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/8.4/sys/kern/subr_sleepqueue.c 236345 2012-05-30 23:42:48Z rstone $");
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 SYSCTL_NODE(_debug, OID_AUTO, sleepq, CTLFLAG_RD, 0, "sleepq profiling");
145 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 if (flags & SLEEPQ_STOP_ON_BDRY)
355 td->td_flags |= TDF_SBDRY;
356 }
357 thread_unlock(td);
358 }
359
360 /*
361 * Sets a timeout that will remove the current thread from the specified
362 * sleep queue after timo ticks if the thread has not already been awakened.
363 */
364 void
365 sleepq_set_timeout(void *wchan, int timo)
366 {
367 struct sleepqueue_chain *sc;
368 struct thread *td;
369
370 td = curthread;
371 sc = SC_LOOKUP(wchan);
372 mtx_assert(&sc->sc_lock, MA_OWNED);
373 MPASS(TD_ON_SLEEPQ(td));
374 MPASS(td->td_sleepqueue == NULL);
375 MPASS(wchan != NULL);
376 callout_reset_curcpu(&td->td_slpcallout, timo, sleepq_timeout, td);
377 }
378
379 /*
380 * Return the number of actual sleepers for the specified queue.
381 */
382 u_int
383 sleepq_sleepcnt(void *wchan, int queue)
384 {
385 struct sleepqueue *sq;
386
387 KASSERT(wchan != NULL, ("%s: invalid NULL wait channel", __func__));
388 MPASS((queue >= 0) && (queue < NR_SLEEPQS));
389 sq = sleepq_lookup(wchan);
390 if (sq == NULL)
391 return (0);
392 return (sq->sq_blockedcnt[queue]);
393 }
394
395 /*
396 * Marks the pending sleep of the current thread as interruptible and
397 * makes an initial check for pending signals before putting a thread
398 * to sleep. Enters and exits with the thread lock held. Thread lock
399 * may have transitioned from the sleepq lock to a run lock.
400 */
401 static int
402 sleepq_catch_signals(void *wchan, int pri)
403 {
404 struct sleepqueue_chain *sc;
405 struct sleepqueue *sq;
406 struct thread *td;
407 struct proc *p;
408 struct sigacts *ps;
409 int sig, ret, stop_allowed;
410
411 td = curthread;
412 p = curproc;
413 sc = SC_LOOKUP(wchan);
414 mtx_assert(&sc->sc_lock, MA_OWNED);
415 MPASS(wchan != NULL);
416 /*
417 * See if there are any pending signals for this thread. If not
418 * we can switch immediately. Otherwise do the signal processing
419 * directly.
420 */
421 thread_lock(td);
422 if ((td->td_flags & (TDF_NEEDSIGCHK | TDF_NEEDSUSPCHK)) == 0) {
423 sleepq_switch(wchan, pri);
424 return (0);
425 }
426 stop_allowed = (td->td_flags & TDF_SBDRY) ? SIG_STOP_NOT_ALLOWED :
427 SIG_STOP_ALLOWED;
428 thread_unlock(td);
429 mtx_unlock_spin(&sc->sc_lock);
430 CTR3(KTR_PROC, "sleepq catching signals: thread %p (pid %ld, %s)",
431 (void *)td, (long)p->p_pid, td->td_name);
432 PROC_LOCK(p);
433 ps = p->p_sigacts;
434 mtx_lock(&ps->ps_mtx);
435 sig = cursig(td, stop_allowed);
436 if (sig == 0) {
437 mtx_unlock(&ps->ps_mtx);
438 ret = thread_suspend_check(1);
439 MPASS(ret == 0 || ret == EINTR || ret == ERESTART);
440 } else {
441 if (SIGISMEMBER(ps->ps_sigintr, sig))
442 ret = EINTR;
443 else
444 ret = ERESTART;
445 mtx_unlock(&ps->ps_mtx);
446 }
447 /*
448 * Lock the per-process spinlock prior to dropping the PROC_LOCK
449 * to avoid a signal delivery race. PROC_LOCK, PROC_SLOCK, and
450 * thread_lock() are currently held in tdsignal().
