2 * Copyright (c) 1982, 1986, 1990, 1991, 1993
3 * The Regents of the University of California. All rights reserved.
4 * (c) UNIX System Laboratories, Inc.
5 * All or some portions of this file are derived from material licensed
6 * to the University of California by American Telephone and Telegraph
7 * Co. or Unix System Laboratories, Inc. and are reproduced herein with
8 * the permission of UNIX System Laboratories, Inc.
10 * Redistribution and use in source and binary forms, with or without
11 * modification, are permitted provided that the following conditions
13 * 1. Redistributions of source code must retain the above copyright
14 * notice, this list of conditions and the following disclaimer.
15 * 2. Redistributions in binary form must reproduce the above copyright
16 * notice, this list of conditions and the following disclaimer in the
17 * documentation and/or other materials provided with the distribution.
18 * 3. All advertising materials mentioning features or use of this software
19 * must display the following acknowledgement:
20 * This product includes software developed by the University of
21 * California, Berkeley and its contributors.
22 * 4. Neither the name of the University nor the names of its contributors
23 * may be used to endorse or promote products derived from this software
24 * without specific prior written permission.
26 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
27 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
28 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
29 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
30 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
31 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
32 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
33 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
34 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
35 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
38 * @(#)kern_synch.c 8.9 (Berkeley) 5/19/95
39 * $FreeBSD: src/sys/kern/kern_synch.c,v 1.87.2.6 2002/10/13 07:29:53 kbyanc Exp $
40 * $DragonFly: src/sys/kern/kern_synch.c,v 1.2 2003/06/17 04:28:41 dillon Exp $
43 #include "opt_ktrace.h"
45 #include <sys/param.h>
46 #include <sys/systm.h>
48 #include <sys/kernel.h>
49 #include <sys/signalvar.h>
50 #include <sys/resourcevar.h>
51 #include <sys/vmmeter.h>
52 #include <sys/sysctl.h>
55 #include <sys/ktrace.h>
58 #include <machine/cpu.h>
59 #include <machine/ipl.h>
60 #include <machine/smp.h>
62 static void sched_setup __P((void *dummy));
63 SYSINIT(sched_setup, SI_SUB_KICK_SCHEDULER, SI_ORDER_FIRST, sched_setup, NULL)
68 int sched_quantum; /* Roundrobin scheduling quantum in ticks. */
70 static struct callout loadav_callout;
72 struct loadavg averunnable =
73 { {0, 0, 0}, FSCALE }; /* load average, of runnable procs */
75 * Constants for averages over 1, 5, and 15 minutes
76 * when sampling at 5 second intervals.
78 static fixpt_t cexp[3] = {
79 0.9200444146293232 * FSCALE, /* exp(-1/12) */
80 0.9834714538216174 * FSCALE, /* exp(-1/60) */
81 0.9944598480048967 * FSCALE, /* exp(-1/180) */
84 static int curpriority_cmp __P((struct proc *p));
85 static void endtsleep __P((void *));
86 static void loadav __P((void *arg));
87 static void maybe_resched __P((struct proc *chk));
88 static void roundrobin __P((void *arg));
89 static void schedcpu __P((void *arg));
90 static void updatepri __P((struct proc *p));
93 sysctl_kern_quantum(SYSCTL_HANDLER_ARGS)
97 new_val = sched_quantum * tick;
98 error = sysctl_handle_int(oidp, &new_val, 0, req);
99 if (error != 0 || req->newptr == NULL)
103 sched_quantum = new_val / tick;
104 hogticks = 2 * sched_quantum;
108 SYSCTL_PROC(_kern, OID_AUTO, quantum, CTLTYPE_INT|CTLFLAG_RW,
109 0, sizeof sched_quantum, sysctl_kern_quantum, "I", "");
112 * Compare priorities. Return:
113 * <0: priority of p < current priority
114 * 0: priority of p == current priority
115 * >0: priority of p > current priority
116 * The priorities are the normal priorities or the normal realtime priorities
117 * if p is on the same scheduler as curproc. Otherwise the process on the
118 * more realtimeish scheduler has lowest priority. As usual, a higher
119 * priority really means a lower priority.
