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_sched.h"
65 #include "opt_stack.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/stack.h>
80 #include <sys/sysctl.h>
90 * Constants for the hash table of sleep queue chains.
91 * SC_TABLESIZE must be a power of two for SC_MASK to work properly.
93 #define SC_TABLESIZE 256 /* Must be power of 2. */
94 #define SC_MASK (SC_TABLESIZE - 1)
96 #define SC_HASH(wc) ((((uintptr_t)(wc) >> SC_SHIFT) ^ (uintptr_t)(wc)) & \
98 #define SC_LOOKUP(wc) &sleepq_chains[SC_HASH(wc)]
101 * There two different lists of sleep queues. Both lists are connected
102 * via the sq_hash entries. The first list is the sleep queue chain list
103 * that a sleep queue is on when it is attached to a wait channel. The
104 * second list is the free list hung off of a sleep queue that is attached
107 * Each sleep queue also contains the wait channel it is attached to, the
108 * list of threads blocked on that wait channel, flags specific to the
109 * wait channel, and the lock used to synchronize with a wait channel.
110 * The flags are used to catch mismatches between the various consumers
111 * of the sleep queue API (e.g. sleep/wakeup and condition variables).
112 * The lock pointer is only used when invariants are enabled for various
116 * c - sleep queue chain lock
119 TAILQ_HEAD(, thread) sq_blocked[NR_SLEEPQS]; /* (c) Blocked threads. */
120 u_int sq_blockedcnt[NR_SLEEPQS]; /* (c) N. of blocked threads. */
121 LIST_ENTRY(sleepqueue) sq_hash; /* (c) Chain and free list. */
122 LIST_HEAD(, sleepqueue) sq_free; /* (c) Free queues. */
123 void *sq_wchan; /* (c) Wait channel. */
124 int sq_type; /* (c) Queue type. */
126 struct lock_object *sq_lock; /* (c) Associated lock. */
130 struct sleepqueue_chain {
131 LIST_HEAD(, sleepqueue) sc_queues; /* List of sleep queues. */
132 struct mtx sc_lock; /* Spin lock for this chain. */
133 #ifdef SLEEPQUEUE_PROFILING
134 u_int sc_depth; /* Length of sc_queues. */
135 u_int sc_max_depth; /* Max length of sc_queues. */
139 #ifdef SLEEPQUEUE_PROFILING
140 u_int sleepq_max_depth;
141 static SYSCTL_NODE(_debug, OID_AUTO, sleepq, CTLFLAG_RD, 0, "sleepq profiling");
142 static SYSCTL_NODE(_debug_sleepq, OID_AUTO, chains, CTLFLAG_RD, 0,
143 "sleepq chain stats");
144 SYSCTL_UINT(_debug_sleepq, OID_AUTO, max_depth, CTLFLAG_RD, &sleepq_max_depth,
145 0, "maxmimum depth achieved of a single chain");
147 static void sleepq_profile(const char *wmesg);
148 static int prof_enabled;
150 static struct sleepqueue_chain sleepq_chains[SC_TABLESIZE];
151 static uma_zone_t sleepq_zone;
154 * Prototypes for non-exported routines.
156 static int sleepq_catch_signals(void *wchan, int pri);
157 static int sleepq_check_signals(void);
158 static int sleepq_check_timeout(void);
160 static void sleepq_dtor(void *mem, int size, void *arg);
162 static int sleepq_init(void *mem, int size, int flags);
163 static int sleepq_resume_thread(struct sleepqueue *sq, struct thread *td,
165 static void sleepq_switch(void *wchan, int pri);
166 static void sleepq_timeout(void *arg);
168 SDT_PROBE_DECLARE(sched, , , sleep);
169 SDT_PROBE_DECLARE(sched, , , wakeup);
172 * Initialize SLEEPQUEUE_PROFILING specific sysctl nodes.
173 * Note that it must happen after sleepinit() has been fully executed, so
174 * it must happen after SI_SUB_KMEM SYSINIT() subsystem setup.
