Home | History | Annotate | Line # | Download | only in kern
kern_synch.c revision 1.55
      1 /*	$NetBSD: kern_synch.c,v 1.55 1999/02/23 02:56:03 ross Exp $	*/
      2 
      3 /*-
      4  * Copyright (c) 1982, 1986, 1990, 1991, 1993
      5  *	The Regents of the University of California.  All rights reserved.
      6  * (c) UNIX System Laboratories, Inc.
      7  * All or some portions of this file are derived from material licensed
      8  * to the University of California by American Telephone and Telegraph
      9  * Co. or Unix System Laboratories, Inc. and are reproduced herein with
     10  * the permission of UNIX System Laboratories, Inc.
     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  * 3. All advertising materials mentioning features or use of this software
     21  *    must display the following acknowledgement:
     22  *	This product includes software developed by the University of
     23  *	California, Berkeley and its contributors.
     24  * 4. Neither the name of the University nor the names of its contributors
     25  *    may be used to endorse or promote products derived from this software
     26  *    without specific prior written permission.
     27  *
     28  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
     29  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
     30  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
     31  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
     32  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
     33  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
     34  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
     35  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
     36  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
     37  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
     38  * SUCH DAMAGE.
     39  *
     40  *	@(#)kern_synch.c	8.9 (Berkeley) 5/19/95
     41  */
     42 
     43 #include "opt_ddb.h"
     44 #include "opt_ktrace.h"
     45 #include "opt_uvm.h"
     46 
     47 #include <sys/param.h>
     48 #include <sys/systm.h>
     49 #include <sys/proc.h>
     50 #include <sys/kernel.h>
     51 #include <sys/buf.h>
     52 #include <sys/signalvar.h>
     53 #include <sys/resourcevar.h>
     54 #include <vm/vm.h>
     55 #include <sys/sched.h>
     56 
     57 #if defined(UVM)
     58 #include <uvm/uvm_extern.h>
     59 #endif
     60 
     61 #ifdef KTRACE
     62 #include <sys/ktrace.h>
     63 #endif
     64 
     65 #define NICE_WEIGHT 2			/* priorities per nice level */
     66 #define	PPQ	(128 / NQS)		/* priorities per queue */
     67 
     68 #define	ESTCPULIM(e) min((e), NICE_WEIGHT * PRIO_MAX - PPQ)
     69 
     70 #include <machine/cpu.h>
     71 
     72 u_char	curpriority;		/* usrpri of curproc */
     73 int	lbolt;			/* once a second sleep address */
     74 
     75 void roundrobin __P((void *));
     76 void schedcpu __P((void *));
     77 void updatepri __P((struct proc *));
     78 void endtsleep __P((void *));
     79 
     80 /*
     81  * Force switch among equal priority processes every 100ms.
     82  */
     83 /* ARGSUSED */
     84 void
     85 roundrobin(arg)
     86 	void *arg;
     87 {
     88 
     89 	need_resched();
     90 	timeout(roundrobin, NULL, hz / 10);
     91 }
     92 
     93 /*
     94  * Constants for digital decay and forget:
     95  *	90% of (p_estcpu) usage in 5 * loadav time
     96  *	95% of (p_pctcpu) usage in 60 seconds (load insensitive)
     97  *          Note that, as ps(1) mentions, this can let percentages
     98  *          total over 100% (I've seen 137.9% for 3 processes).
     99  *
    100  * Note that hardclock updates p_estcpu and p_cpticks independently.
    101  *
    102  * We wish to decay away 90% of p_estcpu in (5 * loadavg) seconds.
    103  * That is, the system wants to compute a value of decay such
    104  * that the following for loop:
    105  * 	for (i = 0; i < (5 * loadavg); i++)
    106  * 		p_estcpu *= decay;
    107  * will compute
    108  * 	p_estcpu *= 0.1;
    109  * for all values of loadavg:
    110  *
    111  * Mathematically this loop can be expressed by saying:
    112  * 	decay ** (5 * loadavg) ~= .1
    113  *
    114  * The system computes decay as:
    115  * 	decay = (2 * loadavg) / (2 * loadavg + 1)
    116  *
    117  * We wish to prove that the system's computation of decay
    118  * will always fulfill the equation:
    119  * 	decay ** (5 * loadavg) ~= .1
    120  *
    121  * If we compute b as:
    122  * 	b = 2 * loadavg
    123  * then
    124  * 	decay = b / (b + 1)
    125  *
    126  * We now need to prove two things:
    127  *	1) Given factor ** (5 * loadavg) ~= .1, prove factor == b/(b+1)
    128  *	2) Given b/(b+1) ** power ~= .1, prove power == (5 * loadavg)
    129  *
    130  * Facts:
    131  *         For x close to zero, exp(x) =~ 1 + x, since
    132  *              exp(x) = 0! + x**1/1! + x**2/2! + ... .