125 int c_class, p_class;
127 c_class = RTP_PRIO_BASE(curproc->p_rtprio.type);
128 p_class = RTP_PRIO_BASE(p->p_rtprio.type);
129 if (p_class != c_class)
130 return (p_class - c_class);
131 if (p_class == RTP_PRIO_NORMAL)
132 return (((int)p->p_priority - (int)curpriority) / PPQ);
133 return ((int)p->p_rtprio.prio - (int)curproc->p_rtprio.prio);
137 * Arrange to reschedule if necessary, taking the priorities and
138 * schedulers into account.
144 struct proc *p = curproc; /* XXX */
147 * XXX idle scheduler still broken because proccess stays on idle
148 * scheduler during waits (such as when getting FS locks). If a
149 * standard process becomes runaway cpu-bound, the system can lockup
150 * due to idle-scheduler processes in wakeup never getting any cpu.
156 } else if (chk == p) {
157 /* We may need to yield if our priority has been raised. */
158 if (curpriority_cmp(chk) > 0)
160 } else if (curpriority_cmp(chk) < 0)
165 roundrobin_interval(void)
167 return (sched_quantum);
171 * Force switch among equal priority processes every 100ms.
179 struct proc *p = curproc; /* XXX */
184 forward_roundrobin();
186 if (p == 0 || RTP_PRIO_NEED_RR(p->p_rtprio.type))
190 timeout(roundrobin, NULL, sched_quantum);
194 * Constants for digital decay and forget:
195 * 90% of (p_estcpu) usage in 5 * loadav time
196 * 95% of (p_pctcpu) usage in 60 seconds (load insensitive)
197 * Note that, as ps(1) mentions, this can let percentages
198 * total over 100% (I've seen 137.9% for 3 processes).
200 * Note that schedclock() updates p_estcpu and p_cpticks asynchronously.
202 * We wish to decay away 90% of p_estcpu in (5 * loadavg) seconds.
203 * That is, the system wants to compute a value of decay such
204 * that the following for loop:
205 * for (i = 0; i < (5 * loadavg); i++)
209 * for all values of loadavg:
211 * Mathematically this loop can be expressed by saying:
212 * decay ** (5 * loadavg) ~= .1
214 * The system computes decay as:
215 * decay = (2 * loadavg) / (2 * loadavg + 1)
217 * We wish to prove that the system's computation of decay
218 * will always fulfill the equation:
219 * decay ** (5 * loadavg) ~= .1
221 * If we compute b as:
224 * decay = b / (b + 1)
226 * We now need to prove two things:
227 * 1) Given factor ** (5 * loadavg) ~= .1, prove factor == b/(b+1)
228 * 2) Given b/(b+1) ** power ~= .1, prove power == (5 * loadavg)
231 * For x close to zero, exp(x) =~ 1 + x, since
232 * exp(x) = 0! + x**1/1! + x**2/2! + ... .
233 * therefore exp(-1/b) =~ 1 - (1/b) = (b-1)/b.
234 * For x close to zero, ln(1+x) =~ x, since
235 * ln(1+x) = x - x**2/2 + x**3/3 - ... -1 < x < 1
236 * therefore ln(b/(b+1)) = ln(1 - 1/(b+1)) =~ -1/(b+1).
