Home | History | Annotate | Line # | Download | only in kern
kern_synch.c revision 1.57.2.1
      1 /*	$NetBSD: kern_synch.c,v 1.57.2.1 1999/10/17 22:29:40 cgd 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 
     46 #include <sys/param.h>
     47 #include <sys/systm.h>
     48 #include <sys/proc.h>
     49 #include <sys/kernel.h>
     50 #include <sys/buf.h>
     51 #include <sys/signalvar.h>
     52 #include <sys/resourcevar.h>
     53 #include <vm/vm.h>
     54 #include <sys/sched.h>
     55 
     56 #include <uvm/uvm_extern.h>
     57 
     58 #ifdef KTRACE
     59 #include <sys/ktrace.h>
     60 #endif
     61 
     62 #define NICE_WEIGHT 2			/* priorities per nice level */
     63 #define	PPQ	(128 / NQS)		/* priorities per queue */
     64 
     65 #define	ESTCPULIM(e) min((e), NICE_WEIGHT * PRIO_MAX - PPQ)
     66 
     67 #include <machine/cpu.h>
     68 
     69 u_char	curpriority;		/* usrpri of curproc */
     70 int	lbolt;			/* once a second sleep address */
     71 
     72 void roundrobin __P((void *));
     73 void schedcpu __P((void *));
     74 void updatepri __P((struct proc *));
     75 void endtsleep __P((void *));
     76 
     77 /*
     78  * Force switch among equal priority processes every 100ms.
     79  */
     80 /* ARGSUSED */
     81 void
     82 roundrobin(arg)
     83 	void *arg;
     84 {
     85 
     86 	need_resched();
     87 	timeout(roundrobin, NULL, hz / 10);
     88 }
     89 
     90 /*
     91  * Constants for digital decay and forget:
     92  *	90% of (p_estcpu) usage in 5 * loadav time
     93  *	95% of (p_pctcpu) usage in 60 seconds (load insensitive)
     94  *          Note that, as ps(1) mentions, this can let percentages
     95  *          total over 100% (I've seen 137.9% for 3 processes).
     96  *
     97  * Note that hardclock updates p_estcpu and p_cpticks independently.
     98  *
     99  * We wish to decay away 90% of p_estcpu in (5 * loadavg) seconds.
    100  * That is, the system wants to compute a value of decay such
    101  * that the following for loop:
    102  * 	for (i = 0; i < (5 * loadavg); i++)
    103  * 		p_estcpu *= decay;
    104  * will compute
    105  * 	p_estcpu *= 0.1;
    106  * for all values of loadavg:
    107  *
    108  * Mathematically this loop can be expressed by saying:
    109  * 	decay ** (5 * loadavg) ~= .1
    110  *
    111  * The system computes decay as:
    112  * 	decay = (2 * loadavg) / (2 * loadavg + 1)
    113  *
    114  * We wish to prove that the system's computation of decay
    115  * will always fulfill the equation:
    116  * 	decay ** (5 * loadavg) ~= .1
    117  *
    118  * If we compute b as:
    119  * 	b = 2 * loadavg
    120  * then
    121  * 	decay = b / (b + 1)
    122  *
    123  * We now need to prove two things:
    124  *	1) Given factor ** (5 * loadavg) ~= .1, prove factor == b/(b+1)
    125  *	2) Given b/(b+1) ** power ~= .1, prove power == (5 * loadavg)
    126  *
    127  * Facts:
    128  *         For x close to zero, exp(x) =~ 1 + x, since
    129  *              exp(x) = 0! + x**1/1! + x**2/2! + ... .
    130  *              therefore exp(-1/b) =~ 1 - (1/b) = (b-1)/b.
    131  *         For x close to zero, ln(1+x) =~ x, since
    132  *              ln(1+x) = x - x**2/2 + x**3/3 - ...     -1 < x < 1
    133  *              therefore ln(b/(b+1)) = ln(1 - 1/(b+1)) =~ -1/(b+1).
