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. Neither the name of the University nor the names of its contributors
19 * may be used to endorse or promote products derived from this software
20 * without specific prior written permission.
22 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
23 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
24 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
25 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
26 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
27 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
28 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
29 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
30 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
31 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
34 * @(#)kern_synch.c 8.9 (Berkeley) 5/19/95
35 * $FreeBSD: src/sys/kern/kern_synch.c,v 1.87.2.6 2002/10/13 07:29:53 kbyanc Exp $
38 #include "opt_ktrace.h"
40 #include <sys/param.h>
41 #include <sys/systm.h>
43 #include <sys/kernel.h>
44 #include <sys/signalvar.h>
45 #include <sys/resourcevar.h>
46 #include <sys/vmmeter.h>
47 #include <sys/sysctl.h>
51 #include <sys/ktrace.h>
54 #include <sys/serialize.h>
56 #include <sys/signal2.h>
57 #include <sys/thread2.h>
58 #include <sys/spinlock2.h>
59 #include <sys/mutex2.h>
61 #include <machine/cpu.h>
62 #include <machine/smp.h>
64 TAILQ_HEAD(tslpque, thread);
66 static void sched_setup (void *dummy);
67 SYSINIT(sched_setup, SI_SUB_KICK_SCHEDULER, SI_ORDER_FIRST, sched_setup, NULL)
71 int sched_quantum; /* Roundrobin scheduling quantum in ticks. */
73 int ncpus2, ncpus2_shift, ncpus2_mask; /* note: mask not cpumask_t */
74 int ncpus_fit, ncpus_fit_mask; /* note: mask not cpumask_t */
77 int tsleep_crypto_dump = 0;
79 static struct callout loadav_callout;
80 static struct callout schedcpu_callout;
81 MALLOC_DEFINE(M_TSLEEP, "tslpque", "tsleep queues");
83 #define __DEALL(ident) __DEQUALIFY(void *, ident)
85 #if !defined(KTR_TSLEEP)
86 #define KTR_TSLEEP KTR_ALL
88 KTR_INFO_MASTER(tsleep);
89 KTR_INFO(KTR_TSLEEP, tsleep, tsleep_beg, 0, "tsleep enter %p", const volatile void *ident);
90 KTR_INFO(KTR_TSLEEP, tsleep, tsleep_end, 1, "tsleep exit");
91 KTR_INFO(KTR_TSLEEP, tsleep, wakeup_beg, 2, "wakeup enter %p", const volatile void *ident);
92 KTR_INFO(KTR_TSLEEP, tsleep, wakeup_end, 3, "wakeup exit");
93 KTR_INFO(KTR_TSLEEP, tsleep, ilockfail, 4, "interlock failed %p", const volatile void *ident);
95 #define logtsleep1(name) KTR_LOG(tsleep_ ## name)
96 #define logtsleep2(name, val) KTR_LOG(tsleep_ ## name, val)
98 struct loadavg averunnable =
99 { {0, 0, 0}, FSCALE }; /* load average, of runnable procs */
101 * Constants for averages over 1, 5, and 15 minutes
102 * when sampling at 5 second intervals.
104 static fixpt_t cexp[3] = {
105 0.9200444146293232 * FSCALE, /* exp(-1/12) */
106 0.9834714538216174 * FSCALE, /* exp(-1/60) */
107 0.9944598480048967 * FSCALE, /* exp(-1/180) */
110 static void endtsleep (void *);
111 static void loadav (void *arg);
112 static void schedcpu (void *arg);
115 * Adjust the scheduler quantum. The quantum is specified in microseconds.
116 * Note that 'tick' is in microseconds per tick.
119 sysctl_kern_quantum(SYSCTL_HANDLER_ARGS)
123 new_val = sched_quantum * ustick;
124 error = sysctl_handle_int(oidp, &new_val, 0, req);
125 if (error != 0 || req->newptr == NULL)
127 if (new_val < ustick)
129 sched_quantum = new_val / ustick;
133 SYSCTL_PROC(_kern, OID_AUTO, quantum, CTLTYPE_INT|CTLFLAG_RW,
134 0, sizeof sched_quantum, sysctl_kern_quantum, "I", "");
136 static int pctcpu_decay = 10;
137 SYSCTL_INT(_kern, OID_AUTO, pctcpu_decay, CTLFLAG_RW, &pctcpu_decay, 0, "");
140 * kernel uses `FSCALE', userland (SHOULD) use kern.fscale
142 int fscale __unused = FSCALE; /* exported to systat */
143 SYSCTL_INT(_kern, OID_AUTO, fscale, CTLFLAG_RD, 0, FSCALE, "");
146 * Recompute process priorities, once a second.
