From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: Received: (majordomo@vger.kernel.org) by vger.kernel.org via listexpand id S262222AbTEZUcc (ORCPT ); Mon, 26 May 2003 16:32:32 -0400 Received: (majordomo@vger.kernel.org) by vger.kernel.org id S262226AbTEZUcc (ORCPT ); Mon, 26 May 2003 16:32:32 -0400 Received: from neon-gw-l3.transmeta.com ([63.209.4.196]:11021 "EHLO neon-gw.transmeta.com") by vger.kernel.org with ESMTP id S262222AbTEZUca (ORCPT ); Mon, 26 May 2003 16:32:30 -0400 Date: Mon, 26 May 2003 13:45:25 -0700 (PDT) From: Linus Torvalds To: James Bottomley cc: Jens Axboe , Linux Kernel Subject: Re: [BK PATCHES] add ata scsi driver In-Reply-To: <1053981380.1768.203.camel@mulgrave> Message-ID: MIME-Version: 1.0 Content-Type: TEXT/PLAIN; charset=US-ASCII Sender: linux-kernel-owner@vger.kernel.org X-Mailing-List: linux-kernel@vger.kernel.org On 26 May 2003, James Bottomley wrote: > > > I'd hate for SATA to pick up these kinds of mistakes from the SCSI layer. > > The elevator is based on linear head movements. Historically, yes. But we've been moving more and more towards a latency-based elevator, or at least one that takes latency into account. Which is exactly why you'd like to leave unfinished requests on the queue, because otherwise your queue doesn't really show what is really going on. In particular, while the higher layers don't actually _do_ this yet, from a latency standpoint the right thing to do when some request seems to get starved is to refuse to feed more tagged requeusts to the device until the starved request has finished. As I mentioned, Andrew actually had some really bad latency numbers to prove that this is a real issue. SCSI with more than 4 tags or so results in potentially _major_ starvation, because the disks themselves are not even trying to avoid it. Also, even aside from the starvation issue with unfair disks, just from a "linear head movement" standpoint, you really want to sort the queue according to what is going on _now_ in the disk. If the disk eats the requests immediately (but doesn't execute them immediately), the sorting has nothing to go on, and you get tons of back-and-forth movements. Basically, if you're trying to do an elevator, you need to know what the outstanding commands are. Think it through on paper, and you'll see what I mean. Linus