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/sched.h>
77 #include <sys/signalvar.h>
78 #include <sys/sleepqueue.h>
79 #include <sys/sysctl.h>
80
81 #include <vm/uma.h>
82
83 #ifdef DDB
84 #include <ddb/ddb.h>
85 #endif
86
87 /*
88 * Constants for the hash table of sleep queue chains. These constants are
89 * the same ones that 4BSD (and possibly earlier versions of BSD) used.
90 * Basically, we ignore the lower 8 bits of the address since most wait
91 * channel pointers are aligned and only look at the next 7 bits for the
92 * hash. SC_TABLESIZE must be a power of two for SC_MASK to work properly.
93 */
94 #define SC_TABLESIZE 128 /* Must be power of 2. */
95 #define SC_MASK (SC_TABLESIZE - 1)
96 #define SC_SHIFT 8
97 #define SC_HASH(wc) (((uintptr_t)(wc) >> SC_SHIFT) & SC_MASK)
98 #define SC_LOOKUP(wc) &sleepq_chains[SC_HASH(wc)]
99 #define NR_SLEEPQS 2
100 /*
101 * There two different lists of sleep queues. Both lists are connected
102 * via the sq_hash entries. The first list is the sleep queue chain list
103 * that a sleep queue is on when it is attached to a wait channel. The
104 * second list is the free list hung off of a sleep queue that is attached
105 * to a wait channel.
106 *
107 * Each sleep queue also contains the wait channel it is attached to, the
108 * list of threads blocked on that wait channel, flags specific to the
109 * wait channel, and the lock used to synchronize with a wait channel.
110 * The flags are used to catch mismatches between the various consumers
111 * of the sleep queue API (e.g. sleep/wakeup and condition variables).
112 * The lock pointer is only used when invariants are enabled for various
113 * debugging checks.
114 *
115 * Locking key:
116 * c - sleep queue chain lock
117 */
118 struct sleepqueue {
119 TAILQ_HEAD(, thread) sq_blocked[NR_SLEEPQS]; /* (c) Blocked threads. */
120 LIST_ENTRY(sleepqueue) sq_hash; /* (c) Chain and free list. */
121 LIST_HEAD(, sleepqueue) sq_free; /* (c) Free queues. */
122 void *sq_wchan; /* (c) Wait channel. */
123 #ifdef INVARIANTS
124 int sq_type; /* (c) Queue type. */
125 struct lock_object *sq_lock; /* (c) Associated lock. */
126 #endif
127 };
128
129 struct sleepqueue_chain {
130 LIST_HEAD(, sleepqueue) sc_queues; /* List of sleep queues. */
131 struct mtx sc_lock; /* Spin lock for this chain. */
132 #ifdef SLEEPQUEUE_PROFILING
133 u_int sc_depth; /* Length of sc_queues. */
134 u_int sc_max_depth; /* Max length of sc_queues. */
135 #endif
136 };
137
138 #ifdef SLEEPQUEUE_PROFILING
139 u_int sleepq_max_depth;
140 SYSCTL_NODE(_debug, OID_AUTO, sleepq, CTLFLAG_RD, 0, "sleepq profiling");
141 SYSCTL_NODE(_debug_sleepq, OID_AUTO, chains, CTLFLAG_RD, 0,
142 "sleepq chain stats");
143 SYSCTL_UINT(_debug_sleepq, OID_AUTO, max_depth, CTLFLAG_RD, &sleepq_max_depth,
144 0, "maxmimum depth achieved of a single chain");
145 #endif
146 static struct sleepqueue_chain sleepq_chains[SC_TABLESIZE];
147 static uma_zone_t sleepq_zone;
148
149 /*
150 * Prototypes for non-exported routines.
