[LSF/MM/BPF TOPIC] Implementing KVM vCPU Priority boosting via BPF

[LSF/MM/BPF TOPIC] Implementing KVM vCPU Priority boosting via BPF

[Joel Fernandes]

We should discuss a new approach for increasing KVM virtual CPU (vCPU) priority
when guests need low latency. The last RFC posting [1] on this is thought to be
too rigid and baking too much policy into the kernel. Incorporating complex
policy logic directly into KVM seems problematic long-term for maintenance. Lets
discuss leveraging BPF programs to offload more scheduling policy decisions to
BPF / userspace.

Specific issues to discuss:

* Add support for enabling BPF programs to share memory and interface with guest.

* Create a kernel function allowing BPF programs to call sched_setscheduler(),
facilitating priority boosting.

* UAPI concerns.

* Challenges with loading BPF programs in guest userspace we don't control.

[1] https://lore.kernel.org/all/20231214024727.3503870-1-vineeth@bitbyteword.org/

Re: [LSF/MM/BPF TOPIC] Implementing KVM vCPU Priority boosting via BPF

[David Vernet]

[-- Attachment #1: Type: text/plain, Size: 1019 bytes --]

On Tue, Jan 23, 2024 at 09:59:50PM -0500, Joel Fernandes wrote:
> We should discuss a new approach for increasing KVM virtual CPU (vCPU) priority
> when guests need low latency. The last RFC posting [1] on this is thought to be
> too rigid and baking too much policy into the kernel. Incorporating complex
> policy logic directly into KVM seems problematic long-term for maintenance. Lets
> discuss leveraging BPF programs to offload more scheduling policy decisions to
> BPF / userspace.
> 
> Specific issues to discuss:
> 
> * Add support for enabling BPF programs to share memory and interface with guest.
> 
> * Create a kernel function allowing BPF programs to call sched_setscheduler(),
> facilitating priority boosting.
> 
> * UAPI concerns.
> 
> * Challenges with loading BPF programs in guest userspace we don't control.
> 
> [1] https://lore.kernel.org/all/20231214024727.3503870-1-vineeth@bitbyteword.org/

+1 to discussing all of the above at LSFMM, ideally as part of the BPF
track.

[-- Attachment #2: signature.asc --]
[-- Type: application/pgp-signature, Size: 228 bytes --]