sys/kern/kern_cpu.c

   1 /*-
   2  * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
   3  *
   4  * Copyright (c) 2004-2007 Nate Lawson (SDG)
   5  * All rights reserved.
   6  *
   7  * Redistribution and use in source and binary forms, with or without
   8  * modification, are permitted provided that the following conditions
   9  * are met:
  10  * 1. Redistributions of source code must retain the above copyright
  11  *    notice, this list of conditions and the following disclaimer.
  12  * 2. Redistributions in binary form must reproduce the above copyright
  13  *    notice, this list of conditions and the following disclaimer in the
  14  *    documentation and/or other materials provided with the distribution.
  15  *
  16  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
  17  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
  18  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
  19  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
  20  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
  21  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
  22  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
  23  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
  24  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
  25  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
  26  * SUCH DAMAGE.
  27  */
  28
  29 #include <sys/cdefs.h>
  30 __FBSDID("$FreeBSD$");
  31
  32 #include <sys/param.h>
  33 #include <sys/bus.h>
  34 #include <sys/cpu.h>
  35 #include <sys/eventhandler.h>
  36 #include <sys/kernel.h>
  37 #include <sys/lock.h>
  38 #include <sys/malloc.h>
  39 #include <sys/module.h>
  40 #include <sys/proc.h>
  41 #include <sys/queue.h>
  42 #include <sys/sbuf.h>
  43 #include <sys/sched.h>
  44 #include <sys/smp.h>
  45 #include <sys/sysctl.h>
  46 #include <sys/systm.h>
  47 #include <sys/sx.h>
  48 #include <sys/timetc.h>
  49 #include <sys/taskqueue.h>
  50
  51 #include "cpufreq_if.h"
  52
  53 /*
  54  * Common CPU frequency glue code.  Drivers for specific hardware can
  55  * attach this interface to allow users to get/set the CPU frequency.
  56  */
  57
  58 /*
  59  * Number of levels we can handle.  Levels are synthesized from settings
  60  * so for M settings and N drivers, there may be M*N levels.
  61  */
  62 #define CF_MAX_LEVELS   256
  63
  64 struct cf_saved_freq {
  65         struct cf_level                 level;
  66         int                             priority;
  67         SLIST_ENTRY(cf_saved_freq)      link;
  68 };
  69
  70 struct cpufreq_softc {
  71         struct sx                       lock;
  72         struct cf_level                 curr_level;
  73         int                             curr_priority;
  74         SLIST_HEAD(, cf_saved_freq)     saved_freq;
  75         struct cf_level_lst             all_levels;
  76         int                             all_count;
  77         int                             max_mhz;
  78         device_t                        dev;
  79         struct sysctl_ctx_list          sysctl_ctx;
  80         struct task                     startup_task;
  81         struct cf_level                 *levels_buf;
  82 };
  83
  84 struct cf_setting_array {
  85         struct cf_setting               sets[MAX_SETTINGS];
  86         int                             count;
  87         TAILQ_ENTRY(cf_setting_array)   link;
  88 };
  89
  90 TAILQ_HEAD(cf_setting_lst, cf_setting_array);
  91
  92 #define CF_MTX_INIT(x)          sx_init((x), "cpufreq lock")
  93 #define CF_MTX_LOCK(x)          sx_xlock((x))
  94 #define CF_MTX_UNLOCK(x)        sx_xunlock((x))
  95 #define CF_MTX_ASSERT(x)        sx_assert((x), SX_XLOCKED)
  96
  97 #define CF_DEBUG(msg...)        do {            \
  98         if (cf_verbose)                         \
  99                 printf("cpufreq: " msg);        \
 100         } while (0)
 101
 102 static int      cpufreq_attach(device_t dev);
 103 static void     cpufreq_startup_task(void *ctx, int pending);
 104 static int      cpufreq_detach(device_t dev);
 105 static int      cf_set_method(device_t dev, const struct cf_level *level,
 106                     int priority);
 107 static int      cf_get_method(device_t dev, struct cf_level *level);
 108 static int      cf_levels_method(device_t dev, struct cf_level *levels,
 109                     int *count);
 110 static int      cpufreq_insert_abs(struct cpufreq_softc *sc,
 111                     struct cf_setting *sets, int count);
 112 static int      cpufreq_expand_set(struct cpufreq_softc *sc,
 113                     struct cf_setting_array *set_arr);
 114 static struct cf_level *cpufreq_dup_set(struct cpufreq_softc *sc,
 115                     struct cf_level *dup, struct cf_setting *set);
 116 static int      cpufreq_curr_sysctl(SYSCTL_HANDLER_ARGS);
 117 static int      cpufreq_levels_sysctl(SYSCTL_HANDLER_ARGS);
 118 static int      cpufreq_settings_sysctl(SYSCTL_HANDLER_ARGS);
 119
 120 static device_method_t cpufreq_methods[] = {
 121         DEVMETHOD(device_probe,         bus_generic_probe),
 122         DEVMETHOD(device_attach,        cpufreq_attach),
