On 18/6/24 03:30, Danilo Krummrich wrote:
On Fri, Jun 14, 2024 at 03:02:09AM +1000, Ben Skeggs wrote:
NVIDIA has been exploring ways to better support the effort for an
upstream kernel mode driver for GPUs that are capable of running GSP-RM
firmware, since the introduction[1] to Nova.
Use cases have been identified for which separating the core GPU
programming out of the full DRM driver stack is a strong requirement
from our key customers.
An upstreamed NVIDIA GPU driver should be able to support current and
emerging customer use cases for vGPU hosts. NVIDIA's vGPU deployments
to date do not support compute or graphics functionality within the
hypervisor host, and have no dependency on the Linux graphics subsystem,
instead implementing the minimal functionality required to run vGPU
guest VMs.
For security-sensitive environments such as cloud infrastructure, it's
important to continue support for running a minimal footprint vGPU host
driver in a stripped-down / barebones kernel environment.
This can be achieved by supporting both VFIO and DRM drivers as clients
of a core driver, without requiring a full-fledged DRM driver (or the
DRM subsystem itself) to be built into the host kernel.
A core driver would be responsible for booting and communicating with
GSP-RM, enumeration of HW configuration, shared/partitioned resource
management, exception handling, and event dispatch.
The DRM driver would do all the standard things a DRM driver does, and
implement GPU memory management (TTM/HMM), KMS, command submission etc,
as well as providing UAPI for userspace clients. These features would
be implemented using HW resources allocated from a core driver, rather
than the DRM driver being directly responsible for HW programming.
As Nouveau's KMD is already split (in the logical sense) along similar
lines, we're using it here for the purposes of this RFC to demonstrate
the feasibility of such an architecture, and open it up for discussion.
Generally, I think that approach is reasonable and I like it. There's only a few
concerns I have for now.
We've already had (and still have) quite a few difficulties due to this split in
Nouveau. Especially when it comes to VMM and handling page tables. There are
cases where the locking architecture must be closely aligned with the upper
layers, i.e. the (VM_BIND) uAPI.
Having a separate (local) locking architecture doesn't work out well in this
case due to the implications of dealing with dma_fences and their signalling
paths.
Unfortunately, we can't even argue that we solved this problem in Nouveau. I
think it's fair to say that we found ways (without rewriting / restructuring a
lot of the VMM code to use a more global locking architecture) to make it work
in practice, but surely there are still conditions that (at least theoretically)
can lock things up.
I'm not saying that it's impossible to work this out, but having a strong
separation is likely to make those things quite a bit more difficult.
Yeah, I think there's a bit of work ahead to determine where exactly all
the pieces should live. For VMM specifically, I'd be looking more at an
architecture where the DRM driver "owns" all the memory remaining after
GSP has booted (or was allocated to its VFIO partition) etc, and manages
it however it sees fit. And, rather than calling into the core driver
for mapping into a VMM, the DRM driver would allocate its own PDB,
inform the core driver of its location, and manage its own page tables
directly from there.
This is similar to how the interface between NVKM and GSP-RM works now
at least.
I've looked a little bit recently into how to approach fixing this in
nouveau, but don't have a solid plan yet.
On the other hand this is a problem we might have to deal with either way, it
shouldn't matter too much having separate modules for VFIO and the GPU core.
Besides that, do we expect semantical changes in the firmware that can
potentially propagate up in the following sense?
[GSP firmware -> Host GPU core driver -> VFIO driver -> Guest GPU core driver]
If so, how do we deal with those? In the context of ensuring compatibility, can
we ensure this can't lead to increasing maintainance and testing effort over
time?
That's a very good question. I suspect it's inevitable that those type
of changes could flow through from FW updates, and we'll need to come up
with a plan for how to deal with them. In general, it seems like the
same kind of problem as with maintaining UAPI compatibility, but I'm not
sure if there's any additional considerations for virt.
Ben.
- Danilo
A link[2] to a tree containing the patches is below.
[1] https://lore.kernel.org/all/3ed356488c9b0ca93845501425d427309f4cf616.camel@xxxxxxxxxx/
[2] https://gitlab.freedesktop.org/bskeggs/nouveau/-/tree/00.03-module
*** BLURB HERE ***
Ben Skeggs (2):
drm/nouveau/nvkm: export symbols needed by the drm driver
drm/nouveau/nvkm: separate out into nvkm.ko
drivers/gpu/drm/nouveau/Kbuild | 4 ++--
drivers/gpu/drm/nouveau/include/nvkm/core/module.h | 3 ---
drivers/gpu/drm/nouveau/nouveau_drm.c | 10 +---------
drivers/gpu/drm/nouveau/nvkm/core/driver.c | 1 +
drivers/gpu/drm/nouveau/nvkm/core/gpuobj.c | 2 ++
drivers/gpu/drm/nouveau/nvkm/core/mm.c | 4 ++++
drivers/gpu/drm/nouveau/nvkm/device/acpi.c | 1 +
drivers/gpu/drm/nouveau/nvkm/engine/gr/base.c | 1 +
drivers/gpu/drm/nouveau/nvkm/module.c | 8 ++++++--
drivers/gpu/drm/nouveau/nvkm/subdev/bios/init.c | 1 +
drivers/gpu/drm/nouveau/nvkm/subdev/bios/pll.c | 1 +
drivers/gpu/drm/nouveau/nvkm/subdev/fb/base.c | 3 +++
drivers/gpu/drm/nouveau/nvkm/subdev/gpio/base.c | 3 +++
drivers/gpu/drm/nouveau/nvkm/subdev/i2c/base.c | 2 ++
drivers/gpu/drm/nouveau/nvkm/subdev/i2c/bus.c | 1 +
drivers/gpu/drm/nouveau/nvkm/subdev/iccsense/base.c | 1 +
drivers/gpu/drm/nouveau/nvkm/subdev/therm/base.c | 1 +
drivers/gpu/drm/nouveau/nvkm/subdev/therm/fan.c | 1 +
drivers/gpu/drm/nouveau/nvkm/subdev/volt/base.c | 1 +
19 files changed, 33 insertions(+), 16 deletions(-)
--
2.44.0
