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

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

On Sat, Jan 09, 2010 at 01:38:25AM +0000, Al Viro wrote:
> On Fri, Jan 08, 2010 at 05:17:14PM -0800, Linus Torvalds wrote:
> > 
> > 
> > On Fri, 8 Jan 2010, Trond Myklebust wrote:
> > >
> > > How about something like the following. I chose to wrap the call to
> > > do_mmap_pgoff() instead of making a special ->pre_mmap(), since that
> > > seems more consistent with the way we handle ->read() and ->write().
> > 
> > I still don't think that you can ever do mmap _and_ readdir on the same 
> > inode, so there's something wrong with the lockdep annotations.
> 
> readdir() is certainly a red herring.  write(), OTOH, is quite real.
> And there we do i_mutex followed by pagefaults.
> 
> I *REALLY* dislike Trond's solution, though.
> 
> Could we please get a sane expalanation of the reasons why nfs mmap
> wants i_mutex in the first place?  Before we add yet another hook
> from hell and complicate already overcomplicated area...

PS: mmap/write deadlock is real, AFAICT - unless something very subtle
prevents it, we can get buggered if we have two threads with the same
VM, mmap 1.4Mb from floppy and do
thread A:
	write(fd_on_nfs, buffer_mmaped_from_floppy, 1440 * 1024);
thread B:
	mmap(..., fd_on_nfs, ...)

It's not even particulary narrow.