xref: /freebsd/sys/kern/kern_clocksource.c (revision 3b8f08459569bf0faa21473e5cec2491e95c9349)
1 /*-
2  * Copyright (c) 2010-2013 Alexander Motin <mav@FreeBSD.org>
3  * All rights reserved.
4  *
5  * Redistribution and use in source and binary forms, with or without
6  * modification, are permitted provided that the following conditions
7  * are met:
8  * 1. Redistributions of source code must retain the above copyright
9  *    notice, this list of conditions and the following disclaimer,
10  *    without modification, immediately at the beginning of the file.
11  * 2. Redistributions in binary form must reproduce the above copyright
12  *    notice, this list of conditions and the following disclaimer in the
13  *    documentation and/or other materials provided with the distribution.
14  *
15  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR
16  * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
17  * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
18  * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT,
19  * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT
20  * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
21  * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
22  * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
23  * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF
24  * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
25  */
26 
27 #include <sys/cdefs.h>
28 __FBSDID("$FreeBSD$");
29 
30 /*
31  * Common routines to manage event timers hardware.
32  */
33 
34 #include "opt_device_polling.h"
35 
36 #include <sys/param.h>
37 #include <sys/systm.h>
38 #include <sys/bus.h>
39 #include <sys/limits.h>
40 #include <sys/lock.h>
41 #include <sys/kdb.h>
42 #include <sys/ktr.h>
43 #include <sys/mutex.h>
44 #include <sys/proc.h>
45 #include <sys/kernel.h>
46 #include <sys/sched.h>
47 #include <sys/smp.h>
48 #include <sys/sysctl.h>
49 #include <sys/timeet.h>
50 #include <sys/timetc.h>
51 
52 #include <machine/atomic.h>
53 #include <machine/clock.h>
54 #include <machine/cpu.h>
55 #include <machine/smp.h>
56 
57 #ifdef KDTRACE_HOOKS
58 #include <sys/dtrace_bsd.h>
59 cyclic_clock_func_t	cyclic_clock_func = NULL;
60 #endif
61 
62 int			cpu_can_deep_sleep = 0;	/* C3 state is available. */
63 int			cpu_disable_deep_sleep = 0; /* Timer dies in C3. */
64 
65 static void		setuptimer(void);
66 static void		loadtimer(sbintime_t now, int first);
67 static int		doconfigtimer(void);
68 static void		configtimer(int start);
69 static int		round_freq(struct eventtimer *et, int freq);
70 
71 static sbintime_t	getnextcpuevent(int idle);
72 static sbintime_t	getnextevent(void);
73 static int		handleevents(sbintime_t now, int fake);
74 
75 static struct mtx	et_hw_mtx;
76 
77 #define	ET_HW_LOCK(state)						\
78 	{								\
79 		if (timer->et_flags & ET_FLAGS_PERCPU)			\
80 			mtx_lock_spin(&(state)->et_hw_mtx);		\
81 		else							\
82 			mtx_lock_spin(&et_hw_mtx);			\
83 	}
84 
85 #define	ET_HW_UNLOCK(state)						\
86 	{								\
87 		if (timer->et_flags & ET_FLAGS_PERCPU)			\
88 			mtx_unlock_spin(&(state)->et_hw_mtx);		\
89 		else							\
90 			mtx_unlock_spin(&et_hw_mtx);			\
91 	}
92 
93 static struct eventtimer *timer = NULL;
94 static sbintime_t	timerperiod;	/* Timer period for periodic mode. */
95 static sbintime_t	statperiod;	/* statclock() events period. */
96 static sbintime_t	profperiod;	/* profclock() events period. */
97 static sbintime_t	nexttick;	/* Next global timer tick time. */
98 static u_int		busy = 1;	/* Reconfiguration is in progress. */
99 static int		profiling = 0;	/* Profiling events enabled. */
100 
101 static char		timername[32];	/* Wanted timer. */
102 TUNABLE_STR("kern.eventtimer.timer", timername, sizeof(timername));
103 
104 static int		singlemul = 0;	/* Multiplier for periodic mode. */
105 TUNABLE_INT("kern.eventtimer.singlemul", &singlemul);
106 SYSCTL_INT(_kern_eventtimer, OID_AUTO, singlemul, CTLFLAG_RW, &singlemul,