 123         DEVMETHOD(device_detach,        cpufreq_detach),
 124
 125         DEVMETHOD(cpufreq_set,          cf_set_method),
 126         DEVMETHOD(cpufreq_get,          cf_get_method),
 127         DEVMETHOD(cpufreq_levels,       cf_levels_method),
 128         {0, 0}
 129 };
 130 static driver_t cpufreq_driver = {
 131         "cpufreq", cpufreq_methods, sizeof(struct cpufreq_softc)
 132 };
 133 static devclass_t cpufreq_dc;
 134 DRIVER_MODULE(cpufreq, cpu, cpufreq_driver, cpufreq_dc, 0, 0);
 135
 136 static int              cf_lowest_freq;
 137 static int              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_RWTUN, &cf_lowest_freq, 1,
 141     "Don't provide levels below this frequency.");
 142 SYSCTL_INT(_debug_cpufreq, OID_AUTO, verbose, CTLFLAG_RWTUN, &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         u_char pri;
 249
 250         sc = device_get_softc(dev);
 251         error = 0;
 252         set = NULL;
 253         saved_freq = NULL;
 254
 255         /* We are going to change levels so notify the pre-change handler. */
 256         EVENTHANDLER_INVOKE(cpufreq_pre_change, level, &error);
 257         if (error != 0) {
 258                 EVENTHANDLER_INVOKE(cpufreq_post_change, level, error);
 259                 return (error);
 260         }
 261
 262         CF_MTX_LOCK(&sc->lock);
 263
 264 #ifdef SMP
 265 #ifdef EARLY_AP_STARTUP
 266         MPASS(mp_ncpus == 1 || smp_started);
 267 #else
 268         /*
 269          * If still booting and secondary CPUs not started yet, don't allow
 270          * changing the frequency until they're online.  This is because we
 271          * can't switch to them using sched_bind() and thus we'd only be
 272          * switching the main CPU.  XXXTODO: Need to think more about how to
 273          * handle having different CPUs at different frequencies.
 274          */
 275         if (mp_ncpus > 1 && !smp_started) {
 276                 device_printf(dev, "rejecting change, SMP not started yet\n");
 277                 error = ENXIO;
 278                 goto out;
 279         }
 280 #endif
 281 #endif /* SMP */
 282
 283         /*
 284          * If the requested level has a lower priority, don't allow
 285          * the new level right now.
 286          */
 287         if (priority < sc->curr_priority) {
 288                 CF_DEBUG("ignoring, curr prio %d less than %d\n", priority,
 289                     sc->curr_priority);
 290                 error = EPERM;
 291                 goto out;
 292         }
 293
 294         /*
 295          * If the caller didn't specify a level and one is saved, prepare to
 296          * restore the saved level.  If none has been saved, return an error.
 297          */
 298         if (level == NULL) {
 299                 saved_freq = SLIST_FIRST(&sc->saved_freq);
 300                 if (saved_freq == NULL) {
 301                         CF_DEBUG("NULL level, no saved level\n");
 302                         error = ENXIO;
 303                         goto out;
 304                 }
 305                 level = &saved_freq->level;
 306                 priority = saved_freq->priority;
 307                 CF_DEBUG("restoring saved level, freq %d prio %d\n",
 308                     level->total_set.freq, priority);
 309         }
 310
 311         /* Reject levels that are below our specified threshold. */
 312         if (level->total_set.freq < cf_lowest_freq) {
 313                 CF_DEBUG("rejecting freq %d, less than %d limit\n",
 314                     level->total_set.freq, cf_lowest_freq);
 315                 error = EINVAL;
 316                 goto out;
 317         }
 318
 319         /* If already at this level, just return. */
 320         if (sc->curr_level.total_set.freq == level->total_set.freq) {
 321                 CF_DEBUG("skipping freq %d, same as current level %d\n",
 322                     level->total_set.freq, sc->curr_level.total_set.freq);
 323                 goto skip;
 324         }
 325
 326         /* First, set the absolute frequency via its driver. */
 327         set = &level->abs_set;
 328         if (set->dev) {
 329                 if (!device_is_attached(set->dev)) {
 330                         error = ENXIO;
 331                         goto out;
 332                 }
 333
 334                 /* Bind to the target CPU before switching. */
 335                 pc = cpu_get_pcpu(set->dev);
 336
 337                 /* Skip settings if CPU is not started. */
 338                 if (pc == NULL) {
 339                         error = 0;
 340                         goto out;
 341                 }
 342                 thread_lock(curthread);
 343                 pri = curthread->td_priority;
 344                 sched_prio(curthread, PRI_MIN);
 345                 sched_bind(curthread, pc->pc_cpuid);
 346                 thread_unlock(curthread);
 347                 CF_DEBUG("setting abs freq %d on %s (cpu %d)\n", set->freq,
 348                     device_get_nameunit(set->dev), PCPU_GET(cpuid));
 349                 error = CPUFREQ_DRV_SET(set->dev, set);
 350                 thread_lock(curthread);
 351                 sched_unbind(curthread);
