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 are 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 its 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>
83 #include <machine/atomic.h>
93 * Constants for the hash table of sleep queue chains.
94 * SC_TABLESIZE must be a power of two for SC_MASK to work properly.
96 #define SC_TABLESIZE 256 /* Must be power of 2. */
97 #define SC_MASK (SC_TABLESIZE - 1)
99 #define SC_HASH(wc) ((((uintptr_t)(wc) >> SC_SHIFT) ^ (uintptr_t)(wc)) & \
101 #define SC_LOOKUP(wc) &sleepq_chains[SC_HASH(wc)]
104 * There are 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
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
119 * c - sleep queue chain lock
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. */
129 struct lock_object *sq_lock; /* (c) Associated lock. */
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. */
142 #ifdef SLEEPQUEUE_PROFILING
143 u_int sleepq_max_depth;
144 static SYSCTL_NODE(_debug, OID_AUTO, sleepq, CTLFLAG_RD, 0, "sleepq profiling");
145 static SYSCTL_NODE(_debug_sleepq, OID_AUTO, chains, CTLFLAG_RD, 0,
146 "sleepq chain stats");
147 SYSCTL_UINT(_debug_sleepq, OID_AUTO, max_depth, CTLFLAG_RD, &sleepq_max_depth,
148 0, "maxmimum depth achieved of a single chain");
150 static void sleepq_profile(const char *wmesg);
151 static int prof_enabled;
153 static struct sleepqueue_chain sleepq_chains[SC_TABLESIZE];
154 static uma_zone_t sleepq_zone;
157 * Prototypes for non-exported routines.
159 static int sleepq_catch_signals(void *wchan, int pri);
160 static int sleepq_check_signals(void);
161 static int sleepq_check_timeout(void);
163 static void sleepq_dtor(void *mem, int size, void *arg);
165 static int sleepq_init(void *mem, int size, int flags);
166 static int sleepq_resume_thread(struct sleepqueue *sq, struct thread *td,
168 static void sleepq_switch(void *wchan, int pri);
169 static void sleepq_timeout(void *arg);
171 SDT_PROBE_DECLARE(sched, , , sleep);
172 SDT_PROBE_DECLARE(sched, , , wakeup);
175 * Initialize SLEEPQUEUE_PROFILING specific sysctl nodes.
176 * Note that it must happen after sleepinit() has been fully executed, so
177 * it must happen after SI_SUB_KMEM SYSINIT() subsystem setup.
179 #ifdef SLEEPQUEUE_PROFILING
181 init_sleepqueue_profiling(void)
184 struct sysctl_oid *chain_oid;
187 for (i = 0; i < SC_TABLESIZE; i++) {
188 snprintf(chain_name, sizeof(chain_name), "%u", i);
189 chain_oid = SYSCTL_ADD_NODE(NULL,
190 SYSCTL_STATIC_CHILDREN(_debug_sleepq_chains), OID_AUTO,
191 chain_name, CTLFLAG_RD, NULL, "sleepq chain stats");
192 SYSCTL_ADD_UINT(NULL, SYSCTL_CHILDREN(chain_oid), OID_AUTO,
193 "depth", CTLFLAG_RD, &sleepq_chains[i].sc_depth, 0, NULL);
194 SYSCTL_ADD_UINT(NULL, SYSCTL_CHILDREN(chain_oid), OID_AUTO,
195 "max_depth", CTLFLAG_RD, &sleepq_chains[i].sc_max_depth, 0,
200 SYSINIT(sleepqueue_profiling, SI_SUB_LOCK, SI_ORDER_ANY,
201 init_sleepqueue_profiling, NULL);
205 * Early initialization of sleep queues that is called from the sleepinit()
209 init_sleepqueues(void)
213 for (i = 0; i < SC_TABLESIZE; i++) {
214 LIST_INIT(&sleepq_chains[i].sc_queues);
215 mtx_init(&sleepq_chains[i].sc_lock, "sleepq chain", NULL,
216 MTX_SPIN | MTX_RECURSE);
218 sleepq_zone = uma_zcreate("SLEEPQUEUE", sizeof(struct sleepqueue),
220 NULL, sleepq_dtor, sleepq_init, NULL, UMA_ALIGN_CACHE, 0);
222 NULL, NULL, sleepq_init, NULL, UMA_ALIGN_CACHE, 0);
225 thread0.td_sleepqueue = sleepq_alloc();
229 * Get a sleep queue for a new thread.
235 return (uma_zalloc(sleepq_zone, M_WAITOK));
239 * Free a sleep queue when a thread is destroyed.
242 sleepq_free(struct sleepqueue *sq)
245 uma_zfree(sleepq_zone, sq);
249 * Lock the sleep queue chain associated with the specified wait channel.
252 sleepq_lock(void *wchan)
254 struct sleepqueue_chain *sc;
256 sc = SC_LOOKUP(wchan);
257 mtx_lock_spin(&sc->sc_lock);
261 * Look up the sleep queue associated with a given wait channel in the hash
262 * table locking the associated sleep queue chain. If no queue is found in
263 * the table, NULL is returned.
