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 1.91 2008/09/09 04:06:13 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>
56 #include <sys/ktrace.h>
58 #include <sys/xwait.h>
60 #include <sys/serialize.h>
62 #include <sys/signal2.h>
63 #include <sys/thread2.h>
64 #include <sys/spinlock2.h>
65 #include <sys/mutex2.h>
67 #include <machine/cpu.h>
68 #include <machine/smp.h>
70 TAILQ_HEAD(tslpque, thread);
72 static void sched_setup (void *dummy);
73 SYSINIT(sched_setup, SI_SUB_KICK_SCHEDULER, SI_ORDER_FIRST, sched_setup, NULL)
78 int sched_quantum; /* Roundrobin scheduling quantum in ticks. */
80 int ncpus2, ncpus2_shift, ncpus2_mask; /* note: mask not cpumask_t */
81 int ncpus_fit, ncpus_fit_mask; /* note: mask not cpumask_t */
84 int tsleep_crypto_dump = 0;
86 static struct callout loadav_callout;
87 static struct callout schedcpu_callout;
88 MALLOC_DEFINE(M_TSLEEP, "tslpque", "tsleep queues");
90 #define __DEALL(ident) __DEQUALIFY(void *, ident)
92 #if !defined(KTR_TSLEEP)
93 #define KTR_TSLEEP KTR_ALL
95 KTR_INFO_MASTER(tsleep);
96 KTR_INFO(KTR_TSLEEP, tsleep, tsleep_beg, 0, "tsleep enter %p", sizeof(void *));
97 KTR_INFO(KTR_TSLEEP, tsleep, tsleep_end, 1, "tsleep exit", 0);
98 KTR_INFO(KTR_TSLEEP, tsleep, wakeup_beg, 2, "wakeup enter %p", sizeof(void *));
99 KTR_INFO(KTR_TSLEEP, tsleep, wakeup_end, 3, "wakeup exit", 0);
100 KTR_INFO(KTR_TSLEEP, tsleep, ilockfail, 4, "interlock failed %p", sizeof(void *));
102 #define logtsleep1(name) KTR_LOG(tsleep_ ## name)
103 #define logtsleep2(name, val) KTR_LOG(tsleep_ ## name, val)
105 struct loadavg averunnable =
106 { {0, 0, 0}, FSCALE }; /* load average, of runnable procs */
108 * Constants for averages over 1, 5, and 15 minutes
109 * when sampling at 5 second intervals.
111 static fixpt_t cexp[3] = {
112 0.9200444146293232 * FSCALE, /* exp(-1/12) */
113 0.9834714538216174 * FSCALE, /* exp(-1/60) */
114 0.9944598480048967 * FSCALE, /* exp(-1/180) */
117 static void endtsleep (void *);
118 static void loadav (void *arg);
119 static void schedcpu (void *arg);
121 static void tsleep_wakeup_remote(struct thread *td);
125 * Adjust the scheduler quantum. The quantum is specified in microseconds.
126 * Note that 'tick' is in microseconds per tick.
129 sysctl_kern_quantum(SYSCTL_HANDLER_ARGS)
133 new_val = sched_quantum * ustick;
134 error = sysctl_handle_int(oidp, &new_val, 0, req);
135 if (error != 0 || req->newptr == NULL)
137 if (new_val < ustick)
139 sched_quantum = new_val / ustick;
140 hogticks = 2 * sched_quantum;
144 SYSCTL_PROC(_kern, OID_AUTO, quantum, CTLTYPE_INT|CTLFLAG_RW,
145 0, sizeof sched_quantum, sysctl_kern_quantum, "I", "");
148 * If `ccpu' is not equal to `exp(-1/20)' and you still want to use the
149 * faster/more-accurate formula, you'll have to estimate CCPU_SHIFT below
150 * and possibly adjust FSHIFT in "param.h" so that (FSHIFT >= CCPU_SHIFT).
152 * To estimate CCPU_SHIFT for exp(-1/20), the following formula was used:
153 * 1 - exp(-1/20) ~= 0.0487 ~= 0.0488 == 1 (fixed pt, *11* bits).
