xref: /openbsd-src/sys/kern/kern_clock.c (revision 94717a92adc4a8cae3f2627ad9d78d50fe2dd4e1)
1 /*	$OpenBSD: kern_clock.c,v 1.84 2013/12/24 01:11:00 tedu 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 #include <sys/timetc.h>
54 
55 
56 #ifdef GPROF
57 #include <sys/gmon.h>
58 #endif
59 
60 /*
61  * Clock handling routines.
62  *
63  * This code is written to operate with two timers that run independently of
64  * each other.  The main clock, running hz times per second, is used to keep
65  * track of real time.  The second timer handles kernel and user profiling,
66  * and does resource use estimation.  If the second timer is programmable,
67  * it is randomized to avoid aliasing between the two clocks.  For example,
68  * the randomization prevents an adversary from always giving up the cpu
69  * just before its quantum expires.  Otherwise, it would never accumulate
70  * cpu ticks.  The mean frequency of the second timer is stathz.
71  *
72  * If no second timer exists, stathz will be zero; in this case we drive
73  * profiling and statistics off the main clock.  This WILL NOT be accurate;
74  * do not do it unless absolutely necessary.
75  *
76  * The statistics clock may (or may not) be run at a higher rate while
77  * profiling.  This profile clock runs at profhz.  We require that profhz
78  * be an integral multiple of stathz.
79  *
80  * If the statistics clock is running fast, it must be divided by the ratio
81  * profhz/stathz for statistics.  (For profiling, every tick counts.)
82  */
83 
84 /*
85  * Bump a timeval by a small number of usec's.
86  */
87 #define BUMPTIME(t, usec) { \
88 	volatile struct timeval *tp = (t); \
89 	long us; \
90  \
91 	tp->tv_usec = us = tp->tv_usec + (usec); \
92 	if (us >= 1000000) { \
93 		tp->tv_usec = us - 1000000; \
94 		tp->tv_sec++; \
95 	} \
96 }
97 
98 int	stathz;
99 int	schedhz;
100 int	profhz;
101 int	profprocs;
102 int	ticks;
103 static int psdiv, pscnt;		/* prof => stat divider */
104 int	psratio;			/* ratio: prof / stat */
105 
106 long cp_time[CPUSTATES];
107 
108 void	*softclock_si;
109 
110 /*
111  * Initialize clock frequencies and start both clocks running.
112  */
113 void
114 initclocks(void)
115 {
116 	int i;
117 
118 	softclock_si = softintr_establish(IPL_SOFTCLOCK, softclock, NULL);
119 	if (softclock_si == NULL)
120 		panic("initclocks: unable to register softclock intr");
121 
122 	/*
123 	 * Set divisors to 1 (normal case) and let the machine-specific
124 	 * code do its bit.
125 	 */
126 	psdiv = pscnt = 1;
127 	cpu_initclocks();
128 
129 	/*
130 	 * Compute profhz/stathz, and fix profhz if needed.
131 	 */
132 	i = stathz ? stathz : hz;
133 	if (profhz == 0)
134 		profhz = i;
135 	psratio = profhz / i;
136 
137 	/* For very large HZ, ensure that division by 0 does not occur later */
138 	if (tickadj == 0)
139 		tickadj = 1;
140 
141 	inittimecounter();
142 }
143 
144 /*
145  * hardclock does the accounting needed for ITIMER_PROF and ITIMER_VIRTUAL.
146  * We don't want to send signals with psignal from hardclock because it makes
147  * MULTIPROCESSOR locking very complicated. Instead, to use an idea from
148  * FreeBSD, we set a flag on the thread and when it goes to return to
149  * userspace it signals itself.
150  */
151 
152 /*
153  * The real-time timer, interrupting hz times per second.
154  */
155 void
156 hardclock(struct clockframe *frame)
157 {
158 	struct proc *p;
159 	struct cpu_info *ci = curcpu();
160 
161 	p = curproc;
162 	if (p && ((p->p_flag & (P_SYSTEM | P_WEXIT)) == 0)) {
163 		struct process *pr = p->p_p;
164 
165 		/*
166 		 * Run current process's virtual and profile time, as needed.
