Re: [RFC v2 03/14] vfio/nvidia-vgpu: introduce vGPU type uploading

[Zhi Wang <zhiw@nvidia.com>]

On Thu, 04 Sep 2025 11:41:03 +0200
"Danilo Krummrich" <dakr@kernel.org> wrote:

> (Cc: Alex, John, Joel, Alistair, nouveau)
> 
> On Thu Sep 4, 2025 at 11:37 AM CEST, Danilo Krummrich wrote:
> > On Thu Sep 4, 2025 at 12:11 AM CEST, Zhi Wang wrote:
> >> diff --git a/drivers/vfio/pci/nvidia-vgpu/include/nvrm/gsp.h b/drivers/vfio/pci/nvidia-vgpu/include/nvrm/gsp.h
> >> new file mode 100644
> >> index 000000000000..c3fb7b299533
> >> --- /dev/null
> >> +++ b/drivers/vfio/pci/nvidia-vgpu/include/nvrm/gsp.h
> >> @@ -0,0 +1,18 @@
> >> +/* SPDX-License-Identifier: MIT */
> >> +#ifndef __NVRM_GSP_H__
> >> +#define __NVRM_GSP_H__
> >> +
> >> +#include <nvrm/nvtypes.h>
> >> +
> >> +/* Excerpt of RM headers from https://github.com/NVIDIA/open-gpu-kernel-modules/tree/570 */
> >> +
> >> +#define NV2080_CTRL_CMD_GSP_GET_FEATURES (0x20803601)
> >> +
> >> +typedef struct NV2080_CTRL_GSP_GET_FEATURES_PARAMS {
> >> +	NvU32  gspFeatures;
> >> +	NvBool bValid;
> >> +	NvBool bDefaultGspRmGpu;
> >> +	NvU8   firmwareVersion[GSP_MAX_BUILD_VERSION_LENGTH];
> >> +} NV2080_CTRL_GSP_GET_FEATURES_PARAMS;
> >> +
> >> +#endif
> >
> > <snip>
> >

The RFC v2 is based on the same architecture of RFC V1 but switching the
core driver from NVKM to nova-core. Yet the new architecture and auxiliary
bus is WIP. So it doesn't represent the final picture, e.g. the rust code
I wrote in the nova-core. The main idea is to demonstrate the progress
of the vGPU development.

> >> +static struct version supported_version_list[] = {
> >> +	{ 18, 1, "570.144" },
> >> +};
> >
> > nova-core won't provide any firmware specific APIs, it is meant to serve as a
> > hardware and firmware abstraction layer for higher level drivers, such as vGPU
> > or nova-drm.
> >
> > As a general rule the interface between nova-core and higher level drivers must
> > not leak any hardware or firmware specific details, but work on a higher level
> > abstraction layer.
> >

It is more a matter of where we are going to place vGPU specific
functionality in the whole picture. In this case, if we are thinking about
the requirement of vGPU type loading, which requires the GSP version
number and checking. Are we leaning towards putting some vGPU specific
functionality also in nova-core?

Regarding not leaking any of the hardware details, is that doable? 
Looking at {nv04 * _fence}.c {chan*}.c in the current NVIF interfaces, I
think we will expose the HW concept somehow.

> > Now, I recognize that at some point it might be necessary to do some kind of
> > versioning in this API anyways. For instance, when the semantics of the firmware
> > API changes too significantly.
> >
> > However, this would be a separte API where nova-core, at the initial handshake,
> > then asks clients to use e.g. v2 of the nova-core API, still hiding any firmware
> > and hardware details from the client.
> >
> > Some more general notes, since I also had a look at the nova-core <-> vGPU
> > interface patches in your tree (even though I'm aware that they're not part of
> > the RFC of course):
> >
> > The interface for the general lifecycle management for any clients attaching to
> > nova-core (VGPU, nova-drm) should be common and not specific to vGPU. (The same
> > goes for interfaces that will be used by vGPU and nova-drm.)
> >
> > The interface nova-core provides for that should be designed in Rust, so we can
> > take advantage of all the features the type system provides us with connecting
> > to Rust clients (nova-drm).
> >
> > For vGPU, we can then monomorphize those types into the corresponding C
> > structures and provide the corresponding functions very easily.
> >
> > Doing it the other way around would be a very bad idea, since the Rust type
> > system is much more powerful and hence it'd be very hard to avoid introducing
> > limitations on the Rust side of things.
> >
> > Hence, I recommend to start with some patches defining the API in nova-core for
> > the general lifecycle (in Rust), so we can take it from there.
> >
> > Another note: I don't see any use of the auxiliary bus in vGPU, any clients
> > should attach via the auxiliary bus API, it provides proper matching where
> > there's more than on compatible GPU in the system. nova-core already registers
> > an auxiliary device for each bound PCI device.
> >
> > Please don't re-implement what the auxiliary bus already does for us.
> >
> > - Danilo
> 
>