xref: /freebsd/sys/kern/kern_synch.c (revision d876124d6ae9d56da5b4ff4c6015efd1d0c9222a)
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 #include "opt_sched.h"
42 
43 #include <sys/param.h>
44 #include <sys/systm.h>
45 #include <sys/condvar.h>
46 #include <sys/kdb.h>
47 #include <sys/kernel.h>
48 #include <sys/ktr.h>
49 #include <sys/lock.h>
50 #include <sys/mutex.h>
51 #include <sys/proc.h>
52 #include <sys/resourcevar.h>
53 #include <sys/sched.h>
54 #include <sys/signalvar.h>
55 #include <sys/sleepqueue.h>
56 #include <sys/smp.h>
57 #include <sys/sx.h>
58 #include <sys/sysctl.h>
59 #include <sys/sysproto.h>
60 #include <sys/vmmeter.h>
61 #ifdef KTRACE
62 #include <sys/uio.h>
63 #include <sys/ktrace.h>
64 #endif
65 
66 #include <machine/cpu.h>
67 
68 static void synch_setup(void *dummy);
69 SYSINIT(synch_setup, SI_SUB_KICK_SCHEDULER, SI_ORDER_FIRST, synch_setup,
70     NULL);
71 
72 int	hogticks;
73 int	lbolt;
74 static int pause_wchan;
75 
76 static struct callout loadav_callout;
77 static struct callout lbolt_callout;
78 
79 struct loadavg averunnable =
80 	{ {0, 0, 0}, FSCALE };	/* load average, of runnable procs */
81 /*
82  * Constants for averages over 1, 5, and 15 minutes
83  * when sampling at 5 second intervals.
84  */
85 static fixpt_t cexp[3] = {
86 	0.9200444146293232 * FSCALE,	/* exp(-1/12) */
87 	0.9834714538216174 * FSCALE,	/* exp(-1/60) */
88 	0.9944598480048967 * FSCALE,	/* exp(-1/180) */
89 };
90 
91 /* kernel uses `FSCALE', userland (SHOULD) use kern.fscale */
92 static int      fscale __unused = FSCALE;
93 SYSCTL_INT(_kern, OID_AUTO, fscale, CTLFLAG_RD, 0, FSCALE, "");
94 
95 static void	loadav(void *arg);
96 static void	lboltcb(void *arg);
97 
98 void
99 sleepinit(void)
100 {
101 
102 	hogticks = (hz / 10) * 2;	/* Default only. */
103 	init_sleepqueues();
104 }
105 
106 /*
107  * General sleep call.  Suspends the current thread until a wakeup is
108  * performed on the specified identifier.  The thread will then be made
109  * runnable with the specified priority.  Sleeps at most timo/hz seconds
110  * (0 means no timeout).  If pri includes PCATCH flag, signals are checked
111  * before and after sleeping, else signals are not checked.  Returns 0 if
112  * awakened, EWOULDBLOCK if the timeout expires.  If PCATCH is set and a
113  * signal needs to be delivered, ERESTART is returned if the current system
114  * call should be restarted if possible, and EINTR is returned if the system
115  * call should be interrupted by the signal (return EINTR).
116  *
117  * The lock argument is unlocked before the caller is suspended, and
118  * re-locked before _sleep() returns.  If priority includes the PDROP
119  * flag the lock is not re-locked before returning.
120  */
121 int
122 _sleep(void *ident, struct lock_object *lock, int priority,
123     const char *wmesg, int timo)
124 {
125 	struct thread *td;
126 	struct proc *p;
127 	struct lock_class *class;
128 	int catch, flags, lock_state, pri, rval;
129 	WITNESS_SAVE_DECL(lock_witness);
130 
131 	td = curthread;
132 	p = td->td_proc;
133 #ifdef KTRACE
134 	if (KTRPOINT(td, KTR_CSW))
135 		ktrcsw(1, 0);
136 #endif
137 	WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, lock,
138 	    "Sleeping on \"%s\"", wmesg);
139 	KASSERT(timo != 0 || mtx_owned(&Giant) || lock != NULL ||
140 	    ident == &lbolt, ("sleeping without a lock"));
141 	KASSERT(p != NULL, ("msleep1"));
142 	KASSERT(ident != NULL && TD_IS_RUNNING(td), ("msleep"));
143 	if (lock != NULL)
144 		class = LOCK_CLASS(lock);
145 	else
146 		class = NULL;
147 
148 	if (cold) {
149 		/*
150 		 * During autoconfiguration, just return;
151 		 * don't run any other threads or panic below,
152 		 * in case this is the idle thread and already asleep.
153 		 * XXX: this used to do "s = splhigh(); splx(safepri);
154 		 * splx(s);" to give interrupts a chance, but there is
155 		 * no way to give interrupts a chance now.
