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"
40 #include <sys/param.h>
41 #include <sys/systm.h>
42 #include <sys/kernel.h>
45 #include <sys/kthread.h>
46 #include <sys/mutex.h>
48 #include <sys/resourcevar.h>
49 #include <sys/sched.h>
51 #include <sys/sysctl.h>
53 #include <sys/turnstile.h>
55 #include <machine/pcb.h>
56 #include <machine/smp.h>
59 #include <sys/pmckern.h>
63 * INVERSE_ESTCPU_WEIGHT is only suitable for statclock() frequencies in
64 * the range 100-256 Hz (approximately).
66 #define ESTCPULIM(e) \
67 min((e), INVERSE_ESTCPU_WEIGHT * (NICE_WEIGHT * (PRIO_MAX - PRIO_MIN) - \
68 RQ_PPQ) + INVERSE_ESTCPU_WEIGHT - 1)
70 #define INVERSE_ESTCPU_WEIGHT (8 * smp_cpus)
72 #define INVERSE_ESTCPU_WEIGHT 8 /* 1 / (priorities per estcpu level). */
74 #define NICE_WEIGHT 1 /* Priorities per nice level. */
77 * The schedulable entity that runs a context.
78 * This is an extension to the thread structure and is tailored to
79 * the requirements of this scheduler
82 TAILQ_ENTRY(td_sched) ts_procq; /* (j/z) Run queue. */
83 struct thread *ts_thread; /* (*) Active associated thread. */
84 fixpt_t ts_pctcpu; /* (j) %cpu during p_swtime. */
85 u_char ts_rqindex; /* (j) Run queue index. */
86 int ts_cpticks; /* (j) Ticks of cpu time. */
87 int ts_slptime; /* (j) Seconds !RUNNING. */
88 struct runq *ts_runq; /* runq the thread is currently on */
91 /* flags kept in td_flags */
92 #define TDF_DIDRUN TDF_SCHED0 /* thread actually ran. */
93 #define TDF_EXIT TDF_SCHED1 /* thread is being killed. */
94 #define TDF_BOUND TDF_SCHED2
96 #define ts_flags ts_thread->td_flags
97 #define TSF_DIDRUN TDF_DIDRUN /* thread actually ran. */
98 #define TSF_EXIT TDF_EXIT /* thread is being killed. */
99 #define TSF_BOUND TDF_BOUND /* stuck to one CPU */
101 #define SKE_RUNQ_PCPU(ts) \
102 ((ts)->ts_runq != 0 && (ts)->ts_runq != &runq)
104 static struct td_sched td_sched0;
105 struct mtx sched_lock;
107 static int sched_tdcnt; /* Total runnable threads in the system. */
108 static int sched_quantum; /* Roundrobin scheduling quantum in ticks. */
109 #define SCHED_QUANTUM (hz / 10) /* Default sched quantum */
111 static void setup_runqs(void);
112 static void schedcpu(void);
113 static void schedcpu_thread(void);
114 static void sched_priority(struct thread *td, u_char prio);
115 static void sched_setup(void *dummy);
116 static void maybe_resched(struct thread *td);
117 static void updatepri(struct thread *td);
118 static void resetpriority(struct thread *td);
119 static void resetpriority_thread(struct thread *td);
121 static int forward_wakeup(int cpunum);
124 static struct kproc_desc sched_kp = {
129 SYSINIT(schedcpu, SI_SUB_RUN_SCHEDULER, SI_ORDER_FIRST, kproc_start, &sched_kp)
130 SYSINIT(sched_setup, SI_SUB_RUN_QUEUE, SI_ORDER_FIRST, sched_setup, NULL)
135 static struct runq runq;
141 static struct runq runq_pcpu[MAXCPU];
150 for (i = 0; i < MAXCPU; ++i)
151 runq_init(&runq_pcpu[i]);
158 sysctl_kern_quantum(SYSCTL_HANDLER_ARGS)
162 new_val = sched_quantum * tick;
163 error = sysctl_handle_int(oidp, &new_val, 0, req);
164 if (error != 0 || req->newptr == NULL)
168 sched_quantum = new_val / tick;
169 hogticks = 2 * sched_quantum;
173 SYSCTL_NODE(_kern, OID_AUTO, sched, CTLFLAG_RD, 0, "Scheduler");
175 SYSCTL_STRING(_kern_sched, OID_AUTO, name, CTLFLAG_RD, "4BSD", 0,
178 SYSCTL_PROC(_kern_sched, OID_AUTO, quantum, CTLTYPE_INT | CTLFLAG_RW,
179 0, sizeof sched_quantum, sysctl_kern_quantum, "I",
