xref: /openbsd-src/sys/kern/kern_clock.c (revision 9bcfcad5a8f89fd60010f3485ee44bf251abde23)
1 /*	$OpenBSD: kern_clock.c,v 1.106 2023/02/04 19:33:03 cheloha 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 	ticks = INT_MAX - (15 * 60 * hz);
102 	jiffies = ULONG_MAX - (10 * 60 * hz);
103 
104 	/*
105 	 * Set divisors to 1 (normal case) and let the machine-specific
106 	 * code do its bit.
107 	 */
108 	psdiv = pscnt = 1;
109 	cpu_initclocks();
110 
111 	/*
112 	 * Compute profhz/stathz.
113 	 */
114 	psratio = profhz / stathz;
115 
116 	inittimecounter();
117 }
118 
119 /*
120  * hardclock does the accounting needed for ITIMER_PROF and ITIMER_VIRTUAL.
121  * We don't want to send signals with psignal from hardclock because it makes
122  * MULTIPROCESSOR locking very complicated. Instead, to use an idea from
123  * FreeBSD, we set a flag on the thread and when it goes to return to
124  * userspace it signals itself.
125  */
126 
127 /*
128  * The real-time timer, interrupting hz times per second.
129  */
130 void
131 hardclock(struct clockframe *frame)
132 {
133 	struct proc *p;
134 	struct cpu_info *ci = curcpu();
135 
136 	p = curproc;
137 	if (p && ((p->p_flag & (P_SYSTEM | P_WEXIT)) == 0)) {
138 		struct process *pr = p->p_p;
139 
140 		/*
141 		 * Run current process's virtual and profile time, as needed.
142 		 */
143 		if (CLKF_USERMODE(frame) &&
144 		    timespecisset(&pr->ps_timer[ITIMER_VIRTUAL].it_value) &&
145 		    itimerdecr(&pr->ps_timer[ITIMER_VIRTUAL], tick_nsec) == 0) {
146 			atomic_setbits_int(&p->p_flag, P_ALRMPEND);
147 			need_proftick(p);
148 		}
149 		if (timespecisset(&pr->ps_timer[ITIMER_PROF].it_value) &&
150 		    itimerdecr(&pr->ps_timer[ITIMER_PROF], tick_nsec) == 0) {
151 			atomic_setbits_int(&p->p_flag, P_PROFPEND);
152 			need_proftick(p);
153 		}
154 	}
155 
156 	if (--ci->ci_schedstate.spc_rrticks <= 0)
157 		roundrobin(ci);
158 
159 #if NDT > 0
160 	DT_ENTER(profile, NULL);
161 	if (CPU_IS_PRIMARY(ci))
162 		DT_ENTER(interval, NULL);
163 #endif
164 
165 	/*
166 	 * If we are not the primary CPU, we're not allowed to do
167 	 * any more work.
168 	 */
169 	if (CPU_IS_PRIMARY(ci) == 0)
170 		return;
171 
172 	tc_ticktock();
173 	ticks++;
174 	jiffies++;
175 
176 	/*
177 	 * Update the timeout wheel.
178 	 */
179 	timeout_hardclock_update();
180 }
181 
182 /*
183  * Compute number of hz in the specified amount of time.
184  */
185 int
186 tvtohz(const struct timeval *tv)
187 {
188 	unsigned long nticks;
189 	time_t sec;
190 	long usec;
191 
192 	/*
193 	 * If the number of usecs in the whole seconds part of the time
194 	 * fits in a long, then the total number of usecs will
195 	 * fit in an unsigned long.  Compute the total and convert it to
196 	 * ticks, rounding up and adding 1 to allow for the current tick
197 	 * to expire.  Rounding also depends on unsigned long arithmetic
198 	 * to avoid overflow.
199 	 *
200 	 * Otherwise, if the number of ticks in the whole seconds part of
201 	 * the time fits in a long, then convert the parts to
202 	 * ticks separately and add, using similar rounding methods and
203 	 * overflow avoidance.  This method would work in the previous
204 	 * case but it is slightly slower and assumes that hz is integral.
205 	 *
206 	 * Otherwise, round the time down to the maximum
207 	 * representable value.
208 	 *
209 	 * If ints have 32 bits, then the maximum value for any timeout in
210 	 * 10ms ticks is 248 days.
211 	 */
212 	sec = tv->tv_sec;
213 	usec = tv->tv_usec;
214 	if (sec < 0 || (sec == 0 && usec <= 0))
215 		nticks = 0;
216 	else if (sec <= LONG_MAX / 1000000)
217 		nticks = (sec * 1000000 + (unsigned long)usec + (tick - 1))
218 		    / tick + 1;
219 	else if (sec <= LONG_MAX / hz)
220 		nticks = sec * hz
221 		    + ((unsigned long)usec + (tick - 1)) / tick + 1;
222 	else
223 		nticks = LONG_MAX;
224 	if (nticks > INT_MAX)
225 		nticks = INT_MAX;
226 	return ((int)nticks);
227 }
228 
229 int
230 tstohz(const struct timespec *ts)
231 {
232 	struct timeval tv;
233 	TIMESPEC_TO_TIMEVAL(&tv, ts);
234 
235 	/* Round up. */
236 	if ((ts->tv_nsec % 1000) != 0) {
237 		tv.tv_usec += 1;
238 		if (tv.tv_usec >= 1000000) {
239 			tv.tv_usec -= 1000000;
240 			tv.tv_sec += 1;
241 		}
242 	}
243 
244 	return (tvtohz(&tv));
245 }
246 
247 /*
248  * Start profiling on a process.
