xref: /dflybsd-src/sys/kern/kern_synch.c (revision 9b5a99654f820c32a9fb8fa9bae3c3b12fe27a0b)
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 1.86 2007/06/08 02:02:27 dillon 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	ncpus_fit, ncpus_fit_mask;
80 int	safepri;
81 int	tsleep_now_works;
82 
83 static struct callout loadav_callout;
84 static struct callout schedcpu_callout;
85 MALLOC_DEFINE(M_TSLEEP, "tslpque", "tsleep queues");
86 
87 #if !defined(KTR_TSLEEP)
88 #define KTR_TSLEEP	KTR_ALL
89 #endif
90 KTR_INFO_MASTER(tsleep);
91 KTR_INFO(KTR_TSLEEP, tsleep, tsleep_beg, 0, "tsleep enter %p", sizeof(void *));
92 KTR_INFO(KTR_TSLEEP, tsleep, tsleep_end, 1, "tsleep exit", 0);
93 KTR_INFO(KTR_TSLEEP, tsleep, wakeup_beg, 2, "wakeup enter %p", sizeof(void *));
94 KTR_INFO(KTR_TSLEEP, tsleep, wakeup_end, 3, "wakeup exit", 0);
95 
96 #define logtsleep1(name)	KTR_LOG(tsleep_ ## name)
97 #define logtsleep2(name, val)	KTR_LOG(tsleep_ ## name, val)
98 
99 struct loadavg averunnable =
100 	{ {0, 0, 0}, FSCALE };	/* load average, of runnable procs */
101 /*
102  * Constants for averages over 1, 5, and 15 minutes
103  * when sampling at 5 second intervals.
104  */
105 static fixpt_t cexp[3] = {
106 	0.9200444146293232 * FSCALE,	/* exp(-1/12) */
107 	0.9834714538216174 * FSCALE,	/* exp(-1/60) */
108 	0.9944598480048967 * FSCALE,	/* exp(-1/180) */
109 };
110 
111 static void	endtsleep (void *);
112 static void	unsleep_and_wakeup_thread(struct thread *td);
113 static void	loadav (void *arg);
114 static void	schedcpu (void *arg);
115 
116 /*
117  * Adjust the scheduler quantum.  The quantum is specified in microseconds.
118  * Note that 'tick' is in microseconds per tick.
119  */
120 static int
121 sysctl_kern_quantum(SYSCTL_HANDLER_ARGS)
122 {
123 	int error, new_val;
124 
125 	new_val = sched_quantum * tick;
126 	error = sysctl_handle_int(oidp, &new_val, 0, req);
127         if (error != 0 || req->newptr == NULL)
128 		return (error);
129 	if (new_val < tick)
130 		return (EINVAL);
131 	sched_quantum = new_val / tick;
132 	hogticks = 2 * sched_quantum;
133 	return (0);
134 }
135 
136 SYSCTL_PROC(_kern, OID_AUTO, quantum, CTLTYPE_INT|CTLFLAG_RW,
137 	0, sizeof sched_quantum, sysctl_kern_quantum, "I", "");
138 
139 /*
140  * If `ccpu' is not equal to `exp(-1/20)' and you still want to use the
141  * faster/more-accurate formula, you'll have to estimate CCPU_SHIFT below
142  * and possibly adjust FSHIFT in "param.h" so that (FSHIFT >= CCPU_SHIFT).
143  *
144  * To estimate CCPU_SHIFT for exp(-1/20), the following formula was used:
145  *     1 - exp(-1/20) ~= 0.0487 ~= 0.0488 == 1 (fixed pt, *11* bits).
146  *
147  * If you don't want to bother with the faster/more-accurate formula, you
148  * can set CCPU_SHIFT to (FSHIFT + 1) which will use a slower/less-accurate
149  * (more general) method of calculating the %age of CPU used by a process.
150  *
151  * decay 95% of `lwp_pctcpu' in 60 seconds; see CCPU_SHIFT before changing
152  */
153 #define CCPU_SHIFT	11
154 
155 static fixpt_t ccpu = 0.95122942450071400909 * FSCALE; /* exp(-1/20) */
156 SYSCTL_INT(_kern, OID_AUTO, ccpu, CTLFLAG_RD, &ccpu, 0, "");
157 
158 /*
159  * kernel uses `FSCALE', userland (SHOULD) use kern.fscale
160  */
161 int     fscale __unused = FSCALE;	/* exported to systat */
162 SYSCTL_INT(_kern, OID_AUTO, fscale, CTLFLAG_RD, 0, FSCALE, "");
163 
164 /*
165  * Recompute process priorities, once a second.
166  *
167  * Since the userland schedulers are typically event oriented, if the
168  * estcpu calculation at wakeup() time is not sufficient to make a
169  * process runnable relative to other processes in the system we have
170  * a 1-second recalc to help out.
171  *
172  * This code also allows us to store sysclock_t data in the process structure
173  * without fear of an overrun, since sysclock_t are guarenteed to hold
174  * several seconds worth of count.
