drivers/thermal/devfreq_cooling.c

   1 /*
   2  * devfreq_cooling: Thermal cooling device implementation for devices using
   3  *                  devfreq
   4  *
   5  * Copyright (C) 2014-2015 ARM Limited
   6  *
   7  * This program is free software; you can redistribute it and/or modify
   8  * it under the terms of the GNU General Public License version 2 as
   9  * published by the Free Software Foundation.
  10  *
  11  * This program is distributed "as is" WITHOUT ANY WARRANTY of any
  12  * kind, whether express or implied; without even the implied warranty
  13  * of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  14  * GNU General Public License for more details.
  15  *
  16  * TODO:
  17  *    - If OPPs are added or removed after devfreq cooling has
  18  *      registered, the devfreq cooling won't react to it.
  19  */
  20
  21 #include <linux/devfreq.h>
  22 #include <linux/devfreq_cooling.h>
  23 #include <linux/export.h>
  24 #include <linux/idr.h>
  25 #include <linux/slab.h>
  26 #include <linux/pm_opp.h>
  27 #include <linux/thermal.h>
  28
  29 #include <trace/events/thermal.h>
  30
  31 static DEFINE_IDA(devfreq_ida);
  32
  33 /**
  34  * struct devfreq_cooling_device - Devfreq cooling device
  35  * @id:         unique integer value corresponding to each
  36  *              devfreq_cooling_device registered.
  37  * @cdev:       Pointer to associated thermal cooling device.
  38  * @devfreq:    Pointer to associated devfreq device.
  39  * @cooling_state:      Current cooling state.
  40  * @power_table:        Pointer to table with maximum power draw for each
  41  *                      cooling state. State is the index into the table, and
  42  *                      the power is in mW.
  43  * @freq_table: Pointer to a table with the frequencies sorted in descending
  44  *              order.  You can index the table by cooling device state
  45  * @freq_table_size:    Size of the @freq_table and @power_table
  46  * @power_ops:  Pointer to devfreq_cooling_power, used to generate the
  47  *              @power_table.
  48  */
  49 struct devfreq_cooling_device {
  50         int id;
  51         struct thermal_cooling_device *cdev;
  52         struct devfreq *devfreq;
  53         unsigned long cooling_state;
  54         u32 *power_table;
  55         u32 *freq_table;
  56         size_t freq_table_size;
  57         struct devfreq_cooling_power *power_ops;
  58 };
  59
  60 /**
  61  * partition_enable_opps() - disable all opps above a given state
  62  * @dfc:        Pointer to devfreq we are operating on
  63  * @cdev_state: cooling device state we're setting
  64  *
  65  * Go through the OPPs of the device, enabling all OPPs until
  66  * @cdev_state and disabling those frequencies above it.
  67  */
  68 static int partition_enable_opps(struct devfreq_cooling_device *dfc,
  69                                  unsigned long cdev_state)
  70 {
  71         int i;
  72         struct device *dev = dfc->devfreq->dev.parent;
  73
  74         for (i = 0; i < dfc->freq_table_size; i++) {
  75                 struct dev_pm_opp *opp;
  76                 int ret = 0;
  77                 unsigned int freq = dfc->freq_table[i];
  78                 bool want_enable = i >= cdev_state ? true : false;
  79
  80                 opp = dev_pm_opp_find_freq_exact(dev, freq, !want_enable);
  81
  82                 if (PTR_ERR(opp) == -ERANGE)
  83                         continue;
  84                 else if (IS_ERR(opp))
  85                         return PTR_ERR(opp);
  86
  87                 dev_pm_opp_put(opp);
  88
  89                 if (want_enable)
  90                         ret = dev_pm_opp_enable(dev, freq);
  91                 else
  92                         ret = dev_pm_opp_disable(dev, freq);
  93
  94                 if (ret)
  95                         return ret;
  96         }
  97
  98         return 0;
  99 }
 100
 101 static int devfreq_cooling_get_max_state(struct thermal_cooling_device *cdev,
 102                                          unsigned long *state)
 103 {
