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>
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.
23 1. What is a CPUFreq Governor?
25 2. Governors In the Linux Kernel
33 3. The Governor Interface in the CPUfreq Core
37 1. What Is A CPUFreq Governor?
38 ==============================
40 Most cpufreq drivers (except the intel_pstate and longrun) or even most
41 cpu frequency scaling algorithms only offer the CPU to be set to one
42 frequency. In order to offer dynamic frequency scaling, the cpufreq
43 core must be able to tell these drivers of a "target frequency". So
44 these specific drivers will be transformed to offer a "->target/target_index"
45 call instead of the existing "->setpolicy" call. For "longrun", all
46 stays the same, though.
48 How to decide what frequency within the CPUfreq policy should be used?
49 That's done using "cpufreq governors". Two are already in this patch
50 -- they're the already existing "powersave" and "performance" which
51 set the frequency statically to the lowest or highest frequency,
52 respectively. At least two more such governors will be ready for
53 addition in the near future, but likely many more as there are various
54 different theories and models about dynamic frequency scaling
55 around. Using such a generic interface as cpufreq offers to scaling
56 governors, these can be tested extensively, and the best one can be
57 selected for each specific use.
59 Basically, it's the following flow graph:
61 CPU can be set to switch independently | CPU can only be set
62 within specific "limits" | to specific frequencies
65 consists of frequency limits (policy->{min,max})
66 and CPUfreq governor to be used
69 / the cpufreq governor decides
70 / (dynamically or statically)
71 / what target_freq to set within
72 / the limits of policy->{min,max}
75 Using the ->setpolicy call, Using the ->target/target_index call,
76 the limits and the the frequency closest
77 "policy" is set. to target_freq is set.
79 is within policy->{min,max}
82 2. Governors In the Linux Kernel
83 ================================
88 The CPUfreq governor "performance" sets the CPU statically to the
89 highest frequency within the borders of scaling_min_freq and
96 The CPUfreq governor "powersave" sets the CPU statically to the
97 lowest frequency within the borders of scaling_min_freq and
104 The CPUfreq governor "userspace" allows the user, or any userspace
105 program running with UID "root", to set the CPU to a specific frequency
106 by making a sysfs file "scaling_setspeed" available in the CPU-device
113 The CPUfreq governor "ondemand" sets the CPU depending on the
114 current usage. To do this the CPU must have the capability to
115 switch the frequency very quickly. There are a number of sysfs file
116 accessible parameters:
118 sampling_rate: measured in uS (10^-6 seconds), this is how often you
119 want the kernel to look at the CPU usage and to make decisions on
120 what to do about the frequency. Typically this is set to values of
121 around '10000' or more. It's default value is (cmp. with users-guide.txt):
122 transition_latency * 1000
123 Be aware that transition latency is in ns and sampling_rate is in us, so you
124 get the same sysfs value by default.
125 Sampling rate should always get adjusted considering the transition latency
126 To set the sampling rate 750 times as high as the transition latency
127 in the bash (as said, 1000 is default), do:
128 echo `$(($(cat cpuinfo_transition_latency) * 750 / 1000)) \
129 >ondemand/sampling_rate
132 The sampling rate is limited by the HW transition latency:
133 transition_latency * 100
134 Or by kernel restrictions:
135 If CONFIG_NO_HZ_COMMON is set, the limit is 10ms fixed.
136 If CONFIG_NO_HZ_COMMON is not set or nohz=off boot parameter is used, the
137 limits depend on the CONFIG_HZ option:
138 HZ=1000: min=20000us (20ms)
139 HZ=250: min=80000us (80ms)
140 HZ=100: min=200000us (200ms)
141 The highest value of kernel and HW latency restrictions is shown and
142 used as the minimum sampling rate.
144 up_threshold: defines what the average CPU usage between the samplings
145 of 'sampling_rate' needs to be for the kernel to make a decision on
146 whether it should increase the frequency. For example when it is set
147 to its default value of '95' it means that between the checking
148 intervals the CPU needs to be on average more than 95% in use to then
149 decide that the CPU frequency needs to be increased.
