xref: /netbsd-src/sys/kern/kern_clock.c (revision 627f7eb200a4419d89b531d55fccd2ee3ffdcde0)
1 /*	$NetBSD: kern_clock.c,v 1.144 2021/01/16 02:20:00 riastradh 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 1.144 2021/01/16 02:20:00 riastradh Exp $");
73 
74 #ifdef _KERNEL_OPT
75 #include "opt_dtrace.h"
76 #include "opt_gprof.h"
77 #endif
78 
79 #include <sys/param.h>
80 #include <sys/systm.h>
81 #include <sys/callout.h>
82 #include <sys/kernel.h>
83 #include <sys/proc.h>
84 #include <sys/resourcevar.h>
85 #include <sys/signalvar.h>
86 #include <sys/sysctl.h>
87 #include <sys/timex.h>
88 #include <sys/sched.h>
89 #include <sys/time.h>
90 #include <sys/timetc.h>
91 #include <sys/cpu.h>
92 #include <sys/atomic.h>
93 #include <sys/rndsource.h>
94 
95 #ifdef GPROF
96 #include <sys/gmon.h>
97 #endif
98 
99 #ifdef KDTRACE_HOOKS
100 #include <sys/dtrace_bsd.h>
101 #include <sys/cpu.h>
102 
103 cyclic_clock_func_t	cyclic_clock_func[MAXCPUS];
104 #endif
105 
106 static int sysctl_kern_clockrate(SYSCTLFN_PROTO);
107 
108 /*
109  * Clock handling routines.
110  *
111  * This code is written to operate with two timers that run independently of
112  * each other.  The main clock, running hz times per second, is used to keep
113  * track of real time.  The second timer handles kernel and user profiling,
114  * and does resource use estimation.  If the second timer is programmable,
115  * it is randomized to avoid aliasing between the two clocks.  For example,
116  * the randomization prevents an adversary from always giving up the CPU
117  * just before its quantum expires.  Otherwise, it would never accumulate
118  * CPU ticks.  The mean frequency of the second timer is stathz.
119  *
120  * If no second timer exists, stathz will be zero; in this case we drive
121  * profiling and statistics off the main clock.  This WILL NOT be accurate;
122  * do not do it unless absolutely necessary.
123  *
124  * The statistics clock may (or may not) be run at a higher rate while
125  * profiling.  This profile clock runs at profhz.  We require that profhz
126  * be an integral multiple of stathz.
127  *
128  * If the statistics clock is running fast, it must be divided by the ratio
129  * profhz/stathz for statistics.  (For profiling, every tick counts.)
130  */
131 
132 int	stathz;
133 int	profhz;
134 int	profsrc;
135 int	schedhz;
136 int	profprocs;
137 int	hardclock_ticks;
138 static int hardscheddiv; /* hard => sched divider (used if schedhz == 0) */
139 static int psdiv;			/* prof => stat divider */
140 int	psratio;			/* ratio: prof / stat */
141 
142 struct clockrnd {
143 	struct krndsource source;
144 	unsigned needed;
145 };
146 
147 static struct clockrnd hardclockrnd __aligned(COHERENCY_UNIT);
148 static struct clockrnd statclockrnd __aligned(COHERENCY_UNIT);
149 
150 static void
151 clockrnd_get(size_t needed, void *cookie)
152 {
153 	struct clockrnd *C = cookie;
154 
155 	/* Start sampling.  */
156 	atomic_store_relaxed(&C->needed, 2*NBBY*needed);
157 }
158 
159 static void
160 clockrnd_sample(struct clockrnd *C)
161 {
162 	struct cpu_info *ci = curcpu();
163 
164 	/* If there's nothing needed right now, stop here.  */
165 	if (__predict_true(C->needed == 0))
166 		return;
167 
168 	/*
169 	 * If we're not the primary core of a package, we're probably
170 	 * driven by the same clock as the primary core, so don't
171 	 * bother.
