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>
53 #include <sys/xwait.h>
55 #include <sys/serialize.h>
57 #include <sys/signal2.h>
58 #include <sys/thread2.h>
59 #include <sys/spinlock2.h>
60 #include <sys/mutex2.h>
62 #include <machine/cpu.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)
72 int sched_quantum; /* Roundrobin scheduling quantum in ticks. */
74 int ncpus2, ncpus2_shift, ncpus2_mask; /* note: mask not cpumask_t */
75 int ncpus_fit, ncpus_fit_mask; /* note: mask not cpumask_t */
78 int tsleep_crypto_dump = 0;
80 static struct callout loadav_callout;
81 static struct callout schedcpu_callout;
82 MALLOC_DEFINE(M_TSLEEP, "tslpque", "tsleep queues");
84 #define __DEALL(ident) __DEQUALIFY(void *, ident)
86 #if !defined(KTR_TSLEEP)
87 #define KTR_TSLEEP KTR_ALL
89 KTR_INFO_MASTER(tsleep);
90 KTR_INFO(KTR_TSLEEP, tsleep, tsleep_beg, 0, "tsleep enter %p", const volatile void *ident);
91 KTR_INFO(KTR_TSLEEP, tsleep, tsleep_end, 1, "tsleep exit");
92 KTR_INFO(KTR_TSLEEP, tsleep, wakeup_beg, 2, "wakeup enter %p", const volatile void *ident);
93 KTR_INFO(KTR_TSLEEP, tsleep, wakeup_end, 3, "wakeup exit");
94 KTR_INFO(KTR_TSLEEP, tsleep, ilockfail, 4, "interlock failed %p", const volatile void *ident);
96 #define logtsleep1(name) KTR_LOG(tsleep_ ## name)
97 #define logtsleep2(name, val) KTR_LOG(tsleep_ ## name, val)
99 struct loadavg averunnable =
100 { {0, 0, 0}, FSCALE }; /* load average, of runnable procs */
102 * Constants for averages over 1, 5, and 15 minutes
103 * when sampling at 5 second intervals.
105 static fixpt_t cexp[3] = {
106 0.9200444146293232 * FSCALE, /* exp(-1/12) */
107 0.9834714538216174 * FSCALE, /* exp(-1/60) */
108 0.9944598480048967 * FSCALE, /* exp(-1/180) */
111 static void endtsleep (void *);
112 static void loadav (void *arg);
113 static void schedcpu (void *arg);
116 * Adjust the scheduler quantum. The quantum is specified in microseconds.
117 * Note that 'tick' is in microseconds per tick.
120 sysctl_kern_quantum(SYSCTL_HANDLER_ARGS)
124 new_val = sched_quantum * ustick;
125 error = sysctl_handle_int(oidp, &new_val, 0, req);
126 if (error != 0 || req->newptr == NULL)
128 if (new_val < ustick)
130 sched_quantum = new_val / ustick;
134 SYSCTL_PROC(_kern, OID_AUTO, quantum, CTLTYPE_INT|CTLFLAG_RW,
135 0, sizeof sched_quantum, sysctl_kern_quantum, "I", "");
137 static int pctcpu_decay = 10;
138 SYSCTL_INT(_kern, OID_AUTO, pctcpu_decay, CTLFLAG_RW, &pctcpu_decay, 0, "");
141 * kernel uses `FSCALE', userland (SHOULD) use kern.fscale
143 int fscale __unused = FSCALE; /* exported to systat */
144 SYSCTL_INT(_kern, OID_AUTO, fscale, CTLFLAG_RD, 0, FSCALE, "");
147 * Recompute process priorities, once a second.
149 * Since the userland schedulers are typically event oriented, if the
150 * estcpu calculation at wakeup() time is not sufficient to make a
151 * process runnable relative to other processes in the system we have
152 * a 1-second recalc to help out.
154 * This code also allows us to store sysclock_t data in the process structure
155 * without fear of an overrun, since sysclock_t are guarenteed to hold
156 * several seconds worth of count.
158 * WARNING! callouts can preempt normal threads. However, they will not
159 * preempt a thread holding a spinlock so we *can* safely use spinlocks.
