Skip to contentSkip to conversation

problem with 'perf script -F weight'

PATCH Jiri Olsa 3 participants 5 messages Sep 28, 2021 – Sep 29, 2021 perf events

lore
Share & tools
lore

Jiri Olsa

author
#0 lore raw

hi,
Joe reported broken -F weight in perf script with (on x86):

  # ./perf mem record 
  # ./perf script -F weight
  Samples for 'dummy:HG' event do not have WEIGHT attribute set. Cannot print 'weight' field.

the problem seems to be introduced with the WEIGHT_STRUCT change:
ea8d0ed6eae3 perf tools: Support PERF_SAMPLE_WEIGHT_STRUCT

which enables WEIGHT_STRUCT sample_type on x86 instead of WEIGHT, and leaves 'perf script -F weight' in the dark

I'm not sure what the fix should look like.. should we allow user to use just 'WEIGHT' sample type? or is there a way to get original weight from 'WEIGHT_STRUCT' data, so we could fix just perf script?

FYI when I switch back the WEIGHT_STRUCT to WEIGHT again it seems to work as before.. as expected ;-)

any thoughts? thanks
jirka

---

tools/perf/arch/x86/util/evsel.c +1 -1
index 2f733cdc8dbb..6bdd744894ad 100644
@@ -4,5 +4,5 @@
 
 void arch_evsel__set_sample_weight(struct evsel *evsel)
 {
-	evsel__set_sample_bit(evsel, WEIGHT_STRUCT);
+	evsel__set_sample_bit(evsel, WEIGHT);
 }
#1 lore raw
Quoted 10 lines · Jiri Olsa

hi,
Joe reported broken -F weight in perf script with (on x86):

   # ./perf mem record
   # ./perf script -F weight
   Samples for 'dummy:HG' event do not have WEIGHT attribute set. Cannot print 'weight' field.

the problem seems to be introduced with the WEIGHT_STRUCT change:
ea8d0ed6eae3 perf tools: Support PERF_SAMPLE_WEIGHT_STRUCT

Quoted 6 lines

which enables WEIGHT_STRUCT sample_type on x86 instead of WEIGHT, and leaves 'perf script -F weight' in the dark

I'm not sure what the fix should look like.. should we allow user to use just 'WEIGHT' sample type? or is there a way to get original weight from 'WEIGHT_STRUCT' data, so we could fix just perf script?

We would need to support both in perf script, to handle both perf.data generated by old and new kernels on both big and little endian

So probably try one first and then the other while parsing

The same is true for all other users, so probably there needs to be a library function for this somewhere.

-Andi

#2 lore raw
In reply to Jiri Olsa
Quoted 11 lines · Jiri Olsa

Joe reported broken -F weight in perf script with (on x86):

   # ./perf mem record
   # ./perf script -F weight
   Samples for 'dummy:HG' event do not have WEIGHT attribute set. Cannot print 'weight' field.

the problem seems to be introduced with the WEIGHT_STRUCT change:
ea8d0ed6eae3 perf tools: Support PERF_SAMPLE_WEIGHT_STRUCT

which enables WEIGHT_STRUCT sample_type on x86 instead of WEIGHT, and leaves 'perf script -F weight' in the dark

Right, I missed the support for the perf script. Because the default perf script worked. It didn't ring the bell. :(

Quoted 4 lines

I'm not sure what the fix should look like.. should we allow user to use just 'WEIGHT' sample type? or is there a way to get original weight from 'WEIGHT_STRUCT' data, so we could fix just perf script?

Just fix the perf script is enough.
Could you please try the below patch?

(I probably need to split the patch into two patches, since it fixed two issues, one is this error, the other is to print the instruction latency. Anyway, let's try it first.)

From 99443f49a6236ef6e3bea3930792967a7863f753 Mon Sep 17 00:00:00 2001
From: Kan Liang <[email protected]>
Date: Tue, 28 Sep 2021 13:57:18 -0700
Subject: [PATCH] perf script: Fix PERF_SAMPLE_WEIGHT_STRUCT support

-F weight in perf script is broken.

