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 $
42 #include "opt_ktrace.h"
44 #include <sys/param.h>
45 #include <sys/systm.h>
47 #include <sys/kernel.h>
48 #include <sys/signalvar.h>
49 #include <sys/resourcevar.h>
50 #include <sys/vmmeter.h>
51 #include <sys/sysctl.h>
55 #include <sys/ktrace.h>
57 #include <sys/xwait.h>
59 #include <sys/serialize.h>
61 #include <sys/signal2.h>
62 #include <sys/thread2.h>
63 #include <sys/spinlock2.h>
64 #include <sys/mutex2.h>
66 #include <machine/cpu.h>
67 #include <machine/smp.h>
69 TAILQ_HEAD(tslpque, thread);
71 static void sched_setup (void *dummy);
72 SYSINIT(sched_setup, SI_SUB_KICK_SCHEDULER, SI_ORDER_FIRST, sched_setup, NULL)
77 int sched_quantum; /* Roundrobin scheduling quantum in ticks. */
79 int ncpus2, ncpus2_shift, ncpus2_mask; /* note: mask not cpumask_t */
80 int ncpus_fit, ncpus_fit_mask; /* note: mask not cpumask_t */
83 int tsleep_crypto_dump = 0;
85 static struct callout loadav_callout;
86 static struct callout schedcpu_callout;
87 MALLOC_DEFINE(M_TSLEEP, "tslpque", "tsleep queues");
89 #define __DEALL(ident) __DEQUALIFY(void *, ident)
91 #if !defined(KTR_TSLEEP)
92 #define KTR_TSLEEP KTR_ALL
94 KTR_INFO_MASTER(tsleep);
95 KTR_INFO(KTR_TSLEEP, tsleep, tsleep_beg, 0, "tsleep enter %p", const volatile void *ident);
96 KTR_INFO(KTR_TSLEEP, tsleep, tsleep_end, 1, "tsleep exit");
97 KTR_INFO(KTR_TSLEEP, tsleep, wakeup_beg, 2, "wakeup enter %p", const volatile void *ident);
98 KTR_INFO(KTR_TSLEEP, tsleep, wakeup_end, 3, "wakeup exit");
99 KTR_INFO(KTR_TSLEEP, tsleep, ilockfail, 4, "interlock failed %p", const volatile void *ident);
101 #define logtsleep1(name) KTR_LOG(tsleep_ ## name)
102 #define logtsleep2(name, val) KTR_LOG(tsleep_ ## name, val)
104 struct loadavg averunnable =
105 { {0, 0, 0}, FSCALE }; /* load average, of runnable procs */
107 * Constants for averages over 1, 5, and 15 minutes
108 * when sampling at 5 second intervals.
110 static fixpt_t cexp[3] = {
111 0.9200444146293232 * FSCALE, /* exp(-1/12) */
112 0.9834714538216174 * FSCALE, /* exp(-1/60) */
113 0.9944598480048967 * FSCALE, /* exp(-1/180) */
116 static void endtsleep (void *);
117 static void loadav (void *arg);
118 static void schedcpu (void *arg);
121 * Adjust the scheduler quantum. The quantum is specified in microseconds.
122 * Note that 'tick' is in microseconds per tick.
125 sysctl_kern_quantum(SYSCTL_HANDLER_ARGS)
129 new_val = sched_quantum * ustick;
130 error = sysctl_handle_int(oidp, &new_val, 0, req);
131 if (error != 0 || req->newptr == NULL)
133 if (new_val < ustick)
135 sched_quantum = new_val / ustick;
136 hogticks = 2 * sched_quantum;
140 SYSCTL_PROC(_kern, OID_AUTO, quantum, CTLTYPE_INT|CTLFLAG_RW,
141 0, sizeof sched_quantum, sysctl_kern_quantum, "I", "");
143 static int pctcpu_decay = 10;
144 SYSCTL_INT(_kern, OID_AUTO, pctcpu_decay, CTLFLAG_RW, &pctcpu_decay, 0, "");
147 * kernel uses `FSCALE', userland (SHOULD) use kern.fscale
149 int fscale __unused = FSCALE; /* exported to systat */
150 SYSCTL_INT(_kern, OID_AUTO, fscale, CTLFLAG_RD, 0, FSCALE, "");
153 * Recompute process priorities, once a second.
