2 * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
4 * Copyright (c) 2004 John Baldwin <jhb@FreeBSD.org>
6 * Redistribution and use in source and binary forms, with or without
7 * modification, are permitted provided that the following conditions
9 * 1. Redistributions of source code must retain the above copyright
10 * notice, this list of conditions and the following disclaimer.
11 * 2. Redistributions in binary form must reproduce the above copyright
12 * notice, this list of conditions and the following disclaimer in the
13 * documentation and/or other materials provided with the distribution.
15 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
16 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
17 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
18 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
19 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
20 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
21 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
22 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
23 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
24 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
29 * Implementation of sleep queues used to hold queue of threads blocked on
30 * a wait channel. Sleep queues are different from turnstiles in that wait
31 * channels are not owned by anyone, so there is no priority propagation.
32 * Sleep queues can also provide a timeout and can also be interrupted by
33 * signals. That said, there are several similarities between the turnstile
34 * and sleep queue implementations. (Note: turnstiles were implemented
35 * first.) For example, both use a hash table of the same size where each
36 * bucket is referred to as a "chain" that contains both a spin lock and
37 * a linked list of queues. An individual queue is located by using a hash
38 * to pick a chain, locking the chain, and then walking the chain searching
39 * for the queue. This means that a wait channel object does not need to
40 * embed its queue head just as locks do not embed their turnstile queue
41 * head. Threads also carry around a sleep queue that they lend to the
42 * wait channel when blocking. Just as in turnstiles, the queue includes
43 * a free list of the sleep queues of other threads blocked on the same
44 * wait channel in the case of multiple waiters.
46 * Some additional functionality provided by sleep queues include the
47 * ability to set a timeout. The timeout is managed using a per-thread
48 * callout that resumes a thread if it is asleep. A thread may also
49 * catch signals while it is asleep (aka an interruptible sleep). The
50 * signal code uses sleepq_abort() to interrupt a sleeping thread. Finally,
51 * sleep queues also provide some extra assertions. One is not allowed to
52 * mix the sleep/wakeup and cv APIs for a given wait channel. Also, one
53 * must consistently use the same lock to synchronize with a wait channel,
54 * though this check is currently only a warning for sleep/wakeup due to
55 * pre-existing abuse of that API. The same lock must also be held when
56 * awakening threads, though that is currently only enforced for condition
60 #include <sys/cdefs.h>
61 __FBSDID("$FreeBSD$");
63 #include "opt_sleepqueue_profiling.h"
65 #include "opt_sched.h"
66 #include "opt_stack.h"
68 #include <sys/param.h>
69 #include <sys/systm.h>
71 #include <sys/kernel.h>
73 #include <sys/mutex.h>
76 #include <sys/sched.h>
78 #include <sys/signalvar.h>
79 #include <sys/sleepqueue.h>
80 #include <sys/stack.h>
81 #include <sys/sysctl.h>
84 #include <sys/epoch.h>
87 #include <machine/atomic.h>
97 * Constants for the hash table of sleep queue chains.
98 * SC_TABLESIZE must be a power of two for SC_MASK to work properly.
101 #define SC_TABLESIZE 256
103 CTASSERT(powerof2(SC_TABLESIZE));
104 #define SC_MASK (SC_TABLESIZE - 1)
106 #define SC_HASH(wc) ((((uintptr_t)(wc) >> SC_SHIFT) ^ (uintptr_t)(wc)) & \
108 #define SC_LOOKUP(wc) &sleepq_chains[SC_HASH(wc)]
111 * There are two different lists of sleep queues. Both lists are connected
112 * via the sq_hash entries. The first list is the sleep queue chain list
113 * that a sleep queue is on when it is attached to a wait channel. The
114 * second list is the free list hung off of a sleep queue that is attached
117 * Each sleep queue also contains the wait channel it is attached to, the
118 * list of threads blocked on that wait channel, flags specific to the
119 * wait channel, and the lock used to synchronize with a wait channel.
120 * The flags are used to catch mismatches between the various consumers
121 * of the sleep queue API (e.g. sleep/wakeup and condition variables).
122 * The lock pointer is only used when invariants are enabled for various
126 * c - sleep queue chain lock
129 struct threadqueue sq_blocked[NR_SLEEPQS]; /* (c) Blocked threads. */
130 u_int sq_blockedcnt[NR_SLEEPQS]; /* (c) N. of blocked threads. */
131 LIST_ENTRY(sleepqueue) sq_hash; /* (c) Chain and free list. */
132 LIST_HEAD(, sleepqueue) sq_free; /* (c) Free queues. */
133 void *sq_wchan; /* (c) Wait channel. */
134 int sq_type; /* (c) Queue type. */
136 struct lock_object *sq_lock; /* (c) Associated lock. */
140 struct sleepqueue_chain {
141 LIST_HEAD(, sleepqueue) sc_queues; /* List of sleep queues. */
142 struct mtx sc_lock; /* Spin lock for this chain. */
143 #ifdef SLEEPQUEUE_PROFILING
144 u_int sc_depth; /* Length of sc_queues. */
145 u_int sc_max_depth; /* Max length of sc_queues. */
147 } __aligned(CACHE_LINE_SIZE);
149 #ifdef SLEEPQUEUE_PROFILING
150 u_int sleepq_max_depth;
151 static SYSCTL_NODE(_debug, OID_AUTO, sleepq, CTLFLAG_RD, 0, "sleepq profiling");
152 static SYSCTL_NODE(_debug_sleepq, OID_AUTO, chains, CTLFLAG_RD, 0,
153 "sleepq chain stats");
154 SYSCTL_UINT(_debug_sleepq, OID_AUTO, max_depth, CTLFLAG_RD, &sleepq_max_depth,
155 0, "maxmimum depth achieved of a single chain");
157 static void sleepq_profile(const char *wmesg);
158 static int prof_enabled;
160 static struct sleepqueue_chain sleepq_chains[SC_TABLESIZE];
161 static uma_zone_t sleepq_zone;
164 * Prototypes for non-exported routines.