167 		 */
168 		if (CLKF_USERMODE(frame) &&
169 		    timerisset(&pr->ps_timer[ITIMER_VIRTUAL].it_value) &&
170 		    itimerdecr(&pr->ps_timer[ITIMER_VIRTUAL], tick) == 0) {
171 			atomic_setbits_int(&p->p_flag, P_ALRMPEND);
172 			need_proftick(p);
173 		}
174 		if (timerisset(&pr->ps_timer[ITIMER_PROF].it_value) &&
175 		    itimerdecr(&pr->ps_timer[ITIMER_PROF], tick) == 0) {
176 			atomic_setbits_int(&p->p_flag, P_PROFPEND);
177 			need_proftick(p);
178 		}
179 	}
180 
181 	/*
182 	 * If no separate statistics clock is available, run it from here.
183 	 */
184 	if (stathz == 0)
185 		statclock(frame);
186 
187 	if (--ci->ci_schedstate.spc_rrticks <= 0)
188 		roundrobin(ci);
189 
190 	/*
191 	 * If we are not the primary CPU, we're not allowed to do
192 	 * any more work.
193 	 */
194 	if (CPU_IS_PRIMARY(ci) == 0)
195 		return;
196 
197 	tc_ticktock();
198 
199 	/*
200 	 * Update real-time timeout queue.
201 	 * Process callouts at a very low cpu priority, so we don't keep the
202 	 * relatively high clock interrupt priority any longer than necessary.
203 	 */
204 	if (timeout_hardclock_update())
205 		softintr_schedule(softclock_si);
206 }
207 
208 /*
209  * Compute number of hz until specified time.  Used to
210  * compute the second argument to timeout_add() from an absolute time.
211  */
212 int
213 hzto(const struct timeval *tv)
214 {
215 	struct timeval now;
216 	unsigned long nticks;
217 	long sec, usec;
218 
219 	/*
220 	 * If the number of usecs in the whole seconds part of the time
221 	 * difference fits in a long, then the total number of usecs will
222 	 * fit in an unsigned long.  Compute the total and convert it to
223 	 * ticks, rounding up and adding 1 to allow for the current tick
224 	 * to expire.  Rounding also depends on unsigned long arithmetic
225 	 * to avoid overflow.
226 	 *
227 	 * Otherwise, if the number of ticks in the whole seconds part of
228 	 * the time difference fits in a long, then convert the parts to
229 	 * ticks separately and add, using similar rounding methods and
230 	 * overflow avoidance.  This method would work in the previous
231 	 * case but it is slightly slower and assumes that hz is integral.
232 	 *
233 	 * Otherwise, round the time difference down to the maximum
234 	 * representable value.
235 	 *
236 	 * If ints have 32 bits, then the maximum value for any timeout in
237 	 * 10ms ticks is 248 days.
238 	 */
239 	getmicrotime(&now);
240 	sec = tv->tv_sec - now.tv_sec;
241 	usec = tv->tv_usec - now.tv_usec;
242 	if (usec < 0) {
243 		sec--;
244 		usec += 1000000;
245 	}
246 	if (sec < 0 || (sec == 0 && usec <= 0)) {
247 		nticks = 0;
248 	} else if (sec <= LONG_MAX / 1000000)
249 		nticks = (sec * 1000000 + (unsigned long)usec + (tick - 1))
250 		    / tick + 1;
251 	else if (sec <= LONG_MAX / hz)
252 		nticks = sec * hz
253 		    + ((unsigned long)usec + (tick - 1)) / tick + 1;
254 	else
255 		nticks = LONG_MAX;
256 	if (nticks > INT_MAX)
257 		nticks = INT_MAX;
258 	return ((int)nticks);
259 }
260 
261 /*
262  * Compute number of hz in the specified amount of time.
