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: releng/9.2/sys/kern/subr_sleepqueue.c 251147 2013-05-30 19:14:34Z jhb $");
64
65 #include "opt_sleepqueue_profiling.h"
66 #include "opt_ddb.h"
67 #include "opt_kdtrace.h"
68 #include "opt_sched.h"
69
70 #include <sys/param.h>
71 #include <sys/systm.h>
72 #include <sys/lock.h>
73 #include <sys/kernel.h>
74 #include <sys/ktr.h>
75 #include <sys/mutex.h>
76 #include <sys/proc.h>
77 #include <sys/sbuf.h>
78 #include <sys/sched.h>
79 #include <sys/sdt.h>
80 #include <sys/signalvar.h>
81 #include <sys/sleepqueue.h>
82 #include <sys/sysctl.h>
83
84 #include <vm/uma.h>
85
86 #ifdef DDB
87 #include <ddb/ddb.h>
88 #endif
89
90 /*
91 * Constants for the hash table of sleep queue chains. These constants are
92 * the same ones that 4BSD (and possibly earlier versions of BSD) used.
93 * Basically, we ignore the lower 8 bits of the address since most wait
94 * channel pointers are aligned and only look at the next 7 bits for the
95 * hash. SC_TABLESIZE must be a power of two for SC_MASK to work properly.
96 */
97 #define SC_TABLESIZE 128 /* Must be power of 2. */
98 #define SC_MASK (SC_TABLESIZE - 1)
99 #define SC_SHIFT 8
100 #define SC_HASH(wc) (((uintptr_t)(wc) >> SC_SHIFT) & SC_MASK)
101 #define SC_LOOKUP(wc) &sleepq_chains[SC_HASH(wc)]
102 #define NR_SLEEPQS 2
103 /*
104 * There two different lists of sleep queues. Both lists are connected
105 * via the sq_hash entries. The first list is the sleep queue chain list
106 * that a sleep queue is on when it is attached to a wait channel. The
107 * second list is the free list hung off of a sleep queue that is attached
108 * to a wait channel.
109 *
110 * Each sleep queue also contains the wait channel it is attached to, the
111 * list of threads blocked on that wait channel, flags specific to the
112 * wait channel, and the lock used to synchronize with a wait channel.
113 * The flags are used to catch mismatches between the various consumers
114 * of the sleep queue API (e.g. sleep/wakeup and condition variables).
115 * The lock pointer is only used when invariants are enabled for various
116 * debugging checks.
117 *
118 * Locking key:
119 * c - sleep queue chain lock
120 */
121 struct sleepqueue {
122 TAILQ_HEAD(, thread) sq_blocked[NR_SLEEPQS]; /* (c) Blocked threads. */
123 u_int sq_blockedcnt[NR_SLEEPQS]; /* (c) N. of blocked threads. */
124 LIST_ENTRY(sleepqueue) sq_hash; /* (c) Chain and free list. */
125 LIST_HEAD(, sleepqueue) sq_free; /* (c) Free queues. */
126 void *sq_wchan; /* (c) Wait channel. */
127 int sq_type; /* (c) Queue type. */
128 #ifdef INVARIANTS
129 struct lock_object *sq_lock; /* (c) Associated lock. */
130 #endif
131 };
132
133 struct sleepqueue_chain {
134 LIST_HEAD(, sleepqueue) sc_queues; /* List of sleep queues. */
135 struct mtx sc_lock; /* Spin lock for this chain. */
136 #ifdef SLEEPQUEUE_PROFILING
137 u_int sc_depth; /* Length of sc_queues. */
138 u_int sc_max_depth; /* Max length of sc_queues. */
139 #endif
140 };
141
142 #ifdef SLEEPQUEUE_PROFILING
143 u_int sleepq_max_depth;
144 static SYSCTL_NODE(_debug, OID_AUTO, sleepq, CTLFLAG_RD, 0, "sleepq profiling");
145 static SYSCTL_NODE(_debug_sleepq, OID_AUTO, chains, CTLFLAG_RD, 0,
146 "sleepq chain stats");
147 SYSCTL_UINT(_debug_sleepq, OID_AUTO, max_depth, CTLFLAG_RD, &sleepq_max_depth,
148 0, "maxmimum depth achieved of a single chain");
149
150 static void sleepq_profile(const char *wmesg);
151 static int prof_enabled;
152 #endif
153 static struct sleepqueue_chain sleepq_chains[SC_TABLESIZE];
154 static uma_zone_t sleepq_zone;
155
156 /*
157 * Prototypes for non-exported routines.
158 */
159 static int sleepq_catch_signals(void *wchan, int pri);
160 static int sleepq_check_signals(void);
161 static int sleepq_check_timeout(void);
162 #ifdef INVARIANTS
163 static void sleepq_dtor(void *mem, int size, void *arg);
164 #endif
165 static int sleepq_init(void *mem, int size, int flags);
166 static int sleepq_resume_thread(struct sleepqueue *sq, struct thread *td,
167 int pri);
168 static void sleepq_switch(void *wchan, int pri);
169 static void sleepq_timeout(void *arg);
170
171 SDT_PROBE_DECLARE(sched, , , sleep);
172 SDT_PROBE_DECLARE(sched, , , wakeup);
173
174 /*
175 * Early initialization of sleep queues that is called from the sleepinit()
176 * SYSINIT.
