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"
41 #include <sys/param.h>
42 #include <sys/systm.h>
43 #include <sys/cpuset.h>
44 #include <sys/kernel.h>
47 #include <sys/kthread.h>
48 #include <sys/mutex.h>
50 #include <sys/resourcevar.h>
51 #include <sys/sched.h>
53 #include <sys/sysctl.h>
55 #include <sys/turnstile.h>
57 #include <machine/pcb.h>
58 #include <machine/smp.h>
61 #include <sys/pmckern.h>
65 * INVERSE_ESTCPU_WEIGHT is only suitable for statclock() frequencies in
66 * the range 100-256 Hz (approximately).
68 #define ESTCPULIM(e) \
69 min((e), INVERSE_ESTCPU_WEIGHT * (NICE_WEIGHT * (PRIO_MAX - PRIO_MIN) - \
70 RQ_PPQ) + INVERSE_ESTCPU_WEIGHT - 1)
72 #define INVERSE_ESTCPU_WEIGHT (8 * smp_cpus)
74 #define INVERSE_ESTCPU_WEIGHT 8 /* 1 / (priorities per estcpu level). */
76 #define NICE_WEIGHT 1 /* Priorities per nice level. */
79 * The schedulable entity that runs a context.
80 * This is an extension to the thread structure and is tailored to
81 * the requirements of this scheduler
84 TAILQ_ENTRY(td_sched) ts_procq; /* (j/z) Run queue. */
85 struct thread *ts_thread; /* (*) Active associated thread. */
86 fixpt_t ts_pctcpu; /* (j) %cpu during p_swtime. */
87 u_char ts_rqindex; /* (j) Run queue index. */
88 int ts_cpticks; /* (j) Ticks of cpu time. */
89 int ts_slptime; /* (j) Seconds !RUNNING. */
90 struct runq *ts_runq; /* runq the thread is currently on */
93 /* flags kept in td_flags */
94 #define TDF_DIDRUN TDF_SCHED0 /* thread actually ran. */
95 #define TDF_EXIT TDF_SCHED1 /* thread is being killed. */
96 #define TDF_BOUND TDF_SCHED2
98 #define ts_flags ts_thread->td_flags
99 #define TSF_DIDRUN TDF_DIDRUN /* thread actually ran. */
100 #define TSF_EXIT TDF_EXIT /* thread is being killed. */
101 #define TSF_BOUND TDF_BOUND /* stuck to one CPU */
103 #define SKE_RUNQ_PCPU(ts) \
104 ((ts)->ts_runq != 0 && (ts)->ts_runq != &runq)
106 static struct td_sched td_sched0;
107 struct mtx sched_lock;
109 static int sched_tdcnt; /* Total runnable threads in the system. */
110 static int sched_quantum; /* Roundrobin scheduling quantum in ticks. */
111 #define SCHED_QUANTUM (hz / 10) /* Default sched quantum */
113 static void setup_runqs(void);
114 static void schedcpu(void);
115 static void schedcpu_thread(void);
116 static void sched_priority(struct thread *td, u_char prio);
117 static void sched_setup(void *dummy);
118 static void maybe_resched(struct thread *td);
119 static void updatepri(struct thread *td);
120 static void resetpriority(struct thread *td);
121 static void resetpriority_thread(struct thread *td);
123 static int forward_wakeup(int cpunum);
126 static struct kproc_desc sched_kp = {
131 SYSINIT(schedcpu, SI_SUB_RUN_SCHEDULER, SI_ORDER_FIRST, kproc_start,
133 SYSINIT(sched_setup, SI_SUB_RUN_QUEUE, SI_ORDER_FIRST, sched_setup, NULL);
138 static struct runq runq;
144 static struct runq runq_pcpu[MAXCPU];
153 for (i = 0; i < MAXCPU; ++i)
154 runq_init(&runq_pcpu[i]);
161 sysctl_kern_quantum(SYSCTL_HANDLER_ARGS)
165 new_val = sched_quantum * tick;
166 error = sysctl_handle_int(oidp, &new_val, 0, req);
167 if (error != 0 || req->newptr == NULL)
171 sched_quantum = new_val / tick;
172 hogticks = 2 * sched_quantum;
176 SYSCTL_NODE(_kern, OID_AUTO, sched, CTLFLAG_RD, 0, "Scheduler");
178 SYSCTL_STRING(_kern_sched, OID_AUTO, name, CTLFLAG_RD, "4BSD", 0,
181 SYSCTL_PROC(_kern_sched, OID_AUTO, quantum, CTLTYPE_INT | CTLFLAG_RW,
182 0, sizeof sched_quantum, sysctl_kern_quantum, "I",
