xref: /freebsd/sys/kern/kern_cpu.c (revision 63a938566d524836885917d95bd491aa4400b181)
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	64
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 	struct sysctl_ctx_list		sysctl_ctx;
80 	struct task			startup_task;
81 	struct cf_level			*levels_buf;
82 };
83 
84 struct cf_setting_array {
85 	struct cf_setting		sets[MAX_SETTINGS];
86 	int				count;
87 	TAILQ_ENTRY(cf_setting_array)	link;
88 };
89 
90 TAILQ_HEAD(cf_setting_lst, cf_setting_array);
91 
92 #define CF_MTX_INIT(x)		sx_init((x), "cpufreq lock")
93 #define CF_MTX_LOCK(x)		sx_xlock((x))
94 #define CF_MTX_UNLOCK(x)	sx_xunlock((x))
95 #define CF_MTX_ASSERT(x)	sx_assert((x), SX_XLOCKED)
96 
97 #define CF_DEBUG(msg...)	do {		\
98 	if (cf_verbose)				\
99 		printf("cpufreq: " msg);	\
100 	} while (0)
101 
102 static int	cpufreq_attach(device_t dev);
103 static void	cpufreq_startup_task(void *ctx, int pending);
104 static int	cpufreq_detach(device_t dev);
105 static int	cf_set_method(device_t dev, const struct cf_level *level,
106 		    int priority);
107 static int	cf_get_method(device_t dev, struct cf_level *level);
108 static int	cf_levels_method(device_t dev, struct cf_level *levels,
109 		    int *count);
110 static int	cpufreq_insert_abs(struct cpufreq_softc *sc,
111 		    struct cf_setting *sets, int count);
112 static int	cpufreq_expand_set(struct cpufreq_softc *sc,
113 		    struct cf_setting_array *set_arr);
114 static struct cf_level *cpufreq_dup_set(struct cpufreq_softc *sc,
115 		    struct cf_level *dup, struct cf_setting *set);
116 static int	cpufreq_curr_sysctl(SYSCTL_HANDLER_ARGS);
117 static int	cpufreq_levels_sysctl(SYSCTL_HANDLER_ARGS);
118 static int	cpufreq_settings_sysctl(SYSCTL_HANDLER_ARGS);
119 
120 static device_method_t cpufreq_methods[] = {
121 	DEVMETHOD(device_probe,		bus_generic_probe),
122 	DEVMETHOD(device_attach,	cpufreq_attach),
123 	DEVMETHOD(device_detach,	cpufreq_detach),
124 
125         DEVMETHOD(cpufreq_set,		cf_set_method),
126         DEVMETHOD(cpufreq_get,		cf_get_method),
127         DEVMETHOD(cpufreq_levels,	cf_levels_method),
128 	{0, 0}
129 };
130 static driver_t cpufreq_driver = {
131 	"cpufreq", cpufreq_methods, sizeof(struct cpufreq_softc)
132 };
133 static devclass_t cpufreq_dc;
134 DRIVER_MODULE(cpufreq, cpu, cpufreq_driver, cpufreq_dc, 0, 0);
135 
136 static int		cf_lowest_freq;
137 static int		cf_verbose;
138 static SYSCTL_NODE(_debug, OID_AUTO, cpufreq, CTLFLAG_RD, NULL,
139     "cpufreq debugging");
140 SYSCTL_INT(_debug_cpufreq, OID_AUTO, lowest, CTLFLAG_RWTUN, &cf_lowest_freq, 1,
141     "Don't provide levels below this frequency.");
142 SYSCTL_INT(_debug_cpufreq, OID_AUTO, verbose, CTLFLAG_RWTUN, &cf_verbose, 1,
143     "Print verbose debugging messages");
144 
145 static int
146 cpufreq_attach(device_t dev)
147 {
148 	struct cpufreq_softc *sc;
149 	struct pcpu *pc;
150 	device_t parent;
151 	uint64_t rate;
152 	int numdevs;
153 
154 	CF_DEBUG("initializing %s\n", device_get_nameunit(dev));
155 	sc = device_get_softc(dev);
156 	parent = device_get_parent(dev);
157 	sc->dev = dev;
158 	sysctl_ctx_init(&sc->sysctl_ctx);
159 	TAILQ_INIT(&sc->all_levels);
160 	CF_MTX_INIT(&sc->lock);
161 	sc->curr_level.total_set.freq = CPUFREQ_VAL_UNKNOWN;
162 	SLIST_INIT(&sc->saved_freq);
163 	/* Try to get nominal CPU freq to use it as maximum later if needed */
164 	sc->max_mhz = cpu_get_nominal_mhz(dev);
165 	/* If that fails, try to measure the current rate */
166 	if (sc->max_mhz <= 0) {
167 		pc = cpu_get_pcpu(dev);
168 		if (cpu_est_clockrate(pc->pc_cpuid, &rate) == 0)
169 			sc->max_mhz = rate / 1000000;
170 		else
171 			sc->max_mhz = CPUFREQ_VAL_UNKNOWN;
172 	}
173 
174 	/*
175 	 * Only initialize one set of sysctls for all CPUs.  In the future,
176 	 * if multiple CPUs can have different settings, we can move these
177 	 * sysctls to be under every CPU instead of just the first one.