107     0, "Multiplier for periodic mode");
108 
109 static u_int		idletick = 0;	/* Run periodic events when idle. */
110 TUNABLE_INT("kern.eventtimer.idletick", &idletick);
111 SYSCTL_UINT(_kern_eventtimer, OID_AUTO, idletick, CTLFLAG_RW, &idletick,
112     0, "Run periodic events when idle");
113 
114 static int		periodic = 0;	/* Periodic or one-shot mode. */
115 static int		want_periodic = 0; /* What mode to prefer. */
116 TUNABLE_INT("kern.eventtimer.periodic", &want_periodic);
117 
118 struct pcpu_state {
119 	struct mtx	et_hw_mtx;	/* Per-CPU timer mutex. */
120 	u_int		action;		/* Reconfiguration requests. */
121 	u_int		handle;		/* Immediate handle resuests. */
122 	sbintime_t	now;		/* Last tick time. */
123 	sbintime_t	nextevent;	/* Next scheduled event on this CPU. */
124 	sbintime_t	nexttick;	/* Next timer tick time. */
125 	sbintime_t	nexthard;	/* Next hardlock() event. */
126 	sbintime_t	nextstat;	/* Next statclock() event. */
127 	sbintime_t	nextprof;	/* Next profclock() event. */
128 	sbintime_t	nextcall;	/* Next callout event. */
129 	sbintime_t	nextcallopt;	/* Next optional callout event. */
130 #ifdef KDTRACE_HOOKS
131 	sbintime_t	nextcyc;	/* Next OpenSolaris cyclics event. */
132 #endif
133 	int		ipi;		/* This CPU needs IPI. */
134 	int		idle;		/* This CPU is in idle mode. */
135 };
136 
137 static DPCPU_DEFINE(struct pcpu_state, timerstate);
138 DPCPU_DEFINE(sbintime_t, hardclocktime);
139 
140 /*
141  * Timer broadcast IPI handler.
142  */
143 int
144 hardclockintr(void)
145 {
146 	sbintime_t now;
147 	struct pcpu_state *state;
148 	int done;
149 
150 	if (doconfigtimer() || busy)
151 		return (FILTER_HANDLED);
152 	state = DPCPU_PTR(timerstate);
153 	now = state->now;
154 	CTR3(KTR_SPARE2, "ipi  at %d:    now  %d.%08x",
155 	    curcpu, (int)(now >> 32), (u_int)(now & 0xffffffff));
156 	done = handleevents(now, 0);
157 	return (done ? FILTER_HANDLED : FILTER_STRAY);
158 }
159 
160 /*
161  * Handle all events for specified time on this CPU
162  */
163 static int
164 handleevents(sbintime_t now, int fake)
165 {
166 	sbintime_t t, *hct;
167 	struct trapframe *frame;
168 	struct pcpu_state *state;
169 	int usermode;
170 	int done, runs;
171 
172 	CTR3(KTR_SPARE2, "handle at %d:  now  %d.%08x",
173 	    curcpu, (int)(now >> 32), (u_int)(now & 0xffffffff));
174 	done = 0;
175 	if (fake) {
176 		frame = NULL;
177 		usermode = 0;
178 	} else {
179 		frame = curthread->td_intr_frame;
180 		usermode = TRAPF_USERMODE(frame);
181 	}
182 
183 	state = DPCPU_PTR(timerstate);
184 
185 	runs = 0;
186 	while (now >= state->nexthard) {
187 		state->nexthard += tick_sbt;
188 		runs++;
189 	}
190 	if (runs) {
191 		hct = DPCPU_PTR(hardclocktime);
192 		*hct = state->nexthard - tick_sbt;
193 		if (fake < 2) {
194 			hardclock_cnt(runs, usermode);
195 			done = 1;
196 		}
197 	}
198 	runs = 0;
199 	while (now >= state->nextstat) {
200 		state->nextstat += statperiod;
201 		runs++;
202 	}
203 	if (runs && fake < 2) {
204 		statclock_cnt(runs, usermode);
205 		done = 1;
206 	}
207 	if (profiling) {
208 		runs = 0;
209 		while (now >= state->nextprof) {
210 			state->nextprof += profperiod;
211 			runs++;
212 		}
213 		if (runs && !fake) {
214 			profclock_cnt(runs, usermode, TRAPF_PC(frame));
215 			done = 1;
216 		}
217 	} else
218 		state->nextprof = state->nextstat;
219 	if (now >= state->nextcallopt) {
220 		state->nextcall = state->nextcallopt = SBT_MAX;
221 		callout_process(now);
222 	}
223 
224 #ifdef KDTRACE_HOOKS
225 	if (fake == 0 && now >= state->nextcyc && cyclic_clock_func != NULL) {
226 		state->nextcyc = SBT_MAX;
227 		(*cyclic_clock_func)(frame);
228 	}
229 #endif
230 
231 	t = getnextcpuevent(0);
232 	ET_HW_LOCK(state);
233 	if (!busy) {
234 		state->idle = 0;
235 		state->nextevent = t;
236 		loadtimer(now, (fake == 2) &&
237 		    (timer->et_flags & ET_FLAGS_PERCPU));
238 	}
239 	ET_HW_UNLOCK(state);
240 	return (done);
241 }
242 
243 /*
244  * Schedule binuptime of the next event on current CPU.