155 * If you don't want to bother with the faster/more-accurate formula, you
156 * can set CCPU_SHIFT to (FSHIFT + 1) which will use a slower/less-accurate
157 * (more general) method of calculating the %age of CPU used by a process.
159 * decay 95% of `lwp_pctcpu' in 60 seconds; see CCPU_SHIFT before changing
161 #define CCPU_SHIFT 11
163 static fixpt_t ccpu = 0.95122942450071400909 * FSCALE; /* exp(-1/20) */
164 SYSCTL_INT(_kern, OID_AUTO, ccpu, CTLFLAG_RD, &ccpu, 0, "");
167 * kernel uses `FSCALE', userland (SHOULD) use kern.fscale
169 int fscale __unused = FSCALE; /* exported to systat */
170 SYSCTL_INT(_kern, OID_AUTO, fscale, CTLFLAG_RD, 0, FSCALE, "");
173 * Recompute process priorities, once a second.
175 * Since the userland schedulers are typically event oriented, if the
176 * estcpu calculation at wakeup() time is not sufficient to make a
177 * process runnable relative to other processes in the system we have
178 * a 1-second recalc to help out.
180 * This code also allows us to store sysclock_t data in the process structure
181 * without fear of an overrun, since sysclock_t are guarenteed to hold
182 * several seconds worth of count.
184 * WARNING! callouts can preempt normal threads. However, they will not
185 * preempt a thread holding a spinlock so we *can* safely use spinlocks.
187 static int schedcpu_stats(struct proc *p, void *data __unused);
188 static int schedcpu_resource(struct proc *p, void *data __unused);
193 allproc_scan(schedcpu_stats, NULL);
194 allproc_scan(schedcpu_resource, NULL);
195 wakeup((caddr_t)&lbolt);
196 wakeup((caddr_t)&lbolt_syncer);
197 callout_reset(&schedcpu_callout, hz, schedcpu, NULL);
201 * General process statistics once a second
204 schedcpu_stats(struct proc *p, void *data __unused)
209 * Threads may not be completely set up if process in SIDL state.
211 if (p->p_stat == SIDL)
215 if (lwkt_trytoken(&p->p_token) == FALSE) {
221 FOREACH_LWP_IN_PROC(lp, p) {
222 if (lp->lwp_stat == LSSLEEP)
226 * Only recalculate processes that are active or have slept
227 * less then 2 seconds. The schedulers understand this.
229 if (lp->lwp_slptime <= 1) {
230 p->p_usched->recalculate(lp);
232 lp->lwp_pctcpu = (lp->lwp_pctcpu * ccpu) >> FSHIFT;
235 lwkt_reltoken(&p->p_token);
241 * Resource checks. XXX break out since ksignal/killproc can block,
242 * limiting us to one process killed per second. There is probably
246 schedcpu_resource(struct proc *p, void *data __unused)
251 if (p->p_stat == SIDL)
255 if (lwkt_trytoken(&p->p_token) == FALSE) {
260 if (p->p_stat == SZOMB || p->p_limit == NULL) {
261 lwkt_reltoken(&p->p_token);
267 FOREACH_LWP_IN_PROC(lp, p) {
269 * We may have caught an lp in the middle of being
270 * created, lwp_thread can be NULL.
272 if (lp->lwp_thread) {
273 ttime += lp->lwp_thread->td_sticks;
274 ttime += lp->lwp_thread->td_uticks;
278 switch(plimit_testcpulimit(p->p_limit, ttime)) {
279 case PLIMIT_TESTCPU_KILL:
280 killproc(p, "exceeded maximum CPU limit");
282 case PLIMIT_TESTCPU_XCPU:
283 if ((p->p_flag & P_XCPU) == 0) {
291 lwkt_reltoken(&p->p_token);
297 * This is only used by ps. Generate a cpu percentage use over
298 * a period of one second.
303 updatepcpu(struct lwp *lp, int cpticks, int ttlticks)
308 acc = (cpticks << FSHIFT) / ttlticks;
309 if (ttlticks >= ESTCPUFREQ) {
310 lp->lwp_pctcpu = acc;
312 remticks = ESTCPUFREQ - ttlticks;
313 lp->lwp_pctcpu = (acc * ttlticks + lp->lwp_pctcpu * remticks) /
319 * tsleep/wakeup hash table parameters. Try to find the sweet spot for
320 * like addresses being slept on.
