Re: [patch 02/10] emm: notifier logic

[Andrea Arcangeli <andrea@qumranet.com>]

On Sun, Apr 06, 2008 at 10:48:56PM -0700, Christoph Lameter wrote:
> On Sat, 5 Apr 2008, Andrea Arcangeli wrote:
> 
> > > +	rcu_assign_pointer(mm->emm_notifier, e);
> > > +	mm_unlock(mm);
> > 
> > My mm_lock solution makes all rcu serialization an unnecessary
> > overhead so you should remove it like I already did in #v11. If it
> > wasn't the case, then mm_lock wouldn't be a definitive fix for the
> > race.
> 
> There still could be junk in the cache of one cpu. If you just read the 
> new pointer but use the earlier content pointed to then you have a 
> problem.

There can't be junk, spinlocks provides semantics of proper memory
barriers, just like rcu, so it's entirely superflous.

There could be junk only if any of the mmu_notifier_* methods would be
invoked _outside_ the i_mmap_lock and _outside_ the anon_vma and
outside the mmap_sem, that is never the case of course.

> So a memory fence / barrier is needed to guarantee that the contents 
> pointed to are fetched after the pointer.

It's not needed... if you were right we could never possibly run a
list_for_each inside any spinlock protected critical section and we'd
always need to use the _rcu version instead. The _rcu version is
needed only when the list walk happens outside the spinlock critical
section of course (rcu = no spinlock cacheline exlusive write
operation in the read side, here the read side takes the spinlock big time).