  # ./perf mem record
  # ./perf script -F weight
  Samples for 'dummy:HG' event do not have WEIGHT attribute set. Cannot
print 'weight' field.

The sample type, PERF_SAMPLE_WEIGHT_STRUCT, is an alternative of the PERF_SAMPLE_WEIGHT sample type. They share the same space, weight. The lower 32 bits are exactly the same for both sample type. The higher 32 bits may be different for different architecture. For a new kernel on x86, the PERF_SAMPLE_WEIGHT_STRUCT is used. For an old kernel or other ARCHs, the PERF_SAMPLE_WEIGHT is used.

With -F weight, current perf script will check the input string "weight" with the corresponding sample type. However, the commit ID ea8d0ed6eae3 ("perf tools: Support PERF_SAMPLE_WEIGHT_STRUCT") didn't update the sample type for perf script. The check fails with a new kernel on x86.

Use PERF_SAMPLE_WEIGHT_TYPE, which supports both sample types, to replace PERF_SAMPLE_WEIGHT.

Also, print the instruction latency for PERF_SAMPLE_WEIGHT_STRUCT type.

Fixes ea8d0ed6eae3 ("perf tools: Support PERF_SAMPLE_WEIGHT_STRUCT")
---
 tools/perf/builtin-script.c | 7 +++++--
 1 file changed, 5 insertions(+), 2 deletions(-)
tools/perf/builtin-script.c +5 -2
index 6211d0b..fd73bbc 100644
@@ -459,7 +459,7 @@ static int evsel__check_attr(struct evsel *evsel, struct perf_session *session)
 		return -EINVAL;
 
 	if (PRINT_FIELD(WEIGHT) &&
-	    evsel__check_stype(evsel, PERF_SAMPLE_WEIGHT, "WEIGHT", PERF_OUTPUT_WEIGHT))
+	    evsel__check_stype(evsel, PERF_SAMPLE_WEIGHT_TYPE, "WEIGHT", PERF_OUTPUT_WEIGHT))
 		return -EINVAL;
 
 	if (PRINT_FIELD(SYM) &&
@@ -2036,8 +2036,11 @@ static void process_event(struct perf_script *script,
 	if (PRINT_FIELD(DATA_SRC))
 		data_src__fprintf(sample->data_src, fp);
 
-	if (PRINT_FIELD(WEIGHT))
+	if (PRINT_FIELD(WEIGHT)) {
 		fprintf(fp, "%16" PRIu64, sample->weight);
+		if (attr->sample_type & PERF_SAMPLE_WEIGHT_STRUCT)
+			fprintf(fp, ",0x%" PRIx16, sample->ins_lat);
+	}
 
 	if (PRINT_FIELD(IP)) {
 		struct callchain_cursor *cursor = NULL;
Signature · 4 lines
2.7.4

Thanks,
Kan
author
#3 lore raw
Quoted 83 lines · Kan
Quoted 11 lines · Jiri Olsa

Joe reported broken -F weight in perf script with (on x86):

   # ./perf mem record
   # ./perf script -F weight
   Samples for 'dummy:HG' event do not have WEIGHT attribute set. Cannot print 'weight' field.

the problem seems to be introduced with the WEIGHT_STRUCT change:
ea8d0ed6eae3 perf tools: Support PERF_SAMPLE_WEIGHT_STRUCT

which enables WEIGHT_STRUCT sample_type on x86 instead of WEIGHT, and leaves 'perf script -F weight' in the dark

Right, I missed the support for the perf script. Because the default perf script worked. It didn't ring the bell. :(

Quoted 4 lines

I'm not sure what the fix should look like.. should we allow user to use just 'WEIGHT' sample type? or is there a way to get original weight from 'WEIGHT_STRUCT' data, so we could fix just perf script?

Just fix the perf script is enough.
Could you please try the below patch?

(I probably need to split the patch into two patches, since it fixed two issues, one is this error, the other is to print the instruction latency. Anyway, let's try it first.)

