xref: /openbsd-src/sys/kern/kern_clock.c (revision c90a81c56dcebd6a1b73fe4aff9b03385b8e63b3)
1 /*	$OpenBSD: kern_clock.c,v 1.97 2018/10/17 12:25:38 bluhm 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/timeout.h>
43 #include <sys/kernel.h>
44 #include <sys/limits.h>
45 #include <sys/proc.h>
46 #include <sys/user.h>
47 #include <sys/resourcevar.h>
48 #include <sys/signalvar.h>
49 #include <sys/sysctl.h>
50 #include <sys/sched.h>
51 #include <sys/timetc.h>
52 
53 
54 #if defined(GPROF) || defined(DDBPROF)
55 #include <sys/gmon.h>
56 #endif
57 
58 /*
59  * Clock handling routines.
60  *
61  * This code is written to operate with two timers that run independently of
62  * each other.  The main clock, running hz times per second, is used to keep
63  * track of real time.  The second timer handles kernel and user profiling,
64  * and does resource use estimation.  If the second timer is programmable,
65  * it is randomized to avoid aliasing between the two clocks.  For example,
66  * the randomization prevents an adversary from always giving up the cpu
67  * just before its quantum expires.  Otherwise, it would never accumulate
68  * cpu ticks.  The mean frequency of the second timer is stathz.
69  *
70  * If no second timer exists, stathz will be zero; in this case we drive
71  * profiling and statistics off the main clock.  This WILL NOT be accurate;
72  * do not do it unless absolutely necessary.
73  *
74  * The statistics clock may (or may not) be run at a higher rate while
75  * profiling.  This profile clock runs at profhz.  We require that profhz
76  * be an integral multiple of stathz.
77  *
78  * If the statistics clock is running fast, it must be divided by the ratio
79  * profhz/stathz for statistics.  (For profiling, every tick counts.)
80  */
81 
82 int	stathz;
83 int	schedhz;
84 int	profhz;
85 int	profprocs;
86 int	ticks;
87 static int psdiv, pscnt;		/* prof => stat divider */
88 int	psratio;			/* ratio: prof / stat */
89 
90 void	*softclock_si;
91 
92 volatile unsigned long jiffies;		/* XXX Linux API for drm(4) */
93 
94 /*
95  * Initialize clock frequencies and start both clocks running.
96  */
97 void
98 initclocks(void)
99 {
100 	int i;
101 
102 	softclock_si = softintr_establish(IPL_SOFTCLOCK, softclock, NULL);
103 	if (softclock_si == NULL)
104 		panic("initclocks: unable to register softclock intr");
105 
106 	ticks = INT_MAX - (15 * 60 * hz);
107 	jiffies = ULONG_MAX - (10 * 60 * hz);
108 
109 	/*
110 	 * Set divisors to 1 (normal case) and let the machine-specific
111 	 * code do its bit.
112 	 */
113 	psdiv = pscnt = 1;
114 	cpu_initclocks();
115 
116 	/*
117 	 * Compute profhz/stathz, and fix profhz if needed.
118 	 */
119 	i = stathz ? stathz : hz;
120 	if (profhz == 0)
121 		profhz = i;
122 	psratio = profhz / i;
123 
124 	/* For very large HZ, ensure that division by 0 does not occur later */
125 	if (tickadj == 0)
126 		tickadj = 1;
127 
128 	inittimecounter();
129 }
130 
131 /*
132  * hardclock does the accounting needed for ITIMER_PROF and ITIMER_VIRTUAL.
133  * We don't want to send signals with psignal from hardclock because it makes
134  * MULTIPROCESSOR locking very complicated. Instead, to use an idea from
135  * FreeBSD, we set a flag on the thread and when it goes to return to
136  * userspace it signals itself.
137  */
138 
139 /*
140  * The real-time timer, interrupting hz times per second.
141  */
142 void
143 hardclock(struct clockframe *frame)
144 {
145 	struct proc *p;
146 	struct cpu_info *ci = curcpu();
147 
148 	p = curproc;
149 	if (p && ((p->p_flag & (P_SYSTEM | P_WEXIT)) == 0)) {
150 		struct process *pr = p->p_p;
151 
152 		/*
153 		 * Run current process's virtual and profile time, as needed.
154 		 */
155 		if (CLKF_USERMODE(frame) &&
156 		    timerisset(&pr->ps_timer[ITIMER_VIRTUAL].it_value) &&
157 		    itimerdecr(&pr->ps_timer[ITIMER_VIRTUAL], tick) == 0) {
158 			atomic_setbits_int(&p->p_flag, P_ALRMPEND);
159 			need_proftick(p);
160 		}
161 		if (timerisset(&pr->ps_timer[ITIMER_PROF].it_value) &&
162 		    itimerdecr(&pr->ps_timer[ITIMER_PROF], tick) == 0) {
163 			atomic_setbits_int(&p->p_flag, P_PROFPEND);
164 			need_proftick(p);
165 		}
166 	}
167 
168 	/*
169 	 * If no separate statistics clock is available, run it from here.
