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
35 #include <sys/cdefs.h>
36 __FBSDID("$FreeBSD$");
38 #include "opt_hwpmc_hooks.h"
39 #include "opt_sched.h"
40 #include "opt_kdtrace.h"
42 #include <sys/param.h>
43 #include <sys/systm.h>
44 #include <sys/cpuset.h>
45 #include <sys/kernel.h>
48 #include <sys/kthread.h>
49 #include <sys/mutex.h>
51 #include <sys/resourcevar.h>
52 #include <sys/sched.h>
55 #include <sys/sysctl.h>
57 #include <sys/turnstile.h>
59 #include <machine/pcb.h>
60 #include <machine/smp.h>
63 #include <sys/pmckern.h>
67 #include <sys/dtrace_bsd.h>
68 int dtrace_vtime_active;
69 dtrace_vtime_switch_func_t dtrace_vtime_switch_func;
73 * INVERSE_ESTCPU_WEIGHT is only suitable for statclock() frequencies in
74 * the range 100-256 Hz (approximately).
76 #define ESTCPULIM(e) \
77 min((e), INVERSE_ESTCPU_WEIGHT * (NICE_WEIGHT * (PRIO_MAX - PRIO_MIN) - \
78 RQ_PPQ) + INVERSE_ESTCPU_WEIGHT - 1)
80 #define INVERSE_ESTCPU_WEIGHT (8 * smp_cpus)
82 #define INVERSE_ESTCPU_WEIGHT 8 /* 1 / (priorities per estcpu level). */
84 #define NICE_WEIGHT 1 /* Priorities per nice level. */
86 #define TS_NAME_LEN (MAXCOMLEN + sizeof(" td ") + sizeof(__XSTRING(UINT_MAX)))
89 * The schedulable entity that runs a context.
90 * This is an extension to the thread structure and is tailored to
91 * the requirements of this scheduler
94 fixpt_t ts_pctcpu; /* (j) %cpu during p_swtime. */
95 int ts_cpticks; /* (j) Ticks of cpu time. */
96 int ts_slptime; /* (j) Seconds !RUNNING. */
98 struct runq *ts_runq; /* runq the thread is currently on */
100 char ts_name[TS_NAME_LEN];
104 /* flags kept in td_flags */
105 #define TDF_DIDRUN TDF_SCHED0 /* thread actually ran. */
106 #define TDF_BOUND TDF_SCHED1 /* Bound to one CPU. */
108 /* flags kept in ts_flags */
109 #define TSF_AFFINITY 0x0001 /* Has a non-"full" CPU set. */
111 #define SKE_RUNQ_PCPU(ts) \
112 ((ts)->ts_runq != 0 && (ts)->ts_runq != &runq)
114 #define THREAD_CAN_SCHED(td, cpu) \
115 CPU_ISSET((cpu), &(td)->td_cpuset->cs_mask)
117 static struct td_sched td_sched0;
118 struct mtx sched_lock;
120 static int sched_tdcnt; /* Total runnable threads in the system. */
121 static int sched_quantum; /* Roundrobin scheduling quantum in ticks. */
122 #define SCHED_QUANTUM (hz / 10) /* Default sched quantum */
124 static void setup_runqs(void);
125 static void schedcpu(void);
126 static void schedcpu_thread(void);
127 static void sched_priority(struct thread *td, u_char prio);
128 static void sched_setup(void *dummy);
129 static void maybe_resched(struct thread *td);
130 static void updatepri(struct thread *td);
131 static void resetpriority(struct thread *td);
132 static void resetpriority_thread(struct thread *td);
134 static int sched_pickcpu(struct thread *td);
135 static int forward_wakeup(int cpunum);
136 static void kick_other_cpu(int pri, int cpuid);
139 static struct kproc_desc sched_kp = {
144 SYSINIT(schedcpu, SI_SUB_RUN_SCHEDULER, SI_ORDER_FIRST, kproc_start,
146 SYSINIT(sched_setup, SI_SUB_RUN_QUEUE, SI_ORDER_FIRST, sched_setup, NULL);
151 static struct runq runq;
157 static struct runq runq_pcpu[MAXCPU];
158 long runq_length[MAXCPU];
167 for (i = 0; i < MAXCPU; ++i)
168 runq_init(&runq_pcpu[i]);
175 sysctl_kern_quantum(SYSCTL_HANDLER_ARGS)
179 new_val = sched_quantum * tick;
180 error = sysctl_handle_int(oidp, &new_val, 0, req);
181 if (error != 0 || req->newptr == NULL)
185 sched_quantum = new_val / tick;
186 hogticks = 2 * sched_quantum;
190 SYSCTL_NODE(_kern, OID_AUTO, sched, CTLFLAG_RD, 0, "Scheduler");
192 SYSCTL_STRING(_kern_sched, OID_AUTO, name, CTLFLAG_RD, "4BSD", 0,
195 SYSCTL_PROC(_kern_sched, OID_AUTO, quantum, CTLTYPE_INT | CTLFLAG_RW,
196 0, sizeof sched_quantum, sysctl_kern_quantum, "I",
197 "Roundrobin scheduling quantum in microseconds");
200 /* Enable forwarding of wakeups to all other cpus */
201 SYSCTL_NODE(_kern_sched, OID_AUTO, ipiwakeup, CTLFLAG_RD, NULL, "Kernel SMP");
203 static int runq_fuzz = 1;
204 SYSCTL_INT(_kern_sched, OID_AUTO, runq_fuzz, CTLFLAG_RW, &runq_fuzz, 0, "");
206 static int forward_wakeup_enabled = 1;
207 SYSCTL_INT(_kern_sched_ipiwakeup, OID_AUTO, enabled, CTLFLAG_RW,
208 &forward_wakeup_enabled, 0,
209 "Forwarding of wakeup to idle CPUs");
211 static int forward_wakeups_requested = 0;
212 SYSCTL_INT(_kern_sched_ipiwakeup, OID_AUTO, requested, CTLFLAG_RD,
213 &forward_wakeups_requested, 0,
214 "Requests for Forwarding of wakeup to idle CPUs");
216 static int forward_wakeups_delivered = 0;
217 SYSCTL_INT(_kern_sched_ipiwakeup, OID_AUTO, delivered, CTLFLAG_RD,
218 &forward_wakeups_delivered, 0,
219 "Completed Forwarding of wakeup to idle CPUs");
221 static int forward_wakeup_use_mask = 1;
222 SYSCTL_INT(_kern_sched_ipiwakeup, OID_AUTO, usemask, CTLFLAG_RW,
