xref: /dflybsd-src/sys/kern/kern_synch.c (revision 744c01d0dc2aa1481a40e5b0988d15691602f5c9)
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 1.71 2007/01/14 07:59:03 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/smp.h>
66 
67 TAILQ_HEAD(tslpque, thread);
68 
69 static void sched_setup (void *dummy);
70 SYSINIT(sched_setup, SI_SUB_KICK_SCHEDULER, SI_ORDER_FIRST, sched_setup, NULL)
71 
72 int	hogticks;
73 int	lbolt;
74 int	lbolt_syncer;
75 int	sched_quantum;		/* Roundrobin scheduling quantum in ticks. */
76 int	ncpus;
77 int	ncpus2, ncpus2_shift, ncpus2_mask;
78 int	safepri;
79 
80 static struct callout loadav_callout;
81 static struct callout schedcpu_callout;
82 MALLOC_DEFINE(M_TSLEEP, "tslpque", "tsleep queues");
83 
84 #if !defined(KTR_TSLEEP)
85 #define KTR_TSLEEP	KTR_ALL
86 #endif
87 KTR_INFO_MASTER(tsleep);
88 KTR_INFO(KTR_TSLEEP, tsleep, tsleep_beg, 0, "tsleep enter", 0);
89 KTR_INFO(KTR_TSLEEP, tsleep, tsleep_end, 0, "tsleep exit", 0);
90 KTR_INFO(KTR_TSLEEP, tsleep, wakeup_beg, 0, "wakeup enter", 0);
91 KTR_INFO(KTR_TSLEEP, tsleep, wakeup_end, 0, "wakeup exit", 0);
92 #define logtsleep(name)	KTR_LOG(tsleep_ ## name)
93 
94 struct loadavg averunnable =
95 	{ {0, 0, 0}, FSCALE };	/* load average, of runnable procs */
96 /*
97  * Constants for averages over 1, 5, and 15 minutes
98  * when sampling at 5 second intervals.
99  */
100 static fixpt_t cexp[3] = {
101 	0.9200444146293232 * FSCALE,	/* exp(-1/12) */
102 	0.9834714538216174 * FSCALE,	/* exp(-1/60) */
103 	0.9944598480048967 * FSCALE,	/* exp(-1/180) */
104 };
105 
106 static void	endtsleep (void *);
107 static void	unsleep_and_wakeup_thread(struct thread *td);
108 static void	loadav (void *arg);
109 static void	schedcpu (void *arg);
110 
111 /*
112  * Adjust the scheduler quantum.  The quantum is specified in microseconds.
113  * Note that 'tick' is in microseconds per tick.
114  */
115 static int
116 sysctl_kern_quantum(SYSCTL_HANDLER_ARGS)
117 {
118 	int error, new_val;
119 
120 	new_val = sched_quantum * tick;
121 	error = sysctl_handle_int(oidp, &new_val, 0, req);
122         if (error != 0 || req->newptr == NULL)
123 		return (error);
124 	if (new_val < tick)
125 		return (EINVAL);
126 	sched_quantum = new_val / tick;
127 	hogticks = 2 * sched_quantum;
128 	return (0);
129 }
130 
131 SYSCTL_PROC(_kern, OID_AUTO, quantum, CTLTYPE_INT|CTLFLAG_RW,
132 	0, sizeof sched_quantum, sysctl_kern_quantum, "I", "");
133 
134 /*
135  * If `ccpu' is not equal to `exp(-1/20)' and you still want to use the
136  * faster/more-accurate formula, you'll have to estimate CCPU_SHIFT below
137  * and possibly adjust FSHIFT in "param.h" so that (FSHIFT >= CCPU_SHIFT).
138  *
139  * To estimate CCPU_SHIFT for exp(-1/20), the following formula was used:
140  *     1 - exp(-1/20) ~= 0.0487 ~= 0.0488 == 1 (fixed pt, *11* bits).
141  *
142  * If you don't want to bother with the faster/more-accurate formula, you
143  * can set CCPU_SHIFT to (FSHIFT + 1) which will use a slower/less-accurate
144  * (more general) method of calculating the %age of CPU used by a process.
145  *
146  * decay 95% of `p_pctcpu' in 60 seconds; see CCPU_SHIFT before changing
147  */
148 #define CCPU_SHIFT	11
149 
150 static fixpt_t ccpu = 0.95122942450071400909 * FSCALE; /* exp(-1/20) */
151 SYSCTL_INT(_kern, OID_AUTO, ccpu, CTLFLAG_RD, &ccpu, 0, "");
152 
153 /*
154  * kernel uses `FSCALE', userland (SHOULD) use kern.fscale
155  */
156 int     fscale __unused = FSCALE;	/* exported to systat */
157 SYSCTL_INT(_kern, OID_AUTO, fscale, CTLFLAG_RD, 0, FSCALE, "");
158 
159 /*
160  * Recompute process priorities, once a second.
