xref: /freebsd/sys/kern/kern_synch.c (revision f856af0466c076beef4ea9b15d088e1119a945b8)
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  * 4. Neither the name of the University nor the names of its contributors
19  *    may be used to endorse or promote products derived from this software
20  *    without specific prior written permission.
21  *
22  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
23  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
24  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
25  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
26  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
27  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
28  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
29  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
30  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
31  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
32  * SUCH DAMAGE.
33  *
34  *	@(#)kern_synch.c	8.9 (Berkeley) 5/19/95
35  */
36 
37 #include <sys/cdefs.h>
38 __FBSDID("$FreeBSD$");
39 
40 #include "opt_ktrace.h"
41 
42 #include <sys/param.h>
43 #include <sys/systm.h>
44 #include <sys/condvar.h>
45 #include <sys/kdb.h>
46 #include <sys/kernel.h>
47 #include <sys/ktr.h>
48 #include <sys/lock.h>
49 #include <sys/mutex.h>
50 #include <sys/proc.h>
51 #include <sys/resourcevar.h>
52 #include <sys/sched.h>
53 #include <sys/signalvar.h>
54 #include <sys/sleepqueue.h>
55 #include <sys/smp.h>
56 #include <sys/sx.h>
57 #include <sys/sysctl.h>
58 #include <sys/sysproto.h>
59 #include <sys/vmmeter.h>
60 #ifdef KTRACE
61 #include <sys/uio.h>
62 #include <sys/ktrace.h>
63 #endif
64 
65 #include <machine/cpu.h>
66 
67 static void synch_setup(void *dummy);
68 SYSINIT(synch_setup, SI_SUB_KICK_SCHEDULER, SI_ORDER_FIRST, synch_setup, NULL)
69 
70 int	hogticks;
71 int	lbolt;
72 
73 static struct callout loadav_callout;
74 static struct callout lbolt_callout;
75 
76 struct loadavg averunnable =
77 	{ {0, 0, 0}, FSCALE };	/* load average, of runnable procs */
78 /*
79  * Constants for averages over 1, 5, and 15 minutes
80  * when sampling at 5 second intervals.
81  */
82 static fixpt_t cexp[3] = {
83 	0.9200444146293232 * FSCALE,	/* exp(-1/12) */
84 	0.9834714538216174 * FSCALE,	/* exp(-1/60) */
85 	0.9944598480048967 * FSCALE,	/* exp(-1/180) */
86 };
87 
88 /* kernel uses `FSCALE', userland (SHOULD) use kern.fscale */
89 static int      fscale __unused = FSCALE;
90 SYSCTL_INT(_kern, OID_AUTO, fscale, CTLFLAG_RD, 0, FSCALE, "");
91 
92 static void	loadav(void *arg);
93 static void	lboltcb(void *arg);
94 
95 void
96 sleepinit(void)
97 {
98 
99 	hogticks = (hz / 10) * 2;	/* Default only. */
100 	init_sleepqueues();
101 }
102 
103 /*
104  * General sleep call.  Suspends the current thread until a wakeup is
105  * performed on the specified identifier.  The thread will then be made
106  * runnable with the specified priority.  Sleeps at most timo/hz seconds
107  * (0 means no timeout).  If pri includes PCATCH flag, signals are checked
108  * before and after sleeping, else signals are not checked.  Returns 0 if
109  * awakened, EWOULDBLOCK if the timeout expires.  If PCATCH is set and a
110  * signal needs to be delivered, ERESTART is returned if the current system
111  * call should be restarted if possible, and EINTR is returned if the system
112  * call should be interrupted by the signal (return EINTR).
113  *
114  * The mutex argument is unlocked before the caller is suspended, and
115  * re-locked before msleep returns.  If priority includes the PDROP
116  * flag the mutex is not re-locked before returning.