183 "Roundrobin scheduling quantum in microseconds");
186 /* Enable forwarding of wakeups to all other cpus */
187 SYSCTL_NODE(_kern_sched, OID_AUTO, ipiwakeup, CTLFLAG_RD, NULL, "Kernel SMP");
189 static int forward_wakeup_enabled = 1;
190 SYSCTL_INT(_kern_sched_ipiwakeup, OID_AUTO, enabled, CTLFLAG_RW,
191 &forward_wakeup_enabled, 0,
192 "Forwarding of wakeup to idle CPUs");
194 static int forward_wakeups_requested = 0;
195 SYSCTL_INT(_kern_sched_ipiwakeup, OID_AUTO, requested, CTLFLAG_RD,
196 &forward_wakeups_requested, 0,
197 "Requests for Forwarding of wakeup to idle CPUs");
199 static int forward_wakeups_delivered = 0;
200 SYSCTL_INT(_kern_sched_ipiwakeup, OID_AUTO, delivered, CTLFLAG_RD,
201 &forward_wakeups_delivered, 0,
202 "Completed Forwarding of wakeup to idle CPUs");
204 static int forward_wakeup_use_mask = 1;
205 SYSCTL_INT(_kern_sched_ipiwakeup, OID_AUTO, usemask, CTLFLAG_RW,
206 &forward_wakeup_use_mask, 0,
207 "Use the mask of idle cpus");
209 static int forward_wakeup_use_loop = 0;
210 SYSCTL_INT(_kern_sched_ipiwakeup, OID_AUTO, useloop, CTLFLAG_RW,
211 &forward_wakeup_use_loop, 0,
212 "Use a loop to find idle cpus");
214 static int forward_wakeup_use_single = 0;
215 SYSCTL_INT(_kern_sched_ipiwakeup, OID_AUTO, onecpu, CTLFLAG_RW,
216 &forward_wakeup_use_single, 0,
217 "Only signal one idle cpu");
219 static int forward_wakeup_use_htt = 0;
220 SYSCTL_INT(_kern_sched_ipiwakeup, OID_AUTO, htt2, CTLFLAG_RW,
221 &forward_wakeup_use_htt, 0,
226 static int sched_followon = 0;
227 SYSCTL_INT(_kern_sched, OID_AUTO, followon, CTLFLAG_RW,
229 "allow threads to share a quantum");
236 CTR1(KTR_SCHED, "global load: %d", sched_tdcnt);
243 CTR1(KTR_SCHED, "global load: %d", sched_tdcnt);
246 * Arrange to reschedule if necessary, taking the priorities and
247 * schedulers into account.
250 maybe_resched(struct thread *td)
253 THREAD_LOCK_ASSERT(td, MA_OWNED);
254 if (td->td_priority < curthread->td_priority)
255 curthread->td_flags |= TDF_NEEDRESCHED;
259 * Constants for digital decay and forget:
260 * 90% of (td_estcpu) usage in 5 * loadav time
261 * 95% of (ts_pctcpu) usage in 60 seconds (load insensitive)
262 * Note that, as ps(1) mentions, this can let percentages
263 * total over 100% (I've seen 137.9% for 3 processes).
265 * Note that schedclock() updates td_estcpu and p_cpticks asynchronously.
267 * We wish to decay away 90% of td_estcpu in (5 * loadavg) seconds.
268 * That is, the system wants to compute a value of decay such
269 * that the following for loop:
270 * for (i = 0; i < (5 * loadavg); i++)
271 * td_estcpu *= decay;
274 * for all values of loadavg:
276 * Mathematically this loop can be expressed by saying:
277 * decay ** (5 * loadavg) ~= .1
279 * The system computes decay as:
280 * decay = (2 * loadavg) / (2 * loadavg + 1)
282 * We wish to prove that the system's computation of decay
283 * will always fulfill the equation:
284 * decay ** (5 * loadavg) ~= .1
286 * If we compute b as:
289 * decay = b / (b + 1)
291 * We now need to prove two things:
292 * 1) Given factor ** (5 * loadavg) ~= .1, prove factor == b/(b+1)
293 * 2) Given b/(b+1) ** power ~= .1, prove power == (5 * loadavg)
296 * For x close to zero, exp(x) =~ 1 + x, since
297 * exp(x) = 0! + x**1/1! + x**2/2! + ... .
298 * therefore exp(-1/b) =~ 1 - (1/b) = (b-1)/b.
299 * For x close to zero, ln(1+x) =~ x, since
300 * ln(1+x) = x - x**2/2 + x**3/3 - ... -1 < x < 1
301 * therefore ln(b/(b+1)) = ln(1 - 1/(b+1)) =~ -1/(b+1).
