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 1.72 2007/02/03 17:05:58 corecode 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>
61 #include <sys/thread2.h>
62 #include <sys/spinlock2.h>
64 #include <machine/cpu.h>
65 #include <machine/smp.h>
67 TAILQ_HEAD(tslpque, thread);
69 static void sched_setup (void *dummy);
70 SYSINIT(sched_setup, SI_SUB_KICK_SCHEDULER, SI_ORDER_FIRST, sched_setup, NULL)
75 int sched_quantum; /* Roundrobin scheduling quantum in ticks. */
77 int ncpus2, ncpus2_shift, ncpus2_mask;
80 static struct callout loadav_callout;
81 static struct callout schedcpu_callout;
82 MALLOC_DEFINE(M_TSLEEP, "tslpque", "tsleep queues");
84 #if !defined(KTR_TSLEEP)
85 #define KTR_TSLEEP KTR_ALL
87 KTR_INFO_MASTER(tsleep);
88 KTR_INFO(KTR_TSLEEP, tsleep, tsleep_beg, 0, "tsleep enter", 0);
89 KTR_INFO(KTR_TSLEEP, tsleep, tsleep_end, 0, "tsleep exit", 0);
90 KTR_INFO(KTR_TSLEEP, tsleep, wakeup_beg, 0, "wakeup enter", 0);
91 KTR_INFO(KTR_TSLEEP, tsleep, wakeup_end, 0, "wakeup exit", 0);
92 #define logtsleep(name) KTR_LOG(tsleep_ ## name)
94 struct loadavg averunnable =
95 { {0, 0, 0}, FSCALE }; /* load average, of runnable procs */
97 * Constants for averages over 1, 5, and 15 minutes
98 * when sampling at 5 second intervals.
100 static fixpt_t cexp[3] = {
101 0.9200444146293232 * FSCALE, /* exp(-1/12) */
102 0.9834714538216174 * FSCALE, /* exp(-1/60) */
103 0.9944598480048967 * FSCALE, /* exp(-1/180) */
106 static void endtsleep (void *);
107 static void unsleep_and_wakeup_thread(struct thread *td);
108 static void loadav (void *arg);
109 static void schedcpu (void *arg);
112 * Adjust the scheduler quantum. The quantum is specified in microseconds.
113 * Note that 'tick' is in microseconds per tick.
116 sysctl_kern_quantum(SYSCTL_HANDLER_ARGS)
120 new_val = sched_quantum * tick;
121 error = sysctl_handle_int(oidp, &new_val, 0, req);
122 if (error != 0 || req->newptr == NULL)
126 sched_quantum = new_val / tick;
127 hogticks = 2 * sched_quantum;
131 SYSCTL_PROC(_kern, OID_AUTO, quantum, CTLTYPE_INT|CTLFLAG_RW,
132 0, sizeof sched_quantum, sysctl_kern_quantum, "I", "");
135 * If `ccpu' is not equal to `exp(-1/20)' and you still want to use the
136 * faster/more-accurate formula, you'll have to estimate CCPU_SHIFT below
137 * and possibly adjust FSHIFT in "param.h" so that (FSHIFT >= CCPU_SHIFT).
139 * To estimate CCPU_SHIFT for exp(-1/20), the following formula was used:
140 * 1 - exp(-1/20) ~= 0.0487 ~= 0.0488 == 1 (fixed pt, *11* bits).
142 * If you don't want to bother with the faster/more-accurate formula, you
143 * can set CCPU_SHIFT to (FSHIFT + 1) which will use a slower/less-accurate
144 * (more general) method of calculating the %age of CPU used by a process.
