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 cf_saved_freq {
  61         struct cf_level                 level;
  62         int                             priority;
  63         SLIST_ENTRY(cf_saved_freq)      link;
  64 };
  65
  66 struct cpufreq_softc {
  67         struct sx                       lock;
  68         struct cf_level                 curr_level;
  69         int                             curr_priority;
  70         SLIST_HEAD(, cf_saved_freq)     saved_freq;
  71         struct cf_level_lst             all_levels;
  72         int                             all_count;
  73         int                             max_mhz;
  74         device_t                        dev;
  75         struct sysctl_ctx_list          sysctl_ctx;
  76 };
  77
  78 struct cf_setting_array {
  79         struct cf_setting               sets[MAX_SETTINGS];
  80         int                             count;
  81         TAILQ_ENTRY(cf_setting_array)   link;
  82 };
  83
  84 TAILQ_HEAD(cf_setting_lst, cf_setting_array);
  85
  86 #define CF_MTX_INIT(x)          sx_init((x), "cpufreq lock")
  87 #define CF_MTX_LOCK(x)          sx_xlock((x))
  88 #define CF_MTX_UNLOCK(x)        sx_xunlock((x))
  89 #define CF_MTX_ASSERT(x)        sx_assert((x), SX_XLOCKED)
  90
  91 #define CF_DEBUG(msg...)        do {            \
  92         if (cf_verbose)                         \
  93                 printf("cpufreq: " msg);        \
  94         } while (0)
  95
  96 static int      cpufreq_attach(device_t dev);
  97 static int      cpufreq_detach(device_t dev);
  98 static void     cpufreq_evaluate(void *arg);
  99 static int      cf_set_method(device_t dev, const struct cf_level *level,
 100                     int priority);
 101 static int      cf_get_method(device_t dev, struct cf_level *level);
 102 static int      cf_levels_method(device_t dev, struct cf_level *levels,
 103                     int *count);
 104 static int      cpufreq_insert_abs(struct cpufreq_softc *sc,
 105                     struct cf_setting *sets, int count);
 106 static int      cpufreq_expand_set(struct cpufreq_softc *sc,
 107                     struct cf_setting_array *set_arr);
 108 static struct cf_level *cpufreq_dup_set(struct cpufreq_softc *sc,
 109                     struct cf_level *dup, struct cf_setting *set);
 110 static int      cpufreq_curr_sysctl(SYSCTL_HANDLER_ARGS);
 111 static int      cpufreq_levels_sysctl(SYSCTL_HANDLER_ARGS);
 112 static int      cpufreq_settings_sysctl(SYSCTL_HANDLER_ARGS);
 113
 114 static device_method_t cpufreq_methods[] = {
 115         DEVMETHOD(device_probe,         bus_generic_probe),
 116         DEVMETHOD(device_attach,        cpufreq_attach),
 117         DEVMETHOD(device_detach,        cpufreq_detach),
 118
 119         DEVMETHOD(cpufreq_set,          cf_set_method),
 120         DEVMETHOD(cpufreq_get,          cf_get_method),
 121         DEVMETHOD(cpufreq_levels,       cf_levels_method),
 122         {0, 0}
 123 };
 124 static driver_t cpufreq_driver = {
 125         "cpufreq", cpufreq_methods, sizeof(struct cpufreq_softc)
 126 };
 127 static devclass_t cpufreq_dc;
 128 DRIVER_MODULE(cpufreq, cpu, cpufreq_driver, cpufreq_dc, 0, 0);
 129
 130 static eventhandler_tag cf_ev_tag;
 131
 132 static int              cf_lowest_freq;
 133 static int              cf_verbose;
 134 TUNABLE_INT("debug.cpufreq.lowest", &cf_lowest_freq);
 135 TUNABLE_INT("debug.cpufreq.verbose", &cf_verbose);
 136 SYSCTL_NODE(_debug, OID_AUTO, cpufreq, CTLFLAG_RD, NULL, "cpufreq debugging");
 137 SYSCTL_INT(_debug_cpufreq, OID_AUTO, lowest, CTLFLAG_RW, &cf_lowest_freq, 1,
 138     "Don't provide levels below this frequency.");
 139 SYSCTL_INT(_debug_cpufreq, OID_AUTO, verbose, CTLFLAG_RW, &cf_verbose, 1,
 140     "Print verbose debugging messages");
 141
 142 static int
 143 cpufreq_attach(device_t dev)
 144 {
 145         struct cpufreq_softc *sc;
 146         device_t parent;
 147         int numdevs;
 148
 149         CF_DEBUG("initializing %s\n", device_get_nameunit(dev));
 150         sc = device_get_softc(dev);
 151         parent = device_get_parent(dev);
 152         sc->dev = dev;
 153         sysctl_ctx_init(&sc->sysctl_ctx);
 154         TAILQ_INIT(&sc->all_levels);
 155         CF_MTX_INIT(&sc->lock);
 156         sc->curr_level.total_set.freq = CPUFREQ_VAL_UNKNOWN;
 157         SLIST_INIT(&sc->saved_freq);
 158         sc->max_mhz = CPUFREQ_VAL_UNKNOWN;
 159
 160         /*
 161          * Only initialize one set of sysctls for all CPUs.  In the future,
 162          * if multiple CPUs can have different settings, we can move these
 163          * sysctls to be under every CPU instead of just the first one.
