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 <machine/atomic.h>
94 * Constants for the hash table of sleep queue chains.
95 * SC_TABLESIZE must be a power of two for SC_MASK to work properly.
98 #define SC_TABLESIZE 256
100 CTASSERT(powerof2(SC_TABLESIZE));
101 #define SC_MASK (SC_TABLESIZE - 1)
103 #define SC_HASH(wc) ((((uintptr_t)(wc) >> SC_SHIFT) ^ (uintptr_t)(wc)) & \
105 #define SC_LOOKUP(wc) &sleepq_chains[SC_HASH(wc)]
108 * There are two different lists of sleep queues. Both lists are connected
109 * via the sq_hash entries. The first list is the sleep queue chain list
110 * that a sleep queue is on when it is attached to a wait channel. The
111 * second list is the free list hung off of a sleep queue that is attached
114 * Each sleep queue also contains the wait channel it is attached to, the
115 * list of threads blocked on that wait channel, flags specific to the
116 * wait channel, and the lock used to synchronize with a wait channel.
117 * The flags are used to catch mismatches between the various consumers
118 * of the sleep queue API (e.g. sleep/wakeup and condition variables).
119 * The lock pointer is only used when invariants are enabled for various
123 * c - sleep queue chain lock
126 struct threadqueue sq_blocked[NR_SLEEPQS]; /* (c) Blocked threads. */
127 u_int sq_blockedcnt[NR_SLEEPQS]; /* (c) N. of blocked threads. */
128 LIST_ENTRY(sleepqueue) sq_hash; /* (c) Chain and free list. */
129 LIST_HEAD(, sleepqueue) sq_free; /* (c) Free queues. */
130 void *sq_wchan; /* (c) Wait channel. */
131 int sq_type; /* (c) Queue type. */
133 struct lock_object *sq_lock; /* (c) Associated lock. */
137 struct sleepqueue_chain {
138 LIST_HEAD(, sleepqueue) sc_queues; /* List of sleep queues. */
139 struct mtx sc_lock; /* Spin lock for this chain. */
140 #ifdef SLEEPQUEUE_PROFILING
141 u_int sc_depth; /* Length of sc_queues. */
142 u_int sc_max_depth; /* Max length of sc_queues. */
144 } __aligned(CACHE_LINE_SIZE);
146 #ifdef SLEEPQUEUE_PROFILING
147 u_int sleepq_max_depth;
148 static SYSCTL_NODE(_debug, OID_AUTO, sleepq, CTLFLAG_RD, 0, "sleepq profiling");
149 static SYSCTL_NODE(_debug_sleepq, OID_AUTO, chains, CTLFLAG_RD, 0,
150 "sleepq chain stats");
151 SYSCTL_UINT(_debug_sleepq, OID_AUTO, max_depth, CTLFLAG_RD, &sleepq_max_depth,
152 0, "maxmimum depth achieved of a single chain");
154 static void sleepq_profile(const char *wmesg);
155 static int prof_enabled;
157 static struct sleepqueue_chain sleepq_chains[SC_TABLESIZE];
158 static uma_zone_t sleepq_zone;
161 * Prototypes for non-exported routines.
163 static int sleepq_catch_signals(void *wchan, int pri);
164 static int sleepq_check_signals(void);
165 static int sleepq_check_timeout(void);
167 static void sleepq_dtor(void *mem, int size, void *arg);
169 static int sleepq_init(void *mem, int size, int flags);
170 static int sleepq_resume_thread(struct sleepqueue *sq, struct thread *td,
172 static void sleepq_switch(void *wchan, int pri);
173 static void sleepq_timeout(void *arg);
175 SDT_PROBE_DECLARE(sched, , , sleep);
176 SDT_PROBE_DECLARE(sched, , , wakeup);
179 * Initialize SLEEPQUEUE_PROFILING specific sysctl nodes.
180 * Note that it must happen after sleepinit() has been fully executed, so
181 * it must happen after SI_SUB_KMEM SYSINIT() subsystem setup.
183 #ifdef SLEEPQUEUE_PROFILING
185 init_sleepqueue_profiling(void)
188 struct sysctl_oid *chain_oid;
191 for (i = 0; i < SC_TABLESIZE; i++) {
192 snprintf(chain_name, sizeof(chain_name), "%u", i);
193 chain_oid = SYSCTL_ADD_NODE(NULL,
194 SYSCTL_STATIC_CHILDREN(_debug_sleepq_chains), OID_AUTO,
195 chain_name, CTLFLAG_RD, NULL, "sleepq chain stats");
196 SYSCTL_ADD_UINT(NULL, SYSCTL_CHILDREN(chain_oid), OID_AUTO,
197 "depth", CTLFLAG_RD, &sleepq_chains[i].sc_depth, 0, NULL);
198 SYSCTL_ADD_UINT(NULL, SYSCTL_CHILDREN(chain_oid), OID_AUTO,
199 "max_depth", CTLFLAG_RD, &sleepq_chains[i].sc_max_depth, 0,
204 SYSINIT(sleepqueue_profiling, SI_SUB_LOCK, SI_ORDER_ANY,
205 init_sleepqueue_profiling, NULL);
209 * Early initialization of sleep queues that is called from the sleepinit()
213 init_sleepqueues(void)
217 for (i = 0; i < SC_TABLESIZE; i++) {
218 LIST_INIT(&sleepq_chains[i].sc_queues);
219 mtx_init(&sleepq_chains[i].sc_lock, "sleepq chain", NULL,
220 MTX_SPIN | MTX_RECURSE);
222 sleepq_zone = uma_zcreate("SLEEPQUEUE", sizeof(struct sleepqueue),
224 NULL, sleepq_dtor, sleepq_init, NULL, UMA_ALIGN_CACHE, 0);
226 NULL, NULL, sleepq_init, NULL, UMA_ALIGN_CACHE, 0);
229 thread0.td_sleepqueue = sleepq_alloc();
233 * Get a sleep queue for a new thread.