161  *
162  * Since the userland schedulers are typically event oriented, if the
163  * estcpu calculation at wakeup() time is not sufficient to make a
164  * process runnable relative to other processes in the system we have
165  * a 1-second recalc to help out.
166  *
167  * This code also allows us to store sysclock_t data in the process structure
168  * without fear of an overrun, since sysclock_t are guarenteed to hold
169  * several seconds worth of count.
170  *
171  * WARNING!  callouts can preempt normal threads.  However, they will not
172  * preempt a thread holding a spinlock so we *can* safely use spinlocks.
173  */
174 static int schedcpu_stats(struct proc *p, void *data __unused);
175 static int schedcpu_resource(struct proc *p, void *data __unused);
176 
177 static void
178 schedcpu(void *arg)
179 {
180 	allproc_scan(schedcpu_stats, NULL);
181 	allproc_scan(schedcpu_resource, NULL);
182 	wakeup((caddr_t)&lbolt);
183 	wakeup((caddr_t)&lbolt_syncer);
184 	callout_reset(&schedcpu_callout, hz, schedcpu, NULL);
185 }
186 
187 /*
188  * General process statistics once a second
189  */
190 static int
191 schedcpu_stats(struct proc *p, void *data __unused)
192 {
193 	crit_enter();
194 	p->p_swtime++;
195 	if (p->p_stat == SSLEEP)
196 		p->p_slptime++;
197 
198 	/*
199 	 * Only recalculate processes that are active or have slept
200 	 * less then 2 seconds.  The schedulers understand this.
201 	 */
202 	if (p->p_slptime <= 1) {
203 		p->p_usched->recalculate(&p->p_lwp);
204 	} else {
205 		p->p_pctcpu = (p->p_pctcpu * ccpu) >> FSHIFT;
206 	}
207 	crit_exit();
208 	return(0);
209 }
210 
211 /*
212  * Resource checks.  XXX break out since ksignal/killproc can block,
213  * limiting us to one process killed per second.  There is probably
214  * a better way.
215  */
216 static int
217 schedcpu_resource(struct proc *p, void *data __unused)
218 {
219 	u_int64_t ttime;
220 
221 	crit_enter();
222 	if (p->p_stat == SIDL ||
223 	    (p->p_flag & P_ZOMBIE) ||
224 	    p->p_limit == NULL ||
225 	    p->p_thread == NULL
226 	) {
227 		crit_exit();
228 		return(0);
229 	}
230 
231 	ttime = p->p_thread->td_sticks + p->p_thread->td_uticks;
232 
233 	switch(plimit_testcpulimit(p->p_limit, ttime)) {
234 	case PLIMIT_TESTCPU_KILL:
235 		killproc(p, "exceeded maximum CPU limit");
236 		break;
237 	case PLIMIT_TESTCPU_XCPU:
238 		if ((p->p_flag & P_XCPU) == 0) {
239 			p->p_flag |= P_XCPU;
240 			ksignal(p, SIGXCPU);
241 		}
242 		break;
243 	default:
244 		break;
245 	}
246 	crit_exit();
247 	return(0);
248 }
249 
250 /*
251  * This is only used by ps.  Generate a cpu percentage use over
252  * a period of one second.
253  *
254  * MPSAFE
255  */
256 void
257 updatepcpu(struct lwp *lp, int cpticks, int ttlticks)
258 {
259 	fixpt_t acc;
260 	int remticks;
261 
262 	acc = (cpticks << FSHIFT) / ttlticks;
263 	if (ttlticks >= ESTCPUFREQ) {
264 		lp->lwp_pctcpu = acc;
265 	} else {
266 		remticks = ESTCPUFREQ - ttlticks;
267 		lp->lwp_pctcpu = (acc * ttlticks + lp->lwp_pctcpu * remticks) /
268 				ESTCPUFREQ;
269 	}
270 }
271 
272 /*
273  * We're only looking at 7 bits of the address; everything is
274  * aligned to 4, lots of things are aligned to greater powers
275  * of 2.  Shift right by 8, i.e. drop the bottom 256 worth.
