sys/kern/kern_cpu.c

   1 /*-
   2  * Copyright (c) 2004-2007 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/sbuf.h>
  41 #include <sys/sched.h>
  42 #include <sys/smp.h>
  43 #include <sys/sysctl.h>
  44 #include <sys/systm.h>
  45 #include <sys/sx.h>
  46 #include <sys/timetc.h>
  47 #include <sys/taskqueue.h>
  48
  49 #include "cpufreq_if.h"
  50
  51 /*
  52  * Common CPU frequency glue code.  Drivers for specific hardware can
  53  * attach this interface to allow users to get/set the CPU frequency.
  54  */
  55
  56 /*
  57  * Number of levels we can handle.  Levels are synthesized from settings
  58  * so for M settings and N drivers, there may be M*N levels.
  59  */
  60 #define CF_MAX_LEVELS   64
  61
  62 struct cf_saved_freq {
  63         struct cf_level                 level;
  64         int                             priority;
  65         SLIST_ENTRY(cf_saved_freq)      link;
  66 };
  67
  68 struct cpufreq_softc {
  69         struct sx                       lock;
  70         struct cf_level                 curr_level;
  71         int                             curr_priority;
  72         SLIST_HEAD(, cf_saved_freq)     saved_freq;
  73         struct cf_level_lst             all_levels;
  74         int                             all_count;
  75         int                             max_mhz;
  76         device_t                        dev;
  77         struct sysctl_ctx_list          sysctl_ctx;
  78         struct task                     startup_task;
  79         struct cf_level                 *levels_buf;
  80 };
  81
  82 struct cf_setting_array {
  83         struct cf_setting               sets[MAX_SETTINGS];
  84         int                             count;
  85         TAILQ_ENTRY(cf_setting_array)   link;
  86 };
  87
  88 TAILQ_HEAD(cf_setting_lst, cf_setting_array);
  89
  90 #define CF_MTX_INIT(x)          sx_init((x), "cpufreq lock")
  91 #define CF_MTX_LOCK(x)          sx_xlock((x))
  92 #define CF_MTX_UNLOCK(x)        sx_xunlock((x))
  93 #define CF_MTX_ASSERT(x)        sx_assert((x), SX_XLOCKED)
  94
  95 #define CF_DEBUG(msg...)        do {            \
  96         if (cf_verbose)                         \
  97                 printf("cpufreq: " msg);        \
  98         } while (0)
  99
 100 static int      cpufreq_attach(device_t dev);
 101 static void     cpufreq_startup_task(void *ctx, int pending);
 102 static int      cpufreq_detach(device_t dev);
 103 static int      cf_set_method(device_t dev, const struct cf_level *level,
 104                     int priority);
 105 static int      cf_get_method(device_t dev, struct cf_level *level);
 106 static int      cf_levels_method(device_t dev, struct cf_level *levels,
 107                     int *count);
 108 static int      cpufreq_insert_abs(struct cpufreq_softc *sc,
 109                     struct cf_setting *sets, int count);
 110 static int      cpufreq_expand_set(struct cpufreq_softc *sc,
 111                     struct cf_setting_array *set_arr);
 112 static struct cf_level *cpufreq_dup_set(struct cpufreq_softc *sc,
 113                     struct cf_level *dup, struct cf_setting *set);
 114 static int      cpufreq_curr_sysctl(SYSCTL_HANDLER_ARGS);
 115 static int      cpufreq_levels_sysctl(SYSCTL_HANDLER_ARGS);
 116 static int      cpufreq_settings_sysctl(SYSCTL_HANDLER_ARGS);
 117
 118 static device_method_t cpufreq_methods[] = {
 119         DEVMETHOD(device_probe,         bus_generic_probe),
 120         DEVMETHOD(device_attach,        cpufreq_attach),
 121         DEVMETHOD(device_detach,        cpufreq_detach),
 122
 123         DEVMETHOD(cpufreq_set,          cf_set_method),
 124         DEVMETHOD(cpufreq_get,          cf_get_method),
 125         DEVMETHOD(cpufreq_levels,       cf_levels_method),
 126         {0, 0}
 127 };
 128 static driver_t cpufreq_driver = {
 129         "cpufreq", cpufreq_methods, sizeof(struct cpufreq_softc)
 130 };
 131 static devclass_t cpufreq_dc;
 132 DRIVER_MODULE(cpufreq, cpu, cpufreq_driver, cpufreq_dc, 0, 0);
 133
 134 static int              cf_lowest_freq;
 135 static int              cf_verbose;
 136 static SYSCTL_NODE(_debug, OID_AUTO, cpufreq, CTLFLAG_RD, NULL,
 137     "cpufreq debugging");
 138 SYSCTL_INT(_debug_cpufreq, OID_AUTO, lowest, CTLFLAG_RWTUN, &cf_lowest_freq, 1,
 139     "Don't provide levels below this frequency.");
 140 SYSCTL_INT(_debug_cpufreq, OID_AUTO, verbose, CTLFLAG_RWTUN, &cf_verbose, 1,
 141     "Print verbose debugging messages");
 142
 143 static int
 144 cpufreq_attach(device_t dev)
 145 {
 146         struct cpufreq_softc *sc;
 147         struct pcpu *pc;
 148         device_t parent;
 149         uint64_t rate;
 150         int numdevs;
 151
 152         CF_DEBUG("initializing %s\n", device_get_nameunit(dev));
 153         sc = device_get_softc(dev);
 154         parent = device_get_parent(dev);
 155         sc->dev = dev;
 156         sysctl_ctx_init(&sc->sysctl_ctx);
 157         TAILQ_INIT(&sc->all_levels);
 158         CF_MTX_INIT(&sc->lock);
 159         sc->curr_level.total_set.freq = CPUFREQ_VAL_UNKNOWN;
 160         SLIST_INIT(&sc->saved_freq);
 161         /* Try to get nominal CPU freq to use it as maximum later if needed */
 162         sc->max_mhz = cpu_get_nominal_mhz(dev);
 163         /* If that fails, try to measure the current rate */
 164         if (sc->max_mhz <= 0) {
 165                 pc = cpu_get_pcpu(dev);
 166                 if (cpu_est_clockrate(pc->pc_cpuid, &rate) == 0)
 167                         sc->max_mhz = rate / 1000000;
 168                 else
 169                         sc->max_mhz = CPUFREQ_VAL_UNKNOWN;
 170         }
 171
 172         /*
 173          * Only initialize one set of sysctls for all CPUs.  In the future,
 174          * if multiple CPUs can have different settings, we can move these
 175          * sysctls to be under every CPU instead of just the first one.
