1 // SPDX-License-Identifier: GPL-2.0
3 * A power allocator to manage temperature
5 * Copyright (C) 2014 ARM Ltd.
9 #define pr_fmt(fmt) "Power allocator: " fmt
11 #include <linux/rculist.h>
12 #include <linux/slab.h>
13 #include <linux/thermal.h>
15 #define CREATE_TRACE_POINTS
16 #include <trace/events/thermal_power_allocator.h>
18 #include "thermal_core.h"
20 #define INVALID_TRIP -1
23 #define int_to_frac(x) ((x) << FRAC_BITS)
24 #define frac_to_int(x) ((x) >> FRAC_BITS)
27 * mul_frac() - multiply two fixed-point numbers
28 * @x: first multiplicand
29 * @y: second multiplicand
31 * Return: the result of multiplying two fixed-point numbers. The
32 * result is also a fixed-point number.
34 static inline s64 mul_frac(s64 x, s64 y)
36 return (x * y) >> FRAC_BITS;
40 * div_frac() - divide two fixed-point numbers
44 * Return: the result of dividing two fixed-point numbers. The
45 * result is also a fixed-point number.
47 static inline s64 div_frac(s64 x, s64 y)
49 return div_s64(x << FRAC_BITS, y);
53 * struct power_allocator_params - parameters for the power allocator governor
54 * @allocated_tzp: whether we have allocated tzp for this thermal zone and
55 * it needs to be freed on unbind
56 * @err_integral: accumulated error in the PID controller.
57 * @prev_err: error in the previous iteration of the PID controller.
58 * Used to calculate the derivative term.
59 * @trip_switch_on: first passive trip point of the thermal zone. The
60 * governor switches on when this trip point is crossed.
61 * If the thermal zone only has one passive trip point,
62 * @trip_switch_on should be INVALID_TRIP.
63 * @trip_max_desired_temperature: last passive trip point of the thermal
64 * zone. The temperature we are
67 struct power_allocator_params {
72 int trip_max_desired_temperature;
76 * estimate_sustainable_power() - Estimate the sustainable power of a thermal zone
77 * @tz: thermal zone we are operating in
79 * For thermal zones that don't provide a sustainable_power in their
80 * thermal_zone_params, estimate one. Calculate it using the minimum
81 * power of all the cooling devices as that gives a valid value that
82 * can give some degree of functionality. For optimal performance of
83 * this governor, provide a sustainable_power in the thermal zone's
84 * thermal_zone_params.
86 static u32 estimate_sustainable_power(struct thermal_zone_device *tz)
88 u32 sustainable_power = 0;
89 struct thermal_instance *instance;
90 struct power_allocator_params *params = tz->governor_data;
92 list_for_each_entry(instance, &tz->thermal_instances, tz_node) {
93 struct thermal_cooling_device *cdev = instance->cdev;
96 if (instance->trip != params->trip_max_desired_temperature)
99 if (!cdev_is_power_actor(cdev))
102 if (cdev->ops->state2power(cdev, instance->upper, &min_power))
105 sustainable_power += min_power;
108 return sustainable_power;
112 * estimate_pid_constants() - Estimate the constants for the PID controller
113 * @tz: thermal zone for which to estimate the constants
114 * @sustainable_power: sustainable power for the thermal zone
115 * @trip_switch_on: trip point number for the switch on temperature
116 * @control_temp: target temperature for the power allocator governor
117 * @force: whether to force the update of the constants
119 * This function is used to update the estimation of the PID
120 * controller constants in struct thermal_zone_parameters.
121 * Sustainable power is provided in case it was estimated. The
122 * estimated sustainable_power should not be stored in the
123 * thermal_zone_parameters so it has to be passed explicitly to this
126 * If @force is not set, the values in the thermal zone's parameters
127 * are preserved if they are not zero. If @force is set, the values
128 * in thermal zone's parameters are overwritten.