148 * Since the userland schedulers are typically event oriented, if the
149 * estcpu calculation at wakeup() time is not sufficient to make a
150 * process runnable relative to other processes in the system we have
151 * a 1-second recalc to help out.
153 * This code also allows us to store sysclock_t data in the process structure
154 * without fear of an overrun, since sysclock_t are guarenteed to hold
155 * several seconds worth of count.
157 * WARNING! callouts can preempt normal threads. However, they will not
158 * preempt a thread holding a spinlock so we *can* safely use spinlocks.
160 static int schedcpu_stats(struct proc *p, void *data __unused);
161 static int schedcpu_resource(struct proc *p, void *data __unused);
166 allproc_scan(schedcpu_stats, NULL);
167 allproc_scan(schedcpu_resource, NULL);
168 wakeup((caddr_t)&lbolt);
169 wakeup(lbolt_syncer);
170 callout_reset(&schedcpu_callout, hz, schedcpu, NULL);
174 * General process statistics once a second
177 schedcpu_stats(struct proc *p, void *data __unused)
182 * Threads may not be completely set up if process in SIDL state.
184 if (p->p_stat == SIDL)
188 if (lwkt_trytoken(&p->p_token) == FALSE) {
194 FOREACH_LWP_IN_PROC(lp, p) {
195 if (lp->lwp_stat == LSSLEEP) {
197 if (lp->lwp_slptime == 1)
198 p->p_usched->uload_update(lp);
202 * Only recalculate processes that are active or have slept
203 * less then 2 seconds. The schedulers understand this.
204 * Otherwise decay by 50% per second.
206 if (lp->lwp_slptime <= 1) {
207 p->p_usched->recalculate(lp);
211 decay = pctcpu_decay;
217 lp->lwp_pctcpu = (lp->lwp_pctcpu * (decay - 1)) / decay;
220 lwkt_reltoken(&p->p_token);
227 * Resource checks. XXX break out since ksignal/killproc can block,
228 * limiting us to one process killed per second. There is probably
232 schedcpu_resource(struct proc *p, void *data __unused)
237 if (p->p_stat == SIDL)
241 if (lwkt_trytoken(&p->p_token) == FALSE) {
246 if (p->p_stat == SZOMB || p->p_limit == NULL) {
247 lwkt_reltoken(&p->p_token);
253 FOREACH_LWP_IN_PROC(lp, p) {
255 * We may have caught an lp in the middle of being
256 * created, lwp_thread can be NULL.
258 if (lp->lwp_thread) {
259 ttime += lp->lwp_thread->td_sticks;
260 ttime += lp->lwp_thread->td_uticks;
264 switch(plimit_testcpulimit(p->p_limit, ttime)) {
265 case PLIMIT_TESTCPU_KILL:
266 killproc(p, "exceeded maximum CPU limit");
268 case PLIMIT_TESTCPU_XCPU:
269 if ((p->p_flags & P_XCPU) == 0) {
270 p->p_flags |= P_XCPU;
277 lwkt_reltoken(&p->p_token);
284 * This is only used by ps. Generate a cpu percentage use over
285 * a period of one second.
288 updatepcpu(struct lwp *lp, int cpticks, int ttlticks)
293 acc = (cpticks << FSHIFT) / ttlticks;
294 if (ttlticks >= ESTCPUFREQ) {
295 lp->lwp_pctcpu = acc;
297 remticks = ESTCPUFREQ - ttlticks;
298 lp->lwp_pctcpu = (acc * ttlticks + lp->lwp_pctcpu * remticks) /
304 * tsleep/wakeup hash table parameters. Try to find the sweet spot for
305 * like addresses being slept on.
307 #define TABLESIZE 4001
308 #define LOOKUP(x) (((u_int)(uintptr_t)(x)) % TABLESIZE)
310 static cpumask_t slpque_cpumasks[TABLESIZE];
313 * General scheduler initialization. We force a reschedule 25 times
314 * a second by default. Note that cpu0 is initialized in early boot and
315 * cannot make any high level calls.