146 * decay 95% of `lwp_pctcpu' in 60 seconds; see CCPU_SHIFT before changing
148 #define CCPU_SHIFT 11
150 static fixpt_t ccpu = 0.95122942450071400909 * FSCALE; /* exp(-1/20) */
151 SYSCTL_INT(_kern, OID_AUTO, ccpu, CTLFLAG_RD, &ccpu, 0, "");
154 * kernel uses `FSCALE', userland (SHOULD) use kern.fscale
156 int fscale __unused = FSCALE; /* exported to systat */
157 SYSCTL_INT(_kern, OID_AUTO, fscale, CTLFLAG_RD, 0, FSCALE, "");
160 * Recompute process priorities, once a second.
162 * Since the userland schedulers are typically event oriented, if the
163 * estcpu calculation at wakeup() time is not sufficient to make a
164 * process runnable relative to other processes in the system we have
165 * a 1-second recalc to help out.
167 * This code also allows us to store sysclock_t data in the process structure
168 * without fear of an overrun, since sysclock_t are guarenteed to hold
169 * several seconds worth of count.
171 * WARNING! callouts can preempt normal threads. However, they will not
172 * preempt a thread holding a spinlock so we *can* safely use spinlocks.
174 static int schedcpu_stats(struct proc *p, void *data __unused);
175 static int schedcpu_resource(struct proc *p, void *data __unused);
180 allproc_scan(schedcpu_stats, NULL);
181 allproc_scan(schedcpu_resource, NULL);
182 wakeup((caddr_t)&lbolt);
183 wakeup((caddr_t)&lbolt_syncer);
184 callout_reset(&schedcpu_callout, hz, schedcpu, NULL);
188 * General process statistics once a second
191 schedcpu_stats(struct proc *p, void *data __unused)
196 lp = FIRST_LWP_IN_PROC(p);
199 if (p->p_stat == SSLEEP)
203 * Only recalculate processes that are active or have slept
204 * less then 2 seconds. The schedulers understand this.
206 if (lp->lwp_slptime <= 1) {
207 p->p_usched->recalculate(lp);
209 lp->lwp_pctcpu = (lp->lwp_pctcpu * ccpu) >> FSHIFT;
216 * Resource checks. XXX break out since ksignal/killproc can block,
217 * limiting us to one process killed per second. There is probably
221 schedcpu_resource(struct proc *p, void *data __unused)
227 lp = FIRST_LWP_IN_PROC(p);
229 if (p->p_stat == SIDL ||
230 (p->p_flag & P_ZOMBIE) ||
231 p->p_limit == NULL ||
232 lp->lwp_thread == NULL
238 ttime = lp->lwp_thread->td_sticks + lp->lwp_thread->td_uticks;
240 switch(plimit_testcpulimit(p->p_limit, ttime)) {
241 case PLIMIT_TESTCPU_KILL:
242 killproc(p, "exceeded maximum CPU limit");
244 case PLIMIT_TESTCPU_XCPU:
245 if ((p->p_flag & P_XCPU) == 0) {
258 * This is only used by ps. Generate a cpu percentage use over
259 * a period of one second.
264 updatepcpu(struct lwp *lp, int cpticks, int ttlticks)
269 acc = (cpticks << FSHIFT) / ttlticks;
270 if (ttlticks >= ESTCPUFREQ) {
271 lp->lwp_pctcpu = acc;
273 remticks = ESTCPUFREQ - ttlticks;
274 lp->lwp_pctcpu = (acc * ttlticks + lp->lwp_pctcpu * remticks) /
280 * We're only looking at 7 bits of the address; everything is
281 * aligned to 4, lots of things are aligned to greater powers
282 * of 2. Shift right by 8, i.e. drop the bottom 256 worth.
284 #define TABLESIZE 128
285 #define LOOKUP(x) (((intptr_t)(x) >> 8) & (TABLESIZE - 1))
287 static cpumask_t slpque_cpumasks[TABLESIZE];
290 * General scheduler initialization. We force a reschedule 25 times
291 * a second by default. Note that cpu0 is initialized in early boot and
292 * cannot make any high level calls.
294 * Each cpu has its own sleep queue.
