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 TUNABLE_INT("debug.cpufreq.lowest", &cf_lowest_freq);
 137 TUNABLE_INT("debug.cpufreq.verbose", &cf_verbose);
 138 SYSCTL_NODE(_debug, OID_AUTO, cpufreq, CTLFLAG_RD, NULL, "cpufreq debugging");
 139 SYSCTL_INT(_debug_cpufreq, OID_AUTO, lowest, CTLFLAG_RW, &cf_lowest_freq, 1,
 140     "Don't provide levels below this frequency.");
 141 SYSCTL_INT(_debug_cpufreq, OID_AUTO, verbose, CTLFLAG_RW, &cf_verbose, 1,
 142     "Print verbose debugging messages");
 143
 144 static int
 145 cpufreq_attach(device_t dev)
 146 {
 147         struct cpufreq_softc *sc;
 148         struct pcpu *pc;
 149         device_t parent;
 150         uint64_t rate;
 151         int numdevs;
 152
 153         CF_DEBUG("initializing %s\n", device_get_nameunit(dev));
 154         sc = device_get_softc(dev);
 155         parent = device_get_parent(dev);
 156         sc->dev = dev;
 157         sysctl_ctx_init(&sc->sysctl_ctx);
 158         TAILQ_INIT(&sc->all_levels);
 159         CF_MTX_INIT(&sc->lock);
 160         sc->curr_level.total_set.freq = CPUFREQ_VAL_UNKNOWN;
 161         SLIST_INIT(&sc->saved_freq);
 162         /* Try to get nominal CPU freq to use it as maximum later if needed */
 163         sc->max_mhz = cpu_get_nominal_mhz(dev);
 164         /* If that fails, try to measure the current rate */
 165         if (sc->max_mhz <= 0) {
 166                 pc = cpu_get_pcpu(dev);
 167                 if (cpu_est_clockrate(pc->pc_cpuid, &rate) == 0)
 168                         sc->max_mhz = rate / 1000000;
 169                 else
 170                         sc->max_mhz = CPUFREQ_VAL_UNKNOWN;
 171         }
 172
 173         /*
 174          * Only initialize one set of sysctls for all CPUs.  In the future,
 175          * if multiple CPUs can have different settings, we can move these
 176          * sysctls to be under every CPU instead of just the first one.
 177          */
 178         numdevs = devclass_get_count(cpufreq_dc);
 179         if (numdevs > 1)
 180                 return (0);
 181
 182         CF_DEBUG("initializing one-time data for %s\n",
 183             device_get_nameunit(dev));
 184         sc->levels_buf = malloc(CF_MAX_LEVELS * sizeof(*sc->levels_buf),
 185             M_DEVBUF, M_WAITOK);
 186         SYSCTL_ADD_PROC(&sc->sysctl_ctx,
 187             SYSCTL_CHILDREN(device_get_sysctl_tree(parent)),
 188             OID_AUTO, "freq", CTLTYPE_INT | CTLFLAG_RW, sc, 0,
 189             cpufreq_curr_sysctl, "I", "Current CPU frequency");
 190         SYSCTL_ADD_PROC(&sc->sysctl_ctx,
 191             SYSCTL_CHILDREN(device_get_sysctl_tree(parent)),
 192             OID_AUTO, "freq_levels", CTLTYPE_STRING | CTLFLAG_RD, sc, 0,
 193             cpufreq_levels_sysctl, "A", "CPU frequency levels");
 194
 195         /*
 196          * Queue a one-shot broadcast that levels have changed.
 197          * It will run once the system has completed booting.