245  */
246 static sbintime_t
247 getnextcpuevent(int idle)
248 {
249 	sbintime_t event;
250 	struct pcpu_state *state;
251 	u_int hardfreq;
252 
253 	state = DPCPU_PTR(timerstate);
254 	/* Handle hardclock() events, skipping some if CPU is idle. */
255 	event = state->nexthard;
256 	if (idle) {
257 		hardfreq = (u_int)hz / 2;
258 		if (tc_min_ticktock_freq > 2
259 #ifdef SMP
260 		    && curcpu == CPU_FIRST()
261 #endif
262 		    )
263 			hardfreq = hz / tc_min_ticktock_freq;
264 		if (hardfreq > 1)
265 			event += tick_sbt * (hardfreq - 1);
266 	}
267 	/* Handle callout events. */
268 	if (event > state->nextcall)
269 		event = state->nextcall;
270 	if (!idle) { /* If CPU is active - handle other types of events. */
271 		if (event > state->nextstat)
272 			event = state->nextstat;
273 		if (profiling && event > state->nextprof)
274 			event = state->nextprof;
275 	}
276 #ifdef KDTRACE_HOOKS
277 	if (event > state->nextcyc)
278 		event = state->nextcyc;
279 #endif
280 	return (event);
281 }
282 
283 /*
284  * Schedule binuptime of the next event on all CPUs.
285  */
286 static sbintime_t
287 getnextevent(void)
288 {
289 	struct pcpu_state *state;
290 	sbintime_t event;
291 #ifdef SMP
292 	int	cpu;
293 #endif
294 	int	c;
295 
296 	state = DPCPU_PTR(timerstate);
297 	event = state->nextevent;
298 	c = -1;
299 #ifdef SMP
300 	if ((timer->et_flags & ET_FLAGS_PERCPU) == 0) {
301 		CPU_FOREACH(cpu) {
302 			state = DPCPU_ID_PTR(cpu, timerstate);
303 			if (event > state->nextevent) {
304 				event = state->nextevent;
305 				c = cpu;
306 			}
307 		}
308 	}
309 #endif
310 	CTR4(KTR_SPARE2, "next at %d:    next %d.%08x by %d",
311 	    curcpu, (int)(event >> 32), (u_int)(event & 0xffffffff), c);
312 	return (event);
313 }
314 
315 /* Hardware timer callback function. */
316 static void
317 timercb(struct eventtimer *et, void *arg)
318 {
319 	sbintime_t now;
320 	sbintime_t *next;
321 	struct pcpu_state *state;
322 #ifdef SMP
323 	int cpu, bcast;
324 #endif
325 
326 	/* Do not touch anything if somebody reconfiguring timers. */
327 	if (busy)
328 		return;
329 	/* Update present and next tick times. */
330 	state = DPCPU_PTR(timerstate);
331 	if (et->et_flags & ET_FLAGS_PERCPU) {
332 		next = &state->nexttick;
333 	} else
334 		next = &nexttick;
335 	now = sbinuptime();
336 	if (periodic)
337 		*next = now + timerperiod;
338 	else
339 		*next = -1;	/* Next tick is not scheduled yet. */
340 	state->now = now;
341 	CTR3(KTR_SPARE2, "intr at %d:    now  %d.%08x",
342 	    curcpu, (int)(now >> 32), (u_int)(now & 0xffffffff));
343 
344 #ifdef SMP
345 	/* Prepare broadcasting to other CPUs for non-per-CPU timers. */
346 	bcast = 0;
347 	if ((et->et_flags & ET_FLAGS_PERCPU) == 0 && smp_started) {
348 		CPU_FOREACH(cpu) {
349 			state = DPCPU_ID_PTR(cpu, timerstate);
350 			ET_HW_LOCK(state);
351 			state->now = now;
352 			if (now >= state->nextevent) {
353 				state->nextevent += SBT_1S;
354 				if (curcpu != cpu) {
355 					state->ipi = 1;
356 					bcast = 1;
357 				}
358 			}
359 			ET_HW_UNLOCK(state);
360 		}
361 	}
362 #endif
363 
364 	/* Handle events for this time on this CPU. */
365 	handleevents(now, 0);
366 
367 #ifdef SMP
368 	/* Broadcast interrupt to other CPUs for non-per-CPU timers. */
369 	if (bcast) {
370 		CPU_FOREACH(cpu) {
371 			if (curcpu == cpu)
372 				continue;
373 			state = DPCPU_ID_PTR(cpu, timerstate);
374 			if (state->ipi) {
375 				state->ipi = 0;
376 				ipi_cpu(cpu, IPI_HARDCLOCK);
377 			}
378 		}
379 	}
380 #endif
381 }
382 
383 /*
384  * Load new value into hardware timer.