 176          */
 177         numdevs = devclass_get_count(cpufreq_dc);
 178         if (numdevs > 1)
 179                 return (0);
 180
 181         CF_DEBUG("initializing one-time data for %s\n",
 182             device_get_nameunit(dev));
 183         sc->levels_buf = malloc(CF_MAX_LEVELS * sizeof(*sc->levels_buf),
 184             M_DEVBUF, M_WAITOK);
 185         SYSCTL_ADD_PROC(&sc->sysctl_ctx,
 186             SYSCTL_CHILDREN(device_get_sysctl_tree(parent)),
 187             OID_AUTO, "freq", CTLTYPE_INT | CTLFLAG_RW, sc, 0,
 188             cpufreq_curr_sysctl, "I", "Current CPU frequency");
 189         SYSCTL_ADD_PROC(&sc->sysctl_ctx,
 190             SYSCTL_CHILDREN(device_get_sysctl_tree(parent)),
 191             OID_AUTO, "freq_levels", CTLTYPE_STRING | CTLFLAG_RD, sc, 0,
 192             cpufreq_levels_sysctl, "A", "CPU frequency levels");
 193
 194         /*
 195          * Queue a one-shot broadcast that levels have changed.
 196          * It will run once the system has completed booting.
 197          */
 198         TASK_INIT(&sc->startup_task, 0, cpufreq_startup_task, dev);
 199         taskqueue_enqueue(taskqueue_thread, &sc->startup_task);
 200
 201         return (0);
 202 }
 203
 204 /* Handle any work to be done for all drivers that attached during boot. */
 205 static void
 206 cpufreq_startup_task(void *ctx, int pending)
 207 {
 208
 209         cpufreq_settings_changed((device_t)ctx);
 210 }
 211
 212 static int
 213 cpufreq_detach(device_t dev)
 214 {
 215         struct cpufreq_softc *sc;
 216         struct cf_saved_freq *saved_freq;
 217         int numdevs;
 218
 219         CF_DEBUG("shutdown %s\n", device_get_nameunit(dev));
 220         sc = device_get_softc(dev);
 221         sysctl_ctx_free(&sc->sysctl_ctx);
 222
 223         while ((saved_freq = SLIST_FIRST(&sc->saved_freq)) != NULL) {
 224                 SLIST_REMOVE_HEAD(&sc->saved_freq, link);
 225                 free(saved_freq, M_TEMP);
 226         }
 227
 228         /* Only clean up these resources when the last device is detaching. */
 229         numdevs = devclass_get_count(cpufreq_dc);
 230         if (numdevs == 1) {
 231                 CF_DEBUG("final shutdown for %s\n", device_get_nameunit(dev));
 232                 free(sc->levels_buf, M_DEVBUF);
 233         }
 234
 235         return (0);
 236 }
 237
 238 static int
 239 cf_set_method(device_t dev, const struct cf_level *level, int priority)
 240 {
 241         struct cpufreq_softc *sc;
 242         const struct cf_setting *set;
 243         struct cf_saved_freq *saved_freq, *curr_freq;
 244         struct pcpu *pc;
 245         int error, i;
 246
 247         sc = device_get_softc(dev);
 248         error = 0;
 249         set = NULL;
 250         saved_freq = NULL;
 251
 252         /* We are going to change levels so notify the pre-change handler. */
 253         EVENTHANDLER_INVOKE(cpufreq_pre_change, level, &error);
 254         if (error != 0) {
 255                 EVENTHANDLER_INVOKE(cpufreq_post_change, level, error);
 256                 return (error);
 257         }
 258
 259         CF_MTX_LOCK(&sc->lock);
 260
 261 #ifdef SMP
 262         /*
 263          * If still booting and secondary CPUs not started yet, don't allow
 264          * changing the frequency until they're online.  This is because we
 265          * can't switch to them using sched_bind() and thus we'd only be
 266          * switching the main CPU.  XXXTODO: Need to think more about how to
 267          * handle having different CPUs at different frequencies.