161 static int schedcpu_stats(struct proc *p, void *data __unused);
162 static int schedcpu_resource(struct proc *p, void *data __unused);
167 allproc_scan(schedcpu_stats, NULL);
168 allproc_scan(schedcpu_resource, NULL);
169 wakeup((caddr_t)&lbolt);
170 wakeup(lbolt_syncer);
171 callout_reset(&schedcpu_callout, hz, schedcpu, NULL);
175 * General process statistics once a second
178 schedcpu_stats(struct proc *p, void *data __unused)
183 * Threads may not be completely set up if process in SIDL state.
185 if (p->p_stat == SIDL)
189 if (lwkt_trytoken(&p->p_token) == FALSE) {
195 FOREACH_LWP_IN_PROC(lp, p) {
196 if (lp->lwp_stat == LSSLEEP) {
198 if (lp->lwp_slptime == 1)
199 p->p_usched->uload_update(lp);
203 * Only recalculate processes that are active or have slept
204 * less then 2 seconds. The schedulers understand this.
205 * Otherwise decay by 50% per second.
207 if (lp->lwp_slptime <= 1) {
208 p->p_usched->recalculate(lp);
212 decay = pctcpu_decay;
218 lp->lwp_pctcpu = (lp->lwp_pctcpu * (decay - 1)) / decay;
221 lwkt_reltoken(&p->p_token);
228 * Resource checks. XXX break out since ksignal/killproc can block,
229 * limiting us to one process killed per second. There is probably
233 schedcpu_resource(struct proc *p, void *data __unused)
238 if (p->p_stat == SIDL)
242 if (lwkt_trytoken(&p->p_token) == FALSE) {
247 if (p->p_stat == SZOMB || p->p_limit == NULL) {
248 lwkt_reltoken(&p->p_token);
254 FOREACH_LWP_IN_PROC(lp, p) {
256 * We may have caught an lp in the middle of being
257 * created, lwp_thread can be NULL.
259 if (lp->lwp_thread) {
260 ttime += lp->lwp_thread->td_sticks;
261 ttime += lp->lwp_thread->td_uticks;
265 switch(plimit_testcpulimit(p->p_limit, ttime)) {
266 case PLIMIT_TESTCPU_KILL:
267 killproc(p, "exceeded maximum CPU limit");
269 case PLIMIT_TESTCPU_XCPU:
270 if ((p->p_flags & P_XCPU) == 0) {
271 p->p_flags |= P_XCPU;
278 lwkt_reltoken(&p->p_token);
285 * This is only used by ps. Generate a cpu percentage use over
286 * a period of one second.
289 updatepcpu(struct lwp *lp, int cpticks, int ttlticks)
294 acc = (cpticks << FSHIFT) / ttlticks;
295 if (ttlticks >= ESTCPUFREQ) {
296 lp->lwp_pctcpu = acc;
298 remticks = ESTCPUFREQ - ttlticks;
299 lp->lwp_pctcpu = (acc * ttlticks + lp->lwp_pctcpu * remticks) /
305 * tsleep/wakeup hash table parameters. Try to find the sweet spot for
306 * like addresses being slept on.
308 #define TABLESIZE 4001
309 #define LOOKUP(x) (((u_int)(uintptr_t)(x)) % TABLESIZE)
311 static cpumask_t slpque_cpumasks[TABLESIZE];
314 * General scheduler initialization. We force a reschedule 25 times
315 * a second by default. Note that cpu0 is initialized in early boot and
316 * cannot make any high level calls.
318 * Each cpu has its own sleep queue.
321 sleep_gdinit(globaldata_t gd)
323 static struct tslpque slpque_cpu0[TABLESIZE];
326 if (gd->gd_cpuid == 0) {
327 sched_quantum = (hz + 24) / 25;
328 gd->gd_tsleep_hash = slpque_cpu0;
330 gd->gd_tsleep_hash = kmalloc(sizeof(slpque_cpu0),
331 M_TSLEEP, M_WAITOK | M_ZERO);
333 for (i = 0; i < TABLESIZE; ++i)
334 TAILQ_INIT(&gd->gd_tsleep_hash[i]);
338 * This is a dandy function that allows us to interlock tsleep/wakeup
339 * operations with unspecified upper level locks, such as lockmgr locks,
340 * simply by holding a critical section. The sequence is:
342 * (acquire upper level lock)
343 * tsleep_interlock(blah)
344 * (release upper level lock)
347 * Basically this functions queues us on the tsleep queue without actually
348 * descheduling us. When tsleep() is later called with PINTERLOCK it
349 * assumes the thread was already queued, otherwise it queues it there.