151 */
152 static int sleepq_catch_signals(void *wchan);
153 static int sleepq_check_signals(void);
154 static int sleepq_check_timeout(void);
155 #ifdef INVARIANTS
156 static void sleepq_dtor(void *mem, int size, void *arg);
157 #endif
158 static int sleepq_init(void *mem, int size, int flags);
159 static int sleepq_resume_thread(struct sleepqueue *sq, struct thread *td,
160 int pri);
161 static void sleepq_switch(void *wchan);
162 static void sleepq_timeout(void *arg);
163
164 /*
165 * Early initialization of sleep queues that is called from the sleepinit()
166 * SYSINIT.
167 */
168 void
169 init_sleepqueues(void)
170 {
171 #ifdef SLEEPQUEUE_PROFILING
172 struct sysctl_oid *chain_oid;
173 char chain_name[10];
174 #endif
175 int i;
176
177 for (i = 0; i < SC_TABLESIZE; i++) {
178 LIST_INIT(&sleepq_chains[i].sc_queues);
179 mtx_init(&sleepq_chains[i].sc_lock, "sleepq chain", NULL,
180 MTX_SPIN | MTX_RECURSE);
181 #ifdef SLEEPQUEUE_PROFILING
182 snprintf(chain_name, sizeof(chain_name), "%d", i);
183 chain_oid = SYSCTL_ADD_NODE(NULL,
184 SYSCTL_STATIC_CHILDREN(_debug_sleepq_chains), OID_AUTO,
185 chain_name, CTLFLAG_RD, NULL, "sleepq chain stats");
186 SYSCTL_ADD_UINT(NULL, SYSCTL_CHILDREN(chain_oid), OID_AUTO,
187 "depth", CTLFLAG_RD, &sleepq_chains[i].sc_depth, 0, NULL);
188 SYSCTL_ADD_UINT(NULL, SYSCTL_CHILDREN(chain_oid), OID_AUTO,
189 "max_depth", CTLFLAG_RD, &sleepq_chains[i].sc_max_depth, 0,
190 NULL);
191 #endif
192 }
193 sleepq_zone = uma_zcreate("SLEEPQUEUE", sizeof(struct sleepqueue),
194 #ifdef INVARIANTS
195 NULL, sleepq_dtor, sleepq_init, NULL, UMA_ALIGN_CACHE, 0);
196 #else
197 NULL, NULL, sleepq_init, NULL, UMA_ALIGN_CACHE, 0);
198 #endif
199
200 thread0.td_sleepqueue = sleepq_alloc();
201 }
202
203 /*
204 * Get a sleep queue for a new thread.
205 */
206 struct sleepqueue *
207 sleepq_alloc(void)
208 {
209
210 return (uma_zalloc(sleepq_zone, M_WAITOK));
211 }
212
213 /*
214 * Free a sleep queue when a thread is destroyed.
215 */
216 void
217 sleepq_free(struct sleepqueue *sq)
218 {
219
220 uma_zfree(sleepq_zone, sq);
221 }
222
223 /*
224 * Lock the sleep queue chain associated with the specified wait channel.
225 */
226 void
227 sleepq_lock(void *wchan)
228 {
229 struct sleepqueue_chain *sc;
230
231 sc = SC_LOOKUP(wchan);
232 mtx_lock_spin(&sc->sc_lock);
233 }
234
235 /*
236 * Look up the sleep queue associated with a given wait channel in the hash
237 * table locking the associated sleep queue chain. If no queue is found in
238 * the table, NULL is returned.
239 */
240 struct sleepqueue *
241 sleepq_lookup(void *wchan)
242 {
243 struct sleepqueue_chain *sc;
244 struct sleepqueue *sq;
245
246 KASSERT(wchan != NULL, ("%s: invalid NULL wait channel", __func__));
247 sc = SC_LOOKUP(wchan);
248 mtx_assert(&sc->sc_lock, MA_OWNED);
249 LIST_FOREACH(sq, &sc->sc_queues, sq_hash)
250 if (sq->sq_wchan == wchan)
251 return (sq);
252 return (NULL);
253 }
254
255 /*
256 * Unlock the sleep queue chain associated with a given wait channel.
