xref: /freebsd/sys/kern/subr_sleepqueue.c (revision 3c6e15bceeab4470243c60c9a4b5b9cafca9abaa)
1 /*-
2  * Copyright (c) 2004 John Baldwin <jhb@FreeBSD.org>
3  * All rights reserved.
4  *
5  * Redistribution and use in source and binary forms, with or without
6  * modification, are permitted provided that the following conditions
7  * are met:
8  * 1. Redistributions of source code must retain the above copyright
9  *    notice, this list of conditions and the following disclaimer.
10  * 2. Redistributions in binary form must reproduce the above copyright
11  *    notice, this list of conditions and the following disclaimer in the
12  *    documentation and/or other materials provided with the distribution.
13  * 3. Neither the name of the author nor the names of any co-contributors
14  *    may be used to endorse or promote products derived from this software
15  *    without specific prior written permission.
16  *
17  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
18  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
19  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
20  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
21  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
22  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
23  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
24  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
25  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
26  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
27  * SUCH DAMAGE.
28  */
29 
30 /*
31  * Implementation of sleep queues used to hold queue of threads blocked on
32  * a wait channel.  Sleep queues different from turnstiles in that wait
33  * channels are not owned by anyone, so there is no priority propagation.
34  * Sleep queues can also provide a timeout and can also be interrupted by
35  * signals.  That said, there are several similarities between the turnstile
36  * and sleep queue implementations.  (Note: turnstiles were implemented
37  * first.)  For example, both use a hash table of the same size where each
38  * bucket is referred to as a "chain" that contains both a spin lock and
39  * a linked list of queues.  An individual queue is located by using a hash
40  * to pick a chain, locking the chain, and then walking the chain searching
41  * for the queue.  This means that a wait channel object does not need to
42  * embed it's queue head just as locks do not embed their turnstile queue
43  * head.  Threads also carry around a sleep queue that they lend to the
44  * wait channel when blocking.  Just as in turnstiles, the queue includes
45  * a free list of the sleep queues of other threads blocked on the same
46  * wait channel in the case of multiple waiters.
47  *
48  * Some additional functionality provided by sleep queues include the
49  * ability to set a timeout.  The timeout is managed using a per-thread
50  * callout that resumes a thread if it is asleep.  A thread may also
51  * catch signals while it is asleep (aka an interruptible sleep).  The
52  * signal code uses sleepq_abort() to interrupt a sleeping thread.  Finally,
53  * sleep queues also provide some extra assertions.  One is not allowed to
54  * mix the sleep/wakeup and cv APIs for a given wait channel.  Also, one
55  * must consistently use the same lock to synchronize with a wait channel,
56  * though this check is currently only a warning for sleep/wakeup due to
57  * pre-existing abuse of that API.  The same lock must also be held when
58  * awakening threads, though that is currently only enforced for condition
59  * variables.
60  */
61 
62 #include <sys/cdefs.h>
63 __FBSDID("$FreeBSD$");
64 
65 #include "opt_sleepqueue_profiling.h"
66 #include "opt_ddb.h"
67 #include "opt_sched.h"
68 
69 #include <sys/param.h>
70 #include <sys/systm.h>
71 #include <sys/lock.h>
72 #include <sys/kernel.h>
73 #include <sys/ktr.h>
74 #include <sys/mutex.h>
75 #include <sys/proc.h>
76 #include <sys/sbuf.h>
77 #include <sys/sched.h>
78 #include <sys/signalvar.h>
79 #include <sys/sleepqueue.h>
80 #include <sys/sysctl.h>
81 
82 #include <vm/uma.h>
83 
84 #ifdef DDB
85 #include <ddb/ddb.h>
86 #endif
87 
88 /*
89  * Constants for the hash table of sleep queue chains.  These constants are
90  * the same ones that 4BSD (and possibly earlier versions of BSD) used.
91  * Basically, we ignore the lower 8 bits of the address since most wait
92  * channel pointers are aligned and only look at the next 7 bits for the
93  * hash.  SC_TABLESIZE must be a power of two for SC_MASK to work properly.
94  */
95 #define	SC_TABLESIZE	128			/* Must be power of 2. */
96 #define	SC_MASK		(SC_TABLESIZE - 1)
97 #define	SC_SHIFT	8
98 #define	SC_HASH(wc)	(((uintptr_t)(wc) >> SC_SHIFT) & SC_MASK)
99 #define	SC_LOOKUP(wc)	&sleepq_chains[SC_HASH(wc)]
100 #define NR_SLEEPQS      2
101 /*
102  * There two different lists of sleep queues.  Both lists are connected
103  * via the sq_hash entries.  The first list is the sleep queue chain list
104  * that a sleep queue is on when it is attached to a wait channel.  The
105  * second list is the free list hung off of a sleep queue that is attached
106  * to a wait channel.
107  *
108  * Each sleep queue also contains the wait channel it is attached to, the
109  * list of threads blocked on that wait channel, flags specific to the
110  * wait channel, and the lock used to synchronize with a wait channel.
111  * The flags are used to catch mismatches between the various consumers
112  * of the sleep queue API (e.g. sleep/wakeup and condition variables).
113  * The lock pointer is only used when invariants are enabled for various
114  * debugging checks.
