xref: /dflybsd-src/sys/kern/kern_synch.c (revision 9bb2a92deb77a8c17f5fcc2740d0e3cc0e1c6c84)
1 /*-
2  * Copyright (c) 1982, 1986, 1990, 1991, 1993
3  *	The Regents of the University of California.  All rights reserved.
4  * (c) UNIX System Laboratories, Inc.
5  * All or some portions of this file are derived from material licensed
6  * to the University of California by American Telephone and Telegraph
7  * Co. or Unix System Laboratories, Inc. and are reproduced herein with
8  * the permission of UNIX System Laboratories, Inc.
9  *
10  * Redistribution and use in source and binary forms, with or without
11  * modification, are permitted provided that the following conditions
12  * are met:
13  * 1. Redistributions of source code must retain the above copyright
14  *    notice, this list of conditions and the following disclaimer.
15  * 2. Redistributions in binary form must reproduce the above copyright
16  *    notice, this list of conditions and the following disclaimer in the
17  *    documentation and/or other materials provided with the distribution.
18  * 3. All advertising materials mentioning features or use of this software
19  *    must display the following acknowledgement:
20  *	This product includes software developed by the University of
21  *	California, Berkeley and its contributors.
22  * 4. Neither the name of the University nor the names of its contributors
23  *    may be used to endorse or promote products derived from this software
24  *    without specific prior written permission.
25  *
26  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
27  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
28  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
29  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
30  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
31  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
32  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
33  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
34  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
35  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
36  * SUCH DAMAGE.
37  *
38  *	@(#)kern_synch.c	8.9 (Berkeley) 5/19/95
39  * $FreeBSD: src/sys/kern/kern_synch.c,v 1.87.2.6 2002/10/13 07:29:53 kbyanc Exp $
40  * $DragonFly: src/sys/kern/kern_synch.c,v 1.29 2004/03/08 03:05:27 dillon Exp $
41  */
42 
43 #include "opt_ktrace.h"
44 
45 #include <sys/param.h>
46 #include <sys/systm.h>
47 #include <sys/proc.h>
48 #include <sys/kernel.h>
49 #include <sys/signalvar.h>
50 #include <sys/resourcevar.h>
51 #include <sys/vmmeter.h>
52 #include <sys/sysctl.h>
53 #include <sys/thread2.h>
54 #ifdef KTRACE
55 #include <sys/uio.h>
56 #include <sys/ktrace.h>
57 #endif
58 #include <sys/xwait.h>
59 
60 #include <machine/cpu.h>
61 #include <machine/ipl.h>
62 #include <machine/smp.h>
63 
64 static void sched_setup (void *dummy);
65 SYSINIT(sched_setup, SI_SUB_KICK_SCHEDULER, SI_ORDER_FIRST, sched_setup, NULL)
66 
67 int	hogticks;
68 int	lbolt;
69 int	sched_quantum;		/* Roundrobin scheduling quantum in ticks. */
70 int	ncpus;
71 int	ncpus2, ncpus2_shift, ncpus2_mask;
72 
73 static struct callout loadav_callout;
74 
75 struct loadavg averunnable =
76 	{ {0, 0, 0}, FSCALE };	/* load average, of runnable procs */
77 /*
78  * Constants for averages over 1, 5, and 15 minutes
79  * when sampling at 5 second intervals.
80  */
81 static fixpt_t cexp[3] = {
82 	0.9200444146293232 * FSCALE,	/* exp(-1/12) */
83 	0.9834714538216174 * FSCALE,	/* exp(-1/60) */
84 	0.9944598480048967 * FSCALE,	/* exp(-1/180) */
85 };
86 
87 static void	endtsleep (void *);
88 static void	loadav (void *arg);
89 static void	roundrobin (void *arg);
90 static void	schedcpu (void *arg);
91 static void	updatepri (struct proc *p);
92 static void	crit_panicints(void);
93 
94 static int
95 sysctl_kern_quantum(SYSCTL_HANDLER_ARGS)
96 {
97 	int error, new_val;
98 
99 	new_val = sched_quantum * tick;
100 	error = sysctl_handle_int(oidp, &new_val, 0, req);
101         if (error != 0 || req->newptr == NULL)
102 		return (error);
103 	if (new_val < tick)
104 		return (EINVAL);
105 	sched_quantum = new_val / tick;
106 	hogticks = 2 * sched_quantum;
107 	return (0);
108 }
109 
110 SYSCTL_PROC(_kern, OID_AUTO, quantum, CTLTYPE_INT|CTLFLAG_RW,
111 	0, sizeof sched_quantum, sysctl_kern_quantum, "I", "");
112 
113 int
114 roundrobin_interval(void)
115 {
116 	return (sched_quantum);
117 }
118 
119 /*
120  * Force switch among equal priority processes every 100ms.
