[PATCH] Mantaining turnstile aligned to 128 bytes in i386 CPUs

[Attilio Rao]

2006/7/26, Attilio Rao <attilio at freebsd.org>:
> 2006/7/25, Attilio Rao <attilio at freebsd.org>:
> > 2006/7/25, John Baldwin <jhb at freebsd.org>:
> > > On Tuesday 25 July 2006 11:14, Attilio Rao wrote:
> > > > 2006/7/25, Attilio Rao <attilio at freebsd.org>:
> > > > > Hi,
> > > > > Intel documentation points out that having a 128-bytes aligned
> > > > > syncronizing primitive  (which fits in a cache line) will minimize the
> > > > > traffic for cache bus, so this patch implements an alignment for i386
> > > > > on turnstiles.
> > > > >
> > > > > Any comments, feedbacks?
> > > >
> > > > Oh, sorry, I've unforgotten the diff.
> > > >
> > > > Attilio
> > >
> > > I think a better approach would be to stick turnstiles (and sleepqueues) in a
> > > UMA zone and specify cache-size alignment to the zone.  However, turnstiles
> > > aren't really sychronization primitives in that you don't spin on a variable
> > > inside the structure, and I think it's the spinning and avoiding bouncing
> > > cache lines around that Intel's documentation is really about.  In that case,
> > > the things you want aligned are things like mutexes, rwlocks, etc.
> >
> > Well, I think that this is referred in particular to the latter issue
> > you mentioned.
> > Spinning is not really concerned to cache bus issues (more, in
> > particular, datapath latency).
> > With this point of view, turnstiles (as sleepqueues) are passed around
> > CPUs more than a mutex/rwlock (or a cv), I guess, so I was thinking
> > that it's better optimizing turnstile than the real syncronizing
> > primitive itself.
>
> This is a patch which let turnstiles/sleepqueues using an UMA zone.
>
> I've tried in my 6.1R branch and it works quite fine, so this HEAD
> version might be alright (I've not tried yet, so please test):
> http://users.gufi.org/~rookie/works/patches/uma_sync.diff
>
> It, obviously, set default alignment for i386 at 128 bytes.
> Any comments, feedbacks, ideas, are welcome.
>
> Attilio
>
> PS: I know that I could simplify *_alloc(), *_free() routines
> implementing init/fini but it is simpler and more optimized having
> things like so.

After some thinking, I think it's better using init/fini methods
(since they hide the sizeof(struct turnstile) with size parameter).

Feedbacks and comments are welcome:
http://users.gufi.org/~rookie/works/patches/uma_sync_init.diff

Thanks,
Attilio


-- 
Peace can only be achieved by understanding - A. Einstein