Skip to contentSkip to conversation

perf, tools: Add PERF_PID

PATCH In review Andi Kleen 3 participants 5 messages Sep 24, 2014 – Oct 1, 2014 perf events

lore
Share & tools
lore

Andi Kleen

author
#0 lore raw

PATCH 1/2 perf, tools: Add PERF_PID

From: Andi Kleen <[email protected]>

It's currently difficult to filter out perf itself using a filter. This can give cascading effects during IO tracing when the IO perf does itself causes more trace output.

The best way to filter is to use the pid. But it's difficult to get the pid of perf without using hacks.

Add a PERF_PID meta variable to the perf filter that contains the current pid.

With this patch the following works

% perf record -e syscalls:sys_enter_write -a --filter 'common_pid != PERF_PID' ...

This won't work for more complex perf pipe lines with multiple processes, but at least solves the problem nicely for a single perf.

v2: Remove debug code. Don't use zero padding (Namhyung)
v3: Correct patch.
v4: Handle arbitary pid length.
v5: Fix a warning

---
 tools/perf/Documentation/perf-record.txt |  2 +-
 tools/perf/util/parse-events.c           | 21 ++++++++++++++++++++-
 2 files changed, 21 insertions(+), 2 deletions(-)
tools/perf/Documentation/perf-record.txt +1 -1
index d460049..b6c5e51 100644
@@ -41,7 +41,7 @@ OPTIONS
           'mem:0x1000:rw'.
 
 --filter=<filter>::
-        Event filter.
+        Event filter. PERF_PID represents the perf pid.
 
 -a::
 --all-cpus::
tools/perf/util/parse-events.c +20 -1
index 1e15df1..9c4bab8 100644
@@ -967,6 +967,9 @@ int parse_filter(const struct option *opt, const char *str,
 {
 	struct perf_evlist *evlist = *(struct perf_evlist **)opt->value;
 	struct perf_evsel *last = NULL;
+	char *pid, buf[30], *o;
+	int len, plen;
+	const char *p;
 
 	if (evlist->nr_entries > 0)
 		last = perf_evlist__last(evlist);
@@ -977,12 +980,28 @@ int parse_filter(const struct option *opt, const char *str,
 		return -1;
 	}
 
-	last->filter = strdup(str);
+	plen = sprintf(buf, "%d", getpid());
+	len = strlen(str) + 1;
+	for (pid = strstr(str, "PERF_PID"); pid;
+	     pid = strstr(pid + 1, "PERF_PID")) {
+		len += plen - 8;
+	}
+
+	last->filter = malloc(len);
 	if (last->filter == NULL) {
 		fprintf(stderr, "not enough memory to hold filter string\n");
 		return -1;
 	}
 
+	o = last->filter;
+	for (p = str; *p; ) {
+		if (*p == 'P' && !strncmp(p, "PERF_PID", 8)) {
+			strcpy(o, buf);
+			o += plen;
+			p += sizeof("PERF_PID") - 1;
+		} else
+			*o++ = *p++;
+	}
 	return 0;
 }
 
Signature · 1 line
1.9.3
author
#1 lore raw

PATCH 2/2 perf tools: apply -F filter to all previous events

From: Andi Kleen <[email protected]>

Right now you would need to duplicate the filter definition for every trace point event. Instead apply it to each previous one that did not have its own filter. If you have some tracepoint events you don't want to give filter you can put them after the last one.

---
 tools/perf/util/parse-events.c | 48 +++++++++++++++++++++++++++---------------
 1 file changed, 31 insertions(+), 17 deletions(-)
tools/perf/util/parse-events.c +31 -17
index 9c4bab8..9bf4173 100644
@@ -962,23 +962,12 @@ int parse_events_option(const struct option *opt, const char *str,
 	return ret;
 }
 