239 return (uma_zalloc(sleepq_zone, M_WAITOK));
243 * Free a sleep queue when a thread is destroyed.
246 sleepq_free(struct sleepqueue *sq)
249 uma_zfree(sleepq_zone, sq);
253 * Lock the sleep queue chain associated with the specified wait channel.
256 sleepq_lock(void *wchan)
258 struct sleepqueue_chain *sc;
260 sc = SC_LOOKUP(wchan);
261 mtx_lock_spin(&sc->sc_lock);
265 * Look up the sleep queue associated with a given wait channel in the hash
266 * table locking the associated sleep queue chain. If no queue is found in
267 * the table, NULL is returned.
270 sleepq_lookup(void *wchan)
272 struct sleepqueue_chain *sc;
273 struct sleepqueue *sq;
275 KASSERT(wchan != NULL, ("%s: invalid NULL wait channel", __func__));
276 sc = SC_LOOKUP(wchan);
277 mtx_assert(&sc->sc_lock, MA_OWNED);
278 LIST_FOREACH(sq, &sc->sc_queues, sq_hash)
279 if (sq->sq_wchan == wchan)
285 * Unlock the sleep queue chain associated with a given wait channel.
288 sleepq_release(void *wchan)
290 struct sleepqueue_chain *sc;
292 sc = SC_LOOKUP(wchan);
293 mtx_unlock_spin(&sc->sc_lock);
297 * Places the current thread on the sleep queue for the specified wait
298 * channel. If INVARIANTS is enabled, then it associates the passed in
299 * lock with the sleepq to make sure it is held when that sleep queue is
303 sleepq_add(void *wchan, struct lock_object *lock, const char *wmesg, int flags,
306 struct sleepqueue_chain *sc;
307 struct sleepqueue *sq;
311 sc = SC_LOOKUP(wchan);
312 mtx_assert(&sc->sc_lock, MA_OWNED);
313 MPASS(td->td_sleepqueue != NULL);
314 MPASS(wchan != NULL);
315 MPASS((queue >= 0) && (queue < NR_SLEEPQS));
317 /* If this thread is not allowed to sleep, die a horrible death. */
318 KASSERT(td->td_no_sleeping == 0,
319 ("%s: td %p to sleep on wchan %p with sleeping prohibited",
320 __func__, td, wchan));
322 /* Look up the sleep queue associated with the wait channel 'wchan'. */
323 sq = sleepq_lookup(wchan);
326 * If the wait channel does not already have a sleep queue, use
327 * this thread's sleep queue. Otherwise, insert the current thread
328 * into the sleep queue already in use by this wait channel.
334 sq = td->td_sleepqueue;
335 for (i = 0; i < NR_SLEEPQS; i++) {
336 KASSERT(TAILQ_EMPTY(&sq->sq_blocked[i]),
337 ("thread's sleep queue %d is not empty", i));
338 KASSERT(sq->sq_blockedcnt[i] == 0,
339 ("thread's sleep queue %d count mismatches", i));
341 KASSERT(LIST_EMPTY(&sq->sq_free),
342 ("thread's sleep queue has a non-empty free list"));
343 KASSERT(sq->sq_wchan == NULL, ("stale sq_wchan pointer"));
346 #ifdef SLEEPQUEUE_PROFILING
348 if (sc->sc_depth > sc->sc_max_depth) {
349 sc->sc_max_depth = sc->sc_depth;
350 if (sc->sc_max_depth > sleepq_max_depth)
351 sleepq_max_depth = sc->sc_max_depth;
354 sq = td->td_sleepqueue;
355 LIST_INSERT_HEAD(&sc->sc_queues, sq, sq_hash);
356 sq->sq_wchan = wchan;
357 sq->sq_type = flags & SLEEPQ_TYPE;
359 MPASS(wchan == sq->sq_wchan);
360 MPASS(lock == sq->sq_lock);
361 MPASS((flags & SLEEPQ_TYPE) == sq->sq_type);
362 LIST_INSERT_HEAD(&sq->sq_free, td->td_sleepqueue, sq_hash);
365 TAILQ_INSERT_TAIL(&sq->sq_blocked[queue], td, td_slpq);
366 sq->sq_blockedcnt[queue]++;
367 td->td_sleepqueue = NULL;
368 td->td_sqqueue = queue;
369 td->td_wchan = wchan;
370 td->td_wmesg = wmesg;
371 if (flags & SLEEPQ_INTERRUPTIBLE) {
372 td->td_flags |= TDF_SINTR;
373 td->td_flags &= ~TDF_SLEEPABORT;
379 * Sets a timeout that will remove the current thread from the specified
380 * sleep queue after timo ticks if the thread has not already been awakened.
