xref: /dflybsd-src/sys/kern/kern_synch.c (revision a6d5e0d89a9c7db991a06391d6a4a6ddadf6af51)
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 1.42 2005/06/06 15:02:28 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 int	safepri;
73 
74 static struct callout loadav_callout;
75 static struct callout roundrobin_callout;
76 static struct callout schedcpu_callout;
77 
78 struct loadavg averunnable =
79 	{ {0, 0, 0}, FSCALE };	/* load average, of runnable procs */
80 /*
81  * Constants for averages over 1, 5, and 15 minutes
82  * when sampling at 5 second intervals.
83  */
84 static fixpt_t cexp[3] = {
85 	0.9200444146293232 * FSCALE,	/* exp(-1/12) */
86 	0.9834714538216174 * FSCALE,	/* exp(-1/60) */
87 	0.9944598480048967 * FSCALE,	/* exp(-1/180) */
88 };
89 
90 static void	endtsleep (void *);
91 static void	loadav (void *arg);
92 static void	roundrobin (void *arg);
93 static void	schedcpu (void *arg);
94 static void	updatepri (struct proc *p);
95 
96 static int
97 sysctl_kern_quantum(SYSCTL_HANDLER_ARGS)
98 {
99 	int error, new_val;
100 
101 	new_val = sched_quantum * tick;
102 	error = sysctl_handle_int(oidp, &new_val, 0, req);
103         if (error != 0 || req->newptr == NULL)
104 		return (error);
105 	if (new_val < tick)
106 		return (EINVAL);
107 	sched_quantum = new_val / tick;
108 	hogticks = 2 * sched_quantum;
109 	return (0);
110 }
111 
112 SYSCTL_PROC(_kern, OID_AUTO, quantum, CTLTYPE_INT|CTLFLAG_RW,
113 	0, sizeof sched_quantum, sysctl_kern_quantum, "I", "");
114 
115 int
116 roundrobin_interval(void)
117 {
118 	return (sched_quantum);
119 }
120 
121 /*
122  * Force switch among equal priority processes every 100ms.
123  *
124  * WARNING!  The MP lock is not held on ipi message remotes.
125  */
126 #ifdef SMP
127 
128 static void
129 roundrobin_remote(void *arg)
130 {
131 	struct proc *p = lwkt_preempted_proc();
132  	if (p == NULL || RTP_PRIO_NEED_RR(p->p_rtprio.type))
133 		need_user_resched();
134 }
135 
136 #endif
137 
138 static void
139 roundrobin(void *arg)
140 {
141 	struct proc *p = lwkt_preempted_proc();
142  	if (p == NULL || RTP_PRIO_NEED_RR(p->p_rtprio.type))
143 		need_user_resched();
144 #ifdef SMP
145 	lwkt_send_ipiq_mask(mycpu->gd_other_cpus, roundrobin_remote, NULL);
146 #endif
147  	callout_reset(&roundrobin_callout, sched_quantum, roundrobin, NULL);
148 }
149 
150 #ifdef SMP
151 
152 void
153 resched_cpus(u_int32_t mask)
154 {
155 	lwkt_send_ipiq_mask(mask, roundrobin_remote, NULL);
156 }
157 
158 #endif
159 
160 /*
161  * The load average is scaled by FSCALE (2048 typ).  The estimated cpu is
162  * incremented at a rate of ESTCPUVFREQ per second (40hz typ), but this is
163  * divided up across all cpu bound processes running in the system so an
164  * individual process will get less under load.  ESTCPULIM typicaly caps
165  * out at ESTCPUMAX (around 376, or 11 nice levels).
166  *
167  * Generally speaking the decay equation needs to break-even on growth
168  * at the limit at all load levels >= 1.0, so if the estimated cpu for
169  * a process increases by (ESTVCPUFREQ / load) per second, then the decay
170  * should reach this value when estcpu reaches ESTCPUMAX.  That calculation
171  * is:
172  *
173  *	ESTCPUMAX * decay = ESTCPUVFREQ / load
174  *	decay = ESTCPUVFREQ / (load * ESTCPUMAX)
175  *	decay = estcpu * 0.053 / load
176  *
177  * If the load is less then 1.0 we assume a load of 1.0.