240 * Solve (factor)**(power) =~ .1 given power (5*loadav):
241 * solving for factor,
242 * ln(factor) =~ (-2.30/5*loadav), or
243 * factor =~ exp(-1/((5/2.30)*loadav)) =~ exp(-1/(2*loadav)) =
244 * exp(-1/b) =~ (b-1)/b =~ b/(b+1). QED
247 * Solve (factor)**(power) =~ .1 given factor == (b/(b+1)):
249 * power*ln(b/(b+1)) =~ -2.30, or
250 * power =~ 2.3 * (b + 1) = 4.6*loadav + 2.3 =~ 5*loadav. QED
252 * Actual power values for the implemented algorithm are as follows:
254 * power: 5.68 10.32 14.94 19.55
257 /* calculations for digital decay to forget 90% of usage in 5*loadav sec */
258 #define loadfactor(loadav) (2 * (loadav))
259 #define decay_cpu(loadfac, cpu) (((loadfac) * (cpu)) / ((loadfac) + FSCALE))
261 /* decay 95% of `p_pctcpu' in 60 seconds; see CCPU_SHIFT before changing */
262 static fixpt_t ccpu = 0.95122942450071400909 * FSCALE; /* exp(-1/20) */
263 SYSCTL_INT(_kern, OID_AUTO, ccpu, CTLFLAG_RD, &ccpu, 0, "");
265 /* kernel uses `FSCALE', userland (SHOULD) use kern.fscale */
266 static int fscale __unused = FSCALE;
267 SYSCTL_INT(_kern, OID_AUTO, fscale, CTLFLAG_RD, 0, FSCALE, "");
270 * If `ccpu' is not equal to `exp(-1/20)' and you still want to use the
271 * faster/more-accurate formula, you'll have to estimate CCPU_SHIFT below
272 * and possibly adjust FSHIFT in "param.h" so that (FSHIFT >= CCPU_SHIFT).
274 * To estimate CCPU_SHIFT for exp(-1/20), the following formula was used:
275 * 1 - exp(-1/20) ~= 0.0487 ~= 0.0488 == 1 (fixed pt, *11* bits).
277 * If you don't want to bother with the faster/more-accurate formula, you
278 * can set CCPU_SHIFT to (FSHIFT + 1) which will use a slower/less-accurate
279 * (more general) method of calculating the %age of CPU used by a process.
281 #define CCPU_SHIFT 11
284 * Recompute process priorities, every hz ticks.
291 register fixpt_t loadfac = loadfactor(averunnable.ldavg[0]);
292 register struct proc *p;
293 register int realstathz, s;
295 realstathz = stathz ? stathz : hz;
296 LIST_FOREACH(p, &allproc, p_list) {
298 * Increment time in/out of memory and sleep time
299 * (if sleeping). We ignore overflow; with 16-bit int's
300 * (remember them?) overflow takes 45 days.
303 if (p->p_stat == SSLEEP || p->p_stat == SSTOP)
305 p->p_pctcpu = (p->p_pctcpu * ccpu) >> FSHIFT;
307 * If the process has slept the entire second,
308 * stop recalculating its priority until it wakes up.
310 if (p->p_slptime > 1)
312 s = splhigh(); /* prevent state changes and protect run queue */
314 * p_pctcpu is only for ps.
316 #if (FSHIFT >= CCPU_SHIFT)
317 p->p_pctcpu += (realstathz == 100)?
318 ((fixpt_t) p->p_cpticks) << (FSHIFT - CCPU_SHIFT):
319 100 * (((fixpt_t) p->p_cpticks)
320 << (FSHIFT - CCPU_SHIFT)) / realstathz;
322 p->p_pctcpu += ((FSCALE - ccpu) *
323 (p->p_cpticks * FSCALE / realstathz)) >> FSHIFT;
326 p->p_estcpu = decay_cpu(loadfac, p->p_estcpu);
328 if (p->p_priority >= PUSER) {
329 if ((p != curproc) &&
331 p->p_oncpu == 0xff && /* idle */
334 (p->p_flag & P_INMEM) &&
335 (p->p_priority / PPQ) != (p->p_usrpri / PPQ)) {
337 p->p_priority = p->p_usrpri;
340 p->p_priority = p->p_usrpri;
344 wakeup((caddr_t)&lbolt);
345 timeout(schedcpu, (void *)0, hz);
349 * Recalculate the priority of a process after it has slept for a while.