383 sleepq_set_timeout_sbt(void *wchan, sbintime_t sbt, sbintime_t pr,
386 struct sleepqueue_chain *sc __unused;
391 sc = SC_LOOKUP(wchan);
392 mtx_assert(&sc->sc_lock, MA_OWNED);
393 MPASS(TD_ON_SLEEPQ(td));
394 MPASS(td->td_sleepqueue == NULL);
395 MPASS(wchan != NULL);
396 if (cold && td == &thread0)
397 panic("timed sleep before timers are working");
398 KASSERT(td->td_sleeptimo == 0, ("td %d %p td_sleeptimo %jx",
399 td->td_tid, td, (uintmax_t)td->td_sleeptimo));
401 callout_when(sbt, pr, flags, &td->td_sleeptimo, &pr1);
403 callout_reset_sbt_on(&td->td_slpcallout, td->td_sleeptimo, pr1,
404 sleepq_timeout, td, PCPU_GET(cpuid), flags | C_PRECALC |
409 * Return the number of actual sleepers for the specified queue.
412 sleepq_sleepcnt(void *wchan, int queue)
414 struct sleepqueue *sq;
416 KASSERT(wchan != NULL, ("%s: invalid NULL wait channel", __func__));
417 MPASS((queue >= 0) && (queue < NR_SLEEPQS));
418 sq = sleepq_lookup(wchan);
421 return (sq->sq_blockedcnt[queue]);
425 * Marks the pending sleep of the current thread as interruptible and
426 * makes an initial check for pending signals before putting a thread
427 * to sleep. Enters and exits with the thread lock held. Thread lock
428 * may have transitioned from the sleepq lock to a run lock.
431 sleepq_catch_signals(void *wchan, int pri)
433 struct sleepqueue_chain *sc;
434 struct sleepqueue *sq;
443 sc = SC_LOOKUP(wchan);
444 mtx_assert(&sc->sc_lock, MA_OWNED);
445 MPASS(wchan != NULL);
446 if ((td->td_pflags & TDP_WAKEUP) != 0) {
447 td->td_pflags &= ~TDP_WAKEUP;
454 * See if there are any pending signals or suspension requests for this
455 * thread. If not, we can switch immediately.
458 if ((td->td_flags & (TDF_NEEDSIGCHK | TDF_NEEDSUSPCHK)) != 0) {
460 mtx_unlock_spin(&sc->sc_lock);
461 CTR3(KTR_PROC, "sleepq catching signals: thread %p (pid %ld, %s)",
462 (void *)td, (long)p->p_pid, td->td_name);
465 * Check for suspension first. Checking for signals and then
466 * suspending could result in a missed signal, since a signal
467 * can be delivered while this thread is suspended.
469 if ((td->td_flags & TDF_NEEDSUSPCHK) != 0) {
470 ret = thread_suspend_check(1);
471 MPASS(ret == 0 || ret == EINTR || ret == ERESTART);
474 mtx_lock_spin(&sc->sc_lock);
479 if ((td->td_flags & TDF_NEEDSIGCHK) != 0) {
481 mtx_lock(&ps->ps_mtx);
484 mtx_unlock(&ps->ps_mtx);
485 KASSERT((td->td_flags & TDF_SBDRY) != 0,
487 KASSERT(TD_SBDRY_INTR(td),
488 ("lost TDF_SERESTART of TDF_SEINTR"));
489 KASSERT((td->td_flags &
490 (TDF_SEINTR | TDF_SERESTART)) !=
491 (TDF_SEINTR | TDF_SERESTART),
492 ("both TDF_SEINTR and TDF_SERESTART"));
493 ret = TD_SBDRY_ERRNO(td);
494 } else if (sig != 0) {
495 ret = SIGISMEMBER(ps->ps_sigintr, sig) ?
497 mtx_unlock(&ps->ps_mtx);
499 mtx_unlock(&ps->ps_mtx);
503 * Do not go into sleep if this thread was the
504 * ptrace(2) attach leader. cursig() consumed
505 * SIGSTOP from PT_ATTACH, but we usually act
506 * on the signal by interrupting sleep, and
507 * should do that here as well.
509 if ((td->td_dbgflags & TDB_FSTP) != 0) {
512 td->td_dbgflags &= ~TDB_FSTP;
516 * Lock the per-process spinlock prior to dropping the PROC_LOCK
517 * to avoid a signal delivery race. PROC_LOCK, PROC_SLOCK, and
518 * thread_lock() are currently held in tdsendsignal().
521 mtx_lock_spin(&sc->sc_lock);
527 sleepq_switch(wchan, pri);
532 * There were pending signals and this thread is still
533 * on the sleep queue, remove it from the sleep queue.
535 if (TD_ON_SLEEPQ(td)) {
536 sq = sleepq_lookup(wchan);
537 if (sleepq_resume_thread(sq, td, 0)) {
540 * This thread hasn't gone to sleep yet, so it
541 * should not be swapped out.
543 panic("not waking up swapper");
547 mtx_unlock_spin(&sc->sc_lock);
548 MPASS(td->td_lock != &sc->sc_lock);
553 * Switches to another thread if we are still asleep on a sleep queue.
554 * Returns with thread lock.
