xref: /openbsd/sys/kern/kern_clock.c (revision 17df1aa7)
1 /*	$OpenBSD: kern_clock.c,v 1.70 2010/01/14 23:12:11 schwarze 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 #ifdef __HAVE_TIMECOUNTER
133 	extern void inittimecounter(void);
134 #endif
135 
136 	softclock_si = softintr_establish(IPL_SOFTCLOCK, softclock, NULL);
137 	if (softclock_si == NULL)
138 		panic("initclocks: unable to register softclock intr");
139 
140 	/*
141 	 * Set divisors to 1 (normal case) and let the machine-specific
142 	 * code do its bit.
143 	 */
144 	psdiv = pscnt = 1;
145 	cpu_initclocks();
146 
147 	/*
148 	 * Compute profhz/stathz, and fix profhz if needed.
149 	 */
150 	i = stathz ? stathz : hz;
151 	if (profhz == 0)
152 		profhz = i;
153 	psratio = profhz / i;
154 
155 	/* For very large HZ, ensure that division by 0 does not occur later */
156 	if (tickadj == 0)
157 		tickadj = 1;
158 
159 #ifdef __HAVE_TIMECOUNTER
160 	inittimecounter();
161 #endif
162 }
163 
164 /*
165  * hardclock does the accounting needed for ITIMER_PROF and ITIMER_VIRTUAL.
166  * We don't want to send signals with psignal from hardclock because it makes
167  * MULTIPROCESSOR locking very complicated. Instead we use a small trick
168  * to send the signals safely and without blocking too many interrupts
169  * while doing that (signal handling can be heavy).
170  *
171  * hardclock detects that the itimer has expired, and schedules a timeout
172  * to deliver the signal. This works because of the following reasons:
173  *  - The timeout structures can be in struct pstats because the timers
174  *    can be only activated on curproc (never swapped). Swapout can
175  *    only happen from a kernel thread and softclock runs before threads
176  *    are scheduled.
177  *  - The timeout can be scheduled with a 1 tick time because we're
178  *    doing it before the timeout processing in hardclock. So it will
179  *    be scheduled to run as soon as possible.
180  *  - The timeout will be run in softclock which will run before we
181  *    return to userland and process pending signals.
182  *  - If the system is so busy that several VIRTUAL/PROF ticks are
183  *    sent before softclock processing, we'll send only one signal.
184  *    But if we'd send the signal from hardclock only one signal would
185  *    be delivered to the user process. So userland will only see one
186  *    signal anyway.
187  */
188 
189 void
190 virttimer_trampoline(void *v)
191 {
192 	struct proc *p = v;
193 
194 	psignal(p, SIGVTALRM);
195 }
196 
197 void
198 proftimer_trampoline(void *v)
199 {
200 	struct proc *p = v;
201 
202 	psignal(p, SIGPROF);
203 }
204 
205 /*
206  * The real-time timer, interrupting hz times per second.
207  */
208 void
209 hardclock(struct clockframe *frame)
210 {
211 	struct proc *p;
212 #ifndef __HAVE_TIMECOUNTER
213 	int delta;
214 	extern int tickdelta;
215 	extern long timedelta;
216 	extern int64_t ntp_tick_permanent;
217 	extern int64_t ntp_tick_acc;
218 #endif
219 	struct cpu_info *ci = curcpu();
220 
221 	p = curproc;
222 	if (p && ((p->p_flag & (P_SYSTEM | P_WEXIT)) == 0)) {
223 		struct pstats *pstats;
224 
225 		/*
226 		 * Run current process's virtual and profile time, as needed.
227 		 */
228 		pstats = p->p_stats;
229 		if (CLKF_USERMODE(frame) &&
230 		    timerisset(&pstats->p_timer[ITIMER_VIRTUAL].it_value) &&
231 		    itimerdecr(&pstats->p_timer[ITIMER_VIRTUAL], tick) == 0)
232 			timeout_add(&pstats->p_virt_to, 1);
233 		if (timerisset(&pstats->p_timer[ITIMER_PROF].it_value) &&
234 		    itimerdecr(&pstats->p_timer[ITIMER_PROF], tick) == 0)
235 			timeout_add(&pstats->p_prof_to, 1);
236 	}
237 
238 	/*
239 	 * If no separate statistics clock is available, run it from here.
240 	 */
241 	if (stathz == 0)
242 		statclock(frame);
243 
244 	if (--ci->ci_schedstate.spc_rrticks <= 0)
245 		roundrobin(ci);
246 
247 	/*
248 	 * If we are not the primary CPU, we're not allowed to do
249 	 * any more work.
