Loading...
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * A power allocator to manage temperature
4 *
5 * Copyright (C) 2014 ARM Ltd.
6 *
7 */
8
9#define pr_fmt(fmt) "Power allocator: " fmt
10
11#include <linux/slab.h>
12#include <linux/thermal.h>
13
14#define CREATE_TRACE_POINTS
15#include "thermal_trace_ipa.h"
16
17#include "thermal_core.h"
18
19#define FRAC_BITS 10
20#define int_to_frac(x) ((x) << FRAC_BITS)
21#define frac_to_int(x) ((x) >> FRAC_BITS)
22
23/**
24 * mul_frac() - multiply two fixed-point numbers
25 * @x: first multiplicand
26 * @y: second multiplicand
27 *
28 * Return: the result of multiplying two fixed-point numbers. The
29 * result is also a fixed-point number.
30 */
31static inline s64 mul_frac(s64 x, s64 y)
32{
33 return (x * y) >> FRAC_BITS;
34}
35
36/**
37 * div_frac() - divide two fixed-point numbers
38 * @x: the dividend
39 * @y: the divisor
40 *
41 * Return: the result of dividing two fixed-point numbers. The
42 * result is also a fixed-point number.
43 */
44static inline s64 div_frac(s64 x, s64 y)
45{
46 return div_s64(x << FRAC_BITS, y);
47}
48
49/**
50 * struct power_actor - internal power information for power actor
51 * @req_power: requested power value (not weighted)
52 * @max_power: max allocatable power for this actor
53 * @granted_power: granted power for this actor
54 * @extra_actor_power: extra power that this actor can receive
55 * @weighted_req_power: weighted requested power as input to IPA
56 */
57struct power_actor {
58 u32 req_power;
59 u32 max_power;
60 u32 granted_power;
61 u32 extra_actor_power;
62 u32 weighted_req_power;
63};
64
65/**
66 * struct power_allocator_params - parameters for the power allocator governor
67 * @allocated_tzp: whether we have allocated tzp for this thermal zone and
68 * it needs to be freed on unbind
69 * @err_integral: accumulated error in the PID controller.
70 * @prev_err: error in the previous iteration of the PID controller.
71 * Used to calculate the derivative term.
72 * @sustainable_power: Sustainable power (heat) that this thermal zone can
73 * dissipate
74 * @trip_switch_on: first passive trip point of the thermal zone. The
75 * governor switches on when this trip point is crossed.
76 * If the thermal zone only has one passive trip point,
77 * @trip_switch_on should be NULL.
78 * @trip_max: last passive trip point of the thermal zone. The
79 * temperature we are controlling for.
80 * @total_weight: Sum of all thermal instances weights
81 * @num_actors: number of cooling devices supporting IPA callbacks
82 * @buffer_size: internal buffer size, to avoid runtime re-calculation
83 * @power: buffer for all power actors internal power information
84 */
85struct power_allocator_params {
86 bool allocated_tzp;
87 s64 err_integral;
88 s32 prev_err;
89 u32 sustainable_power;
90 const struct thermal_trip *trip_switch_on;
91 const struct thermal_trip *trip_max;
92 int total_weight;
93 unsigned int num_actors;
94 unsigned int buffer_size;
95 struct power_actor *power;
96};
97
98static bool power_actor_is_valid(struct power_allocator_params *params,
99 struct thermal_instance *instance)
100{
101 return (instance->trip == params->trip_max &&
102 cdev_is_power_actor(instance->cdev));
103}
104
105/**
106 * estimate_sustainable_power() - Estimate the sustainable power of a thermal zone
107 * @tz: thermal zone we are operating in
108 *
109 * For thermal zones that don't provide a sustainable_power in their
110 * thermal_zone_params, estimate one. Calculate it using the minimum
111 * power of all the cooling devices as that gives a valid value that
112 * can give some degree of functionality. For optimal performance of
113 * this governor, provide a sustainable_power in the thermal zone's
114 * thermal_zone_params.
115 */
116static u32 estimate_sustainable_power(struct thermal_zone_device *tz)
117{
118 struct power_allocator_params *params = tz->governor_data;
119 struct thermal_cooling_device *cdev;
120 struct thermal_instance *instance;
121 u32 sustainable_power = 0;
122 u32 min_power;
123
124 list_for_each_entry(instance, &tz->thermal_instances, tz_node) {
125 if (!power_actor_is_valid(params, instance))
126 continue;
127
128 cdev = instance->cdev;
129 if (cdev->ops->state2power(cdev, instance->upper, &min_power))
130 continue;
131
132 sustainable_power += min_power;
133 }
134
135 return sustainable_power;
136}
137
138/**
139 * estimate_pid_constants() - Estimate the constants for the PID controller
140 * @tz: thermal zone for which to estimate the constants
141 * @sustainable_power: sustainable power for the thermal zone
142 * @trip_switch_on: trip point for the switch on temperature
143 * @control_temp: target temperature for the power allocator governor
144 *
145 * This function is used to update the estimation of the PID
146 * controller constants in struct thermal_zone_parameters.
147 */
148static void estimate_pid_constants(struct thermal_zone_device *tz,
149 u32 sustainable_power,
150 const struct thermal_trip *trip_switch_on,
151 int control_temp)
152{
153 u32 temperature_threshold = control_temp;
154 s32 k_i;
155
156 if (trip_switch_on)
157 temperature_threshold -= trip_switch_on->temperature;
158
159 /*
160 * estimate_pid_constants() tries to find appropriate default
161 * values for thermal zones that don't provide them. If a
162 * system integrator has configured a thermal zone with two
163 * passive trip points at the same temperature, that person
164 * hasn't put any effort to set up the thermal zone properly
165 * so just give up.
