xref: /dflybsd-src/sys/kern/kern_synch.c (revision 269ffd40655c28ffc04742b158e7500aec5b0166)
1 /*-
2  * Copyright (c) 1982, 1986, 1990, 1991, 1993
3  *	The Regents of the University of California.  All rights reserved.
4  * (c) UNIX System Laboratories, Inc.
5  * All or some portions of this file are derived from material licensed
6  * to the University of California by American Telephone and Telegraph
7  * Co. or Unix System Laboratories, Inc. and are reproduced herein with
8  * the permission of UNIX System Laboratories, Inc.
9  *
10  * Redistribution and use in source and binary forms, with or without
11  * modification, are permitted provided that the following conditions
12  * are met:
13  * 1. Redistributions of source code must retain the above copyright
14  *    notice, this list of conditions and the following disclaimer.
15  * 2. Redistributions in binary form must reproduce the above copyright
16  *    notice, this list of conditions and the following disclaimer in the
17  *    documentation and/or other materials provided with the distribution.
18  * 3. All advertising materials mentioning features or use of this software
19  *    must display the following acknowledgement:
20  *	This product includes software developed by the University of
21  *	California, Berkeley and its contributors.
22  * 4. Neither the name of the University nor the names of its contributors
23  *    may be used to endorse or promote products derived from this software
24  *    without specific prior written permission.
25  *
26  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
27  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
28  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
29  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
30  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
31  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
32  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
33  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
34  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
35  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
36  * SUCH DAMAGE.
37  *
38  *	@(#)kern_synch.c	8.9 (Berkeley) 5/19/95
39  * $FreeBSD: src/sys/kern/kern_synch.c,v 1.87.2.6 2002/10/13 07:29:53 kbyanc Exp $
40  * $DragonFly: src/sys/kern/kern_synch.c,v 1.82 2007/03/12 21:08:15 corecode Exp $
41  */
42 
43 #include "opt_ktrace.h"
44 
45 #include <sys/param.h>
46 #include <sys/systm.h>
47 #include <sys/proc.h>
48 #include <sys/kernel.h>
49 #include <sys/signalvar.h>
50 #include <sys/signal2.h>
51 #include <sys/resourcevar.h>
52 #include <sys/vmmeter.h>
53 #include <sys/sysctl.h>
54 #include <sys/lock.h>
55 #ifdef KTRACE
56 #include <sys/uio.h>
57 #include <sys/ktrace.h>
58 #endif
59 #include <sys/xwait.h>
60 #include <sys/ktr.h>
61 
62 #include <sys/thread2.h>
63 #include <sys/spinlock2.h>
64 
65 #include <machine/cpu.h>
66 #include <machine/smp.h>
67 
68 TAILQ_HEAD(tslpque, thread);
69 
70 static void sched_setup (void *dummy);
71 SYSINIT(sched_setup, SI_SUB_KICK_SCHEDULER, SI_ORDER_FIRST, sched_setup, NULL)
72 
73 int	hogticks;
74 int	lbolt;
75 int	lbolt_syncer;
76 int	sched_quantum;		/* Roundrobin scheduling quantum in ticks. */
77 int	ncpus;
78 int	ncpus2, ncpus2_shift, ncpus2_mask;
79 int	safepri;
80 
81 static struct callout loadav_callout;
82 static struct callout schedcpu_callout;
83 MALLOC_DEFINE(M_TSLEEP, "tslpque", "tsleep queues");
84 
85 #if !defined(KTR_TSLEEP)
86 #define KTR_TSLEEP	KTR_ALL
87 #endif
88 KTR_INFO_MASTER(tsleep);
89 KTR_INFO(KTR_TSLEEP, tsleep, tsleep_beg, 0, "tsleep enter", 0);
90 KTR_INFO(KTR_TSLEEP, tsleep, tsleep_end, 0, "tsleep exit", 0);
91 KTR_INFO(KTR_TSLEEP, tsleep, wakeup_beg, 0, "wakeup enter", 0);
92 KTR_INFO(KTR_TSLEEP, tsleep, wakeup_end, 0, "wakeup exit", 0);
93 #define logtsleep(name)	KTR_LOG(tsleep_ ## name)
94 
95 struct loadavg averunnable =
96 	{ {0, 0, 0}, FSCALE };	/* load average, of runnable procs */
97 /*
98  * Constants for averages over 1, 5, and 15 minutes
99  * when sampling at 5 second intervals.
100  */
101 static fixpt_t cexp[3] = {
102 	0.9200444146293232 * FSCALE,	/* exp(-1/12) */
103 	0.9834714538216174 * FSCALE,	/* exp(-1/60) */
104 	0.9944598480048967 * FSCALE,	/* exp(-1/180) */
105 };
106 
107 static void	endtsleep (void *);
108 static void	unsleep_and_wakeup_thread(struct thread *td);
109 static void	loadav (void *arg);
110 static void	schedcpu (void *arg);
111 
112 /*
113  * Adjust the scheduler quantum.  The quantum is specified in microseconds.
114  * Note that 'tick' is in microseconds per tick.
