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                 thread_lock(curthread);
 337                 pri = curthread->td_priority;
 338                 sched_prio(curthread, PRI_MIN);
 339                 sched_bind(curthread, pc->pc_cpuid);
 340                 thread_unlock(curthread);
 341                 CF_DEBUG("setting abs freq %d on %s (cpu %d)\n", set->freq,
 342                     device_get_nameunit(set->dev), PCPU_GET(cpuid));
 343                 error = CPUFREQ_DRV_SET(set->dev, set);
 344                 thread_lock(curthread);
 345                 sched_unbind(curthread);
 346                 sched_prio(curthread, pri);
 347                 thread_unlock(curthread);
 348                 if (error) {
 349                         goto out;
 350                 }
 351         }
 352
 353         /* Next, set any/all relative frequencies via their drivers. */
 354         for (i = 0; i < level->rel_count; i++) {
 355                 set = &level->rel_set[i];
 356                 if (!device_is_attached(set->dev)) {
 357                         error = ENXIO;
 358                         goto out;
 359                 }
 360
 361                 /* Bind to the target CPU before switching. */
 362                 pc = cpu_get_pcpu(set->dev);
 363                 thread_lock(curthread);
 364                 pri = curthread->td_priority;
 365                 sched_prio(curthread, PRI_MIN);
 366                 sched_bind(curthread, pc->pc_cpuid);
 367                 thread_unlock(curthread);
 368                 CF_DEBUG("setting rel freq %d on %s (cpu %d)\n", set->freq,
 369                     device_get_nameunit(set->dev), PCPU_GET(cpuid));
 370                 error = CPUFREQ_DRV_SET(set->dev, set);
 371                 thread_lock(curthread);
 372                 sched_unbind(curthread);
 373                 sched_prio(curthread, pri);
 374                 thread_unlock(curthread);
 375                 if (error) {
 376                         /* XXX Back out any successful setting? */
 377                         goto out;
 378                 }
 379         }
 380
 381 skip:
 382         /*
 383          * Before recording the current level, check if we're going to a
 384          * higher priority.  If so, save the previous level and priority.
 385          */
 386         if (sc->curr_level.total_set.freq != CPUFREQ_VAL_UNKNOWN &&
 387             priority > sc->curr_priority) {
 388                 CF_DEBUG("saving level, freq %d prio %d\n",
 389                     sc->curr_level.total_set.freq, sc->curr_priority);
 390                 curr_freq = malloc(sizeof(*curr_freq), M_TEMP, M_NOWAIT);
 391                 if (curr_freq == NULL) {
 392                         error = ENOMEM;
 393                         goto out;
 394                 }
 395                 curr_freq->level = sc->curr_level;
 396                 curr_freq->priority = sc->curr_priority;
 397                 SLIST_INSERT_HEAD(&sc->saved_freq, curr_freq, link);
 398         }
 399         sc->curr_level = *level;
 400         sc->curr_priority = priority;
 401
 402         /* If we were restoring a saved state, reset it to "unused". */
 403         if (saved_freq != NULL) {
 404                 CF_DEBUG("resetting saved level\n");
 405                 sc->curr_level.total_set.freq = CPUFREQ_VAL_UNKNOWN;
 406                 SLIST_REMOVE_HEAD(&sc->saved_freq, link);
 407                 free(saved_freq, M_TEMP);
 408         }
 409
 410 out:
 411         CF_MTX_UNLOCK(&sc->lock);
 412
 413         /*
 414          * We changed levels (or attempted to) so notify the post-change
 415          * handler of new frequency or error.
 416          */
 417         EVENTHANDLER_INVOKE(cpufreq_post_change, level, error);
 418         if (error && set)
 419                 device_printf(set->dev, "set freq failed, err %d\n", error);
 420
 421         return (error);
 422 }
 423
 424 static int
 425 cf_get_method(device_t dev, struct cf_level *level)
 426 {
 427         struct cpufreq_softc *sc;
 428         struct cf_level *levels;
 429         struct cf_setting *curr_set, set;
 430         struct pcpu *pc;
 431         device_t *devs;
 432         int bdiff, count, diff, error, i, n, numdevs;
 433         uint64_t rate;
 434
 435         sc = device_get_softc(dev);
 436         error = 0;
 437         levels = NULL;
 438
 439         /* If we already know the current frequency, we're done. */
 440         CF_MTX_LOCK(&sc->lock);
 441         curr_set = &sc->curr_level.total_set;
 442         if (curr_set->freq != CPUFREQ_VAL_UNKNOWN) {
 443                 CF_DEBUG("get returning known freq %d\n", curr_set->freq);
 444                 goto out;
 445         }
 446         CF_MTX_UNLOCK(&sc->lock);
 447
 448         /*
 449          * We need to figure out the current level.  Loop through every
 450          * driver, getting the current setting.  Then, attempt to get a best
 451          * match of settings against each level.
