xref: /openbsd-src/sys/kern/kern_clock.c (revision ac9b4aacc1da35008afea06a5d23c2f2dea9b93e)
1 /*	$OpenBSD: kern_clock.c,v 1.75 2012/08/02 03:18:48 guenther Exp $	*/
2 /*	$NetBSD: kern_clock.c,v 1.34 1996/06/09 04:51:03 briggs Exp $	*/
3 
4 /*-
5  * Copyright (c) 1982, 1986, 1991, 1993
6  *	The Regents of the University of California.  All rights reserved.
7  * (c) UNIX System Laboratories, Inc.
8  * All or some portions of this file are derived from material licensed
9  * to the University of California by American Telephone and Telegraph
10  * Co. or Unix System Laboratories, Inc. and are reproduced herein with
11  * the permission of UNIX System Laboratories, Inc.
12  *
13  * Redistribution and use in source and binary forms, with or without
14  * modification, are permitted provided that the following conditions
15  * are met:
16  * 1. Redistributions of source code must retain the above copyright
17  *    notice, this list of conditions and the following disclaimer.
18  * 2. Redistributions in binary form must reproduce the above copyright
19  *    notice, this list of conditions and the following disclaimer in the
20  *    documentation and/or other materials provided with the distribution.
21  * 3. Neither the name of the University nor the names of its contributors
22  *    may be used to endorse or promote products derived from this software
23  *    without specific prior written permission.
24  *
25  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
26  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
27  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
28  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
29  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
30  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
31  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
32  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
33  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
34  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
35  * SUCH DAMAGE.
36  *
37  *	@(#)kern_clock.c	8.5 (Berkeley) 1/21/94
38  */
39 
40 #include <sys/param.h>
41 #include <sys/systm.h>
42 #include <sys/dkstat.h>
43 #include <sys/timeout.h>
44 #include <sys/kernel.h>
45 #include <sys/limits.h>
46 #include <sys/proc.h>
47 #include <sys/user.h>
48 #include <sys/resourcevar.h>
49 #include <sys/signalvar.h>
50 #include <uvm/uvm_extern.h>
51 #include <sys/sysctl.h>
52 #include <sys/sched.h>
53 #ifdef __HAVE_TIMECOUNTER
54 #include <sys/timetc.h>
55 #endif
56 
57 #include <machine/cpu.h>
58 
59 #ifdef GPROF
60 #include <sys/gmon.h>
61 #endif
62 
63 /*
64  * Clock handling routines.
65  *
66  * This code is written to operate with two timers that run independently of
67  * each other.  The main clock, running hz times per second, is used to keep
68  * track of real time.  The second timer handles kernel and user profiling,
69  * and does resource use estimation.  If the second timer is programmable,
70  * it is randomized to avoid aliasing between the two clocks.  For example,
71  * the randomization prevents an adversary from always giving up the cpu
72  * just before its quantum expires.  Otherwise, it would never accumulate
73  * cpu ticks.  The mean frequency of the second timer is stathz.
74  *
75  * If no second timer exists, stathz will be zero; in this case we drive
76  * profiling and statistics off the main clock.  This WILL NOT be accurate;
77  * do not do it unless absolutely necessary.
78  *
79  * The statistics clock may (or may not) be run at a higher rate while
80  * profiling.  This profile clock runs at profhz.  We require that profhz
81  * be an integral multiple of stathz.
82  *
83  * If the statistics clock is running fast, it must be divided by the ratio
84  * profhz/stathz for statistics.  (For profiling, every tick counts.)
85  */
86 
87 /*
88  * Bump a timeval by a small number of usec's.
89  */
90 #define BUMPTIME(t, usec) { \
91 	volatile struct timeval *tp = (t); \
92 	long us; \
93  \
94 	tp->tv_usec = us = tp->tv_usec + (usec); \
95 	if (us >= 1000000) { \
96 		tp->tv_usec = us - 1000000; \
97 		tp->tv_sec++; \
98 	} \
99 }
100 
101 int	stathz;
102 int	schedhz;
103 int	profhz;
104 int	profprocs;
105 int	ticks;
106 static int psdiv, pscnt;		/* prof => stat divider */
107 int	psratio;			/* ratio: prof / stat */
108 
109 long cp_time[CPUSTATES];
110 
111 #ifndef __HAVE_TIMECOUNTER
112 int	tickfix, tickfixinterval;	/* used if tick not really integral */
113 static int tickfixcnt;			/* accumulated fractional error */
114 
115 volatile time_t time_second;
116 volatile time_t time_uptime;
117 
118 volatile struct	timeval time
119 	__attribute__((__aligned__(__alignof__(quad_t))));
120 volatile struct	timeval mono_time;
121 #endif
122 
123 void	*softclock_si;
124 
125 /*
126  * Initialize clock frequencies and start both clocks running.
