Re: [RFC PATCH bpf-next 2/8] sched_ext: Add filter for scx_kfunc_ids_select_cpu

[Juntong Deng]

On 2025/2/6 23:39, Andrea Righi wrote:
> On Wed, Feb 05, 2025 at 07:30:14PM +0000, Juntong Deng wrote:
> ...
> > +static int scx_kfunc_ids_other_rqlocked_filter(const struct bpf_prog *prog, u32 kfunc_id)
> > +{
> > +	u32 moff = prog->aux->attach_st_ops_member_off;
> > +
> > +	if (moff == offsetof(struct sched_ext_ops, runnable) ||
> > +	    moff == offsetof(struct sched_ext_ops, dequeue) ||
> > +	    moff == offsetof(struct sched_ext_ops, stopping) ||
> > +	    moff == offsetof(struct sched_ext_ops, quiescent) ||
> > +	    moff == offsetof(struct sched_ext_ops, yield) ||
> > +	    moff == offsetof(struct sched_ext_ops, cpu_acquire) ||
> > +	    moff == offsetof(struct sched_ext_ops, running) ||
> > +	    moff == offsetof(struct sched_ext_ops, core_sched_before) ||
> > +	    moff == offsetof(struct sched_ext_ops, set_cpumask) ||
> > +	    moff == offsetof(struct sched_ext_ops, update_idle) ||
> > +	    moff == offsetof(struct sched_ext_ops, tick) ||
> > +	    moff == offsetof(struct sched_ext_ops, enable) ||
> > +	    moff == offsetof(struct sched_ext_ops, set_weight) ||
> > +	    moff == offsetof(struct sched_ext_ops, disable) ||
> > +	    moff == offsetof(struct sched_ext_ops, exit_task) ||
> > +	    moff == offsetof(struct sched_ext_ops, dump_task) ||
> > +	    moff == offsetof(struct sched_ext_ops, dump_cpu))
> > +		return 0;
> > +
> > +	return -EACCES;
>
> Actually, do we need this filter at all?
>
> I think the other filters in your patch set should be sufficient to
> establish the correct permissions for all kfuncs, as none of them need to
> be called from any rq-locked operations. Or am I missing something?

Thanks for your reply.

I think I misunderstood SCX_KF_REST.

I incorrectly thought that all but SCX_KF_UNLOCKED belonged to
SCX_KF_REST (including SCX_KF_CPU_RELEASE, SCX_KF_DISPATCH, etc.).

I will remove scx_kfunc_ids_other_rqlocked_filter in the next version.

If you find any other mistakes, please let me know.

> -Andrea