xref: /freebsd/sys/kern/sched_4bsd.c (revision 830940567b49bb0c08dfaed40418999e76616909)
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 static void
161 setup_runqs(void)
162 {
163 #ifdef SMP
164 	int i;
165 
166 	for (i = 0; i < MAXCPU; ++i)
167 		runq_init(&runq_pcpu[i]);
168 #endif
169 
170 	runq_init(&runq);
171 }
172 
173 static int
174 sysctl_kern_quantum(SYSCTL_HANDLER_ARGS)
175 {
176 	int error, new_val;
177 
178 	new_val = sched_quantum * tick;
179 	error = sysctl_handle_int(oidp, &new_val, 0, req);
180         if (error != 0 || req->newptr == NULL)
181 		return (error);
182 	if (new_val < tick)
183 		return (EINVAL);
184 	sched_quantum = new_val / tick;
185 	hogticks = 2 * sched_quantum;
186 	return (0);
187 }
188 
189 SYSCTL_NODE(_kern, OID_AUTO, sched, CTLFLAG_RD, 0, "Scheduler");
190 
191 SYSCTL_STRING(_kern_sched, OID_AUTO, name, CTLFLAG_RD, "4BSD", 0,
192     "Scheduler name");
193 
194 SYSCTL_PROC(_kern_sched, OID_AUTO, quantum, CTLTYPE_INT | CTLFLAG_RW,
195     0, sizeof sched_quantum, sysctl_kern_quantum, "I",
196     "Roundrobin scheduling quantum in microseconds");
197 
198 #ifdef SMP
199 /* Enable forwarding of wakeups to all other cpus */
200 SYSCTL_NODE(_kern_sched, OID_AUTO, ipiwakeup, CTLFLAG_RD, NULL, "Kernel SMP");
201 
202 static int runq_fuzz = 1;
203 SYSCTL_INT(_kern_sched, OID_AUTO, runq_fuzz, CTLFLAG_RW, &runq_fuzz, 0, "");
204 
205 static int forward_wakeup_enabled = 1;
206 SYSCTL_INT(_kern_sched_ipiwakeup, OID_AUTO, enabled, CTLFLAG_RW,
207 	   &forward_wakeup_enabled, 0,
208 	   "Forwarding of wakeup to idle CPUs");
209 
210 static int forward_wakeups_requested = 0;
211 SYSCTL_INT(_kern_sched_ipiwakeup, OID_AUTO, requested, CTLFLAG_RD,
212 	   &forward_wakeups_requested, 0,
213 	   "Requests for Forwarding of wakeup to idle CPUs");
214 
215 static int forward_wakeups_delivered = 0;
216 SYSCTL_INT(_kern_sched_ipiwakeup, OID_AUTO, delivered, CTLFLAG_RD,
217 	   &forward_wakeups_delivered, 0,
218 	   "Completed Forwarding of wakeup to idle CPUs");
219 
220 static int forward_wakeup_use_mask = 1;
221 SYSCTL_INT(_kern_sched_ipiwakeup, OID_AUTO, usemask, CTLFLAG_RW,
222 	   &forward_wakeup_use_mask, 0,
223 	   "Use the mask of idle cpus");
224 
225 static int forward_wakeup_use_loop = 0;
226 SYSCTL_INT(_kern_sched_ipiwakeup, OID_AUTO, useloop, CTLFLAG_RW,
227 	   &forward_wakeup_use_loop, 0,
228 	   "Use a loop to find idle cpus");
229 
230 static int forward_wakeup_use_single = 0;
231 SYSCTL_INT(_kern_sched_ipiwakeup, OID_AUTO, onecpu, CTLFLAG_RW,
232 	   &forward_wakeup_use_single, 0,
233 	   "Only signal one idle cpu");
234 
235 static int forward_wakeup_use_htt = 0;
236 SYSCTL_INT(_kern_sched_ipiwakeup, OID_AUTO, htt2, CTLFLAG_RW,
237 	   &forward_wakeup_use_htt, 0,
238 	   "account for htt");
239 
240 #endif
241 #if 0
242 static int sched_followon = 0;
243 SYSCTL_INT(_kern_sched, OID_AUTO, followon, CTLFLAG_RW,
244 	   &sched_followon, 0,
245 	   "allow threads to share a quantum");
246 #endif
247 
248 static __inline void
249 sched_load_add(void)
250 {
251 
252 	sched_tdcnt++;
253 	KTR_COUNTER0(KTR_SCHED, "load", "global load", sched_tdcnt);
254 }
255 
256 static __inline void
257 sched_load_rem(void)
258 {
259 
260 	sched_tdcnt--;
261 	KTR_COUNTER0(KTR_SCHED, "load", "global load", sched_tdcnt);
262 }
263 /*
264  * Arrange to reschedule if necessary, taking the priorities and
265  * schedulers into account.
266  */
267 static void
268 maybe_resched(struct thread *td)
269 {
270 
271 	THREAD_LOCK_ASSERT(td, MA_OWNED);
272 	if (td->td_priority < curthread->td_priority)
273 		curthread->td_flags |= TDF_NEEDRESCHED;
274 }
275 
276 /*
277  * This function is called when a thread is about to be put on run queue
278  * because it has been made runnable or its priority has been adjusted.  It
279  * determines if the new thread should be immediately preempted to.  If so,
280  * it switches to it and eventually returns true.  If not, it returns false
281  * so that the caller may place the thread on an appropriate run queue.
282  */
283 int
284 maybe_preempt(struct thread *td)
285 {
286 #ifdef PREEMPTION
287 	struct thread *ctd;
288 	int cpri, pri;
289 
290 	/*
291 	 * The new thread should not preempt the current thread if any of the
292 	 * following conditions are true:
293 	 *
294 	 *  - The kernel is in the throes of crashing (panicstr).
295 	 *  - The current thread has a higher (numerically lower) or
296 	 *    equivalent priority.  Note that this prevents curthread from
297 	 *    trying to preempt to itself.
298 	 *  - It is too early in the boot for context switches (cold is set).
299 	 *  - The current thread has an inhibitor set or is in the process of
300 	 *    exiting.  In this case, the current thread is about to switch
301 	 *    out anyways, so there's no point in preempting.  If we did,
302 	 *    the current thread would not be properly resumed as well, so
303 	 *    just avoid that whole landmine.
304 	 *  - If the new thread's priority is not a realtime priority and
305 	 *    the current thread's priority is not an idle priority and
306 	 *    FULL_PREEMPTION is disabled.
307 	 *
308 	 * If all of these conditions are false, but the current thread is in
309 	 * a nested critical section, then we have to defer the preemption
310 	 * until we exit the critical section.  Otherwise, switch immediately
311 	 * to the new thread.
312 	 */
313 	ctd = curthread;
314 	THREAD_LOCK_ASSERT(td, MA_OWNED);
315 	KASSERT((td->td_inhibitors == 0),
316 			("maybe_preempt: trying to run inhibited thread"));
317 	pri = td->td_priority;
318 	cpri = ctd->td_priority;
319 	if (panicstr != NULL || pri >= cpri || cold /* || dumping */ ||
320 	    TD_IS_INHIBITED(ctd))
321 		return (0);
322 #ifndef FULL_PREEMPTION
323 	if (pri > PRI_MAX_ITHD && cpri < PRI_MIN_IDLE)
324 		return (0);
325 #endif
326 
327 	if (ctd->td_critnest > 1) {
328 		CTR1(KTR_PROC, "maybe_preempt: in critical section %d",
329 		    ctd->td_critnest);
330 		ctd->td_owepreempt = 1;
331 		return (0);
332 	}
333 	/*
334 	 * Thread is runnable but not yet put on system run queue.