350 * For all load averages >= 1 and max p_estcpu of 255, sleeping for at
351 * least six times the loadfactor will decay p_estcpu to zero.
355 register struct proc *p;
357 register unsigned int newcpu = p->p_estcpu;
358 register fixpt_t loadfac = loadfactor(averunnable.ldavg[0]);
360 if (p->p_slptime > 5 * loadfac)
363 p->p_slptime--; /* the first time was done in schedcpu */
364 while (newcpu && --p->p_slptime)
365 newcpu = decay_cpu(loadfac, newcpu);
366 p->p_estcpu = newcpu;
372 * We're only looking at 7 bits of the address; everything is
373 * aligned to 4, lots of things are aligned to greater powers
374 * of 2. Shift right by 8, i.e. drop the bottom 256 worth.
376 #define TABLESIZE 128
377 static TAILQ_HEAD(slpquehead, proc) slpque[TABLESIZE];
378 #define LOOKUP(x) (((intptr_t)(x) >> 8) & (TABLESIZE - 1))
381 * During autoconfiguration or after a panic, a sleep will simply
382 * lower the priority briefly to allow interrupts, then return.
383 * The priority to be used (safepri) is machine-dependent, thus this
384 * value is initialized and maintained in the machine-dependent layers.
385 * This priority will typically be 0, or the lowest priority
386 * that is safe for use on the interrupt stack; it can be made
387 * higher to block network software interrupts after panics.
396 sched_quantum = hz/10;
397 hogticks = 2 * sched_quantum;
398 for (i = 0; i < TABLESIZE; i++)
399 TAILQ_INIT(&slpque[i]);
403 * General sleep call. Suspends the current process until a wakeup is
404 * performed on the specified identifier. The process will then be made
405 * runnable with the specified priority. Sleeps at most timo/hz seconds
406 * (0 means no timeout). If pri includes PCATCH flag, signals are checked
407 * before and after sleeping, else signals are not checked. Returns 0 if
408 * awakened, EWOULDBLOCK if the timeout expires. If PCATCH is set and a
409 * signal needs to be delivered, ERESTART is returned if the current system
410 * call should be restarted if possible, and EINTR is returned if the system
411 * call should be interrupted by the signal (return EINTR).
414 tsleep(ident, priority, wmesg, timo)
419 struct proc *p = curproc;
420 int s, sig, catch = priority & PCATCH;
421 struct callout_handle thandle;
424 if (p && KTRPOINT(p, KTR_CSW))
425 ktrcsw(p->p_tracep, 1, 0);
428 if (cold || panicstr) {
430 * After a panic, or during autoconfiguration,
431 * just give interrupts a chance, then just return;
432 * don't run any other procs or panic below,
433 * in case this is the idle process and already asleep.
439 KASSERT(p != NULL, ("tsleep1"));
440 KASSERT(ident != NULL && p->p_stat == SRUN, ("tsleep"));
442 * Process may be sitting on a slpque if asleep() was called, remove
443 * it before re-adding.
445 if (p->p_wchan != NULL)
451 p->p_priority = priority & PRIMASK;
452 TAILQ_INSERT_TAIL(&slpque[LOOKUP(ident)], p, p_procq);
454 thandle = timeout(endtsleep, (void *)p, timo);
456 * We put ourselves on the sleep queue and start our timeout
457 * before calling CURSIG, as we could stop there, and a wakeup
458 * or a SIGCONT (or both) could occur while we were stopped.
459 * A SIGCONT would cause us to be marked as SSLEEP
460 * without resuming us, thus we must be ready for sleep
461 * when CURSIG is called. If the wakeup happens while we're
462 * stopped, p->p_wchan will be 0 upon return from CURSIG.