155 * Since the userland schedulers are typically event oriented, if the
156 * estcpu calculation at wakeup() time is not sufficient to make a
157 * process runnable relative to other processes in the system we have
158 * a 1-second recalc to help out.
160 * This code also allows us to store sysclock_t data in the process structure
161 * without fear of an overrun, since sysclock_t are guarenteed to hold
162 * several seconds worth of count.
164 * WARNING! callouts can preempt normal threads. However, they will not
165 * preempt a thread holding a spinlock so we *can* safely use spinlocks.
167 static int schedcpu_stats(struct proc *p, void *data __unused);
168 static int schedcpu_resource(struct proc *p, void *data __unused);
173 allproc_scan(schedcpu_stats, NULL);
174 allproc_scan(schedcpu_resource, NULL);
175 wakeup((caddr_t)&lbolt);
176 wakeup(lbolt_syncer);
177 callout_reset(&schedcpu_callout, hz, schedcpu, NULL);
181 * General process statistics once a second
184 schedcpu_stats(struct proc *p, void *data __unused)
189 * Threads may not be completely set up if process in SIDL state.
191 if (p->p_stat == SIDL)
195 if (lwkt_trytoken(&p->p_token) == FALSE) {
201 FOREACH_LWP_IN_PROC(lp, p) {
202 if (lp->lwp_stat == LSSLEEP) {
204 if (lp->lwp_slptime == 1)
205 p->p_usched->uload_update(lp);
209 * Only recalculate processes that are active or have slept
210 * less then 2 seconds. The schedulers understand this.
211 * Otherwise decay by 50% per second.
213 if (lp->lwp_slptime <= 1) {
214 p->p_usched->recalculate(lp);
218 decay = pctcpu_decay;
224 lp->lwp_pctcpu = (lp->lwp_pctcpu * (decay - 1)) / decay;
227 lwkt_reltoken(&p->p_token);
234 * Resource checks. XXX break out since ksignal/killproc can block,
235 * limiting us to one process killed per second. There is probably
239 schedcpu_resource(struct proc *p, void *data __unused)
244 if (p->p_stat == SIDL)
248 if (lwkt_trytoken(&p->p_token) == FALSE) {
253 if (p->p_stat == SZOMB || p->p_limit == NULL) {
254 lwkt_reltoken(&p->p_token);
260 FOREACH_LWP_IN_PROC(lp, p) {
262 * We may have caught an lp in the middle of being
263 * created, lwp_thread can be NULL.
265 if (lp->lwp_thread) {
266 ttime += lp->lwp_thread->td_sticks;
267 ttime += lp->lwp_thread->td_uticks;
271 switch(plimit_testcpulimit(p->p_limit, ttime)) {
272 case PLIMIT_TESTCPU_KILL:
273 killproc(p, "exceeded maximum CPU limit");
275 case PLIMIT_TESTCPU_XCPU:
276 if ((p->p_flags & P_XCPU) == 0) {
277 p->p_flags |= P_XCPU;
284 lwkt_reltoken(&p->p_token);
291 * This is only used by ps. Generate a cpu percentage use over
292 * a period of one second.
295 updatepcpu(struct lwp *lp, int cpticks, int ttlticks)
300 acc = (cpticks << FSHIFT) / ttlticks;
301 if (ttlticks >= ESTCPUFREQ) {
302 lp->lwp_pctcpu = acc;
304 remticks = ESTCPUFREQ - ttlticks;
305 lp->lwp_pctcpu = (acc * ttlticks + lp->lwp_pctcpu * remticks) /
311 * tsleep/wakeup hash table parameters. Try to find the sweet spot for
312 * like addresses being slept on.
314 #define TABLESIZE 4001
315 #define LOOKUP(x) (((u_int)(uintptr_t)(x)) % TABLESIZE)
317 static cpumask_t slpque_cpumasks[TABLESIZE];
320 * General scheduler initialization. We force a reschedule 25 times
321 * a second by default. Note that cpu0 is initialized in early boot and
322 * cannot make any high level calls.
324 * Each cpu has its own sleep queue.
