xref: /netbsd-src/sys/kern/kern_synch.c (revision 0df165c04d0a9ca1adde9ed2b890344c937954a6)
1 /*	$NetBSD: kern_synch.c,v 1.207 2007/11/12 23:11:59 ad Exp $	*/
2 
3 /*-
4  * Copyright (c) 1999, 2000, 2004, 2006, 2007 The NetBSD Foundation, Inc.
5  * All rights reserved.
6  *
7  * This code is derived from software contributed to The NetBSD Foundation
8  * by Jason R. Thorpe of the Numerical Aerospace Simulation Facility,
9  * NASA Ames Research Center, by Charles M. Hannum, Andrew Doran and
10  * Daniel Sieger.
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. All advertising materials mentioning features or use of this software
21  *    must display the following acknowledgement:
22  *	This product includes software developed by the NetBSD
23  *	Foundation, Inc. and its contributors.
24  * 4. Neither the name of The NetBSD Foundation nor the names of its
25  *    contributors may be used to endorse or promote products derived
26  *    from this software without specific prior written permission.
27  *
28  * THIS SOFTWARE IS PROVIDED BY THE NETBSD FOUNDATION, INC. AND CONTRIBUTORS
29  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED
30  * TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
31  * PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE FOUNDATION OR CONTRIBUTORS
32  * BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
33  * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
34  * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
35  * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
36  * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
37  * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
38  * POSSIBILITY OF SUCH DAMAGE.
39  */
40 
41 /*-
42  * Copyright (c) 1982, 1986, 1990, 1991, 1993
43  *	The Regents of the University of California.  All rights reserved.
44  * (c) UNIX System Laboratories, Inc.
45  * All or some portions of this file are derived from material licensed
46  * to the University of California by American Telephone and Telegraph
47  * Co. or Unix System Laboratories, Inc. and are reproduced herein with
48  * the permission of UNIX System Laboratories, Inc.
49  *
50  * Redistribution and use in source and binary forms, with or without
51  * modification, are permitted provided that the following conditions
52  * are met:
53  * 1. Redistributions of source code must retain the above copyright
54  *    notice, this list of conditions and the following disclaimer.
55  * 2. Redistributions in binary form must reproduce the above copyright
56  *    notice, this list of conditions and the following disclaimer in the
57  *    documentation and/or other materials provided with the distribution.
58  * 3. Neither the name of the University nor the names of its contributors
59  *    may be used to endorse or promote products derived from this software
60  *    without specific prior written permission.
61  *
62  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
63  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
64  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
65  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
66  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
67  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
68  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
69  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
70  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
71  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
72  * SUCH DAMAGE.
73  *
74  *	@(#)kern_synch.c	8.9 (Berkeley) 5/19/95
75  */
76 
77 #include <sys/cdefs.h>
78 __KERNEL_RCSID(0, "$NetBSD: kern_synch.c,v 1.207 2007/11/12 23:11:59 ad Exp $");
79 
80 #include "opt_kstack.h"
81 #include "opt_lockdebug.h"
82 #include "opt_multiprocessor.h"
83 #include "opt_perfctrs.h"
84 
85 #define	__MUTEX_PRIVATE
86 
87 #include <sys/param.h>
88 #include <sys/systm.h>
89 #include <sys/proc.h>
90 #include <sys/kernel.h>
91 #if defined(PERFCTRS)
92 #include <sys/pmc.h>
93 #endif
94 #include <sys/cpu.h>
95 #include <sys/resourcevar.h>
96 #include <sys/sched.h>
97 #include <sys/syscall_stats.h>
98 #include <sys/sleepq.h>
99 #include <sys/lockdebug.h>
100 #include <sys/evcnt.h>
101 #include <sys/intr.h>
102 #include <sys/lwpctl.h>
103 
104 #include <uvm/uvm_extern.h>
105 
106 callout_t sched_pstats_ch;
107 unsigned int sched_pstats_ticks;
108 
109 kcondvar_t	lbolt;			/* once a second sleep address */
110 
111 static void	sched_unsleep(struct lwp *);
112 static void	sched_changepri(struct lwp *, pri_t);
113 static void	sched_lendpri(struct lwp *, pri_t);
114 
115 syncobj_t sleep_syncobj = {
116 	SOBJ_SLEEPQ_SORTED,
117 	sleepq_unsleep,
118 	sleepq_changepri,
119 	sleepq_lendpri,
120 	syncobj_noowner,
121 };
122 
123 syncobj_t sched_syncobj = {
124 	SOBJ_SLEEPQ_SORTED,
125 	sched_unsleep,
126 	sched_changepri,
127 	sched_lendpri,
128 	syncobj_noowner,
129 };
130 
131 /*
132  * During autoconfiguration or after a panic, a sleep will simply lower the
133  * priority briefly to allow interrupts, then return.  The priority to be
134  * used (safepri) is machine-dependent, thus this value is initialized and
135  * maintained in the machine-dependent layers.  This priority will typically
136  * be 0, or the lowest priority that is safe for use on the interrupt stack;
137  * it can be made higher to block network software interrupts after panics.
