Skip to contentSkip to conversation

perf script: Skip dummy event attr check

PATCH Applied Jiri Olsa 6 participants 10 messages Aug 31, 2022 – Sep 8, 2022 perf events

lore
Share & tools
lore

Jiri Olsa

author reviewer
#0 lore raw

PATCH 1/1 perf script: Skip dummy event attr check

Hongtao Yu reported problem when displaying uregs in perf script for system wide perf.data:

  # perf script -F uregs | head -10
  Samples for 'dummy:HG' event do not have UREGS attribute set. Cannot print 'uregs' field.

The problem is the extra dummy event added for system wide, which does not have proper sample_type setup.

Skipping attr check completely for dummy event as suggested by Namhyung, because it does not have any samples anyway.

---
 tools/perf/builtin-script.c | 2 ++
 1 file changed, 2 insertions(+)
tools/perf/builtin-script.c +2 -0
index 13580a9c50b8..304d234d8e84 100644
@@ -566,6 +566,8 @@ static struct evsel *find_first_output_type(struct evlist *evlist,
 	struct evsel *evsel;
 
 	evlist__for_each_entry(evlist, evsel) {
+		if (evsel__is_dummy_event(evsel))
+			continue;
 		if (output_type(evsel->core.attr.type) == (int)type)
 			return evsel;
 	}
Signature · 1 line
2.37.2
reviewer Acked-by on v1 1/1
#1 lore raw
Quoted 10 lines · Jiri Olsa

Hongtao Yu reported problem when displaying uregs in perf script for system wide perf.data:

  # perf script -F uregs | head -10
  Samples for 'dummy:HG' event do not have UREGS attribute set. Cannot print 'uregs' field.

The problem is the extra dummy event added for system wide, which does not have proper sample_type setup.

Quoted 6 lines

Skipping attr check completely for dummy event as suggested by Namhyung, because it does not have any samples anyway.

Thanks,
Ian

Quoted 20 lines
---
 tools/perf/builtin-script.c | 2 ++
 1 file changed, 2 insertions(+)
tools/perf/builtin-script.c +2 -1
index 13580a9c50b8..304d234d8e84 100644
@@ -566,6 +566,8 @@ static struct evsel *find_first_output_type(struct evlist *evlist,
        struct evsel *evsel;
 
        evlist__for_each_entry(evlist, evsel) {
+               if (evsel__is_dummy_event(evsel))
+                       continue;
                if (output_type(evsel->core.attr.type) == (int)type)
                        return evsel;
        }
--

2.37.2

maintainer Applied
#2 lore raw

Em Wed, Aug 31, 2022 at 09:02:46AM -0700, Ian Rogers escreveu:

Quoted 11 lines · Jiri Olsa
Quoted 10 lines

Hongtao Yu reported problem when displaying uregs in perf script for system wide perf.data:

  # perf script -F uregs | head -10
  Samples for 'dummy:HG' event do not have UREGS attribute set. Cannot print 'uregs' field.

The problem is the extra dummy event added for system wide, which does not have proper sample_type setup.

Quoted 8 lines
Quoted 6 lines

Skipping attr check completely for dummy event as suggested by Namhyung, because it does not have any samples anyway.

Thanks, applied to perf/urgent.

- Arnaldo

#3 lore raw

Hi,

Quoted 19 lines · Arnaldo Carvalho de Melo

Em Wed, Aug 31, 2022 at 09:02:46AM -0700, Ian Rogers escreveu:

Quoted 18 lines · Jiri Olsa
Quoted 16 lines

Hongtao Yu reported problem when displaying uregs in perf script for system wide perf.data:

   # perf script -F uregs | head -10
   Samples for 'dummy:HG' event do not have UREGS attribute set. Cannot print 'uregs' field.

The problem is the extra dummy event added for system wide, which does not have proper sample_type setup.

Skipping attr check completely for dummy event as suggested by Namhyung, because it does not have any samples anyway.

Quoted 5 lines
Quoted 1 line

Thanks, applied to perf/urgent.

- Arnaldo

I have met a similar issue on hybrid systems such as ADL, I apply the patch, and it works OK for the normal mode.

 # ./perf record  --intr-regs=di,r8,dx,cx -e 
br_inst_retired.near_call:p -c 1000 --per-thread true
[ perf record: Woken up 1 times to write data ]
[ perf record: Captured and wrote 0.024 MB perf.data (25 samples) ]
# ./perf script -F iregs |head -5
 ABI:2    CX:0xffff90e19d024ed8    DX:0x1    DI:0xffff90e19d024ed8 
R8:0xffffba5e437e7b30
 ABI:2    CX:0x7f5239783000    DX:0x80000000    DI:0xffff90e1801eea00 
 R8:0x71
 ABI:2    CX:0x40    DX:0x60    DI:0xffffffff9fde5ab8    R8:0x40
 ABI:2    CX:0x0    DX:0xffffffffffffffff    DI:0xffff90e1877408e8 
R8:0x1
 ABI:2    CX:0xcc0    DX:0x1    DI:0xffff90e187d17c60    R8:0x40

But the issue still happened when running the test in the pipe mode. In the pipe mode, it calls process_attr() which still checks the attr for the dummy event, so the issue happened.

