_pmu_name, *pos_leader_pmu_name;
+		bool force_grouped = arch_evsel__must_be_in_group(pos);
+
+		/* Reset index and nr_members. */
+		pos->core.idx = idx++;
+		pos->core.nr_members = 0;
+
+		/*
+		 * Set the group leader respecting the given groupings and that
+		 * groups can't span PMUs.
+		 */
+		if (!cur_leader)
+			cur_leader = pos;
+
+		cur_leader_pmu_name = evsel__pmu_name(cur_leader);
+		if ((cur_leaders_grp != pos->core.leader && !force_grouped) ||
+		    strcmp(cur_leader_pmu_name, pos_pmu_name)) {
+			/* Event is for a different group/PMU than last. */
+			cur_leader = pos;
+			/*
+			 * Remember the leader's group before it is overwritten,
+			 * so that later events match as being in the same
+			 * group.
+			 */
+			cur_leaders_grp = pos->core.leader;
+		}
+		pos_leader_pmu_name = evsel__pmu_name(pos_leader);
+		if (strcmp(pos_leader_pmu_name, pos_pmu_name) || force_grouped) {
+			/*
+			 * Event's PMU differs from its leader's. Groups can't
+			 * span PMUs, so update leader from the group/PMU
+			 * tracker.
+			 */
+			evsel__set_leader(pos, cur_leader);
+		}
+	}
+	list_for_each_entry(pos, list, core.node) {
+		pos->core.leader->nr_members++;
+	}
+}
+
 int __parse_events(struct evlist *evlist, const char *str,
 		   struct parse_events_error *err, struct perf_pmu *fake_pmu)
 {
@@ -2266,6 +2256,8 @@ int __parse_events(struct evlist *evlist, const char *str,
 		return -1;
 	}
 
+	parse_events__sort_events_and_fix_groups(&parse_state.list);
+
 	/*
 	 * Add list to the evlist even with errors to allow callers to clean up.
 	 */
diff --git a/tools/perf/util/parse-events.h b/tools/perf/util/parse-events.h
index 428e72eaafcc..22fc11b0bd59 100644
--- a/tools/perf/util/parse-events.h
+++ b/tools/perf/util/parse-events.h
@@ -200,8 +200,7 @@ int parse_events_copy_term_list(struct list_head *old,
 
 enum perf_pmu_event_symbol_type
 perf_pmu__parse_check(const char *name);
-void parse_events__set_leader(char *name, struct list_head *list,
-			      struct parse_events_state *parse_state);
+void parse_events__set_leader(char *name, struct list_head *list);
 void parse_events_update_lists(struct list_head *list_event,
 			       struct list_head *list_all);
 void parse_events_evlist_error(struct parse_events_state *parse_state,
diff --git a/tools/perf/util/parse-events.y b/tools/perf/util/parse-events.y
index 541b8dde2063..90d12f2bc8be 100644
--- a/tools/perf/util/parse-events.y
+++ b/tools/perf/util/parse-events.y
@@ -203,7 +203,7 @@ PE_NAME '{' events '}'
 
 	inc_group_count(list, _parse_state);
 	/* Takes ownership of $1. */
-	parse_events__set_leader($1, list, _parse_state);
+	parse_events__set_leader($1, list);
 	$$ = list;
 }
 |
@@ -212,7 +212,7 @@ PE_NAME '{' events '}'
 	struct list_head *list = $2;
 
 	inc_group_count(list, _parse_state);
-	parse_events__set_leader(NULL, list, _parse_state);
+	parse_events__set_leader(NULL, list);
 	$$ = list;
 }
 
-- 
2.40.0.rc0.216.gc4246ad0f0-goog

Re: [PATCH v2 08/10] perf parse-events: Sort and group parsed events

[Namhyung Kim]

On Thu, Mar 2, 2023 at 1:26 PM Ian Rogers <irogers@google.com> wrote:
>
> This change is intended to be a no-op for most current cases, the
> default sort order is the order the events were parsed. Where it
> varies is in how groups are handled. Previously an uncore and core
> event that are grouped would most often cause the group to be removed:
>
> ```
> $ perf stat -e '{instructions,uncore_imc_free_running_0/data_total/}' -a sleep 1
> WARNING: grouped events cpus do not match, disabling group:
>   anon group { instructions, uncore_imc_free_running_0/data_total/ }
> ...
> ```
>
> However, when wildcards are used the events should be re-sorted and
> re-grouped in parse_events__set_leader, but this currently fails for
> simple examples:
>
> ```
> $ perf stat -e '{uncore_imc_free_running/data_read/,uncore_imc_free_running/data_write/}' -a sleep 1
>
>  Performance counter stats for 'system wide':
>
>      <not counted> MiB  uncore_imc_free_running/data_read/
>      <not counted> MiB  uncore_imc_free_running/data_write/
>
>        1.000996992 seconds time elapsed
> ```
>
> A futher failure mode, fixed in this patch, is to force topdown events
> into a group.
>
> This change moves sorting the evsels in the evlist after parsing. It
> requires parsing to set up groups. First the evsels are sorted
> respecting the existing groupings and parse order, but also reordering
> to ensure evsels of the same PMU and group appear together. So that
> software and aux events respect groups, their pmu_name is taken from
> the group leader. The sorting is done with list_sort removing a memory
> allocation.
>
> After sorting a pass is done to correct the group leaders and for
> topdown events ensuring they have a group leader.
>
> This fixes the problems seen before:
>
> ```
> $ perf stat -e '{uncore_imc_free_running/data_read/,uncore_imc_free_running/data_write/}' -a sleep 1
>
>  Performance counter stats for 'system wide':
>
>             727.42 MiB  uncore_imc_free_running/data_read/
>              81.84 MiB  uncore_imc_free_running/data_write/
>
>        1.000948615 seconds time elapsed
> ```
>
> As well as making groups not fail for cases like:
>
> ```
> $ perf stat -e '{imc_free_running_0/data_total/,imc_free_running_1/data_total/}' -a sleep 1
>
>  Performance counter stats for 'system wide':
>
>             256.47 MiB  imc_free_running_0/data_total/
>             256.48 MiB  imc_free_running_1/data_total/

I didn't expect we can group events from different PMUs.
Not sure if it can handle multiplexing properly..