138  */
139 int	safepri;
140 
141 /*
142  * OBSOLETE INTERFACE
143  *
144  * General sleep call.  Suspends the current process until a wakeup is
145  * performed on the specified identifier.  The process will then be made
146  * runnable with the specified priority.  Sleeps at most timo/hz seconds (0
147  * means no timeout).  If pri includes PCATCH flag, signals are checked
148  * before and after sleeping, else signals are not checked.  Returns 0 if
149  * awakened, EWOULDBLOCK if the timeout expires.  If PCATCH is set and a
150  * signal needs to be delivered, ERESTART is returned if the current system
151  * call should be restarted if possible, and EINTR is returned if the system
152  * call should be interrupted by the signal (return EINTR).
153  *
154  * The interlock is held until we are on a sleep queue. The interlock will
155  * be locked before returning back to the caller unless the PNORELOCK flag
156  * is specified, in which case the interlock will always be unlocked upon
157  * return.
158  */
159 int
160 ltsleep(wchan_t ident, pri_t priority, const char *wmesg, int timo,
161 	volatile struct simplelock *interlock)
162 {
163 	struct lwp *l = curlwp;
164 	sleepq_t *sq;
165 	int error;
166 
167 	KASSERT((l->l_pflag & LP_INTR) == 0);
168 
169 	if (sleepq_dontsleep(l)) {
170 		(void)sleepq_abort(NULL, 0);
171 		if ((priority & PNORELOCK) != 0)
172 			simple_unlock(interlock);
173 		return 0;
174 	}
175 
176 	l->l_kpriority = true;
177 	sq = sleeptab_lookup(&sleeptab, ident);
178 	sleepq_enter(sq, l);
179 	sleepq_enqueue(sq, ident, wmesg, &sleep_syncobj);
180 
181 	if (interlock != NULL) {
182 		KASSERT(simple_lock_held(interlock));
183 		simple_unlock(interlock);
184 	}
185 
186 	error = sleepq_block(timo, priority & PCATCH);
187 
188 	if (interlock != NULL && (priority & PNORELOCK) == 0)
189 		simple_lock(interlock);
190 
191 	return error;
192 }
193 
194 int
195 mtsleep(wchan_t ident, pri_t priority, const char *wmesg, int timo,
196 	kmutex_t *mtx)
197 {
198 	struct lwp *l = curlwp;
199 	sleepq_t *sq;
200 	int error;
201 
202 	KASSERT((l->l_pflag & LP_INTR) == 0);
203 
204 	if (sleepq_dontsleep(l)) {
205 		(void)sleepq_abort(mtx, (priority & PNORELOCK) != 0);
206 		return 0;
207 	}
208 
209 	l->l_kpriority = true;
210 	sq = sleeptab_lookup(&sleeptab, ident);
211 	sleepq_enter(sq, l);
212 	sleepq_enqueue(sq, ident, wmesg, &sleep_syncobj);
213 	mutex_exit(mtx);
214 	error = sleepq_block(timo, priority & PCATCH);
215 
216 	if ((priority & PNORELOCK) == 0)
217 		mutex_enter(mtx);
218 
219 	return error;
220 }
221 
222 /*
223  * General sleep call for situations where a wake-up is not expected.
224  */
225 int
226 kpause(const char *wmesg, bool intr, int timo, kmutex_t *mtx)
227 {
228 	struct lwp *l = curlwp;
229 	sleepq_t *sq;
230 	int error;
231 
232 	if (sleepq_dontsleep(l))
233 		return sleepq_abort(NULL, 0);
234 
235 	if (mtx != NULL)
236 		mutex_exit(mtx);
237 	l->l_kpriority = true;
238 	sq = sleeptab_lookup(&sleeptab, l);
239 	sleepq_enter(sq, l);
240 	sleepq_enqueue(sq, l, wmesg, &sleep_syncobj);
241 	error = sleepq_block(timo, intr);
242 	if (mtx != NULL)
243 		mutex_enter(mtx);
244 
245 	return error;
246 }
247 
248 /*
249  * OBSOLETE INTERFACE
250  *
251  * Make all processes sleeping on the specified identifier runnable.