117  */
118 int
119 msleep(ident, mtx, priority, wmesg, timo)
120 	void *ident;
121 	struct mtx *mtx;
122 	int priority, timo;
123 	const char *wmesg;
124 {
125 	struct thread *td;
126 	struct proc *p;
127 	int catch, rval, flags, pri;
128 	WITNESS_SAVE_DECL(mtx);
129 
130 	td = curthread;
131 	p = td->td_proc;
132 #ifdef KTRACE
133 	if (KTRPOINT(td, KTR_CSW))
134 		ktrcsw(1, 0);
135 #endif
136 	WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, mtx == NULL ? NULL :
137 	    &mtx->mtx_object, "Sleeping on \"%s\"", wmesg);
138 	KASSERT(timo != 0 || mtx_owned(&Giant) || mtx != NULL ||
139 	    ident == &lbolt, ("sleeping without a mutex"));
140 	KASSERT(p != NULL, ("msleep1"));
141 	KASSERT(ident != NULL && TD_IS_RUNNING(td), ("msleep"));
142 
143 	if (cold) {
144 		/*
145 		 * During autoconfiguration, just return;
146 		 * don't run any other threads or panic below,
147 		 * in case this is the idle thread and already asleep.
148 		 * XXX: this used to do "s = splhigh(); splx(safepri);
149 		 * splx(s);" to give interrupts a chance, but there is
150 		 * no way to give interrupts a chance now.
151 		 */
152 		if (mtx != NULL && priority & PDROP)
153 			mtx_unlock(mtx);
154 		return (0);
155 	}
156 	catch = priority & PCATCH;
157 	rval = 0;
158 
159 	/*
160 	 * If we are already on a sleep queue, then remove us from that
161 	 * sleep queue first.  We have to do this to handle recursive
162 	 * sleeps.
163 	 */
164 	if (TD_ON_SLEEPQ(td))
165 		sleepq_remove(td, td->td_wchan);
166 
167 	flags = SLEEPQ_MSLEEP;
168 	if (catch)
169 		flags |= SLEEPQ_INTERRUPTIBLE;
170 
171 	sleepq_lock(ident);
172 	CTR5(KTR_PROC, "msleep: thread %p (pid %ld, %s) on %s (%p)",
173 	    (void *)td, (long)p->p_pid, p->p_comm, wmesg, ident);
174 
175 	DROP_GIANT();
176 	if (mtx != NULL) {
177 		mtx_assert(mtx, MA_OWNED | MA_NOTRECURSED);
178 		WITNESS_SAVE(&mtx->mtx_object, mtx);
179 		mtx_unlock(mtx);
180 	}
181 
182 	/*
183 	 * We put ourselves on the sleep queue and start our timeout
184 	 * before calling thread_suspend_check, as we could stop there,
185 	 * and a wakeup or a SIGCONT (or both) could occur while we were
186 	 * stopped without resuming us.  Thus, we must be ready for sleep
187 	 * when cursig() is called.  If the wakeup happens while we're
188 	 * stopped, then td will no longer be on a sleep queue upon
189 	 * return from cursig().
190 	 */
191 	sleepq_add(ident, ident == &lbolt ? NULL : &mtx->mtx_object, wmesg,
192 	    flags, 0);
193 	if (timo)
194 		sleepq_set_timeout(ident, timo);
195 
196 	/*
197 	 * Adjust this thread's priority, if necessary.
198 	 */
199 	pri = priority & PRIMASK;
200 	if (pri != 0 && pri != td->td_priority) {
201 		mtx_lock_spin(&sched_lock);
202 		sched_prio(td, pri);
203 		mtx_unlock_spin(&sched_lock);
204 	}
205 
206 	if (timo && catch)
207 		rval = sleepq_timedwait_sig(ident);
208 	else if (timo)
209 		rval = sleepq_timedwait(ident);
210 	else if (catch)
211 		rval = sleepq_wait_sig(ident);
212 	else {
213 		sleepq_wait(ident);
214 		rval = 0;
215 	}
216 #ifdef KTRACE
217 	if (KTRPOINT(td, KTR_CSW))
218 		ktrcsw(0, 0);
219 #endif
220 	PICKUP_GIANT();
221 	if (mtx != NULL && !(priority & PDROP)) {
222 		mtx_lock(mtx);
223 		WITNESS_RESTORE(&mtx->mtx_object, mtx);
224 	}
225 	return (rval);
226 }
227 
228 int
229 msleep_spin(ident, mtx, wmesg, timo)
230 	void *ident;
231 	struct mtx *mtx;
232 	const char *wmesg;
233 	int timo;
234 {
235 	struct thread *td;
236 	struct proc *p;
237 	int rval;
238 	WITNESS_SAVE_DECL(mtx);
239 
240 	td = curthread;
241 	p = td->td_proc;
242 	KASSERT(mtx != NULL, ("sleeping without a mutex"));
243 	KASSERT(p != NULL, ("msleep1"));
244 	KASSERT(ident != NULL && TD_IS_RUNNING(td), ("msleep"));
245 
246 	if (cold) {
247 		/*
248 		 * During autoconfiguration, just return;
249 		 * don't run any other threads or panic below,
250 		 * in case this is the idle thread and already asleep.
