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/signal2.h>
51 #include <sys/resourcevar.h>
52 #include <sys/vmmeter.h>
53 #include <sys/sysctl.h>
57 #include <sys/ktrace.h>
59 #include <sys/xwait.h>
62 #include <sys/thread2.h>
63 #include <sys/spinlock2.h>
64 #include <sys/serialize.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;
80 int ncpus_fit, ncpus_fit_mask;
84 static struct callout loadav_callout;
85 static struct callout schedcpu_callout;
86 MALLOC_DEFINE(M_TSLEEP, "tslpque", "tsleep queues");
88 #if !defined(KTR_TSLEEP)
89 #define KTR_TSLEEP KTR_ALL
91 KTR_INFO_MASTER(tsleep);
92 KTR_INFO(KTR_TSLEEP, tsleep, tsleep_beg, 0, "tsleep enter %p", sizeof(void *));
93 KTR_INFO(KTR_TSLEEP, tsleep, tsleep_end, 1, "tsleep exit", 0);
94 KTR_INFO(KTR_TSLEEP, tsleep, wakeup_beg, 2, "wakeup enter %p", sizeof(void *));
95 KTR_INFO(KTR_TSLEEP, tsleep, wakeup_end, 3, "wakeup exit", 0);
96 KTR_INFO(KTR_TSLEEP, tsleep, ilockfail, 4, "interlock failed %p", sizeof(void *));
98 #define logtsleep1(name) KTR_LOG(tsleep_ ## name)
99 #define logtsleep2(name, val) KTR_LOG(tsleep_ ## name, val)
101 struct loadavg averunnable =
102 { {0, 0, 0}, FSCALE }; /* load average, of runnable procs */
104 * Constants for averages over 1, 5, and 15 minutes
105 * when sampling at 5 second intervals.
107 static fixpt_t cexp[3] = {
108 0.9200444146293232 * FSCALE, /* exp(-1/12) */
109 0.9834714538216174 * FSCALE, /* exp(-1/60) */
110 0.9944598480048967 * FSCALE, /* exp(-1/180) */
113 static void endtsleep (void *);
114 static void unsleep_and_wakeup_thread(struct thread *td);
115 static void loadav (void *arg);
116 static void schedcpu (void *arg);
119 * Adjust the scheduler quantum. The quantum is specified in microseconds.
120 * Note that 'tick' is in microseconds per tick.
123 sysctl_kern_quantum(SYSCTL_HANDLER_ARGS)
127 new_val = sched_quantum * tick;
128 error = sysctl_handle_int(oidp, &new_val, 0, req);
129 if (error != 0 || req->newptr == NULL)
133 sched_quantum = new_val / tick;
134 hogticks = 2 * sched_quantum;
138 SYSCTL_PROC(_kern, OID_AUTO, quantum, CTLTYPE_INT|CTLFLAG_RW,
139 0, sizeof sched_quantum, sysctl_kern_quantum, "I", "");
142 * If `ccpu' is not equal to `exp(-1/20)' and you still want to use the
143 * faster/more-accurate formula, you'll have to estimate CCPU_SHIFT below
144 * and possibly adjust FSHIFT in "param.h" so that (FSHIFT >= CCPU_SHIFT).
146 * To estimate CCPU_SHIFT for exp(-1/20), the following formula was used:
147 * 1 - exp(-1/20) ~= 0.0487 ~= 0.0488 == 1 (fixed pt, *11* bits).
149 * If you don't want to bother with the faster/more-accurate formula, you
150 * can set CCPU_SHIFT to (FSHIFT + 1) which will use a slower/less-accurate
151 * (more general) method of calculating the %age of CPU used by a process.
153 * decay 95% of `lwp_pctcpu' in 60 seconds; see CCPU_SHIFT before changing
155 #define CCPU_SHIFT 11
157 static fixpt_t ccpu = 0.95122942450071400909 * FSCALE; /* exp(-1/20) */
158 SYSCTL_INT(_kern, OID_AUTO, ccpu, CTLFLAG_RD, &ccpu, 0, "");
161 * kernel uses `FSCALE', userland (SHOULD) use kern.fscale
163 int fscale __unused = FSCALE; /* exported to systat */
164 SYSCTL_INT(_kern, OID_AUTO, fscale, CTLFLAG_RD, 0, FSCALE, "");
167 * Recompute process priorities, once a second.
