xref: /freebsd/sys/kern/sched_4bsd.c (revision 66e576525d35c68fcb86f142ebaa5a448555c0c7)
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 #include "opt_sched.h"
40 #include "opt_kdtrace.h"
41 
42 #include <sys/param.h>
43 #include <sys/systm.h>
44 #include <sys/cpuset.h>
45 #include <sys/kernel.h>
46 #include <sys/ktr.h>
47 #include <sys/lock.h>
48 #include <sys/kthread.h>
49 #include <sys/mutex.h>
50 #include <sys/proc.h>
51 #include <sys/resourcevar.h>
52 #include <sys/sched.h>
53 #include <sys/smp.h>
54 #include <sys/sysctl.h>
55 #include <sys/sx.h>
56 #include <sys/turnstile.h>
57 #include <sys/umtx.h>
58 #include <machine/pcb.h>
59 #include <machine/smp.h>
60 
61 #ifdef HWPMC_HOOKS
62 #include <sys/pmckern.h>
63 #endif
64 
65 #ifdef KDTRACE_HOOKS
66 #include <sys/dtrace_bsd.h>
67 int				dtrace_vtime_active;
68 dtrace_vtime_switch_func_t	dtrace_vtime_switch_func;
69 #endif
70 
71 /*
72  * INVERSE_ESTCPU_WEIGHT is only suitable for statclock() frequencies in
73  * the range 100-256 Hz (approximately).
74  */
75 #define	ESTCPULIM(e) \
76     min((e), INVERSE_ESTCPU_WEIGHT * (NICE_WEIGHT * (PRIO_MAX - PRIO_MIN) - \
77     RQ_PPQ) + INVERSE_ESTCPU_WEIGHT - 1)
78 #ifdef SMP
79 #define	INVERSE_ESTCPU_WEIGHT	(8 * smp_cpus)
80 #else
81 #define	INVERSE_ESTCPU_WEIGHT	8	/* 1 / (priorities per estcpu level). */
82 #endif
83 #define	NICE_WEIGHT		1	/* Priorities per nice level. */
84 
85 #define	TS_NAME_LEN (MAXCOMLEN + sizeof(" td ") + sizeof(__XSTRING(UINT_MAX)))
86 
87 /*
88  * The schedulable entity that runs a context.
89  * This is  an extension to the thread structure and is tailored to
90  * the requirements of this scheduler
91  */
92 struct td_sched {
93 	fixpt_t		ts_pctcpu;	/* (j) %cpu during p_swtime. */
94 	int		ts_cpticks;	/* (j) Ticks of cpu time. */
95 	int		ts_slptime;	/* (j) Seconds !RUNNING. */
96 	int		ts_flags;
97 	struct runq	*ts_runq;	/* runq the thread is currently on */
98 #ifdef KTR
99 	char		ts_name[TS_NAME_LEN];
100 #endif
101 };
102 
103 /* flags kept in td_flags */
104 #define TDF_DIDRUN	TDF_SCHED0	/* thread actually ran. */
105 #define TDF_BOUND	TDF_SCHED1	/* Bound to one CPU. */
106 
107 /* flags kept in ts_flags */
108 #define	TSF_AFFINITY	0x0001		/* Has a non-"full" CPU set. */
109 
110 #define SKE_RUNQ_PCPU(ts)						\
111     ((ts)->ts_runq != 0 && (ts)->ts_runq != &runq)
112 
113 #define	THREAD_CAN_SCHED(td, cpu)	\
114     CPU_ISSET((cpu), &(td)->td_cpuset->cs_mask)
115 
116 static struct td_sched td_sched0;
117 struct mtx sched_lock;
118 
119 static int	sched_tdcnt;	/* Total runnable threads in the system. */
120 static int	sched_quantum;	/* Roundrobin scheduling quantum in ticks. */
121 #define	SCHED_QUANTUM	(hz / 10)	/* Default sched quantum */
122 
123 static void	setup_runqs(void);
124 static void	schedcpu(void);
125 static void	schedcpu_thread(void);
126 static void	sched_priority(struct thread *td, u_char prio);
127 static void	sched_setup(void *dummy);
128 static void	maybe_resched(struct thread *td);
129 static void	updatepri(struct thread *td);
130 static void	resetpriority(struct thread *td);
131 static void	resetpriority_thread(struct thread *td);
132 #ifdef SMP
133 static int	sched_pickcpu(struct thread *td);
134 static int	forward_wakeup(int cpunum);
135 static void	kick_other_cpu(int pri, int cpuid);
136 #endif
137 
138 static struct kproc_desc sched_kp = {
139         "schedcpu",
140         schedcpu_thread,
141         NULL
142 };
143 SYSINIT(schedcpu, SI_SUB_RUN_SCHEDULER, SI_ORDER_FIRST, kproc_start,
144     &sched_kp);
145 SYSINIT(sched_setup, SI_SUB_RUN_QUEUE, SI_ORDER_FIRST, sched_setup, NULL);
146 
147 /*
148  * Global run queue.
149  */
150 static struct runq runq;
151 
152 #ifdef SMP
153 /*
154  * Per-CPU run queues
155  */
156 static struct runq runq_pcpu[MAXCPU];
157 long runq_length[MAXCPU];
158 #endif
159 
160 struct pcpuidlestat {
161 	u_int idlecalls;
162 	u_int oldidlecalls;
163 };
164 static DPCPU_DEFINE(struct pcpuidlestat, idlestat);
165 
166 static void
167 setup_runqs(void)
168 {
169 #ifdef SMP
170 	int i;
171 
172 	for (i = 0; i < MAXCPU; ++i)
173 		runq_init(&runq_pcpu[i]);
174 #endif
175 
176 	runq_init(&runq);
177 }
178 
179 static int
180 sysctl_kern_quantum(SYSCTL_HANDLER_ARGS)
181 {
182 	int error, new_val;
183 
184 	new_val = sched_quantum * tick;
185 	error = sysctl_handle_int(oidp, &new_val, 0, req);
186         if (error != 0 || req->newptr == NULL)
187 		return (error);
188 	if (new_val < tick)
189 		return (EINVAL);
190 	sched_quantum = new_val / tick;
191 	hogticks = 2 * sched_quantum;
192 	return (0);
193 }
194 
195 SYSCTL_NODE(_kern, OID_AUTO, sched, CTLFLAG_RD, 0, "Scheduler");
196 
197 SYSCTL_STRING(_kern_sched, OID_AUTO, name, CTLFLAG_RD, "4BSD", 0,
198     "Scheduler name");
199 
200 SYSCTL_PROC(_kern_sched, OID_AUTO, quantum, CTLTYPE_INT | CTLFLAG_RW,
201     0, sizeof sched_quantum, sysctl_kern_quantum, "I",
202     "Roundrobin scheduling quantum in microseconds");
203 
204 #ifdef SMP
205 /* Enable forwarding of wakeups to all other cpus */
206 SYSCTL_NODE(_kern_sched, OID_AUTO, ipiwakeup, CTLFLAG_RD, NULL, "Kernel SMP");
207 
208 static int runq_fuzz = 1;
209 SYSCTL_INT(_kern_sched, OID_AUTO, runq_fuzz, CTLFLAG_RW, &runq_fuzz, 0, "");
210 
211 static int forward_wakeup_enabled = 1;
212 SYSCTL_INT(_kern_sched_ipiwakeup, OID_AUTO, enabled, CTLFLAG_RW,
213 	   &forward_wakeup_enabled, 0,
214 	   "Forwarding of wakeup to idle CPUs");
215 
216 static int forward_wakeups_requested = 0;
217 SYSCTL_INT(_kern_sched_ipiwakeup, OID_AUTO, requested, CTLFLAG_RD,
218 	   &forward_wakeups_requested, 0,
219 	   "Requests for Forwarding of wakeup to idle CPUs");
220 
221 static int forward_wakeups_delivered = 0;
222 SYSCTL_INT(_kern_sched_ipiwakeup, OID_AUTO, delivered, CTLFLAG_RD,
223 	   &forward_wakeups_delivered, 0,
224 	   "Completed Forwarding of wakeup to idle CPUs");
225 
226 static int forward_wakeup_use_mask = 1;
227 SYSCTL_INT(_kern_sched_ipiwakeup, OID_AUTO, usemask, CTLFLAG_RW,
228 	   &forward_wakeup_use_mask, 0,
229 	   "Use the mask of idle cpus");
230 
231 static int forward_wakeup_use_loop = 0;
232 SYSCTL_INT(_kern_sched_ipiwakeup, OID_AUTO, useloop, CTLFLAG_RW,
233 	   &forward_wakeup_use_loop, 0,
234 	   "Use a loop to find idle cpus");
235 
236 static int forward_wakeup_use_single = 0;
237 SYSCTL_INT(_kern_sched_ipiwakeup, OID_AUTO, onecpu, CTLFLAG_RW,
238 	   &forward_wakeup_use_single, 0,
239 	   "Only signal one idle cpu");
240 
241 static int forward_wakeup_use_htt = 0;
242 SYSCTL_INT(_kern_sched_ipiwakeup, OID_AUTO, htt2, CTLFLAG_RW,
243 	   &forward_wakeup_use_htt, 0,
244 	   "account for htt");
245 
246 #endif
247 #if 0
248 static int sched_followon = 0;
249 SYSCTL_INT(_kern_sched, OID_AUTO, followon, CTLFLAG_RW,
250 	   &sched_followon, 0,
251 	   "allow threads to share a quantum");
252 #endif
253 
254 static __inline void
255 sched_load_add(void)
256 {
257 
258 	sched_tdcnt++;
259 	KTR_COUNTER0(KTR_SCHED, "load", "global load", sched_tdcnt);
260 }
261 
262 static __inline void
263 sched_load_rem(void)
264 {
265 
266 	sched_tdcnt--;
267 	KTR_COUNTER0(KTR_SCHED, "load", "global load", sched_tdcnt);
268 }
269 /*
270  * Arrange to reschedule if necessary, taking the priorities and
271  * schedulers into account.
