cpufreq, intel_pstate, Fix rounding errors
Subsystems (2)
v2 cover cpufreq, intel_pstate, Fix rounding errors
I have a Intel (6,63) processor with a "marketing" frequency (from /proc/cpuinfo) of 2100MHz, and a max turbo frequency of 2600MHz. I can execute
cpupower frequency-set -g powersave --min 1200MHz --max 2100MHz
and the max_freq_pct is set to 80. When adding load to the system I noticed that the cpu frequency only reached 2000MHZ and not 2100MHz as expected.
I wrote a little test program to set the frequencies in decrements of 100MHz and compared the targeted frequency (the frequency set through the cpupower command) and the actual frequency (from /proc/cpuinfo), as well as dumping out the value of the MSR_IA32_PERF_CTL.
Target Achieved Difference MSR(0x199)
3300 2900 -400 0x1e00 3200 2900 -300 0x1e00 3100 2900 -200 0x1e00 3000 2900 -100 0x1d00 2900 2800 -100 0x1c00 2800 2700 -100 0x1b00 2700 2600 -100 0x1a00 2600 2500 -100 0x1900 2500 2400 -100 0x1800 2400 2300 -100 0x1700 2300 2200 -100 0x1600 2200 2100 -100 0x1500 2100 2000 -100 0x1400 2000 1900 -100 0x1300 1900 1800 -100 0x1200 1800 1700 -100 0x1100 1700 1600 -100 0x1000 1600 1500 -100 0xf00 1500 1400 -100 0xe00 1400 1300 -100 0xd00 1300 1200 -100 0xc00 1200 1200 0 0xc00
As can be seen the frequencies are consistently off by 100MHz. After some examination I found a rounding error in intel_pstate_set_policy() for the calculation of limits->max_policy_pct which needs to be rounded up to the nearest percentage point. For example, setting a frequency of 2100MHz on this system results in limits->max_policy_pct = ((2100 * 100) / 2600) = 80.
However, ((2100 * 100) / 2600) is actually 80.7, or 81. This is fixed
by expanding the calculation an extra decimal point and rounding to the nearest percentage point.
A second rounding error was found in the calculation of limits->max_perf in intel_pstate_set_policy(), which is used to calculate the max and min pstate values in intel_pstate_get_min_max(). In this case, limits->max_perf is truncated to 2 hex digits such that, for example, 0x169 was incorrectly truncated to 0x16 instead of 0x17. This resulted in the pstate being set one level too low.
After applying these two fixes we consistently reach the targeted frequency.
Target Achieved Difference MSR(0x199)
3300 2900 -400 0x1e00 3200 2900 -300 0x1e00 3100 2900 -200 0x1e00 3000 2900 -100 0x1d00 2900 2900 0 0x1d00 2800 2800 0 0x1c00 2700 2700 0 0x1b00 2600 2600 0 0x1a00 2500 2500 0 0x1900 2400 2400 0 0x1800 2300 2300 0 0x1700 2200 2200 0 0x1600 2100 2100 0 0x1500 2000 2000 0 0x1400 1900 1900 0 0x1300 1800 1800 0 0x1200 1700 1700 0 0x1100 1600 1600 0 0x1000 1500 1500 0 0xf00 1400 1400 0 0xe00 1300 1300 0 0xd00 1200 1200 0 0xc00
Additional tests were run on a (6,78) with HWP enabled and a (6,79) system. Testing on both systems showed that the problem was resolved.
[v2]: Separate into two patches, and rebase onto git://git.kernel.org/pub/scm/linux/kernel/git/rafael/linux-pm.git linux-next Prarit Bhargava (2):
cpufreq, intel_pstate, Fix limits->max_policy_pct rounding error
cpufreq, intel_pstate, fix limits->max_perf rounding error
drivers/cpufreq/intel_pstate.c | 3 +++ 1 file changed, 3 insertions(+)
Signature · 1 line
1.8.3.1
PATCH v2 1/2 cpufreq, intel_pstate, Fix limits->max_policy_pct rounding error
I have a Intel (6,63) processor with a "marketing" frequency (from /proc/cpuinfo) of 2100MHz, and a max turbo frequency of 2600MHz. I can execute
cpupower frequency-set -g powersave --min 1200MHz --max 2100MHz
and the max_freq_pct is set to 80. When adding load to the system I noticed that the cpu frequency only reached 2000MHZ and not 2100MHz as expected.
This is because limits->max_policy_pct is calculated as 2100 * 100 /2600 = 80.7 and is rounded down to 80 when it should be rounded up to 81. This patch adds a DIV_ROUND_UP() which will return the correct value.
