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>
67 #include <machine/cpu.h>
68 #include <machine/smp.h>
70 TAILQ_HEAD(tslpque, thread);
72 static void sched_setup (void *dummy);
73 SYSINIT(sched_setup, SI_SUB_KICK_SCHEDULER, SI_ORDER_FIRST, sched_setup, NULL)
78 int sched_quantum; /* Roundrobin scheduling quantum in ticks. */
80 int ncpus2, ncpus2_shift, ncpus2_mask; /* note: mask not cpumask_t */
81 int ncpus_fit, ncpus_fit_mask; /* note: mask not cpumask_t */
84 int tsleep_crypto_dump = 0;
86 static struct callout loadav_callout;
87 static struct callout schedcpu_callout;
88 MALLOC_DEFINE(M_TSLEEP, "tslpque", "tsleep queues");
90 #define __DEALL(ident) __DEQUALIFY(void *, ident)
92 #if !defined(KTR_TSLEEP)
93 #define KTR_TSLEEP KTR_ALL
95 KTR_INFO_MASTER(tsleep);
96 KTR_INFO(KTR_TSLEEP, tsleep, tsleep_beg, 0, "tsleep enter %p", const volatile void *ident);
97 KTR_INFO(KTR_TSLEEP, tsleep, tsleep_end, 1, "tsleep exit");
98 KTR_INFO(KTR_TSLEEP, tsleep, wakeup_beg, 2, "wakeup enter %p", const volatile void *ident);
99 KTR_INFO(KTR_TSLEEP, tsleep, wakeup_end, 3, "wakeup exit");
100 KTR_INFO(KTR_TSLEEP, tsleep, ilockfail, 4, "interlock failed %p", const volatile void *ident);
102 #define logtsleep1(name) KTR_LOG(tsleep_ ## name)
103 #define logtsleep2(name, val) KTR_LOG(tsleep_ ## name, val)
105 struct loadavg averunnable =
106 { {0, 0, 0}, FSCALE }; /* load average, of runnable procs */
108 * Constants for averages over 1, 5, and 15 minutes
109 * when sampling at 5 second intervals.
111 static fixpt_t cexp[3] = {
112 0.9200444146293232 * FSCALE, /* exp(-1/12) */
113 0.9834714538216174 * FSCALE, /* exp(-1/60) */
114 0.9944598480048967 * FSCALE, /* exp(-1/180) */
117 static void endtsleep (void *);
118 static void loadav (void *arg);
119 static void schedcpu (void *arg);
122 * Adjust the scheduler quantum. The quantum is specified in microseconds.
123 * Note that 'tick' is in microseconds per tick.
126 sysctl_kern_quantum(SYSCTL_HANDLER_ARGS)
130 new_val = sched_quantum * ustick;
131 error = sysctl_handle_int(oidp, &new_val, 0, req);
132 if (error != 0 || req->newptr == NULL)
134 if (new_val < ustick)
136 sched_quantum = new_val / ustick;
137 hogticks = 2 * sched_quantum;
141 SYSCTL_PROC(_kern, OID_AUTO, quantum, CTLTYPE_INT|CTLFLAG_RW,
142 0, sizeof sched_quantum, sysctl_kern_quantum, "I", "");
145 * If `ccpu' is not equal to `exp(-1/20)' and you still want to use the
146 * faster/more-accurate formula, you'll have to estimate CCPU_SHIFT below
147 * and possibly adjust FSHIFT in "param.h" so that (FSHIFT >= CCPU_SHIFT).
149 * To estimate CCPU_SHIFT for exp(-1/20), the following formula was used:
150 * 1 - exp(-1/20) ~= 0.0487 ~= 0.0488 == 1 (fixed pt, *11* bits).
152 * If you don't want to bother with the faster/more-accurate formula, you
153 * can set CCPU_SHIFT to (FSHIFT + 1) which will use a slower/less-accurate
154 * (more general) method of calculating the %age of CPU used by a process.