451 */
452 PROC_SLOCK(p);
453 mtx_lock_spin(&sc->sc_lock);
454 PROC_UNLOCK(p);
455 thread_lock(td);
456 PROC_SUNLOCK(p);
457 if (ret == 0) {
458 sleepq_switch(wchan, pri);
459 return (0);
460 }
461 /*
462 * There were pending signals and this thread is still
463 * on the sleep queue, remove it from the sleep queue.
464 */
465 if (TD_ON_SLEEPQ(td)) {
466 sq = sleepq_lookup(wchan);
467 if (sleepq_resume_thread(sq, td, 0)) {
468 #ifdef INVARIANTS
469 /*
470 * This thread hasn't gone to sleep yet, so it
471 * should not be swapped out.
472 */
473 panic("not waking up swapper");
474 #endif
475 }
476 }
477 mtx_unlock_spin(&sc->sc_lock);
478 MPASS(td->td_lock != &sc->sc_lock);
479 return (ret);
480 }
481
482 /*
483 * Switches to another thread if we are still asleep on a sleep queue.
484 * Returns with thread lock.
485 */
486 static void
487 sleepq_switch(void *wchan, int pri)
488 {
489 struct sleepqueue_chain *sc;
490 struct sleepqueue *sq;
491 struct thread *td;
492
493 td = curthread;
494 sc = SC_LOOKUP(wchan);
495 mtx_assert(&sc->sc_lock, MA_OWNED);
496 THREAD_LOCK_ASSERT(td, MA_OWNED);
497
498 /*
499 * If we have a sleep queue, then we've already been woken up, so
500 * just return.
501 */
502 if (td->td_sleepqueue != NULL) {
503 mtx_unlock_spin(&sc->sc_lock);
504 return;
505 }
506
507 /*
508 * If TDF_TIMEOUT is set, then our sleep has been timed out
509 * already but we are still on the sleep queue, so dequeue the
510 * thread and return.
511 */
512 if (td->td_flags & TDF_TIMEOUT) {
513 MPASS(TD_ON_SLEEPQ(td));
514 sq = sleepq_lookup(wchan);
515 if (sleepq_resume_thread(sq, td, 0)) {
516 #ifdef INVARIANTS
517 /*
518 * This thread hasn't gone to sleep yet, so it
519 * should not be swapped out.
520 */
521 panic("not waking up swapper");
522 #endif
523 }
524 mtx_unlock_spin(&sc->sc_lock);
525 return;
526 }
527 #ifdef SLEEPQUEUE_PROFILING
528 if (prof_enabled)
529 sleepq_profile(td->td_wmesg);
530 #endif
531 MPASS(td->td_sleepqueue == NULL);
532 sched_sleep(td, pri);
533 thread_lock_set(td, &sc->sc_lock);
534 SDT_PROBE0(sched, , , sleep);
535 TD_SET_SLEEPING(td);
536 mi_switch(SW_VOL | SWT_SLEEPQ, NULL);
537 KASSERT(TD_IS_RUNNING(td), ("running but not TDS_RUNNING"));
538 CTR3(KTR_PROC, "sleepq resume: thread %p (pid %ld, %s)",
539 (void *)td, (long)td->td_proc->p_pid, (void *)td->td_name);
540 }
541
542 /*
543 * Check to see if we timed out.
544 */
545 static int
546 sleepq_check_timeout(void)
547 {
548 struct thread *td;
549
550 td = curthread;
551 THREAD_LOCK_ASSERT(td, MA_OWNED);
552
553 /*
554 * If TDF_TIMEOUT is set, we timed out.
555 */
556 if (td->td_flags & TDF_TIMEOUT) {
557 td->td_flags &= ~TDF_TIMEOUT;
558 return (EWOULDBLOCK);
559 }
560
561 /*
562 * If TDF_TIMOFAIL is set, the timeout ran after we had
563 * already been woken up.
564 */
565 if (td->td_flags & TDF_TIMOFAIL)
566 td->td_flags &= ~TDF_TIMOFAIL;
567
568 /*
569 * If callout_stop() fails, then the timeout is running on
570 * another CPU, so synchronize with it to avoid having it
571 * accidentally wake up a subsequent sleep.
