Home | History | Annotate | Line # | Download | only in kern
kern_synch.c revision 1.55
      1  1.55      ross /*	$NetBSD: kern_synch.c,v 1.55 1999/02/23 02:56:03 ross Exp $	*/
      2  1.26       cgd 
      3  1.26       cgd /*-
      4  1.26       cgd  * Copyright (c) 1982, 1986, 1990, 1991, 1993
      5  1.26       cgd  *	The Regents of the University of California.  All rights reserved.
      6  1.26       cgd  * (c) UNIX System Laboratories, Inc.
      7  1.26       cgd  * All or some portions of this file are derived from material licensed
      8  1.26       cgd  * to the University of California by American Telephone and Telegraph
      9  1.26       cgd  * Co. or Unix System Laboratories, Inc. and are reproduced herein with
     10  1.26       cgd  * the permission of UNIX System Laboratories, Inc.
     11  1.26       cgd  *
     12  1.26       cgd  * Redistribution and use in source and binary forms, with or without
     13  1.26       cgd  * modification, are permitted provided that the following conditions
     14  1.26       cgd  * are met:
     15  1.26       cgd  * 1. Redistributions of source code must retain the above copyright
     16  1.26       cgd  *    notice, this list of conditions and the following disclaimer.
     17  1.26       cgd  * 2. Redistributions in binary form must reproduce the above copyright
     18  1.26       cgd  *    notice, this list of conditions and the following disclaimer in the
     19  1.26       cgd  *    documentation and/or other materials provided with the distribution.
     20  1.26       cgd  * 3. All advertising materials mentioning features or use of this software
     21  1.26       cgd  *    must display the following acknowledgement:
     22  1.26       cgd  *	This product includes software developed by the University of
     23  1.26       cgd  *	California, Berkeley and its contributors.
     24  1.26       cgd  * 4. Neither the name of the University nor the names of its contributors
     25  1.26       cgd  *    may be used to endorse or promote products derived from this software
     26  1.26       cgd  *    without specific prior written permission.
     27  1.26       cgd  *
     28  1.26       cgd  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
     29  1.26       cgd  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
     30  1.26       cgd  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
     31  1.26       cgd  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
     32  1.26       cgd  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
     33  1.26       cgd  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
     34  1.26       cgd  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
     35  1.26       cgd  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
     36  1.26       cgd  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
     37  1.26       cgd  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
     38  1.26       cgd  * SUCH DAMAGE.
     39  1.26       cgd  *
     40  1.50      fvdl  *	@(#)kern_synch.c	8.9 (Berkeley) 5/19/95
     41  1.26       cgd  */
     42  1.48       mrg 
     43  1.52  jonathan #include "opt_ddb.h"
     44  1.51   thorpej #include "opt_ktrace.h"
     45  1.48       mrg #include "opt_uvm.h"
     46  1.26       cgd 
     47  1.26       cgd #include <sys/param.h>
     48  1.26       cgd #include <sys/systm.h>
     49  1.26       cgd #include <sys/proc.h>
     50  1.26       cgd #include <sys/kernel.h>
     51  1.26       cgd #include <sys/buf.h>
     52  1.26       cgd #include <sys/signalvar.h>
     53  1.26       cgd #include <sys/resourcevar.h>
     54  1.34  christos #include <vm/vm.h>
     55  1.55      ross #include <sys/sched.h>
     56  1.47       mrg 
     57  1.47       mrg #if defined(UVM)
     58  1.47       mrg #include <uvm/uvm_extern.h>
     59  1.47       mrg #endif
     60  1.47       mrg 
     61  1.26       cgd #ifdef KTRACE
     62  1.26       cgd #include <sys/ktrace.h>
     63  1.26       cgd #endif
     64  1.26       cgd 
     65  1.55      ross #define NICE_WEIGHT 2			/* priorities per nice level */
     66  1.55      ross #define	PPQ	(128 / NQS)		/* priorities per queue */
     67  1.55      ross 
     68  1.55      ross #define	ESTCPULIM(e) min((e), NICE_WEIGHT * PRIO_MAX - PPQ)
     69  1.55      ross 
     70  1.26       cgd #include <machine/cpu.h>
     71  1.34  christos 
     72  1.26       cgd u_char	curpriority;		/* usrpri of curproc */
     73  1.26       cgd int	lbolt;			/* once a second sleep address */
     74  1.26       cgd 
     75  1.34  christos void roundrobin __P((void *));
     76  1.34  christos void schedcpu __P((void *));
     77  1.34  christos void updatepri __P((struct proc *));
     78  1.34  christos void endtsleep __P((void *));
     79  1.34  christos 
     80  1.26       cgd /*
     81  1.26       cgd  * Force switch among equal priority processes every 100ms.
     82  1.26       cgd  */
     83  1.26       cgd /* ARGSUSED */
     84  1.26       cgd void
     85  1.26       cgd roundrobin(arg)
     86  1.26       cgd 	void *arg;
     87  1.26       cgd {
     88  1.26       cgd 
     89  1.26       cgd 	need_resched();
     90  1.26       cgd 	timeout(roundrobin, NULL, hz / 10);
     91  1.26       cgd }
     92  1.26       cgd 
     93  1.26       cgd /*
     94  1.26       cgd  * Constants for digital decay and forget:
     95  1.26       cgd  *	90% of (p_estcpu) usage in 5 * loadav time
     96  1.26       cgd  *	95% of (p_pctcpu) usage in 60 seconds (load insensitive)
     97  1.26       cgd  *          Note that, as ps(1) mentions, this can let percentages
     98  1.26       cgd  *          total over 100% (I've seen 137.9% for 3 processes).
