1 /*
2 * A power allocator to manage temperature
3 *
4 * Copyright (C) 2014 ARM Ltd.
5 *
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 as
8 * published by the Free Software Foundation.
9 *
10 * This program is distributed "as is" WITHOUT ANY WARRANTY of any
11 * kind, whether express or implied; without even the implied warranty
12 * of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
13 * GNU General Public License for more details.
14 */
15
16 #define pr_fmt(fmt) "Power allocator: " fmt
17
18 #include <linux/rculist.h>
19 #include <linux/slab.h>
20 #include <linux/thermal.h>
21
22 #define CREATE_TRACE_POINTS
23 #include <trace/events/thermal_power_allocator.h>
24
25 #include "thermal_core.h"
26
27 #define INVALID_TRIP -1
28
29 #define FRAC_BITS 10
30 #define int_to_frac(x) ((x) << FRAC_BITS)
31 #define frac_to_int(x) ((x) >> FRAC_BITS)
32
33 /**
34 * mul_frac() - multiply two fixed-point numbers
35 * @x: first multiplicand
36 * @y: second multiplicand
37 *
38 * Return: the result of multiplying two fixed-point numbers. The
39 * result is also a fixed-point number.
40 */
mul_frac(s64 x,s64 y)41 static inline s64 mul_frac(s64 x, s64 y)
42 {
43 return (x * y) >> FRAC_BITS;
44 }
45
46 /**
47 * div_frac() - divide two fixed-point numbers
48 * @x: the dividend
49 * @y: the divisor
50 *
51 * Return: the result of dividing two fixed-point numbers. The
52 * result is also a fixed-point number.
53 */
div_frac(s64 x,s64 y)54 static inline s64 div_frac(s64 x, s64 y)
55 {
56 return div_s64(x << FRAC_BITS, y);
57 }
58
59 /**
60 * struct power_allocator_params - parameters for the power allocator governor
61 * @allocated_tzp: whether we have allocated tzp for this thermal zone and
62 * it needs to be freed on unbind
63 * @err_integral: accumulated error in the PID controller.
64 * @prev_err: error in the previous iteration of the PID controller.
65 * Used to calculate the derivative term.
66 * @trip_switch_on: first passive trip point of the thermal zone. The
67 * governor switches on when this trip point is crossed.
68 * If the thermal zone only has one passive trip point,
69 * @trip_switch_on should be INVALID_TRIP.
70 * @trip_max_desired_temperature: last passive trip point of the thermal
71 * zone. The temperature we are
72 * controlling for.
73 */
74 struct power_allocator_params {
75 bool allocated_tzp;
76 s64 err_integral;
77 s32 prev_err;
78 int trip_switch_on;
79 int trip_max_desired_temperature;
80 };
81
82 /**
83 * estimate_sustainable_power() - Estimate the sustainable power of a thermal zone
84 * @tz: thermal zone we are operating in
85 *
86 * For thermal zones that don't provide a sustainable_power in their
87 * thermal_zone_params, estimate one. Calculate it using the minimum
88 * power of all the cooling devices as that gives a valid value that
89 * can give some degree of functionality. For optimal performance of
90 * this governor, provide a sustainable_power in the thermal zone's
91 * thermal_zone_params.
92 */
estimate_sustainable_power(struct thermal_zone_device * tz)93 static u32 estimate_sustainable_power(struct thermal_zone_device *tz)
94 {
95 u32 sustainable_power = 0;
96 struct thermal_instance *instance;
97 struct power_allocator_params *params = tz->governor_data;
98
99 list_for_each_entry(instance, &tz->thermal_instances, tz_node) {
100 struct thermal_cooling_device *cdev = instance->cdev;
101 u32 min_power;
102
103 if (instance->trip != params->trip_max_desired_temperature)
104 continue;
105
106 if (power_actor_get_min_power(cdev, tz, &min_power))
107 continue;
108
109 sustainable_power += min_power;
110 }
111
112 return sustainable_power;
113 }
114
115 /**
116 * estimate_pid_constants() - Estimate the constants for the PID controller
117 * @tz: thermal zone for which to estimate the constants
118 * @sustainable_power: sustainable power for the thermal zone
119 * @trip_switch_on: trip point number for the switch on temperature
120 * @control_temp: target temperature for the power allocator governor
121 * @force: whether to force the update of the constants
122 *
123 * This function is used to update the estimation of the PID
124 * controller constants in struct thermal_zone_parameters.
