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