Return-Path: Received: (majordomo@vger.kernel.org) by vger.kernel.org via listexpand id ; Thu, 19 Sep 2002 15:34:11 -0400 Received: (majordomo@vger.kernel.org) by vger.kernel.org id ; Thu, 19 Sep 2002 15:34:10 -0400 Received: from as8-6-1.ens.s.bonet.se ([217.215.92.25]:33286 "EHLO zoo.weinigel.se") by vger.kernel.org with ESMTP id ; Thu, 19 Sep 2002 15:34:05 -0400 To: Robert Schwebel , Alan Cox Cc: linux-kernel@vger.kernel.org Subject: MediaGX/Geode performance fix, Was: Which processor/board for embedded NTP References: <1032354632.23252.14.camel@venus> <87r8frqech.fsf@zoo.weinigel.se> <20020919060218.GD10773@pengutronix.de> From: Christer Weinigel Organization: Weinigel Ingenjorsbyra AB Date: 19 Sep 2002 21:39:04 +0200 In-Reply-To: <20020919060218.GD10773@pengutronix.de> Message-ID: <873cs5vkfb.fsf_-_@zoo.weinigel.se> User-Agent: Gnus/5.09 (Gnus v5.9.0) Emacs/21.2 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 Content-Length: 3183 Lines: 87 This mail contains a patch to fix a performance problem with many Cyrix MediaGX/NatSemi Geode platforms. The register settings have been officially recommended by NatSemi themselves. The patch is against linux-2.4.20-pre7. Should this be merged into the mainsteam linux kernel? Robert Schwebel writes: > Hi Christer, > > On Wed, Sep 18, 2002 at 09:37:02PM +0200, Christer Weinigel wrote: > > On the newer "IA on a chip" geodes (SC1200, SC2200 and SC3200) there > > is also a high speed timer in the chipset that seems to be quite > > stable. > > Did the realtime behaviour of the newer Geodes also change to the > better? Last time we tested them with RTAI there have been Buslocks > which prevented them from being usefull... Speaking about this, there is a bug in many BIOS:es for the Cyrix MediaGX and Natsemi Geode that according to the BIOS errata will lower the performance of master PCI transfers from 70MB/s to 25MB/s. What happens is that the bit SDBE is set to 1 when it should be set to 0, causing PCI disconnects every 16 bytes which is slow. Index 41h PCI Control Function 2 Register (R/W) 3:2 SDB Slave Disconnect Boundary: PCI slave issues a disconnect with data when it crosses line boundary: 00 = 128 bytes 01 = 256 bytes 10 = 512 bytes 11 = 1024 bytes Works in conjunction with bit 1. 1 SDBE Slave Disconnect Boundary Enable: 0 = PCI slave disconnects on boundaries set by bits [3:2]. 1 = PCI disconnects on cache line boundary which is 16 bytes. So, this patch just sets the SDBE bit to 0 and prints a message to that effect. /Christer diff -ur linux/drivers/pci/quirks.c.orig linux/drivers/pci/quirks.c --- linux/drivers/pci/quirks.c.orig Thu Sep 19 21:34:21 2002 +++ linux/drivers/pci/quirks.c Thu Sep 19 21:35:07 2002 @@ -477,6 +477,24 @@ } /* + * Common misconfiguration of the MediaGX/Geode PCI master that will + * reduce PCI bandwidth from 70MB/s to 25MB/s. See the GXM/GXLV/GX1 + * datasheets found at http://www.national.com/ds/GX for info on what + * these bits do. + */ + +static void __init quirk_mediagx_master(struct pci_dev *dev) +{ + u8 reg; + pci_read_config_byte(dev, 0x41, ®); + if (reg & 2) { + reg &= ~2; + printk(KERN_INFO "PCI: Fixup for MediaGX/Geode Slave Disconnect Boundary (0x41=0x%02x)\n", reg); + pci_write_config_byte(dev, 0x41, reg); + } +} + +/* * The main table of quirks. */ @@ -538,6 +556,8 @@ */ { PCI_FIXUP_HEADER, PCI_VENDOR_ID_INTEL, PCI_DEVICE_ID_INTEL_82380FB, quirk_transparent_bridge }, + { PCI_FIXUP_FINAL, PCI_VENDOR_ID_CYRIX, PCI_DEVICE_ID_CYRIX_PCI_MASTER, quirk_mediagx_master }, + { 0 } }; -- "Just how much can I get away with and still go to heaven?" Freelance consultant specializing in device driver programming for Linux Christer Weinigel http://www.weinigel.se - To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html Please read the FAQ at http://www.tux.org/lkml/