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);
242 * Resource checks. XXX break out since ksignal/killproc can block,
243 * limiting us to one process killed per second. There is probably
247 schedcpu_resource(struct proc *p, void *data __unused)
252 if (p->p_stat == SIDL)
256 if (lwkt_trytoken(&p->p_token) == FALSE) {
261 if (p->p_stat == SZOMB || p->p_limit == NULL) {
262 lwkt_reltoken(&p->p_token);
268 FOREACH_LWP_IN_PROC(lp, p) {
270 * We may have caught an lp in the middle of being
271 * created, lwp_thread can be NULL.
273 if (lp->lwp_thread) {
274 ttime += lp->lwp_thread->td_sticks;
275 ttime += lp->lwp_thread->td_uticks;
279 switch(plimit_testcpulimit(p->p_limit, ttime)) {
280 case PLIMIT_TESTCPU_KILL:
281 killproc(p, "exceeded maximum CPU limit");
283 case PLIMIT_TESTCPU_XCPU:
284 if ((p->p_flag & P_XCPU) == 0) {
292 lwkt_reltoken(&p->p_token);
299 * This is only used by ps. Generate a cpu percentage use over
300 * a period of one second.
305 updatepcpu(struct lwp *lp, int cpticks, int ttlticks)
310 acc = (cpticks << FSHIFT) / ttlticks;
311 if (ttlticks >= ESTCPUFREQ) {
312 lp->lwp_pctcpu = acc;
314 remticks = ESTCPUFREQ - ttlticks;
315 lp->lwp_pctcpu = (acc * ttlticks + lp->lwp_pctcpu * remticks) /
321 * tsleep/wakeup hash table parameters. Try to find the sweet spot for
322 * like addresses being slept on.
324 #define TABLESIZE 4001
325 #define LOOKUP(x) (((u_int)(uintptr_t)(x)) % TABLESIZE)
327 static cpumask_t slpque_cpumasks[TABLESIZE];
330 * General scheduler initialization. We force a reschedule 25 times
331 * a second by default. Note that cpu0 is initialized in early boot and
332 * cannot make any high level calls.
334 * Each cpu has its own sleep queue.
337 sleep_gdinit(globaldata_t gd)
339 static struct tslpque slpque_cpu0[TABLESIZE];
342 if (gd->gd_cpuid == 0) {
343 sched_quantum = (hz + 24) / 25;
344 hogticks = 2 * sched_quantum;
346 gd->gd_tsleep_hash = slpque_cpu0;
348 gd->gd_tsleep_hash = kmalloc(sizeof(slpque_cpu0),
349 M_TSLEEP, M_WAITOK | M_ZERO);
351 for (i = 0; i < TABLESIZE; ++i)
352 TAILQ_INIT(&gd->gd_tsleep_hash[i]);
356 * This is a dandy function that allows us to interlock tsleep/wakeup
357 * operations with unspecified upper level locks, such as lockmgr locks,
358 * simply by holding a critical section. The sequence is:
360 * (acquire upper level lock)
361 * tsleep_interlock(blah)
362 * (release upper level lock)
365 * Basically this functions queues us on the tsleep queue without actually
366 * descheduling us. When tsleep() is later called with PINTERLOCK it
367 * assumes the thread was already queued, otherwise it queues it there.
369 * Thus it is possible to receive the wakeup prior to going to sleep and
370 * the race conditions are covered.
373 _tsleep_interlock(globaldata_t gd, const volatile void *ident, int flags)
375 thread_t td = gd->gd_curthread;
378 crit_enter_quick(td);
379 if (td->td_flags & TDF_TSLEEPQ) {
380 id = LOOKUP(td->td_wchan);
381 TAILQ_REMOVE(&gd->gd_tsleep_hash[id], td, td_sleepq);
382 if (TAILQ_FIRST(&gd->gd_tsleep_hash[id]) == NULL) {
383 atomic_clear_cpumask(&slpque_cpumasks[id],
387 td->td_flags |= TDF_TSLEEPQ;
390 TAILQ_INSERT_TAIL(&gd->gd_tsleep_hash[id], td, td_sleepq);
391 atomic_set_cpumask(&slpque_cpumasks[id], gd->gd_cpumask);
392 td->td_wchan = ident;
393 td->td_wdomain = flags & PDOMAIN_MASK;
398 tsleep_interlock(const volatile void *ident, int flags)
400 _tsleep_interlock(mycpu, ident, flags);
404 * Remove thread from sleepq. Must be called with a critical section held.