    134  *         ln(.1) =~ -2.30
    135  *
    136  * Proof of (1):
    137  *    Solve (factor)**(power) =~ .1 given power (5*loadav):
    138  *	solving for factor,
    139  *      ln(factor) =~ (-2.30/5*loadav), or
    140  *      factor =~ exp(-1/((5/2.30)*loadav)) =~ exp(-1/(2*loadav)) =
    141  *          exp(-1/b) =~ (b-1)/b =~ b/(b+1).                    QED
    142  *
    143  * Proof of (2):
    144  *    Solve (factor)**(power) =~ .1 given factor == (b/(b+1)):
    145  *	solving for power,
    146  *      power*ln(b/(b+1)) =~ -2.30, or
    147  *      power =~ 2.3 * (b + 1) = 4.6*loadav + 2.3 =~ 5*loadav.  QED
    148  *
    149  * Actual power values for the implemented algorithm are as follows:
    150  *      loadav: 1       2       3       4
    151  *      power:  5.68    10.32   14.94   19.55
    152  */
    153 
    154 /* calculations for digital decay to forget 90% of usage in 5*loadav sec */
    155 #define	loadfactor(loadav)	(2 * (loadav))
    156 #define	decay_cpu(loadfac, cpu)	(((loadfac) * (cpu)) / ((loadfac) + FSCALE))
    157 
    158 /* decay 95% of `p_pctcpu' in 60 seconds; see CCPU_SHIFT before changing */
    159 fixpt_t	ccpu = 0.95122942450071400909 * FSCALE;		/* exp(-1/20) */
    160 
    161 /*
    162  * If `ccpu' is not equal to `exp(-1/20)' and you still want to use the
    163  * faster/more-accurate formula, you'll have to estimate CCPU_SHIFT below
    164  * and possibly adjust FSHIFT in "param.h" so that (FSHIFT >= CCPU_SHIFT).
    165  *
    166  * To estimate CCPU_SHIFT for exp(-1/20), the following formula was used:
    167  *	1 - exp(-1/20) ~= 0.0487 ~= 0.0488 == 1 (fixed pt, *11* bits).
    168  *
    169  * If you dont want to bother with the faster/more-accurate formula, you
    170  * can set CCPU_SHIFT to (FSHIFT + 1) which will use a slower/less-accurate
    171  * (more general) method of calculating the %age of CPU used by a process.
    172  */
    173 #define	CCPU_SHIFT	11
    174 
    175 /*
    176  * Recompute process priorities, every hz ticks.
    177  */
    178 /* ARGSUSED */
    179 void
    180 schedcpu(arg)
    181 	void *arg;
    182 {
    183 	register fixpt_t loadfac = loadfactor(averunnable.ldavg[0]);
    184 	register struct proc *p;
    185 	register int s;
    186 	register unsigned int newcpu;
    187 	int clkhz;
    188 
    189 	wakeup((caddr_t)&lbolt);
    190 	for (p = allproc.lh_first; p != 0; p = p->p_list.le_next) {
    191 		/*
    192 		 * Increment time in/out of memory and sleep time
    193 		 * (if sleeping).  We ignore overflow; with 16-bit int's
    194 		 * (remember them?) overflow takes 45 days.
    195 		 */
    196 		p->p_swtime++;
    197 		if (p->p_stat == SSLEEP || p->p_stat == SSTOP)
    198 			p->p_slptime++;
    199 		p->p_pctcpu = (p->p_pctcpu * ccpu) >> FSHIFT;
    200 		/*
    201 		 * If the process has slept the entire second,
    202 		 * stop recalculating its priority until it wakes up.
    203 		 */
    204 		if (p->p_slptime > 1)
    205 			continue;
    206 		s = splstatclock();	/* prevent state changes */
    207 		/*
    208 		 * p_pctcpu is only for ps.
    209 		 */
    210 		clkhz = stathz != 0 ? stathz : hz;
    211 #if	(FSHIFT >= CCPU_SHIFT)
    212 		p->p_pctcpu += (clkhz == 100)?
