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 1.45 2005/06/26 04:36:31 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 (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. */
71 int ncpus2, ncpus2_shift, ncpus2_mask;
74 static struct callout loadav_callout;
75 static struct callout roundrobin_callout;
76 static struct callout schedcpu_callout;
78 struct loadavg averunnable =
79 { {0, 0, 0}, FSCALE }; /* load average, of runnable procs */
81 * Constants for averages over 1, 5, and 15 minutes
82 * when sampling at 5 second intervals.
84 static fixpt_t cexp[3] = {
85 0.9200444146293232 * FSCALE, /* exp(-1/12) */
86 0.9834714538216174 * FSCALE, /* exp(-1/60) */
87 0.9944598480048967 * FSCALE, /* exp(-1/180) */
90 static void endtsleep (void *);
91 static void loadav (void *arg);
92 static void roundrobin (void *arg);
93 static void schedcpu (void *arg);
94 static void updatepri (struct proc *p);
97 sysctl_kern_quantum(SYSCTL_HANDLER_ARGS)
101 new_val = sched_quantum * tick;
102 error = sysctl_handle_int(oidp, &new_val, 0, req);
103 if (error != 0 || req->newptr == NULL)
107 sched_quantum = new_val / tick;
108 hogticks = 2 * sched_quantum;
112 SYSCTL_PROC(_kern, OID_AUTO, quantum, CTLTYPE_INT|CTLFLAG_RW,
113 0, sizeof sched_quantum, sysctl_kern_quantum, "I", "");
116 roundrobin_interval(void)
118 return (sched_quantum);
122 * Force switch among equal priority processes every 100ms.
124 * WARNING! The MP lock is not held on ipi message remotes.
129 roundrobin_remote(void *arg)
131 struct proc *p = lwkt_preempted_proc();
132 if (p == NULL || RTP_PRIO_NEED_RR(p->p_rtprio.type))
139 roundrobin(void *arg)
141 struct proc *p = lwkt_preempted_proc();
142 if (p == NULL || RTP_PRIO_NEED_RR(p->p_rtprio.type))
145 lwkt_send_ipiq_mask(mycpu->gd_other_cpus, roundrobin_remote, NULL);
147 callout_reset(&roundrobin_callout, sched_quantum, roundrobin, NULL);
153 resched_cpus(u_int32_t mask)
155 lwkt_send_ipiq_mask(mask, roundrobin_remote, NULL);
161 * The load average is scaled by FSCALE (2048 typ). The estimated cpu is
162 * incremented at a rate of ESTCPUVFREQ per second (40hz typ), but this is
163 * divided up across all cpu bound processes running in the system so an
164 * individual process will get less under load. ESTCPULIM typicaly caps
165 * out at ESTCPUMAX (around 376, or 11 nice levels).
167 * Generally speaking the decay equation needs to break-even on growth
168 * at the limit at all load levels >= 1.0, so if the estimated cpu for
169 * a process increases by (ESTVCPUFREQ / load) per second, then the decay
170 * should reach this value when estcpu reaches ESTCPUMAX. That calculation
173 * ESTCPUMAX * decay = ESTCPUVFREQ / load
174 * decay = ESTCPUVFREQ / (load * ESTCPUMAX)
175 * decay = estcpu * 0.053 / load
177 * If the load is less then 1.0 we assume a load of 1.0.
180 #define cload(loadav) ((loadav) < FSCALE ? FSCALE : (loadav))
181 #define decay_cpu(loadav,estcpu) \
182 ((estcpu) * (FSCALE * ESTCPUVFREQ / ESTCPUMAX) / cload(loadav))
184 /* decay 95% of `p_pctcpu' in 60 seconds; see CCPU_SHIFT before changing */
185 static fixpt_t ccpu = 0.95122942450071400909 * FSCALE; /* exp(-1/20) */
186 SYSCTL_INT(_kern, OID_AUTO, ccpu, CTLFLAG_RD, &ccpu, 0, "");
188 /* kernel uses `FSCALE', userland (SHOULD) use kern.fscale */
189 static int fscale __unused = FSCALE;
190 SYSCTL_INT(_kern, OID_AUTO, fscale, CTLFLAG_RD, 0, FSCALE, "");
193 * If `ccpu' is not equal to `exp(-1/20)' and you still want to use the
194 * faster/more-accurate formula, you'll have to estimate CCPU_SHIFT below
195 * and possibly adjust FSHIFT in "param.h" so that (FSHIFT >= CCPU_SHIFT).
