drivers/thermal/cpufreq_cooling.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  *  linux/drivers/thermal/cpufreq_cooling.c
   4  *
   5  *  Copyright (C) 2012  Samsung Electronics Co., Ltd(http://www.samsung.com)
   6  *
   7  *  Copyright (C) 2012-2018 Linaro Limited.
   8  *
   9  *  Authors:    Amit Daniel <amit.kachhap@linaro.org>
  10  *              Viresh Kumar <viresh.kumar@linaro.org>
  11  *
  12  */
  13 #include <linux/cpu.h>
  14 #include <linux/cpufreq.h>
  15 #include <linux/cpu_cooling.h>
  16 #include <linux/device.h>
  17 #include <linux/energy_model.h>
  18 #include <linux/err.h>
  19 #include <linux/export.h>
  20 #include <linux/pm_opp.h>
  21 #include <linux/pm_qos.h>
  22 #include <linux/slab.h>
  23 #include <linux/thermal.h>
  24 #include <linux/units.h>
  25
  26 #include <trace/events/thermal.h>
  27
  28 /*
  29  * Cooling state <-> CPUFreq frequency
  30  *
  31  * Cooling states are translated to frequencies throughout this driver and this
  32  * is the relation between them.
  33  *
  34  * Highest cooling state corresponds to lowest possible frequency.
  35  *
  36  * i.e.
  37  *      level 0 --> 1st Max Freq
  38  *      level 1 --> 2nd Max Freq
  39  *      ...
  40  */
  41
  42 /**
  43  * struct time_in_idle - Idle time stats
  44  * @time: previous reading of the absolute time that this cpu was idle
  45  * @timestamp: wall time of the last invocation of get_cpu_idle_time_us()
  46  */
  47 struct time_in_idle {
  48         u64 time;
  49         u64 timestamp;
  50 };
  51
  52 /**
  53  * struct cpufreq_cooling_device - data for cooling device with cpufreq
  54  * @last_load: load measured by the latest call to cpufreq_get_requested_power()
  55  * @cpufreq_state: integer value representing the current state of cpufreq
  56  *      cooling devices.
  57  * @max_level: maximum cooling level. One less than total number of valid
  58  *      cpufreq frequencies.
  59  * @em: Reference on the Energy Model of the device
  60  * @cdev: thermal_cooling_device pointer to keep track of the
  61  *      registered cooling device.
  62  * @policy: cpufreq policy.
  63  * @idle_time: idle time stats
  64  * @qos_req: PM QoS contraint to apply
  65  *
  66  * This structure is required for keeping information of each registered
  67  * cpufreq_cooling_device.
  68  */
  69 struct cpufreq_cooling_device {
  70         u32 last_load;
  71         unsigned int cpufreq_state;
  72         unsigned int max_level;
  73         struct em_perf_domain *em;
  74         struct cpufreq_policy *policy;
  75 #ifndef CONFIG_SMP
  76         struct time_in_idle *idle_time;
  77 #endif
  78         struct freq_qos_request qos_req;
  79 };
  80
  81 #ifdef CONFIG_THERMAL_GOV_POWER_ALLOCATOR
  82 /**
  83  * get_level: Find the level for a particular frequency
  84  * @cpufreq_cdev: cpufreq_cdev for which the property is required
  85  * @freq: Frequency
  86  *
  87  * Return: level corresponding to the frequency.