121  *
122  * WARNING!  The MP lock is not held on ipi message remotes.
123  */
124 #ifdef SMP
125 
126 static void
127 roundrobin_remote(void *arg)
128 {
129 	struct proc *p = lwkt_preempted_proc();
130  	if (p == NULL || RTP_PRIO_NEED_RR(p->p_rtprio.type))
131 		need_resched();
132 }
133 
134 #endif
135 
136 static void
137 roundrobin(void *arg)
138 {
139 	struct proc *p = lwkt_preempted_proc();
140  	if (p == NULL || RTP_PRIO_NEED_RR(p->p_rtprio.type))
141 		need_resched();
142 #ifdef SMP
143 	lwkt_send_ipiq_mask(mycpu->gd_other_cpus, roundrobin_remote, NULL);
144 #endif
145  	timeout(roundrobin, NULL, sched_quantum);
146 }
147 
148 #ifdef SMP
149 
150 void
151 resched_cpus(u_int32_t mask)
152 {
153 	lwkt_send_ipiq_mask(mask, roundrobin_remote, NULL);
154 }
155 
156 #endif
157 
158 /*
159  * Constants for digital decay and forget:
160  *	90% of (p_estcpu) usage in 5 * loadav time
161  *	95% of (p_pctcpu) usage in 60 seconds (load insensitive)
162  *          Note that, as ps(1) mentions, this can let percentages
163  *          total over 100% (I've seen 137.9% for 3 processes).
164  *
165  * Note that schedulerclock() updates p_estcpu and p_cpticks asynchronously.
166  *
167  * We wish to decay away 90% of p_estcpu in (5 * loadavg) seconds.
168  * That is, the system wants to compute a value of decay such
169  * that the following for loop:
170  * 	for (i = 0; i < (5 * loadavg); i++)
171  * 		p_estcpu *= decay;
172  * will compute
173  * 	p_estcpu *= 0.1;
174  * for all values of loadavg:
175  *
176  * Mathematically this loop can be expressed by saying:
177  * 	decay ** (5 * loadavg) ~= .1
178  *
179  * The system computes decay as:
180  * 	decay = (2 * loadavg) / (2 * loadavg + 1)
181  *
182  * We wish to prove that the system's computation of decay
183  * will always fulfill the equation:
184  * 	decay ** (5 * loadavg) ~= .1
185  *
186  * If we compute b as:
187  * 	b = 2 * loadavg
188  * then
189  * 	decay = b / (b + 1)
190  *
191  * We now need to prove two things:
192  *	1) Given factor ** (5 * loadavg) ~= .1, prove factor == b/(b+1)
193  *	2) Given b/(b+1) ** power ~= .1, prove power == (5 * loadavg)
194  *
195  * Facts:
196  *         For x close to zero, exp(x) =~ 1 + x, since
197  *              exp(x) = 0! + x**1/1! + x**2/2! + ... .
198  *              therefore exp(-1/b) =~ 1 - (1/b) = (b-1)/b.
199  *         For x close to zero, ln(1+x) =~ x, since
200  *              ln(1+x) = x - x**2/2 + x**3/3 - ...     -1 < x < 1
201  *              therefore ln(b/(b+1)) = ln(1 - 1/(b+1)) =~ -1/(b+1).
