On Thu, Oct 12, 2017 at 10:53:35AM +0200, Sebastian Andrzej Siewior wrote:
> On 2017-10-10 14:43:13 [-0700], Paul E. McKenney wrote:
> > diff --git a/kernel/rcu/srcutree.c b/kernel/rcu/srcutree.c
> > index 6d5880089ff6..558f9e7b283e 100644
> > --- a/kernel/rcu/srcutree.c
> > +++ b/kernel/rcu/srcutree.c
> > @@ -830,7 +866,7 @@ void __call_srcu(struct srcu_struct *sp, struct
> > rcu_head *rhp,
> > rhp->func = func;
> > local_irq_save(flags);
> > sdp = this_cpu_ptr(sp->sda);
> > - raw_spin_lock_rcu_node(sdp);
> > + spin_lock_rcu_node(sdp);
>
> This and the same thing in srcu_might_be_idle() does not work because
> local_irq_save() + spin_lock() != spin_lock_irqsave()
> but
> local_irq_save() + raw_spinlock = raw_spin_lock_irqsave()
>
> I think that preempt_disable() for a stable this_cpu_ptr() is enough
> here. I replaced local_irq_save() with local_lock_irqsave() on RT which
> provides a per-CPU spinlock (for mutual exclusion) and disables
> interrupts in !RT mode.
>
> I've been testing this for a while and it seems to work. Thank you.
So I keep mainline as is, and the local_irq_save()-to-local_lock_irqsave()
conversion happens in -rt, given that mainline doesn't have a
local_lock_irqsave(), correct?
And just so you know, there is one patchset adding call_srcu() that I am
following up on. Looks to me like it is OK with this change, but if not,
well, back to the drawing board... :-/
Thanx, Paul
> > rcu_segcblist_enqueue(&sdp->srcu_cblist, rhp, false);
> > rcu_segcblist_advance(&sdp->srcu_cblist,
> > rcu_seq_current(&sp->srcu_gp_seq));
>
>
> Sebastian
>
