Skip to contentSkip to conversation

perf stat: fix cvs output format

PATCH Applied Cong Wang 6 participants 14 messages Mar 6, 2018 – Mar 9, 2018 perf events

lore
Share & tools
lore

Cong Wang

author
#0 lore raw

PATCH 1/1 perf stat: fix cvs output format

From: Ilya Pronin <[email protected]>

When printing stats in CSV mode, perf stat appends extra CSV separators when counter is not supported:

<not supported>,,L1-dcache-store-misses,mesos/bd442f34-2b4a-47df-b966-9b281f9f56fc,0,100.00,,,,

which causes a failure of parsing fields. The numbers of separators is fixed for each line, no matter supported or not supported.

Fixes 92a61f6412d3 ("perf stat: Implement CSV metrics output")
Cc Andi Kleen <[email protected]>
Cc Arnaldo Carvalho de Melo <[email protected]>
Cc Jiri Olsa <[email protected]>
---
 tools/perf/builtin-stat.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
tools/perf/builtin-stat.c +1 -1
index 98bf9d32f222..54a4c152edb3 100644
@@ -917,7 +917,7 @@ static void print_metric_csv(void *ctx,
 	char buf[64], *vals, *ends;
 
 	if (unit == NULL || fmt == NULL) {
-		fprintf(out, "%s%s%s%s", csv_sep, csv_sep, csv_sep, csv_sep);
+		fprintf(out, "%s%s", csv_sep, csv_sep);
 		return;
 	}
 	snprintf(buf, sizeof(buf), fmt, val);
Signature · 1 line
2.13.0
Acked-by on v1 1/1
#1 lore raw
Quoted 25 lines · Cong Wang

From: Ilya Pronin <[email protected]>

When printing stats in CSV mode, perf stat appends extra CSV separators when counter is not supported:

<not supported>,,L1-dcache-store-misses,mesos/bd442f34-2b4a-47df-b966-9b281f9f56fc,0,100.00,,,,

which causes a failure of parsing fields. The numbers of separators is fixed for each line, no matter supported or not supported.

Fixes 92a61f6412d3 ("perf stat: Implement CSV metrics output")
Cc Andi Kleen <[email protected]>
Cc Arnaldo Carvalho de Melo <[email protected]>
Cc Jiri Olsa <[email protected]>
---
 tools/perf/builtin-stat.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
tools/perf/builtin-stat.c +0 -0
index 98bf9d32f222..54a4c152edb3 100644

tools/perf/builtin-stat.c static void print_metric_csv(void *ctx, · line 922

Quoted 7 lines
tools/perf/builtin-stat.c +1 -1
@@ -917,7 +917,7 @@ static void print_metric_csv(void *ctx,
 	char buf[64], *vals, *ends;
 
 	if (unit == NULL || fmt == NULL) {
-		fprintf(out, "%s%s%s%s", csv_sep, csv_sep, csv_sep, csv_sep);
+		fprintf(out, "%s%s", csv_sep, csv_sep);
 		return;
 	}

right, the non else legs prints just 2 values:
fprintf(out, "%s%s%s%s", csv_sep, vals, csv_sep, unit);

thanks,
jirka

#2 lore raw

Em Tue, Mar 06, 2018 at 08:58:46AM +0100, Jiri Olsa escreveu:

Quoted 37 lines · Cong Wang
Quoted 25 lines

From: Ilya Pronin <[email protected]>

When printing stats in CSV mode, perf stat appends extra CSV separators when counter is not supported:

<not supported>,,L1-dcache-store-misses,mesos/bd442f34-2b4a-47df-b966-9b281f9f56fc,0,100.00,,,,

which causes a failure of parsing fields. The numbers of separators is fixed for each line, no matter supported or not supported.

Fixes 92a61f6412d3 ("perf stat: Implement CSV metrics output")
Cc Andi Kleen <[email protected]>
Cc Arnaldo Carvalho de Melo <[email protected]>
Cc Jiri Olsa <[email protected]>
---
 tools/perf/builtin-stat.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
tools/perf/builtin-stat.c +0 -0
index 98bf9d32f222..54a4c152edb3 100644

tools/perf/builtin-stat.c static void print_metric_csv(void *ctx, · line 922

Quoted 7 lines
tools/perf/builtin-stat.c +1 -1
@@ -917,7 +917,7 @@ static void print_metric_csv(void *ctx,
 	char buf[64], *vals, *ends;
 
 	if (unit == NULL || fmt == NULL) {
-		fprintf(out, "%s%s%s%s", csv_sep, csv_sep, csv_sep, csv_sep);
+		fprintf(out, "%s%s", csv_sep, csv_sep);
 		return;
 	}

right, the non else legs prints just 2 values:
fprintf(out, "%s%s%s%s", csv_sep, vals, csv_sep, unit);

Quoted 2 lines

Thanks, applied to perf/urgent.

