xref: /freebsd/sys/kern/kern_synch.c (revision 1e413cf93298b5b97441a21d9a50fdcd0ee9945e)
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 static int pause_wchan;
73 
74 static struct callout loadav_callout;
75 static struct callout lbolt_callout;
76 
77 struct loadavg averunnable =
78 	{ {0, 0, 0}, FSCALE };	/* load average, of runnable procs */
79 /*
80  * Constants for averages over 1, 5, and 15 minutes
81  * when sampling at 5 second intervals.
82  */
83 static fixpt_t cexp[3] = {
84 	0.9200444146293232 * FSCALE,	/* exp(-1/12) */
85 	0.9834714538216174 * FSCALE,	/* exp(-1/60) */
86 	0.9944598480048967 * FSCALE,	/* exp(-1/180) */
87 };
88 
89 /* kernel uses `FSCALE', userland (SHOULD) use kern.fscale */
90 static int      fscale __unused = FSCALE;
91 SYSCTL_INT(_kern, OID_AUTO, fscale, CTLFLAG_RD, 0, FSCALE, "");
92 
93 static void	loadav(void *arg);
94 static void	lboltcb(void *arg);
95 
96 void
97 sleepinit(void)
98 {
99 
100 	hogticks = (hz / 10) * 2;	/* Default only. */
101 	init_sleepqueues();
102 }
103 
104 /*
105  * General sleep call.  Suspends the current thread until a wakeup is
106  * performed on the specified identifier.  The thread will then be made
107  * runnable with the specified priority.  Sleeps at most timo/hz seconds
108  * (0 means no timeout).  If pri includes PCATCH flag, signals are checked
109  * before and after sleeping, else signals are not checked.  Returns 0 if
110  * awakened, EWOULDBLOCK if the timeout expires.  If PCATCH is set and a
111  * signal needs to be delivered, ERESTART is returned if the current system
112  * call should be restarted if possible, and EINTR is returned if the system
113  * call should be interrupted by the signal (return EINTR).
114  *
115  * The lock argument is unlocked before the caller is suspended, and
116  * re-locked before _sleep() returns.  If priority includes the PDROP
117  * flag the lock is not re-locked before returning.
118  */
119 int
120 _sleep(ident, lock, priority, wmesg, timo)
121 	void *ident;
122 	struct lock_object *lock;
123 	int priority, timo;
124 	const char *wmesg;
125 {
126 	struct thread *td;
127 	struct proc *p;
128 	struct lock_class *class;
129 	int catch, flags, lock_state, pri, rval;
130 	WITNESS_SAVE_DECL(lock_witness);
131 
132 	td = curthread;
133 	p = td->td_proc;
134 #ifdef KTRACE
135 	if (KTRPOINT(td, KTR_CSW))
136 		ktrcsw(1, 0);
137 #endif
138 	WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, lock,
139 	    "Sleeping on \"%s\"", wmesg);
140 	KASSERT(timo != 0 || mtx_owned(&Giant) || lock != NULL ||
141 	    ident == &lbolt, ("sleeping without a lock"));
142 	KASSERT(p != NULL, ("msleep1"));
143 	KASSERT(ident != NULL && TD_IS_RUNNING(td), ("msleep"));
144 	if (lock != NULL)
145 		class = LOCK_CLASS(lock);
146 	else
147 		class = NULL;
148 
149 	if (cold) {
150 		/*
151 		 * During autoconfiguration, just return;
152 		 * don't run any other threads or panic below,
153 		 * in case this is the idle thread and already asleep.
154 		 * XXX: this used to do "s = splhigh(); splx(safepri);
155 		 * splx(s);" to give interrupts a chance, but there is
156 		 * no way to give interrupts a chance now.
157 		 */
158 		if (lock != NULL && priority & PDROP)
159 			class->lc_unlock(lock);
160 		return (0);
161 	}
162 	catch = priority & PCATCH;
163 	rval = 0;
164 
165 	/*
166 	 * If we are already on a sleep queue, then remove us from that
167 	 * sleep queue first.  We have to do this to handle recursive
168 	 * sleeps.
