xref: /freebsd/sys/kern/kern_synch.c (revision 6829dae12bb055451fa467da4589c43bd03b1e64)
1 /*-
2  * SPDX-License-Identifier: BSD-3-Clause
3  *
4  * Copyright (c) 1982, 1986, 1990, 1991, 1993
5  *	The Regents of the University of California.  All rights reserved.
6  * (c) UNIX System Laboratories, Inc.
7  * All or some portions of this file are derived from material licensed
8  * to the University of California by American Telephone and Telegraph
9  * Co. or Unix System Laboratories, Inc. and are reproduced herein with
10  * the permission of UNIX System Laboratories, Inc.
11  *
12  * Redistribution and use in source and binary forms, with or without
13  * modification, are permitted provided that the following conditions
14  * are met:
15  * 1. Redistributions of source code must retain the above copyright
16  *    notice, this list of conditions and the following disclaimer.
17  * 2. Redistributions in binary form must reproduce the above copyright
18  *    notice, this list of conditions and the following disclaimer in the
19  *    documentation and/or other materials provided with the distribution.
20  * 3. Neither the name of the University nor the names of its contributors
21  *    may be used to endorse or promote products derived from this software
22  *    without specific prior written permission.
23  *
24  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
25  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
26  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
27  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
28  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
29  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
30  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
31  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
32  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
33  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
34  * SUCH DAMAGE.
35  *
36  *	@(#)kern_synch.c	8.9 (Berkeley) 5/19/95
37  */
38 
39 #include <sys/cdefs.h>
40 __FBSDID("$FreeBSD$");
41 
42 #include "opt_ktrace.h"
43 #include "opt_sched.h"
44 
45 #include <sys/param.h>
46 #include <sys/systm.h>
47 #include <sys/condvar.h>
48 #include <sys/kdb.h>
49 #include <sys/kernel.h>
50 #include <sys/ktr.h>
51 #include <sys/lock.h>
52 #include <sys/mutex.h>
53 #include <sys/proc.h>
54 #include <sys/resourcevar.h>
55 #include <sys/sched.h>
56 #include <sys/sdt.h>
57 #include <sys/signalvar.h>
58 #include <sys/sleepqueue.h>
59 #include <sys/smp.h>
60 #include <sys/sx.h>
61 #include <sys/sysctl.h>
62 #include <sys/sysproto.h>
63 #include <sys/vmmeter.h>
64 #ifdef KTRACE
65 #include <sys/uio.h>
66 #include <sys/ktrace.h>
67 #endif
68 
69 #include <machine/cpu.h>
70 
71 static void synch_setup(void *dummy);
72 SYSINIT(synch_setup, SI_SUB_KICK_SCHEDULER, SI_ORDER_FIRST, synch_setup,
73     NULL);
74 
75 int	hogticks;
76 static uint8_t pause_wchan[MAXCPU];
77 
78 static struct callout loadav_callout;
79 
80 struct loadavg averunnable =
81 	{ {0, 0, 0}, FSCALE };	/* load average, of runnable procs */
82 /*
83  * Constants for averages over 1, 5, and 15 minutes
84  * when sampling at 5 second intervals.
85  */
86 static fixpt_t cexp[3] = {
87 	0.9200444146293232 * FSCALE,	/* exp(-1/12) */
88 	0.9834714538216174 * FSCALE,	/* exp(-1/60) */
89 	0.9944598480048967 * FSCALE,	/* exp(-1/180) */
90 };
91 
92 /* kernel uses `FSCALE', userland (SHOULD) use kern.fscale */
93 SYSCTL_INT(_kern, OID_AUTO, fscale, CTLFLAG_RD, SYSCTL_NULL_INT_PTR, FSCALE, "");
94 
95 static void	loadav(void *arg);
96 
97 SDT_PROVIDER_DECLARE(sched);
98 SDT_PROBE_DEFINE(sched, , , preempt);
99 
100 static void
101 sleepinit(void *unused)
102 {
103 
104 	hogticks = (hz / 10) * 2;	/* Default only. */
105 	init_sleepqueues();
106 }
107 
108 /*
109  * vmem tries to lock the sleepq mutexes when free'ing kva, so make sure
110  * it is available.
