subr_sleepqueue.c revision 184653
1139804Simp/*- 2126324Sjhb * Copyright (c) 2004 John Baldwin <jhb@FreeBSD.org> 3126324Sjhb * All rights reserved. 4126324Sjhb * 5126324Sjhb * Redistribution and use in source and binary forms, with or without 6126324Sjhb * modification, are permitted provided that the following conditions 7126324Sjhb * are met: 8126324Sjhb * 1. Redistributions of source code must retain the above copyright 9126324Sjhb * notice, this list of conditions and the following disclaimer. 10126324Sjhb * 2. Redistributions in binary form must reproduce the above copyright 11126324Sjhb * notice, this list of conditions and the following disclaimer in the 12126324Sjhb * documentation and/or other materials provided with the distribution. 13126324Sjhb * 3. Neither the name of the author nor the names of any co-contributors 14126324Sjhb * may be used to endorse or promote products derived from this software 15126324Sjhb * without specific prior written permission. 16126324Sjhb * 17126324Sjhb * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND 18126324Sjhb * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 19126324Sjhb * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 20126324Sjhb * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE 21126324Sjhb * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 22126324Sjhb * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 23126324Sjhb * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 24126324Sjhb * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 25126324Sjhb * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 26126324Sjhb * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 27126324Sjhb * SUCH DAMAGE. 28126324Sjhb */ 29126324Sjhb 30126324Sjhb/* 31126324Sjhb * Implementation of sleep queues used to hold queue of threads blocked on 32126324Sjhb * a wait channel. Sleep queues different from turnstiles in that wait 33126324Sjhb * channels are not owned by anyone, so there is no priority propagation. 34126324Sjhb * Sleep queues can also provide a timeout and can also be interrupted by 35126324Sjhb * signals. That said, there are several similarities between the turnstile 36126324Sjhb * and sleep queue implementations. (Note: turnstiles were implemented 37126324Sjhb * first.) For example, both use a hash table of the same size where each 38126324Sjhb * bucket is referred to as a "chain" that contains both a spin lock and 39126324Sjhb * a linked list of queues. An individual queue is located by using a hash 40126324Sjhb * to pick a chain, locking the chain, and then walking the chain searching 41126324Sjhb * for the queue. This means that a wait channel object does not need to 42126324Sjhb * embed it's queue head just as locks do not embed their turnstile queue 43126324Sjhb * head. Threads also carry around a sleep queue that they lend to the 44126324Sjhb * wait channel when blocking. Just as in turnstiles, the queue includes 45126324Sjhb * a free list of the sleep queues of other threads blocked on the same 46126324Sjhb * wait channel in the case of multiple waiters. 