xref: /freebsd/sys/kern/kern_synch.c (revision 5bf5ca772c6de2d53344a78cf461447cc322ccea)
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, 0);
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 	if (priority & PDROP)
151 		KASSERT(lock != NULL && lock != &Giant.lock_object,
152 		    ("PDROP requires a non-Giant lock"));
153 	if (lock != NULL)
154 		class = LOCK_CLASS(lock);
155 	else
156 		class = NULL;
157 
158 	if (SCHEDULER_STOPPED_TD(td)) {
159 		if (lock != NULL && priority & PDROP)
160 			class->lc_unlock(lock);
161 		return (0);
162 	}
163 	catch = priority & PCATCH;
164 	pri = priority & PRIMASK;
165 
166 	KASSERT(!TD_ON_SLEEPQ(td), ("recursive sleep"));
167 
168 	if ((uint8_t *)ident >= &pause_wchan[0] &&
169 	    (uint8_t *)ident <= &pause_wchan[MAXCPU - 1])
170 		sleepq_flags = SLEEPQ_PAUSE;
171 	else
172 		sleepq_flags = SLEEPQ_SLEEP;
173 	if (catch)
174 		sleepq_flags |= SLEEPQ_INTERRUPTIBLE;
175 
176 	sleepq_lock(ident);
177 	CTR5(KTR_PROC, "sleep: thread %ld (pid %ld, %s) on %s (%p)",
178 	    td->td_tid, td->td_proc->p_pid, td->td_name, wmesg, ident);
179 
180 	if (lock == &Giant.lock_object)
181 		mtx_assert(&Giant, MA_OWNED);
182 	DROP_GIANT();
183 	if (lock != NULL && lock != &Giant.lock_object &&
184 	    !(class->lc_flags & LC_SLEEPABLE)) {
185 		WITNESS_SAVE(lock, lock_witness);
186 		lock_state = class->lc_unlock(lock);
187 	} else
188 		/* GCC needs to follow the Yellow Brick Road */
189 		lock_state = -1;
190 
191 	/*
192 	 * We put ourselves on the sleep queue and start our timeout
193 	 * before calling thread_suspend_check, as we could stop there,
194 	 * and a wakeup or a SIGCONT (or both) could occur while we were
195 	 * stopped without resuming us.  Thus, we must be ready for sleep
196 	 * when cursig() is called.  If the wakeup happens while we're
197 	 * stopped, then td will no longer be on a sleep queue upon
198 	 * return from cursig().
199 	 */
200 	sleepq_add(ident, lock, wmesg, sleepq_flags, 0);
201 	if (sbt != 0)
202 		sleepq_set_timeout_sbt(ident, sbt, pr, flags);
203 	if (lock != NULL && class->lc_flags & LC_SLEEPABLE) {
204 		sleepq_release(ident);
205 		WITNESS_SAVE(lock, lock_witness);
206 		lock_state = class->lc_unlock(lock);
207 		sleepq_lock(ident);
208 	}
209 	if (sbt != 0 && catch)
210 		rval = sleepq_timedwait_sig(ident, pri);
211 	else if (sbt != 0)
212 		rval = sleepq_timedwait(ident, pri);
213 	else if (catch)
214 		rval = sleepq_wait_sig(ident, pri);
215 	else {
216 		sleepq_wait(ident, pri);
217 		rval = 0;
218 	}
219 #ifdef KTRACE
220 	if (KTRPOINT(td, KTR_CSW))
221 		ktrcsw(0, 0, wmesg);
222 #endif
223 	PICKUP_GIANT();
224 	if (lock != NULL && lock != &Giant.lock_object && !(priority & PDROP)) {
225 		class->lc_lock(lock, lock_state);
226 		WITNESS_RESTORE(lock, lock_witness);
227 	}
228 	return (rval);
229 }
230 
231 int
232 msleep_spin_sbt(void *ident, struct mtx *mtx, const char *wmesg,
233     sbintime_t sbt, sbintime_t pr, int flags)
234 {
235 	struct thread *td;
236 	int rval;
237 	WITNESS_SAVE_DECL(mtx);
238 
239 	td = curthread;
240 	KASSERT(mtx != NULL, ("sleeping without a mutex"));
241 	KASSERT(ident != NULL, ("msleep_spin_sbt: NULL ident"));
242 	KASSERT(TD_IS_RUNNING(td), ("msleep_spin_sbt: curthread not running"));
243 
244 	if (SCHEDULER_STOPPED_TD(td))
245 		return (0);
246 
247 	sleepq_lock(ident);
248 	CTR5(KTR_PROC, "msleep_spin: thread %ld (pid %ld, %s) on %s (%p)",
249 	    td->td_tid, td->td_proc->p_pid, td->td_name, wmesg, ident);
250 
251 	DROP_GIANT();
252 	mtx_assert(mtx, MA_OWNED | MA_NOTRECURSED);
253 	WITNESS_SAVE(&mtx->lock_object, mtx);
254 	mtx_unlock_spin(mtx);
255 
256 	/*
257 	 * We put ourselves on the sleep queue and start our timeout.
