Re: [PATCH] Fix the cpumask rewrite

[Vojtech Pavlik <vojtech@suse.cz>]

On Sat, Jun 26, 2004 at 11:54:24AM -0700, Linus Torvalds wrote:

> On Sat, 26 Jun 2004, Vojtech Pavlik wrote:
> > 
> > At least input pretty much relies on the fact that bitops don't need
> > locking and act as memory barriers.
> 
> Well, plain test_bit() has always been more relaxed than the others, and
> has never implied a memory barrier. Only the "test_and_set/clear()" things
> imply memory barriers.

Ouch. I'll have to revisit some code then.

> What we _could_ do (without changing any existing rules) is to add a
> "__test_bit()" that is the relaxed version that doesn't do any of the
> volatile etc. That would match the "__"  versions of the other bit
> operations.
> 
> Then people who know that they use the bits without any volatility issues 
> can use that one, and let the compiler optimize more. 
> 
> Hmm?
 
That makes a lot of sense to me, as we already have the __ variants for
most of the other bitops already.

-- 
Vojtech Pavlik
SuSE Labs, SuSE CR