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
66 * @sustainable_power: Sustainable power (heat) that this thermal zone can
69 struct power_allocator_params {
74 int trip_max_desired_temperature;
75 u32 sustainable_power;
79 * estimate_sustainable_power() - Estimate the sustainable power of a thermal zone
80 * @tz: thermal zone we are operating in
82 * For thermal zones that don't provide a sustainable_power in their
83 * thermal_zone_params, estimate one. Calculate it using the minimum
84 * power of all the cooling devices as that gives a valid value that
85 * can give some degree of functionality. For optimal performance of
86 * this governor, provide a sustainable_power in the thermal zone's
87 * thermal_zone_params.
89 static u32 estimate_sustainable_power(struct thermal_zone_device *tz)
91 u32 sustainable_power = 0;
92 struct thermal_instance *instance;
93 struct power_allocator_params *params = tz->governor_data;
95 list_for_each_entry(instance, &tz->thermal_instances, tz_node) {
96 struct thermal_cooling_device *cdev = instance->cdev;
99 if (instance->trip != params->trip_max_desired_temperature)
102 if (!cdev_is_power_actor(cdev))
105 if (cdev->ops->state2power(cdev, instance->upper, &min_power))
108 sustainable_power += min_power;
111 return sustainable_power;
115 * estimate_pid_constants() - Estimate the constants for the PID controller
116 * @tz: thermal zone for which to estimate the constants
117 * @sustainable_power: sustainable power for the thermal zone
118 * @trip_switch_on: trip point number for the switch on temperature
119 * @control_temp: target temperature for the power allocator governor
121 * This function is used to update the estimation of the PID
122 * controller constants in struct thermal_zone_parameters.
124 static void estimate_pid_constants(struct thermal_zone_device *tz,
125 u32 sustainable_power, int trip_switch_on,
130 u32 temperature_threshold;
133 ret = tz->ops->get_trip_temp(tz, trip_switch_on, &switch_on_temp);
137 temperature_threshold = control_temp - switch_on_temp;
139 * estimate_pid_constants() tries to find appropriate default
140 * values for thermal zones that don't provide them. If a
141 * system integrator has configured a thermal zone with two
142 * passive trip points at the same temperature, that person
143 * hasn't put any effort to set up the thermal zone properly
146 if (!temperature_threshold)
149 tz->tzp->k_po = int_to_frac(sustainable_power) /
150 temperature_threshold;
152 tz->tzp->k_pu = int_to_frac(2 * sustainable_power) /
153 temperature_threshold;
155 k_i = tz->tzp->k_pu / 10;
156 tz->tzp->k_i = k_i > 0 ? k_i : 1;
159 * The default for k_d and integral_cutoff is 0, so we can
160 * leave them as they are.
165 * get_sustainable_power() - Get the right sustainable power
166 * @tz: thermal zone for which to estimate the constants
167 * @params: parameters for the power allocator governor
168 * @control_temp: target temperature for the power allocator governor
170 * This function is used for getting the proper sustainable power value based
171 * on variables which might be updated by the user sysfs interface. If that
172 * happen the new value is going to be estimated and updated. It is also used
173 * after thermal zone binding, where the initial values where set to 0.
175 static u32 get_sustainable_power(struct thermal_zone_device *tz,
176 struct power_allocator_params *params,
179 u32 sustainable_power;
181 if (!tz->tzp->sustainable_power)
182 sustainable_power = estimate_sustainable_power(tz);
184 sustainable_power = tz->tzp->sustainable_power;
186 /* Check if it's init value 0 or there was update via sysfs */
187 if (sustainable_power != params->sustainable_power) {
188 estimate_pid_constants(tz, sustainable_power,
189 params->trip_switch_on, control_temp);
191 /* Do the estimation only once and make available in sysfs */
192 tz->tzp->sustainable_power = sustainable_power;
193 params->sustainable_power = sustainable_power;
196 return sustainable_power;
200 * pid_controller() - PID controller
201 * @tz: thermal zone we are operating in
202 * @control_temp: the target temperature in millicelsius
203 * @max_allocatable_power: maximum allocatable power for this thermal zone
205 * This PID controller increases the available power budget so that the
206 * temperature of the thermal zone gets as close as possible to
207 * @control_temp and limits the power if it exceeds it. k_po is the
208 * proportional term when we are overshooting, k_pu is the
209 * proportional term when we are undershooting. integral_cutoff is a
210 * threshold below which we stop accumulating the error. The
211 * accumulated error is only valid if the requested power will make
212 * the system warmer. If the system is mostly idle, there's no point
213 * in accumulating positive error.
215 * Return: The power budget for the next period.