     99  1.26       cgd  *
    100  1.26       cgd  * Note that hardclock updates p_estcpu and p_cpticks independently.
    101  1.26       cgd  *
    102  1.26       cgd  * We wish to decay away 90% of p_estcpu in (5 * loadavg) seconds.
    103  1.26       cgd  * That is, the system wants to compute a value of decay such
    104  1.26       cgd  * that the following for loop:
    105  1.26       cgd  * 	for (i = 0; i < (5 * loadavg); i++)
    106  1.26       cgd  * 		p_estcpu *= decay;
    107  1.26       cgd  * will compute
    108  1.26       cgd  * 	p_estcpu *= 0.1;
    109  1.26       cgd  * for all values of loadavg:
    110  1.26       cgd  *
    111  1.26       cgd  * Mathematically this loop can be expressed by saying:
    112  1.26       cgd  * 	decay ** (5 * loadavg) ~= .1
    113  1.26       cgd  *
    114  1.26       cgd  * The system computes decay as:
    115  1.26       cgd  * 	decay = (2 * loadavg) / (2 * loadavg + 1)
    116  1.26       cgd  *
    117  1.26       cgd  * We wish to prove that the system's computation of decay
    118  1.26       cgd  * will always fulfill the equation:
    119  1.26       cgd  * 	decay ** (5 * loadavg) ~= .1
    120  1.26       cgd  *
    121  1.26       cgd  * If we compute b as:
    122  1.26       cgd  * 	b = 2 * loadavg
    123  1.26       cgd  * then
    124  1.26       cgd  * 	decay = b / (b + 1)
    125  1.26       cgd  *
    126  1.26       cgd  * We now need to prove two things:
    127  1.26       cgd  *	1) Given factor ** (5 * loadavg) ~= .1, prove factor == b/(b+1)
    128  1.26       cgd  *	2) Given b/(b+1) ** power ~= .1, prove power == (5 * loadavg)
    129  1.26       cgd  *
    130  1.26       cgd  * Facts:
    131  1.26       cgd  *         For x close to zero, exp(x) =~ 1 + x, since
    132  1.26       cgd  *              exp(x) = 0! + x**1/1! + x**2/2! + ... .
    133  1.26       cgd  *              therefore exp(-1/b) =~ 1 - (1/b) = (b-1)/b.
    134  1.26       cgd  *         For x close to zero, ln(1+x) =~ x, since
    135  1.26       cgd  *              ln(1+x) = x - x**2/2 + x**3/3 - ...     -1 < x < 1
    136  1.26       cgd  *              therefore ln(b/(b+1)) = ln(1 - 1/(b+1)) =~ -1/(b+1).
    137  1.26       cgd  *         ln(.1) =~ -2.30
    138  1.26       cgd  *
    139  1.26       cgd  * Proof of (1):
    140  1.26       cgd  *    Solve (factor)**(power) =~ .1 given power (5*loadav):
    141  1.26       cgd  *	solving for factor,
    142  1.26       cgd  *      ln(factor) =~ (-2.30/5*loadav), or
    143  1.26       cgd  *      factor =~ exp(-1/((5/2.30)*loadav)) =~ exp(-1/(2*loadav)) =
    144  1.26       cgd  *          exp(-1/b) =~ (b-1)/b =~ b/(b+1).                    QED
    145  1.26       cgd  *
    146  1.26       cgd  * Proof of (2):
    147  1.26       cgd  *    Solve (factor)**(power) =~ .1 given factor == (b/(b+1)):
    148  1.26       cgd  *	solving for power,
    149  1.26       cgd  *      power*ln(b/(b+1)) =~ -2.30, or
    150  1.26       cgd  *      power =~ 2.3 * (b + 1) = 4.6*loadav + 2.3 =~ 5*loadav.  QED
    151  1.26       cgd  *
    152  1.26       cgd  * Actual power values for the implemented algorithm are as follows:
    153  1.26       cgd  *      loadav: 1       2       3       4
    154  1.26       cgd  *      power:  5.68    10.32   14.94   19.55
    155  1.26       cgd  */
    156  1.26       cgd 
    157  1.26       cgd /* calculations for digital decay to forget 90% of usage in 5*loadav sec */
    158  1.26       cgd #define	loadfactor(loadav)	(2 * (loadav))
    159  1.26       cgd #define	decay_cpu(loadfac, cpu)	(((loadfac) * (cpu)) / ((loadfac) + FSCALE))
    160  1.26       cgd 
    161  1.26       cgd /* decay 95% of `p_pctcpu' in 60 seconds; see CCPU_SHIFT before changing */
    162  1.26       cgd fixpt_t	ccpu = 0.95122942450071400909 * FSCALE;		/* exp(-1/20) */
    163  1.26       cgd 
    164  1.26       cgd /*
    165  1.26       cgd  * If `ccpu' is not equal to `exp(-1/20)' and you still want to use the
    166  1.26       cgd  * faster/more-accurate formula, you'll have to estimate CCPU_SHIFT below
    167  1.26       cgd  * and possibly adjust FSHIFT in "param.h" so that (FSHIFT >= CCPU_SHIFT).
    168  1.26       cgd  *
    169  1.26       cgd  * To estimate CCPU_SHIFT for exp(-1/20), the following formula was used:
    170  1.26       cgd  *	1 - exp(-1/20) ~= 0.0487 ~= 0.0488 == 1 (fixed pt, *11* bits).
