x86-64: double timer interrupts in recent 2.4.x

DISCUSSION Peter Cordes 4 participants 4 messages Jun 16, 2004 – Jun 21, 2004

Share & tools

Peter Cordes

author

2004-06-16 19:32 UTC #0 lore raw

Nobody replied to this message on [email protected], or
[email protected]. Hopefully I've found the right places to send this this time around. Actually, Roland Fehrenbacher saw my message in a list archive and mailed me to confirm that he saw the same double-speed clock problem on two different machines, so it's not just Tyan S2880 boards. He suggested I mail Andi and lkml, so here goes. (I haven't tested again with anything more recent than 2.4.27-pre2, so if this is fixed, sorry.)

-----

I just noticed that on my Opteron cluster, the nodes that are running 64bit
kernels have their clocks ticking at double speed. This happens with Linux 2.4.26, and 2.4.27-pre2, compiled with gcc 3.3.3 (Debian 20040401) in a Debian pure64 chroot. Linux 2.4.25, compiled on Debian Woody + bi-arch gcc 3.3.2 20030908, does _not_ have the problem. The config options were pretty much the same for all kernels, and all the kernels are plain vanilla flavour from www.ca.kernel.org.

If I run ntpdate to set the clock, then 10 seconds later it will be 10
seconds fast. Running date(1), the system time advances 20 seconds in 10 seconds of real time. (I haven't done anything weird with adjtimex(8).) time sleep 10 takes 5 seconds, but bash reports its real time as 10 seconds. The timer interrupt counter is increasing at a rate of 200/real second, so it seems like the system is getting timer interrupts twice as fast as it should. (With 2.4.25, it is 100/sec, same as HZ).

Linux says it is using the PIT and TSC timers. I have HPET enabled in my
Linux config, but I guess Tyan's S2880 mobo doesn't have one. This is a dual-Opteron 240 machine, BTW.

i386 Linux on the same machines has no problems with timekeeping. (But I
haven't tested versions later than 2.4.25 in legacy mode.)

I spent some time poking around the timer code that increments xtime, but I
guess the fact that the timer irqs are coming at double speed indicates that the problem lies elsewhere. Maybe the code that sets up the timer?

$ uname -a
Linux node6.cs.dal.ca 2.4.26 #2 SMP Fri May 14 14:46:42 ADT 2004 x86_64 x86_64 x86_64 GNU/Linux
$ cat /proc/interrupts
           CPU0       CPU1
  0:    4415908          0    IO-APIC-edge  timer
  1:          2          0    IO-APIC-edge  keyboard
  2:          0          0          XT-PIC  cascade
  9:          0          0   IO-APIC-level  acpi
 14:      17861          1    IO-APIC-edge  ide0
 19:          0          0   IO-APIC-level  usb-ohci, usb-ohci
 24:     563942          0   IO-APIC-level  eth0
 25:     564331          0   IO-APIC-level  eth1
NMI:      19097      19097
LOC:    2211090    2211095
ERR:          0
MIS:          0

Only CPU0 is getting the timer interrupt, but at least we know it's not
that both CPUs are getting the timer interrupt. (Both CPUs get 100 LOC:
(local APIC) interrupts/sec, but that happens on the non-buggy 2.4.25, too.)

Thanks for any help,

I'm not subscribed to the lkml, so please CC me on any followups.

Signature · 6 lines

#define X(x,y) x##y
Peter Cordes ;  e-mail: X(peter@cor , des.ca)

"The gods confound the man who first found out how to distinguish the hours!
 Confound him, too, who in this place set up a sundial, to cut and hack
 my day so wretchedly into small pieces!" -- Plautus, 200 BC

Andi Kleen

2004-06-16 19:42 UTC #1 lore raw

Quoted 8 lines · On Wed, 16 Jun 2004 16:28:26 -0300 Peter Cordes <[email protected]> wrote:

Nobody replied to this message on [email protected], or
[email protected]. Hopefully I've found the right places to send this this time around. Actually, Roland Fehrenbacher saw my message in a list archive and mailed me to confirm that he saw the same double-speed clock problem on two different machines, so it's not just Tyan S2880 boards. He suggested I mail Andi and lkml, so here goes. (I haven't tested again with anything more recent than 2.4.27-pre2, so if this is fixed, sorry.)

It would be a good start if you could track down which kernel started causing this behaviour (best with -bk* kernels, -pre* is not fine grained enough).

-Andi

Milan Gabor

2004-06-21 10:46 UTC #2 lore raw

In reply to Peter Cordes

Hi!

I have Suse 9.0 and dual Opteron on MSI K8T Master 2 motherboard.
I also get interrupts only on one cpu and my clock is ticking strange, so I have to synchronize it with NTP server frequently.

