scx_p2dq: fix RCU pointer violation in bpf_for_each dispatch loop

[scrossley11]

The bpf_for_each(scx_dsq, ...) iterator yields a non-trusted task pointer. Accessing p->cpus_ptr directly on this pointer fails BPF verification on 6.16 kernels which require RCU pointers for cpus_ptr.

Replace direct cpus_ptr accesses with the existing safe helpers (peek_cpumask_test_cpu, peek_cpumask_any_distribute) that obtain a trusted reference via bpf_task_from_pid() before accessing cpus_ptr.

This also fixes scx_chaos which includes p2dq's code via main.bpf.inc.

[htejun]

This is for compatibility, right? Newer kernels shouldn't require trusted pointer. Maybe it'd be useful to add comment explaining that or move it to compat.bpf.h?

[scrossley11]

This is for compatibility, right? Newer kernels shouldn't require trusted pointer. Maybe it'd be useful to add comment explaining that or move it to compat.bpf.h?

Newer than 6.16? If not then maybe this is not the right fix

[likewhatevs]

p2dq should be good on stables going back to 6.13 per ci:

https://github.com/sched-ext/scx/actions/runs/23822398184/job/69455861674?pr=3499#step:9:1463

that runs the verifier with the topologies + everything else (i.e. rodata) of nothing (kinda invalid imo, but what we've always done) and of a 9950x (pretty valid).

[hodgesds]

I think the problem is on some backported kernels it has trouble verifying.