176 #ifdef SLEEPQUEUE_PROFILING
178 init_sleepqueue_profiling(void)
181 struct sysctl_oid *chain_oid;
184 for (i = 0; i < SC_TABLESIZE; i++) {
185 snprintf(chain_name, sizeof(chain_name), "%u", i);
186 chain_oid = SYSCTL_ADD_NODE(NULL,
187 SYSCTL_STATIC_CHILDREN(_debug_sleepq_chains), OID_AUTO,
188 chain_name, CTLFLAG_RD, NULL, "sleepq chain stats");
189 SYSCTL_ADD_UINT(NULL, SYSCTL_CHILDREN(chain_oid), OID_AUTO,
190 "depth", CTLFLAG_RD, &sleepq_chains[i].sc_depth, 0, NULL);
191 SYSCTL_ADD_UINT(NULL, SYSCTL_CHILDREN(chain_oid), OID_AUTO,
192 "max_depth", CTLFLAG_RD, &sleepq_chains[i].sc_max_depth, 0,
197 SYSINIT(sleepqueue_profiling, SI_SUB_LOCK, SI_ORDER_ANY,
198 init_sleepqueue_profiling, NULL);
202 * Early initialization of sleep queues that is called from the sleepinit()
206 init_sleepqueues(void)
210 for (i = 0; i < SC_TABLESIZE; i++) {
211 LIST_INIT(&sleepq_chains[i].sc_queues);
212 mtx_init(&sleepq_chains[i].sc_lock, "sleepq chain", NULL,
213 MTX_SPIN | MTX_RECURSE);
215 sleepq_zone = uma_zcreate("SLEEPQUEUE", sizeof(struct sleepqueue),
217 NULL, sleepq_dtor, sleepq_init, NULL, UMA_ALIGN_CACHE, 0);
219 NULL, NULL, sleepq_init, NULL, UMA_ALIGN_CACHE, 0);
222 thread0.td_sleepqueue = sleepq_alloc();
226 * Get a sleep queue for a new thread.
232 return (uma_zalloc(sleepq_zone, M_WAITOK));
236 * Free a sleep queue when a thread is destroyed.
239 sleepq_free(struct sleepqueue *sq)
242 uma_zfree(sleepq_zone, sq);
246 * Lock the sleep queue chain associated with the specified wait channel.
249 sleepq_lock(void *wchan)
251 struct sleepqueue_chain *sc;
253 sc = SC_LOOKUP(wchan);
254 mtx_lock_spin(&sc->sc_lock);
258 * Look up the sleep queue associated with a given wait channel in the hash
259 * table locking the associated sleep queue chain. If no queue is found in
260 * the table, NULL is returned.
263 sleepq_lookup(void *wchan)
265 struct sleepqueue_chain *sc;
266 struct sleepqueue *sq;
268 KASSERT(wchan != NULL, ("%s: invalid NULL wait channel", __func__));
269 sc = SC_LOOKUP(wchan);
270 mtx_assert(&sc->sc_lock, MA_OWNED);
271 LIST_FOREACH(sq, &sc->sc_queues, sq_hash)
272 if (sq->sq_wchan == wchan)
278 * Unlock the sleep queue chain associated with a given wait channel.
281 sleepq_release(void *wchan)
283 struct sleepqueue_chain *sc;
285 sc = SC_LOOKUP(wchan);
286 mtx_unlock_spin(&sc->sc_lock);
290 * Places the current thread on the sleep queue for the specified wait
291 * channel. If INVARIANTS is enabled, then it associates the passed in
292 * lock with the sleepq to make sure it is held when that sleep queue is
296 sleepq_add(void *wchan, struct lock_object *lock, const char *wmesg, int flags,
299 struct sleepqueue_chain *sc;
300 struct sleepqueue *sq;
304 sc = SC_LOOKUP(wchan);
305 mtx_assert(&sc->sc_lock, MA_OWNED);
306 MPASS(td->td_sleepqueue != NULL);
307 MPASS(wchan != NULL);
308 MPASS((queue >= 0) && (queue < NR_SLEEPQS));
310 /* If this thread is not allowed to sleep, die a horrible death. */
311 KASSERT(td->td_no_sleeping == 0,
312 ("%s: td %p to sleep on wchan %p with sleeping prohibited",
313 __func__, td, wchan));
315 /* Look up the sleep queue associated with the wait channel 'wchan'. */
316 sq = sleepq_lookup(wchan);
319 * If the wait channel does not already have a sleep queue, use
320 * this thread's sleep queue. Otherwise, insert the current thread
321 * into the sleep queue already in use by this wait channel.
327 sq = td->td_sleepqueue;
328 for (i = 0; i < NR_SLEEPQS; i++) {
329 KASSERT(TAILQ_EMPTY(&sq->sq_blocked[i]),
330 ("thread's sleep queue %d is not empty", i));
331 KASSERT(sq->sq_blockedcnt[i] == 0,
332 ("thread's sleep queue %d count mismatches", i));
334 KASSERT(LIST_EMPTY(&sq->sq_free),
335 ("thread's sleep queue has a non-empty free list"));
336 KASSERT(sq->sq_wchan == NULL, ("stale sq_wchan pointer"));
339 #ifdef SLEEPQUEUE_PROFILING
341 if (sc->sc_depth > sc->sc_max_depth) {
342 sc->sc_max_depth = sc->sc_depth;
343 if (sc->sc_max_depth > sleepq_max_depth)
344 sleepq_max_depth = sc->sc_max_depth;
347 sq = td->td_sleepqueue;
348 LIST_INSERT_HEAD(&sc->sc_queues, sq, sq_hash);
349 sq->sq_wchan = wchan;
350 sq->sq_type = flags & SLEEPQ_TYPE;
352 MPASS(wchan == sq->sq_wchan);
353 MPASS(lock == sq->sq_lock);
354 MPASS((flags & SLEEPQ_TYPE) == sq->sq_type);
355 LIST_INSERT_HEAD(&sq->sq_free, td->td_sleepqueue, sq_hash);
358 TAILQ_INSERT_TAIL(&sq->sq_blocked[queue], td, td_slpq);
359 sq->sq_blockedcnt[queue]++;
360 td->td_sleepqueue = NULL;
361 td->td_sqqueue = queue;
362 td->td_wchan = wchan;
363 td->td_wmesg = wmesg;
364 if (flags & SLEEPQ_INTERRUPTIBLE) {
365 td->td_flags |= TDF_SINTR;
366 td->td_flags &= ~TDF_SLEEPABORT;
372 * Sets a timeout that will remove the current thread from the specified
373 * sleep queue after timo ticks if the thread has not already been awakened.
