2 * Copyright (c) 2004 John Baldwin <jhb@FreeBSD.org>
5 * Redistribution and use in source and binary forms, with or without
6 * modification, are permitted provided that the following conditions
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.
14 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
15 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
16 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
17 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
18 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
19 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
20 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
21 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
22 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
23 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
28 * Implementation of sleep queues used to hold queue of threads blocked on
29 * a wait channel. Sleep queues different from turnstiles in that wait
30 * channels are not owned by anyone, so there is no priority propagation.
31 * Sleep queues can also provide a timeout and can also be interrupted by
32 * signals. That said, there are several similarities between the turnstile
33 * and sleep queue implementations. (Note: turnstiles were implemented
34 * first.) For example, both use a hash table of the same size where each
35 * bucket is referred to as a "chain" that contains both a spin lock and
36 * a linked list of queues. An individual queue is located by using a hash
37 * to pick a chain, locking the chain, and then walking the chain searching
38 * for the queue. This means that a wait channel object does not need to
39 * embed it's queue head just as locks do not embed their turnstile queue
40 * head. Threads also carry around a sleep queue that they lend to the
41 * wait channel when blocking. Just as in turnstiles, the queue includes
42 * a free list of the sleep queues of other threads blocked on the same
43 * wait channel in the case of multiple waiters.
45 * Some additional functionality provided by sleep queues include the
46 * ability to set a timeout. The timeout is managed using a per-thread
47 * callout that resumes a thread if it is asleep. A thread may also
48 * catch signals while it is asleep (aka an interruptible sleep). The
49 * signal code uses sleepq_abort() to interrupt a sleeping thread. Finally,
50 * sleep queues also provide some extra assertions. One is not allowed to
51 * mix the sleep/wakeup and cv APIs for a given wait channel. Also, one
52 * must consistently use the same lock to synchronize with a wait channel,
53 * though this check is currently only a warning for sleep/wakeup due to
54 * pre-existing abuse of that API. The same lock must also be held when
55 * awakening threads, though that is currently only enforced for condition
59 #include <sys/cdefs.h>
60 __FBSDID("$FreeBSD$");
62 #include "opt_sleepqueue_profiling.h"
64 #include "opt_kdtrace.h"
65 #include "opt_sched.h"
67 #include <sys/param.h>
68 #include <sys/systm.h>
70 #include <sys/kernel.h>
72 #include <sys/mutex.h>
75 #include <sys/sched.h>
77 #include <sys/signalvar.h>
78 #include <sys/sleepqueue.h>
79 #include <sys/sysctl.h>
88 * Constants for the hash table of sleep queue chains.
89 * SC_TABLESIZE must be a power of two for SC_MASK to work properly.
91 #define SC_TABLESIZE 256 /* Must be power of 2. */
92 #define SC_MASK (SC_TABLESIZE - 1)
94 #define SC_HASH(wc) ((((uintptr_t)(wc) >> SC_SHIFT) ^ (uintptr_t)(wc)) & \
96 #define SC_LOOKUP(wc) &sleepq_chains[SC_HASH(wc)]
99 * There two different lists of sleep queues. Both lists are connected
100 * via the sq_hash entries. The first list is the sleep queue chain list
101 * that a sleep queue is on when it is attached to a wait channel. The
102 * second list is the free list hung off of a sleep queue that is attached
105 * Each sleep queue also contains the wait channel it is attached to, the
106 * list of threads blocked on that wait channel, flags specific to the
107 * wait channel, and the lock used to synchronize with a wait channel.
108 * The flags are used to catch mismatches between the various consumers
109 * of the sleep queue API (e.g. sleep/wakeup and condition variables).
110 * The lock pointer is only used when invariants are enabled for various
114 * c - sleep queue chain lock
117 TAILQ_HEAD(, thread) sq_blocked[NR_SLEEPQS]; /* (c) Blocked threads. */
118 u_int sq_blockedcnt[NR_SLEEPQS]; /* (c) N. of blocked threads. */
119 LIST_ENTRY(sleepqueue) sq_hash; /* (c) Chain and free list. */
120 LIST_HEAD(, sleepqueue) sq_free; /* (c) Free queues. */
121 void *sq_wchan; /* (c) Wait channel. */
122 int sq_type; /* (c) Queue type. */
124 struct lock_object *sq_lock; /* (c) Associated lock. */
128 struct sleepqueue_chain {
129 LIST_HEAD(, sleepqueue) sc_queues; /* List of sleep queues. */
130 struct mtx sc_lock; /* Spin lock for this chain. */
131 #ifdef SLEEPQUEUE_PROFILING
132 u_int sc_depth; /* Length of sc_queues. */
133 u_int sc_max_depth; /* Max length of sc_queues. */
137 #ifdef SLEEPQUEUE_PROFILING
138 u_int sleepq_max_depth;
139 static SYSCTL_NODE(_debug, OID_AUTO, sleepq, CTLFLAG_RD, 0, "sleepq profiling");
140 static SYSCTL_NODE(_debug_sleepq, OID_AUTO, chains, CTLFLAG_RD, 0,
141 "sleepq chain stats");
142 SYSCTL_UINT(_debug_sleepq, OID_AUTO, max_depth, CTLFLAG_RD, &sleepq_max_depth,
143 0, "maxmimum depth achieved of a single chain");
145 static void sleepq_profile(const char *wmesg);
146 static int prof_enabled;
148 static struct sleepqueue_chain sleepq_chains[SC_TABLESIZE];
149 static uma_zone_t sleepq_zone;
152 * Prototypes for non-exported routines.