115  */
116 static int
117 sysctl_kern_quantum(SYSCTL_HANDLER_ARGS)
118 {
119 	int error, new_val;
120 
121 	new_val = sched_quantum * tick;
122 	error = sysctl_handle_int(oidp, &new_val, 0, req);
123         if (error != 0 || req->newptr == NULL)
124 		return (error);
125 	if (new_val < tick)
126 		return (EINVAL);
127 	sched_quantum = new_val / tick;
128 	hogticks = 2 * sched_quantum;
129 	return (0);
130 }
131 
132 SYSCTL_PROC(_kern, OID_AUTO, quantum, CTLTYPE_INT|CTLFLAG_RW,
133 	0, sizeof sched_quantum, sysctl_kern_quantum, "I", "");
134 
135 /*
136  * If `ccpu' is not equal to `exp(-1/20)' and you still want to use the
137  * faster/more-accurate formula, you'll have to estimate CCPU_SHIFT below
138  * and possibly adjust FSHIFT in "param.h" so that (FSHIFT >= CCPU_SHIFT).
139  *
140  * To estimate CCPU_SHIFT for exp(-1/20), the following formula was used:
141  *     1 - exp(-1/20) ~= 0.0487 ~= 0.0488 == 1 (fixed pt, *11* bits).
142  *
143  * If you don't want to bother with the faster/more-accurate formula, you
144  * can set CCPU_SHIFT to (FSHIFT + 1) which will use a slower/less-accurate
145  * (more general) method of calculating the %age of CPU used by a process.
146  *
147  * decay 95% of `lwp_pctcpu' in 60 seconds; see CCPU_SHIFT before changing
148  */
149 #define CCPU_SHIFT	11
150 
151 static fixpt_t ccpu = 0.95122942450071400909 * FSCALE; /* exp(-1/20) */
152 SYSCTL_INT(_kern, OID_AUTO, ccpu, CTLFLAG_RD, &ccpu, 0, "");
153 
154 /*
155  * kernel uses `FSCALE', userland (SHOULD) use kern.fscale
156  */
157 int     fscale __unused = FSCALE;	/* exported to systat */
158 SYSCTL_INT(_kern, OID_AUTO, fscale, CTLFLAG_RD, 0, FSCALE, "");
159 
160 /*
161  * Recompute process priorities, once a second.
162  *
163  * Since the userland schedulers are typically event oriented, if the
164  * estcpu calculation at wakeup() time is not sufficient to make a
165  * process runnable relative to other processes in the system we have
166  * a 1-second recalc to help out.
167  *
168  * This code also allows us to store sysclock_t data in the process structure
169  * without fear of an overrun, since sysclock_t are guarenteed to hold
170  * several seconds worth of count.
171  *
172  * WARNING!  callouts can preempt normal threads.  However, they will not
173  * preempt a thread holding a spinlock so we *can* safely use spinlocks.
174  */
175 static int schedcpu_stats(struct proc *p, void *data __unused);
176 static int schedcpu_resource(struct proc *p, void *data __unused);
177 
178 static void
179 schedcpu(void *arg)
180 {
181 	allproc_scan(schedcpu_stats, NULL);
182 	allproc_scan(schedcpu_resource, NULL);
183 	wakeup((caddr_t)&lbolt);
184 	wakeup((caddr_t)&lbolt_syncer);
185 	callout_reset(&schedcpu_callout, hz, schedcpu, NULL);
186 }
187 
188 /*
189  * General process statistics once a second
190  */
191 static int
192 schedcpu_stats(struct proc *p, void *data __unused)
193 {
194 	struct lwp *lp;
195 
196 	crit_enter();
197 	p->p_swtime++;
198 	FOREACH_LWP_IN_PROC(lp, p) {
199 		if (lp->lwp_stat == LSSLEEP)
200 			lp->lwp_slptime++;
201 
202 		/*
203 		 * Only recalculate processes that are active or have slept
204 		 * less then 2 seconds.  The schedulers understand this.
205 		 */
206 		if (lp->lwp_slptime <= 1) {
207 			p->p_usched->recalculate(lp);
208 		} else {
209 			lp->lwp_pctcpu = (lp->lwp_pctcpu * ccpu) >> FSHIFT;
210 		}
211 	}
212 	crit_exit();
213 	return(0);
214 }
215 
216 /*
217  * Resource checks.  XXX break out since ksignal/killproc can block,
218  * limiting us to one process killed per second.  There is probably
219  * a better way.
220  */
221 static int
222 schedcpu_resource(struct proc *p, void *data __unused)
223 {
224 	u_int64_t ttime;
225 	struct lwp *lp;
226 
227 	crit_enter();
228 	if (p->p_stat == SIDL ||
229 	    p->p_stat == SZOMB ||
230 	    p->p_limit == NULL
231 	) {
232 		crit_exit();
233 		return(0);
234 	}
235 
236 	ttime = 0;
237 	FOREACH_LWP_IN_PROC(lp, p) {
238 		ttime += lp->lwp_thread->td_sticks;
239 		ttime += lp->lwp_thread->td_uticks;
240 	}
241 
242 	switch(plimit_testcpulimit(p->p_limit, ttime)) {
243 	case PLIMIT_TESTCPU_KILL:
244 		killproc(p, "exceeded maximum CPU limit");
245 		break;
246 	case PLIMIT_TESTCPU_XCPU:
247 		if ((p->p_flag & P_XCPU) == 0) {
248 			p->p_flag |= P_XCPU;
249 			ksignal(p, SIGXCPU);
250 		}
251 		break;
252 	default:
253 		break;
254 	}
255 	crit_exit();
256 	return(0);
257 }
258 
259 /*
260  * This is only used by ps.  Generate a cpu percentage use over
261  * a period of one second.