557 sleepq_switch(void *wchan, int pri)
559 struct sleepqueue_chain *sc;
560 struct sleepqueue *sq;
565 sc = SC_LOOKUP(wchan);
566 mtx_assert(&sc->sc_lock, MA_OWNED);
567 THREAD_LOCK_ASSERT(td, MA_OWNED);
570 * If we have a sleep queue, then we've already been woken up, so
573 if (td->td_sleepqueue != NULL) {
574 mtx_unlock_spin(&sc->sc_lock);
579 * If TDF_TIMEOUT is set, then our sleep has been timed out
580 * already but we are still on the sleep queue, so dequeue the
583 * Do the same if the real-time clock has been adjusted since this
584 * thread calculated its timeout based on that clock. This handles
585 * the following race:
586 * - The Ts thread needs to sleep until an absolute real-clock time.
587 * It copies the global rtc_generation into curthread->td_rtcgen,
588 * reads the RTC, and calculates a sleep duration based on that time.
589 * See umtxq_sleep() for an example.
590 * - The Tc thread adjusts the RTC, bumps rtc_generation, and wakes
591 * threads that are sleeping until an absolute real-clock time.
592 * See tc_setclock() and the POSIX specification of clock_settime().
593 * - Ts reaches the code below. It holds the sleepqueue chain lock,
594 * so Tc has finished waking, so this thread must test td_rtcgen.
595 * (The declaration of td_rtcgen refers to this comment.)
597 rtc_changed = td->td_rtcgen != 0 && td->td_rtcgen != rtc_generation;
598 if ((td->td_flags & TDF_TIMEOUT) || rtc_changed) {
602 MPASS(TD_ON_SLEEPQ(td));
603 sq = sleepq_lookup(wchan);
604 if (sleepq_resume_thread(sq, td, 0)) {
607 * This thread hasn't gone to sleep yet, so it
608 * should not be swapped out.
610 panic("not waking up swapper");
613 mtx_unlock_spin(&sc->sc_lock);
616 #ifdef SLEEPQUEUE_PROFILING
618 sleepq_profile(td->td_wmesg);
620 MPASS(td->td_sleepqueue == NULL);
621 sched_sleep(td, pri);
622 thread_lock_set(td, &sc->sc_lock);
623 SDT_PROBE0(sched, , , sleep);
625 mi_switch(SW_VOL | SWT_SLEEPQ, NULL);
626 KASSERT(TD_IS_RUNNING(td), ("running but not TDS_RUNNING"));
627 CTR3(KTR_PROC, "sleepq resume: thread %p (pid %ld, %s)",
628 (void *)td, (long)td->td_proc->p_pid, (void *)td->td_name);
632 * Check to see if we timed out.
635 sleepq_check_timeout(void)
641 THREAD_LOCK_ASSERT(td, MA_OWNED);
644 * If TDF_TIMEOUT is set, we timed out. But recheck
645 * td_sleeptimo anyway.
648 if (td->td_sleeptimo != 0) {
649 if (td->td_sleeptimo <= sbinuptime())
651 td->td_sleeptimo = 0;
653 if (td->td_flags & TDF_TIMEOUT)
654 td->td_flags &= ~TDF_TIMEOUT;
657 * We ignore the situation where timeout subsystem was
658 * unable to stop our callout. The struct thread is
659 * type-stable, the callout will use the correct
660 * memory when running. The checks of the
661 * td_sleeptimo value in this function and in
662 * sleepq_timeout() ensure that the thread does not
663 * get spurious wakeups, even if the callout was reset
666 callout_stop(&td->td_slpcallout);
671 * Check to see if we were awoken by a signal.
674 sleepq_check_signals(void)
679 THREAD_LOCK_ASSERT(td, MA_OWNED);
681 /* We are no longer in an interruptible sleep. */
682 if (td->td_flags & TDF_SINTR)
683 td->td_flags &= ~TDF_SINTR;
685 if (td->td_flags & TDF_SLEEPABORT) {
686 td->td_flags &= ~TDF_SLEEPABORT;
687 return (td->td_intrval);
694 * Block the current thread until it is awakened from its sleep queue.
697 sleepq_wait(void *wchan, int pri)
702 MPASS(!(td->td_flags & TDF_SINTR));
704 sleepq_switch(wchan, pri);
709 * Block the current thread until it is awakened from its sleep queue
710 * or it is interrupted by a signal.
713 sleepq_wait_sig(void *wchan, int pri)
718 rcatch = sleepq_catch_signals(wchan, pri);
719 rval = sleepq_check_signals();
720 thread_unlock(curthread);
727 * Block the current thread until it is awakened from its sleep queue
728 * or it times out while waiting.
731 sleepq_timedwait(void *wchan, int pri)
737 MPASS(!(td->td_flags & TDF_SINTR));
739 sleepq_switch(wchan, pri);
740 rval = sleepq_check_timeout();
747 * Block the current thread until it is awakened from its sleep queue,
748 * it is interrupted by a signal, or it times out waiting to be awakened.
751 sleepq_timedwait_sig(void *wchan, int pri)
753 int rcatch, rvalt, rvals;
755 rcatch = sleepq_catch_signals(wchan, pri);
756 rvalt = sleepq_check_timeout();
757 rvals = sleepq_check_signals();
758 thread_unlock(curthread);
767 * Returns the type of sleepqueue given a waitchannel.