 352                 sched_prio(curthread, pri);
 353                 thread_unlock(curthread);
 354                 if (error) {
 355                         goto out;
 356                 }
 357         }
 358
 359         /* Next, set any/all relative frequencies via their drivers. */
 360         for (i = 0; i < level->rel_count; i++) {
 361                 set = &level->rel_set[i];
 362                 if (!device_is_attached(set->dev)) {
 363                         error = ENXIO;
 364                         goto out;
 365                 }
 366
 367                 /* Bind to the target CPU before switching. */
 368                 pc = cpu_get_pcpu(set->dev);
 369                 thread_lock(curthread);
 370                 pri = curthread->td_priority;
 371                 sched_prio(curthread, PRI_MIN);
 372                 sched_bind(curthread, pc->pc_cpuid);
 373                 thread_unlock(curthread);
 374                 CF_DEBUG("setting rel freq %d on %s (cpu %d)\n", set->freq,
 375                     device_get_nameunit(set->dev), PCPU_GET(cpuid));
 376                 error = CPUFREQ_DRV_SET(set->dev, set);
 377                 thread_lock(curthread);
 378                 sched_unbind(curthread);
 379                 sched_prio(curthread, pri);
 380                 thread_unlock(curthread);
 381                 if (error) {
 382                         /* XXX Back out any successful setting? */
 383                         goto out;
 384                 }
 385         }
 386
 387 skip:
 388         /*
 389          * Before recording the current level, check if we're going to a
 390          * higher priority.  If so, save the previous level and priority.
 391          */
 392         if (sc->curr_level.total_set.freq != CPUFREQ_VAL_UNKNOWN &&
 393             priority > sc->curr_priority) {
 394                 CF_DEBUG("saving level, freq %d prio %d\n",
 395                     sc->curr_level.total_set.freq, sc->curr_priority);
 396                 curr_freq = malloc(sizeof(*curr_freq), M_TEMP, M_NOWAIT);
 397                 if (curr_freq == NULL) {
 398                         error = ENOMEM;
 399                         goto out;
 400                 }
 401                 curr_freq->level = sc->curr_level;
 402                 curr_freq->priority = sc->curr_priority;
 403                 SLIST_INSERT_HEAD(&sc->saved_freq, curr_freq, link);
 404         }
 405         sc->curr_level = *level;
 406         sc->curr_priority = priority;
 407
 408         /* If we were restoring a saved state, reset it to "unused". */
 409         if (saved_freq != NULL) {
 410                 CF_DEBUG("resetting saved level\n");
 411                 sc->curr_level.total_set.freq = CPUFREQ_VAL_UNKNOWN;
 412                 SLIST_REMOVE_HEAD(&sc->saved_freq, link);
 413                 free(saved_freq, M_TEMP);
 414         }
 415
 416 out:
 417         CF_MTX_UNLOCK(&sc->lock);
 418
 419         /*
 420          * We changed levels (or attempted to) so notify the post-change
 421          * handler of new frequency or error.
 422          */
 423         EVENTHANDLER_INVOKE(cpufreq_post_change, level, error);
 424         if (error && set)
 425                 device_printf(set->dev, "set freq failed, err %d\n", error);
 426
 427         return (error);
 428 }
 429
 430 static int
 431 cf_get_method(device_t dev, struct cf_level *level)
 432 {
 433         struct cpufreq_softc *sc;
 434         struct cf_level *levels;
 435         struct cf_setting *curr_set, set;
 436         struct pcpu *pc;
 437         device_t *devs;
 438         int bdiff, count, diff, error, i, n, numdevs;
 439         uint64_t rate;
 440
 441         sc = device_get_softc(dev);
 442         error = 0;
 443         levels = NULL;
 444
 445         /* If we already know the current frequency, we're done. */
 446         CF_MTX_LOCK(&sc->lock);
 447         curr_set = &sc->curr_level.total_set;
 448         if (curr_set->freq != CPUFREQ_VAL_UNKNOWN) {
 449                 CF_DEBUG("get returning known freq %d\n", curr_set->freq);
 450                 goto out;
 451         }
 452         CF_MTX_UNLOCK(&sc->lock);
 453
 454         /*
 455          * We need to figure out the current level.  Loop through every
 456          * driver, getting the current setting.  Then, attempt to get a best
 457          * match of settings against each level.
 458          */
 459         count = CF_MAX_LEVELS;
 460         levels = malloc(count * sizeof(*levels), M_TEMP, M_NOWAIT);
 461         if (levels == NULL)
 462                 return (ENOMEM);
 463         error = CPUFREQ_LEVELS(sc->dev, levels, &count);
 464         if (error) {
 465                 if (error == E2BIG)
 466                         printf("cpufreq: need to increase CF_MAX_LEVELS\n");
 467                 free(levels, M_TEMP);
 468                 return (error);
 469         }
 470         error = device_get_children(device_get_parent(dev), &devs, &numdevs);
 471         if (error) {
 472                 free(levels, M_TEMP);
 473                 return (error);
 474         }
 475
 476         /*
 477          * Reacquire the lock and search for the given level.
