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 #ifdef EARLY_AP_STARTUP
 263         MPASS(mp_ncpus == 1 || smp_started);
 264 #else
 265         /*
 266          * If still booting and secondary CPUs not started yet, don't allow
 267          * changing the frequency until they're online.  This is because we
 268          * can't switch to them using sched_bind() and thus we'd only be
 269          * switching the main CPU.  XXXTODO: Need to think more about how to
 270          * handle having different CPUs at different frequencies.
 271          */
 272         if (mp_ncpus > 1 && !smp_started) {
 273                 device_printf(dev, "rejecting change, SMP not started yet\n");
 274                 error = ENXIO;
 275                 goto out;
 276         }
 277 #endif
 278 #endif /* SMP */
 279
 280         /*
 281          * If the requested level has a lower priority, don't allow
 282          * the new level right now.
 283          */
 284         if (priority < sc->curr_priority) {
 285                 CF_DEBUG("ignoring, curr prio %d less than %d\n", priority,
 286                     sc->curr_priority);
 287                 error = EPERM;
 288                 goto out;
 289         }
 290
 291         /*
 292          * If the caller didn't specify a level and one is saved, prepare to
 293          * restore the saved level.  If none has been saved, return an error.
 294          */
 295         if (level == NULL) {
 296                 saved_freq = SLIST_FIRST(&sc->saved_freq);
 297                 if (saved_freq == NULL) {
 298                         CF_DEBUG("NULL level, no saved level\n");
 299                         error = ENXIO;
 300                         goto out;
 301                 }
 302                 level = &saved_freq->level;
 303                 priority = saved_freq->priority;
 304                 CF_DEBUG("restoring saved level, freq %d prio %d\n",
 305                     level->total_set.freq, priority);
 306         }
 307
 308         /* Reject levels that are below our specified threshold. */
 309         if (level->total_set.freq < cf_lowest_freq) {
 310                 CF_DEBUG("rejecting freq %d, less than %d limit\n",
 311                     level->total_set.freq, cf_lowest_freq);
 312                 error = EINVAL;
 313                 goto out;
 314         }
 315
 316         /* If already at this level, just return. */
 317         if (sc->curr_level.total_set.freq == level->total_set.freq) {
 318                 CF_DEBUG("skipping freq %d, same as current level %d\n",
 319                     level->total_set.freq, sc->curr_level.total_set.freq);
 320                 goto skip;
 321         }
 322
 323         /* First, set the absolute frequency via its driver. */
 324         set = &level->abs_set;
 325         if (set->dev) {
 326                 if (!device_is_attached(set->dev)) {
 327                         error = ENXIO;
 328                         goto out;
 329                 }
 330
 331                 /* Bind to the target CPU before switching. */
 332                 pc = cpu_get_pcpu(set->dev);
 333                 thread_lock(curthread);
 334                 sched_bind(curthread, pc->pc_cpuid);
 335                 thread_unlock(curthread);
 336                 CF_DEBUG("setting abs freq %d on %s (cpu %d)\n", set->freq,
 337                     device_get_nameunit(set->dev), PCPU_GET(cpuid));
 338                 error = CPUFREQ_DRV_SET(set->dev, set);
 339                 thread_lock(curthread);
 340                 sched_unbind(curthread);
 341                 thread_unlock(curthread);
 342                 if (error) {
 343                         goto out;
 344                 }
 345         }
 346
 347         /* Next, set any/all relative frequencies via their drivers. */
 348         for (i = 0; i < level->rel_count; i++) {
 349                 set = &level->rel_set[i];
 350                 if (!device_is_attached(set->dev)) {
 351                         error = ENXIO;
 352                         goto out;
 353                 }
 354
 355                 /* Bind to the target CPU before switching. */
 356                 pc = cpu_get_pcpu(set->dev);
 357                 thread_lock(curthread);
 358                 sched_bind(curthread, pc->pc_cpuid);
 359                 thread_unlock(curthread);
 360                 CF_DEBUG("setting rel freq %d on %s (cpu %d)\n", set->freq,
 361                     device_get_nameunit(set->dev), PCPU_GET(cpuid));
 362                 error = CPUFREQ_DRV_SET(set->dev, set);
 363                 thread_lock(curthread);
 364                 sched_unbind(curthread);
 365                 thread_unlock(curthread);
 366                 if (error) {
 367                         /* XXX Back out any successful setting? */
 368                         goto out;
 369                 }
 370         }
 371
 372 skip:
 373         /*
 374          * Before recording the current level, check if we're going to a
 375          * higher priority.  If so, save the previous level and priority.