322 #define TABLESIZE 4001
323 #define LOOKUP(x) (((u_int)(uintptr_t)(x)) % TABLESIZE)
325 static cpumask_t slpque_cpumasks[TABLESIZE];
328 * General scheduler initialization. We force a reschedule 25 times
329 * a second by default. Note that cpu0 is initialized in early boot and
330 * cannot make any high level calls.
332 * Each cpu has its own sleep queue.
335 sleep_gdinit(globaldata_t gd)
337 static struct tslpque slpque_cpu0[TABLESIZE];
340 if (gd->gd_cpuid == 0) {
341 sched_quantum = (hz + 24) / 25;
342 hogticks = 2 * sched_quantum;
344 gd->gd_tsleep_hash = slpque_cpu0;
346 gd->gd_tsleep_hash = kmalloc(sizeof(slpque_cpu0),
347 M_TSLEEP, M_WAITOK | M_ZERO);
349 for (i = 0; i < TABLESIZE; ++i)
350 TAILQ_INIT(&gd->gd_tsleep_hash[i]);
354 * This is a dandy function that allows us to interlock tsleep/wakeup
355 * operations with unspecified upper level locks, such as lockmgr locks,
356 * simply by holding a critical section. The sequence is:
358 * (acquire upper level lock)
359 * tsleep_interlock(blah)
360 * (release upper level lock)
363 * Basically this functions queues us on the tsleep queue without actually
364 * descheduling us. When tsleep() is later called with PINTERLOCK it
365 * assumes the thread was already queued, otherwise it queues it there.
367 * Thus it is possible to receive the wakeup prior to going to sleep and
368 * the race conditions are covered.
371 _tsleep_interlock(globaldata_t gd, const volatile void *ident, int flags)
373 thread_t td = gd->gd_curthread;
376 crit_enter_quick(td);
377 if (td->td_flags & TDF_TSLEEPQ) {
378 id = LOOKUP(td->td_wchan);
379 TAILQ_REMOVE(&gd->gd_tsleep_hash[id], td, td_sleepq);
380 if (TAILQ_FIRST(&gd->gd_tsleep_hash[id]) == NULL) {
381 atomic_clear_cpumask(&slpque_cpumasks[id],
385 td->td_flags |= TDF_TSLEEPQ;
388 TAILQ_INSERT_TAIL(&gd->gd_tsleep_hash[id], td, td_sleepq);
389 atomic_set_cpumask(&slpque_cpumasks[id], gd->gd_cpumask);
390 td->td_wchan = ident;
391 td->td_wdomain = flags & PDOMAIN_MASK;
396 tsleep_interlock(const volatile void *ident, int flags)
398 _tsleep_interlock(mycpu, ident, flags);
402 * Remove thread from sleepq. Must be called with a critical section held.
405 _tsleep_remove(thread_t td)
407 globaldata_t gd = mycpu;
410 KKASSERT(td->td_gd == gd);
411 if (td->td_flags & TDF_TSLEEPQ) {
412 td->td_flags &= ~TDF_TSLEEPQ;
413 id = LOOKUP(td->td_wchan);
414 TAILQ_REMOVE(&gd->gd_tsleep_hash[id], td, td_sleepq);
415 if (TAILQ_FIRST(&gd->gd_tsleep_hash[id]) == NULL)
416 atomic_clear_cpumask(&slpque_cpumasks[id], gd->gd_cpumask);
423 tsleep_remove(thread_t td)
429 * This function removes a thread from the tsleep queue and schedules
430 * it. This function may act asynchronously. The target thread may be
431 * sleeping on a different cpu.