223 &forward_wakeup_use_mask, 0,
224 "Use the mask of idle cpus");
226 static int forward_wakeup_use_loop = 0;
227 SYSCTL_INT(_kern_sched_ipiwakeup, OID_AUTO, useloop, CTLFLAG_RW,
228 &forward_wakeup_use_loop, 0,
229 "Use a loop to find idle cpus");
231 static int forward_wakeup_use_single = 0;
232 SYSCTL_INT(_kern_sched_ipiwakeup, OID_AUTO, onecpu, CTLFLAG_RW,
233 &forward_wakeup_use_single, 0,
234 "Only signal one idle cpu");
236 static int forward_wakeup_use_htt = 0;
237 SYSCTL_INT(_kern_sched_ipiwakeup, OID_AUTO, htt2, CTLFLAG_RW,
238 &forward_wakeup_use_htt, 0,
243 static int sched_followon = 0;
244 SYSCTL_INT(_kern_sched, OID_AUTO, followon, CTLFLAG_RW,
246 "allow threads to share a quantum");
249 SDT_PROVIDER_DEFINE(sched);
251 SDT_PROBE_DEFINE3(sched, , , change_pri, change-pri, "struct thread *",
252 "struct proc *", "uint8_t");
253 SDT_PROBE_DEFINE3(sched, , , dequeue, dequeue, "struct thread *",
254 "struct proc *", "void *");
255 SDT_PROBE_DEFINE4(sched, , , enqueue, enqueue, "struct thread *",
256 "struct proc *", "void *", "int");
257 SDT_PROBE_DEFINE4(sched, , , lend_pri, lend-pri, "struct thread *",
258 "struct proc *", "uint8_t", "struct thread *");
259 SDT_PROBE_DEFINE2(sched, , , load_change, load-change, "int", "int");
260 SDT_PROBE_DEFINE2(sched, , , off_cpu, off-cpu, "struct thread *",
262 SDT_PROBE_DEFINE(sched, , , on_cpu, on-cpu);
263 SDT_PROBE_DEFINE(sched, , , remain_cpu, remain-cpu);
264 SDT_PROBE_DEFINE2(sched, , , surrender, surrender, "struct thread *",
272 KTR_COUNTER0(KTR_SCHED, "load", "global load", sched_tdcnt);
273 SDT_PROBE2(sched, , , load_change, NOCPU, sched_tdcnt);
281 KTR_COUNTER0(KTR_SCHED, "load", "global load", sched_tdcnt);
282 SDT_PROBE2(sched, , , load_change, NOCPU, sched_tdcnt);
285 * Arrange to reschedule if necessary, taking the priorities and
286 * schedulers into account.
289 maybe_resched(struct thread *td)
292 THREAD_LOCK_ASSERT(td, MA_OWNED);
293 if (td->td_priority < curthread->td_priority)
294 curthread->td_flags |= TDF_NEEDRESCHED;
298 * This function is called when a thread is about to be put on run queue
299 * because it has been made runnable or its priority has been adjusted. It
300 * determines if the new thread should be immediately preempted to. If so,
301 * it switches to it and eventually returns true. If not, it returns false
302 * so that the caller may place the thread on an appropriate run queue.
305 maybe_preempt(struct thread *td)
312 * The new thread should not preempt the current thread if any of the
313 * following conditions are true:
315 * - The kernel is in the throes of crashing (panicstr).
316 * - The current thread has a higher (numerically lower) or
317 * equivalent priority. Note that this prevents curthread from
318 * trying to preempt to itself.
319 * - It is too early in the boot for context switches (cold is set).
320 * - The current thread has an inhibitor set or is in the process of
321 * exiting. In this case, the current thread is about to switch
322 * out anyways, so there's no point in preempting. If we did,
323 * the current thread would not be properly resumed as well, so
324 * just avoid that whole landmine.
325 * - If the new thread's priority is not a realtime priority and
326 * the current thread's priority is not an idle priority and
327 * FULL_PREEMPTION is disabled.
329 * If all of these conditions are false, but the current thread is in
330 * a nested critical section, then we have to defer the preemption
331 * until we exit the critical section. Otherwise, switch immediately
335 THREAD_LOCK_ASSERT(td, MA_OWNED);
336 KASSERT((td->td_inhibitors == 0),
337 ("maybe_preempt: trying to run inhibited thread"));
338 pri = td->td_priority;
339 cpri = ctd->td_priority;
340 if (panicstr != NULL || pri >= cpri || cold /* || dumping */ ||
341 TD_IS_INHIBITED(ctd))
343 #ifndef FULL_PREEMPTION
344 if (pri > PRI_MAX_ITHD && cpri < PRI_MIN_IDLE)
348 if (ctd->td_critnest > 1) {
349 CTR1(KTR_PROC, "maybe_preempt: in critical section %d",
351 ctd->td_owepreempt = 1;
355 * Thread is runnable but not yet put on system run queue.
357 MPASS(ctd->td_lock == td->td_lock);
358 MPASS(TD_ON_RUNQ(td));
360 CTR3(KTR_PROC, "preempting to thread %p (pid %d, %s)\n", td,
361 td->td_proc->p_pid, td->td_name);
362 mi_switch(SW_INVOL | SW_PREEMPT | SWT_PREEMPT, td);
364 * td's lock pointer may have changed. We have to return with it
378 * Constants for digital decay and forget:
379 * 90% of (td_estcpu) usage in 5 * loadav time
380 * 95% of (ts_pctcpu) usage in 60 seconds (load insensitive)
381 * Note that, as ps(1) mentions, this can let percentages
382 * total over 100% (I've seen 137.9% for 3 processes).
384 * Note that schedclock() updates td_estcpu and p_cpticks asynchronously.