385  */
386 static void
387 loadtimer(sbintime_t now, int start)
388 {
389 	struct pcpu_state *state;
390 	sbintime_t new;
391 	sbintime_t *next;
392 	uint64_t tmp;
393 	int eq;
394 
395 	if (timer->et_flags & ET_FLAGS_PERCPU) {
396 		state = DPCPU_PTR(timerstate);
397 		next = &state->nexttick;
398 	} else
399 		next = &nexttick;
400 	if (periodic) {
401 		if (start) {
402 			/*
403 			 * Try to start all periodic timers aligned
404 			 * to period to make events synchronous.
405 			 */
406 			tmp = now % timerperiod;
407 			new = timerperiod - tmp;
408 			if (new < tmp)		/* Left less then passed. */
409 				new += timerperiod;
410 			CTR5(KTR_SPARE2, "load p at %d:   now %d.%08x first in %d.%08x",
411 			    curcpu, (int)(now >> 32), (u_int)(now & 0xffffffff),
412 			    (int)(new >> 32), (u_int)(new & 0xffffffff));
413 			*next = new + now;
414 			et_start(timer, new, timerperiod);
415 		}
416 	} else {
417 		new = getnextevent();
418 		eq = (new == *next);
419 		CTR4(KTR_SPARE2, "load at %d:    next %d.%08x eq %d",
420 		    curcpu, (int)(new >> 32), (u_int)(new & 0xffffffff), eq);
421 		if (!eq) {
422 			*next = new;
423 			et_start(timer, new - now, 0);
424 		}
425 	}
426 }
427 
428 /*
429  * Prepare event timer parameters after configuration changes.
430  */
431 static void
432 setuptimer(void)
433 {
434 	int freq;
435 
436 	if (periodic && (timer->et_flags & ET_FLAGS_PERIODIC) == 0)
437 		periodic = 0;
438 	else if (!periodic && (timer->et_flags & ET_FLAGS_ONESHOT) == 0)
439 		periodic = 1;
440 	singlemul = MIN(MAX(singlemul, 1), 20);
441 	freq = hz * singlemul;
442 	while (freq < (profiling ? profhz : stathz))
443 		freq += hz;
444 	freq = round_freq(timer, freq);
445 	timerperiod = SBT_1S / freq;
446 }
447 
448 /*
449  * Reconfigure specified per-CPU timer on other CPU. Called from IPI handler.
450  */
451 static int
452 doconfigtimer(void)
453 {
454 	sbintime_t now;
455 	struct pcpu_state *state;
456 
457 	state = DPCPU_PTR(timerstate);
458 	switch (atomic_load_acq_int(&state->action)) {
459 	case 1:
460 		now = sbinuptime();
461 		ET_HW_LOCK(state);
462 		loadtimer(now, 1);
463 		ET_HW_UNLOCK(state);
464 		state->handle = 0;
465 		atomic_store_rel_int(&state->action, 0);
466 		return (1);
467 	case 2:
468 		ET_HW_LOCK(state);
469 		et_stop(timer);
470 		ET_HW_UNLOCK(state);
471 		state->handle = 0;
472 		atomic_store_rel_int(&state->action, 0);
473 		return (1);
474 	}
475 	if (atomic_readandclear_int(&state->handle) && !busy) {
476 		now = sbinuptime();
477 		handleevents(now, 0);
478 		return (1);
479 	}
480 	return (0);
481 }
482 
483 /*
484  * Reconfigure specified timer.