572 */
573 else if (callout_stop(&td->td_slpcallout) == 0) {
574 td->td_flags |= TDF_TIMEOUT;
575 TD_SET_SLEEPING(td);
576 mi_switch(SW_INVOL | SWT_SLEEPQTIMO, NULL);
577 }
578 return (0);
579 }
580
581 /*
582 * Check to see if we were awoken by a signal.
583 */
584 static int
585 sleepq_check_signals(void)
586 {
587 struct thread *td;
588
589 td = curthread;
590 THREAD_LOCK_ASSERT(td, MA_OWNED);
591
592 /* We are no longer in an interruptible sleep. */
593 if (td->td_flags & TDF_SINTR)
594 td->td_flags &= ~(TDF_SINTR | TDF_SBDRY);
595
596 if (td->td_flags & TDF_SLEEPABORT) {
597 td->td_flags &= ~TDF_SLEEPABORT;
598 return (td->td_intrval);
599 }
600
601 return (0);
602 }
603
604 /*
605 * Block the current thread until it is awakened from its sleep queue.
606 */
607 void
608 sleepq_wait(void *wchan, int pri)
609 {
610 struct thread *td;
611
612 td = curthread;
613 MPASS(!(td->td_flags & TDF_SINTR));
614 thread_lock(td);
615 sleepq_switch(wchan, pri);
616 thread_unlock(td);
617 }
618
619 /*
620 * Block the current thread until it is awakened from its sleep queue
621 * or it is interrupted by a signal.
622 */
623 int
624 sleepq_wait_sig(void *wchan, int pri)
625 {
626 int rcatch;
627 int rval;
628
629 rcatch = sleepq_catch_signals(wchan, pri);
630 rval = sleepq_check_signals();
631 thread_unlock(curthread);
632 if (rcatch)
633 return (rcatch);
634 return (rval);
635 }
636
637 /*
638 * Block the current thread until it is awakened from its sleep queue
639 * or it times out while waiting.
640 */
641 int
642 sleepq_timedwait(void *wchan, int pri)
643 {
644 struct thread *td;
645 int rval;
646
647 td = curthread;
648 MPASS(!(td->td_flags & TDF_SINTR));
649 thread_lock(td);
650 sleepq_switch(wchan, pri);
651 rval = sleepq_check_timeout();
652 thread_unlock(td);
653
654 return (rval);
655 }
656
657 /*
658 * Block the current thread until it is awakened from its sleep queue,
659 * it is interrupted by a signal, or it times out waiting to be awakened.
660 */
661 int
662 sleepq_timedwait_sig(void *wchan, int pri)
663 {
664 int rcatch, rvalt, rvals;
665
666 rcatch = sleepq_catch_signals(wchan, pri);
667 rvalt = sleepq_check_timeout();
668 rvals = sleepq_check_signals();
669 thread_unlock(curthread);
670 if (rcatch)
671 return (rcatch);
672 if (rvals)
673 return (rvals);
674 return (rvalt);
675 }
676
677 /*
678 * Returns the type of sleepqueue given a waitchannel.
679 */
680 int
681 sleepq_type(void *wchan)
682 {
683 struct sleepqueue *sq;
684 int type;
685
686 MPASS(wchan != NULL);
687
688 sleepq_lock(wchan);
689 sq = sleepq_lookup(wchan);
690 if (sq == NULL) {
691 sleepq_release(wchan);
692 return (-1);
693 }
694 type = sq->sq_type;
695 sleepq_release(wchan);
696 return (type);
697 }
698
699 /*
700 * Removes a thread from a sleep queue and makes it
701 * runnable.
