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 * 4. Neither the name of the University nor the names of its contributors
19 * may be used to endorse or promote products derived from this software
20 * without specific prior written permission.
22 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
23 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
24 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
25 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
26 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
27 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
28 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
29 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
30 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
31 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
34 * @(#)kern_synch.c 8.9 (Berkeley) 5/19/95
35 * $FreeBSD: src/sys/kern/kern_synch.c,v 1.87.2.6 2002/10/13 07:29:53 kbyanc Exp $
38 #include "opt_ktrace.h"
40 #include <sys/param.h>
41 #include <sys/systm.h>
43 #include <sys/kernel.h>
44 #include <sys/signalvar.h>
45 #include <sys/resourcevar.h>
46 #include <sys/vmmeter.h>
47 #include <sys/sysctl.h>
51 #include <sys/ktrace.h>
53 #include <sys/xwait.h>
55 #include <sys/serialize.h>
57 #include <sys/signal2.h>
58 #include <sys/thread2.h>
59 #include <sys/spinlock2.h>
60 #include <sys/mutex2.h>
62 #include <machine/cpu.h>
63 #include <machine/smp.h>
65 TAILQ_HEAD(tslpque, thread);
67 static void sched_setup (void *dummy);
68 SYSINIT(sched_setup, SI_SUB_KICK_SCHEDULER, SI_ORDER_FIRST, sched_setup, NULL)
73 int sched_quantum; /* Roundrobin scheduling quantum in ticks. */
75 int ncpus2, ncpus2_shift, ncpus2_mask; /* note: mask not cpumask_t */
76 int ncpus_fit, ncpus_fit_mask; /* note: mask not cpumask_t */
79 int tsleep_crypto_dump = 0;
81 static struct callout loadav_callout;
82 static struct callout schedcpu_callout;
83 MALLOC_DEFINE(M_TSLEEP, "tslpque", "tsleep queues");
85 #define __DEALL(ident) __DEQUALIFY(void *, ident)
87 #if !defined(KTR_TSLEEP)
88 #define KTR_TSLEEP KTR_ALL
90 KTR_INFO_MASTER(tsleep);
91 KTR_INFO(KTR_TSLEEP, tsleep, tsleep_beg, 0, "tsleep enter %p", const volatile void *ident);
92 KTR_INFO(KTR_TSLEEP, tsleep, tsleep_end, 1, "tsleep exit");
93 KTR_INFO(KTR_TSLEEP, tsleep, wakeup_beg, 2, "wakeup enter %p", const volatile void *ident);
94 KTR_INFO(KTR_TSLEEP, tsleep, wakeup_end, 3, "wakeup exit");
95 KTR_INFO(KTR_TSLEEP, tsleep, ilockfail, 4, "interlock failed %p", const volatile void *ident);
97 #define logtsleep1(name) KTR_LOG(tsleep_ ## name)
98 #define logtsleep2(name, val) KTR_LOG(tsleep_ ## name, val)
100 struct loadavg averunnable =
101 { {0, 0, 0}, FSCALE }; /* load average, of runnable procs */
103 * Constants for averages over 1, 5, and 15 minutes
104 * when sampling at 5 second intervals.
106 static fixpt_t cexp[3] = {
107 0.9200444146293232 * FSCALE, /* exp(-1/12) */
108 0.9834714538216174 * FSCALE, /* exp(-1/60) */
109 0.9944598480048967 * FSCALE, /* exp(-1/180) */
112 static void endtsleep (void *);
113 static void loadav (void *arg);
114 static void schedcpu (void *arg);
117 * Adjust the scheduler quantum. The quantum is specified in microseconds.
118 * Note that 'tick' is in microseconds per tick.
121 sysctl_kern_quantum(SYSCTL_HANDLER_ARGS)
125 new_val = sched_quantum * ustick;
126 error = sysctl_handle_int(oidp, &new_val, 0, req);
127 if (error != 0 || req->newptr == NULL)
129 if (new_val < ustick)
131 sched_quantum = new_val / ustick;
132 hogticks = 2 * sched_quantum;
136 SYSCTL_PROC(_kern, OID_AUTO, quantum, CTLTYPE_INT|CTLFLAG_RW,
137 0, sizeof sched_quantum, sysctl_kern_quantum, "I", "");
139 static int pctcpu_decay = 10;
140 SYSCTL_INT(_kern, OID_AUTO, pctcpu_decay, CTLFLAG_RW, &pctcpu_decay, 0, "");
143 * kernel uses `FSCALE', userland (SHOULD) use kern.fscale
145 int fscale __unused = FSCALE; /* exported to systat */
146 SYSCTL_INT(_kern, OID_AUTO, fscale, CTLFLAG_RD, 0, FSCALE, "");
149 * Recompute process priorities, once a second.
