xref: /netbsd-src/sys/kern/sched_4bsd.c (revision cdd498c00dcf861038107c594e5bfe5c0990efb6)
1 /*	$NetBSD: sched_4bsd.c,v 1.47 2025/01/17 04:11:33 mrg Exp $	*/
2 
3 /*
4  * Copyright (c) 1999, 2000, 2004, 2006, 2007, 2008, 2019, 2020
5  *     The NetBSD Foundation, Inc.
6  * All rights reserved.
7  *
8  * This code is derived from software contributed to The NetBSD Foundation
9  * by Jason R. Thorpe of the Numerical Aerospace Simulation Facility,
10  * NASA Ames Research Center, by Charles M. Hannum, Andrew Doran, and
11  * Daniel Sieger.
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, 1990, 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_synch.c	8.9 (Berkeley) 5/19/95
69  */
70 
71 #include <sys/cdefs.h>
72 __KERNEL_RCSID(0, "$NetBSD: sched_4bsd.c,v 1.47 2025/01/17 04:11:33 mrg Exp $");
73 
74 #include "opt_ddb.h"
75 #include "opt_lockdebug.h"
76 
77 #include <sys/param.h>
78 #include <sys/systm.h>
79 #include <sys/callout.h>
80 #include <sys/cpu.h>
81 #include <sys/proc.h>
82 #include <sys/kernel.h>
83 #include <sys/resourcevar.h>
84 #include <sys/sched.h>
85 #include <sys/sysctl.h>
86 #include <sys/lockdebug.h>
87 #include <sys/intr.h>
88 #include <sys/atomic.h>
89 
90 static void updatepri(struct lwp *);
91 static void resetpriority(struct lwp *);
92 
93 /* Number of hardclock ticks per sched_tick() */
94 u_int sched_rrticks __read_mostly;
95 
96 /*
97  * Force switch among equal priority processes every 100ms.
98  * Called from hardclock every hz/10 == sched_rrticks hardclock ticks.
99  */
100 /* ARGSUSED */
101 void
102 sched_tick(struct cpu_info *ci)
103 {
104 	struct schedstate_percpu *spc = &ci->ci_schedstate;
105 	pri_t pri = PRI_NONE;
106 	lwp_t *l;
107 
108 	spc->spc_ticks = sched_rrticks;
109 
110 	if (CURCPU_IDLE_P()) {
111 		spc_lock(ci);
112 		sched_resched_cpu(ci, MAXPRI_KTHREAD, true);
113 		/* spc now unlocked */
114 		return;
115 	}
116 	l = ci->ci_onproc;
117 	if (l == NULL) {
118 		return;
119 	}
120 	/*
121 	 * Can only be spc_lwplock or a turnstile lock at this point
122 	 * (if we interrupted priority inheritance trylock dance).
123 	 */
124 	KASSERT(l->l_mutex != spc->spc_mutex);
125 	switch (l->l_class) {
126 	case SCHED_FIFO:
127 		/* No timeslicing for FIFO jobs. */
128 		break;
129 	case SCHED_RR:
130 		/* Force it into mi_switch() to look for other jobs to run. */
131 		pri = MAXPRI_KERNEL_RT;
132 		break;
133 	default:
134 		if (spc->spc_flags & SPCF_SHOULDYIELD) {
135 			/*
136 			 * Process is stuck in kernel somewhere, probably
137 			 * due to buggy or inefficient code.  Force a
138 			 * kernel preemption.
139 			 */
140 			pri = MAXPRI_KERNEL_RT;
141 		} else if (spc->spc_flags & SPCF_SEENRR) {
142 			/*
143 			 * The process has already been through a roundrobin
144 			 * without switching and may be hogging the CPU.
145 			 * Indicate that the process should yield.