125 * Sustainable power is provided in case it was estimated. The
126 * estimated sustainable_power should not be stored in the
127 * thermal_zone_parameters so it has to be passed explicitly to this
128 * function.
129 *
130 * If @force is not set, the values in the thermal zone's parameters
131 * are preserved if they are not zero. If @force is set, the values
132 * in thermal zone's parameters are overwritten.
133 */
estimate_pid_constants(struct thermal_zone_device * tz,u32 sustainable_power,int trip_switch_on,int control_temp,bool force)134 static void estimate_pid_constants(struct thermal_zone_device *tz,
135 u32 sustainable_power, int trip_switch_on,
136 int control_temp, bool force)
137 {
138 int ret;
139 int switch_on_temp;
140 u32 temperature_threshold;
141
142 ret = tz->ops->get_trip_temp(tz, trip_switch_on, &switch_on_temp);
143 if (ret)
144 switch_on_temp = 0;
145
146 temperature_threshold = control_temp - switch_on_temp;
147 /*
148 * estimate_pid_constants() tries to find appropriate default
149 * values for thermal zones that don't provide them. If a
150 * system integrator has configured a thermal zone with two
151 * passive trip points at the same temperature, that person
152 * hasn't put any effort to set up the thermal zone properly
153 * so just give up.
154 */
155 if (!temperature_threshold)
156 return;
157
158 if (!tz->tzp->k_po || force)
159 tz->tzp->k_po = int_to_frac(sustainable_power) /
160 temperature_threshold;
161
162 if (!tz->tzp->k_pu || force)
163 tz->tzp->k_pu = int_to_frac(2 * sustainable_power) /
164 temperature_threshold;
165
166 if (!tz->tzp->k_i || force)
167 tz->tzp->k_i = int_to_frac(10) / 1000;
168 /*
169 * The default for k_d and integral_cutoff is 0, so we can
170 * leave them as they are.
171 */
172 }
173
174 /**
175 * pid_controller() - PID controller
176 * @tz: thermal zone we are operating in
177 * @control_temp: the target temperature in millicelsius
178 * @max_allocatable_power: maximum allocatable power for this thermal zone
179 *
180 * This PID controller increases the available power budget so that the
181 * temperature of the thermal zone gets as close as possible to
182 * @control_temp and limits the power if it exceeds it. k_po is the
183 * proportional term when we are overshooting, k_pu is the
184 * proportional term when we are undershooting. integral_cutoff is a
185 * threshold below which we stop accumulating the error. The
186 * accumulated error is only valid if the requested power will make
187 * the system warmer. If the system is mostly idle, there's no point
188 * in accumulating positive error.
189 *
190 * Return: The power budget for the next period.
