sys/kern/kern_cpu.c

   1 /*-
   2  * Copyright (c) 2004-2005 Nate Lawson (SDG)
   3  * All rights reserved.
   4  *
   5  * Redistribution and use in source and binary forms, with or without
   6  * modification, are permitted provided that the following conditions
   7  * are met:
   8  * 1. Redistributions of source code must retain the above copyright
   9  *    notice, this list of conditions and the following disclaimer.
  10  * 2. Redistributions in binary form must reproduce the above copyright
  11  *    notice, this list of conditions and the following disclaimer in the
  12  *    documentation and/or other materials provided with the distribution.
  13  *
  14  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
  15  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
  16  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
  17  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
  18  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
  19  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
  20  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
  21  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
  22  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
  23  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
  24  * SUCH DAMAGE.
  25  */
  26
  27 #include <sys/cdefs.h>
  28 __FBSDID("$FreeBSD$");
  29
  30 #include <sys/param.h>
  31 #include <sys/bus.h>
  32 #include <sys/cpu.h>
  33 #include <sys/eventhandler.h>
  34 #include <sys/kernel.h>
  35 #include <sys/lock.h>
  36 #include <sys/malloc.h>
  37 #include <sys/module.h>
  38 #include <sys/proc.h>
  39 #include <sys/queue.h>
  40 #include <sys/sched.h>
  41 #include <sys/sysctl.h>
  42 #include <sys/systm.h>
  43 #include <sys/sbuf.h>
  44 #include <sys/sx.h>
  45 #include <sys/timetc.h>
  46
  47 #include "cpufreq_if.h"
  48
  49 /*
  50  * Common CPU frequency glue code.  Drivers for specific hardware can
  51  * attach this interface to allow users to get/set the CPU frequency.
  52  */
  53
  54 /*
  55  * Number of levels we can handle.  Levels are synthesized from settings
  56  * so for M settings and N drivers, there may be M*N levels.
  57  */
  58 #define CF_MAX_LEVELS   64
  59
  60 struct cpufreq_softc {
  61         struct sx                       lock;
  62         struct cf_level                 curr_level;
  63         int                             curr_priority;
  64         struct cf_level                 saved_level;
  65         int                             saved_priority;
  66         struct cf_level_lst             all_levels;
  67         int                             all_count;
  68         int                             max_mhz;
  69         device_t                        dev;
  70         struct sysctl_ctx_list          sysctl_ctx;
  71 };
  72
  73 struct cf_setting_array {
  74         struct cf_setting               sets[MAX_SETTINGS];
  75         int                             count;
  76         TAILQ_ENTRY(cf_setting_array)   link;
  77 };
  78
  79 TAILQ_HEAD(cf_setting_lst, cf_setting_array);
  80
  81 #define CF_MTX_INIT(x)          sx_init((x), "cpufreq lock")
  82 #define CF_MTX_LOCK(x)          sx_xlock((x))
  83 #define CF_MTX_UNLOCK(x)        sx_xunlock((x))
  84 #define CF_MTX_ASSERT(x)        sx_assert((x), SX_XLOCKED)
  85
  86 #define CF_DEBUG(msg...)        do {            \
  87         if (cf_verbose)                         \
  88                 printf("cpufreq: " msg);        \
  89         } while (0)
  90
  91 static int      cpufreq_attach(device_t dev);
  92 static int      cpufreq_detach(device_t dev);
  93 static void     cpufreq_evaluate(void *arg);
  94 static int      cf_set_method(device_t dev, const struct cf_level *level,
  95                     int priority);
  96 static int      cf_get_method(device_t dev, struct cf_level *level);
  97 static int      cf_levels_method(device_t dev, struct cf_level *levels,
  98                     int *count);
  99 static int      cpufreq_insert_abs(struct cpufreq_softc *sc,
 100                     struct cf_setting *sets, int count);
 101 static int      cpufreq_expand_set(struct cpufreq_softc *sc,
 102                     struct cf_setting_array *set_arr);
 103 static struct cf_level *cpufreq_dup_set(struct cpufreq_softc *sc,
 104                     struct cf_level *dup, struct cf_setting *set);
 105 static int      cpufreq_curr_sysctl(SYSCTL_HANDLER_ARGS);
 106 static int      cpufreq_levels_sysctl(SYSCTL_HANDLER_ARGS);
 107 static int      cpufreq_settings_sysctl(SYSCTL_HANDLER_ARGS);
 108
 109 static device_method_t cpufreq_methods[] = {
 110         DEVMETHOD(device_probe,         bus_generic_probe),
