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>
66 #include <sys/mplock2.h>
68 #include <machine/cpu.h>
69 #include <machine/smp.h>
71 TAILQ_HEAD(tslpque, thread);
73 static void sched_setup (void *dummy);
74 SYSINIT(sched_setup, SI_SUB_KICK_SCHEDULER, SI_ORDER_FIRST, sched_setup, NULL)
79 int sched_quantum; /* Roundrobin scheduling quantum in ticks. */
81 int ncpus2, ncpus2_shift, ncpus2_mask; /* note: mask not cpumask_t */
82 int ncpus_fit, ncpus_fit_mask; /* note: mask not cpumask_t */
85 int tsleep_crypto_dump = 0;
87 static struct callout loadav_callout;
88 static struct callout schedcpu_callout;
89 MALLOC_DEFINE(M_TSLEEP, "tslpque", "tsleep queues");
91 #define __DEALL(ident) __DEQUALIFY(void *, ident)
93 #if !defined(KTR_TSLEEP)
94 #define KTR_TSLEEP KTR_ALL
96 KTR_INFO_MASTER(tsleep);
97 KTR_INFO(KTR_TSLEEP, tsleep, tsleep_beg, 0, "tsleep enter %p", sizeof(void *));
98 KTR_INFO(KTR_TSLEEP, tsleep, tsleep_end, 1, "tsleep exit", 0);
99 KTR_INFO(KTR_TSLEEP, tsleep, wakeup_beg, 2, "wakeup enter %p", sizeof(void *));
100 KTR_INFO(KTR_TSLEEP, tsleep, wakeup_end, 3, "wakeup exit", 0);
101 KTR_INFO(KTR_TSLEEP, tsleep, ilockfail, 4, "interlock failed %p", sizeof(void *));
103 #define logtsleep1(name) KTR_LOG(tsleep_ ## name)
104 #define logtsleep2(name, val) KTR_LOG(tsleep_ ## name, val)
106 struct loadavg averunnable =
107 { {0, 0, 0}, FSCALE }; /* load average, of runnable procs */
109 * Constants for averages over 1, 5, and 15 minutes
110 * when sampling at 5 second intervals.
112 static fixpt_t cexp[3] = {
113 0.9200444146293232 * FSCALE, /* exp(-1/12) */
114 0.9834714538216174 * FSCALE, /* exp(-1/60) */
115 0.9944598480048967 * FSCALE, /* exp(-1/180) */
118 static void endtsleep (void *);
119 static void loadav (void *arg);
120 static void schedcpu (void *arg);
122 static void tsleep_wakeup_remote(struct thread *td);
126 * Adjust the scheduler quantum. The quantum is specified in microseconds.
127 * Note that 'tick' is in microseconds per tick.
130 sysctl_kern_quantum(SYSCTL_HANDLER_ARGS)
134 new_val = sched_quantum * ustick;
135 error = sysctl_handle_int(oidp, &new_val, 0, req);
136 if (error != 0 || req->newptr == NULL)
138 if (new_val < ustick)
140 sched_quantum = new_val / ustick;
141 hogticks = 2 * sched_quantum;
145 SYSCTL_PROC(_kern, OID_AUTO, quantum, CTLTYPE_INT|CTLFLAG_RW,
146 0, sizeof sched_quantum, sysctl_kern_quantum, "I", "");
149 * If `ccpu' is not equal to `exp(-1/20)' and you still want to use the
150 * faster/more-accurate formula, you'll have to estimate CCPU_SHIFT below
151 * and possibly adjust FSHIFT in "param.h" so that (FSHIFT >= CCPU_SHIFT).
153 * To estimate CCPU_SHIFT for exp(-1/20), the following formula was used:
154 * 1 - exp(-1/20) ~= 0.0487 ~= 0.0488 == 1 (fixed pt, *11* bits).
156 * If you don't want to bother with the faster/more-accurate formula, you
157 * can set CCPU_SHIFT to (FSHIFT + 1) which will use a slower/less-accurate
158 * (more general) method of calculating the %age of CPU used by a process.
160 * decay 95% of `lwp_pctcpu' in 60 seconds; see CCPU_SHIFT before changing
162 #define CCPU_SHIFT 11
164 static fixpt_t ccpu = 0.95122942450071400909 * FSCALE; /* exp(-1/20) */
165 SYSCTL_INT(_kern, OID_AUTO, ccpu, CTLFLAG_RD, &ccpu, 0, "");
168 * kernel uses `FSCALE', userland (SHOULD) use kern.fscale
170 int fscale __unused = FSCALE; /* exported to systat */
171 SYSCTL_INT(_kern, OID_AUTO, fscale, CTLFLAG_RD, 0, FSCALE, "");
174 * Recompute process priorities, once a second.
