xref: /netbsd-src/sys/kern/sched_4bsd.c (revision c9496f6b604074a9451a67df576a5b423068e71e)
1 /*	$NetBSD: sched_4bsd.c,v 1.33 2017/07/14 13:23:48 maxv Exp $	*/
2 
3 /*
4  * Copyright (c) 1999, 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, by Charles M. Hannum, Andrew Doran, and
10  * Daniel Sieger.
11  *
12  * Redistribution and use in source and binary forms, with or without
13  * modification, are permitted provided that the following conditions
14  * are met:
15  * 1. Redistributions of source code must retain the above copyright
16  *    notice, this list of conditions and the following disclaimer.
17  * 2. Redistributions in binary form must reproduce the above copyright
18  *    notice, this list of conditions and the following disclaimer in the
19  *    documentation and/or other materials provided with the distribution.
20  *
21  * THIS SOFTWARE IS PROVIDED BY THE NETBSD FOUNDATION, INC. AND CONTRIBUTORS
22  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED
23  * TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
24  * PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE FOUNDATION OR CONTRIBUTORS
25  * BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
26  * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
27  * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
28  * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
29  * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
30  * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
31  * POSSIBILITY OF SUCH DAMAGE.
32  */
33 
34 /*
35  * Copyright (c) 1982, 1986, 1990, 1991, 1993
36  *	The Regents of the University of California.  All rights reserved.
37  * (c) UNIX System Laboratories, Inc.
38  * All or some portions of this file are derived from material licensed
39  * to the University of California by American Telephone and Telegraph
40  * Co. or Unix System Laboratories, Inc. and are reproduced herein with
41  * the permission of UNIX System Laboratories, Inc.
42  *
43  * Redistribution and use in source and binary forms, with or without
44  * modification, are permitted provided that the following conditions
45  * are met:
46  * 1. Redistributions of source code must retain the above copyright
47  *    notice, this list of conditions and the following disclaimer.
48  * 2. Redistributions in binary form must reproduce the above copyright
49  *    notice, this list of conditions and the following disclaimer in the
50  *    documentation and/or other materials provided with the distribution.
51  * 3. Neither the name of the University nor the names of its contributors
52  *    may be used to endorse or promote products derived from this software
53  *    without specific prior written permission.
54  *
55  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
56  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
57  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
58  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
59  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
60  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
61  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
62  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
63  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
64  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
65  * SUCH DAMAGE.
66  *
67  *	@(#)kern_synch.c	8.9 (Berkeley) 5/19/95
68  */
69 
70 #include <sys/cdefs.h>
71 __KERNEL_RCSID(0, "$NetBSD: sched_4bsd.c,v 1.33 2017/07/14 13:23:48 maxv Exp $");
72 
73 #include "opt_ddb.h"
74 #include "opt_lockdebug.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/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 
89 static void updatepri(struct lwp *);
90 static void resetpriority(struct lwp *);
91 
92 extern unsigned int sched_pstats_ticks; /* defined in kern_synch.c */
93 
94 /* Number of hardclock ticks per sched_tick() */
95 static int rrticks __read_mostly;
96 
97 /*
98  * Force switch among equal priority processes every 100ms.
99  * Called from hardclock every hz/10 == rrticks hardclock ticks.
100  *
101  * There's no need to lock anywhere in this routine, as it's
102  * CPU-local and runs at IPL_SCHED (called from clock interrupt).
103  */
104 /* ARGSUSED */
105 void
106 sched_tick(struct cpu_info *ci)
107 {
108 	struct schedstate_percpu *spc = &ci->ci_schedstate;
109 	lwp_t *l;
110 
111 	spc->spc_ticks = rrticks;
112 
113 	if (CURCPU_IDLE_P()) {
114 		cpu_need_resched(ci, 0);
115 		return;
116 	}
117 	l = ci->ci_data.cpu_onproc;
118 	if (l == NULL) {
119 		return;
120 	}
121 	switch (l->l_class) {
122 	case SCHED_FIFO:
123 		/* No timeslicing for FIFO jobs. */
124 		break;
125 	case SCHED_RR:
126 		/* Force it into mi_switch() to look for other jobs to run. */
127 		cpu_need_resched(ci, RESCHED_KPREEMPT);
128 		break;
129 	default:
130 		if (spc->spc_flags & SPCF_SHOULDYIELD) {
131 			/*
132 			 * Process is stuck in kernel somewhere, probably
133 			 * due to buggy or inefficient code.  Force a
134 			 * kernel preemption.
