Re: missing mnt_drop_write() on open error

[Dave Hansen <haveblue@us.ibm.com>]

On Wed, 2007-09-26 at 10:38 +0200, Miklos Szeredi wrote: 
> > On Wed, 2007-09-26 at 01:14 +0200, Miklos Szeredi wrote:
> > > I get this at umount, if there was a failed open():
> > > 
> > > WARNING: at fs/namespace.c:586 __mntput()
> > > 
> > > I think the problem is that may_open() calls mnt_want_write(), but if
> > > open doesn't succeed, mnt_drop_write() will not be called.
> > 
> > Does this help?
> 
> It didn't fix it for me, but the patch looks OK.
> 
> In __dentry_open() there's still a few places where fput() won't be
> called, notably when ->open fails, which is what I'm triggering I
> think.
>
> Also even more horrible things can happen because of the
> nd->intent.open.file thing.  For example if the lookup routine calls
> lookup_instantiate_filp(), and after this, but before may_open() some
> error happens, then release_open_intent() will call fput() on the
> file, which will cause mnt_drop_write() to be called, even though a
> matching mnt_want_write() hasn't yet been called.  Ugly, eh?

I used to have a patch that didn't completely trust that all files with
FMODE_WRITE set to have taken a write on the mnt.  I think I used a flag
to indicate whether or not a particular file had a mnt_want_write() done
on its behalf.  It somewhat artificially keeps the mnt write count
balanced, but I think it will let us detect when things like this go on.

> > I'm also thinking that we should change the open_namei*
> > functions to simply return 'struct file *'.  Those are the only users
> > other than NFS, and forcing the return of a file like that will force
> > users to do the fput() on it if they don't want it any more.  We'd just
> > need to make sure no new may_open() users pop up.  Any thoughts on that?
> 
> Yeah, something needs to be done with open, because currently it's way
> too convoluted.

Sounds like Christoph has some ideas...

-- Dave