Re: [PATCH sched_ext/for-6.18] sched_ext: Use cgroup_lock/unlock() to synchronize against cgroup operations

[Andrea Righi <arighi@nvidia.com>]

On Wed, Sep 03, 2025 at 10:57:39AM -1000, Tejun Heo wrote:
> On Wed, Sep 03, 2025 at 10:48:09PM +0200, Andrea Righi wrote:
> > On Wed, Sep 03, 2025 at 06:39:46AM -1000, Tejun Heo wrote:
> > > Hello,
> > > 
> > > On Wed, Sep 03, 2025 at 02:44:58PM +0200, Andrea Righi wrote:
> > > > >  static void scx_cgroup_lock(void)
> > > > >  {
> > > > > -	percpu_down_write(&scx_cgroup_rwsem);
> > > > > +	percpu_down_write(&scx_cgroup_ops_rwsem);
> > > > > +	cgroup_lock();
> > > > >  }
> > > > 
> > > > Shouldn't we acquire cgroup_lock() before scx_cgroup_ops_rwsem to avoid
> > > > a potential AB-BA deadlock?
> > > 
> > > There's no existing ordering between the two locks, so any order should be
> > > safe. The reason why I put it in this particular order is because any
> > > cgroup_lock() holder has no reason to grab ops_rwsem now or in the future
> > > while the opposite direction is still unlikely but theoretically more
> > > possible.
> > 
> > Isn't scx_group_set_weight() called with cgroup_mutex held? In this case
> > the order is cgroup_lock() -> scx_cgroup_ops_rwsem, or am I missing
> > something?
> 
> Oh, no, cgroup_lock() is only held for operations that change the cgroup
> hierarchy - cgroup creation, deletion, controller enable/disable, task
> migration and so on. Writes to control knobs doesn't acquire cgroup_lock().

Ah! That's the part I was missing, thanks for clarifying it. In that case:

Acked-by: Andrea Righi <arighi@nvidia.com>

-Andrea