178 	 */
179 	numdevs = devclass_get_count(cpufreq_dc);
180 	if (numdevs > 1)
181 		return (0);
182 
183 	CF_DEBUG("initializing one-time data for %s\n",
184 	    device_get_nameunit(dev));
185 	sc->levels_buf = malloc(CF_MAX_LEVELS * sizeof(*sc->levels_buf),
186 	    M_DEVBUF, M_WAITOK);
187 	SYSCTL_ADD_PROC(&sc->sysctl_ctx,
188 	    SYSCTL_CHILDREN(device_get_sysctl_tree(parent)),
189 	    OID_AUTO, "freq", CTLTYPE_INT | CTLFLAG_RW, sc, 0,
190 	    cpufreq_curr_sysctl, "I", "Current CPU frequency");
191 	SYSCTL_ADD_PROC(&sc->sysctl_ctx,
192 	    SYSCTL_CHILDREN(device_get_sysctl_tree(parent)),
193 	    OID_AUTO, "freq_levels", CTLTYPE_STRING | CTLFLAG_RD, sc, 0,
194 	    cpufreq_levels_sysctl, "A", "CPU frequency levels");
195 
196 	/*
197 	 * Queue a one-shot broadcast that levels have changed.
198 	 * It will run once the system has completed booting.
199 	 */
200 	TASK_INIT(&sc->startup_task, 0, cpufreq_startup_task, dev);
201 	taskqueue_enqueue(taskqueue_thread, &sc->startup_task);
202 
203 	return (0);
204 }
205 
206 /* Handle any work to be done for all drivers that attached during boot. */
207 static void
208 cpufreq_startup_task(void *ctx, int pending)
209 {
210 
211 	cpufreq_settings_changed((device_t)ctx);
212 }
213 
214 static int
215 cpufreq_detach(device_t dev)
216 {
217 	struct cpufreq_softc *sc;
218 	struct cf_saved_freq *saved_freq;
219 	int numdevs;
220 
221 	CF_DEBUG("shutdown %s\n", device_get_nameunit(dev));
222 	sc = device_get_softc(dev);
223 	sysctl_ctx_free(&sc->sysctl_ctx);
224 
225 	while ((saved_freq = SLIST_FIRST(&sc->saved_freq)) != NULL) {
226 		SLIST_REMOVE_HEAD(&sc->saved_freq, link);
227 		free(saved_freq, M_TEMP);
228 	}
229 
230 	/* Only clean up these resources when the last device is detaching. */
231 	numdevs = devclass_get_count(cpufreq_dc);
232 	if (numdevs == 1) {
233 		CF_DEBUG("final shutdown for %s\n", device_get_nameunit(dev));
234 		free(sc->levels_buf, M_DEVBUF);
235 	}
236 
237 	return (0);
238 }
239 
240 static int
241 cf_set_method(device_t dev, const struct cf_level *level, int priority)
242 {
243 	struct cpufreq_softc *sc;
244 	const struct cf_setting *set;
245 	struct cf_saved_freq *saved_freq, *curr_freq;
246 	struct pcpu *pc;
247 	int error, i;
248 
249 	sc = device_get_softc(dev);
250 	error = 0;
251 	set = NULL;
252 	saved_freq = NULL;
253 
254 	/* We are going to change levels so notify the pre-change handler. */
255 	EVENTHANDLER_INVOKE(cpufreq_pre_change, level, &error);
256 	if (error != 0) {
257 		EVENTHANDLER_INVOKE(cpufreq_post_change, level, error);
258 		return (error);
259 	}
260 
261 	CF_MTX_LOCK(&sc->lock);
262 
263 #ifdef SMP
264 #ifdef EARLY_AP_STARTUP
265 	MPASS(mp_ncpus == 1 || smp_started);
266 #else
267 	/*
268 	 * If still booting and secondary CPUs not started yet, don't allow
269 	 * changing the frequency until they're online.  This is because we
270 	 * can't switch to them using sched_bind() and thus we'd only be
271 	 * switching the main CPU.  XXXTODO: Need to think more about how to
272 	 * handle having different CPUs at different frequencies.
273 	 */
274 	if (mp_ncpus > 1 && !smp_started) {
275 		device_printf(dev, "rejecting change, SMP not started yet\n");
276 		error = ENXIO;
277 		goto out;
278 	}
279 #endif
280 #endif /* SMP */
281 
282 	/*
283 	 * If the requested level has a lower priority, don't allow
284 	 * the new level right now.
285 	 */
286 	if (priority < sc->curr_priority) {
287 		CF_DEBUG("ignoring, curr prio %d less than %d\n", priority,
288 		    sc->curr_priority);
289 		error = EPERM;
290 		goto out;
291 	}
292 
293 	/*
294 	 * If the caller didn't specify a level and one is saved, prepare to
295 	 * restore the saved level.  If none has been saved, return an error.