166 static int sleepq_catch_signals(void *wchan, int pri);
167 static int sleepq_check_signals(void);
168 static int sleepq_check_timeout(void);
170 static void sleepq_dtor(void *mem, int size, void *arg);
172 static int sleepq_init(void *mem, int size, int flags);
173 static int sleepq_resume_thread(struct sleepqueue *sq, struct thread *td,
175 static void sleepq_switch(void *wchan, int pri);
176 static void sleepq_timeout(void *arg);
178 SDT_PROBE_DECLARE(sched, , , sleep);
179 SDT_PROBE_DECLARE(sched, , , wakeup);
182 * Initialize SLEEPQUEUE_PROFILING specific sysctl nodes.
183 * Note that it must happen after sleepinit() has been fully executed, so
184 * it must happen after SI_SUB_KMEM SYSINIT() subsystem setup.
186 #ifdef SLEEPQUEUE_PROFILING
188 init_sleepqueue_profiling(void)
191 struct sysctl_oid *chain_oid;
194 for (i = 0; i < SC_TABLESIZE; i++) {
195 snprintf(chain_name, sizeof(chain_name), "%u", i);
196 chain_oid = SYSCTL_ADD_NODE(NULL,
197 SYSCTL_STATIC_CHILDREN(_debug_sleepq_chains), OID_AUTO,
198 chain_name, CTLFLAG_RD, NULL, "sleepq chain stats");
199 SYSCTL_ADD_UINT(NULL, SYSCTL_CHILDREN(chain_oid), OID_AUTO,
200 "depth", CTLFLAG_RD, &sleepq_chains[i].sc_depth, 0, NULL);
201 SYSCTL_ADD_UINT(NULL, SYSCTL_CHILDREN(chain_oid), OID_AUTO,
202 "max_depth", CTLFLAG_RD, &sleepq_chains[i].sc_max_depth, 0,
207 SYSINIT(sleepqueue_profiling, SI_SUB_LOCK, SI_ORDER_ANY,
208 init_sleepqueue_profiling, NULL);
212 * Early initialization of sleep queues that is called from the sleepinit()
216 init_sleepqueues(void)
220 for (i = 0; i < SC_TABLESIZE; i++) {
221 LIST_INIT(&sleepq_chains[i].sc_queues);
222 mtx_init(&sleepq_chains[i].sc_lock, "sleepq chain", NULL,
223 MTX_SPIN | MTX_RECURSE);
225 sleepq_zone = uma_zcreate("SLEEPQUEUE", sizeof(struct sleepqueue),
227 NULL, sleepq_dtor, sleepq_init, NULL, UMA_ALIGN_CACHE, 0);
229 NULL, NULL, sleepq_init, NULL, UMA_ALIGN_CACHE, 0);
232 thread0.td_sleepqueue = sleepq_alloc();
236 * Get a sleep queue for a new thread.
242 return (uma_zalloc(sleepq_zone, M_WAITOK));
246 * Free a sleep queue when a thread is destroyed.
249 sleepq_free(struct sleepqueue *sq)
252 uma_zfree(sleepq_zone, sq);
256 * Lock the sleep queue chain associated with the specified wait channel.
259 sleepq_lock(void *wchan)
261 struct sleepqueue_chain *sc;
263 sc = SC_LOOKUP(wchan);
264 mtx_lock_spin(&sc->sc_lock);
268 * Look up the sleep queue associated with a given wait channel in the hash
269 * table locking the associated sleep queue chain. If no queue is found in
270 * the table, NULL is returned.
273 sleepq_lookup(void *wchan)
275 struct sleepqueue_chain *sc;
276 struct sleepqueue *sq;
278 KASSERT(wchan != NULL, ("%s: invalid NULL wait channel", __func__));
279 sc = SC_LOOKUP(wchan);
280 mtx_assert(&sc->sc_lock, MA_OWNED);
281 LIST_FOREACH(sq, &sc->sc_queues, sq_hash)
282 if (sq->sq_wchan == wchan)
288 * Unlock the sleep queue chain associated with a given wait channel.
291 sleepq_release(void *wchan)
293 struct sleepqueue_chain *sc;
295 sc = SC_LOOKUP(wchan);
296 mtx_unlock_spin(&sc->sc_lock);
300 * Places the current thread on the sleep queue for the specified wait
301 * channel. If INVARIANTS is enabled, then it associates the passed in
302 * lock with the sleepq to make sure it is held when that sleep queue is
306 sleepq_add(void *wchan, struct lock_object *lock, const char *wmesg, int flags,
309 struct sleepqueue_chain *sc;
310 struct sleepqueue *sq;
314 sc = SC_LOOKUP(wchan);
315 mtx_assert(&sc->sc_lock, MA_OWNED);
316 MPASS(td->td_sleepqueue != NULL);
317 MPASS(wchan != NULL);
318 MPASS((queue >= 0) && (queue < NR_SLEEPQS));
320 /* If this thread is not allowed to sleep, die a horrible death. */
321 if (__predict_false(!THREAD_CAN_SLEEP())) {
323 epoch_trace_list(curthread);
326 ("%s: td %p to sleep on wchan %p with sleeping prohibited",
327 __func__, td, wchan));
330 /* Look up the sleep queue associated with the wait channel 'wchan'. */
331 sq = sleepq_lookup(wchan);
334 * If the wait channel does not already have a sleep queue, use
335 * this thread's sleep queue. Otherwise, insert the current thread
336 * into the sleep queue already in use by this wait channel.