272  */
273 static void
274 maybe_resched(struct thread *td)
275 {
276 
277 	THREAD_LOCK_ASSERT(td, MA_OWNED);
278 	if (td->td_priority < curthread->td_priority)
279 		curthread->td_flags |= TDF_NEEDRESCHED;
280 }
281 
282 /*
283  * This function is called when a thread is about to be put on run queue
284  * because it has been made runnable or its priority has been adjusted.  It
285  * determines if the new thread should be immediately preempted to.  If so,
286  * it switches to it and eventually returns true.  If not, it returns false
287  * so that the caller may place the thread on an appropriate run queue.
288  */
289 int
290 maybe_preempt(struct thread *td)
291 {
292 #ifdef PREEMPTION
293 	struct thread *ctd;
294 	int cpri, pri;
295 
296 	/*
297 	 * The new thread should not preempt the current thread if any of the
298 	 * following conditions are true:
299 	 *
300 	 *  - The kernel is in the throes of crashing (panicstr).
301 	 *  - The current thread has a higher (numerically lower) or
302 	 *    equivalent priority.  Note that this prevents curthread from
303 	 *    trying to preempt to itself.
304 	 *  - It is too early in the boot for context switches (cold is set).
305 	 *  - The current thread has an inhibitor set or is in the process of
306 	 *    exiting.  In this case, the current thread is about to switch
307 	 *    out anyways, so there's no point in preempting.  If we did,
308 	 *    the current thread would not be properly resumed as well, so
309 	 *    just avoid that whole landmine.
310 	 *  - If the new thread's priority is not a realtime priority and
311 	 *    the current thread's priority is not an idle priority and
312 	 *    FULL_PREEMPTION is disabled.
313 	 *
314 	 * If all of these conditions are false, but the current thread is in
315 	 * a nested critical section, then we have to defer the preemption
316 	 * until we exit the critical section.  Otherwise, switch immediately
317 	 * to the new thread.
318 	 */
319 	ctd = curthread;
320 	THREAD_LOCK_ASSERT(td, MA_OWNED);
321 	KASSERT((td->td_inhibitors == 0),
322 			("maybe_preempt: trying to run inhibited thread"));
323 	pri = td->td_priority;
324 	cpri = ctd->td_priority;
325 	if (panicstr != NULL || pri >= cpri || cold /* || dumping */ ||
326 	    TD_IS_INHIBITED(ctd))
327 		return (0);
328 #ifndef FULL_PREEMPTION
329 	if (pri > PRI_MAX_ITHD && cpri < PRI_MIN_IDLE)
330 		return (0);
331 #endif
332 
333 	if (ctd->td_critnest > 1) {
334 		CTR1(KTR_PROC, "maybe_preempt: in critical section %d",
335 		    ctd->td_critnest);
336 		ctd->td_owepreempt = 1;
337 		return (0);
338 	}
339 	/*
340 	 * Thread is runnable but not yet put on system run queue.
341 	 */
342 	MPASS(ctd->td_lock == td->td_lock);
343 	MPASS(TD_ON_RUNQ(td));
344 	TD_SET_RUNNING(td);
345 	CTR3(KTR_PROC, "preempting to thread %p (pid %d, %s)\n", td,
346 	    td->td_proc->p_pid, td->td_name);
347 	mi_switch(SW_INVOL | SW_PREEMPT | SWT_PREEMPT, td);
348 	/*
349 	 * td's lock pointer may have changed.  We have to return with it
350 	 * locked.
351 	 */
352 	spinlock_enter();
353 	thread_unlock(ctd);
354 	thread_lock(td);
355 	spinlock_exit();
356 	return (1);
357 #else
358 	return (0);
359 #endif
360 }
361 
362 /*
363  * Constants for digital decay and forget:
364  *	90% of (td_estcpu) usage in 5 * loadav time
365  *	95% of (ts_pctcpu) usage in 60 seconds (load insensitive)
366  *          Note that, as ps(1) mentions, this can let percentages
367  *          total over 100% (I've seen 137.9% for 3 processes).
368  *
369  * Note that schedclock() updates td_estcpu and p_cpticks asynchronously.
370  *
371  * We wish to decay away 90% of td_estcpu in (5 * loadavg) seconds.
372  * That is, the system wants to compute a value of decay such
373  * that the following for loop:
374  * 	for (i = 0; i < (5 * loadavg); i++)
375  * 		td_estcpu *= decay;
376  * will compute
377  * 	td_estcpu *= 0.1;
378  * for all values of loadavg:
379  *
380  * Mathematically this loop can be expressed by saying:
381  * 	decay ** (5 * loadavg) ~= .1
382  *
383  * The system computes decay as:
384  * 	decay = (2 * loadavg) / (2 * loadavg + 1)
385  *
386  * We wish to prove that the system's computation of decay
387  * will always fulfill the equation:
388  * 	decay ** (5 * loadavg) ~= .1
389  *
390  * If we compute b as:
391  * 	b = 2 * loadavg
392  * then
393  * 	decay = b / (b + 1)
394  *
395  * We now need to prove two things:
396  *	1) Given factor ** (5 * loadavg) ~= .1, prove factor == b/(b+1)
397  *	2) Given b/(b+1) ** power ~= .1, prove power == (5 * loadavg)
398  *
399  * Facts:
400  *         For x close to zero, exp(x) =~ 1 + x, since
401  *              exp(x) = 0! + x**1/1! + x**2/2! + ... .
402  *              therefore exp(-1/b) =~ 1 - (1/b) = (b-1)/b.
403  *         For x close to zero, ln(1+x) =~ x, since
404  *              ln(1+x) = x - x**2/2 + x**3/3 - ...     -1 < x < 1
405  *              therefore ln(b/(b+1)) = ln(1 - 1/(b+1)) =~ -1/(b+1).