386 * We wish to decay away 90% of td_estcpu in (5 * loadavg) seconds.
387 * That is, the system wants to compute a value of decay such
388 * that the following for loop:
389 * for (i = 0; i < (5 * loadavg); i++)
390 * td_estcpu *= decay;
393 * for all values of loadavg:
395 * Mathematically this loop can be expressed by saying:
396 * decay ** (5 * loadavg) ~= .1
398 * The system computes decay as:
399 * decay = (2 * loadavg) / (2 * loadavg + 1)
401 * We wish to prove that the system's computation of decay
402 * will always fulfill the equation:
403 * decay ** (5 * loadavg) ~= .1
405 * If we compute b as:
408 * decay = b / (b + 1)
410 * We now need to prove two things:
411 * 1) Given factor ** (5 * loadavg) ~= .1, prove factor == b/(b+1)
412 * 2) Given b/(b+1) ** power ~= .1, prove power == (5 * loadavg)
415 * For x close to zero, exp(x) =~ 1 + x, since
416 * exp(x) = 0! + x**1/1! + x**2/2! + ... .
417 * therefore exp(-1/b) =~ 1 - (1/b) = (b-1)/b.
418 * For x close to zero, ln(1+x) =~ x, since
419 * ln(1+x) = x - x**2/2 + x**3/3 - ... -1 < x < 1
420 * therefore ln(b/(b+1)) = ln(1 - 1/(b+1)) =~ -1/(b+1).
424 * Solve (factor)**(power) =~ .1 given power (5*loadav):
425 * solving for factor,
426 * ln(factor) =~ (-2.30/5*loadav), or
427 * factor =~ exp(-1/((5/2.30)*loadav)) =~ exp(-1/(2*loadav)) =
428 * exp(-1/b) =~ (b-1)/b =~ b/(b+1). QED
431 * Solve (factor)**(power) =~ .1 given factor == (b/(b+1)):
433 * power*ln(b/(b+1)) =~ -2.30, or
434 * power =~ 2.3 * (b + 1) = 4.6*loadav + 2.3 =~ 5*loadav. QED
436 * Actual power values for the implemented algorithm are as follows:
438 * power: 5.68 10.32 14.94 19.55
441 /* calculations for digital decay to forget 90% of usage in 5*loadav sec */
442 #define loadfactor(loadav) (2 * (loadav))
443 #define decay_cpu(loadfac, cpu) (((loadfac) * (cpu)) / ((loadfac) + FSCALE))
445 /* decay 95% of `ts_pctcpu' in 60 seconds; see CCPU_SHIFT before changing */
446 static fixpt_t ccpu = 0.95122942450071400909 * FSCALE; /* exp(-1/20) */
447 SYSCTL_INT(_kern, OID_AUTO, ccpu, CTLFLAG_RD, &ccpu, 0, "");
450 * If `ccpu' is not equal to `exp(-1/20)' and you still want to use the
451 * faster/more-accurate formula, you'll have to estimate CCPU_SHIFT below
452 * and possibly adjust FSHIFT in "param.h" so that (FSHIFT >= CCPU_SHIFT).
454 * To estimate CCPU_SHIFT for exp(-1/20), the following formula was used:
455 * 1 - exp(-1/20) ~= 0.0487 ~= 0.0488 == 1 (fixed pt, *11* bits).
457 * If you don't want to bother with the faster/more-accurate formula, you
458 * can set CCPU_SHIFT to (FSHIFT + 1) which will use a slower/less-accurate
459 * (more general) method of calculating the %age of CPU used by a process.
461 #define CCPU_SHIFT 11
464 * Recompute process priorities, every hz ticks.
465 * MP-safe, called without the Giant mutex.
471 register fixpt_t loadfac = loadfactor(averunnable.ldavg[0]);
475 int awake, realstathz;
477 realstathz = stathz ? stathz : hz;
478 sx_slock(&allproc_lock);
479 FOREACH_PROC_IN_SYSTEM(p) {
481 if (p->p_state == PRS_NEW) {
485 FOREACH_THREAD_IN_PROC(p, td) {
490 * Increment sleep time (if sleeping). We
491 * ignore overflow, as above.
494 * The td_sched slptimes are not touched in wakeup
495 * because the thread may not HAVE everything in
496 * memory? XXX I think this is out of date.
498 if (TD_ON_RUNQ(td)) {
500 td->td_flags &= ~TDF_DIDRUN;
501 } else if (TD_IS_RUNNING(td)) {
503 /* Do not clear TDF_DIDRUN */
504 } else if (td->td_flags & TDF_DIDRUN) {
506 td->td_flags &= ~TDF_DIDRUN;
510 * ts_pctcpu is only for ps and ttyinfo().
512 ts->ts_pctcpu = (ts->ts_pctcpu * ccpu) >> FSHIFT;
514 * If the td_sched has been idle the entire second,
515 * stop recalculating its priority until
518 if (ts->ts_cpticks != 0) {
519 #if (FSHIFT >= CCPU_SHIFT)
520 ts->ts_pctcpu += (realstathz == 100)
521 ? ((fixpt_t) ts->ts_cpticks) <<
522 (FSHIFT - CCPU_SHIFT) :
523 100 * (((fixpt_t) ts->ts_cpticks)
524 << (FSHIFT - CCPU_SHIFT)) / realstathz;
526 ts->ts_pctcpu += ((FSCALE - ccpu) *
528 FSCALE / realstathz)) >> FSHIFT;
533 * If there are ANY running threads in this process,
534 * then don't count it as sleeping.
535 * XXX: this is broken.
538 if (ts->ts_slptime > 1) {
540 * In an ideal world, this should not
541 * happen, because whoever woke us
542 * up from the long sleep should have
543 * unwound the slptime and reset our
544 * priority before we run at the stale
545 * priority. Should KASSERT at some
546 * point when all the cases are fixed.
553 if (ts->ts_slptime > 1) {
557 td->td_estcpu = decay_cpu(loadfac, td->td_estcpu);
559 resetpriority_thread(td);
564 sx_sunlock(&allproc_lock);
568 * Main loop for a kthread that executes schedcpu once a second.
571 schedcpu_thread(void)
581 * Recalculate the priority of a process after it has slept for a while.