151 * Since the userland schedulers are typically event oriented, if the
152 * estcpu calculation at wakeup() time is not sufficient to make a
153 * process runnable relative to other processes in the system we have
154 * a 1-second recalc to help out.
156 * This code also allows us to store sysclock_t data in the process structure
157 * without fear of an overrun, since sysclock_t are guarenteed to hold
158 * several seconds worth of count.
160 * WARNING! callouts can preempt normal threads. However, they will not
161 * preempt a thread holding a spinlock so we *can* safely use spinlocks.
163 static int schedcpu_stats(struct proc *p, void *data __unused);
164 static int schedcpu_resource(struct proc *p, void *data __unused);
169 allproc_scan(schedcpu_stats, NULL);
170 allproc_scan(schedcpu_resource, NULL);
171 wakeup((caddr_t)&lbolt);
172 wakeup(lbolt_syncer);
173 callout_reset(&schedcpu_callout, hz, schedcpu, NULL);
177 * General process statistics once a second
180 schedcpu_stats(struct proc *p, void *data __unused)
185 * Threads may not be completely set up if process in SIDL state.
187 if (p->p_stat == SIDL)
191 if (lwkt_trytoken(&p->p_token) == FALSE) {
197 FOREACH_LWP_IN_PROC(lp, p) {
198 if (lp->lwp_stat == LSSLEEP) {
200 if (lp->lwp_slptime == 1)
201 p->p_usched->uload_update(lp);
205 * Only recalculate processes that are active or have slept
206 * less then 2 seconds. The schedulers understand this.
207 * Otherwise decay by 50% per second.
209 if (lp->lwp_slptime <= 1) {
210 p->p_usched->recalculate(lp);
214 decay = pctcpu_decay;
220 lp->lwp_pctcpu = (lp->lwp_pctcpu * (decay - 1)) / decay;
223 lwkt_reltoken(&p->p_token);
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)
240 if (p->p_stat == SIDL)
244 if (lwkt_trytoken(&p->p_token) == FALSE) {
249 if (p->p_stat == SZOMB || p->p_limit == NULL) {
250 lwkt_reltoken(&p->p_token);
256 FOREACH_LWP_IN_PROC(lp, p) {
258 * We may have caught an lp in the middle of being
259 * created, lwp_thread can be NULL.
261 if (lp->lwp_thread) {
262 ttime += lp->lwp_thread->td_sticks;
263 ttime += lp->lwp_thread->td_uticks;
267 switch(plimit_testcpulimit(p->p_limit, ttime)) {
268 case PLIMIT_TESTCPU_KILL:
269 killproc(p, "exceeded maximum CPU limit");
271 case PLIMIT_TESTCPU_XCPU:
272 if ((p->p_flags & P_XCPU) == 0) {
273 p->p_flags |= P_XCPU;
280 lwkt_reltoken(&p->p_token);
287 * This is only used by ps. Generate a cpu percentage use over
288 * a period of one second.
291 updatepcpu(struct lwp *lp, int cpticks, int ttlticks)
296 acc = (cpticks << FSHIFT) / ttlticks;
297 if (ttlticks >= ESTCPUFREQ) {
298 lp->lwp_pctcpu = acc;
300 remticks = ESTCPUFREQ - ttlticks;
301 lp->lwp_pctcpu = (acc * ttlticks + lp->lwp_pctcpu * remticks) /
307 * tsleep/wakeup hash table parameters. Try to find the sweet spot for
308 * like addresses being slept on.
310 #define TABLESIZE 4001
311 #define LOOKUP(x) (((u_int)(uintptr_t)(x)) % TABLESIZE)
313 static cpumask_t slpque_cpumasks[TABLESIZE];
316 * General scheduler initialization. We force a reschedule 25 times
317 * a second by default. Note that cpu0 is initialized in early boot and
318 * cannot make any high level calls.