335 	 */
336 	MPASS(ctd->td_lock == td->td_lock);
337 	MPASS(TD_ON_RUNQ(td));
338 	TD_SET_RUNNING(td);
339 	CTR3(KTR_PROC, "preempting to thread %p (pid %d, %s)\n", td,
340 	    td->td_proc->p_pid, td->td_name);
341 	mi_switch(SW_INVOL | SW_PREEMPT | SWT_PREEMPT, td);
342 	/*
343 	 * td's lock pointer may have changed.  We have to return with it
344 	 * locked.
345 	 */
346 	spinlock_enter();
347 	thread_unlock(ctd);
348 	thread_lock(td);
349 	spinlock_exit();
350 	return (1);
351 #else
352 	return (0);
353 #endif
354 }
355 
356 /*
357  * Constants for digital decay and forget:
358  *	90% of (td_estcpu) usage in 5 * loadav time
359  *	95% of (ts_pctcpu) usage in 60 seconds (load insensitive)
360  *          Note that, as ps(1) mentions, this can let percentages
361  *          total over 100% (I've seen 137.9% for 3 processes).
362  *
363  * Note that schedclock() updates td_estcpu and p_cpticks asynchronously.
364  *
365  * We wish to decay away 90% of td_estcpu in (5 * loadavg) seconds.
366  * That is, the system wants to compute a value of decay such
367  * that the following for loop:
368  * 	for (i = 0; i < (5 * loadavg); i++)
369  * 		td_estcpu *= decay;
370  * will compute
371  * 	td_estcpu *= 0.1;
372  * for all values of loadavg:
373  *
374  * Mathematically this loop can be expressed by saying:
375  * 	decay ** (5 * loadavg) ~= .1
376  *
377  * The system computes decay as:
378  * 	decay = (2 * loadavg) / (2 * loadavg + 1)
379  *
380  * We wish to prove that the system's computation of decay
381  * will always fulfill the equation:
382  * 	decay ** (5 * loadavg) ~= .1
383  *
384  * If we compute b as:
385  * 	b = 2 * loadavg
386  * then
387  * 	decay = b / (b + 1)
388  *
389  * We now need to prove two things:
390  *	1) Given factor ** (5 * loadavg) ~= .1, prove factor == b/(b+1)
391  *	2) Given b/(b+1) ** power ~= .1, prove power == (5 * loadavg)
392  *
393  * Facts:
394  *         For x close to zero, exp(x) =~ 1 + x, since
395  *              exp(x) = 0! + x**1/1! + x**2/2! + ... .
396  *              therefore exp(-1/b) =~ 1 - (1/b) = (b-1)/b.
397  *         For x close to zero, ln(1+x) =~ x, since
398  *              ln(1+x) = x - x**2/2 + x**3/3 - ...     -1 < x < 1
399  *              therefore ln(b/(b+1)) = ln(1 - 1/(b+1)) =~ -1/(b+1).
400  *         ln(.1) =~ -2.30
401  *
402  * Proof of (1):
403  *    Solve (factor)**(power) =~ .1 given power (5*loadav):
404  *	solving for factor,
405  *      ln(factor) =~ (-2.30/5*loadav), or
406  *      factor =~ exp(-1/((5/2.30)*loadav)) =~ exp(-1/(2*loadav)) =
407  *          exp(-1/b) =~ (b-1)/b =~ b/(b+1).                    QED
408  *
409  * Proof of (2):
410  *    Solve (factor)**(power) =~ .1 given factor == (b/(b+1)):
411  *	solving for power,
412  *      power*ln(b/(b+1)) =~ -2.30, or
413  *      power =~ 2.3 * (b + 1) = 4.6*loadav + 2.3 =~ 5*loadav.  QED
414  *
415  * Actual power values for the implemented algorithm are as follows:
416  *      loadav: 1       2       3       4
417  *      power:  5.68    10.32   14.94   19.55
418  */
419 
420 /* calculations for digital decay to forget 90% of usage in 5*loadav sec */
421 #define	loadfactor(loadav)	(2 * (loadav))
422 #define	decay_cpu(loadfac, cpu)	(((loadfac) * (cpu)) / ((loadfac) + FSCALE))
423 
424 /* decay 95% of `ts_pctcpu' in 60 seconds; see CCPU_SHIFT before changing */
425 static fixpt_t	ccpu = 0.95122942450071400909 * FSCALE;	/* exp(-1/20) */
426 SYSCTL_INT(_kern, OID_AUTO, ccpu, CTLFLAG_RD, &ccpu, 0, "");
427 
428 /*
429  * If `ccpu' is not equal to `exp(-1/20)' and you still want to use the
430  * faster/more-accurate formula, you'll have to estimate CCPU_SHIFT below
431  * and possibly adjust FSHIFT in "param.h" so that (FSHIFT >= CCPU_SHIFT).
432  *
433  * To estimate CCPU_SHIFT for exp(-1/20), the following formula was used:
434  *	1 - exp(-1/20) ~= 0.0487 ~= 0.0488 == 1 (fixed pt, *11* bits).
435  *
436  * If you don't want to bother with the faster/more-accurate formula, you
437  * can set CCPU_SHIFT to (FSHIFT + 1) which will use a slower/less-accurate
438  * (more general) method of calculating the %age of CPU used by a process.
439  */
440 #define	CCPU_SHIFT	11
441 
442 /*
443  * Recompute process priorities, every hz ticks.
444  * MP-safe, called without the Giant mutex.
445  */
446 /* ARGSUSED */
447 static void
448 schedcpu(void)
449 {
450 	register fixpt_t loadfac = loadfactor(averunnable.ldavg[0]);
451 	struct thread *td;
452 	struct proc *p;
453 	struct td_sched *ts;
454 	int awake, realstathz;
455 
456 	realstathz = stathz ? stathz : hz;
457 	sx_slock(&allproc_lock);
458 	FOREACH_PROC_IN_SYSTEM(p) {
459 		PROC_LOCK(p);
460 		FOREACH_THREAD_IN_PROC(p, td) {
461 			awake = 0;
462 			thread_lock(td);
463 			ts = td->td_sched;
464 			/*
465 			 * Increment sleep time (if sleeping).  We
466 			 * ignore overflow, as above.
467 			 */
468 			/*
469 			 * The td_sched slptimes are not touched in wakeup
470 			 * because the thread may not HAVE everything in
471 			 * memory? XXX I think this is out of date.
472 			 */
473 			if (TD_ON_RUNQ(td)) {
474 				awake = 1;
475 				td->td_flags &= ~TDF_DIDRUN;
476 			} else if (TD_IS_RUNNING(td)) {
477 				awake = 1;
478 				/* Do not clear TDF_DIDRUN */
479 			} else if (td->td_flags & TDF_DIDRUN) {
480 				awake = 1;
481 				td->td_flags &= ~TDF_DIDRUN;
482 			}
483 
484 			/*
485 			 * ts_pctcpu is only for ps and ttyinfo().