257 */
258 void
259 sleepq_release(void *wchan)
260 {
261 struct sleepqueue_chain *sc;
262
263 sc = SC_LOOKUP(wchan);
264 mtx_unlock_spin(&sc->sc_lock);
265 }
266
267 /*
268 * Places the current thread on the sleep queue for the specified wait
269 * channel. If INVARIANTS is enabled, then it associates the passed in
270 * lock with the sleepq to make sure it is held when that sleep queue is
271 * woken up.
272 */
273 void
274 sleepq_add(void *wchan, struct lock_object *lock, const char *wmesg, int flags,
275 int queue)
276 {
277 struct sleepqueue_chain *sc;
278 struct sleepqueue *sq;
279 struct thread *td;
280
281 td = curthread;
282 sc = SC_LOOKUP(wchan);
283 mtx_assert(&sc->sc_lock, MA_OWNED);
284 MPASS(td->td_sleepqueue != NULL);
285 MPASS(wchan != NULL);
286 MPASS((queue >= 0) && (queue < NR_SLEEPQS));
287
288 /* If this thread is not allowed to sleep, die a horrible death. */
289 KASSERT(!(td->td_pflags & TDP_NOSLEEPING),
290 ("Trying sleep, but thread marked as sleeping prohibited"));
291
292 /* Look up the sleep queue associated with the wait channel 'wchan'. */
293 sq = sleepq_lookup(wchan);
294
295 /*
296 * If the wait channel does not already have a sleep queue, use
297 * this thread's sleep queue. Otherwise, insert the current thread
298 * into the sleep queue already in use by this wait channel.
299 */
300 if (sq == NULL) {
301 #ifdef INVARIANTS
302 int i;
303
304 sq = td->td_sleepqueue;
305 for (i = 0; i < NR_SLEEPQS; i++)
306 KASSERT(TAILQ_EMPTY(&sq->sq_blocked[i]),
307 ("thread's sleep queue %d is not empty", i));
308 KASSERT(LIST_EMPTY(&sq->sq_free),
309 ("thread's sleep queue has a non-empty free list"));
310 KASSERT(sq->sq_wchan == NULL, ("stale sq_wchan pointer"));
311 sq->sq_lock = lock;
312 sq->sq_type = flags & SLEEPQ_TYPE;
313 #endif
314 #ifdef SLEEPQUEUE_PROFILING
315 sc->sc_depth++;
316 if (sc->sc_depth > sc->sc_max_depth) {
317 sc->sc_max_depth = sc->sc_depth;
318 if (sc->sc_max_depth > sleepq_max_depth)
319 sleepq_max_depth = sc->sc_max_depth;
320 }
321 #endif
322 sq = td->td_sleepqueue;
323 LIST_INSERT_HEAD(&sc->sc_queues, sq, sq_hash);
324 sq->sq_wchan = wchan;
325 } else {
326 MPASS(wchan == sq->sq_wchan);
327 MPASS(lock == sq->sq_lock);
328 MPASS((flags & SLEEPQ_TYPE) == sq->sq_type);
329 LIST_INSERT_HEAD(&sq->sq_free, td->td_sleepqueue, sq_hash);
330 }
331 thread_lock(td);
332 TAILQ_INSERT_TAIL(&sq->sq_blocked[queue], td, td_slpq);
333 td->td_sleepqueue = NULL;
334 td->td_sqqueue = queue;
335 td->td_wchan = wchan;
336 td->td_wmesg = wmesg;
337 if (flags & SLEEPQ_INTERRUPTIBLE) {
338 td->td_flags |= TDF_SINTR;
339 td->td_flags &= ~TDF_SLEEPABORT;
340 }
341 thread_unlock(td);
342 }
343
344 /*
345 * Sets a timeout that will remove the current thread from the specified
346 * sleep queue after timo ticks if the thread has not already been awakened.