297 sleep_gdinit(globaldata_t gd)
299 static struct tslpque slpque_cpu0[TABLESIZE];
302 if (gd->gd_cpuid == 0) {
303 sched_quantum = (hz + 24) / 25;
304 hogticks = 2 * sched_quantum;
306 gd->gd_tsleep_hash = slpque_cpu0;
308 gd->gd_tsleep_hash = kmalloc(sizeof(slpque_cpu0),
309 M_TSLEEP, M_WAITOK | M_ZERO);
311 for (i = 0; i < TABLESIZE; ++i)
312 TAILQ_INIT(&gd->gd_tsleep_hash[i]);
316 * General sleep call. Suspends the current process until a wakeup is
317 * performed on the specified identifier. The process will then be made
318 * runnable with the specified priority. Sleeps at most timo/hz seconds
319 * (0 means no timeout). If flags includes PCATCH flag, signals are checked
320 * before and after sleeping, else signals are not checked. Returns 0 if
321 * awakened, EWOULDBLOCK if the timeout expires. If PCATCH is set and a
322 * signal needs to be delivered, ERESTART is returned if the current system
323 * call should be restarted if possible, and EINTR is returned if the system
324 * call should be interrupted by the signal (return EINTR).
326 * Note that if we are a process, we release_curproc() before messing with
327 * the LWKT scheduler.
329 * During autoconfiguration or after a panic, a sleep will simply
330 * lower the priority briefly to allow interrupts, then return.
333 tsleep(void *ident, int flags, const char *wmesg, int timo)
335 struct thread *td = curthread;
336 struct lwp *lp = td->td_lwp;
337 struct proc *p = td->td_proc; /* may be NULL */
344 struct callout thandle;
347 * NOTE: removed KTRPOINT, it could cause races due to blocking
348 * even in stable. Just scrap it for now.
350 if (cold || panicstr) {
352 * After a panic, or during autoconfiguration,
353 * just give interrupts a chance, then just return;
354 * don't run any other procs or panic below,
355 * in case this is the idle process and already asleep.
358 oldpri = td->td_pri & TDPRI_MASK;
359 lwkt_setpri_self(safepri);
361 lwkt_setpri_self(oldpri);
364 logtsleep(tsleep_beg);
366 KKASSERT(td != &gd->gd_idlethread); /* you must be kidding! */
369 * NOTE: all of this occurs on the current cpu, including any
370 * callout-based wakeups, so a critical section is a sufficient
373 * The entire sequence through to where we actually sleep must
374 * run without breaking the critical section.
377 catch = flags & PCATCH;
381 crit_enter_quick(td);
383 KASSERT(ident != NULL, ("tsleep: no ident"));
384 KASSERT(p == NULL || p->p_stat == SRUN, ("tsleep %p %s %d",
385 ident, wmesg, p->p_stat));
388 * Setup for the current process (if this is a process).
393 * Early termination if PCATCH was set and a
394 * signal is pending, interlocked with the
397 * Early termination only occurs when tsleep() is
398 * entered while in a normal SRUN state.
400 if ((sig = CURSIG(lp)) != 0)
404 * Early termination if PCATCH was set and a
405 * mailbox signal was possibly delivered prior to
406 * the system call even being made, in order to
407 * allow the user to interlock without having to
408 * make additional system calls.
410 if (p->p_flag & P_MAILBOX)
414 * Causes ksignal to wake us up when.
416 p->p_flag |= P_SINTR;
420 * Make sure the current process has been untangled from
421 * the userland scheduler and initialize slptime to start
424 if (flags & PNORESCHED)
425 td->td_flags |= TDF_NORESCHED;
426 p->p_usched->release_curproc(lp);
431 * Move our thread to the correct queue and setup our wchan, etc.