From 99443f49a6236ef6e3bea3930792967a7863f753 Mon Sep 17 00:00:00 2001
From: Kan Liang <[email protected]>
Date: Tue, 28 Sep 2021 13:57:18 -0700
Subject: [PATCH] perf script: Fix PERF_SAMPLE_WEIGHT_STRUCT support

-F weight in perf script is broken.

  # ./perf mem record
  # ./perf script -F weight
  Samples for 'dummy:HG' event do not have WEIGHT attribute set. Cannot
print 'weight' field.

The sample type, PERF_SAMPLE_WEIGHT_STRUCT, is an alternative of the PERF_SAMPLE_WEIGHT sample type. They share the same space, weight. The lower 32 bits are exactly the same for both sample type. The higher 32 bits may be different for different architecture. For a new kernel on x86, the PERF_SAMPLE_WEIGHT_STRUCT is used. For an old kernel or other ARCHs, the PERF_SAMPLE_WEIGHT is used.

With -F weight, current perf script will check the input string "weight" with the corresponding sample type. However, the commit ID ea8d0ed6eae3 ("perf tools: Support PERF_SAMPLE_WEIGHT_STRUCT") didn't update the sample type for perf script. The check fails with a new kernel on x86.

Use PERF_SAMPLE_WEIGHT_TYPE, which supports both sample types, to replace PERF_SAMPLE_WEIGHT.

Also, print the instruction latency for PERF_SAMPLE_WEIGHT_STRUCT type.

Fixes ea8d0ed6eae3 ("perf tools: Support PERF_SAMPLE_WEIGHT_STRUCT")
---
 tools/perf/builtin-script.c | 7 +++++--
 1 file changed, 5 insertions(+), 2 deletions(-)
tools/perf/builtin-script.c +1 -1
index 6211d0b..fd73bbc 100644
@@ -459,7 +459,7 @@ static int evsel__check_attr(struct evsel *evsel, struct
 perf_session *session)
 		return -EINVAL;
 
 	if (PRINT_FIELD(WEIGHT) &&
-	    evsel__check_stype(evsel, PERF_SAMPLE_WEIGHT, "WEIGHT",
 PERF_OUTPUT_WEIGHT))
+	    evsel__check_stype(evsel, PERF_SAMPLE_WEIGHT_TYPE, "WEIGHT",
 PERF_OUTPUT_WEIGHT))
tools/perf/builtin-script.c +0 -0
 		return -EINVAL;
 
 	if (PRINT_FIELD(SYM) &&

tools/perf/builtin-script.c static void process_event(struct perf_script *script, · line 2042

Quoted 8 lines
tools/perf/builtin-script.c +3 -1
@@ -2036,8 +2036,11 @@ static void process_event(struct perf_script *script,
 	if (PRINT_FIELD(DATA_SRC))
 		data_src__fprintf(sample->data_src, fp);
 
-	if (PRINT_FIELD(WEIGHT))
+	if (PRINT_FIELD(WEIGHT)) {
 		fprintf(fp, "%16" PRIu64, sample->weight);
+		if (attr->sample_type & PERF_SAMPLE_WEIGHT_STRUCT)
+			fprintf(fp, ",0x%" PRIx16, sample->ins_lat);

this will add extra number to weight.. should we add separate column for that -F 'insn-lat' ? users expect just single value for -F weight

thanks,
jirka

Quoted 10 lines
tools/perf/builtin-script.c +1 -0
+	}
 
 	if (PRINT_FIELD(IP)) {
 		struct callchain_cursor *cursor = NULL;
Signature · 4 lines
2.7.4

Thanks,
Kan
#4 lore raw
Quoted 87 lines · Jiri Olsa
Quoted 86 lines · Kan
Quoted 11 lines · Jiri Olsa

Joe reported broken -F weight in perf script with (on x86):

    # ./perf mem record
    # ./perf script -F weight
    Samples for 'dummy:HG' event do not have WEIGHT attribute set. Cannot print 'weight' field.

the problem seems to be introduced with the WEIGHT_STRUCT change:
ea8d0ed6eae3 perf tools: Support PERF_SAMPLE_WEIGHT_STRUCT

which enables WEIGHT_STRUCT sample_type on x86 instead of WEIGHT, and leaves 'perf script -F weight' in the dark

Right, I missed the support for the perf script. Because the default perf script worked. It didn't ring the bell. :(

Quoted 4 lines

I'm not sure what the fix should look like.. should we allow user to use just 'WEIGHT' sample type? or is there a way to get original weight from 'WEIGHT_STRUCT' data, so we could fix just perf script?