47126324Sjhb * 48126324Sjhb * Some additional functionality provided by sleep queues include the 49126324Sjhb * ability to set a timeout. The timeout is managed using a per-thread 50126324Sjhb * callout that resumes a thread if it is asleep. A thread may also 51126324Sjhb * catch signals while it is asleep (aka an interruptible sleep). The 52126324Sjhb * signal code uses sleepq_abort() to interrupt a sleeping thread. Finally, 53126324Sjhb * sleep queues also provide some extra assertions. One is not allowed to 54126324Sjhb * mix the sleep/wakeup and cv APIs for a given wait channel. Also, one 55126324Sjhb * must consistently use the same lock to synchronize with a wait channel, 56126324Sjhb * though this check is currently only a warning for sleep/wakeup due to 57126324Sjhb * pre-existing abuse of that API. The same lock must also be held when 58126324Sjhb * awakening threads, though that is currently only enforced for condition 59126324Sjhb * variables. 60126324Sjhb */ 61126324Sjhb 62126324Sjhb#include <sys/cdefs.h> 63126324Sjhb__FBSDID("$FreeBSD: head/sys/kern/subr_sleepqueue.c 184653 2008-11-04 19:13:53Z jhb $"); 64126324Sjhb 65154936Sjhb#include "opt_sleepqueue_profiling.h" 66154936Sjhb#include "opt_ddb.h" 67170640Sjeff#include "opt_sched.h" 68154936Sjhb 69126324Sjhb#include <sys/param.h> 70126324Sjhb#include <sys/systm.h> 71126324Sjhb#include <sys/lock.h> 72126324Sjhb#include <sys/kernel.h> 73126324Sjhb#include <sys/ktr.h> 74126324Sjhb#include <sys/mutex.h> 75126324Sjhb#include <sys/proc.h> 76177372Sjeff#include <sys/sbuf.h> 77126324Sjhb#include <sys/sched.h> 78126324Sjhb#include <sys/signalvar.h> 79126324Sjhb#include <sys/sleepqueue.h> 80131259Sjhb#include <sys/sysctl.h> 81126324Sjhb 82169666Sjeff#include <vm/uma.h> 83169666Sjeff 84154936Sjhb#ifdef DDB 85154936Sjhb#include <ddb/ddb.h> 86154936Sjhb#endif 87154936Sjhb 88126324Sjhb/* 89126324Sjhb * Constants for the hash table of sleep queue chains. These constants are 90126324Sjhb * the same ones that 4BSD (and possibly earlier versions of BSD) used. 91126324Sjhb * Basically, we ignore the lower 8 bits of the address since most wait 92126324Sjhb * channel pointers are aligned and only look at the next 7 bits for the 93126324Sjhb * hash. SC_TABLESIZE must be a power of two for SC_MASK to work properly. 94126324Sjhb */ 95126324Sjhb#define SC_TABLESIZE 128 /* Must be power of 2. */ 96126324Sjhb#define SC_MASK (SC_TABLESIZE - 1) 97126324Sjhb#define SC_SHIFT 8 98126324Sjhb#define SC_HASH(wc) (((uintptr_t)(wc) >> SC_SHIFT) & SC_MASK) 99126324Sjhb#define SC_LOOKUP(wc) &sleepq_chains[SC_HASH(wc)] 100165272Skmacy#define NR_SLEEPQS 2 101126324Sjhb/* 102126324Sjhb * There two different lists of sleep queues. Both lists are connected 103126324Sjhb * via the sq_hash entries. The first list is the sleep queue chain list 104126324Sjhb * that a sleep queue is on when it is attached to a wait channel. The 105126324Sjhb * second list is the free list hung off of a sleep queue that is attached 106126324Sjhb * to a wait channel. 