156 * decay 95% of `lwp_pctcpu' in 60 seconds; see CCPU_SHIFT before changing
158 #define CCPU_SHIFT 11
160 static fixpt_t ccpu = 0.95122942450071400909 * FSCALE; /* exp(-1/20) */
161 SYSCTL_INT(_kern, OID_AUTO, ccpu, CTLFLAG_RD, &ccpu, 0, "");
164 * kernel uses `FSCALE', userland (SHOULD) use kern.fscale
166 int fscale __unused = FSCALE; /* exported to systat */
167 SYSCTL_INT(_kern, OID_AUTO, fscale, CTLFLAG_RD, 0, FSCALE, "");
170 * Recompute process priorities, once a second.
172 * Since the userland schedulers are typically event oriented, if the
173 * estcpu calculation at wakeup() time is not sufficient to make a
174 * process runnable relative to other processes in the system we have
175 * a 1-second recalc to help out.
177 * This code also allows us to store sysclock_t data in the process structure
178 * without fear of an overrun, since sysclock_t are guarenteed to hold
179 * several seconds worth of count.
181 * WARNING! callouts can preempt normal threads. However, they will not
182 * preempt a thread holding a spinlock so we *can* safely use spinlocks.
184 static int schedcpu_stats(struct proc *p, void *data __unused);
185 static int schedcpu_resource(struct proc *p, void *data __unused);
190 allproc_scan(schedcpu_stats, NULL);
191 allproc_scan(schedcpu_resource, NULL);
192 wakeup((caddr_t)&lbolt);
193 wakeup(lbolt_syncer);
194 callout_reset(&schedcpu_callout, hz, schedcpu, NULL);
198 * General process statistics once a second
201 schedcpu_stats(struct proc *p, void *data __unused)
206 * Threads may not be completely set up if process in SIDL state.
208 if (p->p_stat == SIDL)
212 if (lwkt_trytoken(&p->p_token) == FALSE) {
218 FOREACH_LWP_IN_PROC(lp, p) {
219 if (lp->lwp_stat == LSSLEEP) {
221 if (lp->lwp_slptime == 1)
222 p->p_usched->uload_update(lp);
226 * Only recalculate processes that are active or have slept
227 * less then 2 seconds. The schedulers understand this.
229 if (lp->lwp_slptime <= 1) {
230 p->p_usched->recalculate(lp);
232 lp->lwp_pctcpu = (lp->lwp_pctcpu * ccpu) >> FSHIFT;
235 lwkt_reltoken(&p->p_token);
242 * Resource checks. XXX break out since ksignal/killproc can block,
243 * limiting us to one process killed per second. There is probably
247 schedcpu_resource(struct proc *p, void *data __unused)
252 if (p->p_stat == SIDL)
256 if (lwkt_trytoken(&p->p_token) == FALSE) {
261 if (p->p_stat == SZOMB || p->p_limit == NULL) {
262 lwkt_reltoken(&p->p_token);
268 FOREACH_LWP_IN_PROC(lp, p) {
270 * We may have caught an lp in the middle of being
271 * created, lwp_thread can be NULL.
273 if (lp->lwp_thread) {
274 ttime += lp->lwp_thread->td_sticks;
275 ttime += lp->lwp_thread->td_uticks;
279 switch(plimit_testcpulimit(p->p_limit, ttime)) {
280 case PLIMIT_TESTCPU_KILL:
281 killproc(p, "exceeded maximum CPU limit");
283 case PLIMIT_TESTCPU_XCPU:
284 if ((p->p_flags & P_XCPU) == 0) {
285 p->p_flags |= P_XCPU;
292 lwkt_reltoken(&p->p_token);
299 * This is only used by ps. Generate a cpu percentage use over
300 * a period of one second.