407 _tsleep_remove(thread_t td)
409 globaldata_t gd = mycpu;
412 KKASSERT(td->td_gd == gd && IN_CRITICAL_SECT(td));
413 if (td->td_flags & TDF_TSLEEPQ) {
414 td->td_flags &= ~TDF_TSLEEPQ;
415 id = LOOKUP(td->td_wchan);
416 TAILQ_REMOVE(&gd->gd_tsleep_hash[id], td, td_sleepq);
417 if (TAILQ_FIRST(&gd->gd_tsleep_hash[id]) == NULL)
418 atomic_clear_cpumask(&slpque_cpumasks[id], gd->gd_cpumask);
425 tsleep_remove(thread_t td)
431 * This function removes a thread from the tsleep queue and schedules
432 * it. This function may act asynchronously. The target thread may be
433 * sleeping on a different cpu.
435 * This function mus be called while in a critical section but if the
436 * target thread is sleeping on a different cpu we cannot safely probe
439 * This function is only called from a different cpu via setrunnable()
440 * when the thread is in a known sleep. However, multiple wakeups are
441 * possible and we must hold the td to prevent a race against the thread
446 _tsleep_wakeup(struct thread *td)
449 globaldata_t gd = mycpu;
451 if (td->td_gd != gd) {
453 lwkt_send_ipiq(td->td_gd, (ipifunc1_t)tsleep_wakeup_remote, td);
458 if (td->td_flags & TDF_TSLEEP_DESCHEDULED) {
459 td->td_flags &= ~TDF_TSLEEP_DESCHEDULED;
467 tsleep_wakeup_remote(struct thread *td)
476 * General sleep call. Suspends the current process until a wakeup is
477 * performed on the specified identifier. The process will then be made
478 * runnable with the specified priority. Sleeps at most timo/hz seconds
479 * (0 means no timeout). If flags includes PCATCH flag, signals are checked
480 * before and after sleeping, else signals are not checked. Returns 0 if
481 * awakened, EWOULDBLOCK if the timeout expires. If PCATCH is set and a
482 * signal needs to be delivered, ERESTART is returned if the current system
483 * call should be restarted if possible, and EINTR is returned if the system
484 * call should be interrupted by the signal (return EINTR).
486 * Note that if we are a process, we release_curproc() before messing with
487 * the LWKT scheduler.
489 * During autoconfiguration or after a panic, a sleep will simply
490 * lower the priority briefly to allow interrupts, then return.
493 tsleep(const volatile void *ident, int flags, const char *wmesg, int timo)
495 struct thread *td = curthread;
496 struct lwp *lp = td->td_lwp;
497 struct proc *p = td->td_proc; /* may be NULL */
504 struct callout thandle;
507 * NOTE: removed KTRPOINT, it could cause races due to blocking
508 * even in stable. Just scrap it for now.
510 if (!tsleep_crypto_dump && (tsleep_now_works == 0 || panicstr)) {
512 * After a panic, or before we actually have an operational
513 * softclock, just give interrupts a chance, then just return;
515 * don't run any other procs or panic below,
516 * in case this is the idle process and already asleep.
520 lwkt_setpri_self(safepri);
522 lwkt_setpri_self(oldpri);
525 logtsleep2(tsleep_beg, ident);
527 KKASSERT(td != &gd->gd_idlethread); /* you must be kidding! */
530 * NOTE: all of this occurs on the current cpu, including any
531 * callout-based wakeups, so a critical section is a sufficient
534 * The entire sequence through to where we actually sleep must
535 * run without breaking the critical section.
537 catch = flags & PCATCH;
541 crit_enter_quick(td);
543 KASSERT(ident != NULL, ("tsleep: no ident"));
544 KASSERT(lp == NULL ||
545 lp->lwp_stat == LSRUN || /* Obvious */
546 lp->lwp_stat == LSSTOP, /* Set in tstop */
548 ident, wmesg, lp->lwp_stat));
551 * We interlock the sleep queue if the caller has not already done
552 * it for us. This must be done before we potentially acquire any
553 * tokens or we can loose the wakeup.
555 if ((flags & PINTERLOCKED) == 0) {
557 _tsleep_interlock(gd, ident, flags);
561 * Setup for the current process (if this is a process).
563 * We hold the process token if lp && catch. The resume
564 * code will release it.
569 * Early termination if PCATCH was set and a
570 * signal is pending, interlocked with the
573 * Early termination only occurs when tsleep() is
574 * entered while in a normal LSRUN state.
