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>
54 #include <sys/sysctl.h>
56 #include <sys/turnstile.h>
58 #include <machine/pcb.h>
59 #include <machine/smp.h>
62 #include <sys/pmckern.h>
66 #include <sys/dtrace_bsd.h>
67 int dtrace_vtime_active;
68 dtrace_vtime_switch_func_t dtrace_vtime_switch_func;
72 * INVERSE_ESTCPU_WEIGHT is only suitable for statclock() frequencies in
73 * the range 100-256 Hz (approximately).
75 #define ESTCPULIM(e) \
76 min((e), INVERSE_ESTCPU_WEIGHT * (NICE_WEIGHT * (PRIO_MAX - PRIO_MIN) - \
77 RQ_PPQ) + INVERSE_ESTCPU_WEIGHT - 1)
79 #define INVERSE_ESTCPU_WEIGHT (8 * smp_cpus)
81 #define INVERSE_ESTCPU_WEIGHT 8 /* 1 / (priorities per estcpu level). */
83 #define NICE_WEIGHT 1 /* Priorities per nice level. */
85 #define TS_NAME_LEN (MAXCOMLEN + sizeof(" td ") + sizeof(__XSTRING(UINT_MAX)))
88 * The schedulable entity that runs a context.
89 * This is an extension to the thread structure and is tailored to
90 * the requirements of this scheduler
93 fixpt_t ts_pctcpu; /* (j) %cpu during p_swtime. */
94 int ts_cpticks; /* (j) Ticks of cpu time. */
95 int ts_slptime; /* (j) Seconds !RUNNING. */
97 struct runq *ts_runq; /* runq the thread is currently on */
99 char ts_name[TS_NAME_LEN];
103 /* flags kept in td_flags */
104 #define TDF_DIDRUN TDF_SCHED0 /* thread actually ran. */
105 #define TDF_BOUND TDF_SCHED1 /* Bound to one CPU. */
107 /* flags kept in ts_flags */
108 #define TSF_AFFINITY 0x0001 /* Has a non-"full" CPU set. */
110 #define SKE_RUNQ_PCPU(ts) \
111 ((ts)->ts_runq != 0 && (ts)->ts_runq != &runq)
113 #define THREAD_CAN_SCHED(td, cpu) \
114 CPU_ISSET((cpu), &(td)->td_cpuset->cs_mask)
116 static struct td_sched td_sched0;
117 struct mtx sched_lock;
119 static int sched_tdcnt; /* Total runnable threads in the system. */
120 static int sched_quantum; /* Roundrobin scheduling quantum in ticks. */
121 #define SCHED_QUANTUM (hz / 10) /* Default sched quantum */
123 static void setup_runqs(void);
124 static void schedcpu(void);
125 static void schedcpu_thread(void);
126 static void sched_priority(struct thread *td, u_char prio);
127 static void sched_setup(void *dummy);
128 static void maybe_resched(struct thread *td);
129 static void updatepri(struct thread *td);
130 static void resetpriority(struct thread *td);
131 static void resetpriority_thread(struct thread *td);
133 static int sched_pickcpu(struct thread *td);
134 static int forward_wakeup(int cpunum);
135 static void kick_other_cpu(int pri, int cpuid);
138 static struct kproc_desc sched_kp = {
143 SYSINIT(schedcpu, SI_SUB_RUN_SCHEDULER, SI_ORDER_FIRST, kproc_start,
145 SYSINIT(sched_setup, SI_SUB_RUN_QUEUE, SI_ORDER_FIRST, sched_setup, NULL);
150 static struct runq runq;
156 static struct runq runq_pcpu[MAXCPU];
157 long runq_length[MAXCPU];
159 static cpuset_t idle_cpus_mask;
162 struct pcpuidlestat {
166 static DPCPU_DEFINE(struct pcpuidlestat, idlestat);
174 for (i = 0; i < MAXCPU; ++i)
175 runq_init(&runq_pcpu[i]);
182 sysctl_kern_quantum(SYSCTL_HANDLER_ARGS)
186 new_val = sched_quantum * tick;
187 error = sysctl_handle_int(oidp, &new_val, 0, req);
188 if (error != 0 || req->newptr == NULL)
192 sched_quantum = new_val / tick;
193 hogticks = 2 * sched_quantum;
197 SYSCTL_NODE(_kern, OID_AUTO, sched, CTLFLAG_RD, 0, "Scheduler");
199 SYSCTL_STRING(_kern_sched, OID_AUTO, name, CTLFLAG_RD, "4BSD", 0,
202 SYSCTL_PROC(_kern_sched, OID_AUTO, quantum, CTLTYPE_INT | CTLFLAG_RW,
203 0, sizeof sched_quantum, sysctl_kern_quantum, "I",
204 "Roundrobin scheduling quantum in microseconds");
207 /* Enable forwarding of wakeups to all other cpus */
208 SYSCTL_NODE(_kern_sched, OID_AUTO, ipiwakeup, CTLFLAG_RD, NULL, "Kernel SMP");
210 static int runq_fuzz = 1;
211 SYSCTL_INT(_kern_sched, OID_AUTO, runq_fuzz, CTLFLAG_RW, &runq_fuzz, 0, "");
213 static int forward_wakeup_enabled = 1;
214 SYSCTL_INT(_kern_sched_ipiwakeup, OID_AUTO, enabled, CTLFLAG_RW,
215 &forward_wakeup_enabled, 0,
216 "Forwarding of wakeup to idle CPUs");
218 static int forward_wakeups_requested = 0;
219 SYSCTL_INT(_kern_sched_ipiwakeup, OID_AUTO, requested, CTLFLAG_RD,
220 &forward_wakeups_requested, 0,
221 "Requests for Forwarding of wakeup to idle CPUs");
223 static int forward_wakeups_delivered = 0;
224 SYSCTL_INT(_kern_sched_ipiwakeup, OID_AUTO, delivered, CTLFLAG_RD,
225 &forward_wakeups_delivered, 0,
226 "Completed Forwarding of wakeup to idle CPUs");
228 static int forward_wakeup_use_mask = 1;
229 SYSCTL_INT(_kern_sched_ipiwakeup, OID_AUTO, usemask, CTLFLAG_RW,
230 &forward_wakeup_use_mask, 0,
231 "Use the mask of idle cpus");
233 static int forward_wakeup_use_loop = 0;
234 SYSCTL_INT(_kern_sched_ipiwakeup, OID_AUTO, useloop, CTLFLAG_RW,
235 &forward_wakeup_use_loop, 0,
236 "Use a loop to find idle cpus");
240 static int sched_followon = 0;
241 SYSCTL_INT(_kern_sched, OID_AUTO, followon, CTLFLAG_RW,
243 "allow threads to share a quantum");
251 KTR_COUNTER0(KTR_SCHED, "load", "global load", sched_tdcnt);
259 KTR_COUNTER0(KTR_SCHED, "load", "global load", sched_tdcnt);
262 * Arrange to reschedule if necessary, taking the priorities and
263 * schedulers into account.
