Skip to contentSkip to conversation

perf: Ignore format attributes with an unknown perf_event_attr field

PATCH In review Rob Herring 3 participants 8 messages Sep 1, 2022 – Sep 9, 2022 perf events

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 perf: Ignore format attributes with an unknown perf_event_attr field

If the kernel exposes a new perf_event_attr field in a format attr, perf will return an error stating the specified PMU can't be found. For example, a format attr with 'config3:0-63' causes an error if config3 is unknown to perf. This causes a compatibility issue between a newer kernel and an older perf tool.

The addition here makes any attr string up to the ':' ignored, but still checks the 'bits' portion.

---
This is the YACC mud I threw and seems to stick. Maybe there's a better way to handle this. It doesn't seem like there's a way to do wildcards (i.e. config.*) in YACC.

This is needed for this series[1]. Unfortunately the best we do to avoid the issue is applying this to stable. I think there's some time before v8.7 h/w is deployed, too.

Rob

[1] https://lore.kernel.org/all/[email protected]/

 tools/perf/util/pmu.y | 3 +++
 1 file changed, 3 insertions(+)
tools/perf/util/pmu.y +3 -0
index bfd7e8509869..3096864ec9b9 100644
@@ -60,6 +60,9 @@ PP_CONFIG2 ':' bits
 				      PERF_PMU_FORMAT_VALUE_CONFIG2,
 				      $3));
 }
+|
+error ':' bits
+{}
 
 bits:
 bits ',' bit_term
Signature · 1 line
2.34.1
#1 lore raw
Quoted 8 lines · Rob Herring

If the kernel exposes a new perf_event_attr field in a format attr, perf will return an error stating the specified PMU can't be found. For example, a format attr with 'config3:0-63' causes an error if config3 is unknown to perf. This causes a compatibility issue between a newer kernel and an older perf tool.

The addition here makes any attr string up to the ':' ignored, but still checks the 'bits' portion.

So, can you please show what is the behavior of the tool, with an actual command line and it's output, before and after your patch?

- Arnaldo

Quoted 32 lines

---
This is the YACC mud I threw and seems to stick. Maybe there's a better way to handle this. It doesn't seem like there's a way to do wildcards (i.e. config.*) in YACC.

This is needed for this series[1]. Unfortunately the best we do to avoid the issue is applying this to stable. I think there's some time before v8.7 h/w is deployed, too.

Rob

[1] https://lore.kernel.org/all/[email protected]/

tools/perf/util/pmu.y | 3 +++

1 file changed, 3 insertions(+)
tools/perf/util/pmu.y +3 -0
index bfd7e8509869..3096864ec9b9 100644
@@ -60,6 +60,9 @@ PP_CONFIG2 ':' bits
 				      PERF_PMU_FORMAT_VALUE_CONFIG2,
 				      $3));
 }
+|
+error ':' bits
+{}
 

bits:
bits ',' bit_term

#2 lore raw
Quoted 10 lines · Arnaldo Carvalho de Melo
Quoted 6 lines · Rob Herring

If the kernel exposes a new perf_event_attr field in a format attr, perf will return an error stating the specified PMU can't be found. For example, a format attr with 'config3:0-63' causes an error if config3 is unknown to perf. This causes a compatibility issue between a newer kernel and an older perf tool.

Quoted 4 lines
Quoted 2 lines

The addition here makes any attr string up to the ':' ignored, but still checks the 'bits' portion.

So, can you please show what is the behavior of the tool, with an actual command line and it's output, before and after your patch?

Before this patch with a kernel adding 'config3' I get:

# perf record -e arm_spe// -- true
event syntax error: 'arm_spe//'
                     \___ Cannot find PMU `arm_spe'. Missing kernel support?
Run 'perf list' for a list of valid events

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

-e, --event <event> event selector. use 'perf list' to list
available events

After this patch, I get:

# perf record -e arm_spe// -- true
[ perf record: Woken up 2 times to write data ]
[ perf record: Captured and wrote 0.091 MB perf.data ]

Rob

maintainer
#3 lore raw
In reply to Rob Herring

Hello,

Quoted 11 lines · Rob Herring

If the kernel exposes a new perf_event_attr field in a format attr, perf will return an error stating the specified PMU can't be found. For example, a format attr with 'config3:0-63' causes an error if config3 is unknown to perf. This causes a compatibility issue between a newer kernel and an older perf tool.

The addition here makes any attr string up to the ':' ignored, but still checks the 'bits' portion.