263  */
264 int
265 tvtohz(const struct timeval *tv)
266 {
267 	unsigned long nticks;
268 	time_t sec;
269 	long usec;
270 
271 	/*
272 	 * If the number of usecs in the whole seconds part of the time
273 	 * fits in a long, then the total number of usecs will
274 	 * fit in an unsigned long.  Compute the total and convert it to
275 	 * ticks, rounding up and adding 1 to allow for the current tick
276 	 * to expire.  Rounding also depends on unsigned long arithmetic
277 	 * to avoid overflow.
278 	 *
279 	 * Otherwise, if the number of ticks in the whole seconds part of
280 	 * the time fits in a long, then convert the parts to
281 	 * ticks separately and add, using similar rounding methods and
282 	 * overflow avoidance.  This method would work in the previous
283 	 * case but it is slightly slower and assumes that hz is integral.
284 	 *
285 	 * Otherwise, round the time down to the maximum
286 	 * representable value.
287 	 *
288 	 * If ints have 32 bits, then the maximum value for any timeout in
289 	 * 10ms ticks is 248 days.
290 	 */
291 	sec = tv->tv_sec;
292 	usec = tv->tv_usec;
293 	if (sec < 0 || (sec == 0 && usec <= 0))
294 		nticks = 0;
295 	else if (sec <= LONG_MAX / 1000000)
296 		nticks = (sec * 1000000 + (unsigned long)usec + (tick - 1))
297 		    / tick + 1;
298 	else if (sec <= LONG_MAX / hz)
299 		nticks = sec * hz
300 		    + ((unsigned long)usec + (tick - 1)) / tick + 1;
301 	else
302 		nticks = LONG_MAX;
303 	if (nticks > INT_MAX)
304 		nticks = INT_MAX;
305 	return ((int)nticks);
306 }
307 
308 int
309 tstohz(const struct timespec *ts)
310 {
311 	struct timeval tv;
312 	TIMESPEC_TO_TIMEVAL(&tv, ts);
313 
314 	/* Round up. */
315 	if ((ts->tv_nsec % 1000) != 0) {
316 		tv.tv_usec += 1;
317 		if (tv.tv_usec >= 1000000) {
318 			tv.tv_usec -= 1000000;
319 			tv.tv_sec += 1;
320 		}
321 	}
322 
323 	return (tvtohz(&tv));
324 }
325 
326 /*
327  * Start profiling on a process.
328  *
329  * Kernel profiling passes proc0 which never exits and hence
330  * keeps the profile clock running constantly.
331  */
332 void
333 startprofclock(struct process *pr)
334 {
335 	int s;
336 
337 	if ((pr->ps_flags & PS_PROFIL) == 0) {
338 		atomic_setbits_int(&pr->ps_flags, PS_PROFIL);
339 		if (++profprocs == 1 && stathz != 0) {
340 			s = splstatclock();
341 			psdiv = pscnt = psratio;
342 			setstatclockrate(profhz);
343 			splx(s);
344 		}
345 	}
346 }
347 
348 /*
349  * Stop profiling on a process.
350  */
351 void
352 stopprofclock(struct process *pr)
353 {
354 	int s;
355 
356 	if (pr->ps_flags & PS_PROFIL) {
357 		atomic_clearbits_int(&pr->ps_flags, PS_PROFIL);
358 		if (--profprocs == 0 && stathz != 0) {
359 			s = splstatclock();
360 			psdiv = pscnt = 1;
361 			setstatclockrate(stathz);
362 			splx(s);
363 		}
364 	}
365 }
366 
367 /*
368  * Statistics clock.  Grab profile sample, and if divider reaches 0,
369  * do process and kernel statistics.
370  */
371 void
372 statclock(struct clockframe *frame)
373 {
374 #ifdef GPROF
375 	struct gmonparam *g;
376 	u_long i;
377 #endif
378 	struct cpu_info *ci = curcpu();
379 	struct schedstate_percpu *spc = &ci->ci_schedstate;
380 	struct proc *p = curproc;
381 	struct process *pr;
382 
383 	/*
384 	 * Notice changes in divisor frequency, and adjust clock
385 	 * frequency accordingly.