486 			 */
487 			ts->ts_pctcpu = (ts->ts_pctcpu * ccpu) >> FSHIFT;
488 			/*
489 			 * If the td_sched has been idle the entire second,
490 			 * stop recalculating its priority until
491 			 * it wakes up.
492 			 */
493 			if (ts->ts_cpticks != 0) {
494 #if	(FSHIFT >= CCPU_SHIFT)
495 				ts->ts_pctcpu += (realstathz == 100)
496 				    ? ((fixpt_t) ts->ts_cpticks) <<
497 				    (FSHIFT - CCPU_SHIFT) :
498 				    100 * (((fixpt_t) ts->ts_cpticks)
499 				    << (FSHIFT - CCPU_SHIFT)) / realstathz;
500 #else
501 				ts->ts_pctcpu += ((FSCALE - ccpu) *
502 				    (ts->ts_cpticks *
503 				    FSCALE / realstathz)) >> FSHIFT;
504 #endif
505 				ts->ts_cpticks = 0;
506 			}
507 			/*
508 			 * If there are ANY running threads in this process,
509 			 * then don't count it as sleeping.
510 			 * XXX: this is broken.
511 			 */
512 			if (awake) {
513 				if (ts->ts_slptime > 1) {
514 					/*
515 					 * In an ideal world, this should not
516 					 * happen, because whoever woke us
517 					 * up from the long sleep should have
518 					 * unwound the slptime and reset our
519 					 * priority before we run at the stale
520 					 * priority.  Should KASSERT at some
521 					 * point when all the cases are fixed.
522 					 */
523 					updatepri(td);
524 				}
525 				ts->ts_slptime = 0;
526 			} else
527 				ts->ts_slptime++;
528 			if (ts->ts_slptime > 1) {
529 				thread_unlock(td);
530 				continue;
531 			}
532 			td->td_estcpu = decay_cpu(loadfac, td->td_estcpu);
533 		      	resetpriority(td);
534 			resetpriority_thread(td);
535 			thread_unlock(td);
536 		}
537 		PROC_UNLOCK(p);
538 	}
539 	sx_sunlock(&allproc_lock);
540 }
541 
542 /*
543  * Main loop for a kthread that executes schedcpu once a second.
544  */
545 static void
546 schedcpu_thread(void)
547 {
548 
549 	for (;;) {
550 		schedcpu();
551 		pause("-", hz);
552 	}
553 }
554 
555 /*
556  * Recalculate the priority of a process after it has slept for a while.
557  * For all load averages >= 1 and max td_estcpu of 255, sleeping for at
558  * least six times the loadfactor will decay td_estcpu to zero.
559  */
560 static void
561 updatepri(struct thread *td)
562 {
563 	struct td_sched *ts;
564 	fixpt_t loadfac;
565 	unsigned int newcpu;
566 
567 	ts = td->td_sched;
568 	loadfac = loadfactor(averunnable.ldavg[0]);
569 	if (ts->ts_slptime > 5 * loadfac)
570 		td->td_estcpu = 0;
571 	else {
572 		newcpu = td->td_estcpu;
573 		ts->ts_slptime--;	/* was incremented in schedcpu() */
574 		while (newcpu && --ts->ts_slptime)
575 			newcpu = decay_cpu(loadfac, newcpu);
576 		td->td_estcpu = newcpu;
577 	}
578 }
579 
580 /*
581  * Compute the priority of a process when running in user mode.
582  * Arrange to reschedule if the resulting priority is better
583  * than that of the current process.
584  */
585 static void
586 resetpriority(struct thread *td)
587 {
588 	register unsigned int newpriority;
589 
590 	if (td->td_pri_class == PRI_TIMESHARE) {
591 		newpriority = PUSER + td->td_estcpu / INVERSE_ESTCPU_WEIGHT +
592 		    NICE_WEIGHT * (td->td_proc->p_nice - PRIO_MIN);
593 		newpriority = min(max(newpriority, PRI_MIN_TIMESHARE),
594 		    PRI_MAX_TIMESHARE);
595 		sched_user_prio(td, newpriority);
596 	}
597 }
598 
599 /*
600  * Update the thread's priority when the associated process's user
601  * priority changes.
602  */
603 static void
604 resetpriority_thread(struct thread *td)
605 {
606 
607 	/* Only change threads with a time sharing user priority. */
608 	if (td->td_priority < PRI_MIN_TIMESHARE ||
609 	    td->td_priority > PRI_MAX_TIMESHARE)
610 		return;
611 
612 	/* XXX the whole needresched thing is broken, but not silly. */
613 	maybe_resched(td);
614 
615 	sched_prio(td, td->td_user_pri);
616 }
617 
618 /* ARGSUSED */
619 static void
620 sched_setup(void *dummy)
621 {
622 	setup_runqs();
623 
624 	if (sched_quantum == 0)
625 		sched_quantum = SCHED_QUANTUM;
626 	hogticks = 2 * sched_quantum;
627 
628 	/* Account for thread0. */
629 	sched_load_add();
630 }
631 
632 /* External interfaces start here */
633 
634 /*
635  * Very early in the boot some setup of scheduler-specific
636  * parts of proc0 and of some scheduler resources needs to be done.
637  * Called from:
638  *  proc0_init()
639  */
640 void
641 schedinit(void)
642 {
643 	/*
644 	 * Set up the scheduler specific parts of proc0.
645 	 */
646 	proc0.p_sched = NULL; /* XXX */
647 	thread0.td_sched = &td_sched0;
648 	thread0.td_lock = &sched_lock;
649 	mtx_init(&sched_lock, "sched lock", NULL, MTX_SPIN | MTX_RECURSE);
650 }
651 
652 int
653 sched_runnable(void)
654 {
655 #ifdef SMP
656 	return runq_check(&runq) + runq_check(&runq_pcpu[PCPU_GET(cpuid)]);
657 #else
658 	return runq_check(&runq);
659 #endif
660 }
661 
662 int
663 sched_rr_interval(void)
664 {
665 	if (sched_quantum == 0)
666 		sched_quantum = SCHED_QUANTUM;
667 	return (sched_quantum);
668 }
669 
670 /*
671  * We adjust the priority of the current process.  The priority of
672  * a process gets worse as it accumulates CPU time.  The cpu usage
673  * estimator (td_estcpu) is increased here.  resetpriority() will
674  * compute a different priority each time td_estcpu increases by
675  * INVERSE_ESTCPU_WEIGHT
676  * (until MAXPRI is reached).  The cpu usage estimator ramps up
677  * quite quickly when the process is running (linearly), and decays
678  * away exponentially, at a rate which is proportionally slower when
679  * the system is busy.  The basic principle is that the system will
680  * 90% forget that the process used a lot of CPU time in 5 * loadav
681  * seconds.  This causes the system to favor processes which haven't
682  * run much recently, and to round-robin among other processes.
683  */
684 void
685 sched_clock(struct thread *td)
686 {
687 	struct td_sched *ts;
688 
689 	THREAD_LOCK_ASSERT(td, MA_OWNED);
690 	ts = td->td_sched;
691 
692 	ts->ts_cpticks++;
693 	td->td_estcpu = ESTCPULIM(td->td_estcpu + 1);
694 	if ((td->td_estcpu % INVERSE_ESTCPU_WEIGHT) == 0) {
695 		resetpriority(td);
696 		resetpriority_thread(td);
697 	}
698 
699 	/*
700 	 * Force a context switch if the current thread has used up a full
701 	 * quantum (default quantum is 100ms).
