xref: /freebsd/sys/kern/sched_4bsd.c (revision 682c9e0fed0115eb6f283e755901c0aac90e86e8)
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 
159 static cpuset_t idle_cpus_mask;
160 #endif
161 
162 struct pcpuidlestat {
163 	u_int idlecalls;
164 	u_int oldidlecalls;
165 };
166 static DPCPU_DEFINE(struct pcpuidlestat, idlestat);
167 
168 static void
169 setup_runqs(void)
170 {
171 #ifdef SMP
172 	int i;
173 
174 	for (i = 0; i < MAXCPU; ++i)
175 		runq_init(&runq_pcpu[i]);
176 #endif
177 
178 	runq_init(&runq);
179 }
180 
181 static int
182 sysctl_kern_quantum(SYSCTL_HANDLER_ARGS)
183 {
184 	int error, new_val;
185 
186 	new_val = sched_quantum * tick;
187 	error = sysctl_handle_int(oidp, &new_val, 0, req);
188         if (error != 0 || req->newptr == NULL)
189 		return (error);
190 	if (new_val < tick)
191 		return (EINVAL);
192 	sched_quantum = new_val / tick;
193 	hogticks = 2 * sched_quantum;
194 	return (0);
195 }
196 
197 SYSCTL_NODE(_kern, OID_AUTO, sched, CTLFLAG_RD, 0, "Scheduler");
198 
199 SYSCTL_STRING(_kern_sched, OID_AUTO, name, CTLFLAG_RD, "4BSD", 0,
200     "Scheduler name");
201 
202 SYSCTL_PROC(_kern_sched, OID_AUTO, quantum, CTLTYPE_INT | CTLFLAG_RW,
203     0, sizeof sched_quantum, sysctl_kern_quantum, "I",
204     "Roundrobin scheduling quantum in microseconds");
205 
206 #ifdef SMP
207 /* Enable forwarding of wakeups to all other cpus */
208 SYSCTL_NODE(_kern_sched, OID_AUTO, ipiwakeup, CTLFLAG_RD, NULL, "Kernel SMP");
209 
210 static int runq_fuzz = 1;
211 SYSCTL_INT(_kern_sched, OID_AUTO, runq_fuzz, CTLFLAG_RW, &runq_fuzz, 0, "");
212 
213 static int forward_wakeup_enabled = 1;
214 SYSCTL_INT(_kern_sched_ipiwakeup, OID_AUTO, enabled, CTLFLAG_RW,
215 	   &forward_wakeup_enabled, 0,
216 	   "Forwarding of wakeup to idle CPUs");
217 
218 static int forward_wakeups_requested = 0;
219 SYSCTL_INT(_kern_sched_ipiwakeup, OID_AUTO, requested, CTLFLAG_RD,
220 	   &forward_wakeups_requested, 0,
221 	   "Requests for Forwarding of wakeup to idle CPUs");
222 
223 static int forward_wakeups_delivered = 0;
224 SYSCTL_INT(_kern_sched_ipiwakeup, OID_AUTO, delivered, CTLFLAG_RD,
225 	   &forward_wakeups_delivered, 0,
226 	   "Completed Forwarding of wakeup to idle CPUs");
227 
228 static int forward_wakeup_use_mask = 1;
229 SYSCTL_INT(_kern_sched_ipiwakeup, OID_AUTO, usemask, CTLFLAG_RW,
230 	   &forward_wakeup_use_mask, 0,
231 	   "Use the mask of idle cpus");
232 
233 static int forward_wakeup_use_loop = 0;
234 SYSCTL_INT(_kern_sched_ipiwakeup, OID_AUTO, useloop, CTLFLAG_RW,
235 	   &forward_wakeup_use_loop, 0,
236 	   "Use a loop to find idle cpus");
237 
238 #endif
239 #if 0
240 static int sched_followon = 0;
241 SYSCTL_INT(_kern_sched, OID_AUTO, followon, CTLFLAG_RW,
242 	   &sched_followon, 0,
243 	   "allow threads to share a quantum");
244 #endif
245 
246 static __inline void
247 sched_load_add(void)
248 {
249 
250 	sched_tdcnt++;
251 	KTR_COUNTER0(KTR_SCHED, "load", "global load", sched_tdcnt);
252 }
253 
254 static __inline void
255 sched_load_rem(void)
256 {
257 
258 	sched_tdcnt--;
259 	KTR_COUNTER0(KTR_SCHED, "load", "global load", sched_tdcnt);
260 }
261 /*
262  * Arrange to reschedule if necessary, taking the priorities and
263  * schedulers into account.
264  */
265 static void
266 maybe_resched(struct thread *td)
267 {
268 
269 	THREAD_LOCK_ASSERT(td, MA_OWNED);
270 	if (td->td_priority < curthread->td_priority)
271 		curthread->td_flags |= TDF_NEEDRESCHED;
272 }
273 
274 /*
275  * This function is called when a thread is about to be put on run queue
276  * because it has been made runnable or its priority has been adjusted.  It
277  * determines if the new thread should be immediately preempted to.  If so,
278  * it switches to it and eventually returns true.  If not, it returns false
279  * so that the caller may place the thread on an appropriate run queue.
280  */
281 int
282 maybe_preempt(struct thread *td)
283 {
284 #ifdef PREEMPTION
285 	struct thread *ctd;
286 	int cpri, pri;
287 
288 	/*
289 	 * The new thread should not preempt the current thread if any of the
290 	 * following conditions are true:
291 	 *
292 	 *  - The kernel is in the throes of crashing (panicstr).
293 	 *  - The current thread has a higher (numerically lower) or
294 	 *    equivalent priority.  Note that this prevents curthread from
295 	 *    trying to preempt to itself.
296 	 *  - It is too early in the boot for context switches (cold is set).
297 	 *  - The current thread has an inhibitor set or is in the process of
298 	 *    exiting.  In this case, the current thread is about to switch
299 	 *    out anyways, so there's no point in preempting.  If we did,
300 	 *    the current thread would not be properly resumed as well, so
301 	 *    just avoid that whole landmine.
302 	 *  - If the new thread's priority is not a realtime priority and
303 	 *    the current thread's priority is not an idle priority and
304 	 *    FULL_PREEMPTION is disabled.
305 	 *
306 	 * If all of these conditions are false, but the current thread is in
307 	 * a nested critical section, then we have to defer the preemption
308 	 * until we exit the critical section.  Otherwise, switch immediately
309 	 * to the new thread.
310 	 */
311 	ctd = curthread;
312 	THREAD_LOCK_ASSERT(td, MA_OWNED);
313 	KASSERT((td->td_inhibitors == 0),
314 			("maybe_preempt: trying to run inhibited thread"));
315 	pri = td->td_priority;
316 	cpri = ctd->td_priority;
317 	if (panicstr != NULL || pri >= cpri || cold /* || dumping */ ||
318 	    TD_IS_INHIBITED(ctd))
319 		return (0);
320 #ifndef FULL_PREEMPTION
321 	if (pri > PRI_MAX_ITHD && cpri < PRI_MIN_IDLE)
322 		return (0);
323 #endif
324 
325 	if (ctd->td_critnest > 1) {
326 		CTR1(KTR_PROC, "maybe_preempt: in critical section %d",
327 		    ctd->td_critnest);
328 		ctd->td_owepreempt = 1;
329 		return (0);
330 	}
331 	/*
332 	 * Thread is runnable but not yet put on system run queue.