327 sleep_gdinit(globaldata_t gd)
329 static struct tslpque slpque_cpu0[TABLESIZE];
332 if (gd->gd_cpuid == 0) {
333 sched_quantum = (hz + 24) / 25;
334 hogticks = 2 * sched_quantum;
336 gd->gd_tsleep_hash = slpque_cpu0;
338 gd->gd_tsleep_hash = kmalloc(sizeof(slpque_cpu0),
339 M_TSLEEP, M_WAITOK | M_ZERO);
341 for (i = 0; i < TABLESIZE; ++i)
342 TAILQ_INIT(&gd->gd_tsleep_hash[i]);
346 * This is a dandy function that allows us to interlock tsleep/wakeup
347 * operations with unspecified upper level locks, such as lockmgr locks,
348 * simply by holding a critical section. The sequence is:
350 * (acquire upper level lock)
351 * tsleep_interlock(blah)
352 * (release upper level lock)
355 * Basically this functions queues us on the tsleep queue without actually
356 * descheduling us. When tsleep() is later called with PINTERLOCK it
357 * assumes the thread was already queued, otherwise it queues it there.
359 * Thus it is possible to receive the wakeup prior to going to sleep and
360 * the race conditions are covered.
363 _tsleep_interlock(globaldata_t gd, const volatile void *ident, int flags)
365 thread_t td = gd->gd_curthread;
368 crit_enter_quick(td);
369 if (td->td_flags & TDF_TSLEEPQ) {
370 id = LOOKUP(td->td_wchan);
371 TAILQ_REMOVE(&gd->gd_tsleep_hash[id], td, td_sleepq);
372 if (TAILQ_FIRST(&gd->gd_tsleep_hash[id]) == NULL) {
373 atomic_clear_cpumask(&slpque_cpumasks[id],
377 td->td_flags |= TDF_TSLEEPQ;
380 TAILQ_INSERT_TAIL(&gd->gd_tsleep_hash[id], td, td_sleepq);
381 atomic_set_cpumask(&slpque_cpumasks[id], gd->gd_cpumask);
382 td->td_wchan = ident;
383 td->td_wdomain = flags & PDOMAIN_MASK;
388 tsleep_interlock(const volatile void *ident, int flags)
390 _tsleep_interlock(mycpu, ident, flags);
394 * Remove thread from sleepq. Must be called with a critical section held.
395 * The thread must not be migrating.
398 _tsleep_remove(thread_t td)
400 globaldata_t gd = mycpu;
403 KKASSERT(td->td_gd == gd && IN_CRITICAL_SECT(td));
404 KKASSERT((td->td_flags & TDF_MIGRATING) == 0);
405 if (td->td_flags & TDF_TSLEEPQ) {
406 td->td_flags &= ~TDF_TSLEEPQ;
407 id = LOOKUP(td->td_wchan);
408 TAILQ_REMOVE(&gd->gd_tsleep_hash[id], td, td_sleepq);
409 if (TAILQ_FIRST(&gd->gd_tsleep_hash[id]) == NULL)
410 atomic_clear_cpumask(&slpque_cpumasks[id], gd->gd_cpumask);
417 tsleep_remove(thread_t td)
423 * General sleep call. Suspends the current process until a wakeup is
424 * performed on the specified identifier. The process will then be made
425 * runnable with the specified priority. Sleeps at most timo/hz seconds
426 * (0 means no timeout). If flags includes PCATCH flag, signals are checked
427 * before and after sleeping, else signals are not checked. Returns 0 if
428 * awakened, EWOULDBLOCK if the timeout expires. If PCATCH is set and a
429 * signal needs to be delivered, ERESTART is returned if the current system
430 * call should be restarted if possible, and EINTR is returned if the system
431 * call should be interrupted by the signal (return EINTR).
433 * Note that if we are a process, we release_curproc() before messing with
434 * the LWKT scheduler.
436 * During autoconfiguration or after a panic, a sleep will simply
437 * lower the priority briefly to allow interrupts, then return.
439 * WARNING! This code can't block (short of switching away), or bad things
440 * will happen. No getting tokens, no blocking locks, etc.