 268          */
 269         if (mp_ncpus > 1 && !smp_started) {
 270                 device_printf(dev, "rejecting change, SMP not started yet\n");
 271                 error = ENXIO;
 272                 goto out;
 273         }
 274 #endif /* SMP */
 275
 276         /*
 277          * If the requested level has a lower priority, don't allow
 278          * the new level right now.
 279          */
 280         if (priority < sc->curr_priority) {
 281                 CF_DEBUG("ignoring, curr prio %d less than %d\n", priority,
 282                     sc->curr_priority);
 283                 error = EPERM;
 284                 goto out;
 285         }
 286
 287         /*
 288          * If the caller didn't specify a level and one is saved, prepare to
 289          * restore the saved level.  If none has been saved, return an error.
 290          */
 291         if (level == NULL) {
 292                 saved_freq = SLIST_FIRST(&sc->saved_freq);
 293                 if (saved_freq == NULL) {
 294                         CF_DEBUG("NULL level, no saved level\n");
 295                         error = ENXIO;
 296                         goto out;
 297                 }
 298                 level = &saved_freq->level;
 299                 priority = saved_freq->priority;
 300                 CF_DEBUG("restoring saved level, freq %d prio %d\n",
 301                     level->total_set.freq, priority);
 302         }
 303
 304         /* Reject levels that are below our specified threshold. */
 305         if (level->total_set.freq < cf_lowest_freq) {
 306                 CF_DEBUG("rejecting freq %d, less than %d limit\n",
 307                     level->total_set.freq, cf_lowest_freq);
 308                 error = EINVAL;
 309                 goto out;
 310         }
 311
 312         /* If already at this level, just return. */
 313         if (sc->curr_level.total_set.freq == level->total_set.freq) {
 314                 CF_DEBUG("skipping freq %d, same as current level %d\n",
 315                     level->total_set.freq, sc->curr_level.total_set.freq);
 316                 goto skip;
 317         }
 318
 319         /* First, set the absolute frequency via its driver. */
 320         set = &level->abs_set;
 321         if (set->dev) {
 322                 if (!device_is_attached(set->dev)) {
 323                         error = ENXIO;
 324                         goto out;
 325                 }
 326
 327                 /* Bind to the target CPU before switching. */
 328                 pc = cpu_get_pcpu(set->dev);
 329                 thread_lock(curthread);
 330                 sched_bind(curthread, pc->pc_cpuid);
 331                 thread_unlock(curthread);
 332                 CF_DEBUG("setting abs freq %d on %s (cpu %d)\n", set->freq,
 333                     device_get_nameunit(set->dev), PCPU_GET(cpuid));
 334                 error = CPUFREQ_DRV_SET(set->dev, set);
 335                 thread_lock(curthread);
 336                 sched_unbind(curthread);
 337                 thread_unlock(curthread);
 338                 if (error) {
 339                         goto out;
 340                 }
 341         }
 342
 343         /* Next, set any/all relative frequencies via their drivers. */
 344         for (i = 0; i < level->rel_count; i++) {
 345                 set = &level->rel_set[i];
 346                 if (!device_is_attached(set->dev)) {
 347                         error = ENXIO;
 348                         goto out;
 349                 }
 350
 351                 /* Bind to the target CPU before switching. */
 352                 pc = cpu_get_pcpu(set->dev);
 353                 thread_lock(curthread);
 354                 sched_bind(curthread, pc->pc_cpuid);
 355                 thread_unlock(curthread);
 356                 CF_DEBUG("setting rel freq %d on %s (cpu %d)\n", set->freq,
 357                     device_get_nameunit(set->dev), PCPU_GET(cpuid));
 358                 error = CPUFREQ_DRV_SET(set->dev, set);
 359                 thread_lock(curthread);
 360                 sched_unbind(curthread);
 361                 thread_unlock(curthread);
 362                 if (error) {
 363                         /* XXX Back out any successful setting? */
 364                         goto out;
 365                 }
 366         }
 367
 368 skip:
 369         /*
 370          * Before recording the current level, check if we're going to a
 371          * higher priority.  If so, save the previous level and priority.
 372          */
 373         if (sc->curr_level.total_set.freq != CPUFREQ_VAL_UNKNOWN &&
 374             priority > sc->curr_priority) {
 375                 CF_DEBUG("saving level, freq %d prio %d\n",
 376                     sc->curr_level.total_set.freq, sc->curr_priority);
 377                 curr_freq = malloc(sizeof(*curr_freq), M_TEMP, M_NOWAIT);
 378                 if (curr_freq == NULL) {
 379                         error = ENOMEM;
 380                         goto out;
 381                 }
 382                 curr_freq->level = sc->curr_level;
 383                 curr_freq->priority = sc->curr_priority;
 384                 SLIST_INSERT_HEAD(&sc->saved_freq, curr_freq, link);
 385         }
 386         sc->curr_level = *level;
 387         sc->curr_priority = priority;
 388
 389         /* If we were restoring a saved state, reset it to "unused". */
 390         if (saved_freq != NULL) {
 391                 CF_DEBUG("resetting saved level\n");
 392                 sc->curr_level.total_set.freq = CPUFREQ_VAL_UNKNOWN;
 393                 SLIST_REMOVE_HEAD(&sc->saved_freq, link);
 394                 free(saved_freq, M_TEMP);
 395         }
 396
 397 out:
 398         CF_MTX_UNLOCK(&sc->lock);
 399
 400         /*
 401          * We changed levels (or attempted to) so notify the post-change
 402          * handler of new frequency or error.
