Re: [patch 2/21] reduced locking in buffer.c

[Andrew Morton <akpm@zip.com.au>]

Christoph Hellwig wrote:
> 
> On Tue, Aug 13, 2002 at 10:53:59AM -0700, Andrew Morton wrote:
> > I have discussed it with David - he said it's OK in 2.5, but
> > not in 2.4, and he has eyeballed the diff.
> >
> > However there's another thing to think about:
> >
> >       local_irq_disable();
> >       atomic_inc();
> >
> > If the architecture implements atomic_inc with spinlocks, this will
> > schedule with interrupts off with CONFIG_PREEMPT=y, I expect.
> >
> > I can fix that with a preempt_disable() in there, but ick.
> 
> Is there a reason you can't just use brlocks?

I didn't use brlocks in the initial code because I wanted the lock
in the same cacheline as the data it's locking.

And this code removes the locking altogether.

I suspect the lock traffic is in the noise compared with all the
get_bh, brelse, set_bit and clear_bit operations but it's a start.
We don't have a tool to measure those other things ;)