xref: /freebsd/sys/kern/kern_switch.c (revision 7842f65e7f8970faa152cdd92cb266a9f558fe36)
1 /*
2  * Copyright (c) 2001 Jake Burkholder <jake@FreeBSD.org>
3  * All rights reserved.
4  *
5  * Redistribution and use in source and binary forms, with or without
6  * modification, are permitted provided that the following conditions
7  * are met:
8  * 1. Redistributions of source code must retain the above copyright
9  *    notice, this list of conditions and the following disclaimer.
10  * 2. Redistributions in binary form must reproduce the above copyright
11  *    notice, this list of conditions and the following disclaimer in the
12  *    documentation and/or other materials provided with the distribution.
13  *
14  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
15  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
16  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
17  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
18  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
19  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
20  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
21  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
22  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
23  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
24  * SUCH DAMAGE.
25  */
26 
27 /***
28 Here is the logic..
29 
30 If there are N processors, then there are at most N KSEs (kernel
31 schedulable entities) working to process threads that belong to a
32 KSEGROUP (kg). If there are X of these KSEs actually running at the
33 moment in question, then there are at most M (N-X) of these KSEs on
34 the run queue, as running KSEs are not on the queue.
35 
36 Runnable threads are queued off the KSEGROUP in priority order.
37 If there are M or more threads runnable, the top M threads
38 (by priority) are 'preassigned' to the M KSEs not running. The KSEs take
39 their priority from those threads and are put on the run queue.
40 
41 The last thread that had a priority high enough to have a KSE associated
42 with it, AND IS ON THE RUN QUEUE is pointed to by
43 kg->kg_last_assigned. If no threads queued off the KSEGROUP have KSEs
44 assigned as all the available KSEs are activly running, or because there
45 are no threads queued, that pointer is NULL.
46 
47 When a KSE is removed from the run queue to become runnable, we know
48 it was associated with the highest priority thread in the queue (at the head
49 of the queue). If it is also the last assigned we know M was 1 and must
50 now be 0. Since the thread is no longer queued that pointer must be
51 removed from it. Since we know there were no more KSEs available,
52 (M was 1 and is now 0) and since we are not FREEING our KSE
53 but using it, we know there are STILL no more KSEs available, we can prove
54 that the next thread in the ksegrp list will not have a KSE to assign to
55 it, so we can show that the pointer must be made 'invalid' (NULL).
56 
57 The pointer exists so that when a new thread is made runnable, it can
58 have its priority compared with the last assigned thread to see if
59 it should 'steal' its KSE or not.. i.e. is it 'earlier'
60 on the list than that thread or later.. If it's earlier, then the KSE is
61 removed from the last assigned (which is now not assigned a KSE)
62 and reassigned to the new thread, which is placed earlier in the list.
63 The pointer is then backed up to the previous thread (which may or may not
64 be the new thread).
65 
66 When a thread sleeps or is removed, the KSE becomes available and if there
67 are queued threads that are not assigned KSEs, the highest priority one of
68 them is assigned the KSE, which is then placed back on the run queue at
69 the approipriate place, and the kg->kg_last_assigned pointer is adjusted down
70 to point to it.
71 
72 The following diagram shows 2 KSEs and 3 threads from a single process.
73 
74  RUNQ: --->KSE---KSE--...    (KSEs queued at priorities from threads)
75               \    \____
76                \        \
77     KSEGROUP---thread--thread--thread    (queued in priority order)
78         \                 /
79          \_______________/
80           (last_assigned)
81 
82 The result of this scheme is that the M available KSEs are always
83 queued at the priorities they have inherrited from the M highest priority
84 threads for that KSEGROUP. If this situation changes, the KSEs are
85 reassigned to keep this true.
86 ***/
87 
88 #include <sys/cdefs.h>
89 __FBSDID("$FreeBSD$");
90 
91 #include "opt_sched.h"
92 
93 #ifndef KERN_SWITCH_INCLUDE
94 #include <sys/param.h>
95 #include <sys/systm.h>
96 #include <sys/kdb.h>
97 #include <sys/kernel.h>
98 #include <sys/ktr.h>
99 #include <sys/lock.h>
100 #include <sys/mutex.h>
101 #include <sys/proc.h>
102 #include <sys/queue.h>
103 #include <sys/sched.h>
104 #else  /* KERN_SWITCH_INCLUDE */
105 #if defined(SMP) && (defined(__i386__) || defined(__amd64__))
106 #include <sys/smp.h>
107 #endif
108 #include <machine/critical.h>
109 #if defined(SMP) && defined(SCHED_4BSD)
110 #include <sys/sysctl.h>
111 #endif
112 
113 #ifdef FULL_PREEMPTION
114 #ifndef PREEMPTION
115 #error "The FULL_PREEMPTION option requires the PREEMPTION option"
116 #endif
117 #endif
118 
119 CTASSERT((RQB_BPW * RQB_LEN) == RQ_NQS);
120 
121 #define td_kse td_sched
122 
123 /************************************************************************
124  * Functions that manipulate runnability from a thread perspective.	*
125  ************************************************************************/
126 /*
127  * Select the KSE that will be run next.  From that find the thread, and
128  * remove it from the KSEGRP's run queue.  If there is thread clustering,
129  * this will be what does it.
