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. These constants are
89 * the same ones that 4BSD (and possibly earlier versions of BSD) used.
90 * Basically, we ignore the lower 8 bits of the address since most wait
91 * channel pointers are aligned and only look at the next 7 bits for the
92 * hash. SC_TABLESIZE must be a power of two for SC_MASK to work properly.
94 #define SC_TABLESIZE 128 /* Must be power of 2. */
95 #define SC_MASK (SC_TABLESIZE - 1)
97 #define SC_HASH(wc) (((uintptr_t)(wc) >> SC_SHIFT) & SC_MASK)
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 * Early initialization of sleep queues that is called from the sleepinit()
176 init_sleepqueues(void)
178 #ifdef SLEEPQUEUE_PROFILING
179 struct sysctl_oid *chain_oid;
184 for (i = 0; i < SC_TABLESIZE; i++) {
185 LIST_INIT(&sleepq_chains[i].sc_queues);
186 mtx_init(&sleepq_chains[i].sc_lock, "sleepq chain", NULL,
187 MTX_SPIN | MTX_RECURSE);
188 #ifdef SLEEPQUEUE_PROFILING
189 snprintf(chain_name, sizeof(chain_name), "%d", i);
190 chain_oid = SYSCTL_ADD_NODE(NULL,
191 SYSCTL_STATIC_CHILDREN(_debug_sleepq_chains), OID_AUTO,
192 chain_name, CTLFLAG_RD, NULL, "sleepq chain stats");
193 SYSCTL_ADD_UINT(NULL, SYSCTL_CHILDREN(chain_oid), OID_AUTO,
194 "depth", CTLFLAG_RD, &sleepq_chains[i].sc_depth, 0, NULL);
195 SYSCTL_ADD_UINT(NULL, SYSCTL_CHILDREN(chain_oid), OID_AUTO,
196 "max_depth", CTLFLAG_RD, &sleepq_chains[i].sc_max_depth, 0,
200 sleepq_zone = uma_zcreate("SLEEPQUEUE", sizeof(struct sleepqueue),
202 NULL, sleepq_dtor, sleepq_init, NULL, UMA_ALIGN_CACHE, 0);
204 NULL, NULL, sleepq_init, NULL, UMA_ALIGN_CACHE, 0);
207 thread0.td_sleepqueue = sleepq_alloc();
211 * Get a sleep queue for a new thread.
217 return (uma_zalloc(sleepq_zone, M_WAITOK));
221 * Free a sleep queue when a thread is destroyed.
224 sleepq_free(struct sleepqueue *sq)
227 uma_zfree(sleepq_zone, sq);
231 * Lock the sleep queue chain associated with the specified wait channel.
234 sleepq_lock(void *wchan)
236 struct sleepqueue_chain *sc;
238 sc = SC_LOOKUP(wchan);
239 mtx_lock_spin(&sc->sc_lock);
243 * Look up the sleep queue associated with a given wait channel in the hash
244 * table locking the associated sleep queue chain. If no queue is found in
245 * the table, NULL is returned.
248 sleepq_lookup(void *wchan)
250 struct sleepqueue_chain *sc;
251 struct sleepqueue *sq;
253 KASSERT(wchan != NULL, ("%s: invalid NULL wait channel", __func__));
254 sc = SC_LOOKUP(wchan);
255 mtx_assert(&sc->sc_lock, MA_OWNED);
256 LIST_FOREACH(sq, &sc->sc_queues, sq_hash)
257 if (sq->sq_wchan == wchan)
263 * Unlock the sleep queue chain associated with a given wait channel.
