xref: /netbsd-src/sys/kern/kern_synch.c (revision da9817918ec7e88db2912a2882967c7570a83f47)
1 /*	$NetBSD: kern_synch.c,v 1.264 2009/04/16 21:19:23 ad Exp $	*/
2 
3 /*-
4  * Copyright (c) 1999, 2000, 2004, 2006, 2007, 2008, 2009
5  *    The NetBSD Foundation, Inc.
6  * All rights reserved.
7  *
8  * This code is derived from software contributed to The NetBSD Foundation
9  * by Jason R. Thorpe of the Numerical Aerospace Simulation Facility,
10  * NASA Ames Research Center, by Charles M. Hannum, Andrew Doran and
11  * Daniel Sieger.
12  *
13  * Redistribution and use in source and binary forms, with or without
14  * modification, are permitted provided that the following conditions
15  * are met:
16  * 1. Redistributions of source code must retain the above copyright
17  *    notice, this list of conditions and the following disclaimer.
18  * 2. Redistributions in binary form must reproduce the above copyright
19  *    notice, this list of conditions and the following disclaimer in the
20  *    documentation and/or other materials provided with the distribution.
21  *
22  * THIS SOFTWARE IS PROVIDED BY THE NETBSD FOUNDATION, INC. AND CONTRIBUTORS
23  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED
24  * TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
25  * PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE FOUNDATION OR CONTRIBUTORS
26  * BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
27  * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
28  * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
29  * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
30  * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
31  * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
32  * POSSIBILITY OF SUCH DAMAGE.
33  */
34 
35 /*-
36  * Copyright (c) 1982, 1986, 1990, 1991, 1993
37  *	The Regents of the University of California.  All rights reserved.
38  * (c) UNIX System Laboratories, Inc.
39  * All or some portions of this file are derived from material licensed
40  * to the University of California by American Telephone and Telegraph
41  * Co. or Unix System Laboratories, Inc. and are reproduced herein with
42  * the permission of UNIX System Laboratories, Inc.
43  *
44  * Redistribution and use in source and binary forms, with or without
45  * modification, are permitted provided that the following conditions
46  * are met:
47  * 1. Redistributions of source code must retain the above copyright
48  *    notice, this list of conditions and the following disclaimer.
49  * 2. Redistributions in binary form must reproduce the above copyright
50  *    notice, this list of conditions and the following disclaimer in the
51  *    documentation and/or other materials provided with the distribution.
52  * 3. Neither the name of the University nor the names of its contributors
53  *    may be used to endorse or promote products derived from this software
54  *    without specific prior written permission.
55  *
56  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
57  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
58  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
59  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
60  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
61  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
62  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
63  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
64  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
65  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
66  * SUCH DAMAGE.
67  *
68  *	@(#)kern_synch.c	8.9 (Berkeley) 5/19/95
69  */
70 
71 #include <sys/cdefs.h>
72 __KERNEL_RCSID(0, "$NetBSD: kern_synch.c,v 1.264 2009/04/16 21:19:23 ad Exp $");
73 
74 #include "opt_kstack.h"
75 #include "opt_perfctrs.h"
76 #include "opt_sa.h"
77 
78 #define	__MUTEX_PRIVATE
79 
80 #include <sys/param.h>
81 #include <sys/systm.h>
82 #include <sys/proc.h>
83 #include <sys/kernel.h>
84 #if defined(PERFCTRS)
85 #include <sys/pmc.h>
86 #endif
87 #include <sys/cpu.h>
88 #include <sys/resourcevar.h>
89 #include <sys/sched.h>
90 #include <sys/sa.h>
91 #include <sys/savar.h>
92 #include <sys/syscall_stats.h>
93 #include <sys/sleepq.h>
94 #include <sys/lockdebug.h>
95 #include <sys/evcnt.h>
96 #include <sys/intr.h>
97 #include <sys/lwpctl.h>
98 #include <sys/atomic.h>
99 #include <sys/simplelock.h>
100 
101 #include <uvm/uvm_extern.h>
102 
103 #include <dev/lockstat.h>
104 
105 static u_int	sched_unsleep(struct lwp *, bool);
106 static void	sched_changepri(struct lwp *, pri_t);
107 static void	sched_lendpri(struct lwp *, pri_t);
108 static void	resched_cpu(struct lwp *);
109 
110 syncobj_t sleep_syncobj = {
111 	SOBJ_SLEEPQ_SORTED,
112 	sleepq_unsleep,
113 	sleepq_changepri,
114 	sleepq_lendpri,
115 	syncobj_noowner,
116 };
117 
118 syncobj_t sched_syncobj = {
119 	SOBJ_SLEEPQ_SORTED,
120 	sched_unsleep,
121 	sched_changepri,
122 	sched_lendpri,
123 	syncobj_noowner,
124 };
125 
126 callout_t 	sched_pstats_ch;
127 unsigned	sched_pstats_ticks;
128 kcondvar_t	lbolt;			/* once a second sleep address */
129 
130 /* Preemption event counters */
131 static struct evcnt kpreempt_ev_crit;
132 static struct evcnt kpreempt_ev_klock;
133 static struct evcnt kpreempt_ev_immed;
134 
135 /*
136  * During autoconfiguration or after a panic, a sleep will simply lower the
137  * priority briefly to allow interrupts, then return.  The priority to be
138  * used (safepri) is machine-dependent, thus this value is initialized and
139  * maintained in the machine-dependent layers.  This priority will typically
140  * be 0, or the lowest priority that is safe for use on the interrupt stack;
141  * it can be made higher to block network software interrupts after panics.