305 updatepcpu(struct lwp *lp, int cpticks, int ttlticks)
310 acc = (cpticks << FSHIFT) / ttlticks;
311 if (ttlticks >= ESTCPUFREQ) {
312 lp->lwp_pctcpu = acc;
314 remticks = ESTCPUFREQ - ttlticks;
315 lp->lwp_pctcpu = (acc * ttlticks + lp->lwp_pctcpu * remticks) /
321 * tsleep/wakeup hash table parameters. Try to find the sweet spot for
322 * like addresses being slept on.
324 #define TABLESIZE 4001
325 #define LOOKUP(x) (((u_int)(uintptr_t)(x)) % TABLESIZE)
327 static cpumask_t slpque_cpumasks[TABLESIZE];
330 * General scheduler initialization. We force a reschedule 25 times
331 * a second by default. Note that cpu0 is initialized in early boot and
332 * cannot make any high level calls.
334 * Each cpu has its own sleep queue.
337 sleep_gdinit(globaldata_t gd)
339 static struct tslpque slpque_cpu0[TABLESIZE];
342 if (gd->gd_cpuid == 0) {
343 sched_quantum = (hz + 24) / 25;
344 hogticks = 2 * sched_quantum;
346 gd->gd_tsleep_hash = slpque_cpu0;
348 gd->gd_tsleep_hash = kmalloc(sizeof(slpque_cpu0),
349 M_TSLEEP, M_WAITOK | M_ZERO);
351 for (i = 0; i < TABLESIZE; ++i)
352 TAILQ_INIT(&gd->gd_tsleep_hash[i]);
356 * This is a dandy function that allows us to interlock tsleep/wakeup
357 * operations with unspecified upper level locks, such as lockmgr locks,
358 * simply by holding a critical section. The sequence is:
360 * (acquire upper level lock)
361 * tsleep_interlock(blah)
362 * (release upper level lock)
365 * Basically this functions queues us on the tsleep queue without actually
366 * descheduling us. When tsleep() is later called with PINTERLOCK it
367 * assumes the thread was already queued, otherwise it queues it there.
369 * Thus it is possible to receive the wakeup prior to going to sleep and
370 * the race conditions are covered.
373 _tsleep_interlock(globaldata_t gd, const volatile void *ident, int flags)
375 thread_t td = gd->gd_curthread;
378 crit_enter_quick(td);
379 if (td->td_flags & TDF_TSLEEPQ) {
380 id = LOOKUP(td->td_wchan);
381 TAILQ_REMOVE(&gd->gd_tsleep_hash[id], td, td_sleepq);
382 if (TAILQ_FIRST(&gd->gd_tsleep_hash[id]) == NULL) {
383 atomic_clear_cpumask(&slpque_cpumasks[id],
387 td->td_flags |= TDF_TSLEEPQ;
390 TAILQ_INSERT_TAIL(&gd->gd_tsleep_hash[id], td, td_sleepq);
391 atomic_set_cpumask(&slpque_cpumasks[id], gd->gd_cpumask);
392 td->td_wchan = ident;
393 td->td_wdomain = flags & PDOMAIN_MASK;
398 tsleep_interlock(const volatile void *ident, int flags)
400 _tsleep_interlock(mycpu, ident, flags);
404 * Remove thread from sleepq. Must be called with a critical section held.
405 * The thread must not be migrating.
408 _tsleep_remove(thread_t td)
410 globaldata_t gd = mycpu;
413 KKASSERT(td->td_gd == gd && IN_CRITICAL_SECT(td));
414 KKASSERT((td->td_flags & TDF_MIGRATING) == 0);
415 if (td->td_flags & TDF_TSLEEPQ) {
416 td->td_flags &= ~TDF_TSLEEPQ;
417 id = LOOKUP(td->td_wchan);
418 TAILQ_REMOVE(&gd->gd_tsleep_hash[id], td, td_sleepq);
419 if (TAILQ_FIRST(&gd->gd_tsleep_hash[id]) == NULL)
420 atomic_clear_cpumask(&slpque_cpumasks[id], gd->gd_cpumask);
427 tsleep_remove(thread_t td)
433 * General sleep call. Suspends the current process until a wakeup is
434 * performed on the specified identifier. The process will then be made
435 * runnable with the specified priority. Sleeps at most timo/hz seconds
436 * (0 means no timeout). If flags includes PCATCH flag, signals are checked
437 * before and after sleeping, else signals are not checked. Returns 0 if
438 * awakened, EWOULDBLOCK if the timeout expires. If PCATCH is set and a
439 * signal needs to be delivered, ERESTART is returned if the current system
440 * call should be restarted if possible, and EINTR is returned if the system
441 * call should be interrupted by the signal (return EINTR).