202  *         ln(.1) =~ -2.30
203  *
204  * Proof of (1):
205  *    Solve (factor)**(power) =~ .1 given power (5*loadav):
206  *	solving for factor,
207  *      ln(factor) =~ (-2.30/5*loadav), or
208  *      factor =~ exp(-1/((5/2.30)*loadav)) =~ exp(-1/(2*loadav)) =
209  *          exp(-1/b) =~ (b-1)/b =~ b/(b+1).                    QED
210  *
211  * Proof of (2):
212  *    Solve (factor)**(power) =~ .1 given factor == (b/(b+1)):
213  *	solving for power,
214  *      power*ln(b/(b+1)) =~ -2.30, or
215  *      power =~ 2.3 * (b + 1) = 4.6*loadav + 2.3 =~ 5*loadav.  QED
216  *
217  * Actual power values for the implemented algorithm are as follows:
218  *      loadav: 1       2       3       4
219  *      power:  5.68    10.32   14.94   19.55
220  */
221 
222 /* calculations for digital decay to forget 90% of usage in 5*loadav sec */
223 #define	loadfactor(loadav)	(2 * (loadav))
224 #define	decay_cpu(loadfac, cpu)	(((loadfac) * (cpu)) / ((loadfac) + FSCALE))
225 
226 /* decay 95% of `p_pctcpu' in 60 seconds; see CCPU_SHIFT before changing */
227 static fixpt_t	ccpu = 0.95122942450071400909 * FSCALE;	/* exp(-1/20) */
228 SYSCTL_INT(_kern, OID_AUTO, ccpu, CTLFLAG_RD, &ccpu, 0, "");
229 
230 /* kernel uses `FSCALE', userland (SHOULD) use kern.fscale */
231 static int	fscale __unused = FSCALE;
232 SYSCTL_INT(_kern, OID_AUTO, fscale, CTLFLAG_RD, 0, FSCALE, "");
233 
234 /*
235  * If `ccpu' is not equal to `exp(-1/20)' and you still want to use the
236  * faster/more-accurate formula, you'll have to estimate CCPU_SHIFT below
237  * and possibly adjust FSHIFT in "param.h" so that (FSHIFT >= CCPU_SHIFT).
238  *
239  * To estimate CCPU_SHIFT for exp(-1/20), the following formula was used:
240  *	1 - exp(-1/20) ~= 0.0487 ~= 0.0488 == 1 (fixed pt, *11* bits).
241  *
242  * If you don't want to bother with the faster/more-accurate formula, you
243  * can set CCPU_SHIFT to (FSHIFT + 1) which will use a slower/less-accurate
244  * (more general) method of calculating the %age of CPU used by a process.
245  */
246 #define	CCPU_SHIFT	11
247 
248 /*
249  * Recompute process priorities, every hz ticks.
250  */
251 /* ARGSUSED */
252 static void
253 schedcpu(void *arg)
254 {
255 	fixpt_t loadfac = loadfactor(averunnable.ldavg[0]);
256 	struct proc *p;
257 	int realstathz, s;
258 
259 	realstathz = stathz ? stathz : hz;
260 	FOREACH_PROC_IN_SYSTEM(p) {
261 		/*
262 		 * Increment time in/out of memory and sleep time
263 		 * (if sleeping).  We ignore overflow; with 16-bit int's
264 		 * (remember them?) overflow takes 45 days.
265 		 */
266 		p->p_swtime++;
267 		if (p->p_stat == SSLEEP || p->p_stat == SSTOP)
268 			p->p_slptime++;
269 		p->p_pctcpu = (p->p_pctcpu * ccpu) >> FSHIFT;
270 		/*
271 		 * If the process has slept the entire second,
272 		 * stop recalculating its priority until it wakes up.
273 		 */
274 		if (p->p_slptime > 1)
275 			continue;
276 		s = splhigh();	/* prevent state changes and protect run queue */
277 		/*
278 		 * p_pctcpu is only for ps.
279 		 */
280 #if	(FSHIFT >= CCPU_SHIFT)
281 		p->p_pctcpu += (realstathz == 100)?
282 			((fixpt_t) p->p_cpticks) << (FSHIFT - CCPU_SHIFT):
283                 	100 * (((fixpt_t) p->p_cpticks)
284 				<< (FSHIFT - CCPU_SHIFT)) / realstathz;
285 #else
286 		p->p_pctcpu += ((FSCALE - ccpu) *
287 			(p->p_cpticks * FSCALE / realstathz)) >> FSHIFT;
288 #endif
289 		p->p_cpticks = 0;
290 		p->p_estcpu = decay_cpu(loadfac, p->p_estcpu);
291 		resetpriority(p);
292 		splx(s);
293 	}
294 	wakeup((caddr_t)&lbolt);
295 	timeout(schedcpu, (void *)0, hz);
296 }
297 
298 /*
299  * Recalculate the priority of a process after it has slept for a while.