443 tsleep(const volatile void *ident, int flags, const char *wmesg, int timo)
445 struct thread *td = curthread;
446 struct lwp *lp = td->td_lwp;
447 struct proc *p = td->td_proc; /* may be NULL */
453 struct callout thandle;
456 * Currently a severe hack. Make sure any delayed wakeups
457 * are flushed before we sleep or we might deadlock on whatever
458 * event we are sleeping on.
460 if (td->td_flags & TDF_DELAYED_WAKEUP)
461 wakeup_end_delayed();
464 * NOTE: removed KTRPOINT, it could cause races due to blocking
465 * even in stable. Just scrap it for now.
467 if (!tsleep_crypto_dump && (tsleep_now_works == 0 || panicstr)) {
469 * After a panic, or before we actually have an operational
470 * softclock, just give interrupts a chance, then just return;
472 * don't run any other procs or panic below,
473 * in case this is the idle process and already asleep.
477 lwkt_setpri_self(safepri);
479 lwkt_setpri_self(oldpri);
482 logtsleep2(tsleep_beg, ident);
484 KKASSERT(td != &gd->gd_idlethread); /* you must be kidding! */
485 td->td_wakefromcpu = -1; /* overwritten by _wakeup */
488 * NOTE: all of this occurs on the current cpu, including any
489 * callout-based wakeups, so a critical section is a sufficient
492 * The entire sequence through to where we actually sleep must
493 * run without breaking the critical section.
495 catch = flags & PCATCH;
499 crit_enter_quick(td);
501 KASSERT(ident != NULL, ("tsleep: no ident"));
502 KASSERT(lp == NULL ||
503 lp->lwp_stat == LSRUN || /* Obvious */
504 lp->lwp_stat == LSSTOP, /* Set in tstop */
506 ident, wmesg, lp->lwp_stat));
509 * We interlock the sleep queue if the caller has not already done
510 * it for us. This must be done before we potentially acquire any
511 * tokens or we can loose the wakeup.
513 if ((flags & PINTERLOCKED) == 0) {
514 _tsleep_interlock(gd, ident, flags);
518 * Setup for the current process (if this is a process). We must
519 * interlock with lwp_token to avoid remote wakeup races via
523 lwkt_gettoken(&lp->lwp_token);
526 * Early termination if PCATCH was set and a
527 * signal is pending, interlocked with the
530 * Early termination only occurs when tsleep() is
531 * entered while in a normal LSRUN state.
533 if ((sig = CURSIG(lp)) != 0)
537 * Causes ksignal to wake us up if a signal is
538 * received (interlocked with p->p_token).
540 lp->lwp_flags |= LWP_SINTR;
547 * Make sure the current process has been untangled from
548 * the userland scheduler and initialize slptime to start
551 * NOTE: td->td_wakefromcpu is pre-set by the release function
552 * for the dfly scheduler, and then adjusted by _wakeup()
555 p->p_usched->release_curproc(lp);
560 * If the interlocked flag is set but our cpu bit in the slpqueue
561 * is no longer set, then a wakeup was processed inbetween the
562 * tsleep_interlock() (ours or the callers), and here. This can
563 * occur under numerous circumstances including when we release the
566 * Extreme loads can cause the sending of an IPI (e.g. wakeup()'s)
567 * to process incoming IPIs, thus draining incoming wakeups.
569 if ((td->td_flags & TDF_TSLEEPQ) == 0) {
570 logtsleep2(ilockfail, ident);
575 * scheduling is blocked while in a critical section. Coincide
576 * the descheduled-by-tsleep flag with the descheduling of the
579 * The timer callout is localized on our cpu and interlocked by
580 * our critical section.
582 lwkt_deschedule_self(td);
583 td->td_flags |= TDF_TSLEEP_DESCHEDULED;
584 td->td_wmesg = wmesg;
587 * Setup the timeout, if any. The timeout is only operable while
588 * the thread is flagged descheduled.
590 KKASSERT((td->td_flags & TDF_TIMEOUT) == 0);
592 callout_init_mp(&thandle);
593 callout_reset(&thandle, timo, endtsleep, td);
601 * Ok, we are sleeping. Place us in the SSLEEP state.
603 KKASSERT((lp->lwp_mpflags & LWP_MP_ONRUNQ) == 0);
606 * tstop() sets LSSTOP, so don't fiddle with that.