702 	 */
703 	if (!TD_IS_IDLETHREAD(td) &&
704 	    ticks - PCPU_GET(switchticks) >= sched_quantum)
705 		td->td_flags |= TDF_NEEDRESCHED;
706 }
707 
708 /*
709  * Charge child's scheduling CPU usage to parent.
710  */
711 void
712 sched_exit(struct proc *p, struct thread *td)
713 {
714 
715 	KTR_STATE1(KTR_SCHED, "thread", sched_tdname(td), "proc exit",
716 	    "prio:td", td->td_priority);
717 
718 	PROC_LOCK_ASSERT(p, MA_OWNED);
719 	sched_exit_thread(FIRST_THREAD_IN_PROC(p), td);
720 }
721 
722 void
723 sched_exit_thread(struct thread *td, struct thread *child)
724 {
725 
726 	KTR_STATE1(KTR_SCHED, "thread", sched_tdname(child), "exit",
727 	    "prio:td", child->td_priority);
728 	thread_lock(td);
729 	td->td_estcpu = ESTCPULIM(td->td_estcpu + child->td_estcpu);
730 	thread_unlock(td);
731 	mtx_lock_spin(&sched_lock);
732 	if ((child->td_proc->p_flag & P_NOLOAD) == 0)
733 		sched_load_rem();
734 	mtx_unlock_spin(&sched_lock);
735 }
736 
737 void
738 sched_fork(struct thread *td, struct thread *childtd)
739 {
740 	sched_fork_thread(td, childtd);
741 }
742 
743 void
744 sched_fork_thread(struct thread *td, struct thread *childtd)
745 {
746 	struct td_sched *ts;
747 
748 	childtd->td_estcpu = td->td_estcpu;
749 	childtd->td_lock = &sched_lock;
750 	childtd->td_cpuset = cpuset_ref(td->td_cpuset);
751 	ts = childtd->td_sched;
752 	bzero(ts, sizeof(*ts));
753 	ts->ts_flags |= (td->td_sched->ts_flags & TSF_AFFINITY);
754 }
755 
756 void
757 sched_nice(struct proc *p, int nice)
758 {
759 	struct thread *td;
760 
761 	PROC_LOCK_ASSERT(p, MA_OWNED);
762 	p->p_nice = nice;
763 	FOREACH_THREAD_IN_PROC(p, td) {
764 		thread_lock(td);
765 		resetpriority(td);
766 		resetpriority_thread(td);
767 		thread_unlock(td);
768 	}
769 }
770 
771 void
772 sched_class(struct thread *td, int class)
773 {
774 	THREAD_LOCK_ASSERT(td, MA_OWNED);
775 	td->td_pri_class = class;
776 }
777 
778 /*
779  * Adjust the priority of a thread.
780  */
781 static void
782 sched_priority(struct thread *td, u_char prio)
783 {
784 
785 
786 	KTR_POINT3(KTR_SCHED, "thread", sched_tdname(td), "priority change",
787 	    "prio:%d", td->td_priority, "new prio:%d", prio, KTR_ATTR_LINKED,
788 	    sched_tdname(curthread));
789 	if (td != curthread && prio > td->td_priority) {
790 		KTR_POINT3(KTR_SCHED, "thread", sched_tdname(curthread),
791 		    "lend prio", "prio:%d", td->td_priority, "new prio:%d",
792 		    prio, KTR_ATTR_LINKED, sched_tdname(td));
793 	}
794 	THREAD_LOCK_ASSERT(td, MA_OWNED);
795 	if (td->td_priority == prio)
796 		return;
797 	td->td_priority = prio;
798 	if (TD_ON_RUNQ(td) && td->td_rqindex != (prio / RQ_PPQ)) {
799 		sched_rem(td);
800 		sched_add(td, SRQ_BORING);
801 	}
802 }
803 
804 /*
805  * Update a thread's priority when it is lent another thread's
806  * priority.
807  */
808 void
809 sched_lend_prio(struct thread *td, u_char prio)
810 {
811 
812 	td->td_flags |= TDF_BORROWING;
813 	sched_priority(td, prio);
814 }
815 
816 /*
817  * Restore a thread's priority when priority propagation is
818  * over.  The prio argument is the minimum priority the thread
819  * needs to have to satisfy other possible priority lending
820  * requests.  If the thread's regulary priority is less
821  * important than prio the thread will keep a priority boost
822  * of prio.
823  */
824 void
825 sched_unlend_prio(struct thread *td, u_char prio)
826 {
827 	u_char base_pri;
828 
829 	if (td->td_base_pri >= PRI_MIN_TIMESHARE &&
830 	    td->td_base_pri <= PRI_MAX_TIMESHARE)
831 		base_pri = td->td_user_pri;
832 	else
833 		base_pri = td->td_base_pri;
834 	if (prio >= base_pri) {
835 		td->td_flags &= ~TDF_BORROWING;
836 		sched_prio(td, base_pri);
837 	} else
838 		sched_lend_prio(td, prio);
839 }
840 
841 void
842 sched_prio(struct thread *td, u_char prio)
843 {
844 	u_char oldprio;
845 
846 	/* First, update the base priority. */
847 	td->td_base_pri = prio;
848 
849 	/*
850 	 * If the thread is borrowing another thread's priority, don't ever
851 	 * lower the priority.
852 	 */
853 	if (td->td_flags & TDF_BORROWING && td->td_priority < prio)
854 		return;
855 
856 	/* Change the real priority. */
857 	oldprio = td->td_priority;
858 	sched_priority(td, prio);
859 
860 	/*
861 	 * If the thread is on a turnstile, then let the turnstile update
862 	 * its state.
863 	 */
864 	if (TD_ON_LOCK(td) && oldprio != prio)
865 		turnstile_adjust(td, oldprio);
866 }
867 
868 void
869 sched_user_prio(struct thread *td, u_char prio)
870 {
871 	u_char oldprio;
872 
873 	THREAD_LOCK_ASSERT(td, MA_OWNED);
874 	td->td_base_user_pri = prio;
875 	if (td->td_flags & TDF_UBORROWING && td->td_user_pri <= prio)
876 		return;
877 	oldprio = td->td_user_pri;
878 	td->td_user_pri = prio;
879 }
880 
881 void
882 sched_lend_user_prio(struct thread *td, u_char prio)
883 {
884 	u_char oldprio;
885 
886 	THREAD_LOCK_ASSERT(td, MA_OWNED);
887 	td->td_flags |= TDF_UBORROWING;
888 	oldprio = td->td_user_pri;
889 	td->td_user_pri = prio;
890 }
891 
892 void
893 sched_unlend_user_prio(struct thread *td, u_char prio)
894 {
895 	u_char base_pri;
896 
897 	THREAD_LOCK_ASSERT(td, MA_OWNED);
898 	base_pri = td->td_base_user_pri;
899 	if (prio >= base_pri) {
900 		td->td_flags &= ~TDF_UBORROWING;
901 		sched_user_prio(td, base_pri);
902 	} else {
903 		sched_lend_user_prio(td, prio);
904 	}
905 }
906 
907 void
908 sched_sleep(struct thread *td, int pri)
909 {
910 
911 	THREAD_LOCK_ASSERT(td, MA_OWNED);
912 	td->td_slptick = ticks;
913 	td->td_sched->ts_slptime = 0;
914 	if (pri)
915 		sched_prio(td, pri);
916 	if (TD_IS_SUSPENDED(td) || pri <= PSOCK)
917 		td->td_flags |= TDF_CANSWAP;
918 }
919 
920 void
921 sched_switch(struct thread *td, struct thread *newtd, int flags)
922 {
923 	struct td_sched *ts;
924 	struct proc *p;
925 
926 	ts = td->td_sched;
927 	p = td->td_proc;
928 
929 	THREAD_LOCK_ASSERT(td, MA_OWNED);
930 
931 	/*
932 	 * Switch to the sched lock to fix things up and pick
933 	 * a new thread.
