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 1.55 2005/12/01 18:30:08 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>
57 #include <sys/ktrace.h>
59 #include <sys/xwait.h>
61 #include <machine/cpu.h>
62 #include <machine/ipl.h>
63 #include <machine/smp.h>
65 TAILQ_HEAD(tslpque, thread);
67 static void sched_setup (void *dummy);
68 SYSINIT(sched_setup, SI_SUB_KICK_SCHEDULER, SI_ORDER_FIRST, sched_setup, NULL)
73 int sched_quantum; /* Roundrobin scheduling quantum in ticks. */
75 int ncpus2, ncpus2_shift, ncpus2_mask;
78 static struct callout loadav_callout;
79 static struct callout schedcpu_callout;
80 MALLOC_DEFINE(M_TSLEEP, "tslpque", "tsleep queues");
82 struct loadavg averunnable =
83 { {0, 0, 0}, FSCALE }; /* load average, of runnable procs */
85 * Constants for averages over 1, 5, and 15 minutes
86 * when sampling at 5 second intervals.
88 static fixpt_t cexp[3] = {
89 0.9200444146293232 * FSCALE, /* exp(-1/12) */
90 0.9834714538216174 * FSCALE, /* exp(-1/60) */
91 0.9944598480048967 * FSCALE, /* exp(-1/180) */
94 static void endtsleep (void *);
95 static void unsleep_and_wakeup_thread(struct thread *td);
96 static void loadav (void *arg);
97 static void schedcpu (void *arg);
100 * Adjust the scheduler quantum. The quantum is specified in microseconds.
101 * Note that 'tick' is in microseconds per tick.
104 sysctl_kern_quantum(SYSCTL_HANDLER_ARGS)
108 new_val = sched_quantum * tick;
109 error = sysctl_handle_int(oidp, &new_val, 0, req);
110 if (error != 0 || req->newptr == NULL)
114 sched_quantum = new_val / tick;
115 hogticks = 2 * sched_quantum;
119 SYSCTL_PROC(_kern, OID_AUTO, quantum, CTLTYPE_INT|CTLFLAG_RW,
120 0, sizeof sched_quantum, sysctl_kern_quantum, "I", "");
123 * If `ccpu' is not equal to `exp(-1/20)' and you still want to use the
124 * faster/more-accurate formula, you'll have to estimate CCPU_SHIFT below
125 * and possibly adjust FSHIFT in "param.h" so that (FSHIFT >= CCPU_SHIFT).
127 * To estimate CCPU_SHIFT for exp(-1/20), the following formula was used:
128 * 1 - exp(-1/20) ~= 0.0487 ~= 0.0488 == 1 (fixed pt, *11* bits).
130 * If you don't want to bother with the faster/more-accurate formula, you
131 * can set CCPU_SHIFT to (FSHIFT + 1) which will use a slower/less-accurate
132 * (more general) method of calculating the %age of CPU used by a process.
134 * decay 95% of `p_pctcpu' in 60 seconds; see CCPU_SHIFT before changing
136 #define CCPU_SHIFT 11
138 static fixpt_t ccpu = 0.95122942450071400909 * FSCALE; /* exp(-1/20) */
139 SYSCTL_INT(_kern, OID_AUTO, ccpu, CTLFLAG_RD, &ccpu, 0, "");
142 * kernel uses `FSCALE', userland (SHOULD) use kern.fscale
144 static int fscale __unused = FSCALE;
145 SYSCTL_INT(_kern, OID_AUTO, fscale, CTLFLAG_RD, 0, FSCALE, "");
148 * Recompute process priorities, once a second.
150 * Since the userland schedulers are typically event oriented, if the
151 * estcpu calculation at wakeup() time is not sufficient to make a
152 * process runnable relative to other processes in the system we have
153 * a 1-second recalc to help out.
155 * This code also allows us to store sysclock_t data in the process structure
156 * without fear of an overrun, since sysclock_t are guarenteed to hold
157 * several seconds worth of count.
168 * General process statistics once a second
170 FOREACH_PROC_IN_SYSTEM(p) {
173 if (p->p_stat == SSLEEP)
177 * Only recalculate processes that are active or have slept
178 * less then 2 seconds. The schedulers understand this.
