Re: [PATCH v2] perf record: Fix coredump with --overwrite and --max-size

[Namhyung Kim]

On Sun, Jan 8, 2023 at 6:47 PM Yang Jihong <yangjihong1@xxxxxxxxxx> wrote:
>
> Hello,
>
> On 2023/1/7 5:12, Namhyung Kim wrote:
> > Hello,
> >
> > On Wed, Jan 4, 2023 at 8:09 PM Yang Jihong <yangjihong1@xxxxxxxxxx> wrote:
> >>
> >> Hello,
> >>
> >> On 2023/1/4 0:50, Namhyung Kim wrote:
> >>> On Mon, Jan 2, 2023 at 8:20 AM Arnaldo Carvalho de Melo <acme@xxxxxxxxxx> wrote:
> >>>>
> >>>> Em Thu, Dec 29, 2022 at 12:47:28PM +0000, Yang Jihong escreveu:
> >>>>> When --overwrite and --max-size options of perf record are used together,
> >>>>> a segmentation fault occurs. The following is an example:
> >>>>>
> >>>>>    # perf record -e sched:sched* --overwrite --max-size 1M -a -- sleep 1
> >>>>>     [ perf record: Woken up 1 times to write data ]
> >>>>>     perf: Segmentation fault
> >>>>>     Obtained 1 stack frames.
> >>>>>     [0xc4c67f]
> >>>>>     Segmentation fault (core dumped)
> >>>>>
> >>>>> backtrace of the core file is as follows:
> >>>>>
> >>>>>     #0  0x0000000000417990 in process_locked_synthesized_event (tool=0x0, event=0x15, sample=0x1de0, machine=0xf8) at builtin-record.c:630
> >>>>>     #1  0x000000000057ee53 in perf_event__synthesize_threads (nr_threads_synthesize=21, mmap_data=<optimized out>, needs_mmap=<optimized out>, machine=0x17ad9b0, process=<optimized out>, tool=0x0) at util/synthetic-events.c:1950
> >>>>>     #2  __machine__synthesize_threads (nr_threads_synthesize=0, data_mmap=<optimized out>, needs_mmap=<optimized out>, process=<optimized out>, threads=0x8, target=0x8, tool=0x0, machine=0x17ad9b0) at util/synthetic-events.c:1936
> >>>>>     #3  machine__synthesize_threads (machine=0x17ad9b0, target=0x8, threads=0x8, needs_mmap=<optimized out>, data_mmap=<optimized out>, nr_threads_synthesize=0) at util/synthetic-events.c:1947
> >>>>>     #4  0x000000000040165d in record__synthesize (tail=<optimized out>, rec=0xbe2520 <record>) at builtin-record.c:2010
> >>>>>     #5  0x0000000000403989 in __cmd_record (argc=<optimized out>, argv=<optimized out>, rec=0xbe2520 <record>) at builtin-record.c:2810
> >>>>>     #6  0x00000000004196ba in record__init_thread_user_masks (rec=0xbe2520 <record>, cpus=0x17a65f0) at builtin-record.c:3837
> >>>>>     #7  record__init_thread_masks (rec=0xbe2520 <record>) at builtin-record.c:3938
> >>>>>     #8  cmd_record (argc=1, argv=0x7ffdd692dc60) at builtin-record.c:4241
> >>>>>     #9  0x00000000004b701d in pager_command_config (var=0x0, value=0x15 <error: Cannot access memory at address 0x15>, data=0x1de0) at perf.c:117
> >>>>>     #10 0x00000000004b732b in get_leaf_frame_caller_aarch64 (sample=0xfffffffb, thread=0x0, usr_idx=<optimized out>) at util/arm64-frame-pointer-unwind-support.c:56
> >>>>>     #11 0x0000000000406331 in execv_dashed_external (argv=0x7ffdd692d9e8) at perf.c:410
> >>>>>     #12 run_argv (argcp=<synthetic pointer>, argv=<synthetic pointer>) at perf.c:431
> >>>>>     #13 main (argc=<optimized out>, argv=0x7ffdd692d9e8) at perf.c:562
> >>>>>
> >>>>> The reason is that record__bytes_written accesses the freed memory rec->thread_data,
> >>>>> The process is as follows:
> >>>>>     __cmd_record
> >>>>>       -> record__free_thread_data
> >>>>>         -> zfree(&rec->thread_data)         // free rec->thread_data
> >>>>>       -> record__synthesize
> >>>>>         -> perf_event__synthesize_id_index
> >>>>>           -> process_synthesized_event
> >>>>>             -> record__write
> >>>>>               -> record__bytes_written     // access rec->thread_data
> >>>>>
> >>>>> we only need to check the value of done first.
> >>>>> Also add variable check in record__bytes_written for code hardening,
> >>>>> and save bytes_written separately to reduce one calculation.
> >>>>>
> >>>>> Fixes: 6d57581659f7 ("perf record: Add support for limit perf output file size")
> >>>>> Signed-off-by: Yang Jihong <yangjihong1@xxxxxxxxxx>
> >>>>> ---
> >>>>>
> >>>>> Changes since v1:
> >>>>>    - Add variable check in record__bytes_written for code hardening.
> >>>>>    - Save bytes_written separately to reduce one calculation.
> >>>>>    - Remove rec->opts.tail_synthesize check.
> >>>>
> >>>> Namhyung, are you ok with this now?
> >>>>
> >>>> - Arnaldo
> >>>>
> >>>>>    tools/perf/builtin-record.c | 26 +++++++++++++++++---------
> >>>>>    1 file changed, 17 insertions(+), 9 deletions(-)
> >>>>>
> >>>>> diff --git a/tools/perf/builtin-record.c b/tools/perf/builtin-record.c
> >>>>> index 29dcd454b8e2..acba9e43e519 100644
> >>>>> --- a/tools/perf/builtin-record.c
> >>>>> +++ b/tools/perf/builtin-record.c
> >>>>> @@ -230,16 +230,29 @@ static u64 record__bytes_written(struct record *rec)
> >>>>>         u64 bytes_written = rec->bytes_written;
> >>>>>         struct record_thread *thread_data = rec->thread_data;
> >>>>>
> >>>>> +     if (thread_data == NULL)
> >>>>> +             return bytes_written;
> >>>>> +
> >>>
> >>> Then it won't count bytes written by threads, right?
> >>> I think it needs to be saved somewhere.
> >>>
> >> I'm not sure here. Can you explain it more clearly, thanks :)
> >> I can modify it accordingly.
> >>
> >> I think if thread_data == NULL, it is not thread data.
> >> In this case, we just return rec->bytes_written.
> >
> > It can be thread data but freed before tail synthesis, right?
> > In that case, I think it needs to add bytes_written by threads
> > to calculate the correct data size.
> Em... In the __cmd_record function, record__stop_threads is called
> before record__free_thread_data, so if the thread has been freed, there
> will be no thread data.
> I think it's okay to ignore the situation you mentioned above.

Right, the thread data is already freed, but we need the size.

I think it didn't (and won't) update to rec->bytes_written for the data
written by the threads (data.X file) because it's only for the main
'data' file.  So record__bytes_written() will return a smaller number
after the threads are gone.  But I think it should return the total
data size.

Thanks,
Namhyung