178  */
179 
180 #define cload(loadav)	((loadav) < FSCALE ? FSCALE : (loadav))
181 #define decay_cpu(loadav,estcpu)	\
182     ((estcpu) * (FSCALE * ESTCPUVFREQ / ESTCPUMAX) / cload(loadav))
183 
184 /* decay 95% of `p_pctcpu' in 60 seconds; see CCPU_SHIFT before changing */
185 static fixpt_t	ccpu = 0.95122942450071400909 * FSCALE;	/* exp(-1/20) */
186 SYSCTL_INT(_kern, OID_AUTO, ccpu, CTLFLAG_RD, &ccpu, 0, "");
187 
188 /* kernel uses `FSCALE', userland (SHOULD) use kern.fscale */
189 static int	fscale __unused = FSCALE;
190 SYSCTL_INT(_kern, OID_AUTO, fscale, CTLFLAG_RD, 0, FSCALE, "");
191 
192 /*
193  * If `ccpu' is not equal to `exp(-1/20)' and you still want to use the
194  * faster/more-accurate formula, you'll have to estimate CCPU_SHIFT below
195  * and possibly adjust FSHIFT in "param.h" so that (FSHIFT >= CCPU_SHIFT).
196  *
197  * To estimate CCPU_SHIFT for exp(-1/20), the following formula was used:
198  *	1 - exp(-1/20) ~= 0.0487 ~= 0.0488 == 1 (fixed pt, *11* bits).
199  *
200  * If you don't want to bother with the faster/more-accurate formula, you
201  * can set CCPU_SHIFT to (FSHIFT + 1) which will use a slower/less-accurate
202  * (more general) method of calculating the %age of CPU used by a process.
203  */
204 #define	CCPU_SHIFT	11
205 
206 /*
207  * Recompute process priorities, once a second.
208  */
209 /* ARGSUSED */
210 static void
211 schedcpu(void *arg)
212 {
213 	fixpt_t loadfac = averunnable.ldavg[0];
214 	struct proc *p;
215 	unsigned int ndecay;
216 
217 	FOREACH_PROC_IN_SYSTEM(p) {
218 		/*
219 		 * Increment time in/out of memory and sleep time
220 		 * (if sleeping).  We ignore overflow; with 16-bit int's
221 		 * (remember them?) overflow takes 45 days.
222 		 */
223 		p->p_swtime++;
224 		if (p->p_stat == SSLEEP || p->p_stat == SSTOP)
225 			p->p_slptime++;
226 		p->p_pctcpu = (p->p_pctcpu * ccpu) >> FSHIFT;
227 
228 		/*
229 		 * If the process has slept the entire second,
230 		 * stop recalculating its priority until it wakes up.
231 		 *
232 		 * Note that interactive calculations do not occur for
233 		 * long sleeps (because that isn't necessarily indicative
234 		 * of an interactive process).
235 		 */
236 		if (p->p_slptime > 1)
237 			continue;
238 		/* prevent state changes and protect run queue */
239 		crit_enter();
240 		/*
241 		 * p_cpticks runs at ESTCPUFREQ but must be divided by the
242 		 * load average for par-100% use.  Higher p_interactive
243 		 * values mean less interactive, lower values mean more
244 		 * interactive.
245 		 */
246 		if ((((fixpt_t)p->p_cpticks * cload(loadfac)) >> FSHIFT)  >
247 		    ESTCPUFREQ / 4) {
248 			if (p->p_interactive < 127)
249 				++p->p_interactive;
250 		} else {
251 			if (p->p_interactive > -127)
252 				--p->p_interactive;
253 		}
254 		/*
255 		 * p_pctcpu is only for ps.