305 * Solve (factor)**(power) =~ .1 given power (5*loadav):
306 * solving for factor,
307 * ln(factor) =~ (-2.30/5*loadav), or
308 * factor =~ exp(-1/((5/2.30)*loadav)) =~ exp(-1/(2*loadav)) =
309 * exp(-1/b) =~ (b-1)/b =~ b/(b+1). QED
312 * Solve (factor)**(power) =~ .1 given factor == (b/(b+1)):
314 * power*ln(b/(b+1)) =~ -2.30, or
315 * power =~ 2.3 * (b + 1) = 4.6*loadav + 2.3 =~ 5*loadav. QED
317 * Actual power values for the implemented algorithm are as follows:
319 * power: 5.68 10.32 14.94 19.55
322 /* calculations for digital decay to forget 90% of usage in 5*loadav sec */
323 #define loadfactor(loadav) (2 * (loadav))
324 #define decay_cpu(loadfac, cpu) (((loadfac) * (cpu)) / ((loadfac) + FSCALE))
326 /* decay 95% of `ts_pctcpu' in 60 seconds; see CCPU_SHIFT before changing */
327 static fixpt_t ccpu = 0.95122942450071400909 * FSCALE; /* exp(-1/20) */
328 SYSCTL_INT(_kern, OID_AUTO, ccpu, CTLFLAG_RD, &ccpu, 0, "");
331 * If `ccpu' is not equal to `exp(-1/20)' and you still want to use the
332 * faster/more-accurate formula, you'll have to estimate CCPU_SHIFT below
333 * and possibly adjust FSHIFT in "param.h" so that (FSHIFT >= CCPU_SHIFT).
335 * To estimate CCPU_SHIFT for exp(-1/20), the following formula was used:
336 * 1 - exp(-1/20) ~= 0.0487 ~= 0.0488 == 1 (fixed pt, *11* bits).
338 * If you don't want to bother with the faster/more-accurate formula, you
339 * can set CCPU_SHIFT to (FSHIFT + 1) which will use a slower/less-accurate
340 * (more general) method of calculating the %age of CPU used by a process.
342 #define CCPU_SHIFT 11
345 * Recompute process priorities, every hz ticks.
346 * MP-safe, called without the Giant mutex.
352 register fixpt_t loadfac = loadfactor(averunnable.ldavg[0]);
356 int awake, realstathz;
358 realstathz = stathz ? stathz : hz;
359 sx_slock(&allproc_lock);
360 FOREACH_PROC_IN_SYSTEM(p) {
362 FOREACH_THREAD_IN_PROC(p, td) {
367 * Increment sleep time (if sleeping). We
368 * ignore overflow, as above.
371 * The td_sched slptimes are not touched in wakeup
372 * because the thread may not HAVE everything in
373 * memory? XXX I think this is out of date.
375 if (TD_ON_RUNQ(td)) {
377 ts->ts_flags &= ~TSF_DIDRUN;
378 } else if (TD_IS_RUNNING(td)) {
380 /* Do not clear TSF_DIDRUN */
381 } else if (ts->ts_flags & TSF_DIDRUN) {
383 ts->ts_flags &= ~TSF_DIDRUN;
387 * ts_pctcpu is only for ps and ttyinfo().
389 ts->ts_pctcpu = (ts->ts_pctcpu * ccpu) >> FSHIFT;
391 * If the td_sched has been idle the entire second,
392 * stop recalculating its priority until
395 if (ts->ts_cpticks != 0) {
396 #if (FSHIFT >= CCPU_SHIFT)
397 ts->ts_pctcpu += (realstathz == 100)
398 ? ((fixpt_t) ts->ts_cpticks) <<
399 (FSHIFT - CCPU_SHIFT) :
400 100 * (((fixpt_t) ts->ts_cpticks)
401 << (FSHIFT - CCPU_SHIFT)) / realstathz;
403 ts->ts_pctcpu += ((FSCALE - ccpu) *
405 FSCALE / realstathz)) >> FSHIFT;
410 * If there are ANY running threads in this process,
411 * then don't count it as sleeping.
416 if (ts->ts_slptime > 1) {
418 * In an ideal world, this should not
419 * happen, because whoever woke us
420 * up from the long sleep should have
421 * unwound the slptime and reset our
422 * priority before we run at the stale
423 * priority. Should KASSERT at some
424 * point when all the cases are fixed.
431 if (ts->ts_slptime > 1) {
435 td->td_estcpu = decay_cpu(loadfac, td->td_estcpu);
437 resetpriority_thread(td);
439 } /* end of thread loop */
441 } /* end of process loop */
442 sx_sunlock(&allproc_lock);
446 * Main loop for a kthread that executes schedcpu once a second.
449 schedcpu_thread(void)
459 * Recalculate the priority of a process after it has slept for a while.
460 * For all load averages >= 1 and max td_estcpu of 255, sleeping for at
461 * least six times the loadfactor will decay td_estcpu to zero.
464 updatepri(struct thread *td)
471 loadfac = loadfactor(averunnable.ldavg[0]);
472 if (ts->ts_slptime > 5 * loadfac)
475 newcpu = td->td_estcpu;
476 ts->ts_slptime--; /* was incremented in schedcpu() */
477 while (newcpu && --ts->ts_slptime)
478 newcpu = decay_cpu(loadfac, newcpu);
479 td->td_estcpu = newcpu;
484 * Compute the priority of a process when running in user mode.
485 * Arrange to reschedule if the resulting priority is better
486 * than that of the current process.
