xref: /openbsd/sys/kern/kern_clock.c (revision 4cfece93)
1 /*	$OpenBSD: kern_clock.c,v 1.101 2020/01/21 16:16:23 mpi 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 #include "dt.h"
59 #if NDT > 0
60 #include <dev/dt/dtvar.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 int	stathz;
88 int	schedhz;
89 int	profhz;
90 int	profprocs;
91 int	ticks;
92 static int psdiv, pscnt;		/* prof => stat divider */
93 int	psratio;			/* ratio: prof / stat */
94 
95 volatile unsigned long jiffies;		/* XXX Linux API for drm(4) */
96 
97 /*
98  * Initialize clock frequencies and start both clocks running.
99  */
100 void
101 initclocks(void)
102 {
103 	int i;
104 
105 	ticks = INT_MAX - (15 * 60 * hz);
106 	jiffies = ULONG_MAX - (10 * 60 * hz);
107 
108 	/*
109 	 * Set divisors to 1 (normal case) and let the machine-specific
110 	 * code do its bit.
111 	 */
112 	psdiv = pscnt = 1;
113 	cpu_initclocks();
114 
115 	/*
116 	 * Compute profhz/stathz, and fix profhz if needed.
117 	 */
118 	i = stathz ? stathz : hz;
119 	if (profhz == 0)
120 		profhz = i;
121 	psratio = profhz / i;
122 
123 	/* For very large HZ, ensure that division by 0 does not occur later */
124 	if (tickadj == 0)
125 		tickadj = 1;
126 
127 	inittimecounter();
128 }
129 
130 /*
131  * hardclock does the accounting needed for ITIMER_PROF and ITIMER_VIRTUAL.
132  * We don't want to send signals with psignal from hardclock because it makes
133  * MULTIPROCESSOR locking very complicated. Instead, to use an idea from
134  * FreeBSD, we set a flag on the thread and when it goes to return to
135  * userspace it signals itself.
136  */
137 
138 /*
139  * The real-time timer, interrupting hz times per second.
140  */
141 void
142 hardclock(struct clockframe *frame)
143 {
144 	struct proc *p;
145 	struct cpu_info *ci = curcpu();
146 
147 	p = curproc;
148 	if (p && ((p->p_flag & (P_SYSTEM | P_WEXIT)) == 0)) {
149 		struct process *pr = p->p_p;
150 
151 		/*
152 		 * Run current process's virtual and profile time, as needed.
153 		 */
154 		if (CLKF_USERMODE(frame) &&
155 		    timespecisset(&pr->ps_timer[ITIMER_VIRTUAL].it_value) &&
156 		    itimerdecr(&pr->ps_timer[ITIMER_VIRTUAL], tick_nsec) == 0) {
157 			atomic_setbits_int(&p->p_flag, P_ALRMPEND);
158 			need_proftick(p);
159 		}
160 		if (timespecisset(&pr->ps_timer[ITIMER_PROF].it_value) &&
161 		    itimerdecr(&pr->ps_timer[ITIMER_PROF], tick_nsec) == 0) {
162 			atomic_setbits_int(&p->p_flag, P_PROFPEND);
163 			need_proftick(p);
164 		}
165 	}
166 
167 	/*
168 	 * If no separate statistics clock is available, run it from here.
169 	 */
170 	if (stathz == 0)
171 		statclock(frame);
172 
173 	if (--ci->ci_schedstate.spc_rrticks <= 0)
174 		roundrobin(ci);
175 
176 #if NDT > 0
177 	DT_ENTER(profile, NULL);
178 	if (CPU_IS_PRIMARY(ci))
179 		DT_ENTER(interval, NULL);
180 #endif
181 
182 	/*
183 	 * If we are not the primary CPU, we're not allowed to do
184 	 * any more work.
185 	 */
186 	if (CPU_IS_PRIMARY(ci) == 0)
187 		return;
188 
189 	tc_ticktock();
190 	ticks++;
191 	jiffies++;
192 
193 	/*
194 	 * Update the timeout wheel.
195 	 */
196 	timeout_hardclock_update();
197 }
198 
199 /*
200  * Compute number of hz in the specified amount of time.