146 			 */
147 			pri = MAXPRI_KTHREAD;
148 			spc->spc_flags |= SPCF_SHOULDYIELD;
149 		} else if (!cpu_is_1stclass(ci)) {
150 			/*
151 			 * For SMT or asymmetric systems push a little
152 			 * harder: if this is not a 1st class CPU, try to
153 			 * find a better one to run this LWP.
154 			 */
155 			pri = MAXPRI_KTHREAD;
156 			spc->spc_flags |= SPCF_SHOULDYIELD;
157 		} else {
158 			spc->spc_flags |= SPCF_SEENRR;
159 		}
160 		break;
161 	}
162 
163 	if (pri != PRI_NONE) {
164 		spc_lock(ci);
165 		sched_resched_cpu(ci, pri, true);
166 		/* spc now unlocked */
167 	}
168 }
169 
170 /*
171  * Why PRIO_MAX - 2? From setpriority(2):
172  *
173  *	prio is a value in the range -20 to 20.  The default priority is
174  *	0; lower priorities cause more favorable scheduling.  A value of
175  *	19 or 20 will schedule a process only when nothing at priority <=
176  *	0 is runnable.
177  *
178  * This gives estcpu influence over 18 priority levels, and leaves nice
179  * with 40 levels.  One way to think about it is that nice has 20 levels
180  * either side of estcpu's 18.
181  */
182 #define	ESTCPU_SHIFT	11
183 #define	ESTCPU_MAX	((PRIO_MAX - 2) << ESTCPU_SHIFT)
184 #define	ESTCPU_ACCUM	(1 << (ESTCPU_SHIFT - 1))
185 #define	ESTCPULIM(e)	uimin((e), ESTCPU_MAX)
186 
187 /*
188  * The main parameter used by this algorithm is 'l_estcpu'. It is an estimate
189  * of the recent CPU utilization of the thread.
190  *
191  * l_estcpu is:
192  *  - increased each time the hardclock ticks and the thread is found to
193  *    be executing, in sched_schedclock() called from hardclock()
194  *  - decreased (filtered) on each sched tick, in sched_pstats_hook()
195  * If the lwp is sleeping for more than a second, we don't touch l_estcpu: it
196  * will be updated in sched_setrunnable() when the lwp wakes up, in burst mode
197  * (ie, we decrease it n times).
198  *
199  * Note that hardclock updates l_estcpu and l_cpticks independently.
200  *
201  * -----------------------------------------------------------------------------
202  *
203  * Here we describe how l_estcpu is decreased.
204  *
205  * Constants for digital decay (filter):
206  *     90% of l_estcpu usage in (5 * loadavg) seconds
207  *
208  * We wish to decay away 90% of l_estcpu in (5 * loadavg) seconds. That is, we
209  * want to compute a value of decay such that the following loop:
210  *     for (i = 0; i < (5 * loadavg); i++)
211  *         l_estcpu *= decay;
212  * will result in
213  *     l_estcpu *= 0.1;
214  * for all values of loadavg.
215  *
216  * Mathematically this loop can be expressed by saying:
217  *     decay ** (5 * loadavg) ~= .1
218  *
219  * And finally, the corresponding value of decay we're using is:
220  *     decay = (2 * loadavg) / (2 * loadavg + 1)
221  *
222  * -----------------------------------------------------------------------------
223  *
224  * Now, let's prove that the value of decay stated above will always fulfill
225  * the equation:
226  *     decay ** (5 * loadavg) ~= .1
227  *
228  * If we compute b as:
229  *     b = 2 * loadavg
230  * then
231  *     decay = b / (b + 1)
232  *
233  * We now need to prove two things:
234  *     1) Given [factor ** (5 * loadavg) =~ .1], prove [factor == b/(b+1)].
235  *     2) Given [b/(b+1) ** power =~ .1], prove [power == (5 * loadavg)].
236  *
237  * Facts:
238  *   * For x real: exp(x) = 0! + x**1/1! + x**2/2! + ...
239  *     Therefore, for x close to zero, exp(x) =~ 1 + x.