175  *
176  * WARNING!  callouts can preempt normal threads.  However, they will not
177  * preempt a thread holding a spinlock so we *can* safely use spinlocks.
178  */
179 static int schedcpu_stats(struct proc *p, void *data __unused);
180 static int schedcpu_resource(struct proc *p, void *data __unused);
181 
182 static void
183 schedcpu(void *arg)
184 {
185 	allproc_scan(schedcpu_stats, NULL);
186 	allproc_scan(schedcpu_resource, NULL);
187 	wakeup((caddr_t)&lbolt);
188 	wakeup((caddr_t)&lbolt_syncer);
189 	callout_reset(&schedcpu_callout, hz, schedcpu, NULL);
190 }
191 
192 /*
193  * General process statistics once a second
194  */
195 static int
196 schedcpu_stats(struct proc *p, void *data __unused)
197 {
198 	struct lwp *lp;
199 
200 	crit_enter();
201 	p->p_swtime++;
202 	FOREACH_LWP_IN_PROC(lp, p) {
203 		if (lp->lwp_stat == LSSLEEP)
204 			lp->lwp_slptime++;
205 
206 		/*
207 		 * Only recalculate processes that are active or have slept
208 		 * less then 2 seconds.  The schedulers understand this.
209 		 */
210 		if (lp->lwp_slptime <= 1) {
211 			p->p_usched->recalculate(lp);
212 		} else {
213 			lp->lwp_pctcpu = (lp->lwp_pctcpu * ccpu) >> FSHIFT;
214 		}
215 	}
216 	crit_exit();
217 	return(0);
218 }
219 
220 /*
221  * Resource checks.  XXX break out since ksignal/killproc can block,
222  * limiting us to one process killed per second.  There is probably
223  * a better way.
224  */
225 static int
226 schedcpu_resource(struct proc *p, void *data __unused)
227 {
228 	u_int64_t ttime;
229 	struct lwp *lp;
230 
231 	crit_enter();
232 	if (p->p_stat == SIDL ||
233 	    p->p_stat == SZOMB ||
234 	    p->p_limit == NULL
235 	) {
236 		crit_exit();
237 		return(0);
238 	}
239 
240 	ttime = 0;
241 	FOREACH_LWP_IN_PROC(lp, p) {
242 		ttime += lp->lwp_thread->td_sticks;
243 		ttime += lp->lwp_thread->td_uticks;
244 	}
245 
246 	switch(plimit_testcpulimit(p->p_limit, ttime)) {
247 	case PLIMIT_TESTCPU_KILL:
248 		killproc(p, "exceeded maximum CPU limit");
249 		break;
250 	case PLIMIT_TESTCPU_XCPU:
251 		if ((p->p_flag & P_XCPU) == 0) {
252 			p->p_flag |= P_XCPU;
253 			ksignal(p, SIGXCPU);
254 		}
255 		break;
256 	default:
257 		break;
258 	}
259 	crit_exit();
260 	return(0);
261 }
262 
263 /*
264  * This is only used by ps.  Generate a cpu percentage use over
265  * a period of one second.
266  *
267  * MPSAFE
268  */
269 void
270 updatepcpu(struct lwp *lp, int cpticks, int ttlticks)
271 {
272 	fixpt_t acc;
273 	int remticks;
274 
275 	acc = (cpticks << FSHIFT) / ttlticks;
276 	if (ttlticks >= ESTCPUFREQ) {
277 		lp->lwp_pctcpu = acc;
278 	} else {
279 		remticks = ESTCPUFREQ - ttlticks;
280 		lp->lwp_pctcpu = (acc * ttlticks + lp->lwp_pctcpu * remticks) /
281 				ESTCPUFREQ;
282 	}
283 }
284 
285 /*
286  * tsleep/wakeup hash table parameters.  Try to find the sweet spot for
287  * like addresses being slept on.
288  */
289 #define TABLESIZE	1024
290 #define LOOKUP(x)	(((intptr_t)(x) >> 6) & (TABLESIZE - 1))
291 
292 static cpumask_t slpque_cpumasks[TABLESIZE];
293 
294 /*
295  * General scheduler initialization.  We force a reschedule 25 times
296  * a second by default.  Note that cpu0 is initialized in early boot and
297  * cannot make any high level calls.
298  *
299  * Each cpu has its own sleep queue.
