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 __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 }; 131 132 static driver_t cpufreq_driver = { 133 "cpufreq", cpufreq_methods, sizeof(struct cpufreq_softc) 134 }; 135 136 DRIVER_MODULE(cpufreq, cpu, cpufreq_driver, 0, 0); 137 138 static int cf_lowest_freq; 139 static int cf_verbose; 140 static SYSCTL_NODE(_debug, OID_AUTO, cpufreq, CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, 141 "cpufreq debugging"); 142 SYSCTL_INT(_debug_cpufreq, OID_AUTO, lowest, CTLFLAG_RWTUN, &cf_lowest_freq, 1, 143 "Don't provide levels below this frequency."); 144 SYSCTL_INT(_debug_cpufreq, OID_AUTO, verbose, CTLFLAG_RWTUN, &cf_verbose, 1, 145 "Print verbose debugging messages"); 146 147 /* 148 * This is called as the result of a hardware specific frequency control driver 149 * calling cpufreq_register. It provides a general interface for system wide 150 * frequency controls and operates on a per cpu basis. 151 */ 152 static int 153 cpufreq_attach(device_t dev) 154 { 155 struct cpufreq_softc *sc; 156 struct pcpu *pc; 157 device_t parent; 158 uint64_t rate; 159 160 CF_DEBUG("initializing %s\n", device_get_nameunit(dev)); 161 sc = device_get_softc(dev); 162 parent = device_get_parent(dev); 163 sc->dev = dev; 164 sysctl_ctx_init(&sc->sysctl_ctx); 165 TAILQ_INIT(&sc->all_levels); 166 CF_MTX_INIT(&sc->lock); 167 sc->curr_level.total_set.freq = CPUFREQ_VAL_UNKNOWN; 168 SLIST_INIT(&sc->saved_freq); 169 /* Try to get nominal CPU freq to use it as maximum later if needed */ 170 sc->max_mhz = cpu_get_nominal_mhz(dev); 171 /* If that fails, try to measure the current rate */ 172 if (sc->max_mhz <= 0) { 173 CF_DEBUG("Unable to obtain nominal frequency.\n"); 174 pc = cpu_get_pcpu(dev); 175 if (cpu_est_clockrate(pc->pc_cpuid, &rate) == 0) 176 sc->max_mhz = rate / 1000000; 177 else 178 sc->max_mhz = CPUFREQ_VAL_UNKNOWN; 179 } 180 181 CF_DEBUG("initializing one-time data for %s\n", 182 device_get_nameunit(dev)); 183 sc->levels_buf = malloc(CF_MAX_LEVELS * sizeof(*sc->levels_buf), 184 M_DEVBUF, M_WAITOK); 185 SYSCTL_ADD_PROC(&sc->sysctl_ctx, 186 SYSCTL_CHILDREN(device_get_sysctl_tree(parent)), 187 OID_AUTO, "freq", CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_NEEDGIANT, 188 sc, 0, cpufreq_curr_sysctl, "I", "Current CPU frequency"); 189 SYSCTL_ADD_PROC(&sc->sysctl_ctx, 190 SYSCTL_CHILDREN(device_get_sysctl_tree(parent)), 191 OID_AUTO, "freq_levels", 192 CTLTYPE_STRING | CTLFLAG_RD | CTLFLAG_NEEDGIANT, sc, 0, 193 cpufreq_levels_sysctl, "A", "CPU frequency levels"); 194 195 /* 196 * Queue a one-shot broadcast that levels have changed. 197 * It will run once the system has completed booting. 198 */ 199 TASK_INIT(&sc->startup_task, 0, cpufreq_startup_task, dev); 200 taskqueue_enqueue(taskqueue_thread, &sc->startup_task); 201 202 return (0); 203 } 204 205 /* Handle any work to be done for all drivers that attached during boot. */ 206 static void 207 cpufreq_startup_task(void *ctx, int pending) 208 { 209 210 cpufreq_settings_changed((device_t)ctx); 211 } 212 213 static int 214 cpufreq_detach(device_t dev) 215 { 216 struct cpufreq_softc *sc; 217 struct cf_saved_freq *saved_freq; 218 219 CF_DEBUG("shutdown %s\n", device_get_nameunit(dev)); 220 sc = device_get_softc(dev); 221 sysctl_ctx_free(&sc->sysctl_ctx); 222 223 while ((saved_freq = SLIST_FIRST(&sc->saved_freq)) != NULL) { 224 SLIST_REMOVE_HEAD(&sc->saved_freq, link); 225 free(saved_freq, M_TEMP); 226 } 227 228 free(sc->levels_buf, M_DEVBUF); 229 230 return (0); 231 } 232 233 static int 234 cf_set_method(device_t dev, const struct cf_level *level, int priority) 235 { 236 struct cpufreq_softc *sc; 237 const struct cf_setting *set; 238 struct cf_saved_freq *saved_freq, *curr_freq; 239 