256 		 */
257 #if	(FSHIFT >= CCPU_SHIFT)
258 		p->p_pctcpu += (ESTCPUFREQ == 100)?
259 			((fixpt_t) p->p_cpticks) << (FSHIFT - CCPU_SHIFT):
260                 	100 * (((fixpt_t) p->p_cpticks)
261 				<< (FSHIFT - CCPU_SHIFT)) / ESTCPUFREQ;
262 #else
263 		p->p_pctcpu += ((FSCALE - ccpu) *
264 			(p->p_cpticks * FSCALE / ESTCPUFREQ)) >> FSHIFT;
265 #endif
266 		p->p_cpticks = 0;
267 		ndecay = decay_cpu(loadfac, p->p_estcpu);
268 		if (p->p_estcpu > ndecay)
269 			p->p_estcpu -= ndecay;
270 		else
271 			p->p_estcpu = 0;
272 		resetpriority(p);
273 		crit_exit();
274 	}
275 	wakeup((caddr_t)&lbolt);
276 	callout_reset(&schedcpu_callout, hz, schedcpu, NULL);
277 }
278 
279 /*
280  * Recalculate the priority of a process after it has slept for a while.
281  * For all load averages >= 1 and max p_estcpu of 255, sleeping for at
282  * least six times the loadfactor will decay p_estcpu to zero.
283  */
284 static void
285 updatepri(struct proc *p)
286 {
287 	unsigned int ndecay;
288 
289 	ndecay = decay_cpu(averunnable.ldavg[0], p->p_estcpu) * p->p_slptime;
290 	if (p->p_estcpu > ndecay)
291 		p->p_estcpu -= ndecay;
292 	else
293 		p->p_estcpu = 0;
294 	resetpriority(p);
295 }
296 
297 /*
298  * We're only looking at 7 bits of the address; everything is
299  * aligned to 4, lots of things are aligned to greater powers
300  * of 2.  Shift right by 8, i.e. drop the bottom 256 worth.
301  */
302 #define TABLESIZE	128
303 static TAILQ_HEAD(slpquehead, thread) slpque[TABLESIZE];
304 #define LOOKUP(x)	(((intptr_t)(x) >> 8) & (TABLESIZE - 1))
305 
306 /*
307  * During autoconfiguration or after a panic, a sleep will simply
308  * lower the priority briefly to allow interrupts, then return.
309  */
310 
311 void
312 sleepinit(void)
313 {
314 	int i;
315 
316 	sched_quantum = hz/10;
317 	hogticks = 2 * sched_quantum;
318 	for (i = 0; i < TABLESIZE; i++)
319 		TAILQ_INIT(&slpque[i]);
320 }
321 
322 /*
323  * General sleep call.  Suspends the current process until a wakeup is
324  * performed on the specified identifier.  The process will then be made
325  * runnable with the specified priority.  Sleeps at most timo/hz seconds
326  * (0 means no timeout).  If flags includes PCATCH flag, signals are checked
327  * before and after sleeping, else signals are not checked.  Returns 0 if
328  * awakened, EWOULDBLOCK if the timeout expires.  If PCATCH is set and a
329  * signal needs to be delivered, ERESTART is returned if the current system
330  * call should be restarted if possible, and EINTR is returned if the system
331  * call should be interrupted by the signal (return EINTR).
332  *
333  * Note that if we are a process, we release_curproc() before messing with
334  * the LWKT scheduler.
335  */
336 int
337 tsleep(void *ident, int flags, const char *wmesg, int timo)
338 {
339 	struct thread *td = curthread;
340 	struct proc *p = td->td_proc;		/* may be NULL */
341 	int sig = 0, catch = flags & PCATCH;
342 	int id = LOOKUP(ident);
343 	int oldpri;
344 	struct callout thandle;
345 
346 	/*
347 	 * NOTE: removed KTRPOINT, it could cause races due to blocking
348 	 * even in stable.  Just scrap it for now.