    213 			((fixpt_t) p->p_cpticks) << (FSHIFT - CCPU_SHIFT):
    214                 	100 * (((fixpt_t) p->p_cpticks)
    215 				<< (FSHIFT - CCPU_SHIFT)) / clkhz;
    216 #else
    217 		p->p_pctcpu += ((FSCALE - ccpu) *
    218 			(p->p_cpticks * FSCALE / clkhz)) >> FSHIFT;
    219 #endif
    220 		p->p_cpticks = 0;
    221 		newcpu = (u_int)decay_cpu(loadfac, p->p_estcpu);
    222 		p->p_estcpu = newcpu;
    223 		resetpriority(p);
    224 		if (p->p_priority >= PUSER) {
    225 			if ((p != curproc) &&
    226 			    p->p_stat == SRUN &&
    227 			    (p->p_flag & P_INMEM) &&
    228 			    (p->p_priority / PPQ) != (p->p_usrpri / PPQ)) {
    229 				remrunqueue(p);
    230 				p->p_priority = p->p_usrpri;
    231 				setrunqueue(p);
    232 			} else
    233 				p->p_priority = p->p_usrpri;
    234 		}
    235 		splx(s);
    236 	}
    237 	uvm_meter();
    238 	timeout(schedcpu, (void *)0, hz);
    239 }
    240 
    241 /*
    242  * Recalculate the priority of a process after it has slept for a while.
    243  * For all load averages >= 1 and max p_estcpu of 255, sleeping for at
    244  * least six times the loadfactor will decay p_estcpu to zero.
    245  */
    246 void
    247 updatepri(p)
    248 	register struct proc *p;
    249 {
    250 	register unsigned int newcpu = p->p_estcpu;
    251 	register fixpt_t loadfac = loadfactor(averunnable.ldavg[0]);
    252 
    253 	if (p->p_slptime > 5 * loadfac)
    254 		p->p_estcpu = 0;
    255 	else {
    256 		p->p_slptime--;	/* the first time was done in schedcpu */
    257 		while (newcpu && --p->p_slptime)
    258 			newcpu = (int) decay_cpu(loadfac, newcpu);
    259 		p->p_estcpu = newcpu;
    260 	}
    261 	resetpriority(p);
    262 }
    263 
    264 /*
    265  * We're only looking at 7 bits of the address; everything is
    266  * aligned to 4, lots of things are aligned to greater powers
    267  * of 2.  Shift right by 8, i.e. drop the bottom 256 worth.
    268  */
    269 #define TABLESIZE	128
    270 #define LOOKUP(x)	(((long)(x) >> 8) & (TABLESIZE - 1))
    271 struct slpque {
    272 	struct proc *sq_head;
    273 	struct proc **sq_tailp;
    274 } slpque[TABLESIZE];
    275 
    276 /*
    277  * During autoconfiguration or after a panic, a sleep will simply
    278  * lower the priority briefly to allow interrupts, then return.
    279  * The priority to be used (safepri) is machine-dependent, thus this
    280  * value is initialized and maintained in the machine-dependent layers.
    281  * This priority will typically be 0, or the lowest priority
    282  * that is safe for use on the interrupt stack; it can be made
    283  * higher to block network software interrupts after panics.
    284  */
    285 int safepri;
    286 
    287 /*
    288  * General sleep call.  Suspends the current process until a wakeup is
    289  * performed on the specified identifier.  The process will then be made
    290  * runnable with the specified priority.  Sleeps at most timo/hz seconds
    291  * (0 means no timeout).  If pri includes PCATCH flag, signals are checked
    292  * before and after sleeping, else signals are not checked.  Returns 0 if
    293  * awakened, EWOULDBLOCK if the timeout expires.  If PCATCH is set and a
    294  * signal needs to be delivered, ERESTART is returned if the current system
    295  * call should be restarted if possible, and EINTR is returned if the system
    296  * call should be interrupted by the signal (return EINTR).