169 	 */
170 	if (TD_ON_SLEEPQ(td))
171 		sleepq_remove(td, td->td_wchan);
172 
173 	if (ident == &pause_wchan)
174 		flags = SLEEPQ_PAUSE;
175 	else
176 		flags = SLEEPQ_SLEEP;
177 	if (catch)
178 		flags |= SLEEPQ_INTERRUPTIBLE;
179 
180 	sleepq_lock(ident);
181 	CTR5(KTR_PROC, "sleep: thread %ld (pid %ld, %s) on %s (%p)",
182 	    td->td_tid, p->p_pid, td->td_name, wmesg, ident);
183 
184 	DROP_GIANT();
185 	if (lock != NULL && !(class->lc_flags & LC_SLEEPABLE)) {
186 		WITNESS_SAVE(lock, lock_witness);
187 		lock_state = class->lc_unlock(lock);
188 	} else
189 		/* GCC needs to follow the Yellow Brick Road */
190 		lock_state = -1;
191 
192 	/*
193 	 * We put ourselves on the sleep queue and start our timeout
194 	 * before calling thread_suspend_check, as we could stop there,
195 	 * and a wakeup or a SIGCONT (or both) could occur while we were
196 	 * stopped without resuming us.  Thus, we must be ready for sleep
197 	 * when cursig() is called.  If the wakeup happens while we're
198 	 * stopped, then td will no longer be on a sleep queue upon
199 	 * return from cursig().
200 	 */
201 	sleepq_add(ident, ident == &lbolt ? NULL : lock, wmesg, flags, 0);
202 	if (timo)
203 		sleepq_set_timeout(ident, timo);
204 	if (lock != NULL && class->lc_flags & LC_SLEEPABLE) {
205 		sleepq_release(ident);
206 		WITNESS_SAVE(lock, lock_witness);
207 		lock_state = class->lc_unlock(lock);
208 		sleepq_lock(ident);
209 	}
210 
211 	/*
212 	 * Adjust this thread's priority, if necessary.
213 	 */
214 	pri = priority & PRIMASK;
215 	if (pri != 0 && pri != td->td_priority) {
216 		thread_lock(td);
217 		sched_prio(td, pri);
218 		thread_unlock(td);
219 	}
220 
221 	if (timo && catch)
222 		rval = sleepq_timedwait_sig(ident);
223 	else if (timo)
224 		rval = sleepq_timedwait(ident);
225 	else if (catch)
226 		rval = sleepq_wait_sig(ident);
227 	else {
228 		sleepq_wait(ident);
229 		rval = 0;
230 	}
231 #ifdef KTRACE
232 	if (KTRPOINT(td, KTR_CSW))
233 		ktrcsw(0, 0);
234 #endif
235 	PICKUP_GIANT();
236 	if (lock != NULL && !(priority & PDROP)) {
237 		class->lc_lock(lock, lock_state);
238 		WITNESS_RESTORE(lock, lock_witness);
239 	}
240 	return (rval);
241 }
242 
243 int
244 msleep_spin(ident, mtx, wmesg, timo)
245 	void *ident;
246 	struct mtx *mtx;
247 	const char *wmesg;
248 	int timo;
249 {
250 	struct thread *td;
251 	struct proc *p;
252 	int rval;
253 	WITNESS_SAVE_DECL(mtx);
254 
255 	td = curthread;
256 	p = td->td_proc;
257 	KASSERT(mtx != NULL, ("sleeping without a mutex"));
258 	KASSERT(p != NULL, ("msleep1"));
259 	KASSERT(ident != NULL && TD_IS_RUNNING(td), ("msleep"));
260 
261 	if (cold) {
262 		/*
263 		 * During autoconfiguration, just return;
264 		 * don't run any other threads or panic below,
265 		 * in case this is the idle thread and already asleep.
