Skip to contentSkip to conversation

cpufreq: constant cpu_khz

BUG Peter Zijlstra 5 participants 8 messages Aug 25, 2008 x86

lore
Share & tools
lore
#0 lore raw

Hi,

I noticed that my sched_clock() was slow on a number of machine, so I started looking at cpufreq. I inserted the following printk in the set_cyc2ns_scale() function, which is called from the cpufreq change notifier:

arch/x86/kernel/tsc.c +3 -0
index 46af716..8dffd82 100644
@@ -261,6 +263,9 @@ static void set_cyc2ns_scale(unsigned long cpu_khz, int cpu)
 	if (cpu_khz)
 		*scale = (NSEC_PER_MSEC << CYC2NS_SCALE_FACTOR)/cpu_khz;
 
+	printk(KERN_DEBUG "set_cyc2ns_scale: %lu %lu %d\n",
+			*scale, cpu_khz, cpu);
+
 	sched_clock_idle_wakeup_event(0);
 	local_irq_restore(flags);
 }

And on my laptop:

root@lappy:~# dmesg | grep cyc2ns | tail -10
[   91.378693] set_cyc2ns_scale: 854 1198980 0
[   91.378713] set_cyc2ns_scale: 854 1198980 0
[  110.548064] set_cyc2ns_scale: 854 1198980 0
[  110.548106] set_cyc2ns_scale: 854 1198980 0
[  110.756121] set_cyc2ns_scale: 854 1198980 0
[  110.756135] set_cyc2ns_scale: 854 1198980 0
[  440.812063] set_cyc2ns_scale: 854 1198980 0
[  440.812106] set_cyc2ns_scale: 854 1198980 0
[  441.593140] set_cyc2ns_scale: 854 1198980 0
[  441.593159] set_cyc2ns_scale: 854 1198980 0

even though I'm sure there were some 800 <-> 1200 MHz changes there esp. as:

root@lappy:~# cat /proc/cpuinfo
processor       : 0
vendor_id       : GenuineIntel
cpu family      : 6
model           : 11
model name      : Intel(R) Pentium(R) III Mobile CPU      1200MHz
stepping        : 1
cpu MHz         : 798.000
cache size      : 512 KB
fdiv_bug        : no
hlt_bug         : no
f00f_bug        : no
coma_bug        : no
fpu             : yes
fpu_exception   : yes
cpuid level     : 2
wp              : yes
flags           : fpu vme de pse tsc msr pae mce cx8 sep mtrr pge mca cmov pat pse36 mmx fxsr sse up nopl
bogomips        : 1598.64
clflush size    : 32
power management:

On my opteron:

[root@opteron ~]# dmesg | grep cyc2ns | tail -10

[    0.000000] set_cyc2ns_scale: 427 2394004 0
[    0.000000] set_cyc2ns_scale: 427 2394004 1
[   14.784154] set_cyc2ns_scale: 427 2394004 0
[   14.788342] set_cyc2ns_scale: 427 2394004 1
[   14.792789] set_cyc2ns_scale: 427 2394004 0
[   14.797028] set_cyc2ns_scale: 427 2394004 1
[   15.841049] set_cyc2ns_scale: 427 2394004 0
[   15.845250] set_cyc2ns_scale: 427 2394004 1
[   15.849503] set_cyc2ns_scale: 427 2394004 0
[   15.853785] set_cyc2ns_scale: 427 2394004 1

[root@opteron ~]# cat /proc/cpuinfo
processor : 0
vendor_id : AuthenticAMD
cpu family : 15
model : 67
model name : Dual-Core AMD Opteron(tm) Processor 1216
stepping : 2
cpu MHz : 1000.000
cache size : 1024 KB
physical id : 0
siblings : 2
core id : 0
cpu cores : 2
apicid : 0
initial apicid : 0
fpu : yes
fpu_exception : yes
cpuid level : 1
wp : yes
flags : fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pat pse36 clflush mmx fxsr sse sse2 ht syscall nx mmxext fxsr_opt rdtscp lm 3dnowext 3dnow rep_good nopl pni cx16 lahf_lm cmp_legacy svm extapic cr8_legacy
bogomips : 1995.00
TLB size : 1024 4K pages
clflush size : 64
cache_alignment : 64
address sizes : 40 bits physical, 48 bits virtual
power management: ts fid vid ttp tm stc

processor : 1
vendor_id : AuthenticAMD
cpu family : 15
model : 67
model name : Dual-Core AMD Opteron(tm) Processor 1216
stepping : 2
cpu MHz : 1000.000
cache size : 1024 KB
physical id : 0
siblings : 2
core id : 1
cpu cores : 2
apicid : 1
initial apicid : 1
fpu : yes
fpu_exception : yes
cpuid level : 1
wp : yes
flags : fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pat pse36 clflush mmx fxsr sse sse2 ht syscall nx mmxext fxsr_opt rdtscp lm 3dnowext 3dnow rep_good nopl pni cx16 lahf_lm cmp_legacy svm extapic cr8_legacy
bogomips : 1995.00
TLB size : 1024 4K pages
clflush size : 64
cache_alignment : 64
address sizes : 40 bits physical, 48 bits virtual
power management: ts fid vid ttp tm stc

#1 lore raw
Quoted 5 lines · Peter Zijlstra

Hi,

I noticed that my sched_clock() was slow on a number of machine, so I started looking at cpufreq.