485  * For per-CPU timers use IPI to make other CPUs to reconfigure.
486  */
487 static void
488 configtimer(int start)
489 {
490 	sbintime_t now, next;
491 	struct pcpu_state *state;
492 	int cpu;
493 
494 	if (start) {
495 		setuptimer();
496 		now = sbinuptime();
497 	} else
498 		now = 0;
499 	critical_enter();
500 	ET_HW_LOCK(DPCPU_PTR(timerstate));
501 	if (start) {
502 		/* Initialize time machine parameters. */
503 		next = now + timerperiod;
504 		if (periodic)
505 			nexttick = next;
506 		else
507 			nexttick = -1;
508 		CPU_FOREACH(cpu) {
509 			state = DPCPU_ID_PTR(cpu, timerstate);
510 			state->now = now;
511 			if (!smp_started && cpu != CPU_FIRST())
512 				state->nextevent = SBT_MAX;
513 			else
514 				state->nextevent = next;
515 			if (periodic)
516 				state->nexttick = next;
517 			else
518 				state->nexttick = -1;
519 			state->nexthard = next;
520 			state->nextstat = next;
521 			state->nextprof = next;
522 			state->nextcall = next;
523 			state->nextcallopt = next;
524 			hardclock_sync(cpu);
525 		}
526 		busy = 0;
527 		/* Start global timer or per-CPU timer of this CPU. */
528 		loadtimer(now, 1);
529 	} else {
530 		busy = 1;
531 		/* Stop global timer or per-CPU timer of this CPU. */
532 		et_stop(timer);
533 	}
534 	ET_HW_UNLOCK(DPCPU_PTR(timerstate));
535 #ifdef SMP
536 	/* If timer is global or there is no other CPUs yet - we are done. */
537 	if ((timer->et_flags & ET_FLAGS_PERCPU) == 0 || !smp_started) {
538 		critical_exit();
539 		return;
540 	}
541 	/* Set reconfigure flags for other CPUs. */
542 	CPU_FOREACH(cpu) {
543 		state = DPCPU_ID_PTR(cpu, timerstate);
544 		atomic_store_rel_int(&state->action,
545 		    (cpu == curcpu) ? 0 : ( start ? 1 : 2));
546 	}
547 	/* Broadcast reconfigure IPI. */
548 	ipi_all_but_self(IPI_HARDCLOCK);
549 	/* Wait for reconfiguration completed. */
550 restart:
551 	cpu_spinwait();
552 	CPU_FOREACH(cpu) {
553 		if (cpu == curcpu)
554 			continue;
555 		state = DPCPU_ID_PTR(cpu, timerstate);
556 		if (atomic_load_acq_int(&state->action))
557 			goto restart;
558 	}
559 #endif
560 	critical_exit();
561 }
562 
563 /*
564  * Calculate nearest frequency supported by hardware timer.
565  */
566 static int
567 round_freq(struct eventtimer *et, int freq)
568 {
569 	uint64_t div;
570 
571 	if (et->et_frequency != 0) {
572 		div = lmax((et->et_frequency + freq / 2) / freq, 1);
573 		if (et->et_flags & ET_FLAGS_POW2DIV)
574 			div = 1 << (flsl(div + div / 2) - 1);
575 		freq = (et->et_frequency + div / 2) / div;
576 	}
577 	if (et->et_min_period > SBT_1S)
578 		panic("Event timer \"%s\" doesn't support sub-second periods!",
579 		    et->et_name);
580 	else if (et->et_min_period != 0)
581 		freq = min(freq, SBT2FREQ(et->et_min_period));
582 	if (et->et_max_period < SBT_1S && et->et_max_period != 0)
583 		freq = max(freq, SBT2FREQ(et->et_max_period));
584 	return (freq);
585 }
586 
587 /*
588  * Configure and start event timers (BSP part).