107126324Sjhb * 108126324Sjhb * Each sleep queue also contains the wait channel it is attached to, the 109126324Sjhb * list of threads blocked on that wait channel, flags specific to the 110126324Sjhb * wait channel, and the lock used to synchronize with a wait channel. 111126324Sjhb * The flags are used to catch mismatches between the various consumers 112126324Sjhb * of the sleep queue API (e.g. sleep/wakeup and condition variables). 113126324Sjhb * The lock pointer is only used when invariants are enabled for various 114126324Sjhb * debugging checks. 115126324Sjhb * 116126324Sjhb * Locking key: 117126324Sjhb * c - sleep queue chain lock 118126324Sjhb */ 119126324Sjhbstruct sleepqueue { 120165272Skmacy TAILQ_HEAD(, thread) sq_blocked[NR_SLEEPQS]; /* (c) Blocked threads. */ 121126324Sjhb LIST_ENTRY(sleepqueue) sq_hash; /* (c) Chain and free list. */ 122126324Sjhb LIST_HEAD(, sleepqueue) sq_free; /* (c) Free queues. */ 123126324Sjhb void *sq_wchan; /* (c) Wait channel. */ 124136445Sjhb#ifdef INVARIANTS 125134013Sjhb int sq_type; /* (c) Queue type. */ 126164325Spjd struct lock_object *sq_lock; /* (c) Associated lock. */ 127126324Sjhb#endif 128126324Sjhb}; 129126324Sjhb 130126324Sjhbstruct sleepqueue_chain { 131126324Sjhb LIST_HEAD(, sleepqueue) sc_queues; /* List of sleep queues. */ 132126324Sjhb struct mtx sc_lock; /* Spin lock for this chain. */ 133131259Sjhb#ifdef SLEEPQUEUE_PROFILING 134131259Sjhb u_int sc_depth; /* Length of sc_queues. */ 135131259Sjhb u_int sc_max_depth; /* Max length of sc_queues. */ 136131259Sjhb#endif 137126324Sjhb}; 138126324Sjhb 139131259Sjhb#ifdef SLEEPQUEUE_PROFILING 140131259Sjhbu_int sleepq_max_depth; 141131259SjhbSYSCTL_NODE(_debug, OID_AUTO, sleepq, CTLFLAG_RD, 0, "sleepq profiling"); 142131259SjhbSYSCTL_NODE(_debug_sleepq, OID_AUTO, chains, CTLFLAG_RD, 0, 143131259Sjhb "sleepq chain stats"); 144131259SjhbSYSCTL_UINT(_debug_sleepq, OID_AUTO, max_depth, CTLFLAG_RD, &sleepq_max_depth, 145131259Sjhb 0, "maxmimum depth achieved of a single chain"); 146177372Sjeff 147177372Sjeffstatic void sleepq_profile(const char *wmesg); 148177372Sjeffstatic int prof_enabled; 149131259Sjhb#endif 150126324Sjhbstatic struct sleepqueue_chain sleepq_chains[SC_TABLESIZE]; 151169666Sjeffstatic uma_zone_t sleepq_zone; 152126324Sjhb 153126324Sjhb/* 154126324Sjhb * Prototypes for non-exported routines. 155126324Sjhb */ 156177085Sjeffstatic int sleepq_catch_signals(void *wchan, int pri); 157165272Skmacystatic int sleepq_check_signals(void); 158126324Sjhbstatic int sleepq_check_timeout(void); 159169666Sjeff#ifdef INVARIANTS 160169666Sjeffstatic void sleepq_dtor(void *mem, int size, void *arg); 161169666Sjeff#endif 162169666Sjeffstatic int sleepq_init(void *mem, int size, int flags); 163181334Sjhbstatic int sleepq_resume_thread(struct sleepqueue *sq, struct thread *td, 164169666Sjeff int pri); 165177085Sjeffstatic void sleepq_switch(void *wchan, int pri); 166126324Sjhbstatic void sleepq_timeout(void *arg); 167126324Sjhb 168126324Sjhb/* 169126324Sjhb * Early initialization of sleep queues that is called from the sleepinit() 170126324Sjhb * SYSINIT. 