300  * For all load averages >= 1 and max p_estcpu of 255, sleeping for at
301  * least six times the loadfactor will decay p_estcpu to zero.
302  */
303 static void
304 updatepri(struct proc *p)
305 {
306 	unsigned int newcpu = p->p_estcpu;
307 	fixpt_t loadfac = loadfactor(averunnable.ldavg[0]);
308 
309 	if (p->p_slptime > 5 * loadfac) {
310 		p->p_estcpu = 0;
311 	} else {
312 		p->p_slptime--;	/* the first time was done in schedcpu */
313 		while (newcpu && --p->p_slptime)
314 			newcpu = decay_cpu(loadfac, newcpu);
315 		p->p_estcpu = newcpu;
316 	}
317 	resetpriority(p);
318 }
319 
320 /*
321  * We're only looking at 7 bits of the address; everything is
322  * aligned to 4, lots of things are aligned to greater powers
323  * of 2.  Shift right by 8, i.e. drop the bottom 256 worth.
324  */
325 #define TABLESIZE	128
326 static TAILQ_HEAD(slpquehead, thread) slpque[TABLESIZE];
327 #define LOOKUP(x)	(((intptr_t)(x) >> 8) & (TABLESIZE - 1))
328 
329 /*
330  * During autoconfiguration or after a panic, a sleep will simply
331  * lower the priority briefly to allow interrupts, then return.
332  * The priority to be used (safepri) is machine-dependent, thus this
333  * value is initialized and maintained in the machine-dependent layers.
334  * This priority will typically be 0, or the lowest priority
335  * that is safe for use on the interrupt stack; it can be made
336  * higher to block network software interrupts after panics.
337  */
338 int safepri;
339 
340 void
341 sleepinit(void)
342 {
343 	int i;
344 
345 	sched_quantum = hz/10;
346 	hogticks = 2 * sched_quantum;
347 	for (i = 0; i < TABLESIZE; i++)
348 		TAILQ_INIT(&slpque[i]);
349 }
350 
351 /*
352  * General sleep call.  Suspends the current process until a wakeup is
353  * performed on the specified identifier.  The process will then be made
354  * runnable with the specified priority.  Sleeps at most timo/hz seconds
355  * (0 means no timeout).  If flags includes PCATCH flag, signals are checked
356  * before and after sleeping, else signals are not checked.  Returns 0 if
357  * awakened, EWOULDBLOCK if the timeout expires.  If PCATCH is set and a
358  * signal needs to be delivered, ERESTART is returned if the current system
359  * call should be restarted if possible, and EINTR is returned if the system
360  * call should be interrupted by the signal (return EINTR).
361  *
362  * If the process has P_CURPROC set mi_switch() will not re-queue it to
363  * the userland scheduler queues because we are in a SSLEEP state.  If
364  * we are not the current process then we have to remove ourselves from
365  * the scheduler queues.
366  *
367  * YYY priority now unused
368  */
369 int
370 tsleep(void *ident, int flags, const char *wmesg, int timo)
371 {
372 	struct thread *td = curthread;
373 	struct proc *p = td->td_proc;		/* may be NULL */
374 	int s, sig = 0, catch = flags & PCATCH;
375 	int id = LOOKUP(ident);
376 	struct callout_handle thandle;
377 
378 	/*
379 	 * NOTE: removed KTRPOINT, it could cause races due to blocking
380 	 * even in stable.  Just scrap it for now.
381 	 */
382 	if (cold || panicstr) {
383 		/*
384 		 * After a panic, or during autoconfiguration,
385 		 * just give interrupts a chance, then just return;
386 		 * don't run any other procs or panic below,
387 		 * in case this is the idle process and already asleep.
388 		 */
389 		crit_panicints();
390 		return (0);
391 	}
392 	KKASSERT(td != &mycpu->gd_idlethread);	/* you must be kidding! */
393 	s = splhigh();
394 	KASSERT(ident != NULL, ("tsleep: no ident"));
395 	KASSERT(p == NULL || p->p_stat == SRUN, ("tsleep %p %s %d",
396 		ident, wmesg, p->p_stat));
397 
398 	crit_enter();
399 	td->td_wchan = ident;
400 	td->td_wmesg = wmesg;
401 	if (p)
402 		p->p_slptime = 0;
403 	lwkt_deschedule_self();
404 	TAILQ_INSERT_TAIL(&slpque[id], td, td_threadq);
405 	if (timo)
406 		thandle = timeout(endtsleep, (void *)td, timo);
407 	/*
408 	 * We put ourselves on the sleep queue and start our timeout
409 	 * before calling CURSIG, as we could stop there, and a wakeup
410 	 * or a SIGCONT (or both) could occur while we were stopped.
