BSD 4_4 release
[unix-history] / usr / src / sys / kern / kern_synch.c
index 9d4f917..609100b 100644 (file)
-/*     kern_synch.c    4.17    82/07/22        */
-
-#include "../h/param.h"
-#include "../h/systm.h"
-#include "../h/dir.h"
-#include "../h/user.h"
-#include "../h/proc.h"
-#include "../h/file.h"
-#include "../h/inode.h"
-#include "../h/vm.h"
-#include "../h/pte.h"
-#include "../h/inline.h"
-#include "../h/mtpr.h"
-#include "../h/quota.h"
+/*-
+ * Copyright (c) 1982, 1986, 1990, 1991, 1993
+ *     The Regents of the University of California.  All rights reserved.
+ * All rights reserved.
+ *
+ * Redistribution and use in source and binary forms, with or without
+ * modification, are permitted provided that the following conditions
+ * are met:
+ * 1. Redistributions of source code must retain the above copyright
+ *    notice, this list of conditions and the following disclaimer.
+ * 2. Redistributions in binary form must reproduce the above copyright
+ *    notice, this list of conditions and the following disclaimer in the
+ *    documentation and/or other materials provided with the distribution.
+ * 3. All advertising materials mentioning features or use of this software
+ *    must display the following acknowledgement:
+ *     This product includes software developed by the University of
+ *     California, Berkeley and its contributors.
+ * 4. Neither the name of the University nor the names of its contributors
+ *    may be used to endorse or promote products derived from this software
+ *    without specific prior written permission.
+ *
+ * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
+ * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+ * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
+ * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+ * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
+ * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
+ * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
+ * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
+ * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
+ * SUCH DAMAGE.
+ *
+ *     @(#)kern_synch.c        8.1 (Berkeley) 6/10/93
+ */
+
+#include <sys/param.h>
+#include <sys/systm.h>
+#include <sys/proc.h>
+#include <sys/kernel.h>
+#include <sys/buf.h>
+#include <sys/signalvar.h>
+#include <sys/resourcevar.h>
+#include <sys/vmmeter.h>
+#ifdef KTRACE
+#include <sys/ktrace.h>
+#endif
+
+#include <machine/cpu.h>
+
+u_char curpri;                 /* usrpri of curproc */
+int    lbolt;                  /* once a second sleep address */
+
+/*
+ * Force switch among equal priority processes every 100ms.
+ */
+/* ARGSUSED */
+void
+roundrobin(arg)
+       void *arg;
+{
+
+       need_resched();
+       timeout(roundrobin, (void *)0, hz / 10);
+}
+
+/*
+ * constants for digital decay and forget
+ *     90% of (p_cpu) usage in 5*loadav time
+ *     95% of (p_pctcpu) usage in 60 seconds (load insensitive)
+ *          Note that, as ps(1) mentions, this can let percentages
+ *          total over 100% (I've seen 137.9% for 3 processes).
+ *
+ * Note that hardclock updates p_cpu and p_cpticks independently.
+ *
+ * We wish to decay away 90% of p_cpu in (5 * loadavg) seconds.
+ * That is, the system wants to compute a value of decay such
+ * that the following for loop:
+ *     for (i = 0; i < (5 * loadavg); i++)
+ *             p_cpu *= decay;
+ * will compute
+ *     p_cpu *= 0.1;
+ * for all values of loadavg:
+ *
+ * Mathematically this loop can be expressed by saying:
+ *     decay ** (5 * loadavg) ~= .1
+ *
+ * The system computes decay as:
+ *     decay = (2 * loadavg) / (2 * loadavg + 1)
+ *
+ * We wish to prove that the system's computation of decay
+ * will always fulfill the equation:
+ *     decay ** (5 * loadavg) ~= .1
+ *
+ * If we compute b as:
+ *     b = 2 * loadavg
+ * then
+ *     decay = b / (b + 1)
+ *
+ * We now need to prove two things:
+ *     1) Given factor ** (5 * loadavg) ~= .1, prove factor == b/(b+1)
+ *     2) Given b/(b+1) ** power ~= .1, prove power == (5 * loadavg)
+ *     
+ * Facts:
+ *         For x close to zero, exp(x) =~ 1 + x, since
+ *              exp(x) = 0! + x**1/1! + x**2/2! + ... .
+ *              therefore exp(-1/b) =~ 1 - (1/b) = (b-1)/b.
+ *         For x close to zero, ln(1+x) =~ x, since
+ *              ln(1+x) = x - x**2/2 + x**3/3 - ...     -1 < x < 1
+ *              therefore ln(b/(b+1)) = ln(1 - 1/(b+1)) =~ -1/(b+1).
+ *         ln(.1) =~ -2.30
+ *
+ * Proof of (1):
+ *    Solve (factor)**(power) =~ .1 given power (5*loadav):
+ *     solving for factor,
+ *      ln(factor) =~ (-2.30/5*loadav), or
+ *      factor =~ exp(-1/((5/2.30)*loadav)) =~ exp(-1/(2*loadav)) =
+ *          exp(-1/b) =~ (b-1)/b =~ b/(b+1).                    QED
