Nick Piggin writes:
 > 
 > 
 > Nikita Danilov wrote:
 > 
 > >Hugh Dickins writes:
 > > > On Wed, 4 Feb 2004, Nikita Danilov wrote:
 > > > > Hugh Dickins writes:
 > > > >  > If you go the writepage-while-mapped route (more general gotchas?
 > > > >  > I forget), you'll have to make an exception for shmem_writepage.
 > > > > 
 > > > > May be one can just call try_to_unmap() from shmem_writepage()?
 > > > 
 > > > That sounds much cleaner.  But I've not yet found what tree your
 > > > p12-dont-unmap-on-pageout.patch applies to, so cannot judge it.
 > >
 > >Whole
 > >ftp://ftp.namesys.com/pub/misc-patches/unsupported/extra/2004.02.04/
 > >applies to the 2.6.2-rc2.
 > >
 > >I just updated p12-dont-unmap-on-pageout.patch in-place.
 > >
 > >  
 > >
 > 
 > Sure, I can give this a try. It makes sense.
 > 

To my surprise I have just found that

ftp://ftp.namesys.com/pub/misc-patches/unsupported/extra/2004.02.04/p10-trasnfer-dirty-on-refill.patch

[yes, I know there is a typo in the name.]

patch improves performance quite measurably. It implements a suggestion
made in the comment in refill_inactive_zone():

 			/*
			 * probably it would be useful to transfer dirty bit
			 * from pte to the @page here.
 			 */

To do this page_is_dirty() function is used (the same one as used by
dont-unmap-on-pageout.patch), which is implemented in
check-pte-dirty.patch.

I ran

$ time build.sh 10 11

(attached) and get following elapsed time:

without patch: 3818.320, with patch: 3368.690 (11% improvement).

As I see it, early transfer of dirtiness to the struct page allows to do
more write-back through ->writepages() which is much more efficient way
than single-page ->writepage.

Nikita.