Re: [PATCH v5 9/9] drm: selftest: convert drm_mm selftest to KUnit

[Maxime Ripard]

Hi David,

On Tue, Jul 25, 2023 at 05:54:32PM +0800, David Gow wrote:
> On Tue, 25 Jul 2023 at 16:38, Maxime Ripard <mripard@xxxxxxxxxx> wrote:
> > On Thu, Apr 27, 2023 at 03:14:39PM +0200, Maxime Ripard wrote:
> > > Hi,
> > >
> > > On Fri, Jul 08, 2022 at 05:30:52PM -0300, Maíra Canal wrote:
> > > > From: Arthur Grillo <arthur.grillo@xxxxxx>
> > > >
> > > > Considering the current adoption of the KUnit framework, convert the
> > > > DRM mm selftest to the KUnit API.
> > > >
> > > > Signed-off-by: Arthur Grillo <arthur.grillo@xxxxxx>
> > > > Tested-by: David Gow <davidgow@xxxxxxxxxx>
> > > > Acked-by: Daniel Latypov <dlatypov@xxxxxxxxxx>
> > > > Reviewed-by: Javier Martinez Canillas <javierm@xxxxxxxxxx>
> > > > Signed-off-by: Maíra Canal <maira.canal@xxxxxx>
> > >
> > > I'm very late to the party, but I'd like to discuss that patch some more.
> > >
> > > Two tests (drm_test_mm_reserve, drm_test_mm_insert) in it take a super
> > > long time to run (about 30s each on my machine).
> > >
> > > If we run all the DRM tests and VC4 tests, each of those two are longer
> > > to run than all the ~300 tests combined. About 100 times longer.
> > >
> > > I don't think that running for so long is reasonable, and for multiple
> > > reasons:
> > >
> > >   - While I don't know drm_mm well, it doesn't look like any of those
> > >     tests do something that really should take this long. I'm especially
> > >     skeptical about the fact that we test each operation 8192 times by
> > >     default.
> > >
> > >   - It makes using kunit more tedious than it should be. Like I said, on
> > >     a very capable machine, running the all the DRM and VC4 tests takes
> > >     about 50s with those two tests, ~0.4s without.
> > >
> > >   - The corollary is that it will get in the way of people that really
> > >     want to use kunit will just remove those tests before doing so,
> > >     defeating the original intent.
> > >
> > >
> > > I understand that it came from selftests initially, but I think we
> > > should rewrite the tests entirely to have smaller, faster tests. It's
> > > not clear to me why those tests are as complicated as they are though.
> > >
> > > Also, going forward we should probably put disencourage tests running
> > > that long. Could Kunit timeout/warn after a while if a test is taking
> > > more than X seconds to run?
> >
> > I'd still like to address this. We spend ~90% of the DRM kunit tests
> > execution time executing those two tests, which doesn't seem like a
> > reasonable thing to do.
> 
> FWIW, KUnit is going to add a "speed" attribute for tests, so that
> it's easy to skip tests which are slow:
> https://lore.kernel.org/linux-kselftest/20230724162834.1354164-3-rmoar@xxxxxxxxxx/T/#u
> 
> This would allow the slow tests to be marked using KUNIT_CASE_SLOW(),
> and then be run via kunit.py --filter "speed>slow".
> 
> It obviously doesn't make the tests themselves any faster, but could
> at least make it possible to run only the fast tests during
> development, and the full, slower set before sending the patches out
> (or in CI), for example.

That's awesome, thanks

Speaking of which, should we detect in kunit.py tests that should be
marked as (super) slow but aren't?

Maxime

Attachment: signature.asc
Description: PGP signature