276  */
277 #define TABLESIZE	128
278 #define LOOKUP(x)	(((intptr_t)(x) >> 8) & (TABLESIZE - 1))
279 
280 static cpumask_t slpque_cpumasks[TABLESIZE];
281 
282 /*
283  * General scheduler initialization.  We force a reschedule 25 times
284  * a second by default.  Note that cpu0 is initialized in early boot and
285  * cannot make any high level calls.
286  *
287  * Each cpu has its own sleep queue.
288  */
289 void
290 sleep_gdinit(globaldata_t gd)
291 {
292 	static struct tslpque slpque_cpu0[TABLESIZE];
293 	int i;
294 
295 	if (gd->gd_cpuid == 0) {
296 		sched_quantum = (hz + 24) / 25;
297 		hogticks = 2 * sched_quantum;
298 
299 		gd->gd_tsleep_hash = slpque_cpu0;
300 	} else {
301 		gd->gd_tsleep_hash = kmalloc(sizeof(slpque_cpu0),
302 					    M_TSLEEP, M_WAITOK | M_ZERO);
303 	}
304 	for (i = 0; i < TABLESIZE; ++i)
305 		TAILQ_INIT(&gd->gd_tsleep_hash[i]);
306 }
307 
308 /*
309  * General sleep call.  Suspends the current process until a wakeup is
310  * performed on the specified identifier.  The process will then be made
311  * runnable with the specified priority.  Sleeps at most timo/hz seconds
312  * (0 means no timeout).  If flags includes PCATCH flag, signals are checked
313  * before and after sleeping, else signals are not checked.  Returns 0 if
314  * awakened, EWOULDBLOCK if the timeout expires.  If PCATCH is set and a
315  * signal needs to be delivered, ERESTART is returned if the current system
316  * call should be restarted if possible, and EINTR is returned if the system
317  * call should be interrupted by the signal (return EINTR).
318  *
319  * Note that if we are a process, we release_curproc() before messing with
320  * the LWKT scheduler.
321  *
322  * During autoconfiguration or after a panic, a sleep will simply
323  * lower the priority briefly to allow interrupts, then return.
324  */
325 int
326 tsleep(void *ident, int flags, const char *wmesg, int timo)
327 {
328 	struct thread *td = curthread;
329 	struct proc *p = td->td_proc;		/* may be NULL */
330 	globaldata_t gd;
331 	int sig;
332 	int catch;
333 	int id;
334 	int error;
335 	int oldpri;
336 	struct callout thandle;
337 
338 	/*
339 	 * NOTE: removed KTRPOINT, it could cause races due to blocking
340 	 * even in stable.  Just scrap it for now.
341 	 */
342 	if (cold || panicstr) {
343 		/*
344 		 * After a panic, or during autoconfiguration,
345 		 * just give interrupts a chance, then just return;
346 		 * don't run any other procs or panic below,
347 		 * in case this is the idle process and already asleep.
348 		 */
349 		splz();
350 		oldpri = td->td_pri & TDPRI_MASK;
351 		lwkt_setpri_self(safepri);
352 		lwkt_switch();
353 		lwkt_setpri_self(oldpri);
354 		return (0);
355 	}
356 	logtsleep(tsleep_beg);
357 	gd = td->td_gd;
358 	KKASSERT(td != &gd->gd_idlethread);	/* you must be kidding! */
359 
360 	/*
361 	 * NOTE: all of this occurs on the current cpu, including any
362 	 * callout-based wakeups, so a critical section is a sufficient
363 	 * interlock.
364 	 *
365 	 * The entire sequence through to where we actually sleep must
366 	 * run without breaking the critical section.
367 	 */
368 	id = LOOKUP(ident);
369 	catch = flags & PCATCH;
370 	error = 0;
371 	sig = 0;
372 
373 	crit_enter_quick(td);
374 
375 	KASSERT(ident != NULL, ("tsleep: no ident"));
376 	KASSERT(p == NULL || p->p_stat == SRUN, ("tsleep %p %s %d",
377 		ident, wmesg, p->p_stat));
378 
379 	/*
380 	 * Setup for the current process (if this is a process).
381 	 */
382 	if (p) {
383 		if (catch) {
384 			/*
385 			 * Early termination if PCATCH was set and a
386 			 * signal is pending, interlocked with the
387 			 * critical section.
388 			 *
389 			 * Early termination only occurs when tsleep() is
390 			 * entered while in a normal SRUN state.
391 			 */
392 			if ((sig = CURSIG(p)) != 0)
393 				goto resume;
394 
395 			/*
396 			 * Early termination if PCATCH was set and a
397 			 * mailbox signal was possibly delivered prior to
398 			 * the system call even being made, in order to
399 			 * allow the user to interlock without having to
400 			 * make additional system calls.