433 * This function mus be called while in a critical section but if the
434 * target thread is sleeping on a different cpu we cannot safely probe
437 * This function is only called from a different cpu via setrunnable()
438 * when the thread is in a known sleep. However, multiple wakeups are
439 * possible and we must hold the td to prevent a race against the thread
444 _tsleep_wakeup(struct thread *td)
447 globaldata_t gd = mycpu;
449 if (td->td_gd != gd) {
451 lwkt_send_ipiq(td->td_gd, (ipifunc1_t)tsleep_wakeup_remote, td);
456 if (td->td_flags & TDF_TSLEEP_DESCHEDULED) {
457 td->td_flags &= ~TDF_TSLEEP_DESCHEDULED;
465 tsleep_wakeup_remote(struct thread *td)
474 * General sleep call. Suspends the current process until a wakeup is
475 * performed on the specified identifier. The process will then be made
476 * runnable with the specified priority. Sleeps at most timo/hz seconds
477 * (0 means no timeout). If flags includes PCATCH flag, signals are checked
478 * before and after sleeping, else signals are not checked. Returns 0 if
479 * awakened, EWOULDBLOCK if the timeout expires. If PCATCH is set and a
480 * signal needs to be delivered, ERESTART is returned if the current system
481 * call should be restarted if possible, and EINTR is returned if the system
482 * call should be interrupted by the signal (return EINTR).
484 * Note that if we are a process, we release_curproc() before messing with
485 * the LWKT scheduler.
487 * During autoconfiguration or after a panic, a sleep will simply
488 * lower the priority briefly to allow interrupts, then return.
491 tsleep(const volatile void *ident, int flags, const char *wmesg, int timo)
493 struct thread *td = curthread;
494 struct lwp *lp = td->td_lwp;
495 struct proc *p = td->td_proc; /* may be NULL */
502 struct callout thandle;
505 * NOTE: removed KTRPOINT, it could cause races due to blocking
506 * even in stable. Just scrap it for now.
508 if (!tsleep_crypto_dump && (tsleep_now_works == 0 || panicstr)) {
510 * After a panic, or before we actually have an operational
511 * softclock, just give interrupts a chance, then just return;
513 * don't run any other procs or panic below,
514 * in case this is the idle process and already asleep.
518 lwkt_setpri_self(safepri);
520 lwkt_setpri_self(oldpri);
523 logtsleep2(tsleep_beg, ident);
525 KKASSERT(td != &gd->gd_idlethread); /* you must be kidding! */
528 * NOTE: all of this occurs on the current cpu, including any
529 * callout-based wakeups, so a critical section is a sufficient
532 * The entire sequence through to where we actually sleep must
533 * run without breaking the critical section.
535 catch = flags & PCATCH;
539 crit_enter_quick(td);
541 KASSERT(ident != NULL, ("tsleep: no ident"));
542 KASSERT(lp == NULL ||
543 lp->lwp_stat == LSRUN || /* Obvious */
544 lp->lwp_stat == LSSTOP, /* Set in tstop */
546 ident, wmesg, lp->lwp_stat));
549 * We interlock the sleep queue if the caller has not already done
550 * it for us. This must be done before we potentially acquire any
551 * tokens or we can loose the wakeup.
553 if ((flags & PINTERLOCKED) == 0) {
555 _tsleep_interlock(gd, ident, flags);
559 * Setup for the current process (if this is a process).
561 * We hold the process token if lp && catch. The resume
562 * code will release it.
567 * Early termination if PCATCH was set and a
568 * signal is pending, interlocked with the
571 * Early termination only occurs when tsleep() is
572 * entered while in a normal LSRUN state.
574 lwkt_gettoken(&p->p_token);
575 if ((sig = CURSIG(lp)) != 0)
579 * Early termination if PCATCH was set and a
580 * mailbox signal was possibly delivered prior to
581 * the system call even being made, in order to
582 * allow the user to interlock without having to
583 * make additional system calls.
585 if (p->p_flag & P_MAILBOX)
589 * Causes ksignal to wake us up if a signal is
590 * received (interlocked with p->p_token).
592 lp->lwp_flag |= LWP_SINTR;
599 * Make sure the current process has been untangled from
600 * the userland scheduler and initialize slptime to start
604 p->p_usched->release_curproc(lp);
609 * If the interlocked flag is set but our cpu bit in the slpqueue
610 * is no longer set, then a wakeup was processed inbetween the
611 * tsleep_interlock() (ours or the callers), and here. This can
612 * occur under numerous circumstances including when we release the
615 * Extreme loads can cause the sending of an IPI (e.g. wakeup()'s)
616 * to process incoming IPIs, thus draining incoming wakeups.
