xref: /freebsd/sys/kern/sched_4bsd.c (revision 349cc55c9796c4596a5b9904cd3281af295f878f)
1 /*-
2  * SPDX-License-Identifier: BSD-3-Clause
3  *
4  * Copyright (c) 1982, 1986, 1990, 1991, 1993
5  *	The Regents of the University of California.  All rights reserved.
6  * (c) UNIX System Laboratories, Inc.
7  * All or some portions of this file are derived from material licensed
8  * to the University of California by American Telephone and Telegraph
9  * Co. or Unix System Laboratories, Inc. and are reproduced herein with
10  * the permission of UNIX System Laboratories, Inc.
11  *
12  * Redistribution and use in source and binary forms, with or without
13  * modification, are permitted provided that the following conditions
14  * are met:
15  * 1. Redistributions of source code must retain the above copyright
16  *    notice, this list of conditions and the following disclaimer.
17  * 2. Redistributions in binary form must reproduce the above copyright
18  *    notice, this list of conditions and the following disclaimer in the
19  *    documentation and/or other materials provided with the distribution.
20  * 3. Neither the name of the University nor the names of its contributors
21  *    may be used to endorse or promote products derived from this software
22  *    without specific prior written permission.
23  *
24  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
25  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
26  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
27  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
28  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
29  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
30  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
31  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
32  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
33  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
34  * SUCH DAMAGE.
35  */
36 
37 #include <sys/cdefs.h>
38 __FBSDID("$FreeBSD$");
39 
40 #include "opt_hwpmc_hooks.h"
41 #include "opt_sched.h"
42 
43 #include <sys/param.h>
44 #include <sys/systm.h>
45 #include <sys/cpuset.h>
46 #include <sys/kernel.h>
47 #include <sys/ktr.h>
48 #include <sys/lock.h>
49 #include <sys/kthread.h>
50 #include <sys/mutex.h>
51 #include <sys/proc.h>
52 #include <sys/resourcevar.h>
53 #include <sys/sched.h>
54 #include <sys/sdt.h>
55 #include <sys/smp.h>
56 #include <sys/sysctl.h>
57 #include <sys/sx.h>
58 #include <sys/turnstile.h>
59 #include <sys/umtxvar.h>
60 #include <machine/pcb.h>
61 #include <machine/smp.h>
62 
63 #ifdef HWPMC_HOOKS
64 #include <sys/pmckern.h>
65 #endif
66 
67 #ifdef KDTRACE_HOOKS
68 #include <sys/dtrace_bsd.h>
69 int __read_mostly		dtrace_vtime_active;
70 dtrace_vtime_switch_func_t	dtrace_vtime_switch_func;
71 #endif
72 
73 /*
74  * INVERSE_ESTCPU_WEIGHT is only suitable for statclock() frequencies in
75  * the range 100-256 Hz (approximately).
76  */
77 #define	ESTCPULIM(e) \
78     min((e), INVERSE_ESTCPU_WEIGHT * (NICE_WEIGHT * (PRIO_MAX - PRIO_MIN) - \
79     RQ_PPQ) + INVERSE_ESTCPU_WEIGHT - 1)
80 #ifdef SMP
81 #define	INVERSE_ESTCPU_WEIGHT	(8 * smp_cpus)
82 #else
83 #define	INVERSE_ESTCPU_WEIGHT	8	/* 1 / (priorities per estcpu level). */
84 #endif
85 #define	NICE_WEIGHT		1	/* Priorities per nice level. */
86 
87 #define	TS_NAME_LEN (MAXCOMLEN + sizeof(" td ") + sizeof(__XSTRING(UINT_MAX)))
88 
89 /*
90  * The schedulable entity that runs a context.
91  * This is  an extension to the thread structure and is tailored to
92  * the requirements of this scheduler.
93  * All fields are protected by the scheduler lock.
94  */
95 struct td_sched {
96 	fixpt_t		ts_pctcpu;	/* %cpu during p_swtime. */
97 	u_int		ts_estcpu;	/* Estimated cpu utilization. */
98 	int		ts_cpticks;	/* Ticks of cpu time. */
99 	int		ts_slptime;	/* Seconds !RUNNING. */
100 	int		ts_slice;	/* Remaining part of time slice. */
101 	int		ts_flags;
102 	struct runq	*ts_runq;	/* runq the thread is currently on */
103 #ifdef KTR
104 	char		ts_name[TS_NAME_LEN];
105 #endif
106 };
107 
108 /* flags kept in td_flags */
109 #define TDF_DIDRUN	TDF_SCHED0	/* thread actually ran. */
110 #define TDF_BOUND	TDF_SCHED1	/* Bound to one CPU. */
111 #define	TDF_SLICEEND	TDF_SCHED2	/* Thread time slice is over. */
112 
113 /* flags kept in ts_flags */
114 #define	TSF_AFFINITY	0x0001		/* Has a non-"full" CPU set. */
115 
116 #define SKE_RUNQ_PCPU(ts)						\
117     ((ts)->ts_runq != 0 && (ts)->ts_runq != &runq)
118 
119 #define	THREAD_CAN_SCHED(td, cpu)	\
120     CPU_ISSET((cpu), &(td)->td_cpuset->cs_mask)
121 
122 _Static_assert(sizeof(struct thread) + sizeof(struct td_sched) <=
123     sizeof(struct thread0_storage),
124     "increase struct thread0_storage.t0st_sched size");
125 
126 static struct mtx sched_lock;
127 
128 static int	realstathz = 127; /* stathz is sometimes 0 and run off of hz. */
129 static int	sched_tdcnt;	/* Total runnable threads in the system. */
130 static int	sched_slice = 12; /* Thread run time before rescheduling. */
131 
132 static void	setup_runqs(void);
133 static void	schedcpu(void);
134 static void	schedcpu_thread(void);
135 static void	sched_priority(struct thread *td, u_char prio);
136 static void	sched_setup(void *dummy);
137 static void	maybe_resched(struct thread *td);
138 static void	updatepri(struct thread *td);
139 static void	resetpriority(struct thread *td);
140 static void	resetpriority_thread(struct thread *td);
141 #ifdef SMP
142 static int	sched_pickcpu(struct thread *td);
143 static int	forward_wakeup(int cpunum);
144 static void	kick_other_cpu(int pri, int cpuid);
145 #endif
146 
147 static struct kproc_desc sched_kp = {
148         "schedcpu",
149         schedcpu_thread,
150         NULL
151 };
152 SYSINIT(schedcpu, SI_SUB_LAST, SI_ORDER_FIRST, kproc_start,
153     &sched_kp);
154 SYSINIT(sched_setup, SI_SUB_RUN_QUEUE, SI_ORDER_FIRST, sched_setup, NULL);
155 
156 static void sched_initticks(void *dummy);
157 SYSINIT(sched_initticks, SI_SUB_CLOCKS, SI_ORDER_THIRD, sched_initticks,
158     NULL);
159 
160 /*
161  * Global run queue.