240  *     In turn, for b large enough, exp(-1/b) =~ 1 - (1/b) = (b-1)/b.
241  *
242  *   * For b large enough, (b-1)/b =~ b/(b+1).
243  *
244  *   * For x belonging to [-1;1[, ln(1-x) = - x - x**2/2 - x**3/3 - ...
245  *     Therefore ln(b/(b+1)) = ln(1 - 1/(b+1)) =~ -1/(b+1).
246  *
247  *   * ln(0.1) =~ -2.30
248  *
249  * Proof of (1):
250  *     factor ** (5 * loadavg) =~ 0.1
251  *  => ln(factor) =~ -2.30 / (5 * loadavg)
252  *  => factor =~ exp(-1 / ((5 / 2.30) * loadavg))
253  *            =~ exp(-1 / (2 * loadavg))
254  *            =~ exp(-1 / b)
255  *            =~ (b - 1) / b
256  *            =~ b / (b + 1)
257  *            =~ (2 * loadavg) / ((2 * loadavg) + 1)
258  *
259  * Proof of (2):
260  *     (b / (b + 1)) ** power =~ .1
261  *  => power * ln(b / (b + 1)) =~ -2.30
262  *  => power * (-1 / (b + 1)) =~ -2.30
263  *  => power =~ 2.30 * (b + 1)
264  *  => power =~ 4.60 * loadavg + 2.30
265  *  => power =~ 5 * loadavg
266  *
267  * Conclusion: decay = (2 * loadavg) / (2 * loadavg + 1)
268  */
269 
270 /* See calculations above */
271 #define	loadfactor(loadavg)  (2 * (loadavg))
272 
273 static fixpt_t
274 decay_cpu(fixpt_t loadfac, fixpt_t estcpu)
275 {
276 
277 	if (estcpu == 0) {
278 		return 0;
279 	}
280 
281 #if !defined(_LP64)
282 	/* avoid 64bit arithmetics. */
283 #define	FIXPT_MAX ((fixpt_t)((UINTMAX_C(1) << sizeof(fixpt_t) * CHAR_BIT) - 1))
284 	if (__predict_true(loadfac <= FIXPT_MAX / ESTCPU_MAX)) {
285 		return estcpu * loadfac / (loadfac + FSCALE);
286 	}
287 #endif
288 
289 	return (uint64_t)estcpu * loadfac / (loadfac + FSCALE);
290 }
291 
292 static fixpt_t
293 decay_cpu_batch(fixpt_t loadfac, fixpt_t estcpu, unsigned int n)
294 {
295 
296 	/*
297 	 * For all load averages >= 1 and max l_estcpu of (255 << ESTCPU_SHIFT),
298 	 * if we slept for at least seven times the loadfactor, we will decay
299 	 * l_estcpu to less than (1 << ESTCPU_SHIFT), and therefore we can
300 	 * return zero directly.
301 	 *
302 	 * Note that our ESTCPU_MAX is actually much smaller than
303 	 * (255 << ESTCPU_SHIFT).
304 	 */
305 	if ((n << FSHIFT) >= 7 * loadfac) {
306 		return 0;
307 	}
308 
309 	while (estcpu != 0 && n > 1) {
310 		estcpu = decay_cpu(loadfac, estcpu);
311 		n--;
312 	}
313 
314 	return estcpu;
315 }
316 
317 /*
318  * sched_pstats_hook:
319  *
320  * Periodically called from sched_pstats(); used to recalculate priorities.
321  */
322 void
323 sched_pstats_hook(struct lwp *l, int batch)
324 {
325 	fixpt_t loadfac;
326 
327 	/*
328 	 * If the LWP has slept an entire second, stop recalculating
329 	 * its priority until it wakes up.
