Re: [RFC] unify semaphore implementations

[Trond Myklebust <trond.myklebust@fys.uio.no>]

fr den 29.04.2005 Klokka 10:14 (-0400) skreiv Benjamin LaHaise:
> On Fri, Apr 29, 2005 at 10:44:17AM +1000, Paul Mackerras wrote:
> > You have made semaphores bigger and slower on the architectures that
> > have load-linked/store-conditional instructions, which is at least
> > ppc, ppc64, sparc64 and alpha.  Did you take the trouble to understand
> > the ppc semaphore implementation?
> 
> The ppc implementation does have some good ideas that are worth using.  
> It's hard to know which of the 23 versions were worth using, but I'm 
> getting a picture where at least 2 variants are need.  The atomic ops 
> variant should use the single counter as ppc does (why did nobody port 
> that to x86?).  A spinlock version is needed at least by parisc.

The PPC implementation would be hard to port to x86, since it relies on
the load-linked/store-conditional stuff to provide a fast primitive for
atomic_dec_if_positive().
The only way I found to implement that on x86 was to use cmpxchg. On my
machine, therefore, a spinlock-based semaphore implementation turns out
to be at least as fast for the "fast" path (and is naturally much more
efficient for the slow path).

Cheers,
  Trond

-- 
Trond Myklebust <trond.myklebust@fys.uio.no>