702 */
703 static int
704 sleepq_resume_thread(struct sleepqueue *sq, struct thread *td, int pri)
705 {
706 struct sleepqueue_chain *sc;
707
708 MPASS(td != NULL);
709 MPASS(sq->sq_wchan != NULL);
710 MPASS(td->td_wchan == sq->sq_wchan);
711 MPASS(td->td_sqqueue < NR_SLEEPQS && td->td_sqqueue >= 0);
712 THREAD_LOCK_ASSERT(td, MA_OWNED);
713 sc = SC_LOOKUP(sq->sq_wchan);
714 mtx_assert(&sc->sc_lock, MA_OWNED);
715
716 SDT_PROBE2(sched, , , wakeup, td, td->td_proc);
717
718 /* Remove the thread from the queue. */
719 sq->sq_blockedcnt[td->td_sqqueue]--;
720 TAILQ_REMOVE(&sq->sq_blocked[td->td_sqqueue], td, td_slpq);
721
722 /*
723 * Get a sleep queue for this thread. If this is the last waiter,
724 * use the queue itself and take it out of the chain, otherwise,
725 * remove a queue from the free list.
726 */
727 if (LIST_EMPTY(&sq->sq_free)) {
728 td->td_sleepqueue = sq;
729 #ifdef INVARIANTS
730 sq->sq_wchan = NULL;
731 #endif
732 #ifdef SLEEPQUEUE_PROFILING
733 sc->sc_depth--;
734 #endif
735 } else
736 td->td_sleepqueue = LIST_FIRST(&sq->sq_free);
737 LIST_REMOVE(td->td_sleepqueue, sq_hash);
738
739 td->td_wmesg = NULL;
740 td->td_wchan = NULL;
741 td->td_flags &= ~(TDF_SINTR | TDF_SBDRY);
742
743 CTR3(KTR_PROC, "sleepq_wakeup: thread %p (pid %ld, %s)",
744 (void *)td, (long)td->td_proc->p_pid, td->td_name);
745
746 /* Adjust priority if requested. */
747 MPASS(pri == 0 || (pri >= PRI_MIN && pri <= PRI_MAX));
748 if (pri != 0 && td->td_priority > pri)
749 sched_prio(td, pri);
750
751 /*
752 * Note that thread td might not be sleeping if it is running
753 * sleepq_catch_signals() on another CPU or is blocked on its
754 * proc lock to check signals. There's no need to mark the
755 * thread runnable in that case.
756 */
757 if (TD_IS_SLEEPING(td)) {
758 TD_CLR_SLEEPING(td);
759 return (setrunnable(td));
760 }
761 return (0);
762 }
763
764 #ifdef INVARIANTS
765 /*
766 * UMA zone item deallocator.
767 */
768 static void
769 sleepq_dtor(void *mem, int size, void *arg)
770 {
771 struct sleepqueue *sq;
772 int i;
773
774 sq = mem;
775 for (i = 0; i < NR_SLEEPQS; i++) {
776 MPASS(TAILQ_EMPTY(&sq->sq_blocked[i]));
777 MPASS(sq->sq_blockedcnt[i] == 0);
778 }
779 }
780 #endif
781
782 /*
783 * UMA zone item initializer.
784 */
785 static int
786 sleepq_init(void *mem, int size, int flags)
787 {
788 struct sleepqueue *sq;
789 int i;
790
791 bzero(mem, size);
792 sq = mem;
793 for (i = 0; i < NR_SLEEPQS; i++) {
794 TAILQ_INIT(&sq->sq_blocked[i]);
795 sq->sq_blockedcnt[i] = 0;
796 }
797 LIST_INIT(&sq->sq_free);
798 return (0);
799 }
800
801 /*
802 * Find the highest priority thread sleeping on a wait channel and resume it.
803 */
804 int
805 sleepq_signal(void *wchan, int flags, int pri, int queue)
806 {
807 struct sleepqueue *sq;
808 struct thread *td, *besttd;
809 int wakeup_swapper;
810
811 CTR2(KTR_PROC, "sleepq_signal(%p, %d)", wchan, flags);
812 KASSERT(wchan != NULL, ("%s: invalid NULL wait channel", __func__));
813 MPASS((queue >= 0) && (queue < NR_SLEEPQS));
814 sq = sleepq_lookup(wchan);
815 if (sq == NULL)
816 return (0);
817 KASSERT(sq->sq_type == (flags & SLEEPQ_TYPE),
818 ("%s: mismatch between sleep/wakeup and cv_*", __func__));
819
820 /*
821 * Find the highest priority thread on the queue. If there is a
822 * tie, use the thread that first appears in the queue as it has
823 * been sleeping the longest since threads are always added to
824 * the tail of sleep queues.
