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/slab.h>
12 #include <linux/thermal.h>
14 #define CREATE_TRACE_POINTS
15 #include <trace/events/thermal_power_allocator.h>
17 #include "thermal_core.h"
19 #define INVALID_TRIP -1
22 #define int_to_frac(x) ((x) << FRAC_BITS)
23 #define frac_to_int(x) ((x) >> FRAC_BITS)
26 * mul_frac() - multiply two fixed-point numbers
27 * @x: first multiplicand
28 * @y: second multiplicand
30 * Return: the result of multiplying two fixed-point numbers. The
31 * result is also a fixed-point number.
33 static inline s64 mul_frac(s64 x, s64 y)
35 return (x * y) >> FRAC_BITS;
39 * div_frac() - divide two fixed-point numbers
43 * Return: the result of dividing two fixed-point numbers. The
44 * result is also a fixed-point number.
46 static inline s64 div_frac(s64 x, s64 y)
48 return div_s64(x << FRAC_BITS, y);
52 * struct power_allocator_params - parameters for the power allocator governor
53 * @allocated_tzp: whether we have allocated tzp for this thermal zone and
54 * it needs to be freed on unbind
55 * @err_integral: accumulated error in the PID controller.
56 * @prev_err: error in the previous iteration of the PID controller.
57 * Used to calculate the derivative term.
58 * @trip_switch_on: first passive trip point of the thermal zone. The
59 * governor switches on when this trip point is crossed.
60 * If the thermal zone only has one passive trip point,
61 * @trip_switch_on should be INVALID_TRIP.
62 * @trip_max_desired_temperature: last passive trip point of the thermal
63 * zone. The temperature we are
65 * @sustainable_power: Sustainable power (heat) that this thermal zone can
68 struct power_allocator_params {
73 int trip_max_desired_temperature;
74 u32 sustainable_power;
78 * estimate_sustainable_power() - Estimate the sustainable power of a thermal zone
79 * @tz: thermal zone we are operating in
81 * For thermal zones that don't provide a sustainable_power in their
82 * thermal_zone_params, estimate one. Calculate it using the minimum
83 * power of all the cooling devices as that gives a valid value that
84 * can give some degree of functionality. For optimal performance of
85 * this governor, provide a sustainable_power in the thermal zone's
86 * thermal_zone_params.
88 static u32 estimate_sustainable_power(struct thermal_zone_device *tz)
90 u32 sustainable_power = 0;
91 struct thermal_instance *instance;
92 struct power_allocator_params *params = tz->governor_data;
94 list_for_each_entry(instance, &tz->thermal_instances, tz_node) {
95 struct thermal_cooling_device *cdev = instance->cdev;
98 if (instance->trip != params->trip_max_desired_temperature)
101 if (!cdev_is_power_actor(cdev))
104 if (cdev->ops->state2power(cdev, instance->upper, &min_power))
107 sustainable_power += min_power;
110 return sustainable_power;
114 * estimate_pid_constants() - Estimate the constants for the PID controller
115 * @tz: thermal zone for which to estimate the constants
116 * @sustainable_power: sustainable power for the thermal zone
117 * @trip_switch_on: trip point number for the switch on temperature
118 * @control_temp: target temperature for the power allocator governor
120 * This function is used to update the estimation of the PID
121 * controller constants in struct thermal_zone_parameters.
123 static void estimate_pid_constants(struct thermal_zone_device *tz,
124 u32 sustainable_power, int trip_switch_on,
129 u32 temperature_threshold;
132 ret = tz->ops->get_trip_temp(tz, trip_switch_on, &switch_on_temp);
136 temperature_threshold = control_temp - switch_on_temp;
138 * estimate_pid_constants() tries to find appropriate default
139 * values for thermal zones that don't provide them. If a
140 * system integrator has configured a thermal zone with two
141 * passive trip points at the same temperature, that person
142 * hasn't put any effort to set up the thermal zone properly
145 if (!temperature_threshold)
148 tz->tzp->k_po = int_to_frac(sustainable_power) /
149 temperature_threshold;
151 tz->tzp->k_pu = int_to_frac(2 * sustainable_power) /
152 temperature_threshold;
154 k_i = tz->tzp->k_pu / 10;
155 tz->tzp->k_i = k_i > 0 ? k_i : 1;
158 * The default for k_d and integral_cutoff is 0, so we can
159 * leave them as they are.