252  */
253 void
254 wakeup(wchan_t ident)
255 {
256 	sleepq_t *sq;
257 
258 	if (cold)
259 		return;
260 
261 	sq = sleeptab_lookup(&sleeptab, ident);
262 	sleepq_wake(sq, ident, (u_int)-1);
263 }
264 
265 /*
266  * OBSOLETE INTERFACE
267  *
268  * Make the highest priority process first in line on the specified
269  * identifier runnable.
270  */
271 void
272 wakeup_one(wchan_t ident)
273 {
274 	sleepq_t *sq;
275 
276 	if (cold)
277 		return;
278 
279 	sq = sleeptab_lookup(&sleeptab, ident);
280 	sleepq_wake(sq, ident, 1);
281 }
282 
283 
284 /*
285  * General yield call.  Puts the current process back on its run queue and
286  * performs a voluntary context switch.  Should only be called when the
287  * current process explicitly requests it (eg sched_yield(2)).
288  */
289 void
290 yield(void)
291 {
292 	struct lwp *l = curlwp;
293 
294 	KERNEL_UNLOCK_ALL(l, &l->l_biglocks);
295 	lwp_lock(l);
296 	KASSERT(lwp_locked(l, &l->l_cpu->ci_schedstate.spc_lwplock));
297 	KASSERT(l->l_stat == LSONPROC);
298 	l->l_kpriority = false;
299 	if (l->l_class == SCHED_OTHER) {
300 		/*
301 		 * Only for timeshared threads.  It will be reset
302 		 * by the scheduler in due course.
303 		 */
304 		l->l_priority = 0;
305 	}
306 	(void)mi_switch(l);
307 	KERNEL_LOCK(l->l_biglocks, l);
308 }
309 
310 /*
311  * General preemption call.  Puts the current process back on its run queue
312  * and performs an involuntary context switch.
313  */
314 void
315 preempt(void)
316 {
317 	struct lwp *l = curlwp;
318 
319 	KERNEL_UNLOCK_ALL(l, &l->l_biglocks);
320 	lwp_lock(l);
321 	KASSERT(lwp_locked(l, &l->l_cpu->ci_schedstate.spc_lwplock));
322 	KASSERT(l->l_stat == LSONPROC);
323 	l->l_kpriority = false;
324 	l->l_nivcsw++;
325 	(void)mi_switch(l);
326 	KERNEL_LOCK(l->l_biglocks, l);
327 }
328 
329 /*
330  * Compute the amount of time during which the current lwp was running.
331  *
332  * - update l_rtime unless it's an idle lwp.
333  */
334 
335 void
336 updatertime(lwp_t *l, const struct timeval *tv)
337 {
338 	long s, u;
339 
340 	if ((l->l_flag & LW_IDLE) != 0)
341 		return;
342 
343 	u = l->l_rtime.tv_usec + (tv->tv_usec - l->l_stime.tv_usec);
344 	s = l->l_rtime.tv_sec + (tv->tv_sec - l->l_stime.tv_sec);
345 	if (u < 0) {
346 		u += 1000000;
347 		s--;
348 	} else if (u >= 1000000) {
349 		u -= 1000000;
350 		s++;
351 	}
352 	l->l_rtime.tv_usec = u;
353 	l->l_rtime.tv_sec = s;
354 }
355 
356 /*
357  * The machine independent parts of context switch.
358  *
359  * Returns 1 if another LWP was actually run.
360  */
361 int
362 mi_switch(lwp_t *l)
363 {
364 	struct schedstate_percpu *spc;
365 	struct lwp *newl;
366 	int retval, oldspl;
367 	struct cpu_info *ci;
368 	struct timeval tv;
369 	bool returning;
370 
371 	KASSERT(lwp_locked(l, NULL));
372 	LOCKDEBUG_BARRIER(l->l_mutex, 1);
373 
374 #ifdef KSTACK_CHECK_MAGIC
375 	kstack_check_magic(l);
376 #endif
377 
378 	microtime(&tv);
379 
380 	/*
381 	 * It's safe to read the per CPU schedstate unlocked here, as all we
382 	 * are after is the run time and that's guarenteed to have been last
383 	 * updated by this CPU.