266 maybe_resched(struct thread *td)
269 THREAD_LOCK_ASSERT(td, MA_OWNED);
270 if (td->td_priority < curthread->td_priority)
271 curthread->td_flags |= TDF_NEEDRESCHED;
275 * This function is called when a thread is about to be put on run queue
276 * because it has been made runnable or its priority has been adjusted. It
277 * determines if the new thread should be immediately preempted to. If so,
278 * it switches to it and eventually returns true. If not, it returns false
279 * so that the caller may place the thread on an appropriate run queue.
282 maybe_preempt(struct thread *td)
289 * The new thread should not preempt the current thread if any of the
290 * following conditions are true:
292 * - The kernel is in the throes of crashing (panicstr).
293 * - The current thread has a higher (numerically lower) or
294 * equivalent priority. Note that this prevents curthread from
295 * trying to preempt to itself.
296 * - It is too early in the boot for context switches (cold is set).
297 * - The current thread has an inhibitor set or is in the process of
298 * exiting. In this case, the current thread is about to switch
299 * out anyways, so there's no point in preempting. If we did,
300 * the current thread would not be properly resumed as well, so
301 * just avoid that whole landmine.
302 * - If the new thread's priority is not a realtime priority and
303 * the current thread's priority is not an idle priority and
304 * FULL_PREEMPTION is disabled.
306 * If all of these conditions are false, but the current thread is in
307 * a nested critical section, then we have to defer the preemption
308 * until we exit the critical section. Otherwise, switch immediately
312 THREAD_LOCK_ASSERT(td, MA_OWNED);
313 KASSERT((td->td_inhibitors == 0),
314 ("maybe_preempt: trying to run inhibited thread"));
315 pri = td->td_priority;
316 cpri = ctd->td_priority;
317 if (panicstr != NULL || pri >= cpri || cold /* || dumping */ ||
318 TD_IS_INHIBITED(ctd))
320 #ifndef FULL_PREEMPTION
321 if (pri > PRI_MAX_ITHD && cpri < PRI_MIN_IDLE)
325 if (ctd->td_critnest > 1) {
326 CTR1(KTR_PROC, "maybe_preempt: in critical section %d",
328 ctd->td_owepreempt = 1;
332 * Thread is runnable but not yet put on system run queue.
334 MPASS(ctd->td_lock == td->td_lock);
335 MPASS(TD_ON_RUNQ(td));
337 CTR3(KTR_PROC, "preempting to thread %p (pid %d, %s)\n", td,
338 td->td_proc->p_pid, td->td_name);
339 mi_switch(SW_INVOL | SW_PREEMPT | SWT_PREEMPT, td);
341 * td's lock pointer may have changed. We have to return with it
355 * Constants for digital decay and forget:
356 * 90% of (td_estcpu) usage in 5 * loadav time
357 * 95% of (ts_pctcpu) usage in 60 seconds (load insensitive)
358 * Note that, as ps(1) mentions, this can let percentages
359 * total over 100% (I've seen 137.9% for 3 processes).
361 * Note that schedclock() updates td_estcpu and p_cpticks asynchronously.
363 * We wish to decay away 90% of td_estcpu in (5 * loadavg) seconds.
364 * That is, the system wants to compute a value of decay such
365 * that the following for loop:
366 * for (i = 0; i < (5 * loadavg); i++)
367 * td_estcpu *= decay;
370 * for all values of loadavg:
372 * Mathematically this loop can be expressed by saying:
373 * decay ** (5 * loadavg) ~= .1
375 * The system computes decay as:
376 * decay = (2 * loadavg) / (2 * loadavg + 1)
378 * We wish to prove that the system's computation of decay
379 * will always fulfill the equation:
380 * decay ** (5 * loadavg) ~= .1
382 * If we compute b as:
385 * decay = b / (b + 1)
387 * We now need to prove two things:
388 * 1) Given factor ** (5 * loadavg) ~= .1, prove factor == b/(b+1)
389 * 2) Given b/(b+1) ** power ~= .1, prove power == (5 * loadavg)
392 * For x close to zero, exp(x) =~ 1 + x, since
393 * exp(x) = 0! + x**1/1! + x**2/2! + ... .
394 * therefore exp(-1/b) =~ 1 - (1/b) = (b-1)/b.
395 * For x close to zero, ln(1+x) =~ x, since
396 * ln(1+x) = x - x**2/2 + x**3/3 - ... -1 < x < 1
397 * therefore ln(b/(b+1)) = ln(1 - 1/(b+1)) =~ -1/(b+1).