266 sleepq_lookup(void *wchan)
268 struct sleepqueue_chain *sc;
269 struct sleepqueue *sq;
271 KASSERT(wchan != NULL, ("%s: invalid NULL wait channel", __func__));
272 sc = SC_LOOKUP(wchan);
273 mtx_assert(&sc->sc_lock, MA_OWNED);
274 LIST_FOREACH(sq, &sc->sc_queues, sq_hash)
275 if (sq->sq_wchan == wchan)
281 * Unlock the sleep queue chain associated with a given wait channel.
284 sleepq_release(void *wchan)
286 struct sleepqueue_chain *sc;
288 sc = SC_LOOKUP(wchan);
289 mtx_unlock_spin(&sc->sc_lock);
293 * Places the current thread on the sleep queue for the specified wait
294 * channel. If INVARIANTS is enabled, then it associates the passed in
295 * lock with the sleepq to make sure it is held when that sleep queue is
299 sleepq_add(void *wchan, struct lock_object *lock, const char *wmesg, int flags,
302 struct sleepqueue_chain *sc;
303 struct sleepqueue *sq;
307 sc = SC_LOOKUP(wchan);
308 mtx_assert(&sc->sc_lock, MA_OWNED);
309 MPASS(td->td_sleepqueue != NULL);
310 MPASS(wchan != NULL);
311 MPASS((queue >= 0) && (queue < NR_SLEEPQS));
313 /* If this thread is not allowed to sleep, die a horrible death. */
314 KASSERT(td->td_no_sleeping == 0,
315 ("%s: td %p to sleep on wchan %p with sleeping prohibited",
316 __func__, td, wchan));
318 /* Look up the sleep queue associated with the wait channel 'wchan'. */
319 sq = sleepq_lookup(wchan);
322 * If the wait channel does not already have a sleep queue, use
323 * this thread's sleep queue. Otherwise, insert the current thread
324 * into the sleep queue already in use by this wait channel.
330 sq = td->td_sleepqueue;
331 for (i = 0; i < NR_SLEEPQS; i++) {
332 KASSERT(TAILQ_EMPTY(&sq->sq_blocked[i]),
333 ("thread's sleep queue %d is not empty", i));
334 KASSERT(sq->sq_blockedcnt[i] == 0,
335 ("thread's sleep queue %d count mismatches", i));
337 KASSERT(LIST_EMPTY(&sq->sq_free),
338 ("thread's sleep queue has a non-empty free list"));
339 KASSERT(sq->sq_wchan == NULL, ("stale sq_wchan pointer"));
342 #ifdef SLEEPQUEUE_PROFILING
344 if (sc->sc_depth > sc->sc_max_depth) {
345 sc->sc_max_depth = sc->sc_depth;
346 if (sc->sc_max_depth > sleepq_max_depth)
347 sleepq_max_depth = sc->sc_max_depth;
350 sq = td->td_sleepqueue;
351 LIST_INSERT_HEAD(&sc->sc_queues, sq, sq_hash);
352 sq->sq_wchan = wchan;
353 sq->sq_type = flags & SLEEPQ_TYPE;
355 MPASS(wchan == sq->sq_wchan);
356 MPASS(lock == sq->sq_lock);
357 MPASS((flags & SLEEPQ_TYPE) == sq->sq_type);
358 LIST_INSERT_HEAD(&sq->sq_free, td->td_sleepqueue, sq_hash);
361 TAILQ_INSERT_TAIL(&sq->sq_blocked[queue], td, td_slpq);
362 sq->sq_blockedcnt[queue]++;
363 td->td_sleepqueue = NULL;
364 td->td_sqqueue = queue;
365 td->td_wchan = wchan;
366 td->td_wmesg = wmesg;
367 if (flags & SLEEPQ_INTERRUPTIBLE) {
368 td->td_flags |= TDF_SINTR;
369 td->td_flags &= ~TDF_SLEEPABORT;
375 * Sets a timeout that will remove the current thread from the specified
376 * sleep queue after timo ticks if the thread has not already been awakened.
379 sleepq_set_timeout_sbt(void *wchan, sbintime_t sbt, sbintime_t pr,
382 struct sleepqueue_chain *sc;
387 sc = SC_LOOKUP(wchan);
388 mtx_assert(&sc->sc_lock, MA_OWNED);
389 MPASS(TD_ON_SLEEPQ(td));
390 MPASS(td->td_sleepqueue == NULL);
391 MPASS(wchan != NULL);
392 if (cold && td == &thread0)
393 panic("timed sleep before timers are working");
394 KASSERT(td->td_sleeptimo == 0, ("td %d %p td_sleeptimo %jx",
395 td->td_tid, td, (uintmax_t)td->td_sleeptimo));
397 callout_when(sbt, pr, flags, &td->td_sleeptimo, &pr1);
399 callout_reset_sbt_on(&td->td_slpcallout, td->td_sleeptimo, pr1,
400 sleepq_timeout, td, PCPU_GET(cpuid), flags | C_PRECALC |
405 * Return the number of actual sleepers for the specified queue.