115  *
116  * Locking key:
117  *  c - sleep queue chain lock
118  */
119 struct sleepqueue {
120 	TAILQ_HEAD(, thread) sq_blocked[NR_SLEEPQS];	/* (c) Blocked threads. */
121 	LIST_ENTRY(sleepqueue) sq_hash;		/* (c) Chain and free list. */
122 	LIST_HEAD(, sleepqueue) sq_free;	/* (c) Free queues. */
123 	void	*sq_wchan;			/* (c) Wait channel. */
124 #ifdef INVARIANTS
125 	int	sq_type;			/* (c) Queue type. */
126 	struct lock_object *sq_lock;		/* (c) Associated lock. */
127 #endif
128 };
129 
130 struct sleepqueue_chain {
131 	LIST_HEAD(, sleepqueue) sc_queues;	/* List of sleep queues. */
132 	struct mtx sc_lock;			/* Spin lock for this chain. */
133 #ifdef SLEEPQUEUE_PROFILING
134 	u_int	sc_depth;			/* Length of sc_queues. */
135 	u_int	sc_max_depth;			/* Max length of sc_queues. */
136 #endif
137 };
138 
139 #ifdef SLEEPQUEUE_PROFILING
140 u_int sleepq_max_depth;
141 SYSCTL_NODE(_debug, OID_AUTO, sleepq, CTLFLAG_RD, 0, "sleepq profiling");
142 SYSCTL_NODE(_debug_sleepq, OID_AUTO, chains, CTLFLAG_RD, 0,
143     "sleepq chain stats");
144 SYSCTL_UINT(_debug_sleepq, OID_AUTO, max_depth, CTLFLAG_RD, &sleepq_max_depth,
145     0, "maxmimum depth achieved of a single chain");
146 
147 static void	sleepq_profile(const char *wmesg);
148 static int	prof_enabled;
149 #endif
150 static struct sleepqueue_chain sleepq_chains[SC_TABLESIZE];
151 static uma_zone_t sleepq_zone;
152 
153 /*
154  * Prototypes for non-exported routines.
155  */
156 static int	sleepq_catch_signals(void *wchan, int pri);
157 static int	sleepq_check_signals(void);
158 static int	sleepq_check_timeout(void);
159 #ifdef INVARIANTS
160 static void	sleepq_dtor(void *mem, int size, void *arg);
161 #endif
162 static int	sleepq_init(void *mem, int size, int flags);
163 static int	sleepq_resume_thread(struct sleepqueue *sq, struct thread *td,
164 		    int pri);
165 static void	sleepq_switch(void *wchan, int pri);
166 static void	sleepq_timeout(void *arg);
167 
168 /*
169  * Early initialization of sleep queues that is called from the sleepinit()
170  * SYSINIT.
171  */
172 void
173 init_sleepqueues(void)
174 {
175 #ifdef SLEEPQUEUE_PROFILING
176 	struct sysctl_oid *chain_oid;
177 	char chain_name[10];
178 #endif
179 	int i;
180 
181 	for (i = 0; i < SC_TABLESIZE; i++) {
182 		LIST_INIT(&sleepq_chains[i].sc_queues);
183 		mtx_init(&sleepq_chains[i].sc_lock, "sleepq chain", NULL,
184 		    MTX_SPIN | MTX_RECURSE);
185 #ifdef SLEEPQUEUE_PROFILING
186 		snprintf(chain_name, sizeof(chain_name), "%d", i);
187 		chain_oid = SYSCTL_ADD_NODE(NULL,
188 		    SYSCTL_STATIC_CHILDREN(_debug_sleepq_chains), OID_AUTO,
189 		    chain_name, CTLFLAG_RD, NULL, "sleepq chain stats");
190 		SYSCTL_ADD_UINT(NULL, SYSCTL_CHILDREN(chain_oid), OID_AUTO,
191 		    "depth", CTLFLAG_RD, &sleepq_chains[i].sc_depth, 0, NULL);
192 		SYSCTL_ADD_UINT(NULL, SYSCTL_CHILDREN(chain_oid), OID_AUTO,
193 		    "max_depth", CTLFLAG_RD, &sleepq_chains[i].sc_max_depth, 0,
194 		    NULL);
195 #endif
196 	}
197 	sleepq_zone = uma_zcreate("SLEEPQUEUE", sizeof(struct sleepqueue),
198 #ifdef INVARIANTS
199 	    NULL, sleepq_dtor, sleepq_init, NULL, UMA_ALIGN_CACHE, 0);
200 #else
201 	    NULL, NULL, sleepq_init, NULL, UMA_ALIGN_CACHE, 0);
202 #endif
203 
204 	thread0.td_sleepqueue = sleepq_alloc();
205 }
206 
207 /*
208  * Get a sleep queue for a new thread.
209  */
210 struct sleepqueue *
211 sleepq_alloc(void)
212 {
213 
214 	return (uma_zalloc(sleepq_zone, M_WAITOK));
215 }
216 
217 /*
218  * Free a sleep queue when a thread is destroyed.
219  */
220 void
221 sleepq_free(struct sleepqueue *sq)
222 {
223 
224 	uma_zfree(sleepq_zone, sq);
225 }
226 
227 /*
228  * Lock the sleep queue chain associated with the specified wait channel.
229  */
230 void
231 sleepq_lock(void *wchan)
232 {
233 	struct sleepqueue_chain *sc;
234 
235 	sc = SC_LOOKUP(wchan);
236 	mtx_lock_spin(&sc->sc_lock);
237 }
238 
239 /*
240  * Look up the sleep queue associated with a given wait channel in the hash
241  * table locking the associated sleep queue chain.  If no queue is found in
242  * the table, NULL is returned.
243  */
244 struct sleepqueue *
245 sleepq_lookup(void *wchan)
246 {
247 	struct sleepqueue_chain *sc;
248 	struct sleepqueue *sq;
249 
250 	KASSERT(wchan != NULL, ("%s: invalid NULL wait channel", __func__));
251 	sc = SC_LOOKUP(wchan);
252 	mtx_assert(&sc->sc_lock, MA_OWNED);
253 	LIST_FOREACH(sq, &sc->sc_queues, sq_hash)
254 		if (sq->sq_wchan == wchan)
255 			return (sq);
256 	return (NULL);
257 }
258 
259 /*
260  * Unlock the sleep queue chain associated with a given wait channel.
261  */
262 void
263 sleepq_release(void *wchan)
264 {
265 	struct sleepqueue_chain *sc;
266 
267 	sc = SC_LOOKUP(wchan);
268 	mtx_unlock_spin(&sc->sc_lock);
269 }
270 
271 /*
272  * Places the current thread on the sleep queue for the specified wait
273  * channel.  If INVARIANTS is enabled, then it associates the passed in
274  * lock with the sleepq to make sure it is held when that sleep queue is
275  * woken up.