608 if (lp->lwp_stat != LSSTOP)
609 lp->lwp_stat = LSSLEEP;
610 lp->lwp_ru.ru_nvcsw++;
611 p->p_usched->uload_update(lp);
615 * And when we are woken up, put us back in LSRUN. If we
616 * slept for over a second, recalculate our estcpu.
618 lp->lwp_stat = LSRUN;
619 if (lp->lwp_slptime) {
620 p->p_usched->uload_update(lp);
621 p->p_usched->recalculate(lp);
629 * Make sure we haven't switched cpus while we were asleep. It's
630 * not supposed to happen. Cleanup our temporary flags.
632 KKASSERT(gd == td->td_gd);
635 * Cleanup the timeout. If the timeout has already occured thandle
636 * has already been stopped, otherwise stop thandle. If the timeout
637 * is running (the callout thread must be blocked trying to get
638 * lwp_token) then wait for us to get scheduled.
641 while (td->td_flags & TDF_TIMEOUT_RUNNING) {
642 lwkt_deschedule_self(td);
643 td->td_wmesg = "tsrace";
645 kprintf("td %p %s: timeout race\n", td, td->td_comm);
647 if (td->td_flags & TDF_TIMEOUT) {
648 td->td_flags &= ~TDF_TIMEOUT;
651 /* does not block when on same cpu */
652 callout_stop(&thandle);
655 td->td_flags &= ~TDF_TSLEEP_DESCHEDULED;
658 * Make sure we have been removed from the sleepq. In most
659 * cases this will have been done for us already but it is
660 * possible for a scheduling IPI to be in-flight from a
661 * previous tsleep/tsleep_interlock() or due to a straight-out
662 * call to lwkt_schedule() (in the case of an interrupt thread),
663 * causing a spurious wakeup.
669 * Figure out the correct error return. If interrupted by a
670 * signal we want to return EINTR or ERESTART.
674 if (catch && error == 0) {
675 if (sig != 0 || (sig = CURSIG(lp))) {
676 if (SIGISMEMBER(p->p_sigacts->ps_sigintr, sig))
682 lp->lwp_flags &= ~LWP_SINTR;
683 lwkt_reltoken(&lp->lwp_token);
685 logtsleep1(tsleep_end);
691 * Interlocked spinlock sleep. An exclusively held spinlock must
692 * be passed to ssleep(). The function will atomically release the
693 * spinlock and tsleep on the ident, then reacquire the spinlock and
696 * This routine is fairly important along the critical path, so optimize it
700 ssleep(const volatile void *ident, struct spinlock *spin, int flags,
701 const char *wmesg, int timo)
703 globaldata_t gd = mycpu;
706 _tsleep_interlock(gd, ident, flags);
707 spin_unlock_quick(gd, spin);
708 error = tsleep(ident, flags | PINTERLOCKED, wmesg, timo);
709 spin_lock_quick(gd, spin);
715 lksleep(const volatile void *ident, struct lock *lock, int flags,
716 const char *wmesg, int timo)
718 globaldata_t gd = mycpu;
721 _tsleep_interlock(gd, ident, flags);
722 lockmgr(lock, LK_RELEASE);
723 error = tsleep(ident, flags | PINTERLOCKED, wmesg, timo);
724 lockmgr(lock, LK_EXCLUSIVE);
730 * Interlocked mutex sleep. An exclusively held mutex must be passed
731 * to mtxsleep(). The function will atomically release the mutex
732 * and tsleep on the ident, then reacquire the mutex and return.
735 mtxsleep(const volatile void *ident, struct mtx *mtx, int flags,
736 const char *wmesg, int timo)
738 globaldata_t gd = mycpu;
741 _tsleep_interlock(gd, ident, flags);
743 error = tsleep(ident, flags | PINTERLOCKED, wmesg, timo);
744 mtx_lock_ex_quick(mtx, wmesg);
750 * Interlocked serializer sleep. An exclusively held serializer must
751 * be passed to zsleep(). The function will atomically release
752 * the serializer and tsleep on the ident, then reacquire the serializer
756 zsleep(const volatile void *ident, struct lwkt_serialize *slz, int flags,
757 const char *wmesg, int timo)
759 globaldata_t gd = mycpu;
762 ASSERT_SERIALIZED(slz);
764 _tsleep_interlock(gd, ident, flags);
765 lwkt_serialize_exit(slz);
766 ret = tsleep(ident, flags | PINTERLOCKED, wmesg, timo);
767 lwkt_serialize_enter(slz);
773 * Directly block on the LWKT thread by descheduling it. This
774 * is much faster then tsleep(), but the only legal way to wake
775 * us up is to directly schedule the thread.