296 	 */
297 	if (level == NULL) {
298 		saved_freq = SLIST_FIRST(&sc->saved_freq);
299 		if (saved_freq == NULL) {
300 			CF_DEBUG("NULL level, no saved level\n");
301 			error = ENXIO;
302 			goto out;
303 		}
304 		level = &saved_freq->level;
305 		priority = saved_freq->priority;
306 		CF_DEBUG("restoring saved level, freq %d prio %d\n",
307 		    level->total_set.freq, priority);
308 	}
309 
310 	/* Reject levels that are below our specified threshold. */
311 	if (level->total_set.freq < cf_lowest_freq) {
312 		CF_DEBUG("rejecting freq %d, less than %d limit\n",
313 		    level->total_set.freq, cf_lowest_freq);
314 		error = EINVAL;
315 		goto out;
316 	}
317 
318 	/* If already at this level, just return. */
319 	if (sc->curr_level.total_set.freq == level->total_set.freq) {
320 		CF_DEBUG("skipping freq %d, same as current level %d\n",
321 		    level->total_set.freq, sc->curr_level.total_set.freq);
322 		goto skip;
323 	}
324 
325 	/* First, set the absolute frequency via its driver. */
326 	set = &level->abs_set;
327 	if (set->dev) {
328 		if (!device_is_attached(set->dev)) {
329 			error = ENXIO;
330 			goto out;
331 		}
332 
333 		/* Bind to the target CPU before switching. */
334 		pc = cpu_get_pcpu(set->dev);
335 		thread_lock(curthread);
336 		sched_bind(curthread, pc->pc_cpuid);
337 		thread_unlock(curthread);
338 		CF_DEBUG("setting abs freq %d on %s (cpu %d)\n", set->freq,
339 		    device_get_nameunit(set->dev), PCPU_GET(cpuid));
340 		error = CPUFREQ_DRV_SET(set->dev, set);
341 		thread_lock(curthread);
342 		sched_unbind(curthread);
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 		sched_bind(curthread, pc->pc_cpuid);
361 		thread_unlock(curthread);
362 		CF_DEBUG("setting rel freq %d on %s (cpu %d)\n", set->freq,
363 		    device_get_nameunit(set->dev), PCPU_GET(cpuid));
364 		error = CPUFREQ_DRV_SET(set->dev, set);
365 		thread_lock(curthread);
366 		sched_unbind(curthread);
367 		thread_unlock(curthread);
368 		if (error) {
369 			/* XXX Back out any successful setting? */
370 			goto out;
371 		}
372 	}
373 
374 skip:
375 	/*
376 	 * Before recording the current level, check if we're going to a
377 	 * higher priority.  If so, save the previous level and priority.
378 	 */
379 	if (sc->curr_level.total_set.freq != CPUFREQ_VAL_UNKNOWN &&
380 	    priority > sc->curr_priority) {
381 		CF_DEBUG("saving level, freq %d prio %d\n",
382 		    sc->curr_level.total_set.freq, sc->curr_priority);
383 		curr_freq = malloc(sizeof(*curr_freq), M_TEMP, M_NOWAIT);
384 		if (curr_freq == NULL) {
385 			error = ENOMEM;
386 			goto out;
387 		}
388 		curr_freq->level = sc->curr_level;
389 		curr_freq->priority = sc->curr_priority;
390 		SLIST_INSERT_HEAD(&sc->saved_freq, curr_freq, link);
391 	}
392 	sc->curr_level = *level;
393 	sc->curr_priority = priority;
394 
395 	/* If we were restoring a saved state, reset it to "unused". */
396 	if (saved_freq != NULL) {
397 		CF_DEBUG("resetting saved level\n");
398 		sc->curr_level.total_set.freq = CPUFREQ_VAL_UNKNOWN;
399 		SLIST_REMOVE_HEAD(&sc->saved_freq, link);
400 		free(saved_freq, M_TEMP);
401 	}
402 
403 out:
404 	CF_MTX_UNLOCK(&sc->lock);
405 
406 	/*
407 	 * We changed levels (or attempted to) so notify the post-change
408 	 * handler of new frequency or error.
409 	 */
410 	EVENTHANDLER_INVOKE(cpufreq_post_change, level, error);
411 	if (error && set)
412 		device_printf(set->dev, "set freq failed, err %d\n", error);
413 
414 	return (error);
415 }
416 
417 static int
418 cf_get_method(device_t dev, struct cf_level *level)
419 {
420 	struct cpufreq_softc *sc;
421 	struct cf_level *levels;
422 	struct cf_setting *curr_set, set;
423 	struct pcpu *pc;
424 	device_t *devs;
425 	int bdiff, count, diff, error, i, n, numdevs;
426 	uint64_t rate;
427 
428 	sc = device_get_softc(dev);
429 	error = 0;
430 	levels = NULL;
431 
432 	/* If we already know the current frequency, we're done. */
433 	CF_MTX_LOCK(&sc->lock);
434 	curr_set = &sc->curr_level.total_set;
435 	if (curr_set->freq != CPUFREQ_VAL_UNKNOWN) {
436 		CF_DEBUG("get returning known freq %d\n", curr_set->freq);
437 		goto out;
438 	}
439 	CF_MTX_UNLOCK(&sc->lock);
440 
441 	/*
442 	 * We need to figure out the current level.  Loop through every
443 	 * driver, getting the current setting.  Then, attempt to get a best
444 	 * match of settings against each level.
