[PATCH 0/3] NVKM GSP RPC fixes

[Zhi Wang]

Hi folks:

Here are some fixes of weird bugs I encountered when I was enabling vGPU [1] on
core-driver aka NVKM. They are exposed because of the new RPCs required by
vGPU.

For testing, I tried to run Uniengine Heaven[2] on my RTX 4060 for 8 hours and
the GL CTS runner[3] (commandline: ./cts-runner --type-gl40) from Khronos
without any problem.

v2:

- Remove the Fixes: tags as the vanilla nouveau aren't going to hit these bugs. (Danilo)
- Test the patchset on VK-GL-CTS. (Danilo)
- Remove the ambiguous empty line in PATCH 2. (Danilo)
- Rename the r535_gsp_msgq_wait to gsp_msgq_recv. (Danilo)
- Introduce a data structure to hold the params of gsp_msgq_recv(). (Danilo)
- Document the params and the states they are related to. (Danilo)

[1] https://lore.kernel.org/kvm/20240922124951.1946072-1-zhiw@nvidia.com/T/#t
[2] https://benchmark.unigine.com/heaven
[3] https://github.com/KhronosGroup/VK-GL-CTS

Zhi Wang (3):
  nvkm/gsp: correctly advance the read pointer of GSP message queue
  nvkm: correctly calculate the available space of the GSP cmdq buffer
  nvkm: handle the return of large RPC

 .../gpu/drm/nouveau/nvkm/subdev/gsp/r535.c    | 251 +++++++++++++-----
 1 file changed, 184 insertions(+), 67 deletions(-)

-- 
2.34.1