 164          */
 165         numdevs = devclass_get_count(cpufreq_dc);
 166         if (numdevs > 1)
 167                 return (0);
 168
 169         CF_DEBUG("initializing one-time data for %s\n",
 170             device_get_nameunit(dev));
 171         SYSCTL_ADD_PROC(&sc->sysctl_ctx,
 172             SYSCTL_CHILDREN(device_get_sysctl_tree(parent)),
 173             OID_AUTO, "freq", CTLTYPE_INT | CTLFLAG_RW, sc, 0,
 174             cpufreq_curr_sysctl, "I", "Current CPU frequency");
 175         SYSCTL_ADD_PROC(&sc->sysctl_ctx,
 176             SYSCTL_CHILDREN(device_get_sysctl_tree(parent)),
 177             OID_AUTO, "freq_levels", CTLTYPE_STRING | CTLFLAG_RD, sc, 0,
 178             cpufreq_levels_sysctl, "A", "CPU frequency levels");
 179         cf_ev_tag = EVENTHANDLER_REGISTER(cpufreq_changed, cpufreq_evaluate,
 180             NULL, EVENTHANDLER_PRI_ANY);
 181
 182         return (0);
 183 }
 184
 185 static int
 186 cpufreq_detach(device_t dev)
 187 {
 188         struct cpufreq_softc *sc;
 189         struct cf_saved_freq *saved_freq;
 190         int numdevs;
 191
 192         CF_DEBUG("shutdown %s\n", device_get_nameunit(dev));
 193         sc = device_get_softc(dev);
 194         sysctl_ctx_free(&sc->sysctl_ctx);
 195
 196         while ((saved_freq = SLIST_FIRST(&sc->saved_freq)) != NULL) {
 197                 SLIST_REMOVE_HEAD(&sc->saved_freq, link);
 198                 free(saved_freq, M_TEMP);
 199         }
 200
 201         /* Only clean up these resources when the last device is detaching. */
 202         numdevs = devclass_get_count(cpufreq_dc);
 203         if (numdevs == 1) {
 204                 CF_DEBUG("final shutdown for %s\n", device_get_nameunit(dev));
 205                 EVENTHANDLER_DEREGISTER(cpufreq_changed, cf_ev_tag);
 206         }
 207
 208         return (0);
 209 }
 210
 211 static void
 212 cpufreq_evaluate(void *arg)
 213 {
 214         /* TODO: Re-evaluate when notified of changes to drivers. */
 215 }
 216
 217 static int
 218 cf_set_method(device_t dev, const struct cf_level *level, int priority)
 219 {
 220         struct cpufreq_softc *sc;
 221         const struct cf_setting *set;
 222         struct cf_saved_freq *saved_freq, *curr_freq;
 223         struct pcpu *pc;
 224         int cpu_id, error, i;
 225         static int once;
 226
 227         sc = device_get_softc(dev);
 228         error = 0;
 229         set = NULL;
 230         saved_freq = NULL;
 231
 232         /*
 233          * Check that the TSC isn't being used as a timecounter.
 234          * If it is, then return EBUSY and refuse to change the
 235          * clock speed.
 236          */
 237         if (strcmp(timecounter->tc_name, "TSC") == 0) {
 238                 if (!once) {
 239                         printf("cpufreq: frequency change with timecounter"
 240                                 " TSC not allowed, see cpufreq(4)\n");
 241                         once = 1;
 242                 }
 243                 return (EBUSY);
 244         }
 245
 246         CF_MTX_LOCK(&sc->lock);
 247
 248         /*
 249          * If the requested level has a lower priority, don't allow
 250          * the new level right now.
 251          */
 252         if (priority < sc->curr_priority) {
 253                 CF_DEBUG("ignoring, curr prio %d less than %d\n", priority,
 254                     sc->curr_priority);
 255                 error = EPERM;
 256                 goto out;
 257         }
 258
 259         /*
 260          * If the caller didn't specify a level and one is saved, prepare to
 261          * restore the saved level.  If none has been saved, return an error.