384 	 */
385 	ci = l->l_cpu;
386 	KDASSERT(ci == curcpu());
387 
388 	/*
389 	 * Process is about to yield the CPU; clear the appropriate
390 	 * scheduling flags.
391 	 */
392 	spc = &ci->ci_schedstate;
393 	returning = false;
394 	newl = NULL;
395 
396 	/*
397 	 * If we have been asked to switch to a specific LWP, then there
398 	 * is no need to inspect the run queues.  If a soft interrupt is
399 	 * blocking, then return to the interrupted thread without adjusting
400 	 * VM context or its start time: neither have been changed in order
401 	 * to take the interrupt.
402 	 */
403 	if (l->l_switchto != NULL) {
404 		if ((l->l_pflag & LP_INTR) != 0) {
405 			returning = true;
406 			softint_block(l);
407 			if ((l->l_flag & LW_TIMEINTR) != 0)
408 				updatertime(l, &tv);
409 		}
410 		newl = l->l_switchto;
411 		l->l_switchto = NULL;
412 	}
413 #ifndef __HAVE_FAST_SOFTINTS
414 	else if (ci->ci_data.cpu_softints != 0) {
415 		/* There are pending soft interrupts, so pick one. */
416 		newl = softint_picklwp();
417 		newl->l_stat = LSONPROC;
418 		newl->l_flag |= LW_RUNNING;
419 	}
420 #endif	/* !__HAVE_FAST_SOFTINTS */
421 
422 	/* Count time spent in current system call */
423 	if (!returning) {
424 		SYSCALL_TIME_SLEEP(l);
425 
426 		/*
427 		 * XXXSMP If we are using h/w performance counters,
428 		 * save context.
429 		 */
430 #if PERFCTRS
431 		if (PMC_ENABLED(l->l_proc)) {
432 			pmc_save_context(l->l_proc);
433 		}
434 #endif
435 		updatertime(l, &tv);
436 	}
437 
438 	/*
439 	 * If on the CPU and we have gotten this far, then we must yield.
440 	 */
441 	mutex_spin_enter(spc->spc_mutex);
442 	KASSERT(l->l_stat != LSRUN);
443 	if (l->l_stat == LSONPROC && l != newl) {
444 		KASSERT(lwp_locked(l, &spc->spc_lwplock));
445 		if ((l->l_flag & LW_IDLE) == 0) {
446 			l->l_stat = LSRUN;
447 			lwp_setlock(l, spc->spc_mutex);
448 			sched_enqueue(l, true);
449 		} else
450 			l->l_stat = LSIDL;
451 	}
452 
453 	/*
454 	 * Let sched_nextlwp() select the LWP to run the CPU next.
455 	 * If no LWP is runnable, switch to the idle LWP.
456 	 * Note that spc_lwplock might not necessary be held.
457 	 */
458 	if (newl == NULL) {
459 		newl = sched_nextlwp();
460 		if (newl != NULL) {
461 			sched_dequeue(newl);
462 			KASSERT(lwp_locked(newl, spc->spc_mutex));
463 			newl->l_stat = LSONPROC;
464 			newl->l_cpu = ci;
465 			newl->l_flag |= LW_RUNNING;
466 			lwp_setlock(newl, &spc->spc_lwplock);
467 		} else {
468 			newl = ci->ci_data.cpu_idlelwp;
469 			newl->l_stat = LSONPROC;
470 			newl->l_flag |= LW_RUNNING;
471 		}
472 		/*
473 		 * Only clear want_resched if there are no
474 		 * pending (slow) software interrupts.
475 		 */
476 		ci->ci_want_resched = ci->ci_data.cpu_softints;
477 		spc->spc_flags &= ~SPCF_SWITCHCLEAR;
478 		spc->spc_curpriority = lwp_eprio(newl);
479 	}
480 
481 	/* Items that must be updated with the CPU locked. */
482 	if (!returning) {
483 		/* Update the new LWP's start time. */
484 		newl->l_stime = tv;
485 
486 		/*
487 		 * ci_curlwp changes when a fast soft interrupt occurs.
488 		 * We use cpu_onproc to keep track of which kernel or
489 		 * user thread is running 'underneath' the software
490 		 * interrupt.  This is important for time accounting,
491 		 * itimers and forcing user threads to preempt (aston).