162  */
163 static struct runq runq;
164 
165 #ifdef SMP
166 /*
167  * Per-CPU run queues
168  */
169 static struct runq runq_pcpu[MAXCPU];
170 long runq_length[MAXCPU];
171 
172 static cpuset_t idle_cpus_mask;
173 #endif
174 
175 struct pcpuidlestat {
176 	u_int idlecalls;
177 	u_int oldidlecalls;
178 };
179 DPCPU_DEFINE_STATIC(struct pcpuidlestat, idlestat);
180 
181 static void
182 setup_runqs(void)
183 {
184 #ifdef SMP
185 	int i;
186 
187 	for (i = 0; i < MAXCPU; ++i)
188 		runq_init(&runq_pcpu[i]);
189 #endif
190 
191 	runq_init(&runq);
192 }
193 
194 static int
195 sysctl_kern_quantum(SYSCTL_HANDLER_ARGS)
196 {
197 	int error, new_val, period;
198 
199 	period = 1000000 / realstathz;
200 	new_val = period * sched_slice;
201 	error = sysctl_handle_int(oidp, &new_val, 0, req);
202 	if (error != 0 || req->newptr == NULL)
203 		return (error);
204 	if (new_val <= 0)
205 		return (EINVAL);
206 	sched_slice = imax(1, (new_val + period / 2) / period);
207 	hogticks = imax(1, (2 * hz * sched_slice + realstathz / 2) /
208 	    realstathz);
209 	return (0);
210 }
211 
212 SYSCTL_NODE(_kern, OID_AUTO, sched, CTLFLAG_RD | CTLFLAG_MPSAFE, 0,
213     "Scheduler");
214 
215 SYSCTL_STRING(_kern_sched, OID_AUTO, name, CTLFLAG_RD, "4BSD", 0,
216     "Scheduler name");
217 SYSCTL_PROC(_kern_sched, OID_AUTO, quantum,
218     CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_MPSAFE, NULL, 0,
219     sysctl_kern_quantum, "I",
220     "Quantum for timeshare threads in microseconds");
221 SYSCTL_INT(_kern_sched, OID_AUTO, slice, CTLFLAG_RW, &sched_slice, 0,
222     "Quantum for timeshare threads in stathz ticks");
223 #ifdef SMP
224 /* Enable forwarding of wakeups to all other cpus */
225 static SYSCTL_NODE(_kern_sched, OID_AUTO, ipiwakeup,
226     CTLFLAG_RD | CTLFLAG_MPSAFE, NULL,
227     "Kernel SMP");
228 
229 static int runq_fuzz = 1;
230 SYSCTL_INT(_kern_sched, OID_AUTO, runq_fuzz, CTLFLAG_RW, &runq_fuzz, 0, "");
231 
232 static int forward_wakeup_enabled = 1;
233 SYSCTL_INT(_kern_sched_ipiwakeup, OID_AUTO, enabled, CTLFLAG_RW,
234 	   &forward_wakeup_enabled, 0,
235 	   "Forwarding of wakeup to idle CPUs");
236 
237 static int forward_wakeups_requested = 0;
238 SYSCTL_INT(_kern_sched_ipiwakeup, OID_AUTO, requested, CTLFLAG_RD,
239 	   &forward_wakeups_requested, 0,
240 	   "Requests for Forwarding of wakeup to idle CPUs");
241 
242 static int forward_wakeups_delivered = 0;
243 SYSCTL_INT(_kern_sched_ipiwakeup, OID_AUTO, delivered, CTLFLAG_RD,
244 	   &forward_wakeups_delivered, 0,
245 	   "Completed Forwarding of wakeup to idle CPUs");
246 
247 static int forward_wakeup_use_mask = 1;
248 SYSCTL_INT(_kern_sched_ipiwakeup, OID_AUTO, usemask, CTLFLAG_RW,
249 	   &forward_wakeup_use_mask, 0,
250 	   "Use the mask of idle cpus");
251 
252 static int forward_wakeup_use_loop = 0;
253 SYSCTL_INT(_kern_sched_ipiwakeup, OID_AUTO, useloop, CTLFLAG_RW,
254 	   &forward_wakeup_use_loop, 0,
255 	   "Use a loop to find idle cpus");
256 
257 #endif
258 #if 0
259 static int sched_followon = 0;
260 SYSCTL_INT(_kern_sched, OID_AUTO, followon, CTLFLAG_RW,
261 	   &sched_followon, 0,
262 	   "allow threads to share a quantum");
263 #endif
264 
265 SDT_PROVIDER_DEFINE(sched);
266 
267 SDT_PROBE_DEFINE3(sched, , , change__pri, "struct thread *",
268     "struct proc *", "uint8_t");
269 SDT_PROBE_DEFINE3(sched, , , dequeue, "struct thread *",
270     "struct proc *", "void *");
271 SDT_PROBE_DEFINE4(sched, , , enqueue, "struct thread *",
272     "struct proc *", "void *", "int");
273 SDT_PROBE_DEFINE4(sched, , , lend__pri, "struct thread *",
274     "struct proc *", "uint8_t", "struct thread *");
275 SDT_PROBE_DEFINE2(sched, , , load__change, "int", "int");
276 SDT_PROBE_DEFINE2(sched, , , off__cpu, "struct thread *",
277     "struct proc *");
278 SDT_PROBE_DEFINE(sched, , , on__cpu);
279 SDT_PROBE_DEFINE(sched, , , remain__cpu);
280 SDT_PROBE_DEFINE2(sched, , , surrender, "struct thread *",
281     "struct proc *");
282 
283 static __inline void
284 sched_load_add(void)
285 {
286 
287 	sched_tdcnt++;
288 	KTR_COUNTER0(KTR_SCHED, "load", "global load", sched_tdcnt);
289 	SDT_PROBE2(sched, , , load__change, NOCPU, sched_tdcnt);
290 }
291 
292 static __inline void
293 sched_load_rem(void)
294 {
295 
296 	sched_tdcnt--;
297 	KTR_COUNTER0(KTR_SCHED, "load", "global load", sched_tdcnt);
298 	SDT_PROBE2(sched, , , load__change, NOCPU, sched_tdcnt);
299 }
300 /*
301  * Arrange to reschedule if necessary, taking the priorities and
302  * schedulers into account.
303  */
304 static void
305 maybe_resched(struct thread *td)
306 {
307 
308 	THREAD_LOCK_ASSERT(td, MA_OWNED);
309 	if (td->td_priority < curthread->td_priority)
310 		curthread->td_flags |= TDF_NEEDRESCHED;
311 }
312 
313 /*
314  * This function is called when a thread is about to be put on run queue
315  * because it has been made runnable or its priority has been adjusted.  It
316  * determines if the new thread should preempt the current thread.  If so,
317  * it sets td_owepreempt to request a preemption.
318  */
319 int
320 maybe_preempt(struct thread *td)
321 {
322 #ifdef PREEMPTION
323 	struct thread *ctd;
324 	int cpri, pri;
325 
326 	/*
327 	 * The new thread should not preempt the current thread if any of the
328 	 * following conditions are true:
329 	 *
330 	 *  - The kernel is in the throes of crashing (panicstr).
331 	 *  - The current thread has a higher (numerically lower) or
332 	 *    equivalent priority.  Note that this prevents curthread from
333 	 *    trying to preempt to itself.
334 	 *  - The current thread has an inhibitor set or is in the process of
335 	 *    exiting.  In this case, the current thread is about to switch
336 	 *    out anyways, so there's no point in preempting.  If we did,
337 	 *    the current thread would not be properly resumed as well, so
338 	 *    just avoid that whole landmine.
339 	 *  - If the new thread's priority is not a realtime priority and
340 	 *    the current thread's priority is not an idle priority and
341 	 *    FULL_PREEMPTION is disabled.
342 	 *
343 	 * If all of these conditions are false, but the current thread is in
344 	 * a nested critical section, then we have to defer the preemption
345 	 * until we exit the critical section.  Otherwise, switch immediately
346 	 * to the new thread.
347 	 */
348 	ctd = curthread;
349 	THREAD_LOCK_ASSERT(td, MA_OWNED);
350 	KASSERT((td->td_inhibitors == 0),
351 			("maybe_preempt: trying to run inhibited thread"));
352 	pri = td->td_priority;
353 	cpri = ctd->td_priority;
354 	if (KERNEL_PANICKED() || pri >= cpri /* || dumping */ ||
355 	    TD_IS_INHIBITED(ctd))
356 		return (0);
357 #ifndef FULL_PREEMPTION
358 	if (pri > PRI_MAX_ITHD && cpri < PRI_MIN_IDLE)
359 		return (0);
360 #endif
361 
362 	CTR0(KTR_PROC, "maybe_preempt: scheduling preemption");
363 	ctd->td_owepreempt = 1;
364 	return (1);
365 #else
366 	return (0);
367 #endif
368 }
369 
370 /*
371  * Constants for digital decay and forget:
372  *	90% of (ts_estcpu) usage in 5 * loadav time
373  *	95% of (ts_pctcpu) usage in 60 seconds (load insensitive)
374  *          Note that, as ps(1) mentions, this can let percentages
375  *          total over 100% (I've seen 137.9% for 3 processes).
376  *
377  * Note that schedclock() updates ts_estcpu and p_cpticks asynchronously.
378  *
379  * We wish to decay away 90% of ts_estcpu in (5 * loadavg) seconds.
380  * That is, the system wants to compute a value of decay such
381  * that the following for loop:
382  * 	for (i = 0; i < (5 * loadavg); i++)
383  * 		ts_estcpu *= decay;
384  * will compute
385  * 	ts_estcpu *= 0.1;
386  * for all values of loadavg:
387  *
388  * Mathematically this loop can be expressed by saying:
389  * 	decay ** (5 * loadavg) ~= .1
390  *
391  * The system computes decay as:
392  * 	decay = (2 * loadavg) / (2 * loadavg + 1)
393  *
394  * We wish to prove that the system's computation of decay
395  * will always fulfill the equation:
396  * 	decay ** (5 * loadavg) ~= .1
397  *
398  * If we compute b as:
399  * 	b = 2 * loadavg
400  * then
401  * 	decay = b / (b + 1)
402  *
403  * We now need to prove two things:
404  *	1) Given factor ** (5 * loadavg) ~= .1, prove factor == b/(b+1)
405  *	2) Given b/(b+1) ** power ~= .1, prove power == (5 * loadavg)
406  *
407  * Facts:
408  *         For x close to zero, exp(x) =~ 1 + x, since
409  *              exp(x) = 0! + x**1/1! + x**2/2! + ... .
410  *              therefore exp(-1/b) =~ 1 - (1/b) = (b-1)/b.
411  *         For x close to zero, ln(1+x) =~ x, since
412  *              ln(1+x) = x - x**2/2 + x**3/3 - ...     -1 < x < 1
413  *              therefore ln(b/(b+1)) = ln(1 - 1/(b+1)) =~ -1/(b+1).