142  */
143 int	safepri;
144 
145 void
146 sched_init(void)
147 {
148 
149 	cv_init(&lbolt, "lbolt");
150 	callout_init(&sched_pstats_ch, CALLOUT_MPSAFE);
151 	callout_setfunc(&sched_pstats_ch, sched_pstats, NULL);
152 
153 	evcnt_attach_dynamic(&kpreempt_ev_crit, EVCNT_TYPE_MISC, NULL,
154 	   "kpreempt", "defer: critical section");
155 	evcnt_attach_dynamic(&kpreempt_ev_klock, EVCNT_TYPE_MISC, NULL,
156 	   "kpreempt", "defer: kernel_lock");
157 	evcnt_attach_dynamic(&kpreempt_ev_immed, EVCNT_TYPE_MISC, NULL,
158 	   "kpreempt", "immediate");
159 
160 	sched_pstats(NULL);
161 }
162 
163 /*
164  * OBSOLETE INTERFACE
165  *
166  * General sleep call.  Suspends the current LWP until a wakeup is
167  * performed on the specified identifier.  The LWP will then be made
168  * runnable with the specified priority.  Sleeps at most timo/hz seconds (0
169  * means no timeout).  If pri includes PCATCH flag, signals are checked
170  * before and after sleeping, else signals are not checked.  Returns 0 if
171  * awakened, EWOULDBLOCK if the timeout expires.  If PCATCH is set and a
172  * signal needs to be delivered, ERESTART is returned if the current system
173  * call should be restarted if possible, and EINTR is returned if the system
174  * call should be interrupted by the signal (return EINTR).
175  *
176  * The interlock is held until we are on a sleep queue. The interlock will
177  * be locked before returning back to the caller unless the PNORELOCK flag
178  * is specified, in which case the interlock will always be unlocked upon
179  * return.
180  */
181 int
182 ltsleep(wchan_t ident, pri_t priority, const char *wmesg, int timo,
183 	volatile struct simplelock *interlock)
184 {
185 	struct lwp *l = curlwp;
186 	sleepq_t *sq;
187 	kmutex_t *mp;
188 	int error;
189 
190 	KASSERT((l->l_pflag & LP_INTR) == 0);
191 
192 	if (sleepq_dontsleep(l)) {
193 		(void)sleepq_abort(NULL, 0);
194 		if ((priority & PNORELOCK) != 0)
195 			simple_unlock(interlock);
196 		return 0;
197 	}
198 
199 	l->l_kpriority = true;
200 	sq = sleeptab_lookup(&sleeptab, ident, &mp);
201 	sleepq_enter(sq, l, mp);
202 	sleepq_enqueue(sq, ident, wmesg, &sleep_syncobj);
203 
204 	if (interlock != NULL) {
205 		KASSERT(simple_lock_held(interlock));
206 		simple_unlock(interlock);
207 	}
208 
209 	error = sleepq_block(timo, priority & PCATCH);
210 
211 	if (interlock != NULL && (priority & PNORELOCK) == 0)
212 		simple_lock(interlock);
213 
214 	return error;
215 }
216 
217 int
218 mtsleep(wchan_t ident, pri_t priority, const char *wmesg, int timo,
219 	kmutex_t *mtx)
220 {
221 	struct lwp *l = curlwp;
222 	sleepq_t *sq;
223 	kmutex_t *mp;
224 	int error;
225 
226 	KASSERT((l->l_pflag & LP_INTR) == 0);
227 
228 	if (sleepq_dontsleep(l)) {
229 		(void)sleepq_abort(mtx, (priority & PNORELOCK) != 0);
230 		return 0;
231 	}
232 
233 	l->l_kpriority = true;
234 	sq = sleeptab_lookup(&sleeptab, ident, &mp);
235 	sleepq_enter(sq, l, mp);
236 	sleepq_enqueue(sq, ident, wmesg, &sleep_syncobj);
237 	mutex_exit(mtx);
238 	error = sleepq_block(timo, priority & PCATCH);
239 
240 	if ((priority & PNORELOCK) == 0)
241 		mutex_enter(mtx);
242 
243 	return error;
244 }
245 
246 /*
247  * General sleep call for situations where a wake-up is not expected.
248  */
249 int
250 kpause(const char *wmesg, bool intr, int timo, kmutex_t *mtx)
251 {
252 	struct lwp *l = curlwp;
253 	kmutex_t *mp;
254 	sleepq_t *sq;
255 	int error;
256 
257 	if (sleepq_dontsleep(l))
258 		return sleepq_abort(NULL, 0);
259 
260 	if (mtx != NULL)
261 		mutex_exit(mtx);
262 	l->l_kpriority = true;
263 	sq = sleeptab_lookup(&sleeptab, l, &mp);
264 	sleepq_enter(sq, l, mp);
265 	sleepq_enqueue(sq, l, wmesg, &sleep_syncobj);
266 	error = sleepq_block(timo, intr);
267 	if (mtx != NULL)
268 		mutex_enter(mtx);
269 
270 	return error;
271 }
272 
273 #ifdef KERN_SA
274 /*
275  * sa_awaken:
276  *
277  *	We believe this lwp is an SA lwp. If it's yielding,
278  * let it know it needs to wake up.
279  *
280  *	We are called and exit with the lwp locked. We are
281  * called in the middle of wakeup operations, so we need
282  * to not touch the locks at all.
283  */
284 void
285 sa_awaken(struct lwp *l)
286 {
287 	/* LOCK_ASSERT(lwp_locked(l, NULL)); */
288 
289 	if (l == l->l_savp->savp_lwp && l->l_flag & LW_SA_YIELD)
290 		l->l_flag &= ~LW_SA_IDLE;
291 }
292 #endif /* KERN_SA */
293 
294 /*
295  * OBSOLETE INTERFACE
296  *
297  * Make all LWPs sleeping on the specified identifier runnable.