347 */
348 void
349 sleepq_set_timeout(void *wchan, int timo)
350 {
351 struct sleepqueue_chain *sc;
352 struct thread *td;
353
354 td = curthread;
355 sc = SC_LOOKUP(wchan);
356 mtx_assert(&sc->sc_lock, MA_OWNED);
357 MPASS(TD_ON_SLEEPQ(td));
358 MPASS(td->td_sleepqueue == NULL);
359 MPASS(wchan != NULL);
360 callout_reset(&td->td_slpcallout, timo, sleepq_timeout, td);
361 }
362
363 /*
364 * Marks the pending sleep of the current thread as interruptible and
365 * makes an initial check for pending signals before putting a thread
366 * to sleep. Enters and exits with the thread lock held. Thread lock
367 * may have transitioned from the sleepq lock to a run lock.
368 */
369 static int
370 sleepq_catch_signals(void *wchan)
371 {
372 struct sleepqueue_chain *sc;
373 struct sleepqueue *sq;
374 struct thread *td;
375 struct proc *p;
376 struct sigacts *ps;
377 int sig, ret;
378
379 td = curthread;
380 p = curproc;
381 sc = SC_LOOKUP(wchan);
382 mtx_assert(&sc->sc_lock, MA_OWNED);
383 MPASS(wchan != NULL);
384 CTR3(KTR_PROC, "sleepq catching signals: thread %p (pid %ld, %s)",
385 (void *)td, (long)p->p_pid, p->p_comm);
386
387 mtx_unlock_spin(&sc->sc_lock);
388
389 /* See if there are any pending signals for this thread. */
390 PROC_LOCK(p);
391 ps = p->p_sigacts;
392 mtx_lock(&ps->ps_mtx);
393 sig = cursig(td);
394 if (sig == 0) {
395 mtx_unlock(&ps->ps_mtx);
396 ret = thread_suspend_check(1);
397 MPASS(ret == 0 || ret == EINTR || ret == ERESTART);
398 } else {
399 if (SIGISMEMBER(ps->ps_sigintr, sig))
400 ret = EINTR;
401 else
402 ret = ERESTART;
403 mtx_unlock(&ps->ps_mtx);
404 }
405 /*
406 * Lock the per-process spinlock prior to dropping the PROC_LOCK
407 * to avoid a signal delivery race. PROC_LOCK, PROC_SLOCK, and
408 * thread_lock() are currently held in tdsignal().
409 */
410 PROC_SLOCK(p);
411 mtx_lock_spin(&sc->sc_lock);
412 PROC_UNLOCK(p);
413 thread_lock(td);
414 PROC_SUNLOCK(p);
415 if (ret == 0) {
416 if (!(td->td_flags & TDF_INTERRUPT)) {
417 sleepq_switch(wchan);
418 return (0);
419 }
420 /* KSE threads tried unblocking us. */
421 ret = td->td_intrval;
422 MPASS(ret == EINTR || ret == ERESTART || ret == EWOULDBLOCK);
423 }
424 /*
425 * There were pending signals and this thread is still
426 * on the sleep queue, remove it from the sleep queue.
427 */
428 if (TD_ON_SLEEPQ(td)) {
429 sq = sleepq_lookup(wchan);
430 if (sleepq_resume_thread(sq, td, -1)) {
431 #ifdef INVARIANTS
432 /*
433 * This thread hasn't gone to sleep yet, so it
434 * should not be swapped out.
435 */
436 panic("not waking up swapper");
437 #endif
438 }
439 }
440 mtx_unlock_spin(&sc->sc_lock);
441 MPASS(td->td_lock != &sc->sc_lock);
442 return (ret);
443 }
444
445 /*
446 * Switches to another thread if we are still asleep on a sleep queue.
447 * Returns with thread lock.