 198          */
 199         TASK_INIT(&sc->startup_task, 0, cpufreq_startup_task, dev);
 200         taskqueue_enqueue(taskqueue_thread, &sc->startup_task);
 201
 202         return (0);
 203 }
 204
 205 /* Handle any work to be done for all drivers that attached during boot. */
 206 static void
 207 cpufreq_startup_task(void *ctx, int pending)
 208 {
 209
 210         cpufreq_settings_changed((device_t)ctx);
 211 }
 212
 213 static int
 214 cpufreq_detach(device_t dev)
 215 {
 216         struct cpufreq_softc *sc;
 217         struct cf_saved_freq *saved_freq;
 218         int numdevs;
 219
 220         CF_DEBUG("shutdown %s\n", device_get_nameunit(dev));
 221         sc = device_get_softc(dev);
 222         sysctl_ctx_free(&sc->sysctl_ctx);
 223
 224         while ((saved_freq = SLIST_FIRST(&sc->saved_freq)) != NULL) {
 225                 SLIST_REMOVE_HEAD(&sc->saved_freq, link);
 226                 free(saved_freq, M_TEMP);
 227         }
 228
 229         /* Only clean up these resources when the last device is detaching. */
 230         numdevs = devclass_get_count(cpufreq_dc);
 231         if (numdevs == 1) {
 232                 CF_DEBUG("final shutdown for %s\n", device_get_nameunit(dev));
 233                 free(sc->levels_buf, M_DEVBUF);
 234         }
 235
 236         return (0);
 237 }
 238
 239 static int
 240 cf_set_method(device_t dev, const struct cf_level *level, int priority)
 241 {
 242         struct cpufreq_softc *sc;
 243         const struct cf_setting *set;
 244         struct cf_saved_freq *saved_freq, *curr_freq;
 245         struct pcpu *pc;
 246         int error, i;
 247
 248         sc = device_get_softc(dev);
 249         error = 0;
 250         set = NULL;
 251         saved_freq = NULL;
 252
 253         /* We are going to change levels so notify the pre-change handler. */
 254         EVENTHANDLER_INVOKE(cpufreq_pre_change, level, &error);
 255         if (error != 0) {
 256                 EVENTHANDLER_INVOKE(cpufreq_post_change, level, error);
 257                 return (error);
 258         }
 259
 260         CF_MTX_LOCK(&sc->lock);
 261
 262 #ifdef SMP
 263         /*
 264          * If still booting and secondary CPUs not started yet, don't allow
 265          * changing the frequency until they're online.  This is because we
 266          * can't switch to them using sched_bind() and thus we'd only be
 267          * switching the main CPU.  XXXTODO: Need to think more about how to
 268          * handle having different CPUs at different frequencies.
 269          */
 270         if (mp_ncpus > 1 && !smp_active) {
 271                 device_printf(dev, "rejecting change, SMP not started yet\n");
 272                 error = ENXIO;
 273                 goto out;
 274         }
 275 #endif /* SMP */
 276
 277         /*
 278          * If the requested level has a lower priority, don't allow
 279          * the new level right now.
 280          */
 281         if (priority < sc->curr_priority) {
 282                 CF_DEBUG("ignoring, curr prio %d less than %d\n", priority,
 283                     sc->curr_priority);
 284                 error = EPERM;
 285                 goto out;
 286         }
 287
 288         /*
 289          * If the caller didn't specify a level and one is saved, prepare to
 290          * restore the saved level.  If none has been saved, return an error.
 291          */
 292         if (level == NULL) {
 293                 saved_freq = SLIST_FIRST(&sc->saved_freq);
 294                 if (saved_freq == NULL) {
 295                         CF_DEBUG("NULL level, no saved level\n");
 296                         error = ENXIO;
 297                         goto out;
 298                 }
 299                 level = &saved_freq->level;
 300                 priority = saved_freq->priority;
 301                 CF_DEBUG("restoring saved level, freq %d prio %d\n",
 302                     level->total_set.freq, priority);
 303         }
 304
 305         /* Reject levels that are below our specified threshold. */
 306         if (level->total_set.freq < cf_lowest_freq) {
 307                 CF_DEBUG("rejecting freq %d, less than %d limit\n",
 308                     level->total_set.freq, cf_lowest_freq);