    171  1.26       cgd  *
    172  1.26       cgd  * If you dont want to bother with the faster/more-accurate formula, you
    173  1.26       cgd  * can set CCPU_SHIFT to (FSHIFT + 1) which will use a slower/less-accurate
    174  1.26       cgd  * (more general) method of calculating the %age of CPU used by a process.
    175  1.26       cgd  */
    176  1.26       cgd #define	CCPU_SHIFT	11
    177  1.26       cgd 
    178  1.26       cgd /*
    179  1.26       cgd  * Recompute process priorities, every hz ticks.
    180  1.26       cgd  */
    181  1.26       cgd /* ARGSUSED */
    182  1.26       cgd void
    183  1.26       cgd schedcpu(arg)
    184  1.26       cgd 	void *arg;
    185  1.26       cgd {
    186  1.26       cgd 	register fixpt_t loadfac = loadfactor(averunnable.ldavg[0]);
    187  1.26       cgd 	register struct proc *p;
    188  1.26       cgd 	register int s;
    189  1.26       cgd 	register unsigned int newcpu;
    190  1.26       cgd 
    191  1.26       cgd 	wakeup((caddr_t)&lbolt);
    192  1.27   mycroft 	for (p = allproc.lh_first; p != 0; p = p->p_list.le_next) {
    193  1.26       cgd 		/*
    194  1.26       cgd 		 * Increment time in/out of memory and sleep time
    195  1.26       cgd 		 * (if sleeping).  We ignore overflow; with 16-bit int's
    196  1.26       cgd 		 * (remember them?) overflow takes 45 days.
    197  1.26       cgd 		 */
    198  1.26       cgd 		p->p_swtime++;
    199  1.26       cgd 		if (p->p_stat == SSLEEP || p->p_stat == SSTOP)
    200  1.26       cgd 			p->p_slptime++;
    201  1.26       cgd 		p->p_pctcpu = (p->p_pctcpu * ccpu) >> FSHIFT;
    202  1.26       cgd 		/*
    203  1.26       cgd 		 * If the process has slept the entire second,
    204  1.26       cgd 		 * stop recalculating its priority until it wakes up.
    205  1.26       cgd 		 */
    206  1.26       cgd 		if (p->p_slptime > 1)
    207  1.26       cgd 			continue;
    208  1.26       cgd 		s = splstatclock();	/* prevent state changes */
    209  1.26       cgd 		/*
    210  1.26       cgd 		 * p_pctcpu is only for ps.
    211  1.26       cgd 		 */
    212  1.55      ross 		KASSERT(profhz);
    213  1.26       cgd #if	(FSHIFT >= CCPU_SHIFT)
    214  1.55      ross 		p->p_pctcpu += (profhz == 100)?
    215  1.26       cgd 			((fixpt_t) p->p_cpticks) << (FSHIFT - CCPU_SHIFT):
    216  1.26       cgd                 	100 * (((fixpt_t) p->p_cpticks)
    217  1.55      ross 				<< (FSHIFT - CCPU_SHIFT)) / profhz;
    218  1.26       cgd #else
    219  1.26       cgd 		p->p_pctcpu += ((FSCALE - ccpu) *
    220  1.55      ross 			(p->p_cpticks * FSCALE / profhz)) >> FSHIFT;
    221  1.26       cgd #endif
    222  1.26       cgd 		p->p_cpticks = 0;
    223  1.55      ross 		newcpu = (u_int)decay_cpu(loadfac, p->p_estcpu);
    224  1.55      ross 		p->p_estcpu = newcpu;
    225  1.26       cgd 		resetpriority(p);
    226  1.26       cgd 		if (p->p_priority >= PUSER) {
    227  1.26       cgd 			if ((p != curproc) &&
    228  1.26       cgd 			    p->p_stat == SRUN &&
    229  1.26       cgd 			    (p->p_flag & P_INMEM) &&
    230  1.26       cgd 			    (p->p_priority / PPQ) != (p->p_usrpri / PPQ)) {
    231  1.43       cgd 				remrunqueue(p);
    232  1.26       cgd 				p->p_priority = p->p_usrpri;
    233  1.26       cgd 				setrunqueue(p);
    234  1.26       cgd 			} else
    235  1.26       cgd 				p->p_priority = p->p_usrpri;
    236  1.26       cgd 		}
    237  1.26       cgd 		splx(s);
    238  1.26       cgd 	}
    239  1.47       mrg #if defined(UVM)
    240  1.47       mrg 	uvm_meter();
    241  1.47       mrg #else
    242  1.26       cgd 	vmmeter();
    243  1.47       mrg #endif
    244  1.26       cgd 	timeout(schedcpu, (void *)0, hz);
    245  1.26       cgd }
    246  1.26       cgd 
    247  1.26       cgd /*
    248  1.26       cgd  * Recalculate the priority of a process after it has slept for a while.
    249  1.26       cgd  * For all load averages >= 1 and max p_estcpu of 255, sleeping for at
    250  1.26       cgd  * least six times the loadfactor will decay p_estcpu to zero.
