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 "opt_sleepqueue_profiling.h"
63
64 #include <sys/cdefs.h>
65 __FBSDID("$FreeBSD: releng/5.3/sys/kern/subr_sleepqueue.c 136588 2004-10-16 08:43:07Z cvs2svn $");
66
67 #include <sys/param.h>
68 #include <sys/systm.h>
69 #include <sys/lock.h>
70 #include <sys/kernel.h>
71 #include <sys/ktr.h>
72 #include <sys/malloc.h>
73 #include <sys/mutex.h>
74 #include <sys/proc.h>
75 #include <sys/sched.h>
76 #include <sys/signalvar.h>
77 #include <sys/sleepqueue.h>
78 #include <sys/sysctl.h>
79
80 /*
81 * Constants for the hash table of sleep queue chains. These constants are
82 * the same ones that 4BSD (and possibly earlier versions of BSD) used.
83 * Basically, we ignore the lower 8 bits of the address since most wait
84 * channel pointers are aligned and only look at the next 7 bits for the
85 * hash. SC_TABLESIZE must be a power of two for SC_MASK to work properly.
86 */
87 #define SC_TABLESIZE 128 /* Must be power of 2. */
88 #define SC_MASK (SC_TABLESIZE - 1)
89 #define SC_SHIFT 8
90 #define SC_HASH(wc) (((uintptr_t)(wc) >> SC_SHIFT) & SC_MASK)
91 #define SC_LOOKUP(wc) &sleepq_chains[SC_HASH(wc)]
92
93 /*
94 * There two different lists of sleep queues. Both lists are connected
95 * via the sq_hash entries. The first list is the sleep queue chain list
96 * that a sleep queue is on when it is attached to a wait channel. The
97 * second list is the free list hung off of a sleep queue that is attached
98 * to a wait channel.
99 *
100 * Each sleep queue also contains the wait channel it is attached to, the
101 * list of threads blocked on that wait channel, flags specific to the
102 * wait channel, and the lock used to synchronize with a wait channel.
103 * The flags are used to catch mismatches between the various consumers
104 * of the sleep queue API (e.g. sleep/wakeup and condition variables).
105 * The lock pointer is only used when invariants are enabled for various
106 * debugging checks.
107 *
108 * Locking key:
109 * c - sleep queue chain lock
110 */
111 struct sleepqueue {
112 TAILQ_HEAD(, thread) sq_blocked; /* (c) Blocked threads. */
113 LIST_ENTRY(sleepqueue) sq_hash; /* (c) Chain and free list. */
114 LIST_HEAD(, sleepqueue) sq_free; /* (c) Free queues. */
115 void *sq_wchan; /* (c) Wait channel. */
116 int sq_type; /* (c) Queue type. */
117 #ifdef INVARIANTS
118 struct mtx *sq_lock; /* (c) Associated lock. */
119 #endif
120 };
121
122 struct sleepqueue_chain {
123 LIST_HEAD(, sleepqueue) sc_queues; /* List of sleep queues. */
124 struct mtx sc_lock; /* Spin lock for this chain. */
125 #ifdef SLEEPQUEUE_PROFILING
126 u_int sc_depth; /* Length of sc_queues. */
127 u_int sc_max_depth; /* Max length of sc_queues. */
128 #endif
129 };
130
131 #ifdef SLEEPQUEUE_PROFILING
132 u_int sleepq_max_depth;
133 SYSCTL_NODE(_debug, OID_AUTO, sleepq, CTLFLAG_RD, 0, "sleepq profiling");
134 SYSCTL_NODE(_debug_sleepq, OID_AUTO, chains, CTLFLAG_RD, 0,
135 "sleepq chain stats");
136 SYSCTL_UINT(_debug_sleepq, OID_AUTO, max_depth, CTLFLAG_RD, &sleepq_max_depth,
137 0, "maxmimum depth achieved of a single chain");
138 #endif
139 static struct sleepqueue_chain sleepq_chains[SC_TABLESIZE];
140
141 MALLOC_DEFINE(M_SLEEPQUEUE, "sleep queues", "sleep queues");
142
143 /*
144 * Prototypes for non-exported routines.
145 */
146 static int sleepq_check_timeout(void);
147 static void sleepq_switch(void *wchan);
148 static void sleepq_timeout(void *arg);
149 static void sleepq_remove_thread(struct sleepqueue *sq, struct thread *td);
150 static void sleepq_resume_thread(struct thread *td, int pri);
151
152 /*
153 * Early initialization of sleep queues that is called from the sleepinit()
154 * SYSINIT.