191 */
pid_controller(struct thermal_zone_device * tz,int control_temp,u32 max_allocatable_power)192 static u32 pid_controller(struct thermal_zone_device *tz,
193 int control_temp,
194 u32 max_allocatable_power)
195 {
196 s64 p, i, d, power_range;
197 s32 err, max_power_frac;
198 u32 sustainable_power;
199 struct power_allocator_params *params = tz->governor_data;
200
201 max_power_frac = int_to_frac(max_allocatable_power);
202
203 if (tz->tzp->sustainable_power) {
204 sustainable_power = tz->tzp->sustainable_power;
205 } else {
206 sustainable_power = estimate_sustainable_power(tz);
207 estimate_pid_constants(tz, sustainable_power,
208 params->trip_switch_on, control_temp,
209 true);
210 }
211
212 err = control_temp - tz->temperature;
213 err = int_to_frac(err);
214
215 /* Calculate the proportional term */
216 p = mul_frac(err < 0 ? tz->tzp->k_po : tz->tzp->k_pu, err);
217
218 /*
219 * Calculate the integral term
220 *
221 * if the error is less than cut off allow integration (but
222 * the integral is limited to max power)
223 */
224 i = mul_frac(tz->tzp->k_i, params->err_integral);
225
226 if (err < int_to_frac(tz->tzp->integral_cutoff)) {
227 s64 i_next = i + mul_frac(tz->tzp->k_i, err);
228
229 if (abs(i_next) < max_power_frac) {
230 i = i_next;
231 params->err_integral += err;
232 }
233 }
234
235 /*
236 * Calculate the derivative term
237 *
238 * We do err - prev_err, so with a positive k_d, a decreasing
239 * error (i.e. driving closer to the line) results in less
240 * power being applied, slowing down the controller)
241 */
242 d = mul_frac(tz->tzp->k_d, err - params->prev_err);
243 d = div_frac(d, tz->passive_delay);
244 params->prev_err = err;
245
246 power_range = p + i + d;
247
248 /* feed-forward the known sustainable dissipatable power */
249 power_range = sustainable_power + frac_to_int(power_range);
250
251 power_range = clamp(power_range, (s64)0, (s64)max_allocatable_power);
252
253 trace_thermal_power_allocator_pid(tz, frac_to_int(err),
254 frac_to_int(params->err_integral),
255 frac_to_int(p), frac_to_int(i),
256 frac_to_int(d), power_range);
257
258 return power_range;
259 }
260
261 /**
262 * divvy_up_power() - divvy the allocated power between the actors
263 * @req_power: each actor's requested power
264 * @max_power: each actor's maximum available power
265 * @num_actors: size of the @req_power, @max_power and @granted_power's array
266 * @total_req_power: sum of @req_power
267 * @power_range: total allocated power
268 * @granted_power: output array: each actor's granted power
269 * @extra_actor_power: an appropriately sized array to be used in the
270 * function as temporary storage of the extra power given
271 * to the actors
272 *
273 * This function divides the total allocated power (@power_range)
274 * fairly between the actors. It first tries to give each actor a
275 * share of the @power_range according to how much power it requested
276 * compared to the rest of the actors. For example, if only one actor
277 * requests power, then it receives all the @power_range. If
278 * three actors each requests 1mW, each receives a third of the
279 * @power_range.
280 *
281 * If any actor received more than their maximum power, then that
282 * surplus is re-divvied among the actors based on how far they are
283 * from their respective maximums.
284 *
285 * Granted power for each actor is written to @granted_power, which
286 * should've been allocated by the calling function.
287 */
divvy_up_power(u32 * req_power,u32 * max_power,int num_actors,u32 total_req_power,u32 power_range,u32 * granted_power,u32 * extra_actor_power)288 static void divvy_up_power(u32 *req_power, u32 *max_power, int num_actors,
289 u32 total_req_power, u32 power_range,
290 u32 *granted_power, u32 *extra_actor_power)
291 {
292 u32 extra_power, capped_extra_power;
293 int i;
294
295 /*
296 * Prevent division by 0 if none of the actors request power.