 309                 error = EINVAL;
 310                 goto out;
 311         }
 312
 313         /* If already at this level, just return. */
 314         if (CPUFREQ_CMP(sc->curr_level.total_set.freq, level->total_set.freq)) {
 315                 CF_DEBUG("skipping freq %d, same as current level %d\n",
 316                     level->total_set.freq, sc->curr_level.total_set.freq);
 317                 goto skip;
 318         }
 319
 320         /* First, set the absolute frequency via its driver. */
 321         set = &level->abs_set;
 322         if (set->dev) {
 323                 if (!device_is_attached(set->dev)) {
 324                         error = ENXIO;
 325                         goto out;
 326                 }
 327
 328                 /* Bind to the target CPU before switching. */
 329                 pc = cpu_get_pcpu(set->dev);
 330                 thread_lock(curthread);
 331                 sched_bind(curthread, pc->pc_cpuid);
 332                 thread_unlock(curthread);
 333                 CF_DEBUG("setting abs freq %d on %s (cpu %d)\n", set->freq,
 334                     device_get_nameunit(set->dev), PCPU_GET(cpuid));
 335                 error = CPUFREQ_DRV_SET(set->dev, set);
 336                 thread_lock(curthread);
 337                 sched_unbind(curthread);
 338                 thread_unlock(curthread);
 339                 if (error) {
 340                         goto out;
 341                 }
 342         }
 343
 344         /* Next, set any/all relative frequencies via their drivers. */
 345         for (i = 0; i < level->rel_count; i++) {
 346                 set = &level->rel_set[i];
 347                 if (!device_is_attached(set->dev)) {
 348                         error = ENXIO;
 349                         goto out;
 350                 }
 351
 352                 /* Bind to the target CPU before switching. */
 353                 pc = cpu_get_pcpu(set->dev);
 354                 thread_lock(curthread);
 355                 sched_bind(curthread, pc->pc_cpuid);
 356                 thread_unlock(curthread);
 357                 CF_DEBUG("setting rel freq %d on %s (cpu %d)\n", set->freq,
 358                     device_get_nameunit(set->dev), PCPU_GET(cpuid));
 359                 error = CPUFREQ_DRV_SET(set->dev, set);
 360                 thread_lock(curthread);
 361                 sched_unbind(curthread);
 362                 thread_unlock(curthread);
 363                 if (error) {
 364                         /* XXX Back out any successful setting? */
 365                         goto out;
 366                 }
 367         }
 368
 369 skip:
 370         /*
 371          * Before recording the current level, check if we're going to a
 372          * higher priority.  If so, save the previous level and priority.
 373          */
 374         if (sc->curr_level.total_set.freq != CPUFREQ_VAL_UNKNOWN &&
 375             priority > sc->curr_priority) {
 376                 CF_DEBUG("saving level, freq %d prio %d\n",
 377                     sc->curr_level.total_set.freq, sc->curr_priority);
 378                 curr_freq = malloc(sizeof(*curr_freq), M_TEMP, M_NOWAIT);
 379                 if (curr_freq == NULL) {
 380                         error = ENOMEM;
 381                         goto out;
 382                 }
 383                 curr_freq->level = sc->curr_level;
 384                 curr_freq->priority = sc->curr_priority;
 385                 SLIST_INSERT_HEAD(&sc->saved_freq, curr_freq, link);
 386         }
 387         sc->curr_level = *level;
 388         sc->curr_priority = priority;
 389
 390         /* If we were restoring a saved state, reset it to "unused". */
 391         if (saved_freq != NULL) {
 392                 CF_DEBUG("resetting saved level\n");
 393                 sc->curr_level.total_set.freq = CPUFREQ_VAL_UNKNOWN;
 394                 SLIST_REMOVE_HEAD(&sc->saved_freq, link);
 395                 free(saved_freq, M_TEMP);
 396         }
 397
 398 out:
 399         CF_MTX_UNLOCK(&sc->lock);
 400
 401         /*
 402          * We changed levels (or attempted to) so notify the post-change
 403          * handler of new frequency or error.