300  */
301 void
302 sleep_gdinit(globaldata_t gd)
303 {
304 	static struct tslpque slpque_cpu0[TABLESIZE];
305 	int i;
306 
307 	if (gd->gd_cpuid == 0) {
308 		sched_quantum = (hz + 24) / 25;
309 		hogticks = 2 * sched_quantum;
310 
311 		gd->gd_tsleep_hash = slpque_cpu0;
312 	} else {
313 		gd->gd_tsleep_hash = kmalloc(sizeof(slpque_cpu0),
314 					    M_TSLEEP, M_WAITOK | M_ZERO);
315 	}
316 	for (i = 0; i < TABLESIZE; ++i)
317 		TAILQ_INIT(&gd->gd_tsleep_hash[i]);
318 }
319 
320 /*
321  * General sleep call.  Suspends the current process until a wakeup is
322  * performed on the specified identifier.  The process will then be made
323  * runnable with the specified priority.  Sleeps at most timo/hz seconds
324  * (0 means no timeout).  If flags includes PCATCH flag, signals are checked
325  * before and after sleeping, else signals are not checked.  Returns 0 if
326  * awakened, EWOULDBLOCK if the timeout expires.  If PCATCH is set and a
327  * signal needs to be delivered, ERESTART is returned if the current system
328  * call should be restarted if possible, and EINTR is returned if the system
329  * call should be interrupted by the signal (return EINTR).
330  *
331  * Note that if we are a process, we release_curproc() before messing with
332  * the LWKT scheduler.
333  *
334  * During autoconfiguration or after a panic, a sleep will simply
335  * lower the priority briefly to allow interrupts, then return.
336  */
337 int
338 tsleep(void *ident, int flags, const char *wmesg, int timo)
339 {
340 	struct thread *td = curthread;
341 	struct lwp *lp = td->td_lwp;
342 	struct proc *p = td->td_proc;		/* may be NULL */
343 	globaldata_t gd;
344 	int sig;
345 	int catch;
346 	int id;
347 	int error;
348 	int oldpri;
349 	struct callout thandle;
350 
351 	/*
352 	 * NOTE: removed KTRPOINT, it could cause races due to blocking
353 	 * even in stable.  Just scrap it for now.
354 	 */
355 	if (tsleep_now_works == 0 || panicstr) {
356 		/*
357 		 * After a panic, or before we actually have an operational
358 		 * softclock, just give interrupts a chance, then just return;
359 		 *
360 		 * don't run any other procs or panic below,
361 		 * in case this is the idle process and already asleep.
362 		 */
363 		splz();
364 		oldpri = td->td_pri & TDPRI_MASK;
365 		lwkt_setpri_self(safepri);
366 		lwkt_switch();
367 		lwkt_setpri_self(oldpri);
368 		return (0);
369 	}
370 	logtsleep2(tsleep_beg, ident);
371 	gd = td->td_gd;
372 	KKASSERT(td != &gd->gd_idlethread);	/* you must be kidding! */
373 
374 	/*
375 	 * NOTE: all of this occurs on the current cpu, including any
376 	 * callout-based wakeups, so a critical section is a sufficient
377 	 * interlock.
378 	 *
379 	 * The entire sequence through to where we actually sleep must
380 	 * run without breaking the critical section.
381 	 */
382 	id = LOOKUP(ident);
383 	catch = flags & PCATCH;
384 	error = 0;
385 	sig = 0;
386 
387 	crit_enter_quick(td);
388 
389 	KASSERT(ident != NULL, ("tsleep: no ident"));
390 	KASSERT(lp == NULL ||
391 		lp->lwp_stat == LSRUN ||	/* Obvious */
392 		lp->lwp_stat == LSSTOP,		/* Set in tstop */
393 		("tsleep %p %s %d",
394 			ident, wmesg, lp->lwp_stat));
395 
396 	/*
397 	 * Setup for the current process (if this is a process).
398 	 */
399 	if (lp) {
400 		if (catch) {
401 			/*
402 			 * Early termination if PCATCH was set and a
403 			 * signal is pending, interlocked with the
404 			 * critical section.
405 			 *
406 			 * Early termination only occurs when tsleep() is
407 			 * entered while in a normal LSRUN state.
408 			 */
409 			if ((sig = CURSIG(lp)) != 0)
410 				goto resume;
411 
412 			/*
413 			 * Early termination if PCATCH was set and a
414 			 * mailbox signal was possibly delivered prior to
415 			 * the system call even being made, in order to
416 			 * allow the user to interlock without having to
417 			 * make additional system calls.
418 			 */
419 			if (p->p_flag & P_MAILBOX)
420 				goto resume;
421 
422 			/*
423 			 * Causes ksignal to wake us up when.
424 			 */
425 			lp->lwp_flag |= LWP_SINTR;
426 		}
427 
428 		/*
429 		 * Make sure the current process has been untangled from
430 		 * the userland scheduler and initialize slptime to start
431 		 * counting.
432 		 */
433 		if (flags & PNORESCHED)
434 			td->td_flags |= TDF_NORESCHED;
435 		p->p_usched->release_curproc(lp);
436 		lp->lwp_slptime = 0;
437 	}
438 
439 	/*
440 	 * Move our thread to the correct queue and setup our wchan, etc.