576 lwkt_gettoken(&lp->lwp_token);
577 if ((sig = CURSIG(lp)) != 0)
581 * Early termination if PCATCH was set and a
582 * mailbox signal was possibly delivered prior to
583 * the system call even being made, in order to
584 * allow the user to interlock without having to
585 * make additional system calls.
587 if (p->p_flag & P_MAILBOX)
591 * Causes ksignal to wake us up if a signal is
592 * received (interlocked with p->p_token).
594 lp->lwp_flag |= LWP_SINTR;
601 * Make sure the current process has been untangled from
602 * the userland scheduler and initialize slptime to start
606 p->p_usched->release_curproc(lp);
611 * If the interlocked flag is set but our cpu bit in the slpqueue
612 * is no longer set, then a wakeup was processed inbetween the
613 * tsleep_interlock() (ours or the callers), and here. This can
614 * occur under numerous circumstances including when we release the
617 * Extreme loads can cause the sending of an IPI (e.g. wakeup()'s)
618 * to process incoming IPIs, thus draining incoming wakeups.
620 if ((td->td_flags & TDF_TSLEEPQ) == 0) {
621 logtsleep2(ilockfail, ident);
626 * scheduling is blocked while in a critical section. Coincide
627 * the descheduled-by-tsleep flag with the descheduling of the
630 * The timer callout is localized on our cpu and interlocked by
631 * our critical section.
633 lwkt_deschedule_self(td);
634 td->td_flags |= TDF_TSLEEP_DESCHEDULED;
635 td->td_wmesg = wmesg;
638 * Setup the timeout, if any. The timeout is only operable while
639 * the thread is flagged descheduled.
641 KKASSERT((td->td_flags & TDF_TIMEOUT) == 0);
643 callout_init_mp(&thandle);
644 callout_reset(&thandle, timo, endtsleep, td);
652 * Ok, we are sleeping. Place us in the SSLEEP state.
654 KKASSERT((lp->lwp_flag & LWP_ONRUNQ) == 0);
656 * tstop() sets LSSTOP, so don't fiddle with that.
658 if (lp->lwp_stat != LSSTOP)
659 lp->lwp_stat = LSSLEEP;
660 lp->lwp_ru.ru_nvcsw++;
662 td->td_flags &= ~TDF_TSLEEP_DESCHEDULED;
665 * And when we are woken up, put us back in LSRUN. If we
666 * slept for over a second, recalculate our estcpu.
668 lp->lwp_stat = LSRUN;
670 p->p_usched->recalculate(lp);
674 td->td_flags &= ~TDF_TSLEEP_DESCHEDULED;
678 * Make sure we haven't switched cpus while we were asleep. It's
679 * not supposed to happen. Cleanup our temporary flags.
681 KKASSERT(gd == td->td_gd);
684 * Cleanup the timeout. If the timeout has already occured thandle
685 * has already been stopped, otherwise stop thandle.
688 if (td->td_flags & TDF_TIMEOUT) {
689 td->td_flags &= ~TDF_TIMEOUT;
692 /* does not block when on same cpu */
693 callout_stop(&thandle);
698 * Make sure we have been removed from the sleepq. In most
699 * cases this will have been done for us already but it is
700 * possible for a scheduling IPI to be in-flight from a
701 * previous tsleep/tsleep_interlock() or due to a straight-out
702 * call to lwkt_schedule() (in the case of an interrupt thread),
703 * causing a spurious wakeup.
709 * Figure out the correct error return. If interrupted by a
710 * signal we want to return EINTR or ERESTART.
712 * If P_MAILBOX is set no automatic system call restart occurs
713 * and we return EINTR. P_MAILBOX is meant to be used as an
714 * interlock, the user must poll it prior to any system call
715 * that it wishes to interlock a mailbox signal against since
716 * the flag is cleared on *any* system call that sleeps.
718 * lp->lwp_token is held in the lwp && catch case.
