xref: /dflybsd-src/sys/kern/kern_synch.c (revision a9cfaf7ce72f185c8b4450485c9415aa0c648bfe)
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 1.63 2006/05/29 03:57:20 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/resourcevar.h>
51 #include <sys/vmmeter.h>
52 #include <sys/sysctl.h>
53 #include <sys/lock.h>
54 #ifdef KTRACE
55 #include <sys/uio.h>
56 #include <sys/ktrace.h>
57 #endif
58 #include <sys/xwait.h>
59 #include <sys/ktr.h>
60 
61 #include <sys/thread2.h>
62 #include <sys/spinlock2.h>
63 
64 #include <machine/cpu.h>
65 #include <machine/ipl.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 `p_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 static int     fscale __unused = FSCALE;
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 	crit_enter();
195 	p->p_swtime++;
196 	if (p->p_stat == SSLEEP)
197 		p->p_slptime++;
198 
199 	/*
200 	 * Only recalculate processes that are active or have slept
201 	 * less then 2 seconds.  The schedulers understand this.
202 	 */
203 	if (p->p_slptime <= 1) {
204 		p->p_usched->recalculate(&p->p_lwp);
205 	} else {
206 		p->p_pctcpu = (p->p_pctcpu * ccpu) >> FSHIFT;
207 	}
208 	crit_exit();
209 	return(0);
210 }
211 
212 /*
213  * Resource checks.  XXX break out since psignal/killproc can block,
214  * limiting us to one process killed per second.  There is probably
215  * a better way.
216  */
217 static int
218 schedcpu_resource(struct proc *p, void *data __unused)
219 {
220 	u_int64_t ttime;
221 
222 	crit_enter();
223 	if (p->p_stat == SIDL ||
224 	    (p->p_flag & P_ZOMBIE) ||
225 	    p->p_limit == NULL ||
226 	    p->p_thread == NULL
227 	) {
228 		crit_exit();
229 		return(0);
230 	}
231 
232 	ttime = p->p_thread->td_sticks + p->p_thread->td_uticks;
233 
234 	switch(plimit_testcpulimit(p->p_limit, ttime)) {
235 	case PLIMIT_TESTCPU_KILL:
236 		killproc(p, "exceeded maximum CPU limit");
237 		break;
238 	case PLIMIT_TESTCPU_XCPU:
239 		if ((p->p_flag & P_XCPU) == 0) {
240 			p->p_flag |= P_XCPU;
241 			psignal(p, SIGXCPU);
242 		}
243 		break;
244 	default:
245 		break;
246 	}
247 	crit_exit();
248 	return(0);
249 }
250 
251 /*
252  * This is only used by ps.  Generate a cpu percentage use over
253  * a period of one second.
254  *
255  * MPSAFE
256  */
257 void
258 updatepcpu(struct lwp *lp, int cpticks, int ttlticks)
259 {
260 	fixpt_t acc;
261 	int remticks;
262 
263 	acc = (cpticks << FSHIFT) / ttlticks;
264 	if (ttlticks >= ESTCPUFREQ) {
265 		lp->lwp_pctcpu = acc;
266 	} else {
267 		remticks = ESTCPUFREQ - ttlticks;
268 		lp->lwp_pctcpu = (acc * ttlticks + lp->lwp_pctcpu * remticks) /
269 				ESTCPUFREQ;
270 	}
271 }
272 
273 /*
274  * We're only looking at 7 bits of the address; everything is
275  * aligned to 4, lots of things are aligned to greater powers
276  * of 2.  Shift right by 8, i.e. drop the bottom 256 worth.
277  */
278 #define TABLESIZE	128
279 #define LOOKUP(x)	(((intptr_t)(x) >> 8) & (TABLESIZE - 1))
280 
281 static cpumask_t slpque_cpumasks[TABLESIZE];
282 
283 /*
284  * General scheduler initialization.  We force a reschedule 25 times
285  * a second by default.  Note that cpu0 is initialized in early boot and
286  * cannot make any high level calls.
287  *
288  * Each cpu has its own sleep queue.