197 * To estimate CCPU_SHIFT for exp(-1/20), the following formula was used:
198 * 1 - exp(-1/20) ~= 0.0487 ~= 0.0488 == 1 (fixed pt, *11* bits).
200 * If you don't want to bother with the faster/more-accurate formula, you
201 * can set CCPU_SHIFT to (FSHIFT + 1) which will use a slower/less-accurate
202 * (more general) method of calculating the %age of CPU used by a process.
204 #define CCPU_SHIFT 11
207 * Recompute process priorities, once a second.
213 fixpt_t loadfac = averunnable.ldavg[0];
217 FOREACH_PROC_IN_SYSTEM(p) {
219 * Increment time in/out of memory and sleep time
220 * (if sleeping). We ignore overflow; with 16-bit int's
221 * (remember them?) overflow takes 45 days.
224 if (p->p_stat == SSLEEP || p->p_stat == SSTOP)
226 p->p_pctcpu = (p->p_pctcpu * ccpu) >> FSHIFT;
229 * If the process has slept the entire second,
230 * stop recalculating its priority until it wakes up.
232 * Note that interactive calculations do not occur for
233 * long sleeps (because that isn't necessarily indicative
234 * of an interactive process).
236 if (p->p_slptime > 1)
238 /* prevent state changes and protect run queue */
242 * p_cpticks runs at ESTCPUFREQ but must be divided by the
243 * load average for par-100% use. Higher p_interactive
244 * values mean less interactive, lower values mean more
247 if ((((fixpt_t)p->p_cpticks * cload(loadfac)) >> FSHIFT) >
249 p->p_usched->heuristic_estcpu(p, 1);
251 p->p_usched->heuristic_estcpu(p, -1);
254 * p_pctcpu is only for ps.
256 #if (FSHIFT >= CCPU_SHIFT)
257 p->p_pctcpu += (ESTCPUFREQ == 100)?
258 ((fixpt_t) p->p_cpticks) << (FSHIFT - CCPU_SHIFT):
259 100 * (((fixpt_t) p->p_cpticks)
260 << (FSHIFT - CCPU_SHIFT)) / ESTCPUFREQ;
262 p->p_pctcpu += ((FSCALE - ccpu) *
263 (p->p_cpticks * FSCALE / ESTCPUFREQ)) >> FSHIFT;
266 ndecay = decay_cpu(loadfac, p->p_estcpu);
267 if (p->p_estcpu > ndecay)
268 p->p_estcpu -= ndecay;
271 p->p_usched->resetpriority(p);
274 wakeup((caddr_t)&lbolt);
275 callout_reset(&schedcpu_callout, hz, schedcpu, NULL);
279 * Recalculate the priority of a process after it has slept for a while.
280 * For all load averages >= 1 and max p_estcpu of 255, sleeping for at
281 * least six times the loadfactor will decay p_estcpu to zero.
284 updatepri(struct proc *p)
288 ndecay = decay_cpu(averunnable.ldavg[0], p->p_estcpu) * p->p_slptime;
289 if (p->p_estcpu > ndecay)
290 p->p_estcpu -= ndecay;
293 p->p_usched->resetpriority(p);
297 * We're only looking at 7 bits of the address; everything is
298 * aligned to 4, lots of things are aligned to greater powers
299 * of 2. Shift right by 8, i.e. drop the bottom 256 worth.
301 #define TABLESIZE 128
302 static TAILQ_HEAD(slpquehead, thread) slpque[TABLESIZE];
303 #define LOOKUP(x) (((intptr_t)(x) >> 8) & (TABLESIZE - 1))
306 * During autoconfiguration or after a panic, a sleep will simply
307 * lower the priority briefly to allow interrupts, then return.
