Re: [PATCH tip/core/rcu 03/17] rcu/tree: Skip entry into the page allocator for PREEMPT_RT

[Joel Fernandes]

On Mon, Jul 06, 2020 at 09:55:57PM +0200, Uladzislau Rezki wrote:
[...]
> > > Another way of fixing it is just dropping the lock letting the page
> > > allocator to do an allocation without our "upper/local" lock. I did a
> > > proposal like that once upon a time, so maybe it is a time to highlight
> > > it again:
> > > <snip>
> > > diff --git a/kernel/rcu/tree.c b/kernel/rcu/tree.c
> > > index 21c2fa5bd8c3..249f10a89bb9 100644
> > > --- a/kernel/rcu/tree.c
> > > +++ b/kernel/rcu/tree.c
> > > @@ -3278,9 +3278,11 @@ static void kfree_rcu_monitor(struct work_struct *work)
> > >  }
> > > 
> > >  static inline bool
> > > -kvfree_call_rcu_add_ptr_to_bulk(struct kfree_rcu_cpu *krcp, void *ptr)
> > > +kvfree_call_rcu_add_ptr_to_bulk(struct kfree_rcu_cpu *krcp,
> > > +       void *ptr, unsigned long *flags)
> > >  {
> > >         struct kvfree_rcu_bulk_data *bnode;
> > > +       struct kfree_rcu_cpu *tmp;
> > >         int idx;
> > > 
> > >         if (unlikely(!krcp->initialized))
> > > @@ -3306,6 +3308,9 @@ kvfree_call_rcu_add_ptr_to_bulk(struct kfree_rcu_cpu *krcp, void *ptr)
> > >                         if (IS_ENABLED(CONFIG_PREEMPT_RT))
> > >                                 return false;
> > > 
> > > +                       migrate_disable();
> > > +                       krc_this_cpu_unlock(krcp, *flags);
> > 
> > If I remember, the issue here is that migrate_disable is not implemented on a
> > non-RT kernel due to issues with starvation.
> > 
> It is implemented. Please have a look linux/preempt.h for regular kernel.

Yeah sorry, I meant it is not implemented in the right way in the sense - it
disables migration with the preempt-disable hammer.

Anyway, I think as we discussed on IRC, your proposal will work for both
kernels while allowing us to keep the internal lock as raw.

thanks,

 - Joel