582 * For all load averages >= 1 and max td_estcpu of 255, sleeping for at
583 * least six times the loadfactor will decay td_estcpu to zero.
586 updatepri(struct thread *td)
593 loadfac = loadfactor(averunnable.ldavg[0]);
594 if (ts->ts_slptime > 5 * loadfac)
597 newcpu = td->td_estcpu;
598 ts->ts_slptime--; /* was incremented in schedcpu() */
599 while (newcpu && --ts->ts_slptime)
600 newcpu = decay_cpu(loadfac, newcpu);
601 td->td_estcpu = newcpu;
606 * Compute the priority of a process when running in user mode.
607 * Arrange to reschedule if the resulting priority is better
608 * than that of the current process.
611 resetpriority(struct thread *td)
613 register unsigned int newpriority;
615 if (td->td_pri_class == PRI_TIMESHARE) {
616 newpriority = PUSER + td->td_estcpu / INVERSE_ESTCPU_WEIGHT +
617 NICE_WEIGHT * (td->td_proc->p_nice - PRIO_MIN);
618 newpriority = min(max(newpriority, PRI_MIN_TIMESHARE),
620 sched_user_prio(td, newpriority);
625 * Update the thread's priority when the associated process's user
629 resetpriority_thread(struct thread *td)
632 /* Only change threads with a time sharing user priority. */
633 if (td->td_priority < PRI_MIN_TIMESHARE ||
634 td->td_priority > PRI_MAX_TIMESHARE)
637 /* XXX the whole needresched thing is broken, but not silly. */
640 sched_prio(td, td->td_user_pri);
645 sched_setup(void *dummy)
649 if (sched_quantum == 0)
650 sched_quantum = SCHED_QUANTUM;
651 hogticks = 2 * sched_quantum;
653 /* Account for thread0. */
657 /* External interfaces start here */
660 * Very early in the boot some setup of scheduler-specific
661 * parts of proc0 and of some scheduler resources needs to be done.
669 * Set up the scheduler specific parts of proc0.
671 proc0.p_sched = NULL; /* XXX */
672 thread0.td_sched = &td_sched0;
673 thread0.td_lock = &sched_lock;
674 mtx_init(&sched_lock, "sched lock", NULL, MTX_SPIN | MTX_RECURSE);
681 return runq_check(&runq) + runq_check(&runq_pcpu[PCPU_GET(cpuid)]);
683 return runq_check(&runq);
688 sched_rr_interval(void)
690 if (sched_quantum == 0)
691 sched_quantum = SCHED_QUANTUM;
692 return (sched_quantum);
696 * We adjust the priority of the current process. The priority of
697 * a process gets worse as it accumulates CPU time. The cpu usage
698 * estimator (td_estcpu) is increased here. resetpriority() will
699 * compute a different priority each time td_estcpu increases by
700 * INVERSE_ESTCPU_WEIGHT
701 * (until MAXPRI is reached). The cpu usage estimator ramps up
702 * quite quickly when the process is running (linearly), and decays
703 * away exponentially, at a rate which is proportionally slower when
704 * the system is busy. The basic principle is that the system will
705 * 90% forget that the process used a lot of CPU time in 5 * loadav
706 * seconds. This causes the system to favor processes which haven't
707 * run much recently, and to round-robin among other processes.
710 sched_clock(struct thread *td)
714 THREAD_LOCK_ASSERT(td, MA_OWNED);
718 td->td_estcpu = ESTCPULIM(td->td_estcpu + 1);
719 if ((td->td_estcpu % INVERSE_ESTCPU_WEIGHT) == 0) {
721 resetpriority_thread(td);
725 * Force a context switch if the current thread has used up a full
726 * quantum (default quantum is 100ms).
728 if (!TD_IS_IDLETHREAD(td) &&
729 ticks - PCPU_GET(switchticks) >= sched_quantum)
730 td->td_flags |= TDF_NEEDRESCHED;
734 * Charge child's scheduling CPU usage to parent.
737 sched_exit(struct proc *p, struct thread *td)
740 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(td), "proc exit",
741 "prio:%d", td->td_priority);
743 PROC_LOCK_ASSERT(p, MA_OWNED);
744 sched_exit_thread(FIRST_THREAD_IN_PROC(p), td);
748 sched_exit_thread(struct thread *td, struct thread *child)
751 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(child), "exit",
752 "prio:%d", child->td_priority);
754 td->td_estcpu = ESTCPULIM(td->td_estcpu + child->td_estcpu);
757 if ((child->td_flags & TDF_NOLOAD) == 0)
759 thread_unlock(child);
763 sched_fork(struct thread *td, struct thread *childtd)
765 sched_fork_thread(td, childtd);
769 sched_fork_thread(struct thread *td, struct thread *childtd)
773 childtd->td_estcpu = td->td_estcpu;
774 childtd->td_lock = &sched_lock;
775 childtd->td_cpuset = cpuset_ref(td->td_cpuset);
776 childtd->td_priority = childtd->td_base_pri;
777 ts = childtd->td_sched;
778 bzero(ts, sizeof(*ts));
779 ts->ts_flags |= (td->td_sched->ts_flags & TSF_AFFINITY);
783 sched_nice(struct proc *p, int nice)
787 PROC_LOCK_ASSERT(p, MA_OWNED);
789 FOREACH_THREAD_IN_PROC(p, td) {
792 resetpriority_thread(td);
798 sched_class(struct thread *td, int class)
800 THREAD_LOCK_ASSERT(td, MA_OWNED);
801 td->td_pri_class = class;
805 * Adjust the priority of a thread.