 111         DEVMETHOD(device_attach,        cpufreq_attach),
 112         DEVMETHOD(device_detach,        cpufreq_detach),
 113
 114         DEVMETHOD(cpufreq_set,          cf_set_method),
 115         DEVMETHOD(cpufreq_get,          cf_get_method),
 116         DEVMETHOD(cpufreq_levels,       cf_levels_method),
 117         {0, 0}
 118 };
 119 static driver_t cpufreq_driver = {
 120         "cpufreq", cpufreq_methods, sizeof(struct cpufreq_softc)
 121 };
 122 static devclass_t cpufreq_dc;
 123 DRIVER_MODULE(cpufreq, cpu, cpufreq_driver, cpufreq_dc, 0, 0);
 124
 125 static eventhandler_tag cf_ev_tag;
 126
 127 static int              cf_lowest_freq;
 128 static int              cf_verbose;
 129 TUNABLE_INT("debug.cpufreq.lowest", &cf_lowest_freq);
 130 TUNABLE_INT("debug.cpufreq.verbose", &cf_verbose);
 131 SYSCTL_NODE(_debug, OID_AUTO, cpufreq, CTLFLAG_RD, NULL, "cpufreq debugging");
 132 SYSCTL_INT(_debug_cpufreq, OID_AUTO, lowest, CTLFLAG_RW, &cf_lowest_freq, 1,
 133     "Don't provide levels below this frequency.");
 134 SYSCTL_INT(_debug_cpufreq, OID_AUTO, verbose, CTLFLAG_RW, &cf_verbose, 1,
 135     "Print verbose debugging messages");
 136
 137 static int
 138 cpufreq_attach(device_t dev)
 139 {
 140         struct cpufreq_softc *sc;
 141         device_t parent;
 142         int numdevs;
 143
 144         CF_DEBUG("initializing %s\n", device_get_nameunit(dev));
 145         sc = device_get_softc(dev);
 146         parent = device_get_parent(dev);
 147         sc->dev = dev;
 148         sysctl_ctx_init(&sc->sysctl_ctx);
 149         TAILQ_INIT(&sc->all_levels);
 150         CF_MTX_INIT(&sc->lock);
 151         sc->curr_level.total_set.freq = CPUFREQ_VAL_UNKNOWN;
 152         sc->saved_level.total_set.freq = CPUFREQ_VAL_UNKNOWN;
 153         sc->max_mhz = CPUFREQ_VAL_UNKNOWN;
 154
 155         /*
 156          * Only initialize one set of sysctls for all CPUs.  In the future,
 157          * if multiple CPUs can have different settings, we can move these
 158          * sysctls to be under every CPU instead of just the first one.
 159          */
 160         numdevs = devclass_get_count(cpufreq_dc);
 161         if (numdevs > 1)
 162                 return (0);
 163
 164         CF_DEBUG("initializing one-time data for %s\n",
 165             device_get_nameunit(dev));
 166         SYSCTL_ADD_PROC(&sc->sysctl_ctx,
 167             SYSCTL_CHILDREN(device_get_sysctl_tree(parent)),
 168             OID_AUTO, "freq", CTLTYPE_INT | CTLFLAG_RW, sc, 0,
 169             cpufreq_curr_sysctl, "I", "Current CPU frequency");
 170         SYSCTL_ADD_PROC(&sc->sysctl_ctx,
 171             SYSCTL_CHILDREN(device_get_sysctl_tree(parent)),
 172             OID_AUTO, "freq_levels", CTLTYPE_STRING | CTLFLAG_RD, sc, 0,
 173             cpufreq_levels_sysctl, "A", "CPU frequency levels");
 174         cf_ev_tag = EVENTHANDLER_REGISTER(cpufreq_changed, cpufreq_evaluate,
 175             NULL, EVENTHANDLER_PRI_ANY);
 176
 177         return (0);
 178 }
 179
 180 static int
 181 cpufreq_detach(device_t dev)
 182 {
 183         struct cpufreq_softc *sc;
 184         int numdevs;
 185
 186         CF_DEBUG("shutdown %s\n", device_get_nameunit(dev));
 187         sc = device_get_softc(dev);
 188         sysctl_ctx_free(&sc->sysctl_ctx);
 189
 190         /* Only clean up these resources when the last device is detaching. */
 191         numdevs = devclass_get_count(cpufreq_dc);
 192         if (numdevs == 1) {
 193                 CF_DEBUG("final shutdown for %s\n", device_get_nameunit(dev));
 194                 EVENTHANDLER_DEREGISTER(cpufreq_changed, cf_ev_tag);
 195         }
 196
 197         return (0);
 198 }
 199
 200 static void
 201 cpufreq_evaluate(void *arg)
 202 {
 203         /* TODO: Re-evaluate when notified of changes to drivers. */
 204 }
 205
 206 static int
 207 cf_set_method(device_t dev, const struct cf_level *level, int priority)
 208 {
 209         struct cpufreq_softc *sc;
 210         const struct cf_setting *set;
 211         struct pcpu *pc;
 212         int cpu_id, error, i;
 213
 214         sc = device_get_softc(dev);
 215         error = 0;
 216         set = NULL;
 217
 218         /*
 219          * Check that the TSC isn't being used as a timecounter.
 220          * If it is, then return EBUSY and refuse to change the
 221          * clock speed.
 222          */
 223         if (strcmp(timecounter->tc_name, "TSC") == 0)
 224                 return (EBUSY);
 225
 226         /*
 227          * If the caller didn't specify a level and one is saved, prepare to
 228          * restore the saved level.  If none has been saved, return an error.