298  */
299 void
300 wakeup(wchan_t ident)
301 {
302 	sleepq_t *sq;
303 	kmutex_t *mp;
304 
305 	if (__predict_false(cold))
306 		return;
307 
308 	sq = sleeptab_lookup(&sleeptab, ident, &mp);
309 	sleepq_wake(sq, ident, (u_int)-1, mp);
310 }
311 
312 /*
313  * OBSOLETE INTERFACE
314  *
315  * Make the highest priority LWP first in line on the specified
316  * identifier runnable.
317  */
318 void
319 wakeup_one(wchan_t ident)
320 {
321 	sleepq_t *sq;
322 	kmutex_t *mp;
323 
324 	if (__predict_false(cold))
325 		return;
326 
327 	sq = sleeptab_lookup(&sleeptab, ident, &mp);
328 	sleepq_wake(sq, ident, 1, mp);
329 }
330 
331 
332 /*
333  * General yield call.  Puts the current LWP back on its run queue and
334  * performs a voluntary context switch.  Should only be called when the
335  * current LWP explicitly requests it (eg sched_yield(2)).
336  */
337 void
338 yield(void)
339 {
340 	struct lwp *l = curlwp;
341 
342 	KERNEL_UNLOCK_ALL(l, &l->l_biglocks);
343 	lwp_lock(l);
344 	KASSERT(lwp_locked(l, l->l_cpu->ci_schedstate.spc_lwplock));
345 	KASSERT(l->l_stat == LSONPROC);
346 	l->l_kpriority = false;
347 	(void)mi_switch(l);
348 	KERNEL_LOCK(l->l_biglocks, l);
349 }
350 
351 /*
352  * General preemption call.  Puts the current LWP back on its run queue
353  * and performs an involuntary context switch.
354  */
355 void
356 preempt(void)
357 {
358 	struct lwp *l = curlwp;
359 
360 	KERNEL_UNLOCK_ALL(l, &l->l_biglocks);
361 	lwp_lock(l);
362 	KASSERT(lwp_locked(l, l->l_cpu->ci_schedstate.spc_lwplock));
363 	KASSERT(l->l_stat == LSONPROC);
364 	l->l_kpriority = false;
365 	l->l_nivcsw++;
366 	(void)mi_switch(l);
367 	KERNEL_LOCK(l->l_biglocks, l);
368 }
369 
370 /*
371  * Handle a request made by another agent to preempt the current LWP
372  * in-kernel.  Usually called when l_dopreempt may be non-zero.
373  *
374  * Character addresses for lockstat only.
375  */
376 static char	in_critical_section;
377 static char	kernel_lock_held;
378 static char	is_softint;
379 static char	cpu_kpreempt_enter_fail;
380 
381 bool
382 kpreempt(uintptr_t where)
383 {
384 	uintptr_t failed;
385 	lwp_t *l;
386 	int s, dop, lsflag;
387 
388 	l = curlwp;
389 	failed = 0;
390 	while ((dop = l->l_dopreempt) != 0) {
391 		if (l->l_stat != LSONPROC) {
392 			/*
393 			 * About to block (or die), let it happen.
394 			 * Doesn't really count as "preemption has
395 			 * been blocked", since we're going to
396 			 * context switch.
397 			 */
398 			l->l_dopreempt = 0;
399 			return true;
400 		}
401 		if (__predict_false((l->l_flag & LW_IDLE) != 0)) {
402 			/* Can't preempt idle loop, don't count as failure. */
403 			l->l_dopreempt = 0;
404 			return true;
405 		}
406 		if (__predict_false(l->l_nopreempt != 0)) {
407 			/* LWP holds preemption disabled, explicitly. */
408 			if ((dop & DOPREEMPT_COUNTED) == 0) {
409 				kpreempt_ev_crit.ev_count++;
410 			}
411 			failed = (uintptr_t)&in_critical_section;
412 			break;
413 		}
414 		if (__predict_false((l->l_pflag & LP_INTR) != 0)) {
415 			/* Can't preempt soft interrupts yet. */
416 			l->l_dopreempt = 0;
417 			failed = (uintptr_t)&is_softint;
418 			break;
419 		}
420 		s = splsched();
421 		if (__predict_false(l->l_blcnt != 0 ||
422 		    curcpu()->ci_biglock_wanted != NULL)) {
423 			/* Hold or want kernel_lock, code is not MT safe. */
424 			splx(s);
425 			if ((dop & DOPREEMPT_COUNTED) == 0) {
426 				kpreempt_ev_klock.ev_count++;
427 			}
428 			failed = (uintptr_t)&kernel_lock_held;
429 			break;
430 		}
431 		if (__predict_false(!cpu_kpreempt_enter(where, s))) {
432 			/*
433 			 * It may be that the IPL is too high.
434 			 * kpreempt_enter() can schedule an
435 			 * interrupt to retry later.