492 		 */
493 		ci->ci_data.cpu_onproc = newl;
494 	}
495 
496 	if (l != newl) {
497 		struct lwp *prevlwp;
498 
499 		/*
500 		 * If the old LWP has been moved to a run queue above,
501 		 * drop the general purpose LWP lock: it's now locked
502 		 * by the scheduler lock.
503 		 *
504 		 * Otherwise, drop the scheduler lock.  We're done with
505 		 * the run queues for now.
506 		 */
507 		if (l->l_mutex == spc->spc_mutex) {
508 			mutex_spin_exit(&spc->spc_lwplock);
509 		} else {
510 			mutex_spin_exit(spc->spc_mutex);
511 		}
512 
513 		/* Unlocked, but for statistics only. */
514 		uvmexp.swtch++;
515 
516 		/*
517 		 * Save old VM context, unless a soft interrupt
518 		 * handler is blocking.
519 		 */
520 		if (!returning)
521 			pmap_deactivate(l);
522 
523 		/* Update status for lwpctl, if present. */
524 	        if (l->l_lwpctl != NULL)
525 			l->l_lwpctl->lc_curcpu = LWPCTL_CPU_NONE;
526 
527 		/* Switch to the new LWP.. */
528 		l->l_ncsw++;
529 		l->l_flag &= ~LW_RUNNING;
530 		oldspl = MUTEX_SPIN_OLDSPL(ci);
531 		prevlwp = cpu_switchto(l, newl, returning);
532 		ci = curcpu();
533 
534 		/*
535 		 * .. we have switched away and are now back so we must
536 		 * be the new curlwp.  prevlwp is who we replaced.
537 		 */
538 		if (prevlwp != NULL) {
539 			ci->ci_mtx_oldspl = oldspl;
540 			lwp_unlock(prevlwp);
541 		} else {
542 			splx(oldspl);
543 		}
544 
545 		/* Restore VM context. */
546 		pmap_activate(l);
547 		retval = 1;
548 
549 		/* Update status for lwpctl, if present. */
550 	        if (l->l_lwpctl != NULL)
551 			l->l_lwpctl->lc_curcpu = (short)ci->ci_data.cpu_index;
552 	} else {
553 		/* Nothing to do - just unlock and return. */
554 		mutex_spin_exit(spc->spc_mutex);
555 		lwp_unlock(l);
556 		retval = 0;
557 	}
558 
559 	KASSERT(l == curlwp);
560 	KASSERT(l->l_stat == LSONPROC);
561 	KASSERT(l->l_cpu == ci);
562 
563 	/*
564 	 * XXXSMP If we are using h/w performance counters, restore context.
565 	 */
566 #if PERFCTRS
567 	if (PMC_ENABLED(l->l_proc)) {
568 		pmc_restore_context(l->l_proc);
569 	}
570 #endif
571 
572 	/*
573 	 * We're running again; record our new start time.  We might
574 	 * be running on a new CPU now, so don't use the cached
575 	 * schedstate_percpu pointer.
576 	 */
577 	SYSCALL_TIME_WAKEUP(l);
578 	KASSERT(curlwp == l);
579 	KDASSERT(l->l_cpu == ci);
580 	LOCKDEBUG_BARRIER(NULL, 1);
581 
582 	return retval;
583 }
584 
585 /*
586  * Change process state to be runnable, placing it on the run queue if it is
587  * in memory, and awakening the swapper if it isn't in memory.
588  *
589  * Call with the process and LWP locked.  Will return with the LWP unlocked.
590  */
591 void
592 setrunnable(struct lwp *l)
593 {
594 	struct proc *p = l->l_proc;
595 	struct cpu_info *ci;
596 	sigset_t *ss;
597 
598 	KASSERT((l->l_flag & LW_IDLE) == 0);
599 	KASSERT(mutex_owned(&p->p_smutex));
600 	KASSERT(lwp_locked(l, NULL));
601 	KASSERT(l->l_mutex != l->l_cpu->ci_schedstate.spc_mutex);
602 
603 	switch (l->l_stat) {
604 	case LSSTOP:
605 		/*
606 		 * If we're being traced (possibly because someone attached us
607 		 * while we were stopped), check for a signal from the debugger.