130  */
131 struct thread *
132 choosethread(void)
133 {
134 	struct kse *ke;
135 	struct thread *td;
136 	struct ksegrp *kg;
137 
138 #if defined(SMP) && (defined(__i386__) || defined(__amd64__))
139 	if (smp_active == 0 && PCPU_GET(cpuid) != 0) {
140 		/* Shutting down, run idlethread on AP's */
141 		td = PCPU_GET(idlethread);
142 		ke = td->td_kse;
143 		CTR1(KTR_RUNQ, "choosethread: td=%p (idle)", td);
144 		ke->ke_flags |= KEF_DIDRUN;
145 		TD_SET_RUNNING(td);
146 		return (td);
147 	}
148 #endif
149 
150 retry:
151 	ke = sched_choose();
152 	if (ke) {
153 		td = ke->ke_thread;
154 		KASSERT((td->td_kse == ke), ("kse/thread mismatch"));
155 		kg = ke->ke_ksegrp;
156 		if (td->td_proc->p_flag & P_HADTHREADS) {
157 			if (kg->kg_last_assigned == td) {
158 				kg->kg_last_assigned = TAILQ_PREV(td,
159 				    threadqueue, td_runq);
160 			}
161 			TAILQ_REMOVE(&kg->kg_runq, td, td_runq);
162 		}
163 		CTR2(KTR_RUNQ, "choosethread: td=%p pri=%d",
164 		    td, td->td_priority);
165 	} else {
166 		/* Simulate runq_choose() having returned the idle thread */
167 		td = PCPU_GET(idlethread);
168 		ke = td->td_kse;
169 		CTR1(KTR_RUNQ, "choosethread: td=%p (idle)", td);
170 	}
171 	ke->ke_flags |= KEF_DIDRUN;
172 
173 	/*
174 	 * If we are in panic, only allow system threads,
175 	 * plus the one we are running in, to be run.
176 	 */
177 	if (panicstr && ((td->td_proc->p_flag & P_SYSTEM) == 0 &&
178 	    (td->td_flags & TDF_INPANIC) == 0)) {
179 		/* note that it is no longer on the run queue */
180 		TD_SET_CAN_RUN(td);
181 		goto retry;
182 	}
183 
184 	TD_SET_RUNNING(td);
185 	return (td);
186 }
187 
188 /*
189  * Given a surplus system slot, try assign a new runnable thread to it.
190  * Called from:
191  *  sched_thread_exit()  (local)
192  *  sched_switch()  (local)
193  *  sched_thread_exit()  (local)
194  *  remrunqueue()  (local)  (not at the moment)
195  */
196 static void
197 slot_fill(struct ksegrp *kg)
198 {
199 	struct thread *td;
200 
201 	mtx_assert(&sched_lock, MA_OWNED);
202 	while (kg->kg_avail_opennings > 0) {
203 		/*
204 		 * Find the first unassigned thread
205 		 */
206 		if ((td = kg->kg_last_assigned) != NULL)
207 			td = TAILQ_NEXT(td, td_runq);
208 		else
209 			td = TAILQ_FIRST(&kg->kg_runq);
210 
211 		/*
212 		 * If we found one, send it to the system scheduler.
213 		 */
214 		if (td) {
215 			kg->kg_last_assigned = td;
216 			sched_add(td, SRQ_YIELDING);
217 			CTR2(KTR_RUNQ, "slot_fill: td%p -> kg%p", td, kg);
218 		} else {
219 			/* no threads to use up the slots. quit now */
220 			break;
221 		}
222 	}
223 }
224 
225 #ifdef	SCHED_4BSD
226 /*
227  * Remove a thread from its KSEGRP's run queue.
228  * This in turn may remove it from a KSE if it was already assigned
229  * to one, possibly causing a new thread to be assigned to the KSE
230  * and the KSE getting a new priority.
231  */
232 static void
233 remrunqueue(struct thread *td)
234 {
235 	struct thread *td2, *td3;
236 	struct ksegrp *kg;
237 	struct kse *ke;
238 
239 	mtx_assert(&sched_lock, MA_OWNED);
240 	KASSERT((TD_ON_RUNQ(td)), ("remrunqueue: Bad state on run queue"));
241 	kg = td->td_ksegrp;
242 	ke = td->td_kse;
243 	CTR1(KTR_RUNQ, "remrunqueue: td%p", td);
244 	TD_SET_CAN_RUN(td);
245 	/*
246 	 * If it is not a threaded process, take the shortcut.