 # ./perf record -o - --intr-regs=di,r8,dx,cx -e 
br_inst_retired.near_call:p -c 1000 --per-thread true 2>/dev/null|./perf script -F iregs |head -5
Samples for 'dummy:HG' event do not have IREGS attribute set. Cannot print 'iregs' field.
0x120 [0x90]: failed to process type: 64

I have one test patch which can fix the pipe mode issue.

./perf record -o - --intr-regs=di,r8,dx,cx -e
br_inst_retired.near_call:p -c 1000 --per-thread true 2>/dev/null|./perf script -F iregs |head -5
ABI:2 CX:0xffff90e18119e278 DX:0x0 DI:0xffff90e18119f858
R8:0xffff90e18119e278

 ABI:2    CX:0x0    DX:0x1    DI:0xfffffa2844a91580    R8:0x402
 ABI:2    CX:0x0    DX:0x0    DI:0x100cca    R8:0x0
 ABI:2    CX:0x0    DX:0x0    DI:0xffffffff9e997ca5    R8:0x0
 ABI:2    CX:0x113ce8000    DX:0xffff90e198f46600 

DI:0xffff90e189de8000 R8:0x290

Fixes b91e5492f9d7 ("perf record: Add a dummy event on hybrid systems to collect metadata records")
---
 tools/perf/builtin-script.c | 2 ++
 1 file changed, 2 insertions(+)
tools/perf/builtin-script.c +2 -0
index 9152e3e45b69..2d863cdb19fe 100644
@@ -2429,6 +2429,8 @@ static int process_attr(struct perf_tool *tool, union perf_event *event,
        }
 
        if (evsel->core.attr.sample_type) {
+               if (evsel__is_dummy_event(evsel))
+                       return 0;
                err = evsel__check_attr(evsel, scr->session);
                if (err)
                        return err;

--
2.25.1

Signature · 1 line
Zhengjun Xing
maintainer
#4 lore raw

Hello,

Quoted 83 lines · Xing Zhengjun

Hi,

Quoted 19 lines · Arnaldo Carvalho de Melo

Em Wed, Aug 31, 2022 at 09:02:46AM -0700, Ian Rogers escreveu:

Quoted 18 lines · Jiri Olsa
Quoted 16 lines

Hongtao Yu reported problem when displaying uregs in perf script for system wide perf.data:

   # perf script -F uregs | head -10
   Samples for 'dummy:HG' event do not have UREGS attribute set. Cannot print 'uregs' field.

The problem is the extra dummy event added for system wide, which does not have proper sample_type setup.

Skipping attr check completely for dummy event as suggested by Namhyung, because it does not have any samples anyway.

Quoted 5 lines
Quoted 1 line

Thanks, applied to perf/urgent.

- Arnaldo

I have met a similar issue on hybrid systems such as ADL, I apply the patch, and it works OK for the normal mode.

  # ./perf record  --intr-regs=di,r8,dx,cx -e
br_inst_retired.near_call:p -c 1000 --per-thread true
[ perf record: Woken up 1 times to write data ]
[ perf record: Captured and wrote 0.024 MB perf.data (25 samples) ]
# ./perf script -F iregs |head -5
  ABI:2    CX:0xffff90e19d024ed8    DX:0x1    DI:0xffff90e19d024ed8
R8:0xffffba5e437e7b30
  ABI:2    CX:0x7f5239783000    DX:0x80000000    DI:0xffff90e1801eea00
  R8:0x71
  ABI:2    CX:0x40    DX:0x60    DI:0xffffffff9fde5ab8    R8:0x40
  ABI:2    CX:0x0    DX:0xffffffffffffffff    DI:0xffff90e1877408e8
R8:0x1
  ABI:2    CX:0xcc0    DX:0x1    DI:0xffff90e187d17c60    R8:0x40

But the issue still happened when running the test in the pipe mode. In the pipe mode, it calls process_attr() which still checks the attr for the dummy event, so the issue happened.

  # ./perf record -o - --intr-regs=di,r8,dx,cx -e
br_inst_retired.near_call:p -c 1000 --per-thread true 2>/dev/null|./perf
script -F iregs |head -5
Samples for 'dummy:HG' event do not have IREGS attribute set. Cannot
print 'iregs' field.
0x120 [0x90]: failed to process type: 64

I have one test patch which can fix the pipe mode issue.

