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/slab.h>
  18 #include <linux/pm_opp.h>
  19 #include <linux/pm_qos.h>
  20 #include <linux/thermal.h>
  21 #include <linux/units.h>
  22
  23 #include "thermal_trace.h"
  24
  25 #define SCALE_ERROR_MITIGATION  100
  26
  27 /**
  28  * struct devfreq_cooling_device - Devfreq cooling device
  29  *              devfreq_cooling_device registered.
  30  * @cdev:       Pointer to associated thermal cooling device.
  31  * @cooling_ops: devfreq callbacks to thermal cooling device ops
  32  * @devfreq:    Pointer to associated devfreq device.
  33  * @cooling_state:      Current cooling state.
  34  * @freq_table: Pointer to a table with the frequencies sorted in descending
  35  *              order.  You can index the table by cooling device state
  36  * @max_state:  It is the last index, that is, one less than the number of the
  37  *              OPPs
  38  * @power_ops:  Pointer to devfreq_cooling_power, a more precised model.
  39  * @res_util:   Resource utilization scaling factor for the power.
  40  *              It is multiplied by 100 to minimize the error. It is used
  41  *              for estimation of the power budget instead of using
  42  *              'utilization' (which is 'busy_time' / 'total_time').
  43  *              The 'res_util' range is from 100 to power * 100 for the
  44  *              corresponding 'state'.
  45  * @capped_state:       index to cooling state with in dynamic power budget
  46  * @req_max_freq:       PM QoS request for limiting the maximum frequency
  47  *                      of the devfreq device.
  48  * @em_pd:              Energy Model for the associated Devfreq device
  49  */
  50 struct devfreq_cooling_device {
  51         struct thermal_cooling_device *cdev;
  52         struct thermal_cooling_device_ops cooling_ops;
  53         struct devfreq *devfreq;
  54         unsigned long cooling_state;
  55         u32 *freq_table;
  56         size_t max_state;
  57         struct devfreq_cooling_power *power_ops;
  58         u32 res_util;
  59         int capped_state;
  60         struct dev_pm_qos_request req_max_freq;
  61         struct em_perf_domain *em_pd;
  62 };
  63
  64 static int devfreq_cooling_get_max_state(struct thermal_cooling_device *cdev,
  65                                          unsigned long *state)
  66 {
  67         struct devfreq_cooling_device *dfc = cdev->devdata;
  68
  69         *state = dfc->max_state;
  70
  71         return 0;
  72 }
  73
  74 static int devfreq_cooling_get_cur_state(struct thermal_cooling_device *cdev,
  75                                          unsigned long *state)
  76 {
  77         struct devfreq_cooling_device *dfc = cdev->devdata;
  78
  79         *state = dfc->cooling_state;
  80
  81         return 0;
  82 }
  83
  84 static int devfreq_cooling_set_cur_state(struct thermal_cooling_device *cdev,
  85                                          unsigned long state)
  86 {
  87         struct devfreq_cooling_device *dfc = cdev->devdata;
  88         struct devfreq *df = dfc->devfreq;
  89         struct device *dev = df->dev.parent;
  90         struct em_perf_state *table;
  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 (dfc->em_pd) {
 103                 perf_idx = dfc->max_state - state;
 104
 105                 rcu_read_lock();
 106                 table = em_perf_state_from_pd(dfc->em_pd);
 107                 freq = table[perf_idx].frequency * 1000;
 108                 rcu_read_unlock();
 109         } else {
 110                 freq = dfc->freq_table[state];
 111         }
 112
 113         dev_pm_qos_update_request(&dfc->req_max_freq,
 114                                   DIV_ROUND_UP(freq, HZ_PER_KHZ));
 115
 116         dfc->cooling_state = state;
 117
 118         return 0;
 119 }
 120
 121 /**
 122  * get_perf_idx() - get the performance index corresponding to a frequency
 123  * @em_pd:      Pointer to device's Energy Model
 124  * @freq:       frequency in kHz
 125  *
 126  * Return: the performance index associated with the @freq, or
 127  * -EINVAL if it wasn't found.