- Arnaldo

#3 lore raw
In reply to Cong Wang
Quoted 8 lines · Cong Wang

From: Ilya Pronin <[email protected]>

When printing stats in CSV mode, perf stat appends extra CSV separators when counter is not supported:

<not supported>,,L1-dcache-store-misses,mesos/bd442f34-2b4a-47df-b966-9b281f9f56fc,0,100.00,,,,

which causes a failure of parsing fields. The numbers of separators

Causes failure in what?

Quoted 1 line

is fixed for each line, no matter supported or not supported.

I don't think they're extra fields, there are cases where they can be filled out for variance, metricvalue, unit. And other code in perf too uses empty fields when something is not available.

- optional usec time stamp in fractions of second (with -I xxx)
- optional CPU, core, or socket identifier
- optional number of logical CPUs aggregated
- counter value
- unit of the counter value or empty
- event name
- run time of counter
- percentage of measurement time the counter was running
- optional variance if multiple values are collected with -r
- optional metric value
- optional unit of metric

-Andi

#4 lore raw

Em Tue, Mar 06, 2018 at 09:00:11AM -0800, Andi Kleen escreveu:

Quoted 8 lines · Cong Wang
Quoted 7 lines

From: Ilya Pronin <[email protected]>

When printing stats in CSV mode, perf stat appends extra CSV separators when counter is not supported:

<not supported>,,L1-dcache-store-misses,mesos/bd442f34-2b4a-47df-b966-9b281f9f56fc,0,100.00,,,,

Quoted 7 lines
Quoted 1 line

which causes a failure of parsing fields. The numbers of separators

Causes failure in what?

Quoted 1 line

is fixed for each line, no matter supported or not supported.

I don't think they're extra fields, there are cases where they can be filled out

My understanding was that at some place there is a if/else

if (supported counters)
fprintf_something with N fields, all filled in
else
fprintf_empty_fields with != N fields

So I think this is not about using things like 'a,b,,,,,,' but about using different number of commas (fields) for supported/unsupported counters, no?

- Arnaldo

Quoted 17 lines

for variance, metricvalue, unit. And other code in perf too uses empty fields when something is not available.

- optional usec time stamp in fractions of second (with -I xxx)
- optional CPU, core, or socket identifier
- optional number of logical CPUs aggregated
- counter value
- unit of the counter value or empty
- event name
- run time of counter
- percentage of measurement time the counter was running
- optional variance if multiple values are collected with -r
- optional metric value
- optional unit of metric

-Andi

#5 lore raw
Quoted 10 lines

My understanding was that at some place there is a if/else

if (supported counters)
fprintf_something with N fields, all filled in
else
fprintf_empty_fields with != N fields

So I think this is not about using things like 'a,b,,,,,,' but about using different number of commas (fields) for supported/unsupported counters, no?

I believe it's only about empty fields at the end. I don't think we ever get the columns wrong.

The original patch just moved the problem because there are still cases where we can output different number of columns.

If a tool looks only at the first row to allocate the number of columns it might error out if there are lines with more columns later.

All outputs have to be padded to the maximum number of columns, so removing columns is never the right fix.

-Andi

author
#6 lore raw
In reply to Andi Kleen
Quoted 11 lines · Andi Kleen
Quoted 8 lines · Cong Wang

From: Ilya Pronin <[email protected]>

When printing stats in CSV mode, perf stat appends extra CSV separators when counter is not supported:

<not supported>,,L1-dcache-store-misses,mesos/bd442f34-2b4a-47df-b966-9b281f9f56fc,0,100.00,,,,

which causes a failure of parsing fields. The numbers of separators

Causes failure in what?

Failed to know how many fields in that line, clearly there are less separators when it is supported.

Quoted 6 lines
Quoted 2 lines

is fixed for each line, no matter supported or not supported.

I don't think they're extra fields, there are cases where they can be filled out for variance, metricvalue, unit. And other code in perf too uses empty fields when something is not available.

Are you saying there should be more fields when it is not supported?

