xref: /freebsd/sys/kern/kern_synch.c (revision 7d0d268b8a67f28ccefdd0b8ce6fb38acac78d80)
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 #ifdef XEN
69 #include <vm/vm.h>
70 #include <vm/vm_param.h>
71 #include <vm/pmap.h>
72 #endif
73 
74 #define	KTDSTATE(td)							\
75 	(((td)->td_inhibitors & TDI_SLEEPING) != 0 ? "sleep"  :		\
76 	((td)->td_inhibitors & TDI_SUSPENDED) != 0 ? "suspended" :	\
77 	((td)->td_inhibitors & TDI_SWAPPED) != 0 ? "swapped" :		\
78 	((td)->td_inhibitors & TDI_LOCK) != 0 ? "blocked" :		\
79 	((td)->td_inhibitors & TDI_IWAIT) != 0 ? "iwait" : "yielding")
80 
81 static void synch_setup(void *dummy);
82 SYSINIT(synch_setup, SI_SUB_KICK_SCHEDULER, SI_ORDER_FIRST, synch_setup,
83     NULL);
84 
85 int	hogticks;
86 static int pause_wchan;
87 
88 static struct callout loadav_callout;
89 
90 struct loadavg averunnable =
91 	{ {0, 0, 0}, FSCALE };	/* load average, of runnable procs */
92 /*
93  * Constants for averages over 1, 5, and 15 minutes
94  * when sampling at 5 second intervals.
95  */
96 static fixpt_t cexp[3] = {
97 	0.9200444146293232 * FSCALE,	/* exp(-1/12) */
98 	0.9834714538216174 * FSCALE,	/* exp(-1/60) */
99 	0.9944598480048967 * FSCALE,	/* exp(-1/180) */
100 };
101 
102 /* kernel uses `FSCALE', userland (SHOULD) use kern.fscale */
103 static int      fscale __unused = FSCALE;
104 SYSCTL_INT(_kern, OID_AUTO, fscale, CTLFLAG_RD, 0, FSCALE, "");
105 
106 static void	loadav(void *arg);
107 
108 void
109 sleepinit(void)
110 {
111 
112 	hogticks = (hz / 10) * 2;	/* Default only. */
113 	init_sleepqueues();
114 }
115 
116 /*
117  * General sleep call.  Suspends the current thread until a wakeup is
118  * performed on the specified identifier.  The thread will then be made
119  * runnable with the specified priority.  Sleeps at most timo/hz seconds
120  * (0 means no timeout).  If pri includes PCATCH flag, signals are checked
121  * before and after sleeping, else signals are not checked.  Returns 0 if
122  * awakened, EWOULDBLOCK if the timeout expires.  If PCATCH is set and a
123  * signal needs to be delivered, ERESTART is returned if the current system
124  * call should be restarted if possible, and EINTR is returned if the system
125  * call should be interrupted by the signal (return EINTR).
126  *
127  * The lock argument is unlocked before the caller is suspended, and
128  * re-locked before _sleep() returns.  If priority includes the PDROP
129  * flag the lock is not re-locked before returning.
130  */
131 int
132 _sleep(void *ident, struct lock_object *lock, int priority,
133     const char *wmesg, int timo)
134 {
135 	struct thread *td;
136 	struct proc *p;
137 	struct lock_class *class;
138 	int catch, flags, lock_state, pri, rval;
139 	WITNESS_SAVE_DECL(lock_witness);
140 
141 	td = curthread;
142 	p = td->td_proc;
143 #ifdef KTRACE
144 	if (KTRPOINT(td, KTR_CSW))
145 		ktrcsw(1, 0);
146 #endif
147 	WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, lock,
148 	    "Sleeping on \"%s\"", wmesg);
149 	KASSERT(timo != 0 || mtx_owned(&Giant) || lock != NULL,
150 	    ("sleeping without a lock"));
151 	KASSERT(p != NULL, ("msleep1"));
152 	KASSERT(ident != NULL && TD_IS_RUNNING(td), ("msleep"));
153 	if (priority & PDROP)
154 		KASSERT(lock != NULL && lock != &Giant.lock_object,
155 		    ("PDROP requires a non-Giant lock"));
156 	if (lock != NULL)
157 		class = LOCK_CLASS(lock);
158 	else
159 		class = NULL;
160 
161 	if (cold) {
162 		/*
163 		 * During autoconfiguration, just return;
164 		 * don't run any other threads or panic below,
165 		 * in case this is the idle thread and already asleep.
166 		 * XXX: this used to do "s = splhigh(); splx(safepri);
167 		 * splx(s);" to give interrupts a chance, but there is
168 		 * no way to give interrupts a chance now.