349 	 */
350 	if (cold || panicstr) {
351 		/*
352 		 * After a panic, or during autoconfiguration,
353 		 * just give interrupts a chance, then just return;
354 		 * don't run any other procs or panic below,
355 		 * in case this is the idle process and already asleep.
356 		 */
357 		splz();
358 		oldpri = td->td_pri & TDPRI_MASK;
359 		lwkt_setpri_self(safepri);
360 		lwkt_switch();
361 		lwkt_setpri_self(oldpri);
362 		return (0);
363 	}
364 	KKASSERT(td != &mycpu->gd_idlethread);	/* you must be kidding! */
365 	crit_enter_quick(td);
366 	KASSERT(ident != NULL, ("tsleep: no ident"));
367 	KASSERT(p == NULL || p->p_stat == SRUN, ("tsleep %p %s %d",
368 		ident, wmesg, p->p_stat));
369 
370 	td->td_wchan = ident;
371 	td->td_wmesg = wmesg;
372 	td->td_wdomain = flags & PDOMAIN_MASK;
373 	if (p) {
374 		if (flags & PNORESCHED)
375 			td->td_flags |= TDF_NORESCHED;
376 		release_curproc(p);
377 		p->p_slptime = 0;
378 	}
379 	lwkt_deschedule_self(td);
380 	TAILQ_INSERT_TAIL(&slpque[id], td, td_threadq);
381 	if (timo) {
382 		callout_init(&thandle);
383 		callout_reset(&thandle, timo, endtsleep, td);
384 	}
385 	/*
386 	 * We put ourselves on the sleep queue and start our timeout
387 	 * before calling CURSIG, as we could stop there, and a wakeup
388 	 * or a SIGCONT (or both) could occur while we were stopped.
389 	 * A SIGCONT would cause us to be marked as SSLEEP
390 	 * without resuming us, thus we must be ready for sleep
391 	 * when CURSIG is called.  If the wakeup happens while we're
392 	 * stopped, td->td_wchan will be 0 upon return from CURSIG.
393 	 */
394 	if (p) {
395 		if (catch) {
396 			p->p_flag |= P_SINTR;
397 			if ((sig = CURSIG(p))) {
398 				if (td->td_wchan) {
399 					unsleep(td);
400 					lwkt_schedule_self(td);
401 				}
402 				p->p_stat = SRUN;
403 				goto resume;
404 			}
405 			if (td->td_wchan == NULL) {
406 				catch = 0;
407 				goto resume;
408 			}
409 		} else {
410 			sig = 0;
411 		}
412 
413 		/*
414 		 * If we are not the current process we have to remove ourself
415 		 * from the run queue.
416 		 */
417 		KASSERT(p->p_stat == SRUN, ("PSTAT NOT SRUN %d %d", p->p_pid, p->p_stat));
418 		/*
419 		 * If this is the current 'user' process schedule another one.
420 		 */
421 		clrrunnable(p, SSLEEP);
422 		p->p_stats->p_ru.ru_nvcsw++;
423 		mi_switch(p);
424 		KASSERT(p->p_stat == SRUN, ("tsleep: stat not srun"));
425 	} else {
426 		lwkt_switch();
427 	}
428 resume:
429 	if (p)
430 		p->p_flag &= ~P_SINTR;
431 	crit_exit_quick(td);
432 	td->td_flags &= ~TDF_NORESCHED;
433 	if (td->td_flags & TDF_TIMEOUT) {
434 		td->td_flags &= ~TDF_TIMEOUT;
435 		if (sig == 0)
436 			return (EWOULDBLOCK);
437 	} else if (timo) {
438 		callout_stop(&thandle);
439 	} else if (td->td_wmesg) {
440 		/*
441 		 * This can happen if a thread is woken up directly.  Clear
442 		 * wmesg to avoid debugging confusion.
