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 1.14 2003/07/03 17:24:02 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>
53 #include <sys/thread2.h>
56 #include <sys/ktrace.h>
58 #include <sys/xwait.h>
60 #include <machine/cpu.h>
61 #include <machine/ipl.h>
62 #include <machine/smp.h>
64 static void sched_setup __P((void *dummy));
65 SYSINIT(sched_setup, SI_SUB_KICK_SCHEDULER, SI_ORDER_FIRST, sched_setup, NULL)
69 int sched_quantum; /* Roundrobin scheduling quantum in ticks. */
72 static struct callout loadav_callout;
74 struct loadavg averunnable =
75 { {0, 0, 0}, FSCALE }; /* load average, of runnable procs */
77 * Constants for averages over 1, 5, and 15 minutes
78 * when sampling at 5 second intervals.
80 static fixpt_t cexp[3] = {
81 0.9200444146293232 * FSCALE, /* exp(-1/12) */
82 0.9834714538216174 * FSCALE, /* exp(-1/60) */
83 0.9944598480048967 * FSCALE, /* exp(-1/180) */
86 static void endtsleep __P((void *));
87 static void loadav __P((void *arg));
88 static void maybe_resched __P((struct proc *chk));
89 static void roundrobin __P((void *arg));
90 static void schedcpu __P((void *arg));
91 static void updatepri __P((struct proc *p));
94 sysctl_kern_quantum(SYSCTL_HANDLER_ARGS)
98 new_val = sched_quantum * tick;
99 error = sysctl_handle_int(oidp, &new_val, 0, req);
100 if (error != 0 || req->newptr == NULL)
104 sched_quantum = new_val / tick;
105 hogticks = 2 * sched_quantum;
109 SYSCTL_PROC(_kern, OID_AUTO, quantum, CTLTYPE_INT|CTLFLAG_RW,
110 0, sizeof sched_quantum, sysctl_kern_quantum, "I", "");
113 * Arrange to reschedule if necessary by checking to see if the current
114 * process is on the highest priority user scheduling queue. This may
115 * be run from an interrupt so we have to follow any preemption chains
116 * back to the original process.
119 maybe_resched(struct proc *chk)
121 struct proc *cur = lwkt_preempted_proc();
127 * Check the user queue (realtime, normal, idle). Lower numbers
128 * indicate higher priority queues. Lower numbers are also better
131 if (chk->p_rtprio.type < cur->p_rtprio.type) {
133 } else if (chk->p_rtprio.type == cur->p_rtprio.type) {
134 if (chk->p_rtprio.type == RTP_PRIO_NORMAL) {
135 if (chk->p_priority / PPQ < cur->p_priority / PPQ)
138 if (chk->p_rtprio.prio < cur->p_rtprio.prio)
145 roundrobin_interval(void)
147 return (sched_quantum);
151 * Force switch among equal priority processes every 100ms.
159 struct proc *p = curproc; /* XXX */
164 forward_roundrobin();
166 if (p == 0 || RTP_PRIO_NEED_RR(p->p_rtprio.type))
170 timeout(roundrobin, NULL, sched_quantum);
174 * Constants for digital decay and forget:
175 * 90% of (p_estcpu) usage in 5 * loadav time
176 * 95% of (p_pctcpu) usage in 60 seconds (load insensitive)
177 * Note that, as ps(1) mentions, this can let percentages
178 * total over 100% (I've seen 137.9% for 3 processes).
180 * Note that schedclock() updates p_estcpu and p_cpticks asynchronously.
182 * We wish to decay away 90% of p_estcpu in (5 * loadavg) seconds.