169 		 */
170 		if (lock != NULL && priority & PDROP)
171 			class->lc_unlock(lock);
172 		return (0);
173 	}
174 	catch = priority & PCATCH;
175 	pri = priority & PRIMASK;
176 	rval = 0;
177 
178 	/*
179 	 * If we are already on a sleep queue, then remove us from that
180 	 * sleep queue first.  We have to do this to handle recursive
181 	 * sleeps.
182 	 */
183 	if (TD_ON_SLEEPQ(td))
184 		sleepq_remove(td, td->td_wchan);
185 
186 	if (ident == &pause_wchan)
187 		flags = SLEEPQ_PAUSE;
188 	else
189 		flags = SLEEPQ_SLEEP;
190 	if (catch)
191 		flags |= SLEEPQ_INTERRUPTIBLE;
192 
193 	sleepq_lock(ident);
194 	CTR5(KTR_PROC, "sleep: thread %ld (pid %ld, %s) on %s (%p)",
195 	    td->td_tid, p->p_pid, td->td_name, wmesg, ident);
196 
197 	if (lock == &Giant.lock_object)
198 		mtx_assert(&Giant, MA_OWNED);
199 	DROP_GIANT();
200 	if (lock != NULL && lock != &Giant.lock_object &&
201 	    !(class->lc_flags & LC_SLEEPABLE)) {
202 		WITNESS_SAVE(lock, lock_witness);
203 		lock_state = class->lc_unlock(lock);
204 	} else
205 		/* GCC needs to follow the Yellow Brick Road */
206 		lock_state = -1;
207 
208 	/*
209 	 * We put ourselves on the sleep queue and start our timeout
210 	 * before calling thread_suspend_check, as we could stop there,
211 	 * and a wakeup or a SIGCONT (or both) could occur while we were
212 	 * stopped without resuming us.  Thus, we must be ready for sleep
213 	 * when cursig() is called.  If the wakeup happens while we're
214 	 * stopped, then td will no longer be on a sleep queue upon
215 	 * return from cursig().
216 	 */
217 	sleepq_add(ident, lock, wmesg, flags, 0);
218 	if (timo)
219 		sleepq_set_timeout(ident, timo);
220 	if (lock != NULL && class->lc_flags & LC_SLEEPABLE) {
221 		sleepq_release(ident);
222 		WITNESS_SAVE(lock, lock_witness);
223 		lock_state = class->lc_unlock(lock);
224 		sleepq_lock(ident);
225 	}
226 	if (timo && catch)
227 		rval = sleepq_timedwait_sig(ident, pri);
228 	else if (timo)
229 		rval = sleepq_timedwait(ident, pri);
230 	else if (catch)
231 		rval = sleepq_wait_sig(ident, pri);
232 	else {
233 		sleepq_wait(ident, pri);
234 		rval = 0;
235 	}
236 #ifdef KTRACE
237 	if (KTRPOINT(td, KTR_CSW))
238 		ktrcsw(0, 0);
239 #endif
240 	PICKUP_GIANT();
241 	if (lock != NULL && lock != &Giant.lock_object && !(priority & PDROP)) {
242 		class->lc_lock(lock, lock_state);
243 		WITNESS_RESTORE(lock, lock_witness);
244 	}
245 	return (rval);
246 }
247 
248 int
249 msleep_spin(void *ident, struct mtx *mtx, const char *wmesg, int timo)
250 {
251 	struct thread *td;
252 	struct proc *p;
253 	int rval;
254 	WITNESS_SAVE_DECL(mtx);
255 
256 	td = curthread;
257 	p = td->td_proc;
258 	KASSERT(mtx != NULL, ("sleeping without a mutex"));
259 	KASSERT(p != NULL, ("msleep1"));
260 	KASSERT(ident != NULL && TD_IS_RUNNING(td), ("msleep"));
261 
262 	if (cold) {
263 		/*
264 		 * During autoconfiguration, just return;
265 		 * don't run any other threads or panic below,
266 		 * in case this is the idle thread and already asleep.
267 		 * XXX: this used to do "s = splhigh(); splx(safepri);
268 		 * splx(s);" to give interrupts a chance, but there is
269 		 * no way to give interrupts a chance now.
270 		 */
271 		return (0);
272 	}
273 
274 	sleepq_lock(ident);
275 	CTR5(KTR_PROC, "msleep_spin: thread %ld (pid %ld, %s) on %s (%p)",
276 	    td->td_tid, p->p_pid, td->td_name, wmesg, ident);
277 
278 	DROP_GIANT();
279 	mtx_assert(mtx, MA_OWNED | MA_NOTRECURSED);
280 	WITNESS_SAVE(&mtx->lock_object, mtx);
281 	mtx_unlock_spin(mtx);
282 
283 	/*
284 	 * We put ourselves on the sleep queue and start our timeout.
