Re: [PATCH 5/6] nptl: Convert some rwlock tests to use libsupport

[Florian Weimer <fweimer at redhat dot com>]

* Mike Crowe:

> On Friday 05 April 2019 at 13:05:55 +0200, Florian Weimer wrote:
>> * Mike Crowe:
>> 
>> >    if (pthread_rwlock_wrlock (&r) != 0)
>> 
>> >    if (pthread_barrier_init (&b, NULL, 2) != 0)
>> 
>> We have xpthread_rwlock_wrlock and xpthread_barrier_init, which will
>> perform the error checking for you.  Similarly for a few more functions.
>
> I swapped a few emails with Adhemerval Zanella about that. To me, it felt
> like the wrapper functions weren't appropriate when they wrap the very
> functions that the test is supposed to be testing. In particular, they
> don't provide line number information for the failure. On that basis, the
> code you quoted would become:
>
>  TEST_COMPARE (pthread_rwlock_wrlock (&r), 0);
>  xpthread_barrier_init (&b, NULL, 2);
>
> because the file in question is testing rwlock, but can assume that
> barriers ought to work.

I don't have a strong opinion here.  I think we should have proper error
testing in tests, and also report a little bit more detail on unexpected
errors.  The TEST_COMPARE approach actually provides more information
than the x* functions (because the latter do not show the location of
the call), so that's fine as well.

> True. I thought that the extra messages were useful, but it looks like
> TEST_COMPARE provides enough context.

Right.

> However, I'm less sure about this code from tst-rwlock7.c(do_test):
>
>   size_t cnt;
>   for (cnt = 0; cnt < sizeof (kind) / sizeof (kind[0]); ++cnt)
>     {
>       pthread_rwlock_t r;
>       pthread_rwlockattr_t a;
>       if (pthread_rwlockattr_init (&a) != 0)
>         FAIL_EXIT1 ("round %Zu: rwlockattr_t failed\n", cnt);
>
>       if (pthread_rwlockattr_setkind_np (&a, kind[cnt]) != 0)
>         FAIL_EXIT1 ("round %Zu: rwlockattr_setkind failed\n", cnt);
>     ...
>
> If the round numbers are important then I need to keep the messages. If
> not, then I can also replace these with TEST_COMPARE. Do you have an
> opinion on that?

I don't think I've ever seen tst-rwlock7 fail, so I don't know if the
round numbers are useful for diagnosing test failures.

Thanks,
Florian