linux/drivers/thermal/power_allocator.c
<<
>>
Prefs
   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 */
  41static 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 */
  54static 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 */
  74struct 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 */
  93static 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 */
 134static 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 * @current_temp:       the current temperature in millicelsius
 178 * @control_temp:       the target temperature in millicelsius
 179 * @max_allocatable_power:      maximum allocatable power for this thermal zone
 180 *
 181 * This PID controller increases the available power budget so that the
 182 * temperature of the thermal zone gets as close as possible to
 183 * @control_temp and limits the power if it exceeds it.  k_po is the
 184 * proportional term when we are overshooting, k_pu is the
 185 * proportional term when we are undershooting.  integral_cutoff is a
 186 * threshold below which we stop accumulating the error.  The
 187 * accumulated error is only valid if the requested power will make
 188 * the system warmer.  If the system is mostly idle, there's no point
 189 * in accumulating positive error.
 190 *
 191 * Return: The power budget for the next period.
 192 */
 193static u32 pid_controller(struct thermal_zone_device *tz,
 194                          int current_temp,
 195                          int control_temp,
 196                          u32 max_allocatable_power)
 197{
 198        s64 p, i, d, power_range;
 199        s32 err, max_power_frac;
 200        u32 sustainable_power;
 201        struct power_allocator_params *params = tz->governor_data;
 202
 203        max_power_frac = int_to_frac(max_allocatable_power);
 204
 205        if (tz->tzp->sustainable_power) {
 206                sustainable_power = tz->tzp->sustainable_power;
 207        } else {
 208                sustainable_power = estimate_sustainable_power(tz);
 209                estimate_pid_constants(tz, sustainable_power,
 210                                       params->trip_switch_on, control_temp,
 211                                       true);
 212        }
 213
 214        err = control_temp - current_temp;
 215        err = int_to_frac(err);
 216
 217        /* Calculate the proportional term */
 218        p = mul_frac(err < 0 ? tz->tzp->k_po : tz->tzp->k_pu, err);
 219
 220        /*
 221         * Calculate the integral term
 222         *
 223         * if the error is less than cut off allow integration (but
 224         * the integral is limited to max power)
 225         */
 226        i = mul_frac(tz->tzp->k_i, params->err_integral);
 227
 228        if (err < int_to_frac(tz->tzp->integral_cutoff)) {
 229                s64 i_next = i + mul_frac(tz->tzp->k_i, err);
 230
 231                if (abs64(i_next) < max_power_frac) {
 232                        i = i_next;
 233                        params->err_integral += err;
 234                }
 235        }
 236
 237        /*
 238         * Calculate the derivative term
 239         *
 240         * We do err - prev_err, so with a positive k_d, a decreasing
 241         * error (i.e. driving closer to the line) results in less
 242         * power being applied, slowing down the controller)
 243         */
 244        d = mul_frac(tz->tzp->k_d, err - params->prev_err);
 245        d = div_frac(d, tz->passive_delay);
 246        params->prev_err = err;
 247
 248        power_range = p + i + d;
 249
 250        /* feed-forward the known sustainable dissipatable power */
 251        power_range = sustainable_power + frac_to_int(power_range);
 252
 253        power_range = clamp(power_range, (s64)0, (s64)max_allocatable_power);
 254
 255        trace_thermal_power_allocator_pid(tz, frac_to_int(err),
 256                                          frac_to_int(params->err_integral),
 257                                          frac_to_int(p), frac_to_int(i),
 258                                          frac_to_int(d), power_range);
 259
 260        return power_range;
 261}
 262
 263/**
 264 * divvy_up_power() - divvy the allocated power between the actors
 265 * @req_power:  each actor's requested power
 266 * @max_power:  each actor's maximum available power
 267 * @num_actors: size of the @req_power, @max_power and @granted_power's array
 268 * @total_req_power: sum of @req_power
 269 * @power_range:        total allocated power
 270 * @granted_power:      output array: each actor's granted power
 271 * @extra_actor_power:  an appropriately sized array to be used in the
 272 *                      function as temporary storage of the extra power given
 273 *                      to the actors
 274 *
 275 * This function divides the total allocated power (@power_range)
 276 * fairly between the actors.  It first tries to give each actor a
 277 * share of the @power_range according to how much power it requested
 278 * compared to the rest of the actors.  For example, if only one actor
 279 * requests power, then it receives all the @power_range.  If
 280 * three actors each requests 1mW, each receives a third of the
 281 * @power_range.