258 	 */
259 	sleepq_add(ident, &mtx->lock_object, wmesg, SLEEPQ_SLEEP, 0);
260 	if (sbt != 0)
261 		sleepq_set_timeout_sbt(ident, sbt, pr, flags);
262 
263 	/*
264 	 * Can't call ktrace with any spin locks held so it can lock the
265 	 * ktrace_mtx lock, and WITNESS_WARN considers it an error to hold
266 	 * any spin lock.  Thus, we have to drop the sleepq spin lock while
267 	 * we handle those requests.  This is safe since we have placed our
268 	 * thread on the sleep queue already.
269 	 */
270 #ifdef KTRACE
271 	if (KTRPOINT(td, KTR_CSW)) {
272 		sleepq_release(ident);
273 		ktrcsw(1, 0, wmesg);
274 		sleepq_lock(ident);
275 	}
276 #endif
277 #ifdef WITNESS
278 	sleepq_release(ident);
279 	WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL, "Sleeping on \"%s\"",
280 	    wmesg);
281 	sleepq_lock(ident);
282 #endif
283 	if (sbt != 0)
284 		rval = sleepq_timedwait(ident, 0);
285 	else {
286 		sleepq_wait(ident, 0);
287 		rval = 0;
288 	}
289 #ifdef KTRACE
290 	if (KTRPOINT(td, KTR_CSW))
291 		ktrcsw(0, 0, wmesg);
292 #endif
293 	PICKUP_GIANT();
294 	mtx_lock_spin(mtx);
295 	WITNESS_RESTORE(&mtx->lock_object, mtx);
296 	return (rval);
297 }
298 
299 /*
300  * pause_sbt() delays the calling thread by the given signed binary
301  * time. During cold bootup, pause_sbt() uses the DELAY() function
302  * instead of the _sleep() function to do the waiting. The "sbt"
303  * argument must be greater than or equal to zero. A "sbt" value of
304  * zero is equivalent to a "sbt" value of one tick.
305  */
306 int
307 pause_sbt(const char *wmesg, sbintime_t sbt, sbintime_t pr, int flags)
308 {
309 	KASSERT(sbt >= 0, ("pause_sbt: timeout must be >= 0"));
310 
311 	/* silently convert invalid timeouts */
312 	if (sbt == 0)
313 		sbt = tick_sbt;
314 
315 	if ((cold && curthread == &thread0) || kdb_active ||
316 	    SCHEDULER_STOPPED()) {
317 		/*
318 		 * We delay one second at a time to avoid overflowing the
319 		 * system specific DELAY() function(s):
320 		 */
321 		while (sbt >= SBT_1S) {
322 			DELAY(1000000);
323 			sbt -= SBT_1S;
324 		}
325 		/* Do the delay remainder, if any */
326 		sbt = howmany(sbt, SBT_1US);
327 		if (sbt > 0)
328 			DELAY(sbt);
329 		return (EWOULDBLOCK);
330 	}
331 	return (_sleep(&pause_wchan[curcpu], NULL,
332 	    (flags & C_CATCH) ? PCATCH : 0, wmesg, sbt, pr, flags));
333 }
334 
335 /*
336  * Make all threads sleeping on the specified identifier runnable.