127  */
128 void
129 initclocks(void)
130 {
131 	int i;
132 
133 	softclock_si = softintr_establish(IPL_SOFTCLOCK, softclock, NULL);
134 	if (softclock_si == NULL)
135 		panic("initclocks: unable to register softclock intr");
136 
137 	/*
138 	 * Set divisors to 1 (normal case) and let the machine-specific
139 	 * code do its bit.
140 	 */
141 	psdiv = pscnt = 1;
142 	cpu_initclocks();
143 
144 	/*
145 	 * Compute profhz/stathz, and fix profhz if needed.
146 	 */
147 	i = stathz ? stathz : hz;
148 	if (profhz == 0)
149 		profhz = i;
150 	psratio = profhz / i;
151 
152 	/* For very large HZ, ensure that division by 0 does not occur later */
153 	if (tickadj == 0)
154 		tickadj = 1;
155 
156 #ifdef __HAVE_TIMECOUNTER
157 	inittimecounter();
158 #endif
159 }
160 
161 /*
162  * hardclock does the accounting needed for ITIMER_PROF and ITIMER_VIRTUAL.
163  * We don't want to send signals with psignal from hardclock because it makes
164  * MULTIPROCESSOR locking very complicated. Instead we use a small trick
165  * to send the signals safely and without blocking too many interrupts
166  * while doing that (signal handling can be heavy).
167  *
168  * hardclock detects that the itimer has expired, and schedules a timeout
169  * to deliver the signal. This works because of the following reasons:
170  *  - The timeout can be scheduled with a 1 tick time because we're
171  *    doing it before the timeout processing in hardclock. So it will
172  *    be scheduled to run as soon as possible.
173  *  - The timeout will be run in softclock which will run before we
174  *    return to userland and process pending signals.
175  *  - If the system is so busy that several VIRTUAL/PROF ticks are
176  *    sent before softclock processing, we'll send only one signal.
177  *    But if we'd send the signal from hardclock only one signal would
178  *    be delivered to the user process. So userland will only see one
179  *    signal anyway.
180  */
181 
182 void
183 virttimer_trampoline(void *v)
184 {
185 	struct process *pr = v;
186 
187 	psignal(pr->ps_mainproc, SIGVTALRM);
188 }
189 
190 void
191 proftimer_trampoline(void *v)
192 {
193 	struct process *pr = v;
194 
195 	psignal(pr->ps_mainproc, SIGPROF);
196 }
197 
198 /*
199  * The real-time timer, interrupting hz times per second.
200  */
201 void
202 hardclock(struct clockframe *frame)
203 {
204 	struct proc *p;
205 #ifndef __HAVE_TIMECOUNTER
206 	int delta;
207 	extern int tickdelta;
208 	extern long timedelta;
209 	extern int64_t ntp_tick_permanent;
210 	extern int64_t ntp_tick_acc;
211 #endif
212 	struct cpu_info *ci = curcpu();
213 
214 	p = curproc;
215 	if (p && ((p->p_flag & (P_SYSTEM | P_WEXIT)) == 0)) {
216 		struct process *pr = p->p_p;
217 
218 		/*
219 		 * Run current process's virtual and profile time, as needed.
220 		 */
221 		if (CLKF_USERMODE(frame) &&
222 		    timerisset(&pr->ps_timer[ITIMER_VIRTUAL].it_value) &&
223 		    itimerdecr(&pr->ps_timer[ITIMER_VIRTUAL], tick) == 0)
224 			timeout_add(&pr->ps_virt_to, 1);
225 		if (timerisset(&pr->ps_timer[ITIMER_PROF].it_value) &&
226 		    itimerdecr(&pr->ps_timer[ITIMER_PROF], tick) == 0)
227 			timeout_add(&pr->ps_prof_to, 1);
228 	}
229 
230 	/*
231 	 * If no separate statistics clock is available, run it from here.
232 	 */
233 	if (stathz == 0)
234 		statclock(frame);
235 
236 	if (--ci->ci_schedstate.spc_rrticks <= 0)
237 		roundrobin(ci);
238 
239 	/*
240 	 * If we are not the primary CPU, we're not allowed to do
241 	 * any more work.