 478          *
 479          * XXX Note: this is not quite right since we really need to go
 480          * through each level and compare both absolute and relative
 481          * settings for each driver in the system before making a match.
 482          * The estimation code below catches this case though.
 483          */
 484         CF_MTX_LOCK(&sc->lock);
 485         for (n = 0; n < numdevs && curr_set->freq == CPUFREQ_VAL_UNKNOWN; n++) {
 486                 if (!device_is_attached(devs[n]))
 487                         continue;
 488                 if (CPUFREQ_DRV_GET(devs[n], &set) != 0)
 489                         continue;
 490                 for (i = 0; i < count; i++) {
 491                         if (set.freq == levels[i].total_set.freq) {
 492                                 sc->curr_level = levels[i];
 493                                 break;
 494                         }
 495                 }
 496         }
 497         free(devs, M_TEMP);
 498         if (curr_set->freq != CPUFREQ_VAL_UNKNOWN) {
 499                 CF_DEBUG("get matched freq %d from drivers\n", curr_set->freq);
 500                 goto out;
 501         }
 502
 503         /*
 504          * We couldn't find an exact match, so attempt to estimate and then
 505          * match against a level.
 506          */
 507         pc = cpu_get_pcpu(dev);
 508         if (pc == NULL) {
 509                 error = ENXIO;
 510                 goto out;
 511         }
 512         cpu_est_clockrate(pc->pc_cpuid, &rate);
 513         rate /= 1000000;
 514         bdiff = 1 << 30;
 515         for (i = 0; i < count; i++) {
 516                 diff = abs(levels[i].total_set.freq - rate);
 517                 if (diff < bdiff) {
 518                         bdiff = diff;
 519                         sc->curr_level = levels[i];
 520                 }
 521         }
 522         CF_DEBUG("get estimated freq %d\n", curr_set->freq);
 523
 524 out:
 525         if (error == 0)
 526                 *level = sc->curr_level;
 527
 528         CF_MTX_UNLOCK(&sc->lock);
 529         if (levels)
 530                 free(levels, M_TEMP);
 531         return (error);
 532 }
 533
 534 static int
 535 cf_levels_method(device_t dev, struct cf_level *levels, int *count)
 536 {
 537         struct cf_setting_array *set_arr;
 538         struct cf_setting_lst rel_sets;
 539         struct cpufreq_softc *sc;
 540         struct cf_level *lev;
 541         struct cf_setting *sets;
 542         struct pcpu *pc;
 543         device_t *devs;
 544         int error, i, numdevs, set_count, type;
 545         uint64_t rate;
 546
 547         if (levels == NULL || count == NULL)
 548                 return (EINVAL);
 549
 550         TAILQ_INIT(&rel_sets);
 551         sc = device_get_softc(dev);
 552         error = device_get_children(device_get_parent(dev), &devs, &numdevs);
 553         if (error)
 554                 return (error);
 555         sets = malloc(MAX_SETTINGS * sizeof(*sets), M_TEMP, M_NOWAIT);
 556         if (sets == NULL) {
 557                 free(devs, M_TEMP);
 558                 return (ENOMEM);
 559         }
 560
 561         /* Get settings from all cpufreq drivers. */
 562         CF_MTX_LOCK(&sc->lock);
 563         for (i = 0; i < numdevs; i++) {
 564                 /* Skip devices that aren't ready. */
 565                 if (!device_is_attached(devs[i]))
 566                         continue;
 567
 568                 /*
 569                  * Get settings, skipping drivers that offer no settings or
 570                  * provide settings for informational purposes only.
 571                  */
 572                 error = CPUFREQ_DRV_TYPE(devs[i], &type);
 573                 if (error || (type & CPUFREQ_FLAG_INFO_ONLY)) {
 574                         if (error == 0) {
 575                                 CF_DEBUG("skipping info-only driver %s\n",
 576                                     device_get_nameunit(devs[i]));
 577                         }
 578                         continue;
 579                 }
 580                 set_count = MAX_SETTINGS;
 581                 error = CPUFREQ_DRV_SETTINGS(devs[i], sets, &set_count);
 582                 if (error || set_count == 0)
 583                         continue;
 584
 585                 /* Add the settings to our absolute/relative lists. */
 586                 switch (type & CPUFREQ_TYPE_MASK) {
 587                 case CPUFREQ_TYPE_ABSOLUTE:
 588                         error = cpufreq_insert_abs(sc, sets, set_count);
 589                         break;
 590                 case CPUFREQ_TYPE_RELATIVE:
 591                         CF_DEBUG("adding %d relative settings\n", set_count);
 592                         set_arr = malloc(sizeof(*set_arr), M_TEMP, M_NOWAIT);
 593                         if (set_arr == NULL) {
 594                                 error = ENOMEM;
 595                                 goto out;
 596                         }
 597                         bcopy(sets, set_arr->sets, set_count * sizeof(*sets));
 598                         set_arr->count = set_count;
 599                         TAILQ_INSERT_TAIL(&rel_sets, set_arr, link);
 600                         break;
 601                 default:
 602                         error = EINVAL;
 603                 }
 604                 if (error)
 605                         goto out;
 606         }
 607
 608         /*
 609          * If there are no absolute levels, create a fake one at 100%.  We
 610          * then cache the clockrate for later use as our base frequency.