 229          * If they did specify one, but the requested level has a lower
 230          * priority, don't allow the new level right now.
 231          */
 232         CF_MTX_LOCK(&sc->lock);
 233         if (level == NULL) {
 234                 if (sc->saved_level.total_set.freq != CPUFREQ_VAL_UNKNOWN) {
 235                         level = &sc->saved_level;
 236                         priority = sc->saved_priority;
 237                         CF_DEBUG("restoring saved level, freq %d prio %d\n",
 238                             level->total_set.freq, priority);
 239                 } else {
 240                         CF_DEBUG("NULL level, no saved level\n");
 241                         error = ENXIO;
 242                         goto out;
 243                 }
 244         } else if (priority < sc->curr_priority) {
 245                 CF_DEBUG("ignoring, curr prio %d less than %d\n", priority,
 246                     sc->curr_priority);
 247                 error = EPERM;
 248                 goto out;
 249         }
 250
 251         /* Reject levels that are below our specified threshold. */
 252         if (level->total_set.freq <= cf_lowest_freq) {
 253                 CF_DEBUG("rejecting freq %d, less than %d limit\n",
 254                     level->total_set.freq, cf_lowest_freq);
 255                 error = EINVAL;
 256                 goto out;
 257         }
 258
 259         /* If already at this level, just return. */
 260         if (CPUFREQ_CMP(sc->curr_level.total_set.freq, level->total_set.freq)) {
 261                 CF_DEBUG("skipping freq %d, same as current level %d\n",
 262                     level->total_set.freq, sc->curr_level.total_set.freq);
 263                 goto out;
 264         }
 265
 266         /* First, set the absolute frequency via its driver. */
 267         set = &level->abs_set;
 268         if (set->dev) {
 269                 if (!device_is_attached(set->dev)) {
 270                         error = ENXIO;
 271                         goto out;
 272                 }
 273
 274                 /* Bind to the target CPU before switching, if necessary. */
 275                 cpu_id = PCPU_GET(cpuid);
 276                 pc = cpu_get_pcpu(set->dev);
 277                 if (cpu_id != pc->pc_cpuid) {
 278                         mtx_lock_spin(&sched_lock);
 279                         sched_bind(curthread, pc->pc_cpuid);
 280                         mtx_unlock_spin(&sched_lock);
 281                 }
 282                 CF_DEBUG("setting abs freq %d on %s (cpu %d)\n", set->freq,
 283                     device_get_nameunit(set->dev), PCPU_GET(cpuid));
 284                 error = CPUFREQ_DRV_SET(set->dev, set);
 285                 if (cpu_id != pc->pc_cpuid) {
 286                         mtx_lock_spin(&sched_lock);
 287                         sched_unbind(curthread);
 288                         mtx_unlock_spin(&sched_lock);
 289                 }
 290                 if (error) {
 291                         goto out;
 292                 }
 293         }
 294
 295         /* Next, set any/all relative frequencies via their drivers. */
 296         for (i = 0; i < level->rel_count; i++) {
 297                 set = &level->rel_set[i];
 298                 if (!device_is_attached(set->dev)) {
 299                         error = ENXIO;
 300                         goto out;
 301                 }
 302
 303                 /* Bind to the target CPU before switching, if necessary. */
 304                 cpu_id = PCPU_GET(cpuid);
 305                 pc = cpu_get_pcpu(set->dev);
 306                 if (cpu_id != pc->pc_cpuid) {
 307                         mtx_lock_spin(&sched_lock);
 308                         sched_bind(curthread, pc->pc_cpuid);
 309                         mtx_unlock_spin(&sched_lock);
 310                 }
 311                 CF_DEBUG("setting rel freq %d on %s (cpu %d)\n", set->freq,
 312                     device_get_nameunit(set->dev), PCPU_GET(cpuid));
 313                 error = CPUFREQ_DRV_SET(set->dev, set);
 314                 if (cpu_id != pc->pc_cpuid) {
 315                         mtx_lock_spin(&sched_lock);
 316                         sched_unbind(curthread);
 317                         mtx_unlock_spin(&sched_lock);
 318                 }
 319                 if (error) {
 320                         /* XXX Back out any successful setting? */
 321                         goto out;
 322                 }
 323         }
 324
 325         /* If we were restoring a saved state, reset it to "unused". */
 326         if (level == &sc->saved_level) {
 327                 CF_DEBUG("resetting saved level\n");
 328                 sc->saved_level.total_set.freq = CPUFREQ_VAL_UNKNOWN;
 329                 sc->saved_priority = 0;
 330         }
 331
 332         /*
 333          * Before recording the current level, check if we're going to a
 334          * higher priority and have not saved a level yet.  If so, save the
 335          * previous level and priority.
