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 };
  80
  81 struct cf_setting_array {
  82         struct cf_setting               sets[MAX_SETTINGS];
  83         int                             count;
  84         TAILQ_ENTRY(cf_setting_array)   link;
  85 };
  86
  87 TAILQ_HEAD(cf_setting_lst, cf_setting_array);
  88
  89 #define CF_MTX_INIT(x)          sx_init((x), "cpufreq lock")
  90 #define CF_MTX_LOCK(x)          sx_xlock((x))
  91 #define CF_MTX_UNLOCK(x)        sx_xunlock((x))
  92 #define CF_MTX_ASSERT(x)        sx_assert((x), SX_XLOCKED)
  93
  94 #define CF_DEBUG(msg...)        do {            \
  95         if (cf_verbose)                         \
  96                 printf("cpufreq: " msg);        \
  97         } while (0)
  98
  99 static int      cpufreq_attach(device_t dev);
 100 static void     cpufreq_startup_task(void *ctx, int pending);
 101 static int      cpufreq_detach(device_t dev);
 102 static int      cf_set_method(device_t dev, const struct cf_level *level,
 103                     int priority);
 104 static int      cf_get_method(device_t dev, struct cf_level *level);
 105 static int      cf_levels_method(device_t dev, struct cf_level *levels,
 106                     int *count);
 107 static int      cpufreq_insert_abs(struct cpufreq_softc *sc,
 108                     struct cf_setting *sets, int count);
 109 static int      cpufreq_expand_set(struct cpufreq_softc *sc,
 110                     struct cf_setting_array *set_arr);
 111 static struct cf_level *cpufreq_dup_set(struct cpufreq_softc *sc,
 112                     struct cf_level *dup, struct cf_setting *set);
 113 static int      cpufreq_curr_sysctl(SYSCTL_HANDLER_ARGS);
 114 static int      cpufreq_levels_sysctl(SYSCTL_HANDLER_ARGS);
 115 static int      cpufreq_settings_sysctl(SYSCTL_HANDLER_ARGS);
 116
 117 static device_method_t cpufreq_methods[] = {
 118         DEVMETHOD(device_probe,         bus_generic_probe),
 119         DEVMETHOD(device_attach,        cpufreq_attach),
 120         DEVMETHOD(device_detach,        cpufreq_detach),
 121
 122         DEVMETHOD(cpufreq_set,          cf_set_method),
 123         DEVMETHOD(cpufreq_get,          cf_get_method),
 124         DEVMETHOD(cpufreq_levels,       cf_levels_method),
 125         {0, 0}
 126 };
 127 static driver_t cpufreq_driver = {
 128         "cpufreq", cpufreq_methods, sizeof(struct cpufreq_softc)
 129 };
 130 static devclass_t cpufreq_dc;
 131 DRIVER_MODULE(cpufreq, cpu, cpufreq_driver, cpufreq_dc, 0, 0);
 132
 133 static int              cf_lowest_freq;
 134 static int              cf_verbose;
 135 TUNABLE_INT("debug.cpufreq.lowest", &cf_lowest_freq);
 136 TUNABLE_INT("debug.cpufreq.verbose", &cf_verbose);
 137 SYSCTL_NODE(_debug, OID_AUTO, cpufreq, CTLFLAG_RD, NULL, "cpufreq debugging");
 138 SYSCTL_INT(_debug_cpufreq, OID_AUTO, lowest, CTLFLAG_RW, &cf_lowest_freq, 1,
 139     "Don't provide levels below this frequency.");
 140 SYSCTL_INT(_debug_cpufreq, OID_AUTO, verbose, CTLFLAG_RW, &cf_verbose, 1,
 141     "Print verbose debugging messages");
 142
 143 static int
 144 cpufreq_attach(device_t dev)
 145 {
 146         struct cpufreq_softc *sc;
 147         struct pcpu *pc;
 148         device_t parent;
 149         uint64_t rate;
 150         int numdevs;
 151
 152         CF_DEBUG("initializing %s\n", device_get_nameunit(dev));
 153         sc = device_get_softc(dev);
 154         parent = device_get_parent(dev);
 155         sc->dev = dev;
 156         sysctl_ctx_init(&sc->sysctl_ctx);
 157         TAILQ_INIT(&sc->all_levels);
 158         CF_MTX_INIT(&sc->lock);
 159         sc->curr_level.total_set.freq = CPUFREQ_VAL_UNKNOWN;
 160         SLIST_INIT(&sc->saved_freq);
 161         /* Try to get nominal CPU freq to use it as maximum later if needed */
 162         sc->max_mhz = cpu_get_nominal_mhz(dev);
 163         /* If that fails, try to measure the current rate */
 164         if (sc->max_mhz <= 0) {
 165                 pc = cpu_get_pcpu(dev);
 166                 if (cpu_est_clockrate(pc->pc_cpuid, &rate) == 0)
 167                         sc->max_mhz = rate / 1000000;
 168                 else
 169                         sc->max_mhz = CPUFREQ_VAL_UNKNOWN;
 170         }
 171
 172         /*
 173          * Only initialize one set of sysctls for all CPUs.  In the future,
 174          * if multiple CPUs can have different settings, we can move these
 175          * sysctls to be under every CPU instead of just the first one.