414  *         ln(.1) =~ -2.30
415  *
416  * Proof of (1):
417  *    Solve (factor)**(power) =~ .1 given power (5*loadav):
418  *	solving for factor,
419  *      ln(factor) =~ (-2.30/5*loadav), or
420  *      factor =~ exp(-1/((5/2.30)*loadav)) =~ exp(-1/(2*loadav)) =
421  *          exp(-1/b) =~ (b-1)/b =~ b/(b+1).                    QED
422  *
423  * Proof of (2):
424  *    Solve (factor)**(power) =~ .1 given factor == (b/(b+1)):
425  *	solving for power,
426  *      power*ln(b/(b+1)) =~ -2.30, or
427  *      power =~ 2.3 * (b + 1) = 4.6*loadav + 2.3 =~ 5*loadav.  QED
428  *
429  * Actual power values for the implemented algorithm are as follows:
430  *      loadav: 1       2       3       4
431  *      power:  5.68    10.32   14.94   19.55
432  */
433 
434 /* calculations for digital decay to forget 90% of usage in 5*loadav sec */
435 #define	loadfactor(loadav)	(2 * (loadav))
436 #define	decay_cpu(loadfac, cpu)	(((loadfac) * (cpu)) / ((loadfac) + FSCALE))
437 
438 /* decay 95% of `ts_pctcpu' in 60 seconds; see CCPU_SHIFT before changing */
439 static fixpt_t	ccpu = 0.95122942450071400909 * FSCALE;	/* exp(-1/20) */
440 SYSCTL_UINT(_kern, OID_AUTO, ccpu, CTLFLAG_RD, &ccpu, 0,
441     "Decay factor used for updating %CPU");
442 
443 /*
444  * If `ccpu' is not equal to `exp(-1/20)' and you still want to use the
445  * faster/more-accurate formula, you'll have to estimate CCPU_SHIFT below
446  * and possibly adjust FSHIFT in "param.h" so that (FSHIFT >= CCPU_SHIFT).
447  *
448  * To estimate CCPU_SHIFT for exp(-1/20), the following formula was used:
449  *	1 - exp(-1/20) ~= 0.0487 ~= 0.0488 == 1 (fixed pt, *11* bits).
450  *
451  * If you don't want to bother with the faster/more-accurate formula, you
452  * can set CCPU_SHIFT to (FSHIFT + 1) which will use a slower/less-accurate
453  * (more general) method of calculating the %age of CPU used by a process.
454  */
455 #define	CCPU_SHIFT	11
456 
457 /*
458  * Recompute process priorities, every hz ticks.
459  * MP-safe, called without the Giant mutex.
460  */
461 /* ARGSUSED */
462 static void
463 schedcpu(void)
464 {
465 	fixpt_t loadfac = loadfactor(averunnable.ldavg[0]);
466 	struct thread *td;
467 	struct proc *p;
468 	struct td_sched *ts;
469 	int awake;
470 
471 	sx_slock(&allproc_lock);
472 	FOREACH_PROC_IN_SYSTEM(p) {
473 		PROC_LOCK(p);
474 		if (p->p_state == PRS_NEW) {
475 			PROC_UNLOCK(p);
476 			continue;
477 		}
478 		FOREACH_THREAD_IN_PROC(p, td) {
479 			awake = 0;
480 			ts = td_get_sched(td);
481 			thread_lock(td);
482 			/*
483 			 * Increment sleep time (if sleeping).  We
484 			 * ignore overflow, as above.
485 			 */
486 			/*
487 			 * The td_sched slptimes are not touched in wakeup
488 			 * because the thread may not HAVE everything in
489 			 * memory? XXX I think this is out of date.
490 			 */
491 			if (TD_ON_RUNQ(td)) {
492 				awake = 1;
493 				td->td_flags &= ~TDF_DIDRUN;
494 			} else if (TD_IS_RUNNING(td)) {
495 				awake = 1;
496 				/* Do not clear TDF_DIDRUN */
497 			} else if (td->td_flags & TDF_DIDRUN) {
498 				awake = 1;
499 				td->td_flags &= ~TDF_DIDRUN;
500 			}
501 
502 			/*
503 			 * ts_pctcpu is only for ps and ttyinfo().
504 			 */
505 			ts->ts_pctcpu = (ts->ts_pctcpu * ccpu) >> FSHIFT;
506 			/*
507 			 * If the td_sched has been idle the entire second,
508 			 * stop recalculating its priority until
509 			 * it wakes up.
510 			 */
511 			if (ts->ts_cpticks != 0) {
512 #if	(FSHIFT >= CCPU_SHIFT)
513 				ts->ts_pctcpu += (realstathz == 100)
514 				    ? ((fixpt_t) ts->ts_cpticks) <<
515 				    (FSHIFT - CCPU_SHIFT) :
516 				    100 * (((fixpt_t) ts->ts_cpticks)
517 				    << (FSHIFT - CCPU_SHIFT)) / realstathz;
518 #else
519 				ts->ts_pctcpu += ((FSCALE - ccpu) *
520 				    (ts->ts_cpticks *
521 				    FSCALE / realstathz)) >> FSHIFT;
522 #endif
523 				ts->ts_cpticks = 0;
524 			}
525 			/*
526 			 * If there are ANY running threads in this process,
527 			 * then don't count it as sleeping.
528 			 * XXX: this is broken.
529 			 */
530 			if (awake) {
531 				if (ts->ts_slptime > 1) {
532 					/*
533 					 * In an ideal world, this should not
534 					 * happen, because whoever woke us
535 					 * up from the long sleep should have
536 					 * unwound the slptime and reset our
537 					 * priority before we run at the stale
538 					 * priority.  Should KASSERT at some
539 					 * point when all the cases are fixed.
540 					 */
541 					updatepri(td);
542 				}
543 				ts->ts_slptime = 0;
544 			} else
545 				ts->ts_slptime++;
546 			if (ts->ts_slptime > 1) {
547 				thread_unlock(td);
548 				continue;
549 			}
550 			ts->ts_estcpu = decay_cpu(loadfac, ts->ts_estcpu);
551 		      	resetpriority(td);
552 			resetpriority_thread(td);
553 			thread_unlock(td);
554 		}
555 		PROC_UNLOCK(p);
556 	}
557 	sx_sunlock(&allproc_lock);
558 }
559 
560 /*
561  * Main loop for a kthread that executes schedcpu once a second.
562  */
563 static void
564 schedcpu_thread(void)
565 {
566 
567 	for (;;) {
568 		schedcpu();
569 		pause("-", hz);
570 	}
571 }
572 
573 /*
574  * Recalculate the priority of a process after it has slept for a while.
575  * For all load averages >= 1 and max ts_estcpu of 255, sleeping for at
576  * least six times the loadfactor will decay ts_estcpu to zero.
577  */
578 static void
579 updatepri(struct thread *td)
580 {
581 	struct td_sched *ts;
582 	fixpt_t loadfac;
583 	unsigned int newcpu;
584 
585 	ts = td_get_sched(td);
586 	loadfac = loadfactor(averunnable.ldavg[0]);
587 	if (ts->ts_slptime > 5 * loadfac)
588 		ts->ts_estcpu = 0;
589 	else {
590 		newcpu = ts->ts_estcpu;
591 		ts->ts_slptime--;	/* was incremented in schedcpu() */
592 		while (newcpu && --ts->ts_slptime)
593 			newcpu = decay_cpu(loadfac, newcpu);
594 		ts->ts_estcpu = newcpu;
595 	}
596 }
597 
598 /*
599  * Compute the priority of a process when running in user mode.
600  * Arrange to reschedule if the resulting priority is better
601  * than that of the current process.
602  */
603 static void
604 resetpriority(struct thread *td)
605 {
606 	u_int newpriority;
607 
608 	if (td->td_pri_class != PRI_TIMESHARE)
609 		return;
610 	newpriority = PUSER +
611 	    td_get_sched(td)->ts_estcpu / INVERSE_ESTCPU_WEIGHT +
612 	    NICE_WEIGHT * (td->td_proc->p_nice - PRIO_MIN);
613 	newpriority = min(max(newpriority, PRI_MIN_TIMESHARE),
614 	    PRI_MAX_TIMESHARE);
615 	sched_user_prio(td, newpriority);
616 }
617 
618 /*
619  * Update the thread's priority when the associated process's user
620  * priority changes.
621  */
622 static void
623 resetpriority_thread(struct thread *td)
624 {
625 
626 	/* Only change threads with a time sharing user priority. */
627 	if (td->td_priority < PRI_MIN_TIMESHARE ||
628 	    td->td_priority > PRI_MAX_TIMESHARE)
629 		return;
630 
631 	/* XXX the whole needresched thing is broken, but not silly. */
632 	maybe_resched(td);
633 
634 	sched_prio(td, td->td_user_pri);
635 }
636 
637 /* ARGSUSED */
638 static void
639 sched_setup(void *dummy)
640 {
641 
642 	setup_runqs();
643 
644 	/* Account for thread0. */
645 	sched_load_add();
646 }
647 
648 /*
649  * This routine determines time constants after stathz and hz are setup.