445 	 */
446 	count = CF_MAX_LEVELS;
447 	levels = malloc(count * sizeof(*levels), M_TEMP, M_NOWAIT);
448 	if (levels == NULL)
449 		return (ENOMEM);
450 	error = CPUFREQ_LEVELS(sc->dev, levels, &count);
451 	if (error) {
452 		if (error == E2BIG)
453 			printf("cpufreq: need to increase CF_MAX_LEVELS\n");
454 		free(levels, M_TEMP);
455 		return (error);
456 	}
457 	error = device_get_children(device_get_parent(dev), &devs, &numdevs);
458 	if (error) {
459 		free(levels, M_TEMP);
460 		return (error);
461 	}
462 
463 	/*
464 	 * Reacquire the lock and search for the given level.
465 	 *
466 	 * XXX Note: this is not quite right since we really need to go
467 	 * through each level and compare both absolute and relative
468 	 * settings for each driver in the system before making a match.
469 	 * The estimation code below catches this case though.
470 	 */
471 	CF_MTX_LOCK(&sc->lock);
472 	for (n = 0; n < numdevs && curr_set->freq == CPUFREQ_VAL_UNKNOWN; n++) {
473 		if (!device_is_attached(devs[n]))
474 			continue;
475 		if (CPUFREQ_DRV_GET(devs[n], &set) != 0)
476 			continue;
477 		for (i = 0; i < count; i++) {
478 			if (set.freq == levels[i].total_set.freq) {
479 				sc->curr_level = levels[i];
480 				break;
481 			}
482 		}
483 	}
484 	free(devs, M_TEMP);
485 	if (curr_set->freq != CPUFREQ_VAL_UNKNOWN) {
486 		CF_DEBUG("get matched freq %d from drivers\n", curr_set->freq);
487 		goto out;
488 	}
489 
490 	/*
491 	 * We couldn't find an exact match, so attempt to estimate and then
492 	 * match against a level.
493 	 */
494 	pc = cpu_get_pcpu(dev);
495 	if (pc == NULL) {
496 		error = ENXIO;
497 		goto out;
498 	}
499 	cpu_est_clockrate(pc->pc_cpuid, &rate);
500 	rate /= 1000000;
501 	bdiff = 1 << 30;
502 	for (i = 0; i < count; i++) {
503 		diff = abs(levels[i].total_set.freq - rate);
504 		if (diff < bdiff) {
505 			bdiff = diff;
506 			sc->curr_level = levels[i];
507 		}
508 	}
509 	CF_DEBUG("get estimated freq %d\n", curr_set->freq);
510 
511 out:
512 	if (error == 0)
513 		*level = sc->curr_level;
514 
515 	CF_MTX_UNLOCK(&sc->lock);
516 	if (levels)
517 		free(levels, M_TEMP);
518 	return (error);
519 }
520 
521 static int
522 cf_levels_method(device_t dev, struct cf_level *levels, int *count)
523 {
524 	struct cf_setting_array *set_arr;
525 	struct cf_setting_lst rel_sets;
526 	struct cpufreq_softc *sc;
527 	struct cf_level *lev;
528 	struct cf_setting *sets;
529 	struct pcpu *pc;
530 	device_t *devs;
531 	int error, i, numdevs, set_count, type;
532 	uint64_t rate;
533 
534 	if (levels == NULL || count == NULL)
535 		return (EINVAL);
536 
537 	TAILQ_INIT(&rel_sets);
538 	sc = device_get_softc(dev);
539 	error = device_get_children(device_get_parent(dev), &devs, &numdevs);
540 	if (error)
541 		return (error);
542 	sets = malloc(MAX_SETTINGS * sizeof(*sets), M_TEMP, M_NOWAIT);
543 	if (sets == NULL) {
544 		free(devs, M_TEMP);
545 		return (ENOMEM);
546 	}
547 
548 	/* Get settings from all cpufreq drivers. */
549 	CF_MTX_LOCK(&sc->lock);
550 	for (i = 0; i < numdevs; i++) {
551 		/* Skip devices that aren't ready. */
552 		if (!device_is_attached(devs[i]))
553 			continue;
554 
555 		/*
556 		 * Get settings, skipping drivers that offer no settings or
557 		 * provide settings for informational purposes only.