This is from my system:
CPU0 CPU1

  0:      30434   16139843    IO-APIC-edge  timer
  1:        944          0    IO-APIC-edge  keyboard
  2:          0          0          XT-PIC  cascade
 14:         30          1    IO-APIC-edge  ide0
 16:     657371          0   IO-APIC-level  eth0
 20:     261267          0   IO-APIC-level  libata
NMI:     694146     873271
LOC:   16167676   16167576
ERR:          1
MIS:          0

Linux www 2.4.21-226-smp #1 SMP Tue Jun 15 09:14:10 UTC 2004 x86_64 x86_64 x86_64 GNU/Linux

I am also running irq_balance and acpi=off set from grub boot menu. Without acpi=off system never boots.

Is there any solution, so clock will work OK and interrupts will be on both CPUs?

MIlan

Quoted 62 lines · Peter Cordes

Nobody replied to this message on [email protected], or
[email protected]. Hopefully I've found the right places to send this this time around. Actually, Roland Fehrenbacher saw my message in a list archive and mailed me to confirm that he saw the same double-speed clock problem on two different machines, so it's not just Tyan S2880 boards. He suggested I mail Andi and lkml, so here goes. (I haven't tested again with anything more recent than 2.4.27-pre2, so if this is fixed, sorry.)
-----
I just noticed that on my Opteron cluster, the nodes that are running 64bit
kernels have their clocks ticking at double speed. This happens with Linux 2.4.26, and 2.4.27-pre2, compiled with gcc 3.3.3 (Debian 20040401) in a Debian pure64 chroot. Linux 2.4.25, compiled on Debian Woody + bi-arch gcc 3.3.2 20030908, does _not_ have the problem. The config options were pretty much the same for all kernels, and all the kernels are plain vanilla flavour from www.ca.kernel.org.
If I run ntpdate to set the clock, then 10 seconds later it will be 10
seconds fast. Running date(1), the system time advances 20 seconds in 10 seconds of real time. (I haven't done anything weird with adjtimex(8).) time sleep 10 takes 5 seconds, but bash reports its real time as 10 seconds. The timer interrupt counter is increasing at a rate of 200/real second, so it seems like the system is getting timer interrupts twice as fast as it should. (With 2.4.25, it is 100/sec, same as HZ).
Linux says it is using the PIT and TSC timers. I have HPET enabled in my
Linux config, but I guess Tyan's S2880 mobo doesn't have one. This is a dual-Opteron 240 machine, BTW.
i386 Linux on the same machines has no problems with timekeeping. (But I
haven't tested versions later than 2.4.25 in legacy mode.)
I spent some time poking around the timer code that increments xtime, but I
guess the fact that the timer irqs are coming at double speed indicates that the problem lies elsewhere. Maybe the code that sets up the timer?
$ uname -a
Linux node6.cs.dal.ca 2.4.26 #2 SMP Fri May 14 14:46:42 ADT 2004 x86_64 x86_64 x86_64 GNU/Linux
$ cat /proc/interrupts
           CPU0       CPU1
  0:    4415908          0    IO-APIC-edge  timer
  1:          2          0    IO-APIC-edge  keyboard
  2:          0          0          XT-PIC  cascade
  9:          0          0   IO-APIC-level  acpi
 14:      17861          1    IO-APIC-edge  ide0
 19:          0          0   IO-APIC-level  usb-ohci, usb-ohci
 24:     563942          0   IO-APIC-level  eth0
 25:     564331          0   IO-APIC-level  eth1
NMI:      19097      19097
LOC:    2211090    2211095
ERR:          0
MIS:          0
Only CPU0 is getting the timer interrupt, but at least we know it's not
that both CPUs are getting the timer interrupt. (Both CPUs get 100 LOC:
(local APIC) interrupts/sec, but that happens on the non-buggy 2.4.25, too.)
Thanks for any help,
I'm not subscribed to the lkml, so please CC me on any followups.

Vojtech Pavlik

2004-06-21 11:04 UTC #3 lore raw

Quoted 23 lines · Milan Gabor

Hi!
I have Suse 9.0 and dual Opteron on MSI K8T Master 2 motherboard. I also get interrupts only on one cpu and my clock is ticking strange, so I have to synchronize it with NTP server frequently.
This is from my system:
CPU0 CPU1
  0:      30434   16139843    IO-APIC-edge  timer
  1:        944          0    IO-APIC-edge  keyboard
  2:          0          0          XT-PIC  cascade
 14:         30          1    IO-APIC-edge  ide0
 16:     657371          0   IO-APIC-level  eth0
 20:     261267          0   IO-APIC-level  libata
NMI:     694146     873271
LOC:   16167676   16167576
ERR:          1
MIS:          0
Linux www 2.4.21-226-smp #1 SMP Tue Jun 15 09:14:10 UTC 2004 x86_64 x86_64 x86_64 GNU/Linux

Quoted 7 lines

I am also running irq_balance and acpi=off set from grub boot menu. Without acpi=off system never boots.
Is there any solution, so clock will work OK and interrupts will be on both CPUs?
MIlan

This patch could fix that (replace i386 with x86_64):

http://marc.theaimsgroup.com/?l=linux-kernel&m=108774225111967&w=2

Signature · 2 lines

Vojtech Pavlik
SuSE Labs, SuSE CR