650  */
651 static void
652 sched_initticks(void *dummy)
653 {
654 
655 	realstathz = stathz ? stathz : hz;
656 	sched_slice = realstathz / 10;	/* ~100ms */
657 	hogticks = imax(1, (2 * hz * sched_slice + realstathz / 2) /
658 	    realstathz);
659 }
660 
661 /* External interfaces start here */
662 
663 /*
664  * Very early in the boot some setup of scheduler-specific
665  * parts of proc0 and of some scheduler resources needs to be done.
666  * Called from:
667  *  proc0_init()
668  */
669 void
670 schedinit(void)
671 {
672 
673 	/*
674 	 * Set up the scheduler specific parts of thread0.
675 	 */
676 	thread0.td_lock = &sched_lock;
677 	td_get_sched(&thread0)->ts_slice = sched_slice;
678 	mtx_init(&sched_lock, "sched lock", NULL, MTX_SPIN);
679 }
680 
681 void
682 schedinit_ap(void)
683 {
684 
685 	/* Nothing needed. */
686 }
687 
688 int
689 sched_runnable(void)
690 {
691 #ifdef SMP
692 	return runq_check(&runq) + runq_check(&runq_pcpu[PCPU_GET(cpuid)]);
693 #else
694 	return runq_check(&runq);
695 #endif
696 }
697 
698 int
699 sched_rr_interval(void)
700 {
701 
702 	/* Convert sched_slice from stathz to hz. */
703 	return (imax(1, (sched_slice * hz + realstathz / 2) / realstathz));
704 }
705 
706 /*
707  * We adjust the priority of the current process.  The priority of a
708  * process gets worse as it accumulates CPU time.  The cpu usage
709  * estimator (ts_estcpu) is increased here.  resetpriority() will
710  * compute a different priority each time ts_estcpu increases by
711  * INVERSE_ESTCPU_WEIGHT (until PRI_MAX_TIMESHARE is reached).  The
712  * cpu usage estimator ramps up quite quickly when the process is
713  * running (linearly), and decays away exponentially, at a rate which
714  * is proportionally slower when the system is busy.  The basic
715  * principle is that the system will 90% forget that the process used
716  * a lot of CPU time in 5 * loadav seconds.  This causes the system to
717  * favor processes which haven't run much recently, and to round-robin
718  * among other processes.
719  */
720 static void
721 sched_clock_tick(struct thread *td)
722 {
723 	struct pcpuidlestat *stat;
724 	struct td_sched *ts;
725 
726 	THREAD_LOCK_ASSERT(td, MA_OWNED);
727 	ts = td_get_sched(td);
728 
729 	ts->ts_cpticks++;
730 	ts->ts_estcpu = ESTCPULIM(ts->ts_estcpu + 1);
731 	if ((ts->ts_estcpu % INVERSE_ESTCPU_WEIGHT) == 0) {
732 		resetpriority(td);
733 		resetpriority_thread(td);
734 	}
735 
736 	/*
737 	 * Force a context switch if the current thread has used up a full
738 	 * time slice (default is 100ms).
739 	 */
740 	if (!TD_IS_IDLETHREAD(td) && --ts->ts_slice <= 0) {
741 		ts->ts_slice = sched_slice;
742 		td->td_flags |= TDF_NEEDRESCHED | TDF_SLICEEND;
743 	}
744 
745 	stat = DPCPU_PTR(idlestat);
746 	stat->oldidlecalls = stat->idlecalls;
747 	stat->idlecalls = 0;
748 }
749 
750 void
751 sched_clock(struct thread *td, int cnt)
752 {
753 
754 	for ( ; cnt > 0; cnt--)
755 		sched_clock_tick(td);
756 }
757 
758 /*
759  * Charge child's scheduling CPU usage to parent.
760  */
761 void
762 sched_exit(struct proc *p, struct thread *td)
763 {
764 
765 	KTR_STATE1(KTR_SCHED, "thread", sched_tdname(td), "proc exit",
766 	    "prio:%d", td->td_priority);
767 
768 	PROC_LOCK_ASSERT(p, MA_OWNED);
769 	sched_exit_thread(FIRST_THREAD_IN_PROC(p), td);
770 }
771 
772 void
773 sched_exit_thread(struct thread *td, struct thread *child)
774 {
775 
776 	KTR_STATE1(KTR_SCHED, "thread", sched_tdname(child), "exit",
777 	    "prio:%d", child->td_priority);
778 	thread_lock(td);
779 	td_get_sched(td)->ts_estcpu = ESTCPULIM(td_get_sched(td)->ts_estcpu +
780 	    td_get_sched(child)->ts_estcpu);
781 	thread_unlock(td);
782 	thread_lock(child);
783 	if ((child->td_flags & TDF_NOLOAD) == 0)
784 		sched_load_rem();
785 	thread_unlock(child);
786 }
787 
788 void
789 sched_fork(struct thread *td, struct thread *childtd)
790 {
791 	sched_fork_thread(td, childtd);
792 }
793 
794 void
795 sched_fork_thread(struct thread *td, struct thread *childtd)
796 {
797 	struct td_sched *ts, *tsc;
798 
799 	childtd->td_oncpu = NOCPU;
800 	childtd->td_lastcpu = NOCPU;
801 	childtd->td_lock = &sched_lock;
802 	childtd->td_cpuset = cpuset_ref(td->td_cpuset);
803 	childtd->td_domain.dr_policy = td->td_cpuset->cs_domain;
804 	childtd->td_priority = childtd->td_base_pri;
805 	ts = td_get_sched(childtd);
806 	bzero(ts, sizeof(*ts));
807 	tsc = td_get_sched(td);
808 	ts->ts_estcpu = tsc->ts_estcpu;
809 	ts->ts_flags |= (tsc->ts_flags & TSF_AFFINITY);
810 	ts->ts_slice = 1;
811 }
812 
813 void
814 sched_nice(struct proc *p, int nice)
815 {
816 	struct thread *td;
817 
818 	PROC_LOCK_ASSERT(p, MA_OWNED);
819 	p->p_nice = nice;
820 	FOREACH_THREAD_IN_PROC(p, td) {
821 		thread_lock(td);
822 		resetpriority(td);
823 		resetpriority_thread(td);
824 		thread_unlock(td);
825 	}
826 }
827 
828 void
829 sched_class(struct thread *td, int class)
830 {
831 	THREAD_LOCK_ASSERT(td, MA_OWNED);
832 	td->td_pri_class = class;
833 }
834 
835 /*
836  * Adjust the priority of a thread.
837  */
838 static void
839 sched_priority(struct thread *td, u_char prio)
840 {
841 
842 	KTR_POINT3(KTR_SCHED, "thread", sched_tdname(td), "priority change",
843 	    "prio:%d", td->td_priority, "new prio:%d", prio, KTR_ATTR_LINKED,
844 	    sched_tdname(curthread));
845 	SDT_PROBE3(sched, , , change__pri, td, td->td_proc, prio);
846 	if (td != curthread && prio > td->td_priority) {
847 		KTR_POINT3(KTR_SCHED, "thread", sched_tdname(curthread),
848 		    "lend prio", "prio:%d", td->td_priority, "new prio:%d",
849 		    prio, KTR_ATTR_LINKED, sched_tdname(td));
850 		SDT_PROBE4(sched, , , lend__pri, td, td->td_proc, prio,
851 		    curthread);
852 	}
853 	THREAD_LOCK_ASSERT(td, MA_OWNED);
854 	if (td->td_priority == prio)
855 		return;
856 	td->td_priority = prio;
857 	if (TD_ON_RUNQ(td) && td->td_rqindex != (prio / RQ_PPQ)) {
858 		sched_rem(td);
859 		sched_add(td, SRQ_BORING | SRQ_HOLDTD);
860 	}
861 }
862 
863 /*
864  * Update a thread's priority when it is lent another thread's
865  * priority.
866  */
867 void
868 sched_lend_prio(struct thread *td, u_char prio)
869 {
870 
871 	td->td_flags |= TDF_BORROWING;
872 	sched_priority(td, prio);
873 }
874 
875 /*
876  * Restore a thread's priority when priority propagation is
877  * over.  The prio argument is the minimum priority the thread
878  * needs to have to satisfy other possible priority lending
879  * requests.  If the thread's regulary priority is less
880  * important than prio the thread will keep a priority boost
881  * of prio.
882  */
883 void
884 sched_unlend_prio(struct thread *td, u_char prio)
885 {
886 	u_char base_pri;
887 
888 	if (td->td_base_pri >= PRI_MIN_TIMESHARE &&
889 	    td->td_base_pri <= PRI_MAX_TIMESHARE)
890 		base_pri = td->td_user_pri;
891 	else
892 		base_pri = td->td_base_pri;
893 	if (prio >= base_pri) {
894 		td->td_flags &= ~TDF_BORROWING;
895 		sched_prio(td, base_pri);
896 	} else
897 		sched_lend_prio(td, prio);
898 }
899 
900 void
901 sched_prio(struct thread *td, u_char prio)
902 {
903 	u_char oldprio;
904 
905 	/* First, update the base priority. */
906 	td->td_base_pri = prio;
907 
908 	/*
909 	 * If the thread is borrowing another thread's priority, don't ever
910 	 * lower the priority.