./perf record -o - --intr-regs=di,r8,dx,cx -e
br_inst_retired.near_call:p -c 1000 --per-thread true 2>/dev/null|./perf script -F iregs |head -5
ABI:2 CX:0xffff90e18119e278 DX:0x0 DI:0xffff90e18119f858
R8:0xffff90e18119e278

  ABI:2    CX:0x0    DX:0x1    DI:0xfffffa2844a91580    R8:0x402
  ABI:2    CX:0x0    DX:0x0    DI:0x100cca    R8:0x0
  ABI:2    CX:0x0    DX:0x0    DI:0xffffffff9e997ca5    R8:0x0
  ABI:2    CX:0x113ce8000    DX:0xffff90e198f46600

DI:0xffff90e189de8000 R8:0x290

Fixes b91e5492f9d7 ("perf record: Add a dummy event on hybrid systems

to collect metadata records")

---
  tools/perf/builtin-script.c | 2 ++
  1 file changed, 2 insertions(+)
tools/perf/builtin-script.c +0 -0
index 9152e3e45b69..2d863cdb19fe 100644

tools/perf/builtin-script.c static int process_attr(struct perf_tool *tool, · line 2434

Quoted 6 lines
tools/perf/builtin-script.c +2 -0
@@ -2429,6 +2429,8 @@ static int process_attr(struct perf_tool *tool,
 union perf_event *event,
         }
 
         if (evsel->core.attr.sample_type) {
+               if (evsel__is_dummy_event(evsel))
+                       return 0;

Maybe we can move this into evsel__check_attr().

Thanks,
Namhyung

Quoted 8 lines
tools/perf/builtin-script.c +0 -0
                 err = evsel__check_attr(evsel, scr->session);
                 if (err)
                         return err;

--
2.25.1

--
Zhengjun Xing

#5 lore raw
Quoted 90 lines · Namhyung Kim

Hello,

Quoted 86 lines · Xing Zhengjun

Hi,

Quoted 19 lines · Arnaldo Carvalho de Melo

Em Wed, Aug 31, 2022 at 09:02:46AM -0700, Ian Rogers escreveu:

Quoted 18 lines · Jiri Olsa
Quoted 16 lines

Hongtao Yu reported problem when displaying uregs in perf script for system wide perf.data:

    # perf script -F uregs | head -10
    Samples for 'dummy:HG' event do not have UREGS attribute set. Cannot print 'uregs' field.

The problem is the extra dummy event added for system wide, which does not have proper sample_type setup.

Skipping attr check completely for dummy event as suggested by Namhyung, because it does not have any samples anyway.

Quoted 5 lines
Quoted 1 line

Thanks, applied to perf/urgent.

- Arnaldo

I have met a similar issue on hybrid systems such as ADL, I apply the patch, and it works OK for the normal mode.

   # ./perf record  --intr-regs=di,r8,dx,cx -e
br_inst_retired.near_call:p -c 1000 --per-thread true
[ perf record: Woken up 1 times to write data ]
[ perf record: Captured and wrote 0.024 MB perf.data (25 samples) ]
# ./perf script -F iregs |head -5
   ABI:2    CX:0xffff90e19d024ed8    DX:0x1    DI:0xffff90e19d024ed8
R8:0xffffba5e437e7b30
   ABI:2    CX:0x7f5239783000    DX:0x80000000    DI:0xffff90e1801eea00
   R8:0x71
   ABI:2    CX:0x40    DX:0x60    DI:0xffffffff9fde5ab8    R8:0x40
   ABI:2    CX:0x0    DX:0xffffffffffffffff    DI:0xffff90e1877408e8
R8:0x1
   ABI:2    CX:0xcc0    DX:0x1    DI:0xffff90e187d17c60    R8:0x40

But the issue still happened when running the test in the pipe mode. In the pipe mode, it calls process_attr() which still checks the attr for the dummy event, so the issue happened.

   # ./perf record -o - --intr-regs=di,r8,dx,cx -e
br_inst_retired.near_call:p -c 1000 --per-thread true 2>/dev/null|./perf
script -F iregs |head -5
Samples for 'dummy:HG' event do not have IREGS attribute set. Cannot
print 'iregs' field.
0x120 [0x90]: failed to process type: 64

I have one test patch which can fix the pipe mode issue.

