Re: [patch 0/9] mutex subsystem, -V4

[Nicolas Pitre <nico@cam.org>]

On Thu, 22 Dec 2005, Andrew Morton wrote:

> Christoph Hellwig <hch@infradead.org> wrote:
> > I really don't get why you hate mutex primitives so much.
> 
> I've just spelled out in considerable detail why this work is premature. 
> How can you not "get" it?  Why do I have to keep saying the same thing in
> different ways?  It's really quite simple.
> 
> 
> So here is permutation #4:
> 
> If we can optimise semaphores for speed and space, the mutexes are
> *unneeded*.

How can't you get the fact that semaphores could _never_ be as simple as 
mutexes?  This is a theoritical impossibility, which maybe turns out not 
to be so true on x86, but which is damn true on ARM where the fast path 
(the common case of a mutex) is significantly more efficient.

Semaphores _require_ an atomic decrement, mutexes do not.  On some 
architectures that makes a huge difference.

> And I think we _should_ optimise semaphores for speed and space.  Don't you?

No one disagrees with that.

> If we can do that, there is no point at all in adding a new lock type which
> has no speed advantage and no space advantage and which has less
> functionality than semaphores.

The very point is that mutexes will always have a speed advantage by 
nature.

> And, repeating myself yet again: if we can demonstrate that it is not
> feasible to optimise semaphores to the same performance and space efficiency
> of mutexes then (and only then) we have a reason for adding mutexes.

I spent the whole week making that demonstration repeatedly.  Why are 
you ignoring me?


Nicolas