    297  */
    298 int
    299 tsleep(ident, priority, wmesg, timo)
    300 	void *ident;
    301 	int priority, timo;
    302 	const char *wmesg;
    303 {
    304 	register struct proc *p = curproc;
    305 	register struct slpque *qp;
    306 	register int s;
    307 	int sig, catch = priority & PCATCH;
    308 	extern int cold;
    309 	void endtsleep __P((void *));
    310 
    311 	if (cold || panicstr) {
    312 		/*
    313 		 * After a panic, or during autoconfiguration,
    314 		 * just give interrupts a chance, then just return;
    315 		 * don't run any other procs or panic below,
    316 		 * in case this is the idle process and already asleep.
    317 		 */
    318 		s = splhigh();
    319 		splx(safepri);
    320 		splx(s);
    321 		return (0);
    322 	}
    323 
    324 #ifdef KTRACE
    325 	if (KTRPOINT(p, KTR_CSW))
    326 		ktrcsw(p->p_tracep, 1, 0);
    327 #endif
    328 	s = splhigh();
    329 
    330 #ifdef DIAGNOSTIC
    331 	if (ident == NULL || p->p_stat != SRUN || p->p_back)
    332 		panic("tsleep");
    333 #endif
    334 	p->p_wchan = ident;
    335 	p->p_wmesg = wmesg;
    336 	p->p_slptime = 0;
    337 	p->p_priority = priority & PRIMASK;
    338 	qp = &slpque[LOOKUP(ident)];
    339 	if (qp->sq_head == 0)
    340 		qp->sq_head = p;
    341 	else
    342 		*qp->sq_tailp = p;
    343 	*(qp->sq_tailp = &p->p_forw) = 0;
    344 	if (timo)
    345 		timeout(endtsleep, (void *)p, timo);
    346 	/*
    347 	 * We put ourselves on the sleep queue and start our timeout
    348 	 * before calling CURSIG, as we could stop there, and a wakeup
    349 	 * or a SIGCONT (or both) could occur while we were stopped.
    350 	 * A SIGCONT would cause us to be marked as SSLEEP
    351 	 * without resuming us, thus we must be ready for sleep
    352 	 * when CURSIG is called.  If the wakeup happens while we're
    353 	 * stopped, p->p_wchan will be 0 upon return from CURSIG.
    354 	 */
    355 	if (catch) {
    356 		p->p_flag |= P_SINTR;
    357 		if ((sig = CURSIG(p)) != 0) {
    358 			if (p->p_wchan)
    359 				unsleep(p);
    360 			p->p_stat = SRUN;
    361 			goto resume;
    362 		}
    363 		if (p->p_wchan == 0) {
    364 			catch = 0;
    365 			goto resume;
    366 		}
    367 	} else
    368 		sig = 0;
    369 	p->p_stat = SSLEEP;
    370 	p->p_stats->p_ru.ru_nvcsw++;
    371 	mi_switch();
    372 #ifdef	DDB
    373 	/* handy breakpoint location after process "wakes" */
    374 	asm(".globl bpendtsleep ; bpendtsleep:");
    375 #endif
    376 resume:
    377 	curpriority = p->p_usrpri;
    378 	splx(s);
    379 	p->p_flag &= ~P_SINTR;
    380 	if (p->p_flag & P_TIMEOUT) {
    381 		p->p_flag &= ~P_TIMEOUT;
    382 		if (sig == 0) {
    383 #ifdef KTRACE
    384 			if (KTRPOINT(p, KTR_CSW))
    385 				ktrcsw(p->p_tracep, 0, 0);
    386 #endif
    387 			return (EWOULDBLOCK);
    388 		}
    389 	} else if (timo)
    390 		untimeout(endtsleep, (void *)p);
    391 	if (catch && (sig != 0 || (sig = CURSIG(p)) != 0)) {
    392 #ifdef KTRACE
    393 		if (KTRPOINT(p, KTR_CSW))
    394 			ktrcsw(p->p_tracep, 0, 0);
    395 #endif
    396 		if ((p->p_sigacts->ps_sigact[sig].sa_flags & SA_RESTART) == 0)
    397 			return (EINTR);
    398 		return (ERESTART);
    399 	}
    400 #ifdef KTRACE
    401 	if (KTRPOINT(p, KTR_CSW))
    402 		ktrcsw(p->p_tracep, 0, 0);
    403 #endif
    404 	return (0);
    405 }
    406 
    407 /*
    408  * Implement timeout for tsleep.