169 * Since the userland schedulers are typically event oriented, if the
170 * estcpu calculation at wakeup() time is not sufficient to make a
171 * process runnable relative to other processes in the system we have
172 * a 1-second recalc to help out.
174 * This code also allows us to store sysclock_t data in the process structure
175 * without fear of an overrun, since sysclock_t are guarenteed to hold
176 * several seconds worth of count.
178 * WARNING! callouts can preempt normal threads. However, they will not
179 * preempt a thread holding a spinlock so we *can* safely use spinlocks.
181 static int schedcpu_stats(struct proc *p, void *data __unused);
182 static int schedcpu_resource(struct proc *p, void *data __unused);
187 allproc_scan(schedcpu_stats, NULL);
188 allproc_scan(schedcpu_resource, NULL);
189 wakeup((caddr_t)&lbolt);
190 wakeup((caddr_t)&lbolt_syncer);
191 callout_reset(&schedcpu_callout, hz, schedcpu, NULL);
195 * General process statistics once a second
198 schedcpu_stats(struct proc *p, void *data __unused)
204 FOREACH_LWP_IN_PROC(lp, p) {
205 if (lp->lwp_stat == LSSLEEP)
209 * Only recalculate processes that are active or have slept
210 * less then 2 seconds. The schedulers understand this.
212 if (lp->lwp_slptime <= 1) {
213 p->p_usched->recalculate(lp);
215 lp->lwp_pctcpu = (lp->lwp_pctcpu * ccpu) >> FSHIFT;
223 * Resource checks. XXX break out since ksignal/killproc can block,
224 * limiting us to one process killed per second. There is probably
228 schedcpu_resource(struct proc *p, void *data __unused)
234 if (p->p_stat == SIDL ||
235 p->p_stat == SZOMB ||
243 FOREACH_LWP_IN_PROC(lp, p) {
245 * We may have caught an lp in the middle of being
246 * created, lwp_thread can be NULL.
248 if (lp->lwp_thread) {
249 ttime += lp->lwp_thread->td_sticks;
250 ttime += lp->lwp_thread->td_uticks;
254 switch(plimit_testcpulimit(p->p_limit, ttime)) {
255 case PLIMIT_TESTCPU_KILL:
256 killproc(p, "exceeded maximum CPU limit");
258 case PLIMIT_TESTCPU_XCPU:
259 if ((p->p_flag & P_XCPU) == 0) {
272 * This is only used by ps. Generate a cpu percentage use over
273 * a period of one second.
278 updatepcpu(struct lwp *lp, int cpticks, int ttlticks)
283 acc = (cpticks << FSHIFT) / ttlticks;
284 if (ttlticks >= ESTCPUFREQ) {
285 lp->lwp_pctcpu = acc;
287 remticks = ESTCPUFREQ - ttlticks;
288 lp->lwp_pctcpu = (acc * ttlticks + lp->lwp_pctcpu * remticks) /
294 * tsleep/wakeup hash table parameters. Try to find the sweet spot for
295 * like addresses being slept on.
297 #define TABLESIZE 1024
298 #define LOOKUP(x) (((intptr_t)(x) >> 6) & (TABLESIZE - 1))
300 static cpumask_t slpque_cpumasks[TABLESIZE];
303 * General scheduler initialization. We force a reschedule 25 times
304 * a second by default. Note that cpu0 is initialized in early boot and
305 * cannot make any high level calls.
307 * Each cpu has its own sleep queue.
