Re: [PATCH] sched/fair: Clear ->h_load_next after hierarchical load

[Krister Johansen <kjlx@templeofstupid.com>]

On Thu, Oct 16, 2025 at 11:06:17AM +0800, Peng Wang wrote:
> On Wed, Oct 15, 2025 at 02:44:22PM +0200, Peter Zijlstra wrote:
> > On Wed, Oct 15, 2025 at 08:19:50PM +0800, Peng Wang wrote:
> > 
> > > We found that the task_group corresponding to the problematic se
> > > is not in the parent task_group’s children list, indicating that
> > > h_load_next points to an invalid address. Consider the following
> > > cgroup and task hierarchy:
> > > 
> > >          A
> > >         / \
> > >        /   \
> > >       B     E
> > >      / \    |
> > >     /   \   t2
> > >    C     D
> > >    |     |
> > >    t0    t1
> > > 
> > > Here follows a timing sequence that may be responsible for triggering
> > > the problem:
> > > 
> > > CPU X                   CPU Y                   CPU Z
> > > wakeup t0
> > > set list A->B->C
> > > traverse A->B->C
> > > t0 exits
> > > destroy C
> > >                         wakeup t2
> > >                         set list A->E           wakeup t1
> > >                                                 set list A->B->D
> > >                         traverse A->B->C
> > >                         panic
> > > 
> > > CPU Z sets ->h_load_next list to A->B->D, but due to arm64 weaker memory
> > > ordering, Y may observe A->B before it sees B->D, then in this time window,
> > > it can traverse A->B->C and reach an invalid se.
> > 
> > Hmm, I rather think we should ensure update_cfs_rq_h_load() is
> > serialized against unregister_fair_sched_group().
> 
> I might be mistaken, but it seems that, even with RCU protection around
> update_cfs_rq_h_load(), there remains a risk of reading stale values.
> 
> 
>  CPU X                   CPU Y                   CPU Z
> 
>  wakeup t0
>  rcu_read_lock()
>  set list A->B->C
>  traverse A->B->C
>  rcu_read_unlock()
>  t0 exits
>  destroy C
> 
> After the prior RCU grace period has elapsed, C has already been reclaimed,
> yet the stale A->B->C remains.
> 
> 
>                          wakeup t2
>                          rcu_read_lock()
>                          set list A->E           wakeup t1
>                                                  rcu_read_lock()
>                                                  set list A->B->D
>                                                       ...
>                          traverse A->B->C
>                          panic
> 
> A subsequent rcu_read_lock() only guarantees that A/B/D/E will not be
> reclaimed while the list is being traversed; C had already been freed
> before the next grace period even began.

FWIW, I've caught arm64 machines running into this problem recently on
6.x kernels.  These particular systems are small enough that they have
just a single memory node and no NUMA balancing enabled.

Would the scheduling experts be willing to consider picking up Peng's
fix while the 6.18 release is still open for bug fixes?

-K