201  */
202 int
203 tvtohz(const struct timeval *tv)
204 {
205 	unsigned long nticks;
206 	time_t sec;
207 	long usec;
208 
209 	/*
210 	 * If the number of usecs in the whole seconds part of the time
211 	 * fits in a long, then the total number of usecs will
212 	 * fit in an unsigned long.  Compute the total and convert it to
213 	 * ticks, rounding up and adding 1 to allow for the current tick
214 	 * to expire.  Rounding also depends on unsigned long arithmetic
215 	 * to avoid overflow.
216 	 *
217 	 * Otherwise, if the number of ticks in the whole seconds part of
218 	 * the time fits in a long, then convert the parts to
219 	 * ticks separately and add, using similar rounding methods and
220 	 * overflow avoidance.  This method would work in the previous
221 	 * case but it is slightly slower and assumes that hz is integral.
222 	 *
223 	 * Otherwise, round the time down to the maximum
224 	 * representable value.
225 	 *
226 	 * If ints have 32 bits, then the maximum value for any timeout in
227 	 * 10ms ticks is 248 days.
228 	 */
229 	sec = tv->tv_sec;
230 	usec = tv->tv_usec;
231 	if (sec < 0 || (sec == 0 && usec <= 0))
232 		nticks = 0;
233 	else if (sec <= LONG_MAX / 1000000)
234 		nticks = (sec * 1000000 + (unsigned long)usec + (tick - 1))
235 		    / tick + 1;
236 	else if (sec <= LONG_MAX / hz)
237 		nticks = sec * hz
238 		    + ((unsigned long)usec + (tick - 1)) / tick + 1;
239 	else
240 		nticks = LONG_MAX;
241 	if (nticks > INT_MAX)
242 		nticks = INT_MAX;
243 	return ((int)nticks);
244 }
245 
246 int
247 tstohz(const struct timespec *ts)
248 {
249 	struct timeval tv;
250 	TIMESPEC_TO_TIMEVAL(&tv, ts);
251 
252 	/* Round up. */
253 	if ((ts->tv_nsec % 1000) != 0) {
254 		tv.tv_usec += 1;
255 		if (tv.tv_usec >= 1000000) {
256 			tv.tv_usec -= 1000000;
257 			tv.tv_sec += 1;
258 		}
259 	}
260 
261 	return (tvtohz(&tv));
262 }
263 
264 /*
265  * Start profiling on a process.
266  *
267  * Kernel profiling passes proc0 which never exits and hence
268  * keeps the profile clock running constantly.
269  */
270 void
271 startprofclock(struct process *pr)
272 {
273 	int s;
274 
275 	if ((pr->ps_flags & PS_PROFIL) == 0) {
276 		atomic_setbits_int(&pr->ps_flags, PS_PROFIL);
277 		if (++profprocs == 1 && stathz != 0) {
278 			s = splstatclock();
279 			psdiv = pscnt = psratio;
280 			setstatclockrate(profhz);
281 			splx(s);
282 		}
283 	}
284 }
285 
286 /*
287  * Stop profiling on a process.
288  */
289 void
290 stopprofclock(struct process *pr)
291 {
292 	int s;
293 
294 	if (pr->ps_flags & PS_PROFIL) {
295 		atomic_clearbits_int(&pr->ps_flags, PS_PROFIL);
296 		if (--profprocs == 0 && stathz != 0) {
297 			s = splstatclock();
298 			psdiv = pscnt = 1;
299 			setstatclockrate(stathz);
300 			splx(s);
301 		}
302 	}
303 }
304 
305 /*
306  * Statistics clock.  Grab profile sample, and if divider reaches 0,
307  * do process and kernel statistics.
308  */
309 void
310 statclock(struct clockframe *frame)
311 {
312 #if defined(GPROF) || defined(DDBPROF)
313 	struct gmonparam *g;
314 	u_long i;
315 #endif
316 	struct cpu_info *ci = curcpu();
317 	struct schedstate_percpu *spc = &ci->ci_schedstate;
318 	struct proc *p = curproc;
319 	struct process *pr;
320 
321 	/*
322 	 * Notice changes in divisor frequency, and adjust clock
323 	 * frequency accordingly.