struct pcpu *pc; 240 int error, i; 241 u_char pri; 242 243 sc = device_get_softc(dev); 244 error = 0; 245 set = NULL; 246 saved_freq = NULL; 247 248 /* We are going to change levels so notify the pre-change handler. */ 249 EVENTHANDLER_INVOKE(cpufreq_pre_change, level, &error); 250 if (error != 0) { 251 EVENTHANDLER_INVOKE(cpufreq_post_change, level, error); 252 return (error); 253 } 254 255 CF_MTX_LOCK(&sc->lock); 256 257 #ifdef SMP 258 #ifdef EARLY_AP_STARTUP 259 MPASS(mp_ncpus == 1 || smp_started); 260 #else 261 /* 262 * If still booting and secondary CPUs not started yet, don't allow 263 * changing the frequency until they're online. This is because we 264 * can't switch to them using sched_bind() and thus we'd only be 265 * switching the main CPU. XXXTODO: Need to think more about how to 266 * handle having different CPUs at different frequencies. 267 */ 268 if (mp_ncpus > 1 && !smp_started) { 269 device_printf(dev, "rejecting change, SMP not started yet\n"); 270 error = ENXIO; 271 goto out; 272 } 273 #endif 274 #endif /* SMP */ 275 276 /* 277 * If the requested level has a lower priority, don't allow 278 * the new level right now. 279 */ 280 if (priority < sc->curr_priority) { 281 CF_DEBUG("ignoring, curr prio %d less than %d\n", priority, 282 sc->curr_priority); 283 error = EPERM; 284 goto out; 285 } 286 287 /* 288 * If the caller didn't specify a level and one is saved, prepare to 289 * restore the saved level. If none has been saved, return an error. 290 */ 291 if (level == NULL) { 292 saved_freq = SLIST_FIRST(&sc->saved_freq); 293 if (saved_freq == NULL) { 294 CF_DEBUG("NULL level, no saved level\n"); 295 error = ENXIO; 296 goto out; 297 } 298 level = &saved_freq->level; 299 priority = saved_freq->priority; 300 CF_DEBUG("restoring saved level, freq %d prio %d\n", 301 level->total_set.freq, priority); 302 } 303 304 /* Reject levels that are below our specified threshold. */ 305 if (level->total_set.freq < cf_lowest_freq) { 306 CF_DEBUG("rejecting freq %d, less than %d limit\n", 307 level->total_set.freq, cf_lowest_freq); 308 error = EINVAL; 309 goto out; 310 } 311 312 /* If already at this level, just return. */ 313 if (sc->curr_level.total_set.freq == level->total_set.freq) { 314 CF_DEBUG("skipping freq %d, same as current level %d\n", 315 level->total_set.freq, sc->curr_level.total_set.freq); 316 goto skip; 317 } 318 319 /* First, set the absolute frequency via its driver. */ 320 set = &level->abs_set; 321 if (set->dev) { 322 if (!device_is_attached(set->dev)) { 323 error = ENXIO; 324 goto out; 325 } 326 327 /* Bind to the target CPU before switching. */ 328 pc = cpu_get_pcpu(set->dev); 329 330 /* Skip settings if CPU is not started. */ 331 if (pc == NULL) { 332 error = 0; 333 goto out; 334 } 335 thread_lock(curthread); 336 pri = curthread->td_priority; 337 sched_prio(curthread, PRI_MIN); 338 sched_bind(curthread, pc->pc_cpuid); 339 thread_unlock(curthread); 340 CF_DEBUG("setting abs freq %d on %s (cpu %d)\n", set->freq, 341 device_get_nameunit(set->dev), PCPU_GET(cpuid)); 342 error = CPUFREQ_DRV_SET(set->dev, set); 343 thread_lock(curthread); 344 sched_unbind(curthread); 345 sched_prio(curthread, pri); 346 thread_unlock(curthread); 347 if (error) { 348 goto out; 349 } 350 } 351 352 /* Next, set any/all relative frequencies via their drivers. */ 353 for (i = 0; i < level->rel_count; i++) { 354 set = &level->rel_set[i]; 355 if (!device_is_attached(set->dev)) { 356 error = ENXIO; 357 goto out; 358 } 359 360 /* Bind to the target CPU before switching. */ 361 pc = cpu_get_pcpu(set->dev); 362 thread_lock(curthread); 363 pri = curthread->td_priority; 364 sched_prio(curthread, PRI_MIN); 365 sched_bind(curthread, pc->pc_cpuid); 366 thread_unlock(curthread); 367 CF_DEBUG("setting rel freq %d on %s (cpu %d)\n", set->freq, 368 device_get_nameunit(set->dev), PCPU_GET(cpuid)); 369 error = CPUFREQ_DRV_SET(set->dev, set); 370 thread_lock(curthread); 371 sched_unbind(curthread); 372 sched_prio(curthread, pri); 373 thread_unlock(curthread); 374 if (error) { 375 /* XXX Back out any successful setting? */ 376 goto out; 377 } 378 } 379 380 skip: 381 /* 382 * Before recording the current level, check if we're going to a 383 * higher priority. If so, save the previous level and priority. 