    133  *              therefore exp(-1/b) =~ 1 - (1/b) = (b-1)/b.
    134  *         For x close to zero, ln(1+x) =~ x, since
    135  *              ln(1+x) = x - x**2/2 + x**3/3 - ...     -1 < x < 1
    136  *              therefore ln(b/(b+1)) = ln(1 - 1/(b+1)) =~ -1/(b+1).
    137  *         ln(.1) =~ -2.30
    138  *
    139  * Proof of (1):
    140  *    Solve (factor)**(power) =~ .1 given power (5*loadav):
    141  *	solving for factor,
    142  *      ln(factor) =~ (-2.30/5*loadav), or
    143  *      factor =~ exp(-1/((5/2.30)*loadav)) =~ exp(-1/(2*loadav)) =
    144  *          exp(-1/b) =~ (b-1)/b =~ b/(b+1).                    QED
    145  *
    146  * Proof of (2):
    147  *    Solve (factor)**(power) =~ .1 given factor == (b/(b+1)):
    148  *	solving for power,
    149  *      power*ln(b/(b+1)) =~ -2.30, or
    150  *      power =~ 2.3 * (b + 1) = 4.6*loadav + 2.3 =~ 5*loadav.  QED
    151  *
    152  * Actual power values for the implemented algorithm are as follows:
    153  *      loadav: 1       2       3       4
    154  *      power:  5.68    10.32   14.94   19.55
    155  */
    156 
    157 /* calculations for digital decay to forget 90% of usage in 5*loadav sec */
    158 #define	loadfactor(loadav)	(2 * (loadav))
    159 #define	decay_cpu(loadfac, cpu)	(((loadfac) * (cpu)) / ((loadfac) + FSCALE))
    160 
    161 /* decay 95% of `p_pctcpu' in 60 seconds; see CCPU_SHIFT before changing */
    162 fixpt_t	ccpu = 0.95122942450071400909 * FSCALE;		/* exp(-1/20) */
    163 
    164 /*
    165  * If `ccpu' is not equal to `exp(-1/20)' and you still want to use the
    166  * faster/more-accurate formula, you'll have to estimate CCPU_SHIFT below
    167  * and possibly adjust FSHIFT in "param.h" so that (FSHIFT >= CCPU_SHIFT).
    168  *
    169  * To estimate CCPU_SHIFT for exp(-1/20), the following formula was used:
    170  *	1 - exp(-1/20) ~= 0.0487 ~= 0.0488 == 1 (fixed pt, *11* bits).
    171  *
    172  * If you dont want to bother with the faster/more-accurate formula, you
    173  * can set CCPU_SHIFT to (FSHIFT + 1) which will use a slower/less-accurate
    174  * (more general) method of calculating the %age of CPU used by a process.
    175  */
    176 #define	CCPU_SHIFT	11
    177 
    178 /*
    179  * Recompute process priorities, every hz ticks.
    180  */
    181 /* ARGSUSED */
    182 void
    183 schedcpu(arg)
    184 	void *arg;
    185 {
    186 	register fixpt_t loadfac = loadfactor(averunnable.ldavg[0]);
    187 	register struct proc *p;
    188 	register int s;
    189 	register unsigned int newcpu;
    190 
    191 	wakeup((caddr_t)&lbolt);
    192 	for (p = allproc.lh_first; p != 0; p = p->p_list.le_next) {
    193 		/*
    194 		 * Increment time in/out of memory and sleep time
    195 		 * (if sleeping).  We ignore overflow; with 16-bit int's
    196 		 * (remember them?) overflow takes 45 days.