351 * Thus it is possible to receive the wakeup prior to going to sleep and
352 * the race conditions are covered.
355 _tsleep_interlock(globaldata_t gd, const volatile void *ident, int flags)
357 thread_t td = gd->gd_curthread;
360 crit_enter_quick(td);
361 if (td->td_flags & TDF_TSLEEPQ) {
362 id = LOOKUP(td->td_wchan);
363 TAILQ_REMOVE(&gd->gd_tsleep_hash[id], td, td_sleepq);
364 if (TAILQ_FIRST(&gd->gd_tsleep_hash[id]) == NULL) {
365 ATOMIC_CPUMASK_NANDBIT(slpque_cpumasks[id],
369 td->td_flags |= TDF_TSLEEPQ;
372 TAILQ_INSERT_TAIL(&gd->gd_tsleep_hash[id], td, td_sleepq);
373 ATOMIC_CPUMASK_ORBIT(slpque_cpumasks[id], gd->gd_cpuid);
374 td->td_wchan = ident;
375 td->td_wdomain = flags & PDOMAIN_MASK;
380 tsleep_interlock(const volatile void *ident, int flags)
382 _tsleep_interlock(mycpu, ident, flags);
386 * Remove thread from sleepq. Must be called with a critical section held.
387 * The thread must not be migrating.
390 _tsleep_remove(thread_t td)
392 globaldata_t gd = mycpu;
395 KKASSERT(td->td_gd == gd && IN_CRITICAL_SECT(td));
396 KKASSERT((td->td_flags & TDF_MIGRATING) == 0);
397 if (td->td_flags & TDF_TSLEEPQ) {
398 td->td_flags &= ~TDF_TSLEEPQ;
399 id = LOOKUP(td->td_wchan);
400 TAILQ_REMOVE(&gd->gd_tsleep_hash[id], td, td_sleepq);
401 if (TAILQ_FIRST(&gd->gd_tsleep_hash[id]) == NULL) {
402 ATOMIC_CPUMASK_NANDBIT(slpque_cpumasks[id],
411 tsleep_remove(thread_t td)
417 * General sleep call. Suspends the current process until a wakeup is
418 * performed on the specified identifier. The process will then be made
419 * runnable with the specified priority. Sleeps at most timo/hz seconds
420 * (0 means no timeout). If flags includes PCATCH flag, signals are checked
421 * before and after sleeping, else signals are not checked. Returns 0 if
422 * awakened, EWOULDBLOCK if the timeout expires. If PCATCH is set and a
423 * signal needs to be delivered, ERESTART is returned if the current system
424 * call should be restarted if possible, and EINTR is returned if the system
425 * call should be interrupted by the signal (return EINTR).
427 * Note that if we are a process, we release_curproc() before messing with
428 * the LWKT scheduler.
430 * During autoconfiguration or after a panic, a sleep will simply
431 * lower the priority briefly to allow interrupts, then return.
433 * WARNING! This code can't block (short of switching away), or bad things
434 * will happen. No getting tokens, no blocking locks, etc.
437 tsleep(const volatile void *ident, int flags, const char *wmesg, int timo)
439 struct thread *td = curthread;
440 struct lwp *lp = td->td_lwp;
441 struct proc *p = td->td_proc; /* may be NULL */
447 struct callout thandle;
450 * Currently a severe hack. Make sure any delayed wakeups
451 * are flushed before we sleep or we might deadlock on whatever
452 * event we are sleeping on.
454 if (td->td_flags & TDF_DELAYED_WAKEUP)
455 wakeup_end_delayed();
458 * NOTE: removed KTRPOINT, it could cause races due to blocking
459 * even in stable. Just scrap it for now.
461 if (!tsleep_crypto_dump && (tsleep_now_works == 0 || panicstr)) {
463 * After a panic, or before we actually have an operational
464 * softclock, just give interrupts a chance, then just return;
466 * don't run any other procs or panic below,
467 * in case this is the idle process and already asleep.
