xref: /freebsd/sys/kern/kern_synch.c (revision 195ebc7e9e4b129de810833791a19dfb4349d6a9)
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 
177 	/*
178 	 * If we are already on a sleep queue, then remove us from that
179 	 * sleep queue first.  We have to do this to handle recursive
180 	 * sleeps.
181 	 */
182 	if (TD_ON_SLEEPQ(td))
183 		sleepq_remove(td, td->td_wchan);
184 
185 	if (ident == &pause_wchan)
186 		flags = SLEEPQ_PAUSE;
187 	else
188 		flags = SLEEPQ_SLEEP;
189 	if (catch)
190 		flags |= SLEEPQ_INTERRUPTIBLE;
191 
192 	sleepq_lock(ident);
193 	CTR5(KTR_PROC, "sleep: thread %ld (pid %ld, %s) on %s (%p)",
194 	    td->td_tid, p->p_pid, td->td_name, wmesg, ident);
195 
196 	if (lock == &Giant.lock_object)
197 		mtx_assert(&Giant, MA_OWNED);
198 	DROP_GIANT();
199 	if (lock != NULL && lock != &Giant.lock_object &&
200 	    !(class->lc_flags & LC_SLEEPABLE)) {
201 		WITNESS_SAVE(lock, lock_witness);
202 		lock_state = class->lc_unlock(lock);
203 	} else
204 		/* GCC needs to follow the Yellow Brick Road */
205 		lock_state = -1;
206 
207 	/*
208 	 * We put ourselves on the sleep queue and start our timeout
209 	 * before calling thread_suspend_check, as we could stop there,
210 	 * and a wakeup or a SIGCONT (or both) could occur while we were
211 	 * stopped without resuming us.  Thus, we must be ready for sleep
212 	 * when cursig() is called.  If the wakeup happens while we're
213 	 * stopped, then td will no longer be on a sleep queue upon
214 	 * return from cursig().
215 	 */
216 	sleepq_add(ident, lock, wmesg, flags, 0);
217 	if (timo)
218 		sleepq_set_timeout(ident, timo);
219 	if (lock != NULL && class->lc_flags & LC_SLEEPABLE) {
220 		sleepq_release(ident);
221 		WITNESS_SAVE(lock, lock_witness);
222 		lock_state = class->lc_unlock(lock);
223 		sleepq_lock(ident);
224 	}
225 	if (timo && catch)
226 		rval = sleepq_timedwait_sig(ident, pri);
227 	else if (timo)
228 		rval = sleepq_timedwait(ident, pri);
229 	else if (catch)
230 		rval = sleepq_wait_sig(ident, pri);
231 	else {
232 		sleepq_wait(ident, pri);
233 		rval = 0;
234 	}
235 #ifdef KTRACE
236 	if (KTRPOINT(td, KTR_CSW))
237 		ktrcsw(0, 0);
238 #endif
239 	PICKUP_GIANT();
240 	if (lock != NULL && lock != &Giant.lock_object && !(priority & PDROP)) {
241 		class->lc_lock(lock, lock_state);
242 		WITNESS_RESTORE(lock, lock_witness);
243 	}
244 	return (rval);
245 }
246 
247 int
248 msleep_spin(void *ident, struct mtx *mtx, const char *wmesg, 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, 0);
311 	else {
312 		sleepq_wait(ident, 0);
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(const char *wmesg, int timo)
333 {
334 
335 	KASSERT(timo != 0, ("pause: timeout required"));
336 	return (tsleep(&pause_wchan, 0, wmesg, timo));
337 }
338 
339 /*
340  * Make all threads sleeping on the specified identifier runnable.
341  */
342 void
343 wakeup(void *ident)
344 {
345 	int wakeup_swapper;
346 
347 	sleepq_lock(ident);
348 	wakeup_swapper = sleepq_broadcast(ident, SLEEPQ_SLEEP, 0, 0);
349 	sleepq_release(ident);
350 	if (wakeup_swapper)
351 		kick_proc0();
352 }
353 
354 /*
355  * Make a thread sleeping on the specified identifier runnable.
356  * May wake more than one thread if a target thread is currently
357  * swapped out.
358  */
359 void
360 wakeup_one(void *ident)
361 {
362 	int wakeup_swapper;
363 
364 	sleepq_lock(ident);
365 	wakeup_swapper = sleepq_signal(ident, SLEEPQ_SLEEP, 0, 0);
366 	sleepq_release(ident);
367 	if (wakeup_swapper)
368 		kick_proc0();
369 }
370 
371 static void
372 kdb_switch(void)
373 {
374 	thread_unlock(curthread);
375 	kdb_backtrace();
376 	kdb_reenter();
377 	panic("%s: did not reenter debugger", __func__);
378 }
379 
380 /*
381  * The machine independent parts of context switching.