276  */
277 void
278 sleepq_add(void *wchan, struct lock_object *lock, const char *wmesg, int flags,
279     int queue)
280 {
281 	struct sleepqueue_chain *sc;
282 	struct sleepqueue *sq;
283 	struct thread *td;
284 
285 	td = curthread;
286 	sc = SC_LOOKUP(wchan);
287 	mtx_assert(&sc->sc_lock, MA_OWNED);
288 	MPASS(td->td_sleepqueue != NULL);
289 	MPASS(wchan != NULL);
290 	MPASS((queue >= 0) && (queue < NR_SLEEPQS));
291 
292 	/* If this thread is not allowed to sleep, die a horrible death. */
293 	KASSERT(!(td->td_pflags & TDP_NOSLEEPING),
294 	    ("Trying sleep, but thread marked as sleeping prohibited"));
295 
296 	/* Look up the sleep queue associated with the wait channel 'wchan'. */
297 	sq = sleepq_lookup(wchan);
298 
299 	/*
300 	 * If the wait channel does not already have a sleep queue, use
301 	 * this thread's sleep queue.  Otherwise, insert the current thread
302 	 * into the sleep queue already in use by this wait channel.
303 	 */
304 	if (sq == NULL) {
305 #ifdef INVARIANTS
306 		int i;
307 
308 		sq = td->td_sleepqueue;
309 		for (i = 0; i < NR_SLEEPQS; i++)
310 			KASSERT(TAILQ_EMPTY(&sq->sq_blocked[i]),
311 				("thread's sleep queue %d is not empty", i));
312 		KASSERT(LIST_EMPTY(&sq->sq_free),
313 		    ("thread's sleep queue has a non-empty free list"));
314 		KASSERT(sq->sq_wchan == NULL, ("stale sq_wchan pointer"));
315 		sq->sq_lock = lock;
316 		sq->sq_type = flags & SLEEPQ_TYPE;
317 #endif
318 #ifdef SLEEPQUEUE_PROFILING
319 		sc->sc_depth++;
320 		if (sc->sc_depth > sc->sc_max_depth) {
321 			sc->sc_max_depth = sc->sc_depth;
322 			if (sc->sc_max_depth > sleepq_max_depth)
323 				sleepq_max_depth = sc->sc_max_depth;
324 		}
325 #endif
326 		sq = td->td_sleepqueue;
327 		LIST_INSERT_HEAD(&sc->sc_queues, sq, sq_hash);
328 		sq->sq_wchan = wchan;
329 	} else {
330 		MPASS(wchan == sq->sq_wchan);
331 		MPASS(lock == sq->sq_lock);
332 		MPASS((flags & SLEEPQ_TYPE) == sq->sq_type);
333 		LIST_INSERT_HEAD(&sq->sq_free, td->td_sleepqueue, sq_hash);
334 	}
335 	thread_lock(td);
336 	TAILQ_INSERT_TAIL(&sq->sq_blocked[queue], td, td_slpq);
337 	td->td_sleepqueue = NULL;
338 	td->td_sqqueue = queue;
339 	td->td_wchan = wchan;
340 	td->td_wmesg = wmesg;
341 	if (flags & SLEEPQ_INTERRUPTIBLE) {
342 		td->td_flags |= TDF_SINTR;
343 		td->td_flags &= ~TDF_SLEEPABORT;
344 	}
345 	thread_unlock(td);
346 }
347 
348 /*
349  * Sets a timeout that will remove the current thread from the specified
350  * sleep queue after timo ticks if the thread has not already been awakened.
351  */
352 void
353 sleepq_set_timeout(void *wchan, int timo)
354 {
355 	struct sleepqueue_chain *sc;
356 	struct thread *td;
357 
358 	td = curthread;
359 	sc = SC_LOOKUP(wchan);
360 	mtx_assert(&sc->sc_lock, MA_OWNED);
361 	MPASS(TD_ON_SLEEPQ(td));
362 	MPASS(td->td_sleepqueue == NULL);
363 	MPASS(wchan != NULL);
364 	callout_reset_curcpu(&td->td_slpcallout, timo, sleepq_timeout, td);
365 }
366 
367 /*
368  * Marks the pending sleep of the current thread as interruptible and
369  * makes an initial check for pending signals before putting a thread
370  * to sleep. Enters and exits with the thread lock held.  Thread lock
371  * may have transitioned from the sleepq lock to a run lock.
372  */
373 static int
374 sleepq_catch_signals(void *wchan, int pri)
375 {
376 	struct sleepqueue_chain *sc;
377 	struct sleepqueue *sq;
378 	struct thread *td;
379 	struct proc *p;
380 	struct sigacts *ps;
381 	int sig, ret;
382 
383 	td = curthread;
384 	p = curproc;
385 	sc = SC_LOOKUP(wchan);
386 	mtx_assert(&sc->sc_lock, MA_OWNED);
387 	MPASS(wchan != NULL);
388 	/*
389 	 * See if there are any pending signals for this thread.  If not
390 	 * we can switch immediately.  Otherwise do the signal processing
391 	 * directly.
392 	 */
393 	thread_lock(td);
394 	if ((td->td_flags & (TDF_NEEDSIGCHK | TDF_NEEDSUSPCHK)) == 0) {
395 		sleepq_switch(wchan, pri);
396 		return (0);
397 	}
398 	thread_unlock(td);
399 	mtx_unlock_spin(&sc->sc_lock);
400 	CTR3(KTR_PROC, "sleepq catching signals: thread %p (pid %ld, %s)",
401 		(void *)td, (long)p->p_pid, td->td_name);
402 	PROC_LOCK(p);
403 	ps = p->p_sigacts;
404 	mtx_lock(&ps->ps_mtx);
405 	sig = cursig(td);
406 	if (sig == 0) {
407 		mtx_unlock(&ps->ps_mtx);
408 		ret = thread_suspend_check(1);
409 		MPASS(ret == 0 || ret == EINTR || ret == ERESTART);
410 	} else {
411 		if (SIGISMEMBER(ps->ps_sigintr, sig))
412 			ret = EINTR;
413 		else
414 			ret = ERESTART;
415 		mtx_unlock(&ps->ps_mtx);
416 	}
417 	/*
418 	 * Lock the per-process spinlock prior to dropping the PROC_LOCK
419 	 * to avoid a signal delivery race.  PROC_LOCK, PROC_SLOCK, and
420 	 * thread_lock() are currently held in tdsignal().