342 sq = td->td_sleepqueue;
343 for (i = 0; i < NR_SLEEPQS; i++) {
344 KASSERT(TAILQ_EMPTY(&sq->sq_blocked[i]),
345 ("thread's sleep queue %d is not empty", i));
346 KASSERT(sq->sq_blockedcnt[i] == 0,
347 ("thread's sleep queue %d count mismatches", i));
349 KASSERT(LIST_EMPTY(&sq->sq_free),
350 ("thread's sleep queue has a non-empty free list"));
351 KASSERT(sq->sq_wchan == NULL, ("stale sq_wchan pointer"));
354 #ifdef SLEEPQUEUE_PROFILING
356 if (sc->sc_depth > sc->sc_max_depth) {
357 sc->sc_max_depth = sc->sc_depth;
358 if (sc->sc_max_depth > sleepq_max_depth)
359 sleepq_max_depth = sc->sc_max_depth;
362 sq = td->td_sleepqueue;
363 LIST_INSERT_HEAD(&sc->sc_queues, sq, sq_hash);
364 sq->sq_wchan = wchan;
365 sq->sq_type = flags & SLEEPQ_TYPE;
367 MPASS(wchan == sq->sq_wchan);
368 MPASS(lock == sq->sq_lock);
369 MPASS((flags & SLEEPQ_TYPE) == sq->sq_type);
370 LIST_INSERT_HEAD(&sq->sq_free, td->td_sleepqueue, sq_hash);
373 TAILQ_INSERT_TAIL(&sq->sq_blocked[queue], td, td_slpq);
374 sq->sq_blockedcnt[queue]++;
375 td->td_sleepqueue = NULL;
376 td->td_sqqueue = queue;
377 td->td_wchan = wchan;
378 td->td_wmesg = wmesg;
379 if (flags & SLEEPQ_INTERRUPTIBLE) {
380 td->td_flags |= TDF_SINTR;
381 td->td_flags &= ~TDF_SLEEPABORT;
387 * Sets a timeout that will remove the current thread from the specified
388 * sleep queue after timo ticks if the thread has not already been awakened.
391 sleepq_set_timeout_sbt(void *wchan, sbintime_t sbt, sbintime_t pr,
394 struct sleepqueue_chain *sc __unused;
399 sc = SC_LOOKUP(wchan);
400 mtx_assert(&sc->sc_lock, MA_OWNED);
401 MPASS(TD_ON_SLEEPQ(td));
402 MPASS(td->td_sleepqueue == NULL);
403 MPASS(wchan != NULL);
404 if (cold && td == &thread0)
405 panic("timed sleep before timers are working");
406 KASSERT(td->td_sleeptimo == 0, ("td %d %p td_sleeptimo %jx",
407 td->td_tid, td, (uintmax_t)td->td_sleeptimo));
409 callout_when(sbt, pr, flags, &td->td_sleeptimo, &pr1);
411 callout_reset_sbt_on(&td->td_slpcallout, td->td_sleeptimo, pr1,
412 sleepq_timeout, td, PCPU_GET(cpuid), flags | C_PRECALC |
417 * Return the number of actual sleepers for the specified queue.
420 sleepq_sleepcnt(void *wchan, int queue)
422 struct sleepqueue *sq;
424 KASSERT(wchan != NULL, ("%s: invalid NULL wait channel", __func__));
425 MPASS((queue >= 0) && (queue < NR_SLEEPQS));
426 sq = sleepq_lookup(wchan);
429 return (sq->sq_blockedcnt[queue]);
433 * Marks the pending sleep of the current thread as interruptible and
434 * makes an initial check for pending signals before putting a thread
435 * to sleep. Enters and exits with the thread lock held. Thread lock
436 * may have transitioned from the sleepq lock to a run lock.
439 sleepq_catch_signals(void *wchan, int pri)
441 struct sleepqueue_chain *sc;
442 struct sleepqueue *sq;
451 sc = SC_LOOKUP(wchan);
452 mtx_assert(&sc->sc_lock, MA_OWNED);
453 MPASS(wchan != NULL);
454 if ((td->td_pflags & TDP_WAKEUP) != 0) {
455 td->td_pflags &= ~TDP_WAKEUP;
462 * See if there are any pending signals or suspension requests for this
463 * thread. If not, we can switch immediately.
466 if ((td->td_flags & (TDF_NEEDSIGCHK | TDF_NEEDSUSPCHK)) != 0) {
468 mtx_unlock_spin(&sc->sc_lock);
469 CTR3(KTR_PROC, "sleepq catching signals: thread %p (pid %ld, %s)",
470 (void *)td, (long)p->p_pid, td->td_name);
473 * Check for suspension first. Checking for signals and then
474 * suspending could result in a missed signal, since a signal
475 * can be delivered while this thread is suspended.
477 if ((td->td_flags & TDF_NEEDSUSPCHK) != 0) {
478 ret = thread_suspend_check(1);
479 MPASS(ret == 0 || ret == EINTR || ret == ERESTART);
482 mtx_lock_spin(&sc->sc_lock);
487 if ((td->td_flags & TDF_NEEDSIGCHK) != 0) {
489 mtx_lock(&ps->ps_mtx);
492 mtx_unlock(&ps->ps_mtx);
493 KASSERT((td->td_flags & TDF_SBDRY) != 0,
495 KASSERT(TD_SBDRY_INTR(td),
496 ("lost TDF_SERESTART of TDF_SEINTR"));
497 KASSERT((td->td_flags &
498 (TDF_SEINTR | TDF_SERESTART)) !=
499 (TDF_SEINTR | TDF_SERESTART),
500 ("both TDF_SEINTR and TDF_SERESTART"));
501 ret = TD_SBDRY_ERRNO(td);
502 } else if (sig != 0) {
503 ret = SIGISMEMBER(ps->ps_sigintr, sig) ?