183 * That is, the system wants to compute a value of decay such
184 * that the following for loop:
185 * for (i = 0; i < (5 * loadavg); i++)
189 * for all values of loadavg:
191 * Mathematically this loop can be expressed by saying:
192 * decay ** (5 * loadavg) ~= .1
194 * The system computes decay as:
195 * decay = (2 * loadavg) / (2 * loadavg + 1)
197 * We wish to prove that the system's computation of decay
198 * will always fulfill the equation:
199 * decay ** (5 * loadavg) ~= .1
201 * If we compute b as:
204 * decay = b / (b + 1)
206 * We now need to prove two things:
207 * 1) Given factor ** (5 * loadavg) ~= .1, prove factor == b/(b+1)
208 * 2) Given b/(b+1) ** power ~= .1, prove power == (5 * loadavg)
211 * For x close to zero, exp(x) =~ 1 + x, since
212 * exp(x) = 0! + x**1/1! + x**2/2! + ... .
213 * therefore exp(-1/b) =~ 1 - (1/b) = (b-1)/b.
214 * For x close to zero, ln(1+x) =~ x, since
215 * ln(1+x) = x - x**2/2 + x**3/3 - ... -1 < x < 1
216 * therefore ln(b/(b+1)) = ln(1 - 1/(b+1)) =~ -1/(b+1).
220 * Solve (factor)**(power) =~ .1 given power (5*loadav):
221 * solving for factor,
222 * ln(factor) =~ (-2.30/5*loadav), or
223 * factor =~ exp(-1/((5/2.30)*loadav)) =~ exp(-1/(2*loadav)) =
224 * exp(-1/b) =~ (b-1)/b =~ b/(b+1). QED
227 * Solve (factor)**(power) =~ .1 given factor == (b/(b+1)):
229 * power*ln(b/(b+1)) =~ -2.30, or
230 * power =~ 2.3 * (b + 1) = 4.6*loadav + 2.3 =~ 5*loadav. QED
232 * Actual power values for the implemented algorithm are as follows:
234 * power: 5.68 10.32 14.94 19.55
237 /* calculations for digital decay to forget 90% of usage in 5*loadav sec */
238 #define loadfactor(loadav) (2 * (loadav))
239 #define decay_cpu(loadfac, cpu) (((loadfac) * (cpu)) / ((loadfac) + FSCALE))
241 /* decay 95% of `p_pctcpu' in 60 seconds; see CCPU_SHIFT before changing */
242 static fixpt_t ccpu = 0.95122942450071400909 * FSCALE; /* exp(-1/20) */
243 SYSCTL_INT(_kern, OID_AUTO, ccpu, CTLFLAG_RD, &ccpu, 0, "");
245 /* kernel uses `FSCALE', userland (SHOULD) use kern.fscale */
246 static int fscale __unused = FSCALE;
247 SYSCTL_INT(_kern, OID_AUTO, fscale, CTLFLAG_RD, 0, FSCALE, "");
250 * If `ccpu' is not equal to `exp(-1/20)' and you still want to use the
251 * faster/more-accurate formula, you'll have to estimate CCPU_SHIFT below
252 * and possibly adjust FSHIFT in "param.h" so that (FSHIFT >= CCPU_SHIFT).
254 * To estimate CCPU_SHIFT for exp(-1/20), the following formula was used:
255 * 1 - exp(-1/20) ~= 0.0487 ~= 0.0488 == 1 (fixed pt, *11* bits).
257 * If you don't want to bother with the faster/more-accurate formula, you
258 * can set CCPU_SHIFT to (FSHIFT + 1) which will use a slower/less-accurate
259 * (more general) method of calculating the %age of CPU used by a process.
261 #define CCPU_SHIFT 11
264 * Recompute process priorities, every hz ticks.
270 fixpt_t loadfac = loadfactor(averunnable.ldavg[0]);
275 curp = lwkt_preempted_proc(); /* YYY temporary hack */
277 realstathz = stathz ? stathz : hz;
278 LIST_FOREACH(p, &allproc, p_list) {
280 * Increment time in/out of memory and sleep time
281 * (if sleeping). We ignore overflow; with 16-bit int's
282 * (remember them?) overflow takes 45 days.