337  */
338 void
339 wakeup(void *ident)
340 {
341 	int wakeup_swapper;
342 
343 	sleepq_lock(ident);
344 	wakeup_swapper = sleepq_broadcast(ident, SLEEPQ_SLEEP, 0, 0);
345 	sleepq_release(ident);
346 	if (wakeup_swapper) {
347 		KASSERT(ident != &proc0,
348 		    ("wakeup and wakeup_swapper and proc0"));
349 		kick_proc0();
350 	}
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(void *ident)
360 {
361 	int wakeup_swapper;
362 
363 	sleepq_lock(ident);
364 	wakeup_swapper = sleepq_signal(ident, SLEEPQ_SLEEP, 0, 0);
365 	sleepq_release(ident);
366 	if (wakeup_swapper)
367 		kick_proc0();
368 }
369 
370 static void
371 kdb_switch(void)
372 {
373 	thread_unlock(curthread);
374 	kdb_backtrace();
375 	kdb_reenter();
376 	panic("%s: did not reenter debugger", __func__);
377 }
378 
379 /*
380  * The machine independent parts of context switching.
381  */
382 void
383 mi_switch(int flags, struct thread *newtd)
384 {
385 	uint64_t runtime, new_switchtime;
386 	struct thread *td;
387 
388 	td = curthread;			/* XXX */
389 	THREAD_LOCK_ASSERT(td, MA_OWNED | MA_NOTRECURSED);
390 	KASSERT(!TD_ON_RUNQ(td), ("mi_switch: called by old code"));
391 #ifdef INVARIANTS
392 	if (!TD_ON_LOCK(td) && !TD_IS_RUNNING(td))
393 		mtx_assert(&Giant, MA_NOTOWNED);
394 #endif
395 	KASSERT(td->td_critnest == 1 || panicstr,
396 	    ("mi_switch: switch in a critical section"));
397 	KASSERT((flags & (SW_INVOL | SW_VOL)) != 0,
398 	    ("mi_switch: switch must be voluntary or involuntary"));
399 	KASSERT(newtd != curthread, ("mi_switch: preempting back to ourself"));
400 
401 	/*
402 	 * Don't perform context switches from the debugger.
403 	 */
404 	if (kdb_active)
405 		kdb_switch();
406 	if (SCHEDULER_STOPPED_TD(td))
407 		return;
408 	if (flags & SW_VOL) {
409 		td->td_ru.ru_nvcsw++;
410 		td->td_swvoltick = ticks;
411 	} else {
412 		td->td_ru.ru_nivcsw++;
413 		td->td_swinvoltick = ticks;
414 	}
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 	VM_CNT_INC(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_get_sched(td), td->td_proc->p_pid, td->td_name);
432 #ifdef KDTRACE_HOOKS
433 	if ((flags & SW_PREEMPT) != 0 || ((flags & SW_INVOL) != 0 &&
434 	    (flags & SW_TYPE_MASK) == SWT_NEEDRESCHED))
435 		SDT_PROBE0(sched, , , preempt);
436 #endif
437 	sched_switch(td, newtd, flags);
438 	CTR4(KTR_PROC, "mi_switch: new thread %ld (td_sched %p, pid %ld, %s)",
439 	    td->td_tid, td_get_sched(td), td->td_proc->p_pid, td->td_name);
440 
441 	/*
442 	 * If the last thread was exiting, finish cleaning it up.
443 	 */
444 	if ((td = PCPU_GET(deadthread))) {
445 		PCPU_SET(deadthread, NULL);
446 		thread_stash(td);
447 	}
448 }
449 
450 /*
451  * Change thread state to be runnable, placing it on the run queue if
452  * it is in memory.  If it is swapped out, return true so our caller
453  * will know to awaken the swapper.