 262          */
 263         if (level == NULL) {
 264                 saved_freq = SLIST_FIRST(&sc->saved_freq);
 265                 if (saved_freq == NULL) {
 266                         CF_DEBUG("NULL level, no saved level\n");
 267                         error = ENXIO;
 268                         goto out;
 269                 }
 270                 level = &saved_freq->level;
 271                 priority = saved_freq->priority;
 272                 CF_DEBUG("restoring saved level, freq %d prio %d\n",
 273                     level->total_set.freq, priority);
 274         }
 275
 276         /* Reject levels that are below our specified threshold. */
 277         if (level->total_set.freq < cf_lowest_freq) {
 278                 CF_DEBUG("rejecting freq %d, less than %d limit\n",
 279                     level->total_set.freq, cf_lowest_freq);
 280                 error = EINVAL;
 281                 goto out;
 282         }
 283
 284         /* If already at this level, just return. */
 285         if (CPUFREQ_CMP(sc->curr_level.total_set.freq, level->total_set.freq)) {
 286                 CF_DEBUG("skipping freq %d, same as current level %d\n",
 287                     level->total_set.freq, sc->curr_level.total_set.freq);
 288                 goto skip;
 289         }
 290
 291         /* First, set the absolute frequency via its driver. */
 292         set = &level->abs_set;
 293         if (set->dev) {
 294                 if (!device_is_attached(set->dev)) {
 295                         error = ENXIO;
 296                         goto out;
 297                 }
 298
 299                 /* Bind to the target CPU before switching, if necessary. */
 300                 cpu_id = PCPU_GET(cpuid);
 301                 pc = cpu_get_pcpu(set->dev);
 302                 if (cpu_id != pc->pc_cpuid) {
 303                         mtx_lock_spin(&sched_lock);
 304                         sched_bind(curthread, pc->pc_cpuid);
 305                         mtx_unlock_spin(&sched_lock);
 306                 }
 307                 CF_DEBUG("setting abs freq %d on %s (cpu %d)\n", set->freq,
 308                     device_get_nameunit(set->dev), PCPU_GET(cpuid));
 309                 error = CPUFREQ_DRV_SET(set->dev, set);
 310                 if (cpu_id != pc->pc_cpuid) {
 311                         mtx_lock_spin(&sched_lock);
 312                         sched_unbind(curthread);
 313                         mtx_unlock_spin(&sched_lock);
 314                 }
 315                 if (error) {
 316                         goto out;
 317                 }
 318         }
 319
 320         /* Next, set any/all relative frequencies via their drivers. */
 321         for (i = 0; i < level->rel_count; i++) {
 322                 set = &level->rel_set[i];
 323                 if (!device_is_attached(set->dev)) {
 324                         error = ENXIO;
 325                         goto out;
 326                 }
 327
 328                 /* Bind to the target CPU before switching, if necessary. */
 329                 cpu_id = PCPU_GET(cpuid);
 330                 pc = cpu_get_pcpu(set->dev);
 331                 if (cpu_id != pc->pc_cpuid) {
 332                         mtx_lock_spin(&sched_lock);
 333                         sched_bind(curthread, pc->pc_cpuid);
 334                         mtx_unlock_spin(&sched_lock);
 335                 }
 336                 CF_DEBUG("setting rel freq %d on %s (cpu %d)\n", set->freq,
 337                     device_get_nameunit(set->dev), PCPU_GET(cpuid));
 338                 error = CPUFREQ_DRV_SET(set->dev, set);
 339                 if (cpu_id != pc->pc_cpuid) {
 340                         mtx_lock_spin(&sched_lock);
 341                         sched_unbind(curthread);
 342                         mtx_unlock_spin(&sched_lock);
 343                 }
 344                 if (error) {
 345                         /* XXX Back out any successful setting? */
 346                         goto out;
 347                 }
 348         }
 349
 350 skip:
 351         /*
 352          * Before recording the current level, check if we're going to a
 353          * higher priority.  If so, save the previous level and priority.
 354          */
 355         if (sc->curr_level.total_set.freq != CPUFREQ_VAL_UNKNOWN &&
 356             priority > sc->curr_priority) {
 357                 CF_DEBUG("saving level, freq %d prio %d\n",
 358                     sc->curr_level.total_set.freq, sc->curr_priority);
 359                 curr_freq = malloc(sizeof(*curr_freq), M_TEMP, M_NOWAIT);
 360                 if (curr_freq == NULL) {
 361                         error = ENOMEM;
 362                         goto out;
 363                 }
 364                 curr_freq->level = sc->curr_level;
 365                 curr_freq->priority = sc->curr_priority;
 366                 SLIST_INSERT_HEAD(&sc->saved_freq, curr_freq, link);
 367         }
 368         sc->curr_level = *level;
 369         sc->curr_priority = priority;
 370
 371         /* If we were restoring a saved state, reset it to "unused". */
 372         if (saved_freq != NULL) {
 373                 CF_DEBUG("resetting saved level\n");
 374                 sc->curr_level.total_set.freq = CPUFREQ_VAL_UNKNOWN;
 375                 SLIST_REMOVE_HEAD(&sc->saved_freq, link);
 376                 free(saved_freq, M_TEMP);
 377         }
 378
 379 out:
 380         CF_MTX_UNLOCK(&sc->lock);
 381         if (error && set)
 382                 device_printf(set->dev, "set freq failed, err %d\n", error);
 383         return (error);
 384 }
 385
 386 static int
 387 cf_get_method(device_t dev, struct cf_level *level)
 388 {
 389         struct cpufreq_softc *sc;
 390         struct cf_level *levels;
 391         struct cf_setting *curr_set, set;
 392         struct pcpu *pc;
 393         device_t *devs;
 394         int count, error, i, numdevs;
 395         uint64_t rate;
 396
 397         sc = device_get_softc(dev);
 398         error = 0;
 399         levels = NULL;
 400
 401         /* If we already know the current frequency, we're done. */
 402         CF_MTX_LOCK(&sc->lock);
 403         curr_set = &sc->curr_level.total_set;
 404         if (curr_set->freq != CPUFREQ_VAL_UNKNOWN) {
 405                 CF_DEBUG("get returning known freq %d\n", curr_set->freq);
 406                 goto out;
 407         }
 408         CF_MTX_UNLOCK(&sc->lock);
 409
 410         /*
 411          * We need to figure out the current level.  Loop through every
 412          * driver, getting the current setting.  Then, attempt to get a best
 413          * match of settings against each level.