376 sleepq_set_timeout_sbt(void *wchan, sbintime_t sbt, sbintime_t pr,
379 struct sleepqueue_chain *sc;
383 sc = SC_LOOKUP(wchan);
384 mtx_assert(&sc->sc_lock, MA_OWNED);
385 MPASS(TD_ON_SLEEPQ(td));
386 MPASS(td->td_sleepqueue == NULL);
387 MPASS(wchan != NULL);
388 callout_reset_sbt_on(&td->td_slpcallout, sbt, pr,
389 sleepq_timeout, td, PCPU_GET(cpuid), flags | C_DIRECT_EXEC);
393 * Return the number of actual sleepers for the specified queue.
396 sleepq_sleepcnt(void *wchan, int queue)
398 struct sleepqueue *sq;
400 KASSERT(wchan != NULL, ("%s: invalid NULL wait channel", __func__));
401 MPASS((queue >= 0) && (queue < NR_SLEEPQS));
402 sq = sleepq_lookup(wchan);
405 return (sq->sq_blockedcnt[queue]);
409 * Marks the pending sleep of the current thread as interruptible and
410 * makes an initial check for pending signals before putting a thread
411 * to sleep. Enters and exits with the thread lock held. Thread lock
412 * may have transitioned from the sleepq lock to a run lock.
415 sleepq_catch_signals(void *wchan, int pri)
417 struct sleepqueue_chain *sc;
418 struct sleepqueue *sq;
426 sc = SC_LOOKUP(wchan);
427 mtx_assert(&sc->sc_lock, MA_OWNED);
428 MPASS(wchan != NULL);
429 if ((td->td_pflags & TDP_WAKEUP) != 0) {
430 td->td_pflags &= ~TDP_WAKEUP;
437 * See if there are any pending signals for this thread. If not
438 * we can switch immediately. Otherwise do the signal processing
442 if ((td->td_flags & (TDF_NEEDSIGCHK | TDF_NEEDSUSPCHK)) == 0) {
443 sleepq_switch(wchan, pri);
447 mtx_unlock_spin(&sc->sc_lock);
448 CTR3(KTR_PROC, "sleepq catching signals: thread %p (pid %ld, %s)",
449 (void *)td, (long)p->p_pid, td->td_name);
452 mtx_lock(&ps->ps_mtx);
455 mtx_unlock(&ps->ps_mtx);
456 ret = thread_suspend_check(1);
457 MPASS(ret == 0 || ret == EINTR || ret == ERESTART);
459 if (SIGISMEMBER(ps->ps_sigintr, sig))
463 mtx_unlock(&ps->ps_mtx);
466 * Lock the per-process spinlock prior to dropping the PROC_LOCK
467 * to avoid a signal delivery race. PROC_LOCK, PROC_SLOCK, and
468 * thread_lock() are currently held in tdsendsignal().
471 mtx_lock_spin(&sc->sc_lock);
476 sleepq_switch(wchan, pri);
481 * There were pending signals and this thread is still
482 * on the sleep queue, remove it from the sleep queue.
484 if (TD_ON_SLEEPQ(td)) {
485 sq = sleepq_lookup(wchan);
486 if (sleepq_resume_thread(sq, td, 0)) {
489 * This thread hasn't gone to sleep yet, so it
490 * should not be swapped out.
492 panic("not waking up swapper");
496 mtx_unlock_spin(&sc->sc_lock);
497 MPASS(td->td_lock != &sc->sc_lock);
502 * Switches to another thread if we are still asleep on a sleep queue.
503 * Returns with thread lock.