722 if (catch && error == 0) {
723 if ((p->p_flag & P_MAILBOX) && sig == 0) {
725 } else if (sig != 0 || (sig = CURSIG(lp))) {
726 if (SIGISMEMBER(p->p_sigacts->ps_sigintr, sig))
732 lp->lwp_flag &= ~(LWP_BREAKTSLEEP | LWP_SINTR);
734 lwkt_reltoken(&lp->lwp_token);
735 if (p->p_flag & P_MAILBOX) {
736 lwkt_gettoken(&p->p_token);
737 p->p_flag &= ~P_MAILBOX;
738 lwkt_reltoken(&p->p_token);
742 logtsleep1(tsleep_end);
748 * Interlocked spinlock sleep. An exclusively held spinlock must
749 * be passed to ssleep(). The function will atomically release the
750 * spinlock and tsleep on the ident, then reacquire the spinlock and
753 * This routine is fairly important along the critical path, so optimize it
757 ssleep(const volatile void *ident, struct spinlock *spin, int flags,
758 const char *wmesg, int timo)
760 globaldata_t gd = mycpu;
763 _tsleep_interlock(gd, ident, flags);
764 spin_unlock_quick(gd, spin);
765 error = tsleep(ident, flags | PINTERLOCKED, wmesg, timo);
766 spin_lock_quick(gd, spin);
772 lksleep(const volatile void *ident, struct lock *lock, int flags,
773 const char *wmesg, int timo)
775 globaldata_t gd = mycpu;
778 _tsleep_interlock(gd, ident, flags);
779 lockmgr(lock, LK_RELEASE);
780 error = tsleep(ident, flags | PINTERLOCKED, wmesg, timo);
781 lockmgr(lock, LK_EXCLUSIVE);
787 * Interlocked mutex sleep. An exclusively held mutex must be passed
788 * to mtxsleep(). The function will atomically release the mutex
789 * and tsleep on the ident, then reacquire the mutex and return.
792 mtxsleep(const volatile void *ident, struct mtx *mtx, int flags,
793 const char *wmesg, int timo)
795 globaldata_t gd = mycpu;
798 _tsleep_interlock(gd, ident, flags);
800 error = tsleep(ident, flags | PINTERLOCKED, wmesg, timo);
801 mtx_lock_ex_quick(mtx, wmesg);
807 * Interlocked serializer sleep. An exclusively held serializer must
808 * be passed to zsleep(). The function will atomically release
809 * the serializer and tsleep on the ident, then reacquire the serializer
813 zsleep(const volatile void *ident, struct lwkt_serialize *slz, int flags,
814 const char *wmesg, int timo)
816 globaldata_t gd = mycpu;
819 ASSERT_SERIALIZED(slz);
821 _tsleep_interlock(gd, ident, flags);
822 lwkt_serialize_exit(slz);
823 ret = tsleep(ident, flags | PINTERLOCKED, wmesg, timo);
824 lwkt_serialize_enter(slz);
830 * Directly block on the LWKT thread by descheduling it. This
831 * is much faster then tsleep(), but the only legal way to wake
832 * us up is to directly schedule the thread.
834 * Setting TDF_SINTR will cause new signals to directly schedule us.
836 * This routine must be called while in a critical section.
839 lwkt_sleep(const char *wmesg, int flags)
841 thread_t td = curthread;
844 if ((flags & PCATCH) == 0 || td->td_lwp == NULL) {
845 td->td_flags |= TDF_BLOCKED;
846 td->td_wmesg = wmesg;
847 lwkt_deschedule_self(td);
850 td->td_flags &= ~TDF_BLOCKED;
853 if ((sig = CURSIG(td->td_lwp)) != 0) {
854 if (SIGISMEMBER(td->td_proc->p_sigacts->ps_sigintr, sig))
860 td->td_flags |= TDF_BLOCKED | TDF_SINTR;
861 td->td_wmesg = wmesg;
862 lwkt_deschedule_self(td);
864 td->td_flags &= ~(TDF_BLOCKED | TDF_SINTR);
870 * Implement the timeout for tsleep.
872 * We set LWP_BREAKTSLEEP to indicate that an event has occured, but
873 * we only call setrunnable if the process is not stopped.
875 * This type of callout timeout is scheduled on the same cpu the process
876 * is sleeping on. Also, at the moment, the MP lock is held.
887 * Do this before we potentially block acquiring the token. Setting
888 * TDF_TIMEOUT tells tsleep that we have already stopped the callout.
891 td->td_flags |= TDF_TIMEOUT;
896 if ((lp = td->td_lwp) != NULL)
897 lwkt_gettoken(&lp->lwp_token);
900 * Only do nominal wakeup processing if TDF_TIMEOUT and
901 * TDF_TSLEEP_DESCHEDULED are both still set. Otherwise
902 * we raced a wakeup or we began executed and raced due to
903 * blocking in the token above, and should do nothing.