 414          */
 415         count = CF_MAX_LEVELS;
 416         levels = malloc(count * sizeof(*levels), M_TEMP, M_NOWAIT);
 417         if (levels == NULL)
 418                 return (ENOMEM);
 419         error = CPUFREQ_LEVELS(sc->dev, levels, &count);
 420         if (error) {
 421                 if (error == E2BIG)
 422                         printf("cpufreq: need to increase CF_MAX_LEVELS\n");
 423                 free(levels, M_TEMP);
 424                 return (error);
 425         }
 426         error = device_get_children(device_get_parent(dev), &devs, &numdevs);
 427         if (error) {
 428                 free(levels, M_TEMP);
 429                 return (error);
 430         }
 431
 432         /*
 433          * Reacquire the lock and search for the given level.
 434          *
 435          * XXX Note: this is not quite right since we really need to go
 436          * through each level and compare both absolute and relative
 437          * settings for each driver in the system before making a match.
 438          * The estimation code below catches this case though.
 439          */
 440         CF_MTX_LOCK(&sc->lock);
 441         for (i = 0; i < numdevs && curr_set->freq == CPUFREQ_VAL_UNKNOWN; i++) {
 442                 if (!device_is_attached(devs[i]))
 443                         continue;
 444                 error = CPUFREQ_DRV_GET(devs[i], &set);
 445                 if (error)
 446                         continue;
 447                 for (i = 0; i < count; i++) {
 448                         if (CPUFREQ_CMP(set.freq, levels[i].total_set.freq)) {
 449                                 sc->curr_level = levels[i];
 450                                 break;
 451                         }
 452                 }
 453         }
 454         free(devs, M_TEMP);
 455         if (curr_set->freq != CPUFREQ_VAL_UNKNOWN) {
 456                 CF_DEBUG("get matched freq %d from drivers\n", curr_set->freq);
 457                 goto out;
 458         }
 459
 460         /*
 461          * We couldn't find an exact match, so attempt to estimate and then
 462          * match against a level.
 463          */
 464         pc = cpu_get_pcpu(dev);
 465         if (pc == NULL) {
 466                 error = ENXIO;
 467                 goto out;
 468         }
 469         cpu_est_clockrate(pc->pc_cpuid, &rate);
 470         rate /= 1000000;
 471         for (i = 0; i < count; i++) {
 472                 if (CPUFREQ_CMP(rate, levels[i].total_set.freq)) {
 473                         sc->curr_level = levels[i];
 474                         CF_DEBUG("get estimated freq %d\n", curr_set->freq);
 475                         break;
 476                 }
 477         }
 478
 479 out:
 480         if (error == 0)
 481                 *level = sc->curr_level;
 482
 483         CF_MTX_UNLOCK(&sc->lock);
 484         if (levels)
 485                 free(levels, M_TEMP);
 486         return (error);
 487 }
 488
 489 static int
 490 cf_levels_method(device_t dev, struct cf_level *levels, int *count)
 491 {
 492         struct cf_setting_array *set_arr;
 493         struct cf_setting_lst rel_sets;
 494         struct cpufreq_softc *sc;
 495         struct cf_level *lev;
 496         struct cf_setting *sets;
 497         struct pcpu *pc;
 498         device_t *devs;
 499         int error, i, numdevs, set_count, type;
 500         uint64_t rate;
 501
 502         if (levels == NULL || count == NULL)
 503                 return (EINVAL);
 504
 505         TAILQ_INIT(&rel_sets);
 506         sc = device_get_softc(dev);
 507         error = device_get_children(device_get_parent(dev), &devs, &numdevs);
 508         if (error)
 509                 return (error);
 510         sets = malloc(MAX_SETTINGS * sizeof(*sets), M_TEMP, M_NOWAIT);
 511         if (sets == NULL) {
 512                 free(devs, M_TEMP);
 513                 return (ENOMEM);
 514         }
 515
 516         /* Get settings from all cpufreq drivers. */
 517         CF_MTX_LOCK(&sc->lock);
 518         for (i = 0; i < numdevs; i++) {
 519                 /* Skip devices that aren't ready. */
 520                 if (!device_is_attached(devs[i]))
 521                         continue;
 522
 523                 /*
 524                  * Get settings, skipping drivers that offer no settings or
 525                  * provide settings for informational purposes only.
