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 "thermal_trace_ipa.h"
17 #include "thermal_core.h"
20 #define int_to_frac(x) ((x) << FRAC_BITS)
21 #define frac_to_int(x) ((x) >> FRAC_BITS)
24 * mul_frac() - multiply two fixed-point numbers
25 * @x: first multiplicand
26 * @y: second multiplicand
28 * Return: the result of multiplying two fixed-point numbers. The
29 * result is also a fixed-point number.
31 static inline s64 mul_frac(s64 x, s64 y)
33 return (x * y) >> FRAC_BITS;
37 * div_frac() - divide two fixed-point numbers
41 * Return: the result of dividing two fixed-point numbers. The
42 * result is also a fixed-point number.
44 static inline s64 div_frac(s64 x, s64 y)
46 return div_s64(x << FRAC_BITS, y);
50 * struct power_allocator_params - parameters for the power allocator governor
51 * @allocated_tzp: whether we have allocated tzp for this thermal zone and
52 * it needs to be freed on unbind
53 * @err_integral: accumulated error in the PID controller.
54 * @prev_err: error in the previous iteration of the PID controller.
55 * Used to calculate the derivative term.
56 * @sustainable_power: Sustainable power (heat) that this thermal zone can
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 NULL.
62 * @trip_max_desired_temperature: last passive trip point of the thermal
63 * zone. The temperature we are
66 struct power_allocator_params {
70 u32 sustainable_power;
71 const struct thermal_trip *trip_switch_on;
72 const struct thermal_trip *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 for the switch on temperature
116 * @control_temp: target temperature for the power allocator governor
118 * This function is used to update the estimation of the PID
119 * controller constants in struct thermal_zone_parameters.
121 static void estimate_pid_constants(struct thermal_zone_device *tz,
122 u32 sustainable_power,
123 const struct thermal_trip *trip_switch_on,
126 u32 temperature_threshold = control_temp;
130 temperature_threshold -= trip_switch_on->temperature;
133 * estimate_pid_constants() tries to find appropriate default
134 * values for thermal zones that don't provide them. If a
135 * system integrator has configured a thermal zone with two
136 * passive trip points at the same temperature, that person
137 * hasn't put any effort to set up the thermal zone properly
140 if (!temperature_threshold)
143 tz->tzp->k_po = int_to_frac(sustainable_power) /
144 temperature_threshold;
146 tz->tzp->k_pu = int_to_frac(2 * sustainable_power) /
147 temperature_threshold;
149 k_i = tz->tzp->k_pu / 10;
150 tz->tzp->k_i = k_i > 0 ? k_i : 1;
153 * The default for k_d and integral_cutoff is 0, so we can
154 * leave them as they are.
159 * get_sustainable_power() - Get the right sustainable power
160 * @tz: thermal zone for which to estimate the constants
161 * @params: parameters for the power allocator governor
162 * @control_temp: target temperature for the power allocator governor
164 * This function is used for getting the proper sustainable power value based
165 * on variables which might be updated by the user sysfs interface. If that
166 * happen the new value is going to be estimated and updated. It is also used
167 * after thermal zone binding, where the initial values where set to 0.
169 static u32 get_sustainable_power(struct thermal_zone_device *tz,
170 struct power_allocator_params *params,
173 u32 sustainable_power;
175 if (!tz->tzp->sustainable_power)
176 sustainable_power = estimate_sustainable_power(tz);
178 sustainable_power = tz->tzp->sustainable_power;
180 /* Check if it's init value 0 or there was update via sysfs */
181 if (sustainable_power != params->sustainable_power) {
182 estimate_pid_constants(tz, sustainable_power,
183 params->trip_switch_on, control_temp);
185 /* Do the estimation only once and make available in sysfs */
186 tz->tzp->sustainable_power = sustainable_power;
187 params->sustainable_power = sustainable_power;
190 return sustainable_power;
194 * pid_controller() - PID controller
195 * @tz: thermal zone we are operating in
196 * @control_temp: the target temperature in millicelsius
197 * @max_allocatable_power: maximum allocatable power for this thermal zone
199 * This PID controller increases the available power budget so that the
200 * temperature of the thermal zone gets as close as possible to
201 * @control_temp and limits the power if it exceeds it. k_po is the
202 * proportional term when we are overshooting, k_pu is the
203 * proportional term when we are undershooting. integral_cutoff is a
204 * threshold below which we stop accumulating the error. The
205 * accumulated error is only valid if the requested power will make
206 * the system warmer. If the system is mostly idle, there's no point
207 * in accumulating positive error.