408 sleepq_sleepcnt(void *wchan, int queue)
410 struct sleepqueue *sq;
412 KASSERT(wchan != NULL, ("%s: invalid NULL wait channel", __func__));
413 MPASS((queue >= 0) && (queue < NR_SLEEPQS));
414 sq = sleepq_lookup(wchan);
417 return (sq->sq_blockedcnt[queue]);
421 * Marks the pending sleep of the current thread as interruptible and
422 * makes an initial check for pending signals before putting a thread
423 * to sleep. Enters and exits with the thread lock held. Thread lock
424 * may have transitioned from the sleepq lock to a run lock.
427 sleepq_catch_signals(void *wchan, int pri)
429 struct sleepqueue_chain *sc;
430 struct sleepqueue *sq;
439 sc = SC_LOOKUP(wchan);
440 mtx_assert(&sc->sc_lock, MA_OWNED);
441 MPASS(wchan != NULL);
442 if ((td->td_pflags & TDP_WAKEUP) != 0) {
443 td->td_pflags &= ~TDP_WAKEUP;
450 * See if there are any pending signals or suspension requests for this
451 * thread. If not, we can switch immediately.
454 if ((td->td_flags & (TDF_NEEDSIGCHK | TDF_NEEDSUSPCHK)) != 0) {
456 mtx_unlock_spin(&sc->sc_lock);
457 CTR3(KTR_PROC, "sleepq catching signals: thread %p (pid %ld, %s)",
458 (void *)td, (long)p->p_pid, td->td_name);
461 * Check for suspension first. Checking for signals and then
462 * suspending could result in a missed signal, since a signal
463 * can be delivered while this thread is suspended.
465 if ((td->td_flags & TDF_NEEDSUSPCHK) != 0) {
466 ret = thread_suspend_check(1);
467 MPASS(ret == 0 || ret == EINTR || ret == ERESTART);
470 mtx_lock_spin(&sc->sc_lock);
475 if ((td->td_flags & TDF_NEEDSIGCHK) != 0) {
477 mtx_lock(&ps->ps_mtx);
480 mtx_unlock(&ps->ps_mtx);
481 KASSERT((td->td_flags & TDF_SBDRY) != 0,
483 KASSERT(TD_SBDRY_INTR(td),
484 ("lost TDF_SERESTART of TDF_SEINTR"));
485 KASSERT((td->td_flags &
486 (TDF_SEINTR | TDF_SERESTART)) !=
487 (TDF_SEINTR | TDF_SERESTART),
488 ("both TDF_SEINTR and TDF_SERESTART"));
489 ret = TD_SBDRY_ERRNO(td);
490 } else if (sig != 0) {
491 ret = SIGISMEMBER(ps->ps_sigintr, sig) ?
493 mtx_unlock(&ps->ps_mtx);
495 mtx_unlock(&ps->ps_mtx);
499 * Lock the per-process spinlock prior to dropping the PROC_LOCK
500 * to avoid a signal delivery race. PROC_LOCK, PROC_SLOCK, and
501 * thread_lock() are currently held in tdsendsignal().
504 mtx_lock_spin(&sc->sc_lock);
510 sleepq_switch(wchan, pri);
515 * There were pending signals and this thread is still
516 * on the sleep queue, remove it from the sleep queue.
518 if (TD_ON_SLEEPQ(td)) {
519 sq = sleepq_lookup(wchan);
520 if (sleepq_resume_thread(sq, td, 0)) {
523 * This thread hasn't gone to sleep yet, so it
524 * should not be swapped out.
526 panic("not waking up swapper");
530 mtx_unlock_spin(&sc->sc_lock);
531 MPASS(td->td_lock != &sc->sc_lock);
536 * Switches to another thread if we are still asleep on a sleep queue.
537 * Returns with thread lock.
540 sleepq_switch(void *wchan, int pri)
542 struct sleepqueue_chain *sc;
543 struct sleepqueue *sq;
548 sc = SC_LOOKUP(wchan);
549 mtx_assert(&sc->sc_lock, MA_OWNED);
550 THREAD_LOCK_ASSERT(td, MA_OWNED);
553 * If we have a sleep queue, then we've already been woken up, so
556 if (td->td_sleepqueue != NULL) {
557 mtx_unlock_spin(&sc->sc_lock);
562 * If TDF_TIMEOUT is set, then our sleep has been timed out
563 * already but we are still on the sleep queue, so dequeue the
566 * Do the same if the real-time clock has been adjusted since this
567 * thread calculated its timeout based on that clock. This handles
568 * the following race:
569 * - The Ts thread needs to sleep until an absolute real-clock time.
570 * It copies the global rtc_generation into curthread->td_rtcgen,
571 * reads the RTC, and calculates a sleep duration based on that time.
572 * See umtxq_sleep() for an example.
573 * - The Tc thread adjusts the RTC, bumps rtc_generation, and wakes
574 * threads that are sleeping until an absolute real-clock time.
575 * See tc_setclock() and the POSIX specification of clock_settime().
