From: Josef Bacik <josef@redhat.com>
Date: Thu, 2 Jul 2009 09:09:41 -0400
Subject: [mm] fix re-read performance regression
Message-id: 20090702130941.GG16713@dhcp231-156.rdu.redhat.com
O-Subject: Re: [PATCH RHEL5.4] fix re-read performance regression UPDATED
Bugzilla: 506511
RH-Acked-by: Eric Sandeen <sandeen@redhat.com>

On Tue, Jun 30, 2009 at 02:33:48PM -0400, Josef Bacik wrote:
> Hello,
>
> This is in reference to bz 506511.  The performance team noticed that there was
> a 10% regression in re-read on one of their iozone tests.  This is because of
> the new aops patches, specifically the one that converted ext3 over to use the
> new aops.  When doing the write_end, we are supposed to unlock_page() and
> page_cache_release() the page.  We also should be doing mark_page_accessed() so
> the page does not get flushed to disk too hastily.  Ext3 was not doing this, and
> still isn't so there is no upstream reference for this yet (I will be sending a
> patch upstream shortly).  This patch makes sure the pages are marked as accessed
> when we exit write_end.  This has been tested with the iozone test and the
> regression is no longer seen.  Thank you,
>

Hello,

After talking with Nick Piggin we've come up with a better way to fix the
problem, and this will handle everybody who uses the new AOPs.  Please take this
patch instead.  Thank you,

Josef

diff --git a/mm/filemap.c b/mm/filemap.c
index 44661d7..6204c0f 100644
--- a/mm/filemap.c
+++ b/mm/filemap.c
@@ -2566,6 +2566,7 @@ again:
 		pagefault_enable();
 		flush_dcache_page(page);
 
+		mark_page_accessed(page);
 		status = a_ops->write_end(file, mapping, pos, bytes, copied,
 						page, fsdata);
 		if (unlikely(status < 0))