sys/kern/kern_cpu.c

   1 /*-
   2  * SPDX-License-Identifier: BSD-2-Clause
   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 #include <sys/param.h>
  31 #include <sys/bus.h>
  32 #include <sys/cpu.h>
  33 #include <sys/eventhandler.h>
  34 #include <sys/kernel.h>
  35 #include <sys/lock.h>
  36 #include <sys/malloc.h>
  37 #include <sys/module.h>
  38 #include <sys/proc.h>
  39 #include <sys/queue.h>
  40 #include <sys/sbuf.h>
  41 #include <sys/sched.h>
  42 #include <sys/smp.h>
  43 #include <sys/sysctl.h>
  44 #include <sys/systm.h>
  45 #include <sys/sx.h>
  46 #include <sys/timetc.h>
  47 #include <sys/taskqueue.h>
  48
  49 #include "cpufreq_if.h"
  50
  51 /*
  52  * Common CPU frequency glue code.  Drivers for specific hardware can
  53  * attach this interface to allow users to get/set the CPU frequency.
  54  */
  55
  56 /*
  57  * Number of levels we can handle.  Levels are synthesized from settings
  58  * so for M settings and N drivers, there may be M*N levels.
  59  */
  60 #define CF_MAX_LEVELS   256
  61
  62 struct cf_saved_freq {
  63         struct cf_level                 level;
  64         int                             priority;
  65         SLIST_ENTRY(cf_saved_freq)      link;
  66 };
  67
  68 struct cpufreq_softc {
  69         struct sx                       lock;
  70         struct cf_level                 curr_level;
  71         int                             curr_priority;
  72         SLIST_HEAD(, cf_saved_freq)     saved_freq;
  73         struct cf_level_lst             all_levels;
  74         int                             all_count;
  75         int                             max_mhz;
  76         device_t                        dev;
  77         device_t                        cf_drv_dev;
  78         struct sysctl_ctx_list          sysctl_ctx;
  79         struct task                     startup_task;
  80         struct cf_level                 *levels_buf;
  81 };
  82
  83 struct cf_setting_array {
  84         struct cf_setting               sets[MAX_SETTINGS];
  85         int                             count;
  86         TAILQ_ENTRY(cf_setting_array)   link;
  87 };
  88
  89 TAILQ_HEAD(cf_setting_lst, cf_setting_array);
  90
  91 #define CF_MTX_INIT(x)          sx_init((x), "cpufreq lock")
  92 #define CF_MTX_LOCK(x)          sx_xlock((x))
  93 #define CF_MTX_UNLOCK(x)        sx_xunlock((x))
  94 #define CF_MTX_ASSERT(x)        sx_assert((x), SX_XLOCKED)
  95
  96 #define CF_DEBUG(msg...)        do {            \
  97         if (cf_verbose)                         \
  98                 printf("cpufreq: " msg);        \
  99         } while (0)
 100
 101 static int      cpufreq_attach(device_t dev);
 102 static void     cpufreq_startup_task(void *ctx, int pending);
 103 static int      cpufreq_detach(device_t dev);
 104 static int      cf_set_method(device_t dev, const struct cf_level *level,
 105                     int priority);
 106 static int      cf_get_method(device_t dev, struct cf_level *level);
 107 static int      cf_levels_method(device_t dev, struct cf_level *levels,
 108                     int *count);
 109 static int      cpufreq_insert_abs(struct cpufreq_softc *sc,
 110                     struct cf_setting *sets, int count);
 111 static int      cpufreq_expand_set(struct cpufreq_softc *sc,
 112                     struct cf_setting_array *set_arr);
 113 static struct cf_level *cpufreq_dup_set(struct cpufreq_softc *sc,
 114                     struct cf_level *dup, struct cf_setting *set);
 115 static int      cpufreq_curr_sysctl(SYSCTL_HANDLER_ARGS);
 116 static int      cpufreq_levels_sysctl(SYSCTL_HANDLER_ARGS);
 117 static int      cpufreq_settings_sysctl(SYSCTL_HANDLER_ARGS);
 118
 119 static device_method_t cpufreq_methods[] = {
 120         DEVMETHOD(device_probe,         bus_generic_probe),
 121         DEVMETHOD(device_attach,        cpufreq_attach),
 122         DEVMETHOD(device_detach,        cpufreq_detach),
 123
 124         DEVMETHOD(cpufreq_set,          cf_set_method),
 125         DEVMETHOD(cpufreq_get,          cf_get_method),
 126         DEVMETHOD(cpufreq_levels,       cf_levels_method),
 127         {0, 0}
 128 };
 129 static driver_t cpufreq_driver = {
 130         "cpufreq", cpufreq_methods, sizeof(struct cpufreq_softc)
 131 };
 132 static devclass_t cpufreq_dc;
 133 DRIVER_MODULE(cpufreq, cpu, cpufreq_driver, cpufreq_dc, 0, 0);
 134
 135 static int              cf_lowest_freq;
 136 static int              cf_verbose;
 137 static SYSCTL_NODE(_debug, OID_AUTO, cpufreq, CTLFLAG_RD | CTLFLAG_MPSAFE, NULL,
 138     "cpufreq debugging");
 139 SYSCTL_INT(_debug_cpufreq, OID_AUTO, lowest, CTLFLAG_RWTUN, &cf_lowest_freq, 1,
 140     "Don't provide levels below this frequency.");
 141 SYSCTL_INT(_debug_cpufreq, OID_AUTO, verbose, CTLFLAG_RWTUN, &cf_verbose, 1,
 142     "Print verbose debugging messages");
 143
 144 /*
 145  * This is called as the result of a hardware specific frequency control driver
 146  * calling cpufreq_register. It provides a general interface for system wide
 147  * frequency controls and operates on a per cpu basis.
