xref: /netbsd-src/sys/kern/kern_clock.c (revision b1c86f5f087524e68db12794ee9c3e3da1ab17a0)
1 /*	$NetBSD: kern_clock.c,v 1.126 2008/10/05 21:57:20 pooka Exp $	*/
2 
3 /*-
4  * Copyright (c) 2000, 2004, 2006, 2007, 2008 The NetBSD Foundation, Inc.
5  * All rights reserved.
6  *
7  * This code is derived from software contributed to The NetBSD Foundation
8  * by Jason R. Thorpe of the Numerical Aerospace Simulation Facility,
9  * NASA Ames Research Center.
10  * This code is derived from software contributed to The NetBSD Foundation
11  * by Charles M. Hannum.
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  *
22  * THIS SOFTWARE IS PROVIDED BY THE NETBSD FOUNDATION, INC. AND CONTRIBUTORS
23  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED
24  * TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
25  * PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE FOUNDATION OR CONTRIBUTORS
26  * BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
27  * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
28  * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
29  * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
30  * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
31  * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
32  * POSSIBILITY OF SUCH DAMAGE.
33  */
34 
35 /*-
36  * Copyright (c) 1982, 1986, 1991, 1993
37  *	The Regents of the University of California.  All rights reserved.
38  * (c) UNIX System Laboratories, Inc.
39  * All or some portions of this file are derived from material licensed
40  * to the University of California by American Telephone and Telegraph
41  * Co. or Unix System Laboratories, Inc. and are reproduced herein with
42  * the permission of UNIX System Laboratories, Inc.
43  *
44  * Redistribution and use in source and binary forms, with or without
45  * modification, are permitted provided that the following conditions
46  * are met:
47  * 1. Redistributions of source code must retain the above copyright
48  *    notice, this list of conditions and the following disclaimer.
49  * 2. Redistributions in binary form must reproduce the above copyright
50  *    notice, this list of conditions and the following disclaimer in the
51  *    documentation and/or other materials provided with the distribution.
52  * 3. Neither the name of the University nor the names of its contributors
53  *    may be used to endorse or promote products derived from this software
54  *    without specific prior written permission.
55  *
56  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
57  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
58  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
59  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
60  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
61  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
62  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
63  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
64  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
65  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
66  * SUCH DAMAGE.
67  *
68  *	@(#)kern_clock.c	8.5 (Berkeley) 1/21/94
69  */
70 
71 #include <sys/cdefs.h>
72 __KERNEL_RCSID(0, "$NetBSD: kern_clock.c,v 1.126 2008/10/05 21:57:20 pooka Exp $");
73 
74 #include "opt_ntp.h"
75 #include "opt_perfctrs.h"
76 
77 #include <sys/param.h>
78 #include <sys/systm.h>
79 #include <sys/callout.h>
80 #include <sys/kernel.h>
81 #include <sys/proc.h>
82 #include <sys/resourcevar.h>
83 #include <sys/signalvar.h>
84 #include <sys/sysctl.h>
85 #include <sys/timex.h>
86 #include <sys/sched.h>
87 #include <sys/time.h>
88 #include <sys/timetc.h>
89 #include <sys/cpu.h>
90 #include <sys/atomic.h>
91 
92 #include <uvm/uvm_extern.h>
93 
94 #ifdef GPROF
95 #include <sys/gmon.h>
96 #endif
97 
98 /*
99  * Clock handling routines.
100  *
101  * This code is written to operate with two timers that run independently of
102  * each other.  The main clock, running hz times per second, is used to keep
103  * track of real time.  The second timer handles kernel and user profiling,
104  * and does resource use estimation.  If the second timer is programmable,
105  * it is randomized to avoid aliasing between the two clocks.  For example,
106  * the randomization prevents an adversary from always giving up the CPU
107  * just before its quantum expires.  Otherwise, it would never accumulate
108  * CPU ticks.  The mean frequency of the second timer is stathz.
