xref: /freebsd/sys/kern/subr_sleepqueue.c (revision 830940567b49bb0c08dfaed40418999e76616909)
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 		if (flags & SLEEPQ_STOP_ON_BDRY)
345 			td->td_flags |= TDF_SBDRY;
346 	}
347 	thread_unlock(td);
348 }
349 
350 /*
351  * Sets a timeout that will remove the current thread from the specified
352  * sleep queue after timo ticks if the thread has not already been awakened.
353  */
354 void
355 sleepq_set_timeout(void *wchan, int timo)
356 {
357 	struct sleepqueue_chain *sc;
358 	struct thread *td;
359 
360 	td = curthread;
361 	sc = SC_LOOKUP(wchan);
362 	mtx_assert(&sc->sc_lock, MA_OWNED);
363 	MPASS(TD_ON_SLEEPQ(td));
364 	MPASS(td->td_sleepqueue == NULL);
365 	MPASS(wchan != NULL);
366 	callout_reset_curcpu(&td->td_slpcallout, timo, sleepq_timeout, td);
367 }
368 
369 /*
370  * Marks the pending sleep of the current thread as interruptible and
371  * makes an initial check for pending signals before putting a thread
372  * to sleep. Enters and exits with the thread lock held.  Thread lock
373  * may have transitioned from the sleepq lock to a run lock.
374  */
375 static int
376 sleepq_catch_signals(void *wchan, int pri)
377 {
378 	struct sleepqueue_chain *sc;
379 	struct sleepqueue *sq;
380 	struct thread *td;
381 	struct proc *p;
382 	struct sigacts *ps;
383 	int sig, ret, stop_allowed;
384 
385 	td = curthread;
386 	p = curproc;
387 	sc = SC_LOOKUP(wchan);
388 	mtx_assert(&sc->sc_lock, MA_OWNED);
389 	MPASS(wchan != NULL);
390 	/*
391 	 * See if there are any pending signals for this thread.  If not
392 	 * we can switch immediately.  Otherwise do the signal processing
393 	 * directly.
394 	 */
395 	thread_lock(td);
396 	if ((td->td_flags & (TDF_NEEDSIGCHK | TDF_NEEDSUSPCHK)) == 0) {
397 		sleepq_switch(wchan, pri);
398 		return (0);
399 	}
400 	stop_allowed = (td->td_flags & TDF_SBDRY) ? SIG_STOP_NOT_ALLOWED :
401 	    SIG_STOP_ALLOWED;
402 	thread_unlock(td);
403 	mtx_unlock_spin(&sc->sc_lock);
404 	CTR3(KTR_PROC, "sleepq catching signals: thread %p (pid %ld, %s)",
405 		(void *)td, (long)p->p_pid, td->td_name);
406 	PROC_LOCK(p);
407 	ps = p->p_sigacts;
408 	mtx_lock(&ps->ps_mtx);
409 	sig = cursig(td, stop_allowed);
410 	if (sig == 0) {
411 		mtx_unlock(&ps->ps_mtx);
412 		ret = thread_suspend_check(1);
413 		MPASS(ret == 0 || ret == EINTR || ret == ERESTART);
414 	} else {
415 		if (SIGISMEMBER(ps->ps_sigintr, sig))
416 			ret = EINTR;
417 		else
418 			ret = ERESTART;
419 		mtx_unlock(&ps->ps_mtx);
420 	}
421 	/*
422 	 * Lock the per-process spinlock prior to dropping the PROC_LOCK
423 	 * to avoid a signal delivery race.  PROC_LOCK, PROC_SLOCK, and
424 	 * thread_lock() are currently held in tdsignal().
425 	 */
426 	PROC_SLOCK(p);
427 	mtx_lock_spin(&sc->sc_lock);
428 	PROC_UNLOCK(p);
429 	thread_lock(td);
430 	PROC_SUNLOCK(p);
431 	if (ret == 0) {
432 		sleepq_switch(wchan, pri);
433 		return (0);
434 	}
435 	/*
436 	 * There were pending signals and this thread is still
437 	 * on the sleep queue, remove it from the sleep queue.