 452          */
 453         count = CF_MAX_LEVELS;
 454         levels = malloc(count * sizeof(*levels), M_TEMP, M_NOWAIT);
 455         if (levels == NULL)
 456                 return (ENOMEM);
 457         error = CPUFREQ_LEVELS(sc->dev, levels, &count);
 458         if (error) {
 459                 if (error == E2BIG)
 460                         printf("cpufreq: need to increase CF_MAX_LEVELS\n");
 461                 free(levels, M_TEMP);
 462                 return (error);
 463         }
 464         error = device_get_children(device_get_parent(dev), &devs, &numdevs);
 465         if (error) {
 466                 free(levels, M_TEMP);
 467                 return (error);
 468         }
 469
 470         /*
 471          * Reacquire the lock and search for the given level.
 472          *
 473          * XXX Note: this is not quite right since we really need to go
 474          * through each level and compare both absolute and relative
 475          * settings for each driver in the system before making a match.
 476          * The estimation code below catches this case though.
 477          */
 478         CF_MTX_LOCK(&sc->lock);
 479         for (n = 0; n < numdevs && curr_set->freq == CPUFREQ_VAL_UNKNOWN; n++) {
 480                 if (!device_is_attached(devs[n]))
 481                         continue;
 482                 if (CPUFREQ_DRV_GET(devs[n], &set) != 0)
 483                         continue;
 484                 for (i = 0; i < count; i++) {
 485                         if (set.freq == levels[i].total_set.freq) {
 486                                 sc->curr_level = levels[i];
 487                                 break;
 488                         }
 489                 }
 490         }
 491         free(devs, M_TEMP);
 492         if (curr_set->freq != CPUFREQ_VAL_UNKNOWN) {
 493                 CF_DEBUG("get matched freq %d from drivers\n", curr_set->freq);
 494                 goto out;
 495         }
 496
 497         /*
 498          * We couldn't find an exact match, so attempt to estimate and then
 499          * match against a level.
 500          */
 501         pc = cpu_get_pcpu(dev);
 502         if (pc == NULL) {
 503                 error = ENXIO;
 504                 goto out;
 505         }
 506         cpu_est_clockrate(pc->pc_cpuid, &rate);
 507         rate /= 1000000;
 508         bdiff = 1 << 30;
 509         for (i = 0; i < count; i++) {
 510                 diff = abs(levels[i].total_set.freq - rate);
 511                 if (diff < bdiff) {
 512                         bdiff = diff;
 513                         sc->curr_level = levels[i];
 514                 }
 515         }
 516         CF_DEBUG("get estimated freq %d\n", curr_set->freq);
 517
 518 out:
 519         if (error == 0)
 520                 *level = sc->curr_level;
 521
 522         CF_MTX_UNLOCK(&sc->lock);
 523         if (levels)
 524                 free(levels, M_TEMP);
 525         return (error);
 526 }
 527
 528 static int
 529 cf_levels_method(device_t dev, struct cf_level *levels, int *count)
 530 {
 531         struct cf_setting_array *set_arr;
 532         struct cf_setting_lst rel_sets;
 533         struct cpufreq_softc *sc;
 534         struct cf_level *lev;
 535         struct cf_setting *sets;
 536         struct pcpu *pc;
 537         device_t *devs;
 538         int error, i, numdevs, set_count, type;
 539         uint64_t rate;
 540
 541         if (levels == NULL || count == NULL)
 542                 return (EINVAL);
 543
 544         TAILQ_INIT(&rel_sets);
 545         sc = device_get_softc(dev);
 546         error = device_get_children(device_get_parent(dev), &devs, &numdevs);
 547         if (error)
 548                 return (error);
 549         sets = malloc(MAX_SETTINGS * sizeof(*sets), M_TEMP, M_NOWAIT);
 550         if (sets == NULL) {
 551                 free(devs, M_TEMP);
 552                 return (ENOMEM);
 553         }
 554
 555         /* Get settings from all cpufreq drivers. */
 556         CF_MTX_LOCK(&sc->lock);
 557         for (i = 0; i < numdevs; i++) {
 558                 /* Skip devices that aren't ready. */
 559                 if (!device_is_attached(devs[i]))
 560                         continue;
 561
 562                 /*
 563                  * Get settings, skipping drivers that offer no settings or
 564                  * provide settings for informational purposes only.