130 static void estimate_pid_constants(struct thermal_zone_device *tz,
131 u32 sustainable_power, int trip_switch_on,
132 int control_temp, bool force)
136 u32 temperature_threshold;
138 ret = tz->ops->get_trip_temp(tz, trip_switch_on, &switch_on_temp);
142 temperature_threshold = control_temp - switch_on_temp;
144 * estimate_pid_constants() tries to find appropriate default
145 * values for thermal zones that don't provide them. If a
146 * system integrator has configured a thermal zone with two
147 * passive trip points at the same temperature, that person
148 * hasn't put any effort to set up the thermal zone properly
151 if (!temperature_threshold)
154 if (!tz->tzp->k_po || force)
155 tz->tzp->k_po = int_to_frac(sustainable_power) /
156 temperature_threshold;
158 if (!tz->tzp->k_pu || force)
159 tz->tzp->k_pu = int_to_frac(2 * sustainable_power) /
160 temperature_threshold;
162 if (!tz->tzp->k_i || force)
163 tz->tzp->k_i = int_to_frac(10) / 1000;
165 * The default for k_d and integral_cutoff is 0, so we can
166 * leave them as they are.
171 * pid_controller() - PID controller
172 * @tz: thermal zone we are operating in
173 * @control_temp: the target temperature in millicelsius
174 * @max_allocatable_power: maximum allocatable power for this thermal zone
176 * This PID controller increases the available power budget so that the
177 * temperature of the thermal zone gets as close as possible to
178 * @control_temp and limits the power if it exceeds it. k_po is the
179 * proportional term when we are overshooting, k_pu is the
180 * proportional term when we are undershooting. integral_cutoff is a
181 * threshold below which we stop accumulating the error. The
182 * accumulated error is only valid if the requested power will make
183 * the system warmer. If the system is mostly idle, there's no point
184 * in accumulating positive error.
186 * Return: The power budget for the next period.
188 static u32 pid_controller(struct thermal_zone_device *tz,
190 u32 max_allocatable_power)
192 s64 p, i, d, power_range;
193 s32 err, max_power_frac;
194 u32 sustainable_power;
195 struct power_allocator_params *params = tz->governor_data;
197 max_power_frac = int_to_frac(max_allocatable_power);
199 if (tz->tzp->sustainable_power) {
200 sustainable_power = tz->tzp->sustainable_power;
202 sustainable_power = estimate_sustainable_power(tz);
203 estimate_pid_constants(tz, sustainable_power,
204 params->trip_switch_on, control_temp,
208 err = control_temp - tz->temperature;
209 err = int_to_frac(err);
211 /* Calculate the proportional term */
212 p = mul_frac(err < 0 ? tz->tzp->k_po : tz->tzp->k_pu, err);
215 * Calculate the integral term
217 * if the error is less than cut off allow integration (but
218 * the integral is limited to max power)
220 i = mul_frac(tz->tzp->k_i, params->err_integral);
222 if (err < int_to_frac(tz->tzp->integral_cutoff)) {
223 s64 i_next = i + mul_frac(tz->tzp->k_i, err);
225 if (abs(i_next) < max_power_frac) {
227 params->err_integral += err;
232 * Calculate the derivative term
234 * We do err - prev_err, so with a positive k_d, a decreasing
235 * error (i.e. driving closer to the line) results in less
236 * power being applied, slowing down the controller)
238 d = mul_frac(tz->tzp->k_d, err - params->prev_err);
239 d = div_frac(d, tz->passive_delay);
240 params->prev_err = err;
242 power_range = p + i + d;
244 /* feed-forward the known sustainable dissipatable power */
245 power_range = sustainable_power + frac_to_int(power_range);
247 power_range = clamp(power_range, (s64)0, (s64)max_allocatable_power);
249 trace_thermal_power_allocator_pid(tz, frac_to_int(err),
250 frac_to_int(params->err_integral),
251 frac_to_int(p), frac_to_int(i),
252 frac_to_int(d), power_range);
258 * power_actor_set_power() - limit the maximum power a cooling device consumes
259 * @cdev: pointer to &thermal_cooling_device
260 * @instance: thermal instance to update
261 * @power: the power in milliwatts
263 * Set the cooling device to consume at most @power milliwatts. The limit is
264 * expected to be a cap at the maximum power consumption.