320 * Each cpu has its own sleep queue.
323 sleep_gdinit(globaldata_t gd)
325 static struct tslpque slpque_cpu0[TABLESIZE];
328 if (gd->gd_cpuid == 0) {
329 sched_quantum = (hz + 24) / 25;
330 hogticks = 2 * sched_quantum;
332 gd->gd_tsleep_hash = slpque_cpu0;
334 gd->gd_tsleep_hash = kmalloc(sizeof(slpque_cpu0),
335 M_TSLEEP, M_WAITOK | M_ZERO);
337 for (i = 0; i < TABLESIZE; ++i)
338 TAILQ_INIT(&gd->gd_tsleep_hash[i]);
342 * This is a dandy function that allows us to interlock tsleep/wakeup
343 * operations with unspecified upper level locks, such as lockmgr locks,
344 * simply by holding a critical section. The sequence is:
346 * (acquire upper level lock)
347 * tsleep_interlock(blah)
348 * (release upper level lock)
351 * Basically this functions queues us on the tsleep queue without actually
352 * descheduling us. When tsleep() is later called with PINTERLOCK it
353 * assumes the thread was already queued, otherwise it queues it there.
355 * Thus it is possible to receive the wakeup prior to going to sleep and
356 * the race conditions are covered.
359 _tsleep_interlock(globaldata_t gd, const volatile void *ident, int flags)
361 thread_t td = gd->gd_curthread;
364 crit_enter_quick(td);
365 if (td->td_flags & TDF_TSLEEPQ) {
366 id = LOOKUP(td->td_wchan);
367 TAILQ_REMOVE(&gd->gd_tsleep_hash[id], td, td_sleepq);
368 if (TAILQ_FIRST(&gd->gd_tsleep_hash[id]) == NULL) {
369 atomic_clear_cpumask(&slpque_cpumasks[id],
373 td->td_flags |= TDF_TSLEEPQ;
376 TAILQ_INSERT_TAIL(&gd->gd_tsleep_hash[id], td, td_sleepq);
377 atomic_set_cpumask(&slpque_cpumasks[id], gd->gd_cpumask);
378 td->td_wchan = ident;
379 td->td_wdomain = flags & PDOMAIN_MASK;
384 tsleep_interlock(const volatile void *ident, int flags)
386 _tsleep_interlock(mycpu, ident, flags);
390 * Remove thread from sleepq. Must be called with a critical section held.
391 * The thread must not be migrating.
394 _tsleep_remove(thread_t td)
396 globaldata_t gd = mycpu;
399 KKASSERT(td->td_gd == gd && IN_CRITICAL_SECT(td));
400 KKASSERT((td->td_flags & TDF_MIGRATING) == 0);
401 if (td->td_flags & TDF_TSLEEPQ) {
402 td->td_flags &= ~TDF_TSLEEPQ;
403 id = LOOKUP(td->td_wchan);
404 TAILQ_REMOVE(&gd->gd_tsleep_hash[id], td, td_sleepq);
405 if (TAILQ_FIRST(&gd->gd_tsleep_hash[id]) == NULL)
406 atomic_clear_cpumask(&slpque_cpumasks[id], gd->gd_cpumask);
413 tsleep_remove(thread_t td)
419 * General sleep call. Suspends the current process until a wakeup is
420 * performed on the specified identifier. The process will then be made
421 * runnable with the specified priority. Sleeps at most timo/hz seconds
422 * (0 means no timeout). If flags includes PCATCH flag, signals are checked
423 * before and after sleeping, else signals are not checked. Returns 0 if
424 * awakened, EWOULDBLOCK if the timeout expires. If PCATCH is set and a
425 * signal needs to be delivered, ERESTART is returned if the current system
426 * call should be restarted if possible, and EINTR is returned if the system
427 * call should be interrupted by the signal (return EINTR).
429 * Note that if we are a process, we release_curproc() before messing with
430 * the LWKT scheduler.
432 * During autoconfiguration or after a panic, a sleep will simply
433 * lower the priority briefly to allow interrupts, then return.
435 * WARNING! This code can't block (short of switching away), or bad things
436 * will happen. No getting tokens, no blocking locks, etc.