401 * Solve (factor)**(power) =~ .1 given power (5*loadav):
402 * solving for factor,
403 * ln(factor) =~ (-2.30/5*loadav), or
404 * factor =~ exp(-1/((5/2.30)*loadav)) =~ exp(-1/(2*loadav)) =
405 * exp(-1/b) =~ (b-1)/b =~ b/(b+1). QED
408 * Solve (factor)**(power) =~ .1 given factor == (b/(b+1)):
410 * power*ln(b/(b+1)) =~ -2.30, or
411 * power =~ 2.3 * (b + 1) = 4.6*loadav + 2.3 =~ 5*loadav. QED
413 * Actual power values for the implemented algorithm are as follows:
415 * power: 5.68 10.32 14.94 19.55
418 /* calculations for digital decay to forget 90% of usage in 5*loadav sec */
419 #define loadfactor(loadav) (2 * (loadav))
420 #define decay_cpu(loadfac, cpu) (((loadfac) * (cpu)) / ((loadfac) + FSCALE))
422 /* decay 95% of `ts_pctcpu' in 60 seconds; see CCPU_SHIFT before changing */
423 static fixpt_t ccpu = 0.95122942450071400909 * FSCALE; /* exp(-1/20) */
424 SYSCTL_UINT(_kern, OID_AUTO, ccpu, CTLFLAG_RD, &ccpu, 0, "");
427 * If `ccpu' is not equal to `exp(-1/20)' and you still want to use the
428 * faster/more-accurate formula, you'll have to estimate CCPU_SHIFT below
429 * and possibly adjust FSHIFT in "param.h" so that (FSHIFT >= CCPU_SHIFT).
431 * To estimate CCPU_SHIFT for exp(-1/20), the following formula was used:
432 * 1 - exp(-1/20) ~= 0.0487 ~= 0.0488 == 1 (fixed pt, *11* bits).
434 * If you don't want to bother with the faster/more-accurate formula, you
435 * can set CCPU_SHIFT to (FSHIFT + 1) which will use a slower/less-accurate
436 * (more general) method of calculating the %age of CPU used by a process.
438 #define CCPU_SHIFT 11
441 * Recompute process priorities, every hz ticks.
442 * MP-safe, called without the Giant mutex.
448 register fixpt_t loadfac = loadfactor(averunnable.ldavg[0]);
452 int awake, realstathz;
454 realstathz = stathz ? stathz : hz;
455 sx_slock(&allproc_lock);
456 FOREACH_PROC_IN_SYSTEM(p) {
458 if (p->p_state == PRS_NEW) {
462 FOREACH_THREAD_IN_PROC(p, td) {
467 * Increment sleep time (if sleeping). We
468 * ignore overflow, as above.
471 * The td_sched slptimes are not touched in wakeup
472 * because the thread may not HAVE everything in
473 * memory? XXX I think this is out of date.
475 if (TD_ON_RUNQ(td)) {
477 td->td_flags &= ~TDF_DIDRUN;
478 } else if (TD_IS_RUNNING(td)) {
480 /* Do not clear TDF_DIDRUN */
481 } else if (td->td_flags & TDF_DIDRUN) {
483 td->td_flags &= ~TDF_DIDRUN;
487 * ts_pctcpu is only for ps and ttyinfo().
489 ts->ts_pctcpu = (ts->ts_pctcpu * ccpu) >> FSHIFT;
491 * If the td_sched has been idle the entire second,
492 * stop recalculating its priority until
495 if (ts->ts_cpticks != 0) {
496 #if (FSHIFT >= CCPU_SHIFT)
497 ts->ts_pctcpu += (realstathz == 100)
498 ? ((fixpt_t) ts->ts_cpticks) <<
499 (FSHIFT - CCPU_SHIFT) :
500 100 * (((fixpt_t) ts->ts_cpticks)
501 << (FSHIFT - CCPU_SHIFT)) / realstathz;
503 ts->ts_pctcpu += ((FSCALE - ccpu) *
505 FSCALE / realstathz)) >> FSHIFT;
510 * If there are ANY running threads in this process,
511 * then don't count it as sleeping.
512 * XXX: this is broken.
515 if (ts->ts_slptime > 1) {
517 * In an ideal world, this should not
518 * happen, because whoever woke us
519 * up from the long sleep should have
520 * unwound the slptime and reset our
521 * priority before we run at the stale
522 * priority. Should KASSERT at some
523 * point when all the cases are fixed.
530 if (ts->ts_slptime > 1) {
534 td->td_estcpu = decay_cpu(loadfac, td->td_estcpu);
536 resetpriority_thread(td);
541 sx_sunlock(&allproc_lock);
545 * Main loop for a kthread that executes schedcpu once a second.
548 schedcpu_thread(void)
558 * Recalculate the priority of a process after it has slept for a while.
559 * For all load averages >= 1 and max td_estcpu of 255, sleeping for at
560 * least six times the loadfactor will decay td_estcpu to zero.
563 updatepri(struct thread *td)
570 loadfac = loadfactor(averunnable.ldavg[0]);
571 if (ts->ts_slptime > 5 * loadfac)
574 newcpu = td->td_estcpu;
575 ts->ts_slptime--; /* was incremented in schedcpu() */
576 while (newcpu && --ts->ts_slptime)
577 newcpu = decay_cpu(loadfac, newcpu);
578 td->td_estcpu = newcpu;
583 * Compute the priority of a process when running in user mode.
584 * Arrange to reschedule if the resulting priority is better
585 * than that of the current process.