247 	 */
248 	if ((td->td_proc->p_flag & P_HADTHREADS) == 0) {
249 		/* remve from sys run queue and free up a slot */
250 		sched_rem(td);
251 		ke->ke_state = KES_THREAD;
252 		return;
253 	}
254    	td3 = TAILQ_PREV(td, threadqueue, td_runq);
255 	TAILQ_REMOVE(&kg->kg_runq, td, td_runq);
256 	if (ke->ke_state == KES_ONRUNQ) {
257 		/*
258 		 * This thread has been assigned to the system run queue.
259 		 * We need to dissociate it and try assign the
260 		 * KSE to the next available thread. Then, we should
261 		 * see if we need to move the KSE in the run queues.
262 		 */
263 		sched_rem(td);
264 		ke->ke_state = KES_THREAD;
265 		td2 = kg->kg_last_assigned;
266 		KASSERT((td2 != NULL), ("last assigned has wrong value"));
267 		if (td2 == td)
268 			kg->kg_last_assigned = td3;
269 		/* slot_fill(kg); */ /* will replace it with another */
270 	}
271 }
272 #endif
273 
274 /*
275  * Change the priority of a thread that is on the run queue.
276  */
277 void
278 adjustrunqueue( struct thread *td, int newpri)
279 {
280 	struct ksegrp *kg;
281 	struct kse *ke;
282 
283 	mtx_assert(&sched_lock, MA_OWNED);
284 	KASSERT((TD_ON_RUNQ(td)), ("adjustrunqueue: Bad state on run queue"));
285 
286 	ke = td->td_kse;
287 	CTR1(KTR_RUNQ, "adjustrunqueue: td%p", td);
288 	/*
289 	 * If it is not a threaded process, take the shortcut.
290 	 */
291 	if ((td->td_proc->p_flag & P_HADTHREADS) == 0) {
292 		/* We only care about the kse in the run queue. */
293 		td->td_priority = newpri;
294 		if (ke->ke_rqindex != (newpri / RQ_PPQ)) {
295 			sched_rem(td);
296 			sched_add(td, SRQ_BORING);
297 		}
298 		return;
299 	}
300 
301 	/* It is a threaded process */
302 	kg = td->td_ksegrp;
303 	if (ke->ke_state == KES_ONRUNQ) {
304 		if (kg->kg_last_assigned == td) {
305 			kg->kg_last_assigned =
306 			    TAILQ_PREV(td, threadqueue, td_runq);
307 		}
308 		sched_rem(td);
309 	}
310 	TAILQ_REMOVE(&kg->kg_runq, td, td_runq);
311 	TD_SET_CAN_RUN(td);
312 	td->td_priority = newpri;
313 	setrunqueue(td, SRQ_BORING);
314 }
315 
316 /*
317  * This function is called when a thread is about to be put on a
318  * ksegrp run queue because it has been made runnable or its
319  * priority has been adjusted and the ksegrp does not have a
320  * free kse slot.  It determines if a thread from the same ksegrp
321  * should be preempted.  If so, it tries to switch threads
322  * if the thread is on the same cpu or notifies another cpu that
323  * it should switch threads.
324  */
325 
326 static void
327 maybe_preempt_in_ksegrp(struct thread *td)
328 #if  !defined(SMP)
329 {
330 	struct thread *running_thread;
331 
332 #ifndef FULL_PREEMPTION
333 	int pri;
334 	pri = td->td_priority;
335 	if (!(pri >= PRI_MIN_ITHD && pri <= PRI_MAX_ITHD))
336 		return;
337 #endif
338 	mtx_assert(&sched_lock, MA_OWNED);
339 	running_thread = curthread;
340 
341 	if (running_thread->td_ksegrp != td->td_ksegrp)
342 		return;
343 
344 	if (td->td_priority > running_thread->td_priority)
345 		return;
346 #ifdef PREEMPTION
347 	if (running_thread->td_critnest > 1)
348 		running_thread->td_pflags |= TDP_OWEPREEMPT;
349 	 else
350 		 mi_switch(SW_INVOL, NULL);
351 
352 #else
353 	running_thread->td_flags |= TDF_NEEDRESCHED;
354 #endif
355 	return;
356 }
357 
358 #else /* SMP */
359 {
360 	struct thread *running_thread;
361 	int worst_pri;
362 	struct ksegrp *kg;
363 	cpumask_t cpumask,dontuse;
364 	struct pcpu *pc;
365 	struct pcpu *best_pcpu;
366 	struct thread *cputhread;
367 
368 #ifndef FULL_PREEMPTION
369 	int pri;
370 	pri = td->td_priority;
371 	if (!(pri >= PRI_MIN_ITHD && pri <= PRI_MAX_ITHD))
372 		return;
373 #endif
374 
375 	mtx_assert(&sched_lock, MA_OWNED);
376 
377 	running_thread = curthread;
378 