777 * Setting TDF_SINTR will cause new signals to directly schedule us.
779 * This routine must be called while in a critical section.
782 lwkt_sleep(const char *wmesg, int flags)
784 thread_t td = curthread;
787 if ((flags & PCATCH) == 0 || td->td_lwp == NULL) {
788 td->td_flags |= TDF_BLOCKED;
789 td->td_wmesg = wmesg;
790 lwkt_deschedule_self(td);
793 td->td_flags &= ~TDF_BLOCKED;
796 if ((sig = CURSIG(td->td_lwp)) != 0) {
797 if (SIGISMEMBER(td->td_proc->p_sigacts->ps_sigintr, sig))
803 td->td_flags |= TDF_BLOCKED | TDF_SINTR;
804 td->td_wmesg = wmesg;
805 lwkt_deschedule_self(td);
807 td->td_flags &= ~(TDF_BLOCKED | TDF_SINTR);
813 * Implement the timeout for tsleep.
815 * This type of callout timeout is scheduled on the same cpu the process
816 * is sleeping on. Also, at the moment, the MP lock is held.
825 * We are going to have to get the lwp_token, which means we might
826 * block. This can race a tsleep getting woken up by other means
827 * so set TDF_TIMEOUT_RUNNING to force the tsleep to wait for our
828 * processing to complete (sorry tsleep!).
830 * We can safely set td_flags because td MUST be on the same cpu
833 KKASSERT(td->td_gd == mycpu);
835 td->td_flags |= TDF_TIMEOUT_RUNNING | TDF_TIMEOUT;
838 * This can block but TDF_TIMEOUT_RUNNING will prevent the thread
839 * from exiting the tsleep on us. The flag is interlocked by virtue
840 * of lp being on the same cpu as we are.
842 if ((lp = td->td_lwp) != NULL)
843 lwkt_gettoken(&lp->lwp_token);
845 KKASSERT(td->td_flags & TDF_TSLEEP_DESCHEDULED);
848 if (lp->lwp_proc->p_stat != SSTOP)
850 lwkt_reltoken(&lp->lwp_token);
855 KKASSERT(td->td_gd == mycpu);
856 td->td_flags &= ~TDF_TIMEOUT_RUNNING;
861 * Make all processes sleeping on the specified identifier runnable.
862 * count may be zero or one only.
864 * The domain encodes the sleep/wakeup domain, flags, plus the originating
867 * This call may run without the MP lock held. We can only manipulate thread
868 * state on the cpu owning the thread. We CANNOT manipulate process state
871 * _wakeup() can be passed to an IPI so we can't use (const volatile
875 _wakeup(void *ident, int domain)
885 logtsleep2(wakeup_beg, ident);
888 qp = &gd->gd_tsleep_hash[id];
890 for (td = TAILQ_FIRST(qp); td != NULL; td = ntd) {
891 ntd = TAILQ_NEXT(td, td_sleepq);
892 if (td->td_wchan == ident &&
893 td->td_wdomain == (domain & PDOMAIN_MASK)
895 KKASSERT(td->td_gd == gd);
897 td->td_wakefromcpu = PWAKEUP_DECODE(domain);
898 if (td->td_flags & TDF_TSLEEP_DESCHEDULED) {
900 if (domain & PWAKEUP_ONE)
908 * We finished checking the current cpu but there still may be
909 * more work to do. Either wakeup_one was requested and no matching
910 * thread was found, or a normal wakeup was requested and we have
911 * to continue checking cpus.
913 * It should be noted that this scheme is actually less expensive then
914 * the old scheme when waking up multiple threads, since we send
915 * only one IPI message per target candidate which may then schedule
916 * multiple threads. Before we could have wound up sending an IPI
917 * message for each thread on the target cpu (!= current cpu) that
918 * needed to be woken up.
