xref: /freebsd/sys/kern/sched_4bsd.c (revision d139ce67c0b39ab6532275f7baff67d220fe8001)
1 /*-
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.
9  *
10  * Redistribution and use in source and binary forms, with or without
11  * modification, are permitted provided that the following conditions
12  * are met:
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.
21  *
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
32  * SUCH DAMAGE.
33  */
34 
35 #include <sys/cdefs.h>
36 __FBSDID("$FreeBSD$");
37 
38 #include "opt_hwpmc_hooks.h"
39 
40 #include <sys/param.h>
41 #include <sys/systm.h>
42 #include <sys/kernel.h>
43 #include <sys/ktr.h>
44 #include <sys/lock.h>
45 #include <sys/kthread.h>
46 #include <sys/mutex.h>
47 #include <sys/proc.h>
48 #include <sys/resourcevar.h>
49 #include <sys/sched.h>
50 #include <sys/smp.h>
51 #include <sys/sysctl.h>
52 #include <sys/sx.h>
53 #include <sys/turnstile.h>
54 #include <sys/umtx.h>
55 #include <machine/pcb.h>
56 #include <machine/smp.h>
57 
58 #ifdef HWPMC_HOOKS
59 #include <sys/pmckern.h>
60 #endif
61 
62 /*
63  * INVERSE_ESTCPU_WEIGHT is only suitable for statclock() frequencies in
64  * the range 100-256 Hz (approximately).
65  */
66 #define	ESTCPULIM(e) \
67     min((e), INVERSE_ESTCPU_WEIGHT * (NICE_WEIGHT * (PRIO_MAX - PRIO_MIN) - \
68     RQ_PPQ) + INVERSE_ESTCPU_WEIGHT - 1)
69 #ifdef SMP
70 #define	INVERSE_ESTCPU_WEIGHT	(8 * smp_cpus)
71 #else
72 #define	INVERSE_ESTCPU_WEIGHT	8	/* 1 / (priorities per estcpu level). */
73 #endif
74 #define	NICE_WEIGHT		1	/* Priorities per nice level. */
75 
76 /*
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
80  */
81 struct td_sched {
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 	struct runq	*ts_runq;	/* runq the thread is currently on */
88 };
89 
90 /* flags kept in td_flags */
91 #define TDF_DIDRUN	TDF_SCHED0	/* thread actually ran. */
92 #define TDF_EXIT	TDF_SCHED1	/* thread is being killed. */
93 #define TDF_BOUND	TDF_SCHED2
94 
95 #define ts_flags	ts_thread->td_flags
96 #define TSF_DIDRUN	TDF_DIDRUN /* thread actually ran. */
97 #define TSF_EXIT	TDF_EXIT /* thread is being killed. */
98 #define TSF_BOUND	TDF_BOUND /* stuck to one CPU */
99 
100 #define SKE_RUNQ_PCPU(ts)						\
101     ((ts)->ts_runq != 0 && (ts)->ts_runq != &runq)
102 
103 static struct td_sched td_sched0;
104 
105 static int	sched_tdcnt;	/* Total runnable threads in the system. */
106 static int	sched_quantum;	/* Roundrobin scheduling quantum in ticks. */
107 #define	SCHED_QUANTUM	(hz / 10)	/* Default sched quantum */
108 
109 static struct callout roundrobin_callout;
110 
111 static void	setup_runqs(void);
112 static void	roundrobin(void *arg);
113 static void	schedcpu(void);
114 static void	schedcpu_thread(void);
115 static void	sched_priority(struct thread *td, u_char prio);
116 static void	sched_setup(void *dummy);
117 static void	maybe_resched(struct thread *td);
118 static void	updatepri(struct thread *td);
119 static void	resetpriority(struct thread *td);
120 static void	resetpriority_thread(struct thread *td);
121 #ifdef SMP
122 static int	forward_wakeup(int  cpunum);
123 #endif
124 
125 static struct kproc_desc sched_kp = {
126         "schedcpu",
127         schedcpu_thread,
128         NULL
129 };
130 SYSINIT(schedcpu, SI_SUB_RUN_SCHEDULER, SI_ORDER_FIRST, kproc_start, &sched_kp)
131 SYSINIT(sched_setup, SI_SUB_RUN_QUEUE, SI_ORDER_FIRST, sched_setup, NULL)
132 
133 /*
134  * Global run queue.
135  */
136 static struct runq runq;
137 
138 #ifdef SMP
139 /*
140  * Per-CPU run queues
141  */
142 static struct runq runq_pcpu[MAXCPU];
143 #endif
144 
145 static void
146 setup_runqs(void)
147 {
148 #ifdef SMP
149 	int i;
150 
151 	for (i = 0; i < MAXCPU; ++i)
152 		runq_init(&runq_pcpu[i]);
153 #endif
154 
155 	runq_init(&runq);
156 }
157 
158 static int
159 sysctl_kern_quantum(SYSCTL_HANDLER_ARGS)
160 {
161 	int error, new_val;
162 
163 	new_val = sched_quantum * tick;
164 	error = sysctl_handle_int(oidp, &new_val, 0, req);
165         if (error != 0 || req->newptr == NULL)
166 		return (error);
167 	if (new_val < tick)
168 		return (EINVAL);
169 	sched_quantum = new_val / tick;
170 	hogticks = 2 * sched_quantum;
171 	return (0);
172 }
173 
174 SYSCTL_NODE(_kern, OID_AUTO, sched, CTLFLAG_RD, 0, "Scheduler");
175 
176 SYSCTL_STRING(_kern_sched, OID_AUTO, name, CTLFLAG_RD, "4BSD", 0,
177     "Scheduler name");
178 
179 SYSCTL_PROC(_kern_sched, OID_AUTO, quantum, CTLTYPE_INT | CTLFLAG_RW,
180     0, sizeof sched_quantum, sysctl_kern_quantum, "I",
181     "Roundrobin scheduling quantum in microseconds");
182 
183 #ifdef SMP
184 /* Enable forwarding of wakeups to all other cpus */