    251  1.26       cgd  */
    252  1.26       cgd void
    253  1.26       cgd updatepri(p)
    254  1.26       cgd 	register struct proc *p;
    255  1.26       cgd {
    256  1.26       cgd 	register unsigned int newcpu = p->p_estcpu;
    257  1.26       cgd 	register fixpt_t loadfac = loadfactor(averunnable.ldavg[0]);
    258  1.26       cgd 
    259  1.26       cgd 	if (p->p_slptime > 5 * loadfac)
    260  1.26       cgd 		p->p_estcpu = 0;
    261  1.26       cgd 	else {
    262  1.26       cgd 		p->p_slptime--;	/* the first time was done in schedcpu */
    263  1.26       cgd 		while (newcpu && --p->p_slptime)
    264  1.26       cgd 			newcpu = (int) decay_cpu(loadfac, newcpu);
    265  1.55      ross 		p->p_estcpu = newcpu;
    266  1.26       cgd 	}
    267  1.26       cgd 	resetpriority(p);
    268  1.26       cgd }
    269  1.26       cgd 
    270  1.26       cgd /*
    271  1.26       cgd  * We're only looking at 7 bits of the address; everything is
    272  1.26       cgd  * aligned to 4, lots of things are aligned to greater powers
    273  1.26       cgd  * of 2.  Shift right by 8, i.e. drop the bottom 256 worth.
    274  1.26       cgd  */
    275  1.26       cgd #define TABLESIZE	128
    276  1.30       cgd #define LOOKUP(x)	(((long)(x) >> 8) & (TABLESIZE - 1))
    277  1.26       cgd struct slpque {
    278  1.26       cgd 	struct proc *sq_head;
    279  1.26       cgd 	struct proc **sq_tailp;
    280  1.26       cgd } slpque[TABLESIZE];
    281  1.26       cgd 
    282  1.26       cgd /*
    283  1.26       cgd  * During autoconfiguration or after a panic, a sleep will simply
    284  1.26       cgd  * lower the priority briefly to allow interrupts, then return.
    285  1.26       cgd  * The priority to be used (safepri) is machine-dependent, thus this
    286  1.26       cgd  * value is initialized and maintained in the machine-dependent layers.
    287  1.26       cgd  * This priority will typically be 0, or the lowest priority
    288  1.26       cgd  * that is safe for use on the interrupt stack; it can be made
    289  1.26       cgd  * higher to block network software interrupts after panics.
    290  1.26       cgd  */
    291  1.26       cgd int safepri;
    292  1.26       cgd 
    293  1.26       cgd /*
    294  1.26       cgd  * General sleep call.  Suspends the current process until a wakeup is
    295  1.26       cgd  * performed on the specified identifier.  The process will then be made
    296  1.26       cgd  * runnable with the specified priority.  Sleeps at most timo/hz seconds
    297  1.26       cgd  * (0 means no timeout).  If pri includes PCATCH flag, signals are checked
    298  1.26       cgd  * before and after sleeping, else signals are not checked.  Returns 0 if
    299  1.26       cgd  * awakened, EWOULDBLOCK if the timeout expires.  If PCATCH is set and a
    300  1.26       cgd  * signal needs to be delivered, ERESTART is returned if the current system
    301  1.26       cgd  * call should be restarted if possible, and EINTR is returned if the system
    302  1.26       cgd  * call should be interrupted by the signal (return EINTR).
    303  1.26       cgd  */
    304  1.26       cgd int
    305  1.26       cgd tsleep(ident, priority, wmesg, timo)
    306  1.26       cgd 	void *ident;
    307  1.26       cgd 	int priority, timo;
    308  1.45   mycroft 	const char *wmesg;
    309  1.26       cgd {
    310  1.26       cgd 	register struct proc *p = curproc;
    311  1.26       cgd 	register struct slpque *qp;
    312  1.49    kleink 	register int s;
    313  1.26       cgd 	int sig, catch = priority & PCATCH;
    314  1.26       cgd 	extern int cold;
    315  1.26       cgd 	void endtsleep __P((void *));
    316  1.26       cgd 
    317  1.26       cgd 	if (cold || panicstr) {
    318  1.26       cgd 		/*
    319  1.26       cgd 		 * After a panic, or during autoconfiguration,
    320  1.26       cgd 		 * just give interrupts a chance, then just return;
    321  1.26       cgd 		 * don't run any other procs or panic below,
    322  1.26       cgd 		 * in case this is the idle process and already asleep.
    323  1.26       cgd 		 */
    324  1.42       cgd 		s = splhigh();
    325  1.26       cgd 		splx(safepri);
    326  1.26       cgd 		splx(s);
    327  1.26       cgd 		return (0);
    328  1.26       cgd 	}
    329  1.42       cgd 
    330  1.42       cgd #ifdef KTRACE
    331  1.42       cgd 	if (KTRPOINT(p, KTR_CSW))
    332  1.42       cgd 		ktrcsw(p->p_tracep, 1, 0);
    333  1.42       cgd #endif
    334  1.42       cgd 	s = splhigh();
    335  1.42       cgd 
    336  1.26       cgd #ifdef DIAGNOSTIC
    337  1.26       cgd 	if (ident == NULL || p->p_stat != SRUN || p->p_back)
    338  1.26       cgd 		panic("tsleep");
    339  1.26       cgd #endif
    340  1.26       cgd 	p->p_wchan = ident;
    341  1.26       cgd 	p->p_wmesg = wmesg;
    342  1.26       cgd 	p->p_slptime = 0;
    343  1.26       cgd 	p->p_priority = priority & PRIMASK;
    344  1.26       cgd 	qp = &slpque[LOOKUP(ident)];
    345  1.26       cgd 	if (qp->sq_head == 0)
    346  1.26       cgd 		qp->sq_head = p;
    347  1.26       cgd 	else
    348  1.26       cgd 		*qp->sq_tailp = p;
    349  1.26       cgd 	*(qp->sq_tailp = &p->p_forw) = 0;
    350  1.26       cgd 	if (timo)
    351  1.26       cgd 		timeout(endtsleep, (void *)p, timo);
    352  1.26       cgd 	/*
    353  1.26       cgd 	 * We put ourselves on the sleep queue and start our timeout
    354  1.26       cgd 	 * before calling CURSIG, as we could stop there, and a wakeup
    355  1.26       cgd 	 * or a SIGCONT (or both) could occur while we were stopped.