   ./perf record -o - --intr-regs=di,r8,dx,cx -e

br_inst_retired.near_call:p -c 1000 --per-thread true 2>/dev/null|./perf script -F iregs |head -5
ABI:2 CX:0xffff90e18119e278 DX:0x0 DI:0xffff90e18119f858
R8:0xffff90e18119e278

   ABI:2    CX:0x0    DX:0x1    DI:0xfffffa2844a91580    R8:0x402
   ABI:2    CX:0x0    DX:0x0    DI:0x100cca    R8:0x0
   ABI:2    CX:0x0    DX:0x0    DI:0xffffffff9e997ca5    R8:0x0
   ABI:2    CX:0x113ce8000    DX:0xffff90e198f46600

DI:0xffff90e189de8000 R8:0x290

Fixes b91e5492f9d7 ("perf record: Add a dummy event on hybrid systems

to collect metadata records")

---
   tools/perf/builtin-script.c | 2 ++
   1 file changed, 2 insertions(+)
tools/perf/builtin-script.c +0 -0
index 9152e3e45b69..2d863cdb19fe 100644
@@ -2429,6 +2429,8 @@ static int process_attr(struct perf_tool *tool,
 union perf_event *event,
          }
 
Quoted 8 lines

tools/perf/builtin-script.c static int process_attr(struct perf_tool *tool, · line 2434

Quoted 3 lines
tools/perf/builtin-script.c +2 -0
          if (evsel->core.attr.sample_type) {
+               if (evsel__is_dummy_event(evsel))
+                       return 0;

Maybe we can move this into evsel__check_attr().

Thanks,
Namhyung

Yes, the following changes in evsel__check_attr() can fix both normal and pipe mode issues, Otherwise, we have to patch everywhere when evsel__check_attr() is called.

tools/perf/builtin-script.c +3 -0
index 13580a9c50b8..fb76e3191858 100644
@@ -451,6 +451,9 @@ static int evsel__check_attr(struct evsel *evsel, struct perf_session *session)
        allow_user_set = perf_header__has_feat(&session->header,
                                               HEADER_AUXTRACE);
 
+       if (evsel__is_dummy_event(evsel))
+               allow_user_set = true;
+
        if (PRINT_FIELD(TRACE) &&
            !perf_session__has_traces(session, "record -R"))
                return -EINVAL;

--
2.25.1

Quoted 10 lines
Quoted 10 lines
tools/perf/builtin-script.c +0 -0
                  err = evsel__check_attr(evsel, scr->session);
                  if (err)
                          return err;

--
2.25.1

--
Zhengjun Xing

Signature · 1 line
Zhengjun Xing
#6 lore raw
Quoted 111 lines · Xing Zhengjun
Quoted 90 lines · Namhyung Kim

Hello,

Quoted 86 lines · Xing Zhengjun

Hi,

Quoted 19 lines · Arnaldo Carvalho de Melo

Em Wed, Aug 31, 2022 at 09:02:46AM -0700, Ian Rogers escreveu:

Quoted 18 lines · Jiri Olsa
Quoted 16 lines

Hongtao Yu reported problem when displaying uregs in perf script for system wide perf.data:

    # perf script -F uregs | head -10
    Samples for 'dummy:HG' event do not have UREGS attribute set. Cannot print 'uregs' field.

The problem is the extra dummy event added for system wide, which does not have proper sample_type setup.

Skipping attr check completely for dummy event as suggested by Namhyung, because it does not have any samples anyway.

Quoted 5 lines
Quoted 1 line

Thanks, applied to perf/urgent.

- Arnaldo

I have met a similar issue on hybrid systems such as ADL, I apply the patch, and it works OK for the normal mode.

   # ./perf record  --intr-regs=di,r8,dx,cx -e
br_inst_retired.near_call:p -c 1000 --per-thread true
[ perf record: Woken up 1 times to write data ]
[ perf record: Captured and wrote 0.024 MB perf.data (25 samples) ]
# ./perf script -F iregs |head -5
   ABI:2    CX:0xffff90e19d024ed8    DX:0x1    DI:0xffff90e19d024ed8
R8:0xffffba5e437e7b30
   ABI:2    CX:0x7f5239783000    DX:0x80000000    DI:0xffff90e1801eea00
   R8:0x71
   ABI:2    CX:0x40    DX:0x60    DI:0xffffffff9fde5ab8    R8:0x40
   ABI:2    CX:0x0    DX:0xffffffffffffffff    DI:0xffff90e1877408e8
R8:0x1
   ABI:2    CX:0xcc0    DX:0x1    DI:0xffff90e187d17c60    R8:0x40

But the issue still happened when running the test in the pipe mode. In the pipe mode, it calls process_attr() which still checks the attr for the dummy event, so the issue happened.

   # ./perf record -o - --intr-regs=di,r8,dx,cx -e
br_inst_retired.near_call:p -c 1000 --per-thread true 2>/dev/null|./perf
script -F iregs |head -5
Samples for 'dummy:HG' event do not have IREGS attribute set. Cannot
print 'iregs' field.
0x120 [0x90]: failed to process type: 64

I have one test patch which can fix the pipe mode issue.