421 	 */
422 	PROC_SLOCK(p);
423 	mtx_lock_spin(&sc->sc_lock);
424 	PROC_UNLOCK(p);
425 	thread_lock(td);
426 	PROC_SUNLOCK(p);
427 	/*
428 	 * There were pending signals and this thread is still
429 	 * on the sleep queue, remove it from the sleep queue.
430 	 */
431 	if (TD_ON_SLEEPQ(td)) {
432 		sq = sleepq_lookup(wchan);
433 		if (sleepq_resume_thread(sq, td, 0)) {
434 #ifdef INVARIANTS
435 			/*
436 			 * This thread hasn't gone to sleep yet, so it
437 			 * should not be swapped out.
438 			 */
439 			panic("not waking up swapper");
440 #endif
441 		}
442 	}
443 	mtx_unlock_spin(&sc->sc_lock);
444 	MPASS(td->td_lock != &sc->sc_lock);
445 	return (ret);
446 }
447 
448 /*
449  * Switches to another thread if we are still asleep on a sleep queue.
450  * Returns with thread lock.
451  */
452 static void
453 sleepq_switch(void *wchan, int pri)
454 {
455 	struct sleepqueue_chain *sc;
456 	struct sleepqueue *sq;
457 	struct thread *td;
458 
459 	td = curthread;
460 	sc = SC_LOOKUP(wchan);
461 	mtx_assert(&sc->sc_lock, MA_OWNED);
462 	THREAD_LOCK_ASSERT(td, MA_OWNED);
463 
464 	/*
465 	 * If we have a sleep queue, then we've already been woken up, so
466 	 * just return.
467 	 */
468 	if (td->td_sleepqueue != NULL) {
469 		mtx_unlock_spin(&sc->sc_lock);
470 		return;
471 	}
472 
473 	/*
474 	 * If TDF_TIMEOUT is set, then our sleep has been timed out
475 	 * already but we are still on the sleep queue, so dequeue the
476 	 * thread and return.
477 	 */
478 	if (td->td_flags & TDF_TIMEOUT) {
479 		MPASS(TD_ON_SLEEPQ(td));
480 		sq = sleepq_lookup(wchan);
481 		if (sleepq_resume_thread(sq, td, 0)) {
482 #ifdef INVARIANTS
483 			/*
484 			 * This thread hasn't gone to sleep yet, so it
485 			 * should not be swapped out.
486 			 */
487 			panic("not waking up swapper");
488 #endif
489 		}
490 		mtx_unlock_spin(&sc->sc_lock);
491 		return;
492 	}
493 #ifdef SLEEPQUEUE_PROFILING
494 	if (prof_enabled)
495 		sleepq_profile(td->td_wmesg);
496 #endif
497 	MPASS(td->td_sleepqueue == NULL);
498 	sched_sleep(td, pri);
499 	thread_lock_set(td, &sc->sc_lock);
500 	TD_SET_SLEEPING(td);
501 	mi_switch(SW_VOL | SWT_SLEEPQ, NULL);
502 	KASSERT(TD_IS_RUNNING(td), ("running but not TDS_RUNNING"));
503 	CTR3(KTR_PROC, "sleepq resume: thread %p (pid %ld, %s)",
504 	    (void *)td, (long)td->td_proc->p_pid, (void *)td->td_name);
505 }
506 
507 /*
508  * Check to see if we timed out.
509  */
510 static int
511 sleepq_check_timeout(void)
512 {
513 	struct thread *td;
514 
515 	td = curthread;
516 	THREAD_LOCK_ASSERT(td, MA_OWNED);
517 
518 	/*
519 	 * If TDF_TIMEOUT is set, we timed out.
520 	 */
521 	if (td->td_flags & TDF_TIMEOUT) {
522 		td->td_flags &= ~TDF_TIMEOUT;
523 		return (EWOULDBLOCK);
524 	}
525 
526 	/*
527 	 * If TDF_TIMOFAIL is set, the timeout ran after we had
528 	 * already been woken up.
529 	 */
530 	if (td->td_flags & TDF_TIMOFAIL)
531 		td->td_flags &= ~TDF_TIMOFAIL;
532 
533 	/*
534 	 * If callout_stop() fails, then the timeout is running on
535 	 * another CPU, so synchronize with it to avoid having it
536 	 * accidentally wake up a subsequent sleep.
537 	 */
538 	else if (callout_stop(&td->td_slpcallout) == 0) {
539 		td->td_flags |= TDF_TIMEOUT;
540 		TD_SET_SLEEPING(td);
541 		mi_switch(SW_INVOL | SWT_SLEEPQTIMO, NULL);
542 	}
543 	return (0);
544 }
545 
546 /*
547  * Check to see if we were awoken by a signal.
548  */
549 static int
550 sleepq_check_signals(void)
551 {
552 	struct thread *td;
553 
554 	td = curthread;
555 	THREAD_LOCK_ASSERT(td, MA_OWNED);
556 
557 	/* We are no longer in an interruptible sleep. */
558 	if (td->td_flags & TDF_SINTR)
559 		td->td_flags &= ~TDF_SINTR;
560 
561 	if (td->td_flags & TDF_SLEEPABORT) {
562 		td->td_flags &= ~TDF_SLEEPABORT;
563 		return (td->td_intrval);
564 	}
565 
566 	return (0);
567 }
568 
569 /*
570  * Block the current thread until it is awakened from its sleep queue.
