[PATCH v3] perf record: add dummy event during system wide synthesis

[PATCH v3] perf record: add dummy event during system wide synthesis

[Ian Rogers]

During the processing of /proc during event synthesis new processes may
start. Add a dummy event if /proc is to be processed, to capture mmaps
for starting processes. This reuses the existing logic for
initial-delay.

v3 fixes the attr test of test-record-C0
v2 fixes the dummy event configuration and a branch stack issue.

Suggested-by: Stephane Eranian <eranian@google.com>
Signed-off-by: Ian Rogers <irogers@google.com>
---
 tools/perf/builtin-record.c             | 19 +++++++---
 tools/perf/tests/attr/system-wide-dummy | 50 +++++++++++++++++++++++++
 tools/perf/tests/attr/test-record-C0    | 12 +++++-
 tools/perf/util/evsel.c                 |  5 ++-
 4 files changed, 78 insertions(+), 8 deletions(-)
 create mode 100644 tools/perf/tests/attr/system-wide-dummy

diff --git a/tools/perf/builtin-record.c b/tools/perf/builtin-record.c
index 1ab349abe904..8d1e93351298 100644
--- a/tools/perf/builtin-record.c
+++ b/tools/perf/builtin-record.c
@@ -805,19 +805,28 @@ static int record__open(struct record *rec)
 	int rc = 0;
 
 	/*
-	 * For initial_delay we need to add a dummy event so that we can track
-	 * PERF_RECORD_MMAP while we wait for the initial delay to enable the
-	 * real events, the ones asked by the user.
+	 * For initial_delay or system wide, we need to add a dummy event so
+	 * that we can track PERF_RECORD_MMAP to cover the delay of waiting or
+	 * event synthesis.
 	 */
-	if (opts->initial_delay) {
+	if (opts->initial_delay || target__has_cpu(&opts->target)) {
 		if (perf_evlist__add_dummy(evlist))
 			return -ENOMEM;
 
+		/* Disable tracking of mmaps on lead event. */
 		pos = evlist__first(evlist);
 		pos->tracking = 0;
+		/* Set up dummy event. */
 		pos = evlist__last(evlist);
 		pos->tracking = 1;
-		pos->core.attr.enable_on_exec = 1;
+		/*
+		 * Enable the dummy event when the process is forked for
+		 * initial_delay, immediately for system wide.
+		 */
+		if (opts->initial_delay)
+			pos->core.attr.enable_on_exec = 1;
+		else
+			pos->immediate = 1;
 	}
 
 	perf_evlist__config(evlist, opts, &callchain_param);
diff --git a/tools/perf/tests/attr/system-wide-dummy b/tools/perf/tests/attr/system-wide-dummy
new file mode 100644
index 000000000000..eba723cc0d38
--- /dev/null
+++ b/tools/perf/tests/attr/system-wide-dummy
@@ -0,0 +1,50 @@
+# Event added by system-wide or CPU perf-record to handle the race of
+# processes starting while /proc is processed.
+[event]
+fd=1
+group_fd=-1
+cpu=*
+pid=-1
+flags=8
+type=1
+size=120
+config=9
+sample_period=4000
+sample_type=455
+read_format=4
+# Event will be enabled right away.
+disabled=0
+inherit=1
+pinned=0
+exclusive=0
+exclude_user=0
+exclude_kernel=0
+exclude_hv=0
+exclude_idle=0
+mmap=1
+comm=1
+freq=1
+inherit_stat=0
+enable_on_exec=0
+task=1
+watermark=0
+precise_ip=0
+mmap_data=0
+sample_id_all=1
+exclude_host=0
+exclude_guest=0
+exclude_callchain_kernel=0
+exclude_callchain_user=0
+mmap2=1
+comm_exec=1
+context_switch=0
+write_backward=0
+namespaces=0
+use_clockid=0
+wakeup_events=0
+bp_type=0
+config1=0
+config2=0
+branch_sample_type=0
+sample_regs_user=0
+sample_stack_user=0
diff --git a/tools/perf/tests/attr/test-record-C0 b/tools/perf/tests/attr/test-record-C0
index 93818054ae20..317730b906dd 100644
--- a/tools/perf/tests/attr/test-record-C0
+++ b/tools/perf/tests/attr/test-record-C0
@@ -9,6 +9,14 @@ cpu=0
 # no enable on exec for CPU attached
 enable_on_exec=0
 
-# PERF_SAMPLE_IP | PERF_SAMPLE_TID PERF_SAMPLE_TIME | # PERF_SAMPLE_PERIOD
+# PERF_SAMPLE_IP | PERF_SAMPLE_TID | PERF_SAMPLE_TIME |
+# PERF_SAMPLE_ID | PERF_SAMPLE_PERIOD
 # + PERF_SAMPLE_CPU added by -C 0
-sample_type=391
+sample_type=455
+
+# Dummy event handles mmaps, comm and task.
+mmap=0
+comm=0
+task=0
+
+[event:system-wide-dummy]
diff --git a/tools/perf/util/evsel.c b/tools/perf/util/evsel.c
index 6a571d322bb2..ca8f9533d8f9 100644
--- a/tools/perf/util/evsel.c
+++ b/tools/perf/util/evsel.c
@@ -1163,11 +1163,14 @@ void perf_evsel__config(struct evsel *evsel, struct record_opts *opts,
 	}
 