310 sleep_gdinit(globaldata_t gd)
312 static struct tslpque slpque_cpu0[TABLESIZE];
315 if (gd->gd_cpuid == 0) {
316 sched_quantum = (hz + 24) / 25;
317 hogticks = 2 * sched_quantum;
319 gd->gd_tsleep_hash = slpque_cpu0;
321 gd->gd_tsleep_hash = kmalloc(sizeof(slpque_cpu0),
322 M_TSLEEP, M_WAITOK | M_ZERO);
324 for (i = 0; i < TABLESIZE; ++i)
325 TAILQ_INIT(&gd->gd_tsleep_hash[i]);
329 * General sleep call. Suspends the current process until a wakeup is
330 * performed on the specified identifier. The process will then be made
331 * runnable with the specified priority. Sleeps at most timo/hz seconds
332 * (0 means no timeout). If flags includes PCATCH flag, signals are checked
333 * before and after sleeping, else signals are not checked. Returns 0 if
334 * awakened, EWOULDBLOCK if the timeout expires. If PCATCH is set and a
335 * signal needs to be delivered, ERESTART is returned if the current system
336 * call should be restarted if possible, and EINTR is returned if the system
337 * call should be interrupted by the signal (return EINTR).
339 * Note that if we are a process, we release_curproc() before messing with
340 * the LWKT scheduler.
342 * During autoconfiguration or after a panic, a sleep will simply
343 * lower the priority briefly to allow interrupts, then return.
346 tsleep(void *ident, int flags, const char *wmesg, int timo)
348 struct thread *td = curthread;
349 struct lwp *lp = td->td_lwp;
350 struct proc *p = td->td_proc; /* may be NULL */
357 struct callout thandle;
360 * NOTE: removed KTRPOINT, it could cause races due to blocking
361 * even in stable. Just scrap it for now.
363 if (tsleep_now_works == 0 || panicstr) {
365 * After a panic, or before we actually have an operational
366 * softclock, just give interrupts a chance, then just return;
368 * don't run any other procs or panic below,
369 * in case this is the idle process and already asleep.
372 oldpri = td->td_pri & TDPRI_MASK;
373 lwkt_setpri_self(safepri);
375 lwkt_setpri_self(oldpri);
378 logtsleep2(tsleep_beg, ident);
380 KKASSERT(td != &gd->gd_idlethread); /* you must be kidding! */
383 * NOTE: all of this occurs on the current cpu, including any
384 * callout-based wakeups, so a critical section is a sufficient
387 * The entire sequence through to where we actually sleep must
388 * run without breaking the critical section.
391 catch = flags & PCATCH;
395 crit_enter_quick(td);
397 KASSERT(ident != NULL, ("tsleep: no ident"));
398 KASSERT(lp == NULL ||
399 lp->lwp_stat == LSRUN || /* Obvious */
400 lp->lwp_stat == LSSTOP, /* Set in tstop */
402 ident, wmesg, lp->lwp_stat));
405 * Setup for the current process (if this is a process).
410 * Early termination if PCATCH was set and a
411 * signal is pending, interlocked with the
414 * Early termination only occurs when tsleep() is
415 * entered while in a normal LSRUN state.
417 if ((sig = CURSIG(lp)) != 0)
421 * Early termination if PCATCH was set and a
422 * mailbox signal was possibly delivered prior to
423 * the system call even being made, in order to
424 * allow the user to interlock without having to
425 * make additional system calls.
427 if (p->p_flag & P_MAILBOX)
431 * Causes ksignal to wake us up when.
433 lp->lwp_flag |= LWP_SINTR;
437 * Make sure the current process has been untangled from
438 * the userland scheduler and initialize slptime to start
441 p->p_usched->release_curproc(lp);
446 * If the interlocked flag is set but our cpu bit in the slpqueue
447 * is no longer set, then a wakeup was processed inbetween the
448 * tsleep_interlock() and here. This can occur under extreme loads
449 * if the IPIQ fills up and gets processed synchronously by, say,
450 * a wakeup() or other IPI sent inbetween the interlock and here.
452 * Even the usched->release function just above can muff it up.
454 if ((flags & PINTERLOCKED) &&
455 (slpque_cpumasks[id] & gd->gd_cpumask) == 0) {
456 logtsleep2(ilockfail, ident);
461 * Move our thread to the correct queue and setup our wchan, etc.
463 lwkt_deschedule_self(td);
464 td->td_flags |= TDF_TSLEEPQ;
465 TAILQ_INSERT_TAIL(&gd->gd_tsleep_hash[id], td, td_threadq);
466 atomic_set_int(&slpque_cpumasks[id], gd->gd_cpumask);
468 td->td_wchan = ident;
469 td->td_wmesg = wmesg;
470 td->td_wdomain = flags & PDOMAIN_MASK;
473 * Setup the timeout, if any
476 callout_init(&thandle);
477 callout_reset(&thandle, timo, endtsleep, td);
485 * Ok, we are sleeping. Place us in the SSLEEP state.