-int parse_filter(const struct option *opt, const char *str,
-		 int unset __maybe_unused)
+static char *gen_filter(const char *str)
 {
-	struct perf_evlist *evlist = *(struct perf_evlist **)opt->value;
-	struct perf_evsel *last = NULL;
 	char *pid, buf[30], *o;
 	int len, plen;
 	const char *p;
-
-	if (evlist->nr_entries > 0)
-		last = perf_evlist__last(evlist);
-
-	if (last == NULL || last->attr.type != PERF_TYPE_TRACEPOINT) {
-		fprintf(stderr,
-			"-F option should follow a -e tracepoint option\n");
-		return -1;
-	}
+	char *filter;
 
 	plen = sprintf(buf, "%d", getpid());
 	len = strlen(str) + 1;
@@ -987,13 +976,13 @@ int parse_filter(const struct option *opt, const char *str,
 		len += plen - 8;
 	}
 
-	last->filter = malloc(len);
-	if (last->filter == NULL) {
+	filter = malloc(len);
+	if (filter == NULL) {
 		fprintf(stderr, "not enough memory to hold filter string\n");
-		return -1;
+		return NULL;
 	}
 
-	o = last->filter;
+	o = filter;
 	for (p = str; *p; ) {
 		if (*p == 'P' && !strncmp(p, "PERF_PID", 8)) {
 			strcpy(o, buf);
@@ -1002,6 +991,31 @@ int parse_filter(const struct option *opt, const char *str,
 		} else
 			*o++ = *p++;
 	}
+	return filter;
+}
+
+int parse_filter(const struct option *opt, const char *str,
+		 int unset __maybe_unused)
+{
+	struct perf_evlist *evlist = *(struct perf_evlist **)opt->value;
+	struct perf_evsel *evsel;
+	int nr_tp = 0;
+
+	evlist__for_each(evlist, evsel) {
+		if (evsel->attr.type == PERF_TYPE_TRACEPOINT &&
+			!evsel->filter) {
+			evsel->filter = gen_filter(str);
+			if (!evsel->filter)
+				return -1;
+			nr_tp++;
+		}
+	}
+
+	if (nr_tp == 0) {
+		fprintf(stderr,
+			"-F option should follow a -e tracepoint option\n");
+		return -1;
+	}
 	return 0;
 }
 
Signature · 1 line
1.9.3
#2 lore raw

Adding a filter to events (instead of replacing one) was Re: [PATCH 1/2] perf, tools: Add PERF_PID

In reply to Andi Kleen

Em Wed, Sep 24, 2014 at 01:51:08PM -0700, Andi Kleen escreveu:

Quoted 4 lines

It's currently difficult to filter out perf itself using a filter. This can give cascading effects during IO tracing when the IO perf does itself causes more trace output.

Quoted 8 lines

The best way to filter is to use the pid. But it's difficult to get the pid of perf without using hacks.

Add a PERF_PID meta variable to the perf filter that contains the current pid.

With this patch the following works

% perf record -e syscalls:sys_enter_write -a --filter 'common_pid != PERF_PID' ...

So I tried this one now and saw the other patch, that applies the --filter to all events, while trying I got:

[root@zoo ~]# perf record --filter "common_pid != PERF_PID" -a -F option should follow a -e tracepoint option

usage: perf record [<options>] [<command>]
or: perf record [<options>] -- <command> [<options>]

--filter <filter>
event filter
[root@zoo ~]#

A bug there, as '-F' is for frequency, will fix, but the message tells that --filter affects just the last event, i.e. whoever implemented this wanted to apply filters to specific events, and those need to be tracepoints (as we can't apply filters to other PERF_TYPE_ events (we should!).

So, more surgery would be needed to keep the existing behaviour, i.e. to allow per event filters, which is what the current does (or should, judging by the error message above), and to provide a new one that would apply the given filter to all suitable events, i.e. all tracepoint events (nowadays, others in the future).

I give a shot at implementing it in 'perf trace', by introducing a perf_evlist__filter_pid() that would then be used by default, but stopped working as to really make it good by default I would have to filter out activity on the gnome-terminal where 'perf trace' was running, to avoid another feedback loop.

But now I think that perf_evlist__filter_pid() thing is what we want here and then we should add a --filter-self option to 'perf record', that would just do:

	perf_evlist__filter_pid(evlist, getpid());

Agreed?