905 if ((td->td_flags & (TDF_TIMEOUT | TDF_TSLEEP_DESCHEDULED)) ==
906 (TDF_TIMEOUT | TDF_TSLEEP_DESCHEDULED)) {
908 lp->lwp_flag |= LWP_BREAKTSLEEP;
909 if (lp->lwp_proc->p_stat != SSTOP)
916 lwkt_reltoken(&lp->lwp_token);
922 * Make all processes sleeping on the specified identifier runnable.
923 * count may be zero or one only.
925 * The domain encodes the sleep/wakeup domain AND the first cpu to check
926 * (which is always the current cpu). As we iterate across cpus
928 * This call may run without the MP lock held. We can only manipulate thread
929 * state on the cpu owning the thread. We CANNOT manipulate process state
932 * _wakeup() can be passed to an IPI so we can't use (const volatile
936 _wakeup(void *ident, int domain)
948 logtsleep2(wakeup_beg, ident);
951 qp = &gd->gd_tsleep_hash[id];
953 for (td = TAILQ_FIRST(qp); td != NULL; td = ntd) {
954 ntd = TAILQ_NEXT(td, td_sleepq);
955 if (td->td_wchan == ident &&
956 td->td_wdomain == (domain & PDOMAIN_MASK)
958 KKASSERT(td->td_gd == gd);
960 if (td->td_flags & TDF_TSLEEP_DESCHEDULED) {
961 td->td_flags &= ~TDF_TSLEEP_DESCHEDULED;
963 if (domain & PWAKEUP_ONE)
972 * We finished checking the current cpu but there still may be
973 * more work to do. Either wakeup_one was requested and no matching
974 * thread was found, or a normal wakeup was requested and we have
975 * to continue checking cpus.
977 * It should be noted that this scheme is actually less expensive then
978 * the old scheme when waking up multiple threads, since we send
979 * only one IPI message per target candidate which may then schedule
980 * multiple threads. Before we could have wound up sending an IPI
981 * message for each thread on the target cpu (!= current cpu) that
982 * needed to be woken up.
984 * NOTE: Wakeups occuring on remote cpus are asynchronous. This
985 * should be ok since we are passing idents in the IPI rather then
988 if ((domain & PWAKEUP_MYCPU) == 0 &&
989 (mask = slpque_cpumasks[id] & gd->gd_other_cpus) != 0) {
990 lwkt_send_ipiq2_mask(mask, _wakeup, ident,
991 domain | PWAKEUP_MYCPU);
995 logtsleep1(wakeup_end);
1000 * Wakeup all threads tsleep()ing on the specified ident, on all cpus
1003 wakeup(const volatile void *ident)
1005 _wakeup(__DEALL(ident), PWAKEUP_ENCODE(0, mycpu->gd_cpuid));
1009 * Wakeup one thread tsleep()ing on the specified ident, on any cpu.
1012 wakeup_one(const volatile void *ident)
1014 /* XXX potentially round-robin the first responding cpu */
1015 _wakeup(__DEALL(ident), PWAKEUP_ENCODE(0, mycpu->gd_cpuid) | PWAKEUP_ONE);
1019 * Wakeup threads tsleep()ing on the specified ident on the current cpu
1023 wakeup_mycpu(const volatile void *ident)
1025 _wakeup(__DEALL(ident), PWAKEUP_MYCPU);
1029 * Wakeup one thread tsleep()ing on the specified ident on the current cpu
1033 wakeup_mycpu_one(const volatile void *ident)
1035 /* XXX potentially round-robin the first responding cpu */
1036 _wakeup(__DEALL(ident), PWAKEUP_MYCPU|PWAKEUP_ONE);
1040 * Wakeup all thread tsleep()ing on the specified ident on the specified cpu
1044 wakeup_oncpu(globaldata_t gd, const volatile void *ident)
1048 _wakeup(__DEALL(ident), PWAKEUP_MYCPU);
1050 lwkt_send_ipiq2(gd, _wakeup, __DEALL(ident), PWAKEUP_MYCPU);
1053 _wakeup(__DEALL(ident), PWAKEUP_MYCPU);
1058 * Wakeup one thread tsleep()ing on the specified ident on the specified cpu
1062 wakeup_oncpu_one(globaldata_t gd, const volatile void *ident)
1066 _wakeup(__DEALL(ident), PWAKEUP_MYCPU | PWAKEUP_ONE);
1068 lwkt_send_ipiq2(gd, _wakeup, __DEALL(ident),
1069 PWAKEUP_MYCPU | PWAKEUP_ONE);
1072 _wakeup(__DEALL(ident), PWAKEUP_MYCPU | PWAKEUP_ONE);
1077 * Wakeup all threads waiting on the specified ident that slept using
1078 * the specified domain, on all cpus.