465 p->p_flag |= P_SINTR;
466 if ((sig = CURSIG(p))) {
472 if (p->p_wchan == 0) {
479 p->p_stats->p_ru.ru_nvcsw++;
482 curpriority = p->p_usrpri;
484 p->p_flag &= ~P_SINTR;
485 if (p->p_flag & P_TIMEOUT) {
486 p->p_flag &= ~P_TIMEOUT;
489 if (KTRPOINT(p, KTR_CSW))
490 ktrcsw(p->p_tracep, 0, 0);
492 return (EWOULDBLOCK);
495 untimeout(endtsleep, (void *)p, thandle);
496 if (catch && (sig != 0 || (sig = CURSIG(p)))) {
498 if (KTRPOINT(p, KTR_CSW))
499 ktrcsw(p->p_tracep, 0, 0);
501 if (SIGISMEMBER(p->p_sigacts->ps_sigintr, sig))
506 if (KTRPOINT(p, KTR_CSW))
507 ktrcsw(p->p_tracep, 0, 0);
513 * asleep() - async sleep call. Place process on wait queue and return
514 * immediately without blocking. The process stays runnable until await()
515 * is called. If ident is NULL, remove process from wait queue if it is still
518 * Only the most recent sleep condition is effective when making successive
519 * calls to asleep() or when calling tsleep().
521 * The timeout, if any, is not initiated until await() is called. The sleep
522 * priority, signal, and timeout is specified in the asleep() call but may be
523 * overriden in the await() call.
525 * <<<<<<<< EXPERIMENTAL, UNTESTED >>>>>>>>>>
529 asleep(void *ident, int priority, const char *wmesg, int timo)
531 struct proc *p = curproc;
535 * splhigh() while manipulating sleep structures and slpque.
537 * Remove preexisting wait condition (if any) and place process
538 * on appropriate slpque, but do not put process to sleep.
543 if (p->p_wchan != NULL)
550 p->p_asleep.as_priority = priority;
551 p->p_asleep.as_timo = timo;
552 TAILQ_INSERT_TAIL(&slpque[LOOKUP(ident)], p, p_procq);
561 * await() - wait for async condition to occur. The process blocks until
562 * wakeup() is called on the most recent asleep() address. If wakeup is called
563 * priority to await(), await() winds up being a NOP.
565 * If await() is called more then once (without an intervening asleep() call),
566 * await() is still effectively a NOP but it calls mi_switch() to give other
567 * processes some cpu before returning. The process is left runnable.
569 * <<<<<<<< EXPERIMENTAL, UNTESTED >>>>>>>>>>
573 await(int priority, int timo)
575 struct proc *p = curproc;
580 if (p->p_wchan != NULL) {
581 struct callout_handle thandle;
586 * The call to await() can override defaults specified in
587 * the original asleep().
590 priority = p->p_asleep.as_priority;
592 timo = p->p_asleep.as_timo;
599 thandle = timeout(endtsleep, (void *)p, timo);
602 catch = priority & PCATCH;
605 p->p_flag |= P_SINTR;
606 if ((sig = CURSIG(p))) {
612 if (p->p_wchan == NULL) {
618 p->p_stats->p_ru.ru_nvcsw++;
621 curpriority = p->p_usrpri;
624 p->p_flag &= ~P_SINTR;
625 if (p->p_flag & P_TIMEOUT) {
626 p->p_flag &= ~P_TIMEOUT;
629 if (KTRPOINT(p, KTR_CSW))
630 ktrcsw(p->p_tracep, 0, 0);
632 return (EWOULDBLOCK);
635 untimeout(endtsleep, (void *)p, thandle);
636 if (catch && (sig != 0 || (sig = CURSIG(p)))) {
638 if (KTRPOINT(p, KTR_CSW))
639 ktrcsw(p->p_tracep, 0, 0);
641 if (SIGISMEMBER(p->p_sigacts->ps_sigintr, sig))
646 if (KTRPOINT(p, KTR_CSW))
647 ktrcsw(p->p_tracep, 0, 0);
651 * If as_priority is 0, await() has been called without an
652 * intervening asleep(). We are still effectively a NOP,
653 * but we call mi_switch() for safety.