317 * Each cpu has its own sleep queue.
320 sleep_gdinit(globaldata_t gd)
322 static struct tslpque slpque_cpu0[TABLESIZE];
325 if (gd->gd_cpuid == 0) {
326 sched_quantum = (hz + 24) / 25;
327 gd->gd_tsleep_hash = slpque_cpu0;
329 gd->gd_tsleep_hash = kmalloc(sizeof(slpque_cpu0),
330 M_TSLEEP, M_WAITOK | M_ZERO);
332 for (i = 0; i < TABLESIZE; ++i)
333 TAILQ_INIT(&gd->gd_tsleep_hash[i]);
337 * This is a dandy function that allows us to interlock tsleep/wakeup
338 * operations with unspecified upper level locks, such as lockmgr locks,
339 * simply by holding a critical section. The sequence is:
341 * (acquire upper level lock)
342 * tsleep_interlock(blah)
343 * (release upper level lock)
346 * Basically this functions queues us on the tsleep queue without actually
347 * descheduling us. When tsleep() is later called with PINTERLOCK it
348 * assumes the thread was already queued, otherwise it queues it there.
350 * Thus it is possible to receive the wakeup prior to going to sleep and
351 * the race conditions are covered.
354 _tsleep_interlock(globaldata_t gd, const volatile void *ident, int flags)
356 thread_t td = gd->gd_curthread;
359 crit_enter_quick(td);
360 if (td->td_flags & TDF_TSLEEPQ) {
361 id = LOOKUP(td->td_wchan);
362 TAILQ_REMOVE(&gd->gd_tsleep_hash[id], td, td_sleepq);
363 if (TAILQ_FIRST(&gd->gd_tsleep_hash[id]) == NULL) {
364 ATOMIC_CPUMASK_NANDBIT(slpque_cpumasks[id],
368 td->td_flags |= TDF_TSLEEPQ;
371 TAILQ_INSERT_TAIL(&gd->gd_tsleep_hash[id], td, td_sleepq);
372 ATOMIC_CPUMASK_ORBIT(slpque_cpumasks[id], gd->gd_cpuid);
373 td->td_wchan = ident;
374 td->td_wdomain = flags & PDOMAIN_MASK;
379 tsleep_interlock(const volatile void *ident, int flags)
381 _tsleep_interlock(mycpu, ident, flags);
385 * Remove thread from sleepq. Must be called with a critical section held.
386 * The thread must not be migrating.
389 _tsleep_remove(thread_t td)
391 globaldata_t gd = mycpu;
394 KKASSERT(td->td_gd == gd && IN_CRITICAL_SECT(td));
395 KKASSERT((td->td_flags & TDF_MIGRATING) == 0);
396 if (td->td_flags & TDF_TSLEEPQ) {
397 td->td_flags &= ~TDF_TSLEEPQ;
398 id = LOOKUP(td->td_wchan);
399 TAILQ_REMOVE(&gd->gd_tsleep_hash[id], td, td_sleepq);
400 if (TAILQ_FIRST(&gd->gd_tsleep_hash[id]) == NULL) {
401 ATOMIC_CPUMASK_NANDBIT(slpque_cpumasks[id],
410 tsleep_remove(thread_t td)
416 * General sleep call. Suspends the current process until a wakeup is
417 * performed on the specified identifier. The process will then be made
418 * runnable with the specified priority. Sleeps at most timo/hz seconds
419 * (0 means no timeout). If flags includes PCATCH flag, signals are checked
420 * before and after sleeping, else signals are not checked. Returns 0 if
421 * awakened, EWOULDBLOCK if the timeout expires. If PCATCH is set and a
422 * signal needs to be delivered, ERESTART is returned if the current system
423 * call should be restarted if possible, and EINTR is returned if the system
424 * call should be interrupted by the signal (return EINTR).
426 * Note that if we are a process, we release_curproc() before messing with
427 * the LWKT scheduler.
429 * During autoconfiguration or after a panic, a sleep will simply
430 * lower the priority briefly to allow interrupts, then return.
432 * WARNING! This code can't block (short of switching away), or bad things
433 * will happen. No getting tokens, no blocking locks, etc.