 148  */
 149 static int
 150 cpufreq_attach(device_t dev)
 151 {
 152         struct cpufreq_softc *sc;
 153         struct pcpu *pc;
 154         device_t parent;
 155         uint64_t rate;
 156
 157         CF_DEBUG("initializing %s\n", device_get_nameunit(dev));
 158         sc = device_get_softc(dev);
 159         parent = device_get_parent(dev);
 160         sc->dev = dev;
 161         sysctl_ctx_init(&sc->sysctl_ctx);
 162         TAILQ_INIT(&sc->all_levels);
 163         CF_MTX_INIT(&sc->lock);
 164         sc->curr_level.total_set.freq = CPUFREQ_VAL_UNKNOWN;
 165         SLIST_INIT(&sc->saved_freq);
 166         /* Try to get nominal CPU freq to use it as maximum later if needed */
 167         sc->max_mhz = cpu_get_nominal_mhz(dev);
 168         /* If that fails, try to measure the current rate */
 169         if (sc->max_mhz <= 0) {
 170                 CF_DEBUG("Unable to obtain nominal frequency.\n");
 171                 pc = cpu_get_pcpu(dev);
 172                 if (cpu_est_clockrate(pc->pc_cpuid, &rate) == 0)
 173                         sc->max_mhz = rate / 1000000;
 174                 else
 175                         sc->max_mhz = CPUFREQ_VAL_UNKNOWN;
 176         }
 177
 178         CF_DEBUG("initializing one-time data for %s\n",
 179             device_get_nameunit(dev));
 180         sc->levels_buf = malloc(CF_MAX_LEVELS * sizeof(*sc->levels_buf),
 181             M_DEVBUF, M_WAITOK);
 182         SYSCTL_ADD_PROC(&sc->sysctl_ctx,
 183             SYSCTL_CHILDREN(device_get_sysctl_tree(parent)),
 184             OID_AUTO, "freq", CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_NEEDGIANT,
 185             sc, 0, cpufreq_curr_sysctl, "I", "Current CPU frequency");
 186         SYSCTL_ADD_PROC(&sc->sysctl_ctx,
 187             SYSCTL_CHILDREN(device_get_sysctl_tree(parent)),
 188             OID_AUTO, "freq_levels",
 189             CTLTYPE_STRING | CTLFLAG_RD | CTLFLAG_NEEDGIANT, sc, 0,
 190             cpufreq_levels_sysctl, "A", "CPU frequency levels");
 191
 192         /*
 193          * Queue a one-shot broadcast that levels have changed.
 194          * It will run once the system has completed booting.
 195          */
 196         TASK_INIT(&sc->startup_task, 0, cpufreq_startup_task, dev);
 197         taskqueue_enqueue(taskqueue_thread, &sc->startup_task);
 198
 199         return (0);
 200 }
 201
 202 /* Handle any work to be done for all drivers that attached during boot. */
 203 static void
 204 cpufreq_startup_task(void *ctx, int pending)
 205 {
 206
 207         cpufreq_settings_changed((device_t)ctx);
 208 }
 209
 210 static int
 211 cpufreq_detach(device_t dev)
 212 {
 213         struct cpufreq_softc *sc;
 214         struct cf_saved_freq *saved_freq;
 215
 216         CF_DEBUG("shutdown %s\n", device_get_nameunit(dev));
 217         sc = device_get_softc(dev);
 218         sysctl_ctx_free(&sc->sysctl_ctx);
 219
 220         while ((saved_freq = SLIST_FIRST(&sc->saved_freq)) != NULL) {
 221                 SLIST_REMOVE_HEAD(&sc->saved_freq, link);
 222                 free(saved_freq, M_TEMP);
 223         }
 224
 225         free(sc->levels_buf, M_DEVBUF);
 226
 227         return (0);
 228 }
 229
 230 static int
 231 cf_set_method(device_t dev, const struct cf_level *level, int priority)
 232 {
 233         struct cpufreq_softc *sc;
 234         const struct cf_setting *set;
 235         struct cf_saved_freq *saved_freq, *curr_freq;
 236         struct pcpu *pc;
 237         int error, i;
 238         u_char pri;
 239
 240         sc = device_get_softc(dev);
 241         error = 0;
 242         set = NULL;
 243         saved_freq = NULL;
 244
 245         /* We are going to change levels so notify the pre-change handler. */
 246         EVENTHANDLER_INVOKE(cpufreq_pre_change, level, &error);
 247         if (error != 0) {
 248                 EVENTHANDLER_INVOKE(cpufreq_post_change, level, error);
 249                 return (error);
 250         }
 251
 252         CF_MTX_LOCK(&sc->lock);
 253
 254 #ifdef SMP
 255 #ifdef EARLY_AP_STARTUP
 256         MPASS(mp_ncpus == 1 || smp_started);
 257 #else
 258         /*
 259          * If still booting and secondary CPUs not started yet, don't allow
 260          * changing the frequency until they're online.  This is because we
 261          * can't switch to them using sched_bind() and thus we'd only be
 262          * switching the main CPU.  XXXTODO: Need to think more about how to
 263          * handle having different CPUs at different frequencies.
 264          */
 265         if (mp_ncpus > 1 && !smp_started) {
 266                 device_printf(dev, "rejecting change, SMP not started yet\n");
 267                 error = ENXIO;
 268                 goto out;
 269         }
 270 #endif
 271 #endif /* SMP */
 272
 273         /*
 274          * If the requested level has a lower priority, don't allow
 275          * the new level right now.