289  */
290 void
291 sleep_gdinit(globaldata_t gd)
292 {
293 	static struct tslpque slpque_cpu0[TABLESIZE];
294 	int i;
295 
296 	if (gd->gd_cpuid == 0) {
297 		sched_quantum = (hz + 24) / 25;
298 		hogticks = 2 * sched_quantum;
299 
300 		gd->gd_tsleep_hash = slpque_cpu0;
301 	} else {
302 		gd->gd_tsleep_hash = malloc(sizeof(slpque_cpu0),
303 					    M_TSLEEP, M_WAITOK | M_ZERO);
304 	}
305 	for (i = 0; i < TABLESIZE; ++i)
306 		TAILQ_INIT(&gd->gd_tsleep_hash[i]);
307 }
308 
309 /*
310  * General sleep call.  Suspends the current process until a wakeup is
311  * performed on the specified identifier.  The process will then be made
312  * runnable with the specified priority.  Sleeps at most timo/hz seconds
313  * (0 means no timeout).  If flags includes PCATCH flag, signals are checked
314  * before and after sleeping, else signals are not checked.  Returns 0 if
315  * awakened, EWOULDBLOCK if the timeout expires.  If PCATCH is set and a
316  * signal needs to be delivered, ERESTART is returned if the current system
317  * call should be restarted if possible, and EINTR is returned if the system
318  * call should be interrupted by the signal (return EINTR).
319  *
320  * Note that if we are a process, we release_curproc() before messing with
321  * the LWKT scheduler.
322  *
323  * During autoconfiguration or after a panic, a sleep will simply
324  * lower the priority briefly to allow interrupts, then return.
325  */
326 int
327 tsleep(void *ident, int flags, const char *wmesg, int timo)
328 {
329 	struct thread *td = curthread;
330 	struct proc *p = td->td_proc;		/* may be NULL */
331 	globaldata_t gd;
332 	int sig;
333 	int catch;
334 	int id;
335 	int error;
336 	int oldpri;
337 	struct callout thandle;
338 
339 	/*
340 	 * NOTE: removed KTRPOINT, it could cause races due to blocking
341 	 * even in stable.  Just scrap it for now.
342 	 */
343 	if (cold || panicstr) {
344 		/*
345 		 * After a panic, or during autoconfiguration,
346 		 * just give interrupts a chance, then just return;
347 		 * don't run any other procs or panic below,
348 		 * in case this is the idle process and already asleep.
349 		 */
350 		splz();
351 		oldpri = td->td_pri & TDPRI_MASK;
352 		lwkt_setpri_self(safepri);
353 		lwkt_switch();
354 		lwkt_setpri_self(oldpri);
355 		return (0);
356 	}
357 	logtsleep(tsleep_beg);
358 	gd = td->td_gd;
359 	KKASSERT(td != &gd->gd_idlethread);	/* you must be kidding! */
360 
361 	/*
362 	 * NOTE: all of this occurs on the current cpu, including any
363 	 * callout-based wakeups, so a critical section is a sufficient
364 	 * interlock.
365 	 *
366 	 * The entire sequence through to where we actually sleep must
367 	 * run without breaking the critical section.
368 	 */
369 	id = LOOKUP(ident);
370 	catch = flags & PCATCH;
371 	error = 0;
372 	sig = 0;
373 
374 	crit_enter_quick(td);
375 
376 	KASSERT(ident != NULL, ("tsleep: no ident"));
377 	KASSERT(p == NULL || p->p_stat == SRUN, ("tsleep %p %s %d",
378 		ident, wmesg, p->p_stat));
379 
380 	/*
381 	 * Setup for the current process (if this is a process).
382 	 */
383 	if (p) {
384 		if (catch) {
385 			/*
386 			 * Early termination if PCATCH was set and a
387 			 * signal is pending, interlocked with the
388 			 * critical section.
389 			 *
390 			 * Early termination only occurs when tsleep() is
391 			 * entered while in a normal SRUN state.
392 			 */
393 			if ((sig = CURSIG(p)) != 0)
394 				goto resume;
395 
396 			/*
397 			 * Causes psignal to wake us up when.
398 			 */
399 			p->p_flag |= P_SINTR;
400 		}
401 
402 		/*
403 		 * Make sure the current process has been untangled from
404 		 * the userland scheduler and initialize slptime to start
405 		 * counting.