436 			 */
437 			splx(s);
438 			failed = (uintptr_t)&cpu_kpreempt_enter_fail;
439 			break;
440 		}
441 		/* Do it! */
442 		if (__predict_true((dop & DOPREEMPT_COUNTED) == 0)) {
443 			kpreempt_ev_immed.ev_count++;
444 		}
445 		lwp_lock(l);
446 		mi_switch(l);
447 		l->l_nopreempt++;
448 		splx(s);
449 
450 		/* Take care of any MD cleanup. */
451 		cpu_kpreempt_exit(where);
452 		l->l_nopreempt--;
453 	}
454 
455 	if (__predict_true(!failed)) {
456 		return false;
457 	}
458 
459 	/* Record preemption failure for reporting via lockstat. */
460 	atomic_or_uint(&l->l_dopreempt, DOPREEMPT_COUNTED);
461 	lsflag = 0;
462 	LOCKSTAT_ENTER(lsflag);
463 	if (__predict_false(lsflag)) {
464 		if (where == 0) {
465 			where = (uintptr_t)__builtin_return_address(0);
466 		}
467 		/* Preemption is on, might recurse, so make it atomic. */
468 		if (atomic_cas_ptr_ni((void *)&l->l_pfailaddr, NULL,
469 		    (void *)where) == NULL) {
470 			LOCKSTAT_START_TIMER(lsflag, l->l_pfailtime);
471 			l->l_pfaillock = failed;
472 		}
473 	}
474 	LOCKSTAT_EXIT(lsflag);
475 	return true;
476 }
477 
478 /*
479  * Return true if preemption is explicitly disabled.
480  */
481 bool
482 kpreempt_disabled(void)
483 {
484 	const lwp_t *l = curlwp;
485 
486 	return l->l_nopreempt != 0 || l->l_stat == LSZOMB ||
487 	    (l->l_flag & LW_IDLE) != 0 || cpu_kpreempt_disabled();
488 }
489 
490 /*
491  * Disable kernel preemption.
492  */
493 void
494 kpreempt_disable(void)
495 {
496 
497 	KPREEMPT_DISABLE(curlwp);
498 }
499 
500 /*
501  * Reenable kernel preemption.
502  */
503 void
504 kpreempt_enable(void)
505 {
506 
507 	KPREEMPT_ENABLE(curlwp);
508 }
509 
510 /*
511  * Compute the amount of time during which the current lwp was running.
512  *
513  * - update l_rtime unless it's an idle lwp.
514  */
515 
516 void
517 updatertime(lwp_t *l, const struct bintime *now)
518 {
519 
520 	if (__predict_false(l->l_flag & LW_IDLE))
521 		return;
522 
523 	/* rtime += now - stime */
524 	bintime_add(&l->l_rtime, now);
525 	bintime_sub(&l->l_rtime, &l->l_stime);
526 }
527 
528 /*
529  * Select next LWP from the current CPU to run..
530  */
531 static inline lwp_t *
532 nextlwp(struct cpu_info *ci, struct schedstate_percpu *spc)
533 {
534 	lwp_t *newl;
535 
536 	/*
537 	 * Let sched_nextlwp() select the LWP to run the CPU next.
538 	 * If no LWP is runnable, select the idle LWP.
539 	 *
540 	 * Note that spc_lwplock might not necessary be held, and
541 	 * new thread would be unlocked after setting the LWP-lock.
542 	 */
543 	newl = sched_nextlwp();
544 	if (newl != NULL) {
545 		sched_dequeue(newl);
546 		KASSERT(lwp_locked(newl, spc->spc_mutex));
547 		newl->l_stat = LSONPROC;
548 		newl->l_cpu = ci;
549 		newl->l_pflag |= LP_RUNNING;
550 		lwp_setlock(newl, spc->spc_lwplock);
551 	} else {
552 		newl = ci->ci_data.cpu_idlelwp;
553 		newl->l_stat = LSONPROC;
554 		newl->l_pflag |= LP_RUNNING;
555 	}
556 
557 	/*
558 	 * Only clear want_resched if there are no pending (slow)
559 	 * software interrupts.
560 	 */
561 	ci->ci_want_resched = ci->ci_data.cpu_softints;
562 	spc->spc_flags &= ~SPCF_SWITCHCLEAR;
563 	spc->spc_curpriority = lwp_eprio(newl);
564 
565 	return newl;
566 }
567 
568 /*
569  * The machine independent parts of context switch.
570  *
571  * Returns 1 if another LWP was actually run.
572  */
573 int
574 mi_switch(lwp_t *l)
575 {
576 	struct cpu_info *ci;
577 	struct schedstate_percpu *spc;
578 	struct lwp *newl;
579 	int retval, oldspl;
580 	struct bintime bt;
581 	bool returning;
582 
583 	KASSERT(lwp_locked(l, NULL));
584 	KASSERT(kpreempt_disabled());
585 	LOCKDEBUG_BARRIER(l->l_mutex, 1);
586 
587 	kstack_check_magic(l);
588 
589 	binuptime(&bt);
590 
591 	KASSERT(l->l_cpu == curcpu());
592 	ci = l->l_cpu;
593 	spc = &ci->ci_schedstate;
594 	returning = false;
595 	newl = NULL;
596 
597 	/*
598 	 * If we have been asked to switch to a specific LWP, then there
599 	 * is no need to inspect the run queues.  If a soft interrupt is
600 	 * blocking, then return to the interrupted thread without adjusting
601 	 * VM context or its start time: neither have been changed in order
602 	 * to take the interrupt.
603 	 */
604 	if (l->l_switchto != NULL) {
605 		if ((l->l_pflag & LP_INTR) != 0) {
606 			returning = true;
607 			softint_block(l);
608 			if ((l->l_pflag & LP_TIMEINTR) != 0)
609 				updatertime(l, &bt);
610 		}
611 		newl = l->l_switchto;
612 		l->l_switchto = NULL;
613 	}
614 #ifndef __HAVE_FAST_SOFTINTS
615 	else if (ci->ci_data.cpu_softints != 0) {
616 		/* There are pending soft interrupts, so pick one. */
617 		newl = softint_picklwp();
618 		newl->l_stat = LSONPROC;
619 		newl->l_pflag |= LP_RUNNING;
620 	}
621 #endif	/* !__HAVE_FAST_SOFTINTS */
622 
623 	/* Count time spent in current system call */
624 	if (!returning) {
625 		SYSCALL_TIME_SLEEP(l);
626 
627 		/*
628 		 * XXXSMP If we are using h/w performance counters,
629 		 * save context.