285 if (p->p_stat == SSLEEP || p->p_stat == SSTOP)
287 p->p_pctcpu = (p->p_pctcpu * ccpu) >> FSHIFT;
289 * If the process has slept the entire second,
290 * stop recalculating its priority until it wakes up.
292 if (p->p_slptime > 1)
294 s = splhigh(); /* prevent state changes and protect run queue */
296 * p_pctcpu is only for ps.
298 #if (FSHIFT >= CCPU_SHIFT)
299 p->p_pctcpu += (realstathz == 100)?
300 ((fixpt_t) p->p_cpticks) << (FSHIFT - CCPU_SHIFT):
301 100 * (((fixpt_t) p->p_cpticks)
302 << (FSHIFT - CCPU_SHIFT)) / realstathz;
304 p->p_pctcpu += ((FSCALE - ccpu) *
305 (p->p_cpticks * FSCALE / realstathz)) >> FSHIFT;
308 p->p_estcpu = decay_cpu(loadfac, p->p_estcpu);
312 wakeup((caddr_t)&lbolt);
313 timeout(schedcpu, (void *)0, hz);
317 * Recalculate the priority of a process after it has slept for a while.
318 * For all load averages >= 1 and max p_estcpu of 255, sleeping for at
319 * least six times the loadfactor will decay p_estcpu to zero.
322 updatepri(struct proc *p)
324 unsigned int newcpu = p->p_estcpu;
325 fixpt_t loadfac = loadfactor(averunnable.ldavg[0]);
327 if (p->p_slptime > 5 * loadfac) {
330 p->p_slptime--; /* the first time was done in schedcpu */
331 while (newcpu && --p->p_slptime)
332 newcpu = decay_cpu(loadfac, newcpu);
333 p->p_estcpu = newcpu;
339 * We're only looking at 7 bits of the address; everything is
340 * aligned to 4, lots of things are aligned to greater powers
341 * of 2. Shift right by 8, i.e. drop the bottom 256 worth.
343 #define TABLESIZE 128
344 static TAILQ_HEAD(slpquehead, thread) slpque[TABLESIZE];
345 #define LOOKUP(x) (((intptr_t)(x) >> 8) & (TABLESIZE - 1))
348 * During autoconfiguration or after a panic, a sleep will simply
349 * lower the priority briefly to allow interrupts, then return.
350 * The priority to be used (safepri) is machine-dependent, thus this
351 * value is initialized and maintained in the machine-dependent layers.
352 * This priority will typically be 0, or the lowest priority
353 * that is safe for use on the interrupt stack; it can be made
354 * higher to block network software interrupts after panics.
363 sched_quantum = hz/10;
364 hogticks = 2 * sched_quantum;
365 for (i = 0; i < TABLESIZE; i++)
366 TAILQ_INIT(&slpque[i]);
370 * General sleep call. Suspends the current process until a wakeup is
371 * performed on the specified identifier. The process will then be made
372 * runnable with the specified priority. Sleeps at most timo/hz seconds
373 * (0 means no timeout). If pri includes PCATCH flag, signals are checked
374 * before and after sleeping, else signals are not checked. Returns 0 if
375 * awakened, EWOULDBLOCK if the timeout expires. If PCATCH is set and a
376 * signal needs to be delivered, ERESTART is returned if the current system
377 * call should be restarted if possible, and EINTR is returned if the system
378 * call should be interrupted by the signal (return EINTR).
380 * If the process has P_CURPROC set mi_switch() will not re-queue it to
381 * the userland scheduler queues because we are in a SSLEEP state. If
382 * we are not the current process then we have to remove ourselves from
383 * the scheduler queues.
385 * YYY priority now unused
388 tsleep(ident, priority, wmesg, timo)
393 struct thread *td = curthread;
394 struct proc *p = td->td_proc; /* may be NULL */
395 int s, sig = 0, catch = priority & PCATCH;
396 int id = LOOKUP(ident);
397 struct callout_handle thandle;
400 * NOTE: removed KTRPOINT, it could cause races due to blocking
401 * even in stable. Just scrap it for now.