406  *         ln(.1) =~ -2.30
407  *
408  * Proof of (1):
409  *    Solve (factor)**(power) =~ .1 given power (5*loadav):
410  *	solving for factor,
411  *      ln(factor) =~ (-2.30/5*loadav), or
412  *      factor =~ exp(-1/((5/2.30)*loadav)) =~ exp(-1/(2*loadav)) =
413  *          exp(-1/b) =~ (b-1)/b =~ b/(b+1).                    QED
414  *
415  * Proof of (2):
416  *    Solve (factor)**(power) =~ .1 given factor == (b/(b+1)):
417  *	solving for power,
418  *      power*ln(b/(b+1)) =~ -2.30, or
419  *      power =~ 2.3 * (b + 1) = 4.6*loadav + 2.3 =~ 5*loadav.  QED
420  *
421  * Actual power values for the implemented algorithm are as follows:
422  *      loadav: 1       2       3       4
423  *      power:  5.68    10.32   14.94   19.55
424  */
425 
426 /* calculations for digital decay to forget 90% of usage in 5*loadav sec */
427 #define	loadfactor(loadav)	(2 * (loadav))
428 #define	decay_cpu(loadfac, cpu)	(((loadfac) * (cpu)) / ((loadfac) + FSCALE))
429 
430 /* decay 95% of `ts_pctcpu' in 60 seconds; see CCPU_SHIFT before changing */
431 static fixpt_t	ccpu = 0.95122942450071400909 * FSCALE;	/* exp(-1/20) */
432 SYSCTL_INT(_kern, OID_AUTO, ccpu, CTLFLAG_RD, &ccpu, 0, "");
433 
434 /*
435  * If `ccpu' is not equal to `exp(-1/20)' and you still want to use the
436  * faster/more-accurate formula, you'll have to estimate CCPU_SHIFT below
437  * and possibly adjust FSHIFT in "param.h" so that (FSHIFT >= CCPU_SHIFT).
438  *
439  * To estimate CCPU_SHIFT for exp(-1/20), the following formula was used:
440  *	1 - exp(-1/20) ~= 0.0487 ~= 0.0488 == 1 (fixed pt, *11* bits).
441  *
442  * If you don't want to bother with the faster/more-accurate formula, you
443  * can set CCPU_SHIFT to (FSHIFT + 1) which will use a slower/less-accurate
444  * (more general) method of calculating the %age of CPU used by a process.
445  */
446 #define	CCPU_SHIFT	11
447 
448 /*
449  * Recompute process priorities, every hz ticks.
450  * MP-safe, called without the Giant mutex.
451  */
452 /* ARGSUSED */
453 static void
454 schedcpu(void)
455 {
456 	register fixpt_t loadfac = loadfactor(averunnable.ldavg[0]);
457 	struct thread *td;
458 	struct proc *p;
459 	struct td_sched *ts;
460 	int awake, realstathz;
461 
462 	realstathz = stathz ? stathz : hz;
463 	sx_slock(&allproc_lock);
464 	FOREACH_PROC_IN_SYSTEM(p) {
465 		PROC_LOCK(p);
466 		FOREACH_THREAD_IN_PROC(p, td) {
467 			awake = 0;
468 			thread_lock(td);
469 			ts = td->td_sched;
470 			/*
471 			 * Increment sleep time (if sleeping).  We
472 			 * ignore overflow, as above.
473 			 */
474 			/*
475 			 * The td_sched slptimes are not touched in wakeup
476 			 * because the thread may not HAVE everything in
477 			 * memory? XXX I think this is out of date.
478 			 */
479 			if (TD_ON_RUNQ(td)) {
480 				awake = 1;
481 				td->td_flags &= ~TDF_DIDRUN;
482 			} else if (TD_IS_RUNNING(td)) {
483 				awake = 1;
484 				/* Do not clear TDF_DIDRUN */
485 			} else if (td->td_flags & TDF_DIDRUN) {
486 				awake = 1;
487 				td->td_flags &= ~TDF_DIDRUN;
488 			}
489 
490 			/*
491 			 * ts_pctcpu is only for ps and ttyinfo().
492 			 */
493 			ts->ts_pctcpu = (ts->ts_pctcpu * ccpu) >> FSHIFT;
494 			/*
495 			 * If the td_sched has been idle the entire second,
496 			 * stop recalculating its priority until
497 			 * it wakes up.
498 			 */
499 			if (ts->ts_cpticks != 0) {
500 #if	(FSHIFT >= CCPU_SHIFT)
501 				ts->ts_pctcpu += (realstathz == 100)
502 				    ? ((fixpt_t) ts->ts_cpticks) <<
503 				    (FSHIFT - CCPU_SHIFT) :
504 				    100 * (((fixpt_t) ts->ts_cpticks)
505 				    << (FSHIFT - CCPU_SHIFT)) / realstathz;
506 #else
507 				ts->ts_pctcpu += ((FSCALE - ccpu) *
508 				    (ts->ts_cpticks *
509 				    FSCALE / realstathz)) >> FSHIFT;
510 #endif
511 				ts->ts_cpticks = 0;
512 			}
513 			/*
514 			 * If there are ANY running threads in this process,
515 			 * then don't count it as sleeping.
516 			 * XXX: this is broken.
517 			 */
518 			if (awake) {
519 				if (ts->ts_slptime > 1) {
520 					/*
521 					 * In an ideal world, this should not
522 					 * happen, because whoever woke us
523 					 * up from the long sleep should have
524 					 * unwound the slptime and reset our
525 					 * priority before we run at the stale
526 					 * priority.  Should KASSERT at some
527 					 * point when all the cases are fixed.
528 					 */
529 					updatepri(td);
530 				}
531 				ts->ts_slptime = 0;
532 			} else
533 				ts->ts_slptime++;
534 			if (ts->ts_slptime > 1) {
535 				thread_unlock(td);
536 				continue;
537 			}
538 			td->td_estcpu = decay_cpu(loadfac, td->td_estcpu);
539 		      	resetpriority(td);
540 			resetpriority_thread(td);
541 			thread_unlock(td);
542 		}
543 		PROC_UNLOCK(p);
544 	}
545 	sx_sunlock(&allproc_lock);
546 }
547 
548 /*
549  * Main loop for a kthread that executes schedcpu once a second.
550  */
551 static void
552 schedcpu_thread(void)
553 {
554 
555 	for (;;) {
556 		schedcpu();
557 		pause("-", hz);
558 	}
559 }
560 
561 /*
562  * Recalculate the priority of a process after it has slept for a while.
563  * For all load averages >= 1 and max td_estcpu of 255, sleeping for at
564  * least six times the loadfactor will decay td_estcpu to zero.
565  */
566 static void
567 updatepri(struct thread *td)
568 {
569 	struct td_sched *ts;
570 	fixpt_t loadfac;
571 	unsigned int newcpu;
572 
573 	ts = td->td_sched;
574 	loadfac = loadfactor(averunnable.ldavg[0]);
575 	if (ts->ts_slptime > 5 * loadfac)
576 		td->td_estcpu = 0;
577 	else {
578 		newcpu = td->td_estcpu;
579 		ts->ts_slptime--;	/* was incremented in schedcpu() */
580 		while (newcpu && --ts->ts_slptime)
581 			newcpu = decay_cpu(loadfac, newcpu);
582 		td->td_estcpu = newcpu;
583 	}
584 }
585 
586 /*
587  * Compute the priority of a process when running in user mode.
588  * Arrange to reschedule if the resulting priority is better
589  * than that of the current process.
590  */
591 static void
592 resetpriority(struct thread *td)
593 {
594 	register unsigned int newpriority;
595 
596 	if (td->td_pri_class == PRI_TIMESHARE) {
597 		newpriority = PUSER + td->td_estcpu / INVERSE_ESTCPU_WEIGHT +
598 		    NICE_WEIGHT * (td->td_proc->p_nice - PRIO_MIN);
599 		newpriority = min(max(newpriority, PRI_MIN_TIMESHARE),
600 		    PRI_MAX_TIMESHARE);
601 		sched_user_prio(td, newpriority);
602 	}
603 }
604 
605 /*
606  * Update the thread's priority when the associated process's user
607  * priority changes.
