1 CPU frequency and voltage scaling code in the Linux(TM) kernel
4 L i n u x C P U F r e q
6 C P U F r e q G o v e r n o r s
8 - information for users and developers -
11 Dominik Brodowski <linux@brodo.de>
12 some additions and corrections by Nico Golde <nico@ngolde.de>
13 Rafael J. Wysocki <rafael.j.wysocki@intel.com>
14 Viresh Kumar <viresh.kumar@linaro.org>
18 Clock scaling allows you to change the clock speed of the CPUs on the
19 fly. This is a nice method to save battery power, because the lower
20 the clock speed, the less power the CPU consumes.
25 1. What is a CPUFreq Governor?
27 2. Governors In the Linux Kernel
35 3. The Governor Interface in the CPUfreq Core
40 1. What Is A CPUFreq Governor?
41 ==============================
43 Most cpufreq drivers (except the intel_pstate and longrun) or even most
44 cpu frequency scaling algorithms only allow the CPU frequency to be set
45 to predefined fixed values. In order to offer dynamic frequency
46 scaling, the cpufreq core must be able to tell these drivers of a
47 "target frequency". So these specific drivers will be transformed to
48 offer a "->target/target_index/fast_switch()" call instead of the
49 "->setpolicy()" call. For set_policy drivers, all stays the same,
52 How to decide what frequency within the CPUfreq policy should be used?
53 That's done using "cpufreq governors".
55 Basically, it's the following flow graph:
57 CPU can be set to switch independently | CPU can only be set
58 within specific "limits" | to specific frequencies
61 consists of frequency limits (policy->{min,max})
62 and CPUfreq governor to be used
65 / the cpufreq governor decides
66 / (dynamically or statically)
67 / what target_freq to set within
68 / the limits of policy->{min,max}
71 Using the ->setpolicy call, Using the ->target/target_index/fast_switch call,
72 the limits and the the frequency closest
73 "policy" is set. to target_freq is set.
75 is within policy->{min,max}
78 2. Governors In the Linux Kernel
79 ================================
84 The CPUfreq governor "performance" sets the CPU statically to the
85 highest frequency within the borders of scaling_min_freq and
92 The CPUfreq governor "powersave" sets the CPU statically to the
93 lowest frequency within the borders of scaling_min_freq and
100 The CPUfreq governor "userspace" allows the user, or any userspace
101 program running with UID "root", to set the CPU to a specific frequency
102 by making a sysfs file "scaling_setspeed" available in the CPU-device
109 The CPUfreq governor "ondemand" sets the CPU frequency depending on the
110 current system load. Load estimation is triggered by the scheduler
111 through the update_util_data->func hook; when triggered, cpufreq checks
112 the CPU-usage statistics over the last period and the governor sets the
113 CPU accordingly. The CPU must have the capability to switch the
114 frequency very quickly.
120 Measured in uS (10^-6 seconds), this is how often you want the kernel
121 to look at the CPU usage and to make decisions on what to do about the
122 frequency. Typically this is set to values of around '10000' or more.
123 It's default value is (cmp. with users-guide.txt): transition_latency
124 * 1000. Be aware that transition latency is in ns and sampling_rate
125 is in us, so you get the same sysfs value by default. Sampling rate
126 should always get adjusted considering the transition latency to set
127 the sampling rate 750 times as high as the transition latency in the
128 bash (as said, 1000 is default), do:
130 $ echo `$(($(cat cpuinfo_transition_latency) * 750 / 1000)) > ondemand/sampling_rate
134 The sampling rate is limited by the HW transition latency:
135 transition_latency * 100
137 Or by kernel restrictions:
138 - If CONFIG_NO_HZ_COMMON is set, the limit is 10ms fixed.
139 - If CONFIG_NO_HZ_COMMON is not set or nohz=off boot parameter is
140 used, the limits depend on the CONFIG_HZ option:
141 HZ=1000: min=20000us (20ms)
142 HZ=250: min=80000us (80ms)
143 HZ=100: min=200000us (200ms)
145 The highest value of kernel and HW latency restrictions is shown and
146 used as the minimum sampling rate.
150 This defines what the average CPU usage between the samplings of
151 'sampling_rate' needs to be for the kernel to make a decision on
152 whether it should increase the frequency. For example when it is set
153 to its default value of '95' it means that between the checking
154 intervals the CPU needs to be on average more than 95% in use to then
155 decide that the CPU frequency needs to be increased.
159 This parameter takes a value of '0' or '1'. When set to '0' (its
160 default), all processes are counted towards the 'cpu utilisation'
161 value. When set to '1', the processes that are run with a 'nice'
162 value will not count (and thus be ignored) in the overall usage
163 calculation. This is useful if you are running a CPU intensive
164 calculation on your laptop that you do not care how long it takes to
165 complete as you can 'nice' it and prevent it from taking part in the
166 deciding process of whether to increase your CPU frequency.