489 resetpriority(struct thread *td)
491 register unsigned int newpriority;
493 if (td->td_pri_class == PRI_TIMESHARE) {
494 newpriority = PUSER + td->td_estcpu / INVERSE_ESTCPU_WEIGHT +
495 NICE_WEIGHT * (td->td_proc->p_nice - PRIO_MIN);
496 newpriority = min(max(newpriority, PRI_MIN_TIMESHARE),
498 sched_user_prio(td, newpriority);
503 * Update the thread's priority when the associated process's user
507 resetpriority_thread(struct thread *td)
510 /* Only change threads with a time sharing user priority. */
511 if (td->td_priority < PRI_MIN_TIMESHARE ||
512 td->td_priority > PRI_MAX_TIMESHARE)
515 /* XXX the whole needresched thing is broken, but not silly. */
518 sched_prio(td, td->td_user_pri);
523 sched_setup(void *dummy)
527 if (sched_quantum == 0)
528 sched_quantum = SCHED_QUANTUM;
529 hogticks = 2 * sched_quantum;
531 /* Account for thread0. */
535 /* External interfaces start here */
537 * Very early in the boot some setup of scheduler-specific
538 * parts of proc0 and of some scheduler resources needs to be done.
546 * Set up the scheduler specific parts of proc0.
548 proc0.p_sched = NULL; /* XXX */
549 thread0.td_sched = &td_sched0;
550 thread0.td_lock = &sched_lock;
551 td_sched0.ts_thread = &thread0;
552 mtx_init(&sched_lock, "sched lock", NULL, MTX_SPIN | MTX_RECURSE);
559 return runq_check(&runq) + runq_check(&runq_pcpu[PCPU_GET(cpuid)]);
561 return runq_check(&runq);
566 sched_rr_interval(void)
568 if (sched_quantum == 0)
569 sched_quantum = SCHED_QUANTUM;
570 return (sched_quantum);
574 * We adjust the priority of the current process. The priority of
575 * a process gets worse as it accumulates CPU time. The cpu usage
576 * estimator (td_estcpu) is increased here. resetpriority() will
577 * compute a different priority each time td_estcpu increases by
578 * INVERSE_ESTCPU_WEIGHT
579 * (until MAXPRI is reached). The cpu usage estimator ramps up
580 * quite quickly when the process is running (linearly), and decays
581 * away exponentially, at a rate which is proportionally slower when
582 * the system is busy. The basic principle is that the system will
583 * 90% forget that the process used a lot of CPU time in 5 * loadav
584 * seconds. This causes the system to favor processes which haven't
585 * run much recently, and to round-robin among other processes.
588 sched_clock(struct thread *td)
592 THREAD_LOCK_ASSERT(td, MA_OWNED);
596 td->td_estcpu = ESTCPULIM(td->td_estcpu + 1);
597 if ((td->td_estcpu % INVERSE_ESTCPU_WEIGHT) == 0) {
599 resetpriority_thread(td);
603 * Force a context switch if the current thread has used up a full
604 * quantum (default quantum is 100ms).
606 if (!TD_IS_IDLETHREAD(td) &&
607 ticks - PCPU_GET(switchticks) >= sched_quantum)
608 td->td_flags |= TDF_NEEDRESCHED;
612 * charge childs scheduling cpu usage to parent.
615 sched_exit(struct proc *p, struct thread *td)
618 CTR3(KTR_SCHED, "sched_exit: %p(%s) prio %d",
619 td, td->td_name, td->td_priority);
620 PROC_LOCK_ASSERT(p, MA_OWNED);
621 sched_exit_thread(FIRST_THREAD_IN_PROC(p), td);
625 sched_exit_thread(struct thread *td, struct thread *child)
628 CTR3(KTR_SCHED, "sched_exit_thread: %p(%s) prio %d",
629 child, child->td_name, child->td_priority);
631 td->td_estcpu = ESTCPULIM(td->td_estcpu + child->td_estcpu);
633 mtx_lock_spin(&sched_lock);
634 if ((child->td_proc->p_flag & P_NOLOAD) == 0)
636 mtx_unlock_spin(&sched_lock);
640 sched_fork(struct thread *td, struct thread *childtd)
642 sched_fork_thread(td, childtd);
646 sched_fork_thread(struct thread *td, struct thread *childtd)
648 childtd->td_estcpu = td->td_estcpu;
649 childtd->td_lock = &sched_lock;
650 childtd->td_cpuset = cpuset_ref(td->td_cpuset);
651 sched_newthread(childtd);
655 sched_nice(struct proc *p, int nice)
659 PROC_LOCK_ASSERT(p, MA_OWNED);
661 FOREACH_THREAD_IN_PROC(p, td) {
664 resetpriority_thread(td);
670 sched_class(struct thread *td, int class)
672 THREAD_LOCK_ASSERT(td, MA_OWNED);
673 td->td_pri_class = class;
677 * Adjust the priority of a thread.
