xref: /freebsd/sys/kern/sched_4bsd.c (revision 5def4c47d4bd90b209b9b4a4ba9faec15846d8fd)
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/umtx.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 int
682 sched_runnable(void)
683 {
684 #ifdef SMP
685 	return runq_check(&runq) + runq_check(&runq_pcpu[PCPU_GET(cpuid)]);
686 #else
687 	return runq_check(&runq);
688 #endif
689 }
690 
691 int
692 sched_rr_interval(void)
693 {
694 
695 	/* Convert sched_slice from stathz to hz. */
696 	return (imax(1, (sched_slice * hz + realstathz / 2) / realstathz));
697 }
698 
699 /*
700  * We adjust the priority of the current process.  The priority of a
701  * process gets worse as it accumulates CPU time.  The cpu usage
702  * estimator (ts_estcpu) is increased here.  resetpriority() will
703  * compute a different priority each time ts_estcpu increases by
704  * INVERSE_ESTCPU_WEIGHT (until PRI_MAX_TIMESHARE is reached).  The
705  * cpu usage estimator ramps up quite quickly when the process is
706  * running (linearly), and decays away exponentially, at a rate which
707  * is proportionally slower when the system is busy.  The basic
708  * principle is that the system will 90% forget that the process used
709  * a lot of CPU time in 5 * loadav seconds.  This causes the system to
710  * favor processes which haven't run much recently, and to round-robin
711  * among other processes.
712  */
713 static void
714 sched_clock_tick(struct thread *td)
715 {
716 	struct pcpuidlestat *stat;
717 	struct td_sched *ts;
718 
719 	THREAD_LOCK_ASSERT(td, MA_OWNED);
720 	ts = td_get_sched(td);
721 
722 	ts->ts_cpticks++;
723 	ts->ts_estcpu = ESTCPULIM(ts->ts_estcpu + 1);
724 	if ((ts->ts_estcpu % INVERSE_ESTCPU_WEIGHT) == 0) {
725 		resetpriority(td);
726 		resetpriority_thread(td);
727 	}
728 
729 	/*
730 	 * Force a context switch if the current thread has used up a full
731 	 * time slice (default is 100ms).
732 	 */
733 	if (!TD_IS_IDLETHREAD(td) && --ts->ts_slice <= 0) {
734 		ts->ts_slice = sched_slice;
735 		td->td_flags |= TDF_NEEDRESCHED | TDF_SLICEEND;
736 	}
737 
738 	stat = DPCPU_PTR(idlestat);
739 	stat->oldidlecalls = stat->idlecalls;
740 	stat->idlecalls = 0;
741 }
742 
743 void
744 sched_clock(struct thread *td, int cnt)
745 {
746 
747 	for ( ; cnt > 0; cnt--)
748 		sched_clock_tick(td);
749 }
750 
751 /*
752  * Charge child's scheduling CPU usage to parent.
753  */
754 void
755 sched_exit(struct proc *p, struct thread *td)
756 {
757 
758 	KTR_STATE1(KTR_SCHED, "thread", sched_tdname(td), "proc exit",
759 	    "prio:%d", td->td_priority);
760 
761 	PROC_LOCK_ASSERT(p, MA_OWNED);
762 	sched_exit_thread(FIRST_THREAD_IN_PROC(p), td);
763 }
764 
765 void
766 sched_exit_thread(struct thread *td, struct thread *child)
767 {
768 
769 	KTR_STATE1(KTR_SCHED, "thread", sched_tdname(child), "exit",
770 	    "prio:%d", child->td_priority);
771 	thread_lock(td);
772 	td_get_sched(td)->ts_estcpu = ESTCPULIM(td_get_sched(td)->ts_estcpu +
773 	    td_get_sched(child)->ts_estcpu);
774 	thread_unlock(td);
775 	thread_lock(child);
776 	if ((child->td_flags & TDF_NOLOAD) == 0)
777 		sched_load_rem();
778 	thread_unlock(child);
779 }
780 
781 void
782 sched_fork(struct thread *td, struct thread *childtd)
783 {
784 	sched_fork_thread(td, childtd);
785 }
786 
787 void
788 sched_fork_thread(struct thread *td, struct thread *childtd)
789 {
790 	struct td_sched *ts, *tsc;
791 
792 	childtd->td_oncpu = NOCPU;
793 	childtd->td_lastcpu = NOCPU;
794 	childtd->td_lock = &sched_lock;
795 	childtd->td_cpuset = cpuset_ref(td->td_cpuset);
796 	childtd->td_domain.dr_policy = td->td_cpuset->cs_domain;
797 	childtd->td_priority = childtd->td_base_pri;
798 	ts = td_get_sched(childtd);
799 	bzero(ts, sizeof(*ts));
800 	tsc = td_get_sched(td);
801 	ts->ts_estcpu = tsc->ts_estcpu;
802 	ts->ts_flags |= (tsc->ts_flags & TSF_AFFINITY);
803 	ts->ts_slice = 1;
804 }
805 
806 void
807 sched_nice(struct proc *p, int nice)
808 {
809 	struct thread *td;
810 
811 	PROC_LOCK_ASSERT(p, MA_OWNED);
812 	p->p_nice = nice;
813 	FOREACH_THREAD_IN_PROC(p, td) {
814 		thread_lock(td);
815 		resetpriority(td);
816 		resetpriority_thread(td);
817 		thread_unlock(td);
818 	}
819 }
820 
821 void
822 sched_class(struct thread *td, int class)
823 {
824 	THREAD_LOCK_ASSERT(td, MA_OWNED);
825 	td->td_pri_class = class;
826 }
827 
828 /*
829  * Adjust the priority of a thread.