558 		 */
559 		error = CPUFREQ_DRV_TYPE(devs[i], &type);
560 		if (error || (type & CPUFREQ_FLAG_INFO_ONLY)) {
561 			if (error == 0) {
562 				CF_DEBUG("skipping info-only driver %s\n",
563 				    device_get_nameunit(devs[i]));
564 			}
565 			continue;
566 		}
567 		set_count = MAX_SETTINGS;
568 		error = CPUFREQ_DRV_SETTINGS(devs[i], sets, &set_count);
569 		if (error || set_count == 0)
570 			continue;
571 
572 		/* Add the settings to our absolute/relative lists. */
573 		switch (type & CPUFREQ_TYPE_MASK) {
574 		case CPUFREQ_TYPE_ABSOLUTE:
575 			error = cpufreq_insert_abs(sc, sets, set_count);
576 			break;
577 		case CPUFREQ_TYPE_RELATIVE:
578 			CF_DEBUG("adding %d relative settings\n", set_count);
579 			set_arr = malloc(sizeof(*set_arr), M_TEMP, M_NOWAIT);
580 			if (set_arr == NULL) {
581 				error = ENOMEM;
582 				goto out;
583 			}
584 			bcopy(sets, set_arr->sets, set_count * sizeof(*sets));
585 			set_arr->count = set_count;
586 			TAILQ_INSERT_TAIL(&rel_sets, set_arr, link);
587 			break;
588 		default:
589 			error = EINVAL;
590 		}
591 		if (error)
592 			goto out;
593 	}
594 
595 	/*
596 	 * If there are no absolute levels, create a fake one at 100%.  We
597 	 * then cache the clockrate for later use as our base frequency.
598 	 */
599 	if (TAILQ_EMPTY(&sc->all_levels)) {
600 		if (sc->max_mhz == CPUFREQ_VAL_UNKNOWN) {
601 			sc->max_mhz = cpu_get_nominal_mhz(dev);
602 			/*
603 			 * If the CPU can't report a rate for 100%, hope
604 			 * the CPU is running at its nominal rate right now,
605 			 * and use that instead.
606 			 */
607 			if (sc->max_mhz <= 0) {
608 				pc = cpu_get_pcpu(dev);
609 				cpu_est_clockrate(pc->pc_cpuid, &rate);
610 				sc->max_mhz = rate / 1000000;
611 			}
612 		}
613 		memset(&sets[0], CPUFREQ_VAL_UNKNOWN, sizeof(*sets));
614 		sets[0].freq = sc->max_mhz;
615 		sets[0].dev = NULL;
616 		error = cpufreq_insert_abs(sc, sets, 1);
617 		if (error)
618 			goto out;
619 	}
620 
621 	/* Create a combined list of absolute + relative levels. */
622 	TAILQ_FOREACH(set_arr, &rel_sets, link)
623 		cpufreq_expand_set(sc, set_arr);
624 
625 	/* If the caller doesn't have enough space, return the actual count. */
626 	if (sc->all_count > *count) {
627 		*count = sc->all_count;
628 		error = E2BIG;
629 		goto out;
630 	}
631 
632 	/* Finally, output the list of levels. */
633 	i = 0;
634 	TAILQ_FOREACH(lev, &sc->all_levels, link) {
635 
636 		/* Skip levels that have a frequency that is too low. */
637 		if (lev->total_set.freq < cf_lowest_freq) {
638 			sc->all_count--;
639 			continue;
640 		}
641 
642 		levels[i] = *lev;
643 		i++;
644 	}
645 	*count = sc->all_count;
646 	error = 0;
647 
648 out:
649 	/* Clear all levels since we regenerate them each time. */
650 	while ((lev = TAILQ_FIRST(&sc->all_levels)) != NULL) {
651 		TAILQ_REMOVE(&sc->all_levels, lev, link);
652 		free(lev, M_TEMP);
653 	}
654 	sc->all_count = 0;
655 
656 	CF_MTX_UNLOCK(&sc->lock);
657 	while ((set_arr = TAILQ_FIRST(&rel_sets)) != NULL) {
658 		TAILQ_REMOVE(&rel_sets, set_arr, link);
659 		free(set_arr, M_TEMP);
660 	}
661 	free(devs, M_TEMP);
662 	free(sets, M_TEMP);
663 	return (error);
664 }
665 
666 /*
667  * Create levels for an array of absolute settings and insert them in
668  * sorted order in the specified list.
669  */
670 static int
671 cpufreq_insert_abs(struct cpufreq_softc *sc, struct cf_setting *sets,
672     int count)
673 {
674 	struct cf_level_lst *list;
675 	struct cf_level *level, *search;
676 	int i;
677 
678 	CF_MTX_ASSERT(&sc->lock);
679 
680 	list = &sc->all_levels;
681 	for (i = 0; i < count; i++) {
682 		level = malloc(sizeof(*level), M_TEMP, M_NOWAIT | M_ZERO);
683 		if (level == NULL)
684 			return (ENOMEM);
685 		level->abs_set = sets[i];
686 		level->total_set = sets[i];
687 		level->total_set.dev = NULL;
688 		sc->all_count++;
689 
690 		if (TAILQ_EMPTY(list)) {
691 			CF_DEBUG("adding abs setting %d at head\n",
692 			    sets[i].freq);
693 			TAILQ_INSERT_HEAD(list, level, link);
694 			continue;
695 		}
696 
697 		TAILQ_FOREACH_REVERSE(search, list, cf_level_lst, link) {
698 			if (sets[i].freq <= search->total_set.freq) {
699 				CF_DEBUG("adding abs setting %d after %d\n",
700 				    sets[i].freq, search->total_set.freq);
701 				TAILQ_INSERT_AFTER(list, search, level, link);
702 				break;
703 			}
704 		}
705 	}
706 	return (0);
707 }
708 
709 /*
710  * Expand a group of relative settings, creating derived levels from them.