172 	 */
173 	if (ci != ci->ci_package1st)
174 		return;
175 
176 	/* Take a sample and enter it into the pool.  */
177 	rnd_add_uint32(&C->source, 0);
178 
179 	/*
180 	 * On the primary CPU, count down.  Using an atomic decrement
181 	 * here isn't really necessary -- on every platform we care
182 	 * about, stores to unsigned int are atomic, and the only other
183 	 * memory operation that could happen here is for another CPU
184 	 * to store a higher value for needed.  But using an atomic
185 	 * decrement avoids giving the impression of data races, and is
186 	 * unlikely to hurt because only one CPU will ever be writing
187 	 * to the location.
188 	 */
189 	if (CPU_IS_PRIMARY(curcpu())) {
190 		unsigned needed __diagused;
191 
192 		needed = atomic_dec_uint_nv(&C->needed);
193 		KASSERT(needed != UINT_MAX);
194 	}
195 }
196 
197 static u_int get_intr_timecount(struct timecounter *);
198 
199 static struct timecounter intr_timecounter = {
200 	.tc_get_timecount	= get_intr_timecount,
201 	.tc_poll_pps		= NULL,
202 	.tc_counter_mask	= ~0u,
203 	.tc_frequency		= 0,
204 	.tc_name		= "clockinterrupt",
205 	/* quality - minimum implementation level for a clock */
206 	.tc_quality		= 0,
207 	.tc_priv		= NULL,
208 };
209 
210 static u_int
211 get_intr_timecount(struct timecounter *tc)
212 {
213 
214 	return (u_int)getticks();
215 }
216 
217 int
218 getticks(void)
219 {
220 	return atomic_load_relaxed(&hardclock_ticks);
221 }
222 
223 /*
224  * Initialize clock frequencies and start both clocks running.
225  */
226 void
227 initclocks(void)
228 {
229 	static struct sysctllog *clog;
230 	int i;
231 
232 	/*
233 	 * Set divisors to 1 (normal case) and let the machine-specific
234 	 * code do its bit.
235 	 */
236 	psdiv = 1;
237 
238 	/*
239 	 * Call cpu_initclocks() before registering the default
240 	 * timecounter, in case it needs to adjust hz.
241 	 */
242 	const int old_hz = hz;
243 	cpu_initclocks();
244 	if (old_hz != hz) {
245 		tick = 1000000 / hz;
246 		tickadj = (240000 / (60 * hz)) ? (240000 / (60 * hz)) : 1;
247 	}
248 
249 	/*
250 	 * provide minimum default time counter
251 	 * will only run at interrupt resolution
252 	 */
253 	intr_timecounter.tc_frequency = hz;
254 	tc_init(&intr_timecounter);
255 
256 	/*
257 	 * Compute profhz and stathz, fix profhz if needed.
258 	 */
259 	i = stathz ? stathz : hz;
260 	if (profhz == 0)
261 		profhz = i;
262 	psratio = profhz / i;
263 	if (schedhz == 0) {
264 		/* 16Hz is best */
265 		hardscheddiv = hz / 16;
266 		if (hardscheddiv <= 0)
267 			panic("hardscheddiv");
268 	}
269 
270 	sysctl_createv(&clog, 0, NULL, NULL,
271 		       CTLFLAG_PERMANENT,
272 		       CTLTYPE_STRUCT, "clockrate",
273 		       SYSCTL_DESCR("Kernel clock rates"),
274 		       sysctl_kern_clockrate, 0, NULL,
275 		       sizeof(struct clockinfo),
276 		       CTL_KERN, KERN_CLOCKRATE, CTL_EOL);
277 	sysctl_createv(&clog, 0, NULL, NULL,
278 		       CTLFLAG_PERMANENT,
279 		       CTLTYPE_INT, "hardclock_ticks",
280 		       SYSCTL_DESCR("Number of hardclock ticks"),
281 		       NULL, 0, &hardclock_ticks, sizeof(hardclock_ticks),
282 		       CTL_KERN, KERN_HARDCLOCK_TICKS, CTL_EOL);
283 
284 	rndsource_setcb(&hardclockrnd.source, clockrnd_get, &hardclockrnd);
285 	rnd_attach_source(&hardclockrnd.source, "hardclock", RND_TYPE_SKEW,
286 	    RND_FLAG_COLLECT_TIME|RND_FLAG_HASCB);
287 	if (stathz) {
288 		rndsource_setcb(&statclockrnd.source, clockrnd_get,
289 		    &statclockrnd);
290 		rnd_attach_source(&statclockrnd.source, "statclock",
291 		    RND_TYPE_SKEW, RND_FLAG_COLLECT_TIME|RND_FLAG_HASCB);
292 	}
293 }
294 
295 /*
296  * The real-time timer, interrupting hz times per second.