177 */
178 void
179 init_sleepqueues(void)
180 {
181 #ifdef SLEEPQUEUE_PROFILING
182 struct sysctl_oid *chain_oid;
183 char chain_name[10];
184 #endif
185 int i;
186
187 for (i = 0; i < SC_TABLESIZE; i++) {
188 LIST_INIT(&sleepq_chains[i].sc_queues);
189 mtx_init(&sleepq_chains[i].sc_lock, "sleepq chain", NULL,
190 MTX_SPIN | MTX_RECURSE);
191 #ifdef SLEEPQUEUE_PROFILING
192 snprintf(chain_name, sizeof(chain_name), "%d", i);
193 chain_oid = SYSCTL_ADD_NODE(NULL,
194 SYSCTL_STATIC_CHILDREN(_debug_sleepq_chains), OID_AUTO,
195 chain_name, CTLFLAG_RD, NULL, "sleepq chain stats");
196 SYSCTL_ADD_UINT(NULL, SYSCTL_CHILDREN(chain_oid), OID_AUTO,
197 "depth", CTLFLAG_RD, &sleepq_chains[i].sc_depth, 0, NULL);
198 SYSCTL_ADD_UINT(NULL, SYSCTL_CHILDREN(chain_oid), OID_AUTO,
199 "max_depth", CTLFLAG_RD, &sleepq_chains[i].sc_max_depth, 0,
200 NULL);
201 #endif
202 }
203 sleepq_zone = uma_zcreate("SLEEPQUEUE", sizeof(struct sleepqueue),
204 #ifdef INVARIANTS
205 NULL, sleepq_dtor, sleepq_init, NULL, UMA_ALIGN_CACHE, 0);
206 #else
207 NULL, NULL, sleepq_init, NULL, UMA_ALIGN_CACHE, 0);
208 #endif
209
210 thread0.td_sleepqueue = sleepq_alloc();
211 }
212
213 /*
214 * Get a sleep queue for a new thread.
215 */
216 struct sleepqueue *
217 sleepq_alloc(void)
218 {
219
220 return (uma_zalloc(sleepq_zone, M_WAITOK));
221 }
222
223 /*
224 * Free a sleep queue when a thread is destroyed.
225 */
226 void
227 sleepq_free(struct sleepqueue *sq)
228 {
229
230 uma_zfree(sleepq_zone, sq);
231 }
232
233 /*
234 * Lock the sleep queue chain associated with the specified wait channel.
235 */
236 void
237 sleepq_lock(void *wchan)
238 {
239 struct sleepqueue_chain *sc;
240
241 sc = SC_LOOKUP(wchan);
242 mtx_lock_spin(&sc->sc_lock);
243 }
244
245 /*
246 * Look up the sleep queue associated with a given wait channel in the hash
247 * table locking the associated sleep queue chain. If no queue is found in
248 * the table, NULL is returned.
249 */
250 struct sleepqueue *
251 sleepq_lookup(void *wchan)
252 {
253 struct sleepqueue_chain *sc;
254 struct sleepqueue *sq;
255
256 KASSERT(wchan != NULL, ("%s: invalid NULL wait channel", __func__));
257 sc = SC_LOOKUP(wchan);
258 mtx_assert(&sc->sc_lock, MA_OWNED);
259 LIST_FOREACH(sq, &sc->sc_queues, sq_hash)
260 if (sq->sq_wchan == wchan)
261 return (sq);
262 return (NULL);
263 }
264
265 /*
266 * Unlock the sleep queue chain associated with a given wait channel.
267 */
268 void
269 sleepq_release(void *wchan)
270 {
271 struct sleepqueue_chain *sc;
272
273 sc = SC_LOOKUP(wchan);
274 mtx_unlock_spin(&sc->sc_lock);
275 }
276
277 /*
278 * Places the current thread on the sleep queue for the specified wait
279 * channel. If INVARIANTS is enabled, then it associates the passed in
280 * lock with the sleepq to make sure it is held when that sleep queue is
281 * woken up.
282 */
283 void
284 sleepq_add(void *wchan, struct lock_object *lock, const char *wmesg, int flags,
285 int queue)
286 {
287 struct sleepqueue_chain *sc;
288 struct sleepqueue *sq;
289 struct thread *td;
290
291 td = curthread;
292 sc = SC_LOOKUP(wchan);
293 mtx_assert(&sc->sc_lock, MA_OWNED);
294 MPASS(td->td_sleepqueue != NULL);
295 MPASS(wchan != NULL);
296 MPASS((queue >= 0) && (queue < NR_SLEEPQS));
297
298 /* If this thread is not allowed to sleep, die a horrible death. */
299 KASSERT(!(td->td_pflags & TDP_NOSLEEPING),
300 ("Trying sleep, but thread marked as sleeping prohibited"));
301
302 /* Look up the sleep queue associated with the wait channel 'wchan'. */
303 sq = sleepq_lookup(wchan);
304
305 /*
306 * If the wait channel does not already have a sleep queue, use
307 * this thread's sleep queue. Otherwise, insert the current thread
308 * into the sleep queue already in use by this wait channel.