920 * NOTE: Wakeups occuring on remote cpus are asynchronous. This
921 * should be ok since we are passing idents in the IPI rather then
924 if ((domain & PWAKEUP_MYCPU) == 0 &&
925 (mask = slpque_cpumasks[id] & gd->gd_other_cpus) != 0) {
926 lwkt_send_ipiq2_mask(mask, _wakeup, ident,
927 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)) {
944 if (!atomic_cmpset_ptr(&gd->gd_delayed_wakeup[0], NULL, ident)) {
945 if (!atomic_cmpset_ptr(&gd->gd_delayed_wakeup[1], NULL, ident))
946 _wakeup(__DEALL(ident), PWAKEUP_ENCODE(0, gd->gd_cpuid));
950 _wakeup(__DEALL(ident), PWAKEUP_ENCODE(0, gd->gd_cpuid));
954 * Wakeup one thread tsleep()ing on the specified ident, on any cpu.
957 wakeup_one(const volatile void *ident)
959 /* XXX potentially round-robin the first responding cpu */
960 _wakeup(__DEALL(ident), PWAKEUP_ENCODE(0, mycpu->gd_cpuid) |
965 * Wakeup threads tsleep()ing on the specified ident on the current cpu
969 wakeup_mycpu(const volatile void *ident)
971 _wakeup(__DEALL(ident), PWAKEUP_ENCODE(0, mycpu->gd_cpuid) |
976 * Wakeup one thread tsleep()ing on the specified ident on the current cpu
980 wakeup_mycpu_one(const volatile void *ident)
982 /* XXX potentially round-robin the first responding cpu */
983 _wakeup(__DEALL(ident), PWAKEUP_ENCODE(0, mycpu->gd_cpuid) |
984 PWAKEUP_MYCPU | PWAKEUP_ONE);
988 * Wakeup all thread tsleep()ing on the specified ident on the specified cpu
992 wakeup_oncpu(globaldata_t gd, const volatile void *ident)
994 globaldata_t mygd = mycpu;
996 _wakeup(__DEALL(ident), PWAKEUP_ENCODE(0, mygd->gd_cpuid) |
999 lwkt_send_ipiq2(gd, _wakeup, __DEALL(ident),
1000 PWAKEUP_ENCODE(0, mygd->gd_cpuid) |
1006 * Wakeup one thread tsleep()ing on the specified ident on the specified cpu
1010 wakeup_oncpu_one(globaldata_t gd, const volatile void *ident)
1012 globaldata_t mygd = mycpu;
1014 _wakeup(__DEALL(ident), PWAKEUP_ENCODE(0, mygd->gd_cpuid) |
1015 PWAKEUP_MYCPU | PWAKEUP_ONE);
1017 lwkt_send_ipiq2(gd, _wakeup, __DEALL(ident),
1018 PWAKEUP_ENCODE(0, mygd->gd_cpuid) |
1019 PWAKEUP_MYCPU | PWAKEUP_ONE);
1024 * Wakeup all threads waiting on the specified ident that slept using
1025 * the specified domain, on all cpus.
1028 wakeup_domain(const volatile void *ident, int domain)
1030 _wakeup(__DEALL(ident), PWAKEUP_ENCODE(domain, mycpu->gd_cpuid));
1034 * Wakeup one thread waiting on the specified ident that slept using
1035 * the specified domain, on any cpu.
1038 wakeup_domain_one(const volatile void *ident, int domain)
1040 /* XXX potentially round-robin the first responding cpu */
1041 _wakeup(__DEALL(ident),
1042 PWAKEUP_ENCODE(domain, mycpu->gd_cpuid) | PWAKEUP_ONE);
1046 wakeup_start_delayed(void)
1048 globaldata_t gd = mycpu;
1051 gd->gd_curthread->td_flags |= TDF_DELAYED_WAKEUP;
1056 wakeup_end_delayed(void)
1058 globaldata_t gd = mycpu;
1060 if (gd->gd_curthread->td_flags & TDF_DELAYED_WAKEUP) {
1062 gd->gd_curthread->td_flags &= ~TDF_DELAYED_WAKEUP;
1063 if (gd->gd_delayed_wakeup[0] || gd->gd_delayed_wakeup[1]) {
1064 if (gd->gd_delayed_wakeup[0]) {
1065 wakeup(gd->gd_delayed_wakeup[0]);
1066 gd->gd_delayed_wakeup[0] = NULL;
1068 if (gd->gd_delayed_wakeup[1]) {
1069 wakeup(gd->gd_delayed_wakeup[1]);
1070 gd->gd_delayed_wakeup[1] = NULL;
1080 * Make a process runnable. lp->lwp_token must be held on call and this
1081 * function must be called from the cpu owning lp.