571  */
572 void
573 sleepq_wait(void *wchan, int pri)
574 {
575 	struct thread *td;
576 
577 	td = curthread;
578 	MPASS(!(td->td_flags & TDF_SINTR));
579 	thread_lock(td);
580 	sleepq_switch(wchan, pri);
581 	thread_unlock(td);
582 }
583 
584 /*
585  * Block the current thread until it is awakened from its sleep queue
586  * or it is interrupted by a signal.
587  */
588 int
589 sleepq_wait_sig(void *wchan, int pri)
590 {
591 	int rcatch;
592 	int rval;
593 
594 	rcatch = sleepq_catch_signals(wchan, pri);
595 	rval = sleepq_check_signals();
596 	thread_unlock(curthread);
597 	if (rcatch)
598 		return (rcatch);
599 	return (rval);
600 }
601 
602 /*
603  * Block the current thread until it is awakened from its sleep queue
604  * or it times out while waiting.
605  */
606 int
607 sleepq_timedwait(void *wchan, int pri)
608 {
609 	struct thread *td;
610 	int rval;
611 
612 	td = curthread;
613 	MPASS(!(td->td_flags & TDF_SINTR));
614 	thread_lock(td);
615 	sleepq_switch(wchan, pri);
616 	rval = sleepq_check_timeout();
617 	thread_unlock(td);
618 
619 	return (rval);
620 }
621 
622 /*
623  * Block the current thread until it is awakened from its sleep queue,
624  * it is interrupted by a signal, or it times out waiting to be awakened.
625  */
626 int
627 sleepq_timedwait_sig(void *wchan, int pri)
628 {
629 	int rcatch, rvalt, rvals;
630 
631 	rcatch = sleepq_catch_signals(wchan, pri);
632 	rvalt = sleepq_check_timeout();
633 	rvals = sleepq_check_signals();
634 	thread_unlock(curthread);
635 	if (rcatch)
636 		return (rcatch);
637 	if (rvals)
638 		return (rvals);
639 	return (rvalt);
640 }
641 
642 /*
643  * Removes a thread from a sleep queue and makes it
644  * runnable.
645  */
646 static int
647 sleepq_resume_thread(struct sleepqueue *sq, struct thread *td, int pri)
648 {
649 	struct sleepqueue_chain *sc;
650 
651 	MPASS(td != NULL);
652 	MPASS(sq->sq_wchan != NULL);
653 	MPASS(td->td_wchan == sq->sq_wchan);
654 	MPASS(td->td_sqqueue < NR_SLEEPQS && td->td_sqqueue >= 0);
655 	THREAD_LOCK_ASSERT(td, MA_OWNED);
656 	sc = SC_LOOKUP(sq->sq_wchan);
657 	mtx_assert(&sc->sc_lock, MA_OWNED);
658 
659 	/* Remove the thread from the queue. */
660 	TAILQ_REMOVE(&sq->sq_blocked[td->td_sqqueue], td, td_slpq);
661 
662 	/*
663 	 * Get a sleep queue for this thread.  If this is the last waiter,
664 	 * use the queue itself and take it out of the chain, otherwise,
665 	 * remove a queue from the free list.
666 	 */
667 	if (LIST_EMPTY(&sq->sq_free)) {
668 		td->td_sleepqueue = sq;
669 #ifdef INVARIANTS
670 		sq->sq_wchan = NULL;
671 #endif
672 #ifdef SLEEPQUEUE_PROFILING
673 		sc->sc_depth--;
674 #endif
675 	} else
676 		td->td_sleepqueue = LIST_FIRST(&sq->sq_free);
677 	LIST_REMOVE(td->td_sleepqueue, sq_hash);
678 
679 	td->td_wmesg = NULL;
680 	td->td_wchan = NULL;
681 	td->td_flags &= ~TDF_SINTR;
682 
683 	CTR3(KTR_PROC, "sleepq_wakeup: thread %p (pid %ld, %s)",
684 	    (void *)td, (long)td->td_proc->p_pid, td->td_name);
685 
686 	/* Adjust priority if requested. */
687 	MPASS(pri == 0 || (pri >= PRI_MIN && pri <= PRI_MAX));
688 	if (pri != 0 && td->td_priority > pri)
689 		sched_prio(td, pri);
690 
691 	/*
692 	 * Note that thread td might not be sleeping if it is running
693 	 * sleepq_catch_signals() on another CPU or is blocked on its
694 	 * proc lock to check signals.  There's no need to mark the
695 	 * thread runnable in that case.
696 	 */
697 	if (TD_IS_SLEEPING(td)) {
698 		TD_CLR_SLEEPING(td);
699 		return (setrunnable(td));
700 	}
701 	return (0);
702 }
703 
704 #ifdef INVARIANTS
705 /*
706  * UMA zone item deallocator.
707  */
708 static void
709 sleepq_dtor(void *mem, int size, void *arg)
710 {
711 	struct sleepqueue *sq;
712 	int i;
713 
714 	sq = mem;
715 	for (i = 0; i < NR_SLEEPQS; i++)
716 		MPASS(TAILQ_EMPTY(&sq->sq_blocked[i]));
717 }
718 #endif
719 
720 /*
721  * UMA zone item initializer.
722  */
723 static int
724 sleepq_init(void *mem, int size, int flags)
725 {
726 	struct sleepqueue *sq;
727 	int i;
728 
729 	bzero(mem, size);
730 	sq = mem;
731 	for (i = 0; i < NR_SLEEPQS; i++)
732 		TAILQ_INIT(&sq->sq_blocked[i]);
733 	LIST_INIT(&sq->sq_free);
734 	return (0);
735 }
736 
737 /*
738  * Find the highest priority thread sleeping on a wait channel and resume it.