309 */
310 if (sq == NULL) {
311 #ifdef INVARIANTS
312 int i;
313
314 sq = td->td_sleepqueue;
315 for (i = 0; i < NR_SLEEPQS; i++) {
316 KASSERT(TAILQ_EMPTY(&sq->sq_blocked[i]),
317 ("thread's sleep queue %d is not empty", i));
318 KASSERT(sq->sq_blockedcnt[i] == 0,
319 ("thread's sleep queue %d count mismatches", i));
320 }
321 KASSERT(LIST_EMPTY(&sq->sq_free),
322 ("thread's sleep queue has a non-empty free list"));
323 KASSERT(sq->sq_wchan == NULL, ("stale sq_wchan pointer"));
324 sq->sq_lock = lock;
325 #endif
326 #ifdef SLEEPQUEUE_PROFILING
327 sc->sc_depth++;
328 if (sc->sc_depth > sc->sc_max_depth) {
329 sc->sc_max_depth = sc->sc_depth;
330 if (sc->sc_max_depth > sleepq_max_depth)
331 sleepq_max_depth = sc->sc_max_depth;
332 }
333 #endif
334 sq = td->td_sleepqueue;
335 LIST_INSERT_HEAD(&sc->sc_queues, sq, sq_hash);
336 sq->sq_wchan = wchan;
337 sq->sq_type = flags & SLEEPQ_TYPE;
338 } else {
339 MPASS(wchan == sq->sq_wchan);
340 MPASS(lock == sq->sq_lock);
341 MPASS((flags & SLEEPQ_TYPE) == sq->sq_type);
342 LIST_INSERT_HEAD(&sq->sq_free, td->td_sleepqueue, sq_hash);
343 }
344 thread_lock(td);
345 TAILQ_INSERT_TAIL(&sq->sq_blocked[queue], td, td_slpq);
346 sq->sq_blockedcnt[queue]++;
347 td->td_sleepqueue = NULL;
348 td->td_sqqueue = queue;
349 td->td_wchan = wchan;
350 td->td_wmesg = wmesg;
351 if (flags & SLEEPQ_INTERRUPTIBLE) {
352 td->td_flags |= TDF_SINTR;
353 td->td_flags &= ~TDF_SLEEPABORT;
354 }
355 thread_unlock(td);
356 }
357
358 /*
359 * Sets a timeout that will remove the current thread from the specified
360 * sleep queue after timo ticks if the thread has not already been awakened.
361 */
362 void
363 sleepq_set_timeout(void *wchan, int timo)
364 {
365 struct sleepqueue_chain *sc;
366 struct thread *td;
367
368 td = curthread;
369 sc = SC_LOOKUP(wchan);
370 mtx_assert(&sc->sc_lock, MA_OWNED);
371 MPASS(TD_ON_SLEEPQ(td));
372 MPASS(td->td_sleepqueue == NULL);
373 MPASS(wchan != NULL);
374 callout_reset_curcpu(&td->td_slpcallout, timo, sleepq_timeout, td);
375 }
376
377 /*
378 * Return the number of actual sleepers for the specified queue.
379 */
380 u_int
381 sleepq_sleepcnt(void *wchan, int queue)
382 {
383 struct sleepqueue *sq;
384
385 KASSERT(wchan != NULL, ("%s: invalid NULL wait channel", __func__));
386 MPASS((queue >= 0) && (queue < NR_SLEEPQS));
387 sq = sleepq_lookup(wchan);
388 if (sq == NULL)
389 return (0);
390 return (sq->sq_blockedcnt[queue]);
391 }
392
393 /*
394 * Marks the pending sleep of the current thread as interruptible and
395 * makes an initial check for pending signals before putting a thread
396 * to sleep. Enters and exits with the thread lock held. Thread lock
397 * may have transitioned from the sleepq lock to a run lock.
398 */
399 static int
400 sleepq_catch_signals(void *wchan, int pri)
401 {
402 struct sleepqueue_chain *sc;
403 struct sleepqueue *sq;
404 struct thread *td;
405 struct proc *p;
406 struct sigacts *ps;
407 int sig, ret, stop_allowed;
408
409 td = curthread;
410 p = curproc;
411 sc = SC_LOOKUP(wchan);
412 mtx_assert(&sc->sc_lock, MA_OWNED);
413 MPASS(wchan != NULL);
414 if ((td->td_pflags & TDP_WAKEUP) != 0) {
415 td->td_pflags &= ~TDP_WAKEUP;
416 ret = EINTR;
417 thread_lock(td);
418 goto out;
419 }
420
421 /*
422 * See if there are any pending signals for this thread. If not
423 * we can switch immediately. Otherwise do the signal processing
424 * directly.
425 */
426 thread_lock(td);
427 if ((td->td_flags & (TDF_NEEDSIGCHK | TDF_NEEDSUSPCHK)) == 0) {
428 sleepq_switch(wchan, pri);
429 return (0);
430 }
431 stop_allowed = (td->td_flags & TDF_SBDRY) ? SIG_STOP_NOT_ALLOWED :
432 SIG_STOP_ALLOWED;
433 thread_unlock(td);
434 mtx_unlock_spin(&sc->sc_lock);
435 CTR3(KTR_PROC, "sleepq catching signals: thread %p (pid %ld, %s)",
436 (void *)td, (long)p->p_pid, td->td_name);
437 PROC_LOCK(p);
438 ps = p->p_sigacts;
439 mtx_lock(&ps->ps_mtx);
440 sig = cursig(td, stop_allowed);
441 if (sig == 0) {
442 mtx_unlock(&ps->ps_mtx);
443 ret = thread_suspend_check(1);
444 MPASS(ret == 0 || ret == EINTR || ret == ERESTART);
445 } else {
446 if (SIGISMEMBER(ps->ps_sigintr, sig))
447 ret = EINTR;
448 else
449 ret = ERESTART;
450 mtx_unlock(&ps->ps_mtx);
451 }
452 /*
453 * Lock the per-process spinlock prior to dropping the PROC_LOCK
454 * to avoid a signal delivery race. PROC_LOCK, PROC_SLOCK, and
455 * thread_lock() are currently held in tdsendsignal().
456 */
457 PROC_SLOCK(p);
458 mtx_lock_spin(&sc->sc_lock);
459 PROC_UNLOCK(p);
460 thread_lock(td);
461 PROC_SUNLOCK(p);
462 if (ret == 0) {
463 sleepq_switch(wchan, pri);
464 return (0);
465 }
466 out:
467 /*
468 * There were pending signals and this thread is still
469 * on the sleep queue, remove it from the sleep queue.