249  *
250  * Kernel profiling passes proc0 which never exits and hence
251  * keeps the profile clock running constantly.
252  */
253 void
254 startprofclock(struct process *pr)
255 {
256 	int s;
257 
258 	if ((pr->ps_flags & PS_PROFIL) == 0) {
259 		atomic_setbits_int(&pr->ps_flags, PS_PROFIL);
260 		if (++profprocs == 1) {
261 			s = splstatclock();
262 			psdiv = pscnt = psratio;
263 			setstatclockrate(profhz);
264 			splx(s);
265 		}
266 	}
267 }
268 
269 /*
270  * Stop profiling on a process.
271  */
272 void
273 stopprofclock(struct process *pr)
274 {
275 	int s;
276 
277 	if (pr->ps_flags & PS_PROFIL) {
278 		atomic_clearbits_int(&pr->ps_flags, PS_PROFIL);
279 		if (--profprocs == 0) {
280 			s = splstatclock();
281 			psdiv = pscnt = 1;
282 			setstatclockrate(stathz);
283 			splx(s);
284 		}
285 	}
286 }
287 
288 /*
289  * Statistics clock.  Grab profile sample, and if divider reaches 0,
290  * do process and kernel statistics.
291  */
292 void
293 statclock(struct clockframe *frame)
294 {
295 #if defined(GPROF) || defined(DDBPROF)
296 	struct gmonparam *g;
297 	u_long i;
298 #endif
299 	struct cpu_info *ci = curcpu();
300 	struct schedstate_percpu *spc = &ci->ci_schedstate;
301 	struct proc *p = curproc;
302 	struct process *pr;
303 
304 	/*
305 	 * Notice changes in divisor frequency, and adjust clock
306 	 * frequency accordingly.
307 	 */
308 	if (spc->spc_psdiv != psdiv) {
309 		spc->spc_psdiv = psdiv;
310 		spc->spc_pscnt = psdiv;
311 		if (psdiv == 1) {
312 			setstatclockrate(stathz);
313 		} else {
314 			setstatclockrate(profhz);
315 		}
316 	}
317 
318 	if (CLKF_USERMODE(frame)) {
319 		pr = p->p_p;
320 		if (pr->ps_flags & PS_PROFIL)
321 			addupc_intr(p, CLKF_PC(frame));
322 		if (--spc->spc_pscnt > 0)
323 			return;
324 		/*
325 		 * Came from user mode; CPU was in user state.
326 		 * If this process is being profiled record the tick.
327 		 */
328 		p->p_uticks++;
329 		if (pr->ps_nice > NZERO)
330 			spc->spc_cp_time[CP_NICE]++;
331 		else
332 			spc->spc_cp_time[CP_USER]++;
333 	} else {
334 #if defined(GPROF) || defined(DDBPROF)
335 		/*
336 		 * Kernel statistics are just like addupc_intr, only easier.
337 		 */
338 		g = ci->ci_gmon;
339 		if (g != NULL && g->state == GMON_PROF_ON) {
340 			i = CLKF_PC(frame) - g->lowpc;
341 			if (i < g->textsize) {
342 				i /= HISTFRACTION * sizeof(*g->kcount);
343 				g->kcount[i]++;
344 			}
345 		}
346 #endif
347 		if (p != NULL && p->p_p->ps_flags & PS_PROFIL)
348 			addupc_intr(p, PROC_PC(p));
349 		if (--spc->spc_pscnt > 0)
350 			return;
351 		/*
352 		 * Came from kernel mode, so we were:
353 		 * - spinning on a lock
354 		 * - handling an interrupt,
355 		 * - doing syscall or trap work on behalf of the current
356 		 *   user process, or
357 		 * - spinning in the idle loop.
358 		 * Whichever it is, charge the time as appropriate.
359 		 * Note that we charge interrupts to the current process,
360 		 * regardless of whether they are ``for'' that process,
361 		 * so that we know how much of its real time was spent
362 		 * in ``non-process'' (i.e., interrupt) work.
363 		 */
364 		if (CLKF_INTR(frame)) {
365 			if (p != NULL)
366 				p->p_iticks++;
367 			spc->spc_cp_time[spc->spc_spinning ?
368 			    CP_SPIN : CP_INTR]++;
369 		} else if (p != NULL && p != spc->spc_idleproc) {
370 			p->p_sticks++;
371 			spc->spc_cp_time[spc->spc_spinning ?
372 			    CP_SPIN : CP_SYS]++;
373 		} else
374 			spc->spc_cp_time[spc->spc_spinning ?
375 			    CP_SPIN : CP_IDLE]++;
376 	}
377 	spc->spc_pscnt = psdiv;
378 
379 	if (p != NULL) {
380 		p->p_cpticks++;
381 		/*
382 		 * If no schedclock is provided, call it here at ~~12-25 Hz;
383 		 * ~~16 Hz is best
384 		 */
385 		if (schedhz == 0) {
386 			if ((++spc->spc_schedticks & 3) == 0)
387 				schedclock(p);
388 		}
389 	}
390 }
391 
392 /*
393  * Return information about system clocks.
394  */
395 int
396 sysctl_clockrate(char *where, size_t *sizep, void *newp)
397 {
398 	struct clockinfo clkinfo;
399 
400 	/*
401 	 * Construct clockinfo structure.
402 	 */
403 	memset(&clkinfo, 0, sizeof clkinfo);
404 	clkinfo.tick = tick;
405 	clkinfo.hz = hz;
406 	clkinfo.profhz = profhz;
407 	clkinfo.stathz = stathz;
408 	return (sysctl_rdstruct(where, sizep, newp, &clkinfo, sizeof(clkinfo)));
409 }
410