 276          */
 277         if (priority < sc->curr_priority) {
 278                 CF_DEBUG("ignoring, curr prio %d less than %d\n", priority,
 279                     sc->curr_priority);
 280                 error = EPERM;
 281                 goto out;
 282         }
 283
 284         /*
 285          * If the caller didn't specify a level and one is saved, prepare to
 286          * restore the saved level.  If none has been saved, return an error.
 287          */
 288         if (level == NULL) {
 289                 saved_freq = SLIST_FIRST(&sc->saved_freq);
 290                 if (saved_freq == NULL) {
 291                         CF_DEBUG("NULL level, no saved level\n");
 292                         error = ENXIO;
 293                         goto out;
 294                 }
 295                 level = &saved_freq->level;
 296                 priority = saved_freq->priority;
 297                 CF_DEBUG("restoring saved level, freq %d prio %d\n",
 298                     level->total_set.freq, priority);
 299         }
 300
 301         /* Reject levels that are below our specified threshold. */
 302         if (level->total_set.freq < cf_lowest_freq) {
 303                 CF_DEBUG("rejecting freq %d, less than %d limit\n",
 304                     level->total_set.freq, cf_lowest_freq);
 305                 error = EINVAL;
 306                 goto out;
 307         }
 308
 309         /* If already at this level, just return. */
 310         if (sc->curr_level.total_set.freq == level->total_set.freq) {
 311                 CF_DEBUG("skipping freq %d, same as current level %d\n",
 312                     level->total_set.freq, sc->curr_level.total_set.freq);
 313                 goto skip;
 314         }
 315
 316         /* First, set the absolute frequency via its driver. */
 317         set = &level->abs_set;
 318         if (set->dev) {
 319                 if (!device_is_attached(set->dev)) {
 320                         error = ENXIO;
 321                         goto out;
 322                 }
 323
 324                 /* Bind to the target CPU before switching. */
 325                 pc = cpu_get_pcpu(set->dev);
 326
 327                 /* Skip settings if CPU is not started. */
 328                 if (pc == NULL) {
 329                         error = 0;
 330                         goto out;
 331                 }
 332                 thread_lock(curthread);
 333                 pri = curthread->td_priority;
 334                 sched_prio(curthread, PRI_MIN);
 335                 sched_bind(curthread, pc->pc_cpuid);
 336                 thread_unlock(curthread);
 337                 CF_DEBUG("setting abs freq %d on %s (cpu %d)\n", set->freq,
 338                     device_get_nameunit(set->dev), PCPU_GET(cpuid));
 339                 error = CPUFREQ_DRV_SET(set->dev, set);
 340                 thread_lock(curthread);
 341                 sched_unbind(curthread);
 342                 sched_prio(curthread, pri);
 343                 thread_unlock(curthread);
 344                 if (error) {
 345                         goto out;
 346                 }
 347         }
 348
 349         /* Next, set any/all relative frequencies via their drivers. */
 350         for (i = 0; i < level->rel_count; i++) {
 351                 set = &level->rel_set[i];
 352                 if (!device_is_attached(set->dev)) {
 353                         error = ENXIO;
 354                         goto out;
 355                 }
 356
 357                 /* Bind to the target CPU before switching. */
 358                 pc = cpu_get_pcpu(set->dev);
 359                 thread_lock(curthread);
 360                 pri = curthread->td_priority;
 361                 sched_prio(curthread, PRI_MIN);
 362                 sched_bind(curthread, pc->pc_cpuid);
 363                 thread_unlock(curthread);
 364                 CF_DEBUG("setting rel freq %d on %s (cpu %d)\n", set->freq,
 365                     device_get_nameunit(set->dev), PCPU_GET(cpuid));
 366                 error = CPUFREQ_DRV_SET(set->dev, set);
 367                 thread_lock(curthread);
 368                 sched_unbind(curthread);
 369                 sched_prio(curthread, pri);
 370                 thread_unlock(curthread);
 371                 if (error) {
 372                         /* XXX Back out any successful setting? */
 373                         goto out;
 374                 }
 375         }
 376
 377 skip:
 378         /*
 379          * Before recording the current level, check if we're going to a
 380          * higher priority.  If so, save the previous level and priority.
 381          */
 382         if (sc->curr_level.total_set.freq != CPUFREQ_VAL_UNKNOWN &&
 383             priority > sc->curr_priority) {
 384                 CF_DEBUG("saving level, freq %d prio %d\n",
 385                     sc->curr_level.total_set.freq, sc->curr_priority);
 386                 curr_freq = malloc(sizeof(*curr_freq), M_TEMP, M_NOWAIT);
 387                 if (curr_freq == NULL) {
 388                         error = ENOMEM;
 389                         goto out;
 390                 }
 391                 curr_freq->level = sc->curr_level;
 392                 curr_freq->priority = sc->curr_priority;
 393                 SLIST_INSERT_HEAD(&sc->saved_freq, curr_freq, link);
 394         }
 395         sc->curr_level = *level;
 396         sc->curr_priority = priority;
 397
 398         /* If we were restoring a saved state, reset it to "unused". */
 399         if (saved_freq != NULL) {
 400                 CF_DEBUG("resetting saved level\n");
 401                 sc->curr_level.total_set.freq = CPUFREQ_VAL_UNKNOWN;
 402                 SLIST_REMOVE_HEAD(&sc->saved_freq, link);
 403                 free(saved_freq, M_TEMP);
 404         }
 405
 406 out:
 407         CF_MTX_UNLOCK(&sc->lock);
 408
 409         /*
 410          * We changed levels (or attempted to) so notify the post-change
 411          * handler of new frequency or error.