470 */
471 if (TD_ON_SLEEPQ(td)) {
472 sq = sleepq_lookup(wchan);
473 if (sleepq_resume_thread(sq, td, 0)) {
474 #ifdef INVARIANTS
475 /*
476 * This thread hasn't gone to sleep yet, so it
477 * should not be swapped out.
478 */
479 panic("not waking up swapper");
480 #endif
481 }
482 }
483 mtx_unlock_spin(&sc->sc_lock);
484 MPASS(td->td_lock != &sc->sc_lock);
485 return (ret);
486 }
487
488 /*
489 * Switches to another thread if we are still asleep on a sleep queue.
490 * Returns with thread lock.
491 */
492 static void
493 sleepq_switch(void *wchan, int pri)
494 {
495 struct sleepqueue_chain *sc;
496 struct sleepqueue *sq;
497 struct thread *td;
498
499 td = curthread;
500 sc = SC_LOOKUP(wchan);
501 mtx_assert(&sc->sc_lock, MA_OWNED);
502 THREAD_LOCK_ASSERT(td, MA_OWNED);
503
504 /*
505 * If we have a sleep queue, then we've already been woken up, so
506 * just return.
507 */
508 if (td->td_sleepqueue != NULL) {
509 mtx_unlock_spin(&sc->sc_lock);
510 return;
511 }
512
513 /*
514 * If TDF_TIMEOUT is set, then our sleep has been timed out
515 * already but we are still on the sleep queue, so dequeue the
516 * thread and return.
517 */
518 if (td->td_flags & TDF_TIMEOUT) {
519 MPASS(TD_ON_SLEEPQ(td));
520 sq = sleepq_lookup(wchan);
521 if (sleepq_resume_thread(sq, td, 0)) {
522 #ifdef INVARIANTS
523 /*
524 * This thread hasn't gone to sleep yet, so it
525 * should not be swapped out.
526 */
527 panic("not waking up swapper");
528 #endif
529 }
530 mtx_unlock_spin(&sc->sc_lock);
531 return;
532 }
533 #ifdef SLEEPQUEUE_PROFILING
534 if (prof_enabled)
535 sleepq_profile(td->td_wmesg);
536 #endif
537 MPASS(td->td_sleepqueue == NULL);
538 sched_sleep(td, pri);
539 thread_lock_set(td, &sc->sc_lock);
540 SDT_PROBE0(sched, , , sleep);
541 TD_SET_SLEEPING(td);
542 mi_switch(SW_VOL | SWT_SLEEPQ, NULL);
543 KASSERT(TD_IS_RUNNING(td), ("running but not TDS_RUNNING"));
544 CTR3(KTR_PROC, "sleepq resume: thread %p (pid %ld, %s)",
545 (void *)td, (long)td->td_proc->p_pid, (void *)td->td_name);
546 }
547
548 /*
549 * Check to see if we timed out.
550 */
551 static int
552 sleepq_check_timeout(void)
553 {
554 struct thread *td;
555
556 td = curthread;
557 THREAD_LOCK_ASSERT(td, MA_OWNED);
558
559 /*
560 * If TDF_TIMEOUT is set, we timed out.
561 */
562 if (td->td_flags & TDF_TIMEOUT) {
563 td->td_flags &= ~TDF_TIMEOUT;
564 return (EWOULDBLOCK);
565 }
566
567 /*
568 * If TDF_TIMOFAIL is set, the timeout ran after we had
569 * already been woken up.
570 */
571 if (td->td_flags & TDF_TIMOFAIL)
572 td->td_flags &= ~TDF_TIMOFAIL;
573
574 /*
575 * If callout_stop() fails, then the timeout is running on
576 * another CPU, so synchronize with it to avoid having it
577 * accidentally wake up a subsequent sleep.
578 */
579 else if (callout_stop(&td->td_slpcallout) == 0) {
580 td->td_flags |= TDF_TIMEOUT;
581 TD_SET_SLEEPING(td);
582 mi_switch(SW_INVOL | SWT_SLEEPQTIMO, NULL);
583 }
584 return (0);
585 }
586
587 /*
588 * Check to see if we were awoken by a signal.
589 */
590 static int
591 sleepq_check_signals(void)
592 {
593 struct thread *td;
594
595 td = curthread;
596 THREAD_LOCK_ASSERT(td, MA_OWNED);
597
598 /* We are no longer in an interruptible sleep. */
599 if (td->td_flags & TDF_SINTR)
600 td->td_flags &= ~TDF_SINTR;
601
602 if (td->td_flags & TDF_SLEEPABORT) {
603 td->td_flags &= ~TDF_SLEEPABORT;
604 return (td->td_intrval);
605 }
606
607 return (0);
608 }
609
610 /*
611 * Block the current thread until it is awakened from its sleep queue.
612 */
613 void
614 sleepq_wait(void *wchan, int pri)
615 {
616 struct thread *td;
617
618 td = curthread;
619 MPASS(!(td->td_flags & TDF_SINTR));
620 thread_lock(td);
621 sleepq_switch(wchan, pri);
622 thread_unlock(td);
623 }
624
625 /*
626 * Block the current thread until it is awakened from its sleep queue
627 * or it is interrupted by a signal.
628 */
629 int
630 sleepq_wait_sig(void *wchan, int pri)
631 {
632 int rcatch;
633 int rval;
634
635 rcatch = sleepq_catch_signals(wchan, pri);
636 rval = sleepq_check_signals();
637 thread_unlock(curthread);
638 if (rcatch)
639 return (rcatch);
640 return (rval);
641 }
642
643 /*
644 * Block the current thread until it is awakened from its sleep queue
645 * or it times out while waiting.