433 lwkt_deschedule_self(td);
434 td->td_flags |= TDF_TSLEEPQ;
435 TAILQ_INSERT_TAIL(&gd->gd_tsleep_hash[id], td, td_threadq);
436 atomic_set_int(&slpque_cpumasks[id], gd->gd_cpumask);
438 td->td_wchan = ident;
439 td->td_wmesg = wmesg;
440 td->td_wdomain = flags & PDOMAIN_MASK;
443 * Setup the timeout, if any
446 callout_init(&thandle);
447 callout_reset(&thandle, timo, endtsleep, td);
455 * Ok, we are sleeping. Place us in the SSLEEP state.
457 KKASSERT((p->p_flag & P_ONRUNQ) == 0);
459 lp->lwp_ru.ru_nvcsw++;
463 * And when we are woken up, put us back in SRUN. If we
464 * slept for over a second, recalculate our estcpu.
468 p->p_usched->recalculate(lp);
475 * Make sure we haven't switched cpus while we were asleep. It's
476 * not supposed to happen. Cleanup our temporary flags.
478 KKASSERT(gd == td->td_gd);
479 td->td_flags &= ~TDF_NORESCHED;
482 * Cleanup the timeout.
485 if (td->td_flags & TDF_TIMEOUT) {
486 td->td_flags &= ~TDF_TIMEOUT;
490 callout_stop(&thandle);
495 * Since td_threadq is used both for our run queue AND for the
496 * tsleep hash queue, we can't still be on it at this point because
497 * we've gotten cpu back.
499 KASSERT((td->td_flags & TDF_TSLEEPQ) == 0, ("tsleep: impossible thread flags %08x", td->td_flags));
505 * Figure out the correct error return. If interrupted by a
506 * signal we want to return EINTR or ERESTART.
508 * If P_MAILBOX is set no automatic system call restart occurs
509 * and we return EINTR. P_MAILBOX is meant to be used as an
510 * interlock, the user must poll it prior to any system call
511 * that it wishes to interlock a mailbox signal against since
512 * the flag is cleared on *any* system call that sleeps.
516 if (catch && error == 0) {
517 if ((p->p_flag & P_MAILBOX) && sig == 0) {
519 } else if (sig != 0 || (sig = CURSIG(lp))) {
520 if (SIGISMEMBER(p->p_sigacts->ps_sigintr, sig))
526 p->p_flag &= ~(P_BREAKTSLEEP | P_SINTR | P_MAILBOX);
528 logtsleep(tsleep_end);
534 * This is a dandy function that allows us to interlock tsleep/wakeup
535 * operations with unspecified upper level locks, such as lockmgr locks,
536 * simply by holding a critical section. The sequence is:
538 * (enter critical section)
539 * (acquire upper level lock)
540 * tsleep_interlock(blah)
541 * (release upper level lock)
543 * (exit critical section)
545 * Basically this function sets our cpumask for the ident which informs
546 * other cpus that our cpu 'might' be waiting (or about to wait on) the
547 * hash index related to the ident. The critical section prevents another
548 * cpu's wakeup() from being processed on our cpu until we are actually
549 * able to enter the tsleep(). Thus, no race occurs between our attempt
550 * to release a resource and sleep, and another cpu's attempt to acquire
551 * a resource and call wakeup.
553 * There isn't much of a point to this function unless you call it while
554 * holding a critical section.
557 _tsleep_interlock(globaldata_t gd, void *ident)
559 int id = LOOKUP(ident);
561 atomic_set_int(&slpque_cpumasks[id], gd->gd_cpumask);
565 tsleep_interlock(void *ident)
567 _tsleep_interlock(mycpu, ident);
571 * Interlocked spinlock sleep. An exclusively held spinlock must
572 * be passed to msleep(). The function will atomically release the
573 * spinlock and tsleep on the ident, then reacquire the spinlock and
576 * This routine is fairly important along the critical path, so optimize it
580 msleep(void *ident, struct spinlock *spin, int flags,
581 const char *wmesg, int timo)
583 globaldata_t gd = mycpu;
587 _tsleep_interlock(gd, ident);
588 spin_unlock_wr_quick(gd, spin);
589 error = tsleep(ident, flags, wmesg, timo);
590 spin_lock_wr_quick(gd, spin);
597 * Implement the timeout for tsleep.