176 * Since the userland schedulers are typically event oriented, if the
177 * estcpu calculation at wakeup() time is not sufficient to make a
178 * process runnable relative to other processes in the system we have
179 * a 1-second recalc to help out.
181 * This code also allows us to store sysclock_t data in the process structure
182 * without fear of an overrun, since sysclock_t are guarenteed to hold
183 * several seconds worth of count.
185 * WARNING! callouts can preempt normal threads. However, they will not
186 * preempt a thread holding a spinlock so we *can* safely use spinlocks.
188 static int schedcpu_stats(struct proc *p, void *data __unused);
189 static int schedcpu_resource(struct proc *p, void *data __unused);
194 allproc_scan(schedcpu_stats, NULL);
195 allproc_scan(schedcpu_resource, NULL);
196 wakeup((caddr_t)&lbolt);
197 wakeup((caddr_t)&lbolt_syncer);
198 callout_reset(&schedcpu_callout, hz, schedcpu, NULL);
202 * General process statistics once a second
205 schedcpu_stats(struct proc *p, void *data __unused)
211 FOREACH_LWP_IN_PROC(lp, p) {
212 if (lp->lwp_stat == LSSLEEP)
216 * Only recalculate processes that are active or have slept
217 * less then 2 seconds. The schedulers understand this.
219 if (lp->lwp_slptime <= 1) {
220 p->p_usched->recalculate(lp);
222 lp->lwp_pctcpu = (lp->lwp_pctcpu * ccpu) >> FSHIFT;
230 * Resource checks. XXX break out since ksignal/killproc can block,
231 * limiting us to one process killed per second. There is probably
235 schedcpu_resource(struct proc *p, void *data __unused)
241 if (p->p_stat == SIDL ||
242 p->p_stat == SZOMB ||
250 FOREACH_LWP_IN_PROC(lp, p) {
252 * We may have caught an lp in the middle of being
253 * created, lwp_thread can be NULL.
255 if (lp->lwp_thread) {
256 ttime += lp->lwp_thread->td_sticks;
257 ttime += lp->lwp_thread->td_uticks;
261 switch(plimit_testcpulimit(p->p_limit, ttime)) {
262 case PLIMIT_TESTCPU_KILL:
263 killproc(p, "exceeded maximum CPU limit");
265 case PLIMIT_TESTCPU_XCPU:
266 if ((p->p_flag & P_XCPU) == 0) {
279 * This is only used by ps. Generate a cpu percentage use over
280 * a period of one second.
285 updatepcpu(struct lwp *lp, int cpticks, int ttlticks)
290 acc = (cpticks << FSHIFT) / ttlticks;
291 if (ttlticks >= ESTCPUFREQ) {
292 lp->lwp_pctcpu = acc;
294 remticks = ESTCPUFREQ - ttlticks;
295 lp->lwp_pctcpu = (acc * ttlticks + lp->lwp_pctcpu * remticks) /
301 * tsleep/wakeup hash table parameters. Try to find the sweet spot for
302 * like addresses being slept on.
304 #define TABLESIZE 1024
305 #define LOOKUP(x) (((intptr_t)(x) >> 6) & (TABLESIZE - 1))
307 static cpumask_t slpque_cpumasks[TABLESIZE];
310 * General scheduler initialization. We force a reschedule 25 times
311 * a second by default. Note that cpu0 is initialized in early boot and
312 * cannot make any high level calls.
314 * Each cpu has its own sleep queue.
317 sleep_gdinit(globaldata_t gd)
319 static struct tslpque slpque_cpu0[TABLESIZE];
322 if (gd->gd_cpuid == 0) {
323 sched_quantum = (hz + 24) / 25;
324 hogticks = 2 * sched_quantum;
326 gd->gd_tsleep_hash = slpque_cpu0;
328 gd->gd_tsleep_hash = kmalloc(sizeof(slpque_cpu0),
329 M_TSLEEP, M_WAITOK | M_ZERO);
331 for (i = 0; i < TABLESIZE; ++i)
332 TAILQ_INIT(&gd->gd_tsleep_hash[i]);
336 * This is a dandy function that allows us to interlock tsleep/wakeup
337 * operations with unspecified upper level locks, such as lockmgr locks,
338 * simply by holding a critical section. The sequence is:
340 * (acquire upper level lock)
341 * tsleep_interlock(blah)
342 * (release upper level lock)
345 * Basically this functions queues us on the tsleep queue without actually
346 * descheduling us. When tsleep() is later called with PINTERLOCK it
347 * assumes the thread was already queued, otherwise it queues it there.
