Re: [PATCH] tst: Provide test for select

[Adhemerval Zanella <adhemerval.zanella@linaro.org>]

On 24/03/2021 18:21, H.J. Lu wrote:
> On Wed, Mar 24, 2021 at 2:14 PM Adhemerval Zanella
> <adhemerval.zanella@linaro.org> wrote:
>>
>>
>>
>> On 24/03/2021 17:13, H.J. Lu wrote:
>>> On Wed, Mar 24, 2021 at 12:47 PM Adhemerval Zanella
>>> <adhemerval.zanella@linaro.org> wrote:
>>>>
>>>>
>>>>
>>>> On 24/03/2021 16:17, H.J. Lu wrote:
>>>>> On Mon, Mar 22, 2021 at 4:33 AM Lukasz Majewski <lukma@denx.de> wrote:
>>>>>>
>>>>>> Dear Community,
>>>>>>
>>>>>>> This change adds new test to assess select()'s timeout related
>>>>>>> functionality (the rdfs set provides valid fd - stderr - but during
>>>>>>> normal program operation there is no data to be read, so one just
>>>>>>> waits for timeout).
>>>>>>>
>>>>>>> To be more specific - two use cases are checked:
>>>>>>> - if select() times out immediately when passed struct timeval has
>>>>>>>   zero values of tv_usec and tv_sec.
>>>>>>> - if select() times out after timeout specified in passed argument
>>>>>>
>>>>>> Do you have any comments regarding this patch?
>>>>>>
>>>>>
>>>>> This test failed on machines with more than 40 cores:
>>>>>
>>>>> tst-select.c:54: 1616610088.851713938s not after 1616610089.851712804s
>>>>> (difference 0.999998866s)
>>>>> error: 1 test failures
>>>>>
>>>>> I was doing 3 "makec -j28 check" in parallel.
>>>>
>>>> I think the nanosecond precision of time accounting is triggering the
>>>> failure, since select only support timeval (the error indicates that
>>>> the nanosecond precision is what is triggering it).
>>>>
>>>> Maybe if we ignore the nanosecond precision:
>>>>
>>>> diff --git a/misc/tst-select.c b/misc/tst-select.c
>>>> index 7c310256c5..4b1791ac8a 100644
>>>> --- a/misc/tst-select.c
>>>> +++ b/misc/tst-select.c
>>>> @@ -45,6 +45,8 @@ test_select_timeout (bool zero_tmo)
>>>>    to.tv_sec = timeout;
>>>>    ts = xclock_now (CLOCK_REALTIME);
>>>>    ts = timespec_add (ts, (struct timespec) { timeout, 0 });
>>>> +  /* Ignore nanosecond precision since select only support microsecond.  */
>>>> +  ts.tv_nsec = (ts.tv_nsec * 1000) / 1000;
>>>
>>> How does it work?  It looks like a NOP to me.
>>
>> The idea is to clear the nanosecond precision from the xclock_now call.
> 
> I tried it.  It doesn't solve the problem.  I opened:
> 
> https://sourceware.org/bugzilla/show_bug.cgi?id=27648

I think we will need to clear out on the TEST_TIMESPEC_NOW_OR_AFTER
as well (by adding another macro).