xref: /dflybsd-src/sys/kern/kern_synch.c (revision c5541aee854b0d32586182b733a9ea4d4c92168b)
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 1.31 2004/03/30 19:14:11 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_user_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_user_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  * The load average is scaled by FSCALE (2048 typ).  The estimated cpu is
160  * incremented at a rate of ESTCPUVFREQ per second (40hz typ), but this is
161  * divided up across all cpu bound processes running in the system so an
162  * individual process will get less under load.  ESTCPULIM typicaly caps
163  * out at ESTCPUMAX (around 376, or 11 nice levels).
164  *
165  * Generally speaking the decay equation needs to break-even on growth
166  * at the limit at all load levels >= 1.0, so if the estimated cpu for
167  * a process increases by (ESTVCPUFREQ / load) per second, then the decay
168  * should reach this value when estcpu reaches ESTCPUMAX.  That calculation
169  * is:
170  *
171  *	ESTCPUMAX * decay = ESTCPUVFREQ / load
172  *	decay = ESTCPUVFREQ / (load * ESTCPUMAX)
173  *	decay = estcpu * 0.053 / load
174  *
175  * If the load is less then 1.0 we assume a load of 1.0.
176  */
177 
178 #define cload(loadav)	((loadav) < FSCALE ? FSCALE : (loadav))
179 #define decay_cpu(loadav,estcpu)	\
180     ((estcpu) * (FSCALE * ESTCPUVFREQ / ESTCPUMAX) / cload(loadav))
181 
182 /* decay 95% of `p_pctcpu' in 60 seconds; see CCPU_SHIFT before changing */
183 static fixpt_t	ccpu = 0.95122942450071400909 * FSCALE;	/* exp(-1/20) */
184 SYSCTL_INT(_kern, OID_AUTO, ccpu, CTLFLAG_RD, &ccpu, 0, "");
185 
186 /* kernel uses `FSCALE', userland (SHOULD) use kern.fscale */
187 static int	fscale __unused = FSCALE;
188 SYSCTL_INT(_kern, OID_AUTO, fscale, CTLFLAG_RD, 0, FSCALE, "");
189 
190 /*
191  * If `ccpu' is not equal to `exp(-1/20)' and you still want to use the
192  * faster/more-accurate formula, you'll have to estimate CCPU_SHIFT below
193  * and possibly adjust FSHIFT in "param.h" so that (FSHIFT >= CCPU_SHIFT).
194  *
195  * To estimate CCPU_SHIFT for exp(-1/20), the following formula was used:
196  *	1 - exp(-1/20) ~= 0.0487 ~= 0.0488 == 1 (fixed pt, *11* bits).
197  *
198  * If you don't want to bother with the faster/more-accurate formula, you
199  * can set CCPU_SHIFT to (FSHIFT + 1) which will use a slower/less-accurate
200  * (more general) method of calculating the %age of CPU used by a process.
201  */
202 #define	CCPU_SHIFT	11
203 
204 /*
205  * Recompute process priorities, once a second.
206  */
207 /* ARGSUSED */
208 static void
209 schedcpu(void *arg)
210 {
211 	fixpt_t loadfac = averunnable.ldavg[0];
212 	struct proc *p;
213 	int s;
214 	unsigned int ndecay;
215 
216 	FOREACH_PROC_IN_SYSTEM(p) {
217 		/*
218 		 * Increment time in/out of memory and sleep time
219 		 * (if sleeping).  We ignore overflow; with 16-bit int's
220 		 * (remember them?) overflow takes 45 days.
221 		 */
222 		p->p_swtime++;
223 		if (p->p_stat == SSLEEP || p->p_stat == SSTOP)
224 			p->p_slptime++;
225 		p->p_pctcpu = (p->p_pctcpu * ccpu) >> FSHIFT;
226 		/*
227 		 * If the process has slept the entire second,
228 		 * stop recalculating its priority until it wakes up.