471 lwkt_setpri_self(safepri);
473 lwkt_setpri_self(oldpri);
476 logtsleep2(tsleep_beg, ident);
478 KKASSERT(td != &gd->gd_idlethread); /* you must be kidding! */
479 td->td_wakefromcpu = -1; /* overwritten by _wakeup */
482 * NOTE: all of this occurs on the current cpu, including any
483 * callout-based wakeups, so a critical section is a sufficient
486 * The entire sequence through to where we actually sleep must
487 * run without breaking the critical section.
489 catch = flags & PCATCH;
493 crit_enter_quick(td);
495 KASSERT(ident != NULL, ("tsleep: no ident"));
496 KASSERT(lp == NULL ||
497 lp->lwp_stat == LSRUN || /* Obvious */
498 lp->lwp_stat == LSSTOP, /* Set in tstop */
500 ident, wmesg, lp->lwp_stat));
503 * We interlock the sleep queue if the caller has not already done
504 * it for us. This must be done before we potentially acquire any
505 * tokens or we can loose the wakeup.
507 if ((flags & PINTERLOCKED) == 0) {
508 _tsleep_interlock(gd, ident, flags);
512 * Setup for the current process (if this is a process). We must
513 * interlock with lwp_token to avoid remote wakeup races via
517 lwkt_gettoken(&lp->lwp_token);
520 * Early termination if PCATCH was set and a
521 * signal is pending, interlocked with the
524 * Early termination only occurs when tsleep() is
525 * entered while in a normal LSRUN state.
527 if ((sig = CURSIG(lp)) != 0)
531 * Causes ksignal to wake us up if a signal is
532 * received (interlocked with p->p_token).
534 lp->lwp_flags |= LWP_SINTR;
541 * Make sure the current process has been untangled from
542 * the userland scheduler and initialize slptime to start
545 * NOTE: td->td_wakefromcpu is pre-set by the release function
546 * for the dfly scheduler, and then adjusted by _wakeup()
549 p->p_usched->release_curproc(lp);
554 * If the interlocked flag is set but our cpu bit in the slpqueue
555 * is no longer set, then a wakeup was processed inbetween the
556 * tsleep_interlock() (ours or the callers), and here. This can
557 * occur under numerous circumstances including when we release the
560 * Extreme loads can cause the sending of an IPI (e.g. wakeup()'s)
561 * to process incoming IPIs, thus draining incoming wakeups.
563 if ((td->td_flags & TDF_TSLEEPQ) == 0) {
564 logtsleep2(ilockfail, ident);
569 * scheduling is blocked while in a critical section. Coincide
570 * the descheduled-by-tsleep flag with the descheduling of the
573 * The timer callout is localized on our cpu and interlocked by
574 * our critical section.
576 lwkt_deschedule_self(td);
577 td->td_flags |= TDF_TSLEEP_DESCHEDULED;
578 td->td_wmesg = wmesg;
581 * Setup the timeout, if any. The timeout is only operable while
582 * the thread is flagged descheduled.
584 KKASSERT((td->td_flags & TDF_TIMEOUT) == 0);
586 callout_init_mp(&thandle);
587 callout_reset(&thandle, timo, endtsleep, td);
595 * Ok, we are sleeping. Place us in the SSLEEP state.
597 KKASSERT((lp->lwp_mpflags & LWP_MP_ONRUNQ) == 0);
600 * tstop() sets LSSTOP, so don't fiddle with that.
602 if (lp->lwp_stat != LSSTOP)
603 lp->lwp_stat = LSSLEEP;
604 lp->lwp_ru.ru_nvcsw++;
605 p->p_usched->uload_update(lp);
609 * And when we are woken up, put us back in LSRUN. If we
610 * slept for over a second, recalculate our estcpu.
612 lp->lwp_stat = LSRUN;
613 if (lp->lwp_slptime) {
614 p->p_usched->uload_update(lp);
615 p->p_usched->recalculate(lp);
623 * Make sure we haven't switched cpus while we were asleep. It's
624 * not supposed to happen. Cleanup our temporary flags.