    197 		 */
    198 		p->p_swtime++;
    199 		if (p->p_stat == SSLEEP || p->p_stat == SSTOP)
    200 			p->p_slptime++;
    201 		p->p_pctcpu = (p->p_pctcpu * ccpu) >> FSHIFT;
    202 		/*
    203 		 * If the process has slept the entire second,
    204 		 * stop recalculating its priority until it wakes up.
    205 		 */
    206 		if (p->p_slptime > 1)
    207 			continue;
    208 		s = splstatclock();	/* prevent state changes */
    209 		/*
    210 		 * p_pctcpu is only for ps.
    211 		 */
    212 		KASSERT(profhz);
    213 #if	(FSHIFT >= CCPU_SHIFT)
    214 		p->p_pctcpu += (profhz == 100)?
    215 			((fixpt_t) p->p_cpticks) << (FSHIFT - CCPU_SHIFT):
    216                 	100 * (((fixpt_t) p->p_cpticks)
    217 				<< (FSHIFT - CCPU_SHIFT)) / profhz;
    218 #else
    219 		p->p_pctcpu += ((FSCALE - ccpu) *
    220 			(p->p_cpticks * FSCALE / profhz)) >> FSHIFT;
    221 #endif
    222 		p->p_cpticks = 0;
    223 		newcpu = (u_int)decay_cpu(loadfac, p->p_estcpu);
    224 		p->p_estcpu = newcpu;
    225 		resetpriority(p);
    226 		if (p->p_priority >= PUSER) {
    227 			if ((p != curproc) &&
    228 			    p->p_stat == SRUN &&
    229 			    (p->p_flag & P_INMEM) &&
    230 			    (p->p_priority / PPQ) != (p->p_usrpri / PPQ)) {
    231 				remrunqueue(p);
    232 				p->p_priority = p->p_usrpri;
    233 				setrunqueue(p);
    234 			} else
    235 				p->p_priority = p->p_usrpri;
    236 		}
    237 		splx(s);
    238 	}
    239 #if defined(UVM)
    240 	uvm_meter();
    241 #else
    242 	vmmeter();
    243 #endif
    244 	timeout(schedcpu, (void *)0, hz);
    245 }
    246 
    247 /*
    248  * Recalculate the priority of a process after it has slept for a while.
    249  * For all load averages >= 1 and max p_estcpu of 255, sleeping for at
    250  * least six times the loadfactor will decay p_estcpu to zero.
    251  */
    252 void
    253 updatepri(p)
    254 	register struct proc *p;
    255 {
    256 	register unsigned int newcpu = p->p_estcpu;
    257 	register fixpt_t loadfac = loadfactor(averunnable.ldavg[0]);
    258 
    259 	if (p->p_slptime > 5 * loadfac)
    260 		p->p_estcpu = 0;
    261 	else {
    262 		p->p_slptime--;	/* the first time was done in schedcpu */
    263 		while (newcpu && --p->p_slptime)
    264 			newcpu = (int) decay_cpu(loadfac, newcpu);
    265 		p->p_estcpu = newcpu;
    266 	}
    267 	resetpriority(p);
    268 }
    269 
    270 /*
    271  * We're only looking at 7 bits of the address; everything is
    272  * aligned to 4, lots of things are aligned to greater powers
    273  * of 2.  Shift right by 8, i.e. drop the bottom 256 worth.
    274  */
    275 #define TABLESIZE	128
    276 #define LOOKUP(x)	(((long)(x) >> 8) & (TABLESIZE - 1))
    277 struct slpque {
    278 	struct proc *sq_head;
    279 	struct proc **sq_tailp;
    280 } slpque[TABLESIZE];
    281 
    282 /*
    283  * During autoconfiguration or after a panic, a sleep will simply
    284  * lower the priority briefly to allow interrupts, then return.
    285  * The priority to be used (safepri) is machine-dependent, thus this
    286  * value is initialized and maintained in the machine-dependent layers.
    287  * This priority will typically be 0, or the lowest priority
    288  * that is safe for use on the interrupt stack; it can be made
    289  * higher to block network software interrupts after panics.