229 		 */
230 		if (p->p_slptime > 1)
231 			continue;
232 		s = splhigh();	/* prevent state changes and protect run queue */
233 		/*
234 		 * p_pctcpu is only for ps.
235 		 */
236 #if	(FSHIFT >= CCPU_SHIFT)
237 		p->p_pctcpu += (ESTCPUFREQ == 100)?
238 			((fixpt_t) p->p_cpticks) << (FSHIFT - CCPU_SHIFT):
239                 	100 * (((fixpt_t) p->p_cpticks)
240 				<< (FSHIFT - CCPU_SHIFT)) / ESTCPUFREQ;
241 #else
242 		p->p_pctcpu += ((FSCALE - ccpu) *
243 			(p->p_cpticks * FSCALE / ESTCPUFREQ)) >> FSHIFT;
244 #endif
245 		p->p_cpticks = 0;
246 		ndecay = decay_cpu(loadfac, p->p_estcpu);
247 		if (p->p_estcpu > ndecay)
248 			p->p_estcpu -= ndecay;
249 		else
250 			p->p_estcpu = 0;
251 		resetpriority(p);
252 		splx(s);
253 	}
254 	wakeup((caddr_t)&lbolt);
255 	timeout(schedcpu, (void *)0, hz);
256 }
257 
258 /*
259  * Recalculate the priority of a process after it has slept for a while.
260  * For all load averages >= 1 and max p_estcpu of 255, sleeping for at
261  * least six times the loadfactor will decay p_estcpu to zero.
262  */
263 static void
264 updatepri(struct proc *p)
265 {
266 	unsigned int ndecay;
267 
268 	ndecay = decay_cpu(averunnable.ldavg[0], p->p_estcpu) * p->p_slptime;
269 	if (p->p_estcpu > ndecay)
270 		p->p_estcpu -= ndecay;
271 	else
272 		p->p_estcpu = 0;
273 	resetpriority(p);
274 }
275 
276 /*
277  * We're only looking at 7 bits of the address; everything is
278  * aligned to 4, lots of things are aligned to greater powers
279  * of 2.  Shift right by 8, i.e. drop the bottom 256 worth.
280  */
281 #define TABLESIZE	128
282 static TAILQ_HEAD(slpquehead, thread) slpque[TABLESIZE];
283 #define LOOKUP(x)	(((intptr_t)(x) >> 8) & (TABLESIZE - 1))
284 
285 /*
286  * During autoconfiguration or after a panic, a sleep will simply
287  * lower the priority briefly to allow interrupts, then return.
288  * The priority to be used (safepri) is machine-dependent, thus this
289  * value is initialized and maintained in the machine-dependent layers.
290  * This priority will typically be 0, or the lowest priority
291  * that is safe for use on the interrupt stack; it can be made
292  * higher to block network software interrupts after panics.
293  */
294 int safepri;
295 
296 void
297 sleepinit(void)
298 {
299 	int i;
300 
301 	sched_quantum = hz/10;
302 	hogticks = 2 * sched_quantum;
303 	for (i = 0; i < TABLESIZE; i++)
304 		TAILQ_INIT(&slpque[i]);
305 }
306 
307 /*
308  * General sleep call.  Suspends the current process until a wakeup is
309  * performed on the specified identifier.  The process will then be made
310  * runnable with the specified priority.  Sleeps at most timo/hz seconds
311  * (0 means no timeout).  If flags includes PCATCH flag, signals are checked
312  * before and after sleeping, else signals are not checked.  Returns 0 if
313  * awakened, EWOULDBLOCK if the timeout expires.  If PCATCH is set and a
314  * signal needs to be delivered, ERESTART is returned if the current system
315  * call should be restarted if possible, and EINTR is returned if the system
316  * call should be interrupted by the signal (return EINTR).
317  *
318  * Note that if we are a process, we release_curproc() before messing with
319  * the LWKT scheduler.