154 static int sleepq_catch_signals(void *wchan, int pri);
155 static int sleepq_check_signals(void);
156 static int sleepq_check_timeout(void);
158 static void sleepq_dtor(void *mem, int size, void *arg);
160 static int sleepq_init(void *mem, int size, int flags);
161 static int sleepq_resume_thread(struct sleepqueue *sq, struct thread *td,
163 static void sleepq_switch(void *wchan, int pri);
164 static void sleepq_timeout(void *arg);
166 SDT_PROBE_DECLARE(sched, , , sleep);
167 SDT_PROBE_DECLARE(sched, , , wakeup);
170 * Early initialization of sleep queues that is called from the sleepinit()
174 init_sleepqueues(void)
176 #ifdef SLEEPQUEUE_PROFILING
177 struct sysctl_oid *chain_oid;
182 for (i = 0; i < SC_TABLESIZE; i++) {
183 LIST_INIT(&sleepq_chains[i].sc_queues);
184 mtx_init(&sleepq_chains[i].sc_lock, "sleepq chain", NULL,
185 MTX_SPIN | MTX_RECURSE);
186 #ifdef SLEEPQUEUE_PROFILING
187 snprintf(chain_name, sizeof(chain_name), "%d", i);
188 chain_oid = SYSCTL_ADD_NODE(NULL,
189 SYSCTL_STATIC_CHILDREN(_debug_sleepq_chains), OID_AUTO,
190 chain_name, CTLFLAG_RD, NULL, "sleepq chain stats");
191 SYSCTL_ADD_UINT(NULL, SYSCTL_CHILDREN(chain_oid), OID_AUTO,
192 "depth", CTLFLAG_RD, &sleepq_chains[i].sc_depth, 0, NULL);
193 SYSCTL_ADD_UINT(NULL, SYSCTL_CHILDREN(chain_oid), OID_AUTO,
194 "max_depth", CTLFLAG_RD, &sleepq_chains[i].sc_max_depth, 0,
198 sleepq_zone = uma_zcreate("SLEEPQUEUE", sizeof(struct sleepqueue),
200 NULL, sleepq_dtor, sleepq_init, NULL, UMA_ALIGN_CACHE, 0);
202 NULL, NULL, sleepq_init, NULL, UMA_ALIGN_CACHE, 0);
205 thread0.td_sleepqueue = sleepq_alloc();
209 * Get a sleep queue for a new thread.
215 return (uma_zalloc(sleepq_zone, M_WAITOK));
219 * Free a sleep queue when a thread is destroyed.
222 sleepq_free(struct sleepqueue *sq)
225 uma_zfree(sleepq_zone, sq);
229 * Lock the sleep queue chain associated with the specified wait channel.
232 sleepq_lock(void *wchan)
234 struct sleepqueue_chain *sc;
236 sc = SC_LOOKUP(wchan);
237 mtx_lock_spin(&sc->sc_lock);
241 * Look up the sleep queue associated with a given wait channel in the hash
242 * table locking the associated sleep queue chain. If no queue is found in
243 * the table, NULL is returned.
246 sleepq_lookup(void *wchan)
248 struct sleepqueue_chain *sc;
249 struct sleepqueue *sq;
251 KASSERT(wchan != NULL, ("%s: invalid NULL wait channel", __func__));
252 sc = SC_LOOKUP(wchan);
253 mtx_assert(&sc->sc_lock, MA_OWNED);
254 LIST_FOREACH(sq, &sc->sc_queues, sq_hash)
255 if (sq->sq_wchan == wchan)
261 * Unlock the sleep queue chain associated with a given wait channel.