 376          */
 377         if (sc->curr_level.total_set.freq != CPUFREQ_VAL_UNKNOWN &&
 378             priority > sc->curr_priority) {
 379                 CF_DEBUG("saving level, freq %d prio %d\n",
 380                     sc->curr_level.total_set.freq, sc->curr_priority);
 381                 curr_freq = malloc(sizeof(*curr_freq), M_TEMP, M_NOWAIT);
 382                 if (curr_freq == NULL) {
 383                         error = ENOMEM;
 384                         goto out;
 385                 }
 386                 curr_freq->level = sc->curr_level;
 387                 curr_freq->priority = sc->curr_priority;
 388                 SLIST_INSERT_HEAD(&sc->saved_freq, curr_freq, link);
 389         }
 390         sc->curr_level = *level;
 391         sc->curr_priority = priority;
 392
 393         /* If we were restoring a saved state, reset it to "unused". */
 394         if (saved_freq != NULL) {
 395                 CF_DEBUG("resetting saved level\n");
 396                 sc->curr_level.total_set.freq = CPUFREQ_VAL_UNKNOWN;
 397                 SLIST_REMOVE_HEAD(&sc->saved_freq, link);
 398                 free(saved_freq, M_TEMP);
 399         }
 400
 401 out:
 402         CF_MTX_UNLOCK(&sc->lock);
 403
 404         /*
 405          * We changed levels (or attempted to) so notify the post-change
 406          * handler of new frequency or error.
 407          */
 408         EVENTHANDLER_INVOKE(cpufreq_post_change, level, error);
 409         if (error && set)
 410                 device_printf(set->dev, "set freq failed, err %d\n", error);
 411
 412         return (error);
 413 }
 414
 415 static int
 416 cf_get_method(device_t dev, struct cf_level *level)
 417 {
 418         struct cpufreq_softc *sc;
 419         struct cf_level *levels;
 420         struct cf_setting *curr_set, set;
 421         struct pcpu *pc;
 422         device_t *devs;
 423         int bdiff, count, diff, error, i, n, numdevs;
 424         uint64_t rate;
 425
 426         sc = device_get_softc(dev);
 427         error = 0;
 428         levels = NULL;
 429
 430         /* If we already know the current frequency, we're done. */
 431         CF_MTX_LOCK(&sc->lock);
 432         curr_set = &sc->curr_level.total_set;
 433         if (curr_set->freq != CPUFREQ_VAL_UNKNOWN) {
 434                 CF_DEBUG("get returning known freq %d\n", curr_set->freq);
 435                 goto out;
 436         }
 437         CF_MTX_UNLOCK(&sc->lock);
 438
 439         /*
 440          * We need to figure out the current level.  Loop through every
 441          * driver, getting the current setting.  Then, attempt to get a best
 442          * match of settings against each level.
 443          */
 444         count = CF_MAX_LEVELS;
 445         levels = malloc(count * sizeof(*levels), M_TEMP, M_NOWAIT);
 446         if (levels == NULL)
 447                 return (ENOMEM);
 448         error = CPUFREQ_LEVELS(sc->dev, levels, &count);
 449         if (error) {
 450                 if (error == E2BIG)
 451                         printf("cpufreq: need to increase CF_MAX_LEVELS\n");
 452                 free(levels, M_TEMP);
 453                 return (error);
 454         }
 455         error = device_get_children(device_get_parent(dev), &devs, &numdevs);
 456         if (error) {
 457                 free(levels, M_TEMP);
 458                 return (error);
 459         }
 460
 461         /*
 462          * Reacquire the lock and search for the given level.