266 * Return: 0 on success, -EINVAL if the cooling device does not
267 * implement the power actor API or -E* for other failures.
270 power_actor_set_power(struct thermal_cooling_device *cdev,
271 struct thermal_instance *instance, u32 power)
276 ret = cdev->ops->power2state(cdev, power, &state);
280 instance->target = clamp_val(state, instance->lower, instance->upper);
281 mutex_lock(&cdev->lock);
282 cdev->updated = false;
283 mutex_unlock(&cdev->lock);
284 thermal_cdev_update(cdev);
290 * divvy_up_power() - divvy the allocated power between the actors
291 * @req_power: each actor's requested power
292 * @max_power: each actor's maximum available power
293 * @num_actors: size of the @req_power, @max_power and @granted_power's array
294 * @total_req_power: sum of @req_power
295 * @power_range: total allocated power
296 * @granted_power: output array: each actor's granted power
297 * @extra_actor_power: an appropriately sized array to be used in the
298 * function as temporary storage of the extra power given
301 * This function divides the total allocated power (@power_range)
302 * fairly between the actors. It first tries to give each actor a
303 * share of the @power_range according to how much power it requested
304 * compared to the rest of the actors. For example, if only one actor
305 * requests power, then it receives all the @power_range. If
306 * three actors each requests 1mW, each receives a third of the
309 * If any actor received more than their maximum power, then that
310 * surplus is re-divvied among the actors based on how far they are
311 * from their respective maximums.
313 * Granted power for each actor is written to @granted_power, which
314 * should've been allocated by the calling function.
316 static void divvy_up_power(u32 *req_power, u32 *max_power, int num_actors,
317 u32 total_req_power, u32 power_range,
318 u32 *granted_power, u32 *extra_actor_power)
320 u32 extra_power, capped_extra_power;
324 * Prevent division by 0 if none of the actors request power.
326 if (!total_req_power)
329 capped_extra_power = 0;
331 for (i = 0; i < num_actors; i++) {
332 u64 req_range = (u64)req_power[i] * power_range;
334 granted_power[i] = DIV_ROUND_CLOSEST_ULL(req_range,
337 if (granted_power[i] > max_power[i]) {
338 extra_power += granted_power[i] - max_power[i];
339 granted_power[i] = max_power[i];
342 extra_actor_power[i] = max_power[i] - granted_power[i];
343 capped_extra_power += extra_actor_power[i];
350 * Re-divvy the reclaimed extra among actors based on
351 * how far they are from the max
353 extra_power = min(extra_power, capped_extra_power);
354 if (capped_extra_power > 0)
355 for (i = 0; i < num_actors; i++)
356 granted_power[i] += (extra_actor_power[i] *
357 extra_power) / capped_extra_power;
360 static int allocate_power(struct thermal_zone_device *tz,
363 struct thermal_instance *instance;
364 struct power_allocator_params *params = tz->governor_data;
365 u32 *req_power, *max_power, *granted_power, *extra_actor_power;
366 u32 *weighted_req_power;
367 u32 total_req_power, max_allocatable_power, total_weighted_req_power;
368 u32 total_granted_power, power_range;
369 int i, num_actors, total_weight, ret = 0;
370 int trip_max_desired_temperature = params->trip_max_desired_temperature;
372 mutex_lock(&tz->lock);
376 list_for_each_entry(instance, &tz->thermal_instances, tz_node) {
377 if ((instance->trip == trip_max_desired_temperature) &&
378 cdev_is_power_actor(instance->cdev)) {
380 total_weight += instance->weight;
390 * We need to allocate five arrays of the same size:
391 * req_power, max_power, granted_power, extra_actor_power and
392 * weighted_req_power. They are going to be needed until this
393 * function returns. Allocate them all in one go to simplify
394 * the allocation and deallocation logic.
