Re: circular locking dependency detected panic in filldir when CONFIG_PROVE_LOCKING=y

[Peter Zijlstra <a.p.zijlstra@chello.nl>]

On Sat, 2009-09-19 at 09:23 +0200, Ingo Molnar wrote:
> * Trond Myklebust <trond.myklebust@fys.uio.no> wrote:
> 
> > On Fri, 2009-09-18 at 17:15 -0400, Joe Korty wrote:
> > > I experienced a might_fault panic from NFS's use of filldir
> > > in a 2.6.31 kernel compiled with CONFIG_PROVE_LOCKING=y.
> > > 
> > > Looking at filldir, I see it is accessing user space with
> > > __put_dir's (which are inatomic) and with one copy_to_user
> > > (which is not inatomic).  It is the single copy_to_user
> > > which is causing the might_fault panic.
> > > 
> > > It doesn't make any sense to be mixing use of inatomic
> > > and non-inatomic services in filldir.  Either all should be
> > > the inatomic version, or none should be.
> > > 
> > > The might_fault condition being reported by the panic looks
> > > real to me, so I suspect the wrong answer is converting
> > > everything to the inatomic version, since that just
> > > suppresses the circular dependency check while leaving
> > > the circular dependency in place.
> > 
> > Yes. This is known... Please see 
> >    http://thread.gmane.org/gmane.linux.nfs/28578
> > and
> >    http://thread.gmane.org/gmane.linux.nfs/27406
> > 
> > I'm still hoping the VM folks can see fit to merge Peter's fix at some
> > point...
> 
> Ouch, those patches at:
> 
>   http://programming.kicks-ass.net/kernel-patches/mmap-vs-nfs/
> 
> .... are 2 years old. Higher intensity prodding needed to get this 
> moving?

No I just need to find a way to clone() myself :-)