443 		 */
444 		td->td_wmesg = NULL;
445 	}
446 	/* inline of iscaught() */
447 	if (p) {
448 		if (catch && (sig != 0 || (sig = CURSIG(p)))) {
449 			if (SIGISMEMBER(p->p_sigacts->ps_sigintr, sig))
450 				return (EINTR);
451 			return (ERESTART);
452 		}
453 	}
454 	return (0);
455 }
456 
457 /*
458  * Implement the timeout for tsleep.  We interlock against
459  * wchan when setting TDF_TIMEOUT.  For processes we remove
460  * the sleep if the process is stopped rather then sleeping,
461  * so it remains stopped.
462  */
463 static void
464 endtsleep(void *arg)
465 {
466 	thread_t td = arg;
467 	struct proc *p;
468 
469 	crit_enter();
470 	if (td->td_wchan) {
471 		td->td_flags |= TDF_TIMEOUT;
472 		if ((p = td->td_proc) != NULL) {
473 			if (p->p_stat == SSLEEP)
474 				setrunnable(p);
475 			else
476 				unsleep(td);
477 		} else {
478 			unsleep(td);
479 			lwkt_schedule(td);
480 		}
481 	}
482 	crit_exit();
483 }
484 
485 /*
486  * Remove a process from its wait queue
487  */
488 void
489 unsleep(struct thread *td)
490 {
491 	crit_enter();
492 	if (td->td_wchan) {
493 #if 0
494 		if (p->p_flag & P_XSLEEP) {
495 			struct xwait *w = p->p_wchan;
496 			TAILQ_REMOVE(&w->waitq, p, p_procq);
497 			p->p_flag &= ~P_XSLEEP;
498 		} else
499 #endif
500 		TAILQ_REMOVE(&slpque[LOOKUP(td->td_wchan)], td, td_threadq);
501 		td->td_wchan = NULL;
502 	}
503 	crit_exit();
504 }
505 
506 #if 0
507 /*
508  * Make all processes sleeping on the explicit lock structure runnable.
509  */
510 void
511 xwakeup(struct xwait *w)
512 {
513 	struct proc *p;
514 
515 	crit_enter();
516 	++w->gen;
517 	while ((p = TAILQ_FIRST(&w->waitq)) != NULL) {
518 		TAILQ_REMOVE(&w->waitq, p, p_procq);
519 		KASSERT(p->p_wchan == w && (p->p_flag & P_XSLEEP),
520 		    ("xwakeup: wchan mismatch for %p (%p/%p) %08x", p, p->p_wchan, w, p->p_flag & P_XSLEEP));
521 		p->p_wchan = NULL;
522 		p->p_flag &= ~P_XSLEEP;
523 		if (p->p_stat == SSLEEP) {
524 			/* OPTIMIZED EXPANSION OF setrunnable(p); */
525 			if (p->p_slptime > 1)
526 				updatepri(p);
527 			p->p_slptime = 0;
528 			p->p_stat = SRUN;
529 			if (p->p_flag & P_INMEM) {
530 				lwkt_schedule(td);
531 			} else {
532 				p->p_flag |= P_SWAPINREQ;
533 				wakeup((caddr_t)&proc0);
534 			}
535 		}
536 	}
537 	crit_exit();
538 }
539 #endif
540 
541 /*
542  * Make all processes sleeping on the specified identifier runnable.
