Re: [PATCH v2 sched_ext/for-7.3 27/36] sched_ext: Route task slice writes through set_task_slice()

[sashiko-bot@kernel.org]

Thank you for your contribution! Sashiko AI review found 1 potential issue(s) to consider:
- [High] Routing the lazy fallback assignment through set_task_slice() in scx_bpf_dsq_insert___v2() and scx_dsq_insert_vtime() incorrectly drops pending out-of-band slice requests when preserving the current slice.
--

commit 127371a444f95f70296cadcc3202bbe1233bead6
Author: Tejun Heo <tj@kernel.org>

sched_ext: Route task slice writes through set_task_slice()

This commit introduces set_task_slice() to centralize writes to p->scx.slice. 
It replaces direct assignments across the sched_ext code in preparation for 
dropping pending out-of-band slice requests on slice writes.

> diff --git a/kernel/sched/ext/ext.c b/kernel/sched/ext/ext.c
> index 9fa4cfbc75c84..ff997ba881b02 100644
> --- a/kernel/sched/ext/ext.c
> +++ b/kernel/sched/ext/ext.c

[ ... ]

> @@ -8021,9 +8027,9 @@ __bpf_kfunc bool scx_bpf_dsq_insert___v2(struct task_struct *p, u64 dsq_id,
>  		return false;
>  
>  	if (slice)
> -		p->scx.slice = slice;
> +		set_task_slice(p, slice);
>  	else
> -		p->scx.slice = p->scx.slice ?: 1;
> +		set_task_slice(p, p->scx.slice ?: 1);

[Severity: High]
Will this unconditional call to set_task_slice() accidentally drop pending
out-of-band requests when slice == 0 and p->scx.slice is already non-zero?

The commit message mentions that a later change will make every write
through set_task_slice() drop pending out-of-band slice requests. If a BPF
scheduler passes slice == 0 to scx_bpf_dsq_insert___v2() intending to
preserve the current slice, routing this fallback through set_task_slice()
performs an explicit in-band write of the old slice.

Could this cause the task to lose its assigned slice budget from an
out-of-band request (e.g. set via scx_bpf_task_set_slice()), reverting to
its old slice or a 1ns slice and potentially causing starvation?

Should this avoid calling set_task_slice() if p->scx.slice is already
non-zero? For example: else if (!p->scx.slice) set_task_slice(p, 1);

>  
>  	scx_dsq_insert_commit(sch, p, dsq_id, enq_flags);
>  
> @@ -8047,9 +8053,9 @@ static bool scx_dsq_insert_vtime(struct scx_sched *sch, struct task_struct *p,
>  		return false;
>  
>  	if (slice)
> -		p->scx.slice = slice;
> +		set_task_slice(p, slice);
>  	else
> -		p->scx.slice = p->scx.slice ?: 1;
> +		set_task_slice(p, p->scx.slice ?: 1);

[Severity: High]
Similarly for scx_dsq_insert_vtime(), does unconditional use of
set_task_slice() here clear the pending out-of-band request when preserving
the existing slice?

-- 
Sashiko AI review · https://sashiko.dev/#/patchset/20260706014058.439853-1-tj@kernel.org?part=27