608  */
609 static void
610 resetpriority_thread(struct thread *td)
611 {
612 
613 	/* Only change threads with a time sharing user priority. */
614 	if (td->td_priority < PRI_MIN_TIMESHARE ||
615 	    td->td_priority > PRI_MAX_TIMESHARE)
616 		return;
617 
618 	/* XXX the whole needresched thing is broken, but not silly. */
619 	maybe_resched(td);
620 
621 	sched_prio(td, td->td_user_pri);
622 }
623 
624 /* ARGSUSED */
625 static void
626 sched_setup(void *dummy)
627 {
628 	setup_runqs();
629 
630 	if (sched_quantum == 0)
631 		sched_quantum = SCHED_QUANTUM;
632 	hogticks = 2 * sched_quantum;
633 
634 	/* Account for thread0. */
635 	sched_load_add();
636 }
637 
638 /* External interfaces start here */
639 
640 /*
641  * Very early in the boot some setup of scheduler-specific
642  * parts of proc0 and of some scheduler resources needs to be done.
643  * Called from:
644  *  proc0_init()
645  */
646 void
647 schedinit(void)
648 {
649 	/*
650 	 * Set up the scheduler specific parts of proc0.
651 	 */
652 	proc0.p_sched = NULL; /* XXX */
653 	thread0.td_sched = &td_sched0;
654 	thread0.td_lock = &sched_lock;
655 	mtx_init(&sched_lock, "sched lock", NULL, MTX_SPIN | MTX_RECURSE);
656 }
657 
658 int
659 sched_runnable(void)
660 {
661 #ifdef SMP
662 	return runq_check(&runq) + runq_check(&runq_pcpu[PCPU_GET(cpuid)]);
663 #else
664 	return runq_check(&runq);
665 #endif
666 }
667 
668 int
669 sched_rr_interval(void)
670 {
671 	if (sched_quantum == 0)
672 		sched_quantum = SCHED_QUANTUM;
673 	return (sched_quantum);
674 }
675 
676 /*
677  * We adjust the priority of the current process.  The priority of
678  * a process gets worse as it accumulates CPU time.  The cpu usage
679  * estimator (td_estcpu) is increased here.  resetpriority() will
680  * compute a different priority each time td_estcpu increases by
681  * INVERSE_ESTCPU_WEIGHT
682  * (until MAXPRI is reached).  The cpu usage estimator ramps up
683  * quite quickly when the process is running (linearly), and decays
684  * away exponentially, at a rate which is proportionally slower when
685  * the system is busy.  The basic principle is that the system will
686  * 90% forget that the process used a lot of CPU time in 5 * loadav
687  * seconds.  This causes the system to favor processes which haven't
688  * run much recently, and to round-robin among other processes.
689  */
690 void
691 sched_clock(struct thread *td)
692 {
693 	struct pcpuidlestat *stat;
694 	struct td_sched *ts;
695 
696 	THREAD_LOCK_ASSERT(td, MA_OWNED);
697 	ts = td->td_sched;
698 
699 	ts->ts_cpticks++;
700 	td->td_estcpu = ESTCPULIM(td->td_estcpu + 1);
701 	if ((td->td_estcpu % INVERSE_ESTCPU_WEIGHT) == 0) {
702 		resetpriority(td);
703 		resetpriority_thread(td);
704 	}
705 
706 	/*
707 	 * Force a context switch if the current thread has used up a full
708 	 * quantum (default quantum is 100ms).
709 	 */
710 	if (!TD_IS_IDLETHREAD(td) &&
711 	    ticks - PCPU_GET(switchticks) >= sched_quantum)
712 		td->td_flags |= TDF_NEEDRESCHED;
713 
714 	stat = DPCPU_PTR(idlestat);
715 	stat->oldidlecalls = stat->idlecalls;
716 	stat->idlecalls = 0;
717 }
718 
719 /*
720  * Charge child's scheduling CPU usage to parent.
721  */
722 void
723 sched_exit(struct proc *p, struct thread *td)
724 {
725 
726 	KTR_STATE1(KTR_SCHED, "thread", sched_tdname(td), "proc exit",
727 	    "prio:td", td->td_priority);
728 
729 	PROC_LOCK_ASSERT(p, MA_OWNED);
730 	sched_exit_thread(FIRST_THREAD_IN_PROC(p), td);
731 }
732 
733 void
734 sched_exit_thread(struct thread *td, struct thread *child)
735 {
736 
737 	KTR_STATE1(KTR_SCHED, "thread", sched_tdname(child), "exit",
738 	    "prio:td", child->td_priority);
739 	thread_lock(td);
740 	td->td_estcpu = ESTCPULIM(td->td_estcpu + child->td_estcpu);
741 	thread_unlock(td);
742 	thread_lock(child);
743 	if ((child->td_flags & TDF_NOLOAD) == 0)
744 		sched_load_rem();
745 	thread_unlock(child);
746 }
747 
748 void
749 sched_fork(struct thread *td, struct thread *childtd)
750 {
751 	sched_fork_thread(td, childtd);
752 }
753 
754 void
755 sched_fork_thread(struct thread *td, struct thread *childtd)
756 {
757 	struct td_sched *ts;
758 
759 	childtd->td_estcpu = td->td_estcpu;
760 	childtd->td_lock = &sched_lock;
761 	childtd->td_cpuset = cpuset_ref(td->td_cpuset);
762 	ts = childtd->td_sched;
763 	bzero(ts, sizeof(*ts));
764 	ts->ts_flags |= (td->td_sched->ts_flags & TSF_AFFINITY);
765 }
766 
767 void
768 sched_nice(struct proc *p, int nice)
769 {
770 	struct thread *td;
771 
772 	PROC_LOCK_ASSERT(p, MA_OWNED);
773 	p->p_nice = nice;
774 	FOREACH_THREAD_IN_PROC(p, td) {
775 		thread_lock(td);
776 		resetpriority(td);
777 		resetpriority_thread(td);
778 		thread_unlock(td);
779 	}
780 }
781 
782 void
783 sched_class(struct thread *td, int class)
784 {
785 	THREAD_LOCK_ASSERT(td, MA_OWNED);
786 	td->td_pri_class = class;
787 }
788 
789 /*
790  * Adjust the priority of a thread.
791  */
792 static void
793 sched_priority(struct thread *td, u_char prio)
794 {
795 
796 
797 	KTR_POINT3(KTR_SCHED, "thread", sched_tdname(td), "priority change",
798 	    "prio:%d", td->td_priority, "new prio:%d", prio, KTR_ATTR_LINKED,
799 	    sched_tdname(curthread));
800 	if (td != curthread && prio > td->td_priority) {
801 		KTR_POINT3(KTR_SCHED, "thread", sched_tdname(curthread),
802 		    "lend prio", "prio:%d", td->td_priority, "new prio:%d",
803 		    prio, KTR_ATTR_LINKED, sched_tdname(td));
804 	}
805 	THREAD_LOCK_ASSERT(td, MA_OWNED);
806 	if (td->td_priority == prio)
807 		return;
808 	td->td_priority = prio;
809 	if (TD_ON_RUNQ(td) && td->td_rqindex != (prio / RQ_PPQ)) {
810 		sched_rem(td);
811 		sched_add(td, SRQ_BORING);
812 	}
813 }
814 
815 /*
816  * Update a thread's priority when it is lent another thread's
817  * priority.
818  */
819 void
820 sched_lend_prio(struct thread *td, u_char prio)
821 {
822 
823 	td->td_flags |= TDF_BORROWING;
824 	sched_priority(td, prio);
825 }
826 
827 /*
828  * Restore a thread's priority when priority propagation is
829  * over.  The prio argument is the minimum priority the thread
830  * needs to have to satisfy other possible priority lending
831  * requests.  If the thread's regulary priority is less
832  * important than prio the thread will keep a priority boost
833  * of prio.
834  */
835 void
836 sched_unlend_prio(struct thread *td, u_char prio)
837 {
838 	u_char base_pri;
839 
840 	if (td->td_base_pri >= PRI_MIN_TIMESHARE &&
841 	    td->td_base_pri <= PRI_MAX_TIMESHARE)
842 		base_pri = td->td_user_pri;
843 	else
844 		base_pri = td->td_base_pri;
845 	if (prio >= base_pri) {
846 		td->td_flags &= ~TDF_BORROWING;
847 		sched_prio(td, base_pri);
848 	} else
849 		sched_lend_prio(td, prio);
850 }
851 
852 void
853 sched_prio(struct thread *td, u_char prio)
854 {
855 	u_char oldprio;
856 
857 	/* First, update the base priority. */
858 	td->td_base_pri = prio;
859 
860 	/*
861 	 * If the thread is borrowing another thread's priority, don't ever
862 	 * lower the priority.