135 			 */
136 			cpu_need_resched(ci, RESCHED_KPREEMPT);
137 		} else if (spc->spc_flags & SPCF_SEENRR) {
138 			/*
139 			 * The process has already been through a roundrobin
140 			 * without switching and may be hogging the CPU.
141 			 * Indicate that the process should yield.
142 			 */
143 			spc->spc_flags |= SPCF_SHOULDYIELD;
144 			cpu_need_resched(ci, 0);
145 		} else {
146 			spc->spc_flags |= SPCF_SEENRR;
147 		}
148 		break;
149 	}
150 }
151 
152 /*
153  * Why PRIO_MAX - 2? From setpriority(2):
154  *
155  *	prio is a value in the range -20 to 20.  The default priority is
156  *	0; lower priorities cause more favorable scheduling.  A value of
157  *	19 or 20 will schedule a process only when nothing at priority <=
158  *	0 is runnable.
159  *
160  * This gives estcpu influence over 18 priority levels, and leaves nice
161  * with 40 levels.  One way to think about it is that nice has 20 levels
162  * either side of estcpu's 18.
163  */
164 #define	ESTCPU_SHIFT	11
165 #define	ESTCPU_MAX	((PRIO_MAX - 2) << ESTCPU_SHIFT)
166 #define	ESTCPU_ACCUM	(1 << (ESTCPU_SHIFT - 1))
167 #define	ESTCPULIM(e)	min((e), ESTCPU_MAX)
168 
169 /*
170  * The main parameter used by this algorithm is 'l_estcpu'. It is an estimate
171  * of the recent CPU utilization of the thread.
172  *
173  * l_estcpu is:
174  *  - increased each time the hardclock ticks and the thread is found to
175  *    be executing, in sched_schedclock() called from hardclock()
176  *  - decreased (filtered) on each sched tick, in sched_pstats_hook()
177  * If the lwp is sleeping for more than a second, we don't touch l_estcpu: it
178  * will be updated in sched_setrunnable() when the lwp wakes up, in burst mode
179  * (ie, we decrease it n times).
180  *
181  * Note that hardclock updates l_estcpu and l_cpticks independently.
182  *
183  * -----------------------------------------------------------------------------
184  *
185  * Here we describe how l_estcpu is decreased.
186  *
187  * Constants for digital decay (filter):
188  *     90% of l_estcpu usage in (5 * loadavg) seconds
189  *
190  * We wish to decay away 90% of l_estcpu in (5 * loadavg) seconds. That is, we
191  * want to compute a value of decay such that the following loop:
192  *     for (i = 0; i < (5 * loadavg); i++)
193  *         l_estcpu *= decay;
194  * will result in
195  *     l_estcpu *= 0.1;
196  * for all values of loadavg.
197  *
198  * Mathematically this loop can be expressed by saying:
199  *     decay ** (5 * loadavg) ~= .1
200  *
201  * And finally, the corresponding value of decay we're using is:
202  *     decay = (2 * loadavg) / (2 * loadavg + 1)
203  *
204  * -----------------------------------------------------------------------------
205  *
206  * Now, let's prove that the value of decay stated above will always fulfill
207  * the equation:
208  *     decay ** (5 * loadavg) ~= .1
209  *
210  * If we compute b as:
211  *     b = 2 * loadavg
212  * then
213  *     decay = b / (b + 1)
214  *
215  * We now need to prove two things:
216  *     1) Given [factor ** (5 * loadavg) =~ .1], prove [factor == b/(b+1)].