406 		 */
407 		if (flags & PNORESCHED)
408 			td->td_flags |= TDF_NORESCHED;
409 		p->p_usched->release_curproc(&p->p_lwp);
410 		p->p_slptime = 0;
411 	}
412 
413 	/*
414 	 * Move our thread to the correct queue and setup our wchan, etc.
415 	 */
416 	lwkt_deschedule_self(td);
417 	td->td_flags |= TDF_TSLEEPQ;
418 	TAILQ_INSERT_TAIL(&gd->gd_tsleep_hash[id], td, td_threadq);
419 	atomic_set_int(&slpque_cpumasks[id], gd->gd_cpumask);
420 
421 	td->td_wchan = ident;
422 	td->td_wmesg = wmesg;
423 	td->td_wdomain = flags & PDOMAIN_MASK;
424 
425 	/*
426 	 * Setup the timeout, if any
427 	 */
428 	if (timo) {
429 		callout_init(&thandle);
430 		callout_reset(&thandle, timo, endtsleep, td);
431 	}
432 
433 	/*
434 	 * Beddy bye bye.
435 	 */
436 	if (p) {
437 		/*
438 		 * Ok, we are sleeping.  Place us in the SSLEEP state.
439 		 */
440 		KKASSERT((p->p_flag & P_ONRUNQ) == 0);
441 		p->p_stat = SSLEEP;
442 		p->p_stats->p_ru.ru_nvcsw++;
443 		lwkt_switch();
444 		p->p_stat = SRUN;
445 	} else {
446 		lwkt_switch();
447 	}
448 
449 	/*
450 	 * Make sure we haven't switched cpus while we were asleep.  It's
451 	 * not supposed to happen.  Cleanup our temporary flags.
452 	 */
453 	KKASSERT(gd == td->td_gd);
454 	td->td_flags &= ~TDF_NORESCHED;
455 
456 	/*
457 	 * Cleanup the timeout.
458 	 */
459 	if (timo) {
460 		if (td->td_flags & TDF_TIMEOUT) {
461 			td->td_flags &= ~TDF_TIMEOUT;
462 			if (sig == 0)
463 				error = EWOULDBLOCK;
464 		} else {
465 			callout_stop(&thandle);
466 		}
467 	}
468 
469 	/*
470 	 * Since td_threadq is used both for our run queue AND for the
471 	 * tsleep hash queue, we can't still be on it at this point because
472 	 * we've gotten cpu back.
473 	 */
474 	KASSERT((td->td_flags & TDF_TSLEEPQ) == 0, ("tsleep: impossible thread flags %08x", td->td_flags));
475 	td->td_wchan = NULL;
476 	td->td_wmesg = NULL;
477 	td->td_wdomain = 0;
478 
479 	/*
480 	 * Figure out the correct error return
481 	 */
482 resume:
483 	if (p) {
484 		p->p_flag &= ~(P_BREAKTSLEEP | P_SINTR);
485 		if (catch && error == 0 && (sig != 0 || (sig = CURSIG(p)))) {
486 			if (SIGISMEMBER(p->p_sigacts->ps_sigintr, sig))
487 				error = EINTR;
488 			else
489 				error = ERESTART;
490 		}
491 	}
492 	logtsleep(tsleep_end);
493 	crit_exit_quick(td);
494 	return (error);
495 }
496 
497 /*
498  * This is a dandy function that allows us to interlock tsleep/wakeup
499  * operations with unspecified upper level locks, such as lockmgr locks,
500  * simply by holding a critical section.  The sequence is:
501  *
502  *	(enter critical section)
503  *	(acquire upper level lock)
504  *	tsleep_interlock(blah)
505  *	(release upper level lock)
506  *	tsleep(blah, ...)
507  *	(exit critical section)
508  *
509  * Basically this function sets our cpumask for the ident which informs
510  * other cpus that our cpu 'might' be waiting (or about to wait on) the
511  * hash index related to the ident.  The critical section prevents another
512  * cpu's wakeup() from being processed on our cpu until we are actually
513  * able to enter the tsleep().  Thus, no race occurs between our attempt
514  * to release a resource and sleep, and another cpu's attempt to acquire
515  * a resource and call wakeup.
516  *
517  * There isn't much of a point to this function unless you call it while
518  * holding a critical section.