266 sleepq_release(void *wchan)
268 struct sleepqueue_chain *sc;
270 sc = SC_LOOKUP(wchan);
271 mtx_unlock_spin(&sc->sc_lock);
275 * Places the current thread on the sleep queue for the specified wait
276 * channel. If INVARIANTS is enabled, then it associates the passed in
277 * lock with the sleepq to make sure it is held when that sleep queue is
281 sleepq_add(void *wchan, struct lock_object *lock, const char *wmesg, int flags,
284 struct sleepqueue_chain *sc;
285 struct sleepqueue *sq;
289 sc = SC_LOOKUP(wchan);
290 mtx_assert(&sc->sc_lock, MA_OWNED);
291 MPASS(td->td_sleepqueue != NULL);
292 MPASS(wchan != NULL);
293 MPASS((queue >= 0) && (queue < NR_SLEEPQS));
295 /* If this thread is not allowed to sleep, die a horrible death. */
296 KASSERT(!(td->td_pflags & TDP_NOSLEEPING),
297 ("Trying sleep, but thread marked as sleeping prohibited"));
299 /* Look up the sleep queue associated with the wait channel 'wchan'. */
300 sq = sleepq_lookup(wchan);
303 * If the wait channel does not already have a sleep queue, use
304 * this thread's sleep queue. Otherwise, insert the current thread
305 * into the sleep queue already in use by this wait channel.
311 sq = td->td_sleepqueue;
312 for (i = 0; i < NR_SLEEPQS; i++) {
313 KASSERT(TAILQ_EMPTY(&sq->sq_blocked[i]),
314 ("thread's sleep queue %d is not empty", i));
315 KASSERT(sq->sq_blockedcnt[i] == 0,
316 ("thread's sleep queue %d count mismatches", i));
318 KASSERT(LIST_EMPTY(&sq->sq_free),
319 ("thread's sleep queue has a non-empty free list"));
320 KASSERT(sq->sq_wchan == NULL, ("stale sq_wchan pointer"));
323 #ifdef SLEEPQUEUE_PROFILING
325 if (sc->sc_depth > sc->sc_max_depth) {
326 sc->sc_max_depth = sc->sc_depth;
327 if (sc->sc_max_depth > sleepq_max_depth)
328 sleepq_max_depth = sc->sc_max_depth;
331 sq = td->td_sleepqueue;
332 LIST_INSERT_HEAD(&sc->sc_queues, sq, sq_hash);
333 sq->sq_wchan = wchan;
334 sq->sq_type = flags & SLEEPQ_TYPE;
336 MPASS(wchan == sq->sq_wchan);
337 MPASS(lock == sq->sq_lock);
338 MPASS((flags & SLEEPQ_TYPE) == sq->sq_type);
339 LIST_INSERT_HEAD(&sq->sq_free, td->td_sleepqueue, sq_hash);
342 TAILQ_INSERT_TAIL(&sq->sq_blocked[queue], td, td_slpq);
343 sq->sq_blockedcnt[queue]++;
344 td->td_sleepqueue = NULL;
345 td->td_sqqueue = queue;
346 td->td_wchan = wchan;
347 td->td_wmesg = wmesg;
348 if (flags & SLEEPQ_INTERRUPTIBLE) {
349 td->td_flags |= TDF_SINTR;
350 td->td_flags &= ~TDF_SLEEPABORT;
356 * Sets a timeout that will remove the current thread from the specified
357 * sleep queue after timo ticks if the thread has not already been awakened.
360 sleepq_set_timeout(void *wchan, int timo)
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_curcpu(&td->td_slpcallout, timo, sleepq_timeout, td);
375 * Return the number of actual sleepers for the specified queue.
378 sleepq_sleepcnt(void *wchan, int queue)
380 struct sleepqueue *sq;
382 KASSERT(wchan != NULL, ("%s: invalid NULL wait channel", __func__));
383 MPASS((queue >= 0) && (queue < NR_SLEEPQS));
384 sq = sleepq_lookup(wchan);
387 return (sq->sq_blockedcnt[queue]);
391 * Marks the pending sleep of the current thread as interruptible and
392 * makes an initial check for pending signals before putting a thread
393 * to sleep. Enters and exits with the thread lock held. Thread lock
394 * may have transitioned from the sleepq lock to a run lock.