401 			 */
402 			if (p->p_flag & P_MAILBOX)
403 				goto resume;
404 
405 			/*
406 			 * Causes ksignal to wake us up when.
407 			 */
408 			p->p_flag |= P_SINTR;
409 		}
410 
411 		/*
412 		 * Make sure the current process has been untangled from
413 		 * the userland scheduler and initialize slptime to start
414 		 * counting.
415 		 */
416 		if (flags & PNORESCHED)
417 			td->td_flags |= TDF_NORESCHED;
418 		p->p_usched->release_curproc(&p->p_lwp);
419 		p->p_slptime = 0;
420 	}
421 
422 	/*
423 	 * Move our thread to the correct queue and setup our wchan, etc.
424 	 */
425 	lwkt_deschedule_self(td);
426 	td->td_flags |= TDF_TSLEEPQ;
427 	TAILQ_INSERT_TAIL(&gd->gd_tsleep_hash[id], td, td_threadq);
428 	atomic_set_int(&slpque_cpumasks[id], gd->gd_cpumask);
429 
430 	td->td_wchan = ident;
431 	td->td_wmesg = wmesg;
432 	td->td_wdomain = flags & PDOMAIN_MASK;
433 
434 	/*
435 	 * Setup the timeout, if any
436 	 */
437 	if (timo) {
438 		callout_init(&thandle);
439 		callout_reset(&thandle, timo, endtsleep, td);
440 	}
441 
442 	/*
443 	 * Beddy bye bye.
444 	 */
445 	if (p) {
446 		/*
447 		 * Ok, we are sleeping.  Place us in the SSLEEP state.
448 		 */
449 		KKASSERT((p->p_flag & P_ONRUNQ) == 0);
450 		p->p_stat = SSLEEP;
451 		p->p_lwp.lwp_ru.ru_nvcsw++;
452 		lwkt_switch();
453 
454 		/*
455 		 * And when we are woken up, put us back in SRUN.  If we
456 		 * slept for over a second, recalculate our estcpu.
457 		 */
458 		p->p_stat = SRUN;
459 		if (p->p_slptime)
460 			p->p_usched->recalculate(&p->p_lwp);
461 		p->p_slptime = 0;
462 	} else {
463 		lwkt_switch();
464 	}
465 
466 	/*
467 	 * Make sure we haven't switched cpus while we were asleep.  It's
468 	 * not supposed to happen.  Cleanup our temporary flags.
469 	 */
470 	KKASSERT(gd == td->td_gd);
471 	td->td_flags &= ~TDF_NORESCHED;
472 
473 	/*
474 	 * Cleanup the timeout.
475 	 */
476 	if (timo) {
477 		if (td->td_flags & TDF_TIMEOUT) {
478 			td->td_flags &= ~TDF_TIMEOUT;
479 			if (sig == 0)
480 				error = EWOULDBLOCK;
481 		} else {
482 			callout_stop(&thandle);
483 		}
484 	}
485 
486 	/*
487 	 * Since td_threadq is used both for our run queue AND for the
488 	 * tsleep hash queue, we can't still be on it at this point because
489 	 * we've gotten cpu back.
490 	 */
491 	KASSERT((td->td_flags & TDF_TSLEEPQ) == 0, ("tsleep: impossible thread flags %08x", td->td_flags));
492 	td->td_wchan = NULL;
493 	td->td_wmesg = NULL;
494 	td->td_wdomain = 0;
495 
496 	/*
497 	 * Figure out the correct error return.  If interrupted by a
498 	 * signal we want to return EINTR or ERESTART.
499 	 *
500 	 * If P_MAILBOX is set no automatic system call restart occurs
501 	 * and we return EINTR.  P_MAILBOX is meant to be used as an
502 	 * interlock, the user must poll it prior to any system call
503 	 * that it wishes to interlock a mailbox signal against since
504 	 * the flag is cleared on *any* system call that sleeps.
505 	 */
506 resume:
507 	if (p) {
508 		if (catch && error == 0) {
509 			if ((p->p_flag & P_MAILBOX) && sig == 0) {
510 				error = EINTR;
511 			} else if ((sig != 0 || (sig = CURSIG(p)))) {
512 				if (SIGISMEMBER(p->p_sigacts->ps_sigintr, sig))
513 					error = EINTR;
514 				else
515 					error = ERESTART;
516 			}
517 		}
518 		p->p_flag &= ~(P_BREAKTSLEEP | P_SINTR | P_MAILBOX);
519 	}
520 	logtsleep(tsleep_end);
521 	crit_exit_quick(td);
522 	return (error);
523 }
524 
525 /*
526  * This is a dandy function that allows us to interlock tsleep/wakeup
527  * operations with unspecified upper level locks, such as lockmgr locks,
528  * simply by holding a critical section.  The sequence is:
529  *
530  *	(enter critical section)
531  *	(acquire upper level lock)
532  *	tsleep_interlock(blah)
533  *	(release upper level lock)
534  *	tsleep(blah, ...)