 176          */
 177         numdevs = devclass_get_count(cpufreq_dc);
 178         if (numdevs > 1)
 179                 return (0);
 180
 181         CF_DEBUG("initializing one-time data for %s\n",
 182             device_get_nameunit(dev));
 183         SYSCTL_ADD_PROC(&sc->sysctl_ctx,
 184             SYSCTL_CHILDREN(device_get_sysctl_tree(parent)),
 185             OID_AUTO, "freq", CTLTYPE_INT | CTLFLAG_RW, sc, 0,
 186             cpufreq_curr_sysctl, "I", "Current CPU frequency");
 187         SYSCTL_ADD_PROC(&sc->sysctl_ctx,
 188             SYSCTL_CHILDREN(device_get_sysctl_tree(parent)),
 189             OID_AUTO, "freq_levels", CTLTYPE_STRING | CTLFLAG_RD, sc, 0,
 190             cpufreq_levels_sysctl, "A", "CPU frequency levels");
 191
 192         /*
 193          * Queue a one-shot broadcast that levels have changed.
 194          * It will run once the system has completed booting.
 195          */
 196         TASK_INIT(&sc->startup_task, 0, cpufreq_startup_task, dev);
 197         taskqueue_enqueue(taskqueue_thread, &sc->startup_task);
 198
 199         return (0);
 200 }
 201
 202 /* Handle any work to be done for all drivers that attached during boot. */
 203 static void
 204 cpufreq_startup_task(void *ctx, int pending)
 205 {
 206
 207         cpufreq_settings_changed((device_t)ctx);
 208 }
 209
 210 static int
 211 cpufreq_detach(device_t dev)
 212 {
 213         struct cpufreq_softc *sc;
 214         struct cf_saved_freq *saved_freq;
 215         int numdevs;
 216
 217         CF_DEBUG("shutdown %s\n", device_get_nameunit(dev));
 218         sc = device_get_softc(dev);
 219         sysctl_ctx_free(&sc->sysctl_ctx);
 220
 221         while ((saved_freq = SLIST_FIRST(&sc->saved_freq)) != NULL) {
 222                 SLIST_REMOVE_HEAD(&sc->saved_freq, link);
 223                 free(saved_freq, M_TEMP);
 224         }
 225
 226         /* Only clean up these resources when the last device is detaching. */
 227         numdevs = devclass_get_count(cpufreq_dc);
 228         if (numdevs == 1) {
 229                 CF_DEBUG("final shutdown for %s\n", device_get_nameunit(dev));
 230         }
 231
 232         return (0);
 233 }
 234
 235 static int
 236 cf_set_method(device_t dev, const struct cf_level *level, int priority)
 237 {
 238         struct cpufreq_softc *sc;
 239         const struct cf_setting *set;
 240         struct cf_saved_freq *saved_freq, *curr_freq;
 241         struct pcpu *pc;
 242         int error, i;
 243
 244         sc = device_get_softc(dev);
 245         error = 0;
 246         set = NULL;
 247         saved_freq = NULL;
 248
 249         /* We are going to change levels so notify the pre-change handler. */
 250         EVENTHANDLER_INVOKE(cpufreq_pre_change, level, &error);
 251         if (error != 0) {
 252                 EVENTHANDLER_INVOKE(cpufreq_post_change, level, error);
 253                 return (error);
 254         }
 255
 256         CF_MTX_LOCK(&sc->lock);
 257
 258 #ifdef SMP
 259         /*
 260          * If still booting and secondary CPUs not started yet, don't allow
 261          * changing the frequency until they're online.  This is because we
 262          * can't switch to them using sched_bind() and thus we'd only be
 263          * switching the main CPU.  XXXTODO: Need to think more about how to
 264          * handle having different CPUs at different frequencies.
 265          */
 266         if (mp_ncpus > 1 && !smp_active) {
 267                 device_printf(dev, "rejecting change, SMP not started yet\n");
 268                 error = ENXIO;
 269                 goto out;
 270         }
 271 #endif /* SMP */
 272
 273         /*
 274          * If the requested level has a lower priority, don't allow
 275          * the new level right now.
 276          */
 277         if (priority < sc->curr_priority) {
 278                 CF_DEBUG("ignoring, curr prio %d less than %d\n", priority,
 279                     sc->curr_priority);
 280                 error = EPERM;
 281                 goto out;
 282         }
 283
 284         /*
 285          * If the caller didn't specify a level and one is saved, prepare to
 286          * restore the saved level.  If none has been saved, return an error.