436 tsleep(const volatile void *ident, int flags, const char *wmesg, int timo)
438 struct thread *td = curthread;
439 struct lwp *lp = td->td_lwp;
440 struct proc *p = td->td_proc; /* may be NULL */
446 struct callout thandle;
449 * Currently a severe hack. Make sure any delayed wakeups
450 * are flushed before we sleep or we might deadlock on whatever
451 * event we are sleeping on.
453 if (td->td_flags & TDF_DELAYED_WAKEUP)
454 wakeup_end_delayed();
457 * NOTE: removed KTRPOINT, it could cause races due to blocking
458 * even in stable. Just scrap it for now.
460 if (!tsleep_crypto_dump && (tsleep_now_works == 0 || panicstr)) {
462 * After a panic, or before we actually have an operational
463 * softclock, just give interrupts a chance, then just return;
465 * don't run any other procs or panic below,
466 * in case this is the idle process and already asleep.
470 lwkt_setpri_self(safepri);
472 lwkt_setpri_self(oldpri);
475 logtsleep2(tsleep_beg, ident);
477 KKASSERT(td != &gd->gd_idlethread); /* you must be kidding! */
478 td->td_wakefromcpu = -1; /* overwritten by _wakeup */
481 * NOTE: all of this occurs on the current cpu, including any
482 * callout-based wakeups, so a critical section is a sufficient
485 * The entire sequence through to where we actually sleep must
486 * run without breaking the critical section.
488 catch = flags & PCATCH;
492 crit_enter_quick(td);
494 KASSERT(ident != NULL, ("tsleep: no ident"));
495 KASSERT(lp == NULL ||
496 lp->lwp_stat == LSRUN || /* Obvious */
497 lp->lwp_stat == LSSTOP, /* Set in tstop */
499 ident, wmesg, lp->lwp_stat));
502 * We interlock the sleep queue if the caller has not already done
503 * it for us. This must be done before we potentially acquire any
504 * tokens or we can loose the wakeup.
506 if ((flags & PINTERLOCKED) == 0) {
507 _tsleep_interlock(gd, ident, flags);
511 * Setup for the current process (if this is a process). We must
512 * interlock with lwp_token to avoid remote wakeup races via
516 lwkt_gettoken(&lp->lwp_token);
519 * Early termination if PCATCH was set and a
520 * signal is pending, interlocked with the
523 * Early termination only occurs when tsleep() is
524 * entered while in a normal LSRUN state.
526 if ((sig = CURSIG(lp)) != 0)
530 * Causes ksignal to wake us up if a signal is
531 * received (interlocked with p->p_token).
533 lp->lwp_flags |= LWP_SINTR;
540 * Make sure the current process has been untangled from
541 * the userland scheduler and initialize slptime to start
544 * NOTE: td->td_wakefromcpu is pre-set by the release function
545 * for the dfly scheduler, and then adjusted by _wakeup()
548 p->p_usched->release_curproc(lp);
553 * If the interlocked flag is set but our cpu bit in the slpqueue
554 * is no longer set, then a wakeup was processed inbetween the
555 * tsleep_interlock() (ours or the callers), and here. This can
556 * occur under numerous circumstances including when we release the
559 * Extreme loads can cause the sending of an IPI (e.g. wakeup()'s)
560 * to process incoming IPIs, thus draining incoming wakeups.
562 if ((td->td_flags & TDF_TSLEEPQ) == 0) {
563 logtsleep2(ilockfail, ident);
568 * scheduling is blocked while in a critical section. Coincide
569 * the descheduled-by-tsleep flag with the descheduling of the
572 * The timer callout is localized on our cpu and interlocked by
573 * our critical section.
575 lwkt_deschedule_self(td);
576 td->td_flags |= TDF_TSLEEP_DESCHEDULED;
577 td->td_wmesg = wmesg;
580 * Setup the timeout, if any. The timeout is only operable while
581 * the thread is flagged descheduled.
583 KKASSERT((td->td_flags & TDF_TIMEOUT) == 0);
585 callout_init_mp(&thandle);
586 callout_reset(&thandle, timo, endtsleep, td);
594 * Ok, we are sleeping. Place us in the SSLEEP state.
596 KKASSERT((lp->lwp_mpflags & LWP_MP_ONRUNQ) == 0);
599 * tstop() sets LSSTOP, so don't fiddle with that.
601 if (lp->lwp_stat != LSSTOP)
602 lp->lwp_stat = LSSLEEP;
603 lp->lwp_ru.ru_nvcsw++;
604 p->p_usched->uload_update(lp);
608 * And when we are woken up, put us back in LSRUN. If we
609 * slept for over a second, recalculate our estcpu.