    290  */
    291 int safepri;
    292 
    293 /*
    294  * General sleep call.  Suspends the current process until a wakeup is
    295  * performed on the specified identifier.  The process will then be made
    296  * runnable with the specified priority.  Sleeps at most timo/hz seconds
    297  * (0 means no timeout).  If pri includes PCATCH flag, signals are checked
    298  * before and after sleeping, else signals are not checked.  Returns 0 if
    299  * awakened, EWOULDBLOCK if the timeout expires.  If PCATCH is set and a
    300  * signal needs to be delivered, ERESTART is returned if the current system
    301  * call should be restarted if possible, and EINTR is returned if the system
    302  * call should be interrupted by the signal (return EINTR).
    303  */
    304 int
    305 tsleep(ident, priority, wmesg, timo)
    306 	void *ident;
    307 	int priority, timo;
    308 	const char *wmesg;
    309 {
    310 	register struct proc *p = curproc;
    311 	register struct slpque *qp;
    312 	register int s;
    313 	int sig, catch = priority & PCATCH;
    314 	extern int cold;
    315 	void endtsleep __P((void *));
    316 
    317 	if (cold || panicstr) {
    318 		/*
    319 		 * After a panic, or during autoconfiguration,
    320 		 * just give interrupts a chance, then just return;
    321 		 * don't run any other procs or panic below,
    322 		 * in case this is the idle process and already asleep.
    323 		 */
    324 		s = splhigh();
    325 		splx(safepri);
    326 		splx(s);
    327 		return (0);
    328 	}
    329 
    330 #ifdef KTRACE
    331 	if (KTRPOINT(p, KTR_CSW))
    332 		ktrcsw(p->p_tracep, 1, 0);
    333 #endif
    334 	s = splhigh();
    335 
    336 #ifdef DIAGNOSTIC
    337 	if (ident == NULL || p->p_stat != SRUN || p->p_back)
    338 		panic("tsleep");
    339 #endif
    340 	p->p_wchan = ident;
    341 	p->p_wmesg = wmesg;
    342 	p->p_slptime = 0;
    343 	p->p_priority = priority & PRIMASK;
    344 	qp = &slpque[LOOKUP(ident)];
    345 	if (qp->sq_head == 0)
    346 		qp->sq_head = p;
    347 	else
    348 		*qp->sq_tailp = p;
    349 	*(qp->sq_tailp = &p->p_forw) = 0;
    350 	if (timo)
    351 		timeout(endtsleep, (void *)p, timo);
    352 	/*
    353 	 * We put ourselves on the sleep queue and start our timeout
    354 	 * before calling CURSIG, as we could stop there, and a wakeup
    355 	 * or a SIGCONT (or both) could occur while we were stopped.
    356 	 * A SIGCONT would cause us to be marked as SSLEEP
    357 	 * without resuming us, thus we must be ready for sleep
    358 	 * when CURSIG is called.  If the wakeup happens while we're
    359 	 * stopped, p->p_wchan will be 0 upon return from CURSIG.
    360 	 */
    361 	if (catch) {
    362 		p->p_flag |= P_SINTR;
    363 		if ((sig = CURSIG(p)) != 0) {
    364 			if (p->p_wchan)
    365 				unsleep(p);
    366 			p->p_stat = SRUN;
    367 			goto resume;
    368 		}
    369 		if (p->p_wchan == 0) {
    370 			catch = 0;
    371 			goto resume;
    372 		}
    373 	} else
    374 		sig = 0;
    375 	p->p_stat = SSLEEP;
    376 	p->p_stats->p_ru.ru_nvcsw++;
    377 	mi_switch();
    378 #ifdef	DDB
    379 	/* handy breakpoint location after process "wakes" */
    380 	asm(".globl bpendtsleep ; bpendtsleep:");
    381 #endif
    382 resume:
    383 	curpriority = p->p_usrpri;
    384 	splx(s);
    385 	p->p_flag &= ~P_SINTR;
    386 	if (p->p_flag & P_TIMEOUT) {
    387 		p->p_flag &= ~P_TIMEOUT;
    388 		if (sig == 0) {
    389 #ifdef KTRACE
    390 			if (KTRPOINT(p, KTR_CSW))
    391 				ktrcsw(p->p_tracep, 0, 0);
    392 #endif
    393 			return (EWOULDBLOCK);
    394 		}
    395 	} else if (timo)
    396 		untimeout(endtsleep, (void *)p);
    397 	if (catch && (sig != 0 || (sig = CURSIG(p)) != 0)) {
    398 #ifdef KTRACE
    399 		if (KTRPOINT(p, KTR_CSW))
    400 			ktrcsw(p->p_tracep, 0, 0);
    401 #endif
    402 		if ((p->p_sigacts->ps_sigact[sig].sa_flags & SA_RESTART) == 0)
    403 			return (EINTR);
    404 		return (ERESTART);
    405 	}
    406 #ifdef KTRACE
    407 	if (KTRPOINT(p, KTR_CSW))
    408 		ktrcsw(p->p_tracep, 0, 0);
    409 #endif
    410 	return (0);
    411 }
    412 
    413 /*
    414  * Implement timeout for tsleep.