 287          */
 288         if (level == NULL) {
 289                 saved_freq = SLIST_FIRST(&sc->saved_freq);
 290                 if (saved_freq == NULL) {
 291                         CF_DEBUG("NULL level, no saved level\n");
 292                         error = ENXIO;
 293                         goto out;
 294                 }
 295                 level = &saved_freq->level;
 296                 priority = saved_freq->priority;
 297                 CF_DEBUG("restoring saved level, freq %d prio %d\n",
 298                     level->total_set.freq, priority);
 299         }
 300
 301         /* Reject levels that are below our specified threshold. */
 302         if (level->total_set.freq < cf_lowest_freq) {
 303                 CF_DEBUG("rejecting freq %d, less than %d limit\n",
 304                     level->total_set.freq, cf_lowest_freq);
 305                 error = EINVAL;
 306                 goto out;
 307         }
 308
 309         /* If already at this level, just return. */
 310         if (CPUFREQ_CMP(sc->curr_level.total_set.freq, level->total_set.freq)) {
 311                 CF_DEBUG("skipping freq %d, same as current level %d\n",
 312                     level->total_set.freq, sc->curr_level.total_set.freq);
 313                 goto skip;
 314         }
 315
 316         /* First, set the absolute frequency via its driver. */
 317         set = &level->abs_set;
 318         if (set->dev) {
 319                 if (!device_is_attached(set->dev)) {
 320                         error = ENXIO;
 321                         goto out;
 322                 }
 323
 324                 /* Bind to the target CPU before switching. */
 325                 pc = cpu_get_pcpu(set->dev);
 326                 thread_lock(curthread);
 327                 sched_bind(curthread, pc->pc_cpuid);
 328                 thread_unlock(curthread);
 329                 CF_DEBUG("setting abs freq %d on %s (cpu %d)\n", set->freq,
 330                     device_get_nameunit(set->dev), PCPU_GET(cpuid));
 331                 error = CPUFREQ_DRV_SET(set->dev, set);
 332                 thread_lock(curthread);
 333                 sched_unbind(curthread);
 334                 thread_unlock(curthread);
 335                 if (error) {
 336                         goto out;
 337                 }
 338         }
 339
 340         /* Next, set any/all relative frequencies via their drivers. */
 341         for (i = 0; i < level->rel_count; i++) {
 342                 set = &level->rel_set[i];
 343                 if (!device_is_attached(set->dev)) {
 344                         error = ENXIO;
 345                         goto out;
 346                 }
 347
 348                 /* Bind to the target CPU before switching. */
 349                 pc = cpu_get_pcpu(set->dev);
 350                 thread_lock(curthread);
 351                 sched_bind(curthread, pc->pc_cpuid);
 352                 thread_unlock(curthread);
 353                 CF_DEBUG("setting rel freq %d on %s (cpu %d)\n", set->freq,
 354                     device_get_nameunit(set->dev), PCPU_GET(cpuid));
 355                 error = CPUFREQ_DRV_SET(set->dev, set);
 356                 thread_lock(curthread);
 357                 sched_unbind(curthread);
 358                 thread_unlock(curthread);
 359                 if (error) {
 360                         /* XXX Back out any successful setting? */
 361                         goto out;
 362                 }
 363         }
 364
 365 skip:
 366         /*
 367          * Before recording the current level, check if we're going to a
 368          * higher priority.  If so, save the previous level and priority.
 369          */
 370         if (sc->curr_level.total_set.freq != CPUFREQ_VAL_UNKNOWN &&
 371             priority > sc->curr_priority) {
 372                 CF_DEBUG("saving level, freq %d prio %d\n",
 373                     sc->curr_level.total_set.freq, sc->curr_priority);
 374                 curr_freq = malloc(sizeof(*curr_freq), M_TEMP, M_NOWAIT);
 375                 if (curr_freq == NULL) {
 376                         error = ENOMEM;
 377                         goto out;
 378                 }
 379                 curr_freq->level = sc->curr_level;
 380                 curr_freq->priority = sc->curr_priority;
 381                 SLIST_INSERT_HEAD(&sc->saved_freq, curr_freq, link);
 382         }
 383         sc->curr_level = *level;
 384         sc->curr_priority = priority;
 385
 386         /* If we were restoring a saved state, reset it to "unused". */
 387         if (saved_freq != NULL) {
 388                 CF_DEBUG("resetting saved level\n");
 389                 sc->curr_level.total_set.freq = CPUFREQ_VAL_UNKNOWN;
 390                 SLIST_REMOVE_HEAD(&sc->saved_freq, link);
 391                 free(saved_freq, M_TEMP);
 392         }
 393
 394 out:
 395         CF_MTX_UNLOCK(&sc->lock);
 396
 397         /*
 398          * We changed levels (or attempted to) so notify the post-change
 399          * handler of new frequency or error.
 400          */
 401         EVENTHANDLER_INVOKE(cpufreq_post_change, level, error);
 402         if (error && set)
 403                 device_printf(set->dev, "set freq failed, err %d\n", error);
 404
 405         return (error);
 406 }
 407
 408 static int
 409 cf_get_method(device_t dev, struct cf_level *level)
 410 {
 411         struct cpufreq_softc *sc;
 412         struct cf_level *levels;
 413         struct cf_setting *curr_set, set;
 414         struct pcpu *pc;
 415         device_t *devs;
 416         int count, error, i, n, numdevs;
 417         uint64_t rate;
 418
 419         sc = device_get_softc(dev);
 420         error = 0;
 421         levels = NULL;
 422
 423         /* If we already know the current frequency, we're done. */
 424         CF_MTX_LOCK(&sc->lock);
 425         curr_set = &sc->curr_level.total_set;
 426         if (curr_set->freq != CPUFREQ_VAL_UNKNOWN) {
 427                 CF_DEBUG("get returning known freq %d\n", curr_set->freq);
 428                 goto out;
 429         }
 430         CF_MTX_UNLOCK(&sc->lock);
 431
 432         /*
 433          * We need to figure out the current level.  Loop through every
 434          * driver, getting the current setting.  Then, attempt to get a best
 435          * match of settings against each level.