+ *
+ * Proof of (2):
+ *    Solve (factor)**(power) =~ .1 given factor == (b/(b+1)):
+ *     solving for power,
+ *      power*ln(b/(b+1)) =~ -2.30, or
+ *      power =~ 2.3 * (b + 1) = 4.6*loadav + 2.3 =~ 5*loadav.  QED
+ *
+ * Actual power values for the implemented algorithm are as follows:
+ *      loadav: 1       2       3       4
+ *      power:  5.68    10.32   14.94   19.55
+ */
+
+/* calculations for digital decay to forget 90% of usage in 5*loadav sec */
+#define        loadfactor(loadav)      (2 * (loadav))
+#define        decay_cpu(loadfac, cpu) (((loadfac) * (cpu)) / ((loadfac) + FSCALE))
+
+/* decay 95% of `p_pctcpu' in 60 seconds; see CCPU_SHIFT before changing */
+fixpt_t        ccpu = 0.95122942450071400909 * FSCALE;         /* exp(-1/20) */
+
+/*
+ * If `ccpu' is not equal to `exp(-1/20)' and you still want to use the
+ * faster/more-accurate formula, you'll have to estimate CCPU_SHIFT below
+ * and possibly adjust FSHIFT in "param.h" so that (FSHIFT >= CCPU_SHIFT).
+ *
+ * To estimate CCPU_SHIFT for exp(-1/20), the following formula was used:
+ *     1 - exp(-1/20) ~= 0.0487 ~= 0.0488 == 1 (fixed pt, *11* bits).
+ *
+ * If you dont want to bother with the faster/more-accurate formula, you
+ * can set CCPU_SHIFT to (FSHIFT + 1) which will use a slower/less-accurate
+ * (more general) method of calculating the %age of CPU used by a process.
+ */
+#define        CCPU_SHIFT      11
+
+/*
+ * Recompute process priorities, once a second
+ */
+/* ARGSUSED */
+void
+schedcpu(arg)
+       void *arg;
+{
+       register fixpt_t loadfac = loadfactor(averunnable.ldavg[0]);
+       register struct proc *p;
+       register int s;
+       register unsigned int newcpu;
+
+       wakeup((caddr_t)&lbolt);
+       for (p = (struct proc *)allproc; p != NULL; p = p->p_nxt) {
+               /*
+                * Increment time in/out of memory and sleep time
+                * (if sleeping).  We ignore overflow; with 16-bit int's
+                * (remember them?) overflow takes 45 days.
+                */
+               p->p_time++;
+               if (p->p_stat == SSLEEP || p->p_stat == SSTOP)
+                       p->p_slptime++;
+               p->p_pctcpu = (p->p_pctcpu * ccpu) >> FSHIFT;
+               /*
+                * If the process has slept the entire second,
+                * stop recalculating its priority until it wakes up.
+                */
+               if (p->p_slptime > 1)
+                       continue;
+               s = splstatclock();     /* prevent state changes */
+               /*
+                * p_pctcpu is only for ps.
+                */
+#if    (FSHIFT >= CCPU_SHIFT)
+               p->p_pctcpu += (hz == 100)?
+                       ((fixpt_t) p->p_cpticks) << (FSHIFT - CCPU_SHIFT):
+                       100 * (((fixpt_t) p->p_cpticks)
+                               << (FSHIFT - CCPU_SHIFT)) / hz;
+#else
+               p->p_pctcpu += ((FSCALE - ccpu) *
+                       (p->p_cpticks * FSCALE / hz)) >> FSHIFT;
+#endif
+               p->p_cpticks = 0;
+               newcpu = (u_int) decay_cpu(loadfac, p->p_cpu) + p->p_nice;
+               p->p_cpu = min(newcpu, UCHAR_MAX);
+               setpri(p);
+               if (p->p_pri >= PUSER) {
+#define        PPQ     (128 / NQS)             /* priorities per queue */
+                       if ((p != curproc) &&
+                           p->p_stat == SRUN &&
+                           (p->p_flag & SLOAD) &&
+                           (p->p_pri / PPQ) != (p->p_usrpri / PPQ)) {
+                               remrq(p);
+                               p->p_pri = p->p_usrpri;
+                               setrq(p);
+                       } else
+                               p->p_pri = p->p_usrpri;
+               }
+               splx(s);
+       }
+       vmmeter();
+       if (bclnlist != NULL)
+               wakeup((caddr_t)pageproc);
+       timeout(schedcpu, (void *)0, hz);
+}
+
+/*
+ * Recalculate the priority of a process after it has slept for a while.
+ * For all load averages >= 1 and max p_cpu of 255, sleeping for at least
+ * six times the loadfactor will decay p_cpu to zero.
+ */
+void
+updatepri(p)
+       register struct proc *p;
+{
+       register unsigned int newcpu = p->p_cpu;
+       register fixpt_t loadfac = loadfactor(averunnable.ldavg[0]);
+
+       if (p->p_slptime > 5 * loadfac)
+               p->p_cpu = 0;
+       else {
+               p->p_slptime--; /* the first time was done in schedcpu */
+               while (newcpu && --p->p_slptime)
+                       newcpu = (int) decay_cpu(loadfac, newcpu);
+               p->p_cpu = min(newcpu, UCHAR_MAX);
+       }
+       setpri(p);
+}
 