863 	 */
864 	if (td->td_flags & TDF_BORROWING && td->td_priority < prio)
865 		return;
866 
867 	/* Change the real priority. */
868 	oldprio = td->td_priority;
869 	sched_priority(td, prio);
870 
871 	/*
872 	 * If the thread is on a turnstile, then let the turnstile update
873 	 * its state.
874 	 */
875 	if (TD_ON_LOCK(td) && oldprio != prio)
876 		turnstile_adjust(td, oldprio);
877 }
878 
879 void
880 sched_user_prio(struct thread *td, u_char prio)
881 {
882 	u_char oldprio;
883 
884 	THREAD_LOCK_ASSERT(td, MA_OWNED);
885 	td->td_base_user_pri = prio;
886 	if (td->td_flags & TDF_UBORROWING && td->td_user_pri <= prio)
887 		return;
888 	oldprio = td->td_user_pri;
889 	td->td_user_pri = prio;
890 }
891 
892 void
893 sched_lend_user_prio(struct thread *td, u_char prio)
894 {
895 	u_char oldprio;
896 
897 	THREAD_LOCK_ASSERT(td, MA_OWNED);
898 	td->td_flags |= TDF_UBORROWING;
899 	oldprio = td->td_user_pri;
900 	td->td_user_pri = prio;
901 }
902 
903 void
904 sched_unlend_user_prio(struct thread *td, u_char prio)
905 {
906 	u_char base_pri;
907 
908 	THREAD_LOCK_ASSERT(td, MA_OWNED);
909 	base_pri = td->td_base_user_pri;
910 	if (prio >= base_pri) {
911 		td->td_flags &= ~TDF_UBORROWING;
912 		sched_user_prio(td, base_pri);
913 	} else {
914 		sched_lend_user_prio(td, prio);
915 	}
916 }
917 
918 void
919 sched_sleep(struct thread *td, int pri)
920 {
921 
922 	THREAD_LOCK_ASSERT(td, MA_OWNED);
923 	td->td_slptick = ticks;
924 	td->td_sched->ts_slptime = 0;
925 	if (pri)
926 		sched_prio(td, pri);
927 	if (TD_IS_SUSPENDED(td) || pri >= PSOCK)
928 		td->td_flags |= TDF_CANSWAP;
929 }
930 
931 void
932 sched_switch(struct thread *td, struct thread *newtd, int flags)
933 {
934 	struct mtx *tmtx;
935 	struct td_sched *ts;
936 	struct proc *p;
937 
938 	tmtx = NULL;
939 	ts = td->td_sched;
940 	p = td->td_proc;
941 
942 	THREAD_LOCK_ASSERT(td, MA_OWNED);
943 
944 	/*
945 	 * Switch to the sched lock to fix things up and pick
946 	 * a new thread.
947 	 * Block the td_lock in order to avoid breaking the critical path.
948 	 */
949 	if (td->td_lock != &sched_lock) {
950 		mtx_lock_spin(&sched_lock);
951 		tmtx = thread_lock_block(td);
952 	}
953 
954 	if ((td->td_flags & TDF_NOLOAD) == 0)
955 		sched_load_rem();
956 
957 	if (newtd) {
958 		MPASS(newtd->td_lock == &sched_lock);
959 		newtd->td_flags |= (td->td_flags & TDF_NEEDRESCHED);
960 	}
961 
962 	td->td_lastcpu = td->td_oncpu;
963 	td->td_flags &= ~TDF_NEEDRESCHED;
964 	td->td_owepreempt = 0;
965 	td->td_oncpu = NOCPU;
966 
967 	/*
968 	 * At the last moment, if this thread is still marked RUNNING,
969 	 * then put it back on the run queue as it has not been suspended
970 	 * or stopped or any thing else similar.  We never put the idle
971 	 * threads on the run queue, however.
972 	 */
973 	if (td->td_flags & TDF_IDLETD) {
974 		TD_SET_CAN_RUN(td);
975 #ifdef SMP
976 		idle_cpus_mask &= ~PCPU_GET(cpumask);
977 #endif
978 	} else {
979 		if (TD_IS_RUNNING(td)) {
980 			/* Put us back on the run queue. */
981 			sched_add(td, (flags & SW_PREEMPT) ?
982 			    SRQ_OURSELF|SRQ_YIELDING|SRQ_PREEMPTED :
983 			    SRQ_OURSELF|SRQ_YIELDING);
984 		}
985 	}
986 	if (newtd) {
987 		/*
988 		 * The thread we are about to run needs to be counted
989 		 * as if it had been added to the run queue and selected.
990 		 * It came from:
991 		 * * A preemption
992 		 * * An upcall
993 		 * * A followon
994 		 */
995 		KASSERT((newtd->td_inhibitors == 0),
996 			("trying to run inhibited thread"));
997 		newtd->td_flags |= TDF_DIDRUN;
998         	TD_SET_RUNNING(newtd);
999 		if ((newtd->td_flags & TDF_NOLOAD) == 0)
1000 			sched_load_add();
1001 	} else {
1002 		newtd = choosethread();
1003 		MPASS(newtd->td_lock == &sched_lock);
1004 	}
1005 
1006 	if (td != newtd) {
1007 #ifdef	HWPMC_HOOKS
1008 		if (PMC_PROC_IS_USING_PMCS(td->td_proc))
1009 			PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_OUT);
1010 #endif
1011                 /* I feel sleepy */
1012 		lock_profile_release_lock(&sched_lock.lock_object);
1013 #ifdef KDTRACE_HOOKS
1014 		/*
1015 		 * If DTrace has set the active vtime enum to anything
1016 		 * other than INACTIVE (0), then it should have set the
1017 		 * function to call.
1018 		 */
1019 		if (dtrace_vtime_active)
1020 			(*dtrace_vtime_switch_func)(newtd);
1021 #endif
1022 
1023 		cpu_switch(td, newtd, tmtx != NULL ? tmtx : td->td_lock);
1024 		lock_profile_obtain_lock_success(&sched_lock.lock_object,
1025 		    0, 0, __FILE__, __LINE__);
1026 		/*
1027 		 * Where am I?  What year is it?
1028 		 * We are in the same thread that went to sleep above,
1029 		 * but any amount of time may have passed. All our context
1030 		 * will still be available as will local variables.
1031 		 * PCPU values however may have changed as we may have
1032 		 * changed CPU so don't trust cached values of them.
1033 		 * New threads will go to fork_exit() instead of here
1034 		 * so if you change things here you may need to change
1035 		 * things there too.
1036 		 *
1037 		 * If the thread above was exiting it will never wake
1038 		 * up again here, so either it has saved everything it
1039 		 * needed to, or the thread_wait() or wait() will
1040 		 * need to reap it.
1041 		 */
1042 #ifdef	HWPMC_HOOKS
1043 		if (PMC_PROC_IS_USING_PMCS(td->td_proc))
1044 			PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_IN);
1045 #endif
1046 	}
1047 
1048 #ifdef SMP
1049 	if (td->td_flags & TDF_IDLETD)
1050 		idle_cpus_mask |= PCPU_GET(cpumask);
1051 #endif
1052 	sched_lock.mtx_lock = (uintptr_t)td;
1053 	td->td_oncpu = PCPU_GET(cpuid);
1054 	MPASS(td->td_lock == &sched_lock);
1055 }
1056 
1057 void
1058 sched_wakeup(struct thread *td)
1059 {
1060 	struct td_sched *ts;
1061 
1062 	THREAD_LOCK_ASSERT(td, MA_OWNED);
1063 	ts = td->td_sched;
1064 	td->td_flags &= ~TDF_CANSWAP;
1065 	if (ts->ts_slptime > 1) {
1066 		updatepri(td);
1067 		resetpriority(td);
1068 	}
1069 	td->td_slptick = 0;
1070 	ts->ts_slptime = 0;
1071 	sched_add(td, SRQ_BORING);
1072 }
1073 
1074 #ifdef SMP
1075 static int
1076 forward_wakeup(int cpunum)
1077 {
1078 	struct pcpu *pc;
1079 	cpumask_t dontuse, id, map, map2, map3, me;
1080 
1081 	mtx_assert(&sched_lock, MA_OWNED);
1082 
1083 	CTR0(KTR_RUNQ, "forward_wakeup()");
1084 
1085 	if ((!forward_wakeup_enabled) ||
1086 	     (forward_wakeup_use_mask == 0 && forward_wakeup_use_loop == 0))
1087 		return (0);
1088 	if (!smp_started || cold || panicstr)
1089 		return (0);
1090 
1091 	forward_wakeups_requested++;
1092 
1093 	/*
1094 	 * Check the idle mask we received against what we calculated
1095 	 * before in the old version.