576 * - Ts reaches the code below. It holds the sleepqueue chain lock,
577 * so Tc has finished waking, so this thread must test td_rtcgen.
578 * (The declaration of td_rtcgen refers to this comment.)
580 rtc_changed = td->td_rtcgen != 0 && td->td_rtcgen != rtc_generation;
581 if ((td->td_flags & TDF_TIMEOUT) || rtc_changed) {
585 MPASS(TD_ON_SLEEPQ(td));
586 sq = sleepq_lookup(wchan);
587 if (sleepq_resume_thread(sq, td, 0)) {
590 * This thread hasn't gone to sleep yet, so it
591 * should not be swapped out.
593 panic("not waking up swapper");
596 mtx_unlock_spin(&sc->sc_lock);
599 #ifdef SLEEPQUEUE_PROFILING
601 sleepq_profile(td->td_wmesg);
603 MPASS(td->td_sleepqueue == NULL);
604 sched_sleep(td, pri);
605 thread_lock_set(td, &sc->sc_lock);
606 SDT_PROBE0(sched, , , sleep);
608 mi_switch(SW_VOL | SWT_SLEEPQ, NULL);
609 KASSERT(TD_IS_RUNNING(td), ("running but not TDS_RUNNING"));
610 CTR3(KTR_PROC, "sleepq resume: thread %p (pid %ld, %s)",
611 (void *)td, (long)td->td_proc->p_pid, (void *)td->td_name);
615 * Check to see if we timed out.
618 sleepq_check_timeout(void)
624 THREAD_LOCK_ASSERT(td, MA_OWNED);
627 * If TDF_TIMEOUT is set, we timed out. But recheck
628 * td_sleeptimo anyway.
631 if (td->td_sleeptimo != 0) {
632 if (td->td_sleeptimo <= sbinuptime())
634 td->td_sleeptimo = 0;
636 if (td->td_flags & TDF_TIMEOUT)
637 td->td_flags &= ~TDF_TIMEOUT;
640 * We ignore the situation where timeout subsystem was
641 * unable to stop our callout. The struct thread is
642 * type-stable, the callout will use the correct
643 * memory when running. The checks of the
644 * td_sleeptimo value in this function and in
645 * sleepq_timeout() ensure that the thread does not
646 * get spurious wakeups, even if the callout was reset
649 callout_stop(&td->td_slpcallout);
654 * Check to see if we were awoken by a signal.
657 sleepq_check_signals(void)
662 THREAD_LOCK_ASSERT(td, MA_OWNED);
664 /* We are no longer in an interruptible sleep. */
665 if (td->td_flags & TDF_SINTR)
666 td->td_flags &= ~TDF_SINTR;
668 if (td->td_flags & TDF_SLEEPABORT) {
669 td->td_flags &= ~TDF_SLEEPABORT;
670 return (td->td_intrval);
677 * Block the current thread until it is awakened from its sleep queue.
680 sleepq_wait(void *wchan, int pri)
685 MPASS(!(td->td_flags & TDF_SINTR));
687 sleepq_switch(wchan, pri);
692 * Block the current thread until it is awakened from its sleep queue
693 * or it is interrupted by a signal.
696 sleepq_wait_sig(void *wchan, int pri)
701 rcatch = sleepq_catch_signals(wchan, pri);
702 rval = sleepq_check_signals();
703 thread_unlock(curthread);
710 * Block the current thread until it is awakened from its sleep queue
711 * or it times out while waiting.
714 sleepq_timedwait(void *wchan, int pri)
720 MPASS(!(td->td_flags & TDF_SINTR));
722 sleepq_switch(wchan, pri);
723 rval = sleepq_check_timeout();
730 * Block the current thread until it is awakened from its sleep queue,
731 * it is interrupted by a signal, or it times out waiting to be awakened.
734 sleepq_timedwait_sig(void *wchan, int pri)
736 int rcatch, rvalt, rvals;
738 rcatch = sleepq_catch_signals(wchan, pri);
739 rvalt = sleepq_check_timeout();
740 rvals = sleepq_check_signals();
741 thread_unlock(curthread);
750 * Returns the type of sleepqueue given a waitchannel.
753 sleepq_type(void *wchan)
755 struct sleepqueue *sq;
758 MPASS(wchan != NULL);
761 sq = sleepq_lookup(wchan);
763 sleepq_release(wchan);
767 sleepq_release(wchan);
772 * Removes a thread from a sleep queue and makes it
776 sleepq_resume_thread(struct sleepqueue *sq, struct thread *td, int pri)
778 struct sleepqueue_chain *sc;
781 MPASS(sq->sq_wchan != NULL);
782 MPASS(td->td_wchan == sq->sq_wchan);
783 MPASS(td->td_sqqueue < NR_SLEEPQS && td->td_sqqueue >= 0);
784 THREAD_LOCK_ASSERT(td, MA_OWNED);
785 sc = SC_LOOKUP(sq->sq_wchan);
786 mtx_assert(&sc->sc_lock, MA_OWNED);
788 SDT_PROBE2(sched, , , wakeup, td, td->td_proc);
790 /* Remove the thread from the queue. */
791 sq->sq_blockedcnt[td->td_sqqueue]--;
792 TAILQ_REMOVE(&sq->sq_blocked[td->td_sqqueue], td, td_slpq);
795 * Get a sleep queue for this thread. If this is the last waiter,
796 * use the queue itself and take it out of the chain, otherwise,
797 * remove a queue from the free list.