 	/*
+	 * A dummy event never triggers any actual counter and therefore
+	 * cannot be used with branch_stack.
+	 *
 	 * For initial_delay, a dummy event is added implicitly.
 	 * The software event will trigger -EOPNOTSUPP error out,
 	 * if BRANCH_STACK bit is set.
 	 */
-	if (opts->initial_delay && is_dummy_event(evsel))
+	if (is_dummy_event(evsel))
 		perf_evsel__reset_sample_bit(evsel, BRANCH_STACK);
 }
 
-- 
2.26.2.303.gf8c07b1a785-goog

Re: [PATCH v3] perf record: add dummy event during system wide synthesis

[Jiri Olsa]

On Wed, Apr 22, 2020 at 10:36:15AM -0700, Ian Rogers wrote:
> During the processing of /proc during event synthesis new processes may
> start. Add a dummy event if /proc is to be processed, to capture mmaps
> for starting processes. This reuses the existing logic for
> initial-delay.
> 
> v3 fixes the attr test of test-record-C0

SNIP

> +config2=0
> +branch_sample_type=0
> +sample_regs_user=0
> +sample_stack_user=0
> diff --git a/tools/perf/tests/attr/test-record-C0 b/tools/perf/tests/attr/test-record-C0
> index 93818054ae20..317730b906dd 100644
> --- a/tools/perf/tests/attr/test-record-C0
> +++ b/tools/perf/tests/attr/test-record-C0
> @@ -9,6 +9,14 @@ cpu=0
>  # no enable on exec for CPU attached
>  enable_on_exec=0
>  
> -# PERF_SAMPLE_IP | PERF_SAMPLE_TID PERF_SAMPLE_TIME | # PERF_SAMPLE_PERIOD
> +# PERF_SAMPLE_IP | PERF_SAMPLE_TID | PERF_SAMPLE_TIME |
> +# PERF_SAMPLE_ID | PERF_SAMPLE_PERIOD
>  # + PERF_SAMPLE_CPU added by -C 0
> -sample_type=391
> +sample_type=455

aah, so because now there's 2 events now, so PERF_SAMPLE_ID was added

Acked-by: Jiri Olsa <jolsa@redhat.com>

thanks,
jirka

Re: [PATCH v3] perf record: add dummy event during system wide synthesis

[Arnaldo Carvalho de Melo]

Em Wed, Apr 22, 2020 at 10:36:15AM -0700, Ian Rogers escreveu:
> During the processing of /proc during event synthesis new processes may
> start. Add a dummy event if /proc is to be processed, to capture mmaps
> for starting processes. This reuses the existing logic for
> initial-delay.
> 
> v3 fixes the attr test of test-record-C0
> v2 fixes the dummy event configuration and a branch stack issue.

Thanks, applied after splitting it up into two patches, one for this
part:

> +++ b/tools/perf/util/evsel.c
> @@ -1163,11 +1163,14 @@ void perf_evsel__config(struct evsel *evsel, struct record_opts *opts,
>       }
>
>       /*
> +      * A dummy event never triggers any actual counter and therefore
> +      * cannot be used with branch_stack.
> +      *
>        * For initial_delay, a dummy event is added implicitly.
>        * The software event will trigger -EOPNOTSUPP error out,
>        * if BRANCH_STACK bit is set.
>        */
> -     if (opts->initial_delay && is_dummy_event(evsel))
> +     if (is_dummy_event(evsel))
>               perf_evsel__reset_sample_bit(evsel, BRANCH_STACK);
>  }

Which isn't related to what is in the subject line for this patch, ok?

We don't need to get more patches, but we need to have patches that do
one thing and just that, to ease with bisecting, reverting things
sometimes, etc.

And thanks for the extra comments :-)