443 * Note that if we are a process, we release_curproc() before messing with
444 * the LWKT scheduler.
446 * During autoconfiguration or after a panic, a sleep will simply
447 * lower the priority briefly to allow interrupts, then return.
449 * WARNING! This code can't block (short of switching away), or bad things
450 * will happen. No getting tokens, no blocking locks, etc.
453 tsleep(const volatile void *ident, int flags, const char *wmesg, int timo)
455 struct thread *td = curthread;
456 struct lwp *lp = td->td_lwp;
457 struct proc *p = td->td_proc; /* may be NULL */
463 struct callout thandle;
466 * NOTE: removed KTRPOINT, it could cause races due to blocking
467 * even in stable. Just scrap it for now.
469 if (!tsleep_crypto_dump && (tsleep_now_works == 0 || panicstr)) {
471 * After a panic, or before we actually have an operational
472 * softclock, just give interrupts a chance, then just return;
474 * don't run any other procs or panic below,
475 * in case this is the idle process and already asleep.
479 lwkt_setpri_self(safepri);
481 lwkt_setpri_self(oldpri);
484 logtsleep2(tsleep_beg, ident);
486 KKASSERT(td != &gd->gd_idlethread); /* you must be kidding! */
487 td->td_wakefromcpu = -1; /* overwritten by _wakeup */
490 * NOTE: all of this occurs on the current cpu, including any
491 * callout-based wakeups, so a critical section is a sufficient
494 * The entire sequence through to where we actually sleep must
495 * run without breaking the critical section.
497 catch = flags & PCATCH;
501 crit_enter_quick(td);
503 KASSERT(ident != NULL, ("tsleep: no ident"));
504 KASSERT(lp == NULL ||
505 lp->lwp_stat == LSRUN || /* Obvious */
506 lp->lwp_stat == LSSTOP, /* Set in tstop */
508 ident, wmesg, lp->lwp_stat));
511 * We interlock the sleep queue if the caller has not already done
512 * it for us. This must be done before we potentially acquire any
513 * tokens or we can loose the wakeup.
515 if ((flags & PINTERLOCKED) == 0) {
516 _tsleep_interlock(gd, ident, flags);
520 * Setup for the current process (if this is a process). We must
521 * interlock with lwp_token to avoid remote wakeup races via
525 lwkt_gettoken(&lp->lwp_token);
528 * Early termination if PCATCH was set and a
529 * signal is pending, interlocked with the
532 * Early termination only occurs when tsleep() is
533 * entered while in a normal LSRUN state.
535 if ((sig = CURSIG(lp)) != 0)
539 * Causes ksignal to wake us up if a signal is
540 * received (interlocked with p->p_token).
542 lp->lwp_flags |= LWP_SINTR;
549 * Make sure the current process has been untangled from
550 * the userland scheduler and initialize slptime to start
553 * NOTE: td->td_wakefromcpu is pre-set by the release function
554 * for the dfly scheduler, and then adjusted by _wakeup()
557 p->p_usched->release_curproc(lp);
562 * If the interlocked flag is set but our cpu bit in the slpqueue
563 * is no longer set, then a wakeup was processed inbetween the
564 * tsleep_interlock() (ours or the callers), and here. This can
565 * occur under numerous circumstances including when we release the
568 * Extreme loads can cause the sending of an IPI (e.g. wakeup()'s)
569 * to process incoming IPIs, thus draining incoming wakeups.