405 if (cold || panicstr) {
407 * After a panic, or during autoconfiguration,
408 * just give interrupts a chance, then just return;
409 * don't run any other procs or panic below,
410 * in case this is the idle process and already asleep.
416 KASSERT(ident != NULL, ("tsleep: no ident"));
417 KASSERT(p == NULL || p->p_stat == SRUN, ("tsleep %p %s %d",
418 ident, wmesg, p->p_stat));
421 td->td_wchan = ident;
422 td->td_wmesg = wmesg;
425 lwkt_deschedule_self();
426 TAILQ_INSERT_TAIL(&slpque[id], td, td_threadq);
428 thandle = timeout(endtsleep, (void *)td, timo);
430 * We put ourselves on the sleep queue and start our timeout
431 * before calling CURSIG, as we could stop there, and a wakeup
432 * or a SIGCONT (or both) could occur while we were stopped.
433 * A SIGCONT would cause us to be marked as SSLEEP
434 * without resuming us, thus we must be ready for sleep
435 * when CURSIG is called. If the wakeup happens while we're
436 * stopped, p->p_wchan will be 0 upon return from CURSIG.
440 p->p_flag |= P_SINTR;
441 if ((sig = CURSIG(p))) {
444 lwkt_schedule_self();
449 if (p->p_wchan == NULL) {
458 * If we are not the current process we have to remove ourself
459 * from the run queue.
461 KASSERT(p->p_stat == SRUN, ("PSTAT NOT SRUN %d %d", p->p_pid, p->p_stat));
463 * If this is the current 'user' process schedule another one.
465 clrrunnable(p, SSLEEP);
466 p->p_stats->p_ru.ru_nvcsw++;
468 KASSERT(p->p_stat == SRUN, ("tsleep: stat not srun"));
475 p->p_flag &= ~P_SINTR;
477 if (td->td_flags & TDF_TIMEOUT) {
478 td->td_flags &= ~TDF_TIMEOUT;
480 return (EWOULDBLOCK);
482 untimeout(endtsleep, (void *)td, thandle);
485 if (catch && (sig != 0 || (sig = CURSIG(p)))) {
486 if (SIGISMEMBER(p->p_sigacts->ps_sigintr, sig))
497 * General sleep call. Suspends the current process until a wakeup is
498 * performed on the specified xwait structure. The process will then be made
499 * runnable with the specified priority. Sleeps at most timo/hz seconds
500 * (0 means no timeout). If pri includes PCATCH flag, signals are checked
501 * before and after sleeping, else signals are not checked. Returns 0 if
502 * awakened, EWOULDBLOCK if the timeout expires. If PCATCH is set and a
503 * signal needs to be delivered, ERESTART is returned if the current system
504 * call should be restarted if possible, and EINTR is returned if the system
505 * call should be interrupted by the signal (return EINTR).
507 * If the passed generation number is different from the generation number
508 * in the xwait, return immediately.
511 xsleep(struct xwait *w, int priority, const char *wmesg, int timo, int *gen)
513 struct thread *td = curthread;
514 struct proc *p = td->td_proc;
515 int s, sig, catch = priority & PCATCH;
516 struct callout_handle thandle;
519 if (KTRPOINT(td, KTR_CSW))
520 ktrcsw(p->p_tracep, 1, 0);
524 if (cold || panicstr) {
526 * After a panic, or during autoconfiguration,
527 * just give interrupts a chance, then just return;
528 * don't run any other procs or panic below,
529 * in case this is the idle process and already asleep.
535 KASSERT(p != NULL, ("xsleep1"));
536 KASSERT(w != NULL && p->p_stat == SRUN, ("xsleep"));
539 * If the generation number does not match we return immediately.