297 */
298 if (!total_req_power)
299 total_req_power = 1;
300
301 capped_extra_power = 0;
302 extra_power = 0;
303 for (i = 0; i < num_actors; i++) {
304 u64 req_range = (u64)req_power[i] * power_range;
305
306 granted_power[i] = DIV_ROUND_CLOSEST_ULL(req_range,
307 total_req_power);
308
309 if (granted_power[i] > max_power[i]) {
310 extra_power += granted_power[i] - max_power[i];
311 granted_power[i] = max_power[i];
312 }
313
314 extra_actor_power[i] = max_power[i] - granted_power[i];
315 capped_extra_power += extra_actor_power[i];
316 }
317
318 if (!extra_power)
319 return;
320
321 /*
322 * Re-divvy the reclaimed extra among actors based on
323 * how far they are from the max
324 */
325 extra_power = min(extra_power, capped_extra_power);
326 if (capped_extra_power > 0)
327 for (i = 0; i < num_actors; i++)
328 granted_power[i] += (extra_actor_power[i] *
329 extra_power) / capped_extra_power;
330 }
331
allocate_power(struct thermal_zone_device * tz,int control_temp)332 static int allocate_power(struct thermal_zone_device *tz,
333 int control_temp)
334 {
335 struct thermal_instance *instance;
336 struct power_allocator_params *params = tz->governor_data;
337 u32 *req_power, *max_power, *granted_power, *extra_actor_power;
338 u32 *weighted_req_power;
339 u32 total_req_power, max_allocatable_power, total_weighted_req_power;
340 u32 total_granted_power, power_range;
341 int i, num_actors, total_weight, ret = 0;
342 int trip_max_desired_temperature = params->trip_max_desired_temperature;
343
344 mutex_lock(&tz->lock);
345
346 num_actors = 0;
347 total_weight = 0;
348 list_for_each_entry(instance, &tz->thermal_instances, tz_node) {
349 if ((instance->trip == trip_max_desired_temperature) &&
350 cdev_is_power_actor(instance->cdev)) {
351 num_actors++;
352 total_weight += instance->weight;
353 }
354 }
355
356 if (!num_actors) {
357 ret = -ENODEV;
358 goto unlock;
359 }
360
361 /*
362 * We need to allocate five arrays of the same size:
363 * req_power, max_power, granted_power, extra_actor_power and
364 * weighted_req_power. They are going to be needed until this
365 * function returns. Allocate them all in one go to simplify
366 * the allocation and deallocation logic.
367 */
368 BUILD_BUG_ON(sizeof(*req_power) != sizeof(*max_power));
369 BUILD_BUG_ON(sizeof(*req_power) != sizeof(*granted_power));
370 BUILD_BUG_ON(sizeof(*req_power) != sizeof(*extra_actor_power));
371 BUILD_BUG_ON(sizeof(*req_power) != sizeof(*weighted_req_power));
372 req_power = kcalloc(num_actors * 5, sizeof(*req_power), GFP_KERNEL);
373 if (!req_power) {
374 ret = -ENOMEM;
375 goto unlock;
376 }
377
378 max_power = &req_power[num_actors];
379 granted_power = &req_power[2 * num_actors];
380 extra_actor_power = &req_power[3 * num_actors];
381 weighted_req_power = &req_power[4 * num_actors];
382
383 i = 0;
384 total_weighted_req_power = 0;
385 total_req_power = 0;
386 max_allocatable_power = 0;
387
388 list_for_each_entry(instance, &tz->thermal_instances, tz_node) {
389 int weight;
390 struct thermal_cooling_device *cdev = instance->cdev;
391
392 if (instance->trip != trip_max_desired_temperature)
393 continue;
394
395 if (!cdev_is_power_actor(cdev))
396 continue;
397
398 if (cdev->ops->get_requested_power(cdev, tz, &req_power[i]))
399 continue;
400
401 if (!total_weight)
402 weight = 1 << FRAC_BITS;
403 else
404 weight = instance->weight;
405
406 weighted_req_power[i] = frac_to_int(weight * req_power[i]);
407
408 if (power_actor_get_max_power(cdev, tz, &max_power[i]))
409 continue;
410
411 total_req_power += req_power[i];
412 max_allocatable_power += max_power[i];
413 total_weighted_req_power += weighted_req_power[i];
414
415 i++;
416 }
417
418 power_range = pid_controller(tz, control_temp, max_allocatable_power);
419
420 divvy_up_power(weighted_req_power, max_power, num_actors,
421 total_weighted_req_power, power_range, granted_power,