 403          */
 404         EVENTHANDLER_INVOKE(cpufreq_post_change, level, error);
 405         if (error && set)
 406                 device_printf(set->dev, "set freq failed, err %d\n", error);
 407
 408         return (error);
 409 }
 410
 411 static int
 412 cf_get_method(device_t dev, struct cf_level *level)
 413 {
 414         struct cpufreq_softc *sc;
 415         struct cf_level *levels;
 416         struct cf_setting *curr_set, set;
 417         struct pcpu *pc;
 418         device_t *devs;
 419         int bdiff, count, diff, error, i, n, numdevs;
 420         uint64_t rate;
 421
 422         sc = device_get_softc(dev);
 423         error = 0;
 424         levels = NULL;
 425
 426         /* If we already know the current frequency, we're done. */
 427         CF_MTX_LOCK(&sc->lock);
 428         curr_set = &sc->curr_level.total_set;
 429         if (curr_set->freq != CPUFREQ_VAL_UNKNOWN) {
 430                 CF_DEBUG("get returning known freq %d\n", curr_set->freq);
 431                 goto out;
 432         }
 433         CF_MTX_UNLOCK(&sc->lock);
 434
 435         /*
 436          * We need to figure out the current level.  Loop through every
 437          * driver, getting the current setting.  Then, attempt to get a best
 438          * match of settings against each level.
 439          */
 440         count = CF_MAX_LEVELS;
 441         levels = malloc(count * sizeof(*levels), M_TEMP, M_NOWAIT);
 442         if (levels == NULL)
 443                 return (ENOMEM);
 444         error = CPUFREQ_LEVELS(sc->dev, levels, &count);
 445         if (error) {
 446                 if (error == E2BIG)
 447                         printf("cpufreq: need to increase CF_MAX_LEVELS\n");
 448                 free(levels, M_TEMP);
 449                 return (error);
 450         }
 451         error = device_get_children(device_get_parent(dev), &devs, &numdevs);
 452         if (error) {
 453                 free(levels, M_TEMP);
 454                 return (error);
 455         }
 456
 457         /*
 458          * Reacquire the lock and search for the given level.
 459          *
 460          * XXX Note: this is not quite right since we really need to go
 461          * through each level and compare both absolute and relative
 462          * settings for each driver in the system before making a match.
 463          * The estimation code below catches this case though.
 464          */
 465         CF_MTX_LOCK(&sc->lock);
 466         for (n = 0; n < numdevs && curr_set->freq == CPUFREQ_VAL_UNKNOWN; n++) {
 467                 if (!device_is_attached(devs[n]))
 468                         continue;
 469                 if (CPUFREQ_DRV_GET(devs[n], &set) != 0)
 470                         continue;
 471                 for (i = 0; i < count; i++) {
 472                         if (set.freq == levels[i].total_set.freq) {
 473                                 sc->curr_level = levels[i];
 474                                 break;
 475                         }
 476                 }
 477         }
 478         free(devs, M_TEMP);
 479         if (curr_set->freq != CPUFREQ_VAL_UNKNOWN) {
 480                 CF_DEBUG("get matched freq %d from drivers\n", curr_set->freq);
 481                 goto out;
 482         }
 483
 484         /*
 485          * We couldn't find an exact match, so attempt to estimate and then
 486          * match against a level.
 487          */
 488         pc = cpu_get_pcpu(dev);
 489         if (pc == NULL) {
 490                 error = ENXIO;
 491                 goto out;
 492         }
 493         cpu_est_clockrate(pc->pc_cpuid, &rate);
 494         rate /= 1000000;
 495         bdiff = 1 << 30;
 496         for (i = 0; i < count; i++) {
 497                 diff = abs(levels[i].total_set.freq - rate);
 498                 if (diff < bdiff) {
 499                         bdiff = diff;
 500                         sc->curr_level = levels[i];
 501                 }
 502         }
 503         CF_DEBUG("get estimated freq %d\n", curr_set->freq);
 504
 505 out:
 506         if (error == 0)
 507                 *level = sc->curr_level;
 508
 509         CF_MTX_UNLOCK(&sc->lock);
 510         if (levels)
 511                 free(levels, M_TEMP);
 512         return (error);
 513 }
 514
 515 static int
 516 cf_levels_method(device_t dev, struct cf_level *levels, int *count)
 517 {
 518         struct cf_setting_array *set_arr;
 519         struct cf_setting_lst rel_sets;
 520         struct cpufreq_softc *sc;
 521         struct cf_level *lev;
 522         struct cf_setting *sets;
 523         struct pcpu *pc;
 524         device_t *devs;
 525         int error, i, numdevs, set_count, type;
 526         uint64_t rate;
 527
 528         if (levels == NULL || count == NULL)
 529                 return (EINVAL);
 530
 531         TAILQ_INIT(&rel_sets);
 532         sc = device_get_softc(dev);
 533         error = device_get_children(device_get_parent(dev), &devs, &numdevs);
 534         if (error)
 535                 return (error);
 536         sets = malloc(MAX_SETTINGS * sizeof(*sets), M_TEMP, M_NOWAIT);
 537         if (sets == NULL) {
 538                 free(devs, M_TEMP);
 539                 return (ENOMEM);
 540         }
 541
 542         /* Get settings from all cpufreq drivers. */
 543         CF_MTX_LOCK(&sc->lock);
 544         for (i = 0; i < numdevs; i++) {
 545                 /* Skip devices that aren't ready. */
 546                 if (!device_is_attached(devs[i]))
 547                         continue;
 548
 549                 /*
 550                  * Get settings, skipping drivers that offer no settings or
 551                  * provide settings for informational purposes only.