  88  */
  89 static unsigned long get_level(struct cpufreq_cooling_device *cpufreq_cdev,
  90                                unsigned int freq)
  91 {
  92         int i;
  93
  94         for (i = cpufreq_cdev->max_level - 1; i >= 0; i--) {
  95                 if (freq > cpufreq_cdev->em->table[i].frequency)
  96                         break;
  97         }
  98
  99         return cpufreq_cdev->max_level - i - 1;
 100 }
 101
 102 static u32 cpu_freq_to_power(struct cpufreq_cooling_device *cpufreq_cdev,
 103                              u32 freq)
 104 {
 105         unsigned long power_mw;
 106         int i;
 107
 108         for (i = cpufreq_cdev->max_level - 1; i >= 0; i--) {
 109                 if (freq > cpufreq_cdev->em->table[i].frequency)
 110                         break;
 111         }
 112
 113         power_mw = cpufreq_cdev->em->table[i + 1].power;
 114         power_mw /= MICROWATT_PER_MILLIWATT;
 115
 116         return power_mw;
 117 }
 118
 119 static u32 cpu_power_to_freq(struct cpufreq_cooling_device *cpufreq_cdev,
 120                              u32 power)
 121 {
 122         unsigned long em_power_mw;
 123         int i;
 124
 125         for (i = cpufreq_cdev->max_level; i > 0; i--) {
 126                 /* Convert EM power to milli-Watts to make safe comparison */
 127                 em_power_mw = cpufreq_cdev->em->table[i].power;
 128                 em_power_mw /= MICROWATT_PER_MILLIWATT;
 129                 if (power >= em_power_mw)
 130                         break;
 131         }
 132
 133         return cpufreq_cdev->em->table[i].frequency;
 134 }
 135
 136 /**
 137  * get_load() - get load for a cpu
 138  * @cpufreq_cdev: struct cpufreq_cooling_device for the cpu
 139  * @cpu: cpu number
 140  * @cpu_idx: index of the cpu in time_in_idle array
 141  *
 142  * Return: The average load of cpu @cpu in percentage since this
 143  * function was last called.
 144  */
 145 #ifdef CONFIG_SMP
 146 static u32 get_load(struct cpufreq_cooling_device *cpufreq_cdev, int cpu,
 147                     int cpu_idx)
 148 {
 149         unsigned long util = sched_cpu_util(cpu);
 150
 151         return (util * 100) / arch_scale_cpu_capacity(cpu);
 152 }
 153 #else /* !CONFIG_SMP */
 154 static u32 get_load(struct cpufreq_cooling_device *cpufreq_cdev, int cpu,
 155                     int cpu_idx)
 156 {
 157         u32 load;
 158         u64 now, now_idle, delta_time, delta_idle;
 159         struct time_in_idle *idle_time = &cpufreq_cdev->idle_time[cpu_idx];
 160
 161         now_idle = get_cpu_idle_time(cpu, &now, 0);
 162         delta_idle = now_idle - idle_time->time;
 163         delta_time = now - idle_time->timestamp;
 164
 165         if (delta_time <= delta_idle)
 166                 load = 0;
 167         else
 168                 load = div64_u64(100 * (delta_time - delta_idle), delta_time);
 169
 170         idle_time->time = now_idle;
 171         idle_time->timestamp = now;
 172
 173         return load;
 174 }
 175 #endif /* CONFIG_SMP */
 176
 177 /**
 178  * get_dynamic_power() - calculate the dynamic power
 179  * @cpufreq_cdev:       &cpufreq_cooling_device for this cdev
 180  * @freq:       current frequency
 181  *
 182  * Return: the dynamic power consumed by the cpus described by
 183  * @cpufreq_cdev.
 184  */
 185 static u32 get_dynamic_power(struct cpufreq_cooling_device *cpufreq_cdev,
 186                              unsigned long freq)
 187 {
 188         u32 raw_cpu_power;
 189
 190         raw_cpu_power = cpu_freq_to_power(cpufreq_cdev, freq);
 191         return (raw_cpu_power * cpufreq_cdev->last_load) / 100;
 192 }
 193
 194 /**
 195  * cpufreq_get_requested_power() - get the current power
 196  * @cdev:       &thermal_cooling_device pointer
 197  * @power:      pointer in which to store the resulting power
 198  *
 199  * Calculate the current power consumption of the cpus in milliwatts
 200  * and store it in @power.  This function should actually calculate
 201  * the requested power, but it's hard to get the frequency that
 202  * cpufreq would have assigned if there were no thermal limits.
 203  * Instead, we calculate the current power on the assumption that the
 204  * immediate future will look like the immediate past.