571 if ((td->td_flags & TDF_TSLEEPQ) == 0) {
572 logtsleep2(ilockfail, ident);
577 * scheduling is blocked while in a critical section. Coincide
578 * the descheduled-by-tsleep flag with the descheduling of the
581 * The timer callout is localized on our cpu and interlocked by
582 * our critical section.
584 lwkt_deschedule_self(td);
585 td->td_flags |= TDF_TSLEEP_DESCHEDULED;
586 td->td_wmesg = wmesg;
589 * Setup the timeout, if any. The timeout is only operable while
590 * the thread is flagged descheduled.
592 KKASSERT((td->td_flags & TDF_TIMEOUT) == 0);
594 callout_init_mp(&thandle);
595 callout_reset(&thandle, timo, endtsleep, td);
603 * Ok, we are sleeping. Place us in the SSLEEP state.
605 KKASSERT((lp->lwp_mpflags & LWP_MP_ONRUNQ) == 0);
608 * tstop() sets LSSTOP, so don't fiddle with that.
610 if (lp->lwp_stat != LSSTOP)
611 lp->lwp_stat = LSSLEEP;
612 lp->lwp_ru.ru_nvcsw++;
613 p->p_usched->uload_update(lp);
617 * And when we are woken up, put us back in LSRUN. If we
618 * slept for over a second, recalculate our estcpu.
620 lp->lwp_stat = LSRUN;
621 if (lp->lwp_slptime) {
622 p->p_usched->uload_update(lp);
623 p->p_usched->recalculate(lp);
631 * Make sure we haven't switched cpus while we were asleep. It's
632 * not supposed to happen. Cleanup our temporary flags.
634 KKASSERT(gd == td->td_gd);
637 * Cleanup the timeout. If the timeout has already occured thandle
638 * has already been stopped, otherwise stop thandle. If the timeout
639 * is running (the callout thread must be blocked trying to get
640 * lwp_token) then wait for us to get scheduled.
643 while (td->td_flags & TDF_TIMEOUT_RUNNING) {
644 lwkt_deschedule_self(td);
645 td->td_wmesg = "tsrace";
647 kprintf("td %p %s: timeout race\n", td, td->td_comm);
649 if (td->td_flags & TDF_TIMEOUT) {
650 td->td_flags &= ~TDF_TIMEOUT;
653 /* does not block when on same cpu */
654 callout_stop(&thandle);
657 td->td_flags &= ~TDF_TSLEEP_DESCHEDULED;
660 * Make sure we have been removed from the sleepq. In most
661 * cases this will have been done for us already but it is
662 * possible for a scheduling IPI to be in-flight from a
663 * previous tsleep/tsleep_interlock() or due to a straight-out
664 * call to lwkt_schedule() (in the case of an interrupt thread),
665 * causing a spurious wakeup.
671 * Figure out the correct error return. If interrupted by a
672 * signal we want to return EINTR or ERESTART.
676 if (catch && error == 0) {
677 if (sig != 0 || (sig = CURSIG(lp))) {
678 if (SIGISMEMBER(p->p_sigacts->ps_sigintr, sig))
684 lp->lwp_flags &= ~LWP_SINTR;
685 lwkt_reltoken(&lp->lwp_token);
687 logtsleep1(tsleep_end);
693 * Interlocked spinlock sleep. An exclusively held spinlock must
694 * be passed to ssleep(). The function will atomically release the
695 * spinlock and tsleep on the ident, then reacquire the spinlock and
698 * This routine is fairly important along the critical path, so optimize it
702 ssleep(const volatile void *ident, struct spinlock *spin, int flags,
703 const char *wmesg, int timo)
705 globaldata_t gd = mycpu;
708 _tsleep_interlock(gd, ident, flags);
709 spin_unlock_quick(gd, spin);
710 error = tsleep(ident, flags | PINTERLOCKED, wmesg, timo);
711 spin_lock_quick(gd, spin);
717 lksleep(const volatile void *ident, struct lock *lock, int flags,
718 const char *wmesg, int timo)
720 globaldata_t gd = mycpu;
723 _tsleep_interlock(gd, ident, flags);
724 lockmgr(lock, LK_RELEASE);
725 error = tsleep(ident, flags | PINTERLOCKED, wmesg, timo);
726 lockmgr(lock, LK_EXCLUSIVE);
732 * Interlocked mutex sleep. An exclusively held mutex must be passed
733 * to mtxsleep(). The function will atomically release the mutex
734 * and tsleep on the ident, then reacquire the mutex and return.
