[PATCH] perf parse-events: Don't report unreadable tracepoint globs as unknown

[PATCH] perf parse-events: Don't report unreadable tracepoint globs as unknown

[Hang Zhengyang via B4 Relay]

From: Hang Zhengyang <hang.zhengyang1010@gmail.com>

When the events directory can't be read, a tracepoint glob like
'syscalls:*' is reported as an unknown tracepoint, with a hint about a
missing CONFIG_ option.  A plain tracepoint name reports the permission
problem instead:

  $ perf stat -e 'syscalls:*' true
  event syntax error: 'syscalls:*'
                       \___ unknown tracepoint

  Error:  File /sys/kernel/tracing//events/syscalls/* not found.

add_tracepoint_multi_sys() reports ENOENT whenever nothing matched and
ignores the error returned while reading the directories, which got
lost when this code moved to tp_pmu.c.  Report that error when there
is one.  As a normal user 'syscalls:*' now gets "can't access trace
events" and the remount hint, and perf test 5.1 is skipped again
instead of failing.

Fixes: d002aab87de8 ("perf tp_pmu: Factor existing tracepoint logic to new file")
Assisted-by: Claude:claude-opus-5-5
Signed-off-by: Hang Zhengyang <hang.zhengyang1010@gmail.com>
---
Found running perf test as a normal user.  Tested on perf-tools-next
13a8d580433a.  As a normal user, 'syscalls:*', 'sched:*',
'syscalls:sys_enter_open*' and '*:sys_enter_openat' now report the
permission problem, perf test 5 passes with 5.1 skipped, and the rest
of perf test is unchanged.  As root, 'nosuchsys:*' still reports an
unknown tracepoint, 'syscalls:sys_enter_open*' counts, and perf test
5.1 passes.

The issue was found and the fix and changelog drafted with Claude; I
reproduced the issue and reviewed and tested the change.
---
 tools/perf/util/parse-events.c | 6 ++++--
 1 file changed, 4 insertions(+), 2 deletions(-)

diff --git a/tools/perf/util/parse-events.c b/tools/perf/util/parse-events.c
index f3ba70d4f76b..8a6d5a13863d 100644
--- a/tools/perf/util/parse-events.c
+++ b/tools/perf/util/parse-events.c
@@ -572,8 +572,10 @@ static int add_tracepoint_multi_sys(struct parse_events_state *parse_state,
 		ret = tp_pmu__for_each_tp_sys(&args, add_tracepoint_multi_sys_cb);
 	}
 	if (args.found == 0) {
-		tracepoint_error(err, ENOENT, sys_glob, evt_glob, loc->first_column);
-		return -ENOENT;
+		if (!ret)
+			ret = -ENOENT;
+		tracepoint_error(err, ret, sys_glob, evt_glob, loc->first_column);
+		return ret;
 	}
 	return ret;
 }

---
base-commit: 13a8d580433a3b72b10be34f0e26787fe5659d3c
change-id: 20261110-tp-glob-errno-672ea1eb786b

Best regards,
-- 
Hang Zhengyang <hang.zhengyang1010@gmail.com>

Re: [PATCH] perf parse-events: Don't report unreadable tracepoint globs as unknown

[sashiko-bot]

Thank you for your contribution! Sashiko AI review found no new issues
in this patch, but noticed 1 related pre-existing issue it does not
introduce:
- [Medium] perf parse-events: event enumeration abort in
  add_tracepoint_multi_sys_cb() [1]
--

commit a63217b652c98d06df0265b3b5da9302408c378e
Author: Hang Zhengyang <hang.zhengyang1010@gmail.com>

perf parse-events: Don't report unreadable tracepoint globs as unknown

This commit preserves the directory read error in add_tracepoint_multi_sys()
when no tracepoints match a glob, rather than always reporting -ENOENT.

> diff --git a/tools/perf/util/parse-events.c b/tools/perf/util/parse-events.c
> index f3ba70d4f76b2..8a6d5a13863d2 100644
> --- a/tools/perf/util/parse-events.c
> +++ b/tools/perf/util/parse-events.c
> @@ -572,8 +572,10 @@ static int add_tracepoint_multi_sys(struct parse_events_state *parse_state,
>  		ret = tp_pmu__for_each_tp_sys(&args, add_tracepoint_multi_sys_cb);
>  	}
>  	if (args.found == 0) {
> -		tracepoint_error(err, ENOENT, sys_glob, evt_glob, loc->first_column);
> -		return -ENOENT;
> +		if (!ret)
> +			ret = -ENOENT;
> +		tracepoint_error(err, ret, sys_glob, evt_glob, loc->first_column);
> +		return ret;
>  	}
>  	return ret;
>  }

[Severity: Medium]
This problem wasn't introduced by this patch, but can concurrent removal of a
tracepoint subsystem directory cause add_tracepoint_multi_sys_cb() to abort
the entire enumeration [1]?