349 * Thus it is possible to receive the wakeup prior to going to sleep and
350 * the race conditions are covered.
353 _tsleep_interlock(globaldata_t gd, const volatile void *ident, int flags)
355 thread_t td = gd->gd_curthread;
358 crit_enter_quick(td);
359 if (td->td_flags & TDF_TSLEEPQ) {
360 id = LOOKUP(td->td_wchan);
361 TAILQ_REMOVE(&gd->gd_tsleep_hash[id], td, td_sleepq);
362 if (TAILQ_FIRST(&gd->gd_tsleep_hash[id]) == NULL)
363 atomic_clear_cpumask(&slpque_cpumasks[id], gd->gd_cpumask);
365 td->td_flags |= TDF_TSLEEPQ;
368 TAILQ_INSERT_TAIL(&gd->gd_tsleep_hash[id], td, td_sleepq);
369 atomic_set_cpumask(&slpque_cpumasks[id], gd->gd_cpumask);
370 td->td_wchan = ident;
371 td->td_wdomain = flags & PDOMAIN_MASK;
376 tsleep_interlock(const volatile void *ident, int flags)
378 _tsleep_interlock(mycpu, ident, flags);
382 * Remove thread from sleepq. Must be called with a critical section held.
385 _tsleep_remove(thread_t td)
387 globaldata_t gd = mycpu;
390 KKASSERT(td->td_gd == gd);
391 if (td->td_flags & TDF_TSLEEPQ) {
392 td->td_flags &= ~TDF_TSLEEPQ;
393 id = LOOKUP(td->td_wchan);
394 TAILQ_REMOVE(&gd->gd_tsleep_hash[id], td, td_sleepq);
395 if (TAILQ_FIRST(&gd->gd_tsleep_hash[id]) == NULL)
396 atomic_clear_cpumask(&slpque_cpumasks[id], gd->gd_cpumask);
403 tsleep_remove(thread_t td)
409 * This function removes a thread from the tsleep queue and schedules
410 * it. This function may act asynchronously. The target thread may be
411 * sleeping on a different cpu.
413 * This function mus be called while in a critical section but if the
414 * target thread is sleeping on a different cpu we cannot safely probe
417 * This function is only called from a different cpu via setrunnable()
418 * when the thread is in a known sleep. However, multiple wakeups are
419 * possible and we must hold the td to prevent a race against the thread
424 _tsleep_wakeup(struct thread *td)
427 globaldata_t gd = mycpu;
429 if (td->td_gd != gd) {
431 lwkt_send_ipiq(td->td_gd, (ipifunc1_t)tsleep_wakeup_remote, td);
436 if (td->td_flags & TDF_TSLEEP_DESCHEDULED) {
437 td->td_flags &= ~TDF_TSLEEP_DESCHEDULED;
445 tsleep_wakeup_remote(struct thread *td)
454 * General sleep call. Suspends the current process until a wakeup is
455 * performed on the specified identifier. The process will then be made
456 * runnable with the specified priority. Sleeps at most timo/hz seconds
457 * (0 means no timeout). If flags includes PCATCH flag, signals are checked
458 * before and after sleeping, else signals are not checked. Returns 0 if
459 * awakened, EWOULDBLOCK if the timeout expires. If PCATCH is set and a
460 * signal needs to be delivered, ERESTART is returned if the current system
461 * call should be restarted if possible, and EINTR is returned if the system
462 * call should be interrupted by the signal (return EINTR).
464 * Note that if we are a process, we release_curproc() before messing with
465 * the LWKT scheduler.
467 * During autoconfiguration or after a panic, a sleep will simply
468 * lower the priority briefly to allow interrupts, then return.
471 tsleep(const volatile void *ident, int flags, const char *wmesg, int timo)
473 struct thread *td = curthread;
474 struct lwp *lp = td->td_lwp;
475 struct proc *p = td->td_proc; /* may be NULL */
482 struct callout thandle;
485 * NOTE: removed KTRPOINT, it could cause races due to blocking
486 * even in stable. Just scrap it for now.