711  */
712 static int
713 cpufreq_expand_set(struct cpufreq_softc *sc, struct cf_setting_array *set_arr)
714 {
715 	struct cf_level *fill, *search;
716 	struct cf_setting *set;
717 	int i;
718 
719 	CF_MTX_ASSERT(&sc->lock);
720 
721 	/*
722 	 * Walk the set of all existing levels in reverse.  This is so we
723 	 * create derived states from the lowest absolute settings first
724 	 * and discard duplicates created from higher absolute settings.
725 	 * For instance, a level of 50 Mhz derived from 100 Mhz + 50% is
726 	 * preferable to 200 Mhz + 25% because absolute settings are more
727 	 * efficient since they often change the voltage as well.
728 	 */
729 	TAILQ_FOREACH_REVERSE(search, &sc->all_levels, cf_level_lst, link) {
730 		/* Add each setting to the level, duplicating if necessary. */
731 		for (i = 0; i < set_arr->count; i++) {
732 			set = &set_arr->sets[i];
733 
734 			/*
735 			 * If this setting is less than 100%, split the level
736 			 * into two and add this setting to the new level.
737 			 */
738 			fill = search;
739 			if (set->freq < 10000) {
740 				fill = cpufreq_dup_set(sc, search, set);
741 
742 				/*
743 				 * The new level was a duplicate of an existing
744 				 * level or its absolute setting is too high
745 				 * so we freed it.  For example, we discard a
746 				 * derived level of 1000 MHz/25% if a level
747 				 * of 500 MHz/100% already exists.
748 				 */
749 				if (fill == NULL)
750 					break;
751 			}
752 
753 			/* Add this setting to the existing or new level. */
754 			KASSERT(fill->rel_count < MAX_SETTINGS,
755 			    ("cpufreq: too many relative drivers (%d)",
756 			    MAX_SETTINGS));
757 			fill->rel_set[fill->rel_count] = *set;
758 			fill->rel_count++;
759 			CF_DEBUG(
760 			"expand set added rel setting %d%% to %d level\n",
761 			    set->freq / 100, fill->total_set.freq);
762 		}
763 	}
764 
765 	return (0);
766 }
767 
768 static struct cf_level *
769 cpufreq_dup_set(struct cpufreq_softc *sc, struct cf_level *dup,
770     struct cf_setting *set)
771 {
772 	struct cf_level_lst *list;
773 	struct cf_level *fill, *itr;
774 	struct cf_setting *fill_set, *itr_set;
775 	int i;
776 
777 	CF_MTX_ASSERT(&sc->lock);
778 
779 	/*
780 	 * Create a new level, copy it from the old one, and update the
781 	 * total frequency and power by the percentage specified in the
782 	 * relative setting.
783 	 */
784 	fill = malloc(sizeof(*fill), M_TEMP, M_NOWAIT);
785 	if (fill == NULL)
786 		return (NULL);
787 	*fill = *dup;
788 	fill_set = &fill->total_set;
789 	fill_set->freq =
790 	    ((uint64_t)fill_set->freq * set->freq) / 10000;
791 	if (fill_set->power != CPUFREQ_VAL_UNKNOWN) {
792 		fill_set->power = ((uint64_t)fill_set->power * set->freq)
793 		    / 10000;
794 	}
795 	if (set->lat != CPUFREQ_VAL_UNKNOWN) {
796 		if (fill_set->lat != CPUFREQ_VAL_UNKNOWN)
797 			fill_set->lat += set->lat;
798 		else
799 			fill_set->lat = set->lat;
800 	}
801 	CF_DEBUG("dup set considering derived setting %d\n", fill_set->freq);
802 
803 	/*
804 	 * If we copied an old level that we already modified (say, at 100%),
805 	 * we need to remove that setting before adding this one.  Since we
806 	 * process each setting array in order, we know any settings for this
807 	 * driver will be found at the end.
808 	 */
809 	for (i = fill->rel_count; i != 0; i--) {
810 		if (fill->rel_set[i - 1].dev != set->dev)
811 			break;
812 		CF_DEBUG("removed last relative driver: %s\n",
813 		    device_get_nameunit(set->dev));
814 		fill->rel_count--;
815 	}
816 
817 	/*
818 	 * Insert the new level in sorted order.  If it is a duplicate of an
819 	 * existing level (1) or has an absolute setting higher than the
820 	 * existing level (2), do not add it.  We can do this since any such
821 	 * level is guaranteed use less power.  For example (1), a level with
822 	 * one absolute setting of 800 Mhz uses less power than one composed
823 	 * of an absolute setting of 1600 Mhz and a relative setting at 50%.