180 "Roundrobin scheduling quantum in microseconds");
183 /* Enable forwarding of wakeups to all other cpus */
184 SYSCTL_NODE(_kern_sched, OID_AUTO, ipiwakeup, CTLFLAG_RD, NULL, "Kernel SMP");
186 static int forward_wakeup_enabled = 1;
187 SYSCTL_INT(_kern_sched_ipiwakeup, OID_AUTO, enabled, CTLFLAG_RW,
188 &forward_wakeup_enabled, 0,
189 "Forwarding of wakeup to idle CPUs");
191 static int forward_wakeups_requested = 0;
192 SYSCTL_INT(_kern_sched_ipiwakeup, OID_AUTO, requested, CTLFLAG_RD,
193 &forward_wakeups_requested, 0,
194 "Requests for Forwarding of wakeup to idle CPUs");
196 static int forward_wakeups_delivered = 0;
197 SYSCTL_INT(_kern_sched_ipiwakeup, OID_AUTO, delivered, CTLFLAG_RD,
198 &forward_wakeups_delivered, 0,
199 "Completed Forwarding of wakeup to idle CPUs");
201 static int forward_wakeup_use_mask = 1;
202 SYSCTL_INT(_kern_sched_ipiwakeup, OID_AUTO, usemask, CTLFLAG_RW,
203 &forward_wakeup_use_mask, 0,
204 "Use the mask of idle cpus");
206 static int forward_wakeup_use_loop = 0;
207 SYSCTL_INT(_kern_sched_ipiwakeup, OID_AUTO, useloop, CTLFLAG_RW,
208 &forward_wakeup_use_loop, 0,
209 "Use a loop to find idle cpus");
211 static int forward_wakeup_use_single = 0;
212 SYSCTL_INT(_kern_sched_ipiwakeup, OID_AUTO, onecpu, CTLFLAG_RW,
213 &forward_wakeup_use_single, 0,
214 "Only signal one idle cpu");
216 static int forward_wakeup_use_htt = 0;
217 SYSCTL_INT(_kern_sched_ipiwakeup, OID_AUTO, htt2, CTLFLAG_RW,
218 &forward_wakeup_use_htt, 0,
223 static int sched_followon = 0;
224 SYSCTL_INT(_kern_sched, OID_AUTO, followon, CTLFLAG_RW,
226 "allow threads to share a quantum");
233 CTR1(KTR_SCHED, "global load: %d", sched_tdcnt);
240 CTR1(KTR_SCHED, "global load: %d", sched_tdcnt);
243 * Arrange to reschedule if necessary, taking the priorities and
244 * schedulers into account.
247 maybe_resched(struct thread *td)
250 THREAD_LOCK_ASSERT(td, MA_OWNED);
251 if (td->td_priority < curthread->td_priority)
252 curthread->td_flags |= TDF_NEEDRESCHED;
256 * Constants for digital decay and forget:
257 * 90% of (td_estcpu) usage in 5 * loadav time
258 * 95% of (ts_pctcpu) usage in 60 seconds (load insensitive)
259 * Note that, as ps(1) mentions, this can let percentages
260 * total over 100% (I've seen 137.9% for 3 processes).
262 * Note that schedclock() updates td_estcpu and p_cpticks asynchronously.
264 * We wish to decay away 90% of td_estcpu in (5 * loadavg) seconds.
265 * That is, the system wants to compute a value of decay such
266 * that the following for loop:
267 * for (i = 0; i < (5 * loadavg); i++)
268 * td_estcpu *= decay;
271 * for all values of loadavg:
273 * Mathematically this loop can be expressed by saying:
274 * decay ** (5 * loadavg) ~= .1
276 * The system computes decay as:
277 * decay = (2 * loadavg) / (2 * loadavg + 1)
279 * We wish to prove that the system's computation of decay
280 * will always fulfill the equation:
281 * decay ** (5 * loadavg) ~= .1
283 * If we compute b as:
286 * decay = b / (b + 1)
288 * We now need to prove two things:
289 * 1) Given factor ** (5 * loadavg) ~= .1, prove factor == b/(b+1)
290 * 2) Given b/(b+1) ** power ~= .1, prove power == (5 * loadavg)
293 * For x close to zero, exp(x) =~ 1 + x, since
294 * exp(x) = 0! + x**1/1! + x**2/2! + ... .
295 * therefore exp(-1/b) =~ 1 - (1/b) = (b-1)/b.
296 * For x close to zero, ln(1+x) =~ x, since
297 * ln(1+x) = x - x**2/2 + x**3/3 - ... -1 < x < 1
298 * therefore ln(b/(b+1)) = ln(1 - 1/(b+1)) =~ -1/(b+1).