630 		 */
631 #if PERFCTRS
632 		if (PMC_ENABLED(l->l_proc)) {
633 			pmc_save_context(l->l_proc);
634 		}
635 #endif
636 		updatertime(l, &bt);
637 	}
638 
639 	/* Lock the runqueue */
640 	KASSERT(l->l_stat != LSRUN);
641 	mutex_spin_enter(spc->spc_mutex);
642 
643 	/*
644 	 * If on the CPU and we have gotten this far, then we must yield.
645 	 */
646 	if (l->l_stat == LSONPROC && l != newl) {
647 		KASSERT(lwp_locked(l, spc->spc_lwplock));
648 		if ((l->l_flag & LW_IDLE) == 0) {
649 			l->l_stat = LSRUN;
650 			lwp_setlock(l, spc->spc_mutex);
651 			sched_enqueue(l, true);
652 			/* Handle migration case */
653 			KASSERT(spc->spc_migrating == NULL);
654 			if (l->l_target_cpu !=  NULL) {
655 				spc->spc_migrating = l;
656 			}
657 		} else
658 			l->l_stat = LSIDL;
659 	}
660 
661 	/* Pick new LWP to run. */
662 	if (newl == NULL) {
663 		newl = nextlwp(ci, spc);
664 	}
665 
666 	/* Items that must be updated with the CPU locked. */
667 	if (!returning) {
668 		/* Update the new LWP's start time. */
669 		newl->l_stime = bt;
670 
671 		/*
672 		 * ci_curlwp changes when a fast soft interrupt occurs.
673 		 * We use cpu_onproc to keep track of which kernel or
674 		 * user thread is running 'underneath' the software
675 		 * interrupt.  This is important for time accounting,
676 		 * itimers and forcing user threads to preempt (aston).
677 		 */
678 		ci->ci_data.cpu_onproc = newl;
679 	}
680 
681 	/*
682 	 * Preemption related tasks.  Must be done with the current
683 	 * CPU locked.
684 	 */
685 	cpu_did_resched(l);
686 	l->l_dopreempt = 0;
687 	if (__predict_false(l->l_pfailaddr != 0)) {
688 		LOCKSTAT_FLAG(lsflag);
689 		LOCKSTAT_ENTER(lsflag);
690 		LOCKSTAT_STOP_TIMER(lsflag, l->l_pfailtime);
691 		LOCKSTAT_EVENT_RA(lsflag, l->l_pfaillock, LB_NOPREEMPT|LB_SPIN,
692 		    1, l->l_pfailtime, l->l_pfailaddr);
693 		LOCKSTAT_EXIT(lsflag);
694 		l->l_pfailtime = 0;
695 		l->l_pfaillock = 0;
696 		l->l_pfailaddr = 0;
697 	}
698 
699 	if (l != newl) {
700 		struct lwp *prevlwp;
701 
702 		/* Release all locks, but leave the current LWP locked */
703 		if (l->l_mutex == spc->spc_mutex) {
704 			/*
705 			 * Drop spc_lwplock, if the current LWP has been moved
706 			 * to the run queue (it is now locked by spc_mutex).
707 			 */
708 			mutex_spin_exit(spc->spc_lwplock);
709 		} else {
710 			/*
711 			 * Otherwise, drop the spc_mutex, we are done with the
712 			 * run queues.
713 			 */
714 			mutex_spin_exit(spc->spc_mutex);
715 		}
716 
717 		/*
718 		 * Mark that context switch is going to be performed
719 		 * for this LWP, to protect it from being switched
720 		 * to on another CPU.
721 		 */
722 		KASSERT(l->l_ctxswtch == 0);
723 		l->l_ctxswtch = 1;
724 		l->l_ncsw++;
725 		l->l_pflag &= ~LP_RUNNING;
726 
727 		/*
728 		 * Increase the count of spin-mutexes before the release
729 		 * of the last lock - we must remain at IPL_SCHED during
730 		 * the context switch.
731 		 */
732 		oldspl = MUTEX_SPIN_OLDSPL(ci);
733 		ci->ci_mtx_count--;
734 		lwp_unlock(l);
735 
736 		/* Count the context switch on this CPU. */
737 		ci->ci_data.cpu_nswtch++;
738 
739 		/* Update status for lwpctl, if present. */
740 		if (l->l_lwpctl != NULL)
741 			l->l_lwpctl->lc_curcpu = LWPCTL_CPU_NONE;
742 
743 		/*
744 		 * Save old VM context, unless a soft interrupt
745 		 * handler is blocking.
746 		 */
747 		if (!returning)
748 			pmap_deactivate(l);
749 
750 		/*
751 		 * We may need to spin-wait for if 'newl' is still
752 		 * context switching on another CPU.
753 		 */
754 		if (__predict_false(newl->l_ctxswtch != 0)) {
755 			u_int count;
756 			count = SPINLOCK_BACKOFF_MIN;
757 			while (newl->l_ctxswtch)
758 				SPINLOCK_BACKOFF(count);
759 		}
760 
761 		/* Switch to the new LWP.. */
762 		prevlwp = cpu_switchto(l, newl, returning);
763 		ci = curcpu();
764 
765 		/*
766 		 * Switched away - we have new curlwp.
767 		 * Restore VM context and IPL.