Here is the output from your own commit:

      423470,,stalled-cycles-frontend,509102,100.00,65.69,frontend cycles idle
      <not supported>,,stalled-cycles-backend,0,100.00,,,,

so line 1 has 7 fields, line 2 has 9 fields, and this is expected?

#7 lore raw
Quoted 6 lines

Here is the output from your own commit:

      423470,,stalled-cycles-frontend,509102,100.00,65.69,frontend cycles idle
      <not supported>,,stalled-cycles-backend,0,100.00,,,,

so line 1 has 7 fields, line 2 has 9 fields, and this is expected?

If you had metrics on line 1 it would be correct.

So you just shifted it to break that case.

If you always want to have the same number of fields you need to add two empty fields to the normal output when there are no metrics.

-Andi

author
#8 lore raw
Quoted 7 lines · Andi Kleen
Quoted 6 lines

Here is the output from your own commit:

      423470,,stalled-cycles-frontend,509102,100.00,65.69,frontend cycles idle
      <not supported>,,stalled-cycles-backend,0,100.00,,,,

so line 1 has 7 fields, line 2 has 9 fields, and this is expected?

Quoted 7 lines

If you had metrics on line 1 it would be correct.

So you just shifted it to break that case.

If you always want to have the same number of fields you need to add two empty fields to the normal output when there are no metrics.

The number of separators is the only way to learn the number of fields, therefore it must be a fixed number.

Yeah, it could be the other way that supported ones have less separators than it should. If we look at print_metric_csv() alone, it should produce a same number of separators for all cases, otherwise hard to count.

So I believe we need an additional patch, like the one attached, to make it complete? Note, I only spot the cgroup field here.

tools/perf/builtin-stat.c +7 -8
index 54a4c152edb3..6896e739ae4e 100644
@@ -1044,8 +1044,7 @@ static void nsec_printout(int id, int nr, struct perf_evsel *evsel, double avg)
 
 	fprintf(output, fmt_n, name);
 
-	if (evsel->cgrp)
-		fprintf(output, "%s%s", csv_sep, evsel->cgrp->name);
+	fprintf(output, "%s%s", csv_sep, evsel->cgrp ? evsel->cgrp->name : "");
 }
 
 static int first_shadow_cpu(struct perf_evsel *evsel, int id)
@@ -1092,12 +1091,13 @@ static void abs_printout(int id, int nr, struct perf_evsel *evsel, double avg)
 	if (evsel->unit)
 		fprintf(output, "%-*s%s",
 			csv_output ? 0 : unit_width,
-			evsel->unit, csv_sep);
+			evsel->unit ? evsel->unit : "", csv_sep);
+	else
+		fprintf(output, "%s", csv_sep);
 
 	fprintf(output, "%-*s", csv_output ? 0 : 25, perf_evsel__name(evsel));
 
-	if (evsel->cgrp)
-		fprintf(output, "%s%s", csv_sep, evsel->cgrp->name);
+	fprintf(output, "%s%s", csv_sep, evsel->cgrp ? evsel->cgrp->name : "");
 }
 
 static void printout(int id, int nr, struct perf_evsel *counter, double uval,
@@ -1163,9 +1163,8 @@ static void printout(int id, int nr, struct perf_evsel *counter, double uval,
 			csv_output ? 0 : -25,
 			perf_evsel__name(counter));
 
-		if (counter->cgrp)
-			fprintf(stat_config.output, "%s%s",
-				csv_sep, counter->cgrp->name);
+		fprintf(stat_config.output, "%s%s", csv_sep,
+			counter->cgrp ? counter->cgrp->name : "");
 
 		if (!csv_output)
 			pm(&os, NULL, NULL, "", 0);
#9 lore raw

Speaking from the user's seat. An optional (not just empty) cgroup field is fine as long it consistently appears when requested with -G option. The problem with print_metric_csv() was that in the case of unsupported counters 2 additional empty fields in the output are completely unexpected and not documented anywhere.

Andi, in the output example in your commit 92a61f6412d3a09d6462252a522fa79c9290f405 stalled-cycles-backend event has counter run time field, counter run time percentage field, empty metric value, empty metric unit, and then 2 other empty fields. Are they expected? If yes, what are they and why other events, e.g. stalled-cycles-frontend, don't have them? Am I missing something here?

Quoted 26 lines · Cong Wang
Quoted 7 lines · Andi Kleen
Quoted 6 lines

Here is the output from your own commit:

      423470,,stalled-cycles-frontend,509102,100.00,65.69,frontend cycles idle
      <not supported>,,stalled-cycles-backend,0,100.00,,,,

so line 1 has 7 fields, line 2 has 9 fields, and this is expected?