384 */ 385 if (sc->curr_level.total_set.freq != CPUFREQ_VAL_UNKNOWN && 386 priority > sc->curr_priority) { 387 CF_DEBUG("saving level, freq %d prio %d\n", 388 sc->curr_level.total_set.freq, sc->curr_priority); 389 curr_freq = malloc(sizeof(*curr_freq), M_TEMP, M_NOWAIT); 390 if (curr_freq == NULL) { 391 error = ENOMEM; 392 goto out; 393 } 394 curr_freq->level = sc->curr_level; 395 curr_freq->priority = sc->curr_priority; 396 SLIST_INSERT_HEAD(&sc->saved_freq, curr_freq, link); 397 } 398 sc->curr_level = *level; 399 sc->curr_priority = priority; 400 401 /* If we were restoring a saved state, reset it to "unused". */ 402 if (saved_freq != NULL) { 403 CF_DEBUG("resetting saved level\n"); 404 sc->curr_level.total_set.freq = CPUFREQ_VAL_UNKNOWN; 405 SLIST_REMOVE_HEAD(&sc->saved_freq, link); 406 free(saved_freq, M_TEMP); 407 } 408 409 out: 410 CF_MTX_UNLOCK(&sc->lock); 411 412 /* 413 * We changed levels (or attempted to) so notify the post-change 414 * handler of new frequency or error. 415 */ 416 EVENTHANDLER_INVOKE(cpufreq_post_change, level, error); 417 if (error && set) 418 device_printf(set->dev, "set freq failed, err %d\n", error); 419 420 return (error); 421 } 422 423 static int 424 cpufreq_get_frequency(device_t dev) 425 { 426 struct cf_setting set; 427 428 if (CPUFREQ_DRV_GET(dev, &set) != 0) 429 return (-1); 430 431 return (set.freq); 432 } 433 434 /* Returns the index into *levels with the match */ 435 static int 436 cpufreq_get_level(device_t dev, struct cf_level *levels, int count) 437 { 438 int i, freq; 439 440 if ((freq = cpufreq_get_frequency(dev)) < 0) 441 return (-1); 442 for (i = 0; i < count; i++) 443 if (freq == levels[i].total_set.freq) 444 return (i); 445 446 return (-1); 447 } 448 449 /* 450 * Used by the cpufreq core, this function will populate *level with the current 451 * frequency as either determined by a cached value sc->curr_level, or in the 452 * case the lower level driver has set the CPUFREQ_FLAG_UNCACHED flag, it will 453 * obtain the frequency from the driver itself. 454 */ 455 static int 456 cf_get_method(device_t dev, struct cf_level *level) 457 { 458 struct cpufreq_softc *sc; 459 struct cf_level *levels; 460 struct cf_setting *curr_set; 461 struct pcpu *pc; 462 int bdiff, count, diff, error, i, type; 463 uint64_t rate; 464 465 sc = device_get_softc(dev); 466 error = 0; 467 levels = NULL; 468 469 /* 470 * If we already know the current frequency, and the driver didn't ask 471 * for uncached usage, we're done. 472 */ 473 CF_MTX_LOCK(&sc->lock); 474 curr_set = &sc->curr_level.total_set; 475 error = CPUFREQ_DRV_TYPE(sc->cf_drv_dev, &type); 476 if (error == 0 && (type & CPUFREQ_FLAG_UNCACHED)) { 477 struct cf_setting set; 478 479 /* 480 * If the driver wants to always report back the real frequency, 481 * first try the driver and if that fails, fall back to 482 * estimating. 483 */ 484 if (CPUFREQ_DRV_GET(sc->cf_drv_dev, &set) == 0) { 485 sc->curr_level.total_set = set; 486 CF_DEBUG("get returning immediate freq %d\n", 487 curr_set->freq); 488 goto out; 489 } 490 } else if (curr_set->freq != CPUFREQ_VAL_UNKNOWN) { 491 CF_DEBUG("get returning known freq %d\n", curr_set->freq); 492 error = 0; 493 goto out; 494 } 495 CF_MTX_UNLOCK(&sc->lock); 496 497 /* 498 * We need to figure out the current level. Loop through every 499 * driver, getting the current setting. Then, attempt to get a best 500 * match of settings against each level. 