302 * Solve (factor)**(power) =~ .1 given power (5*loadav):
303 * solving for factor,
304 * ln(factor) =~ (-2.30/5*loadav), or
305 * factor =~ exp(-1/((5/2.30)*loadav)) =~ exp(-1/(2*loadav)) =
306 * exp(-1/b) =~ (b-1)/b =~ b/(b+1). QED
309 * Solve (factor)**(power) =~ .1 given factor == (b/(b+1)):
311 * power*ln(b/(b+1)) =~ -2.30, or
312 * power =~ 2.3 * (b + 1) = 4.6*loadav + 2.3 =~ 5*loadav. QED
314 * Actual power values for the implemented algorithm are as follows:
316 * power: 5.68 10.32 14.94 19.55
319 /* calculations for digital decay to forget 90% of usage in 5*loadav sec */
320 #define loadfactor(loadav) (2 * (loadav))
321 #define decay_cpu(loadfac, cpu) (((loadfac) * (cpu)) / ((loadfac) + FSCALE))
323 /* decay 95% of `ts_pctcpu' in 60 seconds; see CCPU_SHIFT before changing */
324 static fixpt_t ccpu = 0.95122942450071400909 * FSCALE; /* exp(-1/20) */
325 SYSCTL_INT(_kern, OID_AUTO, ccpu, CTLFLAG_RD, &ccpu, 0, "");
328 * If `ccpu' is not equal to `exp(-1/20)' and you still want to use the
329 * faster/more-accurate formula, you'll have to estimate CCPU_SHIFT below
330 * and possibly adjust FSHIFT in "param.h" so that (FSHIFT >= CCPU_SHIFT).
332 * To estimate CCPU_SHIFT for exp(-1/20), the following formula was used:
333 * 1 - exp(-1/20) ~= 0.0487 ~= 0.0488 == 1 (fixed pt, *11* bits).
335 * If you don't want to bother with the faster/more-accurate formula, you
336 * can set CCPU_SHIFT to (FSHIFT + 1) which will use a slower/less-accurate
337 * (more general) method of calculating the %age of CPU used by a process.
339 #define CCPU_SHIFT 11
342 * Recompute process priorities, every hz ticks.
343 * MP-safe, called without the Giant mutex.
349 register fixpt_t loadfac = loadfactor(averunnable.ldavg[0]);
353 int awake, realstathz;
355 realstathz = stathz ? stathz : hz;
356 sx_slock(&allproc_lock);
357 FOREACH_PROC_IN_SYSTEM(p) {
359 FOREACH_THREAD_IN_PROC(p, td) {
364 * Increment sleep time (if sleeping). We
365 * ignore overflow, as above.
368 * The td_sched slptimes are not touched in wakeup
369 * because the thread may not HAVE everything in
370 * memory? XXX I think this is out of date.
372 if (TD_ON_RUNQ(td)) {
374 ts->ts_flags &= ~TSF_DIDRUN;
375 } else if (TD_IS_RUNNING(td)) {
377 /* Do not clear TSF_DIDRUN */
378 } else if (ts->ts_flags & TSF_DIDRUN) {
380 ts->ts_flags &= ~TSF_DIDRUN;
384 * ts_pctcpu is only for ps and ttyinfo().
385 * Do it per td_sched, and add them up at the end?
388 ts->ts_pctcpu = (ts->ts_pctcpu * ccpu) >> FSHIFT;
390 * If the td_sched has been idle the entire second,
391 * stop recalculating its priority until
394 if (ts->ts_cpticks != 0) {
395 #if (FSHIFT >= CCPU_SHIFT)
396 ts->ts_pctcpu += (realstathz == 100)
397 ? ((fixpt_t) ts->ts_cpticks) <<
398 (FSHIFT - CCPU_SHIFT) :
399 100 * (((fixpt_t) ts->ts_cpticks)
400 << (FSHIFT - CCPU_SHIFT)) / realstathz;
402 ts->ts_pctcpu += ((FSCALE - ccpu) *
404 FSCALE / realstathz)) >> FSHIFT;
409 * If there are ANY running threads in this process,
410 * then don't count it as sleeping.
415 if (ts->ts_slptime > 1) {
417 * In an ideal world, this should not
418 * happen, because whoever woke us
419 * up from the long sleep should have
420 * unwound the slptime and reset our
421 * priority before we run at the stale
422 * priority. Should KASSERT at some
423 * point when all the cases are fixed.
430 if (ts->ts_slptime > 1) {
434 td->td_estcpu = decay_cpu(loadfac, td->td_estcpu);
436 resetpriority_thread(td);
438 } /* end of thread loop */
440 } /* end of process loop */
441 sx_sunlock(&allproc_lock);
445 * Main loop for a kthread that executes schedcpu once a second.
448 schedcpu_thread(void)
458 * Recalculate the priority of a process after it has slept for a while.
459 * For all load averages >= 1 and max td_estcpu of 255, sleeping for at
460 * least six times the loadfactor will decay td_estcpu to zero.
463 updatepri(struct thread *td)
470 loadfac = loadfactor(averunnable.ldavg[0]);
471 if (ts->ts_slptime > 5 * loadfac)
474 newcpu = td->td_estcpu;
475 ts->ts_slptime--; /* was incremented in schedcpu() */
476 while (newcpu && --ts->ts_slptime)
477 newcpu = decay_cpu(loadfac, newcpu);
478 td->td_estcpu = newcpu;
483 * Compute the priority of a process when running in user mode.
484 * Arrange to reschedule if the resulting priority is better
485 * than that of the current process.
