Timer subsystem broken for Pentium M / early XEON / P6 family
Hi everybody,
Have a look at the system log of a kernel compile on a Pentium M 750 system. Speedstep is enabled, the kernel was booted with cpufreq.debug=2, the kernel is configured to give timing informations on printks. Timer frequency is 1000 Hz.
Kernel is 2.6.15-git7 + some changes unrelated to the timer subsystem.
Jan 21 10:14:42 linux kernel: [ 1260.867085] speedstep-centrino: no change needed - msr was and needs to be 612 Jan 21 10:14:43 linux kernel: [ 1261.081077] speedstep-centrino: no change needed - msr was and needs to be 612 Jan 21 10:14:45 linux kernel: [ 1262.150884] speedstep-centrino: target=1080050kHz old=800000 new=1067000 msr=0817 Jan 21 10:14:46 linux kernel: [ 947.401994] speedstep-centrino: no change needed - msr was and needs to be 817 Jan 21 10:14:46 linux kernel: [ 947.616092] speedstep-centrino: no change needed - msr was and needs to be 817 Jan 21 10:14:47 linux kernel: [ 947.830189] speedstep-centrino: target=1360100kHz old=1067000 new=1333000 msr=0a1c Jan 21 10:14:47 linux kernel: [ 757.097509] speedstep-centrino: no change needed - msr was and needs to be a1c Jan 21 10:14:48 linux kernel: [ 757.311216] speedstep-centrino: no change needed - msr was and needs to be a1c Jan 21 10:14:48 linux kernel: [ 757.524922] speedstep-centrino: target=1640150kHz old=1333000 new=1600000 msr=0c21 Jan 21 10:14:52 linux kernel: [ 632.382043] speedstep-centrino: no change needed - msr was and needs to be c21 Jan 21 10:14:52 linux kernel: [ 632.595621] speedstep-centrino: no change needed - msr was and needs to be c21 Jan 21 10:14:53 linux kernel: [ 632.809197] speedstep-centrino: target=1867000kHz old=1600000 new=1867000 msr=0e26 Jan 21 10:32:46 linux kernel: [ 1001.352174] speedstep-centrino: target=1773650kHz old=1867000 new=1600000 msr=0c21 Jan 21 10:32:47 linux kernel: [ 1167.961176] speedstep-centrino: target=1867000kHz old=1600000 new=1867000 msr=0e26 Jan 21 10:32:53 linux kernel: [ 1004.496699] speedstep-centrino: target=1773650kHz old=1867000 new=1600000 msr=0c21 Jan 21 10:32:55 linux kernel: [ 1172.054711] speedstep-centrino: no change needed - msr was and needs to be c21 Jan 21 10:32:57 linux kernel: [ 1172.909018] speedstep-centrino: target=1586950kHz old=1600000 new=1333000 msr=0a1c Jan 21 10:33:01 linux kernel: [ 1410.056608] speedstep-centrino: no change needed - msr was and needs to be a1c Jan 21 10:33:03 linux kernel: [ 1410.911434] speedstep-centrino: no change needed - msr was and needs to be a1c Jan 21 10:33:05 linux kernel: [ 1411.766261] speedstep-centrino: target=1306900kHz old=1333000 new=1067000 msr=0817 Jan 21 10:33:06 linux kernel: [ 1768.140232] speedstep-centrino: target=1400250kHz old=1067000 new=1333000 msr=0a1c Jan 21 10:33:08 linux kernel: [ 1412.792040] speedstep-centrino: target=1306900kHz old=1333000 new=1067000 msr=0817 Jan 21 10:33:10 linux kernel: [ 1770.067082] speedstep-centrino: no change needed - msr was and needs to be 817 Jan 21 10:33:12 linux kernel: [ 1770.923466] speedstep-centrino: no change needed - msr was and needs to be 817 Jan 21 10:33:14 linux kernel: [ 1771.779852] speedstep-centrino: target=1026850kHz old=1067000 new=800000 msr=0612 Jan 21 10:33:16 linux kernel: [ 2361.795520] speedstep-centrino: no change needed - msr was and needs to be 612 Jan 21 10:33:18 linux kernel: [ 2362.651387] speedstep-centrino: no change needed - msr was and needs to be 612
Pentium 4 and Xeon model 3 and above increment the time stamp counter at a constant rate independent of the current cpu frequency. Pentium M, Xeon up to model 2 and the P6 family increment with every internal processor cycle. This behaviour is documented in chapter 18.8 of IA-32 Intel® Architecture Software Developer’s Manual, Volume 3B: System Programming Guide, Part 2, Order Number: 253669-018, January 2006.
The timing information on printks should be monotone, it should be accurate, and it should be somewhat related to the system time.
Obviously printk time jumps. But it is also inaccurate compared against system time: 10:33:18 is 1116 seconds after 10:14:42, and so the printk time at 10:33:18 should be 1260+1116==2376 and not 2362. That means printk time lost 14 seconds against system time in less that 20 Minutes.
Someone should have a close look at the code.
cu,
Knut
Quoted 7 lines · Knut Petersen
Obviously printk time jumps. But it is also inaccurate compared against
system time: 10:33:18 is
1116 seconds after 10:14:42, and so the printk time at 10:33:18 should be 1260+1116==2376 and not 2362.
That means printk time lost 14 seconds against system time in less that 20 Minutes.
printk-time is really just a debugging/development thing - it doesn't try to be serious.
We _could_ use some more accurate time function for this, but most of them take locks, and taking locks inside printk isn't a great idea.
Plus syslogd already adds timestamps.