501 */ 502 count = CF_MAX_LEVELS; 503 levels = malloc(count * sizeof(*levels), M_TEMP, M_NOWAIT); 504 if (levels == NULL) 505 return (ENOMEM); 506 error = CPUFREQ_LEVELS(sc->dev, levels, &count); 507 if (error) { 508 if (error == E2BIG) 509 printf("cpufreq: need to increase CF_MAX_LEVELS\n"); 510 free(levels, M_TEMP); 511 return (error); 512 } 513 514 /* 515 * Reacquire the lock and search for the given level. 516 * 517 * XXX Note: this is not quite right since we really need to go 518 * through each level and compare both absolute and relative 519 * settings for each driver in the system before making a match. 520 * The estimation code below catches this case though. 521 */ 522 CF_MTX_LOCK(&sc->lock); 523 i = cpufreq_get_level(sc->cf_drv_dev, levels, count); 524 if (i >= 0) 525 sc->curr_level = levels[i]; 526 else 527 CF_DEBUG("Couldn't find supported level for %s\n", 528 device_get_nameunit(sc->cf_drv_dev)); 529 530 if (curr_set->freq != CPUFREQ_VAL_UNKNOWN) { 531 CF_DEBUG("get matched freq %d from drivers\n", curr_set->freq); 532 goto out; 533 } 534 535 /* 536 * We couldn't find an exact match, so attempt to estimate and then 537 * match against a level. 538 */ 539 pc = cpu_get_pcpu(dev); 540 if (pc == NULL) { 541 error = ENXIO; 542 goto out; 543 } 544 cpu_est_clockrate(pc->pc_cpuid, &rate); 545 rate /= 1000000; 546 bdiff = 1 << 30; 547 for (i = 0; i < count; i++) { 548 diff = abs(levels[i].total_set.freq - rate); 549 if (diff < bdiff) { 550 bdiff = diff; 551 sc->curr_level = levels[i]; 552 } 553 } 554 CF_DEBUG("get estimated freq %d\n", curr_set->freq); 555 556 out: 557 if (error == 0) 558 *level = sc->curr_level; 559 560 CF_MTX_UNLOCK(&sc->lock); 561 if (levels) 562 free(levels, M_TEMP); 563 return (error); 564 } 565 566 /* 567 * Either directly obtain settings from the cpufreq driver, or build a list of 568 * relative settings to be integrated later against an absolute max. 569 */ 570 static int 571 cpufreq_add_levels(device_t cf_dev, struct cf_setting_lst *rel_sets) 572 { 573 struct cf_setting_array *set_arr; 574 struct cf_setting *sets; 575 device_t dev; 576 struct cpufreq_softc *sc; 577 int type, set_count, error; 578 579 sc = device_get_softc(cf_dev); 580 dev = sc->cf_drv_dev; 581 582 /* Skip devices that aren't ready. */ 583 if (!device_is_attached(cf_dev)) 584 return (0); 585 586 /* 587 * Get settings, skipping drivers that offer no settings or 588 * provide settings for informational purposes only. 589 */ 590 error = CPUFREQ_DRV_TYPE(dev, &type); 591 if (error != 0 || (type & CPUFREQ_FLAG_INFO_ONLY)) { 592 if (error == 0) { 593 CF_DEBUG("skipping info-only driver %s\n", 594 device_get_nameunit(cf_dev)); 595 } 596 return (error); 597 } 598 599 sets = malloc(MAX_SETTINGS * sizeof(*sets), M_TEMP, M_NOWAIT); 600 if (sets == NULL) 601 return (ENOMEM); 602 603 set_count = MAX_SETTINGS; 604 error = CPUFREQ_DRV_SETTINGS(dev, sets, &set_count); 605 if (error != 0 || set_count == 0) 606 goto out; 607 608 /* Add the settings to our absolute/relative lists. */ 609 switch (type & CPUFREQ_TYPE_MASK) { 610 case CPUFREQ_TYPE_ABSOLUTE: 611 error = cpufreq_insert_abs(sc, sets, set_count); 612 break; 613 case CPUFREQ_TYPE_RELATIVE: 614 CF_DEBUG("adding %d relative settings\n", set_count); 615 set_arr = malloc(sizeof(*set_arr), M_TEMP, M_NOWAIT); 616 if (set_arr == NULL) { 617 error = ENOMEM; 618 goto out; 619 } 620 bcopy(sets, set_arr->sets, set_count * sizeof(*sets)); 621 set_arr->count = set_count; 622 TAILQ_INSERT_TAIL(rel_sets, set_arr, link); 623 break; 624 default: 