825 */
826 besttd = NULL;
827 TAILQ_FOREACH(td, &sq->sq_blocked[queue], td_slpq) {
828 if (besttd == NULL || td->td_priority < besttd->td_priority)
829 besttd = td;
830 }
831 MPASS(besttd != NULL);
832 thread_lock(besttd);
833 wakeup_swapper = sleepq_resume_thread(sq, besttd, pri);
834 thread_unlock(besttd);
835 return (wakeup_swapper);
836 }
837
838 /*
839 * Resume all threads sleeping on a specified wait channel.
840 */
841 int
842 sleepq_broadcast(void *wchan, int flags, int pri, int queue)
843 {
844 struct sleepqueue *sq;
845 struct thread *td, *tdn;
846 int wakeup_swapper;
847
848 CTR2(KTR_PROC, "sleepq_broadcast(%p, %d)", wchan, flags);
849 KASSERT(wchan != NULL, ("%s: invalid NULL wait channel", __func__));
850 MPASS((queue >= 0) && (queue < NR_SLEEPQS));
851 sq = sleepq_lookup(wchan);
852 if (sq == NULL)
853 return (0);
854 KASSERT(sq->sq_type == (flags & SLEEPQ_TYPE),
855 ("%s: mismatch between sleep/wakeup and cv_*", __func__));
856
857 /* Resume all blocked threads on the sleep queue. */
858 wakeup_swapper = 0;
859 TAILQ_FOREACH_SAFE(td, &sq->sq_blocked[queue], td_slpq, tdn) {
860 thread_lock(td);
861 if (sleepq_resume_thread(sq, td, pri))
862 wakeup_swapper = 1;
863 thread_unlock(td);
864 }
865 return (wakeup_swapper);
866 }
867
868 /*
869 * Time sleeping threads out. When the timeout expires, the thread is
870 * removed from the sleep queue and made runnable if it is still asleep.
871 */
872 static void
873 sleepq_timeout(void *arg)
874 {
875 struct sleepqueue_chain *sc;
876 struct sleepqueue *sq;
877 struct thread *td;
878 void *wchan;
879 int wakeup_swapper;
880
881 td = arg;
882 wakeup_swapper = 0;
883 CTR3(KTR_PROC, "sleepq_timeout: thread %p (pid %ld, %s)",
884 (void *)td, (long)td->td_proc->p_pid, (void *)td->td_name);
885
886 /*
887 * First, see if the thread is asleep and get the wait channel if
888 * it is.
889 */
890 thread_lock(td);
891 if (TD_IS_SLEEPING(td) && TD_ON_SLEEPQ(td)) {
892 wchan = td->td_wchan;
893 sc = SC_LOOKUP(wchan);
894 THREAD_LOCKPTR_ASSERT(td, &sc->sc_lock);
895 sq = sleepq_lookup(wchan);
896 MPASS(sq != NULL);
897 td->td_flags |= TDF_TIMEOUT;
898 wakeup_swapper = sleepq_resume_thread(sq, td, 0);
899 thread_unlock(td);
900 if (wakeup_swapper)
901 kick_proc0();
902 return;
903 }
904
905 /*
906 * If the thread is on the SLEEPQ but isn't sleeping yet, it
907 * can either be on another CPU in between sleepq_add() and
908 * one of the sleepq_*wait*() routines or it can be in
909 * sleepq_catch_signals().
910 */
911 if (TD_ON_SLEEPQ(td)) {
912 td->td_flags |= TDF_TIMEOUT;
913 thread_unlock(td);
914 return;
915 }
916
917 /*
918 * Now check for the edge cases. First, if TDF_TIMEOUT is set,
919 * then the other thread has already yielded to us, so clear
920 * the flag and resume it. If TDF_TIMEOUT is not set, then the
921 * we know that the other thread is not on a sleep queue, but it
922 * hasn't resumed execution yet. In that case, set TDF_TIMOFAIL
923 * to let it know that the timeout has already run and doesn't
924 * need to be canceled.