599 * We set P_BREAKTSLEEP to indicate that an event has occured, but
600 * we only call setrunnable if the process is not stopped.
602 * This type of callout timeout is scheduled on the same cpu the process
603 * is sleeping on. Also, at the moment, the MP lock is held.
611 ASSERT_MP_LOCK_HELD(curthread);
615 * cpu interlock. Thread flags are only manipulated on
616 * the cpu owning the thread. proc flags are only manipulated
617 * by the older of the MP lock. We have both.
619 if (td->td_flags & TDF_TSLEEPQ) {
620 td->td_flags |= TDF_TIMEOUT;
622 if ((p = td->td_proc) != NULL) {
623 p->p_flag |= P_BREAKTSLEEP;
624 if ((p->p_flag & P_STOPPED) == 0)
627 unsleep_and_wakeup_thread(td);
634 * Unsleep and wakeup a thread. This function runs without the MP lock
635 * which means that it can only manipulate thread state on the owning cpu,
636 * and cannot touch the process state at all.
640 unsleep_and_wakeup_thread(struct thread *td)
642 globaldata_t gd = mycpu;
646 if (td->td_gd != gd) {
647 lwkt_send_ipiq(td->td_gd, (ipifunc1_t)unsleep_and_wakeup_thread, td);
652 if (td->td_flags & TDF_TSLEEPQ) {
653 td->td_flags &= ~TDF_TSLEEPQ;
654 id = LOOKUP(td->td_wchan);
655 TAILQ_REMOVE(&gd->gd_tsleep_hash[id], td, td_threadq);
656 if (TAILQ_FIRST(&gd->gd_tsleep_hash[id]) == NULL)
657 atomic_clear_int(&slpque_cpumasks[id], gd->gd_cpumask);
664 * Make all processes sleeping on the specified identifier runnable.
665 * count may be zero or one only.
667 * The domain encodes the sleep/wakeup domain AND the first cpu to check
668 * (which is always the current cpu). As we iterate across cpus
670 * This call may run without the MP lock held. We can only manipulate thread
671 * state on the cpu owning the thread. We CANNOT manipulate process state
675 _wakeup(void *ident, int domain)
690 logtsleep(wakeup_beg);
693 qp = &gd->gd_tsleep_hash[id];
695 for (td = TAILQ_FIRST(qp); td != NULL; td = ntd) {
696 ntd = TAILQ_NEXT(td, td_threadq);
697 if (td->td_wchan == ident &&
698 td->td_wdomain == (domain & PDOMAIN_MASK)
700 KKASSERT(td->td_flags & TDF_TSLEEPQ);
701 td->td_flags &= ~TDF_TSLEEPQ;
702 TAILQ_REMOVE(qp, td, td_threadq);
703 if (TAILQ_FIRST(qp) == NULL) {
704 atomic_clear_int(&slpque_cpumasks[id],
708 if (domain & PWAKEUP_ONE)
716 * We finished checking the current cpu but there still may be
717 * more work to do. Either wakeup_one was requested and no matching
718 * thread was found, or a normal wakeup was requested and we have
719 * to continue checking cpus.
721 * The cpu that started the wakeup sequence is encoded in the domain.
722 * We use this information to determine which cpus still need to be
723 * checked, locate a candidate cpu, and chain the wakeup
724 * asynchronously with an IPI message.
726 * It should be noted that this scheme is actually less expensive then
727 * the old scheme when waking up multiple threads, since we send
728 * only one IPI message per target candidate which may then schedule
729 * multiple threads. Before we could have wound up sending an IPI
730 * message for each thread on the target cpu (!= current cpu) that
731 * needed to be woken up.