808 sched_priority(struct thread *td, u_char prio)
812 KTR_POINT3(KTR_SCHED, "thread", sched_tdname(td), "priority change",
813 "prio:%d", td->td_priority, "new prio:%d", prio, KTR_ATTR_LINKED,
814 sched_tdname(curthread));
815 SDT_PROBE3(sched, , , change_pri, td, td->td_proc, prio);
816 if (td != curthread && prio > td->td_priority) {
817 KTR_POINT3(KTR_SCHED, "thread", sched_tdname(curthread),
818 "lend prio", "prio:%d", td->td_priority, "new prio:%d",
819 prio, KTR_ATTR_LINKED, sched_tdname(td));
820 SDT_PROBE4(sched, , , lend_pri, td, td->td_proc, prio,
823 THREAD_LOCK_ASSERT(td, MA_OWNED);
824 if (td->td_priority == prio)
826 td->td_priority = prio;
827 if (TD_ON_RUNQ(td) && td->td_rqindex != (prio / RQ_PPQ)) {
829 sched_add(td, SRQ_BORING);
834 * Update a thread's priority when it is lent another thread's
838 sched_lend_prio(struct thread *td, u_char prio)
841 td->td_flags |= TDF_BORROWING;
842 sched_priority(td, prio);
846 * Restore a thread's priority when priority propagation is
847 * over. The prio argument is the minimum priority the thread
848 * needs to have to satisfy other possible priority lending
849 * requests. If the thread's regulary priority is less
850 * important than prio the thread will keep a priority boost
854 sched_unlend_prio(struct thread *td, u_char prio)
858 if (td->td_base_pri >= PRI_MIN_TIMESHARE &&
859 td->td_base_pri <= PRI_MAX_TIMESHARE)
860 base_pri = td->td_user_pri;
862 base_pri = td->td_base_pri;
863 if (prio >= base_pri) {
864 td->td_flags &= ~TDF_BORROWING;
865 sched_prio(td, base_pri);
867 sched_lend_prio(td, prio);
871 sched_prio(struct thread *td, u_char prio)
875 /* First, update the base priority. */
876 td->td_base_pri = prio;
879 * If the thread is borrowing another thread's priority, don't ever
880 * lower the priority.
882 if (td->td_flags & TDF_BORROWING && td->td_priority < prio)
885 /* Change the real priority. */
886 oldprio = td->td_priority;
887 sched_priority(td, prio);
890 * If the thread is on a turnstile, then let the turnstile update
893 if (TD_ON_LOCK(td) && oldprio != prio)
894 turnstile_adjust(td, oldprio);
898 sched_user_prio(struct thread *td, u_char prio)
902 THREAD_LOCK_ASSERT(td, MA_OWNED);
903 td->td_base_user_pri = prio;
904 if (td->td_flags & TDF_UBORROWING && td->td_user_pri <= prio)
906 oldprio = td->td_user_pri;
907 td->td_user_pri = prio;
911 sched_lend_user_prio(struct thread *td, u_char prio)
915 THREAD_LOCK_ASSERT(td, MA_OWNED);
916 td->td_flags |= TDF_UBORROWING;
917 oldprio = td->td_user_pri;
918 td->td_user_pri = prio;
922 sched_unlend_user_prio(struct thread *td, u_char prio)
926 THREAD_LOCK_ASSERT(td, MA_OWNED);
927 base_pri = td->td_base_user_pri;
928 if (prio >= base_pri) {
929 td->td_flags &= ~TDF_UBORROWING;
930 sched_user_prio(td, base_pri);
932 sched_lend_user_prio(td, prio);
937 sched_sleep(struct thread *td, int pri)
940 THREAD_LOCK_ASSERT(td, MA_OWNED);
941 td->td_slptick = ticks;
942 td->td_sched->ts_slptime = 0;
945 if (TD_IS_SUSPENDED(td) || pri >= PSOCK)
946 td->td_flags |= TDF_CANSWAP;
950 sched_switch(struct thread *td, struct thread *newtd, int flags)
960 THREAD_LOCK_ASSERT(td, MA_OWNED);
963 * Switch to the sched lock to fix things up and pick
965 * Block the td_lock in order to avoid breaking the critical path.
967 if (td->td_lock != &sched_lock) {
968 mtx_lock_spin(&sched_lock);
969 tmtx = thread_lock_block(td);
972 if ((td->td_flags & TDF_NOLOAD) == 0)
975 td->td_lastcpu = td->td_oncpu;
976 if (!(flags & SW_PREEMPT))
977 td->td_flags &= ~TDF_NEEDRESCHED;
978 td->td_owepreempt = 0;
979 td->td_oncpu = NOCPU;
982 * At the last moment, if this thread is still marked RUNNING,
983 * then put it back on the run queue as it has not been suspended
984 * or stopped or any thing else similar. We never put the idle
985 * threads on the run queue, however.
987 if (td->td_flags & TDF_IDLETD) {
990 idle_cpus_mask &= ~PCPU_GET(cpumask);
993 if (TD_IS_RUNNING(td)) {
994 /* Put us back on the run queue. */
995 sched_add(td, (flags & SW_PREEMPT) ?
996 SRQ_OURSELF|SRQ_YIELDING|SRQ_PREEMPTED :
997 SRQ_OURSELF|SRQ_YIELDING);
1002 * The thread we are about to run needs to be counted
1003 * as if it had been added to the run queue and selected.
1009 KASSERT((newtd->td_inhibitors == 0),
1010 ("trying to run inhibited thread"));
1011 newtd->td_flags |= TDF_DIDRUN;
1012 TD_SET_RUNNING(newtd);
1013 if ((newtd->td_flags & TDF_NOLOAD) == 0)
1016 newtd = choosethread();
1017 MPASS(newtd->td_lock == &sched_lock);
1022 if (PMC_PROC_IS_USING_PMCS(td->td_proc))
1023 PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_OUT);
1026 SDT_PROBE2(sched, , , off_cpu, td, td->td_proc);
1029 lock_profile_release_lock(&sched_lock.lock_object);
1030 #ifdef KDTRACE_HOOKS
1032 * If DTrace has set the active vtime enum to anything
1033 * other than INACTIVE (0), then it should have set the
1036 if (dtrace_vtime_active)
1037 (*dtrace_vtime_switch_func)(newtd);
1040 cpu_switch(td, newtd, tmtx != NULL ? tmtx : td->td_lock);
1041 lock_profile_obtain_lock_success(&sched_lock.lock_object,
1042 0, 0, __FILE__, __LINE__);
1044 * Where am I? What year is it?
1045 * We are in the same thread that went to sleep above,
1046 * but any amount of time may have passed. All our context
1047 * will still be available as will local variables.