543  */
544 static void
545 _wakeup(void *ident, int domain, int count)
546 {
547 	struct slpquehead *qp;
548 	struct thread *td;
549 	struct thread *ntd;
550 	struct proc *p;
551 	int id = LOOKUP(ident);
552 
553 	crit_enter();
554 	qp = &slpque[id];
555 restart:
556 	for (td = TAILQ_FIRST(qp); td != NULL; td = ntd) {
557 		ntd = TAILQ_NEXT(td, td_threadq);
558 		if (td->td_wchan == ident && td->td_wdomain == domain) {
559 			TAILQ_REMOVE(qp, td, td_threadq);
560 			td->td_wchan = NULL;
561 			if ((p = td->td_proc) != NULL && p->p_stat == SSLEEP) {
562 				/* OPTIMIZED EXPANSION OF setrunnable(p); */
563 				if (p->p_slptime > 1)
564 					updatepri(p);
565 				p->p_slptime = 0;
566 				p->p_stat = SRUN;
567 				if (p->p_flag & P_INMEM) {
568 					/*
569 					 * LWKT scheduled now, there is no
570 					 * userland runq interaction until
571 					 * the thread tries to return to user
572 					 * mode.
573 					 *
574 					 * setrunqueue(p);
575 					 */
576 					lwkt_schedule(td);
577 				} else {
578 					p->p_flag |= P_SWAPINREQ;
579 					wakeup((caddr_t)&proc0);
580 				}
581 				/* END INLINE EXPANSION */
582 			} else if (p == NULL) {
583 				lwkt_schedule(td);
584 			}
585 			if (--count == 0)
586 				break;
587 			goto restart;
588 		}
589 	}
590 	crit_exit();
591 }
592 
593 void
594 wakeup(void *ident)
595 {
596     _wakeup(ident, 0, 0);
597 }
598 
599 void
600 wakeup_one(void *ident)
601 {
602     _wakeup(ident, 0, 1);
603 }
604 
605 void
606 wakeup_domain(void *ident, int domain)
607 {
608     _wakeup(ident, domain, 0);
609 }
610 
611 void
612 wakeup_domain_one(void *ident, int domain)
613 {
614     _wakeup(ident, domain, 1);
615 }
616 
617 /*
618  * The machine independent parts of mi_switch().
619  *
620  * 'p' must be the current process.
621  */
622 void
623 mi_switch(struct proc *p)
624 {
625 	thread_t td = p->p_thread;
626 	struct rlimit *rlim;
627 	u_int64_t ttime;
628 
629 	KKASSERT(td == mycpu->gd_curthread);
630 
631 	crit_enter_quick(td);
632 
633 	/*
634 	 * Check if the process exceeds its cpu resource allocation.
635 	 * If over max, kill it.  Time spent in interrupts is not
636 	 * included.  YYY 64 bit match is expensive.  Ick.
637 	 */
638 	ttime = td->td_sticks + td->td_uticks;
639 	if (p->p_stat != SZOMB && p->p_limit->p_cpulimit != RLIM_INFINITY &&
640 	    ttime > p->p_limit->p_cpulimit) {
641 		rlim = &p->p_rlimit[RLIMIT_CPU];
642 		if (ttime / (rlim_t)1000000 >= rlim->rlim_max) {
643 			killproc(p, "exceeded maximum CPU limit");
644 		} else {
645 			psignal(p, SIGXCPU);
646 			if (rlim->rlim_cur < rlim->rlim_max) {
647 				/* XXX: we should make a private copy */
648 				rlim->rlim_cur += 5;
649 			}
650 		}
651 	}
652 
653 	/*
654 	 * If we are in a SSTOPped state we deschedule ourselves.
655 	 * YYY this needs to be cleaned up, remember that LWKTs stay on
656 	 * their run queue which works differently then the user scheduler
657 	 * which removes the process from the runq when it runs it.
658 	 */
659 	mycpu->gd_cnt.v_swtch++;
660 	if (p->p_stat == SSTOP)
661 		lwkt_deschedule_self(td);
662 	lwkt_switch();
663 	crit_exit_quick(td);
664 }
665 
666 /*
667  * Change process state to be runnable,
668  * placing it on the run queue if it is in memory,
669  * and awakening the swapper if it isn't in memory.