 412          */
 413         EVENTHANDLER_INVOKE(cpufreq_post_change, level, error);
 414         if (error && set)
 415                 device_printf(set->dev, "set freq failed, err %d\n", error);
 416
 417         return (error);
 418 }
 419
 420 static int
 421 cpufreq_get_frequency(device_t dev)
 422 {
 423         struct cf_setting set;
 424
 425         if (CPUFREQ_DRV_GET(dev, &set) != 0)
 426                 return (-1);
 427
 428         return (set.freq);
 429 }
 430
 431 /* Returns the index into *levels with the match */
 432 static int
 433 cpufreq_get_level(device_t dev, struct cf_level *levels, int count)
 434 {
 435         int i, freq;
 436
 437         if ((freq = cpufreq_get_frequency(dev)) < 0)
 438                 return (-1);
 439         for (i = 0; i < count; i++)
 440                 if (freq == levels[i].total_set.freq)
 441                         return (i);
 442
 443         return (-1);
 444 }
 445
 446 /*
 447  * Used by the cpufreq core, this function will populate *level with the current
 448  * frequency as either determined by a cached value sc->curr_level, or in the
 449  * case the lower level driver has set the CPUFREQ_FLAG_UNCACHED flag, it will
 450  * obtain the frequency from the driver itself.
 451  */
 452 static int
 453 cf_get_method(device_t dev, struct cf_level *level)
 454 {
 455         struct cpufreq_softc *sc;
 456         struct cf_level *levels;
 457         struct cf_setting *curr_set;
 458         struct pcpu *pc;
 459         int bdiff, count, diff, error, i, type;
 460         uint64_t rate;
 461
 462         sc = device_get_softc(dev);
 463         error = 0;
 464         levels = NULL;
 465
 466         /*
 467          * If we already know the current frequency, and the driver didn't ask
 468          * for uncached usage, we're done.
 469          */
 470         CF_MTX_LOCK(&sc->lock);
 471         curr_set = &sc->curr_level.total_set;
 472         error = CPUFREQ_DRV_TYPE(sc->cf_drv_dev, &type);
 473         if (error == 0 && (type & CPUFREQ_FLAG_UNCACHED)) {
 474                 struct cf_setting set;
 475
 476                 /*
 477                  * If the driver wants to always report back the real frequency,
 478                  * first try the driver and if that fails, fall back to
 479                  * estimating.
 480                  */
 481                 if (CPUFREQ_DRV_GET(sc->cf_drv_dev, &set) == 0) {
 482                         sc->curr_level.total_set = set;
 483                         CF_DEBUG("get returning immediate freq %d\n",
 484                             curr_set->freq);
 485                         goto out;
 486                 }
 487         } else if (curr_set->freq != CPUFREQ_VAL_UNKNOWN) {
 488                 CF_DEBUG("get returning known freq %d\n", curr_set->freq);
 489                 error = 0;
 490                 goto out;
 491         }
 492         CF_MTX_UNLOCK(&sc->lock);
 493
 494         /*
 495          * We need to figure out the current level.  Loop through every
 496          * driver, getting the current setting.  Then, attempt to get a best
 497          * match of settings against each level.
 498          */
 499         count = CF_MAX_LEVELS;
 500         levels = malloc(count * sizeof(*levels), M_TEMP, M_NOWAIT);
 501         if (levels == NULL)
 502                 return (ENOMEM);
 503         error = CPUFREQ_LEVELS(sc->dev, levels, &count);
 504         if (error) {
 505                 if (error == E2BIG)
 506                         printf("cpufreq: need to increase CF_MAX_LEVELS\n");
 507                 free(levels, M_TEMP);
 508                 return (error);
 509         }
 510
 511         /*
 512          * Reacquire the lock and search for the given level.
 513          *
 514          * XXX Note: this is not quite right since we really need to go
 515          * through each level and compare both absolute and relative
 516          * settings for each driver in the system before making a match.
 517          * The estimation code below catches this case though.
 518          */
 519         CF_MTX_LOCK(&sc->lock);
 520         i = cpufreq_get_level(sc->cf_drv_dev, levels, count);
 521         if (i >= 0)
 522                 sc->curr_level = levels[i];
 523         else
 524                 CF_DEBUG("Couldn't find supported level for %s\n",
 525                     device_get_nameunit(sc->cf_drv_dev));
 526
 527         if (curr_set->freq != CPUFREQ_VAL_UNKNOWN) {
 528                 CF_DEBUG("get matched freq %d from drivers\n", curr_set->freq);
 529                 goto out;
 530         }
 531
 532         /*
 533          * We couldn't find an exact match, so attempt to estimate and then
 534          * match against a level.
 535          */
 536         pc = cpu_get_pcpu(dev);
 537         if (pc == NULL) {
 538                 error = ENXIO;
 539                 goto out;
 540         }
 541         cpu_est_clockrate(pc->pc_cpuid, &rate);
 542         rate /= 1000000;
 543         bdiff = 1 << 30;
 544         for (i = 0; i < count; i++) {
 545                 diff = abs(levels[i].total_set.freq - rate);
 546                 if (diff < bdiff) {
 547                         bdiff = diff;
 548                         sc->curr_level = levels[i];
 549                 }
 550         }
 551         CF_DEBUG("get estimated freq %d\n", curr_set->freq);
 552
 553 out:
 554         if (error == 0)
 555                 *level = sc->curr_level;
 556
 557         CF_MTX_UNLOCK(&sc->lock);
 558         if (levels)
 559                 free(levels, M_TEMP);
 560         return (error);
 561 }
 562
 563 /*
 564  * Either directly obtain settings from the cpufreq driver, or build a list of
 565  * relative settings to be integrated later against an absolute max.