Thanks,
Namhyung


>
>        1.001165442 seconds time elapsed
> ```
>
> Signed-off-by: Ian Rogers <irogers@google.com>
> ---
>  tools/perf/arch/x86/util/evlist.c |  39 ++---
>  tools/perf/util/evlist.h          |   2 +-
>  tools/perf/util/parse-events.c    | 240 +++++++++++++++---------------
>  tools/perf/util/parse-events.h    |   3 +-
>  tools/perf/util/parse-events.y    |   4 +-
>  5 files changed, 136 insertions(+), 152 deletions(-)
>
> diff --git a/tools/perf/arch/x86/util/evlist.c b/tools/perf/arch/x86/util/evlist.c
> index 8a7ae4162563..d4193479a364 100644
> --- a/tools/perf/arch/x86/util/evlist.c
> +++ b/tools/perf/arch/x86/util/evlist.c
> @@ -65,29 +65,22 @@ int arch_evlist__add_default_attrs(struct evlist *evlist,
>         return ___evlist__add_default_attrs(evlist, attrs, nr_attrs);
>  }
>
> -struct evsel *arch_evlist__leader(struct list_head *list)
> +int arch_evlist__cmp(const struct evsel *lhs, const struct evsel *rhs)
>  {
> -       struct evsel *evsel, *first, *slots = NULL;
> -       bool has_topdown = false;
> -
> -       first = list_first_entry(list, struct evsel, core.node);
> -
> -       if (!topdown_sys_has_perf_metrics())
> -               return first;
> -
> -       /* If there is a slots event and a topdown event then the slots event comes first. */
> -       __evlist__for_each_entry(list, evsel) {
> -               if (evsel->pmu_name && !strncmp(evsel->pmu_name, "cpu", 3) && evsel->name) {
> -                       if (strcasestr(evsel->name, "slots")) {
> -                               slots = evsel;
> -                               if (slots == first)
> -                                       return first;
> -                       }
> -                       if (strcasestr(evsel->name, "topdown"))
> -                               has_topdown = true;
> -                       if (slots && has_topdown)
> -                               return slots;
> -               }
> +       if (topdown_sys_has_perf_metrics() &&
> +           (!lhs->pmu_name || !strncmp(lhs->pmu_name, "cpu", 3))) {
> +               /* Ensure the topdown slots comes first. */
> +               if (strcasestr(lhs->name, "slots"))
> +                       return -1;
> +               if (strcasestr(rhs->name, "slots"))
> +                       return 1;
> +               /* Followed by topdown events. */
> +               if (strcasestr(lhs->name, "topdown") && !strcasestr(rhs->name, "topdown"))
> +                       return -1;
> +               if (!strcasestr(lhs->name, "topdown") && strcasestr(rhs->name, "topdown"))
> +                       return 1;
>         }
> -       return first;
> +
> +       /* Default ordering by insertion index. */
> +       return lhs->core.idx - rhs->core.idx;
>  }
> diff --git a/tools/perf/util/evlist.h b/tools/perf/util/evlist.h
> index 01fa9d592c5a..d89d8f92802b 100644
> --- a/tools/perf/util/evlist.h
> +++ b/tools/perf/util/evlist.h
> @@ -119,7 +119,7 @@ int arch_evlist__add_default_attrs(struct evlist *evlist,
>  #define evlist__add_default_attrs(evlist, array) \
>         arch_evlist__add_default_attrs(evlist, array, ARRAY_SIZE(array))
>
> -struct evsel *arch_evlist__leader(struct list_head *list);
> +int arch_evlist__cmp(const struct evsel *lhs, const struct evsel *rhs);
>
>  int evlist__add_dummy(struct evlist *evlist);
>  struct evsel *evlist__add_aux_dummy(struct evlist *evlist, bool system_wide);
> diff --git a/tools/perf/util/parse-events.c b/tools/perf/util/parse-events.c
> index 1be454697d57..d6eb0a54dd2d 100644
> --- a/tools/perf/util/parse-events.c
> +++ b/tools/perf/util/parse-events.c
> @@ -1,6 +1,7 @@
>  // SPDX-License-Identifier: GPL-2.0
>  #include <linux/hw_breakpoint.h>
>  #include <linux/err.h>
> +#include <linux/list_sort.h>
>  #include <linux/zalloc.h>
>  #include <dirent.h>
>  #include <errno.h>
> @@ -1655,125 +1656,7 @@ int parse_events__modifier_group(struct list_head *list,
>         return parse_events__modifier_event(list, event_mod, true);
>  }
>
> -/*
> - * Check if the two uncore PMUs are from the same uncore block
> - * The format of the uncore PMU name is uncore_#blockname_#pmuidx
> - */
> -static bool is_same_uncore_block(const char *pmu_name_a, const char *pmu_name_b)
> -{
> -       char *end_a, *end_b;
> -
> -       end_a = strrchr(pmu_name_a, '_');
> -       end_b = strrchr(pmu_name_b, '_');
> -
> -       if (!end_a || !end_b)
> -               return false;
> -
> -       if ((end_a - pmu_name_a) != (end_b - pmu_name_b))
> -               return false;
> -
> -       return (strncmp(pmu_name_a, pmu_name_b, end_a - pmu_name_a) == 0);
> -}
> -
> -static int
> -parse_events__set_leader_for_uncore_aliase(char *name, struct list_head *list,
> -                                          struct parse_events_state *parse_state)
> -{
> -       struct evsel *evsel, *leader;
> -       uintptr_t *leaders;
> -       bool is_leader = true;
> -       int i, nr_pmu = 0, total_members, ret = 0;
> -
> -       leader = list_first_entry(list, struct evsel, core.node);
> -       evsel = list_last_entry(list, struct evsel, core.node);
> -       total_members = evsel->core.idx - leader->core.idx + 1;
> -
> -       leaders = calloc(total_members, sizeof(uintptr_t));
> -       if (WARN_ON(!leaders))
> -               return 0;
> -
> -       /*
> -        * Going through the whole group and doing sanity check.
> -        * All members must use alias, and be from the same uncore block.
> -        * Also, storing the leader events in an array.
> -        */
> -       __evlist__for_each_entry(list, evsel) {
> -
> -               /* Only split the uncore group which members use alias */
> -               if (!evsel->use_uncore_alias)
> -                       goto out;
> -
> -               /* The events must be from the same uncore block */
> -               if (!is_same_uncore_block(leader->pmu_name, evsel->pmu_name))
> -                       goto out;
> -
> -               if (!is_leader)
> -                       continue;
> -               /*
> -                * If the event's PMU name starts to repeat, it must be a new
> -                * event. That can be used to distinguish the leader from
> -                * other members, even they have the same event name.
> -                */
> -               if ((leader != evsel) &&
> -                   !strcmp(leader->pmu_name, evsel->pmu_name)) {
> -                       is_leader = false;
> -                       continue;
> -               }
> -
> -               /* Store the leader event for each PMU */
> -               leaders[nr_pmu++] = (uintptr_t) evsel;
> -       }
> -
> -       /* only one event alias */
> -       if (nr_pmu == total_members) {
> -               parse_state->nr_groups--;
> -               goto handled;
> -       }
> -
> -       /*
> -        * An uncore event alias is a joint name which means the same event
> -        * runs on all PMUs of a block.
> -        * Perf doesn't support mixed events from different PMUs in the same
> -        * group. The big group has to be split into multiple small groups
> -        * which only include the events from the same PMU.
> -        *
> -        * Here the uncore event aliases must be from the same uncore block.
> -        * The number of PMUs must be same for each alias. The number of new
> -        * small groups equals to the number of PMUs.
> -        * Setting the leader event for corresponding members in each group.
> -        */
> -       i = 0;
> -       __evlist__for_each_entry(list, evsel) {
> -               if (i >= nr_pmu)
> -                       i = 0;
> -               evsel__set_leader(evsel, (struct evsel *) leaders[i++]);
> -       }
> -
> -       /* The number of members and group name are same for each group */
> -       for (i = 0; i < nr_pmu; i++) {
> -               evsel = (struct evsel *) leaders[i];
> -               evsel->core.nr_members = total_members / nr_pmu;
> -               evsel->group_name = name ? strdup(name) : NULL;
> -       }
> -
> -       /* Take the new small groups into account */
> -       parse_state->nr_groups += nr_pmu - 1;
> -
> -handled:
> -       ret = 1;
> -       free(name);
> -out:
> -       free(leaders);
> -       return ret;
> -}
> -
> -__weak struct evsel *arch_evlist__leader(struct list_head *list)
> -{
> -       return list_first_entry(list, struct evsel, core.node);
> -}
> -
> -void parse_events__set_leader(char *name, struct list_head *list,
> -                             struct parse_events_state *parse_state)
> +void parse_events__set_leader(char *name, struct list_head *list)
>  {
>         struct evsel *leader;
>
> @@ -1782,13 +1665,9 @@ void parse_events__set_leader(char *name, struct list_head *list,
>                 return;
>         }
>
> -       if (parse_events__set_leader_for_uncore_aliase(name, list, parse_state))
> -               return;
> -
> -       leader = arch_evlist__leader(list);
> +       leader = list_first_entry(list, struct evsel, core.node);
>         __perf_evlist__set_leader(list, &leader->core);
>         leader->group_name = name;
> -       list_move(&leader->core.node, list);
>  }
>
>  /* list_event is assumed to point to malloc'ed memory */
> @@ -2245,6 +2124,117 @@ static int parse_events__with_hybrid_pmu(struct parse_events_state *parse_state,
>         return ret;
>  }
>
> +__weak int arch_evlist__cmp(const struct evsel *lhs, const struct evsel *rhs)
> +{
> +       /* Order by insertion index. */
> +       return lhs->core.idx - rhs->core.idx;
> +}
> +
> +static int evlist__cmp(void *state, const struct list_head *l, const struct list_head *r)
> +{
> +       const struct perf_evsel *lhs_core = container_of(l, struct perf_evsel, node);
> +       const struct evsel *lhs = container_of(lhs_core, struct evsel, core);
> +       const struct perf_evsel *rhs_core = container_of(r, struct perf_evsel, node);
> +       const struct evsel *rhs = container_of(rhs_core, struct evsel, core);
> +       int *leader_idx = state;
> +       int lhs_leader_idx = *leader_idx, rhs_leader_idx = *leader_idx, ret;
> +       const char *lhs_pmu_name, *rhs_pmu_name;
> +
> +       /*
> +        * First sort by grouping/leader. Read the leader idx only if the evsel
> +        * is part of a group, as -1 indicates no group.
> +        */
> +       if (lhs_core->leader != lhs_core || lhs_core->nr_members > 1)
> +               lhs_leader_idx = lhs_core->leader->idx;
> +       if (rhs_core->leader != rhs_core || rhs_core->nr_members > 1)
> +               rhs_leader_idx = rhs_core->leader->idx;
> +
> +       if (lhs_leader_idx != rhs_leader_idx)
> +               return lhs_leader_idx - rhs_leader_idx;
> +
> +       /* Group by PMU. Groups can't span PMUs. */
> +       lhs_pmu_name = evsel__pmu_name(lhs);
> +       rhs_pmu_name = evsel__pmu_name(rhs);
> +       ret = strcmp(lhs_pmu_name, rhs_pmu_name);
> +       if (ret)
> +               return ret;
> +
> +       /* Architecture specific sorting. */
> +       return arch_evlist__cmp(lhs, rhs);
> +}
> +
> +static void parse_events__sort_events_and_fix_groups(struct list_head *list)
> +{
> +       int idx = -1;
> +       struct evsel *pos, *cur_leader = NULL;
> +       struct perf_evsel *cur_leaders_grp = NULL;
> +
> +       /*
> +        * Compute index to insert ungrouped events at. Place them where the
> +        * first ungrouped event appears.
> +        */
> +       list_for_each_entry(pos, list, core.node) {
> +               const struct evsel *pos_leader = evsel__leader(pos);
> +
> +               if (pos != pos_leader || pos->core.nr_members > 1)
> +                       continue;
> +
> +               idx = pos->core.idx;
> +               break;
> +       }
> +
> +       /* Sort events. */
> +       list_sort(&idx, list, evlist__cmp);
> +
> +       /*
> +        * Recompute groups, splitting for PMUs and adding groups for events
> +        * that require them.
> +        */
> +       idx = 0;
> +       list_for_each_entry(pos, list, core.node) {
> +               const struct evsel *pos_leader = evsel__leader(pos);
> +               const char *pos_pmu_name = evsel__pmu_name(pos);
> +               const char *cur_leader_pmu_name, *pos_leader_pmu_name;
> +               bool force_grouped = arch_evsel__must_be_in_group(pos);
> +
> +               /* Reset index and nr_members. */
> +               pos->core.idx = idx++;
> +               pos->core.nr_members = 0;
> +
> +               /*
> +                * Set the group leader respecting the given groupings and that
> +                * groups can't span PMUs.
> +                */
> +               if (!cur_leader)
> +                       cur_leader = pos;
> +
> +               cur_leader_pmu_name = evsel__pmu_name(cur_leader);
> +               if ((cur_leaders_grp != pos->core.leader && !force_grouped) ||
> +                   strcmp(cur_leader_pmu_name, pos_pmu_name)) {
> +                       /* Event is for a different group/PMU than last. */
> +                       cur_leader = pos;
> +                       /*
> +                        * Remember the leader's group before it is overwritten,
> +                        * so that later events match as being in the same
> +                        * group.
> +                        */
> +                       cur_leaders_grp = pos->core.leader;
> +               }
> +               pos_leader_pmu_name = evsel__pmu_name(pos_leader);
> +               if (strcmp(pos_leader_pmu_name, pos_pmu_name) || force_grouped) {
> +                       /*
> +                        * Event's PMU differs from its leader's. Groups can't
> +                        * span PMUs, so update leader from the group/PMU
> +                        * tracker.
> +                        */
> +                       evsel__set_leader(pos, cur_leader);
> +               }
> +       }
> +       list_for_each_entry(pos, list, core.node) {
> +               pos->core.leader->nr_members++;
> +       }
> +}
> +
>  int __parse_events(struct evlist *evlist, const char *str,
>                    struct parse_events_error *err, struct perf_pmu *fake_pmu)
>  {
> @@ -2266,6 +2256,8 @@ int __parse_events(struct evlist *evlist, const char *str,
>                 return -1;
>         }
>
> +       parse_events__sort_events_and_fix_groups(&parse_state.list);
> +
>         /*
>          * Add list to the evlist even with errors to allow callers to clean up.
>          */
> diff --git a/tools/perf/util/parse-events.h b/tools/perf/util/parse-events.h
> index 428e72eaafcc..22fc11b0bd59 100644
> --- a/tools/perf/util/parse-events.h
> +++ b/tools/perf/util/parse-events.h
> @@ -200,8 +200,7 @@ int parse_events_copy_term_list(struct list_head *old,
>
>  enum perf_pmu_event_symbol_type
>  perf_pmu__parse_check(const char *name);
> -void parse_events__set_leader(char *name, struct list_head *list,
> -                             struct parse_events_state *parse_state);
> +void parse_events__set_leader(char *name, struct list_head *list);
>  void parse_events_update_lists(struct list_head *list_event,
>                                struct list_head *list_all);
>  void parse_events_evlist_error(struct parse_events_state *parse_state,
> diff --git a/tools/perf/util/parse-events.y b/tools/perf/util/parse-events.y
> index 541b8dde2063..90d12f2bc8be 100644
> --- a/tools/perf/util/parse-events.y
> +++ b/tools/perf/util/parse-events.y
> @@ -203,7 +203,7 @@ PE_NAME '{' events '}'
>
>         inc_group_count(list, _parse_state);
>         /* Takes ownership of $1. */
> -       parse_events__set_leader($1, list, _parse_state);
> +       parse_events__set_leader($1, list);
>         $$ = list;
>  }
>  |
> @@ -212,7 +212,7 @@ PE_NAME '{' events '}'
>         struct list_head *list = $2;
>
>         inc_group_count(list, _parse_state);
> -       parse_events__set_leader(NULL, list, _parse_state);
> +       parse_events__set_leader(NULL, list);
>         $$ = list;
>  }
>
> --
> 2.40.0.rc0.216.gc4246ad0f0-goog
>