589  */
590 void
591 cpu_initclocks_bsp(void)
592 {
593 	struct pcpu_state *state;
594 	int base, div, cpu;
595 
596 	mtx_init(&et_hw_mtx, "et_hw_mtx", NULL, MTX_SPIN);
597 	CPU_FOREACH(cpu) {
598 		state = DPCPU_ID_PTR(cpu, timerstate);
599 		mtx_init(&state->et_hw_mtx, "et_hw_mtx", NULL, MTX_SPIN);
600 #ifdef KDTRACE_HOOKS
601 		state->nextcyc = SBT_MAX;
602 #endif
603 		state->nextcall = SBT_MAX;
604 		state->nextcallopt = SBT_MAX;
605 	}
606 	periodic = want_periodic;
607 	/* Grab requested timer or the best of present. */
608 	if (timername[0])
609 		timer = et_find(timername, 0, 0);
610 	if (timer == NULL && periodic) {
611 		timer = et_find(NULL,
612 		    ET_FLAGS_PERIODIC, ET_FLAGS_PERIODIC);
613 	}
614 	if (timer == NULL) {
615 		timer = et_find(NULL,
616 		    ET_FLAGS_ONESHOT, ET_FLAGS_ONESHOT);
617 	}
618 	if (timer == NULL && !periodic) {
619 		timer = et_find(NULL,
620 		    ET_FLAGS_PERIODIC, ET_FLAGS_PERIODIC);
621 	}
622 	if (timer == NULL)
623 		panic("No usable event timer found!");
624 	et_init(timer, timercb, NULL, NULL);
625 
626 	/* Adapt to timer capabilities. */
627 	if (periodic && (timer->et_flags & ET_FLAGS_PERIODIC) == 0)
628 		periodic = 0;
629 	else if (!periodic && (timer->et_flags & ET_FLAGS_ONESHOT) == 0)
630 		periodic = 1;
631 	if (timer->et_flags & ET_FLAGS_C3STOP)
632 		cpu_disable_deep_sleep++;
633 
634 	/*
635 	 * We honor the requested 'hz' value.
636 	 * We want to run stathz in the neighborhood of 128hz.
637 	 * We would like profhz to run as often as possible.
638 	 */
639 	if (singlemul <= 0 || singlemul > 20) {
640 		if (hz >= 1500 || (hz % 128) == 0)
641 			singlemul = 1;
642 		else if (hz >= 750)
643 			singlemul = 2;
644 		else
645 			singlemul = 4;
646 	}
647 	if (periodic) {
648 		base = round_freq(timer, hz * singlemul);
649 		singlemul = max((base + hz / 2) / hz, 1);
650 		hz = (base + singlemul / 2) / singlemul;
651 		if (base <= 128)
652 			stathz = base;
653 		else {
654 			div = base / 128;
655 			if (div >= singlemul && (div % singlemul) == 0)
656 				div++;
657 			stathz = base / div;
658 		}
659 		profhz = stathz;
660 		while ((profhz + stathz) <= 128 * 64)
661 			profhz += stathz;
662 		profhz = round_freq(timer, profhz);
663 	} else {
664 		hz = round_freq(timer, hz);
665 		stathz = round_freq(timer, 127);
666 		profhz = round_freq(timer, stathz * 64);
667 	}
668 	tick = 1000000 / hz;
669 	tick_sbt = SBT_1S / hz;
670 	tick_bt = sbttobt(tick_sbt);
671 	statperiod = SBT_1S / stathz;
672 	profperiod = SBT_1S / profhz;
673 	ET_LOCK();
674 	configtimer(1);
675 	ET_UNLOCK();
676 }
677 
678 /*
679  * Start per-CPU event timers on APs.
680  */
681 void
682 cpu_initclocks_ap(void)
683 {
684 	sbintime_t now;
685 	struct pcpu_state *state;
686 	struct thread *td;
687 
688 	state = DPCPU_PTR(timerstate);
689 	now = sbinuptime();
690 	ET_HW_LOCK(state);
691 	state->now = now;
692 	hardclock_sync(curcpu);
693 	spinlock_enter();
694 	ET_HW_UNLOCK(state);
695 	td = curthread;
696 	td->td_intr_nesting_level++;
697 	handleevents(state->now, 2);
698 	td->td_intr_nesting_level--;
699 	spinlock_exit();
700 }
701 
702 /*
703  * Switch to profiling clock rates.
704  */
705 void
706 cpu_startprofclock(void)
707 {
708 
709 	ET_LOCK();
710 	if (profiling == 0) {
711 		if (periodic) {
712 			configtimer(0);
713 			profiling = 1;
714 			configtimer(1);
715 		} else
716 			profiling = 1;
717 	} else
718 		profiling++;
719 	ET_UNLOCK();
720 }
721 
722 /*
723  * Switch to regular clock rates.