166 */
167 if (!temperature_threshold)
168 return;
169
170 tz->tzp->k_po = int_to_frac(sustainable_power) /
171 temperature_threshold;
172
173 tz->tzp->k_pu = int_to_frac(2 * sustainable_power) /
174 temperature_threshold;
175
176 k_i = tz->tzp->k_pu / 10;
177 tz->tzp->k_i = k_i > 0 ? k_i : 1;
178
179 /*
180 * The default for k_d and integral_cutoff is 0, so we can
181 * leave them as they are.
182 */
183}
184
185/**
186 * get_sustainable_power() - Get the right sustainable power
187 * @tz: thermal zone for which to estimate the constants
188 * @params: parameters for the power allocator governor
189 * @control_temp: target temperature for the power allocator governor
190 *
191 * This function is used for getting the proper sustainable power value based
192 * on variables which might be updated by the user sysfs interface. If that
193 * happen the new value is going to be estimated and updated. It is also used
194 * after thermal zone binding, where the initial values where set to 0.
195 */
196static u32 get_sustainable_power(struct thermal_zone_device *tz,
197 struct power_allocator_params *params,
198 int control_temp)
199{
200 u32 sustainable_power;
201
202 if (!tz->tzp->sustainable_power)
203 sustainable_power = estimate_sustainable_power(tz);
204 else
205 sustainable_power = tz->tzp->sustainable_power;
206
207 /* Check if it's init value 0 or there was update via sysfs */
208 if (sustainable_power != params->sustainable_power) {
209 estimate_pid_constants(tz, sustainable_power,
210 params->trip_switch_on, control_temp);
211
212 /* Do the estimation only once and make available in sysfs */
213 tz->tzp->sustainable_power = sustainable_power;
214 params->sustainable_power = sustainable_power;
215 }
216
217 return sustainable_power;
218}
219
220/**
221 * pid_controller() - PID controller
222 * @tz: thermal zone we are operating in
223 * @control_temp: the target temperature in millicelsius
224 * @max_allocatable_power: maximum allocatable power for this thermal zone
225 *
226 * This PID controller increases the available power budget so that the
227 * temperature of the thermal zone gets as close as possible to
228 * @control_temp and limits the power if it exceeds it. k_po is the
229 * proportional term when we are overshooting, k_pu is the
230 * proportional term when we are undershooting. integral_cutoff is a
231 * threshold below which we stop accumulating the error. The
232 * accumulated error is only valid if the requested power will make
233 * the system warmer. If the system is mostly idle, there's no point
234 * in accumulating positive error.
235 *
236 * Return: The power budget for the next period.
237 */
238static u32 pid_controller(struct thermal_zone_device *tz,
239 int control_temp,
240 u32 max_allocatable_power)
241{
242 struct power_allocator_params *params = tz->governor_data;
243 s64 p, i, d, power_range;
244 s32 err, max_power_frac;
245 u32 sustainable_power;
246
247 max_power_frac = int_to_frac(max_allocatable_power);
248
249 sustainable_power = get_sustainable_power(tz, params, control_temp);
250
251 err = control_temp - tz->temperature;
252 err = int_to_frac(err);
253
254 /* Calculate the proportional term */
255 p = mul_frac(err < 0 ? tz->tzp->k_po : tz->tzp->k_pu, err);
256
257 /*
258 * Calculate the integral term
259 *
260 * if the error is less than cut off allow integration (but
261 * the integral is limited to max power)
262 */
263 i = mul_frac(tz->tzp->k_i, params->err_integral);
264
265 if (err < int_to_frac(tz->tzp->integral_cutoff)) {
266 s64 i_next = i + mul_frac(tz->tzp->k_i, err);
267
268 if (abs(i_next) < max_power_frac) {
269 i = i_next;
270 params->err_integral += err;
271 }
272 }
273
274 /*
275 * Calculate the derivative term
276 *
277 * We do err - prev_err, so with a positive k_d, a decreasing
278 * error (i.e. driving closer to the line) results in less
279 * power being applied, slowing down the controller)
280 */
281 d = mul_frac(tz->tzp->k_d, err - params->prev_err);
282 d = div_frac(d, jiffies_to_msecs(tz->passive_delay_jiffies));
283 params->prev_err = err;
284
285 power_range = p + i + d;
286
287 /* feed-forward the known sustainable dissipatable power */
288 power_range = sustainable_power + frac_to_int(power_range);
289
290 power_range = clamp(power_range, (s64)0, (s64)max_allocatable_power);
291
292 trace_thermal_power_allocator_pid(tz, frac_to_int(err),
293 frac_to_int(params->err_integral),
294 frac_to_int(p), frac_to_int(i),
295 frac_to_int(d), power_range);
296
297 return power_range;
298}
299
300/**
301 * power_actor_set_power() - limit the maximum power a cooling device consumes
302 * @cdev: pointer to &thermal_cooling_device
303 * @instance: thermal instance to update
304 * @power: the power in milliwatts
305 *
306 * Set the cooling device to consume at most @power milliwatts. The limit is
307 * expected to be a cap at the maximum power consumption.
308 *
309 * Return: 0 on success, -EINVAL if the cooling device does not
310 * implement the power actor API or -E* for other failures.