439 tsleep(const volatile void *ident, int flags, const char *wmesg, int timo)
441 struct thread *td = curthread;
442 struct lwp *lp = td->td_lwp;
443 struct proc *p = td->td_proc; /* may be NULL */
449 struct callout thandle;
452 * Currently a severe hack. Make sure any delayed wakeups
453 * are flushed before we sleep or we might deadlock on whatever
454 * event we are sleeping on.
456 if (td->td_flags & TDF_DELAYED_WAKEUP)
457 wakeup_end_delayed();
460 * NOTE: removed KTRPOINT, it could cause races due to blocking
461 * even in stable. Just scrap it for now.
463 if (!tsleep_crypto_dump && (tsleep_now_works == 0 || panicstr)) {
465 * After a panic, or before we actually have an operational
466 * softclock, just give interrupts a chance, then just return;
468 * don't run any other procs or panic below,
469 * in case this is the idle process and already asleep.
473 lwkt_setpri_self(safepri);
475 lwkt_setpri_self(oldpri);
478 logtsleep2(tsleep_beg, ident);
480 KKASSERT(td != &gd->gd_idlethread); /* you must be kidding! */
481 td->td_wakefromcpu = -1; /* overwritten by _wakeup */
484 * NOTE: all of this occurs on the current cpu, including any
485 * callout-based wakeups, so a critical section is a sufficient
488 * The entire sequence through to where we actually sleep must
489 * run without breaking the critical section.
491 catch = flags & PCATCH;
495 crit_enter_quick(td);
497 KASSERT(ident != NULL, ("tsleep: no ident"));
498 KASSERT(lp == NULL ||
499 lp->lwp_stat == LSRUN || /* Obvious */
500 lp->lwp_stat == LSSTOP, /* Set in tstop */
502 ident, wmesg, lp->lwp_stat));
505 * We interlock the sleep queue if the caller has not already done
506 * it for us. This must be done before we potentially acquire any
507 * tokens or we can loose the wakeup.
509 if ((flags & PINTERLOCKED) == 0) {
510 _tsleep_interlock(gd, ident, flags);
514 * Setup for the current process (if this is a process). We must
515 * interlock with lwp_token to avoid remote wakeup races via
519 lwkt_gettoken(&lp->lwp_token);
522 * Early termination if PCATCH was set and a
523 * signal is pending, interlocked with the
526 * Early termination only occurs when tsleep() is
527 * entered while in a normal LSRUN state.
529 if ((sig = CURSIG(lp)) != 0)
533 * Causes ksignal to wake us up if a signal is
534 * received (interlocked with p->p_token).
536 lp->lwp_flags |= LWP_SINTR;
543 * Make sure the current process has been untangled from
544 * the userland scheduler and initialize slptime to start
547 * NOTE: td->td_wakefromcpu is pre-set by the release function
548 * for the dfly scheduler, and then adjusted by _wakeup()
551 p->p_usched->release_curproc(lp);
556 * If the interlocked flag is set but our cpu bit in the slpqueue
557 * is no longer set, then a wakeup was processed inbetween the
558 * tsleep_interlock() (ours or the callers), and here. This can
559 * occur under numerous circumstances including when we release the
562 * Extreme loads can cause the sending of an IPI (e.g. wakeup()'s)
563 * to process incoming IPIs, thus draining incoming wakeups.
565 if ((td->td_flags & TDF_TSLEEPQ) == 0) {
566 logtsleep2(ilockfail, ident);
571 * scheduling is blocked while in a critical section. Coincide
572 * the descheduled-by-tsleep flag with the descheduling of the
575 * The timer callout is localized on our cpu and interlocked by
576 * our critical section.
578 lwkt_deschedule_self(td);
579 td->td_flags |= TDF_TSLEEP_DESCHEDULED;
580 td->td_wmesg = wmesg;
583 * Setup the timeout, if any. The timeout is only operable while
584 * the thread is flagged descheduled.
586 KKASSERT((td->td_flags & TDF_TIMEOUT) == 0);
588 callout_init_mp(&thandle);
589 callout_reset(&thandle, timo, endtsleep, td);
597 * Ok, we are sleeping. Place us in the SSLEEP state.
599 KKASSERT((lp->lwp_mpflags & LWP_MP_ONRUNQ) == 0);
602 * tstop() sets LSSTOP, so don't fiddle with that.