- Arnaldo
 
> Suggested-by: Stephane Eranian <eranian@google.com>
> Signed-off-by: Ian Rogers <irogers@google.com>
> ---
>  tools/perf/builtin-record.c             | 19 +++++++---
>  tools/perf/tests/attr/system-wide-dummy | 50 +++++++++++++++++++++++++
>  tools/perf/tests/attr/test-record-C0    | 12 +++++-
>  tools/perf/util/evsel.c                 |  5 ++-
>  4 files changed, 78 insertions(+), 8 deletions(-)
>  create mode 100644 tools/perf/tests/attr/system-wide-dummy
> 
> diff --git a/tools/perf/builtin-record.c b/tools/perf/builtin-record.c
> index 1ab349abe904..8d1e93351298 100644
> --- a/tools/perf/builtin-record.c
> +++ b/tools/perf/builtin-record.c
> @@ -805,19 +805,28 @@ static int record__open(struct record *rec)
>  	int rc = 0;
>  
>  	/*
> -	 * For initial_delay we need to add a dummy event so that we can track
> -	 * PERF_RECORD_MMAP while we wait for the initial delay to enable the
> -	 * real events, the ones asked by the user.
> +	 * For initial_delay or system wide, we need to add a dummy event so
> +	 * that we can track PERF_RECORD_MMAP to cover the delay of waiting or
> +	 * event synthesis.
>  	 */
> -	if (opts->initial_delay) {
> +	if (opts->initial_delay || target__has_cpu(&opts->target)) {
>  		if (perf_evlist__add_dummy(evlist))
>  			return -ENOMEM;
>  
> +		/* Disable tracking of mmaps on lead event. */
>  		pos = evlist__first(evlist);
>  		pos->tracking = 0;
> +		/* Set up dummy event. */
>  		pos = evlist__last(evlist);
>  		pos->tracking = 1;
> -		pos->core.attr.enable_on_exec = 1;
> +		/*
> +		 * Enable the dummy event when the process is forked for
> +		 * initial_delay, immediately for system wide.
> +		 */
> +		if (opts->initial_delay)
> +			pos->core.attr.enable_on_exec = 1;
> +		else
> +			pos->immediate = 1;
>  	}
>  
>  	perf_evlist__config(evlist, opts, &callchain_param);
> diff --git a/tools/perf/tests/attr/system-wide-dummy b/tools/perf/tests/attr/system-wide-dummy
> new file mode 100644
> index 000000000000..eba723cc0d38
> --- /dev/null
> +++ b/tools/perf/tests/attr/system-wide-dummy
> @@ -0,0 +1,50 @@
> +# Event added by system-wide or CPU perf-record to handle the race of
> +# processes starting while /proc is processed.
> +[event]
> +fd=1
> +group_fd=-1
> +cpu=*
> +pid=-1
> +flags=8
> +type=1
> +size=120
> +config=9
> +sample_period=4000
> +sample_type=455
> +read_format=4
> +# Event will be enabled right away.
> +disabled=0
> +inherit=1
> +pinned=0
> +exclusive=0
> +exclude_user=0
> +exclude_kernel=0
> +exclude_hv=0
> +exclude_idle=0
> +mmap=1
> +comm=1
> +freq=1
> +inherit_stat=0
> +enable_on_exec=0
> +task=1
> +watermark=0
> +precise_ip=0
> +mmap_data=0
> +sample_id_all=1
> +exclude_host=0
> +exclude_guest=0
> +exclude_callchain_kernel=0
> +exclude_callchain_user=0
> +mmap2=1
> +comm_exec=1
> +context_switch=0
> +write_backward=0
> +namespaces=0
> +use_clockid=0
> +wakeup_events=0
> +bp_type=0
> +config1=0
> +config2=0
> +branch_sample_type=0
> +sample_regs_user=0
> +sample_stack_user=0
> diff --git a/tools/perf/tests/attr/test-record-C0 b/tools/perf/tests/attr/test-record-C0
> index 93818054ae20..317730b906dd 100644
> --- a/tools/perf/tests/attr/test-record-C0
> +++ b/tools/perf/tests/attr/test-record-C0
> @@ -9,6 +9,14 @@ cpu=0
>  # no enable on exec for CPU attached
>  enable_on_exec=0
>  
> -# PERF_SAMPLE_IP | PERF_SAMPLE_TID PERF_SAMPLE_TIME | # PERF_SAMPLE_PERIOD
> +# PERF_SAMPLE_IP | PERF_SAMPLE_TID | PERF_SAMPLE_TIME |
> +# PERF_SAMPLE_ID | PERF_SAMPLE_PERIOD
>  # + PERF_SAMPLE_CPU added by -C 0
> -sample_type=391
> +sample_type=455
> +
> +# Dummy event handles mmaps, comm and task.
> +mmap=0
> +comm=0
> +task=0
> +
> +[event:system-wide-dummy]
> diff --git a/tools/perf/util/evsel.c b/tools/perf/util/evsel.c
> index 6a571d322bb2..ca8f9533d8f9 100644
> --- a/tools/perf/util/evsel.c
> +++ b/tools/perf/util/evsel.c
> @@ -1163,11 +1163,14 @@ void perf_evsel__config(struct evsel *evsel, struct record_opts *opts,
>  	}
>  
>  	/*
> +	 * A dummy event never triggers any actual counter and therefore
> +	 * cannot be used with branch_stack.
> +	 *
>  	 * For initial_delay, a dummy event is added implicitly.
>  	 * The software event will trigger -EOPNOTSUPP error out,
>  	 * if BRANCH_STACK bit is set.
>  	 */
> -	if (opts->initial_delay && is_dummy_event(evsel))
> +	if (is_dummy_event(evsel))
>  		perf_evsel__reset_sample_bit(evsel, BRANCH_STACK);
>  }
>  
> -- 
> 2.26.2.303.gf8c07b1a785-goog
> 

-- 

- Arnaldo

Re: [PATCH v3] perf record: add dummy event during system wide synthesis

[Ian Rogers]