505 mtx_unlock(&ps->ps_mtx);
507 mtx_unlock(&ps->ps_mtx);
511 * Do not go into sleep if this thread was the
512 * ptrace(2) attach leader. cursig() consumed
513 * SIGSTOP from PT_ATTACH, but we usually act
514 * on the signal by interrupting sleep, and
515 * should do that here as well.
517 if ((td->td_dbgflags & TDB_FSTP) != 0) {
520 td->td_dbgflags &= ~TDB_FSTP;
524 * Lock the per-process spinlock prior to dropping the PROC_LOCK
525 * to avoid a signal delivery race. PROC_LOCK, PROC_SLOCK, and
526 * thread_lock() are currently held in tdsendsignal().
529 mtx_lock_spin(&sc->sc_lock);
535 sleepq_switch(wchan, pri);
540 * There were pending signals and this thread is still
541 * on the sleep queue, remove it from the sleep queue.
543 if (TD_ON_SLEEPQ(td)) {
544 sq = sleepq_lookup(wchan);
545 if (sleepq_resume_thread(sq, td, 0)) {
548 * This thread hasn't gone to sleep yet, so it
549 * should not be swapped out.
551 panic("not waking up swapper");
555 mtx_unlock_spin(&sc->sc_lock);
556 MPASS(td->td_lock != &sc->sc_lock);
561 * Switches to another thread if we are still asleep on a sleep queue.
562 * Returns with thread lock.
565 sleepq_switch(void *wchan, int pri)
567 struct sleepqueue_chain *sc;
568 struct sleepqueue *sq;
573 sc = SC_LOOKUP(wchan);
574 mtx_assert(&sc->sc_lock, MA_OWNED);
575 THREAD_LOCK_ASSERT(td, MA_OWNED);
578 * If we have a sleep queue, then we've already been woken up, so
581 if (td->td_sleepqueue != NULL) {
582 mtx_unlock_spin(&sc->sc_lock);
587 * If TDF_TIMEOUT is set, then our sleep has been timed out
588 * already but we are still on the sleep queue, so dequeue the
591 * Do the same if the real-time clock has been adjusted since this
592 * thread calculated its timeout based on that clock. This handles
593 * the following race:
594 * - The Ts thread needs to sleep until an absolute real-clock time.
595 * It copies the global rtc_generation into curthread->td_rtcgen,
596 * reads the RTC, and calculates a sleep duration based on that time.
597 * See umtxq_sleep() for an example.
598 * - The Tc thread adjusts the RTC, bumps rtc_generation, and wakes
599 * threads that are sleeping until an absolute real-clock time.
600 * See tc_setclock() and the POSIX specification of clock_settime().
601 * - Ts reaches the code below. It holds the sleepqueue chain lock,
602 * so Tc has finished waking, so this thread must test td_rtcgen.
603 * (The declaration of td_rtcgen refers to this comment.)
605 rtc_changed = td->td_rtcgen != 0 && td->td_rtcgen != rtc_generation;
606 if ((td->td_flags & TDF_TIMEOUT) || rtc_changed) {
610 MPASS(TD_ON_SLEEPQ(td));
611 sq = sleepq_lookup(wchan);
612 if (sleepq_resume_thread(sq, td, 0)) {
615 * This thread hasn't gone to sleep yet, so it
616 * should not be swapped out.
618 panic("not waking up swapper");
621 mtx_unlock_spin(&sc->sc_lock);
624 #ifdef SLEEPQUEUE_PROFILING
626 sleepq_profile(td->td_wmesg);
628 MPASS(td->td_sleepqueue == NULL);
629 sched_sleep(td, pri);
630 thread_lock_set(td, &sc->sc_lock);
631 SDT_PROBE0(sched, , , sleep);
633 mi_switch(SW_VOL | SWT_SLEEPQ, NULL);
634 KASSERT(TD_IS_RUNNING(td), ("running but not TDS_RUNNING"));
635 CTR3(KTR_PROC, "sleepq resume: thread %p (pid %ld, %s)",
636 (void *)td, (long)td->td_proc->p_pid, (void *)td->td_name);
640 * Check to see if we timed out.
643 sleepq_check_timeout(void)
649 THREAD_LOCK_ASSERT(td, MA_OWNED);
652 * If TDF_TIMEOUT is set, we timed out. But recheck
653 * td_sleeptimo anyway.
656 if (td->td_sleeptimo != 0) {
657 if (td->td_sleeptimo <= sbinuptime())
659 td->td_sleeptimo = 0;
661 if (td->td_flags & TDF_TIMEOUT)
662 td->td_flags &= ~TDF_TIMEOUT;
665 * We ignore the situation where timeout subsystem was
666 * unable to stop our callout. The struct thread is
667 * type-stable, the callout will use the correct
668 * memory when running. The checks of the
669 * td_sleeptimo value in this function and in
670 * sleepq_timeout() ensure that the thread does not
671 * get spurious wakeups, even if the callout was reset
674 callout_stop(&td->td_slpcallout);
679 * Check to see if we were awoken by a signal.
682 sleepq_check_signals(void)
687 THREAD_LOCK_ASSERT(td, MA_OWNED);
689 /* We are no longer in an interruptible sleep. */
690 if (td->td_flags & TDF_SINTR)
691 td->td_flags &= ~TDF_SINTR;
693 if (td->td_flags & TDF_SLEEPABORT) {
694 td->td_flags &= ~TDF_SLEEPABORT;
695 return (td->td_intrval);
702 * Block the current thread until it is awakened from its sleep queue.