171126324Sjhb */ 172126324Sjhbvoid 173126324Sjhbinit_sleepqueues(void) 174126324Sjhb{ 175131259Sjhb#ifdef SLEEPQUEUE_PROFILING 176131259Sjhb struct sysctl_oid *chain_oid; 177131259Sjhb char chain_name[10]; 178131259Sjhb#endif 179126324Sjhb int i; 180126324Sjhb 181126324Sjhb for (i = 0; i < SC_TABLESIZE; i++) { 182126324Sjhb LIST_INIT(&sleepq_chains[i].sc_queues); 183126324Sjhb mtx_init(&sleepq_chains[i].sc_lock, "sleepq chain", NULL, 184176258Sjhb MTX_SPIN | MTX_RECURSE); 185131259Sjhb#ifdef SLEEPQUEUE_PROFILING 186131259Sjhb snprintf(chain_name, sizeof(chain_name), "%d", i); 187131259Sjhb chain_oid = SYSCTL_ADD_NODE(NULL, 188131259Sjhb SYSCTL_STATIC_CHILDREN(_debug_sleepq_chains), OID_AUTO, 189131259Sjhb chain_name, CTLFLAG_RD, NULL, "sleepq chain stats"); 190131259Sjhb SYSCTL_ADD_UINT(NULL, SYSCTL_CHILDREN(chain_oid), OID_AUTO, 191131259Sjhb "depth", CTLFLAG_RD, &sleepq_chains[i].sc_depth, 0, NULL); 192131259Sjhb SYSCTL_ADD_UINT(NULL, SYSCTL_CHILDREN(chain_oid), OID_AUTO, 193131259Sjhb "max_depth", CTLFLAG_RD, &sleepq_chains[i].sc_max_depth, 0, 194131259Sjhb NULL); 195131259Sjhb#endif 196126324Sjhb } 197169666Sjeff sleepq_zone = uma_zcreate("SLEEPQUEUE", sizeof(struct sleepqueue), 198169666Sjeff#ifdef INVARIANTS 199169666Sjeff NULL, sleepq_dtor, sleepq_init, NULL, UMA_ALIGN_CACHE, 0); 200169666Sjeff#else 201169666Sjeff NULL, NULL, sleepq_init, NULL, UMA_ALIGN_CACHE, 0); 202169666Sjeff#endif 203169666Sjeff 204126324Sjhb thread0.td_sleepqueue = sleepq_alloc(); 205126324Sjhb} 206126324Sjhb 207126324Sjhb/* 208169666Sjeff * Get a sleep queue for a new thread. 209126324Sjhb */ 210126324Sjhbstruct sleepqueue * 211126324Sjhbsleepq_alloc(void) 212126324Sjhb{ 213126324Sjhb 214169666Sjeff return (uma_zalloc(sleepq_zone, M_WAITOK)); 215126324Sjhb} 216126324Sjhb 217126324Sjhb/* 218126324Sjhb * Free a sleep queue when a thread is destroyed. 219126324Sjhb */ 220126324Sjhbvoid 221126324Sjhbsleepq_free(struct sleepqueue *sq) 222126324Sjhb{ 223126324Sjhb 224169666Sjeff uma_zfree(sleepq_zone, sq); 225126324Sjhb} 226126324Sjhb 227126324Sjhb/* 228136445Sjhb * Lock the sleep queue chain associated with the specified wait channel. 229136445Sjhb */ 230136445Sjhbvoid 231136445Sjhbsleepq_lock(void *wchan) 232136445Sjhb{ 233136445Sjhb struct sleepqueue_chain *sc; 234136445Sjhb 235136445Sjhb sc = SC_LOOKUP(wchan); 236136445Sjhb mtx_lock_spin(&sc->sc_lock); 237136445Sjhb} 238136445Sjhb 239136445Sjhb/* 240126324Sjhb * Look up the sleep queue associated with a given wait channel in the hash 241136445Sjhb * table locking the associated sleep queue chain. If no queue is found in 242136445Sjhb * the table, NULL is returned. 