 436          */
 437         count = CF_MAX_LEVELS;
 438         levels = malloc(count * sizeof(*levels), M_TEMP, M_NOWAIT);
 439         if (levels == NULL)
 440                 return (ENOMEM);
 441         error = CPUFREQ_LEVELS(sc->dev, levels, &count);
 442         if (error) {
 443                 if (error == E2BIG)
 444                         printf("cpufreq: need to increase CF_MAX_LEVELS\n");
 445                 free(levels, M_TEMP);
 446                 return (error);
 447         }
 448         error = device_get_children(device_get_parent(dev), &devs, &numdevs);
 449         if (error) {
 450                 free(levels, M_TEMP);
 451                 return (error);
 452         }
 453
 454         /*
 455          * Reacquire the lock and search for the given level.
 456          *
 457          * XXX Note: this is not quite right since we really need to go
 458          * through each level and compare both absolute and relative
 459          * settings for each driver in the system before making a match.
 460          * The estimation code below catches this case though.
 461          */
 462         CF_MTX_LOCK(&sc->lock);
 463         for (n = 0; n < numdevs && curr_set->freq == CPUFREQ_VAL_UNKNOWN; n++) {
 464                 if (!device_is_attached(devs[n]))
 465                         continue;
 466                 if (CPUFREQ_DRV_GET(devs[n], &set) != 0)
 467                         continue;
 468                 for (i = 0; i < count; i++) {
 469                         if (CPUFREQ_CMP(set.freq, levels[i].total_set.freq)) {
 470                                 sc->curr_level = levels[i];
 471                                 break;
 472                         }
 473                 }
 474         }
 475         free(devs, M_TEMP);
 476         if (curr_set->freq != CPUFREQ_VAL_UNKNOWN) {
 477                 CF_DEBUG("get matched freq %d from drivers\n", curr_set->freq);
 478                 goto out;
 479         }
 480
 481         /*
 482          * We couldn't find an exact match, so attempt to estimate and then
 483          * match against a level.
 484          */
 485         pc = cpu_get_pcpu(dev);
 486         if (pc == NULL) {
 487                 error = ENXIO;
 488                 goto out;
 489         }
 490         cpu_est_clockrate(pc->pc_cpuid, &rate);
 491         rate /= 1000000;
 492         for (i = 0; i < count; i++) {
 493                 if (CPUFREQ_CMP(rate, levels[i].total_set.freq)) {
 494                         sc->curr_level = levels[i];
 495                         CF_DEBUG("get estimated freq %d\n", curr_set->freq);
 496                         goto out;
 497                 }
 498         }
 499         error = ENXIO;
 500
 501 out:
 502         if (error == 0)
 503                 *level = sc->curr_level;
 504
 505         CF_MTX_UNLOCK(&sc->lock);
 506         if (levels)
 507                 free(levels, M_TEMP);
 508         return (error);
 509 }
 510
 511 static int
 512 cf_levels_method(device_t dev, struct cf_level *levels, int *count)
 513 {
 514         struct cf_setting_array *set_arr;
 515         struct cf_setting_lst rel_sets;
 516         struct cpufreq_softc *sc;
 517         struct cf_level *lev;
 518         struct cf_setting *sets;
 519         struct pcpu *pc;
 520         device_t *devs;
 521         int error, i, numdevs, set_count, type;
 522         uint64_t rate;
 523
 524         if (levels == NULL || count == NULL)
 525                 return (EINVAL);
 526
 527         TAILQ_INIT(&rel_sets);
 528         sc = device_get_softc(dev);
 529         error = device_get_children(device_get_parent(dev), &devs, &numdevs);
 530         if (error)
 531                 return (error);
 532         sets = malloc(MAX_SETTINGS * sizeof(*sets), M_TEMP, M_NOWAIT);
 533         if (sets == NULL) {
 534                 free(devs, M_TEMP);
 535                 return (ENOMEM);
 536         }
 537
 538         /* Get settings from all cpufreq drivers. */
 539         CF_MTX_LOCK(&sc->lock);
 540         for (i = 0; i < numdevs; i++) {
 541                 /* Skip devices that aren't ready. */
 542                 if (!device_is_attached(devs[i]))
 543                         continue;
 544
 545                 /*
 546                  * Get settings, skipping drivers that offer no settings or
 547                  * provide settings for informational purposes only.