535  *	(exit critical section)
536  *
537  * Basically this function sets our cpumask for the ident which informs
538  * other cpus that our cpu 'might' be waiting (or about to wait on) the
539  * hash index related to the ident.  The critical section prevents another
540  * cpu's wakeup() from being processed on our cpu until we are actually
541  * able to enter the tsleep().  Thus, no race occurs between our attempt
542  * to release a resource and sleep, and another cpu's attempt to acquire
543  * a resource and call wakeup.
544  *
545  * There isn't much of a point to this function unless you call it while
546  * holding a critical section.
547  */
548 static __inline void
549 _tsleep_interlock(globaldata_t gd, void *ident)
550 {
551 	int id = LOOKUP(ident);
552 
553 	atomic_set_int(&slpque_cpumasks[id], gd->gd_cpumask);
554 }
555 
556 void
557 tsleep_interlock(void *ident)
558 {
559 	_tsleep_interlock(mycpu, ident);
560 }
561 
562 /*
563  * Interlocked spinlock sleep.  An exclusively held spinlock must
564  * be passed to msleep().  The function will atomically release the
565  * spinlock and tsleep on the ident, then reacquire the spinlock and
566  * return.
567  *
568  * This routine is fairly important along the critical path, so optimize it
569  * heavily.
570  */
571 int
572 msleep(void *ident, struct spinlock *spin, int flags,
573        const char *wmesg, int timo)
574 {
575 	globaldata_t gd = mycpu;
576 	int error;
577 
578 	crit_enter_gd(gd);
579 	_tsleep_interlock(gd, ident);
580 	spin_unlock_wr_quick(gd, spin);
581 	error = tsleep(ident, flags, wmesg, timo);
582 	spin_lock_wr_quick(gd, spin);
583 	crit_exit_gd(gd);
584 
585 	return (error);
586 }
587 
588 /*
589  * Implement the timeout for tsleep.
590  *
591  * We set P_BREAKTSLEEP to indicate that an event has occured, but
592  * we only call setrunnable if the process is not stopped.
593  *
594  * This type of callout timeout is scheduled on the same cpu the process
595  * is sleeping on.  Also, at the moment, the MP lock is held.
596  */
597 static void
598 endtsleep(void *arg)
599 {
600 	thread_t td = arg;
601 	struct proc *p;
602 
603 	ASSERT_MP_LOCK_HELD(curthread);
604 	crit_enter();
605 
606 	/*
607 	 * cpu interlock.  Thread flags are only manipulated on
608 	 * the cpu owning the thread.  proc flags are only manipulated
609 	 * by the older of the MP lock.  We have both.
610 	 */
611 	if (td->td_flags & TDF_TSLEEPQ) {
612 		td->td_flags |= TDF_TIMEOUT;
613 
614 		if ((p = td->td_proc) != NULL) {
615 			p->p_flag |= P_BREAKTSLEEP;
616 			if ((p->p_flag & P_STOPPED) == 0)
617 				setrunnable(p);
618 		} else {
619 			unsleep_and_wakeup_thread(td);
620 		}
621 	}
622 	crit_exit();
623 }
624 
625 /*
626  * Unsleep and wakeup a thread.  This function runs without the MP lock
627  * which means that it can only manipulate thread state on the owning cpu,
628  * and cannot touch the process state at all.
629  */
630 static
631 void
632 unsleep_and_wakeup_thread(struct thread *td)
633 {
634 	globaldata_t gd = mycpu;
635 	int id;
636 
637 #ifdef SMP
638 	if (td->td_gd != gd) {
639 		lwkt_send_ipiq(td->td_gd, (ipifunc1_t)unsleep_and_wakeup_thread, td);
640 		return;
641 	}
642 #endif
643 	crit_enter();
644 	if (td->td_flags & TDF_TSLEEPQ) {
645 		td->td_flags &= ~TDF_TSLEEPQ;
646 		id = LOOKUP(td->td_wchan);
647 		TAILQ_REMOVE(&gd->gd_tsleep_hash[id], td, td_threadq);
648 		if (TAILQ_FIRST(&gd->gd_tsleep_hash[id]) == NULL)
649 			atomic_clear_int(&slpque_cpumasks[id], gd->gd_cpumask);
650 		lwkt_schedule(td);
651 	}
652 	crit_exit();
653 }
654 
655 /*
656  * Make all processes sleeping on the specified identifier runnable.