250 	 */
251 	if (CPU_IS_PRIMARY(ci) == 0)
252 		return;
253 
254 #ifndef __HAVE_TIMECOUNTER
255 	/*
256 	 * Increment the time-of-day.  The increment is normally just
257 	 * ``tick''.  If the machine is one which has a clock frequency
258 	 * such that ``hz'' would not divide the second evenly into
259 	 * milliseconds, a periodic adjustment must be applied.  Finally,
260 	 * if we are still adjusting the time (see adjtime()),
261 	 * ``tickdelta'' may also be added in.
262 	 */
263 
264 	delta = tick;
265 
266 	if (tickfix) {
267 		tickfixcnt += tickfix;
268 		if (tickfixcnt >= tickfixinterval) {
269 			delta++;
270 			tickfixcnt -= tickfixinterval;
271 		}
272 	}
273 	/* Imprecise 4bsd adjtime() handling */
274 	if (timedelta != 0) {
275 		delta += tickdelta;
276 		timedelta -= tickdelta;
277 	}
278 
279 	/*
280 	 * ntp_tick_permanent accumulates the clock correction each
281 	 * tick. The unit is ns per tick shifted left 32 bits. If we have
282 	 * accumulated more than 1us, we bump delta in the right
283 	 * direction. Use a loop to avoid long long div; typically
284 	 * the loops will be executed 0 or 1 iteration.
285 	 */
286 	if (ntp_tick_permanent != 0) {
287 		ntp_tick_acc += ntp_tick_permanent;
288 		while (ntp_tick_acc >= (1000LL << 32)) {
289 			delta++;
290 			ntp_tick_acc -= (1000LL << 32);
291 		}
292 		while (ntp_tick_acc <= -(1000LL << 32)) {
293 			delta--;
294 			ntp_tick_acc += (1000LL << 32);
295 		}
296 	}
297 
298 	BUMPTIME(&time, delta);
299 	BUMPTIME(&mono_time, delta);
300 	time_second = time.tv_sec;
301 	time_uptime = mono_time.tv_sec;
302 #else
303 	tc_ticktock();
304 #endif
305 
306 	/*
307 	 * Update real-time timeout queue.
308 	 * Process callouts at a very low cpu priority, so we don't keep the
309 	 * relatively high clock interrupt priority any longer than necessary.
310 	 */
311 	if (timeout_hardclock_update())
312 		softintr_schedule(softclock_si);
313 }
314 
315 /*
316  * Compute number of hz until specified time.  Used to
317  * compute the second argument to timeout_add() from an absolute time.
318  */
319 int
320 hzto(struct timeval *tv)
321 {
322 	struct timeval now;
323 	unsigned long ticks;
324 	long sec, usec;
325 
326 	/*
327 	 * If the number of usecs in the whole seconds part of the time
328 	 * difference fits in a long, then the total number of usecs will
329 	 * fit in an unsigned long.  Compute the total and convert it to
330 	 * ticks, rounding up and adding 1 to allow for the current tick
331 	 * to expire.  Rounding also depends on unsigned long arithmetic
332 	 * to avoid overflow.
333 	 *
334 	 * Otherwise, if the number of ticks in the whole seconds part of
335 	 * the time difference fits in a long, then convert the parts to
336 	 * ticks separately and add, using similar rounding methods and
337 	 * overflow avoidance.  This method would work in the previous
338 	 * case but it is slightly slower and assumes that hz is integral.
339 	 *
340 	 * Otherwise, round the time difference down to the maximum
341 	 * representable value.
342 	 *
343 	 * If ints have 32 bits, then the maximum value for any timeout in
344 	 * 10ms ticks is 248 days.
345 	 */
346 	getmicrotime(&now);
347 	sec = tv->tv_sec - now.tv_sec;
348 	usec = tv->tv_usec - now.tv_usec;
349 	if (usec < 0) {
350 		sec--;
351 		usec += 1000000;
352 	}
353 	if (sec < 0 || (sec == 0 && usec <= 0)) {
354 		ticks = 0;
355 	} else if (sec <= LONG_MAX / 1000000)
356 		ticks = (sec * 1000000 + (unsigned long)usec + (tick - 1))
357 		    / tick + 1;
358 	else if (sec <= LONG_MAX / hz)
359 		ticks = sec * hz
360 		    + ((unsigned long)usec + (tick - 1)) / tick + 1;
361 	else
362 		ticks = LONG_MAX;
363 	if (ticks > INT_MAX)
364 		ticks = INT_MAX;
365 	return ((int)ticks);
366 }
367 
368 /*
369  * Compute number of hz in the specified amount of time.
