Re: [patch 0/4] mm: de-skew page refcount

[Linus Torvalds <torvalds@osdl.org>]

On Wed, 18 Jan 2006, Nick Piggin wrote:
>
> The following patchset (against 2.6.16-rc1 + migrate race fixes) uses the new
> atomic ops to do away with the offset page refcounting, and simplify the race
> that it was designed to cover.
> 
> This allows some nice optimisations

Why?

The real downside is that "atomic_inc_nonzero()" is a lot more expensive 
than checking for zero on x86 (and x86-64).

The reason it's offset is that on architectures that automatically test 
the _result_ of an atomic op (ie x86[-64]), it's easy to see when 
something _becomes_ negative or _becomes_ zero, and that's what

	atomic_add_negative
	atomic_inc_and_test

are optimized for (there's also "atomic_dec_and_test()" which reacts on 
the count becoming zero, but that doesn't have a pairing: there's no way 
to react to the count becoming one for the increment operation, so the 
"atomic_dec_and_test()" is used for things where zero means "free it").

Nothing else can be done that fast on x86. Everything else requires an 
insane "load, update, cmpxchg" sequence.

So I disagree with this patch series. It has real downsides. There's a 
reason we have the offset.

I suspect that whatever "nice optimizations" you have are quite doable 
without doing this count pessimization.

		Linus