657  * count may be zero or one only.
658  *
659  * The domain encodes the sleep/wakeup domain AND the first cpu to check
660  * (which is always the current cpu).  As we iterate across cpus
661  *
662  * This call may run without the MP lock held.  We can only manipulate thread
663  * state on the cpu owning the thread.  We CANNOT manipulate process state
664  * at all.
665  */
666 static void
667 _wakeup(void *ident, int domain)
668 {
669 	struct tslpque *qp;
670 	struct thread *td;
671 	struct thread *ntd;
672 	globaldata_t gd;
673 #ifdef SMP
674 	cpumask_t mask;
675 	cpumask_t tmask;
676 	int startcpu;
677 	int nextcpu;
678 #endif
679 	int id;
680 
681 	crit_enter();
682 	logtsleep(wakeup_beg);
683 	gd = mycpu;
684 	id = LOOKUP(ident);
685 	qp = &gd->gd_tsleep_hash[id];
686 restart:
687 	for (td = TAILQ_FIRST(qp); td != NULL; td = ntd) {
688 		ntd = TAILQ_NEXT(td, td_threadq);
689 		if (td->td_wchan == ident &&
690 		    td->td_wdomain == (domain & PDOMAIN_MASK)
691 		) {
692 			KKASSERT(td->td_flags & TDF_TSLEEPQ);
693 			td->td_flags &= ~TDF_TSLEEPQ;
694 			TAILQ_REMOVE(qp, td, td_threadq);
695 			if (TAILQ_FIRST(qp) == NULL) {
696 				atomic_clear_int(&slpque_cpumasks[id],
697 						 gd->gd_cpumask);
698 			}
699 			lwkt_schedule(td);
700 			if (domain & PWAKEUP_ONE)
701 				goto done;
702 			goto restart;
703 		}
704 	}
705 
706 #ifdef SMP
707 	/*
708 	 * We finished checking the current cpu but there still may be
709 	 * more work to do.  Either wakeup_one was requested and no matching
710 	 * thread was found, or a normal wakeup was requested and we have
711 	 * to continue checking cpus.
712 	 *
713 	 * The cpu that started the wakeup sequence is encoded in the domain.
714 	 * We use this information to determine which cpus still need to be
715 	 * checked, locate a candidate cpu, and chain the wakeup
716 	 * asynchronously with an IPI message.
717 	 *
718 	 * It should be noted that this scheme is actually less expensive then
719 	 * the old scheme when waking up multiple threads, since we send
720 	 * only one IPI message per target candidate which may then schedule
721 	 * multiple threads.  Before we could have wound up sending an IPI
722 	 * message for each thread on the target cpu (!= current cpu) that
723 	 * needed to be woken up.
724 	 *
725 	 * NOTE: Wakeups occuring on remote cpus are asynchronous.  This
726 	 * should be ok since we are passing idents in the IPI rather then
727 	 * thread pointers.
728 	 */
729 	if ((domain & PWAKEUP_MYCPU) == 0 &&
730 	    (mask = slpque_cpumasks[id]) != 0
731 	) {
732 		/*
733 		 * Look for a cpu that might have work to do.  Mask out cpus
734 		 * which have already been processed.
735 		 *
736 		 * 31xxxxxxxxxxxxxxxxxxxxxxxxxxxxx0
737 		 *        ^        ^           ^
738 		 *      start   currentcpu    start
739 		 *      case2                 case1
740 		 *        *        *           *
741 		 * 11111111111111110000000000000111	case1
742 		 * 00000000111111110000000000000000	case2
743 		 *
744 		 * case1:  We started at start_case1 and processed through
745 		 *  	   to the current cpu.  We have to check any bits
746 		 *	   after the current cpu, then check bits before
747 		 *         the starting cpu.
748 		 *
749 		 * case2:  We have already checked all the bits from
750 		 *         start_case2 to the end, and from 0 to the current
751 		 *         cpu.  We just have the bits from the current cpu
752 		 *         to start_case2 left to check.