488 resetpriority(struct thread *td)
490 register unsigned int newpriority;
492 if (td->td_pri_class == PRI_TIMESHARE) {
493 newpriority = PUSER + td->td_estcpu / INVERSE_ESTCPU_WEIGHT +
494 NICE_WEIGHT * (td->td_proc->p_nice - PRIO_MIN);
495 newpriority = min(max(newpriority, PRI_MIN_TIMESHARE),
497 sched_user_prio(td, newpriority);
502 * Update the thread's priority when the associated process's user
506 resetpriority_thread(struct thread *td)
509 /* Only change threads with a time sharing user priority. */
510 if (td->td_priority < PRI_MIN_TIMESHARE ||
511 td->td_priority > PRI_MAX_TIMESHARE)
514 /* XXX the whole needresched thing is broken, but not silly. */
517 sched_prio(td, td->td_user_pri);
522 sched_setup(void *dummy)
526 if (sched_quantum == 0)
527 sched_quantum = SCHED_QUANTUM;
528 hogticks = 2 * sched_quantum;
530 /* Account for thread0. */
534 /* External interfaces start here */
536 * Very early in the boot some setup of scheduler-specific
537 * parts of proc0 and of some scheduler resources needs to be done.
545 * Set up the scheduler specific parts of proc0.
547 proc0.p_sched = NULL; /* XXX */
548 thread0.td_sched = &td_sched0;
549 thread0.td_lock = &sched_lock;
550 td_sched0.ts_thread = &thread0;
551 mtx_init(&sched_lock, "sched lock", NULL, MTX_SPIN | MTX_RECURSE);
558 return runq_check(&runq) + runq_check(&runq_pcpu[PCPU_GET(cpuid)]);
560 return runq_check(&runq);
565 sched_rr_interval(void)
567 if (sched_quantum == 0)
568 sched_quantum = SCHED_QUANTUM;
569 return (sched_quantum);
573 * We adjust the priority of the current process. The priority of
574 * a process gets worse as it accumulates CPU time. The cpu usage
575 * estimator (td_estcpu) is increased here. resetpriority() will
576 * compute a different priority each time td_estcpu increases by
577 * INVERSE_ESTCPU_WEIGHT
578 * (until MAXPRI is reached). The cpu usage estimator ramps up
579 * quite quickly when the process is running (linearly), and decays
580 * away exponentially, at a rate which is proportionally slower when
581 * the system is busy. The basic principle is that the system will
582 * 90% forget that the process used a lot of CPU time in 5 * loadav
583 * seconds. This causes the system to favor processes which haven't
584 * run much recently, and to round-robin among other processes.
587 sched_clock(struct thread *td)
591 THREAD_LOCK_ASSERT(td, MA_OWNED);
595 td->td_estcpu = ESTCPULIM(td->td_estcpu + 1);
596 if ((td->td_estcpu % INVERSE_ESTCPU_WEIGHT) == 0) {
598 resetpriority_thread(td);
602 * Force a context switch if the current thread has used up a full
603 * quantum (default quantum is 100ms).
605 if (!TD_IS_IDLETHREAD(td) &&
606 ticks - PCPU_GET(switchticks) >= sched_quantum)
607 td->td_flags |= TDF_NEEDRESCHED;
611 * charge childs scheduling cpu usage to parent.
614 sched_exit(struct proc *p, struct thread *td)
617 CTR3(KTR_SCHED, "sched_exit: %p(%s) prio %d",
618 td, td->td_name, td->td_priority);
619 PROC_SLOCK_ASSERT(p, MA_OWNED);
620 sched_exit_thread(FIRST_THREAD_IN_PROC(p), td);
624 sched_exit_thread(struct thread *td, struct thread *child)
627 CTR3(KTR_SCHED, "sched_exit_thread: %p(%s) prio %d",
628 child, child->td_name, child->td_priority);
630 td->td_estcpu = ESTCPULIM(td->td_estcpu + child->td_estcpu);
632 mtx_lock_spin(&sched_lock);
633 if ((child->td_proc->p_flag & P_NOLOAD) == 0)
635 mtx_unlock_spin(&sched_lock);
639 sched_fork(struct thread *td, struct thread *childtd)
641 sched_fork_thread(td, childtd);
645 sched_fork_thread(struct thread *td, struct thread *childtd)
647 childtd->td_estcpu = td->td_estcpu;
648 childtd->td_lock = &sched_lock;
649 sched_newthread(childtd);
653 sched_nice(struct proc *p, int nice)
657 PROC_LOCK_ASSERT(p, MA_OWNED);
658 PROC_SLOCK_ASSERT(p, MA_OWNED);
660 FOREACH_THREAD_IN_PROC(p, td) {
663 resetpriority_thread(td);
669 sched_class(struct thread *td, int class)
671 THREAD_LOCK_ASSERT(td, MA_OWNED);
672 td->td_pri_class = class;
676 * Adjust the priority of a thread.