 205  *
 206  * We use the current frequency and the average load since this
 207  * function was last called.  In reality, there could have been
 208  * multiple opps since this function was last called and that affects
 209  * the load calculation.  While it's not perfectly accurate, this
 210  * simplification is good enough and works.  REVISIT this, as more
 211  * complex code may be needed if experiments show that it's not
 212  * accurate enough.
 213  *
 214  * Return: 0 on success, -E* if getting the static power failed.
 215  */
 216 static int cpufreq_get_requested_power(struct thermal_cooling_device *cdev,
 217                                        u32 *power)
 218 {
 219         unsigned long freq;
 220         int i = 0, cpu;
 221         u32 total_load = 0;
 222         struct cpufreq_cooling_device *cpufreq_cdev = cdev->devdata;
 223         struct cpufreq_policy *policy = cpufreq_cdev->policy;
 224         u32 *load_cpu = NULL;
 225
 226         freq = cpufreq_quick_get(policy->cpu);
 227
 228         if (trace_thermal_power_cpu_get_power_enabled()) {
 229                 u32 ncpus = cpumask_weight(policy->related_cpus);
 230
 231                 load_cpu = kcalloc(ncpus, sizeof(*load_cpu), GFP_KERNEL);
 232         }
 233
 234         for_each_cpu(cpu, policy->related_cpus) {
 235                 u32 load;
 236
 237                 if (cpu_online(cpu))
 238                         load = get_load(cpufreq_cdev, cpu, i);
 239                 else
 240                         load = 0;
 241
 242                 total_load += load;
 243                 if (load_cpu)
 244                         load_cpu[i] = load;
 245
 246                 i++;
 247         }
 248
 249         cpufreq_cdev->last_load = total_load;
 250
 251         *power = get_dynamic_power(cpufreq_cdev, freq);
 252
 253         if (load_cpu) {
 254                 trace_thermal_power_cpu_get_power(policy->related_cpus, freq,
 255                                                   load_cpu, i, *power);
 256
 257                 kfree(load_cpu);
 258         }
 259
 260         return 0;
 261 }
 262
 263 /**
 264  * cpufreq_state2power() - convert a cpu cdev state to power consumed
 265  * @cdev:       &thermal_cooling_device pointer
 266  * @state:      cooling device state to be converted
 267  * @power:      pointer in which to store the resulting power
 268  *
 269  * Convert cooling device state @state into power consumption in
 270  * milliwatts assuming 100% load.  Store the calculated power in
 271  * @power.
 272  *
 273  * Return: 0 on success, -EINVAL if the cooling device state could not
 274  * be converted into a frequency or other -E* if there was an error
 275  * when calculating the static power.
 276  */
 277 static int cpufreq_state2power(struct thermal_cooling_device *cdev,
 278                                unsigned long state, u32 *power)
 279 {
 280         unsigned int freq, num_cpus, idx;
 281         struct cpufreq_cooling_device *cpufreq_cdev = cdev->devdata;
 282
 283         /* Request state should be less than max_level */
 284         if (state > cpufreq_cdev->max_level)
 285                 return -EINVAL;
 286
 287         num_cpus = cpumask_weight(cpufreq_cdev->policy->cpus);
 288
 289         idx = cpufreq_cdev->max_level - state;
 290         freq = cpufreq_cdev->em->table[idx].frequency;
 291         *power = cpu_freq_to_power(cpufreq_cdev, freq) * num_cpus;
 292
 293         return 0;
 294 }
 295
 296 /**
 297  * cpufreq_power2state() - convert power to a cooling device state
 298  * @cdev:       &thermal_cooling_device pointer
 299  * @power:      power in milliwatts to be converted
 300  * @state:      pointer in which to store the resulting state
 301  *
 302  * Calculate a cooling device state for the cpus described by @cdev
 303  * that would allow them to consume at most @power mW and store it in
 304  * @state.  Note that this calculation depends on external factors
 305  * such as the cpu load or the current static power.  Calling this
 306  * function with the same power as input can yield different cooling
 307  * device states depending on those external factors.
 308  *
 309  * Return: 0 on success, -ENODEV if no cpus are online or -EINVAL if
 310  * the calculated frequency could not be converted to a valid state.