 104         struct devfreq_cooling_device *dfc = cdev->devdata;
 105
 106         *state = dfc->freq_table_size - 1;
 107
 108         return 0;
 109 }
 110
 111 static int devfreq_cooling_get_cur_state(struct thermal_cooling_device *cdev,
 112                                          unsigned long *state)
 113 {
 114         struct devfreq_cooling_device *dfc = cdev->devdata;
 115
 116         *state = dfc->cooling_state;
 117
 118         return 0;
 119 }
 120
 121 static int devfreq_cooling_set_cur_state(struct thermal_cooling_device *cdev,
 122                                          unsigned long state)
 123 {
 124         struct devfreq_cooling_device *dfc = cdev->devdata;
 125         struct devfreq *df = dfc->devfreq;
 126         struct device *dev = df->dev.parent;
 127         int ret;
 128
 129         if (state == dfc->cooling_state)
 130                 return 0;
 131
 132         dev_dbg(dev, "Setting cooling state %lu\n", state);
 133
 134         if (state >= dfc->freq_table_size)
 135                 return -EINVAL;
 136
 137         ret = partition_enable_opps(dfc, state);
 138         if (ret)
 139                 return ret;
 140
 141         dfc->cooling_state = state;
 142
 143         return 0;
 144 }
 145
 146 /**
 147  * freq_get_state() - get the cooling state corresponding to a frequency
 148  * @dfc:        Pointer to devfreq cooling device
 149  * @freq:       frequency in Hz
 150  *
 151  * Return: the cooling state associated with the @freq, or
 152  * THERMAL_CSTATE_INVALID if it wasn't found.
 153  */
 154 static unsigned long
 155 freq_get_state(struct devfreq_cooling_device *dfc, unsigned long freq)
 156 {
 157         int i;
 158
 159         for (i = 0; i < dfc->freq_table_size; i++) {
 160                 if (dfc->freq_table[i] == freq)
 161                         return i;
 162         }
 163
 164         return THERMAL_CSTATE_INVALID;
 165 }
 166
 167 /**
 168  * get_static_power() - calculate the static power
 169  * @dfc:        Pointer to devfreq cooling device
 170  * @freq:       Frequency in Hz
 171  *
 172  * Calculate the static power in milliwatts using the supplied
 173  * get_static_power().  The current voltage is calculated using the
 174  * OPP library.  If no get_static_power() was supplied, assume the
 175  * static power is negligible.
 176  */
 177 static unsigned long
 178 get_static_power(struct devfreq_cooling_device *dfc, unsigned long freq)
 179 {
 180         struct devfreq *df = dfc->devfreq;
 181         struct device *dev = df->dev.parent;
 182         unsigned long voltage;
 183         struct dev_pm_opp *opp;
 184
 185         if (!dfc->power_ops->get_static_power)
 186                 return 0;
 187
 188         opp = dev_pm_opp_find_freq_exact(dev, freq, true);
 189         if (PTR_ERR(opp) == -ERANGE)
 190                 opp = dev_pm_opp_find_freq_exact(dev, freq, false);
 191
 192         if (IS_ERR(opp)) {
 193                 dev_err_ratelimited(dev, "Failed to find OPP for frequency %lu: %ld\n",
 194                                     freq, PTR_ERR(opp));
 195                 return 0;
 196         }
 197
 198         voltage = dev_pm_opp_get_voltage(opp) / 1000; /* mV */
 199         dev_pm_opp_put(opp);
 200
 201         if (voltage == 0) {
 202                 dev_err_ratelimited(dev,
 203                                     "Failed to get voltage for frequency %lu\n",
 204                                     freq);
 205                 return 0;
 206         }
 207
 208         return dfc->power_ops->get_static_power(df, voltage);
 209 }
 210
 211 /**
 212  * get_dynamic_power - calculate the dynamic power
 213  * @dfc:        Pointer to devfreq cooling device
 214  * @freq:       Frequency in Hz
 215  * @voltage:    Voltage in millivolts
 216  *
 217  * Calculate the dynamic power in milliwatts consumed by the device at
 218  * frequency @freq and voltage @voltage.  If the get_dynamic_power()
 219  * was supplied as part of the devfreq_cooling_power struct, then that
 220  * function is used.  Otherwise, a simple power model (Pdyn = Coeff *
 221  * Voltage^2 * Frequency) is used.