 #define SQSIZE 0100    /* Must be power of 2 */
 #define HASH(x)        (( (int) x >> 5) & (SQSIZE-1))
 
 #define SQSIZE 0100    /* Must be power of 2 */
 #define HASH(x)        (( (int) x >> 5) & (SQSIZE-1))
-struct proc *slpque[SQSIZE];
+struct slpque {
+       struct proc *sq_head;
+       struct proc **sq_tailp;
+} slpque[SQSIZE];
 
 /*
 
 /*
- * Give up the processor till a wakeup occurs
- * on chan, at which time the process
- * enters the scheduling queue at priority pri.
- * The most important effect of pri is that when
- * pri<=PZERO a signal cannot disturb the sleep;
- * if pri>PZERO signals will be processed.
- * Callers of this routine must be prepared for
- * premature return, and check that the reason for
- * sleeping has gone away.
+ * During autoconfiguration or after a panic, a sleep will simply
+ * lower the priority briefly to allow interrupts, then return.
+ * The priority to be used (safepri) is machine-dependent, thus this
+ * value is initialized and maintained in the machine-dependent layers.
+ * This priority will typically be 0, or the lowest priority
+ * that is safe for use on the interrupt stack; it can be made
+ * higher to block network software interrupts after panics.
  */
  */
-sleep(chan, pri)
-caddr_t chan;
+int safepri;
+
+/*
+ * General sleep call.
+ * Suspends current process until a wakeup is made on chan.
+ * The process will then be made runnable with priority pri.
+ * Sleeps at most timo/hz seconds (0 means no timeout).
+ * If pri includes PCATCH flag, signals are checked
+ * before and after sleeping, else signals are not checked.
+ * Returns 0 if awakened, EWOULDBLOCK if the timeout expires.
+ * If PCATCH is set and a signal needs to be delivered,
+ * ERESTART is returned if the current system call should be restarted
+ * if possible, and EINTR is returned if the system call should
+ * be interrupted by the signal (return EINTR).
+ */
+int
+tsleep(chan, pri, wmesg, timo)
+       void *chan;
+       int pri;
+       char *wmesg;
+       int timo;
 {
 {
-       register struct proc *rp, **hp;
+       register struct proc *p = curproc;
+       register struct slpque *qp;
        register s;
        register s;
+       int sig, catch = pri & PCATCH;
+       extern int cold;
+       void endtsleep __P((void *));
 