The below seems to fix the problem for me, but I don't have enough clue to know if its the correct fix, please advise.

arch/x86/kernel/tsc.c +0 -0
index 46af716..2f10f8c 100644
@@ -314,7 +319,7 @@ static int time_cpufreq_notifier(struct

notifier_block *nb, unsigned long val,
mark_tsc_unstable("cpufreq changes");
}

- set_cyc2ns_scale(tsc_khz_ref, freq->cpu);
+ set_cyc2ns_scale(tsc_khz, freq->cpu);

return 0;
}

#2 lore raw

Review

Quoted 8 lines · Peter Zijlstra
Quoted 2 lines

I noticed that my sched_clock() was slow on a number of machine, so I started looking at cpufreq.

The below seems to fix the problem for me, but I don't have enough clue to know if its the correct fix, please advise.

(quoted diff) +1 -1
-	set_cyc2ns_scale(tsc_khz_ref, freq->cpu);
+	set_cyc2ns_scale(tsc_khz, freq->cpu);

hm, that too is due to the tsc.c unification - Alok Cc:-ed. Applied your fix to x86/urgent.

Ingo

#3 lore raw
Quoted 5 lines · Ingo Molnar
Quoted 3 lines · Peter Zijlstra
Quoted 2 lines

I noticed that my sched_clock() was slow on a number of machine, so I started looking at cpufreq.

Quoted 9 lines

Review

Quoted 5 lines

The below seems to fix the problem for me, but I don't have enough clue to know if its the correct fix, please advise.

(quoted diff) +1 -1
-     set_cyc2ns_scale(tsc_khz_ref, freq->cpu);
+     set_cyc2ns_scale(tsc_khz, freq->cpu);

hm, that too is due to the tsc.c unification - Alok Cc:-ed. Applied your fix to x86/urgent.

Yes the fix is correct, tsc_khz can change in the loop above. Looking at tsc_64.c before the integration, that too had this bug but 32bit was good. So this is the right fix.

Thanks,
Alok

Quoted 1 line

Ingo

Acked-by
#4 lore raw
In reply to Ingo Molnar
Quoted 6 lines · Ingo Molnar
Quoted 3 lines · Peter Zijlstra
Quoted 2 lines

I noticed that my sched_clock() was slow on a number of machine, so I started looking at cpufreq.

Quoted 8 lines

Review

Quoted 5 lines

The below seems to fix the problem for me, but I don't have enough clue to know if its the correct fix, please advise.

(quoted diff) +1 -1
-	set_cyc2ns_scale(tsc_khz_ref, freq->cpu);
+	set_cyc2ns_scale(tsc_khz, freq->cpu);

hm, that too is due to the tsc.c unification - Alok Cc:-ed. Applied your fix to x86/urgent.

Good catch Peter. I'm puzzled how that bug was latent on 64bit for so long with no-one realising though.

Dave

Signature · 1 line
http://www.codemonkey.org.uk
#5 lore raw
Quoted 7 lines · Dave Jones
Quoted 2 lines

hm, that too is due to the tsc.c unification - Alok Cc:-ed. Applied your fix to x86/urgent.

Good catch Peter. I'm puzzled how that bug was latent on 64bit for so long with no-one realising though.

i think it's the combination of these two factors:

- bootup frequently is typically full-speed, so we calibrate things right

- cpufreq events are relatively slow-scale - and when they trigger the system is definitely not under load. So how precisely the scheduler functions isnt all that important in such scenarios - there's tons of CPU power available.

but it's indeed nice to have this fixed, and i just pushed the fix to Linus.

Ingo

#6 lore raw
Quoted 13 lines · On Mon, 25 Aug 2008 19:57:12 +0200 Ingo Molnar <[email protected]> wrote:
Quoted 7 lines · Dave Jones
Quoted 2 lines

hm, that too is due to the tsc.c unification - Alok Cc:-ed. Applied your fix to x86/urgent.

Good catch Peter. I'm puzzled how that bug was latent on 64bit for so long with no-one realising though.

i think it's the combination of these two factors:

Quoted 8 lines

- bootup frequently is typically full-speed, so we calibrate things right

- cpufreq events are relatively slow-scale - and when they trigger
the system is definitely not under load. So how precisely the scheduler functions isnt all that important in such scenarios - there's tons of CPU power available.

- many many of the 64 bit capable cpus are constant-tsc anyway

#7 lore raw
Quoted 19 lines · Arjan van de Ven
Quoted 16 lines · On Mon, 25 Aug 2008 19:57:12 +0200 Ingo Molnar <[email protected]> wrote:
Quoted 7 lines · Dave Jones
Quoted 2 lines

hm, that too is due to the tsc.c unification - Alok Cc:-ed. Applied your fix to x86/urgent.

Good catch Peter. I'm puzzled how that bug was latent on 64bit for so long with no-one realising though.

i think it's the combination of these two factors:

- bootup frequently is typically full-speed, so we calibrate things right

Quoted 7 lines
Quoted 5 lines

- cpufreq events are relatively slow-scale - and when they trigger
the system is definitely not under load. So how precisely the scheduler functions isnt all that important in such scenarios - there's tons of CPU power available.

- many many of the 64 bit capable cpus are constant-tsc anyway

Ironically I noticed it first on 64bit ;-)