753 		 */
754 		startcpu = PWAKEUP_DECODE(domain);
755 		if (gd->gd_cpuid >= startcpu) {
756 			/*
757 			 * CASE1
758 			 */
759 			tmask = mask & ~((gd->gd_cpumask << 1) - 1);
760 			if (mask & tmask) {
761 				nextcpu = bsfl(mask & tmask);
762 				lwkt_send_ipiq2(globaldata_find(nextcpu),
763 						_wakeup, ident, domain);
764 			} else {
765 				tmask = (1 << startcpu) - 1;
766 				if (mask & tmask) {
767 					nextcpu = bsfl(mask & tmask);
768 					lwkt_send_ipiq2(
769 						    globaldata_find(nextcpu),
770 						    _wakeup, ident, domain);
771 				}
772 			}
773 		} else {
774 			/*
775 			 * CASE2
776 			 */
777 			tmask = ~((gd->gd_cpumask << 1) - 1) &
778 				 ((1 << startcpu) - 1);
779 			if (mask & tmask) {
780 				nextcpu = bsfl(mask & tmask);
781 				lwkt_send_ipiq2(globaldata_find(nextcpu),
782 						_wakeup, ident, domain);
783 			}
784 		}
785 	}
786 #endif
787 done:
788 	logtsleep(wakeup_end);
789 	crit_exit();
790 }
791 
792 /*
793  * Wakeup all threads tsleep()ing on the specified ident, on all cpus
794  */
795 void
796 wakeup(void *ident)
797 {
798     _wakeup(ident, PWAKEUP_ENCODE(0, mycpu->gd_cpuid));
799 }
800 
801 /*
802  * Wakeup one thread tsleep()ing on the specified ident, on any cpu.
803  */
804 void
805 wakeup_one(void *ident)
806 {
807     /* XXX potentially round-robin the first responding cpu */
808     _wakeup(ident, PWAKEUP_ENCODE(0, mycpu->gd_cpuid) | PWAKEUP_ONE);
809 }
810 
811 /*
812  * Wakeup threads tsleep()ing on the specified ident on the current cpu
813  * only.
814  */
815 void
816 wakeup_mycpu(void *ident)
817 {
818     _wakeup(ident, PWAKEUP_MYCPU);
819 }
820 
821 /*
822  * Wakeup one thread tsleep()ing on the specified ident on the current cpu
823  * only.
824  */
825 void
826 wakeup_mycpu_one(void *ident)
827 {
828     /* XXX potentially round-robin the first responding cpu */
829     _wakeup(ident, PWAKEUP_MYCPU|PWAKEUP_ONE);
830 }
831 
832 /*
833  * Wakeup all thread tsleep()ing on the specified ident on the specified cpu
834  * only.
835  */
836 void
837 wakeup_oncpu(globaldata_t gd, void *ident)
838 {
839 #ifdef SMP
840     if (gd == mycpu) {
841 	_wakeup(ident, PWAKEUP_MYCPU);
842     } else {
843 	lwkt_send_ipiq2(gd, _wakeup, ident, PWAKEUP_MYCPU);
844     }
845 #else
846     _wakeup(ident, PWAKEUP_MYCPU);
847 #endif
848 }
849 
850 /*
851  * Wakeup one thread tsleep()ing on the specified ident on the specified cpu
852  * only.
853  */
854 void
855 wakeup_oncpu_one(globaldata_t gd, void *ident)
856 {
857 #ifdef SMP
858     if (gd == mycpu) {
859 	_wakeup(ident, PWAKEUP_MYCPU | PWAKEUP_ONE);
860     } else {
861 	lwkt_send_ipiq2(gd, _wakeup, ident, PWAKEUP_MYCPU | PWAKEUP_ONE);
862     }
863 #else
864     _wakeup(ident, PWAKEUP_MYCPU | PWAKEUP_ONE);
865 #endif
866 }
867 
868 /*
869  * Wakeup all threads waiting on the specified ident that slept using
870  * the specified domain, on all cpus.
871  */
872 void
873 wakeup_domain(void *ident, int domain)
874 {
875     _wakeup(ident, PWAKEUP_ENCODE(domain, mycpu->gd_cpuid));
876 }
877 
878 /*
879  * Wakeup one thread waiting on the specified ident that slept using
880  * the specified  domain, on any cpu.
881  */
882 void
883 wakeup_domain_one(void *ident, int domain)
884 {
885     /* XXX potentially round-robin the first responding cpu */
886     _wakeup(ident, PWAKEUP_ENCODE(domain, mycpu->gd_cpuid) | PWAKEUP_ONE);
887 }
888 
889 /*
890  * setrunnable()
891  *
892  * Make a process runnable.  The MP lock must be held on call.  This only
893  * has an effect if we are in SSLEEP.  We only break out of the
894  * tsleep if P_BREAKTSLEEP is set, otherwise we just fix-up the state.