830  */
831 static void
832 sched_priority(struct thread *td, u_char prio)
833 {
834 
835 	KTR_POINT3(KTR_SCHED, "thread", sched_tdname(td), "priority change",
836 	    "prio:%d", td->td_priority, "new prio:%d", prio, KTR_ATTR_LINKED,
837 	    sched_tdname(curthread));
838 	SDT_PROBE3(sched, , , change__pri, td, td->td_proc, prio);
839 	if (td != curthread && prio > td->td_priority) {
840 		KTR_POINT3(KTR_SCHED, "thread", sched_tdname(curthread),
841 		    "lend prio", "prio:%d", td->td_priority, "new prio:%d",
842 		    prio, KTR_ATTR_LINKED, sched_tdname(td));
843 		SDT_PROBE4(sched, , , lend__pri, td, td->td_proc, prio,
844 		    curthread);
845 	}
846 	THREAD_LOCK_ASSERT(td, MA_OWNED);
847 	if (td->td_priority == prio)
848 		return;
849 	td->td_priority = prio;
850 	if (TD_ON_RUNQ(td) && td->td_rqindex != (prio / RQ_PPQ)) {
851 		sched_rem(td);
852 		sched_add(td, SRQ_BORING | SRQ_HOLDTD);
853 	}
854 }
855 
856 /*
857  * Update a thread's priority when it is lent another thread's
858  * priority.
859  */
860 void
861 sched_lend_prio(struct thread *td, u_char prio)
862 {
863 
864 	td->td_flags |= TDF_BORROWING;
865 	sched_priority(td, prio);
866 }
867 
868 /*
869  * Restore a thread's priority when priority propagation is
870  * over.  The prio argument is the minimum priority the thread
871  * needs to have to satisfy other possible priority lending
872  * requests.  If the thread's regulary priority is less
873  * important than prio the thread will keep a priority boost
874  * of prio.
875  */
876 void
877 sched_unlend_prio(struct thread *td, u_char prio)
878 {
879 	u_char base_pri;
880 
881 	if (td->td_base_pri >= PRI_MIN_TIMESHARE &&
882 	    td->td_base_pri <= PRI_MAX_TIMESHARE)
883 		base_pri = td->td_user_pri;
884 	else
885 		base_pri = td->td_base_pri;
886 	if (prio >= base_pri) {
887 		td->td_flags &= ~TDF_BORROWING;
888 		sched_prio(td, base_pri);
889 	} else
890 		sched_lend_prio(td, prio);
891 }
892 
893 void
894 sched_prio(struct thread *td, u_char prio)
895 {
896 	u_char oldprio;
897 
898 	/* First, update the base priority. */
899 	td->td_base_pri = prio;
900 
901 	/*
902 	 * If the thread is borrowing another thread's priority, don't ever
903 	 * lower the priority.
904 	 */
905 	if (td->td_flags & TDF_BORROWING && td->td_priority < prio)
906 		return;
907 
908 	/* Change the real priority. */
909 	oldprio = td->td_priority;
910 	sched_priority(td, prio);
911 
912 	/*
913 	 * If the thread is on a turnstile, then let the turnstile update
914 	 * its state.
915 	 */
916 	if (TD_ON_LOCK(td) && oldprio != prio)
917 		turnstile_adjust(td, oldprio);
918 }
919 
920 void
921 sched_user_prio(struct thread *td, u_char prio)
922 {
923 
924 	THREAD_LOCK_ASSERT(td, MA_OWNED);
925 	td->td_base_user_pri = prio;
926 	if (td->td_lend_user_pri <= prio)
927 		return;
928 	td->td_user_pri = prio;
929 }
930 
931 void
932 sched_lend_user_prio(struct thread *td, u_char prio)
933 {
934 
935 	THREAD_LOCK_ASSERT(td, MA_OWNED);
936 	td->td_lend_user_pri = prio;
937 	td->td_user_pri = min(prio, td->td_base_user_pri);
938 	if (td->td_priority > td->td_user_pri)
939 		sched_prio(td, td->td_user_pri);
940 	else if (td->td_priority != td->td_user_pri)
941 		td->td_flags |= TDF_NEEDRESCHED;
942 }
943 
944 /*
945  * Like the above but first check if there is anything to do.
946  */
947 void
948 sched_lend_user_prio_cond(struct thread *td, u_char prio)
949 {
950 
951 	if (td->td_lend_user_pri != prio)
952 		goto lend;
953 	if (td->td_user_pri != min(prio, td->td_base_user_pri))
954 		goto lend;
955 	if (td->td_priority != td->td_user_pri)
956 		goto lend;
957 	return;
958 
959 lend:
960 	thread_lock(td);
961 	sched_lend_user_prio(td, prio);
962 	thread_unlock(td);
963 }
964 
965 void
966 sched_sleep(struct thread *td, int pri)
967 {
968 
969 	THREAD_LOCK_ASSERT(td, MA_OWNED);
970 	td->td_slptick = ticks;
971 	td_get_sched(td)->ts_slptime = 0;
972 	if (pri != 0 && PRI_BASE(td->td_pri_class) == PRI_TIMESHARE)
973 		sched_prio(td, pri);
974 	if (TD_IS_SUSPENDED(td) || pri >= PSOCK)
975 		td->td_flags |= TDF_CANSWAP;
976 }
977 
978 void
979 sched_switch(struct thread *td, int flags)
980 {
981 	struct thread *newtd;
982 	struct mtx *tmtx;
983 	struct td_sched *ts;
984 	struct proc *p;
985 	int preempted;
986 
987 	tmtx = &sched_lock;
988 	ts = td_get_sched(td);
989 	p = td->td_proc;
990 
991 	THREAD_LOCK_ASSERT(td, MA_OWNED);
992 
993 	td->td_lastcpu = td->td_oncpu;
994 	preempted = (td->td_flags & TDF_SLICEEND) == 0 &&
995 	    (flags & SW_PREEMPT) != 0;
996 	td->td_flags &= ~(TDF_NEEDRESCHED | TDF_SLICEEND);
997 	td->td_owepreempt = 0;
998 	td->td_oncpu = NOCPU;
999 
1000 	/*
1001 	 * At the last moment, if this thread is still marked RUNNING,
1002 	 * then put it back on the run queue as it has not been suspended
1003 	 * or stopped or any thing else similar.  We never put the idle
1004 	 * threads on the run queue, however.