541 if (*gen != w->gen) {
545 if (KTRPOINT(td, KTR_CSW))
546 ktrcsw(p->p_tracep, 0, 0);
554 p->p_flag |= P_XSLEEP;
555 TAILQ_INSERT_TAIL(&w->waitq, p, p_procq);
557 thandle = timeout(endtsleep, (void *)p, timo);
559 * We put ourselves on the sleep queue and start our timeout
560 * before calling CURSIG, as we could stop there, and a wakeup
561 * or a SIGCONT (or both) could occur while we were stopped.
562 * A SIGCONT would cause us to be marked as SSLEEP
563 * without resuming us, thus we must be ready for sleep
564 * when CURSIG is called. If the wakeup happens while we're
565 * stopped, p->p_wchan will be 0 upon return from CURSIG.
568 p->p_flag |= P_SINTR;
569 if ((sig = CURSIG(p))) {
572 lwkt_schedule_self();
577 if (p->p_wchan == NULL) {
584 clrrunnable(p, SSLEEP);
585 p->p_stats->p_ru.ru_nvcsw++;
588 *gen = w->gen; /* update generation number */
590 p->p_flag &= ~P_SINTR;
591 if (p->p_flag & P_TIMEOUT) {
592 p->p_flag &= ~P_TIMEOUT;
595 if (KTRPOINT(td, KTR_CSW))
596 ktrcsw(p->p_tracep, 0, 0);
598 return (EWOULDBLOCK);
601 untimeout(endtsleep, (void *)p, thandle);
602 if (catch && (sig != 0 || (sig = CURSIG(p)))) {
604 if (KTRPOINT(td, KTR_CSW))
605 ktrcsw(p->p_tracep, 0, 0);
607 if (SIGISMEMBER(p->p_sigacts->ps_sigintr, sig))
612 if (KTRPOINT(td, KTR_CSW))
613 ktrcsw(p->p_tracep, 0, 0);
621 * Implement the timeout for tsleep. We interlock against
622 * wchan when setting TDF_TIMEOUT. For processes we remove
623 * the sleep if the process is stopped rather then sleeping,
624 * so it remains stopped.
635 td->td_flags |= TDF_TIMEOUT;
636 if ((p = td->td_proc) != NULL) {
637 if (p->p_stat == SSLEEP)
650 * Remove a process from its wait queue
653 unsleep(struct thread *td)
660 if (p->p_flag & P_XSLEEP) {
661 struct xwait *w = p->p_wchan;
662 TAILQ_REMOVE(&w->waitq, p, p_procq);
663 p->p_flag &= ~P_XSLEEP;
666 TAILQ_REMOVE(&slpque[LOOKUP(td->td_wchan)], td, td_threadq);
674 * Make all processes sleeping on the explicit lock structure runnable.
677 xwakeup(struct xwait *w)
684 while ((p = TAILQ_FIRST(&w->waitq)) != NULL) {
685 TAILQ_REMOVE(&w->waitq, p, p_procq);
686 KASSERT(p->p_wchan == w && (p->p_flag & P_XSLEEP),
687 ("xwakeup: wchan mismatch for %p (%p/%p) %08x", p, p->p_wchan, w, p->p_flag & P_XSLEEP));
689 p->p_flag &= ~P_XSLEEP;
690 if (p->p_stat == SSLEEP) {
691 /* OPTIMIZED EXPANSION OF setrunnable(p); */
692 if (p->p_slptime > 1)
696 if (p->p_flag & P_INMEM) {
700 p->p_flag |= P_SWAPINREQ;
701 wakeup((caddr_t)&proc0);
710 * Make all processes sleeping on the specified identifier runnable.