506 sleepq_switch(void *wchan, int pri)
508 struct sleepqueue_chain *sc;
509 struct sleepqueue *sq;
513 sc = SC_LOOKUP(wchan);
514 mtx_assert(&sc->sc_lock, MA_OWNED);
515 THREAD_LOCK_ASSERT(td, MA_OWNED);
518 * If we have a sleep queue, then we've already been woken up, so
521 if (td->td_sleepqueue != NULL) {
522 mtx_unlock_spin(&sc->sc_lock);
527 * If TDF_TIMEOUT is set, then our sleep has been timed out
528 * already but we are still on the sleep queue, so dequeue the
531 if (td->td_flags & TDF_TIMEOUT) {
532 MPASS(TD_ON_SLEEPQ(td));
533 sq = sleepq_lookup(wchan);
534 if (sleepq_resume_thread(sq, td, 0)) {
537 * This thread hasn't gone to sleep yet, so it
538 * should not be swapped out.
540 panic("not waking up swapper");
543 mtx_unlock_spin(&sc->sc_lock);
546 #ifdef SLEEPQUEUE_PROFILING
548 sleepq_profile(td->td_wmesg);
550 MPASS(td->td_sleepqueue == NULL);
551 sched_sleep(td, pri);
552 thread_lock_set(td, &sc->sc_lock);
553 SDT_PROBE0(sched, , , sleep);
555 mi_switch(SW_VOL | SWT_SLEEPQ, NULL);
556 KASSERT(TD_IS_RUNNING(td), ("running but not TDS_RUNNING"));
557 CTR3(KTR_PROC, "sleepq resume: thread %p (pid %ld, %s)",
558 (void *)td, (long)td->td_proc->p_pid, (void *)td->td_name);
562 * Check to see if we timed out.
565 sleepq_check_timeout(void)
570 THREAD_LOCK_ASSERT(td, MA_OWNED);
573 * If TDF_TIMEOUT is set, we timed out.
575 if (td->td_flags & TDF_TIMEOUT) {
576 td->td_flags &= ~TDF_TIMEOUT;
577 return (EWOULDBLOCK);
581 * If TDF_TIMOFAIL is set, the timeout ran after we had
582 * already been woken up.
584 if (td->td_flags & TDF_TIMOFAIL)
585 td->td_flags &= ~TDF_TIMOFAIL;
588 * If callout_stop() fails, then the timeout is running on
589 * another CPU, so synchronize with it to avoid having it
590 * accidentally wake up a subsequent sleep.
592 else if (_callout_stop_safe(&td->td_slpcallout, CS_MIGRBLOCK, NULL)
594 td->td_flags |= TDF_TIMEOUT;
596 mi_switch(SW_INVOL | SWT_SLEEPQTIMO, NULL);
602 * Check to see if we were awoken by a signal.
605 sleepq_check_signals(void)
610 THREAD_LOCK_ASSERT(td, MA_OWNED);
612 /* We are no longer in an interruptible sleep. */
613 if (td->td_flags & TDF_SINTR)
614 td->td_flags &= ~TDF_SINTR;
616 if (td->td_flags & TDF_SLEEPABORT) {
617 td->td_flags &= ~TDF_SLEEPABORT;
618 return (td->td_intrval);
625 * Block the current thread until it is awakened from its sleep queue.
628 sleepq_wait(void *wchan, int pri)
633 MPASS(!(td->td_flags & TDF_SINTR));
635 sleepq_switch(wchan, pri);
640 * Block the current thread until it is awakened from its sleep queue
641 * or it is interrupted by a signal.
644 sleepq_wait_sig(void *wchan, int pri)
649 rcatch = sleepq_catch_signals(wchan, pri);
650 rval = sleepq_check_signals();
651 thread_unlock(curthread);
658 * Block the current thread until it is awakened from its sleep queue
659 * or it times out while waiting.
662 sleepq_timedwait(void *wchan, int pri)
668 MPASS(!(td->td_flags & TDF_SINTR));
670 sleepq_switch(wchan, pri);
671 rval = sleepq_check_timeout();
678 * Block the current thread until it is awakened from its sleep queue,
679 * it is interrupted by a signal, or it times out waiting to be awakened.
682 sleepq_timedwait_sig(void *wchan, int pri)
684 int rcatch, rvalt, rvals;
686 rcatch = sleepq_catch_signals(wchan, pri);
687 rvalt = sleepq_check_timeout();
688 rvals = sleepq_check_signals();
689 thread_unlock(curthread);
698 * Returns the type of sleepqueue given a waitchannel.
701 sleepq_type(void *wchan)
703 struct sleepqueue *sq;
706 MPASS(wchan != NULL);
709 sq = sleepq_lookup(wchan);
711 sleepq_release(wchan);
715 sleepq_release(wchan);
720 * Removes a thread from a sleep queue and makes it
724 sleepq_resume_thread(struct sleepqueue *sq, struct thread *td, int pri)
726 struct sleepqueue_chain *sc;
729 MPASS(sq->sq_wchan != NULL);
730 MPASS(td->td_wchan == sq->sq_wchan);
731 MPASS(td->td_sqqueue < NR_SLEEPQS && td->td_sqqueue >= 0);
732 THREAD_LOCK_ASSERT(td, MA_OWNED);
733 sc = SC_LOOKUP(sq->sq_wchan);
734 mtx_assert(&sc->sc_lock, MA_OWNED);
736 SDT_PROBE2(sched, , , wakeup, td, td->td_proc);
738 /* Remove the thread from the queue. */
739 sq->sq_blockedcnt[td->td_sqqueue]--;
740 TAILQ_REMOVE(&sq->sq_blocked[td->td_sqqueue], td, td_slpq);
743 * Get a sleep queue for this thread. If this is the last waiter,
744 * use the queue itself and take it out of the chain, otherwise,
745 * remove a queue from the free list.