251 		 * XXX: this used to do "s = splhigh(); splx(safepri);
252 		 * splx(s);" to give interrupts a chance, but there is
253 		 * no way to give interrupts a chance now.
254 		 */
255 		return (0);
256 	}
257 
258 	sleepq_lock(ident);
259 	CTR5(KTR_PROC, "msleep_spin: thread %p (pid %ld, %s) on %s (%p)",
260 	    (void *)td, (long)p->p_pid, p->p_comm, wmesg, ident);
261 
262 	DROP_GIANT();
263 	mtx_assert(mtx, MA_OWNED | MA_NOTRECURSED);
264 	WITNESS_SAVE(&mtx->mtx_object, mtx);
265 	mtx_unlock_spin(mtx);
266 
267 	/*
268 	 * We put ourselves on the sleep queue and start our timeout.
269 	 */
270 	sleepq_add(ident, &mtx->mtx_object, wmesg, SLEEPQ_MSLEEP, 0);
271 	if (timo)
272 		sleepq_set_timeout(ident, timo);
273 
274 	/*
275 	 * Can't call ktrace with any spin locks held so it can lock the
276 	 * ktrace_mtx lock, and WITNESS_WARN considers it an error to hold
277 	 * any spin lock.  Thus, we have to drop the sleepq spin lock while
278 	 * we handle those requests.  This is safe since we have placed our
279 	 * thread on the sleep queue already.
280 	 */
281 #ifdef KTRACE
282 	if (KTRPOINT(td, KTR_CSW)) {
283 		sleepq_release(ident);
284 		ktrcsw(1, 0);
285 		sleepq_lock(ident);
286 	}
287 #endif
288 #ifdef WITNESS
289 	sleepq_release(ident);
290 	WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL, "Sleeping on \"%s\"",
291 	    wmesg);
292 	sleepq_lock(ident);
293 #endif
294 	if (timo)
295 		rval = sleepq_timedwait(ident);
296 	else {
297 		sleepq_wait(ident);
298 		rval = 0;
299 	}
300 #ifdef KTRACE
301 	if (KTRPOINT(td, KTR_CSW))
302 		ktrcsw(0, 0);
303 #endif
304 	PICKUP_GIANT();
305 	mtx_lock_spin(mtx);
306 	WITNESS_RESTORE(&mtx->mtx_object, mtx);
307 	return (rval);
308 }
309 
310 /*
311  * Make all threads sleeping on the specified identifier runnable.
312  */
313 void
314 wakeup(ident)
315 	register void *ident;
316 {
317 
318 	sleepq_lock(ident);
319 	sleepq_broadcast(ident, SLEEPQ_MSLEEP, -1, 0);
320 }
321 
322 /*
323  * Make a thread sleeping on the specified identifier runnable.
324  * May wake more than one thread if a target thread is currently
325  * swapped out.
326  */
327 void
328 wakeup_one(ident)
329 	register void *ident;
330 {
331 
332 	sleepq_lock(ident);
333 	sleepq_signal(ident, SLEEPQ_MSLEEP, -1, 0);
334 }
335 
336 /*
337  * The machine independent parts of context switching.