739  */
740 int
741 sleepq_signal(void *wchan, int flags, int pri, int queue)
742 {
743 	struct sleepqueue *sq;
744 	struct thread *td, *besttd;
745 	int wakeup_swapper;
746 
747 	CTR2(KTR_PROC, "sleepq_signal(%p, %d)", wchan, flags);
748 	KASSERT(wchan != NULL, ("%s: invalid NULL wait channel", __func__));
749 	MPASS((queue >= 0) && (queue < NR_SLEEPQS));
750 	sq = sleepq_lookup(wchan);
751 	if (sq == NULL)
752 		return (0);
753 	KASSERT(sq->sq_type == (flags & SLEEPQ_TYPE),
754 	    ("%s: mismatch between sleep/wakeup and cv_*", __func__));
755 
756 	/*
757 	 * Find the highest priority thread on the queue.  If there is a
758 	 * tie, use the thread that first appears in the queue as it has
759 	 * been sleeping the longest since threads are always added to
760 	 * the tail of sleep queues.
761 	 */
762 	besttd = NULL;
763 	TAILQ_FOREACH(td, &sq->sq_blocked[queue], td_slpq) {
764 		if (besttd == NULL || td->td_priority < besttd->td_priority)
765 			besttd = td;
766 	}
767 	MPASS(besttd != NULL);
768 	thread_lock(besttd);
769 	wakeup_swapper = sleepq_resume_thread(sq, besttd, pri);
770 	thread_unlock(besttd);
771 	return (wakeup_swapper);
772 }
773 
774 /*
775  * Resume all threads sleeping on a specified wait channel.
776  */
777 int
778 sleepq_broadcast(void *wchan, int flags, int pri, int queue)
779 {
780 	struct sleepqueue *sq;
781 	struct thread *td, *tdn;
782 	int wakeup_swapper;
783 
784 	CTR2(KTR_PROC, "sleepq_broadcast(%p, %d)", wchan, flags);
785 	KASSERT(wchan != NULL, ("%s: invalid NULL wait channel", __func__));
786 	MPASS((queue >= 0) && (queue < NR_SLEEPQS));
787 	sq = sleepq_lookup(wchan);
788 	if (sq == NULL)
789 		return (0);
790 	KASSERT(sq->sq_type == (flags & SLEEPQ_TYPE),
791 	    ("%s: mismatch between sleep/wakeup and cv_*", __func__));
792 
793 	/* Resume all blocked threads on the sleep queue. */
794 	wakeup_swapper = 0;
795 	TAILQ_FOREACH_SAFE(td, &sq->sq_blocked[queue], td_slpq, tdn) {
796 		thread_lock(td);
797 		if (sleepq_resume_thread(sq, td, pri))
798 			wakeup_swapper = 1;
799 		thread_unlock(td);
800 	}
801 	return (wakeup_swapper);
802 }
803 
804 /*
805  * Time sleeping threads out.  When the timeout expires, the thread is
806  * removed from the sleep queue and made runnable if it is still asleep.
807  */
808 static void
809 sleepq_timeout(void *arg)
810 {
811 	struct sleepqueue_chain *sc;
812 	struct sleepqueue *sq;
813 	struct thread *td;
814 	void *wchan;
815 	int wakeup_swapper;
816 
817 	td = arg;
818 	wakeup_swapper = 0;
819 	CTR3(KTR_PROC, "sleepq_timeout: thread %p (pid %ld, %s)",
820 	    (void *)td, (long)td->td_proc->p_pid, (void *)td->td_name);
821 
822 	/*
823 	 * First, see if the thread is asleep and get the wait channel if
824 	 * it is.
825 	 */
826 	thread_lock(td);
827 	if (TD_IS_SLEEPING(td) && TD_ON_SLEEPQ(td)) {
828 		wchan = td->td_wchan;
829 		sc = SC_LOOKUP(wchan);
830 		THREAD_LOCKPTR_ASSERT(td, &sc->sc_lock);
831 		sq = sleepq_lookup(wchan);
832 		MPASS(sq != NULL);
833 		td->td_flags |= TDF_TIMEOUT;
834 		wakeup_swapper = sleepq_resume_thread(sq, td, 0);
835 		thread_unlock(td);
836 		if (wakeup_swapper)
837 			kick_proc0();
838 		return;
839 	}
840 
841 	/*
842 	 * If the thread is on the SLEEPQ but isn't sleeping yet, it
843 	 * can either be on another CPU in between sleepq_add() and
844 	 * one of the sleepq_*wait*() routines or it can be in
845 	 * sleepq_catch_signals().
846 	 */
847 	if (TD_ON_SLEEPQ(td)) {
848 		td->td_flags |= TDF_TIMEOUT;
849 		thread_unlock(td);
850 		return;
851 	}
852 
853 	/*
854 	 * Now check for the edge cases.  First, if TDF_TIMEOUT is set,
855 	 * then the other thread has already yielded to us, so clear
856 	 * the flag and resume it.  If TDF_TIMEOUT is not set, then the
857 	 * we know that the other thread is not on a sleep queue, but it
858 	 * hasn't resumed execution yet.  In that case, set TDF_TIMOFAIL
859 	 * to let it know that the timeout has already run and doesn't
860 	 * need to be canceled.
861 	 */
862 	if (td->td_flags & TDF_TIMEOUT) {
863 		MPASS(TD_IS_SLEEPING(td));
864 		td->td_flags &= ~TDF_TIMEOUT;
865 		TD_CLR_SLEEPING(td);
866 		wakeup_swapper = setrunnable(td);
867 	} else
868 		td->td_flags |= TDF_TIMOFAIL;
869 	thread_unlock(td);
870 	if (wakeup_swapper)
871 		kick_proc0();
872 }
873 
874 /*
875  * Resumes a specific thread from the sleep queue associated with a specific
876  * wait channel if it is on that queue.
877  */
878 void
879 sleepq_remove(struct thread *td, void *wchan)
880 {
881 	struct sleepqueue *sq;
882 	int wakeup_swapper;
883 
884 	/*
885 	 * Look up the sleep queue for this wait channel, then re-check
886 	 * that the thread is asleep on that channel, if it is not, then
887 	 * bail.