1096 	 */
1097 	me = PCPU_GET(cpumask);
1098 
1099 	/* Don't bother if we should be doing it ourself. */
1100 	if ((me & idle_cpus_mask) && (cpunum == NOCPU || me == (1 << cpunum)))
1101 		return (0);
1102 
1103 	dontuse = me | stopped_cpus | hlt_cpus_mask;
1104 	map3 = 0;
1105 	if (forward_wakeup_use_loop) {
1106 		SLIST_FOREACH(pc, &cpuhead, pc_allcpu) {
1107 			id = pc->pc_cpumask;
1108 			if ((id & dontuse) == 0 &&
1109 			    pc->pc_curthread == pc->pc_idlethread) {
1110 				map3 |= id;
1111 			}
1112 		}
1113 	}
1114 
1115 	if (forward_wakeup_use_mask) {
1116 		map = 0;
1117 		map = idle_cpus_mask & ~dontuse;
1118 
1119 		/* If they are both on, compare and use loop if different. */
1120 		if (forward_wakeup_use_loop) {
1121 			if (map != map3) {
1122 				printf("map (%02X) != map3 (%02X)\n", map,
1123 				    map3);
1124 				map = map3;
1125 			}
1126 		}
1127 	} else {
1128 		map = map3;
1129 	}
1130 
1131 	/* If we only allow a specific CPU, then mask off all the others. */
1132 	if (cpunum != NOCPU) {
1133 		KASSERT((cpunum <= mp_maxcpus),("forward_wakeup: bad cpunum."));
1134 		map &= (1 << cpunum);
1135 	} else {
1136 		/* Try choose an idle die. */
1137 		if (forward_wakeup_use_htt) {
1138 			map2 =  (map & (map >> 1)) & 0x5555;
1139 			if (map2) {
1140 				map = map2;
1141 			}
1142 		}
1143 
1144 		/* Set only one bit. */
1145 		if (forward_wakeup_use_single) {
1146 			map = map & ((~map) + 1);
1147 		}
1148 	}
1149 	if (map) {
1150 		forward_wakeups_delivered++;
1151 		SLIST_FOREACH(pc, &cpuhead, pc_allcpu) {
1152 			id = pc->pc_cpumask;
1153 			if ((map & id) == 0)
1154 				continue;
1155 			if (cpu_idle_wakeup(pc->pc_cpuid))
1156 				map &= ~id;
1157 		}
1158 		if (map)
1159 			ipi_selected(map, IPI_AST);
1160 		return (1);
1161 	}
1162 	if (cpunum == NOCPU)
1163 		printf("forward_wakeup: Idle processor not found\n");
1164 	return (0);
1165 }
1166 
1167 static void
1168 kick_other_cpu(int pri, int cpuid)
1169 {
1170 	struct pcpu *pcpu;
1171 	int cpri;
1172 
1173 	pcpu = pcpu_find(cpuid);
1174 	if (idle_cpus_mask & pcpu->pc_cpumask) {
1175 		forward_wakeups_delivered++;
1176 		if (!cpu_idle_wakeup(cpuid))
1177 			ipi_cpu(cpuid, IPI_AST);
1178 		return;
1179 	}
1180 
1181 	cpri = pcpu->pc_curthread->td_priority;
1182 	if (pri >= cpri)
1183 		return;
1184 
1185 #if defined(IPI_PREEMPTION) && defined(PREEMPTION)
1186 #if !defined(FULL_PREEMPTION)
1187 	if (pri <= PRI_MAX_ITHD)
1188 #endif /* ! FULL_PREEMPTION */
1189 	{
1190 		ipi_cpu(cpuid, IPI_PREEMPT);
1191 		return;
1192 	}
1193 #endif /* defined(IPI_PREEMPTION) && defined(PREEMPTION) */
1194 
1195 	pcpu->pc_curthread->td_flags |= TDF_NEEDRESCHED;
1196 	ipi_cpu(cpuid, IPI_AST);
1197 	return;
1198 }
1199 #endif /* SMP */
1200 
1201 #ifdef SMP
1202 static int
1203 sched_pickcpu(struct thread *td)
1204 {
1205 	int best, cpu;
1206 
1207 	mtx_assert(&sched_lock, MA_OWNED);
1208 
1209 	if (THREAD_CAN_SCHED(td, td->td_lastcpu))
1210 		best = td->td_lastcpu;
1211 	else
1212 		best = NOCPU;
1213 	CPU_FOREACH(cpu) {
1214 		if (!THREAD_CAN_SCHED(td, cpu))
1215 			continue;
1216 
1217 		if (best == NOCPU)
1218 			best = cpu;
1219 		else if (runq_length[cpu] < runq_length[best])
1220 			best = cpu;
1221 	}
1222 	KASSERT(best != NOCPU, ("no valid CPUs"));
1223 
1224 	return (best);
1225 }
1226 #endif
1227 
1228 void
1229 sched_add(struct thread *td, int flags)
1230 #ifdef SMP
1231 {
1232 	struct td_sched *ts;
1233 	int forwarded = 0;
1234 	int cpu;
1235 	int single_cpu = 0;
1236 
1237 	ts = td->td_sched;
1238 	THREAD_LOCK_ASSERT(td, MA_OWNED);
1239 	KASSERT((td->td_inhibitors == 0),
1240 	    ("sched_add: trying to run inhibited thread"));
1241 	KASSERT((TD_CAN_RUN(td) || TD_IS_RUNNING(td)),
1242 	    ("sched_add: bad thread state"));
1243 	KASSERT(td->td_flags & TDF_INMEM,
1244 	    ("sched_add: thread swapped out"));
1245 
1246 	KTR_STATE2(KTR_SCHED, "thread", sched_tdname(td), "runq add",
1247 	    "prio:%d", td->td_priority, KTR_ATTR_LINKED,
1248 	    sched_tdname(curthread));
1249 	KTR_POINT1(KTR_SCHED, "thread", sched_tdname(curthread), "wokeup",
1250 	    KTR_ATTR_LINKED, sched_tdname(td));
1251 
1252 
1253 	/*
1254 	 * Now that the thread is moving to the run-queue, set the lock
1255 	 * to the scheduler's lock.