724  */
725 void
726 cpu_stopprofclock(void)
727 {
728 
729 	ET_LOCK();
730 	if (profiling == 1) {
731 		if (periodic) {
732 			configtimer(0);
733 			profiling = 0;
734 			configtimer(1);
735 		} else
736 		profiling = 0;
737 	} else
738 		profiling--;
739 	ET_UNLOCK();
740 }
741 
742 /*
743  * Switch to idle mode (all ticks handled).
744  */
745 sbintime_t
746 cpu_idleclock(void)
747 {
748 	sbintime_t now, t;
749 	struct pcpu_state *state;
750 
751 	if (idletick || busy ||
752 	    (periodic && (timer->et_flags & ET_FLAGS_PERCPU))
753 #ifdef DEVICE_POLLING
754 	    || curcpu == CPU_FIRST()
755 #endif
756 	    )
757 		return (-1);
758 	state = DPCPU_PTR(timerstate);
759 	if (periodic)
760 		now = state->now;
761 	else
762 		now = sbinuptime();
763 	CTR3(KTR_SPARE2, "idle at %d:    now  %d.%08x",
764 	    curcpu, (int)(now >> 32), (u_int)(now & 0xffffffff));
765 	t = getnextcpuevent(1);
766 	ET_HW_LOCK(state);
767 	state->idle = 1;
768 	state->nextevent = t;
769 	if (!periodic)
770 		loadtimer(now, 0);
771 	ET_HW_UNLOCK(state);
772 	return (MAX(t - now, 0));
773 }
774 
775 /*
776  * Switch to active mode (skip empty ticks).
777  */
778 void
779 cpu_activeclock(void)
780 {
781 	sbintime_t now;
782 	struct pcpu_state *state;
783 	struct thread *td;
784 
785 	state = DPCPU_PTR(timerstate);
786 	if (state->idle == 0 || busy)
787 		return;
788 	if (periodic)
789 		now = state->now;
790 	else
791 		now = sbinuptime();
792 	CTR3(KTR_SPARE2, "active at %d:  now  %d.%08x",
793 	    curcpu, (int)(now >> 32), (u_int)(now & 0xffffffff));
794 	spinlock_enter();
795 	td = curthread;
796 	td->td_intr_nesting_level++;
797 	handleevents(now, 1);
798 	td->td_intr_nesting_level--;
799 	spinlock_exit();
800 }
801 
802 /*
803  * Change the frequency of the given timer.  This changes et->et_frequency and
804  * if et is the active timer it reconfigures the timer on all CPUs.  This is
805  * intended to be a private interface for the use of et_change_frequency() only.
806  */
807 void
808 cpu_et_frequency(struct eventtimer *et, uint64_t newfreq)
809 {
810 
811 	ET_LOCK();
812 	if (et == timer) {
813 		configtimer(0);
814 		et->et_frequency = newfreq;
815 		configtimer(1);
816 	} else
817 		et->et_frequency = newfreq;
818 	ET_UNLOCK();
819 }
820 
821 #ifdef KDTRACE_HOOKS
822 void
823 clocksource_cyc_set(const struct bintime *bt)
824 {
825 	sbintime_t now, t;
826 	struct pcpu_state *state;
827 
828 	/* Do not touch anything if somebody reconfiguring timers. */
829 	if (busy)
830 		return;
831 	t = bttosbt(*bt);
832 	state = DPCPU_PTR(timerstate);
833 	if (periodic)
834 		now = state->now;
835 	else
836 		now = sbinuptime();
837 
838 	CTR5(KTR_SPARE2, "set_cyc at %d:  now  %d.%08x  t  %d.%08x",
839 	    curcpu, (int)(now >> 32), (u_int)(now & 0xffffffff),
840 	    (int)(t >> 32), (u_int)(t & 0xffffffff));
841 
842 	ET_HW_LOCK(state);
843 	if (t == state->nextcyc)
844 		goto done;
845 	state->nextcyc = t;
846 	if (t >= state->nextevent)
847 		goto done;
848 	state->nextevent = t;
849 	if (!periodic)
850 		loadtimer(now, 0);
851 done:
852 	ET_HW_UNLOCK(state);
853 }
854 #endif
855 
856 void
857 cpu_new_callout(int cpu, sbintime_t bt, sbintime_t bt_opt)
858 {
859 	struct pcpu_state *state;
860 
861 	/* Do not touch anything if somebody reconfiguring timers. */
862 	if (busy)
863 		return;
864 	CTR6(KTR_SPARE2, "new co at %d:    on %d at %d.%08x - %d.%08x",
865 	    curcpu, cpu, (int)(bt_opt >> 32), (u_int)(bt_opt & 0xffffffff),
866 	    (int)(bt >> 32), (u_int)(bt & 0xffffffff));
867 	state = DPCPU_ID_PTR(cpu, timerstate);
868 	ET_HW_LOCK(state);
869 
870 	/*
871 	 * If there is callout time already set earlier -- do nothing.