180 if (p->p_slptime <= 1) {
181 p->p_usched->recalculate(&p->p_lwp);
183 p->p_pctcpu = (p->p_pctcpu * ccpu) >> FSHIFT;
189 * Resource checks. XXX break out since psignal/killproc can block,
190 * limiting us to one process killed per second. There is probably
193 FOREACH_PROC_IN_SYSTEM(p) {
195 if (p->p_stat == SIDL ||
196 (p->p_flag & P_ZOMBIE) ||
197 p->p_limit == NULL ||
203 ttime = p->p_thread->td_sticks + p->p_thread->td_uticks;
204 if (p->p_limit->p_cpulimit != RLIM_INFINITY &&
205 ttime > p->p_limit->p_cpulimit
207 rlim = &p->p_rlimit[RLIMIT_CPU];
208 if (ttime / (rlim_t)1000000 >= rlim->rlim_max) {
209 killproc(p, "exceeded maximum CPU limit");
212 if (rlim->rlim_cur < rlim->rlim_max) {
213 /* XXX: we should make a private copy */
223 wakeup((caddr_t)&lbolt);
224 wakeup((caddr_t)&lbolt_syncer);
225 callout_reset(&schedcpu_callout, hz, schedcpu, NULL);
229 * This is only used by ps. Generate a cpu percentage use over
230 * a period of one second.
233 updatepcpu(struct lwp *lp, int cpticks, int ttlticks)
238 acc = (cpticks << FSHIFT) / ttlticks;
239 if (ttlticks >= ESTCPUFREQ) {
240 lp->lwp_pctcpu = acc;
242 remticks = ESTCPUFREQ - ttlticks;
243 lp->lwp_pctcpu = (acc * ttlticks + lp->lwp_pctcpu * remticks) /
249 * We're only looking at 7 bits of the address; everything is
250 * aligned to 4, lots of things are aligned to greater powers
251 * of 2. Shift right by 8, i.e. drop the bottom 256 worth.
253 #define TABLESIZE 128
254 #define LOOKUP(x) (((intptr_t)(x) >> 8) & (TABLESIZE - 1))
256 static cpumask_t slpque_cpumasks[TABLESIZE];
259 * General scheduler initialization. We force a reschedule 25 times
260 * a second by default. Note that cpu0 is initialized in early boot and
261 * cannot make any high level calls.
263 * Each cpu has its own sleep queue.
266 sleep_gdinit(globaldata_t gd)
268 static struct tslpque slpque_cpu0[TABLESIZE];
271 if (gd->gd_cpuid == 0) {
272 sched_quantum = (hz + 24) / 25;
273 hogticks = 2 * sched_quantum;
275 gd->gd_tsleep_hash = slpque_cpu0;
277 gd->gd_tsleep_hash = malloc(sizeof(slpque_cpu0),
278 M_TSLEEP, M_WAITOK | M_ZERO);
280 for (i = 0; i < TABLESIZE; ++i)
281 TAILQ_INIT(&gd->gd_tsleep_hash[i]);
285 * General sleep call. Suspends the current process until a wakeup is
286 * performed on the specified identifier. The process will then be made
287 * runnable with the specified priority. Sleeps at most timo/hz seconds
288 * (0 means no timeout). If flags includes PCATCH flag, signals are checked
289 * before and after sleeping, else signals are not checked. Returns 0 if
290 * awakened, EWOULDBLOCK if the timeout expires. If PCATCH is set and a
291 * signal needs to be delivered, ERESTART is returned if the current system
292 * call should be restarted if possible, and EINTR is returned if the system
293 * call should be interrupted by the signal (return EINTR).
295 * Note that if we are a process, we release_curproc() before messing with
296 * the LWKT scheduler.
298 * During autoconfiguration or after a panic, a sleep will simply
299 * lower the priority briefly to allow interrupts, then return.
302 tsleep(void *ident, int flags, const char *wmesg, int timo)
304 struct thread *td = curthread;
305 struct proc *p = td->td_proc; /* may be NULL */
312 struct callout thandle;
315 * NOTE: removed KTRPOINT, it could cause races due to blocking
316 * even in stable. Just scrap it for now.
318 if (cold || panicstr) {
320 * After a panic, or during autoconfiguration,
321 * just give interrupts a chance, then just return;
322 * don't run any other procs or panic below,
323 * in case this is the idle process and already asleep.
326 oldpri = td->td_pri & TDPRI_MASK;
327 lwkt_setpri_self(safepri);
329 lwkt_setpri_self(oldpri);
333 KKASSERT(td != &gd->gd_idlethread); /* you must be kidding! */
336 * NOTE: all of this occurs on the current cpu, including any
337 * callout-based wakeups, so a critical section is a sufficient
340 * The entire sequence through to where we actually sleep must
341 * run without breaking the critical section.
