Re: Oi. NFS people. Read this.

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

On Wed, May 07, 2008 at 03:10:27PM -0700, Trond Myklebust wrote:
> On Wed, 2008-05-07 at 14:00 -0600, Matthew Wilcox wrote:
> > On Wed, May 07, 2008 at 12:44:48PM -0700, Linus Torvalds wrote:
> > > On Wed, 7 May 2008, Matthew Wilcox wrote:
> > > > 
> > > > One patch I'd still like Yanmin to test is my one from yesterday which
> > > > removes the BKL from fs/locks.c.
> > > 
> > > And I'd personally rather have the network-fs people test and comment on 
> > > that one ;)
> > > 
> > > I think that patch is worth looking at regardless, but the problems with 
> > > that one aren't about performance, but about what the implications are for 
> > > the filesystems (if any)...
> > 
> > Oh, well, they don't seem interested.
> 
> Poor timing: we're all preparing for and travelling to the annual
> Connectathon interoperability testing conference which starts tomorrow.
> 
> > I can comment on some of the problems though.
> > 
> > fs/lockd/svcsubs.c, fs/nfs/delegation.c, fs/nfs/nfs4state.c,
> > fs/nfsd/nfs4state.c all walk the i_flock list under the BKL.  That won't
> > protect them against locks.c any more.  That's probably OK for fs/nfs/*
> > since they'll be protected by their own data structures (Someone please
> > check me on that?), but it's a bad idea for lockd/nfsd which are walking
> > the lists for filesystems.
> 
> Yes. fs/nfs is just reusing the code in fs/locks.c in order to track the
> locks it holds on the server. We could alternatively have coded a
> private lock implementation, but this seemed easier.

So, assuming nfs is taking care of its own locking (I don't know if
that's right), that leaves nlm_traverse_locks() and nlm_file_inuse()
(both in fs/lockd/svcsubs.c) as the problem spots.

> > Are we going to have to export the file_lock_lock?  I'd rather not.  But
> > we need to keep nfsd/lockd from tripping over locks.c.
> > 
> > Maybe we could come up with a decent API that lockd could use?  It all
> > seems a bit complex at the moment ... maybe lockd should be keeping
> > track of the locks it owns anyway (since surely the posix deadlock
> > detection code can't work properly if it's just passing all the locks
> > through).
> 
> I'm not sure what you mean when you talk about lockd keeping track of
> the locks it owns. It has to keep those locks on inode->i_flock in order
> to make them visible to the host filesystem...
> 
> All lockd really needs, is the ability to find a lock it owns, and then
> obtain a copy.

That sounds right.

--b.

> As for the nfs client, I suspect we can make do with
> something similar...