 548                  */
 549                 error = CPUFREQ_DRV_TYPE(devs[i], &type);
 550                 if (error || (type & CPUFREQ_FLAG_INFO_ONLY)) {
 551                         if (error == 0) {
 552                                 CF_DEBUG("skipping info-only driver %s\n",
 553                                     device_get_nameunit(devs[i]));
 554                         }
 555                         continue;
 556                 }
 557                 set_count = MAX_SETTINGS;
 558                 error = CPUFREQ_DRV_SETTINGS(devs[i], sets, &set_count);
 559                 if (error || set_count == 0)
 560                         continue;
 561
 562                 /* Add the settings to our absolute/relative lists. */
 563                 switch (type & CPUFREQ_TYPE_MASK) {
 564                 case CPUFREQ_TYPE_ABSOLUTE:
 565                         error = cpufreq_insert_abs(sc, sets, set_count);
 566                         break;
 567                 case CPUFREQ_TYPE_RELATIVE:
 568                         CF_DEBUG("adding %d relative settings\n", set_count);
 569                         set_arr = malloc(sizeof(*set_arr), M_TEMP, M_NOWAIT);
 570                         if (set_arr == NULL) {
 571                                 error = ENOMEM;
 572                                 goto out;
 573                         }
 574                         bcopy(sets, set_arr->sets, set_count * sizeof(*sets));
 575                         set_arr->count = set_count;
 576                         TAILQ_INSERT_TAIL(&rel_sets, set_arr, link);
 577                         break;
 578                 default:
 579                         error = EINVAL;
 580                 }
 581                 if (error)
 582                         goto out;
 583         }
 584
 585         /*
 586          * If there are no absolute levels, create a fake one at 100%.  We
 587          * then cache the clockrate for later use as our base frequency.
 588          */
 589         if (TAILQ_EMPTY(&sc->all_levels)) {
 590                 if (sc->max_mhz == CPUFREQ_VAL_UNKNOWN) {
 591                         sc->max_mhz = cpu_get_nominal_mhz(dev);
 592                         /*
 593                          * If the CPU can't report a rate for 100%, hope
 594                          * the CPU is running at its nominal rate right now,
 595                          * and use that instead.
 596                          */
 597                         if (sc->max_mhz <= 0) {
 598                                 pc = cpu_get_pcpu(dev);
 599                                 cpu_est_clockrate(pc->pc_cpuid, &rate);
 600                                 sc->max_mhz = rate / 1000000;
 601                         }
 602                 }
 603                 memset(&sets[0], CPUFREQ_VAL_UNKNOWN, sizeof(*sets));
 604                 sets[0].freq = sc->max_mhz;
 605                 sets[0].dev = NULL;
 606                 error = cpufreq_insert_abs(sc, sets, 1);
 607                 if (error)
 608                         goto out;
 609         }
 610
 611         /* Create a combined list of absolute + relative levels. */
 612         TAILQ_FOREACH(set_arr, &rel_sets, link)
 613                 cpufreq_expand_set(sc, set_arr);
 614
 615         /* If the caller doesn't have enough space, return the actual count. */
 616         if (sc->all_count > *count) {
 617                 *count = sc->all_count;
 618                 error = E2BIG;
 619                 goto out;
 620         }
 621
 622         /* Finally, output the list of levels. */
 623         i = 0;
 624         TAILQ_FOREACH(lev, &sc->all_levels, link) {
 625                 /*
 626                  * Skip levels that are too close in frequency to the
 627                  * previous levels.  Some systems report bogus duplicate
 628                  * settings (i.e., for acpi_perf).
 629                  */
 630                 if (i > 0 && CPUFREQ_CMP(lev->total_set.freq,
 631                     levels[i - 1].total_set.freq)) {
 632                         sc->all_count--;
 633                         continue;
 634                 }
 635
 636                 /* Skip levels that have a frequency that is too low. */
 637                 if (lev->total_set.freq < cf_lowest_freq) {
 638                         sc->all_count--;
 639                         continue;
 640                 }
 641
 642                 levels[i] = *lev;
 643                 i++;
 644         }
 645         *count = sc->all_count;
 646         error = 0;
 647
 648 out:
 649         /* Clear all levels since we regenerate them each time. */
 650         while ((lev = TAILQ_FIRST(&sc->all_levels)) != NULL) {
 651                 TAILQ_REMOVE(&sc->all_levels, lev, link);
 652                 free(lev, M_TEMP);
 653         }
 654         sc->all_count = 0;
 655
 656         CF_MTX_UNLOCK(&sc->lock);
 657         while ((set_arr = TAILQ_FIRST(&rel_sets)) != NULL) {
 658                 TAILQ_REMOVE(&rel_sets, set_arr, link);
 659                 free(set_arr, M_TEMP);
 660         }
 661         free(devs, M_TEMP);
 662         free(sets, M_TEMP);
 663         return (error);
 664 }
 665
 666 /*
 667  * Create levels for an array of absolute settings and insert them in
 668  * sorted order in the specified list.