438 	 */
439 	if (TD_ON_SLEEPQ(td)) {
440 		sq = sleepq_lookup(wchan);
441 		if (sleepq_resume_thread(sq, td, 0)) {
442 #ifdef INVARIANTS
443 			/*
444 			 * This thread hasn't gone to sleep yet, so it
445 			 * should not be swapped out.
446 			 */
447 			panic("not waking up swapper");
448 #endif
449 		}
450 	}
451 	mtx_unlock_spin(&sc->sc_lock);
452 	MPASS(td->td_lock != &sc->sc_lock);
453 	return (ret);
454 }
455 
456 /*
457  * Switches to another thread if we are still asleep on a sleep queue.
458  * Returns with thread lock.
459  */
460 static void
461 sleepq_switch(void *wchan, int pri)
462 {
463 	struct sleepqueue_chain *sc;
464 	struct sleepqueue *sq;
465 	struct thread *td;
466 
467 	td = curthread;
468 	sc = SC_LOOKUP(wchan);
469 	mtx_assert(&sc->sc_lock, MA_OWNED);
470 	THREAD_LOCK_ASSERT(td, MA_OWNED);
471 
472 	/*
473 	 * If we have a sleep queue, then we've already been woken up, so
474 	 * just return.
475 	 */
476 	if (td->td_sleepqueue != NULL) {
477 		mtx_unlock_spin(&sc->sc_lock);
478 		return;
479 	}
480 
481 	/*
482 	 * If TDF_TIMEOUT is set, then our sleep has been timed out
483 	 * already but we are still on the sleep queue, so dequeue the
484 	 * thread and return.
485 	 */
486 	if (td->td_flags & TDF_TIMEOUT) {
487 		MPASS(TD_ON_SLEEPQ(td));
488 		sq = sleepq_lookup(wchan);
489 		if (sleepq_resume_thread(sq, td, 0)) {
490 #ifdef INVARIANTS
491 			/*
492 			 * This thread hasn't gone to sleep yet, so it
493 			 * should not be swapped out.
494 			 */
495 			panic("not waking up swapper");
496 #endif
497 		}
498 		mtx_unlock_spin(&sc->sc_lock);
499 		return;
500 	}
501 #ifdef SLEEPQUEUE_PROFILING
502 	if (prof_enabled)
503 		sleepq_profile(td->td_wmesg);
504 #endif
505 	MPASS(td->td_sleepqueue == NULL);
506 	sched_sleep(td, pri);
507 	thread_lock_set(td, &sc->sc_lock);
508 	TD_SET_SLEEPING(td);
509 	mi_switch(SW_VOL | SWT_SLEEPQ, NULL);
510 	KASSERT(TD_IS_RUNNING(td), ("running but not TDS_RUNNING"));
511 	CTR3(KTR_PROC, "sleepq resume: thread %p (pid %ld, %s)",
512 	    (void *)td, (long)td->td_proc->p_pid, (void *)td->td_name);
513 }
514 
515 /*
516  * Check to see if we timed out.
517  */
518 static int
519 sleepq_check_timeout(void)
520 {
521 	struct thread *td;
522 
523 	td = curthread;
524 	THREAD_LOCK_ASSERT(td, MA_OWNED);
525 
526 	/*
527 	 * If TDF_TIMEOUT is set, we timed out.
528 	 */
529 	if (td->td_flags & TDF_TIMEOUT) {
530 		td->td_flags &= ~TDF_TIMEOUT;
531 		return (EWOULDBLOCK);
532 	}
533 
534 	/*
535 	 * If TDF_TIMOFAIL is set, the timeout ran after we had
536 	 * already been woken up.
537 	 */
538 	if (td->td_flags & TDF_TIMOFAIL)
539 		td->td_flags &= ~TDF_TIMOFAIL;
540 
541 	/*
542 	 * If callout_stop() fails, then the timeout is running on
543 	 * another CPU, so synchronize with it to avoid having it
544 	 * accidentally wake up a subsequent sleep.
545 	 */
546 	else if (callout_stop(&td->td_slpcallout) == 0) {
547 		td->td_flags |= TDF_TIMEOUT;
548 		TD_SET_SLEEPING(td);
549 		mi_switch(SW_INVOL | SWT_SLEEPQTIMO, NULL);
550 	}
551 	return (0);
552 }
553 
554 /*
555  * Check to see if we were awoken by a signal.