618 if ((td->td_flags & TDF_TSLEEPQ) == 0) {
619 logtsleep2(ilockfail, ident);
624 * scheduling is blocked while in a critical section. Coincide
625 * the descheduled-by-tsleep flag with the descheduling of the
628 * The timer callout is localized on our cpu and interlocked by
629 * our critical section.
631 lwkt_deschedule_self(td);
632 td->td_flags |= TDF_TSLEEP_DESCHEDULED;
633 td->td_wmesg = wmesg;
636 * Setup the timeout, if any. The timeout is only operable while
637 * the thread is flagged descheduled.
639 KKASSERT((td->td_flags & TDF_TIMEOUT) == 0);
641 callout_init_mp(&thandle);
642 callout_reset(&thandle, timo, endtsleep, td);
650 * Ok, we are sleeping. Place us in the SSLEEP state.
652 KKASSERT((lp->lwp_flag & LWP_ONRUNQ) == 0);
654 * tstop() sets LSSTOP, so don't fiddle with that.
656 if (lp->lwp_stat != LSSTOP)
657 lp->lwp_stat = LSSLEEP;
658 lp->lwp_ru.ru_nvcsw++;
660 td->td_flags &= ~TDF_TSLEEP_DESCHEDULED;
663 * And when we are woken up, put us back in LSRUN. If we
664 * slept for over a second, recalculate our estcpu.
666 lp->lwp_stat = LSRUN;
668 p->p_usched->recalculate(lp);
672 td->td_flags &= ~TDF_TSLEEP_DESCHEDULED;
676 * Make sure we haven't switched cpus while we were asleep. It's
677 * not supposed to happen. Cleanup our temporary flags.
679 KKASSERT(gd == td->td_gd);
682 * Cleanup the timeout. If the timeout has already occured thandle
683 * has already been stopped, otherwise stop thandle.
686 if (td->td_flags & TDF_TIMEOUT) {
687 td->td_flags &= ~TDF_TIMEOUT;
690 /* does not block when on same cpu */
691 callout_stop(&thandle);
696 * Make sure we have been removed from the sleepq. In most
697 * cases this will have been done for us already but it is
698 * possible for a scheduling IPI to be in-flight from a
699 * previous tsleep/tsleep_interlock() or due to a straight-out
700 * call to lwkt_schedule() (in the case of an interrupt thread),
701 * causing a spurious wakeup.
707 * Figure out the correct error return. If interrupted by a
708 * signal we want to return EINTR or ERESTART.
710 * If P_MAILBOX is set no automatic system call restart occurs
711 * and we return EINTR. P_MAILBOX is meant to be used as an
712 * interlock, the user must poll it prior to any system call
713 * that it wishes to interlock a mailbox signal against since
714 * the flag is cleared on *any* system call that sleeps.
716 * p->p_token is held in the p && catch case.
720 if (catch && error == 0) {
721 if ((p->p_flag & P_MAILBOX) && sig == 0) {
723 } else if (sig != 0 || (sig = CURSIG(lp))) {
724 if (SIGISMEMBER(p->p_sigacts->ps_sigintr, sig))
731 lwkt_reltoken(&p->p_token);
732 lp->lwp_flag &= ~(LWP_BREAKTSLEEP | LWP_SINTR);
733 p->p_flag &= ~P_MAILBOX;
735 logtsleep1(tsleep_end);
741 * Interlocked spinlock sleep. An exclusively held spinlock must
742 * be passed to ssleep(). The function will atomically release the
743 * spinlock and tsleep on the ident, then reacquire the spinlock and
746 * This routine is fairly important along the critical path, so optimize it
750 ssleep(const volatile void *ident, struct spinlock *spin, int flags,
751 const char *wmesg, int timo)
753 globaldata_t gd = mycpu;
756 _tsleep_interlock(gd, ident, flags);
757 spin_unlock_quick(gd, spin);
758 error = tsleep(ident, flags | PINTERLOCKED, wmesg, timo);
759 spin_lock_quick(gd, spin);
765 lksleep(const volatile void *ident, struct lock *lock, int flags,
766 const char *wmesg, int timo)
768 globaldata_t gd = mycpu;
771 _tsleep_interlock(gd, ident, flags);
772 lockmgr(lock, LK_RELEASE);
773 error = tsleep(ident, flags | PINTERLOCKED, wmesg, timo);
774 lockmgr(lock, LK_EXCLUSIVE);
780 * Interlocked mutex sleep. An exclusively held mutex must be passed
781 * to mtxsleep(). The function will atomically release the mutex
782 * and tsleep on the ident, then reacquire the mutex and return.