    356  1.26       cgd 	 * A SIGCONT would cause us to be marked as SSLEEP
    357  1.26       cgd 	 * without resuming us, thus we must be ready for sleep
    358  1.26       cgd 	 * when CURSIG is called.  If the wakeup happens while we're
    359  1.26       cgd 	 * stopped, p->p_wchan will be 0 upon return from CURSIG.
    360  1.26       cgd 	 */
    361  1.26       cgd 	if (catch) {
    362  1.26       cgd 		p->p_flag |= P_SINTR;
    363  1.34  christos 		if ((sig = CURSIG(p)) != 0) {
    364  1.26       cgd 			if (p->p_wchan)
    365  1.26       cgd 				unsleep(p);
    366  1.26       cgd 			p->p_stat = SRUN;
    367  1.26       cgd 			goto resume;
    368  1.26       cgd 		}
    369  1.26       cgd 		if (p->p_wchan == 0) {
    370  1.26       cgd 			catch = 0;
    371  1.26       cgd 			goto resume;
    372  1.26       cgd 		}
    373  1.26       cgd 	} else
    374  1.26       cgd 		sig = 0;
    375  1.26       cgd 	p->p_stat = SSLEEP;
    376  1.26       cgd 	p->p_stats->p_ru.ru_nvcsw++;
    377  1.26       cgd 	mi_switch();
    378  1.26       cgd #ifdef	DDB
    379  1.26       cgd 	/* handy breakpoint location after process "wakes" */
    380  1.26       cgd 	asm(".globl bpendtsleep ; bpendtsleep:");
    381  1.26       cgd #endif
    382  1.26       cgd resume:
    383  1.26       cgd 	curpriority = p->p_usrpri;
    384  1.26       cgd 	splx(s);
    385  1.26       cgd 	p->p_flag &= ~P_SINTR;
    386  1.26       cgd 	if (p->p_flag & P_TIMEOUT) {
    387  1.26       cgd 		p->p_flag &= ~P_TIMEOUT;
    388  1.26       cgd 		if (sig == 0) {
    389  1.26       cgd #ifdef KTRACE
    390  1.26       cgd 			if (KTRPOINT(p, KTR_CSW))
    391  1.26       cgd 				ktrcsw(p->p_tracep, 0, 0);
    392  1.26       cgd #endif
    393  1.26       cgd 			return (EWOULDBLOCK);
    394  1.26       cgd 		}
    395  1.26       cgd 	} else if (timo)
    396  1.26       cgd 		untimeout(endtsleep, (void *)p);
    397  1.34  christos 	if (catch && (sig != 0 || (sig = CURSIG(p)) != 0)) {
    398  1.26       cgd #ifdef KTRACE
    399  1.26       cgd 		if (KTRPOINT(p, KTR_CSW))
    400  1.26       cgd 			ktrcsw(p->p_tracep, 0, 0);
    401  1.26       cgd #endif
    402  1.53   mycroft 		if ((p->p_sigacts->ps_sigact[sig].sa_flags & SA_RESTART) == 0)
    403  1.26       cgd 			return (EINTR);
    404  1.26       cgd 		return (ERESTART);
    405  1.26       cgd 	}
    406  1.26       cgd #ifdef KTRACE
    407  1.26       cgd 	if (KTRPOINT(p, KTR_CSW))
    408  1.26       cgd 		ktrcsw(p->p_tracep, 0, 0);
    409  1.26       cgd #endif
    410  1.26       cgd 	return (0);
    411  1.26       cgd }
    412  1.26       cgd 
    413  1.26       cgd /*
    414  1.26       cgd  * Implement timeout for tsleep.
    415  1.26       cgd  * If process hasn't been awakened (wchan non-zero),
    416  1.26       cgd  * set timeout flag and undo the sleep.  If proc
    417  1.26       cgd  * is stopped, just unsleep so it will remain stopped.
    418  1.26       cgd  */
    419  1.26       cgd void
    420  1.26       cgd endtsleep(arg)
    421  1.26       cgd 	void *arg;
    422  1.26       cgd {
    423  1.26       cgd 	register struct proc *p;
    424  1.26       cgd 	int s;
    425  1.26       cgd 
    426  1.26       cgd 	p = (struct proc *)arg;
    427  1.26       cgd 	s = splhigh();
    428  1.26       cgd 	if (p->p_wchan) {
    429  1.26       cgd 		if (p->p_stat == SSLEEP)
    430  1.26       cgd 			setrunnable(p);
    431  1.26       cgd 		else
    432  1.26       cgd 			unsleep(p);
    433  1.26       cgd 		p->p_flag |= P_TIMEOUT;
    434  1.26       cgd 	}
    435  1.26       cgd 	splx(s);
    436  1.26       cgd }
    437  1.26       cgd 
    438  1.26       cgd /*
    439  1.26       cgd  * Short-term, non-interruptable sleep.
