Re: [PATCH] block: remove plugging at buffered write time

[Jan Kara <jack@suse.cz>]

On Fri 10-02-12 10:47:16, Wu Fengguang wrote:
> On Fri, Feb 10, 2012 at 09:52:18AM +0800, Wu Fengguang wrote:
> > On Thu, Feb 09, 2012 at 01:30:27PM -0500, Chris Mason wrote:
> > > On Thu, Feb 09, 2012 at 01:06:35PM -0500, Christoph Hellwig wrote:
> > > > On Thu, Feb 09, 2012 at 04:02:24PM +0800, Wu Fengguang wrote:
> > > > > On Thu, Feb 09, 2012 at 10:27:19AM +1100, Dave Chinner wrote:
> > > > > > On Wed, Feb 08, 2012 at 07:01:44PM +0800, Wu Fengguang wrote:
> > > > > > > Buffered write(2) is not directly tied to IO, so it's not suitable to
> > > > > > > handle plug in generic_file_aio_write().
> > > > > > 
> > > > > > But generic_sync_write() does issue IO for O_SYNC writes, so unless
> > > > > > there is plugging at a lower layer in the writeback code then it
> > > > > > appears to me that plugging is still necessary (at least inside the
> > > > > > sync branch)....
> > > > > 
> > > > > Good catch! It looks that generic_write_sync() eventually calls into
> > > > > vfs_fsync_range() which further calls ->fsync(). We may add plugging
> > > > > around it:
> > > > 
> > > > 
> > > > NAK, please keep the plugging down in the fs, or the libraries used but
> > > > not common VFS code.
> > > 
> > > Please, what Christoph said.  At least for btrfs plugging here is wrong.
> > 
> > OK, I get the point: the fs knows best when to unplug. Since any
> > higher level plug nesting will turn such low level efforts into no-op,
> > it's highly undesirable to do it in the high level.
> 
> It's actually wrong to do plugging around vfs_fsync_range().
> 
> Because these call paths
> 
>         write() with O_SYNC
>           generic_write_sync()
>             vfs_fsync_range()
>               ->fsync()
>               generic_file_fsync()
> 
>         fsync()
>           do_fsync()
>             vfs_fsync()
>               vfs_fsync_range()
> 
> pass arbitrary @size arguments, which may be much larger than the
> preferable I/O size, or may cross extent/device boundaries.
> 
> generic_file_fsync() starts with a filemap_write_and_wait_range()
> call, which already has proper plugging somewhere underneath. Then
> followed by metadata writes, which has plugging inside
> fsync_buffers_list(). At last, sync_inode_metadata() calls into
> ->write_inode() which may or may not care plugging.
> 
> The other fs specific ->fsync() do similar steps, varying in the
> metadata and fs specific housekeeping part.
> 
> I'll just drop this code. Shall the fs specific metadata I/O be
> plugged accordingly? I'm afraid this is beyond my knowledge base...
  The filesystems I know (ext?, ocfs2, reiserfs, udf) either don't do any
metadata io from ->fsync (it happens from a journalling thread) or the io
is random so plugging is not desirable anyway AFAIU (well,
mpage_writepages() is clever enough to submit metadata which is interleaved
with data in one sequential stream together with the data so metadata that
remain are mostly random).

								Honza
-- 
Jan Kara <jack@suse.cz>
SUSE Labs, CR