262  *
263  * MPSAFE
264  */
265 void
266 updatepcpu(struct lwp *lp, int cpticks, int ttlticks)
267 {
268 	fixpt_t acc;
269 	int remticks;
270 
271 	acc = (cpticks << FSHIFT) / ttlticks;
272 	if (ttlticks >= ESTCPUFREQ) {
273 		lp->lwp_pctcpu = acc;
274 	} else {
275 		remticks = ESTCPUFREQ - ttlticks;
276 		lp->lwp_pctcpu = (acc * ttlticks + lp->lwp_pctcpu * remticks) /
277 				ESTCPUFREQ;
278 	}
279 }
280 
281 /*
282  * We're only looking at 7 bits of the address; everything is
283  * aligned to 4, lots of things are aligned to greater powers
284  * of 2.  Shift right by 8, i.e. drop the bottom 256 worth.
285  */
286 #define TABLESIZE	128
287 #define LOOKUP(x)	(((intptr_t)(x) >> 8) & (TABLESIZE - 1))
288 
289 static cpumask_t slpque_cpumasks[TABLESIZE];
290 
291 /*
292  * General scheduler initialization.  We force a reschedule 25 times
293  * a second by default.  Note that cpu0 is initialized in early boot and
294  * cannot make any high level calls.
295  *
296  * Each cpu has its own sleep queue.
297  */
298 void
299 sleep_gdinit(globaldata_t gd)
300 {
301 	static struct tslpque slpque_cpu0[TABLESIZE];
302 	int i;
303 
304 	if (gd->gd_cpuid == 0) {
305 		sched_quantum = (hz + 24) / 25;
306 		hogticks = 2 * sched_quantum;
307 
308 		gd->gd_tsleep_hash = slpque_cpu0;
309 	} else {
310 		gd->gd_tsleep_hash = kmalloc(sizeof(slpque_cpu0),
311 					    M_TSLEEP, M_WAITOK | M_ZERO);
312 	}
313 	for (i = 0; i < TABLESIZE; ++i)
314 		TAILQ_INIT(&gd->gd_tsleep_hash[i]);
315 }
316 
317 /*
318  * General sleep call.  Suspends the current process until a wakeup is
319  * performed on the specified identifier.  The process will then be made
320  * runnable with the specified priority.  Sleeps at most timo/hz seconds
321  * (0 means no timeout).  If flags includes PCATCH flag, signals are checked
322  * before and after sleeping, else signals are not checked.  Returns 0 if
323  * awakened, EWOULDBLOCK if the timeout expires.  If PCATCH is set and a
324  * signal needs to be delivered, ERESTART is returned if the current system
325  * call should be restarted if possible, and EINTR is returned if the system
326  * call should be interrupted by the signal (return EINTR).
327  *
328  * Note that if we are a process, we release_curproc() before messing with
329  * the LWKT scheduler.
330  *
331  * During autoconfiguration or after a panic, a sleep will simply
332  * lower the priority briefly to allow interrupts, then return.
333  */
334 int
335 tsleep(void *ident, int flags, const char *wmesg, int timo)
336 {
337 	struct thread *td = curthread;
338 	struct lwp *lp = td->td_lwp;
339 	struct proc *p = td->td_proc;		/* may be NULL */
340 	globaldata_t gd;
341 	int sig;
342 	int catch;
343 	int id;
344 	int error;
345 	int oldpri;
346 	struct callout thandle;
347 
348 	/*
349 	 * NOTE: removed KTRPOINT, it could cause races due to blocking
350 	 * even in stable.  Just scrap it for now.
351 	 */
352 	if (cold || panicstr) {
353 		/*
354 		 * After a panic, or during autoconfiguration,
355 		 * just give interrupts a chance, then just return;
356 		 * don't run any other procs or panic below,
357 		 * in case this is the idle process and already asleep.
358 		 */
359 		splz();
360 		oldpri = td->td_pri & TDPRI_MASK;
361 		lwkt_setpri_self(safepri);
362 		lwkt_switch();
363 		lwkt_setpri_self(oldpri);
364 		return (0);
365 	}
366 	logtsleep(tsleep_beg);
367 	gd = td->td_gd;
368 	KKASSERT(td != &gd->gd_idlethread);	/* you must be kidding! */
369 
370 	/*
371 	 * NOTE: all of this occurs on the current cpu, including any
372 	 * callout-based wakeups, so a critical section is a sufficient
373 	 * interlock.
374 	 *
375 	 * The entire sequence through to where we actually sleep must
376 	 * run without breaking the critical section.