 222  */
 223 static unsigned long
 224 get_dynamic_power(struct devfreq_cooling_device *dfc, unsigned long freq,
 225                   unsigned long voltage)
 226 {
 227         u64 power;
 228         u32 freq_mhz;
 229         struct devfreq_cooling_power *dfc_power = dfc->power_ops;
 230
 231         if (dfc_power->get_dynamic_power)
 232                 return dfc_power->get_dynamic_power(dfc->devfreq, freq,
 233                                                     voltage);
 234
 235         freq_mhz = freq / 1000000;
 236         power = (u64)dfc_power->dyn_power_coeff * freq_mhz * voltage * voltage;
 237         do_div(power, 1000000000);
 238
 239         return power;
 240 }
 241
 242 static int devfreq_cooling_get_requested_power(struct thermal_cooling_device *cdev,
 243                                                struct thermal_zone_device *tz,
 244                                                u32 *power)
 245 {
 246         struct devfreq_cooling_device *dfc = cdev->devdata;
 247         struct devfreq *df = dfc->devfreq;
 248         struct devfreq_dev_status *status = &df->last_status;
 249         unsigned long state;
 250         unsigned long freq = status->current_frequency;
 251         u32 dyn_power, static_power;
 252
 253         /* Get dynamic power for state */
 254         state = freq_get_state(dfc, freq);
 255         if (state == THERMAL_CSTATE_INVALID)
 256                 return -EAGAIN;
 257
 258         dyn_power = dfc->power_table[state];
 259
 260         /* Scale dynamic power for utilization */
 261         dyn_power = (dyn_power * status->busy_time) / status->total_time;
 262
 263         /* Get static power */
 264         static_power = get_static_power(dfc, freq);
 265
 266         trace_thermal_power_devfreq_get_power(cdev, status, freq, dyn_power,
 267                                               static_power);
 268
 269         *power = dyn_power + static_power;
 270
 271         return 0;
 272 }
 273
 274 static int devfreq_cooling_state2power(struct thermal_cooling_device *cdev,
 275                                        struct thermal_zone_device *tz,
 276                                        unsigned long state,
 277                                        u32 *power)
 278 {
 279         struct devfreq_cooling_device *dfc = cdev->devdata;
 280         unsigned long freq;
 281         u32 static_power;
 282
 283         if (state >= dfc->freq_table_size)
 284                 return -EINVAL;
 285
 286         freq = dfc->freq_table[state];
 287         static_power = get_static_power(dfc, freq);
 288
 289         *power = dfc->power_table[state] + static_power;
 290         return 0;
 291 }
 292
 293 static int devfreq_cooling_power2state(struct thermal_cooling_device *cdev,
 294                                        struct thermal_zone_device *tz,
 295                                        u32 power, unsigned long *state)
 296 {
 297         struct devfreq_cooling_device *dfc = cdev->devdata;
 298         struct devfreq *df = dfc->devfreq;
 299         struct devfreq_dev_status *status = &df->last_status;
 300         unsigned long freq = status->current_frequency;
 301         unsigned long busy_time;
 302         s32 dyn_power;
 303         u32 static_power;
 304         int i;
 305
 306         static_power = get_static_power(dfc, freq);
 307
 308         dyn_power = power - static_power;
 309         dyn_power = dyn_power > 0 ? dyn_power : 0;
 310
 311         /* Scale dynamic power for utilization */
 312         busy_time = status->busy_time ?: 1;
 313         dyn_power = (dyn_power * status->total_time) / busy_time;
 314
 315         /*
 316          * Find the first cooling state that is within the power
 317          * budget for dynamic power.
