Re: [RFC PATCH 2/2] NFS: Fix a potential deadlock in nfs_file_mmap()

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

On Fri, Jan 08, 2010 at 07:56:24PM -0500, Trond Myklebust wrote:
> We cannot call nfs_invalidate_mapping() inside file->f_ops->mmap(), since
> this would cause us to grab the inode->i_mutex while already holding the
> current->mm->mmap_sem (thus causing a potential ABBA deadlock with the file
> write code, which can grab those locks in the opposite order).
> 
> We can fix this situation for the mmap() system call by using the new
> mmap_pgoff() callback, which is called prior to taking the
> current->mm->mmap_sem mutex.
> 
> We also add ensure that open() invalidates the mapping if the inode data is
> stale so that other users of mmap() (mainly the exec and uselib system
> calls) get up to date data too.
 
> +	status = nfs_revalidate_mapping(inode, file->f_mapping);
> +	if (status < 0)
> +		return status;
> +
> +	return generic_file_mmap_pgoff(file, addr, len, prot, flags, pgoff);

This is completely bogus.  Why do you need i_mutex for that and what
the <expletives> does that really prevent?  You might wait for a _loong_
time waiting for that mmap_sem, so what is really going on there?