556  */
557 static int
558 sleepq_check_signals(void)
559 {
560 	struct thread *td;
561 
562 	td = curthread;
563 	THREAD_LOCK_ASSERT(td, MA_OWNED);
564 
565 	/* We are no longer in an interruptible sleep. */
566 	if (td->td_flags & TDF_SINTR)
567 		td->td_flags &= ~(TDF_SINTR | TDF_SBDRY);
568 
569 	if (td->td_flags & TDF_SLEEPABORT) {
570 		td->td_flags &= ~TDF_SLEEPABORT;
571 		return (td->td_intrval);
572 	}
573 
574 	return (0);
575 }
576 
577 /*
578  * Block the current thread until it is awakened from its sleep queue.
579  */
580 void
581 sleepq_wait(void *wchan, int pri)
582 {
583 	struct thread *td;
584 
585 	td = curthread;
586 	MPASS(!(td->td_flags & TDF_SINTR));
587 	thread_lock(td);
588 	sleepq_switch(wchan, pri);
589 	thread_unlock(td);
590 }
591 
592 /*
593  * Block the current thread until it is awakened from its sleep queue
594  * or it is interrupted by a signal.
595  */
596 int
597 sleepq_wait_sig(void *wchan, int pri)
598 {
599 	int rcatch;
600 	int rval;
601 
602 	rcatch = sleepq_catch_signals(wchan, pri);
603 	rval = sleepq_check_signals();
604 	thread_unlock(curthread);
605 	if (rcatch)
606 		return (rcatch);
607 	return (rval);
608 }
609 
610 /*
611  * Block the current thread until it is awakened from its sleep queue
612  * or it times out while waiting.
613  */
614 int
615 sleepq_timedwait(void *wchan, int pri)
616 {
617 	struct thread *td;
618 	int rval;
619 
620 	td = curthread;
621 	MPASS(!(td->td_flags & TDF_SINTR));
622 	thread_lock(td);
623 	sleepq_switch(wchan, pri);
624 	rval = sleepq_check_timeout();
625 	thread_unlock(td);
626 
627 	return (rval);
628 }
629 
630 /*
631  * Block the current thread until it is awakened from its sleep queue,
632  * it is interrupted by a signal, or it times out waiting to be awakened.
633  */
634 int
635 sleepq_timedwait_sig(void *wchan, int pri)
636 {
637 	int rcatch, rvalt, rvals;
638 
639 	rcatch = sleepq_catch_signals(wchan, pri);
640 	rvalt = sleepq_check_timeout();
641 	rvals = sleepq_check_signals();
642 	thread_unlock(curthread);
643 	if (rcatch)
644 		return (rcatch);
645 	if (rvals)
646 		return (rvals);
647 	return (rvalt);
648 }
649 
650 /*
651  * Removes a thread from a sleep queue and makes it
652  * runnable.
653  */
654 static int
655 sleepq_resume_thread(struct sleepqueue *sq, struct thread *td, int pri)
656 {
657 	struct sleepqueue_chain *sc;
658 
659 	MPASS(td != NULL);
660 	MPASS(sq->sq_wchan != NULL);
661 	MPASS(td->td_wchan == sq->sq_wchan);
662 	MPASS(td->td_sqqueue < NR_SLEEPQS && td->td_sqqueue >= 0);
663 	THREAD_LOCK_ASSERT(td, MA_OWNED);
664 	sc = SC_LOOKUP(sq->sq_wchan);
665 	mtx_assert(&sc->sc_lock, MA_OWNED);
666 
667 	/* Remove the thread from the queue. */
668 	TAILQ_REMOVE(&sq->sq_blocked[td->td_sqqueue], td, td_slpq);
669 
670 	/*
671 	 * Get a sleep queue for this thread.  If this is the last waiter,
672 	 * use the queue itself and take it out of the chain, otherwise,
673 	 * remove a queue from the free list.