604 if (lp->lwp_stat != LSSTOP)
605 lp->lwp_stat = LSSLEEP;
606 lp->lwp_ru.ru_nvcsw++;
607 p->p_usched->uload_update(lp);
611 * And when we are woken up, put us back in LSRUN. If we
612 * slept for over a second, recalculate our estcpu.
614 lp->lwp_stat = LSRUN;
615 if (lp->lwp_slptime) {
616 p->p_usched->uload_update(lp);
617 p->p_usched->recalculate(lp);
625 * Make sure we haven't switched cpus while we were asleep. It's
626 * not supposed to happen. Cleanup our temporary flags.
628 KKASSERT(gd == td->td_gd);
631 * Cleanup the timeout. If the timeout has already occured thandle
632 * has already been stopped, otherwise stop thandle. If the timeout
633 * is running (the callout thread must be blocked trying to get
634 * lwp_token) then wait for us to get scheduled.
637 while (td->td_flags & TDF_TIMEOUT_RUNNING) {
638 lwkt_deschedule_self(td);
639 td->td_wmesg = "tsrace";
641 kprintf("td %p %s: timeout race\n", td, td->td_comm);
643 if (td->td_flags & TDF_TIMEOUT) {
644 td->td_flags &= ~TDF_TIMEOUT;
647 /* does not block when on same cpu */
648 callout_stop(&thandle);
651 td->td_flags &= ~TDF_TSLEEP_DESCHEDULED;
654 * Make sure we have been removed from the sleepq. In most
655 * cases this will have been done for us already but it is
656 * possible for a scheduling IPI to be in-flight from a
657 * previous tsleep/tsleep_interlock() or due to a straight-out
658 * call to lwkt_schedule() (in the case of an interrupt thread),
659 * causing a spurious wakeup.
665 * Figure out the correct error return. If interrupted by a
666 * signal we want to return EINTR or ERESTART.
670 if (catch && error == 0) {
671 if (sig != 0 || (sig = CURSIG(lp))) {
672 if (SIGISMEMBER(p->p_sigacts->ps_sigintr, sig))
678 lp->lwp_flags &= ~LWP_SINTR;
679 lwkt_reltoken(&lp->lwp_token);
681 logtsleep1(tsleep_end);
687 * Interlocked spinlock sleep. An exclusively held spinlock must
688 * be passed to ssleep(). The function will atomically release the
689 * spinlock and tsleep on the ident, then reacquire the spinlock and
692 * This routine is fairly important along the critical path, so optimize it
696 ssleep(const volatile void *ident, struct spinlock *spin, int flags,
697 const char *wmesg, int timo)
699 globaldata_t gd = mycpu;
702 _tsleep_interlock(gd, ident, flags);
703 spin_unlock_quick(gd, spin);
704 error = tsleep(ident, flags | PINTERLOCKED, wmesg, timo);
705 spin_lock_quick(gd, spin);
711 lksleep(const volatile void *ident, struct lock *lock, int flags,
712 const char *wmesg, int timo)
714 globaldata_t gd = mycpu;
717 _tsleep_interlock(gd, ident, flags);
718 lockmgr(lock, LK_RELEASE);
719 error = tsleep(ident, flags | PINTERLOCKED, wmesg, timo);
720 lockmgr(lock, LK_EXCLUSIVE);
726 * Interlocked mutex sleep. An exclusively held mutex must be passed
727 * to mtxsleep(). The function will atomically release the mutex
728 * and tsleep on the ident, then reacquire the mutex and return.
731 mtxsleep(const volatile void *ident, struct mtx *mtx, int flags,
732 const char *wmesg, int timo)
734 globaldata_t gd = mycpu;
737 _tsleep_interlock(gd, ident, flags);
739 error = tsleep(ident, flags | PINTERLOCKED, wmesg, timo);
740 mtx_lock_ex_quick(mtx, wmesg);
746 * Interlocked serializer sleep. An exclusively held serializer must
747 * be passed to zsleep(). The function will atomically release
748 * the serializer and tsleep on the ident, then reacquire the serializer
752 zsleep(const volatile void *ident, struct lwkt_serialize *slz, int flags,
753 const char *wmesg, int timo)
755 globaldata_t gd = mycpu;
758 ASSERT_SERIALIZED(slz);
760 _tsleep_interlock(gd, ident, flags);
761 lwkt_serialize_exit(slz);
762 ret = tsleep(ident, flags | PINTERLOCKED, wmesg, timo);
763 lwkt_serialize_enter(slz);
769 * Directly block on the LWKT thread by descheduling it. This
770 * is much faster then tsleep(), but the only legal way to wake
771 * us up is to directly schedule the thread.