487 KKASSERT((lp->lwp_flag & LWP_ONRUNQ) == 0);
489 * tstop() sets LSSTOP, so don't fiddle with that.
491 if (lp->lwp_stat != LSSTOP)
492 lp->lwp_stat = LSSLEEP;
493 lp->lwp_ru.ru_nvcsw++;
497 * And when we are woken up, put us back in LSRUN. If we
498 * slept for over a second, recalculate our estcpu.
500 lp->lwp_stat = LSRUN;
502 p->p_usched->recalculate(lp);
509 * Make sure we haven't switched cpus while we were asleep. It's
510 * not supposed to happen. Cleanup our temporary flags.
512 KKASSERT(gd == td->td_gd);
515 * Cleanup the timeout.
518 if (td->td_flags & TDF_TIMEOUT) {
519 td->td_flags &= ~TDF_TIMEOUT;
522 callout_stop(&thandle);
527 * Since td_threadq is used both for our run queue AND for the
528 * tsleep hash queue, we can't still be on it at this point because
529 * we've gotten cpu back.
531 KASSERT((td->td_flags & TDF_TSLEEPQ) == 0, ("tsleep: impossible thread flags %08x", td->td_flags));
537 * Figure out the correct error return. If interrupted by a
538 * signal we want to return EINTR or ERESTART.
540 * If P_MAILBOX is set no automatic system call restart occurs
541 * and we return EINTR. P_MAILBOX is meant to be used as an
542 * interlock, the user must poll it prior to any system call
543 * that it wishes to interlock a mailbox signal against since
544 * the flag is cleared on *any* system call that sleeps.
548 if (catch && error == 0) {
549 if ((p->p_flag & P_MAILBOX) && sig == 0) {
551 } else if (sig != 0 || (sig = CURSIG(lp))) {
552 if (SIGISMEMBER(p->p_sigacts->ps_sigintr, sig))
558 lp->lwp_flag &= ~(LWP_BREAKTSLEEP | LWP_SINTR);
559 p->p_flag &= ~P_MAILBOX;
561 logtsleep1(tsleep_end);
567 * This is a dandy function that allows us to interlock tsleep/wakeup
568 * operations with unspecified upper level locks, such as lockmgr locks,
569 * simply by holding a critical section. The sequence is:
571 * (enter critical section)
572 * (acquire upper level lock)
573 * tsleep_interlock(blah)
574 * (release upper level lock)
576 * (exit critical section)
578 * Basically this function sets our cpumask for the ident which informs
579 * other cpus that our cpu 'might' be waiting (or about to wait on) the
580 * hash index related to the ident. The critical section prevents another
581 * cpu's wakeup() from being processed on our cpu until we are actually
582 * able to enter the tsleep(). Thus, no race occurs between our attempt
583 * to release a resource and sleep, and another cpu's attempt to acquire
584 * a resource and call wakeup.
586 * There isn't much of a point to this function unless you call it while
587 * holding a critical section.
590 _tsleep_interlock(globaldata_t gd, void *ident)
592 int id = LOOKUP(ident);
594 atomic_set_int(&slpque_cpumasks[id], gd->gd_cpumask);
598 tsleep_interlock(void *ident)
600 _tsleep_interlock(mycpu, ident);
604 * Interlocked spinlock sleep. An exclusively held spinlock must
605 * be passed to msleep(). The function will atomically release the
606 * spinlock and tsleep on the ident, then reacquire the spinlock and
609 * This routine is fairly important along the critical path, so optimize it
613 msleep(void *ident, struct spinlock *spin, int flags,
614 const char *wmesg, int timo)
616 globaldata_t gd = mycpu;
620 _tsleep_interlock(gd, ident);
621 spin_unlock_wr_quick(gd, spin);
622 error = tsleep(ident, flags | PINTERLOCKED, wmesg, timo);
623 spin_lock_wr_quick(gd, spin);
630 * Interlocked serializer sleep. An exclusively held serializer must
631 * be passed to serialize_sleep(). The function will atomically release
632 * the serializer and tsleep on the ident, then reacquire the serializer
636 serialize_sleep(void *ident, struct lwkt_serialize *slz, int flags,
637 const char *wmesg, int timo)
641 ASSERT_SERIALIZED(slz);
644 tsleep_interlock(ident);
645 lwkt_serialize_exit(slz);
646 ret = tsleep(ident, flags | PINTERLOCKED, wmesg, timo);
647 lwkt_serialize_enter(slz);
654 * Directly block on the LWKT thread by descheduling it. This
655 * is much faster then tsleep(), but the only legal way to wake
656 * us up is to directly schedule the thread.