Re: [PATCH v2 08/10] perf parse-events: Sort and group parsed events

[Ian Rogers]

On Thu, Mar 2, 2023 at 4:37 PM Namhyung Kim <namhyung@kernel.org> wrote:
>
> On Thu, Mar 2, 2023 at 1:26 PM Ian Rogers <irogers@google.com> wrote:
> >
> > This change is intended to be a no-op for most current cases, the
> > default sort order is the order the events were parsed. Where it
> > varies is in how groups are handled. Previously an uncore and core
> > event that are grouped would most often cause the group to be removed:
> >
> > ```
> > $ perf stat -e '{instructions,uncore_imc_free_running_0/data_total/}' -a sleep 1
> > WARNING: grouped events cpus do not match, disabling group:
> >   anon group { instructions, uncore_imc_free_running_0/data_total/ }
> > ...
> > ```
> >
> > However, when wildcards are used the events should be re-sorted and
> > re-grouped in parse_events__set_leader, but this currently fails for
> > simple examples:
> >
> > ```
> > $ perf stat -e '{uncore_imc_free_running/data_read/,uncore_imc_free_running/data_write/}' -a sleep 1
> >
> >  Performance counter stats for 'system wide':
> >
> >      <not counted> MiB  uncore_imc_free_running/data_read/
> >      <not counted> MiB  uncore_imc_free_running/data_write/
> >
> >        1.000996992 seconds time elapsed
> > ```
> >
> > A futher failure mode, fixed in this patch, is to force topdown events
> > into a group.
> >
> > This change moves sorting the evsels in the evlist after parsing. It
> > requires parsing to set up groups. First the evsels are sorted
> > respecting the existing groupings and parse order, but also reordering
> > to ensure evsels of the same PMU and group appear together. So that
> > software and aux events respect groups, their pmu_name is taken from
> > the group leader. The sorting is done with list_sort removing a memory
> > allocation.
> >
> > After sorting a pass is done to correct the group leaders and for
> > topdown events ensuring they have a group leader.
> >
> > This fixes the problems seen before:
> >
> > ```
> > $ perf stat -e '{uncore_imc_free_running/data_read/,uncore_imc_free_running/data_write/}' -a sleep 1
> >
> >  Performance counter stats for 'system wide':
> >
> >             727.42 MiB  uncore_imc_free_running/data_read/
> >              81.84 MiB  uncore_imc_free_running/data_write/
> >
> >        1.000948615 seconds time elapsed
> > ```
> >
> > As well as making groups not fail for cases like:
> >
> > ```
> > $ perf stat -e '{imc_free_running_0/data_total/,imc_free_running_1/data_total/}' -a sleep 1
> >
> >  Performance counter stats for 'system wide':
> >
> >             256.47 MiB  imc_free_running_0/data_total/
> >             256.48 MiB  imc_free_running_1/data_total/
>
> I didn't expect we can group events from different PMUs.
> Not sure if it can handle multiplexing properly..

You are right, this example is now working as the sorting and
regrouping breaks the events into two groups. The rules around
grouping are complex and Arnaldo mentioned that maybe cases like this
should be warned about. The problem then is that wildcard and metric
expansion may naturally produce these cases and we don't want the
warning. It is something of a shame that the grouping information in
the perf stat output isn't clearer.

