Re: xfs cluster rewrites is broken?

[David Chinner]

Hi Alexander,

On Fri, Mar 17, 2006 at 03:32:03PM +0300, Alexander Y. Fomichev wrote:
> Hello,
> 
> Two days ago i've try 2.6.16-rc5 on 2-way dual-core Opteron server
> and faced with a strange system behaviour. 
.....
>     [XFS] cluster rewrites      We can cluster mapped pages aswell, this 
> improves
>     performances on rewrites since we can reduce the number of allocator
>     calls.

FYI, prior to this mod, XFS never clustered rewrites, so it's not surprising
that this particular issue has only recently come to light.

> diff -urN a/fs/xfs/linux-2.6/xfs_aops.c b/fs/xfs/linux-2.6/xfs_aops.c
> --- a/fs/xfs/linux-2.6/xfs_aops.c	2006-03-17 13:13:53.000000000 +0300
> +++ b/fs/xfs/linux-2.6/xfs_aops.c	2006-03-17 15:12:12.000000000 +0300
> @@ -616,8 +616,6 @@
>  				acceptable = (type == IOMAP_UNWRITTEN);
>  			else if (buffer_delay(bh))
>  				acceptable = (type == IOMAP_DELAY);
> -			else if (buffer_mapped(bh))
> -				acceptable = (type == 0);
>  			else
>  				break;
>  		} while ((bh = bh->b_this_page) != head);

Well, that switches off rewrite clustering altogether, so it's
not surprising that it fixed your problem. It also points out the
problem as well - we don't every check if the buffer is dirty before
declaring that the page is acceptible for write clustering.

The other cases checked here (buffer_unwritten() and buffer_delay())
are, by defintition, dirty buffers and so they only ever cluster
dirty pages. buffer_mapped(), OTOH, could be clean or dirty.....

Can you try the patch below, and see if that fixes the problem
you are seeing?

Cheers,

Dave.
-- 
Dave Chinner
R&D Software Enginner
SGI Australian Software Group


Index: 2.6.x-xfs-new/fs/xfs/linux-2.6/xfs_aops.c
===================================================================
--- 2.6.x-xfs-new.orig/fs/xfs/linux-2.6/xfs_aops.c	2006-03-17 13:16:13.000000000 +1100
+++ 2.6.x-xfs-new/fs/xfs/linux-2.6/xfs_aops.c	2006-03-20 10:51:36.906723758 +1100
@@ -647,7 +647,7 @@ xfs_is_delayed_page(
 				acceptable = (type == IOMAP_UNWRITTEN);
 			else if (buffer_delay(bh))
 				acceptable = (type == IOMAP_DELAY);
-			else if (buffer_mapped(bh))
+			else if (buffer_mapped(bh) && buffer_dirty(bh))
 				acceptable = (type == 0);
 			else
 				break;
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/