164 * get_sustainable_power() - Get the right sustainable power
165 * @tz: thermal zone for which to estimate the constants
166 * @params: parameters for the power allocator governor
167 * @control_temp: target temperature for the power allocator governor
169 * This function is used for getting the proper sustainable power value based
170 * on variables which might be updated by the user sysfs interface. If that
171 * happen the new value is going to be estimated and updated. It is also used
172 * after thermal zone binding, where the initial values where set to 0.
174 static u32 get_sustainable_power(struct thermal_zone_device *tz,
175 struct power_allocator_params *params,
178 u32 sustainable_power;
180 if (!tz->tzp->sustainable_power)
181 sustainable_power = estimate_sustainable_power(tz);
183 sustainable_power = tz->tzp->sustainable_power;
185 /* Check if it's init value 0 or there was update via sysfs */
186 if (sustainable_power != params->sustainable_power) {
187 estimate_pid_constants(tz, sustainable_power,
188 params->trip_switch_on, control_temp);
190 /* Do the estimation only once and make available in sysfs */
191 tz->tzp->sustainable_power = sustainable_power;
192 params->sustainable_power = sustainable_power;
195 return sustainable_power;
199 * pid_controller() - PID controller
200 * @tz: thermal zone we are operating in
201 * @control_temp: the target temperature in millicelsius
202 * @max_allocatable_power: maximum allocatable power for this thermal zone
204 * This PID controller increases the available power budget so that the
205 * temperature of the thermal zone gets as close as possible to
206 * @control_temp and limits the power if it exceeds it. k_po is the
207 * proportional term when we are overshooting, k_pu is the
208 * proportional term when we are undershooting. integral_cutoff is a
209 * threshold below which we stop accumulating the error. The
210 * accumulated error is only valid if the requested power will make
211 * the system warmer. If the system is mostly idle, there's no point
212 * in accumulating positive error.
214 * Return: The power budget for the next period.
216 static u32 pid_controller(struct thermal_zone_device *tz,
218 u32 max_allocatable_power)
220 s64 p, i, d, power_range;
221 s32 err, max_power_frac;
222 u32 sustainable_power;
223 struct power_allocator_params *params = tz->governor_data;
225 max_power_frac = int_to_frac(max_allocatable_power);
227 sustainable_power = get_sustainable_power(tz, params, control_temp);
229 err = control_temp - tz->temperature;
230 err = int_to_frac(err);
232 /* Calculate the proportional term */
233 p = mul_frac(err < 0 ? tz->tzp->k_po : tz->tzp->k_pu, err);
236 * Calculate the integral term
238 * if the error is less than cut off allow integration (but
239 * the integral is limited to max power)
241 i = mul_frac(tz->tzp->k_i, params->err_integral);
243 if (err < int_to_frac(tz->tzp->integral_cutoff)) {
244 s64 i_next = i + mul_frac(tz->tzp->k_i, err);
246 if (abs(i_next) < max_power_frac) {
248 params->err_integral += err;
253 * Calculate the derivative term
255 * We do err - prev_err, so with a positive k_d, a decreasing
256 * error (i.e. driving closer to the line) results in less
257 * power being applied, slowing down the controller)
259 d = mul_frac(tz->tzp->k_d, err - params->prev_err);
260 d = div_frac(d, jiffies_to_msecs(tz->passive_delay_jiffies));
261 params->prev_err = err;
263 power_range = p + i + d;
265 /* feed-forward the known sustainable dissipatable power */
266 power_range = sustainable_power + frac_to_int(power_range);
268 power_range = clamp(power_range, (s64)0, (s64)max_allocatable_power);
270 trace_thermal_power_allocator_pid(tz, frac_to_int(err),
271 frac_to_int(params->err_integral),
272 frac_to_int(p), frac_to_int(i),
273 frac_to_int(d), power_range);
279 * power_actor_set_power() - limit the maximum power a cooling device consumes
280 * @cdev: pointer to &thermal_cooling_device
281 * @instance: thermal instance to update
282 * @power: the power in milliwatts
284 * Set the cooling device to consume at most @power milliwatts. The limit is
285 * expected to be a cap at the maximum power consumption.