185 SYSCTL_NODE(_kern_sched, OID_AUTO, ipiwakeup, CTLFLAG_RD, NULL, "Kernel SMP");
186 
187 static int forward_wakeup_enabled = 1;
188 SYSCTL_INT(_kern_sched_ipiwakeup, OID_AUTO, enabled, CTLFLAG_RW,
189 	   &forward_wakeup_enabled, 0,
190 	   "Forwarding of wakeup to idle CPUs");
191 
192 static int forward_wakeups_requested = 0;
193 SYSCTL_INT(_kern_sched_ipiwakeup, OID_AUTO, requested, CTLFLAG_RD,
194 	   &forward_wakeups_requested, 0,
195 	   "Requests for Forwarding of wakeup to idle CPUs");
196 
197 static int forward_wakeups_delivered = 0;
198 SYSCTL_INT(_kern_sched_ipiwakeup, OID_AUTO, delivered, CTLFLAG_RD,
199 	   &forward_wakeups_delivered, 0,
200 	   "Completed Forwarding of wakeup to idle CPUs");
201 
202 static int forward_wakeup_use_mask = 1;
203 SYSCTL_INT(_kern_sched_ipiwakeup, OID_AUTO, usemask, CTLFLAG_RW,
204 	   &forward_wakeup_use_mask, 0,
205 	   "Use the mask of idle cpus");
206 
207 static int forward_wakeup_use_loop = 0;
208 SYSCTL_INT(_kern_sched_ipiwakeup, OID_AUTO, useloop, CTLFLAG_RW,
209 	   &forward_wakeup_use_loop, 0,
210 	   "Use a loop to find idle cpus");
211 
212 static int forward_wakeup_use_single = 0;
213 SYSCTL_INT(_kern_sched_ipiwakeup, OID_AUTO, onecpu, CTLFLAG_RW,
214 	   &forward_wakeup_use_single, 0,
215 	   "Only signal one idle cpu");
216 
217 static int forward_wakeup_use_htt = 0;
218 SYSCTL_INT(_kern_sched_ipiwakeup, OID_AUTO, htt2, CTLFLAG_RW,
219 	   &forward_wakeup_use_htt, 0,
220 	   "account for htt");
221 
222 #endif
223 #if 0
224 static int sched_followon = 0;
225 SYSCTL_INT(_kern_sched, OID_AUTO, followon, CTLFLAG_RW,
226 	   &sched_followon, 0,
227 	   "allow threads to share a quantum");
228 #endif
229 
230 static __inline void
231 sched_load_add(void)
232 {
233 	sched_tdcnt++;
234 	CTR1(KTR_SCHED, "global load: %d", sched_tdcnt);
235 }
236 
237 static __inline void
238 sched_load_rem(void)
239 {
240 	sched_tdcnt--;
241 	CTR1(KTR_SCHED, "global load: %d", sched_tdcnt);
242 }
243 /*
244  * Arrange to reschedule if necessary, taking the priorities and
245  * schedulers into account.
246  */
247 static void
248 maybe_resched(struct thread *td)
249 {
250 
251 	mtx_assert(&sched_lock, MA_OWNED);
252 	if (td->td_priority < curthread->td_priority)
253 		curthread->td_flags |= TDF_NEEDRESCHED;
254 }
255 
256 /*
257  * Force switch among equal priority processes every 100ms.
258  * We don't actually need to force a context switch of the current process.
259  * The act of firing the event triggers a context switch to softclock() and
260  * then switching back out again which is equivalent to a preemption, thus
261  * no further work is needed on the local CPU.
262  */
263 /* ARGSUSED */
264 static void
265 roundrobin(void *arg)
266 {
267 
268 #ifdef SMP
269 	mtx_lock_spin(&sched_lock);
270 	forward_roundrobin();
271 	mtx_unlock_spin(&sched_lock);
272 #endif
273 
274 	callout_reset(&roundrobin_callout, sched_quantum, roundrobin, NULL);
275 }
276 
277 /*
278  * Constants for digital decay and forget:
279  *	90% of (td_estcpu) usage in 5 * loadav time
280  *	95% of (ts_pctcpu) usage in 60 seconds (load insensitive)
281  *          Note that, as ps(1) mentions, this can let percentages
282  *          total over 100% (I've seen 137.9% for 3 processes).
283  *
284  * Note that schedclock() updates td_estcpu and p_cpticks asynchronously.
285  *
286  * We wish to decay away 90% of td_estcpu in (5 * loadavg) seconds.
287  * That is, the system wants to compute a value of decay such
288  * that the following for loop:
289  * 	for (i = 0; i < (5 * loadavg); i++)
290  * 		td_estcpu *= decay;
291  * will compute
292  * 	td_estcpu *= 0.1;
293  * for all values of loadavg:
294  *
295  * Mathematically this loop can be expressed by saying:
296  * 	decay ** (5 * loadavg) ~= .1
297  *
298  * The system computes decay as:
299  * 	decay = (2 * loadavg) / (2 * loadavg + 1)
300  *
301  * We wish to prove that the system's computation of decay
302  * will always fulfill the equation:
303  * 	decay ** (5 * loadavg) ~= .1
304  *
305  * If we compute b as:
306  * 	b = 2 * loadavg
307  * then
308  * 	decay = b / (b + 1)
309  *
310  * We now need to prove two things:
311  *	1) Given factor ** (5 * loadavg) ~= .1, prove factor == b/(b+1)
312  *	2) Given b/(b+1) ** power ~= .1, prove power == (5 * loadavg)
313  *
314  * Facts:
315  *         For x close to zero, exp(x) =~ 1 + x, since
316  *              exp(x) = 0! + x**1/1! + x**2/2! + ... .
317  *              therefore exp(-1/b) =~ 1 - (1/b) = (b-1)/b.
318  *         For x close to zero, ln(1+x) =~ x, since
319  *              ln(1+x) = x - x**2/2 + x**3/3 - ...     -1 < x < 1
320  *              therefore ln(b/(b+1)) = ln(1 - 1/(b+1)) =~ -1/(b+1).