1256 	 */
1257 	if (td->td_lock != &sched_lock) {
1258 		mtx_lock_spin(&sched_lock);
1259 		thread_lock_set(td, &sched_lock);
1260 	}
1261 	TD_SET_RUNQ(td);
1262 
1263 	if (td->td_pinned != 0) {
1264 		cpu = td->td_lastcpu;
1265 		ts->ts_runq = &runq_pcpu[cpu];
1266 		single_cpu = 1;
1267 		CTR3(KTR_RUNQ,
1268 		    "sched_add: Put td_sched:%p(td:%p) on cpu%d runq", ts, td,
1269 		    cpu);
1270 	} else if (td->td_flags & TDF_BOUND) {
1271 		/* Find CPU from bound runq. */
1272 		KASSERT(SKE_RUNQ_PCPU(ts),
1273 		    ("sched_add: bound td_sched not on cpu runq"));
1274 		cpu = ts->ts_runq - &runq_pcpu[0];
1275 		single_cpu = 1;
1276 		CTR3(KTR_RUNQ,
1277 		    "sched_add: Put td_sched:%p(td:%p) on cpu%d runq", ts, td,
1278 		    cpu);
1279 	} else if (ts->ts_flags & TSF_AFFINITY) {
1280 		/* Find a valid CPU for our cpuset */
1281 		cpu = sched_pickcpu(td);
1282 		ts->ts_runq = &runq_pcpu[cpu];
1283 		single_cpu = 1;
1284 		CTR3(KTR_RUNQ,
1285 		    "sched_add: Put td_sched:%p(td:%p) on cpu%d runq", ts, td,
1286 		    cpu);
1287 	} else {
1288 		CTR2(KTR_RUNQ,
1289 		    "sched_add: adding td_sched:%p (td:%p) to gbl runq", ts,
1290 		    td);
1291 		cpu = NOCPU;
1292 		ts->ts_runq = &runq;
1293 	}
1294 
1295 	if (single_cpu && (cpu != PCPU_GET(cpuid))) {
1296 	        kick_other_cpu(td->td_priority, cpu);
1297 	} else {
1298 		if (!single_cpu) {
1299 			cpumask_t me = PCPU_GET(cpumask);
1300 			cpumask_t idle = idle_cpus_mask & me;
1301 
1302 			if (!idle && ((flags & SRQ_INTR) == 0) &&
1303 			    (idle_cpus_mask & ~(hlt_cpus_mask | me)))
1304 				forwarded = forward_wakeup(cpu);
1305 		}
1306 
1307 		if (!forwarded) {
1308 			if ((flags & SRQ_YIELDING) == 0 && maybe_preempt(td))
1309 				return;
1310 			else
1311 				maybe_resched(td);
1312 		}
1313 	}
1314 
1315 	if ((td->td_flags & TDF_NOLOAD) == 0)
1316 		sched_load_add();
1317 	runq_add(ts->ts_runq, td, flags);
1318 	if (cpu != NOCPU)
1319 		runq_length[cpu]++;
1320 }
1321 #else /* SMP */
1322 {
1323 	struct td_sched *ts;
1324 
1325 	ts = td->td_sched;
1326 	THREAD_LOCK_ASSERT(td, MA_OWNED);
1327 	KASSERT((td->td_inhibitors == 0),
1328 	    ("sched_add: trying to run inhibited thread"));
1329 	KASSERT((TD_CAN_RUN(td) || TD_IS_RUNNING(td)),
1330 	    ("sched_add: bad thread state"));
1331 	KASSERT(td->td_flags & TDF_INMEM,
1332 	    ("sched_add: thread swapped out"));
1333 	KTR_STATE2(KTR_SCHED, "thread", sched_tdname(td), "runq add",
1334 	    "prio:%d", td->td_priority, KTR_ATTR_LINKED,
1335 	    sched_tdname(curthread));
1336 	KTR_POINT1(KTR_SCHED, "thread", sched_tdname(curthread), "wokeup",
1337 	    KTR_ATTR_LINKED, sched_tdname(td));
1338 
1339 	/*
1340 	 * Now that the thread is moving to the run-queue, set the lock
1341 	 * to the scheduler's lock.
1342 	 */
1343 	if (td->td_lock != &sched_lock) {
1344 		mtx_lock_spin(&sched_lock);
1345 		thread_lock_set(td, &sched_lock);
1346 	}
1347 	TD_SET_RUNQ(td);
1348 	CTR2(KTR_RUNQ, "sched_add: adding td_sched:%p (td:%p) to runq", ts, td);
1349 	ts->ts_runq = &runq;
1350 
1351 	/*
1352 	 * If we are yielding (on the way out anyhow) or the thread
1353 	 * being saved is US, then don't try be smart about preemption
1354 	 * or kicking off another CPU as it won't help and may hinder.
1355 	 * In the YIEDLING case, we are about to run whoever is being
1356 	 * put in the queue anyhow, and in the OURSELF case, we are
1357 	 * puting ourself on the run queue which also only happens
1358 	 * when we are about to yield.
1359 	 */
1360 	if ((flags & SRQ_YIELDING) == 0) {
1361 		if (maybe_preempt(td))
1362 			return;
1363 	}
1364 	if ((td->td_flags & TDF_NOLOAD) == 0)
1365 		sched_load_add();
1366 	runq_add(ts->ts_runq, td, flags);
1367 	maybe_resched(td);
1368 }
1369 #endif /* SMP */
1370 
1371 void
1372 sched_rem(struct thread *td)
1373 {
1374 	struct td_sched *ts;
1375 
1376 	ts = td->td_sched;
1377 	KASSERT(td->td_flags & TDF_INMEM,
1378 	    ("sched_rem: thread swapped out"));
1379 	KASSERT(TD_ON_RUNQ(td),
1380 	    ("sched_rem: thread not on run queue"));
1381 	mtx_assert(&sched_lock, MA_OWNED);
1382 	KTR_STATE2(KTR_SCHED, "thread", sched_tdname(td), "runq rem",
1383 	    "prio:%d", td->td_priority, KTR_ATTR_LINKED,
1384 	    sched_tdname(curthread));
1385 
1386 	if ((td->td_flags & TDF_NOLOAD) == 0)
1387 		sched_load_rem();
1388 #ifdef SMP
1389 	if (ts->ts_runq != &runq)
1390 		runq_length[ts->ts_runq - runq_pcpu]--;
1391 #endif
1392 	runq_remove(ts->ts_runq, td);
1393 	TD_SET_CAN_RUN(td);
1394 }
1395 
1396 /*
1397  * Select threads to run.  Note that running threads still consume a
1398  * slot.
1399  */
1400 struct thread *
1401 sched_choose(void)
1402 {
1403 	struct thread *td;
1404 	struct runq *rq;
1405 
1406 	mtx_assert(&sched_lock,  MA_OWNED);
1407 #ifdef SMP
1408 	struct thread *tdcpu;
1409 
1410 	rq = &runq;
1411 	td = runq_choose_fuzz(&runq, runq_fuzz);
1412 	tdcpu = runq_choose(&runq_pcpu[PCPU_GET(cpuid)]);
1413 
1414 	if (td == NULL ||
1415 	    (tdcpu != NULL &&
1416 	     tdcpu->td_priority < td->td_priority)) {
1417 		CTR2(KTR_RUNQ, "choosing td %p from pcpu runq %d", tdcpu,
1418 		     PCPU_GET(cpuid));
1419 		td = tdcpu;
1420 		rq = &runq_pcpu[PCPU_GET(cpuid)];
1421 	} else {
1422 		CTR1(KTR_RUNQ, "choosing td_sched %p from main runq", td);
1423 	}
1424 
1425 #else
1426 	rq = &runq;
1427 	td = runq_choose(&runq);
1428 #endif
1429 
1430 	if (td) {
1431 #ifdef SMP
1432 		if (td == tdcpu)
1433 			runq_length[PCPU_GET(cpuid)]--;
1434 #endif
1435 		runq_remove(rq, td);
1436 		td->td_flags |= TDF_DIDRUN;
1437 
1438 		KASSERT(td->td_flags & TDF_INMEM,
1439 		    ("sched_choose: thread swapped out"));
1440 		return (td);
1441 	}
1442 	return (PCPU_GET(idlethread));
1443 }
1444 
1445 void
1446 sched_preempt(struct thread *td)
1447 {
1448 	thread_lock(td);
1449 	if (td->td_critnest > 1)
1450 		td->td_owepreempt = 1;
1451 	else
1452 		mi_switch(SW_INVOL | SW_PREEMPT | SWT_PREEMPT, NULL);
1453 	thread_unlock(td);
1454 }
1455 
1456 void
1457 sched_userret(struct thread *td)
1458 {
1459 	/*
1460 	 * XXX we cheat slightly on the locking here to avoid locking in
1461 	 * the usual case.  Setting td_priority here is essentially an
1462 	 * incomplete workaround for not setting it properly elsewhere.
1463 	 * Now that some interrupt handlers are threads, not setting it
1464 	 * properly elsewhere can clobber it in the window between setting
1465 	 * it here and returning to user mode, so don't waste time setting
1466 	 * it perfectly here.