 336          */
 337         if (sc->curr_level.total_set.freq != CPUFREQ_VAL_UNKNOWN &&
 338             sc->saved_level.total_set.freq == CPUFREQ_VAL_UNKNOWN &&
 339             priority > sc->curr_priority) {
 340                 CF_DEBUG("saving level, freq %d prio %d\n",
 341                     sc->curr_level.total_set.freq, sc->curr_priority);
 342                 sc->saved_level = sc->curr_level;
 343                 sc->saved_priority = sc->curr_priority;
 344         }
 345         sc->curr_level = *level;
 346         sc->curr_priority = priority;
 347         error = 0;
 348
 349 out:
 350         CF_MTX_UNLOCK(&sc->lock);
 351         if (error && set)
 352                 device_printf(set->dev, "set freq failed, err %d\n", error);
 353         return (error);
 354 }
 355
 356 static int
 357 cf_get_method(device_t dev, struct cf_level *level)
 358 {
 359         struct cpufreq_softc *sc;
 360         struct cf_level *levels;
 361         struct cf_setting *curr_set, set;
 362         struct pcpu *pc;
 363         device_t *devs;
 364         int count, error, i, numdevs;
 365         uint64_t rate;
 366
 367         sc = device_get_softc(dev);
 368         error = 0;
 369         levels = NULL;
 370
 371         /* If we already know the current frequency, we're done. */
 372         CF_MTX_LOCK(&sc->lock);
 373         curr_set = &sc->curr_level.total_set;
 374         if (curr_set->freq != CPUFREQ_VAL_UNKNOWN) {
 375                 CF_DEBUG("get returning known freq %d\n", curr_set->freq);
 376                 goto out;
 377         }
 378         CF_MTX_UNLOCK(&sc->lock);
 379
 380         /*
 381          * We need to figure out the current level.  Loop through every
 382          * driver, getting the current setting.  Then, attempt to get a best
 383          * match of settings against each level.
 384          */
 385         count = CF_MAX_LEVELS;
 386         levels = malloc(count * sizeof(*levels), M_TEMP, M_NOWAIT);
 387         if (levels == NULL)
 388                 return (ENOMEM);
 389         error = CPUFREQ_LEVELS(sc->dev, levels, &count);
 390         if (error) {
 391                 if (error == E2BIG)
 392                         printf("cpufreq: need to increase CF_MAX_LEVELS\n");
 393                 free(levels, M_TEMP);
 394                 return (error);
 395         }
 396         error = device_get_children(device_get_parent(dev), &devs, &numdevs);
 397         if (error) {
 398                 free(levels, M_TEMP);
 399                 return (error);
 400         }
 401
 402         /*
 403          * Reacquire the lock and search for the given level.
 404          *
 405          * XXX Note: this is not quite right since we really need to go
 406          * through each level and compare both absolute and relative
 407          * settings for each driver in the system before making a match.
 408          * The estimation code below catches this case though.
 409          */
 410         CF_MTX_LOCK(&sc->lock);
 411         for (i = 0; i < numdevs && curr_set->freq == CPUFREQ_VAL_UNKNOWN; i++) {
 412                 if (!device_is_attached(devs[i]))
 413                         continue;
 414                 error = CPUFREQ_DRV_GET(devs[i], &set);
 415                 if (error)
 416                         continue;
 417                 for (i = 0; i < count; i++) {
 418                         if (CPUFREQ_CMP(set.freq, levels[i].total_set.freq)) {
 419                                 sc->curr_level = levels[i];
 420                                 break;
 421                         }
 422                 }
 423         }
 424         free(devs, M_TEMP);
 425         if (curr_set->freq != CPUFREQ_VAL_UNKNOWN) {
 426                 CF_DEBUG("get matched freq %d from drivers\n", curr_set->freq);
 427                 goto out;
 428         }
 429
 430         /*
 431          * We couldn't find an exact match, so attempt to estimate and then
 432          * match against a level.
 433          */
 434         pc = cpu_get_pcpu(dev);
 435         if (pc == NULL) {
 436                 error = ENXIO;
 437                 goto out;
 438         }
 439         cpu_est_clockrate(pc->pc_cpuid, &rate);
 440         rate /= 1000000;
 441         for (i = 0; i < count; i++) {
 442                 if (CPUFREQ_CMP(rate, levels[i].total_set.freq)) {
 443                         sc->curr_level = levels[i];
 444                         CF_DEBUG("get estimated freq %d\n", curr_set->freq);
 445                         break;
 446                 }
 447         }
 448
 449 out:
 450         if (error == 0)
 451                 *level = sc->curr_level;
 452
 453         CF_MTX_UNLOCK(&sc->lock);
 454         if (levels)
 455                 free(levels, M_TEMP);
 456         return (error);
 457 }
 458
 459 static int
 460 cf_levels_method(device_t dev, struct cf_level *levels, int *count)
 461 {
 462         struct cf_setting_array *set_arr;
 463         struct cf_setting_lst rel_sets;
 464         struct cpufreq_softc *sc;
 465         struct cf_level *lev;
 466         struct cf_setting *sets;
 467         struct pcpu *pc;
 468         device_t *devs;
 469         int error, i, numdevs, set_count, type;
 470         uint64_t rate;
 471
 472         if (levels == NULL || count == NULL)
 473                 return (EINVAL);
 474
 475         TAILQ_INIT(&rel_sets);
 476         sc = device_get_softc(dev);
 477         error = device_get_children(device_get_parent(dev), &devs, &numdevs);
 478         if (error)
 479                 return (error);
 480         sets = malloc(MAX_SETTINGS * sizeof(*sets), M_TEMP, M_NOWAIT);
 481         if (sets == NULL) {
 482                 free(devs, M_TEMP);
 483                 return (ENOMEM);
 484         }
 485
 486         /* Get settings from all cpufreq drivers. */
 487         CF_MTX_LOCK(&sc->lock);
 488         for (i = 0; i < numdevs; i++) {
 489                 /* Skip devices that aren't ready. */
 490                 if (!device_is_attached(devs[i]))
 491                         continue;
 492
 493                 /*
 494                  * Get settings, skipping drivers that offer no settings or
 495                  * provide settings for informational purposes only.
