drivers/thermal/devfreq_cooling.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * devfreq_cooling: Thermal cooling device implementation for devices using
   4  *                  devfreq
   5  *
   6  * Copyright (C) 2014-2015 ARM Limited
   7  *
   8  * TODO:
   9  *    - If OPPs are added or removed after devfreq cooling has
  10  *      registered, the devfreq cooling won't react to it.
  11  */
  12
  13 #include <linux/devfreq.h>
  14 #include <linux/devfreq_cooling.h>
  15 #include <linux/energy_model.h>
  16 #include <linux/export.h>
  17 #include <linux/idr.h>
  18 #include <linux/slab.h>
  19 #include <linux/pm_opp.h>
  20 #include <linux/pm_qos.h>
  21 #include <linux/thermal.h>
  22
  23 #include <trace/events/thermal.h>
  24
  25 #define HZ_PER_KHZ              1000
  26 #define SCALE_ERROR_MITIGATION  100
  27
  28 static DEFINE_IDA(devfreq_ida);
  29
  30 /**
  31  * struct devfreq_cooling_device - Devfreq cooling device
  32  * @id:         unique integer value corresponding to each
  33  *              devfreq_cooling_device registered.
  34  * @cdev:       Pointer to associated thermal cooling device.
  35  * @devfreq:    Pointer to associated devfreq device.
  36  * @cooling_state:      Current cooling state.
  37  * @freq_table: Pointer to a table with the frequencies sorted in descending
  38  *              order.  You can index the table by cooling device state
  39  * @max_state:  It is the last index, that is, one less than the number of the
  40  *              OPPs
  41  * @power_ops:  Pointer to devfreq_cooling_power, a more precised model.
  42  * @res_util:   Resource utilization scaling factor for the power.
  43  *              It is multiplied by 100 to minimize the error. It is used
  44  *              for estimation of the power budget instead of using
  45  *              'utilization' (which is 'busy_time' / 'total_time').
  46  *              The 'res_util' range is from 100 to power * 100 for the
  47  *              corresponding 'state'.
  48  * @capped_state:       index to cooling state with in dynamic power budget
  49  * @req_max_freq:       PM QoS request for limiting the maximum frequency
  50  *                      of the devfreq device.
  51  * @em_pd:              Energy Model for the associated Devfreq device
  52  */
  53 struct devfreq_cooling_device {
  54         int id;
  55         struct thermal_cooling_device *cdev;
  56         struct devfreq *devfreq;
  57         unsigned long cooling_state;
  58         u32 *freq_table;
  59         size_t max_state;
  60         struct devfreq_cooling_power *power_ops;
  61         u32 res_util;
  62         int capped_state;
  63         struct dev_pm_qos_request req_max_freq;
  64         struct em_perf_domain *em_pd;
  65 };
  66
  67 static int devfreq_cooling_get_max_state(struct thermal_cooling_device *cdev,
  68                                          unsigned long *state)
  69 {
  70         struct devfreq_cooling_device *dfc = cdev->devdata;
  71
  72         *state = dfc->max_state;
  73
  74         return 0;
  75 }
  76
  77 static int devfreq_cooling_get_cur_state(struct thermal_cooling_device *cdev,
  78                                          unsigned long *state)
  79 {
  80         struct devfreq_cooling_device *dfc = cdev->devdata;
  81
  82         *state = dfc->cooling_state;
  83
  84         return 0;
  85 }
  86
  87 static int devfreq_cooling_set_cur_state(struct thermal_cooling_device *cdev,
  88                                          unsigned long state)
  89 {
  90         struct devfreq_cooling_device *dfc = cdev->devdata;
  91         struct devfreq *df = dfc->devfreq;
  92         struct device *dev = df->dev.parent;
  93         unsigned long freq;
  94         int perf_idx;
  95
  96         if (state == dfc->cooling_state)
  97                 return 0;
  98
  99         dev_dbg(dev, "Setting cooling state %lu\n", state);
 100
 101         if (state > dfc->max_state)
 102                 return -EINVAL;
 103
 104         if (dfc->em_pd) {
 105                 perf_idx = dfc->max_state - state;
 106                 freq = dfc->em_pd->table[perf_idx].frequency * 1000;
 107         } else {
 108                 freq = dfc->freq_table[state];
 109         }
 110
 111         dev_pm_qos_update_request(&dfc->req_max_freq,
 112                                   DIV_ROUND_UP(freq, HZ_PER_KHZ));
 113
 114         dfc->cooling_state = state;
 115
 116         return 0;
 117 }
 118
 119 /**
 120  * get_perf_idx() - get the performance index corresponding to a frequency
 121  * @em_pd:      Pointer to device's Energy Model
 122  * @freq:       frequency in kHz
 123  *
 124  * Return: the performance index associated with the @freq, or
 125  * -EINVAL if it wasn't found.