321  *         ln(.1) =~ -2.30
322  *
323  * Proof of (1):
324  *    Solve (factor)**(power) =~ .1 given power (5*loadav):
325  *	solving for factor,
326  *      ln(factor) =~ (-2.30/5*loadav), or
327  *      factor =~ exp(-1/((5/2.30)*loadav)) =~ exp(-1/(2*loadav)) =
328  *          exp(-1/b) =~ (b-1)/b =~ b/(b+1).                    QED
329  *
330  * Proof of (2):
331  *    Solve (factor)**(power) =~ .1 given factor == (b/(b+1)):
332  *	solving for power,
333  *      power*ln(b/(b+1)) =~ -2.30, or
334  *      power =~ 2.3 * (b + 1) = 4.6*loadav + 2.3 =~ 5*loadav.  QED
335  *
336  * Actual power values for the implemented algorithm are as follows:
337  *      loadav: 1       2       3       4
338  *      power:  5.68    10.32   14.94   19.55
339  */
340 
341 /* calculations for digital decay to forget 90% of usage in 5*loadav sec */
342 #define	loadfactor(loadav)	(2 * (loadav))
343 #define	decay_cpu(loadfac, cpu)	(((loadfac) * (cpu)) / ((loadfac) + FSCALE))
344 
345 /* decay 95% of `ts_pctcpu' in 60 seconds; see CCPU_SHIFT before changing */
346 static fixpt_t	ccpu = 0.95122942450071400909 * FSCALE;	/* exp(-1/20) */
347 SYSCTL_INT(_kern, OID_AUTO, ccpu, CTLFLAG_RD, &ccpu, 0, "");
348 
349 /*
350  * If `ccpu' is not equal to `exp(-1/20)' and you still want to use the
351  * faster/more-accurate formula, you'll have to estimate CCPU_SHIFT below
352  * and possibly adjust FSHIFT in "param.h" so that (FSHIFT >= CCPU_SHIFT).
353  *
354  * To estimate CCPU_SHIFT for exp(-1/20), the following formula was used:
355  *	1 - exp(-1/20) ~= 0.0487 ~= 0.0488 == 1 (fixed pt, *11* bits).
356  *
357  * If you don't want to bother with the faster/more-accurate formula, you
358  * can set CCPU_SHIFT to (FSHIFT + 1) which will use a slower/less-accurate
359  * (more general) method of calculating the %age of CPU used by a process.
360  */
361 #define	CCPU_SHIFT	11
362 
363 /*
364  * Recompute process priorities, every hz ticks.
365  * MP-safe, called without the Giant mutex.
366  */
367 /* ARGSUSED */
368 static void
369 schedcpu(void)
370 {
371 	register fixpt_t loadfac = loadfactor(averunnable.ldavg[0]);
372 	struct thread *td;
373 	struct proc *p;
374 	struct td_sched *ts;
375 	int awake, realstathz;
376 
377 	realstathz = stathz ? stathz : hz;
378 	sx_slock(&allproc_lock);
379 	FOREACH_PROC_IN_SYSTEM(p) {
380 		/*
381 		 * Prevent state changes and protect run queue.
382 		 */
383 		mtx_lock_spin(&sched_lock);
384 		/*
385 		 * Increment time in/out of memory.  We ignore overflow; with
386 		 * 16-bit int's (remember them?) overflow takes 45 days.
387 		 */
388 		p->p_swtime++;
389 		FOREACH_THREAD_IN_PROC(p, td) {
390 			awake = 0;
391 			ts = td->td_sched;
392 			/*
393 			 * Increment sleep time (if sleeping).  We
394 			 * ignore overflow, as above.
395 			 */
396 			/*
397 			 * The td_sched slptimes are not touched in wakeup
398 			 * because the thread may not HAVE everything in
399 			 * memory? XXX I think this is out of date.
400 			 */
401 			if (TD_ON_RUNQ(td)) {
402 				awake = 1;
403 				ts->ts_flags &= ~TSF_DIDRUN;
404 			} else if (TD_IS_RUNNING(td)) {
405 				awake = 1;
406 				/* Do not clear TSF_DIDRUN */
407 			} else if (ts->ts_flags & TSF_DIDRUN) {
408 				awake = 1;
409 				ts->ts_flags &= ~TSF_DIDRUN;
410 			}
411 
412 			/*
413 			 * ts_pctcpu is only for ps and ttyinfo().
414 			 * Do it per td_sched, and add them up at the end?
415 			 * XXXKSE
416 			 */
417 			ts->ts_pctcpu = (ts->ts_pctcpu * ccpu) >> FSHIFT;
418 			/*
419 			 * If the td_sched has been idle the entire second,
420 			 * stop recalculating its priority until
421 			 * it wakes up.
422 			 */
423 			if (ts->ts_cpticks != 0) {
424 #if	(FSHIFT >= CCPU_SHIFT)
425 				ts->ts_pctcpu += (realstathz == 100)
426 				    ? ((fixpt_t) ts->ts_cpticks) <<
427 				    (FSHIFT - CCPU_SHIFT) :
428 				    100 * (((fixpt_t) ts->ts_cpticks)
429 				    << (FSHIFT - CCPU_SHIFT)) / realstathz;
430 #else
431 				ts->ts_pctcpu += ((FSCALE - ccpu) *
432 				    (ts->ts_cpticks *
433 				    FSCALE / realstathz)) >> FSHIFT;
434 #endif
435 				ts->ts_cpticks = 0;
436 			}
437 			/*
438 			 * If there are ANY running threads in this process,
439 			 * then don't count it as sleeping.
440 XXX  this is broken
441 
442 			 */
443 			if (awake) {
444 				if (td->td_slptime > 1) {
445 					/*
446 					 * In an ideal world, this should not
447 					 * happen, because whoever woke us
448 					 * up from the long sleep should have
449 					 * unwound the slptime and reset our
450 					 * priority before we run at the stale
451 					 * priority.  Should KASSERT at some
452 					 * point when all the cases are fixed.
453 					 */
454 					updatepri(td);
455 				}
456 				td->td_slptime = 0;
457 			} else
458 				td->td_slptime++;
459 			if (td->td_slptime > 1)
460 				continue;
461 			td->td_estcpu = decay_cpu(loadfac, td->td_estcpu);
462 		      	resetpriority(td);
463 			resetpriority_thread(td);
464 		} /* end of thread loop */
465 		mtx_unlock_spin(&sched_lock);
466 	} /* end of process loop */
467 	sx_sunlock(&allproc_lock);
468 }
469 
470 /*
471  * Main loop for a kthread that executes schedcpu once a second.
472  */
473 static void
474 schedcpu_thread(void)
475 {
476 	int nowake;
477 
478 	for (;;) {
479 		schedcpu();
480 		tsleep(&nowake, 0, "-", hz);
481 	}
482 }
483 
484 /*
485  * Recalculate the priority of a process after it has slept for a while.
486  * For all load averages >= 1 and max td_estcpu of 255, sleeping for at
487  * least six times the loadfactor will decay td_estcpu to zero.