Just fix the perf script is enough.
Could you please try the below patch?

(I probably need to split the patch into two patches, since it fixed two issues, one is this error, the other is to print the instruction latency. Anyway, let's try it first.)

From 99443f49a6236ef6e3bea3930792967a7863f753 Mon Sep 17 00:00:00 2001
From: Kan Liang <[email protected]>
Date: Tue, 28 Sep 2021 13:57:18 -0700
Subject: [PATCH] perf script: Fix PERF_SAMPLE_WEIGHT_STRUCT support

-F weight in perf script is broken.

   # ./perf mem record
   # ./perf script -F weight
   Samples for 'dummy:HG' event do not have WEIGHT attribute set. Cannot
print 'weight' field.

The sample type, PERF_SAMPLE_WEIGHT_STRUCT, is an alternative of the PERF_SAMPLE_WEIGHT sample type. They share the same space, weight. The lower 32 bits are exactly the same for both sample type. The higher 32 bits may be different for different architecture. For a new kernel on x86, the PERF_SAMPLE_WEIGHT_STRUCT is used. For an old kernel or other ARCHs, the PERF_SAMPLE_WEIGHT is used.

With -F weight, current perf script will check the input string "weight" with the corresponding sample type. However, the commit ID ea8d0ed6eae3 ("perf tools: Support PERF_SAMPLE_WEIGHT_STRUCT") didn't update the sample type for perf script. The check fails with a new kernel on x86.

Use PERF_SAMPLE_WEIGHT_TYPE, which supports both sample types, to replace PERF_SAMPLE_WEIGHT.

Also, print the instruction latency for PERF_SAMPLE_WEIGHT_STRUCT type.

Fixes ea8d0ed6eae3 ("perf tools: Support PERF_SAMPLE_WEIGHT_STRUCT")
---
  tools/perf/builtin-script.c | 7 +++++--
  1 file changed, 5 insertions(+), 2 deletions(-)
tools/perf/builtin-script.c +1 -1
index 6211d0b..fd73bbc 100644
@@ -459,7 +459,7 @@ static int evsel__check_attr(struct evsel *evsel, struct
 perf_session *session)
  		return -EINVAL;
 
  	if (PRINT_FIELD(WEIGHT) &&
-	    evsel__check_stype(evsel, PERF_SAMPLE_WEIGHT, "WEIGHT",
 PERF_OUTPUT_WEIGHT))
+	    evsel__check_stype(evsel, PERF_SAMPLE_WEIGHT_TYPE, "WEIGHT",
 PERF_OUTPUT_WEIGHT))

return -EINVAL;

if (PRINT_FIELD(SYM) &&

tools/perf/builtin-script.c +0 -0
@@ -2036,8 +2036,11 @@ static void process_event(struct perf_script *script,
  	if (PRINT_FIELD(DATA_SRC))
  		data_src__fprintf(sample->data_src, fp);
 
Quoted 9 lines

tools/perf/builtin-script.c static void process_event(struct perf_script *script, · line 2042

Quoted 5 lines
tools/perf/builtin-script.c +3 -1
-	if (PRINT_FIELD(WEIGHT))
+	if (PRINT_FIELD(WEIGHT)) {
  		fprintf(fp, "%16" PRIu64, sample->weight);
+		if (attr->sample_type & PERF_SAMPLE_WEIGHT_STRUCT)
+			fprintf(fp, ",0x%" PRIx16, sample->ins_lat);

this will add extra number to weight.. should we add separate column for that -F 'insn-lat' ? users expect just single value for -F weight

Sure, I will send out a patch set to LKML for review shortly.

Thanks,
Kan