285 	 */
286 	sleepq_add(ident, &mtx->lock_object, wmesg, SLEEPQ_SLEEP, 0);
287 	if (timo)
288 		sleepq_set_timeout(ident, timo);
289 
290 	/*
291 	 * Can't call ktrace with any spin locks held so it can lock the
292 	 * ktrace_mtx lock, and WITNESS_WARN considers it an error to hold
293 	 * any spin lock.  Thus, we have to drop the sleepq spin lock while
294 	 * we handle those requests.  This is safe since we have placed our
295 	 * thread on the sleep queue already.
296 	 */
297 #ifdef KTRACE
298 	if (KTRPOINT(td, KTR_CSW)) {
299 		sleepq_release(ident);
300 		ktrcsw(1, 0);
301 		sleepq_lock(ident);
302 	}
303 #endif
304 #ifdef WITNESS
305 	sleepq_release(ident);
306 	WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL, "Sleeping on \"%s\"",
307 	    wmesg);
308 	sleepq_lock(ident);
309 #endif
310 	if (timo)
311 		rval = sleepq_timedwait(ident, 0);
312 	else {
313 		sleepq_wait(ident, 0);
314 		rval = 0;
315 	}
316 #ifdef KTRACE
317 	if (KTRPOINT(td, KTR_CSW))
318 		ktrcsw(0, 0);
319 #endif
320 	PICKUP_GIANT();
321 	mtx_lock_spin(mtx);
322 	WITNESS_RESTORE(&mtx->lock_object, mtx);
323 	return (rval);
324 }
325 
326 /*
327  * pause() is like tsleep() except that the intention is to not be
328  * explicitly woken up by another thread.  Instead, the current thread
329  * simply wishes to sleep until the timeout expires.  It is
330  * implemented using a dummy wait channel.
331  */
332 int
333 pause(const char *wmesg, int timo)
334 {
335 
336 	KASSERT(timo != 0, ("pause: timeout required"));
337 	return (tsleep(&pause_wchan, 0, wmesg, timo));
338 }
339 
340 /*
341  * Make all threads sleeping on the specified identifier runnable.
342  */
343 void
344 wakeup(void *ident)
345 {
346 	int wakeup_swapper;
347 
348 	sleepq_lock(ident);
349 	wakeup_swapper = sleepq_broadcast(ident, SLEEPQ_SLEEP, 0, 0);
350 	sleepq_release(ident);
351 	if (wakeup_swapper)
352 		kick_proc0();
353 }
354 
355 /*
356  * Make a thread sleeping on the specified identifier runnable.
357  * May wake more than one thread if a target thread is currently
358  * swapped out.
359  */
360 void
361 wakeup_one(void *ident)
362 {
363 	int wakeup_swapper;
364 
365 	sleepq_lock(ident);
366 	wakeup_swapper = sleepq_signal(ident, SLEEPQ_SLEEP, 0, 0);
367 	sleepq_release(ident);
368 	if (wakeup_swapper)
369 		kick_proc0();
370 }
371 
372 static void
373 kdb_switch(void)
374 {
375 	thread_unlock(curthread);
376 	kdb_backtrace();
377 	kdb_reenter();
378 	panic("%s: did not reenter debugger", __func__);
379 }
380 
381 /*
382  * The machine independent parts of context switching.
383  */
384 void
385 mi_switch(int flags, struct thread *newtd)
386 {
387 	uint64_t runtime, new_switchtime;
388 	struct thread *td;
389 	struct proc *p;
390 
391 	td = curthread;			/* XXX */
392 	THREAD_LOCK_ASSERT(td, MA_OWNED | MA_NOTRECURSED);
393 	p = td->td_proc;		/* XXX */
394 	KASSERT(!TD_ON_RUNQ(td), ("mi_switch: called by old code"));
395 #ifdef INVARIANTS
396 	if (!TD_ON_LOCK(td) && !TD_IS_RUNNING(td))
397 		mtx_assert(&Giant, MA_NOTOWNED);
398 #endif
399 	KASSERT(td->td_critnest == 1 || (td->td_critnest == 2 &&
400 	    (td->td_owepreempt) && (flags & SW_INVOL) != 0 &&
401 	    newtd == NULL) || panicstr,
402 	    ("mi_switch: switch in a critical section"));
403 	KASSERT((flags & (SW_INVOL | SW_VOL)) != 0,
404 	    ("mi_switch: switch must be voluntary or involuntary"));
405 	KASSERT(newtd != curthread, ("mi_switch: preempting back to ourself"));
406 
407 	/*
408 	 * Don't perform context switches from the debugger.