156 		 */
157 		if (lock != NULL && priority & PDROP)
158 			class->lc_unlock(lock);
159 		return (0);
160 	}
161 	catch = priority & PCATCH;
162 	pri = priority & PRIMASK;
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 	if (timo && catch)
211 		rval = sleepq_timedwait_sig(ident, pri);
212 	else if (timo)
213 		rval = sleepq_timedwait(ident, pri);
214 	else if (catch)
215 		rval = sleepq_wait_sig(ident, pri);
216 	else {
217 		sleepq_wait(ident, pri);
218 		rval = 0;
219 	}
220 #ifdef KTRACE
221 	if (KTRPOINT(td, KTR_CSW))
222 		ktrcsw(0, 0);
223 #endif
224 	PICKUP_GIANT();
225 	if (lock != NULL && !(priority & PDROP)) {
226 		class->lc_lock(lock, lock_state);
227 		WITNESS_RESTORE(lock, lock_witness);
228 	}
229 	return (rval);
230 }
231 
232 int
233 msleep_spin(void *ident, struct mtx *mtx, const char *wmesg, 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 %ld (pid %ld, %s) on %s (%p)",
260 	    td->td_tid, p->p_pid, td->td_name, wmesg, ident);
261 
262 	DROP_GIANT();
263 	mtx_assert(mtx, MA_OWNED | MA_NOTRECURSED);
264 	WITNESS_SAVE(&mtx->lock_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->lock_object, wmesg, SLEEPQ_SLEEP, 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, 0);
296 	else {
297 		sleepq_wait(ident, 0);
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->lock_object, mtx);
307 	return (rval);
308 }
309 
310 /*
311  * pause() is like tsleep() except that the intention is to not be
312  * explicitly woken up by another thread.  Instead, the current thread
313  * simply wishes to sleep until the timeout expires.  It is
314  * implemented using a dummy wait channel.
315  */
316 int
317 pause(const char *wmesg, int timo)
318 {
319 
320 	KASSERT(timo != 0, ("pause: timeout required"));
321 	return (tsleep(&pause_wchan, 0, wmesg, timo));
322 }
323 
324 /*
325  * Make all threads sleeping on the specified identifier runnable.
326  */
327 void
328 wakeup(void *ident)
329 {
330 
331 	sleepq_lock(ident);
332 	sleepq_broadcast(ident, SLEEPQ_SLEEP, 0, 0);
333 	sleepq_release(ident);
334 }
335 
336 /*
337  * Make a thread sleeping on the specified identifier runnable.
338  * May wake more than one thread if a target thread is currently
339  * swapped out.
340  */
341 void
342 wakeup_one(void *ident)
343 {
344 
345 	sleepq_lock(ident);
346 	sleepq_signal(ident, SLEEPQ_SLEEP, 0, 0);
347 	sleepq_release(ident);
348 }
349 
350 static void
351 kdb_switch(void)
352 {
353 	thread_unlock(curthread);
354 	kdb_backtrace();
355 	kdb_reenter();
356 	panic("%s: did not reenter debugger", __func__);
357 }
358 
359 /*
360  * The machine independent parts of context switching.
361  */
362 void
363 mi_switch(int flags, struct thread *newtd)
364 {
365 	uint64_t runtime, new_switchtime;
366 	struct thread *td;
367 	struct proc *p;
368 
369 	td = curthread;			/* XXX */
370 	THREAD_LOCK_ASSERT(td, MA_OWNED | MA_NOTRECURSED);
371 	p = td->td_proc;		/* XXX */
372 	KASSERT(!TD_ON_RUNQ(td), ("mi_switch: called by old code"));
373 #ifdef INVARIANTS
374 	if (!TD_ON_LOCK(td) && !TD_IS_RUNNING(td))
375 		mtx_assert(&Giant, MA_NOTOWNED);
376 #endif
377 	KASSERT(td->td_critnest == 1 || (td->td_critnest == 2 &&
378 	    (td->td_owepreempt) && (flags & SW_INVOL) != 0 &&
379 	    newtd == NULL) || panicstr,
380 	    ("mi_switch: switch in a critical section"));
381 	KASSERT((flags & (SW_INVOL | SW_VOL)) != 0,
382 	    ("mi_switch: switch must be voluntary or involuntary"));
383 	KASSERT(newtd != curthread, ("mi_switch: preempting back to ourself"));
384 
385 	/*
386 	 * Don't perform context switches from the debugger.
387 	 */
388 	if (kdb_active)
389 		kdb_switch();
390 	if (flags & SW_VOL)
391 		td->td_ru.ru_nvcsw++;
392 	else
393 		td->td_ru.ru_nivcsw++;
394 #ifdef SCHED_STATS
395 	SCHED_STAT_INC(sched_switch_stats[flags & SW_TYPE_MASK]);
396 #endif
397 	/*
398 	 * Compute the amount of time during which the current
399 	 * thread was running, and add that to its total so far.