 565                  */
 566                 error = CPUFREQ_DRV_TYPE(devs[i], &type);
 567                 if (error || (type & CPUFREQ_FLAG_INFO_ONLY)) {
 568                         if (error == 0) {
 569                                 CF_DEBUG("skipping info-only driver %s\n",
 570                                     device_get_nameunit(devs[i]));
 571                         }
 572                         continue;
 573                 }
 574                 set_count = MAX_SETTINGS;
 575                 error = CPUFREQ_DRV_SETTINGS(devs[i], sets, &set_count);
 576                 if (error || set_count == 0)
 577                         continue;
 578
 579                 /* Add the settings to our absolute/relative lists. */
 580                 switch (type & CPUFREQ_TYPE_MASK) {
 581                 case CPUFREQ_TYPE_ABSOLUTE:
 582                         error = cpufreq_insert_abs(sc, sets, set_count);
 583                         break;
 584                 case CPUFREQ_TYPE_RELATIVE:
 585                         CF_DEBUG("adding %d relative settings\n", set_count);
 586                         set_arr = malloc(sizeof(*set_arr), M_TEMP, M_NOWAIT);
 587                         if (set_arr == NULL) {
 588                                 error = ENOMEM;
 589                                 goto out;
 590                         }
 591                         bcopy(sets, set_arr->sets, set_count * sizeof(*sets));
 592                         set_arr->count = set_count;
 593                         TAILQ_INSERT_TAIL(&rel_sets, set_arr, link);
 594                         break;
 595                 default:
 596                         error = EINVAL;
 597                 }
 598                 if (error)
 599                         goto out;
 600         }
 601
 602         /*
 603          * If there are no absolute levels, create a fake one at 100%.  We
 604          * then cache the clockrate for later use as our base frequency.
 605          */
 606         if (TAILQ_EMPTY(&sc->all_levels)) {
 607                 if (sc->max_mhz == CPUFREQ_VAL_UNKNOWN) {
 608                         sc->max_mhz = cpu_get_nominal_mhz(dev);
 609                         /*
 610                          * If the CPU can't report a rate for 100%, hope
 611                          * the CPU is running at its nominal rate right now,
 612                          * and use that instead.
 613                          */
 614                         if (sc->max_mhz <= 0) {
 615                                 pc = cpu_get_pcpu(dev);
 616                                 cpu_est_clockrate(pc->pc_cpuid, &rate);
 617                                 sc->max_mhz = rate / 1000000;
 618                         }
 619                 }
 620                 memset(&sets[0], CPUFREQ_VAL_UNKNOWN, sizeof(*sets));
 621                 sets[0].freq = sc->max_mhz;
 622                 sets[0].dev = NULL;
 623                 error = cpufreq_insert_abs(sc, sets, 1);
 624                 if (error)
 625                         goto out;
 626         }
 627
 628         /* Create a combined list of absolute + relative levels. */
 629         TAILQ_FOREACH(set_arr, &rel_sets, link)
 630                 cpufreq_expand_set(sc, set_arr);
 631
 632         /* If the caller doesn't have enough space, return the actual count. */
 633         if (sc->all_count > *count) {
 634                 *count = sc->all_count;
 635                 error = E2BIG;
 636                 goto out;
 637         }
 638
 639         /* Finally, output the list of levels. */
 640         i = 0;
 641         TAILQ_FOREACH(lev, &sc->all_levels, link) {
 642
 643                 /* Skip levels that have a frequency that is too low. */
 644                 if (lev->total_set.freq < cf_lowest_freq) {
 645                         sc->all_count--;
 646                         continue;
 647                 }
 648
 649                 levels[i] = *lev;
 650                 i++;
 651         }
 652         *count = sc->all_count;
 653         error = 0;
 654
 655 out:
 656         /* Clear all levels since we regenerate them each time. */
 657         while ((lev = TAILQ_FIRST(&sc->all_levels)) != NULL) {
 658                 TAILQ_REMOVE(&sc->all_levels, lev, link);
 659                 free(lev, M_TEMP);
 660         }
 661         sc->all_count = 0;
 662
 663         CF_MTX_UNLOCK(&sc->lock);
 664         while ((set_arr = TAILQ_FIRST(&rel_sets)) != NULL) {
 665                 TAILQ_REMOVE(&rel_sets, set_arr, link);
 666                 free(set_arr, M_TEMP);
 667         }
 668         free(devs, M_TEMP);
 669         free(sets, M_TEMP);
 670         return (error);
 671 }
 672
 673 /*
 674  * Create levels for an array of absolute settings and insert them in
 675  * sorted order in the specified list.