 463          *
 464          * XXX Note: this is not quite right since we really need to go
 465          * through each level and compare both absolute and relative
 466          * settings for each driver in the system before making a match.
 467          * The estimation code below catches this case though.
 468          */
 469         CF_MTX_LOCK(&sc->lock);
 470         for (n = 0; n < numdevs && curr_set->freq == CPUFREQ_VAL_UNKNOWN; n++) {
 471                 if (!device_is_attached(devs[n]))
 472                         continue;
 473                 if (CPUFREQ_DRV_GET(devs[n], &set) != 0)
 474                         continue;
 475                 for (i = 0; i < count; i++) {
 476                         if (set.freq == levels[i].total_set.freq) {
 477                                 sc->curr_level = levels[i];
 478                                 break;
 479                         }
 480                 }
 481         }
 482         free(devs, M_TEMP);
 483         if (curr_set->freq != CPUFREQ_VAL_UNKNOWN) {
 484                 CF_DEBUG("get matched freq %d from drivers\n", curr_set->freq);
 485                 goto out;
 486         }
 487
 488         /*
 489          * We couldn't find an exact match, so attempt to estimate and then
 490          * match against a level.
 491          */
 492         pc = cpu_get_pcpu(dev);
 493         if (pc == NULL) {
 494                 error = ENXIO;
 495                 goto out;
 496         }
 497         cpu_est_clockrate(pc->pc_cpuid, &rate);
 498         rate /= 1000000;
 499         bdiff = 1 << 30;
 500         for (i = 0; i < count; i++) {
 501                 diff = abs(levels[i].total_set.freq - rate);
 502                 if (diff < bdiff) {
 503                         bdiff = diff;
 504                         sc->curr_level = levels[i];
 505                 }
 506         }
 507         CF_DEBUG("get estimated freq %d\n", curr_set->freq);
 508
 509 out:
 510         if (error == 0)
 511                 *level = sc->curr_level;
 512
 513         CF_MTX_UNLOCK(&sc->lock);
 514         if (levels)
 515                 free(levels, M_TEMP);
 516         return (error);
 517 }
 518
 519 static int
 520 cf_levels_method(device_t dev, struct cf_level *levels, int *count)
 521 {
 522         struct cf_setting_array *set_arr;
 523         struct cf_setting_lst rel_sets;
 524         struct cpufreq_softc *sc;
 525         struct cf_level *lev;
 526         struct cf_setting *sets;
 527         struct pcpu *pc;
 528         device_t *devs;
 529         int error, i, numdevs, set_count, type;
 530         uint64_t rate;
 531
 532         if (levels == NULL || count == NULL)
 533                 return (EINVAL);
 534
 535         TAILQ_INIT(&rel_sets);
 536         sc = device_get_softc(dev);
 537         error = device_get_children(device_get_parent(dev), &devs, &numdevs);
 538         if (error)
 539                 return (error);
 540         sets = malloc(MAX_SETTINGS * sizeof(*sets), M_TEMP, M_NOWAIT);
 541         if (sets == NULL) {
 542                 free(devs, M_TEMP);
 543                 return (ENOMEM);
 544         }
 545
 546         /* Get settings from all cpufreq drivers. */
 547         CF_MTX_LOCK(&sc->lock);
 548         for (i = 0; i < numdevs; i++) {
 549                 /* Skip devices that aren't ready. */
 550                 if (!device_is_attached(devs[i]))
 551                         continue;
 552
 553                 /*
 554                  * Get settings, skipping drivers that offer no settings or
 555                  * provide settings for informational purposes only.