768 		 */
769 		pmap_activate(l);
770 		if (prevlwp != NULL) {
771 			/* Normalize the count of the spin-mutexes */
772 			ci->ci_mtx_count++;
773 			/* Unmark the state of context switch */
774 			membar_exit();
775 			prevlwp->l_ctxswtch = 0;
776 		}
777 
778 		/* Update status for lwpctl, if present. */
779 		if (l->l_lwpctl != NULL) {
780 			l->l_lwpctl->lc_curcpu = (int)cpu_index(ci);
781 			l->l_lwpctl->lc_pctr++;
782 		}
783 
784 		KASSERT(l->l_cpu == ci);
785 		splx(oldspl);
786 		retval = 1;
787 	} else {
788 		/* Nothing to do - just unlock and return. */
789 		mutex_spin_exit(spc->spc_mutex);
790 		lwp_unlock(l);
791 		retval = 0;
792 	}
793 
794 	KASSERT(l == curlwp);
795 	KASSERT(l->l_stat == LSONPROC);
796 
797 	/*
798 	 * XXXSMP If we are using h/w performance counters, restore context.
799 	 * XXXSMP preemption problem.
800 	 */
801 #if PERFCTRS
802 	if (PMC_ENABLED(l->l_proc)) {
803 		pmc_restore_context(l->l_proc);
804 	}
805 #endif
806 	SYSCALL_TIME_WAKEUP(l);
807 	LOCKDEBUG_BARRIER(NULL, 1);
808 
809 	return retval;
810 }
811 
812 /*
813  * The machine independent parts of context switch to oblivion.
814  * Does not return.  Call with the LWP unlocked.
815  */
816 void
817 lwp_exit_switchaway(lwp_t *l)
818 {
819 	struct cpu_info *ci;
820 	struct lwp *newl;
821 	struct bintime bt;
822 
823 	ci = l->l_cpu;
824 
825 	KASSERT(kpreempt_disabled());
826 	KASSERT(l->l_stat == LSZOMB || l->l_stat == LSIDL);
827 	KASSERT(ci == curcpu());
828 	LOCKDEBUG_BARRIER(NULL, 0);
829 
830 	kstack_check_magic(l);
831 
832 	/* Count time spent in current system call */
833 	SYSCALL_TIME_SLEEP(l);
834 	binuptime(&bt);
835 	updatertime(l, &bt);
836 
837 	/* Must stay at IPL_SCHED even after releasing run queue lock. */
838 	(void)splsched();
839 
840 	/*
841 	 * Let sched_nextlwp() select the LWP to run the CPU next.
842 	 * If no LWP is runnable, select the idle LWP.
843 	 *
844 	 * Note that spc_lwplock might not necessary be held, and
845 	 * new thread would be unlocked after setting the LWP-lock.
846 	 */
847 	spc_lock(ci);
848 #ifndef __HAVE_FAST_SOFTINTS
849 	if (ci->ci_data.cpu_softints != 0) {
850 		/* There are pending soft interrupts, so pick one. */
851 		newl = softint_picklwp();
852 		newl->l_stat = LSONPROC;
853 		newl->l_pflag |= LP_RUNNING;
854 	} else
855 #endif	/* !__HAVE_FAST_SOFTINTS */
856 	{
857 		newl = nextlwp(ci, &ci->ci_schedstate);
858 	}
859 
860 	/* Update the new LWP's start time. */
861 	newl->l_stime = bt;
862 	l->l_pflag &= ~LP_RUNNING;
863 
864 	/*
865 	 * ci_curlwp changes when a fast soft interrupt occurs.
866 	 * We use cpu_onproc to keep track of which kernel or
867 	 * user thread is running 'underneath' the software
868 	 * interrupt.  This is important for time accounting,
869 	 * itimers and forcing user threads to preempt (aston).
870 	 */
871 	ci->ci_data.cpu_onproc = newl;
872 
873 	/*
874 	 * Preemption related tasks.  Must be done with the current
875 	 * CPU locked.
876 	 */
877 	cpu_did_resched(l);
878 
879 	/* Unlock the run queue. */
880 	spc_unlock(ci);
881 
882 	/* Count the context switch on this CPU. */
883 	ci->ci_data.cpu_nswtch++;
884 
885 	/* Update status for lwpctl, if present. */
886 	if (l->l_lwpctl != NULL)
887 		l->l_lwpctl->lc_curcpu = LWPCTL_CPU_EXITED;
888 
889 	/*
890 	 * We may need to spin-wait for if 'newl' is still
891 	 * context switching on another CPU.
892 	 */
893 	if (__predict_false(newl->l_ctxswtch != 0)) {
894 		u_int count;
895 		count = SPINLOCK_BACKOFF_MIN;
896 		while (newl->l_ctxswtch)
897 			SPINLOCK_BACKOFF(count);
898 	}
899 
900 	/* Switch to the new LWP.. */
901 	(void)cpu_switchto(NULL, newl, false);
902 
903 	for (;;) continue;	/* XXX: convince gcc about "noreturn" */
904 	/* NOTREACHED */
905 }
906 
907 /*
908  * Change LWP state to be runnable, placing it on the run queue if it is
909  * in memory, and awakening the swapper if it isn't in memory.
910  *
911  * Call with the process and LWP locked.  Will return with the LWP unlocked.
912  */
913 void
914 setrunnable(struct lwp *l)
915 {
916 	struct proc *p = l->l_proc;
917 	struct cpu_info *ci;
918 
919 	KASSERT((l->l_flag & LW_IDLE) == 0);
920 	KASSERT(mutex_owned(p->p_lock));
921 	KASSERT(lwp_locked(l, NULL));
922 	KASSERT(l->l_mutex != l->l_cpu->ci_schedstate.spc_mutex);
923 
924 	switch (l->l_stat) {
925 	case LSSTOP:
926 		/*
927 		 * If we're being traced (possibly because someone attached us
928 		 * while we were stopped), check for a signal from the debugger.