370  */
371 int
372 tvtohz(struct timeval *tv)
373 {
374 	unsigned long ticks;
375 	long sec, usec;
376 
377 	/*
378 	 * If the number of usecs in the whole seconds part of the time
379 	 * fits in a long, then the total number of usecs will
380 	 * fit in an unsigned long.  Compute the total and convert it to
381 	 * ticks, rounding up and adding 1 to allow for the current tick
382 	 * to expire.  Rounding also depends on unsigned long arithmetic
383 	 * to avoid overflow.
384 	 *
385 	 * Otherwise, if the number of ticks in the whole seconds part of
386 	 * the time fits in a long, then convert the parts to
387 	 * ticks separately and add, using similar rounding methods and
388 	 * overflow avoidance.  This method would work in the previous
389 	 * case but it is slightly slower and assumes that hz is integral.
390 	 *
391 	 * Otherwise, round the time down to the maximum
392 	 * representable value.
393 	 *
394 	 * If ints have 32 bits, then the maximum value for any timeout in
395 	 * 10ms ticks is 248 days.
396 	 */
397 	sec = tv->tv_sec;
398 	usec = tv->tv_usec;
399 	if (sec < 0 || (sec == 0 && usec <= 0))
400 		ticks = 0;
401 	else if (sec <= LONG_MAX / 1000000)
402 		ticks = (sec * 1000000 + (unsigned long)usec + (tick - 1))
403 		    / tick + 1;
404 	else if (sec <= LONG_MAX / hz)
405 		ticks = sec * hz
406 		    + ((unsigned long)usec + (tick - 1)) / tick + 1;
407 	else
408 		ticks = LONG_MAX;
409 	if (ticks > INT_MAX)
410 		ticks = INT_MAX;
411 	return ((int)ticks);
412 }
413 
414 /*
415  * Start profiling on a process.
416  *
417  * Kernel profiling passes proc0 which never exits and hence
418  * keeps the profile clock running constantly.
419  */
420 void
421 startprofclock(struct proc *p)
422 {
423 	int s;
424 
425 	if ((p->p_flag & P_PROFIL) == 0) {
426 		atomic_setbits_int(&p->p_flag, P_PROFIL);
427 		if (++profprocs == 1 && stathz != 0) {
428 			s = splstatclock();
429 			psdiv = pscnt = psratio;
430 			setstatclockrate(profhz);
431 			splx(s);
432 		}
433 	}
434 }
435 
436 /*
437  * Stop profiling on a process.
438  */
439 void
440 stopprofclock(struct proc *p)
441 {
442 	int s;
443 
444 	if (p->p_flag & P_PROFIL) {
445 		atomic_clearbits_int(&p->p_flag, P_PROFIL);
446 		if (--profprocs == 0 && stathz != 0) {
447 			s = splstatclock();
448 			psdiv = pscnt = 1;
449 			setstatclockrate(stathz);
450 			splx(s);
451 		}
452 	}
453 }
454 
455 /*
456  * Statistics clock.  Grab profile sample, and if divider reaches 0,
457  * do process and kernel statistics.
458  */
459 void
460 statclock(struct clockframe *frame)
461 {
462 #ifdef GPROF
463 	struct gmonparam *g;
464 	u_long i;
465 #endif
466 	struct cpu_info *ci = curcpu();
467 	struct schedstate_percpu *spc = &ci->ci_schedstate;
468 	struct proc *p = curproc;
469 
470 	/*
471 	 * Notice changes in divisor frequency, and adjust clock
472 	 * frequency accordingly.
473 	 */
474 	if (spc->spc_psdiv != psdiv) {
475 		spc->spc_psdiv = psdiv;
476 		spc->spc_pscnt = psdiv;
477 		if (psdiv == 1) {
478 			setstatclockrate(stathz);
479 		} else {
480 			setstatclockrate(profhz);
481 		}
482 	}
483 
484 	if (CLKF_USERMODE(frame)) {
485 		if (p->p_flag & P_PROFIL)
486 			addupc_intr(p, CLKF_PC(frame));
487 		if (--spc->spc_pscnt > 0)
488 			return;
489 		/*
490 		 * Came from user mode; CPU was in user state.
491 		 * If this process is being profiled record the tick.
492 		 */
493 		p->p_uticks++;
494 		if (p->p_nice > NZERO)
495 			spc->spc_cp_time[CP_NICE]++;
496 		else
497 			spc->spc_cp_time[CP_USER]++;
498 	} else {
499 #ifdef GPROF
500 		/*
501 		 * Kernel statistics are just like addupc_intr, only easier.