 126  */
 127 static int get_perf_idx(struct em_perf_domain *em_pd, unsigned long freq)
 128 {
 129         int i;
 130
 131         for (i = 0; i < em_pd->nr_perf_states; i++) {
 132                 if (em_pd->table[i].frequency == freq)
 133                         return i;
 134         }
 135
 136         return -EINVAL;
 137 }
 138
 139 static unsigned long get_voltage(struct devfreq *df, unsigned long freq)
 140 {
 141         struct device *dev = df->dev.parent;
 142         unsigned long voltage;
 143         struct dev_pm_opp *opp;
 144
 145         opp = dev_pm_opp_find_freq_exact(dev, freq, true);
 146         if (PTR_ERR(opp) == -ERANGE)
 147                 opp = dev_pm_opp_find_freq_exact(dev, freq, false);
 148
 149         if (IS_ERR(opp)) {
 150                 dev_err_ratelimited(dev, "Failed to find OPP for frequency %lu: %ld\n",
 151                                     freq, PTR_ERR(opp));
 152                 return 0;
 153         }
 154
 155         voltage = dev_pm_opp_get_voltage(opp) / 1000; /* mV */
 156         dev_pm_opp_put(opp);
 157
 158         if (voltage == 0) {
 159                 dev_err_ratelimited(dev,
 160                                     "Failed to get voltage for frequency %lu\n",
 161                                     freq);
 162         }
 163
 164         return voltage;
 165 }
 166
 167 static void _normalize_load(struct devfreq_dev_status *status)
 168 {
 169         if (status->total_time > 0xfffff) {
 170                 status->total_time >>= 10;
 171                 status->busy_time >>= 10;
 172         }
 173
 174         status->busy_time <<= 10;
 175         status->busy_time /= status->total_time ? : 1;
 176
 177         status->busy_time = status->busy_time ? : 1;
 178         status->total_time = 1024;
 179 }
 180
 181 static int devfreq_cooling_get_requested_power(struct thermal_cooling_device *cdev,
 182                                                u32 *power)
 183 {
 184         struct devfreq_cooling_device *dfc = cdev->devdata;
 185         struct devfreq *df = dfc->devfreq;
 186         struct devfreq_dev_status status;
 187         unsigned long state;
 188         unsigned long freq;
 189         unsigned long voltage;
 190         int res, perf_idx;
 191
 192         mutex_lock(&df->lock);
 193         status = df->last_status;
 194         mutex_unlock(&df->lock);
 195
 196         freq = status.current_frequency;
 197
 198         if (dfc->power_ops && dfc->power_ops->get_real_power) {
 199                 voltage = get_voltage(df, freq);
 200                 if (voltage == 0) {
 201                         res = -EINVAL;
 202                         goto fail;
 203                 }
 204
 205                 res = dfc->power_ops->get_real_power(df, power, freq, voltage);
 206                 if (!res) {
 207                         state = dfc->capped_state;
 208                         dfc->res_util = dfc->em_pd->table[state].power;
 209                         dfc->res_util *= SCALE_ERROR_MITIGATION;
 210
 211                         if (*power > 1)
 212                                 dfc->res_util /= *power;
 213                 } else {
 214                         goto fail;
 215                 }
 216         } else {
 217                 /* Energy Model frequencies are in kHz */
 218                 perf_idx = get_perf_idx(dfc->em_pd, freq / 1000);
 219                 if (perf_idx < 0) {
 220                         res = -EAGAIN;
 221                         goto fail;
 222                 }
 223
 224                 _normalize_load(&status);
 225
 226                 /* Scale power for utilization */
 227                 *power = dfc->em_pd->table[perf_idx].power;
 228                 *power *= status.busy_time;
 229                 *power >>= 10;
 230         }
 231
 232         trace_thermal_power_devfreq_get_power(cdev, &status, freq, *power);
 233
 234         return 0;
 235 fail:
 236         /* It is safe to set max in this case */
 237         dfc->res_util = SCALE_ERROR_MITIGATION;