--- drivers/cpufreq/intel_pstate.c | 2 ++ 1 file changed, 2 insertions(+)
drivers/cpufreq/intel_pstate.c +2 -0
@@ -1110,6 +1110,8 @@ static int intel_pstate_set_policy(struct cpufreq_policy *policy) limits->min_policy_pct =clamp_t (int, limits->min_policy_pct,0 ,100 ); limits->max_policy_pct = (policy->max *100 ) / policy->cpuinfo.max_freq; limits->max_policy_pct =clamp_t (int, limits->max_policy_pct,0 ,100 ); + limits->max_policy_pct =DIV_ROUND_UP (policy->max *100 , + policy->cpuinfo.max_freq);/* Normalize user input to [min_policy_pct, max_policy_pct] */ limits->min_perf_pct = max(limits->min_policy_pct,
Signature · 1 line
1.8.3.1
Viresh Kumar
maintainerRe: [PATCH 1/2] cpufreq, intel_pstate, Fix limits->max_policy_pct rounding error
Quoted 30 lines · Prarit Bhargava
I have a Intel (6,63) processor with a "marketing" frequency (from
/proc/cpuinfo) of 2100MHz, and a max turbo frequency of 2600MHz. I can executecpupower frequency-set -g powersave --min 1200MHz --max 2100MHz
and the
max_freq_pctis set to 80. When adding load to the system I noticed that the cpu frequency only reached 2000MHZ and not 2100MHz as expected.This is because
limits->max_policy_pctis calculated as 2100 * 100 /2600 = 80.7 and is rounded down to 80 when it should be rounded up to 81. This patch adds aDIV_ROUND_UP()which will return the correct value.Cc Srinivas Pandruvada <[email protected]>Cc Len Brown <[email protected]>Cc Alexandra Yates <[email protected]>Cc Kristen Carlson Accardi <[email protected]>Cc "Rafael J. Wysocki" <[email protected]>Cc Viresh Kumar <[email protected]>Signed-off-by Prarit Bhargava <[email protected]>--- drivers/cpufreq/intel_pstate.c | 2 ++ 1 file changed, 2 insertions(+)drivers/cpufreq/intel_pstate.c +0
-0
drivers/cpufreq/intel_pstate.c static int intel_pstate_set_policy(struct cpufreq_policy *policy) · line 1111
Quoted 2 lines
drivers/cpufreq/intel_pstate.c +0
-0@@ -1110,6 +1110,8 @@ static int intel_pstate_set_policy(struct cpufreq_policy *policy) limits->min_policy_pct =clamp_t (int, limits->min_policy_pct,0 ,100 ); limits->max_policy_pct = (policy->max *100 ) / policy->cpuinfo.max_freq;
Forgot to remove this line ?
drivers/cpufreq/intel_pstate.c static int intel_pstate_set_policy(struct cpufreq_policy *policy) · line 1112
Quoted 1 line
drivers/cpufreq/intel_pstate.c +0
-0limits->max_policy_pct =clamp_t (int, limits->max_policy_pct,0 ,100 );
And put this after the later one ?
drivers/cpufreq/intel_pstate.c static int intel_pstate_set_policy(struct cpufreq_policy *policy) · line 1117
Quoted 5 lines
drivers/cpufreq/intel_pstate.c +2
-0+ limits->max_policy_pct =DIV_ROUND_UP (policy->max *100 , + policy->cpuinfo.max_freq);/* Normalize user input to [min_policy_pct, max_policy_pct] */ limits->min_perf_pct = max(limits->min_policy_pct,
Sure you tested it ? :)
Signature · 1 line
viresh
Re: [PATCH 1/2] cpufreq, intel_pstate, Fix limits->max_policy_pct rounding error
Quoted 40 lines · Viresh Kumar
Quoted 30 lines · Prarit Bhargava
I have a Intel (6,63) processor with a "marketing" frequency (from
/proc/cpuinfo) of 2100MHz, and a max turbo frequency of 2600MHz. I can executecpupower frequency-set -g powersave --min 1200MHz --max 2100MHz
and the
max_freq_pctis set to 80. When adding load to the system I noticed that the cpu frequency only reached 2000MHZ and not 2100MHz as expected.This is because
limits->max_policy_pctis calculated as 2100 * 100 /2600 = 80.7 and is rounded down to 80 when it should be rounded up to 81. This patch adds aDIV_ROUND_UP()which will return the correct value.Cc Srinivas Pandruvada <[email protected]>Cc Len Brown <[email protected]>Cc Alexandra Yates <[email protected]>Cc Kristen Carlson Accardi <[email protected]>Cc "Rafael J. Wysocki" <[email protected]>Cc Viresh Kumar <[email protected]>Signed-off-by Prarit Bhargava <[email protected]>--- drivers/cpufreq/intel_pstate.c | 2 ++ 1 file changed, 2 insertions(+)drivers/cpufreq/intel_pstate.c +0
-0drivers/cpufreq/intel_pstate.c static int intel_pstate_set_policy(struct cpufreq_policy *policy) · line 1111
Quoted 2 lines
drivers/cpufreq/intel_pstate.c +0
-0@@ -1110,6 +1110,8 @@ static int intel_pstate_set_policy(struct cpufreq_policy *policy) limits->min_policy_pct =clamp_t (int, limits->min_policy_pct,0 ,100 ); limits->max_policy_pct = (policy->max *100 ) / policy->cpuinfo.max_freq;Forgot to remove this line ?