911 	 */
912 	if (td->td_flags & TDF_BORROWING && td->td_priority < prio)
913 		return;
914 
915 	/* Change the real priority. */
916 	oldprio = td->td_priority;
917 	sched_priority(td, prio);
918 
919 	/*
920 	 * If the thread is on a turnstile, then let the turnstile update
921 	 * its state.
922 	 */
923 	if (TD_ON_LOCK(td) && oldprio != prio)
924 		turnstile_adjust(td, oldprio);
925 }
926 
927 void
928 sched_user_prio(struct thread *td, u_char prio)
929 {
930 
931 	THREAD_LOCK_ASSERT(td, MA_OWNED);
932 	td->td_base_user_pri = prio;
933 	if (td->td_lend_user_pri <= prio)
934 		return;
935 	td->td_user_pri = prio;
936 }
937 
938 void
939 sched_lend_user_prio(struct thread *td, u_char prio)
940 {
941 
942 	THREAD_LOCK_ASSERT(td, MA_OWNED);
943 	td->td_lend_user_pri = prio;
944 	td->td_user_pri = min(prio, td->td_base_user_pri);
945 	if (td->td_priority > td->td_user_pri)
946 		sched_prio(td, td->td_user_pri);
947 	else if (td->td_priority != td->td_user_pri)
948 		td->td_flags |= TDF_NEEDRESCHED;
949 }
950 
951 /*
952  * Like the above but first check if there is anything to do.
953  */
954 void
955 sched_lend_user_prio_cond(struct thread *td, u_char prio)
956 {
957 
958 	if (td->td_lend_user_pri != prio)
959 		goto lend;
960 	if (td->td_user_pri != min(prio, td->td_base_user_pri))
961 		goto lend;
962 	if (td->td_priority != td->td_user_pri)
963 		goto lend;
964 	return;
965 
966 lend:
967 	thread_lock(td);
968 	sched_lend_user_prio(td, prio);
969 	thread_unlock(td);
970 }
971 
972 void
973 sched_sleep(struct thread *td, int pri)
974 {
975 
976 	THREAD_LOCK_ASSERT(td, MA_OWNED);
977 	td->td_slptick = ticks;
978 	td_get_sched(td)->ts_slptime = 0;
979 	if (pri != 0 && PRI_BASE(td->td_pri_class) == PRI_TIMESHARE)
980 		sched_prio(td, pri);
981 	if (TD_IS_SUSPENDED(td) || pri >= PSOCK)
982 		td->td_flags |= TDF_CANSWAP;
983 }
984 
985 void
986 sched_switch(struct thread *td, int flags)
987 {
988 	struct thread *newtd;
989 	struct mtx *tmtx;
990 	int preempted;
991 
992 	tmtx = &sched_lock;
993 
994 	THREAD_LOCK_ASSERT(td, MA_OWNED);
995 
996 	td->td_lastcpu = td->td_oncpu;
997 	preempted = (td->td_flags & TDF_SLICEEND) == 0 &&
998 	    (flags & SW_PREEMPT) != 0;
999 	td->td_flags &= ~(TDF_NEEDRESCHED | TDF_SLICEEND);
1000 	td->td_owepreempt = 0;
1001 	td->td_oncpu = NOCPU;
1002 
1003 	/*
1004 	 * At the last moment, if this thread is still marked RUNNING,
1005 	 * then put it back on the run queue as it has not been suspended
1006 	 * or stopped or any thing else similar.  We never put the idle
1007 	 * threads on the run queue, however.
1008 	 */
1009 	if (td->td_flags & TDF_IDLETD) {
1010 		TD_SET_CAN_RUN(td);
1011 #ifdef SMP
1012 		CPU_CLR(PCPU_GET(cpuid), &idle_cpus_mask);
1013 #endif
1014 	} else {
1015 		if (TD_IS_RUNNING(td)) {
1016 			/* Put us back on the run queue. */
1017 			sched_add(td, preempted ?
1018 			    SRQ_HOLDTD|SRQ_OURSELF|SRQ_YIELDING|SRQ_PREEMPTED :
1019 			    SRQ_HOLDTD|SRQ_OURSELF|SRQ_YIELDING);
1020 		}
1021 	}
1022 
1023 	/*
1024 	 * Switch to the sched lock to fix things up and pick
1025 	 * a new thread.  Block the td_lock in order to avoid
1026 	 * breaking the critical path.
1027 	 */
1028 	if (td->td_lock != &sched_lock) {
1029 		mtx_lock_spin(&sched_lock);
1030 		tmtx = thread_lock_block(td);
1031 		mtx_unlock_spin(tmtx);
1032 	}
1033 
1034 	if ((td->td_flags & TDF_NOLOAD) == 0)
1035 		sched_load_rem();
1036 
1037 	newtd = choosethread();
1038 	MPASS(newtd->td_lock == &sched_lock);
1039 
1040 #if (KTR_COMPILE & KTR_SCHED) != 0
1041 	if (TD_IS_IDLETHREAD(td))
1042 		KTR_STATE1(KTR_SCHED, "thread", sched_tdname(td), "idle",
1043 		    "prio:%d", td->td_priority);
1044 	else
1045 		KTR_STATE3(KTR_SCHED, "thread", sched_tdname(td), KTDSTATE(td),
1046 		    "prio:%d", td->td_priority, "wmesg:\"%s\"", td->td_wmesg,
1047 		    "lockname:\"%s\"", td->td_lockname);
1048 #endif
1049 
1050 	if (td != newtd) {
1051 #ifdef	HWPMC_HOOKS
1052 		if (PMC_PROC_IS_USING_PMCS(td->td_proc))
1053 			PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_OUT);
1054 #endif
1055 
1056 		SDT_PROBE2(sched, , , off__cpu, newtd, newtd->td_proc);
1057 
1058                 /* I feel sleepy */
1059 		lock_profile_release_lock(&sched_lock.lock_object, true);
1060 #ifdef KDTRACE_HOOKS
1061 		/*
1062 		 * If DTrace has set the active vtime enum to anything
1063 		 * other than INACTIVE (0), then it should have set the
1064 		 * function to call.
1065 		 */
1066 		if (dtrace_vtime_active)
1067 			(*dtrace_vtime_switch_func)(newtd);
1068 #endif
1069 
1070 		cpu_switch(td, newtd, tmtx);
1071 		lock_profile_obtain_lock_success(&sched_lock.lock_object, true,
1072 		    0, 0, __FILE__, __LINE__);
1073 		/*
1074 		 * Where am I?  What year is it?
1075 		 * We are in the same thread that went to sleep above,
1076 		 * but any amount of time may have passed. All our context
1077 		 * will still be available as will local variables.
1078 		 * PCPU values however may have changed as we may have
1079 		 * changed CPU so don't trust cached values of them.
1080 		 * New threads will go to fork_exit() instead of here
1081 		 * so if you change things here you may need to change
1082 		 * things there too.
1083 		 *
1084 		 * If the thread above was exiting it will never wake
1085 		 * up again here, so either it has saved everything it
1086 		 * needed to, or the thread_wait() or wait() will
1087 		 * need to reap it.
1088 		 */
1089 
1090 		SDT_PROBE0(sched, , , on__cpu);
1091 #ifdef	HWPMC_HOOKS
1092 		if (PMC_PROC_IS_USING_PMCS(td->td_proc))
1093 			PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_IN);
1094 #endif
1095 	} else {
1096 		td->td_lock = &sched_lock;
1097 		SDT_PROBE0(sched, , , remain__cpu);
1098 	}
1099 
1100 	KTR_STATE1(KTR_SCHED, "thread", sched_tdname(td), "running",
1101 	    "prio:%d", td->td_priority);
1102 
1103 #ifdef SMP
1104 	if (td->td_flags & TDF_IDLETD)
1105 		CPU_SET(PCPU_GET(cpuid), &idle_cpus_mask);
1106 #endif
1107 	sched_lock.mtx_lock = (uintptr_t)td;
1108 	td->td_oncpu = PCPU_GET(cpuid);
1109 	spinlock_enter();
1110 	mtx_unlock_spin(&sched_lock);
1111 }
1112 
1113 void
1114 sched_wakeup(struct thread *td, int srqflags)
1115 {
1116 	struct td_sched *ts;
1117 
1118 	THREAD_LOCK_ASSERT(td, MA_OWNED);
1119 	ts = td_get_sched(td);
1120 	td->td_flags &= ~TDF_CANSWAP;
1121 	if (ts->ts_slptime > 1) {
1122 		updatepri(td);
1123 		resetpriority(td);
1124 	}
1125 	td->td_slptick = 0;
1126 	ts->ts_slptime = 0;
1127 	ts->ts_slice = sched_slice;
1128 	sched_add(td, srqflags);
1129 }
1130 
1131 #ifdef SMP
1132 static int
1133 forward_wakeup(int cpunum)
1134 {
1135 	struct pcpu *pc;
1136 	cpuset_t dontuse, map, map2;
1137 	u_int id, me;
1138 	int iscpuset;
1139 
1140 	mtx_assert(&sched_lock, MA_OWNED);
1141 
1142 	CTR0(KTR_RUNQ, "forward_wakeup()");
1143 
1144 	if ((!forward_wakeup_enabled) ||
1145 	     (forward_wakeup_use_mask == 0 && forward_wakeup_use_loop == 0))
1146 		return (0);
1147 	if (!smp_started || KERNEL_PANICKED())
1148 		return (0);
1149 
1150 	forward_wakeups_requested++;
1151 
1152 	/*
1153 	 * Check the idle mask we received against what we calculated
1154 	 * before in the old version.