379 #if !defined(KSEG_PEEMPT_BEST_CPU)
380 	if (running_thread->td_ksegrp != td->td_ksegrp) {
381 #endif
382 		kg = td->td_ksegrp;
383 
384 		/* if someone is ahead of this thread, wait our turn */
385 		if (td != TAILQ_FIRST(&kg->kg_runq))
386 			return;
387 
388 		worst_pri = td->td_priority;
389 		best_pcpu = NULL;
390 		dontuse   = stopped_cpus | idle_cpus_mask;
391 
392 		/*
393 		 * Find a cpu with the worst priority that runs at thread from
394 		 * the same  ksegrp - if multiple exist give first the last run
395 		 * cpu and then the current cpu priority
396 		 */
397 
398 		SLIST_FOREACH(pc, &cpuhead, pc_allcpu) {
399 			cpumask   = pc->pc_cpumask;
400 			cputhread = pc->pc_curthread;
401 
402 			if ((cpumask & dontuse)  ||
403 			    cputhread->td_ksegrp != kg)
404 				continue;
405 
406 			if (cputhread->td_priority > worst_pri) {
407 				worst_pri = cputhread->td_priority;
408 				best_pcpu = pc;
409 				continue;
410 			}
411 
412 			if (cputhread->td_priority == worst_pri &&
413 			    best_pcpu != NULL &&
414 			    (td->td_lastcpu == pc->pc_cpuid ||
415 				(PCPU_GET(cpumask) == cpumask &&
416 				    td->td_lastcpu != best_pcpu->pc_cpuid)))
417 			    best_pcpu = pc;
418 		}
419 
420 		/* Check if we need to preempt someone */
421 		if (best_pcpu == NULL)
422 			return;
423 
424 		if (PCPU_GET(cpuid) != best_pcpu->pc_cpuid) {
425 			best_pcpu->pc_curthread->td_flags |= TDF_NEEDRESCHED;
426 			ipi_selected(best_pcpu->pc_cpumask, IPI_AST);
427 			return;
428 		}
429 #if !defined(KSEG_PEEMPT_BEST_CPU)
430 	}
431 #endif
432 
433 	if (td->td_priority > running_thread->td_priority)
434 		return;
435 #ifdef PREEMPTION
436 	if (running_thread->td_critnest > 1)
437 		running_thread->td_pflags |= TDP_OWEPREEMPT;
438 	 else
439 		 mi_switch(SW_INVOL, NULL);
440 
441 #else
442 	running_thread->td_flags |= TDF_NEEDRESCHED;
443 #endif
444 	return;
445 }
446 #endif /* !SMP */
447 
448 
449 int limitcount;
450 void
451 setrunqueue(struct thread *td, int flags)
452 {
453 	struct ksegrp *kg;
454 	struct thread *td2;
455 	struct thread *tda;
456 
457 	CTR3(KTR_RUNQ, "setrunqueue: td:%p kg:%p pid:%d",
458 	    td, td->td_ksegrp, td->td_proc->p_pid);
459 	mtx_assert(&sched_lock, MA_OWNED);
460 	KASSERT((td->td_inhibitors == 0),
461 			("setrunqueue: trying to run inhibitted thread"));
462 	KASSERT((TD_CAN_RUN(td) || TD_IS_RUNNING(td)),
463 	    ("setrunqueue: bad thread state"));
464 	TD_SET_RUNQ(td);
465 	kg = td->td_ksegrp;
466 	if ((td->td_proc->p_flag & P_HADTHREADS) == 0) {
467 		/*
468 		 * Common path optimisation: Only one of everything
469 		 * and the KSE is always already attached.
470 		 * Totally ignore the ksegrp run queue.
471 		 */
472 		if (kg->kg_avail_opennings != 1) {
473 			if (limitcount < 1) {
474 				limitcount++;
475 				printf("pid %d: corrected slot count (%d->1)\n",
476 				    td->td_proc->p_pid, kg->kg_avail_opennings);
477 
478 			}
479 			kg->kg_avail_opennings = 1;
480 		}
481 		sched_add(td, flags);
482 		return;
483 	}
484 
485 	/*
486 	 * If the concurrency has reduced, and we would go in the
487 	 * assigned section, then keep removing entries from the
488 	 * system run queue, until we are not in that section
489 	 * or there is room for us to be put in that section.
490 	 * What we MUST avoid is the case where there are threads of less
491 	 * priority than the new one scheduled, but it can not
492 	 * be scheduled itself. That would lead to a non contiguous set
493 	 * of scheduled threads, and everything would break.
494 	 */
495 	tda = kg->kg_last_assigned;
496 	while ((kg->kg_avail_opennings <= 0) &&
497 	    (tda && (tda->td_priority > td->td_priority))) {
498 		/*
499 		 * None free, but there is one we can commandeer.
