- When invalidating pages, take care to shoot down any ptes which map them
  as well.

  This ensures that the next mmap access to the page will generate a major
  fault, so NFS's server-side modifications are picked up.

  This also allows us to call invalidate_complete_page() on all pages, so
  filesytems such as ext3 get a chance to invalidate the buffer_heads.

- Don't mark in-pagetable pages as non-uptodate any more.  That broke a
  previous guarantee that mapped-into-user-process pages are always uptodate.

- Check the return value of invalidate_complete_page().  It can fail if
  someone redirties a page after generic_file_direct_IO() write it back.

But we still have a problem.  If invalidate_inode_pages2() calls
unmap_mapping_range(), that can cause zap_pte_range() to dirty the pagecache
pages.  That will redirty the page's buffers and will cause
invalidate_complete_page() to fail.

So, in generic_file_direct_IO() we do a complete pte shootdown on the file
up-front, prior to writing back dirty pagecache.  This is only done for
O_DIRECT writes.  It _could_ be done for O_DIRECT reads too, providing full
mmap-vs-direct-IO coherency for both O_DIRECT reads and O_DIRECT writes, but
permitting the pte shootdown on O_DIRECT reads trivially allows people to nuke
other people's mapped pagecache.

NFS also uses invalidate_inode_pages2() for handling server-side modification
notifications.  But in the NFS case the clear_page_dirty() in
invalidate_inode_pages2() is sufficient, because NFS doesn't have to worry
about the "dirty buffers against a clean page" problem. (I think)
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>