168 * sampling_down_factor:
170 This parameter controls the rate at which the kernel makes a decision
171 on when to decrease the frequency while running at top speed. When set
172 to 1 (the default) decisions to reevaluate load are made at the same
173 interval regardless of current clock speed. But when set to greater
174 than 1 (e.g. 100) it acts as a multiplier for the scheduling interval
175 for reevaluating load when the CPU is at its top speed due to high
176 load. This improves performance by reducing the overhead of load
177 evaluation and helping the CPU stay at its top speed when truly busy,
178 rather than shifting back and forth in speed. This tunable has no
179 effect on behavior at lower speeds/lower CPU loads.
183 This parameter takes a value between 0 to 1000. It defines the
184 percentage (times 10) value of the target frequency that will be
185 shaved off of the target. For example, when set to 100 -- 10%, when
186 ondemand governor would have targeted 1000 MHz, it will target
187 1000 MHz - (10% of 1000 MHz) = 900 MHz instead. This is set to 0
188 (disabled) by default.
190 When AMD frequency sensitivity powersave bias driver --
191 drivers/cpufreq/amd_freq_sensitivity.c is loaded, this parameter
192 defines the workload frequency sensitivity threshold in which a lower
193 frequency is chosen instead of ondemand governor's original target.
194 The frequency sensitivity is a hardware reported (on AMD Family 16h
195 Processors and above) value between 0 to 100% that tells software how
196 the performance of the workload running on a CPU will change when
197 frequency changes. A workload with sensitivity of 0% (memory/IO-bound)
198 will not perform any better on higher core frequency, whereas a
199 workload with sensitivity of 100% (CPU-bound) will perform better
200 higher the frequency. When the driver is loaded, this is set to 400 by
201 default -- for CPUs running workloads with sensitivity value below
202 40%, a lower frequency is chosen. Unloading the driver or writing 0
203 will disable this feature.
209 The CPUfreq governor "conservative", much like the "ondemand"
210 governor, sets the CPU frequency depending on the current usage. It
211 differs in behaviour in that it gracefully increases and decreases the
212 CPU speed rather than jumping to max speed the moment there is any load
213 on the CPU. This behaviour is more suitable in a battery powered
214 environment. The governor is tweaked in the same manner as the
215 "ondemand" governor through sysfs with the addition of:
219 This describes what percentage steps the cpu freq should be increased
220 and decreased smoothly by. By default the cpu frequency will increase
221 in 5% chunks of your maximum cpu frequency. You can change this value
222 to anywhere between 0 and 100 where '0' will effectively lock your CPU
223 at a speed regardless of its load whilst '100' will, in theory, make
224 it behave identically to the "ondemand" governor.
228 Same as the 'up_threshold' found for the "ondemand" governor but for
229 the opposite direction. For example when set to its default value of
230 '20' it means that if the CPU usage needs to be below 20% between
231 samples to have the frequency decreased.
233 * sampling_down_factor:
235 Similar functionality as in "ondemand" governor. But in
236 "conservative", it controls the rate at which the kernel makes a
237 decision on when to decrease the frequency while running in any speed.
238 Load for frequency increase is still evaluated every sampling rate.
244 The "schedutil" governor aims at better integration with the Linux
245 kernel scheduler. Load estimation is achieved through the scheduler's
246 Per-Entity Load Tracking (PELT) mechanism, which also provides
247 information about the recent load [1]. This governor currently does
248 load based DVFS only for tasks managed by CFS. RT and DL scheduler tasks
249 are always run at the highest frequency. Unlike all the other
250 governors, the code is located under the kernel/sched/ directory.
256 This contains a value in microseconds. The governor waits for
257 rate_limit_us time before reevaluating the load again, after it has
258 evaluated the load once.
260 For an in-depth comparison with the other governors refer to [2].
263 3. The Governor Interface in the CPUfreq Core
264 =============================================
266 A new governor must register itself with the CPUfreq core using
267 "cpufreq_register_governor". The struct cpufreq_governor, which has to
268 be passed to that function, must contain the following values:
270 governor->name - A unique name for this governor.
271 governor->owner - .THIS_MODULE for the governor module (if appropriate).
273 plus a set of hooks to the functions implementing the governor's logic.
275 The CPUfreq governor may call the CPU processor driver using one of
278 int cpufreq_driver_target(struct cpufreq_policy *policy,
279 unsigned int target_freq,
280 unsigned int relation);
282 int __cpufreq_driver_target(struct cpufreq_policy *policy,
283 unsigned int target_freq,
284 unsigned int relation);
286 target_freq must be within policy->min and policy->max, of course.
287 What's the difference between these two functions? When your governor is
288 in a direct code path of a call to governor callbacks, like
289 governor->start(), the policy->rwsem is still held in the cpufreq core,
290 and there's no need to lock it again (in fact, this would cause a
291 deadlock). So use __cpufreq_driver_target only in these cases. In all
292 other cases (for example, when there's a "daemonized" function that
293 wakes up every second), use cpufreq_driver_target to take policy->rwsem
294 before the command is passed to the cpufreq driver.
299 [1] Per-entity load tracking: https://lwn.net/Articles/531853/
300 [2] Improvements in CPU frequency management: https://lwn.net/Articles/682391/