[Intel-gfx] [igt-dev] [PATH i-g-t] igt: Test tagging support

[Tvrtko Ursulin]

On 12/09/2018 10:07, Chris Wilson wrote:
> Quoting Tvrtko Ursulin (2018-09-12 09:48:00)
>>
>> On 07/09/2018 12:43, Chris Wilson wrote:
>>> Quoting Tvrtko Ursulin (2018-09-07 12:14:20)
>>>> From: Tvrtko Ursulin <tvrtko.ursulin at intel.com>
>>>>
>>>> Proposal to add test tags as a replacement for separate test
>>>> list which can be difficult to maintain and get out of date.
>>>>
>>>> Putting this maintanenace inline with tests makes it easier
>>>> to remember to update the (now implicit) lists, and also enables
>>>> richer test selection possibilities for the test runner.
>>>>
>>>> Current method of implying tags from test/subtest names has a
>>>> couple of problems one of which is where some names can use
>>>> the same token for different meanings. (One example is the
>>>> "default" token.) It also creates a name clash between naming
>>>> and tagging.
>>>>
>>>> Test tags are strings of tokens separated by spaces, meaning of
>>>> which we decide by creating a convetion and helped by the
>>>> suitable helper macros.
>>>>
>>>> For example tags can be things like: gem, kms, basic, guc, flip,
>>>> semaphore, bz12345678, gt4e, reset, etc..
>>>>
>>>> At runtime this would look something like this (abbreviated for
>>>> readability):
>>>>
>>>>     @ tests/gem_sync --list-subtests-with-tags
>>>>     ...
>>>>     forked-each TAGS="gem "
>>>>     forked-store-each TAGS="gem "
>>>>     basic-all TAGS="gem basic "
>>>>     basic-store-all TAGS="gem basic "
>>>>
>>>>     @ tests/gem_concurrent_blit --list-subtests-with-tags
>>>>     ...
>>>>     16MiB-swap-gpuX-render-write-read-bcs-bomb TAGS="gem stress "
>>>>     16MiB-swap-gpuX-render-write-read-rcs-bomb TAGS="gem stress "
>>>>     16MiB-swap-gpuX-render-gpu-read-after-write-bomb TAGS="gem stress "
>>>>
>>>>     @ tests/kms_flip --list-subtests-with-tags | grep basic
>>>>     basic-plain-flip TAGS="kms basic "
>>>>     basic-flip-vs-dpms TAGS="kms basic "
>>>>
>>>> Test runner could then enable usages like:
>>>>
>>>>     ./run-tests --include gem --exclude stress
>>>
>>> Minor comment, I like some basic boolean algebra here
>>> --include "gem AND smoketest NOT stress"
>>> :)
>>
>> That's what my hypothetical examples showed just with a different syntax!
>>
>>> I'd probably tag everything according to which ioctls they use. I feel my
>>> endgoal is still to list tests by their kernel functions (which can and
>>> should be automatically derived), and their hw function which is the
>>> open problem.
>>
>> If we want to do that automatically then tagging in this flavour
>> probably doesn't make sense and it should instead be an external database.
>>
>> If we go on the ioctl granularity it can probably mostly have one
>> version, and it should be static enough to be able to live in the tree,
>> but if we go more granular, like something derived from kcov, then
>> that's out of the window. Since it will be tied both to a platform and
>> kernel version.
>>
>> So I think tagging as proposed here is not appropriate for either, but
>> only if we wanted to tag on coarser functional areas and such.
> 
> Yeah, I think the same as well, that tags should be "smoke", "stress".
> (But one man's stress is another's minimal workload :(
> 
> But both of those are in essence a duration parameter, and I'd prefer if
> we expressed those as configurable parameters. At which point there is a
> meta level of test scripts to tag ;)
> 
> I feel that "gem", "kms" is better expressed in lower level terms, so
> what's left to tag? Clients? "display", "opencl", "media", "opengl" ?
> Even using client specs for features (e.g. EGL_IMG_context_priority)?

If we overcomplicate and try to change too much at once it is deemed to 
fail. GEM, KMS, etc, have this advantage of being very established.

So I was thinking these basic keywords and then something high level 
like reset, rps, stolen, and similar. So for instance we can remove the 
igt_skip_on_simulation, which can be time consuming to answer why it is 
there, and replace it with --exclude reset,... when running under the 
simulator.

True it is manual work to keep the tags up to date. Is it more work than 
the current scheme is TBD.

> Who and why would use those? From a selfish perspective, I want lcov
> with tests sorted in order of increasing "stress" (start with the
> precise X does Y test, finish with can it survive pathological client
> behaviour for 48 hours).

Who and why would use which ones? Tags in general, tags in the style the 
patch proposes, or tags in the flavour you described?

My initial idea was that it would make easy for developers to run an 
approximate sweep when touching a feature. But with good CI and trybot 
maybe no one would use tags and just rely on "outsourced" test runs? 
Could be.. but hey, you convinced me to re-send this. Or was it Joonas? :)

Regards,

Tvrtko