 238         return res;
 239 }
 240
 241 static int devfreq_cooling_state2power(struct thermal_cooling_device *cdev,
 242                                        unsigned long state, u32 *power)
 243 {
 244         struct devfreq_cooling_device *dfc = cdev->devdata;
 245         int perf_idx;
 246
 247         if (state > dfc->max_state)
 248                 return -EINVAL;
 249
 250         perf_idx = dfc->max_state - state;
 251         *power = dfc->em_pd->table[perf_idx].power;
 252
 253         return 0;
 254 }
 255
 256 static int devfreq_cooling_power2state(struct thermal_cooling_device *cdev,
 257                                        u32 power, unsigned long *state)
 258 {
 259         struct devfreq_cooling_device *dfc = cdev->devdata;
 260         struct devfreq *df = dfc->devfreq;
 261         struct devfreq_dev_status status;
 262         unsigned long freq;
 263         s32 est_power;
 264         int i;
 265
 266         mutex_lock(&df->lock);
 267         status = df->last_status;
 268         mutex_unlock(&df->lock);
 269
 270         freq = status.current_frequency;
 271
 272         if (dfc->power_ops && dfc->power_ops->get_real_power) {
 273                 /* Scale for resource utilization */
 274                 est_power = power * dfc->res_util;
 275                 est_power /= SCALE_ERROR_MITIGATION;
 276         } else {
 277                 /* Scale dynamic power for utilization */
 278                 _normalize_load(&status);
 279                 est_power = power << 10;
 280                 est_power /= status.busy_time;
 281         }
 282
 283         /*
 284          * Find the first cooling state that is within the power
 285          * budget. The EM power table is sorted ascending.
 286          */
 287         for (i = dfc->max_state; i > 0; i--)
 288                 if (est_power >= dfc->em_pd->table[i].power)
 289                         break;
 290
 291         *state = dfc->max_state - i;
 292         dfc->capped_state = *state;
 293
 294         trace_thermal_power_devfreq_limit(cdev, freq, *state, power);
 295         return 0;
 296 }
 297
 298 static struct thermal_cooling_device_ops devfreq_cooling_ops = {
 299         .get_max_state = devfreq_cooling_get_max_state,
 300         .get_cur_state = devfreq_cooling_get_cur_state,
 301         .set_cur_state = devfreq_cooling_set_cur_state,
 302 };
 303
 304 /**
 305  * devfreq_cooling_gen_tables() - Generate frequency table.
 306  * @dfc:        Pointer to devfreq cooling device.
 307  * @num_opps:   Number of OPPs
 308  *
 309  * Generate frequency table which holds the frequencies in descending
 310  * order. That way its indexed by cooling device state. This is for
 311  * compatibility with drivers which do not register Energy Model.
 312  *
 313  * Return: 0 on success, negative error code on failure.
 314  */
 315 static int devfreq_cooling_gen_tables(struct devfreq_cooling_device *dfc,
 316                                       int num_opps)
 317 {
 318         struct devfreq *df = dfc->devfreq;
 319         struct device *dev = df->dev.parent;
 320         unsigned long freq;
 321         int i;
 322
 323         dfc->freq_table = kcalloc(num_opps, sizeof(*dfc->freq_table),
 324                              GFP_KERNEL);
 325         if (!dfc->freq_table)
 326                 return -ENOMEM;
 327
 328         for (i = 0, freq = ULONG_MAX; i < num_opps; i++, freq--) {
 329                 struct dev_pm_opp *opp;
 330
 331                 opp = dev_pm_opp_find_freq_floor(dev, &freq);
 332                 if (IS_ERR(opp)) {
 333                         kfree(dfc->freq_table);
 334                         return PTR_ERR(opp);
 335                 }
 336
 337                 dev_pm_opp_put(opp);
 338                 dfc->freq_table[i] = freq;
 339         }
 340
 341         return 0;
 342 }
 343
 344 /**
 345  * of_devfreq_cooling_register_power() - Register devfreq cooling device,
 346  *                                      with OF and power information.