799 if (LIST_EMPTY(&sq->sq_free)) {
800 td->td_sleepqueue = sq;
804 #ifdef SLEEPQUEUE_PROFILING
808 td->td_sleepqueue = LIST_FIRST(&sq->sq_free);
809 LIST_REMOVE(td->td_sleepqueue, sq_hash);
813 td->td_flags &= ~TDF_SINTR;
815 CTR3(KTR_PROC, "sleepq_wakeup: thread %p (pid %ld, %s)",
816 (void *)td, (long)td->td_proc->p_pid, td->td_name);
818 /* Adjust priority if requested. */
819 MPASS(pri == 0 || (pri >= PRI_MIN && pri <= PRI_MAX));
820 if (pri != 0 && td->td_priority > pri &&
821 PRI_BASE(td->td_pri_class) == PRI_TIMESHARE)
825 * Note that thread td might not be sleeping if it is running
826 * sleepq_catch_signals() on another CPU or is blocked on its
827 * proc lock to check signals. There's no need to mark the
828 * thread runnable in that case.
830 if (TD_IS_SLEEPING(td)) {
832 return (setrunnable(td));
839 * UMA zone item deallocator.
842 sleepq_dtor(void *mem, int size, void *arg)
844 struct sleepqueue *sq;
848 for (i = 0; i < NR_SLEEPQS; i++) {
849 MPASS(TAILQ_EMPTY(&sq->sq_blocked[i]));
850 MPASS(sq->sq_blockedcnt[i] == 0);
856 * UMA zone item initializer.
859 sleepq_init(void *mem, int size, int flags)
861 struct sleepqueue *sq;
866 for (i = 0; i < NR_SLEEPQS; i++) {
867 TAILQ_INIT(&sq->sq_blocked[i]);
868 sq->sq_blockedcnt[i] = 0;
870 LIST_INIT(&sq->sq_free);
875 * Find the highest priority thread sleeping on a wait channel and resume it.
878 sleepq_signal(void *wchan, int flags, int pri, int queue)
880 struct sleepqueue *sq;
881 struct thread *td, *besttd;
884 CTR2(KTR_PROC, "sleepq_signal(%p, %d)", wchan, flags);
885 KASSERT(wchan != NULL, ("%s: invalid NULL wait channel", __func__));
886 MPASS((queue >= 0) && (queue < NR_SLEEPQS));
887 sq = sleepq_lookup(wchan);
890 KASSERT(sq->sq_type == (flags & SLEEPQ_TYPE),
891 ("%s: mismatch between sleep/wakeup and cv_*", __func__));
894 * Find the highest priority thread on the queue. If there is a
895 * tie, use the thread that first appears in the queue as it has
896 * been sleeping the longest since threads are always added to
897 * the tail of sleep queues.
899 besttd = TAILQ_FIRST(&sq->sq_blocked[queue]);
900 TAILQ_FOREACH(td, &sq->sq_blocked[queue], td_slpq) {
901 if (td->td_priority < besttd->td_priority)
904 MPASS(besttd != NULL);
906 wakeup_swapper = sleepq_resume_thread(sq, besttd, pri);
907 thread_unlock(besttd);
908 return (wakeup_swapper);
912 match_any(struct thread *td __unused)
919 * Resume all threads sleeping on a specified wait channel.
922 sleepq_broadcast(void *wchan, int flags, int pri, int queue)
924 struct sleepqueue *sq;
926 CTR2(KTR_PROC, "sleepq_broadcast(%p, %d)", wchan, flags);
927 KASSERT(wchan != NULL, ("%s: invalid NULL wait channel", __func__));
928 MPASS((queue >= 0) && (queue < NR_SLEEPQS));
929 sq = sleepq_lookup(wchan);
932 KASSERT(sq->sq_type == (flags & SLEEPQ_TYPE),
933 ("%s: mismatch between sleep/wakeup and cv_*", __func__));
935 return (sleepq_remove_matching(sq, queue, match_any, pri));
939 * Resume threads on the sleep queue that match the given predicate.
942 sleepq_remove_matching(struct sleepqueue *sq, int queue,
943 bool (*matches)(struct thread *), int pri)
945 struct thread *td, *tdn;
949 * The last thread will be given ownership of sq and may
950 * re-enqueue itself before sleepq_resume_thread() returns,
951 * so we must cache the "next" queue item at the beginning
952 * of the final iteration.
955 TAILQ_FOREACH_SAFE(td, &sq->sq_blocked[queue], td_slpq, tdn) {
958 wakeup_swapper |= sleepq_resume_thread(sq, td, pri);
962 return (wakeup_swapper);
966 * Time sleeping threads out. When the timeout expires, the thread is
967 * removed from the sleep queue and made runnable if it is still asleep.
