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