On Fri, May 8, 2020 at 9:29 AM Arnaldo Carvalho de Melo
<arnaldo.melo@gmail.com> wrote:
>
> Em Wed, Apr 22, 2020 at 10:36:15AM -0700, Ian Rogers escreveu:
> > During the processing of /proc during event synthesis new processes may
> > start. Add a dummy event if /proc is to be processed, to capture mmaps
> > for starting processes. This reuses the existing logic for
> > initial-delay.
> >
> > v3 fixes the attr test of test-record-C0
> > v2 fixes the dummy event configuration and a branch stack issue.
>
> Thanks, applied after splitting it up into two patches, one for this
> part:
>
> > +++ b/tools/perf/util/evsel.c
> > @@ -1163,11 +1163,14 @@ void perf_evsel__config(struct evsel *evsel, struct record_opts *opts,
> >       }
> >
> >       /*
> > +      * A dummy event never triggers any actual counter and therefore
> > +      * cannot be used with branch_stack.
> > +      *
> >        * For initial_delay, a dummy event is added implicitly.
> >        * The software event will trigger -EOPNOTSUPP error out,
> >        * if BRANCH_STACK bit is set.
> >        */
> > -     if (opts->initial_delay && is_dummy_event(evsel))
> > +     if (is_dummy_event(evsel))
> >               perf_evsel__reset_sample_bit(evsel, BRANCH_STACK);
> >  }
>
> Which isn't related to what is in the subject line for this patch, ok?
>
> We don't need to get more patches, but we need to have patches that do
> one thing and just that, to ease with bisecting, reverting things
> sometimes, etc.

Thanks for merging! Won't that separation break system wide mode? I
guess you can do the evsel change first and then the system wide
change, but the effect of 1 change won't be felt without the other.
I'm looking to keep patches separate but when one patch can only be
tested with another in place the separation is hard to work with for
testing. There is also the issue of trying to understand the context a
change was made, which is broken by separation. Anyway, I want to
minimize your work in merging changes so I'll try to fit your needs
with future patches.