 311  * The latter should not happen unless the frequencies available to
 312  * cpufreq have changed since the initialization of the cpu cooling
 313  * device.
 314  */
 315 static int cpufreq_power2state(struct thermal_cooling_device *cdev,
 316                                u32 power, unsigned long *state)
 317 {
 318         unsigned int target_freq;
 319         u32 last_load, normalised_power;
 320         struct cpufreq_cooling_device *cpufreq_cdev = cdev->devdata;
 321         struct cpufreq_policy *policy = cpufreq_cdev->policy;
 322
 323         last_load = cpufreq_cdev->last_load ?: 1;
 324         normalised_power = (power * 100) / last_load;
 325         target_freq = cpu_power_to_freq(cpufreq_cdev, normalised_power);
 326
 327         *state = get_level(cpufreq_cdev, target_freq);
 328         trace_thermal_power_cpu_limit(policy->related_cpus, target_freq, *state,
 329                                       power);
 330         return 0;
 331 }
 332
 333 static inline bool em_is_sane(struct cpufreq_cooling_device *cpufreq_cdev,
 334                               struct em_perf_domain *em) {
 335         struct cpufreq_policy *policy;
 336         unsigned int nr_levels;
 337
 338         if (!em || em_is_artificial(em))
 339                 return false;
 340
 341         policy = cpufreq_cdev->policy;
 342         if (!cpumask_equal(policy->related_cpus, em_span_cpus(em))) {
 343                 pr_err("The span of pd %*pbl is misaligned with cpufreq policy %*pbl\n",
 344                         cpumask_pr_args(em_span_cpus(em)),
 345                         cpumask_pr_args(policy->related_cpus));
 346                 return false;
 347         }
 348
 349         nr_levels = cpufreq_cdev->max_level + 1;
 350         if (em_pd_nr_perf_states(em) != nr_levels) {
 351                 pr_err("The number of performance states in pd %*pbl (%u) doesn't match the number of cooling levels (%u)\n",
 352                         cpumask_pr_args(em_span_cpus(em)),
 353                         em_pd_nr_perf_states(em), nr_levels);
 354                 return false;
 355         }
 356
 357         return true;
 358 }
 359 #endif /* CONFIG_THERMAL_GOV_POWER_ALLOCATOR */
 360
 361 #ifdef CONFIG_SMP
 362 static inline int allocate_idle_time(struct cpufreq_cooling_device *cpufreq_cdev)
 363 {
 364         return 0;
 365 }
 366
 367 static inline void free_idle_time(struct cpufreq_cooling_device *cpufreq_cdev)
 368 {
 369 }
 370 #else
 371 static int allocate_idle_time(struct cpufreq_cooling_device *cpufreq_cdev)
 372 {
 373         unsigned int num_cpus = cpumask_weight(cpufreq_cdev->policy->related_cpus);
 374
 375         cpufreq_cdev->idle_time = kcalloc(num_cpus,
 376                                           sizeof(*cpufreq_cdev->idle_time),
 377                                           GFP_KERNEL);
 378         if (!cpufreq_cdev->idle_time)
 379                 return -ENOMEM;
 380
 381         return 0;
 382 }
 383
 384 static void free_idle_time(struct cpufreq_cooling_device *cpufreq_cdev)
 385 {
 386         kfree(cpufreq_cdev->idle_time);
 387         cpufreq_cdev->idle_time = NULL;
 388 }
 389 #endif /* CONFIG_SMP */
 390
 391 static unsigned int get_state_freq(struct cpufreq_cooling_device *cpufreq_cdev,
 392                                    unsigned long state)
 393 {
 394         struct cpufreq_policy *policy;
 395         unsigned long idx;
 396
 397 #ifdef CONFIG_THERMAL_GOV_POWER_ALLOCATOR
 398         /* Use the Energy Model table if available */
 399         if (cpufreq_cdev->em) {
 400                 idx = cpufreq_cdev->max_level - state;
 401                 return cpufreq_cdev->em->table[idx].frequency;
 402         }
 403 #endif
 404
 405         /* Otherwise, fallback on the CPUFreq table */
 406         policy = cpufreq_cdev->policy;
 407         if (policy->freq_table_sorted == CPUFREQ_TABLE_SORTED_ASCENDING)
 408                 idx = cpufreq_cdev->max_level - state;
 409         else
 410                 idx = state;
 411
 412         return policy->freq_table[idx].frequency;
 413 }
 414
 415 /* cpufreq cooling device callback functions are defined below */
 416
 417 /**
 418  * cpufreq_get_max_state - callback function to get the max cooling state.