155 */
156 void
157 init_sleepqueues(void)
158 {
159 #ifdef SLEEPQUEUE_PROFILING
160 struct sysctl_oid *chain_oid;
161 char chain_name[10];
162 #endif
163 int i;
164
165 for (i = 0; i < SC_TABLESIZE; i++) {
166 LIST_INIT(&sleepq_chains[i].sc_queues);
167 mtx_init(&sleepq_chains[i].sc_lock, "sleepq chain", NULL,
168 MTX_SPIN);
169 #ifdef SLEEPQUEUE_PROFILING
170 snprintf(chain_name, sizeof(chain_name), "%d", i);
171 chain_oid = SYSCTL_ADD_NODE(NULL,
172 SYSCTL_STATIC_CHILDREN(_debug_sleepq_chains), OID_AUTO,
173 chain_name, CTLFLAG_RD, NULL, "sleepq chain stats");
174 SYSCTL_ADD_UINT(NULL, SYSCTL_CHILDREN(chain_oid), OID_AUTO,
175 "depth", CTLFLAG_RD, &sleepq_chains[i].sc_depth, 0, NULL);
176 SYSCTL_ADD_UINT(NULL, SYSCTL_CHILDREN(chain_oid), OID_AUTO,
177 "max_depth", CTLFLAG_RD, &sleepq_chains[i].sc_max_depth, 0,
178 NULL);
179 #endif
180 }
181 thread0.td_sleepqueue = sleepq_alloc();
182 }
183
184 /*
185 * Malloc and initialize a new sleep queue for a new thread.
186 */
187 struct sleepqueue *
188 sleepq_alloc(void)
189 {
190 struct sleepqueue *sq;
191
192 sq = malloc(sizeof(struct sleepqueue), M_SLEEPQUEUE, M_WAITOK | M_ZERO);
193 TAILQ_INIT(&sq->sq_blocked);
194 LIST_INIT(&sq->sq_free);
195 return (sq);
196 }
197
198 /*
199 * Free a sleep queue when a thread is destroyed.
200 */
201 void
202 sleepq_free(struct sleepqueue *sq)
203 {
204
205 MPASS(sq != NULL);
206 MPASS(TAILQ_EMPTY(&sq->sq_blocked));
207 free(sq, M_SLEEPQUEUE);
208 }
209
210 /*
211 * Look up the sleep queue associated with a given wait channel in the hash
212 * table locking the associated sleep queue chain. Return holdind the sleep
213 * queue chain lock. If no queue is found in the table, NULL is returned.
214 */
215 struct sleepqueue *
216 sleepq_lookup(void *wchan)
217 {
218 struct sleepqueue_chain *sc;
219 struct sleepqueue *sq;
220
221 KASSERT(wchan != NULL, ("%s: invalid NULL wait channel", __func__));
222 sc = SC_LOOKUP(wchan);
223 mtx_lock_spin(&sc->sc_lock);
224 LIST_FOREACH(sq, &sc->sc_queues, sq_hash)
225 if (sq->sq_wchan == wchan)
226 return (sq);
227 return (NULL);
228 }
229
230 /*
231 * Unlock the sleep queue chain associated with a given wait channel.
232 */
233 void
234 sleepq_release(void *wchan)
235 {
236 struct sleepqueue_chain *sc;
237
238 sc = SC_LOOKUP(wchan);
239 mtx_unlock_spin(&sc->sc_lock);
240 }
241
242 /*
243 * Places the current thread on the sleepqueue for the specified wait
244 * channel. If INVARIANTS is enabled, then it associates the passed in
245 * lock with the sleepq to make sure it is held when that sleep queue is
246 * woken up.