1155 	 */
1156 	me = PCPU_GET(cpuid);
1157 
1158 	/* Don't bother if we should be doing it ourself. */
1159 	if (CPU_ISSET(me, &idle_cpus_mask) &&
1160 	    (cpunum == NOCPU || me == cpunum))
1161 		return (0);
1162 
1163 	CPU_SETOF(me, &dontuse);
1164 	CPU_OR(&dontuse, &dontuse, &stopped_cpus);
1165 	CPU_OR(&dontuse, &dontuse, &hlt_cpus_mask);
1166 	CPU_ZERO(&map2);
1167 	if (forward_wakeup_use_loop) {
1168 		STAILQ_FOREACH(pc, &cpuhead, pc_allcpu) {
1169 			id = pc->pc_cpuid;
1170 			if (!CPU_ISSET(id, &dontuse) &&
1171 			    pc->pc_curthread == pc->pc_idlethread) {
1172 				CPU_SET(id, &map2);
1173 			}
1174 		}
1175 	}
1176 
1177 	if (forward_wakeup_use_mask) {
1178 		map = idle_cpus_mask;
1179 		CPU_ANDNOT(&map, &map, &dontuse);
1180 
1181 		/* If they are both on, compare and use loop if different. */
1182 		if (forward_wakeup_use_loop) {
1183 			if (CPU_CMP(&map, &map2)) {
1184 				printf("map != map2, loop method preferred\n");
1185 				map = map2;
1186 			}
1187 		}
1188 	} else {
1189 		map = map2;
1190 	}
1191 
1192 	/* If we only allow a specific CPU, then mask off all the others. */
1193 	if (cpunum != NOCPU) {
1194 		KASSERT((cpunum <= mp_maxcpus),("forward_wakeup: bad cpunum."));
1195 		iscpuset = CPU_ISSET(cpunum, &map);
1196 		if (iscpuset == 0)
1197 			CPU_ZERO(&map);
1198 		else
1199 			CPU_SETOF(cpunum, &map);
1200 	}
1201 	if (!CPU_EMPTY(&map)) {
1202 		forward_wakeups_delivered++;
1203 		STAILQ_FOREACH(pc, &cpuhead, pc_allcpu) {
1204 			id = pc->pc_cpuid;
1205 			if (!CPU_ISSET(id, &map))
1206 				continue;
1207 			if (cpu_idle_wakeup(pc->pc_cpuid))
1208 				CPU_CLR(id, &map);
1209 		}
1210 		if (!CPU_EMPTY(&map))
1211 			ipi_selected(map, IPI_AST);
1212 		return (1);
1213 	}
1214 	if (cpunum == NOCPU)
1215 		printf("forward_wakeup: Idle processor not found\n");
1216 	return (0);
1217 }
1218 
1219 static void
1220 kick_other_cpu(int pri, int cpuid)
1221 {
1222 	struct pcpu *pcpu;
1223 	int cpri;
1224 
1225 	pcpu = pcpu_find(cpuid);
1226 	if (CPU_ISSET(cpuid, &idle_cpus_mask)) {
1227 		forward_wakeups_delivered++;
1228 		if (!cpu_idle_wakeup(cpuid))
1229 			ipi_cpu(cpuid, IPI_AST);
1230 		return;
1231 	}
1232 
1233 	cpri = pcpu->pc_curthread->td_priority;
1234 	if (pri >= cpri)
1235 		return;
1236 
1237 #if defined(IPI_PREEMPTION) && defined(PREEMPTION)
1238 #if !defined(FULL_PREEMPTION)
1239 	if (pri <= PRI_MAX_ITHD)
1240 #endif /* ! FULL_PREEMPTION */
1241 	{
1242 		ipi_cpu(cpuid, IPI_PREEMPT);
1243 		return;
1244 	}
1245 #endif /* defined(IPI_PREEMPTION) && defined(PREEMPTION) */
1246 
1247 	pcpu->pc_curthread->td_flags |= TDF_NEEDRESCHED;
1248 	ipi_cpu(cpuid, IPI_AST);
1249 	return;
1250 }
1251 #endif /* SMP */
1252 
1253 #ifdef SMP
1254 static int
1255 sched_pickcpu(struct thread *td)
1256 {
1257 	int best, cpu;
1258 
1259 	mtx_assert(&sched_lock, MA_OWNED);
1260 
1261 	if (td->td_lastcpu != NOCPU && THREAD_CAN_SCHED(td, td->td_lastcpu))
1262 		best = td->td_lastcpu;
1263 	else
1264 		best = NOCPU;
1265 	CPU_FOREACH(cpu) {
1266 		if (!THREAD_CAN_SCHED(td, cpu))
1267 			continue;
1268 
1269 		if (best == NOCPU)
1270 			best = cpu;
1271 		else if (runq_length[cpu] < runq_length[best])
1272 			best = cpu;
1273 	}
1274 	KASSERT(best != NOCPU, ("no valid CPUs"));
1275 
1276 	return (best);
1277 }
1278 #endif
1279 
1280 void
1281 sched_add(struct thread *td, int flags)
1282 #ifdef SMP
1283 {
1284 	cpuset_t tidlemsk;
1285 	struct td_sched *ts;
1286 	u_int cpu, cpuid;
1287 	int forwarded = 0;
1288 	int single_cpu = 0;
1289 
1290 	ts = td_get_sched(td);
1291 	THREAD_LOCK_ASSERT(td, MA_OWNED);
1292 	KASSERT((td->td_inhibitors == 0),
1293 	    ("sched_add: trying to run inhibited thread"));
1294 	KASSERT((TD_CAN_RUN(td) || TD_IS_RUNNING(td)),
1295 	    ("sched_add: bad thread state"));
1296 	KASSERT(td->td_flags & TDF_INMEM,
1297 	    ("sched_add: thread swapped out"));
1298 
1299 	KTR_STATE2(KTR_SCHED, "thread", sched_tdname(td), "runq add",
1300 	    "prio:%d", td->td_priority, KTR_ATTR_LINKED,
1301 	    sched_tdname(curthread));
1302 	KTR_POINT1(KTR_SCHED, "thread", sched_tdname(curthread), "wokeup",
1303 	    KTR_ATTR_LINKED, sched_tdname(td));
1304 	SDT_PROBE4(sched, , , enqueue, td, td->td_proc, NULL,
1305 	    flags & SRQ_PREEMPTED);
1306 
1307 	/*
1308 	 * Now that the thread is moving to the run-queue, set the lock
1309 	 * to the scheduler's lock.
1310 	 */
1311 	if (td->td_lock != &sched_lock) {
1312 		mtx_lock_spin(&sched_lock);
1313 		if ((flags & SRQ_HOLD) != 0)
1314 			td->td_lock = &sched_lock;
1315 		else
1316 			thread_lock_set(td, &sched_lock);
1317 	}
1318 	TD_SET_RUNQ(td);
1319 
1320 	/*
1321 	 * If SMP is started and the thread is pinned or otherwise limited to
1322 	 * a specific set of CPUs, queue the thread to a per-CPU run queue.
1323 	 * Otherwise, queue the thread to the global run queue.
1324 	 *
1325 	 * If SMP has not yet been started we must use the global run queue
1326 	 * as per-CPU state may not be initialized yet and we may crash if we
1327 	 * try to access the per-CPU run queues.
