From mboxrd@z Thu Jan 1 00:00:00 1970 From: Andrew Morton Subject: Re: Possible explanation for SCSI benchmark problems in 2.5 Date: Sun, 29 Sep 2002 13:53:56 -0700 Sender: linux-scsi-owner@vger.kernel.org Message-ID: <3D976864.65C82D7B@digeo.com> References: <200209292048.g8TKmN424799@localhost.localdomain> Mime-Version: 1.0 Content-Type: text/plain; charset=us-ascii Content-Transfer-Encoding: 7bit Return-path: Received: from digeo-nav01.digeo.com (digeo-nav01.digeo.com [192.168.1.233]) by packet.digeo.com (8.9.3+Sun/8.9.3) with SMTP id NAA13207 for ; Sun, 29 Sep 2002 13:53:59 -0700 (PDT) List-Id: linux-scsi@vger.kernel.org To: James Bottomley Cc: linux-scsi@vger.kernel.org, Jens Axboe James Bottomley wrote: > > Hi All, > > While looking at other code (OK, how to get the MCA drivers not to use bounce > buffers...) I discovered that the way scsi_scan.c calls > scsi_initialize_merge_fn() guarantees (on an x86) that the > blk_queue_bounce_limit is always called with BLK_BOUNCE_HIGH. This will > probably hurt benchmarks on machines with 1Gb or more of memory. > > The problem is that scsi_initialize_merge_fn() checks sdev->type, but this is > always -1 in scsi_alloc_sdev, and thus it never uses the pci dma mask for disk > devices. > > The attached patch moves the initialisation to after sdev->type has the > correct value. > argh, sorry. I've had such a patch in my tree for ages, and I do all testing with it in place. But high-IO is bust for IDE as well at present, so at least we're being fair ;)