247 */
248 void
249 sleepq_add(struct sleepqueue *sq, void *wchan, struct mtx *lock,
250 const char *wmesg, int flags)
251 {
252 struct sleepqueue_chain *sc;
253 struct thread *td, *td1;
254
255 td = curthread;
256 sc = SC_LOOKUP(wchan);
257 mtx_assert(&sc->sc_lock, MA_OWNED);
258 MPASS(td->td_sleepqueue != NULL);
259 MPASS(wchan != NULL);
260
261 /* If the passed in sleep queue is NULL, use this thread's queue. */
262 if (sq == NULL) {
263 #ifdef SLEEPQUEUE_PROFILING
264 sc->sc_depth++;
265 if (sc->sc_depth > sc->sc_max_depth) {
266 sc->sc_max_depth = sc->sc_depth;
267 if (sc->sc_max_depth > sleepq_max_depth)
268 sleepq_max_depth = sc->sc_max_depth;
269 }
270 #endif
271 sq = td->td_sleepqueue;
272 LIST_INSERT_HEAD(&sc->sc_queues, sq, sq_hash);
273 KASSERT(TAILQ_EMPTY(&sq->sq_blocked),
274 ("thread's sleep queue has a non-empty queue"));
275 KASSERT(LIST_EMPTY(&sq->sq_free),
276 ("thread's sleep queue has a non-empty free list"));
277 KASSERT(sq->sq_wchan == NULL, ("stale sq_wchan pointer"));
278 sq->sq_wchan = wchan;
279 #ifdef INVARIANTS
280 sq->sq_lock = lock;
281 #endif
282 sq->sq_type = flags & SLEEPQ_TYPE;
283 TAILQ_INSERT_TAIL(&sq->sq_blocked, td, td_slpq);
284 } else {
285 MPASS(wchan == sq->sq_wchan);
286 MPASS(lock == sq->sq_lock);
287 TAILQ_FOREACH(td1, &sq->sq_blocked, td_slpq)
288 if (td1->td_priority > td->td_priority)
289 break;
290 if (td1 != NULL)
291 TAILQ_INSERT_BEFORE(td1, td, td_slpq);
292 else
293 TAILQ_INSERT_TAIL(&sq->sq_blocked, td, td_slpq);
294 LIST_INSERT_HEAD(&sq->sq_free, td->td_sleepqueue, sq_hash);
295 }
296 td->td_sleepqueue = NULL;
297 mtx_lock_spin(&sched_lock);
298 td->td_wchan = wchan;
299 td->td_wmesg = wmesg;
300 if (flags & SLEEPQ_INTERRUPTIBLE)
301 td->td_flags |= TDF_SINTR;
302 mtx_unlock_spin(&sched_lock);
303 }
304
305 /*
306 * Sets a timeout that will remove the current thread from the specified
307 * sleep queue after timo ticks if the thread has not already been awakened.
308 */
309 void
310 sleepq_set_timeout(void *wchan, int timo)
311 {
312 struct sleepqueue_chain *sc;
313 struct thread *td;
314
315 td = curthread;
316 sc = SC_LOOKUP(wchan);
317 mtx_assert(&sc->sc_lock, MA_OWNED);
318 MPASS(TD_ON_SLEEPQ(td));
319 MPASS(td->td_sleepqueue == NULL);
320 MPASS(wchan != NULL);
321 callout_reset(&td->td_slpcallout, timo, sleepq_timeout, td);
322 }
323
324 /*
325 * Marks the pending sleep of the current thread as interruptible and
326 * makes an initial check for pending signals before putting a thread
327 * to sleep.
328 */
329 int
330 sleepq_catch_signals(void *wchan)
331 {
332 struct sleepqueue_chain *sc;
333 struct sleepqueue *sq;
334 struct thread *td;
335 struct proc *p;
336 int do_upcall;
337 int sig;
338
339 do_upcall = 0;
340 td = curthread;
341 p = td->td_proc;
342 sc = SC_LOOKUP(wchan);
343 mtx_assert(&sc->sc_lock, MA_OWNED);
344 MPASS(td->td_sleepqueue == NULL);
345 MPASS(wchan != NULL);
346 CTR3(KTR_PROC, "sleepq catching signals: thread %p (pid %ld, %s)",
347 (void *)td, (long)p->p_pid, p->p_comm);
348
349 /* Mark thread as being in an interruptible sleep. */
350 MPASS(td->td_flags & TDF_SINTR);
351 MPASS(TD_ON_SLEEPQ(td));
352 sleepq_release(wchan);
353
354 /* See if there are any pending signals for this thread. */
355 PROC_LOCK(p);
356 mtx_lock(&p->p_sigacts->ps_mtx);
357 sig = cursig(td);
358 mtx_unlock(&p->p_sigacts->ps_mtx);
359 if (sig == 0 && thread_suspend_check(1))
360 sig = SIGSTOP;
361 else
362 do_upcall = thread_upcall_check(td);
363 PROC_UNLOCK(p);
364
365 /*
366 * If there were pending signals and this thread is still on
367 * the sleep queue, remove it from the sleep queue. If the
368 * thread was removed from the sleep queue while we were blocked
369 * above, then clear TDF_SINTR before returning.