 419  * @cdev: thermal cooling device pointer.
 420  * @state: fill this variable with the max cooling state.
 421  *
 422  * Callback for the thermal cooling device to return the cpufreq
 423  * max cooling state.
 424  *
 425  * Return: 0 on success, an error code otherwise.
 426  */
 427 static int cpufreq_get_max_state(struct thermal_cooling_device *cdev,
 428                                  unsigned long *state)
 429 {
 430         struct cpufreq_cooling_device *cpufreq_cdev = cdev->devdata;
 431
 432         *state = cpufreq_cdev->max_level;
 433         return 0;
 434 }
 435
 436 /**
 437  * cpufreq_get_cur_state - callback function to get the current cooling state.
 438  * @cdev: thermal cooling device pointer.
 439  * @state: fill this variable with the current cooling state.
 440  *
 441  * Callback for the thermal cooling device to return the cpufreq
 442  * current cooling state.
 443  *
 444  * Return: 0 on success, an error code otherwise.
 445  */
 446 static int cpufreq_get_cur_state(struct thermal_cooling_device *cdev,
 447                                  unsigned long *state)
 448 {
 449         struct cpufreq_cooling_device *cpufreq_cdev = cdev->devdata;
 450
 451         *state = cpufreq_cdev->cpufreq_state;
 452
 453         return 0;
 454 }
 455
 456 /**
 457  * cpufreq_set_cur_state - callback function to set the current cooling state.
 458  * @cdev: thermal cooling device pointer.
 459  * @state: set this variable to the current cooling state.
 460  *
 461  * Callback for the thermal cooling device to change the cpufreq
 462  * current cooling state.
 463  *
 464  * Return: 0 on success, an error code otherwise.
 465  */
 466 static int cpufreq_set_cur_state(struct thermal_cooling_device *cdev,
 467                                  unsigned long state)
 468 {
 469         struct cpufreq_cooling_device *cpufreq_cdev = cdev->devdata;
 470         struct cpumask *cpus;
 471         unsigned int frequency;
 472         int ret;
 473
 474         /* Request state should be less than max_level */
 475         if (state > cpufreq_cdev->max_level)
 476                 return -EINVAL;
 477
 478         /* Check if the old cooling action is same as new cooling action */
 479         if (cpufreq_cdev->cpufreq_state == state)
 480                 return 0;
 481
 482         frequency = get_state_freq(cpufreq_cdev, state);
 483
 484         ret = freq_qos_update_request(&cpufreq_cdev->qos_req, frequency);
 485         if (ret >= 0) {
 486                 cpufreq_cdev->cpufreq_state = state;
 487                 cpus = cpufreq_cdev->policy->related_cpus;
 488                 arch_update_thermal_pressure(cpus, frequency);
 489                 ret = 0;
 490         }
 491
 492         return ret;
 493 }
 494
 495 /* Bind cpufreq callbacks to thermal cooling device ops */
 496
 497 static struct thermal_cooling_device_ops cpufreq_cooling_ops = {
 498         .get_max_state          = cpufreq_get_max_state,
 499         .get_cur_state          = cpufreq_get_cur_state,
 500         .set_cur_state          = cpufreq_set_cur_state,
 501 };
 502
 503 /**
 504  * __cpufreq_cooling_register - helper function to create cpufreq cooling device
 505  * @np: a valid struct device_node to the cooling device device tree node
 506  * @policy: cpufreq policy
 507  * Normally this should be same as cpufreq policy->related_cpus.