733 * NOTE: Wakeups occuring on remote cpus are asynchronous. This
734 * should be ok since we are passing idents in the IPI rather then
737 if ((domain & PWAKEUP_MYCPU) == 0 &&
738 (mask = slpque_cpumasks[id]) != 0
741 * Look for a cpu that might have work to do. Mask out cpus
742 * which have already been processed.
744 * 31xxxxxxxxxxxxxxxxxxxxxxxxxxxxx0
746 * start currentcpu start
749 * 11111111111111110000000000000111 case1
750 * 00000000111111110000000000000000 case2
752 * case1: We started at start_case1 and processed through
753 * to the current cpu. We have to check any bits
754 * after the current cpu, then check bits before
757 * case2: We have already checked all the bits from
758 * start_case2 to the end, and from 0 to the current
759 * cpu. We just have the bits from the current cpu
760 * to start_case2 left to check.
762 startcpu = PWAKEUP_DECODE(domain);
763 if (gd->gd_cpuid >= startcpu) {
767 tmask = mask & ~((gd->gd_cpumask << 1) - 1);
769 nextcpu = bsfl(mask & tmask);
770 lwkt_send_ipiq2(globaldata_find(nextcpu),
771 _wakeup, ident, domain);
773 tmask = (1 << startcpu) - 1;
775 nextcpu = bsfl(mask & tmask);
777 globaldata_find(nextcpu),
778 _wakeup, ident, domain);
785 tmask = ~((gd->gd_cpumask << 1) - 1) &
786 ((1 << startcpu) - 1);
788 nextcpu = bsfl(mask & tmask);
789 lwkt_send_ipiq2(globaldata_find(nextcpu),
790 _wakeup, ident, domain);
796 logtsleep(wakeup_end);
801 * Wakeup all threads tsleep()ing on the specified ident, on all cpus
806 _wakeup(ident, PWAKEUP_ENCODE(0, mycpu->gd_cpuid));
810 * Wakeup one thread tsleep()ing on the specified ident, on any cpu.
813 wakeup_one(void *ident)
815 /* XXX potentially round-robin the first responding cpu */
816 _wakeup(ident, PWAKEUP_ENCODE(0, mycpu->gd_cpuid) | PWAKEUP_ONE);
820 * Wakeup threads tsleep()ing on the specified ident on the current cpu
824 wakeup_mycpu(void *ident)
826 _wakeup(ident, PWAKEUP_MYCPU);
830 * Wakeup one thread tsleep()ing on the specified ident on the current cpu
834 wakeup_mycpu_one(void *ident)
836 /* XXX potentially round-robin the first responding cpu */
837 _wakeup(ident, PWAKEUP_MYCPU|PWAKEUP_ONE);
841 * Wakeup all thread tsleep()ing on the specified ident on the specified cpu
845 wakeup_oncpu(globaldata_t gd, void *ident)
849 _wakeup(ident, PWAKEUP_MYCPU);
851 lwkt_send_ipiq2(gd, _wakeup, ident, PWAKEUP_MYCPU);
854 _wakeup(ident, PWAKEUP_MYCPU);
859 * Wakeup one thread tsleep()ing on the specified ident on the specified cpu
863 wakeup_oncpu_one(globaldata_t gd, void *ident)
867 _wakeup(ident, PWAKEUP_MYCPU | PWAKEUP_ONE);
869 lwkt_send_ipiq2(gd, _wakeup, ident, PWAKEUP_MYCPU | PWAKEUP_ONE);
872 _wakeup(ident, PWAKEUP_MYCPU | PWAKEUP_ONE);
877 * Wakeup all threads waiting on the specified ident that slept using
878 * the specified domain, on all cpus.
881 wakeup_domain(void *ident, int domain)
883 _wakeup(ident, PWAKEUP_ENCODE(domain, mycpu->gd_cpuid));
887 * Wakeup one thread waiting on the specified ident that slept using
888 * the specified domain, on any cpu.