370 */
371 sq = sleepq_lookup(wchan);
372 mtx_lock_spin(&sched_lock);
373 if (TD_ON_SLEEPQ(td) && (sig != 0 || do_upcall != 0)) {
374 mtx_unlock_spin(&sched_lock);
375 sleepq_remove_thread(sq, td);
376 } else {
377 if (!TD_ON_SLEEPQ(td) && sig == 0)
378 td->td_flags &= ~TDF_SINTR;
379 mtx_unlock_spin(&sched_lock);
380 }
381 return (sig);
382 }
383
384 /*
385 * Switches to another thread if we are still asleep on a sleep queue and
386 * drop the lock on the sleepqueue chain. Returns with sched_lock held.
387 */
388 static void
389 sleepq_switch(void *wchan)
390 {
391 struct sleepqueue_chain *sc;
392 struct thread *td;
393
394 td = curthread;
395 sc = SC_LOOKUP(wchan);
396 mtx_assert(&sc->sc_lock, MA_OWNED);
397
398 /*
399 * If we have a sleep queue, then we've already been woken up, so
400 * just return.
401 */
402 if (td->td_sleepqueue != NULL) {
403 MPASS(!TD_ON_SLEEPQ(td));
404 mtx_unlock_spin(&sc->sc_lock);
405 mtx_lock_spin(&sched_lock);
406 return;
407 }
408
409 /*
410 * Otherwise, actually go to sleep.
411 */
412 mtx_lock_spin(&sched_lock);
413 mtx_unlock_spin(&sc->sc_lock);
414
415 sched_sleep(td);
416 TD_SET_SLEEPING(td);
417 mi_switch(SW_VOL, NULL);
418 KASSERT(TD_IS_RUNNING(td), ("running but not TDS_RUNNING"));
419 CTR3(KTR_PROC, "sleepq resume: thread %p (pid %ld, %s)",
420 (void *)td, (long)td->td_proc->p_pid, (void *)td->td_proc->p_comm);
421 }
422
423 /*
424 * Check to see if we timed out.
425 */
426 static int
427 sleepq_check_timeout(void)
428 {
429 struct thread *td;
430
431 mtx_assert(&sched_lock, MA_OWNED);
432 td = curthread;
433
434 /*
435 * If TDF_TIMEOUT is set, we timed out.
436 */
437 if (td->td_flags & TDF_TIMEOUT) {
438 td->td_flags &= ~TDF_TIMEOUT;
439 return (EWOULDBLOCK);
440 }
441
442 /*
443 * If TDF_TIMOFAIL is set, the timeout ran after we had
444 * already been woken up.
445 */
446 if (td->td_flags & TDF_TIMOFAIL)
447 td->td_flags &= ~TDF_TIMOFAIL;
448
449 /*
450 * If callout_stop() fails, then the timeout is running on
451 * another CPU, so synchronize with it to avoid having it
452 * accidentally wake up a subsequent sleep.
453 */
454 else if (callout_stop(&td->td_slpcallout) == 0) {
455 td->td_flags |= TDF_TIMEOUT;
456 TD_SET_SLEEPING(td);
457 mi_switch(SW_INVOL, NULL);
458 }
459 return (0);
460 }
461
462 /*
463 * Check to see if we were awoken by a signal.
464 */
465 static int
466 sleepq_check_signals(void)
467 {
468 struct thread *td;
469
470 mtx_assert(&sched_lock, MA_OWNED);
471 td = curthread;
472
473 /*
474 * If TDF_SINTR is clear, then we were awakened while executing
475 * sleepq_catch_signals().
476 */
477 if (!(td->td_flags & TDF_SINTR))
478 return (0);
479
480 /* We are no longer in an interruptible sleep. */
481 td->td_flags &= ~TDF_SINTR;
482
483 if (td->td_flags & TDF_INTERRUPT)
484 return (td->td_intrval);
485 return (0);
486 }
487
488 /*
489 * If we were in an interruptible sleep and we weren't interrupted and
490 * didn't timeout, check to see if there are any pending signals and
491 * which return value we should use if so. The return value from an
492 * earlier call to sleepq_catch_signals() should be passed in as the
493 * argument.
