Re: [PATCH] tmscsim: 64-bit cleanup

[Kurt Garloff <garloff@suse.de>]

[-- Attachment #1: Type: text/plain, Size: 1243 bytes --]

On Sat, May 22, 2004 at 10:43:12AM -0500, James Bottomley wrote:
> On Sat, 2004-05-22 at 03:09, Kurt Garloff wrote:
> > The AM53C974 hardware does not have hardware support for scatter-gather.
> > (Yes, this is sad, you get an IRQ after every segment ...)
> 
> But then you should set sg_tablesize to one.  There's little point the
> elevator wasting time to coalesce if you can't process the request in a
> single go.  Doing that would also allow for a simpler driver design,
> because now you know you can complete individual requests since they
> only yield a single sg element.  You no longer have to track multiple
> card completions to the request completion.

It's still cheaper to send down one command for a large transfer, let
the card interrupt the CPU to get a new DMA address every segment.
(When we're lucky, we might even have segments with contiguous bus
 addresses -- this is certainly the case if there's a IOMMU in the
 system) 

The per-command overhead in SCSI is significant, so I would hate to
restrict the driver to process 4k commands.

Regards,
-- 
Kurt Garloff  <garloff@suse.de>                            Cologne, DE 
SUSE LINUX AG / Novell, Nuernberg, DE               Director SUSE Labs

[-- Attachment #2: Type: application/pgp-signature, Size: 189 bytes --]