[igt-dev] [PATCH i-g-t] tests/perf: add a test for OA data polling reads using "small" buffers

[Dixit, Ashutosh]

On Tue, 31 Mar 2020 00:48:45 -0700, Dixit, Ashutosh wrote:
>
> On Tue, 31 Mar 2020 00:36:54 -0700, Lionel Landwerlin wrote:
> >
> > On 31/03/2020 09:06, Dixit, Ashutosh wrote:
> > > On Fri, 27 Mar 2020 12:49:22 -0700, Dixit, Ashutosh wrote:
> > >> On Fri, 27 Mar 2020 12:06:13 -0700, Lionel Landwerlin wrote:
> > >>> On 27/03/2020 21:03, Dixit, Ashutosh wrote:
> > >>>> On Fri, 27 Mar 2020 09:09:41 -0700, Lionel Landwerlin wrote:
> > >>>>> On 27/03/2020 06:50, Dixit, Ashutosh wrote:
> > >>>>>> On Thu, 26 Mar 2020 21:42:50 -0700, Ashutosh Dixit wrote:
> > >>>>>>> diff --git a/tests/perf.c b/tests/perf.c
> > >>>>>>> index 724f6f809..3dc757c3b 100644
> > >>>>>>> --- a/tests/perf.c
> > >>>>>>> +++ b/tests/perf.c
> > >>>>>>> +static void test_polling_small_buf(void)
> > >>>>>>> +{
> > >>>>>> /snip/
> > >>>>>>
> > >>>>>>> +
> > >>>>>>> +	igt_assert(abs(n_expect_read_bytes - n_bytes_read) < (0.10 * n_expect_read_bytes));
> > >>>>>>> +}
> > >>>>>>> +
> > >>>>>> I'd be wary of a 90% match on slow platforms like Atom? Maybe 80% is safer?
> > >>>>> Do we have any experiment showing them behaving differently?
> > >>>> No I don't have any data, but considering that in previous stable versions
> > >>>> we can only read < 10% of the data, I think it should be ok to go down to
> > >>>> 80%? So that we don't start getting unnecessary false alarms in CI, even
> > >>>> when the issue is fixed.
> > >>> Okay, for the record I get somewhere between 93~95% of expected reports on
> > >>> KBLGT2.
> > >> Yes I tried it and saw that. I already gave a R-b so we could probably
> > >> merge the patch after making that change (0.2 instead of 0.1 above), or do
> > >> you want me to post a new version with the change? Thanks!
> > >
> > > Actually there has been some change in the kernel, earlier like you I was
> > > also getting around 94% with a 1 KB buffer, now I am getting about
> > > 87%. I am getting 94% with a 1 MB buffer. Does the amount of expected data
> > > in the test need to be modified? I can try to bisect tomorrow and see what
> > > has done this, unless you already know. Thanks!

I haven't been able to easily narrow this down, I am not sure if it is
worth spending more time on it. However since v3 of the test only requires
a 80% match the test passes with kernel patch : "drm/i915/perf: Do not
clear pollin for small user read buffers" and will fail on prior/stable
kernels. Thanks!

> >
> > Ah, that's probably the read() bug you're fixing...
> >
> > Are you running with the kernel patch : "drm/i915/perf: Do not clear pollin
> > for small user read buffers" ?
>
> Yes, I was just testing the patch before posting it and I chanced on
> this. Thanks!