441 	 */
442 	lwkt_deschedule_self(td);
443 	td->td_flags |= TDF_TSLEEPQ;
444 	TAILQ_INSERT_TAIL(&gd->gd_tsleep_hash[id], td, td_threadq);
445 	atomic_set_int(&slpque_cpumasks[id], gd->gd_cpumask);
446 
447 	td->td_wchan = ident;
448 	td->td_wmesg = wmesg;
449 	td->td_wdomain = flags & PDOMAIN_MASK;
450 
451 	/*
452 	 * Setup the timeout, if any
453 	 */
454 	if (timo) {
455 		callout_init(&thandle);
456 		callout_reset(&thandle, timo, endtsleep, td);
457 	}
458 
459 	/*
460 	 * Beddy bye bye.
461 	 */
462 	if (lp) {
463 		/*
464 		 * Ok, we are sleeping.  Place us in the SSLEEP state.
465 		 */
466 		KKASSERT((lp->lwp_flag & LWP_ONRUNQ) == 0);
467 		/*
468 		 * tstop() sets LSSTOP, so don't fiddle with that.
469 		 */
470 		if (lp->lwp_stat != LSSTOP)
471 			lp->lwp_stat = LSSLEEP;
472 		lp->lwp_ru.ru_nvcsw++;
473 		lwkt_switch();
474 
475 		/*
476 		 * And when we are woken up, put us back in LSRUN.  If we
477 		 * slept for over a second, recalculate our estcpu.
478 		 */
479 		lp->lwp_stat = LSRUN;
480 		if (lp->lwp_slptime)
481 			p->p_usched->recalculate(lp);
482 		lp->lwp_slptime = 0;
483 	} else {
484 		lwkt_switch();
485 	}
486 
487 	/*
488 	 * Make sure we haven't switched cpus while we were asleep.  It's
489 	 * not supposed to happen.  Cleanup our temporary flags.
490 	 */
491 	KKASSERT(gd == td->td_gd);
492 	td->td_flags &= ~TDF_NORESCHED;
493 
494 	/*
495 	 * Cleanup the timeout.
496 	 */
497 	if (timo) {
498 		if (td->td_flags & TDF_TIMEOUT) {
499 			td->td_flags &= ~TDF_TIMEOUT;
500 			error = EWOULDBLOCK;
501 		} else {
502 			callout_stop(&thandle);
503 		}
504 	}
505 
506 	/*
507 	 * Since td_threadq is used both for our run queue AND for the
508 	 * tsleep hash queue, we can't still be on it at this point because
509 	 * we've gotten cpu back.
510 	 */
511 	KASSERT((td->td_flags & TDF_TSLEEPQ) == 0, ("tsleep: impossible thread flags %08x", td->td_flags));
512 	td->td_wchan = NULL;
513 	td->td_wmesg = NULL;
514 	td->td_wdomain = 0;
515 
516 	/*
517 	 * Figure out the correct error return.  If interrupted by a
518 	 * signal we want to return EINTR or ERESTART.
519 	 *
520 	 * If P_MAILBOX is set no automatic system call restart occurs
521 	 * and we return EINTR.  P_MAILBOX is meant to be used as an
522 	 * interlock, the user must poll it prior to any system call
523 	 * that it wishes to interlock a mailbox signal against since
524 	 * the flag is cleared on *any* system call that sleeps.
525 	 */
526 resume:
527 	if (p) {
528 		if (catch && error == 0) {
529 			if ((p->p_flag & P_MAILBOX) && sig == 0) {
530 				error = EINTR;
531 			} else if (sig != 0 || (sig = CURSIG(lp))) {
532 				if (SIGISMEMBER(p->p_sigacts->ps_sigintr, sig))
533 					error = EINTR;
534 				else
535 					error = ERESTART;
536 			}
537 		}
538 		lp->lwp_flag &= ~(LWP_BREAKTSLEEP | LWP_SINTR);
539 		p->p_flag &= ~P_MAILBOX;
540 	}
541 	logtsleep1(tsleep_end);
542 	crit_exit_quick(td);
543 	return (error);
544 }
545 
546 /*
547  * This is a dandy function that allows us to interlock tsleep/wakeup
548  * operations with unspecified upper level locks, such as lockmgr locks,
549  * simply by holding a critical section.  The sequence is:
550  *
551  *	(enter critical section)
552  *	(acquire upper level lock)
553  *	tsleep_interlock(blah)
554  *	(release upper level lock)
555  *	tsleep(blah, ...)
556  *	(exit critical section)
557  *
558  * Basically this function sets our cpumask for the ident which informs
559  * other cpus that our cpu 'might' be waiting (or about to wait on) the
560  * hash index related to the ident.  The critical section prevents another
561  * cpu's wakeup() from being processed on our cpu until we are actually
562  * able to enter the tsleep().  Thus, no race occurs between our attempt
563  * to release a resource and sleep, and another cpu's attempt to acquire
564  * a resource and call wakeup.
565  *
566  * There isn't much of a point to this function unless you call it while
567  * holding a critical section.