Thanks,
Ian

> And thanks for the extra comments :-)
>
> - Arnaldo
>
> > Suggested-by: Stephane Eranian <eranian@google.com>
> > Signed-off-by: Ian Rogers <irogers@google.com>
> > ---
> >  tools/perf/builtin-record.c             | 19 +++++++---
> >  tools/perf/tests/attr/system-wide-dummy | 50 +++++++++++++++++++++++++
> >  tools/perf/tests/attr/test-record-C0    | 12 +++++-
> >  tools/perf/util/evsel.c                 |  5 ++-
> >  4 files changed, 78 insertions(+), 8 deletions(-)
> >  create mode 100644 tools/perf/tests/attr/system-wide-dummy
> >
> > diff --git a/tools/perf/builtin-record.c b/tools/perf/builtin-record.c
> > index 1ab349abe904..8d1e93351298 100644
> > --- a/tools/perf/builtin-record.c
> > +++ b/tools/perf/builtin-record.c
> > @@ -805,19 +805,28 @@ static int record__open(struct record *rec)
> >       int rc = 0;
> >
> >       /*
> > -      * For initial_delay we need to add a dummy event so that we can track
> > -      * PERF_RECORD_MMAP while we wait for the initial delay to enable the
> > -      * real events, the ones asked by the user.
> > +      * For initial_delay or system wide, we need to add a dummy event so
> > +      * that we can track PERF_RECORD_MMAP to cover the delay of waiting or
> > +      * event synthesis.
> >        */
> > -     if (opts->initial_delay) {
> > +     if (opts->initial_delay || target__has_cpu(&opts->target)) {
> >               if (perf_evlist__add_dummy(evlist))
> >                       return -ENOMEM;
> >
> > +             /* Disable tracking of mmaps on lead event. */
> >               pos = evlist__first(evlist);
> >               pos->tracking = 0;
> > +             /* Set up dummy event. */
> >               pos = evlist__last(evlist);
> >               pos->tracking = 1;
> > -             pos->core.attr.enable_on_exec = 1;
> > +             /*
> > +              * Enable the dummy event when the process is forked for
> > +              * initial_delay, immediately for system wide.
> > +              */
> > +             if (opts->initial_delay)
> > +                     pos->core.attr.enable_on_exec = 1;
> > +             else
> > +                     pos->immediate = 1;
> >       }
> >
> >       perf_evlist__config(evlist, opts, &callchain_param);
> > diff --git a/tools/perf/tests/attr/system-wide-dummy b/tools/perf/tests/attr/system-wide-dummy
> > new file mode 100644
> > index 000000000000..eba723cc0d38
> > --- /dev/null
> > +++ b/tools/perf/tests/attr/system-wide-dummy
> > @@ -0,0 +1,50 @@
> > +# Event added by system-wide or CPU perf-record to handle the race of
> > +# processes starting while /proc is processed.
> > +[event]
> > +fd=1
> > +group_fd=-1
> > +cpu=*
> > +pid=-1
> > +flags=8
> > +type=1
> > +size=120
> > +config=9
> > +sample_period=4000
> > +sample_type=455
> > +read_format=4
> > +# Event will be enabled right away.
> > +disabled=0
> > +inherit=1
> > +pinned=0
> > +exclusive=0
> > +exclude_user=0
> > +exclude_kernel=0
> > +exclude_hv=0
> > +exclude_idle=0
> > +mmap=1
> > +comm=1
> > +freq=1
> > +inherit_stat=0
> > +enable_on_exec=0
> > +task=1
> > +watermark=0
> > +precise_ip=0
> > +mmap_data=0
> > +sample_id_all=1
> > +exclude_host=0
> > +exclude_guest=0
> > +exclude_callchain_kernel=0
> > +exclude_callchain_user=0
> > +mmap2=1
> > +comm_exec=1
> > +context_switch=0
> > +write_backward=0
> > +namespaces=0
> > +use_clockid=0
> > +wakeup_events=0
> > +bp_type=0
> > +config1=0
> > +config2=0
> > +branch_sample_type=0
> > +sample_regs_user=0
> > +sample_stack_user=0
> > diff --git a/tools/perf/tests/attr/test-record-C0 b/tools/perf/tests/attr/test-record-C0
> > index 93818054ae20..317730b906dd 100644
> > --- a/tools/perf/tests/attr/test-record-C0
> > +++ b/tools/perf/tests/attr/test-record-C0
> > @@ -9,6 +9,14 @@ cpu=0
> >  # no enable on exec for CPU attached
> >  enable_on_exec=0
> >
> > -# PERF_SAMPLE_IP | PERF_SAMPLE_TID PERF_SAMPLE_TIME | # PERF_SAMPLE_PERIOD
> > +# PERF_SAMPLE_IP | PERF_SAMPLE_TID | PERF_SAMPLE_TIME |
> > +# PERF_SAMPLE_ID | PERF_SAMPLE_PERIOD
> >  # + PERF_SAMPLE_CPU added by -C 0
> > -sample_type=391
> > +sample_type=455
> > +
> > +# Dummy event handles mmaps, comm and task.
> > +mmap=0
> > +comm=0
> > +task=0
> > +
> > +[event:system-wide-dummy]
> > diff --git a/tools/perf/util/evsel.c b/tools/perf/util/evsel.c
> > index 6a571d322bb2..ca8f9533d8f9 100644
> > --- a/tools/perf/util/evsel.c
> > +++ b/tools/perf/util/evsel.c
> > @@ -1163,11 +1163,14 @@ void perf_evsel__config(struct evsel *evsel, struct record_opts *opts,
> >       }
> >
> >       /*
> > +      * A dummy event never triggers any actual counter and therefore
> > +      * cannot be used with branch_stack.
> > +      *
> >        * For initial_delay, a dummy event is added implicitly.
> >        * The software event will trigger -EOPNOTSUPP error out,
> >        * if BRANCH_STACK bit is set.
> >        */
> > -     if (opts->initial_delay && is_dummy_event(evsel))
> > +     if (is_dummy_event(evsel))
> >               perf_evsel__reset_sample_bit(evsel, BRANCH_STACK);
> >  }
> >
> > --
> > 2.26.2.303.gf8c07b1a785-goog
> >
>
> --
>
> - Arnaldo

Re: [PATCH v3] perf record: add dummy event during system wide synthesis

[Arnaldo Carvalho de Melo]

Em Wed, Apr 22, 2020 at 10:36:15AM -0700, Ian Rogers escreveu:
> During the processing of /proc during event synthesis new processes may
> start. Add a dummy event if /proc is to be processed, to capture mmaps
> for starting processes. This reuses the existing logic for
> initial-delay.
> 
> v3 fixes the attr test of test-record-C0
> v2 fixes the dummy event configuration and a branch stack issue.

Something I noticed only now is that this ends up in the perf.data file,
and we don't need it at all there, i.e.

  # perf record -I

I.e. system wide, asking for registers now ends up with:

[root@quaco ~]# perf record -I
^C[ perf record: Woken up 1 times to write data ]
[ perf record: Captured and wrote 2.855 MB perf.data (4902 samples) ]
[root@quaco ~]# perf evlist
cycles
dummy:HG
[root@quaco ~]# perf evlist -v
cycles: size: 120, { sample_period, sample_freq }: 4000, sample_type: IP|TID|TIME|ID|CPU|PERIOD|REGS_INTR, read_format: ID, disabled: 1, inherit: 1, freq: 1, precise_ip: 3, sample_id_all: 1, exclude_guest: 1, sample_regs_intr: 0xff0fff
dummy:HG: type: 1, size: 120, config: 0x9, { sample_period, sample_freq }: 4000, sample_type: IP|TID|TIME|ID|CPU|PERIOD|REGS_INTR, read_format: ID, inherit: 1, mmap: 1, comm: 1, freq: 1, task: 1, sample_id_all: 1, mmap2: 1, comm_exec: 1, ksymbol: 1, bpf_event: 1, sample_regs_intr: 0xff0fff
[root@quaco ~]#

For perf top is ok to reuse the main evlist, as those are not going to
hit the disk, but for 'perf record' it pollutes the perf.data file with
that dummy event.