411 	 * A SIGCONT would cause us to be marked as SSLEEP
412 	 * without resuming us, thus we must be ready for sleep
413 	 * when CURSIG is called.  If the wakeup happens while we're
414 	 * stopped, td->td_wchan will be 0 upon return from CURSIG.
415 	 */
416 	if (p) {
417 		if (catch) {
418 			p->p_flag |= P_SINTR;
419 			if ((sig = CURSIG(p))) {
420 				if (td->td_wchan) {
421 					unsleep(td);
422 					lwkt_schedule_self();
423 				}
424 				p->p_stat = SRUN;
425 				goto resume;
426 			}
427 			if (td->td_wchan == NULL) {
428 				catch = 0;
429 				goto resume;
430 			}
431 		} else {
432 			sig = 0;
433 		}
434 
435 		/*
436 		 * If we are not the current process we have to remove ourself
437 		 * from the run queue.
438 		 */
439 		KASSERT(p->p_stat == SRUN, ("PSTAT NOT SRUN %d %d", p->p_pid, p->p_stat));
440 		/*
441 		 * If this is the current 'user' process schedule another one.
442 		 */
443 		clrrunnable(p, SSLEEP);
444 		p->p_stats->p_ru.ru_nvcsw++;
445 		KKASSERT(td->td_release || (p->p_flag & P_CURPROC) == 0);
446 		mi_switch();
447 		KASSERT(p->p_stat == SRUN, ("tsleep: stat not srun"));
448 	} else {
449 		lwkt_switch();
450 	}
451 resume:
452 	crit_exit();
453 	if (p)
454 		p->p_flag &= ~P_SINTR;
455 	splx(s);
456 	if (td->td_flags & TDF_TIMEOUT) {
457 		td->td_flags &= ~TDF_TIMEOUT;
458 		if (sig == 0)
459 			return (EWOULDBLOCK);
460 	} else if (timo) {
461 		untimeout(endtsleep, (void *)td, thandle);
462 	} else if (td->td_wmesg) {
463 		/*
464 		 * This can happen if a thread is woken up directly.  Clear
465 		 * wmesg to avoid debugging confusion.
466 		 */
467 		td->td_wmesg = NULL;
468 	}
469 	/* inline of iscaught() */
470 	if (p) {
471 		if (catch && (sig != 0 || (sig = CURSIG(p)))) {
472 			if (SIGISMEMBER(p->p_sigacts->ps_sigintr, sig))
473 				return (EINTR);
474 			return (ERESTART);
475 		}
476 	}
477 	return (0);
478 }
479 
480 /*
481  * Implement the timeout for tsleep.  We interlock against
482  * wchan when setting TDF_TIMEOUT.  For processes we remove
483  * the sleep if the process is stopped rather then sleeping,
484  * so it remains stopped.
485  */
486 static void
487 endtsleep(void *arg)
488 {
489 	thread_t td = arg;
490 	struct proc *p;
491 	int s;
492 
493 	s = splhigh();
494 	if (td->td_wchan) {
495 		td->td_flags |= TDF_TIMEOUT;
496 		if ((p = td->td_proc) != NULL) {
497 			if (p->p_stat == SSLEEP)
498 				setrunnable(p);
499 			else
500 				unsleep(td);
501 		} else {
502 			unsleep(td);
503 			lwkt_schedule(td);
504 		}
505 	}
506 	splx(s);
507 }
508 
509 /*
510  * Remove a process from its wait queue
511  */
512 void
513 unsleep(struct thread *td)
514 {
515 	int s;
516 
517 	s = splhigh();
518 	if (td->td_wchan) {
519 #if 0
520 		if (p->p_flag & P_XSLEEP) {
521 			struct xwait *w = p->p_wchan;
522 			TAILQ_REMOVE(&w->waitq, p, p_procq);
523 			p->p_flag &= ~P_XSLEEP;
524 		} else
525 #endif
526 		TAILQ_REMOVE(&slpque[LOOKUP(td->td_wchan)], td, td_threadq);
527 		td->td_wchan = NULL;
528 	}
529 	splx(s);
530 }
531 
532 #if 0
533 /*
534  * Make all processes sleeping on the explicit lock structure runnable.