 676  */
 677 static int
 678 cpufreq_insert_abs(struct cpufreq_softc *sc, struct cf_setting *sets,
 679     int count)
 680 {
 681         struct cf_level_lst *list;
 682         struct cf_level *level, *search;
 683         int i, inserted;
 684
 685         CF_MTX_ASSERT(&sc->lock);
 686
 687         list = &sc->all_levels;
 688         for (i = 0; i < count; i++) {
 689                 level = malloc(sizeof(*level), M_TEMP, M_NOWAIT | M_ZERO);
 690                 if (level == NULL)
 691                         return (ENOMEM);
 692                 level->abs_set = sets[i];
 693                 level->total_set = sets[i];
 694                 level->total_set.dev = NULL;
 695                 sc->all_count++;
 696                 inserted = 0;
 697
 698                 if (TAILQ_EMPTY(list)) {
 699                         CF_DEBUG("adding abs setting %d at head\n",
 700                             sets[i].freq);
 701                         TAILQ_INSERT_HEAD(list, level, link);
 702                         continue;
 703                 }
 704
 705                 TAILQ_FOREACH_REVERSE(search, list, cf_level_lst, link)
 706                         if (sets[i].freq <= search->total_set.freq) {
 707                                 CF_DEBUG("adding abs setting %d after %d\n",
 708                                     sets[i].freq, search->total_set.freq);
 709                                 TAILQ_INSERT_AFTER(list, search, level, link);
 710                                 inserted = 1;
 711                                 break;
 712                         }
 713
 714                 if (inserted == 0) {
 715                         TAILQ_FOREACH(search, list, link)
 716                                 if (sets[i].freq >= search->total_set.freq) {
 717                                         CF_DEBUG("adding abs setting %d before %d\n",
 718                                             sets[i].freq, search->total_set.freq);
 719                                         TAILQ_INSERT_BEFORE(search, level, link);
 720                                         break;
 721                                 }
 722                 }
 723         }
 724
 725         return (0);
 726 }
 727
 728 /*
 729  * Expand a group of relative settings, creating derived levels from them.
 730  */
 731 static int
 732 cpufreq_expand_set(struct cpufreq_softc *sc, struct cf_setting_array *set_arr)
 733 {
 734         struct cf_level *fill, *search;
 735         struct cf_setting *set;
 736         int i;
 737
 738         CF_MTX_ASSERT(&sc->lock);
 739
 740         /*
 741          * Walk the set of all existing levels in reverse.  This is so we
 742          * create derived states from the lowest absolute settings first
 743          * and discard duplicates created from higher absolute settings.
 744          * For instance, a level of 50 Mhz derived from 100 Mhz + 50% is
 745          * preferable to 200 Mhz + 25% because absolute settings are more
 746          * efficient since they often change the voltage as well.
 747          */
 748         TAILQ_FOREACH_REVERSE(search, &sc->all_levels, cf_level_lst, link) {
 749                 /* Add each setting to the level, duplicating if necessary. */
 750                 for (i = 0; i < set_arr->count; i++) {
 751                         set = &set_arr->sets[i];
 752
 753                         /*
 754                          * If this setting is less than 100%, split the level
 755                          * into two and add this setting to the new level.
 756                          */
 757                         fill = search;
 758                         if (set->freq < 10000) {
 759                                 fill = cpufreq_dup_set(sc, search, set);
 760
 761                                 /*
 762                                  * The new level was a duplicate of an existing
 763                                  * level or its absolute setting is too high
 764                                  * so we freed it.  For example, we discard a
 765                                  * derived level of 1000 MHz/25% if a level
 766                                  * of 500 MHz/100% already exists.