320  */
321 int
322 tsleep(void *ident, int flags, const char *wmesg, int timo)
323 {
324 	struct thread *td = curthread;
325 	struct proc *p = td->td_proc;		/* may be NULL */
326 	int s, sig = 0, catch = flags & PCATCH;
327 	int id = LOOKUP(ident);
328 	struct callout_handle thandle;
329 
330 	/*
331 	 * NOTE: removed KTRPOINT, it could cause races due to blocking
332 	 * even in stable.  Just scrap it for now.
333 	 */
334 	if (cold || panicstr) {
335 		/*
336 		 * After a panic, or during autoconfiguration,
337 		 * just give interrupts a chance, then just return;
338 		 * don't run any other procs or panic below,
339 		 * in case this is the idle process and already asleep.
340 		 */
341 		crit_panicints();
342 		return (0);
343 	}
344 	KKASSERT(td != &mycpu->gd_idlethread);	/* you must be kidding! */
345 	s = splhigh();
346 	KASSERT(ident != NULL, ("tsleep: no ident"));
347 	KASSERT(p == NULL || p->p_stat == SRUN, ("tsleep %p %s %d",
348 		ident, wmesg, p->p_stat));
349 
350 	crit_enter();
351 	td->td_wchan = ident;
352 	td->td_wmesg = wmesg;
353 	if (p) {
354 		if (flags & PNORESCHED)
355 			td->td_flags |= TDF_NORESCHED;
356 		release_curproc(p);
357 		p->p_slptime = 0;
358 	}
359 	lwkt_deschedule_self();
360 	TAILQ_INSERT_TAIL(&slpque[id], td, td_threadq);
361 	if (timo)
362 		thandle = timeout(endtsleep, (void *)td, timo);
363 	/*
364 	 * We put ourselves on the sleep queue and start our timeout
365 	 * before calling CURSIG, as we could stop there, and a wakeup
366 	 * or a SIGCONT (or both) could occur while we were stopped.
367 	 * A SIGCONT would cause us to be marked as SSLEEP
368 	 * without resuming us, thus we must be ready for sleep
369 	 * when CURSIG is called.  If the wakeup happens while we're
370 	 * stopped, td->td_wchan will be 0 upon return from CURSIG.
371 	 */
372 	if (p) {
373 		if (catch) {
374 			p->p_flag |= P_SINTR;
375 			if ((sig = CURSIG(p))) {
376 				if (td->td_wchan) {
377 					unsleep(td);
378 					lwkt_schedule_self();
379 				}
380 				p->p_stat = SRUN;
381 				goto resume;
382 			}
383 			if (td->td_wchan == NULL) {
384 				catch = 0;
385 				goto resume;
386 			}
387 		} else {
388 			sig = 0;
389 		}
390 
391 		/*
392 		 * If we are not the current process we have to remove ourself
393 		 * from the run queue.
394 		 */
395 		KASSERT(p->p_stat == SRUN, ("PSTAT NOT SRUN %d %d", p->p_pid, p->p_stat));
396 		/*
397 		 * If this is the current 'user' process schedule another one.
398 		 */
399 		clrrunnable(p, SSLEEP);
400 		p->p_stats->p_ru.ru_nvcsw++;
401 		mi_switch();
402 		KASSERT(p->p_stat == SRUN, ("tsleep: stat not srun"));
403 	} else {
404 		lwkt_switch();
405 	}
406 resume:
407 	crit_exit();
408 	if (p)
409 		p->p_flag &= ~P_SINTR;
410 	splx(s);
411 	td->td_flags &= ~TDF_NORESCHED;
412 	if (td->td_flags & TDF_TIMEOUT) {
413 		td->td_flags &= ~TDF_TIMEOUT;
414 		if (sig == 0)
415 			return (EWOULDBLOCK);
416 	} else if (timo) {
417 		untimeout(endtsleep, (void *)td, thandle);
418 	} else if (td->td_wmesg) {
419 		/*
420 		 * This can happen if a thread is woken up directly.  Clear
421 		 * wmesg to avoid debugging confusion.
