Re: [PATCH sched_ext/for-6.16] sched_ext: Add RCU protection to scx_root in DSQ iterator

[Andrea Righi <arighi@nvidia.com>]

On Wed, Apr 30, 2025 at 08:59:06AM -1000, Tejun Heo wrote:
> On Wed, Apr 30, 2025 at 11:09:30AM +0200, Andrea Righi wrote:
> > @@ -6907,28 +6907,39 @@ __bpf_kfunc int bpf_iter_scx_dsq_new(struct bpf_iter_scx_dsq *it, u64 dsq_id,
> >  {
> >  	struct bpf_iter_scx_dsq_kern *kit = (void *)it;
> >  	struct scx_sched *sch;
> > +	int ret = 0;
> >  
> >  	BUILD_BUG_ON(sizeof(struct bpf_iter_scx_dsq_kern) >
> >  		     sizeof(struct bpf_iter_scx_dsq));
> >  	BUILD_BUG_ON(__alignof__(struct bpf_iter_scx_dsq_kern) !=
> >  		     __alignof__(struct bpf_iter_scx_dsq));
> >  
> > +	rcu_read_lock();
> >  	sch = rcu_dereference(scx_root);
> 
> Can you instead try changing rcu_derefernece(scx_root) to
> rcu_derefernece_check(scx_root, rcu_read_lock_bh_held())? The warning is
> because this can be called both from rcu-locked and bh contexts and while
> grabbing rcu_read_lock() once works, telling rcu that the dereference can be
> in both contexts is more accurate here.

Yeah, that seems better, also considering that bpf_iter_scx_dsq_next() is
marked with KF_RCU_PROTECTED, so checking for rcu_read_lock_bh_held()
should cover all the possible cases.

Will test that and send a v2.

Thanks!
-Andrea