679 sched_priority(struct thread *td, u_char prio)
681 CTR6(KTR_SCHED, "sched_prio: %p(%s) prio %d newprio %d by %p(%s)",
682 td, td->td_name, td->td_priority, prio, curthread,
685 THREAD_LOCK_ASSERT(td, MA_OWNED);
686 if (td->td_priority == prio)
688 td->td_priority = prio;
689 if (TD_ON_RUNQ(td) &&
690 td->td_sched->ts_rqindex != (prio / RQ_PPQ)) {
692 sched_add(td, SRQ_BORING);
697 * Update a thread's priority when it is lent another thread's
701 sched_lend_prio(struct thread *td, u_char prio)
704 td->td_flags |= TDF_BORROWING;
705 sched_priority(td, prio);
709 * Restore a thread's priority when priority propagation is
710 * over. The prio argument is the minimum priority the thread
711 * needs to have to satisfy other possible priority lending
712 * requests. If the thread's regulary priority is less
713 * important than prio the thread will keep a priority boost
717 sched_unlend_prio(struct thread *td, u_char prio)
721 if (td->td_base_pri >= PRI_MIN_TIMESHARE &&
722 td->td_base_pri <= PRI_MAX_TIMESHARE)
723 base_pri = td->td_user_pri;
725 base_pri = td->td_base_pri;
726 if (prio >= base_pri) {
727 td->td_flags &= ~TDF_BORROWING;
728 sched_prio(td, base_pri);
730 sched_lend_prio(td, prio);
734 sched_prio(struct thread *td, u_char prio)
738 /* First, update the base priority. */
739 td->td_base_pri = prio;
742 * If the thread is borrowing another thread's priority, don't ever
743 * lower the priority.
745 if (td->td_flags & TDF_BORROWING && td->td_priority < prio)
748 /* Change the real priority. */
749 oldprio = td->td_priority;
750 sched_priority(td, prio);
753 * If the thread is on a turnstile, then let the turnstile update
756 if (TD_ON_LOCK(td) && oldprio != prio)
757 turnstile_adjust(td, oldprio);
761 sched_user_prio(struct thread *td, u_char prio)
765 THREAD_LOCK_ASSERT(td, MA_OWNED);
766 td->td_base_user_pri = prio;
767 if (td->td_flags & TDF_UBORROWING && td->td_user_pri <= prio)
769 oldprio = td->td_user_pri;
770 td->td_user_pri = prio;
774 sched_lend_user_prio(struct thread *td, u_char prio)
778 THREAD_LOCK_ASSERT(td, MA_OWNED);
779 td->td_flags |= TDF_UBORROWING;
780 oldprio = td->td_user_pri;
781 td->td_user_pri = prio;
785 sched_unlend_user_prio(struct thread *td, u_char prio)
789 THREAD_LOCK_ASSERT(td, MA_OWNED);
790 base_pri = td->td_base_user_pri;
791 if (prio >= base_pri) {
792 td->td_flags &= ~TDF_UBORROWING;
793 sched_user_prio(td, base_pri);
795 sched_lend_user_prio(td, prio);
800 sched_sleep(struct thread *td)
803 THREAD_LOCK_ASSERT(td, MA_OWNED);
804 td->td_slptick = ticks;
805 td->td_sched->ts_slptime = 0;
809 sched_switch(struct thread *td, struct thread *newtd, int flags)
817 THREAD_LOCK_ASSERT(td, MA_OWNED);
819 * Switch to the sched lock to fix things up and pick
822 if (td->td_lock != &sched_lock) {
823 mtx_lock_spin(&sched_lock);
827 if ((p->p_flag & P_NOLOAD) == 0)
831 newtd->td_flags |= (td->td_flags & TDF_NEEDRESCHED);
833 td->td_lastcpu = td->td_oncpu;
834 td->td_flags &= ~TDF_NEEDRESCHED;
835 td->td_owepreempt = 0;
836 td->td_oncpu = NOCPU;
838 * At the last moment, if this thread is still marked RUNNING,
839 * then put it back on the run queue as it has not been suspended
840 * or stopped or any thing else similar. We never put the idle
841 * threads on the run queue, however.
843 if (td->td_flags & TDF_IDLETD) {
846 idle_cpus_mask &= ~PCPU_GET(cpumask);
849 if (TD_IS_RUNNING(td)) {
850 /* Put us back on the run queue. */
851 sched_add(td, (flags & SW_PREEMPT) ?
852 SRQ_OURSELF|SRQ_YIELDING|SRQ_PREEMPTED :
853 SRQ_OURSELF|SRQ_YIELDING);
858 * The thread we are about to run needs to be counted
859 * as if it had been added to the run queue and selected.