151 ignore_nice_load: this parameter takes a value of '0' or '1'. When
152 set to '0' (its default), all processes are counted towards the
153 'cpu utilisation' value. When set to '1', the processes that are
154 run with a 'nice' value will not count (and thus be ignored) in the
155 overall usage calculation. This is useful if you are running a CPU
156 intensive calculation on your laptop that you do not care how long it
157 takes to complete as you can 'nice' it and prevent it from taking part
158 in the deciding process of whether to increase your CPU frequency.
160 sampling_down_factor: this parameter controls the rate at which the
161 kernel makes a decision on when to decrease the frequency while running
162 at top speed. When set to 1 (the default) decisions to reevaluate load
163 are made at the same interval regardless of current clock speed. But
164 when set to greater than 1 (e.g. 100) it acts as a multiplier for the
165 scheduling interval for reevaluating load when the CPU is at its top
166 speed due to high load. This improves performance by reducing the overhead
167 of load evaluation and helping the CPU stay at its top speed when truly
168 busy, rather than shifting back and forth in speed. This tunable has no
169 effect on behavior at lower speeds/lower CPU loads.
171 powersave_bias: this parameter takes a value between 0 to 1000. It
172 defines the percentage (times 10) value of the target frequency that
173 will be shaved off of the target. For example, when set to 100 -- 10%,
174 when ondemand governor would have targeted 1000 MHz, it will target
175 1000 MHz - (10% of 1000 MHz) = 900 MHz instead. This is set to 0
176 (disabled) by default.
177 When AMD frequency sensitivity powersave bias driver --
178 drivers/cpufreq/amd_freq_sensitivity.c is loaded, this parameter
179 defines the workload frequency sensitivity threshold in which a lower
180 frequency is chosen instead of ondemand governor's original target.
181 The frequency sensitivity is a hardware reported (on AMD Family 16h
182 Processors and above) value between 0 to 100% that tells software how
183 the performance of the workload running on a CPU will change when
184 frequency changes. A workload with sensitivity of 0% (memory/IO-bound)
185 will not perform any better on higher core frequency, whereas a
186 workload with sensitivity of 100% (CPU-bound) will perform better
187 higher the frequency. When the driver is loaded, this is set to 400
188 by default -- for CPUs running workloads with sensitivity value below
189 40%, a lower frequency is chosen. Unloading the driver or writing 0
190 will disable this feature.
196 The CPUfreq governor "conservative", much like the "ondemand"
197 governor, sets the CPU depending on the current usage. It differs in
198 behaviour in that it gracefully increases and decreases the CPU speed
199 rather than jumping to max speed the moment there is any load on the
200 CPU. This behaviour more suitable in a battery powered environment.
201 The governor is tweaked in the same manner as the "ondemand" governor
202 through sysfs with the addition of:
204 freq_step: this describes what percentage steps the cpu freq should be
205 increased and decreased smoothly by. By default the cpu frequency will
206 increase in 5% chunks of your maximum cpu frequency. You can change this
207 value to anywhere between 0 and 100 where '0' will effectively lock your
208 CPU at a speed regardless of its load whilst '100' will, in theory, make
209 it behave identically to the "ondemand" governor.
211 down_threshold: same as the 'up_threshold' found for the "ondemand"
212 governor but for the opposite direction. For example when set to its
213 default value of '20' it means that if the CPU usage needs to be below
214 20% between samples to have the frequency decreased.
216 sampling_down_factor: similar functionality as in "ondemand" governor.
217 But in "conservative", it controls the rate at which the kernel makes
218 a decision on when to decrease the frequency while running in any
219 speed. Load for frequency increase is still evaluated every
225 The CPUfreq governor "interactive" is designed for latency-sensitive,
226 interactive workloads. This governor sets the CPU speed depending on
227 usage, similar to "ondemand" and "conservative" governors, but with a
228 different set of configurable behaviors.
