Re: [PATCH v2 1/2] perf parse-events: fix memory leaks found on parse_events

From: Arnaldo Carvalho de Melo
Date: Sat May 02 2020 - 11:11:55 EST


Em Thu, Apr 30, 2020 at 02:39:14PM -0700, Ian Rogers escreveu:
> On Wed, Apr 29, 2020 at 10:54 AM Arnaldo Carvalho de Melo
> <arnaldo.melo@xxxxxxxxx> wrote:
> >
> > Em Wed, Mar 18, 2020 at 07:31:00PM -0700, Ian Rogers escreveu:
> > > Memory leaks found by applying LLVM's libfuzzer on the parse_events
> > > function.
> > >
> > > Signed-off-by: Ian Rogers <irogers@xxxxxxxxxx>
> > > ---
> > > tools/perf/util/parse-events.c | 2 ++
> > > tools/perf/util/parse-events.y | 3 ++-
> > > 2 files changed, 4 insertions(+), 1 deletion(-)
> > >
> > > diff --git a/tools/perf/util/parse-events.c b/tools/perf/util/parse-events.c
> > > index 593b6b03785d..1e0bec5c0846 100644
> > > --- a/tools/perf/util/parse-events.c
> > > +++ b/tools/perf/util/parse-events.c
> > > @@ -1482,6 +1482,8 @@ int parse_events_add_pmu(struct parse_events_state *parse_state,
> > >
> > > list_for_each_entry_safe(pos, tmp, &config_terms, list) {
> > > list_del_init(&pos->list);
> > > + if (pos->free_str)
> > > + free(pos->val.str);
> >
> > I'm applying it but only after changing it to zfree(&pos->free_str), to
> > make sure that any othe rcode that may still hold a pointer to pos will
> > see a NULL in ->free_str and crash sooner rather than later.
> >
> > > free(pos);
> > > }
> > > return -EINVAL;
> >
> > And the following should be in a different patch
> >
> > > diff --git a/tools/perf/util/parse-events.y b/tools/perf/util/parse-events.y
> > > index 94f8bcd83582..8212cc771667 100644
> > > --- a/tools/perf/util/parse-events.y
> > > +++ b/tools/perf/util/parse-events.y
> > > @@ -44,7 +44,7 @@ static void free_list_evsel(struct list_head* list_evsel)
> > >
> > > list_for_each_entry_safe(evsel, tmp, list_evsel, core.node) {
> > > list_del_init(&evsel->core.node);
> > > - perf_evsel__delete(evsel);
> > > + evsel__delete(evsel);
> > > }
> > > free(list_evsel);
> > > }
> >
> > And this one in another, I'll fix this up, but please try in the future
> > to provide different patches for different fixes, so that if we
> > eventually find out that one of the unrelated fixes is wrong, then we
> > can revert the patch more easily with 'git revert' instead of having to
> > do a patch that reverts just part of the bigger hodge-podge patch.
> >
> > If you go and have a track record of doing this as piecemeal as
> > possible, I will in turn feel more confident of processing your patches
> > in a faster fashion ;-) :-)
>
> Thanks, at some point I'd like to get libfuzzer with asan working for
> more than just me so that we don't backslide. It'd also make the
> reproductions easier to share.

If we can detect the presence of the needed components, libraries,
compiler with the right feature set, yeah, a 'perf test' built under
such environment surely would benefit from having further tests,

- Arnaldo