568  */
569 static __inline void
570 _tsleep_interlock(globaldata_t gd, void *ident)
571 {
572 	int id = LOOKUP(ident);
573 
574 	atomic_set_int(&slpque_cpumasks[id], gd->gd_cpumask);
575 }
576 
577 void
578 tsleep_interlock(void *ident)
579 {
580 	_tsleep_interlock(mycpu, ident);
581 }
582 
583 /*
584  * Interlocked spinlock sleep.  An exclusively held spinlock must
585  * be passed to msleep().  The function will atomically release the
586  * spinlock and tsleep on the ident, then reacquire the spinlock and
587  * return.
588  *
589  * This routine is fairly important along the critical path, so optimize it
590  * heavily.
591  */
592 int
593 msleep(void *ident, struct spinlock *spin, int flags,
594        const char *wmesg, int timo)
595 {
596 	globaldata_t gd = mycpu;
597 	int error;
598 
599 	crit_enter_gd(gd);
600 	_tsleep_interlock(gd, ident);
601 	spin_unlock_wr_quick(gd, spin);
602 	error = tsleep(ident, flags, wmesg, timo);
603 	spin_lock_wr_quick(gd, spin);
604 	crit_exit_gd(gd);
605 
606 	return (error);
607 }
608 
609 /*
610  * Directly block on the LWKT thread by descheduling it.  This
611  * is much faster then tsleep(), but the only legal way to wake
612  * us up is to directly schedule the thread.
613  *
614  * Setting TDF_SINTR will cause new signals to directly schedule us.
615  *
616  * This routine is typically called while in a critical section.
617  */
618 int
619 lwkt_sleep(const char *wmesg, int flags)
620 {
621 	thread_t td = curthread;
622 	int sig;
623 
624 	if ((flags & PCATCH) == 0 || td->td_lwp == NULL) {
625 		td->td_flags |= TDF_BLOCKED;
626 		td->td_wmesg = wmesg;
627 		lwkt_deschedule_self(td);
628 		lwkt_switch();
629 		td->td_wmesg = NULL;
630 		td->td_flags &= ~TDF_BLOCKED;
631 		return(0);
632 	}
633 	if ((sig = CURSIG(td->td_lwp)) != 0) {
634 		if (SIGISMEMBER(td->td_proc->p_sigacts->ps_sigintr, sig))
635 			return(EINTR);
636 		else
637 			return(ERESTART);
638 
639 	}
640 	td->td_flags |= TDF_BLOCKED | TDF_SINTR;
641 	td->td_wmesg = wmesg;
642 	lwkt_deschedule_self(td);
643 	lwkt_switch();
644 	td->td_flags &= ~(TDF_BLOCKED | TDF_SINTR);
645 	td->td_wmesg = NULL;
646 	return(0);
647 }
648 
649 /*
650  * Implement the timeout for tsleep.
651  *
652  * We set LWP_BREAKTSLEEP to indicate that an event has occured, but
653  * we only call setrunnable if the process is not stopped.
654  *
655  * This type of callout timeout is scheduled on the same cpu the process
656  * is sleeping on.  Also, at the moment, the MP lock is held.
657  */
658 static void
659 endtsleep(void *arg)
660 {
661 	thread_t td = arg;
662 	struct lwp *lp;
663 
664 	ASSERT_MP_LOCK_HELD(curthread);
665 	crit_enter();
666 
667 	/*
668 	 * cpu interlock.  Thread flags are only manipulated on
669 	 * the cpu owning the thread.  proc flags are only manipulated
670 	 * by the older of the MP lock.  We have both.
671 	 */
672 	if (td->td_flags & TDF_TSLEEPQ) {
673 		td->td_flags |= TDF_TIMEOUT;
674 
675 		if ((lp = td->td_lwp) != NULL) {
676 			lp->lwp_flag |= LWP_BREAKTSLEEP;
677 			if (lp->lwp_proc->p_stat != SSTOP)
678 				setrunnable(lp);
679 		} else {
680 			unsleep_and_wakeup_thread(td);
681 		}
682 	}
683 	crit_exit();
684 }
685 
686 /*
687  * Unsleep and wakeup a thread.  This function runs without the MP lock
688  * which means that it can only manipulate thread state on the owning cpu,
689  * and cannot touch the process state at all.
690  */
691 static
692 void
693 unsleep_and_wakeup_thread(struct thread *td)
694 {
695 	globaldata_t gd = mycpu;
696 	int id;
697 
698 #ifdef SMP
699 	if (td->td_gd != gd) {
700 		lwkt_send_ipiq(td->td_gd, (ipifunc1_t)unsleep_and_wakeup_thread, td);
701 		return;
702 	}
703 #endif
704 	crit_enter();
705 	if (td->td_flags & TDF_TSLEEPQ) {
706 		td->td_flags &= ~TDF_TSLEEPQ;
707 		id = LOOKUP(td->td_wchan);
708 		TAILQ_REMOVE(&gd->gd_tsleep_hash[id], td, td_threadq);
709 		if (TAILQ_FIRST(&gd->gd_tsleep_hash[id]) == NULL)
710 			atomic_clear_int(&slpque_cpumasks[id], gd->gd_cpumask);
711 		lwkt_schedule(td);
712 	}
713 	crit_exit();
714 }
715 
716 /*
717  * Make all processes sleeping on the specified identifier runnable.