382  */
383 void
384 mi_switch(int flags, struct thread *newtd)
385 {
386 	uint64_t runtime, new_switchtime;
387 	struct thread *td;
388 	struct proc *p;
389 
390 	td = curthread;			/* XXX */
391 	THREAD_LOCK_ASSERT(td, MA_OWNED | MA_NOTRECURSED);
392 	p = td->td_proc;		/* XXX */
393 	KASSERT(!TD_ON_RUNQ(td), ("mi_switch: called by old code"));
394 #ifdef INVARIANTS
395 	if (!TD_ON_LOCK(td) && !TD_IS_RUNNING(td))
396 		mtx_assert(&Giant, MA_NOTOWNED);
397 #endif
398 	KASSERT(td->td_critnest == 1 || (td->td_critnest == 2 &&
399 	    (td->td_owepreempt) && (flags & SW_INVOL) != 0 &&
400 	    newtd == NULL) || panicstr,
401 	    ("mi_switch: switch in a critical section"));
402 	KASSERT((flags & (SW_INVOL | SW_VOL)) != 0,
403 	    ("mi_switch: switch must be voluntary or involuntary"));
404 	KASSERT(newtd != curthread, ("mi_switch: preempting back to ourself"));
405 
406 	/*
407 	 * Don't perform context switches from the debugger.
408 	 */
409 	if (kdb_active)
410 		kdb_switch();
411 	if (flags & SW_VOL)
412 		td->td_ru.ru_nvcsw++;
413 	else
414 		td->td_ru.ru_nivcsw++;
415 #ifdef SCHED_STATS
416 	SCHED_STAT_INC(sched_switch_stats[flags & SW_TYPE_MASK]);
417 #endif
418 	/*
419 	 * Compute the amount of time during which the current
420 	 * thread was running, and add that to its total so far.
421 	 */
422 	new_switchtime = cpu_ticks();
423 	runtime = new_switchtime - PCPU_GET(switchtime);
424 	td->td_runtime += runtime;
425 	td->td_incruntime += runtime;
426 	PCPU_SET(switchtime, new_switchtime);
427 	td->td_generation++;	/* bump preempt-detect counter */
428 	PCPU_INC(cnt.v_swtch);
429 	PCPU_SET(switchticks, ticks);
430 	CTR4(KTR_PROC, "mi_switch: old thread %ld (td_sched %p, pid %ld, %s)",
431 	    td->td_tid, td->td_sched, p->p_pid, td->td_name);
432 #if (KTR_COMPILE & KTR_SCHED) != 0
433 	if (TD_IS_IDLETHREAD(td))
434 		KTR_STATE1(KTR_SCHED, "thread", sched_tdname(td), "idle",
435 		    "prio:%d", td->td_priority);
436 	else
437 		KTR_STATE3(KTR_SCHED, "thread", sched_tdname(td), KTDSTATE(td),
438 		    "prio:%d", td->td_priority, "wmesg:\"%s\"", td->td_wmesg,
439 		    "lockname:\"%s\"", td->td_lockname);
440 #endif
441 #ifdef XEN
442 	PT_UPDATES_FLUSH();
443 #endif
444 	sched_switch(td, newtd, flags);
445 	KTR_STATE1(KTR_SCHED, "thread", sched_tdname(td), "running",
446 	    "prio:%d", td->td_priority);
447 
448 	CTR4(KTR_PROC, "mi_switch: new thread %ld (td_sched %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 thread state to be runnable, placing it on the run queue if
462  * it is in memory.  If it is swapped out, return true so our caller
463  * will know to awaken the swapper.
464  */
465 int
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 (0);
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 (0);
483 		/* FALLTHROUGH */
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 			return (1);
494 		}
495 	} else
496 		sched_wakeup(td);
497 	return (0);
498 }
499 
500 /*
501  * Compute a tenex style load average of a quantity on
502  * 1, 5 and 15 minute intervals.
503  */
504 static void
505 loadav(void *arg)
506 {
507 	int i, nrun;
508 	struct loadavg *avg;
509 
510 	nrun = sched_load();
511 	avg = &averunnable;
512 
513 	for (i = 0; i < 3; i++)
514 		avg->ldavg[i] = (cexp[i] * avg->ldavg[i] +
515 		    nrun * FSCALE * (FSCALE - cexp[i])) >> FSHIFT;
516 
517 	/*
518 	 * Schedule the next update to occur after 5 seconds, but add a
519 	 * random variation to avoid synchronisation with processes that
520 	 * run at regular intervals.
521 	 */
522 	callout_reset(&loadav_callout, hz * 4 + (int)(random() % (hz * 2 + 1)),
523 	    loadav, NULL);
524 }
525 
526 /* ARGSUSED */
527 static void
528 synch_setup(void *dummy)
529 {
530 	callout_init(&loadav_callout, CALLOUT_MPSAFE);
531 
532 	/* Kick off timeout driven events by calling first time. */
533 	loadav(NULL);
534 }
535 
536 /*
537  * General purpose yield system call.
538  */
539 int
540 yield(struct thread *td, struct yield_args *uap)
541 {
542 
543 	thread_lock(td);
544 	sched_prio(td, PRI_MAX_TIMESHARE);
545 	mi_switch(SW_VOL | SWT_RELINQUISH, NULL);
546 	thread_unlock(td);
547 	td->td_retval[0] = 0;
548 	return (0);
549 }
550