On 09/11/2026 10:21, Philipp Stanner wrote:
On Fri, 2026-11-06 at 14:56 +0000, Tvrtko Ursulin wrote:
[First four patches are just my queue to avoid conflicts. I am still hoping I
can bring them in a state where all is good since all these together simplify
things.]

I'm confused, so is this an RFC for AMD to take a first look?

Yes, but as the cover letter later re-iterates, skip to patches 5 and 6 for spsc removal. Prologue is due my local patch ordering and otherwise there would be conflicts.

So as requested by Philipp and as the title says - replacing the SPSC queue with
a list and a spinlock. A first cut which passes unit tests and smoke testing on
AMD.

Thx for working on this, it's one of the most important chunks of tech
debt!

But it wasn't easy, since we can't really get away from some of the deeply
ingrained scheduler design points.

What would those be?

See patch 6. Unlocked peek from the run job singleton due entity<->rq lock inversion and sleeping ->prepare_job. In this RFC peek is implemented via list RCU helpers which make it safe because the single consumer is still true.

Regards,

Tvrtko

Reply via email to