891 wakeup_domain_one(void *ident, int domain)
893 /* XXX potentially round-robin the first responding cpu */
894 _wakeup(ident, PWAKEUP_ENCODE(domain, mycpu->gd_cpuid) | PWAKEUP_ONE);
900 * Make a process runnable. The MP lock must be held on call. This only
901 * has an effect if we are in SSLEEP. We only break out of the
902 * tsleep if P_BREAKTSLEEP is set, otherwise we just fix-up the state.
904 * NOTE: With the MP lock held we can only safely manipulate the process
905 * structure. We cannot safely manipulate the thread structure.
908 setrunnable(struct proc *p)
911 struct lwp *lp = FIRST_LWP_IN_PROC(p);
913 ASSERT_MP_LOCK_HELD(curthread);
914 p->p_flag &= ~P_STOPPED;
915 if (p->p_stat == SSLEEP && (p->p_flag & P_BREAKTSLEEP)) {
916 unsleep_and_wakeup_thread(lp->lwp_thread);
922 * The process is stopped due to some condition, usually because P_STOPPED
923 * is set but also possibly due to being traced.
925 * NOTE! If the caller sets P_STOPPED, the caller must also clear P_WAITED
926 * because the parent may check the child's status before the child actually
927 * gets to this routine.
929 * This routine is called with the current process only, typically just
930 * before returning to userland.
932 * Setting P_BREAKTSLEEP before entering the tsleep will cause a passive
933 * SIGCONT to break out of the tsleep.
936 tstop(struct proc *p)
938 wakeup((caddr_t)p->p_pptr);
939 p->p_flag |= P_BREAKTSLEEP;
940 tsleep(p, 0, "stop", 0);
944 * Yield / synchronous reschedule. This is a bit tricky because the trap
945 * code might have set a lazy release on the switch function. Setting
946 * P_PASSIVE_ACQ will ensure that the lazy release executes when we call
947 * switch, and that we are given a greater chance of affinity with our
950 * We call lwkt_setpri_self() to rotate our thread to the end of the lwkt
951 * run queue. lwkt_switch() will also execute any assigned passive release
952 * (which usually calls release_curproc()), allowing a same/higher priority
953 * process to be designated as the current process.
955 * While it is possible for a lower priority process to be designated,
956 * it's call to lwkt_maybe_switch() in acquire_curproc() will likely
957 * round-robin back to us and we will be able to re-acquire the current
958 * process designation.
963 struct thread *td = curthread;
964 struct proc *p = td->td_proc;
966 lwkt_setpri_self(td->td_pri & TDPRI_MASK);
968 p->p_flag |= P_PASSIVE_ACQ;
970 p->p_flag &= ~P_PASSIVE_ACQ;
977 * Compute a tenex style load average of a quantity on
978 * 1, 5 and 15 minute intervals.
980 static int loadav_count_runnable(struct proc *p, void *data);
989 allproc_scan(loadav_count_runnable, &nrun);
991 for (i = 0; i < 3; i++) {
992 avg->ldavg[i] = (cexp[i] * avg->ldavg[i] +
993 nrun * FSCALE * (FSCALE - cexp[i])) >> FSHIFT;
997 * Schedule the next update to occur after 5 seconds, but add a
998 * random variation to avoid synchronisation with processes that
999 * run at regular intervals.
1001 callout_reset(&loadav_callout, hz * 4 + (int)(krandom() % (hz * 2 + 1)),
1006 loadav_count_runnable(struct proc *p, void *data)
1013 lp = FIRST_LWP_IN_PROC(p);
1014 switch (p->p_stat) {
1016 if ((td = lp->lwp_thread) == NULL)
1018 if (td->td_flags & TDF_BLOCKED)
1032 sched_setup(void *dummy)
1034 callout_init(&loadav_callout);
1035 callout_init(&schedcpu_callout);
1037 /* Kick off timeout driven events by calling first time. */