Re: Hashed locking for the namecache

[yamt%mwd.biglobe.ne.jp@localhost (YAMAMOTO Takashi)]

> On Sun, Apr 06, 2008 at 09:43:15PM +0900, YAMAMOTO Takashi wrote:
> > > Hi,
> > > 
> > > http://www.netbsd.org/~ad/cache.diff
> > > 
> > > This patch uses a hash of locks to cover forward lookup in the namecache.
> > > On an 8-core box it cuts the time for 'build.sh -j16 release' by about 5%.
> > > We could probably do something a lot more efficient with garbage 
> > > collection.
> > > But, this patch is simple and I think it could be a stepping stone towards
> > > a more better scheme.
> > > 
> > > Comments?
> > > 
> > > Thanks,
> > > Andrew
> > 
> > cache_lookup() sometimes modifies caches.  eg. "goto remove" cases.
> 
> Sigh - I didn't see that. Ok, here is a first cut of a patch that uses
> garbage collection to batch removals of entries from the forward lookup
> hash. The reclaim mechanism isn't very good and I think it may still have
> bugs.
> 
>       http://www.netbsd.org/~ad/cache2.diff
> 
> Andrew

it seems missing cpu_data.h diff.
i prefer to have a limit in cache_enter to prevent hogging the whole kva.
otherwise seems reasonable to me.

YAMAMOTO Takashi