519  */
520 static __inline void
521 _tsleep_interlock(globaldata_t gd, void *ident)
522 {
523 	int id = LOOKUP(ident);
524 
525 	atomic_set_int(&slpque_cpumasks[id], gd->gd_cpumask);
526 }
527 
528 void
529 tsleep_interlock(void *ident)
530 {
531 	_tsleep_interlock(mycpu, ident);
532 }
533 
534 /*
535  * Interlocked spinlock sleep.  An exclusively held spinlock must
536  * be passed to msleep().  The function will atomically release the
537  * spinlock and tsleep on the ident, then reacquire the spinlock and
538  * return.
539  *
540  * This routine is fairly important along the critical path, so optimize it
541  * heavily.
542  */
543 int
544 msleep(void *ident, struct spinlock *spin, int flags,
545        const char *wmesg, int timo)
546 {
547 	globaldata_t gd = mycpu;
548 	int error;
549 
550 	crit_enter_gd(gd);
551 	_tsleep_interlock(gd, ident);
552 	spin_unlock_wr_quick(gd, spin);
553 	error = tsleep(ident, flags, wmesg, timo);
554 	spin_lock_wr_quick(gd, spin);
555 	crit_exit_gd(gd);
556 
557 	return (error);
558 }
559 
560 /*
561  * Implement the timeout for tsleep.
562  *
563  * We set P_BREAKTSLEEP to indicate that an event has occured, but
564  * we only call setrunnable if the process is not stopped.
565  *
566  * This type of callout timeout is scheduled on the same cpu the process
567  * is sleeping on.  Also, at the moment, the MP lock is held.
568  */
569 static void
570 endtsleep(void *arg)
571 {
572 	thread_t td = arg;
573 	struct proc *p;
574 
575 	ASSERT_MP_LOCK_HELD(curthread);
576 	crit_enter();
577 
578 	/*
579 	 * cpu interlock.  Thread flags are only manipulated on
580 	 * the cpu owning the thread.  proc flags are only manipulated
581 	 * by the older of the MP lock.  We have both.
582 	 */
583 	if (td->td_flags & TDF_TSLEEPQ) {
584 		td->td_flags |= TDF_TIMEOUT;
585 
586 		if ((p = td->td_proc) != NULL) {
587 			p->p_flag |= P_BREAKTSLEEP;
588 			if ((p->p_flag & P_STOPPED) == 0)
589 				setrunnable(p);
590 		} else {
591 			unsleep_and_wakeup_thread(td);
592 		}
593 	}
594 	crit_exit();
595 }
596 
597 /*
598  * Unsleep and wakeup a thread.  This function runs without the MP lock
599  * which means that it can only manipulate thread state on the owning cpu,
600  * and cannot touch the process state at all.
601  */
602 static
603 void
604 unsleep_and_wakeup_thread(struct thread *td)
605 {
606 	globaldata_t gd = mycpu;
607 	int id;
608 
609 #ifdef SMP
610 	if (td->td_gd != gd) {
611 		lwkt_send_ipiq(td->td_gd, (ipifunc1_t)unsleep_and_wakeup_thread, td);
612 		return;
613 	}
614 #endif
615 	crit_enter();
616 	if (td->td_flags & TDF_TSLEEPQ) {
617 		td->td_flags &= ~TDF_TSLEEPQ;
618 		id = LOOKUP(td->td_wchan);
619 		TAILQ_REMOVE(&gd->gd_tsleep_hash[id], td, td_threadq);
620 		if (TAILQ_FIRST(&gd->gd_tsleep_hash[id]) == NULL)
621 			atomic_clear_int(&slpque_cpumasks[id], gd->gd_cpumask);
622 		lwkt_schedule(td);
623 	}
624 	crit_exit();
625 }
626 
627 /*
628  * Make all processes sleeping on the specified identifier runnable.
629  * count may be zero or one only.
630  *
631  * The domain encodes the sleep/wakeup domain AND the first cpu to check
632  * (which is always the current cpu).  As we iterate across cpus
633  *
634  * This call may run without the MP lock held.  We can only manipulate thread
635  * state on the cpu owning the thread.  We CANNOT manipulate process state
636  * at all.