109  *
110  * If no second timer exists, stathz will be zero; in this case we drive
111  * profiling and statistics off the main clock.  This WILL NOT be accurate;
112  * do not do it unless absolutely necessary.
113  *
114  * The statistics clock may (or may not) be run at a higher rate while
115  * profiling.  This profile clock runs at profhz.  We require that profhz
116  * be an integral multiple of stathz.
117  *
118  * If the statistics clock is running fast, it must be divided by the ratio
119  * profhz/stathz for statistics.  (For profiling, every tick counts.)
120  */
121 
122 int	stathz;
123 int	profhz;
124 int	profsrc;
125 int	schedhz;
126 int	profprocs;
127 int	hardclock_ticks;
128 static int hardscheddiv; /* hard => sched divider (used if schedhz == 0) */
129 static int psdiv;			/* prof => stat divider */
130 int	psratio;			/* ratio: prof / stat */
131 
132 static u_int get_intr_timecount(struct timecounter *);
133 
134 static struct timecounter intr_timecounter = {
135 	get_intr_timecount,	/* get_timecount */
136 	0,			/* no poll_pps */
137 	~0u,			/* counter_mask */
138 	0,		        /* frequency */
139 	"clockinterrupt",	/* name */
140 	0,			/* quality - minimum implementation level for a clock */
141 	NULL,			/* prev */
142 	NULL,			/* next */
143 };
144 
145 static u_int
146 get_intr_timecount(struct timecounter *tc)
147 {
148 
149 	return (u_int)hardclock_ticks;
150 }
151 
152 /*
153  * Initialize clock frequencies and start both clocks running.
154  */
155 void
156 initclocks(void)
157 {
158 	int i;
159 
160 	/*
161 	 * Set divisors to 1 (normal case) and let the machine-specific
162 	 * code do its bit.
163 	 */
164 	psdiv = 1;
165 	/*
166 	 * provide minimum default time counter
167 	 * will only run at interrupt resolution
168 	 */
169 	intr_timecounter.tc_frequency = hz;
170 	tc_init(&intr_timecounter);
171 	cpu_initclocks();
172 
173 	/*
174 	 * Compute profhz and stathz, fix profhz if needed.
175 	 */
176 	i = stathz ? stathz : hz;
177 	if (profhz == 0)
178 		profhz = i;
179 	psratio = profhz / i;
180 	if (schedhz == 0) {
181 		/* 16Hz is best */
182 		hardscheddiv = hz / 16;
183 		if (hardscheddiv <= 0)
184 			panic("hardscheddiv");
185 	}
186 
187 }
188 
189 /*
190  * The real-time timer, interrupting hz times per second.
191  */
192 void
193 hardclock(struct clockframe *frame)
194 {
195 	struct lwp *l;
196 	struct cpu_info *ci;
197 
198 	ci = curcpu();
199 	l = ci->ci_data.cpu_onproc;
200 
201 	timer_tick(l, CLKF_USERMODE(frame));
202 
203 	/*
204 	 * If no separate statistics clock is available, run it from here.
205 	 */
206 	if (stathz == 0)
207 		statclock(frame);
208 	/*
209 	 * If no separate schedclock is provided, call it here
210 	 * at about 16 Hz.
211 	 */
212 	if (schedhz == 0) {
213 		if ((int)(--ci->ci_schedstate.spc_schedticks) <= 0) {
214 			schedclock(l);
215 			ci->ci_schedstate.spc_schedticks = hardscheddiv;
216 		}
217 	}
218 	if ((--ci->ci_schedstate.spc_ticks) <= 0)
219 		sched_tick(ci);
220 
221 	if (CPU_IS_PRIMARY(ci)) {
222 		hardclock_ticks++;
223 		tc_ticktock();
224 	}
225 
226 	/*
227 	 * Update real-time timeout queue.
228 	 */
229 	callout_hardclock();
230 }
231 
232 /*
233  * Start profiling on a process.