448 */
449 static void
450 sleepq_switch(void *wchan)
451 {
452 struct sleepqueue_chain *sc;
453 struct sleepqueue *sq;
454 struct thread *td;
455
456 td = curthread;
457 sc = SC_LOOKUP(wchan);
458 mtx_assert(&sc->sc_lock, MA_OWNED);
459 THREAD_LOCK_ASSERT(td, MA_OWNED);
460
461 /*
462 * If we have a sleep queue, then we've already been woken up, so
463 * just return.
464 */
465 if (td->td_sleepqueue != NULL) {
466 mtx_unlock_spin(&sc->sc_lock);
467 return;
468 }
469
470 /*
471 * If TDF_TIMEOUT is set, then our sleep has been timed out
472 * already but we are still on the sleep queue, so dequeue the
473 * thread and return.
474 */
475 if (td->td_flags & TDF_TIMEOUT) {
476 MPASS(TD_ON_SLEEPQ(td));
477 sq = sleepq_lookup(wchan);
478 if (sleepq_resume_thread(sq, td, -1)) {
479 #ifdef INVARIANTS
480 /*
481 * This thread hasn't gone to sleep yet, so it
482 * should not be swapped out.
483 */
484 panic("not waking up swapper");
485 #endif
486 }
487 mtx_unlock_spin(&sc->sc_lock);
488 return;
489 }
490
491 thread_lock_set(td, &sc->sc_lock);
492
493 MPASS(td->td_sleepqueue == NULL);
494 sched_sleep(td);
495 TD_SET_SLEEPING(td);
496 SCHED_STAT_INC(switch_sleepq);
497 mi_switch(SW_VOL, NULL);
498 KASSERT(TD_IS_RUNNING(td), ("running but not TDS_RUNNING"));
499 CTR3(KTR_PROC, "sleepq resume: thread %p (pid %ld, %s)",
500 (void *)td, (long)td->td_proc->p_pid, (void *)td->td_proc->p_comm);
501 }
502
503 /*
504 * Check to see if we timed out.
505 */
506 static int
507 sleepq_check_timeout(void)
508 {
509 struct thread *td;
510
511 td = curthread;
512 THREAD_LOCK_ASSERT(td, MA_OWNED);
513
514 /*
515 * If TDF_TIMEOUT is set, we timed out.
516 */
517 if (td->td_flags & TDF_TIMEOUT) {
518 td->td_flags &= ~TDF_TIMEOUT;
519 return (EWOULDBLOCK);
520 }
521
522 /*
523 * If TDF_TIMOFAIL is set, the timeout ran after we had
524 * already been woken up.
525 */
526 if (td->td_flags & TDF_TIMOFAIL)
527 td->td_flags &= ~TDF_TIMOFAIL;
528
529 /*
530 * If callout_stop() fails, then the timeout is running on
531 * another CPU, so synchronize with it to avoid having it
532 * accidentally wake up a subsequent sleep.
533 */
534 else if (callout_stop(&td->td_slpcallout) == 0) {
535 td->td_flags |= TDF_TIMEOUT;
536 TD_SET_SLEEPING(td);
537 SCHED_STAT_INC(switch_sleepqtimo);
538 mi_switch(SW_INVOL, NULL);
539 }
540 return (0);
541 }
542
543 /*
544 * Check to see if we were awoken by a signal.