111  */
112 SYSINIT(sleepinit, SI_SUB_KMEM, SI_ORDER_ANY, sleepinit, NULL);
113 
114 /*
115  * General sleep call.  Suspends the current thread until a wakeup is
116  * performed on the specified identifier.  The thread will then be made
117  * runnable with the specified priority.  Sleeps at most sbt units of time
118  * (0 means no timeout).  If pri includes the PCATCH flag, let signals
119  * interrupt the sleep, otherwise ignore them while sleeping.  Returns 0 if
120  * awakened, EWOULDBLOCK if the timeout expires.  If PCATCH is set and a
121  * signal becomes pending, ERESTART is returned if the current system
122  * call should be restarted if possible, and EINTR is returned if the system
123  * call should be interrupted by the signal (return EINTR).
124  *
125  * The lock argument is unlocked before the caller is suspended, and
126  * re-locked before _sleep() returns.  If priority includes the PDROP
127  * flag the lock is not re-locked before returning.
128  */
129 int
130 _sleep(void *ident, struct lock_object *lock, int priority,
131     const char *wmesg, sbintime_t sbt, sbintime_t pr, int flags)
132 {
133 	struct thread *td;
134 	struct lock_class *class;
135 	uintptr_t lock_state;
136 	int catch, pri, rval, sleepq_flags;
137 	WITNESS_SAVE_DECL(lock_witness);
138 
139 	td = curthread;
140 #ifdef KTRACE
141 	if (KTRPOINT(td, KTR_CSW))
142 		ktrcsw(1, 0, wmesg);
143 #endif
144 	WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, lock,
145 	    "Sleeping on \"%s\"", wmesg);
146 	KASSERT(sbt != 0 || mtx_owned(&Giant) || lock != NULL,
147 	    ("sleeping without a lock"));
148 	KASSERT(ident != NULL, ("_sleep: NULL ident"));
149 	KASSERT(TD_IS_RUNNING(td), ("_sleep: curthread not running"));
150 	KASSERT(td->td_epochnest == 0, ("sleeping in an epoch section"));
151 	if (priority & PDROP)
152 		KASSERT(lock != NULL && lock != &Giant.lock_object,
153 		    ("PDROP requires a non-Giant lock"));
154 	if (lock != NULL)
155 		class = LOCK_CLASS(lock);
156 	else
157 		class = NULL;
158 
159 	if (SCHEDULER_STOPPED_TD(td)) {
160 		if (lock != NULL && priority & PDROP)
161 			class->lc_unlock(lock);
162 		return (0);
163 	}
164 	catch = priority & PCATCH;
165 	pri = priority & PRIMASK;
166 
167 	KASSERT(!TD_ON_SLEEPQ(td), ("recursive sleep"));
168 
169 	if ((uint8_t *)ident >= &pause_wchan[0] &&
170 	    (uint8_t *)ident <= &pause_wchan[MAXCPU - 1])
171 		sleepq_flags = SLEEPQ_PAUSE;
172 	else
173 		sleepq_flags = SLEEPQ_SLEEP;
174 	if (catch)
175 		sleepq_flags |= SLEEPQ_INTERRUPTIBLE;
176 
177 	sleepq_lock(ident);
178 	CTR5(KTR_PROC, "sleep: thread %ld (pid %ld, %s) on %s (%p)",
179 	    td->td_tid, td->td_proc->p_pid, td->td_name, wmesg, ident);
180 
181 	if (lock == &Giant.lock_object)
182 		mtx_assert(&Giant, MA_OWNED);
183 	DROP_GIANT();
184 	if (lock != NULL && lock != &Giant.lock_object &&