234  *
235  * Kernel profiling passes proc0 which never exits and hence
236  * keeps the profile clock running constantly.
237  */
238 void
239 startprofclock(struct proc *p)
240 {
241 
242 	KASSERT(mutex_owned(&p->p_stmutex));
243 
244 	if ((p->p_stflag & PST_PROFIL) == 0) {
245 		p->p_stflag |= PST_PROFIL;
246 		/*
247 		 * This is only necessary if using the clock as the
248 		 * profiling source.
249 		 */
250 		if (++profprocs == 1 && stathz != 0)
251 			psdiv = psratio;
252 	}
253 }
254 
255 /*
256  * Stop profiling on a process.
257  */
258 void
259 stopprofclock(struct proc *p)
260 {
261 
262 	KASSERT(mutex_owned(&p->p_stmutex));
263 
264 	if (p->p_stflag & PST_PROFIL) {
265 		p->p_stflag &= ~PST_PROFIL;
266 		/*
267 		 * This is only necessary if using the clock as the
268 		 * profiling source.
269 		 */
270 		if (--profprocs == 0 && stathz != 0)
271 			psdiv = 1;
272 	}
273 }
274 
275 #if defined(PERFCTRS)
276 /*
277  * Independent profiling "tick" in case we're using a separate
278  * clock or profiling event source.  Currently, that's just
279  * performance counters--hence the wrapper.
280  */
281 void
282 proftick(struct clockframe *frame)
283 {
284 #ifdef GPROF
285         struct gmonparam *g;
286         intptr_t i;
287 #endif
288 	struct lwp *l;
289 	struct proc *p;
290 
291 	l = curcpu()->ci_data.cpu_onproc;
292 	p = (l ? l->l_proc : NULL);
293 	if (CLKF_USERMODE(frame)) {
294 		mutex_spin_enter(&p->p_stmutex);
295 		if (p->p_stflag & PST_PROFIL)
296 			addupc_intr(l, CLKF_PC(frame));
297 		mutex_spin_exit(&p->p_stmutex);
298 	} else {
299 #ifdef GPROF
300 		g = &_gmonparam;
301 		if (g->state == GMON_PROF_ON) {
302 			i = CLKF_PC(frame) - g->lowpc;
303 			if (i < g->textsize) {
304 				i /= HISTFRACTION * sizeof(*g->kcount);
305 				g->kcount[i]++;
306 			}
307 		}
308 #endif
309 #ifdef LWP_PC
310 		if (p != NULL && (p->p_stflag & PST_PROFIL) != 0)
311 			addupc_intr(l, LWP_PC(l));
312 #endif
313 	}
314 }
315 #endif
316 
317 void
318 schedclock(struct lwp *l)
319 {
320 	struct cpu_info *ci;
321 
322 	ci = l->l_cpu;
323 
324 	/* Accumulate syscall and context switch counts. */
325 	atomic_add_int((unsigned *)&uvmexp.swtch, ci->ci_data.cpu_nswtch);
326 	ci->ci_data.cpu_nswtch = 0;
327 	atomic_add_int((unsigned *)&uvmexp.syscalls, ci->ci_data.cpu_nsyscall);
328 	ci->ci_data.cpu_nsyscall = 0;
329 	atomic_add_int((unsigned *)&uvmexp.traps, ci->ci_data.cpu_ntrap);
330 	ci->ci_data.cpu_ntrap = 0;
331 
332 	if ((l->l_flag & LW_IDLE) != 0)
333 		return;
334 
335 	sched_schedclock(l);
336 }
337 
338 /*
339  * Statistics clock.  Grab profile sample, and if divider reaches 0,
340  * do process and kernel statistics.