264 sleepq_release(void *wchan)
266 struct sleepqueue_chain *sc;
268 sc = SC_LOOKUP(wchan);
269 mtx_unlock_spin(&sc->sc_lock);
273 * Places the current thread on the sleep queue for the specified wait
274 * channel. If INVARIANTS is enabled, then it associates the passed in
275 * lock with the sleepq to make sure it is held when that sleep queue is
279 sleepq_add(void *wchan, struct lock_object *lock, const char *wmesg, int flags,
282 struct sleepqueue_chain *sc;
283 struct sleepqueue *sq;
287 sc = SC_LOOKUP(wchan);
288 mtx_assert(&sc->sc_lock, MA_OWNED);
289 MPASS(td->td_sleepqueue != NULL);
290 MPASS(wchan != NULL);
291 MPASS((queue >= 0) && (queue < NR_SLEEPQS));
293 /* If this thread is not allowed to sleep, die a horrible death. */
294 KASSERT(td->td_no_sleeping == 0,
295 ("%s: td %p to sleep on wchan %p with sleeping prohibited",
296 __func__, td, wchan));
298 /* Look up the sleep queue associated with the wait channel 'wchan'. */
299 sq = sleepq_lookup(wchan);
302 * If the wait channel does not already have a sleep queue, use
303 * this thread's sleep queue. Otherwise, insert the current thread
304 * into the sleep queue already in use by this wait channel.
310 sq = td->td_sleepqueue;
311 for (i = 0; i < NR_SLEEPQS; i++) {
312 KASSERT(TAILQ_EMPTY(&sq->sq_blocked[i]),
313 ("thread's sleep queue %d is not empty", i));
314 KASSERT(sq->sq_blockedcnt[i] == 0,
315 ("thread's sleep queue %d count mismatches", i));
317 KASSERT(LIST_EMPTY(&sq->sq_free),
318 ("thread's sleep queue has a non-empty free list"));
319 KASSERT(sq->sq_wchan == NULL, ("stale sq_wchan pointer"));
322 #ifdef SLEEPQUEUE_PROFILING
324 if (sc->sc_depth > sc->sc_max_depth) {
325 sc->sc_max_depth = sc->sc_depth;
326 if (sc->sc_max_depth > sleepq_max_depth)
327 sleepq_max_depth = sc->sc_max_depth;
330 sq = td->td_sleepqueue;
331 LIST_INSERT_HEAD(&sc->sc_queues, sq, sq_hash);
332 sq->sq_wchan = wchan;
333 sq->sq_type = flags & SLEEPQ_TYPE;
335 MPASS(wchan == sq->sq_wchan);
336 MPASS(lock == sq->sq_lock);
337 MPASS((flags & SLEEPQ_TYPE) == sq->sq_type);
338 LIST_INSERT_HEAD(&sq->sq_free, td->td_sleepqueue, sq_hash);
341 TAILQ_INSERT_TAIL(&sq->sq_blocked[queue], td, td_slpq);
342 sq->sq_blockedcnt[queue]++;
343 td->td_sleepqueue = NULL;
344 td->td_sqqueue = queue;
345 td->td_wchan = wchan;
346 td->td_wmesg = wmesg;
347 if (flags & SLEEPQ_INTERRUPTIBLE) {
348 td->td_flags |= TDF_SINTR;
349 td->td_flags &= ~TDF_SLEEPABORT;
355 * Sets a timeout that will remove the current thread from the specified
356 * sleep queue after timo ticks if the thread has not already been awakened.
359 sleepq_set_timeout_sbt(void *wchan, sbintime_t sbt, sbintime_t pr,
362 struct sleepqueue_chain *sc;
366 sc = SC_LOOKUP(wchan);
367 mtx_assert(&sc->sc_lock, MA_OWNED);
368 MPASS(TD_ON_SLEEPQ(td));
369 MPASS(td->td_sleepqueue == NULL);
370 MPASS(wchan != NULL);
371 callout_reset_sbt_on(&td->td_slpcallout, sbt, pr,
372 sleepq_timeout, td, PCPU_GET(cpuid), flags | C_DIRECT_EXEC);
376 * Return the number of actual sleepers for the specified queue.
379 sleepq_sleepcnt(void *wchan, int queue)
381 struct sleepqueue *sq;
383 KASSERT(wchan != NULL, ("%s: invalid NULL wait channel", __func__));
384 MPASS((queue >= 0) && (queue < NR_SLEEPQS));
385 sq = sleepq_lookup(wchan);
388 return (sq->sq_blockedcnt[queue]);
392 * Marks the pending sleep of the current thread as interruptible and
393 * makes an initial check for pending signals before putting a thread
394 * to sleep. Enters and exits with the thread lock held. Thread lock
395 * may have transitioned from the sleepq lock to a run lock.