680 sched_priority(struct thread *td, u_char prio)
682 CTR6(KTR_SCHED, "sched_prio: %p(%s) prio %d newprio %d by %p(%s)",
683 td, td->td_name, td->td_priority, prio, curthread,
686 THREAD_LOCK_ASSERT(td, MA_OWNED);
687 if (td->td_priority == prio)
689 td->td_priority = prio;
690 if (TD_ON_RUNQ(td) &&
691 td->td_sched->ts_rqindex != (prio / RQ_PPQ)) {
693 sched_add(td, SRQ_BORING);
698 * Update a thread's priority when it is lent another thread's
702 sched_lend_prio(struct thread *td, u_char prio)
705 td->td_flags |= TDF_BORROWING;
706 sched_priority(td, prio);
710 * Restore a thread's priority when priority propagation is
711 * over. The prio argument is the minimum priority the thread
712 * needs to have to satisfy other possible priority lending
713 * requests. If the thread's regulary priority is less
714 * important than prio the thread will keep a priority boost
718 sched_unlend_prio(struct thread *td, u_char prio)
722 if (td->td_base_pri >= PRI_MIN_TIMESHARE &&
723 td->td_base_pri <= PRI_MAX_TIMESHARE)
724 base_pri = td->td_user_pri;
726 base_pri = td->td_base_pri;
727 if (prio >= base_pri) {
728 td->td_flags &= ~TDF_BORROWING;
729 sched_prio(td, base_pri);
731 sched_lend_prio(td, prio);
735 sched_prio(struct thread *td, u_char prio)
739 /* First, update the base priority. */
740 td->td_base_pri = prio;
743 * If the thread is borrowing another thread's priority, don't ever
744 * lower the priority.
746 if (td->td_flags & TDF_BORROWING && td->td_priority < prio)
749 /* Change the real priority. */
750 oldprio = td->td_priority;
751 sched_priority(td, prio);
754 * If the thread is on a turnstile, then let the turnstile update
757 if (TD_ON_LOCK(td) && oldprio != prio)
758 turnstile_adjust(td, oldprio);
762 sched_user_prio(struct thread *td, u_char prio)
766 THREAD_LOCK_ASSERT(td, MA_OWNED);
767 td->td_base_user_pri = prio;
768 if (td->td_flags & TDF_UBORROWING && td->td_user_pri <= prio)
770 oldprio = td->td_user_pri;
771 td->td_user_pri = prio;
775 sched_lend_user_prio(struct thread *td, u_char prio)
779 THREAD_LOCK_ASSERT(td, MA_OWNED);
780 td->td_flags |= TDF_UBORROWING;
781 oldprio = td->td_user_pri;
782 td->td_user_pri = prio;
786 sched_unlend_user_prio(struct thread *td, u_char prio)
790 THREAD_LOCK_ASSERT(td, MA_OWNED);
791 base_pri = td->td_base_user_pri;
792 if (prio >= base_pri) {
793 td->td_flags &= ~TDF_UBORROWING;
794 sched_user_prio(td, base_pri);
796 sched_lend_user_prio(td, prio);
801 sched_sleep(struct thread *td, int pri)
804 THREAD_LOCK_ASSERT(td, MA_OWNED);
805 td->td_slptick = ticks;
806 td->td_sched->ts_slptime = 0;
809 if (TD_IS_SUSPENDED(td) || pri <= PSOCK)
810 td->td_flags |= TDF_CANSWAP;
814 sched_switch(struct thread *td, struct thread *newtd, int flags)
822 THREAD_LOCK_ASSERT(td, MA_OWNED);
824 * Switch to the sched lock to fix things up and pick
827 if (td->td_lock != &sched_lock) {
828 mtx_lock_spin(&sched_lock);
832 if ((p->p_flag & P_NOLOAD) == 0)
836 newtd->td_flags |= (td->td_flags & TDF_NEEDRESCHED);
838 td->td_lastcpu = td->td_oncpu;
839 td->td_flags &= ~TDF_NEEDRESCHED;
840 td->td_owepreempt = 0;
841 td->td_oncpu = NOCPU;
843 * At the last moment, if this thread is still marked RUNNING,
844 * then put it back on the run queue as it has not been suspended
845 * or stopped or any thing else similar. We never put the idle
846 * threads on the run queue, however.
848 if (td->td_flags & TDF_IDLETD) {
851 idle_cpus_mask &= ~PCPU_GET(cpumask);
854 if (TD_IS_RUNNING(td)) {
855 /* Put us back on the run queue. */
856 sched_add(td, (flags & SW_PREEMPT) ?
857 SRQ_OURSELF|SRQ_YIELDING|SRQ_PREEMPTED :
858 SRQ_OURSELF|SRQ_YIELDING);
863 * The thread we are about to run needs to be counted
864 * as if it had been added to the run queue and selected.
870 KASSERT((newtd->td_inhibitors == 0),
871 ("trying to run inhibited thread"));
872 newtd->td_sched->ts_flags |= TSF_DIDRUN;
873 TD_SET_RUNNING(newtd);
874 if ((newtd->td_proc->p_flag & P_NOLOAD) == 0)
877 newtd = choosethread();
879 MPASS(newtd->td_lock == &sched_lock);
883 if (PMC_PROC_IS_USING_PMCS(td->td_proc))
884 PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_OUT);
887 lock_profile_release_lock(&sched_lock.lock_object);
888 cpu_switch(td, newtd, td->td_lock);
889 lock_profile_obtain_lock_success(&sched_lock.lock_object,
890 0, 0, __FILE__, __LINE__);
892 * Where am I? What year is it?