330 	 */
331 	KASSERT(lwp_locked(l, NULL));
332 	if (l->l_stat == LSSLEEP || l->l_stat == LSSTOP ||
333 	    l->l_stat == LSSUSPENDED) {
334 		if (l->l_slptime > 1) {
335 			return;
336 		}
337 	}
338 
339 	loadfac = loadfactor(averunnable.ldavg[0]);
340 	l->l_estcpu = decay_cpu(loadfac, l->l_estcpu);
341 	resetpriority(l);
342 }
343 
344 /*
345  * Recalculate the priority of an LWP after it has slept for a while.
346  */
347 static void
348 updatepri(struct lwp *l)
349 {
350 	fixpt_t loadfac;
351 
352 	KASSERT(lwp_locked(l, NULL));
353 	KASSERT(l->l_slptime > 1);
354 
355 	loadfac = loadfactor(averunnable.ldavg[0]);
356 
357 	l->l_slptime--; /* the first time was done in sched_pstats */
358 	l->l_estcpu = decay_cpu_batch(loadfac, l->l_estcpu, l->l_slptime);
359 	resetpriority(l);
360 }
361 
362 void
363 sched_rqinit(void)
364 {
365 
366 }
367 
368 void
369 sched_setrunnable(struct lwp *l)
370 {
371 
372  	if (l->l_slptime > 1)
373  		updatepri(l);
374 }
375 
376 void
377 sched_nice(struct proc *p, int n)
378 {
379 	struct lwp *l;
380 
381 	KASSERT(mutex_owned(p->p_lock));
382 
383 	p->p_nice = n;
384 	LIST_FOREACH(l, &p->p_lwps, l_sibling) {
385 		lwp_lock(l);
386 		resetpriority(l);
387 		lwp_unlock(l);
388 	}
389 }
390 
391 /*
392  * Recompute the priority of an LWP.  Arrange to reschedule if
393  * the resulting priority is better than that of the current LWP.
394  */
395 static void
396 resetpriority(struct lwp *l)
397 {
398 	pri_t pri;
399 	struct proc *p = l->l_proc;
400 
401 	KASSERT(lwp_locked(l, NULL));
402 
403 	if (l->l_class != SCHED_OTHER)
404 		return;
405 
406 	/* See comments above ESTCPU_SHIFT definition. */
407 	pri = (PRI_KERNEL - 1) - (l->l_estcpu >> ESTCPU_SHIFT) - p->p_nice;
408 	pri = imax(pri, 0);
409 	if (pri != l->l_priority)
410 		lwp_changepri(l, pri);
411 }
412 
413 /*
414  * We adjust the priority of the current LWP.  The priority of a LWP
415  * gets worse as it accumulates CPU time.  The CPU usage estimator (l_estcpu)
416  * is increased here.  The formula for computing priorities will compute a
417  * different value each time l_estcpu increases. This can cause a switch,
418  * but unless the priority crosses a PPQ boundary the actual queue will not
419  * change.  The CPU usage estimator ramps up quite quickly when the process
420  * is running (linearly), and decays away exponentially, at a rate which is
421  * proportionally slower when the system is busy.  The basic principle is
422  * that the system will 90% forget that the process used a lot of CPU time
423  * in (5 * loadavg) seconds.  This causes the system to favor processes which
424  * haven't run much recently, and to round-robin among other processes.
425  */
426 void
427 sched_schedclock(struct lwp *l)
428 {
429 
430 	if (l->l_class != SCHED_OTHER)
431 		return;
432 
433 	KASSERT(!CURCPU_IDLE_P());
434 	l->l_estcpu = ESTCPULIM(l->l_estcpu + ESTCPU_ACCUM);
435 	lwp_lock(l);
436 	resetpriority(l);
437 	lwp_unlock(l);
438 }
439 
440 /*
441  * sched_proc_fork:
442  *
443  *	Inherit the parent's scheduler history.