 282 *
 283 * If any actor received more than their maximum power, then that
 284 * surplus is re-divvied among the actors based on how far they are
 285 * from their respective maximums.
 286 *
 287 * Granted power for each actor is written to @granted_power, which
 288 * should've been allocated by the calling function.
 289 */
 290static void divvy_up_power(u32 *req_power, u32 *max_power, int num_actors,
 291                           u32 total_req_power, u32 power_range,
 292                           u32 *granted_power, u32 *extra_actor_power)
 293{
 294        u32 extra_power, capped_extra_power;
 295        int i;
 296
 297        /*
 298         * Prevent division by 0 if none of the actors request power.
 299         */
 300        if (!total_req_power)
 301                total_req_power = 1;
 302
 303        capped_extra_power = 0;
 304        extra_power = 0;
 305        for (i = 0; i < num_actors; i++) {
 306                u64 req_range = req_power[i] * power_range;
 307
 308                granted_power[i] = DIV_ROUND_CLOSEST_ULL(req_range,
 309                                                         total_req_power);
 310
 311                if (granted_power[i] > max_power[i]) {
 312                        extra_power += granted_power[i] - max_power[i];
 313                        granted_power[i] = max_power[i];
 314                }
 315
 316                extra_actor_power[i] = max_power[i] - granted_power[i];
 317                capped_extra_power += extra_actor_power[i];
 318        }
 319
 320        if (!extra_power)
 321                return;
 322
 323        /*
 324         * Re-divvy the reclaimed extra among actors based on
 325         * how far they are from the max
 326         */
 327        extra_power = min(extra_power, capped_extra_power);
 328        if (capped_extra_power > 0)
 329                for (i = 0; i < num_actors; i++)
 330                        granted_power[i] += (extra_actor_power[i] *
 331                                        extra_power) / capped_extra_power;
 332}
 333
 334static int allocate_power(struct thermal_zone_device *tz,
 335                          int current_temp,
 336                          int control_temp)
 337{
 338        struct thermal_instance *instance;
 339        struct power_allocator_params *params = tz->governor_data;
 340        u32 *req_power, *max_power, *granted_power, *extra_actor_power;
 341        u32 *weighted_req_power;
 342        u32 total_req_power, max_allocatable_power, total_weighted_req_power;
 343        u32 total_granted_power, power_range;
 344        int i, num_actors, total_weight, ret = 0;
 345        int trip_max_desired_temperature = params->trip_max_desired_temperature;
 346
 347        mutex_lock(&tz->lock);
 348
 349        num_actors = 0;
 350        total_weight = 0;
 351        list_for_each_entry(instance, &tz->thermal_instances, tz_node) {
 352                if ((instance->trip == trip_max_desired_temperature) &&
 353                    cdev_is_power_actor(instance->cdev)) {
 354                        num_actors++;
 355                        total_weight += instance->weight;
 356                }
 357        }
 358
 359        if (!num_actors) {
 360                ret = -ENODEV;
 361                goto unlock;
 362        }
 363
 364        /*
 365         * We need to allocate five arrays of the same size:
 366         * req_power, max_power, granted_power, extra_actor_power and
 367         * weighted_req_power.  They are going to be needed until this
 368         * function returns.  Allocate them all in one go to simplify
 369         * the allocation and deallocation logic.