377 	 */
378 	id = LOOKUP(ident);
379 	catch = flags & PCATCH;
380 	error = 0;
381 	sig = 0;
382 
383 	crit_enter_quick(td);
384 
385 	KASSERT(ident != NULL, ("tsleep: no ident"));
386 	KASSERT(lp == NULL ||
387 		lp->lwp_stat == LSRUN ||	/* Obvious */
388 		lp->lwp_stat == LSSTOP,		/* Set in tstop */
389 		("tsleep %p %s %d",
390 			ident, wmesg, lp->lwp_stat));
391 
392 	/*
393 	 * Setup for the current process (if this is a process).
394 	 */
395 	if (lp) {
396 		if (catch) {
397 			/*
398 			 * Early termination if PCATCH was set and a
399 			 * signal is pending, interlocked with the
400 			 * critical section.
401 			 *
402 			 * Early termination only occurs when tsleep() is
403 			 * entered while in a normal LSRUN state.
404 			 */
405 			if ((sig = CURSIG(lp)) != 0)
406 				goto resume;
407 
408 			/*
409 			 * Early termination if PCATCH was set and a
410 			 * mailbox signal was possibly delivered prior to
411 			 * the system call even being made, in order to
412 			 * allow the user to interlock without having to
413 			 * make additional system calls.
414 			 */
415 			if (p->p_flag & P_MAILBOX)
416 				goto resume;
417 
418 			/*
419 			 * Causes ksignal to wake us up when.
420 			 */
421 			lp->lwp_flag |= LWP_SINTR;
422 		}
423 
424 		/*
425 		 * Make sure the current process has been untangled from
426 		 * the userland scheduler and initialize slptime to start
427 		 * counting.
428 		 */
429 		if (flags & PNORESCHED)
430 			td->td_flags |= TDF_NORESCHED;
431 		p->p_usched->release_curproc(lp);
432 		lp->lwp_slptime = 0;
433 	}
434 
435 	/*
436 	 * Move our thread to the correct queue and setup our wchan, etc.
437 	 */
438 	lwkt_deschedule_self(td);
439 	td->td_flags |= TDF_TSLEEPQ;
440 	TAILQ_INSERT_TAIL(&gd->gd_tsleep_hash[id], td, td_threadq);
441 	atomic_set_int(&slpque_cpumasks[id], gd->gd_cpumask);
442 
443 	td->td_wchan = ident;
444 	td->td_wmesg = wmesg;
445 	td->td_wdomain = flags & PDOMAIN_MASK;
446 
447 	/*
448 	 * Setup the timeout, if any
449 	 */
450 	if (timo) {
451 		callout_init(&thandle);
452 		callout_reset(&thandle, timo, endtsleep, td);
453 	}
454 
455 	/*
456 	 * Beddy bye bye.
457 	 */
458 	if (lp) {
459 		/*
460 		 * Ok, we are sleeping.  Place us in the SSLEEP state.
461 		 */
462 		KKASSERT((lp->lwp_flag & LWP_ONRUNQ) == 0);
463 		/*
464 		 * tstop() sets LSSTOP, so don't fiddle with that.
465 		 */
466 		if (lp->lwp_stat != LSSTOP)
467 			lp->lwp_stat = LSSLEEP;
468 		lp->lwp_ru.ru_nvcsw++;
469 		lwkt_switch();
470 
471 		/*
472 		 * And when we are woken up, put us back in LSRUN.  If we
473 		 * slept for over a second, recalculate our estcpu.
474 		 */
475 		lp->lwp_stat = LSRUN;
476 		if (lp->lwp_slptime)
477 			p->p_usched->recalculate(lp);
478 		lp->lwp_slptime = 0;
479 	} else {
480 		lwkt_switch();
481 	}
482 
483 	/*
484 	 * Make sure we haven't switched cpus while we were asleep.  It's
485 	 * not supposed to happen.  Cleanup our temporary flags.
486 	 */
487 	KKASSERT(gd == td->td_gd);
488 	td->td_flags &= ~TDF_NORESCHED;
489 
490 	/*
491 	 * Cleanup the timeout.
492 	 */
493 	if (timo) {
494 		if (td->td_flags & TDF_TIMEOUT) {
495 			td->td_flags &= ~TDF_TIMEOUT;
496 			error = EWOULDBLOCK;
497 		} else {
498 			callout_stop(&thandle);
499 		}
500 	}
501 
502 	/*
503 	 * Since td_threadq is used both for our run queue AND for the
504 	 * tsleep hash queue, we can't still be on it at this point because
505 	 * we've gotten cpu back.
506 	 */
507 	KASSERT((td->td_flags & TDF_TSLEEPQ) == 0, ("tsleep: impossible thread flags %08x", td->td_flags));
508 	td->td_wchan = NULL;
509 	td->td_wmesg = NULL;
510 	td->td_wdomain = 0;
511 
512 	/*
513 	 * Figure out the correct error return.  If interrupted by a
514 	 * signal we want to return EINTR or ERESTART.