488 if (!tsleep_crypto_dump && (tsleep_now_works == 0 || panicstr)) {
490 * After a panic, or before we actually have an operational
491 * softclock, just give interrupts a chance, then just return;
493 * don't run any other procs or panic below,
494 * in case this is the idle process and already asleep.
498 lwkt_setpri_self(safepri);
500 lwkt_setpri_self(oldpri);
503 logtsleep2(tsleep_beg, ident);
505 KKASSERT(td != &gd->gd_idlethread); /* you must be kidding! */
508 * NOTE: all of this occurs on the current cpu, including any
509 * callout-based wakeups, so a critical section is a sufficient
512 * The entire sequence through to where we actually sleep must
513 * run without breaking the critical section.
515 catch = flags & PCATCH;
519 crit_enter_quick(td);
521 KASSERT(ident != NULL, ("tsleep: no ident"));
522 KASSERT(lp == NULL ||
523 lp->lwp_stat == LSRUN || /* Obvious */
524 lp->lwp_stat == LSSTOP, /* Set in tstop */
526 ident, wmesg, lp->lwp_stat));
529 * We interlock the sleep queue if the caller has not already done
530 * it for us. This must be done before we potentially acquire any
531 * tokens or we can loose the wakeup.
533 if ((flags & PINTERLOCKED) == 0) {
535 _tsleep_interlock(gd, ident, flags);
539 * Setup for the current process (if this is a process).
541 * We hold the process token if lp && catch. The resume
542 * code will release it.
547 * Early termination if PCATCH was set and a
548 * signal is pending, interlocked with the
551 * Early termination only occurs when tsleep() is
552 * entered while in a normal LSRUN state.
554 lwkt_gettoken(&p->p_token);
555 if ((sig = CURSIG(lp)) != 0)
559 * Early termination if PCATCH was set and a
560 * mailbox signal was possibly delivered prior to
561 * the system call even being made, in order to
562 * allow the user to interlock without having to
563 * make additional system calls.
565 if (p->p_flag & P_MAILBOX)
569 * Causes ksignal to wake us up if a signal is
570 * received (interlocked with p->p_token).
572 lp->lwp_flag |= LWP_SINTR;
579 * Make sure the current process has been untangled from
580 * the userland scheduler and initialize slptime to start
584 p->p_usched->release_curproc(lp);
589 * If the interlocked flag is set but our cpu bit in the slpqueue
590 * is no longer set, then a wakeup was processed inbetween the
591 * tsleep_interlock() (ours or the callers), and here. This can
592 * occur under numerous circumstances including when we release the
595 * Extreme loads can cause the sending of an IPI (e.g. wakeup()'s)
596 * to process incoming IPIs, thus draining incoming wakeups.
598 if ((td->td_flags & TDF_TSLEEPQ) == 0) {
599 logtsleep2(ilockfail, ident);
604 * scheduling is blocked while in a critical section. Coincide
605 * the descheduled-by-tsleep flag with the descheduling of the
608 lwkt_deschedule_self(td);
609 td->td_flags |= TDF_TSLEEP_DESCHEDULED;
610 td->td_wmesg = wmesg;
613 * Setup the timeout, if any
616 callout_init(&thandle);
617 callout_reset(&thandle, timo, endtsleep, td);
625 * Ok, we are sleeping. Place us in the SSLEEP state.
627 KKASSERT((lp->lwp_flag & LWP_ONRUNQ) == 0);
629 * tstop() sets LSSTOP, so don't fiddle with that.
631 if (lp->lwp_stat != LSSTOP)
632 lp->lwp_stat = LSSLEEP;
633 lp->lwp_ru.ru_nvcsw++;
637 * And when we are woken up, put us back in LSRUN. If we
638 * slept for over a second, recalculate our estcpu.
640 lp->lwp_stat = LSRUN;
642 p->p_usched->recalculate(lp);
649 * Make sure we haven't switched cpus while we were asleep. It's
650 * not supposed to happen. Cleanup our temporary flags.
652 KKASSERT(gd == td->td_gd);
655 * Cleanup the timeout.
658 if (td->td_flags & TDF_TIMEOUT) {
659 td->td_flags &= ~TDF_TIMEOUT;
662 callout_stop(&thandle);
667 * Make sure we have been removed from the sleepq. This should
668 * have been done for us already.