670  */
671 void
672 setrunnable(struct proc *p)
673 {
674 	crit_enter();
675 
676 	switch (p->p_stat) {
677 	case 0:
678 	case SRUN:
679 	case SZOMB:
680 	default:
681 		panic("setrunnable");
682 	case SSTOP:
683 	case SSLEEP:
684 		unsleep(p->p_thread);	/* e.g. when sending signals */
685 		break;
686 
687 	case SIDL:
688 		break;
689 	}
690 	p->p_stat = SRUN;
691 
692 	/*
693 	 * The process is controlled by LWKT at this point, we do not mess
694 	 * around with the userland scheduler until the thread tries to
695 	 * return to user mode.
696 	 */
697 #if 0
698 	if (p->p_flag & P_INMEM)
699 		setrunqueue(p);
700 #endif
701 	if (p->p_flag & P_INMEM)
702 		lwkt_schedule(p->p_thread);
703 	crit_exit();
704 	if (p->p_slptime > 1)
705 		updatepri(p);
706 	p->p_slptime = 0;
707 	if ((p->p_flag & P_INMEM) == 0) {
708 		p->p_flag |= P_SWAPINREQ;
709 		wakeup((caddr_t)&proc0);
710 	}
711 }
712 
713 /*
714  * Change the process state to NOT be runnable, removing it from the run
715  * queue.
716  */
717 void
718 clrrunnable(struct proc *p, int stat)
719 {
720 	crit_enter_quick(p->p_thread);
721 	if (p->p_stat == SRUN && (p->p_flag & P_ONRUNQ))
722 		remrunqueue(p);
723 	p->p_stat = stat;
724 	crit_exit_quick(p->p_thread);
725 }
726 
727 /*
728  * Compute the priority of a process when running in user mode.
729  * Arrange to reschedule if the resulting priority is better
730  * than that of the current process.
731  */
732 void
733 resetpriority(struct proc *p)
734 {
735 	int newpriority;
736 	int interactive;
737 	int opq;
738 	int npq;
739 
740 	/*
741 	 * Set p_priority for general process comparisons
742 	 */
743 	switch(p->p_rtprio.type) {
744 	case RTP_PRIO_REALTIME:
745 		p->p_priority = PRIBASE_REALTIME + p->p_rtprio.prio;
746 		return;
747 	case RTP_PRIO_NORMAL:
748 		break;
749 	case RTP_PRIO_IDLE:
750 		p->p_priority = PRIBASE_IDLE + p->p_rtprio.prio;
751 		return;
752 	case RTP_PRIO_THREAD:
753 		p->p_priority = PRIBASE_THREAD + p->p_rtprio.prio;
754 		return;
755 	}
756 
757 	/*
758 	 * NORMAL priorities fall through.  These are based on niceness
759 	 * and cpu use.  Lower numbers == higher priorities.
760 	 */
761 	newpriority = (int)(NICE_ADJUST(p->p_nice - PRIO_MIN) +
762 			p->p_estcpu / ESTCPURAMP);
763 
764 	/*
765 	 * p_interactive is -128 to +127 and represents very long term
766 	 * interactivity or batch (whereas estcpu is a much faster variable).
767 	 * Interactivity can modify the priority by up to 8 units either way.
768 	 * (8 units == approximately 4 nice levels).
769 	 */
770 	interactive = p->p_interactive / 10;
771 	newpriority += interactive;
772 
773 	newpriority = MIN(newpriority, MAXPRI);
774 	newpriority = MAX(newpriority, 0);
775 	npq = newpriority / PPQ;
776 	crit_enter();
777 	opq = (p->p_priority & PRIMASK) / PPQ;
778 	if (p->p_stat == SRUN && (p->p_flag & P_ONRUNQ) && opq != npq) {
779 		/*
780 		 * We have to move the process to another queue
781 		 */
782 		remrunqueue(p);
783 		p->p_priority = PRIBASE_NORMAL + newpriority;
784 		setrunqueue(p);
785 	} else {
786 		/*
787 		 * We can just adjust the priority and it will be picked
788 		 * up later.
