Re: [RFC] Refcounting of objects part of a lockfree collection

[Dipankar Sarma <dipankar@in.ibm.com>]

On Wed, Jul 14, 2004 at 10:08:00AM -0700, Greg KH wrote:
> On Wed, Jul 14, 2004 at 02:27:58PM +0530, Ravikiran G Thirumalai wrote:
> > might be a bit yucky to assume atomic_t internals, but it is just one header
> > file :) <ducks>
> 
> I still think you need to fix this, manipulating atomic_t variables by
> hand is not always guaranteed to work on all arches, from what I
> remember.

AFAICS, the hash-locked refcounting grabs a spin lock for all
operations on the atomic_t. Any reason why that should not be safe ?
Of course, I can't see why we can't have two versions of the
reference counter depending on __HAVE_ARCH_CMPXCHG. Kiran ?

> 
> And what arches do not support cmpxchg?  How does this change affect the
> performance of them?

mips64, smp arm ?? ;-)

With a hashed lock, it should not be all that bad in low-end SMPs.
Besides we already use such a thing in gettimeofday implementation
with a global lock. However this is a valid issue and performance #s
from those arch users would be useful.

Thanks
Dipankar