 556                  */
 557                 error = CPUFREQ_DRV_TYPE(devs[i], &type);
 558                 if (error || (type & CPUFREQ_FLAG_INFO_ONLY)) {
 559                         if (error == 0) {
 560                                 CF_DEBUG("skipping info-only driver %s\n",
 561                                     device_get_nameunit(devs[i]));
 562                         }
 563                         continue;
 564                 }
 565                 set_count = MAX_SETTINGS;
 566                 error = CPUFREQ_DRV_SETTINGS(devs[i], sets, &set_count);
 567                 if (error || set_count == 0)
 568                         continue;
 569
 570                 /* Add the settings to our absolute/relative lists. */
 571                 switch (type & CPUFREQ_TYPE_MASK) {
 572                 case CPUFREQ_TYPE_ABSOLUTE:
 573                         error = cpufreq_insert_abs(sc, sets, set_count);
 574                         break;
 575                 case CPUFREQ_TYPE_RELATIVE:
 576                         CF_DEBUG("adding %d relative settings\n", set_count);
 577                         set_arr = malloc(sizeof(*set_arr), M_TEMP, M_NOWAIT);
 578                         if (set_arr == NULL) {
 579                                 error = ENOMEM;
 580                                 goto out;
 581                         }
 582                         bcopy(sets, set_arr->sets, set_count * sizeof(*sets));
 583                         set_arr->count = set_count;
 584                         TAILQ_INSERT_TAIL(&rel_sets, set_arr, link);
 585                         break;
 586                 default:
 587                         error = EINVAL;
 588                 }
 589                 if (error)
 590                         goto out;
 591         }
 592
 593         /*
 594          * If there are no absolute levels, create a fake one at 100%.  We
 595          * then cache the clockrate for later use as our base frequency.
 596          */
 597         if (TAILQ_EMPTY(&sc->all_levels)) {
 598                 if (sc->max_mhz == CPUFREQ_VAL_UNKNOWN) {
 599                         sc->max_mhz = cpu_get_nominal_mhz(dev);
 600                         /*
 601                          * If the CPU can't report a rate for 100%, hope
 602                          * the CPU is running at its nominal rate right now,
 603                          * and use that instead.
 604                          */
 605                         if (sc->max_mhz <= 0) {
 606                                 pc = cpu_get_pcpu(dev);
 607                                 cpu_est_clockrate(pc->pc_cpuid, &rate);
 608                                 sc->max_mhz = rate / 1000000;
 609                         }
 610                 }
 611                 memset(&sets[0], CPUFREQ_VAL_UNKNOWN, sizeof(*sets));
 612                 sets[0].freq = sc->max_mhz;
 613                 sets[0].dev = NULL;
 614                 error = cpufreq_insert_abs(sc, sets, 1);
 615                 if (error)
 616                         goto out;
 617         }
 618
 619         /* Create a combined list of absolute + relative levels. */
 620         TAILQ_FOREACH(set_arr, &rel_sets, link)
 621                 cpufreq_expand_set(sc, set_arr);
 622
 623         /* If the caller doesn't have enough space, return the actual count. */
 624         if (sc->all_count > *count) {
 625                 *count = sc->all_count;
 626                 error = E2BIG;
 627                 goto out;
 628         }
 629
 630         /* Finally, output the list of levels. */
 631         i = 0;
 632         TAILQ_FOREACH(lev, &sc->all_levels, link) {
 633
 634                 /* Skip levels that have a frequency that is too low. */
 635                 if (lev->total_set.freq < cf_lowest_freq) {
 636                         sc->all_count--;
 637                         continue;
 638                 }
 639
 640                 levels[i] = *lev;
 641                 i++;
 642         }
 643         *count = sc->all_count;
 644         error = 0;
 645
 646 out:
 647         /* Clear all levels since we regenerate them each time. */
 648         while ((lev = TAILQ_FIRST(&sc->all_levels)) != NULL) {
 649                 TAILQ_REMOVE(&sc->all_levels, lev, link);
 650                 free(lev, M_TEMP);
 651         }
 652         sc->all_count = 0;
 653
 654         CF_MTX_UNLOCK(&sc->lock);
 655         while ((set_arr = TAILQ_FIRST(&rel_sets)) != NULL) {
 656                 TAILQ_REMOVE(&rel_sets, set_arr, link);
 657                 free(set_arr, M_TEMP);
 658         }
 659         free(devs, M_TEMP);
 660         free(sets, M_TEMP);
 661         return (error);
 662 }
 663
 664 /*
 665  * Create levels for an array of absolute settings and insert them in
 666  * sorted order in the specified list.
