Re: [RFC PATCH 0/2] Fix up the NFS mmap code

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

On Fri, Jan 08, 2010 at 05:57:27PM -0800, Linus Torvalds wrote:
> 
> 
> On Sat, 9 Jan 2010, Al Viro wrote:
> > 
> > readdir() is certainly a red herring.
> 
> That's the one that lockdep reports, though. I still don't see why. Afaik, 
> the only place where NFS gets an inode is nfs_fhget(), and that seems to 
> do things correctly.

Well, sure - it steps on i_mutex-before-mmmap_sem first from ls somewhere and
records the ordering for posterity.  Then NFS steps into mmap() (on a
different inode) and gets conflicting ordering.
 
It would be a false positive if rules for NFS *really* had been different
and it could safely grab i_mutex on NFS inodes inside mmap_sem.  It can't.
The rules really are the same.  And readdir is just the earliest case of
kernel stepping on mmap_sem while holding *some* i_mutex.  write() is
another and there i_mutex can very well be the same as in case of mmap().

lockdep doesn't make a distinction (and really, how many paths reinforcing
the normal lock ordering would you record?), but if we'd given i_mutex of
NFS regular files a class of its own, we'd see a warning with nfs write
instead of readdir...