625 error = EINVAL; 626 } 627 628 out: 629 free(sets, M_TEMP); 630 return (error); 631 } 632 633 static int 634 cf_levels_method(device_t dev, struct cf_level *levels, int *count) 635 { 636 struct cf_setting_array *set_arr; 637 struct cf_setting_lst rel_sets; 638 struct cpufreq_softc *sc; 639 struct cf_level *lev; 640 struct pcpu *pc; 641 int error, i; 642 uint64_t rate; 643 644 if (levels == NULL || count == NULL) 645 return (EINVAL); 646 647 TAILQ_INIT(&rel_sets); 648 sc = device_get_softc(dev); 649 650 CF_MTX_LOCK(&sc->lock); 651 error = cpufreq_add_levels(sc->dev, &rel_sets); 652 if (error) 653 goto out; 654 655 /* 656 * If there are no absolute levels, create a fake one at 100%. We 657 * then cache the clockrate for later use as our base frequency. 658 */ 659 if (TAILQ_EMPTY(&sc->all_levels)) { 660 struct cf_setting set; 661 662 CF_DEBUG("No absolute levels returned by driver\n"); 663 664 if (sc->max_mhz == CPUFREQ_VAL_UNKNOWN) { 665 sc->max_mhz = cpu_get_nominal_mhz(dev); 666 /* 667 * If the CPU can't report a rate for 100%, hope 668 * the CPU is running at its nominal rate right now, 669 * and use that instead. 670 */ 671 if (sc->max_mhz <= 0) { 672 pc = cpu_get_pcpu(dev); 673 cpu_est_clockrate(pc->pc_cpuid, &rate); 674 sc->max_mhz = rate / 1000000; 675 } 676 } 677 memset(&set, CPUFREQ_VAL_UNKNOWN, sizeof(set)); 678 set.freq = sc->max_mhz; 679 set.dev = NULL; 680 error = cpufreq_insert_abs(sc, &set, 1); 681 if (error) 682 goto out; 683 } 684 685 /* Create a combined list of absolute + relative levels. */ 686 TAILQ_FOREACH(set_arr, &rel_sets, link) 687 cpufreq_expand_set(sc, set_arr); 688 689 /* If the caller doesn't have enough space, return the actual count. */ 690 if (sc->all_count > *count) { 691 *count = sc->all_count; 692 error = E2BIG; 693 goto out; 694 } 695 696 /* Finally, output the list of levels. */ 697 i = 0; 698 TAILQ_FOREACH(lev, &sc->all_levels, link) { 699 /* Skip levels that have a frequency that is too low. */ 700 if (lev->total_set.freq < cf_lowest_freq) { 701 sc->all_count--; 702 continue; 703 } 704 705 levels[i] = *lev; 706 i++; 707 } 708 *count = sc->all_count; 709 error = 0; 710 711 out: 712 /* Clear all levels since we regenerate them each time. */ 713 while ((lev = TAILQ_FIRST(&sc->all_levels)) != NULL) { 714 TAILQ_REMOVE(&sc->all_levels, lev, link); 715 free(lev, M_TEMP); 716 } 717 sc->all_count = 0; 718 719 CF_MTX_UNLOCK(&sc->lock); 720 while ((set_arr = TAILQ_FIRST(&rel_sets)) != NULL) { 721 TAILQ_REMOVE(&rel_sets, set_arr, link); 722 free(set_arr, M_TEMP); 723 } 724 return (error); 725 } 726 727 /* 728 * Create levels for an array of absolute settings and insert them in 729 * sorted order in the specified list. 730 */ 731 static int 732 cpufreq_insert_abs(struct cpufreq_softc *sc, struct cf_setting *sets, 733 int count) 734 { 735 struct cf_level_lst *list; 736 struct cf_level *level, *search; 737 int i, inserted; 738 739 CF_MTX_ASSERT(&sc->lock); 740 741 list = &sc->all_levels; 742 for (i = 0; i < count; i++) { 743 level = malloc(sizeof(*level), M_TEMP, M_NOWAIT | M_ZERO); 744 if (level == NULL) 745 return (ENOMEM); 746 level->abs_set = sets[i]; 747 level->total_set = sets[i]; 748 level->total_set.dev = NULL; 749 sc->all_count++; 750 inserted = 0; 751 752 if (TAILQ_EMPTY(list)) { 753 CF_DEBUG("adding abs setting %d at head\n", 754 sets[i].freq); 755 TAILQ_INSERT_HEAD(list, level, link); 756 continue; 757 } 758 759 TAILQ_FOREACH_REVERSE(search, list, cf_level_lst, link) 760 if (sets[i].freq <= search->total_set.freq) { 761 CF_DEBUG("adding abs setting %d after %d\n", 762 sets[i].freq, search->total_set.freq); 763 TAILQ_INSERT_AFTER(list, search, level, link); 764 inserted = 1; 765 break; 766 } 767 768 if (inserted == 0) { 769 