333 	 */
334 	MPASS(ctd->td_lock == td->td_lock);
335 	MPASS(TD_ON_RUNQ(td));
336 	TD_SET_RUNNING(td);
337 	CTR3(KTR_PROC, "preempting to thread %p (pid %d, %s)\n", td,
338 	    td->td_proc->p_pid, td->td_name);
339 	mi_switch(SW_INVOL | SW_PREEMPT | SWT_PREEMPT, td);
340 	/*
341 	 * td's lock pointer may have changed.  We have to return with it
342 	 * locked.
343 	 */
344 	spinlock_enter();
345 	thread_unlock(ctd);
346 	thread_lock(td);
347 	spinlock_exit();
348 	return (1);
349 #else
350 	return (0);
351 #endif
352 }
353 
354 /*
355  * Constants for digital decay and forget:
356  *	90% of (td_estcpu) usage in 5 * loadav time
357  *	95% of (ts_pctcpu) usage in 60 seconds (load insensitive)
358  *          Note that, as ps(1) mentions, this can let percentages
359  *          total over 100% (I've seen 137.9% for 3 processes).
360  *
361  * Note that schedclock() updates td_estcpu and p_cpticks asynchronously.
362  *
363  * We wish to decay away 90% of td_estcpu in (5 * loadavg) seconds.
364  * That is, the system wants to compute a value of decay such
365  * that the following for loop:
366  * 	for (i = 0; i < (5 * loadavg); i++)
367  * 		td_estcpu *= decay;
368  * will compute
369  * 	td_estcpu *= 0.1;
370  * for all values of loadavg:
371  *
372  * Mathematically this loop can be expressed by saying:
373  * 	decay ** (5 * loadavg) ~= .1
374  *
375  * The system computes decay as:
376  * 	decay = (2 * loadavg) / (2 * loadavg + 1)
377  *
378  * We wish to prove that the system's computation of decay
379  * will always fulfill the equation:
380  * 	decay ** (5 * loadavg) ~= .1
381  *
382  * If we compute b as:
383  * 	b = 2 * loadavg
384  * then
385  * 	decay = b / (b + 1)
386  *
387  * We now need to prove two things:
388  *	1) Given factor ** (5 * loadavg) ~= .1, prove factor == b/(b+1)
389  *	2) Given b/(b+1) ** power ~= .1, prove power == (5 * loadavg)
390  *
391  * Facts:
392  *         For x close to zero, exp(x) =~ 1 + x, since
393  *              exp(x) = 0! + x**1/1! + x**2/2! + ... .
394  *              therefore exp(-1/b) =~ 1 - (1/b) = (b-1)/b.
395  *         For x close to zero, ln(1+x) =~ x, since
396  *              ln(1+x) = x - x**2/2 + x**3/3 - ...     -1 < x < 1
397  *              therefore ln(b/(b+1)) = ln(1 - 1/(b+1)) =~ -1/(b+1).
398  *         ln(.1) =~ -2.30
399  *
400  * Proof of (1):
401  *    Solve (factor)**(power) =~ .1 given power (5*loadav):
402  *	solving for factor,
403  *      ln(factor) =~ (-2.30/5*loadav), or
404  *      factor =~ exp(-1/((5/2.30)*loadav)) =~ exp(-1/(2*loadav)) =
405  *          exp(-1/b) =~ (b-1)/b =~ b/(b+1).                    QED
406  *
407  * Proof of (2):
408  *    Solve (factor)**(power) =~ .1 given factor == (b/(b+1)):
409  *	solving for power,
410  *      power*ln(b/(b+1)) =~ -2.30, or
411  *      power =~ 2.3 * (b + 1) = 4.6*loadav + 2.3 =~ 5*loadav.  QED
412  *
413  * Actual power values for the implemented algorithm are as follows:
414  *      loadav: 1       2       3       4
415  *      power:  5.68    10.32   14.94   19.55
416  */
417 
418 /* calculations for digital decay to forget 90% of usage in 5*loadav sec */
419 #define	loadfactor(loadav)	(2 * (loadav))
420 #define	decay_cpu(loadfac, cpu)	(((loadfac) * (cpu)) / ((loadfac) + FSCALE))
421 
422 /* decay 95% of `ts_pctcpu' in 60 seconds; see CCPU_SHIFT before changing */
423 static fixpt_t	ccpu = 0.95122942450071400909 * FSCALE;	/* exp(-1/20) */
424 SYSCTL_UINT(_kern, OID_AUTO, ccpu, CTLFLAG_RD, &ccpu, 0, "");
425 
426 /*
427  * If `ccpu' is not equal to `exp(-1/20)' and you still want to use the
428  * faster/more-accurate formula, you'll have to estimate CCPU_SHIFT below
429  * and possibly adjust FSHIFT in "param.h" so that (FSHIFT >= CCPU_SHIFT).
430  *
431  * To estimate CCPU_SHIFT for exp(-1/20), the following formula was used:
432  *	1 - exp(-1/20) ~= 0.0487 ~= 0.0488 == 1 (fixed pt, *11* bits).
433  *
434  * If you don't want to bother with the faster/more-accurate formula, you
435  * can set CCPU_SHIFT to (FSHIFT + 1) which will use a slower/less-accurate
436  * (more general) method of calculating the %age of CPU used by a process.
437  */
438 #define	CCPU_SHIFT	11
439 
440 /*
441  * Recompute process priorities, every hz ticks.
442  * MP-safe, called without the Giant mutex.
443  */
444 /* ARGSUSED */
445 static void
446 schedcpu(void)
447 {
448 	register fixpt_t loadfac = loadfactor(averunnable.ldavg[0]);
449 	struct thread *td;
450 	struct proc *p;
451 	struct td_sched *ts;
452 	int awake, realstathz;
453 
454 	realstathz = stathz ? stathz : hz;
455 	sx_slock(&allproc_lock);
456 	FOREACH_PROC_IN_SYSTEM(p) {
457 		PROC_LOCK(p);
458 		if (p->p_state == PRS_NEW) {
459 			PROC_UNLOCK(p);
460 			continue;
461 		}
462 		FOREACH_THREAD_IN_PROC(p, td) {
463 			awake = 0;
464 			thread_lock(td);
465 			ts = td->td_sched;
466 			/*
467 			 * Increment sleep time (if sleeping).  We
468 			 * ignore overflow, as above.
469 			 */
470 			/*
471 			 * The td_sched slptimes are not touched in wakeup
472 			 * because the thread may not HAVE everything in
473 			 * memory? XXX I think this is out of date.
474 			 */
475 			if (TD_ON_RUNQ(td)) {
476 				awake = 1;
477 				td->td_flags &= ~TDF_DIDRUN;
478 			} else if (TD_IS_RUNNING(td)) {
479 				awake = 1;
480 				/* Do not clear TDF_DIDRUN */
481 			} else if (td->td_flags & TDF_DIDRUN) {
482 				awake = 1;
483 				td->td_flags &= ~TDF_DIDRUN;
484 			}
485 
486 			/*
487 			 * ts_pctcpu is only for ps and ttyinfo().
488 			 */
489 			ts->ts_pctcpu = (ts->ts_pctcpu * ccpu) >> FSHIFT;
490 			/*
491 			 * If the td_sched has been idle the entire second,
492 			 * stop recalculating its priority until
493 			 * it wakes up.