230 The tuneable values for this governor are:
232 target_loads: CPU load values used to adjust speed to influence the
233 current CPU load toward that value. In general, the lower the target
234 load, the more often the governor will raise CPU speeds to bring load
235 below the target. The format is a single target load, optionally
236 followed by pairs of CPU speeds and CPU loads to target at or above
237 those speeds. Colons can be used between the speeds and associated
238 target loads for readability. For example:
240 85 1000000:90 1700000:99
242 targets CPU load 85% below speed 1GHz, 90% at or above 1GHz, until
243 1.7GHz and above, at which load 99% is targeted. If speeds are
244 specified these must appear in ascending order. Higher target load
245 values are typically specified for higher speeds, that is, target load
246 values also usually appear in an ascending order. The default is
247 target load 90% for all speeds.
249 min_sample_time: The minimum amount of time to spend at the current
250 frequency before ramping down. Default is 80000 uS.
252 hispeed_freq: An intermediate "hi speed" at which to initially ramp
253 when CPU load hits the value specified in go_hispeed_load. If load
254 stays high for the amount of time specified in above_hispeed_delay,
255 then speed may be bumped higher. Default is the maximum speed
256 allowed by the policy at governor initialization time.
258 go_hispeed_load: The CPU load at which to ramp to hispeed_freq.
261 above_hispeed_delay: When speed is at or above hispeed_freq, wait for
262 this long before raising speed in response to continued high load.
263 The format is a single delay value, optionally followed by pairs of
264 CPU speeds and the delay to use at or above those speeds. Colons can
265 be used between the speeds and associated delays for readability. For
268 80000 1300000:200000 1500000:40000
270 uses delay 80000 uS until CPU speed 1.3 GHz, at which speed delay
271 200000 uS is used until speed 1.5 GHz, at which speed (and above)
272 delay 40000 uS is used. If speeds are specified these must appear in
273 ascending order. Default is 20000 uS.
275 timer_rate: Sample rate for reevaluating CPU load when the CPU is not
276 idle. A deferrable timer is used, such that the CPU will not be woken
277 from idle to service this timer until something else needs to run.
278 (The maximum time to allow deferring this timer when not running at
279 minimum speed is configurable via timer_slack.) Default is 20000 uS.
281 timer_slack: Maximum additional time to defer handling the governor
282 sampling timer beyond timer_rate when running at speeds above the
283 minimum. For platforms that consume additional power at idle when
284 CPUs are running at speeds greater than minimum, this places an upper
285 bound on how long the timer will be deferred prior to re-evaluating
286 load and dropping speed. For example, if timer_rate is 20000uS and
287 timer_slack is 10000uS then timers will be deferred for up to 30msec
288 when not at lowest speed. A value of -1 means defer timers
289 indefinitely at all speeds. Default is 80000 uS.
291 boost: If non-zero, immediately boost speed of all CPUs to at least
292 hispeed_freq until zero is written to this attribute. If zero, allow
293 CPU speeds to drop below hispeed_freq according to load as usual.
296 boostpulse: On each write, immediately boost speed of all CPUs to
297 hispeed_freq for at least the period of time specified by
298 boostpulse_duration, after which speeds are allowed to drop below
299 hispeed_freq according to load as usual.
301 boostpulse_duration: Length of time to hold CPU speed at hispeed_freq
302 on a write to boostpulse, before allowing speed to drop according to
303 load as usual. Default is 80000 uS.
305 align_windows: If non-zero, align governor timer window to fire at
306 multiples of number of jiffies timer_rate converts to.
308 use_sched_load: If non-zero, query scheduler for CPU busy time,
309 instead of collecting it directly in governor. This would allow
310 scheduler to adjust the busy time of each CPU to account for known
311 information such as migration. If non-zero, this also implies governor
312 sampling windows are aligned across CPUs, with same timer_rate,
313 regardless what align_windows is set to. Default is zero.