 
-       rp = u.u_procp;
-       s = spl6();
-       if (chan==0 || rp->p_stat != SRUN || rp->p_rlink)
-               panic("sleep");
-       rp->p_wchan = chan;
-       rp->p_slptime = 0;
-       rp->p_pri = pri;
-       hp = &slpque[HASH(chan)];
-       rp->p_link = *hp;
-       *hp = rp;
-       if (pri > PZERO) {
-               if (ISSIG(rp)) {
-                       if (rp->p_wchan)
-                               unsleep(rp);
-                       rp->p_stat = SRUN;
-                       (void) spl0();
-                       goto psig;
-               }
-               if (rp->p_wchan == 0)
-                       goto out;
-               rp->p_stat = SSLEEP;
-               (void) spl0();
-               swtch();
-               if (ISSIG(rp))
-                       goto psig;
-       } else {
-               rp->p_stat = SSLEEP;
-               (void) spl0();
-               swtch();
+#ifdef KTRACE
+       if (KTRPOINT(p, KTR_CSW))
+               ktrcsw(p->p_tracep, 1, 0);
+#endif
+       s = splhigh();
+       if (cold || panicstr) {
+               /*
+                * After a panic, or during autoconfiguration,
+                * just give interrupts a chance, then just return;
+                * don't run any other procs or panic below,
+                * in case this is the idle process and already asleep.
+                */
+               splx(safepri);
+               splx(s);
+               return (0);
        }
        }
-out:
-       splx(s);
-       return;
-
+#ifdef DIAGNOSTIC
+       if (chan == NULL || p->p_stat != SRUN || p->p_rlink)
+               panic("tsleep");
+#endif
+       p->p_wchan = chan;
+       p->p_wmesg = wmesg;
+       p->p_slptime = 0;
+       p->p_pri = pri & PRIMASK;
+       qp = &slpque[HASH(chan)];
+       if (qp->sq_head == 0)
+               qp->sq_head = p;
+       else
+               *qp->sq_tailp = p;
+       *(qp->sq_tailp = &p->p_link) = 0;
+       if (timo)
+               timeout(endtsleep, (void *)p, timo);
        /*
        /*
-        * If priority was low (>PZERO) and
-        * there has been a signal, execute non-local goto through
-        * u.u_qsav, aborting the system call in progress (see trap.c)
-        * (or finishing a tsleep, see below)
+        * We put ourselves on the sleep queue and start our timeout
+        * before calling CURSIG, as we could stop there, and a wakeup
+        * or a SIGCONT (or both) could occur while we were stopped.
+        * A SIGCONT would cause us to be marked as SSLEEP
+        * without resuming us, thus we must be ready for sleep
+        * when CURSIG is called.  If the wakeup happens while we're
+        * stopped, p->p_wchan will be 0 upon return from CURSIG.
         */
         */
-psig:
-       longjmp(u.u_qsav);
-       /*NOTREACHED*/
+       if (catch) {
+               p->p_flag |= SSINTR;
+               if (sig = CURSIG(p)) {
+                       if (p->p_wchan)
+                               unsleep(p);
+                       p->p_stat = SRUN;
+                       goto resume;
+               }
+               if (p->p_wchan == 0) {
+                       catch = 0;
+                       goto resume;
+               }
+       } else
+               sig = 0;
+       p->p_stat = SSLEEP;
+       p->p_stats->p_ru.ru_nvcsw++;
+       swtch();
+resume:
+       curpri = p->p_usrpri;
+       splx(s);
+       p->p_flag &= ~SSINTR;
+       if (p->p_flag & STIMO) {
+               p->p_flag &= ~STIMO;
+               if (sig == 0) {
+#ifdef KTRACE
+                       if (KTRPOINT(p, KTR_CSW))
+                               ktrcsw(p->p_tracep, 0, 0);
+#endif
+                       return (EWOULDBLOCK);
+               }
+       } else if (timo)
+               untimeout(endtsleep, (void *)p);
+       if (catch && (sig != 0 || (sig = CURSIG(p)))) {
+#ifdef KTRACE
+               if (KTRPOINT(p, KTR_CSW))
+                       ktrcsw(p->p_tracep, 0, 0);
+#endif
+               if (p->p_sigacts->ps_sigintr & sigmask(sig))
+                       return (EINTR);
+               return (ERESTART);
+       }
+#ifdef KTRACE
+       if (KTRPOINT(p, KTR_CSW))
+               ktrcsw(p->p_tracep, 0, 0);
+#endif
+       return (0);
 }
 