658 * Setting TDF_SINTR will cause new signals to directly schedule us.
660 * This routine is typically called while in a critical section.
663 lwkt_sleep(const char *wmesg, int flags)
665 thread_t td = curthread;
668 if ((flags & PCATCH) == 0 || td->td_lwp == NULL) {
669 td->td_flags |= TDF_BLOCKED;
670 td->td_wmesg = wmesg;
671 lwkt_deschedule_self(td);
674 td->td_flags &= ~TDF_BLOCKED;
677 if ((sig = CURSIG(td->td_lwp)) != 0) {
678 if (SIGISMEMBER(td->td_proc->p_sigacts->ps_sigintr, sig))
684 td->td_flags |= TDF_BLOCKED | TDF_SINTR;
685 td->td_wmesg = wmesg;
686 lwkt_deschedule_self(td);
688 td->td_flags &= ~(TDF_BLOCKED | TDF_SINTR);
694 * Implement the timeout for tsleep.
696 * We set LWP_BREAKTSLEEP to indicate that an event has occured, but
697 * we only call setrunnable if the process is not stopped.
699 * This type of callout timeout is scheduled on the same cpu the process
700 * is sleeping on. Also, at the moment, the MP lock is held.
708 ASSERT_MP_LOCK_HELD(curthread);
712 * cpu interlock. Thread flags are only manipulated on
713 * the cpu owning the thread. proc flags are only manipulated
714 * by the older of the MP lock. We have both.
716 if (td->td_flags & TDF_TSLEEPQ) {
717 td->td_flags |= TDF_TIMEOUT;
719 if ((lp = td->td_lwp) != NULL) {
720 lp->lwp_flag |= LWP_BREAKTSLEEP;
721 if (lp->lwp_proc->p_stat != SSTOP)
724 unsleep_and_wakeup_thread(td);
731 * Unsleep and wakeup a thread. This function runs without the MP lock
732 * which means that it can only manipulate thread state on the owning cpu,
733 * and cannot touch the process state at all.
737 unsleep_and_wakeup_thread(struct thread *td)
739 globaldata_t gd = mycpu;
743 if (td->td_gd != gd) {
744 lwkt_send_ipiq(td->td_gd, (ipifunc1_t)unsleep_and_wakeup_thread, td);
749 if (td->td_flags & TDF_TSLEEPQ) {
750 td->td_flags &= ~TDF_TSLEEPQ;
751 id = LOOKUP(td->td_wchan);
752 TAILQ_REMOVE(&gd->gd_tsleep_hash[id], td, td_threadq);
753 if (TAILQ_FIRST(&gd->gd_tsleep_hash[id]) == NULL)
754 atomic_clear_int(&slpque_cpumasks[id], gd->gd_cpumask);
761 * Make all processes sleeping on the specified identifier runnable.
762 * count may be zero or one only.
764 * The domain encodes the sleep/wakeup domain AND the first cpu to check
765 * (which is always the current cpu). As we iterate across cpus
767 * This call may run without the MP lock held. We can only manipulate thread
768 * state on the cpu owning the thread. We CANNOT manipulate process state
772 _wakeup(void *ident, int domain)
784 logtsleep2(wakeup_beg, ident);
787 qp = &gd->gd_tsleep_hash[id];
789 for (td = TAILQ_FIRST(qp); td != NULL; td = ntd) {
790 ntd = TAILQ_NEXT(td, td_threadq);
791 if (td->td_wchan == ident &&
792 td->td_wdomain == (domain & PDOMAIN_MASK)
794 KKASSERT(td->td_flags & TDF_TSLEEPQ);
795 td->td_flags &= ~TDF_TSLEEPQ;
796 TAILQ_REMOVE(qp, td, td_threadq);
797 if (TAILQ_FIRST(qp) == NULL) {
798 atomic_clear_int(&slpque_cpumasks[id],
802 if (domain & PWAKEUP_ONE)
810 * We finished checking the current cpu but there still may be
811 * more work to do. Either wakeup_one was requested and no matching
812 * thread was found, or a normal wakeup was requested and we have
813 * to continue checking cpus.