 404          */
 405         EVENTHANDLER_INVOKE(cpufreq_post_change, level, error);
 406         if (error && set)
 407                 device_printf(set->dev, "set freq failed, err %d\n", error);
 408
 409         return (error);
 410 }
 411
 412 static int
 413 cf_get_method(device_t dev, struct cf_level *level)
 414 {
 415         struct cpufreq_softc *sc;
 416         struct cf_level *levels;
 417         struct cf_setting *curr_set, set;
 418         struct pcpu *pc;
 419         device_t *devs;
 420         int count, error, i, n, numdevs;
 421         uint64_t rate;
 422
 423         sc = device_get_softc(dev);
 424         error = 0;
 425         levels = NULL;
 426
 427         /* If we already know the current frequency, we're done. */
 428         CF_MTX_LOCK(&sc->lock);
 429         curr_set = &sc->curr_level.total_set;
 430         if (curr_set->freq != CPUFREQ_VAL_UNKNOWN) {
 431                 CF_DEBUG("get returning known freq %d\n", curr_set->freq);
 432                 goto out;
 433         }
 434         CF_MTX_UNLOCK(&sc->lock);
 435
 436         /*
 437          * We need to figure out the current level.  Loop through every
 438          * driver, getting the current setting.  Then, attempt to get a best
 439          * match of settings against each level.
 440          */
 441         count = CF_MAX_LEVELS;
 442         levels = malloc(count * sizeof(*levels), M_TEMP, M_NOWAIT);
 443         if (levels == NULL)
 444                 return (ENOMEM);
 445         error = CPUFREQ_LEVELS(sc->dev, levels, &count);
 446         if (error) {
 447                 if (error == E2BIG)
 448                         printf("cpufreq: need to increase CF_MAX_LEVELS\n");
 449                 free(levels, M_TEMP);
 450                 return (error);
 451         }
 452         error = device_get_children(device_get_parent(dev), &devs, &numdevs);
 453         if (error) {
 454                 free(levels, M_TEMP);
 455                 return (error);
 456         }
 457
 458         /*
 459          * Reacquire the lock and search for the given level.
 460          *
 461          * XXX Note: this is not quite right since we really need to go
 462          * through each level and compare both absolute and relative
 463          * settings for each driver in the system before making a match.
 464          * The estimation code below catches this case though.
 465          */
 466         CF_MTX_LOCK(&sc->lock);
 467         for (n = 0; n < numdevs && curr_set->freq == CPUFREQ_VAL_UNKNOWN; n++) {
 468                 if (!device_is_attached(devs[n]))
 469                         continue;
 470                 if (CPUFREQ_DRV_GET(devs[n], &set) != 0)
 471                         continue;
 472                 for (i = 0; i < count; i++) {
 473                         if (CPUFREQ_CMP(set.freq, levels[i].total_set.freq)) {
 474                                 sc->curr_level = levels[i];
 475                                 break;
 476                         }
 477                 }
 478         }
 479         free(devs, M_TEMP);
 480         if (curr_set->freq != CPUFREQ_VAL_UNKNOWN) {
 481                 CF_DEBUG("get matched freq %d from drivers\n", curr_set->freq);
 482                 goto out;
 483         }
 484
 485         /*
 486          * We couldn't find an exact match, so attempt to estimate and then
 487          * match against a level.
 488          */
 489         pc = cpu_get_pcpu(dev);
 490         if (pc == NULL) {
 491                 error = ENXIO;
 492                 goto out;
 493         }
 494         cpu_est_clockrate(pc->pc_cpuid, &rate);
 495         rate /= 1000000;
 496         for (i = 0; i < count; i++) {
 497                 if (CPUFREQ_CMP(rate, levels[i].total_set.freq)) {
 498                         sc->curr_level = levels[i];
 499                         CF_DEBUG("get estimated freq %d\n", curr_set->freq);
 500                         goto out;
 501                 }
 502         }
 503         error = ENXIO;
 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 are too close in frequency to the
 631                  * previous levels.  Some systems report bogus duplicate
 632                  * settings (i.e., for acpi_perf).