 496                  */
 497                 error = CPUFREQ_DRV_TYPE(devs[i], &type);
 498                 if (error || (type & CPUFREQ_FLAG_INFO_ONLY)) {
 499                         if (error == 0) {
 500                                 CF_DEBUG("skipping info-only driver %s\n",
 501                                     device_get_nameunit(devs[i]));
 502                         }
 503                         continue;
 504                 }
 505                 set_count = MAX_SETTINGS;
 506                 error = CPUFREQ_DRV_SETTINGS(devs[i], sets, &set_count);
 507                 if (error || set_count == 0)
 508                         continue;
 509
 510                 /* Add the settings to our absolute/relative lists. */
 511                 switch (type & CPUFREQ_TYPE_MASK) {
 512                 case CPUFREQ_TYPE_ABSOLUTE:
 513                         error = cpufreq_insert_abs(sc, sets, set_count);
 514                         break;
 515                 case CPUFREQ_TYPE_RELATIVE:
 516                         CF_DEBUG("adding %d relative settings\n", set_count);
 517                         set_arr = malloc(sizeof(*set_arr), M_TEMP, M_NOWAIT);
 518                         if (set_arr == NULL) {
 519                                 error = ENOMEM;
 520                                 goto out;
 521                         }
 522                         bcopy(sets, set_arr->sets, set_count * sizeof(*sets));
 523                         set_arr->count = set_count;
 524                         TAILQ_INSERT_TAIL(&rel_sets, set_arr, link);
 525                         break;
 526                 default:
 527                         error = EINVAL;
 528                 }
 529                 if (error)
 530                         goto out;
 531         }
 532
 533         /*
 534          * If there are no absolute levels, create a fake one at 100%.  We
 535          * then cache the clockrate for later use as our base frequency.
 536          *
 537          * XXX This assumes that the first time through, if we only have
 538          * relative drivers, the CPU is currently running at 100%.
 539          */
 540         if (TAILQ_EMPTY(&sc->all_levels)) {
 541                 if (sc->max_mhz == CPUFREQ_VAL_UNKNOWN) {
 542                         pc = cpu_get_pcpu(dev);
 543                         cpu_est_clockrate(pc->pc_cpuid, &rate);
 544                         sc->max_mhz = rate / 1000000;
 545                 }
 546                 memset(&sets[0], CPUFREQ_VAL_UNKNOWN, sizeof(*sets));
 547                 sets[0].freq = sc->max_mhz;
 548                 sets[0].dev = NULL;
 549                 error = cpufreq_insert_abs(sc, sets, 1);
 550                 if (error)
 551                         goto out;
 552         }
 553
 554         /* Create a combined list of absolute + relative levels. */
 555         TAILQ_FOREACH(set_arr, &rel_sets, link)
 556                 cpufreq_expand_set(sc, set_arr);
 557
 558         /* If the caller doesn't have enough space, return the actual count. */
 559         if (sc->all_count > *count) {
 560                 *count = sc->all_count;
 561                 error = E2BIG;
 562                 goto out;
 563         }
 564
 565         /* Finally, output the list of levels. */
 566         i = 0;
 567         TAILQ_FOREACH(lev, &sc->all_levels, link) {
 568                 /* Skip levels that have a frequency that is too low. */
 569                 if (lev->total_set.freq <= cf_lowest_freq) {
 570                         sc->all_count--;
 571                         continue;
 572                 }
 573
 574                 levels[i] = *lev;
 575                 i++;
 576         }
 577         *count = sc->all_count;
 578         error = 0;
 579
 580 out:
 581         /* Clear all levels since we regenerate them each time. */
 582         while ((lev = TAILQ_FIRST(&sc->all_levels)) != NULL) {
 583                 TAILQ_REMOVE(&sc->all_levels, lev, link);
 584                 free(lev, M_TEMP);
 585         }
 586         sc->all_count = 0;
 587
 588         CF_MTX_UNLOCK(&sc->lock);
 589         while ((set_arr = TAILQ_FIRST(&rel_sets)) != NULL) {
 590                 TAILQ_REMOVE(&rel_sets, set_arr, link);
 591                 free(set_arr, M_TEMP);
 592         }
 593         free(devs, M_TEMP);
 594         free(sets, M_TEMP);
 595         return (error);
 596 }
 597
 598 /*
 599  * Create levels for an array of absolute settings and insert them in
 600  * sorted order in the specified list.