747 if (LIST_EMPTY(&sq->sq_free)) {
748 td->td_sleepqueue = sq;
752 #ifdef SLEEPQUEUE_PROFILING
756 td->td_sleepqueue = LIST_FIRST(&sq->sq_free);
757 LIST_REMOVE(td->td_sleepqueue, sq_hash);
761 td->td_flags &= ~TDF_SINTR;
763 CTR3(KTR_PROC, "sleepq_wakeup: thread %p (pid %ld, %s)",
764 (void *)td, (long)td->td_proc->p_pid, td->td_name);
766 /* Adjust priority if requested. */
767 MPASS(pri == 0 || (pri >= PRI_MIN && pri <= PRI_MAX));
768 if (pri != 0 && td->td_priority > pri &&
769 PRI_BASE(td->td_pri_class) == PRI_TIMESHARE)
773 * Note that thread td might not be sleeping if it is running
774 * sleepq_catch_signals() on another CPU or is blocked on its
775 * proc lock to check signals. There's no need to mark the
776 * thread runnable in that case.
778 if (TD_IS_SLEEPING(td)) {
780 return (setrunnable(td));
787 * UMA zone item deallocator.
790 sleepq_dtor(void *mem, int size, void *arg)
792 struct sleepqueue *sq;
796 for (i = 0; i < NR_SLEEPQS; i++) {
797 MPASS(TAILQ_EMPTY(&sq->sq_blocked[i]));
798 MPASS(sq->sq_blockedcnt[i] == 0);
804 * UMA zone item initializer.
807 sleepq_init(void *mem, int size, int flags)
809 struct sleepqueue *sq;
814 for (i = 0; i < NR_SLEEPQS; i++) {
815 TAILQ_INIT(&sq->sq_blocked[i]);
816 sq->sq_blockedcnt[i] = 0;
818 LIST_INIT(&sq->sq_free);
823 * Find the highest priority thread sleeping on a wait channel and resume it.
826 sleepq_signal(void *wchan, int flags, int pri, int queue)
828 struct sleepqueue *sq;
829 struct thread *td, *besttd;
832 CTR2(KTR_PROC, "sleepq_signal(%p, %d)", wchan, flags);
833 KASSERT(wchan != NULL, ("%s: invalid NULL wait channel", __func__));
834 MPASS((queue >= 0) && (queue < NR_SLEEPQS));
835 sq = sleepq_lookup(wchan);
838 KASSERT(sq->sq_type == (flags & SLEEPQ_TYPE),
839 ("%s: mismatch between sleep/wakeup and cv_*", __func__));
842 * Find the highest priority thread on the queue. If there is a
843 * tie, use the thread that first appears in the queue as it has
844 * been sleeping the longest since threads are always added to
845 * the tail of sleep queues.
848 TAILQ_FOREACH(td, &sq->sq_blocked[queue], td_slpq) {
849 if (besttd == NULL || td->td_priority < besttd->td_priority)
852 MPASS(besttd != NULL);
854 wakeup_swapper = sleepq_resume_thread(sq, besttd, pri);
855 thread_unlock(besttd);
856 return (wakeup_swapper);
860 * Resume all threads sleeping on a specified wait channel.
863 sleepq_broadcast(void *wchan, int flags, int pri, int queue)
865 struct sleepqueue *sq;
866 struct thread *td, *tdn;
869 CTR2(KTR_PROC, "sleepq_broadcast(%p, %d)", wchan, flags);
870 KASSERT(wchan != NULL, ("%s: invalid NULL wait channel", __func__));
871 MPASS((queue >= 0) && (queue < NR_SLEEPQS));
872 sq = sleepq_lookup(wchan);
875 KASSERT(sq->sq_type == (flags & SLEEPQ_TYPE),
876 ("%s: mismatch between sleep/wakeup and cv_*", __func__));
878 /* Resume all blocked threads on the sleep queue. */
880 TAILQ_FOREACH_SAFE(td, &sq->sq_blocked[queue], td_slpq, tdn) {
882 if (sleepq_resume_thread(sq, td, pri))
886 return (wakeup_swapper);
890 * Time sleeping threads out. When the timeout expires, the thread is
891 * removed from the sleep queue and made runnable if it is still asleep.