 667  */
 668 static int
 669 cpufreq_insert_abs(struct cpufreq_softc *sc, struct cf_setting *sets,
 670     int count)
 671 {
 672         struct cf_level_lst *list;
 673         struct cf_level *level, *search;
 674         int i;
 675
 676         CF_MTX_ASSERT(&sc->lock);
 677
 678         list = &sc->all_levels;
 679         for (i = 0; i < count; i++) {
 680                 level = malloc(sizeof(*level), M_TEMP, M_NOWAIT | M_ZERO);
 681                 if (level == NULL)
 682                         return (ENOMEM);
 683                 level->abs_set = sets[i];
 684                 level->total_set = sets[i];
 685                 level->total_set.dev = NULL;
 686                 sc->all_count++;
 687
 688                 if (TAILQ_EMPTY(list)) {
 689                         CF_DEBUG("adding abs setting %d at head\n",
 690                             sets[i].freq);
 691                         TAILQ_INSERT_HEAD(list, level, link);
 692                         continue;
 693                 }
 694
 695                 TAILQ_FOREACH_REVERSE(search, list, cf_level_lst, link) {
 696                         if (sets[i].freq <= search->total_set.freq) {
 697                                 CF_DEBUG("adding abs setting %d after %d\n",
 698                                     sets[i].freq, search->total_set.freq);
 699                                 TAILQ_INSERT_AFTER(list, search, level, link);
 700                                 break;
 701                         }
 702                 }
 703         }
 704         return (0);
 705 }
 706
 707 /*
 708  * Expand a group of relative settings, creating derived levels from them.
 709  */
 710 static int
 711 cpufreq_expand_set(struct cpufreq_softc *sc, struct cf_setting_array *set_arr)
 712 {
 713         struct cf_level *fill, *search;
 714         struct cf_setting *set;
 715         int i;
 716
 717         CF_MTX_ASSERT(&sc->lock);
 718
 719         /*
 720          * Walk the set of all existing levels in reverse.  This is so we
 721          * create derived states from the lowest absolute settings first
 722          * and discard duplicates created from higher absolute settings.
 723          * For instance, a level of 50 Mhz derived from 100 Mhz + 50% is
 724          * preferable to 200 Mhz + 25% because absolute settings are more
 725          * efficient since they often change the voltage as well.
 726          */
 727         TAILQ_FOREACH_REVERSE(search, &sc->all_levels, cf_level_lst, link) {
 728                 /* Add each setting to the level, duplicating if necessary. */
 729                 for (i = 0; i < set_arr->count; i++) {
 730                         set = &set_arr->sets[i];
 731
 732                         /*
 733                          * If this setting is less than 100%, split the level
 734                          * into two and add this setting to the new level.
 735                          */
 736                         fill = search;
 737                         if (set->freq < 10000) {
 738                                 fill = cpufreq_dup_set(sc, search, set);
 739
 740                                 /*
 741                                  * The new level was a duplicate of an existing
 742                                  * level or its absolute setting is too high
 743                                  * so we freed it.  For example, we discard a
 744                                  * derived level of 1000 MHz/25% if a level
 745                                  * of 500 MHz/100% already exists.
 746                                  */
 747                                 if (fill == NULL)
 748                                         break;
 749                         }
 750
 751                         /* Add this setting to the existing or new level. */
 752                         KASSERT(fill->rel_count < MAX_SETTINGS,
 753                             ("cpufreq: too many relative drivers (%d)",
 754                             MAX_SETTINGS));
 755                         fill->rel_set[fill->rel_count] = *set;
 756                         fill->rel_count++;
 757                         CF_DEBUG(
 758                         "expand set added rel setting %d%% to %d level\n",
 759                             set->freq / 100, fill->total_set.freq);
 760                 }
 761         }
 762
 763         return (0);
 764 }
 765
 766 static struct cf_level *
 767 cpufreq_dup_set(struct cpufreq_softc *sc, struct cf_level *dup,
 768     struct cf_setting *set)
 769 {
 770         struct cf_level_lst *list;
 771         struct cf_level *fill, *itr;
 772         struct cf_setting *fill_set, *itr_set;
 773         int i;
 774
 775         CF_MTX_ASSERT(&sc->lock);
 776
 777         /*
 778          * Create a new level, copy it from the old one, and update the
 779          * total frequency and power by the percentage specified in the
 780          * relative setting.
