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 1.47 2005/06/27 18:37:57 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 schedcpu_callout;
77 struct loadavg averunnable =
78 { {0, 0, 0}, FSCALE }; /* load average, of runnable procs */
80 * Constants for averages over 1, 5, and 15 minutes
81 * when sampling at 5 second intervals.
83 static fixpt_t cexp[3] = {
84 0.9200444146293232 * FSCALE, /* exp(-1/12) */
85 0.9834714538216174 * FSCALE, /* exp(-1/60) */
86 0.9944598480048967 * FSCALE, /* exp(-1/180) */
89 static void endtsleep (void *);
90 static void loadav (void *arg);
91 static void schedcpu (void *arg);
94 * Adjust the scheduler quantum. The quantum is specified in microseconds.
95 * Note that 'tick' is in microseconds per tick.
98 sysctl_kern_quantum(SYSCTL_HANDLER_ARGS)
102 new_val = sched_quantum * tick;
103 error = sysctl_handle_int(oidp, &new_val, 0, req);
104 if (error != 0 || req->newptr == NULL)
108 sched_quantum = new_val / tick;
109 hogticks = 2 * sched_quantum;
113 SYSCTL_PROC(_kern, OID_AUTO, quantum, CTLTYPE_INT|CTLFLAG_RW,
114 0, sizeof sched_quantum, sysctl_kern_quantum, "I", "");
117 * If `ccpu' is not equal to `exp(-1/20)' and you still want to use the
118 * faster/more-accurate formula, you'll have to estimate CCPU_SHIFT below
119 * and possibly adjust FSHIFT in "param.h" so that (FSHIFT >= CCPU_SHIFT).
121 * To estimate CCPU_SHIFT for exp(-1/20), the following formula was used:
122 * 1 - exp(-1/20) ~= 0.0487 ~= 0.0488 == 1 (fixed pt, *11* bits).
124 * If you don't want to bother with the faster/more-accurate formula, you
125 * can set CCPU_SHIFT to (FSHIFT + 1) which will use a slower/less-accurate
126 * (more general) method of calculating the %age of CPU used by a process.
128 * decay 95% of `p_pctcpu' in 60 seconds; see CCPU_SHIFT before changing
130 #define CCPU_SHIFT 11
132 static fixpt_t ccpu = 0.95122942450071400909 * FSCALE; /* exp(-1/20) */
133 SYSCTL_INT(_kern, OID_AUTO, ccpu, CTLFLAG_RD, &ccpu, 0, "");
136 * kernel uses `FSCALE', userland (SHOULD) use kern.fscale
138 static int fscale __unused = FSCALE;
139 SYSCTL_INT(_kern, OID_AUTO, fscale, CTLFLAG_RD, 0, FSCALE, "");
142 * Recompute process priorities, once a second.
144 * Since the userland schedulers are typically event oriented, if the
145 * estcpu calculation at wakeup() time is not sufficient to make a
146 * process runnable relative to other processes in the system we have
147 * a 1-second recalc to help out.
149 * This code also allows us to store sysclock_t data in the process structure
150 * without fear of an overrun, since sysclock_t are guarenteed to hold
151 * several seconds worth of count.
159 FOREACH_PROC_IN_SYSTEM(p) {
161 * Increment time in/out of memory and sleep time
162 * (if sleeping). We ignore overflow; with 16-bit int's
163 * (remember them?) overflow takes 45 days.
167 if (p->p_stat == SSLEEP || p->p_stat == SSTOP)
171 * Only recalculate processes that are active or have slept
172 * less then 2 seconds. The schedulers understand this.
174 if (p->p_slptime <= 1) {
175 p->p_usched->recalculate(p);
177 p->p_pctcpu = (p->p_pctcpu * ccpu) >> FSHIFT;
181 wakeup((caddr_t)&lbolt);
182 callout_reset(&schedcpu_callout, hz, schedcpu, NULL);
186 * This is only used by ps. Generate a cpu percentage use over
187 * a period of one second.
190 updatepcpu(struct proc *p, int cpticks, int ttlticks)
195 acc = (cpticks << FSHIFT) / ttlticks;
196 if (ttlticks >= ESTCPUFREQ) {
199 remticks = ESTCPUFREQ - ttlticks;
200 p->p_pctcpu = (acc * ttlticks + p->p_pctcpu * remticks) /
207 * We're only looking at 7 bits of the address; everything is
208 * aligned to 4, lots of things are aligned to greater powers
209 * of 2. Shift right by 8, i.e. drop the bottom 256 worth.
211 #define TABLESIZE 128
212 static TAILQ_HEAD(slpquehead, thread) slpque[TABLESIZE];
213 #define LOOKUP(x) (((intptr_t)(x) >> 8) & (TABLESIZE - 1))
216 * General scheduler initialization. We force a reschedule 25 times
217 * a second by default.