626 KKASSERT(gd == td->td_gd);
629 * Cleanup the timeout. If the timeout has already occured thandle
630 * has already been stopped, otherwise stop thandle. If the timeout
631 * is running (the callout thread must be blocked trying to get
632 * lwp_token) then wait for us to get scheduled.
635 while (td->td_flags & TDF_TIMEOUT_RUNNING) {
636 lwkt_deschedule_self(td);
637 td->td_wmesg = "tsrace";
639 kprintf("td %p %s: timeout race\n", td, td->td_comm);
641 if (td->td_flags & TDF_TIMEOUT) {
642 td->td_flags &= ~TDF_TIMEOUT;
645 /* does not block when on same cpu */
646 callout_stop(&thandle);
649 td->td_flags &= ~TDF_TSLEEP_DESCHEDULED;
652 * Make sure we have been removed from the sleepq. In most
653 * cases this will have been done for us already but it is
654 * possible for a scheduling IPI to be in-flight from a
655 * previous tsleep/tsleep_interlock() or due to a straight-out
656 * call to lwkt_schedule() (in the case of an interrupt thread),
657 * causing a spurious wakeup.
663 * Figure out the correct error return. If interrupted by a
664 * signal we want to return EINTR or ERESTART.
668 if (catch && error == 0) {
669 if (sig != 0 || (sig = CURSIG(lp))) {
670 if (SIGISMEMBER(p->p_sigacts->ps_sigintr, sig))
676 lp->lwp_flags &= ~LWP_SINTR;
677 lwkt_reltoken(&lp->lwp_token);
679 logtsleep1(tsleep_end);
685 * Interlocked spinlock sleep. An exclusively held spinlock must
686 * be passed to ssleep(). The function will atomically release the
687 * spinlock and tsleep on the ident, then reacquire the spinlock and
690 * This routine is fairly important along the critical path, so optimize it
694 ssleep(const volatile void *ident, struct spinlock *spin, int flags,
695 const char *wmesg, int timo)
697 globaldata_t gd = mycpu;
700 _tsleep_interlock(gd, ident, flags);
701 spin_unlock_quick(gd, spin);
702 error = tsleep(ident, flags | PINTERLOCKED, wmesg, timo);
703 _spin_lock_quick(gd, spin, wmesg);
709 lksleep(const volatile void *ident, struct lock *lock, int flags,
710 const char *wmesg, int timo)
712 globaldata_t gd = mycpu;
715 _tsleep_interlock(gd, ident, flags);
716 lockmgr(lock, LK_RELEASE);
717 error = tsleep(ident, flags | PINTERLOCKED, wmesg, timo);
718 lockmgr(lock, LK_EXCLUSIVE);
724 * Interlocked mutex sleep. An exclusively held mutex must be passed
725 * to mtxsleep(). The function will atomically release the mutex
726 * and tsleep on the ident, then reacquire the mutex and return.
729 mtxsleep(const volatile void *ident, struct mtx *mtx, int flags,
730 const char *wmesg, int timo)
732 globaldata_t gd = mycpu;
735 _tsleep_interlock(gd, ident, flags);
737 error = tsleep(ident, flags | PINTERLOCKED, wmesg, timo);
738 mtx_lock_ex_quick(mtx, wmesg);
744 * Interlocked serializer sleep. An exclusively held serializer must
745 * be passed to zsleep(). The function will atomically release
746 * the serializer and tsleep on the ident, then reacquire the serializer
750 zsleep(const volatile void *ident, struct lwkt_serialize *slz, int flags,
751 const char *wmesg, int timo)
753 globaldata_t gd = mycpu;
756 ASSERT_SERIALIZED(slz);
758 _tsleep_interlock(gd, ident, flags);
759 lwkt_serialize_exit(slz);
760 ret = tsleep(ident, flags | PINTERLOCKED, wmesg, timo);
761 lwkt_serialize_enter(slz);
767 * Directly block on the LWKT thread by descheduling it. This
768 * is much faster then tsleep(), but the only legal way to wake
769 * us up is to directly schedule the thread.
771 * Setting TDF_SINTR will cause new signals to directly schedule us.
773 * This routine must be called while in a critical section.