718  * count may be zero or one only.
719  *
720  * The domain encodes the sleep/wakeup domain AND the first cpu to check
721  * (which is always the current cpu).  As we iterate across cpus
722  *
723  * This call may run without the MP lock held.  We can only manipulate thread
724  * state on the cpu owning the thread.  We CANNOT manipulate process state
725  * at all.
726  */
727 static void
728 _wakeup(void *ident, int domain)
729 {
730 	struct tslpque *qp;
731 	struct thread *td;
732 	struct thread *ntd;
733 	globaldata_t gd;
734 #ifdef SMP
735 	cpumask_t mask;
736 	cpumask_t tmask;
737 	int startcpu;
738 	int nextcpu;
739 #endif
740 	int id;
741 
742 	crit_enter();
743 	logtsleep2(wakeup_beg, ident);
744 	gd = mycpu;
745 	id = LOOKUP(ident);
746 	qp = &gd->gd_tsleep_hash[id];
747 restart:
748 	for (td = TAILQ_FIRST(qp); td != NULL; td = ntd) {
749 		ntd = TAILQ_NEXT(td, td_threadq);
750 		if (td->td_wchan == ident &&
751 		    td->td_wdomain == (domain & PDOMAIN_MASK)
752 		) {
753 			KKASSERT(td->td_flags & TDF_TSLEEPQ);
754 			td->td_flags &= ~TDF_TSLEEPQ;
755 			TAILQ_REMOVE(qp, td, td_threadq);
756 			if (TAILQ_FIRST(qp) == NULL) {
757 				atomic_clear_int(&slpque_cpumasks[id],
758 						 gd->gd_cpumask);
759 			}
760 			lwkt_schedule(td);
761 			if (domain & PWAKEUP_ONE)
762 				goto done;
763 			goto restart;
764 		}
765 	}
766 
767 #ifdef SMP
768 	/*
769 	 * We finished checking the current cpu but there still may be
770 	 * more work to do.  Either wakeup_one was requested and no matching
771 	 * thread was found, or a normal wakeup was requested and we have
772 	 * to continue checking cpus.
773 	 *
774 	 * The cpu that started the wakeup sequence is encoded in the domain.
775 	 * We use this information to determine which cpus still need to be
776 	 * checked, locate a candidate cpu, and chain the wakeup
777 	 * asynchronously with an IPI message.
778 	 *
779 	 * It should be noted that this scheme is actually less expensive then
780 	 * the old scheme when waking up multiple threads, since we send
781 	 * only one IPI message per target candidate which may then schedule
782 	 * multiple threads.  Before we could have wound up sending an IPI
783 	 * message for each thread on the target cpu (!= current cpu) that
784 	 * needed to be woken up.
785 	 *
786 	 * NOTE: Wakeups occuring on remote cpus are asynchronous.  This
787 	 * should be ok since we are passing idents in the IPI rather then
788 	 * thread pointers.
789 	 */
790 	if ((domain & PWAKEUP_MYCPU) == 0 &&
791 	    (mask = slpque_cpumasks[id]) != 0
792 	) {
793 		/*
794 		 * Look for a cpu that might have work to do.  Mask out cpus
795 		 * which have already been processed.
796 		 *
797 		 * 31xxxxxxxxxxxxxxxxxxxxxxxxxxxxx0
798 		 *        ^        ^           ^
799 		 *      start   currentcpu    start
800 		 *      case2                 case1
801 		 *        *        *           *
802 		 * 11111111111111110000000000000111	case1
803 		 * 00000000111111110000000000000000	case2
804 		 *
805 		 * case1:  We started at start_case1 and processed through
806 		 *  	   to the current cpu.  We have to check any bits
807 		 *	   after the current cpu, then check bits before
808 		 *         the starting cpu.
809 		 *
810 		 * case2:  We have already checked all the bits from
811 		 *         start_case2 to the end, and from 0 to the current
812 		 *         cpu.  We just have the bits from the current cpu
813 		 *         to start_case2 left to check.
814 		 */
815 		startcpu = PWAKEUP_DECODE(domain);
816 		if (gd->gd_cpuid >= startcpu) {
817 			/*
818 			 * CASE1
819 			 */
820 			tmask = mask & ~((gd->gd_cpumask << 1) - 1);
821 			if (mask & tmask) {
822 				nextcpu = bsfl(mask & tmask);
823 				lwkt_send_ipiq2(globaldata_find(nextcpu),
824 						_wakeup, ident, domain);
825 			} else {
826 				tmask = (1 << startcpu) - 1;
827 				if (mask & tmask) {
828 					nextcpu = bsfl(mask & tmask);
829 					lwkt_send_ipiq2(
830 						    globaldata_find(nextcpu),
831 						    _wakeup, ident, domain);
832 				}
833 			}
834 		} else {
835 			/*
836 			 * CASE2
837 			 */
838 			tmask = ~((gd->gd_cpumask << 1) - 1) &
839 				 ((1 << startcpu) - 1);
840 			if (mask & tmask) {
841 				nextcpu = bsfl(mask & tmask);
842 				lwkt_send_ipiq2(globaldata_find(nextcpu),
843 						_wakeup, ident, domain);
844 			}
845 		}
846 	}
847 #endif
848 done:
849 	logtsleep1(wakeup_end);
850 	crit_exit();
851 }
852 
853 /*
854  * Wakeup all threads tsleep()ing on the specified ident, on all cpus
855  */
856 void
857 wakeup(void *ident)
858 {
859     _wakeup(ident, PWAKEUP_ENCODE(0, mycpu->gd_cpuid));
860 }
861 
862 /*
863  * Wakeup one thread tsleep()ing on the specified ident, on any cpu.