TAILQ_FOREACH(search, list, link) 770 if (sets[i].freq >= search->total_set.freq) { 771 CF_DEBUG("adding abs setting %d before %d\n", 772 sets[i].freq, search->total_set.freq); 773 TAILQ_INSERT_BEFORE(search, level, link); 774 break; 775 } 776 } 777 } 778 779 return (0); 780 } 781 782 /* 783 * Expand a group of relative settings, creating derived levels from them. 784 */ 785 static int 786 cpufreq_expand_set(struct cpufreq_softc *sc, struct cf_setting_array *set_arr) 787 { 788 struct cf_level *fill, *search; 789 struct cf_setting *set; 790 int i; 791 792 CF_MTX_ASSERT(&sc->lock); 793 794 /* 795 * Walk the set of all existing levels in reverse. This is so we 796 * create derived states from the lowest absolute settings first 797 * and discard duplicates created from higher absolute settings. 798 * For instance, a level of 50 Mhz derived from 100 Mhz + 50% is 799 * preferable to 200 Mhz + 25% because absolute settings are more 800 * efficient since they often change the voltage as well. 801 */ 802 TAILQ_FOREACH_REVERSE(search, &sc->all_levels, cf_level_lst, link) { 803 /* Add each setting to the level, duplicating if necessary. */ 804 for (i = 0; i < set_arr->count; i++) { 805 set = &set_arr->sets[i]; 806 807 /* 808 * If this setting is less than 100%, split the level 809 * into two and add this setting to the new level. 810 */ 811 fill = search; 812 if (set->freq < 10000) { 813 fill = cpufreq_dup_set(sc, search, set); 814 815 /* 816 * The new level was a duplicate of an existing 817 * level or its absolute setting is too high 818 * so we freed it. For example, we discard a 819 * derived level of 1000 MHz/25% if a level 820 * of 500 MHz/100% already exists. 821 */ 822 if (fill == NULL) 823 break; 824 } 825 826 /* Add this setting to the existing or new level. */ 827 KASSERT(fill->rel_count < MAX_SETTINGS, 828 ("cpufreq: too many relative drivers (%d)", 829 MAX_SETTINGS)); 830 fill->rel_set[fill->rel_count] = *set; 831 fill->rel_count++; 832 CF_DEBUG( 833 "expand set added rel setting %d%% to %d level\n", 834 set->freq / 100, fill->total_set.freq); 835 } 836 } 837 838 return (0); 839 } 840 841 static struct cf_level * 842 cpufreq_dup_set(struct cpufreq_softc *sc, struct cf_level *dup, 843 struct cf_setting *set) 844 { 845 struct cf_level_lst *list; 846 struct cf_level *fill, *itr; 847 struct cf_setting *fill_set, *itr_set; 848 int i; 849 850 CF_MTX_ASSERT(&sc->lock); 851 852 /* 853 * Create a new level, copy it from the old one, and update the 854 * total frequency and power by the percentage specified in the 855 * relative setting. 856 */ 857 fill = malloc(sizeof(*fill), M_TEMP, M_NOWAIT); 858 if (fill == NULL) 859 return (NULL); 860 *fill = *dup; 861 fill_set = &fill->total_set; 862 fill_set->freq = 863 ((uint64_t)fill_set->freq * set->freq) / 10000; 864 if (fill_set->power != CPUFREQ_VAL_UNKNOWN) { 865 fill_set->power = ((uint64_t)fill_set->power * set->freq) 866 / 10000; 867 } 868 if (set->lat != CPUFREQ_VAL_UNKNOWN) { 869 if (fill_set->lat != CPUFREQ_VAL_UNKNOWN) 870 fill_set->lat += set->lat; 871 else 872 fill_set->lat = set->lat; 873 } 874 CF_DEBUG("dup set considering derived setting %d\n", fill_set->freq); 875 876 /* 877 * If we copied an old level that we already modified (say, at 100%), 878 * we need to remove that setting before adding this one. Since we 879 * process each setting array in order, we know any settings for this 880 * driver will be found at the end. 881 */ 882 for (i = fill->rel_count; i != 0; i--) { 883 if (fill->rel_set[i - 1].dev != set->dev) 884 break; 885 CF_DEBUG("removed last relative driver: %s\n", 886 device_get_nameunit(set->dev)); 887 fill->rel_count--; 888 } 889 890 /* 891 * Insert the new level in sorted order. If it is a duplicate of an 892 * existing level (1) or has an absolute setting higher than the 893 * existing level (2), do not add it. We can do this since any such 894 * level is guaranteed use less power. For example (1), a level with 895 * one absolute setting of 800 Mhz uses less power than one composed 896 * of an absolute setting of 1600 Mhz and a relative setting at 50%. 