1 CPU frequency and voltage scaling code in the Linux(TM) kernel 2 3 4 L i n u x C P U F r e q 5 6 C P U F r e q G o v e r n o r s 7 8 - information for users and developers - 9 10 11 Dominik Brodowski <linux@brodo.de> 12 some additions and corrections by Nico Golde <nico@ngolde.de> 13 14 15 16 Clock scaling allows you to change the clock speed of the CPUs on the 17 fly. This is a nice method to save battery power, because the lower 18 the clock speed, the less power the CPU consumes. 19 20 21Contents: 22--------- 231. What is a CPUFreq Governor? 24 252. Governors In the Linux Kernel 262.1 Performance 272.2 Powersave 282.3 Userspace 292.4 Ondemand 302.5 Conservative 31 323. The Governor Interface in the CPUfreq Core 33 34 35 361. What Is A CPUFreq Governor? 37============================== 38 39Most cpufreq drivers (in fact, all except one, longrun) or even most 40cpu frequency scaling algorithms only offer the CPU to be set to one 41frequency. In order to offer dynamic frequency scaling, the cpufreq 42core must be able to tell these drivers of a "target frequency". So 43these specific drivers will be transformed to offer a "->target/target_index" 44call instead of the existing "->setpolicy" call. For "longrun", all 45stays the same, though. 46 47How to decide what frequency within the CPUfreq policy should be used? 48That's done using "cpufreq governors". Two are already in this patch 49-- they're the already existing "powersave" and "performance" which 50set the frequency statically to the lowest or highest frequency, 51respectively. At least two more such governors will be ready for 52addition in the near future, but likely many more as there are various 53different theories and models about dynamic frequency scaling 54around. Using such a generic interface as cpufreq offers to scaling 55governors, these can be tested extensively, and the best one can be 56selected for each specific use. 57 58Basically, it's the following flow graph: 59 60CPU can be set to switch independently | CPU can only be set 61 within specific "limits" | to specific frequencies 62 63 "CPUfreq policy" 64 consists of frequency limits (policy->{min,max}) 65 and CPUfreq governor to be used 66 / \ 67 / \ 68 / the cpufreq governor decides 69 / (dynamically or statically) 70 / what target_freq to set within 71 / the limits of policy->{min,max} 72 / \ 73 / \ 74 Using the ->setpolicy call, Using the ->target/target_index call, 75 the limits and the the frequency closest 76 "policy" is set. to target_freq is set. 77 It is assured that it 78 is within policy->{min,max} 79 80 812. Governors In the Linux Kernel 82================================ 83 842.1 Performance 85--------------- 86 87The CPUfreq governor "performance" sets the CPU statically to the 88highest frequency within the borders of scaling_min_freq and 89scaling_max_freq. 90 91 922.2 Powersave 93------------- 94 95The CPUfreq governor "powersave" sets the CPU statically to the 96lowest frequency within the borders of scaling_min_freq and 97scaling_max_freq. 98 99 1002.3 Userspace 101------------- 102 103The CPUfreq governor "userspace" allows the user, or any userspace 104program running with UID "root", to set the CPU to a specific frequency 105by making a sysfs file "scaling_setspeed" available in the CPU-device 106directory. 107 108 1092.4 Ondemand 110------------ 111 112The CPUfreq governor "ondemand" sets the CPU depending on the 113current usage. To do this the CPU must have the capability to 114switch the frequency very quickly. There are a number of sysfs file 115accessible parameters: 116 117sampling_rate: measured in uS (10^-6 seconds), this is how often you 118want the kernel to look at the CPU usage and to make decisions on 119what to do about the frequency. Typically this is set to values of 120around '10000' or more. It's default value is (cmp. with users-guide.txt): 121transition_latency * 1000 122Be aware that transition latency is in ns and sampling_rate is in us, so you 123get the same sysfs value by default. 124Sampling rate should always get adjusted considering the transition latency 125To set the sampling rate 750 times as high as the transition latency 126in the bash (as said, 1000 is default), do: 127echo `$(($(cat cpuinfo_transition_latency) * 750 / 1000)) \ 128 >ondemand/sampling_rate 129 130sampling_rate_min: 131The sampling rate is limited by the HW transition latency: 132transition_latency * 100 133Or by kernel restrictions: 134If CONFIG_NO_HZ_COMMON is set, the limit is 10ms fixed. 135If CONFIG_NO_HZ_COMMON is not set or nohz=off boot parameter is used, the 136limits depend on the CONFIG_HZ option: 137HZ=1000: min=20000us (20ms) 138HZ=250: min=80000us (80ms) 139HZ=100: min=200000us (200ms) 140The highest value of kernel and HW latency restrictions is shown and 141used as the minimum sampling rate. 142 143up_threshold: defines what the average CPU usage between the samplings 144of 'sampling_rate' needs to be for the kernel to make a decision on 145whether it should increase the frequency. For example when it is set 146to its default value of '95' it means that between the checking 147intervals the CPU needs to be on average more than 95% in use to then 148decide that the CPU frequency needs to be increased. 149 150ignore_nice_load: this parameter takes a value of '0' or '1'. When 151set to '0' (its default), all processes are counted towards the 152'cpu utilisation' value. When set to '1', the processes that are 153run with a 'nice' value will not count (and thus be ignored) in the 154overall usage calculation. This is useful if you are running a CPU 155intensive calculation on your laptop that you do not care how long it 156takes to complete as you can 'nice' it and prevent it from taking part 157in the deciding process of whether to increase your CPU frequency. 