1467 	 */
1468 	KASSERT((td->td_flags & TDF_BORROWING) == 0,
1469 	    ("thread with borrowed priority returning to userland"));
1470 	if (td->td_priority != td->td_user_pri) {
1471 		thread_lock(td);
1472 		td->td_priority = td->td_user_pri;
1473 		td->td_base_pri = td->td_user_pri;
1474 		thread_unlock(td);
1475 	}
1476 }
1477 
1478 void
1479 sched_bind(struct thread *td, int cpu)
1480 {
1481 	struct td_sched *ts;
1482 
1483 	THREAD_LOCK_ASSERT(td, MA_OWNED|MA_NOTRECURSED);
1484 	KASSERT(td == curthread, ("sched_bind: can only bind curthread"));
1485 
1486 	ts = td->td_sched;
1487 
1488 	td->td_flags |= TDF_BOUND;
1489 #ifdef SMP
1490 	ts->ts_runq = &runq_pcpu[cpu];
1491 	if (PCPU_GET(cpuid) == cpu)
1492 		return;
1493 
1494 	mi_switch(SW_VOL, NULL);
1495 #endif
1496 }
1497 
1498 void
1499 sched_unbind(struct thread* td)
1500 {
1501 	THREAD_LOCK_ASSERT(td, MA_OWNED);
1502 	KASSERT(td == curthread, ("sched_unbind: can only bind curthread"));
1503 	td->td_flags &= ~TDF_BOUND;
1504 }
1505 
1506 int
1507 sched_is_bound(struct thread *td)
1508 {
1509 	THREAD_LOCK_ASSERT(td, MA_OWNED);
1510 	return (td->td_flags & TDF_BOUND);
1511 }
1512 
1513 void
1514 sched_relinquish(struct thread *td)
1515 {
1516 	thread_lock(td);
1517 	mi_switch(SW_VOL | SWT_RELINQUISH, NULL);
1518 	thread_unlock(td);
1519 }
1520 
1521 int
1522 sched_load(void)
1523 {
1524 	return (sched_tdcnt);
1525 }
1526 
1527 int
1528 sched_sizeof_proc(void)
1529 {
1530 	return (sizeof(struct proc));
1531 }
1532 
1533 int
1534 sched_sizeof_thread(void)
1535 {
1536 	return (sizeof(struct thread) + sizeof(struct td_sched));
1537 }
1538 
1539 fixpt_t
1540 sched_pctcpu(struct thread *td)
1541 {
1542 	struct td_sched *ts;
1543 
1544 	THREAD_LOCK_ASSERT(td, MA_OWNED);
1545 	ts = td->td_sched;
1546 	return (ts->ts_pctcpu);
1547 }
1548 
1549 void
1550 sched_tick(int cnt)
1551 {
1552 }
1553 
1554 /*
1555  * The actual idle process.
1556  */
1557 void
1558 sched_idletd(void *dummy)
1559 {
1560 	struct pcpuidlestat *stat;
1561 
1562 	stat = DPCPU_PTR(idlestat);
1563 	for (;;) {
1564 		mtx_assert(&Giant, MA_NOTOWNED);
1565 
1566 		while (sched_runnable() == 0) {
1567 			cpu_idle(stat->idlecalls + stat->oldidlecalls > 64);
1568 			stat->idlecalls++;
1569 		}
1570 
1571 		mtx_lock_spin(&sched_lock);
1572 		mi_switch(SW_VOL | SWT_IDLE, NULL);
1573 		mtx_unlock_spin(&sched_lock);
1574 	}
1575 }
1576 
1577 /*
1578  * A CPU is entering for the first time or a thread is exiting.
1579  */
1580 void
1581 sched_throw(struct thread *td)
1582 {
1583 	/*
1584 	 * Correct spinlock nesting.  The idle thread context that we are
1585 	 * borrowing was created so that it would start out with a single
1586 	 * spin lock (sched_lock) held in fork_trampoline().  Since we've
1587 	 * explicitly acquired locks in this function, the nesting count
1588 	 * is now 2 rather than 1.  Since we are nested, calling
1589 	 * spinlock_exit() will simply adjust the counts without allowing
1590 	 * spin lock using code to interrupt us.
1591 	 */
1592 	if (td == NULL) {
1593 		mtx_lock_spin(&sched_lock);
1594 		spinlock_exit();
1595 	} else {
1596 		lock_profile_release_lock(&sched_lock.lock_object);
1597 		MPASS(td->td_lock == &sched_lock);
1598 	}
1599 	mtx_assert(&sched_lock, MA_OWNED);
1600 	KASSERT(curthread->td_md.md_spinlock_count == 1, ("invalid count"));
1601 	PCPU_SET(switchtime, cpu_ticks());
1602 	PCPU_SET(switchticks, ticks);
1603 	cpu_throw(td, choosethread());	/* doesn't return */
1604 }
1605 
1606 void
1607 sched_fork_exit(struct thread *td)
1608 {
1609 
1610 	/*
1611 	 * Finish setting up thread glue so that it begins execution in a
1612 	 * non-nested critical section with sched_lock held but not recursed.
1613 	 */
1614 	td->td_oncpu = PCPU_GET(cpuid);
1615 	sched_lock.mtx_lock = (uintptr_t)td;
1616 	lock_profile_obtain_lock_success(&sched_lock.lock_object,
1617 	    0, 0, __FILE__, __LINE__);
1618 	THREAD_LOCK_ASSERT(td, MA_OWNED | MA_NOTRECURSED);
1619 }
1620 
1621 char *
1622 sched_tdname(struct thread *td)
1623 {
1624 #ifdef KTR
1625 	struct td_sched *ts;
1626 
1627 	ts = td->td_sched;
1628 	if (ts->ts_name[0] == '\0')
1629 		snprintf(ts->ts_name, sizeof(ts->ts_name),
1630 		    "%s tid %d", td->td_name, td->td_tid);
1631 	return (ts->ts_name);
1632 #else
1633 	return (td->td_name);
1634 #endif
1635 }
1636 
1637 void
1638 sched_affinity(struct thread *td)
1639 {
1640 #ifdef SMP
1641 	struct td_sched *ts;
1642 	int cpu;
1643 
1644 	THREAD_LOCK_ASSERT(td, MA_OWNED);
1645 
1646 	/*
1647 	 * Set the TSF_AFFINITY flag if there is at least one CPU this
1648 	 * thread can't run on.
1649 	 */
1650 	ts = td->td_sched;
1651 	ts->ts_flags &= ~TSF_AFFINITY;
1652 	CPU_FOREACH(cpu) {
1653 		if (!THREAD_CAN_SCHED(td, cpu)) {
1654 			ts->ts_flags |= TSF_AFFINITY;
1655 			break;
1656 		}
1657 	}
1658 
1659 	/*
1660 	 * If this thread can run on all CPUs, nothing else to do.
1661 	 */
1662 	if (!(ts->ts_flags & TSF_AFFINITY))
1663 		return;
1664 
1665 	/* Pinned threads and bound threads should be left alone. */
1666 	if (td->td_pinned != 0 || td->td_flags & TDF_BOUND)
1667 		return;
1668 
1669 	switch (td->td_state) {
1670 	case TDS_RUNQ:
1671 		/*
1672 		 * If we are on a per-CPU runqueue that is in the set,
1673 		 * then nothing needs to be done.
1674 		 */
1675 		if (ts->ts_runq != &runq &&
1676 		    THREAD_CAN_SCHED(td, ts->ts_runq - runq_pcpu))
1677 			return;
1678 
1679 		/* Put this thread on a valid per-CPU runqueue. */
1680 		sched_rem(td);
1681 		sched_add(td, SRQ_BORING);
1682 		break;
1683 	case TDS_RUNNING:
1684 		/*
1685 		 * See if our current CPU is in the set.  If not, force a
1686 		 * context switch.
1687 		 */
1688 		if (THREAD_CAN_SCHED(td, td->td_oncpu))
1689 			return;
1690 
1691 		td->td_flags |= TDF_NEEDRESCHED;
1692 		if (td != curthread)
1693 			ipi_cpu(cpu, IPI_AST);
1694 		break;
1695 	default:
1696 		break;
1697 	}
1698 #endif
1699 }
1700