515 	 *
516 	 * If P_MAILBOX is set no automatic system call restart occurs
517 	 * and we return EINTR.  P_MAILBOX is meant to be used as an
518 	 * interlock, the user must poll it prior to any system call
519 	 * that it wishes to interlock a mailbox signal against since
520 	 * the flag is cleared on *any* system call that sleeps.
521 	 */
522 resume:
523 	if (p) {
524 		if (catch && error == 0) {
525 			if ((p->p_flag & P_MAILBOX) && sig == 0) {
526 				error = EINTR;
527 			} else if (sig != 0 || (sig = CURSIG(lp))) {
528 				if (SIGISMEMBER(p->p_sigacts->ps_sigintr, sig))
529 					error = EINTR;
530 				else
531 					error = ERESTART;
532 			}
533 		}
534 		lp->lwp_flag &= ~(LWP_BREAKTSLEEP | LWP_SINTR);
535 		p->p_flag &= ~P_MAILBOX;
536 	}
537 	logtsleep(tsleep_end);
538 	crit_exit_quick(td);
539 	return (error);
540 }
541 
542 /*
543  * This is a dandy function that allows us to interlock tsleep/wakeup
544  * operations with unspecified upper level locks, such as lockmgr locks,
545  * simply by holding a critical section.  The sequence is:
546  *
547  *	(enter critical section)
548  *	(acquire upper level lock)
549  *	tsleep_interlock(blah)
550  *	(release upper level lock)
551  *	tsleep(blah, ...)
552  *	(exit critical section)
553  *
554  * Basically this function sets our cpumask for the ident which informs
555  * other cpus that our cpu 'might' be waiting (or about to wait on) the
556  * hash index related to the ident.  The critical section prevents another
557  * cpu's wakeup() from being processed on our cpu until we are actually
558  * able to enter the tsleep().  Thus, no race occurs between our attempt
559  * to release a resource and sleep, and another cpu's attempt to acquire
560  * a resource and call wakeup.
561  *
562  * There isn't much of a point to this function unless you call it while
563  * holding a critical section.
564  */
565 static __inline void
566 _tsleep_interlock(globaldata_t gd, void *ident)
567 {
568 	int id = LOOKUP(ident);
569 
570 	atomic_set_int(&slpque_cpumasks[id], gd->gd_cpumask);
571 }
572 
573 void
574 tsleep_interlock(void *ident)
575 {
576 	_tsleep_interlock(mycpu, ident);
577 }
578 
579 /*
580  * Interlocked spinlock sleep.  An exclusively held spinlock must
581  * be passed to msleep().  The function will atomically release the
582  * spinlock and tsleep on the ident, then reacquire the spinlock and
583  * return.
584  *
585  * This routine is fairly important along the critical path, so optimize it
586  * heavily.
587  */
588 int
589 msleep(void *ident, struct spinlock *spin, int flags,
590        const char *wmesg, int timo)
591 {
592 	globaldata_t gd = mycpu;
593 	int error;
594 
595 	crit_enter_gd(gd);
596 	_tsleep_interlock(gd, ident);
597 	spin_unlock_wr_quick(gd, spin);
598 	error = tsleep(ident, flags, wmesg, timo);
599 	spin_lock_wr_quick(gd, spin);
600 	crit_exit_gd(gd);
601 
602 	return (error);
603 }
604 
605 /*
606  * Implement the timeout for tsleep.
607  *
608  * We set LWP_BREAKTSLEEP to indicate that an event has occured, but
609  * we only call setrunnable if the process is not stopped.
610  *
611  * This type of callout timeout is scheduled on the same cpu the process
612  * is sleeping on.  Also, at the moment, the MP lock is held.
613  */
614 static void
615 endtsleep(void *arg)
616 {
617 	thread_t td = arg;
618 	struct lwp *lp;
619 
620 	ASSERT_MP_LOCK_HELD(curthread);
621 	crit_enter();
622 
623 	/*
624 	 * cpu interlock.  Thread flags are only manipulated on
625 	 * the cpu owning the thread.  proc flags are only manipulated
626 	 * by the older of the MP lock.  We have both.
627 	 */
628 	if (td->td_flags & TDF_TSLEEPQ) {
629 		td->td_flags |= TDF_TIMEOUT;
630 
631 		if ((lp = td->td_lwp) != NULL) {
632 			lp->lwp_flag |= LWP_BREAKTSLEEP;
633 			if (lp->lwp_proc->p_stat != SSTOP)
634 				setrunnable(lp);
635 		} else {
636 			unsleep_and_wakeup_thread(td);
637 		}
638 	}
639 	crit_exit();
640 }
641 
642 /*
643  * Unsleep and wakeup a thread.  This function runs without the MP lock
644  * which means that it can only manipulate thread state on the owning cpu,
645  * and cannot touch the process state at all.