535  */
536 void
537 xwakeup(struct xwait *w)
538 {
539 	struct proc *p;
540 	int s;
541 
542 	s = splhigh();
543 	++w->gen;
544 	while ((p = TAILQ_FIRST(&w->waitq)) != NULL) {
545 		TAILQ_REMOVE(&w->waitq, p, p_procq);
546 		KASSERT(p->p_wchan == w && (p->p_flag & P_XSLEEP),
547 		    ("xwakeup: wchan mismatch for %p (%p/%p) %08x", p, p->p_wchan, w, p->p_flag & P_XSLEEP));
548 		p->p_wchan = NULL;
549 		p->p_flag &= ~P_XSLEEP;
550 		if (p->p_stat == SSLEEP) {
551 			/* OPTIMIZED EXPANSION OF setrunnable(p); */
552 			if (p->p_slptime > 1)
553 				updatepri(p);
554 			p->p_slptime = 0;
555 			p->p_stat = SRUN;
556 			if (p->p_flag & P_INMEM) {
557 				setrunqueue(p);
558 			} else {
559 				p->p_flag |= P_SWAPINREQ;
560 				wakeup((caddr_t)&proc0);
561 			}
562 		}
563 	}
564 	splx(s);
565 }
566 #endif
567 
568 /*
569  * Make all processes sleeping on the specified identifier runnable.
570  */
571 static void
572 _wakeup(void *ident, int count)
573 {
574 	struct slpquehead *qp;
575 	struct thread *td;
576 	struct thread *ntd;
577 	struct proc *p;
578 	int s;
579 	int id = LOOKUP(ident);
580 
581 	s = splhigh();
582 	qp = &slpque[id];
583 restart:
584 	for (td = TAILQ_FIRST(qp); td != NULL; td = ntd) {
585 		ntd = TAILQ_NEXT(td, td_threadq);
586 		if (td->td_wchan == ident) {
587 			TAILQ_REMOVE(qp, td, td_threadq);
588 			td->td_wchan = NULL;
589 			if ((p = td->td_proc) != NULL && p->p_stat == SSLEEP) {
590 				/* OPTIMIZED EXPANSION OF setrunnable(p); */
591 				if (p->p_slptime > 1)
592 					updatepri(p);
593 				p->p_slptime = 0;
594 				p->p_stat = SRUN;
595 				if (p->p_flag & P_INMEM) {
596 					setrunqueue(p);
597 				} else {
598 					p->p_flag |= P_SWAPINREQ;
599 					wakeup((caddr_t)&proc0);
600 				}
601 				/* END INLINE EXPANSION */
602 			} else if (p == NULL) {
603 				lwkt_schedule(td);
604 			}
605 			if (--count == 0)
606 				break;
607 			goto restart;
608 		}
609 	}
610 	splx(s);
611 }
612 
613 void
614 wakeup(void *ident)
615 {
616     _wakeup(ident, 0);
617 }
618 
619 void
620 wakeup_one(void *ident)
621 {
622     _wakeup(ident, 1);
623 }
624 
625 /*
626  * The machine independent parts of mi_switch().
627  * Must be called at splstatclock() or higher.
628  */
629 void
630 mi_switch()
631 {
632 	struct thread *td = curthread;
633 	struct proc *p = td->td_proc;	/* XXX */
634 	struct rlimit *rlim;
635 	int x;
636 	u_int64_t ttime;
637 
638 	/*
639 	 * XXX this spl is almost unnecessary.  It is partly to allow for
640 	 * sloppy callers that don't do it (issignal() via CURSIG() is the
641 	 * main offender).  It is partly to work around a bug in the i386
642 	 * cpu_switch() (the ipl is not preserved).  We ran for years
643 	 * without it.  I think there was only a interrupt latency problem.
644 	 * The main caller, tsleep(), does an splx() a couple of instructions
645 	 * after calling here.  The buggy caller, issignal(), usually calls
646 	 * here at spl0() and sometimes returns at splhigh().  The process
647 	 * then runs for a little too long at splhigh().  The ipl gets fixed
648 	 * when the process returns to user mode (or earlier).