 669  */
 670 static int
 671 cpufreq_insert_abs(struct cpufreq_softc *sc, struct cf_setting *sets,
 672     int count)
 673 {
 674         struct cf_level_lst *list;
 675         struct cf_level *level, *search;
 676         int i;
 677
 678         CF_MTX_ASSERT(&sc->lock);
 679
 680         list = &sc->all_levels;
 681         for (i = 0; i < count; i++) {
 682                 level = malloc(sizeof(*level), M_TEMP, M_NOWAIT | M_ZERO);
 683                 if (level == NULL)
 684                         return (ENOMEM);
 685                 level->abs_set = sets[i];
 686                 level->total_set = sets[i];
 687                 level->total_set.dev = NULL;
 688                 sc->all_count++;
 689
 690                 if (TAILQ_EMPTY(list)) {
 691                         CF_DEBUG("adding abs setting %d at head\n",
 692                             sets[i].freq);
 693                         TAILQ_INSERT_HEAD(list, level, link);
 694                         continue;
 695                 }
 696
 697                 TAILQ_FOREACH_REVERSE(search, list, cf_level_lst, link) {
 698                         if (sets[i].freq <= search->total_set.freq) {
 699                                 CF_DEBUG("adding abs setting %d after %d\n",
 700                                     sets[i].freq, search->total_set.freq);
 701                                 TAILQ_INSERT_AFTER(list, search, level, link);
 702                                 break;
 703                         }
 704                 }
 705         }
 706         return (0);
 707 }
 708
 709 /*
 710  * Expand a group of relative settings, creating derived levels from them.
 711  */
 712 static int
 713 cpufreq_expand_set(struct cpufreq_softc *sc, struct cf_setting_array *set_arr)
 714 {
 715         struct cf_level *fill, *search;
 716         struct cf_setting *set;
 717         int i;
 718
 719         CF_MTX_ASSERT(&sc->lock);
 720
 721         /*
 722          * Walk the set of all existing levels in reverse.  This is so we
 723          * create derived states from the lowest absolute settings first
 724          * and discard duplicates created from higher absolute settings.
 725          * For instance, a level of 50 Mhz derived from 100 Mhz + 50% is
 726          * preferable to 200 Mhz + 25% because absolute settings are more
 727          * efficient since they often change the voltage as well.
 728          */
 729         TAILQ_FOREACH_REVERSE(search, &sc->all_levels, cf_level_lst, link) {
 730                 /* Add each setting to the level, duplicating if necessary. */
 731                 for (i = 0; i < set_arr->count; i++) {
 732                         set = &set_arr->sets[i];
 733
 734                         /*
 735                          * If this setting is less than 100%, split the level
 736                          * into two and add this setting to the new level.
 737                          */
 738                         fill = search;
 739                         if (set->freq < 10000) {
 740                                 fill = cpufreq_dup_set(sc, search, set);
 741
 742                                 /*
 743                                  * The new level was a duplicate of an existing
 744                                  * level or its absolute setting is too high
 745                                  * so we freed it.  For example, we discard a
 746                                  * derived level of 1000 MHz/25% if a level
 747                                  * of 500 MHz/100% already exists.
 748                                  */
 749                                 if (fill == NULL)
 750                                         break;
 751                         }
 752
 753                         /* Add this setting to the existing or new level. */
 754                         KASSERT(fill->rel_count < MAX_SETTINGS,
 755                             ("cpufreq: too many relative drivers (%d)",
 756                             MAX_SETTINGS));
 757                         fill->rel_set[fill->rel_count] = *set;
 758                         fill->rel_count++;
 759                         CF_DEBUG(
 760                         "expand set added rel setting %d%% to %d level\n",
 761                             set->freq / 100, fill->total_set.freq);
 762                 }
 763         }
 764
 765         return (0);
 766 }
 767
 768 static struct cf_level *
 769 cpufreq_dup_set(struct cpufreq_softc *sc, struct cf_level *dup,
 770     struct cf_setting *set)
 771 {
 772         struct cf_level_lst *list;
 773         struct cf_level *fill, *itr;
 774         struct cf_setting *fill_set, *itr_set;
 775         int i;
 776
 777         CF_MTX_ASSERT(&sc->lock);
 778
 779         /*
 780          * Create a new level, copy it from the old one, and update the
 781          * total frequency and power by the percentage specified in the
 782          * relative setting.
 783          */
 784         fill = malloc(sizeof(*fill), M_TEMP, M_NOWAIT);
 785         if (fill == NULL)
 786                 return (NULL);
 787         *fill = *dup;
 788         fill_set = &fill->total_set;
 789         fill_set->freq =
 790             ((uint64_t)fill_set->freq * set->freq) / 10000;
 791         if (fill_set->power != CPUFREQ_VAL_UNKNOWN) {
 792                 fill_set->power = ((uint64_t)fill_set->power * set->freq)
 793                     / 10000;
 794         }
 795         if (set->lat != CPUFREQ_VAL_UNKNOWN) {
 796                 if (fill_set->lat != CPUFREQ_VAL_UNKNOWN)
 797                         fill_set->lat += set->lat;
 798                 else
 799                         fill_set->lat = set->lat;
 800         }
 801         CF_DEBUG("dup set considering derived setting %d\n", fill_set->freq);
 802
 803         /*
 804          * If we copied an old level that we already modified (say, at 100%),
 805          * we need to remove that setting before adding this one.  Since we
 806          * process each setting array in order, we know any settings for this
 807          * driver will be found at the end.