209 * Return: The power budget for the next period.
211 static u32 pid_controller(struct thermal_zone_device *tz,
213 u32 max_allocatable_power)
215 s64 p, i, d, power_range;
216 s32 err, max_power_frac;
217 u32 sustainable_power;
218 struct power_allocator_params *params = tz->governor_data;
220 max_power_frac = int_to_frac(max_allocatable_power);
222 sustainable_power = get_sustainable_power(tz, params, control_temp);
224 err = control_temp - tz->temperature;
225 err = int_to_frac(err);
227 /* Calculate the proportional term */
228 p = mul_frac(err < 0 ? tz->tzp->k_po : tz->tzp->k_pu, err);
231 * Calculate the integral term
233 * if the error is less than cut off allow integration (but
234 * the integral is limited to max power)
236 i = mul_frac(tz->tzp->k_i, params->err_integral);
238 if (err < int_to_frac(tz->tzp->integral_cutoff)) {
239 s64 i_next = i + mul_frac(tz->tzp->k_i, err);
241 if (abs(i_next) < max_power_frac) {
243 params->err_integral += err;
248 * Calculate the derivative term
250 * We do err - prev_err, so with a positive k_d, a decreasing
251 * error (i.e. driving closer to the line) results in less
252 * power being applied, slowing down the controller)
254 d = mul_frac(tz->tzp->k_d, err - params->prev_err);
255 d = div_frac(d, jiffies_to_msecs(tz->passive_delay_jiffies));
256 params->prev_err = err;
258 power_range = p + i + d;
260 /* feed-forward the known sustainable dissipatable power */
261 power_range = sustainable_power + frac_to_int(power_range);
263 power_range = clamp(power_range, (s64)0, (s64)max_allocatable_power);
265 trace_thermal_power_allocator_pid(tz, frac_to_int(err),
266 frac_to_int(params->err_integral),
267 frac_to_int(p), frac_to_int(i),
268 frac_to_int(d), power_range);
274 * power_actor_set_power() - limit the maximum power a cooling device consumes
275 * @cdev: pointer to &thermal_cooling_device
276 * @instance: thermal instance to update
277 * @power: the power in milliwatts
279 * Set the cooling device to consume at most @power milliwatts. The limit is
280 * expected to be a cap at the maximum power consumption.
282 * Return: 0 on success, -EINVAL if the cooling device does not
283 * implement the power actor API or -E* for other failures.
286 power_actor_set_power(struct thermal_cooling_device *cdev,
287 struct thermal_instance *instance, u32 power)
292 ret = cdev->ops->power2state(cdev, power, &state);
296 instance->target = clamp_val(state, instance->lower, instance->upper);
297 mutex_lock(&cdev->lock);
298 __thermal_cdev_update(cdev);
299 mutex_unlock(&cdev->lock);
305 * divvy_up_power() - divvy the allocated power between the actors
306 * @req_power: each actor's requested power
307 * @max_power: each actor's maximum available power
308 * @num_actors: size of the @req_power, @max_power and @granted_power's array
309 * @total_req_power: sum of @req_power
310 * @power_range: total allocated power
311 * @granted_power: output array: each actor's granted power
312 * @extra_actor_power: an appropriately sized array to be used in the
313 * function as temporary storage of the extra power given
316 * This function divides the total allocated power (@power_range)
317 * fairly between the actors. It first tries to give each actor a
318 * share of the @power_range according to how much power it requested
319 * compared to the rest of the actors. For example, if only one actor
320 * requests power, then it receives all the @power_range. If
321 * three actors each requests 1mW, each receives a third of the
324 * If any actor received more than their maximum power, then that
325 * surplus is re-divvied among the actors based on how far they are
326 * from their respective maximums.
328 * Granted power for each actor is written to @granted_power, which
329 * should've been allocated by the calling function.
331 static void divvy_up_power(u32 *req_power, u32 *max_power, int num_actors,
332 u32 total_req_power, u32 power_range,
333 u32 *granted_power, u32 *extra_actor_power)
335 u32 extra_power, capped_extra_power;
339 * Prevent division by 0 if none of the actors request power.
