Re: [PATCH 7/7] perf record: Introduce --switch-output-event

[Arnaldo Carvalho de Melo]

Em Tue, Apr 28, 2020 at 11:48:39AM +0200, Jiri Olsa escreveu:
> On Mon, Apr 27, 2020 at 06:19:35PM -0300, Arnaldo Carvalho de Melo wrote:
> 
> SNIP
> 
> > ---
> >  tools/perf/Documentation/perf-record.txt | 13 ++++++++
> >  tools/perf/builtin-record.c              | 40 +++++++++++++++++++++---
> >  2 files changed, 49 insertions(+), 4 deletions(-)
> > 
> > diff --git a/tools/perf/Documentation/perf-record.txt b/tools/perf/Documentation/perf-record.txt
> > index 6e8b4649307c..561ef55743e2 100644
> > --- a/tools/perf/Documentation/perf-record.txt
> > +++ b/tools/perf/Documentation/perf-record.txt
> > @@ -556,6 +556,19 @@ overhead. You can still switch them on with:
> >  
> >    --switch-output --no-no-buildid  --no-no-buildid-cache
> >  
> > +--switch-output-event::
> > +Events that will cause the switch of the perf.data file, auto-selecting
> > +--switch-output=signal, the results are similar as internally the side band
> > +thread will also send a SIGUSR2 to the main one.
> > +
> > +Uses the same syntax as --event, it will just not be recorded, serving only to
> > +switch the perf.data file as soon as the --switch-output event is processed by
> > +a separate sideband thread.
> > +
> > +This sideband thread is also used to other purposes, like processing the
> > +PERF_RECORD_BPF_EVENT records as they happen, asking the kernel for extra BPF
> > +information, etc.
> 
> first I thought we should follow the --switch-output 'xxx' way,
> but then you need to specify an event, so I guess --switch-output-event
> is better

Yeah, otherwise we'd have to play a bit more tricky games with
--switch-output parsing, to support something like:

--switch-output=event(event_desc_str)

and pass that event_desc_str to parse_events_option_new_evlist(), do you
think that sounds better?
 
> > +
> >  --switch-max-files=N::
> >  
> >  When rotating perf.data with --switch-output, only keep N files.
> > diff --git a/tools/perf/builtin-record.c b/tools/perf/builtin-record.c
> > index ed2244847400..8ff5eaec26e9 100644
> > --- a/tools/perf/builtin-record.c
> > +++ b/tools/perf/builtin-record.c
> > @@ -87,6 +87,7 @@ struct record {
> >  	struct evlist	*evlist;
> >  	struct perf_session	*session;
> >  	struct evlist		*sb_evlist;
> > +	pthread_t		thread_id;
> >  	int			realtime_prio;
> >  	bool			no_buildid;
> >  	bool			no_buildid_set;
> > @@ -1436,6 +1437,13 @@ static int record__synthesize(struct record *rec, bool tail)
> >  	return err;
> >  }
> >  
> > +static int record__process_signal_event(union perf_event *event __maybe_unused, void *data)
> > +{
> > +	struct record *rec = data;
> > +	pthread_kill(rec->thread_id, SIGUSR2);
> > +	return 0;
> > +}
> > +
> >  static int __cmd_record(struct record *rec, int argc, const char **argv)
> >  {
> >  	int err;
> > @@ -1580,12 +1588,24 @@ static int __cmd_record(struct record *rec, int argc, const char **argv)
> >  		goto out_child;
> >  	}
> >  
> > -	if (!opts->no_bpf_event) {
> > -		rec->sb_evlist = evlist__new();
> > +	if (rec->sb_evlist != NULL) {
> > +		/*
> > +		 * We get here if --switch-output-event populated the
> > +		 * sb_evlist, so associate a callback that will send a SIGUSR2
> > +		 * to the main thread.
> > +		 */
> > +		evlist__set_cb(rec->sb_evlist, record__process_signal_event, rec);
> > +		rec->thread_id = pthread_self();
> > +	}
> >  
> > +	if (!opts->no_bpf_event) {
> >  		if (rec->sb_evlist == NULL) {
> > -			pr_err("Couldn't create side band evlist.\n.");
> > -			goto out_child;
> > +			rec->sb_evlist = evlist__new();
> > +
> > +			if (rec->sb_evlist == NULL) {
> > +				pr_err("Couldn't create side band evlist.\n.");
> > +				goto out_child;
> > +			}
> >  		}
> >  
> >  		if (evlist__add_bpf_sb_event(rec->sb_evlist, &session->header.env)) {
> 
> it's getting bigger, I wonder we should put all the sb_* setup in
> separated functions like sb_start/sb_stop

Well, the rec->thread_id = pthread_self(); part is just for reusing a
'perf record' specific mechanism, what to do when the event appears in
the side band thread ring buffer, the evlist__set_cb() also is related
to that, moving thread_id to evlist seems too much at this time.