158 159sampling_down_factor: this parameter controls the rate at which the 160kernel makes a decision on when to decrease the frequency while running 161at top speed. When set to 1 (the default) decisions to reevaluate load 162are made at the same interval regardless of current clock speed. But 163when set to greater than 1 (e.g. 100) it acts as a multiplier for the 164scheduling interval for reevaluating load when the CPU is at its top 165speed due to high load. This improves performance by reducing the overhead 166of load evaluation and helping the CPU stay at its top speed when truly 167busy, rather than shifting back and forth in speed. This tunable has no 168effect on behavior at lower speeds/lower CPU loads. 169 170powersave_bias: this parameter takes a value between 0 to 1000. It 171defines the percentage (times 10) value of the target frequency that 172will be shaved off of the target. For example, when set to 100 -- 10%, 173when ondemand governor would have targeted 1000 MHz, it will target 1741000 MHz - (10% of 1000 MHz) = 900 MHz instead. This is set to 0 175(disabled) by default. 176When AMD frequency sensitivity powersave bias driver -- 177drivers/cpufreq/amd_freq_sensitivity.c is loaded, this parameter 178defines the workload frequency sensitivity threshold in which a lower 179frequency is chosen instead of ondemand governor's original target. 180The frequency sensitivity is a hardware reported (on AMD Family 16h 181Processors and above) value between 0 to 100% that tells software how 182the performance of the workload running on a CPU will change when 183frequency changes. A workload with sensitivity of 0% (memory/IO-bound) 184will not perform any better on higher core frequency, whereas a 185workload with sensitivity of 100% (CPU-bound) will perform better 186higher the frequency. When the driver is loaded, this is set to 400 187by default -- for CPUs running workloads with sensitivity value below 18840%, a lower frequency is chosen. Unloading the driver or writing 0 189will disable this feature. 190 191 1922.5 Conservative 193---------------- 194 195The CPUfreq governor "conservative", much like the "ondemand" 196governor, sets the CPU depending on the current usage. It differs in 197behaviour in that it gracefully increases and decreases the CPU speed 198rather than jumping to max speed the moment there is any load on the 199CPU. This behaviour more suitable in a battery powered environment. 200The governor is tweaked in the same manner as the "ondemand" governor 201through sysfs with the addition of: 202 203freq_step: this describes what percentage steps the cpu freq should be 204increased and decreased smoothly by. By default the cpu frequency will 205increase in 5% chunks of your maximum cpu frequency. You can change this 206value to anywhere between 0 and 100 where '0' will effectively lock your 207CPU at a speed regardless of its load whilst '100' will, in theory, make 208it behave identically to the "ondemand" governor. 209 210down_threshold: same as the 'up_threshold' found for the "ondemand" 211governor but for the opposite direction. For example when set to its 212default value of '20' it means that if the CPU usage needs to be below 21320% between samples to have the frequency decreased. 214 215sampling_down_factor: similar functionality as in "ondemand" governor. 216But in "conservative", it controls the rate at which the kernel makes 217a decision on when to decrease the frequency while running in any 218speed. Load for frequency increase is still evaluated every 219sampling rate. 220 2213. The Governor Interface in the CPUfreq Core 222============================================= 223 224A new governor must register itself with the CPUfreq core using 225"cpufreq_register_governor". The struct cpufreq_governor, which has to 226be passed to that function, must contain the following values: 227 228governor->name - A unique name for this governor 229governor->governor - The governor callback function 230governor->owner - .THIS_MODULE for the governor module (if 231 appropriate) 232 233The governor->governor callback is called with the current (or to-be-set) 234cpufreq_policy struct for that CPU, and an unsigned int event. The 235following events are currently defined: 236 237CPUFREQ_GOV_START: This governor shall start its duty for the CPU 238 policy->cpu 239CPUFREQ_GOV_STOP: This governor shall end its duty for the CPU 240 policy->cpu 241CPUFREQ_GOV_LIMITS: The limits for CPU policy->cpu have changed to 242 policy->min and policy->max. 243 244If you need other "events" externally of your driver, _only_ use the 245cpufreq_governor_l(unsigned int cpu, unsigned int event) call to the 246CPUfreq core to ensure proper locking. 247 248 249The CPUfreq governor may call the CPU processor driver using one of 250these two functions: 251 252int cpufreq_driver_target(struct cpufreq_policy *policy, 253 unsigned int target_freq, 254 unsigned int relation); 255 256int __cpufreq_driver_target(struct cpufreq_policy *policy, 257 unsigned int target_freq, 258 unsigned int relation); 259 260target_freq must be within policy->min and policy->max, of course. 261What's the difference between these two functions? When your governor 262still is in a direct code path of a call to governor->governor, the 263per-CPU cpufreq lock is still held in the cpufreq core, and there's 264no need to lock it again (in fact, this would cause a deadlock). So 265use __cpufreq_driver_target only in these cases. In all other cases 266(for example, when there's a "daemonized" function that wakes up 267every second), use cpufreq_driver_target to lock the cpufreq per-CPU 268lock before the command is passed to the cpufreq processor driver. 269 270