Re: [RFC v1 0/6] kunit: Add test attributes API

[Rae Moar]

On Sat, Jun 10, 2023 at 4:29 AM David Gow <davidgow@xxxxxxxxxx> wrote:
>
> On Sat, 10 Jun 2023 at 08:52, Rae Moar <rmoar@xxxxxxxxxx> wrote:
> >
> > Hello everyone,
> >
> > This is an RFC patch series to propose the addition of a test attributes
> > framework to KUnit.
> >
> > There has been interest in filtering out "slow" KUnit tests. Most notably,
> > a new config, CONFIG_MEMCPY_SLOW_KUNIT_TEST, has been added to exclude
> > particularly slow memcpy tests
> > (https://lore.kernel.org/all/20230118200653.give.574-kees@xxxxxxxxxx/).
>
> Awesome: this is a long overdue feature.
>

Hi David!

Thank you for all the comments!

> > This proposed attributes framework would be used to save and access test
> > associated data, including whether a test is slow. These attributes would
> > be reportable (via KTAP and command line output) and some will be
> > filterable.
>
> Why wouldn't they all be filterable? I guess I can imagine some where
> filtering wouldn't be useful, but I can't think of a technical reason
> why the filter shouldn't work.

I am definitely open to all attributes being filterable. My
reservation is that I can imagine an attribute with a complex data
type that would cause the filtering method to be difficult to
implement. If the attribute does not benefit much from being
filterable, I wonder if it is worth requiring the filtering method to
be implemented in that case.

Perhaps for now all attributes are filterable and then if this becomes
the case, this is addressed then?

>
> Also, as I understand it, I think this could also work with data which
> is not "saved" in this kunit_attributes struct you define. So we could
> have attributes which are generated automatically from other
> information about the test.
>  I could definitely see value in being able to filter on things like
> "is_parameterised" or "runs_at_init" or similar.
>

Yes! This is a great benefit of this flexible structure for
attributes. I would definitely be interested in implementing
"is_parameterised" and "runs_at_init" in future patches.

> Finally, it'd be really great if these attributes could apply to
> individual parameters in parameterised tests, in which case we could
> have and filter on the parameter value. That seems like it could be
> incredibly useful.
>

Yes, this would be an exciting extension for this project. I have
started thinking about this as potentially a follow up project.

> >
> > This framework is designed to allow for the addition of other attributes in
> > the future. These attributes could include whether the test is flaky,
> > associated test files, etc.
>
> A small part of me is hesitant to add this much framework code for
> only one attribute, so it'd be nice to look into at least having an
> RFC for some of these. Hopefully we don't actually have flaky tests,
> but "is_deterministic" would be nice (alongside a future ability to
> inject a random seed, or similar). Other properties like
> "is_threadsafe", "is_reentrant", etc could be useful for future
> features. And I'm sure there could be some more subsystem-specific
> things which would be useful to filter on, too.
>

I understand the reservations to add this large framework for one
attribute. I would definitely consider adding an additional attribute
to this RFC or creating a separate RFC.

I would be happy to go ahead and add "is_deterministic" if there is
interest. As well as potentially "is_threadsafe", "is_reentrant", etc
in future patches.

> Some of these could probably replace the need for custom code to make
> the test skip itself if dependencies aren't met, too, which would be
> fun.

This would be great!

>
> I'm not sure "associated test files" quite gels perfectly with this
> system as-is (assuming I understand what that refers to). If it's the
> ability to "attach" extra data (logs, etc) to the KTAP output, that's
> possibly best injected at runtime, or added by the separate parser,
> rather than hardcoded in the kernel.
>

Hmm I see what you are saying here. If the associated test files of
interest are best injected at runtime I am happy to scrap this idea
for now.

> >
> > Note that this could intersect with the discussions on how to format
> > test-associated data in KTAP v2 that I am also involved in
> > (https://lore.kernel.org/all/20230420205734.1288498-1-rmoar@xxxxxxxxxx/).
> >
> I definitely need to re-read and respond to that. I'm not 100%
> thrilled with the output format here, and I think the goal with KTAP
> "test associated data" is, as you say, related, but not identical to
> this. There'd definitely be data which doesn't make sense as a KUnit
> attribute which we might want to add to the output (e.g., data only
> calcuated while the test runs), and attributes which we might not want
> to always print out with the results.
>

I have thought much about the differences between the two concepts. My
current understanding with KTAP metadata and KUnit attributes is that
they are not going to be perfect mirrors of each other but the KUnit
attributes framework can help to save and output some of the KTAP
metadata.

I am happy to be flexible on the output format or see more discussion
on KTAP metadata in general. What part of the output is most
concerning?

> > If the overall idea seems good, I'll make sure to add tests/documentation,
> > and more patches marking existing tests as slow to the patch series.
> >
>
> I think the series is good overall. If no-one else objects, let's move
> forward with it.
> I'd definitely prefer to see a few more tests and some documentation.
> Having another attribute would be great, too, though I can certainly
> live with that being a separate series.

Great! Yes I will add documentation and more tests in the next
versions. I will also work on the implementation of another attribute.


>
>
> > Thanks!
> > Rae
> >
> > Rae Moar (6):
> >   kunit: Add test attributes API structure
> >   kunit: Add speed attribute
> >   kunit: Add ability to filter attributes
> >   kunit: tool: Add command line interface to filter and report
> >     attributes
> >   kunit: memcpy: Mark tests as slow using test attributes
> >   kunit: time: Mark test as slow using test attributes
> >
> >  include/kunit/attributes.h             |  41 ++++
> >  include/kunit/test.h                   |  62 ++++++
> >  kernel/time/time_test.c                |   2 +-
> >  lib/kunit/Makefile                     |   3 +-
> >  lib/kunit/attributes.c                 | 280 +++++++++++++++++++++++++
> >  lib/kunit/executor.c                   |  89 ++++++--
> >  lib/kunit/executor_test.c              |   8 +-
> >  lib/kunit/kunit-example-test.c         |   9 +
> >  lib/kunit/test.c                       |  17 +-
> >  lib/memcpy_kunit.c                     |   8 +-
> >  tools/testing/kunit/kunit.py           |  34 ++-
> >  tools/testing/kunit/kunit_kernel.py    |   6 +-
> >  tools/testing/kunit/kunit_tool_test.py |  41 ++--
> >  13 files changed, 536 insertions(+), 64 deletions(-)
> >  create mode 100644 include/kunit/attributes.h
> >  create mode 100644 lib/kunit/attributes.c
> >
> >
> > base-commit: fefdb43943c1a0d87e1b43ae4d03e5f9a1d058f4
> > --
> > 2.41.0.162.gfafddb0af9-goog
> >