670 * However, it is possible for a scheduling IPI to be in flight
671 * from a previous tsleep/tsleep_interlock or due to a straight-out
672 * call to lwkt_schedule() (in the case of an interrupt thread).
673 * So don't complain if DESCHEDULED is still set.
677 if (td->td_flags & TDF_TSLEEP_DESCHEDULED) {
678 td->td_flags &= ~TDF_TSLEEP_DESCHEDULED;
682 * Figure out the correct error return. If interrupted by a
683 * signal we want to return EINTR or ERESTART.
685 * If P_MAILBOX is set no automatic system call restart occurs
686 * and we return EINTR. P_MAILBOX is meant to be used as an
687 * interlock, the user must poll it prior to any system call
688 * that it wishes to interlock a mailbox signal against since
689 * the flag is cleared on *any* system call that sleeps.
691 * p->p_token is held in the p && catch case.
695 if (catch && error == 0) {
696 if ((p->p_flag & P_MAILBOX) && sig == 0) {
698 } else if (sig != 0 || (sig = CURSIG(lp))) {
699 if (SIGISMEMBER(p->p_sigacts->ps_sigintr, sig))
706 lwkt_reltoken(&p->p_token);
707 lp->lwp_flag &= ~(LWP_BREAKTSLEEP | LWP_SINTR);
708 p->p_flag &= ~P_MAILBOX;
710 logtsleep1(tsleep_end);
716 * Interlocked spinlock sleep. An exclusively held spinlock must
717 * be passed to ssleep(). The function will atomically release the
718 * spinlock and tsleep on the ident, then reacquire the spinlock and
721 * This routine is fairly important along the critical path, so optimize it
725 ssleep(const volatile void *ident, struct spinlock *spin, int flags,
726 const char *wmesg, int timo)
728 globaldata_t gd = mycpu;
731 _tsleep_interlock(gd, ident, flags);
732 spin_unlock_quick(gd, spin);
733 error = tsleep(ident, flags | PINTERLOCKED, wmesg, timo);
734 spin_lock_quick(gd, spin);
740 lksleep(const volatile void *ident, struct lock *lock, int flags,
741 const char *wmesg, int timo)
743 globaldata_t gd = mycpu;
746 _tsleep_interlock(gd, ident, flags);
747 lockmgr(lock, LK_RELEASE);
748 error = tsleep(ident, flags | PINTERLOCKED, wmesg, timo);
749 lockmgr(lock, LK_EXCLUSIVE);
755 * Interlocked mutex sleep. An exclusively held mutex must be passed
756 * to mtxsleep(). The function will atomically release the mutex
757 * and tsleep on the ident, then reacquire the mutex and return.
760 mtxsleep(const volatile void *ident, struct mtx *mtx, int flags,
761 const char *wmesg, int timo)
763 globaldata_t gd = mycpu;
766 _tsleep_interlock(gd, ident, flags);
768 error = tsleep(ident, flags | PINTERLOCKED, wmesg, timo);
769 mtx_lock_ex_quick(mtx, wmesg);
775 * Interlocked serializer sleep. An exclusively held serializer must
776 * be passed to zsleep(). The function will atomically release
777 * the serializer and tsleep on the ident, then reacquire the serializer
781 zsleep(const volatile void *ident, struct lwkt_serialize *slz, int flags,
782 const char *wmesg, int timo)
784 globaldata_t gd = mycpu;
787 ASSERT_SERIALIZED(slz);
789 _tsleep_interlock(gd, ident, flags);
790 lwkt_serialize_exit(slz);
791 ret = tsleep(ident, flags | PINTERLOCKED, wmesg, timo);
792 lwkt_serialize_enter(slz);
798 * Directly block on the LWKT thread by descheduling it. This
799 * is much faster then tsleep(), but the only legal way to wake
800 * us up is to directly schedule the thread.
802 * Setting TDF_SINTR will cause new signals to directly schedule us.
804 * This routine must be called while in a critical section.
