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         device_t                        cf_drv_dev;
  80         struct sysctl_ctx_list          sysctl_ctx;
  81         struct task                     startup_task;
  82         struct cf_level                 *levels_buf;
  83 };
  84
  85 struct cf_setting_array {
  86         struct cf_setting               sets[MAX_SETTINGS];
  87         int                             count;
  88         TAILQ_ENTRY(cf_setting_array)   link;
  89 };
  90
  91 TAILQ_HEAD(cf_setting_lst, cf_setting_array);
  92
  93 #define CF_MTX_INIT(x)          sx_init((x), "cpufreq lock")
  94 #define CF_MTX_LOCK(x)          sx_xlock((x))
  95 #define CF_MTX_UNLOCK(x)        sx_xunlock((x))
  96 #define CF_MTX_ASSERT(x)        sx_assert((x), SX_XLOCKED)
  97
  98 #define CF_DEBUG(msg...)        do {            \
  99         if (cf_verbose)                         \
 100                 printf("cpufreq: " msg);        \
 101         } while (0)
 102
 103 static int      cpufreq_attach(device_t dev);
 104 static void     cpufreq_startup_task(void *ctx, int pending);
 105 static int      cpufreq_detach(device_t dev);
 106 static int      cf_set_method(device_t dev, const struct cf_level *level,
 107                     int priority);
 108 static int      cf_get_method(device_t dev, struct cf_level *level);
 109 static int      cf_levels_method(device_t dev, struct cf_level *levels,
 110                     int *count);
 111 static int      cpufreq_insert_abs(struct cpufreq_softc *sc,
 112                     struct cf_setting *sets, int count);
 113 static int      cpufreq_expand_set(struct cpufreq_softc *sc,
 114                     struct cf_setting_array *set_arr);
 115 static struct cf_level *cpufreq_dup_set(struct cpufreq_softc *sc,
 116                     struct cf_level *dup, struct cf_setting *set);
 117 static int      cpufreq_curr_sysctl(SYSCTL_HANDLER_ARGS);
 118 static int      cpufreq_levels_sysctl(SYSCTL_HANDLER_ARGS);
 119 static int      cpufreq_settings_sysctl(SYSCTL_HANDLER_ARGS);
 120
 121 static device_method_t cpufreq_methods[] = {
 122         DEVMETHOD(device_probe,         bus_generic_probe),
 123         DEVMETHOD(device_attach,        cpufreq_attach),
 124         DEVMETHOD(device_detach,        cpufreq_detach),
 125
 126         DEVMETHOD(cpufreq_set,          cf_set_method),
 127         DEVMETHOD(cpufreq_get,          cf_get_method),
 128         DEVMETHOD(cpufreq_levels,       cf_levels_method),
 129         {0, 0}
 130 };
 131 static driver_t cpufreq_driver = {
 132         "cpufreq", cpufreq_methods, sizeof(struct cpufreq_softc)
 133 };
 134 static devclass_t cpufreq_dc;
 135 DRIVER_MODULE(cpufreq, cpu, cpufreq_driver, cpufreq_dc, 0, 0);
 136
 137 static int              cf_lowest_freq;
 138 static int              cf_verbose;
 139 static SYSCTL_NODE(_debug, OID_AUTO, cpufreq, CTLFLAG_RD | CTLFLAG_MPSAFE, NULL,
 140     "cpufreq debugging");
 141 SYSCTL_INT(_debug_cpufreq, OID_AUTO, lowest, CTLFLAG_RWTUN, &cf_lowest_freq, 1,
 142     "Don't provide levels below this frequency.");
 143 SYSCTL_INT(_debug_cpufreq, OID_AUTO, verbose, CTLFLAG_RWTUN, &cf_verbose, 1,
 144     "Print verbose debugging messages");
 145
 146 /*
 147  * This is called as the result of a hardware specific frequency control driver
 148  * calling cpufreq_register. It provides a general interface for system wide
 149  * frequency controls and operates on a per cpu basis.
 150  */
 151 static int
 152 cpufreq_attach(device_t dev)
 153 {
 154         struct cpufreq_softc *sc;
 155         struct pcpu *pc;
 156         device_t parent;
 157         uint64_t rate;
 158
 159         CF_DEBUG("initializing %s\n", device_get_nameunit(dev));
 160         sc = device_get_softc(dev);
 161         parent = device_get_parent(dev);
 162         sc->dev = dev;
 163         sysctl_ctx_init(&sc->sysctl_ctx);
 164         TAILQ_INIT(&sc->all_levels);
 165         CF_MTX_INIT(&sc->lock);
 166         sc->curr_level.total_set.freq = CPUFREQ_VAL_UNKNOWN;
 167         SLIST_INIT(&sc->saved_freq);
 168         /* Try to get nominal CPU freq to use it as maximum later if needed */
 169         sc->max_mhz = cpu_get_nominal_mhz(dev);
 170         /* If that fails, try to measure the current rate */
 171         if (sc->max_mhz <= 0) {
 172                 CF_DEBUG("Unable to obtain nominal frequency.\n");
 173                 pc = cpu_get_pcpu(dev);
 174                 if (cpu_est_clockrate(pc->pc_cpuid, &rate) == 0)
 175                         sc->max_mhz = rate / 1000000;
 176                 else
 177                         sc->max_mhz = CPUFREQ_VAL_UNKNOWN;
 178         }
 179
 180         CF_DEBUG("initializing one-time data for %s\n",
 181             device_get_nameunit(dev));
 182         sc->levels_buf = malloc(CF_MAX_LEVELS * sizeof(*sc->levels_buf),
 183             M_DEVBUF, M_WAITOK);
 184         SYSCTL_ADD_PROC(&sc->sysctl_ctx,
 185             SYSCTL_CHILDREN(device_get_sysctl_tree(parent)),
 186             OID_AUTO, "freq", CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_NEEDGIANT,
 187             sc, 0, cpufreq_curr_sysctl, "I", "Current CPU frequency");
 188         SYSCTL_ADD_PROC(&sc->sysctl_ctx,
 189             SYSCTL_CHILDREN(device_get_sysctl_tree(parent)),
 190             OID_AUTO, "freq_levels",
 191             CTLTYPE_STRING | CTLFLAG_RD | CTLFLAG_NEEDGIANT, sc, 0,
 192             cpufreq_levels_sysctl, "A", "CPU frequency levels");
 193
 194         /*
 195          * Queue a one-shot broadcast that levels have changed.