242 	 */
243 	if (CPU_IS_PRIMARY(ci) == 0)
244 		return;
245 
246 #ifndef __HAVE_TIMECOUNTER
247 	/*
248 	 * Increment the time-of-day.  The increment is normally just
249 	 * ``tick''.  If the machine is one which has a clock frequency
250 	 * such that ``hz'' would not divide the second evenly into
251 	 * milliseconds, a periodic adjustment must be applied.  Finally,
252 	 * if we are still adjusting the time (see adjtime()),
253 	 * ``tickdelta'' may also be added in.
254 	 */
255 
256 	delta = tick;
257 
258 	if (tickfix) {
259 		tickfixcnt += tickfix;
260 		if (tickfixcnt >= tickfixinterval) {
261 			delta++;
262 			tickfixcnt -= tickfixinterval;
263 		}
264 	}
265 	/* Imprecise 4bsd adjtime() handling */
266 	if (timedelta != 0) {
267 		delta += tickdelta;
268 		timedelta -= tickdelta;
269 	}
270 
271 	/*
272 	 * ntp_tick_permanent accumulates the clock correction each
273 	 * tick. The unit is ns per tick shifted left 32 bits. If we have
274 	 * accumulated more than 1us, we bump delta in the right
275 	 * direction. Use a loop to avoid long long div; typically
276 	 * the loops will be executed 0 or 1 iteration.
277 	 */
278 	if (ntp_tick_permanent != 0) {
279 		ntp_tick_acc += ntp_tick_permanent;
280 		while (ntp_tick_acc >= (1000LL << 32)) {
281 			delta++;
282 			ntp_tick_acc -= (1000LL << 32);
283 		}
284 		while (ntp_tick_acc <= -(1000LL << 32)) {
285 			delta--;
286 			ntp_tick_acc += (1000LL << 32);
287 		}
288 	}
289 
290 	BUMPTIME(&time, delta);
291 	BUMPTIME(&mono_time, delta);
292 	time_second = time.tv_sec;
293 	time_uptime = mono_time.tv_sec;
294 #else
295 	tc_ticktock();
296 #endif
297 
298 	/*
299 	 * Update real-time timeout queue.
300 	 * Process callouts at a very low cpu priority, so we don't keep the
301 	 * relatively high clock interrupt priority any longer than necessary.
302 	 */
303 	if (timeout_hardclock_update())
304 		softintr_schedule(softclock_si);
305 }
306 
307 /*
308  * Compute number of hz until specified time.  Used to
309  * compute the second argument to timeout_add() from an absolute time.
310  */
311 int
312 hzto(const struct timeval *tv)
313 {
314 	struct timeval now;
315 	unsigned long ticks;
316 	long sec, usec;
317 
318 	/*
319 	 * If the number of usecs in the whole seconds part of the time
320 	 * difference fits in a long, then the total number of usecs will
321 	 * fit in an unsigned long.  Compute the total and convert it to
322 	 * ticks, rounding up and adding 1 to allow for the current tick
323 	 * to expire.  Rounding also depends on unsigned long arithmetic
324 	 * to avoid overflow.
325 	 *
326 	 * Otherwise, if the number of ticks in the whole seconds part of
327 	 * the time difference fits in a long, then convert the parts to
328 	 * ticks separately and add, using similar rounding methods and
329 	 * overflow avoidance.  This method would work in the previous
330 	 * case but it is slightly slower and assumes that hz is integral.
331 	 *
332 	 * Otherwise, round the time difference down to the maximum
333 	 * representable value.
334 	 *
335 	 * If ints have 32 bits, then the maximum value for any timeout in
336 	 * 10ms ticks is 248 days.
337 	 */
338 	getmicrotime(&now);
339 	sec = tv->tv_sec - now.tv_sec;
340 	usec = tv->tv_usec - now.tv_usec;
341 	if (usec < 0) {
342 		sec--;
343 		usec += 1000000;
344 	}
345 	if (sec < 0 || (sec == 0 && usec <= 0)) {
346 		ticks = 0;
347 	} else if (sec <= LONG_MAX / 1000000)
348 		ticks = (sec * 1000000 + (unsigned long)usec + (tick - 1))
349 		    / tick + 1;
350 	else if (sec <= LONG_MAX / hz)
351 		ticks = sec * hz
352 		    + ((unsigned long)usec + (tick - 1)) / tick + 1;
353 	else
354 		ticks = LONG_MAX;
355 	if (ticks > INT_MAX)
356 		ticks = INT_MAX;
357 	return ((int)ticks);
358 }
359 
360 /*
361  * Compute number of hz in the specified amount of time.
