Re: [PATCH] adapt page_lock_anon_vma() to PREEMPT_RCU

[Oleg Nesterov <oleg@tv-sign.ru>]

On 02/27, Andrew Morton wrote:
>
> > On Sun, 25 Feb 2007 23:06:21 +0300 Oleg Nesterov <oleg@tv-sign.ru> wrote:
> > page_lock_anon_vma() uses spin_lock() to block RCU. This doesn't work with
> > PREEMPT_RCU, we have to do rcu_read_lock() explicitely. Otherwise, it is
> > theoretically possible that slab returns anon_vma's memory to the system
> > before we do spin_unlock(&anon_vma->lock).
> > 
> > ...
> >
> > +static void page_unlock_anon_vma(struct anon_vma *anon_vma)
> > +{
> > +	spin_unlock(&anon_vma->lock);
> > +	rcu_read_unlock();
> >  }
> 
> It's a bit sad doing a double preempt_disable() for non-PREEMPT_RCU builds.

Actually, we don't in this case. This patch in essence moves "preempt_enable"
from "lock" to "unlock" side. Zero impact for non-PREEMPT_RCU builds, except
.text grows a bit.

Before this patch, page_lock_anon_vma() does preempt_enable() before return,
but this can't help because ->preempt_count was incremented by spin_lock().

> Perhaps we would benefit from a new rcu_read_lock_preempt_rcu() which is a
> no-op if !PREEMPT_RCU.

I also thought about things like

	rcu_read_lock_when_we_know_that_preemption_disabled()
	rcu_read_lock_when_we_know_that_irqs_disabled()

which are noops when !PREEMPT_RCU.

Oleg.