217  *     2) Given [b/(b+1) ** power =~ .1], prove [power == (5 * loadavg)].
218  *
219  * Facts:
220  *   * For x real: exp(x) = 0! + x**1/1! + x**2/2! + ...
221  *     Therefore, for x close to zero, exp(x) =~ 1 + x.
222  *     In turn, for b large enough, exp(-1/b) =~ 1 - (1/b) = (b-1)/b.
223  *
224  *   * For b large enough, (b-1)/b =~ b/(b+1).
225  *
226  *   * For x belonging to [-1;1[, ln(1-x) = - x - x**2/2 - x**3/3 - ...
227  *     Therefore ln(b/(b+1)) = ln(1 - 1/(b+1)) =~ -1/(b+1).
228  *
229  *   * ln(0.1) =~ -2.30
230  *
231  * Proof of (1):
232  *     factor ** (5 * loadavg) =~ 0.1
233  *  => ln(factor) =~ -2.30 / (5 * loadavg)
234  *  => factor =~ exp(-1 / ((5 / 2.30) * loadavg))
235  *            =~ exp(-1 / (2 * loadavg))
236  *            =~ exp(-1 / b)
237  *            =~ (b - 1) / b
238  *            =~ b / (b + 1)
239  *            =~ (2 * loadavg) / ((2 * loadavg) + 1)
240  *
241  * Proof of (2):
242  *     (b / (b + 1)) ** power =~ .1
243  *  => power * ln(b / (b + 1)) =~ -2.30
244  *  => power * (-1 / (b + 1)) =~ -2.30
245  *  => power =~ 2.30 * (b + 1)
246  *  => power =~ 4.60 * loadavg + 2.30
247  *  => power =~ 5 * loadavg
248  *
249  * Conclusion: decay = (2 * loadavg) / (2 * loadavg + 1)
250  */
251 
252 /* See calculations above */
253 #define	loadfactor(loadavg)  (2 * (loadavg))
254 
255 static fixpt_t
256 decay_cpu(fixpt_t loadfac, fixpt_t estcpu)
257 {
258 
259 	if (estcpu == 0) {
260 		return 0;
261 	}
262 
263 #if !defined(_LP64)
264 	/* avoid 64bit arithmetics. */
265 #define	FIXPT_MAX ((fixpt_t)((UINTMAX_C(1) << sizeof(fixpt_t) * CHAR_BIT) - 1))
266 	if (__predict_true(loadfac <= FIXPT_MAX / ESTCPU_MAX)) {
267 		return estcpu * loadfac / (loadfac + FSCALE);
268 	}
269 #endif
270 
271 	return (uint64_t)estcpu * loadfac / (loadfac + FSCALE);
272 }
273 
274 static fixpt_t
275 decay_cpu_batch(fixpt_t loadfac, fixpt_t estcpu, unsigned int n)
276 {
277 
278 	/*
279 	 * For all load averages >= 1 and max l_estcpu of (255 << ESTCPU_SHIFT),
280 	 * if we slept for at least seven times the loadfactor, we will decay
281 	 * l_estcpu to less than (1 << ESTCPU_SHIFT), and therefore we can
282 	 * return zero directly.
283 	 *
284 	 * Note that our ESTCPU_MAX is actually much smaller than
285 	 * (255 << ESTCPU_SHIFT).
286 	 */
287 	if ((n << FSHIFT) >= 7 * loadfac) {
288 		return 0;
289 	}
290 
291 	while (estcpu != 0 && n > 1) {
292 		estcpu = decay_cpu(loadfac, estcpu);
293 		n--;
294 	}
295 
296 	return estcpu;
297 }
298 
299 /*
300  * sched_pstats_hook:
301  *
302  * Periodically called from sched_pstats(); used to recalculate priorities.
303  */
304 void
305 sched_pstats_hook(struct lwp *l, int batch)
306 {
307 	fixpt_t loadfac;
308 
309 	/*
310 	 * If the LWP has slept an entire second, stop recalculating
311 	 * its priority until it wakes up.