311 */
312static int
313power_actor_set_power(struct thermal_cooling_device *cdev,
314 struct thermal_instance *instance, u32 power)
315{
316 unsigned long state;
317 int ret;
318
319 ret = cdev->ops->power2state(cdev, power, &state);
320 if (ret)
321 return ret;
322
323 instance->target = clamp_val(state, instance->lower, instance->upper);
324 mutex_lock(&cdev->lock);
325 __thermal_cdev_update(cdev);
326 mutex_unlock(&cdev->lock);
327
328 return 0;
329}
330
331/**
332 * divvy_up_power() - divvy the allocated power between the actors
333 * @power: buffer for all power actors internal power information
334 * @num_actors: number of power actors in this thermal zone
335 * @total_req_power: sum of all weighted requested power for all actors
336 * @power_range: total allocated power
337 *
338 * This function divides the total allocated power (@power_range)
339 * fairly between the actors. It first tries to give each actor a
340 * share of the @power_range according to how much power it requested
341 * compared to the rest of the actors. For example, if only one actor
342 * requests power, then it receives all the @power_range. If
343 * three actors each requests 1mW, each receives a third of the
344 * @power_range.
345 *
346 * If any actor received more than their maximum power, then that
347 * surplus is re-divvied among the actors based on how far they are
348 * from their respective maximums.
349 */
350static void divvy_up_power(struct power_actor *power, int num_actors,
351 u32 total_req_power, u32 power_range)
352{
353 u32 capped_extra_power = 0;
354 u32 extra_power = 0;
355 int i;
356
357 /*
358 * Prevent division by 0 if none of the actors request power.
359 */
360 if (!total_req_power)
361 total_req_power = 1;
362
363 for (i = 0; i < num_actors; i++) {
364 struct power_actor *pa = &power[i];
365 u64 req_range = (u64)pa->req_power * power_range;
366
367 pa->granted_power = DIV_ROUND_CLOSEST_ULL(req_range,
368 total_req_power);
369
370 if (pa->granted_power > pa->max_power) {
371 extra_power += pa->granted_power - pa->max_power;
372 pa->granted_power = pa->max_power;
373 }
374
375 pa->extra_actor_power = pa->max_power - pa->granted_power;
376 capped_extra_power += pa->extra_actor_power;
377 }
378
379 if (!extra_power || !capped_extra_power)
380 return;
381
382 /*
383 * Re-divvy the reclaimed extra among actors based on
384 * how far they are from the max
385 */
386 extra_power = min(extra_power, capped_extra_power);
387
388 for (i = 0; i < num_actors; i++) {
389 struct power_actor *pa = &power[i];
390 u64 extra_range = pa->extra_actor_power;
391
392 extra_range *= extra_power;
393 pa->granted_power += DIV_ROUND_CLOSEST_ULL(extra_range,
394 capped_extra_power);
395 }
396}
397
398static int allocate_power(struct thermal_zone_device *tz, int control_temp)
399{
400 struct power_allocator_params *params = tz->governor_data;
401 unsigned int num_actors = params->num_actors;
402 struct power_actor *power = params->power;
403 struct thermal_cooling_device *cdev;
404 struct thermal_instance *instance;
405 u32 total_weighted_req_power = 0;
406 u32 max_allocatable_power = 0;
407 u32 total_granted_power = 0;
408 u32 total_req_power = 0;
409 u32 power_range, weight;
410 int i = 0, ret;
411
412 if (!num_actors)
413 return -ENODEV;
414
415 /* Clean all buffers for new power estimations */
416 memset(power, 0, params->buffer_size);
417
418 list_for_each_entry(instance, &tz->thermal_instances, tz_node) {
419 struct power_actor *pa = &power[i];
420
421 if (!power_actor_is_valid(params, instance))
422 continue;
423
424 cdev = instance->cdev;
425
426 ret = cdev->ops->get_requested_power(cdev, &pa->req_power);
427 if (ret)
428 continue;
429
430 if (!params->total_weight)
431 weight = 1 << FRAC_BITS;
432 else
433 weight = instance->weight;
434
435 pa->weighted_req_power = frac_to_int(weight * pa->req_power);
436
437 ret = cdev->ops->state2power(cdev, instance->lower,
438 &pa->max_power);
439 if (ret)
440 continue;
441
442 total_req_power += pa->req_power;
443 max_allocatable_power += pa->max_power;
444 total_weighted_req_power += pa->weighted_req_power;
445
446 i++;
447 }
448
449 power_range = pid_controller(tz, control_temp, max_allocatable_power);
450
451 divvy_up_power(power, num_actors, total_weighted_req_power,
452 power_range);
453
454 i = 0;
455 list_for_each_entry(instance, &tz->thermal_instances, tz_node) {
456 struct power_actor *pa = &power[i];
457
458 if (!power_actor_is_valid(params, instance))
459 continue;
460
461 power_actor_set_power(instance->cdev, instance,
462 pa->granted_power);
463 total_granted_power += pa->granted_power;
464
465 trace_thermal_power_actor(tz, i, pa->req_power,
466 pa->granted_power);
467 i++;
468 }
469
470 trace_thermal_power_allocator(tz, total_req_power, total_granted_power,
471 num_actors, power_range,
472 max_allocatable_power, tz->temperature,
473 control_temp - tz->temperature);
474
475 return 0;
476}
477
478/**
479 * get_governor_trips() - get the two trip points that are key for this governor
480 * @tz: thermal zone to operate on
481 * @params: pointer to private data for this governor
482 *
483 * The power allocator governor works optimally with two trips points:
484 * a "switch on" trip point and a "maximum desired temperature". These
485 * are defined as the first and last passive trip points.
486 *
487 * If there is only one trip point, then that's considered to be the
488 * "maximum desired temperature" trip point and the governor is always
489 * on. If there are no passive or active trip points, then the
490 * governor won't do anything. In fact, its throttle function
491 * won't be called at all.