 196          * It will run once the system has completed booting.
 197          */
 198         TASK_INIT(&sc->startup_task, 0, cpufreq_startup_task, dev);
 199         taskqueue_enqueue(taskqueue_thread, &sc->startup_task);
 200
 201         return (0);
 202 }
 203
 204 /* Handle any work to be done for all drivers that attached during boot. */
 205 static void
 206 cpufreq_startup_task(void *ctx, int pending)
 207 {
 208
 209         cpufreq_settings_changed((device_t)ctx);
 210 }
 211
 212 static int
 213 cpufreq_detach(device_t dev)
 214 {
 215         struct cpufreq_softc *sc;
 216         struct cf_saved_freq *saved_freq;
 217
 218         CF_DEBUG("shutdown %s\n", device_get_nameunit(dev));
 219         sc = device_get_softc(dev);
 220         sysctl_ctx_free(&sc->sysctl_ctx);
 221
 222         while ((saved_freq = SLIST_FIRST(&sc->saved_freq)) != NULL) {
 223                 SLIST_REMOVE_HEAD(&sc->saved_freq, link);
 224                 free(saved_freq, M_TEMP);
 225         }
 226
 227         free(sc->levels_buf, M_DEVBUF);
 228
 229         return (0);
 230 }
 231
 232 static int
 233 cf_set_method(device_t dev, const struct cf_level *level, int priority)
 234 {
 235         struct cpufreq_softc *sc;
 236         const struct cf_setting *set;
 237         struct cf_saved_freq *saved_freq, *curr_freq;
 238         struct pcpu *pc;
 239         int error, i;
 240         u_char pri;
 241
 242         sc = device_get_softc(dev);
 243         error = 0;
 244         set = NULL;
 245         saved_freq = NULL;
 246
 247         /* We are going to change levels so notify the pre-change handler. */
 248         EVENTHANDLER_INVOKE(cpufreq_pre_change, level, &error);
 249         if (error != 0) {
 250                 EVENTHANDLER_INVOKE(cpufreq_post_change, level, error);
 251                 return (error);
 252         }
 253
 254         CF_MTX_LOCK(&sc->lock);
 255
 256 #ifdef SMP
 257 #ifdef EARLY_AP_STARTUP
 258         MPASS(mp_ncpus == 1 || smp_started);
 259 #else
 260         /*
 261          * If still booting and secondary CPUs not started yet, don't allow
 262          * changing the frequency until they're online.  This is because we
 263          * can't switch to them using sched_bind() and thus we'd only be
 264          * switching the main CPU.  XXXTODO: Need to think more about how to
 265          * handle having different CPUs at different frequencies.
 266          */
 267         if (mp_ncpus > 1 && !smp_started) {
 268                 device_printf(dev, "rejecting change, SMP not started yet\n");
 269                 error = ENXIO;
 270                 goto out;
 271         }
 272 #endif
 273 #endif /* SMP */
 274
 275         /*
 276          * If the requested level has a lower priority, don't allow
 277          * the new level right now.
 278          */
 279         if (priority < sc->curr_priority) {
 280                 CF_DEBUG("ignoring, curr prio %d less than %d\n", priority,
 281                     sc->curr_priority);
 282                 error = EPERM;
 283                 goto out;
 284         }
 285
 286         /*
 287          * If the caller didn't specify a level and one is saved, prepare to
 288          * restore the saved level.  If none has been saved, return an error.