185 	    !(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, lock, wmesg, sleepq_flags, 0);
202 	if (sbt != 0)
203 		sleepq_set_timeout_sbt(ident, sbt, pr, flags);
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 (sbt != 0 && catch)
211 		rval = sleepq_timedwait_sig(ident, pri);
212 	else if (sbt != 0)
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, wmesg);
223 #endif
224 	PICKUP_GIANT();
225 	if (lock != NULL && lock != &Giant.lock_object && !(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_sbt(void *ident, struct mtx *mtx, const char *wmesg,
234     sbintime_t sbt, sbintime_t pr, int flags)
235 {
236 	struct thread *td;
237 	int rval;
238 	WITNESS_SAVE_DECL(mtx);
239 
240 	td = curthread;
241 	KASSERT(mtx != NULL, ("sleeping without a mutex"));
242 	KASSERT(ident != NULL, ("msleep_spin_sbt: NULL ident"));
243 	KASSERT(TD_IS_RUNNING(td), ("msleep_spin_sbt: curthread not running"));
244 
245 	if (SCHEDULER_STOPPED_TD(td))
246 		return (0);
247 
248 	sleepq_lock(ident);
249 	CTR5(KTR_PROC, "msleep_spin: thread %ld (pid %ld, %s) on %s (%p)",
250 	    td->td_tid, td->td_proc->p_pid, td->td_name, wmesg, ident);
251 
252 	DROP_GIANT();
253 	mtx_assert(mtx, MA_OWNED | MA_NOTRECURSED);
254 	WITNESS_SAVE(&mtx->lock_object, mtx);
255 	mtx_unlock_spin(mtx);
256 
257 	/*
258 	 * We put ourselves on the sleep queue and start our timeout.
259 	 */
260 	sleepq_add(ident, &mtx->lock_object, wmesg, SLEEPQ_SLEEP, 0);
261 	if (sbt != 0)
262 		sleepq_set_timeout_sbt(ident, sbt, pr, flags);
263 
264 	/*
265 	 * Can't call ktrace with any spin locks held so it can lock the
266 	 * ktrace_mtx lock, and WITNESS_WARN considers it an error to hold
267 	 * any spin lock.  Thus, we have to drop the sleepq spin lock while
268 	 * we handle those requests.  This is safe since we have placed our
269 	 * thread on the sleep queue already.
270 	 */
271 #ifdef KTRACE
272 	if (KTRPOINT(td, KTR_CSW)) {
273 		sleepq_release(ident);
274 		ktrcsw(1, 0, wmesg);
275 		sleepq_lock(ident);
276 	}
277 #endif
278 #ifdef WITNESS
279 	sleepq_release(ident);
280 	WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL, "Sleeping on \"%s\"",
281 	    wmesg);
282 	sleepq_lock(ident);
283 #endif
284 	if (sbt != 0)
285 		rval = sleepq_timedwait(ident, 0);
286 	else {
287 		sleepq_wait(ident, 0);
288 		rval = 0;
289 	}
290 #ifdef KTRACE
291 	if (KTRPOINT(td, KTR_CSW))
292 		ktrcsw(0, 0, wmesg);
293 #endif
294 	PICKUP_GIANT();
295 	mtx_lock_spin(mtx);
296 	WITNESS_RESTORE(&mtx->lock_object, mtx);
297 	return (rval);
298 }
299 
300 /*
301  * pause_sbt() delays the calling thread by the given signed binary
302  * time. During cold bootup, pause_sbt() uses the DELAY() function
303  * instead of the _sleep() function to do the waiting. The "sbt"
304  * argument must be greater than or equal to zero. A "sbt" value of
305  * zero is equivalent to a "sbt" value of one tick.