344 catch = flags & PCATCH;
348 crit_enter_quick(td);
350 KASSERT(ident != NULL, ("tsleep: no ident"));
351 KASSERT(p == NULL || p->p_stat == SRUN, ("tsleep %p %s %d",
352 ident, wmesg, p->p_stat));
355 * Setup for the current process (if this is a process).
360 * Early termination if PCATCH was set and a
361 * signal is pending, interlocked with the
364 * Early termination only occurs when tsleep() is
365 * entered while in a normal SRUN state.
367 if ((sig = CURSIG(p)) != 0)
371 * Causes psignal to wake us up when.
373 p->p_flag |= P_SINTR;
377 * Make sure the current process has been untangled from
378 * the userland scheduler and initialize slptime to start
381 if (flags & PNORESCHED)
382 td->td_flags |= TDF_NORESCHED;
383 p->p_usched->release_curproc(&p->p_lwp);
388 * Move our thread to the correct queue and setup our wchan, etc.
390 lwkt_deschedule_self(td);
391 td->td_flags |= TDF_TSLEEPQ;
392 TAILQ_INSERT_TAIL(&gd->gd_tsleep_hash[id], td, td_threadq);
393 atomic_set_int(&slpque_cpumasks[id], gd->gd_cpumask);
395 td->td_wchan = ident;
396 td->td_wmesg = wmesg;
397 td->td_wdomain = flags & PDOMAIN_MASK;
400 * Setup the timeout, if any
403 callout_init(&thandle);
404 callout_reset(&thandle, timo, endtsleep, td);
412 * Ok, we are sleeping. Remove us from the userland runq
413 * and place us in the SSLEEP state.
415 if (p->p_flag & P_ONRUNQ)
416 p->p_usched->remrunqueue(&p->p_lwp);
418 p->p_stats->p_ru.ru_nvcsw++;
426 * Make sure we haven't switched cpus while we were asleep. It's
427 * not supposed to happen. Cleanup our temporary flags.
429 KKASSERT(gd == td->td_gd);
430 td->td_flags &= ~TDF_NORESCHED;
433 * Cleanup the timeout.
436 if (td->td_flags & TDF_TIMEOUT) {
437 td->td_flags &= ~TDF_TIMEOUT;
441 callout_stop(&thandle);
446 * Since td_threadq is used both for our run queue AND for the
447 * tsleep hash queue, we can't still be on it at this point because
448 * we've gotten cpu back.
450 KKASSERT((td->td_flags & TDF_TSLEEPQ) == 0);
456 * Figure out the correct error return
460 p->p_flag &= ~(P_BREAKTSLEEP | P_SINTR);
461 if (catch && error == 0 && (sig != 0 || (sig = CURSIG(p)))) {
462 if (SIGISMEMBER(p->p_sigacts->ps_sigintr, sig))
473 * This is a dandy function that allows us to interlock tsleep/wakeup
474 * operations with unspecified upper level locks, such as lockmgr locks,
475 * simply by holding a critical section. The sequence is:
477 * (enter critical section)
478 * (acquire upper level lock)
479 * tsleep_interlock(blah)
480 * (release upper level lock)
482 * (exit critical section)
484 * Basically this function sets our cpumask for the ident which informs
485 * other cpus that our cpu 'might' be waiting (or about to wait on) the
486 * hash index related to the ident. The critical section prevents another
487 * cpu's wakeup() from being processed on our cpu until we are actually
488 * able to enter the tsleep(). Thus, no race occurs between our attempt
489 * to release a resource and sleep, and another cpu's attempt to acquire
490 * a resource and call wakeup.
492 * There isn't much of a point to this function unless you call it while
493 * holding a critical section.
496 tsleep_interlock(void *ident)
498 int id = LOOKUP(ident);
500 atomic_set_int(&slpque_cpumasks[id], mycpu->gd_cpumask);
504 * Implement the timeout for tsleep.
506 * We set P_BREAKTSLEEP to indicate that an event has occured, but
507 * we only call setrunnable if the process is not stopped.
509 * This type of callout timeout is scheduled on the same cpu the process
510 * is sleeping on. Also, at the moment, the MP lock is held.
518 ASSERT_MP_LOCK_HELD(curthread);
522 * cpu interlock. Thread flags are only manipulated on
523 * the cpu owning the thread. proc flags are only manipulated
524 * by the older of the MP lock. We have both.