785 mtxsleep(const volatile void *ident, struct mtx *mtx, int flags,
786 const char *wmesg, int timo)
788 globaldata_t gd = mycpu;
791 _tsleep_interlock(gd, ident, flags);
793 error = tsleep(ident, flags | PINTERLOCKED, wmesg, timo);
794 mtx_lock_ex_quick(mtx, wmesg);
800 * Interlocked serializer sleep. An exclusively held serializer must
801 * be passed to zsleep(). The function will atomically release
802 * the serializer and tsleep on the ident, then reacquire the serializer
806 zsleep(const volatile void *ident, struct lwkt_serialize *slz, int flags,
807 const char *wmesg, int timo)
809 globaldata_t gd = mycpu;
812 ASSERT_SERIALIZED(slz);
814 _tsleep_interlock(gd, ident, flags);
815 lwkt_serialize_exit(slz);
816 ret = tsleep(ident, flags | PINTERLOCKED, wmesg, timo);
817 lwkt_serialize_enter(slz);
823 * Directly block on the LWKT thread by descheduling it. This
824 * is much faster then tsleep(), but the only legal way to wake
825 * us up is to directly schedule the thread.
827 * Setting TDF_SINTR will cause new signals to directly schedule us.
829 * This routine must be called while in a critical section.
832 lwkt_sleep(const char *wmesg, int flags)
834 thread_t td = curthread;
837 if ((flags & PCATCH) == 0 || td->td_lwp == NULL) {
838 td->td_flags |= TDF_BLOCKED;
839 td->td_wmesg = wmesg;
840 lwkt_deschedule_self(td);
843 td->td_flags &= ~TDF_BLOCKED;
846 if ((sig = CURSIG(td->td_lwp)) != 0) {
847 if (SIGISMEMBER(td->td_proc->p_sigacts->ps_sigintr, sig))
853 td->td_flags |= TDF_BLOCKED | TDF_SINTR;
854 td->td_wmesg = wmesg;
855 lwkt_deschedule_self(td);
857 td->td_flags &= ~(TDF_BLOCKED | TDF_SINTR);
863 * Implement the timeout for tsleep.
865 * We set LWP_BREAKTSLEEP to indicate that an event has occured, but
866 * we only call setrunnable if the process is not stopped.
868 * This type of callout timeout is scheduled on the same cpu the process
869 * is sleeping on. Also, at the moment, the MP lock is held.
880 * Do this before we potentially block acquiring the token. Setting
881 * TDF_TIMEOUT tells tsleep that we have already stopped the callout.
884 td->td_flags |= TDF_TIMEOUT;
889 if ((lp = td->td_lwp) != NULL)
890 lwkt_gettoken(&lp->lwp_proc->p_token);
893 * Only do nominal wakeup processing if TDF_TIMEOUT and
894 * TDF_TSLEEP_DESCHEDULED are both still set. Otherwise
895 * we raced a wakeup or we began executed and raced due to
896 * blocking in the token above, and should do nothing.
898 if ((td->td_flags & (TDF_TIMEOUT | TDF_TSLEEP_DESCHEDULED)) ==
899 (TDF_TIMEOUT | TDF_TSLEEP_DESCHEDULED)) {
901 lp->lwp_flag |= LWP_BREAKTSLEEP;
902 if (lp->lwp_proc->p_stat != SSTOP)
909 lwkt_reltoken(&lp->lwp_proc->p_token);
915 * Make all processes sleeping on the specified identifier runnable.
916 * count may be zero or one only.