306  */
307 int
308 pause_sbt(const char *wmesg, sbintime_t sbt, sbintime_t pr, int flags)
309 {
310 	KASSERT(sbt >= 0, ("pause_sbt: timeout must be >= 0"));
311 
312 	/* silently convert invalid timeouts */
313 	if (sbt == 0)
314 		sbt = tick_sbt;
315 
316 	if ((cold && curthread == &thread0) || kdb_active ||
317 	    SCHEDULER_STOPPED()) {
318 		/*
319 		 * We delay one second at a time to avoid overflowing the
320 		 * system specific DELAY() function(s):
321 		 */
322 		while (sbt >= SBT_1S) {
323 			DELAY(1000000);
324 			sbt -= SBT_1S;
325 		}
326 		/* Do the delay remainder, if any */
327 		sbt = howmany(sbt, SBT_1US);
328 		if (sbt > 0)
329 			DELAY(sbt);
330 		return (EWOULDBLOCK);
331 	}
332 	return (_sleep(&pause_wchan[curcpu], NULL,
333 	    (flags & C_CATCH) ? PCATCH : 0, wmesg, sbt, pr, flags));
334 }
335 
336 /*
337  * Make all threads sleeping on the specified identifier runnable.
338  */
339 void
340 wakeup(void *ident)
341 {
342 	int wakeup_swapper;
343 
344 	sleepq_lock(ident);
345 	wakeup_swapper = sleepq_broadcast(ident, SLEEPQ_SLEEP, 0, 0);
346 	sleepq_release(ident);
347 	if (wakeup_swapper) {
348 		KASSERT(ident != &proc0,
349 		    ("wakeup and wakeup_swapper and proc0"));
350 		kick_proc0();
351 	}
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 
389 	td = curthread;			/* XXX */
390 	THREAD_LOCK_ASSERT(td, MA_OWNED | MA_NOTRECURSED);
391 	KASSERT(!TD_ON_RUNQ(td), ("mi_switch: called by old code"));
392 #ifdef INVARIANTS
393 	if (!TD_ON_LOCK(td) && !TD_IS_RUNNING(td))
394 		mtx_assert(&Giant, MA_NOTOWNED);
395 #endif
396 	KASSERT(td->td_critnest == 1 || panicstr,
397 	    ("mi_switch: switch in a critical section"));
398 	KASSERT((flags & (SW_INVOL | SW_VOL)) != 0,
399 	    ("mi_switch: switch must be voluntary or involuntary"));
400 	KASSERT(newtd != curthread, ("mi_switch: preempting back to ourself"));
401 
402 	/*
403 	 * Don't perform context switches from the debugger.
404 	 */
405 	if (kdb_active)
406 		kdb_switch();
407 	if (SCHEDULER_STOPPED_TD(td))
408 		return;
409 	if (flags & SW_VOL) {
410 		td->td_ru.ru_nvcsw++;
411 		td->td_swvoltick = ticks;
412 	} else {
413 		td->td_ru.ru_nivcsw++;
414 		td->td_swinvoltick = ticks;
415 	}
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 	VM_CNT_INC(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_get_sched(td), td->td_proc->p_pid, td->td_name);
433 #ifdef KDTRACE_HOOKS
434 	if (SDT_PROBES_ENABLED() &&
435 	    ((flags & SW_PREEMPT) != 0 || ((flags & SW_INVOL) != 0 &&
436 	    (flags & SW_TYPE_MASK) == SWT_NEEDRESCHED)))
437 		SDT_PROBE0(sched, , , preempt);
438 #endif
439 	sched_switch(td, newtd, flags);
440 	CTR4(KTR_PROC, "mi_switch: new thread %ld (td_sched %p, pid %ld, %s)",
441 	    td->td_tid, td_get_sched(td), td->td_proc->p_pid, td->td_name);
442 
443 	/*
444 	 * If the last thread was exiting, finish cleaning it up.
445 	 */
446 	if ((td = PCPU_GET(deadthread))) {
447 		PCPU_SET(deadthread, NULL);
448 		thread_stash(td);
449 	}
450 }
451 
452 /*
453  * Change thread state to be runnable, placing it on the run queue if
454  * it is in memory.  If it is swapped out, return true so our caller
455  * will know to awaken the swapper.