 781          */
 782         fill = malloc(sizeof(*fill), M_TEMP, M_NOWAIT);
 783         if (fill == NULL)
 784                 return (NULL);
 785         *fill = *dup;
 786         fill_set = &fill->total_set;
 787         fill_set->freq =
 788             ((uint64_t)fill_set->freq * set->freq) / 10000;
 789         if (fill_set->power != CPUFREQ_VAL_UNKNOWN) {
 790                 fill_set->power = ((uint64_t)fill_set->power * set->freq)
 791                     / 10000;
 792         }
 793         if (set->lat != CPUFREQ_VAL_UNKNOWN) {
 794                 if (fill_set->lat != CPUFREQ_VAL_UNKNOWN)
 795                         fill_set->lat += set->lat;
 796                 else
 797                         fill_set->lat = set->lat;
 798         }
 799         CF_DEBUG("dup set considering derived setting %d\n", fill_set->freq);
 800
 801         /*
 802          * If we copied an old level that we already modified (say, at 100%),
 803          * we need to remove that setting before adding this one.  Since we
 804          * process each setting array in order, we know any settings for this
 805          * driver will be found at the end.
 806          */
 807         for (i = fill->rel_count; i != 0; i--) {
 808                 if (fill->rel_set[i - 1].dev != set->dev)
 809                         break;
 810                 CF_DEBUG("removed last relative driver: %s\n",
 811                     device_get_nameunit(set->dev));
 812                 fill->rel_count--;
 813         }
 814
 815         /*
 816          * Insert the new level in sorted order.  If it is a duplicate of an
 817          * existing level (1) or has an absolute setting higher than the
 818          * existing level (2), do not add it.  We can do this since any such
 819          * level is guaranteed use less power.  For example (1), a level with
 820          * one absolute setting of 800 Mhz uses less power than one composed
 821          * of an absolute setting of 1600 Mhz and a relative setting at 50%.
 822          * Also for example (2), a level of 800 Mhz/75% is preferable to
 823          * 1600 Mhz/25% even though the latter has a lower total frequency.
 824          */
 825         list = &sc->all_levels;
 826         KASSERT(!TAILQ_EMPTY(list), ("all levels list empty in dup set"));
 827         TAILQ_FOREACH_REVERSE(itr, list, cf_level_lst, link) {
 828                 itr_set = &itr->total_set;
 829                 if (CPUFREQ_CMP(fill_set->freq, itr_set->freq)) {
 830                         CF_DEBUG("dup set rejecting %d (dupe)\n",
 831                             fill_set->freq);
 832                         itr = NULL;
 833                         break;
 834                 } else if (fill_set->freq < itr_set->freq) {
 835                         if (fill->abs_set.freq <= itr->abs_set.freq) {
 836                                 CF_DEBUG(
 837                         "dup done, inserting new level %d after %d\n",
 838                                     fill_set->freq, itr_set->freq);
 839                                 TAILQ_INSERT_AFTER(list, itr, fill, link);
 840                                 sc->all_count++;
 841                         } else {
 842                                 CF_DEBUG("dup set rejecting %d (abs too big)\n",
 843                                     fill_set->freq);
 844                                 itr = NULL;
 845                         }
 846                         break;
 847                 }
 848         }
 849
 850         /* We didn't find a good place for this new level so free it. */
 851         if (itr == NULL) {
 852                 CF_DEBUG("dup set freeing new level %d (not optimal)\n",
 853                     fill_set->freq);
 854                 free(fill, M_TEMP);
 855                 fill = NULL;
 856         }
 857
 858         return (fill);
 859 }
 860
 861 static int
 862 cpufreq_curr_sysctl(SYSCTL_HANDLER_ARGS)
 863 {
 864         struct cpufreq_softc *sc;
 865         struct cf_level *levels;
 866         int best, count, diff, bdiff, devcount, error, freq, i, n;
 867         device_t *devs;
 868
 869         devs = NULL;
 870         sc = oidp->oid_arg1;
 871         levels = sc->levels_buf;
 872
 873         error = CPUFREQ_GET(sc->dev, &levels[0]);
 874         if (error)
 875                 goto out;
 876         freq = levels[0].total_set.freq;
 877         error = sysctl_handle_int(oidp, &freq, 0, req);
 878         if (error != 0 || req->newptr == NULL)
 879                 goto out;
 880
 881         /*
 882          * While we only call cpufreq_get() on one device (assuming all
 883          * CPUs have equal levels), we call cpufreq_set() on all CPUs.