500 		 */
501 		CTR2(KTR_RUNQ,
502 		    "setrunqueue: kg:%p: take slot from td: %p", kg, tda);
503 		sched_rem(tda);
504 		tda = kg->kg_last_assigned =
505 		    TAILQ_PREV(tda, threadqueue, td_runq);
506 	}
507 
508 	/*
509 	 * Add the thread to the ksegrp's run queue at
510 	 * the appropriate place.
511 	 */
512 	TAILQ_FOREACH(td2, &kg->kg_runq, td_runq) {
513 		if (td2->td_priority > td->td_priority) {
514 			TAILQ_INSERT_BEFORE(td2, td, td_runq);
515 			break;
516 		}
517 	}
518 	if (td2 == NULL) {
519 		/* We ran off the end of the TAILQ or it was empty. */
520 		TAILQ_INSERT_TAIL(&kg->kg_runq, td, td_runq);
521 	}
522 
523 	/*
524 	 * If we have a slot to use, then put the thread on the system
525 	 * run queue and if needed, readjust the last_assigned pointer.
526 	 * it may be that we need to schedule something anyhow
527 	 * even if the availabel slots are -ve so that
528 	 * all the items < last_assigned are scheduled.
529 	 */
530 	if (kg->kg_avail_opennings > 0) {
531 		if (tda == NULL) {
532 			/*
533 			 * No pre-existing last assigned so whoever is first
534 			 * gets the slot.. (maybe us)
535 			 */
536 			td2 = TAILQ_FIRST(&kg->kg_runq);
537 			kg->kg_last_assigned = td2;
538 		} else if (tda->td_priority > td->td_priority) {
539 			td2 = td;
540 		} else {
541 			/*
542 			 * We are past last_assigned, so
543 			 * give the next slot to whatever is next,
544 			 * which may or may not be us.
545 			 */
546 			td2 = TAILQ_NEXT(tda, td_runq);
547 			kg->kg_last_assigned = td2;
548 		}
549 		sched_add(td2, flags);
550 	} else {
551 		CTR3(KTR_RUNQ, "setrunqueue: held: td%p kg%p pid%d",
552 			td, td->td_ksegrp, td->td_proc->p_pid);
553 		if ((flags & SRQ_YIELDING) == 0)
554 			maybe_preempt_in_ksegrp(td);
555 	}
556 }
557 
558 /*
559  * Kernel thread preemption implementation.  Critical sections mark
560  * regions of code in which preemptions are not allowed.
561  */
562 void
563 critical_enter(void)
564 {
565 	struct thread *td;
566 
567 	td = curthread;
568 	if (td->td_critnest == 0)
569 		cpu_critical_enter(td);
570 	td->td_critnest++;
571 	CTR4(KTR_CRITICAL, "critical_enter by thread %p (%ld, %s) to %d", td,
572 	    (long)td->td_proc->p_pid, td->td_proc->p_comm, td->td_critnest);
573 }
574 
575 void
576 critical_exit(void)
577 {
578 	struct thread *td;
579 
580 	td = curthread;
581 	KASSERT(td->td_critnest != 0,
582 	    ("critical_exit: td_critnest == 0"));
583 	if (td->td_critnest == 1) {
584 		if (td->td_pflags & TDP_WAKEPROC0) {
585 			td->td_pflags &= ~TDP_WAKEPROC0;
586 			wakeup(&proc0);
587 		}
588 #ifdef PREEMPTION
589 		mtx_assert(&sched_lock, MA_NOTOWNED);
590 		if (td->td_pflags & TDP_OWEPREEMPT) {
591 			mtx_lock_spin(&sched_lock);
592 			mi_switch(SW_INVOL, NULL);
593 			mtx_unlock_spin(&sched_lock);
594 		}
595 #endif
596 		td->td_critnest = 0;
597 		cpu_critical_exit(td);
598 	} else {
599 		td->td_critnest--;
600 	}
601 	CTR4(KTR_CRITICAL, "critical_exit by thread %p (%ld, %s) to %d", td,
602 	    (long)td->td_proc->p_pid, td->td_proc->p_comm, td->td_critnest);
603 }
604 
605 /*
606  * This function is called when a thread is about to be put on run queue
607  * because it has been made runnable or its priority has been adjusted.  It
608  * determines if the new thread should be immediately preempted to.  If so,
609  * it switches to it and eventually returns true.  If not, it returns false
610  * so that the caller may place the thread on an appropriate run queue.
611  */
612 int
613 maybe_preempt(struct thread *td)
614 {
615 #ifdef PREEMPTION
616 	struct thread *ctd;
617 	int cpri, pri;
618 #endif
619 
620 	mtx_assert(&sched_lock, MA_OWNED);
621 #ifdef PREEMPTION
622 	/*
623 	 * The new thread should not preempt the current thread if any of the
624 	 * following conditions are true:
625 	 *
626 	 *  - The current thread has a higher (numerically lower) or
627 	 *    equivalent priority.  Note that this prevents curthread from
628 	 *    trying to preempt to itself.