646 */
647 int
648 sleepq_timedwait(void *wchan, int pri)
649 {
650 struct thread *td;
651 int rval;
652
653 td = curthread;
654 MPASS(!(td->td_flags & TDF_SINTR));
655 thread_lock(td);
656 sleepq_switch(wchan, pri);
657 rval = sleepq_check_timeout();
658 thread_unlock(td);
659
660 return (rval);
661 }
662
663 /*
664 * Block the current thread until it is awakened from its sleep queue,
665 * it is interrupted by a signal, or it times out waiting to be awakened.
666 */
667 int
668 sleepq_timedwait_sig(void *wchan, int pri)
669 {
670 int rcatch, rvalt, rvals;
671
672 rcatch = sleepq_catch_signals(wchan, pri);
673 rvalt = sleepq_check_timeout();
674 rvals = sleepq_check_signals();
675 thread_unlock(curthread);
676 if (rcatch)
677 return (rcatch);
678 if (rvals)
679 return (rvals);
680 return (rvalt);
681 }
682
683 /*
684 * Returns the type of sleepqueue given a waitchannel.
685 */
686 int
687 sleepq_type(void *wchan)
688 {
689 struct sleepqueue *sq;
690 int type;
691
692 MPASS(wchan != NULL);
693
694 sleepq_lock(wchan);
695 sq = sleepq_lookup(wchan);
696 if (sq == NULL) {
697 sleepq_release(wchan);
698 return (-1);
699 }
700 type = sq->sq_type;
701 sleepq_release(wchan);
702 return (type);
703 }
704
705 /*
706 * Removes a thread from a sleep queue and makes it
707 * runnable.
708 */
709 static int
710 sleepq_resume_thread(struct sleepqueue *sq, struct thread *td, int pri)
711 {
712 struct sleepqueue_chain *sc;
713
714 MPASS(td != NULL);
715 MPASS(sq->sq_wchan != NULL);
716 MPASS(td->td_wchan == sq->sq_wchan);
717 MPASS(td->td_sqqueue < NR_SLEEPQS && td->td_sqqueue >= 0);
718 THREAD_LOCK_ASSERT(td, MA_OWNED);
719 sc = SC_LOOKUP(sq->sq_wchan);
720 mtx_assert(&sc->sc_lock, MA_OWNED);
721
722 SDT_PROBE2(sched, , , wakeup, td, td->td_proc);
723
724 /* Remove the thread from the queue. */
725 sq->sq_blockedcnt[td->td_sqqueue]--;
726 TAILQ_REMOVE(&sq->sq_blocked[td->td_sqqueue], td, td_slpq);
727
728 /*
729 * Get a sleep queue for this thread. If this is the last waiter,
730 * use the queue itself and take it out of the chain, otherwise,
731 * remove a queue from the free list.
732 */
733 if (LIST_EMPTY(&sq->sq_free)) {
734 td->td_sleepqueue = sq;
735 #ifdef INVARIANTS
736 sq->sq_wchan = NULL;
737 #endif
738 #ifdef SLEEPQUEUE_PROFILING
739 sc->sc_depth--;
740 #endif
741 } else
742 td->td_sleepqueue = LIST_FIRST(&sq->sq_free);
743 LIST_REMOVE(td->td_sleepqueue, sq_hash);
744
745 td->td_wmesg = NULL;
746 td->td_wchan = NULL;
747 td->td_flags &= ~TDF_SINTR;
748
749 CTR3(KTR_PROC, "sleepq_wakeup: thread %p (pid %ld, %s)",
750 (void *)td, (long)td->td_proc->p_pid, td->td_name);
751
752 /* Adjust priority if requested. */
753 MPASS(pri == 0 || (pri >= PRI_MIN && pri <= PRI_MAX));
754 if (pri != 0 && td->td_priority > pri &&
755 PRI_BASE(td->td_pri_class) == PRI_TIMESHARE)
756 sched_prio(td, pri);
757
758 /*
759 * Note that thread td might not be sleeping if it is running
760 * sleepq_catch_signals() on another CPU or is blocked on its
761 * proc lock to check signals. There's no need to mark the
762 * thread runnable in that case.
763 */
764 if (TD_IS_SLEEPING(td)) {
765 TD_CLR_SLEEPING(td);
766 return (setrunnable(td));
767 }
768 return (0);
769 }
770
771 #ifdef INVARIANTS
772 /*
773 * UMA zone item deallocator.
774 */
775 static void
776 sleepq_dtor(void *mem, int size, void *arg)
777 {
778 struct sleepqueue *sq;
779 int i;
780
781 sq = mem;
782 for (i = 0; i < NR_SLEEPQS; i++) {
783 MPASS(TAILQ_EMPTY(&sq->sq_blocked[i]));
784 MPASS(sq->sq_blockedcnt[i] == 0);
785 }
786 }
787 #endif
788
789 /*
790 * UMA zone item initializer.
791 */
792 static int
793 sleepq_init(void *mem, int size, int flags)
794 {
795 struct sleepqueue *sq;
796 int i;
797
798 bzero(mem, size);
799 sq = mem;
800 for (i = 0; i < NR_SLEEPQS; i++) {
801 TAILQ_INIT(&sq->sq_blocked[i]);
802 sq->sq_blockedcnt[i] = 0;
803 }
804 LIST_INIT(&sq->sq_free);
805 return (0);
806 }
807
808 /*
809 * Find the highest priority thread sleeping on a wait channel and resume it.