243126324Sjhb */ 244126324Sjhbstruct sleepqueue * 245126324Sjhbsleepq_lookup(void *wchan) 246126324Sjhb{ 247126324Sjhb struct sleepqueue_chain *sc; 248126324Sjhb struct sleepqueue *sq; 249126324Sjhb 250126324Sjhb KASSERT(wchan != NULL, ("%s: invalid NULL wait channel", __func__)); 251126324Sjhb sc = SC_LOOKUP(wchan); 252136445Sjhb mtx_assert(&sc->sc_lock, MA_OWNED); 253126324Sjhb LIST_FOREACH(sq, &sc->sc_queues, sq_hash) 254126324Sjhb if (sq->sq_wchan == wchan) 255126324Sjhb return (sq); 256126324Sjhb return (NULL); 257126324Sjhb} 258126324Sjhb 259126324Sjhb/* 260126324Sjhb * Unlock the sleep queue chain associated with a given wait channel. 261126324Sjhb */ 262126324Sjhbvoid 263126324Sjhbsleepq_release(void *wchan) 264126324Sjhb{ 265126324Sjhb struct sleepqueue_chain *sc; 266126324Sjhb 267126324Sjhb sc = SC_LOOKUP(wchan); 268126324Sjhb mtx_unlock_spin(&sc->sc_lock); 269126324Sjhb} 270126324Sjhb 271126324Sjhb/* 272137277Sjhb * Places the current thread on the sleep queue for the specified wait 273126324Sjhb * channel. If INVARIANTS is enabled, then it associates the passed in 274126324Sjhb * lock with the sleepq to make sure it is held when that sleep queue is 275126324Sjhb * woken up. 276126324Sjhb */ 277126324Sjhbvoid 278165272Skmacysleepq_add(void *wchan, struct lock_object *lock, const char *wmesg, int flags, 279165272Skmacy int queue) 280126324Sjhb{ 281126324Sjhb struct sleepqueue_chain *sc; 282136445Sjhb struct sleepqueue *sq; 283137277Sjhb struct thread *td; 284126324Sjhb 285126324Sjhb td = curthread; 286126324Sjhb sc = SC_LOOKUP(wchan); 287126324Sjhb mtx_assert(&sc->sc_lock, MA_OWNED); 288126324Sjhb MPASS(td->td_sleepqueue != NULL); 289126324Sjhb MPASS(wchan != NULL); 290165272Skmacy MPASS((queue >= 0) && (queue < NR_SLEEPQS)); 291126324Sjhb 292150177Sjhb /* If this thread is not allowed to sleep, die a horrible death. */ 293150177Sjhb KASSERT(!(td->td_pflags & TDP_NOSLEEPING), 294152221Simp ("Trying sleep, but thread marked as sleeping prohibited")); 295150177Sjhb 296136445Sjhb /* Look up the sleep queue associated with the wait channel 'wchan'. */ 297136445Sjhb sq = sleepq_lookup(wchan); 298136445Sjhb 299136445Sjhb /* 300136445Sjhb * If the wait channel does not already have a sleep queue, use 301136445Sjhb * this thread's sleep queue. Otherwise, insert the current thread 302136445Sjhb * into the sleep queue already in use by this wait channel. 303136445Sjhb */ 304126324Sjhb if (sq == NULL) { 305165272Skmacy#ifdef INVARIANTS 306165292Skmacy int i; 307165291Sache 308165292Skmacy sq = td->td_sleepqueue; 309165292Skmacy for (i = 0; i < NR_SLEEPQS; i++) 310165292Skmacy KASSERT(TAILQ_EMPTY(&sq->sq_blocked[i]), 311165272Skmacy ("thread's sleep queue %d is not empty", i)); 312165272Skmacy KASSERT(LIST_EMPTY(&sq->sq_free), 313165272Skmacy ("thread's sleep queue has a non-empty free list")); 314165272Skmacy KASSERT(sq->sq_wchan == NULL, ("stale sq_wchan pointer")); 315165292Skmacy sq->sq_lock = lock; 316165292Skmacy sq->sq_type = flags & SLEEPQ_TYPE; 317165272Skmacy#endif 318131259Sjhb#ifdef SLEEPQUEUE_PROFILING 319131259Sjhb sc->sc_depth++; 320131259Sjhb