 611          */
 612         if (TAILQ_EMPTY(&sc->all_levels)) {
 613                 if (sc->max_mhz == CPUFREQ_VAL_UNKNOWN) {
 614                         sc->max_mhz = cpu_get_nominal_mhz(dev);
 615                         /*
 616                          * If the CPU can't report a rate for 100%, hope
 617                          * the CPU is running at its nominal rate right now,
 618                          * and use that instead.
 619                          */
 620                         if (sc->max_mhz <= 0) {
 621                                 pc = cpu_get_pcpu(dev);
 622                                 cpu_est_clockrate(pc->pc_cpuid, &rate);
 623                                 sc->max_mhz = rate / 1000000;
 624                         }
 625                 }
 626                 memset(&sets[0], CPUFREQ_VAL_UNKNOWN, sizeof(*sets));
 627                 sets[0].freq = sc->max_mhz;
 628                 sets[0].dev = NULL;
 629                 error = cpufreq_insert_abs(sc, sets, 1);
 630                 if (error)
 631                         goto out;
 632         }
 633
 634         /* Create a combined list of absolute + relative levels. */
 635         TAILQ_FOREACH(set_arr, &rel_sets, link)
 636                 cpufreq_expand_set(sc, set_arr);
 637
 638         /* If the caller doesn't have enough space, return the actual count. */
 639         if (sc->all_count > *count) {
 640                 *count = sc->all_count;
 641                 error = E2BIG;
 642                 goto out;
 643         }
 644
 645         /* Finally, output the list of levels. */
 646         i = 0;
 647         TAILQ_FOREACH(lev, &sc->all_levels, link) {
 648
 649                 /* Skip levels that have a frequency that is too low. */
 650                 if (lev->total_set.freq < cf_lowest_freq) {
 651                         sc->all_count--;
 652                         continue;
 653                 }
 654
 655                 levels[i] = *lev;
 656                 i++;
 657         }
 658         *count = sc->all_count;
 659         error = 0;
 660
 661 out:
 662         /* Clear all levels since we regenerate them each time. */
 663         while ((lev = TAILQ_FIRST(&sc->all_levels)) != NULL) {
 664                 TAILQ_REMOVE(&sc->all_levels, lev, link);
 665                 free(lev, M_TEMP);
 666         }
 667         sc->all_count = 0;
 668
 669         CF_MTX_UNLOCK(&sc->lock);
 670         while ((set_arr = TAILQ_FIRST(&rel_sets)) != NULL) {
 671                 TAILQ_REMOVE(&rel_sets, set_arr, link);
 672                 free(set_arr, M_TEMP);
 673         }
 674         free(devs, M_TEMP);
 675         free(sets, M_TEMP);
 676         return (error);
 677 }
 678
 679 /*
 680  * Create levels for an array of absolute settings and insert them in
 681  * sorted order in the specified list.
 682  */
 683 static int
 684 cpufreq_insert_abs(struct cpufreq_softc *sc, struct cf_setting *sets,
 685     int count)
 686 {
 687         struct cf_level_lst *list;
 688         struct cf_level *level, *search;
 689         int i, inserted;
 690
 691         CF_MTX_ASSERT(&sc->lock);
 692
 693         list = &sc->all_levels;
 694         for (i = 0; i < count; i++) {
 695                 level = malloc(sizeof(*level), M_TEMP, M_NOWAIT | M_ZERO);
 696                 if (level == NULL)
 697                         return (ENOMEM);
 698                 level->abs_set = sets[i];
 699                 level->total_set = sets[i];
 700                 level->total_set.dev = NULL;
 701                 sc->all_count++;
 702                 inserted = 0;
 703
 704                 if (TAILQ_EMPTY(list)) {
 705                         CF_DEBUG("adding abs setting %d at head\n",
 706                             sets[i].freq);
 707                         TAILQ_INSERT_HEAD(list, level, link);
 708                         continue;
 709                 }
 710
 711                 TAILQ_FOREACH_REVERSE(search, list, cf_level_lst, link)
 712                         if (sets[i].freq <= search->total_set.freq) {
 713                                 CF_DEBUG("adding abs setting %d after %d\n",
 714                                     sets[i].freq, search->total_set.freq);
 715                                 TAILQ_INSERT_AFTER(list, search, level, link);
 716                                 inserted = 1;
 717                                 break;
 718                         }
 719
 720                 if (inserted == 0) {
 721                         TAILQ_FOREACH(search, list, link)
 722                                 if (sets[i].freq >= search->total_set.freq) {
 723                                         CF_DEBUG("adding abs setting %d before %d\n",
 724                                             sets[i].freq, search->total_set.freq);
 725                                         TAILQ_INSERT_BEFORE(search, level, link);
 726                                         break;
 727                                 }
 728                 }
 729         }
 730
 731         return (0);
 732 }
 733
 734 /*
 735  * Expand a group of relative settings, creating derived levels from them.