815 * It should be noted that this scheme is actually less expensive then
816 * the old scheme when waking up multiple threads, since we send
817 * only one IPI message per target candidate which may then schedule
818 * multiple threads. Before we could have wound up sending an IPI
819 * message for each thread on the target cpu (!= current cpu) that
820 * needed to be woken up.
822 * NOTE: Wakeups occuring on remote cpus are asynchronous. This
823 * should be ok since we are passing idents in the IPI rather then
826 if ((domain & PWAKEUP_MYCPU) == 0 &&
827 (mask = slpque_cpumasks[id] & gd->gd_other_cpus) != 0) {
828 lwkt_send_ipiq2_mask(mask, _wakeup, ident,
829 domain | PWAKEUP_MYCPU);
833 logtsleep1(wakeup_end);
838 * Wakeup all threads tsleep()ing on the specified ident, on all cpus
843 _wakeup(ident, PWAKEUP_ENCODE(0, mycpu->gd_cpuid));
847 * Wakeup one thread tsleep()ing on the specified ident, on any cpu.
850 wakeup_one(void *ident)
852 /* XXX potentially round-robin the first responding cpu */
853 _wakeup(ident, PWAKEUP_ENCODE(0, mycpu->gd_cpuid) | PWAKEUP_ONE);
857 * Wakeup threads tsleep()ing on the specified ident on the current cpu
861 wakeup_mycpu(void *ident)
863 _wakeup(ident, PWAKEUP_MYCPU);
867 * Wakeup one thread tsleep()ing on the specified ident on the current cpu
871 wakeup_mycpu_one(void *ident)
873 /* XXX potentially round-robin the first responding cpu */
874 _wakeup(ident, PWAKEUP_MYCPU|PWAKEUP_ONE);
878 * Wakeup all thread tsleep()ing on the specified ident on the specified cpu
882 wakeup_oncpu(globaldata_t gd, void *ident)
886 _wakeup(ident, PWAKEUP_MYCPU);
888 lwkt_send_ipiq2(gd, _wakeup, ident, PWAKEUP_MYCPU);
891 _wakeup(ident, PWAKEUP_MYCPU);
896 * Wakeup one thread tsleep()ing on the specified ident on the specified cpu
900 wakeup_oncpu_one(globaldata_t gd, void *ident)
904 _wakeup(ident, PWAKEUP_MYCPU | PWAKEUP_ONE);
906 lwkt_send_ipiq2(gd, _wakeup, ident, PWAKEUP_MYCPU | PWAKEUP_ONE);
909 _wakeup(ident, PWAKEUP_MYCPU | PWAKEUP_ONE);
914 * Wakeup all threads waiting on the specified ident that slept using
915 * the specified domain, on all cpus.
918 wakeup_domain(void *ident, int domain)
920 _wakeup(ident, PWAKEUP_ENCODE(domain, mycpu->gd_cpuid));
924 * Wakeup one thread waiting on the specified ident that slept using
925 * the specified domain, on any cpu.
928 wakeup_domain_one(void *ident, int domain)
930 /* XXX potentially round-robin the first responding cpu */
931 _wakeup(ident, PWAKEUP_ENCODE(domain, mycpu->gd_cpuid) | PWAKEUP_ONE);
937 * Make a process runnable. The MP lock must be held on call. This only
938 * has an effect if we are in SSLEEP. We only break out of the
939 * tsleep if LWP_BREAKTSLEEP is set, otherwise we just fix-up the state.
941 * NOTE: With the MP lock held we can only safely manipulate the process
942 * structure. We cannot safely manipulate the thread structure.