611 lp->lwp_stat = LSRUN;
612 if (lp->lwp_slptime) {
613 p->p_usched->uload_update(lp);
614 p->p_usched->recalculate(lp);
622 * Make sure we haven't switched cpus while we were asleep. It's
623 * not supposed to happen. Cleanup our temporary flags.
625 KKASSERT(gd == td->td_gd);
628 * Cleanup the timeout. If the timeout has already occured thandle
629 * has already been stopped, otherwise stop thandle. If the timeout
630 * is running (the callout thread must be blocked trying to get
631 * lwp_token) then wait for us to get scheduled.
634 while (td->td_flags & TDF_TIMEOUT_RUNNING) {
635 lwkt_deschedule_self(td);
636 td->td_wmesg = "tsrace";
638 kprintf("td %p %s: timeout race\n", td, td->td_comm);
640 if (td->td_flags & TDF_TIMEOUT) {
641 td->td_flags &= ~TDF_TIMEOUT;
644 /* does not block when on same cpu */
645 callout_stop(&thandle);
648 td->td_flags &= ~TDF_TSLEEP_DESCHEDULED;
651 * Make sure we have been removed from the sleepq. In most
652 * cases this will have been done for us already but it is
653 * possible for a scheduling IPI to be in-flight from a
654 * previous tsleep/tsleep_interlock() or due to a straight-out
655 * call to lwkt_schedule() (in the case of an interrupt thread),
656 * causing a spurious wakeup.
662 * Figure out the correct error return. If interrupted by a
663 * signal we want to return EINTR or ERESTART.
667 if (catch && error == 0) {
668 if (sig != 0 || (sig = CURSIG(lp))) {
669 if (SIGISMEMBER(p->p_sigacts->ps_sigintr, sig))
675 lp->lwp_flags &= ~LWP_SINTR;
676 lwkt_reltoken(&lp->lwp_token);
678 logtsleep1(tsleep_end);
684 * Interlocked spinlock sleep. An exclusively held spinlock must
685 * be passed to ssleep(). The function will atomically release the
686 * spinlock and tsleep on the ident, then reacquire the spinlock and
689 * This routine is fairly important along the critical path, so optimize it
693 ssleep(const volatile void *ident, struct spinlock *spin, int flags,
694 const char *wmesg, int timo)
696 globaldata_t gd = mycpu;
699 _tsleep_interlock(gd, ident, flags);
700 spin_unlock_quick(gd, spin);
701 error = tsleep(ident, flags | PINTERLOCKED, wmesg, timo);
702 _spin_lock_quick(gd, spin, wmesg);
708 lksleep(const volatile void *ident, struct lock *lock, int flags,
709 const char *wmesg, int timo)
711 globaldata_t gd = mycpu;
714 _tsleep_interlock(gd, ident, flags);
715 lockmgr(lock, LK_RELEASE);
716 error = tsleep(ident, flags | PINTERLOCKED, wmesg, timo);
717 lockmgr(lock, LK_EXCLUSIVE);
723 * Interlocked mutex sleep. An exclusively held mutex must be passed
724 * to mtxsleep(). The function will atomically release the mutex
725 * and tsleep on the ident, then reacquire the mutex and return.
728 mtxsleep(const volatile void *ident, struct mtx *mtx, int flags,
729 const char *wmesg, int timo)
731 globaldata_t gd = mycpu;
734 _tsleep_interlock(gd, ident, flags);
736 error = tsleep(ident, flags | PINTERLOCKED, wmesg, timo);
737 mtx_lock_ex_quick(mtx, wmesg);
743 * Interlocked serializer sleep. An exclusively held serializer must
744 * be passed to zsleep(). The function will atomically release
745 * the serializer and tsleep on the ident, then reacquire the serializer
749 zsleep(const volatile void *ident, struct lwkt_serialize *slz, int flags,
750 const char *wmesg, int timo)
752 globaldata_t gd = mycpu;
755 ASSERT_SERIALIZED(slz);
757 _tsleep_interlock(gd, ident, flags);
758 lwkt_serialize_exit(slz);
759 ret = tsleep(ident, flags | PINTERLOCKED, wmesg, timo);
760 lwkt_serialize_enter(slz);
766 * Directly block on the LWKT thread by descheduling it. This
767 * is much faster then tsleep(), but the only legal way to wake
768 * us up is to directly schedule the thread.