 884          * This is needed for some MP systems.
 885          */
 886         error = devclass_get_devices(cpufreq_dc, &devs, &devcount);
 887         if (error)
 888                 goto out;
 889         for (n = 0; n < devcount; n++) {
 890                 count = CF_MAX_LEVELS;
 891                 error = CPUFREQ_LEVELS(devs[n], levels, &count);
 892                 if (error) {
 893                         if (error == E2BIG)
 894                                 printf(
 895                         "cpufreq: need to increase CF_MAX_LEVELS\n");
 896                         break;
 897                 }
 898                 best = 0;
 899                 bdiff = 1 << 30;
 900                 for (i = 0; i < count; i++) {
 901                         diff = abs(levels[i].total_set.freq - freq);
 902                         if (diff < bdiff) {
 903                                 bdiff = diff;
 904                                 best = i;
 905                         }
 906                 }
 907                 error = CPUFREQ_SET(devs[n], &levels[best], CPUFREQ_PRIO_USER);
 908         }
 909
 910 out:
 911         if (devs)
 912                 free(devs, M_TEMP);
 913         return (error);
 914 }
 915
 916 static int
 917 cpufreq_levels_sysctl(SYSCTL_HANDLER_ARGS)
 918 {
 919         struct cpufreq_softc *sc;
 920         struct cf_level *levels;
 921         struct cf_setting *set;
 922         struct sbuf sb;
 923         int count, error, i;
 924
 925         sc = oidp->oid_arg1;
 926         sbuf_new(&sb, NULL, 128, SBUF_AUTOEXTEND);
 927
 928         /* Get settings from the device and generate the output string. */
 929         count = CF_MAX_LEVELS;
 930         levels = sc->levels_buf;
 931         if (levels == NULL) {
 932                 sbuf_delete(&sb);
 933                 return (ENOMEM);
 934         }
 935         error = CPUFREQ_LEVELS(sc->dev, levels, &count);
 936         if (error) {
 937                 if (error == E2BIG)
 938                         printf("cpufreq: need to increase CF_MAX_LEVELS\n");
 939                 goto out;
 940         }
 941         if (count) {
 942                 for (i = 0; i < count; i++) {
 943                         set = &levels[i].total_set;
 944                         sbuf_printf(&sb, "%d/%d ", set->freq, set->power);
 945                 }
 946         } else
 947                 sbuf_cpy(&sb, "0");
 948         sbuf_trim(&sb);
 949         sbuf_finish(&sb);
 950         error = sysctl_handle_string(oidp, sbuf_data(&sb), sbuf_len(&sb), req);
 951
 952 out:
 953         sbuf_delete(&sb);
 954         return (error);
 955 }
 956
 957 static int
 958 cpufreq_settings_sysctl(SYSCTL_HANDLER_ARGS)
 959 {
 960         device_t dev;
 961         struct cf_setting *sets;
 962         struct sbuf sb;
 963         int error, i, set_count;
 964
 965         dev = oidp->oid_arg1;
 966         sbuf_new(&sb, NULL, 128, SBUF_AUTOEXTEND);
 967
 968         /* Get settings from the device and generate the output string. */
 969         set_count = MAX_SETTINGS;
 970         sets = malloc(set_count * sizeof(*sets), M_TEMP, M_NOWAIT);
 971         if (sets == NULL) {
 972                 sbuf_delete(&sb);
 973                 return (ENOMEM);
 974         }
 975         error = CPUFREQ_DRV_SETTINGS(dev, sets, &set_count);