287 * Return: 0 on success, -EINVAL if the cooling device does not
288 * implement the power actor API or -E* for other failures.
291 power_actor_set_power(struct thermal_cooling_device *cdev,
292 struct thermal_instance *instance, u32 power)
297 ret = cdev->ops->power2state(cdev, power, &state);
301 instance->target = clamp_val(state, instance->lower, instance->upper);
302 mutex_lock(&cdev->lock);
303 __thermal_cdev_update(cdev);
304 mutex_unlock(&cdev->lock);
310 * divvy_up_power() - divvy the allocated power between the actors
311 * @req_power: each actor's requested power
312 * @max_power: each actor's maximum available power
313 * @num_actors: size of the @req_power, @max_power and @granted_power's array
314 * @total_req_power: sum of @req_power
315 * @power_range: total allocated power
316 * @granted_power: output array: each actor's granted power
317 * @extra_actor_power: an appropriately sized array to be used in the
318 * function as temporary storage of the extra power given
321 * This function divides the total allocated power (@power_range)
322 * fairly between the actors. It first tries to give each actor a
323 * share of the @power_range according to how much power it requested
324 * compared to the rest of the actors. For example, if only one actor
325 * requests power, then it receives all the @power_range. If
326 * three actors each requests 1mW, each receives a third of the
329 * If any actor received more than their maximum power, then that
330 * surplus is re-divvied among the actors based on how far they are
331 * from their respective maximums.
333 * Granted power for each actor is written to @granted_power, which
334 * should've been allocated by the calling function.
336 static void divvy_up_power(u32 *req_power, u32 *max_power, int num_actors,
337 u32 total_req_power, u32 power_range,
338 u32 *granted_power, u32 *extra_actor_power)
340 u32 extra_power, capped_extra_power;
344 * Prevent division by 0 if none of the actors request power.
346 if (!total_req_power)
349 capped_extra_power = 0;
351 for (i = 0; i < num_actors; i++) {
352 u64 req_range = (u64)req_power[i] * power_range;
354 granted_power[i] = DIV_ROUND_CLOSEST_ULL(req_range,
357 if (granted_power[i] > max_power[i]) {
358 extra_power += granted_power[i] - max_power[i];
359 granted_power[i] = max_power[i];
362 extra_actor_power[i] = max_power[i] - granted_power[i];
363 capped_extra_power += extra_actor_power[i];
370 * Re-divvy the reclaimed extra among actors based on
371 * how far they are from the max
373 extra_power = min(extra_power, capped_extra_power);
374 if (capped_extra_power > 0)
375 for (i = 0; i < num_actors; i++) {
376 u64 extra_range = (u64)extra_actor_power[i] * extra_power;
377 granted_power[i] += DIV_ROUND_CLOSEST_ULL(extra_range,
382 static int allocate_power(struct thermal_zone_device *tz,
385 struct thermal_instance *instance;
386 struct power_allocator_params *params = tz->governor_data;
387 u32 *req_power, *max_power, *granted_power, *extra_actor_power;
388 u32 *weighted_req_power;
389 u32 total_req_power, max_allocatable_power, total_weighted_req_power;
390 u32 total_granted_power, power_range;
391 int i, num_actors, total_weight, ret = 0;
392 int trip_max_desired_temperature = params->trip_max_desired_temperature;
396 list_for_each_entry(instance, &tz->thermal_instances, tz_node) {
397 if ((instance->trip == trip_max_desired_temperature) &&
398 cdev_is_power_actor(instance->cdev)) {
400 total_weight += instance->weight;
408 * We need to allocate five arrays of the same size:
409 * req_power, max_power, granted_power, extra_actor_power and
410 * weighted_req_power. They are going to be needed until this
411 * function returns. Allocate them all in one go to simplify
412 * the allocation and deallocation logic.