 526                  */
 527                 error = CPUFREQ_DRV_TYPE(devs[i], &type);
 528                 if (error || (type & CPUFREQ_FLAG_INFO_ONLY)) {
 529                         if (error == 0) {
 530                                 CF_DEBUG("skipping info-only driver %s\n",
 531                                     device_get_nameunit(devs[i]));
 532                         }
 533                         continue;
 534                 }
 535                 set_count = MAX_SETTINGS;
 536                 error = CPUFREQ_DRV_SETTINGS(devs[i], sets, &set_count);
 537                 if (error || set_count == 0)
 538                         continue;
 539
 540                 /* Add the settings to our absolute/relative lists. */
 541                 switch (type & CPUFREQ_TYPE_MASK) {
 542                 case CPUFREQ_TYPE_ABSOLUTE:
 543                         error = cpufreq_insert_abs(sc, sets, set_count);
 544                         break;
 545                 case CPUFREQ_TYPE_RELATIVE:
 546                         CF_DEBUG("adding %d relative settings\n", set_count);
 547                         set_arr = malloc(sizeof(*set_arr), M_TEMP, M_NOWAIT);
 548                         if (set_arr == NULL) {
 549                                 error = ENOMEM;
 550                                 goto out;
 551                         }
 552                         bcopy(sets, set_arr->sets, set_count * sizeof(*sets));
 553                         set_arr->count = set_count;
 554                         TAILQ_INSERT_TAIL(&rel_sets, set_arr, link);
 555                         break;
 556                 default:
 557                         error = EINVAL;
 558                 }
 559                 if (error)
 560                         goto out;
 561         }
 562
 563         /*
 564          * If there are no absolute levels, create a fake one at 100%.  We
 565          * then cache the clockrate for later use as our base frequency.
 566          *
 567          * XXX This assumes that the first time through, if we only have
 568          * relative drivers, the CPU is currently running at 100%.
 569          */
 570         if (TAILQ_EMPTY(&sc->all_levels)) {
 571                 if (sc->max_mhz == CPUFREQ_VAL_UNKNOWN) {
 572                         pc = cpu_get_pcpu(dev);
 573                         cpu_est_clockrate(pc->pc_cpuid, &rate);
 574                         sc->max_mhz = rate / 1000000;
 575                 }
 576                 memset(&sets[0], CPUFREQ_VAL_UNKNOWN, sizeof(*sets));
 577                 sets[0].freq = sc->max_mhz;
 578                 sets[0].dev = NULL;
 579                 error = cpufreq_insert_abs(sc, sets, 1);
 580                 if (error)
 581                         goto out;
 582         }
 583
 584         /* Create a combined list of absolute + relative levels. */
 585         TAILQ_FOREACH(set_arr, &rel_sets, link)
 586                 cpufreq_expand_set(sc, set_arr);
 587
 588         /* If the caller doesn't have enough space, return the actual count. */
 589         if (sc->all_count > *count) {
 590                 *count = sc->all_count;
 591                 error = E2BIG;
 592                 goto out;
 593         }
 594
 595         /* Finally, output the list of levels. */
 596         i = 0;
 597         TAILQ_FOREACH(lev, &sc->all_levels, link) {
 598                 /* Skip levels that have a frequency that is too low. */
 599                 if (lev->total_set.freq < cf_lowest_freq) {
 600                         sc->all_count--;
 601                         continue;
 602                 }
 603
 604                 levels[i] = *lev;
 605                 i++;
 606         }
 607         *count = sc->all_count;
 608         error = 0;
 609
 610 out:
 611         /* Clear all levels since we regenerate them each time. */
 612         while ((lev = TAILQ_FIRST(&sc->all_levels)) != NULL) {
 613                 TAILQ_REMOVE(&sc->all_levels, lev, link);
 614                 free(lev, M_TEMP);
 615         }
 616         sc->all_count = 0;
 617
 618         CF_MTX_UNLOCK(&sc->lock);
 619         while ((set_arr = TAILQ_FIRST(&rel_sets)) != NULL) {
 620                 TAILQ_REMOVE(&rel_sets, set_arr, link);
 621                 free(set_arr, M_TEMP);
 622         }
 623         free(devs, M_TEMP);
 624         free(sets, M_TEMP);
 625         return (error);
 626 }
 627
 628 /*
 629  * Create levels for an array of absolute settings and insert them in
 630  * sorted order in the specified list.