396 BUILD_BUG_ON(sizeof(*req_power) != sizeof(*max_power));
397 BUILD_BUG_ON(sizeof(*req_power) != sizeof(*granted_power));
398 BUILD_BUG_ON(sizeof(*req_power) != sizeof(*extra_actor_power));
399 BUILD_BUG_ON(sizeof(*req_power) != sizeof(*weighted_req_power));
400 req_power = kcalloc(num_actors * 5, sizeof(*req_power), GFP_KERNEL);
406 max_power = &req_power[num_actors];
407 granted_power = &req_power[2 * num_actors];
408 extra_actor_power = &req_power[3 * num_actors];
409 weighted_req_power = &req_power[4 * num_actors];
412 total_weighted_req_power = 0;
414 max_allocatable_power = 0;
416 list_for_each_entry(instance, &tz->thermal_instances, tz_node) {
418 struct thermal_cooling_device *cdev = instance->cdev;
420 if (instance->trip != trip_max_desired_temperature)
423 if (!cdev_is_power_actor(cdev))
426 if (cdev->ops->get_requested_power(cdev, &req_power[i]))
430 weight = 1 << FRAC_BITS;
432 weight = instance->weight;
434 weighted_req_power[i] = frac_to_int(weight * req_power[i]);
436 if (cdev->ops->state2power(cdev, instance->lower,
440 total_req_power += req_power[i];
441 max_allocatable_power += max_power[i];
442 total_weighted_req_power += weighted_req_power[i];
447 power_range = pid_controller(tz, control_temp, max_allocatable_power);
449 divvy_up_power(weighted_req_power, max_power, num_actors,
450 total_weighted_req_power, power_range, granted_power,
453 total_granted_power = 0;
455 list_for_each_entry(instance, &tz->thermal_instances, tz_node) {
456 if (instance->trip != trip_max_desired_temperature)
459 if (!cdev_is_power_actor(instance->cdev))
462 power_actor_set_power(instance->cdev, instance,
464 total_granted_power += granted_power[i];
469 trace_thermal_power_allocator(tz, req_power, total_req_power,
470 granted_power, total_granted_power,
471 num_actors, power_range,
472 max_allocatable_power, tz->temperature,
473 control_temp - tz->temperature);
477 mutex_unlock(&tz->lock);
483 * get_governor_trips() - get the number of the two trip points that are key for this governor
484 * @tz: thermal zone to operate on
485 * @params: pointer to private data for this governor
487 * The power allocator governor works optimally with two trips points:
488 * a "switch on" trip point and a "maximum desired temperature". These
489 * are defined as the first and last passive trip points.
491 * If there is only one trip point, then that's considered to be the
492 * "maximum desired temperature" trip point and the governor is always
493 * on. If there are no passive or active trip points, then the
494 * governor won't do anything. In fact, its throttle function
495 * won't be called at all.