386 	 */
387 	if (spc->spc_psdiv != psdiv) {
388 		spc->spc_psdiv = psdiv;
389 		spc->spc_pscnt = psdiv;
390 		if (psdiv == 1) {
391 			setstatclockrate(stathz);
392 		} else {
393 			setstatclockrate(profhz);
394 		}
395 	}
396 
397 	if (CLKF_USERMODE(frame)) {
398 		pr = p->p_p;
399 		if (pr->ps_flags & PS_PROFIL)
400 			addupc_intr(p, CLKF_PC(frame));
401 		if (--spc->spc_pscnt > 0)
402 			return;
403 		/*
404 		 * Came from user mode; CPU was in user state.
405 		 * If this process is being profiled record the tick.
406 		 */
407 		p->p_uticks++;
408 		if (pr->ps_nice > NZERO)
409 			spc->spc_cp_time[CP_NICE]++;
410 		else
411 			spc->spc_cp_time[CP_USER]++;
412 	} else {
413 #ifdef GPROF
414 		/*
415 		 * Kernel statistics are just like addupc_intr, only easier.
416 		 */
417 		g = ci->ci_gmon;
418 		if (g != NULL && g->state == GMON_PROF_ON) {
419 			i = CLKF_PC(frame) - g->lowpc;
420 			if (i < g->textsize) {
421 				i /= HISTFRACTION * sizeof(*g->kcount);
422 				g->kcount[i]++;
423 			}
424 		}
425 #endif
426 #if defined(PROC_PC)
427 		if (p != NULL && p->p_p->ps_flags & PS_PROFIL)
428 			addupc_intr(p, PROC_PC(p));
429 #endif
430 		if (--spc->spc_pscnt > 0)
431 			return;
432 		/*
433 		 * Came from kernel mode, so we were:
434 		 * - handling an interrupt,
435 		 * - doing syscall or trap work on behalf of the current
436 		 *   user process, or
437 		 * - spinning in the idle loop.
438 		 * Whichever it is, charge the time as appropriate.
439 		 * Note that we charge interrupts to the current process,
440 		 * regardless of whether they are ``for'' that process,
441 		 * so that we know how much of its real time was spent
442 		 * in ``non-process'' (i.e., interrupt) work.
443 		 */
444 		if (CLKF_INTR(frame)) {
445 			if (p != NULL)
446 				p->p_iticks++;
447 			spc->spc_cp_time[CP_INTR]++;
448 		} else if (p != NULL && p != spc->spc_idleproc) {
449 			p->p_sticks++;
450 			spc->spc_cp_time[CP_SYS]++;
451 		} else
452 			spc->spc_cp_time[CP_IDLE]++;
453 	}
454 	spc->spc_pscnt = psdiv;
455 
456 	if (p != NULL) {
457 		p->p_cpticks++;
458 		/*
459 		 * If no schedclock is provided, call it here at ~~12-25 Hz;
460 		 * ~~16 Hz is best
461 		 */
462 		if (schedhz == 0) {
463 			if ((++curcpu()->ci_schedstate.spc_schedticks & 3) ==
464 			    0)
465 				schedclock(p);
466 		}
467 	}
468 }
469 
470 /*
471  * Return information about system clocks.
472  */
473 int
474 sysctl_clockrate(char *where, size_t *sizep, void *newp)
475 {
476 	struct clockinfo clkinfo;
477 
478 	/*
479 	 * Construct clockinfo structure.
480 	 */
481 	clkinfo.tick = tick;
482 	clkinfo.tickadj = tickadj;
483 	clkinfo.hz = hz;
484 	clkinfo.profhz = profhz;
485 	clkinfo.stathz = stathz ? stathz : hz;
486 	return (sysctl_rdstruct(where, sizep, newp, &clkinfo, sizeof(clkinfo)));
487 }
488