789 		 */
790 		KKASSERT(opq == npq || (p->p_flag & P_ONRUNQ) == 0);
791 		p->p_priority = PRIBASE_NORMAL + newpriority;
792 	}
793 	crit_exit();
794 }
795 
796 /*
797  * Compute a tenex style load average of a quantity on
798  * 1, 5 and 15 minute intervals.
799  */
800 static void
801 loadav(void *arg)
802 {
803 	int i, nrun;
804 	struct loadavg *avg;
805 	struct proc *p;
806 	thread_t td;
807 
808 	avg = &averunnable;
809 	nrun = 0;
810 	FOREACH_PROC_IN_SYSTEM(p) {
811 		switch (p->p_stat) {
812 		case SRUN:
813 			if ((td = p->p_thread) == NULL)
814 				break;
815 			if (td->td_flags & TDF_BLOCKED)
816 				break;
817 			/* fall through */
818 		case SIDL:
819 			nrun++;
820 			break;
821 		default:
822 			break;
823 		}
824 	}
825 	for (i = 0; i < 3; i++)
826 		avg->ldavg[i] = (cexp[i] * avg->ldavg[i] +
827 		    nrun * FSCALE * (FSCALE - cexp[i])) >> FSHIFT;
828 
829 	/*
830 	 * Schedule the next update to occur after 5 seconds, but add a
831 	 * random variation to avoid synchronisation with processes that
832 	 * run at regular intervals.
833 	 */
834 	callout_reset(&loadav_callout, hz * 4 + (int)(random() % (hz * 2 + 1)),
835 	    loadav, NULL);
836 }
837 
838 /* ARGSUSED */
839 static void
840 sched_setup(void *dummy)
841 {
842 	callout_init(&loadav_callout);
843 	callout_init(&roundrobin_callout);
844 	callout_init(&schedcpu_callout);
845 
846 	/* Kick off timeout driven events by calling first time. */
847 	roundrobin(NULL);
848 	schedcpu(NULL);
849 	loadav(NULL);
850 }
851 
852 /*
853  * We adjust the priority of the current process.  The priority of
854  * a process gets worse as it accumulates CPU time.  The cpu usage
855  * estimator (p_estcpu) is increased here.  resetpriority() will
856  * compute a different priority each time p_estcpu increases by
857  * INVERSE_ESTCPU_WEIGHT * (until MAXPRI is reached).
858  *
859  * The cpu usage estimator ramps up quite quickly when the process is
860  * running (linearly), and decays away exponentially, at a rate which
861  * is proportionally slower when the system is busy.  The basic principle
862  * is that the system will 90% forget that the process used a lot of CPU
863  * time in 5 * loadav seconds.  This causes the system to favor processes
864  * which haven't run much recently, and to round-robin among other processes.
865  *
866  * The actual schedulerclock interrupt rate is ESTCPUFREQ, but we generally
867  * want to ramp-up at a faster rate, ESTCPUVFREQ, so p_estcpu is scaled
868  * by (ESTCPUVFREQ / ESTCPUFREQ).  You can control the ramp-up/ramp-down
869  * rate by adjusting ESTCPUVFREQ in sys/proc.h in integer multiples
870  * of ESTCPUFREQ.
871  *
872  * WARNING! called from a fast-int or an IPI, the MP lock MIGHT NOT BE HELD
873  * and we cannot block.
874  */
875 void
876 schedulerclock(void *dummy)
877 {
878 	struct thread *td;
879 	struct proc *p;
880 
881 	td = curthread;
882 	if ((p = td->td_proc) != NULL) {
883 		p->p_cpticks++;		/* cpticks runs at ESTCPUFREQ */
884 		p->p_estcpu = ESTCPULIM(p->p_estcpu + ESTCPUVFREQ / ESTCPUFREQ);
885 		if (try_mplock()) {
886 			resetpriority(p);
887 			rel_mplock();
888 		}
889 	}
890 }
891 
892