929 		 */
930 		if ((p->p_slflag & PSL_TRACED) != 0 && p->p_xstat != 0)
931 			signotify(l);
932 		p->p_nrlwps++;
933 		break;
934 	case LSSUSPENDED:
935 		l->l_flag &= ~LW_WSUSPEND;
936 		p->p_nrlwps++;
937 		cv_broadcast(&p->p_lwpcv);
938 		break;
939 	case LSSLEEP:
940 		KASSERT(l->l_wchan != NULL);
941 		break;
942 	default:
943 		panic("setrunnable: lwp %p state was %d", l, l->l_stat);
944 	}
945 
946 #ifdef KERN_SA
947 	if (l->l_proc->p_sa)
948 		sa_awaken(l);
949 #endif /* KERN_SA */
950 
951 	/*
952 	 * If the LWP was sleeping interruptably, then it's OK to start it
953 	 * again.  If not, mark it as still sleeping.
954 	 */
955 	if (l->l_wchan != NULL) {
956 		l->l_stat = LSSLEEP;
957 		/* lwp_unsleep() will release the lock. */
958 		lwp_unsleep(l, true);
959 		return;
960 	}
961 
962 	/*
963 	 * If the LWP is still on the CPU, mark it as LSONPROC.  It may be
964 	 * about to call mi_switch(), in which case it will yield.
965 	 */
966 	if ((l->l_pflag & LP_RUNNING) != 0) {
967 		l->l_stat = LSONPROC;
968 		l->l_slptime = 0;
969 		lwp_unlock(l);
970 		return;
971 	}
972 
973 	/*
974 	 * Look for a CPU to run.
975 	 * Set the LWP runnable.
976 	 */
977 	ci = sched_takecpu(l);
978 	l->l_cpu = ci;
979 	spc_lock(ci);
980 	lwp_unlock_to(l, ci->ci_schedstate.spc_mutex);
981 	sched_setrunnable(l);
982 	l->l_stat = LSRUN;
983 	l->l_slptime = 0;
984 
985 	/*
986 	 * If thread is swapped out - wake the swapper to bring it back in.
987 	 * Otherwise, enter it into a run queue.
988 	 */
989 	if (l->l_flag & LW_INMEM) {
990 		sched_enqueue(l, false);
991 		resched_cpu(l);
992 		lwp_unlock(l);
993 	} else {
994 		lwp_unlock(l);
995 		uvm_kick_scheduler();
996 	}
997 }
998 
999 /*
1000  * suspendsched:
1001  *
1002  *	Convert all non-L_SYSTEM LSSLEEP or LSRUN LWPs to LSSUSPENDED.
1003  */
1004 void
1005 suspendsched(void)
1006 {
1007 	CPU_INFO_ITERATOR cii;
1008 	struct cpu_info *ci;
1009 	struct lwp *l;
1010 	struct proc *p;
1011 
1012 	/*
1013 	 * We do this by process in order not to violate the locking rules.
1014 	 */
1015 	mutex_enter(proc_lock);
1016 	PROCLIST_FOREACH(p, &allproc) {
1017 		if ((p->p_flag & PK_MARKER) != 0)
1018 			continue;
1019 
1020 		mutex_enter(p->p_lock);
1021 		if ((p->p_flag & PK_SYSTEM) != 0) {
1022 			mutex_exit(p->p_lock);
1023 			continue;
1024 		}
1025 
1026 		p->p_stat = SSTOP;
1027 
1028 		LIST_FOREACH(l, &p->p_lwps, l_sibling) {
1029 			if (l == curlwp)
1030 				continue;
1031 
1032 			lwp_lock(l);
1033 
1034 			/*
1035 			 * Set L_WREBOOT so that the LWP will suspend itself
1036 			 * when it tries to return to user mode.  We want to
1037 			 * try and get to get as many LWPs as possible to
1038 			 * the user / kernel boundary, so that they will
1039 			 * release any locks that they hold.
1040 			 */
1041 			l->l_flag |= (LW_WREBOOT | LW_WSUSPEND);
1042 
1043 			if (l->l_stat == LSSLEEP &&
1044 			    (l->l_flag & LW_SINTR) != 0) {
1045 				/* setrunnable() will release the lock. */
1046 				setrunnable(l);
1047 				continue;
1048 			}
1049 
1050 			lwp_unlock(l);
1051 		}
1052 
1053 		mutex_exit(p->p_lock);
1054 	}
1055 	mutex_exit(proc_lock);
1056 
1057 	/*
1058 	 * Kick all CPUs to make them preempt any LWPs running in user mode.
1059 	 * They'll trap into the kernel and suspend themselves in userret().
1060 	 */
1061 	for (CPU_INFO_FOREACH(cii, ci)) {
1062 		spc_lock(ci);
1063 		cpu_need_resched(ci, RESCHED_IMMED);
1064 		spc_unlock(ci);
1065 	}
1066 }
1067 
1068 /*
1069  * sched_unsleep:
1070  *
1071  *	The is called when the LWP has not been awoken normally but instead
1072  *	interrupted: for example, if the sleep timed out.  Because of this,
1073  *	it's not a valid action for running or idle LWPs.