170 	 */
171 	if (stathz == 0)
172 		statclock(frame);
173 
174 	if (--ci->ci_schedstate.spc_rrticks <= 0)
175 		roundrobin(ci);
176 
177 	/*
178 	 * If we are not the primary CPU, we're not allowed to do
179 	 * any more work.
180 	 */
181 	if (CPU_IS_PRIMARY(ci) == 0)
182 		return;
183 
184 	tc_ticktock();
185 	ticks++;
186 	jiffies++;
187 
188 	/*
189 	 * Update real-time timeout queue.
190 	 * Process callouts at a very low cpu priority, so we don't keep the
191 	 * relatively high clock interrupt priority any longer than necessary.
192 	 */
193 	if (timeout_hardclock_update())
194 		softintr_schedule(softclock_si);
195 }
196 
197 /*
198  * Compute number of hz in the specified amount of time.
199  */
200 int
201 tvtohz(const struct timeval *tv)
202 {
203 	unsigned long nticks;
204 	time_t sec;
205 	long usec;
206 
207 	/*
208 	 * If the number of usecs in the whole seconds part of the time
209 	 * fits in a long, then the total number of usecs will
210 	 * fit in an unsigned long.  Compute the total and convert it to
211 	 * ticks, rounding up and adding 1 to allow for the current tick
212 	 * to expire.  Rounding also depends on unsigned long arithmetic
213 	 * to avoid overflow.
214 	 *
215 	 * Otherwise, if the number of ticks in the whole seconds part of
216 	 * the time fits in a long, then convert the parts to
217 	 * ticks separately and add, using similar rounding methods and
218 	 * overflow avoidance.  This method would work in the previous
219 	 * case but it is slightly slower and assumes that hz is integral.
220 	 *
221 	 * Otherwise, round the time down to the maximum
222 	 * representable value.
223 	 *
224 	 * If ints have 32 bits, then the maximum value for any timeout in
225 	 * 10ms ticks is 248 days.
226 	 */
227 	sec = tv->tv_sec;
228 	usec = tv->tv_usec;
229 	if (sec < 0 || (sec == 0 && usec <= 0))
230 		nticks = 0;
231 	else if (sec <= LONG_MAX / 1000000)
232 		nticks = (sec * 1000000 + (unsigned long)usec + (tick - 1))
233 		    / tick + 1;
234 	else if (sec <= LONG_MAX / hz)
235 		nticks = sec * hz
236 		    + ((unsigned long)usec + (tick - 1)) / tick + 1;
237 	else
238 		nticks = LONG_MAX;
239 	if (nticks > INT_MAX)
240 		nticks = INT_MAX;
241 	return ((int)nticks);
242 }
243 
244 int
245 tstohz(const struct timespec *ts)
246 {
247 	struct timeval tv;
248 	TIMESPEC_TO_TIMEVAL(&tv, ts);
249 
250 	/* Round up. */
251 	if ((ts->tv_nsec % 1000) != 0) {
252 		tv.tv_usec += 1;
253 		if (tv.tv_usec >= 1000000) {
254 			tv.tv_usec -= 1000000;
255 			tv.tv_sec += 1;
256 		}
257 	}
258 
259 	return (tvtohz(&tv));
260 }
261 
262 /*
263  * Start profiling on a process.
264  *
265  * Kernel profiling passes proc0 which never exits and hence
266  * keeps the profile clock running constantly.
267  */
268 void
269 startprofclock(struct process *pr)
270 {
271 	int s;
272 
273 	if ((pr->ps_flags & PS_PROFIL) == 0) {
274 		atomic_setbits_int(&pr->ps_flags, PS_PROFIL);
275 		if (++profprocs == 1 && stathz != 0) {
276 			s = splstatclock();
277 			psdiv = pscnt = psratio;
278 			setstatclockrate(profhz);
279 			splx(s);
280 		}
281 	}
282 }
283 
284 /*
285  * Stop profiling on a process.
286  */
287 void
288 stopprofclock(struct process *pr)
289 {
290 	int s;
291 
292 	if (pr->ps_flags & PS_PROFIL) {
293 		atomic_clearbits_int(&pr->ps_flags, PS_PROFIL);
294 		if (--profprocs == 0 && stathz != 0) {
295 			s = splstatclock();
296 			psdiv = pscnt = 1;
297 			setstatclockrate(stathz);
298 			splx(s);
299 		}
300 	}
301 }
302 
303 /*
304  * Statistics clock.  Grab profile sample, and if divider reaches 0,
305  * do process and kernel statistics.