865 KASSERT((newtd->td_inhibitors == 0),
866 ("trying to run inhibited thread"));
867 newtd->td_sched->ts_flags |= TSF_DIDRUN;
868 TD_SET_RUNNING(newtd);
869 if ((newtd->td_proc->p_flag & P_NOLOAD) == 0)
872 newtd = choosethread();
874 MPASS(newtd->td_lock == &sched_lock);
878 if (PMC_PROC_IS_USING_PMCS(td->td_proc))
879 PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_OUT);
882 lock_profile_release_lock(&sched_lock.lock_object);
883 cpu_switch(td, newtd, td->td_lock);
884 lock_profile_obtain_lock_success(&sched_lock.lock_object,
885 0, 0, __FILE__, __LINE__);
887 * Where am I? What year is it?
888 * We are in the same thread that went to sleep above,
889 * but any amount of time may have passed. All out context
890 * will still be available as will local variables.
891 * PCPU values however may have changed as we may have
892 * changed CPU so don't trust cached values of them.
893 * New threads will go to fork_exit() instead of here
894 * so if you change things here you may need to change
896 * If the thread above was exiting it will never wake
897 * up again here, so either it has saved everything it
898 * needed to, or the thread_wait() or wait() will
902 if (PMC_PROC_IS_USING_PMCS(td->td_proc))
903 PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_IN);
908 if (td->td_flags & TDF_IDLETD)
909 idle_cpus_mask |= PCPU_GET(cpumask);
911 sched_lock.mtx_lock = (uintptr_t)td;
912 td->td_oncpu = PCPU_GET(cpuid);
913 MPASS(td->td_lock == &sched_lock);
917 sched_wakeup(struct thread *td)
921 THREAD_LOCK_ASSERT(td, MA_OWNED);
923 if (ts->ts_slptime > 1) {
927 td->td_slptick = ticks;
929 sched_add(td, SRQ_BORING);
933 /* enable HTT_2 if you have a 2-way HTT cpu.*/
935 forward_wakeup(int cpunum)
937 cpumask_t map, me, dontuse;
942 mtx_assert(&sched_lock, MA_OWNED);
944 CTR0(KTR_RUNQ, "forward_wakeup()");
946 if ((!forward_wakeup_enabled) ||
947 (forward_wakeup_use_mask == 0 && forward_wakeup_use_loop == 0))
949 if (!smp_started || cold || panicstr)
952 forward_wakeups_requested++;
955 * check the idle mask we received against what we calculated before
956 * in the old version.
958 me = PCPU_GET(cpumask);
960 * don't bother if we should be doing it ourself..
962 if ((me & idle_cpus_mask) && (cpunum == NOCPU || me == (1 << cpunum)))
965 dontuse = me | stopped_cpus | hlt_cpus_mask;
967 if (forward_wakeup_use_loop) {
968 SLIST_FOREACH(pc, &cpuhead, pc_allcpu) {
970 if ( (id & dontuse) == 0 &&
971 pc->pc_curthread == pc->pc_idlethread) {
977 if (forward_wakeup_use_mask) {
979 map = idle_cpus_mask & ~dontuse;
981 /* If they are both on, compare and use loop if different */
982 if (forward_wakeup_use_loop) {
984 printf("map (%02X) != map3 (%02X)\n",
992 /* If we only allow a specific CPU, then mask off all the others */
993 if (cpunum != NOCPU) {
994 KASSERT((cpunum <= mp_maxcpus),("forward_wakeup: bad cpunum."));
995 map &= (1 << cpunum);
997 /* Try choose an idle die. */
998 if (forward_wakeup_use_htt) {
999 map2 = (map & (map >> 1)) & 0x5555;
1005 /* set only one bit */
1006 if (forward_wakeup_use_single) {
1007 map = map & ((~map) + 1);
1011 forward_wakeups_delivered++;
1012 ipi_selected(map, IPI_AST);
1015 if (cpunum == NOCPU)
1016 printf("forward_wakeup: Idle processor not found\n");
1022 static void kick_other_cpu(int pri,int cpuid);
1025 kick_other_cpu(int pri,int cpuid)
1027 struct pcpu * pcpu = pcpu_find(cpuid);
1028 int cpri = pcpu->pc_curthread->td_priority;
1030 if (idle_cpus_mask & pcpu->pc_cpumask) {
1031 forward_wakeups_delivered++;
1032 ipi_selected(pcpu->pc_cpumask, IPI_AST);
1039 #if defined(IPI_PREEMPTION) && defined(PREEMPTION)
1040 #if !defined(FULL_PREEMPTION)
1041 if (pri <= PRI_MAX_ITHD)
1042 #endif /* ! FULL_PREEMPTION */
1044 ipi_selected(pcpu->pc_cpumask, IPI_PREEMPT);
1047 #endif /* defined(IPI_PREEMPTION) && defined(PREEMPTION) */
1049 pcpu->pc_curthread->td_flags |= TDF_NEEDRESCHED;
1050 ipi_selected( pcpu->pc_cpumask , IPI_AST);
1056 sched_add(struct thread *td, int flags)
1059 struct td_sched *ts;
1065 THREAD_LOCK_ASSERT(td, MA_OWNED);
1066 KASSERT((td->td_inhibitors == 0),
1067 ("sched_add: trying to run inhibited thread"));
1068 KASSERT((TD_CAN_RUN(td) || TD_IS_RUNNING(td)),
1069 ("sched_add: bad thread state"));
1070 KASSERT(td->td_flags & TDF_INMEM,
1071 ("sched_add: thread swapped out"));
1072 CTR5(KTR_SCHED, "sched_add: %p(%s) prio %d by %p(%s)",
1073 td, td->td_name, td->td_priority, curthread,
1074 curthread->td_name);
1076 * Now that the thread is moving to the run-queue, set the lock
1077 * to the scheduler's lock.