 566  */
 567 static int
 568 cpufreq_add_levels(device_t cf_dev, struct cf_setting_lst *rel_sets)
 569 {
 570         struct cf_setting_array *set_arr;
 571         struct cf_setting *sets;
 572         device_t dev;
 573         struct cpufreq_softc *sc;
 574         int type, set_count, error;
 575
 576         sc = device_get_softc(cf_dev);
 577         dev = sc->cf_drv_dev;
 578
 579         /* Skip devices that aren't ready. */
 580         if (!device_is_attached(cf_dev))
 581                 return (0);
 582
 583         /*
 584          * Get settings, skipping drivers that offer no settings or
 585          * provide settings for informational purposes only.
 586          */
 587         error = CPUFREQ_DRV_TYPE(dev, &type);
 588         if (error != 0 || (type & CPUFREQ_FLAG_INFO_ONLY)) {
 589                 if (error == 0) {
 590                         CF_DEBUG("skipping info-only driver %s\n",
 591                             device_get_nameunit(cf_dev));
 592                 }
 593                 return (error);
 594         }
 595
 596         sets = malloc(MAX_SETTINGS * sizeof(*sets), M_TEMP, M_NOWAIT);
 597         if (sets == NULL)
 598                 return (ENOMEM);
 599
 600         set_count = MAX_SETTINGS;
 601         error = CPUFREQ_DRV_SETTINGS(dev, sets, &set_count);
 602         if (error != 0 || set_count == 0)
 603                 goto out;
 604
 605         /* Add the settings to our absolute/relative lists. */
 606         switch (type & CPUFREQ_TYPE_MASK) {
 607         case CPUFREQ_TYPE_ABSOLUTE:
 608                 error = cpufreq_insert_abs(sc, sets, set_count);
 609                 break;
 610         case CPUFREQ_TYPE_RELATIVE:
 611                 CF_DEBUG("adding %d relative settings\n", set_count);
 612                 set_arr = malloc(sizeof(*set_arr), M_TEMP, M_NOWAIT);
 613                 if (set_arr == NULL) {
 614                         error = ENOMEM;
 615                         goto out;
 616                 }
 617                 bcopy(sets, set_arr->sets, set_count * sizeof(*sets));
 618                 set_arr->count = set_count;
 619                 TAILQ_INSERT_TAIL(rel_sets, set_arr, link);
 620                 break;
 621         default:
 622                 error = EINVAL;
 623         }
 624
 625 out:
 626         free(sets, M_TEMP);
 627         return (error);
 628 }
 629
 630 static int
 631 cf_levels_method(device_t dev, struct cf_level *levels, int *count)
 632 {
 633         struct cf_setting_array *set_arr;
 634         struct cf_setting_lst rel_sets;
 635         struct cpufreq_softc *sc;
 636         struct cf_level *lev;
 637         struct pcpu *pc;
 638         int error, i;
 639         uint64_t rate;
 640
 641         if (levels == NULL || count == NULL)
 642                 return (EINVAL);
 643
 644         TAILQ_INIT(&rel_sets);
 645         sc = device_get_softc(dev);
 646
 647         CF_MTX_LOCK(&sc->lock);
 648         error = cpufreq_add_levels(sc->dev, &rel_sets);
 649         if (error)
 650                 goto out;
 651
 652         /*
 653          * If there are no absolute levels, create a fake one at 100%.  We
 654          * then cache the clockrate for later use as our base frequency.
 655          */
 656         if (TAILQ_EMPTY(&sc->all_levels)) {
 657                 struct cf_setting set;
 658
 659                 CF_DEBUG("No absolute levels returned by driver\n");
 660
 661                 if (sc->max_mhz == CPUFREQ_VAL_UNKNOWN) {
 662                         sc->max_mhz = cpu_get_nominal_mhz(dev);
 663                         /*
 664                          * If the CPU can't report a rate for 100%, hope
 665                          * the CPU is running at its nominal rate right now,
 666                          * and use that instead.
 667                          */
 668                         if (sc->max_mhz <= 0) {
 669                                 pc = cpu_get_pcpu(dev);
 670                                 cpu_est_clockrate(pc->pc_cpuid, &rate);
 671                                 sc->max_mhz = rate / 1000000;
 672                         }
 673                 }
 674                 memset(&set, CPUFREQ_VAL_UNKNOWN, sizeof(set));
 675                 set.freq = sc->max_mhz;
 676                 set.dev = NULL;
 677                 error = cpufreq_insert_abs(sc, &set, 1);
 678                 if (error)
 679                         goto out;
 680         }
 681
 682         /* Create a combined list of absolute + relative levels. */
 683         TAILQ_FOREACH(set_arr, &rel_sets, link)
 684                 cpufreq_expand_set(sc, set_arr);
 685
 686         /* If the caller doesn't have enough space, return the actual count. */
 687         if (sc->all_count > *count) {
 688                 *count = sc->all_count;
 689                 error = E2BIG;
 690                 goto out;
 691         }
 692
 693         /* Finally, output the list of levels. */
 694         i = 0;
 695         TAILQ_FOREACH(lev, &sc->all_levels, link) {
 696                 /* Skip levels that have a frequency that is too low. */
 697                 if (lev->total_set.freq < cf_lowest_freq) {
 698                         sc->all_count--;
 699                         continue;
 700                 }
 701
 702                 levels[i] = *lev;
 703                 i++;
 704         }
 705         *count = sc->all_count;
 706         error = 0;
 707
 708 out:
 709         /* Clear all levels since we regenerate them each time. */
 710         while ((lev = TAILQ_FIRST(&sc->all_levels)) != NULL) {
 711                 TAILQ_REMOVE(&sc->all_levels, lev, link);
 712                 free(lev, M_TEMP);
 713         }
 714         sc->all_count = 0;
 715
 716         CF_MTX_UNLOCK(&sc->lock);
 717         while ((set_arr = TAILQ_FIRST(&rel_sets)) != NULL) {
 718                 TAILQ_REMOVE(&rel_sets, set_arr, link);
 719                 free(set_arr, M_TEMP);
 720         }
 721         return (error);
 722 }
 723
 724 /*
 725  * Create levels for an array of absolute settings and insert them in
 726  * sorted order in the specified list.