494 			 */
495 			if (ts->ts_cpticks != 0) {
496 #if	(FSHIFT >= CCPU_SHIFT)
497 				ts->ts_pctcpu += (realstathz == 100)
498 				    ? ((fixpt_t) ts->ts_cpticks) <<
499 				    (FSHIFT - CCPU_SHIFT) :
500 				    100 * (((fixpt_t) ts->ts_cpticks)
501 				    << (FSHIFT - CCPU_SHIFT)) / realstathz;
502 #else
503 				ts->ts_pctcpu += ((FSCALE - ccpu) *
504 				    (ts->ts_cpticks *
505 				    FSCALE / realstathz)) >> FSHIFT;
506 #endif
507 				ts->ts_cpticks = 0;
508 			}
509 			/*
510 			 * If there are ANY running threads in this process,
511 			 * then don't count it as sleeping.
512 			 * XXX: this is broken.
513 			 */
514 			if (awake) {
515 				if (ts->ts_slptime > 1) {
516 					/*
517 					 * In an ideal world, this should not
518 					 * happen, because whoever woke us
519 					 * up from the long sleep should have
520 					 * unwound the slptime and reset our
521 					 * priority before we run at the stale
522 					 * priority.  Should KASSERT at some
523 					 * point when all the cases are fixed.
524 					 */
525 					updatepri(td);
526 				}
527 				ts->ts_slptime = 0;
528 			} else
529 				ts->ts_slptime++;
530 			if (ts->ts_slptime > 1) {
531 				thread_unlock(td);
532 				continue;
533 			}
534 			td->td_estcpu = decay_cpu(loadfac, td->td_estcpu);
535 		      	resetpriority(td);
536 			resetpriority_thread(td);
537 			thread_unlock(td);
538 		}
539 		PROC_UNLOCK(p);
540 	}
541 	sx_sunlock(&allproc_lock);
542 }
543 
544 /*
545  * Main loop for a kthread that executes schedcpu once a second.
546  */
547 static void
548 schedcpu_thread(void)
549 {
550 
551 	for (;;) {
552 		schedcpu();
553 		pause("-", hz);
554 	}
555 }
556 
557 /*
558  * Recalculate the priority of a process after it has slept for a while.
559  * For all load averages >= 1 and max td_estcpu of 255, sleeping for at
560  * least six times the loadfactor will decay td_estcpu to zero.
561  */
562 static void
563 updatepri(struct thread *td)
564 {
565 	struct td_sched *ts;
566 	fixpt_t loadfac;
567 	unsigned int newcpu;
568 
569 	ts = td->td_sched;
570 	loadfac = loadfactor(averunnable.ldavg[0]);
571 	if (ts->ts_slptime > 5 * loadfac)
572 		td->td_estcpu = 0;
573 	else {
574 		newcpu = td->td_estcpu;
575 		ts->ts_slptime--;	/* was incremented in schedcpu() */
576 		while (newcpu && --ts->ts_slptime)
577 			newcpu = decay_cpu(loadfac, newcpu);
578 		td->td_estcpu = newcpu;
579 	}
580 }
581 
582 /*
583  * Compute the priority of a process when running in user mode.
584  * Arrange to reschedule if the resulting priority is better
585  * than that of the current process.
586  */
587 static void
588 resetpriority(struct thread *td)
589 {
590 	register unsigned int newpriority;
591 
592 	if (td->td_pri_class == PRI_TIMESHARE) {
593 		newpriority = PUSER + td->td_estcpu / INVERSE_ESTCPU_WEIGHT +
594 		    NICE_WEIGHT * (td->td_proc->p_nice - PRIO_MIN);
595 		newpriority = min(max(newpriority, PRI_MIN_TIMESHARE),
596 		    PRI_MAX_TIMESHARE);
597 		sched_user_prio(td, newpriority);
598 	}
599 }
600 
601 /*
602  * Update the thread's priority when the associated process's user
603  * priority changes.
604  */
605 static void
606 resetpriority_thread(struct thread *td)
607 {
608 
609 	/* Only change threads with a time sharing user priority. */
610 	if (td->td_priority < PRI_MIN_TIMESHARE ||
611 	    td->td_priority > PRI_MAX_TIMESHARE)
612 		return;
613 
614 	/* XXX the whole needresched thing is broken, but not silly. */
615 	maybe_resched(td);
616 
617 	sched_prio(td, td->td_user_pri);
618 }
619 
620 /* ARGSUSED */
621 static void
622 sched_setup(void *dummy)
623 {
624 	setup_runqs();
625 
626 	if (sched_quantum == 0)
627 		sched_quantum = SCHED_QUANTUM;
628 	hogticks = 2 * sched_quantum;
629 
630 	/* Account for thread0. */
631 	sched_load_add();
632 }
633 
634 /* External interfaces start here */
635 
636 /*
637  * Very early in the boot some setup of scheduler-specific
638  * parts of proc0 and of some scheduler resources needs to be done.
639  * Called from:
640  *  proc0_init()
641  */
642 void
643 schedinit(void)
644 {
645 	/*
646 	 * Set up the scheduler specific parts of proc0.
647 	 */
648 	proc0.p_sched = NULL; /* XXX */
649 	thread0.td_sched = &td_sched0;
650 	thread0.td_lock = &sched_lock;
651 	mtx_init(&sched_lock, "sched lock", NULL, MTX_SPIN | MTX_RECURSE);
652 }
653 
654 int
655 sched_runnable(void)
656 {
657 #ifdef SMP
658 	return runq_check(&runq) + runq_check(&runq_pcpu[PCPU_GET(cpuid)]);
659 #else
660 	return runq_check(&runq);
661 #endif
662 }
663 
664 int
665 sched_rr_interval(void)
666 {
667 	if (sched_quantum == 0)
668 		sched_quantum = SCHED_QUANTUM;
669 	return (sched_quantum);
670 }
671 
672 /*
673  * We adjust the priority of the current process.  The priority of
674  * a process gets worse as it accumulates CPU time.  The cpu usage
675  * estimator (td_estcpu) is increased here.  resetpriority() will
676  * compute a different priority each time td_estcpu increases by
677  * INVERSE_ESTCPU_WEIGHT
678  * (until MAXPRI is reached).  The cpu usage estimator ramps up
679  * quite quickly when the process is running (linearly), and decays
680  * away exponentially, at a rate which is proportionally slower when
681  * the system is busy.  The basic principle is that the system will
682  * 90% forget that the process used a lot of CPU time in 5 * loadav
683  * seconds.  This causes the system to favor processes which haven't
684  * run much recently, and to round-robin among other processes.
685  */
686 void
687 sched_clock(struct thread *td)
688 {
689 	struct pcpuidlestat *stat;
690 	struct td_sched *ts;
691 
692 	THREAD_LOCK_ASSERT(td, MA_OWNED);
693 	ts = td->td_sched;
694 
695 	ts->ts_cpticks++;
696 	td->td_estcpu = ESTCPULIM(td->td_estcpu + 1);
697 	if ((td->td_estcpu % INVERSE_ESTCPU_WEIGHT) == 0) {
698 		resetpriority(td);
699 		resetpriority_thread(td);
700 	}
701 
702 	/*
703 	 * Force a context switch if the current thread has used up a full
704 	 * quantum (default quantum is 100ms).
705 	 */
706 	if (!TD_IS_IDLETHREAD(td) &&
707 	    ticks - PCPU_GET(switchticks) >= sched_quantum)
708 		td->td_flags |= TDF_NEEDRESCHED;
709 
710 	stat = DPCPU_PTR(idlestat);
711 	stat->oldidlecalls = stat->idlecalls;
712 	stat->idlecalls = 0;
713 }
714 
715 /*
716  * Charge child's scheduling CPU usage to parent.
