Skip to contentSkip to conversation

[2.4] Nforce2 oops and occasional hang (tried the lockups patch, no difference)

DISCUSSION Ross Dickson 5 participants 8 messages Dec 13, 2003 – Dec 20, 2003

lore
Share & tools
lore
#0 lore raw

Oh, and the modules list:
Module Size Used by Tainted: P
i2c-dev 4548 0 (unused)
i2c-core 13604 0 [i2c-dev]
<snip>

I am not certain your problems are nforce2 type specific. Standard response: I don't suppose you can try a different stick of ram?

The reason I say that is that oops were very uncommon on either the epox 8rga+ or albatron km18G-pro MOBOS upon which I developed my patches. Hard lockups were pretty much all I experienced prior to the patches except for an occasional X fail. Base OS flavour I use is Suse 8.2 including gcc version (web updates utilised)

The udma patches are really just a cleanup on the address setup timing so I do not think that they are a factor.

The local apic ack delay timing patch needs athlon cpu and amd/nvidia ide on in kern config to kick in. If you are using it then I highly recommend uniprocessor ioapic config as well to go with it to route the 8254 timer irq0 through pin 0 of ioapic as using the apic config alone leaves a lot of ints generated on irq7 which can cause problems. (Reason for 8259 making them spurious on irq7 is explained in 8259A data sheet)

Also I now use a small patch to fixup proc info - only if you are using the 64 bit jiffies var hz patch, avail here:

http://linux.derkeiler.com/Mailing-Lists/Kernel/2003-12/0838.html

If you try acpi=off on boot and it is then not very stable then I think it has little to do with lockups patch as that is my fallback mode when I am playing with apic ioapic code.

Another fallback I use at times is

hdparm -Xudma3 /dev/hda

Hope this helps the confusion

Regards
Ross.

#1 lore raw
Quoted 8 lines · Ross Dickson

Oh, and the modules list: Module Size Used by Tainted: P i2c-dev 4548 0 (unused) i2c-core 13604 0 [i2c-dev] <snip>

I am not certain your problems are nforce2 type specific. Standard response: I don't suppose you can try a different stick of ram?

Yes, and stock settings with tested ram may be necessary with nforce2, possibly related to our timing-related voodoo culture(might overclock later on in life as timing-related patches evolve).

I have a via board that only recognizes two of four generic ram sticks, but second stick will cause an oops soon after. Another via setup will oops if any fast ram settings(4-way,csl2 etc) is attempted, though only using tested cas2 ram. "Try a different stick of ram".

On nforce2 I'm able to use bios "performance" ram timing but if I manually tweak all the ram settings up like I can do on other systems, I get mem-related OOPS with nforce2.

acpi apic lapic amd pre-empt nforce2ide (once you start you have to go all the way)

Quoted 26 lines

The reason I say that is that oops were very uncommon on either the epox 8rga+ or albatron km18G-pro MOBOS upon which I developed my patches. Hard lockups were pretty much all I experienced prior to the patches except for an occasional X fail. Base OS flavour I use is Suse 8.2 including gcc version (web updates utilised)

The udma patches are really just a cleanup on the address setup timing so I do not think that they are a factor.

The local apic ack delay timing patch needs athlon cpu and amd/nvidia ide on in kern config to kick in. If you are using it then I highly recommend uniprocessor ioapic config as well to go with it to route the 8254 timer irq0 through pin 0 of ioapic as using the apic config alone leaves a lot of ints generated on irq7 which can cause problems. (Reason for 8259 making them spurious on irq7 is explained in 8259A data sheet)

Also I now use a small patch to fixup proc info - only if you are using the 64 bit jiffies var hz patch, avail here:

http://linux.derkeiler.com/Mailing-Lists/Kernel/2003-12/0838.html

If you try acpi=off on boot and it is then not very stable then I think it has little to do with lockups patch as that is my fallback mode when I am playing with apic ioapic code.

Another fallback I use at times is

hdparm -Xudma3 /dev/hda

Hope this helps the confusion

Regards
Ross

#2 lore raw
In reply to Ross Dickson

Thanks greatly for the tips, btw. Its much appreciated :)

Also, I think I forgot to mention (doh) its an Epox 8rda+.

Quoted 2 lines · Ross Dickson

I am not certain your problems are nforce2 type specific. Standard response: I don't suppose you can try a different stick of ram?

Unfortunately not - its a single kingston hyper-x 3200 stick. (In theory I'll be moving it up to 700M or a gig in the next couple months, at which point I'll be in position to swap ram around and so forth.) I did get the bios all tuned and run memtest-mmx on it for 24 hours before the system installation though, and it passed. What I did yesterday is turn the memory frequency down from 200 to 166. (Which leaves the cpu overclocked by about 33 mhz, something I think it will survive just fine ;) ..)

Quoted 6 lines

The local apic ack delay timing patch needs athlon cpu and amd/nvidia ide on in kern config to kick in. If you are using it then I highly recommend uniprocessor ioapic config as well to go with it to route the 8254 timer irq0 through pin 0 of ioapic as using the apic config alone leaves a lot of ints generated on irq7 which can cause problems. (Reason for 8259 making them spurious on irq7 is explained in 8259A data sheet)

Thats how I'm running it now - its gone about 1 day without any oopses. (In the past it would go anywhere from hours to about a week, so the results aren't in yet.)