810 */
811 int
812 sleepq_signal(void *wchan, int flags, int pri, int queue)
813 {
814 struct sleepqueue *sq;
815 struct thread *td, *besttd;
816 int wakeup_swapper;
817
818 CTR2(KTR_PROC, "sleepq_signal(%p, %d)", wchan, flags);
819 KASSERT(wchan != NULL, ("%s: invalid NULL wait channel", __func__));
820 MPASS((queue >= 0) && (queue < NR_SLEEPQS));
821 sq = sleepq_lookup(wchan);
822 if (sq == NULL)
823 return (0);
824 KASSERT(sq->sq_type == (flags & SLEEPQ_TYPE),
825 ("%s: mismatch between sleep/wakeup and cv_*", __func__));
826
827 /*
828 * Find the highest priority thread on the queue. If there is a
829 * tie, use the thread that first appears in the queue as it has
830 * been sleeping the longest since threads are always added to
831 * the tail of sleep queues.
832 */
833 besttd = NULL;
834 TAILQ_FOREACH(td, &sq->sq_blocked[queue], td_slpq) {
835 if (besttd == NULL || td->td_priority < besttd->td_priority)
836 besttd = td;
837 }
838 MPASS(besttd != NULL);
839 thread_lock(besttd);
840 wakeup_swapper = sleepq_resume_thread(sq, besttd, pri);
841 thread_unlock(besttd);
842 return (wakeup_swapper);
843 }
844
845 /*
846 * Resume all threads sleeping on a specified wait channel.
847 */
848 int
849 sleepq_broadcast(void *wchan, int flags, int pri, int queue)
850 {
851 struct sleepqueue *sq;
852 struct thread *td, *tdn;
853 int wakeup_swapper;
854
855 CTR2(KTR_PROC, "sleepq_broadcast(%p, %d)", wchan, flags);
856 KASSERT(wchan != NULL, ("%s: invalid NULL wait channel", __func__));
857 MPASS((queue >= 0) && (queue < NR_SLEEPQS));
858 sq = sleepq_lookup(wchan);
859 if (sq == NULL)
860 return (0);
861 KASSERT(sq->sq_type == (flags & SLEEPQ_TYPE),
862 ("%s: mismatch between sleep/wakeup and cv_*", __func__));
863
864 /* Resume all blocked threads on the sleep queue. */
865 wakeup_swapper = 0;
866 TAILQ_FOREACH_SAFE(td, &sq->sq_blocked[queue], td_slpq, tdn) {
867 thread_lock(td);
868 if (sleepq_resume_thread(sq, td, pri))
869 wakeup_swapper = 1;
870 thread_unlock(td);
871 }
872 return (wakeup_swapper);
873 }
874
875 /*
876 * Time sleeping threads out. When the timeout expires, the thread is
877 * removed from the sleep queue and made runnable if it is still asleep.
878 */
879 static void
880 sleepq_timeout(void *arg)
881 {
882 struct sleepqueue_chain *sc;
883 struct sleepqueue *sq;
884 struct thread *td;
885 void *wchan;
886 int wakeup_swapper;
887
888 td = arg;
889 wakeup_swapper = 0;
890 CTR3(KTR_PROC, "sleepq_timeout: thread %p (pid %ld, %s)",
891 (void *)td, (long)td->td_proc->p_pid, (void *)td->td_name);
892
893 /*
894 * First, see if the thread is asleep and get the wait channel if
895 * it is.
896 */
897 thread_lock(td);
898 if (TD_IS_SLEEPING(td) && TD_ON_SLEEPQ(td)) {
899 wchan = td->td_wchan;
900 sc = SC_LOOKUP(wchan);
901 THREAD_LOCKPTR_ASSERT(td, &sc->sc_lock);
902 sq = sleepq_lookup(wchan);
903 MPASS(sq != NULL);
904 td->td_flags |= TDF_TIMEOUT;
905 wakeup_swapper = sleepq_resume_thread(sq, td, 0);
906 thread_unlock(td);
907 if (wakeup_swapper)
908 kick_proc0();
909 return;
910 }
911
912 /*
913 * If the thread is on the SLEEPQ but isn't sleeping yet, it
914 * can either be on another CPU in between sleepq_add() and
915 * one of the sleepq_*wait*() routines or it can be in
916 * sleepq_catch_signals().
917 */
918 if (TD_ON_SLEEPQ(td)) {
919 td->td_flags |= TDF_TIMEOUT;
920 thread_unlock(td);
921 return;
922 }
923
924 /*
925 * Now check for the edge cases. First, if TDF_TIMEOUT is set,
926 * then the other thread has already yielded to us, so clear
927 * the flag and resume it. If TDF_TIMEOUT is not set, then the
928 * we know that the other thread is not on a sleep queue, but it
929 * hasn't resumed execution yet. In that case, set TDF_TIMOFAIL
930 * to let it know that the timeout has already run and doesn't
931 * need to be canceled.
932 */
933 if (td->td_flags & TDF_TIMEOUT) {
934 MPASS(TD_IS_SLEEPING(td));
935 td->td_flags &= ~TDF_TIMEOUT;
936 TD_CLR_SLEEPING(td);
937 wakeup_swapper = setrunnable(td);
938 } else
939 td->td_flags |= TDF_TIMOFAIL;
940 thread_unlock(td);
941 if (wakeup_swapper)
942 kick_proc0();
943 }
944
945 /*
946 * Resumes a specific thread from the sleep queue associated with a specific
947 * wait channel if it is on that queue.
948 */
949 void
950 sleepq_remove(struct thread *td, void *wchan)
951 {
952 struct sleepqueue *sq;
953 int wakeup_swapper;
954
955 /*
956 * Look up the sleep queue for this wait channel, then re-check
957 * that the thread is asleep on that channel, if it is not, then
958 * bail.