1005 	 */
1006 	if (td->td_flags & TDF_IDLETD) {
1007 		TD_SET_CAN_RUN(td);
1008 #ifdef SMP
1009 		CPU_CLR(PCPU_GET(cpuid), &idle_cpus_mask);
1010 #endif
1011 	} else {
1012 		if (TD_IS_RUNNING(td)) {
1013 			/* Put us back on the run queue. */
1014 			sched_add(td, preempted ?
1015 			    SRQ_HOLDTD|SRQ_OURSELF|SRQ_YIELDING|SRQ_PREEMPTED :
1016 			    SRQ_HOLDTD|SRQ_OURSELF|SRQ_YIELDING);
1017 		}
1018 	}
1019 
1020 	/*
1021 	 * Switch to the sched lock to fix things up and pick
1022 	 * a new thread.  Block the td_lock in order to avoid
1023 	 * breaking the critical path.
1024 	 */
1025 	if (td->td_lock != &sched_lock) {
1026 		mtx_lock_spin(&sched_lock);
1027 		tmtx = thread_lock_block(td);
1028 		mtx_unlock_spin(tmtx);
1029 	}
1030 
1031 	if ((td->td_flags & TDF_NOLOAD) == 0)
1032 		sched_load_rem();
1033 
1034 	newtd = choosethread();
1035 	MPASS(newtd->td_lock == &sched_lock);
1036 
1037 #if (KTR_COMPILE & KTR_SCHED) != 0
1038 	if (TD_IS_IDLETHREAD(td))
1039 		KTR_STATE1(KTR_SCHED, "thread", sched_tdname(td), "idle",
1040 		    "prio:%d", td->td_priority);
1041 	else
1042 		KTR_STATE3(KTR_SCHED, "thread", sched_tdname(td), KTDSTATE(td),
1043 		    "prio:%d", td->td_priority, "wmesg:\"%s\"", td->td_wmesg,
1044 		    "lockname:\"%s\"", td->td_lockname);
1045 #endif
1046 
1047 	if (td != newtd) {
1048 #ifdef	HWPMC_HOOKS
1049 		if (PMC_PROC_IS_USING_PMCS(td->td_proc))
1050 			PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_OUT);
1051 #endif
1052 
1053 		SDT_PROBE2(sched, , , off__cpu, newtd, newtd->td_proc);
1054 
1055                 /* I feel sleepy */
1056 		lock_profile_release_lock(&sched_lock.lock_object, true);
1057 #ifdef KDTRACE_HOOKS
1058 		/*
1059 		 * If DTrace has set the active vtime enum to anything
1060 		 * other than INACTIVE (0), then it should have set the
1061 		 * function to call.
1062 		 */
1063 		if (dtrace_vtime_active)
1064 			(*dtrace_vtime_switch_func)(newtd);
1065 #endif
1066 
1067 		cpu_switch(td, newtd, tmtx);
1068 		lock_profile_obtain_lock_success(&sched_lock.lock_object, true,
1069 		    0, 0, __FILE__, __LINE__);
1070 		/*
1071 		 * Where am I?  What year is it?
1072 		 * We are in the same thread that went to sleep above,
1073 		 * but any amount of time may have passed. All our context
1074 		 * will still be available as will local variables.
1075 		 * PCPU values however may have changed as we may have
1076 		 * changed CPU so don't trust cached values of them.
1077 		 * New threads will go to fork_exit() instead of here
1078 		 * so if you change things here you may need to change
1079 		 * things there too.
1080 		 *
1081 		 * If the thread above was exiting it will never wake
1082 		 * up again here, so either it has saved everything it
1083 		 * needed to, or the thread_wait() or wait() will
1084 		 * need to reap it.