 128  */
 129 static int get_perf_idx(struct em_perf_domain *em_pd, unsigned long freq)
 130 {
 131         struct em_perf_state *table;
 132         int i, idx = -EINVAL;
 133
 134         rcu_read_lock();
 135         table = em_perf_state_from_pd(em_pd);
 136         for (i = 0; i < em_pd->nr_perf_states; i++) {
 137                 if (table[i].frequency != freq)
 138                         continue;
 139
 140                 idx = i;
 141                 break;
 142         }
 143         rcu_read_unlock();
 144
 145         return idx;
 146 }
 147
 148 static unsigned long get_voltage(struct devfreq *df, unsigned long freq)
 149 {
 150         struct device *dev = df->dev.parent;
 151         unsigned long voltage;
 152         struct dev_pm_opp *opp;
 153
 154         opp = dev_pm_opp_find_freq_exact(dev, freq, true);
 155         if (PTR_ERR(opp) == -ERANGE)
 156                 opp = dev_pm_opp_find_freq_exact(dev, freq, false);
 157
 158         if (IS_ERR(opp)) {
 159                 dev_err_ratelimited(dev, "Failed to find OPP for frequency %lu: %ld\n",
 160                                     freq, PTR_ERR(opp));
 161                 return 0;
 162         }
 163
 164         voltage = dev_pm_opp_get_voltage(opp) / 1000; /* mV */
 165         dev_pm_opp_put(opp);
 166
 167         if (voltage == 0) {
 168                 dev_err_ratelimited(dev,
 169                                     "Failed to get voltage for frequency %lu\n",
 170                                     freq);
 171         }
 172
 173         return voltage;
 174 }
 175
 176 static void _normalize_load(struct devfreq_dev_status *status)
 177 {
 178         if (status->total_time > 0xfffff) {
 179                 status->total_time >>= 10;
 180                 status->busy_time >>= 10;
 181         }
 182
 183         status->busy_time <<= 10;
 184         status->busy_time /= status->total_time ? : 1;
 185
 186         status->busy_time = status->busy_time ? : 1;
 187         status->total_time = 1024;
 188 }
 189
 190 static int devfreq_cooling_get_requested_power(struct thermal_cooling_device *cdev,
 191                                                u32 *power)
 192 {
 193         struct devfreq_cooling_device *dfc = cdev->devdata;
 194         struct devfreq *df = dfc->devfreq;
 195         struct devfreq_dev_status status;
 196         struct em_perf_state *table;
 197         unsigned long state;
 198         unsigned long freq;
 199         unsigned long voltage;
 200         int res, perf_idx;
 201
 202         mutex_lock(&df->lock);
 203         status = df->last_status;
 204         mutex_unlock(&df->lock);
 205
 206         freq = status.current_frequency;
 207
 208         if (dfc->power_ops && dfc->power_ops->get_real_power) {
 209                 voltage = get_voltage(df, freq);
 210                 if (voltage == 0) {
 211                         res = -EINVAL;
 212                         goto fail;
 213                 }
 214
 215                 res = dfc->power_ops->get_real_power(df, power, freq, voltage);
 216                 if (!res) {
 217                         state = dfc->capped_state;
 218
 219                         /* Convert EM power into milli-Watts first */
 220                         rcu_read_lock();
 221                         table = em_perf_state_from_pd(dfc->em_pd);
 222                         dfc->res_util = table[state].power;
 223                         rcu_read_unlock();
 224
 225                         dfc->res_util /= MICROWATT_PER_MILLIWATT;
 226
 227                         dfc->res_util *= SCALE_ERROR_MITIGATION;
 228
 229                         if (*power > 1)
 230                                 dfc->res_util /= *power;
 231                 } else {
 232                         goto fail;
 233                 }
 234         } else {
 235                 /* Energy Model frequencies are in kHz */
 236                 perf_idx = get_perf_idx(dfc->em_pd, freq / 1000);
 237                 if (perf_idx < 0) {
 238                         res = -EAGAIN;
 239                         goto fail;
 240                 }
 241
 242                 _normalize_load(&status);
 243
 244                 /* Convert EM power into milli-Watts first */
 245                 rcu_read_lock();
 246                 table = em_perf_state_from_pd(dfc->em_pd);
 247                 *power = table[perf_idx].power;
 248                 rcu_read_unlock();
 249
 250                 *power /= MICROWATT_PER_MILLIWATT;
 251                 /* Scale power for utilization */
 252                 *power *= status.busy_time;