717  */
718 void
719 sched_exit(struct proc *p, struct thread *td)
720 {
721 
722 	KTR_STATE1(KTR_SCHED, "thread", sched_tdname(td), "proc exit",
723 	    "prio:td", td->td_priority);
724 
725 	PROC_LOCK_ASSERT(p, MA_OWNED);
726 	sched_exit_thread(FIRST_THREAD_IN_PROC(p), td);
727 }
728 
729 void
730 sched_exit_thread(struct thread *td, struct thread *child)
731 {
732 
733 	KTR_STATE1(KTR_SCHED, "thread", sched_tdname(child), "exit",
734 	    "prio:td", child->td_priority);
735 	thread_lock(td);
736 	td->td_estcpu = ESTCPULIM(td->td_estcpu + child->td_estcpu);
737 	thread_unlock(td);
738 	thread_lock(child);
739 	if ((child->td_flags & TDF_NOLOAD) == 0)
740 		sched_load_rem();
741 	thread_unlock(child);
742 }
743 
744 void
745 sched_fork(struct thread *td, struct thread *childtd)
746 {
747 	sched_fork_thread(td, childtd);
748 }
749 
750 void
751 sched_fork_thread(struct thread *td, struct thread *childtd)
752 {
753 	struct td_sched *ts;
754 
755 	childtd->td_estcpu = td->td_estcpu;
756 	childtd->td_lock = &sched_lock;
757 	childtd->td_cpuset = cpuset_ref(td->td_cpuset);
758 	childtd->td_priority = childtd->td_base_pri;
759 	ts = childtd->td_sched;
760 	bzero(ts, sizeof(*ts));
761 	ts->ts_flags |= (td->td_sched->ts_flags & TSF_AFFINITY);
762 }
763 
764 void
765 sched_nice(struct proc *p, int nice)
766 {
767 	struct thread *td;
768 
769 	PROC_LOCK_ASSERT(p, MA_OWNED);
770 	p->p_nice = nice;
771 	FOREACH_THREAD_IN_PROC(p, td) {
772 		thread_lock(td);
773 		resetpriority(td);
774 		resetpriority_thread(td);
775 		thread_unlock(td);
776 	}
777 }
778 
779 void
780 sched_class(struct thread *td, int class)
781 {
782 	THREAD_LOCK_ASSERT(td, MA_OWNED);
783 	td->td_pri_class = class;
784 }
785 
786 /*
787  * Adjust the priority of a thread.
788  */
789 static void
790 sched_priority(struct thread *td, u_char prio)
791 {
792 
793 
794 	KTR_POINT3(KTR_SCHED, "thread", sched_tdname(td), "priority change",
795 	    "prio:%d", td->td_priority, "new prio:%d", prio, KTR_ATTR_LINKED,
796 	    sched_tdname(curthread));
797 	if (td != curthread && prio > td->td_priority) {
798 		KTR_POINT3(KTR_SCHED, "thread", sched_tdname(curthread),
799 		    "lend prio", "prio:%d", td->td_priority, "new prio:%d",
800 		    prio, KTR_ATTR_LINKED, sched_tdname(td));
801 	}
802 	THREAD_LOCK_ASSERT(td, MA_OWNED);
803 	if (td->td_priority == prio)
804 		return;
805 	td->td_priority = prio;
806 	if (TD_ON_RUNQ(td) && td->td_rqindex != (prio / RQ_PPQ)) {
807 		sched_rem(td);
808 		sched_add(td, SRQ_BORING);
809 	}
810 }
811 
812 /*
813  * Update a thread's priority when it is lent another thread's
814  * priority.
815  */
816 void
817 sched_lend_prio(struct thread *td, u_char prio)
818 {
819 
820 	td->td_flags |= TDF_BORROWING;
821 	sched_priority(td, prio);
822 }
823 
824 /*
825  * Restore a thread's priority when priority propagation is
826  * over.  The prio argument is the minimum priority the thread
827  * needs to have to satisfy other possible priority lending
828  * requests.  If the thread's regulary priority is less
829  * important than prio the thread will keep a priority boost
830  * of prio.
831  */
832 void
833 sched_unlend_prio(struct thread *td, u_char prio)
834 {
835 	u_char base_pri;
836 
837 	if (td->td_base_pri >= PRI_MIN_TIMESHARE &&
838 	    td->td_base_pri <= PRI_MAX_TIMESHARE)
839 		base_pri = td->td_user_pri;
840 	else
841 		base_pri = td->td_base_pri;
842 	if (prio >= base_pri) {
843 		td->td_flags &= ~TDF_BORROWING;
844 		sched_prio(td, base_pri);
845 	} else
846 		sched_lend_prio(td, prio);
847 }
848 
849 void
850 sched_prio(struct thread *td, u_char prio)
851 {
852 	u_char oldprio;
853 
854 	/* First, update the base priority. */
855 	td->td_base_pri = prio;
856 
857 	/*
858 	 * If the thread is borrowing another thread's priority, don't ever
859 	 * lower the priority.
860 	 */
861 	if (td->td_flags & TDF_BORROWING && td->td_priority < prio)
862 		return;
863 
864 	/* Change the real priority. */
865 	oldprio = td->td_priority;
866 	sched_priority(td, prio);
867 
868 	/*
869 	 * If the thread is on a turnstile, then let the turnstile update
870 	 * its state.
871 	 */
872 	if (TD_ON_LOCK(td) && oldprio != prio)
873 		turnstile_adjust(td, oldprio);
874 }
875 
876 void
877 sched_user_prio(struct thread *td, u_char prio)
878 {
879 
880 	THREAD_LOCK_ASSERT(td, MA_OWNED);
881 	td->td_base_user_pri = prio;
882 	if (td->td_lend_user_pri <= prio)
883 		return;
884 	td->td_user_pri = prio;
885 }
886 
887 void
888 sched_lend_user_prio(struct thread *td, u_char prio)
889 {
890 
891 	THREAD_LOCK_ASSERT(td, MA_OWNED);
892 	td->td_lend_user_pri = prio;
893 	td->td_user_pri = min(prio, td->td_base_user_pri);
894 	if (td->td_priority > td->td_user_pri)
895 		sched_prio(td, td->td_user_pri);
896 	else if (td->td_priority != td->td_user_pri)
897 		td->td_flags |= TDF_NEEDRESCHED;
898 }
899 
900 void
901 sched_sleep(struct thread *td, int pri)
902 {
903 
904 	THREAD_LOCK_ASSERT(td, MA_OWNED);
905 	td->td_slptick = ticks;
906 	td->td_sched->ts_slptime = 0;
907 	if (pri != 0 && PRI_BASE(td->td_pri_class) == PRI_TIMESHARE)
908 		sched_prio(td, pri);
909 	if (TD_IS_SUSPENDED(td) || pri >= PSOCK)
910 		td->td_flags |= TDF_CANSWAP;
911 }
912 
913 void
914 sched_switch(struct thread *td, struct thread *newtd, int flags)
915 {
916 	struct mtx *tmtx;
917 	struct td_sched *ts;
918 	struct proc *p;
919 
920 	tmtx = NULL;
921 	ts = td->td_sched;
922 	p = td->td_proc;
923 
924 	THREAD_LOCK_ASSERT(td, MA_OWNED);
925 
926 	/*
927 	 * Switch to the sched lock to fix things up and pick
928 	 * a new thread.