770 * Setting TDF_SINTR will cause new signals to directly schedule us.
772 * This routine must be called while in a critical section.
775 lwkt_sleep(const char *wmesg, int flags)
777 thread_t td = curthread;
780 if ((flags & PCATCH) == 0 || td->td_lwp == NULL) {
781 td->td_flags |= TDF_BLOCKED;
782 td->td_wmesg = wmesg;
783 lwkt_deschedule_self(td);
786 td->td_flags &= ~TDF_BLOCKED;
789 if ((sig = CURSIG(td->td_lwp)) != 0) {
790 if (SIGISMEMBER(td->td_proc->p_sigacts->ps_sigintr, sig))
796 td->td_flags |= TDF_BLOCKED | TDF_SINTR;
797 td->td_wmesg = wmesg;
798 lwkt_deschedule_self(td);
800 td->td_flags &= ~(TDF_BLOCKED | TDF_SINTR);
806 * Implement the timeout for tsleep.
808 * This type of callout timeout is scheduled on the same cpu the process
809 * is sleeping on. Also, at the moment, the MP lock is held.
818 * We are going to have to get the lwp_token, which means we might
819 * block. This can race a tsleep getting woken up by other means
820 * so set TDF_TIMEOUT_RUNNING to force the tsleep to wait for our
821 * processing to complete (sorry tsleep!).
823 * We can safely set td_flags because td MUST be on the same cpu
826 KKASSERT(td->td_gd == mycpu);
828 td->td_flags |= TDF_TIMEOUT_RUNNING | TDF_TIMEOUT;
831 * This can block but TDF_TIMEOUT_RUNNING will prevent the thread
832 * from exiting the tsleep on us. The flag is interlocked by virtue
833 * of lp being on the same cpu as we are.
835 if ((lp = td->td_lwp) != NULL)
836 lwkt_gettoken(&lp->lwp_token);
838 KKASSERT(td->td_flags & TDF_TSLEEP_DESCHEDULED);
842 * callout timer should never be set in tstop() because
843 * it passes a timeout of 0.
845 KKASSERT(lp->lwp_stat != LSSTOP);
847 lwkt_reltoken(&lp->lwp_token);
852 KKASSERT(td->td_gd == mycpu);
853 td->td_flags &= ~TDF_TIMEOUT_RUNNING;
858 * Make all processes sleeping on the specified identifier runnable.
859 * count may be zero or one only.
861 * The domain encodes the sleep/wakeup domain, flags, plus the originating
864 * This call may run without the MP lock held. We can only manipulate thread
865 * state on the cpu owning the thread. We CANNOT manipulate process state
868 * _wakeup() can be passed to an IPI so we can't use (const volatile
872 _wakeup(void *ident, int domain)
882 logtsleep2(wakeup_beg, ident);
885 qp = &gd->gd_tsleep_hash[id];
887 for (td = TAILQ_FIRST(qp); td != NULL; td = ntd) {
888 ntd = TAILQ_NEXT(td, td_sleepq);
889 if (td->td_wchan == ident &&
890 td->td_wdomain == (domain & PDOMAIN_MASK)
892 KKASSERT(td->td_gd == gd);
894 td->td_wakefromcpu = PWAKEUP_DECODE(domain);
895 if (td->td_flags & TDF_TSLEEP_DESCHEDULED) {
897 if (domain & PWAKEUP_ONE)
905 * We finished checking the current cpu but there still may be
906 * more work to do. Either wakeup_one was requested and no matching
907 * thread was found, or a normal wakeup was requested and we have
908 * to continue checking cpus.
910 * It should be noted that this scheme is actually less expensive then
911 * the old scheme when waking up multiple threads, since we send
912 * only one IPI message per target candidate which may then schedule
913 * multiple threads. Before we could have wound up sending an IPI
914 * message for each thread on the target cpu (!= current cpu) that
915 * needed to be woken up.