397 sleepq_catch_signals(void *wchan, int pri)
399 struct sleepqueue_chain *sc;
400 struct sleepqueue *sq;
404 int sig, ret, stop_allowed;
408 sc = SC_LOOKUP(wchan);
409 mtx_assert(&sc->sc_lock, MA_OWNED);
410 MPASS(wchan != NULL);
411 if ((td->td_pflags & TDP_WAKEUP) != 0) {
412 td->td_pflags &= ~TDP_WAKEUP;
419 * See if there are any pending signals for this thread. If not
420 * we can switch immediately. Otherwise do the signal processing
424 if ((td->td_flags & (TDF_NEEDSIGCHK | TDF_NEEDSUSPCHK)) == 0) {
425 sleepq_switch(wchan, pri);
428 stop_allowed = (td->td_flags & TDF_SBDRY) ? SIG_STOP_NOT_ALLOWED :
431 mtx_unlock_spin(&sc->sc_lock);
432 CTR3(KTR_PROC, "sleepq catching signals: thread %p (pid %ld, %s)",
433 (void *)td, (long)p->p_pid, td->td_name);
436 mtx_lock(&ps->ps_mtx);
437 sig = cursig(td, stop_allowed);
439 mtx_unlock(&ps->ps_mtx);
440 ret = thread_suspend_check(1);
441 MPASS(ret == 0 || ret == EINTR || ret == ERESTART);
443 if (SIGISMEMBER(ps->ps_sigintr, sig))
447 mtx_unlock(&ps->ps_mtx);
450 * Lock the per-process spinlock prior to dropping the PROC_LOCK
451 * to avoid a signal delivery race. PROC_LOCK, PROC_SLOCK, and
452 * thread_lock() are currently held in tdsendsignal().
455 mtx_lock_spin(&sc->sc_lock);
460 sleepq_switch(wchan, pri);
465 * There were pending signals and this thread is still
466 * on the sleep queue, remove it from the sleep queue.
468 if (TD_ON_SLEEPQ(td)) {
469 sq = sleepq_lookup(wchan);
470 if (sleepq_resume_thread(sq, td, 0)) {
473 * This thread hasn't gone to sleep yet, so it
474 * should not be swapped out.
476 panic("not waking up swapper");
480 mtx_unlock_spin(&sc->sc_lock);
481 MPASS(td->td_lock != &sc->sc_lock);
486 * Switches to another thread if we are still asleep on a sleep queue.
487 * Returns with thread lock.
490 sleepq_switch(void *wchan, int pri)
492 struct sleepqueue_chain *sc;
493 struct sleepqueue *sq;
497 sc = SC_LOOKUP(wchan);
498 mtx_assert(&sc->sc_lock, MA_OWNED);
499 THREAD_LOCK_ASSERT(td, MA_OWNED);
502 * If we have a sleep queue, then we've already been woken up, so
505 if (td->td_sleepqueue != NULL) {
506 mtx_unlock_spin(&sc->sc_lock);
511 * If TDF_TIMEOUT is set, then our sleep has been timed out
512 * already but we are still on the sleep queue, so dequeue the
515 if (td->td_flags & TDF_TIMEOUT) {
516 MPASS(TD_ON_SLEEPQ(td));
517 sq = sleepq_lookup(wchan);
518 if (sleepq_resume_thread(sq, td, 0)) {
521 * This thread hasn't gone to sleep yet, so it
522 * should not be swapped out.
524 panic("not waking up swapper");
527 mtx_unlock_spin(&sc->sc_lock);
530 #ifdef SLEEPQUEUE_PROFILING
532 sleepq_profile(td->td_wmesg);
534 MPASS(td->td_sleepqueue == NULL);
535 sched_sleep(td, pri);
536 thread_lock_set(td, &sc->sc_lock);
537 SDT_PROBE0(sched, , , sleep);
539 mi_switch(SW_VOL | SWT_SLEEPQ, NULL);
540 KASSERT(TD_IS_RUNNING(td), ("running but not TDS_RUNNING"));
541 CTR3(KTR_PROC, "sleepq resume: thread %p (pid %ld, %s)",
542 (void *)td, (long)td->td_proc->p_pid, (void *)td->td_name);
546 * Check to see if we timed out.