 318          */
 319         for (i = 0; i < dfc->freq_table_size - 1; i++)
 320                 if (dyn_power >= dfc->power_table[i])
 321                         break;
 322
 323         *state = i;
 324         trace_thermal_power_devfreq_limit(cdev, freq, *state, power);
 325         return 0;
 326 }
 327
 328 static struct thermal_cooling_device_ops devfreq_cooling_ops = {
 329         .get_max_state = devfreq_cooling_get_max_state,
 330         .get_cur_state = devfreq_cooling_get_cur_state,
 331         .set_cur_state = devfreq_cooling_set_cur_state,
 332 };
 333
 334 /**
 335  * devfreq_cooling_gen_tables() - Generate power and freq tables.
 336  * @dfc: Pointer to devfreq cooling device.
 337  *
 338  * Generate power and frequency tables: the power table hold the
 339  * device's maximum power usage at each cooling state (OPP).  The
 340  * static and dynamic power using the appropriate voltage and
 341  * frequency for the state, is acquired from the struct
 342  * devfreq_cooling_power, and summed to make the maximum power draw.
 343  *
 344  * The frequency table holds the frequencies in descending order.
 345  * That way its indexed by cooling device state.
 346  *
 347  * The tables are malloced, and pointers put in dfc.  They must be
 348  * freed when unregistering the devfreq cooling device.
 349  *
 350  * Return: 0 on success, negative error code on failure.
 351  */
 352 static int devfreq_cooling_gen_tables(struct devfreq_cooling_device *dfc)
 353 {
 354         struct devfreq *df = dfc->devfreq;
 355         struct device *dev = df->dev.parent;
 356         int ret, num_opps;
 357         unsigned long freq;
 358         u32 *power_table = NULL;
 359         u32 *freq_table;
 360         int i;
 361
 362         num_opps = dev_pm_opp_get_opp_count(dev);
 363
 364         if (dfc->power_ops) {
 365                 power_table = kcalloc(num_opps, sizeof(*power_table),
 366                                       GFP_KERNEL);
 367                 if (!power_table)
 368                         return -ENOMEM;
 369         }
 370
 371         freq_table = kcalloc(num_opps, sizeof(*freq_table),
 372                              GFP_KERNEL);
 373         if (!freq_table) {
 374                 ret = -ENOMEM;
 375                 goto free_power_table;
 376         }
 377
 378         for (i = 0, freq = ULONG_MAX; i < num_opps; i++, freq--) {
 379                 unsigned long power_dyn, voltage;
 380                 struct dev_pm_opp *opp;
 381
 382                 opp = dev_pm_opp_find_freq_floor(dev, &freq);
 383                 if (IS_ERR(opp)) {
 384                         ret = PTR_ERR(opp);
 385                         goto free_tables;
 386                 }
 387
 388                 voltage = dev_pm_opp_get_voltage(opp) / 1000; /* mV */
 389                 dev_pm_opp_put(opp);
 390
 391                 if (dfc->power_ops) {
 392                         power_dyn = get_dynamic_power(dfc, freq, voltage);
 393
 394                         dev_dbg(dev, "Dynamic power table: %lu MHz @ %lu mV: %lu = %lu mW\n",
 395                                 freq / 1000000, voltage, power_dyn, power_dyn);
 396
 397                         power_table[i] = power_dyn;
 398                 }
 399
 400                 freq_table[i] = freq;
 401         }
 402
 403         if (dfc->power_ops)
 404                 dfc->power_table = power_table;
 405
 406         dfc->freq_table = freq_table;
 407         dfc->freq_table_size = num_opps;
 408
 409         return 0;
 410
 411 free_tables:
 412         kfree(freq_table);
 413 free_power_table:
 414         kfree(power_table);
 415
 416         return ret;
 417 }
 418
 419 /**
 420  * of_devfreq_cooling_register_power() - Register devfreq cooling device,
 421  *                                      with OF and power information.
 422  * @np: Pointer to OF device_node.
 423  * @df: Pointer to devfreq device.
 424  * @dfc_power:  Pointer to devfreq_cooling_power.
 425  *
 426  * Register a devfreq cooling device.  The available OPPs must be
 427  * registered on the device.
 428  *
 429  * If @dfc_power is provided, the cooling device is registered with the
 430  * power extensions.  For the power extensions to work correctly,
 431  * devfreq should use the simple_ondemand governor, other governors
 432  * are not currently supported.