 808          */
 809         for (i = fill->rel_count; i != 0; i--) {
 810                 if (fill->rel_set[i - 1].dev != set->dev)
 811                         break;
 812                 CF_DEBUG("removed last relative driver: %s\n",
 813                     device_get_nameunit(set->dev));
 814                 fill->rel_count--;
 815         }
 816
 817         /*
 818          * Insert the new level in sorted order.  If it is a duplicate of an
 819          * existing level (1) or has an absolute setting higher than the
 820          * existing level (2), do not add it.  We can do this since any such
 821          * level is guaranteed use less power.  For example (1), a level with
 822          * one absolute setting of 800 Mhz uses less power than one composed
 823          * of an absolute setting of 1600 Mhz and a relative setting at 50%.
 824          * Also for example (2), a level of 800 Mhz/75% is preferable to
 825          * 1600 Mhz/25% even though the latter has a lower total frequency.
 826          */
 827         list = &sc->all_levels;
 828         KASSERT(!TAILQ_EMPTY(list), ("all levels list empty in dup set"));
 829         TAILQ_FOREACH_REVERSE(itr, list, cf_level_lst, link) {
 830                 itr_set = &itr->total_set;
 831                 if (CPUFREQ_CMP(fill_set->freq, itr_set->freq)) {
 832                         CF_DEBUG("dup set rejecting %d (dupe)\n",
 833                             fill_set->freq);
 834                         itr = NULL;
 835                         break;
 836                 } else if (fill_set->freq < itr_set->freq) {
 837                         if (fill->abs_set.freq <= itr->abs_set.freq) {
 838                                 CF_DEBUG(
 839                         "dup done, inserting new level %d after %d\n",
 840                                     fill_set->freq, itr_set->freq);
 841                                 TAILQ_INSERT_AFTER(list, itr, fill, link);
 842                                 sc->all_count++;
 843                         } else {
 844                                 CF_DEBUG("dup set rejecting %d (abs too big)\n",
 845                                     fill_set->freq);
 846                                 itr = NULL;
 847                         }
 848                         break;
 849                 }
 850         }
 851
 852         /* We didn't find a good place for this new level so free it. */
 853         if (itr == NULL) {
 854                 CF_DEBUG("dup set freeing new level %d (not optimal)\n",
 855                     fill_set->freq);
 856                 free(fill, M_TEMP);
 857                 fill = NULL;
 858         }
 859
 860         return (fill);
 861 }
 862
 863 static int
 864 cpufreq_curr_sysctl(SYSCTL_HANDLER_ARGS)
 865 {
 866         struct cpufreq_softc *sc;
 867         struct cf_level *levels;
 868         int count, devcount, error, freq, i, n;
 869         device_t *devs;
 870
 871         devs = NULL;
 872         sc = oidp->oid_arg1;
 873         levels = malloc(CF_MAX_LEVELS * sizeof(*levels), M_TEMP, M_NOWAIT);
 874         if (levels == NULL)
 875                 return (ENOMEM);
 876
 877         error = CPUFREQ_GET(sc->dev, &levels[0]);
 878         if (error)
 879                 goto out;
 880         freq = levels[0].total_set.freq;
 881         error = sysctl_handle_int(oidp, &freq, 0, req);
 882         if (error != 0 || req->newptr == NULL)
 883                 goto out;
 884
 885         /*
 886          * While we only call cpufreq_get() on one device (assuming all
 887          * CPUs have equal levels), we call cpufreq_set() on all CPUs.
 888          * This is needed for some MP systems.
 889          */
 890         error = devclass_get_devices(cpufreq_dc, &devs, &devcount);
 891         if (error)
 892                 goto out;
 893         for (n = 0; n < devcount; n++) {
 894                 count = CF_MAX_LEVELS;
 895                 error = CPUFREQ_LEVELS(devs[n], levels, &count);
 896                 if (error) {
 897                         if (error == E2BIG)
 898                                 printf(
 899                         "cpufreq: need to increase CF_MAX_LEVELS\n");
 900                         break;
 901                 }
 902                 for (i = 0; i < count; i++) {
 903                         if (CPUFREQ_CMP(levels[i].total_set.freq, freq)) {
 904                                 error = CPUFREQ_SET(devs[n], &levels[i],
 905                                     CPUFREQ_PRIO_USER);
 906                                 break;
 907                         }
 908                 }
 909                 if (i == count) {
 910                         error = EINVAL;
 911                         break;
 912                 }
 913         }
 914
 915 out:
 916         if (devs)
 917                 free(devs, M_TEMP);
 918         if (levels)
 919                 free(levels, M_TEMP);
 920         return (error);
 921 }
 922
 923 static int
 924 cpufreq_levels_sysctl(SYSCTL_HANDLER_ARGS)
 925 {
 926         struct cpufreq_softc *sc;
 927         struct cf_level *levels;
 928         struct cf_setting *set;
 929         struct sbuf sb;
 930         int count, error, i;
 931