959 */
960 MPASS(wchan != NULL);
961 sleepq_lock(wchan);
962 sq = sleepq_lookup(wchan);
963 /*
964 * We can not lock the thread here as it may be sleeping on a
965 * different sleepq. However, holding the sleepq lock for this
966 * wchan can guarantee that we do not miss a wakeup for this
967 * channel. The asserts below will catch any false positives.
968 */
969 if (!TD_ON_SLEEPQ(td) || td->td_wchan != wchan) {
970 sleepq_release(wchan);
971 return;
972 }
973 /* Thread is asleep on sleep queue sq, so wake it up. */
974 thread_lock(td);
975 MPASS(sq != NULL);
976 MPASS(td->td_wchan == wchan);
977 wakeup_swapper = sleepq_resume_thread(sq, td, 0);
978 thread_unlock(td);
979 sleepq_release(wchan);
980 if (wakeup_swapper)
981 kick_proc0();
982 }
983
984 /*
985 * Abort a thread as if an interrupt had occurred. Only abort
986 * interruptible waits (unfortunately it isn't safe to abort others).
987 */
988 int
989 sleepq_abort(struct thread *td, int intrval)
990 {
991 struct sleepqueue *sq;
992 void *wchan;
993
994 THREAD_LOCK_ASSERT(td, MA_OWNED);
995 MPASS(TD_ON_SLEEPQ(td));
996 MPASS(td->td_flags & TDF_SINTR);
997 MPASS(intrval == EINTR || intrval == ERESTART);
998
999 /*
1000 * If the TDF_TIMEOUT flag is set, just leave. A
1001 * timeout is scheduled anyhow.
1002 */
1003 if (td->td_flags & TDF_TIMEOUT)
1004 return (0);
1005
1006 CTR3(KTR_PROC, "sleepq_abort: thread %p (pid %ld, %s)",
1007 (void *)td, (long)td->td_proc->p_pid, (void *)td->td_name);
1008 td->td_intrval = intrval;
1009 td->td_flags |= TDF_SLEEPABORT;
1010 /*
1011 * If the thread has not slept yet it will find the signal in
1012 * sleepq_catch_signals() and call sleepq_resume_thread. Otherwise
1013 * we have to do it here.
1014 */
1015 if (!TD_IS_SLEEPING(td))
1016 return (0);
1017 wchan = td->td_wchan;
1018 MPASS(wchan != NULL);
1019 sq = sleepq_lookup(wchan);
1020 MPASS(sq != NULL);
1021
1022 /* Thread is asleep on sleep queue sq, so wake it up. */
1023 return (sleepq_resume_thread(sq, td, 0));
1024 }
1025
1026 #ifdef SLEEPQUEUE_PROFILING
1027 #define SLEEPQ_PROF_LOCATIONS 1024
1028 #define SLEEPQ_SBUFSIZE 512
1029 struct sleepq_prof {
1030 LIST_ENTRY(sleepq_prof) sp_link;
1031 const char *sp_wmesg;
1032 long sp_count;
1033 };
1034
1035 LIST_HEAD(sqphead, sleepq_prof);
1036
1037 struct sqphead sleepq_prof_free;
1038 struct sqphead sleepq_hash[SC_TABLESIZE];
1039 static struct sleepq_prof sleepq_profent[SLEEPQ_PROF_LOCATIONS];
1040 static struct mtx sleepq_prof_lock;
1041 MTX_SYSINIT(sleepq_prof_lock, &sleepq_prof_lock, "sleepq_prof", MTX_SPIN);
1042
1043 static void
1044 sleepq_profile(const char *wmesg)
1045 {
1046 struct sleepq_prof *sp;
1047
1048 mtx_lock_spin(&sleepq_prof_lock);
1049 if (prof_enabled == 0)
1050 goto unlock;
1051 LIST_FOREACH(sp, &sleepq_hash[SC_HASH(wmesg)], sp_link)
1052 if (sp->sp_wmesg == wmesg)
1053 goto done;
1054 sp = LIST_FIRST(&sleepq_prof_free);
1055 if (sp == NULL)
1056 goto unlock;
1057 sp->sp_wmesg = wmesg;
1058 LIST_REMOVE(sp, sp_link);
1059 LIST_INSERT_HEAD(&sleepq_hash[SC_HASH(wmesg)], sp, sp_link);
1060 done:
1061 sp->sp_count++;
1062 unlock:
1063 mtx_unlock_spin(&sleepq_prof_lock);
1064 return;
1065 }
1066
1067 static void
1068 sleepq_prof_reset(void)
1069 {
1070 struct sleepq_prof *sp;
1071 int enabled;
1072 int i;
1073
1074 mtx_lock_spin(&sleepq_prof_lock);
1075 enabled = prof_enabled;
1076 prof_enabled = 0;
1077 for (i = 0; i < SC_TABLESIZE; i++)
1078 LIST_INIT(&sleepq_hash[i]);
1079 LIST_INIT(&sleepq_prof_free);
1080 for (i = 0; i < SLEEPQ_PROF_LOCATIONS; i++) {
1081 sp = &sleepq_profent[i];
1082 sp->sp_wmesg = NULL;
1083 sp->sp_count = 0;
1084 LIST_INSERT_HEAD(&sleepq_prof_free, sp, sp_link);
1085 }
1086 prof_enabled = enabled;
1087 mtx_unlock_spin(&sleepq_prof_lock);
1088 }
1089
1090 static int
1091 enable_sleepq_prof(SYSCTL_HANDLER_ARGS)
1092 {
1093 int error, v;
1094
1095 v = prof_enabled;
1096 error = sysctl_handle_int(oidp, &v, v, req);