324 	 */
325 	if (spc->spc_psdiv != psdiv) {
326 		spc->spc_psdiv = psdiv;
327 		spc->spc_pscnt = psdiv;
328 		if (psdiv == 1) {
329 			setstatclockrate(stathz);
330 		} else {
331 			setstatclockrate(profhz);
332 		}
333 	}
334 
335 	if (CLKF_USERMODE(frame)) {
336 		pr = p->p_p;
337 		if (pr->ps_flags & PS_PROFIL)
338 			addupc_intr(p, CLKF_PC(frame));
339 		if (--spc->spc_pscnt > 0)
340 			return;
341 		/*
342 		 * Came from user mode; CPU was in user state.
343 		 * If this process is being profiled record the tick.
344 		 */
345 		p->p_uticks++;
346 		if (pr->ps_nice > NZERO)
347 			spc->spc_cp_time[CP_NICE]++;
348 		else
349 			spc->spc_cp_time[CP_USER]++;
350 	} else {
351 #if defined(GPROF) || defined(DDBPROF)
352 		/*
353 		 * Kernel statistics are just like addupc_intr, only easier.
354 		 */
355 		g = ci->ci_gmon;
356 		if (g != NULL && g->state == GMON_PROF_ON) {
357 			i = CLKF_PC(frame) - g->lowpc;
358 			if (i < g->textsize) {
359 				i /= HISTFRACTION * sizeof(*g->kcount);
360 				g->kcount[i]++;
361 			}
362 		}
363 #endif
364 #if defined(PROC_PC)
365 		if (p != NULL && p->p_p->ps_flags & PS_PROFIL)
366 			addupc_intr(p, PROC_PC(p));
367 #endif
368 		if (--spc->spc_pscnt > 0)
369 			return;
370 		/*
371 		 * Came from kernel mode, so we were:
372 		 * - spinning on a lock
373 		 * - handling an interrupt,
374 		 * - doing syscall or trap work on behalf of the current
375 		 *   user process, or
376 		 * - spinning in the idle loop.
377 		 * Whichever it is, charge the time as appropriate.
378 		 * Note that we charge interrupts to the current process,
379 		 * regardless of whether they are ``for'' that process,
380 		 * so that we know how much of its real time was spent
381 		 * in ``non-process'' (i.e., interrupt) work.
382 		 */
383 		if (CLKF_INTR(frame)) {
384 			if (p != NULL)
385 				p->p_iticks++;
386 			spc->spc_cp_time[spc->spc_spinning ?
387 			    CP_SPIN : CP_INTR]++;
388 		} else if (p != NULL && p != spc->spc_idleproc) {
389 			p->p_sticks++;
390 			spc->spc_cp_time[spc->spc_spinning ?
391 			    CP_SPIN : CP_SYS]++;
392 		} else
393 			spc->spc_cp_time[spc->spc_spinning ?
394 			    CP_SPIN : CP_IDLE]++;
395 	}
396 	spc->spc_pscnt = psdiv;
397 
398 	if (p != NULL) {
399 		p->p_cpticks++;
400 		/*
401 		 * If no schedclock is provided, call it here at ~~12-25 Hz;
402 		 * ~~16 Hz is best
403 		 */
404 		if (schedhz == 0) {
405 			if ((++spc->spc_schedticks & 3) == 0)
406 				schedclock(p);
407 		}
408 	}
409 }
410 
411 /*
412  * Return information about system clocks.
413  */
414 int
415 sysctl_clockrate(char *where, size_t *sizep, void *newp)
416 {
417 	struct clockinfo clkinfo;
418 
419 	/*
420 	 * Construct clockinfo structure.
421 	 */
422 	memset(&clkinfo, 0, sizeof clkinfo);
423 	clkinfo.tick = tick;
424 	clkinfo.tickadj = tickadj;
425 	clkinfo.hz = hz;
426 	clkinfo.profhz = profhz;
427 	clkinfo.stathz = stathz ? stathz : hz;
428 	return (sysctl_rdstruct(where, sizep, newp, &clkinfo, sizeof(clkinfo)));
429 }
430