    415  * If process hasn't been awakened (wchan non-zero),
    416  * set timeout flag and undo the sleep.  If proc
    417  * is stopped, just unsleep so it will remain stopped.
    418  */
    419 void
    420 endtsleep(arg)
    421 	void *arg;
    422 {
    423 	register struct proc *p;
    424 	int s;
    425 
    426 	p = (struct proc *)arg;
    427 	s = splhigh();
    428 	if (p->p_wchan) {
    429 		if (p->p_stat == SSLEEP)
    430 			setrunnable(p);
    431 		else
    432 			unsleep(p);
    433 		p->p_flag |= P_TIMEOUT;
    434 	}
    435 	splx(s);
    436 }
    437 
    438 /*
    439  * Short-term, non-interruptable sleep.
    440  */
    441 void
    442 sleep(ident, priority)
    443 	void *ident;
    444 	int priority;
    445 {
    446 	register struct proc *p = curproc;
    447 	register struct slpque *qp;
    448 	register int s;
    449 	extern int cold;
    450 
    451 #ifdef DIAGNOSTIC
    452 	if (priority > PZERO) {
    453 		printf("sleep called with priority %d > PZERO, wchan: %p\n",
    454 		    priority, ident);
    455 		panic("old sleep");
    456 	}
    457 #endif
    458 	s = splhigh();
    459 	if (cold || panicstr) {
    460 		/*
    461 		 * After a panic, or during autoconfiguration,
    462 		 * just give interrupts a chance, then just return;
    463 		 * don't run any other procs or panic below,
    464 		 * in case this is the idle process and already asleep.
    465 		 */
    466 		splx(safepri);
    467 		splx(s);
    468 		return;
    469 	}
    470 #ifdef DIAGNOSTIC
    471 	if (ident == NULL || p->p_stat != SRUN || p->p_back)
    472 		panic("sleep");
    473 #endif
    474 	p->p_wchan = ident;
    475 	p->p_wmesg = NULL;
    476 	p->p_slptime = 0;
    477 	p->p_priority = priority;
    478 	qp = &slpque[LOOKUP(ident)];
    479 	if (qp->sq_head == 0)
    480 		qp->sq_head = p;
    481 	else
    482 		*qp->sq_tailp = p;
    483 	*(qp->sq_tailp = &p->p_forw) = 0;
    484 	p->p_stat = SSLEEP;
    485 	p->p_stats->p_ru.ru_nvcsw++;
    486 #ifdef KTRACE
    487 	if (KTRPOINT(p, KTR_CSW))
    488 		ktrcsw(p->p_tracep, 1, 0);
    489 #endif
    490 	mi_switch();
    491 #ifdef	DDB
    492 	/* handy breakpoint location after process "wakes" */
    493 	asm(".globl bpendsleep ; bpendsleep:");
    494 #endif
    495 #ifdef KTRACE
    496 	if (KTRPOINT(p, KTR_CSW))
    497 		ktrcsw(p->p_tracep, 0, 0);
    498 #endif
    499 	curpriority = p->p_usrpri;
    500 	splx(s);
    501 }
    502 
    503 /*
    504  * Remove a process from its wait queue
    505  */
    506 void
    507 unsleep(p)
    508 	register struct proc *p;
    509 {
    510 	register struct slpque *qp;
    511 	register struct proc **hp;
    512 	int s;
    513 
    514 	s = splhigh();
    515 	if (p->p_wchan) {
    516 		hp = &(qp = &slpque[LOOKUP(p->p_wchan)])->sq_head;
    517 		while (*hp != p)
    518 			hp = &(*hp)->p_forw;
    519 		*hp = p->p_forw;
    520 		if (qp->sq_tailp == &p->p_forw)
    521 			qp->sq_tailp = hp;
    522 		p->p_wchan = 0;
    523 	}
    524 	splx(s);
    525 }
    526 
    527 /*
    528  * Make all processes sleeping on the specified identifier runnable.