   ./perf record -o - --intr-regs=di,r8,dx,cx -e

br_inst_retired.near_call:p -c 1000 --per-thread true 2>/dev/null|./perf script -F iregs |head -5
ABI:2 CX:0xffff90e18119e278 DX:0x0 DI:0xffff90e18119f858
R8:0xffff90e18119e278

   ABI:2    CX:0x0    DX:0x1    DI:0xfffffa2844a91580    R8:0x402
   ABI:2    CX:0x0    DX:0x0    DI:0x100cca    R8:0x0
   ABI:2    CX:0x0    DX:0x0    DI:0xffffffff9e997ca5    R8:0x0
   ABI:2    CX:0x113ce8000    DX:0xffff90e198f46600

DI:0xffff90e189de8000 R8:0x290

Fixes b91e5492f9d7 ("perf record: Add a dummy event on hybrid systems

to collect metadata records")

---
   tools/perf/builtin-script.c | 2 ++
   1 file changed, 2 insertions(+)
tools/perf/builtin-script.c +0 -0
index 9152e3e45b69..2d863cdb19fe 100644
@@ -2429,6 +2429,8 @@ static int process_attr(struct perf_tool *tool,
 union perf_event *event,
          }
 
Quoted 8 lines

tools/perf/builtin-script.c static int process_attr(struct perf_tool *tool, · line 2434

Quoted 3 lines
tools/perf/builtin-script.c +2 -0
          if (evsel->core.attr.sample_type) {
+               if (evsel__is_dummy_event(evsel))
+                       return 0;

Maybe we can move this into evsel__check_attr().

Thanks,
Namhyung

Yes, the following changes in evsel__check_attr() can fix both normal and pipe mode issues, Otherwise, we have to patch everywhere when evsel__check_attr() is called.

tools/perf/builtin-script.c +0 -0
index 13580a9c50b8..fb76e3191858 100644

tools/perf/builtin-script.c static int evsel__check_attr(struct evsel *evsel, struct · line 457

Quoted 7 lines
tools/perf/builtin-script.c +3 -0
@@ -451,6 +451,9 @@ static int evsel__check_attr(struct evsel *evsel, struct
 perf_session *session)
        allow_user_set = perf_header__has_feat(&session->header,
                                               HEADER_AUXTRACE);
 
+       if (evsel__is_dummy_event(evsel))
+               allow_user_set = true;
+

hm, do you need to pass allow_user_set to UREGS check then?

jirka

---

tools/perf/builtin-script.c +4 -1
index 13580a9c50b8..30c9cbdee98d 100644
@@ -451,6 +451,9 @@ static int evsel__check_attr(struct evsel *evsel, struct perf_session *session)
 	allow_user_set = perf_header__has_feat(&session->header,
 					       HEADER_AUXTRACE);
 
+	if (evsel__is_dummy_event(evsel))
+		allow_user_set = true;
+
 	if (PRINT_FIELD(TRACE) &&
 	    !perf_session__has_traces(session, "record -R"))
 		return -EINVAL;
@@ -517,7 +520,7 @@ static int evsel__check_attr(struct evsel *evsel, struct perf_session *session)
 		return -EINVAL;
 
 	if (PRINT_FIELD(UREGS) &&
-	    evsel__check_stype(evsel, PERF_SAMPLE_REGS_USER, "UREGS", PERF_OUTPUT_UREGS))
+	    evsel__do_check_stype(evsel, PERF_SAMPLE_REGS_USER, "UREGS", PERF_OUTPUT_UREGS, allow_user_set))
 		return -EINVAL;
 
 	if (PRINT_FIELD(PHYS_ADDR) &&
maintainer
#7 lore raw
Quoted 117 lines · Jiri Olsa
Quoted 115 lines · Xing Zhengjun
Quoted 90 lines · Namhyung Kim

Hello,

Quoted 86 lines · Xing Zhengjun

Hi,

Quoted 19 lines · Arnaldo Carvalho de Melo

Em Wed, Aug 31, 2022 at 09:02:46AM -0700, Ian Rogers escreveu:

Quoted 18 lines · Jiri Olsa
Quoted 16 lines

Hongtao Yu reported problem when displaying uregs in perf script for system wide perf.data:

    # perf script -F uregs | head -10
    Samples for 'dummy:HG' event do not have UREGS attribute set. Cannot print 'uregs' field.

The problem is the extra dummy event added for system wide, which does not have proper sample_type setup.

Skipping attr check completely for dummy event as suggested by Namhyung, because it does not have any samples anyway.

Quoted 5 lines
Quoted 1 line

Thanks, applied to perf/urgent.

- Arnaldo

I have met a similar issue on hybrid systems such as ADL, I apply the patch, and it works OK for the normal mode.