776 lwkt_sleep(const char *wmesg, int flags)
778 thread_t td = curthread;
781 if ((flags & PCATCH) == 0 || td->td_lwp == NULL) {
782 td->td_flags |= TDF_BLOCKED;
783 td->td_wmesg = wmesg;
784 lwkt_deschedule_self(td);
787 td->td_flags &= ~TDF_BLOCKED;
790 if ((sig = CURSIG(td->td_lwp)) != 0) {
791 if (SIGISMEMBER(td->td_proc->p_sigacts->ps_sigintr, sig))
797 td->td_flags |= TDF_BLOCKED | TDF_SINTR;
798 td->td_wmesg = wmesg;
799 lwkt_deschedule_self(td);
801 td->td_flags &= ~(TDF_BLOCKED | TDF_SINTR);
807 * Implement the timeout for tsleep.
809 * This type of callout timeout is scheduled on the same cpu the process
810 * is sleeping on. Also, at the moment, the MP lock is held.
819 * We are going to have to get the lwp_token, which means we might
820 * block. This can race a tsleep getting woken up by other means
821 * so set TDF_TIMEOUT_RUNNING to force the tsleep to wait for our
822 * processing to complete (sorry tsleep!).
824 * We can safely set td_flags because td MUST be on the same cpu
827 KKASSERT(td->td_gd == mycpu);
829 td->td_flags |= TDF_TIMEOUT_RUNNING | TDF_TIMEOUT;
832 * This can block but TDF_TIMEOUT_RUNNING will prevent the thread
833 * from exiting the tsleep on us. The flag is interlocked by virtue
834 * of lp being on the same cpu as we are.
836 if ((lp = td->td_lwp) != NULL)
837 lwkt_gettoken(&lp->lwp_token);
839 KKASSERT(td->td_flags & TDF_TSLEEP_DESCHEDULED);
843 * callout timer should never be set in tstop() because
844 * it passes a timeout of 0.
846 KKASSERT(lp->lwp_stat != LSSTOP);
848 lwkt_reltoken(&lp->lwp_token);
853 KKASSERT(td->td_gd == mycpu);
854 td->td_flags &= ~TDF_TIMEOUT_RUNNING;
859 * Make all processes sleeping on the specified identifier runnable.
860 * count may be zero or one only.
862 * The domain encodes the sleep/wakeup domain, flags, plus the originating
865 * This call may run without the MP lock held. We can only manipulate thread
866 * state on the cpu owning the thread. We CANNOT manipulate process state
869 * _wakeup() can be passed to an IPI so we can't use (const volatile
873 _wakeup(void *ident, int domain)
883 logtsleep2(wakeup_beg, ident);
886 qp = &gd->gd_tsleep_hash[id];
888 for (td = TAILQ_FIRST(qp); td != NULL; td = ntd) {
889 ntd = TAILQ_NEXT(td, td_sleepq);
890 if (td->td_wchan == ident &&
891 td->td_wdomain == (domain & PDOMAIN_MASK)
893 KKASSERT(td->td_gd == gd);
895 td->td_wakefromcpu = PWAKEUP_DECODE(domain);
896 if (td->td_flags & TDF_TSLEEP_DESCHEDULED) {
898 if (domain & PWAKEUP_ONE)
906 * We finished checking the current cpu but there still may be
907 * more work to do. Either wakeup_one was requested and no matching
908 * thread was found, or a normal wakeup was requested and we have
909 * to continue checking cpus.
911 * It should be noted that this scheme is actually less expensive then
912 * the old scheme when waking up multiple threads, since we send
913 * only one IPI message per target candidate which may then schedule
914 * multiple threads. Before we could have wound up sending an IPI
915 * message for each thread on the target cpu (!= current cpu) that
916 * needed to be woken up.