Quoted 7 lines

If you had metrics on line 1 it would be correct.

So you just shifted it to break that case.

If you always want to have the same number of fields you need to add two empty fields to the normal output when there are no metrics.

The number of separators is the only way to learn the number of fields, therefore it must be a fixed number.

Yeah, it could be the other way that supported ones have less separators than it should. If we look at print_metric_csv() alone, it should produce a same number of separators for all cases, otherwise hard to count.

So I believe we need an additional patch, like the one attached, to make it complete? Note, I only spot the cgroup field here.

Signature · 1 line
Ilya Pronin
#10 lore raw
Quoted 11 lines · Ilya Pronin

Speaking from the user's seat. An optional (not just empty) cgroup field is fine as long it consistently appears when requested with -G option. The problem with print_metric_csv() was that in the case of unsupported counters 2 additional empty fields in the output are completely unexpected and not documented anywhere.

Andi, in the output example in your commit 92a61f6412d3a09d6462252a522fa79c9290f405 stalled-cycles-backend event has counter run time field, counter run time percentage field, empty metric value, empty metric unit, and then 2 other empty fields. Are they expected? If yes, what are they and why other events, e.g.

No two extra empty fields are not expected. All lines should have the same number of fields so that a tool that looks at the first like can keep using the same number.

But I don't think that was it what the patch fixed, or did I misread it?

-Andi

#11 lore raw

The patch merely eliminated those 2 unexpected fields. But from what I saw they were violating the assumption that all lines have the same number of fields.

Quoted 21 lines · Andi Kleen
Quoted 11 lines · Ilya Pronin

Speaking from the user's seat. An optional (not just empty) cgroup field is fine as long it consistently appears when requested with -G option. The problem with print_metric_csv() was that in the case of unsupported counters 2 additional empty fields in the output are completely unexpected and not documented anywhere.

Andi, in the output example in your commit 92a61f6412d3a09d6462252a522fa79c9290f405 stalled-cycles-backend event has counter run time field, counter run time percentage field, empty metric value, empty metric unit, and then 2 other empty fields. Are they expected? If yes, what are they and why other events, e.g.

No two extra empty fields are not expected. All lines should have the same number of fields so that a tool that looks at the first like can keep using the same number.

But I don't think that was it what the patch fixed, or did I misread it?

-Andi

Signature · 1 line
Ilya
#12 lore raw
Quoted 3 lines · Ilya Pronin

The patch merely eliminated those 2 unexpected fields. But from what I saw they were violating the assumption that all lines have the same number of fields.

Ok then I must have misread the patch. Sorry about that. It's fine for me then. I think Arnaldo has it already merged anyways.

-Andi

#13 lore raw

[tip:perf/urgent] perf stat: Fix CVS output format for non-supported counters

In reply to Cong Wang
perf stat: Fix CVS output format for non-supported counters
Commit-ID:  40c21898ba5372c14ef71717040529794a91ccc2
Gitweb:     https://git.kernel.org/tip/40c21898ba5372c14ef71717040529794a91ccc2
Author:     Ilya Pronin <[email protected]>
AuthorDate: Mon, 5 Mar 2018 22:43:53 -0800
Committer:  Arnaldo Carvalho de Melo <[email protected]>
CommitDate: Tue, 6 Mar 2018 10:53:52 -0300

perf stat: Fix CVS output format for non-supported counters

When printing stats in CSV mode, 'perf stat' appends extra separators when a counter is not supported:

<not supported>,,L1-dcache-store-misses,mesos/bd442f34-2b4a-47df-b966-9b281f9f56fc,0,100.00,,,,

Which causes a failure when parsing fields. The numbers of separators should be the same for each line, no matter if the counter is or not supported.

---
 tools/perf/builtin-stat.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
tools/perf/builtin-stat.c +1 -1
index 98bf9d32f222..54a4c152edb3 100644
@@ -917,7 +917,7 @@ static void print_metric_csv(void *ctx,
 	char buf[64], *vals, *ends;
 
 	if (unit == NULL || fmt == NULL) {
-		fprintf(out, "%s%s%s%s", csv_sep, csv_sep, csv_sep, csv_sep);
+		fprintf(out, "%s%s", csv_sep, csv_sep);
 		return;
 	}
 	snprintf(buf, sizeof(buf), fmt, val);