934 	 */
935 	if (td->td_lock != &sched_lock) {
936 		mtx_lock_spin(&sched_lock);
937 		thread_unlock(td);
938 	}
939 
940 	if ((p->p_flag & P_NOLOAD) == 0)
941 		sched_load_rem();
942 
943 	if (newtd)
944 		newtd->td_flags |= (td->td_flags & TDF_NEEDRESCHED);
945 
946 	td->td_lastcpu = td->td_oncpu;
947 	td->td_flags &= ~TDF_NEEDRESCHED;
948 	td->td_owepreempt = 0;
949 	td->td_oncpu = NOCPU;
950 
951 	/*
952 	 * At the last moment, if this thread is still marked RUNNING,
953 	 * then put it back on the run queue as it has not been suspended
954 	 * or stopped or any thing else similar.  We never put the idle
955 	 * threads on the run queue, however.
956 	 */
957 	if (td->td_flags & TDF_IDLETD) {
958 		TD_SET_CAN_RUN(td);
959 #ifdef SMP
960 		idle_cpus_mask &= ~PCPU_GET(cpumask);
961 #endif
962 	} else {
963 		if (TD_IS_RUNNING(td)) {
964 			/* Put us back on the run queue. */
965 			sched_add(td, (flags & SW_PREEMPT) ?
966 			    SRQ_OURSELF|SRQ_YIELDING|SRQ_PREEMPTED :
967 			    SRQ_OURSELF|SRQ_YIELDING);
968 		}
969 	}
970 	if (newtd) {
971 		/*
972 		 * The thread we are about to run needs to be counted
973 		 * as if it had been added to the run queue and selected.
974 		 * It came from:
975 		 * * A preemption
976 		 * * An upcall
977 		 * * A followon
978 		 */
979 		KASSERT((newtd->td_inhibitors == 0),
980 			("trying to run inhibited thread"));
981 		newtd->td_flags |= TDF_DIDRUN;
982         	TD_SET_RUNNING(newtd);
983 		if ((newtd->td_proc->p_flag & P_NOLOAD) == 0)
984 			sched_load_add();
985 	} else {
986 		newtd = choosethread();
987 	}
988 	MPASS(newtd->td_lock == &sched_lock);
989 
990 	if (td != newtd) {
991 #ifdef	HWPMC_HOOKS
992 		if (PMC_PROC_IS_USING_PMCS(td->td_proc))
993 			PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_OUT);
994 #endif
995                 /* I feel sleepy */
996 		lock_profile_release_lock(&sched_lock.lock_object);
997 #ifdef KDTRACE_HOOKS
998 		/*
999 		 * If DTrace has set the active vtime enum to anything
1000 		 * other than INACTIVE (0), then it should have set the
1001 		 * function to call.
1002 		 */
1003 		if (dtrace_vtime_active)
1004 			(*dtrace_vtime_switch_func)(newtd);
1005 #endif
1006 
1007 		cpu_switch(td, newtd, td->td_lock);
1008 		lock_profile_obtain_lock_success(&sched_lock.lock_object,
1009 		    0, 0, __FILE__, __LINE__);
1010 		/*
1011 		 * Where am I?  What year is it?
1012 		 * We are in the same thread that went to sleep above,
1013 		 * but any amount of time may have passed. All our context
1014 		 * will still be available as will local variables.
1015 		 * PCPU values however may have changed as we may have
1016 		 * changed CPU so don't trust cached values of them.
1017 		 * New threads will go to fork_exit() instead of here
1018 		 * so if you change things here you may need to change
1019 		 * things there too.
1020 		 *
1021 		 * If the thread above was exiting it will never wake
1022 		 * up again here, so either it has saved everything it
1023 		 * needed to, or the thread_wait() or wait() will
1024 		 * need to reap it.
1025 		 */
1026 #ifdef	HWPMC_HOOKS
1027 		if (PMC_PROC_IS_USING_PMCS(td->td_proc))
1028 			PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_IN);
1029 #endif
1030 	}
1031 
1032 #ifdef SMP
1033 	if (td->td_flags & TDF_IDLETD)
1034 		idle_cpus_mask |= PCPU_GET(cpumask);
1035 #endif
1036 	sched_lock.mtx_lock = (uintptr_t)td;
1037 	td->td_oncpu = PCPU_GET(cpuid);
1038 	MPASS(td->td_lock == &sched_lock);
1039 }
1040 
1041 void
1042 sched_wakeup(struct thread *td)
1043 {
1044 	struct td_sched *ts;
1045 
1046 	THREAD_LOCK_ASSERT(td, MA_OWNED);
1047 	ts = td->td_sched;
1048 	td->td_flags &= ~TDF_CANSWAP;
1049 	if (ts->ts_slptime > 1) {
1050 		updatepri(td);
1051 		resetpriority(td);
1052 	}
1053 	td->td_slptick = ticks;
1054 	ts->ts_slptime = 0;
1055 	sched_add(td, SRQ_BORING);
1056 }
1057 
1058 #ifdef SMP
1059 static int
1060 forward_wakeup(int cpunum)
1061 {
1062 	struct pcpu *pc;
1063 	cpumask_t dontuse, id, map, map2, map3, me;
1064 
1065 	mtx_assert(&sched_lock, MA_OWNED);
1066 
1067 	CTR0(KTR_RUNQ, "forward_wakeup()");
1068 
1069 	if ((!forward_wakeup_enabled) ||
1070 	     (forward_wakeup_use_mask == 0 && forward_wakeup_use_loop == 0))
1071 		return (0);
1072 	if (!smp_started || cold || panicstr)
1073 		return (0);
1074 
1075 	forward_wakeups_requested++;
1076 
1077 	/*
1078 	 * Check the idle mask we received against what we calculated
1079 	 * before in the old version.
