Re: [patch] bitops: lock bitops

[Nick Piggin <npiggin@suse.de>]

On Wed, Jul 11, 2007 at 08:38:11PM -0700, David Miller wrote:
> From: Nick Piggin <npiggin@suse.de>
> Date: Thu, 12 Jul 2007 05:14:19 +0200
> 
> > Here is a set of patches that aims to mitigate some of the lock_page
> > overhead on powerpc introduced in the fault path by another set.
> > Fortunately it also improves various other things too :)
> > 
> > After this set, a dd if=./big-sparse-file of=/dev/null on my G5
> > goes from 563MB/s to 575MB/s, or about 80ns less time per page.
> > However I won't post the full set until after getting some acks
> > from the arch people now, because it is a fair bit of churn in core
> > code (eg. renaming !TestSetPageLocked to trylock_page).
> > 
> > Not sure who else can take advantage of these. Sparc64 probably.
> 
> What I would code up on sparc64 would be basically equivalent to your
> generic versions.
> 
> Or, more simply, I could just call test_and_clear_bit() et al.
> directly and ignore the return value.  There wouldn't be any
> difference.

Oh OK, I stand corrected then. I thought you could avoid the
PRE barrier on your test_and_set, and the POST barrier on
test_and_clear, however I don't know exactly what sparc's
barriers work.