[RFC PATCH v3 0/3] new subsystem for compute accelerator devices

[Christopher Friedt]

Hi Oded,

On Sun, Nov 6, 2022 at 4:03 PM Oded Gabbay <ogabbay at kernel.org> wrote:
> The patches are in the following repo:
> https://git.kernel.org/pub/scm/linux/kernel/git/ogabbay/accel.git/log/?h=accel_v3
>
> As in v2, The HEAD of that branch is a commit adding a dummy driver that
> registers an accel device using the new framework. This can be served
> as a simple reference. I have checked inserting and removing the dummy driver,
> and opening and closing /dev/accel/accel0 and nothing got broken :)
>
> v1 cover letter:
> https://lkml.org/lkml/2022/10/22/544
>
> v2 cover letter:
> https://lore.kernel.org/lkml/20221102203405.1797491-1-ogabbay@kernel.org/T/

I was in the room at Plumbers when a lot of this was discussed (in
2022 and also 2019), but I haven't really had an opportunity to
provide feedback until now. In general, I think it's great and thanks
for pushing it forward and getting feedback.

The v1 cover letter mentioned RAS (reliability, availability,
serviceability) and Dave also mentioned it here [1]. There was a
suggestion to use Netlink. It's an area that I'm fairly interested in
because I do a lot of development on the firmware side (and
specifically, with Zephyr).

Personally, I think Netlink could be one option for serializing and
deserializing RAS information but it would be helpful for that
interface to be somewhat flexible, like a void * and length, and to
provide userspace the capability of querying which RAS formats are
supported.

For example, AntMicro used OpenAMP + rpmsg in their NVMe accelerator,
and gave a talk on it at ZDS and Plumbers this year [2][3].

In Zephyr, the LGPL license for Netlink might be a non-starter
(although I'm no lawyer). However, Zephyr does already support
OpenAMP, protobufs, json, and will soon support Thrift.

Some companies might prefer to use Netlink. Others might prefer to use
ASN.1. Some companies might prefer to use key-value pairs and limit
the parameters and messages to uint32s. Some might handle all of the
RAS details in-kernel, while others might want the kernel to act more
like a transport to firmware.

Companies already producing accelerators may have a particular
preference for serialization / deserialization in their own
datacenters.

With that, it would be helpful to be able to query RAS capabilities via ioctl.

#define ACCEL_CAP_RAS_KEY_VAL_32 BIT(0)
#define ACCEL_CAP_RAS_NETLINK BIT(1)
#define ACCEL_CAP_RAS_JSON BIT(2)
#define ACCEL_CAP_RAS_PROTOBUF BIT(3)
#define ACCEL_CAP_RAS_GRPC BIT(4)
#define ACCEL_CAP_RAS_THRIFT BIT(5)
#define ACCEL_CAP_RAS_JSON BIT(6)
#define ACCEL_CAP_RAS_ASN1 BIT(7)

or something along those lines. Anyway, just putting the idea out there.

I'm sure there are a lot of opinions on this topic and that there are
a lot of implications of using this or that serialization format.
Obviously there can be security implications as well.

Apologies if I've already missed some of this discussion.

Cheers,

C

[1] https://airlied.blogspot.com/2022/09/accelerators-bof-outcomes-summary.html
[2] https://zephyr2022.sched.com/event/10CFD/open-source-nvme-ai-accelerator-platform-with-zephyr-karol-gugala-antmicro
[3] https://lpc.events/event/16/contributions/1245/