674 	 */
675 	if (LIST_EMPTY(&sq->sq_free)) {
676 		td->td_sleepqueue = sq;
677 #ifdef INVARIANTS
678 		sq->sq_wchan = NULL;
679 #endif
680 #ifdef SLEEPQUEUE_PROFILING
681 		sc->sc_depth--;
682 #endif
683 	} else
684 		td->td_sleepqueue = LIST_FIRST(&sq->sq_free);
685 	LIST_REMOVE(td->td_sleepqueue, sq_hash);
686 
687 	td->td_wmesg = NULL;
688 	td->td_wchan = NULL;
689 	td->td_flags &= ~(TDF_SINTR | TDF_SBDRY);
690 
691 	CTR3(KTR_PROC, "sleepq_wakeup: thread %p (pid %ld, %s)",
692 	    (void *)td, (long)td->td_proc->p_pid, td->td_name);
693 
694 	/* Adjust priority if requested. */
695 	MPASS(pri == 0 || (pri >= PRI_MIN && pri <= PRI_MAX));
696 	if (pri != 0 && td->td_priority > pri)
697 		sched_prio(td, pri);
698 
699 	/*
700 	 * Note that thread td might not be sleeping if it is running
701 	 * sleepq_catch_signals() on another CPU or is blocked on its
702 	 * proc lock to check signals.  There's no need to mark the
703 	 * thread runnable in that case.
704 	 */
705 	if (TD_IS_SLEEPING(td)) {
706 		TD_CLR_SLEEPING(td);
707 		return (setrunnable(td));
708 	}
709 	return (0);
710 }
711 
712 #ifdef INVARIANTS
713 /*
714  * UMA zone item deallocator.
715  */
716 static void
717 sleepq_dtor(void *mem, int size, void *arg)
718 {
719 	struct sleepqueue *sq;
720 	int i;
721 
722 	sq = mem;
723 	for (i = 0; i < NR_SLEEPQS; i++)
724 		MPASS(TAILQ_EMPTY(&sq->sq_blocked[i]));
725 }
726 #endif
727 
728 /*
729  * UMA zone item initializer.
730  */
731 static int
732 sleepq_init(void *mem, int size, int flags)
733 {
734 	struct sleepqueue *sq;
735 	int i;
736 
737 	bzero(mem, size);
738 	sq = mem;
739 	for (i = 0; i < NR_SLEEPQS; i++)
740 		TAILQ_INIT(&sq->sq_blocked[i]);
741 	LIST_INIT(&sq->sq_free);
742 	return (0);
743 }
744 
745 /*
746  * Find the highest priority thread sleeping on a wait channel and resume it.
747  */
748 int
749 sleepq_signal(void *wchan, int flags, int pri, int queue)
750 {
751 	struct sleepqueue *sq;
752 	struct thread *td, *besttd;
753 	int wakeup_swapper;
754 
755 	CTR2(KTR_PROC, "sleepq_signal(%p, %d)", wchan, flags);
756 	KASSERT(wchan != NULL, ("%s: invalid NULL wait channel", __func__));
757 	MPASS((queue >= 0) && (queue < NR_SLEEPQS));
758 	sq = sleepq_lookup(wchan);
759 	if (sq == NULL)
760 		return (0);
761 	KASSERT(sq->sq_type == (flags & SLEEPQ_TYPE),
762 	    ("%s: mismatch between sleep/wakeup and cv_*", __func__));
763 
764 	/*
765 	 * Find the highest priority thread on the queue.  If there is a
766 	 * tie, use the thread that first appears in the queue as it has
767 	 * been sleeping the longest since threads are always added to
768 	 * the tail of sleep queues.
769 	 */
770 	besttd = NULL;
771 	TAILQ_FOREACH(td, &sq->sq_blocked[queue], td_slpq) {
772 		if (besttd == NULL || td->td_priority < besttd->td_priority)
773 			besttd = td;
774 	}
775 	MPASS(besttd != NULL);
776 	thread_lock(besttd);
777 	wakeup_swapper = sleepq_resume_thread(sq, besttd, pri);
778 	thread_unlock(besttd);
779 	return (wakeup_swapper);
780 }
781 
782 /*
783  * Resume all threads sleeping on a specified wait channel.