545 */
546 static int
547 sleepq_check_signals(void)
548 {
549 struct thread *td;
550
551 td = curthread;
552 THREAD_LOCK_ASSERT(td, MA_OWNED);
553
554 /* We are no longer in an interruptible sleep. */
555 if (td->td_flags & TDF_SINTR)
556 td->td_flags &= ~TDF_SINTR;
557
558 if (td->td_flags & TDF_SLEEPABORT) {
559 td->td_flags &= ~TDF_SLEEPABORT;
560 return (td->td_intrval);
561 }
562
563 if (td->td_flags & TDF_INTERRUPT)
564 return (td->td_intrval);
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)
574 {
575 struct thread *td;
576
577 td = curthread;
578 MPASS(!(td->td_flags & TDF_SINTR));
579 thread_lock(td);
580 sleepq_switch(wchan);
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)
590 {
591 int rcatch;
592 int rval;
593
594 rcatch = sleepq_catch_signals(wchan);
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)
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);
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)
628 {
629 int rcatch, rvalt, rvals;
630
631 rcatch = sleepq_catch_signals(wchan);
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 /*
684 * Note that thread td might not be sleeping if it is running
685 * sleepq_catch_signals() on another CPU or is blocked on
686 * its proc lock to check signals. It doesn't hurt to clear
687 * the sleeping flag if it isn't set though, so we just always
688 * do it. However, we can't assert that it is set.
689 */
690 CTR3(KTR_PROC, "sleepq_wakeup: thread %p (pid %ld, %s)",
691 (void *)td, (long)td->td_proc->p_pid, td->td_proc->p_comm);
692 TD_CLR_SLEEPING(td);
693
694 /* Adjust priority if requested. */
695 MPASS(pri == -1 || (pri >= PRI_MIN && pri <= PRI_MAX));
696 if (pri != -1 && td->td_priority > pri)
697 sched_prio(td, pri);
698 return (setrunnable(td));
699 }
700
701 #ifdef INVARIANTS
702 /*
703 * UMA zone item deallocator.
704 */
705 static void
706 sleepq_dtor(void *mem, int size, void *arg)
707 {
708 struct sleepqueue *sq;
709 int i;
710
711 sq = mem;
712 for (i = 0; i < NR_SLEEPQS; i++)
713 MPASS(TAILQ_EMPTY(&sq->sq_blocked[i]));
714 }
715 #endif
716
717 /*
718 * UMA zone item initializer.
719 */
720 static int
721 sleepq_init(void *mem, int size, int flags)
722 {
723 struct sleepqueue *sq;
724 int i;
725
726 bzero(mem, size);
727 sq = mem;
728 for (i = 0; i < NR_SLEEPQS; i++)
729 TAILQ_INIT(&sq->sq_blocked[i]);
730 LIST_INIT(&sq->sq_free);
731 return (0);
732 }
733
734 /*
735 * Find the highest priority thread sleeping on a wait channel and resume it.
736 */
737 int
738 sleepq_signal(void *wchan, int flags, int pri, int queue)
739 {
740 struct sleepqueue *sq;
741 struct thread *td, *besttd;
742 int wakeup_swapper;
743
744 CTR2(KTR_PROC, "sleepq_signal(%p, %d)", wchan, flags);
745 KASSERT(wchan != NULL, ("%s: invalid NULL wait channel", __func__));
746 MPASS((queue >= 0) && (queue < NR_SLEEPQS));
747 sq = sleepq_lookup(wchan);
748 if (sq == NULL)
749 return (0);
750 KASSERT(sq->sq_type == (flags & SLEEPQ_TYPE),
751 ("%s: mismatch between sleep/wakeup and cv_*", __func__));
752
753 /*
754 * Find the highest priority thread on the queue. If there is a
755 * tie, use the thread that first appears in the queue as it has
756 * been sleeping the longest since threads are always added to
757 * the tail of sleep queues.
758 */
759 besttd = NULL;
760 TAILQ_FOREACH(td, &sq->sq_blocked[queue], td_slpq) {
761 if (besttd == NULL || td->td_priority < besttd->td_priority)
762 besttd = td;
763 }
764 MPASS(besttd != NULL);
765 thread_lock(besttd);
766 wakeup_swapper = sleepq_resume_thread(sq, besttd, pri);
767 thread_unlock(besttd);
768 return (wakeup_swapper);
769 }
770
771 /*
772 * Resume all threads sleeping on a specified wait channel.
