Re: cron job: media_tree daily build: ERRORS

[Hans Verkuil <hverkuil-cisco@xs4all.nl>]

Hi all,

On 10/12/22 05:13, Hans Verkuil wrote:
> This message is generated daily by a cron job that builds media_tree for
> the kernels and architectures in the list below.

<snip>

> virtme: ERRORS: Final Summary: 3080, Succeeded: 3069, Failed: 11, Warnings: 0
> virtme-32: ERRORS: Final Summary: 3193, Succeeded: 3170, Failed: 23, Warnings: 0

For quite a long time now the regression tests kept failing in mysterious ways,
and I hope I have finally found the cause.

The test-media script that runs the regression has to load vivid and initialize
it in a few places:

	modprobe vivid n_devs=3 multiplanar=1,2,2 cache_hints=1,0,0
	sleep 3
	v4l2_ctl -z platform:vivid-002 -d vivid-002-vid-cap -i3 -v width=3840,height=2160,pixelformat=NV24
	v4l2_ctl -z platform:vivid-002 -d vivid-002-vid-out -o1 -x width=3840,height=2160,pixelformat=NM16

The two v4l2-ctl commands configure the third vivid instance so that it can be
used to export dmabufs that are large enough for use as import buffers elsewhere
in the scripts.

v4l2-ctl finds the media device and reads the topology to find the major/minor
numbers of the video device to use. Then it opens /dev/char/major:minor.

However, that can fail if CONFIG_DEBUG_KOBJECT_RELEASE=y. That makes everything
a lot slower, so after the 'sleep 3' above the /dev/char/major:minor symlink
hasn't been created yet by udev.

So even though the module was loaded, the udev rules weren't finished yet and
the v4l2-ctl commands failed. Therefor the tests that tried to export/import
dmabufs would fail because the buffers were too small. All this is of course
dependent on the speed of the CPU, phase of the moon, etc.

Increasing the sleep should fix this issue. Fingers crossed.

Regards,

	Hans