217 static u32 pid_controller(struct thermal_zone_device *tz,
219 u32 max_allocatable_power)
221 s64 p, i, d, power_range;
222 s32 err, max_power_frac;
223 u32 sustainable_power;
224 struct power_allocator_params *params = tz->governor_data;
226 max_power_frac = int_to_frac(max_allocatable_power);
228 sustainable_power = get_sustainable_power(tz, params, control_temp);
230 err = control_temp - tz->temperature;
231 err = int_to_frac(err);
233 /* Calculate the proportional term */
234 p = mul_frac(err < 0 ? tz->tzp->k_po : tz->tzp->k_pu, err);
237 * Calculate the integral term
239 * if the error is less than cut off allow integration (but
240 * the integral is limited to max power)
242 i = mul_frac(tz->tzp->k_i, params->err_integral);
244 if (err < int_to_frac(tz->tzp->integral_cutoff)) {
245 s64 i_next = i + mul_frac(tz->tzp->k_i, err);
247 if (abs(i_next) < max_power_frac) {
249 params->err_integral += err;
254 * Calculate the derivative term
256 * We do err - prev_err, so with a positive k_d, a decreasing
257 * error (i.e. driving closer to the line) results in less
258 * power being applied, slowing down the controller)
260 d = mul_frac(tz->tzp->k_d, err - params->prev_err);
261 d = div_frac(d, jiffies_to_msecs(tz->passive_delay_jiffies));
262 params->prev_err = err;
264 power_range = p + i + d;
266 /* feed-forward the known sustainable dissipatable power */
267 power_range = sustainable_power + frac_to_int(power_range);
269 power_range = clamp(power_range, (s64)0, (s64)max_allocatable_power);
271 trace_thermal_power_allocator_pid(tz, frac_to_int(err),
272 frac_to_int(params->err_integral),
273 frac_to_int(p), frac_to_int(i),
274 frac_to_int(d), power_range);
280 * power_actor_set_power() - limit the maximum power a cooling device consumes
281 * @cdev: pointer to &thermal_cooling_device
282 * @instance: thermal instance to update
283 * @power: the power in milliwatts
285 * Set the cooling device to consume at most @power milliwatts. The limit is
286 * expected to be a cap at the maximum power consumption.
288 * Return: 0 on success, -EINVAL if the cooling device does not
289 * implement the power actor API or -E* for other failures.
292 power_actor_set_power(struct thermal_cooling_device *cdev,
293 struct thermal_instance *instance, u32 power)
298 ret = cdev->ops->power2state(cdev, power, &state);
302 instance->target = clamp_val(state, instance->lower, instance->upper);
303 mutex_lock(&cdev->lock);
304 __thermal_cdev_update(cdev);
305 mutex_unlock(&cdev->lock);
311 * divvy_up_power() - divvy the allocated power between the actors
312 * @req_power: each actor's requested power
313 * @max_power: each actor's maximum available power
314 * @num_actors: size of the @req_power, @max_power and @granted_power's array
315 * @total_req_power: sum of @req_power
316 * @power_range: total allocated power
317 * @granted_power: output array: each actor's granted power
318 * @extra_actor_power: an appropriately sized array to be used in the
319 * function as temporary storage of the extra power given
322 * This function divides the total allocated power (@power_range)
323 * fairly between the actors. It first tries to give each actor a
324 * share of the @power_range according to how much power it requested
325 * compared to the rest of the actors. For example, if only one actor
326 * requests power, then it receives all the @power_range. If
327 * three actors each requests 1mW, each receives a third of the
330 * If any actor received more than their maximum power, then that
331 * surplus is re-divvied among the actors based on how far they are
332 * from their respective maximums.
334 * Granted power for each actor is written to @granted_power, which
335 * should've been allocated by the calling function.
337 static void divvy_up_power(u32 *req_power, u32 *max_power, int num_actors,
338 u32 total_req_power, u32 power_range,
339 u32 *granted_power, u32 *extra_actor_power)
341 u32 extra_power, capped_extra_power;
345 * Prevent division by 0 if none of the actors request power.