705 sleepq_wait(void *wchan, int pri)
710 MPASS(!(td->td_flags & TDF_SINTR));
712 sleepq_switch(wchan, pri);
717 * Block the current thread until it is awakened from its sleep queue
718 * or it is interrupted by a signal.
721 sleepq_wait_sig(void *wchan, int pri)
726 rcatch = sleepq_catch_signals(wchan, pri);
727 rval = sleepq_check_signals();
728 thread_unlock(curthread);
735 * Block the current thread until it is awakened from its sleep queue
736 * or it times out while waiting.
739 sleepq_timedwait(void *wchan, int pri)
745 MPASS(!(td->td_flags & TDF_SINTR));
747 sleepq_switch(wchan, pri);
748 rval = sleepq_check_timeout();
755 * Block the current thread until it is awakened from its sleep queue,
756 * it is interrupted by a signal, or it times out waiting to be awakened.
759 sleepq_timedwait_sig(void *wchan, int pri)
761 int rcatch, rvalt, rvals;
763 rcatch = sleepq_catch_signals(wchan, pri);
764 rvalt = sleepq_check_timeout();
765 rvals = sleepq_check_signals();
766 thread_unlock(curthread);
775 * Returns the type of sleepqueue given a waitchannel.
778 sleepq_type(void *wchan)
780 struct sleepqueue *sq;
783 MPASS(wchan != NULL);
786 sq = sleepq_lookup(wchan);
788 sleepq_release(wchan);
792 sleepq_release(wchan);
797 * Removes a thread from a sleep queue and makes it
801 sleepq_resume_thread(struct sleepqueue *sq, struct thread *td, int pri)
803 struct sleepqueue_chain *sc __unused;
806 MPASS(sq->sq_wchan != NULL);
807 MPASS(td->td_wchan == sq->sq_wchan);
808 MPASS(td->td_sqqueue < NR_SLEEPQS && td->td_sqqueue >= 0);
809 THREAD_LOCK_ASSERT(td, MA_OWNED);
810 sc = SC_LOOKUP(sq->sq_wchan);
811 mtx_assert(&sc->sc_lock, MA_OWNED);
813 SDT_PROBE2(sched, , , wakeup, td, td->td_proc);
815 /* Remove the thread from the queue. */
816 sq->sq_blockedcnt[td->td_sqqueue]--;
817 TAILQ_REMOVE(&sq->sq_blocked[td->td_sqqueue], td, td_slpq);
820 * Get a sleep queue for this thread. If this is the last waiter,
821 * use the queue itself and take it out of the chain, otherwise,
822 * remove a queue from the free list.
824 if (LIST_EMPTY(&sq->sq_free)) {
825 td->td_sleepqueue = sq;
829 #ifdef SLEEPQUEUE_PROFILING
833 td->td_sleepqueue = LIST_FIRST(&sq->sq_free);
834 LIST_REMOVE(td->td_sleepqueue, sq_hash);
838 td->td_flags &= ~TDF_SINTR;
840 CTR3(KTR_PROC, "sleepq_wakeup: thread %p (pid %ld, %s)",
841 (void *)td, (long)td->td_proc->p_pid, td->td_name);
843 /* Adjust priority if requested. */
844 MPASS(pri == 0 || (pri >= PRI_MIN && pri <= PRI_MAX));
845 if (pri != 0 && td->td_priority > pri &&
846 PRI_BASE(td->td_pri_class) == PRI_TIMESHARE)
850 * Note that thread td might not be sleeping if it is running
851 * sleepq_catch_signals() on another CPU or is blocked on its
852 * proc lock to check signals. There's no need to mark the
853 * thread runnable in that case.
855 if (TD_IS_SLEEPING(td)) {
857 return (setrunnable(td));
864 * UMA zone item deallocator.
867 sleepq_dtor(void *mem, int size, void *arg)
869 struct sleepqueue *sq;
873 for (i = 0; i < NR_SLEEPQS; i++) {
874 MPASS(TAILQ_EMPTY(&sq->sq_blocked[i]));
875 MPASS(sq->sq_blockedcnt[i] == 0);
881 * UMA zone item initializer.
884 sleepq_init(void *mem, int size, int flags)
886 struct sleepqueue *sq;
891 for (i = 0; i < NR_SLEEPQS; i++) {
892 TAILQ_INIT(&sq->sq_blocked[i]);
893 sq->sq_blockedcnt[i] = 0;
895 LIST_INIT(&sq->sq_free);
900 * Find thread sleeping on a wait channel and resume it.
903 sleepq_signal(void *wchan, int flags, int pri, int queue)
905 struct sleepqueue_chain *sc;
906 struct sleepqueue *sq;
907 struct threadqueue *head;
908 struct thread *td, *besttd;
911 CTR2(KTR_PROC, "sleepq_signal(%p, %d)", wchan, flags);
912 KASSERT(wchan != NULL, ("%s: invalid NULL wait channel", __func__));
913 MPASS((queue >= 0) && (queue < NR_SLEEPQS));
914 sq = sleepq_lookup(wchan);
917 KASSERT(sq->sq_type == (flags & SLEEPQ_TYPE),
918 ("%s: mismatch between sleep/wakeup and cv_*", __func__));
920 head = &sq->sq_blocked[queue];
921 if (flags & SLEEPQ_UNFAIR) {
923 * Find the most recently sleeping thread, but try to
924 * skip threads still in process of context switch to
925 * avoid spinning on the thread lock.
927 sc = SC_LOOKUP(wchan);
928 besttd = TAILQ_LAST_FAST(head, thread, td_slpq);
929 while (besttd->td_lock != &sc->sc_lock) {
930 td = TAILQ_PREV_FAST(besttd, head, thread, td_slpq);
937 * Find the highest priority thread on the queue. If there
938 * is a tie, use the thread that first appears in the queue
939 * as it has been sleeping the longest since threads are
940 * always added to the tail of sleep queues.