549 sleepq_check_timeout(void)
554 THREAD_LOCK_ASSERT(td, MA_OWNED);
557 * If TDF_TIMEOUT is set, we timed out.
559 if (td->td_flags & TDF_TIMEOUT) {
560 td->td_flags &= ~TDF_TIMEOUT;
561 return (EWOULDBLOCK);
565 * If TDF_TIMOFAIL is set, the timeout ran after we had
566 * already been woken up.
568 if (td->td_flags & TDF_TIMOFAIL)
569 td->td_flags &= ~TDF_TIMOFAIL;
572 * If callout_stop() fails, then the timeout is running on
573 * another CPU, so synchronize with it to avoid having it
574 * accidentally wake up a subsequent sleep.
576 else if (callout_stop(&td->td_slpcallout) == 0) {
577 td->td_flags |= TDF_TIMEOUT;
579 mi_switch(SW_INVOL | SWT_SLEEPQTIMO, NULL);
585 * Check to see if we were awoken by a signal.
588 sleepq_check_signals(void)
593 THREAD_LOCK_ASSERT(td, MA_OWNED);
595 /* We are no longer in an interruptible sleep. */
596 if (td->td_flags & TDF_SINTR)
597 td->td_flags &= ~TDF_SINTR;
599 if (td->td_flags & TDF_SLEEPABORT) {
600 td->td_flags &= ~TDF_SLEEPABORT;
601 return (td->td_intrval);
608 * Block the current thread until it is awakened from its sleep queue.
611 sleepq_wait(void *wchan, int pri)
616 MPASS(!(td->td_flags & TDF_SINTR));
618 sleepq_switch(wchan, pri);
623 * Block the current thread until it is awakened from its sleep queue
624 * or it is interrupted by a signal.
627 sleepq_wait_sig(void *wchan, int pri)
632 rcatch = sleepq_catch_signals(wchan, pri);
633 rval = sleepq_check_signals();
634 thread_unlock(curthread);
641 * Block the current thread until it is awakened from its sleep queue
642 * or it times out while waiting.
645 sleepq_timedwait(void *wchan, int pri)
651 MPASS(!(td->td_flags & TDF_SINTR));
653 sleepq_switch(wchan, pri);
654 rval = sleepq_check_timeout();
661 * Block the current thread until it is awakened from its sleep queue,
662 * it is interrupted by a signal, or it times out waiting to be awakened.
665 sleepq_timedwait_sig(void *wchan, int pri)
667 int rcatch, rvalt, rvals;
669 rcatch = sleepq_catch_signals(wchan, pri);
670 rvalt = sleepq_check_timeout();
671 rvals = sleepq_check_signals();
672 thread_unlock(curthread);
681 * Returns the type of sleepqueue given a waitchannel.
684 sleepq_type(void *wchan)
686 struct sleepqueue *sq;
689 MPASS(wchan != NULL);
692 sq = sleepq_lookup(wchan);
694 sleepq_release(wchan);
698 sleepq_release(wchan);
703 * Removes a thread from a sleep queue and makes it
707 sleepq_resume_thread(struct sleepqueue *sq, struct thread *td, int pri)
709 struct sleepqueue_chain *sc;
712 MPASS(sq->sq_wchan != NULL);
713 MPASS(td->td_wchan == sq->sq_wchan);
714 MPASS(td->td_sqqueue < NR_SLEEPQS && td->td_sqqueue >= 0);
715 THREAD_LOCK_ASSERT(td, MA_OWNED);
716 sc = SC_LOOKUP(sq->sq_wchan);
717 mtx_assert(&sc->sc_lock, MA_OWNED);
719 SDT_PROBE2(sched, , , wakeup, td, td->td_proc);
721 /* Remove the thread from the queue. */
722 sq->sq_blockedcnt[td->td_sqqueue]--;
723 TAILQ_REMOVE(&sq->sq_blocked[td->td_sqqueue], td, td_slpq);
726 * Get a sleep queue for this thread. If this is the last waiter,
727 * use the queue itself and take it out of the chain, otherwise,
728 * remove a queue from the free list.