737 mtxsleep(const volatile void *ident, struct mtx *mtx, int flags,
738 const char *wmesg, int timo)
740 globaldata_t gd = mycpu;
743 _tsleep_interlock(gd, ident, flags);
745 error = tsleep(ident, flags | PINTERLOCKED, wmesg, timo);
746 mtx_lock_ex_quick(mtx, wmesg);
752 * Interlocked serializer sleep. An exclusively held serializer must
753 * be passed to zsleep(). The function will atomically release
754 * the serializer and tsleep on the ident, then reacquire the serializer
758 zsleep(const volatile void *ident, struct lwkt_serialize *slz, int flags,
759 const char *wmesg, int timo)
761 globaldata_t gd = mycpu;
764 ASSERT_SERIALIZED(slz);
766 _tsleep_interlock(gd, ident, flags);
767 lwkt_serialize_exit(slz);
768 ret = tsleep(ident, flags | PINTERLOCKED, wmesg, timo);
769 lwkt_serialize_enter(slz);
775 * Directly block on the LWKT thread by descheduling it. This
776 * is much faster then tsleep(), but the only legal way to wake
777 * us up is to directly schedule the thread.
779 * Setting TDF_SINTR will cause new signals to directly schedule us.
781 * This routine must be called while in a critical section.
784 lwkt_sleep(const char *wmesg, int flags)
786 thread_t td = curthread;
789 if ((flags & PCATCH) == 0 || td->td_lwp == NULL) {
790 td->td_flags |= TDF_BLOCKED;
791 td->td_wmesg = wmesg;
792 lwkt_deschedule_self(td);
795 td->td_flags &= ~TDF_BLOCKED;
798 if ((sig = CURSIG(td->td_lwp)) != 0) {
799 if (SIGISMEMBER(td->td_proc->p_sigacts->ps_sigintr, sig))
805 td->td_flags |= TDF_BLOCKED | TDF_SINTR;
806 td->td_wmesg = wmesg;
807 lwkt_deschedule_self(td);
809 td->td_flags &= ~(TDF_BLOCKED | TDF_SINTR);
815 * Implement the timeout for tsleep.
817 * This type of callout timeout is scheduled on the same cpu the process
818 * is sleeping on. Also, at the moment, the MP lock is held.
827 * We are going to have to get the lwp_token, which means we might
828 * block. This can race a tsleep getting woken up by other means
829 * so set TDF_TIMEOUT_RUNNING to force the tsleep to wait for our
830 * processing to complete (sorry tsleep!).
832 * We can safely set td_flags because td MUST be on the same cpu
835 KKASSERT(td->td_gd == mycpu);
837 td->td_flags |= TDF_TIMEOUT_RUNNING | TDF_TIMEOUT;
840 * This can block but TDF_TIMEOUT_RUNNING will prevent the thread
841 * from exiting the tsleep on us. The flag is interlocked by virtue
842 * of lp being on the same cpu as we are.