 976         if (error)
 977                 goto out;
 978         if (set_count) {
 979                 for (i = 0; i < set_count; i++)
 980                         sbuf_printf(&sb, "%d/%d ", sets[i].freq, sets[i].power);
 981         } else
 982                 sbuf_cpy(&sb, "0");
 983         sbuf_trim(&sb);
 984         sbuf_finish(&sb);
 985         error = sysctl_handle_string(oidp, sbuf_data(&sb), sbuf_len(&sb), req);
 986
 987 out:
 988         free(sets, M_TEMP);
 989         sbuf_delete(&sb);
 990         return (error);
 991 }
 992
 993 int
 994 cpufreq_register(device_t dev)
 995 {
 996         struct cpufreq_softc *sc;
 997         device_t cf_dev, cpu_dev;
 998
 999         /* Add a sysctl to get each driver's settings separately. */
1000         SYSCTL_ADD_PROC(device_get_sysctl_ctx(dev),
1001             SYSCTL_CHILDREN(device_get_sysctl_tree(dev)),
1002             OID_AUTO, "freq_settings", CTLTYPE_STRING | CTLFLAG_RD, dev, 0,
1003             cpufreq_settings_sysctl, "A", "CPU frequency driver settings");
1004
1005         /*
1006          * Add only one cpufreq device to each CPU.  Currently, all CPUs
1007          * must offer the same levels and be switched at the same time.
1008          */
1009         cpu_dev = device_get_parent(dev);
1010         if ((cf_dev = device_find_child(cpu_dev, "cpufreq", -1))) {
1011                 sc = device_get_softc(cf_dev);
1012                 sc->max_mhz = CPUFREQ_VAL_UNKNOWN;
1013                 return (0);
1014         }
1015
1016         /* Add the child device and possibly sysctls. */
1017         cf_dev = BUS_ADD_CHILD(cpu_dev, 0, "cpufreq", -1);
1018         if (cf_dev == NULL)
1019                 return (ENOMEM);
1020         device_quiet(cf_dev);
1021
1022         return (device_probe_and_attach(cf_dev));
1023 }
1024
1025 int
1026 cpufreq_unregister(device_t dev)
1027 {
1028         device_t cf_dev, *devs;
1029         int cfcount, devcount, error, i, type;
1030
1031         /*
1032          * If this is the last cpufreq child device, remove the control
1033          * device as well.  We identify cpufreq children by calling a method
1034          * they support.
1035          */
1036         error = device_get_children(device_get_parent(dev), &devs, &devcount);
1037         if (error)
1038                 return (error);
1039         cf_dev = device_find_child(device_get_parent(dev), "cpufreq", -1);
1040         if (cf_dev == NULL) {
1041                 device_printf(dev,
1042         "warning: cpufreq_unregister called with no cpufreq device active\n");
1043                 free(devs, M_TEMP);
1044                 return (0);
1045         }
1046         cfcount = 0;
1047         for (i = 0; i < devcount; i++) {
1048                 if (!device_is_attached(devs[i]))
1049                         continue;
1050                 if (CPUFREQ_DRV_TYPE(devs[i], &type) == 0)
1051                         cfcount++;
1052         }
1053         if (cfcount <= 1)
1054                 device_delete_child(device_get_parent(cf_dev), cf_dev);
1055         free(devs, M_TEMP);
1056
1057         return (0);
1058 }
1059
1060 int
1061 cpufreq_settings_changed(device_t dev)
1062 {
1063
1064         EVENTHANDLER_INVOKE(cpufreq_levels_changed,
1065             device_get_unit(device_get_parent(dev)));
1066         return (0);
1067 }