 727  */
 728 static int
 729 cpufreq_insert_abs(struct cpufreq_softc *sc, struct cf_setting *sets,
 730     int count)
 731 {
 732         struct cf_level_lst *list;
 733         struct cf_level *level, *search;
 734         int i, inserted;
 735
 736         CF_MTX_ASSERT(&sc->lock);
 737
 738         list = &sc->all_levels;
 739         for (i = 0; i < count; i++) {
 740                 level = malloc(sizeof(*level), M_TEMP, M_NOWAIT | M_ZERO);
 741                 if (level == NULL)
 742                         return (ENOMEM);
 743                 level->abs_set = sets[i];
 744                 level->total_set = sets[i];
 745                 level->total_set.dev = NULL;
 746                 sc->all_count++;
 747                 inserted = 0;
 748
 749                 if (TAILQ_EMPTY(list)) {
 750                         CF_DEBUG("adding abs setting %d at head\n",
 751                             sets[i].freq);
 752                         TAILQ_INSERT_HEAD(list, level, link);
 753                         continue;
 754                 }
 755
 756                 TAILQ_FOREACH_REVERSE(search, list, cf_level_lst, link)
 757                         if (sets[i].freq <= search->total_set.freq) {
 758                                 CF_DEBUG("adding abs setting %d after %d\n",
 759                                     sets[i].freq, search->total_set.freq);
 760                                 TAILQ_INSERT_AFTER(list, search, level, link);
 761                                 inserted = 1;
 762                                 break;
 763                         }
 764
 765                 if (inserted == 0) {
 766                         TAILQ_FOREACH(search, list, link)
 767                                 if (sets[i].freq >= search->total_set.freq) {
 768                                         CF_DEBUG("adding abs setting %d before %d\n",
 769                                             sets[i].freq, search->total_set.freq);
 770                                         TAILQ_INSERT_BEFORE(search, level, link);
 771                                         break;
 772                                 }
 773                 }
 774         }
 775
 776         return (0);
 777 }
 778
 779 /*
 780  * Expand a group of relative settings, creating derived levels from them.
 781  */
 782 static int
 783 cpufreq_expand_set(struct cpufreq_softc *sc, struct cf_setting_array *set_arr)
 784 {
 785         struct cf_level *fill, *search;
 786         struct cf_setting *set;
 787         int i;
 788
 789         CF_MTX_ASSERT(&sc->lock);
 790
 791         /*
 792          * Walk the set of all existing levels in reverse.  This is so we
 793          * create derived states from the lowest absolute settings first
 794          * and discard duplicates created from higher absolute settings.
 795          * For instance, a level of 50 Mhz derived from 100 Mhz + 50% is
 796          * preferable to 200 Mhz + 25% because absolute settings are more
 797          * efficient since they often change the voltage as well.
 798          */
 799         TAILQ_FOREACH_REVERSE(search, &sc->all_levels, cf_level_lst, link) {
 800                 /* Add each setting to the level, duplicating if necessary. */
 801                 for (i = 0; i < set_arr->count; i++) {
 802                         set = &set_arr->sets[i];
 803
 804                         /*
 805                          * If this setting is less than 100%, split the level
 806                          * into two and add this setting to the new level.
 807                          */
 808                         fill = search;
 809                         if (set->freq < 10000) {
 810                                 fill = cpufreq_dup_set(sc, search, set);
 811
 812                                 /*
 813                                  * The new level was a duplicate of an existing
 814                                  * level or its absolute setting is too high
 815                                  * so we freed it.  For example, we discard a
 816                                  * derived level of 1000 MHz/25% if a level
 817                                  * of 500 MHz/100% already exists.
 818                                  */
 819                                 if (fill == NULL)
 820                                         break;
 821                         }
 822
 823                         /* Add this setting to the existing or new level. */
 824                         KASSERT(fill->rel_count < MAX_SETTINGS,
 825                             ("cpufreq: too many relative drivers (%d)",
 826                             MAX_SETTINGS));
 827                         fill->rel_set[fill->rel_count] = *set;
 828                         fill->rel_count++;
 829                         CF_DEBUG(
 830                         "expand set added rel setting %d%% to %d level\n",
 831                             set->freq / 100, fill->total_set.freq);
 832                 }
 833         }
 834
 835         return (0);
 836 }
 837
 838 static struct cf_level *
 839 cpufreq_dup_set(struct cpufreq_softc *sc, struct cf_level *dup,
 840     struct cf_setting *set)
 841 {
 842         struct cf_level_lst *list;
 843         struct cf_level *fill, *itr;
 844         struct cf_setting *fill_set, *itr_set;
 845         int i;
 846
 847         CF_MTX_ASSERT(&sc->lock);
 848
 849         /*
 850          * Create a new level, copy it from the old one, and update the
 851          * total frequency and power by the percentage specified in the
 852          * relative setting.
 853          */
 854         fill = malloc(sizeof(*fill), M_TEMP, M_NOWAIT);
 855         if (fill == NULL)
 856                 return (NULL);
 857         *fill = *dup;
 858         fill_set = &fill->total_set;
 859         fill_set->freq =
 860             ((uint64_t)fill_set->freq * set->freq) / 10000;
 861         if (fill_set->power != CPUFREQ_VAL_UNKNOWN) {
 862                 fill_set->power = ((uint64_t)fill_set->power * set->freq)
 863                     / 10000;
 864         }
 865         if (set->lat != CPUFREQ_VAL_UNKNOWN) {
 866                 if (fill_set->lat != CPUFREQ_VAL_UNKNOWN)
 867                         fill_set->lat += set->lat;
 868                 else
 869                         fill_set->lat = set->lat;
 870         }
 871         CF_DEBUG("dup set considering derived setting %d\n", fill_set->freq);
 872
 873         /*
 874          * If we copied an old level that we already modified (say, at 100%),
 875          * we need to remove that setting before adding this one.  Since we
 876          * process each setting array in order, we know any settings for this
 877          * driver will be found at the end.