1085 		 */
1086 
1087 		SDT_PROBE0(sched, , , on__cpu);
1088 #ifdef	HWPMC_HOOKS
1089 		if (PMC_PROC_IS_USING_PMCS(td->td_proc))
1090 			PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_IN);
1091 #endif
1092 	} else {
1093 		td->td_lock = &sched_lock;
1094 		SDT_PROBE0(sched, , , remain__cpu);
1095 	}
1096 
1097 	KTR_STATE1(KTR_SCHED, "thread", sched_tdname(td), "running",
1098 	    "prio:%d", td->td_priority);
1099 
1100 #ifdef SMP
1101 	if (td->td_flags & TDF_IDLETD)
1102 		CPU_SET(PCPU_GET(cpuid), &idle_cpus_mask);
1103 #endif
1104 	sched_lock.mtx_lock = (uintptr_t)td;
1105 	td->td_oncpu = PCPU_GET(cpuid);
1106 	spinlock_enter();
1107 	mtx_unlock_spin(&sched_lock);
1108 }
1109 
1110 void
1111 sched_wakeup(struct thread *td, int srqflags)
1112 {
1113 	struct td_sched *ts;
1114 
1115 	THREAD_LOCK_ASSERT(td, MA_OWNED);
1116 	ts = td_get_sched(td);
1117 	td->td_flags &= ~TDF_CANSWAP;
1118 	if (ts->ts_slptime > 1) {
1119 		updatepri(td);
1120 		resetpriority(td);
1121 	}
1122 	td->td_slptick = 0;
1123 	ts->ts_slptime = 0;
1124 	ts->ts_slice = sched_slice;
1125 	sched_add(td, srqflags);
1126 }
1127 
1128 #ifdef SMP
1129 static int
1130 forward_wakeup(int cpunum)
1131 {
1132 	struct pcpu *pc;
1133 	cpuset_t dontuse, map, map2;
1134 	u_int id, me;
1135 	int iscpuset;
1136 
1137 	mtx_assert(&sched_lock, MA_OWNED);
1138 
1139 	CTR0(KTR_RUNQ, "forward_wakeup()");
1140 
1141 	if ((!forward_wakeup_enabled) ||
1142 	     (forward_wakeup_use_mask == 0 && forward_wakeup_use_loop == 0))
1143 		return (0);
1144 	if (!smp_started || KERNEL_PANICKED())
1145 		return (0);
1146 
1147 	forward_wakeups_requested++;
1148 
1149 	/*
1150 	 * Check the idle mask we received against what we calculated
1151 	 * before in the old version.
1152 	 */
1153 	me = PCPU_GET(cpuid);
1154 
1155 	/* Don't bother if we should be doing it ourself. */
1156 	if (CPU_ISSET(me, &idle_cpus_mask) &&
1157 	    (cpunum == NOCPU || me == cpunum))
1158 		return (0);
1159 
1160 	CPU_SETOF(me, &dontuse);
1161 	CPU_OR(&dontuse, &stopped_cpus);
1162 	CPU_OR(&dontuse, &hlt_cpus_mask);
1163 	CPU_ZERO(&map2);
1164 	if (forward_wakeup_use_loop) {
1165 		STAILQ_FOREACH(pc, &cpuhead, pc_allcpu) {
1166 			id = pc->pc_cpuid;
1167 			if (!CPU_ISSET(id, &dontuse) &&
1168 			    pc->pc_curthread == pc->pc_idlethread) {
1169 				CPU_SET(id, &map2);
1170 			}
1171 		}
1172 	}
1173 
1174 	if (forward_wakeup_use_mask) {
1175 		map = idle_cpus_mask;
1176 		CPU_ANDNOT(&map, &dontuse);
1177 
1178 		/* If they are both on, compare and use loop if different. */
1179 		if (forward_wakeup_use_loop) {
1180 			if (CPU_CMP(&map, &map2)) {
1181 				printf("map != map2, loop method preferred\n");
1182 				map = map2;
1183 			}
1184 		}
1185 	} else {
1186 		map = map2;
1187 	}
1188 
1189 	/* If we only allow a specific CPU, then mask off all the others. */
1190 	if (cpunum != NOCPU) {
1191 		KASSERT((cpunum <= mp_maxcpus),("forward_wakeup: bad cpunum."));
1192 		iscpuset = CPU_ISSET(cpunum, &map);
1193 		if (iscpuset == 0)
1194 			CPU_ZERO(&map);
1195 		else
1196 			CPU_SETOF(cpunum, &map);
1197 	}
1198 	if (!CPU_EMPTY(&map)) {
1199 		forward_wakeups_delivered++;
1200 		STAILQ_FOREACH(pc, &cpuhead, pc_allcpu) {
1201 			id = pc->pc_cpuid;
1202 			if (!CPU_ISSET(id, &map))
1203 				continue;
1204 			if (cpu_idle_wakeup(pc->pc_cpuid))
1205 				CPU_CLR(id, &map);
1206 		}
1207 		if (!CPU_EMPTY(&map))
1208 			ipi_selected(map, IPI_AST);
1209 		return (1);
1210 	}
1211 	if (cpunum == NOCPU)
1212 		printf("forward_wakeup: Idle processor not found\n");
1213 	return (0);
1214 }
1215 
1216 static void
1217 kick_other_cpu(int pri, int cpuid)