649 	 *
650 	 * It would probably be better to always call here at spl0(). Callers
651 	 * are prepared to give up control to another process, so they must
652 	 * be prepared to be interrupted.  The clock stuff here may not
653 	 * actually need splstatclock().
654 	 */
655 	x = splstatclock();
656 	clear_resched();
657 
658 	/*
659 	 * Check if the process exceeds its cpu resource allocation.
660 	 * If over max, kill it.  Time spent in interrupts is not
661 	 * included.  YYY 64 bit match is expensive.  Ick.
662 	 */
663 	ttime = td->td_sticks + td->td_uticks;
664 	if (p->p_stat != SZOMB && p->p_limit->p_cpulimit != RLIM_INFINITY &&
665 	    ttime > p->p_limit->p_cpulimit) {
666 		rlim = &p->p_rlimit[RLIMIT_CPU];
667 		if (ttime / (rlim_t)1000000 >= rlim->rlim_max) {
668 			killproc(p, "exceeded maximum CPU limit");
669 		} else {
670 			psignal(p, SIGXCPU);
671 			if (rlim->rlim_cur < rlim->rlim_max) {
672 				/* XXX: we should make a private copy */
673 				rlim->rlim_cur += 5;
674 			}
675 		}
676 	}
677 
678 	/*
679 	 * Pick a new current process and record its start time.  If we
680 	 * are in a SSTOPped state we deschedule ourselves.  YYY this needs
681 	 * to be cleaned up, remember that LWKTs stay on their run queue
682 	 * which works differently then the user scheduler which removes
683 	 * the process from the runq when it runs it.
684 	 */
685 	mycpu->gd_cnt.v_swtch++;
686 	if (p->p_stat == SSTOP)
687 		lwkt_deschedule_self();
688 	lwkt_switch();
689 
690 	splx(x);
691 }
692 
693 /*
694  * Change process state to be runnable,
695  * placing it on the run queue if it is in memory,
696  * and awakening the swapper if it isn't in memory.
697  */
698 void
699 setrunnable(struct proc *p)
700 {
701 	int s;
702 
703 	s = splhigh();
704 	switch (p->p_stat) {
705 	case 0:
706 	case SRUN:
707 	case SZOMB:
708 	default:
709 		panic("setrunnable");
710 	case SSTOP:
711 	case SSLEEP:
712 		unsleep(p->p_thread);	/* e.g. when sending signals */
713 		break;
714 
715 	case SIDL:
716 		break;
717 	}
718 	p->p_stat = SRUN;
719 	if (p->p_flag & P_INMEM)
720 		setrunqueue(p);
721 	splx(s);
722 	if (p->p_slptime > 1)
723 		updatepri(p);
724 	p->p_slptime = 0;
725 	if ((p->p_flag & P_INMEM) == 0) {
726 		p->p_flag |= P_SWAPINREQ;
727 		wakeup((caddr_t)&proc0);
728 	}
729 }
730 
731 /*
732  * Change the process state to NOT be runnable, removing it from the run
733  * queue.  If P_CURPROC is not set and we are in SRUN the process is on the
734  * run queue (If P_INMEM is not set then it isn't because it is swapped).
735  */
736 void
737 clrrunnable(struct proc *p, int stat)
738 {
739 	int s;
740 
741 	s = splhigh();
742 	switch(p->p_stat) {
743 	case SRUN:
744 		if (p->p_flag & P_ONRUNQ)
745 			remrunqueue(p);
746 		break;
747 	default:
748 		break;
749 	}
750 	p->p_stat = stat;
751 	splx(s);
752 }
753 
754 /*
755  * Compute the priority of a process when running in user mode.
756  * Arrange to reschedule if the resulting priority is better
757  * than that of the current process.
758  */
759 void
760 resetpriority(struct proc *p)
761 {
762 	unsigned int newpriority;
763 	int opq;
764 	int npq;
765 
766 	/*
767 	 * Set p_priority for general process comparisons
768 	 */
769 	switch(p->p_rtprio.type) {
770 	case RTP_PRIO_REALTIME:
771 		p->p_priority = PRIBASE_REALTIME + p->p_rtprio.prio;
772 		return;
773 	case RTP_PRIO_NORMAL:
774 		break;
775 	case RTP_PRIO_IDLE:
776 		p->p_priority = PRIBASE_IDLE + p->p_rtprio.prio;
777 		return;
778 	case RTP_PRIO_THREAD:
779 		p->p_priority = PRIBASE_THREAD + p->p_rtprio.prio;
780 		return;
781 	}
782 
783 	/*
784 	 * NORMAL priorities fall through.  These are based on niceness
785 	 * and cpu use.