 736  */
 737 static int
 738 cpufreq_expand_set(struct cpufreq_softc *sc, struct cf_setting_array *set_arr)
 739 {
 740         struct cf_level *fill, *search;
 741         struct cf_setting *set;
 742         int i;
 743
 744         CF_MTX_ASSERT(&sc->lock);
 745
 746         /*
 747          * Walk the set of all existing levels in reverse.  This is so we
 748          * create derived states from the lowest absolute settings first
 749          * and discard duplicates created from higher absolute settings.
 750          * For instance, a level of 50 Mhz derived from 100 Mhz + 50% is
 751          * preferable to 200 Mhz + 25% because absolute settings are more
 752          * efficient since they often change the voltage as well.
 753          */
 754         TAILQ_FOREACH_REVERSE(search, &sc->all_levels, cf_level_lst, link) {
 755                 /* Add each setting to the level, duplicating if necessary. */
 756                 for (i = 0; i < set_arr->count; i++) {
 757                         set = &set_arr->sets[i];
 758
 759                         /*
 760                          * If this setting is less than 100%, split the level
 761                          * into two and add this setting to the new level.
 762                          */
 763                         fill = search;
 764                         if (set->freq < 10000) {
 765                                 fill = cpufreq_dup_set(sc, search, set);
 766
 767                                 /*
 768                                  * The new level was a duplicate of an existing
 769                                  * level or its absolute setting is too high
 770                                  * so we freed it.  For example, we discard a
 771                                  * derived level of 1000 MHz/25% if a level
 772                                  * of 500 MHz/100% already exists.
 773                                  */
 774                                 if (fill == NULL)
 775                                         break;
 776                         }
 777
 778                         /* Add this setting to the existing or new level. */
 779                         KASSERT(fill->rel_count < MAX_SETTINGS,
 780                             ("cpufreq: too many relative drivers (%d)",
 781                             MAX_SETTINGS));
 782                         fill->rel_set[fill->rel_count] = *set;
 783                         fill->rel_count++;
 784                         CF_DEBUG(
 785                         "expand set added rel setting %d%% to %d level\n",
 786                             set->freq / 100, fill->total_set.freq);
 787                 }
 788         }
 789
 790         return (0);
 791 }
 792
 793 static struct cf_level *
 794 cpufreq_dup_set(struct cpufreq_softc *sc, struct cf_level *dup,
 795     struct cf_setting *set)
 796 {
 797         struct cf_level_lst *list;
 798         struct cf_level *fill, *itr;
 799         struct cf_setting *fill_set, *itr_set;
 800         int i;
 801
 802         CF_MTX_ASSERT(&sc->lock);
 803
 804         /*
 805          * Create a new level, copy it from the old one, and update the
 806          * total frequency and power by the percentage specified in the
 807          * relative setting.
 808          */
 809         fill = malloc(sizeof(*fill), M_TEMP, M_NOWAIT);
 810         if (fill == NULL)
 811                 return (NULL);
 812         *fill = *dup;
 813         fill_set = &fill->total_set;
 814         fill_set->freq =
 815             ((uint64_t)fill_set->freq * set->freq) / 10000;
 816         if (fill_set->power != CPUFREQ_VAL_UNKNOWN) {
 817                 fill_set->power = ((uint64_t)fill_set->power * set->freq)
 818                     / 10000;
 819         }
 820         if (set->lat != CPUFREQ_VAL_UNKNOWN) {
 821                 if (fill_set->lat != CPUFREQ_VAL_UNKNOWN)
 822                         fill_set->lat += set->lat;
 823                 else
 824                         fill_set->lat = set->lat;
 825         }
 826         CF_DEBUG("dup set considering derived setting %d\n", fill_set->freq);
 827
 828         /*
 829          * If we copied an old level that we already modified (say, at 100%),
 830          * we need to remove that setting before adding this one.  Since we
 831          * process each setting array in order, we know any settings for this
 832          * driver will be found at the end.
 833          */
 834         for (i = fill->rel_count; i != 0; i--) {
 835                 if (fill->rel_set[i - 1].dev != set->dev)
 836                         break;
 837                 CF_DEBUG("removed last relative driver: %s\n",
 838                     device_get_nameunit(set->dev));
 839                 fill->rel_count--;
 840         }
 841
 842         /*
 843          * Insert the new level in sorted order.  If it is a duplicate of an
 844          * existing level (1) or has an absolute setting higher than the
 845          * existing level (2), do not add it.  We can do this since any such
 846          * level is guaranteed use less power.  For example (1), a level with
 847          * one absolute setting of 800 Mhz uses less power than one composed
 848          * of an absolute setting of 1600 Mhz and a relative setting at 50%.