492 */
493static void get_governor_trips(struct thermal_zone_device *tz,
494 struct power_allocator_params *params)
495{
496 const struct thermal_trip *first_passive = NULL;
497 const struct thermal_trip *last_passive = NULL;
498 const struct thermal_trip *last_active = NULL;
499 const struct thermal_trip *trip;
500
501 for_each_trip(tz, trip) {
502 switch (trip->type) {
503 case THERMAL_TRIP_PASSIVE:
504 if (!first_passive) {
505 first_passive = trip;
506 break;
507 }
508 last_passive = trip;
509 break;
510 case THERMAL_TRIP_ACTIVE:
511 last_active = trip;
512 break;
513 default:
514 break;
515 }
516 }
517
518 if (last_passive) {
519 params->trip_switch_on = first_passive;
520 params->trip_max = last_passive;
521 } else if (first_passive) {
522 params->trip_switch_on = NULL;
523 params->trip_max = first_passive;
524 } else {
525 params->trip_switch_on = NULL;
526 params->trip_max = last_active;
527 }
528}
529
530static void reset_pid_controller(struct power_allocator_params *params)
531{
532 params->err_integral = 0;
533 params->prev_err = 0;
534}
535
536static void allow_maximum_power(struct thermal_zone_device *tz, bool update)
537{
538 struct power_allocator_params *params = tz->governor_data;
539 struct thermal_cooling_device *cdev;
540 struct thermal_instance *instance;
541 u32 req_power;
542
543 list_for_each_entry(instance, &tz->thermal_instances, tz_node) {
544 if (!power_actor_is_valid(params, instance))
545 continue;
546
547 cdev = instance->cdev;
548
549 instance->target = 0;
550 mutex_lock(&cdev->lock);
551 /*
552 * Call for updating the cooling devices local stats and avoid
553 * periods of dozen of seconds when those have not been
554 * maintained.
555 */
556 cdev->ops->get_requested_power(cdev, &req_power);
557
558 if (update)
559 __thermal_cdev_update(cdev);
560
561 mutex_unlock(&cdev->lock);
562 }
563}
564
565/**
566 * check_power_actors() - Check all cooling devices and warn when they are
567 * not power actors
568 * @tz: thermal zone to operate on
569 * @params: power allocator private data
570 *
571 * Check all cooling devices in the @tz and warn every time they are missing
572 * power actor API. The warning should help to investigate the issue, which
573 * could be e.g. lack of Energy Model for a given device.
574 *
575 * If all of the cooling devices currently attached to @tz implement the power
576 * actor API, return the number of them (which may be 0, because some cooling
577 * devices may be attached later). Otherwise, return -EINVAL.
578 */
579static int check_power_actors(struct thermal_zone_device *tz,
580 struct power_allocator_params *params)
581{
582 struct thermal_instance *instance;
583 int ret = 0;
584
585 list_for_each_entry(instance, &tz->thermal_instances, tz_node) {
586 if (instance->trip != params->trip_max)
587 continue;
588
589 if (!cdev_is_power_actor(instance->cdev)) {
590 dev_warn(&tz->device, "power_allocator: %s is not a power actor\n",
591 instance->cdev->type);
592 return -EINVAL;
593 }
594 ret++;
595 }
596
597 return ret;
598}
599
600static int allocate_actors_buffer(struct power_allocator_params *params,
601 int num_actors)
602{
603 int ret;
604
605 kfree(params->power);
606
607 /* There might be no cooling devices yet. */
608 if (!num_actors) {
609 ret = -EINVAL;
610 goto clean_state;
611 }
612
613 params->power = kcalloc(num_actors, sizeof(struct power_actor),
614 GFP_KERNEL);
615 if (!params->power) {
616 ret = -ENOMEM;
617 goto clean_state;
618 }
619
620 params->num_actors = num_actors;
621 params->buffer_size = num_actors * sizeof(struct power_actor);
622
623 return 0;
624
625clean_state:
626 params->num_actors = 0;
627 params->buffer_size = 0;
628 params->power = NULL;
629 return ret;
630}
631
632static void power_allocator_update_tz(struct thermal_zone_device *tz,
633 enum thermal_notify_event reason)
634{
635 struct power_allocator_params *params = tz->governor_data;
636 struct thermal_instance *instance;
637 int num_actors = 0;
638
639 switch (reason) {
640 case THERMAL_TZ_BIND_CDEV:
641 case THERMAL_TZ_UNBIND_CDEV:
642 list_for_each_entry(instance, &tz->thermal_instances, tz_node)
643 if (power_actor_is_valid(params, instance))
644 num_actors++;
645
646 if (num_actors == params->num_actors)
647 return;
648
649 allocate_actors_buffer(params, num_actors);
650 break;
651 case THERMAL_INSTANCE_WEIGHT_CHANGED:
652 params->total_weight = 0;
653 list_for_each_entry(instance, &tz->thermal_instances, tz_node)
654 if (power_actor_is_valid(params, instance))
655 params->total_weight += instance->weight;
656 break;
657 default:
658 break;
659 }
660}
661
662/**
663 * power_allocator_bind() - bind the power_allocator governor to a thermal zone
664 * @tz: thermal zone to bind it to
665 *
666 * Initialize the PID controller parameters and bind it to the thermal
667 * zone.
668 *
669 * Return: 0 on success, or -ENOMEM if we ran out of memory, or -EINVAL
670 * when there are unsupported cooling devices in the @tz.