While tp_pmu__for_each_tp_sys() iterates over subsystem directories,
add_tracepoint_multi_sys_cb() calls add_tracepoint_multi_event() ->
tp_pmu__for_each_tp_event(). If a matching subsystem directory is removed
before open(), tp_pmu__for_each_tp_event() returns -ENOENT, which stops the
loop in tp_pmu__for_each_tp_sys() instead of skipping the missing directory
and continuing to enumerate the remaining tracepoints.

[1] https://sashiko.dev/#/bug/linux-0b2ec2f2-377b-4c58-91d0-b8dd34e637ac

-- 
Sashiko AI review · https://sashiko.dev/#/patchset/20261110-tp-glob-errno-v1-1-6c2cb609c590@gmail.com?part=1

Re: [PATCH] perf parse-events: Don't report unreadable tracepoint globs as unknown

[Ian Rogers]

On Mon, Nov 9, 2026 at 7:05 PM Hang Zhengyang via B4 Relay
<devnull+hang.zhengyang1010.gmail.com@kernel.org> wrote:
>
> From: Hang Zhengyang <hang.zhengyang1010@gmail.com>
>
> When the events directory can't be read, a tracepoint glob like
> 'syscalls:*' is reported as an unknown tracepoint, with a hint about a
> missing CONFIG_ option.  A plain tracepoint name reports the permission
> problem instead:
>
>   $ perf stat -e 'syscalls:*' true
>   event syntax error: 'syscalls:*'
>                        \___ unknown tracepoint
>
>   Error:  File /sys/kernel/tracing//events/syscalls/* not found.

Thanks Hang! The Sashiko feedback seems like overkill to fix here.
Could you give a before and after comparison of the console output for
this change?

Thanks,
Ian

> add_tracepoint_multi_sys() reports ENOENT whenever nothing matched and
> ignores the error returned while reading the directories, which got
> lost when this code moved to tp_pmu.c.  Report that error when there
> is one.  As a normal user 'syscalls:*' now gets "can't access trace
> events" and the remount hint, and perf test 5.1 is skipped again
> instead of failing.
>
> Fixes: d002aab87de8 ("perf tp_pmu: Factor existing tracepoint logic to new file")
> Assisted-by: Claude:claude-opus-5-5
> Signed-off-by: Hang Zhengyang <hang.zhengyang1010@gmail.com>
> ---
> Found running perf test as a normal user.  Tested on perf-tools-next
> 13a8d580433a.  As a normal user, 'syscalls:*', 'sched:*',
> 'syscalls:sys_enter_open*' and '*:sys_enter_openat' now report the
> permission problem, perf test 5 passes with 5.1 skipped, and the rest
> of perf test is unchanged.  As root, 'nosuchsys:*' still reports an
> unknown tracepoint, 'syscalls:sys_enter_open*' counts, and perf test
> 5.1 passes.
>
> The issue was found and the fix and changelog drafted with Claude; I
> reproduced the issue and reviewed and tested the change.
> ---
>  tools/perf/util/parse-events.c | 6 ++++--
>  1 file changed, 4 insertions(+), 2 deletions(-)
>
> diff --git a/tools/perf/util/parse-events.c b/tools/perf/util/parse-events.c
> index f3ba70d4f76b..8a6d5a13863d 100644
> --- a/tools/perf/util/parse-events.c
> +++ b/tools/perf/util/parse-events.c
> @@ -572,8 +572,10 @@ static int add_tracepoint_multi_sys(struct parse_events_state *parse_state,
>                 ret = tp_pmu__for_each_tp_sys(&args, add_tracepoint_multi_sys_cb);
>         }
>         if (args.found == 0) {
> -               tracepoint_error(err, ENOENT, sys_glob, evt_glob, loc->first_column);
> -               return -ENOENT;
> +               if (!ret)
> +                       ret = -ENOENT;
> +               tracepoint_error(err, ret, sys_glob, evt_glob, loc->first_column);
> +               return ret;
>         }
>         return ret;
>  }
>
> ---
> base-commit: 13a8d580433a3b72b10be34f0e26787fe5659d3c
> change-id: 20261110-tp-glob-errno-672ea1eb786b
>
> Best regards,
> --
> Hang Zhengyang <hang.zhengyang1010@gmail.com>
>
>