 253                 *power >>= 10;
 254         }
 255
 256         trace_thermal_power_devfreq_get_power(cdev, &status, freq, *power);
 257
 258         return 0;
 259 fail:
 260         /* It is safe to set max in this case */
 261         dfc->res_util = SCALE_ERROR_MITIGATION;
 262         return res;
 263 }
 264
 265 static int devfreq_cooling_state2power(struct thermal_cooling_device *cdev,
 266                                        unsigned long state, u32 *power)
 267 {
 268         struct devfreq_cooling_device *dfc = cdev->devdata;
 269         struct em_perf_state *table;
 270         int perf_idx;
 271
 272         if (state > dfc->max_state)
 273                 return -EINVAL;
 274
 275         perf_idx = dfc->max_state - state;
 276
 277         rcu_read_lock();
 278         table = em_perf_state_from_pd(dfc->em_pd);
 279         *power = table[perf_idx].power;
 280         rcu_read_unlock();
 281
 282         *power /= MICROWATT_PER_MILLIWATT;
 283
 284         return 0;
 285 }
 286
 287 static int devfreq_cooling_power2state(struct thermal_cooling_device *cdev,
 288                                        u32 power, unsigned long *state)
 289 {
 290         struct devfreq_cooling_device *dfc = cdev->devdata;
 291         struct devfreq *df = dfc->devfreq;
 292         struct devfreq_dev_status status;
 293         unsigned long freq, em_power_mw;
 294         struct em_perf_state *table;
 295         s32 est_power;
 296         int i;
 297
 298         mutex_lock(&df->lock);
 299         status = df->last_status;
 300         mutex_unlock(&df->lock);
 301
 302         freq = status.current_frequency;
 303
 304         if (dfc->power_ops && dfc->power_ops->get_real_power) {
 305                 /* Scale for resource utilization */
 306                 est_power = power * dfc->res_util;
 307                 est_power /= SCALE_ERROR_MITIGATION;
 308         } else {
 309                 /* Scale dynamic power for utilization */
 310                 _normalize_load(&status);
 311                 est_power = power << 10;
 312                 est_power /= status.busy_time;
 313         }
 314
 315         /*
 316          * Find the first cooling state that is within the power
 317          * budget. The EM power table is sorted ascending.
 318          */
 319         rcu_read_lock();
 320         table = em_perf_state_from_pd(dfc->em_pd);
 321         for (i = dfc->max_state; i > 0; i--) {
 322                 /* Convert EM power to milli-Watts to make safe comparison */
 323                 em_power_mw = table[i].power;
 324                 em_power_mw /= MICROWATT_PER_MILLIWATT;
 325                 if (est_power >= em_power_mw)
 326                         break;
 327         }
 328         rcu_read_unlock();
 329
 330         *state = dfc->max_state - i;
 331         dfc->capped_state = *state;
 332
 333         trace_thermal_power_devfreq_limit(cdev, freq, *state, power);
 334         return 0;
 335 }
 336
 337 /**
 338  * devfreq_cooling_gen_tables() - Generate frequency table.
 339  * @dfc:        Pointer to devfreq cooling device.
 340  * @num_opps:   Number of OPPs
 341  *
 342  * Generate frequency table which holds the frequencies in descending
 343  * order. That way its indexed by cooling device state. This is for
 344  * compatibility with drivers which do not register Energy Model.
 345  *
 346  * Return: 0 on success, negative error code on failure.
 347  */
 348 static int devfreq_cooling_gen_tables(struct devfreq_cooling_device *dfc,
 349                                       int num_opps)
 350 {
 351         struct devfreq *df = dfc->devfreq;
 352         struct device *dev = df->dev.parent;
 353         unsigned long freq;
 354         int i;
 355
 356         dfc->freq_table = kcalloc(num_opps, sizeof(*dfc->freq_table),
 357                              GFP_KERNEL);
 358         if (!dfc->freq_table)
 359                 return -ENOMEM;
 360
 361         for (i = 0, freq = ULONG_MAX; i < num_opps; i++, freq--) {
 362                 struct dev_pm_opp *opp;
 363
 364                 opp = dev_pm_opp_find_freq_floor(dev, &freq);
 365                 if (IS_ERR(opp)) {
 366                         kfree(dfc->freq_table);
 367                         return PTR_ERR(opp);
 368                 }
 369
 370                 dev_pm_opp_put(opp);
 371                 dfc->freq_table[i] = freq;
 372         }
 373
 374         return 0;
 375 }
 376
 377 /**
 378  * of_devfreq_cooling_register_power() - Register devfreq cooling device,
 379  *                                      with OF and power information.