This was a problem introduced with initial-delay, that IIRC predates the
side band thread tho, I'll have to think about it, just writing this
down to revisit this, as may raise some eyebrows by now being more
exposed.

- Arnaldo
 
> Suggested-by: Stephane Eranian <eranian@google.com>
> Signed-off-by: Ian Rogers <irogers@google.com>
> ---
>  tools/perf/builtin-record.c             | 19 +++++++---
>  tools/perf/tests/attr/system-wide-dummy | 50 +++++++++++++++++++++++++
>  tools/perf/tests/attr/test-record-C0    | 12 +++++-
>  tools/perf/util/evsel.c                 |  5 ++-
>  4 files changed, 78 insertions(+), 8 deletions(-)
>  create mode 100644 tools/perf/tests/attr/system-wide-dummy
> 
> diff --git a/tools/perf/builtin-record.c b/tools/perf/builtin-record.c
> index 1ab349abe904..8d1e93351298 100644
> --- a/tools/perf/builtin-record.c
> +++ b/tools/perf/builtin-record.c
> @@ -805,19 +805,28 @@ static int record__open(struct record *rec)
>  	int rc = 0;
>  
>  	/*
> -	 * For initial_delay we need to add a dummy event so that we can track
> -	 * PERF_RECORD_MMAP while we wait for the initial delay to enable the
> -	 * real events, the ones asked by the user.
> +	 * For initial_delay or system wide, we need to add a dummy event so
> +	 * that we can track PERF_RECORD_MMAP to cover the delay of waiting or
> +	 * event synthesis.
>  	 */
> -	if (opts->initial_delay) {
> +	if (opts->initial_delay || target__has_cpu(&opts->target)) {
>  		if (perf_evlist__add_dummy(evlist))
>  			return -ENOMEM;
>  
> +		/* Disable tracking of mmaps on lead event. */
>  		pos = evlist__first(evlist);
>  		pos->tracking = 0;
> +		/* Set up dummy event. */
>  		pos = evlist__last(evlist);
>  		pos->tracking = 1;
> -		pos->core.attr.enable_on_exec = 1;
> +		/*
> +		 * Enable the dummy event when the process is forked for
> +		 * initial_delay, immediately for system wide.
> +		 */
> +		if (opts->initial_delay)
> +			pos->core.attr.enable_on_exec = 1;
> +		else
> +			pos->immediate = 1;
>  	}
>  
>  	perf_evlist__config(evlist, opts, &callchain_param);
> diff --git a/tools/perf/tests/attr/system-wide-dummy b/tools/perf/tests/attr/system-wide-dummy
> new file mode 100644
> index 000000000000..eba723cc0d38
> --- /dev/null
> +++ b/tools/perf/tests/attr/system-wide-dummy
> @@ -0,0 +1,50 @@
> +# Event added by system-wide or CPU perf-record to handle the race of
> +# processes starting while /proc is processed.
> +[event]
> +fd=1
> +group_fd=-1
> +cpu=*
> +pid=-1
> +flags=8
> +type=1
> +size=120
> +config=9
> +sample_period=4000
> +sample_type=455
> +read_format=4
> +# Event will be enabled right away.
> +disabled=0
> +inherit=1
> +pinned=0
> +exclusive=0
> +exclude_user=0
> +exclude_kernel=0
> +exclude_hv=0
> +exclude_idle=0
> +mmap=1
> +comm=1
> +freq=1
> +inherit_stat=0
> +enable_on_exec=0
> +task=1
> +watermark=0
> +precise_ip=0
> +mmap_data=0
> +sample_id_all=1
> +exclude_host=0
> +exclude_guest=0
> +exclude_callchain_kernel=0
> +exclude_callchain_user=0
> +mmap2=1
> +comm_exec=1
> +context_switch=0
> +write_backward=0
> +namespaces=0
> +use_clockid=0
> +wakeup_events=0
> +bp_type=0
> +config1=0
> +config2=0
> +branch_sample_type=0
> +sample_regs_user=0
> +sample_stack_user=0
> diff --git a/tools/perf/tests/attr/test-record-C0 b/tools/perf/tests/attr/test-record-C0
> index 93818054ae20..317730b906dd 100644
> --- a/tools/perf/tests/attr/test-record-C0
> +++ b/tools/perf/tests/attr/test-record-C0
> @@ -9,6 +9,14 @@ cpu=0
>  # no enable on exec for CPU attached
>  enable_on_exec=0
>  
> -# PERF_SAMPLE_IP | PERF_SAMPLE_TID PERF_SAMPLE_TIME | # PERF_SAMPLE_PERIOD
> +# PERF_SAMPLE_IP | PERF_SAMPLE_TID | PERF_SAMPLE_TIME |
> +# PERF_SAMPLE_ID | PERF_SAMPLE_PERIOD
>  # + PERF_SAMPLE_CPU added by -C 0
> -sample_type=391
> +sample_type=455
> +
> +# Dummy event handles mmaps, comm and task.
> +mmap=0
> +comm=0
> +task=0
> +
> +[event:system-wide-dummy]
> diff --git a/tools/perf/util/evsel.c b/tools/perf/util/evsel.c
> index 6a571d322bb2..ca8f9533d8f9 100644
> --- a/tools/perf/util/evsel.c
> +++ b/tools/perf/util/evsel.c
> @@ -1163,11 +1163,14 @@ void perf_evsel__config(struct evsel *evsel, struct record_opts *opts,
>  	}
>  
>  	/*
> +	 * A dummy event never triggers any actual counter and therefore
> +	 * cannot be used with branch_stack.
> +	 *
>  	 * For initial_delay, a dummy event is added implicitly.
>  	 * The software event will trigger -EOPNOTSUPP error out,
>  	 * if BRANCH_STACK bit is set.
>  	 */
> -	if (opts->initial_delay && is_dummy_event(evsel))
> +	if (is_dummy_event(evsel))
>  		perf_evsel__reset_sample_bit(evsel, BRANCH_STACK);
>  }
>  
> -- 
> 2.26.2.303.gf8c07b1a785-goog
> 