637  */
638 static void
639 _wakeup(void *ident, int domain)
640 {
641 	struct tslpque *qp;
642 	struct thread *td;
643 	struct thread *ntd;
644 	globaldata_t gd;
645 #ifdef SMP
646 	cpumask_t mask;
647 	cpumask_t tmask;
648 	int startcpu;
649 	int nextcpu;
650 #endif
651 	int id;
652 
653 	crit_enter();
654 	logtsleep(wakeup_beg);
655 	gd = mycpu;
656 	id = LOOKUP(ident);
657 	qp = &gd->gd_tsleep_hash[id];
658 restart:
659 	for (td = TAILQ_FIRST(qp); td != NULL; td = ntd) {
660 		ntd = TAILQ_NEXT(td, td_threadq);
661 		if (td->td_wchan == ident &&
662 		    td->td_wdomain == (domain & PDOMAIN_MASK)
663 		) {
664 			KKASSERT(td->td_flags & TDF_TSLEEPQ);
665 			td->td_flags &= ~TDF_TSLEEPQ;
666 			TAILQ_REMOVE(qp, td, td_threadq);
667 			if (TAILQ_FIRST(qp) == NULL) {
668 				atomic_clear_int(&slpque_cpumasks[id],
669 						 gd->gd_cpumask);
670 			}
671 			lwkt_schedule(td);
672 			if (domain & PWAKEUP_ONE)
673 				goto done;
674 			goto restart;
675 		}
676 	}
677 
678 #ifdef SMP
679 	/*
680 	 * We finished checking the current cpu but there still may be
681 	 * more work to do.  Either wakeup_one was requested and no matching
682 	 * thread was found, or a normal wakeup was requested and we have
683 	 * to continue checking cpus.
684 	 *
685 	 * The cpu that started the wakeup sequence is encoded in the domain.
686 	 * We use this information to determine which cpus still need to be
687 	 * checked, locate a candidate cpu, and chain the wakeup
688 	 * asynchronously with an IPI message.
689 	 *
690 	 * It should be noted that this scheme is actually less expensive then
691 	 * the old scheme when waking up multiple threads, since we send
692 	 * only one IPI message per target candidate which may then schedule
693 	 * multiple threads.  Before we could have wound up sending an IPI
694 	 * message for each thread on the target cpu (!= current cpu) that
695 	 * needed to be woken up.
696 	 *
697 	 * NOTE: Wakeups occuring on remote cpus are asynchronous.  This
698 	 * should be ok since we are passing idents in the IPI rather then
699 	 * thread pointers.
700 	 */
701 	if ((domain & PWAKEUP_MYCPU) == 0 &&
702 	    (mask = slpque_cpumasks[id]) != 0
703 	) {
704 		/*
705 		 * Look for a cpu that might have work to do.  Mask out cpus
706 		 * which have already been processed.
707 		 *
708 		 * 31xxxxxxxxxxxxxxxxxxxxxxxxxxxxx0
709 		 *        ^        ^           ^
710 		 *      start   currentcpu    start
711 		 *      case2                 case1
712 		 *        *        *           *
713 		 * 11111111111111110000000000000111	case1
714 		 * 00000000111111110000000000000000	case2
715 		 *
716 		 * case1:  We started at start_case1 and processed through
717 		 *  	   to the current cpu.  We have to check any bits
718 		 *	   after the current cpu, then check bits before
719 		 *         the starting cpu.
720 		 *
721 		 * case2:  We have already checked all the bits from
722 		 *         start_case2 to the end, and from 0 to the current
723 		 *         cpu.  We just have the bits from the current cpu
724 		 *         to start_case2 left to check.
725 		 */
726 		startcpu = PWAKEUP_DECODE(domain);
727 		if (gd->gd_cpuid >= startcpu) {
728 			/*
729 			 * CASE1
730 			 */
731 			tmask = mask & ~((gd->gd_cpumask << 1) - 1);
732 			if (mask & tmask) {
733 				nextcpu = bsfl(mask & tmask);
734 				lwkt_send_ipiq2(globaldata_find(nextcpu),
735 						_wakeup, ident, domain);
736 			} else {
737 				tmask = (1 << startcpu) - 1;
738 				if (mask & tmask) {
739 					nextcpu = bsfl(mask & tmask);
740 					lwkt_send_ipiq2(
741 						    globaldata_find(nextcpu),
742 						    _wakeup, ident, domain);
743 				}
744 			}
745 		} else {
746 			/*
747 			 * CASE2
748 			 */
749 			tmask = ~((gd->gd_cpumask << 1) - 1) &
750 				 ((1 << startcpu) - 1);
751 			if (mask & tmask) {
752 				nextcpu = bsfl(mask & tmask);
753 				lwkt_send_ipiq2(globaldata_find(nextcpu),
754 						_wakeup, ident, domain);
755 			}
756 		}
757 	}
758 #endif
759 done:
760 	logtsleep(wakeup_end);
761 	crit_exit();
762 }
763 
764 /*
765  * Wakeup all threads tsleep()ing on the specified ident, on all cpus
766  */
767 void
768 wakeup(void *ident)
769 {
770     _wakeup(ident, PWAKEUP_ENCODE(0, mycpu->gd_cpuid));
771 }
772 
773 /*
774  * Wakeup one thread tsleep()ing on the specified ident, on any cpu.