297  */
298 void
299 hardclock(struct clockframe *frame)
300 {
301 	struct lwp *l;
302 	struct cpu_info *ci;
303 
304 	clockrnd_sample(&hardclockrnd);
305 
306 	ci = curcpu();
307 	l = ci->ci_onproc;
308 
309 	ptimer_tick(l, CLKF_USERMODE(frame));
310 
311 	/*
312 	 * If no separate statistics clock is available, run it from here.
313 	 */
314 	if (stathz == 0)
315 		statclock(frame);
316 	/*
317 	 * If no separate schedclock is provided, call it here
318 	 * at about 16 Hz.
319 	 */
320 	if (schedhz == 0) {
321 		if ((int)(--ci->ci_schedstate.spc_schedticks) <= 0) {
322 			schedclock(l);
323 			ci->ci_schedstate.spc_schedticks = hardscheddiv;
324 		}
325 	}
326 	if ((--ci->ci_schedstate.spc_ticks) <= 0)
327 		sched_tick(ci);
328 
329 	if (CPU_IS_PRIMARY(ci)) {
330 		atomic_store_relaxed(&hardclock_ticks,
331 		    atomic_load_relaxed(&hardclock_ticks) + 1);
332 		tc_ticktock();
333 	}
334 
335 	/*
336 	 * Update real-time timeout queue.
337 	 */
338 	callout_hardclock();
339 }
340 
341 /*
342  * Start profiling on a process.
343  *
344  * Kernel profiling passes proc0 which never exits and hence
345  * keeps the profile clock running constantly.
346  */
347 void
348 startprofclock(struct proc *p)
349 {
350 
351 	KASSERT(mutex_owned(&p->p_stmutex));
352 
353 	if ((p->p_stflag & PST_PROFIL) == 0) {
354 		p->p_stflag |= PST_PROFIL;
355 		/*
356 		 * This is only necessary if using the clock as the
357 		 * profiling source.
358 		 */
359 		if (++profprocs == 1 && stathz != 0)
360 			psdiv = psratio;
361 	}
362 }
363 
364 /*
365  * Stop profiling on a process.
366  */
367 void
368 stopprofclock(struct proc *p)
369 {
370 
371 	KASSERT(mutex_owned(&p->p_stmutex));
372 
373 	if (p->p_stflag & PST_PROFIL) {
374 		p->p_stflag &= ~PST_PROFIL;
375 		/*
376 		 * This is only necessary if using the clock as the
377 		 * profiling source.
378 		 */
379 		if (--profprocs == 0 && stathz != 0)
380 			psdiv = 1;
381 	}
382 }
383 
384 void
385 schedclock(struct lwp *l)
386 {
387 	if ((l->l_flag & LW_IDLE) != 0)
388 		return;
389 
390 	sched_schedclock(l);
391 }
392 
393 /*
394  * Statistics clock.  Grab profile sample, and if divider reaches 0,
395  * do process and kernel statistics.
396  */
397 void
398 statclock(struct clockframe *frame)
399 {
400 #ifdef GPROF
401 	struct gmonparam *g;
402 	intptr_t i;
403 #endif
404 	struct cpu_info *ci = curcpu();
405 	struct schedstate_percpu *spc = &ci->ci_schedstate;
406 	struct proc *p;
407 	struct lwp *l;
408 
409 	if (stathz)
410 		clockrnd_sample(&statclockrnd);
411 
412 	/*
413 	 * Notice changes in divisor frequency, and adjust clock
414 	 * frequency accordingly.
415 	 */
416 	if (spc->spc_psdiv != psdiv) {
417 		spc->spc_psdiv = psdiv;
418 		spc->spc_pscnt = psdiv;
419 		if (psdiv == 1) {
420 			setstatclockrate(stathz);
421 		} else {
422 			setstatclockrate(profhz);
423 		}
424 	}
425 	l = ci->ci_onproc;
426 	if ((l->l_flag & LW_IDLE) != 0) {
427 		/*
428 		 * don't account idle lwps as swapper.
