Skip to contentSkip to conversation

perf tools: Don't set inherit bit for system wide evsel

RFC PATCH In review Wang Nan 2 participants 7 messages Oct 26, 2015 – Oct 27, 2015 perf events

lore
Share & tools
lore

Wang Nan

author
#0 lore raw

v2 cover perf tools: Don't set inherit bit for system wide evsel

Previous version can be found from [1].

To make bpf_perf_event_output() and bpf_perf_event_read() easier to be used, these two patches set inherit to 0 for system wide evsels. After applying them, users are possible to pass system wide events to BPF programs in following command, and don't need to consider the setting of inherit:

 # perf record -a -e evt=cycles -e ./test__map.c/maps.pmu_map.event=evt/ ...

Patch 1/2 set evsel->system_wide for system wide target. I have checked in current evlist.c and evsel.c. The only behavior change I found is in evlist.c:

		/*
		 * The system_wide flag causes a selected event to be opened
		 * always without a pid.  Consequently it will never get a
		 * POLLHUP, but it is used for tracking in combination with
		 * other events, so it should not need to be polled anyway.
		 * Therefore don't add it for polling.
		 */
		if (!evsel->system_wide &&
		    __perf_evlist__add_pollfd(evlist, fd, idx) < 0) {
			perf_evlist__mmap_put(evlist, idx);
			return -1;
		}

With patch 1/2, __perf_evlist__add_pollfd() won't be called. From the comment I pasted this modification looks harmless.

[1] http://lkml.kernel.org/r/[email protected]

Wang Nan (2):
perf tools: Set evsel->system_wide field for global system wide
recording
perf tools: Don't set inherit bit for system wide evsel

 tools/perf/util/evsel.c | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)
Signature · 1 line
1.8.3.4
author
#1 lore raw

RFC PATCH v2 1/2 perf tools: Set evsel->system_wide field for global system wide recording

evsel->system_wide is introduced by commit bf8e8f4b832972c76d64ab2e2837 (perf evlist: Add 'system_wide' option), which is used for mixing evsels that aren't system-wide with ones that are [1]. However, for global system wide recording (perf record -a ...), evsel->system_wide is set to false, which is confusion.

This patch set evsel->system_wide to true if the target.system_wide is set, which makes evsel->system_wide a reliable way to describe whether itself is system_wide or not.

[1] http://lkml.kernel.org/r/[email protected]

Cc Adrian Hunter <[email protected]>
Cc Arnaldo Carvalho de Melo <[email protected]>
Cc Alexei Starovoitov <[email protected]>
Cc Peter Zijlstra <[email protected]>
Cc Li Zefan <[email protected]>
---
 tools/perf/util/evsel.c | 1 +
 1 file changed, 1 insertion(+)
tools/perf/util/evsel.c +1 -0
index 3ac4ee9c..36ecf0e 100644
@@ -734,6 +734,7 @@ void perf_evsel__config(struct perf_evsel *evsel, struct record_opts *opts)
 	int track = evsel->tracking;
 	bool per_cpu = opts->target.default_per_cpu && !opts->target.per_thread;
 
+	evsel->system_wide = opts->target.system_wide;
 	attr->sample_id_all = perf_missing_features.sample_id_all ? 0 : 1;
 	attr->inherit	    = !opts->no_inherit;
 
Signature · 1 line
1.8.3.4
reviewer
#2 lore raw

Re: [RFC PATCH v2 1/2] perf tools: Set evsel->system_wide field for global system wide recording

Quoted 29 lines · Wang Nan

evsel->system_wide is introduced by commit bf8e8f4b832972c76d64ab2e2837 (perf evlist: Add 'system_wide' option), which is used for mixing evsels that aren't system-wide with ones that are [1]. However, for global system wide recording (perf record -a ...), evsel->system_wide is set to false, which is confusion.

This patch set evsel->system_wide to true if the target.system_wide is set, which makes evsel->system_wide a reliable way to describe whether itself is system_wide or not.

[1] http://lkml.kernel.org/r/[email protected]

Cc Adrian Hunter <[email protected]>
Cc Arnaldo Carvalho de Melo <[email protected]>
Cc Alexei Starovoitov <[email protected]>
Cc Peter Zijlstra <[email protected]>
Cc Li Zefan <[email protected]>
---
 tools/perf/util/evsel.c | 1 +
 1 file changed, 1 insertion(+)
tools/perf/util/evsel.c +0 -0
index 3ac4ee9c..36ecf0e 100644

tools/perf/util/evsel.c void perf_evsel__config(struct perf_evsel *evsel, struct record_opts *opts) · line 737

Quoted 4 lines
tools/perf/util/evsel.c +1 -0
@@ -734,6 +734,7 @@ void perf_evsel__config(struct perf_evsel *evsel, struct record_opts *opts)
 	int track = evsel->tracking;
 	bool per_cpu = opts->target.default_per_cpu && !opts->target.per_thread;
 
+	evsel->system_wide = opts->target.system_wide;

Well that breaks the way evsel->system_wide is used i.e. it is a parameter to the evsel and here you just overwrote it.