422 		 */
423 		td->td_wmesg = NULL;
424 	}
425 	/* inline of iscaught() */
426 	if (p) {
427 		if (catch && (sig != 0 || (sig = CURSIG(p)))) {
428 			if (SIGISMEMBER(p->p_sigacts->ps_sigintr, sig))
429 				return (EINTR);
430 			return (ERESTART);
431 		}
432 	}
433 	return (0);
434 }
435 
436 /*
437  * Implement the timeout for tsleep.  We interlock against
438  * wchan when setting TDF_TIMEOUT.  For processes we remove
439  * the sleep if the process is stopped rather then sleeping,
440  * so it remains stopped.
441  */
442 static void
443 endtsleep(void *arg)
444 {
445 	thread_t td = arg;
446 	struct proc *p;
447 	int s;
448 
449 	s = splhigh();
450 	if (td->td_wchan) {
451 		td->td_flags |= TDF_TIMEOUT;
452 		if ((p = td->td_proc) != NULL) {
453 			if (p->p_stat == SSLEEP)
454 				setrunnable(p);
455 			else
456 				unsleep(td);
457 		} else {
458 			unsleep(td);
459 			lwkt_schedule(td);
460 		}
461 	}
462 	splx(s);
463 }
464 
465 /*
466  * Remove a process from its wait queue
467  */
468 void
469 unsleep(struct thread *td)
470 {
471 	int s;
472 
473 	s = splhigh();
474 	if (td->td_wchan) {
475 #if 0
476 		if (p->p_flag & P_XSLEEP) {
477 			struct xwait *w = p->p_wchan;
478 			TAILQ_REMOVE(&w->waitq, p, p_procq);
479 			p->p_flag &= ~P_XSLEEP;
480 		} else
481 #endif
482 		TAILQ_REMOVE(&slpque[LOOKUP(td->td_wchan)], td, td_threadq);
483 		td->td_wchan = NULL;
484 	}
485 	splx(s);
486 }
487 
488 #if 0
489 /*
490  * Make all processes sleeping on the explicit lock structure runnable.
491  */
492 void
493 xwakeup(struct xwait *w)
494 {
495 	struct proc *p;
496 	int s;
497 
498 	s = splhigh();
499 	++w->gen;
500 	while ((p = TAILQ_FIRST(&w->waitq)) != NULL) {
501 		TAILQ_REMOVE(&w->waitq, p, p_procq);
502 		KASSERT(p->p_wchan == w && (p->p_flag & P_XSLEEP),
503 		    ("xwakeup: wchan mismatch for %p (%p/%p) %08x", p, p->p_wchan, w, p->p_flag & P_XSLEEP));
504 		p->p_wchan = NULL;
505 		p->p_flag &= ~P_XSLEEP;
506 		if (p->p_stat == SSLEEP) {
507 			/* OPTIMIZED EXPANSION OF setrunnable(p); */
508 			if (p->p_slptime > 1)
509 				updatepri(p);
510 			p->p_slptime = 0;
511 			p->p_stat = SRUN;
512 			if (p->p_flag & P_INMEM) {
513 				setrunqueue(p);
514 			} else {
515 				p->p_flag |= P_SWAPINREQ;
516 				wakeup((caddr_t)&proc0);
517 			}
518 		}
519 	}
520 	splx(s);
521 }
522 #endif
523 
524 /*
525  * Make all processes sleeping on the specified identifier runnable.
