Re: perf seg fault

From: Jiri Olsa
Date: Tue May 12 2020 - 11:56:01 EST


On Tue, May 12, 2020 at 06:45:24PM +0300, Adrian Hunter wrote:
> On 12/05/20 6:10 pm, Jiri Olsa wrote:
> > On Tue, May 12, 2020 at 05:58:29PM +0300, Adrian Hunter wrote:
> >> Forgot to cc mailing list
> >>
> >> On 12/05/20 5:50 pm, Adrian Hunter wrote:
> >>> Hi
> >>>
> >>> I am getting a seg fault from your perf/core branch, as follows:
> >>>
> >>> # perf record uname
> >>> Linux
> >>> [ perf record: Woken up 1 times to write data ]
> >>> [ perf record: Captured and wrote 0.005 MB perf.data (7 samples) ]
> >>> perf: Segmentation fault
> >>> Obtained 6 stack frames.
> >>> [0x4e75b4]
> >>> [0x5d1ad0]
> >>> [0x5c9860]
> >>> [0x4a6e5c]
> >>> [0x5cb39b]
> >>> [0x76c89f]
> >>> Segmentation fault
> >>>
> >>> It goes away with --no-bpf-event:
> >>>
> >>> # perf record --no-bpf-event uname
> >>> Linux
> >>> [ perf record: Woken up 1 times to write data ]
> >>> [ perf record: Captured and wrote 0.005 MB perf.data (7 samples) ]
> >>> #
> >>>
> >>> kernel is from the same branch
> >>>
> >>> # uname -a
> >>> Linux buildroot 5.7.0-rc2-00028-g0fdddf5a583a #165 SMP Tue May 12 16:27:53
> >>> EEST 2020 x86_64 GNU/Linux
> >>>
> >>> # perf version --build-options
> >>> perf version 5.6.g0fdddf5a583a
> >>> dwarf: [ OFF ] # HAVE_DWARF_SUPPORT
> >>> dwarf_getlocations: [ OFF ] # HAVE_DWARF_GETLOCATIONS_SUPPORT
> >>> glibc: [ on ] # HAVE_GLIBC_SUPPORT
> >>> gtk2: [ OFF ] # HAVE_GTK2_SUPPORT
> >>> syscall_table: [ on ] # HAVE_SYSCALL_TABLE_SUPPORT
> >>> libbfd: [ on ] # HAVE_LIBBFD_SUPPORT
> >>> libelf: [ on ] # HAVE_LIBELF_SUPPORT
> >>> libnuma: [ OFF ] # HAVE_LIBNUMA_SUPPORT
> >>> numa_num_possible_cpus: [ OFF ] # HAVE_LIBNUMA_SUPPORT
> >>> libperl: [ OFF ] # HAVE_LIBPERL_SUPPORT
> >>> libpython: [ OFF ] # HAVE_LIBPYTHON_SUPPORT
> >>> libslang: [ on ] # HAVE_SLANG_SUPPORT
> >>> libcrypto: [ on ] # HAVE_LIBCRYPTO_SUPPORT
> >>> libunwind: [ OFF ] # HAVE_LIBUNWIND_SUPPORT
> >>> libdw-dwarf-unwind: [ OFF ] # HAVE_DWARF_SUPPORT
> >>> zlib: [ OFF ] # HAVE_ZLIB_SUPPORT
> >>> lzma: [ OFF ] # HAVE_LZMA_SUPPORT
> >>> get_cpuid: [ on ] # HAVE_AUXTRACE_SUPPORT
> >>> bpf: [ OFF ] # HAVE_LIBBPF_SUPPORT
> >>> aio: [ on ] # HAVE_AIO_SUPPORT
> >>> zstd: [ OFF ] # HAVE_ZSTD_SUPPORT
> >>>
> >>>
> >>> Any thoughts?
> >
> > hum, I don't see that, do you reproduce with DEBUG=1?
> > to get more verbose backtrace
>
> It will require a kernel with support for bpf events otherwise the
> --no-bpf-event option would have no effect.

ok, reproduced.. I think we need to think about better way,
but this fixes the issue for me now

jirka


---
diff --git a/tools/perf/builtin-record.c b/tools/perf/builtin-record.c
index 4d4502b7fea0..173012a25df3 100644
--- a/tools/perf/builtin-record.c
+++ b/tools/perf/builtin-record.c
@@ -2564,6 +2564,10 @@ int cmd_record(int argc, const char **argv)
if (quiet)
perf_quiet_option();

+#ifndef HAVE_LIBBPF_SUPPORT
+ rec->opts.no_bpf_event = true;
+#endif
+
/* Make system wide (-a) the default target. */
if (!argc && target__none(&rec->opts.target))
rec->opts.target.system_wide = true;