629 	 *  - It is too early in the boot for context switches (cold is set).
630 	 *  - The current thread has an inhibitor set or is in the process of
631 	 *    exiting.  In this case, the current thread is about to switch
632 	 *    out anyways, so there's no point in preempting.  If we did,
633 	 *    the current thread would not be properly resumed as well, so
634 	 *    just avoid that whole landmine.
635 	 *  - If the new thread's priority is not a realtime priority and
636 	 *    the current thread's priority is not an idle priority and
637 	 *    FULL_PREEMPTION is disabled.
638 	 *
639 	 * If all of these conditions are false, but the current thread is in
640 	 * a nested critical section, then we have to defer the preemption
641 	 * until we exit the critical section.  Otherwise, switch immediately
642 	 * to the new thread.
643 	 */
644 	ctd = curthread;
645 	KASSERT ((ctd->td_kse != NULL && ctd->td_kse->ke_thread == ctd),
646 	  ("thread has no (or wrong) sched-private part."));
647 	KASSERT((td->td_inhibitors == 0),
648 			("maybe_preempt: trying to run inhibitted thread"));
649 	pri = td->td_priority;
650 	cpri = ctd->td_priority;
651 	if (pri >= cpri || cold /* || dumping */ || TD_IS_INHIBITED(ctd) ||
652 	    td->td_kse->ke_state != KES_THREAD)
653 		return (0);
654 #ifndef FULL_PREEMPTION
655 	if (!(pri >= PRI_MIN_ITHD && pri <= PRI_MAX_ITHD) &&
656 	    !(cpri >= PRI_MIN_IDLE))
657 		return (0);
658 #endif
659 	if (ctd->td_critnest > 1) {
660 		CTR1(KTR_PROC, "maybe_preempt: in critical section %d",
661 		    ctd->td_critnest);
662 		ctd->td_pflags |= TDP_OWEPREEMPT;
663 		return (0);
664 	}
665 
666 	/*
667 	 * Thread is runnable but not yet put on system run queue.
668 	 */
669 	MPASS(TD_ON_RUNQ(td));
670 	MPASS(td->td_sched->ke_state != KES_ONRUNQ);
671 	if (td->td_proc->p_flag & P_HADTHREADS) {
672 		/*
673 		 * If this is a threaded process we actually ARE on the
674 		 * ksegrp run queue so take it off that first.
675 		 * Also undo any damage done to the last_assigned pointer.
676 		 * XXX Fix setrunqueue so this isn't needed
677 		 */
678 		struct ksegrp *kg;
679 
680 		kg = td->td_ksegrp;
681 		if (kg->kg_last_assigned == td)
682 			kg->kg_last_assigned =
683 			    TAILQ_PREV(td, threadqueue, td_runq);
684 		TAILQ_REMOVE(&kg->kg_runq, td, td_runq);
685 	}
686 
687 	TD_SET_RUNNING(td);
688 	CTR3(KTR_PROC, "preempting to thread %p (pid %d, %s)\n", td,
689 	    td->td_proc->p_pid, td->td_proc->p_comm);
690 	mi_switch(SW_INVOL|SW_PREEMPT, td);
691 	return (1);
692 #else
693 	return (0);
694 #endif
695 }
696 
697 #if 0
698 #ifndef PREEMPTION
699 /* XXX: There should be a non-static version of this. */
700 static void
701 printf_caddr_t(void *data)
702 {
703 	printf("%s", (char *)data);
704 }
705 static char preempt_warning[] =
706     "WARNING: Kernel preemption is disabled, expect reduced performance.\n";
707 SYSINIT(preempt_warning, SI_SUB_COPYRIGHT, SI_ORDER_ANY, printf_caddr_t,
708     preempt_warning)
709 #endif
710 #endif
711 
712 /************************************************************************
713  * SYSTEM RUN QUEUE manipulations and tests				*
714  ************************************************************************/
715 /*
716  * Initialize a run structure.
717  */
718 void
719 runq_init(struct runq *rq)
720 {
721 	int i;
722 
723 	bzero(rq, sizeof *rq);
724 	for (i = 0; i < RQ_NQS; i++)
725 		TAILQ_INIT(&rq->rq_queues[i]);
726 }
727 
728 /*
729  * Clear the status bit of the queue corresponding to priority level pri,
730  * indicating that it is empty.
731  */
732 static __inline void
733 runq_clrbit(struct runq *rq, int pri)
734 {
735 	struct rqbits *rqb;
736 
737 	rqb = &rq->rq_status;
738 	CTR4(KTR_RUNQ, "runq_clrbit: bits=%#x %#x bit=%#x word=%d",
739 	    rqb->rqb_bits[RQB_WORD(pri)],
740 	    rqb->rqb_bits[RQB_WORD(pri)] & ~RQB_BIT(pri),
741 	    RQB_BIT(pri), RQB_WORD(pri));
742 	rqb->rqb_bits[RQB_WORD(pri)] &= ~RQB_BIT(pri);
743 }
744 
745 /*
746  * Find the index of the first non-empty run queue.  This is done by
747  * scanning the status bits, a set bit indicates a non-empty queue.