942 besttd = td = TAILQ_FIRST(head);
943 while ((td = TAILQ_NEXT(td, td_slpq)) != NULL) {
944 if (td->td_priority < besttd->td_priority)
948 MPASS(besttd != NULL);
950 wakeup_swapper = sleepq_resume_thread(sq, besttd, pri);
951 thread_unlock(besttd);
952 return (wakeup_swapper);
956 match_any(struct thread *td __unused)
963 * Resume all threads sleeping on a specified wait channel.
966 sleepq_broadcast(void *wchan, int flags, int pri, int queue)
968 struct sleepqueue *sq;
970 CTR2(KTR_PROC, "sleepq_broadcast(%p, %d)", wchan, flags);
971 KASSERT(wchan != NULL, ("%s: invalid NULL wait channel", __func__));
972 MPASS((queue >= 0) && (queue < NR_SLEEPQS));
973 sq = sleepq_lookup(wchan);
976 KASSERT(sq->sq_type == (flags & SLEEPQ_TYPE),
977 ("%s: mismatch between sleep/wakeup and cv_*", __func__));
979 return (sleepq_remove_matching(sq, queue, match_any, pri));
983 * Resume threads on the sleep queue that match the given predicate.
986 sleepq_remove_matching(struct sleepqueue *sq, int queue,
987 bool (*matches)(struct thread *), int pri)
989 struct thread *td, *tdn;
993 * The last thread will be given ownership of sq and may
994 * re-enqueue itself before sleepq_resume_thread() returns,
995 * so we must cache the "next" queue item at the beginning
996 * of the final iteration.
999 TAILQ_FOREACH_SAFE(td, &sq->sq_blocked[queue], td_slpq, tdn) {
1002 wakeup_swapper |= sleepq_resume_thread(sq, td, pri);
1006 return (wakeup_swapper);
1010 * Time sleeping threads out. When the timeout expires, the thread is
1011 * removed from the sleep queue and made runnable if it is still asleep.
1014 sleepq_timeout(void *arg)
1016 struct sleepqueue_chain *sc __unused;
1017 struct sleepqueue *sq;
1024 CTR3(KTR_PROC, "sleepq_timeout: thread %p (pid %ld, %s)",
1025 (void *)td, (long)td->td_proc->p_pid, (void *)td->td_name);
1029 if (td->td_sleeptimo > sbinuptime() || td->td_sleeptimo == 0) {
1031 * The thread does not want a timeout (yet).
1033 } else if (TD_IS_SLEEPING(td) && TD_ON_SLEEPQ(td)) {
1035 * See if the thread is asleep and get the wait
1038 wchan = td->td_wchan;
1039 sc = SC_LOOKUP(wchan);
1040 THREAD_LOCKPTR_ASSERT(td, &sc->sc_lock);
1041 sq = sleepq_lookup(wchan);
1043 td->td_flags |= TDF_TIMEOUT;
1044 wakeup_swapper = sleepq_resume_thread(sq, td, 0);
1045 } else if (TD_ON_SLEEPQ(td)) {
1047 * If the thread is on the SLEEPQ but isn't sleeping
1048 * yet, it can either be on another CPU in between
1049 * sleepq_add() and one of the sleepq_*wait*()
1050 * routines or it can be in sleepq_catch_signals().
1052 td->td_flags |= TDF_TIMEOUT;
1061 * Resumes a specific thread from the sleep queue associated with a specific
1062 * wait channel if it is on that queue.
1065 sleepq_remove(struct thread *td, void *wchan)
1067 struct sleepqueue *sq;
1071 * Look up the sleep queue for this wait channel, then re-check
1072 * that the thread is asleep on that channel, if it is not, then
1075 MPASS(wchan != NULL);
1077 sq = sleepq_lookup(wchan);
1079 * We can not lock the thread here as it may be sleeping on a
1080 * different sleepq. However, holding the sleepq lock for this
1081 * wchan can guarantee that we do not miss a wakeup for this
1082 * channel. The asserts below will catch any false positives.
1084 if (!TD_ON_SLEEPQ(td) || td->td_wchan != wchan) {
1085 sleepq_release(wchan);
1088 /* Thread is asleep on sleep queue sq, so wake it up. */
1091 MPASS(td->td_wchan == wchan);
1092 wakeup_swapper = sleepq_resume_thread(sq, td, 0);
1094 sleepq_release(wchan);
1100 * Abort a thread as if an interrupt had occurred. Only abort
1101 * interruptible waits (unfortunately it isn't safe to abort others).
1104 sleepq_abort(struct thread *td, int intrval)
1106 struct sleepqueue *sq;
1109 THREAD_LOCK_ASSERT(td, MA_OWNED);
1110 MPASS(TD_ON_SLEEPQ(td));
1111 MPASS(td->td_flags & TDF_SINTR);
1112 MPASS(intrval == EINTR || intrval == ERESTART);
1115 * If the TDF_TIMEOUT flag is set, just leave. A
1116 * timeout is scheduled anyhow.
1118 if (td->td_flags & TDF_TIMEOUT)
1121 CTR3(KTR_PROC, "sleepq_abort: thread %p (pid %ld, %s)",
1122 (void *)td, (long)td->td_proc->p_pid, (void *)td->td_name);
1123 td->td_intrval = intrval;
1124 td->td_flags |= TDF_SLEEPABORT;
1126 * If the thread has not slept yet it will find the signal in
1127 * sleepq_catch_signals() and call sleepq_resume_thread. Otherwise
1128 * we have to do it here.