315 sched_quantum = hz/10;
316 hogticks = 2 * sched_quantum;
317 for (i = 0; i < TABLESIZE; i++)
318 TAILQ_INIT(&slpque[i]);
322 * General sleep call. Suspends the current process until a wakeup is
323 * performed on the specified identifier. The process will then be made
324 * runnable with the specified priority. Sleeps at most timo/hz seconds
325 * (0 means no timeout). If flags includes PCATCH flag, signals are checked
326 * before and after sleeping, else signals are not checked. Returns 0 if
327 * awakened, EWOULDBLOCK if the timeout expires. If PCATCH is set and a
328 * signal needs to be delivered, ERESTART is returned if the current system
329 * call should be restarted if possible, and EINTR is returned if the system
330 * call should be interrupted by the signal (return EINTR).
332 * Note that if we are a process, we release_curproc() before messing with
333 * the LWKT scheduler.
336 tsleep(void *ident, int flags, const char *wmesg, int timo)
338 struct thread *td = curthread;
339 struct proc *p = td->td_proc; /* may be NULL */
340 int sig = 0, catch = flags & PCATCH;
341 int id = LOOKUP(ident);
343 struct callout thandle;
346 * NOTE: removed KTRPOINT, it could cause races due to blocking
347 * even in stable. Just scrap it for now.
349 if (cold || panicstr) {
351 * After a panic, or during autoconfiguration,
352 * just give interrupts a chance, then just return;
353 * don't run any other procs or panic below,
354 * in case this is the idle process and already asleep.
357 oldpri = td->td_pri & TDPRI_MASK;
358 lwkt_setpri_self(safepri);
360 lwkt_setpri_self(oldpri);
363 KKASSERT(td != &mycpu->gd_idlethread); /* you must be kidding! */
364 crit_enter_quick(td);
365 KASSERT(ident != NULL, ("tsleep: no ident"));
366 KASSERT(p == NULL || p->p_stat == SRUN, ("tsleep %p %s %d",
367 ident, wmesg, p->p_stat));
369 td->td_wchan = ident;
370 td->td_wmesg = wmesg;
371 td->td_wdomain = flags & PDOMAIN_MASK;
373 if (flags & PNORESCHED)
374 td->td_flags |= TDF_NORESCHED;
375 p->p_usched->release_curproc(p);
378 lwkt_deschedule_self(td);
379 TAILQ_INSERT_TAIL(&slpque[id], td, td_threadq);
381 callout_init(&thandle);
382 callout_reset(&thandle, timo, endtsleep, td);
385 * We put ourselves on the sleep queue and start our timeout
386 * before calling CURSIG, as we could stop there, and a wakeup
387 * or a SIGCONT (or both) could occur while we were stopped.
388 * A SIGCONT would cause us to be marked as SSLEEP
389 * without resuming us, thus we must be ready for sleep
390 * when CURSIG is called. If the wakeup happens while we're
391 * stopped, td->td_wchan will be 0 upon return from CURSIG.
395 p->p_flag |= P_SINTR;
396 if ((sig = CURSIG(p))) {
399 lwkt_schedule_self(td);
404 if (td->td_wchan == NULL) {
413 * If we are not the current process we have to remove ourself
414 * from the run queue.
416 KASSERT(p->p_stat == SRUN, ("PSTAT NOT SRUN %d %d", p->p_pid, p->p_stat));
418 * If this is the current 'user' process schedule another one.
420 clrrunnable(p, SSLEEP);
421 p->p_stats->p_ru.ru_nvcsw++;
423 KASSERT(p->p_stat == SRUN, ("tsleep: stat not srun"));
429 p->p_flag &= ~P_SINTR;
431 td->td_flags &= ~TDF_NORESCHED;
432 if (td->td_flags & TDF_TIMEOUT) {
433 td->td_flags &= ~TDF_TIMEOUT;
435 return (EWOULDBLOCK);
437 callout_stop(&thandle);
438 } else if (td->td_wmesg) {
440 * This can happen if a thread is woken up directly. Clear
441 * wmesg to avoid debugging confusion.