608 		 */
609 		if ((p->p_slflag & PSL_TRACED) != 0 && p->p_xstat != 0) {
610 			if ((sigprop[p->p_xstat] & SA_TOLWP) != 0)
611 				ss = &l->l_sigpend.sp_set;
612 			else
613 				ss = &p->p_sigpend.sp_set;
614 			sigaddset(ss, p->p_xstat);
615 			signotify(l);
616 		}
617 		p->p_nrlwps++;
618 		break;
619 	case LSSUSPENDED:
620 		l->l_flag &= ~LW_WSUSPEND;
621 		p->p_nrlwps++;
622 		cv_broadcast(&p->p_lwpcv);
623 		break;
624 	case LSSLEEP:
625 		KASSERT(l->l_wchan != NULL);
626 		break;
627 	default:
628 		panic("setrunnable: lwp %p state was %d", l, l->l_stat);
629 	}
630 
631 	/*
632 	 * If the LWP was sleeping interruptably, then it's OK to start it
633 	 * again.  If not, mark it as still sleeping.
634 	 */
635 	if (l->l_wchan != NULL) {
636 		l->l_stat = LSSLEEP;
637 		/* lwp_unsleep() will release the lock. */
638 		lwp_unsleep(l);
639 		return;
640 	}
641 
642 	/*
643 	 * If the LWP is still on the CPU, mark it as LSONPROC.  It may be
644 	 * about to call mi_switch(), in which case it will yield.
645 	 */
646 	if ((l->l_flag & LW_RUNNING) != 0) {
647 		l->l_stat = LSONPROC;
648 		l->l_slptime = 0;
649 		lwp_unlock(l);
650 		return;
651 	}
652 
653 	/*
654 	 * Look for a CPU to run.
655 	 * Set the LWP runnable.
656 	 */
657 	ci = sched_takecpu(l);
658 	l->l_cpu = ci;
659 	if (l->l_mutex != l->l_cpu->ci_schedstate.spc_mutex) {
660 		lwp_unlock_to(l, ci->ci_schedstate.spc_mutex);
661 		lwp_lock(l);
662 	}
663 	sched_setrunnable(l);
664 	l->l_stat = LSRUN;
665 	l->l_slptime = 0;
666 
667 	/*
668 	 * If thread is swapped out - wake the swapper to bring it back in.
669 	 * Otherwise, enter it into a run queue.
670 	 */
671 	if (l->l_flag & LW_INMEM) {
672 		sched_enqueue(l, false);
673 		resched_cpu(l);
674 		lwp_unlock(l);
675 	} else {
676 		lwp_unlock(l);
677 		uvm_kick_scheduler();
678 	}
679 }
680 
681 /*
682  * suspendsched:
683  *
684  *	Convert all non-L_SYSTEM LSSLEEP or LSRUN LWPs to LSSUSPENDED.
685  */
686 void
687 suspendsched(void)
688 {
689 	CPU_INFO_ITERATOR cii;
690 	struct cpu_info *ci;
691 	struct lwp *l;
692 	struct proc *p;
693 
694 	/*
695 	 * We do this by process in order not to violate the locking rules.
696 	 */
697 	mutex_enter(&proclist_lock);
698 	PROCLIST_FOREACH(p, &allproc) {
699 		mutex_enter(&p->p_smutex);
700 
701 		if ((p->p_flag & PK_SYSTEM) != 0) {
702 			mutex_exit(&p->p_smutex);
703 			continue;
704 		}
705 
706 		p->p_stat = SSTOP;
707 
708 		LIST_FOREACH(l, &p->p_lwps, l_sibling) {
709 			if (l == curlwp)
710 				continue;
711 
712 			lwp_lock(l);
713 
714 			/*
715 			 * Set L_WREBOOT so that the LWP will suspend itself
716 			 * when it tries to return to user mode.  We want to
717 			 * try and get to get as many LWPs as possible to
718 			 * the user / kernel boundary, so that they will
719 			 * release any locks that they hold.
720 			 */
721 			l->l_flag |= (LW_WREBOOT | LW_WSUSPEND);
722 
723 			if (l->l_stat == LSSLEEP &&
724 			    (l->l_flag & LW_SINTR) != 0) {
725 				/* setrunnable() will release the lock. */
726 				setrunnable(l);
727 				continue;
728 			}
729 
730 			lwp_unlock(l);
731 		}
732 
733 		mutex_exit(&p->p_smutex);
734 	}
735 	mutex_exit(&proclist_lock);
736 
737 	/*
738 	 * Kick all CPUs to make them preempt any LWPs running in user mode.
739 	 * They'll trap into the kernel and suspend themselves in userret().