770 sleepq_type(void *wchan)
772 struct sleepqueue *sq;
775 MPASS(wchan != NULL);
778 sq = sleepq_lookup(wchan);
780 sleepq_release(wchan);
784 sleepq_release(wchan);
789 * Removes a thread from a sleep queue and makes it
793 sleepq_resume_thread(struct sleepqueue *sq, struct thread *td, int pri)
795 struct sleepqueue_chain *sc __unused;
798 MPASS(sq->sq_wchan != NULL);
799 MPASS(td->td_wchan == sq->sq_wchan);
800 MPASS(td->td_sqqueue < NR_SLEEPQS && td->td_sqqueue >= 0);
801 THREAD_LOCK_ASSERT(td, MA_OWNED);
802 sc = SC_LOOKUP(sq->sq_wchan);
803 mtx_assert(&sc->sc_lock, MA_OWNED);
805 SDT_PROBE2(sched, , , wakeup, td, td->td_proc);
807 /* Remove the thread from the queue. */
808 sq->sq_blockedcnt[td->td_sqqueue]--;
809 TAILQ_REMOVE(&sq->sq_blocked[td->td_sqqueue], td, td_slpq);
812 * Get a sleep queue for this thread. If this is the last waiter,
813 * use the queue itself and take it out of the chain, otherwise,
814 * remove a queue from the free list.
816 if (LIST_EMPTY(&sq->sq_free)) {
817 td->td_sleepqueue = sq;
821 #ifdef SLEEPQUEUE_PROFILING
825 td->td_sleepqueue = LIST_FIRST(&sq->sq_free);
826 LIST_REMOVE(td->td_sleepqueue, sq_hash);
830 td->td_flags &= ~TDF_SINTR;
832 CTR3(KTR_PROC, "sleepq_wakeup: thread %p (pid %ld, %s)",
833 (void *)td, (long)td->td_proc->p_pid, td->td_name);
835 /* Adjust priority if requested. */
836 MPASS(pri == 0 || (pri >= PRI_MIN && pri <= PRI_MAX));
837 if (pri != 0 && td->td_priority > pri &&
838 PRI_BASE(td->td_pri_class) == PRI_TIMESHARE)
842 * Note that thread td might not be sleeping if it is running
843 * sleepq_catch_signals() on another CPU or is blocked on its
844 * proc lock to check signals. There's no need to mark the
845 * thread runnable in that case.
847 if (TD_IS_SLEEPING(td)) {
849 return (setrunnable(td));
856 * UMA zone item deallocator.
859 sleepq_dtor(void *mem, int size, void *arg)
861 struct sleepqueue *sq;
865 for (i = 0; i < NR_SLEEPQS; i++) {
866 MPASS(TAILQ_EMPTY(&sq->sq_blocked[i]));
867 MPASS(sq->sq_blockedcnt[i] == 0);
873 * UMA zone item initializer.
876 sleepq_init(void *mem, int size, int flags)
878 struct sleepqueue *sq;
883 for (i = 0; i < NR_SLEEPQS; i++) {
884 TAILQ_INIT(&sq->sq_blocked[i]);
885 sq->sq_blockedcnt[i] = 0;
887 LIST_INIT(&sq->sq_free);
892 * Find thread sleeping on a wait channel and resume it.
895 sleepq_signal(void *wchan, int flags, int pri, int queue)
897 struct sleepqueue_chain *sc;
898 struct sleepqueue *sq;
899 struct threadqueue *head;
900 struct thread *td, *besttd;
903 CTR2(KTR_PROC, "sleepq_signal(%p, %d)", wchan, flags);
904 KASSERT(wchan != NULL, ("%s: invalid NULL wait channel", __func__));
905 MPASS((queue >= 0) && (queue < NR_SLEEPQS));
906 sq = sleepq_lookup(wchan);
909 KASSERT(sq->sq_type == (flags & SLEEPQ_TYPE),
910 ("%s: mismatch between sleep/wakeup and cv_*", __func__));
912 head = &sq->sq_blocked[queue];
913 if (flags & SLEEPQ_UNFAIR) {
915 * Find the most recently sleeping thread, but try to
916 * skip threads still in process of context switch to
917 * avoid spinning on the thread lock.
919 sc = SC_LOOKUP(wchan);
920 besttd = TAILQ_LAST_FAST(head, thread, td_slpq);
921 while (besttd->td_lock != &sc->sc_lock) {
922 td = TAILQ_PREV_FAST(besttd, head, thread, td_slpq);
929 * Find the highest priority thread on the queue. If there
930 * is a tie, use the thread that first appears in the queue
931 * as it has been sleeping the longest since threads are
932 * always added to the tail of sleep queues.
934 besttd = td = TAILQ_FIRST(head);
935 while ((td = TAILQ_NEXT(td, td_slpq)) != NULL) {
936 if (td->td_priority < besttd->td_priority)
940 MPASS(besttd != NULL);
942 wakeup_swapper = sleepq_resume_thread(sq, besttd, pri);
943 thread_unlock(besttd);
944 return (wakeup_swapper);
948 match_any(struct thread *td __unused)
955 * Resume all threads sleeping on a specified wait channel.
958 sleepq_broadcast(void *wchan, int flags, int pri, int queue)
960 struct sleepqueue *sq;
962 CTR2(KTR_PROC, "sleepq_broadcast(%p, %d)", wchan, flags);
963 KASSERT(wchan != NULL, ("%s: invalid NULL wait channel", __func__));
964 MPASS((queue >= 0) && (queue < NR_SLEEPQS));
965 sq = sleepq_lookup(wchan);
968 KASSERT(sq->sq_type == (flags & SLEEPQ_TYPE),
969 ("%s: mismatch between sleep/wakeup and cv_*", __func__));
971 return (sleepq_remove_matching(sq, queue, match_any, pri));
975 * Resume threads on the sleep queue that match the given predicate.
