Re: [Qemu-devel] Remaining CI failures

["Alex Bennée" <alex.bennee@linaro.org>]

Emilio G. Cota <cota@braap.org> writes:

> On Fri, Jan 11, 2019 at 19:10:07 +0000, Alex Bennée wrote:
>> So trying to narrow down the remaining failures in the CI system. There
>> is one with a patch in flight (use g_usleep instead of sleep) but there
>> remains two failure modes, both erratic.
>>
>> tests/qht-par:
>>
>> I can trigger this on my dev machine with a gprof enabled build:
>>
>>   # QEMU configure log Fri Jan 11 14:10:45 GMT 2019
>>   # Configured with: './configure' '--disable-tools' '--disable-docs' '--enable-gprof' '--enable-gcov'
>>
>> I only seem to be able to trigger it when running via the wrapper in the
>> make system:
>>
>>   retry.py -n 30 --invert make check-tests/test-qht-par
>>
>> Eventually this crashes with:
>>
>>   ERROR:tests/test-qht-par.c:20:test_qht: assertion failed (rc == 0): (35584 == 0)
>>
>> Leaving a core dump for the child:
>
> I can't replicate this on my machine :-(
>
> I suspect this is probably related to the fact that gprof
> is not even meant to support multi-threaded programs (like qht-bench).
>
> Given that we've fixed the sleep issue (thanks!) and that there is no use
> in running test-qht-par under gprof, I propose to permanently skip
> it under gprof, e.g.:
>
> --- a/tests/Makefile.include
> +++ b/tests/Makefile.include
> @@ -88,7 +88,8 @@ check-unit-y += tests/test-rcu-simpleq$(EXESUF)
>  check-unit-y += tests/test-rcu-tailq$(EXESUF)
>  check-unit-y += tests/test-qdist$(EXESUF)
>  check-unit-y += tests/test-qht$(EXESUF)
> -# FIXME: {test-qht-par + gprof} often break on Travis CI
> +# test-qht-par invokes qht-bench, which is multi-threaded.
> +# gprof doesn't support multi-threaded programs, so skip this test under gprof.
>  check-unit-$(call lnot,$(CONFIG_GPROF)) += tests/test-qht-par$(EXESUF)
>  check-unit-y += tests/test-bitops$(EXESUF)
>  check-unit-y += tests/test-bitcnt$(EXESUF)
>
> If you agree, I can submit a proper patch with the above.

Heh, that fix is already in although I reverted the work-around in my PR
as it seemed a little too hacky given it seems to be fixed now.

--
Alex Bennée