drivers/cpufreq/intel_pstate.c static int intel_pstate_set_policy(struct cpufreq_policy *policy) · line 1112
Quoted 1 line
drivers/cpufreq/intel_pstate.c +0
-0limits->max_policy_pct =clamp_t (int, limits->max_policy_pct,0 ,100 );And put this after the later one ?
Quoted 7 lines
drivers/cpufreq/intel_pstate.c static int intel_pstate_set_policy(struct cpufreq_policy *policy) · line 1117
Quoted 5 lines
drivers/cpufreq/intel_pstate.c +2
-0+ limits->max_policy_pct =DIV_ROUND_UP (policy->max *100 , + policy->cpuinfo.max_freq);/* Normalize user input to [min_policy_pct, max_policy_pct] */ limits->min_perf_pct = max(limits->min_policy_pct,Sure you tested it ? :)
Oops -- and yeah, tested. It works because I rewrite the value of max_policy_pct :). I'll repost shortly.
P.
Quoted 1 line
Viresh Kumar
maintainerRe: [PATCH 1/2] cpufreq, intel_pstate, Fix limits->max_policy_pct rounding error
Quoted 7 lines · Prarit Bhargava
Quoted 7 lines
Quoted 1 line
limits->max_policy_pct=clamp_t(int, limits->max_policy_pct, 0 , 100);And put this after the later one ?
Quoted 3 lines
drivers/cpufreq/intel_pstate.c +2
-0+ limits->max_policy_pct =DIV_ROUND_UP (policy->max *100 , + policy->cpuinfo.max_freq);
Quoted 7 lines
Quoted 4 lines
drivers/cpufreq/intel_pstate.c static int intel_pstate_set_policy(struct cpufreq_policy *policy) · line 1117
Quoted 2 lines
drivers/cpufreq/intel_pstate.c +0
-0/* Normalize user input to [min_policy_pct, max_policy_pct] */ limits->min_perf_pct = max(limits->min_policy_pct,Sure you tested it ? :)
Oops -- and yeah, tested. It works because I rewrite the value of
max_policy_pct:). I'll repost shortly.
But we aren't doing below anymore, doesn't this change the calculations at all?
limits->max_policy_pct = clamp_t(int, limits->max_policy_pct, 0 , 100);
Signature · 1 line
viresh
Re: [PATCH 1/2] cpufreq, intel_pstate, Fix limits->max_policy_pct rounding error
Quoted 13 lines · Viresh Kumar
Quoted 12 lines · Prarit Bhargava
Quoted 11 lines
Quoted 1 line
limits->max_policy_pct=clamp_t(int, limits->max_policy_pct, 0 , 100);And put this after the later one ?
drivers/cpufreq/intel_pstate.c static int intel_pstate_set_policy(struct cpufreq_policy *policy) · line 1117
Quoted 5 lines
drivers/cpufreq/intel_pstate.c +2
-0+ limits->max_policy_pct =DIV_ROUND_UP (policy->max *100 , + policy->cpuinfo.max_freq);/* Normalize user input to [min_policy_pct, max_policy_pct] */ limits->min_perf_pct = max(limits->min_policy_pct,Sure you tested it ? :)
Quoted 7 lines
Quoted 2 lines
Oops -- and yeah, tested. It works because I rewrite the value of
max_policy_pct:). I'll repost shortly.But we aren't doing below anymore, doesn't this change the calculations at all?
limits->max_policy_pct = clamp_t(int, limits->max_policy_pct, 0 , 100);
The clamp only confines the max_policy between 0 and 100. AFAIK it doesn't make any change tothe value of limits->max_policy_pct unless it was outside of that range.
P.