526 if (td->td_flags & TDF_TSLEEPQ) {
527 td->td_flags |= TDF_TIMEOUT;
529 if ((p = td->td_proc) != NULL) {
530 p->p_flag |= P_BREAKTSLEEP;
531 if ((p->p_flag & P_STOPPED) == 0)
534 unsleep_and_wakeup_thread(td);
541 * Unsleep and wakeup a thread. This function runs without the MP lock
542 * which means that it can only manipulate thread state on the owning cpu,
543 * and cannot touch the process state at all.
547 unsleep_and_wakeup_thread(struct thread *td)
549 globaldata_t gd = mycpu;
553 if (td->td_gd != gd) {
554 lwkt_send_ipiq(td->td_gd, (ipifunc1_t)unsleep_and_wakeup_thread, td);
559 if (td->td_flags & TDF_TSLEEPQ) {
560 td->td_flags &= ~TDF_TSLEEPQ;
561 id = LOOKUP(td->td_wchan);
562 TAILQ_REMOVE(&gd->gd_tsleep_hash[id], td, td_threadq);
563 if (TAILQ_FIRST(&gd->gd_tsleep_hash[id]) == NULL)
564 atomic_clear_int(&slpque_cpumasks[id], gd->gd_cpumask);
571 * Make all processes sleeping on the specified identifier runnable.
572 * count may be zero or one only.
574 * The domain encodes the sleep/wakeup domain AND the first cpu to check
575 * (which is always the current cpu). As we iterate across cpus
577 * This call may run without the MP lock held. We can only manipulate thread
578 * state on the cpu owning the thread. We CANNOT manipulate process state
582 _wakeup(void *ident, int domain)
599 qp = &gd->gd_tsleep_hash[id];
601 for (td = TAILQ_FIRST(qp); td != NULL; td = ntd) {
602 ntd = TAILQ_NEXT(td, td_threadq);
603 if (td->td_wchan == ident &&
604 td->td_wdomain == (domain & PDOMAIN_MASK)
606 KKASSERT(td->td_flags & TDF_TSLEEPQ);
607 td->td_flags &= ~TDF_TSLEEPQ;
608 TAILQ_REMOVE(qp, td, td_threadq);
609 if (TAILQ_FIRST(qp) == NULL) {
610 atomic_clear_int(&slpque_cpumasks[id],
614 if (domain & PWAKEUP_ONE)
622 * We finished checking the current cpu but there still may be
623 * more work to do. Either wakeup_one was requested and no matching
624 * thread was found, or a normal wakeup was requested and we have
625 * to continue checking cpus.
627 * The cpu that started the wakeup sequence is encoded in the domain.
628 * We use this information to determine which cpus still need to be
629 * checked, locate a candidate cpu, and chain the wakeup
630 * asynchronously with an IPI message.
632 * It should be noted that this scheme is actually less expensive then
633 * the old scheme when waking up multiple threads, since we send
634 * only one IPI message per target candidate which may then schedule
635 * multiple threads. Before we could have wound up sending an IPI
636 * message for each thread on the target cpu (!= current cpu) that
637 * needed to be woken up.
639 * NOTE: Wakeups occuring on remote cpus are asynchronous. This
640 * should be ok since we are passing idents in the IPI rather then
643 if ((mask = slpque_cpumasks[id]) != 0) {
645 * Look for a cpu that might have work to do. Mask out cpus
646 * which have already been processed.
648 * 31xxxxxxxxxxxxxxxxxxxxxxxxxxxxx0
650 * start currentcpu start
653 * 11111111111111110000000000000111 case1
654 * 00000000111111110000000000000000 case2
656 * case1: We started at start_case1 and processed through
657 * to the current cpu. We have to check any bits
658 * after the current cpu, then check bits before
661 * case2: We have already checked all the bits from
662 * start_case2 to the end, and from 0 to the current
663 * cpu. We just have the bits from the current cpu
664 * to start_case2 left to check.