 631  */
 632 static int
 633 cpufreq_insert_abs(struct cpufreq_softc *sc, struct cf_setting *sets,
 634     int count)
 635 {
 636         struct cf_level_lst *list;
 637         struct cf_level *level, *search;
 638         int i;
 639
 640         CF_MTX_ASSERT(&sc->lock);
 641
 642         list = &sc->all_levels;
 643         for (i = 0; i < count; i++) {
 644                 level = malloc(sizeof(*level), M_TEMP, M_NOWAIT | M_ZERO);
 645                 if (level == NULL)
 646                         return (ENOMEM);
 647                 level->abs_set = sets[i];
 648                 level->total_set = sets[i];
 649                 level->total_set.dev = NULL;
 650                 sc->all_count++;
 651
 652                 if (TAILQ_EMPTY(list)) {
 653                         CF_DEBUG("adding abs setting %d at head\n",
 654                             sets[i].freq);
 655                         TAILQ_INSERT_HEAD(list, level, link);
 656                         continue;
 657                 }
 658
 659                 TAILQ_FOREACH_REVERSE(search, list, cf_level_lst, link) {
 660                         if (sets[i].freq <= search->total_set.freq) {
 661                                 CF_DEBUG("adding abs setting %d after %d\n",
 662                                     sets[i].freq, search->total_set.freq);
 663                                 TAILQ_INSERT_AFTER(list, search, level, link);
 664                                 break;
 665                         }
 666                 }
 667         }
 668         return (0);
 669 }
 670
 671 /*
 672  * Expand a group of relative settings, creating derived levels from them.
 673  */
 674 static int
 675 cpufreq_expand_set(struct cpufreq_softc *sc, struct cf_setting_array *set_arr)
 676 {
 677         struct cf_level *fill, *search;
 678         struct cf_setting *set;
 679         int i;
 680
 681         CF_MTX_ASSERT(&sc->lock);
 682
 683         /*
 684          * Walk the set of all existing levels in reverse.  This is so we
 685          * create derived states from the lowest absolute settings first
 686          * and discard duplicates created from higher absolute settings.
 687          * For instance, a level of 50 Mhz derived from 100 Mhz + 50% is
 688          * preferable to 200 Mhz + 25% because absolute settings are more
 689          * efficient since they often change the voltage as well.
 690          */
 691         TAILQ_FOREACH_REVERSE(search, &sc->all_levels, cf_level_lst, link) {
 692                 /* Add each setting to the level, duplicating if necessary. */
 693                 for (i = 0; i < set_arr->count; i++) {
 694                         set = &set_arr->sets[i];
 695
 696                         /*
 697                          * If this setting is less than 100%, split the level
 698                          * into two and add this setting to the new level.
 699                          */
 700                         fill = search;
 701                         if (set->freq < 10000) {
 702                                 fill = cpufreq_dup_set(sc, search, set);
 703
 704                                 /*
 705                                  * The new level was a duplicate of an existing
 706                                  * level or its absolute setting is too high
 707                                  * so we freed it.  For example, we discard a
 708                                  * derived level of 1000 MHz/25% if a level
 709                                  * of 500 MHz/100% already exists.
 710                                  */
 711                                 if (fill == NULL)
 712                                         break;
 713                         }
 714
 715                         /* Add this setting to the existing or new level. */
 716                         KASSERT(fill->rel_count < MAX_SETTINGS,
 717                             ("cpufreq: too many relative drivers (%d)",
 718                             MAX_SETTINGS));
 719                         fill->rel_set[fill->rel_count] = *set;
 720                         fill->rel_count++;
 721                         CF_DEBUG(
 722                         "expand set added rel setting %d%% to %d level\n",
 723                             set->freq / 100, fill->total_set.freq);
 724                 }
 725         }
 726
 727         return (0);
 728 }
 729
 730 static struct cf_level *
 731 cpufreq_dup_set(struct cpufreq_softc *sc, struct cf_level *dup,
 732     struct cf_setting *set)
 733 {
 734         struct cf_level_lst *list;
 735         struct cf_level *fill, *itr;
 736         struct cf_setting *fill_set, *itr_set;
 737         int i;
 738
 739         CF_MTX_ASSERT(&sc->lock);
 740
 741         /*
 742          * Create a new level, copy it from the old one, and update the
 743          * total frequency and power by the percentage specified in the
 744          * relative setting.
 745          */
 746         fill = malloc(sizeof(*fill), M_TEMP, M_NOWAIT);
 747         if (fill == NULL)
 748                 return (NULL);
 749         *fill = *dup;
 750         fill_set = &fill->total_set;
 751         fill_set->freq =
 752             ((uint64_t)fill_set->freq * set->freq) / 10000;
 753         if (fill_set->power != CPUFREQ_VAL_UNKNOWN) {
 754                 fill_set->power = ((uint64_t)fill_set->power * set->freq)
 755                     / 10000;
 756         }
 757         if (set->lat != CPUFREQ_VAL_UNKNOWN) {
 758                 if (fill_set->lat != CPUFREQ_VAL_UNKNOWN)
 759                         fill_set->lat += set->lat;
 760                 else
 761                         fill_set->lat = set->lat;
 762         }
 763         CF_DEBUG("dup set considering derived setting %d\n", fill_set->freq);
 764
 765         /*
 766          * If we copied an old level that we already modified (say, at 100%),
 767          * we need to remove that setting before adding this one.  Since we
 768          * process each setting array in order, we know any settings for this
 769          * driver will be found at the end.