 633                  */
 634                 if (i > 0 && CPUFREQ_CMP(lev->total_set.freq,
 635                     levels[i - 1].total_set.freq)) {
 636                         sc->all_count--;
 637                         continue;
 638                 }
 639
 640                 /* Skip levels that have a frequency that is too low. */
 641                 if (lev->total_set.freq < cf_lowest_freq) {
 642                         sc->all_count--;
 643                         continue;
 644                 }
 645
 646                 levels[i] = *lev;
 647                 i++;
 648         }
 649         *count = sc->all_count;
 650         error = 0;
 651
 652 out:
 653         /* Clear all levels since we regenerate them each time. */
 654         while ((lev = TAILQ_FIRST(&sc->all_levels)) != NULL) {
 655                 TAILQ_REMOVE(&sc->all_levels, lev, link);
 656                 free(lev, M_TEMP);
 657         }
 658         sc->all_count = 0;
 659
 660         CF_MTX_UNLOCK(&sc->lock);
 661         while ((set_arr = TAILQ_FIRST(&rel_sets)) != NULL) {
 662                 TAILQ_REMOVE(&rel_sets, set_arr, link);
 663                 free(set_arr, M_TEMP);
 664         }
 665         free(devs, M_TEMP);
 666         free(sets, M_TEMP);
 667         return (error);
 668 }
 669
 670 /*
 671  * Create levels for an array of absolute settings and insert them in
 672  * sorted order in the specified list.
 673  */
 674 static int
 675 cpufreq_insert_abs(struct cpufreq_softc *sc, struct cf_setting *sets,
 676     int count)
 677 {
 678         struct cf_level_lst *list;
 679         struct cf_level *level, *search;
 680         int i;
 681
 682         CF_MTX_ASSERT(&sc->lock);
 683
 684         list = &sc->all_levels;
 685         for (i = 0; i < count; i++) {
 686                 level = malloc(sizeof(*level), M_TEMP, M_NOWAIT | M_ZERO);
 687                 if (level == NULL)
 688                         return (ENOMEM);
 689                 level->abs_set = sets[i];
 690                 level->total_set = sets[i];
 691                 level->total_set.dev = NULL;
 692                 sc->all_count++;
 693
 694                 if (TAILQ_EMPTY(list)) {
 695                         CF_DEBUG("adding abs setting %d at head\n",
 696                             sets[i].freq);
 697                         TAILQ_INSERT_HEAD(list, level, link);
 698                         continue;
 699                 }
 700
 701                 TAILQ_FOREACH_REVERSE(search, list, cf_level_lst, link) {
 702                         if (sets[i].freq <= search->total_set.freq) {
 703                                 CF_DEBUG("adding abs setting %d after %d\n",
 704                                     sets[i].freq, search->total_set.freq);
 705                                 TAILQ_INSERT_AFTER(list, search, level, link);
 706                                 break;
 707                         }
 708                 }
 709         }
 710         return (0);
 711 }
 712
 713 /*
 714  * Expand a group of relative settings, creating derived levels from them.
 715  */
 716 static int
 717 cpufreq_expand_set(struct cpufreq_softc *sc, struct cf_setting_array *set_arr)
 718 {
 719         struct cf_level *fill, *search;
 720         struct cf_setting *set;
 721         int i;
 722
 723         CF_MTX_ASSERT(&sc->lock);
 724
 725         /*
 726          * Walk the set of all existing levels in reverse.  This is so we
 727          * create derived states from the lowest absolute settings first
 728          * and discard duplicates created from higher absolute settings.
 729          * For instance, a level of 50 Mhz derived from 100 Mhz + 50% is
 730          * preferable to 200 Mhz + 25% because absolute settings are more
 731          * efficient since they often change the voltage as well.
 732          */
 733         TAILQ_FOREACH_REVERSE(search, &sc->all_levels, cf_level_lst, link) {
 734                 /* Add each setting to the level, duplicating if necessary. */
 735                 for (i = 0; i < set_arr->count; i++) {
 736                         set = &set_arr->sets[i];
 737
 738                         /*
 739                          * If this setting is less than 100%, split the level
 740                          * into two and add this setting to the new level.
 741                          */
 742                         fill = search;
 743                         if (set->freq < 10000) {
 744                                 fill = cpufreq_dup_set(sc, search, set);
 745
 746                                 /*
 747                                  * The new level was a duplicate of an existing
 748                                  * level or its absolute setting is too high
 749                                  * so we freed it.  For example, we discard a
 750                                  * derived level of 1000 MHz/25% if a level
 751                                  * of 500 MHz/100% already exists.