773 * Setting TDF_SINTR will cause new signals to directly schedule us.
775 * This routine must be called while in a critical section.
778 lwkt_sleep(const char *wmesg, int flags)
780 thread_t td = curthread;
783 if ((flags & PCATCH) == 0 || td->td_lwp == NULL) {
784 td->td_flags |= TDF_BLOCKED;
785 td->td_wmesg = wmesg;
786 lwkt_deschedule_self(td);
789 td->td_flags &= ~TDF_BLOCKED;
792 if ((sig = CURSIG(td->td_lwp)) != 0) {
793 if (SIGISMEMBER(td->td_proc->p_sigacts->ps_sigintr, sig))
799 td->td_flags |= TDF_BLOCKED | TDF_SINTR;
800 td->td_wmesg = wmesg;
801 lwkt_deschedule_self(td);
803 td->td_flags &= ~(TDF_BLOCKED | TDF_SINTR);
809 * Implement the timeout for tsleep.
811 * This type of callout timeout is scheduled on the same cpu the process
812 * is sleeping on. Also, at the moment, the MP lock is held.
821 * We are going to have to get the lwp_token, which means we might
822 * block. This can race a tsleep getting woken up by other means
823 * so set TDF_TIMEOUT_RUNNING to force the tsleep to wait for our
824 * processing to complete (sorry tsleep!).
826 * We can safely set td_flags because td MUST be on the same cpu
829 KKASSERT(td->td_gd == mycpu);
831 td->td_flags |= TDF_TIMEOUT_RUNNING | TDF_TIMEOUT;
834 * This can block but TDF_TIMEOUT_RUNNING will prevent the thread
835 * from exiting the tsleep on us. The flag is interlocked by virtue
836 * of lp being on the same cpu as we are.
838 if ((lp = td->td_lwp) != NULL)
839 lwkt_gettoken(&lp->lwp_token);
841 KKASSERT(td->td_flags & TDF_TSLEEP_DESCHEDULED);
844 if (lp->lwp_proc->p_stat != SSTOP)
846 lwkt_reltoken(&lp->lwp_token);
851 KKASSERT(td->td_gd == mycpu);
852 td->td_flags &= ~TDF_TIMEOUT_RUNNING;
857 * Make all processes sleeping on the specified identifier runnable.
858 * count may be zero or one only.
860 * The domain encodes the sleep/wakeup domain, flags, plus the originating
863 * This call may run without the MP lock held. We can only manipulate thread
864 * state on the cpu owning the thread. We CANNOT manipulate process state
867 * _wakeup() can be passed to an IPI so we can't use (const volatile
871 _wakeup(void *ident, int domain)
881 logtsleep2(wakeup_beg, ident);
884 qp = &gd->gd_tsleep_hash[id];
886 for (td = TAILQ_FIRST(qp); td != NULL; td = ntd) {
887 ntd = TAILQ_NEXT(td, td_sleepq);
888 if (td->td_wchan == ident &&
889 td->td_wdomain == (domain & PDOMAIN_MASK)
891 KKASSERT(td->td_gd == gd);
893 td->td_wakefromcpu = PWAKEUP_DECODE(domain);
894 if (td->td_flags & TDF_TSLEEP_DESCHEDULED) {
896 if (domain & PWAKEUP_ONE)
904 * We finished checking the current cpu but there still may be
905 * more work to do. Either wakeup_one was requested and no matching
906 * thread was found, or a normal wakeup was requested and we have
907 * to continue checking cpus.
909 * It should be noted that this scheme is actually less expensive then
910 * the old scheme when waking up multiple threads, since we send
911 * only one IPI message per target candidate which may then schedule
912 * multiple threads. Before we could have wound up sending an IPI
913 * message for each thread on the target cpu (!= current cpu) that
914 * needed to be woken up.