429 		 */
430 		p = NULL;
431 	} else {
432 		p = l->l_proc;
433 		mutex_spin_enter(&p->p_stmutex);
434 	}
435 
436 	if (CLKF_USERMODE(frame)) {
437 		KASSERT(p != NULL);
438 		if ((p->p_stflag & PST_PROFIL) && profsrc == PROFSRC_CLOCK)
439 			addupc_intr(l, CLKF_PC(frame));
440 		if (--spc->spc_pscnt > 0) {
441 			mutex_spin_exit(&p->p_stmutex);
442 			return;
443 		}
444 
445 		/*
446 		 * Came from user mode; CPU was in user state.
447 		 * If this process is being profiled record the tick.
448 		 */
449 		p->p_uticks++;
450 		if (p->p_nice > NZERO)
451 			spc->spc_cp_time[CP_NICE]++;
452 		else
453 			spc->spc_cp_time[CP_USER]++;
454 	} else {
455 #ifdef GPROF
456 		/*
457 		 * Kernel statistics are just like addupc_intr, only easier.
458 		 */
459 		g = &_gmonparam;
460 		if (profsrc == PROFSRC_CLOCK && g->state == GMON_PROF_ON) {
461 			i = CLKF_PC(frame) - g->lowpc;
462 			if (i < g->textsize) {
463 				i /= HISTFRACTION * sizeof(*g->kcount);
464 				g->kcount[i]++;
465 			}
466 		}
467 #endif
468 #ifdef LWP_PC
469 		if (p != NULL && profsrc == PROFSRC_CLOCK &&
470 		    (p->p_stflag & PST_PROFIL)) {
471 			addupc_intr(l, LWP_PC(l));
472 		}
473 #endif
474 		if (--spc->spc_pscnt > 0) {
475 			if (p != NULL)
476 				mutex_spin_exit(&p->p_stmutex);
477 			return;
478 		}
479 		/*
480 		 * Came from kernel mode, so we were:
481 		 * - handling an interrupt,
482 		 * - doing syscall or trap work on behalf of the current
483 		 *   user process, or
484 		 * - spinning in the idle loop.
485 		 * Whichever it is, charge the time as appropriate.
486 		 * Note that we charge interrupts to the current process,
487 		 * regardless of whether they are ``for'' that process,
488 		 * so that we know how much of its real time was spent
489 		 * in ``non-process'' (i.e., interrupt) work.
490 		 */
491 		if (CLKF_INTR(frame) || (curlwp->l_pflag & LP_INTR) != 0) {
492 			if (p != NULL) {
493 				p->p_iticks++;
494 			}
495 			spc->spc_cp_time[CP_INTR]++;
496 		} else if (p != NULL) {
497 			p->p_sticks++;
498 			spc->spc_cp_time[CP_SYS]++;
499 		} else {
500 			spc->spc_cp_time[CP_IDLE]++;
501 		}
502 	}
503 	spc->spc_pscnt = psdiv;
504 
505 	if (p != NULL) {
506 		atomic_inc_uint(&l->l_cpticks);
507 		mutex_spin_exit(&p->p_stmutex);
508 	}
509 
510 #ifdef KDTRACE_HOOKS
511 	cyclic_clock_func_t func = cyclic_clock_func[cpu_index(ci)];
512 	if (func) {
513 		(*func)((struct clockframe *)frame);
514 	}
515 #endif
516 }
517 
518 /*
519  * sysctl helper routine for kern.clockrate. Assembles a struct on
520  * the fly to be returned to the caller.
521  */
522 static int
523 sysctl_kern_clockrate(SYSCTLFN_ARGS)
524 {
525 	struct clockinfo clkinfo;
526 	struct sysctlnode node;
527 
528 	clkinfo.tick = tick;
529 	clkinfo.tickadj = tickadj;
530 	clkinfo.hz = hz;
531 	clkinfo.profhz = profhz;
532 	clkinfo.stathz = stathz ? stathz : hz;
533 
534 	node = *rnode;
535 	node.sysctl_data = &clkinfo;
536 	return (sysctl_lookup(SYSCTLFN_CALL(&node)));
537 }
538