Re: nfsd changes for 2.6.37

[Christoph Hellwig <hch@infradead.org>]

On Wed, Oct 27, 2010 at 11:16:46AM -0400, J. Bruce Fields wrote:
> On Wed, Oct 27, 2010 at 10:59:29AM -0400, Christoph Hellwig wrote:
> > On Wed, Oct 27, 2010 at 10:55:39AM -0400, J. Bruce Fields wrote:
> > > Hm, two problems:
> > > 	- We introduce the possibility of fcntl(fd, F_SETLEASE, F_UNLCK)
> > > 	  failing with ENOMEM.
> > 
> > splitt ->setlease into ->add_least and ->delete_lease.  No need to pass
> > in a structure for the later.  No need to return one either.
> 
> Sounds fine to me.
> 
> > 
> > > 	- fasync_helper(.,.,1,.) sleeps.  Argh.
> > 
> > That's not new..
> 
> So we could do
> 
> 	unlock_flocks();
> 	error = fasync_helper(fd, filp, 1, &fl->fl_fasync);
> 	lock_flocks();
> 
> and say, hey, we didn't introduce any new bug there.  But....
> 
> I don't know, maybe add a version of fasync_add_entry() that takes a
> preallocated fasync_struct??

Or just convert the lock to a sleeping mutex.  Now that we have adaptive
spinning the horrible behaviour that Willy saw years ago might not be
that bad any more.  That'll need some benchmarking, though.