409 	 */
410 	if (kdb_active)
411 		kdb_switch();
412 	if (flags & SW_VOL)
413 		td->td_ru.ru_nvcsw++;
414 	else
415 		td->td_ru.ru_nivcsw++;
416 #ifdef SCHED_STATS
417 	SCHED_STAT_INC(sched_switch_stats[flags & SW_TYPE_MASK]);
418 #endif
419 	/*
420 	 * Compute the amount of time during which the current
421 	 * thread was running, and add that to its total so far.
422 	 */
423 	new_switchtime = cpu_ticks();
424 	runtime = new_switchtime - PCPU_GET(switchtime);
425 	td->td_runtime += runtime;
426 	td->td_incruntime += runtime;
427 	PCPU_SET(switchtime, new_switchtime);
428 	td->td_generation++;	/* bump preempt-detect counter */
429 	PCPU_INC(cnt.v_swtch);
430 	PCPU_SET(switchticks, ticks);
431 	CTR4(KTR_PROC, "mi_switch: old thread %ld (td_sched %p, pid %ld, %s)",
432 	    td->td_tid, td->td_sched, p->p_pid, td->td_name);
433 #if (KTR_COMPILE & KTR_SCHED) != 0
434 	if (TD_IS_IDLETHREAD(td))
435 		KTR_STATE1(KTR_SCHED, "thread", sched_tdname(td), "idle",
436 		    "prio:%d", td->td_priority);
437 	else
438 		KTR_STATE3(KTR_SCHED, "thread", sched_tdname(td), KTDSTATE(td),
439 		    "prio:%d", td->td_priority, "wmesg:\"%s\"", td->td_wmesg,
440 		    "lockname:\"%s\"", td->td_lockname);
441 #endif
442 #ifdef XEN
443 	PT_UPDATES_FLUSH();
444 #endif
445 	sched_switch(td, newtd, flags);
446 	KTR_STATE1(KTR_SCHED, "thread", sched_tdname(td), "running",
447 	    "prio:%d", td->td_priority);
448 
449 	CTR4(KTR_PROC, "mi_switch: new thread %ld (td_sched %p, pid %ld, %s)",
450 	    td->td_tid, td->td_sched, p->p_pid, td->td_name);
451 
452 	/*
453 	 * If the last thread was exiting, finish cleaning it up.
454 	 */
455 	if ((td = PCPU_GET(deadthread))) {
456 		PCPU_SET(deadthread, NULL);
457 		thread_stash(td);
458 	}
459 }
460 
461 /*
462  * Change thread state to be runnable, placing it on the run queue if
463  * it is in memory.  If it is swapped out, return true so our caller
464  * will know to awaken the swapper.
465  */
466 int
467 setrunnable(struct thread *td)
468 {
469 
470 	THREAD_LOCK_ASSERT(td, MA_OWNED);
471 	KASSERT(td->td_proc->p_state != PRS_ZOMBIE,
472 	    ("setrunnable: pid %d is a zombie", td->td_proc->p_pid));
473 	switch (td->td_state) {
474 	case TDS_RUNNING:
475 	case TDS_RUNQ:
476 		return (0);
477 	case TDS_INHIBITED:
478 		/*
479 		 * If we are only inhibited because we are swapped out
480 		 * then arange to swap in this process. Otherwise just return.
481 		 */
482 		if (td->td_inhibitors != TDI_SWAPPED)
483 			return (0);
484 		/* FALLTHROUGH */
485 	case TDS_CAN_RUN:
486 		break;
487 	default:
488 		printf("state is 0x%x", td->td_state);
489 		panic("setrunnable(2)");
490 	}
491 	if ((td->td_flags & TDF_INMEM) == 0) {
492 		if ((td->td_flags & TDF_SWAPINREQ) == 0) {
493 			td->td_flags |= TDF_SWAPINREQ;
494 			return (1);
495 		}
496 	} else
497 		sched_wakeup(td);
498 	return (0);
499 }
500 
501 /*
502  * Compute a tenex style load average of a quantity on
503  * 1, 5 and 15 minute intervals.
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 /* ARGSUSED */
528 static void
529 synch_setup(void *dummy)
530 {
531 	callout_init(&loadav_callout, CALLOUT_MPSAFE);
532 
533 	/* Kick off timeout driven events by calling first time. */
534 	loadav(NULL);
535 }
536 
537 /*
538  * General purpose yield system call.
539  */
540 int
541 yield(struct thread *td, struct yield_args *uap)
542 {
543 
544 	thread_lock(td);
545 	sched_prio(td, PRI_MAX_TIMESHARE);
546 	mi_switch(SW_VOL | SWT_RELINQUISH, NULL);
547 	thread_unlock(td);
548 	td->td_retval[0] = 0;
549 	return (0);
550 }
551