978 sleepq_remove_matching(struct sleepqueue *sq, int queue,
979 bool (*matches)(struct thread *), int pri)
981 struct thread *td, *tdn;
985 * The last thread will be given ownership of sq and may
986 * re-enqueue itself before sleepq_resume_thread() returns,
987 * so we must cache the "next" queue item at the beginning
988 * of the final iteration.
991 TAILQ_FOREACH_SAFE(td, &sq->sq_blocked[queue], td_slpq, tdn) {
994 wakeup_swapper |= sleepq_resume_thread(sq, td, pri);
998 return (wakeup_swapper);
1002 * Time sleeping threads out. When the timeout expires, the thread is
1003 * removed from the sleep queue and made runnable if it is still asleep.
1006 sleepq_timeout(void *arg)
1008 struct sleepqueue_chain *sc __unused;
1009 struct sleepqueue *sq;
1016 CTR3(KTR_PROC, "sleepq_timeout: thread %p (pid %ld, %s)",
1017 (void *)td, (long)td->td_proc->p_pid, (void *)td->td_name);
1021 if (td->td_sleeptimo > sbinuptime() || td->td_sleeptimo == 0) {
1023 * The thread does not want a timeout (yet).
1025 } else if (TD_IS_SLEEPING(td) && TD_ON_SLEEPQ(td)) {
1027 * See if the thread is asleep and get the wait
1030 wchan = td->td_wchan;
1031 sc = SC_LOOKUP(wchan);
1032 THREAD_LOCKPTR_ASSERT(td, &sc->sc_lock);
1033 sq = sleepq_lookup(wchan);
1035 td->td_flags |= TDF_TIMEOUT;
1036 wakeup_swapper = sleepq_resume_thread(sq, td, 0);
1037 } else if (TD_ON_SLEEPQ(td)) {
1039 * If the thread is on the SLEEPQ but isn't sleeping
1040 * yet, it can either be on another CPU in between
1041 * sleepq_add() and one of the sleepq_*wait*()
1042 * routines or it can be in sleepq_catch_signals().
1044 td->td_flags |= TDF_TIMEOUT;
1053 * Resumes a specific thread from the sleep queue associated with a specific
1054 * wait channel if it is on that queue.
1057 sleepq_remove(struct thread *td, void *wchan)
1059 struct sleepqueue *sq;
1063 * Look up the sleep queue for this wait channel, then re-check
1064 * that the thread is asleep on that channel, if it is not, then
1067 MPASS(wchan != NULL);
1069 sq = sleepq_lookup(wchan);
1071 * We can not lock the thread here as it may be sleeping on a
1072 * different sleepq. However, holding the sleepq lock for this
1073 * wchan can guarantee that we do not miss a wakeup for this
1074 * channel. The asserts below will catch any false positives.
1076 if (!TD_ON_SLEEPQ(td) || td->td_wchan != wchan) {
1077 sleepq_release(wchan);
1080 /* Thread is asleep on sleep queue sq, so wake it up. */
1083 MPASS(td->td_wchan == wchan);
1084 wakeup_swapper = sleepq_resume_thread(sq, td, 0);
1086 sleepq_release(wchan);
1092 * Abort a thread as if an interrupt had occurred. Only abort
1093 * interruptible waits (unfortunately it isn't safe to abort others).
1096 sleepq_abort(struct thread *td, int intrval)
1098 struct sleepqueue *sq;
1101 THREAD_LOCK_ASSERT(td, MA_OWNED);
1102 MPASS(TD_ON_SLEEPQ(td));
1103 MPASS(td->td_flags & TDF_SINTR);
1104 MPASS(intrval == EINTR || intrval == ERESTART);
1107 * If the TDF_TIMEOUT flag is set, just leave. A
1108 * timeout is scheduled anyhow.
1110 if (td->td_flags & TDF_TIMEOUT)
1113 CTR3(KTR_PROC, "sleepq_abort: thread %p (pid %ld, %s)",
1114 (void *)td, (long)td->td_proc->p_pid, (void *)td->td_name);
1115 td->td_intrval = intrval;
1116 td->td_flags |= TDF_SLEEPABORT;
1118 * If the thread has not slept yet it will find the signal in
1119 * sleepq_catch_signals() and call sleepq_resume_thread. Otherwise
1120 * we have to do it here.
1122 if (!TD_IS_SLEEPING(td))
1124 wchan = td->td_wchan;
1125 MPASS(wchan != NULL);
1126 sq = sleepq_lookup(wchan);
1129 /* Thread is asleep on sleep queue sq, so wake it up. */
1130 return (sleepq_resume_thread(sq, td, 0));
1134 sleepq_chains_remove_matching(bool (*matches)(struct thread *))
1136 struct sleepqueue_chain *sc;
1137 struct sleepqueue *sq, *sq1;
1138 int i, wakeup_swapper;
1141 for (sc = &sleepq_chains[0]; sc < sleepq_chains + SC_TABLESIZE; ++sc) {
1142 if (LIST_EMPTY(&sc->sc_queues)) {
1145 mtx_lock_spin(&sc->sc_lock);
1146 LIST_FOREACH_SAFE(sq, &sc->sc_queues, sq_hash, sq1) {
1147 for (i = 0; i < NR_SLEEPQS; ++i) {
1148 wakeup_swapper |= sleepq_remove_matching(sq, i,
1152 mtx_unlock_spin(&sc->sc_lock);
1154 if (wakeup_swapper) {
1160 * Prints the stacks of all threads presently sleeping on wchan/queue to
1161 * the sbuf sb. Sets count_stacks_printed to the number of stacks actually
1162 * printed. Typically, this will equal the number of threads sleeping on the
1163 * queue, but may be less if sb overflowed before all stacks were printed.