 552                  */
 553                 error = CPUFREQ_DRV_TYPE(devs[i], &type);
 554                 if (error || (type & CPUFREQ_FLAG_INFO_ONLY)) {
 555                         if (error == 0) {
 556                                 CF_DEBUG("skipping info-only driver %s\n",
 557                                     device_get_nameunit(devs[i]));
 558                         }
 559                         continue;
 560                 }
 561                 set_count = MAX_SETTINGS;
 562                 error = CPUFREQ_DRV_SETTINGS(devs[i], sets, &set_count);
 563                 if (error || set_count == 0)
 564                         continue;
 565
 566                 /* Add the settings to our absolute/relative lists. */
 567                 switch (type & CPUFREQ_TYPE_MASK) {
 568                 case CPUFREQ_TYPE_ABSOLUTE:
 569                         error = cpufreq_insert_abs(sc, sets, set_count);
 570                         break;
 571                 case CPUFREQ_TYPE_RELATIVE:
 572                         CF_DEBUG("adding %d relative settings\n", set_count);
 573                         set_arr = malloc(sizeof(*set_arr), M_TEMP, M_NOWAIT);
 574                         if (set_arr == NULL) {
 575                                 error = ENOMEM;
 576                                 goto out;
 577                         }
 578                         bcopy(sets, set_arr->sets, set_count * sizeof(*sets));
 579                         set_arr->count = set_count;
 580                         TAILQ_INSERT_TAIL(&rel_sets, set_arr, link);
 581                         break;
 582                 default:
 583                         error = EINVAL;
 584                 }
 585                 if (error)
 586                         goto out;
 587         }
 588
 589         /*
 590          * If there are no absolute levels, create a fake one at 100%.  We
 591          * then cache the clockrate for later use as our base frequency.
 592          */
 593         if (TAILQ_EMPTY(&sc->all_levels)) {
 594                 if (sc->max_mhz == CPUFREQ_VAL_UNKNOWN) {
 595                         sc->max_mhz = cpu_get_nominal_mhz(dev);
 596                         /*
 597                          * If the CPU can't report a rate for 100%, hope
 598                          * the CPU is running at its nominal rate right now,
 599                          * and use that instead.
 600                          */
 601                         if (sc->max_mhz <= 0) {
 602                                 pc = cpu_get_pcpu(dev);
 603                                 cpu_est_clockrate(pc->pc_cpuid, &rate);
 604                                 sc->max_mhz = rate / 1000000;
 605                         }
 606                 }
 607                 memset(&sets[0], CPUFREQ_VAL_UNKNOWN, sizeof(*sets));
 608                 sets[0].freq = sc->max_mhz;
 609                 sets[0].dev = NULL;
 610                 error = cpufreq_insert_abs(sc, sets, 1);
 611                 if (error)
 612                         goto out;
 613         }
 614
 615         /* Create a combined list of absolute + relative levels. */
 616         TAILQ_FOREACH(set_arr, &rel_sets, link)
 617                 cpufreq_expand_set(sc, set_arr);
 618
 619         /* If the caller doesn't have enough space, return the actual count. */
 620         if (sc->all_count > *count) {
 621                 *count = sc->all_count;
 622                 error = E2BIG;
 623                 goto out;
 624         }
 625
 626         /* Finally, output the list of levels. */
 627         i = 0;
 628         TAILQ_FOREACH(lev, &sc->all_levels, link) {
 629
 630                 /* Skip levels that have a frequency that is too low. */
 631                 if (lev->total_set.freq < cf_lowest_freq) {
 632                         sc->all_count--;
 633                         continue;
 634                 }
 635
 636                 levels[i] = *lev;
 637                 i++;
 638         }
 639         *count = sc->all_count;
 640         error = 0;
 641
 642 out:
 643         /* Clear all levels since we regenerate them each time. */
 644         while ((lev = TAILQ_FIRST(&sc->all_levels)) != NULL) {
 645                 TAILQ_REMOVE(&sc->all_levels, lev, link);
 646                 free(lev, M_TEMP);
 647         }
 648         sc->all_count = 0;
 649
 650         CF_MTX_UNLOCK(&sc->lock);
 651         while ((set_arr = TAILQ_FIRST(&rel_sets)) != NULL) {
 652                 TAILQ_REMOVE(&rel_sets, set_arr, link);
 653                 free(set_arr, M_TEMP);
 654         }
 655         free(devs, M_TEMP);
 656         free(sets, M_TEMP);
 657         return (error);
 658 }
 659
 660 /*
 661  * Create levels for an array of absolute settings and insert them in
 662  * sorted order in the specified list.