893 * We are in the same thread that went to sleep above,
894 * but any amount of time may have passed. All out context
895 * will still be available as will local variables.
896 * PCPU values however may have changed as we may have
897 * changed CPU so don't trust cached values of them.
898 * New threads will go to fork_exit() instead of here
899 * so if you change things here you may need to change
901 * If the thread above was exiting it will never wake
902 * up again here, so either it has saved everything it
903 * needed to, or the thread_wait() or wait() will
907 if (PMC_PROC_IS_USING_PMCS(td->td_proc))
908 PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_IN);
913 if (td->td_flags & TDF_IDLETD)
914 idle_cpus_mask |= PCPU_GET(cpumask);
916 sched_lock.mtx_lock = (uintptr_t)td;
917 td->td_oncpu = PCPU_GET(cpuid);
918 MPASS(td->td_lock == &sched_lock);
922 sched_wakeup(struct thread *td)
926 THREAD_LOCK_ASSERT(td, MA_OWNED);
928 td->td_flags &= ~TDF_CANSWAP;
929 if (ts->ts_slptime > 1) {
933 td->td_slptick = ticks;
935 sched_add(td, SRQ_BORING);
939 /* enable HTT_2 if you have a 2-way HTT cpu.*/
941 forward_wakeup(int cpunum)
943 cpumask_t map, me, dontuse;
948 mtx_assert(&sched_lock, MA_OWNED);
950 CTR0(KTR_RUNQ, "forward_wakeup()");
952 if ((!forward_wakeup_enabled) ||
953 (forward_wakeup_use_mask == 0 && forward_wakeup_use_loop == 0))
955 if (!smp_started || cold || panicstr)
958 forward_wakeups_requested++;
961 * check the idle mask we received against what we calculated before
962 * in the old version.
964 me = PCPU_GET(cpumask);
966 * don't bother if we should be doing it ourself..
968 if ((me & idle_cpus_mask) && (cpunum == NOCPU || me == (1 << cpunum)))
971 dontuse = me | stopped_cpus | hlt_cpus_mask;
973 if (forward_wakeup_use_loop) {
974 SLIST_FOREACH(pc, &cpuhead, pc_allcpu) {
976 if ( (id & dontuse) == 0 &&
977 pc->pc_curthread == pc->pc_idlethread) {
983 if (forward_wakeup_use_mask) {
985 map = idle_cpus_mask & ~dontuse;
987 /* If they are both on, compare and use loop if different */
988 if (forward_wakeup_use_loop) {
990 printf("map (%02X) != map3 (%02X)\n",
998 /* If we only allow a specific CPU, then mask off all the others */
999 if (cpunum != NOCPU) {
1000 KASSERT((cpunum <= mp_maxcpus),("forward_wakeup: bad cpunum."));
1001 map &= (1 << cpunum);
1003 /* Try choose an idle die. */
1004 if (forward_wakeup_use_htt) {
1005 map2 = (map & (map >> 1)) & 0x5555;
1011 /* set only one bit */
1012 if (forward_wakeup_use_single) {
1013 map = map & ((~map) + 1);
1017 forward_wakeups_delivered++;
1018 ipi_selected(map, IPI_AST);
1021 if (cpunum == NOCPU)
1022 printf("forward_wakeup: Idle processor not found\n");
1028 static void kick_other_cpu(int pri,int cpuid);
1031 kick_other_cpu(int pri,int cpuid)
1033 struct pcpu * pcpu = pcpu_find(cpuid);
1034 int cpri = pcpu->pc_curthread->td_priority;
1036 if (idle_cpus_mask & pcpu->pc_cpumask) {
1037 forward_wakeups_delivered++;
1038 ipi_selected(pcpu->pc_cpumask, IPI_AST);
1045 #if defined(IPI_PREEMPTION) && defined(PREEMPTION)
1046 #if !defined(FULL_PREEMPTION)
1047 if (pri <= PRI_MAX_ITHD)
1048 #endif /* ! FULL_PREEMPTION */
1050 ipi_selected(pcpu->pc_cpumask, IPI_PREEMPT);
1053 #endif /* defined(IPI_PREEMPTION) && defined(PREEMPTION) */
1055 pcpu->pc_curthread->td_flags |= TDF_NEEDRESCHED;
1056 ipi_selected( pcpu->pc_cpumask , IPI_AST);
1062 sched_add(struct thread *td, int flags)
1065 struct td_sched *ts;
1071 THREAD_LOCK_ASSERT(td, MA_OWNED);
1072 KASSERT((td->td_inhibitors == 0),
1073 ("sched_add: trying to run inhibited thread"));
1074 KASSERT((TD_CAN_RUN(td) || TD_IS_RUNNING(td)),
1075 ("sched_add: bad thread state"));
1076 KASSERT(td->td_flags & TDF_INMEM,
1077 ("sched_add: thread swapped out"));
1078 CTR5(KTR_SCHED, "sched_add: %p(%s) prio %d by %p(%s)",
1079 td, td->td_name, td->td_priority, curthread,
1080 curthread->td_name);
1082 * Now that the thread is moving to the run-queue, set the lock
1083 * to the scheduler's lock.
