Re: [PATCH bpf-next] selftests/bpf: fix usleep() implementation

[Andrii Nakryiko <andrii.nakryiko@gmail.com>]

On Thu, Mar 12, 2020 at 11:45 PM Alexei Starovoitov
<alexei.starovoitov@gmail.com> wrote:
>
> On Thu, Mar 12, 2020 at 11:18:37PM -0700, Andrii Nakryiko wrote:
> > nanosleep syscall expects pointer to struct timespec, not nanoseconds
> > directly. Current implementation fulfills its purpose of invoking nanosleep
> > syscall, but doesn't really provide sleeping capabilities, which can cause
> > flakiness for tests relying on usleep() to wait for something.
> >
> > Fixes: ec12a57b822c ("selftests/bpf: Guarantee that useep() calls nanosleep() syscall")
> > Signed-off-by: Andrii Nakryiko <andriin@fb.com>
> > ---
> >  tools/testing/selftests/bpf/test_progs.c | 11 ++++++++++-
> >  1 file changed, 10 insertions(+), 1 deletion(-)
> >
> > diff --git a/tools/testing/selftests/bpf/test_progs.c b/tools/testing/selftests/bpf/test_progs.c
> > index 2b0bc1171c9c..b6201dd82edf 100644
> > --- a/tools/testing/selftests/bpf/test_progs.c
> > +++ b/tools/testing/selftests/bpf/test_progs.c
> > @@ -35,7 +35,16 @@ struct prog_test_def {
> >   */
> >  int usleep(useconds_t usec)
> >  {
> > -     return syscall(__NR_nanosleep, usec * 1000UL);
> > +     struct timespec ts;
> > +
> > +     if (usec > 999999) {
> > +             ts.tv_sec = usec / 1000000;
> > +             ts.tv_nsec = usec % 1000000;
> > +     } else {
> > +             ts.tv_sec = 0;
> > +             ts.tv_nsec = usec;
> > +     }
> > +     return nanosleep(&ts, NULL);
> >  }
>
> Is this a copy-paste from somewhere?

nope, my very own prematurely optimized implementation :)

> Above 'if' looks like premature optimization.
> I applied it anyway, since it fixes flakiness in test_progs -n 24.
> Now pin*tp* tests are stable.
>

Great, I hoped as much.

> But the other one is still flaky:
> server_thread:FAIL:237
> Failed to accept client: Resource temporarily unavailable
> #64 tcp_rtt:FAIL
> Note that if I run the test alone (test_progs -n 64) it is stable.
> It fails only when run as part of bigger test_progs.
> test_progs -n 30-64 sporadically fails (most of the time)
> test_progs -n 40-64 consistently passes
> Haven't bisected further.

Okey, I'll get to it once I'm done fixing a bunch of other problems.
Seems like tcp_rtt needs some more love, sigh... :(