918 * The domain encodes the sleep/wakeup domain AND the first cpu to check
919 * (which is always the current cpu). As we iterate across cpus
921 * This call may run without the MP lock held. We can only manipulate thread
922 * state on the cpu owning the thread. We CANNOT manipulate process state
925 * _wakeup() can be passed to an IPI so we can't use (const volatile
929 _wakeup(void *ident, int domain)
941 logtsleep2(wakeup_beg, ident);
944 qp = &gd->gd_tsleep_hash[id];
946 for (td = TAILQ_FIRST(qp); td != NULL; td = ntd) {
947 ntd = TAILQ_NEXT(td, td_sleepq);
948 if (td->td_wchan == ident &&
949 td->td_wdomain == (domain & PDOMAIN_MASK)
951 KKASSERT(td->td_gd == gd);
953 if (td->td_flags & TDF_TSLEEP_DESCHEDULED) {
954 td->td_flags &= ~TDF_TSLEEP_DESCHEDULED;
956 if (domain & PWAKEUP_ONE)
965 * We finished checking the current cpu but there still may be
966 * more work to do. Either wakeup_one was requested and no matching
967 * thread was found, or a normal wakeup was requested and we have
968 * to continue checking cpus.
970 * It should be noted that this scheme is actually less expensive then
971 * the old scheme when waking up multiple threads, since we send
972 * only one IPI message per target candidate which may then schedule
973 * multiple threads. Before we could have wound up sending an IPI
974 * message for each thread on the target cpu (!= current cpu) that
975 * needed to be woken up.
977 * NOTE: Wakeups occuring on remote cpus are asynchronous. This
978 * should be ok since we are passing idents in the IPI rather then
981 if ((domain & PWAKEUP_MYCPU) == 0 &&
982 (mask = slpque_cpumasks[id] & gd->gd_other_cpus) != 0) {
983 lwkt_send_ipiq2_mask(mask, _wakeup, ident,
984 domain | PWAKEUP_MYCPU);
988 logtsleep1(wakeup_end);
993 * Wakeup all threads tsleep()ing on the specified ident, on all cpus
996 wakeup(const volatile void *ident)
998 _wakeup(__DEALL(ident), PWAKEUP_ENCODE(0, mycpu->gd_cpuid));
1002 * Wakeup one thread tsleep()ing on the specified ident, on any cpu.
1005 wakeup_one(const volatile void *ident)
1007 /* XXX potentially round-robin the first responding cpu */
1008 _wakeup(__DEALL(ident), PWAKEUP_ENCODE(0, mycpu->gd_cpuid) | PWAKEUP_ONE);
1012 * Wakeup threads tsleep()ing on the specified ident on the current cpu
1016 wakeup_mycpu(const volatile void *ident)
1018 _wakeup(__DEALL(ident), PWAKEUP_MYCPU);
1022 * Wakeup one thread tsleep()ing on the specified ident on the current cpu
1026 wakeup_mycpu_one(const volatile void *ident)
1028 /* XXX potentially round-robin the first responding cpu */
1029 _wakeup(__DEALL(ident), PWAKEUP_MYCPU|PWAKEUP_ONE);
1033 * Wakeup all thread tsleep()ing on the specified ident on the specified cpu
1037 wakeup_oncpu(globaldata_t gd, const volatile void *ident)
1041 _wakeup(__DEALL(ident), PWAKEUP_MYCPU);
1043 lwkt_send_ipiq2(gd, _wakeup, __DEALL(ident), PWAKEUP_MYCPU);
1046 _wakeup(__DEALL(ident), PWAKEUP_MYCPU);
1051 * Wakeup one thread tsleep()ing on the specified ident on the specified cpu
1055 wakeup_oncpu_one(globaldata_t gd, const volatile void *ident)
1059 _wakeup(__DEALL(ident), PWAKEUP_MYCPU | PWAKEUP_ONE);
1061 lwkt_send_ipiq2(gd, _wakeup, __DEALL(ident),
1062 PWAKEUP_MYCPU | PWAKEUP_ONE);
1065 _wakeup(__DEALL(ident), PWAKEUP_MYCPU | PWAKEUP_ONE);
1070 * Wakeup all threads waiting on the specified ident that slept using
1071 * the specified domain, on all cpus.
1074 wakeup_domain(const volatile void *ident, int domain)
1076 _wakeup(__DEALL(ident), PWAKEUP_ENCODE(domain, mycpu->gd_cpuid));
1080 * Wakeup one thread waiting on the specified ident that slept using
1081 * the specified domain, on any cpu.