398 sleepq_catch_signals(void *wchan, int pri)
400 struct sleepqueue_chain *sc;
401 struct sleepqueue *sq;
409 sc = SC_LOOKUP(wchan);
410 mtx_assert(&sc->sc_lock, MA_OWNED);
411 MPASS(wchan != NULL);
412 if ((td->td_pflags & TDP_WAKEUP) != 0) {
413 td->td_pflags &= ~TDP_WAKEUP;
420 * See if there are any pending signals for this thread. If not
421 * we can switch immediately. Otherwise do the signal processing
425 if ((td->td_flags & (TDF_NEEDSIGCHK | TDF_NEEDSUSPCHK)) == 0) {
426 sleepq_switch(wchan, pri);
430 mtx_unlock_spin(&sc->sc_lock);
431 CTR3(KTR_PROC, "sleepq catching signals: thread %p (pid %ld, %s)",
432 (void *)td, (long)p->p_pid, td->td_name);
435 mtx_lock(&ps->ps_mtx);
438 mtx_unlock(&ps->ps_mtx);
439 ret = thread_suspend_check(1);
440 MPASS(ret == 0 || ret == EINTR || ret == ERESTART);
442 if (SIGISMEMBER(ps->ps_sigintr, sig))
446 mtx_unlock(&ps->ps_mtx);
449 * Lock the per-process spinlock prior to dropping the PROC_LOCK
450 * to avoid a signal delivery race. PROC_LOCK, PROC_SLOCK, and
451 * thread_lock() are currently held in tdsendsignal().
454 mtx_lock_spin(&sc->sc_lock);
459 sleepq_switch(wchan, pri);
464 * There were pending signals and this thread is still
465 * on the sleep queue, remove it from the sleep queue.
467 if (TD_ON_SLEEPQ(td)) {
468 sq = sleepq_lookup(wchan);
469 if (sleepq_resume_thread(sq, td, 0)) {
472 * This thread hasn't gone to sleep yet, so it
473 * should not be swapped out.
475 panic("not waking up swapper");
479 mtx_unlock_spin(&sc->sc_lock);
480 MPASS(td->td_lock != &sc->sc_lock);
485 * Switches to another thread if we are still asleep on a sleep queue.
486 * Returns with thread lock.
489 sleepq_switch(void *wchan, int pri)
491 struct sleepqueue_chain *sc;
492 struct sleepqueue *sq;
496 sc = SC_LOOKUP(wchan);
497 mtx_assert(&sc->sc_lock, MA_OWNED);
498 THREAD_LOCK_ASSERT(td, MA_OWNED);
501 * If we have a sleep queue, then we've already been woken up, so
504 if (td->td_sleepqueue != NULL) {
505 mtx_unlock_spin(&sc->sc_lock);
510 * If TDF_TIMEOUT is set, then our sleep has been timed out
511 * already but we are still on the sleep queue, so dequeue the
514 if (td->td_flags & TDF_TIMEOUT) {
515 MPASS(TD_ON_SLEEPQ(td));
516 sq = sleepq_lookup(wchan);
517 if (sleepq_resume_thread(sq, td, 0)) {
520 * This thread hasn't gone to sleep yet, so it
521 * should not be swapped out.
523 panic("not waking up swapper");
526 mtx_unlock_spin(&sc->sc_lock);
529 #ifdef SLEEPQUEUE_PROFILING
531 sleepq_profile(td->td_wmesg);
533 MPASS(td->td_sleepqueue == NULL);
534 sched_sleep(td, pri);
535 thread_lock_set(td, &sc->sc_lock);
536 SDT_PROBE0(sched, , , sleep);
538 mi_switch(SW_VOL | SWT_SLEEPQ, NULL);
539 KASSERT(TD_IS_RUNNING(td), ("running but not TDS_RUNNING"));
540 CTR3(KTR_PROC, "sleepq resume: thread %p (pid %ld, %s)",
541 (void *)td, (long)td->td_proc->p_pid, (void *)td->td_name);
545 * Check to see if we timed out.
548 sleepq_check_timeout(void)
553 THREAD_LOCK_ASSERT(td, MA_OWNED);
556 * If TDF_TIMEOUT is set, we timed out.
558 if (td->td_flags & TDF_TIMEOUT) {
559 td->td_flags &= ~TDF_TIMEOUT;
560 return (EWOULDBLOCK);
564 * If TDF_TIMOFAIL is set, the timeout ran after we had
565 * already been woken up.
567 if (td->td_flags & TDF_TIMOFAIL)
568 td->td_flags &= ~TDF_TIMOFAIL;
571 * If callout_stop() fails, then the timeout is running on
572 * another CPU, so synchronize with it to avoid having it
573 * accidentally wake up a subsequent sleep.