844 if ((lp = td->td_lwp) != NULL)
845 lwkt_gettoken(&lp->lwp_token);
847 KKASSERT(td->td_flags & TDF_TSLEEP_DESCHEDULED);
850 if (lp->lwp_proc->p_stat != SSTOP)
852 lwkt_reltoken(&lp->lwp_token);
857 KKASSERT(td->td_gd == mycpu);
858 td->td_flags &= ~TDF_TIMEOUT_RUNNING;
863 * Make all processes sleeping on the specified identifier runnable.
864 * count may be zero or one only.
866 * The domain encodes the sleep/wakeup domain, flags, plus the originating
869 * This call may run without the MP lock held. We can only manipulate thread
870 * state on the cpu owning the thread. We CANNOT manipulate process state
873 * _wakeup() can be passed to an IPI so we can't use (const volatile
877 _wakeup(void *ident, int domain)
889 logtsleep2(wakeup_beg, ident);
892 qp = &gd->gd_tsleep_hash[id];
894 for (td = TAILQ_FIRST(qp); td != NULL; td = ntd) {
895 ntd = TAILQ_NEXT(td, td_sleepq);
896 if (td->td_wchan == ident &&
897 td->td_wdomain == (domain & PDOMAIN_MASK)
899 KKASSERT(td->td_gd == gd);
901 td->td_wakefromcpu = PWAKEUP_DECODE(domain);
902 if (td->td_flags & TDF_TSLEEP_DESCHEDULED) {
904 if (domain & PWAKEUP_ONE)
913 * We finished checking the current cpu but there still may be
914 * more work to do. Either wakeup_one was requested and no matching
915 * thread was found, or a normal wakeup was requested and we have
916 * to continue checking cpus.
918 * It should be noted that this scheme is actually less expensive then
919 * the old scheme when waking up multiple threads, since we send
920 * only one IPI message per target candidate which may then schedule
921 * multiple threads. Before we could have wound up sending an IPI
922 * message for each thread on the target cpu (!= current cpu) that
923 * needed to be woken up.
925 * NOTE: Wakeups occuring on remote cpus are asynchronous. This
926 * should be ok since we are passing idents in the IPI rather then
929 if ((domain & PWAKEUP_MYCPU) == 0 &&
930 (mask = slpque_cpumasks[id] & gd->gd_other_cpus) != 0) {
931 lwkt_send_ipiq2_mask(mask, _wakeup, ident,
932 domain | PWAKEUP_MYCPU);
936 logtsleep1(wakeup_end);
941 * Wakeup all threads tsleep()ing on the specified ident, on all cpus
944 wakeup(const volatile void *ident)
946 _wakeup(__DEALL(ident), PWAKEUP_ENCODE(0, mycpu->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)
991 globaldata_t mygd = mycpu;
993 _wakeup(__DEALL(ident), PWAKEUP_ENCODE(0, mygd->gd_cpuid) |
996 lwkt_send_ipiq2(gd, _wakeup, __DEALL(ident),
997 PWAKEUP_ENCODE(0, mygd->gd_cpuid) |
1001 _wakeup(__DEALL(ident), PWAKEUP_MYCPU);
1006 * Wakeup one thread tsleep()ing on the specified ident on the specified cpu
1010 wakeup_oncpu_one(globaldata_t gd, const volatile void *ident)
1013 globaldata_t mygd = mycpu;
1015 _wakeup(__DEALL(ident), PWAKEUP_ENCODE(0, mygd->gd_cpuid) |
1016 PWAKEUP_MYCPU | PWAKEUP_ONE);
1018 lwkt_send_ipiq2(gd, _wakeup, __DEALL(ident),
1019 PWAKEUP_ENCODE(0, mygd->gd_cpuid) |
1020 PWAKEUP_MYCPU | PWAKEUP_ONE);
1023 _wakeup(__DEALL(ident), PWAKEUP_MYCPU | PWAKEUP_ONE);
1028 * Wakeup all threads waiting on the specified ident that slept using
1029 * the specified domain, on all cpus.