445 /* inline of iscaught() */
447 if (catch && (sig != 0 || (sig = CURSIG(p)))) {
448 if (SIGISMEMBER(p->p_sigacts->ps_sigintr, sig))
457 * Implement the timeout for tsleep. We interlock against
458 * wchan when setting TDF_TIMEOUT. For processes we remove
459 * the sleep if the process is stopped rather then sleeping,
460 * so it remains stopped.
470 td->td_flags |= TDF_TIMEOUT;
471 if ((p = td->td_proc) != NULL) {
472 if (p->p_stat == SSLEEP)
485 * Remove a process from its wait queue
488 unsleep(struct thread *td)
492 TAILQ_REMOVE(&slpque[LOOKUP(td->td_wchan)], td, td_threadq);
499 * Make all processes sleeping on the specified identifier runnable.
502 _wakeup(void *ident, int domain, int count)
504 struct slpquehead *qp;
508 int id = LOOKUP(ident);
513 for (td = TAILQ_FIRST(qp); td != NULL; td = ntd) {
514 ntd = TAILQ_NEXT(td, td_threadq);
515 if (td->td_wchan == ident && td->td_wdomain == domain) {
516 TAILQ_REMOVE(qp, td, td_threadq);
518 if ((p = td->td_proc) != NULL && p->p_stat == SSLEEP) {
519 /* OPTIMIZED EXPANSION OF setrunnable(p); */
520 if (p->p_slptime > 1)
524 if (p->p_flag & P_INMEM) {
526 * LWKT scheduled now, there is no
527 * userland runq interaction until
528 * the thread tries to return to user
535 p->p_flag |= P_SWAPINREQ;
536 wakeup((caddr_t)&proc0);
538 /* END INLINE EXPANSION */
539 } else if (p == NULL) {
553 _wakeup(ident, 0, 0);
557 wakeup_one(void *ident)
559 _wakeup(ident, 0, 1);
563 wakeup_domain(void *ident, int domain)
565 _wakeup(ident, domain, 0);
569 wakeup_domain_one(void *ident, int domain)
571 _wakeup(ident, domain, 1);
575 * The machine independent parts of mi_switch().
577 * 'p' must be the current process.
580 mi_switch(struct proc *p)
582 thread_t td = p->p_thread;
586 KKASSERT(td == mycpu->gd_curthread);
588 crit_enter_quick(td);
591 * Check if the process exceeds its cpu resource allocation.
592 * If over max, kill it. Time spent in interrupts is not
593 * included. YYY 64 bit match is expensive. Ick.
595 ttime = td->td_sticks + td->td_uticks;
596 if (p->p_stat != SZOMB && p->p_limit->p_cpulimit != RLIM_INFINITY &&
597 ttime > p->p_limit->p_cpulimit) {
598 rlim = &p->p_rlimit[RLIMIT_CPU];
599 if (ttime / (rlim_t)1000000 >= rlim->rlim_max) {
600 killproc(p, "exceeded maximum CPU limit");
603 if (rlim->rlim_cur < rlim->rlim_max) {
604 /* XXX: we should make a private copy */
611 * If we are in a SSTOPped state we deschedule ourselves.
612 * YYY this needs to be cleaned up, remember that LWKTs stay on
613 * their run queue which works differently then the user scheduler
614 * which removes the process from the runq when it runs it.
616 mycpu->gd_cnt.v_swtch++;
617 if (p->p_stat == SSTOP)
618 lwkt_deschedule_self(td);
624 * Change process state to be runnable,
625 * placing it on the run queue if it is in memory,
626 * and awakening the swapper if it isn't in memory.
629 setrunnable(struct proc *p)
638 panic("setrunnable");
641 unsleep(p->p_thread); /* e.g. when sending signals */
650 * The process is controlled by LWKT at this point, we do not mess
651 * around with the userland scheduler until the thread tries to
652 * return to user mode.