456  */
457 int
458 setrunnable(struct thread *td)
459 {
460 
461 	THREAD_LOCK_ASSERT(td, MA_OWNED);
462 	KASSERT(td->td_proc->p_state != PRS_ZOMBIE,
463 	    ("setrunnable: pid %d is a zombie", td->td_proc->p_pid));
464 	switch (td->td_state) {
465 	case TDS_RUNNING:
466 	case TDS_RUNQ:
467 		return (0);
468 	case TDS_INHIBITED:
469 		/*
470 		 * If we are only inhibited because we are swapped out
471 		 * then arange to swap in this process. Otherwise just return.
472 		 */
473 		if (td->td_inhibitors != TDI_SWAPPED)
474 			return (0);
475 		/* FALLTHROUGH */
476 	case TDS_CAN_RUN:
477 		break;
478 	default:
479 		printf("state is 0x%x", td->td_state);
480 		panic("setrunnable(2)");
481 	}
482 	if ((td->td_flags & TDF_INMEM) == 0) {
483 		if ((td->td_flags & TDF_SWAPINREQ) == 0) {
484 			td->td_flags |= TDF_SWAPINREQ;
485 			return (1);
486 		}
487 	} else
488 		sched_wakeup(td);
489 	return (0);
490 }
491 
492 /*
493  * Compute a tenex style load average of a quantity on
494  * 1, 5 and 15 minute intervals.
495  */
496 static void
497 loadav(void *arg)
498 {
499 	int i, nrun;
500 	struct loadavg *avg;
501 
502 	nrun = sched_load();
503 	avg = &averunnable;
504 
505 	for (i = 0; i < 3; i++)
506 		avg->ldavg[i] = (cexp[i] * avg->ldavg[i] +
507 		    nrun * FSCALE * (FSCALE - cexp[i])) >> FSHIFT;
508 
509 	/*
510 	 * Schedule the next update to occur after 5 seconds, but add a
511 	 * random variation to avoid synchronisation with processes that
512 	 * run at regular intervals.
513 	 */
514 	callout_reset_sbt(&loadav_callout,
515 	    SBT_1US * (4000000 + (int)(random() % 2000001)), SBT_1US,
516 	    loadav, NULL, C_DIRECT_EXEC | C_PREL(32));
517 }
518 
519 /* ARGSUSED */
520 static void
521 synch_setup(void *dummy)
522 {
523 	callout_init(&loadav_callout, 1);
524 
525 	/* Kick off timeout driven events by calling first time. */
526 	loadav(NULL);
527 }
528 
529 int
530 should_yield(void)
531 {
532 
533 	return ((u_int)ticks - (u_int)curthread->td_swvoltick >= hogticks);
534 }
535 
536 void
537 maybe_yield(void)
538 {
539 
540 	if (should_yield())
541 		kern_yield(PRI_USER);
542 }
543 
544 void
545 kern_yield(int prio)
546 {
547 	struct thread *td;
548 
549 	td = curthread;
550 	DROP_GIANT();
551 	thread_lock(td);
552 	if (prio == PRI_USER)
553 		prio = td->td_user_pri;
554 	if (prio >= 0)
555 		sched_prio(td, prio);
556 	mi_switch(SW_VOL | SWT_RELINQUISH, NULL);
557 	thread_unlock(td);
558 	PICKUP_GIANT();
559 }
560 
561 /*
562  * General purpose yield system call.
563  */
564 int
565 sys_yield(struct thread *td, struct yield_args *uap)
566 {
567 
568 	thread_lock(td);
569 	if (PRI_BASE(td->td_pri_class) == PRI_TIMESHARE)
570 		sched_prio(td, PRI_MAX_TIMESHARE);
571 	mi_switch(SW_VOL | SWT_RELINQUISH, NULL);
572 	thread_unlock(td);
573 	td->td_retval[0] = 0;
574 	return (0);
575 }
576