Re: [patch 04/10] fs: fs_struct rwlock to spinlock

[Nick Piggin <npiggin@kernel.dk>]

On Tue, Aug 17, 2010 at 07:14:42PM -0400, Valerie Aurora wrote:
> On Wed, Aug 18, 2010 at 04:37:33AM +1000, Nick Piggin wrote:
> > -	read_lock(&current->fs->lock);
> > +	spin_lock(&current->fs->lock);
> >  	root = dget(current->fs->root.dentry);
> > -	read_unlock(&current->fs->lock);
> > +	spin_unlock(&current->fs->lock);
> >  
> >  	spin_lock(&dcache_lock);
> 
> Your reasoning makes sense to me.  Shared reader access seems very
> unlikely whereas the cost of taking the lock is certain.

Yes, shared reader will only help if we have multiple threads inside
the critical section at the same time, and if they can actually get
any extra parallelism (which they can't, because all they're doing here
is just hitting another contended cacheline).

I doubt this will change scalability at all, but it will improve single
threaded performance a little bit.

With the store-free path walk, I can actually do some tricks to entirely
remove the fs->lock spinlock in the common case here (using a seqlock
instead). Combined with avoiding refcounts to the cwd, so it avoids all
the contention between threads.