Re: [PATCH 3/3] sched_ext, scx_qmap: Add and use SCX_ENQ_CPU_SELECTED

[Andrea Righi]

On Fri, Sep 27, 2024 at 01:46:13PM -1000, Tejun Heo wrote:
> scx_qmap and other schedulers in the SCX repo are using SCX_ENQ_WAKEUP to
> tell whether ops.select_cpu() was called. This is incorrect as
> ops.select_cpu() can be skipped in the wakeup path and leads to e.g.
> incorrectly skipping direct dispatch for tasks that are bound to a single
> CPU.
> 
> sched core has been udpated to specify ENQUEUE_RQ_SELECTED if
> ->select_task_rq() was called. Map it to SCX_ENQ_CPU_SELECTED and update
> scx_qmap to test it instead of SCX_ENQ_WAKEUP.

Even if it's quite convenient to have the SCX_ENQ_CPU_SELECTED flag
provided by kernel, I was wondering if we could just delegate this whole
logic to BPF and avoid introducing this extra flag in the kernel.

In theory we could track when ops.select_cpu() is called by setting a
flag in the BPF task context (BPF_MAP_TYPE_TASK_STORAGE). Specifically,
the flag could be set in ops.select_cpu() and cleared in ops.stopping().
Then, when ops.enqueue() is called, we can check the flag to determine
whether ops.select_cpu() was skipped or not.

Since most of the scx schedulers already implement their own task
context, this shouldn't add too much complexity/overhead to the BPF
code, it'd be fully backward-compatible and it doesn't depend on the
particular kernel logic that calls ->select_task_rq(). WDYT?

Thanks,
-Andrea

> 
> Signed-off-by: Tejun Heo <tj@xxxxxxxxxx>
> Cc: David Vernet <void@xxxxxxxxxxxxx>
> Cc: Daniel Hodges <hodges.daniel.scott@xxxxxxxxx>
> Cc: Changwoo Min <multics69@xxxxxxxxx>
> Cc: Andrea Righi <andrea.righi@xxxxxxxxx>
> Cc: Dan Schatzberg <schatzberg.dan@xxxxxxxxx>
> ---
>  kernel/sched/ext.c             | 1 +
>  tools/sched_ext/scx_qmap.bpf.c | 4 ++--
>  2 files changed, 3 insertions(+), 2 deletions(-)
> 
> diff --git a/kernel/sched/ext.c b/kernel/sched/ext.c
> index c09e3dc38c34..9f00c8b629f1 100644
> --- a/kernel/sched/ext.c
> +++ b/kernel/sched/ext.c
> @@ -691,6 +691,7 @@ enum scx_enq_flags {
>  	/* expose select ENQUEUE_* flags as enums */
>  	SCX_ENQ_WAKEUP		= ENQUEUE_WAKEUP,
>  	SCX_ENQ_HEAD		= ENQUEUE_HEAD,
> +	SCX_ENQ_CPU_SELECTED	= ENQUEUE_RQ_SELECTED,
>  
>  	/* high 32bits are SCX specific */
>  
> diff --git a/tools/sched_ext/scx_qmap.bpf.c b/tools/sched_ext/scx_qmap.bpf.c
> index 83c8f54c1e31..588b7dce44fa 100644
> --- a/tools/sched_ext/scx_qmap.bpf.c
> +++ b/tools/sched_ext/scx_qmap.bpf.c
> @@ -230,8 +230,8 @@ void BPF_STRUCT_OPS(qmap_enqueue, struct task_struct *p, u64 enq_flags)
>  		return;
>  	}
>  
> -	/* if !WAKEUP, select_cpu() wasn't called, try direct dispatch */
> -	if (!(enq_flags & SCX_ENQ_WAKEUP) &&
> +	/* if select_cpu() wasn't called, try direct dispatch */
> +	if (!(enq_flags & SCX_ENQ_CPU_SELECTED) &&
>  	    (cpu = pick_direct_dispatch_cpu(p, scx_bpf_task_cpu(p))) >= 0) {
>  		__sync_fetch_and_add(&nr_ddsp_from_enq, 1);
>  		scx_bpf_dispatch(p, SCX_DSQ_LOCAL_ON | cpu, slice_ns, enq_flags);
> -- 
> 2.46.2
>