Thanks,
Ian

> Thanks,
> Namhyung
>
>
> >
> >        1.001165442 seconds time elapsed
> > ```
> >
> > Signed-off-by: Ian Rogers <irogers@google.com>
> > ---
> >  tools/perf/arch/x86/util/evlist.c |  39 ++---
> >  tools/perf/util/evlist.h          |   2 +-
> >  tools/perf/util/parse-events.c    | 240 +++++++++++++++---------------
> >  tools/perf/util/parse-events.h    |   3 +-
> >  tools/perf/util/parse-events.y    |   4 +-
> >  5 files changed, 136 insertions(+), 152 deletions(-)
> >
> > diff --git a/tools/perf/arch/x86/util/evlist.c b/tools/perf/arch/x86/util/evlist.c
> > index 8a7ae4162563..d4193479a364 100644
> > --- a/tools/perf/arch/x86/util/evlist.c
> > +++ b/tools/perf/arch/x86/util/evlist.c
> > @@ -65,29 +65,22 @@ int arch_evlist__add_default_attrs(struct evlist *evlist,
> >         return ___evlist__add_default_attrs(evlist, attrs, nr_attrs);
> >  }
> >
> > -struct evsel *arch_evlist__leader(struct list_head *list)
> > +int arch_evlist__cmp(const struct evsel *lhs, const struct evsel *rhs)
> >  {
> > -       struct evsel *evsel, *first, *slots = NULL;
> > -       bool has_topdown = false;
> > -
> > -       first = list_first_entry(list, struct evsel, core.node);
> > -
> > -       if (!topdown_sys_has_perf_metrics())
> > -               return first;
> > -
> > -       /* If there is a slots event and a topdown event then the slots event comes first. */
> > -       __evlist__for_each_entry(list, evsel) {
> > -               if (evsel->pmu_name && !strncmp(evsel->pmu_name, "cpu", 3) && evsel->name) {
> > -                       if (strcasestr(evsel->name, "slots")) {
> > -                               slots = evsel;
> > -                               if (slots == first)
> > -                                       return first;
> > -                       }
> > -                       if (strcasestr(evsel->name, "topdown"))
> > -                               has_topdown = true;
> > -                       if (slots && has_topdown)
> > -                               return slots;
> > -               }
> > +       if (topdown_sys_has_perf_metrics() &&
> > +           (!lhs->pmu_name || !strncmp(lhs->pmu_name, "cpu", 3))) {
> > +               /* Ensure the topdown slots comes first. */
> > +               if (strcasestr(lhs->name, "slots"))
> > +                       return -1;
> > +               if (strcasestr(rhs->name, "slots"))
> > +                       return 1;
> > +               /* Followed by topdown events. */
> > +               if (strcasestr(lhs->name, "topdown") && !strcasestr(rhs->name, "topdown"))
> > +                       return -1;
> > +               if (!strcasestr(lhs->name, "topdown") && strcasestr(rhs->name, "topdown"))
> > +                       return 1;
> >         }
> > -       return first;
> > +
> > +       /* Default ordering by insertion index. */
> > +       return lhs->core.idx - rhs->core.idx;
> >  }
> > diff --git a/tools/perf/util/evlist.h b/tools/perf/util/evlist.h
> > index 01fa9d592c5a..d89d8f92802b 100644
> > --- a/tools/perf/util/evlist.h
> > +++ b/tools/perf/util/evlist.h
> > @@ -119,7 +119,7 @@ int arch_evlist__add_default_attrs(struct evlist *evlist,
> >  #define evlist__add_default_attrs(evlist, array) \
> >         arch_evlist__add_default_attrs(evlist, array, ARRAY_SIZE(array))
> >
> > -struct evsel *arch_evlist__leader(struct list_head *list);
> > +int arch_evlist__cmp(const struct evsel *lhs, const struct evsel *rhs);
> >
> >  int evlist__add_dummy(struct evlist *evlist);
> >  struct evsel *evlist__add_aux_dummy(struct evlist *evlist, bool system_wide);
> > diff --git a/tools/perf/util/parse-events.c b/tools/perf/util/parse-events.c
> > index 1be454697d57..d6eb0a54dd2d 100644
> > --- a/tools/perf/util/parse-events.c
> > +++ b/tools/perf/util/parse-events.c
> > @@ -1,6 +1,7 @@
> >  // SPDX-License-Identifier: GPL-2.0
> >  #include <linux/hw_breakpoint.h>
> >  #include <linux/err.h>
> > +#include <linux/list_sort.h>
> >  #include <linux/zalloc.h>
> >  #include <dirent.h>
> >  #include <errno.h>
> > @@ -1655,125 +1656,7 @@ int parse_events__modifier_group(struct list_head *list,
> >         return parse_events__modifier_event(list, event_mod, true);
> >  }
> >
> > -/*
> > - * Check if the two uncore PMUs are from the same uncore block
> > - * The format of the uncore PMU name is uncore_#blockname_#pmuidx
> > - */
> > -static bool is_same_uncore_block(const char *pmu_name_a, const char *pmu_name_b)
> > -{
> > -       char *end_a, *end_b;
> > -
> > -       end_a = strrchr(pmu_name_a, '_');
> > -       end_b = strrchr(pmu_name_b, '_');
> > -
> > -       if (!end_a || !end_b)
> > -               return false;
> > -
> > -       if ((end_a - pmu_name_a) != (end_b - pmu_name_b))
> > -               return false;
> > -
> > -       return (strncmp(pmu_name_a, pmu_name_b, end_a - pmu_name_a) == 0);
> > -}
> > -
> > -static int
> > -parse_events__set_leader_for_uncore_aliase(char *name, struct list_head *list,
> > -                                          struct parse_events_state *parse_state)
> > -{
> > -       struct evsel *evsel, *leader;
> > -       uintptr_t *leaders;
> > -       bool is_leader = true;
> > -       int i, nr_pmu = 0, total_members, ret = 0;
> > -
> > -       leader = list_first_entry(list, struct evsel, core.node);
> > -       evsel = list_last_entry(list, struct evsel, core.node);
> > -       total_members = evsel->core.idx - leader->core.idx + 1;
> > -
> > -       leaders = calloc(total_members, sizeof(uintptr_t));
> > -       if (WARN_ON(!leaders))
> > -               return 0;
> > -
> > -       /*
> > -        * Going through the whole group and doing sanity check.
> > -        * All members must use alias, and be from the same uncore block.
> > -        * Also, storing the leader events in an array.
> > -        */
> > -       __evlist__for_each_entry(list, evsel) {
> > -
> > -               /* Only split the uncore group which members use alias */
> > -               if (!evsel->use_uncore_alias)
> > -                       goto out;
> > -
> > -               /* The events must be from the same uncore block */
> > -               if (!is_same_uncore_block(leader->pmu_name, evsel->pmu_name))
> > -                       goto out;
> > -
> > -               if (!is_leader)
> > -                       continue;
> > -               /*
> > -                * If the event's PMU name starts to repeat, it must be a new
> > -                * event. That can be used to distinguish the leader from
> > -                * other members, even they have the same event name.
> > -                */
> > -               if ((leader != evsel) &&
> > -                   !strcmp(leader->pmu_name, evsel->pmu_name)) {
> > -                       is_leader = false;
> > -                       continue;
> > -               }
> > -
> > -               /* Store the leader event for each PMU */
> > -               leaders[nr_pmu++] = (uintptr_t) evsel;
> > -       }
> > -
> > -       /* only one event alias */
> > -       if (nr_pmu == total_members) {
> > -               parse_state->nr_groups--;
> > -               goto handled;
> > -       }
> > -
> > -       /*
> > -        * An uncore event alias is a joint name which means the same event
> > -        * runs on all PMUs of a block.
> > -        * Perf doesn't support mixed events from different PMUs in the same
> > -        * group. The big group has to be split into multiple small groups
> > -        * which only include the events from the same PMU.
> > -        *
> > -        * Here the uncore event aliases must be from the same uncore block.
> > -        * The number of PMUs must be same for each alias. The number of new
> > -        * small groups equals to the number of PMUs.
> > -        * Setting the leader event for corresponding members in each group.
> > -        */
> > -       i = 0;
> > -       __evlist__for_each_entry(list, evsel) {
> > -               if (i >= nr_pmu)
> > -                       i = 0;
> > -               evsel__set_leader(evsel, (struct evsel *) leaders[i++]);
> > -       }
> > -
> > -       /* The number of members and group name are same for each group */
> > -       for (i = 0; i < nr_pmu; i++) {
> > -               evsel = (struct evsel *) leaders[i];
> > -               evsel->core.nr_members = total_members / nr_pmu;
> > -               evsel->group_name = name ? strdup(name) : NULL;
> > -       }
> > -
> > -       /* Take the new small groups into account */
> > -       parse_state->nr_groups += nr_pmu - 1;
> > -
> > -handled:
> > -       ret = 1;
> > -       free(name);
> > -out:
> > -       free(leaders);
> > -       return ret;
> > -}
> > -
> > -__weak struct evsel *arch_evlist__leader(struct list_head *list)
> > -{
> > -       return list_first_entry(list, struct evsel, core.node);
> > -}
> > -
> > -void parse_events__set_leader(char *name, struct list_head *list,
> > -                             struct parse_events_state *parse_state)
> > +void parse_events__set_leader(char *name, struct list_head *list)
> >  {
> >         struct evsel *leader;
> >
> > @@ -1782,13 +1665,9 @@ void parse_events__set_leader(char *name, struct list_head *list,
> >                 return;
> >         }
> >
> > -       if (parse_events__set_leader_for_uncore_aliase(name, list, parse_state))
> > -               return;
> > -
> > -       leader = arch_evlist__leader(list);
> > +       leader = list_first_entry(list, struct evsel, core.node);
> >         __perf_evlist__set_leader(list, &leader->core);
> >         leader->group_name = name;
> > -       list_move(&leader->core.node, list);
> >  }
> >
> >  /* list_event is assumed to point to malloc'ed memory */
> > @@ -2245,6 +2124,117 @@ static int parse_events__with_hybrid_pmu(struct parse_events_state *parse_state,
> >         return ret;
> >  }
> >
> > +__weak int arch_evlist__cmp(const struct evsel *lhs, const struct evsel *rhs)
> > +{
> > +       /* Order by insertion index. */
> > +       return lhs->core.idx - rhs->core.idx;
> > +}
> > +
> > +static int evlist__cmp(void *state, const struct list_head *l, const struct list_head *r)
> > +{
> > +       const struct perf_evsel *lhs_core = container_of(l, struct perf_evsel, node);
> > +       const struct evsel *lhs = container_of(lhs_core, struct evsel, core);
> > +       const struct perf_evsel *rhs_core = container_of(r, struct perf_evsel, node);
> > +       const struct evsel *rhs = container_of(rhs_core, struct evsel, core);
> > +       int *leader_idx = state;
> > +       int lhs_leader_idx = *leader_idx, rhs_leader_idx = *leader_idx, ret;
> > +       const char *lhs_pmu_name, *rhs_pmu_name;
> > +
> > +       /*
> > +        * First sort by grouping/leader. Read the leader idx only if the evsel
> > +        * is part of a group, as -1 indicates no group.
> > +        */
> > +       if (lhs_core->leader != lhs_core || lhs_core->nr_members > 1)
> > +               lhs_leader_idx = lhs_core->leader->idx;
> > +       if (rhs_core->leader != rhs_core || rhs_core->nr_members > 1)
> > +               rhs_leader_idx = rhs_core->leader->idx;
> > +
> > +       if (lhs_leader_idx != rhs_leader_idx)
> > +               return lhs_leader_idx - rhs_leader_idx;
> > +
> > +       /* Group by PMU. Groups can't span PMUs. */
> > +       lhs_pmu_name = evsel__pmu_name(lhs);
> > +       rhs_pmu_name = evsel__pmu_name(rhs);
> > +       ret = strcmp(lhs_pmu_name, rhs_pmu_name);
> > +       if (ret)
> > +               return ret;
> > +
> > +       /* Architecture specific sorting. */
> > +       return arch_evlist__cmp(lhs, rhs);
> > +}
> > +
> > +static void parse_events__sort_events_and_fix_groups(struct list_head *list)
> > +{
> > +       int idx = -1;
> > +       struct evsel *pos, *cur_leader = NULL;
> > +       struct perf_evsel *cur_leaders_grp = NULL;
> > +
> > +       /*
> > +        * Compute index to insert ungrouped events at. Place them where the
> > +        * first ungrouped event appears.
> > +        */
> > +       list_for_each_entry(pos, list, core.node) {
> > +               const struct evsel *pos_leader = evsel__leader(pos);
> > +
> > +               if (pos != pos_leader || pos->core.nr_members > 1)
> > +                       continue;
> > +
> > +               idx = pos->core.idx;
> > +               break;
> > +       }
> > +
> > +       /* Sort events. */
> > +       list_sort(&idx, list, evlist__cmp);
> > +
> > +       /*
> > +        * Recompute groups, splitting for PMUs and adding groups for events
> > +        * that require them.
> > +        */
> > +       idx = 0;
> > +       list_for_each_entry(pos, list, core.node) {
> > +               const struct evsel *pos_leader = evsel__leader(pos);
> > +               const char *pos_pmu_name = evsel__pmu_name(pos);
> > +               const char *cur_leader_pmu_name, *pos_leader_pmu_name;
> > +               bool force_grouped = arch_evsel__must_be_in_group(pos);
> > +
> > +               /* Reset index and nr_members. */
> > +               pos->core.idx = idx++;
> > +               pos->core.nr_members = 0;
> > +
> > +               /*
> > +                * Set the group leader respecting the given groupings and that
> > +                * groups can't span PMUs.
> > +                */
> > +               if (!cur_leader)
> > +                       cur_leader = pos;
> > +
> > +               cur_leader_pmu_name = evsel__pmu_name(cur_leader);
> > +               if ((cur_leaders_grp != pos->core.leader && !force_grouped) ||
> > +                   strcmp(cur_leader_pmu_name, pos_pmu_name)) {
> > +                       /* Event is for a different group/PMU than last. */
> > +                       cur_leader = pos;
> > +                       /*
> > +                        * Remember the leader's group before it is overwritten,
> > +                        * so that later events match as being in the same
> > +                        * group.
> > +                        */
> > +                       cur_leaders_grp = pos->core.leader;
> > +               }
> > +               pos_leader_pmu_name = evsel__pmu_name(pos_leader);
> > +               if (strcmp(pos_leader_pmu_name, pos_pmu_name) || force_grouped) {
> > +                       /*
> > +                        * Event's PMU differs from its leader's. Groups can't
> > +                        * span PMUs, so update leader from the group/PMU
> > +                        * tracker.
> > +                        */
> > +                       evsel__set_leader(pos, cur_leader);
> > +               }
> > +       }
> > +       list_for_each_entry(pos, list, core.node) {
> > +               pos->core.leader->nr_members++;
> > +       }
> > +}
> > +
> >  int __parse_events(struct evlist *evlist, const char *str,
> >                    struct parse_events_error *err, struct perf_pmu *fake_pmu)
> >  {
> > @@ -2266,6 +2256,8 @@ int __parse_events(struct evlist *evlist, const char *str,
> >                 return -1;
> >         }
> >
> > +       parse_events__sort_events_and_fix_groups(&parse_state.list);
> > +
> >         /*
> >          * Add list to the evlist even with errors to allow callers to clean up.
> >          */
> > diff --git a/tools/perf/util/parse-events.h b/tools/perf/util/parse-events.h
> > index 428e72eaafcc..22fc11b0bd59 100644
> > --- a/tools/perf/util/parse-events.h
> > +++ b/tools/perf/util/parse-events.h
> > @@ -200,8 +200,7 @@ int parse_events_copy_term_list(struct list_head *old,
> >
> >  enum perf_pmu_event_symbol_type
> >  perf_pmu__parse_check(const char *name);
> > -void parse_events__set_leader(char *name, struct list_head *list,
> > -                             struct parse_events_state *parse_state);
> > +void parse_events__set_leader(char *name, struct list_head *list);
> >  void parse_events_update_lists(struct list_head *list_event,
> >                                struct list_head *list_all);
> >  void parse_events_evlist_error(struct parse_events_state *parse_state,
> > diff --git a/tools/perf/util/parse-events.y b/tools/perf/util/parse-events.y
> > index 541b8dde2063..90d12f2bc8be 100644
> > --- a/tools/perf/util/parse-events.y
> > +++ b/tools/perf/util/parse-events.y
> > @@ -203,7 +203,7 @@ PE_NAME '{' events '}'
> >
> >         inc_group_count(list, _parse_state);
> >         /* Takes ownership of $1. */
> > -       parse_events__set_leader($1, list, _parse_state);
> > +       parse_events__set_leader($1, list);
> >         $$ = list;
> >  }
> >  |
> > @@ -212,7 +212,7 @@ PE_NAME '{' events '}'
> >         struct list_head *list = $2;
> >
> >         inc_group_count(list, _parse_state);
> > -       parse_events__set_leader(NULL, list, _parse_state);
> > +       parse_events__set_leader(NULL, list);
> >         $$ = list;
> >  }
> >
> > --
> > 2.40.0.rc0.216.gc4246ad0f0-goog
> >