730 if (LIST_EMPTY(&sq->sq_free)) {
731 td->td_sleepqueue = sq;
735 #ifdef SLEEPQUEUE_PROFILING
739 td->td_sleepqueue = LIST_FIRST(&sq->sq_free);
740 LIST_REMOVE(td->td_sleepqueue, sq_hash);
744 td->td_flags &= ~TDF_SINTR;
746 CTR3(KTR_PROC, "sleepq_wakeup: thread %p (pid %ld, %s)",
747 (void *)td, (long)td->td_proc->p_pid, td->td_name);
749 /* Adjust priority if requested. */
750 MPASS(pri == 0 || (pri >= PRI_MIN && pri <= PRI_MAX));
751 if (pri != 0 && td->td_priority > pri &&
752 PRI_BASE(td->td_pri_class) == PRI_TIMESHARE)
756 * Note that thread td might not be sleeping if it is running
757 * sleepq_catch_signals() on another CPU or is blocked on its
758 * proc lock to check signals. There's no need to mark the
759 * thread runnable in that case.
761 if (TD_IS_SLEEPING(td)) {
763 return (setrunnable(td));
770 * UMA zone item deallocator.
773 sleepq_dtor(void *mem, int size, void *arg)
775 struct sleepqueue *sq;
779 for (i = 0; i < NR_SLEEPQS; i++) {
780 MPASS(TAILQ_EMPTY(&sq->sq_blocked[i]));
781 MPASS(sq->sq_blockedcnt[i] == 0);
787 * UMA zone item initializer.
790 sleepq_init(void *mem, int size, int flags)
792 struct sleepqueue *sq;
797 for (i = 0; i < NR_SLEEPQS; i++) {
798 TAILQ_INIT(&sq->sq_blocked[i]);
799 sq->sq_blockedcnt[i] = 0;
801 LIST_INIT(&sq->sq_free);
806 * Find the highest priority thread sleeping on a wait channel and resume it.
809 sleepq_signal(void *wchan, int flags, int pri, int queue)
811 struct sleepqueue *sq;
812 struct thread *td, *besttd;
815 CTR2(KTR_PROC, "sleepq_signal(%p, %d)", wchan, flags);
816 KASSERT(wchan != NULL, ("%s: invalid NULL wait channel", __func__));
817 MPASS((queue >= 0) && (queue < NR_SLEEPQS));
818 sq = sleepq_lookup(wchan);
821 KASSERT(sq->sq_type == (flags & SLEEPQ_TYPE),
822 ("%s: mismatch between sleep/wakeup and cv_*", __func__));
825 * Find the highest priority thread on the queue. If there is a
826 * tie, use the thread that first appears in the queue as it has
827 * been sleeping the longest since threads are always added to
828 * the tail of sleep queues.
831 TAILQ_FOREACH(td, &sq->sq_blocked[queue], td_slpq) {
832 if (besttd == NULL || td->td_priority < besttd->td_priority)
835 MPASS(besttd != NULL);
837 wakeup_swapper = sleepq_resume_thread(sq, besttd, pri);
838 thread_unlock(besttd);
839 return (wakeup_swapper);
843 * Resume all threads sleeping on a specified wait channel.