488  */
489 static void
490 updatepri(struct thread *td)
491 {
492 	register fixpt_t loadfac;
493 	register unsigned int newcpu;
494 
495 	loadfac = loadfactor(averunnable.ldavg[0]);
496 	if (td->td_slptime > 5 * loadfac)
497 		td->td_estcpu = 0;
498 	else {
499 		newcpu = td->td_estcpu;
500 		td->td_slptime--;	/* was incremented in schedcpu() */
501 		while (newcpu && --td->td_slptime)
502 			newcpu = decay_cpu(loadfac, newcpu);
503 		td->td_estcpu = newcpu;
504 	}
505 }
506 
507 /*
508  * Compute the priority of a process when running in user mode.
509  * Arrange to reschedule if the resulting priority is better
510  * than that of the current process.
511  */
512 static void
513 resetpriority(struct thread *td)
514 {
515 	register unsigned int newpriority;
516 
517 	if (td->td_pri_class == PRI_TIMESHARE) {
518 		newpriority = PUSER + td->td_estcpu / INVERSE_ESTCPU_WEIGHT +
519 		    NICE_WEIGHT * (td->td_proc->p_nice - PRIO_MIN);
520 		newpriority = min(max(newpriority, PRI_MIN_TIMESHARE),
521 		    PRI_MAX_TIMESHARE);
522 		sched_user_prio(td, newpriority);
523 	}
524 }
525 
526 /*
527  * Update the thread's priority when the associated process's user
528  * priority changes.
529  */
530 static void
531 resetpriority_thread(struct thread *td)
532 {
533 
534 	/* Only change threads with a time sharing user priority. */
535 	if (td->td_priority < PRI_MIN_TIMESHARE ||
536 	    td->td_priority > PRI_MAX_TIMESHARE)
537 		return;
538 
539 	/* XXX the whole needresched thing is broken, but not silly. */
540 	maybe_resched(td);
541 
542 	sched_prio(td, td->td_user_pri);
543 }
544 
545 /* ARGSUSED */
546 static void
547 sched_setup(void *dummy)
548 {
549 	setup_runqs();
550 
551 	if (sched_quantum == 0)
552 		sched_quantum = SCHED_QUANTUM;
553 	hogticks = 2 * sched_quantum;
554 
555 	callout_init(&roundrobin_callout, CALLOUT_MPSAFE);
556 
557 	/* Kick off timeout driven events by calling first time. */
558 	roundrobin(NULL);
559 
560 	/* Account for thread0. */
561 	sched_load_add();
562 }
563 
564 /* External interfaces start here */
565 /*
566  * Very early in the boot some setup of scheduler-specific
567  * parts of proc0 and of some scheduler resources needs to be done.
568  * Called from:
569  *  proc0_init()
570  */
571 void
572 schedinit(void)
573 {
574 	/*
575 	 * Set up the scheduler specific parts of proc0.
576 	 */
577 	proc0.p_sched = NULL; /* XXX */
578 	thread0.td_sched = &td_sched0;
579 	td_sched0.ts_thread = &thread0;
580 }
581 
582 int
583 sched_runnable(void)
584 {
585 #ifdef SMP
586 	return runq_check(&runq) + runq_check(&runq_pcpu[PCPU_GET(cpuid)]);
587 #else
588 	return runq_check(&runq);
589 #endif
590 }
591 
592 int
593 sched_rr_interval(void)
594 {
595 	if (sched_quantum == 0)
596 		sched_quantum = SCHED_QUANTUM;
597 	return (sched_quantum);
598 }
599 
600 /*
601  * We adjust the priority of the current process.  The priority of
602  * a process gets worse as it accumulates CPU time.  The cpu usage
603  * estimator (td_estcpu) is increased here.  resetpriority() will
604  * compute a different priority each time td_estcpu increases by
605  * INVERSE_ESTCPU_WEIGHT
606  * (until MAXPRI is reached).  The cpu usage estimator ramps up
607  * quite quickly when the process is running (linearly), and decays
608  * away exponentially, at a rate which is proportionally slower when
609  * the system is busy.  The basic principle is that the system will
610  * 90% forget that the process used a lot of CPU time in 5 * loadav
611  * seconds.  This causes the system to favor processes which haven't
612  * run much recently, and to round-robin among other processes.
613  */
614 void
615 sched_clock(struct thread *td)
616 {
617 	struct td_sched *ts;
618 
619 	mtx_assert(&sched_lock, MA_OWNED);
620 	ts = td->td_sched;
621 
622 	ts->ts_cpticks++;
623 	td->td_estcpu = ESTCPULIM(td->td_estcpu + 1);
624 	if ((td->td_estcpu % INVERSE_ESTCPU_WEIGHT) == 0) {
625 		resetpriority(td);
626 		resetpriority_thread(td);
627 	}
628 }
629 
630 /*
631  * charge childs scheduling cpu usage to parent.
632  */
633 void
634 sched_exit(struct proc *p, struct thread *td)
635 {
636 
637 	CTR3(KTR_SCHED, "sched_exit: %p(%s) prio %d",
638 	    td, td->td_proc->p_comm, td->td_priority);
639 
640 	sched_exit_thread(FIRST_THREAD_IN_PROC(p), td);
641 }
642 
643 void
644 sched_exit_thread(struct thread *td, struct thread *child)
645 {
646 	struct proc *childproc = child->td_proc;
647 
648 	CTR3(KTR_SCHED, "sched_exit_thread: %p(%s) prio %d",
649 	    child, childproc->p_comm, child->td_priority);
650 	td->td_estcpu = ESTCPULIM(td->td_estcpu + child->td_estcpu);
651 	childproc->p_estcpu = ESTCPULIM(childproc->p_estcpu +
652 		child->td_estcpu);
653 	if ((child->td_proc->p_flag & P_NOLOAD) == 0)
654 		sched_load_rem();
655 }
656 
657 void
658 sched_fork(struct thread *td, struct thread *childtd)
659 {
660 	sched_fork_thread(td, childtd);
661 }
662 
663 void
664 sched_fork_thread(struct thread *td, struct thread *childtd)
665 {
666 	childtd->td_estcpu = td->td_estcpu;
667 	sched_newthread(childtd);
668 }
669 
670 void
671 sched_nice(struct proc *p, int nice)
672 {
673 	struct thread *td;
674 
675 	PROC_LOCK_ASSERT(p, MA_OWNED);
676 	mtx_assert(&sched_lock, MA_OWNED);
677 	p->p_nice = nice;
678 	FOREACH_THREAD_IN_PROC(p, td) {
679 		resetpriority(td);
680 		resetpriority_thread(td);
681 	}
682 }
683 
684 void
685 sched_class(struct thread *td, int class)
686 {
687 	mtx_assert(&sched_lock, MA_OWNED);
688 	td->td_pri_class = class;
689 }
690 
691 /*
692  * Adjust the priority of a thread.