-- 

- Arnaldo

Re: [PATCH v3] perf record: add dummy event during system wide synthesis

[Ian Rogers]

On Tue, May 19, 2020 at 6:54 PM Arnaldo Carvalho de Melo
<arnaldo.melo@gmail.com> wrote:
>
> Em Wed, Apr 22, 2020 at 10:36:15AM -0700, Ian Rogers escreveu:
> > During the processing of /proc during event synthesis new processes may
> > start. Add a dummy event if /proc is to be processed, to capture mmaps
> > for starting processes. This reuses the existing logic for
> > initial-delay.
> >
> > v3 fixes the attr test of test-record-C0
> > v2 fixes the dummy event configuration and a branch stack issue.
>
> Something I noticed only now is that this ends up in the perf.data file,
> and we don't need it at all there, i.e.
>
>   # perf record -I
>
> I.e. system wide, asking for registers now ends up with:
>
> [root@quaco ~]# perf record -I
> ^C[ perf record: Woken up 1 times to write data ]
> [ perf record: Captured and wrote 2.855 MB perf.data (4902 samples) ]
> [root@quaco ~]# perf evlist
> cycles
> dummy:HG
> [root@quaco ~]# perf evlist -v
> cycles: size: 120, { sample_period, sample_freq }: 4000, sample_type: IP|TID|TIME|ID|CPU|PERIOD|REGS_INTR, read_format: ID, disabled: 1, inherit: 1, freq: 1, precise_ip: 3, sample_id_all: 1, exclude_guest: 1, sample_regs_intr: 0xff0fff
> dummy:HG: type: 1, size: 120, config: 0x9, { sample_period, sample_freq }: 4000, sample_type: IP|TID|TIME|ID|CPU|PERIOD|REGS_INTR, read_format: ID, inherit: 1, mmap: 1, comm: 1, freq: 1, task: 1, sample_id_all: 1, mmap2: 1, comm_exec: 1, ksymbol: 1, bpf_event: 1, sample_regs_intr: 0xff0fff
> [root@quaco ~]#
>
> For perf top is ok to reuse the main evlist, as those are not going to
> hit the disk, but for 'perf record' it pollutes the perf.data file with
> that dummy event.
>
> This was a problem introduced with initial-delay, that IIRC predates the
> side band thread tho, I'll have to think about it, just writing this
> down to revisit this, as may raise some eyebrows by now being more
> exposed.

Agreed. We've had to adjust some tooling like the protobuf convertor
because of this:
https://github.com/google/perf_data_converter/pull/88