846 sleepq_broadcast(void *wchan, int flags, int pri, int queue)
848 struct sleepqueue *sq;
849 struct thread *td, *tdn;
852 CTR2(KTR_PROC, "sleepq_broadcast(%p, %d)", wchan, flags);
853 KASSERT(wchan != NULL, ("%s: invalid NULL wait channel", __func__));
854 MPASS((queue >= 0) && (queue < NR_SLEEPQS));
855 sq = sleepq_lookup(wchan);
858 KASSERT(sq->sq_type == (flags & SLEEPQ_TYPE),
859 ("%s: mismatch between sleep/wakeup and cv_*", __func__));
861 /* Resume all blocked threads on the sleep queue. */
863 TAILQ_FOREACH_SAFE(td, &sq->sq_blocked[queue], td_slpq, tdn) {
865 if (sleepq_resume_thread(sq, td, pri))
869 return (wakeup_swapper);
873 * Time sleeping threads out. When the timeout expires, the thread is
874 * removed from the sleep queue and made runnable if it is still asleep.
877 sleepq_timeout(void *arg)
879 struct sleepqueue_chain *sc;
880 struct sleepqueue *sq;
887 CTR3(KTR_PROC, "sleepq_timeout: thread %p (pid %ld, %s)",
888 (void *)td, (long)td->td_proc->p_pid, (void *)td->td_name);
891 * First, see if the thread is asleep and get the wait channel if
895 if (TD_IS_SLEEPING(td) && TD_ON_SLEEPQ(td)) {
896 wchan = td->td_wchan;
897 sc = SC_LOOKUP(wchan);
898 THREAD_LOCKPTR_ASSERT(td, &sc->sc_lock);
899 sq = sleepq_lookup(wchan);
901 td->td_flags |= TDF_TIMEOUT;
902 wakeup_swapper = sleepq_resume_thread(sq, td, 0);
910 * If the thread is on the SLEEPQ but isn't sleeping yet, it
911 * can either be on another CPU in between sleepq_add() and
912 * one of the sleepq_*wait*() routines or it can be in
913 * sleepq_catch_signals().
915 if (TD_ON_SLEEPQ(td)) {
916 td->td_flags |= TDF_TIMEOUT;
922 * Now check for the edge cases. First, if TDF_TIMEOUT is set,
923 * then the other thread has already yielded to us, so clear
924 * the flag and resume it. If TDF_TIMEOUT is not set, then the
925 * we know that the other thread is not on a sleep queue, but it
926 * hasn't resumed execution yet. In that case, set TDF_TIMOFAIL
927 * to let it know that the timeout has already run and doesn't
928 * need to be canceled.
930 if (td->td_flags & TDF_TIMEOUT) {
931 MPASS(TD_IS_SLEEPING(td));
932 td->td_flags &= ~TDF_TIMEOUT;
934 wakeup_swapper = setrunnable(td);
936 td->td_flags |= TDF_TIMOFAIL;
943 * Resumes a specific thread from the sleep queue associated with a specific
944 * wait channel if it is on that queue.
947 sleepq_remove(struct thread *td, void *wchan)
949 struct sleepqueue *sq;
953 * Look up the sleep queue for this wait channel, then re-check
954 * that the thread is asleep on that channel, if it is not, then
957 MPASS(wchan != NULL);
959 sq = sleepq_lookup(wchan);
961 * We can not lock the thread here as it may be sleeping on a
962 * different sleepq. However, holding the sleepq lock for this
963 * wchan can guarantee that we do not miss a wakeup for this
964 * channel. The asserts below will catch any false positives.
966 if (!TD_ON_SLEEPQ(td) || td->td_wchan != wchan) {
967 sleepq_release(wchan);
970 /* Thread is asleep on sleep queue sq, so wake it up. */
973 MPASS(td->td_wchan == wchan);
974 wakeup_swapper = sleepq_resume_thread(sq, td, 0);
976 sleepq_release(wchan);
982 * Abort a thread as if an interrupt had occurred. Only abort
983 * interruptible waits (unfortunately it isn't safe to abort others).