588 resetpriority(struct thread *td)
590 register unsigned int newpriority;
592 if (td->td_pri_class == PRI_TIMESHARE) {
593 newpriority = PUSER + td->td_estcpu / INVERSE_ESTCPU_WEIGHT +
594 NICE_WEIGHT * (td->td_proc->p_nice - PRIO_MIN);
595 newpriority = min(max(newpriority, PRI_MIN_TIMESHARE),
597 sched_user_prio(td, newpriority);
602 * Update the thread's priority when the associated process's user
606 resetpriority_thread(struct thread *td)
609 /* Only change threads with a time sharing user priority. */
610 if (td->td_priority < PRI_MIN_TIMESHARE ||
611 td->td_priority > PRI_MAX_TIMESHARE)
614 /* XXX the whole needresched thing is broken, but not silly. */
617 sched_prio(td, td->td_user_pri);
622 sched_setup(void *dummy)
626 if (sched_quantum == 0)
627 sched_quantum = SCHED_QUANTUM;
628 hogticks = 2 * sched_quantum;
630 /* Account for thread0. */
634 /* External interfaces start here */
637 * Very early in the boot some setup of scheduler-specific
638 * parts of proc0 and of some scheduler resources needs to be done.
646 * Set up the scheduler specific parts of proc0.
648 proc0.p_sched = NULL; /* XXX */
649 thread0.td_sched = &td_sched0;
650 thread0.td_lock = &sched_lock;
651 mtx_init(&sched_lock, "sched lock", NULL, MTX_SPIN | MTX_RECURSE);
658 return runq_check(&runq) + runq_check(&runq_pcpu[PCPU_GET(cpuid)]);
660 return runq_check(&runq);
665 sched_rr_interval(void)
667 if (sched_quantum == 0)
668 sched_quantum = SCHED_QUANTUM;
669 return (sched_quantum);
673 * We adjust the priority of the current process. The priority of
674 * a process gets worse as it accumulates CPU time. The cpu usage
675 * estimator (td_estcpu) is increased here. resetpriority() will
676 * compute a different priority each time td_estcpu increases by
677 * INVERSE_ESTCPU_WEIGHT
678 * (until MAXPRI is reached). The cpu usage estimator ramps up
679 * quite quickly when the process is running (linearly), and decays
680 * away exponentially, at a rate which is proportionally slower when
681 * the system is busy. The basic principle is that the system will
682 * 90% forget that the process used a lot of CPU time in 5 * loadav
683 * seconds. This causes the system to favor processes which haven't
684 * run much recently, and to round-robin among other processes.
687 sched_clock(struct thread *td)
689 struct pcpuidlestat *stat;
692 THREAD_LOCK_ASSERT(td, MA_OWNED);
696 td->td_estcpu = ESTCPULIM(td->td_estcpu + 1);
697 if ((td->td_estcpu % INVERSE_ESTCPU_WEIGHT) == 0) {
699 resetpriority_thread(td);
703 * Force a context switch if the current thread has used up a full
704 * quantum (default quantum is 100ms).
706 if (!TD_IS_IDLETHREAD(td) &&
707 ticks - PCPU_GET(switchticks) >= sched_quantum)
708 td->td_flags |= TDF_NEEDRESCHED;
710 stat = DPCPU_PTR(idlestat);
711 stat->oldidlecalls = stat->idlecalls;
716 * Charge child's scheduling CPU usage to parent.
719 sched_exit(struct proc *p, struct thread *td)
722 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(td), "proc exit",
723 "prio:%d", td->td_priority);
725 PROC_LOCK_ASSERT(p, MA_OWNED);
726 sched_exit_thread(FIRST_THREAD_IN_PROC(p), td);
730 sched_exit_thread(struct thread *td, struct thread *child)
733 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(child), "exit",
734 "prio:%d", child->td_priority);
736 td->td_estcpu = ESTCPULIM(td->td_estcpu + child->td_estcpu);
739 if ((child->td_flags & TDF_NOLOAD) == 0)
741 thread_unlock(child);
745 sched_fork(struct thread *td, struct thread *childtd)
747 sched_fork_thread(td, childtd);
751 sched_fork_thread(struct thread *td, struct thread *childtd)
755 childtd->td_estcpu = td->td_estcpu;
756 childtd->td_lock = &sched_lock;
757 childtd->td_cpuset = cpuset_ref(td->td_cpuset);
758 childtd->td_priority = childtd->td_base_pri;
759 ts = childtd->td_sched;
760 bzero(ts, sizeof(*ts));
761 ts->ts_flags |= (td->td_sched->ts_flags & TSF_AFFINITY);
765 sched_nice(struct proc *p, int nice)
769 PROC_LOCK_ASSERT(p, MA_OWNED);
771 FOREACH_THREAD_IN_PROC(p, td) {
774 resetpriority_thread(td);
780 sched_class(struct thread *td, int class)
782 THREAD_LOCK_ASSERT(td, MA_OWNED);
783 td->td_pri_class = class;
787 * Adjust the priority of a thread.
790 sched_priority(struct thread *td, u_char prio)
794 KTR_POINT3(KTR_SCHED, "thread", sched_tdname(td), "priority change",
795 "prio:%d", td->td_priority, "new prio:%d", prio, KTR_ATTR_LINKED,
796 sched_tdname(curthread));
797 if (td != curthread && prio > td->td_priority) {
798 KTR_POINT3(KTR_SCHED, "thread", sched_tdname(curthread),
799 "lend prio", "prio:%d", td->td_priority, "new prio:%d",
800 prio, KTR_ATTR_LINKED, sched_tdname(td));
802 THREAD_LOCK_ASSERT(td, MA_OWNED);
803 if (td->td_priority == prio)
805 td->td_priority = prio;
806 if (TD_ON_RUNQ(td) && td->td_rqindex != (prio / RQ_PPQ)) {
808 sched_add(td, SRQ_BORING);
813 * Update a thread's priority when it is lent another thread's
817 sched_lend_prio(struct thread *td, u_char prio)
820 td->td_flags |= TDF_BORROWING;
821 sched_priority(td, prio);
825 * Restore a thread's priority when priority propagation is
826 * over. The prio argument is the minimum priority the thread
827 * needs to have to satisfy other possible priority lending
828 * requests. If the thread's regulary priority is less
829 * important than prio the thread will keep a priority boost
833 sched_unlend_prio(struct thread *td, u_char prio)
837 if (td->td_base_pri >= PRI_MIN_TIMESHARE &&
838 td->td_base_pri <= PRI_MAX_TIMESHARE)
839 base_pri = td->td_user_pri;
841 base_pri = td->td_base_pri;
842 if (prio >= base_pri) {
843 td->td_flags &= ~TDF_BORROWING;
844 sched_prio(td, base_pri);
846 sched_lend_prio(td, prio);
850 sched_prio(struct thread *td, u_char prio)
854 /* First, update the base priority. */
855 td->td_base_pri = prio;
858 * If the thread is borrowing another thread's priority, don't ever
859 * lower the priority.