 767                                  */
 768                                 if (fill == NULL)
 769                                         break;
 770                         }
 771
 772                         /* Add this setting to the existing or new level. */
 773                         KASSERT(fill->rel_count < MAX_SETTINGS,
 774                             ("cpufreq: too many relative drivers (%d)",
 775                             MAX_SETTINGS));
 776                         fill->rel_set[fill->rel_count] = *set;
 777                         fill->rel_count++;
 778                         CF_DEBUG(
 779                         "expand set added rel setting %d%% to %d level\n",
 780                             set->freq / 100, fill->total_set.freq);
 781                 }
 782         }
 783
 784         return (0);
 785 }
 786
 787 static struct cf_level *
 788 cpufreq_dup_set(struct cpufreq_softc *sc, struct cf_level *dup,
 789     struct cf_setting *set)
 790 {
 791         struct cf_level_lst *list;
 792         struct cf_level *fill, *itr;
 793         struct cf_setting *fill_set, *itr_set;
 794         int i;
 795
 796         CF_MTX_ASSERT(&sc->lock);
 797
 798         /*
 799          * Create a new level, copy it from the old one, and update the
 800          * total frequency and power by the percentage specified in the
 801          * relative setting.
 802          */
 803         fill = malloc(sizeof(*fill), M_TEMP, M_NOWAIT);
 804         if (fill == NULL)
 805                 return (NULL);
 806         *fill = *dup;
 807         fill_set = &fill->total_set;
 808         fill_set->freq =
 809             ((uint64_t)fill_set->freq * set->freq) / 10000;
 810         if (fill_set->power != CPUFREQ_VAL_UNKNOWN) {
 811                 fill_set->power = ((uint64_t)fill_set->power * set->freq)
 812                     / 10000;
 813         }
 814         if (set->lat != CPUFREQ_VAL_UNKNOWN) {
 815                 if (fill_set->lat != CPUFREQ_VAL_UNKNOWN)
 816                         fill_set->lat += set->lat;
 817                 else
 818                         fill_set->lat = set->lat;
 819         }
 820         CF_DEBUG("dup set considering derived setting %d\n", fill_set->freq);
 821
 822         /*
 823          * If we copied an old level that we already modified (say, at 100%),
 824          * we need to remove that setting before adding this one.  Since we
 825          * process each setting array in order, we know any settings for this
 826          * driver will be found at the end.
 827          */
 828         for (i = fill->rel_count; i != 0; i--) {
 829                 if (fill->rel_set[i - 1].dev != set->dev)
 830                         break;
 831                 CF_DEBUG("removed last relative driver: %s\n",
 832                     device_get_nameunit(set->dev));
 833                 fill->rel_count--;
 834         }
 835
 836         /*
 837          * Insert the new level in sorted order.  If it is a duplicate of an
 838          * existing level (1) or has an absolute setting higher than the
 839          * existing level (2), do not add it.  We can do this since any such
 840          * level is guaranteed use less power.  For example (1), a level with
 841          * one absolute setting of 800 Mhz uses less power than one composed
 842          * of an absolute setting of 1600 Mhz and a relative setting at 50%.
 843          * Also for example (2), a level of 800 Mhz/75% is preferable to
 844          * 1600 Mhz/25% even though the latter has a lower total frequency.