 508  * @em: Energy Model of the cpufreq policy
 509  *
 510  * This interface function registers the cpufreq cooling device with the name
 511  * "thermal-cpufreq-%x". This api can support multiple instances of cpufreq
 512  * cooling devices. It also gives the opportunity to link the cooling device
 513  * with a device tree node, in order to bind it via the thermal DT code.
 514  *
 515  * Return: a valid struct thermal_cooling_device pointer on success,
 516  * on failure, it returns a corresponding ERR_PTR().
 517  */
 518 static struct thermal_cooling_device *
 519 __cpufreq_cooling_register(struct device_node *np,
 520                         struct cpufreq_policy *policy,
 521                         struct em_perf_domain *em)
 522 {
 523         struct thermal_cooling_device *cdev;
 524         struct cpufreq_cooling_device *cpufreq_cdev;
 525         unsigned int i;
 526         struct device *dev;
 527         int ret;
 528         struct thermal_cooling_device_ops *cooling_ops;
 529         char *name;
 530
 531         dev = get_cpu_device(policy->cpu);
 532         if (unlikely(!dev)) {
 533                 pr_warn("No cpu device for cpu %d\n", policy->cpu);
 534                 return ERR_PTR(-ENODEV);
 535         }
 536
 537         if (IS_ERR_OR_NULL(policy)) {
 538                 pr_err("%s: cpufreq policy isn't valid: %p\n", __func__, policy);
 539                 return ERR_PTR(-EINVAL);
 540         }
 541
 542         i = cpufreq_table_count_valid_entries(policy);
 543         if (!i) {
 544                 pr_debug("%s: CPUFreq table not found or has no valid entries\n",
 545                          __func__);
 546                 return ERR_PTR(-ENODEV);
 547         }
 548
 549         cpufreq_cdev = kzalloc(sizeof(*cpufreq_cdev), GFP_KERNEL);
 550         if (!cpufreq_cdev)
 551                 return ERR_PTR(-ENOMEM);
 552
 553         cpufreq_cdev->policy = policy;
 554
 555         ret = allocate_idle_time(cpufreq_cdev);
 556         if (ret) {
 557                 cdev = ERR_PTR(ret);
 558                 goto free_cdev;
 559         }
 560
 561         /* max_level is an index, not a counter */
 562         cpufreq_cdev->max_level = i - 1;
 563
 564         cooling_ops = &cpufreq_cooling_ops;
 565
 566 #ifdef CONFIG_THERMAL_GOV_POWER_ALLOCATOR
 567         if (em_is_sane(cpufreq_cdev, em)) {
 568                 cpufreq_cdev->em = em;
 569                 cooling_ops->get_requested_power = cpufreq_get_requested_power;
 570                 cooling_ops->state2power = cpufreq_state2power;
 571                 cooling_ops->power2state = cpufreq_power2state;
 572         } else
 573 #endif
 574         if (policy->freq_table_sorted == CPUFREQ_TABLE_UNSORTED) {
 575                 pr_err("%s: unsorted frequency tables are not supported\n",
 576                        __func__);
 577                 cdev = ERR_PTR(-EINVAL);
 578                 goto free_idle_time;
 579         }
 580
 581         ret = freq_qos_add_request(&policy->constraints,
 582                                    &cpufreq_cdev->qos_req, FREQ_QOS_MAX,
 583                                    get_state_freq(cpufreq_cdev, 0));
 584         if (ret < 0) {
 585                 pr_err("%s: Failed to add freq constraint (%d)\n", __func__,
 586                        ret);
 587                 cdev = ERR_PTR(ret);
 588                 goto free_idle_time;
 589         }
 590
 591         cdev = ERR_PTR(-ENOMEM);
 592         name = kasprintf(GFP_KERNEL, "cpufreq-%s", dev_name(dev));
 593         if (!name)
 594                 goto remove_qos_req;
 595
 596         cdev = thermal_of_cooling_device_register(np, name, cpufreq_cdev,
 597                                                   cooling_ops);
 598         kfree(name);
 599
 600         if (IS_ERR(cdev))
 601                 goto remove_qos_req;
 602
 603         return cdev;
 604
 605 remove_qos_req:
 606         freq_qos_remove_request(&cpufreq_cdev->qos_req);
 607 free_idle_time:
 608         free_idle_time(cpufreq_cdev);
 609 free_cdev:
 610         kfree(cpufreq_cdev);
 611         return cdev;
 612 }
 613
 614 /**
 615  * cpufreq_cooling_register - function to create cpufreq cooling device.