861 if (td->td_flags & TDF_BORROWING && td->td_priority < prio)
864 /* Change the real priority. */
865 oldprio = td->td_priority;
866 sched_priority(td, prio);
869 * If the thread is on a turnstile, then let the turnstile update
872 if (TD_ON_LOCK(td) && oldprio != prio)
873 turnstile_adjust(td, oldprio);
877 sched_user_prio(struct thread *td, u_char prio)
880 THREAD_LOCK_ASSERT(td, MA_OWNED);
881 td->td_base_user_pri = prio;
882 if (td->td_lend_user_pri <= prio)
884 td->td_user_pri = prio;
888 sched_lend_user_prio(struct thread *td, u_char prio)
891 THREAD_LOCK_ASSERT(td, MA_OWNED);
892 td->td_lend_user_pri = prio;
893 td->td_user_pri = min(prio, td->td_base_user_pri);
894 if (td->td_priority > td->td_user_pri)
895 sched_prio(td, td->td_user_pri);
896 else if (td->td_priority != td->td_user_pri)
897 td->td_flags |= TDF_NEEDRESCHED;
901 sched_sleep(struct thread *td, int pri)
904 THREAD_LOCK_ASSERT(td, MA_OWNED);
905 td->td_slptick = ticks;
906 td->td_sched->ts_slptime = 0;
907 if (pri != 0 && PRI_BASE(td->td_pri_class) == PRI_TIMESHARE)
909 if (TD_IS_SUSPENDED(td) || pri >= PSOCK)
910 td->td_flags |= TDF_CANSWAP;
914 sched_switch(struct thread *td, struct thread *newtd, int flags)
924 THREAD_LOCK_ASSERT(td, MA_OWNED);
927 * Switch to the sched lock to fix things up and pick
929 * Block the td_lock in order to avoid breaking the critical path.
931 if (td->td_lock != &sched_lock) {
932 mtx_lock_spin(&sched_lock);
933 tmtx = thread_lock_block(td);
936 if ((td->td_flags & TDF_NOLOAD) == 0)
939 td->td_lastcpu = td->td_oncpu;
940 if (!(flags & SW_PREEMPT))
941 td->td_flags &= ~TDF_NEEDRESCHED;
942 td->td_owepreempt = 0;
943 td->td_oncpu = NOCPU;
946 * At the last moment, if this thread is still marked RUNNING,
947 * then put it back on the run queue as it has not been suspended
948 * or stopped or any thing else similar. We never put the idle
949 * threads on the run queue, however.
951 if (td->td_flags & TDF_IDLETD) {
954 CPU_CLR(PCPU_GET(cpuid), &idle_cpus_mask);
957 if (TD_IS_RUNNING(td)) {
958 /* Put us back on the run queue. */
959 sched_add(td, (flags & SW_PREEMPT) ?
960 SRQ_OURSELF|SRQ_YIELDING|SRQ_PREEMPTED :
961 SRQ_OURSELF|SRQ_YIELDING);
966 * The thread we are about to run needs to be counted
967 * as if it had been added to the run queue and selected.
973 KASSERT((newtd->td_inhibitors == 0),
974 ("trying to run inhibited thread"));
975 newtd->td_flags |= TDF_DIDRUN;
976 TD_SET_RUNNING(newtd);
977 if ((newtd->td_flags & TDF_NOLOAD) == 0)
980 newtd = choosethread();
981 MPASS(newtd->td_lock == &sched_lock);
986 if (PMC_PROC_IS_USING_PMCS(td->td_proc))
987 PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_OUT);
990 lock_profile_release_lock(&sched_lock.lock_object);
993 * If DTrace has set the active vtime enum to anything
994 * other than INACTIVE (0), then it should have set the
997 if (dtrace_vtime_active)
998 (*dtrace_vtime_switch_func)(newtd);
1001 cpu_switch(td, newtd, tmtx != NULL ? tmtx : td->td_lock);
1002 lock_profile_obtain_lock_success(&sched_lock.lock_object,
1003 0, 0, __FILE__, __LINE__);
1005 * Where am I? What year is it?
1006 * We are in the same thread that went to sleep above,
1007 * but any amount of time may have passed. All our context
1008 * will still be available as will local variables.
1009 * PCPU values however may have changed as we may have
1010 * changed CPU so don't trust cached values of them.
1011 * New threads will go to fork_exit() instead of here
1012 * so if you change things here you may need to change
1015 * If the thread above was exiting it will never wake
1016 * up again here, so either it has saved everything it
1017 * needed to, or the thread_wait() or wait() will
1021 if (PMC_PROC_IS_USING_PMCS(td->td_proc))
1022 PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_IN);
1027 if (td->td_flags & TDF_IDLETD)
1028 CPU_SET(PCPU_GET(cpuid), &idle_cpus_mask);
1030 sched_lock.mtx_lock = (uintptr_t)td;
1031 td->td_oncpu = PCPU_GET(cpuid);
1032 MPASS(td->td_lock == &sched_lock);
1036 sched_wakeup(struct thread *td)
1038 struct td_sched *ts;
1040 THREAD_LOCK_ASSERT(td, MA_OWNED);
1042 td->td_flags &= ~TDF_CANSWAP;
1043 if (ts->ts_slptime > 1) {
1049 sched_add(td, SRQ_BORING);
1054 forward_wakeup(int cpunum)
1057 cpuset_t dontuse, map, map2;
1061 mtx_assert(&sched_lock, MA_OWNED);
1063 CTR0(KTR_RUNQ, "forward_wakeup()");
1065 if ((!forward_wakeup_enabled) ||
1066 (forward_wakeup_use_mask == 0 && forward_wakeup_use_loop == 0))
1068 if (!smp_started || cold || panicstr)
1071 forward_wakeups_requested++;
1074 * Check the idle mask we received against what we calculated
1075 * before in the old version.