1218 {
1219 	struct pcpu *pcpu;
1220 	int cpri;
1221 
1222 	pcpu = pcpu_find(cpuid);
1223 	if (CPU_ISSET(cpuid, &idle_cpus_mask)) {
1224 		forward_wakeups_delivered++;
1225 		if (!cpu_idle_wakeup(cpuid))
1226 			ipi_cpu(cpuid, IPI_AST);
1227 		return;
1228 	}
1229 
1230 	cpri = pcpu->pc_curthread->td_priority;
1231 	if (pri >= cpri)
1232 		return;
1233 
1234 #if defined(IPI_PREEMPTION) && defined(PREEMPTION)
1235 #if !defined(FULL_PREEMPTION)
1236 	if (pri <= PRI_MAX_ITHD)
1237 #endif /* ! FULL_PREEMPTION */
1238 	{
1239 		ipi_cpu(cpuid, IPI_PREEMPT);
1240 		return;
1241 	}
1242 #endif /* defined(IPI_PREEMPTION) && defined(PREEMPTION) */
1243 
1244 	pcpu->pc_curthread->td_flags |= TDF_NEEDRESCHED;
1245 	ipi_cpu(cpuid, IPI_AST);
1246 	return;
1247 }
1248 #endif /* SMP */
1249 
1250 #ifdef SMP
1251 static int
1252 sched_pickcpu(struct thread *td)
1253 {
1254 	int best, cpu;
1255 
1256 	mtx_assert(&sched_lock, MA_OWNED);
1257 
1258 	if (td->td_lastcpu != NOCPU && THREAD_CAN_SCHED(td, td->td_lastcpu))
1259 		best = td->td_lastcpu;
1260 	else
1261 		best = NOCPU;
1262 	CPU_FOREACH(cpu) {
1263 		if (!THREAD_CAN_SCHED(td, cpu))
1264 			continue;
1265 
1266 		if (best == NOCPU)
1267 			best = cpu;
1268 		else if (runq_length[cpu] < runq_length[best])
1269 			best = cpu;
1270 	}
1271 	KASSERT(best != NOCPU, ("no valid CPUs"));
1272 
1273 	return (best);
1274 }
1275 #endif
1276 
1277 void
1278 sched_add(struct thread *td, int flags)
1279 #ifdef SMP
1280 {
1281 	cpuset_t tidlemsk;
1282 	struct td_sched *ts;
1283 	u_int cpu, cpuid;
1284 	int forwarded = 0;
1285 	int single_cpu = 0;
1286 
1287 	ts = td_get_sched(td);
1288 	THREAD_LOCK_ASSERT(td, MA_OWNED);
1289 	KASSERT((td->td_inhibitors == 0),
1290 	    ("sched_add: trying to run inhibited thread"));
1291 	KASSERT((TD_CAN_RUN(td) || TD_IS_RUNNING(td)),
1292 	    ("sched_add: bad thread state"));
1293 	KASSERT(td->td_flags & TDF_INMEM,
1294 	    ("sched_add: thread swapped out"));
1295 
1296 	KTR_STATE2(KTR_SCHED, "thread", sched_tdname(td), "runq add",
1297 	    "prio:%d", td->td_priority, KTR_ATTR_LINKED,
1298 	    sched_tdname(curthread));
1299 	KTR_POINT1(KTR_SCHED, "thread", sched_tdname(curthread), "wokeup",
1300 	    KTR_ATTR_LINKED, sched_tdname(td));
1301 	SDT_PROBE4(sched, , , enqueue, td, td->td_proc, NULL,
1302 	    flags & SRQ_PREEMPTED);
1303 
1304 	/*
1305 	 * Now that the thread is moving to the run-queue, set the lock
1306 	 * to the scheduler's lock.
1307 	 */
1308 	if (td->td_lock != &sched_lock) {
1309 		mtx_lock_spin(&sched_lock);
1310 		if ((flags & SRQ_HOLD) != 0)
1311 			td->td_lock = &sched_lock;
1312 		else
1313 			thread_lock_set(td, &sched_lock);
1314 	}
1315 	TD_SET_RUNQ(td);
1316 
1317 	/*
1318 	 * If SMP is started and the thread is pinned or otherwise limited to
1319 	 * a specific set of CPUs, queue the thread to a per-CPU run queue.
1320 	 * Otherwise, queue the thread to the global run queue.
1321 	 *
1322 	 * If SMP has not yet been started we must use the global run queue
1323 	 * as per-CPU state may not be initialized yet and we may crash if we
1324 	 * try to access the per-CPU run queues.
1325 	 */
1326 	if (smp_started && (td->td_pinned != 0 || td->td_flags & TDF_BOUND ||
1327 	    ts->ts_flags & TSF_AFFINITY)) {
1328 		if (td->td_pinned != 0)
1329 			cpu = td->td_lastcpu;
1330 		else if (td->td_flags & TDF_BOUND) {
1331 			/* Find CPU from bound runq. */
1332 			KASSERT(SKE_RUNQ_PCPU(ts),
1333 			    ("sched_add: bound td_sched not on cpu runq"));
1334 			cpu = ts->ts_runq - &runq_pcpu[0];
1335 		} else
1336 			/* Find a valid CPU for our cpuset */