 616  * @policy: cpufreq policy
 617  *
 618  * This interface function registers the cpufreq cooling device with the name
 619  * "thermal-cpufreq-%x". This api can support multiple instances of cpufreq
 620  * cooling devices.
 621  *
 622  * Return: a valid struct thermal_cooling_device pointer on success,
 623  * on failure, it returns a corresponding ERR_PTR().
 624  */
 625 struct thermal_cooling_device *
 626 cpufreq_cooling_register(struct cpufreq_policy *policy)
 627 {
 628         return __cpufreq_cooling_register(NULL, policy, NULL);
 629 }
 630 EXPORT_SYMBOL_GPL(cpufreq_cooling_register);
 631
 632 /**
 633  * of_cpufreq_cooling_register - function to create cpufreq cooling device.
 634  * @policy: cpufreq policy
 635  *
 636  * This interface function registers the cpufreq cooling device with the name
 637  * "thermal-cpufreq-%x". This api can support multiple instances of cpufreq
 638  * cooling devices. Using this API, the cpufreq cooling device will be
 639  * linked to the device tree node provided.
 640  *
 641  * Using this function, the cooling device will implement the power
 642  * extensions by using a simple cpu power model.  The cpus must have
 643  * registered their OPPs using the OPP library.
 644  *
 645  * It also takes into account, if property present in policy CPU node, the
 646  * static power consumed by the cpu.
 647  *
 648  * Return: a valid struct thermal_cooling_device pointer on success,
 649  * and NULL on failure.
 650  */
 651 struct thermal_cooling_device *
 652 of_cpufreq_cooling_register(struct cpufreq_policy *policy)
 653 {
 654         struct device_node *np = of_get_cpu_node(policy->cpu, NULL);
 655         struct thermal_cooling_device *cdev = NULL;
 656
 657         if (!np) {
 658                 pr_err("cpufreq_cooling: OF node not available for cpu%d\n",
 659                        policy->cpu);
 660                 return NULL;
 661         }
 662
 663         if (of_find_property(np, "#cooling-cells", NULL)) {
 664                 struct em_perf_domain *em = em_cpu_get(policy->cpu);
 665
 666                 cdev = __cpufreq_cooling_register(np, policy, em);
 667                 if (IS_ERR(cdev)) {
 668                         pr_err("cpufreq_cooling: cpu%d failed to register as cooling device: %ld\n",
 669                                policy->cpu, PTR_ERR(cdev));
 670                         cdev = NULL;
 671                 }
 672         }
 673
 674         of_node_put(np);
 675         return cdev;
 676 }
 677 EXPORT_SYMBOL_GPL(of_cpufreq_cooling_register);
 678
 679 /**
 680  * cpufreq_cooling_unregister - function to remove cpufreq cooling device.
 681  * @cdev: thermal cooling device pointer.
 682  *
 683  * This interface function unregisters the "thermal-cpufreq-%x" cooling device.
 684  */
 685 void cpufreq_cooling_unregister(struct thermal_cooling_device *cdev)
 686 {
 687         struct cpufreq_cooling_device *cpufreq_cdev;
 688
 689         if (!cdev)
 690                 return;
 691
 692         cpufreq_cdev = cdev->devdata;
 693
 694         thermal_cooling_device_unregister(cdev);
 695         freq_qos_remove_request(&cpufreq_cdev->qos_req);
 696         free_idle_time(cpufreq_cdev);
 697         kfree(cpufreq_cdev);
 698 }
 699 EXPORT_SYMBOL_GPL(cpufreq_cooling_unregister);