986 sleepq_abort(struct thread *td, int intrval)
988 struct sleepqueue *sq;
991 THREAD_LOCK_ASSERT(td, MA_OWNED);
992 MPASS(TD_ON_SLEEPQ(td));
993 MPASS(td->td_flags & TDF_SINTR);
994 MPASS(intrval == EINTR || intrval == ERESTART);
997 * If the TDF_TIMEOUT flag is set, just leave. A
998 * timeout is scheduled anyhow.
1000 if (td->td_flags & TDF_TIMEOUT)
1003 CTR3(KTR_PROC, "sleepq_abort: thread %p (pid %ld, %s)",
1004 (void *)td, (long)td->td_proc->p_pid, (void *)td->td_name);
1005 td->td_intrval = intrval;
1006 td->td_flags |= TDF_SLEEPABORT;
1008 * If the thread has not slept yet it will find the signal in
1009 * sleepq_catch_signals() and call sleepq_resume_thread. Otherwise
1010 * we have to do it here.
1012 if (!TD_IS_SLEEPING(td))
1014 wchan = td->td_wchan;
1015 MPASS(wchan != NULL);
1016 sq = sleepq_lookup(wchan);
1019 /* Thread is asleep on sleep queue sq, so wake it up. */
1020 return (sleepq_resume_thread(sq, td, 0));
1023 #ifdef SLEEPQUEUE_PROFILING
1024 #define SLEEPQ_PROF_LOCATIONS 1024
1025 #define SLEEPQ_SBUFSIZE 512
1026 struct sleepq_prof {
1027 LIST_ENTRY(sleepq_prof) sp_link;
1028 const char *sp_wmesg;
1032 LIST_HEAD(sqphead, sleepq_prof);
1034 struct sqphead sleepq_prof_free;
1035 struct sqphead sleepq_hash[SC_TABLESIZE];
1036 static struct sleepq_prof sleepq_profent[SLEEPQ_PROF_LOCATIONS];
1037 static struct mtx sleepq_prof_lock;
1038 MTX_SYSINIT(sleepq_prof_lock, &sleepq_prof_lock, "sleepq_prof", MTX_SPIN);
1041 sleepq_profile(const char *wmesg)
1043 struct sleepq_prof *sp;
1045 mtx_lock_spin(&sleepq_prof_lock);
1046 if (prof_enabled == 0)
1048 LIST_FOREACH(sp, &sleepq_hash[SC_HASH(wmesg)], sp_link)
1049 if (sp->sp_wmesg == wmesg)
1051 sp = LIST_FIRST(&sleepq_prof_free);
1054 sp->sp_wmesg = wmesg;
1055 LIST_REMOVE(sp, sp_link);
1056 LIST_INSERT_HEAD(&sleepq_hash[SC_HASH(wmesg)], sp, sp_link);
1060 mtx_unlock_spin(&sleepq_prof_lock);
1065 sleepq_prof_reset(void)
1067 struct sleepq_prof *sp;
1071 mtx_lock_spin(&sleepq_prof_lock);
1072 enabled = prof_enabled;
1074 for (i = 0; i < SC_TABLESIZE; i++)
1075 LIST_INIT(&sleepq_hash[i]);
1076 LIST_INIT(&sleepq_prof_free);
1077 for (i = 0; i < SLEEPQ_PROF_LOCATIONS; i++) {
1078 sp = &sleepq_profent[i];
1079 sp->sp_wmesg = NULL;
1081 LIST_INSERT_HEAD(&sleepq_prof_free, sp, sp_link);
1083 prof_enabled = enabled;
1084 mtx_unlock_spin(&sleepq_prof_lock);
1088 enable_sleepq_prof(SYSCTL_HANDLER_ARGS)
1093 error = sysctl_handle_int(oidp, &v, v, req);
1096 if (req->newptr == NULL)
1098 if (v == prof_enabled)
1101 sleepq_prof_reset();
1102 mtx_lock_spin(&sleepq_prof_lock);
1104 mtx_unlock_spin(&sleepq_prof_lock);
