nvidia chipsets + Core 2 Duo = uniprocessor [solved]

jesnow · Guru Joined: 26 Apr 2006 Posts: 394

Some nvidia NFORCE chipsets appear to be pretty thoroughly broken for use with
SMP under linux. Probably under windows too, but of course the problem is
better hidden there. I have an ASUS P5NSLI board that shows this problem,
but it appears that other boards show the same problem.

Sumptom 1: During boot you get messages like:

ntrl · n00b Joined: 26 Sep 2004 Posts: 34 Location: SPb, Russia

Hi.

try pci=bios
_________________
--
WBR

jesnow · Guru Joined: 26 Apr 2006 Posts: 394

ntrl · n00b Joined: 26 Sep 2004 Posts: 34 Location: SPb, Russia

What MB? Asus, gigabyte? May be flash new bios?
_________________
--
WBR

jesnow · Guru Joined: 26 Apr 2006 Posts: 394

jesnow · Guru Joined: 26 Apr 2006 Posts: 394

I have submitted a bug report on kernel bugzilla, will report back.

Jon.

jesnow · Guru Joined: 26 Apr 2006 Posts: 394

Someone in Kernel bugzilla had this to say:

ZeroDivide · Tux's lil' helper Joined: 15 Mar 2003 Posts: 83

I was having the same problems with an Epox 9NPA3-SLI motherboard.

Eventually I found that disabling ACPI caused APIC to work properly. So I just have to pass "acpi=no" as a kernel parameter and smp appears to work fine.

jesnow · Guru Joined: 26 Apr 2006 Posts: 394

Rikai · Posted: Tue Apr 03, 2007 7:40 am Post subject:

If there's a bad ACPI table entry, you might try fixing it youself with a custom DSDT?

At the very least, I don't think it can hurt.

jesnow · Guru Joined: 26 Apr 2006 Posts: 394

transsib · Guru Joined: 26 Jul 2003 Posts: 536

I haven't looked so hard at the boot messages on booting. My mobo is a asus p5n32-e sli with nforce680 chipset and a pentium 4 .
I don't know whether it is related to that problem but I had to give the kernel pci=nomsi on boot to make net.eth0 actually work. I've tried several kernels and for each of them I had to write that into grub.

Does anyone know what pci=nomsi does and how it affects other pci traffic on the board?

I have a beta bios and intend to flash it to the same bios rev. because it just got out of beta status. The system is all in all pretty fast and has no stability problems but I haven't pushed it to its limits yet. It is only 4 weeks old.

My /proc/interrupts looks like this

jesnow · Guru Joined: 26 Apr 2006 Posts: 394

jesnow · Guru Joined: 26 Apr 2006 Posts: 394

I'm a little upset that this thread is petering out and that there is no action on the kernel.org bug I submitted. I think this may go back to the general linux mentality of "it boots, it works". There have to be a lot of people out there with broken SMP who don't know it, and they wonder why their machine is less responsive under linux than under that other OS.

Is it really OK fo a large fraction of machines out there to not process interrupts on both CPU's? What happens when the primary advance in computing power is by adding cores? That's happening now.

At the very least can people with Core 2 processors run cat /proc/interrupts and post here if their machine works (say what your mobo and chipset are).

Cheers,

Jon.

arkhan_jg · Posted: Mon Apr 09, 2007 8:25 am Post subject:

It may also be that it's not strictly a fault in the kernel; if the ACPI table is broken on the motherboard, i.e. ASUS only tested that their BIOS worked with the incredibly out-of-spec tolerating windows, and the nforce chipset drivers aren't happy with a badly implemented ACPI table under linux...

I presume you've tried disabling ACPI in the BIOS directly rather than just passing ACPI=no in the kernel arguments? APM will do the job for a desktop, ACPI causes more problems than it solves IMO for desktops, the implementations are usually so badly written, which is a shame is it does solve some IRQ issues.

I don't have linux installed on any of my core 2 at home, but I have ubuntu on my nforce 4 sli x16 (asus p5n32-sli se deluxe) core 2 at work; I'll have a look at that one when I come back off holiday.
Also, I thought the MCP55 chipset was an athlon AM2 one? The P5NSLI normally has an nforce 570.
_________________
make menuconfig not war

xanas3712 · Guru Joined: 15 Oct 2004 Posts: 455

I'll try to check on this when I get home, but I'm not running nforce and both cores seem to work fine, but I'm curious of what the output is supposed to look like by comparison since I've not looked into this before.

jesnow · Guru Joined: 26 Apr 2006 Posts: 394

aceFruchtsaft · Posted: Tue Apr 10, 2007 8:05 am Post subject:

My output is also similar and I have an Asus P5B Deluxe with an Intel P965 Chipset:

drescherjm · Posted: Tue Apr 10, 2007 8:07 am Post subject:

Here is my result for a dual processor Opteron on a 2.6.19 vserver sources kernel:

DirtyHairy · Posted: Tue Apr 10, 2007 9:49 am Post subject:

Sorry folks, I might be wrong on that one, but imho balancing interrupts is irrelevant where speed is concerned (at least for an usual desktop configuration). The speed improvement with multicore systems comes from the ability of balancing multiple threads between several physical CPUs, nothing else. If this works (and you can check this as easily as firing up several time-consuming processes and then looking at top output), your SMP is working fine. For example,

Cyker · Veteran Joined: 15 Jun 2006 Posts: 1427

Just for kicks, here's my S3 VirgeDX-powered headless server

xanas3712 · Guru Joined: 15 Oct 2004 Posts: 455

I'm not sure honestly. I do know the system runs better even before the interrupt change than the single cores here. And I've not noticed an issue until it was brought up about it.

Anyhow, I didn't say my chipset earlier and forgot what motherboard I was using for some strange reason. It's an intel 965P. Gigabyte-965P-S3 if I recall correctly.

DirtyHairy · Posted: Tue Apr 10, 2007 11:49 am Post subject:

Running irqbalance changed things for me a bit

Cyker · Veteran Joined: 15 Jun 2006 Posts: 1427

Yeah... Linux already has a fantastic scheduler, and most Linux users don't really need two cores anyway so I think for the most part the performance difference will be negligible.
If you're doing heavy processing (Like this Folding@Home thing which seems to be popular

), then the interrupts used will not be as bad since it actually does only very little I/O.

I suspect on my system it might have more of an effect because I *am* doing ridiculous amounts of I/O; Between the RAID array and the two gigabit cards espescially, and on systems that heavily use both cores for real-time stuff (I'm thinking FPS games and such) then it could also suck pretty bad.

xanas3712 · Guru Joined: 15 Oct 2004 Posts: 455

Maybe newer games, I'll only be able to test it maybe when unreal 2007 comes out. With an 8800 GTS quake 4 runs flawlessly, but there aren't too many other linux games with high end graphics (and that one is aging quickly) to really know for sure.