575 else if (callout_stop(&td->td_slpcallout) == 0) {
576 td->td_flags |= TDF_TIMEOUT;
578 mi_switch(SW_INVOL | SWT_SLEEPQTIMO, NULL);
584 * Check to see if we were awoken by a signal.
587 sleepq_check_signals(void)
592 THREAD_LOCK_ASSERT(td, MA_OWNED);
594 /* We are no longer in an interruptible sleep. */
595 if (td->td_flags & TDF_SINTR)
596 td->td_flags &= ~TDF_SINTR;
598 if (td->td_flags & TDF_SLEEPABORT) {
599 td->td_flags &= ~TDF_SLEEPABORT;
600 return (td->td_intrval);
607 * Block the current thread until it is awakened from its sleep queue.
610 sleepq_wait(void *wchan, int pri)
615 MPASS(!(td->td_flags & TDF_SINTR));
617 sleepq_switch(wchan, pri);
622 * Block the current thread until it is awakened from its sleep queue
623 * or it is interrupted by a signal.
626 sleepq_wait_sig(void *wchan, int pri)
631 rcatch = sleepq_catch_signals(wchan, pri);
632 rval = sleepq_check_signals();
633 thread_unlock(curthread);
640 * Block the current thread until it is awakened from its sleep queue
641 * or it times out while waiting.
644 sleepq_timedwait(void *wchan, int pri)
650 MPASS(!(td->td_flags & TDF_SINTR));
652 sleepq_switch(wchan, pri);
653 rval = sleepq_check_timeout();
660 * Block the current thread until it is awakened from its sleep queue,
661 * it is interrupted by a signal, or it times out waiting to be awakened.
664 sleepq_timedwait_sig(void *wchan, int pri)
666 int rcatch, rvalt, rvals;
668 rcatch = sleepq_catch_signals(wchan, pri);
669 rvalt = sleepq_check_timeout();
670 rvals = sleepq_check_signals();
671 thread_unlock(curthread);
680 * Returns the type of sleepqueue given a waitchannel.
683 sleepq_type(void *wchan)
685 struct sleepqueue *sq;
688 MPASS(wchan != NULL);
691 sq = sleepq_lookup(wchan);
693 sleepq_release(wchan);
697 sleepq_release(wchan);
702 * Removes a thread from a sleep queue and makes it
706 sleepq_resume_thread(struct sleepqueue *sq, struct thread *td, int pri)
708 struct sleepqueue_chain *sc;
711 MPASS(sq->sq_wchan != NULL);
712 MPASS(td->td_wchan == sq->sq_wchan);
713 MPASS(td->td_sqqueue < NR_SLEEPQS && td->td_sqqueue >= 0);
714 THREAD_LOCK_ASSERT(td, MA_OWNED);
715 sc = SC_LOOKUP(sq->sq_wchan);
716 mtx_assert(&sc->sc_lock, MA_OWNED);
718 SDT_PROBE2(sched, , , wakeup, td, td->td_proc);
720 /* Remove the thread from the queue. */
721 sq->sq_blockedcnt[td->td_sqqueue]--;
722 TAILQ_REMOVE(&sq->sq_blocked[td->td_sqqueue], td, td_slpq);
725 * Get a sleep queue for this thread. If this is the last waiter,
726 * use the queue itself and take it out of the chain, otherwise,
727 * remove a queue from the free list.
729 if (LIST_EMPTY(&sq->sq_free)) {
730 td->td_sleepqueue = sq;
734 #ifdef SLEEPQUEUE_PROFILING
738 td->td_sleepqueue = LIST_FIRST(&sq->sq_free);
739 LIST_REMOVE(td->td_sleepqueue, sq_hash);
743 td->td_flags &= ~TDF_SINTR;
745 CTR3(KTR_PROC, "sleepq_wakeup: thread %p (pid %ld, %s)",
746 (void *)td, (long)td->td_proc->p_pid, td->td_name);
748 /* Adjust priority if requested. */
749 MPASS(pri == 0 || (pri >= PRI_MIN && pri <= PRI_MAX));
750 if (pri != 0 && td->td_priority > pri &&
751 PRI_BASE(td->td_pri_class) == PRI_TIMESHARE)
755 * Note that thread td might not be sleeping if it is running
756 * sleepq_catch_signals() on another CPU or is blocked on its
757 * proc lock to check signals. There's no need to mark the
758 * thread runnable in that case.
760 if (TD_IS_SLEEPING(td)) {
762 return (setrunnable(td));
769 * UMA zone item deallocator.