    409  * If process hasn't been awakened (wchan non-zero),
    410  * set timeout flag and undo the sleep.  If proc
    411  * is stopped, just unsleep so it will remain stopped.
    412  */
    413 void
    414 endtsleep(arg)
    415 	void *arg;
    416 {
    417 	register struct proc *p;
    418 	int s;
    419 
    420 	p = (struct proc *)arg;
    421 	s = splhigh();
    422 	if (p->p_wchan) {
    423 		if (p->p_stat == SSLEEP)
    424 			setrunnable(p);
    425 		else
    426 			unsleep(p);
    427 		p->p_flag |= P_TIMEOUT;
    428 	}
    429 	splx(s);
    430 }
    431 
    432 /*
    433  * Short-term, non-interruptable sleep.
    434  */
    435 void
    436 sleep(ident, priority)
    437 	void *ident;
    438 	int priority;
    439 {
    440 	register struct proc *p = curproc;
    441 	register struct slpque *qp;
    442 	register int s;
    443 	extern int cold;
    444 
    445 #ifdef DIAGNOSTIC
    446 	if (priority > PZERO) {
    447 		printf("sleep called with priority %d > PZERO, wchan: %p\n",
    448 		    priority, ident);
    449 		panic("old sleep");
    450 	}
    451 #endif
    452 	s = splhigh();
    453 	if (cold || panicstr) {
    454 		/*
    455 		 * After a panic, or during autoconfiguration,
    456 		 * just give interrupts a chance, then just return;
    457 		 * don't run any other procs or panic below,
    458 		 * in case this is the idle process and already asleep.
    459 		 */
    460 		splx(safepri);
    461 		splx(s);
    462 		return;
    463 	}
    464 #ifdef DIAGNOSTIC
    465 	if (ident == NULL || p->p_stat != SRUN || p->p_back)
    466 		panic("sleep");
    467 #endif
    468 	p->p_wchan = ident;
    469 	p->p_wmesg = NULL;
    470 	p->p_slptime = 0;
    471 	p->p_priority = priority;
    472 	qp = &slpque[LOOKUP(ident)];
    473 	if (qp->sq_head == 0)
    474 		qp->sq_head = p;
    475 	else
    476 		*qp->sq_tailp = p;
    477 	*(qp->sq_tailp = &p->p_forw) = 0;
    478 	p->p_stat = SSLEEP;
    479 	p->p_stats->p_ru.ru_nvcsw++;
    480 #ifdef KTRACE
    481 	if (KTRPOINT(p, KTR_CSW))
    482 		ktrcsw(p->p_tracep, 1, 0);
    483 #endif
    484 	mi_switch();
    485 #ifdef	DDB
    486 	/* handy breakpoint location after process "wakes" */
    487 	asm(".globl bpendsleep ; bpendsleep:");
    488 #endif
    489 #ifdef KTRACE
    490 	if (KTRPOINT(p, KTR_CSW))
    491 		ktrcsw(p->p_tracep, 0, 0);
    492 #endif
    493 	curpriority = p->p_usrpri;
    494 	splx(s);
    495 }
    496 
    497 /*
    498  * Remove a process from its wait queue
    499  */
    500 void
    501 unsleep(p)
    502 	register struct proc *p;
    503 {
    504 	register struct slpque *qp;
    505 	register struct proc **hp;
    506 	int s;
    507 
    508 	s = splhigh();
    509 	if (p->p_wchan) {
    510 		hp = &(qp = &slpque[LOOKUP(p->p_wchan)])->sq_head;
    511 		while (*hp != p)
    512 			hp = &(*hp)->p_forw;
    513 		*hp = p->p_forw;
    514 		if (qp->sq_tailp == &p->p_forw)
    515 			qp->sq_tailp = hp;
    516 		p->p_wchan = 0;
    517 	}
    518 	splx(s);
    519 }
    520 
    521 /*
    522  * Make all processes sleeping on the specified identifier runnable.