if (sc->sc_depth > sc->sc_max_depth) { 321131259Sjhb sc->sc_max_depth = sc->sc_depth; 322131259Sjhb if (sc->sc_max_depth > sleepq_max_depth) 323131259Sjhb sleepq_max_depth = sc->sc_max_depth; 324131259Sjhb } 325131259Sjhb#endif 326165292Skmacy sq = td->td_sleepqueue; 327126324Sjhb LIST_INSERT_HEAD(&sc->sc_queues, sq, sq_hash); 328126324Sjhb sq->sq_wchan = wchan; 329126324Sjhb } else { 330126324Sjhb MPASS(wchan == sq->sq_wchan); 331126488Sjhb MPASS(lock == sq->sq_lock); 332136445Sjhb MPASS((flags & SLEEPQ_TYPE) == sq->sq_type); 333126324Sjhb LIST_INSERT_HEAD(&sq->sq_free, td->td_sleepqueue, sq_hash); 334126324Sjhb } 335172155Sattilio thread_lock(td); 336165272Skmacy TAILQ_INSERT_TAIL(&sq->sq_blocked[queue], td, td_slpq); 337126324Sjhb td->td_sleepqueue = NULL; 338165272Skmacy td->td_sqqueue = queue; 339126324Sjhb td->td_wchan = wchan; 340126324Sjhb td->td_wmesg = wmesg; 341155741Sdavidxu if (flags & SLEEPQ_INTERRUPTIBLE) { 342134013Sjhb td->td_flags |= TDF_SINTR; 343155741Sdavidxu td->td_flags &= ~TDF_SLEEPABORT; 344155741Sdavidxu } 345172155Sattilio thread_unlock(td); 346126324Sjhb} 347126324Sjhb 348126324Sjhb/* 349126324Sjhb * Sets a timeout that will remove the current thread from the specified 350126324Sjhb * sleep queue after timo ticks if the thread has not already been awakened. 351126324Sjhb */ 352126324Sjhbvoid 353126885Sjhbsleepq_set_timeout(void *wchan, int timo) 354126324Sjhb{ 355126324Sjhb struct sleepqueue_chain *sc; 356126324Sjhb struct thread *td; 357126324Sjhb 358126324Sjhb td = curthread; 359126324Sjhb sc = SC_LOOKUP(wchan); 360126324Sjhb mtx_assert(&sc->sc_lock, MA_OWNED); 361126324Sjhb MPASS(TD_ON_SLEEPQ(td)); 362126324Sjhb MPASS(td->td_sleepqueue == NULL); 363126324Sjhb MPASS(wchan != NULL); 364177860Sjeff callout_reset_curcpu(&td->td_slpcallout, timo, sleepq_timeout, td); 365126324Sjhb} 366126324Sjhb 367126324Sjhb/* 368126324Sjhb * Marks the pending sleep of the current thread as interruptible and 369126324Sjhb * makes an initial check for pending signals before putting a thread 370170294Sjeff * to sleep. Enters and exits with the thread lock held. Thread lock 371170294Sjeff * may have transitioned from the sleepq lock to a run lock. 372126324Sjhb */ 373155741Sdavidxustatic int 374177085Sjeffsleepq_catch_signals(void *wchan, int pri) 375126324Sjhb{ 376126324Sjhb struct sleepqueue_chain *sc; 377126324Sjhb struct sleepqueue *sq; 378126324Sjhb struct thread *td; 379126324Sjhb struct proc *p; 380155741Sdavidxu struct sigacts *ps; 381155741Sdavidxu int sig, ret; 382126324Sjhb 383126324Sjhb td = curthread; 384155741Sdavidxu p = curproc; 385126324Sjhb sc = SC_LOOKUP(wchan); 386126324Sjhb mtx_assert(&sc->sc_lock, MA_OWNED); 387126324Sjhb MPASS(wchan != NULL); 388177375Sjeff /* 389177375Sjeff * See if there are any pending signals for this thread. If not 390177375Sjeff * we can switch immediately. Otherwise do the signal processing 391177375Sjeff * directly. 