    440  1.26       cgd  */
    441  1.26       cgd void
    442  1.26       cgd sleep(ident, priority)
    443  1.26       cgd 	void *ident;
    444  1.26       cgd 	int priority;
    445  1.26       cgd {
    446  1.26       cgd 	register struct proc *p = curproc;
    447  1.26       cgd 	register struct slpque *qp;
    448  1.49    kleink 	register int s;
    449  1.26       cgd 	extern int cold;
    450  1.26       cgd 
    451  1.26       cgd #ifdef DIAGNOSTIC
    452  1.26       cgd 	if (priority > PZERO) {
    453  1.41  christos 		printf("sleep called with priority %d > PZERO, wchan: %p\n",
    454  1.26       cgd 		    priority, ident);
    455  1.26       cgd 		panic("old sleep");
    456  1.26       cgd 	}
    457  1.26       cgd #endif
    458  1.26       cgd 	s = splhigh();
    459  1.26       cgd 	if (cold || panicstr) {
    460  1.26       cgd 		/*
    461  1.26       cgd 		 * After a panic, or during autoconfiguration,
    462  1.26       cgd 		 * just give interrupts a chance, then just return;
    463  1.26       cgd 		 * don't run any other procs or panic below,
    464  1.26       cgd 		 * in case this is the idle process and already asleep.
    465  1.26       cgd 		 */
    466  1.26       cgd 		splx(safepri);
    467  1.26       cgd 		splx(s);
    468  1.26       cgd 		return;
    469  1.26       cgd 	}
    470  1.26       cgd #ifdef DIAGNOSTIC
    471  1.26       cgd 	if (ident == NULL || p->p_stat != SRUN || p->p_back)
    472  1.26       cgd 		panic("sleep");
    473  1.26       cgd #endif
    474  1.26       cgd 	p->p_wchan = ident;
    475  1.26       cgd 	p->p_wmesg = NULL;
    476  1.26       cgd 	p->p_slptime = 0;
    477  1.26       cgd 	p->p_priority = priority;
    478  1.26       cgd 	qp = &slpque[LOOKUP(ident)];
    479  1.26       cgd 	if (qp->sq_head == 0)
    480  1.26       cgd 		qp->sq_head = p;
    481  1.26       cgd 	else
    482  1.26       cgd 		*qp->sq_tailp = p;
    483  1.26       cgd 	*(qp->sq_tailp = &p->p_forw) = 0;
    484  1.26       cgd 	p->p_stat = SSLEEP;
    485  1.26       cgd 	p->p_stats->p_ru.ru_nvcsw++;
    486  1.26       cgd #ifdef KTRACE
    487  1.26       cgd 	if (KTRPOINT(p, KTR_CSW))
    488  1.26       cgd 		ktrcsw(p->p_tracep, 1, 0);
    489  1.26       cgd #endif
    490  1.26       cgd 	mi_switch();
    491  1.26       cgd #ifdef	DDB
    492  1.26       cgd 	/* handy breakpoint location after process "wakes" */
    493  1.26       cgd 	asm(".globl bpendsleep ; bpendsleep:");
    494  1.26       cgd #endif
    495  1.26       cgd #ifdef KTRACE
    496  1.26       cgd 	if (KTRPOINT(p, KTR_CSW))
    497  1.26       cgd 		ktrcsw(p->p_tracep, 0, 0);
    498  1.26       cgd #endif
    499  1.26       cgd 	curpriority = p->p_usrpri;
    500  1.26       cgd 	splx(s);
    501  1.26       cgd }
    502  1.26       cgd 
    503  1.26       cgd /*
    504  1.26       cgd  * Remove a process from its wait queue
    505  1.26       cgd  */
    506  1.26       cgd void
    507  1.26       cgd unsleep(p)
    508  1.26       cgd 	register struct proc *p;
    509  1.26       cgd {
    510  1.26       cgd 	register struct slpque *qp;
    511  1.26       cgd 	register struct proc **hp;
    512  1.26       cgd 	int s;
    513  1.26       cgd 
    514  1.26       cgd 	s = splhigh();
    515  1.26       cgd 	if (p->p_wchan) {
    516  1.26       cgd 		hp = &(qp = &slpque[LOOKUP(p->p_wchan)])->sq_head;
    517  1.26       cgd 		while (*hp != p)
    518  1.26       cgd 			hp = &(*hp)->p_forw;
    519  1.26       cgd 		*hp = p->p_forw;
    520  1.26       cgd 		if (qp->sq_tailp == &p->p_forw)
    521  1.26       cgd 			qp->sq_tailp = hp;
    522  1.26       cgd 		p->p_wchan = 0;
    523  1.26       cgd 	}
    524  1.26       cgd 	splx(s);
    525  1.26       cgd }
    526  1.26       cgd 
    527  1.26       cgd /*
    528  1.26       cgd  * Make all processes sleeping on the specified identifier runnable.
    529  1.26       cgd  */
    530  1.26       cgd void
    531  1.26       cgd wakeup(ident)
    532  1.26       cgd 	register void *ident;
    533  1.26       cgd {
    534  1.26       cgd 	register struct slpque *qp;
    535  1.26       cgd 	register struct proc *p, **q;
    536  1.26       cgd 	int s;
    537  1.26       cgd 
    538  1.26       cgd 	s = splhigh();
    539  1.26       cgd 	qp = &slpque[LOOKUP(ident)];
    540  1.26       cgd restart:
    541  1.34  christos 	for (q = &qp->sq_head; (p = *q) != NULL; ) {
    542  1.26       cgd #ifdef DIAGNOSTIC
    543  1.34  christos 		if (p->p_back || (p->p_stat != SSLEEP && p->p_stat != SSTOP))
    544  1.26       cgd 			panic("wakeup");
    545  1.26       cgd #endif
    546  1.26       cgd 		if (p->p_wchan == ident) {
    547  1.26       cgd 			p->p_wchan = 0;
    548  1.26       cgd 			*q = p->p_forw;
    549  1.26       cgd 			if (qp->sq_tailp == &p->p_forw)
    550  1.26       cgd 				qp->sq_tailp = q;
    551  1.26       cgd 			if (p->p_stat == SSLEEP) {
    552  1.26       cgd 				/* OPTIMIZED EXPANSION OF setrunnable(p); */
    553  1.26       cgd 				if (p->p_slptime > 1)
    554  1.26       cgd 					updatepri(p);
    555  1.26       cgd 				p->p_slptime = 0;
    556  1.26       cgd 				p->p_stat = SRUN;
    557  1.26       cgd 				if (p->p_flag & P_INMEM)
    558  1.26       cgd 					setrunqueue(p);
    559  1.26       cgd 				/*
    560  1.26       cgd 				 * Since curpriority is a user priority,
    561  1.26       cgd 				 * p->p_priority is always better than
    562  1.26       cgd 				 * curpriority.