646  */
647 static
648 void
649 unsleep_and_wakeup_thread(struct thread *td)
650 {
651 	globaldata_t gd = mycpu;
652 	int id;
653 
654 #ifdef SMP
655 	if (td->td_gd != gd) {
656 		lwkt_send_ipiq(td->td_gd, (ipifunc1_t)unsleep_and_wakeup_thread, td);
657 		return;
658 	}
659 #endif
660 	crit_enter();
661 	if (td->td_flags & TDF_TSLEEPQ) {
662 		td->td_flags &= ~TDF_TSLEEPQ;
663 		id = LOOKUP(td->td_wchan);
664 		TAILQ_REMOVE(&gd->gd_tsleep_hash[id], td, td_threadq);
665 		if (TAILQ_FIRST(&gd->gd_tsleep_hash[id]) == NULL)
666 			atomic_clear_int(&slpque_cpumasks[id], gd->gd_cpumask);
667 		lwkt_schedule(td);
668 	}
669 	crit_exit();
670 }
671 
672 /*
673  * Make all processes sleeping on the specified identifier runnable.
674  * count may be zero or one only.
675  *
676  * The domain encodes the sleep/wakeup domain AND the first cpu to check
677  * (which is always the current cpu).  As we iterate across cpus
678  *
679  * This call may run without the MP lock held.  We can only manipulate thread
680  * state on the cpu owning the thread.  We CANNOT manipulate process state
681  * at all.
682  */
683 static void
684 _wakeup(void *ident, int domain)
685 {
686 	struct tslpque *qp;
687 	struct thread *td;
688 	struct thread *ntd;
689 	globaldata_t gd;
690 #ifdef SMP
691 	cpumask_t mask;
692 	cpumask_t tmask;
693 	int startcpu;
694 	int nextcpu;
695 #endif
696 	int id;
697 
698 	crit_enter();
699 	logtsleep(wakeup_beg);
700 	gd = mycpu;
701 	id = LOOKUP(ident);
702 	qp = &gd->gd_tsleep_hash[id];
703 restart:
704 	for (td = TAILQ_FIRST(qp); td != NULL; td = ntd) {
705 		ntd = TAILQ_NEXT(td, td_threadq);
706 		if (td->td_wchan == ident &&
707 		    td->td_wdomain == (domain & PDOMAIN_MASK)
708 		) {
709 			KKASSERT(td->td_flags & TDF_TSLEEPQ);
710 			td->td_flags &= ~TDF_TSLEEPQ;
711 			TAILQ_REMOVE(qp, td, td_threadq);
712 			if (TAILQ_FIRST(qp) == NULL) {
713 				atomic_clear_int(&slpque_cpumasks[id],
714 						 gd->gd_cpumask);
715 			}
716 			lwkt_schedule(td);
717 			if (domain & PWAKEUP_ONE)
718 				goto done;
719 			goto restart;
720 		}
721 	}
722 
723 #ifdef SMP
724 	/*
725 	 * We finished checking the current cpu but there still may be
726 	 * more work to do.  Either wakeup_one was requested and no matching
727 	 * thread was found, or a normal wakeup was requested and we have
728 	 * to continue checking cpus.
729 	 *
730 	 * The cpu that started the wakeup sequence is encoded in the domain.
731 	 * We use this information to determine which cpus still need to be
732 	 * checked, locate a candidate cpu, and chain the wakeup
733 	 * asynchronously with an IPI message.
734 	 *
735 	 * It should be noted that this scheme is actually less expensive then
736 	 * the old scheme when waking up multiple threads, since we send
737 	 * only one IPI message per target candidate which may then schedule
738 	 * multiple threads.  Before we could have wound up sending an IPI
739 	 * message for each thread on the target cpu (!= current cpu) that
740 	 * needed to be woken up.
741 	 *
742 	 * NOTE: Wakeups occuring on remote cpus are asynchronous.  This
743 	 * should be ok since we are passing idents in the IPI rather then
744 	 * thread pointers.
745 	 */
746 	if ((domain & PWAKEUP_MYCPU) == 0 &&
747 	    (mask = slpque_cpumasks[id]) != 0
748 	) {
749 		/*
750 		 * Look for a cpu that might have work to do.  Mask out cpus
751 		 * which have already been processed.
752 		 *
753 		 * 31xxxxxxxxxxxxxxxxxxxxxxxxxxxxx0
754 		 *        ^        ^           ^
755 		 *      start   currentcpu    start
756 		 *      case2                 case1
757 		 *        *        *           *
758 		 * 11111111111111110000000000000111	case1
759 		 * 00000000111111110000000000000000	case2
760 		 *
761 		 * case1:  We started at start_case1 and processed through
762 		 *  	   to the current cpu.  We have to check any bits
763 		 *	   after the current cpu, then check bits before
764 		 *         the starting cpu.
765 		 *
766 		 * case2:  We have already checked all the bits from
767 		 *         start_case2 to the end, and from 0 to the current
768 		 *         cpu.  We just have the bits from the current cpu
769 		 *         to start_case2 left to check.