970 sleepq_timeout(void *arg)
972 struct sleepqueue_chain *sc;
973 struct sleepqueue *sq;
980 CTR3(KTR_PROC, "sleepq_timeout: thread %p (pid %ld, %s)",
981 (void *)td, (long)td->td_proc->p_pid, (void *)td->td_name);
985 if (td->td_sleeptimo > sbinuptime() || td->td_sleeptimo == 0) {
987 * The thread does not want a timeout (yet).
989 } else if (TD_IS_SLEEPING(td) && TD_ON_SLEEPQ(td)) {
991 * See if the thread is asleep and get the wait
994 wchan = td->td_wchan;
995 sc = SC_LOOKUP(wchan);
996 THREAD_LOCKPTR_ASSERT(td, &sc->sc_lock);
997 sq = sleepq_lookup(wchan);
999 td->td_flags |= TDF_TIMEOUT;
1000 wakeup_swapper = sleepq_resume_thread(sq, td, 0);
1001 } else if (TD_ON_SLEEPQ(td)) {
1003 * If the thread is on the SLEEPQ but isn't sleeping
1004 * yet, it can either be on another CPU in between
1005 * sleepq_add() and one of the sleepq_*wait*()
1006 * routines or it can be in sleepq_catch_signals().
1008 td->td_flags |= TDF_TIMEOUT;
1017 * Resumes a specific thread from the sleep queue associated with a specific
1018 * wait channel if it is on that queue.
1021 sleepq_remove(struct thread *td, void *wchan)
1023 struct sleepqueue *sq;
1027 * Look up the sleep queue for this wait channel, then re-check
1028 * that the thread is asleep on that channel, if it is not, then
1031 MPASS(wchan != NULL);
1033 sq = sleepq_lookup(wchan);
1035 * We can not lock the thread here as it may be sleeping on a
1036 * different sleepq. However, holding the sleepq lock for this
1037 * wchan can guarantee that we do not miss a wakeup for this
1038 * channel. The asserts below will catch any false positives.
1040 if (!TD_ON_SLEEPQ(td) || td->td_wchan != wchan) {
1041 sleepq_release(wchan);
1044 /* Thread is asleep on sleep queue sq, so wake it up. */
1047 MPASS(td->td_wchan == wchan);
1048 wakeup_swapper = sleepq_resume_thread(sq, td, 0);
1050 sleepq_release(wchan);
1056 * Abort a thread as if an interrupt had occurred. Only abort
1057 * interruptible waits (unfortunately it isn't safe to abort others).
1060 sleepq_abort(struct thread *td, int intrval)
1062 struct sleepqueue *sq;
1065 THREAD_LOCK_ASSERT(td, MA_OWNED);
1066 MPASS(TD_ON_SLEEPQ(td));
1067 MPASS(td->td_flags & TDF_SINTR);
1068 MPASS(intrval == EINTR || intrval == ERESTART);
1071 * If the TDF_TIMEOUT flag is set, just leave. A
1072 * timeout is scheduled anyhow.
1074 if (td->td_flags & TDF_TIMEOUT)
1077 CTR3(KTR_PROC, "sleepq_abort: thread %p (pid %ld, %s)",
1078 (void *)td, (long)td->td_proc->p_pid, (void *)td->td_name);
1079 td->td_intrval = intrval;
1080 td->td_flags |= TDF_SLEEPABORT;
1082 * If the thread has not slept yet it will find the signal in
1083 * sleepq_catch_signals() and call sleepq_resume_thread. Otherwise
1084 * we have to do it here.
1086 if (!TD_IS_SLEEPING(td))
1088 wchan = td->td_wchan;
1089 MPASS(wchan != NULL);
1090 sq = sleepq_lookup(wchan);
1093 /* Thread is asleep on sleep queue sq, so wake it up. */
1094 return (sleepq_resume_thread(sq, td, 0));
1098 sleepq_chains_remove_matching(bool (*matches)(struct thread *))
1100 struct sleepqueue_chain *sc;
1101 struct sleepqueue *sq;
1102 int i, wakeup_swapper;
1105 for (sc = &sleepq_chains[0]; sc < sleepq_chains + SC_TABLESIZE; ++sc) {
1106 if (LIST_EMPTY(&sc->sc_queues)) {
1109 mtx_lock_spin(&sc->sc_lock);
1110 LIST_FOREACH(sq, &sc->sc_queues, sq_hash) {
1111 for (i = 0; i < NR_SLEEPQS; ++i) {
1112 wakeup_swapper |= sleepq_remove_matching(sq, i,
1116 mtx_unlock_spin(&sc->sc_lock);
1118 if (wakeup_swapper) {
1124 * Prints the stacks of all threads presently sleeping on wchan/queue to
1125 * the sbuf sb. Sets count_stacks_printed to the number of stacks actually
1126 * printed. Typically, this will equal the number of threads sleeping on the
1127 * queue, but may be less if sb overflowed before all stacks were printed.