918 * NOTE: Wakeups occuring on remote cpus are asynchronous. This
919 * should be ok since we are passing idents in the IPI rather then
922 if ((domain & PWAKEUP_MYCPU) == 0) {
923 mask = slpque_cpumasks[id];
924 CPUMASK_ANDMASK(mask, gd->gd_other_cpus);
925 if (CPUMASK_TESTNZERO(mask)) {
926 lwkt_send_ipiq2_mask(mask, _wakeup, ident,
927 domain | PWAKEUP_MYCPU);
931 logtsleep1(wakeup_end);
936 * Wakeup all threads tsleep()ing on the specified ident, on all cpus
939 wakeup(const volatile void *ident)
941 globaldata_t gd = mycpu;
942 thread_t td = gd->gd_curthread;
944 if (td && (td->td_flags & TDF_DELAYED_WAKEUP)) {
946 * If we are in a delayed wakeup section, record up to two wakeups in
947 * a per-CPU queue and issue them when we block or exit the delayed
950 if (atomic_cmpset_ptr(&gd->gd_delayed_wakeup[0], NULL, ident))
952 if (atomic_cmpset_ptr(&gd->gd_delayed_wakeup[1], NULL, ident))
955 ident = atomic_swap_ptr(__DEQUALIFY(volatile void **, &gd->gd_delayed_wakeup[1]),
957 ident = atomic_swap_ptr(__DEQUALIFY(volatile void **, &gd->gd_delayed_wakeup[0]),
961 _wakeup(__DEALL(ident), PWAKEUP_ENCODE(0, gd->gd_cpuid));
965 * Wakeup one thread tsleep()ing on the specified ident, on any cpu.
968 wakeup_one(const volatile void *ident)
970 /* XXX potentially round-robin the first responding cpu */
971 _wakeup(__DEALL(ident), PWAKEUP_ENCODE(0, mycpu->gd_cpuid) |
976 * Wakeup threads tsleep()ing on the specified ident on the current cpu
980 wakeup_mycpu(const volatile void *ident)
982 _wakeup(__DEALL(ident), PWAKEUP_ENCODE(0, mycpu->gd_cpuid) |
987 * Wakeup one thread tsleep()ing on the specified ident on the current cpu
991 wakeup_mycpu_one(const volatile void *ident)
993 /* XXX potentially round-robin the first responding cpu */
994 _wakeup(__DEALL(ident), PWAKEUP_ENCODE(0, mycpu->gd_cpuid) |
995 PWAKEUP_MYCPU | PWAKEUP_ONE);
999 * Wakeup all thread tsleep()ing on the specified ident on the specified cpu
1003 wakeup_oncpu(globaldata_t gd, const volatile void *ident)
1005 globaldata_t mygd = mycpu;
1007 _wakeup(__DEALL(ident), PWAKEUP_ENCODE(0, mygd->gd_cpuid) |
1010 lwkt_send_ipiq2(gd, _wakeup, __DEALL(ident),
1011 PWAKEUP_ENCODE(0, mygd->gd_cpuid) |
1017 * Wakeup one thread tsleep()ing on the specified ident on the specified cpu
1021 wakeup_oncpu_one(globaldata_t gd, const volatile void *ident)
1023 globaldata_t mygd = mycpu;
1025 _wakeup(__DEALL(ident), PWAKEUP_ENCODE(0, mygd->gd_cpuid) |
1026 PWAKEUP_MYCPU | PWAKEUP_ONE);
1028 lwkt_send_ipiq2(gd, _wakeup, __DEALL(ident),
1029 PWAKEUP_ENCODE(0, mygd->gd_cpuid) |
1030 PWAKEUP_MYCPU | PWAKEUP_ONE);
1035 * Wakeup all threads waiting on the specified ident that slept using
1036 * the specified domain, on all cpus.
1039 wakeup_domain(const volatile void *ident, int domain)
1041 _wakeup(__DEALL(ident), PWAKEUP_ENCODE(domain, mycpu->gd_cpuid));
1045 * Wakeup one thread waiting on the specified ident that slept using
1046 * the specified domain, on any cpu.
1049 wakeup_domain_one(const volatile void *ident, int domain)
1051 /* XXX potentially round-robin the first responding cpu */
1052 _wakeup(__DEALL(ident),
1053 PWAKEUP_ENCODE(domain, mycpu->gd_cpuid) | PWAKEUP_ONE);
1057 wakeup_start_delayed(void)
1059 globaldata_t gd = mycpu;
1062 gd->gd_curthread->td_flags |= TDF_DELAYED_WAKEUP;
1067 wakeup_end_delayed(void)
1069 globaldata_t gd = mycpu;
1071 if (gd->gd_curthread->td_flags & TDF_DELAYED_WAKEUP) {
1073 gd->gd_curthread->td_flags &= ~TDF_DELAYED_WAKEUP;
1074 if (gd->gd_delayed_wakeup[0] || gd->gd_delayed_wakeup[1]) {
1075 if (gd->gd_delayed_wakeup[0]) {
1076 wakeup(gd->gd_delayed_wakeup[0]);
1077 gd->gd_delayed_wakeup[0] = NULL;
1079 if (gd->gd_delayed_wakeup[1]) {
1080 wakeup(gd->gd_delayed_wakeup[1]);
1081 gd->gd_delayed_wakeup[1] = NULL;
1091 * Make a process runnable. lp->lwp_token must be held on call and this
1092 * function must be called from the cpu owning lp.