526  */
527 static void
528 _wakeup(void *ident, int count)
529 {
530 	struct slpquehead *qp;
531 	struct thread *td;
532 	struct thread *ntd;
533 	struct proc *p;
534 	int s;
535 	int id = LOOKUP(ident);
536 
537 	s = splhigh();
538 	qp = &slpque[id];
539 restart:
540 	for (td = TAILQ_FIRST(qp); td != NULL; td = ntd) {
541 		ntd = TAILQ_NEXT(td, td_threadq);
542 		if (td->td_wchan == ident) {
543 			TAILQ_REMOVE(qp, td, td_threadq);
544 			td->td_wchan = NULL;
545 			if ((p = td->td_proc) != NULL && p->p_stat == SSLEEP) {
546 				/* OPTIMIZED EXPANSION OF setrunnable(p); */
547 				if (p->p_slptime > 1)
548 					updatepri(p);
549 				p->p_slptime = 0;
550 				p->p_stat = SRUN;
551 				if (p->p_flag & P_INMEM) {
552 					setrunqueue(p);
553 				} else {
554 					p->p_flag |= P_SWAPINREQ;
555 					wakeup((caddr_t)&proc0);
556 				}
557 				/* END INLINE EXPANSION */
558 			} else if (p == NULL) {
559 				lwkt_schedule(td);
560 			}
561 			if (--count == 0)
562 				break;
563 			goto restart;
564 		}
565 	}
566 	splx(s);
567 }
568 
569 void
570 wakeup(void *ident)
571 {
572     _wakeup(ident, 0);
573 }
574 
575 void
576 wakeup_one(void *ident)
577 {
578     _wakeup(ident, 1);
579 }
580 
581 /*
582  * The machine independent parts of mi_switch().
583  * Must be called at splstatclock() or higher.
584  */
585 void
586 mi_switch()
587 {
588 	struct thread *td = curthread;
589 	struct proc *p = td->td_proc;	/* XXX */
590 	struct rlimit *rlim;
591 	int x;
592 	u_int64_t ttime;
593 
594 	/*
595 	 * XXX this spl is almost unnecessary.  It is partly to allow for
596 	 * sloppy callers that don't do it (issignal() via CURSIG() is the
597 	 * main offender).  It is partly to work around a bug in the i386
598 	 * cpu_switch() (the ipl is not preserved).  We ran for years
599 	 * without it.  I think there was only a interrupt latency problem.
600 	 * The main caller, tsleep(), does an splx() a couple of instructions
601 	 * after calling here.  The buggy caller, issignal(), usually calls
602 	 * here at spl0() and sometimes returns at splhigh().  The process
603 	 * then runs for a little too long at splhigh().  The ipl gets fixed
604 	 * when the process returns to user mode (or earlier).
605 	 *
606 	 * It would probably be better to always call here at spl0(). Callers
607 	 * are prepared to give up control to another process, so they must
608 	 * be prepared to be interrupted.  The clock stuff here may not
609 	 * actually need splstatclock().
610 	 */
611 	x = splstatclock();
612 
613 	/*
614 	 * Check if the process exceeds its cpu resource allocation.
615 	 * If over max, kill it.  Time spent in interrupts is not
616 	 * included.  YYY 64 bit match is expensive.  Ick.
617 	 */
618 	ttime = td->td_sticks + td->td_uticks;
619 	if (p->p_stat != SZOMB && p->p_limit->p_cpulimit != RLIM_INFINITY &&
620 	    ttime > p->p_limit->p_cpulimit) {
621 		rlim = &p->p_rlimit[RLIMIT_CPU];
622 		if (ttime / (rlim_t)1000000 >= rlim->rlim_max) {
623 			killproc(p, "exceeded maximum CPU limit");
624 		} else {
625 			psignal(p, SIGXCPU);
626 			if (rlim->rlim_cur < rlim->rlim_max) {
627 				/* XXX: we should make a private copy */
628 				rlim->rlim_cur += 5;
629 			}
630 		}
631 	}
632 
633 	/*
634 	 * Pick a new current process and record its start time.  If we
635 	 * are in a SSTOPped state we deschedule ourselves.  YYY this needs
636 	 * to be cleaned up, remember that LWKTs stay on their run queue
637 	 * which works differently then the user scheduler which removes
638 	 * the process from the runq when it runs it.