671 */
672static int power_allocator_bind(struct thermal_zone_device *tz)
673{
674 struct power_allocator_params *params;
675 int ret;
676
677 params = kzalloc(sizeof(*params), GFP_KERNEL);
678 if (!params)
679 return -ENOMEM;
680
681 get_governor_trips(tz, params);
682 if (!params->trip_max) {
683 dev_warn(&tz->device, "power_allocator: missing trip_max\n");
684 kfree(params);
685 return -EINVAL;
686 }
687
688 ret = check_power_actors(tz, params);
689 if (ret < 0) {
690 dev_warn(&tz->device, "power_allocator: binding failed\n");
691 kfree(params);
692 return ret;
693 }
694
695 ret = allocate_actors_buffer(params, ret);
696 if (ret) {
697 dev_warn(&tz->device, "power_allocator: allocation failed\n");
698 kfree(params);
699 return ret;
700 }
701
702 if (!tz->tzp) {
703 tz->tzp = kzalloc(sizeof(*tz->tzp), GFP_KERNEL);
704 if (!tz->tzp) {
705 ret = -ENOMEM;
706 goto free_params;
707 }
708
709 params->allocated_tzp = true;
710 }
711
712 if (!tz->tzp->sustainable_power)
713 dev_warn(&tz->device, "power_allocator: sustainable_power will be estimated\n");
714
715 estimate_pid_constants(tz, tz->tzp->sustainable_power,
716 params->trip_switch_on,
717 params->trip_max->temperature);
718
719 reset_pid_controller(params);
720
721 tz->governor_data = params;
722
723 return 0;
724
725free_params:
726 kfree(params->power);
727 kfree(params);
728
729 return ret;
730}
731
732static void power_allocator_unbind(struct thermal_zone_device *tz)
733{
734 struct power_allocator_params *params = tz->governor_data;
735
736 dev_dbg(&tz->device, "Unbinding from thermal zone %d\n", tz->id);
737
738 if (params->allocated_tzp) {
739 kfree(tz->tzp);
740 tz->tzp = NULL;
741 }
742
743 kfree(params->power);
744 kfree(tz->governor_data);
745 tz->governor_data = NULL;
746}
747
748static int power_allocator_throttle(struct thermal_zone_device *tz,
749 const struct thermal_trip *trip)
750{
751 struct power_allocator_params *params = tz->governor_data;
752 bool update;
753
754 lockdep_assert_held(&tz->lock);
755
756 /*
757 * We get called for every trip point but we only need to do
758 * our calculations once
759 */
760 if (trip != params->trip_max)
761 return 0;
762
763 trip = params->trip_switch_on;
764 if (trip && tz->temperature < trip->temperature) {
765 update = tz->passive;
766 tz->passive = 0;
767 reset_pid_controller(params);
768 allow_maximum_power(tz, update);
769 return 0;
770 }
771
772 tz->passive = 1;
773
774 return allocate_power(tz, params->trip_max->temperature);
775}
776
777static struct thermal_governor thermal_gov_power_allocator = {
778 .name = "power_allocator",
779 .bind_to_tz = power_allocator_bind,
780 .unbind_from_tz = power_allocator_unbind,
781 .throttle = power_allocator_throttle,
782 .update_tz = power_allocator_update_tz,
783};
784THERMAL_GOVERNOR_DECLARE(thermal_gov_power_allocator);
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * A power allocator to manage temperature
4 *
5 * Copyright (C) 2014 ARM Ltd.
6 *
7 */
8
9#define pr_fmt(fmt) "Power allocator: " fmt
10
11#include <linux/rculist.h>
12#include <linux/slab.h>
13#include <linux/thermal.h>
14
15#define CREATE_TRACE_POINTS
16#include <trace/events/thermal_power_allocator.h>
17
18#include "thermal_core.h"
19
20#define INVALID_TRIP -1
21
22#define FRAC_BITS 10
23#define int_to_frac(x) ((x) << FRAC_BITS)
24#define frac_to_int(x) ((x) >> FRAC_BITS)
25
26/**
27 * mul_frac() - multiply two fixed-point numbers
28 * @x: first multiplicand
29 * @y: second multiplicand
30 *
31 * Return: the result of multiplying two fixed-point numbers. The
32 * result is also a fixed-point number.
33 */
34static inline s64 mul_frac(s64 x, s64 y)
35{
36 return (x * y) >> FRAC_BITS;
37}
38
39/**
40 * div_frac() - divide two fixed-point numbers
41 * @x: the dividend
42 * @y: the divisor
43 *
44 * Return: the result of dividing two fixed-point numbers. The
45 * result is also a fixed-point number.
46 */
47static inline s64 div_frac(s64 x, s64 y)
48{
49 return div_s64(x << FRAC_BITS, y);
50}
51
52/**
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
65 * controlling for.
66 * @sustainable_power: Sustainable power (heat) that this thermal zone can
67 * dissipate
68 */
69struct power_allocator_params {
70 bool allocated_tzp;
71 s64 err_integral;
72 s32 prev_err;
73 int trip_switch_on;
74 int trip_max_desired_temperature;
75 u32 sustainable_power;
76};
77
78/**
79 * estimate_sustainable_power() - Estimate the sustainable power of a thermal zone
80 * @tz: thermal zone we are operating in
81 *
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.
88 */
89static u32 estimate_sustainable_power(struct thermal_zone_device *tz)
90{
91 u32 sustainable_power = 0;
92 struct thermal_instance *instance;
93 struct power_allocator_params *params = tz->governor_data;
94
95 list_for_each_entry(instance, &tz->thermal_instances, tz_node) {
96 struct thermal_cooling_device *cdev = instance->cdev;
97 u32 min_power;
98
99 if (instance->trip != params->trip_max_desired_temperature)
100 continue;
101
102 if (!cdev_is_power_actor(cdev))
103 continue;
104
105 if (cdev->ops->state2power(cdev, instance->upper, &min_power))
106 continue;
107
108 sustainable_power += min_power;
109 }
110
111 return sustainable_power;
112}
113
114/**
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
120 *
121 * This function is used to update the estimation of the PID
122 * controller constants in struct thermal_zone_parameters.
