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