Quoted 8 lines

---
This is the YACC mud I threw and seems to stick. Maybe there's a better way to handle this. It doesn't seem like there's a way to do wildcards (i.e. config.*) in YACC.

This is needed for this series[1]. Unfortunately the best we do to avoid the issue is applying this to stable. I think there's some time before v8.7 h/w is deployed, too.

Maybe you could change the format_term rule to take an identifier instead of PP_CONFIG* directly and pass it to perf_pmu__new_format(). Then it could check the string and create an appropriate PERF_PMU_FORMAT_VALUE_* or ignore it according to the PERF_ATTR_SIZE_VER*.

Thanks,
Namhyung

Quoted 25 lines

Rob

[1] https://lore.kernel.org/all/[email protected]/

 tools/perf/util/pmu.y | 3 +++
 1 file changed, 3 insertions(+)
tools/perf/util/pmu.y +3 -1
index bfd7e8509869..3096864ec9b9 100644
@@ -60,6 +60,9 @@ PP_CONFIG2 ':' bits
                                      PERF_PMU_FORMAT_VALUE_CONFIG2,
                                      $3));
 }
+|
+error ':' bits
+{}
 
 bits:
 bits ',' bit_term
--

2.34.1

#4 lore raw
Quoted 20 lines · Namhyung Kim

Hello,

Quoted 16 lines · Rob Herring

If the kernel exposes a new perf_event_attr field in a format attr, perf will return an error stating the specified PMU can't be found. For example, a format attr with 'config3:0-63' causes an error if config3 is unknown to perf. This causes a compatibility issue between a newer kernel and an older perf tool.

The addition here makes any attr string up to the ':' ignored, but still checks the 'bits' portion.

---
This is the YACC mud I threw and seems to stick. Maybe there's a better way to handle this. It doesn't seem like there's a way to do wildcards (i.e. config.*) in YACC.

Quoted 8 lines
Quoted 3 lines

This is needed for this series[1]. Unfortunately the best we do to avoid the issue is applying this to stable. I think there's some time before v8.7 h/w is deployed, too.

Maybe you could change the format_term rule to take an identifier instead of PP_CONFIG* directly and pass it to perf_pmu__new_format(). Then it could check the string and create an appropriate PERF_PMU_FORMAT_VALUE_* or ignore it according to the PERF_ATTR_SIZE_VER*.

That only moves parsing of configN from YACC to strcmp in C. In doing so, we'd be left with just the 'error' token case which seems a bit odd (if there's another way to do it, I don't know. yacc is not my thing). Is that really better? Unless there is some way to retrieve the PERF_ATTR_SIZE_VER* from the kernel at runtime?

Rob

maintainer
#5 lore raw
Quoted 31 lines · Rob Herring
Quoted 20 lines · Namhyung Kim

Hello,

Quoted 16 lines · Rob Herring

If the kernel exposes a new perf_event_attr field in a format attr, perf will return an error stating the specified PMU can't be found. For example, a format attr with 'config3:0-63' causes an error if config3 is unknown to perf. This causes a compatibility issue between a newer kernel and an older perf tool.

The addition here makes any attr string up to the ':' ignored, but still checks the 'bits' portion.

---
This is the YACC mud I threw and seems to stick. Maybe there's a better way to handle this. It doesn't seem like there's a way to do wildcards (i.e. config.*) in YACC.

Quoted 8 lines
Quoted 3 lines

This is needed for this series[1]. Unfortunately the best we do to avoid the issue is applying this to stable. I think there's some time before v8.7 h/w is deployed, too.

Maybe you could change the format_term rule to take an identifier instead of PP_CONFIG* directly and pass it to perf_pmu__new_format(). Then it could check the string and create an appropriate PERF_PMU_FORMAT_VALUE_* or ignore it according to the PERF_ATTR_SIZE_VER*.

Quoted 4 lines

That only moves parsing of configN from YACC to strcmp in C. In doing so, we'd be left with just the 'error' token case which seems a bit odd (if there's another way to do it, I don't know. yacc is not my thing). Is that really better?

I thought we could do more flexible handling and detailed error reporting in the C code. But it could be done in the lex/yacc as well..

I think the general idea is that we want to run a more recent version of perf tools than the kernel. So if it detects the tool is older, it can show a warning message like:

"config3 is not in the perf_event_attr.. skipping.
Maybe you're running on a newer kernel. Please upgrade the perf tool."

Quoted 2 lines

Unless there is some way to retrieve
the PERF_ATTR_SIZE_VER* from the kernel at runtime?

Even if it can retrieve the info at runtime, perf tool might not know how to use the new config term.