1328 	 */
1329 	if (smp_started && (td->td_pinned != 0 || td->td_flags & TDF_BOUND ||
1330 	    ts->ts_flags & TSF_AFFINITY)) {
1331 		if (td->td_pinned != 0)
1332 			cpu = td->td_lastcpu;
1333 		else if (td->td_flags & TDF_BOUND) {
1334 			/* Find CPU from bound runq. */
1335 			KASSERT(SKE_RUNQ_PCPU(ts),
1336 			    ("sched_add: bound td_sched not on cpu runq"));
1337 			cpu = ts->ts_runq - &runq_pcpu[0];
1338 		} else
1339 			/* Find a valid CPU for our cpuset */
1340 			cpu = sched_pickcpu(td);
1341 		ts->ts_runq = &runq_pcpu[cpu];
1342 		single_cpu = 1;
1343 		CTR3(KTR_RUNQ,
1344 		    "sched_add: Put td_sched:%p(td:%p) on cpu%d runq", ts, td,
1345 		    cpu);
1346 	} else {
1347 		CTR2(KTR_RUNQ,
1348 		    "sched_add: adding td_sched:%p (td:%p) to gbl runq", ts,
1349 		    td);
1350 		cpu = NOCPU;
1351 		ts->ts_runq = &runq;
1352 	}
1353 
1354 	if ((td->td_flags & TDF_NOLOAD) == 0)
1355 		sched_load_add();
1356 	runq_add(ts->ts_runq, td, flags);
1357 	if (cpu != NOCPU)
1358 		runq_length[cpu]++;
1359 
1360 	cpuid = PCPU_GET(cpuid);
1361 	if (single_cpu && cpu != cpuid) {
1362 	        kick_other_cpu(td->td_priority, cpu);
1363 	} else {
1364 		if (!single_cpu) {
1365 			tidlemsk = idle_cpus_mask;
1366 			CPU_ANDNOT(&tidlemsk, &tidlemsk, &hlt_cpus_mask);
1367 			CPU_CLR(cpuid, &tidlemsk);
1368 
1369 			if (!CPU_ISSET(cpuid, &idle_cpus_mask) &&
1370 			    ((flags & SRQ_INTR) == 0) &&
1371 			    !CPU_EMPTY(&tidlemsk))
1372 				forwarded = forward_wakeup(cpu);
1373 		}
1374 
1375 		if (!forwarded) {
1376 			if (!maybe_preempt(td))
1377 				maybe_resched(td);
1378 		}
1379 	}
1380 	if ((flags & SRQ_HOLDTD) == 0)
1381 		thread_unlock(td);
1382 }
1383 #else /* SMP */
1384 {
1385 	struct td_sched *ts;
1386 
1387 	ts = td_get_sched(td);
1388 	THREAD_LOCK_ASSERT(td, MA_OWNED);
1389 	KASSERT((td->td_inhibitors == 0),
1390 	    ("sched_add: trying to run inhibited thread"));
1391 	KASSERT((TD_CAN_RUN(td) || TD_IS_RUNNING(td)),
1392 	    ("sched_add: bad thread state"));
1393 	KASSERT(td->td_flags & TDF_INMEM,
1394 	    ("sched_add: thread swapped out"));
1395 	KTR_STATE2(KTR_SCHED, "thread", sched_tdname(td), "runq add",
1396 	    "prio:%d", td->td_priority, KTR_ATTR_LINKED,
1397 	    sched_tdname(curthread));
1398 	KTR_POINT1(KTR_SCHED, "thread", sched_tdname(curthread), "wokeup",
1399 	    KTR_ATTR_LINKED, sched_tdname(td));
1400 	SDT_PROBE4(sched, , , enqueue, td, td->td_proc, NULL,
1401 	    flags & SRQ_PREEMPTED);
1402 
1403 	/*
1404 	 * Now that the thread is moving to the run-queue, set the lock
1405 	 * to the scheduler's lock.
1406 	 */
1407 	if (td->td_lock != &sched_lock) {
1408 		mtx_lock_spin(&sched_lock);
1409 		if ((flags & SRQ_HOLD) != 0)
1410 			td->td_lock = &sched_lock;
1411 		else
1412 			thread_lock_set(td, &sched_lock);
1413 	}
1414 	TD_SET_RUNQ(td);
1415 	CTR2(KTR_RUNQ, "sched_add: adding td_sched:%p (td:%p) to runq", ts, td);
1416 	ts->ts_runq = &runq;
1417 
1418 	if ((td->td_flags & TDF_NOLOAD) == 0)
1419 		sched_load_add();
1420 	runq_add(ts->ts_runq, td, flags);
1421 	if (!maybe_preempt(td))
1422 		maybe_resched(td);
1423 	if ((flags & SRQ_HOLDTD) == 0)
1424 		thread_unlock(td);
1425 }
1426 #endif /* SMP */
1427 
1428 void
1429 sched_rem(struct thread *td)
1430 {
1431 	struct td_sched *ts;
1432 
1433 	ts = td_get_sched(td);
1434 	KASSERT(td->td_flags & TDF_INMEM,
1435 	    ("sched_rem: thread swapped out"));
1436 	KASSERT(TD_ON_RUNQ(td),
1437 	    ("sched_rem: thread not on run queue"));
1438 	mtx_assert(&sched_lock, MA_OWNED);
1439 	KTR_STATE2(KTR_SCHED, "thread", sched_tdname(td), "runq rem",
1440 	    "prio:%d", td->td_priority, KTR_ATTR_LINKED,
1441 	    sched_tdname(curthread));
1442 	SDT_PROBE3(sched, , , dequeue, td, td->td_proc, NULL);
1443 
1444 	if ((td->td_flags & TDF_NOLOAD) == 0)
1445 		sched_load_rem();
1446 #ifdef SMP
1447 	if (ts->ts_runq != &runq)
1448 		runq_length[ts->ts_runq - runq_pcpu]--;
1449 #endif
1450 	runq_remove(ts->ts_runq, td);
1451 	TD_SET_CAN_RUN(td);
1452 }
1453 
1454 /*
1455  * Select threads to run.  Note that running threads still consume a
1456  * slot.
1457  */
1458 struct thread *
1459 sched_choose(void)
1460 {
1461 	struct thread *td;
1462 	struct runq *rq;
1463 
1464 	mtx_assert(&sched_lock,  MA_OWNED);
1465 #ifdef SMP
1466 	struct thread *tdcpu;
1467 
1468 	rq = &runq;
1469 	td = runq_choose_fuzz(&runq, runq_fuzz);
1470 	tdcpu = runq_choose(&runq_pcpu[PCPU_GET(cpuid)]);
1471 
1472 	if (td == NULL ||
1473 	    (tdcpu != NULL &&
1474 	     tdcpu->td_priority < td->td_priority)) {
1475 		CTR2(KTR_RUNQ, "choosing td %p from pcpu runq %d", tdcpu,
1476 		     PCPU_GET(cpuid));
1477 		td = tdcpu;
1478 		rq = &runq_pcpu[PCPU_GET(cpuid)];
1479 	} else {
1480 		CTR1(KTR_RUNQ, "choosing td_sched %p from main runq", td);
1481 	}
1482 
1483 #else
1484 	rq = &runq;
1485 	td = runq_choose(&runq);
1486 #endif
1487 
1488 	if (td) {
1489 #ifdef SMP
1490 		if (td == tdcpu)
1491 			runq_length[PCPU_GET(cpuid)]--;
1492 #endif
1493 		runq_remove(rq, td);
1494 		td->td_flags |= TDF_DIDRUN;
1495 
1496 		KASSERT(td->td_flags & TDF_INMEM,
1497 		    ("sched_choose: thread swapped out"));
1498 		return (td);
1499 	}
1500 	return (PCPU_GET(idlethread));
1501 }
1502 
1503 void
1504 sched_preempt(struct thread *td)
1505 {
1506 
1507 	SDT_PROBE2(sched, , , surrender, td, td->td_proc);
1508 	if (td->td_critnest > 1) {
1509 		td->td_owepreempt = 1;
1510 	} else {
1511 		thread_lock(td);
1512 		mi_switch(SW_INVOL | SW_PREEMPT | SWT_PREEMPT);
1513 	}
1514 }
1515 
1516 void
1517 sched_userret_slowpath(struct thread *td)
1518 {
1519 
1520 	thread_lock(td);
1521 	td->td_priority = td->td_user_pri;
1522 	td->td_base_pri = td->td_user_pri;
1523 	thread_unlock(td);
1524 }
1525 
1526 void
1527 sched_bind(struct thread *td, int cpu)
1528 {
1529 #ifdef SMP
1530 	struct td_sched *ts = td_get_sched(td);
1531 #endif
1532 
1533 	THREAD_LOCK_ASSERT(td, MA_OWNED|MA_NOTRECURSED);
1534 	KASSERT(td == curthread, ("sched_bind: can only bind curthread"));
1535 
1536 	td->td_flags |= TDF_BOUND;
1537 #ifdef SMP
1538 	ts->ts_runq = &runq_pcpu[cpu];
1539 	if (PCPU_GET(cpuid) == cpu)
1540 		return;
1541 
1542 	mi_switch(SW_VOL);
1543 	thread_lock(td);
1544 #endif
1545 }
1546 
1547 void
1548 sched_unbind(struct thread* td)
1549 {
1550 	THREAD_LOCK_ASSERT(td, MA_OWNED);
1551 	KASSERT(td == curthread, ("sched_unbind: can only bind curthread"));
1552 	td->td_flags &= ~TDF_BOUND;
1553 }
1554 
1555 int
1556 sched_is_bound(struct thread *td)
1557 {
1558 	THREAD_LOCK_ASSERT(td, MA_OWNED);
1559 	return (td->td_flags & TDF_BOUND);
1560 }
1561 
1562 void
1563 sched_relinquish(struct thread *td)
1564 {
1565 	thread_lock(td);
1566 	mi_switch(SW_VOL | SWT_RELINQUISH);
1567 }
1568 
1569 int
1570 sched_load(void)
1571 {
1572 	return (sched_tdcnt);
1573 }
1574 
1575 int
1576 sched_sizeof_proc(void)
1577 {
1578 	return (sizeof(struct proc));
1579 }
1580 
1581 int
1582 sched_sizeof_thread(void)
1583 {
1584 	return (sizeof(struct thread) + sizeof(struct td_sched));
1585 }
1586 
1587 fixpt_t
1588 sched_pctcpu(struct thread *td)
1589 {
1590 	struct td_sched *ts;
1591 
1592 	THREAD_LOCK_ASSERT(td, MA_OWNED);
1593 	ts = td_get_sched(td);
1594 	return (ts->ts_pctcpu);
1595 }
1596 
1597 #ifdef RACCT
1598 /*
1599  * Calculates the contribution to the thread cpu usage for the latest
1600  * (unfinished) second.