 601  */
 602 static int
 603 cpufreq_insert_abs(struct cpufreq_softc *sc, struct cf_setting *sets,
 604     int count)
 605 {
 606         struct cf_level_lst *list;
 607         struct cf_level *level, *search;
 608         int i;
 609
 610         CF_MTX_ASSERT(&sc->lock);
 611
 612         list = &sc->all_levels;
 613         for (i = 0; i < count; i++) {
 614                 level = malloc(sizeof(*level), M_TEMP, M_NOWAIT | M_ZERO);
 615                 if (level == NULL)
 616                         return (ENOMEM);
 617                 level->abs_set = sets[i];
 618                 level->total_set = sets[i];
 619                 level->total_set.dev = NULL;
 620                 sc->all_count++;
 621
 622                 if (TAILQ_EMPTY(list)) {
 623                         CF_DEBUG("adding abs setting %d at head\n",
 624                             sets[i].freq);
 625                         TAILQ_INSERT_HEAD(list, level, link);
 626                         continue;
 627                 }
 628
 629                 TAILQ_FOREACH_REVERSE(search, list, cf_level_lst, link) {
 630                         if (sets[i].freq <= search->total_set.freq) {
 631                                 CF_DEBUG("adding abs setting %d after %d\n",
 632                                     sets[i].freq, search->total_set.freq);
 633                                 TAILQ_INSERT_AFTER(list, search, level, link);
 634                                 break;
 635                         }
 636                 }
 637         }
 638         return (0);
 639 }
 640
 641 /*
 642  * Expand a group of relative settings, creating derived levels from them.
 643  */
 644 static int
 645 cpufreq_expand_set(struct cpufreq_softc *sc, struct cf_setting_array *set_arr)
 646 {
 647         struct cf_level *fill, *search;
 648         struct cf_setting *set;
 649         int i;
 650
 651         CF_MTX_ASSERT(&sc->lock);
 652
 653         TAILQ_FOREACH(search, &sc->all_levels, link) {
 654                 /* Skip this level if we've already modified it. */
 655                 for (i = 0; i < search->rel_count; i++) {
 656                         if (search->rel_set[i].dev == set_arr->sets[0].dev)
 657                                 break;
 658                 }
 659                 if (i != search->rel_count) {
 660                         CF_DEBUG("skipping modified level, freq %d (dev %s)\n",
 661                             search->total_set.freq,
 662                             device_get_nameunit(search->rel_set[i].dev));
 663                         continue;
 664                 }
 665
 666                 /* Add each setting to the level, duplicating if necessary. */
 667                 for (i = 0; i < set_arr->count; i++) {
 668                         set = &set_arr->sets[i];
 669
 670                         /*
 671                          * If this setting is less than 100%, split the level
 672                          * into two and add this setting to the new level.
 673                          */
 674                         fill = search;
 675                         if (set->freq < 10000)
 676                                 fill = cpufreq_dup_set(sc, search, set);
 677
 678                         /*
 679                          * The new level was a duplicate of an existing level
 680                          * so we freed it.  Go to the next setting.
 681                          */
 682                         if (fill == NULL)
 683                                 continue;
 684
 685                         /* Add this setting to the existing or new level. */
 686                         KASSERT(fill->rel_count < MAX_SETTINGS,
 687                             ("cpufreq: too many relative drivers (%d)",
 688                             MAX_SETTINGS));
 689                         fill->rel_set[fill->rel_count] = *set;
 690                         fill->rel_count++;
 691                         CF_DEBUG(
 692                         "expand set added rel setting %d%% to %d level\n",
 693                             set->freq / 100, fill->total_set.freq);
 694                 }
 695         }
 696
 697         return (0);
 698 }
 699
 700 static struct cf_level *
 701 cpufreq_dup_set(struct cpufreq_softc *sc, struct cf_level *dup,
 702     struct cf_setting *set)
 703 {
 704         struct cf_level_lst *list;
 705         struct cf_level *fill, *itr;
 706         struct cf_setting *fill_set, *itr_set;
 707         int i;
 708
 709         CF_MTX_ASSERT(&sc->lock);
 710
 711         /*
 712          * Create a new level, copy it from the old one, and update the
 713          * total frequency and power by the percentage specified in the
 714          * relative setting.
 715          */
 716         fill = malloc(sizeof(*fill), M_TEMP, M_NOWAIT);
 717         if (fill == NULL)
 718                 return (NULL);
 719         *fill = *dup;
 720         fill_set = &fill->total_set;
 721         fill_set->freq =
 722             ((uint64_t)fill_set->freq * set->freq) / 10000;
 723         if (fill_set->power != CPUFREQ_VAL_UNKNOWN) {
 724                 fill_set->power = ((uint64_t)fill_set->power * set->freq)
 725                     / 10000;
 726         }
 727         if (set->lat != CPUFREQ_VAL_UNKNOWN) {
 728                 if (fill_set->lat != CPUFREQ_VAL_UNKNOWN)
 729                         fill_set->lat += set->lat;
 730                 else
 731                         fill_set->lat = set->lat;
 732         }
 733         CF_DEBUG("dup set considering derived setting %d\n", fill_set->freq);
 734
 735         /*
 736          * If we copied an old level that we already modified (say, at 100%),
 737          * we need to remove that setting before adding this one.  Since we
 738          * process each setting array in order, we know any settings for this
 739          * driver will be found at the end.
