Re: [RFC][PATCH] Avoid vmtruncate/mmap-page-fault race

[Paul E. McKenney]

On Thu, May 29, 2003 at 07:39:47PM +0200, Daniel Phillips wrote:
> On Thursday 29 May 2003 19:15, Daniel Phillips wrote:
> > On Thursday 29 May 2003 18:33, you wrote:
> > > Me?  I much preferred your original, much sparer, nopagedone patch
> > > (labelled "uglyh as hell" by hch).
> >
> > "me too".
> 
> Oh wait, I mispoke... there is another formulation of the patch that hasn't 
> yet been posted for review.  Instead of having the nopagedone hook, it turns 
> the entire do_no_page into a hook, per hch's suggestion, but leaves in the 
> ->nopage hook, which makes the patch small and obviously right.  I need to 
> post that version for comparison, please bear with me.
> 
> IMHO, it's nicer than the ->nopagedone form.

I put together something like this, but the problem with it is that
do_anonymous_page() needs the mm->page_table_lock held, but the
->nopage functions want this lock not to be held.  One could require
that all the lock be held on entry to all ->nopage functions, but
this would require almost all ->nopage functions to drop the lock
immediately upon entry.  This seemed error-prone to me, but could
certainly be done...

Thoughts?  Me, I don't care as long as there is some reasonable
way for distributed filesystems to safely resolve the race between
page faults and invalidation requests from other nodes.  ;-)

						Thanx, Paul