929 	 * Block the td_lock in order to avoid breaking the critical path.
930 	 */
931 	if (td->td_lock != &sched_lock) {
932 		mtx_lock_spin(&sched_lock);
933 		tmtx = thread_lock_block(td);
934 	}
935 
936 	if ((td->td_flags & TDF_NOLOAD) == 0)
937 		sched_load_rem();
938 
939 	td->td_lastcpu = td->td_oncpu;
940 	if (!(flags & SW_PREEMPT))
941 		td->td_flags &= ~TDF_NEEDRESCHED;
942 	td->td_owepreempt = 0;
943 	td->td_oncpu = NOCPU;
944 
945 	/*
946 	 * At the last moment, if this thread is still marked RUNNING,
947 	 * then put it back on the run queue as it has not been suspended
948 	 * or stopped or any thing else similar.  We never put the idle
949 	 * threads on the run queue, however.
950 	 */
951 	if (td->td_flags & TDF_IDLETD) {
952 		TD_SET_CAN_RUN(td);
953 #ifdef SMP
954 		CPU_CLR(PCPU_GET(cpuid), &idle_cpus_mask);
955 #endif
956 	} else {
957 		if (TD_IS_RUNNING(td)) {
958 			/* Put us back on the run queue. */
959 			sched_add(td, (flags & SW_PREEMPT) ?
960 			    SRQ_OURSELF|SRQ_YIELDING|SRQ_PREEMPTED :
961 			    SRQ_OURSELF|SRQ_YIELDING);
962 		}
963 	}
964 	if (newtd) {
965 		/*
966 		 * The thread we are about to run needs to be counted
967 		 * as if it had been added to the run queue and selected.
968 		 * It came from:
969 		 * * A preemption
970 		 * * An upcall
971 		 * * A followon
972 		 */
973 		KASSERT((newtd->td_inhibitors == 0),
974 			("trying to run inhibited thread"));
975 		newtd->td_flags |= TDF_DIDRUN;
976         	TD_SET_RUNNING(newtd);
977 		if ((newtd->td_flags & TDF_NOLOAD) == 0)
978 			sched_load_add();
979 	} else {
980 		newtd = choosethread();
981 		MPASS(newtd->td_lock == &sched_lock);
982 	}
983 
984 	if (td != newtd) {
985 #ifdef	HWPMC_HOOKS
986 		if (PMC_PROC_IS_USING_PMCS(td->td_proc))
987 			PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_OUT);
988 #endif
989                 /* I feel sleepy */
990 		lock_profile_release_lock(&sched_lock.lock_object);
991 #ifdef KDTRACE_HOOKS
992 		/*
993 		 * If DTrace has set the active vtime enum to anything
994 		 * other than INACTIVE (0), then it should have set the
995 		 * function to call.
996 		 */
997 		if (dtrace_vtime_active)
998 			(*dtrace_vtime_switch_func)(newtd);
999 #endif
1000 
1001 		cpu_switch(td, newtd, tmtx != NULL ? tmtx : td->td_lock);
1002 		lock_profile_obtain_lock_success(&sched_lock.lock_object,
1003 		    0, 0, __FILE__, __LINE__);
1004 		/*
1005 		 * Where am I?  What year is it?
1006 		 * We are in the same thread that went to sleep above,
1007 		 * but any amount of time may have passed. All our context
1008 		 * will still be available as will local variables.
1009 		 * PCPU values however may have changed as we may have
1010 		 * changed CPU so don't trust cached values of them.
1011 		 * New threads will go to fork_exit() instead of here
1012 		 * so if you change things here you may need to change
1013 		 * things there too.
1014 		 *
1015 		 * If the thread above was exiting it will never wake
1016 		 * up again here, so either it has saved everything it
1017 		 * needed to, or the thread_wait() or wait() will
1018 		 * need to reap it.
1019 		 */
1020 #ifdef	HWPMC_HOOKS
1021 		if (PMC_PROC_IS_USING_PMCS(td->td_proc))
1022 			PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_IN);
1023 #endif
1024 	}
1025 
1026 #ifdef SMP
1027 	if (td->td_flags & TDF_IDLETD)
1028 		CPU_SET(PCPU_GET(cpuid), &idle_cpus_mask);
1029 #endif
1030 	sched_lock.mtx_lock = (uintptr_t)td;
1031 	td->td_oncpu = PCPU_GET(cpuid);
1032 	MPASS(td->td_lock == &sched_lock);
1033 }
1034 
1035 void
1036 sched_wakeup(struct thread *td)
1037 {
1038 	struct td_sched *ts;
1039 
1040 	THREAD_LOCK_ASSERT(td, MA_OWNED);
1041 	ts = td->td_sched;
1042 	td->td_flags &= ~TDF_CANSWAP;
1043 	if (ts->ts_slptime > 1) {
1044 		updatepri(td);
1045 		resetpriority(td);
1046 	}
1047 	td->td_slptick = 0;
1048 	ts->ts_slptime = 0;
1049 	sched_add(td, SRQ_BORING);
1050 }
1051 
1052 #ifdef SMP
1053 static int
1054 forward_wakeup(int cpunum)
1055 {
1056 	struct pcpu *pc;
1057 	cpuset_t dontuse, map, map2;
1058 	u_int id, me;
1059 	int iscpuset;
1060 
1061 	mtx_assert(&sched_lock, MA_OWNED);
1062 
1063 	CTR0(KTR_RUNQ, "forward_wakeup()");
1064 
1065 	if ((!forward_wakeup_enabled) ||
1066 	     (forward_wakeup_use_mask == 0 && forward_wakeup_use_loop == 0))
1067 		return (0);
1068 	if (!smp_started || cold || panicstr)
1069 		return (0);
1070 
1071 	forward_wakeups_requested++;
1072 
1073 	/*
1074 	 * Check the idle mask we received against what we calculated
1075 	 * before in the old version.