 380  * @np: Pointer to OF device_node.
 381  * @df: Pointer to devfreq device.
 382  * @dfc_power:  Pointer to devfreq_cooling_power.
 383  *
 384  * Register a devfreq cooling device.  The available OPPs must be
 385  * registered on the device.
 386  *
 387  * If @dfc_power is provided, the cooling device is registered with the
 388  * power extensions.  For the power extensions to work correctly,
 389  * devfreq should use the simple_ondemand governor, other governors
 390  * are not currently supported.
 391  */
 392 struct thermal_cooling_device *
 393 of_devfreq_cooling_register_power(struct device_node *np, struct devfreq *df,
 394                                   struct devfreq_cooling_power *dfc_power)
 395 {
 396         struct thermal_cooling_device *cdev;
 397         struct device *dev = df->dev.parent;
 398         struct devfreq_cooling_device *dfc;
 399         struct em_perf_domain *em;
 400         struct thermal_cooling_device_ops *ops;
 401         char *name;
 402         int err, num_opps;
 403
 404
 405         dfc = kzalloc(sizeof(*dfc), GFP_KERNEL);
 406         if (!dfc)
 407                 return ERR_PTR(-ENOMEM);
 408
 409         dfc->devfreq = df;
 410
 411         ops = &dfc->cooling_ops;
 412         ops->get_max_state = devfreq_cooling_get_max_state;
 413         ops->get_cur_state = devfreq_cooling_get_cur_state;
 414         ops->set_cur_state = devfreq_cooling_set_cur_state;
 415
 416         em = em_pd_get(dev);
 417         if (em && !em_is_artificial(em)) {
 418                 dfc->em_pd = em;
 419                 ops->get_requested_power =
 420                         devfreq_cooling_get_requested_power;
 421                 ops->state2power = devfreq_cooling_state2power;
 422                 ops->power2state = devfreq_cooling_power2state;
 423
 424                 dfc->power_ops = dfc_power;
 425
 426                 num_opps = em_pd_nr_perf_states(dfc->em_pd);
 427         } else {
 428                 /* Backward compatibility for drivers which do not use IPA */
 429                 dev_dbg(dev, "missing proper EM for cooling device\n");
 430
 431                 num_opps = dev_pm_opp_get_opp_count(dev);
 432
 433                 err = devfreq_cooling_gen_tables(dfc, num_opps);
 434                 if (err)
 435                         goto free_dfc;
 436         }
 437
 438         if (num_opps <= 0) {
 439                 err = -EINVAL;
 440                 goto free_dfc;
 441         }
 442
 443         /* max_state is an index, not a counter */
 444         dfc->max_state = num_opps - 1;
 445
 446         err = dev_pm_qos_add_request(dev, &dfc->req_max_freq,
 447                                      DEV_PM_QOS_MAX_FREQUENCY,
 448                                      PM_QOS_MAX_FREQUENCY_DEFAULT_VALUE);
 449         if (err < 0)
 450                 goto free_table;
 451
 452         err = -ENOMEM;
 453         name = kasprintf(GFP_KERNEL, "devfreq-%s", dev_name(dev));
 454         if (!name)
 455                 goto remove_qos_req;
 456
 457         cdev = thermal_of_cooling_device_register(np, name, dfc, ops);
 458         kfree(name);
 459
 460         if (IS_ERR(cdev)) {
 461                 err = PTR_ERR(cdev);
 462                 dev_err(dev,
 463                         "Failed to register devfreq cooling device (%d)\n",
 464                         err);
 465                 goto remove_qos_req;
 466         }
 467
 468         dfc->cdev = cdev;
 469
 470         return cdev;
 471
 472 remove_qos_req:
 473         dev_pm_qos_remove_request(&dfc->req_max_freq);
 474 free_table:
 475         kfree(dfc->freq_table);
 476 free_dfc:
 477         kfree(dfc);
 478
 479         return ERR_PTR(err);
 480 }
 481 EXPORT_SYMBOL_GPL(of_devfreq_cooling_register_power);
 482
 483 /**
 484  * of_devfreq_cooling_register() - Register devfreq cooling device,
 485  *                                with OF information.