454  */
455 int
456 setrunnable(struct thread *td)
457 {
458 
459 	THREAD_LOCK_ASSERT(td, MA_OWNED);
460 	KASSERT(td->td_proc->p_state != PRS_ZOMBIE,
461 	    ("setrunnable: pid %d is a zombie", td->td_proc->p_pid));
462 	switch (td->td_state) {
463 	case TDS_RUNNING:
464 	case TDS_RUNQ:
465 		return (0);
466 	case TDS_INHIBITED:
467 		/*
468 		 * If we are only inhibited because we are swapped out
469 		 * then arange to swap in this process. Otherwise just return.
470 		 */
471 		if (td->td_inhibitors != TDI_SWAPPED)
472 			return (0);
473 		/* FALLTHROUGH */
474 	case TDS_CAN_RUN:
475 		break;
476 	default:
477 		printf("state is 0x%x", td->td_state);
478 		panic("setrunnable(2)");
479 	}
480 	if ((td->td_flags & TDF_INMEM) == 0) {
481 		if ((td->td_flags & TDF_SWAPINREQ) == 0) {
482 			td->td_flags |= TDF_SWAPINREQ;
483 			return (1);
484 		}
485 	} else
486 		sched_wakeup(td);
487 	return (0);
488 }
489 
490 /*
491  * Compute a tenex style load average of a quantity on
492  * 1, 5 and 15 minute intervals.
493  */
494 static void
495 loadav(void *arg)
496 {
497 	int i, nrun;
498 	struct loadavg *avg;
499 
500 	nrun = sched_load();
501 	avg = &averunnable;
502 
503 	for (i = 0; i < 3; i++)
504 		avg->ldavg[i] = (cexp[i] * avg->ldavg[i] +
505 		    nrun * FSCALE * (FSCALE - cexp[i])) >> FSHIFT;
506 
507 	/*
508 	 * Schedule the next update to occur after 5 seconds, but add a
509 	 * random variation to avoid synchronisation with processes that
510 	 * run at regular intervals.
511 	 */
512 	callout_reset_sbt(&loadav_callout,
513 	    SBT_1US * (4000000 + (int)(random() % 2000001)), SBT_1US,
514 	    loadav, NULL, C_DIRECT_EXEC | C_PREL(32));
515 }
516 
517 /* ARGSUSED */
518 static void
519 synch_setup(void *dummy)
520 {
521 	callout_init(&loadav_callout, 1);
522 
523 	/* Kick off timeout driven events by calling first time. */
524 	loadav(NULL);
525 }
526 
527 int
528 should_yield(void)
529 {
530 
531 	return ((u_int)ticks - (u_int)curthread->td_swvoltick >= hogticks);
532 }
533 
534 void
535 maybe_yield(void)
536 {
537 
538 	if (should_yield())
539 		kern_yield(PRI_USER);
540 }
541 
542 void
543 kern_yield(int prio)
544 {
545 	struct thread *td;
546 
547 	td = curthread;
548 	DROP_GIANT();
549 	thread_lock(td);
550 	if (prio == PRI_USER)
551 		prio = td->td_user_pri;
552 	if (prio >= 0)
553 		sched_prio(td, prio);
554 	mi_switch(SW_VOL | SWT_RELINQUISH, NULL);
555 	thread_unlock(td);
556 	PICKUP_GIANT();
557 }
558 
559 /*
560  * General purpose yield system call.
561  */
562 int
563 sys_yield(struct thread *td, struct yield_args *uap)
564 {
565 
566 	thread_lock(td);
567 	if (PRI_BASE(td->td_pri_class) == PRI_TIMESHARE)
568 		sched_prio(td, PRI_MAX_TIMESHARE);
569 	mi_switch(SW_VOL | SWT_RELINQUISH, NULL);
570 	thread_unlock(td);
571 	td->td_retval[0] = 0;
572 	return (0);
573 }
574