807 lwkt_sleep(const char *wmesg, int flags)
809 thread_t td = curthread;
812 if ((flags & PCATCH) == 0 || td->td_lwp == NULL) {
813 td->td_flags |= TDF_BLOCKED;
814 td->td_wmesg = wmesg;
815 lwkt_deschedule_self(td);
818 td->td_flags &= ~TDF_BLOCKED;
821 if ((sig = CURSIG(td->td_lwp)) != 0) {
822 if (SIGISMEMBER(td->td_proc->p_sigacts->ps_sigintr, sig))
828 td->td_flags |= TDF_BLOCKED | TDF_SINTR;
829 td->td_wmesg = wmesg;
830 lwkt_deschedule_self(td);
832 td->td_flags &= ~(TDF_BLOCKED | TDF_SINTR);
838 * Implement the timeout for tsleep.
840 * We set LWP_BREAKTSLEEP to indicate that an event has occured, but
841 * we only call setrunnable if the process is not stopped.
843 * This type of callout timeout is scheduled on the same cpu the process
844 * is sleeping on. Also, at the moment, the MP lock is held.
856 lwkt_gettoken(&lp->lwp_proc->p_token);
859 * cpu interlock. Thread flags are only manipulated on
860 * the cpu owning the thread. proc flags are only manipulated
861 * by the holder of p->p_token. We have both.
863 if (td->td_flags & TDF_TSLEEP_DESCHEDULED) {
864 td->td_flags |= TDF_TIMEOUT;
867 lp->lwp_flag |= LWP_BREAKTSLEEP;
868 if (lp->lwp_proc->p_stat != SSTOP)
875 lwkt_reltoken(&lp->lwp_proc->p_token);
880 * Make all processes sleeping on the specified identifier runnable.
881 * count may be zero or one only.
883 * The domain encodes the sleep/wakeup domain AND the first cpu to check
884 * (which is always the current cpu). As we iterate across cpus
886 * This call may run without the MP lock held. We can only manipulate thread
887 * state on the cpu owning the thread. We CANNOT manipulate process state
890 * _wakeup() can be passed to an IPI so we can't use (const volatile
894 _wakeup(void *ident, int domain)
906 logtsleep2(wakeup_beg, ident);
909 qp = &gd->gd_tsleep_hash[id];
911 for (td = TAILQ_FIRST(qp); td != NULL; td = ntd) {
912 ntd = TAILQ_NEXT(td, td_sleepq);
913 if (td->td_wchan == ident &&
914 td->td_wdomain == (domain & PDOMAIN_MASK)
916 KKASSERT(td->td_gd == gd);
918 if (td->td_flags & TDF_TSLEEP_DESCHEDULED) {
919 td->td_flags &= ~TDF_TSLEEP_DESCHEDULED;
921 if (domain & PWAKEUP_ONE)
930 * We finished checking the current cpu but there still may be
931 * more work to do. Either wakeup_one was requested and no matching
932 * thread was found, or a normal wakeup was requested and we have
933 * to continue checking cpus.
935 * It should be noted that this scheme is actually less expensive then
936 * the old scheme when waking up multiple threads, since we send
937 * only one IPI message per target candidate which may then schedule
938 * multiple threads. Before we could have wound up sending an IPI
939 * message for each thread on the target cpu (!= current cpu) that
940 * needed to be woken up.
942 * NOTE: Wakeups occuring on remote cpus are asynchronous. This
943 * should be ok since we are passing idents in the IPI rather then
946 if ((domain & PWAKEUP_MYCPU) == 0 &&
947 (mask = slpque_cpumasks[id] & gd->gd_other_cpus) != 0) {
948 lwkt_send_ipiq2_mask(mask, _wakeup, ident,
949 domain | PWAKEUP_MYCPU);
953 logtsleep1(wakeup_end);
958 * Wakeup all threads tsleep()ing on the specified ident, on all cpus
961 wakeup(const volatile void *ident)
963 _wakeup(__DEALL(ident), PWAKEUP_ENCODE(0, mycpu->gd_cpuid));
967 * Wakeup one thread tsleep()ing on the specified ident, on any cpu.