740 	 */
741 	for (CPU_INFO_FOREACH(cii, ci)) {
742 		spc_lock(ci);
743 		cpu_need_resched(ci, RESCHED_IMMED);
744 		spc_unlock(ci);
745 	}
746 }
747 
748 /*
749  * sched_kpri:
750  *
751  *	Scale a priority level to a kernel priority level, usually
752  *	for an LWP that is about to sleep.
753  */
754 pri_t
755 sched_kpri(struct lwp *l)
756 {
757 	pri_t pri;
758 
759 #ifndef __HAVE_FAST_SOFTINTS
760 	/*
761 	 * Hack: if a user thread is being used to run a soft
762 	 * interrupt, we need to boost the priority here.
763 	 */
764 	if ((l->l_pflag & LP_INTR) != 0 && l->l_priority < PRI_KERNEL_RT)
765 		return softint_kpri(l);
766 #endif
767 
768 	/*
769 	 * Scale user priorities (0 -> 63) up to kernel priorities
770 	 * in the range (64 -> 95).  This makes assumptions about
771 	 * the priority space and so should be kept in sync with
772 	 * param.h.
773 	 */
774 	if ((pri = l->l_priority) >= PRI_KERNEL)
775 		return pri;
776 	return (pri >> 1) + PRI_KERNEL;
777 }
778 
779 /*
780  * sched_unsleep:
781  *
782  *	The is called when the LWP has not been awoken normally but instead
783  *	interrupted: for example, if the sleep timed out.  Because of this,
784  *	it's not a valid action for running or idle LWPs.
785  */
786 static void
787 sched_unsleep(struct lwp *l)
788 {
789 
790 	lwp_unlock(l);
791 	panic("sched_unsleep");
792 }
793 
794 void
795 resched_cpu(struct lwp *l)
796 {
797 	struct cpu_info *ci;
798 
799 	/*
800 	 * XXXSMP
801 	 * Since l->l_cpu persists across a context switch,
802 	 * this gives us *very weak* processor affinity, in
803 	 * that we notify the CPU on which the process last
804 	 * ran that it should try to switch.
805 	 *
806 	 * This does not guarantee that the process will run on
807 	 * that processor next, because another processor might
808 	 * grab it the next time it performs a context switch.
809 	 *
810 	 * This also does not handle the case where its last
811 	 * CPU is running a higher-priority process, but every
812 	 * other CPU is running a lower-priority process.  There
813 	 * are ways to handle this situation, but they're not
814 	 * currently very pretty, and we also need to weigh the
815 	 * cost of moving a process from one CPU to another.
816 	 */
817 	ci = l->l_cpu;
818 	if (lwp_eprio(l) > ci->ci_schedstate.spc_curpriority)
819 		cpu_need_resched(ci, 0);
820 }
821 
822 static void
823 sched_changepri(struct lwp *l, pri_t pri)
824 {
825 
826 	KASSERT(lwp_locked(l, NULL));
827 
828 	if (l->l_stat == LSRUN && (l->l_flag & LW_INMEM) != 0) {
829 		KASSERT(lwp_locked(l, l->l_cpu->ci_schedstate.spc_mutex));
830 		sched_dequeue(l);
831 		l->l_priority = pri;
832 		sched_enqueue(l, false);
833 	} else {
834 		l->l_priority = pri;
835 	}
836 	resched_cpu(l);
837 }
838 
839 static void
840 sched_lendpri(struct lwp *l, pri_t pri)
841 {
842 
843 	KASSERT(lwp_locked(l, NULL));
844 
845 	if (l->l_stat == LSRUN && (l->l_flag & LW_INMEM) != 0) {
846 		KASSERT(lwp_locked(l, l->l_cpu->ci_schedstate.spc_mutex));
847 		sched_dequeue(l);
848 		l->l_inheritedprio = pri;
849 		sched_enqueue(l, false);
850 	} else {
851 		l->l_inheritedprio = pri;
852 	}
853 	resched_cpu(l);
854 }
855 
856 struct lwp *
857 syncobj_noowner(wchan_t wchan)
858 {
859 
860 	return NULL;
861 }
862 
863 
864 /* decay 95% of `p_pctcpu' in 60 seconds; see CCPU_SHIFT before changing */
865 fixpt_t	ccpu = 0.95122942450071400909 * FSCALE;		/* exp(-1/20) */
866 
867 /*
868  * If `ccpu' is not equal to `exp(-1/20)' and you still want to use the
869  * faster/more-accurate formula, you'll have to estimate CCPU_SHIFT below
870  * and possibly adjust FSHIFT in "param.h" so that (FSHIFT >= CCPU_SHIFT).
