Skip to contentSkip to conversation

Serverworks LE and MTRR write-combining question

DISCUSSION Mike Russo 2 participants 3 messages Apr 14, 2005 – Apr 15, 2005

lore
Share & tools
lore

Mike Russo

author
#0 lore raw

Hello,

I'd like to begin by humbly thanking everyone who works on the kernel for their time and patience and energy. Without you my company would have had to spend a lot more on software and gotten nowhere near as much value in return!

I have a question regarding my motherboard's serverworks LE chipset and MTRR write-combining. /usr/src/linux/arch/i386/kernel/cpu/mtrr/main.c will not allow write-combining because of the potential for data corruption, but according to the following old LKML post, this is only true for certain old revisions of the motherboard:

http://www.ussg.iu.edu/hypermail/linux/kernel/0104.3/1007.html

This person even submitted a patch which looked like it was going to be accepted, but nothing happened after that. I checked the output of lspci and my revision (06) was the first one where the fix was included, according to this person's post. I therefore disabled the check and recompiled my kernel (2.6.12-rc2-mm3, which is noticeably faster than the default fedora core 3 kernel, with nearly the same configuration). X was able to setup a write-combining range (I had to disable one that was setup already but it didn't seem to affect anything else) and I've been working over four hours (wow!) without any corruption or lockups. Just wondering if anyone had any updates on this issue, and if not, hey, that's why the source is there -- for me to screw around with. ;)

Signature · 9 lines
Mike Russo
ReadQ Systems, Inc.
(212) 425 3680 x105

Random quote of the day:
We are sorry.  We cannot complete your call as dialed.  Please check
the number and dial again or ask your operator for assistance.

This is a recording.
#1 lore raw

On Thu, 2005-04-14 at 19:00 -0400, Mike Russo wrote:Just

Quoted 3 lines

wondering if anyone had any updates on this issue, and if not, hey, that's why the source is there -- for me to screw around with. ;)

I think it didn't go in just because no one bothered to repost the patch with the comment fixed, as requested in that old thread.

Here's the patch from that thread against 2.6.12-rc2. It also fixes an unrelated typo in nearby code. Obviously untested, as I don't have the hardware ;-)

Summary: Enable write combining for server works LE rev > 6 per
http://www.ussg.iu.edu/hypermail/linux/kernel/0104.3/1007.html

arch/i386/kernel/cpu/mtrr/main.c renamed from arch/i386/kernel/cpu/mtrr/main.c~ +9 -5
@@ -72,17 +72,21 @@
 static int have_wrcomb(void)
 {
 	struct pci_dev *dev;
+	u8 rev;
 	
 	if ((dev = pci_get_class(PCI_CLASS_BRIDGE_HOST << 8, NULL)) != NULL) {
-		/* ServerWorks LE chipsets have problems with write-combining 
+		/* ServerWorks LE chipsets < rev 6 have problems with write-combining 
 		   Don't allow it and leave room for other chipsets to be tagged */
 		if (dev->vendor == PCI_VENDOR_ID_SERVERWORKS &&
 		    dev->device == PCI_DEVICE_ID_SERVERWORKS_LE) {
-			printk(KERN_INFO "mtrr: Serverworks LE detected. Write-combining disabled.\n");
-			pci_dev_put(dev);
-			return 0;
+			pci_read_config_byte(dev, PCI_CLASS_REVISION, &rev);
+			if (rev <= 5) {
+				printk(KERN_INFO "mtrr: Serverworks LE rev < 6 detected. Write-combining disabled.\n");
+				pci_dev_put(dev);
+				return 0;
+			}
 		}
-		/* Intel 450NX errata # 23. Non ascending cachline evictions to
+		/* Intel 450NX errata # 23. Non ascending cacheline evictions to
 		   write combining memory may resulting in data corruption */
 		if (dev->vendor == PCI_VENDOR_ID_INTEL &&
 		    dev->device == PCI_DEVICE_ID_INTEL_82451NX) {
#2 lore raw
Quoted 6 lines · Lee Revell

Here's the patch from that thread against 2.6.12-rc2. It also fixes an unrelated typo in nearby code. Obviously untested, as I don't have the hardware ;-)

Summary: Enable write combining for server works LE rev > 6 per
http://www.ussg.iu.edu/hypermail/linux/kernel/0104.3/1007.html

OK, Mike has tested my patch with a rev 6 chipset and report that write combining works fine. Andrew, please apply.

Lee