341  */
342 void
343 statclock(struct clockframe *frame)
344 {
345 #ifdef GPROF
346 	struct gmonparam *g;
347 	intptr_t i;
348 #endif
349 	struct cpu_info *ci = curcpu();
350 	struct schedstate_percpu *spc = &ci->ci_schedstate;
351 	struct proc *p;
352 	struct lwp *l;
353 
354 	/*
355 	 * Notice changes in divisor frequency, and adjust clock
356 	 * frequency accordingly.
357 	 */
358 	if (spc->spc_psdiv != psdiv) {
359 		spc->spc_psdiv = psdiv;
360 		spc->spc_pscnt = psdiv;
361 		if (psdiv == 1) {
362 			setstatclockrate(stathz);
363 		} else {
364 			setstatclockrate(profhz);
365 		}
366 	}
367 	l = ci->ci_data.cpu_onproc;
368 	if ((l->l_flag & LW_IDLE) != 0) {
369 		/*
370 		 * don't account idle lwps as swapper.
371 		 */
372 		p = NULL;
373 	} else {
374 		p = l->l_proc;
375 		mutex_spin_enter(&p->p_stmutex);
376 	}
377 
378 	if (CLKF_USERMODE(frame)) {
379 		if ((p->p_stflag & PST_PROFIL) && profsrc == PROFSRC_CLOCK)
380 			addupc_intr(l, CLKF_PC(frame));
381 		if (--spc->spc_pscnt > 0) {
382 			mutex_spin_exit(&p->p_stmutex);
383 			return;
384 		}
385 
386 		/*
387 		 * Came from user mode; CPU was in user state.
388 		 * If this process is being profiled record the tick.
389 		 */
390 		p->p_uticks++;
391 		if (p->p_nice > NZERO)
392 			spc->spc_cp_time[CP_NICE]++;
393 		else
394 			spc->spc_cp_time[CP_USER]++;
395 	} else {
396 #ifdef GPROF
397 		/*
398 		 * Kernel statistics are just like addupc_intr, only easier.
399 		 */
400 		g = &_gmonparam;
401 		if (profsrc == PROFSRC_CLOCK && g->state == GMON_PROF_ON) {
402 			i = CLKF_PC(frame) - g->lowpc;
403 			if (i < g->textsize) {
404 				i /= HISTFRACTION * sizeof(*g->kcount);
405 				g->kcount[i]++;
406 			}
407 		}
408 #endif
409 #ifdef LWP_PC
410 		if (p != NULL && profsrc == PROFSRC_CLOCK &&
411 		    (p->p_stflag & PST_PROFIL)) {
412 			addupc_intr(l, LWP_PC(l));
413 		}
414 #endif
415 		if (--spc->spc_pscnt > 0) {
416 			if (p != NULL)
417 				mutex_spin_exit(&p->p_stmutex);
418 			return;
419 		}
420 		/*
421 		 * Came from kernel mode, so we were:
422 		 * - handling an interrupt,
423 		 * - doing syscall or trap work on behalf of the current
424 		 *   user process, or
425 		 * - spinning in the idle loop.
426 		 * Whichever it is, charge the time as appropriate.
427 		 * Note that we charge interrupts to the current process,
428 		 * regardless of whether they are ``for'' that process,
429 		 * so that we know how much of its real time was spent
430 		 * in ``non-process'' (i.e., interrupt) work.
431 		 */
432 		if (CLKF_INTR(frame) || (curlwp->l_pflag & LP_INTR) != 0) {
433 			if (p != NULL) {
434 				p->p_iticks++;
435 			}
436 			spc->spc_cp_time[CP_INTR]++;
437 		} else if (p != NULL) {
438 			p->p_sticks++;
439 			spc->spc_cp_time[CP_SYS]++;
440 		} else {
441 			spc->spc_cp_time[CP_IDLE]++;
442 		}
443 	}
444 	spc->spc_pscnt = psdiv;
445 
446 	if (p != NULL) {
447 		atomic_inc_uint(&l->l_cpticks);
448 		mutex_spin_exit(&p->p_stmutex);
449 	}
450 }
451