1048 * PCPU values however may have changed as we may have
1049 * changed CPU so don't trust cached values of them.
1050 * New threads will go to fork_exit() instead of here
1051 * so if you change things here you may need to change
1054 * If the thread above was exiting it will never wake
1055 * up again here, so either it has saved everything it
1056 * needed to, or the thread_wait() or wait() will
1060 SDT_PROBE0(sched, , , on_cpu);
1062 if (PMC_PROC_IS_USING_PMCS(td->td_proc))
1063 PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_IN);
1066 SDT_PROBE0(sched, , , remain_cpu);
1069 if (td->td_flags & TDF_IDLETD)
1070 idle_cpus_mask |= PCPU_GET(cpumask);
1072 sched_lock.mtx_lock = (uintptr_t)td;
1073 td->td_oncpu = PCPU_GET(cpuid);
1074 MPASS(td->td_lock == &sched_lock);
1078 sched_wakeup(struct thread *td)
1080 struct td_sched *ts;
1082 THREAD_LOCK_ASSERT(td, MA_OWNED);
1084 td->td_flags &= ~TDF_CANSWAP;
1085 if (ts->ts_slptime > 1) {
1091 sched_add(td, SRQ_BORING);
1096 forward_wakeup(int cpunum)
1099 cpumask_t dontuse, id, map, map2, map3, me;
1101 mtx_assert(&sched_lock, MA_OWNED);
1103 CTR0(KTR_RUNQ, "forward_wakeup()");
1105 if ((!forward_wakeup_enabled) ||
1106 (forward_wakeup_use_mask == 0 && forward_wakeup_use_loop == 0))
1108 if (!smp_started || cold || panicstr)
1111 forward_wakeups_requested++;
1114 * Check the idle mask we received against what we calculated
1115 * before in the old version.
1117 me = PCPU_GET(cpumask);
1119 /* Don't bother if we should be doing it ourself. */
1120 if ((me & idle_cpus_mask) && (cpunum == NOCPU || me == (1 << cpunum)))
1123 dontuse = me | stopped_cpus | hlt_cpus_mask;
1125 if (forward_wakeup_use_loop) {
1126 SLIST_FOREACH(pc, &cpuhead, pc_allcpu) {
1127 id = pc->pc_cpumask;
1128 if ((id & dontuse) == 0 &&
1129 pc->pc_curthread == pc->pc_idlethread) {
1135 if (forward_wakeup_use_mask) {
1137 map = idle_cpus_mask & ~dontuse;
1139 /* If they are both on, compare and use loop if different. */
1140 if (forward_wakeup_use_loop) {
1142 printf("map (%02X) != map3 (%02X)\n", map,
1151 /* If we only allow a specific CPU, then mask off all the others. */
1152 if (cpunum != NOCPU) {
1153 KASSERT((cpunum <= mp_maxcpus),("forward_wakeup: bad cpunum."));
1154 map &= (1 << cpunum);
1156 /* Try choose an idle die. */
1157 if (forward_wakeup_use_htt) {
1158 map2 = (map & (map >> 1)) & 0x5555;
1164 /* Set only one bit. */
1165 if (forward_wakeup_use_single) {
1166 map = map & ((~map) + 1);
1170 forward_wakeups_delivered++;
1171 ipi_selected(map, IPI_AST);
1174 if (cpunum == NOCPU)
1175 printf("forward_wakeup: Idle processor not found\n");
1180 kick_other_cpu(int pri, int cpuid)
1185 pcpu = pcpu_find(cpuid);
1186 if (idle_cpus_mask & pcpu->pc_cpumask) {
1187 forward_wakeups_delivered++;
1188 ipi_cpu(cpuid, IPI_AST);
1192 cpri = pcpu->pc_curthread->td_priority;
1196 #if defined(IPI_PREEMPTION) && defined(PREEMPTION)
1197 #if !defined(FULL_PREEMPTION)
1198 if (pri <= PRI_MAX_ITHD)
1199 #endif /* ! FULL_PREEMPTION */
1201 ipi_cpu(cpuid, IPI_PREEMPT);
1204 #endif /* defined(IPI_PREEMPTION) && defined(PREEMPTION) */
1206 pcpu->pc_curthread->td_flags |= TDF_NEEDRESCHED;
1207 ipi_cpu(cpuid, IPI_AST);
1214 sched_pickcpu(struct thread *td)
1218 mtx_assert(&sched_lock, MA_OWNED);
1220 if (THREAD_CAN_SCHED(td, td->td_lastcpu))
1221 best = td->td_lastcpu;
1225 if (!THREAD_CAN_SCHED(td, cpu))
1230 else if (runq_length[cpu] < runq_length[best])
1233 KASSERT(best != NOCPU, ("no valid CPUs"));
1240 sched_add(struct thread *td, int flags)
1243 struct td_sched *ts;
1249 THREAD_LOCK_ASSERT(td, MA_OWNED);
1250 KASSERT((td->td_inhibitors == 0),
1251 ("sched_add: trying to run inhibited thread"));
1252 KASSERT((TD_CAN_RUN(td) || TD_IS_RUNNING(td)),
1253 ("sched_add: bad thread state"));
1254 KASSERT(td->td_flags & TDF_INMEM,
1255 ("sched_add: thread swapped out"));
1257 KTR_STATE2(KTR_SCHED, "thread", sched_tdname(td), "runq add",
1258 "prio:%d", td->td_priority, KTR_ATTR_LINKED,
1259 sched_tdname(curthread));
1260 KTR_POINT1(KTR_SCHED, "thread", sched_tdname(curthread), "wokeup",
1261 KTR_ATTR_LINKED, sched_tdname(td));
1262 SDT_PROBE4(sched, , , enqueue, td, td->td_proc, NULL,
1263 flags & SRQ_PREEMPTED);
1267 * Now that the thread is moving to the run-queue, set the lock
1268 * to the scheduler's lock.
1270 if (td->td_lock != &sched_lock) {
1271 mtx_lock_spin(&sched_lock);
1272 thread_lock_set(td, &sched_lock);
1277 * If SMP is started and the thread is pinned or otherwise limited to
1278 * a specific set of CPUs, queue the thread to a per-CPU run queue.