1032 wakeup_domain(const volatile void *ident, int domain)
1034 _wakeup(__DEALL(ident), PWAKEUP_ENCODE(domain, mycpu->gd_cpuid));
1038 * Wakeup one thread waiting on the specified ident that slept using
1039 * the specified domain, on any cpu.
1042 wakeup_domain_one(const volatile void *ident, int domain)
1044 /* XXX potentially round-robin the first responding cpu */
1045 _wakeup(__DEALL(ident),
1046 PWAKEUP_ENCODE(domain, mycpu->gd_cpuid) | PWAKEUP_ONE);
1052 * Make a process runnable. lp->lwp_token must be held on call and this
1053 * function must be called from the cpu owning lp.
1055 * This only has an effect if we are in LSSTOP or LSSLEEP.
1058 setrunnable(struct lwp *lp)
1060 thread_t td = lp->lwp_thread;
1062 ASSERT_LWKT_TOKEN_HELD(&lp->lwp_token);
1063 KKASSERT(td->td_gd == mycpu);
1065 if (lp->lwp_stat == LSSTOP)
1066 lp->lwp_stat = LSSLEEP;
1067 if (lp->lwp_stat == LSSLEEP) {
1070 } else if (td->td_flags & TDF_SINTR) {
1077 * The process is stopped due to some condition, usually because p_stat is
1078 * set to SSTOP, but also possibly due to being traced.
1080 * Caller must hold p->p_token
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 if the process state is detected as
1088 * possibly being in a stopped state.
1093 struct lwp *lp = curthread->td_lwp;
1094 struct proc *p = lp->lwp_proc;
1097 lwkt_gettoken(&lp->lwp_token);
1101 * If LWP_MP_WSTOP is set, we were sleeping
1102 * while our process was stopped. At this point
1103 * we were already counted as stopped.
1105 if ((lp->lwp_mpflags & LWP_MP_WSTOP) == 0) {
1107 * If we're the last thread to stop, signal
1111 atomic_set_int(&lp->lwp_mpflags, LWP_MP_WSTOP);
1112 wakeup(&p->p_nstopped);
1113 if (p->p_nstopped == p->p_nthreads) {
1115 * Token required to interlock kern_wait()
1119 lwkt_gettoken(&q->p_token);
1120 p->p_flags &= ~P_WAITED;
1122 if ((q->p_sigacts->ps_flag & PS_NOCLDSTOP) == 0)
1123 ksignal(q, SIGCHLD);
1124 lwkt_reltoken(&q->p_token);
1128 while (p->p_stat == SSTOP) {
1129 lp->lwp_stat = LSSTOP;
1130 tsleep(p, 0, "stop", 0);
1133 atomic_clear_int(&lp->lwp_mpflags, LWP_MP_WSTOP);
1135 lwkt_reltoken(&lp->lwp_token);
1139 * Compute a tenex style load average of a quantity on
1140 * 1, 5 and 15 minute intervals.
1142 static int loadav_count_runnable(struct lwp *p, void *data);
1147 struct loadavg *avg;
1151 alllwp_scan(loadav_count_runnable, &nrun);
1153 for (i = 0; i < 3; i++) {
1154 avg->ldavg[i] = (cexp[i] * avg->ldavg[i] +
1155 nrun * FSCALE * (FSCALE - cexp[i])) >> FSHIFT;
1159 * Schedule the next update to occur after 5 seconds, but add a
1160 * random variation to avoid synchronisation with processes that
1161 * run at regular intervals.
1163 callout_reset(&loadav_callout, hz * 4 + (int)(krandom() % (hz * 2 + 1)),
1168 loadav_count_runnable(struct lwp *lp, void *data)
1173 switch (lp->lwp_stat) {
1175 if ((td = lp->lwp_thread) == NULL)
1177 if (td->td_flags & TDF_BLOCKED)
1190 sched_setup(void *dummy)
1192 callout_init_mp(&loadav_callout);
1193 callout_init_mp(&schedcpu_callout);
1195 /* Kick off timeout driven events by calling first time. */