 740          */
 741         for (i = fill->rel_count; i != 0; i--) {
 742                 if (fill->rel_set[i - 1].dev != set->dev)
 743                         break;
 744                 CF_DEBUG("removed last relative driver: %s\n",
 745                     device_get_nameunit(set->dev));
 746                 fill->rel_count--;
 747         }
 748
 749         /*
 750          * Insert the new level in sorted order.  If we find a duplicate,
 751          * free the new level.  We can do this since any existing level will
 752          * be guaranteed to have the same or less settings and thus consume
 753          * less power.  For example, a level with one absolute setting of
 754          * 800 Mhz uses less power than one composed of an absolute setting
 755          * of 1600 Mhz and a relative setting at 50%.
 756          */
 757         list = &sc->all_levels;
 758         if (TAILQ_EMPTY(list)) {
 759                 CF_DEBUG("dup done, inserted %d at head\n", fill_set->freq);
 760                 TAILQ_INSERT_HEAD(list, fill, link);
 761         } else {
 762                 TAILQ_FOREACH_REVERSE(itr, list, cf_level_lst, link) {
 763                         itr_set = &itr->total_set;
 764                         if (CPUFREQ_CMP(fill_set->freq, itr_set->freq)) {
 765                                 CF_DEBUG(
 766                         "dup done, freeing new level %d, matches %d\n",
 767                                     fill_set->freq, itr_set->freq);
 768                                 free(fill, M_TEMP);
 769                                 fill = NULL;
 770                                 break;
 771                         } else if (fill_set->freq < itr_set->freq) {
 772                                 CF_DEBUG(
 773                         "dup done, inserting new level %d after %d\n",
 774                                     fill_set->freq, itr_set->freq);
 775                                 TAILQ_INSERT_AFTER(list, itr, fill, link);
 776                                 sc->all_count++;
 777                                 break;
 778                         }
 779                 }
 780         }
 781
 782         return (fill);
 783 }
 784
 785 static int
 786 cpufreq_curr_sysctl(SYSCTL_HANDLER_ARGS)
 787 {
 788         struct cpufreq_softc *sc;
 789         struct cf_level *levels;
 790         int count, devcount, error, freq, i, n;
 791         device_t *devs;
 792
 793         devs = NULL;
 794         sc = oidp->oid_arg1;
 795         levels = malloc(CF_MAX_LEVELS * sizeof(*levels), M_TEMP, M_NOWAIT);
 796         if (levels == NULL)
 797                 return (ENOMEM);
 798
 799         error = CPUFREQ_GET(sc->dev, &levels[0]);
 800         if (error)
 801                 goto out;
 802         freq = levels[0].total_set.freq;
 803         error = sysctl_handle_int(oidp, &freq, 0, req);
 804         if (error != 0 || req->newptr == NULL)
 805                 goto out;
 806
 807         /*
 808          * While we only call cpufreq_get() on one device (assuming all
 809          * CPUs have equal levels), we call cpufreq_set() on all CPUs.
 810          * This is needed for some MP systems.
 811          */
 812         error = devclass_get_devices(cpufreq_dc, &devs, &devcount);
 813         if (error)
 814                 goto out;
 815         for (n = 0; n < devcount; n++) {
 816                 count = CF_MAX_LEVELS;
 817                 error = CPUFREQ_LEVELS(devs[n], levels, &count);
 818                 if (error) {
 819                         if (error == E2BIG)
 820                                 printf(
 821                         "cpufreq: need to increase CF_MAX_LEVELS\n");
 822                         break;
 823                 }
 824                 for (i = 0; i < count; i++) {
 825                         if (CPUFREQ_CMP(levels[i].total_set.freq, freq)) {
 826                                 error = CPUFREQ_SET(devs[n], &levels[i],
 827                                     CPUFREQ_PRIO_USER);
 828                                 break;
 829                         }
 830                 }
 831                 if (i == count) {
 832                         error = EINVAL;
 833                         break;
 834                 }
 835         }
 836
 837 out:
 838         if (devs)
 839                 free(devs, M_TEMP);
 840         if (levels)
 841                 free(levels, M_TEMP);
 842         return (error);
 843 }
 844
 845 static int
 846 cpufreq_levels_sysctl(SYSCTL_HANDLER_ARGS)
 847 {
 848         struct cpufreq_softc *sc;
 849         struct cf_level *levels;
 850         struct cf_setting *set;
 851         struct sbuf sb;
 852         int count, error, i;
 853
 854         sc = oidp->oid_arg1;
 855         sbuf_new(&sb, NULL, 128, SBUF_AUTOEXTEND);