894 sleepq_timeout(void *arg)
896 struct sleepqueue_chain *sc;
897 struct sleepqueue *sq;
904 CTR3(KTR_PROC, "sleepq_timeout: thread %p (pid %ld, %s)",
905 (void *)td, (long)td->td_proc->p_pid, (void *)td->td_name);
908 * First, see if the thread is asleep and get the wait channel if
912 if (TD_IS_SLEEPING(td) && TD_ON_SLEEPQ(td)) {
913 wchan = td->td_wchan;
914 sc = SC_LOOKUP(wchan);
915 THREAD_LOCKPTR_ASSERT(td, &sc->sc_lock);
916 sq = sleepq_lookup(wchan);
918 td->td_flags |= TDF_TIMEOUT;
919 wakeup_swapper = sleepq_resume_thread(sq, td, 0);
927 * If the thread is on the SLEEPQ but isn't sleeping yet, it
928 * can either be on another CPU in between sleepq_add() and
929 * one of the sleepq_*wait*() routines or it can be in
930 * sleepq_catch_signals().
932 if (TD_ON_SLEEPQ(td)) {
933 td->td_flags |= TDF_TIMEOUT;
939 * Now check for the edge cases. First, if TDF_TIMEOUT is set,
940 * then the other thread has already yielded to us, so clear
941 * the flag and resume it. If TDF_TIMEOUT is not set, then the
942 * we know that the other thread is not on a sleep queue, but it
943 * hasn't resumed execution yet. In that case, set TDF_TIMOFAIL
944 * to let it know that the timeout has already run and doesn't
945 * need to be canceled.
947 if (td->td_flags & TDF_TIMEOUT) {
948 MPASS(TD_IS_SLEEPING(td));
949 td->td_flags &= ~TDF_TIMEOUT;
951 wakeup_swapper = setrunnable(td);
953 td->td_flags |= TDF_TIMOFAIL;
960 * Resumes a specific thread from the sleep queue associated with a specific
961 * wait channel if it is on that queue.
964 sleepq_remove(struct thread *td, void *wchan)
966 struct sleepqueue *sq;
970 * Look up the sleep queue for this wait channel, then re-check
971 * that the thread is asleep on that channel, if it is not, then
974 MPASS(wchan != NULL);
976 sq = sleepq_lookup(wchan);
978 * We can not lock the thread here as it may be sleeping on a
979 * different sleepq. However, holding the sleepq lock for this
980 * wchan can guarantee that we do not miss a wakeup for this
981 * channel. The asserts below will catch any false positives.
983 if (!TD_ON_SLEEPQ(td) || td->td_wchan != wchan) {
984 sleepq_release(wchan);
987 /* Thread is asleep on sleep queue sq, so wake it up. */
990 MPASS(td->td_wchan == wchan);
991 wakeup_swapper = sleepq_resume_thread(sq, td, 0);
993 sleepq_release(wchan);
999 * Abort a thread as if an interrupt had occurred. Only abort
1000 * interruptible waits (unfortunately it isn't safe to abort others).
1003 sleepq_abort(struct thread *td, int intrval)
1005 struct sleepqueue *sq;
1008 THREAD_LOCK_ASSERT(td, MA_OWNED);
1009 MPASS(TD_ON_SLEEPQ(td));
1010 MPASS(td->td_flags & TDF_SINTR);
1011 MPASS(intrval == EINTR || intrval == ERESTART);
1014 * If the TDF_TIMEOUT flag is set, just leave. A
1015 * timeout is scheduled anyhow.
1017 if (td->td_flags & TDF_TIMEOUT)
1020 CTR3(KTR_PROC, "sleepq_abort: thread %p (pid %ld, %s)",
1021 (void *)td, (long)td->td_proc->p_pid, (void *)td->td_name);
1022 td->td_intrval = intrval;
1023 td->td_flags |= TDF_SLEEPABORT;
1025 * If the thread has not slept yet it will find the signal in
1026 * sleepq_catch_signals() and call sleepq_resume_thread. Otherwise
1027 * we have to do it here.
1029 if (!TD_IS_SLEEPING(td))
1031 wchan = td->td_wchan;
1032 MPASS(wchan != NULL);
1033 sq = sleepq_lookup(wchan);
1036 /* Thread is asleep on sleep queue sq, so wake it up. */
1037 return (sleepq_resume_thread(sq, td, 0));
1041 * Prints the stacks of all threads presently sleeping on wchan/queue to
1042 * the sbuf sb. Sets count_stacks_printed to the number of stacks actually
1043 * printed. Typically, this will equal the number of threads sleeping on the
1044 * queue, but may be less if sb overflowed before all stacks were printed.