338  */
339 void
340 mi_switch(int flags, struct thread *newtd)
341 {
342 	uint64_t new_switchtime;
343 	struct thread *td;
344 	struct proc *p;
345 
346 	mtx_assert(&sched_lock, MA_OWNED | MA_NOTRECURSED);
347 	td = curthread;			/* XXX */
348 	p = td->td_proc;		/* XXX */
349 	KASSERT(!TD_ON_RUNQ(td), ("mi_switch: called by old code"));
350 #ifdef INVARIANTS
351 	if (!TD_ON_LOCK(td) && !TD_IS_RUNNING(td))
352 		mtx_assert(&Giant, MA_NOTOWNED);
353 #endif
354 	KASSERT(td->td_critnest == 1 || (td->td_critnest == 2 &&
355 	    (td->td_owepreempt) && (flags & SW_INVOL) != 0 &&
356 	    newtd == NULL) || panicstr,
357 	    ("mi_switch: switch in a critical section"));
358 	KASSERT((flags & (SW_INVOL | SW_VOL)) != 0,
359 	    ("mi_switch: switch must be voluntary or involuntary"));
360 	KASSERT(newtd != curthread, ("mi_switch: preempting back to ourself"));
361 
362 	/*
363 	 * Don't perform context switches from the debugger.
364 	 */
365 	if (kdb_active) {
366 		mtx_unlock_spin(&sched_lock);
367 		kdb_backtrace();
368 		kdb_reenter();
369 		panic("%s: did not reenter debugger", __func__);
370 	}
371 
372 	if (flags & SW_VOL)
373 		p->p_stats->p_ru.ru_nvcsw++;
374 	else
375 		p->p_stats->p_ru.ru_nivcsw++;
376 
377 	/*
378 	 * Compute the amount of time during which the current
379 	 * process was running, and add that to its total so far.
380 	 */
381 	new_switchtime = cpu_ticks();
382 	p->p_rux.rux_runtime += (new_switchtime - PCPU_GET(switchtime));
383 	p->p_rux.rux_uticks += td->td_uticks;
384 	td->td_uticks = 0;
385 	p->p_rux.rux_iticks += td->td_iticks;
386 	td->td_iticks = 0;
387 	p->p_rux.rux_sticks += td->td_sticks;
388 	td->td_sticks = 0;
389 
390 	td->td_generation++;	/* bump preempt-detect counter */
391 
392 	/*
393 	 * Check if the process exceeds its cpu resource allocation.  If
394 	 * it reaches the max, arrange to kill the process in ast().
395 	 */
396 	if (p->p_cpulimit != RLIM_INFINITY &&
397 	    p->p_rux.rux_runtime >= p->p_cpulimit * cpu_tickrate()) {
398 		p->p_sflag |= PS_XCPU;
399 		td->td_flags |= TDF_ASTPENDING;
400 	}
401 
402 	/*
403 	 * Finish up stats for outgoing thread.
404 	 */
405 	cnt.v_swtch++;
406 	PCPU_SET(switchtime, new_switchtime);
407 	PCPU_SET(switchticks, ticks);
408 	CTR4(KTR_PROC, "mi_switch: old thread %p (kse %p, pid %ld, %s)",
409 	    (void *)td, td->td_sched, (long)p->p_pid, p->p_comm);
410 #if (KTR_COMPILE & KTR_SCHED) != 0
411 	if (td == PCPU_GET(idlethread))
412 		CTR3(KTR_SCHED, "mi_switch: %p(%s) prio %d idle",
413 		    td, td->td_proc->p_comm, td->td_priority);
414 	else if (newtd != NULL)
415 		CTR5(KTR_SCHED,
416 		    "mi_switch: %p(%s) prio %d preempted by %p(%s)",
417 		    td, td->td_proc->p_comm, td->td_priority, newtd,
418 		    newtd->td_proc->p_comm);
419 	else
420 		CTR6(KTR_SCHED,
421 		    "mi_switch: %p(%s) prio %d inhibit %d wmesg %s lock %s",
422 		    td, td->td_proc->p_comm, td->td_priority,
423 		    td->td_inhibitors, td->td_wmesg, td->td_lockname);
424 #endif
425 	/*
426 	 * We call thread_switchout after the KTR_SCHED prints above so kse
427 	 * selecting a new thread to run does not show up as a preemption.