1085 if (td->td_lock != &sched_lock) {
1086 mtx_lock_spin(&sched_lock);
1087 thread_lock_set(td, &sched_lock);
1091 if (td->td_pinned != 0) {
1092 cpu = td->td_lastcpu;
1093 ts->ts_runq = &runq_pcpu[cpu];
1096 "sched_add: Put td_sched:%p(td:%p) on cpu%d runq", ts, td, cpu);
1097 } else if ((ts)->ts_flags & TSF_BOUND) {
1098 /* Find CPU from bound runq */
1099 KASSERT(SKE_RUNQ_PCPU(ts),("sched_add: bound td_sched not on cpu runq"));
1100 cpu = ts->ts_runq - &runq_pcpu[0];
1103 "sched_add: Put td_sched:%p(td:%p) on cpu%d runq", ts, td, cpu);
1106 "sched_add: adding td_sched:%p (td:%p) to gbl runq", ts, td);
1108 ts->ts_runq = &runq;
1111 if (single_cpu && (cpu != PCPU_GET(cpuid))) {
1112 kick_other_cpu(td->td_priority,cpu);
1116 cpumask_t me = PCPU_GET(cpumask);
1117 int idle = idle_cpus_mask & me;
1119 if (!idle && ((flags & SRQ_INTR) == 0) &&
1120 (idle_cpus_mask & ~(hlt_cpus_mask | me)))
1121 forwarded = forward_wakeup(cpu);
1125 if ((flags & SRQ_YIELDING) == 0 && maybe_preempt(td))
1132 if ((td->td_proc->p_flag & P_NOLOAD) == 0)
1134 runq_add(ts->ts_runq, ts, flags);
1138 struct td_sched *ts;
1140 THREAD_LOCK_ASSERT(td, MA_OWNED);
1141 KASSERT((td->td_inhibitors == 0),
1142 ("sched_add: trying to run inhibited thread"));
1143 KASSERT((TD_CAN_RUN(td) || TD_IS_RUNNING(td)),
1144 ("sched_add: bad thread state"));
1145 KASSERT(td->td_flags & TDF_INMEM,
1146 ("sched_add: thread swapped out"));
1147 CTR5(KTR_SCHED, "sched_add: %p(%s) prio %d by %p(%s)",
1148 td, td->td_name, td->td_priority, curthread,
1149 curthread->td_name);
1151 * Now that the thread is moving to the run-queue, set the lock
1152 * to the scheduler's lock.
1154 if (td->td_lock != &sched_lock) {
1155 mtx_lock_spin(&sched_lock);
1156 thread_lock_set(td, &sched_lock);
1159 CTR2(KTR_RUNQ, "sched_add: adding td_sched:%p (td:%p) to runq", ts, td);
1160 ts->ts_runq = &runq;
1163 * If we are yielding (on the way out anyhow)
1164 * or the thread being saved is US,
1165 * then don't try be smart about preemption
1166 * or kicking off another CPU
1167 * as it won't help and may hinder.
1168 * In the YIEDLING case, we are about to run whoever is
1169 * being put in the queue anyhow, and in the
1170 * OURSELF case, we are puting ourself on the run queue
1171 * which also only happens when we are about to yield.
1173 if((flags & SRQ_YIELDING) == 0) {
1174 if (maybe_preempt(td))
1177 if ((td->td_proc->p_flag & P_NOLOAD) == 0)
1179 runq_add(ts->ts_runq, ts, flags);
1185 sched_rem(struct thread *td)
1187 struct td_sched *ts;
1190 KASSERT(td->td_flags & TDF_INMEM,
1191 ("sched_rem: thread swapped out"));
1192 KASSERT(TD_ON_RUNQ(td),
1193 ("sched_rem: thread not on run queue"));
1194 mtx_assert(&sched_lock, MA_OWNED);
1195 CTR5(KTR_SCHED, "sched_rem: %p(%s) prio %d by %p(%s)",
1196 td, td->td_name, td->td_priority, curthread,
1197 curthread->td_name);
1199 if ((td->td_proc->p_flag & P_NOLOAD) == 0)
1201 runq_remove(ts->ts_runq, ts);
1206 * Select threads to run.
1207 * Notice that the running threads still consume a slot.