925 */
926 if (td->td_flags & TDF_TIMEOUT) {
927 MPASS(TD_IS_SLEEPING(td));
928 td->td_flags &= ~TDF_TIMEOUT;
929 TD_CLR_SLEEPING(td);
930 wakeup_swapper = setrunnable(td);
931 } else
932 td->td_flags |= TDF_TIMOFAIL;
933 thread_unlock(td);
934 if (wakeup_swapper)
935 kick_proc0();
936 }
937
938 /*
939 * Resumes a specific thread from the sleep queue associated with a specific
940 * wait channel if it is on that queue.
941 */
942 void
943 sleepq_remove(struct thread *td, void *wchan)
944 {
945 struct sleepqueue *sq;
946 int wakeup_swapper;
947
948 /*
949 * Look up the sleep queue for this wait channel, then re-check
950 * that the thread is asleep on that channel, if it is not, then
951 * bail.
952 */
953 MPASS(wchan != NULL);
954 sleepq_lock(wchan);
955 sq = sleepq_lookup(wchan);
956 /*
957 * We can not lock the thread here as it may be sleeping on a
958 * different sleepq. However, holding the sleepq lock for this
959 * wchan can guarantee that we do not miss a wakeup for this
960 * channel. The asserts below will catch any false positives.
961 */
962 if (!TD_ON_SLEEPQ(td) || td->td_wchan != wchan) {
963 sleepq_release(wchan);
964 return;
965 }
966 /* Thread is asleep on sleep queue sq, so wake it up. */
967 thread_lock(td);
968 MPASS(sq != NULL);
969 MPASS(td->td_wchan == wchan);
970 wakeup_swapper = sleepq_resume_thread(sq, td, 0);
971 thread_unlock(td);
972 sleepq_release(wchan);
973 if (wakeup_swapper)
974 kick_proc0();
975 }
976
977 /*
978 * Abort a thread as if an interrupt had occurred. Only abort
979 * interruptible waits (unfortunately it isn't safe to abort others).
980 */
981 int
982 sleepq_abort(struct thread *td, int intrval)
983 {
984 struct sleepqueue *sq;
985 void *wchan;
986
987 THREAD_LOCK_ASSERT(td, MA_OWNED);
988 MPASS(TD_ON_SLEEPQ(td));
989 MPASS(td->td_flags & TDF_SINTR);
990 MPASS(intrval == EINTR || intrval == ERESTART);
991
992 /*
993 * If the TDF_TIMEOUT flag is set, just leave. A
994 * timeout is scheduled anyhow.
995 */
996 if (td->td_flags & TDF_TIMEOUT)
997 return (0);
998
999 CTR3(KTR_PROC, "sleepq_abort: thread %p (pid %ld, %s)",
1000 (void *)td, (long)td->td_proc->p_pid, (void *)td->td_name);
1001 td->td_intrval = intrval;
1002 td->td_flags |= TDF_SLEEPABORT;
1003 /*
1004 * If the thread has not slept yet it will find the signal in
1005 * sleepq_catch_signals() and call sleepq_resume_thread. Otherwise
1006 * we have to do it here.
