Re: [PATCH v6 03/20] perf record: Introduce thread local variable

[Namhyung Kim]

Hi Riccardo,

On Thu, Jun 3, 2021 at 3:56 PM Riccardo Mancini <rickyman7@xxxxxxxxx> wrote:
>
> Hi,
>
> thank you very much for your work for adding threading capabilites to perf
> record.
> I did some testing on your entire patchset, especially checking for memory
> issues using ASan. This is just the first of a couple of emails to point out
> some issues I found.
> I will also do additional tests in the future.
>
> On Wed, 2021-05-26 at 13:52 +0300, Alexey Bayduraev wrote:
> SNIP
> > @@ -2220,18 +2275,20 @@ static int __cmd_record(struct record *rec, int argc,
> > const char **argv)
> >                 goto out_child;
> >         }
> >
> > -       if (!quiet)
> > -               fprintf(stderr, "[ perf record: Woken up %ld times to write data
> > ]\n", waking);
> > -
> >         if (target__none(&rec->opts.target))
> >                 record__synthesize_workload(rec, true);
> >
> >  out_child:
> > +       record__stop_threads(rec, &waking);
> > +out_free_threads:
> >         record__free_thread_data(rec);
> >         evlist__finalize_ctlfd(rec->evlist);
> >         record__mmap_read_all(rec, true);
> >         record__aio_mmap_read_sync(rec);
>
> record__mmap_read_all should be moved before record__free_thread_data since it
> uses the thread_data that's just been freed.
> Furthermore, record__mmap_read_all should also be moved before the
> out_free_threads label, since it cannot be called unless record__start_threads
> succeeded, otherwise thread would be NULL and will cause a segfault (it happens
> if there is an error somewhere else in perf, for example).
>
> In my tests the following order works, but it should be double checked for
> possible side-effects of this order change.
>
> out_child:
>         record__stop_threads(rec, &waking);
>         record__mmap_read_all(rec, true);
> out_free_threads:
>         record__free_thread_data(rec);
>         evlist__finalize_ctlfd(rec->evlist);
>         record__aio_mmap_read_sync(rec);

I wonder how it worked before.. maybe we should place
record__free_thread_data() far below.

Thanks,
Namhyung