Re: [RFC] NUMA futex hashing

[Eric Dumazet <dada1@cosmosbay.com>]

On Tuesday 08 August 2006 18:49, Eric Dumazet wrote:
> On Tuesday 08 August 2006 18:34, Nick Piggin wrote:
> > Eric Dumazet wrote:
> > > We certainly can. But if you insist of using mmap sem at all, then we
> > > have a problem.
> > >
> > > rbtree would not reduce cacheline bouncing, so :
> > >
> > > We could use a hashtable (allocated on demand) of size N, N depending
> > > on NR_CPUS for example. each chain protected by a private spinlock. If
> > > N is well chosen, we might reduce lock cacheline bouncing. (different
> > > threads fighting on different private futexes would have a good chance
> > > to get different cachelines in this hashtable)
> >
> > See other mail. We already have a hash table ;)
>
> Yes but still you want at FUTEX_WAIT time to tell the kernel the futex is
> private to this process.
>
> Giving the same info at FUTEX_WAKE time could avoid the kernel to make the
> second pass (using only a private futex lookup), avoiding again the
> mmap_sem touch in case no threads are waiting anymore on this futex.

After looking at kernel/futex.c, I realize we also can avoid the atomic ops 
(and another cacheline bouncing) done in get_key_refs()/drop_key_refs(), 
touching the inode i_count or mm_count refcounter)

Eric