Re: [PATCH v2 08/10] perf parse-events: Sort and group parsed events

[Namhyung Kim]

On Thu, Mar 2, 2023 at 5:39 PM Ian Rogers <irogers@google.com> wrote:
>
> On Thu, Mar 2, 2023 at 4:37 PM Namhyung Kim <namhyung@kernel.org> wrote:
> >
> > On Thu, Mar 2, 2023 at 1:26 PM Ian Rogers <irogers@google.com> wrote:
> > >
> > > This change is intended to be a no-op for most current cases, the
> > > default sort order is the order the events were parsed. Where it
> > > varies is in how groups are handled. Previously an uncore and core
> > > event that are grouped would most often cause the group to be removed:
> > >
> > > ```
> > > $ perf stat -e '{instructions,uncore_imc_free_running_0/data_total/}' -a sleep 1
> > > WARNING: grouped events cpus do not match, disabling group:
> > >   anon group { instructions, uncore_imc_free_running_0/data_total/ }
> > > ...
> > > ```
> > >
> > > However, when wildcards are used the events should be re-sorted and
> > > re-grouped in parse_events__set_leader, but this currently fails for
> > > simple examples:
> > >
> > > ```
> > > $ perf stat -e '{uncore_imc_free_running/data_read/,uncore_imc_free_running/data_write/}' -a sleep 1
> > >
> > >  Performance counter stats for 'system wide':
> > >
> > >      <not counted> MiB  uncore_imc_free_running/data_read/
> > >      <not counted> MiB  uncore_imc_free_running/data_write/
> > >
> > >        1.000996992 seconds time elapsed
> > > ```
> > >
> > > A futher failure mode, fixed in this patch, is to force topdown events
> > > into a group.
> > >
> > > This change moves sorting the evsels in the evlist after parsing. It
> > > requires parsing to set up groups. First the evsels are sorted
> > > respecting the existing groupings and parse order, but also reordering
> > > to ensure evsels of the same PMU and group appear together. So that
> > > software and aux events respect groups, their pmu_name is taken from
> > > the group leader. The sorting is done with list_sort removing a memory
> > > allocation.
> > >
> > > After sorting a pass is done to correct the group leaders and for
> > > topdown events ensuring they have a group leader.
> > >
> > > This fixes the problems seen before:
> > >
> > > ```
> > > $ perf stat -e '{uncore_imc_free_running/data_read/,uncore_imc_free_running/data_write/}' -a sleep 1
> > >
> > >  Performance counter stats for 'system wide':
> > >
> > >             727.42 MiB  uncore_imc_free_running/data_read/
> > >              81.84 MiB  uncore_imc_free_running/data_write/
> > >
> > >        1.000948615 seconds time elapsed
> > > ```
> > >
> > > As well as making groups not fail for cases like:
> > >
> > > ```
> > > $ perf stat -e '{imc_free_running_0/data_total/,imc_free_running_1/data_total/}' -a sleep 1
> > >
> > >  Performance counter stats for 'system wide':
> > >
> > >             256.47 MiB  imc_free_running_0/data_total/
> > >             256.48 MiB  imc_free_running_1/data_total/
> >
> > I didn't expect we can group events from different PMUs.
> > Not sure if it can handle multiplexing properly..
>
> You are right, this example is now working as the sorting and
> regrouping breaks the events into two groups. The rules around
> grouping are complex and Arnaldo mentioned that maybe cases like this
> should be warned about. The problem then is that wildcard and metric
> expansion may naturally produce these cases and we don't want the
> warning. It is something of a shame that the grouping information in
> the perf stat output isn't clearer.