494 */
495 int
496 sleepq_calc_signal_retval(int sig)
497 {
498 struct thread *td;
499 struct proc *p;
500 int rval;
501
502 td = curthread;
503 p = td->td_proc;
504 PROC_LOCK(p);
505 mtx_lock(&p->p_sigacts->ps_mtx);
506 /* XXX: Should we always be calling cursig()? */
507 if (sig == 0)
508 sig = cursig(td);
509 if (sig != 0) {
510 if (SIGISMEMBER(p->p_sigacts->ps_sigintr, sig))
511 rval = EINTR;
512 else
513 rval = ERESTART;
514 } else
515 rval = 0;
516 mtx_unlock(&p->p_sigacts->ps_mtx);
517 PROC_UNLOCK(p);
518 return (rval);
519 }
520
521 /*
522 * Block the current thread until it is awakened from its sleep queue.
523 */
524 void
525 sleepq_wait(void *wchan)
526 {
527
528 MPASS(!(curthread->td_flags & TDF_SINTR));
529 sleepq_switch(wchan);
530 mtx_unlock_spin(&sched_lock);
531 }
532
533 /*
534 * Block the current thread until it is awakened from its sleep queue
535 * or it is interrupted by a signal.
536 */
537 int
538 sleepq_wait_sig(void *wchan)
539 {
540 int rval;
541
542 sleepq_switch(wchan);
543 rval = sleepq_check_signals();
544 mtx_unlock_spin(&sched_lock);
545 return (rval);
546 }
547
548 /*
549 * Block the current thread until it is awakened from its sleep queue
550 * or it times out while waiting.
551 */
552 int
553 sleepq_timedwait(void *wchan)
554 {
555 int rval;
556
557 MPASS(!(curthread->td_flags & TDF_SINTR));
558 sleepq_switch(wchan);
559 rval = sleepq_check_timeout();
560 mtx_unlock_spin(&sched_lock);
561 return (rval);
562 }
563
564 /*
565 * Block the current thread until it is awakened from its sleep queue,
566 * it is interrupted by a signal, or it times out waiting to be awakened.
567 */
568 int
569 sleepq_timedwait_sig(void *wchan, int signal_caught)
570 {
571 int rvalt, rvals;
572
573 sleepq_switch(wchan);
574 rvalt = sleepq_check_timeout();
575 rvals = sleepq_check_signals();
576 mtx_unlock_spin(&sched_lock);
577 if (signal_caught || rvalt == 0)
578 return (rvals);
579 else
580 return (rvalt);
581 }
582
583 /*
584 * Removes a thread from a sleep queue.
585 */
586 static void
587 sleepq_remove_thread(struct sleepqueue *sq, struct thread *td)
588 {
589 struct sleepqueue_chain *sc;
590
591 MPASS(td != NULL);
592 MPASS(sq->sq_wchan != NULL);
593 MPASS(td->td_wchan == sq->sq_wchan);
594 sc = SC_LOOKUP(sq->sq_wchan);
595 mtx_assert(&sc->sc_lock, MA_OWNED);
596
597 /* Remove the thread from the queue. */
598 TAILQ_REMOVE(&sq->sq_blocked, td, td_slpq);
599
600 /*
601 * Get a sleep queue for this thread. If this is the last waiter,
602 * use the queue itself and take it out of the chain, otherwise,
603 * remove a queue from the free list.
604 */
605 if (LIST_EMPTY(&sq->sq_free)) {
606 td->td_sleepqueue = sq;
607 #ifdef INVARIANTS
608 sq->sq_wchan = NULL;
609 #endif
610 #ifdef SLEEPQUEUE_PROFILING
611 sc->sc_depth--;
612 #endif
613 } else
614 td->td_sleepqueue = LIST_FIRST(&sq->sq_free);
615 LIST_REMOVE(td->td_sleepqueue, sq_hash);
616
617 mtx_lock_spin(&sched_lock);
618 td->td_wmesg = NULL;
619 td->td_wchan = NULL;
620 mtx_unlock_spin(&sched_lock);
621 }
622
623 /*
624 * Resumes a thread that was asleep on a queue.
625 */
626 static void
627 sleepq_resume_thread(struct thread *td, int pri)
628 {
629
630 /*
631 * Note that thread td might not be sleeping if it is running
632 * sleepq_catch_signals() on another CPU or is blocked on
633 * its proc lock to check signals. It doesn't hurt to clear
634 * the sleeping flag if it isn't set though, so we just always
635 * do it. However, we can't assert that it is set.