306  */
307 void
308 statclock(struct clockframe *frame)
309 {
310 #if defined(GPROF) || defined(DDBPROF)
311 	struct gmonparam *g;
312 	u_long i;
313 #endif
314 	struct cpu_info *ci = curcpu();
315 	struct schedstate_percpu *spc = &ci->ci_schedstate;
316 	struct proc *p = curproc;
317 	struct process *pr;
318 
319 	/*
320 	 * Notice changes in divisor frequency, and adjust clock
321 	 * frequency accordingly.
322 	 */
323 	if (spc->spc_psdiv != psdiv) {
324 		spc->spc_psdiv = psdiv;
325 		spc->spc_pscnt = psdiv;
326 		if (psdiv == 1) {
327 			setstatclockrate(stathz);
328 		} else {
329 			setstatclockrate(profhz);
330 		}
331 	}
332 
333 	if (CLKF_USERMODE(frame)) {
334 		pr = p->p_p;
335 		if (pr->ps_flags & PS_PROFIL)
336 			addupc_intr(p, CLKF_PC(frame));
337 		if (--spc->spc_pscnt > 0)
338 			return;
339 		/*
340 		 * Came from user mode; CPU was in user state.
341 		 * If this process is being profiled record the tick.
342 		 */
343 		p->p_uticks++;
344 		if (pr->ps_nice > NZERO)
345 			spc->spc_cp_time[CP_NICE]++;
346 		else
347 			spc->spc_cp_time[CP_USER]++;
348 	} else {
349 #if defined(GPROF) || defined(DDBPROF)
350 		/*
351 		 * Kernel statistics are just like addupc_intr, only easier.
352 		 */
353 		g = ci->ci_gmon;
354 		if (g != NULL && g->state == GMON_PROF_ON) {
355 			i = CLKF_PC(frame) - g->lowpc;
356 			if (i < g->textsize) {
357 				i /= HISTFRACTION * sizeof(*g->kcount);
358 				g->kcount[i]++;
359 			}
360 		}
361 #endif
362 #if defined(PROC_PC)
363 		if (p != NULL && p->p_p->ps_flags & PS_PROFIL)
364 			addupc_intr(p, PROC_PC(p));
365 #endif
366 		if (--spc->spc_pscnt > 0)
367 			return;
368 		/*
369 		 * Came from kernel mode, so we were:
370 		 * - spinning on a lock
371 		 * - handling an interrupt,
372 		 * - doing syscall or trap work on behalf of the current
373 		 *   user process, or
374 		 * - spinning in the idle loop.
375 		 * Whichever it is, charge the time as appropriate.
376 		 * Note that we charge interrupts to the current process,
377 		 * regardless of whether they are ``for'' that process,
378 		 * so that we know how much of its real time was spent
379 		 * in ``non-process'' (i.e., interrupt) work.
380 		 */
381 		if (CLKF_INTR(frame)) {
382 			if (p != NULL)
383 				p->p_iticks++;
384 			spc->spc_cp_time[spc->spc_spinning ?
385 			    CP_SPIN : CP_INTR]++;
386 		} else if (p != NULL && p != spc->spc_idleproc) {
387 			p->p_sticks++;
388 			spc->spc_cp_time[spc->spc_spinning ?
389 			    CP_SPIN : CP_SYS]++;
390 		} else
391 			spc->spc_cp_time[spc->spc_spinning ?
392 			    CP_SPIN : CP_IDLE]++;
393 	}
394 	spc->spc_pscnt = psdiv;
395 
396 	if (p != NULL) {
397 		p->p_cpticks++;
398 		/*
399 		 * If no schedclock is provided, call it here at ~~12-25 Hz;
400 		 * ~~16 Hz is best
401 		 */
402 		if (schedhz == 0) {
403 			if ((++curcpu()->ci_schedstate.spc_schedticks & 3) ==
404 			    0)
405 				schedclock(p);
406 		}
407 	}
408 }
409 
410 /*
411  * Return information about system clocks.
412  */
413 int
414 sysctl_clockrate(char *where, size_t *sizep, void *newp)
415 {
416 	struct clockinfo clkinfo;
417 
418 	/*
419 	 * Construct clockinfo structure.
420 	 */
421 	memset(&clkinfo, 0, sizeof clkinfo);
422 	clkinfo.tick = tick;
423 	clkinfo.tickadj = tickadj;
424 	clkinfo.hz = hz;
425 	clkinfo.profhz = profhz;
426 	clkinfo.stathz = stathz ? stathz : hz;
427 	return (sysctl_rdstruct(where, sizep, newp, &clkinfo, sizeof(clkinfo)));
428 }
429