 /*
 }
 
 /*
- * Sleep on chan at pri.
- * Return in no more than the indicated number of seconds.
- * (If seconds==0, no timeout implied)
- * Return      TS_OK if chan was awakened normally
- *             TS_TIME if timeout occurred
- *             TS_SIG if asynchronous signal occurred
- *
- * SHOULD HAVE OPTION TO SLEEP TO ABSOLUTE TIME OR AN
- * INCREMENT IN MILLISECONDS!
+ * Implement timeout for tsleep.
+ * If process hasn't been awakened (wchan non-zero),
+ * set timeout flag and undo the sleep.  If proc
+ * is stopped, just unsleep so it will remain stopped.
  */
  */
-tsleep(chan, pri, seconds)
-       caddr_t chan;
-       int pri, seconds;
+void
+endtsleep(arg)
+       void *arg;
 {
 {
-       label_t lqsav;
-       register struct proc *pp;
-       register sec, n, rval, sov;
-
-       pp = u.u_procp;
-       n = spl7();
-       sec = 0;
-       rval = 0;
-       sov = 0;
-       if (pp->p_clktim && pp->p_clktim<seconds)
-               if (pri > PZERO)
-                       seconds = 0;
+       register struct proc *p;
+       int s;
+
+       p = (struct proc *)arg;
+       s = splhigh();
+       if (p->p_wchan) {
+               if (p->p_stat == SSLEEP)
+                       setrun(p);
                else
                else
-                       sov++;
-       if (seconds) {
-               pp->p_flag |= STIMO;
-               sec = pp->p_clktim-seconds;
-               pp->p_clktim = seconds;
+                       unsleep(p);
+               p->p_flag |= STIMO;
        }
        }
-       bcopy((caddr_t)u.u_qsav, (caddr_t)lqsav, sizeof (label_t));
-       if (setjmp(u.u_qsav))
-               rval = TS_SIG;
-       else {
-               sleep(chan, pri);
-               if ((pp->p_flag&STIMO)==0 && seconds)
-                       rval = TS_TIME;
-               else
-                       rval = TS_OK;
+       splx(s);
+}
+
+/*
+ * Short-term, non-interruptable sleep.
+ */
+void
+sleep(chan, pri)
+       void *chan;
+       int pri;
+{
+       register struct proc *p = curproc;
+       register struct slpque *qp;
+       register s;
+       extern int cold;
+
+#ifdef DIAGNOSTIC
+       if (pri > PZERO) {
+               printf("sleep called with pri %d > PZERO, wchan: %x\n",
+                   pri, chan);
+               panic("old sleep");
        }
        }
-       pp->p_flag &= ~STIMO;
-       bcopy((caddr_t)lqsav, (caddr_t)u.u_qsav, sizeof (label_t));
-       if (sec > 0)
-               pp->p_clktim += sec;
-       else if (sov) {
-               if ((pp->p_clktim += sec) <= 0) {
-                       pp->p_clktim = 0;
-                       psignal(pp, SIGALRM);
-               }
-       } else
-               pp->p_clktim = 0;
-       splx(n);
-       return (rval);
+#endif
+       s = splhigh();
+       if (cold || panicstr) {
+               /*
+                * After a panic, or during autoconfiguration,
+                * just give interrupts a chance, then just return;
+                * don't run any other procs or panic below,
+                * in case this is the idle process and already asleep.
+                */
+               splx(safepri);
+               splx(s);
+               return;
+       }
+#ifdef DIAGNOSTIC
+       if (chan == NULL || p->p_stat != SRUN || p->p_rlink)
+               panic("sleep");
+#endif
+       p->p_wchan = chan;
+       p->p_wmesg = NULL;
+       p->p_slptime = 0;
+       p->p_pri = pri;
+       qp = &slpque[HASH(chan)];
+       if (qp->sq_head == 0)
+               qp->sq_head = p;
+       else
+               *qp->sq_tailp = p;
+       *(qp->sq_tailp = &p->p_link) = 0;
+       p->p_stat = SSLEEP;
+       p->p_stats->p_ru.ru_nvcsw++;
+#ifdef KTRACE
+       if (KTRPOINT(p, KTR_CSW))
+               ktrcsw(p->p_tracep, 1, 0);
+#endif
+       swtch();
+#ifdef KTRACE
+       if (KTRPOINT(p, KTR_CSW))
+               ktrcsw(p->p_tracep, 0, 0);
+#endif
+       curpri = p->p_usrpri;
+       splx(s);
 }
 
 /*
  * Remove a process from its wait queue
  */
 }
 
 /*
  * Remove a process from its wait queue
  */
+void
 unsleep(p)
        register struct proc *p;
 {
 unsleep(p)
        register struct proc *p;
 {
+       register struct slpque *qp;
        register struct proc **hp;
        register struct proc **hp;
-       register s;
+       int s;
 
 
-       s = spl6();
+       s = splhigh();
        if (p->p_wchan) {
        if (p->p_wchan) {
-               hp = &slpque[HASH(p->p_wchan)];
+               hp = &(qp = &slpque[HASH(p->p_wchan)])->sq_head;
                while (*hp != p)
                        hp = &(*hp)->p_link;
                *hp = p->p_link;
                while (*hp != p)
                        hp = &(*hp)->p_link;
                *hp = p->p_link;
+               if (qp->sq_tailp == &p->p_link)
+                       qp->sq_tailp = hp;
                p->p_wchan = 0;
        }
        splx(s);
 }
 