824 	 * Also for example (2), a level of 800 Mhz/75% is preferable to
825 	 * 1600 Mhz/25% even though the latter has a lower total frequency.
826 	 */
827 	list = &sc->all_levels;
828 	KASSERT(!TAILQ_EMPTY(list), ("all levels list empty in dup set"));
829 	TAILQ_FOREACH_REVERSE(itr, list, cf_level_lst, link) {
830 		itr_set = &itr->total_set;
831 		if (CPUFREQ_CMP(fill_set->freq, itr_set->freq)) {
832 			CF_DEBUG("dup set rejecting %d (dupe)\n",
833 			    fill_set->freq);
834 			itr = NULL;
835 			break;
836 		} else if (fill_set->freq < itr_set->freq) {
837 			if (fill->abs_set.freq <= itr->abs_set.freq) {
838 				CF_DEBUG(
839 			"dup done, inserting new level %d after %d\n",
840 				    fill_set->freq, itr_set->freq);
841 				TAILQ_INSERT_AFTER(list, itr, fill, link);
842 				sc->all_count++;
843 			} else {
844 				CF_DEBUG("dup set rejecting %d (abs too big)\n",
845 				    fill_set->freq);
846 				itr = NULL;
847 			}
848 			break;
849 		}
850 	}
851 
852 	/* We didn't find a good place for this new level so free it. */
853 	if (itr == NULL) {
854 		CF_DEBUG("dup set freeing new level %d (not optimal)\n",
855 		    fill_set->freq);
856 		free(fill, M_TEMP);
857 		fill = NULL;
858 	}
859 
860 	return (fill);
861 }
862 
863 static int
864 cpufreq_curr_sysctl(SYSCTL_HANDLER_ARGS)
865 {
866 	struct cpufreq_softc *sc;
867 	struct cf_level *levels;
868 	int best, count, diff, bdiff, devcount, error, freq, i, n;
869 	device_t *devs;
870 
871 	devs = NULL;
872 	sc = oidp->oid_arg1;
873 	levels = sc->levels_buf;
874 
875 	error = CPUFREQ_GET(sc->dev, &levels[0]);
876 	if (error)
877 		goto out;
878 	freq = levels[0].total_set.freq;
879 	error = sysctl_handle_int(oidp, &freq, 0, req);
880 	if (error != 0 || req->newptr == NULL)
881 		goto out;
882 
883 	/*
884 	 * While we only call cpufreq_get() on one device (assuming all
885 	 * CPUs have equal levels), we call cpufreq_set() on all CPUs.
886 	 * This is needed for some MP systems.
887 	 */
888 	error = devclass_get_devices(cpufreq_dc, &devs, &devcount);
889 	if (error)
890 		goto out;
891 	for (n = 0; n < devcount; n++) {
892 		count = CF_MAX_LEVELS;
893 		error = CPUFREQ_LEVELS(devs[n], levels, &count);
894 		if (error) {
895 			if (error == E2BIG)
896 				printf(
897 			"cpufreq: need to increase CF_MAX_LEVELS\n");
898 			break;
899 		}
900 		best = 0;
901 		bdiff = 1 << 30;
902 		for (i = 0; i < count; i++) {
903 			diff = abs(levels[i].total_set.freq - freq);
904 			if (diff < bdiff) {
905 				bdiff = diff;
906 				best = i;
907 			}
908 		}
909 		error = CPUFREQ_SET(devs[n], &levels[best], CPUFREQ_PRIO_USER);
910 	}
911 
912 out:
913 	if (devs)
914 		free(devs, M_TEMP);
915 	return (error);
916 }
917 
918 static int
919 cpufreq_levels_sysctl(SYSCTL_HANDLER_ARGS)
920 {
921 	struct cpufreq_softc *sc;
922 	struct cf_level *levels;
923 	struct cf_setting *set;
924 	struct sbuf sb;
925 	int count, error, i;
926 
927 	sc = oidp->oid_arg1;
928 	sbuf_new(&sb, NULL, 128, SBUF_AUTOEXTEND);
929 
930 	/* Get settings from the device and generate the output string. */
931 	count = CF_MAX_LEVELS;
932 	levels = sc->levels_buf;
933 	if (levels == NULL) {
934 		sbuf_delete(&sb);
935 		return (ENOMEM);
936 	}
937 	error = CPUFREQ_LEVELS(sc->dev, levels, &count);
938 	if (error) {
939 		if (error == E2BIG)
940 			printf("cpufreq: need to increase CF_MAX_LEVELS\n");
941 		goto out;
942 	}
943 	if (count) {
944 		for (i = 0; i < count; i++) {
945 			set = &levels[i].total_set;
946 			sbuf_printf(&sb, "%d/%d ", set->freq, set->power);