    529  */
    530 void
    531 wakeup(ident)
    532 	register void *ident;
    533 {
    534 	register struct slpque *qp;
    535 	register struct proc *p, **q;
    536 	int s;
    537 
    538 	s = splhigh();
    539 	qp = &slpque[LOOKUP(ident)];
    540 restart:
    541 	for (q = &qp->sq_head; (p = *q) != NULL; ) {
    542 #ifdef DIAGNOSTIC
    543 		if (p->p_back || (p->p_stat != SSLEEP && p->p_stat != SSTOP))
    544 			panic("wakeup");
    545 #endif
    546 		if (p->p_wchan == ident) {
    547 			p->p_wchan = 0;
    548 			*q = p->p_forw;
    549 			if (qp->sq_tailp == &p->p_forw)
    550 				qp->sq_tailp = q;
    551 			if (p->p_stat == SSLEEP) {
    552 				/* OPTIMIZED EXPANSION OF setrunnable(p); */
    553 				if (p->p_slptime > 1)
    554 					updatepri(p);
    555 				p->p_slptime = 0;
    556 				p->p_stat = SRUN;
    557 				if (p->p_flag & P_INMEM)
    558 					setrunqueue(p);
    559 				/*
    560 				 * Since curpriority is a user priority,
    561 				 * p->p_priority is always better than
    562 				 * curpriority.
    563 				 */
    564 				if ((p->p_flag & P_INMEM) == 0)
    565 					wakeup((caddr_t)&proc0);
    566 				else
    567 					need_resched();
    568 				/* END INLINE EXPANSION */
    569 				goto restart;
    570 			}
    571 		} else
    572 			q = &p->p_forw;
    573 	}
    574 	splx(s);
    575 }
    576 
    577 /*
    578  * The machine independent parts of mi_switch().
    579  * Must be called at splstatclock() or higher.
    580  */
    581 void
    582 mi_switch()
    583 {
    584 	register struct proc *p = curproc;	/* XXX */
    585 	register struct rlimit *rlim;
    586 	register long s, u;
    587 	struct timeval tv;
    588 
    589 #ifdef DEBUG
    590 	if (p->p_simple_locks) {
    591 		printf("p->p_simple_locks %d\n", p->p_simple_locks);
    592 #ifdef LOCKDEBUG
    593 		simple_lock_dump();
    594 #endif
    595 		panic("sleep: holding simple lock");
    596 	}
    597 #endif
    598 	/*
    599 	 * Compute the amount of time during which the current
    600 	 * process was running, and add that to its total so far.
    601 	 */
    602 	microtime(&tv);
    603 	u = p->p_rtime.tv_usec + (tv.tv_usec - runtime.tv_usec);
    604 	s = p->p_rtime.tv_sec + (tv.tv_sec - runtime.tv_sec);
    605 	if (u < 0) {
    606 		u += 1000000;
    607 		s--;
    608 	} else if (u >= 1000000) {
    609 		u -= 1000000;
    610 		s++;
    611 	}
    612 	p->p_rtime.tv_usec = u;
    613 	p->p_rtime.tv_sec = s;
    614 
    615 	/*
    616 	 * Check if the process exceeds its cpu resource allocation.
    617 	 * If over max, kill it.  In any case, if it has run for more
    618 	 * than 10 minutes, reduce priority to give others a chance.
    619 	 */
    620 	rlim = &p->p_rlimit[RLIMIT_CPU];
    621 	if (s >= rlim->rlim_cur) {
    622 		if (s >= rlim->rlim_max)
    623 			psignal(p, SIGKILL);
    624 		else {
    625 			psignal(p, SIGXCPU);
    626 			if (rlim->rlim_cur < rlim->rlim_max)
    627 				rlim->rlim_cur += 5;
    628 		}
    629 	}
    630 	if (autonicetime && s > autonicetime && p->p_ucred->cr_uid && p->p_nice == NZERO) {
    631 		p->p_nice = autoniceval + NZERO;
    632 		resetpriority(p);
    633 	}
    634 
    635 	/*
    636 	 * Pick a new current process and record its start time.