 878          */
 879         for (i = fill->rel_count; i != 0; i--) {
 880                 if (fill->rel_set[i - 1].dev != set->dev)
 881                         break;
 882                 CF_DEBUG("removed last relative driver: %s\n",
 883                     device_get_nameunit(set->dev));
 884                 fill->rel_count--;
 885         }
 886
 887         /*
 888          * Insert the new level in sorted order.  If it is a duplicate of an
 889          * existing level (1) or has an absolute setting higher than the
 890          * existing level (2), do not add it.  We can do this since any such
 891          * level is guaranteed use less power.  For example (1), a level with
 892          * one absolute setting of 800 Mhz uses less power than one composed
 893          * of an absolute setting of 1600 Mhz and a relative setting at 50%.
 894          * Also for example (2), a level of 800 Mhz/75% is preferable to
 895          * 1600 Mhz/25% even though the latter has a lower total frequency.
 896          */
 897         list = &sc->all_levels;
 898         KASSERT(!TAILQ_EMPTY(list), ("all levels list empty in dup set"));
 899         TAILQ_FOREACH_REVERSE(itr, list, cf_level_lst, link) {
 900                 itr_set = &itr->total_set;
 901                 if (CPUFREQ_CMP(fill_set->freq, itr_set->freq)) {
 902                         CF_DEBUG("dup set rejecting %d (dupe)\n",
 903                             fill_set->freq);
 904                         itr = NULL;
 905                         break;
 906                 } else if (fill_set->freq < itr_set->freq) {
 907                         if (fill->abs_set.freq <= itr->abs_set.freq) {
 908                                 CF_DEBUG(
 909                         "dup done, inserting new level %d after %d\n",
 910                                     fill_set->freq, itr_set->freq);
 911                                 TAILQ_INSERT_AFTER(list, itr, fill, link);
 912                                 sc->all_count++;
 913                         } else {
 914                                 CF_DEBUG("dup set rejecting %d (abs too big)\n",
 915                                     fill_set->freq);
 916                                 itr = NULL;
 917                         }
 918                         break;
 919                 }
 920         }
 921
 922         /* We didn't find a good place for this new level so free it. */
 923         if (itr == NULL) {
 924                 CF_DEBUG("dup set freeing new level %d (not optimal)\n",
 925                     fill_set->freq);
 926                 free(fill, M_TEMP);
 927                 fill = NULL;
 928         }
 929
 930         return (fill);
 931 }
 932
 933 static int
 934 cpufreq_curr_sysctl(SYSCTL_HANDLER_ARGS)
 935 {
 936         struct cpufreq_softc *sc;
 937         struct cf_level *levels;
 938         int best, count, diff, bdiff, devcount, error, freq, i, n;
 939         device_t *devs;
 940
 941         devs = NULL;
 942         sc = oidp->oid_arg1;
 943         levels = sc->levels_buf;
 944
 945         error = CPUFREQ_GET(sc->dev, &levels[0]);
 946         if (error)
 947                 goto out;
 948         freq = levels[0].total_set.freq;
 949         error = sysctl_handle_int(oidp, &freq, 0, req);
 950         if (error != 0 || req->newptr == NULL)
 951                 goto out;
 952
 953         /*
 954          * While we only call cpufreq_get() on one device (assuming all
 955          * CPUs have equal levels), we call cpufreq_set() on all CPUs.
 956          * This is needed for some MP systems.
 957          */
 958         error = devclass_get_devices(cpufreq_dc, &devs, &devcount);
 959         if (error)
 960                 goto out;
 961         for (n = 0; n < devcount; n++) {
 962                 count = CF_MAX_LEVELS;
 963                 error = CPUFREQ_LEVELS(devs[n], levels, &count);
 964                 if (error) {
 965                         if (error == E2BIG)
 966                                 printf(
 967                         "cpufreq: need to increase CF_MAX_LEVELS\n");
 968                         break;
 969                 }
 970                 best = 0;
 971                 bdiff = 1 << 30;
 972                 for (i = 0; i < count; i++) {
 973                         diff = abs(levels[i].total_set.freq - freq);
 974                         if (diff < bdiff) {
 975                                 bdiff = diff;
 976                                 best = i;
 977                         }
 978                 }
 979                 error = CPUFREQ_SET(devs[n], &levels[best], CPUFREQ_PRIO_USER);
 980         }
 981
 982 out:
 983         if (devs)
 984                 free(devs, M_TEMP);
 985         return (error);
 986 }
 987
 988 static int
 989 cpufreq_levels_sysctl(SYSCTL_HANDLER_ARGS)
 990 {
 991         struct cpufreq_softc *sc;
 992         struct cf_level *levels;
 993         struct cf_setting *set;
 994         struct sbuf sb;
 995         int count, error, i;
 996
 997         sc = oidp->oid_arg1;
 998         sbuf_new(&sb, NULL, 128, SBUF_AUTOEXTEND);
 999
1000         /* Get settings from the device and generate the output string. */
1001         count = CF_MAX_LEVELS;
1002         levels = sc->levels_buf;