1601  */
1602 fixpt_t
1603 sched_pctcpu_delta(struct thread *td)
1604 {
1605 	struct td_sched *ts;
1606 	fixpt_t delta;
1607 	int realstathz;
1608 
1609 	THREAD_LOCK_ASSERT(td, MA_OWNED);
1610 	ts = td_get_sched(td);
1611 	delta = 0;
1612 	realstathz = stathz ? stathz : hz;
1613 	if (ts->ts_cpticks != 0) {
1614 #if	(FSHIFT >= CCPU_SHIFT)
1615 		delta = (realstathz == 100)
1616 		    ? ((fixpt_t) ts->ts_cpticks) <<
1617 		    (FSHIFT - CCPU_SHIFT) :
1618 		    100 * (((fixpt_t) ts->ts_cpticks)
1619 		    << (FSHIFT - CCPU_SHIFT)) / realstathz;
1620 #else
1621 		delta = ((FSCALE - ccpu) *
1622 		    (ts->ts_cpticks *
1623 		    FSCALE / realstathz)) >> FSHIFT;
1624 #endif
1625 	}
1626 
1627 	return (delta);
1628 }
1629 #endif
1630 
1631 u_int
1632 sched_estcpu(struct thread *td)
1633 {
1634 
1635 	return (td_get_sched(td)->ts_estcpu);
1636 }
1637 
1638 /*
1639  * The actual idle process.
1640  */
1641 void
1642 sched_idletd(void *dummy)
1643 {
1644 	struct pcpuidlestat *stat;
1645 
1646 	THREAD_NO_SLEEPING();
1647 	stat = DPCPU_PTR(idlestat);
1648 	for (;;) {
1649 		mtx_assert(&Giant, MA_NOTOWNED);
1650 
1651 		while (sched_runnable() == 0) {
1652 			cpu_idle(stat->idlecalls + stat->oldidlecalls > 64);
1653 			stat->idlecalls++;
1654 		}
1655 
1656 		mtx_lock_spin(&sched_lock);
1657 		mi_switch(SW_VOL | SWT_IDLE);
1658 	}
1659 }
1660 
1661 static void
1662 sched_throw_tail(struct thread *td)
1663 {
1664 
1665 	mtx_assert(&sched_lock, MA_OWNED);
1666 	KASSERT(curthread->td_md.md_spinlock_count == 1, ("invalid count"));
1667 	cpu_throw(td, choosethread());	/* doesn't return */
1668 }
1669 
1670 /*
1671  * A CPU is entering for the first time.
1672  */
1673 void
1674 sched_ap_entry(void)
1675 {
1676 
1677 	/*
1678 	 * Correct spinlock nesting.  The idle thread context that we are
1679 	 * borrowing was created so that it would start out with a single
1680 	 * spin lock (sched_lock) held in fork_trampoline().  Since we've
1681 	 * explicitly acquired locks in this function, the nesting count
1682 	 * is now 2 rather than 1.  Since we are nested, calling
1683 	 * spinlock_exit() will simply adjust the counts without allowing
1684 	 * spin lock using code to interrupt us.
1685 	 */
1686 	mtx_lock_spin(&sched_lock);
1687 	spinlock_exit();
1688 	PCPU_SET(switchtime, cpu_ticks());
1689 	PCPU_SET(switchticks, ticks);
1690 
1691 	sched_throw_tail(NULL);
1692 }
1693 
1694 /*
1695  * A thread is exiting.
1696  */
1697 void
1698 sched_throw(struct thread *td)
1699 {
1700 
1701 	MPASS(td != NULL);
1702 	MPASS(td->td_lock == &sched_lock);
1703 
1704 	lock_profile_release_lock(&sched_lock.lock_object, true);
1705 	td->td_lastcpu = td->td_oncpu;
1706 	td->td_oncpu = NOCPU;
1707 
1708 	sched_throw_tail(td);
1709 }
1710 
1711 void
1712 sched_fork_exit(struct thread *td)
1713 {
1714 
1715 	/*
1716 	 * Finish setting up thread glue so that it begins execution in a
1717 	 * non-nested critical section with sched_lock held but not recursed.
1718 	 */
1719 	td->td_oncpu = PCPU_GET(cpuid);
1720 	sched_lock.mtx_lock = (uintptr_t)td;
1721 	lock_profile_obtain_lock_success(&sched_lock.lock_object, true,
1722 	    0, 0, __FILE__, __LINE__);
1723 	THREAD_LOCK_ASSERT(td, MA_OWNED | MA_NOTRECURSED);
1724 
1725 	KTR_STATE1(KTR_SCHED, "thread", sched_tdname(td), "running",
1726 	    "prio:%d", td->td_priority);
1727 	SDT_PROBE0(sched, , , on__cpu);
1728 }
1729 
1730 char *
1731 sched_tdname(struct thread *td)
1732 {
1733 #ifdef KTR
1734 	struct td_sched *ts;
1735 
1736 	ts = td_get_sched(td);
1737 	if (ts->ts_name[0] == '\0')
1738 		snprintf(ts->ts_name, sizeof(ts->ts_name),
1739 		    "%s tid %d", td->td_name, td->td_tid);
1740 	return (ts->ts_name);
1741 #else
1742 	return (td->td_name);
1743 #endif
1744 }
1745 
1746 #ifdef KTR
1747 void
1748 sched_clear_tdname(struct thread *td)
1749 {
1750 	struct td_sched *ts;
1751 
1752 	ts = td_get_sched(td);
1753 	ts->ts_name[0] = '\0';
1754 }
1755 #endif
1756 
1757 void
1758 sched_affinity(struct thread *td)
1759 {
1760 #ifdef SMP
1761 	struct td_sched *ts;
1762 	int cpu;
1763 
1764 	THREAD_LOCK_ASSERT(td, MA_OWNED);
1765 
1766 	/*
1767 	 * Set the TSF_AFFINITY flag if there is at least one CPU this
1768 	 * thread can't run on.
1769 	 */
1770 	ts = td_get_sched(td);
1771 	ts->ts_flags &= ~TSF_AFFINITY;
1772 	CPU_FOREACH(cpu) {
1773 		if (!THREAD_CAN_SCHED(td, cpu)) {
1774 			ts->ts_flags |= TSF_AFFINITY;
1775 			break;
1776 		}
1777 	}
1778 
1779 	/*
1780 	 * If this thread can run on all CPUs, nothing else to do.
1781 	 */
1782 	if (!(ts->ts_flags & TSF_AFFINITY))
1783 		return;
1784 
1785 	/* Pinned threads and bound threads should be left alone. */
1786 	if (td->td_pinned != 0 || td->td_flags & TDF_BOUND)
1787 		return;
1788 
1789 	switch (TD_GET_STATE(td)) {
1790 	case TDS_RUNQ:
1791 		/*
1792 		 * If we are on a per-CPU runqueue that is in the set,
1793 		 * then nothing needs to be done.
1794 		 */
1795 		if (ts->ts_runq != &runq &&
1796 		    THREAD_CAN_SCHED(td, ts->ts_runq - runq_pcpu))
1797 			return;
1798 
1799 		/* Put this thread on a valid per-CPU runqueue. */
1800 		sched_rem(td);
1801 		sched_add(td, SRQ_HOLDTD | SRQ_BORING);
1802 		break;
1803 	case TDS_RUNNING:
1804 		/*
1805 		 * See if our current CPU is in the set.  If not, force a
1806 		 * context switch.
1807 		 */
1808 		if (THREAD_CAN_SCHED(td, td->td_oncpu))
1809 			return;
1810 
1811 		td->td_flags |= TDF_NEEDRESCHED;
1812 		if (td != curthread)
1813 			ipi_cpu(cpu, IPI_AST);
1814 		break;
1815 	default:
1816 		break;
1817 	}
1818 #endif
1819 }
1820