770 		 */
771 		startcpu = PWAKEUP_DECODE(domain);
772 		if (gd->gd_cpuid >= startcpu) {
773 			/*
774 			 * CASE1
775 			 */
776 			tmask = mask & ~((gd->gd_cpumask << 1) - 1);
777 			if (mask & tmask) {
778 				nextcpu = bsfl(mask & tmask);
779 				lwkt_send_ipiq2(globaldata_find(nextcpu),
780 						_wakeup, ident, domain);
781 			} else {
782 				tmask = (1 << startcpu) - 1;
783 				if (mask & tmask) {
784 					nextcpu = bsfl(mask & tmask);
785 					lwkt_send_ipiq2(
786 						    globaldata_find(nextcpu),
787 						    _wakeup, ident, domain);
788 				}
789 			}
790 		} else {
791 			/*
792 			 * CASE2
793 			 */
794 			tmask = ~((gd->gd_cpumask << 1) - 1) &
795 				 ((1 << startcpu) - 1);
796 			if (mask & tmask) {
797 				nextcpu = bsfl(mask & tmask);
798 				lwkt_send_ipiq2(globaldata_find(nextcpu),
799 						_wakeup, ident, domain);
800 			}
801 		}
802 	}
803 #endif
804 done:
805 	logtsleep(wakeup_end);
806 	crit_exit();
807 }
808 
809 /*
810  * Wakeup all threads tsleep()ing on the specified ident, on all cpus
811  */
812 void
813 wakeup(void *ident)
814 {
815     _wakeup(ident, PWAKEUP_ENCODE(0, mycpu->gd_cpuid));
816 }
817 
818 /*
819  * Wakeup one thread tsleep()ing on the specified ident, on any cpu.
820  */
821 void
822 wakeup_one(void *ident)
823 {
824     /* XXX potentially round-robin the first responding cpu */
825     _wakeup(ident, PWAKEUP_ENCODE(0, mycpu->gd_cpuid) | PWAKEUP_ONE);
826 }
827 
828 /*
829  * Wakeup threads tsleep()ing on the specified ident on the current cpu
830  * only.
831  */
832 void
833 wakeup_mycpu(void *ident)
834 {
835     _wakeup(ident, PWAKEUP_MYCPU);
836 }
837 
838 /*
839  * Wakeup one thread tsleep()ing on the specified ident on the current cpu
840  * only.
841  */
842 void
843 wakeup_mycpu_one(void *ident)
844 {
845     /* XXX potentially round-robin the first responding cpu */
846     _wakeup(ident, PWAKEUP_MYCPU|PWAKEUP_ONE);
847 }
848 
849 /*
850  * Wakeup all thread tsleep()ing on the specified ident on the specified cpu
851  * only.
852  */
853 void
854 wakeup_oncpu(globaldata_t gd, void *ident)
855 {
856 #ifdef SMP
857     if (gd == mycpu) {
858 	_wakeup(ident, PWAKEUP_MYCPU);
859     } else {
860 	lwkt_send_ipiq2(gd, _wakeup, ident, PWAKEUP_MYCPU);
861     }
862 #else
863     _wakeup(ident, PWAKEUP_MYCPU);
864 #endif
865 }
866 
867 /*
868  * Wakeup one thread tsleep()ing on the specified ident on the specified cpu
869  * only.
870  */
871 void
872 wakeup_oncpu_one(globaldata_t gd, void *ident)
873 {
874 #ifdef SMP
875     if (gd == mycpu) {
876 	_wakeup(ident, PWAKEUP_MYCPU | PWAKEUP_ONE);
877     } else {
878 	lwkt_send_ipiq2(gd, _wakeup, ident, PWAKEUP_MYCPU | PWAKEUP_ONE);
879     }
880 #else
881     _wakeup(ident, PWAKEUP_MYCPU | PWAKEUP_ONE);
882 #endif
883 }
884 
885 /*
886  * Wakeup all threads waiting on the specified ident that slept using
887  * the specified domain, on all cpus.
888  */
889 void
890 wakeup_domain(void *ident, int domain)
891 {
892     _wakeup(ident, PWAKEUP_ENCODE(domain, mycpu->gd_cpuid));
893 }
894 
895 /*
896  * Wakeup one thread waiting on the specified ident that slept using
897  * the specified  domain, on any cpu.
898  */
899 void
900 wakeup_domain_one(void *ident, int domain)
901 {
902     /* XXX potentially round-robin the first responding cpu */
903     _wakeup(ident, PWAKEUP_ENCODE(domain, mycpu->gd_cpuid) | PWAKEUP_ONE);
904 }
905 
906 /*
907  * setrunnable()
908  *
909  * Make a process runnable.  The MP lock must be held on call.  This only
910  * has an effect if we are in SSLEEP.  We only break out of the
911  * tsleep if LWP_BREAKTSLEEP is set, otherwise we just fix-up the state.
912  *
913  * NOTE: With the MP lock held we can only safely manipulate the process
914  * structure.  We cannot safely manipulate the thread structure.
915  */
916 void
917 setrunnable(struct lwp *lp)
918 {
919 	crit_enter();
920 	ASSERT_MP_LOCK_HELD(curthread);
921 	if (lp->lwp_stat == LSSTOP)
922 		lp->lwp_stat = LSSLEEP;
923 	if (lp->lwp_stat == LSSLEEP && (lp->lwp_flag & LWP_BREAKTSLEEP))
924 		unsleep_and_wakeup_thread(lp->lwp_thread);
925 	crit_exit();
926 }
927 
928 /*
929  * The process is stopped due to some condition, usually because p_stat is
930  * set to SSTOP, but also possibly due to being traced.
