xref: /openbsd-src/sys/kern/kern_clock.c (revision 3374c67d44f9b75b98444cbf63020f777792342e)
1 /*	$OpenBSD: kern_clock.c,v 1.105 2022/08/14 01:58:27 jsg 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/sysctl.h>
49 #include <sys/sched.h>
50 #include <sys/timetc.h>
51 
52 #if defined(GPROF) || defined(DDBPROF)
53 #include <sys/gmon.h>
54 #endif
55 
56 #include "dt.h"
57 #if NDT > 0
58 #include <dev/dt/dtvar.h>
59 #endif
60 
61 /*
62  * Clock handling routines.
63  *
64  * This code is written to operate with two timers that run independently of
65  * each other.  The main clock, running hz times per second, is used to keep
66  * track of real time.  The second timer handles kernel and user profiling,
67  * and does resource use estimation.  If the second timer is programmable,
68  * it is randomized to avoid aliasing between the two clocks.  For example,
69  * the randomization prevents an adversary from always giving up the cpu
70  * just before its quantum expires.  Otherwise, it would never accumulate
71  * cpu ticks.  The mean frequency of the second timer is stathz.
72  *
73  * If no second timer exists, stathz will be zero; in this case we drive
74  * profiling and statistics off the main clock.  This WILL NOT be accurate;
75  * do not do it unless absolutely necessary.
76  *
77  * The statistics clock may (or may not) be run at a higher rate while
78  * profiling.  This profile clock runs at profhz.  We require that profhz
79  * be an integral multiple of stathz.
80  *
81  * If the statistics clock is running fast, it must be divided by the ratio
82  * profhz/stathz for statistics.  (For profiling, every tick counts.)
83  */
84 
85 int	stathz;
86 int	schedhz;
87 int	profhz;
88 int	profprocs;
89 int	ticks;
90 static int psdiv, pscnt;		/* prof => stat divider */
91 int	psratio;			/* ratio: prof / stat */
92 
93 volatile unsigned long jiffies;		/* XXX Linux API for drm(4) */
94 
95 /*
96  * Initialize clock frequencies and start both clocks running.
97  */
98 void
99 initclocks(void)
100 {
101 	int i;
102 
103 	ticks = INT_MAX - (15 * 60 * hz);
104 	jiffies = ULONG_MAX - (10 * 60 * hz);
105 
106 	/*
107 	 * Set divisors to 1 (normal case) and let the machine-specific
108 	 * code do its bit.
109 	 */
110 	psdiv = pscnt = 1;
111 	cpu_initclocks();
112 
113 	/*
114 	 * Compute profhz/stathz, and fix profhz if needed.
115 	 */
116 	i = stathz ? stathz : hz;
117 	if (profhz == 0)
118 		profhz = i;
119 	psratio = profhz / i;
120 
121 	inittimecounter();
122 }
123 
124 /*
125  * hardclock does the accounting needed for ITIMER_PROF and ITIMER_VIRTUAL.
126  * We don't want to send signals with psignal from hardclock because it makes
127  * MULTIPROCESSOR locking very complicated. Instead, to use an idea from
128  * FreeBSD, we set a flag on the thread and when it goes to return to
129  * userspace it signals itself.
130  */
131 
132 /*
133  * The real-time timer, interrupting hz times per second.
134  */
135 void
136 hardclock(struct clockframe *frame)
137 {
138 	struct proc *p;
139 	struct cpu_info *ci = curcpu();
140 
141 	p = curproc;
142 	if (p && ((p->p_flag & (P_SYSTEM | P_WEXIT)) == 0)) {
143 		struct process *pr = p->p_p;
144 
145 		/*
146 		 * Run current process's virtual and profile time, as needed.
147 		 */
148 		if (CLKF_USERMODE(frame) &&
149 		    timespecisset(&pr->ps_timer[ITIMER_VIRTUAL].it_value) &&
150 		    itimerdecr(&pr->ps_timer[ITIMER_VIRTUAL], tick_nsec) == 0) {
151 			atomic_setbits_int(&p->p_flag, P_ALRMPEND);
152 			need_proftick(p);
153 		}
154 		if (timespecisset(&pr->ps_timer[ITIMER_PROF].it_value) &&
155 		    itimerdecr(&pr->ps_timer[ITIMER_PROF], tick_nsec) == 0) {
156 			atomic_setbits_int(&p->p_flag, P_PROFPEND);
157 			need_proftick(p);
158 		}
159 	}
160 
161 	/*
162 	 * If no separate statistics clock is available, run it from here.