947 		}
948 	} else
949 		sbuf_cpy(&sb, "0");
950 	sbuf_trim(&sb);
951 	sbuf_finish(&sb);
952 	error = sysctl_handle_string(oidp, sbuf_data(&sb), sbuf_len(&sb), req);
953 
954 out:
955 	sbuf_delete(&sb);
956 	return (error);
957 }
958 
959 static int
960 cpufreq_settings_sysctl(SYSCTL_HANDLER_ARGS)
961 {
962 	device_t dev;
963 	struct cf_setting *sets;
964 	struct sbuf sb;
965 	int error, i, set_count;
966 
967 	dev = oidp->oid_arg1;
968 	sbuf_new(&sb, NULL, 128, SBUF_AUTOEXTEND);
969 
970 	/* Get settings from the device and generate the output string. */
971 	set_count = MAX_SETTINGS;
972 	sets = malloc(set_count * sizeof(*sets), M_TEMP, M_NOWAIT);
973 	if (sets == NULL) {
974 		sbuf_delete(&sb);
975 		return (ENOMEM);
976 	}
977 	error = CPUFREQ_DRV_SETTINGS(dev, sets, &set_count);
978 	if (error)
979 		goto out;
980 	if (set_count) {
981 		for (i = 0; i < set_count; i++)
982 			sbuf_printf(&sb, "%d/%d ", sets[i].freq, sets[i].power);
983 	} else
984 		sbuf_cpy(&sb, "0");
985 	sbuf_trim(&sb);
986 	sbuf_finish(&sb);
987 	error = sysctl_handle_string(oidp, sbuf_data(&sb), sbuf_len(&sb), req);
988 
989 out:
990 	free(sets, M_TEMP);
991 	sbuf_delete(&sb);
992 	return (error);
993 }
994 
995 int
996 cpufreq_register(device_t dev)
997 {
998 	struct cpufreq_softc *sc;
999 	device_t cf_dev, cpu_dev;
1000 
1001 	/* Add a sysctl to get each driver's settings separately. */
1002 	SYSCTL_ADD_PROC(device_get_sysctl_ctx(dev),
1003 	    SYSCTL_CHILDREN(device_get_sysctl_tree(dev)),
1004 	    OID_AUTO, "freq_settings", CTLTYPE_STRING | CTLFLAG_RD, dev, 0,
1005 	    cpufreq_settings_sysctl, "A", "CPU frequency driver settings");
1006 
1007 	/*
1008 	 * Add only one cpufreq device to each CPU.  Currently, all CPUs
1009 	 * must offer the same levels and be switched at the same time.
1010 	 */
1011 	cpu_dev = device_get_parent(dev);
1012 	if ((cf_dev = device_find_child(cpu_dev, "cpufreq", -1))) {
1013 		sc = device_get_softc(cf_dev);
1014 		sc->max_mhz = CPUFREQ_VAL_UNKNOWN;
1015 		return (0);
1016 	}
1017 
1018 	/* Add the child device and possibly sysctls. */
1019 	cf_dev = BUS_ADD_CHILD(cpu_dev, 0, "cpufreq", -1);
1020 	if (cf_dev == NULL)
1021 		return (ENOMEM);
1022 	device_quiet(cf_dev);
1023 
1024 	return (device_probe_and_attach(cf_dev));
1025 }
1026 
1027 int
1028 cpufreq_unregister(device_t dev)
1029 {
1030 	device_t cf_dev, *devs;
1031 	int cfcount, devcount, error, i, type;
1032 
1033 	/*
1034 	 * If this is the last cpufreq child device, remove the control
1035 	 * device as well.  We identify cpufreq children by calling a method
1036 	 * they support.
1037 	 */
1038 	error = device_get_children(device_get_parent(dev), &devs, &devcount);
1039 	if (error)
1040 		return (error);
1041 	cf_dev = device_find_child(device_get_parent(dev), "cpufreq", -1);
1042 	if (cf_dev == NULL) {
1043 		device_printf(dev,
1044 	"warning: cpufreq_unregister called with no cpufreq device active\n");
1045 		free(devs, M_TEMP);
1046 		return (0);
1047 	}
1048 	cfcount = 0;
1049 	for (i = 0; i < devcount; i++) {
1050 		if (!device_is_attached(devs[i]))
1051 			continue;
1052 		if (CPUFREQ_DRV_TYPE(devs[i], &type) == 0)
1053 			cfcount++;
1054 	}
1055 	if (cfcount <= 1)
1056 		device_delete_child(device_get_parent(cf_dev), cf_dev);
1057 	free(devs, M_TEMP);
1058 
1059 	return (0);
1060 }
1061 
1062 int
1063 cpufreq_settings_changed(device_t dev)
1064 {
1065 
1066 	EVENTHANDLER_INVOKE(cpufreq_levels_changed,
1067 	    device_get_unit(device_get_parent(dev)));
1068 	return (0);
1069 }
1070