693  */
694 static void
695 sched_priority(struct thread *td, u_char prio)
696 {
697 	CTR6(KTR_SCHED, "sched_prio: %p(%s) prio %d newprio %d by %p(%s)",
698 	    td, td->td_proc->p_comm, td->td_priority, prio, curthread,
699 	    curthread->td_proc->p_comm);
700 
701 	mtx_assert(&sched_lock, MA_OWNED);
702 	if (td->td_priority == prio)
703 		return;
704 	td->td_priority = prio;
705 	if (TD_ON_RUNQ(td) &&
706 	    td->td_sched->ts_rqindex != (prio / RQ_PPQ)) {
707 		sched_rem(td);
708 		sched_add(td, SRQ_BORING);
709 	}
710 }
711 
712 /*
713  * Update a thread's priority when it is lent another thread's
714  * priority.
715  */
716 void
717 sched_lend_prio(struct thread *td, u_char prio)
718 {
719 
720 	td->td_flags |= TDF_BORROWING;
721 	sched_priority(td, prio);
722 }
723 
724 /*
725  * Restore a thread's priority when priority propagation is
726  * over.  The prio argument is the minimum priority the thread
727  * needs to have to satisfy other possible priority lending
728  * requests.  If the thread's regulary priority is less
729  * important than prio the thread will keep a priority boost
730  * of prio.
731  */
732 void
733 sched_unlend_prio(struct thread *td, u_char prio)
734 {
735 	u_char base_pri;
736 
737 	if (td->td_base_pri >= PRI_MIN_TIMESHARE &&
738 	    td->td_base_pri <= PRI_MAX_TIMESHARE)
739 		base_pri = td->td_user_pri;
740 	else
741 		base_pri = td->td_base_pri;
742 	if (prio >= base_pri) {
743 		td->td_flags &= ~TDF_BORROWING;
744 		sched_prio(td, base_pri);
745 	} else
746 		sched_lend_prio(td, prio);
747 }
748 
749 void
750 sched_prio(struct thread *td, u_char prio)
751 {
752 	u_char oldprio;
753 
754 	/* First, update the base priority. */
755 	td->td_base_pri = prio;
756 
757 	/*
758 	 * If the thread is borrowing another thread's priority, don't ever
759 	 * lower the priority.
760 	 */
761 	if (td->td_flags & TDF_BORROWING && td->td_priority < prio)
762 		return;
763 
764 	/* Change the real priority. */
765 	oldprio = td->td_priority;
766 	sched_priority(td, prio);
767 
768 	/*
769 	 * If the thread is on a turnstile, then let the turnstile update
770 	 * its state.
771 	 */
772 	if (TD_ON_LOCK(td) && oldprio != prio)
773 		turnstile_adjust(td, oldprio);
774 }
775 
776 void
777 sched_user_prio(struct thread *td, u_char prio)
778 {
779 	u_char oldprio;
780 
781 	td->td_base_user_pri = prio;
782 	if (td->td_flags & TDF_UBORROWING && td->td_user_pri <= prio)
783 		return;
784 	oldprio = td->td_user_pri;
785 	td->td_user_pri = prio;
786 
787 	if (TD_ON_UPILOCK(td) && oldprio != prio)
788 		umtx_pi_adjust(td, oldprio);
789 }
790 
791 void
792 sched_lend_user_prio(struct thread *td, u_char prio)
793 {
794 	u_char oldprio;
795 
796 	td->td_flags |= TDF_UBORROWING;
797 
798 	oldprio = td->td_user_pri;
799 	td->td_user_pri = prio;
800 
801 	if (TD_ON_UPILOCK(td) && oldprio != prio)
802 		umtx_pi_adjust(td, oldprio);
803 }
804 
805 void
806 sched_unlend_user_prio(struct thread *td, u_char prio)
807 {
808 	u_char base_pri;
809 
810 	base_pri = td->td_base_user_pri;
811 	if (prio >= base_pri) {
812 		td->td_flags &= ~TDF_UBORROWING;
813 		sched_user_prio(td, base_pri);
814 	} else
815 		sched_lend_user_prio(td, prio);
816 }
817 
818 void
819 sched_sleep(struct thread *td)
820 {
821 
822 	mtx_assert(&sched_lock, MA_OWNED);
823 	td->td_slptime = 0;
824 }
825 
826 void
827 sched_switch(struct thread *td, struct thread *newtd, int flags)
828 {
829 	struct td_sched *ts;
830 	struct proc *p;
831 
832 	ts = td->td_sched;
833 	p = td->td_proc;
834 
835 	mtx_assert(&sched_lock, MA_OWNED);
836 
837 	if ((p->p_flag & P_NOLOAD) == 0)
838 		sched_load_rem();
839 #if 0
840 	/*
841 	 * We are volunteering to switch out so we get to nominate
842 	 * a successor for the rest of our quantum
843 	 * First try another thread in our process
844 	 *
845 	 * this is too expensive to do without per process run queues
846 	 * so skip it for now.
847 	 * XXX keep this comment as a marker.
848 	 */
849 	if (sched_followon &&
850 	    (p->p_flag & P_HADTHREADS) &&
851 	    (flags & SW_VOL) &&
852 	    newtd == NULL)
853 		newtd = mumble();
854 #endif
855 
856 	if (newtd)
857 		newtd->td_flags |= (td->td_flags & TDF_NEEDRESCHED);
858 
859 	td->td_lastcpu = td->td_oncpu;
860 	td->td_flags &= ~TDF_NEEDRESCHED;
861 	td->td_owepreempt = 0;
862 	td->td_oncpu = NOCPU;
863 	/*
864 	 * At the last moment, if this thread is still marked RUNNING,
865 	 * then put it back on the run queue as it has not been suspended
866 	 * or stopped or any thing else similar.  We never put the idle
867 	 * threads on the run queue, however.