 289          */
 290         if (level == NULL) {
 291                 saved_freq = SLIST_FIRST(&sc->saved_freq);
 292                 if (saved_freq == NULL) {
 293                         CF_DEBUG("NULL level, no saved level\n");
 294                         error = ENXIO;
 295                         goto out;
 296                 }
 297                 level = &saved_freq->level;
 298                 priority = saved_freq->priority;
 299                 CF_DEBUG("restoring saved level, freq %d prio %d\n",
 300                     level->total_set.freq, priority);
 301         }
 302
 303         /* Reject levels that are below our specified threshold. */
 304         if (level->total_set.freq < cf_lowest_freq) {
 305                 CF_DEBUG("rejecting freq %d, less than %d limit\n",
 306                     level->total_set.freq, cf_lowest_freq);
 307                 error = EINVAL;
 308                 goto out;
 309         }
 310
 311         /* If already at this level, just return. */
 312         if (sc->curr_level.total_set.freq == level->total_set.freq) {
 313                 CF_DEBUG("skipping freq %d, same as current level %d\n",
 314                     level->total_set.freq, sc->curr_level.total_set.freq);
 315                 goto skip;
 316         }
 317
 318         /* First, set the absolute frequency via its driver. */
 319         set = &level->abs_set;
 320         if (set->dev) {
 321                 if (!device_is_attached(set->dev)) {
 322                         error = ENXIO;
 323                         goto out;
 324                 }
 325
 326                 /* Bind to the target CPU before switching. */
 327                 pc = cpu_get_pcpu(set->dev);
 328
 329                 /* Skip settings if CPU is not started. */
 330                 if (pc == NULL) {
 331                         error = 0;
 332                         goto out;
 333                 }
 334                 thread_lock(curthread);
 335                 pri = curthread->td_priority;
 336                 sched_prio(curthread, PRI_MIN);
 337                 sched_bind(curthread, pc->pc_cpuid);
 338                 thread_unlock(curthread);
 339                 CF_DEBUG("setting abs freq %d on %s (cpu %d)\n", set->freq,
 340                     device_get_nameunit(set->dev), PCPU_GET(cpuid));
 341                 error = CPUFREQ_DRV_SET(set->dev, set);
 342                 thread_lock(curthread);
 343                 sched_unbind(curthread);
 344                 sched_prio(curthread, pri);
 345                 thread_unlock(curthread);
 346                 if (error) {
 347                         goto out;
 348                 }
 349         }
 350
 351         /* Next, set any/all relative frequencies via their drivers. */
 352         for (i = 0; i < level->rel_count; i++) {
 353                 set = &level->rel_set[i];
 354                 if (!device_is_attached(set->dev)) {
 355                         error = ENXIO;
 356                         goto out;
 357                 }
 358
 359                 /* Bind to the target CPU before switching. */
 360                 pc = cpu_get_pcpu(set->dev);
 361                 thread_lock(curthread);
 362                 pri = curthread->td_priority;
 363                 sched_prio(curthread, PRI_MIN);
 364                 sched_bind(curthread, pc->pc_cpuid);
 365                 thread_unlock(curthread);
 366                 CF_DEBUG("setting rel freq %d on %s (cpu %d)\n", set->freq,
 367                     device_get_nameunit(set->dev), PCPU_GET(cpuid));
 368                 error = CPUFREQ_DRV_SET(set->dev, set);
 369                 thread_lock(curthread);
 370                 sched_unbind(curthread);
 371                 sched_prio(curthread, pri);
 372                 thread_unlock(curthread);
 373                 if (error) {
 374                         /* XXX Back out any successful setting? */
 375                         goto out;
 376                 }
 377         }
 378
 379 skip:
 380         /*
 381          * Before recording the current level, check if we're going to a
 382          * higher priority.  If so, save the previous level and priority.
 383          */
 384         if (sc->curr_level.total_set.freq != CPUFREQ_VAL_UNKNOWN &&
 385             priority > sc->curr_priority) {
 386                 CF_DEBUG("saving level, freq %d prio %d\n",
 387                     sc->curr_level.total_set.freq, sc->curr_priority);
 388                 curr_freq = malloc(sizeof(*curr_freq), M_TEMP, M_NOWAIT);
 389                 if (curr_freq == NULL) {
 390                         error = ENOMEM;
 391                         goto out;
 392                 }
 393                 curr_freq->level = sc->curr_level;
 394                 curr_freq->priority = sc->curr_priority;
 395                 SLIST_INSERT_HEAD(&sc->saved_freq, curr_freq, link);
 396         }
 397         sc->curr_level = *level;
 398         sc->curr_priority = priority;
 399
 400         /* If we were restoring a saved state, reset it to "unused". */
 401         if (saved_freq != NULL) {
 402                 CF_DEBUG("resetting saved level\n");
 403                 sc->curr_level.total_set.freq = CPUFREQ_VAL_UNKNOWN;
 404                 SLIST_REMOVE_HEAD(&sc->saved_freq, link);
 405                 free(saved_freq, M_TEMP);
 406         }
 407
 408 out:
 409         CF_MTX_UNLOCK(&sc->lock);
 410
 411         /*
 412          * We changed levels (or attempted to) so notify the post-change
 413          * handler of new frequency or error.
 414          */
 415         EVENTHANDLER_INVOKE(cpufreq_post_change, level, error);
 416         if (error && set)
 417                 device_printf(set->dev, "set freq failed, err %d\n", error);
 418
 419         return (error);
 420 }
 421
 422 static int
 423 cpufreq_get_frequency(device_t dev)
 424 {
 425         struct cf_setting set;
 426
 427         if (CPUFREQ_DRV_GET(dev, &set) != 0)
 428                 return (-1);
 429
 430         return (set.freq);
 431 }
 432
 433 /* Returns the index into *levels with the match */
 434 static int
 435 cpufreq_get_level(device_t dev, struct cf_level *levels, int count)
 436 {
 437         int i, freq;
 438
 439         if ((freq = cpufreq_get_frequency(dev)) < 0)
 440                 return (-1);
 441         for (i = 0; i < count; i++)
 442                 if (freq == levels[i].total_set.freq)
 443                         return (i);
 444
 445         return (-1);
 446 }
 447
 448 /*
 449  * Used by the cpufreq core, this function will populate *level with the current
 450  * frequency as either determined by a cached value sc->curr_level, or in the
 451  * case the lower level driver has set the CPUFREQ_FLAG_UNCACHED flag, it will
 452  * obtain the frequency from the driver itself.