656 if (p->p_asleep.as_priority == 0) {
657 p->p_stats->p_ru.ru_nvcsw++;
664 * clear p_asleep.as_priority as an indication that await() has been
665 * called. If await() is called again without an intervening asleep(),
666 * await() is still effectively a NOP but the above mi_switch() code
667 * is triggered as a safety.
669 p->p_asleep.as_priority = 0;
675 * Implement timeout for tsleep or asleep()/await()
677 * If process hasn't been awakened (wchan non-zero),
678 * set timeout flag and undo the sleep. If proc
679 * is stopped, just unsleep so it will remain stopped.
685 register struct proc *p;
688 p = (struct proc *)arg;
691 if (p->p_stat == SSLEEP)
695 p->p_flag |= P_TIMEOUT;
701 * Remove a process from its wait queue
705 register struct proc *p;
711 TAILQ_REMOVE(&slpque[LOOKUP(p->p_wchan)], p, p_procq);
718 * Make all processes sleeping on the specified identifier runnable.
722 register void *ident;
724 register struct slpquehead *qp;
725 register struct proc *p;
730 qp = &slpque[LOOKUP(ident)];
732 for (p = TAILQ_FIRST(qp); p != NULL; p = np) {
733 np = TAILQ_NEXT(p, p_procq);
734 if (p->p_wchan == ident) {
735 TAILQ_REMOVE(qp, p, p_procq);
737 if (p->p_stat == SSLEEP) {
738 /* OPTIMIZED EXPANSION OF setrunnable(p); */
739 if (p->p_slptime > 1)
743 if (p->p_flag & P_INMEM) {
747 p->p_flag |= P_SWAPINREQ;
748 wakeup((caddr_t)&proc0);
750 /* END INLINE EXPANSION */
759 * Make a process sleeping on the specified identifier runnable.
760 * May wake more than one process if a target process is currently
765 register void *ident;
767 register struct slpquehead *qp;
768 register struct proc *p;
773 qp = &slpque[LOOKUP(ident)];
776 for (p = TAILQ_FIRST(qp); p != NULL; p = np) {
777 np = TAILQ_NEXT(p, p_procq);
778 if (p->p_wchan == ident) {
779 TAILQ_REMOVE(qp, p, p_procq);
781 if (p->p_stat == SSLEEP) {
782 /* OPTIMIZED EXPANSION OF setrunnable(p); */
783 if (p->p_slptime > 1)
787 if (p->p_flag & P_INMEM) {
792 p->p_flag |= P_SWAPINREQ;
793 wakeup((caddr_t)&proc0);
795 /* END INLINE EXPANSION */
804 * The machine independent parts of mi_switch().
805 * Must be called at splstatclock() or higher.
810 struct timeval new_switchtime;
811 register struct proc *p = curproc; /* XXX */
812 register struct rlimit *rlim;
816 * XXX this spl is almost unnecessary. It is partly to allow for
817 * sloppy callers that don't do it (issignal() via CURSIG() is the
818 * main offender). It is partly to work around a bug in the i386
819 * cpu_switch() (the ipl is not preserved). We ran for years
820 * without it. I think there was only a interrupt latency problem.
821 * The main caller, tsleep(), does an splx() a couple of instructions
822 * after calling here. The buggy caller, issignal(), usually calls
823 * here at spl0() and sometimes returns at splhigh(). The process
824 * then runs for a little too long at splhigh(). The ipl gets fixed
825 * when the process returns to user mode (or earlier).
827 * It would probably be better to always call here at spl0(). Callers
828 * are prepared to give up control to another process, so they must
829 * be prepared to be interrupted. The clock stuff here may not
830 * actually need splstatclock().
834 #ifdef SIMPLELOCK_DEBUG
835 if (p->p_simple_locks)
836 printf("sleep: holding simple lock\n");
839 * Compute the amount of time during which the current
840 * process was running, and add that to its total so far.