1131 sleepq_sbuf_print_stacks(struct sbuf *sb, void *wchan, int queue,
1132 int *count_stacks_printed)
1134 struct thread *td, *td_next;
1135 struct sleepqueue *sq;
1137 struct sbuf **td_infos;
1138 int i, stack_idx, error, stacks_to_allocate;
1139 bool finished, partial_print;
1143 partial_print = false;
1145 KASSERT(wchan != NULL, ("%s: invalid NULL wait channel", __func__));
1146 MPASS((queue >= 0) && (queue < NR_SLEEPQS));
1148 stacks_to_allocate = 10;
1149 for (i = 0; i < 3 && !finished ; i++) {
1150 /* We cannot malloc while holding the queue's spinlock, so
1151 * we do our mallocs now, and hope it is enough. If it
1152 * isn't, we will free these, drop the lock, malloc more,
1153 * and try again, up to a point. After that point we will
1154 * give up and report ENOMEM. We also cannot write to sb
1155 * during this time since the client may have set the
1156 * SBUF_AUTOEXTEND flag on their sbuf, which could cause a
1157 * malloc as we print to it. So we defer actually printing
1158 * to sb until after we drop the spinlock.
1161 /* Where we will store the stacks. */
1162 st = malloc(sizeof(struct stack *) * stacks_to_allocate,
1164 for (stack_idx = 0; stack_idx < stacks_to_allocate;
1166 st[stack_idx] = stack_create();
1168 /* Where we will store the td name, tid, etc. */
1169 td_infos = malloc(sizeof(struct sbuf *) * stacks_to_allocate,
1171 for (stack_idx = 0; stack_idx < stacks_to_allocate;
1173 td_infos[stack_idx] = sbuf_new(NULL, NULL,
1174 MAXCOMLEN + sizeof(struct thread *) * 2 + 40,
1178 sq = sleepq_lookup(wchan);
1180 /* This sleepq does not exist; exit and return ENOENT. */
1183 sleepq_release(wchan);
1188 /* Save thread info */
1189 TAILQ_FOREACH_SAFE(td, &sq->sq_blocked[queue], td_slpq,
1191 if (stack_idx >= stacks_to_allocate)
1194 /* Note the td_lock is equal to the sleepq_lock here. */
1195 stack_save_td(st[stack_idx], td);
1197 sbuf_printf(td_infos[stack_idx], "%d: %s %p",
1198 td->td_tid, td->td_name, td);
1204 sleepq_release(wchan);
1206 /* Print the stacks */
1207 for (i = 0; i < stack_idx; i++) {
1208 sbuf_finish(td_infos[i]);
1209 sbuf_printf(sb, "--- thread %s: ---\n", sbuf_data(td_infos[i]));
1210 stack_sbuf_print(sb, st[i]);
1211 sbuf_printf(sb, "\n");
1213 error = sbuf_error(sb);
1215 *count_stacks_printed = stack_idx;
1220 sleepq_release(wchan);
1221 for (stack_idx = 0; stack_idx < stacks_to_allocate;
1223 stack_destroy(st[stack_idx]);
1224 for (stack_idx = 0; stack_idx < stacks_to_allocate;
1226 sbuf_delete(td_infos[stack_idx]);
1228 free(td_infos, M_TEMP);
1229 stacks_to_allocate *= 10;
1232 if (!finished && error == 0)
1239 #ifdef SLEEPQUEUE_PROFILING
1240 #define SLEEPQ_PROF_LOCATIONS 1024
1241 #define SLEEPQ_SBUFSIZE 512
1242 struct sleepq_prof {
1243 LIST_ENTRY(sleepq_prof) sp_link;
1244 const char *sp_wmesg;
1248 LIST_HEAD(sqphead, sleepq_prof);
1250 struct sqphead sleepq_prof_free;
1251 struct sqphead sleepq_hash[SC_TABLESIZE];
1252 static struct sleepq_prof sleepq_profent[SLEEPQ_PROF_LOCATIONS];
1253 static struct mtx sleepq_prof_lock;
1254 MTX_SYSINIT(sleepq_prof_lock, &sleepq_prof_lock, "sleepq_prof", MTX_SPIN);
1257 sleepq_profile(const char *wmesg)
1259 struct sleepq_prof *sp;
1261 mtx_lock_spin(&sleepq_prof_lock);
1262 if (prof_enabled == 0)
1264 LIST_FOREACH(sp, &sleepq_hash[SC_HASH(wmesg)], sp_link)
1265 if (sp->sp_wmesg == wmesg)
1267 sp = LIST_FIRST(&sleepq_prof_free);
1270 sp->sp_wmesg = wmesg;
1271 LIST_REMOVE(sp, sp_link);