    523  */
    524 void
    525 wakeup(ident)
    526 	register void *ident;
    527 {
    528 	register struct slpque *qp;
    529 	register struct proc *p, **q;
    530 	int s;
    531 
    532 	s = splhigh();
    533 	qp = &slpque[LOOKUP(ident)];
    534 restart:
    535 	for (q = &qp->sq_head; (p = *q) != NULL; ) {
    536 #ifdef DIAGNOSTIC
    537 		if (p->p_back || (p->p_stat != SSLEEP && p->p_stat != SSTOP))
    538 			panic("wakeup");
    539 #endif
    540 		if (p->p_wchan == ident) {
    541 			p->p_wchan = 0;
    542 			*q = p->p_forw;
    543 			if (qp->sq_tailp == &p->p_forw)
    544 				qp->sq_tailp = q;
    545 			if (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 				/*
    554 				 * Since curpriority is a user priority,
    555 				 * p->p_priority is always better than
    556 				 * curpriority.
    557 				 */
    558 				if ((p->p_flag & P_INMEM) == 0)
    559 					wakeup((caddr_t)&proc0);
    560 				else
    561 					need_resched();
    562 				/* END INLINE EXPANSION */
    563 				goto restart;
    564 			}
    565 		} else
    566 			q = &p->p_forw;
    567 	}
    568 	splx(s);
    569 }
    570 
    571 /*
    572  * The machine independent parts of mi_switch().
    573  * Must be called at splstatclock() or higher.
    574  */
    575 void
    576 mi_switch()
    577 {
    578 	register struct proc *p = curproc;	/* XXX */
    579 	register struct rlimit *rlim;
    580 	register long s, u;
    581 	struct timeval tv;
    582 
    583 #ifdef DEBUG
    584 	if (p->p_simple_locks) {
    585 		printf("p->p_simple_locks %d\n", p->p_simple_locks);
    586 #ifdef LOCKDEBUG
    587 		simple_lock_dump();
    588 #endif
    589 		panic("sleep: holding simple lock");
    590 	}
    591 #endif
    592 	/*
    593 	 * Compute the amount of time during which the current
    594 	 * process was running, and add that to its total so far.
    595 	 */
    596 	microtime(&tv);
    597 	u = p->p_rtime.tv_usec + (tv.tv_usec - runtime.tv_usec);
    598 	s = p->p_rtime.tv_sec + (tv.tv_sec - runtime.tv_sec);
    599 	if (u < 0) {
    600 		u += 1000000;
    601 		s--;
    602 	} else if (u >= 1000000) {
    603 		u -= 1000000;
    604 		s++;
    605 	}
    606 	p->p_rtime.tv_usec = u;
    607 	p->p_rtime.tv_sec = s;
    608 
    609 	/*
    610 	 * Check if the process exceeds its cpu resource allocation.
    611 	 * If over max, kill it.  In any case, if it has run for more
    612 	 * than 10 minutes, reduce priority to give others a chance.
    613 	 */
    614 	rlim = &p->p_rlimit[RLIMIT_CPU];
    615 	if (s >= rlim->rlim_cur) {
    616 		if (s >= rlim->rlim_max)
    617 			psignal(p, SIGKILL);
    618 		else {
    619 			psignal(p, SIGXCPU);
    620 			if (rlim->rlim_cur < rlim->rlim_max)
    621 				rlim->rlim_cur += 5;
    622 		}
    623 	}
    624 	if (autonicetime && s > autonicetime && p->p_ucred->cr_uid && p->p_nice == NZERO) {
    625 		p->p_nice = autoniceval + NZERO;
    626 		resetpriority(p);
    627 	}
    628 
    629 	/*
    630 	 * Pick a new current process and record its start time.