1279 * Otherwise, queue the thread to the global run queue.
1281 * If SMP has not yet been started we must use the global run queue
1282 * as per-CPU state may not be initialized yet and we may crash if we
1283 * try to access the per-CPU run queues.
1285 if (smp_started && (td->td_pinned != 0 || td->td_flags & TDF_BOUND ||
1286 ts->ts_flags & TSF_AFFINITY)) {
1287 if (td->td_pinned != 0)
1288 cpu = td->td_lastcpu;
1289 else if (td->td_flags & TDF_BOUND) {
1290 /* Find CPU from bound runq. */
1291 KASSERT(SKE_RUNQ_PCPU(ts),
1292 ("sched_add: bound td_sched not on cpu runq"));
1293 cpu = ts->ts_runq - &runq_pcpu[0];
1295 /* Find a valid CPU for our cpuset */
1296 cpu = sched_pickcpu(td);
1297 ts->ts_runq = &runq_pcpu[cpu];
1300 "sched_add: Put td_sched:%p(td:%p) on cpu%d runq", ts, td,
1304 "sched_add: adding td_sched:%p (td:%p) to gbl runq", ts,
1307 ts->ts_runq = &runq;
1310 if (single_cpu && (cpu != PCPU_GET(cpuid))) {
1311 kick_other_cpu(td->td_priority, cpu);
1314 cpumask_t me = PCPU_GET(cpumask);
1315 cpumask_t idle = idle_cpus_mask & me;
1317 if (!idle && ((flags & SRQ_INTR) == 0) &&
1318 (idle_cpus_mask & ~(hlt_cpus_mask | me)))
1319 forwarded = forward_wakeup(cpu);
1323 if ((flags & SRQ_YIELDING) == 0 && maybe_preempt(td))
1330 if ((td->td_flags & TDF_NOLOAD) == 0)
1332 runq_add(ts->ts_runq, td, flags);
1338 struct td_sched *ts;
1341 THREAD_LOCK_ASSERT(td, MA_OWNED);
1342 KASSERT((td->td_inhibitors == 0),
1343 ("sched_add: trying to run inhibited thread"));
1344 KASSERT((TD_CAN_RUN(td) || TD_IS_RUNNING(td)),
1345 ("sched_add: bad thread state"));
1346 KASSERT(td->td_flags & TDF_INMEM,
1347 ("sched_add: thread swapped out"));
1348 KTR_STATE2(KTR_SCHED, "thread", sched_tdname(td), "runq add",
1349 "prio:%d", td->td_priority, KTR_ATTR_LINKED,
1350 sched_tdname(curthread));
1351 KTR_POINT1(KTR_SCHED, "thread", sched_tdname(curthread), "wokeup",
1352 KTR_ATTR_LINKED, sched_tdname(td));
1353 SDT_PROBE4(sched, , , enqueue, td, td->td_proc, NULL,
1354 flags & SRQ_PREEMPTED);
1357 * Now that the thread is moving to the run-queue, set the lock
1358 * to the scheduler's lock.
1360 if (td->td_lock != &sched_lock) {
1361 mtx_lock_spin(&sched_lock);
1362 thread_lock_set(td, &sched_lock);
1365 CTR2(KTR_RUNQ, "sched_add: adding td_sched:%p (td:%p) to runq", ts, td);
1366 ts->ts_runq = &runq;
1369 * If we are yielding (on the way out anyhow) or the thread
1370 * being saved is US, then don't try be smart about preemption
1371 * or kicking off another CPU as it won't help and may hinder.
1372 * In the YIEDLING case, we are about to run whoever is being
1373 * put in the queue anyhow, and in the OURSELF case, we are
1374 * puting ourself on the run queue which also only happens
1375 * when we are about to yield.
1377 if ((flags & SRQ_YIELDING) == 0) {
1378 if (maybe_preempt(td))
1381 if ((td->td_flags & TDF_NOLOAD) == 0)
1383 runq_add(ts->ts_runq, td, flags);
1389 sched_rem(struct thread *td)
1391 struct td_sched *ts;
1394 KASSERT(td->td_flags & TDF_INMEM,
1395 ("sched_rem: thread swapped out"));
1396 KASSERT(TD_ON_RUNQ(td),
1397 ("sched_rem: thread not on run queue"));
1398 mtx_assert(&sched_lock, MA_OWNED);
1399 KTR_STATE2(KTR_SCHED, "thread", sched_tdname(td), "runq rem",
1400 "prio:%d", td->td_priority, KTR_ATTR_LINKED,
1401 sched_tdname(curthread));
1402 SDT_PROBE3(sched, , , dequeue, td, td->td_proc, NULL);
1404 if ((td->td_flags & TDF_NOLOAD) == 0)
1407 if (ts->ts_runq != &runq)
1408 runq_length[ts->ts_runq - runq_pcpu]--;
1410 runq_remove(ts->ts_runq, td);
1415 * Select threads to run. Note that running threads still consume a
1424 mtx_assert(&sched_lock, MA_OWNED);
1426 struct thread *tdcpu;
1429 td = runq_choose_fuzz(&runq, runq_fuzz);
1430 tdcpu = runq_choose(&runq_pcpu[PCPU_GET(cpuid)]);
1434 tdcpu->td_priority < td->td_priority)) {
1435 CTR2(KTR_RUNQ, "choosing td %p from pcpu runq %d", tdcpu,
1438 rq = &runq_pcpu[PCPU_GET(cpuid)];
1440 CTR1(KTR_RUNQ, "choosing td_sched %p from main runq", td);
1445 td = runq_choose(&runq);
1451 runq_length[PCPU_GET(cpuid)]--;
1453 runq_remove(rq, td);
1454 td->td_flags |= TDF_DIDRUN;
1456 KASSERT(td->td_flags & TDF_INMEM,
1457 ("sched_choose: thread swapped out"));
1460 return (PCPU_GET(idlethread));
1464 sched_preempt(struct thread *td)
1467 SDT_PROBE2(sched, , , surrender, td, td->td_proc);
1469 if (td->td_critnest > 1)
1470 td->td_owepreempt = 1;
1472 mi_switch(SW_INVOL | SW_PREEMPT | SWT_PREEMPT, NULL);
1477 sched_userret(struct thread *td)
1480 * XXX we cheat slightly on the locking here to avoid locking in
1481 * the usual case. Setting td_priority here is essentially an
1482 * incomplete workaround for not setting it properly elsewhere.
