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

[Arnd Bergmann <arnd@arndb.de>]

On Wednesday 15 September 2010, Sage Weil wrote:
> > @@ -2362,7 +2362,7 @@ static int encode_caps_cb(struct inode *inode, struct ceph_cap *cap,
> >       if (recon_state->flock) {
> >               int num_fcntl_locks, num_flock_locks;
> >  
> > -             lock_kernel();
> > +             lock_flocks();
> >               ceph_count_locks(inode, &num_fcntl_locks, &num_flock_locks);
> >               rec.v2.flock_len = (2*sizeof(u32) +
> >                                   (num_fcntl_locks+num_flock_locks) *
> > @@ -2373,7 +2373,7 @@ static int encode_caps_cb(struct inode *inode, struct ceph_cap *cap,
> >                       err = ceph_encode_locks(inode, pagelist,
> >                                               num_fcntl_locks,
> >                                               num_flock_locks);
> > -             unlock_kernel();
> > +             unlock_flocks();
> >       }
> 
> The Ceph code won't currently behave with lock_flocks() taking a spinlock.  
> We're preparing a patch to fix that now.  As long as there is a window 
> between lock_flocks() being defined and the spinlock conversion, I can 
> send the fix upstream then and avoid any breakage.  Or send the patches 
> your way to include in your tree, whatever you prefer!

I'd be happy to just integrate the fix in this patch, or as a separate patch
in the series.

I certainly don't want to break any file system in the middle of the series,
I'm sure we can find a way to do it right.

What is the problem? I just saw ceph_pagelist_addpage potentially sleeping,
is that what you are thinking of?

> > +void unlock_flocks(void)
> > +{
> > +     unlock_kernel();
> > +}
> 
> EXPORT_SYMBOL(unlock_flocks); ?

Right, thanks!

	Arnd