786 	 */
787 	newpriority = NICE_ADJUST(p->p_nice - PRIO_MIN) +
788 			p->p_estcpu / ESTCPURAMP;
789 	newpriority = min(newpriority, MAXPRI);
790 	npq = newpriority / PPQ;
791 	crit_enter();
792 	opq = (p->p_priority & PRIMASK) / PPQ;
793 	if (p->p_stat == SRUN && (p->p_flag & P_ONRUNQ) && opq != npq) {
794 		/*
795 		 * We have to move the process to another queue
796 		 */
797 		remrunqueue(p);
798 		p->p_priority = PRIBASE_NORMAL + newpriority;
799 		setrunqueue(p);
800 	} else {
801 		/*
802 		 * We can just adjust the priority and it will be picked
803 		 * up later.
804 		 */
805 		KKASSERT(opq == npq || (p->p_flag & P_ONRUNQ) == 0);
806 		p->p_priority = PRIBASE_NORMAL + newpriority;
807 	}
808 	crit_exit();
809 }
810 
811 /*
812  * Compute a tenex style load average of a quantity on
813  * 1, 5 and 15 minute intervals.
814  */
815 static void
816 loadav(void *arg)
817 {
818 	int i, nrun;
819 	struct loadavg *avg;
820 	struct proc *p;
821 
822 	avg = &averunnable;
823 	nrun = 0;
824 	FOREACH_PROC_IN_SYSTEM(p) {
825 		switch (p->p_stat) {
826 		case SRUN:
827 		case SIDL:
828 			nrun++;
829 		}
830 	}
831 	for (i = 0; i < 3; i++)
832 		avg->ldavg[i] = (cexp[i] * avg->ldavg[i] +
833 		    nrun * FSCALE * (FSCALE - cexp[i])) >> FSHIFT;
834 
835 	/*
836 	 * Schedule the next update to occur after 5 seconds, but add a
837 	 * random variation to avoid synchronisation with processes that
838 	 * run at regular intervals.
839 	 */
840 	callout_reset(&loadav_callout, hz * 4 + (int)(random() % (hz * 2 + 1)),
841 	    loadav, NULL);
842 }
843 
844 /* ARGSUSED */
845 static void
846 sched_setup(void *dummy)
847 {
848 
849 	callout_init(&loadav_callout);
850 
851 	/* Kick off timeout driven events by calling first time. */
852 	roundrobin(NULL);
853 	schedcpu(NULL);
854 	loadav(NULL);
855 }
856 
857 /*
858  * We adjust the priority of the current process.  The priority of
859  * a process gets worse as it accumulates CPU time.  The cpu usage
860  * estimator (p_estcpu) is increased here.  resetpriority() will
861  * compute a different priority each time p_estcpu increases by
862  * INVERSE_ESTCPU_WEIGHT * (until MAXPRI is reached).
863  *
864  * The cpu usage estimator ramps up quite quickly when the process is
865  * running (linearly), and decays away exponentially, at a rate which
866  * is proportionally slower when the system is busy.  The basic principle
867  * is that the system will 90% forget that the process used a lot of CPU
868  * time in 5 * loadav seconds.  This causes the system to favor processes
869  * which haven't run much recently, and to round-robin among other processes.
870  *
871  * WARNING! called from a fast-int or an IPI, the MP lock MIGHT NOT BE HELD
872  * and we cannot block.
873  */
874 void
875 schedulerclock(void *dummy)
876 {
877 	struct thread *td;
878 	struct proc *p;
879 
880 	td = curthread;
881 	if ((p = td->td_proc) != NULL) {
882 		p->p_cpticks++;
883 		p->p_estcpu = ESTCPULIM(p->p_estcpu + 1);
884 		if ((p->p_estcpu % PPQ) == 0 && try_mplock()) {
885 			resetpriority(p);
886 			rel_mplock();
887 		}
888 	}
889 }
890 
891 static
892 void
893 crit_panicints(void)
894 {
895     int s;
896     int cpri;
897 
898     s = splhigh();
899     cpri = crit_panic_save();
900     splx(safepri);
901     crit_panic_restore(cpri);
902     splx(s);
903 }
904 
905