1081 wakeup_domain(const volatile void *ident, int domain)
1083 _wakeup(__DEALL(ident), PWAKEUP_ENCODE(domain, mycpu->gd_cpuid));
1087 * Wakeup one thread waiting on the specified ident that slept using
1088 * the specified domain, on any cpu.
1091 wakeup_domain_one(const volatile void *ident, int domain)
1093 /* XXX potentially round-robin the first responding cpu */
1094 _wakeup(__DEALL(ident),
1095 PWAKEUP_ENCODE(domain, mycpu->gd_cpuid) | PWAKEUP_ONE);
1101 * Make a process runnable. lp->lwp_proc->p_token must be held on call.
1102 * This only has an effect if we are in SSLEEP. We only break out of the
1103 * tsleep if LWP_BREAKTSLEEP is set, otherwise we just fix-up the state.
1105 * NOTE: With p_token held we can only safely manipulate the process
1106 * structure and the lp's lwp_stat.
1109 setrunnable(struct lwp *lp)
1111 ASSERT_LWKT_TOKEN_HELD(&lp->lwp_token);
1113 if (lp->lwp_stat == LSSTOP)
1114 lp->lwp_stat = LSSLEEP;
1115 if (lp->lwp_stat == LSSLEEP && (lp->lwp_flag & LWP_BREAKTSLEEP))
1116 _tsleep_wakeup(lp->lwp_thread);
1121 * The process is stopped due to some condition, usually because p_stat is
1122 * set to SSTOP, but also possibly due to being traced.
1124 * NOTE! If the caller sets SSTOP, the caller must also clear P_WAITED
1125 * because the parent may check the child's status before the child actually
1126 * gets to this routine.
1128 * This routine is called with the current lwp only, typically just
1129 * before returning to userland.
1131 * Setting LWP_BREAKTSLEEP before entering the tsleep will cause a passive
1132 * SIGCONT to break out of the tsleep.
1137 struct lwp *lp = curthread->td_lwp;
1138 struct proc *p = lp->lwp_proc;
1143 * If LWP_WSTOP is set, we were sleeping
1144 * while our process was stopped. At this point
1145 * we were already counted as stopped.
1147 if ((lp->lwp_flag & LWP_WSTOP) == 0) {
1149 * If we're the last thread to stop, signal
1153 lp->lwp_flag |= LWP_WSTOP;
1154 wakeup(&p->p_nstopped);
1155 if (p->p_nstopped == p->p_nthreads) {
1157 * Token required to interlock kern_wait()
1161 lwkt_gettoken(&q->p_token);
1162 p->p_flag &= ~P_WAITED;
1164 if ((q->p_sigacts->ps_flag & PS_NOCLDSTOP) == 0)
1165 ksignal(q, SIGCHLD);
1166 lwkt_reltoken(&q->p_token);
1170 while (p->p_stat == SSTOP) {
1171 lp->lwp_flag |= LWP_BREAKTSLEEP;
1172 lp->lwp_stat = LSSTOP;
1173 tsleep(p, 0, "stop", 0);
1176 lp->lwp_flag &= ~LWP_WSTOP;
1181 * Compute a tenex style load average of a quantity on
1182 * 1, 5 and 15 minute intervals.
1184 static int loadav_count_runnable(struct lwp *p, void *data);
1189 struct loadavg *avg;
1193 alllwp_scan(loadav_count_runnable, &nrun);
1195 for (i = 0; i < 3; i++) {
1196 avg->ldavg[i] = (cexp[i] * avg->ldavg[i] +
1197 nrun * FSCALE * (FSCALE - cexp[i])) >> FSHIFT;
1201 * Schedule the next update to occur after 5 seconds, but add a
1202 * random variation to avoid synchronisation with processes that
1203 * run at regular intervals.
1205 callout_reset(&loadav_callout, hz * 4 + (int)(krandom() % (hz * 2 + 1)),
1210 loadav_count_runnable(struct lwp *lp, void *data)
1215 switch (lp->lwp_stat) {
1217 if ((td = lp->lwp_thread) == NULL)
1219 if (td->td_flags & TDF_BLOCKED)
1232 sched_setup(void *dummy)
1234 callout_init_mp(&loadav_callout);
1235 callout_init_mp(&schedcpu_callout);
1237 /* Kick off timeout driven events by calling first time. */