1167 sleepq_sbuf_print_stacks(struct sbuf *sb, void *wchan, int queue,
1168 int *count_stacks_printed)
1170 struct thread *td, *td_next;
1171 struct sleepqueue *sq;
1173 struct sbuf **td_infos;
1174 int i, stack_idx, error, stacks_to_allocate;
1180 KASSERT(wchan != NULL, ("%s: invalid NULL wait channel", __func__));
1181 MPASS((queue >= 0) && (queue < NR_SLEEPQS));
1183 stacks_to_allocate = 10;
1184 for (i = 0; i < 3 && !finished ; i++) {
1185 /* We cannot malloc while holding the queue's spinlock, so
1186 * we do our mallocs now, and hope it is enough. If it
1187 * isn't, we will free these, drop the lock, malloc more,
1188 * and try again, up to a point. After that point we will
1189 * give up and report ENOMEM. We also cannot write to sb
1190 * during this time since the client may have set the
1191 * SBUF_AUTOEXTEND flag on their sbuf, which could cause a
1192 * malloc as we print to it. So we defer actually printing
1193 * to sb until after we drop the spinlock.
1196 /* Where we will store the stacks. */
1197 st = malloc(sizeof(struct stack *) * stacks_to_allocate,
1199 for (stack_idx = 0; stack_idx < stacks_to_allocate;
1201 st[stack_idx] = stack_create(M_WAITOK);
1203 /* Where we will store the td name, tid, etc. */
1204 td_infos = malloc(sizeof(struct sbuf *) * stacks_to_allocate,
1206 for (stack_idx = 0; stack_idx < stacks_to_allocate;
1208 td_infos[stack_idx] = sbuf_new(NULL, NULL,
1209 MAXCOMLEN + sizeof(struct thread *) * 2 + 40,
1213 sq = sleepq_lookup(wchan);
1215 /* This sleepq does not exist; exit and return ENOENT. */
1218 sleepq_release(wchan);
1223 /* Save thread info */
1224 TAILQ_FOREACH_SAFE(td, &sq->sq_blocked[queue], td_slpq,
1226 if (stack_idx >= stacks_to_allocate)
1229 /* Note the td_lock is equal to the sleepq_lock here. */
1230 stack_save_td(st[stack_idx], td);
1232 sbuf_printf(td_infos[stack_idx], "%d: %s %p",
1233 td->td_tid, td->td_name, td);
1239 sleepq_release(wchan);
1241 /* Print the stacks */
1242 for (i = 0; i < stack_idx; i++) {
1243 sbuf_finish(td_infos[i]);
1244 sbuf_printf(sb, "--- thread %s: ---\n", sbuf_data(td_infos[i]));
1245 stack_sbuf_print(sb, st[i]);
1246 sbuf_printf(sb, "\n");
1248 error = sbuf_error(sb);
1250 *count_stacks_printed = stack_idx;
1255 sleepq_release(wchan);
1256 for (stack_idx = 0; stack_idx < stacks_to_allocate;
1258 stack_destroy(st[stack_idx]);
1259 for (stack_idx = 0; stack_idx < stacks_to_allocate;
1261 sbuf_delete(td_infos[stack_idx]);
1263 free(td_infos, M_TEMP);
1264 stacks_to_allocate *= 10;
1267 if (!finished && error == 0)
1274 #ifdef SLEEPQUEUE_PROFILING
1275 #define SLEEPQ_PROF_LOCATIONS 1024
1276 #define SLEEPQ_SBUFSIZE 512
1277 struct sleepq_prof {
1278 LIST_ENTRY(sleepq_prof) sp_link;
1279 const char *sp_wmesg;
1283 LIST_HEAD(sqphead, sleepq_prof);
1285 struct sqphead sleepq_prof_free;
1286 struct sqphead sleepq_hash[SC_TABLESIZE];
1287 static struct sleepq_prof sleepq_profent[SLEEPQ_PROF_LOCATIONS];
1288 static struct mtx sleepq_prof_lock;
1289 MTX_SYSINIT(sleepq_prof_lock, &sleepq_prof_lock, "sleepq_prof", MTX_SPIN);
1292 sleepq_profile(const char *wmesg)
1294 struct sleepq_prof *sp;
1296 mtx_lock_spin(&sleepq_prof_lock);
1297 if (prof_enabled == 0)
1299 LIST_FOREACH(sp, &sleepq_hash[SC_HASH(wmesg)], sp_link)
1300 if (sp->sp_wmesg == wmesg)
1302 sp = LIST_FIRST(&sleepq_prof_free);
1305 sp->sp_wmesg = wmesg;
1306 LIST_REMOVE(sp, sp_link);
1307 LIST_INSERT_HEAD(&sleepq_hash[SC_HASH(wmesg)], sp, sp_link);