224 sched_quantum = (hz + 24) / 25;
225 hogticks = 2 * sched_quantum;
226 for (i = 0; i < TABLESIZE; i++)
227 TAILQ_INIT(&slpque[i]);
231 * General sleep call. Suspends the current process until a wakeup is
232 * performed on the specified identifier. The process will then be made
233 * runnable with the specified priority. Sleeps at most timo/hz seconds
234 * (0 means no timeout). If flags includes PCATCH flag, signals are checked
235 * before and after sleeping, else signals are not checked. Returns 0 if
236 * awakened, EWOULDBLOCK if the timeout expires. If PCATCH is set and a
237 * signal needs to be delivered, ERESTART is returned if the current system
238 * call should be restarted if possible, and EINTR is returned if the system
239 * call should be interrupted by the signal (return EINTR).
241 * Note that if we are a process, we release_curproc() before messing with
242 * the LWKT scheduler.
244 * During autoconfiguration or after a panic, a sleep will simply
245 * lower the priority briefly to allow interrupts, then return.
248 tsleep(void *ident, int flags, const char *wmesg, int timo)
250 struct thread *td = curthread;
251 struct proc *p = td->td_proc; /* may be NULL */
252 int sig = 0, catch = flags & PCATCH;
253 int id = LOOKUP(ident);
255 struct callout thandle;
258 * NOTE: removed KTRPOINT, it could cause races due to blocking
259 * even in stable. Just scrap it for now.
261 if (cold || panicstr) {
263 * After a panic, or during autoconfiguration,
264 * just give interrupts a chance, then just return;
265 * don't run any other procs or panic below,
266 * in case this is the idle process and already asleep.
269 oldpri = td->td_pri & TDPRI_MASK;
270 lwkt_setpri_self(safepri);
272 lwkt_setpri_self(oldpri);
275 KKASSERT(td != &mycpu->gd_idlethread); /* you must be kidding! */
276 crit_enter_quick(td);
277 KASSERT(ident != NULL, ("tsleep: no ident"));
278 KASSERT(p == NULL || p->p_stat == SRUN, ("tsleep %p %s %d",
279 ident, wmesg, p->p_stat));
281 td->td_wchan = ident;
282 td->td_wmesg = wmesg;
283 td->td_wdomain = flags & PDOMAIN_MASK;
285 if (flags & PNORESCHED)
286 td->td_flags |= TDF_NORESCHED;
287 p->p_usched->release_curproc(p);
290 lwkt_deschedule_self(td);
291 TAILQ_INSERT_TAIL(&slpque[id], td, td_threadq);
293 callout_init(&thandle);
294 callout_reset(&thandle, timo, endtsleep, td);
297 * We put ourselves on the sleep queue and start our timeout
298 * before calling CURSIG, as we could stop there, and a wakeup
299 * or a SIGCONT (or both) could occur while we were stopped.
300 * A SIGCONT would cause us to be marked as SSLEEP
301 * without resuming us, thus we must be ready for sleep
302 * when CURSIG is called. If the wakeup happens while we're
303 * stopped, td->td_wchan will be 0 upon return from CURSIG.
307 p->p_flag |= P_SINTR;
308 if ((sig = CURSIG(p))) {
311 lwkt_schedule_self(td);
316 if (td->td_wchan == NULL) {
325 * If we are not the current process we have to remove ourself
326 * from the run queue.
328 KASSERT(p->p_stat == SRUN, ("PSTAT NOT SRUN %d %d", p->p_pid, p->p_stat));
330 * If this is the current 'user' process schedule another one.
332 clrrunnable(p, SSLEEP);
333 p->p_stats->p_ru.ru_nvcsw++;
335 KASSERT(p->p_stat == SRUN, ("tsleep: stat not srun"));
341 p->p_flag &= ~P_SINTR;
343 td->td_flags &= ~TDF_NORESCHED;
344 if (td->td_flags & TDF_TIMEOUT) {
345 td->td_flags &= ~TDF_TIMEOUT;
347 return (EWOULDBLOCK);
349 callout_stop(&thandle);
350 } else if (td->td_wmesg) {
352 * This can happen if a thread is woken up directly. Clear
353 * wmesg to avoid debugging confusion.
357 /* inline of iscaught() */
359 if (catch && (sig != 0 || (sig = CURSIG(p)))) {
360 if (SIGISMEMBER(p->p_sigacts->ps_sigintr, sig))
369 * Implement the timeout for tsleep. We interlock against
370 * wchan when setting TDF_TIMEOUT. For processes we remove
371 * the sleep if the process is stopped rather then sleeping,
372 * so it remains stopped.
382 td->td_flags |= TDF_TIMEOUT;
383 if ((p = td->td_proc) != NULL) {
384 if (p->p_stat == SSLEEP)
397 * Remove a process from its wait queue
400 unsleep(struct thread *td)
404 TAILQ_REMOVE(&slpque[LOOKUP(td->td_wchan)], td, td_threadq);
411 * Make all processes sleeping on the specified identifier runnable.
414 _wakeup(void *ident, int domain, int count)
416 struct slpquehead *qp;
420 int id = LOOKUP(ident);
425 for (td = TAILQ_FIRST(qp); td != NULL; td = ntd) {
426 ntd = TAILQ_NEXT(td, td_threadq);
427 if (td->td_wchan == ident && td->td_wdomain == domain) {
428 TAILQ_REMOVE(qp, td, td_threadq);
430 if ((p = td->td_proc) != NULL && p->p_stat == SSLEEP) {
432 if (p->p_flag & P_INMEM) {
434 * LWKT scheduled now, there is no
435 * userland runq interaction until
436 * the thread tries to return to user
437 * mode. We do NOT call setrunqueue().