 770          */
 771         for (i = fill->rel_count; i != 0; i--) {
 772                 if (fill->rel_set[i - 1].dev != set->dev)
 773                         break;
 774                 CF_DEBUG("removed last relative driver: %s\n",
 775                     device_get_nameunit(set->dev));
 776                 fill->rel_count--;
 777         }
 778
 779         /*
 780          * Insert the new level in sorted order.  If it is a duplicate of an
 781          * existing level (1) or has an absolute setting higher than the
 782          * existing level (2), do not add it.  We can do this since any such
 783          * level is guaranteed use less power.  For example (1), a level with
 784          * one absolute setting of 800 Mhz uses less power than one composed
 785          * of an absolute setting of 1600 Mhz and a relative setting at 50%.
 786          * Also for example (2), a level of 800 Mhz/75% is preferable to
 787          * 1600 Mhz/25% even though the latter has a lower total frequency.
 788          */
 789         list = &sc->all_levels;
 790         KASSERT(!TAILQ_EMPTY(list), ("all levels list empty in dup set"));
 791         TAILQ_FOREACH_REVERSE(itr, list, cf_level_lst, link) {
 792                 itr_set = &itr->total_set;
 793                 if (CPUFREQ_CMP(fill_set->freq, itr_set->freq)) {
 794                         CF_DEBUG("dup set rejecting %d (dupe)\n",
 795                             fill_set->freq);
 796                         itr = NULL;
 797                         break;
 798                 } else if (fill_set->freq < itr_set->freq) {
 799                         if (fill->abs_set.freq <= itr->abs_set.freq) {
 800                                 CF_DEBUG(
 801                         "dup done, inserting new level %d after %d\n",
 802                                     fill_set->freq, itr_set->freq);
 803                                 TAILQ_INSERT_AFTER(list, itr, fill, link);
 804                                 sc->all_count++;
 805                         } else {
 806                                 CF_DEBUG("dup set rejecting %d (abs too big)\n",
 807                                     fill_set->freq);
 808                                 itr = NULL;
 809                         }
 810                         break;
 811                 }
 812         }
 813
 814         /* We didn't find a good place for this new level so free it. */
 815         if (itr == NULL) {
 816                 CF_DEBUG("dup set freeing new level %d (not optimal)\n",
 817                     fill_set->freq);
 818                 free(fill, M_TEMP);
 819                 fill = NULL;
 820         }
 821
 822         return (fill);
 823 }
 824
 825 static int
 826 cpufreq_curr_sysctl(SYSCTL_HANDLER_ARGS)
 827 {
 828         struct cpufreq_softc *sc;
 829         struct cf_level *levels;
 830         int count, devcount, error, freq, i, n;
 831         device_t *devs;
 832
 833         devs = NULL;
 834         sc = oidp->oid_arg1;
 835         levels = malloc(CF_MAX_LEVELS * sizeof(*levels), M_TEMP, M_NOWAIT);
 836         if (levels == NULL)
 837                 return (ENOMEM);
 838
 839         error = CPUFREQ_GET(sc->dev, &levels[0]);
 840         if (error)
 841                 goto out;
 842         freq = levels[0].total_set.freq;
 843         error = sysctl_handle_int(oidp, &freq, 0, req);
 844         if (error != 0 || req->newptr == NULL)
 845                 goto out;
 846
 847         /*
 848          * While we only call cpufreq_get() on one device (assuming all
 849          * CPUs have equal levels), we call cpufreq_set() on all CPUs.
 850          * This is needed for some MP systems.
 851          */
 852         error = devclass_get_devices(cpufreq_dc, &devs, &devcount);
 853         if (error)
 854                 goto out;
 855         for (n = 0; n < devcount; n++) {
 856                 count = CF_MAX_LEVELS;
 857                 error = CPUFREQ_LEVELS(devs[n], levels, &count);
 858                 if (error) {
 859                         if (error == E2BIG)
 860                                 printf(
 861                         "cpufreq: need to increase CF_MAX_LEVELS\n");
 862                         break;
 863                 }
 864                 for (i = 0; i < count; i++) {
 865                         if (CPUFREQ_CMP(levels[i].total_set.freq, freq)) {
 866                                 error = CPUFREQ_SET(devs[n], &levels[i],
 867                                     CPUFREQ_PRIO_USER);
 868                                 break;
 869                         }
 870                 }
 871                 if (i == count) {
 872                         error = EINVAL;
 873                         break;
 874                 }
 875         }
 876
 877 out:
 878         if (devs)
 879                 free(devs, M_TEMP);
 880         if (levels)
 881                 free(levels, M_TEMP);
 882         return (error);
 883 }
 884
 885 static int
 886 cpufreq_levels_sysctl(SYSCTL_HANDLER_ARGS)
 887 {