497 static void get_governor_trips(struct thermal_zone_device *tz,
498 struct power_allocator_params *params)
500 int i, last_active, last_passive;
501 bool found_first_passive;
503 found_first_passive = false;
504 last_active = INVALID_TRIP;
505 last_passive = INVALID_TRIP;
507 for (i = 0; i < tz->trips; i++) {
508 enum thermal_trip_type type;
511 ret = tz->ops->get_trip_type(tz, i, &type);
513 dev_warn(&tz->device,
514 "Failed to get trip point %d type: %d\n", i,
519 if (type == THERMAL_TRIP_PASSIVE) {
520 if (!found_first_passive) {
521 params->trip_switch_on = i;
522 found_first_passive = true;
526 } else if (type == THERMAL_TRIP_ACTIVE) {
533 if (last_passive != INVALID_TRIP) {
534 params->trip_max_desired_temperature = last_passive;
535 } else if (found_first_passive) {
536 params->trip_max_desired_temperature = params->trip_switch_on;
537 params->trip_switch_on = INVALID_TRIP;
539 params->trip_switch_on = INVALID_TRIP;
540 params->trip_max_desired_temperature = last_active;
544 static void reset_pid_controller(struct power_allocator_params *params)
546 params->err_integral = 0;
547 params->prev_err = 0;
550 static void allow_maximum_power(struct thermal_zone_device *tz)
552 struct thermal_instance *instance;
553 struct power_allocator_params *params = tz->governor_data;
555 mutex_lock(&tz->lock);
556 list_for_each_entry(instance, &tz->thermal_instances, tz_node) {
557 if ((instance->trip != params->trip_max_desired_temperature) ||
558 (!cdev_is_power_actor(instance->cdev)))
561 instance->target = 0;
562 mutex_lock(&instance->cdev->lock);
563 instance->cdev->updated = false;
564 mutex_unlock(&instance->cdev->lock);
565 thermal_cdev_update(instance->cdev);
567 mutex_unlock(&tz->lock);
571 * power_allocator_bind() - bind the power_allocator governor to a thermal zone
572 * @tz: thermal zone to bind it to
574 * Initialize the PID controller parameters and bind it to the thermal
577 * Return: 0 on success, or -ENOMEM if we ran out of memory.
579 static int power_allocator_bind(struct thermal_zone_device *tz)
582 struct power_allocator_params *params;
585 params = kzalloc(sizeof(*params), GFP_KERNEL);
590 tz->tzp = kzalloc(sizeof(*tz->tzp), GFP_KERNEL);
596 params->allocated_tzp = true;
599 if (!tz->tzp->sustainable_power)
600 dev_warn(&tz->device, "power_allocator: sustainable_power will be estimated\n");
602 get_governor_trips(tz, params);
605 ret = tz->ops->get_trip_temp(tz,
606 params->trip_max_desired_temperature,
609 estimate_pid_constants(tz, tz->tzp->sustainable_power,
610 params->trip_switch_on,
611 control_temp, false);
614 reset_pid_controller(params);
616 tz->governor_data = params;
626 static void power_allocator_unbind(struct thermal_zone_device *tz)
628 struct power_allocator_params *params = tz->governor_data;
630 dev_dbg(&tz->device, "Unbinding from thermal zone %d\n", tz->id);
632 if (params->allocated_tzp) {
637 kfree(tz->governor_data);
638 tz->governor_data = NULL;
641 static int power_allocator_throttle(struct thermal_zone_device *tz, int trip)
644 int switch_on_temp, control_temp;
645 struct power_allocator_params *params = tz->governor_data;
648 * We get called for every trip point but we only need to do
649 * our calculations once
651 if (trip != params->trip_max_desired_temperature)
654 ret = tz->ops->get_trip_temp(tz, params->trip_switch_on,
656 if (!ret && (tz->temperature < switch_on_temp)) {
658 reset_pid_controller(params);
659 allow_maximum_power(tz);
665 ret = tz->ops->get_trip_temp(tz, params->trip_max_desired_temperature,
668 dev_warn(&tz->device,
669 "Failed to get the maximum desired temperature: %d\n",
674 return allocate_power(tz, control_temp);
677 static struct thermal_governor thermal_gov_power_allocator = {
678 .name = "power_allocator",
679 .bind_to_tz = power_allocator_bind,
680 .unbind_from_tz = power_allocator_unbind,
681 .throttle = power_allocator_throttle,
683 THERMAL_GOVERNOR_DECLARE(thermal_gov_power_allocator);