123 */
124static void estimate_pid_constants(struct thermal_zone_device *tz,
125 u32 sustainable_power, int trip_switch_on,
126 int control_temp)
127{
128 int ret;
129 int switch_on_temp;
130 u32 temperature_threshold;
131 s32 k_i;
132
133 ret = tz->ops->get_trip_temp(tz, trip_switch_on, &switch_on_temp);
134 if (ret)
135 switch_on_temp = 0;
136
137 temperature_threshold = control_temp - switch_on_temp;
138 /*
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
144 * so just give up.
145 */
146 if (!temperature_threshold)
147 return;
148
149 tz->tzp->k_po = int_to_frac(sustainable_power) /
150 temperature_threshold;
151
152 tz->tzp->k_pu = int_to_frac(2 * sustainable_power) /
153 temperature_threshold;
154
155 k_i = tz->tzp->k_pu / 10;
156 tz->tzp->k_i = k_i > 0 ? k_i : 1;
157
158 /*
159 * The default for k_d and integral_cutoff is 0, so we can
160 * leave them as they are.
161 */
162}
163
164/**
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
169 *
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.
174 */
175static u32 get_sustainable_power(struct thermal_zone_device *tz,
176 struct power_allocator_params *params,
177 int control_temp)
178{
179 u32 sustainable_power;
180
181 if (!tz->tzp->sustainable_power)
182 sustainable_power = estimate_sustainable_power(tz);
183 else
184 sustainable_power = tz->tzp->sustainable_power;
185
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);
190
191 /* Do the estimation only once and make available in sysfs */
192 tz->tzp->sustainable_power = sustainable_power;
193 params->sustainable_power = sustainable_power;
194 }
195
196 return sustainable_power;
197}
198
199/**
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
204 *
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.
214 *
215 * Return: The power budget for the next period.
216 */
217static u32 pid_controller(struct thermal_zone_device *tz,
218 int control_temp,
219 u32 max_allocatable_power)
220{
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;
225
226 max_power_frac = int_to_frac(max_allocatable_power);
227
228 sustainable_power = get_sustainable_power(tz, params, control_temp);
229
230 err = control_temp - tz->temperature;
231 err = int_to_frac(err);
232
233 /* Calculate the proportional term */
234 p = mul_frac(err < 0 ? tz->tzp->k_po : tz->tzp->k_pu, err);
235
236 /*
237 * Calculate the integral term
238 *
239 * if the error is less than cut off allow integration (but
240 * the integral is limited to max power)
241 */
242 i = mul_frac(tz->tzp->k_i, params->err_integral);
243
244 if (err < int_to_frac(tz->tzp->integral_cutoff)) {
245 s64 i_next = i + mul_frac(tz->tzp->k_i, err);
246
247 if (abs(i_next) < max_power_frac) {
248 i = i_next;
249 params->err_integral += err;
250 }
251 }
252
253 /*
254 * Calculate the derivative term
255 *
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)
259 */
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;
263
264 power_range = p + i + d;
265
266 /* feed-forward the known sustainable dissipatable power */
267 power_range = sustainable_power + frac_to_int(power_range);
268
269 power_range = clamp(power_range, (s64)0, (s64)max_allocatable_power);
270
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);
275
276 return power_range;
277}
278
279/**
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
284 *
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.
287 *
288 * Return: 0 on success, -EINVAL if the cooling device does not
289 * implement the power actor API or -E* for other failures.
290 */
291static int
292power_actor_set_power(struct thermal_cooling_device *cdev,
293 struct thermal_instance *instance, u32 power)
294{
295 unsigned long state;
296 int ret;
297
298 ret = cdev->ops->power2state(cdev, power, &state);
299 if (ret)
300 return ret;
301
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);
306
307 return 0;
308}
309
310/**
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
320 * to the actors
321 *
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
328 * @power_range.
329 *
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.
333 *
334 * Granted power for each actor is written to @granted_power, which
335 * should've been allocated by the calling function.
336 */
337static 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)
340{
341 u32 extra_power, capped_extra_power;
342 int i;
343
344 /*
345 * Prevent division by 0 if none of the actors request power.
346 */
347 if (!total_req_power)
348 total_req_power = 1;
349
350 capped_extra_power = 0;
351 extra_power = 0;
352 for (i = 0; i < num_actors; i++) {
353 u64 req_range = (u64)req_power[i] * power_range;
354
355 granted_power[i] = DIV_ROUND_CLOSEST_ULL(req_range,
356 total_req_power);
357
358 if (granted_power[i] > max_power[i]) {
359 extra_power += granted_power[i] - max_power[i];
360 granted_power[i] = max_power[i];
361 }
362
363 extra_actor_power[i] = max_power[i] - granted_power[i];
364 capped_extra_power += extra_actor_power[i];
365 }
366
367 if (!extra_power)
368 return;
369
370 /*
371 * Re-divvy the reclaimed extra among actors based on
372 * how far they are from the max
373 */
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,
379 capped_extra_power);
380 }
381}
382
383static int allocate_power(struct thermal_zone_device *tz,
384 int control_temp)
385{
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;
394
395 mutex_lock(&tz->lock);
396
397 num_actors = 0;
398 total_weight = 0;
399 list_for_each_entry(instance, &tz->thermal_instances, tz_node) {
400 if ((instance->trip == trip_max_desired_temperature) &&
401 cdev_is_power_actor(instance->cdev)) {
402 num_actors++;
403 total_weight += instance->weight;
404 }
405 }
406
407 if (!num_actors) {
408 ret = -ENODEV;
409 goto unlock;
410 }
411
412 /*
413 * We need to allocate five arrays of the same size:
414 * req_power, max_power, granted_power, extra_actor_power and
415 * weighted_req_power. They are going to be needed until this
416 * function returns. Allocate them all in one go to simplify
417 * the allocation and deallocation logic.