775  */
776 void
777 wakeup_one(void *ident)
778 {
779     /* XXX potentially round-robin the first responding cpu */
780     _wakeup(ident, PWAKEUP_ENCODE(0, mycpu->gd_cpuid) | PWAKEUP_ONE);
781 }
782 
783 /*
784  * Wakeup threads tsleep()ing on the specified ident on the current cpu
785  * only.
786  */
787 void
788 wakeup_mycpu(void *ident)
789 {
790     _wakeup(ident, PWAKEUP_MYCPU);
791 }
792 
793 /*
794  * Wakeup one thread tsleep()ing on the specified ident on the current cpu
795  * only.
796  */
797 void
798 wakeup_mycpu_one(void *ident)
799 {
800     /* XXX potentially round-robin the first responding cpu */
801     _wakeup(ident, PWAKEUP_MYCPU|PWAKEUP_ONE);
802 }
803 
804 /*
805  * Wakeup all thread tsleep()ing on the specified ident on the specified cpu
806  * only.
807  */
808 void
809 wakeup_oncpu(globaldata_t gd, void *ident)
810 {
811 #ifdef SMP
812     if (gd == mycpu) {
813 	_wakeup(ident, PWAKEUP_MYCPU);
814     } else {
815 	lwkt_send_ipiq2(gd, _wakeup, ident, PWAKEUP_MYCPU);
816     }
817 #else
818     _wakeup(ident, PWAKEUP_MYCPU);
819 #endif
820 }
821 
822 /*
823  * Wakeup one thread tsleep()ing on the specified ident on the specified cpu
824  * only.
825  */
826 void
827 wakeup_oncpu_one(globaldata_t gd, void *ident)
828 {
829 #ifdef SMP
830     if (gd == mycpu) {
831 	_wakeup(ident, PWAKEUP_MYCPU | PWAKEUP_ONE);
832     } else {
833 	lwkt_send_ipiq2(gd, _wakeup, ident, PWAKEUP_MYCPU | PWAKEUP_ONE);
834     }
835 #else
836     _wakeup(ident, PWAKEUP_MYCPU | PWAKEUP_ONE);
837 #endif
838 }
839 
840 /*
841  * Wakeup all threads waiting on the specified ident that slept using
842  * the specified domain, on all cpus.
843  */
844 void
845 wakeup_domain(void *ident, int domain)
846 {
847     _wakeup(ident, PWAKEUP_ENCODE(domain, mycpu->gd_cpuid));
848 }
849 
850 /*
851  * Wakeup one thread waiting on the specified ident that slept using
852  * the specified  domain, on any cpu.
853  */
854 void
855 wakeup_domain_one(void *ident, int domain)
856 {
857     /* XXX potentially round-robin the first responding cpu */
858     _wakeup(ident, PWAKEUP_ENCODE(domain, mycpu->gd_cpuid) | PWAKEUP_ONE);
859 }
860 
861 /*
862  * setrunnable()
863  *
864  * Make a process runnable.  The MP lock must be held on call.  This only
865  * has an effect if we are in SSLEEP.  We only break out of the
866  * tsleep if P_BREAKTSLEEP is set, otherwise we just fix-up the state.
867  *
868  * NOTE: With the MP lock held we can only safely manipulate the process
869  * structure.  We cannot safely manipulate the thread structure.