1084 wakeup_domain_one(const volatile void *ident, int domain)
1086 /* XXX potentially round-robin the first responding cpu */
1087 _wakeup(__DEALL(ident),
1088 PWAKEUP_ENCODE(domain, mycpu->gd_cpuid) | PWAKEUP_ONE);
1094 * Make a process runnable. lp->lwp_proc->p_token must be held on call.
1095 * This only has an effect if we are in SSLEEP. We only break out of the
1096 * tsleep if LWP_BREAKTSLEEP is set, otherwise we just fix-up the state.
1098 * NOTE: With p_token held we can only safely manipulate the process
1099 * structure and the lp's lwp_stat.
1102 setrunnable(struct lwp *lp)
1104 ASSERT_LWKT_TOKEN_HELD(&lp->lwp_proc->p_token);
1106 if (lp->lwp_stat == LSSTOP)
1107 lp->lwp_stat = LSSLEEP;
1108 if (lp->lwp_stat == LSSLEEP && (lp->lwp_flag & LWP_BREAKTSLEEP))
1109 _tsleep_wakeup(lp->lwp_thread);
1114 * The process is stopped due to some condition, usually because p_stat is
1115 * set to SSTOP, but also possibly due to being traced.
1117 * NOTE! If the caller sets SSTOP, the caller must also clear P_WAITED
1118 * because the parent may check the child's status before the child actually
1119 * gets to this routine.
1121 * This routine is called with the current lwp only, typically just
1122 * before returning to userland.
1124 * Setting LWP_BREAKTSLEEP before entering the tsleep will cause a passive
1125 * SIGCONT to break out of the tsleep.
1130 struct lwp *lp = curthread->td_lwp;
1131 struct proc *p = lp->lwp_proc;
1136 * If LWP_WSTOP is set, we were sleeping
1137 * while our process was stopped. At this point
1138 * we were already counted as stopped.
1140 if ((lp->lwp_flag & LWP_WSTOP) == 0) {
1142 * If we're the last thread to stop, signal
1146 lp->lwp_flag |= LWP_WSTOP;
1147 wakeup(&p->p_nstopped);
1148 if (p->p_nstopped == p->p_nthreads) {
1150 * Token required to interlock kern_wait()
1154 lwkt_gettoken(&q->p_token);
1155 p->p_flag &= ~P_WAITED;
1157 if ((q->p_sigacts->ps_flag & PS_NOCLDSTOP) == 0)
1158 ksignal(q, SIGCHLD);
1159 lwkt_reltoken(&q->p_token);
1163 while (p->p_stat == SSTOP) {
1164 lp->lwp_flag |= LWP_BREAKTSLEEP;
1165 lp->lwp_stat = LSSTOP;
1166 tsleep(p, 0, "stop", 0);
1169 lp->lwp_flag &= ~LWP_WSTOP;
1174 * Compute a tenex style load average of a quantity on
1175 * 1, 5 and 15 minute intervals.
1177 static int loadav_count_runnable(struct lwp *p, void *data);
1182 struct loadavg *avg;
1186 alllwp_scan(loadav_count_runnable, &nrun);
1188 for (i = 0; i < 3; i++) {
1189 avg->ldavg[i] = (cexp[i] * avg->ldavg[i] +
1190 nrun * FSCALE * (FSCALE - cexp[i])) >> FSHIFT;
1194 * Schedule the next update to occur after 5 seconds, but add a
1195 * random variation to avoid synchronisation with processes that
1196 * run at regular intervals.
1198 callout_reset(&loadav_callout, hz * 4 + (int)(krandom() % (hz * 2 + 1)),
1203 loadav_count_runnable(struct lwp *lp, void *data)
1208 switch (lp->lwp_stat) {
1210 if ((td = lp->lwp_thread) == NULL)
1212 if (td->td_flags & TDF_BLOCKED)
1224 sched_setup(void *dummy)
1226 callout_init_mp(&loadav_callout);
1227 callout_init_mp(&schedcpu_callout);
1229 /* Kick off timeout driven events by calling first time. */