772 sleepq_dtor(void *mem, int size, void *arg)
774 struct sleepqueue *sq;
778 for (i = 0; i < NR_SLEEPQS; i++) {
779 MPASS(TAILQ_EMPTY(&sq->sq_blocked[i]));
780 MPASS(sq->sq_blockedcnt[i] == 0);
786 * UMA zone item initializer.
789 sleepq_init(void *mem, int size, int flags)
791 struct sleepqueue *sq;
796 for (i = 0; i < NR_SLEEPQS; i++) {
797 TAILQ_INIT(&sq->sq_blocked[i]);
798 sq->sq_blockedcnt[i] = 0;
800 LIST_INIT(&sq->sq_free);
805 * Find the highest priority thread sleeping on a wait channel and resume it.
808 sleepq_signal(void *wchan, int flags, int pri, int queue)
810 struct sleepqueue *sq;
811 struct thread *td, *besttd;
814 CTR2(KTR_PROC, "sleepq_signal(%p, %d)", wchan, flags);
815 KASSERT(wchan != NULL, ("%s: invalid NULL wait channel", __func__));
816 MPASS((queue >= 0) && (queue < NR_SLEEPQS));
817 sq = sleepq_lookup(wchan);
820 KASSERT(sq->sq_type == (flags & SLEEPQ_TYPE),
821 ("%s: mismatch between sleep/wakeup and cv_*", __func__));
824 * Find the highest priority thread on the queue. If there is a
825 * tie, use the thread that first appears in the queue as it has
826 * been sleeping the longest since threads are always added to
827 * the tail of sleep queues.
830 TAILQ_FOREACH(td, &sq->sq_blocked[queue], td_slpq) {
831 if (besttd == NULL || td->td_priority < besttd->td_priority)
834 MPASS(besttd != NULL);
836 wakeup_swapper = sleepq_resume_thread(sq, besttd, pri);
837 thread_unlock(besttd);
838 return (wakeup_swapper);
842 * Resume all threads sleeping on a specified wait channel.
845 sleepq_broadcast(void *wchan, int flags, int pri, int queue)
847 struct sleepqueue *sq;
848 struct thread *td, *tdn;
851 CTR2(KTR_PROC, "sleepq_broadcast(%p, %d)", wchan, flags);
852 KASSERT(wchan != NULL, ("%s: invalid NULL wait channel", __func__));
853 MPASS((queue >= 0) && (queue < NR_SLEEPQS));
854 sq = sleepq_lookup(wchan);
857 KASSERT(sq->sq_type == (flags & SLEEPQ_TYPE),
858 ("%s: mismatch between sleep/wakeup and cv_*", __func__));
860 /* Resume all blocked threads on the sleep queue. */
862 TAILQ_FOREACH_SAFE(td, &sq->sq_blocked[queue], td_slpq, tdn) {
864 if (sleepq_resume_thread(sq, td, pri))
868 return (wakeup_swapper);
872 * Time sleeping threads out. When the timeout expires, the thread is
873 * removed from the sleep queue and made runnable if it is still asleep.
876 sleepq_timeout(void *arg)
878 struct sleepqueue_chain *sc;
879 struct sleepqueue *sq;
886 CTR3(KTR_PROC, "sleepq_timeout: thread %p (pid %ld, %s)",
887 (void *)td, (long)td->td_proc->p_pid, (void *)td->td_name);
890 * First, see if the thread is asleep and get the wait channel if
894 if (TD_IS_SLEEPING(td) && TD_ON_SLEEPQ(td)) {
895 wchan = td->td_wchan;
896 sc = SC_LOOKUP(wchan);
897 THREAD_LOCKPTR_ASSERT(td, &sc->sc_lock);
898 sq = sleepq_lookup(wchan);
900 td->td_flags |= TDF_TIMEOUT;
901 wakeup_swapper = sleepq_resume_thread(sq, td, 0);
909 * If the thread is on the SLEEPQ but isn't sleeping yet, it
910 * can either be on another CPU in between sleepq_add() and
911 * one of the sleepq_*wait*() routines or it can be in
912 * sleepq_catch_signals().
914 if (TD_ON_SLEEPQ(td)) {
915 td->td_flags |= TDF_TIMEOUT;
921 * Now check for the edge cases. First, if TDF_TIMEOUT is set,
922 * then the other thread has already yielded to us, so clear
923 * the flag and resume it. If TDF_TIMEOUT is not set, then the
924 * we know that the other thread is not on a sleep queue, but it
925 * hasn't resumed execution yet. In that case, set TDF_TIMOFAIL
926 * to let it know that the timeout has already run and doesn't
927 * need to be canceled.