713 _wakeup(void *ident, int count)
715 struct slpquehead *qp;
720 int id = LOOKUP(ident);
725 for (td = TAILQ_FIRST(qp); td != NULL; td = ntd) {
726 ntd = TAILQ_NEXT(td, td_threadq);
727 if (td->td_wchan == ident) {
728 TAILQ_REMOVE(qp, td, td_threadq);
730 if ((p = td->td_proc) != NULL && p->p_stat == SSLEEP) {
731 /* OPTIMIZED EXPANSION OF setrunnable(p); */
732 if (p->p_slptime > 1)
736 if (p->p_flag & P_INMEM) {
740 p->p_flag |= P_SWAPINREQ;
741 wakeup((caddr_t)&proc0);
743 /* END INLINE EXPANSION */
744 } else if (p == NULL) {
762 wakeup_one(void *ident)
768 * Release the P_CURPROC designation on a process in order to allow the
769 * userland scheduler to schedule another one. This places a runnable
770 * process back on the userland scheduler's run queue.
772 * Note that losing P_CURPROC does not effect LWKT scheduling, you can
773 * still tsleep/wakeup after having lost P_CURPROC, but userret() will
774 * not return to user mode until it gets it back.
778 _relscurproc(struct proc *p)
783 if (p->p_flag & P_CURPROC) {
784 p->p_flag &= ~P_CURPROC;
785 lwkt_deschedule_self();
786 if (p->p_stat == SRUN && (p->p_flag & P_INMEM)) {
789 if ((np = chooseproc()) != NULL) {
790 np->p_flag |= P_CURPROC;
791 lwkt_schedule(np->p_thread);
793 KKASSERT(mycpu->gd_uprocscheduled == 1);
794 mycpu->gd_uprocscheduled = 0;
801 relscurproc(struct proc *p)
807 * The machine independent parts of mi_switch().
808 * Must be called at splstatclock() or higher.
813 struct thread *td = curthread;
814 struct proc *p = td->td_proc; /* XXX */
820 * XXX this spl is almost unnecessary. It is partly to allow for
821 * sloppy callers that don't do it (issignal() via CURSIG() is the
822 * main offender). It is partly to work around a bug in the i386
823 * cpu_switch() (the ipl is not preserved). We ran for years
824 * without it. I think there was only a interrupt latency problem.
825 * The main caller, tsleep(), does an splx() a couple of instructions
826 * after calling here. The buggy caller, issignal(), usually calls
827 * here at spl0() and sometimes returns at splhigh(). The process
828 * then runs for a little too long at splhigh(). The ipl gets fixed
829 * when the process returns to user mode (or earlier).
831 * It would probably be better to always call here at spl0(). Callers
832 * are prepared to give up control to another process, so they must
833 * be prepared to be interrupted. The clock stuff here may not
834 * actually need splstatclock().
840 * If the process being switched out is the 'current' process then
841 * we have to lose the P_CURPROC designation and choose a new
842 * process. If the process is not being LWKT managed and it is in
843 * SRUN we have to setrunqueue it.
847 #ifdef SIMPLELOCK_DEBUG
848 if (p->p_simple_locks)
849 printf("sleep: holding simple lock\n");
853 * Check if the process exceeds its cpu resource allocation.
854 * If over max, kill it. Time spent in interrupts is not
855 * included. YYY 64 bit match is expensive. Ick.
857 ttime = td->td_sticks + td->td_uticks;
858 if (p->p_stat != SZOMB && p->p_limit->p_cpulimit != RLIM_INFINITY &&
859 ttime > p->p_limit->p_cpulimit) {
860 rlim = &p->p_rlimit[RLIMIT_CPU];
861 if (ttime / (rlim_t)1000000 >= rlim->rlim_max) {
862 killproc(p, "exceeded maximum CPU limit");
865 if (rlim->rlim_cur < rlim->rlim_max) {
866 /* XXX: we should make a private copy */
873 * Pick a new current process and record its start time.
874 * YYY lwkt_switch() will run the heavy weight process restoration
875 * code, which removes the target thread and process from their
876 * respective run queues to temporarily mimic 5.x behavior.
877 * YYY the userland scheduler should pick only one user process
878 * at a time to run per cpu.
880 mycpu->gd_cnt.v_swtch++;
887 * Change process state to be runnable,
888 * placing it on the run queue if it is in memory,
889 * and awakening the swapper if it isn't in memory.