970 wakeup_one(const volatile void *ident)
972 /* XXX potentially round-robin the first responding cpu */
973 _wakeup(__DEALL(ident), PWAKEUP_ENCODE(0, mycpu->gd_cpuid) | PWAKEUP_ONE);
977 * Wakeup threads tsleep()ing on the specified ident on the current cpu
981 wakeup_mycpu(const volatile void *ident)
983 _wakeup(__DEALL(ident), PWAKEUP_MYCPU);
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_MYCPU|PWAKEUP_ONE);
998 * Wakeup all thread tsleep()ing on the specified ident on the specified cpu
1002 wakeup_oncpu(globaldata_t gd, const volatile void *ident)
1006 _wakeup(__DEALL(ident), PWAKEUP_MYCPU);
1008 lwkt_send_ipiq2(gd, _wakeup, __DEALL(ident), PWAKEUP_MYCPU);
1011 _wakeup(__DEALL(ident), PWAKEUP_MYCPU);
1016 * Wakeup one thread tsleep()ing on the specified ident on the specified cpu
1020 wakeup_oncpu_one(globaldata_t gd, const volatile void *ident)
1024 _wakeup(__DEALL(ident), PWAKEUP_MYCPU | PWAKEUP_ONE);
1026 lwkt_send_ipiq2(gd, _wakeup, __DEALL(ident),
1027 PWAKEUP_MYCPU | PWAKEUP_ONE);
1030 _wakeup(__DEALL(ident), 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);
1059 * Make a process runnable. lp->lwp_proc->p_token must be held on call.
1060 * This only has an effect if we are in SSLEEP. We only break out of the
1061 * tsleep if LWP_BREAKTSLEEP is set, otherwise we just fix-up the state.
1063 * NOTE: With p_token held we can only safely manipulate the process
1064 * structure and the lp's lwp_stat.
1067 setrunnable(struct lwp *lp)
1069 ASSERT_LWKT_TOKEN_HELD(&lp->lwp_proc->p_token);
1071 if (lp->lwp_stat == LSSTOP)
1072 lp->lwp_stat = LSSLEEP;
1073 if (lp->lwp_stat == LSSLEEP && (lp->lwp_flag & LWP_BREAKTSLEEP))
1074 _tsleep_wakeup(lp->lwp_thread);
1079 * The process is stopped due to some condition, usually because p_stat is
1080 * set to SSTOP, but also possibly due to being traced.
1082 * NOTE! If the caller sets SSTOP, the caller must also clear P_WAITED
1083 * because the parent may check the child's status before the child actually
1084 * gets to this routine.
1086 * This routine is called with the current lwp only, typically just
1087 * before returning to userland.
1089 * Setting LWP_BREAKTSLEEP before entering the tsleep will cause a passive
1090 * SIGCONT to break out of the tsleep.
1095 struct lwp *lp = curthread->td_lwp;
1096 struct proc *p = lp->lwp_proc;
1100 * If LWP_WSTOP is set, we were sleeping
1101 * while our process was stopped. At this point
1102 * we were already counted as stopped.
1104 if ((lp->lwp_flag & LWP_WSTOP) == 0) {
1106 * If we're the last thread to stop, signal
1110 lp->lwp_flag |= LWP_WSTOP;
1111 wakeup(&p->p_nstopped);
1112 if (p->p_nstopped == p->p_nthreads) {
1113 p->p_flag &= ~P_WAITED;
1115 if ((p->p_pptr->p_sigacts->ps_flag & PS_NOCLDSTOP) == 0)
1116 ksignal(p->p_pptr, SIGCHLD);
1119 while (p->p_stat == SSTOP) {
1120 lp->lwp_flag |= LWP_BREAKTSLEEP;
1121 lp->lwp_stat = LSSTOP;
1122 tsleep(p, 0, "stop", 0);
1125 lp->lwp_flag &= ~LWP_WSTOP;
1130 * Compute a tenex style load average of a quantity on
1131 * 1, 5 and 15 minute intervals.
1133 static int loadav_count_runnable(struct lwp *p, void *data);
1138 struct loadavg *avg;
1142 alllwp_scan(loadav_count_runnable, &nrun);
1144 for (i = 0; i < 3; i++) {
1145 avg->ldavg[i] = (cexp[i] * avg->ldavg[i] +
1146 nrun * FSCALE * (FSCALE - cexp[i])) >> FSHIFT;
1150 * Schedule the next update to occur after 5 seconds, but add a
1151 * random variation to avoid synchronisation with processes that
1152 * run at regular intervals.
1154 callout_reset(&loadav_callout, hz * 4 + (int)(krandom() % (hz * 2 + 1)),
1159 loadav_count_runnable(struct lwp *lp, void *data)
1164 switch (lp->lwp_stat) {
1166 if ((td = lp->lwp_thread) == NULL)
1168 if (td->td_flags & TDF_BLOCKED)
1180 sched_setup(void *dummy)
1182 callout_init(&loadav_callout);
1183 callout_init(&schedcpu_callout);
1185 /* Kick off timeout driven events by calling first time. */