1097 if (error)
1098 return (error);
1099 if (req->newptr == NULL)
1100 return (error);
1101 if (v == prof_enabled)
1102 return (0);
1103 if (v == 1)
1104 sleepq_prof_reset();
1105 mtx_lock_spin(&sleepq_prof_lock);
1106 prof_enabled = !!v;
1107 mtx_unlock_spin(&sleepq_prof_lock);
1108
1109 return (0);
1110 }
1111
1112 static int
1113 reset_sleepq_prof_stats(SYSCTL_HANDLER_ARGS)
1114 {
1115 int error, v;
1116
1117 v = 0;
1118 error = sysctl_handle_int(oidp, &v, 0, req);
1119 if (error)
1120 return (error);
1121 if (req->newptr == NULL)
1122 return (error);
1123 if (v == 0)
1124 return (0);
1125 sleepq_prof_reset();
1126
1127 return (0);
1128 }
1129
1130 static int
1131 dump_sleepq_prof_stats(SYSCTL_HANDLER_ARGS)
1132 {
1133 struct sleepq_prof *sp;
1134 struct sbuf *sb;
1135 int enabled;
1136 int error;
1137 int i;
1138
1139 error = sysctl_wire_old_buffer(req, 0);
1140 if (error != 0)
1141 return (error);
1142 sb = sbuf_new_for_sysctl(NULL, NULL, SLEEPQ_SBUFSIZE, req);
1143 sbuf_printf(sb, "\nwmesg\tcount\n");
1144 enabled = prof_enabled;
1145 mtx_lock_spin(&sleepq_prof_lock);
1146 prof_enabled = 0;
1147 mtx_unlock_spin(&sleepq_prof_lock);
1148 for (i = 0; i < SC_TABLESIZE; i++) {
1149 LIST_FOREACH(sp, &sleepq_hash[i], sp_link) {
1150 sbuf_printf(sb, "%s\t%ld\n",
1151 sp->sp_wmesg, sp->sp_count);
1152 }
1153 }
1154 mtx_lock_spin(&sleepq_prof_lock);
1155 prof_enabled = enabled;
1156 mtx_unlock_spin(&sleepq_prof_lock);
1157
1158 error = sbuf_finish(sb);
1159 sbuf_delete(sb);
1160 return (error);
1161 }
1162
1163 SYSCTL_PROC(_debug_sleepq, OID_AUTO, stats, CTLTYPE_STRING | CTLFLAG_RD,
1164 NULL, 0, dump_sleepq_prof_stats, "A", "Sleepqueue profiling statistics");
1165 SYSCTL_PROC(_debug_sleepq, OID_AUTO, reset, CTLTYPE_INT | CTLFLAG_RW,
1166 NULL, 0, reset_sleepq_prof_stats, "I",
1167 "Reset sleepqueue profiling statistics");
1168 SYSCTL_PROC(_debug_sleepq, OID_AUTO, enable, CTLTYPE_INT | CTLFLAG_RW,
1169 NULL, 0, enable_sleepq_prof, "I", "Enable sleepqueue profiling");
1170 #endif
1171
1172 #ifdef DDB
1173 DB_SHOW_COMMAND(sleepq, db_show_sleepqueue)
1174 {
1175 struct sleepqueue_chain *sc;
1176 struct sleepqueue *sq;
1177 #ifdef INVARIANTS
1178 struct lock_object *lock;
1179 #endif
1180 struct thread *td;
1181 void *wchan;
1182 int i;
1183
1184 if (!have_addr)
1185 return;
1186
1187 /*
1188 * First, see if there is an active sleep queue for the wait channel
1189 * indicated by the address.
1190 */
1191 wchan = (void *)addr;
1192 sc = SC_LOOKUP(wchan);
1193 LIST_FOREACH(sq, &sc->sc_queues, sq_hash)
1194 if (sq->sq_wchan == wchan)
1195 goto found;
1196
1197 /*
1198 * Second, see if there is an active sleep queue at the address
1199 * indicated.
1200 */
1201 for (i = 0; i < SC_TABLESIZE; i++)
1202 LIST_FOREACH(sq, &sleepq_chains[i].sc_queues, sq_hash) {
1203 if (sq == (struct sleepqueue *)addr)
1204 goto found;
1205 }
1206
1207 db_printf("Unable to locate a sleep queue via %p\n", (void *)addr);
1208 return;
1209 found:
1210 db_printf("Wait channel: %p\n", sq->sq_wchan);
1211 db_printf("Queue type: %d\n", sq->sq_type);
1212 #ifdef INVARIANTS
1213 if (sq->sq_lock) {
1214 lock = sq->sq_lock;
1215 db_printf("Associated Interlock: %p - (%s) %s\n", lock,
1216 LOCK_CLASS(lock)->lc_name, lock->lo_name);
1217 }
1218 #endif
1219 db_printf("Blocked threads:\n");
1220 for (i = 0; i < NR_SLEEPQS; i++) {
1221 db_printf("\nQueue[%d]:\n", i);
1222 if (TAILQ_EMPTY(&sq->sq_blocked[i]))
1223 db_printf("\tempty\n");
1224 else
1225 TAILQ_FOREACH(td, &sq->sq_blocked[0],
1226 td_slpq) {
1227 db_printf("\t%p (tid %d, pid %d, \"%s\")\n", td,
1228 td->td_tid, td->td_proc->p_pid,
1229 td->td_name);
1230 }
1231 db_printf("(expected: %u)\n", sq->sq_blockedcnt[i]);
1232 }
1233 }
1234
1235 /* Alias 'show sleepqueue' to 'show sleepq'. */
1236 DB_SHOW_ALIAS(sleepqueue, db_show_sleepqueue);
1237 #endif
Cache object: c9d956b73fcc9545f301a787de26d85e
|