 370         */
 371        BUILD_BUG_ON(sizeof(*req_power) != sizeof(*max_power));
 372        BUILD_BUG_ON(sizeof(*req_power) != sizeof(*granted_power));
 373        BUILD_BUG_ON(sizeof(*req_power) != sizeof(*extra_actor_power));
 374        BUILD_BUG_ON(sizeof(*req_power) != sizeof(*weighted_req_power));
 375        req_power = kcalloc(num_actors * 5, sizeof(*req_power), GFP_KERNEL);
 376        if (!req_power) {
 377                ret = -ENOMEM;
 378                goto unlock;
 379        }
 380
 381        max_power = &req_power[num_actors];
 382        granted_power = &req_power[2 * num_actors];
 383        extra_actor_power = &req_power[3 * num_actors];
 384        weighted_req_power = &req_power[4 * num_actors];
 385
 386        i = 0;
 387        total_weighted_req_power = 0;
 388        total_req_power = 0;
 389        max_allocatable_power = 0;
 390
 391        list_for_each_entry(instance, &tz->thermal_instances, tz_node) {
 392                int weight;
 393                struct thermal_cooling_device *cdev = instance->cdev;
 394
 395                if (instance->trip != trip_max_desired_temperature)
 396                        continue;
 397
 398                if (!cdev_is_power_actor(cdev))
 399                        continue;
 400
 401                if (cdev->ops->get_requested_power(cdev, tz, &req_power[i]))
 402                        continue;
 403
 404                if (!total_weight)
 405                        weight = 1 << FRAC_BITS;
 406                else
 407                        weight = instance->weight;
 408
 409                weighted_req_power[i] = frac_to_int(weight * req_power[i]);
 410
 411                if (power_actor_get_max_power(cdev, tz, &max_power[i]))
 412                        continue;
 413
 414                total_req_power += req_power[i];
 415                max_allocatable_power += max_power[i];
 416                total_weighted_req_power += weighted_req_power[i];
 417
 418                i++;
 419        }
 420
 421        power_range = pid_controller(tz, current_temp, control_temp,
 422                                     max_allocatable_power);
 423
 424        divvy_up_power(weighted_req_power, max_power, num_actors,
 425                       total_weighted_req_power, power_range, granted_power,
 426                       extra_actor_power);
 427
 428        total_granted_power = 0;
 429        i = 0;
 430        list_for_each_entry(instance, &tz->thermal_instances, tz_node) {
 431                if (instance->trip != trip_max_desired_temperature)
 432                        continue;
 433
 434                if (!cdev_is_power_actor(instance->cdev))
 435                        continue;
 436
 437                power_actor_set_power(instance->cdev, instance,
 438                                      granted_power[i]);
 439                total_granted_power += granted_power[i];
 440
 441                i++;
 442        }
 443
 444        trace_thermal_power_allocator(tz, req_power, total_req_power,
 445                                      granted_power, total_granted_power,
 446                                      num_actors, power_range,
 447                                      max_allocatable_power, current_temp,
 448                                      control_temp - current_temp);
 449
 450        kfree(req_power);
 451unlock:
 452        mutex_unlock(&tz->lock);
 453
 454        return ret;
 455}
 456
 457/**
 458 * get_governor_trips() - get the number of the two trip points that are key for this governor
 459 * @tz: thermal zone to operate on
 460 * @params:     pointer to private data for this governor
 461 *
 462 * The power allocator governor works optimally with two trips points:
 463 * a "switch on" trip point and a "maximum desired temperature".  These
 464 * are defined as the first and last passive trip points.
 465 *
 466 * If there is only one trip point, then that's considered to be the
 467 * "maximum desired temperature" trip point and the governor is always
 468 * on.  If there are no passive or active trip points, then the
 469 * governor won't do anything.  In fact, its throttle function
 470 * won't be called at all.