Quoted 1 line
Re: [PATCH 1/2] cpufreq, intel_pstate, Fix limits->max_policy_pct rounding error
Quoted 21 lines · Prarit Bhargava
Quoted 19 lines · Viresh Kumar
Quoted 7 lines · Prarit Bhargava
Quoted 7 lines
Quoted 1 line
limits->max_policy_pct=clamp_t(int, limits->max_policy_pct, 0 , 100);And put this after the later one ?
Quoted 3 lines
(quoted diff) +2
-0+ limits->max_policy_pct = DIV_ROUND_UP(policy->max * 100, + policy->cpuinfo.max_freq);Quoted 7 lines
Quoted 4 lines
Review
Quoted 2 lines
(quoted diff) +0
-0/* Normalize user input to [min_policy_pct, max_policy_pct] */ limits->min_perf_pct = max(limits->min_policy_pct,Sure you tested it ? :)
Oops -- and yeah, tested. It works because I rewrite the value of
max_policy_pct:). I'll repost shortly.But we aren't doing below anymore, doesn't this change the calculations at all?
Quoted 8 lines
Quoted 1 line
limits->max_policy_pct = clamp_t(int, limits->max_policy_pct, 0 , 100);The clamp only confines the
max_policybetween 0 and 100. AFAIK it doesn't make any change tothe value oflimits->max_policy_pctunless it was outside of that range.P.
Quoted 1 line
With the changes below (as suggested above), I did tests. Except two cases, it did correct. Those two are in turbo range, so I am OK with that.
(quoted diff) +0 -0
b/drivers/cpufreq/intel_pstate.c
index b78abe9..c3bcca4 100644
drivers/cpufreq/intel_pstate.c +0 -0
@@ -1111,9 +1111,9 @@ static int intel_pstate_set_policy(struct
cpufreq_policy *policy)
limits = &powersave_limits;
limits->min_policy_pct = (policy->min * 100) /policy->cpuinfo.max_freq;
limits->min_policy_pct = clamp_t(int, limits->min_policy_pct, 0 ,
100);
- limits->max_policy_pct = (policy->max * 100) /policy->cpuinfo.max_freq;
+ limits->max_policy_pct = DIV_ROUND_UP(policy->max * 100,
+ policy->cpuinfo.max_freq);
limits->max_policy_pct = clamp_t(int, limits->max_policy_pct, 0 ,
100);
-
/* Normalize user input to [min_policy_pct, max_policy_pct] */ limits->min_perf_pct = max(limits->min_policy_pct,
limits->min_sysfs_pct);
@@ -1131,7 +1131,7 @@ static int intel_pstate_set_policy(struct cpufreq_policy *policy)
int_tofp(100));
limits->max_perf = div_fp(int_tofp(limits->max_perf_pct),
int_tofp(100));
-
+ limits->max_perf = round_up(limits->max_perf, 8);
if (hwp_active)
intel_pstate_hwp_set();
3300 : OK
3200 : 1 less
3100 : 1 less
3000 : 1 less
2900 : OK
2800 : OK
2700 : OK
2600 : OK
2500 : OK
2400 : OK
2300 : OK
2200 : OK
2100 : OK
2000 : OK
1900 : OK
1800 : OK
1700 : OK
1600 : OK
1500 : OK
1400 : OK
1300 : OK
1200 : OK
1100 : OK
1000 : OK
900 : OK
800 : OK
Thanks,
Srinivas
Re: [PATCH 1/2] cpufreq, intel_pstate, Fix limits->max_policy_pct rounding error
Quoted 69 lines · Srinivas
Quoted 21 lines · Prarit Bhargava
Quoted 19 lines · Viresh Kumar
Quoted 7 lines · Prarit Bhargava
Quoted 7 lines
Quoted 1 line
limits->max_policy_pct=clamp_t(int, limits->max_policy_pct, 0 , 100);And put this after the later one ?
Quoted 3 lines
(quoted diff) +2
-0+ limits->max_policy_pct = DIV_ROUND_UP(policy->max * 100, + policy->cpuinfo.max_freq);Quoted 7 lines
Quoted 4 lines
Review
Quoted 2 lines
(quoted diff) +0
-0/* Normalize user input to [min_policy_pct, max_policy_pct] */ limits->min_perf_pct = max(limits->min_policy_pct,Sure you tested it ? :)
Oops -- and yeah, tested. It works because I rewrite the value of
max_policy_pct:). I'll repost shortly.But we aren't doing below anymore, doesn't this change the calculations at all?
Quoted 8 lines
Quoted 1 line
limits->max_policy_pct = clamp_t(int, limits->max_policy_pct, 0 , 100);The clamp only confines the
max_policybetween 0 and 100. AFAIK it doesn't make any change tothe value oflimits->max_policy_pctunless it was outside of that range.P.