444  */
445 void
446 sched_proc_fork(struct proc *parent, struct proc *child)
447 {
448 	lwp_t *pl;
449 
450 	KASSERT(mutex_owned(parent->p_lock));
451 
452 	pl = LIST_FIRST(&parent->p_lwps);
453 	child->p_estcpu_inherited = pl->l_estcpu;
454 	child->p_forktime = sched_pstats_ticks;
455 }
456 
457 /*
458  * sched_proc_exit:
459  *
460  *	Chargeback parents for the sins of their children.
461  */
462 void
463 sched_proc_exit(struct proc *parent, struct proc *child)
464 {
465 	fixpt_t loadfac = loadfactor(averunnable.ldavg[0]);
466 	fixpt_t estcpu;
467 	lwp_t *pl, *cl;
468 
469 	/* XXX Only if parent != init?? */
470 
471 	mutex_enter(parent->p_lock);
472 	pl = LIST_FIRST(&parent->p_lwps);
473 	cl = LIST_FIRST(&child->p_lwps);
474 	estcpu = decay_cpu_batch(loadfac, child->p_estcpu_inherited,
475 	    sched_pstats_ticks - child->p_forktime);
476 	if (cl->l_estcpu > estcpu) {
477 		lwp_lock(pl);
478 		pl->l_estcpu = ESTCPULIM(pl->l_estcpu + cl->l_estcpu - estcpu);
479 		lwp_unlock(pl);
480 	}
481 	mutex_exit(parent->p_lock);
482 }
483 
484 void
485 sched_wakeup(struct lwp *l)
486 {
487 
488 }
489 
490 void
491 sched_slept(struct lwp *l)
492 {
493 
494 }
495 
496 void
497 sched_lwp_fork(struct lwp *l1, struct lwp *l2)
498 {
499 
500 	l2->l_estcpu = l1->l_estcpu;
501 }
502 
503 void
504 sched_lwp_collect(struct lwp *t)
505 {
506 	lwp_t *l;
507 
508 	/* Absorb estcpu value of collected LWP. */
509 	l = curlwp;
510 	lwp_lock(l);
511 	l->l_estcpu += t->l_estcpu;
512 	lwp_unlock(l);
513 }
514 
515 void
516 sched_oncpu(lwp_t *l)
517 {
518 
519 }
520 
521 void
522 sched_newts(lwp_t *l)
523 {
524 
525 }
526 
527 /*
528  * Sysctl nodes and initialization.
529  */
530 
531 static int
532 sysctl_sched_rtts(SYSCTLFN_ARGS)
533 {
534 	struct sysctlnode node;
535 	int rttsms = hztoms(sched_rrticks);
536 
537 	node = *rnode;
538 	node.sysctl_data = &rttsms;
539 	return sysctl_lookup(SYSCTLFN_CALL(&node));
540 }
541 
542 SYSCTL_SETUP(sysctl_sched_4bsd_setup, "sysctl sched setup")
543 {
544 	const struct sysctlnode *node = NULL;
545 
546 	sysctl_createv(clog, 0, NULL, &node,
547 		CTLFLAG_PERMANENT,
548 		CTLTYPE_NODE, "sched",
549 		SYSCTL_DESCR("Scheduler options"),
550 		NULL, 0, NULL, 0,
551 		CTL_KERN, CTL_CREATE, CTL_EOL);
552 
553 	if (node == NULL)
554 		return;
555 
556 	sched_rrticks = hz / 10;
557 
558 	sysctl_createv(NULL, 0, &node, NULL,
559 		CTLFLAG_PERMANENT,
560 		CTLTYPE_STRING, "name", NULL,
561 		NULL, 0, __UNCONST("4.4BSD"), 0,
562 		CTL_CREATE, CTL_EOL);
563 	sysctl_createv(NULL, 0, &node, NULL,
564 		CTLFLAG_PERMANENT,
565 		CTLTYPE_INT, "rtts",
566 		SYSCTL_DESCR("Round-robin time quantum (in milliseconds)"),
567 		sysctl_sched_rtts, 0, NULL, 0,
568 		CTL_CREATE, CTL_EOL);
569 }
570