1007 */
1008 if (!TD_IS_SLEEPING(td))
1009 return (0);
1010 wchan = td->td_wchan;
1011 MPASS(wchan != NULL);
1012 sq = sleepq_lookup(wchan);
1013 MPASS(sq != NULL);
1014
1015 /* Thread is asleep on sleep queue sq, so wake it up. */
1016 return (sleepq_resume_thread(sq, td, 0));
1017 }
1018
1019 #ifdef SLEEPQUEUE_PROFILING
1020 #define SLEEPQ_PROF_LOCATIONS 1024
1021 #define SLEEPQ_SBUFSIZE (40 * 512)
1022 struct sleepq_prof {
1023 LIST_ENTRY(sleepq_prof) sp_link;
1024 const char *sp_wmesg;
1025 long sp_count;
1026 };
1027
1028 LIST_HEAD(sqphead, sleepq_prof);
1029
1030 struct sqphead sleepq_prof_free;
1031 struct sqphead sleepq_hash[SC_TABLESIZE];
1032 static struct sleepq_prof sleepq_profent[SLEEPQ_PROF_LOCATIONS];
1033 static struct mtx sleepq_prof_lock;
1034 MTX_SYSINIT(sleepq_prof_lock, &sleepq_prof_lock, "sleepq_prof", MTX_SPIN);
1035
1036 static void
1037 sleepq_profile(const char *wmesg)
1038 {
1039 struct sleepq_prof *sp;
1040
1041 mtx_lock_spin(&sleepq_prof_lock);
1042 if (prof_enabled == 0)
1043 goto unlock;
1044 LIST_FOREACH(sp, &sleepq_hash[SC_HASH(wmesg)], sp_link)
1045 if (sp->sp_wmesg == wmesg)
1046 goto done;
1047 sp = LIST_FIRST(&sleepq_prof_free);
1048 if (sp == NULL)
1049 goto unlock;
1050 sp->sp_wmesg = wmesg;
1051 LIST_REMOVE(sp, sp_link);
1052 LIST_INSERT_HEAD(&sleepq_hash[SC_HASH(wmesg)], sp, sp_link);
1053 done:
1054 sp->sp_count++;
1055 unlock:
1056 mtx_unlock_spin(&sleepq_prof_lock);
1057 return;
1058 }
1059
1060 static void
1061 sleepq_prof_reset(void)
1062 {
1063 struct sleepq_prof *sp;
1064 int enabled;
1065 int i;
1066
1067 mtx_lock_spin(&sleepq_prof_lock);
1068 enabled = prof_enabled;
1069 prof_enabled = 0;
1070 for (i = 0; i < SC_TABLESIZE; i++)
1071 LIST_INIT(&sleepq_hash[i]);
1072 LIST_INIT(&sleepq_prof_free);
1073 for (i = 0; i < SLEEPQ_PROF_LOCATIONS; i++) {
1074 sp = &sleepq_profent[i];
1075 sp->sp_wmesg = NULL;
1076 sp->sp_count = 0;
1077 LIST_INSERT_HEAD(&sleepq_prof_free, sp, sp_link);
1078 }
1079 prof_enabled = enabled;
1080 mtx_unlock_spin(&sleepq_prof_lock);
1081 }
1082
1083 static int
1084 enable_sleepq_prof(SYSCTL_HANDLER_ARGS)
1085 {
1086 int error, v;
1087
1088 v = prof_enabled;
1089 error = sysctl_handle_int(oidp, &v, v, req);
1090 if (error)
1091 return (error);
1092 if (req->newptr == NULL)
1093 return (error);
1094 if (v == prof_enabled)
1095 return (0);
1096 if (v == 1)
1097 sleepq_prof_reset();
1098 mtx_lock_spin(&sleepq_prof_lock);
1099 prof_enabled = !!v;
1100 mtx_unlock_spin(&sleepq_prof_lock);
1101
1102 return (0);
1103 }
1104
1105 static int
1106 reset_sleepq_prof_stats(SYSCTL_HANDLER_ARGS)
1107 {
1108 int error, v;
1109
1110 v = 0;
1111 error = sysctl_handle_int(oidp, &v, 0, req);
1112 if (error)
1113 return (error);
1114 if (req->newptr == NULL)
1115 return (error);
1116 if (v == 0)
1117 return (0);
1118 sleepq_prof_reset();
1119
1120 return (0);
1121 }
1122
1123 static int
1124 dump_sleepq_prof_stats(SYSCTL_HANDLER_ARGS)
1125 {
1126 static int multiplier = 1;
1127 struct sleepq_prof *sp;
1128 struct sbuf *sb;
1129 int enabled;
1130 int error;
1131 int i;
1132
1133 retry_sbufops:
1134 sb = sbuf_new(NULL, NULL, SLEEPQ_SBUFSIZE * multiplier, SBUF_FIXEDLEN);
1135 sbuf_printf(sb, "\nwmesg\tcount\n");
1136 enabled = prof_enabled;
1137 mtx_lock_spin(&sleepq_prof_lock);
1138 prof_enabled = 0;
1139 mtx_unlock_spin(&sleepq_prof_lock);
1140 for (i = 0; i < SC_TABLESIZE; i++) {
1141 LIST_FOREACH(sp, &sleepq_hash[i], sp_link) {
1142 sbuf_printf(sb, "%s\t%ld\n",
1143 sp->sp_wmesg, sp->sp_count);
1144 if (sbuf_overflowed(sb)) {
1145 sbuf_delete(sb);
1146 multiplier++;
1147 goto retry_sbufops;
1148 }
1149 }
1150 }
1151 mtx_lock_spin(&sleepq_prof_lock);
1152 prof_enabled = enabled;
1153 mtx_unlock_spin(&sleepq_prof_lock);
1154
1155 sbuf_finish(sb);
1156 error = SYSCTL_OUT(req, sbuf_data(sb), sbuf_len(sb) + 1);
1157 sbuf_delete(sb);
1158 return (error);
1159 }
1160
1161 SYSCTL_PROC(_debug_sleepq, OID_AUTO, stats, CTLTYPE_STRING | CTLFLAG_RD,
1162 NULL, 0, dump_sleepq_prof_stats, "A", "Sleepqueue profiling statistics");
1163 SYSCTL_PROC(_debug_sleepq, OID_AUTO, reset, CTLTYPE_INT | CTLFLAG_RW,
1164 NULL, 0, reset_sleepq_prof_stats, "I",
1165 "Reset sleepqueue profiling statistics");
1166 SYSCTL_PROC(_debug_sleepq, OID_AUTO, enable, CTLTYPE_INT | CTLFLAG_RW,
1167 NULL, 0, enable_sleepq_prof, "I", "Enable sleepqueue profiling");
1168 #endif
1169
1170 #ifdef DDB
1171 DB_SHOW_COMMAND(sleepq, db_show_sleepqueue)
1172 {
1173 struct sleepqueue_chain *sc;
1174 struct sleepqueue *sq;
1175 #ifdef INVARIANTS
1176 struct lock_object *lock;
1177 #endif
1178 struct thread *td;
1179 void *wchan;
1180 int i;
1181
1182 if (!have_addr)
1183 return;
1184
1185 /*
1186 * First, see if there is an active sleep queue for the wait channel
1187 * indicated by the address.
1188 */
1189 wchan = (void *)addr;
1190 sc = SC_LOOKUP(wchan);
1191 LIST_FOREACH(sq, &sc->sc_queues, sq_hash)
1192 if (sq->sq_wchan == wchan)
1193 goto found;
1194
1195 /*
1196 * Second, see if there is an active sleep queue at the address
1197 * indicated.
1198 */
1199 for (i = 0; i < SC_TABLESIZE; i++)
1200 LIST_FOREACH(sq, &sleepq_chains[i].sc_queues, sq_hash) {
1201 if (sq == (struct sleepqueue *)addr)
1202 goto found;
1203 }
1204
1205 db_printf("Unable to locate a sleep queue via %p\n", (void *)addr);
1206 return;
1207 found:
1208 db_printf("Wait channel: %p\n", sq->sq_wchan);
1209 db_printf("Queue type: %d\n", sq->sq_type);
1210 #ifdef INVARIANTS
1211 if (sq->sq_lock) {
1212 lock = sq->sq_lock;
1213 db_printf("Associated Interlock: %p - (%s) %s\n", lock,
1214 LOCK_CLASS(lock)->lc_name, lock->lo_name);
1215 }
1216 #endif
1217 db_printf("Blocked threads:\n");
1218 for (i = 0; i < NR_SLEEPQS; i++) {
1219 db_printf("\nQueue[%d]:\n", i);
1220 if (TAILQ_EMPTY(&sq->sq_blocked[i]))
1221 db_printf("\tempty\n");
1222 else
1223 TAILQ_FOREACH(td, &sq->sq_blocked[0],
1224 td_slpq) {
1225 db_printf("\t%p (tid %d, pid %d, \"%s\")\n", td,
1226 td->td_tid, td->td_proc->p_pid,
1227 td->td_name);
1228 }
1229 db_printf("(expected: %u)\n", sq->sq_blockedcnt[i]);
1230 }
1231 }
1232
1233 /* Alias 'show sleepqueue' to 'show sleepq'. */
1234 DB_SHOW_ALIAS(sleepqueue, db_show_sleepqueue);
1235 #endif
Cache object: a415012df72e1be581e9f3ef4e853450
|