916 * NOTE: Wakeups occuring on remote cpus are asynchronous. This
917 * should be ok since we are passing idents in the IPI rather then
920 if ((domain & PWAKEUP_MYCPU) == 0 &&
921 (mask = slpque_cpumasks[id] & gd->gd_other_cpus) != 0) {
922 lwkt_send_ipiq2_mask(mask, _wakeup, ident,
923 domain | PWAKEUP_MYCPU);
926 logtsleep1(wakeup_end);
931 * Wakeup all threads tsleep()ing on the specified ident, on all cpus
934 wakeup(const volatile void *ident)
936 globaldata_t gd = mycpu;
937 thread_t td = gd->gd_curthread;
939 if (td && (td->td_flags & TDF_DELAYED_WAKEUP)) {
940 if (!atomic_cmpset_ptr(&gd->gd_delayed_wakeup[0], NULL, ident)) {
941 if (!atomic_cmpset_ptr(&gd->gd_delayed_wakeup[1], NULL, ident))
942 _wakeup(__DEALL(ident), PWAKEUP_ENCODE(0, gd->gd_cpuid));
946 _wakeup(__DEALL(ident), PWAKEUP_ENCODE(0, gd->gd_cpuid));
950 * Wakeup one thread tsleep()ing on the specified ident, on any cpu.
953 wakeup_one(const volatile void *ident)
955 /* XXX potentially round-robin the first responding cpu */
956 _wakeup(__DEALL(ident), PWAKEUP_ENCODE(0, mycpu->gd_cpuid) |
961 * Wakeup threads tsleep()ing on the specified ident on the current cpu
965 wakeup_mycpu(const volatile void *ident)
967 _wakeup(__DEALL(ident), PWAKEUP_ENCODE(0, mycpu->gd_cpuid) |
972 * Wakeup one thread tsleep()ing on the specified ident on the current cpu
976 wakeup_mycpu_one(const volatile void *ident)
978 /* XXX potentially round-robin the first responding cpu */
979 _wakeup(__DEALL(ident), PWAKEUP_ENCODE(0, mycpu->gd_cpuid) |
980 PWAKEUP_MYCPU | PWAKEUP_ONE);
984 * Wakeup all thread tsleep()ing on the specified ident on the specified cpu
988 wakeup_oncpu(globaldata_t gd, const volatile void *ident)
990 globaldata_t mygd = mycpu;
992 _wakeup(__DEALL(ident), PWAKEUP_ENCODE(0, mygd->gd_cpuid) |
995 lwkt_send_ipiq2(gd, _wakeup, __DEALL(ident),
996 PWAKEUP_ENCODE(0, mygd->gd_cpuid) |
1002 * Wakeup one thread tsleep()ing on the specified ident on the specified cpu
1006 wakeup_oncpu_one(globaldata_t gd, const volatile void *ident)
1008 globaldata_t mygd = mycpu;
1010 _wakeup(__DEALL(ident), PWAKEUP_ENCODE(0, mygd->gd_cpuid) |
1011 PWAKEUP_MYCPU | PWAKEUP_ONE);
1013 lwkt_send_ipiq2(gd, _wakeup, __DEALL(ident),
1014 PWAKEUP_ENCODE(0, mygd->gd_cpuid) |
1015 PWAKEUP_MYCPU | PWAKEUP_ONE);
1020 * Wakeup all threads waiting on the specified ident that slept using
1021 * the specified domain, on all cpus.
1024 wakeup_domain(const volatile void *ident, int domain)
1026 _wakeup(__DEALL(ident), PWAKEUP_ENCODE(domain, mycpu->gd_cpuid));
1030 * Wakeup one thread waiting on the specified ident that slept using
1031 * the specified domain, on any cpu.