 888         struct cpufreq_softc *sc;
 889         struct cf_level *levels;
 890         struct cf_setting *set;
 891         struct sbuf sb;
 892         int count, error, i;
 893
 894         sc = oidp->oid_arg1;
 895         sbuf_new(&sb, NULL, 128, SBUF_AUTOEXTEND);
 896
 897         /* Get settings from the device and generate the output string. */
 898         count = CF_MAX_LEVELS;
 899         levels = malloc(count * sizeof(*levels), M_TEMP, M_NOWAIT);
 900         if (levels == NULL)
 901                 return (ENOMEM);
 902         error = CPUFREQ_LEVELS(sc->dev, levels, &count);
 903         if (error) {
 904                 if (error == E2BIG)
 905                         printf("cpufreq: need to increase CF_MAX_LEVELS\n");
 906                 goto out;
 907         }
 908         if (count) {
 909                 for (i = 0; i < count; i++) {
 910                         set = &levels[i].total_set;
 911                         sbuf_printf(&sb, "%d/%d ", set->freq, set->power);
 912                 }
 913         } else
 914                 sbuf_cpy(&sb, "0");
 915         sbuf_trim(&sb);
 916         sbuf_finish(&sb);
 917         error = sysctl_handle_string(oidp, sbuf_data(&sb), sbuf_len(&sb), req);
 918
 919 out:
 920         free(levels, M_TEMP);
 921         sbuf_delete(&sb);
 922         return (error);
 923 }
 924
 925 static int
 926 cpufreq_settings_sysctl(SYSCTL_HANDLER_ARGS)
 927 {
 928         device_t dev;
 929         struct cf_setting *sets;
 930         struct sbuf sb;
 931         int error, i, set_count;
 932
 933         dev = oidp->oid_arg1;
 934         sbuf_new(&sb, NULL, 128, SBUF_AUTOEXTEND);
 935
 936         /* Get settings from the device and generate the output string. */
 937         set_count = MAX_SETTINGS;
 938         sets = malloc(set_count * sizeof(*sets), M_TEMP, M_NOWAIT);
 939         if (sets == NULL)
 940                 return (ENOMEM);
 941         error = CPUFREQ_DRV_SETTINGS(dev, sets, &set_count);
 942         if (error)
 943                 goto out;
 944         if (set_count) {
 945                 for (i = 0; i < set_count; i++)
 946                         sbuf_printf(&sb, "%d/%d ", sets[i].freq, sets[i].power);
 947         } else
 948                 sbuf_cpy(&sb, "0");
 949         sbuf_trim(&sb);
 950         sbuf_finish(&sb);
 951         error = sysctl_handle_string(oidp, sbuf_data(&sb), sbuf_len(&sb), req);
 952
 953 out:
 954         free(sets, M_TEMP);
 955         sbuf_delete(&sb);
 956         return (error);
 957 }
 958
 959 int
 960 cpufreq_register(device_t dev)
 961 {
 962         struct cpufreq_softc *sc;
 963         device_t cf_dev, cpu_dev;
 964
 965         /* Add a sysctl to get each driver's settings separately. */
 966         SYSCTL_ADD_PROC(device_get_sysctl_ctx(dev),
 967             SYSCTL_CHILDREN(device_get_sysctl_tree(dev)),
 968             OID_AUTO, "freq_settings", CTLTYPE_STRING | CTLFLAG_RD, dev, 0,
 969             cpufreq_settings_sysctl, "A", "CPU frequency driver settings");
 970
 971         /*
 972          * Add only one cpufreq device to each CPU.  Currently, all CPUs
 973          * must offer the same levels and be switched at the same time.
 974          */
 975         cpu_dev = device_get_parent(dev);
 976         if ((cf_dev = device_find_child(cpu_dev, "cpufreq", -1))) {
 977                 sc = device_get_softc(cf_dev);
 978                 sc->max_mhz = CPUFREQ_VAL_UNKNOWN;
 979                 return (0);
 980         }
 981
 982         /* Add the child device and possibly sysctls. */
 983         cf_dev = BUS_ADD_CHILD(cpu_dev, 0, "cpufreq", -1);
 984         if (cf_dev == NULL)
 985                 return (ENOMEM);
 986         device_quiet(cf_dev);
 987
 988         return (device_probe_and_attach(cf_dev));
 989 }
 990
 991 int
 992 cpufreq_unregister(device_t dev)
 993 {
 994         device_t cf_dev, *devs;
 995         int cfcount, devcount, error, i, type;
 996
 997         /*
 998          * If this is the last cpufreq child device, remove the control
 999          * device as well.  We identify cpufreq children by calling a method
1000          * they support.
1001          */
1002         error = device_get_children(device_get_parent(dev), &devs, &devcount);
1003         if (error)
1004                 return (error);
1005         cf_dev = device_find_child(device_get_parent(dev), "cpufreq", -1);
1006         if (cf_dev == NULL) {
1007                 device_printf(dev,
1008         "warning: cpufreq_unregister called with no cpufreq device active\n");
1009                 return (0);
1010         }
1011         cfcount = 0;
1012         for (i = 0; i < devcount; i++) {
1013                 if (!device_is_attached(devs[i]))
1014                         continue;
1015                 if (CPUFREQ_DRV_TYPE(devs[i], &type) == 0)
1016                         cfcount++;
1017         }
1018         if (cfcount <= 1)
1019                 device_delete_child(device_get_parent(cf_dev), cf_dev);
1020         free(devs, M_TEMP);
1021
1022         return (0);
1023 }