[PATCH i-g-t v2] lib/xe/xe_query: Do not require QUERY_OA_UNITS for all Xe tests

[Dixit, Ashutosh]

On Thu, 25 Jul 2024 07:03:19 -0700, Peter Senna Tschudin wrote:
>

Hi Peter,

> Changes xe_query_oa_units_new() to return NULL instead of aborting
> due to assertion failure when querying for DRM_XE_DEVICE_QUERY_OA_UNITS.
>
> It adds a skip to xe_oa.c to make sure it will not try to run the tests if the
> query has failed before.
>
> The motivation is as follows:
>  - The only test that requires oa_units is tests/intel/xe_oa.c, however current
>    code will prevent all Xe tests from running if OA_UNITS is not supported
>  - Before this patch, trying to run IGT on the Xe module from Fedora 40
>    (6.9.9-200.fc40.x86_64) fails
>  - After this patch, we can run some tests, such as xe_exec_basic
>    successfully

As far as I understand the purpose of IGT is to test the latest drm-tip
kernels. Also, I would think testing older kernels is of dubious value,
since bugs you would find there would typically have been fixed in later
kernels.

I also consider this patch sort of temporary, I would imagine it can be
reverted after distro kernels have moved to 6.11 where the uapi in this
patch will be available.

With that out of the way, I will approve this patch for merging but I have
a couple of nit-picks below.

>
> Here is an example of the impact of this patch:
>  1. Use Fedora 40 with the following kernel command line arguments:
>     "xe.force_probe=* i915.force_probe=!* modprobe.blacklist=i915,ast,snd_hda_intel,xe"
>
>  2. Running IGT without this patch will fail:
>      $ sudo ./build/tests/xe_exec_basic
>      $ echo $?
>      98
>
>  3. Running after fix applied will succeed:
>     sudo ./build/tests/xe_exec_basic
>      $ sudo ./build/tests/xe_exec_basic
>      $ echo $?
>      0
>
> V2:
>  - Cleanup headers from message body
>  - Add 'lib/xe/xe_query:' to subject line

Also, note that, assuming you are new here, for IGT and kernel, the commit
message should explain the "why" or the reason for the patch. So I would
think the above commit message can be condensed, but that is up to you,
it's ok with me.

>
> Signed-off-by: Peter Senna Tschudin <peter.senna at intel.com>
> ---
>  lib/xe/xe_query.c   | 6 +++++-
>  tests/intel/xe_oa.c | 9 ++++++++-
>  2 files changed, 13 insertions(+), 2 deletions(-)
>
> diff --git a/lib/xe/xe_query.c b/lib/xe/xe_query.c
> index 84eaaac96..f0643e1bc 100644
> --- a/lib/xe/xe_query.c
> +++ b/lib/xe/xe_query.c
> @@ -124,7 +124,11 @@ static struct drm_xe_query_oa_units *xe_query_oa_units_new(int fd)
>		.data = 0,
>	};
>
> -	igt_assert_eq(igt_ioctl(fd, DRM_IOCTL_XE_DEVICE_QUERY, &query), 0);
> +	/**

I am seeing, in IGT, the "/**" comment pattern used only for test or
subtest description. So we should just use this pattern for multi-line
comments:

/*
 * This is a multi-line comment
 */

> +	 * Only tests/intel/xe_oa.c needs OA_UNITS

This comment is misleading and needs to be changed. Please change it to
something like (no need for multi-line comment either):

	/* Support older kernels where this uapi is not yet available */

Since this is the real reason for this change.

> +	 */
> +	if (igt_ioctl(fd, DRM_IOCTL_XE_DEVICE_QUERY, &query) != 0)

No need for "!= 0", just do:

	if (igt_ioctl(fd, DRM_IOCTL_XE_DEVICE_QUERY, &query))

> +		return NULL;
>
>	oa_units = malloc(query.size);
>	igt_assert(oa_units);
> diff --git a/tests/intel/xe_oa.c b/tests/intel/xe_oa.c
> index ff2218300..dd02f8608 100644
> --- a/tests/intel/xe_oa.c
> +++ b/tests/intel/xe_oa.c
> @@ -4503,6 +4503,7 @@ static const char *xe_engine_class_name(uint32_t engine_class)
>  igt_main
>  {
>	struct drm_xe_engine_class_instance *hwe = NULL;
> +	struct xe_device *xe_dev;
>
>	igt_fixture {
>		struct stat sb;
> @@ -4527,7 +4528,13 @@ igt_main
>		igt_assert_eq(drm_fd, -1);
>
>		drm_fd = drm_open_driver(DRIVER_XE);
> -		xe_device_get(drm_fd);
> +		xe_dev = xe_device_get(drm_fd);
> +
> +		/**
> +		 * This test needs OA_UNITS to run, see xe_query_oa_units_new()
> +		 * on lib/xe/xe_query.c

Here just say:
		/* See xe_query_oa_units_new() */

Kernel style is to minimize comments and also minimize unnecessary
verbosity in comments.

> +		 */
> +		igt_require(xe_dev->oa_units);
>
>		devid = intel_get_drm_devid(drm_fd);
>		sysfs = igt_sysfs_open(drm_fd);
> --
> 2.34.1
>

Please fix these things up and resend. Then I will will R-b and merge the
patch.

Thanks.
--
Ashutosh