945 setrunnable(struct lwp *lp)
948 ASSERT_MP_LOCK_HELD(curthread);
949 if (lp->lwp_stat == LSSTOP)
950 lp->lwp_stat = LSSLEEP;
951 if (lp->lwp_stat == LSSLEEP && (lp->lwp_flag & LWP_BREAKTSLEEP))
952 unsleep_and_wakeup_thread(lp->lwp_thread);
957 * The process is stopped due to some condition, usually because p_stat is
958 * set to SSTOP, but also possibly due to being traced.
960 * NOTE! If the caller sets SSTOP, the caller must also clear P_WAITED
961 * because the parent may check the child's status before the child actually
962 * gets to this routine.
964 * This routine is called with the current lwp only, typically just
965 * before returning to userland.
967 * Setting LWP_BREAKTSLEEP before entering the tsleep will cause a passive
968 * SIGCONT to break out of the tsleep.
973 struct lwp *lp = curthread->td_lwp;
974 struct proc *p = lp->lwp_proc;
978 * If LWP_WSTOP is set, we were sleeping
979 * while our process was stopped. At this point
980 * we were already counted as stopped.
982 if ((lp->lwp_flag & LWP_WSTOP) == 0) {
984 * If we're the last thread to stop, signal
988 lp->lwp_flag |= LWP_WSTOP;
989 wakeup(&p->p_nstopped);
990 if (p->p_nstopped == p->p_nthreads) {
991 p->p_flag &= ~P_WAITED;
993 if ((p->p_pptr->p_sigacts->ps_flag & PS_NOCLDSTOP) == 0)
994 ksignal(p->p_pptr, SIGCHLD);
997 while (p->p_stat == SSTOP) {
998 lp->lwp_flag |= LWP_BREAKTSLEEP;
999 lp->lwp_stat = LSSTOP;
1000 tsleep(p, 0, "stop", 0);
1003 lp->lwp_flag &= ~LWP_WSTOP;
1008 * Yield / synchronous reschedule. This is a bit tricky because the trap
1009 * code might have set a lazy release on the switch function. Setting
1010 * P_PASSIVE_ACQ will ensure that the lazy release executes when we call
1011 * switch, and that we are given a greater chance of affinity with our
1014 * We call lwkt_setpri_self() to rotate our thread to the end of the lwkt
1015 * run queue. lwkt_switch() will also execute any assigned passive release
1016 * (which usually calls release_curproc()), allowing a same/higher priority
1017 * process to be designated as the current process.
1019 * While it is possible for a lower priority process to be designated,
1020 * it's call to lwkt_maybe_switch() in acquire_curproc() will likely
1021 * round-robin back to us and we will be able to re-acquire the current
1022 * process designation.
1027 struct thread *td = curthread;
1028 struct proc *p = td->td_proc;
1030 lwkt_setpri_self(td->td_pri & TDPRI_MASK);
1032 p->p_flag |= P_PASSIVE_ACQ;
1034 p->p_flag &= ~P_PASSIVE_ACQ;
1041 * Compute a tenex style load average of a quantity on
1042 * 1, 5 and 15 minute intervals.
1044 static int loadav_count_runnable(struct lwp *p, void *data);
1049 struct loadavg *avg;
1053 alllwp_scan(loadav_count_runnable, &nrun);
1055 for (i = 0; i < 3; i++) {
1056 avg->ldavg[i] = (cexp[i] * avg->ldavg[i] +
1057 nrun * FSCALE * (FSCALE - cexp[i])) >> FSHIFT;
1061 * Schedule the next update to occur after 5 seconds, but add a
1062 * random variation to avoid synchronisation with processes that
1063 * run at regular intervals.
1065 callout_reset(&loadav_callout, hz * 4 + (int)(krandom() % (hz * 2 + 1)),
1070 loadav_count_runnable(struct lwp *lp, void *data)
1075 switch (lp->lwp_stat) {
1077 if ((td = lp->lwp_thread) == NULL)
1079 if (td->td_flags & TDF_BLOCKED)
1091 sched_setup(void *dummy)
1093 callout_init(&loadav_callout);
1094 callout_init(&schedcpu_callout);
1096 /* Kick off timeout driven events by calling first time. */