   # ./perf record  --intr-regs=di,r8,dx,cx -e
br_inst_retired.near_call:p -c 1000 --per-thread true
[ perf record: Woken up 1 times to write data ]
[ perf record: Captured and wrote 0.024 MB perf.data (25 samples) ]
# ./perf script -F iregs |head -5
   ABI:2    CX:0xffff90e19d024ed8    DX:0x1    DI:0xffff90e19d024ed8
R8:0xffffba5e437e7b30
   ABI:2    CX:0x7f5239783000    DX:0x80000000    DI:0xffff90e1801eea00
   R8:0x71
   ABI:2    CX:0x40    DX:0x60    DI:0xffffffff9fde5ab8    R8:0x40
   ABI:2    CX:0x0    DX:0xffffffffffffffff    DI:0xffff90e1877408e8
R8:0x1
   ABI:2    CX:0xcc0    DX:0x1    DI:0xffff90e187d17c60    R8:0x40

But the issue still happened when running the test in the pipe mode. In the pipe mode, it calls process_attr() which still checks the attr for the dummy event, so the issue happened.

   # ./perf record -o - --intr-regs=di,r8,dx,cx -e
br_inst_retired.near_call:p -c 1000 --per-thread true 2>/dev/null|./perf
script -F iregs |head -5
Samples for 'dummy:HG' event do not have IREGS attribute set. Cannot
print 'iregs' field.
0x120 [0x90]: failed to process type: 64

I have one test patch which can fix the pipe mode issue.

   ./perf record -o - --intr-regs=di,r8,dx,cx -e

br_inst_retired.near_call:p -c 1000 --per-thread true 2>/dev/null|./perf script -F iregs |head -5
ABI:2 CX:0xffff90e18119e278 DX:0x0 DI:0xffff90e18119f858
R8:0xffff90e18119e278

   ABI:2    CX:0x0    DX:0x1    DI:0xfffffa2844a91580    R8:0x402
   ABI:2    CX:0x0    DX:0x0    DI:0x100cca    R8:0x0
   ABI:2    CX:0x0    DX:0x0    DI:0xffffffff9e997ca5    R8:0x0
   ABI:2    CX:0x113ce8000    DX:0xffff90e198f46600

DI:0xffff90e189de8000 R8:0x290

Fixes b91e5492f9d7 ("perf record: Add a dummy event on hybrid systems

to collect metadata records")

---
   tools/perf/builtin-script.c | 2 ++
   1 file changed, 2 insertions(+)
tools/perf/builtin-script.c +0 -0
index 9152e3e45b69..2d863cdb19fe 100644
@@ -2429,6 +2429,8 @@ static int process_attr(struct perf_tool *tool,
 union perf_event *event,
          }
 
Quoted 8 lines

tools/perf/builtin-script.c static int process_attr(struct perf_tool *tool, · line 2434

Quoted 3 lines
tools/perf/builtin-script.c +2 -0
          if (evsel->core.attr.sample_type) {
+               if (evsel__is_dummy_event(evsel))
+                       return 0;

Maybe we can move this into evsel__check_attr().

Thanks,
Namhyung

Yes, the following changes in evsel__check_attr() can fix both normal and pipe mode issues, Otherwise, we have to patch everywhere when evsel__check_attr() is called.

tools/perf/builtin-script.c +0 -0
index 13580a9c50b8..fb76e3191858 100644
@@ -451,6 +451,9 @@ static int evsel__check_attr(struct evsel *evsel, struct
 perf_session *session)
        allow_user_set = perf_header__has_feat(&session->header,
                                               HEADER_AUXTRACE);
 
Quoted 5 lines

tools/perf/builtin-script.c static int evsel__check_attr(struct evsel *evsel, struct perf_session *session) · line 456

Quoted 3 lines
tools/perf/builtin-script.c +3 -0
+       if (evsel__is_dummy_event(evsel))
+               allow_user_set = true;
+

hm, do you need to pass allow_user_set to UREGS check then?

Well.. actually I thought we can simply return 0 for a dummy event.

Thanks,
Namhyung

#8 lore raw
Quoted 118 lines · Namhyung Kim
Quoted 117 lines · Jiri Olsa
Quoted 115 lines · Xing Zhengjun
Quoted 90 lines · Namhyung Kim

Hello,

Quoted 86 lines · Xing Zhengjun

Hi,

Quoted 19 lines · Arnaldo Carvalho de Melo

Em Wed, Aug 31, 2022 at 09:02:46AM -0700, Ian Rogers escreveu:

Quoted 18 lines · Jiri Olsa
Quoted 16 lines

Hongtao Yu reported problem when displaying uregs in perf script for system wide perf.data:

    # perf script -F uregs | head -10
    Samples for 'dummy:HG' event do not have UREGS attribute set. Cannot print 'uregs' field.

The problem is the extra dummy event added for system wide, which does not have proper sample_type setup.

Skipping attr check completely for dummy event as suggested by Namhyung, because it does not have any samples anyway.