 453  */
 454 static int
 455 cf_get_method(device_t dev, struct cf_level *level)
 456 {
 457         struct cpufreq_softc *sc;
 458         struct cf_level *levels;
 459         struct cf_setting *curr_set;
 460         struct pcpu *pc;
 461         int bdiff, count, diff, error, i, type;
 462         uint64_t rate;
 463
 464         sc = device_get_softc(dev);
 465         error = 0;
 466         levels = NULL;
 467
 468         /*
 469          * If we already know the current frequency, and the driver didn't ask
 470          * for uncached usage, we're done.
 471          */
 472         CF_MTX_LOCK(&sc->lock);
 473         curr_set = &sc->curr_level.total_set;
 474         error = CPUFREQ_DRV_TYPE(sc->cf_drv_dev, &type);
 475         if (error == 0 && (type & CPUFREQ_FLAG_UNCACHED)) {
 476                 struct cf_setting set;
 477
 478                 /*
 479                  * If the driver wants to always report back the real frequency,
 480                  * first try the driver and if that fails, fall back to
 481                  * estimating.
 482                  */
 483                 if (CPUFREQ_DRV_GET(sc->cf_drv_dev, &set) == 0) {
 484                         sc->curr_level.total_set = set;
 485                         CF_DEBUG("get returning immediate freq %d\n",
 486                             curr_set->freq);
 487                         goto out;
 488                 }
 489         } else if (curr_set->freq != CPUFREQ_VAL_UNKNOWN) {
 490                 CF_DEBUG("get returning known freq %d\n", curr_set->freq);
 491                 error = 0;
 492                 goto out;
 493         }
 494         CF_MTX_UNLOCK(&sc->lock);
 495
 496         /*
 497          * We need to figure out the current level.  Loop through every
 498          * driver, getting the current setting.  Then, attempt to get a best
 499          * match of settings against each level.
 500          */
 501         count = CF_MAX_LEVELS;
 502         levels = malloc(count * sizeof(*levels), M_TEMP, M_NOWAIT);
 503         if (levels == NULL)
 504                 return (ENOMEM);
 505         error = CPUFREQ_LEVELS(sc->dev, levels, &count);
 506         if (error) {
 507                 if (error == E2BIG)
 508                         printf("cpufreq: need to increase CF_MAX_LEVELS\n");
 509                 free(levels, M_TEMP);
 510                 return (error);
 511         }
 512
 513         /*
 514          * Reacquire the lock and search for the given level.
 515          *
 516          * XXX Note: this is not quite right since we really need to go
 517          * through each level and compare both absolute and relative
 518          * settings for each driver in the system before making a match.
 519          * The estimation code below catches this case though.
 520          */
 521         CF_MTX_LOCK(&sc->lock);
 522         i = cpufreq_get_level(sc->cf_drv_dev, levels, count);
 523         if (i >= 0)
 524                 sc->curr_level = levels[i];
 525         else
 526                 CF_DEBUG("Couldn't find supported level for %s\n",
 527                     device_get_nameunit(sc->cf_drv_dev));
 528
 529         if (curr_set->freq != CPUFREQ_VAL_UNKNOWN) {
 530                 CF_DEBUG("get matched freq %d from drivers\n", curr_set->freq);
 531                 goto out;
 532         }
 533
 534         /*
 535          * We couldn't find an exact match, so attempt to estimate and then
 536          * match against a level.
 537          */
 538         pc = cpu_get_pcpu(dev);
 539         if (pc == NULL) {
 540                 error = ENXIO;
 541                 goto out;
 542         }
 543         cpu_est_clockrate(pc->pc_cpuid, &rate);
 544         rate /= 1000000;
 545         bdiff = 1 << 30;
 546         for (i = 0; i < count; i++) {
 547                 diff = abs(levels[i].total_set.freq - rate);
 548                 if (diff < bdiff) {
 549                         bdiff = diff;
 550                         sc->curr_level = levels[i];
 551                 }
 552         }
 553         CF_DEBUG("get estimated freq %d\n", curr_set->freq);
 554
 555 out:
 556         if (error == 0)
 557                 *level = sc->curr_level;
 558
 559         CF_MTX_UNLOCK(&sc->lock);
 560         if (levels)
 561                 free(levels, M_TEMP);
 562         return (error);
 563 }
 564
 565 /*
 566  * Either directly obtain settings from the cpufreq driver, or build a list of
 567  * relative settings to be integrated later against an absolute max.
 568  */
 569 static int
 570 cpufreq_add_levels(device_t cf_dev, struct cf_setting_lst *rel_sets)
 571 {
 572         struct cf_setting_array *set_arr;
 573         struct cf_setting *sets;
 574         device_t dev;
 575         struct cpufreq_softc *sc;
 576         int type, set_count, error;
 577
 578         sc = device_get_softc(cf_dev);
 579         dev = sc->cf_drv_dev;
 580
 581         /* Skip devices that aren't ready. */
 582         if (!device_is_attached(cf_dev))
 583                 return (0);
 584
 585         /*
 586          * Get settings, skipping drivers that offer no settings or
 587          * provide settings for informational purposes only.