1076 	 */
1077 	me = PCPU_GET(cpuid);
1078 
1079 	/* Don't bother if we should be doing it ourself. */
1080 	if (CPU_ISSET(me, &idle_cpus_mask) &&
1081 	    (cpunum == NOCPU || me == cpunum))
1082 		return (0);
1083 
1084 	CPU_SETOF(me, &dontuse);
1085 	CPU_OR(&dontuse, &stopped_cpus);
1086 	CPU_OR(&dontuse, &hlt_cpus_mask);
1087 	CPU_ZERO(&map2);
1088 	if (forward_wakeup_use_loop) {
1089 		STAILQ_FOREACH(pc, &cpuhead, pc_allcpu) {
1090 			id = pc->pc_cpuid;
1091 			if (!CPU_ISSET(id, &dontuse) &&
1092 			    pc->pc_curthread == pc->pc_idlethread) {
1093 				CPU_SET(id, &map2);
1094 			}
1095 		}
1096 	}
1097 
1098 	if (forward_wakeup_use_mask) {
1099 		map = idle_cpus_mask;
1100 		CPU_NAND(&map, &dontuse);
1101 
1102 		/* If they are both on, compare and use loop if different. */
1103 		if (forward_wakeup_use_loop) {
1104 			if (CPU_CMP(&map, &map2)) {
1105 				printf("map != map2, loop method preferred\n");
1106 				map = map2;
1107 			}
1108 		}
1109 	} else {
1110 		map = map2;
1111 	}
1112 
1113 	/* If we only allow a specific CPU, then mask off all the others. */
1114 	if (cpunum != NOCPU) {
1115 		KASSERT((cpunum <= mp_maxcpus),("forward_wakeup: bad cpunum."));
1116 		iscpuset = CPU_ISSET(cpunum, &map);
1117 		if (iscpuset == 0)
1118 			CPU_ZERO(&map);
1119 		else
1120 			CPU_SETOF(cpunum, &map);
1121 	}
1122 	if (!CPU_EMPTY(&map)) {
1123 		forward_wakeups_delivered++;
1124 		STAILQ_FOREACH(pc, &cpuhead, pc_allcpu) {
1125 			id = pc->pc_cpuid;
1126 			if (!CPU_ISSET(id, &map))
1127 				continue;
1128 			if (cpu_idle_wakeup(pc->pc_cpuid))
1129 				CPU_CLR(id, &map);
1130 		}
1131 		if (!CPU_EMPTY(&map))
1132 			ipi_selected(map, IPI_AST);
1133 		return (1);
1134 	}
1135 	if (cpunum == NOCPU)
1136 		printf("forward_wakeup: Idle processor not found\n");
1137 	return (0);
1138 }
1139 
1140 static void
1141 kick_other_cpu(int pri, int cpuid)
1142 {
1143 	struct pcpu *pcpu;
1144 	int cpri;
1145 
1146 	pcpu = pcpu_find(cpuid);
1147 	if (CPU_ISSET(cpuid, &idle_cpus_mask)) {
1148 		forward_wakeups_delivered++;
1149 		if (!cpu_idle_wakeup(cpuid))
1150 			ipi_cpu(cpuid, IPI_AST);
1151 		return;
1152 	}
1153 
1154 	cpri = pcpu->pc_curthread->td_priority;
1155 	if (pri >= cpri)
1156 		return;
1157 
1158 #if defined(IPI_PREEMPTION) && defined(PREEMPTION)
1159 #if !defined(FULL_PREEMPTION)
1160 	if (pri <= PRI_MAX_ITHD)
1161 #endif /* ! FULL_PREEMPTION */
1162 	{
1163 		ipi_cpu(cpuid, IPI_PREEMPT);
1164 		return;
1165 	}
1166 #endif /* defined(IPI_PREEMPTION) && defined(PREEMPTION) */
1167 
1168 	pcpu->pc_curthread->td_flags |= TDF_NEEDRESCHED;
1169 	ipi_cpu(cpuid, IPI_AST);
1170 	return;
1171 }
1172 #endif /* SMP */
1173 
1174 #ifdef SMP
1175 static int
1176 sched_pickcpu(struct thread *td)
1177 {
1178 	int best, cpu;
1179 
1180 	mtx_assert(&sched_lock, MA_OWNED);
1181 
1182 	if (THREAD_CAN_SCHED(td, td->td_lastcpu))
1183 		best = td->td_lastcpu;
1184 	else
1185 		best = NOCPU;
1186 	CPU_FOREACH(cpu) {
1187 		if (!THREAD_CAN_SCHED(td, cpu))
1188 			continue;
1189 
1190 		if (best == NOCPU)
1191 			best = cpu;
1192 		else if (runq_length[cpu] < runq_length[best])
1193 			best = cpu;
1194 	}
1195 	KASSERT(best != NOCPU, ("no valid CPUs"));
1196 
1197 	return (best);
1198 }
1199 #endif
1200 
1201 void
1202 sched_add(struct thread *td, int flags)
1203 #ifdef SMP
1204 {
1205 	cpuset_t tidlemsk;
1206 	struct td_sched *ts;
1207 	u_int cpu, cpuid;
1208 	int forwarded = 0;
1209 	int single_cpu = 0;
1210 
1211 	ts = td->td_sched;
1212 	THREAD_LOCK_ASSERT(td, MA_OWNED);
1213 	KASSERT((td->td_inhibitors == 0),
1214 	    ("sched_add: trying to run inhibited thread"));
1215 	KASSERT((TD_CAN_RUN(td) || TD_IS_RUNNING(td)),
1216 	    ("sched_add: bad thread state"));
1217 	KASSERT(td->td_flags & TDF_INMEM,
1218 	    ("sched_add: thread swapped out"));
1219 
1220 	KTR_STATE2(KTR_SCHED, "thread", sched_tdname(td), "runq add",
1221 	    "prio:%d", td->td_priority, KTR_ATTR_LINKED,
1222 	    sched_tdname(curthread));
1223 	KTR_POINT1(KTR_SCHED, "thread", sched_tdname(curthread), "wokeup",
1224 	    KTR_ATTR_LINKED, sched_tdname(td));
1225 
1226 
1227 	/*
1228 	 * Now that the thread is moving to the run-queue, set the lock
1229 	 * to the scheduler's lock.
1230 	 */
1231 	if (td->td_lock != &sched_lock) {
1232 		mtx_lock_spin(&sched_lock);
1233 		thread_lock_set(td, &sched_lock);
1234 	}
1235 	TD_SET_RUNQ(td);
1236 
1237 	/*
1238 	 * If SMP is started and the thread is pinned or otherwise limited to
1239 	 * a specific set of CPUs, queue the thread to a per-CPU run queue.
1240 	 * Otherwise, queue the thread to the global run queue.
1241 	 *
1242 	 * If SMP has not yet been started we must use the global run queue
1243 	 * as per-CPU state may not be initialized yet and we may crash if we
1244 	 * try to access the per-CPU run queues.
1245 	 */
1246 	if (smp_started && (td->td_pinned != 0 || td->td_flags & TDF_BOUND ||
1247 	    ts->ts_flags & TSF_AFFINITY)) {
1248 		if (td->td_pinned != 0)
1249 			cpu = td->td_lastcpu;
1250 		else if (td->td_flags & TDF_BOUND) {
1251 			/* Find CPU from bound runq. */
1252 			KASSERT(SKE_RUNQ_PCPU(ts),
1253 			    ("sched_add: bound td_sched not on cpu runq"));
1254 			cpu = ts->ts_runq - &runq_pcpu[0];
1255 		} else
1256 			/* Find a valid CPU for our cpuset */
1257 			cpu = sched_pickcpu(td);
1258 		ts->ts_runq = &runq_pcpu[cpu];
1259 		single_cpu = 1;
1260 		CTR3(KTR_RUNQ,
1261 		    "sched_add: Put td_sched:%p(td:%p) on cpu%d runq", ts, td,
1262 		    cpu);
1263 	} else {
1264 		CTR2(KTR_RUNQ,
1265 		    "sched_add: adding td_sched:%p (td:%p) to gbl runq", ts,
1266 		    td);
1267 		cpu = NOCPU;
1268 		ts->ts_runq = &runq;
1269 	}
1270 
1271 	cpuid = PCPU_GET(cpuid);
1272 	if (single_cpu && cpu != cpuid) {
1273 	        kick_other_cpu(td->td_priority, cpu);
1274 	} else {
1275 		if (!single_cpu) {
1276 			tidlemsk = idle_cpus_mask;
1277 			CPU_NAND(&tidlemsk, &hlt_cpus_mask);
1278 			CPU_CLR(cpuid, &tidlemsk);
1279 
1280 			if (!CPU_ISSET(cpuid, &idle_cpus_mask) &&
1281 			    ((flags & SRQ_INTR) == 0) &&
1282 			    !CPU_EMPTY(&tidlemsk))
1283 				forwarded = forward_wakeup(cpu);
1284 		}
1285 
1286 		if (!forwarded) {
1287 			if ((flags & SRQ_YIELDING) == 0 && maybe_preempt(td))
1288 				return;
1289 			else
1290 				maybe_resched(td);
1291 		}
1292 	}
1293 
1294 	if ((td->td_flags & TDF_NOLOAD) == 0)
1295 		sched_load_add();
1296 	runq_add(ts->ts_runq, td, flags);
1297 	if (cpu != NOCPU)
1298 		runq_length[cpu]++;
1299 }
1300 #else /* SMP */
1301 {
1302 	struct td_sched *ts;
1303 
1304 	ts = td->td_sched;
1305 	THREAD_LOCK_ASSERT(td, MA_OWNED);
1306 	KASSERT((td->td_inhibitors == 0),
1307 	    ("sched_add: trying to run inhibited thread"));
1308 	KASSERT((TD_CAN_RUN(td) || TD_IS_RUNNING(td)),
1309 	    ("sched_add: bad thread state"));
1310 	KASSERT(td->td_flags & TDF_INMEM,
1311 	    ("sched_add: thread swapped out"));
1312 	KTR_STATE2(KTR_SCHED, "thread", sched_tdname(td), "runq add",
1313 	    "prio:%d", td->td_priority, KTR_ATTR_LINKED,
1314 	    sched_tdname(curthread));
1315 	KTR_POINT1(KTR_SCHED, "thread", sched_tdname(curthread), "wokeup",
1316 	    KTR_ATTR_LINKED, sched_tdname(td));
1317 
1318 	/*
1319 	 * Now that the thread is moving to the run-queue, set the lock
1320 	 * to the scheduler's lock.