    637 	 */
    638 #if defined(UVM)
    639 	uvmexp.swtch++;
    640 #else
    641 	cnt.v_swtch++;
    642 #endif
    643 	cpu_switch(p);
    644 	microtime(&runtime);
    645 }
    646 
    647 /*
    648  * Initialize the (doubly-linked) run queues
    649  * to be empty.
    650  */
    651 void
    652 rqinit()
    653 {
    654 	register int i;
    655 
    656 	for (i = 0; i < NQS; i++)
    657 		qs[i].ph_link = qs[i].ph_rlink = (struct proc *)&qs[i];
    658 }
    659 
    660 /*
    661  * Change process state to be runnable,
    662  * placing it on the run queue if it is in memory,
    663  * and awakening the swapper if it isn't in memory.
    664  */
    665 void
    666 setrunnable(p)
    667 	register struct proc *p;
    668 {
    669 	register int s;
    670 
    671 	s = splhigh();
    672 	switch (p->p_stat) {
    673 	case 0:
    674 	case SRUN:
    675 	case SZOMB:
    676 	default:
    677 		panic("setrunnable");
    678 	case SSTOP:
    679 		/*
    680 		 * If we're being traced (possibly because someone attached us
    681 		 * while we were stopped), check for a signal from the debugger.
    682 		 */
    683 		if ((p->p_flag & P_TRACED) != 0 && p->p_xstat != 0) {
    684 			sigaddset(&p->p_siglist, p->p_xstat);
    685 			p->p_sigcheck = 1;
    686 		}
    687 	case SSLEEP:
    688 		unsleep(p);		/* e.g. when sending signals */
    689 		break;
    690 
    691 	case SIDL:
    692 		break;
    693 	}
    694 	p->p_stat = SRUN;
    695 	if (p->p_flag & P_INMEM)
    696 		setrunqueue(p);
    697 	splx(s);
    698 	if (p->p_slptime > 1)
    699 		updatepri(p);
    700 	p->p_slptime = 0;
    701 	if ((p->p_flag & P_INMEM) == 0)
    702 		wakeup((caddr_t)&proc0);
    703 	else if (p->p_priority < curpriority)
    704 		need_resched();
    705 }
    706 
    707 /*
    708  * Compute the priority of a process when running in user mode.
    709  * Arrange to reschedule if the resulting priority is better
    710  * than that of the current process.
    711  */
    712 void
    713 resetpriority(p)
    714 	register struct proc *p;
    715 {
    716 	register unsigned int newpriority;
    717 
    718 	newpriority = PUSER + p->p_estcpu + NICE_WEIGHT * (p->p_nice - NZERO);
    719 	newpriority = min(newpriority, MAXPRI);
    720 	p->p_usrpri = newpriority;
    721 	if (newpriority < curpriority)
    722 		need_resched();
    723 }
    724 
    725 /*
    726  * We adjust the priority of the current process.  The priority of
    727  * a process gets worse as it accumulates CPU time.  The cpu usage
    728  * estimator (p_estcpu) is increased here.  The formula for computing
    729  * priorities (in kern_synch.c) will compute a different value each
    730  * time p_estcpu increases by 4.  The cpu usage estimator ramps up
    731  * quite quickly when the process is running (linearly), and decays
    732  * away exponentially, at a rate which is proportionally slower when
    733  * the system is busy.  The basic principal is that the system will
    734  * 90% forget that the process used a lot of CPU time in 5 * loadav
    735  * seconds.  This causes the system to favor processes which haven't
    736  * run much recently, and to round-robin among other processes.
    737  */
    738 
    739 void
    740 schedclk(p)
    741 	struct proc *p;
    742 {
    743 	p->p_estcpu = ESTCPULIM(p->p_estcpu + 1);
    744 	resetpriority(p);
    745 	if (p->p_priority >= PUSER)
    746 		p->p_priority = p->p_usrpri;
    747 }
    748