1077 me = PCPU_GET(cpuid);
1079 /* Don't bother if we should be doing it ourself. */
1080 if (CPU_ISSET(me, &idle_cpus_mask) &&
1081 (cpunum == NOCPU || me == cpunum))
1084 CPU_SETOF(me, &dontuse);
1085 CPU_OR(&dontuse, &stopped_cpus);
1086 CPU_OR(&dontuse, &hlt_cpus_mask);
1088 if (forward_wakeup_use_loop) {
1089 STAILQ_FOREACH(pc, &cpuhead, pc_allcpu) {
1091 if (!CPU_ISSET(id, &dontuse) &&
1092 pc->pc_curthread == pc->pc_idlethread) {
1098 if (forward_wakeup_use_mask) {
1099 map = idle_cpus_mask;
1100 CPU_NAND(&map, &dontuse);
1102 /* If they are both on, compare and use loop if different. */
1103 if (forward_wakeup_use_loop) {
1104 if (CPU_CMP(&map, &map2)) {
1105 printf("map != map2, loop method preferred\n");
1113 /* If we only allow a specific CPU, then mask off all the others. */
1114 if (cpunum != NOCPU) {
1115 KASSERT((cpunum <= mp_maxcpus),("forward_wakeup: bad cpunum."));
1116 iscpuset = CPU_ISSET(cpunum, &map);
1120 CPU_SETOF(cpunum, &map);
1122 if (!CPU_EMPTY(&map)) {
1123 forward_wakeups_delivered++;
1124 STAILQ_FOREACH(pc, &cpuhead, pc_allcpu) {
1126 if (!CPU_ISSET(id, &map))
1128 if (cpu_idle_wakeup(pc->pc_cpuid))
1131 if (!CPU_EMPTY(&map))
1132 ipi_selected(map, IPI_AST);
1135 if (cpunum == NOCPU)
1136 printf("forward_wakeup: Idle processor not found\n");
1141 kick_other_cpu(int pri, int cpuid)
1146 pcpu = pcpu_find(cpuid);
1147 if (CPU_ISSET(cpuid, &idle_cpus_mask)) {
1148 forward_wakeups_delivered++;
1149 if (!cpu_idle_wakeup(cpuid))
1150 ipi_cpu(cpuid, IPI_AST);
1154 cpri = pcpu->pc_curthread->td_priority;
1158 #if defined(IPI_PREEMPTION) && defined(PREEMPTION)
1159 #if !defined(FULL_PREEMPTION)
1160 if (pri <= PRI_MAX_ITHD)
1161 #endif /* ! FULL_PREEMPTION */
1163 ipi_cpu(cpuid, IPI_PREEMPT);
1166 #endif /* defined(IPI_PREEMPTION) && defined(PREEMPTION) */
1168 pcpu->pc_curthread->td_flags |= TDF_NEEDRESCHED;
1169 ipi_cpu(cpuid, IPI_AST);
1176 sched_pickcpu(struct thread *td)
1180 mtx_assert(&sched_lock, MA_OWNED);
1182 if (THREAD_CAN_SCHED(td, td->td_lastcpu))
1183 best = td->td_lastcpu;
1187 if (!THREAD_CAN_SCHED(td, cpu))
1192 else if (runq_length[cpu] < runq_length[best])
1195 KASSERT(best != NOCPU, ("no valid CPUs"));
1202 sched_add(struct thread *td, int flags)
1206 struct td_sched *ts;
1212 THREAD_LOCK_ASSERT(td, MA_OWNED);
1213 KASSERT((td->td_inhibitors == 0),
1214 ("sched_add: trying to run inhibited thread"));
1215 KASSERT((TD_CAN_RUN(td) || TD_IS_RUNNING(td)),
1216 ("sched_add: bad thread state"));
1217 KASSERT(td->td_flags & TDF_INMEM,
1218 ("sched_add: thread swapped out"));
1220 KTR_STATE2(KTR_SCHED, "thread", sched_tdname(td), "runq add",
1221 "prio:%d", td->td_priority, KTR_ATTR_LINKED,
1222 sched_tdname(curthread));
1223 KTR_POINT1(KTR_SCHED, "thread", sched_tdname(curthread), "wokeup",
1224 KTR_ATTR_LINKED, sched_tdname(td));
1228 * Now that the thread is moving to the run-queue, set the lock
1229 * to the scheduler's lock.
1231 if (td->td_lock != &sched_lock) {
1232 mtx_lock_spin(&sched_lock);
1233 thread_lock_set(td, &sched_lock);
1238 * If SMP is started and the thread is pinned or otherwise limited to
1239 * a specific set of CPUs, queue the thread to a per-CPU run queue.
1240 * Otherwise, queue the thread to the global run queue.
1242 * If SMP has not yet been started we must use the global run queue
1243 * as per-CPU state may not be initialized yet and we may crash if we
1244 * try to access the per-CPU run queues.