 486  * @np: Pointer to OF device_node.
 487  * @df: Pointer to devfreq device.
 488  */
 489 struct thermal_cooling_device *
 490 of_devfreq_cooling_register(struct device_node *np, struct devfreq *df)
 491 {
 492         return of_devfreq_cooling_register_power(np, df, NULL);
 493 }
 494 EXPORT_SYMBOL_GPL(of_devfreq_cooling_register);
 495
 496 /**
 497  * devfreq_cooling_register() - Register devfreq cooling device.
 498  * @df: Pointer to devfreq device.
 499  */
 500 struct thermal_cooling_device *devfreq_cooling_register(struct devfreq *df)
 501 {
 502         return of_devfreq_cooling_register(NULL, df);
 503 }
 504 EXPORT_SYMBOL_GPL(devfreq_cooling_register);
 505
 506 /**
 507  * devfreq_cooling_em_register() - Register devfreq cooling device with
 508  *              power information and automatically register Energy Model (EM)
 509  * @df:         Pointer to devfreq device.
 510  * @dfc_power:  Pointer to devfreq_cooling_power.
 511  *
 512  * Register a devfreq cooling device and automatically register EM. The
 513  * available OPPs must be registered for the device.
 514  *
 515  * If @dfc_power is provided, the cooling device is registered with the
 516  * power extensions. It is using the simple Energy Model which requires
 517  * "dynamic-power-coefficient" a devicetree property. To not break drivers
 518  * which miss that DT property, the function won't bail out when the EM
 519  * registration failed. The cooling device will be registered if everything
 520  * else is OK.
 521  */
 522 struct thermal_cooling_device *
 523 devfreq_cooling_em_register(struct devfreq *df,
 524                             struct devfreq_cooling_power *dfc_power)
 525 {
 526         struct thermal_cooling_device *cdev;
 527         struct device *dev;
 528         int ret;
 529
 530         if (IS_ERR_OR_NULL(df))
 531                 return ERR_PTR(-EINVAL);
 532
 533         dev = df->dev.parent;
 534
 535         ret = dev_pm_opp_of_register_em(dev, NULL);
 536         if (ret)
 537                 dev_dbg(dev, "Unable to register EM for devfreq cooling device (%d)\n",
 538                         ret);
 539
 540         cdev = of_devfreq_cooling_register_power(dev->of_node, df, dfc_power);
 541
 542         if (IS_ERR_OR_NULL(cdev))
 543                 em_dev_unregister_perf_domain(dev);
 544
 545         return cdev;
 546 }
 547 EXPORT_SYMBOL_GPL(devfreq_cooling_em_register);
 548
 549 /**
 550  * devfreq_cooling_unregister() - Unregister devfreq cooling device.
 551  * @cdev: Pointer to devfreq cooling device to unregister.
 552  *
 553  * Unregisters devfreq cooling device and related Energy Model if it was
 554  * present.
 555  */
 556 void devfreq_cooling_unregister(struct thermal_cooling_device *cdev)
 557 {
 558         struct devfreq_cooling_device *dfc;
 559         struct device *dev;
 560
 561         if (IS_ERR_OR_NULL(cdev))
 562                 return;
 563
 564         dfc = cdev->devdata;
 565         dev = dfc->devfreq->dev.parent;
 566
 567         thermal_cooling_device_unregister(dfc->cdev);
 568         dev_pm_qos_remove_request(&dfc->req_max_freq);
 569
 570         em_dev_unregister_perf_domain(dev);
 571
 572         kfree(dfc->freq_table);
 573         kfree(dfc);
 574 }
 575 EXPORT_SYMBOL_GPL(devfreq_cooling_unregister);