784  */
785 int
786 sleepq_broadcast(void *wchan, int flags, int pri, int queue)
787 {
788 	struct sleepqueue *sq;
789 	struct thread *td, *tdn;
790 	int wakeup_swapper;
791 
792 	CTR2(KTR_PROC, "sleepq_broadcast(%p, %d)", wchan, flags);
793 	KASSERT(wchan != NULL, ("%s: invalid NULL wait channel", __func__));
794 	MPASS((queue >= 0) && (queue < NR_SLEEPQS));
795 	sq = sleepq_lookup(wchan);
796 	if (sq == NULL)
797 		return (0);
798 	KASSERT(sq->sq_type == (flags & SLEEPQ_TYPE),
799 	    ("%s: mismatch between sleep/wakeup and cv_*", __func__));
800 
801 	/* Resume all blocked threads on the sleep queue. */
802 	wakeup_swapper = 0;
803 	TAILQ_FOREACH_SAFE(td, &sq->sq_blocked[queue], td_slpq, tdn) {
804 		thread_lock(td);
805 		if (sleepq_resume_thread(sq, td, pri))
806 			wakeup_swapper = 1;
807 		thread_unlock(td);
808 	}
809 	return (wakeup_swapper);
810 }
811 
812 /*
813  * Time sleeping threads out.  When the timeout expires, the thread is
814  * removed from the sleep queue and made runnable if it is still asleep.
815  */
816 static void
817 sleepq_timeout(void *arg)
818 {
819 	struct sleepqueue_chain *sc;
820 	struct sleepqueue *sq;
821 	struct thread *td;
822 	void *wchan;
823 	int wakeup_swapper;
824 
825 	td = arg;
826 	wakeup_swapper = 0;
827 	CTR3(KTR_PROC, "sleepq_timeout: thread %p (pid %ld, %s)",
828 	    (void *)td, (long)td->td_proc->p_pid, (void *)td->td_name);
829 
830 	/*
831 	 * First, see if the thread is asleep and get the wait channel if
832 	 * it is.
833 	 */
834 	thread_lock(td);
835 	if (TD_IS_SLEEPING(td) && TD_ON_SLEEPQ(td)) {
836 		wchan = td->td_wchan;
837 		sc = SC_LOOKUP(wchan);
838 		THREAD_LOCKPTR_ASSERT(td, &sc->sc_lock);
839 		sq = sleepq_lookup(wchan);
840 		MPASS(sq != NULL);
841 		td->td_flags |= TDF_TIMEOUT;
842 		wakeup_swapper = sleepq_resume_thread(sq, td, 0);
843 		thread_unlock(td);
844 		if (wakeup_swapper)
845 			kick_proc0();
846 		return;
847 	}
848 
849 	/*
850 	 * If the thread is on the SLEEPQ but isn't sleeping yet, it
851 	 * can either be on another CPU in between sleepq_add() and
852 	 * one of the sleepq_*wait*() routines or it can be in
853 	 * sleepq_catch_signals().
854 	 */
855 	if (TD_ON_SLEEPQ(td)) {
856 		td->td_flags |= TDF_TIMEOUT;
857 		thread_unlock(td);
858 		return;
859 	}
860 
861 	/*
862 	 * Now check for the edge cases.  First, if TDF_TIMEOUT is set,
863 	 * then the other thread has already yielded to us, so clear
864 	 * the flag and resume it.  If TDF_TIMEOUT is not set, then the
865 	 * we know that the other thread is not on a sleep queue, but it
866 	 * hasn't resumed execution yet.  In that case, set TDF_TIMOFAIL
867 	 * to let it know that the timeout has already run and doesn't
868 	 * need to be canceled.
869 	 */
870 	if (td->td_flags & TDF_TIMEOUT) {
871 		MPASS(TD_IS_SLEEPING(td));
872 		td->td_flags &= ~TDF_TIMEOUT;
873 		TD_CLR_SLEEPING(td);
874 		wakeup_swapper = setrunnable(td);
875 	} else
876 		td->td_flags |= TDF_TIMOFAIL;
877 	thread_unlock(td);
878 	if (wakeup_swapper)
879 		kick_proc0();
880 }
881 
882 /*
883  * Resumes a specific thread from the sleep queue associated with a specific
884  * wait channel if it is on that queue.