266 		 * XXX: this used to do "s = splhigh(); splx(safepri);
267 		 * splx(s);" to give interrupts a chance, but there is
268 		 * no way to give interrupts a chance now.
269 		 */
270 		return (0);
271 	}
272 
273 	sleepq_lock(ident);
274 	CTR5(KTR_PROC, "msleep_spin: thread %ld (pid %ld, %s) on %s (%p)",
275 	    td->td_tid, p->p_pid, td->td_name, wmesg, ident);
276 
277 	DROP_GIANT();
278 	mtx_assert(mtx, MA_OWNED | MA_NOTRECURSED);
279 	WITNESS_SAVE(&mtx->lock_object, mtx);
280 	mtx_unlock_spin(mtx);
281 
282 	/*
283 	 * We put ourselves on the sleep queue and start our timeout.
284 	 */
285 	sleepq_add(ident, &mtx->lock_object, wmesg, SLEEPQ_SLEEP, 0);
286 	if (timo)
287 		sleepq_set_timeout(ident, timo);
288 
289 	/*
290 	 * Can't call ktrace with any spin locks held so it can lock the
291 	 * ktrace_mtx lock, and WITNESS_WARN considers it an error to hold
292 	 * any spin lock.  Thus, we have to drop the sleepq spin lock while
293 	 * we handle those requests.  This is safe since we have placed our
294 	 * thread on the sleep queue already.
295 	 */
296 #ifdef KTRACE
297 	if (KTRPOINT(td, KTR_CSW)) {
298 		sleepq_release(ident);
299 		ktrcsw(1, 0);
300 		sleepq_lock(ident);
301 	}
302 #endif
303 #ifdef WITNESS
304 	sleepq_release(ident);
305 	WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL, "Sleeping on \"%s\"",
306 	    wmesg);
307 	sleepq_lock(ident);
308 #endif
309 	if (timo)
310 		rval = sleepq_timedwait(ident);
311 	else {
312 		sleepq_wait(ident);
313 		rval = 0;
314 	}
315 #ifdef KTRACE
316 	if (KTRPOINT(td, KTR_CSW))
317 		ktrcsw(0, 0);
318 #endif
319 	PICKUP_GIANT();
320 	mtx_lock_spin(mtx);
321 	WITNESS_RESTORE(&mtx->lock_object, mtx);
322 	return (rval);
323 }
324 
325 /*
326  * pause() is like tsleep() except that the intention is to not be
327  * explicitly woken up by another thread.  Instead, the current thread
328  * simply wishes to sleep until the timeout expires.  It is
329  * implemented using a dummy wait channel.
330  */
331 int
332 pause(wmesg, timo)
333 	const char *wmesg;
334 	int timo;
335 {
336 
337 	KASSERT(timo != 0, ("pause: timeout required"));
338 	return (tsleep(&pause_wchan, 0, wmesg, timo));
339 }
340 
341 /*
342  * Make all threads sleeping on the specified identifier runnable.
343  */
344 void
345 wakeup(ident)
346 	register void *ident;
347 {
348 
349 	sleepq_lock(ident);
350 	sleepq_broadcast(ident, SLEEPQ_SLEEP, -1, 0);
351 }
352 
353 /*
354  * Make a thread sleeping on the specified identifier runnable.
355  * May wake more than one thread if a target thread is currently
356  * swapped out.
357  */
358 void
359 wakeup_one(ident)
360 	register void *ident;
361 {
362 
363 	sleepq_lock(ident);
364 	sleepq_signal(ident, SLEEPQ_SLEEP, -1, 0);
365 	sleepq_release(ident);
366 }
367 
368 /*
369  * The machine independent parts of context switching.