636 */
637 mtx_lock_spin(&sched_lock);
638 CTR3(KTR_PROC, "sleepq_wakeup: thread %p (pid %ld, %s)",
639 (void *)td, (long)td->td_proc->p_pid, td->td_proc->p_comm);
640 TD_CLR_SLEEPING(td);
641
642 /* Adjust priority if requested. */
643 MPASS(pri == -1 || (pri >= PRI_MIN && pri <= PRI_MAX));
644 if (pri != -1 && td->td_priority > pri)
645 sched_prio(td, pri);
646 setrunnable(td);
647 mtx_unlock_spin(&sched_lock);
648 }
649
650 /*
651 * Find the highest priority thread sleeping on a wait channel and resume it.
652 */
653 void
654 sleepq_signal(void *wchan, int flags, int pri)
655 {
656 struct sleepqueue *sq;
657 struct thread *td;
658
659 CTR2(KTR_PROC, "sleepq_signal(%p, %d)", wchan, flags);
660 KASSERT(wchan != NULL, ("%s: invalid NULL wait channel", __func__));
661 sq = sleepq_lookup(wchan);
662 if (sq == NULL) {
663 sleepq_release(wchan);
664 return;
665 }
666 KASSERT(sq->sq_type == (flags & SLEEPQ_TYPE),
667 ("%s: mismatch between sleep/wakeup and cv_*", __func__));
668 /* XXX: Do for all sleep queues eventually. */
669 if (flags & SLEEPQ_CONDVAR)
670 mtx_assert(sq->sq_lock, MA_OWNED);
671
672 /* Remove first thread from queue and awaken it. */
673 td = TAILQ_FIRST(&sq->sq_blocked);
674 sleepq_remove_thread(sq, td);
675 sleepq_release(wchan);
676 sleepq_resume_thread(td, pri);
677 }
678
679 /*
680 * Resume all threads sleeping on a specified wait channel.
681 */
682 void
683 sleepq_broadcast(void *wchan, int flags, int pri)
684 {
685 TAILQ_HEAD(, thread) list;
686 struct sleepqueue *sq;
687 struct thread *td;
688
689 CTR2(KTR_PROC, "sleepq_broadcast(%p, %d)", wchan, flags);
690 KASSERT(wchan != NULL, ("%s: invalid NULL wait channel", __func__));
691 sq = sleepq_lookup(wchan);
692 if (sq == NULL) {
693 sleepq_release(wchan);
694 return;
695 }
696 KASSERT(sq->sq_type == (flags & SLEEPQ_TYPE),
697 ("%s: mismatch between sleep/wakeup and cv_*", __func__));
698 /* XXX: Do for all sleep queues eventually. */
699 if (flags & SLEEPQ_CONDVAR)
700 mtx_assert(sq->sq_lock, MA_OWNED);
701
702 /* Move blocked threads from the sleep queue to a temporary list. */
703 TAILQ_INIT(&list);
704 while (!TAILQ_EMPTY(&sq->sq_blocked)) {
705 td = TAILQ_FIRST(&sq->sq_blocked);
706 sleepq_remove_thread(sq, td);
707 TAILQ_INSERT_TAIL(&list, td, td_slpq);
708 }
709 sleepq_release(wchan);
710
711 /* Resume all the threads on the temporary list. */
712 while (!TAILQ_EMPTY(&list)) {
713 td = TAILQ_FIRST(&list);
714 TAILQ_REMOVE(&list, td, td_slpq);
715 sleepq_resume_thread(td, pri);
716 }
717 }
718
719 /*
720 * Time sleeping threads out. When the timeout expires, the thread is
721 * removed from the sleep queue and made runnable if it is still asleep.
722 */
723 static void
724 sleepq_timeout(void *arg)
725 {
726 struct sleepqueue *sq;
727 struct thread *td;
728 void *wchan;
729
730 td = arg;
731 CTR3(KTR_PROC, "sleepq_timeout: thread %p (pid %ld, %s)",
732 (void *)td, (long)td->td_proc->p_pid, (void *)td->td_proc->p_comm);
733
734 /*
735 * First, see if the thread is asleep and get the wait channel if
736 * it is.
