Re: [RFC PATCH v2 01/15] misc/arm-cla: Add driver skeleton and documentation

From: Linus Walleij

Date: Fri Nov 20 2026 - 04:48:01 EST


On Thu, Oct 15, 2026 at 4:26 PM Ryan Roberts <ryan.roberts@xxxxxxx> wrote:

> +Inter-Accelerator Communication
> +-------------------------------
> +
> +On some platforms, multiple accelerators, each attached to a separate CLA within
> +a cluster, are also directly connected to each other via a shared bus to
> +accelerate cooperation between accelerators. The accelerators sharing a bus
> +cannot be isolated from each other. When collaborative operations are launched
> +on each of the participating accelerators, they synchronize over the bus,
> +stalling until all are ready.

What the patch series needs to do is tell us how this topology is discovered.
I.e. will it be discoverable from the hardware itself, or will this topology be
described in device tree or ACPI for example?

The kernel must certainly know this topology to be able to synchronize
work across several cores? What happens if an asynchronous acceleration
task is running on some cores and no actual CPU load is on any of them
and the kernel want to put them to idle or shut them down? Same problem
on the cluster level?

Further since this appears to be hardware topology userspace must obtain
this information from the kernel, not the other way around: we cannot
have a situation where userspace "just knows" stuff about the hardware,
keeping the kernel in the dark.

I think this is a wider question about the discovery of these CLA:s:
there must be a coherent story about how userspace finds it and maps
it topology and that information must come from the kernel, so the
kernel and userspace have the same map of the world.

Right now IIUC the driver probes from discovery in the hardware from
CPU-specific registers so does this discovery also include the topology?

Yours,
Linus Walleij