897 * Also for example (2), a level of 800 Mhz/75% is preferable to 898 * 1600 Mhz/25% even though the latter has a lower total frequency. 899 */ 900 list = &sc->all_levels; 901 KASSERT(!TAILQ_EMPTY(list), ("all levels list empty in dup set")); 902 TAILQ_FOREACH_REVERSE(itr, list, cf_level_lst, link) { 903 itr_set = &itr->total_set; 904 if (CPUFREQ_CMP(fill_set->freq, itr_set->freq)) { 905 CF_DEBUG("dup set rejecting %d (dupe)\n", 906 fill_set->freq); 907 itr = NULL; 908 break; 909 } else if (fill_set->freq < itr_set->freq) { 910 if (fill->abs_set.freq <= itr->abs_set.freq) { 911 CF_DEBUG( 912 "dup done, inserting new level %d after %d\n", 913 fill_set->freq, itr_set->freq); 914 TAILQ_INSERT_AFTER(list, itr, fill, link); 915 sc->all_count++; 916 } else { 917 CF_DEBUG("dup set rejecting %d (abs too big)\n", 918 fill_set->freq); 919 itr = NULL; 920 } 921 break; 922 } 923 } 924 925 /* We didn't find a good place for this new level so free it. */ 926 if (itr == NULL) { 927 CF_DEBUG("dup set freeing new level %d (not optimal)\n", 928 fill_set->freq); 929 free(fill, M_TEMP); 930 fill = NULL; 931 } 932 933 return (fill); 934 } 935 936 static int 937 cpufreq_curr_sysctl(SYSCTL_HANDLER_ARGS) 938 { 939 struct cpufreq_softc *sc; 940 struct cf_level *levels; 941 int best, count, diff, bdiff, devcount, error, freq, i, n; 942 device_t *devs; 943 944 devs = NULL; 945 sc = oidp->oid_arg1; 946 levels = sc->levels_buf; 947 948 error = CPUFREQ_GET(sc->dev, &levels[0]); 949 if (error) 950 goto out; 951 freq = levels[0].total_set.freq; 952 error = sysctl_handle_int(oidp, &freq, 0, req); 953 if (error != 0 || req->newptr == NULL) 954 goto out; 955 956 /* 957 * While we only call cpufreq_get() on one device (assuming all 958 * CPUs have equal levels), we call cpufreq_set() on all CPUs. 959 * This is needed for some MP systems. 960 */ 961 error = devclass_get_devices(devclass_find("cpufreq"), &devs, &devcount); 962 if (error) 963 goto out; 964 for (n = 0; n < devcount; n++) { 965 count = CF_MAX_LEVELS; 966 error = CPUFREQ_LEVELS(devs[n], levels, &count); 967 if (error) { 968 if (error == E2BIG) 969 printf( 970 "cpufreq: need to increase CF_MAX_LEVELS\n"); 971 break; 972 } 973 best = 0; 974 bdiff = 1 << 30; 975 for (i = 0; i < count; i++) { 976 diff = abs(levels[i].total_set.freq - freq); 977 if (diff < bdiff) { 978 bdiff = diff; 979 best = i; 980 } 981 } 982 error = CPUFREQ_SET(devs[n], &levels[best], CPUFREQ_PRIO_USER); 983 } 984 985 out: 986 if (devs) 987 free(devs, M_TEMP); 988 return (error); 989 } 990 991 static int 992 cpufreq_levels_sysctl(SYSCTL_HANDLER_ARGS) 993 { 994 struct cpufreq_softc *sc; 995 struct cf_level *levels; 996 struct cf_setting *set; 997 struct sbuf sb; 998 int count, error, i; 999 1000 sc = oidp->oid_arg1; 1001 sbuf_new(&sb, NULL, 128, SBUF_AUTOEXTEND); 1002 1003 /* Get settings from the device and generate the output string. */ 1004 count = CF_MAX_LEVELS; 1005 levels = sc->levels_buf; 1006 if (levels == NULL) { 1007 sbuf_delete(&sb); 1008 return (ENOMEM); 1009 } 1010 error = CPUFREQ_LEVELS(sc->dev, levels, &count); 1011 if (error) { 1012 if (error == E2BIG) 1013 printf("cpufreq: need to increase CF_MAX_LEVELS\n"); 1014 goto out; 1015 } 1016 if (count) { 1017 for (i = 0; i < count; i++) { 1018 set = &levels[i].total_set; 1019 sbuf_printf(&sb, "%d/%d ", set->freq, set->power); 1020 } 1021 } else 1022 