1337 			cpu = sched_pickcpu(td);
1338 		ts->ts_runq = &runq_pcpu[cpu];
1339 		single_cpu = 1;
1340 		CTR3(KTR_RUNQ,
1341 		    "sched_add: Put td_sched:%p(td:%p) on cpu%d runq", ts, td,
1342 		    cpu);
1343 	} else {
1344 		CTR2(KTR_RUNQ,
1345 		    "sched_add: adding td_sched:%p (td:%p) to gbl runq", ts,
1346 		    td);
1347 		cpu = NOCPU;
1348 		ts->ts_runq = &runq;
1349 	}
1350 
1351 	if ((td->td_flags & TDF_NOLOAD) == 0)
1352 		sched_load_add();
1353 	runq_add(ts->ts_runq, td, flags);
1354 	if (cpu != NOCPU)
1355 		runq_length[cpu]++;
1356 
1357 	cpuid = PCPU_GET(cpuid);
1358 	if (single_cpu && cpu != cpuid) {
1359 	        kick_other_cpu(td->td_priority, cpu);
1360 	} else {
1361 		if (!single_cpu) {
1362 			tidlemsk = idle_cpus_mask;
1363 			CPU_ANDNOT(&tidlemsk, &hlt_cpus_mask);
1364 			CPU_CLR(cpuid, &tidlemsk);
1365 
1366 			if (!CPU_ISSET(cpuid, &idle_cpus_mask) &&
1367 			    ((flags & SRQ_INTR) == 0) &&
1368 			    !CPU_EMPTY(&tidlemsk))
1369 				forwarded = forward_wakeup(cpu);
1370 		}
1371 
1372 		if (!forwarded) {
1373 			if (!maybe_preempt(td))
1374 				maybe_resched(td);
1375 		}
1376 	}
1377 	if ((flags & SRQ_HOLDTD) == 0)
1378 		thread_unlock(td);
1379 }
1380 #else /* SMP */
1381 {
1382 	struct td_sched *ts;
1383 
1384 	ts = td_get_sched(td);
1385 	THREAD_LOCK_ASSERT(td, MA_OWNED);
1386 	KASSERT((td->td_inhibitors == 0),
1387 	    ("sched_add: trying to run inhibited thread"));
1388 	KASSERT((TD_CAN_RUN(td) || TD_IS_RUNNING(td)),
1389 	    ("sched_add: bad thread state"));
1390 	KASSERT(td->td_flags & TDF_INMEM,
1391 	    ("sched_add: thread swapped out"));
1392 	KTR_STATE2(KTR_SCHED, "thread", sched_tdname(td), "runq add",
1393 	    "prio:%d", td->td_priority, KTR_ATTR_LINKED,
1394 	    sched_tdname(curthread));
1395 	KTR_POINT1(KTR_SCHED, "thread", sched_tdname(curthread), "wokeup",
1396 	    KTR_ATTR_LINKED, sched_tdname(td));
1397 	SDT_PROBE4(sched, , , enqueue, td, td->td_proc, NULL,
1398 	    flags & SRQ_PREEMPTED);
1399 
1400 	/*
1401 	 * Now that the thread is moving to the run-queue, set the lock
1402 	 * to the scheduler's lock.
1403 	 */
1404 	if (td->td_lock != &sched_lock) {
1405 		mtx_lock_spin(&sched_lock);
1406 		if ((flags & SRQ_HOLD) != 0)
1407 			td->td_lock = &sched_lock;
1408 		else
1409 			thread_lock_set(td, &sched_lock);
1410 	}
1411 	TD_SET_RUNQ(td);
1412 	CTR2(KTR_RUNQ, "sched_add: adding td_sched:%p (td:%p) to runq", ts, td);
1413 	ts->ts_runq = &runq;
1414 
1415 	if ((td->td_flags & TDF_NOLOAD) == 0)
1416 		sched_load_add();
1417 	runq_add(ts->ts_runq, td, flags);
1418 	if (!maybe_preempt(td))
1419 		maybe_resched(td);
1420 	if ((flags & SRQ_HOLDTD) == 0)
1421 		thread_unlock(td);
1422 }
1423 #endif /* SMP */
1424 
1425 void
1426 sched_rem(struct thread *td)
1427 {
1428 	struct td_sched *ts;
1429 
1430 	ts = td_get_sched(td);
1431 	KASSERT(td->td_flags & TDF_INMEM,
1432 	    ("sched_rem: thread swapped out"));
1433 	KASSERT(TD_ON_RUNQ(td),
1434 	    ("sched_rem: thread not on run queue"));
1435 	mtx_assert(&sched_lock, MA_OWNED);
1436 	KTR_STATE2(KTR_SCHED, "thread", sched_tdname(td), "runq rem",
1437 	    "prio:%d", td->td_priority, KTR_ATTR_LINKED,
1438 	    sched_tdname(curthread));
1439 	SDT_PROBE3(sched, , , dequeue, td, td->td_proc, NULL);
1440 
1441 	if ((td->td_flags & TDF_NOLOAD) == 0)
1442 		sched_load_rem();
1443 #ifdef SMP
1444 	if (ts->ts_runq != &runq)
1445 		runq_length[ts->ts_runq - runq_pcpu]--;
1446 #endif
1447 	runq_remove(ts->ts_runq, td);
1448 	TD_SET_CAN_RUN(td);
1449 }
1450 
1451 /*
1452  * Select threads to run.  Note that running threads still consume a
1453  * slot.