414 BUILD_BUG_ON(sizeof(*req_power) != sizeof(*max_power));
415 BUILD_BUG_ON(sizeof(*req_power) != sizeof(*granted_power));
416 BUILD_BUG_ON(sizeof(*req_power) != sizeof(*extra_actor_power));
417 BUILD_BUG_ON(sizeof(*req_power) != sizeof(*weighted_req_power));
418 req_power = kcalloc(num_actors * 5, sizeof(*req_power), GFP_KERNEL);
422 max_power = &req_power[num_actors];
423 granted_power = &req_power[2 * num_actors];
424 extra_actor_power = &req_power[3 * num_actors];
425 weighted_req_power = &req_power[4 * num_actors];
428 total_weighted_req_power = 0;
430 max_allocatable_power = 0;
432 list_for_each_entry(instance, &tz->thermal_instances, tz_node) {
434 struct thermal_cooling_device *cdev = instance->cdev;
436 if (instance->trip != trip_max_desired_temperature)
439 if (!cdev_is_power_actor(cdev))
442 if (cdev->ops->get_requested_power(cdev, &req_power[i]))
446 weight = 1 << FRAC_BITS;
448 weight = instance->weight;
450 weighted_req_power[i] = frac_to_int(weight * req_power[i]);
452 if (cdev->ops->state2power(cdev, instance->lower,
456 total_req_power += req_power[i];
457 max_allocatable_power += max_power[i];
458 total_weighted_req_power += weighted_req_power[i];
463 power_range = pid_controller(tz, control_temp, max_allocatable_power);
465 divvy_up_power(weighted_req_power, max_power, num_actors,
466 total_weighted_req_power, power_range, granted_power,
469 total_granted_power = 0;
471 list_for_each_entry(instance, &tz->thermal_instances, tz_node) {
472 if (instance->trip != trip_max_desired_temperature)
475 if (!cdev_is_power_actor(instance->cdev))
478 power_actor_set_power(instance->cdev, instance,
480 total_granted_power += granted_power[i];
485 trace_thermal_power_allocator(tz, req_power, total_req_power,
486 granted_power, total_granted_power,
487 num_actors, power_range,
488 max_allocatable_power, tz->temperature,
489 control_temp - tz->temperature);
497 * get_governor_trips() - get the number of the two trip points that are key for this governor
498 * @tz: thermal zone to operate on
499 * @params: pointer to private data for this governor
501 * The power allocator governor works optimally with two trips points:
502 * a "switch on" trip point and a "maximum desired temperature". These
503 * are defined as the first and last passive trip points.
505 * If there is only one trip point, then that's considered to be the
506 * "maximum desired temperature" trip point and the governor is always
507 * on. If there are no passive or active trip points, then the
508 * governor won't do anything. In fact, its throttle function
509 * won't be called at all.
511 static void get_governor_trips(struct thermal_zone_device *tz,
512 struct power_allocator_params *params)
514 int i, last_active, last_passive;
515 bool found_first_passive;
517 found_first_passive = false;
518 last_active = INVALID_TRIP;
519 last_passive = INVALID_TRIP;
521 for (i = 0; i < tz->num_trips; i++) {
522 enum thermal_trip_type type;
525 ret = tz->ops->get_trip_type(tz, i, &type);
527 dev_warn(&tz->device,
528 "Failed to get trip point %d type: %d\n", i,
533 if (type == THERMAL_TRIP_PASSIVE) {
534 if (!found_first_passive) {
535 params->trip_switch_on = i;
536 found_first_passive = true;
540 } else if (type == THERMAL_TRIP_ACTIVE) {
547 if (last_passive != INVALID_TRIP) {
548 params->trip_max_desired_temperature = last_passive;
549 } else if (found_first_passive) {
550 params->trip_max_desired_temperature = params->trip_switch_on;
551 params->trip_switch_on = INVALID_TRIP;
553 params->trip_switch_on = INVALID_TRIP;
554 params->trip_max_desired_temperature = last_active;
558 static void reset_pid_controller(struct power_allocator_params *params)
560 params->err_integral = 0;
561 params->prev_err = 0;
564 static void allow_maximum_power(struct thermal_zone_device *tz, bool update)
566 struct thermal_instance *instance;
567 struct power_allocator_params *params = tz->governor_data;
570 list_for_each_entry(instance, &tz->thermal_instances, tz_node) {
571 struct thermal_cooling_device *cdev = instance->cdev;
573 if ((instance->trip != params->trip_max_desired_temperature) ||
574 (!cdev_is_power_actor(instance->cdev)))
577 instance->target = 0;
578 mutex_lock(&instance->cdev->lock);
580 * Call for updating the cooling devices local stats and avoid
581 * periods of dozen of seconds when those have not been
584 cdev->ops->get_requested_power(cdev, &req_power);
587 __thermal_cdev_update(instance->cdev);
589 mutex_unlock(&instance->cdev->lock);
594 * check_power_actors() - Check all cooling devices and warn when they are
596 * @tz: thermal zone to operate on
598 * Check all cooling devices in the @tz and warn every time they are missing
599 * power actor API. The warning should help to investigate the issue, which
600 * could be e.g. lack of Energy Model for a given device.