 347  * @np: Pointer to OF device_node.
 348  * @df: Pointer to devfreq device.
 349  * @dfc_power:  Pointer to devfreq_cooling_power.
 350  *
 351  * Register a devfreq cooling device.  The available OPPs must be
 352  * registered on the device.
 353  *
 354  * If @dfc_power is provided, the cooling device is registered with the
 355  * power extensions.  For the power extensions to work correctly,
 356  * devfreq should use the simple_ondemand governor, other governors
 357  * are not currently supported.
 358  */
 359 struct thermal_cooling_device *
 360 of_devfreq_cooling_register_power(struct device_node *np, struct devfreq *df,
 361                                   struct devfreq_cooling_power *dfc_power)
 362 {
 363         struct thermal_cooling_device *cdev;
 364         struct device *dev = df->dev.parent;
 365         struct devfreq_cooling_device *dfc;
 366         char dev_name[THERMAL_NAME_LENGTH];
 367         int err, num_opps;
 368
 369         dfc = kzalloc(sizeof(*dfc), GFP_KERNEL);
 370         if (!dfc)
 371                 return ERR_PTR(-ENOMEM);
 372
 373         dfc->devfreq = df;
 374
 375         dfc->em_pd = em_pd_get(dev);
 376         if (dfc->em_pd) {
 377                 devfreq_cooling_ops.get_requested_power =
 378                         devfreq_cooling_get_requested_power;
 379                 devfreq_cooling_ops.state2power = devfreq_cooling_state2power;
 380                 devfreq_cooling_ops.power2state = devfreq_cooling_power2state;
 381
 382                 dfc->power_ops = dfc_power;
 383
 384                 num_opps = em_pd_nr_perf_states(dfc->em_pd);
 385         } else {
 386                 /* Backward compatibility for drivers which do not use IPA */
 387                 dev_dbg(dev, "missing EM for cooling device\n");
 388
 389                 num_opps = dev_pm_opp_get_opp_count(dev);
 390
 391                 err = devfreq_cooling_gen_tables(dfc, num_opps);
 392                 if (err)
 393                         goto free_dfc;
 394         }
 395
 396         if (num_opps <= 0) {
 397                 err = -EINVAL;
 398                 goto free_dfc;
 399         }
 400
 401         /* max_state is an index, not a counter */
 402         dfc->max_state = num_opps - 1;
 403
 404         err = dev_pm_qos_add_request(dev, &dfc->req_max_freq,
 405                                      DEV_PM_QOS_MAX_FREQUENCY,
 406                                      PM_QOS_MAX_FREQUENCY_DEFAULT_VALUE);
 407         if (err < 0)
 408                 goto free_table;
 409
 410         err = ida_simple_get(&devfreq_ida, 0, 0, GFP_KERNEL);
 411         if (err < 0)
 412                 goto remove_qos_req;
 413
 414         dfc->id = err;
 415
 416         snprintf(dev_name, sizeof(dev_name), "thermal-devfreq-%d", dfc->id);
 417
 418         cdev = thermal_of_cooling_device_register(np, dev_name, dfc,
 419                                                   &devfreq_cooling_ops);
 420         if (IS_ERR(cdev)) {
 421                 err = PTR_ERR(cdev);
 422                 dev_err(dev,
 423                         "Failed to register devfreq cooling device (%d)\n",
 424                         err);
 425                 goto release_ida;
 426         }
 427
 428         dfc->cdev = cdev;
 429
 430         return cdev;
 431
 432 release_ida:
 433         ida_simple_remove(&devfreq_ida, dfc->id);
 434 remove_qos_req:
 435         dev_pm_qos_remove_request(&dfc->req_max_freq);
 436 free_table:
 437         kfree(dfc->freq_table);
 438 free_dfc:
 439         kfree(dfc);
 440
 441         return ERR_PTR(err);
 442 }
 443 EXPORT_SYMBOL_GPL(of_devfreq_cooling_register_power);
 444
 445 /**
 446  * of_devfreq_cooling_register() - Register devfreq cooling device,
 447  *                                with OF information.