639 	 */
640 	mycpu->gd_cnt.v_swtch++;
641 	if (p->p_stat == SSTOP)
642 		lwkt_deschedule_self();
643 	lwkt_switch();
644 
645 	splx(x);
646 }
647 
648 /*
649  * Change process state to be runnable,
650  * placing it on the run queue if it is in memory,
651  * and awakening the swapper if it isn't in memory.
652  */
653 void
654 setrunnable(struct proc *p)
655 {
656 	int s;
657 
658 	s = splhigh();
659 	switch (p->p_stat) {
660 	case 0:
661 	case SRUN:
662 	case SZOMB:
663 	default:
664 		panic("setrunnable");
665 	case SSTOP:
666 	case SSLEEP:
667 		unsleep(p->p_thread);	/* e.g. when sending signals */
668 		break;
669 
670 	case SIDL:
671 		break;
672 	}
673 	p->p_stat = SRUN;
674 	if (p->p_flag & P_INMEM)
675 		setrunqueue(p);
676 	splx(s);
677 	if (p->p_slptime > 1)
678 		updatepri(p);
679 	p->p_slptime = 0;
680 	if ((p->p_flag & P_INMEM) == 0) {
681 		p->p_flag |= P_SWAPINREQ;
682 		wakeup((caddr_t)&proc0);
683 	}
684 }
685 
686 /*
687  * Change the process state to NOT be runnable, removing it from the run
688  * queue.
689  */
690 void
691 clrrunnable(struct proc *p, int stat)
692 {
693 	crit_enter_quick(p->p_thread);
694 	if (p->p_stat == SRUN && (p->p_flag & P_ONRUNQ))
695 		remrunqueue(p);
696 	p->p_stat = stat;
697 	crit_exit_quick(p->p_thread);
698 }
699 
700 /*
701  * Compute the priority of a process when running in user mode.
702  * Arrange to reschedule if the resulting priority is better
703  * than that of the current process.
704  */
705 void
706 resetpriority(struct proc *p)
707 {
708 	unsigned int newpriority;
709 	int opq;
710 	int npq;
711 
712 	/*
713 	 * Set p_priority for general process comparisons
714 	 */
715 	switch(p->p_rtprio.type) {
716 	case RTP_PRIO_REALTIME:
717 		p->p_priority = PRIBASE_REALTIME + p->p_rtprio.prio;
718 		return;
719 	case RTP_PRIO_NORMAL:
720 		break;
721 	case RTP_PRIO_IDLE:
722 		p->p_priority = PRIBASE_IDLE + p->p_rtprio.prio;
723 		return;
724 	case RTP_PRIO_THREAD:
725 		p->p_priority = PRIBASE_THREAD + p->p_rtprio.prio;
726 		return;
727 	}
728 
729 	/*
730 	 * NORMAL priorities fall through.  These are based on niceness
731 	 * and cpu use.
732 	 */
733 	newpriority = NICE_ADJUST(p->p_nice - PRIO_MIN) +
734 			p->p_estcpu / ESTCPURAMP;
735 	newpriority = min(newpriority, MAXPRI);
736 	npq = newpriority / PPQ;
737 	crit_enter();
738 	opq = (p->p_priority & PRIMASK) / PPQ;
739 	if (p->p_stat == SRUN && (p->p_flag & P_ONRUNQ) && opq != npq) {
740 		/*
741 		 * We have to move the process to another queue
742 		 */
743 		remrunqueue(p);
744 		p->p_priority = PRIBASE_NORMAL + newpriority;
745 		setrunqueue(p);
746 	} else {
747 		/*
748 		 * We can just adjust the priority and it will be picked
749 		 * up later.