1094 * This only has an effect if we are in LSSTOP or LSSLEEP.
1097 setrunnable(struct lwp *lp)
1099 thread_t td = lp->lwp_thread;
1101 ASSERT_LWKT_TOKEN_HELD(&lp->lwp_token);
1102 KKASSERT(td->td_gd == mycpu);
1104 if (lp->lwp_stat == LSSTOP)
1105 lp->lwp_stat = LSSLEEP;
1106 if (lp->lwp_stat == LSSLEEP) {
1109 } else if (td->td_flags & TDF_SINTR) {
1116 * The process is stopped due to some condition, usually because p_stat is
1117 * set to SSTOP, but also possibly due to being traced.
1119 * Caller must hold p->p_token
1121 * NOTE! If the caller sets SSTOP, the caller must also clear P_WAITED
1122 * because the parent may check the child's status before the child actually
1123 * gets to this routine.
1125 * This routine is called with the current lwp only, typically just
1126 * before returning to userland if the process state is detected as
1127 * possibly being in a stopped state.
1132 struct lwp *lp = curthread->td_lwp;
1133 struct proc *p = lp->lwp_proc;
1136 lwkt_gettoken(&lp->lwp_token);
1140 * If LWP_MP_WSTOP is set, we were sleeping
1141 * while our process was stopped. At this point
1142 * we were already counted as stopped.
1144 if ((lp->lwp_mpflags & LWP_MP_WSTOP) == 0) {
1146 * If we're the last thread to stop, signal
1150 atomic_set_int(&lp->lwp_mpflags, LWP_MP_WSTOP);
1151 wakeup(&p->p_nstopped);
1152 if (p->p_nstopped == p->p_nthreads) {
1154 * Token required to interlock kern_wait()
1158 lwkt_gettoken(&q->p_token);
1159 p->p_flags &= ~P_WAITED;
1161 if ((q->p_sigacts->ps_flag & PS_NOCLDSTOP) == 0)
1162 ksignal(q, SIGCHLD);
1163 lwkt_reltoken(&q->p_token);
1167 while (p->p_stat == SSTOP) {
1168 lp->lwp_stat = LSSTOP;
1169 tsleep(p, 0, "stop", 0);
1172 atomic_clear_int(&lp->lwp_mpflags, LWP_MP_WSTOP);
1174 lwkt_reltoken(&lp->lwp_token);
1178 * Compute a tenex style load average of a quantity on
1179 * 1, 5 and 15 minute intervals.
1181 static int loadav_count_runnable(struct lwp *p, void *data);
1186 struct loadavg *avg;
1190 alllwp_scan(loadav_count_runnable, &nrun);
1192 for (i = 0; i < 3; i++) {
1193 avg->ldavg[i] = (cexp[i] * avg->ldavg[i] +
1194 nrun * FSCALE * (FSCALE - cexp[i])) >> FSHIFT;
1198 * Schedule the next update to occur after 5 seconds, but add a
1199 * random variation to avoid synchronisation with processes that
1200 * run at regular intervals.
1202 callout_reset(&loadav_callout, hz * 4 + (int)(krandom() % (hz * 2 + 1)),
1207 loadav_count_runnable(struct lwp *lp, void *data)
1212 switch (lp->lwp_stat) {
1214 if ((td = lp->lwp_thread) == NULL)
1216 if (td->td_flags & TDF_BLOCKED)
1229 sched_setup(void *dummy)
1231 callout_init_mp(&loadav_callout);
1232 callout_init_mp(&schedcpu_callout);
1234 /* Kick off timeout driven events by calling first time. */