362  */
363 int
364 tvtohz(const struct timeval *tv)
365 {
366 	unsigned long ticks;
367 	long sec, usec;
368 
369 	/*
370 	 * If the number of usecs in the whole seconds part of the time
371 	 * fits in a long, then the total number of usecs will
372 	 * fit in an unsigned long.  Compute the total and convert it to
373 	 * ticks, rounding up and adding 1 to allow for the current tick
374 	 * to expire.  Rounding also depends on unsigned long arithmetic
375 	 * to avoid overflow.
376 	 *
377 	 * Otherwise, if the number of ticks in the whole seconds part of
378 	 * the time fits in a long, then convert the parts to
379 	 * ticks separately and add, using similar rounding methods and
380 	 * overflow avoidance.  This method would work in the previous
381 	 * case but it is slightly slower and assumes that hz is integral.
382 	 *
383 	 * Otherwise, round the time down to the maximum
384 	 * representable value.
385 	 *
386 	 * If ints have 32 bits, then the maximum value for any timeout in
387 	 * 10ms ticks is 248 days.
388 	 */
389 	sec = tv->tv_sec;
390 	usec = tv->tv_usec;
391 	if (sec < 0 || (sec == 0 && usec <= 0))
392 		ticks = 0;
393 	else if (sec <= LONG_MAX / 1000000)
394 		ticks = (sec * 1000000 + (unsigned long)usec + (tick - 1))
395 		    / tick + 1;
396 	else if (sec <= LONG_MAX / hz)
397 		ticks = sec * hz
398 		    + ((unsigned long)usec + (tick - 1)) / tick + 1;
399 	else
400 		ticks = LONG_MAX;
401 	if (ticks > INT_MAX)
402 		ticks = INT_MAX;
403 	return ((int)ticks);
404 }
405 
406 /*
407  * Start profiling on a process.
408  *
409  * Kernel profiling passes proc0 which never exits and hence
410  * keeps the profile clock running constantly.
411  */
412 void
413 startprofclock(struct process *pr)
414 {
415 	int s;
416 
417 	if ((pr->ps_flags & PS_PROFIL) == 0) {
418 		atomic_setbits_int(&pr->ps_flags, PS_PROFIL);
419 		if (++profprocs == 1 && stathz != 0) {
420 			s = splstatclock();
421 			psdiv = pscnt = psratio;
422 			setstatclockrate(profhz);
423 			splx(s);
424 		}
425 	}
426 }
427 
428 /*
429  * Stop profiling on a process.
430  */
431 void
432 stopprofclock(struct process *pr)
433 {
434 	int s;
435 
436 	if (pr->ps_flags & PS_PROFIL) {
437 		atomic_clearbits_int(&pr->ps_flags, PS_PROFIL);
438 		if (--profprocs == 0 && stathz != 0) {
439 			s = splstatclock();
440 			psdiv = pscnt = 1;
441 			setstatclockrate(stathz);
442 			splx(s);
443 		}
444 	}
445 }
446 
447 /*
448  * Statistics clock.  Grab profile sample, and if divider reaches 0,
449  * do process and kernel statistics.
450  */
451 void
452 statclock(struct clockframe *frame)
453 {
454 #ifdef GPROF
455 	struct gmonparam *g;
456 	u_long i;
457 #endif
458 	struct cpu_info *ci = curcpu();
459 	struct schedstate_percpu *spc = &ci->ci_schedstate;
460 	struct proc *p = curproc;
461 	struct process *pr;
462 
463 	/*
464 	 * Notice changes in divisor frequency, and adjust clock
465 	 * frequency accordingly.
466 	 */
467 	if (spc->spc_psdiv != psdiv) {
468 		spc->spc_psdiv = psdiv;
469 		spc->spc_pscnt = psdiv;
470 		if (psdiv == 1) {
471 			setstatclockrate(stathz);
472 		} else {
473 			setstatclockrate(profhz);
474 		}
475 	}
476 
477 	if (CLKF_USERMODE(frame)) {
478 		pr = p->p_p;
479 		if (pr->ps_flags & PS_PROFIL)
480 			addupc_intr(p, CLKF_PC(frame));
481 		if (--spc->spc_pscnt > 0)
482 			return;
483 		/*
484 		 * Came from user mode; CPU was in user state.
485 		 * If this process is being profiled record the tick.
486 		 */
487 		p->p_uticks++;
488 		if (pr->ps_nice > NZERO)
489 			spc->spc_cp_time[CP_NICE]++;
490 		else
491 			spc->spc_cp_time[CP_USER]++;
492 	} else {
493 #ifdef GPROF
494 		/*
495 		 * Kernel statistics are just like addupc_intr, only easier.