 849          * Also for example (2), a level of 800 Mhz/75% is preferable to
 850          * 1600 Mhz/25% even though the latter has a lower total frequency.
 851          */
 852         list = &sc->all_levels;
 853         KASSERT(!TAILQ_EMPTY(list), ("all levels list empty in dup set"));
 854         TAILQ_FOREACH_REVERSE(itr, list, cf_level_lst, link) {
 855                 itr_set = &itr->total_set;
 856                 if (CPUFREQ_CMP(fill_set->freq, itr_set->freq)) {
 857                         CF_DEBUG("dup set rejecting %d (dupe)\n",
 858                             fill_set->freq);
 859                         itr = NULL;
 860                         break;
 861                 } else if (fill_set->freq < itr_set->freq) {
 862                         if (fill->abs_set.freq <= itr->abs_set.freq) {
 863                                 CF_DEBUG(
 864                         "dup done, inserting new level %d after %d\n",
 865                                     fill_set->freq, itr_set->freq);
 866                                 TAILQ_INSERT_AFTER(list, itr, fill, link);
 867                                 sc->all_count++;
 868                         } else {
 869                                 CF_DEBUG("dup set rejecting %d (abs too big)\n",
 870                                     fill_set->freq);
 871                                 itr = NULL;
 872                         }
 873                         break;
 874                 }
 875         }
 876
 877         /* We didn't find a good place for this new level so free it. */
 878         if (itr == NULL) {
 879                 CF_DEBUG("dup set freeing new level %d (not optimal)\n",
 880                     fill_set->freq);
 881                 free(fill, M_TEMP);
 882                 fill = NULL;
 883         }
 884
 885         return (fill);
 886 }
 887
 888 static int
 889 cpufreq_curr_sysctl(SYSCTL_HANDLER_ARGS)
 890 {
 891         struct cpufreq_softc *sc;
 892         struct cf_level *levels;
 893         int best, count, diff, bdiff, devcount, error, freq, i, n;
 894         device_t *devs;
 895
 896         devs = NULL;
 897         sc = oidp->oid_arg1;
 898         levels = sc->levels_buf;
 899
 900         error = CPUFREQ_GET(sc->dev, &levels[0]);
 901         if (error)
 902                 goto out;
 903         freq = levels[0].total_set.freq;
 904         error = sysctl_handle_int(oidp, &freq, 0, req);
 905         if (error != 0 || req->newptr == NULL)
 906                 goto out;
 907
 908         /*
 909          * While we only call cpufreq_get() on one device (assuming all
 910          * CPUs have equal levels), we call cpufreq_set() on all CPUs.
 911          * This is needed for some MP systems.
 912          */
 913         error = devclass_get_devices(cpufreq_dc, &devs, &devcount);
 914         if (error)
 915                 goto out;
 916         for (n = 0; n < devcount; n++) {
 917                 count = CF_MAX_LEVELS;
 918                 error = CPUFREQ_LEVELS(devs[n], levels, &count);
 919                 if (error) {
 920                         if (error == E2BIG)
 921                                 printf(
 922                         "cpufreq: need to increase CF_MAX_LEVELS\n");
 923                         break;
 924                 }
 925                 best = 0;
 926                 bdiff = 1 << 30;
 927                 for (i = 0; i < count; i++) {
 928                         diff = abs(levels[i].total_set.freq - freq);
 929                         if (diff < bdiff) {
 930                                 bdiff = diff;
 931                                 best = i;
 932                         }
 933                 }
 934                 error = CPUFREQ_SET(devs[n], &levels[best], CPUFREQ_PRIO_USER);
 935         }
 936
 937 out:
 938         if (devs)
 939                 free(devs, M_TEMP);
 940         return (error);
 941 }
 942
 943 static int
 944 cpufreq_levels_sysctl(SYSCTL_HANDLER_ARGS)
 945 {
 946         struct cpufreq_softc *sc;
 947         struct cf_level *levels;
 948         struct cf_setting *set;
 949         struct sbuf sb;
 950         int count, error, i;
 951
 952         sc = oidp->oid_arg1;
 953         sbuf_new(&sb, NULL, 128, SBUF_AUTOEXTEND);
 954
 955         /* Get settings from the device and generate the output string. */
 956         count = CF_MAX_LEVELS;
 957         levels = sc->levels_buf;
 958         if (levels == NULL) {
 959                 sbuf_delete(&sb);
 960                 return (ENOMEM);
 961         }
 962         error = CPUFREQ_LEVELS(sc->dev, levels, &count);
 963         if (error) {
 964                 if (error == E2BIG)
