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