885  */
886 void
887 sleepq_remove(struct thread *td, void *wchan)
888 {
889 	struct sleepqueue *sq;
890 	int wakeup_swapper;
891 
892 	/*
893 	 * Look up the sleep queue for this wait channel, then re-check
894 	 * that the thread is asleep on that channel, if it is not, then
895 	 * bail.
896 	 */
897 	MPASS(wchan != NULL);
898 	sleepq_lock(wchan);
899 	sq = sleepq_lookup(wchan);
900 	/*
901 	 * We can not lock the thread here as it may be sleeping on a
902 	 * different sleepq.  However, holding the sleepq lock for this
903 	 * wchan can guarantee that we do not miss a wakeup for this
904 	 * channel.  The asserts below will catch any false positives.
905 	 */
906 	if (!TD_ON_SLEEPQ(td) || td->td_wchan != wchan) {
907 		sleepq_release(wchan);
908 		return;
909 	}
910 	/* Thread is asleep on sleep queue sq, so wake it up. */
911 	thread_lock(td);
912 	MPASS(sq != NULL);
913 	MPASS(td->td_wchan == wchan);
914 	wakeup_swapper = sleepq_resume_thread(sq, td, 0);
915 	thread_unlock(td);
916 	sleepq_release(wchan);
917 	if (wakeup_swapper)
918 		kick_proc0();
919 }
920 
921 /*
922  * Abort a thread as if an interrupt had occurred.  Only abort
923  * interruptible waits (unfortunately it isn't safe to abort others).
924  */
925 int
926 sleepq_abort(struct thread *td, int intrval)
927 {
928 	struct sleepqueue *sq;
929 	void *wchan;
930 
931 	THREAD_LOCK_ASSERT(td, MA_OWNED);
932 	MPASS(TD_ON_SLEEPQ(td));
933 	MPASS(td->td_flags & TDF_SINTR);
934 	MPASS(intrval == EINTR || intrval == ERESTART);
935 
936 	/*
937 	 * If the TDF_TIMEOUT flag is set, just leave. A
938 	 * timeout is scheduled anyhow.
939 	 */
940 	if (td->td_flags & TDF_TIMEOUT)
941 		return (0);
942 
943 	CTR3(KTR_PROC, "sleepq_abort: thread %p (pid %ld, %s)",
944 	    (void *)td, (long)td->td_proc->p_pid, (void *)td->td_name);
945 	td->td_intrval = intrval;
946 	td->td_flags |= TDF_SLEEPABORT;
947 	/*
948 	 * If the thread has not slept yet it will find the signal in
949 	 * sleepq_catch_signals() and call sleepq_resume_thread.  Otherwise
950 	 * we have to do it here.
951 	 */
952 	if (!TD_IS_SLEEPING(td))
953 		return (0);
954 	wchan = td->td_wchan;
955 	MPASS(wchan != NULL);
956 	sq = sleepq_lookup(wchan);
957 	MPASS(sq != NULL);
958 
959 	/* Thread is asleep on sleep queue sq, so wake it up. */
960 	return (sleepq_resume_thread(sq, td, 0));
961 }
962 
963 #ifdef SLEEPQUEUE_PROFILING
964 #define	SLEEPQ_PROF_LOCATIONS	1024
965 #define	SLEEPQ_SBUFSIZE		(40 * 512)
966 struct sleepq_prof {
967 	LIST_ENTRY(sleepq_prof) sp_link;
968 	const char	*sp_wmesg;
969 	long		sp_count;
970 };
971 
972 LIST_HEAD(sqphead, sleepq_prof);
973 
974 struct sqphead sleepq_prof_free;
975 struct sqphead sleepq_hash[SC_TABLESIZE];
976 static struct sleepq_prof sleepq_profent[SLEEPQ_PROF_LOCATIONS];
977 static struct mtx sleepq_prof_lock;