1321 	 */
1322 	if (td->td_lock != &sched_lock) {
1323 		mtx_lock_spin(&sched_lock);
1324 		thread_lock_set(td, &sched_lock);
1325 	}
1326 	TD_SET_RUNQ(td);
1327 	CTR2(KTR_RUNQ, "sched_add: adding td_sched:%p (td:%p) to runq", ts, td);
1328 	ts->ts_runq = &runq;
1329 
1330 	/*
1331 	 * If we are yielding (on the way out anyhow) or the thread
1332 	 * being saved is US, then don't try be smart about preemption
1333 	 * or kicking off another CPU as it won't help and may hinder.
1334 	 * In the YIEDLING case, we are about to run whoever is being
1335 	 * put in the queue anyhow, and in the OURSELF case, we are
1336 	 * puting ourself on the run queue which also only happens
1337 	 * when we are about to yield.
1338 	 */
1339 	if ((flags & SRQ_YIELDING) == 0) {
1340 		if (maybe_preempt(td))
1341 			return;
1342 	}
1343 	if ((td->td_flags & TDF_NOLOAD) == 0)
1344 		sched_load_add();
1345 	runq_add(ts->ts_runq, td, flags);
1346 	maybe_resched(td);
1347 }
1348 #endif /* SMP */
1349 
1350 void
1351 sched_rem(struct thread *td)
1352 {
1353 	struct td_sched *ts;
1354 
1355 	ts = td->td_sched;
1356 	KASSERT(td->td_flags & TDF_INMEM,
1357 	    ("sched_rem: thread swapped out"));
1358 	KASSERT(TD_ON_RUNQ(td),
1359 	    ("sched_rem: thread not on run queue"));
1360 	mtx_assert(&sched_lock, MA_OWNED);
1361 	KTR_STATE2(KTR_SCHED, "thread", sched_tdname(td), "runq rem",
1362 	    "prio:%d", td->td_priority, KTR_ATTR_LINKED,
1363 	    sched_tdname(curthread));
1364 
1365 	if ((td->td_flags & TDF_NOLOAD) == 0)
1366 		sched_load_rem();
1367 #ifdef SMP
1368 	if (ts->ts_runq != &runq)
1369 		runq_length[ts->ts_runq - runq_pcpu]--;
1370 #endif
1371 	runq_remove(ts->ts_runq, td);
1372 	TD_SET_CAN_RUN(td);
1373 }
1374 
1375 /*
1376  * Select threads to run.  Note that running threads still consume a
1377  * slot.
1378  */
1379 struct thread *
1380 sched_choose(void)
1381 {
1382 	struct thread *td;
1383 	struct runq *rq;
1384 
1385 	mtx_assert(&sched_lock,  MA_OWNED);
1386 #ifdef SMP
1387 	struct thread *tdcpu;
1388 
1389 	rq = &runq;
1390 	td = runq_choose_fuzz(&runq, runq_fuzz);
1391 	tdcpu = runq_choose(&runq_pcpu[PCPU_GET(cpuid)]);
1392 
1393 	if (td == NULL ||
1394 	    (tdcpu != NULL &&
1395 	     tdcpu->td_priority < td->td_priority)) {
1396 		CTR2(KTR_RUNQ, "choosing td %p from pcpu runq %d", tdcpu,
1397 		     PCPU_GET(cpuid));
1398 		td = tdcpu;
1399 		rq = &runq_pcpu[PCPU_GET(cpuid)];
1400 	} else {
1401 		CTR1(KTR_RUNQ, "choosing td_sched %p from main runq", td);
1402 	}
1403 
1404 #else
1405 	rq = &runq;
1406 	td = runq_choose(&runq);
1407 #endif
1408 
1409 	if (td) {
1410 #ifdef SMP
1411 		if (td == tdcpu)
1412 			runq_length[PCPU_GET(cpuid)]--;
1413 #endif
1414 		runq_remove(rq, td);
1415 		td->td_flags |= TDF_DIDRUN;
1416 
1417 		KASSERT(td->td_flags & TDF_INMEM,
1418 		    ("sched_choose: thread swapped out"));
1419 		return (td);
1420 	}
1421 	return (PCPU_GET(idlethread));
1422 }
1423 
1424 void
1425 sched_preempt(struct thread *td)
1426 {
1427 	thread_lock(td);
1428 	if (td->td_critnest > 1)
1429 		td->td_owepreempt = 1;
1430 	else
1431 		mi_switch(SW_INVOL | SW_PREEMPT | SWT_PREEMPT, NULL);
1432 	thread_unlock(td);
1433 }
1434 
1435 void
1436 sched_userret(struct thread *td)
1437 {
1438 	/*
1439 	 * XXX we cheat slightly on the locking here to avoid locking in
1440 	 * the usual case.  Setting td_priority here is essentially an
1441 	 * incomplete workaround for not setting it properly elsewhere.
1442 	 * Now that some interrupt handlers are threads, not setting it
1443 	 * properly elsewhere can clobber it in the window between setting
1444 	 * it here and returning to user mode, so don't waste time setting
1445 	 * it perfectly here.