 663  */
 664 static int
 665 cpufreq_insert_abs(struct cpufreq_softc *sc, struct cf_setting *sets,
 666     int count)
 667 {
 668         struct cf_level_lst *list;
 669         struct cf_level *level, *search;
 670         int i;
 671
 672         CF_MTX_ASSERT(&sc->lock);
 673
 674         list = &sc->all_levels;
 675         for (i = 0; i < count; i++) {
 676                 level = malloc(sizeof(*level), M_TEMP, M_NOWAIT | M_ZERO);
 677                 if (level == NULL)
 678                         return (ENOMEM);
 679                 level->abs_set = sets[i];
 680                 level->total_set = sets[i];
 681                 level->total_set.dev = NULL;
 682                 sc->all_count++;
 683
 684                 if (TAILQ_EMPTY(list)) {
 685                         CF_DEBUG("adding abs setting %d at head\n",
 686                             sets[i].freq);
 687                         TAILQ_INSERT_HEAD(list, level, link);
 688                         continue;
 689                 }
 690
 691                 TAILQ_FOREACH_REVERSE(search, list, cf_level_lst, link) {
 692                         if (sets[i].freq <= search->total_set.freq) {
 693                                 CF_DEBUG("adding abs setting %d after %d\n",
 694                                     sets[i].freq, search->total_set.freq);
 695                                 TAILQ_INSERT_AFTER(list, search, level, link);
 696                                 break;
 697                         }
 698                 }
 699         }
 700         return (0);
 701 }
 702
 703 /*
 704  * Expand a group of relative settings, creating derived levels from them.
 705  */
 706 static int
 707 cpufreq_expand_set(struct cpufreq_softc *sc, struct cf_setting_array *set_arr)
 708 {
 709         struct cf_level *fill, *search;
 710         struct cf_setting *set;
 711         int i;
 712
 713         CF_MTX_ASSERT(&sc->lock);
 714
 715         /*
 716          * Walk the set of all existing levels in reverse.  This is so we
 717          * create derived states from the lowest absolute settings first
 718          * and discard duplicates created from higher absolute settings.
 719          * For instance, a level of 50 Mhz derived from 100 Mhz + 50% is
 720          * preferable to 200 Mhz + 25% because absolute settings are more
 721          * efficient since they often change the voltage as well.
 722          */
 723         TAILQ_FOREACH_REVERSE(search, &sc->all_levels, cf_level_lst, link) {
 724                 /* Add each setting to the level, duplicating if necessary. */
 725                 for (i = 0; i < set_arr->count; i++) {
 726                         set = &set_arr->sets[i];
 727
 728                         /*
 729                          * If this setting is less than 100%, split the level
 730                          * into two and add this setting to the new level.
 731                          */
 732                         fill = search;
 733                         if (set->freq < 10000) {
 734                                 fill = cpufreq_dup_set(sc, search, set);
 735
 736                                 /*
 737                                  * The new level was a duplicate of an existing
 738                                  * level or its absolute setting is too high
 739                                  * so we freed it.  For example, we discard a
 740                                  * derived level of 1000 MHz/25% if a level
 741                                  * of 500 MHz/100% already exists.
 742                                  */
 743                                 if (fill == NULL)
 744                                         break;
 745                         }
 746
 747                         /* Add this setting to the existing or new level. */
 748                         KASSERT(fill->rel_count < MAX_SETTINGS,
 749                             ("cpufreq: too many relative drivers (%d)",
 750                             MAX_SETTINGS));
 751                         fill->rel_set[fill->rel_count] = *set;
 752                         fill->rel_count++;
 753                         CF_DEBUG(
 754                         "expand set added rel setting %d%% to %d level\n",
 755                             set->freq / 100, fill->total_set.freq);
 756                 }
 757         }
 758
 759         return (0);
 760 }
 761
 762 static struct cf_level *
 763 cpufreq_dup_set(struct cpufreq_softc *sc, struct cf_level *dup,
 764     struct cf_setting *set)
 765 {
 766         struct cf_level_lst *list;
 767         struct cf_level *fill, *itr;
 768         struct cf_setting *fill_set, *itr_set;
 769         int i;
 770
 771         CF_MTX_ASSERT(&sc->lock);
 772
 773         /*
 774          * Create a new level, copy it from the old one, and update the
 775          * total frequency and power by the percentage specified in the
 776          * relative setting.
 777          */
 778         fill = malloc(sizeof(*fill), M_TEMP, M_NOWAIT);
 779         if (fill == NULL)
 780                 return (NULL);
 781         *fill = *dup;
 782         fill_set = &fill->total_set;
 783         fill_set->freq =
 784             ((uint64_t)fill_set->freq * set->freq) / 10000;
 785         if (fill_set->power != CPUFREQ_VAL_UNKNOWN) {
 786                 fill_set->power = ((uint64_t)fill_set->power * set->freq)
 787                     / 10000;
 788         }
 789         if (set->lat != CPUFREQ_VAL_UNKNOWN) {
 790                 if (fill_set->lat != CPUFREQ_VAL_UNKNOWN)
 791                         fill_set->lat += set->lat;
 792                 else
 793                         fill_set->lat = set->lat;
 794         }
 795         CF_DEBUG("dup set considering derived setting %d\n", fill_set->freq);
 796
 797         /*
 798          * If we copied an old level that we already modified (say, at 100%),
 799          * we need to remove that setting before adding this one.  Since we
 800          * process each setting array in order, we know any settings for this
 801          * driver will be found at the end.