773 */
774 int
775 sleepq_broadcast(void *wchan, int flags, int pri, int queue)
776 {
777 struct sleepqueue *sq;
778 struct thread *td, *tdn;
779 int wakeup_swapper;
780
781 CTR2(KTR_PROC, "sleepq_broadcast(%p, %d)", wchan, flags);
782 KASSERT(wchan != NULL, ("%s: invalid NULL wait channel", __func__));
783 MPASS((queue >= 0) && (queue < NR_SLEEPQS));
784 sq = sleepq_lookup(wchan);
785 if (sq == NULL) {
786 sleepq_release(wchan);
787 return (0);
788 }
789 KASSERT(sq->sq_type == (flags & SLEEPQ_TYPE),
790 ("%s: mismatch between sleep/wakeup and cv_*", __func__));
791
792 /* Resume all blocked threads on the sleep queue. */
793 wakeup_swapper = 0;
794 TAILQ_FOREACH_SAFE(td, &sq->sq_blocked[queue], td_slpq, tdn) {
795 thread_lock(td);
796 if (sleepq_resume_thread(sq, td, pri))
797 wakeup_swapper = 1;
798 thread_unlock(td);
799 }
800 sleepq_release(wchan);
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_proc->p_comm);
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 MPASS(td->td_lock == &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, -1);
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, -1);
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_proc->p_comm);
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, -1));
953 }
954
955 #ifdef DDB
956 DB_SHOW_COMMAND(sleepq, db_show_sleepqueue)
957 {
958 struct sleepqueue_chain *sc;
959 struct sleepqueue *sq;
960 #ifdef INVARIANTS
961 struct lock_object *lock;
962 #endif
963 struct thread *td;
964 void *wchan;
965 int i;
966
967 if (!have_addr)
968 return;
969
970 /*
971 * First, see if there is an active sleep queue for the wait channel
972 * indicated by the address.
973 */
974 wchan = (void *)addr;
975 sc = SC_LOOKUP(wchan);
976 LIST_FOREACH(sq, &sc->sc_queues, sq_hash)
977 if (sq->sq_wchan == wchan)
978 goto found;
979
980 /*
981 * Second, see if there is an active sleep queue at the address
982 * indicated.
983 */
984 for (i = 0; i < SC_TABLESIZE; i++)
985 LIST_FOREACH(sq, &sleepq_chains[i].sc_queues, sq_hash) {
986 if (sq == (struct sleepqueue *)addr)
987 goto found;
988 }
989
990 db_printf("Unable to locate a sleep queue via %p\n", (void *)addr);
991 return;
992 found:
993 db_printf("Wait channel: %p\n", sq->sq_wchan);
994 #ifdef INVARIANTS
995 db_printf("Queue type: %d\n", sq->sq_type);
996 if (sq->sq_lock) {
997 lock = sq->sq_lock;
998 db_printf("Associated Interlock: %p - (%s) %s\n", lock,
999 LOCK_CLASS(lock)->lc_name, lock->lo_name);
1000 }
1001 #endif
1002 db_printf("Blocked threads:\n");
1003 for (i = 0; i < NR_SLEEPQS; i++) {
1004 db_printf("\nQueue[%d]:\n", i);
1005 if (TAILQ_EMPTY(&sq->sq_blocked[i]))
1006 db_printf("\tempty\n");
1007 else
1008 TAILQ_FOREACH(td, &sq->sq_blocked[0],
1009 td_slpq) {
1010 db_printf("\t%p (tid %d, pid %d, \"%s\")\n", td,
1011 td->td_tid, td->td_proc->p_pid,
1012 td->td_name[i] != '\0' ? td->td_name :
1013 td->td_proc->p_comm);
1014 }
1015 }
1016 }
1017
1018 /* Alias 'show sleepqueue' to 'show sleepq'. */
1019 DB_SET(sleepqueue, db_show_sleepqueue, db_show_cmd_set, 0, NULL);
1020 #endif
Cache object: c14dcbd30eab48fed9af23a1a79df68f
|