 965                         printf("cpufreq: need to increase CF_MAX_LEVELS\n");
 966                 goto out;
 967         }
 968         if (count) {
 969                 for (i = 0; i < count; i++) {
 970                         set = &levels[i].total_set;
 971                         sbuf_printf(&sb, "%d/%d ", set->freq, set->power);
 972                 }
 973         } else
 974                 sbuf_cpy(&sb, "0");
 975         sbuf_trim(&sb);
 976         sbuf_finish(&sb);
 977         error = sysctl_handle_string(oidp, sbuf_data(&sb), sbuf_len(&sb), req);
 978
 979 out:
 980         sbuf_delete(&sb);
 981         return (error);
 982 }
 983
 984 static int
 985 cpufreq_settings_sysctl(SYSCTL_HANDLER_ARGS)
 986 {
 987         device_t dev;
 988         struct cf_setting *sets;
 989         struct sbuf sb;
 990         int error, i, set_count;
 991
 992         dev = oidp->oid_arg1;
 993         sbuf_new(&sb, NULL, 128, SBUF_AUTOEXTEND);
 994
 995         /* Get settings from the device and generate the output string. */
 996         set_count = MAX_SETTINGS;
 997         sets = malloc(set_count * sizeof(*sets), M_TEMP, M_NOWAIT);
 998         if (sets == NULL) {
 999                 sbuf_delete(&sb);
1000                 return (ENOMEM);
1001         }
1002         error = CPUFREQ_DRV_SETTINGS(dev, sets, &set_count);
1003         if (error)
1004                 goto out;
1005         if (set_count) {
1006                 for (i = 0; i < set_count; i++)
1007                         sbuf_printf(&sb, "%d/%d ", sets[i].freq, sets[i].power);
1008         } else
1009                 sbuf_cpy(&sb, "0");
1010         sbuf_trim(&sb);
1011         sbuf_finish(&sb);
1012         error = sysctl_handle_string(oidp, sbuf_data(&sb), sbuf_len(&sb), req);
1013
1014 out:
1015         free(sets, M_TEMP);
1016         sbuf_delete(&sb);
1017         return (error);
1018 }
1019
1020 int
1021 cpufreq_register(device_t dev)
1022 {
1023         struct cpufreq_softc *sc;
1024         device_t cf_dev, cpu_dev;
1025
1026         /* Add a sysctl to get each driver's settings separately. */
1027         SYSCTL_ADD_PROC(device_get_sysctl_ctx(dev),
1028             SYSCTL_CHILDREN(device_get_sysctl_tree(dev)),
1029             OID_AUTO, "freq_settings", CTLTYPE_STRING | CTLFLAG_RD, dev, 0,
1030             cpufreq_settings_sysctl, "A", "CPU frequency driver settings");
1031
1032         /*
1033          * Add only one cpufreq device to each CPU.  Currently, all CPUs
1034          * must offer the same levels and be switched at the same time.
1035          */
1036         cpu_dev = device_get_parent(dev);
1037         if ((cf_dev = device_find_child(cpu_dev, "cpufreq", -1))) {
1038                 sc = device_get_softc(cf_dev);
1039                 sc->max_mhz = CPUFREQ_VAL_UNKNOWN;
1040                 return (0);
1041         }
1042
1043         /* Add the child device and possibly sysctls. */
1044         cf_dev = BUS_ADD_CHILD(cpu_dev, 0, "cpufreq", -1);
1045         if (cf_dev == NULL)
1046                 return (ENOMEM);
1047         device_quiet(cf_dev);
1048
1049         return (device_probe_and_attach(cf_dev));
1050 }
1051
1052 int
1053 cpufreq_unregister(device_t dev)
1054 {
1055         device_t cf_dev, *devs;
1056         int cfcount, devcount, error, i, type;
1057
1058         /*
1059          * If this is the last cpufreq child device, remove the control
1060          * device as well.  We identify cpufreq children by calling a method
1061          * they support.
1062          */
1063         error = device_get_children(device_get_parent(dev), &devs, &devcount);
1064         if (error)
1065                 return (error);
1066         cf_dev = device_find_child(device_get_parent(dev), "cpufreq", -1);
1067         if (cf_dev == NULL) {
1068                 device_printf(dev,
1069         "warning: cpufreq_unregister called with no cpufreq device active\n");
1070                 free(devs, M_TEMP);
1071                 return (0);
1072         }
1073         cfcount = 0;
1074         for (i = 0; i < devcount; i++) {
1075                 if (!device_is_attached(devs[i]))
1076                         continue;
1077                 if (CPUFREQ_DRV_TYPE(devs[i], &type) == 0)
1078                         cfcount++;
1079         }
1080         if (cfcount <= 1)
1081                 device_delete_child(device_get_parent(cf_dev), cf_dev);
1082         free(devs, M_TEMP);
1083
1084         return (0);
1085 }
1086
1087 int
1088 cpufreq_settings_changed(device_t dev)
1089 {
1090
1091         EVENTHANDLER_INVOKE(cpufreq_levels_changed,
1092             device_get_unit(device_get_parent(dev)));
1093         return (0);
1094 }