Re: [PATCH 2/2] [RFC] Remove BKL from fs/locks.c

["J. Bruce Fields" <bfields@fieldses.org>]

On Wed, Apr 14, 2010 at 04:52:14PM -0400, Trond Myklebust wrote:
> On Wed, 2010-04-14 at 22:36 +0200, Arnd Bergmann wrote: 
> > From: Matthew Wilcox <willy@linux.intel.com>
> > 
> > I've taken a patch originally written by Matthew Wilcox and
> > ported it to the current version. It seems that there were
> > originally concerns that this breaks NFS, but since Trond
> > has recently removed the BKL from NFS, my naive assumption
> > would be that it's all good now, despite not having tried to
> > understand what it does.
> 
> Hi Arnd,
> 
> We still need to fix up the bits in NFS that dereference inode->i_flock.
> On the client side, those are mainly the bits that deal with lock
> recovery when the NFS server has rebooted or restarted.
> 
> AFAICS, there are two places in the NFSv4 client that need to be changed
> to call lock_flocks(): nfs_delegation_claim_locks(), and
> nfs4_reclaim_locks(). In both cases, the replacement is trivial.
> 
> For NFSv3, I think we are already safe, since AFAICS the host->h_rwsem
> already provides exclusion between file locking and lock recovery
> attempts. I think we should therefore be able to immediately remove the
> BKL in fs/lockd/clntlock.c:reclaimer().
> 
> I'm not as sure about how sensitive the NFS server is to the switch from
> BKL -> lock_flocks(). Perhaps Bruce can comment...

There's a callback from the lease code, the implementation of which
sleeps in the NFSv4 server case.  I've got patches for the next merge
window which remove that sleep (by just deferring work to a workqueue).

release_lockowner() traverses the lock list.  I need to fix that.

The lockd thread is entirely under the BKL--it takes it at the top of
fs/lockd/svc.c:lockd(), and drops it at the bottom.  It's possible
there may be code that lockd runs that assumes mutual exclusion with
code in locks.c, but I don't know.

--b.