1212 struct td_sched *ts;
1215 mtx_assert(&sched_lock, MA_OWNED);
1217 struct td_sched *kecpu;
1220 ts = runq_choose(&runq);
1221 kecpu = runq_choose(&runq_pcpu[PCPU_GET(cpuid)]);
1225 kecpu->ts_thread->td_priority < ts->ts_thread->td_priority)) {
1226 CTR2(KTR_RUNQ, "choosing td_sched %p from pcpu runq %d", kecpu,
1229 rq = &runq_pcpu[PCPU_GET(cpuid)];
1231 CTR1(KTR_RUNQ, "choosing td_sched %p from main runq", ts);
1236 ts = runq_choose(&runq);
1240 runq_remove(rq, ts);
1241 ts->ts_flags |= TSF_DIDRUN;
1243 KASSERT(ts->ts_thread->td_flags & TDF_INMEM,
1244 ("sched_choose: thread swapped out"));
1245 return (ts->ts_thread);
1247 return (PCPU_GET(idlethread));
1251 sched_preempt(struct thread *td)
1254 if (td->td_critnest > 1)
1255 td->td_owepreempt = 1;
1257 mi_switch(SW_INVOL | SW_PREEMPT, NULL);
1262 sched_userret(struct thread *td)
1265 * XXX we cheat slightly on the locking here to avoid locking in
1266 * the usual case. Setting td_priority here is essentially an
1267 * incomplete workaround for not setting it properly elsewhere.
1268 * Now that some interrupt handlers are threads, not setting it
1269 * properly elsewhere can clobber it in the window between setting
1270 * it here and returning to user mode, so don't waste time setting
1271 * it perfectly here.
1273 KASSERT((td->td_flags & TDF_BORROWING) == 0,
1274 ("thread with borrowed priority returning to userland"));
1275 if (td->td_priority != td->td_user_pri) {
1277 td->td_priority = td->td_user_pri;
1278 td->td_base_pri = td->td_user_pri;
1284 sched_bind(struct thread *td, int cpu)
1286 struct td_sched *ts;
1288 THREAD_LOCK_ASSERT(td, MA_OWNED);
1289 KASSERT(TD_IS_RUNNING(td),
1290 ("sched_bind: cannot bind non-running thread"));
1294 ts->ts_flags |= TSF_BOUND;
1296 ts->ts_runq = &runq_pcpu[cpu];
1297 if (PCPU_GET(cpuid) == cpu)
1300 mi_switch(SW_VOL, NULL);
1305 sched_unbind(struct thread* td)
1307 THREAD_LOCK_ASSERT(td, MA_OWNED);
1308 td->td_sched->ts_flags &= ~TSF_BOUND;
1312 sched_is_bound(struct thread *td)
1314 THREAD_LOCK_ASSERT(td, MA_OWNED);
1315 return (td->td_sched->ts_flags & TSF_BOUND);
1319 sched_relinquish(struct thread *td)
1322 SCHED_STAT_INC(switch_relinquish);
1323 mi_switch(SW_VOL, NULL);
1330 return (sched_tdcnt);
1334 sched_sizeof_proc(void)
1336 return (sizeof(struct proc));
1340 sched_sizeof_thread(void)
1342 return (sizeof(struct thread) + sizeof(struct td_sched));
1346 sched_pctcpu(struct thread *td)
1348 struct td_sched *ts;
1351 return (ts->ts_pctcpu);
1360 * The actual idle process.
1363 sched_idletd(void *dummy)
1367 mtx_assert(&Giant, MA_NOTOWNED);
1369 while (sched_runnable() == 0)
1372 mtx_lock_spin(&sched_lock);
1373 mi_switch(SW_VOL, NULL);
1374 mtx_unlock_spin(&sched_lock);
1379 * A CPU is entering for the first time or a thread is exiting.
1382 sched_throw(struct thread *td)
1385 * Correct spinlock nesting. The idle thread context that we are
1386 * borrowing was created so that it would start out with a single
1387 * spin lock (sched_lock) held in fork_trampoline(). Since we've
1388 * explicitly acquired locks in this function, the nesting count
1389 * is now 2 rather than 1. Since we are nested, calling
1390 * spinlock_exit() will simply adjust the counts without allowing
1391 * spin lock using code to interrupt us.
1394 mtx_lock_spin(&sched_lock);
1397 lock_profile_release_lock(&sched_lock.lock_object);
1398 MPASS(td->td_lock == &sched_lock);
1400 mtx_assert(&sched_lock, MA_OWNED);
1401 KASSERT(curthread->td_md.md_spinlock_count == 1, ("invalid count"));
1402 PCPU_SET(switchtime, cpu_ticks());
1403 PCPU_SET(switchticks, ticks);
1404 cpu_throw(td, choosethread()); /* doesn't return */
1408 sched_fork_exit(struct thread *td)
1412 * Finish setting up thread glue so that it begins execution in a
1413 * non-nested critical section with sched_lock held but not recursed.
1415 td->td_oncpu = PCPU_GET(cpuid);
1416 sched_lock.mtx_lock = (uintptr_t)td;
1417 lock_profile_obtain_lock_success(&sched_lock.lock_object,
1418 0, 0, __FILE__, __LINE__);
1419 THREAD_LOCK_ASSERT(td, MA_OWNED | MA_NOTRECURSED);
1423 sched_affinity(struct thread *td)
1427 #define KERN_SWITCH_INCLUDE 1
1428 #include "kern/kern_switch.c"