418 */
419 BUILD_BUG_ON(sizeof(*req_power) != sizeof(*max_power));
420 BUILD_BUG_ON(sizeof(*req_power) != sizeof(*granted_power));
421 BUILD_BUG_ON(sizeof(*req_power) != sizeof(*extra_actor_power));
422 BUILD_BUG_ON(sizeof(*req_power) != sizeof(*weighted_req_power));
423 req_power = kcalloc(num_actors * 5, sizeof(*req_power), GFP_KERNEL);
424 if (!req_power) {
425 ret = -ENOMEM;
426 goto unlock;
427 }
428
429 max_power = &req_power[num_actors];
430 granted_power = &req_power[2 * num_actors];
431 extra_actor_power = &req_power[3 * num_actors];
432 weighted_req_power = &req_power[4 * num_actors];
433
434 i = 0;
435 total_weighted_req_power = 0;
436 total_req_power = 0;
437 max_allocatable_power = 0;
438
439 list_for_each_entry(instance, &tz->thermal_instances, tz_node) {
440 int weight;
441 struct thermal_cooling_device *cdev = instance->cdev;
442
443 if (instance->trip != trip_max_desired_temperature)
444 continue;
445
446 if (!cdev_is_power_actor(cdev))
447 continue;
448
449 if (cdev->ops->get_requested_power(cdev, &req_power[i]))
450 continue;
451
452 if (!total_weight)
453 weight = 1 << FRAC_BITS;
454 else
455 weight = instance->weight;
456
457 weighted_req_power[i] = frac_to_int(weight * req_power[i]);
458
459 if (cdev->ops->state2power(cdev, instance->lower,
460 &max_power[i]))
461 continue;
462
463 total_req_power += req_power[i];
464 max_allocatable_power += max_power[i];
465 total_weighted_req_power += weighted_req_power[i];
466
467 i++;
468 }
469
470 power_range = pid_controller(tz, control_temp, max_allocatable_power);
471
472 divvy_up_power(weighted_req_power, max_power, num_actors,
473 total_weighted_req_power, power_range, granted_power,
474 extra_actor_power);
475
476 total_granted_power = 0;
477 i = 0;
478 list_for_each_entry(instance, &tz->thermal_instances, tz_node) {
479 if (instance->trip != trip_max_desired_temperature)
480 continue;
481
482 if (!cdev_is_power_actor(instance->cdev))
483 continue;
484
485 power_actor_set_power(instance->cdev, instance,
486 granted_power[i]);
487 total_granted_power += granted_power[i];
488
489 i++;
490 }
491
492 trace_thermal_power_allocator(tz, req_power, total_req_power,
493 granted_power, total_granted_power,
494 num_actors, power_range,
495 max_allocatable_power, tz->temperature,
496 control_temp - tz->temperature);
497
498 kfree(req_power);
499unlock:
500 mutex_unlock(&tz->lock);
501
502 return ret;
503}
504
505/**
506 * get_governor_trips() - get the number of the two trip points that are key for this governor
507 * @tz: thermal zone to operate on
508 * @params: pointer to private data for this governor
509 *
510 * The power allocator governor works optimally with two trips points:
511 * a "switch on" trip point and a "maximum desired temperature". These
512 * are defined as the first and last passive trip points.
513 *
514 * If there is only one trip point, then that's considered to be the
515 * "maximum desired temperature" trip point and the governor is always
516 * on. If there are no passive or active trip points, then the
517 * governor won't do anything. In fact, its throttle function
518 * won't be called at all.
519 */
520static void get_governor_trips(struct thermal_zone_device *tz,
521 struct power_allocator_params *params)
522{
523 int i, last_active, last_passive;
524 bool found_first_passive;
525
526 found_first_passive = false;
527 last_active = INVALID_TRIP;
528 last_passive = INVALID_TRIP;
529
530 for (i = 0; i < tz->trips; i++) {
531 enum thermal_trip_type type;
532 int ret;
533
534 ret = tz->ops->get_trip_type(tz, i, &type);
535 if (ret) {
536 dev_warn(&tz->device,
537 "Failed to get trip point %d type: %d\n", i,
538 ret);
539 continue;
540 }
541
542 if (type == THERMAL_TRIP_PASSIVE) {
543 if (!found_first_passive) {
544 params->trip_switch_on = i;
545 found_first_passive = true;
546 } else {
547 last_passive = i;
548 }
549 } else if (type == THERMAL_TRIP_ACTIVE) {
550 last_active = i;
551 } else {
552 break;
553 }
554 }
555
556 if (last_passive != INVALID_TRIP) {
557 params->trip_max_desired_temperature = last_passive;
558 } else if (found_first_passive) {
559 params->trip_max_desired_temperature = params->trip_switch_on;
560 params->trip_switch_on = INVALID_TRIP;
561 } else {
562 params->trip_switch_on = INVALID_TRIP;
563 params->trip_max_desired_temperature = last_active;
564 }
565}
566
567static void reset_pid_controller(struct power_allocator_params *params)
568{
569 params->err_integral = 0;
570 params->prev_err = 0;
571}
572
573static void allow_maximum_power(struct thermal_zone_device *tz, bool update)
574{
575 struct thermal_instance *instance;
576 struct power_allocator_params *params = tz->governor_data;
577 u32 req_power;
578
579 mutex_lock(&tz->lock);
580 list_for_each_entry(instance, &tz->thermal_instances, tz_node) {
581 struct thermal_cooling_device *cdev = instance->cdev;
582
583 if ((instance->trip != params->trip_max_desired_temperature) ||
584 (!cdev_is_power_actor(instance->cdev)))
585 continue;
586
587 instance->target = 0;
588 mutex_lock(&instance->cdev->lock);
589 /*
590 * Call for updating the cooling devices local stats and avoid
591 * periods of dozen of seconds when those have not been
592 * maintained.
