Message-ID: <eed27aaf-fd0a-4609-a30b-68e7c5c11890@linux.alibaba.com>
Date: Wed, 29 Oct 2025 20:55:14 +0800
From: Shuai Xue <xueshuai@...ux.alibaba.com>
To: Ian Rogers <irogers@...gle.com>
Cc: alexander.shishkin@...ux.intel.com, peterz@...radead.org,
 james.clark@....com, leo.yan@...aro.org, mingo@...hat.com,
 baolin.wang@...ux.alibaba.com, acme@...nel.org, mark.rutland@....com,
 jolsa@...nel.org, namhyung@...nel.org, adrian.hunter@...el.com,
 linux-perf-users@...r.kernel.org, linux-kernel@...r.kernel.org,
 nathan@...nel.org, bpf@...r.kernel.org
Subject: Re: [PATCH] perf record: skip synthesize event when open evsel failed



在 2025/10/24 10:45, Shuai Xue 写道:
> 
> 
> 在 2025/10/24 00:08, Ian Rogers 写道:
>> On Wed, Oct 22, 2025 at 6:50 PM Shuai Xue <xueshuai@...ux.alibaba.com> wrote:
>>>
>>> When using perf record with the `--overwrite` option, a segmentation fault
>>> occurs if an event fails to open. For example:
>>>
>>>    perf record -e cycles-ct -F 1000 -a --overwrite
>>>    Error:
>>>    cycles-ct:H: PMU Hardware doesn't support sampling/overflow-interrupts. Try 'perf stat'
>>>    perf: Segmentation fault
>>>        #0 0x6466b6 in dump_stack debug.c:366
>>>        #1 0x646729 in sighandler_dump_stack debug.c:378
>>>        #2 0x453fd1 in sigsegv_handler builtin-record.c:722
>>>        #3 0x7f8454e65090 in __restore_rt libc-2.32.so[54090]
>>>        #4 0x6c5671 in __perf_event__synthesize_id_index synthetic-events.c:1862
>>>        #5 0x6c5ac0 in perf_event__synthesize_id_index synthetic-events.c:1943
>>>        #6 0x458090 in record__synthesize builtin-record.c:2075
>>>        #7 0x45a85a in __cmd_record builtin-record.c:2888
>>>        #8 0x45deb6 in cmd_record builtin-record.c:4374
>>>        #9 0x4e5e33 in run_builtin perf.c:349
>>>        #10 0x4e60bf in handle_internal_command perf.c:401
>>>        #11 0x4e6215 in run_argv perf.c:448
>>>        #12 0x4e653a in main perf.c:555
>>>        #13 0x7f8454e4fa72 in __libc_start_main libc-2.32.so[3ea72]
>>>        #14 0x43a3ee in _start ??:0
>>>
>>> The --overwrite option implies --tail-synthesize, which collects non-sample
>>> events reflecting the system status when recording finishes. However, when
>>> evsel opening fails (e.g., unsupported event 'cycles-ct'), session->evlist
>>> is not initialized and remains NULL. The code unconditionally calls
>>> record__synthesize() in the error path, which iterates through the NULL
>>> evlist pointer and causes a segfault.
>>>
>>> To fix it, move the record__synthesize() call inside the error check block, so
>>> it's only called when there was no error during recording, ensuring that evlist
>>> is properly initialized.
>>>
>>> Fixes: 4ea648aec019 ("perf record: Add --tail-synthesize option")
>>> Signed-off-by: Shuai Xue <xueshuai@...ux.alibaba.com>
>>
>> This looks great! I wonder if we can add a test, perhaps here:
>> https://web.git.kernel.org/pub/scm/linux/kernel/git/perf/perf-tools-next.git/tree/tools/perf/tests/shell/record.sh?h=perf-tools-next#n435
>> something like:
>> ```
>> $ perf record -e foobar -F 1000 -a --overwrite -o /dev/null -- sleep 0.1
>> ```
>> in a new test subsection for test_overwrite? foobar would be an event
>> that we could assume isn't present. Could you help with a test
>> covering the problems you've uncovered and perhaps related flags?
>>
> 
> Hi, Ian,
> 
> Good suggestion, I'd like to add a test. But foobar may not a good case.
> 
> Regarding your example:
> 
>    perf record -e foobar -a --overwrite -o /dev/null -- sleep 0.1
>    event syntax error: 'foobar'
>                         \___ Bad event name
> 
>    Unable to find event on a PMU of 'foobar'
>    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
> 
> 
> The issue with using foobar is that it's an invalid event name, and the
> perf parser will reject it much earlier. This means the test would exit
> before reaching the part of the code path we want to verify (where
> record__synthesize() could be called).
> 
> A potential alternative could be testing an error case such as EACCES:
> 
>    perf record -e cycles -C 0 --overwrite -o /dev/null -- sleep 0.1
> 
> This could reproduce the scenario of a failure when attempting to access
> a valid event, such as due to permission restrictions. However, the
> limitation here is that users may override
> /proc/sys/kernel/perf_event_paranoid, which affects whether or not this
> test would succeed in triggering an EACCES error.
> 
> 
> If you have any other suggestions or ideas for a better way to simulate
> this situation, I'd love to hear them.
> 
> Thanks.
> Shuai

Hi, Ian,

Gentle ping.

Thanks.
Shuai

Powered by blists - more mailing lists