The patch is attached, but it is still incomplete, as I haven't checked if we set a filter we end up reseting whatever filter is in place, or if we have to use some syntax (prefixing the filter with '+'?) to add a filter instead of replacing what is there.

- Arnaldo

Quoted 73 lines

This won't work for more complex perf pipe lines with multiple processes, but at least solves the problem nicely for a single perf.

v2: Remove debug code. Don't use zero padding (Namhyung)
v3: Correct patch.
v4: Handle arbitary pid length.
v5: Fix a warning

---
 tools/perf/Documentation/perf-record.txt |  2 +-
 tools/perf/util/parse-events.c           | 21 ++++++++++++++++++++-
 2 files changed, 21 insertions(+), 2 deletions(-)
tools/perf/Documentation/perf-record.txt +1 -1
index d460049..b6c5e51 100644
@@ -41,7 +41,7 @@ OPTIONS
           'mem:0x1000:rw'.
 
 --filter=<filter>::
-        Event filter.
+        Event filter. PERF_PID represents the perf pid.
 
 -a::
 --all-cpus::
tools/perf/util/parse-events.c +20 -1
index 1e15df1..9c4bab8 100644
@@ -967,6 +967,9 @@ int parse_filter(const struct option *opt, const char *str,
 {
 	struct perf_evlist *evlist = *(struct perf_evlist **)opt->value;
 	struct perf_evsel *last = NULL;
+	char *pid, buf[30], *o;
+	int len, plen;
+	const char *p;
 
 	if (evlist->nr_entries > 0)
 		last = perf_evlist__last(evlist);
@@ -977,12 +980,28 @@ int parse_filter(const struct option *opt, const char *str,
 		return -1;
 	}
 
-	last->filter = strdup(str);
+	plen = sprintf(buf, "%d", getpid());
+	len = strlen(str) + 1;
+	for (pid = strstr(str, "PERF_PID"); pid;
+	     pid = strstr(pid + 1, "PERF_PID")) {
+		len += plen - 8;
+	}
+
+	last->filter = malloc(len);
 	if (last->filter == NULL) {
 		fprintf(stderr, "not enough memory to hold filter string\n");
 		return -1;
 	}
 
+	o = last->filter;
+	for (p = str; *p; ) {
+		if (*p == 'P' && !strncmp(p, "PERF_PID", 8)) {
+			strcpy(o, buf);
+			o += plen;
+			p += sizeof("PERF_PID") - 1;
+		} else
+			*o++ = *p++;
+	}
 	return 0;
 }
 
Signature · 1 line
1.9.3
tools/perf/builtin-trace.c +9 -0
index 09bcf2393910..d64e8c619ff6 100644
@@ -2105,6 +2105,10 @@ static int trace__run(struct trace *trace, int argc, const char **argv)
 	if (err < 0)
 		goto out_error_open;
 
+	err = perf_evlist__set_filter_pid(evlist, getpid());
+	if (err < 0)
+		goto out_error_filter;
+
 	err = perf_evlist__mmap(evlist, trace->opts.mmap_pages, false);
 	if (err < 0) {
 		fprintf(trace->output, "Couldn't mmap the events: %s\n",
@@ -2216,6 +2220,11 @@ out_error_open:
 out_error:
 	fprintf(trace->output, "%s\n", errbuf);
 	goto out_delete_evlist;
+
+out_error_filter:
+	fprintf(trace->output, "Couldn't set 'perf trace' pid filter: %s\n",
+                        strerror_r(errno, errbuf, sizeof(errbuf)));
+	goto out_delete_evlist;
 }
 }
 
tools/perf/util/evlist.c +21 -0
index 3cebc9a8d52e..06e54304f63f 100644
@@ -1041,6 +1041,27 @@ int perf_evlist__set_filter(struct perf_evlist *evlist, const char *filter)
 	return err;
 }
 