Thanks,
Ian

> - Arnaldo
>
> > Suggested-by: Stephane Eranian <eranian@google.com>
> > Signed-off-by: Ian Rogers <irogers@google.com>
> > ---
> >  tools/perf/builtin-record.c             | 19 +++++++---
> >  tools/perf/tests/attr/system-wide-dummy | 50 +++++++++++++++++++++++++
> >  tools/perf/tests/attr/test-record-C0    | 12 +++++-
> >  tools/perf/util/evsel.c                 |  5 ++-
> >  4 files changed, 78 insertions(+), 8 deletions(-)
> >  create mode 100644 tools/perf/tests/attr/system-wide-dummy
> >
> > diff --git a/tools/perf/builtin-record.c b/tools/perf/builtin-record.c
> > index 1ab349abe904..8d1e93351298 100644
> > --- a/tools/perf/builtin-record.c
> > +++ b/tools/perf/builtin-record.c
> > @@ -805,19 +805,28 @@ static int record__open(struct record *rec)
> >       int rc = 0;
> >
> >       /*
> > -      * For initial_delay we need to add a dummy event so that we can track
> > -      * PERF_RECORD_MMAP while we wait for the initial delay to enable the
> > -      * real events, the ones asked by the user.
> > +      * For initial_delay or system wide, we need to add a dummy event so
> > +      * that we can track PERF_RECORD_MMAP to cover the delay of waiting or
> > +      * event synthesis.
> >        */
> > -     if (opts->initial_delay) {
> > +     if (opts->initial_delay || target__has_cpu(&opts->target)) {
> >               if (perf_evlist__add_dummy(evlist))
> >                       return -ENOMEM;
> >
> > +             /* Disable tracking of mmaps on lead event. */
> >               pos = evlist__first(evlist);
> >               pos->tracking = 0;
> > +             /* Set up dummy event. */
> >               pos = evlist__last(evlist);
> >               pos->tracking = 1;
> > -             pos->core.attr.enable_on_exec = 1;
> > +             /*
> > +              * Enable the dummy event when the process is forked for
> > +              * initial_delay, immediately for system wide.
> > +              */
> > +             if (opts->initial_delay)
> > +                     pos->core.attr.enable_on_exec = 1;
> > +             else
> > +                     pos->immediate = 1;
> >       }
> >
> >       perf_evlist__config(evlist, opts, &callchain_param);
> > diff --git a/tools/perf/tests/attr/system-wide-dummy b/tools/perf/tests/attr/system-wide-dummy
> > new file mode 100644
> > index 000000000000..eba723cc0d38
> > --- /dev/null
> > +++ b/tools/perf/tests/attr/system-wide-dummy
> > @@ -0,0 +1,50 @@
> > +# Event added by system-wide or CPU perf-record to handle the race of
> > +# processes starting while /proc is processed.
> > +[event]
> > +fd=1
> > +group_fd=-1
> > +cpu=*
> > +pid=-1
> > +flags=8
> > +type=1
> > +size=120
> > +config=9
> > +sample_period=4000
> > +sample_type=455
> > +read_format=4
> > +# Event will be enabled right away.
> > +disabled=0
> > +inherit=1
> > +pinned=0
> > +exclusive=0
> > +exclude_user=0
> > +exclude_kernel=0
> > +exclude_hv=0
> > +exclude_idle=0
> > +mmap=1
> > +comm=1
> > +freq=1
> > +inherit_stat=0
> > +enable_on_exec=0
> > +task=1
> > +watermark=0
> > +precise_ip=0
> > +mmap_data=0
> > +sample_id_all=1
> > +exclude_host=0
> > +exclude_guest=0
> > +exclude_callchain_kernel=0
> > +exclude_callchain_user=0
> > +mmap2=1
> > +comm_exec=1
> > +context_switch=0
> > +write_backward=0
> > +namespaces=0
> > +use_clockid=0
> > +wakeup_events=0
> > +bp_type=0
> > +config1=0
> > +config2=0
> > +branch_sample_type=0
> > +sample_regs_user=0
> > +sample_stack_user=0
> > diff --git a/tools/perf/tests/attr/test-record-C0 b/tools/perf/tests/attr/test-record-C0
> > index 93818054ae20..317730b906dd 100644
> > --- a/tools/perf/tests/attr/test-record-C0
> > +++ b/tools/perf/tests/attr/test-record-C0
> > @@ -9,6 +9,14 @@ cpu=0
> >  # no enable on exec for CPU attached
> >  enable_on_exec=0
> >
> > -# PERF_SAMPLE_IP | PERF_SAMPLE_TID PERF_SAMPLE_TIME | # PERF_SAMPLE_PERIOD
> > +# PERF_SAMPLE_IP | PERF_SAMPLE_TID | PERF_SAMPLE_TIME |
> > +# PERF_SAMPLE_ID | PERF_SAMPLE_PERIOD
> >  # + PERF_SAMPLE_CPU added by -C 0
> > -sample_type=391
> > +sample_type=455
> > +
> > +# Dummy event handles mmaps, comm and task.
> > +mmap=0
> > +comm=0
> > +task=0
> > +
> > +[event:system-wide-dummy]
> > diff --git a/tools/perf/util/evsel.c b/tools/perf/util/evsel.c
> > index 6a571d322bb2..ca8f9533d8f9 100644
> > --- a/tools/perf/util/evsel.c
> > +++ b/tools/perf/util/evsel.c
> > @@ -1163,11 +1163,14 @@ void perf_evsel__config(struct evsel *evsel, struct record_opts *opts,
> >       }
> >
> >       /*
> > +      * A dummy event never triggers any actual counter and therefore
> > +      * cannot be used with branch_stack.
> > +      *
> >        * For initial_delay, a dummy event is added implicitly.
> >        * The software event will trigger -EOPNOTSUPP error out,
> >        * if BRANCH_STACK bit is set.
> >        */
> > -     if (opts->initial_delay && is_dummy_event(evsel))
> > +     if (is_dummy_event(evsel))
> >               perf_evsel__reset_sample_bit(evsel, BRANCH_STACK);
> >  }
> >
> > --
> > 2.26.2.303.gf8c07b1a785-goog
> >
>
> --
>
> - Arnaldo