868 	 */
869 	if (td->td_flags & TDF_IDLETD) {
870 		TD_SET_CAN_RUN(td);
871 #ifdef SMP
872 		idle_cpus_mask &= ~PCPU_GET(cpumask);
873 #endif
874 	} else {
875 		if (TD_IS_RUNNING(td)) {
876 			/* Put us back on the run queue. */
877 			sched_add(td, (flags & SW_PREEMPT) ?
878 			    SRQ_OURSELF|SRQ_YIELDING|SRQ_PREEMPTED :
879 			    SRQ_OURSELF|SRQ_YIELDING);
880 		}
881 	}
882 	if (newtd) {
883 		/*
884 		 * The thread we are about to run needs to be counted
885 		 * as if it had been added to the run queue and selected.
886 		 * It came from:
887 		 * * A preemption
888 		 * * An upcall
889 		 * * A followon
890 		 */
891 		KASSERT((newtd->td_inhibitors == 0),
892 			("trying to run inhibited thread"));
893 		newtd->td_sched->ts_flags |= TSF_DIDRUN;
894         	TD_SET_RUNNING(newtd);
895 		if ((newtd->td_proc->p_flag & P_NOLOAD) == 0)
896 			sched_load_add();
897 	} else {
898 		newtd = choosethread();
899 	}
900 
901 	if (td != newtd) {
902 #ifdef	HWPMC_HOOKS
903 		if (PMC_PROC_IS_USING_PMCS(td->td_proc))
904 			PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_OUT);
905 #endif
906 
907                 /* I feel sleepy */
908 		cpu_switch(td, newtd);
909 		/*
910 		 * Where am I?  What year is it?
911 		 * We are in the same thread that went to sleep above,
912 		 * but any amount of time may have passed. All out context
913 		 * will still be available as will local variables.
914 		 * PCPU values however may have changed as we may have
915 		 * changed CPU so don't trust cached values of them.
916 		 * New threads will go to fork_exit() instead of here
917 		 * so if you change things here you may need to change
918 		 * things there too.
919 		 * If the thread above was exiting it will never wake
920 		 * up again here, so either it has saved everything it
921 		 * needed to, or the thread_wait() or wait() will
922 		 * need to reap it.
923 		 */
924 #ifdef	HWPMC_HOOKS
925 		if (PMC_PROC_IS_USING_PMCS(td->td_proc))
926 			PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_IN);
927 #endif
928 	}
929 
930 #ifdef SMP
931 	if (td->td_flags & TDF_IDLETD)
932 		idle_cpus_mask |= PCPU_GET(cpumask);
933 #endif
934 	sched_lock.mtx_lock = (uintptr_t)td;
935 	td->td_oncpu = PCPU_GET(cpuid);
936 }
937 
938 void
939 sched_wakeup(struct thread *td)
940 {
941 	mtx_assert(&sched_lock, MA_OWNED);
942 	if (td->td_slptime > 1) {
943 		updatepri(td);
944 		resetpriority(td);
945 	}
946 	td->td_slptime = 0;
947 	sched_add(td, SRQ_BORING);
948 }
949 
950 #ifdef SMP
951 /* enable HTT_2 if you have a 2-way HTT cpu.*/
952 static int
953 forward_wakeup(int  cpunum)
954 {
955 	cpumask_t map, me, dontuse;
956 	cpumask_t map2;
957 	struct pcpu *pc;
958 	cpumask_t id, map3;
959 
960 	mtx_assert(&sched_lock, MA_OWNED);
961 
962 	CTR0(KTR_RUNQ, "forward_wakeup()");
963 
964 	if ((!forward_wakeup_enabled) ||
965 	     (forward_wakeup_use_mask == 0 && forward_wakeup_use_loop == 0))
966 		return (0);
967 	if (!smp_started || cold || panicstr)
968 		return (0);
969 
970 	forward_wakeups_requested++;
971 
972 /*
973  * check the idle mask we received against what we calculated before
974  * in the old version.
975  */
976 	me = PCPU_GET(cpumask);
977 	/*
978 	 * don't bother if we should be doing it ourself..
979 	 */
980 	if ((me & idle_cpus_mask) && (cpunum == NOCPU || me == (1 << cpunum)))
981 		return (0);
982 
983 	dontuse = me | stopped_cpus | hlt_cpus_mask;
984 	map3 = 0;
985 	if (forward_wakeup_use_loop) {
986 		SLIST_FOREACH(pc, &cpuhead, pc_allcpu) {
987 			id = pc->pc_cpumask;
988 			if ( (id & dontuse) == 0 &&
989 			    pc->pc_curthread == pc->pc_idlethread) {
990 				map3 |= id;
991 			}
992 		}
993 	}
994 
995 	if (forward_wakeup_use_mask) {
996 		map = 0;
997 		map = idle_cpus_mask & ~dontuse;
998 
999 		/* If they are both on, compare and use loop if different */
1000 		if (forward_wakeup_use_loop) {
1001 			if (map != map3) {
1002 				printf("map (%02X) != map3 (%02X)\n",
1003 						map, map3);
1004 				map = map3;
1005 			}
1006 		}
1007 	} else {
1008 		map = map3;
1009 	}
1010 	/* If we only allow a specific CPU, then mask off all the others */
1011 	if (cpunum != NOCPU) {
1012 		KASSERT((cpunum <= mp_maxcpus),("forward_wakeup: bad cpunum."));
1013 		map &= (1 << cpunum);
1014 	} else {
1015 		/* Try choose an idle die. */
1016 		if (forward_wakeup_use_htt) {
1017 			map2 =  (map & (map >> 1)) & 0x5555;
1018 			if (map2) {
1019 				map = map2;
1020 			}
1021 		}
1022 
1023 		/* set only one bit */
1024 		if (forward_wakeup_use_single) {