 433  */
 434 struct thermal_cooling_device *
 435 of_devfreq_cooling_register_power(struct device_node *np, struct devfreq *df,
 436                                   struct devfreq_cooling_power *dfc_power)
 437 {
 438         struct thermal_cooling_device *cdev;
 439         struct devfreq_cooling_device *dfc;
 440         char dev_name[THERMAL_NAME_LENGTH];
 441         int err;
 442
 443         dfc = kzalloc(sizeof(*dfc), GFP_KERNEL);
 444         if (!dfc)
 445                 return ERR_PTR(-ENOMEM);
 446
 447         dfc->devfreq = df;
 448
 449         if (dfc_power) {
 450                 dfc->power_ops = dfc_power;
 451
 452                 devfreq_cooling_ops.get_requested_power =
 453                         devfreq_cooling_get_requested_power;
 454                 devfreq_cooling_ops.state2power = devfreq_cooling_state2power;
 455                 devfreq_cooling_ops.power2state = devfreq_cooling_power2state;
 456         }
 457
 458         err = devfreq_cooling_gen_tables(dfc);
 459         if (err)
 460                 goto free_dfc;
 461
 462         err = ida_simple_get(&devfreq_ida, 0, 0, GFP_KERNEL);
 463         if (err < 0)
 464                 goto free_tables;
 465         dfc->id = err;
 466
 467         snprintf(dev_name, sizeof(dev_name), "thermal-devfreq-%d", dfc->id);
 468
 469         cdev = thermal_of_cooling_device_register(np, dev_name, dfc,
 470                                                   &devfreq_cooling_ops);
 471         if (IS_ERR(cdev)) {
 472                 err = PTR_ERR(cdev);
 473                 dev_err(df->dev.parent,
 474                         "Failed to register devfreq cooling device (%d)\n",
 475                         err);
 476                 goto release_ida;
 477         }
 478
 479         dfc->cdev = cdev;
 480
 481         return cdev;
 482
 483 release_ida:
 484         ida_simple_remove(&devfreq_ida, dfc->id);
 485 free_tables:
 486         kfree(dfc->power_table);
 487         kfree(dfc->freq_table);
 488 free_dfc:
 489         kfree(dfc);
 490
 491         return ERR_PTR(err);
 492 }
 493 EXPORT_SYMBOL_GPL(of_devfreq_cooling_register_power);
 494
 495 /**
 496  * of_devfreq_cooling_register() - Register devfreq cooling device,
 497  *                                with OF information.
 498  * @np: Pointer to OF device_node.
 499  * @df: Pointer to devfreq device.
 500  */
 501 struct thermal_cooling_device *
 502 of_devfreq_cooling_register(struct device_node *np, struct devfreq *df)
 503 {
 504         return of_devfreq_cooling_register_power(np, df, NULL);
 505 }
 506 EXPORT_SYMBOL_GPL(of_devfreq_cooling_register);
 507
 508 /**
 509  * devfreq_cooling_register() - Register devfreq cooling device.
 510  * @df: Pointer to devfreq device.
 511  */
 512 struct thermal_cooling_device *devfreq_cooling_register(struct devfreq *df)
 513 {
 514         return of_devfreq_cooling_register(NULL, df);
 515 }
 516 EXPORT_SYMBOL_GPL(devfreq_cooling_register);
 517
 518 /**
 519  * devfreq_cooling_unregister() - Unregister devfreq cooling device.
 520  * @dfc: Pointer to devfreq cooling device to unregister.
 521  */
 522 void devfreq_cooling_unregister(struct thermal_cooling_device *cdev)
 523 {
 524         struct devfreq_cooling_device *dfc;
 525
 526         if (!cdev)
 527                 return;
 528
 529         dfc = cdev->devdata;
 530
 531         thermal_cooling_device_unregister(dfc->cdev);
 532         ida_simple_remove(&devfreq_ida, dfc->id);
 533         kfree(dfc->power_table);
 534         kfree(dfc->freq_table);
 535
 536         kfree(dfc);
 537 }
 538 EXPORT_SYMBOL_GPL(devfreq_cooling_unregister);