Oh, that means the events are not in a group in this case.
Yeah.. it can be somewhat confusing.  It seems the wildcard
is a kind of exception.  Then we can warn if there's no wildcard?

Thanks,
Namhyung

[PATCH v2 09/10] perf evsel: Remove use_uncore_alias

[Ian Rogers]

This flag used to be used when regrouping uncore events in particular
due to wildcard matches. This is now handled by sorting evlist and so
the flag is redundant.

Signed-off-by: Ian Rogers <irogers@google.com>
---
 tools/perf/util/evsel.c        |  1 -
 tools/perf/util/evsel.h        |  1 -
 tools/perf/util/parse-events.c | 12 +++---------
 tools/perf/util/parse-events.h |  3 +--
 tools/perf/util/parse-events.y | 11 +++++++----
 5 files changed, 11 insertions(+), 17 deletions(-)

diff --git a/tools/perf/util/evsel.c b/tools/perf/util/evsel.c
index 9c6b486f8bd4..5446128be03b 100644
--- a/tools/perf/util/evsel.c
+++ b/tools/perf/util/evsel.c
@@ -458,7 +458,6 @@ struct evsel *evsel__clone(struct evsel *orig)
 	evsel->per_pkg = orig->per_pkg;
 	evsel->percore = orig->percore;
 	evsel->precise_max = orig->precise_max;
-	evsel->use_uncore_alias = orig->use_uncore_alias;
 	evsel->is_libpfm_event = orig->is_libpfm_event;
 
 	evsel->exclude_GH = orig->exclude_GH;
diff --git a/tools/perf/util/evsel.h b/tools/perf/util/evsel.h
index 72121194d3b1..9a8d08fcad1c 100644
--- a/tools/perf/util/evsel.h
+++ b/tools/perf/util/evsel.h
@@ -89,7 +89,6 @@ struct evsel {
 		bool			per_pkg;
 		bool			percore;
 		bool			precise_max;
-		bool			use_uncore_alias;
 		bool			is_libpfm_event;
 		bool			auto_merge_stats;
 		bool			collect_stat;
diff --git a/tools/perf/util/parse-events.c b/tools/perf/util/parse-events.c
index d6eb0a54dd2d..ac7709c1c5b7 100644
--- a/tools/perf/util/parse-events.c
+++ b/tools/perf/util/parse-events.c
@@ -1445,15 +1445,13 @@ static int parse_events__inside_hybrid_pmu(struct parse_events_state *parse_stat
 int parse_events_add_pmu(struct parse_events_state *parse_state,
 			 struct list_head *list, char *name,
 			 struct list_head *head_config,
-			 bool auto_merge_stats,
-			 bool use_alias)
+			 bool auto_merge_stats)
 {
 	struct perf_event_attr attr;
 	struct perf_pmu_info info;
 	struct perf_pmu *pmu;
 	struct evsel *evsel;
 	struct parse_events_error *err = parse_state->error;
-	bool use_uncore_alias;
 	LIST_HEAD(config_terms);
 
 	pmu = parse_state->fake_pmu ?: perf_pmu__find(name);
@@ -1488,8 +1486,6 @@ int parse_events_add_pmu(struct parse_events_state *parse_state,
 		memset(&attr, 0, sizeof(attr));
 	}
 
-	use_uncore_alias = (pmu->is_uncore && use_alias);
-
 	if (!head_config) {
 		attr.type = pmu->type;
 		evsel = __add_event(list, &parse_state->idx, &attr,
@@ -1499,7 +1495,6 @@ int parse_events_add_pmu(struct parse_events_state *parse_state,
 				    /*cpu_list=*/NULL);
 		if (evsel) {
 			evsel->pmu_name = name ? strdup(name) : NULL;
-			evsel->use_uncore_alias = use_uncore_alias;
 			return 0;
 		} else {
 			return -ENOMEM;
@@ -1560,7 +1555,6 @@ int parse_events_add_pmu(struct parse_events_state *parse_state,
 		evsel->use_config_name = true;
 
 	evsel->pmu_name = name ? strdup(name) : NULL;
-	evsel->use_uncore_alias = use_uncore_alias;
 	evsel->percore = config_term_percore(&evsel->config_terms);
 
 	if (parse_state->fake_pmu)
@@ -1622,7 +1616,7 @@ int parse_events_multi_pmu_add(struct parse_events_state *parse_state,
 				parse_events_copy_term_list(head, &orig_head);
 				if (!parse_events_add_pmu(parse_state, list,
 							  pmu->name, orig_head,
-							  true, true)) {
+							  /*auto_merge_stats=*/true)) {
 					pr_debug("%s -> %s/%s/\n", str,
 						 pmu->name, alias->str);
 					ok++;
@@ -1634,7 +1628,7 @@ int parse_events_multi_pmu_add(struct parse_events_state *parse_state,
 
 	if (parse_state->fake_pmu) {
 		if (!parse_events_add_pmu(parse_state, list, str, head,
-					  true, true)) {
+					  /*auto_merge_stats=*/true)) {
 			pr_debug("%s -> %s/%s/\n", str, "fake_pmu", str);
 			ok++;
 		}
diff --git a/tools/perf/util/parse-events.h b/tools/perf/util/parse-events.h
index 22fc11b0bd59..fdac44dc696b 100644
--- a/tools/perf/util/parse-events.h
+++ b/tools/perf/util/parse-events.h
@@ -183,8 +183,7 @@ int parse_events_add_breakpoint(struct list_head *list, int *idx,
 int parse_events_add_pmu(struct parse_events_state *parse_state,
 			 struct list_head *list, char *name,
 			 struct list_head *head_config,
-			 bool auto_merge_stats,
-			 bool use_alias);
+			 bool auto_merge_stats);
 
 struct evsel *parse_events__add_event(int idx, struct perf_event_attr *attr,
 				      const char *name, const char *metric_id,
diff --git a/tools/perf/util/parse-events.y b/tools/perf/util/parse-events.y
index 90d12f2bc8be..f1b153c72d67 100644
--- a/tools/perf/util/parse-events.y
+++ b/tools/perf/util/parse-events.y
@@ -313,7 +313,7 @@ event_pmu_name opt_pmu_config
 	list = alloc_list();
 	if (!list)
 		CLEANUP_YYABORT;
-	if (parse_events_add_pmu(_parse_state, list, $1, $2, false, false)) {
+	if (parse_events_add_pmu(_parse_state, list, $1, $2, /*auto_merge_stats=*/false)) {
 		struct perf_pmu *pmu = NULL;
 		int ok = 0;
 
@@ -330,8 +330,10 @@ event_pmu_name opt_pmu_config
 			    !perf_pmu__match(pattern, pmu->alias_name, $1)) {
 				if (parse_events_copy_term_list(orig_terms, &terms))
 					CLEANUP_YYABORT;
-				if (!parse_events_add_pmu(_parse_state, list, pmu->name, terms, true, false))
+				if (!parse_events_add_pmu(_parse_state, list, pmu->name, terms,
+							  /*auto_merge_stats=*/true)) {
 					ok++;
+				}
 				parse_events_terms__delete(terms);
 			}
 		}
@@ -407,7 +409,8 @@ PE_PMU_EVENT_FAKE sep_dc
 	if (!list)
 		YYABORT;
 
-	err = parse_events_add_pmu(_parse_state, list, $1, NULL, false, false);
+	err = parse_events_add_pmu(_parse_state, list, $1, /*head_config=*/NULL,
+				   /*auto_merge_stats=*/false);
 	free($1);
 	if (err < 0) {
 		free(list);
@@ -425,7 +428,7 @@ PE_PMU_EVENT_FAKE opt_pmu_config
 	if (!list)
 		YYABORT;
 
-	err = parse_events_add_pmu(_parse_state, list, $1, $2, false, false);
+	err = parse_events_add_pmu(_parse_state, list, $1, $2, /*auto_merge_stats=*/false);
 	free($1);
 	parse_events_terms__delete($2);
 	if (err < 0) {
-- 
2.40.0.rc0.216.gc4246ad0f0-goog

[PATCH v2 10/10] perf evlist: Remove nr_groups

[Ian Rogers]

Maintaining the number of groups during event parsing is problematic
and since changing to sort/regroup events can only be computed by a
linear pass over the evlist. As the value is generally only used in
tests, rather than hold it in a variable compute it by passing over
the evlist when necessary.

This change highlights that libpfm's counting of groups with a single
entry disagreed with regular event parsing. The libpfm tests are
updated accordingly.