1025 			map = map & ((~map) + 1);
1026 		}
1027 	}
1028 	if (map) {
1029 		forward_wakeups_delivered++;
1030 		ipi_selected(map, IPI_AST);
1031 		return (1);
1032 	}
1033 	if (cpunum == NOCPU)
1034 		printf("forward_wakeup: Idle processor not found\n");
1035 	return (0);
1036 }
1037 #endif
1038 
1039 #ifdef SMP
1040 static void kick_other_cpu(int pri,int cpuid);
1041 
1042 static void
1043 kick_other_cpu(int pri,int cpuid)
1044 {
1045 	struct pcpu * pcpu = pcpu_find(cpuid);
1046 	int cpri = pcpu->pc_curthread->td_priority;
1047 
1048 	if (idle_cpus_mask & pcpu->pc_cpumask) {
1049 		forward_wakeups_delivered++;
1050 		ipi_selected(pcpu->pc_cpumask, IPI_AST);
1051 		return;
1052 	}
1053 
1054 	if (pri >= cpri)
1055 		return;
1056 
1057 #if defined(IPI_PREEMPTION) && defined(PREEMPTION)
1058 #if !defined(FULL_PREEMPTION)
1059 	if (pri <= PRI_MAX_ITHD)
1060 #endif /* ! FULL_PREEMPTION */
1061 	{
1062 		ipi_selected(pcpu->pc_cpumask, IPI_PREEMPT);
1063 		return;
1064 	}
1065 #endif /* defined(IPI_PREEMPTION) && defined(PREEMPTION) */
1066 
1067 	pcpu->pc_curthread->td_flags |= TDF_NEEDRESCHED;
1068 	ipi_selected( pcpu->pc_cpumask , IPI_AST);
1069 	return;
1070 }
1071 #endif /* SMP */
1072 
1073 void
1074 sched_add(struct thread *td, int flags)
1075 #ifdef SMP
1076 {
1077 	struct td_sched *ts;
1078 	int forwarded = 0;
1079 	int cpu;
1080 	int single_cpu = 0;
1081 
1082 	ts = td->td_sched;
1083 	mtx_assert(&sched_lock, MA_OWNED);
1084 	KASSERT((td->td_inhibitors == 0),
1085 	    ("sched_add: trying to run inhibited thread"));
1086 	KASSERT((TD_CAN_RUN(td) || TD_IS_RUNNING(td)),
1087 	    ("sched_add: bad thread state"));
1088 	KASSERT(td->td_proc->p_sflag & PS_INMEM,
1089 	    ("sched_add: process swapped out"));
1090 	CTR5(KTR_SCHED, "sched_add: %p(%s) prio %d by %p(%s)",
1091 	    td, td->td_proc->p_comm, td->td_priority, curthread,
1092 	    curthread->td_proc->p_comm);
1093 	TD_SET_RUNQ(td);
1094 
1095 	if (td->td_pinned != 0) {
1096 		cpu = td->td_lastcpu;
1097 		ts->ts_runq = &runq_pcpu[cpu];
1098 		single_cpu = 1;
1099 		CTR3(KTR_RUNQ,
1100 		    "sched_add: Put td_sched:%p(td:%p) on cpu%d runq", ts, td, cpu);
1101 	} else if ((ts)->ts_flags & TSF_BOUND) {
1102 		/* Find CPU from bound runq */
1103 		KASSERT(SKE_RUNQ_PCPU(ts),("sched_add: bound td_sched not on cpu runq"));
1104 		cpu = ts->ts_runq - &runq_pcpu[0];
1105 		single_cpu = 1;
1106 		CTR3(KTR_RUNQ,
1107 		    "sched_add: Put td_sched:%p(td:%p) on cpu%d runq", ts, td, cpu);
1108 	} else {
1109 		CTR2(KTR_RUNQ,
1110 		    "sched_add: adding td_sched:%p (td:%p) to gbl runq", ts, td);
1111 		cpu = NOCPU;
1112 		ts->ts_runq = &runq;
1113 	}
1114 
1115 	if (single_cpu && (cpu != PCPU_GET(cpuid))) {
1116 	        kick_other_cpu(td->td_priority,cpu);
1117 	} else {
1118 
1119 		if (!single_cpu) {
1120 			cpumask_t me = PCPU_GET(cpumask);
1121 			int idle = idle_cpus_mask & me;
1122 
1123 			if (!idle && ((flags & SRQ_INTR) == 0) &&
1124 			    (idle_cpus_mask & ~(hlt_cpus_mask | me)))
1125 				forwarded = forward_wakeup(cpu);
1126 		}
1127 
1128 		if (!forwarded) {
1129 			if ((flags & SRQ_YIELDING) == 0 && maybe_preempt(td))
1130 				return;
1131 			else
1132 				maybe_resched(td);
1133 		}
1134 	}
1135 
1136 	if ((td->td_proc->p_flag & P_NOLOAD) == 0)
1137 		sched_load_add();
1138 	runq_add(ts->ts_runq, ts, flags);
1139 }
1140 #else /* SMP */
1141 {
1142 	struct td_sched *ts;
1143 	ts = td->td_sched;
1144 	mtx_assert(&sched_lock, MA_OWNED);
1145 	KASSERT((td->td_inhibitors == 0),
1146 	    ("sched_add: trying to run inhibited thread"));
1147 	KASSERT((TD_CAN_RUN(td) || TD_IS_RUNNING(td)),
1148 	    ("sched_add: bad thread state"));
1149 	KASSERT(td->td_proc->p_sflag & PS_INMEM,
1150 	    ("sched_add: process swapped out"));
1151 	CTR5(KTR_SCHED, "sched_add: %p(%s) prio %d by %p(%s)",
1152 	    td, td->td_proc->p_comm, td->td_priority, curthread,
1153 	    curthread->td_proc->p_comm);
1154 	TD_SET_RUNQ(td);
1155 	CTR2(KTR_RUNQ, "sched_add: adding td_sched:%p (td:%p) to runq", ts, td);
1156 	ts->ts_runq = &runq;
1157 
1158 	/*
1159 	 * If we are yielding (on the way out anyhow)
1160 	 * or the thread being saved is US,
1161 	 * then don't try be smart about preemption
1162 	 * or kicking off another CPU
1163 	 * as it won't help and may hinder.
1164 	 * In the YIEDLING case, we are about to run whoever is
1165 	 * being put in the queue anyhow, and in the
1166 	 * OURSELF case, we are puting ourself on the run queue
1167 	 * which also only happens when we are about to yield.