 752                                  */
 753                                 if (fill == NULL)
 754                                         break;
 755                         }
 756
 757                         /* Add this setting to the existing or new level. */
 758                         KASSERT(fill->rel_count < MAX_SETTINGS,
 759                             ("cpufreq: too many relative drivers (%d)",
 760                             MAX_SETTINGS));
 761                         fill->rel_set[fill->rel_count] = *set;
 762                         fill->rel_count++;
 763                         CF_DEBUG(
 764                         "expand set added rel setting %d%% to %d level\n",
 765                             set->freq / 100, fill->total_set.freq);
 766                 }
 767         }
 768
 769         return (0);
 770 }
 771
 772 static struct cf_level *
 773 cpufreq_dup_set(struct cpufreq_softc *sc, struct cf_level *dup,
 774     struct cf_setting *set)
 775 {
 776         struct cf_level_lst *list;
 777         struct cf_level *fill, *itr;
 778         struct cf_setting *fill_set, *itr_set;
 779         int i;
 780
 781         CF_MTX_ASSERT(&sc->lock);
 782
 783         /*
 784          * Create a new level, copy it from the old one, and update the
 785          * total frequency and power by the percentage specified in the
 786          * relative setting.
 787          */
 788         fill = malloc(sizeof(*fill), M_TEMP, M_NOWAIT);
 789         if (fill == NULL)
 790                 return (NULL);
 791         *fill = *dup;
 792         fill_set = &fill->total_set;
 793         fill_set->freq =
 794             ((uint64_t)fill_set->freq * set->freq) / 10000;
 795         if (fill_set->power != CPUFREQ_VAL_UNKNOWN) {
 796                 fill_set->power = ((uint64_t)fill_set->power * set->freq)
 797                     / 10000;
 798         }
 799         if (set->lat != CPUFREQ_VAL_UNKNOWN) {
 800                 if (fill_set->lat != CPUFREQ_VAL_UNKNOWN)
 801                         fill_set->lat += set->lat;
 802                 else
 803                         fill_set->lat = set->lat;
 804         }
 805         CF_DEBUG("dup set considering derived setting %d\n", fill_set->freq);
 806
 807         /*
 808          * If we copied an old level that we already modified (say, at 100%),
 809          * we need to remove that setting before adding this one.  Since we
 810          * process each setting array in order, we know any settings for this
 811          * driver will be found at the end.
 812          */
 813         for (i = fill->rel_count; i != 0; i--) {
 814                 if (fill->rel_set[i - 1].dev != set->dev)
 815                         break;
 816                 CF_DEBUG("removed last relative driver: %s\n",
 817                     device_get_nameunit(set->dev));
 818                 fill->rel_count--;
 819         }
 820
 821         /*
 822          * Insert the new level in sorted order.  If it is a duplicate of an
 823          * existing level (1) or has an absolute setting higher than the
 824          * existing level (2), do not add it.  We can do this since any such
 825          * level is guaranteed use less power.  For example (1), a level with
 826          * one absolute setting of 800 Mhz uses less power than one composed
 827          * of an absolute setting of 1600 Mhz and a relative setting at 50%.
 828          * Also for example (2), a level of 800 Mhz/75% is preferable to
 829          * 1600 Mhz/25% even though the latter has a lower total frequency.