1079 if (td->td_lock != &sched_lock) {
1080 mtx_lock_spin(&sched_lock);
1081 thread_lock_set(td, &sched_lock);
1085 if (td->td_pinned != 0) {
1086 cpu = td->td_lastcpu;
1087 ts->ts_runq = &runq_pcpu[cpu];
1090 "sched_add: Put td_sched:%p(td:%p) on cpu%d runq", ts, td, cpu);
1091 } else if ((ts)->ts_flags & TSF_BOUND) {
1092 /* Find CPU from bound runq */
1093 KASSERT(SKE_RUNQ_PCPU(ts),("sched_add: bound td_sched not on cpu runq"));
1094 cpu = ts->ts_runq - &runq_pcpu[0];
1097 "sched_add: Put td_sched:%p(td:%p) on cpu%d runq", ts, td, cpu);
1100 "sched_add: adding td_sched:%p (td:%p) to gbl runq", ts, td);
1102 ts->ts_runq = &runq;
1105 if (single_cpu && (cpu != PCPU_GET(cpuid))) {
1106 kick_other_cpu(td->td_priority,cpu);
1110 cpumask_t me = PCPU_GET(cpumask);
1111 int idle = idle_cpus_mask & me;
1113 if (!idle && ((flags & SRQ_INTR) == 0) &&
1114 (idle_cpus_mask & ~(hlt_cpus_mask | me)))
1115 forwarded = forward_wakeup(cpu);
1119 if ((flags & SRQ_YIELDING) == 0 && maybe_preempt(td))
1126 if ((td->td_proc->p_flag & P_NOLOAD) == 0)
1128 runq_add(ts->ts_runq, ts, flags);
1132 struct td_sched *ts;
1134 THREAD_LOCK_ASSERT(td, MA_OWNED);
1135 KASSERT((td->td_inhibitors == 0),
1136 ("sched_add: trying to run inhibited thread"));
1137 KASSERT((TD_CAN_RUN(td) || TD_IS_RUNNING(td)),
1138 ("sched_add: bad thread state"));
1139 KASSERT(td->td_flags & TDF_INMEM,
1140 ("sched_add: thread swapped out"));
1141 CTR5(KTR_SCHED, "sched_add: %p(%s) prio %d by %p(%s)",
1142 td, td->td_name, td->td_priority, curthread,
1143 curthread->td_name);
1145 * Now that the thread is moving to the run-queue, set the lock
1146 * to the scheduler's lock.
1148 if (td->td_lock != &sched_lock) {
1149 mtx_lock_spin(&sched_lock);
1150 thread_lock_set(td, &sched_lock);
1153 CTR2(KTR_RUNQ, "sched_add: adding td_sched:%p (td:%p) to runq", ts, td);
1154 ts->ts_runq = &runq;
1157 * If we are yielding (on the way out anyhow)
1158 * or the thread being saved is US,
1159 * then don't try be smart about preemption
1160 * or kicking off another CPU
1161 * as it won't help and may hinder.
1162 * In the YIEDLING case, we are about to run whoever is
1163 * being put in the queue anyhow, and in the
1164 * OURSELF case, we are puting ourself on the run queue
1165 * which also only happens when we are about to yield.
1167 if((flags & SRQ_YIELDING) == 0) {
1168 if (maybe_preempt(td))
1171 if ((td->td_proc->p_flag & P_NOLOAD) == 0)
1173 runq_add(ts->ts_runq, ts, flags);
1179 sched_rem(struct thread *td)
1181 struct td_sched *ts;
1184 KASSERT(td->td_flags & TDF_INMEM,
1185 ("sched_rem: thread swapped out"));
1186 KASSERT(TD_ON_RUNQ(td),
1187 ("sched_rem: thread not on run queue"));
1188 mtx_assert(&sched_lock, MA_OWNED);
1189 CTR5(KTR_SCHED, "sched_rem: %p(%s) prio %d by %p(%s)",
1190 td, td->td_name, td->td_priority, curthread,
1191 curthread->td_name);
1193 if ((td->td_proc->p_flag & P_NOLOAD) == 0)
1195 runq_remove(ts->ts_runq, ts);
1200 * Select threads to run.
1201 * Notice that the running threads still consume a slot.