1454  */
1455 struct thread *
1456 sched_choose(void)
1457 {
1458 	struct thread *td;
1459 	struct runq *rq;
1460 
1461 	mtx_assert(&sched_lock,  MA_OWNED);
1462 #ifdef SMP
1463 	struct thread *tdcpu;
1464 
1465 	rq = &runq;
1466 	td = runq_choose_fuzz(&runq, runq_fuzz);
1467 	tdcpu = runq_choose(&runq_pcpu[PCPU_GET(cpuid)]);
1468 
1469 	if (td == NULL ||
1470 	    (tdcpu != NULL &&
1471 	     tdcpu->td_priority < td->td_priority)) {
1472 		CTR2(KTR_RUNQ, "choosing td %p from pcpu runq %d", tdcpu,
1473 		     PCPU_GET(cpuid));
1474 		td = tdcpu;
1475 		rq = &runq_pcpu[PCPU_GET(cpuid)];
1476 	} else {
1477 		CTR1(KTR_RUNQ, "choosing td_sched %p from main runq", td);
1478 	}
1479 
1480 #else
1481 	rq = &runq;
1482 	td = runq_choose(&runq);
1483 #endif
1484 
1485 	if (td) {
1486 #ifdef SMP
1487 		if (td == tdcpu)
1488 			runq_length[PCPU_GET(cpuid)]--;
1489 #endif
1490 		runq_remove(rq, td);
1491 		td->td_flags |= TDF_DIDRUN;
1492 
1493 		KASSERT(td->td_flags & TDF_INMEM,
1494 		    ("sched_choose: thread swapped out"));
1495 		return (td);
1496 	}
1497 	return (PCPU_GET(idlethread));
1498 }
1499 
1500 void
1501 sched_preempt(struct thread *td)
1502 {
1503 
1504 	SDT_PROBE2(sched, , , surrender, td, td->td_proc);
1505 	if (td->td_critnest > 1) {
1506 		td->td_owepreempt = 1;
1507 	} else {
1508 		thread_lock(td);
1509 		mi_switch(SW_INVOL | SW_PREEMPT | SWT_PREEMPT);
1510 	}
1511 }
1512 
1513 void
1514 sched_userret_slowpath(struct thread *td)
1515 {
1516 
1517 	thread_lock(td);
1518 	td->td_priority = td->td_user_pri;
1519 	td->td_base_pri = td->td_user_pri;
1520 	thread_unlock(td);
1521 }
1522 
1523 void
1524 sched_bind(struct thread *td, int cpu)
1525 {
1526 	struct td_sched *ts;
1527 
1528 	THREAD_LOCK_ASSERT(td, MA_OWNED|MA_NOTRECURSED);
1529 	KASSERT(td == curthread, ("sched_bind: can only bind curthread"));
1530 
1531 	ts = td_get_sched(td);
1532 
1533 	td->td_flags |= TDF_BOUND;
1534 #ifdef SMP
1535 	ts->ts_runq = &runq_pcpu[cpu];
1536 	if (PCPU_GET(cpuid) == cpu)
1537 		return;
1538 
1539 	mi_switch(SW_VOL);
1540 	thread_lock(td);
1541 #endif
1542 }
1543 
1544 void
1545 sched_unbind(struct thread* td)
1546 {
1547 	THREAD_LOCK_ASSERT(td, MA_OWNED);
1548 	KASSERT(td == curthread, ("sched_unbind: can only bind curthread"));
1549 	td->td_flags &= ~TDF_BOUND;
1550 }
1551 
1552 int
1553 sched_is_bound(struct thread *td)
1554 {
1555 	THREAD_LOCK_ASSERT(td, MA_OWNED);
1556 	return (td->td_flags & TDF_BOUND);
1557 }
1558 
1559 void
1560 sched_relinquish(struct thread *td)
1561 {
1562 	thread_lock(td);
1563 	mi_switch(SW_VOL | SWT_RELINQUISH);
1564 }
1565 
1566 int
1567 sched_load(void)
1568 {
1569 	return (sched_tdcnt);
1570 }
1571 
1572 int
1573 sched_sizeof_proc(void)
1574 {
1575 	return (sizeof(struct proc));
1576 }
1577 
1578 int
1579 sched_sizeof_thread(void)
1580 {
1581 	return (sizeof(struct thread) + sizeof(struct td_sched));
1582 }
1583 
1584 fixpt_t
1585 sched_pctcpu(struct thread *td)
1586 {
1587 	struct td_sched *ts;
1588 
1589 	THREAD_LOCK_ASSERT(td, MA_OWNED);
1590 	ts = td_get_sched(td);
1591 	return (ts->ts_pctcpu);
1592 }
1593 
1594 #ifdef RACCT
1595 /*
1596  * Calculates the contribution to the thread cpu usage for the latest
1597  * (unfinished) second.
1598  */
1599 fixpt_t
1600 sched_pctcpu_delta(struct thread *td)
1601 {
1602 	struct td_sched *ts;
1603 	fixpt_t delta;
1604 	int realstathz;
1605 
1606 	THREAD_LOCK_ASSERT(td, MA_OWNED);
1607 	ts = td_get_sched(td);
1608 	delta = 0;
1609 	realstathz = stathz ? stathz : hz;
1610 	if (ts->ts_cpticks != 0) {
1611 #if	(FSHIFT >= CCPU_SHIFT)
1612 		delta = (realstathz == 100)
1613 		    ? ((fixpt_t) ts->ts_cpticks) <<
1614 		    (FSHIFT - CCPU_SHIFT) :
1615 		    100 * (((fixpt_t) ts->ts_cpticks)
1616 		    << (FSHIFT - CCPU_SHIFT)) / realstathz;
1617 #else
1618 		delta = ((FSCALE - ccpu) *
1619 		    (ts->ts_cpticks *
1620 		    FSCALE / realstathz)) >> FSHIFT;
1621 #endif
1622 	}
1623 
1624 	return (delta);
1625 }
1626 #endif
1627 
1628 u_int
1629 sched_estcpu(struct thread *td)
1630 {
1631 
1632 	return (td_get_sched(td)->ts_estcpu);
1633 }
1634 
1635 /*
1636  * The actual idle process.
1637  */
1638 void
1639 sched_idletd(void *dummy)
1640 {
1641 	struct pcpuidlestat *stat;
1642 
1643 	THREAD_NO_SLEEPING();
1644 	stat = DPCPU_PTR(idlestat);
1645 	for (;;) {
1646 		mtx_assert(&Giant, MA_NOTOWNED);
1647 
1648 		while (sched_runnable() == 0) {
1649 			cpu_idle(stat->idlecalls + stat->oldidlecalls > 64);
1650 			stat->idlecalls++;
1651 		}
1652 
1653 		mtx_lock_spin(&sched_lock);
1654 		mi_switch(SW_VOL | SWT_IDLE);
1655 	}
1656 }
1657 
1658 /*
1659  * A CPU is entering for the first time or a thread is exiting.