1003         if (levels == NULL) {
1004                 sbuf_delete(&sb);
1005                 return (ENOMEM);
1006         }
1007         error = CPUFREQ_LEVELS(sc->dev, levels, &count);
1008         if (error) {
1009                 if (error == E2BIG)
1010                         printf("cpufreq: need to increase CF_MAX_LEVELS\n");
1011                 goto out;
1012         }
1013         if (count) {
1014                 for (i = 0; i < count; i++) {
1015                         set = &levels[i].total_set;
1016                         sbuf_printf(&sb, "%d/%d ", set->freq, set->power);
1017                 }
1018         } else
1019                 sbuf_cpy(&sb, "0");
1020         sbuf_trim(&sb);
1021         sbuf_finish(&sb);
1022         error = sysctl_handle_string(oidp, sbuf_data(&sb), sbuf_len(&sb), req);
1023
1024 out:
1025         sbuf_delete(&sb);
1026         return (error);
1027 }
1028
1029 static int
1030 cpufreq_settings_sysctl(SYSCTL_HANDLER_ARGS)
1031 {
1032         device_t dev;
1033         struct cf_setting *sets;
1034         struct sbuf sb;
1035         int error, i, set_count;
1036
1037         dev = oidp->oid_arg1;
1038         sbuf_new(&sb, NULL, 128, SBUF_AUTOEXTEND);
1039
1040         /* Get settings from the device and generate the output string. */
1041         set_count = MAX_SETTINGS;
1042         sets = malloc(set_count * sizeof(*sets), M_TEMP, M_NOWAIT);
1043         if (sets == NULL) {
1044                 sbuf_delete(&sb);
1045                 return (ENOMEM);
1046         }
1047         error = CPUFREQ_DRV_SETTINGS(dev, sets, &set_count);
1048         if (error)
1049                 goto out;
1050         if (set_count) {
1051                 for (i = 0; i < set_count; i++)
1052                         sbuf_printf(&sb, "%d/%d ", sets[i].freq, sets[i].power);
1053         } else
1054                 sbuf_cpy(&sb, "0");
1055         sbuf_trim(&sb);
1056         sbuf_finish(&sb);
1057         error = sysctl_handle_string(oidp, sbuf_data(&sb), sbuf_len(&sb), req);
1058
1059 out:
1060         free(sets, M_TEMP);
1061         sbuf_delete(&sb);
1062         return (error);
1063 }
1064
1065 static void
1066 cpufreq_add_freq_driver_sysctl(device_t cf_dev)
1067 {
1068         struct cpufreq_softc *sc;
1069
1070         sc = device_get_softc(cf_dev);
1071         SYSCTL_ADD_CONST_STRING(&sc->sysctl_ctx,
1072             SYSCTL_CHILDREN(device_get_sysctl_tree(cf_dev)), OID_AUTO,
1073             "freq_driver", CTLFLAG_RD, device_get_nameunit(sc->cf_drv_dev),
1074             "cpufreq driver used by this cpu");
1075 }
1076
1077 int
1078 cpufreq_register(device_t dev)
1079 {
1080         struct cpufreq_softc *sc;
1081         device_t cf_dev, cpu_dev;
1082         int error;
1083
1084         /* Add a sysctl to get each driver's settings separately. */
1085         SYSCTL_ADD_PROC(device_get_sysctl_ctx(dev),
1086             SYSCTL_CHILDREN(device_get_sysctl_tree(dev)),
1087             OID_AUTO, "freq_settings",
1088             CTLTYPE_STRING | CTLFLAG_RD | CTLFLAG_NEEDGIANT, dev, 0,
1089             cpufreq_settings_sysctl, "A", "CPU frequency driver settings");
1090
1091         /*
1092          * Add only one cpufreq device to each CPU.  Currently, all CPUs
1093          * must offer the same levels and be switched at the same time.
1094          */
1095         cpu_dev = device_get_parent(dev);
1096         if ((cf_dev = device_find_child(cpu_dev, "cpufreq", -1))) {
1097                 sc = device_get_softc(cf_dev);
1098                 sc->max_mhz = CPUFREQ_VAL_UNKNOWN;
1099                 MPASS(sc->cf_drv_dev != NULL);
1100                 return (0);
1101         }
1102
1103         /* Add the child device and possibly sysctls. */
1104         cf_dev = BUS_ADD_CHILD(cpu_dev, 0, "cpufreq", device_get_unit(cpu_dev));
1105         if (cf_dev == NULL)
1106                 return (ENOMEM);
1107         device_quiet(cf_dev);
1108
1109         error = device_probe_and_attach(cf_dev);
1110         if (error)
1111                 return (error);
1112
1113         sc = device_get_softc(cf_dev);
1114         sc->cf_drv_dev = dev;
1115         cpufreq_add_freq_driver_sysctl(cf_dev);
1116         return (error);
1117 }
1118
1119 int
1120 cpufreq_unregister(device_t dev)
1121 {
1122         device_t cf_dev;
1123         struct cpufreq_softc *sc;
1124
1125         /*
1126          * If this is the last cpufreq child device, remove the control
1127          * device as well.  We identify cpufreq children by calling a method
1128          * they support.
1129          */
1130         cf_dev = device_find_child(device_get_parent(dev), "cpufreq", -1);
1131         if (cf_dev == NULL) {
1132                 device_printf(dev,
1133         "warning: cpufreq_unregister called with no cpufreq device active\n");
1134                 return (0);
1135         }
1136         sc = device_get_softc(cf_dev);
1137         MPASS(sc->cf_drv_dev == dev);
1138         device_delete_child(device_get_parent(cf_dev), cf_dev);
1139
1140         return (0);
1141 }
1142
1143 int
1144 cpufreq_settings_changed(device_t dev)
1145 {
1146
1147         EVENTHANDLER_INVOKE(cpufreq_levels_changed,
1148             device_get_unit(device_get_parent(dev)));
1149         return (0);
1150 }