422 extra_actor_power);
423
424 total_granted_power = 0;
425 i = 0;
426 list_for_each_entry(instance, &tz->thermal_instances, tz_node) {
427 if (instance->trip != trip_max_desired_temperature)
428 continue;
429
430 if (!cdev_is_power_actor(instance->cdev))
431 continue;
432
433 power_actor_set_power(instance->cdev, instance,
434 granted_power[i]);
435 total_granted_power += granted_power[i];
436
437 i++;
438 }
439
440 trace_thermal_power_allocator(tz, req_power, total_req_power,
441 granted_power, total_granted_power,
442 num_actors, power_range,
443 max_allocatable_power, tz->temperature,
444 control_temp - tz->temperature);
445
446 kfree(req_power);
447 unlock:
448 mutex_unlock(&tz->lock);
449
450 return ret;
451 }
452
453 /**
454 * get_governor_trips() - get the number of the two trip points that are key for this governor
455 * @tz: thermal zone to operate on
456 * @params: pointer to private data for this governor
457 *
458 * The power allocator governor works optimally with two trips points:
459 * a "switch on" trip point and a "maximum desired temperature". These
460 * are defined as the first and last passive trip points.
461 *
462 * If there is only one trip point, then that's considered to be the
463 * "maximum desired temperature" trip point and the governor is always
464 * on. If there are no passive or active trip points, then the
465 * governor won't do anything. In fact, its throttle function
466 * won't be called at all.
467 */
get_governor_trips(struct thermal_zone_device * tz,struct power_allocator_params * params)468 static void get_governor_trips(struct thermal_zone_device *tz,
469 struct power_allocator_params *params)
470 {
471 int i, last_active, last_passive;
472 bool found_first_passive;
473
474 found_first_passive = false;
475 last_active = INVALID_TRIP;
476 last_passive = INVALID_TRIP;
477
478 for (i = 0; i < tz->trips; i++) {
479 enum thermal_trip_type type;
480 int ret;
481
482 ret = tz->ops->get_trip_type(tz, i, &type);
483 if (ret) {
484 dev_warn(&tz->device,
485 "Failed to get trip point %d type: %d\n", i,
486 ret);
487 continue;
488 }
489
490 if (type == THERMAL_TRIP_PASSIVE) {
491 if (!found_first_passive) {
492 params->trip_switch_on = i;
493 found_first_passive = true;
494 } else {
495 last_passive = i;
496 }
497 } else if (type == THERMAL_TRIP_ACTIVE) {
498 last_active = i;
499 } else {
500 break;
501 }
502 }
503
504 if (last_passive != INVALID_TRIP) {
505 params->trip_max_desired_temperature = last_passive;
506 } else if (found_first_passive) {
507 params->trip_max_desired_temperature = params->trip_switch_on;
508 params->trip_switch_on = INVALID_TRIP;
509 } else {
510 params->trip_switch_on = INVALID_TRIP;
511 params->trip_max_desired_temperature = last_active;
512 }
513 }
514
reset_pid_controller(struct power_allocator_params * params)515 static void reset_pid_controller(struct power_allocator_params *params)
516 {
517 params->err_integral = 0;
518 params->prev_err = 0;
519 }
520
allow_maximum_power(struct thermal_zone_device * tz)521 static void allow_maximum_power(struct thermal_zone_device *tz)
522 {
523 struct thermal_instance *instance;
524 struct power_allocator_params *params = tz->governor_data;
525
526 mutex_lock(&tz->lock);
527 list_for_each_entry(instance, &tz->thermal_instances, tz_node) {
528 if ((instance->trip != params->trip_max_desired_temperature) ||
529 (!cdev_is_power_actor(instance->cdev)))
530 continue;
531
532 instance->target = 0;
533 mutex_lock(&instance->cdev->lock);
534 instance->cdev->updated = false;
535 mutex_unlock(&instance->cdev->lock);
536 thermal_cdev_update(instance->cdev);
537 }
538 mutex_unlock(&tz->lock);
539 }
540
541 /**
542 * power_allocator_bind() - bind the power_allocator governor to a thermal zone
543 * @tz: thermal zone to bind it to
544 *
545 * Initialize the PID controller parameters and bind it to the thermal
546 * zone.
