Re: [PATCH 1/2][RESEND] improve generic_file_buffered_write()

[Christoph Hellwig <hch@infradead.org>]

On Fri, Sep 07, 2007 at 08:52:38PM +0200, Bernd Schubert wrote:
> No further response to our patches yet, so we are sending them again, 
> re-diffed against 2.6.23-rc5
> 
> Hi,
> 
> recently we discovered writing to a nfs-exported lustre filesystem is rather 
> slow (20-40 MB/s writing, but over 200 MB/s reading).
> 
> As I already explained on the nfs mailing list, this happens since there is an 
> offset on the very first page due to the nfs header.
> 
> http://sourceforge.net/mailarchive/forum.php?thread_name=200708312003.30446.bernd-schubert%40gmx.de&forum_name=nfs
> 
> While this especially effects lustre, Olaf Kirch also noticed it on another 
> filesystem before and wrote a nfs patch for it. This patch has two 
> disadvantages  - it requires to move all data within the pages, IMHO rather 
> cpu time consuming, furthermore, it presently causes data corruption when 
> more than one nfs thread is running.
> 
> After thinking it over and over again we (Goswin and I) believe it would be 
> best to improve generic_file_buffered_write().
> If there is sufficient data now, as it is usual for aio writes, 
> generic_file_buffered_write() will now fill each page as much as possible and 
> only then prepare/commit it. Before generic_file_buffered_write() commited 
> chunks of pages even though there were still more data.

While the idea is sound in general the code your touching is almost entirely
gone in -mm and hopefully in 2.6.24.  Can you take a look at the Nick's changes
in -mm that introduce  begin_write and end_write methods replacing prepare_write
and commit_write and see if they improve your situation already.  If not they
should at least provide a framework to deal with it in a slightly cleaner way.