 /*
                p->p_wchan = 0;
        }
        splx(s);
 }
 
 /*
- * Wake up all processes sleeping on chan.
+ * Wakeup on "chan"; set all processes
+ * sleeping on chan to run state.
  */
  */
+void
 wakeup(chan)
 wakeup(chan)
-       register caddr_t chan;
+       register void *chan;
 {
 {
-       register struct proc *p, **q, **h;
+       register struct slpque *qp;
+       register struct proc *p, **q;
        int s;
 
        int s;
 
-       s = spl6();
-       h = &slpque[HASH(chan)];
+       s = splhigh();
+       qp = &slpque[HASH(chan)];
 restart:
 restart:
-       for (q = h; p = *q; ) {
+       for (q = &qp->sq_head; p = *q; ) {
+#ifdef DIAGNOSTIC
                if (p->p_rlink || p->p_stat != SSLEEP && p->p_stat != SSTOP)
                        panic("wakeup");
                if (p->p_rlink || p->p_stat != SSLEEP && p->p_stat != SSTOP)
                        panic("wakeup");
-               if (p->p_wchan==chan) {
+#endif
+               if (p->p_wchan == chan) {
                        p->p_wchan = 0;
                        *q = p->p_link;
                        p->p_wchan = 0;
                        *q = p->p_link;
-                       p->p_slptime = 0;
+                       if (qp->sq_tailp == &p->p_link)
+                               qp->sq_tailp = q;
                        if (p->p_stat == SSLEEP) {
                                /* OPTIMIZED INLINE EXPANSION OF setrun(p) */
                        if (p->p_stat == SSLEEP) {
                                /* OPTIMIZED INLINE EXPANSION OF setrun(p) */
+                               if (p->p_slptime > 1)
+                                       updatepri(p);
+                               p->p_slptime = 0;
                                p->p_stat = SRUN;
                                if (p->p_flag & SLOAD)
                                        setrq(p);
                                p->p_stat = SRUN;
                                if (p->p_flag & SLOAD)
                                        setrq(p);
-                               if (p->p_pri < curpri) {
-                                       runrun++;
-                                       aston();
-                               }
-                               if ((p->p_flag&SLOAD) == 0) {
-                                       if (runout != 0) {
-                                               runout = 0;
-                                               wakeup((caddr_t)&runout);
-                                       }
-                                       wantin++;
-                               }
+                               /*
+                                * Since curpri is a usrpri,
+                                * p->p_pri is always better than curpri.
+                                */
+                               if ((p->p_flag&SLOAD) == 0)
+                                       wakeup((caddr_t)&proc0);
+                               else
+                                       need_resched();
                                /* END INLINE EXPANSION */
                                goto restart;
                        }
                                /* END INLINE EXPANSION */
                                goto restart;
                        }
@@ -202,6 +533,63 @@ restart:
        splx(s);
 }
 
        splx(s);
 }
 
+/*
+ * The machine independent parts of swtch().
+ * Must be called at splstatclock() or higher.
+ */
+void
+swtch()
+{
+       register struct proc *p = curproc;      /* XXX */
+       register struct rlimit *rlim;
+       register long s, u;
+       struct timeval tv;
+
+       /*
+        * Compute the amount of time during which the current
+        * process was running, and add that to its total so far.
+        */
+       microtime(&tv);
+       u = p->p_rtime.tv_usec + (tv.tv_usec - runtime.tv_usec);
+       s = p->p_rtime.tv_sec + (tv.tv_sec - runtime.tv_sec);
+       if (u < 0) {
+               u += 1000000;
+               s--;
+       } else if (u >= 1000000) {
+               u -= 1000000;
+               s++;
+       }
+       p->p_rtime.tv_usec = u;
+       p->p_rtime.tv_sec = s;
+
+       /*
+        * Check if the process exceeds its cpu resource allocation.
+        * If over max, kill it.  In any case, if it has run for more
+        * than 10 minutes, reduce priority to give others a chance.
+        */
+       rlim = &p->p_rlimit[RLIMIT_CPU];
+       if (s >= rlim->rlim_cur) {
+               if (s >= rlim->rlim_max)
+                       psignal(p, SIGKILL);
+               else {
+                       psignal(p, SIGXCPU);
+                       if (rlim->rlim_cur < rlim->rlim_max)
+                               rlim->rlim_cur += 5;
+               }
+       }
+       if (s > 10 * 60 && p->p_ucred->cr_uid && p->p_nice == NZERO) {
+               p->p_nice = NZERO + 4;
+               setpri(p);
+       }
+
+       /*
+        * Pick a new current process and record its start time.
+        */
+       cnt.v_swtch++;
+       cpu_swtch(p);
+       microtime(&runtime);
+}
+
 /*
  * Initialize the (doubly-linked) run queues
  * to be empty.
 /*
  * Initialize the (doubly-linked) run queues
  * to be empty.
@@ -215,15 +603,17 @@ rqinit()
 }
 