1206 struct td_sched *ts;
1209 mtx_assert(&sched_lock, MA_OWNED);
1211 struct td_sched *kecpu;
1214 ts = runq_choose(&runq);
1215 kecpu = runq_choose(&runq_pcpu[PCPU_GET(cpuid)]);
1219 kecpu->ts_thread->td_priority < ts->ts_thread->td_priority)) {
1220 CTR2(KTR_RUNQ, "choosing td_sched %p from pcpu runq %d", kecpu,
1223 rq = &runq_pcpu[PCPU_GET(cpuid)];
1225 CTR1(KTR_RUNQ, "choosing td_sched %p from main runq", ts);
1230 ts = runq_choose(&runq);
1234 runq_remove(rq, ts);
1235 ts->ts_flags |= TSF_DIDRUN;
1237 KASSERT(ts->ts_thread->td_flags & TDF_INMEM,
1238 ("sched_choose: thread swapped out"));
1239 return (ts->ts_thread);
1241 return (PCPU_GET(idlethread));
1245 sched_userret(struct thread *td)
1248 * XXX we cheat slightly on the locking here to avoid locking in
1249 * the usual case. Setting td_priority here is essentially an
1250 * incomplete workaround for not setting it properly elsewhere.
1251 * Now that some interrupt handlers are threads, not setting it
1252 * properly elsewhere can clobber it in the window between setting
1253 * it here and returning to user mode, so don't waste time setting
1254 * it perfectly here.
1256 KASSERT((td->td_flags & TDF_BORROWING) == 0,
1257 ("thread with borrowed priority returning to userland"));
1258 if (td->td_priority != td->td_user_pri) {
1260 td->td_priority = td->td_user_pri;
1261 td->td_base_pri = td->td_user_pri;
1267 sched_bind(struct thread *td, int cpu)
1269 struct td_sched *ts;
1271 THREAD_LOCK_ASSERT(td, MA_OWNED);
1272 KASSERT(TD_IS_RUNNING(td),
1273 ("sched_bind: cannot bind non-running thread"));
1277 ts->ts_flags |= TSF_BOUND;
1279 ts->ts_runq = &runq_pcpu[cpu];
1280 if (PCPU_GET(cpuid) == cpu)
1283 mi_switch(SW_VOL, NULL);
1288 sched_unbind(struct thread* td)
1290 THREAD_LOCK_ASSERT(td, MA_OWNED);
1291 td->td_sched->ts_flags &= ~TSF_BOUND;
1295 sched_is_bound(struct thread *td)
1297 THREAD_LOCK_ASSERT(td, MA_OWNED);
1298 return (td->td_sched->ts_flags & TSF_BOUND);
1302 sched_relinquish(struct thread *td)
1305 SCHED_STAT_INC(switch_relinquish);
1306 mi_switch(SW_VOL, NULL);
1313 return (sched_tdcnt);
1317 sched_sizeof_proc(void)
1319 return (sizeof(struct proc));
1323 sched_sizeof_thread(void)
1325 return (sizeof(struct thread) + sizeof(struct td_sched));
1329 sched_pctcpu(struct thread *td)
1331 struct td_sched *ts;
1334 return (ts->ts_pctcpu);
1343 * The actual idle process.
1346 sched_idletd(void *dummy)
1350 mtx_assert(&Giant, MA_NOTOWNED);
1352 while (sched_runnable() == 0)
1355 mtx_lock_spin(&sched_lock);
1356 mi_switch(SW_VOL, NULL);
1357 mtx_unlock_spin(&sched_lock);
1362 * A CPU is entering for the first time or a thread is exiting.
1365 sched_throw(struct thread *td)
1368 * Correct spinlock nesting. The idle thread context that we are
1369 * borrowing was created so that it would start out with a single
1370 * spin lock (sched_lock) held in fork_trampoline(). Since we've
1371 * explicitly acquired locks in this function, the nesting count
1372 * is now 2 rather than 1. Since we are nested, calling
1373 * spinlock_exit() will simply adjust the counts without allowing
1374 * spin lock using code to interrupt us.
1377 mtx_lock_spin(&sched_lock);
1380 lock_profile_release_lock(&sched_lock.lock_object);
1381 MPASS(td->td_lock == &sched_lock);
1383 mtx_assert(&sched_lock, MA_OWNED);
1384 KASSERT(curthread->td_md.md_spinlock_count == 1, ("invalid count"));
1385 PCPU_SET(switchtime, cpu_ticks());
1386 PCPU_SET(switchticks, ticks);
1387 cpu_throw(td, choosethread()); /* doesn't return */
1391 sched_fork_exit(struct thread *td)
1395 * Finish setting up thread glue so that it begins execution in a
1396 * non-nested critical section with sched_lock held but not recursed.
1398 td->td_oncpu = PCPU_GET(cpuid);
1399 sched_lock.mtx_lock = (uintptr_t)td;
1400 lock_profile_obtain_lock_success(&sched_lock.lock_object,
1401 0, 0, __FILE__, __LINE__);
1402 THREAD_LOCK_ASSERT(td, MA_OWNED | MA_NOTRECURSED);
1405 #define KERN_SWITCH_INCLUDE 1
1406 #include "kern/kern_switch.c"