892 setrunnable(struct proc *p)
902 panic("setrunnable");
905 unsleep(p->p_thread); /* e.g. when sending signals */
912 if (p->p_flag & P_INMEM)
915 if (p->p_slptime > 1)
918 if ((p->p_flag & P_INMEM) == 0) {
919 p->p_flag |= P_SWAPINREQ;
920 wakeup((caddr_t)&proc0);
927 * Change the process state to NOT be runnable, removing it from the run
928 * queue. If P_CURPROC is not set and we are in SRUN the process is on the
929 * run queue (If P_INMEM is not set then it isn't because it is swapped).
932 clrrunnable(struct proc *p, int stat)
939 if ((p->p_flag & (P_INMEM|P_CURPROC)) == P_INMEM)
950 * yield / synchronous reschedule
952 * Simply calling mi_switch() has the effect we want. mi_switch will
953 * deschedule the current thread, make sure the current process is on
954 * the run queue, and then choose and reschedule another process.
959 struct proc *p = curproc;
964 KKASSERT(p->p_stat == SRUN);
965 if ((p->p_flag & (P_INMEM|P_CURPROC)) == (P_INMEM|P_CURPROC))
967 lwkt_deschedule_self();
969 p->p_stats->p_ru.ru_nivcsw++;
975 * Compute the priority of a process when running in user mode.
976 * Arrange to reschedule if the resulting priority is better
977 * than that of the current process.
979 * YYY real time / idle procs do not use p_priority XXX
982 resetpriority(struct proc *p)
984 unsigned int newpriority;
988 if (p->p_rtprio.type != RTP_PRIO_NORMAL)
990 newpriority = PUSER + p->p_estcpu / INVERSE_ESTCPU_WEIGHT +
991 NICE_WEIGHT * p->p_nice;
992 newpriority = min(newpriority, MAXPRI);
993 npq = newpriority / PPQ;
995 opq = p->p_priority / PPQ;
996 if (p->p_stat == SRUN && (p->p_flag & (P_CURPROC|P_INMEM)) == P_INMEM
999 * We have to move the process to another queue
1002 p->p_priority = newpriority;
1006 * Not on a queue or is on the same queue, we can just
1010 p->p_priority = newpriority;
1017 * Compute a tenex style load average of a quantity on
1018 * 1, 5 and 15 minute intervals.
1024 struct loadavg *avg;
1029 LIST_FOREACH(p, &allproc, p_list) {
1030 switch (p->p_stat) {
1036 for (i = 0; i < 3; i++)
1037 avg->ldavg[i] = (cexp[i] * avg->ldavg[i] +
1038 nrun * FSCALE * (FSCALE - cexp[i])) >> FSHIFT;
1041 * Schedule the next update to occur after 5 seconds, but add a
1042 * random variation to avoid synchronisation with processes that
1043 * run at regular intervals.
1045 callout_reset(&loadav_callout, hz * 4 + (int)(random() % (hz * 2 + 1)),
1055 callout_init(&loadav_callout);
1057 /* Kick off timeout driven events by calling first time. */
1064 * We adjust the priority of the current process. The priority of
1065 * a process gets worse as it accumulates CPU time. The cpu usage
1066 * estimator (p_estcpu) is increased here. resetpriority() will
1067 * compute a different priority each time p_estcpu increases by
1068 * INVERSE_ESTCPU_WEIGHT
1069 * (until MAXPRI is reached). The cpu usage estimator ramps up
1070 * quite quickly when the process is running (linearly), and decays
1071 * away exponentially, at a rate which is proportionally slower when
1072 * the system is busy. The basic principle is that the system will
1073 * 90% forget that the process used a lot of CPU time in 5 * loadav
1074 * seconds. This causes the system to favor processes which haven't
1075 * run much recently, and to round-robin among other processes.
1083 p->p_estcpu = ESTCPULIM(p->p_estcpu + 1);
1084 if ((p->p_estcpu % INVERSE_ESTCPU_WEIGHT) == 0)