917 * NOTE: Wakeups occuring on remote cpus are asynchronous. This
918 * should be ok since we are passing idents in the IPI rather then
921 if ((domain & PWAKEUP_MYCPU) == 0) {
922 mask = slpque_cpumasks[id];
923 CPUMASK_ANDMASK(mask, gd->gd_other_cpus);
924 if (CPUMASK_TESTNZERO(mask)) {
925 lwkt_send_ipiq2_mask(mask, _wakeup, ident,
926 domain | PWAKEUP_MYCPU);
930 logtsleep1(wakeup_end);
935 * Wakeup all threads tsleep()ing on the specified ident, on all cpus
938 wakeup(const volatile void *ident)
940 globaldata_t gd = mycpu;
941 thread_t td = gd->gd_curthread;
943 if (td && (td->td_flags & TDF_DELAYED_WAKEUP)) {
945 * If we are in a delayed wakeup section, record up to two wakeups in
946 * a per-CPU queue and issue them when we block or exit the delayed
949 if (atomic_cmpset_ptr(&gd->gd_delayed_wakeup[0], NULL, ident))
951 if (atomic_cmpset_ptr(&gd->gd_delayed_wakeup[1], NULL, ident))
954 ident = atomic_swap_ptr(__DEQUALIFY(volatile void **, &gd->gd_delayed_wakeup[1]),
956 ident = atomic_swap_ptr(__DEQUALIFY(volatile void **, &gd->gd_delayed_wakeup[0]),
960 _wakeup(__DEALL(ident), PWAKEUP_ENCODE(0, gd->gd_cpuid));
964 * Wakeup one thread tsleep()ing on the specified ident, on any cpu.
967 wakeup_one(const volatile void *ident)
969 /* XXX potentially round-robin the first responding cpu */
970 _wakeup(__DEALL(ident), PWAKEUP_ENCODE(0, mycpu->gd_cpuid) |
975 * Wakeup threads tsleep()ing on the specified ident on the current cpu
979 wakeup_mycpu(const volatile void *ident)
981 _wakeup(__DEALL(ident), PWAKEUP_ENCODE(0, mycpu->gd_cpuid) |
986 * Wakeup one thread tsleep()ing on the specified ident on the current cpu
990 wakeup_mycpu_one(const volatile void *ident)
992 /* XXX potentially round-robin the first responding cpu */
993 _wakeup(__DEALL(ident), PWAKEUP_ENCODE(0, mycpu->gd_cpuid) |
994 PWAKEUP_MYCPU | PWAKEUP_ONE);
998 * Wakeup all thread tsleep()ing on the specified ident on the specified cpu
1002 wakeup_oncpu(globaldata_t gd, const volatile void *ident)
1004 globaldata_t mygd = mycpu;
1006 _wakeup(__DEALL(ident), PWAKEUP_ENCODE(0, mygd->gd_cpuid) |
1009 lwkt_send_ipiq2(gd, _wakeup, __DEALL(ident),
1010 PWAKEUP_ENCODE(0, mygd->gd_cpuid) |
1016 * Wakeup one thread tsleep()ing on the specified ident on the specified cpu
1020 wakeup_oncpu_one(globaldata_t gd, const volatile void *ident)
1022 globaldata_t mygd = mycpu;
1024 _wakeup(__DEALL(ident), PWAKEUP_ENCODE(0, mygd->gd_cpuid) |
1025 PWAKEUP_MYCPU | PWAKEUP_ONE);
1027 lwkt_send_ipiq2(gd, _wakeup, __DEALL(ident),
1028 PWAKEUP_ENCODE(0, mygd->gd_cpuid) |
1029 PWAKEUP_MYCPU | PWAKEUP_ONE);
1034 * Wakeup all threads waiting on the specified ident that slept using
1035 * the specified domain, on all cpus.
1038 wakeup_domain(const volatile void *ident, int domain)
1040 _wakeup(__DEALL(ident), PWAKEUP_ENCODE(domain, mycpu->gd_cpuid));
1044 * Wakeup one thread waiting on the specified ident that slept using
1045 * the specified domain, on any cpu.