871  *
872  * To estimate CCPU_SHIFT for exp(-1/20), the following formula was used:
873  *	1 - exp(-1/20) ~= 0.0487 ~= 0.0488 == 1 (fixed pt, *11* bits).
874  *
875  * If you dont want to bother with the faster/more-accurate formula, you
876  * can set CCPU_SHIFT to (FSHIFT + 1) which will use a slower/less-accurate
877  * (more general) method of calculating the %age of CPU used by a process.
878  */
879 #define	CCPU_SHIFT	(FSHIFT + 1)
880 
881 /*
882  * sched_pstats:
883  *
884  * Update process statistics and check CPU resource allocation.
885  * Call scheduler-specific hook to eventually adjust process/LWP
886  * priorities.
887  */
888 /* ARGSUSED */
889 void
890 sched_pstats(void *arg)
891 {
892 	struct rlimit *rlim;
893 	struct lwp *l;
894 	struct proc *p;
895 	int sig, clkhz;
896 	long runtm;
897 
898 	sched_pstats_ticks++;
899 
900 	mutex_enter(&proclist_mutex);
901 	PROCLIST_FOREACH(p, &allproc) {
902 		/*
903 		 * Increment time in/out of memory and sleep time (if
904 		 * sleeping).  We ignore overflow; with 16-bit int's
905 		 * (remember them?) overflow takes 45 days.
906 		 */
907 		mutex_enter(&p->p_smutex);
908 		mutex_spin_enter(&p->p_stmutex);
909 		runtm = p->p_rtime.tv_sec;
910 		LIST_FOREACH(l, &p->p_lwps, l_sibling) {
911 			if ((l->l_flag & LW_IDLE) != 0)
912 				continue;
913 			lwp_lock(l);
914 			runtm += l->l_rtime.tv_sec;
915 			l->l_swtime++;
916 			sched_pstats_hook(l);
917 			lwp_unlock(l);
918 
919 			/*
920 			 * p_pctcpu is only for ps.
921 			 */
922 			l->l_pctcpu = (l->l_pctcpu * ccpu) >> FSHIFT;
923 			if (l->l_slptime < 1) {
924 				clkhz = stathz != 0 ? stathz : hz;
925 #if	(FSHIFT >= CCPU_SHIFT)
926 				l->l_pctcpu += (clkhz == 100) ?
927 				    ((fixpt_t)l->l_cpticks) <<
928 				        (FSHIFT - CCPU_SHIFT) :
929 				    100 * (((fixpt_t) p->p_cpticks)
930 				        << (FSHIFT - CCPU_SHIFT)) / clkhz;
931 #else
932 				l->l_pctcpu += ((FSCALE - ccpu) *
933 				    (l->l_cpticks * FSCALE / clkhz)) >> FSHIFT;
934 #endif
935 				l->l_cpticks = 0;
936 			}
937 		}
938 		p->p_pctcpu = (p->p_pctcpu * ccpu) >> FSHIFT;
939 		mutex_spin_exit(&p->p_stmutex);
940 
941 		/*
942 		 * Check if the process exceeds its CPU resource allocation.
943 		 * If over max, kill it.
944 		 */
945 		rlim = &p->p_rlimit[RLIMIT_CPU];
946 		sig = 0;
947 		if (runtm >= rlim->rlim_cur) {
948 			if (runtm >= rlim->rlim_max)
949 				sig = SIGKILL;
950 			else {
951 				sig = SIGXCPU;
952 				if (rlim->rlim_cur < rlim->rlim_max)
953 					rlim->rlim_cur += 5;
954 			}
955 		}
956 		mutex_exit(&p->p_smutex);
957 		if (sig) {
958 			psignal(p, sig);
959 		}
960 	}
961 	mutex_exit(&proclist_mutex);
962 	uvm_meter();
963 	cv_wakeup(&lbolt);
964 	callout_schedule(&sched_pstats_ch, hz);
965 }
966 
967 void
968 sched_init(void)
969 {
970 
971 	callout_init(&sched_pstats_ch, 0);
972 	callout_setfunc(&sched_pstats_ch, sched_pstats, NULL);
973 	sched_setup();
974 	sched_pstats(NULL);
975 }
976