370  */
371 void
372 mi_switch(int flags, struct thread *newtd)
373 {
374 	uint64_t runtime, new_switchtime;
375 	struct thread *td;
376 	struct proc *p;
377 
378 	td = curthread;			/* XXX */
379 	THREAD_LOCK_ASSERT(td, MA_OWNED | MA_NOTRECURSED);
380 	p = td->td_proc;		/* XXX */
381 	KASSERT(!TD_ON_RUNQ(td), ("mi_switch: called by old code"));
382 #ifdef INVARIANTS
383 	if (!TD_ON_LOCK(td) && !TD_IS_RUNNING(td))
384 		mtx_assert(&Giant, MA_NOTOWNED);
385 #endif
386 	KASSERT(td->td_critnest == 1 || (td->td_critnest == 2 &&
387 	    (td->td_owepreempt) && (flags & SW_INVOL) != 0 &&
388 	    newtd == NULL) || panicstr,
389 	    ("mi_switch: switch in a critical section"));
390 	KASSERT((flags & (SW_INVOL | SW_VOL)) != 0,
391 	    ("mi_switch: switch must be voluntary or involuntary"));
392 	KASSERT(newtd != curthread, ("mi_switch: preempting back to ourself"));
393 
394 	/*
395 	 * Don't perform context switches from the debugger.
396 	 */
397 	if (kdb_active) {
398 		thread_unlock(td);
399 		kdb_backtrace();
400 		kdb_reenter();
401 		panic("%s: did not reenter debugger", __func__);
402 	}
403 	if (flags & SW_VOL)
404 		td->td_ru.ru_nvcsw++;
405 	else
406 		td->td_ru.ru_nivcsw++;
407 	/*
408 	 * Compute the amount of time during which the current
409 	 * thread was running, and add that to its total so far.
410 	 */
411 	new_switchtime = cpu_ticks();
412 	runtime = new_switchtime - PCPU_GET(switchtime);
413 	td->td_runtime += runtime;
414 	td->td_incruntime += runtime;
415 	PCPU_SET(switchtime, new_switchtime);
416 	td->td_generation++;	/* bump preempt-detect counter */
417 	PCPU_INC(cnt.v_swtch);
418 	PCPU_SET(switchticks, ticks);
419 	CTR4(KTR_PROC, "mi_switch: old thread %ld (kse %p, pid %ld, %s)",
420 	    td->td_tid, td->td_sched, p->p_pid, td->td_name);
421 #if (KTR_COMPILE & KTR_SCHED) != 0
422 	if (TD_IS_IDLETHREAD(td))
423 		CTR3(KTR_SCHED, "mi_switch: %p(%s) prio %d idle",
424 		    td, td->td_name, td->td_priority);
425 	else if (newtd != NULL)
426 		CTR5(KTR_SCHED,
427 		    "mi_switch: %p(%s) prio %d preempted by %p(%s)",
428 		    td, td->td_name, td->td_priority, newtd,
429 		    newtd->td_name);
430 	else
431 		CTR6(KTR_SCHED,
432 		    "mi_switch: %p(%s) prio %d inhibit %d wmesg %s lock %s",
433 		    td, td->td_name, td->td_priority,
434 		    td->td_inhibitors, td->td_wmesg, td->td_lockname);
435 #endif
436 	/*
437 	 * We call thread_switchout after the KTR_SCHED prints above so kse
438 	 * selecting a new thread to run does not show up as a preemption.
439 	 */
440 #ifdef KSE
441 	if ((flags & SW_VOL) && (td->td_proc->p_flag & P_SA))
442 		newtd = thread_switchout(td, flags, newtd);
443 #endif
444 	sched_switch(td, newtd, flags);
445 	CTR3(KTR_SCHED, "mi_switch: running %p(%s) prio %d",
446 	    td, td->td_name, td->td_priority);
447 
448 	CTR4(KTR_PROC, "mi_switch: new thread %ld (kse %p, pid %ld, %s)",
449 	    td->td_tid, td->td_sched, p->p_pid, td->td_name);
450 
451 	/*
452 	 * If the last thread was exiting, finish cleaning it up.