748  */
749 static __inline int
750 runq_findbit(struct runq *rq)
751 {
752 	struct rqbits *rqb;
753 	int pri;
754 	int i;
755 
756 	rqb = &rq->rq_status;
757 	for (i = 0; i < RQB_LEN; i++)
758 		if (rqb->rqb_bits[i]) {
759 			pri = RQB_FFS(rqb->rqb_bits[i]) + (i << RQB_L2BPW);
760 			CTR3(KTR_RUNQ, "runq_findbit: bits=%#x i=%d pri=%d",
761 			    rqb->rqb_bits[i], i, pri);
762 			return (pri);
763 		}
764 
765 	return (-1);
766 }
767 
768 /*
769  * Set the status bit of the queue corresponding to priority level pri,
770  * indicating that it is non-empty.
771  */
772 static __inline void
773 runq_setbit(struct runq *rq, int pri)
774 {
775 	struct rqbits *rqb;
776 
777 	rqb = &rq->rq_status;
778 	CTR4(KTR_RUNQ, "runq_setbit: bits=%#x %#x bit=%#x word=%d",
779 	    rqb->rqb_bits[RQB_WORD(pri)],
780 	    rqb->rqb_bits[RQB_WORD(pri)] | RQB_BIT(pri),
781 	    RQB_BIT(pri), RQB_WORD(pri));
782 	rqb->rqb_bits[RQB_WORD(pri)] |= RQB_BIT(pri);
783 }
784 
785 /*
786  * Add the KSE to the queue specified by its priority, and set the
787  * corresponding status bit.
788  */
789 void
790 runq_add(struct runq *rq, struct kse *ke, int flags)
791 {
792 	struct rqhead *rqh;
793 	int pri;
794 
795 	pri = ke->ke_thread->td_priority / RQ_PPQ;
796 	ke->ke_rqindex = pri;
797 	runq_setbit(rq, pri);
798 	rqh = &rq->rq_queues[pri];
799 	CTR5(KTR_RUNQ, "runq_add: td=%p ke=%p pri=%d %d rqh=%p",
800 	    ke->ke_thread, ke, ke->ke_thread->td_priority, pri, rqh);
801 	if (flags & SRQ_PREEMPTED) {
802 		TAILQ_INSERT_HEAD(rqh, ke, ke_procq);
803 	} else {
804 		TAILQ_INSERT_TAIL(rqh, ke, ke_procq);
805 	}
806 }
807 
808 /*
809  * Return true if there are runnable processes of any priority on the run
810  * queue, false otherwise.  Has no side effects, does not modify the run
811  * queue structure.
812  */
813 int
814 runq_check(struct runq *rq)
815 {
816 	struct rqbits *rqb;
817 	int i;
818 
819 	rqb = &rq->rq_status;
820 	for (i = 0; i < RQB_LEN; i++)
821 		if (rqb->rqb_bits[i]) {
822 			CTR2(KTR_RUNQ, "runq_check: bits=%#x i=%d",
823 			    rqb->rqb_bits[i], i);
824 			return (1);
825 		}
826 	CTR0(KTR_RUNQ, "runq_check: empty");
827 
828 	return (0);
829 }
830 
831 #if defined(SMP) && defined(SCHED_4BSD)
832 int runq_fuzz = 1;
833 SYSCTL_INT(_kern_sched, OID_AUTO, runq_fuzz, CTLFLAG_RW, &runq_fuzz, 0, "");
834 #endif
835 
836 /*
837  * Find the highest priority process on the run queue.
838  */
839 struct kse *
840 runq_choose(struct runq *rq)
841 {
842 	struct rqhead *rqh;
843 	struct kse *ke;
844 	int pri;
845 
846 	mtx_assert(&sched_lock, MA_OWNED);
847 	while ((pri = runq_findbit(rq)) != -1) {
848 		rqh = &rq->rq_queues[pri];
849 #if defined(SMP) && defined(SCHED_4BSD)
850 		/* fuzz == 1 is normal.. 0 or less are ignored */
851 		if (runq_fuzz > 1) {
852 			/*
853 			 * In the first couple of entries, check if
854 			 * there is one for our CPU as a preference.