870  */
871 void
872 setrunnable(struct proc *p)
873 {
874 	crit_enter();
875 	ASSERT_MP_LOCK_HELD(curthread);
876 	p->p_flag &= ~P_STOPPED;
877 	if (p->p_stat == SSLEEP && (p->p_flag & P_BREAKTSLEEP)) {
878 		unsleep_and_wakeup_thread(p->p_thread);
879 	}
880 	crit_exit();
881 }
882 
883 /*
884  * The process is stopped due to some condition, usually because P_STOPPED
885  * is set but also possibly due to being traced.
886  *
887  * NOTE!  If the caller sets P_STOPPED, the caller must also clear P_WAITED
888  * because the parent may check the child's status before the child actually
889  * gets to this routine.
890  *
891  * This routine is called with the current process only, typically just
892  * before returning to userland.
893  *
894  * Setting P_BREAKTSLEEP before entering the tsleep will cause a passive
895  * SIGCONT to break out of the tsleep.
896  */
897 void
898 tstop(struct proc *p)
899 {
900 	wakeup((caddr_t)p->p_pptr);
901 	p->p_flag |= P_BREAKTSLEEP;
902 	tsleep(p, 0, "stop", 0);
903 }
904 
905 /*
906  * Yield / synchronous reschedule.  This is a bit tricky because the trap
907  * code might have set a lazy release on the switch function.   Setting
908  * P_PASSIVE_ACQ will ensure that the lazy release executes when we call
909  * switch, and that we are given a greater chance of affinity with our
910  * current cpu.
911  *
912  * We call lwkt_setpri_self() to rotate our thread to the end of the lwkt
913  * run queue.  lwkt_switch() will also execute any assigned passive release
914  * (which usually calls release_curproc()), allowing a same/higher priority
915  * process to be designated as the current process.
916  *
917  * While it is possible for a lower priority process to be designated,
918  * it's call to lwkt_maybe_switch() in acquire_curproc() will likely
919  * round-robin back to us and we will be able to re-acquire the current
920  * process designation.
921  */
922 void
923 uio_yield(void)
924 {
925 	struct thread *td = curthread;
926 	struct proc *p = td->td_proc;
927 
928 	lwkt_setpri_self(td->td_pri & TDPRI_MASK);
929 	if (p) {
930 		p->p_flag |= P_PASSIVE_ACQ;
931 		lwkt_switch();
932 		p->p_flag &= ~P_PASSIVE_ACQ;
933 	} else {
934 		lwkt_switch();
935 	}
936 }
937 
938 /*
939  * Compute a tenex style load average of a quantity on
940  * 1, 5 and 15 minute intervals.
941  */
942 static int loadav_count_runnable(struct proc *p, void *data);
943 
944 static void
945 loadav(void *arg)
946 {
947 	struct loadavg *avg;
948 	int i, nrun;
949 
950 	nrun = 0;
951 	allproc_scan(loadav_count_runnable, &nrun);
952 	avg = &averunnable;
953 	for (i = 0; i < 3; i++) {
954 		avg->ldavg[i] = (cexp[i] * avg->ldavg[i] +
955 		    nrun * FSCALE * (FSCALE - cexp[i])) >> FSHIFT;
956 	}
957 
958 	/*
959 	 * Schedule the next update to occur after 5 seconds, but add a
960 	 * random variation to avoid synchronisation with processes that
961 	 * run at regular intervals.
962 	 */
963 	callout_reset(&loadav_callout, hz * 4 + (int)(random() % (hz * 2 + 1)),
964 		      loadav, NULL);
965 }
966 
967 static int
968 loadav_count_runnable(struct proc *p, void *data)
969 {
970 	int *nrunp = data;
971 	thread_t td;
972 
973 	switch (p->p_stat) {
974 	case SRUN:
975 		if ((td = p->p_thread) == NULL)
976 			break;
977 		if (td->td_flags & TDF_BLOCKED)
978 			break;
979 		/* fall through */
980 	case SIDL:
981 		++*nrunp;
982 		break;
983 	default:
984 		break;
985 	}
986 	return(0);
987 }
988 
989 /* ARGSUSED */
990 static void
991 sched_setup(void *dummy)
992 {
993 	callout_init(&loadav_callout);
994 	callout_init(&schedcpu_callout);
995 
996 	/* Kick off timeout driven events by calling first time. */
997 	schedcpu(NULL);
998 	loadav(NULL);
999 }
1000 
1001