978 MTX_SYSINIT(sleepq_prof_lock, &sleepq_prof_lock, "sleepq_prof", MTX_SPIN);
979 
980 static void
981 sleepq_profile(const char *wmesg)
982 {
983 	struct sleepq_prof *sp;
984 
985 	mtx_lock_spin(&sleepq_prof_lock);
986 	if (prof_enabled == 0)
987 		goto unlock;
988 	LIST_FOREACH(sp, &sleepq_hash[SC_HASH(wmesg)], sp_link)
989 		if (sp->sp_wmesg == wmesg)
990 			goto done;
991 	sp = LIST_FIRST(&sleepq_prof_free);
992 	if (sp == NULL)
993 		goto unlock;
994 	sp->sp_wmesg = wmesg;
995 	LIST_REMOVE(sp, sp_link);
996 	LIST_INSERT_HEAD(&sleepq_hash[SC_HASH(wmesg)], sp, sp_link);
997 done:
998 	sp->sp_count++;
999 unlock:
1000 	mtx_unlock_spin(&sleepq_prof_lock);
1001 	return;
1002 }
1003 
1004 static void
1005 sleepq_prof_reset(void)
1006 {
1007 	struct sleepq_prof *sp;
1008 	int enabled;
1009 	int i;
1010 
1011 	mtx_lock_spin(&sleepq_prof_lock);
1012 	enabled = prof_enabled;
1013 	prof_enabled = 0;
1014 	for (i = 0; i < SC_TABLESIZE; i++)
1015 		LIST_INIT(&sleepq_hash[i]);
1016 	LIST_INIT(&sleepq_prof_free);
1017 	for (i = 0; i < SLEEPQ_PROF_LOCATIONS; i++) {
1018 		sp = &sleepq_profent[i];
1019 		sp->sp_wmesg = NULL;
1020 		sp->sp_count = 0;
1021 		LIST_INSERT_HEAD(&sleepq_prof_free, sp, sp_link);
1022 	}
1023 	prof_enabled = enabled;
1024 	mtx_unlock_spin(&sleepq_prof_lock);
1025 }
1026 
1027 static int
1028 enable_sleepq_prof(SYSCTL_HANDLER_ARGS)
1029 {
1030 	int error, v;
1031 
1032 	v = prof_enabled;
1033 	error = sysctl_handle_int(oidp, &v, v, req);
1034 	if (error)
1035 		return (error);
1036 	if (req->newptr == NULL)
1037 		return (error);
1038 	if (v == prof_enabled)
1039 		return (0);
1040 	if (v == 1)
1041 		sleepq_prof_reset();
1042 	mtx_lock_spin(&sleepq_prof_lock);
1043 	prof_enabled = !!v;
1044 	mtx_unlock_spin(&sleepq_prof_lock);
1045 
1046 	return (0);
1047 }
1048 
1049 static int
1050 reset_sleepq_prof_stats(SYSCTL_HANDLER_ARGS)
1051 {
1052 	int error, v;
1053 
1054 	v = 0;
1055 	error = sysctl_handle_int(oidp, &v, 0, req);
1056 	if (error)
1057 		return (error);
1058 	if (req->newptr == NULL)
1059 		return (error);
1060 	if (v == 0)
1061 		return (0);
1062 	sleepq_prof_reset();
1063 
1064 	return (0);
1065 }
1066 
1067 static int
1068 dump_sleepq_prof_stats(SYSCTL_HANDLER_ARGS)
1069 {
1070 	static int multiplier = 1;
1071 	struct sleepq_prof *sp;
1072 	struct sbuf *sb;
1073 	int enabled;
1074 	int error;
1075 	int i;
1076 
1077 retry_sbufops:
1078 	sb = sbuf_new(NULL, NULL, SLEEPQ_SBUFSIZE * multiplier, SBUF_FIXEDLEN);
1079 	sbuf_printf(sb, "\nwmesg\tcount\n");
1080 	enabled = prof_enabled;
1081 	mtx_lock_spin(&sleepq_prof_lock);
1082 	prof_enabled = 0;
1083 	mtx_unlock_spin(&sleepq_prof_lock);
1084 	for (i = 0; i < SC_TABLESIZE; i++) {
1085 		LIST_FOREACH(sp, &sleepq_hash[i], sp_link) {
1086 			sbuf_printf(sb, "%s\t%ld\n",