888 	 */
889 	MPASS(wchan != NULL);
890 	sleepq_lock(wchan);
891 	sq = sleepq_lookup(wchan);
892 	/*
893 	 * We can not lock the thread here as it may be sleeping on a
894 	 * different sleepq.  However, holding the sleepq lock for this
895 	 * wchan can guarantee that we do not miss a wakeup for this
896 	 * channel.  The asserts below will catch any false positives.
897 	 */
898 	if (!TD_ON_SLEEPQ(td) || td->td_wchan != wchan) {
899 		sleepq_release(wchan);
900 		return;
901 	}
902 	/* Thread is asleep on sleep queue sq, so wake it up. */
903 	thread_lock(td);
904 	MPASS(sq != NULL);
905 	MPASS(td->td_wchan == wchan);
906 	wakeup_swapper = sleepq_resume_thread(sq, td, 0);
907 	thread_unlock(td);
908 	sleepq_release(wchan);
909 	if (wakeup_swapper)
910 		kick_proc0();
911 }
912 
913 /*
914  * Abort a thread as if an interrupt had occurred.  Only abort
915  * interruptible waits (unfortunately it isn't safe to abort others).
916  */
917 int
918 sleepq_abort(struct thread *td, int intrval)
919 {
920 	struct sleepqueue *sq;
921 	void *wchan;
922 
923 	THREAD_LOCK_ASSERT(td, MA_OWNED);
924 	MPASS(TD_ON_SLEEPQ(td));
925 	MPASS(td->td_flags & TDF_SINTR);
926 	MPASS(intrval == EINTR || intrval == ERESTART);
927 
928 	/*
929 	 * If the TDF_TIMEOUT flag is set, just leave. A
930 	 * timeout is scheduled anyhow.
931 	 */
932 	if (td->td_flags & TDF_TIMEOUT)
933 		return (0);
934 
935 	CTR3(KTR_PROC, "sleepq_abort: thread %p (pid %ld, %s)",
936 	    (void *)td, (long)td->td_proc->p_pid, (void *)td->td_name);
937 	td->td_intrval = intrval;
938 	td->td_flags |= TDF_SLEEPABORT;
939 	/*
940 	 * If the thread has not slept yet it will find the signal in
941 	 * sleepq_catch_signals() and call sleepq_resume_thread.  Otherwise
942 	 * we have to do it here.
943 	 */
944 	if (!TD_IS_SLEEPING(td))
945 		return (0);
946 	wchan = td->td_wchan;
947 	MPASS(wchan != NULL);
948 	sq = sleepq_lookup(wchan);
949 	MPASS(sq != NULL);
950 
951 	/* Thread is asleep on sleep queue sq, so wake it up. */
952 	return (sleepq_resume_thread(sq, td, 0));
953 }
954 
955 #ifdef SLEEPQUEUE_PROFILING
956 #define	SLEEPQ_PROF_LOCATIONS	1024
957 #define	SLEEPQ_SBUFSIZE		(40 * 512)
958 struct sleepq_prof {
959 	LIST_ENTRY(sleepq_prof) sp_link;
960 	const char	*sp_wmesg;
961 	long		sp_count;
962 };
963 
964 LIST_HEAD(sqphead, sleepq_prof);
965 
966 struct sqphead sleepq_prof_free;
967 struct sqphead sleepq_hash[SC_TABLESIZE];
968 static struct sleepq_prof sleepq_profent[SLEEPQ_PROF_LOCATIONS];
969 static struct mtx sleepq_prof_lock;
970 MTX_SYSINIT(sleepq_prof_lock, &sleepq_prof_lock, "sleepq_prof", MTX_SPIN);
971 
972 static void
973 sleepq_profile(const char *wmesg)
974 {
975 	struct sleepq_prof *sp;
976 
977 	mtx_lock_spin(&sleepq_prof_lock);
978 	if (prof_enabled == 0)
979 		goto unlock;
980 	LIST_FOREACH(sp, &sleepq_hash[SC_HASH(wmesg)], sp_link)
981 		if (sp->sp_wmesg == wmesg)
982 			goto done;
983 	sp = LIST_FIRST(&sleepq_prof_free);
984 	if (sp == NULL)
985 		goto unlock;
986 	sp->sp_wmesg = wmesg;
987 	LIST_REMOVE(sp, sp_link);
988 	LIST_INSERT_HEAD(&sleepq_hash[SC_HASH(wmesg)], sp, sp_link);
989 done:
990 	sp->sp_count++;
991 unlock:
992 	mtx_unlock_spin(&sleepq_prof_lock);
993 	return;
994 }
995 
996 static void
997 sleepq_prof_reset(void)
998 {
999 	struct sleepq_prof *sp;
1000 	int enabled;
1001 	int i;
1002 
1003 	mtx_lock_spin(&sleepq_prof_lock);
1004 	enabled = prof_enabled;
1005 	prof_enabled = 0;
1006 	for (i = 0; i < SC_TABLESIZE; i++)
1007 		LIST_INIT(&sleepq_hash[i]);
1008 	LIST_INIT(&sleepq_prof_free);
1009 	for (i = 0; i < SLEEPQ_PROF_LOCATIONS; i++) {
1010 		sp = &sleepq_profent[i];
1011 		sp->sp_wmesg = NULL;
1012 		sp->sp_count = 0;
1013 		LIST_INSERT_HEAD(&sleepq_prof_free, sp, sp_link);
1014 	}
1015 	prof_enabled = enabled;
1016 	mtx_unlock_spin(&sleepq_prof_lock);
1017 }
1018 
1019 static int
1020 enable_sleepq_prof(SYSCTL_HANDLER_ARGS)
1021 {
1022 	int error, v;
1023 
1024 	v = prof_enabled;
1025 	error = sysctl_handle_int(oidp, &v, v, req);
1026 	if (error)
1027 		return (error);
1028 	if (req->newptr == NULL)