Quoted 5 lines
Quoted 1 line

Thanks, applied to perf/urgent.

- Arnaldo

I have met a similar issue on hybrid systems such as ADL, I apply the patch, and it works OK for the normal mode.

   # ./perf record  --intr-regs=di,r8,dx,cx -e
br_inst_retired.near_call:p -c 1000 --per-thread true
[ perf record: Woken up 1 times to write data ]
[ perf record: Captured and wrote 0.024 MB perf.data (25 samples) ]
# ./perf script -F iregs |head -5
   ABI:2    CX:0xffff90e19d024ed8    DX:0x1    DI:0xffff90e19d024ed8
R8:0xffffba5e437e7b30
   ABI:2    CX:0x7f5239783000    DX:0x80000000    DI:0xffff90e1801eea00
   R8:0x71
   ABI:2    CX:0x40    DX:0x60    DI:0xffffffff9fde5ab8    R8:0x40
   ABI:2    CX:0x0    DX:0xffffffffffffffff    DI:0xffff90e1877408e8
R8:0x1
   ABI:2    CX:0xcc0    DX:0x1    DI:0xffff90e187d17c60    R8:0x40

But the issue still happened when running the test in the pipe mode. In the pipe mode, it calls process_attr() which still checks the attr for the dummy event, so the issue happened.

   # ./perf record -o - --intr-regs=di,r8,dx,cx -e
br_inst_retired.near_call:p -c 1000 --per-thread true 2>/dev/null|./perf
script -F iregs |head -5
Samples for 'dummy:HG' event do not have IREGS attribute set. Cannot
print 'iregs' field.
0x120 [0x90]: failed to process type: 64

I have one test patch which can fix the pipe mode issue.

   ./perf record -o - --intr-regs=di,r8,dx,cx -e

br_inst_retired.near_call:p -c 1000 --per-thread true 2>/dev/null|./perf script -F iregs |head -5
ABI:2 CX:0xffff90e18119e278 DX:0x0 DI:0xffff90e18119f858
R8:0xffff90e18119e278

   ABI:2    CX:0x0    DX:0x1    DI:0xfffffa2844a91580    R8:0x402
   ABI:2    CX:0x0    DX:0x0    DI:0x100cca    R8:0x0
   ABI:2    CX:0x0    DX:0x0    DI:0xffffffff9e997ca5    R8:0x0
   ABI:2    CX:0x113ce8000    DX:0xffff90e198f46600

DI:0xffff90e189de8000 R8:0x290

Fixes b91e5492f9d7 ("perf record: Add a dummy event on hybrid systems

to collect metadata records")

---
   tools/perf/builtin-script.c | 2 ++
   1 file changed, 2 insertions(+)
tools/perf/builtin-script.c +0 -0
index 9152e3e45b69..2d863cdb19fe 100644
@@ -2429,6 +2429,8 @@ static int process_attr(struct perf_tool *tool,
 union perf_event *event,
          }
 
Quoted 8 lines

tools/perf/builtin-script.c static int process_attr(struct perf_tool *tool, · line 2434

Quoted 3 lines
tools/perf/builtin-script.c +2 -0
          if (evsel->core.attr.sample_type) {
+               if (evsel__is_dummy_event(evsel))
+                       return 0;

Maybe we can move this into evsel__check_attr().

Thanks,
Namhyung

Yes, the following changes in evsel__check_attr() can fix both normal and pipe mode issues, Otherwise, we have to patch everywhere when evsel__check_attr() is called.

tools/perf/builtin-script.c +0 -0
index 13580a9c50b8..fb76e3191858 100644
@@ -451,6 +451,9 @@ static int evsel__check_attr(struct evsel *evsel, struct
 perf_session *session)
        allow_user_set = perf_header__has_feat(&session->header,
                                               HEADER_AUXTRACE);
 
Quoted 7 lines
Quoted 5 lines

tools/perf/builtin-script.c static int evsel__check_attr(struct evsel *evsel, struct perf_session *session) · line 456

Quoted 3 lines
tools/perf/builtin-script.c +3 -0
+       if (evsel__is_dummy_event(evsel))
+               allow_user_set = true;
+

hm, do you need to pass allow_user_set to UREGS check then?

Well.. actually I thought we can simply return 0 for a dummy event.

true :-)

jirka

#9 lore raw
Quoted 125 lines · Jiri Olsa
Quoted 124 lines · Namhyung Kim
Quoted 122 lines · Jiri Olsa
Quoted 111 lines · Xing Zhengjun
Quoted 90 lines · Namhyung Kim

Hello,

Quoted 86 lines · Xing Zhengjun

Hi,

Quoted 19 lines · Arnaldo Carvalho de Melo

Em Wed, Aug 31, 2022 at 09:02:46AM -0700, Ian Rogers escreveu:

Quoted 18 lines · Jiri Olsa
Quoted 16 lines

Hongtao Yu reported problem when displaying uregs in perf script for system wide perf.data:

     # perf script -F uregs | head -10
     Samples for 'dummy:HG' event do not have UREGS attribute set. Cannot print 'uregs' field.