1246 if (smp_started && (td->td_pinned != 0 || td->td_flags & TDF_BOUND ||
1247 ts->ts_flags & TSF_AFFINITY)) {
1248 if (td->td_pinned != 0)
1249 cpu = td->td_lastcpu;
1250 else if (td->td_flags & TDF_BOUND) {
1251 /* Find CPU from bound runq. */
1252 KASSERT(SKE_RUNQ_PCPU(ts),
1253 ("sched_add: bound td_sched not on cpu runq"));
1254 cpu = ts->ts_runq - &runq_pcpu[0];
1256 /* Find a valid CPU for our cpuset */
1257 cpu = sched_pickcpu(td);
1258 ts->ts_runq = &runq_pcpu[cpu];
1261 "sched_add: Put td_sched:%p(td:%p) on cpu%d runq", ts, td,
1265 "sched_add: adding td_sched:%p (td:%p) to gbl runq", ts,
1268 ts->ts_runq = &runq;
1271 cpuid = PCPU_GET(cpuid);
1272 if (single_cpu && cpu != cpuid) {
1273 kick_other_cpu(td->td_priority, cpu);
1276 tidlemsk = idle_cpus_mask;
1277 CPU_NAND(&tidlemsk, &hlt_cpus_mask);
1278 CPU_CLR(cpuid, &tidlemsk);
1280 if (!CPU_ISSET(cpuid, &idle_cpus_mask) &&
1281 ((flags & SRQ_INTR) == 0) &&
1282 !CPU_EMPTY(&tidlemsk))
1283 forwarded = forward_wakeup(cpu);
1287 if ((flags & SRQ_YIELDING) == 0 && maybe_preempt(td))
1294 if ((td->td_flags & TDF_NOLOAD) == 0)
1296 runq_add(ts->ts_runq, td, flags);
1302 struct td_sched *ts;
1305 THREAD_LOCK_ASSERT(td, MA_OWNED);
1306 KASSERT((td->td_inhibitors == 0),
1307 ("sched_add: trying to run inhibited thread"));
1308 KASSERT((TD_CAN_RUN(td) || TD_IS_RUNNING(td)),
1309 ("sched_add: bad thread state"));
1310 KASSERT(td->td_flags & TDF_INMEM,
1311 ("sched_add: thread swapped out"));
1312 KTR_STATE2(KTR_SCHED, "thread", sched_tdname(td), "runq add",
1313 "prio:%d", td->td_priority, KTR_ATTR_LINKED,
1314 sched_tdname(curthread));
1315 KTR_POINT1(KTR_SCHED, "thread", sched_tdname(curthread), "wokeup",
1316 KTR_ATTR_LINKED, sched_tdname(td));
1319 * Now that the thread is moving to the run-queue, set the lock
1320 * to the scheduler's lock.
1322 if (td->td_lock != &sched_lock) {
1323 mtx_lock_spin(&sched_lock);
1324 thread_lock_set(td, &sched_lock);
1327 CTR2(KTR_RUNQ, "sched_add: adding td_sched:%p (td:%p) to runq", ts, td);
1328 ts->ts_runq = &runq;
1331 * If we are yielding (on the way out anyhow) or the thread
1332 * being saved is US, then don't try be smart about preemption
1333 * or kicking off another CPU as it won't help and may hinder.
1334 * In the YIEDLING case, we are about to run whoever is being
1335 * put in the queue anyhow, and in the OURSELF case, we are
1336 * puting ourself on the run queue which also only happens
1337 * when we are about to yield.
1339 if ((flags & SRQ_YIELDING) == 0) {
1340 if (maybe_preempt(td))
1343 if ((td->td_flags & TDF_NOLOAD) == 0)
1345 runq_add(ts->ts_runq, td, flags);
1351 sched_rem(struct thread *td)
1353 struct td_sched *ts;
1356 KASSERT(td->td_flags & TDF_INMEM,
1357 ("sched_rem: thread swapped out"));
1358 KASSERT(TD_ON_RUNQ(td),
1359 ("sched_rem: thread not on run queue"));
1360 mtx_assert(&sched_lock, MA_OWNED);
1361 KTR_STATE2(KTR_SCHED, "thread", sched_tdname(td), "runq rem",
1362 "prio:%d", td->td_priority, KTR_ATTR_LINKED,
1363 sched_tdname(curthread));
1365 if ((td->td_flags & TDF_NOLOAD) == 0)
1368 if (ts->ts_runq != &runq)
1369 runq_length[ts->ts_runq - runq_pcpu]--;
1371 runq_remove(ts->ts_runq, td);
1376 * Select threads to run. Note that running threads still consume a
1385 mtx_assert(&sched_lock, MA_OWNED);
1387 struct thread *tdcpu;
1390 td = runq_choose_fuzz(&runq, runq_fuzz);
1391 tdcpu = runq_choose(&runq_pcpu[PCPU_GET(cpuid)]);
1395 tdcpu->td_priority < td->td_priority)) {
1396 CTR2(KTR_RUNQ, "choosing td %p from pcpu runq %d", tdcpu,
1399 rq = &runq_pcpu[PCPU_GET(cpuid)];
1401 CTR1(KTR_RUNQ, "choosing td_sched %p from main runq", td);
1406 td = runq_choose(&runq);
1412 runq_length[PCPU_GET(cpuid)]--;
1414 runq_remove(rq, td);
1415 td->td_flags |= TDF_DIDRUN;
1417 KASSERT(td->td_flags & TDF_INMEM,
1418 ("sched_choose: thread swapped out"));
1421 return (PCPU_GET(idlethread));
1425 sched_preempt(struct thread *td)
1428 if (td->td_critnest > 1)
1429 td->td_owepreempt = 1;
1431 mi_switch(SW_INVOL | SW_PREEMPT | SWT_PREEMPT, NULL);
1436 sched_userret(struct thread *td)
1439 * XXX we cheat slightly on the locking here to avoid locking in
1440 * the usual case. Setting td_priority here is essentially an
1441 * incomplete workaround for not setting it properly elsewhere.