666 startcpu = PWAKEUP_DECODE(domain);
667 if (gd->gd_cpuid >= startcpu) {
671 tmask = mask & ~((gd->gd_cpumask << 1) - 1);
673 nextcpu = bsfl(mask & tmask);
674 lwkt_send_ipiq2(globaldata_find(nextcpu),
675 _wakeup, ident, domain);
677 tmask = (1 << startcpu) - 1;
679 nextcpu = bsfl(mask & tmask);
681 globaldata_find(nextcpu),
682 _wakeup, ident, domain);
689 tmask = ~((gd->gd_cpumask << 1) - 1) &
690 ((1 << startcpu) - 1);
692 nextcpu = bsfl(mask & tmask);
693 lwkt_send_ipiq2(globaldata_find(nextcpu),
694 _wakeup, ident, domain);
706 _wakeup(ident, PWAKEUP_ENCODE(0, mycpu->gd_cpuid));
710 wakeup_one(void *ident)
712 /* XXX potentially round-robin the first responding cpu */
713 _wakeup(ident, PWAKEUP_ENCODE(0, mycpu->gd_cpuid) | PWAKEUP_ONE);
717 wakeup_domain(void *ident, int domain)
719 _wakeup(ident, PWAKEUP_ENCODE(domain, mycpu->gd_cpuid));
723 wakeup_domain_one(void *ident, int domain)
725 /* XXX potentially round-robin the first responding cpu */
726 _wakeup(ident, PWAKEUP_ENCODE(domain, mycpu->gd_cpuid) | PWAKEUP_ONE);
732 * Make a process runnable. The MP lock must be held on call. This only
733 * has an effect if we are in SSLEEP. We only break out of the
734 * tsleep if P_BREAKTSLEEP is set, otherwise we just fix-up the state.
736 * NOTE: With the MP lock held we can only safely manipulate the process
737 * structure. We cannot safely manipulate the thread structure.
740 setrunnable(struct proc *p)
743 ASSERT_MP_LOCK_HELD(curthread);
744 p->p_flag &= ~P_STOPPED;
745 if (p->p_stat == SSLEEP && (p->p_flag & P_BREAKTSLEEP)) {
746 unsleep_and_wakeup_thread(p->p_thread);
752 * The process is stopped due to some condition, usually because P_STOPPED
753 * is set but also possibly due to being traced.
755 * NOTE! If the caller sets P_STOPPED, the caller must also clear P_WAITED
756 * because the parent may check the child's status before the child actually
757 * gets to this routine.
759 * This routine is called with the current process only, typically just
760 * before returning to userland.
762 * Setting P_BREAKTSLEEP before entering the tsleep will cause a passive
763 * SIGCONT to break out of the tsleep.
766 tstop(struct proc *p)
768 wakeup((caddr_t)p->p_pptr);
769 p->p_flag |= P_BREAKTSLEEP;
770 tsleep(p, 0, "stop", 0);
774 * Yield / synchronous reschedule. This is a bit tricky because the trap
775 * code might have set a lazy release on the switch function. Setting
776 * P_PASSIVE_ACQ will ensure that the lazy release executes when we call
777 * switch, and that we are given a greater chance of affinity with our
780 * We call lwkt_setpri_self() to rotate our thread to the end of the lwkt
781 * run queue. lwkt_switch() will also execute any assigned passive release
782 * (which usually calls release_curproc()), allowing a same/higher priority
783 * process to be designated as the current process.
785 * While it is possible for a lower priority process to be designated,
786 * it's call to lwkt_maybe_switch() in acquire_curproc() will likely
787 * round-robin back to us and we will be able to re-acquire the current
788 * process designation.
793 struct thread *td = curthread;
794 struct proc *p = td->td_proc;
796 lwkt_setpri_self(td->td_pri & TDPRI_MASK);
798 p->p_flag |= P_PASSIVE_ACQ;
800 p->p_flag &= ~P_PASSIVE_ACQ;
807 * Compute a tenex style load average of a quantity on
808 * 1, 5 and 15 minute intervals.
820 FOREACH_PROC_IN_SYSTEM(p) {
823 if ((td = p->p_thread) == NULL)
825 if (td->td_flags & TDF_BLOCKED)
835 for (i = 0; i < 3; i++)
836 avg->ldavg[i] = (cexp[i] * avg->ldavg[i] +
837 nrun * FSCALE * (FSCALE - cexp[i])) >> FSHIFT;
840 * Schedule the next update to occur after 5 seconds, but add a
841 * random variation to avoid synchronisation with processes that
842 * run at regular intervals.
844 callout_reset(&loadav_callout, hz * 4 + (int)(random() % (hz * 2 + 1)),
850 sched_setup(void *dummy)
852 callout_init(&loadav_callout);
853 callout_init(&schedcpu_callout);
855 /* Kick off timeout driven events by calling first time. */