1034 wakeup_domain_one(const volatile void *ident, int domain)
1036 /* XXX potentially round-robin the first responding cpu */
1037 _wakeup(__DEALL(ident),
1038 PWAKEUP_ENCODE(domain, mycpu->gd_cpuid) | PWAKEUP_ONE);
1042 wakeup_start_delayed(void)
1044 globaldata_t gd = mycpu;
1047 gd->gd_curthread->td_flags |= TDF_DELAYED_WAKEUP;
1052 wakeup_end_delayed(void)
1054 globaldata_t gd = mycpu;
1056 if (gd->gd_curthread->td_flags & TDF_DELAYED_WAKEUP) {
1058 gd->gd_curthread->td_flags &= ~TDF_DELAYED_WAKEUP;
1059 if (gd->gd_delayed_wakeup[0] || gd->gd_delayed_wakeup[1]) {
1060 if (gd->gd_delayed_wakeup[0]) {
1061 wakeup(gd->gd_delayed_wakeup[0]);
1062 gd->gd_delayed_wakeup[0] = NULL;
1064 if (gd->gd_delayed_wakeup[1]) {
1065 wakeup(gd->gd_delayed_wakeup[1]);
1066 gd->gd_delayed_wakeup[1] = NULL;
1076 * Make a process runnable. lp->lwp_token must be held on call and this
1077 * function must be called from the cpu owning lp.
1079 * This only has an effect if we are in LSSTOP or LSSLEEP.
1082 setrunnable(struct lwp *lp)
1084 thread_t td = lp->lwp_thread;
1086 ASSERT_LWKT_TOKEN_HELD(&lp->lwp_token);
1087 KKASSERT(td->td_gd == mycpu);
1089 if (lp->lwp_stat == LSSTOP)
1090 lp->lwp_stat = LSSLEEP;
1091 if (lp->lwp_stat == LSSLEEP) {
1094 } else if (td->td_flags & TDF_SINTR) {
1101 * The process is stopped due to some condition, usually because p_stat is
1102 * set to SSTOP, but also possibly due to being traced.
1104 * Caller must hold p->p_token
1106 * NOTE! If the caller sets SSTOP, the caller must also clear P_WAITED
1107 * because the parent may check the child's status before the child actually
1108 * gets to this routine.
1110 * This routine is called with the current lwp only, typically just
1111 * before returning to userland if the process state is detected as
1112 * possibly being in a stopped state.
1117 struct lwp *lp = curthread->td_lwp;
1118 struct proc *p = lp->lwp_proc;
1121 lwkt_gettoken(&lp->lwp_token);
1125 * If LWP_MP_WSTOP is set, we were sleeping
1126 * while our process was stopped. At this point
1127 * we were already counted as stopped.
1129 if ((lp->lwp_mpflags & LWP_MP_WSTOP) == 0) {
1131 * If we're the last thread to stop, signal
1135 atomic_set_int(&lp->lwp_mpflags, LWP_MP_WSTOP);
1136 wakeup(&p->p_nstopped);
1137 if (p->p_nstopped == p->p_nthreads) {
1139 * Token required to interlock kern_wait()
1143 lwkt_gettoken(&q->p_token);
1144 p->p_flags &= ~P_WAITED;
1146 if ((q->p_sigacts->ps_flag & PS_NOCLDSTOP) == 0)
1147 ksignal(q, SIGCHLD);
1148 lwkt_reltoken(&q->p_token);
1152 while (p->p_stat == SSTOP) {
1153 lp->lwp_stat = LSSTOP;
1154 tsleep(p, 0, "stop", 0);
1157 atomic_clear_int(&lp->lwp_mpflags, LWP_MP_WSTOP);
1159 lwkt_reltoken(&lp->lwp_token);
1163 * Compute a tenex style load average of a quantity on
1164 * 1, 5 and 15 minute intervals.
1166 static int loadav_count_runnable(struct lwp *p, void *data);
1171 struct loadavg *avg;
1175 alllwp_scan(loadav_count_runnable, &nrun);
1177 for (i = 0; i < 3; i++) {
1178 avg->ldavg[i] = (cexp[i] * avg->ldavg[i] +
1179 nrun * FSCALE * (FSCALE - cexp[i])) >> FSHIFT;
1183 * Schedule the next update to occur after 5 seconds, but add a
1184 * random variation to avoid synchronisation with processes that
1185 * run at regular intervals.
1187 callout_reset(&loadav_callout, hz * 4 + (int)(krandom() % (hz * 2 + 1)),
1192 loadav_count_runnable(struct lwp *lp, void *data)
1197 switch (lp->lwp_stat) {
1199 if ((td = lp->lwp_thread) == NULL)
1201 if (td->td_flags & TDF_BLOCKED)
1214 sched_setup(void *dummy)
1216 callout_init_mp(&loadav_callout);
1217 callout_init_mp(&schedcpu_callout);
1219 /* Kick off timeout driven events by calling first time. */