929 if (td->td_flags & TDF_TIMEOUT) {
930 MPASS(TD_IS_SLEEPING(td));
931 td->td_flags &= ~TDF_TIMEOUT;
933 wakeup_swapper = setrunnable(td);
935 td->td_flags |= TDF_TIMOFAIL;
942 * Resumes a specific thread from the sleep queue associated with a specific
943 * wait channel if it is on that queue.
946 sleepq_remove(struct thread *td, void *wchan)
948 struct sleepqueue *sq;
952 * Look up the sleep queue for this wait channel, then re-check
953 * that the thread is asleep on that channel, if it is not, then
956 MPASS(wchan != NULL);
958 sq = sleepq_lookup(wchan);
960 * We can not lock the thread here as it may be sleeping on a
961 * different sleepq. However, holding the sleepq lock for this
962 * wchan can guarantee that we do not miss a wakeup for this
963 * channel. The asserts below will catch any false positives.
965 if (!TD_ON_SLEEPQ(td) || td->td_wchan != wchan) {
966 sleepq_release(wchan);
969 /* Thread is asleep on sleep queue sq, so wake it up. */
972 MPASS(td->td_wchan == wchan);
973 wakeup_swapper = sleepq_resume_thread(sq, td, 0);
975 sleepq_release(wchan);
981 * Abort a thread as if an interrupt had occurred. Only abort
982 * interruptible waits (unfortunately it isn't safe to abort others).
985 sleepq_abort(struct thread *td, int intrval)
987 struct sleepqueue *sq;
990 THREAD_LOCK_ASSERT(td, MA_OWNED);
991 MPASS(TD_ON_SLEEPQ(td));
992 MPASS(td->td_flags & TDF_SINTR);
993 MPASS(intrval == EINTR || intrval == ERESTART);
996 * If the TDF_TIMEOUT flag is set, just leave. A
997 * timeout is scheduled anyhow.
999 if (td->td_flags & TDF_TIMEOUT)
1002 CTR3(KTR_PROC, "sleepq_abort: thread %p (pid %ld, %s)",
1003 (void *)td, (long)td->td_proc->p_pid, (void *)td->td_name);
1004 td->td_intrval = intrval;
1005 td->td_flags |= TDF_SLEEPABORT;
1007 * If the thread has not slept yet it will find the signal in
1008 * sleepq_catch_signals() and call sleepq_resume_thread. Otherwise
1009 * we have to do it here.
1011 if (!TD_IS_SLEEPING(td))
1013 wchan = td->td_wchan;
1014 MPASS(wchan != NULL);
1015 sq = sleepq_lookup(wchan);
1018 /* Thread is asleep on sleep queue sq, so wake it up. */
1019 return (sleepq_resume_thread(sq, td, 0));
1022 #ifdef SLEEPQUEUE_PROFILING
1023 #define SLEEPQ_PROF_LOCATIONS 1024
1024 #define SLEEPQ_SBUFSIZE 512
1025 struct sleepq_prof {
1026 LIST_ENTRY(sleepq_prof) sp_link;
1027 const char *sp_wmesg;
1031 LIST_HEAD(sqphead, sleepq_prof);
1033 struct sqphead sleepq_prof_free;
1034 struct sqphead sleepq_hash[SC_TABLESIZE];
1035 static struct sleepq_prof sleepq_profent[SLEEPQ_PROF_LOCATIONS];
1036 static struct mtx sleepq_prof_lock;
1037 MTX_SYSINIT(sleepq_prof_lock, &sleepq_prof_lock, "sleepq_prof", MTX_SPIN);
1040 sleepq_profile(const char *wmesg)
1042 struct sleepq_prof *sp;
1044 mtx_lock_spin(&sleepq_prof_lock);
1045 if (prof_enabled == 0)
1047 LIST_FOREACH(sp, &sleepq_hash[SC_HASH(wmesg)], sp_link)
1048 if (sp->sp_wmesg == wmesg)
1050 sp = LIST_FIRST(&sleepq_prof_free);
1053 sp->sp_wmesg = wmesg;
1054 LIST_REMOVE(sp, sp_link);
1055 LIST_INSERT_HEAD(&sleepq_hash[SC_HASH(wmesg)], sp, sp_link);
1059 mtx_unlock_spin(&sleepq_prof_lock);
1064 sleepq_prof_reset(void)
1066 struct sleepq_prof *sp;
1070 mtx_lock_spin(&sleepq_prof_lock);
1071 enabled = prof_enabled;