 448  * @np: Pointer to OF device_node.
 449  * @df: Pointer to devfreq device.
 450  */
 451 struct thermal_cooling_device *
 452 of_devfreq_cooling_register(struct device_node *np, struct devfreq *df)
 453 {
 454         return of_devfreq_cooling_register_power(np, df, NULL);
 455 }
 456 EXPORT_SYMBOL_GPL(of_devfreq_cooling_register);
 457
 458 /**
 459  * devfreq_cooling_register() - Register devfreq cooling device.
 460  * @df: Pointer to devfreq device.
 461  */
 462 struct thermal_cooling_device *devfreq_cooling_register(struct devfreq *df)
 463 {
 464         return of_devfreq_cooling_register(NULL, df);
 465 }
 466 EXPORT_SYMBOL_GPL(devfreq_cooling_register);
 467
 468 /**
 469  * devfreq_cooling_em_register_power() - Register devfreq cooling device with
 470  *              power information and automatically register Energy Model (EM)
 471  * @df:         Pointer to devfreq device.
 472  * @dfc_power:  Pointer to devfreq_cooling_power.
 473  *
 474  * Register a devfreq cooling device and automatically register EM. The
 475  * available OPPs must be registered for the device.
 476  *
 477  * If @dfc_power is provided, the cooling device is registered with the
 478  * power extensions. It is using the simple Energy Model which requires
 479  * "dynamic-power-coefficient" a devicetree property. To not break drivers
 480  * which miss that DT property, the function won't bail out when the EM
 481  * registration failed. The cooling device will be registered if everything
 482  * else is OK.
 483  */
 484 struct thermal_cooling_device *
 485 devfreq_cooling_em_register(struct devfreq *df,
 486                             struct devfreq_cooling_power *dfc_power)
 487 {
 488         struct thermal_cooling_device *cdev;
 489         struct device *dev;
 490         int ret;
 491
 492         if (IS_ERR_OR_NULL(df))
 493                 return ERR_PTR(-EINVAL);
 494
 495         dev = df->dev.parent;
 496
 497         ret = dev_pm_opp_of_register_em(dev, NULL);
 498         if (ret)
 499                 dev_dbg(dev, "Unable to register EM for devfreq cooling device (%d)\n",
 500                         ret);
 501
 502         cdev = of_devfreq_cooling_register_power(dev->of_node, df, dfc_power);
 503
 504         if (IS_ERR_OR_NULL(cdev))
 505                 em_dev_unregister_perf_domain(dev);
 506
 507         return cdev;
 508 }
 509 EXPORT_SYMBOL_GPL(devfreq_cooling_em_register);
 510
 511 /**
 512  * devfreq_cooling_unregister() - Unregister devfreq cooling device.
 513  * @cdev: Pointer to devfreq cooling device to unregister.
 514  *
 515  * Unregisters devfreq cooling device and related Energy Model if it was
 516  * present.
 517  */
 518 void devfreq_cooling_unregister(struct thermal_cooling_device *cdev)
 519 {
 520         struct devfreq_cooling_device *dfc;
 521         struct device *dev;
 522
 523         if (IS_ERR_OR_NULL(cdev))
 524                 return;
 525
 526         dfc = cdev->devdata;
 527         dev = dfc->devfreq->dev.parent;
 528
 529         thermal_cooling_device_unregister(dfc->cdev);
 530         ida_simple_remove(&devfreq_ida, dfc->id);
 531         dev_pm_qos_remove_request(&dfc->req_max_freq);
 532
 533         em_dev_unregister_perf_domain(dev);
 534
 535         kfree(dfc->freq_table);
 536         kfree(dfc);
 537 }
 538 EXPORT_SYMBOL_GPL(devfreq_cooling_unregister);