 830          */
 831         list = &sc->all_levels;
 832         KASSERT(!TAILQ_EMPTY(list), ("all levels list empty in dup set"));
 833         TAILQ_FOREACH_REVERSE(itr, list, cf_level_lst, link) {
 834                 itr_set = &itr->total_set;
 835                 if (CPUFREQ_CMP(fill_set->freq, itr_set->freq)) {
 836                         CF_DEBUG("dup set rejecting %d (dupe)\n",
 837                             fill_set->freq);
 838                         itr = NULL;
 839                         break;
 840                 } else if (fill_set->freq < itr_set->freq) {
 841                         if (fill->abs_set.freq <= itr->abs_set.freq) {
 842                                 CF_DEBUG(
 843                         "dup done, inserting new level %d after %d\n",
 844                                     fill_set->freq, itr_set->freq);
 845                                 TAILQ_INSERT_AFTER(list, itr, fill, link);
 846                                 sc->all_count++;
 847                         } else {
 848                                 CF_DEBUG("dup set rejecting %d (abs too big)\n",
 849                                     fill_set->freq);
 850                                 itr = NULL;
 851                         }
 852                         break;
 853                 }
 854         }
 855
 856         /* We didn't find a good place for this new level so free it. */
 857         if (itr == NULL) {
 858                 CF_DEBUG("dup set freeing new level %d (not optimal)\n",
 859                     fill_set->freq);
 860                 free(fill, M_TEMP);
 861                 fill = NULL;
 862         }
 863
 864         return (fill);
 865 }
 866
 867 static int
 868 cpufreq_curr_sysctl(SYSCTL_HANDLER_ARGS)
 869 {
 870         struct cpufreq_softc *sc;
 871         struct cf_level *levels;
 872         int count, devcount, error, freq, i, n;
 873         device_t *devs;
 874
 875         devs = NULL;
 876         sc = oidp->oid_arg1;
 877         levels = sc->levels_buf;
 878
 879         error = CPUFREQ_GET(sc->dev, &levels[0]);
 880         if (error)
 881                 goto out;
 882         freq = levels[0].total_set.freq;
 883         error = sysctl_handle_int(oidp, &freq, 0, req);
 884         if (error != 0 || req->newptr == NULL)
 885                 goto out;
 886
 887         /*
 888          * While we only call cpufreq_get() on one device (assuming all
 889          * CPUs have equal levels), we call cpufreq_set() on all CPUs.
 890          * This is needed for some MP systems.
 891          */
 892         error = devclass_get_devices(cpufreq_dc, &devs, &devcount);
 893         if (error)
 894                 goto out;
 895         for (n = 0; n < devcount; n++) {
 896                 count = CF_MAX_LEVELS;
 897                 error = CPUFREQ_LEVELS(devs[n], levels, &count);
 898                 if (error) {
 899                         if (error == E2BIG)
 900                                 printf(
 901                         "cpufreq: need to increase CF_MAX_LEVELS\n");
 902                         break;
 903                 }
 904                 for (i = 0; i < count; i++) {
 905                         if (CPUFREQ_CMP(levels[i].total_set.freq, freq)) {
 906                                 error = CPUFREQ_SET(devs[n], &levels[i],
 907                                     CPUFREQ_PRIO_USER);
 908                                 break;
 909                         }
 910                 }
 911                 if (i == count) {
 912                         error = EINVAL;
 913                         break;
 914                 }
 915         }
 916
 917 out:
 918         if (devs)
 919                 free(devs, M_TEMP);
 920         return (error);
 921 }
 922
 923 static int
 924 cpufreq_levels_sysctl(SYSCTL_HANDLER_ARGS)
 925 {
 926         struct cpufreq_softc *sc;
 927         struct cf_level *levels;
 928         struct cf_setting *set;
 929         struct sbuf sb;
 930         int count, error, i;
 931
 932         sc = oidp->oid_arg1;
 933         sbuf_new(&sb, NULL, 128, SBUF_AUTOEXTEND);
 934
 935         /* Get settings from the device and generate the output string. */
 936         count = CF_MAX_LEVELS;
 937         levels = sc->levels_buf;
 938         if (levels == NULL) {
 939                 sbuf_delete(&sb);
 940                 return (ENOMEM);
 941         }
 942         error = CPUFREQ_LEVELS(sc->dev, levels, &count);
 943         if (error) {
 944                 if (error == E2BIG)
 945                         printf("cpufreq: need to increase CF_MAX_LEVELS\n");