312 	 */
313 	KASSERT(lwp_locked(l, NULL));
314 	if (l->l_stat == LSSLEEP || l->l_stat == LSSTOP ||
315 	    l->l_stat == LSSUSPENDED) {
316 		if (l->l_slptime > 1) {
317 			return;
318 		}
319 	}
320 
321 	loadfac = loadfactor(averunnable.ldavg[0]);
322 	l->l_estcpu = decay_cpu(loadfac, l->l_estcpu);
323 	resetpriority(l);
324 }
325 
326 /*
327  * Recalculate the priority of an LWP after it has slept for a while.
328  */
329 static void
330 updatepri(struct lwp *l)
331 {
332 	fixpt_t loadfac;
333 
334 	KASSERT(lwp_locked(l, NULL));
335 	KASSERT(l->l_slptime > 1);
336 
337 	loadfac = loadfactor(averunnable.ldavg[0]);
338 
339 	l->l_slptime--; /* the first time was done in sched_pstats */
340 	l->l_estcpu = decay_cpu_batch(loadfac, l->l_estcpu, l->l_slptime);
341 	resetpriority(l);
342 }
343 
344 void
345 sched_rqinit(void)
346 {
347 
348 }
349 
350 void
351 sched_setrunnable(struct lwp *l)
352 {
353 
354  	if (l->l_slptime > 1)
355  		updatepri(l);
356 }
357 
358 void
359 sched_nice(struct proc *p, int n)
360 {
361 	struct lwp *l;
362 
363 	KASSERT(mutex_owned(p->p_lock));
364 
365 	p->p_nice = n;
366 	LIST_FOREACH(l, &p->p_lwps, l_sibling) {
367 		lwp_lock(l);
368 		resetpriority(l);
369 		lwp_unlock(l);
370 	}
371 }
372 
373 /*
374  * Recompute the priority of an LWP.  Arrange to reschedule if
375  * the resulting priority is better than that of the current LWP.
376  */
377 static void
378 resetpriority(struct lwp *l)
379 {
380 	pri_t pri;
381 	struct proc *p = l->l_proc;
382 
383 	KASSERT(lwp_locked(l, NULL));
384 
385 	if (l->l_class != SCHED_OTHER)
386 		return;
387 
388 	/* See comments above ESTCPU_SHIFT definition. */
389 	pri = (PRI_KERNEL - 1) - (l->l_estcpu >> ESTCPU_SHIFT) - p->p_nice;
390 	pri = imax(pri, 0);
391 	if (pri != l->l_priority)
392 		lwp_changepri(l, pri);
393 }
394 
395 /*
396  * We adjust the priority of the current LWP.  The priority of a LWP
397  * gets worse as it accumulates CPU time.  The CPU usage estimator (l_estcpu)
398  * is increased here.  The formula for computing priorities will compute a
399  * different value each time l_estcpu increases. This can cause a switch,
400  * but unless the priority crosses a PPQ boundary the actual queue will not
401  * change.  The CPU usage estimator ramps up quite quickly when the process
402  * is running (linearly), and decays away exponentially, at a rate which is
403  * proportionally slower when the system is busy.  The basic principle is
404  * that the system will 90% forget that the process used a lot of CPU time
405  * in (5 * loadavg) seconds.  This causes the system to favor processes which
406  * haven't run much recently, and to round-robin among other processes.
407  */
408 void
409 sched_schedclock(struct lwp *l)
410 {
411 
412 	if (l->l_class != SCHED_OTHER)
413 		return;
414 
415 	KASSERT(!CURCPU_IDLE_P());
416 	l->l_estcpu = ESTCPULIM(l->l_estcpu + ESTCPU_ACCUM);
417 	lwp_lock(l);
418 	resetpriority(l);
419 	lwp_unlock(l);
420 }
421 
422 /*
423  * sched_proc_fork:
424  *
425  *	Inherit the parent's scheduler history.