 588          */
 589         error = CPUFREQ_DRV_TYPE(dev, &type);
 590         if (error != 0 || (type & CPUFREQ_FLAG_INFO_ONLY)) {
 591                 if (error == 0) {
 592                         CF_DEBUG("skipping info-only driver %s\n",
 593                             device_get_nameunit(cf_dev));
 594                 }
 595                 return (error);
 596         }
 597
 598         sets = malloc(MAX_SETTINGS * sizeof(*sets), M_TEMP, M_NOWAIT);
 599         if (sets == NULL)
 600                 return (ENOMEM);
 601
 602         set_count = MAX_SETTINGS;
 603         error = CPUFREQ_DRV_SETTINGS(dev, sets, &set_count);
 604         if (error != 0 || set_count == 0)
 605                 goto out;
 606
 607         /* Add the settings to our absolute/relative lists. */
 608         switch (type & CPUFREQ_TYPE_MASK) {
 609         case CPUFREQ_TYPE_ABSOLUTE:
 610                 error = cpufreq_insert_abs(sc, sets, set_count);
 611                 break;
 612         case CPUFREQ_TYPE_RELATIVE:
 613                 CF_DEBUG("adding %d relative settings\n", set_count);
 614                 set_arr = malloc(sizeof(*set_arr), M_TEMP, M_NOWAIT);
 615                 if (set_arr == NULL) {
 616                         error = ENOMEM;
 617                         goto out;
 618                 }
 619                 bcopy(sets, set_arr->sets, set_count * sizeof(*sets));
 620                 set_arr->count = set_count;
 621                 TAILQ_INSERT_TAIL(rel_sets, set_arr, link);
 622                 break;
 623         default:
 624                 error = EINVAL;
 625         }
 626
 627 out:
 628         free(sets, M_TEMP);
 629         return (error);
 630 }
 631
 632 static int
 633 cf_levels_method(device_t dev, struct cf_level *levels, int *count)
 634 {
 635         struct cf_setting_array *set_arr;
 636         struct cf_setting_lst rel_sets;
 637         struct cpufreq_softc *sc;
 638         struct cf_level *lev;
 639         struct pcpu *pc;
 640         int error, i;
 641         uint64_t rate;
 642
 643         if (levels == NULL || count == NULL)
 644                 return (EINVAL);
 645
 646         TAILQ_INIT(&rel_sets);
 647         sc = device_get_softc(dev);
 648
 649         CF_MTX_LOCK(&sc->lock);
 650         error = cpufreq_add_levels(sc->dev, &rel_sets);
 651         if (error)
 652                 goto out;
 653
 654         /*
 655          * If there are no absolute levels, create a fake one at 100%.  We
 656          * then cache the clockrate for later use as our base frequency.
 657          */
 658         if (TAILQ_EMPTY(&sc->all_levels)) {
 659                 struct cf_setting set;
 660
 661                 CF_DEBUG("No absolute levels returned by driver\n");
 662
 663                 if (sc->max_mhz == CPUFREQ_VAL_UNKNOWN) {
 664                         sc->max_mhz = cpu_get_nominal_mhz(dev);
 665                         /*
 666                          * If the CPU can't report a rate for 100%, hope
 667                          * the CPU is running at its nominal rate right now,
 668                          * and use that instead.
 669                          */
 670                         if (sc->max_mhz <= 0) {
 671                                 pc = cpu_get_pcpu(dev);
 672                                 cpu_est_clockrate(pc->pc_cpuid, &rate);
 673                                 sc->max_mhz = rate / 1000000;
 674                         }
 675                 }
 676                 memset(&set, CPUFREQ_VAL_UNKNOWN, sizeof(set));
 677                 set.freq = sc->max_mhz;
 678                 set.dev = NULL;
 679                 error = cpufreq_insert_abs(sc, &set, 1);
 680                 if (error)
 681                         goto out;
 682         }
 683
 684         /* Create a combined list of absolute + relative levels. */
 685         TAILQ_FOREACH(set_arr, &rel_sets, link)
 686                 cpufreq_expand_set(sc, set_arr);
 687
 688         /* If the caller doesn't have enough space, return the actual count. */
 689         if (sc->all_count > *count) {
 690                 *count = sc->all_count;
 691                 error = E2BIG;
 692                 goto out;
 693         }
 694
 695         /* Finally, output the list of levels. */
 696         i = 0;
 697         TAILQ_FOREACH(lev, &sc->all_levels, link) {
 698                 /* Skip levels that have a frequency that is too low. */
 699                 if (lev->total_set.freq < cf_lowest_freq) {
 700                         sc->all_count--;
 701                         continue;
 702                 }
 703
 704                 levels[i] = *lev;
 705                 i++;
 706         }
 707         *count = sc->all_count;
 708         error = 0;
 709
 710 out:
 711         /* Clear all levels since we regenerate them each time. */
 712         while ((lev = TAILQ_FIRST(&sc->all_levels)) != NULL) {
 713                 TAILQ_REMOVE(&sc->all_levels, lev, link);
 714                 free(lev, M_TEMP);
 715         }
 716         sc->all_count = 0;
 717
 718         CF_MTX_UNLOCK(&sc->lock);
 719         while ((set_arr = TAILQ_FIRST(&rel_sets)) != NULL) {
 720                 TAILQ_REMOVE(&rel_sets, set_arr, link);
 721                 free(set_arr, M_TEMP);
 722         }
 723         return (error);
 724 }
 725
 726 /*
 727  * Create levels for an array of absolute settings and insert them in
 728  * sorted order in the specified list.
