Re: [PATCH sched_ext/for-6.18] sched_ext: Acquire task reference in scx_bpf_cpu_curr()

[Andrea Righi <arighi@nvidia.com>]

Hi Tejun,

On Tue, Sep 09, 2025 at 10:01:16AM -1000, Tejun Heo wrote:
> Hello, Andrea.
> 
> On Tue, Sep 09, 2025 at 09:57:09PM +0200, Andrea Righi wrote:
> > scx_bpf_cpu_curr() has been introduced to retrieve the current task of a
> > given runqueue, allowing schedulers to interact with that task.
> > 
> > The kfunc assumes that it is always called in an RCU context, but this
> > is not always guaranteed and some BPF schedulers can trigger the
> > following warning:
> > 
> >   WARNING: suspicious RCU usage
> >   sched_ext: BPF scheduler "cosmos_1.0.2_gd0e71ca_x86_64_unknown_linux_gnu_debug" enabled
> >   6.17.0-rc1 #1-NixOS Not tainted
> >   -----------------------------
> >   kernel/sched/ext.c:6415 suspicious rcu_dereference_check() usage!
> > 
> > The correct behavior is to acquire a reference to the returned task, so
> > the scheduler can safely access it and then release it with
> > bpf_task_release().
> > 
> > Update the kfunc and the corresponding compatibility helper to implement
> > reference acquisition and prevent potential RCU warnings.
> 
> I think KF_RCU likely fits better for peeking kernel data structures than
> having to acquire/release them. Can you post the full backtrace? Is it being
> called from a sleepable bpf prog? Or is it that we just need to expand the
> rcu check scope to cover regular rcu, bh and sched? And, everything aside,
> if KF_RCU, should we be tripping on rcu_dereference() in the first place?

For the records, as discussed offline, we should be fine marking the kfunc
as KF_RCU_PROTECTED instead of acquiring the reference to the task.

Right now the kfunc is marked as KF_RCU, which is not really necessary,
because KF_RCU ensures the kfunc *arguments* are either RCU-protected or
trusted.

KF_RCU_PROTECTED, instead, should ensure that the kfunc is called inside an
RCU read-side critical section, that is what we need.

In this way the kfunc can safely return a pointer to the task and sleepable
BPF programs can wrap the call in a bpf_rcu_read_lock/unlock() section.
This should prevent the RCU warning while still letting schedulers safely
use the returned task.

I'll send a new patch with a proper fix.

Thanks,
-Andrea