 946                 goto out;
 947         }
 948         if (count) {
 949                 for (i = 0; i < count; i++) {
 950                         set = &levels[i].total_set;
 951                         sbuf_printf(&sb, "%d/%d ", set->freq, set->power);
 952                 }
 953         } else
 954                 sbuf_cpy(&sb, "0");
 955         sbuf_trim(&sb);
 956         sbuf_finish(&sb);
 957         error = sysctl_handle_string(oidp, sbuf_data(&sb), sbuf_len(&sb), req);
 958
 959 out:
 960         sbuf_delete(&sb);
 961         return (error);
 962 }
 963
 964 static int
 965 cpufreq_settings_sysctl(SYSCTL_HANDLER_ARGS)
 966 {
 967         device_t dev;
 968         struct cf_setting *sets;
 969         struct sbuf sb;
 970         int error, i, set_count;
 971
 972         dev = oidp->oid_arg1;
 973         sbuf_new(&sb, NULL, 128, SBUF_AUTOEXTEND);
 974
 975         /* Get settings from the device and generate the output string. */
 976         set_count = MAX_SETTINGS;
 977         sets = malloc(set_count * sizeof(*sets), M_TEMP, M_NOWAIT);
 978         if (sets == NULL) {
 979                 sbuf_delete(&sb);
 980                 return (ENOMEM);
 981         }
 982         error = CPUFREQ_DRV_SETTINGS(dev, sets, &set_count);
 983         if (error)
 984                 goto out;
 985         if (set_count) {
 986                 for (i = 0; i < set_count; i++)
 987                         sbuf_printf(&sb, "%d/%d ", sets[i].freq, sets[i].power);
 988         } else
 989                 sbuf_cpy(&sb, "0");
 990         sbuf_trim(&sb);
 991         sbuf_finish(&sb);
 992         error = sysctl_handle_string(oidp, sbuf_data(&sb), sbuf_len(&sb), req);
 993
 994 out:
 995         free(sets, M_TEMP);
 996         sbuf_delete(&sb);
 997         return (error);
 998 }
 999
1000 int
1001 cpufreq_register(device_t dev)
1002 {
1003         struct cpufreq_softc *sc;
1004         device_t cf_dev, cpu_dev;
1005
1006         /* Add a sysctl to get each driver's settings separately. */
1007         SYSCTL_ADD_PROC(device_get_sysctl_ctx(dev),
1008             SYSCTL_CHILDREN(device_get_sysctl_tree(dev)),
1009             OID_AUTO, "freq_settings", CTLTYPE_STRING | CTLFLAG_RD, dev, 0,
1010             cpufreq_settings_sysctl, "A", "CPU frequency driver settings");
1011
1012         /*
1013          * Add only one cpufreq device to each CPU.  Currently, all CPUs
1014          * must offer the same levels and be switched at the same time.
1015          */
1016         cpu_dev = device_get_parent(dev);
1017         if ((cf_dev = device_find_child(cpu_dev, "cpufreq", -1))) {
1018                 sc = device_get_softc(cf_dev);
1019                 sc->max_mhz = CPUFREQ_VAL_UNKNOWN;
1020                 return (0);
1021         }
1022
1023         /* Add the child device and possibly sysctls. */
1024         cf_dev = BUS_ADD_CHILD(cpu_dev, 0, "cpufreq", -1);
1025         if (cf_dev == NULL)
1026                 return (ENOMEM);
1027         device_quiet(cf_dev);
1028
1029         return (device_probe_and_attach(cf_dev));
1030 }
1031
1032 int
1033 cpufreq_unregister(device_t dev)
1034 {
1035         device_t cf_dev, *devs;
1036         int cfcount, devcount, error, i, type;
1037
1038         /*
1039          * If this is the last cpufreq child device, remove the control
1040          * device as well.  We identify cpufreq children by calling a method
1041          * they support.
1042          */
1043         error = device_get_children(device_get_parent(dev), &devs, &devcount);
1044         if (error)
1045                 return (error);
1046         cf_dev = device_find_child(device_get_parent(dev), "cpufreq", -1);
1047         if (cf_dev == NULL) {
1048                 device_printf(dev,
1049         "warning: cpufreq_unregister called with no cpufreq device active\n");
1050                 return (0);
1051         }
1052         cfcount = 0;
1053         for (i = 0; i < devcount; i++) {
1054                 if (!device_is_attached(devs[i]))
1055                         continue;
1056                 if (CPUFREQ_DRV_TYPE(devs[i], &type) == 0)
1057                         cfcount++;
1058         }
1059         if (cfcount <= 1)
1060                 device_delete_child(device_get_parent(cf_dev), cf_dev);
1061         free(devs, M_TEMP);
1062
1063         return (0);
1064 }
1065
1066 int
1067 cpufreq_settings_changed(device_t dev)
1068 {
1069
1070         EVENTHANDLER_INVOKE(cpufreq_levels_changed,
1071             device_get_unit(device_get_parent(dev)));
1072         return (0);
1073 }