453 	 */
454 	if ((td = PCPU_GET(deadthread))) {
455 		PCPU_SET(deadthread, NULL);
456 		thread_stash(td);
457 	}
458 }
459 
460 /*
461  * Change process state to be runnable,
462  * placing it on the run queue if it is in memory,
463  * and awakening the swapper if it isn't in memory.
464  */
465 void
466 setrunnable(struct thread *td)
467 {
468 
469 	THREAD_LOCK_ASSERT(td, MA_OWNED);
470 	KASSERT(td->td_proc->p_state != PRS_ZOMBIE,
471 	    ("setrunnable: pid %d is a zombie", td->td_proc->p_pid));
472 	switch (td->td_state) {
473 	case TDS_RUNNING:
474 	case TDS_RUNQ:
475 		return;
476 	case TDS_INHIBITED:
477 		/*
478 		 * If we are only inhibited because we are swapped out
479 		 * then arange to swap in this process. Otherwise just return.
480 		 */
481 		if (td->td_inhibitors != TDI_SWAPPED)
482 			return;
483 		/* XXX: intentional fall-through ? */
484 	case TDS_CAN_RUN:
485 		break;
486 	default:
487 		printf("state is 0x%x", td->td_state);
488 		panic("setrunnable(2)");
489 	}
490 	if ((td->td_flags & TDF_INMEM) == 0) {
491 		if ((td->td_flags & TDF_SWAPINREQ) == 0) {
492 			td->td_flags |= TDF_SWAPINREQ;
493 			/*
494 			 * due to a LOR between the thread lock and
495 			 * the sleepqueue chain locks, use
496 			 * lower level scheduling functions.
497 			 */
498 			kick_proc0();
499 		}
500 	} else
501 		sched_wakeup(td);
502 }
503 
504 /*
505  * Compute a tenex style load average of a quantity on
506  * 1, 5 and 15 minute intervals.
507  * XXXKSE   Needs complete rewrite when correct info is available.
508  * Completely Bogus.. only works with 1:1 (but compiles ok now :-)
509  */
510 static void
511 loadav(void *arg)
512 {
513 	int i, nrun;
514 	struct loadavg *avg;
515 
516 	nrun = sched_load();
517 	avg = &averunnable;
518 
519 	for (i = 0; i < 3; i++)
520 		avg->ldavg[i] = (cexp[i] * avg->ldavg[i] +
521 		    nrun * FSCALE * (FSCALE - cexp[i])) >> FSHIFT;
522 
523 	/*
524 	 * Schedule the next update to occur after 5 seconds, but add a
525 	 * random variation to avoid synchronisation with processes that
526 	 * run at regular intervals.
527 	 */
528 	callout_reset(&loadav_callout, hz * 4 + (int)(random() % (hz * 2 + 1)),
529 	    loadav, NULL);
530 }
531 
532 static void
533 lboltcb(void *arg)
534 {
535 	wakeup(&lbolt);
536 	callout_reset(&lbolt_callout, hz, lboltcb, NULL);
537 }
538 
539 /* ARGSUSED */
540 static void
541 synch_setup(dummy)
542 	void *dummy;
543 {
544 	callout_init(&loadav_callout, CALLOUT_MPSAFE);
545 	callout_init(&lbolt_callout, CALLOUT_MPSAFE);
546 
547 	/* Kick off timeout driven events by calling first time. */
548 	loadav(NULL);
549 	lboltcb(NULL);
550 }
551 
552 /*
553  * General purpose yield system call.
554  */
555 int
556 yield(struct thread *td, struct yield_args *uap)
557 {
558 
559 	thread_lock(td);
560 	sched_prio(td, PRI_MAX_TIMESHARE);
561 	mi_switch(SW_VOL, NULL);
562 	thread_unlock(td);
563 	td->td_retval[0] = 0;
564 	return (0);
565 }
566