441 p->p_flag |= P_SWAPINREQ;
442 wakeup((caddr_t)&proc0);
444 /* END INLINE EXPANSION */
445 } else if (p == NULL) {
459 _wakeup(ident, 0, 0);
463 wakeup_one(void *ident)
465 _wakeup(ident, 0, 1);
469 wakeup_domain(void *ident, int domain)
471 _wakeup(ident, domain, 0);
475 wakeup_domain_one(void *ident, int domain)
477 _wakeup(ident, domain, 1);
481 * The machine independent parts of mi_switch().
483 * 'p' must be the current process.
486 mi_switch(struct proc *p)
488 thread_t td = p->p_thread;
492 KKASSERT(td == mycpu->gd_curthread);
494 crit_enter_quick(td);
497 * Check if the process exceeds its cpu resource allocation.
498 * If over max, kill it. Time spent in interrupts is not
499 * included. YYY 64 bit match is expensive. Ick.
501 ttime = td->td_sticks + td->td_uticks;
502 if (p->p_stat != SZOMB && p->p_limit->p_cpulimit != RLIM_INFINITY &&
503 ttime > p->p_limit->p_cpulimit) {
504 rlim = &p->p_rlimit[RLIMIT_CPU];
505 if (ttime / (rlim_t)1000000 >= rlim->rlim_max) {
506 killproc(p, "exceeded maximum CPU limit");
509 if (rlim->rlim_cur < rlim->rlim_max) {
510 /* XXX: we should make a private copy */
517 * If we are in a SSTOPped state we deschedule ourselves.
518 * YYY this needs to be cleaned up, remember that LWKTs stay on
519 * their run queue which works differently then the user scheduler
520 * which removes the process from the runq when it runs it.
522 mycpu->gd_cnt.v_swtch++;
523 if (p->p_stat == SSTOP)
524 lwkt_deschedule_self(td);
530 * Change process state to be runnable,
531 * placing it on the run queue if it is in memory,
532 * and awakening the swapper if it isn't in memory.
535 setrunnable(struct proc *p)
544 panic("setrunnable");
547 unsleep(p->p_thread); /* e.g. when sending signals */
556 * The process is controlled by LWKT at this point, we do not mess
557 * around with the userland scheduler until the thread tries to
558 * return to user mode. We do not clear p_slptime or call
561 if (p->p_flag & P_INMEM) {
562 lwkt_schedule(p->p_thread);
564 p->p_flag |= P_SWAPINREQ;
565 wakeup((caddr_t)&proc0);
571 * Yield / synchronous reschedule. This is a bit tricky because the trap
572 * code might have set a lazy release on the switch function. Setting
573 * P_PASSIVE_ACQ will ensure that the lazy release executes when we call
574 * switch, and that we are given a greater chance of affinity with our
577 * We call lwkt_setpri_self() to rotate our thread to the end of the lwkt
578 * run queue. lwkt_switch() will also execute any assigned passive release
579 * (which usually calls release_curproc()), allowing a same/higher priority
580 * process to be designated as the current process.
582 * While it is possible for a lower priority process to be designated,
583 * it's call to lwkt_maybe_switch() in acquire_curproc() will likely
584 * round-robin back to us and we will be able to re-acquire the current
585 * process designation.
590 struct thread *td = curthread;
591 struct proc *p = td->td_proc;
593 lwkt_setpri_self(td->td_pri & TDPRI_MASK);
595 p->p_flag |= P_PASSIVE_ACQ;
597 p->p_flag &= ~P_PASSIVE_ACQ;
604 * Change the process state to NOT be runnable, removing it from the run
608 clrrunnable(struct proc *p, int stat)
610 crit_enter_quick(p->p_thread);
611 if (p->p_stat == SRUN && (p->p_flag & P_ONRUNQ))
612 p->p_usched->remrunqueue(p);
614 crit_exit_quick(p->p_thread);
618 * Compute a tenex style load average of a quantity on
619 * 1, 5 and 15 minute intervals.
631 FOREACH_PROC_IN_SYSTEM(p) {
634 if ((td = p->p_thread) == NULL)
636 if (td->td_flags & TDF_BLOCKED)
646 for (i = 0; i < 3; i++)
647 avg->ldavg[i] = (cexp[i] * avg->ldavg[i] +
648 nrun * FSCALE * (FSCALE - cexp[i])) >> FSHIFT;
651 * Schedule the next update to occur after 5 seconds, but add a
652 * random variation to avoid synchronisation with processes that
653 * run at regular intervals.
655 callout_reset(&loadav_callout, hz * 4 + (int)(random() % (hz * 2 + 1)),
661 sched_setup(void *dummy)
663 callout_init(&loadav_callout);
664 callout_init(&schedcpu_callout);
666 /* Kick off timeout driven events by calling first time. */