842 microuptime(&new_switchtime);
843 if (timevalcmp(&new_switchtime, &switchtime, <)) {
844 printf("microuptime() went backwards (%ld.%06ld -> %ld.%06ld)\n",
845 switchtime.tv_sec, switchtime.tv_usec,
846 new_switchtime.tv_sec, new_switchtime.tv_usec);
847 new_switchtime = switchtime;
849 p->p_runtime += (new_switchtime.tv_usec - switchtime.tv_usec) +
850 (new_switchtime.tv_sec - switchtime.tv_sec) * (int64_t)1000000;
854 * Check if the process exceeds its cpu resource allocation.
855 * If over max, kill it.
857 if (p->p_stat != SZOMB && p->p_limit->p_cpulimit != RLIM_INFINITY &&
858 p->p_runtime > p->p_limit->p_cpulimit) {
859 rlim = &p->p_rlimit[RLIMIT_CPU];
860 if (p->p_runtime / (rlim_t)1000000 >= rlim->rlim_max) {
861 killproc(p, "exceeded maximum CPU limit");
864 if (rlim->rlim_cur < rlim->rlim_max) {
865 /* XXX: we should make a private copy */
872 * Pick a new current process and record its start time.
875 switchtime = new_switchtime;
877 if (switchtime.tv_sec == 0)
878 microuptime(&switchtime);
885 * Change process state to be runnable,
886 * placing it on the run queue if it is in memory,
887 * and awakening the swapper if it isn't in memory.
891 register struct proc *p;
901 panic("setrunnable");
904 unsleep(p); /* e.g. when sending signals */
911 if (p->p_flag & P_INMEM)
914 if (p->p_slptime > 1)
917 if ((p->p_flag & P_INMEM) == 0) {
918 p->p_flag |= P_SWAPINREQ;
919 wakeup((caddr_t)&proc0);
926 * Compute the priority of a process when running in user mode.
927 * Arrange to reschedule if the resulting priority is better
928 * than that of the current process.
932 register struct proc *p;
934 register unsigned int newpriority;
936 if (p->p_rtprio.type == RTP_PRIO_NORMAL) {
937 newpriority = PUSER + p->p_estcpu / INVERSE_ESTCPU_WEIGHT +
938 NICE_WEIGHT * p->p_nice;
939 newpriority = min(newpriority, MAXPRI);
940 p->p_usrpri = newpriority;
946 * Compute a tenex style load average of a quantity on
947 * 1, 5 and 15 minute intervals.
958 LIST_FOREACH(p, &allproc, p_list) {
965 for (i = 0; i < 3; i++)
966 avg->ldavg[i] = (cexp[i] * avg->ldavg[i] +
967 nrun * FSCALE * (FSCALE - cexp[i])) >> FSHIFT;
970 * Schedule the next update to occur after 5 seconds, but add a
971 * random variation to avoid synchronisation with processes that
972 * run at regular intervals.
974 callout_reset(&loadav_callout, hz * 4 + (int)(random() % (hz * 2 + 1)),
984 callout_init(&loadav_callout);
986 /* Kick off timeout driven events by calling first time. */
993 * We adjust the priority of the current process. The priority of
994 * a process gets worse as it accumulates CPU time. The cpu usage
995 * estimator (p_estcpu) is increased here. resetpriority() will
996 * compute a different priority each time p_estcpu increases by
997 * INVERSE_ESTCPU_WEIGHT
998 * (until MAXPRI is reached). The cpu usage estimator ramps up
999 * quite quickly when the process is running (linearly), and decays
1000 * away exponentially, at a rate which is proportionally slower when
1001 * the system is busy. The basic principle is that the system will
1002 * 90% forget that the process used a lot of CPU time in 5 * loadav
1003 * seconds. This causes the system to favor processes which haven't
1004 * run much recently, and to round-robin among other processes.
1012 p->p_estcpu = ESTCPULIM(p->p_estcpu + 1);
1013 if ((p->p_estcpu % INVERSE_ESTCPU_WEIGHT) == 0) {
1015 if (p->p_priority >= PUSER)
1016 p->p_priority = p->p_usrpri;