593 */
594 cdev->ops->get_requested_power(cdev, &req_power);
595
596 if (update)
597 __thermal_cdev_update(instance->cdev);
598
599 mutex_unlock(&instance->cdev->lock);
600 }
601 mutex_unlock(&tz->lock);
602}
603
604/**
605 * check_power_actors() - Check all cooling devices and warn when they are
606 * not power actors
607 * @tz: thermal zone to operate on
608 *
609 * Check all cooling devices in the @tz and warn every time they are missing
610 * power actor API. The warning should help to investigate the issue, which
611 * could be e.g. lack of Energy Model for a given device.
612 *
613 * Return: 0 on success, -EINVAL if any cooling device does not implement
614 * the power actor API.
615 */
616static int check_power_actors(struct thermal_zone_device *tz)
617{
618 struct thermal_instance *instance;
619 int ret = 0;
620
621 list_for_each_entry(instance, &tz->thermal_instances, tz_node) {
622 if (!cdev_is_power_actor(instance->cdev)) {
623 dev_warn(&tz->device, "power_allocator: %s is not a power actor\n",
624 instance->cdev->type);
625 ret = -EINVAL;
626 }
627 }
628
629 return ret;
630}
631
632/**
633 * power_allocator_bind() - bind the power_allocator governor to a thermal zone
634 * @tz: thermal zone to bind it to
635 *
636 * Initialize the PID controller parameters and bind it to the thermal
637 * zone.
638 *
639 * Return: 0 on success, or -ENOMEM if we ran out of memory, or -EINVAL
640 * when there are unsupported cooling devices in the @tz.
641 */
642static int power_allocator_bind(struct thermal_zone_device *tz)
643{
644 int ret;
645 struct power_allocator_params *params;
646 int control_temp;
647
648 ret = check_power_actors(tz);
649 if (ret)
650 return ret;
651
652 params = kzalloc(sizeof(*params), GFP_KERNEL);
653 if (!params)
654 return -ENOMEM;
655
656 if (!tz->tzp) {
657 tz->tzp = kzalloc(sizeof(*tz->tzp), GFP_KERNEL);
658 if (!tz->tzp) {
659 ret = -ENOMEM;
660 goto free_params;
661 }
662
663 params->allocated_tzp = true;
664 }
665
666 if (!tz->tzp->sustainable_power)
667 dev_warn(&tz->device, "power_allocator: sustainable_power will be estimated\n");
668
669 get_governor_trips(tz, params);
670
671 if (tz->trips > 0) {
672 ret = tz->ops->get_trip_temp(tz,
673 params->trip_max_desired_temperature,
674 &control_temp);
675 if (!ret)
676 estimate_pid_constants(tz, tz->tzp->sustainable_power,
677 params->trip_switch_on,
678 control_temp);
679 }
680
681 reset_pid_controller(params);
682
683 tz->governor_data = params;
684
685 return 0;
686
687free_params:
688 kfree(params);
689
690 return ret;
691}
692
693static void power_allocator_unbind(struct thermal_zone_device *tz)
694{
695 struct power_allocator_params *params = tz->governor_data;
696
697 dev_dbg(&tz->device, "Unbinding from thermal zone %d\n", tz->id);
698
699 if (params->allocated_tzp) {
700 kfree(tz->tzp);
701 tz->tzp = NULL;
702 }
703
704 kfree(tz->governor_data);
705 tz->governor_data = NULL;
706}
707
708static int power_allocator_throttle(struct thermal_zone_device *tz, int trip)
709{
710 int ret;
711 int switch_on_temp, control_temp;
712 struct power_allocator_params *params = tz->governor_data;
713 bool update;
714
715 /*
716 * We get called for every trip point but we only need to do
717 * our calculations once
718 */
719 if (trip != params->trip_max_desired_temperature)
720 return 0;
721
722 ret = tz->ops->get_trip_temp(tz, params->trip_switch_on,
723 &switch_on_temp);
724 if (!ret && (tz->temperature < switch_on_temp)) {
725 update = (tz->last_temperature >= switch_on_temp);
726 tz->passive = 0;
727 reset_pid_controller(params);
728 allow_maximum_power(tz, update);
729 return 0;
730 }
731
732 tz->passive = 1;
733
734 ret = tz->ops->get_trip_temp(tz, params->trip_max_desired_temperature,
735 &control_temp);
736 if (ret) {
737 dev_warn(&tz->device,
738 "Failed to get the maximum desired temperature: %d\n",
739 ret);
740 return ret;
741 }
742
743 return allocate_power(tz, control_temp);
744}
745
746static struct thermal_governor thermal_gov_power_allocator = {
747 .name = "power_allocator",
748 .bind_to_tz = power_allocator_bind,
749 .unbind_from_tz = power_allocator_unbind,
750 .throttle = power_allocator_throttle,
751};
752THERMAL_GOVERNOR_DECLARE(thermal_gov_power_allocator);