sbuf_cpy(&sb, "0"); 1023 sbuf_trim(&sb); 1024 sbuf_finish(&sb); 1025 error = sysctl_handle_string(oidp, sbuf_data(&sb), sbuf_len(&sb), req); 1026 1027 out: 1028 sbuf_delete(&sb); 1029 return (error); 1030 } 1031 1032 static int 1033 cpufreq_settings_sysctl(SYSCTL_HANDLER_ARGS) 1034 { 1035 device_t dev; 1036 struct cf_setting *sets; 1037 struct sbuf sb; 1038 int error, i, set_count; 1039 1040 dev = oidp->oid_arg1; 1041 sbuf_new(&sb, NULL, 128, SBUF_AUTOEXTEND); 1042 1043 /* Get settings from the device and generate the output string. */ 1044 set_count = MAX_SETTINGS; 1045 sets = malloc(set_count * sizeof(*sets), M_TEMP, M_NOWAIT); 1046 if (sets == NULL) { 1047 sbuf_delete(&sb); 1048 return (ENOMEM); 1049 } 1050 error = CPUFREQ_DRV_SETTINGS(dev, sets, &set_count); 1051 if (error) 1052 goto out; 1053 if (set_count) { 1054 for (i = 0; i < set_count; i++) 1055 sbuf_printf(&sb, "%d/%d ", sets[i].freq, sets[i].power); 1056 } else 1057 sbuf_cpy(&sb, "0"); 1058 sbuf_trim(&sb); 1059 sbuf_finish(&sb); 1060 error = sysctl_handle_string(oidp, sbuf_data(&sb), sbuf_len(&sb), req); 1061 1062 out: 1063 free(sets, M_TEMP); 1064 sbuf_delete(&sb); 1065 return (error); 1066 } 1067 1068 static void 1069 cpufreq_add_freq_driver_sysctl(device_t cf_dev) 1070 { 1071 struct cpufreq_softc *sc; 1072 1073 sc = device_get_softc(cf_dev); 1074 SYSCTL_ADD_CONST_STRING(&sc->sysctl_ctx, 1075 SYSCTL_CHILDREN(device_get_sysctl_tree(cf_dev)), OID_AUTO, 1076 "freq_driver", CTLFLAG_RD, device_get_nameunit(sc->cf_drv_dev), 1077 "cpufreq driver used by this cpu"); 1078 } 1079 1080 int 1081 cpufreq_register(device_t dev) 1082 { 1083 struct cpufreq_softc *sc; 1084 device_t cf_dev, cpu_dev; 1085 int error; 1086 1087 /* Add a sysctl to get each driver's settings separately. */ 1088 SYSCTL_ADD_PROC(device_get_sysctl_ctx(dev), 1089 SYSCTL_CHILDREN(device_get_sysctl_tree(dev)), 1090 OID_AUTO, "freq_settings", 1091 CTLTYPE_STRING | CTLFLAG_RD | CTLFLAG_NEEDGIANT, dev, 0, 1092 cpufreq_settings_sysctl, "A", "CPU frequency driver settings"); 1093 1094 /* 1095 * Add only one cpufreq device to each CPU. Currently, all CPUs 1096 * must offer the same levels and be switched at the same time. 1097 */ 1098 cpu_dev = device_get_parent(dev); 1099 if ((cf_dev = device_find_child(cpu_dev, "cpufreq", -1))) { 1100 sc = device_get_softc(cf_dev); 1101 sc->max_mhz = CPUFREQ_VAL_UNKNOWN; 1102 MPASS(sc->cf_drv_dev != NULL); 1103 return (0); 1104 } 1105 1106 /* Add the child device and possibly sysctls. */ 1107 cf_dev = BUS_ADD_CHILD(cpu_dev, 0, "cpufreq", device_get_unit(cpu_dev)); 1108 if (cf_dev == NULL) 1109 return (ENOMEM); 1110 device_quiet(cf_dev); 1111 1112 error = device_probe_and_attach(cf_dev); 1113 if (error) 1114 return (error); 1115 1116 sc = device_get_softc(cf_dev); 1117 sc->cf_drv_dev = dev; 1118 cpufreq_add_freq_driver_sysctl(cf_dev); 1119 return (error); 1120 } 1121 1122 int 1123 cpufreq_unregister(device_t dev) 1124 { 1125 device_t cf_dev; 1126 struct cpufreq_softc *sc __diagused; 1127 1128 /* 1129 * If this is the last cpufreq child device, remove the control 1130 * device as well. We identify cpufreq children by calling a method 1131 * they support. 1132 */ 1133 cf_dev = device_find_child(device_get_parent(dev), "cpufreq", -1); 1134 if (cf_dev == NULL) { 1135 device_printf(dev, 1136 "warning: cpufreq_unregister called with no cpufreq device active\n"); 1137 return (0); 1138 } 1139 sc = device_get_softc(cf_dev); 1140 MPASS(sc->cf_drv_dev == dev); 1141 device_delete_child(device_get_parent(cf_dev), cf_dev); 1142 1143 return (0); 1144 } 1145 1146 int 1147 cpufreq_settings_changed(device_t dev) 1148 { 1149 1150 EVENTHANDLER_INVOKE(cpufreq_levels_changed, 1151 device_get_unit(device_get_parent(dev))); 1152 return (0); 1153 } 1154