1130 if (!TD_IS_SLEEPING(td))
1132 wchan = td->td_wchan;
1133 MPASS(wchan != NULL);
1134 sq = sleepq_lookup(wchan);
1137 /* Thread is asleep on sleep queue sq, so wake it up. */
1138 return (sleepq_resume_thread(sq, td, 0));
1142 sleepq_chains_remove_matching(bool (*matches)(struct thread *))
1144 struct sleepqueue_chain *sc;
1145 struct sleepqueue *sq, *sq1;
1146 int i, wakeup_swapper;
1149 for (sc = &sleepq_chains[0]; sc < sleepq_chains + SC_TABLESIZE; ++sc) {
1150 if (LIST_EMPTY(&sc->sc_queues)) {
1153 mtx_lock_spin(&sc->sc_lock);
1154 LIST_FOREACH_SAFE(sq, &sc->sc_queues, sq_hash, sq1) {
1155 for (i = 0; i < NR_SLEEPQS; ++i) {
1156 wakeup_swapper |= sleepq_remove_matching(sq, i,
1160 mtx_unlock_spin(&sc->sc_lock);
1162 if (wakeup_swapper) {
1168 * Prints the stacks of all threads presently sleeping on wchan/queue to
1169 * the sbuf sb. Sets count_stacks_printed to the number of stacks actually
1170 * printed. Typically, this will equal the number of threads sleeping on the
1171 * queue, but may be less if sb overflowed before all stacks were printed.
1175 sleepq_sbuf_print_stacks(struct sbuf *sb, void *wchan, int queue,
1176 int *count_stacks_printed)
1178 struct thread *td, *td_next;
1179 struct sleepqueue *sq;
1181 struct sbuf **td_infos;
1182 int i, stack_idx, error, stacks_to_allocate;
1188 KASSERT(wchan != NULL, ("%s: invalid NULL wait channel", __func__));
1189 MPASS((queue >= 0) && (queue < NR_SLEEPQS));
1191 stacks_to_allocate = 10;
1192 for (i = 0; i < 3 && !finished ; i++) {
1193 /* We cannot malloc while holding the queue's spinlock, so
1194 * we do our mallocs now, and hope it is enough. If it
1195 * isn't, we will free these, drop the lock, malloc more,
1196 * and try again, up to a point. After that point we will
1197 * give up and report ENOMEM. We also cannot write to sb
1198 * during this time since the client may have set the
1199 * SBUF_AUTOEXTEND flag on their sbuf, which could cause a
1200 * malloc as we print to it. So we defer actually printing
1201 * to sb until after we drop the spinlock.
1204 /* Where we will store the stacks. */
1205 st = malloc(sizeof(struct stack *) * stacks_to_allocate,
1207 for (stack_idx = 0; stack_idx < stacks_to_allocate;
1209 st[stack_idx] = stack_create(M_WAITOK);
1211 /* Where we will store the td name, tid, etc. */
1212 td_infos = malloc(sizeof(struct sbuf *) * stacks_to_allocate,
1214 for (stack_idx = 0; stack_idx < stacks_to_allocate;
1216 td_infos[stack_idx] = sbuf_new(NULL, NULL,
1217 MAXCOMLEN + sizeof(struct thread *) * 2 + 40,
1221 sq = sleepq_lookup(wchan);
1223 /* This sleepq does not exist; exit and return ENOENT. */
1226 sleepq_release(wchan);
1231 /* Save thread info */
1232 TAILQ_FOREACH_SAFE(td, &sq->sq_blocked[queue], td_slpq,
1234 if (stack_idx >= stacks_to_allocate)
1237 /* Note the td_lock is equal to the sleepq_lock here. */
1238 stack_save_td(st[stack_idx], td);
1240 sbuf_printf(td_infos[stack_idx], "%d: %s %p",
1241 td->td_tid, td->td_name, td);
1247 sleepq_release(wchan);
1249 /* Print the stacks */
1250 for (i = 0; i < stack_idx; i++) {
1251 sbuf_finish(td_infos[i]);
1252 sbuf_printf(sb, "--- thread %s: ---\n", sbuf_data(td_infos[i]));
1253 stack_sbuf_print(sb, st[i]);
1254 sbuf_printf(sb, "\n");
1256 error = sbuf_error(sb);
1258 *count_stacks_printed = stack_idx;
1263 sleepq_release(wchan);
1264 for (stack_idx = 0; stack_idx < stacks_to_allocate;
1266 stack_destroy(st[stack_idx]);
1267 for (stack_idx = 0; stack_idx < stacks_to_allocate;
1269 sbuf_delete(td_infos[stack_idx]);
1271 free(td_infos, M_TEMP);
1272 stacks_to_allocate *= 10;
1275 if (!finished && error == 0)
1282 #ifdef SLEEPQUEUE_PROFILING
1283 #define SLEEPQ_PROF_LOCATIONS 1024
1284 #define SLEEPQ_SBUFSIZE 512
1285 struct sleepq_prof {
1286 LIST_ENTRY(sleepq_prof) sp_link;
1287 const char *sp_wmesg;
1291 LIST_HEAD(sqphead, sleepq_prof);
1293 struct sqphead sleepq_prof_free;
1294 struct sqphead sleepq_hash[SC_TABLESIZE];