737 */
738 mtx_lock_spin(&sched_lock);
739 if (TD_ON_SLEEPQ(td)) {
740 wchan = td->td_wchan;
741 mtx_unlock_spin(&sched_lock);
742 sq = sleepq_lookup(wchan);
743 mtx_lock_spin(&sched_lock);
744 } else {
745 wchan = NULL;
746 sq = NULL;
747 }
748
749 /*
750 * At this point, if the thread is still on the sleep queue,
751 * we have that sleep queue locked as it cannot migrate sleep
752 * queues while we dropped sched_lock. If it had resumed and
753 * was on another CPU while the lock was dropped, it would have
754 * seen that TDF_TIMEOUT and TDF_TIMOFAIL are clear and the
755 * call to callout_stop() to stop this routine would have failed
756 * meaning that it would have already set TDF_TIMEOUT to
757 * synchronize with this function.
758 */
759 if (TD_ON_SLEEPQ(td)) {
760 MPASS(td->td_wchan == wchan);
761 MPASS(sq != NULL);
762 td->td_flags |= TDF_TIMEOUT;
763 mtx_unlock_spin(&sched_lock);
764 sleepq_remove_thread(sq, td);
765 sleepq_release(wchan);
766 sleepq_resume_thread(td, -1);
767 return;
768 } else if (wchan != NULL)
769 sleepq_release(wchan);
770
771 /*
772 * Now check for the edge cases. First, if TDF_TIMEOUT is set,
773 * then the other thread has already yielded to us, so clear
774 * the flag and resume it. If TDF_TIMEOUT is not set, then the
775 * we know that the other thread is not on a sleep queue, but it
776 * hasn't resumed execution yet. In that case, set TDF_TIMOFAIL
777 * to let it know that the timeout has already run and doesn't
778 * need to be canceled.
779 */
780 if (td->td_flags & TDF_TIMEOUT) {
781 MPASS(TD_IS_SLEEPING(td));
782 td->td_flags &= ~TDF_TIMEOUT;
783 TD_CLR_SLEEPING(td);
784 setrunnable(td);
785 } else
786 td->td_flags |= TDF_TIMOFAIL;
787 mtx_unlock_spin(&sched_lock);
788 }
789
790 /*
791 * Resumes a specific thread from the sleep queue associated with a specific
792 * wait channel if it is on that queue.
793 */
794 void
795 sleepq_remove(struct thread *td, void *wchan)
796 {
797 struct sleepqueue *sq;
798
799 /*
800 * Look up the sleep queue for this wait channel, then re-check
801 * that the thread is asleep on that channel, if it is not, then
802 * bail.
803 */
804 MPASS(wchan != NULL);
805 sq = sleepq_lookup(wchan);
806 mtx_lock_spin(&sched_lock);
807 if (!TD_ON_SLEEPQ(td) || td->td_wchan != wchan) {
808 mtx_unlock_spin(&sched_lock);
809 sleepq_release(wchan);
810 return;
811 }
812 mtx_unlock_spin(&sched_lock);
813 MPASS(sq != NULL);
814
815 /* Thread is asleep on sleep queue sq, so wake it up. */
816 sleepq_remove_thread(sq, td);
817 sleepq_release(wchan);
818 sleepq_resume_thread(td, -1);
819 }
820
821 /*
822 * Abort a thread as if an interrupt had occurred. Only abort
823 * interruptible waits (unfortunately it isn't safe to abort others).
824 *
825 * XXX: What in the world does the comment below mean?
826 * Also, whatever the signal code does...
827 */
828 void
829 sleepq_abort(struct thread *td)
830 {
831 void *wchan;
832
833 mtx_assert(&sched_lock, MA_OWNED);
834 MPASS(TD_ON_SLEEPQ(td));
835 MPASS(td->td_flags & TDF_SINTR);
836
837 /*
838 * If the TDF_TIMEOUT flag is set, just leave. A
839 * timeout is scheduled anyhow.
840 */
841 if (td->td_flags & TDF_TIMEOUT)
842 return;
843
844 CTR3(KTR_PROC, "sleepq_abort: thread %p (pid %ld, %s)",
845 (void *)td, (long)td->td_proc->p_pid, (void *)td->td_proc->p_comm);
846 wchan = td->td_wchan;
847 mtx_unlock_spin(&sched_lock);
848 sleepq_remove(td, wchan);
849 mtx_lock_spin(&sched_lock);
850 }
Cache object: 028c22c1664f0d8a143220a58b93611c
|