On the basis of it being a memory issue I poked around the epox site and noticed something I hadn't seen before - they recommend what might be different memory timings (I'll have to check if/when it crashes again):

If your PC3200 memory is not stable try the following BIOS settings:

Memory Frequency = 100%
Memory Timing = Expert
T(RAS) = 7
T(RCD) = 3
T(RP) = 3
CAS Latency = 2.5

Adjust the memory frequency above until you reach the resulting frequency of 200MHz (for PC3200).

Signature · 1 line
Disconnect <[email protected]>
#3 lore raw

Tues evening it hardlocked - no messages, no blinking keyboard lights, nothing. Yesterday it did it again, and this time it wouldn't get more than 5-10 minutes of uptime without hanging no matter what I did to memory/cpu timings. (Even underclocked it to 133 and 1G with no change.) So its unfortunately back on a stock 2.4.23-pre9 with noapic/noacpi. (It disables one of the sets of usb ports, as I recall, but it mostly works...)

In theory I'll have the ram upgrades and such tonight, so I can swap the services to another machine and stress it a bit with different sticks, new drives, different controller, etc. Other than memtest86 (and some kernel builds, bonnie, etc), are there any recommended tests?

Quoted 43 lines · Disconnect

Thanks greatly for the tips, btw. Its much appreciated :)

Also, I think I forgot to mention (doh) its an Epox 8rda+.

Quoted 2 lines · Ross Dickson

I am not certain your problems are nforce2 type specific. Standard response: I don't suppose you can try a different stick of ram?

Unfortunately not - its a single kingston hyper-x 3200 stick. (In theory I'll be moving it up to 700M or a gig in the next couple months, at which point I'll be in position to swap ram around and so forth.) I did get the bios all tuned and run memtest-mmx on it for 24 hours before the system installation though, and it passed. What I did yesterday is turn the memory frequency down from 200 to 166. (Which leaves the cpu overclocked by about 33 mhz, something I think it will survive just fine ;) ..)

Quoted 6 lines

The local apic ack delay timing patch needs athlon cpu and amd/nvidia ide on in kern config to kick in. If you are using it then I highly recommend uniprocessor ioapic config as well to go with it to route the 8254 timer irq0 through pin 0 of ioapic as using the apic config alone leaves a lot of ints generated on irq7 which can cause problems. (Reason for 8259 making them spurious on irq7 is explained in 8259A data sheet)

Thats how I'm running it now - its gone about 1 day without any oopses. (In the past it would go anywhere from hours to about a week, so the results aren't in yet.)

On the basis of it being a memory issue I poked around the epox site and noticed something I hadn't seen before - they recommend what might be different memory timings (I'll have to check if/when it crashes again):

If your PC3200 memory is not stable try the following BIOS settings:

Memory Frequency = 100%
Memory Timing = Expert
T(RAS) = 7
T(RCD) = 3
T(RP) = 3
CAS Latency = 2.5

Adjust the memory frequency above until you reach the resulting frequency of 200MHz (for PC3200).

Signature · 1 line
Disconnect <[email protected]>
#4 lore raw
Quoted 4 lines · Disconnect

memory/cpu timings. (Even underclocked it to 133 and 1G with no change.) So its unfortunately back on a stock 2.4.23-pre9 with noapic/noacpi. (It disables one of the sets of usb ports, as I recall, but it mostly works...)

Update: Underclocked from 1.8G to 1.2G (whups, meant to go down only 2-300mhz) and its been vaguely stable for about 1.5 days. I don't have another week (yet..) to run it under its normal load and wait for a crash, so what I'm going to do is:
- Move the workload (web/mail/..) to a different machine so this one
can be down for an extended period
- Replace the ram with new sticks (they arrived this morning)
- Reclock everything to stock (1.83G cpu, 200mhz ram and verify the
timings from kingston)
- Replace the video card
- Memtest86 until it cries
- If it passes, bonnie++ on the new drives
- If that passes, usb/acpi/apic testing with the associated patches

Anyone still watching this? Tips and suggestions on what else might be useful/informative are more than welcome. The tests above mostly replicate what I did when building this box, and it passed them then..

Recap:
Epox 8rda+ nforce2 mobo
AMD Athlon XP 2500+ (Barton) 1.83G
Kingston HyperX PC3200
WD Caviar WD1200JB 8M/UDMA100
Antec case w/ 350W AMD-certified PSU

Oopses and occasional hangs, usually in do_generic_file_read, using stock kernel.org 2.4.2x kernels. Hardware passed testing (memtest86, bonnie++) before I put Linux on it.

Signature · 1 line
Disconnect <[email protected]>
#5 lore raw
Quoted 31 lines · Disconnect
Quoted 4 lines · Disconnect

memory/cpu timings. (Even underclocked it to 133 and 1G with no change.) So its unfortunately back on a stock 2.4.23-pre9 with noapic/noacpi. (It disables one of the sets of usb ports, as I recall, but it mostly works...)