1295 static struct sleepq_prof sleepq_profent[SLEEPQ_PROF_LOCATIONS];
1296 static struct mtx sleepq_prof_lock;
1297 MTX_SYSINIT(sleepq_prof_lock, &sleepq_prof_lock, "sleepq_prof", MTX_SPIN);
1300 sleepq_profile(const char *wmesg)
1302 struct sleepq_prof *sp;
1304 mtx_lock_spin(&sleepq_prof_lock);
1305 if (prof_enabled == 0)
1307 LIST_FOREACH(sp, &sleepq_hash[SC_HASH(wmesg)], sp_link)
1308 if (sp->sp_wmesg == wmesg)
1310 sp = LIST_FIRST(&sleepq_prof_free);
1313 sp->sp_wmesg = wmesg;
1314 LIST_REMOVE(sp, sp_link);
1315 LIST_INSERT_HEAD(&sleepq_hash[SC_HASH(wmesg)], sp, sp_link);
1319 mtx_unlock_spin(&sleepq_prof_lock);
1324 sleepq_prof_reset(void)
1326 struct sleepq_prof *sp;
1330 mtx_lock_spin(&sleepq_prof_lock);
1331 enabled = prof_enabled;
1333 for (i = 0; i < SC_TABLESIZE; i++)
1334 LIST_INIT(&sleepq_hash[i]);
1335 LIST_INIT(&sleepq_prof_free);
1336 for (i = 0; i < SLEEPQ_PROF_LOCATIONS; i++) {
1337 sp = &sleepq_profent[i];
1338 sp->sp_wmesg = NULL;
1340 LIST_INSERT_HEAD(&sleepq_prof_free, sp, sp_link);
1342 prof_enabled = enabled;
1343 mtx_unlock_spin(&sleepq_prof_lock);
1347 enable_sleepq_prof(SYSCTL_HANDLER_ARGS)
1352 error = sysctl_handle_int(oidp, &v, v, req);
1355 if (req->newptr == NULL)
1357 if (v == prof_enabled)
1360 sleepq_prof_reset();
1361 mtx_lock_spin(&sleepq_prof_lock);
1363 mtx_unlock_spin(&sleepq_prof_lock);
1369 reset_sleepq_prof_stats(SYSCTL_HANDLER_ARGS)
1374 error = sysctl_handle_int(oidp, &v, 0, req);
1377 if (req->newptr == NULL)
1381 sleepq_prof_reset();
1387 dump_sleepq_prof_stats(SYSCTL_HANDLER_ARGS)
1389 struct sleepq_prof *sp;
1395 error = sysctl_wire_old_buffer(req, 0);
1398 sb = sbuf_new_for_sysctl(NULL, NULL, SLEEPQ_SBUFSIZE, req);
1399 sbuf_printf(sb, "\nwmesg\tcount\n");
1400 enabled = prof_enabled;
1401 mtx_lock_spin(&sleepq_prof_lock);
1403 mtx_unlock_spin(&sleepq_prof_lock);
1404 for (i = 0; i < SC_TABLESIZE; i++) {
1405 LIST_FOREACH(sp, &sleepq_hash[i], sp_link) {
1406 sbuf_printf(sb, "%s\t%ld\n",
1407 sp->sp_wmesg, sp->sp_count);
1410 mtx_lock_spin(&sleepq_prof_lock);
1411 prof_enabled = enabled;
1412 mtx_unlock_spin(&sleepq_prof_lock);
1414 error = sbuf_finish(sb);
1419 SYSCTL_PROC(_debug_sleepq, OID_AUTO, stats, CTLTYPE_STRING | CTLFLAG_RD,
1420 NULL, 0, dump_sleepq_prof_stats, "A", "Sleepqueue profiling statistics");
1421 SYSCTL_PROC(_debug_sleepq, OID_AUTO, reset, CTLTYPE_INT | CTLFLAG_RW,
1422 NULL, 0, reset_sleepq_prof_stats, "I",
1423 "Reset sleepqueue profiling statistics");
1424 SYSCTL_PROC(_debug_sleepq, OID_AUTO, enable, CTLTYPE_INT | CTLFLAG_RW,
1425 NULL, 0, enable_sleepq_prof, "I", "Enable sleepqueue profiling");
1429 DB_SHOW_COMMAND(sleepq, db_show_sleepqueue)
1431 struct sleepqueue_chain *sc;
1432 struct sleepqueue *sq;
1434 struct lock_object *lock;
1444 * First, see if there is an active sleep queue for the wait channel
1445 * indicated by the address.
1447 wchan = (void *)addr;
1448 sc = SC_LOOKUP(wchan);
1449 LIST_FOREACH(sq, &sc->sc_queues, sq_hash)
1450 if (sq->sq_wchan == wchan)
1454 * Second, see if there is an active sleep queue at the address
1457 for (i = 0; i < SC_TABLESIZE; i++)
1458 LIST_FOREACH(sq, &sleepq_chains[i].sc_queues, sq_hash) {
1459 if (sq == (struct sleepqueue *)addr)
1463 db_printf("Unable to locate a sleep queue via %p\n", (void *)addr);
1466 db_printf("Wait channel: %p\n", sq->sq_wchan);
1467 db_printf("Queue type: %d\n", sq->sq_type);
1471 db_printf("Associated Interlock: %p - (%s) %s\n", lock,
1472 LOCK_CLASS(lock)->lc_name, lock->lo_name);
1475 db_printf("Blocked threads:\n");
1476 for (i = 0; i < NR_SLEEPQS; i++) {
1477 db_printf("\nQueue[%d]:\n", i);
1478 if (TAILQ_EMPTY(&sq->sq_blocked[i]))
1479 db_printf("\tempty\n");
1481 TAILQ_FOREACH(td, &sq->sq_blocked[i],
1483 db_printf("\t%p (tid %d, pid %d, \"%s\")\n", td,
1484 td->td_tid, td->td_proc->p_pid,
1487 db_printf("(expected: %u)\n", sq->sq_blockedcnt[i]);
1491 /* Alias 'show sleepqueue' to 'show sleepq'. */
1492 DB_SHOW_ALIAS(sleepqueue, db_show_sleepqueue);