+int perf_evlist__set_filter_pid(struct perf_evlist *evlist, pid_t pid)
+{
+	struct perf_evsel *evsel;
+	int err = 0;
+	const int ncpus = cpu_map__nr(evlist->cpus),
+		  nthreads = thread_map__nr(evlist->threads);
+	char filter[64];
+
+	scnprintf(filter, sizeof(filter), "common_pid != %d", pid);
+
+	evlist__for_each(evlist, evsel) {
+		if (evsel->attr.type != PERF_TYPE_TRACEPOINT)
+			continue;
+		err = perf_evsel__set_filter(evsel, ncpus, nthreads, filter);
+		if (err)
+			break;
+	}
+
+	return err;
+}
+
 bool perf_evlist__valid_sample_type(struct perf_evlist *evlist)
 {
 	struct perf_evsel *pos;
tools/perf/util/evlist.h +1 -0
index bd312b01e876..9c4b9410d511 100644
@@ -77,6 +77,7 @@ int perf_evlist__add_newtp(struct perf_evlist *evlist,
 			   const char *sys, const char *name, void *handler);
 
 int perf_evlist__set_filter(struct perf_evlist *evlist, const char *filter);
+int perf_evlist__set_filter_pid(struct perf_evlist *evlist, pid_t pid);
 
 struct perf_evsel *
 perf_evlist__find_tracepoint_by_id(struct perf_evlist *evlist, int id);
#3 lore raw

Re: Adding a filter to events (instead of replacing one) was Re: [PATCH 1/2] perf, tools: Add PERF_PID

Quoted 9 lines · Arnaldo Carvalho de Melo

Em Wed, Sep 24, 2014 at 01:51:08PM -0700, Andi Kleen escreveu:

Quoted 6 lines

It's currently difficult to filter out perf itself using a filter. This can give cascading effects during IO tracing when the IO perf does itself causes more trace output.

The best way to filter is to use the pid. But it's difficult to get the pid of perf without using hacks.

Quoted 8 lines
Quoted 5 lines

Add a PERF_PID meta variable to the perf filter that contains the current pid.

With this patch the following works

% perf record -e syscalls:sys_enter_write -a --filter 'common_pid != PERF_PID' ...

So I tried this one now and saw the other patch, that applies the --filter to all events, while trying I got:

Patch seems reasonable to me.

However adding PERF_PID and sanitizing --filter are really two different things and should probably not be mixed in a patch.

-Andi

#4 lore raw

Re: Adding a filter to events (instead of replacing one) was Re: [PATCH 1/2] perf, tools: Add PERF_PID

Em Wed, Oct 01, 2014 at 03:02:18PM -0700, Andi Kleen escreveu:

Quoted 16 lines · Arnaldo Carvalho de Melo
Quoted 14 lines

Em Wed, Sep 24, 2014 at 01:51:08PM -0700, Andi Kleen escreveu:

Quoted 12 lines

It's currently difficult to filter out perf itself using a filter. This can give cascading effects during IO tracing when the IO perf does itself causes more trace output.

The best way to filter is to use the pid. But it's difficult to get the pid of perf without using hacks.

Add a PERF_PID meta variable to the perf filter that contains the current pid.

With this patch the following works

% perf record -e syscalls:sys_enter_write -a --filter 'common_pid != PERF_PID' ...

Quoted 7 lines
Quoted 2 lines

So I tried this one now and saw the other patch, that applies the --filter to all events, while trying I got:

Patch seems reasonable to me.

However adding PERF_PID and sanitizing --filter are really two different things and should probably not be mixed in a patch.

Yes, there are two things, but what seems to be wanted first is a way to exclude 'perf record' samples, for all events.

Being able to specify PERF_PID on a filter so that some events can include 'perf record' samples and some not seems to be something not really needed at this point, i.e. just doing:

perf record --filter-self -e syscalls:sys_enter_write -a ...

Is shorter and doesn't breaks the current --filter semantic (apply just to the last tracepoint informed in the cmdline, not to all tracepoints) like in your second patch.

So probably what is best to do is to finish the patch I sent here, which will cover the filter-out-perf-tooling-samples usecase, and then, if people still think it is needed, introduce the PERF_PID meta filter variable.

- Arnaldo