400 	 */
401 	new_switchtime = cpu_ticks();
402 	runtime = new_switchtime - PCPU_GET(switchtime);
403 	td->td_runtime += runtime;
404 	td->td_incruntime += runtime;
405 	PCPU_SET(switchtime, new_switchtime);
406 	td->td_generation++;	/* bump preempt-detect counter */
407 	PCPU_INC(cnt.v_swtch);
408 	PCPU_SET(switchticks, ticks);
409 	CTR4(KTR_PROC, "mi_switch: old thread %ld (td_sched %p, pid %ld, %s)",
410 	    td->td_tid, td->td_sched, p->p_pid, td->td_name);
411 #if (KTR_COMPILE & KTR_SCHED) != 0
412 	if (TD_IS_IDLETHREAD(td))
413 		CTR3(KTR_SCHED, "mi_switch: %p(%s) prio %d idle",
414 		    td, td->td_name, td->td_priority);
415 	else if (newtd != NULL)
416 		CTR5(KTR_SCHED,
417 		    "mi_switch: %p(%s) prio %d preempted by %p(%s)",
418 		    td, td->td_name, td->td_priority, newtd,
419 		    newtd->td_name);
420 	else
421 		CTR6(KTR_SCHED,
422 		    "mi_switch: %p(%s) prio %d inhibit %d wmesg %s lock %s",
423 		    td, td->td_name, td->td_priority,
424 		    td->td_inhibitors, td->td_wmesg, td->td_lockname);
425 #endif
426 	sched_switch(td, newtd, flags);
427 	CTR3(KTR_SCHED, "mi_switch: running %p(%s) prio %d",
428 	    td, td->td_name, td->td_priority);
429 
430 	CTR4(KTR_PROC, "mi_switch: new thread %ld (td_sched %p, pid %ld, %s)",
431 	    td->td_tid, td->td_sched, p->p_pid, td->td_name);
432 
433 	/*
434 	 * If the last thread was exiting, finish cleaning it up.
435 	 */
436 	if ((td = PCPU_GET(deadthread))) {
437 		PCPU_SET(deadthread, NULL);
438 		thread_stash(td);
439 	}
440 }
441 
442 /*
443  * Change process state to be runnable,
444  * placing it on the run queue if it is in memory,
445  * and awakening the swapper if it isn't in memory.
446  */
447 void
448 setrunnable(struct thread *td)
449 {
450 
451 	THREAD_LOCK_ASSERT(td, MA_OWNED);
452 	KASSERT(td->td_proc->p_state != PRS_ZOMBIE,
453 	    ("setrunnable: pid %d is a zombie", td->td_proc->p_pid));
454 	switch (td->td_state) {
455 	case TDS_RUNNING:
456 	case TDS_RUNQ:
457 		return;
458 	case TDS_INHIBITED:
459 		/*
460 		 * If we are only inhibited because we are swapped out
461 		 * then arange to swap in this process. Otherwise just return.
462 		 */
463 		if (td->td_inhibitors != TDI_SWAPPED)
464 			return;
465 		/* XXX: intentional fall-through ? */
466 	case TDS_CAN_RUN:
467 		break;
468 	default:
469 		printf("state is 0x%x", td->td_state);
470 		panic("setrunnable(2)");
471 	}
472 	if ((td->td_flags & TDF_INMEM) == 0) {
473 		if ((td->td_flags & TDF_SWAPINREQ) == 0) {
474 			td->td_flags |= TDF_SWAPINREQ;
475 			/*
476 			 * due to a LOR between the thread lock and
477 			 * the sleepqueue chain locks, use
478 			 * lower level scheduling functions.
479 			 */
480 			kick_proc0();
481 		}
482 	} else
483 		sched_wakeup(td);
484 }
485 
486 /*
487  * Compute a tenex style load average of a quantity on
488  * 1, 5 and 15 minute intervals.
489  */
490 static void
491 loadav(void *arg)
492 {
493 	int i, nrun;
494 	struct loadavg *avg;
495 
496 	nrun = sched_load();
497 	avg = &averunnable;
498 
499 	for (i = 0; i < 3; i++)
500 		avg->ldavg[i] = (cexp[i] * avg->ldavg[i] +
501 		    nrun * FSCALE * (FSCALE - cexp[i])) >> FSHIFT;
502 
503 	/*
504 	 * Schedule the next update to occur after 5 seconds, but add a
505 	 * random variation to avoid synchronisation with processes that
506 	 * run at regular intervals.
507 	 */
508 	callout_reset(&loadav_callout, hz * 4 + (int)(random() % (hz * 2 + 1)),
509 	    loadav, NULL);
510 }
511 
512 static void
513 lboltcb(void *arg)
514 {
515 	wakeup(&lbolt);
516 	callout_reset(&lbolt_callout, hz, lboltcb, NULL);
517 }
518 
519 /* ARGSUSED */
520 static void
521 synch_setup(void *dummy)
522 {
523 	callout_init(&loadav_callout, CALLOUT_MPSAFE);
524 	callout_init(&lbolt_callout, CALLOUT_MPSAFE);
525 
526 	/* Kick off timeout driven events by calling first time. */
527 	loadav(NULL);
528 	lboltcb(NULL);
529 }
530 
531 /*
532  * General purpose yield system call.
533  */
534 int
535 yield(struct thread *td, struct yield_args *uap)
536 {
537 
538 	thread_lock(td);
539 	sched_prio(td, PRI_MAX_TIMESHARE);
540 	mi_switch(SW_VOL | SWT_RELINQUISH, NULL);
541 	thread_unlock(td);
542 	td->td_retval[0] = 0;
543 	return (0);
544 }
545