Thanks,
Namhyung

#6 lore raw
Quoted 41 lines · Namhyung Kim
Quoted 31 lines · Rob Herring
Quoted 20 lines · Namhyung Kim

Hello,

Quoted 16 lines · Rob Herring

If the kernel exposes a new perf_event_attr field in a format attr, perf will return an error stating the specified PMU can't be found. For example, a format attr with 'config3:0-63' causes an error if config3 is unknown to perf. This causes a compatibility issue between a newer kernel and an older perf tool.

The addition here makes any attr string up to the ':' ignored, but still checks the 'bits' portion.

---
This is the YACC mud I threw and seems to stick. Maybe there's a better way to handle this. It doesn't seem like there's a way to do wildcards (i.e. config.*) in YACC.

Quoted 8 lines
Quoted 3 lines

This is needed for this series[1]. Unfortunately the best we do to avoid the issue is applying this to stable. I think there's some time before v8.7 h/w is deployed, too.

Maybe you could change the format_term rule to take an identifier instead of PP_CONFIG* directly and pass it to perf_pmu__new_format(). Then it could check the string and create an appropriate PERF_PMU_FORMAT_VALUE_* or ignore it according to the PERF_ATTR_SIZE_VER*.

Quoted 4 lines

That only moves parsing of configN from YACC to strcmp in C. In doing so, we'd be left with just the 'error' token case which seems a bit odd (if there's another way to do it, I don't know. yacc is not my thing). Is that really better?

I thought we could do more flexible handling and detailed error reporting in the C code. But it could be done in the lex/yacc as well..

Quoted 6 lines

I think the general idea is that we want to run a more recent version of perf tools than the kernel. So if it detects the tool is older, it can show a warning message like:

"config3 is not in the perf_event_attr.. skipping.
Maybe you're running on a newer kernel. Please upgrade the perf tool."

I figured out how to simplify the yacc code and add a warning. However, one thing to note is that we'll always get the warning if any PMU has an unsupported format attr because all the PMUs are scanned. For example, just this gives a warning even though the SPE PMU is not used:

perf record -e cycles -- true

So the warning might be misleading. On the flip side, new additions are rare.

Rob

maintainer
#7 lore raw
Quoted 47 lines · Rob Herring
Quoted 45 lines · Namhyung Kim
Quoted 31 lines · Rob Herring
Quoted 20 lines · Namhyung Kim

Hello,

Quoted 16 lines · Rob Herring

If the kernel exposes a new perf_event_attr field in a format attr, perf will return an error stating the specified PMU can't be found. For example, a format attr with 'config3:0-63' causes an error if config3 is unknown to perf. This causes a compatibility issue between a newer kernel and an older perf tool.

The addition here makes any attr string up to the ':' ignored, but still checks the 'bits' portion.

---
This is the YACC mud I threw and seems to stick. Maybe there's a better way to handle this. It doesn't seem like there's a way to do wildcards (i.e. config.*) in YACC.

Quoted 8 lines
Quoted 3 lines

This is needed for this series[1]. Unfortunately the best we do to avoid the issue is applying this to stable. I think there's some time before v8.7 h/w is deployed, too.

Maybe you could change the format_term rule to take an identifier instead of PP_CONFIG* directly and pass it to perf_pmu__new_format(). Then it could check the string and create an appropriate PERF_PMU_FORMAT_VALUE_* or ignore it according to the PERF_ATTR_SIZE_VER*.

Quoted 4 lines

That only moves parsing of configN from YACC to strcmp in C. In doing so, we'd be left with just the 'error' token case which seems a bit odd (if there's another way to do it, I don't know. yacc is not my thing). Is that really better?

I thought we could do more flexible handling and detailed error reporting in the C code. But it could be done in the lex/yacc as well..

I think the general idea is that we want to run a more recent version of perf tools than the kernel. So if it detects the tool is older, it can show a warning message like:

Quoted 6 lines
Quoted 2 lines

"config3 is not in the perf_event_attr.. skipping.
Maybe you're running on a newer kernel. Please upgrade the perf tool."

I figured out how to simplify the yacc code and add a warning. However, one thing to note is that we'll always get the warning if any PMU has an unsupported format attr because all the PMUs are scanned.

Right, I think we need to change this behavior.

Quoted 6 lines

For example, just this gives a warning even though the SPE PMU is not used:

perf record -e cycles -- true

So the warning might be misleading. On the flip side, new additions are rare.

Yeah, we should not warn at parsing, do when it's actually used.

Thanks,
Namhyung