sched_ext: ops.sub_ecaps_updated() dispatch context is set up on the wrong CPU under core scheduling
Hello,
scx_process_sync_ecaps() looks like a missed instance of the bug class 3dd52416e44a ("sched_ext: Fix this_rq() assumptions in dispatch kfuncs")
fixed: before invoking ops.sub_ecaps_updated() it does
struct scx_dsp_ctx *dspc = &pcpu->dsp_ctx; dspc->rq = rq;
where @pcpu is the *target* CPU's context (recovered from the llist node), while the dispatch kfuncs resolve theirs with this_cpu_ptr(sch->pcpu). Identical until dispatch_core_pick() started running balance_one() for sibling rqs - now a sync processed by a sibling calls the op with the executing CPU's dsp_ctx.rq unset.
Nothing ever writes a sub's dsp_ctx.rq on a CPU before the first scx_bpf_sub_dispatch() there, so it is NULL at attach time - exactly when the initial cap-grant notifications fire. If the op then calls e.g. scx_bpf_dsq_move_to_local() (it has SCX_KF_ALLOW_DISPATCH) on a non-empty DSQ, scx_consume_dispatch_q(sch, NULL, ...) reaches task_can_run_on_remote_rq() -> cpu_of(NULL) and oopses in the pick path. A leftover rq from an earlier dispatch instead consumes toward the wrong rq.
Fix seems to be a one-liner matching scx_dispatch_sched():
- struct scx_dsp_ctx *dspc = &pcpu->dsp_ctx; + struct scx_dsp_ctx *dspc = + &this_cpu_ptr(pcpu->sch->pcpu)->dsp_ctx;
with Fixes: b81a6c018cde ("sched_ext: Add sub_ecaps_updated() effective-cap change notifier"). Happy to send a patch if nobody is on it, or feel free to just fix it directly.
Cheers !