428 	 */
429 #ifdef KSE
430 	if ((flags & SW_VOL) && (td->td_proc->p_flag & P_SA))
431 		newtd = thread_switchout(td, flags, newtd);
432 #endif
433 	sched_switch(td, newtd, flags);
434 	CTR3(KTR_SCHED, "mi_switch: running %p(%s) prio %d",
435 	    td, td->td_proc->p_comm, td->td_priority);
436 
437 	CTR4(KTR_PROC, "mi_switch: new thread %p (kse %p, pid %ld, %s)",
438 	    (void *)td, td->td_sched, (long)p->p_pid, p->p_comm);
439 
440 	/*
441 	 * If the last thread was exiting, finish cleaning it up.
442 	 */
443 	if ((td = PCPU_GET(deadthread))) {
444 		PCPU_SET(deadthread, NULL);
445 		thread_stash(td);
446 	}
447 }
448 
449 /*
450  * Change process state to be runnable,
451  * placing it on the run queue if it is in memory,
452  * and awakening the swapper if it isn't in memory.
453  */
454 void
455 setrunnable(struct thread *td)
456 {
457 	struct proc *p;
458 
459 	p = td->td_proc;
460 	mtx_assert(&sched_lock, MA_OWNED);
461 	switch (p->p_state) {
462 	case PRS_ZOMBIE:
463 		panic("setrunnable(1)");
464 	default:
465 		break;
466 	}
467 	switch (td->td_state) {
468 	case TDS_RUNNING:
469 	case TDS_RUNQ:
470 		return;
471 	case TDS_INHIBITED:
472 		/*
473 		 * If we are only inhibited because we are swapped out
474 		 * then arange to swap in this process. Otherwise just return.
475 		 */
476 		if (td->td_inhibitors != TDI_SWAPPED)
477 			return;
478 		/* XXX: intentional fall-through ? */
479 	case TDS_CAN_RUN:
480 		break;
481 	default:
482 		printf("state is 0x%x", td->td_state);
483 		panic("setrunnable(2)");
484 	}
485 	if ((p->p_sflag & PS_INMEM) == 0) {
486 		if ((p->p_sflag & PS_SWAPPINGIN) == 0) {
487 			p->p_sflag |= PS_SWAPINREQ;
488 			/*
489 			 * due to a LOR between sched_lock and
490 			 * the sleepqueue chain locks, use
491 			 * lower level scheduling functions.
492 			 */
493 			kick_proc0();
494 		}
495 	} else
496 		sched_wakeup(td);
497 }
498 
499 /*
500  * Compute a tenex style load average of a quantity on
501  * 1, 5 and 15 minute intervals.
502  * XXXKSE   Needs complete rewrite when correct info is available.
503  * Completely Bogus.. only works with 1:1 (but compiles ok now :-)
504  */
505 static void
506 loadav(void *arg)
507 {
508 	int i, nrun;
509 	struct loadavg *avg;
510 
511 	nrun = sched_load();
512 	avg = &averunnable;
513 
514 	for (i = 0; i < 3; i++)
515 		avg->ldavg[i] = (cexp[i] * avg->ldavg[i] +
516 		    nrun * FSCALE * (FSCALE - cexp[i])) >> FSHIFT;
517 
518 	/*
519 	 * Schedule the next update to occur after 5 seconds, but add a
520 	 * random variation to avoid synchronisation with processes that
521 	 * run at regular intervals.
522 	 */
523 	callout_reset(&loadav_callout, hz * 4 + (int)(random() % (hz * 2 + 1)),
524 	    loadav, NULL);
525 }
526 
527 static void
528 lboltcb(void *arg)
529 {
530 	wakeup(&lbolt);
531 	callout_reset(&lbolt_callout, hz, lboltcb, NULL);
532 }
533 
534 /* ARGSUSED */
535 static void
536 synch_setup(dummy)
537 	void *dummy;
538 {
539 	callout_init(&loadav_callout, CALLOUT_MPSAFE);
540 	callout_init(&lbolt_callout, CALLOUT_MPSAFE);
541 
542 	/* Kick off timeout driven events by calling first time. */
543 	loadav(NULL);
544 	lboltcb(NULL);
545 }
546 
547 /*
548  * General purpose yield system call
549  */
550 int
551 yield(struct thread *td, struct yield_args *uap)
552 {
553 	mtx_assert(&Giant, MA_NOTOWNED);
554 	(void)uap;
555 	sched_relinquish(td);
556 	return (0);
557 }
558