426  */
427 void
428 sched_proc_fork(struct proc *parent, struct proc *child)
429 {
430 	lwp_t *pl;
431 
432 	KASSERT(mutex_owned(parent->p_lock));
433 
434 	pl = LIST_FIRST(&parent->p_lwps);
435 	child->p_estcpu_inherited = pl->l_estcpu;
436 	child->p_forktime = sched_pstats_ticks;
437 }
438 
439 /*
440  * sched_proc_exit:
441  *
442  *	Chargeback parents for the sins of their children.
443  */
444 void
445 sched_proc_exit(struct proc *parent, struct proc *child)
446 {
447 	fixpt_t loadfac = loadfactor(averunnable.ldavg[0]);
448 	fixpt_t estcpu;
449 	lwp_t *pl, *cl;
450 
451 	/* XXX Only if parent != init?? */
452 
453 	mutex_enter(parent->p_lock);
454 	pl = LIST_FIRST(&parent->p_lwps);
455 	cl = LIST_FIRST(&child->p_lwps);
456 	estcpu = decay_cpu_batch(loadfac, child->p_estcpu_inherited,
457 	    sched_pstats_ticks - child->p_forktime);
458 	if (cl->l_estcpu > estcpu) {
459 		lwp_lock(pl);
460 		pl->l_estcpu = ESTCPULIM(pl->l_estcpu + cl->l_estcpu - estcpu);
461 		lwp_unlock(pl);
462 	}
463 	mutex_exit(parent->p_lock);
464 }
465 
466 void
467 sched_wakeup(struct lwp *l)
468 {
469 
470 }
471 
472 void
473 sched_slept(struct lwp *l)
474 {
475 
476 }
477 
478 void
479 sched_lwp_fork(struct lwp *l1, struct lwp *l2)
480 {
481 
482 	l2->l_estcpu = l1->l_estcpu;
483 }
484 
485 void
486 sched_lwp_collect(struct lwp *t)
487 {
488 	lwp_t *l;
489 
490 	/* Absorb estcpu value of collected LWP. */
491 	l = curlwp;
492 	lwp_lock(l);
493 	l->l_estcpu += t->l_estcpu;
494 	lwp_unlock(l);
495 }
496 
497 void
498 sched_oncpu(lwp_t *l)
499 {
500 
501 }
502 
503 void
504 sched_newts(lwp_t *l)
505 {
506 
507 }
508 
509 /*
510  * Sysctl nodes and initialization.
511  */
512 
513 static int
514 sysctl_sched_rtts(SYSCTLFN_ARGS)
515 {
516 	struct sysctlnode node;
517 	int rttsms = hztoms(rrticks);
518 
519 	node = *rnode;
520 	node.sysctl_data = &rttsms;
521 	return sysctl_lookup(SYSCTLFN_CALL(&node));
522 }
523 
524 SYSCTL_SETUP(sysctl_sched_4bsd_setup, "sysctl sched setup")
525 {
526 	const struct sysctlnode *node = NULL;
527 
528 	sysctl_createv(clog, 0, NULL, &node,
529 		CTLFLAG_PERMANENT,
530 		CTLTYPE_NODE, "sched",
531 		SYSCTL_DESCR("Scheduler options"),
532 		NULL, 0, NULL, 0,
533 		CTL_KERN, CTL_CREATE, CTL_EOL);
534 
535 	if (node == NULL)
536 		return;
537 
538 	rrticks = hz / 10;
539 
540 	sysctl_createv(NULL, 0, &node, NULL,
541 		CTLFLAG_PERMANENT,
542 		CTLTYPE_STRING, "name", NULL,
543 		NULL, 0, __UNCONST("4.4BSD"), 0,
544 		CTL_CREATE, CTL_EOL);
545 	sysctl_createv(NULL, 0, &node, NULL,
546 		CTLFLAG_PERMANENT,
547 		CTLTYPE_INT, "rtts",
548 		SYSCTL_DESCR("Round-robin time quantum (in milliseconds)"),
549 		sysctl_sched_rtts, 0, NULL, 0,
550 		CTL_CREATE, CTL_EOL);
551 }
552