392177375Sjeff */ 393177375Sjeff thread_lock(td); 394177471Sjeff if ((td->td_flags & (TDF_NEEDSIGCHK | TDF_NEEDSUSPCHK)) == 0) { 395177375Sjeff sleepq_switch(wchan, pri); 396177375Sjeff return (0); 397177375Sjeff } 398184199Sdavidxu 399177375Sjeff thread_unlock(td); 400177375Sjeff mtx_unlock_spin(&sc->sc_lock); 401129241Sbde CTR3(KTR_PROC, "sleepq catching signals: thread %p (pid %ld, %s)", 402173601Sjulian (void *)td, (long)p->p_pid, td->td_name); 403126324Sjhb PROC_LOCK(p); 404155741Sdavidxu ps = p->p_sigacts; 405155741Sdavidxu mtx_lock(&ps->ps_mtx); 406126324Sjhb sig = cursig(td); 407155741Sdavidxu if (sig == 0) { 408155741Sdavidxu mtx_unlock(&ps->ps_mtx); 409155741Sdavidxu ret = thread_suspend_check(1); 410155741Sdavidxu MPASS(ret == 0 || ret == EINTR || ret == ERESTART); 411155741Sdavidxu } else { 412155741Sdavidxu if (SIGISMEMBER(ps->ps_sigintr, sig)) 413155741Sdavidxu ret = EINTR; 414155741Sdavidxu else 415155741Sdavidxu ret = ERESTART; 416155741Sdavidxu mtx_unlock(&ps->ps_mtx); 417155741Sdavidxu } 418184199Sdavidxu 419170294Sjeff mtx_lock_spin(&sc->sc_lock); 420170294Sjeff thread_lock(td); 421184216Sdavidxu PROC_UNLOCK(p); 422184216Sdavidxu if (ret == 0) { 423184216Sdavidxu sleepq_switch(wchan, pri); 424184216Sdavidxu return (0); 425184216Sdavidxu } 426184199Sdavidxu 427155936Sdavidxu /* 428155936Sdavidxu * There were pending signals and this thread is still 429155936Sdavidxu * on the sleep queue, remove it from the sleep queue. 430155936Sdavidxu */ 431170294Sjeff if (TD_ON_SLEEPQ(td)) { 432170294Sjeff sq = sleepq_lookup(wchan); 433181334Sjhb if (sleepq_resume_thread(sq, td, 0)) { 434181334Sjhb#ifdef INVARIANTS 435181334Sjhb /* 436181334Sjhb * This thread hasn't gone to sleep yet, so it 437181334Sjhb * should not be swapped out. 438181334Sjhb */ 439181334Sjhb panic("not waking up swapper"); 440181334Sjhb#endif 441181334Sjhb } 442170294Sjeff } 443170294Sjeff mtx_unlock_spin(&sc->sc_lock); 444170294Sjeff MPASS(td->td_lock != &sc->sc_lock); 445155741Sdavidxu return (ret); 446126324Sjhb} 447126324Sjhb 448126324Sjhb/* 449170294Sjeff * Switches to another thread if we are still asleep on a sleep queue. 450170294Sjeff * Returns with thread lock. 451126324Sjhb */ 452126324Sjhbstatic void 453177085Sjeffsleepq_switch(void *wchan, int pri) 454126324Sjhb{ 455126324Sjhb struct sleepqueue_chain *sc; 456175654Sjhb struct sleepqueue *sq; 457126324Sjhb struct thread *td; 458126324Sjhb 459126324Sjhb td = curthread; 460126324Sjhb sc = SC_LOOKUP(wchan); 461126324Sjhb mtx_assert(&sc->sc_lock, MA_OWNED); 462170294Sjeff THREAD_LOCK_ASSERT(td, MA_OWNED); 463175654Sjhb 464175654Sjhb /* 465175654Sjhb * If we have a sleep queue, then we've already been woken up, so 466175654Sjhb * just return. 467175654Sjhb */ 468126324Sjhb if (td->td_sleepqueue != NULL) { 469126324Sjhb mtx_unlock_spin(&sc->sc_lock); 470126324Sjhb return; 471126324Sjhb } 472175654Sjhb 473175654Sjhb /* 474175654Sjhb * If TDF_TIMEOUT is set, then our sleep has been timed out 475175654Sjhb * already but we are still on the sleep queue, so dequeue the 476175654Sjhb * thread and return. 