 729  */
 730 static int
 731 cpufreq_insert_abs(struct cpufreq_softc *sc, struct cf_setting *sets,
 732     int count)
 733 {
 734         struct cf_level_lst *list;
 735         struct cf_level *level, *search;
 736         int i, inserted;
 737
 738         CF_MTX_ASSERT(&sc->lock);
 739
 740         list = &sc->all_levels;
 741         for (i = 0; i < count; i++) {
 742                 level = malloc(sizeof(*level), M_TEMP, M_NOWAIT | M_ZERO);
 743                 if (level == NULL)
 744                         return (ENOMEM);
 745                 level->abs_set = sets[i];
 746                 level->total_set = sets[i];
 747                 level->total_set.dev = NULL;
 748                 sc->all_count++;
 749                 inserted = 0;
 750
 751                 if (TAILQ_EMPTY(list)) {
 752                         CF_DEBUG("adding abs setting %d at head\n",
 753                             sets[i].freq);
 754                         TAILQ_INSERT_HEAD(list, level, link);
 755                         continue;
 756                 }
 757
 758                 TAILQ_FOREACH_REVERSE(search, list, cf_level_lst, link)
 759                         if (sets[i].freq <= search->total_set.freq) {
 760                                 CF_DEBUG("adding abs setting %d after %d\n",
 761                                     sets[i].freq, search->total_set.freq);
 762                                 TAILQ_INSERT_AFTER(list, search, level, link);
 763                                 inserted = 1;
 764                                 break;
 765                         }
 766
 767                 if (inserted == 0) {
 768                         TAILQ_FOREACH(search, list, link)
 769                                 if (sets[i].freq >= search->total_set.freq) {
 770                                         CF_DEBUG("adding abs setting %d before %d\n",
 771                                             sets[i].freq, search->total_set.freq);
 772                                         TAILQ_INSERT_BEFORE(search, level, link);
 773                                         break;
 774                                 }
 775                 }
 776         }
 777
 778         return (0);
 779 }
 780
 781 /*
 782  * Expand a group of relative settings, creating derived levels from them.
 783  */
 784 static int
 785 cpufreq_expand_set(struct cpufreq_softc *sc, struct cf_setting_array *set_arr)
 786 {
 787         struct cf_level *fill, *search;
 788         struct cf_setting *set;
 789         int i;
 790
 791         CF_MTX_ASSERT(&sc->lock);
 792
 793         /*
 794          * Walk the set of all existing levels in reverse.  This is so we
 795          * create derived states from the lowest absolute settings first
 796          * and discard duplicates created from higher absolute settings.
 797          * For instance, a level of 50 Mhz derived from 100 Mhz + 50% is
 798          * preferable to 200 Mhz + 25% because absolute settings are more
 799          * efficient since they often change the voltage as well.
 800          */
 801         TAILQ_FOREACH_REVERSE(search, &sc->all_levels, cf_level_lst, link) {
 802                 /* Add each setting to the level, duplicating if necessary. */
 803                 for (i = 0; i < set_arr->count; i++) {
 804                         set = &set_arr->sets[i];
 805
 806                         /*
 807                          * If this setting is less than 100%, split the level
 808                          * into two and add this setting to the new level.
 809                          */
 810                         fill = search;
 811                         if (set->freq < 10000) {
 812                                 fill = cpufreq_dup_set(sc, search, set);
 813
 814                                 /*
 815                                  * The new level was a duplicate of an existing
 816                                  * level or its absolute setting is too high
 817                                  * so we freed it.  For example, we discard a
 818                                  * derived level of 1000 MHz/25% if a level
 819                                  * of 500 MHz/100% already exists.
 820                                  */
 821                                 if (fill == NULL)
 822                                         break;
 823                         }
 824
 825                         /* Add this setting to the existing or new level. */
 826                         KASSERT(fill->rel_count < MAX_SETTINGS,
 827                             ("cpufreq: too many relative drivers (%d)",
 828                             MAX_SETTINGS));
 829                         fill->rel_set[fill->rel_count] = *set;
 830                         fill->rel_count++;
 831                         CF_DEBUG(
 832                         "expand set added rel setting %d%% to %d level\n",
 833                             set->freq / 100, fill->total_set.freq);
 834                 }
 835         }
 836
 837         return (0);
 838 }
 839
 840 static struct cf_level *
 841 cpufreq_dup_set(struct cpufreq_softc *sc, struct cf_level *dup,
 842     struct cf_setting *set)
 843 {
 844         struct cf_level_lst *list;
 845         struct cf_level *fill, *itr;
 846         struct cf_setting *fill_set, *itr_set;
 847         int i;
 848
 849         CF_MTX_ASSERT(&sc->lock);
 850
 851         /*
 852          * Create a new level, copy it from the old one, and update the
 853          * total frequency and power by the percentage specified in the
 854          * relative setting.
 855          */
 856         fill = malloc(sizeof(*fill), M_TEMP, M_NOWAIT);
 857         if (fill == NULL)
 858                 return (NULL);
 859         *fill = *dup;
 860         fill_set = &fill->total_set;
 861         fill_set->freq =
 862             ((uint64_t)fill_set->freq * set->freq) / 10000;
 863         if (fill_set->power != CPUFREQ_VAL_UNKNOWN) {
 864                 fill_set->power = ((uint64_t)fill_set->power * set->freq)
 865                     / 10000;
 866         }
 867         if (set->lat != CPUFREQ_VAL_UNKNOWN) {
 868                 if (fill_set->lat != CPUFREQ_VAL_UNKNOWN)
 869                         fill_set->lat += set->lat;
 870                 else
 871                         fill_set->lat = set->lat;
 872         }
 873         CF_DEBUG("dup set considering derived setting %d\n", fill_set->freq);
 874
 875         /*
 876          * If we copied an old level that we already modified (say, at 100%),
 877          * we need to remove that setting before adding this one.  Since we
 878          * process each setting array in order, we know any settings for this
 879          * driver will be found at the end.
