Re: dentry bloat.

["Paul E. McKenney" <paulmck@us.ibm.com>]

On Mon, May 10, 2004 at 09:54:04AM -0500, Matt Mackall wrote:
> On Mon, May 10, 2004 at 02:46:58AM -0700, Andrew Morton wrote:
> > Arjan van de Ven <arjanv@redhat.com> wrote:
> > >
> > > On Mon, 2004-05-10 at 10:27, Helge Hafting wrote:
> > >  > Matt Mackall wrote:
> > >  > 
> > >  > >One also wonders about whether all the RCU stuff is needed on UP. I'm
> > >  > >not sure if I grok all the finepoints here, but it looks like the
> > >  > >answer is no and that we can make struct_rcu head empty and have
> > >  > >call_rcu fall directly through to the callback. This would save
> > >  > >something like 16-32 bytes (32/64bit), not to mention a bunch of
> > >  > >dinking around with lists and whatnot.
> > >  > >
> > >  > >So what am I missing?
> > >  > >  
> > >  > >
> > >  > Preempt can happen anytime, I believe.
> > > 
> > >  ok so for UP-non-preempt we can still get those 16 bytes back from the
> > >  dentry....
> > 
> > I suppose so.  And on small SMP, really.  We chose not to play those games
> > early on so the code got the best testing coverage.
> 
> Ok, I can spin something up. I'll start with a generic no-RCU-on-UP
> and then we can think about the small SMP case a bit later.

Hello, Matt,

You may wish to start with the dcache portion of Dipankar's earlier patch:

http://sourceforge.net/mailarchive/forum.php?thread_id=3644163&forum_id=730

It does not remove the extra rcu_head from the dentry, but does the
rest of the work.

						Thanx, Paul