1048 sleepq_sbuf_print_stacks(struct sbuf *sb, void *wchan, int queue,
1049 int *count_stacks_printed)
1051 struct thread *td, *td_next;
1052 struct sleepqueue *sq;
1054 struct sbuf **td_infos;
1055 int i, stack_idx, error, stacks_to_allocate;
1056 bool finished, partial_print;
1060 partial_print = false;
1062 KASSERT(wchan != NULL, ("%s: invalid NULL wait channel", __func__));
1063 MPASS((queue >= 0) && (queue < NR_SLEEPQS));
1065 stacks_to_allocate = 10;
1066 for (i = 0; i < 3 && !finished ; i++) {
1067 /* We cannot malloc while holding the queue's spinlock, so
1068 * we do our mallocs now, and hope it is enough. If it
1069 * isn't, we will free these, drop the lock, malloc more,
1070 * and try again, up to a point. After that point we will
1071 * give up and report ENOMEM. We also cannot write to sb
1072 * during this time since the client may have set the
1073 * SBUF_AUTOEXTEND flag on their sbuf, which could cause a
1074 * malloc as we print to it. So we defer actually printing
1075 * to sb until after we drop the spinlock.
1078 /* Where we will store the stacks. */
1079 st = malloc(sizeof(struct stack *) * stacks_to_allocate,
1081 for (stack_idx = 0; stack_idx < stacks_to_allocate;
1083 st[stack_idx] = stack_create();
1085 /* Where we will store the td name, tid, etc. */
1086 td_infos = malloc(sizeof(struct sbuf *) * stacks_to_allocate,
1088 for (stack_idx = 0; stack_idx < stacks_to_allocate;
1090 td_infos[stack_idx] = sbuf_new(NULL, NULL,
1091 MAXCOMLEN + sizeof(struct thread *) * 2 + 40,
1095 sq = sleepq_lookup(wchan);
1097 /* This sleepq does not exist; exit and return ENOENT. */
1100 sleepq_release(wchan);
1105 /* Save thread info */
1106 TAILQ_FOREACH_SAFE(td, &sq->sq_blocked[queue], td_slpq,
1108 if (stack_idx >= stacks_to_allocate)
1111 /* Note the td_lock is equal to the sleepq_lock here. */
1112 stack_save_td(st[stack_idx], td);
1114 sbuf_printf(td_infos[stack_idx], "%d: %s %p",
1115 td->td_tid, td->td_name, td);
1121 sleepq_release(wchan);
1123 /* Print the stacks */
1124 for (i = 0; i < stack_idx; i++) {
1125 sbuf_finish(td_infos[i]);
1126 sbuf_printf(sb, "--- thread %s: ---\n", sbuf_data(td_infos[i]));
1127 stack_sbuf_print(sb, st[i]);
1128 sbuf_printf(sb, "\n");
1130 error = sbuf_error(sb);
1132 *count_stacks_printed = stack_idx;
1137 sleepq_release(wchan);
1138 for (stack_idx = 0; stack_idx < stacks_to_allocate;
1140 stack_destroy(st[stack_idx]);
1141 for (stack_idx = 0; stack_idx < stacks_to_allocate;
1143 sbuf_delete(td_infos[stack_idx]);
1145 free(td_infos, M_TEMP);
1146 stacks_to_allocate *= 10;
1149 if (!finished && error == 0)
1156 #ifdef SLEEPQUEUE_PROFILING
1157 #define SLEEPQ_PROF_LOCATIONS 1024
1158 #define SLEEPQ_SBUFSIZE 512
1159 struct sleepq_prof {
1160 LIST_ENTRY(sleepq_prof) sp_link;
1161 const char *sp_wmesg;
1165 LIST_HEAD(sqphead, sleepq_prof);
1167 struct sqphead sleepq_prof_free;
1168 struct sqphead sleepq_hash[SC_TABLESIZE];
1169 static struct sleepq_prof sleepq_profent[SLEEPQ_PROF_LOCATIONS];
1170 static struct mtx sleepq_prof_lock;
1171 MTX_SYSINIT(sleepq_prof_lock, &sleepq_prof_lock, "sleepq_prof", MTX_SPIN);
1174 sleepq_profile(const char *wmesg)
1176 struct sleepq_prof *sp;
1178 mtx_lock_spin(&sleepq_prof_lock);
1179 if (prof_enabled == 0)
1181 LIST_FOREACH(sp, &sleepq_hash[SC_HASH(wmesg)], sp_link)
1182 if (sp->sp_wmesg == wmesg)
1184 sp = LIST_FIRST(&sleepq_prof_free);
1187 sp->sp_wmesg = wmesg;
1188 LIST_REMOVE(sp, sp_link);
