Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 2620:137:e000::1:20 as permitted sender) client-ip=2620:137:e000::1:20;
MIME-Version: 1.0
References: <20230614151625.2077-1-yangjihong1@huawei.com> <CAP-5=fUf0+7HwZ+AHUR0nRD5QnfPn9_CPMEdJZP_5goPfrPB+Q@mail.gmail.com>
 <CAP-5=fVOXjjcusjv858SOGrnNgE2w2sb7zS=0sZUpdFfR1T_GA@mail.gmail.com> <668a6159-b7a8-ed25-d8fa-5584a4c04d37@huawei.com>
In-Reply-To: <668a6159-b7a8-ed25-d8fa-5584a4c04d37@huawei.com>
From:   Ian Rogers <irogers@google.com>
Date:   Wed, 14 Jun 2023 19:04:23 -0700
Message-ID: <CAP-5=fX6nWRboZ2dWc1h_y0pe5TAgs0QC0qwCeaLcALfJ+5YEA@mail.gmail.com>
Subject: Re: [PATCH] perf top & record: Fix segfault when default cycles event
 is not supported
To:     Yang Jihong <yangjihong1@huawei.com>
Cc:     peterz@infradead.org, mingo@redhat.com, acme@kernel.org,
        mark.rutland@arm.com, alexander.shishkin@linux.intel.com,
        jolsa@kernel.org, namhyung@kernel.org, adrian.hunter@intel.com,
        kan.liang@linux.intel.com, linux-perf-users@vger.kernel.org,
        linux-kernel@vger.kernel.org
Content-Type: text/plain; charset="UTF-8"
Content-Transfer-Encoding: quoted-printable
Precedence: bulk

On Wed, Jun 14, 2023 at 6:55=E2=80=AFPM Yang Jihong <yangjihong1@huawei.com=
> wrote:
>
> Hello,
>
> On 2023/6/15 6:03, Ian Rogers wrote:
> > On Wed, Jun 14, 2023 at 9:18=E2=80=AFAM Ian Rogers <irogers@google.com>=
 wrote:
> >>
> >> On Wed, Jun 14, 2023 at 8:18=E2=80=AFAM Yang Jihong <yangjihong1@huawe=
i.com> wrote:
> >>>
> >>> The perf-record and perf-top call parse_event() to add a cycles event=
 to
> >>> an empty evlist. For the system that does not support hardware cycles
> >>> event, such as QEMU, the evlist is empty due to the following code pr=
ocess:
> >>>
> >>>    parse_event(evlist, "cycles:P" or ""cycles:Pu")
> >>>      parse_events(evlist, "cycles:P")
> >>>        __parse_events
> >>>          ...
> >>>          ret =3D parse_events__scanner(str, &parse_state);
> >>>          // ret =3D 0
> >>>          ...
> >>>          ret2 =3D parse_events__sort_events_and_fix_groups()
> >>>          if (ret2 < 0)
> >>>            return ret;
> >>>          // The cycles event is not supported, here ret2 =3D -EINVAL,
> >>>          // Here return 0.
> >>>          ...
> >>>          evlist__splice_list_tail(evlist)
> >>>          // The code here does not execute to, so the evlist is still=
 empty.
> >>>
> >>> A null pointer occurs when the content in the evlist is accessed late=
r.
> >>>
> >>> Before:
> >>>
> >>>    # perf list hw
> >>>
> >>>    List of pre-defined events (to be used in -e or -M):
> >>>
> >>>    # perf record true
> >>>    libperf: Miscounted nr_mmaps 0 vs 1
> >>>    WARNING: No sample_id_all support, falling back to unordered proce=
ssing
> >>>    perf: Segmentation fault
> >>>    Obtained 1 stack frames.
> >>>    [0xc5beff]
> >>>    Segmentation fault
> >>>
> >>> Solution:
> >>>    If cycles event is not supported, try to fall back to cpu-clock ev=
ent.
> >>>
> >>> After:
> >>>    # perf record true
> >>>    [ perf record: Woken up 1 times to write data ]
> >>>    [ perf record: Captured and wrote 0.006 MB perf.data ]
> >>>    #
> >>>
> >>> Fixes: 7b100989b4f6 ("perf evlist: Remove __evlist__add_default")
> >>> Signed-off-by: Yang Jihong <yangjihong1@huawei.com>
> >>
> >> Thanks, useful addition. The cpu-clock fall back wasn't present before
> >> 7b100989b4f6 so is the fixes tag correct?
> >
> > Hmm... it should be coming from evsel__fallback:
> > https://git.kernel.org/pub/scm/linux/kernel/git/acme/linux.git/tree/too=
ls/perf/util/evsel.c?h=3Dtmp.perf-tools-next#n2840
> > so we shouldn't duplicate that logic. The question is why we're not
> > doing the fallback.
> >
>
> Yes, it's a bit of the same logic as evsel__fallback, or we can call
> evlist__add_default() as before, simply create an evsel of hardware
> cycles and add it directly to evlist.
>
> Please confirm whether this solution is feasible. If it is feasible, I
> will send a v2 version.

The previous evlist__add_default logic didn't handle wildcard PMUs for
cycles, hence wanting to reuse the parse events logic. The error is
that the logic now isn't doing the fallback properly. I think an
evlist__add_cycles which uses evsel__fallback makes sense matching the
previous logic. I'd be happy if you took a look. I'll write a patch so
that the perf_pmus list of core PMUs is never empty.

Thanks,
Ian

> Thanks,
> Yang