163 	 */
164 	if (stathz == 0)
165 		statclock(frame);
166 
167 	if (--ci->ci_schedstate.spc_rrticks <= 0)
168 		roundrobin(ci);
169 
170 #if NDT > 0
171 	DT_ENTER(profile, NULL);
172 	if (CPU_IS_PRIMARY(ci))
173 		DT_ENTER(interval, NULL);
174 #endif
175 
176 	/*
177 	 * If we are not the primary CPU, we're not allowed to do
178 	 * any more work.
179 	 */
180 	if (CPU_IS_PRIMARY(ci) == 0)
181 		return;
182 
183 	tc_ticktock();
184 	ticks++;
185 	jiffies++;
186 
187 	/*
188 	 * Update the timeout wheel.
189 	 */
190 	timeout_hardclock_update();
191 }
192 
193 /*
194  * Compute number of hz in the specified amount of time.
195  */
196 int
197 tvtohz(const struct timeval *tv)
198 {
199 	unsigned long nticks;
200 	time_t sec;
201 	long usec;
202 
203 	/*
204 	 * If the number of usecs in the whole seconds part of the time
205 	 * fits in a long, then the total number of usecs will
206 	 * fit in an unsigned long.  Compute the total and convert it to
207 	 * ticks, rounding up and adding 1 to allow for the current tick
208 	 * to expire.  Rounding also depends on unsigned long arithmetic
209 	 * to avoid overflow.
210 	 *
211 	 * Otherwise, if the number of ticks in the whole seconds part of
212 	 * the time fits in a long, then convert the parts to
213 	 * ticks separately and add, using similar rounding methods and
214 	 * overflow avoidance.  This method would work in the previous
215 	 * case but it is slightly slower and assumes that hz is integral.
216 	 *
217 	 * Otherwise, round the time down to the maximum
218 	 * representable value.
219 	 *
220 	 * If ints have 32 bits, then the maximum value for any timeout in
221 	 * 10ms ticks is 248 days.
222 	 */
223 	sec = tv->tv_sec;
224 	usec = tv->tv_usec;
225 	if (sec < 0 || (sec == 0 && usec <= 0))
226 		nticks = 0;
227 	else if (sec <= LONG_MAX / 1000000)
228 		nticks = (sec * 1000000 + (unsigned long)usec + (tick - 1))
229 		    / tick + 1;
230 	else if (sec <= LONG_MAX / hz)
231 		nticks = sec * hz
232 		    + ((unsigned long)usec + (tick - 1)) / tick + 1;
233 	else
234 		nticks = LONG_MAX;
235 	if (nticks > INT_MAX)
236 		nticks = INT_MAX;
237 	return ((int)nticks);
238 }
239 
240 int
241 tstohz(const struct timespec *ts)
242 {
243 	struct timeval tv;
244 	TIMESPEC_TO_TIMEVAL(&tv, ts);
245 
246 	/* Round up. */
247 	if ((ts->tv_nsec % 1000) != 0) {
248 		tv.tv_usec += 1;
249 		if (tv.tv_usec >= 1000000) {
250 			tv.tv_usec -= 1000000;
251 			tv.tv_sec += 1;
252 		}
253 	}
254 
255 	return (tvtohz(&tv));
256 }
257 
258 /*
259  * Start profiling on a process.
260  *
261  * Kernel profiling passes proc0 which never exits and hence
262  * keeps the profile clock running constantly.
263  */
264 void
265 startprofclock(struct process *pr)
266 {
267 	int s;
268 
269 	if ((pr->ps_flags & PS_PROFIL) == 0) {
270 		atomic_setbits_int(&pr->ps_flags, PS_PROFIL);
271 		if (++profprocs == 1 && stathz != 0) {
272 			s = splstatclock();
273 			psdiv = pscnt = psratio;
274 			setstatclockrate(profhz);
275 			splx(s);
276 		}
277 	}
278 }
279 
280 /*
281  * Stop profiling on a process.
282  */
283 void
284 stopprofclock(struct process *pr)
285 {
286 	int s;
287 
288 	if (pr->ps_flags & PS_PROFIL) {
289 		atomic_clearbits_int(&pr->ps_flags, PS_PROFIL);
290 		if (--profprocs == 0 && stathz != 0) {
291 			s = splstatclock();
292 			psdiv = pscnt = 1;
293 			setstatclockrate(stathz);
294 			splx(s);
295 		}
296 	}
297 }
298 
299 /*
300  * Statistics clock.  Grab profile sample, and if divider reaches 0,
301  * do process and kernel statistics.