 845          */
 846         list = &sc->all_levels;
 847         KASSERT(!TAILQ_EMPTY(list), ("all levels list empty in dup set"));
 848         TAILQ_FOREACH_REVERSE(itr, list, cf_level_lst, link) {
 849                 itr_set = &itr->total_set;
 850                 if (CPUFREQ_CMP(fill_set->freq, itr_set->freq)) {
 851                         CF_DEBUG("dup set rejecting %d (dupe)\n",
 852                             fill_set->freq);
 853                         itr = NULL;
 854                         break;
 855                 } else if (fill_set->freq < itr_set->freq) {
 856                         if (fill->abs_set.freq <= itr->abs_set.freq) {
 857                                 CF_DEBUG(
 858                         "dup done, inserting new level %d after %d\n",
 859                                     fill_set->freq, itr_set->freq);
 860                                 TAILQ_INSERT_AFTER(list, itr, fill, link);
 861                                 sc->all_count++;
 862                         } else {
 863                                 CF_DEBUG("dup set rejecting %d (abs too big)\n",
 864                                     fill_set->freq);
 865                                 itr = NULL;
 866                         }
 867                         break;
 868                 }
 869         }
 870
 871         /* We didn't find a good place for this new level so free it. */
 872         if (itr == NULL) {
 873                 CF_DEBUG("dup set freeing new level %d (not optimal)\n",
 874                     fill_set->freq);
 875                 free(fill, M_TEMP);
 876                 fill = NULL;
 877         }
 878
 879         return (fill);
 880 }
 881
 882 static int
 883 cpufreq_curr_sysctl(SYSCTL_HANDLER_ARGS)
 884 {
 885         struct cpufreq_softc *sc;
 886         struct cf_level *levels;
 887         int best, count, diff, bdiff, devcount, error, freq, i, n;
 888         device_t *devs;
 889
 890         devs = NULL;
 891         sc = oidp->oid_arg1;
 892         levels = sc->levels_buf;
 893
 894         error = CPUFREQ_GET(sc->dev, &levels[0]);
 895         if (error)
 896                 goto out;
 897         freq = levels[0].total_set.freq;
 898         error = sysctl_handle_int(oidp, &freq, 0, req);
 899         if (error != 0 || req->newptr == NULL)
 900                 goto out;
 901
 902         /*
 903          * While we only call cpufreq_get() on one device (assuming all
 904          * CPUs have equal levels), we call cpufreq_set() on all CPUs.
 905          * This is needed for some MP systems.
 906          */
 907         error = devclass_get_devices(cpufreq_dc, &devs, &devcount);
 908         if (error)
 909                 goto out;
 910         for (n = 0; n < devcount; n++) {
 911                 count = CF_MAX_LEVELS;
 912                 error = CPUFREQ_LEVELS(devs[n], levels, &count);
 913                 if (error) {
 914                         if (error == E2BIG)
 915                                 printf(
 916                         "cpufreq: need to increase CF_MAX_LEVELS\n");
 917                         break;
 918                 }
 919                 best = 0;
 920                 bdiff = 1 << 30;
 921                 for (i = 0; i < count; i++) {
 922                         diff = abs(levels[i].total_set.freq - freq);
 923                         if (diff < bdiff) {
 924                                 bdiff = diff;
 925                                 best = i;
 926                         }
 927                 }
 928                 error = CPUFREQ_SET(devs[n], &levels[best], CPUFREQ_PRIO_USER);
 929         }
 930
 931 out:
 932         if (devs)
 933                 free(devs, M_TEMP);
 934         return (error);
 935 }
 936
 937 static int
 938 cpufreq_levels_sysctl(SYSCTL_HANDLER_ARGS)
 939 {
 940         struct cpufreq_softc *sc;
 941         struct cf_level *levels;
 942         struct cf_setting *set;
 943         struct sbuf sb;
 944         int count, error, i;
 945
 946         sc = oidp->oid_arg1;
 947         sbuf_new(&sb, NULL, 128, SBUF_AUTOEXTEND);
 948
 949         /* Get settings from the device and generate the output string. */
 950         count = CF_MAX_LEVELS;
 951         levels = sc->levels_buf;
 952         if (levels == NULL) {
 953                 sbuf_delete(&sb);
 954                 return (ENOMEM);
 955         }
 956         error = CPUFREQ_LEVELS(sc->dev, levels, &count);
 957         if (error) {
 958                 if (error == E2BIG)
 959                         printf("cpufreq: need to increase CF_MAX_LEVELS\n");
