[dpdk-dev] Random failure in service_autotest

[Aaron Conole]

David Marchand <david.marchand at redhat.com> writes:

> On Wed, Jul 15, 2020 at 3:09 PM Lukasz Wojciechowski
> <l.wojciechow at partner.samsung.com> wrote:
>>
>>
>> W dniu 15.07.2020 o 15:02, David Marchand pisze:
>> > On Wed, Jul 15, 2020 at 2:56 PM Aaron Conole <aconole at redhat.com> wrote:
>> >> I guess the service_lcore_attr_get failed, but with no useful
>> Why do you suspect service_lcore_attr_get() ?
>
> https://travis-ci.com/github/ovsrobot/dpdk/jobs/361097992#L18697
>
> RTE>>service_autotest
>  + ------------------------------------------------------- +
>  + Test Suite : service core test suite
>  + ------------------------------------------------------- +
>  + TestCase [ 0] : unregister_all succeeded
>  + TestCase [ 1] : service_name succeeded
>  + TestCase [ 2] : service_get_by_name succeeded
> Service dummy_service Summary
>   dummy_service: stats 1 calls 0 cycles 0 avg: 0
> Service dummy_service Summary
>   dummy_service: stats 0 calls 0 cycles 0 avg: 0
>  + TestCase [ 3] : service_dump succeeded
>  + TestCase [ 4] : service_attr_get succeeded
>
> ***
>  + TestCase [ 5] : service_lcore_attr_get failed
> ***
>
>  + TestCase [ 6] : service_probe_capability succeeded
>  + TestCase [ 7] : service_start_stop succeeded
>  + TestCase [ 8] : service_lcore_add_del skipped
>  + TestCase [ 9] : service_lcore_start_stop succeeded
>  + TestCase [10] : service_lcore_en_dis_able succeeded
>  + TestCase [11] : service_mt_unsafe_poll skipped
>  + TestCase [12] : service_mt_safe_poll skipped
> perf test for MT Safe: 56.9 cycles per call
>  + TestCase [13] : service_app_lcore_mt_safe succeeded
> perf test for MT Unsafe: 83.4 cycles per call
>  + TestCase [14] : service_app_lcore_mt_unsafe succeeded
>  + TestCase [15] : service_may_be_active succeeded
>  + TestCase [16] : service_active_two_cores skipped
>  + ------------------------------------------------------- +
>
>
>> >> information.  We should have the test suite app turn the log level all
>> >> the way up.  I realize it will increase the log data even higher, but
>> >> without it, we have no idea why this test failed.
>> >>
>> >> Something like the following
>> >> ---
>> >>
>> >> diff --git a/app/test/test.c b/app/test/test.c
>> >> index 94d26ab1f6..c47cb075f9 100644
>> >> --- a/app/test/test.c
>> >> +++ b/app/test/test.c
>> >> @@ -150,6 +150,9 @@ main(int argc, char **argv)
>> >>
>> >>          prgname = argv[0];
>> >>
>> >> +       rte_log_set_global_level(RTE_LOG_DEBUG);
>> >> +       rte_log_set_level(RTE_LOGTYPE_EAL, RTE_LOG_DEBUG);
>> >> +
>> >>          recursive_call = getenv(RECURSIVE_ENV_VAR);
>> >>          if (recursive_call != NULL) {
>> >>                  ret = do_recursive_call();
>> >> ---
>> >>
>> >> That way we can at least debug when it happens.
>> > Debugging is one thing.
>> > But here service_lcore_attr_get() has a lot of asserts that should
>> > trigger a straight error.
>> >
>> Yes, but without debugs enabled, the assert message won't be printed out.
>
> Changing the whole debug levels could have side effects on the
> libraries being tested: maybe hide races (too bad) or reveal races
> (that would be interesting ;-)).

6 of one, 1/2 dozen of the other.

Maybe there are bugs where people run code inside log level tests that
get omitted otherwise.

> On the other hand, what I am saying is that using debug level logs for
> test asserts might not be the best solution.

Yeah, that's also a good idea :)