341 if (!total_req_power)
344 capped_extra_power = 0;
346 for (i = 0; i < num_actors; i++) {
347 u64 req_range = (u64)req_power[i] * power_range;
349 granted_power[i] = DIV_ROUND_CLOSEST_ULL(req_range,
352 if (granted_power[i] > max_power[i]) {
353 extra_power += granted_power[i] - max_power[i];
354 granted_power[i] = max_power[i];
357 extra_actor_power[i] = max_power[i] - granted_power[i];
358 capped_extra_power += extra_actor_power[i];
365 * Re-divvy the reclaimed extra among actors based on
366 * how far they are from the max
368 extra_power = min(extra_power, capped_extra_power);
369 if (capped_extra_power > 0)
370 for (i = 0; i < num_actors; i++) {
371 u64 extra_range = (u64)extra_actor_power[i] * extra_power;
372 granted_power[i] += DIV_ROUND_CLOSEST_ULL(extra_range,
377 static int allocate_power(struct thermal_zone_device *tz,
380 struct thermal_instance *instance;
381 struct power_allocator_params *params = tz->governor_data;
382 const struct thermal_trip *trip_max_desired_temperature =
383 params->trip_max_desired_temperature;
384 u32 *req_power, *max_power, *granted_power, *extra_actor_power;
385 u32 *weighted_req_power;
386 u32 total_req_power, max_allocatable_power, total_weighted_req_power;
387 u32 total_granted_power, power_range;
388 int i, num_actors, total_weight, ret = 0;
392 list_for_each_entry(instance, &tz->thermal_instances, tz_node) {
393 if ((instance->trip == trip_max_desired_temperature) &&
394 cdev_is_power_actor(instance->cdev)) {
396 total_weight += instance->weight;
404 * We need to allocate five arrays of the same size:
405 * req_power, max_power, granted_power, extra_actor_power and
406 * weighted_req_power. They are going to be needed until this
407 * function returns. Allocate them all in one go to simplify
408 * the allocation and deallocation logic.
410 BUILD_BUG_ON(sizeof(*req_power) != sizeof(*max_power));
411 BUILD_BUG_ON(sizeof(*req_power) != sizeof(*granted_power));
412 BUILD_BUG_ON(sizeof(*req_power) != sizeof(*extra_actor_power));
413 BUILD_BUG_ON(sizeof(*req_power) != sizeof(*weighted_req_power));
414 req_power = kcalloc(num_actors * 5, sizeof(*req_power), GFP_KERNEL);
418 max_power = &req_power[num_actors];
419 granted_power = &req_power[2 * num_actors];
420 extra_actor_power = &req_power[3 * num_actors];
421 weighted_req_power = &req_power[4 * num_actors];
424 total_weighted_req_power = 0;
426 max_allocatable_power = 0;
428 list_for_each_entry(instance, &tz->thermal_instances, tz_node) {
430 struct thermal_cooling_device *cdev = instance->cdev;
432 if (instance->trip != trip_max_desired_temperature)
435 if (!cdev_is_power_actor(cdev))
438 if (cdev->ops->get_requested_power(cdev, &req_power[i]))
442 weight = 1 << FRAC_BITS;
444 weight = instance->weight;
446 weighted_req_power[i] = frac_to_int(weight * req_power[i]);
448 if (cdev->ops->state2power(cdev, instance->lower,
452 total_req_power += req_power[i];
453 max_allocatable_power += max_power[i];
454 total_weighted_req_power += weighted_req_power[i];
459 power_range = pid_controller(tz, control_temp, max_allocatable_power);
461 divvy_up_power(weighted_req_power, max_power, num_actors,
462 total_weighted_req_power, power_range, granted_power,
465 total_granted_power = 0;
467 list_for_each_entry(instance, &tz->thermal_instances, tz_node) {
468 if (instance->trip != trip_max_desired_temperature)
471 if (!cdev_is_power_actor(instance->cdev))
474 power_actor_set_power(instance->cdev, instance,
476 total_granted_power += granted_power[i];
481 trace_thermal_power_allocator(tz, req_power, total_req_power,
482 granted_power, total_granted_power,
483 num_actors, power_range,
484 max_allocatable_power, tz->temperature,
485 control_temp - tz->temperature);
493 * get_governor_trips() - get the two trip points that are key for this governor
494 * @tz: thermal zone to operate on
495 * @params: pointer to private data for this governor
497 * The power allocator governor works optimally with two trips points:
498 * a "switch on" trip point and a "maximum desired temperature". These
499 * are defined as the first and last passive trip points.
501 * If there is only one trip point, then that's considered to be the
502 * "maximum desired temperature" trip point and the governor is always
503 * on. If there are no passive or active trip points, then the
504 * governor won't do anything. In fact, its throttle function
505 * won't be called at all.