1446 	 */
1447 	KASSERT((td->td_flags & TDF_BORROWING) == 0,
1448 	    ("thread with borrowed priority returning to userland"));
1449 	if (td->td_priority != td->td_user_pri) {
1450 		thread_lock(td);
1451 		td->td_priority = td->td_user_pri;
1452 		td->td_base_pri = td->td_user_pri;
1453 		thread_unlock(td);
1454 	}
1455 }
1456 
1457 void
1458 sched_bind(struct thread *td, int cpu)
1459 {
1460 	struct td_sched *ts;
1461 
1462 	THREAD_LOCK_ASSERT(td, MA_OWNED|MA_NOTRECURSED);
1463 	KASSERT(td == curthread, ("sched_bind: can only bind curthread"));
1464 
1465 	ts = td->td_sched;
1466 
1467 	td->td_flags |= TDF_BOUND;
1468 #ifdef SMP
1469 	ts->ts_runq = &runq_pcpu[cpu];
1470 	if (PCPU_GET(cpuid) == cpu)
1471 		return;
1472 
1473 	mi_switch(SW_VOL, NULL);
1474 #endif
1475 }
1476 
1477 void
1478 sched_unbind(struct thread* td)
1479 {
1480 	THREAD_LOCK_ASSERT(td, MA_OWNED);
1481 	KASSERT(td == curthread, ("sched_unbind: can only bind curthread"));
1482 	td->td_flags &= ~TDF_BOUND;
1483 }
1484 
1485 int
1486 sched_is_bound(struct thread *td)
1487 {
1488 	THREAD_LOCK_ASSERT(td, MA_OWNED);
1489 	return (td->td_flags & TDF_BOUND);
1490 }
1491 
1492 void
1493 sched_relinquish(struct thread *td)
1494 {
1495 	thread_lock(td);
1496 	mi_switch(SW_VOL | SWT_RELINQUISH, NULL);
1497 	thread_unlock(td);
1498 }
1499 
1500 int
1501 sched_load(void)
1502 {
1503 	return (sched_tdcnt);
1504 }
1505 
1506 int
1507 sched_sizeof_proc(void)
1508 {
1509 	return (sizeof(struct proc));
1510 }
1511 
1512 int
1513 sched_sizeof_thread(void)
1514 {
1515 	return (sizeof(struct thread) + sizeof(struct td_sched));
1516 }
1517 
1518 fixpt_t
1519 sched_pctcpu(struct thread *td)
1520 {
1521 	struct td_sched *ts;
1522 
1523 	THREAD_LOCK_ASSERT(td, MA_OWNED);
1524 	ts = td->td_sched;
1525 	return (ts->ts_pctcpu);
1526 }
1527 
1528 void
1529 sched_tick(int cnt)
1530 {
1531 }
1532 
1533 /*
1534  * The actual idle process.
1535  */
1536 void
1537 sched_idletd(void *dummy)
1538 {
1539 	struct pcpuidlestat *stat;
1540 
1541 	stat = DPCPU_PTR(idlestat);
1542 	for (;;) {
1543 		mtx_assert(&Giant, MA_NOTOWNED);
1544 
1545 		while (sched_runnable() == 0) {
1546 			cpu_idle(stat->idlecalls + stat->oldidlecalls > 64);
1547 			stat->idlecalls++;
1548 		}
1549 
1550 		mtx_lock_spin(&sched_lock);
1551 		mi_switch(SW_VOL | SWT_IDLE, NULL);
1552 		mtx_unlock_spin(&sched_lock);
1553 	}
1554 }
1555 
1556 /*
1557  * A CPU is entering for the first time or a thread is exiting.
1558  */
1559 void
1560 sched_throw(struct thread *td)
1561 {
1562 	/*
1563 	 * Correct spinlock nesting.  The idle thread context that we are
1564 	 * borrowing was created so that it would start out with a single
1565 	 * spin lock (sched_lock) held in fork_trampoline().  Since we've
1566 	 * explicitly acquired locks in this function, the nesting count
1567 	 * is now 2 rather than 1.  Since we are nested, calling
1568 	 * spinlock_exit() will simply adjust the counts without allowing
1569 	 * spin lock using code to interrupt us.
1570 	 */
1571 	if (td == NULL) {
1572 		mtx_lock_spin(&sched_lock);
1573 		spinlock_exit();
1574 	} else {
1575 		lock_profile_release_lock(&sched_lock.lock_object);
1576 		MPASS(td->td_lock == &sched_lock);
1577 	}
1578 	mtx_assert(&sched_lock, MA_OWNED);
1579 	KASSERT(curthread->td_md.md_spinlock_count == 1, ("invalid count"));
1580 	PCPU_SET(switchtime, cpu_ticks());
1581 	PCPU_SET(switchticks, ticks);
1582 	cpu_throw(td, choosethread());	/* doesn't return */
1583 }
1584 
1585 void
1586 sched_fork_exit(struct thread *td)
1587 {
1588 
1589 	/*
1590 	 * Finish setting up thread glue so that it begins execution in a
1591 	 * non-nested critical section with sched_lock held but not recursed.
1592 	 */
1593 	td->td_oncpu = PCPU_GET(cpuid);
1594 	sched_lock.mtx_lock = (uintptr_t)td;
1595 	lock_profile_obtain_lock_success(&sched_lock.lock_object,
1596 	    0, 0, __FILE__, __LINE__);
1597 	THREAD_LOCK_ASSERT(td, MA_OWNED | MA_NOTRECURSED);
1598 }
1599 
1600 char *
1601 sched_tdname(struct thread *td)
1602 {
1603 #ifdef KTR
1604 	struct td_sched *ts;
1605 
1606 	ts = td->td_sched;
1607 	if (ts->ts_name[0] == '\0')
1608 		snprintf(ts->ts_name, sizeof(ts->ts_name),
1609 		    "%s tid %d", td->td_name, td->td_tid);
1610 	return (ts->ts_name);
1611 #else
1612 	return (td->td_name);
1613 #endif
1614 }
1615 
1616 void
1617 sched_affinity(struct thread *td)
1618 {
1619 #ifdef SMP
1620 	struct td_sched *ts;
1621 	int cpu;
1622 
1623 	THREAD_LOCK_ASSERT(td, MA_OWNED);
1624 
1625 	/*
1626 	 * Set the TSF_AFFINITY flag if there is at least one CPU this
1627 	 * thread can't run on.
1628 	 */
1629 	ts = td->td_sched;
1630 	ts->ts_flags &= ~TSF_AFFINITY;
1631 	CPU_FOREACH(cpu) {
1632 		if (!THREAD_CAN_SCHED(td, cpu)) {
1633 			ts->ts_flags |= TSF_AFFINITY;
1634 			break;
1635 		}
1636 	}
1637 
1638 	/*
1639 	 * If this thread can run on all CPUs, nothing else to do.
1640 	 */
1641 	if (!(ts->ts_flags & TSF_AFFINITY))
1642 		return;
1643 
1644 	/* Pinned threads and bound threads should be left alone. */
1645 	if (td->td_pinned != 0 || td->td_flags & TDF_BOUND)
1646 		return;
1647 
1648 	switch (td->td_state) {
1649 	case TDS_RUNQ:
1650 		/*
1651 		 * If we are on a per-CPU runqueue that is in the set,
1652 		 * then nothing needs to be done.
1653 		 */
1654 		if (ts->ts_runq != &runq &&
1655 		    THREAD_CAN_SCHED(td, ts->ts_runq - runq_pcpu))
1656 			return;
1657 
1658 		/* Put this thread on a valid per-CPU runqueue. */
1659 		sched_rem(td);
1660 		sched_add(td, SRQ_BORING);
1661 		break;
1662 	case TDS_RUNNING:
1663 		/*
1664 		 * See if our current CPU is in the set.  If not, force a
1665 		 * context switch.
1666 		 */
1667 		if (THREAD_CAN_SCHED(td, td->td_oncpu))
1668 			return;
1669 
1670 		td->td_flags |= TDF_NEEDRESCHED;
1671 		if (td != curthread)
1672 			ipi_cpu(cpu, IPI_AST);
1673 		break;
1674 	default:
1675 		break;
1676 	}
1677 #endif
1678 }
1679