    563  1.26       cgd 				 */
    564  1.26       cgd 				if ((p->p_flag & P_INMEM) == 0)
    565  1.26       cgd 					wakeup((caddr_t)&proc0);
    566  1.26       cgd 				else
    567  1.26       cgd 					need_resched();
    568  1.26       cgd 				/* END INLINE EXPANSION */
    569  1.26       cgd 				goto restart;
    570  1.26       cgd 			}
    571  1.26       cgd 		} else
    572  1.26       cgd 			q = &p->p_forw;
    573  1.26       cgd 	}
    574  1.26       cgd 	splx(s);
    575  1.26       cgd }
    576  1.26       cgd 
    577  1.26       cgd /*
    578  1.26       cgd  * The machine independent parts of mi_switch().
    579  1.26       cgd  * Must be called at splstatclock() or higher.
    580  1.26       cgd  */
    581  1.26       cgd void
    582  1.26       cgd mi_switch()
    583  1.26       cgd {
    584  1.26       cgd 	register struct proc *p = curproc;	/* XXX */
    585  1.26       cgd 	register struct rlimit *rlim;
    586  1.26       cgd 	register long s, u;
    587  1.26       cgd 	struct timeval tv;
    588  1.26       cgd 
    589  1.50      fvdl #ifdef DEBUG
    590  1.54       chs 	if (p->p_simple_locks) {
    591  1.54       chs 		printf("p->p_simple_locks %d\n", p->p_simple_locks);
    592  1.54       chs #ifdef LOCKDEBUG
    593  1.54       chs 		simple_lock_dump();
    594  1.54       chs #endif
    595  1.50      fvdl 		panic("sleep: holding simple lock");
    596  1.54       chs 	}
    597  1.50      fvdl #endif
    598  1.26       cgd 	/*
    599  1.26       cgd 	 * Compute the amount of time during which the current
    600  1.26       cgd 	 * process was running, and add that to its total so far.
    601  1.26       cgd 	 */
    602  1.26       cgd 	microtime(&tv);
    603  1.26       cgd 	u = p->p_rtime.tv_usec + (tv.tv_usec - runtime.tv_usec);
    604  1.26       cgd 	s = p->p_rtime.tv_sec + (tv.tv_sec - runtime.tv_sec);
    605  1.26       cgd 	if (u < 0) {
    606  1.26       cgd 		u += 1000000;
    607  1.26       cgd 		s--;
    608  1.26       cgd 	} else if (u >= 1000000) {
    609  1.26       cgd 		u -= 1000000;
    610  1.26       cgd 		s++;
    611  1.26       cgd 	}
    612  1.26       cgd 	p->p_rtime.tv_usec = u;
    613  1.26       cgd 	p->p_rtime.tv_sec = s;
    614  1.26       cgd 
    615  1.26       cgd 	/*
    616  1.26       cgd 	 * Check if the process exceeds its cpu resource allocation.
    617  1.26       cgd 	 * If over max, kill it.  In any case, if it has run for more
    618  1.26       cgd 	 * than 10 minutes, reduce priority to give others a chance.
    619  1.26       cgd 	 */
    620  1.26       cgd 	rlim = &p->p_rlimit[RLIMIT_CPU];
    621  1.26       cgd 	if (s >= rlim->rlim_cur) {
    622  1.26       cgd 		if (s >= rlim->rlim_max)
    623  1.26       cgd 			psignal(p, SIGKILL);
    624  1.26       cgd 		else {
    625  1.26       cgd 			psignal(p, SIGXCPU);
    626  1.26       cgd 			if (rlim->rlim_cur < rlim->rlim_max)
    627  1.26       cgd 				rlim->rlim_cur += 5;
    628  1.26       cgd 		}
    629  1.26       cgd 	}
    630  1.38  explorer 	if (autonicetime && s > autonicetime && p->p_ucred->cr_uid && p->p_nice == NZERO) {
    631  1.39        ws 		p->p_nice = autoniceval + NZERO;
    632  1.26       cgd 		resetpriority(p);
    633  1.26       cgd 	}
    634  1.26       cgd 
    635  1.26       cgd 	/*
    636  1.26       cgd 	 * Pick a new current process and record its start time.
    637  1.26       cgd 	 */
    638  1.47       mrg #if defined(UVM)
    639  1.47       mrg 	uvmexp.swtch++;
    640  1.47       mrg #else
    641  1.26       cgd 	cnt.v_swtch++;
    642  1.47       mrg #endif
    643  1.26       cgd 	cpu_switch(p);
    644  1.26       cgd 	microtime(&runtime);
    645  1.26       cgd }
    646  1.26       cgd 
    647  1.26       cgd /*
    648  1.26       cgd  * Initialize the (doubly-linked) run queues
    649  1.26       cgd  * to be empty.