864  */
865 void
866 wakeup_one(void *ident)
867 {
868     /* XXX potentially round-robin the first responding cpu */
869     _wakeup(ident, PWAKEUP_ENCODE(0, mycpu->gd_cpuid) | PWAKEUP_ONE);
870 }
871 
872 /*
873  * Wakeup threads tsleep()ing on the specified ident on the current cpu
874  * only.
875  */
876 void
877 wakeup_mycpu(void *ident)
878 {
879     _wakeup(ident, PWAKEUP_MYCPU);
880 }
881 
882 /*
883  * Wakeup one thread tsleep()ing on the specified ident on the current cpu
884  * only.
885  */
886 void
887 wakeup_mycpu_one(void *ident)
888 {
889     /* XXX potentially round-robin the first responding cpu */
890     _wakeup(ident, PWAKEUP_MYCPU|PWAKEUP_ONE);
891 }
892 
893 /*
894  * Wakeup all thread tsleep()ing on the specified ident on the specified cpu
895  * only.
896  */
897 void
898 wakeup_oncpu(globaldata_t gd, void *ident)
899 {
900 #ifdef SMP
901     if (gd == mycpu) {
902 	_wakeup(ident, PWAKEUP_MYCPU);
903     } else {
904 	lwkt_send_ipiq2(gd, _wakeup, ident, PWAKEUP_MYCPU);
905     }
906 #else
907     _wakeup(ident, PWAKEUP_MYCPU);
908 #endif
909 }
910 
911 /*
912  * Wakeup one thread tsleep()ing on the specified ident on the specified cpu
913  * only.
914  */
915 void
916 wakeup_oncpu_one(globaldata_t gd, void *ident)
917 {
918 #ifdef SMP
919     if (gd == mycpu) {
920 	_wakeup(ident, PWAKEUP_MYCPU | PWAKEUP_ONE);
921     } else {
922 	lwkt_send_ipiq2(gd, _wakeup, ident, PWAKEUP_MYCPU | PWAKEUP_ONE);
923     }
924 #else
925     _wakeup(ident, PWAKEUP_MYCPU | PWAKEUP_ONE);
926 #endif
927 }
928 
929 /*
930  * Wakeup all threads waiting on the specified ident that slept using
931  * the specified domain, on all cpus.
932  */
933 void
934 wakeup_domain(void *ident, int domain)
935 {
936     _wakeup(ident, PWAKEUP_ENCODE(domain, mycpu->gd_cpuid));
937 }
938 
939 /*
940  * Wakeup one thread waiting on the specified ident that slept using
941  * the specified  domain, on any cpu.
942  */
943 void
944 wakeup_domain_one(void *ident, int domain)
945 {
946     /* XXX potentially round-robin the first responding cpu */
947     _wakeup(ident, PWAKEUP_ENCODE(domain, mycpu->gd_cpuid) | PWAKEUP_ONE);
948 }
949 
950 /*
951  * setrunnable()
952  *
953  * Make a process runnable.  The MP lock must be held on call.  This only
954  * has an effect if we are in SSLEEP.  We only break out of the
955  * tsleep if LWP_BREAKTSLEEP is set, otherwise we just fix-up the state.
956  *
957  * NOTE: With the MP lock held we can only safely manipulate the process
958  * structure.  We cannot safely manipulate the thread structure.
959  */
960 void
961 setrunnable(struct lwp *lp)
962 {
963 	crit_enter();
964 	ASSERT_MP_LOCK_HELD(curthread);
965 	if (lp->lwp_stat == LSSTOP)
966 		lp->lwp_stat = LSSLEEP;
967 	if (lp->lwp_stat == LSSLEEP && (lp->lwp_flag & LWP_BREAKTSLEEP))
968 		unsleep_and_wakeup_thread(lp->lwp_thread);
969 	crit_exit();
970 }
971 
972 /*
973  * The process is stopped due to some condition, usually because p_stat is
974  * set to SSTOP, but also possibly due to being traced.
975  *
976  * NOTE!  If the caller sets SSTOP, the caller must also clear P_WAITED
977  * because the parent may check the child's status before the child actually
978  * gets to this routine.
