Re: [PATCH 2/2] mm: change mmap_sem over to the scalable rw_mutex

[Andrew Morton <akpm@linux-foundation.org>]

On Fri, 11 May 2007 19:12:16 +0200
Peter Zijlstra <a.p.zijlstra@chello.nl> wrote:

> (now with reply-all)
> 
> On Fri, 2007-05-11 at 09:17 -0700, Andrew Morton wrote:
> > On Fri, 11 May 2007 15:15:43 +0200 Peter Zijlstra <a.p.zijlstra@chello.nl> wrote:
> > 
> > > -	down_write(&current->mm->mmap_sem);
> > > +	rw_mutex_write_lock(&current->mm->mmap_lock);
> > 
> > y'know, this is such an important lock and people have had such problems
> > with it and so many different schemes and ideas have popped up that I'm
> > kinda thinking that we should wrap it:
> > 
> > 	write_lock_mm(struct mm_struct *mm);
> > 	write_unlock_mm(struct mm_struct *mm);
> > 	read_lock_mm(struct mm_struct *mm);
> > 	read_unlock_mm(struct mm_struct *mm);
> > 
> > so that further experimentations become easier?
> 
> Sure, can do; it'd require a few more functions than these, but its not
> too many. However, what is the best way to go about such massive rename
> actions? Just push them through quickly, and make everybody cope?

Well, if we _do_ decide to do this (is anyone howling?) then we can do

static inline void write_lock_mm(struct mm_struct *mm)
{
	down_write(&mm->mmap_sem);
}

and then let the conversions trickle into the tree in an orderly fashion.

Once we think all the conversions have landed, we rename mmap_sem to
_mmap_sem to avoid any backpedalling.