 471 */
 472static void get_governor_trips(struct thermal_zone_device *tz,
 473                               struct power_allocator_params *params)
 474{
 475        int i, last_active, last_passive;
 476        bool found_first_passive;
 477
 478        found_first_passive = false;
 479        last_active = INVALID_TRIP;
 480        last_passive = INVALID_TRIP;
 481
 482        for (i = 0; i < tz->trips; i++) {
 483                enum thermal_trip_type type;
 484                int ret;
 485
 486                ret = tz->ops->get_trip_type(tz, i, &type);
 487                if (ret) {
 488                        dev_warn(&tz->device,
 489                                 "Failed to get trip point %d type: %d\n", i,
 490                                 ret);
 491                        continue;
 492                }
 493
 494                if (type == THERMAL_TRIP_PASSIVE) {
 495                        if (!found_first_passive) {
 496                                params->trip_switch_on = i;
 497                                found_first_passive = true;
 498                        } else  {
 499                                last_passive = i;
 500                        }
 501                } else if (type == THERMAL_TRIP_ACTIVE) {
 502                        last_active = i;
 503                } else {
 504                        break;
 505                }
 506        }
 507
 508        if (last_passive != INVALID_TRIP) {
 509                params->trip_max_desired_temperature = last_passive;
 510        } else if (found_first_passive) {
 511                params->trip_max_desired_temperature = params->trip_switch_on;
 512                params->trip_switch_on = INVALID_TRIP;
 513        } else {
 514                params->trip_switch_on = INVALID_TRIP;
 515                params->trip_max_desired_temperature = last_active;
 516        }
 517}
 518
 519static void reset_pid_controller(struct power_allocator_params *params)
 520{
 521        params->err_integral = 0;
 522        params->prev_err = 0;
 523}
 524
 525static void allow_maximum_power(struct thermal_zone_device *tz)
 526{
 527        struct thermal_instance *instance;
 528        struct power_allocator_params *params = tz->governor_data;
 529
 530        list_for_each_entry(instance, &tz->thermal_instances, tz_node) {
 531                if ((instance->trip != params->trip_max_desired_temperature) ||
 532                    (!cdev_is_power_actor(instance->cdev)))
 533                        continue;
 534
 535                instance->target = 0;
 536                instance->cdev->updated = false;
 537                thermal_cdev_update(instance->cdev);
 538        }
 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 */
 550static 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
 591free_params:
 592        kfree(params);
 593
 594        return ret;
 595}
 596
 597static 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
 612static int power_allocator_throttle(struct thermal_zone_device *tz, int trip)
 613{
 614        int ret;
 615        int switch_on_temp, control_temp, current_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 = thermal_zone_get_temp(tz, &current_temp);
 626        if (ret) {
 627                dev_warn(&tz->device, "Failed to get temperature: %d\n", ret);
 628                return ret;
 629        }
 630
 631        ret = tz->ops->get_trip_temp(tz, params->trip_switch_on,
 632                                     &switch_on_temp);
 633        if (!ret && (current_temp < switch_on_temp)) {
 634                tz->passive = 0;
 635                reset_pid_controller(params);
 636                allow_maximum_power(tz);
 637                return 0;
 638        }
 639
 640        tz->passive = 1;
 641
 642        ret = tz->ops->get_trip_temp(tz, params->trip_max_desired_temperature,
 643                                &control_temp);
 644        if (ret) {
 645                dev_warn(&tz->device,
 646                         "Failed to get the maximum desired temperature: %d\n",
 647                         ret);
 648                return ret;
 649        }
 650
 651        return allocate_power(tz, current_temp, control_temp);
 652}
 653
 654static struct thermal_governor thermal_gov_power_allocator = {
 655        .name           = "power_allocator",
 656        .bind_to_tz     = power_allocator_bind,
 657        .unbind_from_tz = power_allocator_unbind,
 658        .throttle       = power_allocator_throttle,
 659};
 660
 661int thermal_gov_power_allocator_register(void)
 662{
 663        return thermal_register_governor(&thermal_gov_power_allocator);
 664}
 665
 666void thermal_gov_power_allocator_unregister(void)
 667{
 668        thermal_unregister_governor(&thermal_gov_power_allocator);
 669}
 670