1073 for (i = 0; i < SC_TABLESIZE; i++)
1074 LIST_INIT(&sleepq_hash[i]);
1075 LIST_INIT(&sleepq_prof_free);
1076 for (i = 0; i < SLEEPQ_PROF_LOCATIONS; i++) {
1077 sp = &sleepq_profent[i];
1078 sp->sp_wmesg = NULL;
1080 LIST_INSERT_HEAD(&sleepq_prof_free, sp, sp_link);
1082 prof_enabled = enabled;
1083 mtx_unlock_spin(&sleepq_prof_lock);
1087 enable_sleepq_prof(SYSCTL_HANDLER_ARGS)
1092 error = sysctl_handle_int(oidp, &v, v, req);
1095 if (req->newptr == NULL)
1097 if (v == prof_enabled)
1100 sleepq_prof_reset();
1101 mtx_lock_spin(&sleepq_prof_lock);
1103 mtx_unlock_spin(&sleepq_prof_lock);
1109 reset_sleepq_prof_stats(SYSCTL_HANDLER_ARGS)
1114 error = sysctl_handle_int(oidp, &v, 0, req);
1117 if (req->newptr == NULL)
1121 sleepq_prof_reset();
1127 dump_sleepq_prof_stats(SYSCTL_HANDLER_ARGS)
1129 struct sleepq_prof *sp;
1135 error = sysctl_wire_old_buffer(req, 0);
1138 sb = sbuf_new_for_sysctl(NULL, NULL, SLEEPQ_SBUFSIZE, req);
1139 sbuf_printf(sb, "\nwmesg\tcount\n");
1140 enabled = prof_enabled;
1141 mtx_lock_spin(&sleepq_prof_lock);
1143 mtx_unlock_spin(&sleepq_prof_lock);
1144 for (i = 0; i < SC_TABLESIZE; i++) {
1145 LIST_FOREACH(sp, &sleepq_hash[i], sp_link) {
1146 sbuf_printf(sb, "%s\t%ld\n",
1147 sp->sp_wmesg, sp->sp_count);
1150 mtx_lock_spin(&sleepq_prof_lock);
1151 prof_enabled = enabled;
1152 mtx_unlock_spin(&sleepq_prof_lock);
1154 error = sbuf_finish(sb);
1159 SYSCTL_PROC(_debug_sleepq, OID_AUTO, stats, CTLTYPE_STRING | CTLFLAG_RD,
1160 NULL, 0, dump_sleepq_prof_stats, "A", "Sleepqueue profiling statistics");
1161 SYSCTL_PROC(_debug_sleepq, OID_AUTO, reset, CTLTYPE_INT | CTLFLAG_RW,
1162 NULL, 0, reset_sleepq_prof_stats, "I",
1163 "Reset sleepqueue profiling statistics");
1164 SYSCTL_PROC(_debug_sleepq, OID_AUTO, enable, CTLTYPE_INT | CTLFLAG_RW,
1165 NULL, 0, enable_sleepq_prof, "I", "Enable sleepqueue profiling");
1169 DB_SHOW_COMMAND(sleepq, db_show_sleepqueue)
1171 struct sleepqueue_chain *sc;
1172 struct sleepqueue *sq;
1174 struct lock_object *lock;
1184 * First, see if there is an active sleep queue for the wait channel
1185 * indicated by the address.
1187 wchan = (void *)addr;
1188 sc = SC_LOOKUP(wchan);
1189 LIST_FOREACH(sq, &sc->sc_queues, sq_hash)
1190 if (sq->sq_wchan == wchan)
1194 * Second, see if there is an active sleep queue at the address
1197 for (i = 0; i < SC_TABLESIZE; i++)
1198 LIST_FOREACH(sq, &sleepq_chains[i].sc_queues, sq_hash) {
1199 if (sq == (struct sleepqueue *)addr)
1203 db_printf("Unable to locate a sleep queue via %p\n", (void *)addr);
1206 db_printf("Wait channel: %p\n", sq->sq_wchan);
1207 db_printf("Queue type: %d\n", sq->sq_type);
1211 db_printf("Associated Interlock: %p - (%s) %s\n", lock,
1212 LOCK_CLASS(lock)->lc_name, lock->lo_name);
1215 db_printf("Blocked threads:\n");
1216 for (i = 0; i < NR_SLEEPQS; i++) {
1217 db_printf("\nQueue[%d]:\n", i);
1218 if (TAILQ_EMPTY(&sq->sq_blocked[i]))
1219 db_printf("\tempty\n");
1221 TAILQ_FOREACH(td, &sq->sq_blocked[0],
1223 db_printf("\t%p (tid %d, pid %d, \"%s\")\n", td,
1224 td->td_tid, td->td_proc->p_pid,
1227 db_printf("(expected: %u)\n", sq->sq_blockedcnt[i]);
1231 /* Alias 'show sleepqueue' to 'show sleepq'. */
1232 DB_SHOW_ALIAS(sleepqueue, db_show_sleepqueue);