895  *
896  * NOTE: With the MP lock held we can only safely manipulate the process
897  * structure.  We cannot safely manipulate the thread structure.
898  */
899 void
900 setrunnable(struct proc *p)
901 {
902 	crit_enter();
903 	ASSERT_MP_LOCK_HELD(curthread);
904 	p->p_flag &= ~P_STOPPED;
905 	if (p->p_stat == SSLEEP && (p->p_flag & P_BREAKTSLEEP)) {
906 		unsleep_and_wakeup_thread(p->p_thread);
907 	}
908 	crit_exit();
909 }
910 
911 /*
912  * The process is stopped due to some condition, usually because P_STOPPED
913  * is set but also possibly due to being traced.
914  *
915  * NOTE!  If the caller sets P_STOPPED, the caller must also clear P_WAITED
916  * because the parent may check the child's status before the child actually
917  * gets to this routine.
918  *
919  * This routine is called with the current process only, typically just
920  * before returning to userland.
921  *
922  * Setting P_BREAKTSLEEP before entering the tsleep will cause a passive
923  * SIGCONT to break out of the tsleep.
924  */
925 void
926 tstop(struct proc *p)
927 {
928 	wakeup((caddr_t)p->p_pptr);
929 	p->p_flag |= P_BREAKTSLEEP;
930 	tsleep(p, 0, "stop", 0);
931 }
932 
933 /*
934  * Yield / synchronous reschedule.  This is a bit tricky because the trap
935  * code might have set a lazy release on the switch function.   Setting
936  * P_PASSIVE_ACQ will ensure that the lazy release executes when we call
937  * switch, and that we are given a greater chance of affinity with our
938  * current cpu.
939  *
940  * We call lwkt_setpri_self() to rotate our thread to the end of the lwkt
941  * run queue.  lwkt_switch() will also execute any assigned passive release
942  * (which usually calls release_curproc()), allowing a same/higher priority
943  * process to be designated as the current process.
944  *
945  * While it is possible for a lower priority process to be designated,
946  * it's call to lwkt_maybe_switch() in acquire_curproc() will likely
947  * round-robin back to us and we will be able to re-acquire the current
948  * process designation.
949  */
950 void
951 uio_yield(void)
952 {
953 	struct thread *td = curthread;
954 	struct proc *p = td->td_proc;
955 
956 	lwkt_setpri_self(td->td_pri & TDPRI_MASK);
957 	if (p) {
958 		p->p_flag |= P_PASSIVE_ACQ;
959 		lwkt_switch();
960 		p->p_flag &= ~P_PASSIVE_ACQ;
961 	} else {
962 		lwkt_switch();
963 	}
964 }
965 
966 /*
967  * Compute a tenex style load average of a quantity on
968  * 1, 5 and 15 minute intervals.
969  */
970 static int loadav_count_runnable(struct proc *p, void *data);
971 
972 static void
973 loadav(void *arg)
974 {
975 	struct loadavg *avg;
976 	int i, nrun;
977 
978 	nrun = 0;
979 	allproc_scan(loadav_count_runnable, &nrun);
980 	avg = &averunnable;
981 	for (i = 0; i < 3; i++) {
982 		avg->ldavg[i] = (cexp[i] * avg->ldavg[i] +
983 		    nrun * FSCALE * (FSCALE - cexp[i])) >> FSHIFT;
984 	}
985 
986 	/*
987 	 * Schedule the next update to occur after 5 seconds, but add a
988 	 * random variation to avoid synchronisation with processes that
989 	 * run at regular intervals.
990 	 */
991 	callout_reset(&loadav_callout, hz * 4 + (int)(krandom() % (hz * 2 + 1)),
992 		      loadav, NULL);
993 }
994 
995 static int
996 loadav_count_runnable(struct proc *p, void *data)
997 {
998 	int *nrunp = data;
999 	thread_t td;
1000 
1001 	switch (p->p_stat) {
1002 	case SRUN:
1003 		if ((td = p->p_thread) == NULL)
1004 			break;
1005 		if (td->td_flags & TDF_BLOCKED)
1006 			break;
1007 		/* fall through */
1008 	case SIDL:
1009 		++*nrunp;
1010 		break;
1011 	default:
1012 		break;
1013 	}
1014 	return(0);
1015 }
1016 
1017 /* ARGSUSED */
1018 static void
1019 sched_setup(void *dummy)
1020 {
1021 	callout_init(&loadav_callout);
1022 	callout_init(&schedcpu_callout);
1023 
1024 	/* Kick off timeout driven events by calling first time. */
1025 	schedcpu(NULL);
1026 	loadav(NULL);
1027 }
1028 
1029