1083 * This only has an effect if we are in LSSTOP or LSSLEEP.
1086 setrunnable(struct lwp *lp)
1088 thread_t td = lp->lwp_thread;
1090 ASSERT_LWKT_TOKEN_HELD(&lp->lwp_token);
1091 KKASSERT(td->td_gd == mycpu);
1093 if (lp->lwp_stat == LSSTOP)
1094 lp->lwp_stat = LSSLEEP;
1095 if (lp->lwp_stat == LSSLEEP) {
1098 } else if (td->td_flags & TDF_SINTR) {
1105 * The process is stopped due to some condition, usually because p_stat is
1106 * set to SSTOP, but also possibly due to being traced.
1108 * Caller must hold p->p_token
1110 * NOTE! If the caller sets SSTOP, the caller must also clear P_WAITED
1111 * because the parent may check the child's status before the child actually
1112 * gets to this routine.
1114 * This routine is called with the current lwp only, typically just
1115 * before returning to userland if the process state is detected as
1116 * possibly being in a stopped state.
1121 struct lwp *lp = curthread->td_lwp;
1122 struct proc *p = lp->lwp_proc;
1125 lwkt_gettoken(&lp->lwp_token);
1129 * If LWP_MP_WSTOP is set, we were sleeping
1130 * while our process was stopped. At this point
1131 * we were already counted as stopped.
1133 if ((lp->lwp_mpflags & LWP_MP_WSTOP) == 0) {
1135 * If we're the last thread to stop, signal
1139 atomic_set_int(&lp->lwp_mpflags, LWP_MP_WSTOP);
1140 wakeup(&p->p_nstopped);
1141 if (p->p_nstopped == p->p_nthreads) {
1143 * Token required to interlock kern_wait()
1147 lwkt_gettoken(&q->p_token);
1148 p->p_flags &= ~P_WAITED;
1150 if ((q->p_sigacts->ps_flag & PS_NOCLDSTOP) == 0)
1151 ksignal(q, SIGCHLD);
1152 lwkt_reltoken(&q->p_token);
1156 while (p->p_stat == SSTOP) {
1157 lp->lwp_stat = LSSTOP;
1158 tsleep(p, 0, "stop", 0);
1161 atomic_clear_int(&lp->lwp_mpflags, LWP_MP_WSTOP);
1163 lwkt_reltoken(&lp->lwp_token);
1167 * Compute a tenex style load average of a quantity on
1168 * 1, 5 and 15 minute intervals.
1170 static int loadav_count_runnable(struct lwp *p, void *data);
1175 struct loadavg *avg;
1179 alllwp_scan(loadav_count_runnable, &nrun);
1181 for (i = 0; i < 3; i++) {
1182 avg->ldavg[i] = (cexp[i] * avg->ldavg[i] +
1183 nrun * FSCALE * (FSCALE - cexp[i])) >> FSHIFT;
1187 * Schedule the next update to occur after 5 seconds, but add a
1188 * random variation to avoid synchronisation with processes that
1189 * run at regular intervals.
1191 callout_reset(&loadav_callout, hz * 4 + (int)(krandom() % (hz * 2 + 1)),
1196 loadav_count_runnable(struct lwp *lp, void *data)
1201 switch (lp->lwp_stat) {
1203 if ((td = lp->lwp_thread) == NULL)
1205 if (td->td_flags & TDF_BLOCKED)
1218 sched_setup(void *dummy)
1220 callout_init_mp(&loadav_callout);
1221 callout_init_mp(&schedcpu_callout);
1223 /* Kick off timeout driven events by calling first time. */