    650  1.26       cgd  */
    651  1.26       cgd void
    652  1.26       cgd rqinit()
    653  1.26       cgd {
    654  1.26       cgd 	register int i;
    655  1.26       cgd 
    656  1.26       cgd 	for (i = 0; i < NQS; i++)
    657  1.26       cgd 		qs[i].ph_link = qs[i].ph_rlink = (struct proc *)&qs[i];
    658  1.26       cgd }
    659  1.26       cgd 
    660  1.26       cgd /*
    661  1.26       cgd  * Change process state to be runnable,
    662  1.26       cgd  * placing it on the run queue if it is in memory,
    663  1.26       cgd  * and awakening the swapper if it isn't in memory.
    664  1.26       cgd  */
    665  1.26       cgd void
    666  1.26       cgd setrunnable(p)
    667  1.26       cgd 	register struct proc *p;
    668  1.26       cgd {
    669  1.26       cgd 	register int s;
    670  1.26       cgd 
    671  1.26       cgd 	s = splhigh();
    672  1.26       cgd 	switch (p->p_stat) {
    673  1.26       cgd 	case 0:
    674  1.26       cgd 	case SRUN:
    675  1.26       cgd 	case SZOMB:
    676  1.26       cgd 	default:
    677  1.26       cgd 		panic("setrunnable");
    678  1.26       cgd 	case SSTOP:
    679  1.33   mycroft 		/*
    680  1.33   mycroft 		 * If we're being traced (possibly because someone attached us
    681  1.33   mycroft 		 * while we were stopped), check for a signal from the debugger.
    682  1.33   mycroft 		 */
    683  1.53   mycroft 		if ((p->p_flag & P_TRACED) != 0 && p->p_xstat != 0) {
    684  1.53   mycroft 			sigaddset(&p->p_siglist, p->p_xstat);
    685  1.53   mycroft 			p->p_sigcheck = 1;
    686  1.53   mycroft 		}
    687  1.26       cgd 	case SSLEEP:
    688  1.26       cgd 		unsleep(p);		/* e.g. when sending signals */
    689  1.26       cgd 		break;
    690  1.26       cgd 
    691  1.26       cgd 	case SIDL:
    692  1.26       cgd 		break;
    693  1.26       cgd 	}
    694  1.26       cgd 	p->p_stat = SRUN;
    695  1.26       cgd 	if (p->p_flag & P_INMEM)
    696  1.26       cgd 		setrunqueue(p);
    697  1.26       cgd 	splx(s);
    698  1.26       cgd 	if (p->p_slptime > 1)
    699  1.26       cgd 		updatepri(p);
    700  1.26       cgd 	p->p_slptime = 0;
    701  1.26       cgd 	if ((p->p_flag & P_INMEM) == 0)
    702  1.26       cgd 		wakeup((caddr_t)&proc0);
    703  1.26       cgd 	else if (p->p_priority < curpriority)
    704  1.26       cgd 		need_resched();
    705  1.26       cgd }
    706  1.26       cgd 
    707  1.26       cgd /*
    708  1.26       cgd  * Compute the priority of a process when running in user mode.
    709  1.26       cgd  * Arrange to reschedule if the resulting priority is better
    710  1.26       cgd  * than that of the current process.
    711  1.26       cgd  */
    712  1.26       cgd void
    713  1.26       cgd resetpriority(p)
    714  1.26       cgd 	register struct proc *p;
    715  1.26       cgd {
    716  1.26       cgd 	register unsigned int newpriority;
    717  1.26       cgd 
    718  1.55      ross 	newpriority = PUSER + p->p_estcpu + NICE_WEIGHT * (p->p_nice - NZERO);
    719  1.26       cgd 	newpriority = min(newpriority, MAXPRI);
    720  1.26       cgd 	p->p_usrpri = newpriority;
    721  1.26       cgd 	if (newpriority < curpriority)
    722  1.26       cgd 		need_resched();
    723  1.55      ross }
    724  1.55      ross 
    725  1.55      ross /*
    726  1.55      ross  * We adjust the priority of the current process.  The priority of
    727  1.55      ross  * a process gets worse as it accumulates CPU time.  The cpu usage
    728  1.55      ross  * estimator (p_estcpu) is increased here.  The formula for computing
    729  1.55      ross  * priorities (in kern_synch.c) will compute a different value each
    730  1.55      ross  * time p_estcpu increases by 4.  The cpu usage estimator ramps up
    731  1.55      ross  * quite quickly when the process is running (linearly), and decays
    732  1.55      ross  * away exponentially, at a rate which is proportionally slower when
    733  1.55      ross  * the system is busy.  The basic principal is that the system will
    734  1.55      ross  * 90% forget that the process used a lot of CPU time in 5 * loadav
    735  1.55      ross  * seconds.  This causes the system to favor processes which haven't
    736  1.55      ross  * run much recently, and to round-robin among other processes.
    737  1.55      ross  */
    738  1.55      ross 
    739  1.55      ross void
    740  1.55      ross schedclk(p)
    741  1.55      ross 	struct proc *p;
    742  1.55      ross {
    743  1.55      ross 	p->p_estcpu = ESTCPULIM(p->p_estcpu + 1);
    744  1.55      ross 	resetpriority(p);
    745  1.55      ross 	if (p->p_priority >= PUSER)
    746  1.55      ross 		p->p_priority = p->p_usrpri;
    747  1.26       cgd }
    748