1074  */
1075 static u_int
1076 sched_unsleep(struct lwp *l, bool cleanup)
1077 {
1078 
1079 	lwp_unlock(l);
1080 	panic("sched_unsleep");
1081 }
1082 
1083 static void
1084 resched_cpu(struct lwp *l)
1085 {
1086 	struct cpu_info *ci = ci = l->l_cpu;
1087 
1088 	KASSERT(lwp_locked(l, NULL));
1089 	if (lwp_eprio(l) > ci->ci_schedstate.spc_curpriority)
1090 		cpu_need_resched(ci, 0);
1091 }
1092 
1093 static void
1094 sched_changepri(struct lwp *l, pri_t pri)
1095 {
1096 
1097 	KASSERT(lwp_locked(l, NULL));
1098 
1099 	if (l->l_stat == LSRUN && (l->l_flag & LW_INMEM) != 0) {
1100 		KASSERT(lwp_locked(l, l->l_cpu->ci_schedstate.spc_mutex));
1101 		sched_dequeue(l);
1102 		l->l_priority = pri;
1103 		sched_enqueue(l, false);
1104 	} else {
1105 		l->l_priority = pri;
1106 	}
1107 	resched_cpu(l);
1108 }
1109 
1110 static void
1111 sched_lendpri(struct lwp *l, pri_t pri)
1112 {
1113 
1114 	KASSERT(lwp_locked(l, NULL));
1115 
1116 	if (l->l_stat == LSRUN && (l->l_flag & LW_INMEM) != 0) {
1117 		KASSERT(lwp_locked(l, l->l_cpu->ci_schedstate.spc_mutex));
1118 		sched_dequeue(l);
1119 		l->l_inheritedprio = pri;
1120 		sched_enqueue(l, false);
1121 	} else {
1122 		l->l_inheritedprio = pri;
1123 	}
1124 	resched_cpu(l);
1125 }
1126 
1127 struct lwp *
1128 syncobj_noowner(wchan_t wchan)
1129 {
1130 
1131 	return NULL;
1132 }
1133 
1134 /* Decay 95% of proc::p_pctcpu in 60 seconds, ccpu = exp(-1/20) */
1135 const fixpt_t	ccpu = 0.95122942450071400909 * FSCALE;
1136 
1137 /*
1138  * sched_pstats:
1139  *
1140  * Update process statistics and check CPU resource allocation.
1141  * Call scheduler-specific hook to eventually adjust process/LWP
1142  * priorities.
1143  */
1144 /* ARGSUSED */
1145 void
1146 sched_pstats(void *arg)
1147 {
1148 	const int clkhz = (stathz != 0 ? stathz : hz);
1149 	static bool backwards;
1150 	struct rlimit *rlim;
1151 	struct lwp *l;
1152 	struct proc *p;
1153 	long runtm;
1154 	fixpt_t lpctcpu;
1155 	u_int lcpticks;
1156 	int sig;
1157 
1158 	sched_pstats_ticks++;
1159 
1160 	mutex_enter(proc_lock);
1161 	PROCLIST_FOREACH(p, &allproc) {
1162 		if (__predict_false((p->p_flag & PK_MARKER) != 0))
1163 			continue;
1164 
1165 		/*
1166 		 * Increment time in/out of memory and sleep
1167 		 * time (if sleeping), ignore overflow.
1168 		 */
1169 		mutex_enter(p->p_lock);
1170 		runtm = p->p_rtime.sec;
1171 		LIST_FOREACH(l, &p->p_lwps, l_sibling) {
1172 			if (__predict_false((l->l_flag & LW_IDLE) != 0))
1173 				continue;
1174 			lwp_lock(l);
1175 			runtm += l->l_rtime.sec;
1176 			l->l_swtime++;
1177 			sched_lwp_stats(l);
1178 			lwp_unlock(l);
1179 
1180 			l->l_pctcpu = (l->l_pctcpu * ccpu) >> FSHIFT;
1181 			if (l->l_slptime != 0)
1182 				continue;
1183 
1184 			lpctcpu = l->l_pctcpu;
1185 			lcpticks = atomic_swap_uint(&l->l_cpticks, 0);
1186 			lpctcpu += ((FSCALE - ccpu) *
1187 			    (lcpticks * FSCALE / clkhz)) >> FSHIFT;
1188 			l->l_pctcpu = lpctcpu;
1189 		}
1190 		/* Calculating p_pctcpu only for ps(1) */
1191 		p->p_pctcpu = (p->p_pctcpu * ccpu) >> FSHIFT;
1192 
1193 		/*
1194 		 * Check if the process exceeds its CPU resource allocation.
1195 		 * If over max, kill it.
1196 		 */
1197 		rlim = &p->p_rlimit[RLIMIT_CPU];
1198 		sig = 0;
1199 		if (__predict_false(runtm >= rlim->rlim_cur)) {
1200 			if (runtm >= rlim->rlim_max)
1201 				sig = SIGKILL;
1202 			else {
1203 				sig = SIGXCPU;
1204 				if (rlim->rlim_cur < rlim->rlim_max)
1205 					rlim->rlim_cur += 5;
1206 			}
1207 		}
1208 		mutex_exit(p->p_lock);
1209 		if (__predict_false(runtm < 0)) {
1210 			if (!backwards) {
1211 				backwards = true;
1212 				printf("WARNING: negative runtime; "
1213 				    "monotonic clock has gone backwards\n");
1214 			}
1215 		} else if (__predict_false(sig)) {
1216 			KASSERT((p->p_flag & PK_SYSTEM) == 0);
1217 			psignal(p, sig);
1218 		}
1219 	}
1220 	mutex_exit(proc_lock);
1221 	uvm_meter();
1222 	cv_wakeup(&lbolt);
1223 	callout_schedule(&sched_pstats_ch, hz);
1224 }
1225