602 * Return: 0 on success, -EINVAL if any cooling device does not implement
603 * the power actor API.
605 static int check_power_actors(struct thermal_zone_device *tz)
607 struct thermal_instance *instance;
610 list_for_each_entry(instance, &tz->thermal_instances, tz_node) {
611 if (!cdev_is_power_actor(instance->cdev)) {
612 dev_warn(&tz->device, "power_allocator: %s is not a power actor\n",
613 instance->cdev->type);
622 * power_allocator_bind() - bind the power_allocator governor to a thermal zone
623 * @tz: thermal zone to bind it to
625 * Initialize the PID controller parameters and bind it to the thermal
628 * Return: 0 on success, or -ENOMEM if we ran out of memory, or -EINVAL
629 * when there are unsupported cooling devices in the @tz.
631 static int power_allocator_bind(struct thermal_zone_device *tz)
634 struct power_allocator_params *params;
637 ret = check_power_actors(tz);
641 params = kzalloc(sizeof(*params), GFP_KERNEL);
646 tz->tzp = kzalloc(sizeof(*tz->tzp), GFP_KERNEL);
652 params->allocated_tzp = true;
655 if (!tz->tzp->sustainable_power)
656 dev_warn(&tz->device, "power_allocator: sustainable_power will be estimated\n");
658 get_governor_trips(tz, params);
660 if (tz->num_trips > 0) {
661 ret = tz->ops->get_trip_temp(tz,
662 params->trip_max_desired_temperature,
665 estimate_pid_constants(tz, tz->tzp->sustainable_power,
666 params->trip_switch_on,
670 reset_pid_controller(params);
672 tz->governor_data = params;
682 static void power_allocator_unbind(struct thermal_zone_device *tz)
684 struct power_allocator_params *params = tz->governor_data;
686 dev_dbg(&tz->device, "Unbinding from thermal zone %d\n", tz->id);
688 if (params->allocated_tzp) {
693 kfree(tz->governor_data);
694 tz->governor_data = NULL;
697 static int power_allocator_throttle(struct thermal_zone_device *tz, int trip)
700 int switch_on_temp, control_temp;
701 struct power_allocator_params *params = tz->governor_data;
704 lockdep_assert_held(&tz->lock);
707 * We get called for every trip point but we only need to do
708 * our calculations once
710 if (trip != params->trip_max_desired_temperature)
713 ret = tz->ops->get_trip_temp(tz, params->trip_switch_on,
715 if (!ret && (tz->temperature < switch_on_temp)) {
716 update = (tz->last_temperature >= switch_on_temp);
718 reset_pid_controller(params);
719 allow_maximum_power(tz, update);
725 ret = tz->ops->get_trip_temp(tz, params->trip_max_desired_temperature,
728 dev_warn(&tz->device,
729 "Failed to get the maximum desired temperature: %d\n",
734 return allocate_power(tz, control_temp);
737 static struct thermal_governor thermal_gov_power_allocator = {
738 .name = "power_allocator",
739 .bind_to_tz = power_allocator_bind,
740 .unbind_from_tz = power_allocator_unbind,
741 .throttle = power_allocator_throttle,
743 THERMAL_GOVERNOR_DECLARE(thermal_gov_power_allocator);