Quoted 4 lines
tools/perf/util/evsel.c +0 -0
 	attr->sample_id_all = perf_missing_features.sample_id_all ? 0 : 1;
 	attr->inherit	    = !opts->no_inherit;
 
#3 lore raw

Re: [RFC PATCH v2 1/2] perf tools: Set evsel->system_wide field for global system wide recording

Quoted 33 lines · Adrian Hunter
Quoted 32 lines · Wang Nan

evsel->system_wide is introduced by commit bf8e8f4b832972c76d64ab2e2837 (perf evlist: Add 'system_wide' option), which is used for mixing evsels that aren't system-wide with ones that are [1]. However, for global system wide recording (perf record -a ...), evsel->system_wide is set to false, which is confusion.

This patch set evsel->system_wide to true if the target.system_wide is set, which makes evsel->system_wide a reliable way to describe whether itself is system_wide or not.

[1] http://lkml.kernel.org/r/[email protected]

Cc Adrian Hunter <[email protected]>
Cc Arnaldo Carvalho de Melo <[email protected]>
Cc Alexei Starovoitov <[email protected]>
Cc Peter Zijlstra <[email protected]>
Cc Li Zefan <[email protected]>
---
  tools/perf/util/evsel.c | 1 +
  1 file changed, 1 insertion(+)
tools/perf/util/evsel.c +0 -0
index 3ac4ee9c..36ecf0e 100644
@@ -734,6 +734,7 @@ void perf_evsel__config(struct perf_evsel *evsel, struct record_opts *opts)
  	int track = evsel->tracking;
  	bool per_cpu = opts->target.default_per_cpu && !opts->target.per_thread;
 
Quoted 3 lines

tools/perf/util/evsel.c void perf_evsel__config(struct perf_evsel *evsel, struct record_opts *opts) · line 737

Quoted 1 line
tools/perf/util/evsel.c +1 -0
+	evsel->system_wide = opts->target.system_wide;

Well that breaks the way evsel->system_wide is used i.e. it is a parameter to the evsel and here you just overwrote it.

Currently the only user of evsel->system_wide is IPT:

auxtrace_record__options -> intel_pt_recording_options

and it only set it to true.

So I think changing to this should make it safe:

evsel->system_wide = (evsel->system_wide || opt->target.system_wide);

Thought?

If we want to add further config terms we can put it to apply_config_terms(), where we can implement something like:

 # perf record -e cycles/system-wide/ -e instruction/no-system-wide/ ...

But currently I don't have such requirement.

Thank you.

reviewer
#4 lore raw

Re: [RFC PATCH v2 1/2] perf tools: Set evsel->system_wide field for global system wide recording

Quoted 45 lines · Wangnan (F)
Quoted 25 lines · Adrian Hunter
Quoted 24 lines · Wang Nan

evsel->system_wide is introduced by commit bf8e8f4b832972c76d64ab2e2837 (perf evlist: Add 'system_wide' option), which is used for mixing evsels that aren't system-wide with ones that are [1]. However, for global system wide recording (perf record -a ...), evsel->system_wide is set to false, which is confusion.

This patch set evsel->system_wide to true if the target.system_wide is set, which makes evsel->system_wide a reliable way to describe whether itself is system_wide or not.

[1] http://lkml.kernel.org/r/[email protected]

Cc Adrian Hunter <[email protected]>
Cc Arnaldo Carvalho de Melo <[email protected]>
Cc Alexei Starovoitov <[email protected]>
Cc Peter Zijlstra <[email protected]>
Cc Li Zefan <[email protected]>
---
  tools/perf/util/evsel.c | 1 +
  1 file changed, 1 insertion(+)
Quoted 12 lines

tools/perf/util/evsel.c void perf_evsel__config(struct perf_evsel *evsel, · line 738

Quoted 10 lines
tools/perf/util/evsel.c +0 -0
index 3ac4ee9c..36ecf0e 100644
@@ -734,6 +734,7 @@ void perf_evsel__config(struct perf_evsel *evsel,
 struct record_opts *opts)
      int track = evsel->tracking;
      bool per_cpu = opts->target.default_per_cpu &&
 !opts->target.per_thread;
  +    evsel->system_wide = opts->target.system_wide;

Well that breaks the way evsel->system_wide is used i.e. it is a parameter to the evsel and here you just overwrote it.