1483 * Now that some interrupt handlers are threads, not setting it
1484 * properly elsewhere can clobber it in the window between setting
1485 * it here and returning to user mode, so don't waste time setting
1486 * it perfectly here.
1488 KASSERT((td->td_flags & TDF_BORROWING) == 0,
1489 ("thread with borrowed priority returning to userland"));
1490 if (td->td_priority != td->td_user_pri) {
1492 td->td_priority = td->td_user_pri;
1493 td->td_base_pri = td->td_user_pri;
1499 sched_bind(struct thread *td, int cpu)
1501 struct td_sched *ts;
1503 THREAD_LOCK_ASSERT(td, MA_OWNED|MA_NOTRECURSED);
1504 KASSERT(td == curthread, ("sched_bind: can only bind curthread"));
1508 td->td_flags |= TDF_BOUND;
1510 ts->ts_runq = &runq_pcpu[cpu];
1511 if (PCPU_GET(cpuid) == cpu)
1514 mi_switch(SW_VOL, NULL);
1519 sched_unbind(struct thread* td)
1521 THREAD_LOCK_ASSERT(td, MA_OWNED);
1522 KASSERT(td == curthread, ("sched_unbind: can only bind curthread"));
1523 td->td_flags &= ~TDF_BOUND;
1527 sched_is_bound(struct thread *td)
1529 THREAD_LOCK_ASSERT(td, MA_OWNED);
1530 return (td->td_flags & TDF_BOUND);
1534 sched_relinquish(struct thread *td)
1537 mi_switch(SW_VOL | SWT_RELINQUISH, NULL);
1544 return (sched_tdcnt);
1548 sched_sizeof_proc(void)
1550 return (sizeof(struct proc));
1554 sched_sizeof_thread(void)
1556 return (sizeof(struct thread) + sizeof(struct td_sched));
1560 sched_pctcpu(struct thread *td)
1562 struct td_sched *ts;
1564 THREAD_LOCK_ASSERT(td, MA_OWNED);
1566 return (ts->ts_pctcpu);
1575 * The actual idle process.
1578 sched_idletd(void *dummy)
1582 mtx_assert(&Giant, MA_NOTOWNED);
1584 while (sched_runnable() == 0)
1587 mtx_lock_spin(&sched_lock);
1588 mi_switch(SW_VOL | SWT_IDLE, NULL);
1589 mtx_unlock_spin(&sched_lock);
1594 * A CPU is entering for the first time or a thread is exiting.
1597 sched_throw(struct thread *td)
1600 * Correct spinlock nesting. The idle thread context that we are
1601 * borrowing was created so that it would start out with a single
1602 * spin lock (sched_lock) held in fork_trampoline(). Since we've
1603 * explicitly acquired locks in this function, the nesting count
1604 * is now 2 rather than 1. Since we are nested, calling
1605 * spinlock_exit() will simply adjust the counts without allowing
1606 * spin lock using code to interrupt us.
1609 mtx_lock_spin(&sched_lock);
1611 PCPU_SET(switchtime, cpu_ticks());
1612 PCPU_SET(switchticks, ticks);
1614 lock_profile_release_lock(&sched_lock.lock_object);
1615 MPASS(td->td_lock == &sched_lock);
1617 mtx_assert(&sched_lock, MA_OWNED);
1618 KASSERT(curthread->td_md.md_spinlock_count == 1, ("invalid count"));
1619 cpu_throw(td, choosethread()); /* doesn't return */
1623 sched_fork_exit(struct thread *td)
1627 * Finish setting up thread glue so that it begins execution in a
1628 * non-nested critical section with sched_lock held but not recursed.
1630 td->td_oncpu = PCPU_GET(cpuid);
1631 sched_lock.mtx_lock = (uintptr_t)td;
1632 lock_profile_obtain_lock_success(&sched_lock.lock_object,
1633 0, 0, __FILE__, __LINE__);
1634 THREAD_LOCK_ASSERT(td, MA_OWNED | MA_NOTRECURSED);
1638 sched_tdname(struct thread *td)
1641 struct td_sched *ts;
1644 if (ts->ts_name[0] == '\0')
1645 snprintf(ts->ts_name, sizeof(ts->ts_name),
1646 "%s tid %d", td->td_name, td->td_tid);
1647 return (ts->ts_name);
1649 return (td->td_name);
1655 sched_clear_tdname(struct thread *td)
1657 struct td_sched *ts;
1660 ts->ts_name[0] = '\0';
1665 sched_affinity(struct thread *td)
1668 struct td_sched *ts;
1671 THREAD_LOCK_ASSERT(td, MA_OWNED);
1674 * Set the TSF_AFFINITY flag if there is at least one CPU this
1675 * thread can't run on.
1678 ts->ts_flags &= ~TSF_AFFINITY;
1680 if (!THREAD_CAN_SCHED(td, cpu)) {
1681 ts->ts_flags |= TSF_AFFINITY;
1687 * If this thread can run on all CPUs, nothing else to do.
1689 if (!(ts->ts_flags & TSF_AFFINITY))
1692 /* Pinned threads and bound threads should be left alone. */
1693 if (td->td_pinned != 0 || td->td_flags & TDF_BOUND)
1696 switch (td->td_state) {
1699 * If we are on a per-CPU runqueue that is in the set,
1700 * then nothing needs to be done.
1702 if (ts->ts_runq != &runq &&
1703 THREAD_CAN_SCHED(td, ts->ts_runq - runq_pcpu))
1706 /* Put this thread on a valid per-CPU runqueue. */
1708 sched_add(td, SRQ_BORING);
1712 * See if our current CPU is in the set. If not, force a
1715 if (THREAD_CAN_SCHED(td, td->td_oncpu))
1718 td->td_flags |= TDF_NEEDRESCHED;
1719 if (td != curthread)
1720 ipi_cpu(cpu, IPI_AST);