1087 			    sp->sp_wmesg, sp->sp_count);
1088 			if (sbuf_overflowed(sb)) {
1089 				sbuf_delete(sb);
1090 				multiplier++;
1091 				goto retry_sbufops;
1092 			}
1093 		}
1094 	}
1095 	mtx_lock_spin(&sleepq_prof_lock);
1096 	prof_enabled = enabled;
1097 	mtx_unlock_spin(&sleepq_prof_lock);
1098 
1099 	sbuf_finish(sb);
1100 	error = SYSCTL_OUT(req, sbuf_data(sb), sbuf_len(sb) + 1);
1101 	sbuf_delete(sb);
1102 	return (error);
1103 }
1104 
1105 SYSCTL_PROC(_debug_sleepq, OID_AUTO, stats, CTLTYPE_STRING | CTLFLAG_RD,
1106     NULL, 0, dump_sleepq_prof_stats, "A", "Sleepqueue profiling statistics");
1107 SYSCTL_PROC(_debug_sleepq, OID_AUTO, reset, CTLTYPE_INT | CTLFLAG_RW,
1108     NULL, 0, reset_sleepq_prof_stats, "I",
1109     "Reset sleepqueue profiling statistics");
1110 SYSCTL_PROC(_debug_sleepq, OID_AUTO, enable, CTLTYPE_INT | CTLFLAG_RW,
1111     NULL, 0, enable_sleepq_prof, "I", "Enable sleepqueue profiling");
1112 #endif
1113 
1114 #ifdef DDB
1115 DB_SHOW_COMMAND(sleepq, db_show_sleepqueue)
1116 {
1117 	struct sleepqueue_chain *sc;
1118 	struct sleepqueue *sq;
1119 #ifdef INVARIANTS
1120 	struct lock_object *lock;
1121 #endif
1122 	struct thread *td;
1123 	void *wchan;
1124 	int i;
1125 
1126 	if (!have_addr)
1127 		return;
1128 
1129 	/*
1130 	 * First, see if there is an active sleep queue for the wait channel
1131 	 * indicated by the address.
1132 	 */
1133 	wchan = (void *)addr;
1134 	sc = SC_LOOKUP(wchan);
1135 	LIST_FOREACH(sq, &sc->sc_queues, sq_hash)
1136 		if (sq->sq_wchan == wchan)
1137 			goto found;
1138 
1139 	/*
1140 	 * Second, see if there is an active sleep queue at the address
1141 	 * indicated.
1142 	 */
1143 	for (i = 0; i < SC_TABLESIZE; i++)
1144 		LIST_FOREACH(sq, &sleepq_chains[i].sc_queues, sq_hash) {
1145 			if (sq == (struct sleepqueue *)addr)
1146 				goto found;
1147 		}
1148 
1149 	db_printf("Unable to locate a sleep queue via %p\n", (void *)addr);
1150 	return;
1151 found:
1152 	db_printf("Wait channel: %p\n", sq->sq_wchan);
1153 #ifdef INVARIANTS
1154 	db_printf("Queue type: %d\n", sq->sq_type);
1155 	if (sq->sq_lock) {
1156 		lock = sq->sq_lock;
1157 		db_printf("Associated Interlock: %p - (%s) %s\n", lock,
1158 		    LOCK_CLASS(lock)->lc_name, lock->lo_name);
1159 	}
1160 #endif
1161 	db_printf("Blocked threads:\n");
1162 	for (i = 0; i < NR_SLEEPQS; i++) {
1163 		db_printf("\nQueue[%d]:\n", i);
1164 		if (TAILQ_EMPTY(&sq->sq_blocked[i]))
1165 			db_printf("\tempty\n");
1166 		else
1167 			TAILQ_FOREACH(td, &sq->sq_blocked[0],
1168 				      td_slpq) {
1169 				db_printf("\t%p (tid %d, pid %d, \"%s\")\n", td,
1170 					  td->td_tid, td->td_proc->p_pid,
1171 					  td->td_name);
1172 			}
1173 	}
1174 }
1175 
1176 /* Alias 'show sleepqueue' to 'show sleepq'. */
1177 DB_SHOW_ALIAS(sleepqueue, db_show_sleepqueue);
1178 #endif
1179