1080 	 */
1081 	me = PCPU_GET(cpumask);
1082 
1083 	/* Don't bother if we should be doing it ourself. */
1084 	if ((me & idle_cpus_mask) && (cpunum == NOCPU || me == (1 << cpunum)))
1085 		return (0);
1086 
1087 	dontuse = me | stopped_cpus | hlt_cpus_mask;
1088 	map3 = 0;
1089 	if (forward_wakeup_use_loop) {
1090 		SLIST_FOREACH(pc, &cpuhead, pc_allcpu) {
1091 			id = pc->pc_cpumask;
1092 			if ((id & dontuse) == 0 &&
1093 			    pc->pc_curthread == pc->pc_idlethread) {
1094 				map3 |= id;
1095 			}
1096 		}
1097 	}
1098 
1099 	if (forward_wakeup_use_mask) {
1100 		map = 0;
1101 		map = idle_cpus_mask & ~dontuse;
1102 
1103 		/* If they are both on, compare and use loop if different. */
1104 		if (forward_wakeup_use_loop) {
1105 			if (map != map3) {
1106 				printf("map (%02X) != map3 (%02X)\n", map,
1107 				    map3);
1108 				map = map3;
1109 			}
1110 		}
1111 	} else {
1112 		map = map3;
1113 	}
1114 
1115 	/* If we only allow a specific CPU, then mask off all the others. */
1116 	if (cpunum != NOCPU) {
1117 		KASSERT((cpunum <= mp_maxcpus),("forward_wakeup: bad cpunum."));
1118 		map &= (1 << cpunum);
1119 	} else {
1120 		/* Try choose an idle die. */
1121 		if (forward_wakeup_use_htt) {
1122 			map2 =  (map & (map >> 1)) & 0x5555;
1123 			if (map2) {
1124 				map = map2;
1125 			}
1126 		}
1127 
1128 		/* Set only one bit. */
1129 		if (forward_wakeup_use_single) {
1130 			map = map & ((~map) + 1);
1131 		}
1132 	}
1133 	if (map) {
1134 		forward_wakeups_delivered++;
1135 		ipi_selected(map, IPI_AST);
1136 		return (1);
1137 	}
1138 	if (cpunum == NOCPU)
1139 		printf("forward_wakeup: Idle processor not found\n");
1140 	return (0);
1141 }
1142 
1143 static void
1144 kick_other_cpu(int pri, int cpuid)
1145 {
1146 	struct pcpu *pcpu;
1147 	int cpri;
1148 
1149 	pcpu = pcpu_find(cpuid);
1150 	if (idle_cpus_mask & pcpu->pc_cpumask) {
1151 		forward_wakeups_delivered++;
1152 		ipi_selected(pcpu->pc_cpumask, IPI_AST);
1153 		return;
1154 	}
1155 
1156 	cpri = pcpu->pc_curthread->td_priority;
1157 	if (pri >= cpri)
1158 		return;
1159 
1160 #if defined(IPI_PREEMPTION) && defined(PREEMPTION)
1161 #if !defined(FULL_PREEMPTION)
1162 	if (pri <= PRI_MAX_ITHD)
1163 #endif /* ! FULL_PREEMPTION */
1164 	{
1165 		ipi_selected(pcpu->pc_cpumask, IPI_PREEMPT);
1166 		return;
1167 	}
1168 #endif /* defined(IPI_PREEMPTION) && defined(PREEMPTION) */
1169 
1170 	pcpu->pc_curthread->td_flags |= TDF_NEEDRESCHED;
1171 	ipi_selected(pcpu->pc_cpumask, IPI_AST);
1172 	return;
1173 }
1174 #endif /* SMP */
1175 
1176 #ifdef SMP
1177 static int
1178 sched_pickcpu(struct thread *td)
1179 {
1180 	int best, cpu;
1181 
1182 	mtx_assert(&sched_lock, MA_OWNED);
1183 
1184 	if (THREAD_CAN_SCHED(td, td->td_lastcpu))
1185 		best = td->td_lastcpu;
1186 	else
1187 		best = NOCPU;
1188 	for (cpu = 0; cpu <= mp_maxid; cpu++) {
1189 		if (CPU_ABSENT(cpu))
1190 			continue;
1191 		if (!THREAD_CAN_SCHED(td, cpu))
1192 			continue;
1193 
1194 		if (best == NOCPU)
1195 			best = cpu;
1196 		else if (runq_length[cpu] < runq_length[best])
1197 			best = cpu;
1198 	}
1199 	KASSERT(best != NOCPU, ("no valid CPUs"));
1200 
1201 	return (best);
1202 }
1203 #endif
1204 
1205 void
1206 sched_add(struct thread *td, int flags)
1207 #ifdef SMP
1208 {
1209 	struct td_sched *ts;
1210 	int forwarded = 0;
1211 	int cpu;
1212 	int single_cpu = 0;
1213 
1214 	ts = td->td_sched;
1215 	THREAD_LOCK_ASSERT(td, MA_OWNED);
1216 	KASSERT((td->td_inhibitors == 0),
1217 	    ("sched_add: trying to run inhibited thread"));
1218 	KASSERT((TD_CAN_RUN(td) || TD_IS_RUNNING(td)),
1219 	    ("sched_add: bad thread state"));
1220 	KASSERT(td->td_flags & TDF_INMEM,
1221 	    ("sched_add: thread swapped out"));
1222 
1223 	KTR_STATE2(KTR_SCHED, "thread", sched_tdname(td), "runq add",
1224 	    "prio:%d", td->td_priority, KTR_ATTR_LINKED,
1225 	    sched_tdname(curthread));
1226 	KTR_POINT1(KTR_SCHED, "thread", sched_tdname(curthread), "wokeup",
1227 	    KTR_ATTR_LINKED, sched_tdname(td));
1228 
1229 
1230 	/*
1231 	 * Now that the thread is moving to the run-queue, set the lock
1232 	 * to the scheduler's lock.
1233 	 */
1234 	if (td->td_lock != &sched_lock) {
1235 		mtx_lock_spin(&sched_lock);
1236 		thread_lock_set(td, &sched_lock);
1237 	}
1238 	TD_SET_RUNQ(td);
1239 
1240 	if (td->td_pinned != 0) {
1241 		cpu = td->td_lastcpu;
1242 		ts->ts_runq = &runq_pcpu[cpu];
1243 		single_cpu = 1;
1244 		CTR3(KTR_RUNQ,
1245 		    "sched_add: Put td_sched:%p(td:%p) on cpu%d runq", ts, td,
1246 		    cpu);
1247 	} else if (td->td_flags & TDF_BOUND) {
1248 		/* Find CPU from bound runq. */
1249 		KASSERT(SKE_RUNQ_PCPU(ts),
1250 		    ("sched_add: bound td_sched not on cpu runq"));
1251 		cpu = ts->ts_runq - &runq_pcpu[0];
1252 		single_cpu = 1;
1253 		CTR3(KTR_RUNQ,
1254 		    "sched_add: Put td_sched:%p(td:%p) on cpu%d runq", ts, td,
1255 		    cpu);
1256 	} else if (ts->ts_flags & TSF_AFFINITY) {
1257 		/* Find a valid CPU for our cpuset */
1258 		cpu = sched_pickcpu(td);
1259 		ts->ts_runq = &runq_pcpu[cpu];
1260 		single_cpu = 1;
1261 		CTR3(KTR_RUNQ,
1262 		    "sched_add: Put td_sched:%p(td:%p) on cpu%d runq", ts, td,
1263 		    cpu);
1264 	} else {
1265 		CTR2(KTR_RUNQ,
1266 		    "sched_add: adding td_sched:%p (td:%p) to gbl runq", ts,
1267 		    td);
1268 		cpu = NOCPU;
1269 		ts->ts_runq = &runq;
1270 	}
1271 
1272 	if (single_cpu && (cpu != PCPU_GET(cpuid))) {
1273 	        kick_other_cpu(td->td_priority, cpu);
1274 	} else {
1275 		if (!single_cpu) {
1276 			cpumask_t me = PCPU_GET(cpumask);
1277 			cpumask_t idle = idle_cpus_mask & me;
1278 
1279 			if (!idle && ((flags & SRQ_INTR) == 0) &&
1280 			    (idle_cpus_mask & ~(hlt_cpus_mask | me)))
1281 				forwarded = forward_wakeup(cpu);
1282 		}
1283 
1284 		if (!forwarded) {
1285 			if ((flags & SRQ_YIELDING) == 0 && maybe_preempt(td))
1286 				return;
1287 			else
1288 				maybe_resched(td);
1289 		}
1290 	}
1291 
1292 	if ((td->td_proc->p_flag & P_NOLOAD) == 0)
1293 		sched_load_add();
1294 	runq_add(ts->ts_runq, td, flags);
1295 	if (cpu != NOCPU)
1296 		runq_length[cpu]++;
1297 }
1298 #else /* SMP */
1299 {
1300 	struct td_sched *ts;
1301 
1302 	ts = td->td_sched;
1303 	THREAD_LOCK_ASSERT(td, MA_OWNED);
1304 	KASSERT((td->td_inhibitors == 0),
1305 	    ("sched_add: trying to run inhibited thread"));
1306 	KASSERT((TD_CAN_RUN(td) || TD_IS_RUNNING(td)),
1307 	    ("sched_add: bad thread state"));
1308 	KASSERT(td->td_flags & TDF_INMEM,
1309 	    ("sched_add: thread swapped out"));
1310 	KTR_STATE2(KTR_SCHED, "thread", sched_tdname(td), "runq add",
1311 	    "prio:%d", td->td_priority, KTR_ATTR_LINKED,
1312 	    sched_tdname(curthread));
1313 	KTR_POINT1(KTR_SCHED, "thread", sched_tdname(curthread), "wokeup",
1314 	    KTR_ATTR_LINKED, sched_tdname(td));
1315 
1316 	/*
1317 	 * Now that the thread is moving to the run-queue, set the lock
1318 	 * to the scheduler's lock.