 856
 857         /* Get settings from the device and generate the output string. */
 858         count = CF_MAX_LEVELS;
 859         levels = malloc(count * sizeof(*levels), M_TEMP, M_NOWAIT);
 860         if (levels == NULL)
 861                 return (ENOMEM);
 862         error = CPUFREQ_LEVELS(sc->dev, levels, &count);
 863         if (error) {
 864                 if (error == E2BIG)
 865                         printf("cpufreq: need to increase CF_MAX_LEVELS\n");
 866                 goto out;
 867         }
 868         if (count) {
 869                 for (i = 0; i < count; i++) {
 870                         set = &levels[i].total_set;
 871                         sbuf_printf(&sb, "%d/%d ", set->freq, set->power);
 872                 }
 873         } else
 874                 sbuf_cpy(&sb, "0");
 875         sbuf_trim(&sb);
 876         sbuf_finish(&sb);
 877         error = sysctl_handle_string(oidp, sbuf_data(&sb), sbuf_len(&sb), req);
 878
 879 out:
 880         free(levels, M_TEMP);
 881         sbuf_delete(&sb);
 882         return (error);
 883 }
 884
 885 static int
 886 cpufreq_settings_sysctl(SYSCTL_HANDLER_ARGS)
 887 {
 888         device_t dev;
 889         struct cf_setting *sets;
 890         struct sbuf sb;
 891         int error, i, set_count;
 892
 893         dev = oidp->oid_arg1;
 894         sbuf_new(&sb, NULL, 128, SBUF_AUTOEXTEND);
 895
 896         /* Get settings from the device and generate the output string. */
 897         set_count = MAX_SETTINGS;
 898         sets = malloc(set_count * sizeof(*sets), M_TEMP, M_NOWAIT);
 899         if (sets == NULL)
 900                 return (ENOMEM);
 901         error = CPUFREQ_DRV_SETTINGS(dev, sets, &set_count);
 902         if (error)
 903                 goto out;
 904         if (set_count) {
 905                 for (i = 0; i < set_count; i++)
 906                         sbuf_printf(&sb, "%d/%d ", sets[i].freq, sets[i].power);
 907         } else
 908                 sbuf_cpy(&sb, "0");
 909         sbuf_trim(&sb);
 910         sbuf_finish(&sb);
 911         error = sysctl_handle_string(oidp, sbuf_data(&sb), sbuf_len(&sb), req);
 912
 913 out:
 914         free(sets, M_TEMP);
 915         sbuf_delete(&sb);
 916         return (error);
 917 }
 918
 919 int
 920 cpufreq_register(device_t dev)
 921 {
 922         struct cpufreq_softc *sc;
 923         device_t cf_dev, cpu_dev;
 924
 925         /* Add a sysctl to get each driver's settings separately. */
 926         SYSCTL_ADD_PROC(device_get_sysctl_ctx(dev),
 927             SYSCTL_CHILDREN(device_get_sysctl_tree(dev)),
 928             OID_AUTO, "freq_settings", CTLTYPE_STRING | CTLFLAG_RD, dev, 0,
 929             cpufreq_settings_sysctl, "A", "CPU frequency driver settings");
 930
 931         /*
 932          * Add only one cpufreq device to each CPU.  Currently, all CPUs
 933          * must offer the same levels and be switched at the same time.
 934          */
 935         cpu_dev = device_get_parent(dev);
 936         if ((cf_dev = device_find_child(cpu_dev, "cpufreq", -1))) {
 937                 sc = device_get_softc(cf_dev);
 938                 sc->max_mhz = CPUFREQ_VAL_UNKNOWN;
 939                 return (0);
 940         }
 941
 942         /* Add the child device and possibly sysctls. */
 943         cf_dev = BUS_ADD_CHILD(cpu_dev, 0, "cpufreq", -1);
 944         if (cf_dev == NULL)
 945                 return (ENOMEM);
 946         device_quiet(cf_dev);
 947
 948         return (device_probe_and_attach(cf_dev));
 949 }
 950
 951 int
 952 cpufreq_unregister(device_t dev)
 953 {
 954         device_t cf_dev, *devs;
 955         int cfcount, devcount, error, i, type;
 956
 957         /*
 958          * If this is the last cpufreq child device, remove the control
 959          * device as well.  We identify cpufreq children by calling a method
 960          * they support.
 961          */
 962         error = device_get_children(device_get_parent(dev), &devs, &devcount);
 963         if (error)
 964                 return (error);
 965         cf_dev = device_find_child(device_get_parent(dev), "cpufreq", -1);
 966         if (cf_dev == NULL) {
 967                 device_printf(dev,
 968         "warning: cpufreq_unregister called with no cpufreq device active\n");
 969                 return (0);
 970         }
 971         cfcount = 0;
 972         for (i = 0; i < devcount; i++) {
 973                 if (!device_is_attached(devs[i]))
 974                         continue;
 975                 if (CPUFREQ_DRV_TYPE(devs[i], &type) == 0)
 976                         cfcount++;
 977         }
 978         if (cfcount <= 1)
 979                 device_delete_child(device_get_parent(cf_dev), cf_dev);
 980         free(devs, M_TEMP);
 981
 982         return (0);
 983 }