 802          */
 803         for (i = fill->rel_count; i != 0; i--) {
 804                 if (fill->rel_set[i - 1].dev != set->dev)
 805                         break;
 806                 CF_DEBUG("removed last relative driver: %s\n",
 807                     device_get_nameunit(set->dev));
 808                 fill->rel_count--;
 809         }
 810
 811         /*
 812          * Insert the new level in sorted order.  If it is a duplicate of an
 813          * existing level (1) or has an absolute setting higher than the
 814          * existing level (2), do not add it.  We can do this since any such
 815          * level is guaranteed use less power.  For example (1), a level with
 816          * one absolute setting of 800 Mhz uses less power than one composed
 817          * of an absolute setting of 1600 Mhz and a relative setting at 50%.
 818          * Also for example (2), a level of 800 Mhz/75% is preferable to
 819          * 1600 Mhz/25% even though the latter has a lower total frequency.
 820          */
 821         list = &sc->all_levels;
 822         KASSERT(!TAILQ_EMPTY(list), ("all levels list empty in dup set"));
 823         TAILQ_FOREACH_REVERSE(itr, list, cf_level_lst, link) {
 824                 itr_set = &itr->total_set;
 825                 if (CPUFREQ_CMP(fill_set->freq, itr_set->freq)) {
 826                         CF_DEBUG("dup set rejecting %d (dupe)\n",
 827                             fill_set->freq);
 828                         itr = NULL;
 829                         break;
 830                 } else if (fill_set->freq < itr_set->freq) {
 831                         if (fill->abs_set.freq <= itr->abs_set.freq) {
 832                                 CF_DEBUG(
 833                         "dup done, inserting new level %d after %d\n",
 834                                     fill_set->freq, itr_set->freq);
 835                                 TAILQ_INSERT_AFTER(list, itr, fill, link);
 836                                 sc->all_count++;
 837                         } else {
 838                                 CF_DEBUG("dup set rejecting %d (abs too big)\n",
 839                                     fill_set->freq);
 840                                 itr = NULL;
 841                         }
 842                         break;
 843                 }
 844         }
 845
 846         /* We didn't find a good place for this new level so free it. */
 847         if (itr == NULL) {
 848                 CF_DEBUG("dup set freeing new level %d (not optimal)\n",
 849                     fill_set->freq);
 850                 free(fill, M_TEMP);
 851                 fill = NULL;
 852         }
 853
 854         return (fill);
 855 }
 856
 857 static int
 858 cpufreq_curr_sysctl(SYSCTL_HANDLER_ARGS)
 859 {
 860         struct cpufreq_softc *sc;
 861         struct cf_level *levels;
 862         int best, count, diff, bdiff, devcount, error, freq, i, n;
 863         device_t *devs;
 864
 865         devs = NULL;
 866         sc = oidp->oid_arg1;
 867         levels = sc->levels_buf;
 868
 869         error = CPUFREQ_GET(sc->dev, &levels[0]);
 870         if (error)
 871                 goto out;
 872         freq = levels[0].total_set.freq;
 873         error = sysctl_handle_int(oidp, &freq, 0, req);
 874         if (error != 0 || req->newptr == NULL)
 875                 goto out;
 876
 877         /*
 878          * While we only call cpufreq_get() on one device (assuming all
 879          * CPUs have equal levels), we call cpufreq_set() on all CPUs.
 880          * This is needed for some MP systems.
 881          */
 882         error = devclass_get_devices(cpufreq_dc, &devs, &devcount);
 883         if (error)
 884                 goto out;
 885         for (n = 0; n < devcount; n++) {
 886                 count = CF_MAX_LEVELS;
 887                 error = CPUFREQ_LEVELS(devs[n], levels, &count);
 888                 if (error) {
 889                         if (error == E2BIG)
 890                                 printf(
 891                         "cpufreq: need to increase CF_MAX_LEVELS\n");
 892                         break;
 893                 }
 894                 best = 0;
 895                 bdiff = 1 << 30;
 896                 for (i = 0; i < count; i++) {
 897                         diff = abs(levels[i].total_set.freq - freq);
 898                         if (diff < bdiff) {
 899                                 bdiff = diff;
 900                                 best = i;
 901                         }
 902                 }
 903                 error = CPUFREQ_SET(devs[n], &levels[best], CPUFREQ_PRIO_USER);
 904         }
 905
 906 out:
 907         if (devs)
 908                 free(devs, M_TEMP);
 909         return (error);
 910 }
 911
 912 static int
 913 cpufreq_levels_sysctl(SYSCTL_HANDLER_ARGS)
 914 {
 915         struct cpufreq_softc *sc;
 916         struct cf_level *levels;
 917         struct cf_setting *set;
 918         struct sbuf sb;
 919         int count, error, i;
 920
 921         sc = oidp->oid_arg1;
 922         sbuf_new(&sb, NULL, 128, SBUF_AUTOEXTEND);