654 if (p->p_flag & P_INMEM)
655 lwkt_schedule(p->p_thread);
657 if (p->p_slptime > 1)
660 if ((p->p_flag & P_INMEM) == 0) {
661 p->p_flag |= P_SWAPINREQ;
662 wakeup((caddr_t)&proc0);
667 * Yield / synchronous reschedule. This is a bit tricky because the trap
668 * code might have set a lazy release on the switch function. Setting
669 * P_PASSIVE_ACQ will ensure that the lazy release executes when we call
670 * switch, and that we are given a greater chance of affinity with our
673 * We call lwkt_setpri_self() to rotate our thread to the end of the lwkt
674 * run queue. lwkt_switch() will also execute any assigned passive release
675 * (which usually calls release_curproc()), allowing a same/higher priority
676 * process to be designated as the current process.
678 * While it is possible for a lower priority process to be designated,
679 * it's call to lwkt_maybe_switch() in acquire_curproc() will likely
680 * round-robin back to us and we will be able to re-acquire the current
681 * process designation.
686 struct thread *td = curthread;
687 struct proc *p = td->td_proc;
689 lwkt_setpri_self(td->td_pri & TDPRI_MASK);
691 p->p_flag |= P_PASSIVE_ACQ;
693 p->p_flag &= ~P_PASSIVE_ACQ;
700 * Change the process state to NOT be runnable, removing it from the run
704 clrrunnable(struct proc *p, int stat)
706 crit_enter_quick(p->p_thread);
707 if (p->p_stat == SRUN && (p->p_flag & P_ONRUNQ))
708 p->p_usched->remrunqueue(p);
710 crit_exit_quick(p->p_thread);
714 * Compute a tenex style load average of a quantity on
715 * 1, 5 and 15 minute intervals.
727 FOREACH_PROC_IN_SYSTEM(p) {
730 if ((td = p->p_thread) == NULL)
732 if (td->td_flags & TDF_BLOCKED)
742 for (i = 0; i < 3; i++)
743 avg->ldavg[i] = (cexp[i] * avg->ldavg[i] +
744 nrun * FSCALE * (FSCALE - cexp[i])) >> FSHIFT;
747 * Schedule the next update to occur after 5 seconds, but add a
748 * random variation to avoid synchronisation with processes that
749 * run at regular intervals.
751 callout_reset(&loadav_callout, hz * 4 + (int)(random() % (hz * 2 + 1)),
757 sched_setup(void *dummy)
759 callout_init(&loadav_callout);
760 callout_init(&roundrobin_callout);
761 callout_init(&schedcpu_callout);
763 /* Kick off timeout driven events by calling first time. */
770 * We adjust the priority of the current process. The priority of
771 * a process gets worse as it accumulates CPU time. The cpu usage
772 * estimator (p_estcpu) is increased here. resetpriority() will
773 * compute a different priority each time p_estcpu increases by
774 * INVERSE_ESTCPU_WEIGHT * (until MAXPRI is reached).
776 * The cpu usage estimator ramps up quite quickly when the process is
777 * running (linearly), and decays away exponentially, at a rate which
778 * is proportionally slower when the system is busy. The basic principle
779 * is that the system will 90% forget that the process used a lot of CPU
780 * time in 5 * loadav seconds. This causes the system to favor processes
781 * which haven't run much recently, and to round-robin among other processes.
783 * The actual schedulerclock interrupt rate is ESTCPUFREQ, but we generally
784 * want to ramp-up at a faster rate, ESTCPUVFREQ, so p_estcpu is scaled
785 * by (ESTCPUVFREQ / ESTCPUFREQ). You can control the ramp-up/ramp-down
786 * rate by adjusting ESTCPUVFREQ in sys/proc.h in integer multiples
789 * WARNING! called from a fast-int or an IPI, the MP lock MIGHT NOT BE HELD
790 * and we cannot block.
793 schedulerclock(void *dummy)
799 if ((p = td->td_proc) != NULL) {
800 p->p_cpticks++; /* cpticks runs at ESTCPUFREQ */
801 p->p_estcpu = ESTCPULIM(p->p_estcpu + ESTCPUVFREQ / ESTCPUFREQ);
803 p->p_usched->resetpriority(p);