 960                 goto out;
 961         }
 962         if (count) {
 963                 for (i = 0; i < count; i++) {
 964                         set = &levels[i].total_set;
 965                         sbuf_printf(&sb, "%d/%d ", set->freq, set->power);
 966                 }
 967         } else
 968                 sbuf_cpy(&sb, "0");
 969         sbuf_trim(&sb);
 970         sbuf_finish(&sb);
 971         error = sysctl_handle_string(oidp, sbuf_data(&sb), sbuf_len(&sb), req);
 972
 973 out:
 974         sbuf_delete(&sb);
 975         return (error);
 976 }
 977
 978 static int
 979 cpufreq_settings_sysctl(SYSCTL_HANDLER_ARGS)
 980 {
 981         device_t dev;
 982         struct cf_setting *sets;
 983         struct sbuf sb;
 984         int error, i, set_count;
 985
 986         dev = oidp->oid_arg1;
 987         sbuf_new(&sb, NULL, 128, SBUF_AUTOEXTEND);
 988
 989         /* Get settings from the device and generate the output string. */
 990         set_count = MAX_SETTINGS;
 991         sets = malloc(set_count * sizeof(*sets), M_TEMP, M_NOWAIT);
 992         if (sets == NULL) {
 993                 sbuf_delete(&sb);
 994                 return (ENOMEM);
 995         }
 996         error = CPUFREQ_DRV_SETTINGS(dev, sets, &set_count);
 997         if (error)
 998                 goto out;
 999         if (set_count) {
1000                 for (i = 0; i < set_count; i++)
1001                         sbuf_printf(&sb, "%d/%d ", sets[i].freq, sets[i].power);
1002         } else
1003                 sbuf_cpy(&sb, "0");
1004         sbuf_trim(&sb);
1005         sbuf_finish(&sb);
1006         error = sysctl_handle_string(oidp, sbuf_data(&sb), sbuf_len(&sb), req);
1007
1008 out:
1009         free(sets, M_TEMP);
1010         sbuf_delete(&sb);
1011         return (error);
1012 }
1013
1014 int
1015 cpufreq_register(device_t dev)
1016 {
1017         struct cpufreq_softc *sc;
1018         device_t cf_dev, cpu_dev;
1019
1020         /* Add a sysctl to get each driver's settings separately. */
1021         SYSCTL_ADD_PROC(device_get_sysctl_ctx(dev),
1022             SYSCTL_CHILDREN(device_get_sysctl_tree(dev)),
1023             OID_AUTO, "freq_settings", CTLTYPE_STRING | CTLFLAG_RD, dev, 0,
1024             cpufreq_settings_sysctl, "A", "CPU frequency driver settings");
1025
1026         /*
1027          * Add only one cpufreq device to each CPU.  Currently, all CPUs
1028          * must offer the same levels and be switched at the same time.
1029          */
1030         cpu_dev = device_get_parent(dev);
1031         if ((cf_dev = device_find_child(cpu_dev, "cpufreq", -1))) {
1032                 sc = device_get_softc(cf_dev);
1033                 sc->max_mhz = CPUFREQ_VAL_UNKNOWN;
1034                 return (0);
1035         }
1036
1037         /* Add the child device and possibly sysctls. */
1038         cf_dev = BUS_ADD_CHILD(cpu_dev, 0, "cpufreq", -1);
1039         if (cf_dev == NULL)
1040                 return (ENOMEM);
1041         device_quiet(cf_dev);
1042
1043         return (device_probe_and_attach(cf_dev));
1044 }
1045
1046 int
1047 cpufreq_unregister(device_t dev)
1048 {
1049         device_t cf_dev, *devs;
1050         int cfcount, devcount, error, i, type;
1051
1052         /*
1053          * If this is the last cpufreq child device, remove the control
1054          * device as well.  We identify cpufreq children by calling a method
1055          * they support.
1056          */
1057         error = device_get_children(device_get_parent(dev), &devs, &devcount);
1058         if (error)
1059                 return (error);
1060         cf_dev = device_find_child(device_get_parent(dev), "cpufreq", -1);
1061         if (cf_dev == NULL) {
1062                 device_printf(dev,
1063         "warning: cpufreq_unregister called with no cpufreq device active\n");
1064                 free(devs, M_TEMP);
1065                 return (0);
1066         }
1067         cfcount = 0;
1068         for (i = 0; i < devcount; i++) {
1069                 if (!device_is_attached(devs[i]))
1070                         continue;
1071                 if (CPUFREQ_DRV_TYPE(devs[i], &type) == 0)
1072                         cfcount++;
1073         }
1074         if (cfcount <= 1)
1075                 device_delete_child(device_get_parent(cf_dev), cf_dev);
1076         free(devs, M_TEMP);
1077
1078         return (0);
1079 }
1080
1081 int
1082 cpufreq_settings_changed(device_t dev)
1083 {
1084
1085         EVENTHANDLER_INVOKE(cpufreq_levels_changed,
1086             device_get_unit(device_get_parent(dev)));
1087         return (0);
1088 }