1442 * Now that some interrupt handlers are threads, not setting it
1443 * properly elsewhere can clobber it in the window between setting
1444 * it here and returning to user mode, so don't waste time setting
1445 * it perfectly here.
1447 KASSERT((td->td_flags & TDF_BORROWING) == 0,
1448 ("thread with borrowed priority returning to userland"));
1449 if (td->td_priority != td->td_user_pri) {
1451 td->td_priority = td->td_user_pri;
1452 td->td_base_pri = td->td_user_pri;
1458 sched_bind(struct thread *td, int cpu)
1460 struct td_sched *ts;
1462 THREAD_LOCK_ASSERT(td, MA_OWNED|MA_NOTRECURSED);
1463 KASSERT(td == curthread, ("sched_bind: can only bind curthread"));
1467 td->td_flags |= TDF_BOUND;
1469 ts->ts_runq = &runq_pcpu[cpu];
1470 if (PCPU_GET(cpuid) == cpu)
1473 mi_switch(SW_VOL, NULL);
1478 sched_unbind(struct thread* td)
1480 THREAD_LOCK_ASSERT(td, MA_OWNED);
1481 KASSERT(td == curthread, ("sched_unbind: can only bind curthread"));
1482 td->td_flags &= ~TDF_BOUND;
1486 sched_is_bound(struct thread *td)
1488 THREAD_LOCK_ASSERT(td, MA_OWNED);
1489 return (td->td_flags & TDF_BOUND);
1493 sched_relinquish(struct thread *td)
1496 mi_switch(SW_VOL | SWT_RELINQUISH, NULL);
1503 return (sched_tdcnt);
1507 sched_sizeof_proc(void)
1509 return (sizeof(struct proc));
1513 sched_sizeof_thread(void)
1515 return (sizeof(struct thread) + sizeof(struct td_sched));
1519 sched_pctcpu(struct thread *td)
1521 struct td_sched *ts;
1523 THREAD_LOCK_ASSERT(td, MA_OWNED);
1525 return (ts->ts_pctcpu);
1534 * The actual idle process.
1537 sched_idletd(void *dummy)
1539 struct pcpuidlestat *stat;
1541 stat = DPCPU_PTR(idlestat);
1543 mtx_assert(&Giant, MA_NOTOWNED);
1545 while (sched_runnable() == 0) {
1546 cpu_idle(stat->idlecalls + stat->oldidlecalls > 64);
1550 mtx_lock_spin(&sched_lock);
1551 mi_switch(SW_VOL | SWT_IDLE, NULL);
1552 mtx_unlock_spin(&sched_lock);
1557 * A CPU is entering for the first time or a thread is exiting.
1560 sched_throw(struct thread *td)
1563 * Correct spinlock nesting. The idle thread context that we are
1564 * borrowing was created so that it would start out with a single
1565 * spin lock (sched_lock) held in fork_trampoline(). Since we've
1566 * explicitly acquired locks in this function, the nesting count
1567 * is now 2 rather than 1. Since we are nested, calling
1568 * spinlock_exit() will simply adjust the counts without allowing
1569 * spin lock using code to interrupt us.
1572 mtx_lock_spin(&sched_lock);
1575 lock_profile_release_lock(&sched_lock.lock_object);
1576 MPASS(td->td_lock == &sched_lock);
1578 mtx_assert(&sched_lock, MA_OWNED);
1579 KASSERT(curthread->td_md.md_spinlock_count == 1, ("invalid count"));
1580 PCPU_SET(switchtime, cpu_ticks());
1581 PCPU_SET(switchticks, ticks);
1582 cpu_throw(td, choosethread()); /* doesn't return */
1586 sched_fork_exit(struct thread *td)
1590 * Finish setting up thread glue so that it begins execution in a
1591 * non-nested critical section with sched_lock held but not recursed.
1593 td->td_oncpu = PCPU_GET(cpuid);
1594 sched_lock.mtx_lock = (uintptr_t)td;
1595 lock_profile_obtain_lock_success(&sched_lock.lock_object,
1596 0, 0, __FILE__, __LINE__);
1597 THREAD_LOCK_ASSERT(td, MA_OWNED | MA_NOTRECURSED);
1601 sched_tdname(struct thread *td)
1604 struct td_sched *ts;
1607 if (ts->ts_name[0] == '\0')
1608 snprintf(ts->ts_name, sizeof(ts->ts_name),
1609 "%s tid %d", td->td_name, td->td_tid);
1610 return (ts->ts_name);
1612 return (td->td_name);
1617 sched_affinity(struct thread *td)
1620 struct td_sched *ts;
1623 THREAD_LOCK_ASSERT(td, MA_OWNED);
1626 * Set the TSF_AFFINITY flag if there is at least one CPU this
1627 * thread can't run on.
1630 ts->ts_flags &= ~TSF_AFFINITY;
1632 if (!THREAD_CAN_SCHED(td, cpu)) {
1633 ts->ts_flags |= TSF_AFFINITY;
1639 * If this thread can run on all CPUs, nothing else to do.
1641 if (!(ts->ts_flags & TSF_AFFINITY))
1644 /* Pinned threads and bound threads should be left alone. */
1645 if (td->td_pinned != 0 || td->td_flags & TDF_BOUND)
1648 switch (td->td_state) {
1651 * If we are on a per-CPU runqueue that is in the set,
1652 * then nothing needs to be done.
1654 if (ts->ts_runq != &runq &&
1655 THREAD_CAN_SCHED(td, ts->ts_runq - runq_pcpu))
1658 /* Put this thread on a valid per-CPU runqueue. */
1660 sched_add(td, SRQ_BORING);
1664 * See if our current CPU is in the set. If not, force a
1667 if (THREAD_CAN_SCHED(td, td->td_oncpu))
1670 td->td_flags |= TDF_NEEDRESCHED;
1671 if (td != curthread)
1672 ipi_cpu(cpu, IPI_AST);