1189 LIST_INSERT_HEAD(&sleepq_hash[SC_HASH(wmesg)], sp, sp_link);
1193 mtx_unlock_spin(&sleepq_prof_lock);
1198 sleepq_prof_reset(void)
1200 struct sleepq_prof *sp;
1204 mtx_lock_spin(&sleepq_prof_lock);
1205 enabled = prof_enabled;
1207 for (i = 0; i < SC_TABLESIZE; i++)
1208 LIST_INIT(&sleepq_hash[i]);
1209 LIST_INIT(&sleepq_prof_free);
1210 for (i = 0; i < SLEEPQ_PROF_LOCATIONS; i++) {
1211 sp = &sleepq_profent[i];
1212 sp->sp_wmesg = NULL;
1214 LIST_INSERT_HEAD(&sleepq_prof_free, sp, sp_link);
1216 prof_enabled = enabled;
1217 mtx_unlock_spin(&sleepq_prof_lock);
1221 enable_sleepq_prof(SYSCTL_HANDLER_ARGS)
1226 error = sysctl_handle_int(oidp, &v, v, req);
1229 if (req->newptr == NULL)
1231 if (v == prof_enabled)
1234 sleepq_prof_reset();
1235 mtx_lock_spin(&sleepq_prof_lock);
1237 mtx_unlock_spin(&sleepq_prof_lock);
1243 reset_sleepq_prof_stats(SYSCTL_HANDLER_ARGS)
1248 error = sysctl_handle_int(oidp, &v, 0, req);
1251 if (req->newptr == NULL)
1255 sleepq_prof_reset();
1261 dump_sleepq_prof_stats(SYSCTL_HANDLER_ARGS)
1263 struct sleepq_prof *sp;
1269 error = sysctl_wire_old_buffer(req, 0);
1272 sb = sbuf_new_for_sysctl(NULL, NULL, SLEEPQ_SBUFSIZE, req);
1273 sbuf_printf(sb, "\nwmesg\tcount\n");
1274 enabled = prof_enabled;
1275 mtx_lock_spin(&sleepq_prof_lock);
1277 mtx_unlock_spin(&sleepq_prof_lock);
1278 for (i = 0; i < SC_TABLESIZE; i++) {
1279 LIST_FOREACH(sp, &sleepq_hash[i], sp_link) {
1280 sbuf_printf(sb, "%s\t%ld\n",
1281 sp->sp_wmesg, sp->sp_count);
1284 mtx_lock_spin(&sleepq_prof_lock);
1285 prof_enabled = enabled;
1286 mtx_unlock_spin(&sleepq_prof_lock);
1288 error = sbuf_finish(sb);
1293 SYSCTL_PROC(_debug_sleepq, OID_AUTO, stats, CTLTYPE_STRING | CTLFLAG_RD,
1294 NULL, 0, dump_sleepq_prof_stats, "A", "Sleepqueue profiling statistics");
1295 SYSCTL_PROC(_debug_sleepq, OID_AUTO, reset, CTLTYPE_INT | CTLFLAG_RW,
1296 NULL, 0, reset_sleepq_prof_stats, "I",
1297 "Reset sleepqueue profiling statistics");
1298 SYSCTL_PROC(_debug_sleepq, OID_AUTO, enable, CTLTYPE_INT | CTLFLAG_RW,
1299 NULL, 0, enable_sleepq_prof, "I", "Enable sleepqueue profiling");
1303 DB_SHOW_COMMAND(sleepq, db_show_sleepqueue)
1305 struct sleepqueue_chain *sc;
1306 struct sleepqueue *sq;
1308 struct lock_object *lock;
1318 * First, see if there is an active sleep queue for the wait channel
1319 * indicated by the address.
1321 wchan = (void *)addr;
1322 sc = SC_LOOKUP(wchan);
1323 LIST_FOREACH(sq, &sc->sc_queues, sq_hash)
1324 if (sq->sq_wchan == wchan)
1328 * Second, see if there is an active sleep queue at the address
1331 for (i = 0; i < SC_TABLESIZE; i++)
1332 LIST_FOREACH(sq, &sleepq_chains[i].sc_queues, sq_hash) {
1333 if (sq == (struct sleepqueue *)addr)
1337 db_printf("Unable to locate a sleep queue via %p\n", (void *)addr);
1340 db_printf("Wait channel: %p\n", sq->sq_wchan);
1341 db_printf("Queue type: %d\n", sq->sq_type);
1345 db_printf("Associated Interlock: %p - (%s) %s\n", lock,
1346 LOCK_CLASS(lock)->lc_name, lock->lo_name);
1349 db_printf("Blocked threads:\n");
1350 for (i = 0; i < NR_SLEEPQS; i++) {
1351 db_printf("\nQueue[%d]:\n", i);
1352 if (TAILQ_EMPTY(&sq->sq_blocked[i]))
1353 db_printf("\tempty\n");
1355 TAILQ_FOREACH(td, &sq->sq_blocked[0],
1357 db_printf("\t%p (tid %d, pid %d, \"%s\")\n", td,
1358 td->td_tid, td->td_proc->p_pid,
1361 db_printf("(expected: %u)\n", sq->sq_blockedcnt[i]);
1365 /* Alias 'show sleepqueue' to 'show sleepq'. */
1366 DB_SHOW_ALIAS(sleepqueue, db_show_sleepqueue);