We need to find out why the kernel starts returning -EINVAL from the page flip
and wait for vblank ioctls.

Can you try writing 255 to /sys/module/drm/parameters/debug before reproducing
the problem, and attach the resulting dmesg output? (Note that writing non-0 to
/sys/module/drm/parameters/debug will cause a lot of debugging output to be
generated, so you'll want to write 0 to it again soon afterwards)

In case that doesn't reveal why -EINVAL is being returned, are you able to
recompile the drm.ko / amdgpu.ko kernel modules with a patch applied?