Quoted 1 line
With the changes below (as suggested above), I did tests. Except two cases, it did correct. Those two are in turbo range, so I am OK with that.
drivers/cpufreq/intel_pstate.c +2
-1@@ -1111,9 +1111,9 @@ static int intel_pstate_set_policy(struct cpufreq_policy *policy) limits = &powersave_limits; limits->min_policy_pct = (policy->min *100 ) / policy->cpuinfo.max_freq; limits->min_policy_pct =clamp_t (int, limits->min_policy_pct,0 ,100 ); - limits->max_policy_pct = (policy->max *100 ) / policy->cpuinfo.max_freq; + limits->max_policy_pct =DIV_ROUND_UP (policy->max *100 , + policy->cpuinfo.max_freq); limits->max_policy_pct = clamp_t(int, limits->max_policy_pct, 0 ,100);
drivers/cpufreq/intel_pstate.c +1
-2-/* Normalize user input to [min_policy_pct, max_policy_pct] */ limits->min_perf_pct =max (limits->min_policy_pct, limits->min_sysfs_pct);@@ -1131,7 +1131,7 @@ static int intel_pstate_set_policy(struct cpufreq_policy *policy) int_tofp(100)); limits->max_perf =div_fp (int_tofp (limits->max_perf_pct),int_tofp (100 )); - + limits->max_perf =round_up (limits->max_perf,8 );if (hwp_active)intel_pstate_hwp_set ();
Quoted 4 lines
3300 : OK
3200 : 1 less
3100 : 1 less
3000 : 1 less
The -1 difference here is not unexpected given the other probable rounding errors in the frequency code. I have a feeling that no one really has done an in depth review to find the errors. I'm not going to because I'm pretty sure I/we can convince users that 3200 == 3199.98 ;). FWIW, I've also wondered if the difference between the marketing frequency and the TSC frequency (which in theory equals the marketing frequency) can cause this sort of error.
OOC did you try loading the system down (with a kernel build) and switching between frequencies? That's a good way to see the effect of the turbo states. I would expect that the turbo state hits a maximum of about 75% of the max turbo state value (based on experiment) so the differences should be larger at the high end.
P.
Re: [PATCH 1/2] cpufreq, intel_pstate, Fix limits->max_policy_pct rounding error
Quoted 1 line · Prarit Bhargava
[cut]
Quoted 2 lines
The -1 difference here is not unexpected given the other probable rounding errors in the frequency code.
Yes. Intel P state cpufreq interface is not optimal. We even debate whether we should have this interface at all.
Quoted 5 lines
I have a feeling that no one really has done an
in depth review to find the errors. I'm not going to because I'm pretty sure I/we can convince users that 3200 == 3199.98 ;). FWIW, I've also wondered if the difference between the marketing frequency and the TSC frequency (which in theory equals the marketing frequency) can cause this sort of error.
We don't even request correct pstate here, so we will not get that. But in this case in turbo region is not controllable (after Sandybridge ) above something called turbo activation ratio. So not a big deal. As long as we can set at lower end we are fine.
Quoted 1 line
Thanks,
Srinivas
Quoted 7 lines
OOC did you try loading the system down (with a kernel build) and switching between frequencies? That's a good way to see the effect of the turbo states. I would expect that the turbo state hits a maximum of about 75% of the max turbo state value (based on experiment) so the differences should be larger at the high end.
P.
PATCH v2 2/2 cpufreq, intel_pstate, fix limits->max_perf rounding error
In reply to Prarit BhargavaA rounding error was found in the calculation of limits->max_perf in intel_pstate_set_policy(), which is used to calculate the max and min pstate values in intel_pstate_get_min_max(). In that code, limits->max_perf is truncated to 2 hex digits such that, for example, 0x169 was incorrectly calculated to 0x16 instead of 0x17. This resulted in the pstate being set one level too low. This patch rounds the value of limits->max_perf up instead of down so that the correct max pstate can be reached.
--- drivers/cpufreq/intel_pstate.c | 1 + 1 file changed, 1 insertion(+)
drivers/cpufreq/intel_pstate.c +1 -0
@@ -1122,6 +1122,7 @@ static int intel_pstate_set_policy(struct cpufreq_policy *policy) limits->max_sysfs_pct); limits->max_perf_pct =max (limits->min_policy_pct, limits->max_perf_pct); + limits->max_perf =round_up (limits->max_perf,8 );/* Make sure min_perf_pct <= max_perf_pct */ limits->min_perf_pct =min (limits->max_perf_pct, limits->min_perf_pct);
Signature · 1 line
1.8.3.1