 /*
 }
 
 /*
- * Set the process running;
- * arrange for it to be swapped in if necessary.
+ * Change process state to be runnable,
+ * placing it on the run queue if it is in memory,
+ * and awakening the swapper if it isn't in memory.
  */
  */
+void
 setrun(p)
        register struct proc *p;
 {
        register int s;
 
 setrun(p)
        register struct proc *p;
 {
        register int s;
 
-       s = spl6();
+       s = splhigh();
        switch (p->p_stat) {
 
        case 0:
        switch (p->p_stat) {
 
        case 0:
@@ -245,201 +635,29 @@ setrun(p)
        if (p->p_flag & SLOAD)
                setrq(p);
        splx(s);
        if (p->p_flag & SLOAD)
                setrq(p);
        splx(s);
-       if (p->p_pri < curpri) {
-               runrun++;
-               aston();
-       }
-       if ((p->p_flag&SLOAD) == 0) {
-               if (runout != 0) {
-                       runout = 0;
-                       wakeup((caddr_t)&runout);
-               }
-               wantin++;
-       }
+       if (p->p_slptime > 1)
+               updatepri(p);
+       p->p_slptime = 0;
+       if ((p->p_flag&SLOAD) == 0)
+               wakeup((caddr_t)&proc0);
+       else if (p->p_pri < curpri)
+               need_resched();
 }
 