502 		 */
503 		g = &_gmonparam;
504 		if (g->state == GMON_PROF_ON) {
505 			i = CLKF_PC(frame) - g->lowpc;
506 			if (i < g->textsize) {
507 				i /= HISTFRACTION * sizeof(*g->kcount);
508 				g->kcount[i]++;
509 			}
510 		}
511 #endif
512 #if defined(PROC_PC)
513 		if (p != NULL && p->p_flag & P_PROFIL)
514 			addupc_intr(p, PROC_PC(p));
515 #endif
516 		if (--spc->spc_pscnt > 0)
517 			return;
518 		/*
519 		 * Came from kernel mode, so we were:
520 		 * - handling an interrupt,
521 		 * - doing syscall or trap work on behalf of the current
522 		 *   user process, or
523 		 * - spinning in the idle loop.
524 		 * Whichever it is, charge the time as appropriate.
525 		 * Note that we charge interrupts to the current process,
526 		 * regardless of whether they are ``for'' that process,
527 		 * so that we know how much of its real time was spent
528 		 * in ``non-process'' (i.e., interrupt) work.
529 		 */
530 		if (CLKF_INTR(frame)) {
531 			if (p != NULL)
532 				p->p_iticks++;
533 			spc->spc_cp_time[CP_INTR]++;
534 		} else if (p != NULL && p != spc->spc_idleproc) {
535 			p->p_sticks++;
536 			spc->spc_cp_time[CP_SYS]++;
537 		} else
538 			spc->spc_cp_time[CP_IDLE]++;
539 	}
540 	spc->spc_pscnt = psdiv;
541 
542 	if (p != NULL) {
543 		p->p_cpticks++;
544 		/*
545 		 * If no schedclock is provided, call it here at ~~12-25 Hz;
546 		 * ~~16 Hz is best
547 		 */
548 		if (schedhz == 0) {
549 			if ((++curcpu()->ci_schedstate.spc_schedticks & 3) ==
550 			    0)
551 				schedclock(p);
552 		}
553 	}
554 }
555 
556 /*
557  * Return information about system clocks.
558  */
559 int
560 sysctl_clockrate(char *where, size_t *sizep, void *newp)
561 {
562 	struct clockinfo clkinfo;
563 
564 	/*
565 	 * Construct clockinfo structure.
566 	 */
567 	clkinfo.tick = tick;
568 	clkinfo.tickadj = tickadj;
569 	clkinfo.hz = hz;
570 	clkinfo.profhz = profhz;
571 	clkinfo.stathz = stathz ? stathz : hz;
572 	return (sysctl_rdstruct(where, sizep, newp, &clkinfo, sizeof(clkinfo)));
573 }
574 
575 #ifndef __HAVE_TIMECOUNTER
576 /*
577  * Placeholders until everyone uses the timecounters code.
578  * Won't improve anything except maybe removing a bunch of bugs in fixed code.
579  */
580 
581 void
582 getmicrotime(struct timeval *tvp)
583 {
584 	int s;
585 
586 	s = splhigh();
587 	*tvp = time;
588 	splx(s);
589 }
590 
591 void
592 nanotime(struct timespec *tsp)
593 {
594 	struct timeval tv;
595 
596 	microtime(&tv);
597 	TIMEVAL_TO_TIMESPEC(&tv, tsp);
598 }
599 
600 void
601 getnanotime(struct timespec *tsp)
602 {
603 	struct timeval tv;
604 
605 	getmicrotime(&tv);
606 	TIMEVAL_TO_TIMESPEC(&tv, tsp);
607 }
608 
609 void
610 nanouptime(struct timespec *tsp)
611 {
612 	struct timeval tv;
613 
614 	microuptime(&tv);
615 	TIMEVAL_TO_TIMESPEC(&tv, tsp);
616 }
617 
618 
619 void
620 getnanouptime(struct timespec *tsp)
621 {
622 	struct timeval tv;
623 
624 	getmicrouptime(&tv);
625 	TIMEVAL_TO_TIMESPEC(&tv, tsp);
626 }
627 
628 void
629 microuptime(struct timeval *tvp)
630 {
631 	struct timeval tv;
632 
633 	microtime(&tv);
634 	timersub(&tv, &boottime, tvp);
635 }
636 
637 void
638 getmicrouptime(struct timeval *tvp)
639 {
640 	int s;
641 
642 	s = splhigh();
643 	*tvp = mono_time;
644 	splx(s);
645 }
646 #endif /* __HAVE_TIMECOUNTER */
647