Signed-off-by: Ian Rogers <irogers@google.com>
---
 tools/lib/perf/evlist.c                  | 18 +++++++++++++++++-
 tools/lib/perf/include/internal/evlist.h |  1 -
 tools/lib/perf/include/perf/evlist.h     |  1 +
 tools/perf/builtin-record.c              |  2 +-
 tools/perf/builtin-report.c              |  2 +-
 tools/perf/tests/bpf.c                   |  1 -
 tools/perf/tests/parse-events.c          | 22 +++++++++++-----------
 tools/perf/tests/pfm.c                   | 12 ++++++------
 tools/perf/util/evlist.c                 |  2 +-
 tools/perf/util/evlist.h                 |  6 ++++++
 tools/perf/util/header.c                 |  3 +--
 tools/perf/util/parse-events.c           |  1 -
 tools/perf/util/parse-events.h           |  1 -
 tools/perf/util/parse-events.y           | 10 ----------
 tools/perf/util/pfm.c                    |  1 -
 15 files changed, 45 insertions(+), 38 deletions(-)

diff --git a/tools/lib/perf/evlist.c b/tools/lib/perf/evlist.c
index 2d6121e89ccb..81e8b5fcd8ba 100644
--- a/tools/lib/perf/evlist.c
+++ b/tools/lib/perf/evlist.c
@@ -703,7 +703,23 @@ void perf_evlist__set_leader(struct perf_evlist *evlist)
 		struct perf_evsel *first = list_entry(evlist->entries.next,
 						struct perf_evsel, node);
 
-		evlist->nr_groups = evlist->nr_entries > 1 ? 1 : 0;
 		__perf_evlist__set_leader(&evlist->entries, first);
 	}
 }
+
+int perf_evlist__nr_groups(struct perf_evlist *evlist)
+{
+	struct perf_evsel *evsel;
+	int nr_groups = 0;
+
+	perf_evlist__for_each_evsel(evlist, evsel) {
+		/*
+		 * evsels by default have a nr_members of 1, and they are their
+		 * own leader. If the nr_members is >1 then this is an
+		 * indication of a group.
+		 */
+		if (evsel->leader == evsel && evsel->nr_members > 1)
+			nr_groups++;
+	}
+	return nr_groups;
+}
diff --git a/tools/lib/perf/include/internal/evlist.h b/tools/lib/perf/include/internal/evlist.h
index 850f07070036..3339bc2f1765 100644
--- a/tools/lib/perf/include/internal/evlist.h
+++ b/tools/lib/perf/include/internal/evlist.h
@@ -17,7 +17,6 @@ struct perf_mmap_param;
 struct perf_evlist {
 	struct list_head	 entries;
 	int			 nr_entries;
-	int			 nr_groups;
 	bool			 has_user_cpus;
 	bool			 needs_map_propagation;
 	/**
diff --git a/tools/lib/perf/include/perf/evlist.h b/tools/lib/perf/include/perf/evlist.h
index 9ca399d49bb4..e894b770779e 100644
--- a/tools/lib/perf/include/perf/evlist.h
+++ b/tools/lib/perf/include/perf/evlist.h
@@ -47,4 +47,5 @@ LIBPERF_API struct perf_mmap *perf_evlist__next_mmap(struct perf_evlist *evlist,
 	     (pos) = perf_evlist__next_mmap((evlist), (pos), overwrite))
 
 LIBPERF_API void perf_evlist__set_leader(struct perf_evlist *evlist);
+LIBPERF_API int perf_evlist__nr_groups(struct perf_evlist *evlist);
 #endif /* __LIBPERF_EVLIST_H */
diff --git a/tools/perf/builtin-record.c b/tools/perf/builtin-record.c
index a0870c076dc0..234b371b7b9f 100644
--- a/tools/perf/builtin-record.c
+++ b/tools/perf/builtin-record.c
@@ -2474,7 +2474,7 @@ static int __cmd_record(struct record *rec, int argc, const char **argv)
 		rec->tool.ordered_events = false;
 	}
 
-	if (!rec->evlist->core.nr_groups)
+	if (evlist__nr_groups(rec->evlist) == 0)
 		perf_header__clear_feat(&session->header, HEADER_GROUP_DESC);
 
 	if (data->is_pipe) {
diff --git a/tools/perf/builtin-report.c b/tools/perf/builtin-report.c
index 2ee2ecca208e..6400615b5e98 100644
--- a/tools/perf/builtin-report.c
+++ b/tools/perf/builtin-report.c
@@ -1481,7 +1481,7 @@ int cmd_report(int argc, const char **argv)
 
 	setup_forced_leader(&report, session->evlist);
 
-	if (symbol_conf.group_sort_idx && !session->evlist->core.nr_groups) {
+	if (symbol_conf.group_sort_idx && evlist__nr_groups(session->evlist) == 0) {
 		parse_options_usage(NULL, options, "group-sort-idx", 0);
 		ret = -EINVAL;
 		goto error;
diff --git a/tools/perf/tests/bpf.c b/tools/perf/tests/bpf.c
index ae9223f27cfb..8beb46066034 100644
--- a/tools/perf/tests/bpf.c
+++ b/tools/perf/tests/bpf.c
@@ -153,7 +153,6 @@ static int do_test(struct bpf_object *obj, int (*func)(void),
 	}
 
 	evlist__splice_list_tail(evlist, &parse_state.list);
-	evlist->core.nr_groups = parse_state.nr_groups;
 
 	evlist__config(evlist, &opts, NULL);
 
diff --git a/tools/perf/tests/parse-events.c b/tools/perf/tests/parse-events.c
index 71a5cb343311..ffa6f0a90741 100644
--- a/tools/perf/tests/parse-events.c
+++ b/tools/perf/tests/parse-events.c
@@ -53,7 +53,7 @@ static int test__checkevent_tracepoint(struct evlist *evlist)
 	struct evsel *evsel = evlist__first(evlist);
 
 	TEST_ASSERT_VAL("wrong number of entries", 1 == evlist->core.nr_entries);
-	TEST_ASSERT_VAL("wrong number of groups", 0 == evlist->core.nr_groups);
+	TEST_ASSERT_VAL("wrong number of groups", 0 == evlist__nr_groups(evlist));
 	TEST_ASSERT_VAL("wrong type", PERF_TYPE_TRACEPOINT == evsel->core.attr.type);
 	TEST_ASSERT_VAL("wrong sample_type",
 		PERF_TP_SAMPLE_TYPE == evsel->core.attr.sample_type);
@@ -66,7 +66,7 @@ static int test__checkevent_tracepoint_multi(struct evlist *evlist)
 	struct evsel *evsel;
 
 	TEST_ASSERT_VAL("wrong number of entries", evlist->core.nr_entries > 1);
-	TEST_ASSERT_VAL("wrong number of groups", 0 == evlist->core.nr_groups);
+	TEST_ASSERT_VAL("wrong number of groups", 0 == evlist__nr_groups(evlist));
 
 	evlist__for_each_entry(evlist, evsel) {
 		TEST_ASSERT_VAL("wrong type",
@@ -677,7 +677,7 @@ static int test__group1(struct evlist *evlist)
 	struct evsel *evsel, *leader;
 
 	TEST_ASSERT_VAL("wrong number of entries", 2 == evlist->core.nr_entries);
-	TEST_ASSERT_VAL("wrong number of groups", 1 == evlist->core.nr_groups);
+	TEST_ASSERT_VAL("wrong number of groups", 1 == evlist__nr_groups(evlist));
 
 	/* instructions:k */
 	evsel = leader = evlist__first(evlist);
@@ -719,7 +719,7 @@ static int test__group2(struct evlist *evlist)
 	struct evsel *evsel, *leader;
 
 	TEST_ASSERT_VAL("wrong number of entries", 3 == evlist->core.nr_entries);
-	TEST_ASSERT_VAL("wrong number of groups", 1 == evlist->core.nr_groups);
+	TEST_ASSERT_VAL("wrong number of groups", 1 == evlist__nr_groups(evlist));
 
 	/* faults + :ku modifier */
 	evsel = leader = evlist__first(evlist);
@@ -775,7 +775,7 @@ static int test__group3(struct evlist *evlist __maybe_unused)
 	struct evsel *evsel, *leader;
 
 	TEST_ASSERT_VAL("wrong number of entries", 5 == evlist->core.nr_entries);
-	TEST_ASSERT_VAL("wrong number of groups", 2 == evlist->core.nr_groups);
+	TEST_ASSERT_VAL("wrong number of groups", 2 == evlist__nr_groups(evlist));
 
 	/* group1 syscalls:sys_enter_openat:H */
 	evsel = leader = evlist__first(evlist);
@@ -868,7 +868,7 @@ static int test__group4(struct evlist *evlist __maybe_unused)
 	struct evsel *evsel, *leader;
 
 	TEST_ASSERT_VAL("wrong number of entries", 2 == evlist->core.nr_entries);
-	TEST_ASSERT_VAL("wrong number of groups", 1 == evlist->core.nr_groups);
+	TEST_ASSERT_VAL("wrong number of groups", 1 == evlist__nr_groups(evlist));
 
 	/* cycles:u + p */
 	evsel = leader = evlist__first(evlist);
@@ -912,7 +912,7 @@ static int test__group5(struct evlist *evlist __maybe_unused)
 	struct evsel *evsel, *leader;
 
 	TEST_ASSERT_VAL("wrong number of entries", 5 == evlist->core.nr_entries);
-	TEST_ASSERT_VAL("wrong number of groups", 2 == evlist->core.nr_groups);
+	TEST_ASSERT_VAL("wrong number of groups", 2 == evlist__nr_groups(evlist));
 
 	/* cycles + G */
 	evsel = leader = evlist__first(evlist);
@@ -998,7 +998,7 @@ static int test__group_gh1(struct evlist *evlist)
 	struct evsel *evsel, *leader;
 