> > @@ -2179,10 +2199,19 @@ static int switch_output_setup(struct record *rec)
> >  	};
> >  	unsigned long val;
> >  
> > +	/*
> > +	 * If we're using --switch-output-events, then we imply its 
> > +	 * --switch-output=signal, as we'll send a SIGUSR2 from the side band
> > +	 *  thread to its parent.
> > +	 */
> > +	if (rec->sb_evlist != NULL)
> > +		goto do_signal;
> > +
> >  	if (!s->set)
> >  		return 0;
 
> hum, it looks like this jump is not necessay and can be avoided
> by some bool checks.. could we add some bool when --switch-output-event
> is used, so we don't depend on wether rec->sb_evlist was allocated for
> whatever reason?

If rec->sb_evlist is NULL, then there was no --switch-output-event
passed... The only advantage in adding the complexity below would be if
we had rec->switch_output_event_set which would clarify that sb_evlist
is not used only for --switch-output-event, to make things clearer.

And this still leaves us with the jump, otherwise we would have to test
it twice, right?

I think I'll separate the patch adding OPT_CALLBACK_SET(), then fold the
switch_output_event_set addition to builtin-record, ok?

- Arnaldo

diff --git a/tools/lib/subcmd/parse-options.h b/tools/lib/subcmd/parse-options.h
index af9def589863..d2414144eb8c 100644
--- a/tools/lib/subcmd/parse-options.h
+++ b/tools/lib/subcmd/parse-options.h
@@ -151,6 +151,8 @@ struct option {
 	{ .type = OPTION_CALLBACK, .short_name = (s), .long_name = (l), .value = (v), .argh = "time", .help = (h), .callback = parse_opt_approxidate_cb }
 #define OPT_CALLBACK(s, l, v, a, h, f) \
 	{ .type = OPTION_CALLBACK, .short_name = (s), .long_name = (l), .value = (v), .argh = (a), .help = (h), .callback = (f) }
+#define OPT_CALLBACK_SET(s, l, v, os, a, h, f) \
+	{ .type = OPTION_CALLBACK, .short_name = (s), .long_name = (l), .value = (v), .argh = (a), .help = (h), .callback = (f), .set = check_vtype(os, bool *)}
 #define OPT_CALLBACK_NOOPT(s, l, v, a, h, f) \
 	{ .type = OPTION_CALLBACK, .short_name = (s), .long_name = (l), .value = (v), .argh = (a), .help = (h), .callback = (f), .flags = PARSE_OPT_NOARG }
 #define OPT_CALLBACK_DEFAULT(s, l, v, a, h, f, d) \
diff --git a/tools/perf/builtin-record.c b/tools/perf/builtin-record.c
index 8ff5eaec26e9..7a6a89972691 100644
--- a/tools/perf/builtin-record.c
+++ b/tools/perf/builtin-record.c
@@ -89,6 +89,7 @@ struct record {
 	struct evlist		*sb_evlist;
 	pthread_t		thread_id;
 	int			realtime_prio;
+	bool			switch_output_event_set;
 	bool			no_buildid;
 	bool			no_buildid_set;
 	bool			no_buildid_cache;
@@ -2204,7 +2205,7 @@ static int switch_output_setup(struct record *rec)
 	 * --switch-output=signal, as we'll send a SIGUSR2 from the side band
 	 *  thread to its parent.
 	 */
-	if (rec->sb_evlist != NULL)
+	if (rec->switch_output_event_set)
 		goto do_signal;
 
 	if (!s->set)
@@ -2469,9 +2470,9 @@ static struct option __record_options[] = {
 			  &record.switch_output.set, "signal or size[BKMG] or time[smhd]",
 			  "Switch output when receiving SIGUSR2 (signal) or cross a size or time threshold",
 			  "signal"),
-	OPT_CALLBACK(0, "switch-output-event", &record.sb_evlist, "switch output event",
-		     "switch output event selector. use 'perf list' to list available events",
-		     parse_events_option_new_evlist),
+	OPT_CALLBACK_SET(0, "switch-output-event", &record.sb_evlist, &record.switch_output_event_set, "switch output event",
+			 "switch output event selector. use 'perf list' to list available events",
+			 parse_events_option_new_evlist),
 	OPT_INTEGER(0, "switch-max-files", &record.switch_output.num_files,
 		   "Limit number of switch output generated files"),
 	OPT_BOOLEAN(0, "dry-run", &dry_run,