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