750 		 */
751 		KKASSERT(opq == npq || (p->p_flag & P_ONRUNQ) == 0);
752 		p->p_priority = PRIBASE_NORMAL + newpriority;
753 	}
754 	crit_exit();
755 }
756 
757 /*
758  * Compute a tenex style load average of a quantity on
759  * 1, 5 and 15 minute intervals.
760  */
761 static void
762 loadav(void *arg)
763 {
764 	int i, nrun;
765 	struct loadavg *avg;
766 	struct proc *p;
767 
768 	avg = &averunnable;
769 	nrun = 0;
770 	FOREACH_PROC_IN_SYSTEM(p) {
771 		switch (p->p_stat) {
772 		case SRUN:
773 		case SIDL:
774 			nrun++;
775 		}
776 	}
777 	for (i = 0; i < 3; i++)
778 		avg->ldavg[i] = (cexp[i] * avg->ldavg[i] +
779 		    nrun * FSCALE * (FSCALE - cexp[i])) >> FSHIFT;
780 
781 	/*
782 	 * Schedule the next update to occur after 5 seconds, but add a
783 	 * random variation to avoid synchronisation with processes that
784 	 * run at regular intervals.
785 	 */
786 	callout_reset(&loadav_callout, hz * 4 + (int)(random() % (hz * 2 + 1)),
787 	    loadav, NULL);
788 }
789 
790 /* ARGSUSED */
791 static void
792 sched_setup(void *dummy)
793 {
794 
795 	callout_init(&loadav_callout);
796 
797 	/* Kick off timeout driven events by calling first time. */
798 	roundrobin(NULL);
799 	schedcpu(NULL);
800 	loadav(NULL);
801 }
802 
803 /*
804  * We adjust the priority of the current process.  The priority of
805  * a process gets worse as it accumulates CPU time.  The cpu usage
806  * estimator (p_estcpu) is increased here.  resetpriority() will
807  * compute a different priority each time p_estcpu increases by
808  * INVERSE_ESTCPU_WEIGHT * (until MAXPRI is reached).
809  *
810  * The cpu usage estimator ramps up quite quickly when the process is
811  * running (linearly), and decays away exponentially, at a rate which
812  * is proportionally slower when the system is busy.  The basic principle
813  * is that the system will 90% forget that the process used a lot of CPU
814  * time in 5 * loadav seconds.  This causes the system to favor processes
815  * which haven't run much recently, and to round-robin among other processes.
816  *
817  * The actual schedulerclock interrupt rate is ESTCPUFREQ, but we generally
818  * want to ramp-up at a faster rate, ESTCPUVFREQ, so p_estcpu is scaled
819  * by (ESTCPUVFREQ / ESTCPUFREQ).  You can control the ramp-up/ramp-down
820  * rate by adjusting ESTCPUVFREQ in sys/proc.h in integer multiples
821  * of ESTCPUFREQ.
822  *
823  * WARNING! called from a fast-int or an IPI, the MP lock MIGHT NOT BE HELD
824  * and we cannot block.
825  */
826 void
827 schedulerclock(void *dummy)
828 {
829 	struct thread *td;
830 	struct proc *p;
831 
832 	td = curthread;
833 	if ((p = td->td_proc) != NULL) {
834 		p->p_cpticks++;		/* cpticks runs at ESTCPUFREQ */
835 		p->p_estcpu = ESTCPULIM(p->p_estcpu + ESTCPUVFREQ / ESTCPUFREQ);
836 		if (try_mplock()) {
837 			resetpriority(p);
838 			rel_mplock();
839 		}
840 	}
841 }
842 
843 static
844 void
845 crit_panicints(void)
846 {
847     int s;
848     int cpri;
849 
850     s = splhigh();
851     cpri = crit_panic_save();
852     splx(safepri);
853     crit_panic_restore(cpri);
854     splx(s);
855 }
856 
857