507 static void get_governor_trips(struct thermal_zone_device *tz,
508 struct power_allocator_params *params)
510 const struct thermal_trip *first_passive = NULL;
511 const struct thermal_trip *last_passive = NULL;
512 const struct thermal_trip *last_active = NULL;
513 const struct thermal_trip *trip;
515 for_each_trip(tz, trip) {
516 switch (trip->type) {
517 case THERMAL_TRIP_PASSIVE:
518 if (!first_passive) {
519 first_passive = trip;
524 case THERMAL_TRIP_ACTIVE:
533 params->trip_switch_on = first_passive;
534 params->trip_max_desired_temperature = last_passive;
535 } else if (first_passive) {
536 params->trip_switch_on = NULL;
537 params->trip_max_desired_temperature = first_passive;
539 params->trip_switch_on = NULL;
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, bool update)
552 struct thermal_instance *instance;
553 struct power_allocator_params *params = tz->governor_data;
556 list_for_each_entry(instance, &tz->thermal_instances, tz_node) {
557 struct thermal_cooling_device *cdev = instance->cdev;
559 if (instance->trip != params->trip_max_desired_temperature ||
560 (!cdev_is_power_actor(instance->cdev)))
563 instance->target = 0;
564 mutex_lock(&instance->cdev->lock);
566 * Call for updating the cooling devices local stats and avoid
567 * periods of dozen of seconds when those have not been
570 cdev->ops->get_requested_power(cdev, &req_power);
573 __thermal_cdev_update(instance->cdev);
575 mutex_unlock(&instance->cdev->lock);
580 * check_power_actors() - Check all cooling devices and warn when they are
582 * @tz: thermal zone to operate on
584 * Check all cooling devices in the @tz and warn every time they are missing
585 * power actor API. The warning should help to investigate the issue, which
586 * could be e.g. lack of Energy Model for a given device.
588 * Return: 0 on success, -EINVAL if any cooling device does not implement
589 * the power actor API.
591 static int check_power_actors(struct thermal_zone_device *tz)
593 struct thermal_instance *instance;
596 list_for_each_entry(instance, &tz->thermal_instances, tz_node) {
597 if (!cdev_is_power_actor(instance->cdev)) {
598 dev_warn(&tz->device, "power_allocator: %s is not a power actor\n",
599 instance->cdev->type);
608 * power_allocator_bind() - bind the power_allocator governor to a thermal zone
609 * @tz: thermal zone to bind it to
611 * Initialize the PID controller parameters and bind it to the thermal
614 * Return: 0 on success, or -ENOMEM if we ran out of memory, or -EINVAL
615 * when there are unsupported cooling devices in the @tz.
617 static int power_allocator_bind(struct thermal_zone_device *tz)
620 struct power_allocator_params *params;
622 ret = check_power_actors(tz);
626 params = kzalloc(sizeof(*params), GFP_KERNEL);
631 tz->tzp = kzalloc(sizeof(*tz->tzp), GFP_KERNEL);
637 params->allocated_tzp = true;
640 if (!tz->tzp->sustainable_power)
641 dev_warn(&tz->device, "power_allocator: sustainable_power will be estimated\n");
643 get_governor_trips(tz, params);
645 if (params->trip_max_desired_temperature) {
646 int temp = params->trip_max_desired_temperature->temperature;
648 estimate_pid_constants(tz, tz->tzp->sustainable_power,
649 params->trip_switch_on, temp);
652 reset_pid_controller(params);
654 tz->governor_data = params;
664 static void power_allocator_unbind(struct thermal_zone_device *tz)
666 struct power_allocator_params *params = tz->governor_data;
668 dev_dbg(&tz->device, "Unbinding from thermal zone %d\n", tz->id);
670 if (params->allocated_tzp) {
675 kfree(tz->governor_data);
676 tz->governor_data = NULL;
679 static int power_allocator_throttle(struct thermal_zone_device *tz,
680 const struct thermal_trip *trip)
682 struct power_allocator_params *params = tz->governor_data;
685 lockdep_assert_held(&tz->lock);
688 * We get called for every trip point but we only need to do
689 * our calculations once
691 if (trip != params->trip_max_desired_temperature)
694 trip = params->trip_switch_on;
695 if (trip && tz->temperature < trip->temperature) {
696 update = tz->last_temperature >= trip->temperature;
698 reset_pid_controller(params);
699 allow_maximum_power(tz, update);
705 return allocate_power(tz, params->trip_max_desired_temperature->temperature);
708 static struct thermal_governor thermal_gov_power_allocator = {
709 .name = "power_allocator",
710 .bind_to_tz = power_allocator_bind,
711 .unbind_from_tz = power_allocator_unbind,
712 .throttle = power_allocator_throttle,
714 THERMAL_GOVERNOR_DECLARE(thermal_gov_power_allocator);