Message-ID: <CAP-5=fUvwokP=MYmS7kZqjCk+ZYs8A-9G+i3zt-zvjdZA6E_Jg@mail.gmail.com>
Date: Fri, 31 Oct 2025 09:04:38 -0700
From: Ian Rogers <irogers@...gle.com>
To: Shuai Xue <xueshuai@...ux.alibaba.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

On Thu, Oct 30, 2025 at 7:36 PM Shuai Xue <xueshuai@...ux.alibaba.com> wrote:
>
> 在 2025/10/31 01:32, Ian Rogers 写道:
> > On Wed, Oct 29, 2025 at 5:55 AM Shuai Xue <xueshuai@...ux.alibaba.com> wrote:
> >>
> >>
> >>
> >> 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.
> >
> > Sorry, for the delay. I was trying to think of a better way given the
> > problems you mention and then got distracted. I wonder if a legacy
> > event that core PMUs never implement would be a good candidate to
> > test. For example, the event "node-prefetch-misses" is for "Local
> > memory prefetch misses" but the memory controller tends to be a
> > separate PMU and this event is never implemented to my knowledge.
> > Running this locally I see:
> >
> > ```
> > $ perf record -e node-prefetch-misses -a --overwrite -o /dev/null -- sleep 0.1
> > Lowering default frequency rate from 4000 to 1750.
> > Please consider tweaking /proc/sys/kernel/perf_event_max_sample_rate.
> > Error:
> > Failure to open event 'cpu_atom/node-prefetch-misses/' on PMU
> > 'cpu_atom' which will be removed.
> > No fallback found for 'cpu_atom/node-prefetch-misses/' for error 2
> > Error:
> > Failure to open event 'cpu_core/node-prefetch-misses/' on PMU
> > 'cpu_core' which will be removed.
> > No fallback found for 'cpu_core/node-prefetch-misses/' for error 2
> > Error:
> > Failure to open any events for recording.
> > perf: Segmentation fault
> >     #0 0x55a487ad8b87 in dump_stack debug.c:366
> >     #1 0x55a487ad8bfd in sighandler_dump_stack debug.c:378
> >     #2 0x55a4878c6f94 in sigsegv_handler builtin-record.c:722
> >     #3 0x7f72aae49df0 in __restore_rt libc_sigaction.c:0
> >     #4 0x55a487b57ef8 in __perf_event__synthesize_id_index
> > synthetic-events.c:1862
> >     #5 0x55a487b58346 in perf_event__synthesize_id_index synthetic-events.c:1943
> >     #6 0x55a4878cb2a3 in record__synthesize builtin-record.c:2150
> >     #7 0x55a4878cdada in __cmd_record builtin-record.c:2963
> >     #8 0x55a4878d11ca in cmd_record builtin-record.c:4453
> >     #9 0x55a48795b3cc in run_builtin perf.c:349
> >     #10 0x55a48795b664 in handle_internal_command perf.c:401
> >     #11 0x55a48795b7bd in run_argv perf.c:448
> >     #12 0x55a48795bb06 in main perf.c:555
> >     #13 0x7f72aae33ca8 in __libc_start_call_main libc_start_call_main.h:74
> >     #14 0x7f72aae33d65 in __libc_start_main_alias_2 libc-start.c:128
> >     #15 0x55a4878acf41 in _start perf[52f41]
> > Segmentation fault
> > ```
>
>
> Hi, Ian，
>
> Is node-prefetch-misses a platform specific event? Running it on ARM Yitian 710
> and Intel SPR platform, I see:
>
> $sudo perf record -e node-prefetch-misses
> Error:
> The node-prefetch-misses event is not supported.

Hi Shuai,

So node-prefetch-misses is a legacy event. Perf has a notion of events
that are inbuilt to the kernel/PMU driver and get special fixed
encodings. That said, the PMU driver in the kernel can just fail to
support the events and I think that's uniformly the case for
node-prefetch-misses. As shown by my reproduction of the crash, which
I hope this suffices for a test - i.e. it is an event that parses but
one that is never supported.

Thanks,
Ian

Powered by blists - more mailing lists