 923
 924         /* Get settings from the device and generate the output string. */
 925         count = CF_MAX_LEVELS;
 926         levels = sc->levels_buf;
 927         if (levels == NULL) {
 928                 sbuf_delete(&sb);
 929                 return (ENOMEM);
 930         }
 931         error = CPUFREQ_LEVELS(sc->dev, levels, &count);
 932         if (error) {
 933                 if (error == E2BIG)
 934                         printf("cpufreq: need to increase CF_MAX_LEVELS\n");
 935                 goto out;
 936         }
 937         if (count) {
 938                 for (i = 0; i < count; i++) {
 939                         set = &levels[i].total_set;
 940                         sbuf_printf(&sb, "%d/%d ", set->freq, set->power);
 941                 }
 942         } else
 943                 sbuf_cpy(&sb, "0");
 944         sbuf_trim(&sb);
 945         sbuf_finish(&sb);
 946         error = sysctl_handle_string(oidp, sbuf_data(&sb), sbuf_len(&sb), req);
 947
 948 out:
 949         sbuf_delete(&sb);
 950         return (error);
 951 }
 952
 953 static int
 954 cpufreq_settings_sysctl(SYSCTL_HANDLER_ARGS)
 955 {
 956         device_t dev;
 957         struct cf_setting *sets;
 958         struct sbuf sb;
 959         int error, i, set_count;
 960
 961         dev = oidp->oid_arg1;
 962         sbuf_new(&sb, NULL, 128, SBUF_AUTOEXTEND);
 963
 964         /* Get settings from the device and generate the output string. */
 965         set_count = MAX_SETTINGS;
 966         sets = malloc(set_count * sizeof(*sets), M_TEMP, M_NOWAIT);
 967         if (sets == NULL) {
 968                 sbuf_delete(&sb);
 969                 return (ENOMEM);
 970         }
 971         error = CPUFREQ_DRV_SETTINGS(dev, sets, &set_count);
 972         if (error)
 973                 goto out;
 974         if (set_count) {
 975                 for (i = 0; i < set_count; i++)
 976                         sbuf_printf(&sb, "%d/%d ", sets[i].freq, sets[i].power);
 977         } else
 978                 sbuf_cpy(&sb, "0");
 979         sbuf_trim(&sb);
 980         sbuf_finish(&sb);
 981         error = sysctl_handle_string(oidp, sbuf_data(&sb), sbuf_len(&sb), req);
 982
 983 out:
 984         free(sets, M_TEMP);
 985         sbuf_delete(&sb);
 986         return (error);
 987 }
 988
 989 int
 990 cpufreq_register(device_t dev)
 991 {
 992         struct cpufreq_softc *sc;
 993         device_t cf_dev, cpu_dev;
 994
 995         /* Add a sysctl to get each driver's settings separately. */
 996         SYSCTL_ADD_PROC(device_get_sysctl_ctx(dev),
 997             SYSCTL_CHILDREN(device_get_sysctl_tree(dev)),
 998             OID_AUTO, "freq_settings", CTLTYPE_STRING | CTLFLAG_RD, dev, 0,
 999             cpufreq_settings_sysctl, "A", "CPU frequency driver settings");
1000
1001         /*
1002          * Add only one cpufreq device to each CPU.  Currently, all CPUs
1003          * must offer the same levels and be switched at the same time.
1004          */
1005         cpu_dev = device_get_parent(dev);
1006         if ((cf_dev = device_find_child(cpu_dev, "cpufreq", -1))) {
1007                 sc = device_get_softc(cf_dev);
1008                 sc->max_mhz = CPUFREQ_VAL_UNKNOWN;
1009                 return (0);
1010         }
1011
1012         /* Add the child device and possibly sysctls. */
1013         cf_dev = BUS_ADD_CHILD(cpu_dev, 0, "cpufreq", -1);
1014         if (cf_dev == NULL)
1015                 return (ENOMEM);
1016         device_quiet(cf_dev);
1017
1018         return (device_probe_and_attach(cf_dev));
1019 }
1020
1021 int
1022 cpufreq_unregister(device_t dev)
1023 {
1024         device_t cf_dev, *devs;
1025         int cfcount, devcount, error, i, type;
1026
1027         /*
1028          * If this is the last cpufreq child device, remove the control
1029          * device as well.  We identify cpufreq children by calling a method
1030          * they support.
1031          */
1032         error = device_get_children(device_get_parent(dev), &devs, &devcount);
1033         if (error)
1034                 return (error);
1035         cf_dev = device_find_child(device_get_parent(dev), "cpufreq", -1);
1036         if (cf_dev == NULL) {
1037                 device_printf(dev,
1038         "warning: cpufreq_unregister called with no cpufreq device active\n");
1039                 free(devs, M_TEMP);
1040                 return (0);
1041         }
1042         cfcount = 0;
1043         for (i = 0; i < devcount; i++) {
1044                 if (!device_is_attached(devs[i]))
1045                         continue;
1046                 if (CPUFREQ_DRV_TYPE(devs[i], &type) == 0)
1047                         cfcount++;
1048         }
1049         if (cfcount <= 1)
1050                 device_delete_child(device_get_parent(cf_dev), cf_dev);
1051         free(devs, M_TEMP);
1052
1053         return (0);
1054 }
1055
1056 int
1057 cpufreq_settings_changed(device_t dev)
1058 {
1059
1060         EVENTHANDLER_INVOKE(cpufreq_levels_changed,
1061             device_get_unit(device_get_parent(dev)));
1062         return (0);
1063 }