1110 reset_sleepq_prof_stats(SYSCTL_HANDLER_ARGS)
1115 error = sysctl_handle_int(oidp, &v, 0, req);
1118 if (req->newptr == NULL)
1122 sleepq_prof_reset();
1128 dump_sleepq_prof_stats(SYSCTL_HANDLER_ARGS)
1130 struct sleepq_prof *sp;
1136 error = sysctl_wire_old_buffer(req, 0);
1139 sb = sbuf_new_for_sysctl(NULL, NULL, SLEEPQ_SBUFSIZE, req);
1140 sbuf_printf(sb, "\nwmesg\tcount\n");
1141 enabled = prof_enabled;
1142 mtx_lock_spin(&sleepq_prof_lock);
1144 mtx_unlock_spin(&sleepq_prof_lock);
1145 for (i = 0; i < SC_TABLESIZE; i++) {
1146 LIST_FOREACH(sp, &sleepq_hash[i], sp_link) {
1147 sbuf_printf(sb, "%s\t%ld\n",
1148 sp->sp_wmesg, sp->sp_count);
1151 mtx_lock_spin(&sleepq_prof_lock);
1152 prof_enabled = enabled;
1153 mtx_unlock_spin(&sleepq_prof_lock);
1155 error = sbuf_finish(sb);
1160 SYSCTL_PROC(_debug_sleepq, OID_AUTO, stats, CTLTYPE_STRING | CTLFLAG_RD,
1161 NULL, 0, dump_sleepq_prof_stats, "A", "Sleepqueue profiling statistics");
1162 SYSCTL_PROC(_debug_sleepq, OID_AUTO, reset, CTLTYPE_INT | CTLFLAG_RW,
1163 NULL, 0, reset_sleepq_prof_stats, "I",
1164 "Reset sleepqueue profiling statistics");
1165 SYSCTL_PROC(_debug_sleepq, OID_AUTO, enable, CTLTYPE_INT | CTLFLAG_RW,
1166 NULL, 0, enable_sleepq_prof, "I", "Enable sleepqueue profiling");
1170 DB_SHOW_COMMAND(sleepq, db_show_sleepqueue)
1172 struct sleepqueue_chain *sc;
1173 struct sleepqueue *sq;
1175 struct lock_object *lock;
1185 * First, see if there is an active sleep queue for the wait channel
1186 * indicated by the address.
1188 wchan = (void *)addr;
1189 sc = SC_LOOKUP(wchan);
1190 LIST_FOREACH(sq, &sc->sc_queues, sq_hash)
1191 if (sq->sq_wchan == wchan)
1195 * Second, see if there is an active sleep queue at the address
1198 for (i = 0; i < SC_TABLESIZE; i++)
1199 LIST_FOREACH(sq, &sleepq_chains[i].sc_queues, sq_hash) {
1200 if (sq == (struct sleepqueue *)addr)
1204 db_printf("Unable to locate a sleep queue via %p\n", (void *)addr);
1207 db_printf("Wait channel: %p\n", sq->sq_wchan);
1208 db_printf("Queue type: %d\n", sq->sq_type);
1212 db_printf("Associated Interlock: %p - (%s) %s\n", lock,
1213 LOCK_CLASS(lock)->lc_name, lock->lo_name);
1216 db_printf("Blocked threads:\n");
1217 for (i = 0; i < NR_SLEEPQS; i++) {
1218 db_printf("\nQueue[%d]:\n", i);
1219 if (TAILQ_EMPTY(&sq->sq_blocked[i]))
1220 db_printf("\tempty\n");
1222 TAILQ_FOREACH(td, &sq->sq_blocked[0],
1224 db_printf("\t%p (tid %d, pid %d, \"%s\")\n", td,
1225 td->td_tid, td->td_proc->p_pid,
1228 db_printf("(expected: %u)\n", sq->sq_blockedcnt[i]);
1232 /* Alias 'show sleepqueue' to 'show sleepq'. */
1233 DB_SHOW_ALIAS(sleepqueue, db_show_sleepqueue);