The problem is the extra dummy event added for system wide, which does not have proper sample_type setup.

Skipping attr check completely for dummy event as suggested by Namhyung, because it does not have any samples anyway.

Quoted 5 lines
Quoted 1 line

Thanks, applied to perf/urgent.

- Arnaldo

I have met a similar issue on hybrid systems such as ADL, I apply the patch, and it works OK for the normal mode.

    # ./perf record  --intr-regs=di,r8,dx,cx -e
br_inst_retired.near_call:p -c 1000 --per-thread true
[ perf record: Woken up 1 times to write data ]
[ perf record: Captured and wrote 0.024 MB perf.data (25 samples) ]
# ./perf script -F iregs |head -5
    ABI:2    CX:0xffff90e19d024ed8    DX:0x1    DI:0xffff90e19d024ed8
R8:0xffffba5e437e7b30
    ABI:2    CX:0x7f5239783000    DX:0x80000000    DI:0xffff90e1801eea00
    R8:0x71
    ABI:2    CX:0x40    DX:0x60    DI:0xffffffff9fde5ab8    R8:0x40
    ABI:2    CX:0x0    DX:0xffffffffffffffff    DI:0xffff90e1877408e8
R8:0x1
    ABI:2    CX:0xcc0    DX:0x1    DI:0xffff90e187d17c60    R8:0x40

But the issue still happened when running the test in the pipe mode. In the pipe mode, it calls process_attr() which still checks the attr for the dummy event, so the issue happened.

    # ./perf record -o - --intr-regs=di,r8,dx,cx -e
br_inst_retired.near_call:p -c 1000 --per-thread true 2>/dev/null|./perf
script -F iregs |head -5
Samples for 'dummy:HG' event do not have IREGS attribute set. Cannot
print 'iregs' field.
0x120 [0x90]: failed to process type: 64

I have one test patch which can fix the pipe mode issue.

    ./perf record -o - --intr-regs=di,r8,dx,cx -e
br_inst_retired.near_call:p -c 1000 --per-thread true 2>/dev/null|./perf
script -F iregs |head -5
    ABI:2    CX:0xffff90e18119e278    DX:0x0    DI:0xffff90e18119f858
R8:0xffff90e18119e278
    ABI:2    CX:0x0    DX:0x1    DI:0xfffffa2844a91580    R8:0x402
    ABI:2    CX:0x0    DX:0x0    DI:0x100cca    R8:0x0
    ABI:2    CX:0x0    DX:0x0    DI:0xffffffff9e997ca5    R8:0x0
    ABI:2    CX:0x113ce8000    DX:0xffff90e198f46600

DI:0xffff90e189de8000 R8:0x290

Fixes b91e5492f9d7 ("perf record: Add a dummy event on hybrid systems

to collect metadata records")

---
    tools/perf/builtin-script.c | 2 ++
    1 file changed, 2 insertions(+)
tools/perf/builtin-script.c +0 -0
index 9152e3e45b69..2d863cdb19fe 100644
@@ -2429,6 +2429,8 @@ static int process_attr(struct perf_tool *tool,
 union perf_event *event,
           }
 
Quoted 8 lines

tools/perf/builtin-script.c static int process_attr(struct perf_tool *tool, · line 2434

Quoted 3 lines
tools/perf/builtin-script.c +2 -0
           if (evsel->core.attr.sample_type) {
+               if (evsel__is_dummy_event(evsel))
+                       return 0;

Maybe we can move this into evsel__check_attr().

Thanks,
Namhyung

Yes, the following changes in evsel__check_attr() can fix both normal and pipe mode issues, Otherwise, we have to patch everywhere when evsel__check_attr() is called.

tools/perf/builtin-script.c +0 -0
index 13580a9c50b8..fb76e3191858 100644

tools/perf/builtin-script.c static int evsel__check_attr(struct evsel *evsel, struct · line 457

Quoted 7 lines
tools/perf/builtin-script.c +3 -0
@@ -451,6 +451,9 @@ static int evsel__check_attr(struct evsel *evsel, struct
 perf_session *session)
         allow_user_set = perf_header__has_feat(&session->header,
                                                HEADER_AUXTRACE);
 
+       if (evsel__is_dummy_event(evsel))
+               allow_user_set = true;
+

hm, do you need to pass allow_user_set to UREGS check then?

Quoted 5 lines
Quoted 1 line

Well.. actually I thought we can simply return 0 for a dummy event.

true :-)

jirka

Thanks all, I will post a patch based on it soon.

Signature · 1 line
Zhengjun Xing