Re: [PATCH 13/14] sched: Add {DE,EN}QUEUE_LOCKED

[Peter Zijlstra <peterz@infradead.org>]

On Sat, Sep 13, 2025 at 12:32:27PM -1000, Tejun Heo wrote:
> Hello,
> 
> On Fri, Sep 12, 2025 at 06:32:32AM -1000, Tejun Heo wrote:
> > Yeah, or I can make scx_tasks iteration smarter so that it can skip through
> > the list for tasks which aren't runnable. As long as it doesn't do lock ops
> > on every task, it should be fine. I think this is solvable one way or
> > another. Let's continue in the other subthread.
> 
> Thought more about it. There's another use case for this runnable list,
> which is the watchdog. As in the migration synchronization, I think the
> right thing to do here is just adding a nested lock. That doesn't add any
> overhead or complications to other sched classes and from sched_ext POV
> given how expensive migrations can be, if we make that a bit cheaper (and I
> believe we will with changes being discussed), added up, the outcome would
> likely be lower overhead.

I really don't see how you could possibly retain that runnable_list.

pick_next_task() must be able to migrate a task from a shared runqueue
to a local runqueue. It must do this without taking a random other
per-cpu runqueue. Therefore, a task on a DSQ must have no 'local' state.

This very much means the runnable_list cannot be per cpu.

No per-task lock is going to help with that.

The watchdog will have to go iterate DSQs or something like that.