 /*
 }
 
 /*
- * Set user priority.
- * The rescheduling flag (runrun)
- * is set if the priority is better
- * than the currently running process.
+ * Compute priority of process when running in user mode.
+ * Arrange to reschedule if the resulting priority
+ * is better than that of the current process.
  */
  */
-setpri(pp)
-       register struct proc *pp;
-{
-       register int p;
-
-       p = (pp->p_cpu & 0377)/4;
-       p += PUSER + 2*(pp->p_nice - NZERO);
-       if (pp->p_rssize > pp->p_maxrss && freemem < desfree)
-               p += 2*4;       /* effectively, nice(4) */
-       if (p > 127)
-               p = 127;
-       if (p < curpri) {
-               runrun++;
-               aston();
-       }
-       pp->p_usrpri = p;
-       return (p);
-}
-
-/*
- * Create a new process-- the internal version of
- * sys fork.
- * It returns 1 in the new process, 0 in the old.
- */
-newproc(isvfork)
-       int isvfork;
-{
+void
+setpri(p)
        register struct proc *p;
        register struct proc *p;
-       register struct proc *rpp, *rip;
-       register int n;
-
-       p = NULL;
-       /*
-        * First, just locate a slot for a process
-        * and copy the useful info from this process into it.
-        * The panic "cannot happen" because fork has already
-        * checked for the existence of a slot.
-        */
-retry:
-       mpid++;
-       if (mpid >= 30000) {
-               mpid = 0;
-               goto retry;
-       }
-       for (rpp = proc; rpp < procNPROC; rpp++) {
-               if (rpp->p_stat == NULL && p==NULL)
-                       p = rpp;
-               if (rpp->p_pid==mpid || rpp->p_pgrp==mpid)
-                       goto retry;
-       }
-       if ((rpp = p) == NULL)
-               panic("no procs");
-
-       /*
-        * Make a proc table entry for the new process.
-        */
-       rip = u.u_procp;
-#ifdef QUOTA
-       (rpp->p_quota = rip->p_quota)->q_cnt++;
-#endif
-       rpp->p_stat = SIDL;
-       rpp->p_clktim = 0;
-       rpp->p_flag = SLOAD | (rip->p_flag & (SPAGI|SNUSIG));
-       if (isvfork) {
-               rpp->p_flag |= SVFORK;
-               rpp->p_ndx = rip->p_ndx;
-       } else
-               rpp->p_ndx = rpp - proc;
-       rpp->p_uid = rip->p_uid;
-       rpp->p_pgrp = rip->p_pgrp;
-       rpp->p_nice = rip->p_nice;
-       rpp->p_textp = isvfork ? 0 : rip->p_textp;
-       rpp->p_pid = mpid;
-       rpp->p_ppid = rip->p_pid;
-       rpp->p_pptr = rip;
-       rpp->p_osptr = rip->p_cptr;
-       if (rip->p_cptr)
-               rip->p_cptr->p_ysptr = rpp;
-       rpp->p_ysptr = NULL;
-       rpp->p_cptr = NULL;
-       rip->p_cptr = rpp;
-       rpp->p_time = 0;
-       rpp->p_cpu = 0;
-       rpp->p_siga0 = rip->p_siga0;
-       rpp->p_siga1 = rip->p_siga1;
-       /* take along any pending signals, like stops? */
-       if (isvfork) {
-               rpp->p_tsize = rpp->p_dsize = rpp->p_ssize = 0;
-               rpp->p_szpt = clrnd(ctopt(UPAGES));
-               forkstat.cntvfork++;
-               forkstat.sizvfork += rip->p_dsize + rip->p_ssize;
-       } else {
-               rpp->p_tsize = rip->p_tsize;
-               rpp->p_dsize = rip->p_dsize;
-               rpp->p_ssize = rip->p_ssize;
-               rpp->p_szpt = rip->p_szpt;
-               forkstat.cntfork++;
-               forkstat.sizfork += rip->p_dsize + rip->p_ssize;
-       }
-       rpp->p_rssize = 0;
-       rpp->p_maxrss = rip->p_maxrss;
-       rpp->p_wchan = 0;
-       rpp->p_slptime = 0;
-       rpp->p_pctcpu = 0;
-       rpp->p_cpticks = 0;
-       n = PIDHASH(rpp->p_pid);
-       p->p_idhash = pidhash[n];
-       pidhash[n] = rpp - proc;
-       multprog++;
-
-       /*
-        * Increase reference counts on shared objects.
-        */
-       for(n=0; n<NOFILE; n++)
-               if (u.u_ofile[n] != NULL)
-                       u.u_ofile[n]->f_count++;
-       u.u_cdir->i_count++;
-       if (u.u_rdir)
-               u.u_rdir->i_count++;
-
-       /*
-        * Partially simulate the environment
-        * of the new process so that when it is actually
-        * created (by copying) it will look right.
-        * This begins the section where we must prevent the parent
-        * from being swapped.
-        */
-       rip->p_flag |= SKEEP;
-       if (procdup(rpp, isvfork))
-               return (1);
-
-       /*
-        * Make child runnable and add to run queue.
-        */
-       (void) spl6();
-       rpp->p_stat = SRUN;
-       setrq(rpp);
-       (void) spl0();
-
-       /*
-        * Cause child to take a non-local goto as soon as it runs.
-        * On older systems this was done with SSWAP bit in proc
-        * table; on VAX we use u.u_pcb.pcb_sswap so don't need
-        * to do rpp->p_flag |= SSWAP.  Actually do nothing here.
-        */
-       /* rpp->p_flag |= SSWAP; */
-
-       /*
-        * Now can be swapped.
-        */
-       rip->p_flag &= ~SKEEP;
-
-       /*
-        * If vfork make chain from parent process to child
-        * (where virtal memory is temporarily).  Wait for
-        * child to finish, steal virtual memory back,
-        * and wakeup child to let it die.
-        */
-       if (isvfork) {
-               u.u_procp->p_xlink = rpp;
-               u.u_procp->p_flag |= SNOVM;
-               while (rpp->p_flag & SVFORK)
-                       sleep((caddr_t)rpp, PZERO - 1);
-               if ((rpp->p_flag & SLOAD) == 0)
-                       panic("newproc vfork");
-               uaccess(rpp, Vfmap, &vfutl);
-               u.u_procp->p_xlink = 0;
-               vpassvm(rpp, u.u_procp, &vfutl, &u, Vfmap);
-               u.u_procp->p_flag &= ~SNOVM;
-               rpp->p_ndx = rpp - proc;
-               rpp->p_flag |= SVFDONE;
-               wakeup((caddr_t)rpp);
-       }
+{
+       register unsigned int newpri;
 
 
-       /*
-        * 0 return means parent.
-        */
-       return (0);
+       newpri = PUSER + p->p_cpu / 4 + 2 * p->p_nice;
+       newpri = min(newpri, MAXPRI);
+       p->p_usrpri = newpri;
+       if (newpri < curpri)
+               need_resched();
 }
 }