1029 		return (error);
1030 	if (v == prof_enabled)
1031 		return (0);
1032 	if (v == 1)
1033 		sleepq_prof_reset();
1034 	mtx_lock_spin(&sleepq_prof_lock);
1035 	prof_enabled = !!v;
1036 	mtx_unlock_spin(&sleepq_prof_lock);
1037 
1038 	return (0);
1039 }
1040 
1041 static int
1042 reset_sleepq_prof_stats(SYSCTL_HANDLER_ARGS)
1043 {
1044 	int error, v;
1045 
1046 	v = 0;
1047 	error = sysctl_handle_int(oidp, &v, 0, req);
1048 	if (error)
1049 		return (error);
1050 	if (req->newptr == NULL)
1051 		return (error);
1052 	if (v == 0)
1053 		return (0);
1054 	sleepq_prof_reset();
1055 
1056 	return (0);
1057 }
1058 
1059 static int
1060 dump_sleepq_prof_stats(SYSCTL_HANDLER_ARGS)
1061 {
1062 	static int multiplier = 1;
1063 	struct sleepq_prof *sp;
1064 	struct sbuf *sb;
1065 	int enabled;
1066 	int error;
1067 	int i;
1068 
1069 retry_sbufops:
1070 	sb = sbuf_new(NULL, NULL, SLEEPQ_SBUFSIZE * multiplier, SBUF_FIXEDLEN);
1071 	sbuf_printf(sb, "\nwmesg\tcount\n");
1072 	enabled = prof_enabled;
1073 	mtx_lock_spin(&sleepq_prof_lock);
1074 	prof_enabled = 0;
1075 	mtx_unlock_spin(&sleepq_prof_lock);
1076 	for (i = 0; i < SC_TABLESIZE; i++) {
1077 		LIST_FOREACH(sp, &sleepq_hash[i], sp_link) {
1078 			sbuf_printf(sb, "%s\t%ld\n",
1079 			    sp->sp_wmesg, sp->sp_count);
1080 			if (sbuf_overflowed(sb)) {
1081 				sbuf_delete(sb);
1082 				multiplier++;
1083 				goto retry_sbufops;
1084 			}
1085 		}
1086 	}
1087 	mtx_lock_spin(&sleepq_prof_lock);
1088 	prof_enabled = enabled;
1089 	mtx_unlock_spin(&sleepq_prof_lock);
1090 
1091 	sbuf_finish(sb);
1092 	error = SYSCTL_OUT(req, sbuf_data(sb), sbuf_len(sb) + 1);
1093 	sbuf_delete(sb);
1094 	return (error);
1095 }
1096 
1097 SYSCTL_PROC(_debug_sleepq, OID_AUTO, stats, CTLTYPE_STRING | CTLFLAG_RD,
1098     NULL, 0, dump_sleepq_prof_stats, "A", "Sleepqueue profiling statistics");
1099 SYSCTL_PROC(_debug_sleepq, OID_AUTO, reset, CTLTYPE_INT | CTLFLAG_RW,
1100     NULL, 0, reset_sleepq_prof_stats, "I",
1101     "Reset sleepqueue profiling statistics");
1102 SYSCTL_PROC(_debug_sleepq, OID_AUTO, enable, CTLTYPE_INT | CTLFLAG_RW,
1103     NULL, 0, enable_sleepq_prof, "I", "Enable sleepqueue profiling");
1104 #endif
1105 
1106 #ifdef DDB
1107 DB_SHOW_COMMAND(sleepq, db_show_sleepqueue)
1108 {
1109 	struct sleepqueue_chain *sc;
1110 	struct sleepqueue *sq;
1111 #ifdef INVARIANTS
1112 	struct lock_object *lock;
1113 #endif
1114 	struct thread *td;
1115 	void *wchan;
1116 	int i;
1117 
1118 	if (!have_addr)
1119 		return;
1120 
1121 	/*
1122 	 * First, see if there is an active sleep queue for the wait channel
1123 	 * indicated by the address.
1124 	 */
1125 	wchan = (void *)addr;
1126 	sc = SC_LOOKUP(wchan);
1127 	LIST_FOREACH(sq, &sc->sc_queues, sq_hash)
1128 		if (sq->sq_wchan == wchan)
1129 			goto found;
1130 
1131 	/*
1132 	 * Second, see if there is an active sleep queue at the address
1133 	 * indicated.
1134 	 */
1135 	for (i = 0; i < SC_TABLESIZE; i++)
1136 		LIST_FOREACH(sq, &sleepq_chains[i].sc_queues, sq_hash) {
1137 			if (sq == (struct sleepqueue *)addr)
1138 				goto found;
1139 		}
1140 
1141 	db_printf("Unable to locate a sleep queue via %p\n", (void *)addr);
1142 	return;
1143 found:
1144 	db_printf("Wait channel: %p\n", sq->sq_wchan);
1145 #ifdef INVARIANTS
1146 	db_printf("Queue type: %d\n", sq->sq_type);
1147 	if (sq->sq_lock) {
1148 		lock = sq->sq_lock;
1149 		db_printf("Associated Interlock: %p - (%s) %s\n", lock,
1150 		    LOCK_CLASS(lock)->lc_name, lock->lo_name);
1151 	}
1152 #endif
1153 	db_printf("Blocked threads:\n");
1154 	for (i = 0; i < NR_SLEEPQS; i++) {
1155 		db_printf("\nQueue[%d]:\n", i);
1156 		if (TAILQ_EMPTY(&sq->sq_blocked[i]))
1157 			db_printf("\tempty\n");
1158 		else
1159 			TAILQ_FOREACH(td, &sq->sq_blocked[0],
1160 				      td_slpq) {
1161 				db_printf("\t%p (tid %d, pid %d, \"%s\")\n", td,
1162 					  td->td_tid, td->td_proc->p_pid,
1163 					  td->td_name);
1164 			}
1165 	}
1166 }
1167 
1168 /* Alias 'show sleepqueue' to 'show sleepq'. */
1169 DB_SHOW_ALIAS(sleepqueue, db_show_sleepqueue);
1170 #endif
1171