855 			 */
856 			int count = runq_fuzz;
857 			int cpu = PCPU_GET(cpuid);
858 			struct kse *ke2;
859 			ke2 = ke = TAILQ_FIRST(rqh);
860 
861 			while (count-- && ke2) {
862 				if (ke->ke_thread->td_lastcpu == cpu) {
863 					ke = ke2;
864 					break;
865 				}
866 				ke2 = TAILQ_NEXT(ke2, ke_procq);
867 			}
868 		} else
869 #endif
870 			ke = TAILQ_FIRST(rqh);
871 		KASSERT(ke != NULL, ("runq_choose: no proc on busy queue"));
872 		CTR3(KTR_RUNQ,
873 		    "runq_choose: pri=%d kse=%p rqh=%p", pri, ke, rqh);
874 		return (ke);
875 	}
876 	CTR1(KTR_RUNQ, "runq_choose: idleproc pri=%d", pri);
877 
878 	return (NULL);
879 }
880 
881 /*
882  * Remove the KSE from the queue specified by its priority, and clear the
883  * corresponding status bit if the queue becomes empty.
884  * Caller must set ke->ke_state afterwards.
885  */
886 void
887 runq_remove(struct runq *rq, struct kse *ke)
888 {
889 	struct rqhead *rqh;
890 	int pri;
891 
892 	KASSERT(ke->ke_proc->p_sflag & PS_INMEM,
893 		("runq_remove: process swapped out"));
894 	pri = ke->ke_rqindex;
895 	rqh = &rq->rq_queues[pri];
896 	CTR5(KTR_RUNQ, "runq_remove: td=%p, ke=%p pri=%d %d rqh=%p",
897 	    ke->ke_thread, ke, ke->ke_thread->td_priority, pri, rqh);
898 	KASSERT(ke != NULL, ("runq_remove: no proc on busy queue"));
899 	TAILQ_REMOVE(rqh, ke, ke_procq);
900 	if (TAILQ_EMPTY(rqh)) {
901 		CTR0(KTR_RUNQ, "runq_remove: empty");
902 		runq_clrbit(rq, pri);
903 	}
904 }
905 
906 /****** functions that are temporarily here ***********/
907 #include <vm/uma.h>
908 extern struct mtx kse_zombie_lock;
909 
910 /*
911  *  Allocate scheduler specific per-process resources.
912  * The thread and ksegrp have already been linked in.
913  * In this case just set the default concurrency value.
914  *
915  * Called from:
916  *  proc_init() (UMA init method)
917  */
918 void
919 sched_newproc(struct proc *p, struct ksegrp *kg, struct thread *td)
920 {
921 
922 	/* This can go in sched_fork */
923 	sched_init_concurrency(kg);
924 }
925 
926 /*
927  * thread is being either created or recycled.
928  * Fix up the per-scheduler resources associated with it.
929  * Called from:
930  *  sched_fork_thread()
931  *  thread_dtor()  (*may go away)
932  *  thread_init()  (*may go away)
933  */
934 void
935 sched_newthread(struct thread *td)
936 {
937 	struct td_sched *ke;
938 
939 	ke = (struct td_sched *) (td + 1);
940 	bzero(ke, sizeof(*ke));
941 	td->td_sched     = ke;
942 	ke->ke_thread	= td;
943 	ke->ke_state	= KES_THREAD;
944 }
945 
946 /*
947  * Set up an initial concurrency of 1
948  * and set the given thread (if given) to be using that
949  * concurrency slot.
950  * May be used "offline"..before the ksegrp is attached to the world
951  * and thus wouldn't need schedlock in that case.
952  * Called from:
953  *  thr_create()
954  *  proc_init() (UMA) via sched_newproc()
955  */
956 void
957 sched_init_concurrency(struct ksegrp *kg)
958 {
959 
960 	CTR1(KTR_RUNQ,"kg %p init slots and concurrency to 1", kg);
961 	kg->kg_concurrency = 1;
962 	kg->kg_avail_opennings = 1;
963 }
964 
965 /*
966  * Change the concurrency of an existing ksegrp to N
967  * Called from:
968  *  kse_create()
969  *  kse_exit()
970  *  thread_exit()
971  *  thread_single()
972  */
973 void
974 sched_set_concurrency(struct ksegrp *kg, int concurrency)
975 {
976 
977 	CTR4(KTR_RUNQ,"kg %p set concurrency to %d, slots %d -> %d",
978 	    kg,
979 	    concurrency,
980 	    kg->kg_avail_opennings,
981 	    kg->kg_avail_opennings + (concurrency - kg->kg_concurrency));
982 	kg->kg_avail_opennings += (concurrency - kg->kg_concurrency);
983 	kg->kg_concurrency = concurrency;
984 }
985 
986 /*
987  * Called from thread_exit() for all exiting thread
988  *
989  * Not to be confused with sched_exit_thread()
990  * that is only called from thread_exit() for threads exiting
991  * without the rest of the process exiting because it is also called from
992  * sched_exit() and we wouldn't want to call it twice.
993  * XXX This can probably be fixed.
994  */
995 void
996 sched_thread_exit(struct thread *td)
997 {
998 
999 	SLOT_RELEASE(td->td_ksegrp);
1000 	slot_fill(td->td_ksegrp);
1001 }
1002 
1003 #endif /* KERN_SWITCH_INCLUDE */
1004