1660  */
1661 void
1662 sched_throw(struct thread *td)
1663 {
1664 	/*
1665 	 * Correct spinlock nesting.  The idle thread context that we are
1666 	 * borrowing was created so that it would start out with a single
1667 	 * spin lock (sched_lock) held in fork_trampoline().  Since we've
1668 	 * explicitly acquired locks in this function, the nesting count
1669 	 * is now 2 rather than 1.  Since we are nested, calling
1670 	 * spinlock_exit() will simply adjust the counts without allowing
1671 	 * spin lock using code to interrupt us.
1672 	 */
1673 	if (td == NULL) {
1674 		mtx_lock_spin(&sched_lock);
1675 		spinlock_exit();
1676 		PCPU_SET(switchtime, cpu_ticks());
1677 		PCPU_SET(switchticks, ticks);
1678 	} else {
1679 		lock_profile_release_lock(&sched_lock.lock_object, true);
1680 		MPASS(td->td_lock == &sched_lock);
1681 		td->td_lastcpu = td->td_oncpu;
1682 		td->td_oncpu = NOCPU;
1683 	}
1684 	mtx_assert(&sched_lock, MA_OWNED);
1685 	KASSERT(curthread->td_md.md_spinlock_count == 1, ("invalid count"));
1686 	cpu_throw(td, choosethread());	/* doesn't return */
1687 }
1688 
1689 void
1690 sched_fork_exit(struct thread *td)
1691 {
1692 
1693 	/*
1694 	 * Finish setting up thread glue so that it begins execution in a
1695 	 * non-nested critical section with sched_lock held but not recursed.
1696 	 */
1697 	td->td_oncpu = PCPU_GET(cpuid);
1698 	sched_lock.mtx_lock = (uintptr_t)td;
1699 	lock_profile_obtain_lock_success(&sched_lock.lock_object, true,
1700 	    0, 0, __FILE__, __LINE__);
1701 	THREAD_LOCK_ASSERT(td, MA_OWNED | MA_NOTRECURSED);
1702 
1703 	KTR_STATE1(KTR_SCHED, "thread", sched_tdname(td), "running",
1704 	    "prio:%d", td->td_priority);
1705 	SDT_PROBE0(sched, , , on__cpu);
1706 }
1707 
1708 char *
1709 sched_tdname(struct thread *td)
1710 {
1711 #ifdef KTR
1712 	struct td_sched *ts;
1713 
1714 	ts = td_get_sched(td);
1715 	if (ts->ts_name[0] == '\0')
1716 		snprintf(ts->ts_name, sizeof(ts->ts_name),
1717 		    "%s tid %d", td->td_name, td->td_tid);
1718 	return (ts->ts_name);
1719 #else
1720 	return (td->td_name);
1721 #endif
1722 }
1723 
1724 #ifdef KTR
1725 void
1726 sched_clear_tdname(struct thread *td)
1727 {
1728 	struct td_sched *ts;
1729 
1730 	ts = td_get_sched(td);
1731 	ts->ts_name[0] = '\0';
1732 }
1733 #endif
1734 
1735 void
1736 sched_affinity(struct thread *td)
1737 {
1738 #ifdef SMP
1739 	struct td_sched *ts;
1740 	int cpu;
1741 
1742 	THREAD_LOCK_ASSERT(td, MA_OWNED);
1743 
1744 	/*
1745 	 * Set the TSF_AFFINITY flag if there is at least one CPU this
1746 	 * thread can't run on.
1747 	 */
1748 	ts = td_get_sched(td);
1749 	ts->ts_flags &= ~TSF_AFFINITY;
1750 	CPU_FOREACH(cpu) {
1751 		if (!THREAD_CAN_SCHED(td, cpu)) {
1752 			ts->ts_flags |= TSF_AFFINITY;
1753 			break;
1754 		}
1755 	}
1756 
1757 	/*
1758 	 * If this thread can run on all CPUs, nothing else to do.
1759 	 */
1760 	if (!(ts->ts_flags & TSF_AFFINITY))
1761 		return;
1762 
1763 	/* Pinned threads and bound threads should be left alone. */
1764 	if (td->td_pinned != 0 || td->td_flags & TDF_BOUND)
1765 		return;
1766 
1767 	switch (TD_GET_STATE(td)) {
1768 	case TDS_RUNQ:
1769 		/*
1770 		 * If we are on a per-CPU runqueue that is in the set,
1771 		 * then nothing needs to be done.
1772 		 */
1773 		if (ts->ts_runq != &runq &&
1774 		    THREAD_CAN_SCHED(td, ts->ts_runq - runq_pcpu))
1775 			return;
1776 
1777 		/* Put this thread on a valid per-CPU runqueue. */
1778 		sched_rem(td);
1779 		sched_add(td, SRQ_HOLDTD | SRQ_BORING);
1780 		break;
1781 	case TDS_RUNNING:
1782 		/*
1783 		 * See if our current CPU is in the set.  If not, force a
1784 		 * context switch.
1785 		 */
1786 		if (THREAD_CAN_SCHED(td, td->td_oncpu))
1787 			return;
1788 
1789 		td->td_flags |= TDF_NEEDRESCHED;
1790 		if (td != curthread)
1791 			ipi_cpu(cpu, IPI_AST);
1792 		break;
1793 	default:
1794 		break;
1795 	}
1796 #endif
1797 }
1798