1319 	 */
1320 	if (td->td_lock != &sched_lock) {
1321 		mtx_lock_spin(&sched_lock);
1322 		thread_lock_set(td, &sched_lock);
1323 	}
1324 	TD_SET_RUNQ(td);
1325 	CTR2(KTR_RUNQ, "sched_add: adding td_sched:%p (td:%p) to runq", ts, td);
1326 	ts->ts_runq = &runq;
1327 
1328 	/*
1329 	 * If we are yielding (on the way out anyhow) or the thread
1330 	 * being saved is US, then don't try be smart about preemption
1331 	 * or kicking off another CPU as it won't help and may hinder.
1332 	 * In the YIEDLING case, we are about to run whoever is being
1333 	 * put in the queue anyhow, and in the OURSELF case, we are
1334 	 * puting ourself on the run queue which also only happens
1335 	 * when we are about to yield.
1336 	 */
1337 	if ((flags & SRQ_YIELDING) == 0) {
1338 		if (maybe_preempt(td))
1339 			return;
1340 	}
1341 	if ((td->td_proc->p_flag & P_NOLOAD) == 0)
1342 		sched_load_add();
1343 	runq_add(ts->ts_runq, td, flags);
1344 	maybe_resched(td);
1345 }
1346 #endif /* SMP */
1347 
1348 void
1349 sched_rem(struct thread *td)
1350 {
1351 	struct td_sched *ts;
1352 
1353 	ts = td->td_sched;
1354 	KASSERT(td->td_flags & TDF_INMEM,
1355 	    ("sched_rem: thread swapped out"));
1356 	KASSERT(TD_ON_RUNQ(td),
1357 	    ("sched_rem: thread not on run queue"));
1358 	mtx_assert(&sched_lock, MA_OWNED);
1359 	KTR_STATE2(KTR_SCHED, "thread", sched_tdname(td), "runq rem",
1360 	    "prio:%d", td->td_priority, KTR_ATTR_LINKED,
1361 	    sched_tdname(curthread));
1362 
1363 	if ((td->td_proc->p_flag & P_NOLOAD) == 0)
1364 		sched_load_rem();
1365 #ifdef SMP
1366 	if (ts->ts_runq != &runq)
1367 		runq_length[ts->ts_runq - runq_pcpu]--;
1368 #endif
1369 	runq_remove(ts->ts_runq, td);
1370 	TD_SET_CAN_RUN(td);
1371 }
1372 
1373 /*
1374  * Select threads to run.  Note that running threads still consume a
1375  * slot.
1376  */
1377 struct thread *
1378 sched_choose(void)
1379 {
1380 	struct thread *td;
1381 	struct runq *rq;
1382 
1383 	mtx_assert(&sched_lock,  MA_OWNED);
1384 #ifdef SMP
1385 	struct thread *tdcpu;
1386 
1387 	rq = &runq;
1388 	td = runq_choose_fuzz(&runq, runq_fuzz);
1389 	tdcpu = runq_choose(&runq_pcpu[PCPU_GET(cpuid)]);
1390 
1391 	if (td == NULL ||
1392 	    (tdcpu != NULL &&
1393 	     tdcpu->td_priority < td->td_priority)) {
1394 		CTR2(KTR_RUNQ, "choosing td %p from pcpu runq %d", tdcpu,
1395 		     PCPU_GET(cpuid));
1396 		td = tdcpu;
1397 		rq = &runq_pcpu[PCPU_GET(cpuid)];
1398 	} else {
1399 		CTR1(KTR_RUNQ, "choosing td_sched %p from main runq", td);
1400 	}
1401 
1402 #else
1403 	rq = &runq;
1404 	td = runq_choose(&runq);
1405 #endif
1406 
1407 	if (td) {
1408 #ifdef SMP
1409 		if (td == tdcpu)
1410 			runq_length[PCPU_GET(cpuid)]--;
1411 #endif
1412 		runq_remove(rq, td);
1413 		td->td_flags |= TDF_DIDRUN;
1414 
1415 		KASSERT(td->td_flags & TDF_INMEM,
1416 		    ("sched_choose: thread swapped out"));
1417 		return (td);
1418 	}
1419 	return (PCPU_GET(idlethread));
1420 }
1421 
1422 void
1423 sched_preempt(struct thread *td)
1424 {
1425 	thread_lock(td);
1426 	if (td->td_critnest > 1)
1427 		td->td_owepreempt = 1;
1428 	else
1429 		mi_switch(SW_INVOL | SW_PREEMPT | SWT_PREEMPT, NULL);
1430 	thread_unlock(td);
1431 }
1432 
1433 void
1434 sched_userret(struct thread *td)
1435 {
1436 	/*
1437 	 * XXX we cheat slightly on the locking here to avoid locking in
1438 	 * the usual case.  Setting td_priority here is essentially an
1439 	 * incomplete workaround for not setting it properly elsewhere.
1440 	 * Now that some interrupt handlers are threads, not setting it
1441 	 * properly elsewhere can clobber it in the window between setting
1442 	 * it here and returning to user mode, so don't waste time setting
1443 	 * it perfectly here.