547 *
548 * Return: 0 on success, or -ENOMEM if we ran out of memory.
549 */
power_allocator_bind(struct thermal_zone_device * tz)550 static int power_allocator_bind(struct thermal_zone_device *tz)
551 {
552 int ret;
553 struct power_allocator_params *params;
554 int control_temp;
555
556 params = kzalloc(sizeof(*params), GFP_KERNEL);
557 if (!params)
558 return -ENOMEM;
559
560 if (!tz->tzp) {
561 tz->tzp = kzalloc(sizeof(*tz->tzp), GFP_KERNEL);
562 if (!tz->tzp) {
563 ret = -ENOMEM;
564 goto free_params;
565 }
566
567 params->allocated_tzp = true;
568 }
569
570 if (!tz->tzp->sustainable_power)
571 dev_warn(&tz->device, "power_allocator: sustainable_power will be estimated\n");
572
573 get_governor_trips(tz, params);
574
575 if (tz->trips > 0) {
576 ret = tz->ops->get_trip_temp(tz,
577 params->trip_max_desired_temperature,
578 &control_temp);
579 if (!ret)
580 estimate_pid_constants(tz, tz->tzp->sustainable_power,
581 params->trip_switch_on,
582 control_temp, false);
583 }
584
585 reset_pid_controller(params);
586
587 tz->governor_data = params;
588
589 return 0;
590
591 free_params:
592 kfree(params);
593
594 return ret;
595 }
596
power_allocator_unbind(struct thermal_zone_device * tz)597 static void power_allocator_unbind(struct thermal_zone_device *tz)
598 {
599 struct power_allocator_params *params = tz->governor_data;
600
601 dev_dbg(&tz->device, "Unbinding from thermal zone %d\n", tz->id);
602
603 if (params->allocated_tzp) {
604 kfree(tz->tzp);
605 tz->tzp = NULL;
606 }
607
608 kfree(tz->governor_data);
609 tz->governor_data = NULL;
610 }
611
power_allocator_throttle(struct thermal_zone_device * tz,int trip)612 static int power_allocator_throttle(struct thermal_zone_device *tz, int trip)
613 {
614 int ret;
615 int switch_on_temp, control_temp;
616 struct power_allocator_params *params = tz->governor_data;
617
618 /*
619 * We get called for every trip point but we only need to do
620 * our calculations once
621 */
622 if (trip != params->trip_max_desired_temperature)
623 return 0;
624
625 ret = tz->ops->get_trip_temp(tz, params->trip_switch_on,
626 &switch_on_temp);
627 if (!ret && (tz->temperature < switch_on_temp)) {
628 tz->passive = 0;
629 reset_pid_controller(params);
630 allow_maximum_power(tz);
631 return 0;
632 }
633
634 tz->passive = 1;
635
636 ret = tz->ops->get_trip_temp(tz, params->trip_max_desired_temperature,
637 &control_temp);
638 if (ret) {
639 dev_warn(&tz->device,
640 "Failed to get the maximum desired temperature: %d\n",
641 ret);
642 return ret;
643 }
644
645 return allocate_power(tz, control_temp);
646 }
647
648 static struct thermal_governor thermal_gov_power_allocator = {
649 .name = "power_allocator",
650 .bind_to_tz = power_allocator_bind,
651 .unbind_from_tz = power_allocator_unbind,
652 .throttle = power_allocator_throttle,
653 };
654 THERMAL_GOVERNOR_DECLARE(thermal_gov_power_allocator);
655