 	TEST_ASSERT_VAL("wrong number of entries", 2 == evlist->core.nr_entries);
-	TEST_ASSERT_VAL("wrong number of groups", 1 == evlist->core.nr_groups);
+	TEST_ASSERT_VAL("wrong number of groups", 1 == evlist__nr_groups(evlist));
 
 	/* cycles + :H group modifier */
 	evsel = leader = evlist__first(evlist);
@@ -1038,7 +1038,7 @@ static int test__group_gh2(struct evlist *evlist)
 	struct evsel *evsel, *leader;
 
 	TEST_ASSERT_VAL("wrong number of entries", 2 == evlist->core.nr_entries);
-	TEST_ASSERT_VAL("wrong number of groups", 1 == evlist->core.nr_groups);
+	TEST_ASSERT_VAL("wrong number of groups", 1 == evlist__nr_groups(evlist));
 
 	/* cycles + :G group modifier */
 	evsel = leader = evlist__first(evlist);
@@ -1078,7 +1078,7 @@ static int test__group_gh3(struct evlist *evlist)
 	struct evsel *evsel, *leader;
 
 	TEST_ASSERT_VAL("wrong number of entries", 2 == evlist->core.nr_entries);
-	TEST_ASSERT_VAL("wrong number of groups", 1 == evlist->core.nr_groups);
+	TEST_ASSERT_VAL("wrong number of groups", 1 == evlist__nr_groups(evlist));
 
 	/* cycles:G + :u group modifier */
 	evsel = leader = evlist__first(evlist);
@@ -1118,7 +1118,7 @@ static int test__group_gh4(struct evlist *evlist)
 	struct evsel *evsel, *leader;
 
 	TEST_ASSERT_VAL("wrong number of entries", 2 == evlist->core.nr_entries);
-	TEST_ASSERT_VAL("wrong number of groups", 1 == evlist->core.nr_groups);
+	TEST_ASSERT_VAL("wrong number of groups", 1 == evlist__nr_groups(evlist));
 
 	/* cycles:G + :uG group modifier */
 	evsel = leader = evlist__first(evlist);
diff --git a/tools/perf/tests/pfm.c b/tools/perf/tests/pfm.c
index 71b76deb1f92..2e38dfa34b6c 100644
--- a/tools/perf/tests/pfm.c
+++ b/tools/perf/tests/pfm.c
@@ -76,7 +76,7 @@ static int test__pfm_events(struct test_suite *test __maybe_unused,
 				count_pfm_events(&evlist->core),
 				table[i].nr_events);
 		TEST_ASSERT_EQUAL(table[i].events,
-				evlist->core.nr_groups,
+				evlist__nr_groups(evlist),
 				0);
 
 		evlist__delete(evlist);
@@ -103,22 +103,22 @@ static int test__pfm_group(struct test_suite *test __maybe_unused,
 		{
 			.events = "{instructions}",
 			.nr_events = 1,
-			.nr_groups = 1,
+			.nr_groups = 0,
 		},
 		{
 			.events = "{instructions},{}",
 			.nr_events = 1,
-			.nr_groups = 1,
+			.nr_groups = 0,
 		},
 		{
 			.events = "{},{instructions}",
 			.nr_events = 1,
-			.nr_groups = 1,
+			.nr_groups = 0,
 		},
 		{
 			.events = "{instructions},{instructions}",
 			.nr_events = 2,
-			.nr_groups = 2,
+			.nr_groups = 0,
 		},
 		{
 			.events = "{instructions,cycles},{instructions,cycles}",
@@ -161,7 +161,7 @@ static int test__pfm_group(struct test_suite *test __maybe_unused,
 				count_pfm_events(&evlist->core),
 				table[i].nr_events);
 		TEST_ASSERT_EQUAL(table[i].events,
-				evlist->core.nr_groups,
+				evlist__nr_groups(evlist),
 				table[i].nr_groups);
 
 		evlist__delete(evlist);
diff --git a/tools/perf/util/evlist.c b/tools/perf/util/evlist.c
index 817df2504a1e..a5f406c468f8 100644
--- a/tools/perf/util/evlist.c
+++ b/tools/perf/util/evlist.c
@@ -1777,7 +1777,7 @@ bool evlist__exclude_kernel(struct evlist *evlist)
  */
 void evlist__force_leader(struct evlist *evlist)
 {
-	if (!evlist->core.nr_groups) {
+	if (evlist__nr_groups(evlist) == 0) {
 		struct evsel *leader = evlist__first(evlist);
 
 		evlist__set_leader(evlist);
diff --git a/tools/perf/util/evlist.h b/tools/perf/util/evlist.h
index d89d8f92802b..46cf402add93 100644
--- a/tools/perf/util/evlist.h
+++ b/tools/perf/util/evlist.h
@@ -9,6 +9,7 @@
 #include <api/fd/array.h>
 #include <internal/evlist.h>
 #include <internal/evsel.h>
+#include <perf/evlist.h>
 #include "events_stats.h"
 #include "evsel.h"
 #include <pthread.h>
@@ -255,6 +256,11 @@ static inline struct evsel *evlist__last(struct evlist *evlist)
 	return container_of(evsel, struct evsel, core);
 }
 
+static inline int evlist__nr_groups(struct evlist *evlist)
+{
+	return perf_evlist__nr_groups(&evlist->core);
+}
+
 int evlist__strerror_open(struct evlist *evlist, int err, char *buf, size_t size);
 int evlist__strerror_mmap(struct evlist *evlist, int err, char *buf, size_t size);
 
diff --git a/tools/perf/util/header.c b/tools/perf/util/header.c
index 404d816ca124..276870221ce0 100644
--- a/tools/perf/util/header.c
+++ b/tools/perf/util/header.c
@@ -786,7 +786,7 @@ static int write_pmu_mappings(struct feat_fd *ff,
 static int write_group_desc(struct feat_fd *ff,
 			    struct evlist *evlist)
 {
-	u32 nr_groups = evlist->core.nr_groups;
+	u32 nr_groups = evlist__nr_groups(evlist);
 	struct evsel *evsel;
 	int ret;
 
@@ -2807,7 +2807,6 @@ static int process_group_desc(struct feat_fd *ff, void *data __maybe_unused)
 	 * Rebuild group relationship based on the group_desc
 	 */
 	session = container_of(ff->ph, struct perf_session, header);
-	session->evlist->core.nr_groups = nr_groups;
 
 	i = nr = 0;
 	evlist__for_each_entry(session->evlist, evsel) {
diff --git a/tools/perf/util/parse-events.c b/tools/perf/util/parse-events.c
index ac7709c1c5b7..c63ec035ecfa 100644
--- a/tools/perf/util/parse-events.c
+++ b/tools/perf/util/parse-events.c
@@ -2260,7 +2260,6 @@ int __parse_events(struct evlist *evlist, const char *str,
 	if (!ret) {
 		struct evsel *last;
 
-		evlist->core.nr_groups += parse_state.nr_groups;
 		last = evlist__last(evlist);
 		last->cmdline_group_boundary = true;
 
diff --git a/tools/perf/util/parse-events.h b/tools/perf/util/parse-events.h
index fdac44dc696b..767ad1729228 100644
--- a/tools/perf/util/parse-events.h
+++ b/tools/perf/util/parse-events.h
@@ -122,7 +122,6 @@ struct parse_events_error {
 struct parse_events_state {
 	struct list_head	   list;
 	int			   idx;
-	int			   nr_groups;
 	struct parse_events_error *error;
 	struct evlist		  *evlist;
 	struct list_head	  *terms;
diff --git a/tools/perf/util/parse-events.y b/tools/perf/util/parse-events.y
index f1b153c72d67..3a04602d2982 100644
--- a/tools/perf/util/parse-events.y
+++ b/tools/perf/util/parse-events.y
@@ -49,14 +49,6 @@ static void free_list_evsel(struct list_head* list_evsel)
 	free(list_evsel);
 }
 
-static void inc_group_count(struct list_head *list,
-		       struct parse_events_state *parse_state)
-{
-	/* Count groups only have more than 1 members */
-	if (!list_is_last(list->next, list))
-		parse_state->nr_groups++;
-}
-
 %}
 
 %token PE_START_EVENTS PE_START_TERMS
@@ -201,7 +193,6 @@ PE_NAME '{' events '}'
 {
 	struct list_head *list = $3;
 
-	inc_group_count(list, _parse_state);
 	/* Takes ownership of $1. */
 	parse_events__set_leader($1, list);
 	$$ = list;
@@ -211,7 +202,6 @@ PE_NAME '{' events '}'
 {
 	struct list_head *list = $2;
 
-	inc_group_count(list, _parse_state);
 	parse_events__set_leader(NULL, list);
 	$$ = list;
 }
diff --git a/tools/perf/util/pfm.c b/tools/perf/util/pfm.c
index b59ba825ddc9..6c11914c179f 100644
--- a/tools/perf/util/pfm.c
+++ b/tools/perf/util/pfm.c
@@ -112,7 +112,6 @@ int parse_libpfm_events_option(const struct option *opt, const char *str,
 				   "cannot close a non-existing event group\n");
 				goto error;
 			}
-			evlist->core.nr_groups++;
 			grp_leader = NULL;
 			grp_evt = -1;
 		}
-- 
2.40.0.rc0.216.gc4246ad0f0-goog