302  */
303 void
304 statclock(struct clockframe *frame)
305 {
306 #if defined(GPROF) || defined(DDBPROF)
307 	struct gmonparam *g;
308 	u_long i;
309 #endif
310 	struct cpu_info *ci = curcpu();
311 	struct schedstate_percpu *spc = &ci->ci_schedstate;
312 	struct proc *p = curproc;
313 	struct process *pr;
314 
315 	/*
316 	 * Notice changes in divisor frequency, and adjust clock
317 	 * frequency accordingly.
318 	 */
319 	if (spc->spc_psdiv != psdiv) {
320 		spc->spc_psdiv = psdiv;
321 		spc->spc_pscnt = psdiv;
322 		if (psdiv == 1) {
323 			setstatclockrate(stathz);
324 		} else {
325 			setstatclockrate(profhz);
326 		}
327 	}
328 
329 	if (CLKF_USERMODE(frame)) {
330 		pr = p->p_p;
331 		if (pr->ps_flags & PS_PROFIL)
332 			addupc_intr(p, CLKF_PC(frame));
333 		if (--spc->spc_pscnt > 0)
334 			return;
335 		/*
336 		 * Came from user mode; CPU was in user state.
337 		 * If this process is being profiled record the tick.
338 		 */
339 		p->p_uticks++;
340 		if (pr->ps_nice > NZERO)
341 			spc->spc_cp_time[CP_NICE]++;
342 		else
343 			spc->spc_cp_time[CP_USER]++;
344 	} else {
345 #if defined(GPROF) || defined(DDBPROF)
346 		/*
347 		 * Kernel statistics are just like addupc_intr, only easier.
348 		 */
349 		g = ci->ci_gmon;
350 		if (g != NULL && g->state == GMON_PROF_ON) {
351 			i = CLKF_PC(frame) - g->lowpc;
352 			if (i < g->textsize) {
353 				i /= HISTFRACTION * sizeof(*g->kcount);
354 				g->kcount[i]++;
355 			}
356 		}
357 #endif
358 		if (p != NULL && p->p_p->ps_flags & PS_PROFIL)
359 			addupc_intr(p, PROC_PC(p));
360 		if (--spc->spc_pscnt > 0)
361 			return;
362 		/*
363 		 * Came from kernel mode, so we were:
364 		 * - spinning on a lock
365 		 * - handling an interrupt,
366 		 * - doing syscall or trap work on behalf of the current
367 		 *   user process, or
368 		 * - spinning in the idle loop.
369 		 * Whichever it is, charge the time as appropriate.
370 		 * Note that we charge interrupts to the current process,
371 		 * regardless of whether they are ``for'' that process,
372 		 * so that we know how much of its real time was spent
373 		 * in ``non-process'' (i.e., interrupt) work.
374 		 */
375 		if (CLKF_INTR(frame)) {
376 			if (p != NULL)
377 				p->p_iticks++;
378 			spc->spc_cp_time[spc->spc_spinning ?
379 			    CP_SPIN : CP_INTR]++;
380 		} else if (p != NULL && p != spc->spc_idleproc) {
381 			p->p_sticks++;
382 			spc->spc_cp_time[spc->spc_spinning ?
383 			    CP_SPIN : CP_SYS]++;
384 		} else
385 			spc->spc_cp_time[spc->spc_spinning ?
386 			    CP_SPIN : CP_IDLE]++;
387 	}
388 	spc->spc_pscnt = psdiv;
389 
390 	if (p != NULL) {
391 		p->p_cpticks++;
392 		/*
393 		 * If no schedclock is provided, call it here at ~~12-25 Hz;
394 		 * ~~16 Hz is best
395 		 */
396 		if (schedhz == 0) {
397 			if ((++spc->spc_schedticks & 3) == 0)
398 				schedclock(p);
399 		}
400 	}
401 }
402 
403 /*
404  * Return information about system clocks.
405  */
406 int
407 sysctl_clockrate(char *where, size_t *sizep, void *newp)
408 {
409 	struct clockinfo clkinfo;
410 
411 	/*
412 	 * Construct clockinfo structure.
413 	 */
414 	memset(&clkinfo, 0, sizeof clkinfo);
415 	clkinfo.tick = tick;
416 	clkinfo.hz = hz;
417 	clkinfo.profhz = profhz;
418 	clkinfo.stathz = stathz ? stathz : hz;
419 	return (sysctl_rdstruct(where, sizep, newp, &clkinfo, sizeof(clkinfo)));
420 }
421