347 if (!total_req_power)
350 capped_extra_power = 0;
352 for (i = 0; i < num_actors; i++) {
353 u64 req_range = (u64)req_power[i] * power_range;
355 granted_power[i] = DIV_ROUND_CLOSEST_ULL(req_range,
358 if (granted_power[i] > max_power[i]) {
359 extra_power += granted_power[i] - max_power[i];
360 granted_power[i] = max_power[i];
363 extra_actor_power[i] = max_power[i] - granted_power[i];
364 capped_extra_power += extra_actor_power[i];
371 * Re-divvy the reclaimed extra among actors based on
372 * how far they are from the max
374 extra_power = min(extra_power, capped_extra_power);
375 if (capped_extra_power > 0)
376 for (i = 0; i < num_actors; i++) {
377 u64 extra_range = (u64)extra_actor_power[i] * extra_power;
378 granted_power[i] += DIV_ROUND_CLOSEST_ULL(extra_range,
383 static int allocate_power(struct thermal_zone_device *tz,
386 struct thermal_instance *instance;
387 struct power_allocator_params *params = tz->governor_data;
388 u32 *req_power, *max_power, *granted_power, *extra_actor_power;
389 u32 *weighted_req_power;
390 u32 total_req_power, max_allocatable_power, total_weighted_req_power;
391 u32 total_granted_power, power_range;
392 int i, num_actors, total_weight, ret = 0;
393 int trip_max_desired_temperature = params->trip_max_desired_temperature;
397 list_for_each_entry(instance, &tz->thermal_instances, tz_node) {
398 if ((instance->trip == trip_max_desired_temperature) &&
399 cdev_is_power_actor(instance->cdev)) {
401 total_weight += instance->weight;
409 * We need to allocate five arrays of the same size:
410 * req_power, max_power, granted_power, extra_actor_power and
411 * weighted_req_power. They are going to be needed until this
412 * function returns. Allocate them all in one go to simplify
413 * the allocation and deallocation logic.
415 BUILD_BUG_ON(sizeof(*req_power) != sizeof(*max_power));
416 BUILD_BUG_ON(sizeof(*req_power) != sizeof(*granted_power));
417 BUILD_BUG_ON(sizeof(*req_power) != sizeof(*extra_actor_power));
418 BUILD_BUG_ON(sizeof(*req_power) != sizeof(*weighted_req_power));
419 req_power = kcalloc(num_actors * 5, sizeof(*req_power), GFP_KERNEL);
423 max_power = &req_power[num_actors];
424 granted_power = &req_power[2 * num_actors];
425 extra_actor_power = &req_power[3 * num_actors];
426 weighted_req_power = &req_power[4 * num_actors];
429 total_weighted_req_power = 0;
431 max_allocatable_power = 0;
433 list_for_each_entry(instance, &tz->thermal_instances, tz_node) {
435 struct thermal_cooling_device *cdev = instance->cdev;
437 if (instance->trip != trip_max_desired_temperature)
440 if (!cdev_is_power_actor(cdev))
443 if (cdev->ops->get_requested_power(cdev, &req_power[i]))
447 weight = 1 << FRAC_BITS;
449 weight = instance->weight;
451 weighted_req_power[i] = frac_to_int(weight * req_power[i]);
453 if (cdev->ops->state2power(cdev, instance->lower,
457 total_req_power += req_power[i];
458 max_allocatable_power += max_power[i];
459 total_weighted_req_power += weighted_req_power[i];
464 power_range = pid_controller(tz, control_temp, max_allocatable_power);
466 divvy_up_power(weighted_req_power, max_power, num_actors,
467 total_weighted_req_power, power_range, granted_power,
470 total_granted_power = 0;
472 list_for_each_entry(instance, &tz->thermal_instances, tz_node) {
473 if (instance->trip != trip_max_desired_temperature)
476 if (!cdev_is_power_actor(instance->cdev))
479 power_actor_set_power(instance->cdev, instance,
481 total_granted_power += granted_power[i];
486 trace_thermal_power_allocator(tz, req_power, total_req_power,
487 granted_power, total_granted_power,
488 num_actors, power_range,
489 max_allocatable_power, tz->temperature,
490 control_temp - tz->temperature);
498 * get_governor_trips() - get the number of the two trip points that are key for this governor
499 * @tz: thermal zone to operate on
500 * @params: pointer to private data for this governor
502 * The power allocator governor works optimally with two trips points:
503 * a "switch on" trip point and a "maximum desired temperature". These
504 * are defined as the first and last passive trip points.
506 * If there is only one trip point, then that's considered to be the
507 * "maximum desired temperature" trip point and the governor is always
508 * on. If there are no passive or active trip points, then the
509 * governor won't do anything. In fact, its throttle function
510 * won't be called at all.