 932         sc = oidp->oid_arg1;
 933         sbuf_new(&sb, NULL, 128, SBUF_AUTOEXTEND);
 934
 935         /* Get settings from the device and generate the output string. */
 936         count = CF_MAX_LEVELS;
 937         levels = malloc(count * sizeof(*levels), M_TEMP, M_NOWAIT);
 938         if (levels == NULL) {
 939                 sbuf_delete(&sb);
 940                 return (ENOMEM);
 941         }
 942         error = CPUFREQ_LEVELS(sc->dev, levels, &count);
 943         if (error) {
 944                 if (error == E2BIG)
 945                         printf("cpufreq: need to increase CF_MAX_LEVELS\n");
 946                 goto out;
 947         }
 948         if (count) {
 949                 for (i = 0; i < count; i++) {
 950                         set = &levels[i].total_set;
 951                         sbuf_printf(&sb, "%d/%d ", set->freq, set->power);
 952                 }
 953         } else
 954                 sbuf_cpy(&sb, "0");
 955         sbuf_trim(&sb);
 956         sbuf_finish(&sb);
 957         error = sysctl_handle_string(oidp, sbuf_data(&sb), sbuf_len(&sb), req);
 958
 959 out:
 960         free(levels, M_TEMP);
 961         sbuf_delete(&sb);
 962         return (error);
 963 }
 964
 965 static int
 966 cpufreq_settings_sysctl(SYSCTL_HANDLER_ARGS)
 967 {
 968         device_t dev;
 969         struct cf_setting *sets;
 970         struct sbuf sb;
 971         int error, i, set_count;
 972
 973         dev = oidp->oid_arg1;
 974         sbuf_new(&sb, NULL, 128, SBUF_AUTOEXTEND);
 975
 976         /* Get settings from the device and generate the output string. */
 977         set_count = MAX_SETTINGS;
 978         sets = malloc(set_count * sizeof(*sets), M_TEMP, M_NOWAIT);
 979         if (sets == NULL) {
 980                 sbuf_delete(&sb);
 981                 return (ENOMEM);
 982         }
 983         error = CPUFREQ_DRV_SETTINGS(dev, sets, &set_count);
 984         if (error)
 985                 goto out;
 986         if (set_count) {
 987                 for (i = 0; i < set_count; i++)
 988                         sbuf_printf(&sb, "%d/%d ", sets[i].freq, sets[i].power);
 989         } else
 990                 sbuf_cpy(&sb, "0");
 991         sbuf_trim(&sb);
 992         sbuf_finish(&sb);
 993         error = sysctl_handle_string(oidp, sbuf_data(&sb), sbuf_len(&sb), req);
 994
 995 out:
 996         free(sets, M_TEMP);
 997         sbuf_delete(&sb);
 998         return (error);
 999 }
1000
1001 int
1002 cpufreq_register(device_t dev)
1003 {
1004         struct cpufreq_softc *sc;
1005         device_t cf_dev, cpu_dev;
1006
1007         /* Add a sysctl to get each driver's settings separately. */
1008         SYSCTL_ADD_PROC(device_get_sysctl_ctx(dev),
1009             SYSCTL_CHILDREN(device_get_sysctl_tree(dev)),
1010             OID_AUTO, "freq_settings", CTLTYPE_STRING | CTLFLAG_RD, dev, 0,
1011             cpufreq_settings_sysctl, "A", "CPU frequency driver settings");
1012
1013         /*
1014          * Add only one cpufreq device to each CPU.  Currently, all CPUs
1015          * must offer the same levels and be switched at the same time.
1016          */
1017         cpu_dev = device_get_parent(dev);
1018         if ((cf_dev = device_find_child(cpu_dev, "cpufreq", -1))) {
1019                 sc = device_get_softc(cf_dev);
1020                 sc->max_mhz = CPUFREQ_VAL_UNKNOWN;
1021                 return (0);
1022         }
1023
1024         /* Add the child device and possibly sysctls. */
1025         cf_dev = BUS_ADD_CHILD(cpu_dev, 0, "cpufreq", -1);
1026         if (cf_dev == NULL)
1027                 return (ENOMEM);
1028         device_quiet(cf_dev);
1029
1030         return (device_probe_and_attach(cf_dev));
1031 }
1032
1033 int
1034 cpufreq_unregister(device_t dev)
1035 {
1036         device_t cf_dev, *devs;
1037         int cfcount, devcount, error, i, type;
1038
1039         /*
1040          * If this is the last cpufreq child device, remove the control
1041          * device as well.  We identify cpufreq children by calling a method
1042          * they support.
1043          */
1044         error = device_get_children(device_get_parent(dev), &devs, &devcount);
1045         if (error)
1046                 return (error);
1047         cf_dev = device_find_child(device_get_parent(dev), "cpufreq", -1);
1048         if (cf_dev == NULL) {
1049                 device_printf(dev,
1050         "warning: cpufreq_unregister called with no cpufreq device active\n");
1051                 return (0);
1052         }
1053         cfcount = 0;
1054         for (i = 0; i < devcount; i++) {
1055                 if (!device_is_attached(devs[i]))
1056                         continue;
1057                 if (CPUFREQ_DRV_TYPE(devs[i], &type) == 0)
1058                         cfcount++;
1059         }
1060         if (cfcount <= 1)
1061                 device_delete_child(device_get_parent(cf_dev), cf_dev);
1062         free(devs, M_TEMP);
1063
1064         return (0);
1065 }
1066
1067 int
1068 cpufreq_settings_changed(device_t dev)
1069 {
1070
1071         EVENTHANDLER_INVOKE(cpufreq_levels_changed,
1072             device_get_unit(device_get_parent(dev)));
1073         return (0);
1074 }