1168 	 */
1169 	if((flags & SRQ_YIELDING) == 0) {
1170 		if (maybe_preempt(td))
1171 			return;
1172 	}
1173 	if ((td->td_proc->p_flag & P_NOLOAD) == 0)
1174 		sched_load_add();
1175 	runq_add(ts->ts_runq, ts, flags);
1176 	maybe_resched(td);
1177 }
1178 #endif /* SMP */
1179 
1180 void
1181 sched_rem(struct thread *td)
1182 {
1183 	struct td_sched *ts;
1184 
1185 	ts = td->td_sched;
1186 	KASSERT(td->td_proc->p_sflag & PS_INMEM,
1187 	    ("sched_rem: process swapped out"));
1188 	KASSERT(TD_ON_RUNQ(td),
1189 	    ("sched_rem: thread not on run queue"));
1190 	mtx_assert(&sched_lock, MA_OWNED);
1191 	CTR5(KTR_SCHED, "sched_rem: %p(%s) prio %d by %p(%s)",
1192 	    td, td->td_proc->p_comm, td->td_priority, curthread,
1193 	    curthread->td_proc->p_comm);
1194 
1195 	if ((td->td_proc->p_flag & P_NOLOAD) == 0)
1196 		sched_load_rem();
1197 	runq_remove(ts->ts_runq, ts);
1198 	TD_SET_CAN_RUN(td);
1199 }
1200 
1201 /*
1202  * Select threads to run.
1203  * Notice that the running threads still consume a slot.
1204  */
1205 struct thread *
1206 sched_choose(void)
1207 {
1208 	struct td_sched *ts;
1209 	struct runq *rq;
1210 
1211 #ifdef SMP
1212 	struct td_sched *kecpu;
1213 
1214 	rq = &runq;
1215 	ts = runq_choose(&runq);
1216 	kecpu = runq_choose(&runq_pcpu[PCPU_GET(cpuid)]);
1217 
1218 	if (ts == NULL ||
1219 	    (kecpu != NULL &&
1220 	     kecpu->ts_thread->td_priority < ts->ts_thread->td_priority)) {
1221 		CTR2(KTR_RUNQ, "choosing td_sched %p from pcpu runq %d", kecpu,
1222 		     PCPU_GET(cpuid));
1223 		ts = kecpu;
1224 		rq = &runq_pcpu[PCPU_GET(cpuid)];
1225 	} else {
1226 		CTR1(KTR_RUNQ, "choosing td_sched %p from main runq", ts);
1227 	}
1228 
1229 #else
1230 	rq = &runq;
1231 	ts = runq_choose(&runq);
1232 #endif
1233 
1234 	if (ts) {
1235 		runq_remove(rq, ts);
1236 		ts->ts_flags |= TSF_DIDRUN;
1237 
1238 		KASSERT(ts->ts_thread->td_proc->p_sflag & PS_INMEM,
1239 		    ("sched_choose: process swapped out"));
1240 		return (ts->ts_thread);
1241 	}
1242 	return (PCPU_GET(idlethread));
1243 }
1244 
1245 void
1246 sched_userret(struct thread *td)
1247 {
1248 	/*
1249 	 * XXX we cheat slightly on the locking here to avoid locking in
1250 	 * the usual case.  Setting td_priority here is essentially an
1251 	 * incomplete workaround for not setting it properly elsewhere.
1252 	 * Now that some interrupt handlers are threads, not setting it
1253 	 * properly elsewhere can clobber it in the window between setting
1254 	 * it here and returning to user mode, so don't waste time setting
1255 	 * it perfectly here.
1256 	 */
1257 	KASSERT((td->td_flags & TDF_BORROWING) == 0,
1258 	    ("thread with borrowed priority returning to userland"));
1259 	if (td->td_priority != td->td_user_pri) {
1260 		mtx_lock_spin(&sched_lock);
1261 		td->td_priority = td->td_user_pri;
1262 		td->td_base_pri = td->td_user_pri;
1263 		mtx_unlock_spin(&sched_lock);
1264 	}
1265 }
1266 
1267 void
1268 sched_bind(struct thread *td, int cpu)
1269 {
1270 	struct td_sched *ts;
1271 
1272 	mtx_assert(&sched_lock, MA_OWNED);
1273 	KASSERT(TD_IS_RUNNING(td),
1274 	    ("sched_bind: cannot bind non-running thread"));
1275 
1276 	ts = td->td_sched;
1277 
1278 	ts->ts_flags |= TSF_BOUND;
1279 #ifdef SMP
1280 	ts->ts_runq = &runq_pcpu[cpu];
1281 	if (PCPU_GET(cpuid) == cpu)
1282 		return;
1283 
1284 	mi_switch(SW_VOL, NULL);
1285 #endif
1286 }
1287 
1288 void
1289 sched_unbind(struct thread* td)
1290 {
1291 	mtx_assert(&sched_lock, MA_OWNED);
1292 	td->td_sched->ts_flags &= ~TSF_BOUND;
1293 }
1294 
1295 int
1296 sched_is_bound(struct thread *td)
1297 {
1298 	mtx_assert(&sched_lock, MA_OWNED);
1299 	return (td->td_sched->ts_flags & TSF_BOUND);
1300 }
1301 
1302 void
1303 sched_relinquish(struct thread *td)
1304 {
1305 	mtx_lock_spin(&sched_lock);
1306 	if (td->td_pri_class == PRI_TIMESHARE)
1307 		sched_prio(td, PRI_MAX_TIMESHARE);
1308 	mi_switch(SW_VOL, NULL);
1309 	mtx_unlock_spin(&sched_lock);
1310 }
1311 
1312 int
1313 sched_load(void)
1314 {
1315 	return (sched_tdcnt);
1316 }
1317 
1318 int
1319 sched_sizeof_proc(void)
1320 {
1321 	return (sizeof(struct proc));
1322 }
1323 
1324 int
1325 sched_sizeof_thread(void)
1326 {
1327 	return (sizeof(struct thread) + sizeof(struct td_sched));
1328 }
1329 
1330 fixpt_t
1331 sched_pctcpu(struct thread *td)
1332 {
1333 	struct td_sched *ts;
1334 
1335 	ts = td->td_sched;
1336 	return (ts->ts_pctcpu);
1337 }
1338 
1339 void
1340 sched_tick(void)
1341 {
1342 }
1343 
1344 /*
1345  * The actual idle process.
1346  */
1347 void
1348 sched_idletd(void *dummy)
1349 {
1350 	struct proc *p;
1351 	struct thread *td;
1352 
1353 	td = curthread;
1354 	p = td->td_proc;
1355 	for (;;) {
1356 		mtx_assert(&Giant, MA_NOTOWNED);
1357 
1358 		while (sched_runnable() == 0)
1359 			cpu_idle();
1360 
1361 		mtx_lock_spin(&sched_lock);
1362 		mi_switch(SW_VOL, NULL);
1363 		mtx_unlock_spin(&sched_lock);
1364 	}
1365 }
1366 
1367 #define KERN_SWITCH_INCLUDE 1
1368 #include "kern/kern_switch.c"
1369