Re: INFO: possible circular locking dependency detected 3.1.0-rc2-00190-g3210d19

[Al Viro <viro@ZenIV.linux.org.uk>]

On Mon, Aug 22, 2011 at 02:16:21PM +0100, Al Viro wrote:
> On Mon, Aug 22, 2011 at 09:09:14AM -0400, Josh Boyer wrote:
> 
> > We've had a report of this on 3.0.1 as well.  Slightly different
> > scenario and fs, but the locks in question are the same.
> > https://bugzilla.redhat.com/show_bug.cgi?id=730998
> > 
> > It seems that with CONFIG_PROVE_LOCKING on, might_fault will always
> > attempt to grab mm->mmap_sem.  The common flow here is that getdents
> > calls filldir, which calls copy_to_user, which is what is calling
> > might_fault.
> > 
> > Beyond that, I'm a bit over my head at the moment because I don't know
> > if the VFS is right and we just need some more lockdep annotations or
> > if there really is a problem.
> 
> Don't grab ->i_mutex in ->evict_inode().  Why are you doing that, anyway?

Note, BTW, that readdir() is a red herring here; there is a much more
relevant reason for that ranking.  Namely, write() doing copy_from_user()
when the file we are writing into has i_mutex held by us.  That can fault
and in this case we have a non-directory inode.  While you can't have
directory mmapped, regular files can be mmapped just fine.