    631 	 */
    632 	uvmexp.swtch++;
    633 	cpu_switch(p);
    634 	microtime(&runtime);
    635 }
    636 
    637 /*
    638  * Initialize the (doubly-linked) run queues
    639  * to be empty.
    640  */
    641 void
    642 rqinit()
    643 {
    644 	register int i;
    645 
    646 	for (i = 0; i < NQS; i++)
    647 		qs[i].ph_link = qs[i].ph_rlink = (struct proc *)&qs[i];
    648 }
    649 
    650 /*
    651  * Change process state to be runnable,
    652  * placing it on the run queue if it is in memory,
    653  * and awakening the swapper if it isn't in memory.
    654  */
    655 void
    656 setrunnable(p)
    657 	register struct proc *p;
    658 {
    659 	register int s;
    660 
    661 	s = splhigh();
    662 	switch (p->p_stat) {
    663 	case 0:
    664 	case SRUN:
    665 	case SZOMB:
    666 	default:
    667 		panic("setrunnable");
    668 	case SSTOP:
    669 		/*
    670 		 * If we're being traced (possibly because someone attached us
    671 		 * while we were stopped), check for a signal from the debugger.
    672 		 */
    673 		if ((p->p_flag & P_TRACED) != 0 && p->p_xstat != 0) {
    674 			sigaddset(&p->p_siglist, p->p_xstat);
    675 			p->p_sigcheck = 1;
    676 		}
    677 	case SSLEEP:
    678 		unsleep(p);		/* e.g. when sending signals */
    679 		break;
    680 
    681 	case SIDL:
    682 		break;
    683 	}
    684 	p->p_stat = SRUN;
    685 	if (p->p_flag & P_INMEM)
    686 		setrunqueue(p);
    687 	splx(s);
    688 	if (p->p_slptime > 1)
    689 		updatepri(p);
    690 	p->p_slptime = 0;
    691 	if ((p->p_flag & P_INMEM) == 0)
    692 		wakeup((caddr_t)&proc0);
    693 	else if (p->p_priority < curpriority)
    694 		need_resched();
    695 }
    696 
    697 /*
    698  * Compute the priority of a process when running in user mode.
    699  * Arrange to reschedule if the resulting priority is better
    700  * than that of the current process.
    701  */
    702 void
    703 resetpriority(p)
    704 	register struct proc *p;
    705 {
    706 	register unsigned int newpriority;
    707 
    708 	newpriority = PUSER + p->p_estcpu + NICE_WEIGHT * (p->p_nice - NZERO);
    709 	newpriority = min(newpriority, MAXPRI);
    710 	p->p_usrpri = newpriority;
    711 	if (newpriority < curpriority)
    712 		need_resched();
    713 }
    714 
    715 /*
    716  * We adjust the priority of the current process.  The priority of a process
    717  * gets worse as it accumulates CPU time.  The cpu usage estimator (p_estcpu)
    718  * is increased here.  The formula for computing priorities (in kern_synch.c)
    719  * will compute a different value each time p_estcpu increases. This can
    720  * cause a switch, but unless the priority crosses a PPQ boundary the actual
    721  * queue will not change.  The cpu usage estimator ramps up quite quickly
    722  * when the process is running (linearly), and decays away exponentially, at
    723  * a rate which is proportionally slower when the system is busy.  The basic
    724  * principal is that the system will 90% forget that the process used a lot
    725  * of CPU time in 5 * loadav seconds.  This causes the system to favor
    726  * processes which haven't run much recently, and to round-robin among other
    727  * processes.
    728  */
    729 
    730 void
    731 schedclock(p)
    732 	struct proc *p;
    733 {
    734 	p->p_estcpu = ESTCPULIM(p->p_estcpu + 1);
    735 	resetpriority(p);
    736 	if (p->p_priority >= PUSER)
    737 		p->p_priority = p->p_usrpri;
    738 }
    739