872 	 * This check may appear redundant because we check already in
873 	 * callout_process() but this double check guarantees we're safe
874 	 * with respect to race conditions between interrupts execution
875 	 * and scheduling.
876 	 */
877 	state->nextcallopt = bt_opt;
878 	if (bt >= state->nextcall)
879 		goto done;
880 	state->nextcall = bt;
881 	/* If there is some other event set earlier -- do nothing. */
882 	if (bt >= state->nextevent)
883 		goto done;
884 	state->nextevent = bt;
885 	/* If timer is periodic -- there is nothing to reprogram. */
886 	if (periodic)
887 		goto done;
888 	/* If timer is global or of the current CPU -- reprogram it. */
889 	if ((timer->et_flags & ET_FLAGS_PERCPU) == 0 || cpu == curcpu) {
890 		loadtimer(sbinuptime(), 0);
891 done:
892 		ET_HW_UNLOCK(state);
893 		return;
894 	}
895 	/* Otherwise make other CPU to reprogram it. */
896 	state->handle = 1;
897 	ET_HW_UNLOCK(state);
898 #ifdef SMP
899 	ipi_cpu(cpu, IPI_HARDCLOCK);
900 #endif
901 }
902 
903 /*
904  * Report or change the active event timers hardware.
905  */
906 static int
907 sysctl_kern_eventtimer_timer(SYSCTL_HANDLER_ARGS)
908 {
909 	char buf[32];
910 	struct eventtimer *et;
911 	int error;
912 
913 	ET_LOCK();
914 	et = timer;
915 	snprintf(buf, sizeof(buf), "%s", et->et_name);
916 	ET_UNLOCK();
917 	error = sysctl_handle_string(oidp, buf, sizeof(buf), req);
918 	ET_LOCK();
919 	et = timer;
920 	if (error != 0 || req->newptr == NULL ||
921 	    strcasecmp(buf, et->et_name) == 0) {
922 		ET_UNLOCK();
923 		return (error);
924 	}
925 	et = et_find(buf, 0, 0);
926 	if (et == NULL) {
927 		ET_UNLOCK();
928 		return (ENOENT);
929 	}
930 	configtimer(0);
931 	et_free(timer);
932 	if (et->et_flags & ET_FLAGS_C3STOP)
933 		cpu_disable_deep_sleep++;
934 	if (timer->et_flags & ET_FLAGS_C3STOP)
935 		cpu_disable_deep_sleep--;
936 	periodic = want_periodic;
937 	timer = et;
938 	et_init(timer, timercb, NULL, NULL);
939 	configtimer(1);
940 	ET_UNLOCK();
941 	return (error);
942 }
943 SYSCTL_PROC(_kern_eventtimer, OID_AUTO, timer,
944     CTLTYPE_STRING | CTLFLAG_RW | CTLFLAG_MPSAFE,
945     0, 0, sysctl_kern_eventtimer_timer, "A", "Chosen event timer");
946 
947 /*
948  * Report or change the active event timer periodicity.
949  */
950 static int
951 sysctl_kern_eventtimer_periodic(SYSCTL_HANDLER_ARGS)
952 {
953 	int error, val;
954 
955 	val = periodic;
956 	error = sysctl_handle_int(oidp, &val, 0, req);
957 	if (error != 0 || req->newptr == NULL)
958 		return (error);
959 	ET_LOCK();
960 	configtimer(0);
961 	periodic = want_periodic = val;
962 	configtimer(1);
963 	ET_UNLOCK();
964 	return (error);
965 }
966 SYSCTL_PROC(_kern_eventtimer, OID_AUTO, periodic,
967     CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_MPSAFE,
968     0, 0, sysctl_kern_eventtimer_periodic, "I", "Enable event timer periodic mode");
969