512 static void get_governor_trips(struct thermal_zone_device *tz,
513 struct power_allocator_params *params)
515 int i, last_active, last_passive;
516 bool found_first_passive;
518 found_first_passive = false;
519 last_active = INVALID_TRIP;
520 last_passive = INVALID_TRIP;
522 for (i = 0; i < tz->num_trips; i++) {
523 enum thermal_trip_type type;
526 ret = tz->ops->get_trip_type(tz, i, &type);
528 dev_warn(&tz->device,
529 "Failed to get trip point %d type: %d\n", i,
534 if (type == THERMAL_TRIP_PASSIVE) {
535 if (!found_first_passive) {
536 params->trip_switch_on = i;
537 found_first_passive = true;
541 } else if (type == THERMAL_TRIP_ACTIVE) {
548 if (last_passive != INVALID_TRIP) {
549 params->trip_max_desired_temperature = last_passive;
550 } else if (found_first_passive) {
551 params->trip_max_desired_temperature = params->trip_switch_on;
552 params->trip_switch_on = INVALID_TRIP;
554 params->trip_switch_on = INVALID_TRIP;
555 params->trip_max_desired_temperature = last_active;
559 static void reset_pid_controller(struct power_allocator_params *params)
561 params->err_integral = 0;
562 params->prev_err = 0;
565 static void allow_maximum_power(struct thermal_zone_device *tz, bool update)
567 struct thermal_instance *instance;
568 struct power_allocator_params *params = tz->governor_data;
571 list_for_each_entry(instance, &tz->thermal_instances, tz_node) {
572 struct thermal_cooling_device *cdev = instance->cdev;
574 if ((instance->trip != params->trip_max_desired_temperature) ||
575 (!cdev_is_power_actor(instance->cdev)))
578 instance->target = 0;
579 mutex_lock(&instance->cdev->lock);
581 * Call for updating the cooling devices local stats and avoid
582 * periods of dozen of seconds when those have not been
585 cdev->ops->get_requested_power(cdev, &req_power);
588 __thermal_cdev_update(instance->cdev);
590 mutex_unlock(&instance->cdev->lock);
595 * check_power_actors() - Check all cooling devices and warn when they are
597 * @tz: thermal zone to operate on
599 * Check all cooling devices in the @tz and warn every time they are missing
600 * power actor API. The warning should help to investigate the issue, which
601 * could be e.g. lack of Energy Model for a given device.
603 * Return: 0 on success, -EINVAL if any cooling device does not implement
604 * the power actor API.
606 static int check_power_actors(struct thermal_zone_device *tz)
608 struct thermal_instance *instance;
611 list_for_each_entry(instance, &tz->thermal_instances, tz_node) {
612 if (!cdev_is_power_actor(instance->cdev)) {
613 dev_warn(&tz->device, "power_allocator: %s is not a power actor\n",
614 instance->cdev->type);
623 * power_allocator_bind() - bind the power_allocator governor to a thermal zone
624 * @tz: thermal zone to bind it to
626 * Initialize the PID controller parameters and bind it to the thermal
629 * Return: 0 on success, or -ENOMEM if we ran out of memory, or -EINVAL
630 * when there are unsupported cooling devices in the @tz.
632 static int power_allocator_bind(struct thermal_zone_device *tz)
635 struct power_allocator_params *params;
638 ret = check_power_actors(tz);
642 params = kzalloc(sizeof(*params), GFP_KERNEL);
647 tz->tzp = kzalloc(sizeof(*tz->tzp), GFP_KERNEL);
653 params->allocated_tzp = true;
656 if (!tz->tzp->sustainable_power)
657 dev_warn(&tz->device, "power_allocator: sustainable_power will be estimated\n");
659 get_governor_trips(tz, params);
661 if (tz->num_trips > 0) {
662 ret = tz->ops->get_trip_temp(tz,
663 params->trip_max_desired_temperature,
666 estimate_pid_constants(tz, tz->tzp->sustainable_power,
667 params->trip_switch_on,
671 reset_pid_controller(params);
673 tz->governor_data = params;
683 static void power_allocator_unbind(struct thermal_zone_device *tz)
685 struct power_allocator_params *params = tz->governor_data;
687 dev_dbg(&tz->device, "Unbinding from thermal zone %d\n", tz->id);
689 if (params->allocated_tzp) {
694 kfree(tz->governor_data);
695 tz->governor_data = NULL;
698 static int power_allocator_throttle(struct thermal_zone_device *tz, int trip)
701 int switch_on_temp, control_temp;
702 struct power_allocator_params *params = tz->governor_data;
705 lockdep_assert_held(&tz->lock);
708 * We get called for every trip point but we only need to do
709 * our calculations once
711 if (trip != params->trip_max_desired_temperature)
714 ret = tz->ops->get_trip_temp(tz, params->trip_switch_on,
716 if (!ret && (tz->temperature < switch_on_temp)) {
717 update = (tz->last_temperature >= switch_on_temp);
719 reset_pid_controller(params);
720 allow_maximum_power(tz, update);
726 ret = tz->ops->get_trip_temp(tz, params->trip_max_desired_temperature,
729 dev_warn(&tz->device,
730 "Failed to get the maximum desired temperature: %d\n",
735 return allocate_power(tz, control_temp);
738 static struct thermal_governor thermal_gov_power_allocator = {
739 .name = "power_allocator",
740 .bind_to_tz = power_allocator_bind,
741 .unbind_from_tz = power_allocator_unbind,
742 .throttle = power_allocator_throttle,
744 THERMAL_GOVERNOR_DECLARE(thermal_gov_power_allocator);