931  *
932  * NOTE!  If the caller sets SSTOP, the caller must also clear P_WAITED
933  * because the parent may check the child's status before the child actually
934  * gets to this routine.
935  *
936  * This routine is called with the current lwp only, typically just
937  * before returning to userland.
938  *
939  * Setting LWP_BREAKTSLEEP before entering the tsleep will cause a passive
940  * SIGCONT to break out of the tsleep.
941  */
942 void
943 tstop(void)
944 {
945 	struct lwp *lp = curthread->td_lwp;
946 	struct proc *p = lp->lwp_proc;
947 
948 	lp->lwp_flag |= LWP_BREAKTSLEEP;
949 	lp->lwp_stat = LSSTOP;
950 	crit_enter();
951 	/*
952 	 * If LWP_WSTOP is set, we were sleeping
953 	 * while our process was stopped.  At this point
954 	 * we were already counted as stopped.
955 	 */
956 	if ((lp->lwp_flag & LWP_WSTOP) == 0) {
957 		/*
958 		 * If we're the last thread to stop, signal
959 		 * our parent.
960 		 */
961 		p->p_nstopped++;
962 		lp->lwp_flag |= LWP_WSTOP;
963 		if (p->p_nstopped == p->p_nthreads) {
964 			p->p_flag &= ~P_WAITED;
965 			wakeup(p->p_pptr);
966 			if ((p->p_pptr->p_sigacts->ps_flag & PS_NOCLDSTOP) == 0)
967 				ksignal(p->p_pptr, SIGCHLD);
968 		}
969 	}
970 	tsleep(lp->lwp_proc, 0, "stop", 0);
971 	p->p_nstopped--;
972 	crit_exit();
973 }
974 
975 /*
976  * Yield / synchronous reschedule.  This is a bit tricky because the trap
977  * code might have set a lazy release on the switch function.   Setting
978  * P_PASSIVE_ACQ will ensure that the lazy release executes when we call
979  * switch, and that we are given a greater chance of affinity with our
980  * current cpu.
981  *
982  * We call lwkt_setpri_self() to rotate our thread to the end of the lwkt
983  * run queue.  lwkt_switch() will also execute any assigned passive release
984  * (which usually calls release_curproc()), allowing a same/higher priority
985  * process to be designated as the current process.
986  *
987  * While it is possible for a lower priority process to be designated,
988  * it's call to lwkt_maybe_switch() in acquire_curproc() will likely
989  * round-robin back to us and we will be able to re-acquire the current
990  * process designation.
991  */
992 void
993 uio_yield(void)
994 {
995 	struct thread *td = curthread;
996 	struct proc *p = td->td_proc;
997 
998 	lwkt_setpri_self(td->td_pri & TDPRI_MASK);
999 	if (p) {
1000 		p->p_flag |= P_PASSIVE_ACQ;
1001 		lwkt_switch();
1002 		p->p_flag &= ~P_PASSIVE_ACQ;
1003 	} else {
1004 		lwkt_switch();
1005 	}
1006 }
1007 
1008 /*
1009  * Compute a tenex style load average of a quantity on
1010  * 1, 5 and 15 minute intervals.
1011  */
1012 static int loadav_count_runnable(struct lwp *p, void *data);
1013 
1014 static void
1015 loadav(void *arg)
1016 {
1017 	struct loadavg *avg;
1018 	int i, nrun;
1019 
1020 	nrun = 0;
1021 	alllwp_scan(loadav_count_runnable, &nrun);
1022 	avg = &averunnable;
1023 	for (i = 0; i < 3; i++) {
1024 		avg->ldavg[i] = (cexp[i] * avg->ldavg[i] +
1025 		    nrun * FSCALE * (FSCALE - cexp[i])) >> FSHIFT;
1026 	}
1027 
1028 	/*
1029 	 * Schedule the next update to occur after 5 seconds, but add a
1030 	 * random variation to avoid synchronisation with processes that
1031 	 * run at regular intervals.
1032 	 */
1033 	callout_reset(&loadav_callout, hz * 4 + (int)(krandom() % (hz * 2 + 1)),
1034 		      loadav, NULL);
1035 }
1036 
1037 static int
1038 loadav_count_runnable(struct lwp *lp, void *data)
1039 {
1040 	int *nrunp = data;
1041 	thread_t td;
1042 
1043 	switch (lp->lwp_stat) {
1044 	case LSRUN:
1045 		if ((td = lp->lwp_thread) == NULL)
1046 			break;
1047 		if (td->td_flags & TDF_BLOCKED)
1048 			break;
1049 		++*nrunp;
1050 		break;
1051 	default:
1052 		break;
1053 	}
1054 	return(0);
1055 }
1056 
1057 /* ARGSUSED */
1058 static void
1059 sched_setup(void *dummy)
1060 {
1061 	callout_init(&loadav_callout);
1062 	callout_init(&schedcpu_callout);
1063 
1064 	/* Kick off timeout driven events by calling first time. */
1065 	schedcpu(NULL);
1066 	loadav(NULL);
1067 }
1068 
1069