Currently the only user of evsel->system_wide is IPT:

auxtrace_record__options -> intel_pt_recording_options

Quoted 7 lines

and it only set it to true.

So I think changing to this should make it safe:

evsel->system_wide = (evsel->system_wide || opt->target.system_wide);

Thought?

That would work although it looks like a kludge.

Have you looked at properly validating opts->no_inherit ?

Also, aren't you neglecting target->cpu_list ?

Quoted 11 lines

If we want to add further config terms we can put it to apply_config_terms(), where we can implement something like:

 # perf record -e cycles/system-wide/ -e instruction/no-system-wide/ ...

But currently I don't have such requirement.

Thank you.

#5 lore raw

Re: [RFC PATCH v2 1/2] perf tools: Set evsel->system_wide field for global system wide recording

Quoted 48 lines · Adrian Hunter
Quoted 47 lines · Wangnan (F)
Quoted 25 lines · Adrian Hunter
Quoted 24 lines · Wang Nan

evsel->system_wide is introduced by commit bf8e8f4b832972c76d64ab2e2837 (perf evlist: Add 'system_wide' option), which is used for mixing evsels that aren't system-wide with ones that are [1]. However, for global system wide recording (perf record -a ...), evsel->system_wide is set to false, which is confusion.

This patch set evsel->system_wide to true if the target.system_wide is set, which makes evsel->system_wide a reliable way to describe whether itself is system_wide or not.

[1] http://lkml.kernel.org/r/[email protected]

Cc Adrian Hunter <[email protected]>
Cc Arnaldo Carvalho de Melo <[email protected]>
Cc Alexei Starovoitov <[email protected]>
Cc Peter Zijlstra <[email protected]>
Cc Li Zefan <[email protected]>
---
   tools/perf/util/evsel.c | 1 +
   1 file changed, 1 insertion(+)
Quoted 12 lines

tools/perf/util/evsel.c void perf_evsel__config(struct perf_evsel *evsel, · line 738

Quoted 10 lines
tools/perf/util/evsel.c +0 -0
index 3ac4ee9c..36ecf0e 100644
@@ -734,6 +734,7 @@ void perf_evsel__config(struct perf_evsel *evsel,
 struct record_opts *opts)
       int track = evsel->tracking;
       bool per_cpu = opts->target.default_per_cpu &&
 !opts->target.per_thread;
   +    evsel->system_wide = opts->target.system_wide;

Well that breaks the way evsel->system_wide is used i.e. it is a parameter to the evsel and here you just overwrote it.

Currently the only user of evsel->system_wide is IPT:

auxtrace_record__options -> intel_pt_recording_options

and it only set it to true.

So I think changing to this should make it safe:

Quoted 8 lines
Quoted 3 lines

evsel->system_wide = (evsel->system_wide || opt->target.system_wide);

Thought?

That would work although it looks like a kludge.

Have you looked at properly validating opts->no_inherit ?

Also, aren't you neglecting target->cpu_list ?

I need this because Alexei's bpf_perf_event_output() helper needs perf side support, and it only accept no-inherit perf event. I have already made a patch[1] to support setting inherit bit per-event, so even without this and next patch I can continue my work on bpf output. I'd like to rethink this patch after I finish perf support for bpf output.

Thank you.

[1] http://lkml.kernel.org/g/[email protected]

author
#6 lore raw

RFC PATCH v2 2/2 perf tools: Don't set inherit bit for system wide evsel

In reply to Wang Nan

Inherit bit is useless for a system wide evsel [1]. Further kernel improvements are giving more constrain [2] on inherit events. This patch set inherit bit to 0 for system wide evsels to avoid potential constrains.

[1] http://lkml.kernel.org/r/[email protected]
[2] http://lkml.kernel.org/r/[email protected]

Cc Adrian Hunter <[email protected]>
Cc Arnaldo Carvalho de Melo <[email protected]>
Cc Alexei Starovoitov <[email protected]>
Cc Peter Zijlstra <[email protected]>
Cc Li Zefan <[email protected]>
---
 tools/perf/util/evsel.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
tools/perf/util/evsel.c +1 -1
index 36ecf0e..2a5f0f8 100644
@@ -736,7 +736,7 @@ void perf_evsel__config(struct perf_evsel *evsel, struct record_opts *opts)
 
 	evsel->system_wide = opts->target.system_wide;
 	attr->sample_id_all = perf_missing_features.sample_id_all ? 0 : 1;
-	attr->inherit	    = !opts->no_inherit;
+	attr->inherit	    = !opts->no_inherit && !evsel->system_wide;
 
 	perf_evsel__set_sample_bit(evsel, IP);
 	perf_evsel__set_sample_bit(evsel, TID);
Signature · 1 line
1.8.3.4