Re: 2.5.32 IO performance issues

[Patrick Mansfield <patmans@us.ibm.com>]

On Thu, Aug 29, 2002 at 01:55:40PM -0700, Badari Pulavarty wrote:
> > 
> > block-highmem is bust for scsi. (aic7xxx at least).  Does
> > http://www.zip.com.au/~akpm/linux/patches/2.5/2.5.32/2.5.32-mm2/broken-out/scsi_hack.patch
> > fix it?
> 
> Hmm !! This patch fixed it. I remember you gave me this patch for 2.5.31. But 2.5.31 
> was doing fine without it. But 2.5.32 seem to need it.
> 

The above patch works fine to get back to the previous (pre-2.5.32) state.
But, it makes no sense to modify the bounce_limit based on the type of
storage that is attached to an adapter.

We want to allow high mem for block devices other than SCSI direct access
devices (TYPE_DISK), such as CD ROM (SDpnt->type TYPE_ROM), WORM devices
(TYPE_WORM), and optical disks (TYPE_MOD).

So it is better to patch scsi_initialize_merge_fn:

--- 1.16/drivers/scsi/scsi_merge.c	Fri Jul  5 09:43:00 2002
+++ edited/drivers/scsi/scsi_merge.c	Thu Aug 29 14:30:12 2002
@@ -140,7 +140,7 @@
 	 * Enable highmem I/O, if appropriate.
 	 */
 	bounce_limit = BLK_BOUNCE_HIGH;
-	if (SHpnt->highmem_io && (SDpnt->type == TYPE_DISK)) {
+	if (SHpnt->highmem_io) {
 		if (!PCI_DMA_BUS_IS_PHYS)
 			/* Platforms with virtual-DMA translation
  			 * hardware have no practical limit.

-- Patrick Mansfield