496 		 */
497 		g = &_gmonparam;
498 		if (g->state == GMON_PROF_ON) {
499 			i = CLKF_PC(frame) - g->lowpc;
500 			if (i < g->textsize) {
501 				i /= HISTFRACTION * sizeof(*g->kcount);
502 				g->kcount[i]++;
503 			}
504 		}
505 #endif
506 #if defined(PROC_PC)
507 		if (p != NULL && p->p_p->ps_flags & PS_PROFIL)
508 			addupc_intr(p, PROC_PC(p));
509 #endif
510 		if (--spc->spc_pscnt > 0)
511 			return;
512 		/*
513 		 * Came from kernel mode, so we were:
514 		 * - handling an interrupt,
515 		 * - doing syscall or trap work on behalf of the current
516 		 *   user process, or
517 		 * - spinning in the idle loop.
518 		 * Whichever it is, charge the time as appropriate.
519 		 * Note that we charge interrupts to the current process,
520 		 * regardless of whether they are ``for'' that process,
521 		 * so that we know how much of its real time was spent
522 		 * in ``non-process'' (i.e., interrupt) work.
523 		 */
524 		if (CLKF_INTR(frame)) {
525 			if (p != NULL)
526 				p->p_iticks++;
527 			spc->spc_cp_time[CP_INTR]++;
528 		} else if (p != NULL && p != spc->spc_idleproc) {
529 			p->p_sticks++;
530 			spc->spc_cp_time[CP_SYS]++;
531 		} else
532 			spc->spc_cp_time[CP_IDLE]++;
533 	}
534 	spc->spc_pscnt = psdiv;
535 
536 	if (p != NULL) {
537 		p->p_cpticks++;
538 		/*
539 		 * If no schedclock is provided, call it here at ~~12-25 Hz;
540 		 * ~~16 Hz is best
541 		 */
542 		if (schedhz == 0) {
543 			if ((++curcpu()->ci_schedstate.spc_schedticks & 3) ==
544 			    0)
545 				schedclock(p);
546 		}
547 	}
548 }
549 
550 /*
551  * Return information about system clocks.
552  */
553 int
554 sysctl_clockrate(char *where, size_t *sizep, void *newp)
555 {
556 	struct clockinfo clkinfo;
557 
558 	/*
559 	 * Construct clockinfo structure.
560 	 */
561 	clkinfo.tick = tick;
562 	clkinfo.tickadj = tickadj;
563 	clkinfo.hz = hz;
564 	clkinfo.profhz = profhz;
565 	clkinfo.stathz = stathz ? stathz : hz;
566 	return (sysctl_rdstruct(where, sizep, newp, &clkinfo, sizeof(clkinfo)));
567 }
568 
569 #ifndef __HAVE_TIMECOUNTER
570 /*
571  * Placeholders until everyone uses the timecounters code.
572  * Won't improve anything except maybe removing a bunch of bugs in fixed code.
573  */
574 
575 void
576 getmicrotime(struct timeval *tvp)
577 {
578 	int s;
579 
580 	s = splhigh();
581 	*tvp = time;
582 	splx(s);
583 }
584 
585 void
586 nanotime(struct timespec *tsp)
587 {
588 	struct timeval tv;
589 
590 	microtime(&tv);
591 	TIMEVAL_TO_TIMESPEC(&tv, tsp);
592 }
593 
594 void
595 getnanotime(struct timespec *tsp)
596 {
597 	struct timeval tv;
598 
599 	getmicrotime(&tv);
600 	TIMEVAL_TO_TIMESPEC(&tv, tsp);
601 }
602 
603 void
604 nanouptime(struct timespec *tsp)
605 {
606 	struct timeval tv;
607 
608 	microuptime(&tv);
609 	TIMEVAL_TO_TIMESPEC(&tv, tsp);
610 }
611 
612 
613 void
614 getnanouptime(struct timespec *tsp)
615 {
616 	struct timeval tv;
617 
618 	getmicrouptime(&tv);
619 	TIMEVAL_TO_TIMESPEC(&tv, tsp);
620 }
621 
622 void
623 microuptime(struct timeval *tvp)
624 {
625 	struct timeval tv;
626 
627 	microtime(&tv);
628 	timersub(&tv, &boottime, tvp);
629 }
630 
631 void
632 getmicrouptime(struct timeval *tvp)
633 {
634 	int s;
635 
636 	s = splhigh();
637 	*tvp = mono_time;
638 	splx(s);
639 }
640 #endif /* __HAVE_TIMECOUNTER */
641