--On Thursday, June 12, 2003 14:44:18 -0700 Andrew Morton <akpm@digeo.com>
wrote:

> Well it is not "worse".  Futzing with i_sem in do_no_page() is pretty
> gross. You could add vm_ops->prevalidate() or something if it worries you.

Actually I've been studying the logic.  I don't think we need to serialize
with i_sem at all.  i_size has already been changed, so just doing a retry
immediately will be safe.  Distributed filesystems should also be safe as
long as they mark the page invalid before they call invalidate_mmap_range().

I like your idea of doing an atomic_t instead of a seqlock.  The original
idea from Andrea implied there was a range of time it was unstable, but
with this scheme a single increment is sufficient.

I also think if we can solve both the vmtruncate and the distributed file
system races without adding any vm_ops, we should.

Here's a new patch.  Does this look better?

Dave

======================================================================
Dave McCracken          IBM Linux Base Kernel Team      1-512-838-3059
dmccr@us.ibm.com                                        T/L   678-3059