1311 mtx_unlock_spin(&sleepq_prof_lock);
1316 sleepq_prof_reset(void)
1318 struct sleepq_prof *sp;
1322 mtx_lock_spin(&sleepq_prof_lock);
1323 enabled = prof_enabled;
1325 for (i = 0; i < SC_TABLESIZE; i++)
1326 LIST_INIT(&sleepq_hash[i]);
1327 LIST_INIT(&sleepq_prof_free);
1328 for (i = 0; i < SLEEPQ_PROF_LOCATIONS; i++) {
1329 sp = &sleepq_profent[i];
1330 sp->sp_wmesg = NULL;
1332 LIST_INSERT_HEAD(&sleepq_prof_free, sp, sp_link);
1334 prof_enabled = enabled;
1335 mtx_unlock_spin(&sleepq_prof_lock);
1339 enable_sleepq_prof(SYSCTL_HANDLER_ARGS)
1344 error = sysctl_handle_int(oidp, &v, v, req);
1347 if (req->newptr == NULL)
1349 if (v == prof_enabled)
1352 sleepq_prof_reset();
1353 mtx_lock_spin(&sleepq_prof_lock);
1355 mtx_unlock_spin(&sleepq_prof_lock);
1361 reset_sleepq_prof_stats(SYSCTL_HANDLER_ARGS)
1366 error = sysctl_handle_int(oidp, &v, 0, req);
1369 if (req->newptr == NULL)
1373 sleepq_prof_reset();
1379 dump_sleepq_prof_stats(SYSCTL_HANDLER_ARGS)
1381 struct sleepq_prof *sp;
1387 error = sysctl_wire_old_buffer(req, 0);
1390 sb = sbuf_new_for_sysctl(NULL, NULL, SLEEPQ_SBUFSIZE, req);
1391 sbuf_printf(sb, "\nwmesg\tcount\n");
1392 enabled = prof_enabled;
1393 mtx_lock_spin(&sleepq_prof_lock);
1395 mtx_unlock_spin(&sleepq_prof_lock);
1396 for (i = 0; i < SC_TABLESIZE; i++) {
1397 LIST_FOREACH(sp, &sleepq_hash[i], sp_link) {
1398 sbuf_printf(sb, "%s\t%ld\n",
1399 sp->sp_wmesg, sp->sp_count);
1402 mtx_lock_spin(&sleepq_prof_lock);
1403 prof_enabled = enabled;
1404 mtx_unlock_spin(&sleepq_prof_lock);
1406 error = sbuf_finish(sb);
1411 SYSCTL_PROC(_debug_sleepq, OID_AUTO, stats, CTLTYPE_STRING | CTLFLAG_RD,
1412 NULL, 0, dump_sleepq_prof_stats, "A", "Sleepqueue profiling statistics");
1413 SYSCTL_PROC(_debug_sleepq, OID_AUTO, reset, CTLTYPE_INT | CTLFLAG_RW,
1414 NULL, 0, reset_sleepq_prof_stats, "I",
1415 "Reset sleepqueue profiling statistics");
1416 SYSCTL_PROC(_debug_sleepq, OID_AUTO, enable, CTLTYPE_INT | CTLFLAG_RW,
1417 NULL, 0, enable_sleepq_prof, "I", "Enable sleepqueue profiling");
1421 DB_SHOW_COMMAND(sleepq, db_show_sleepqueue)
1423 struct sleepqueue_chain *sc;
1424 struct sleepqueue *sq;
1426 struct lock_object *lock;
1436 * First, see if there is an active sleep queue for the wait channel
1437 * indicated by the address.
1439 wchan = (void *)addr;
1440 sc = SC_LOOKUP(wchan);
1441 LIST_FOREACH(sq, &sc->sc_queues, sq_hash)
1442 if (sq->sq_wchan == wchan)
1446 * Second, see if there is an active sleep queue at the address
1449 for (i = 0; i < SC_TABLESIZE; i++)
1450 LIST_FOREACH(sq, &sleepq_chains[i].sc_queues, sq_hash) {
1451 if (sq == (struct sleepqueue *)addr)
1455 db_printf("Unable to locate a sleep queue via %p\n", (void *)addr);
1458 db_printf("Wait channel: %p\n", sq->sq_wchan);
1459 db_printf("Queue type: %d\n", sq->sq_type);
1463 db_printf("Associated Interlock: %p - (%s) %s\n", lock,
1464 LOCK_CLASS(lock)->lc_name, lock->lo_name);
1467 db_printf("Blocked threads:\n");
1468 for (i = 0; i < NR_SLEEPQS; i++) {
1469 db_printf("\nQueue[%d]:\n", i);
1470 if (TAILQ_EMPTY(&sq->sq_blocked[i]))
1471 db_printf("\tempty\n");
1473 TAILQ_FOREACH(td, &sq->sq_blocked[i],
1475 db_printf("\t%p (tid %d, pid %d, \"%s\")\n", td,
1476 td->td_tid, td->td_proc->p_pid,
1479 db_printf("(expected: %u)\n", sq->sq_blockedcnt[i]);
1483 /* Alias 'show sleepqueue' to 'show sleepq'. */
1484 DB_SHOW_ALIAS(sleepqueue, db_show_sleepqueue);