1444 	 */
1445 	KASSERT((td->td_flags & TDF_BORROWING) == 0,
1446 	    ("thread with borrowed priority returning to userland"));
1447 	if (td->td_priority != td->td_user_pri) {
1448 		thread_lock(td);
1449 		td->td_priority = td->td_user_pri;
1450 		td->td_base_pri = td->td_user_pri;
1451 		thread_unlock(td);
1452 	}
1453 }
1454 
1455 void
1456 sched_bind(struct thread *td, int cpu)
1457 {
1458 	struct td_sched *ts;
1459 
1460 	THREAD_LOCK_ASSERT(td, MA_OWNED);
1461 	KASSERT(TD_IS_RUNNING(td),
1462 	    ("sched_bind: cannot bind non-running thread"));
1463 
1464 	ts = td->td_sched;
1465 
1466 	td->td_flags |= TDF_BOUND;
1467 #ifdef SMP
1468 	ts->ts_runq = &runq_pcpu[cpu];
1469 	if (PCPU_GET(cpuid) == cpu)
1470 		return;
1471 
1472 	mi_switch(SW_VOL, NULL);
1473 #endif
1474 }
1475 
1476 void
1477 sched_unbind(struct thread* td)
1478 {
1479 	THREAD_LOCK_ASSERT(td, MA_OWNED);
1480 	td->td_flags &= ~TDF_BOUND;
1481 }
1482 
1483 int
1484 sched_is_bound(struct thread *td)
1485 {
1486 	THREAD_LOCK_ASSERT(td, MA_OWNED);
1487 	return (td->td_flags & TDF_BOUND);
1488 }
1489 
1490 void
1491 sched_relinquish(struct thread *td)
1492 {
1493 	thread_lock(td);
1494 	mi_switch(SW_VOL | SWT_RELINQUISH, NULL);
1495 	thread_unlock(td);
1496 }
1497 
1498 int
1499 sched_load(void)
1500 {
1501 	return (sched_tdcnt);
1502 }
1503 
1504 int
1505 sched_sizeof_proc(void)
1506 {
1507 	return (sizeof(struct proc));
1508 }
1509 
1510 int
1511 sched_sizeof_thread(void)
1512 {
1513 	return (sizeof(struct thread) + sizeof(struct td_sched));
1514 }
1515 
1516 fixpt_t
1517 sched_pctcpu(struct thread *td)
1518 {
1519 	struct td_sched *ts;
1520 
1521 	ts = td->td_sched;
1522 	return (ts->ts_pctcpu);
1523 }
1524 
1525 void
1526 sched_tick(void)
1527 {
1528 }
1529 
1530 /*
1531  * The actual idle process.
1532  */
1533 void
1534 sched_idletd(void *dummy)
1535 {
1536 
1537 	for (;;) {
1538 		mtx_assert(&Giant, MA_NOTOWNED);
1539 
1540 		while (sched_runnable() == 0)
1541 			cpu_idle(0);
1542 
1543 		mtx_lock_spin(&sched_lock);
1544 		mi_switch(SW_VOL | SWT_IDLE, NULL);
1545 		mtx_unlock_spin(&sched_lock);
1546 	}
1547 }
1548 
1549 /*
1550  * A CPU is entering for the first time or a thread is exiting.
1551  */
1552 void
1553 sched_throw(struct thread *td)
1554 {
1555 	/*
1556 	 * Correct spinlock nesting.  The idle thread context that we are
1557 	 * borrowing was created so that it would start out with a single
1558 	 * spin lock (sched_lock) held in fork_trampoline().  Since we've
1559 	 * explicitly acquired locks in this function, the nesting count
1560 	 * is now 2 rather than 1.  Since we are nested, calling
1561 	 * spinlock_exit() will simply adjust the counts without allowing
1562 	 * spin lock using code to interrupt us.
1563 	 */
1564 	if (td == NULL) {
1565 		mtx_lock_spin(&sched_lock);
1566 		spinlock_exit();
1567 	} else {
1568 		lock_profile_release_lock(&sched_lock.lock_object);
1569 		MPASS(td->td_lock == &sched_lock);
1570 	}
1571 	mtx_assert(&sched_lock, MA_OWNED);
1572 	KASSERT(curthread->td_md.md_spinlock_count == 1, ("invalid count"));
1573 	PCPU_SET(switchtime, cpu_ticks());
1574 	PCPU_SET(switchticks, ticks);
1575 	cpu_throw(td, choosethread());	/* doesn't return */
1576 }
1577 
1578 void
1579 sched_fork_exit(struct thread *td)
1580 {
1581 
1582 	/*
1583 	 * Finish setting up thread glue so that it begins execution in a
1584 	 * non-nested critical section with sched_lock held but not recursed.
1585 	 */
1586 	td->td_oncpu = PCPU_GET(cpuid);
1587 	sched_lock.mtx_lock = (uintptr_t)td;
1588 	lock_profile_obtain_lock_success(&sched_lock.lock_object,
1589 	    0, 0, __FILE__, __LINE__);
1590 	THREAD_LOCK_ASSERT(td, MA_OWNED | MA_NOTRECURSED);
1591 }
1592 
1593 char *
1594 sched_tdname(struct thread *td)
1595 {
1596 #ifdef KTR
1597 	struct td_sched *ts;
1598 
1599 	ts = td->td_sched;
1600 	if (ts->ts_name[0] == '\0')
1601 		snprintf(ts->ts_name, sizeof(ts->ts_name),
1602 		    "%s tid %d", td->td_name, td->td_tid);
1603 	return (ts->ts_name);
1604 #else
1605 	return (td->td_name);
1606 #endif
1607 }
1608 
1609 void
1610 sched_affinity(struct thread *td)
1611 {
1612 #ifdef SMP
1613 	struct td_sched *ts;
1614 	int cpu;
1615 
1616 	THREAD_LOCK_ASSERT(td, MA_OWNED);
1617 
1618 	/*
1619 	 * Set the TSF_AFFINITY flag if there is at least one CPU this
1620 	 * thread can't run on.
1621 	 */
1622 	ts = td->td_sched;
1623 	ts->ts_flags &= ~TSF_AFFINITY;
1624 	for (cpu = 0; cpu <= mp_maxid; cpu++) {
1625 		if (CPU_ABSENT(cpu))
1626 			continue;
1627 		if (!THREAD_CAN_SCHED(td, cpu)) {
1628 			ts->ts_flags |= TSF_AFFINITY;
1629 			break;
1630 		}
1631 	}
1632 
1633 	/*
1634 	 * If this thread can run on all CPUs, nothing else to do.
1635 	 */
1636 	if (!(ts->ts_flags & TSF_AFFINITY))
1637 		return;
1638 
1639 	/* Pinned threads and bound threads should be left alone. */
1640 	if (td->td_pinned != 0 || td->td_flags & TDF_BOUND)
1641 		return;
1642 
1643 	switch (td->td_state) {
1644 	case TDS_RUNQ:
1645 		/*
1646 		 * If we are on a per-CPU runqueue that is in the set,
1647 		 * then nothing needs to be done.
1648 		 */
1649 		if (ts->ts_runq != &runq &&
1650 		    THREAD_CAN_SCHED(td, ts->ts_runq - runq_pcpu))
1651 			return;
1652 
1653 		/* Put this thread on a valid per-CPU runqueue. */
1654 		sched_rem(td);
1655 		sched_add(td, SRQ_BORING);
1656 		break;
1657 	case TDS_RUNNING:
1658 		/*
1659 		 * See if our current CPU is in the set.  If not, force a
1660 		 * context switch.
1661 		 */
1662 		if (THREAD_CAN_SCHED(td, td->td_oncpu))
1663 			return;
1664 
1665 		td->td_flags |= TDF_NEEDRESCHED;
1666 		if (td != curthread)
1667 			ipi_selected(1 << cpu, IPI_AST);
1668 		break;
1669 	default:
1670 		break;
1671 	}
1672 #endif
1673 }
1674