 880          */
 881         for (i = fill->rel_count; i != 0; i--) {
 882                 if (fill->rel_set[i - 1].dev != set->dev)
 883                         break;
 884                 CF_DEBUG("removed last relative driver: %s\n",
 885                     device_get_nameunit(set->dev));
 886                 fill->rel_count--;
 887         }
 888
 889         /*
 890          * Insert the new level in sorted order.  If it is a duplicate of an
 891          * existing level (1) or has an absolute setting higher than the
 892          * existing level (2), do not add it.  We can do this since any such
 893          * level is guaranteed use less power.  For example (1), a level with
 894          * one absolute setting of 800 Mhz uses less power than one composed
 895          * of an absolute setting of 1600 Mhz and a relative setting at 50%.
 896          * Also for example (2), a level of 800 Mhz/75% is preferable to
 897          * 1600 Mhz/25% even though the latter has a lower total frequency.
 898          */
 899         list = &sc->all_levels;
 900         KASSERT(!TAILQ_EMPTY(list), ("all levels list empty in dup set"));
 901         TAILQ_FOREACH_REVERSE(itr, list, cf_level_lst, link) {
 902                 itr_set = &itr->total_set;
 903                 if (CPUFREQ_CMP(fill_set->freq, itr_set->freq)) {
 904                         CF_DEBUG("dup set rejecting %d (dupe)\n",
 905                             fill_set->freq);
 906                         itr = NULL;
 907                         break;
 908                 } else if (fill_set->freq < itr_set->freq) {
 909                         if (fill->abs_set.freq <= itr->abs_set.freq) {
 910                                 CF_DEBUG(
 911                         "dup done, inserting new level %d after %d\n",
 912                                     fill_set->freq, itr_set->freq);
 913                                 TAILQ_INSERT_AFTER(list, itr, fill, link);
 914                                 sc->all_count++;
 915                         } else {
 916                                 CF_DEBUG("dup set rejecting %d (abs too big)\n",
 917                                     fill_set->freq);
 918                                 itr = NULL;
 919                         }
 920                         break;
 921                 }
 922         }
 923
 924         /* We didn't find a good place for this new level so free it. */
 925         if (itr == NULL) {
 926                 CF_DEBUG("dup set freeing new level %d (not optimal)\n",
 927                     fill_set->freq);
 928                 free(fill, M_TEMP);
 929                 fill = NULL;
 930         }
 931
 932         return (fill);
 933 }
 934
 935 static int
 936 cpufreq_curr_sysctl(SYSCTL_HANDLER_ARGS)
 937 {
 938         struct cpufreq_softc *sc;
 939         struct cf_level *levels;
 940         int best, count, diff, bdiff, devcount, error, freq, i, n;
 941         device_t *devs;
 942
 943         devs = NULL;
 944         sc = oidp->oid_arg1;
 945         levels = sc->levels_buf;
 946
 947         error = CPUFREQ_GET(sc->dev, &levels[0]);
 948         if (error)
 949                 goto out;
 950         freq = levels[0].total_set.freq;
 951         error = sysctl_handle_int(oidp, &freq, 0, req);
 952         if (error != 0 || req->newptr == NULL)
 953                 goto out;
 954
 955         /*
 956          * While we only call cpufreq_get() on one device (assuming all
 957          * CPUs have equal levels), we call cpufreq_set() on all CPUs.
 958          * This is needed for some MP systems.
 959          */
 960         error = devclass_get_devices(cpufreq_dc, &devs, &devcount);
 961         if (error)
 962                 goto out;
 963         for (n = 0; n < devcount; n++) {
 964                 count = CF_MAX_LEVELS;
 965                 error = CPUFREQ_LEVELS(devs[n], levels, &count);
 966                 if (error) {
 967                         if (error == E2BIG)
 968                                 printf(
 969                         "cpufreq: need to increase CF_MAX_LEVELS\n");
 970                         break;
 971                 }
 972                 best = 0;
 973                 bdiff = 1 << 30;
 974                 for (i = 0; i < count; i++) {
 975                         diff = abs(levels[i].total_set.freq - freq);
 976                         if (diff < bdiff) {
 977                                 bdiff = diff;
 978                                 best = i;
 979                         }
 980                 }
 981                 error = CPUFREQ_SET(devs[n], &levels[best], CPUFREQ_PRIO_USER);
 982         }
 983
 984 out:
 985         if (devs)
 986                 free(devs, M_TEMP);
 987         return (error);
 988 }
 989
 990 static int
 991 cpufreq_levels_sysctl(SYSCTL_HANDLER_ARGS)
 992 {
 993         struct cpufreq_softc *sc;
 994         struct cf_level *levels;
 995         struct cf_setting *set;
 996         struct sbuf sb;
 997         int count, error, i;
 998
 999         sc = oidp->oid_arg1;
1000         sbuf_new(&sb, NULL, 128, SBUF_AUTOEXTEND);
1001
1002         /* Get settings from the device and generate the output string. */
1003         count = CF_MAX_LEVELS;
1004         levels = sc->levels_buf;