1048 wakeup_domain_one(const volatile void *ident, int domain)
1050 /* XXX potentially round-robin the first responding cpu */
1051 _wakeup(__DEALL(ident),
1052 PWAKEUP_ENCODE(domain, mycpu->gd_cpuid) | PWAKEUP_ONE);
1056 wakeup_start_delayed(void)
1058 globaldata_t gd = mycpu;
1061 gd->gd_curthread->td_flags |= TDF_DELAYED_WAKEUP;
1066 wakeup_end_delayed(void)
1068 globaldata_t gd = mycpu;
1070 if (gd->gd_curthread->td_flags & TDF_DELAYED_WAKEUP) {
1072 gd->gd_curthread->td_flags &= ~TDF_DELAYED_WAKEUP;
1073 if (gd->gd_delayed_wakeup[0] || gd->gd_delayed_wakeup[1]) {
1074 if (gd->gd_delayed_wakeup[0]) {
1075 wakeup(gd->gd_delayed_wakeup[0]);
1076 gd->gd_delayed_wakeup[0] = NULL;
1078 if (gd->gd_delayed_wakeup[1]) {
1079 wakeup(gd->gd_delayed_wakeup[1]);
1080 gd->gd_delayed_wakeup[1] = NULL;
1090 * Make a process runnable. lp->lwp_token must be held on call and this
1091 * function must be called from the cpu owning lp.
1093 * This only has an effect if we are in LSSTOP or LSSLEEP.
1096 setrunnable(struct lwp *lp)
1098 thread_t td = lp->lwp_thread;
1100 ASSERT_LWKT_TOKEN_HELD(&lp->lwp_token);
1101 KKASSERT(td->td_gd == mycpu);
1103 if (lp->lwp_stat == LSSTOP)
1104 lp->lwp_stat = LSSLEEP;
1105 if (lp->lwp_stat == LSSLEEP) {
1108 } else if (td->td_flags & TDF_SINTR) {
1115 * The process is stopped due to some condition, usually because p_stat is
1116 * set to SSTOP, but also possibly due to being traced.
1118 * Caller must hold p->p_token
1120 * NOTE! If the caller sets SSTOP, the caller must also clear P_WAITED
1121 * because the parent may check the child's status before the child actually
1122 * gets to this routine.
1124 * This routine is called with the current lwp only, typically just
1125 * before returning to userland if the process state is detected as
1126 * possibly being in a stopped state.
1131 struct lwp *lp = curthread->td_lwp;
1132 struct proc *p = lp->lwp_proc;
1135 lwkt_gettoken(&lp->lwp_token);
1139 * If LWP_MP_WSTOP is set, we were sleeping
1140 * while our process was stopped. At this point
1141 * we were already counted as stopped.
1143 if ((lp->lwp_mpflags & LWP_MP_WSTOP) == 0) {
1145 * If we're the last thread to stop, signal
1149 atomic_set_int(&lp->lwp_mpflags, LWP_MP_WSTOP);
1150 wakeup(&p->p_nstopped);
1151 if (p->p_nstopped == p->p_nthreads) {
1153 * Token required to interlock kern_wait()
1157 lwkt_gettoken(&q->p_token);
1158 p->p_flags &= ~P_WAITED;
1160 if ((q->p_sigacts->ps_flag & PS_NOCLDSTOP) == 0)
1161 ksignal(q, SIGCHLD);
1162 lwkt_reltoken(&q->p_token);
1166 while (p->p_stat == SSTOP) {
1167 lp->lwp_stat = LSSTOP;
1168 tsleep(p, 0, "stop", 0);
1171 atomic_clear_int(&lp->lwp_mpflags, LWP_MP_WSTOP);
1173 lwkt_reltoken(&lp->lwp_token);
1177 * Compute a tenex style load average of a quantity on
1178 * 1, 5 and 15 minute intervals.
1180 static int loadav_count_runnable(struct lwp *p, void *data);
1185 struct loadavg *avg;
1189 alllwp_scan(loadav_count_runnable, &nrun);
1191 for (i = 0; i < 3; i++) {
1192 avg->ldavg[i] = (cexp[i] * avg->ldavg[i] +
1193 nrun * FSCALE * (FSCALE - cexp[i])) >> FSHIFT;
1197 * Schedule the next update to occur after 5 seconds, but add a
1198 * random variation to avoid synchronisation with processes that
1199 * run at regular intervals.
1201 callout_reset(&loadav_callout, hz * 4 + (int)(krandom() % (hz * 2 + 1)),
1206 loadav_count_runnable(struct lwp *lp, void *data)
1211 switch (lp->lwp_stat) {
1213 if ((td = lp->lwp_thread) == NULL)
1215 if (td->td_flags & TDF_BLOCKED)
1228 sched_setup(void *dummy)
1230 callout_init_mp(&loadav_callout);
1231 callout_init_mp(&schedcpu_callout);
1233 /* Kick off timeout driven events by calling first time. */