Update: Underclocked from 1.8G to 1.2G (whups, meant to go down only 2-300mhz) and its been vaguely stable for about 1.5 days. I don't have another week (yet..) to run it under its normal load and wait for a crash, so what I'm going to do is:
- Move the workload (web/mail/..) to a different machine so this one
can be down for an extended period
- Replace the ram with new sticks (they arrived this morning)
- Reclock everything to stock (1.83G cpu, 200mhz ram and verify the
timings from kingston)
- Replace the video card
- Memtest86 until it cries
- If it passes, bonnie++ on the new drives
- If that passes, usb/acpi/apic testing with the associated patches

Anyone still watching this? Tips and suggestions on what else might be useful/informative are more than welcome. The tests above mostly replicate what I did when building this box, and it passed them then..

Recap:
Epox 8rda+ nforce2 mobo
AMD Athlon XP 2500+ (Barton) 1.83G
Kingston HyperX PC3200
WD Caviar WD1200JB 8M/UDMA100
Antec case w/ 350W AMD-certified PSU

Quoted 3 lines

Oopses and occasional hangs, usually in do_generic_file_read, using stock kernel.org 2.4.2x kernels. Hardware passed testing (memtest86, bonnie++) before I put Linux on it.

Does this not relate directly to the APIC/IOAPIC issues with 2.6 kernel and nforce chipset motherboards?

Craig

#6 lore raw
Quoted 2 lines · Craig Bradney

Does this not relate directly to the APIC/IOAPIC issues with 2.6 kernel and nforce chipset motherboards?

Not when apic is disabled. (And I'm guessing you mean 2.4, right? Although IIRC the patches are available for both.)

The patches for enabling/using apic/io-apic on nforce2 work fine, except instead of oopsing it hardlocks semi-randomly. (It ran fine for a few days, then hardlocked, and upon reboot it only made it about 10 minutes, then another 30 minutes, so I went back to an unpatched 2.4.23 with 'noapic'. Now that the hardware has arrived I can move production elsewhere and do further testing.) I was under the impression that nmi-watchdog was supposed to prevent the hardlocks (well, turn them into oopses), but no such luck here.

Signature · 1 line
Disconnect <[email protected]>
#7 lore raw
In reply to Craig Bradney
Quoted 21 lines · Craig Bradney
Quoted 20 lines · Disconnect
Quoted 4 lines · Disconnect

memory/cpu timings. (Even underclocked it to 133 and 1G with no change.) So its unfortunately back on a stock 2.4.23-pre9 with noapic/noacpi. (It disables one of the sets of usb ports, as I recall, but it mostly works...)

Update: Underclocked from 1.8G to 1.2G (whups, meant to go down only 2-300mhz) and its been vaguely stable for about 1.5 days. I don't have another week (yet..) to run it under its normal load and wait for a crash, so what I'm going to do is:
- Move the workload (web/mail/..) to a different machine so this one
can be down for an extended period
- Replace the ram with new sticks (they arrived this morning)
- Reclock everything to stock (1.83G cpu, 200mhz ram and verify the
timings from kingston)
- Replace the video card
- Memtest86 until it cries
- If it passes, bonnie++ on the new drives
- If that passes, usb/acpi/apic testing with the associated patches

Quoted 6 lines
Quoted 6 lines

Anyone still watching this? Tips and suggestions on what else might be useful/informative are more than welcome. The tests above mostly replicate what I did when building this box, and it passed them then..

Recap:
Epox 8rda+ nforce2 mobo

I have Epox 8rda3+

Quoted 1 line
Quoted 1 line

AMD Athlon XP 2500+ (Barton) 1.83G

same

Quoted 1 line
Quoted 1 line

Kingston HyperX PC3200

corsair twinx 512 (2 stiks of 256)

Quoted 1 line
Quoted 1 line

WD Caviar WD1200JB 8M/UDMA100

seagate

Quoted 1 line
Quoted 1 line

Antec case w/ 350W AMD-certified PSU

black case (not something special)
My system works with 2.4 and 2.6 even overclocked to 10x190 (1900 Mhz) Cheers

Quoted 16 lines
Quoted 4 lines

Oopses and occasional hangs, usually in do_generic_file_read, using stock kernel.org 2.4.2x kernels. Hardware passed testing (memtest86, bonnie++) before I put Linux on it.

Does this not relate directly to the APIC/IOAPIC issues with 2.6 kernel and nforce chipset motherboards?

Craig

-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/

Signature · 16 lines
Liviu Voicu
Assistant Programmer and network support
Computation Center, Mount Scopus
Hebrew University of Jerusalem
Tel: 972(2)-5881253
E-mail: "Liviu Voicu"<[email protected]>

/**
 * cat /usr/src/linux/arch/i386/boot/bzImage > /dev/dsp
 * ( and the voice of God will be heard! )
 *
 */

Click here to see my GPG signature:
----------------------------------
	http://search.keyserver.net:11371/pks/lookup?template=netensearch%2Cnetennomatch%2Cnetenerror&search=pacman%40mscc.huji.ac.il&op=vindex&fingerprint=on&submit=Get+List