1005         if (levels == NULL) {
1006                 sbuf_delete(&sb);
1007                 return (ENOMEM);
1008         }
1009         error = CPUFREQ_LEVELS(sc->dev, levels, &count);
1010         if (error) {
1011                 if (error == E2BIG)
1012                         printf("cpufreq: need to increase CF_MAX_LEVELS\n");
1013                 goto out;
1014         }
1015         if (count) {
1016                 for (i = 0; i < count; i++) {
1017                         set = &levels[i].total_set;
1018                         sbuf_printf(&sb, "%d/%d ", set->freq, set->power);
1019                 }
1020         } else
1021                 sbuf_cpy(&sb, "0");
1022         sbuf_trim(&sb);
1023         sbuf_finish(&sb);
1024         error = sysctl_handle_string(oidp, sbuf_data(&sb), sbuf_len(&sb), req);
1025
1026 out:
1027         sbuf_delete(&sb);
1028         return (error);
1029 }
1030
1031 static int
1032 cpufreq_settings_sysctl(SYSCTL_HANDLER_ARGS)
1033 {
1034         device_t dev;
1035         struct cf_setting *sets;
1036         struct sbuf sb;
1037         int error, i, set_count;
1038
1039         dev = oidp->oid_arg1;
1040         sbuf_new(&sb, NULL, 128, SBUF_AUTOEXTEND);
1041
1042         /* Get settings from the device and generate the output string. */
1043         set_count = MAX_SETTINGS;
1044         sets = malloc(set_count * sizeof(*sets), M_TEMP, M_NOWAIT);
1045         if (sets == NULL) {
1046                 sbuf_delete(&sb);
1047                 return (ENOMEM);
1048         }
1049         error = CPUFREQ_DRV_SETTINGS(dev, sets, &set_count);
1050         if (error)
1051                 goto out;
1052         if (set_count) {
1053                 for (i = 0; i < set_count; i++)
1054                         sbuf_printf(&sb, "%d/%d ", sets[i].freq, sets[i].power);
1055         } else
1056                 sbuf_cpy(&sb, "0");
1057         sbuf_trim(&sb);
1058         sbuf_finish(&sb);
1059         error = sysctl_handle_string(oidp, sbuf_data(&sb), sbuf_len(&sb), req);
1060
1061 out:
1062         free(sets, M_TEMP);
1063         sbuf_delete(&sb);
1064         return (error);
1065 }
1066
1067 static void
1068 cpufreq_add_freq_driver_sysctl(device_t cf_dev)
1069 {
1070         struct cpufreq_softc *sc;
1071
1072         sc = device_get_softc(cf_dev);
1073         SYSCTL_ADD_CONST_STRING(&sc->sysctl_ctx,
1074             SYSCTL_CHILDREN(device_get_sysctl_tree(cf_dev)), OID_AUTO,
1075             "freq_driver", CTLFLAG_RD, device_get_nameunit(sc->cf_drv_dev),
1076             "cpufreq driver used by this cpu");
1077 }
1078
1079 int
1080 cpufreq_register(device_t dev)
1081 {
1082         struct cpufreq_softc *sc;
1083         device_t cf_dev, cpu_dev;
1084         int error;
1085
1086         /* Add a sysctl to get each driver's settings separately. */
1087         SYSCTL_ADD_PROC(device_get_sysctl_ctx(dev),
1088             SYSCTL_CHILDREN(device_get_sysctl_tree(dev)),
1089             OID_AUTO, "freq_settings",
1090             CTLTYPE_STRING | CTLFLAG_RD | CTLFLAG_NEEDGIANT, dev, 0,
1091             cpufreq_settings_sysctl, "A", "CPU frequency driver settings");
1092
1093         /*
1094          * Add only one cpufreq device to each CPU.  Currently, all CPUs
1095          * must offer the same levels and be switched at the same time.
1096          */
1097         cpu_dev = device_get_parent(dev);
1098         if ((cf_dev = device_find_child(cpu_dev, "cpufreq", -1))) {
1099                 sc = device_get_softc(cf_dev);
1100                 sc->max_mhz = CPUFREQ_VAL_UNKNOWN;
1101                 MPASS(sc->cf_drv_dev != NULL);
1102                 return (0);
1103         }
1104
1105         /* Add the child device and possibly sysctls. */
1106         cf_dev = BUS_ADD_CHILD(cpu_dev, 0, "cpufreq", -1);
1107         if (cf_dev == NULL)
1108                 return (ENOMEM);
1109         device_quiet(cf_dev);
1110
1111         error = device_probe_and_attach(cf_dev);
1112         if (error)
1113                 return (error);
1114
1115         sc = device_get_softc(cf_dev);
1116         sc->cf_drv_dev = dev;
1117         cpufreq_add_freq_driver_sysctl(cf_dev);
1118         return (error);
1119 }
1120
1121 int
1122 cpufreq_unregister(device_t dev)
1123 {
1124         device_t cf_dev;
1125         struct cpufreq_softc *sc;
1126
1127         /*
1128          * If this is the last cpufreq child device, remove the control
1129          * device as well.  We identify cpufreq children by calling a method
1130          * they support.
1131          */
1132         cf_dev = device_find_child(device_get_parent(dev), "cpufreq", -1);
1133         if (cf_dev == NULL) {
1134                 device_printf(dev,
1135         "warning: cpufreq_unregister called with no cpufreq device active\n");
1136                 return (0);
1137         }
1138         sc = device_get_softc(cf_dev);
1139         MPASS(sc->cf_drv_dev == dev);
1140         device_delete_child(device_get_parent(cf_dev), cf_dev);
1141
1142         return (0);
1143 }
1144
1145 int
1146 cpufreq_settings_changed(device_t dev)
1147 {
1148
1149         EVENTHANDLER_INVOKE(cpufreq_levels_changed,
1150             device_get_unit(device_get_parent(dev)));
1151         return (0);
1152 }