1272 LIST_INSERT_HEAD(&sleepq_hash[SC_HASH(wmesg)], sp, sp_link);
1276 mtx_unlock_spin(&sleepq_prof_lock);
1281 sleepq_prof_reset(void)
1283 struct sleepq_prof *sp;
1287 mtx_lock_spin(&sleepq_prof_lock);
1288 enabled = prof_enabled;
1290 for (i = 0; i < SC_TABLESIZE; i++)
1291 LIST_INIT(&sleepq_hash[i]);
1292 LIST_INIT(&sleepq_prof_free);
1293 for (i = 0; i < SLEEPQ_PROF_LOCATIONS; i++) {
1294 sp = &sleepq_profent[i];
1295 sp->sp_wmesg = NULL;
1297 LIST_INSERT_HEAD(&sleepq_prof_free, sp, sp_link);
1299 prof_enabled = enabled;
1300 mtx_unlock_spin(&sleepq_prof_lock);
1304 enable_sleepq_prof(SYSCTL_HANDLER_ARGS)
1309 error = sysctl_handle_int(oidp, &v, v, req);
1312 if (req->newptr == NULL)
1314 if (v == prof_enabled)
1317 sleepq_prof_reset();
1318 mtx_lock_spin(&sleepq_prof_lock);
1320 mtx_unlock_spin(&sleepq_prof_lock);
1326 reset_sleepq_prof_stats(SYSCTL_HANDLER_ARGS)
1331 error = sysctl_handle_int(oidp, &v, 0, req);
1334 if (req->newptr == NULL)
1338 sleepq_prof_reset();
1344 dump_sleepq_prof_stats(SYSCTL_HANDLER_ARGS)
1346 struct sleepq_prof *sp;
1352 error = sysctl_wire_old_buffer(req, 0);
1355 sb = sbuf_new_for_sysctl(NULL, NULL, SLEEPQ_SBUFSIZE, req);
1356 sbuf_printf(sb, "\nwmesg\tcount\n");
1357 enabled = prof_enabled;
1358 mtx_lock_spin(&sleepq_prof_lock);
1360 mtx_unlock_spin(&sleepq_prof_lock);
1361 for (i = 0; i < SC_TABLESIZE; i++) {
1362 LIST_FOREACH(sp, &sleepq_hash[i], sp_link) {
1363 sbuf_printf(sb, "%s\t%ld\n",
1364 sp->sp_wmesg, sp->sp_count);
1367 mtx_lock_spin(&sleepq_prof_lock);
1368 prof_enabled = enabled;
1369 mtx_unlock_spin(&sleepq_prof_lock);
1371 error = sbuf_finish(sb);
1376 SYSCTL_PROC(_debug_sleepq, OID_AUTO, stats, CTLTYPE_STRING | CTLFLAG_RD,
1377 NULL, 0, dump_sleepq_prof_stats, "A", "Sleepqueue profiling statistics");
1378 SYSCTL_PROC(_debug_sleepq, OID_AUTO, reset, CTLTYPE_INT | CTLFLAG_RW,
1379 NULL, 0, reset_sleepq_prof_stats, "I",
1380 "Reset sleepqueue profiling statistics");
1381 SYSCTL_PROC(_debug_sleepq, OID_AUTO, enable, CTLTYPE_INT | CTLFLAG_RW,
1382 NULL, 0, enable_sleepq_prof, "I", "Enable sleepqueue profiling");
1386 DB_SHOW_COMMAND(sleepq, db_show_sleepqueue)
1388 struct sleepqueue_chain *sc;
1389 struct sleepqueue *sq;
1391 struct lock_object *lock;
1401 * First, see if there is an active sleep queue for the wait channel
1402 * indicated by the address.
1404 wchan = (void *)addr;
1405 sc = SC_LOOKUP(wchan);
1406 LIST_FOREACH(sq, &sc->sc_queues, sq_hash)
1407 if (sq->sq_wchan == wchan)
1411 * Second, see if there is an active sleep queue at the address
1414 for (i = 0; i < SC_TABLESIZE; i++)
1415 LIST_FOREACH(sq, &sleepq_chains[i].sc_queues, sq_hash) {
1416 if (sq == (struct sleepqueue *)addr)
1420 db_printf("Unable to locate a sleep queue via %p\n", (void *)addr);
1423 db_printf("Wait channel: %p\n", sq->sq_wchan);
1424 db_printf("Queue type: %d\n", sq->sq_type);
1428 db_printf("Associated Interlock: %p - (%s) %s\n", lock,
1429 LOCK_CLASS(lock)->lc_name, lock->lo_name);
1432 db_printf("Blocked threads:\n");
1433 for (i = 0; i < NR_SLEEPQS; i++) {
1434 db_printf("\nQueue[%d]:\n", i);
1435 if (TAILQ_EMPTY(&sq->sq_blocked[i]))
1436 db_printf("\tempty\n");
1438 TAILQ_FOREACH(td, &sq->sq_blocked[0],
1440 db_printf("\t%p (tid %d, pid %d, \"%s\")\n", td,
1441 td->td_tid, td->td_proc->p_pid,
1444 db_printf("(expected: %u)\n", sq->sq_blockedcnt[i]);
1448 /* Alias 'show sleepqueue' to 'show sleepq'. */
1449 DB_SHOW_ALIAS(sleepqueue, db_show_sleepqueue);