315 use_migration_notif: If non-zero, schedule hrtimer to fire in 1ms
316 to reevaluate frequency of notified CPU, unless the hrtimer is already
317 pending. If zero, ignore scheduler notification. Default is zero.
319 max_freq_hysteresis: Each time freq evaluation chooses policy->max,
320 next max_freq_hysteresis is considered as hysteresis period. During
321 this period, frequency target will not drop below hispeed_freq, no
322 matter how light actual workload is. If CPU load of any sampling
323 window exceeds go_hispeed_load during this period, governor will
324 directly increase frequency back to policy->max. Default is 0 uS.
326 ignore_hispeed_on_notif: If non-zero, do not apply hispeed related
327 logic if frequency evaluation is triggered by scheduler notification.
328 This includes ignoring go_hispeed_load, hispeed_freq in frequency
329 selection, and ignoring above_hispeed_delay that prevents frequency
330 ramp up. For evaluation triggered by timer, hispeed logic is still
331 always applied. ignore_hispeed_on_notif has no effect if
332 use_migration_notif is set to zero. Default is zero.
334 fast_ramp_down: If non-zero, do not apply min_sample_time if
335 frequency evaluation is triggered by scheduler notification. For
336 evaluation triggered by timer, min_sample_time is still always
337 enforced. fast_ramp_down has no effect if use_migration_notif is
338 set to zero. Default is zero.
340 enable_prediction: If non-zero, two frequencies will be calculated
341 during each sampling period: one based on busy time in previous sampling
342 period (f_prev), and the other based on prediction provided by scheduler
343 (f_pred). Max of both will be selected as final frequency. Hispeed
344 related logic, including both frequency selection and delay is ignored
345 if enable_prediction is set. If only f_pred but not f_prev picked
346 policy->max, max_freq_hysteresis period is not started/extended.
347 use_sched_load must be turned on before enabling this feature.
350 3. The Governor Interface in the CPUfreq Core
351 =============================================
353 A new governor must register itself with the CPUfreq core using
354 "cpufreq_register_governor". The struct cpufreq_governor, which has to
355 be passed to that function, must contain the following values:
357 governor->name - A unique name for this governor
358 governor->governor - The governor callback function
359 governor->owner - .THIS_MODULE for the governor module (if
362 The governor->governor callback is called with the current (or to-be-set)
363 cpufreq_policy struct for that CPU, and an unsigned int event. The
364 following events are currently defined:
366 CPUFREQ_GOV_START: This governor shall start its duty for the CPU
368 CPUFREQ_GOV_STOP: This governor shall end its duty for the CPU
370 CPUFREQ_GOV_LIMITS: The limits for CPU policy->cpu have changed to
371 policy->min and policy->max.
373 If you need other "events" externally of your driver, _only_ use the
374 cpufreq_governor_l(unsigned int cpu, unsigned int event) call to the
375 CPUfreq core to ensure proper locking.
378 The CPUfreq governor may call the CPU processor driver using one of
381 int cpufreq_driver_target(struct cpufreq_policy *policy,
382 unsigned int target_freq,
383 unsigned int relation);
385 int __cpufreq_driver_target(struct cpufreq_policy *policy,
386 unsigned int target_freq,
387 unsigned int relation);
389 target_freq must be within policy->min and policy->max, of course.
390 What's the difference between these two functions? When your governor
391 still is in a direct code path of a call to governor->governor, the
392 per-CPU cpufreq lock is still held in the cpufreq core, and there's
393 no need to lock it again (in fact, this would cause a deadlock). So
394 use __cpufreq_driver_target only in these cases. In all other cases
395 (for example, when there's a "daemonized" function that wakes up
396 every second), use cpufreq_driver_target to lock the cpufreq per-CPU
397 lock before the command is passed to the cpufreq processor driver.
OSDN Git Service
