Re: [PATCH] perf trace: Fix SIGSEGV when processing augmented args

[Arnaldo Carvalho de Melo <acme@kernel.org>]

Em Wed, Jul 06, 2022 at 06:01:29PM +0530, Naveen N. Rao escreveu:
> Hi Arnaldo,
> 
> Arnaldo Carvalho de Melo wrote:
> > Em Tue, Mar 15, 2022 at 02:52:05PM -0300, Arnaldo Carvalho de Melo escreveu:
> > > Em Tue, Mar 15, 2022 at 10:57:57PM +0530, Naveen N. Rao escreveu:
> > > > Yes, it looks like the current check in 'perf' isn't working. The below
> > > > patch also resolves the crash we are seeing:
> > > > diff --git a/tools/perf/builtin-trace.c b/tools/perf/builtin-trace.c
> > > > index 2f1d20553a0aa3..86b459f4ebdd61 100644
> > > > --- a/tools/perf/builtin-trace.c
> > > > +++ b/tools/perf/builtin-trace.c
> > > > @@ -2326,7 +2326,7 @@ static int trace__sys_enter(struct trace *trace, struct evsel *evsel,
> > > >         * thinking that the extra 2 u64 args are the augmented filename, so just check
> > > >         * here and avoid using augmented syscalls when the evsel is the raw_syscalls one.
> > > >         */
> > > > -       if (evsel != trace->syscalls.events.sys_enter)
> > > > +       if (strcmp(evsel__name(evsel), "raw_syscalls:sys_enter"))
> > > >                augmented_args = syscall__augmented_args(sc, sample, &augmented_args_size, trace->raw_augmented_syscalls_args_size);
> > > >        ttrace->entry_time = sample->time;
> > > >        msg = ttrace->entry_str;
> > > 
> > > Interesting, that should be equivalent :-\ humm, not really, understood,
> > > when processing perf.data files we don't setup
> > > trace->syscalls.events.sys_enter...
> > > 
> > > switching from strcmp() to something cheaper but equivalent should be
> > > the fix for now.
> > 
> > I'll add a trace->use_augmented_args boolean that will do this test
> > once, and then use it in this case and will audit to check if this
> > should be used in other places.
> 
> Does something like the below look reasonable?
> 
> I know this isn't quite what you proposed, but it fixes the problem for me
> while avoiding the need for a string comparison. I also think this addresses
> all uses in 'perf trace', though I didn't audit the need for a similar fix
> elsewhere in 'perf'.

I'll check and test, but I like it, two-liner even :-)

- Arnaldo
 
> 
> Thanks,
> Naveen
> 
> ---
> diff --git a/tools/perf/builtin-trace.c b/tools/perf/builtin-trace.c
> index d2de2a4073e7eb..352b88a51dec2d 100644
> --- a/tools/perf/builtin-trace.c
> +++ b/tools/perf/builtin-trace.c
> @@ -4291,6 +4292,8 @@ static int trace__replay(struct trace *trace)
>                goto out;
>        }
> 
> +       trace->syscalls.events.sys_enter = evsel;
> +
>        evsel = evlist__find_tracepoint_by_name(session->evlist, "raw_syscalls:sys_exit");
>        if (evsel == NULL)
>                evsel = evlist__find_tracepoint_by_name(session->evlist, "syscalls:sys_exit");
> @@ -4301,6 +4304,8 @@ static int trace__replay(struct trace *trace)
>                goto out;
>        }
> 
> +       trace->syscalls.events.sys_exit  = evsel;
> +
>        evlist__for_each_entry(session->evlist, evsel) {
>                if (evsel->core.attr.type == PERF_TYPE_SOFTWARE &&
>                    (evsel->core.attr.config == PERF_COUNT_SW_PAGE_FAULTS_MAJ ||
> 
> 

-- 

- Arnaldo