979  *
980  * This routine is called with the current lwp only, typically just
981  * before returning to userland.
982  *
983  * Setting LWP_BREAKTSLEEP before entering the tsleep will cause a passive
984  * SIGCONT to break out of the tsleep.
985  */
986 void
987 tstop(void)
988 {
989 	struct lwp *lp = curthread->td_lwp;
990 	struct proc *p = lp->lwp_proc;
991 
992 	lp->lwp_flag |= LWP_BREAKTSLEEP;
993 	lp->lwp_stat = LSSTOP;
994 	crit_enter();
995 	/*
996 	 * If LWP_WSTOP is set, we were sleeping
997 	 * while our process was stopped.  At this point
998 	 * we were already counted as stopped.
999 	 */
1000 	if ((lp->lwp_flag & LWP_WSTOP) == 0) {
1001 		/*
1002 		 * If we're the last thread to stop, signal
1003 		 * our parent.
1004 		 */
1005 		p->p_nstopped++;
1006 		lp->lwp_flag |= LWP_WSTOP;
1007 		if (p->p_nstopped == p->p_nthreads) {
1008 			p->p_flag &= ~P_WAITED;
1009 			wakeup(p->p_pptr);
1010 			if ((p->p_pptr->p_sigacts->ps_flag & PS_NOCLDSTOP) == 0)
1011 				ksignal(p->p_pptr, SIGCHLD);
1012 		}
1013 	}
1014 	tsleep(lp->lwp_proc, 0, "stop", 0);
1015 	p->p_nstopped--;
1016 	crit_exit();
1017 }
1018 
1019 /*
1020  * Yield / synchronous reschedule.  This is a bit tricky because the trap
1021  * code might have set a lazy release on the switch function.   Setting
1022  * P_PASSIVE_ACQ will ensure that the lazy release executes when we call
1023  * switch, and that we are given a greater chance of affinity with our
1024  * current cpu.
1025  *
1026  * We call lwkt_setpri_self() to rotate our thread to the end of the lwkt
1027  * run queue.  lwkt_switch() will also execute any assigned passive release
1028  * (which usually calls release_curproc()), allowing a same/higher priority
1029  * process to be designated as the current process.
1030  *
1031  * While it is possible for a lower priority process to be designated,
1032  * it's call to lwkt_maybe_switch() in acquire_curproc() will likely
1033  * round-robin back to us and we will be able to re-acquire the current
1034  * process designation.
1035  */
1036 void
1037 uio_yield(void)
1038 {
1039 	struct thread *td = curthread;
1040 	struct proc *p = td->td_proc;
1041 
1042 	lwkt_setpri_self(td->td_pri & TDPRI_MASK);
1043 	if (p) {
1044 		p->p_flag |= P_PASSIVE_ACQ;
1045 		lwkt_switch();
1046 		p->p_flag &= ~P_PASSIVE_ACQ;
1047 	} else {
1048 		lwkt_switch();
1049 	}
1050 }
1051 
1052 /*
1053  * Compute a tenex style load average of a quantity on
1054  * 1, 5 and 15 minute intervals.
1055  */
1056 static int loadav_count_runnable(struct lwp *p, void *data);
1057 
1058 static void
1059 loadav(void *arg)
1060 {
1061 	struct loadavg *avg;
1062 	int i, nrun;
1063 
1064 	nrun = 0;
1065 	alllwp_scan(loadav_count_runnable, &nrun);
1066 	avg = &averunnable;
1067 	for (i = 0; i < 3; i++) {
1068 		avg->ldavg[i] = (cexp[i] * avg->ldavg[i] +
1069 		    nrun * FSCALE * (FSCALE - cexp[i])) >> FSHIFT;
1070 	}
1071 
1072 	/*
1073 	 * Schedule the next update to occur after 5 seconds, but add a
1074 	 * random variation to avoid synchronisation with processes that
1075 	 * run at regular intervals.
1076 	 */
1077 	callout_reset(&loadav_callout, hz * 4 + (int)(krandom() % (hz * 2 + 1)),
1078 		      loadav, NULL);
1079 }
1080 
1081 static int
1082 loadav_count_runnable(struct lwp *lp, void *data)
1083 {
1084 	int *nrunp = data;
1085 	thread_t td;
1086 
1087 	switch (lp->lwp_stat) {
1088 	case LSRUN:
1089 		if ((td = lp->lwp_thread) == NULL)
1090 			break;
1091 		if (td->td_flags & TDF_BLOCKED)
1092 			break;
1093 		++*nrunp;
1094 		break;
1095 	default:
1096 		break;
1097 	}
1098 	return(0);
1099 }
1100 
1101 /* ARGSUSED */
1102 static void
1103 sched_setup(void *dummy)
1104 {
1105 	callout_init(&loadav_callout);
1106 	callout_init(&schedcpu_callout);
1107 
1108 	/* Kick off timeout driven events by calling first time. */
1109 	schedcpu(NULL);
1110 	loadav(NULL);
1111 }
1112 
1113