477175654Sjhb */ 478175654Sjhb if (td->td_flags & TDF_TIMEOUT) { 479175654Sjhb MPASS(TD_ON_SLEEPQ(td)); 480175654Sjhb sq = sleepq_lookup(wchan); 481181334Sjhb if (sleepq_resume_thread(sq, td, 0)) { 482181334Sjhb#ifdef INVARIANTS 483181334Sjhb /* 484181334Sjhb * This thread hasn't gone to sleep yet, so it 485181334Sjhb * should not be swapped out. 486181334Sjhb */ 487181334Sjhb panic("not waking up swapper"); 488181334Sjhb#endif 489181334Sjhb } 490175654Sjhb mtx_unlock_spin(&sc->sc_lock); 491175654Sjhb return; 492175654Sjhb } 493177372Sjeff#ifdef SLEEPQUEUE_PROFILING 494177372Sjeff if (prof_enabled) 495177372Sjeff sleepq_profile(td->td_wmesg); 496177372Sjeff#endif 497177085Sjeff MPASS(td->td_sleepqueue == NULL); 498177085Sjeff sched_sleep(td, pri); 499170294Sjeff thread_lock_set(td, &sc->sc_lock); 500126324Sjhb TD_SET_SLEEPING(td); 501178272Sjeff mi_switch(SW_VOL | SWT_SLEEPQ, NULL); 502126324Sjhb KASSERT(TD_IS_RUNNING(td), ("running but not TDS_RUNNING")); 503129241Sbde CTR3(KTR_PROC, "sleepq resume: thread %p (pid %ld, %s)", 504173600Sjulian (void *)td, (long)td->td_proc->p_pid, (void *)td->td_name); 505126324Sjhb} 506126324Sjhb 507126324Sjhb/* 508126324Sjhb * Check to see if we timed out. 509126324Sjhb */ 510126324Sjhbstatic int 511126324Sjhbsleepq_check_timeout(void) 512126324Sjhb{ 513126324Sjhb struct thread *td; 514126324Sjhb 515126324Sjhb td = curthread; 516170294Sjeff THREAD_LOCK_ASSERT(td, MA_OWNED); 517126324Sjhb 518126324Sjhb /* 519126324Sjhb * If TDF_TIMEOUT is set, we timed out. 520126324Sjhb */ 521126324Sjhb if (td->td_flags & TDF_TIMEOUT) { 522126324Sjhb td->td_flags &= ~TDF_TIMEOUT; 523126324Sjhb return (EWOULDBLOCK); 524126324Sjhb } 525126324Sjhb 526126324Sjhb /* 527126324Sjhb * If TDF_TIMOFAIL is set, the timeout ran after we had 528126324Sjhb * already been woken up. 529126324Sjhb */ 530126324Sjhb if (td->td_flags & TDF_TIMOFAIL) 531126324Sjhb td->td_flags &= ~TDF_TIMOFAIL; 532126324Sjhb 533126324Sjhb /* 534126324Sjhb * If callout_stop() fails, then the timeout is running on 535126324Sjhb * another CPU, so synchronize with it to avoid having it 536126324Sjhb * accidentally wake up a subsequent sleep. 537126324Sjhb */ 538126324Sjhb else if (callout_stop(&td->td_slpcallout) == 0) { 539126324Sjhb td->td_flags |= TDF_TIMEOUT; 540126324Sjhb TD_SET_SLEEPING(td); 541178272Sjeff mi_switch(SW_INVOL | SWT_SLEEPQTIMO, NULL); 542126324Sjhb } 543126324Sjhb return (0); 544126324Sjhb} 545126324Sjhb 546126324Sjhb/* 547126324Sjhb * Check to see if we were awoken by a signal. 548126324Sjhb */ 549126324Sjhbstatic int 550126324Sjhbsleepq_check_signals(void) 551126324Sjhb{ 552126324Sjhb struct thread *td; 553126324Sjhb 554126324Sjhb td = curthread; 555170294Sjeff THREAD_LOCK_ASSERT(td, MA_OWNED); 556126324Sjhb 557126324Sjhb /* We are no longer in an interruptible sleep. */ 558155741Sdavidxu if (td->td_flags & TDF_SINTR) 559155741Sdavidxu td->td_flags &= ~TDF_SINTR; 560126324Sjhb 561155741Sdavidxu if (td->td_flags & TDF_SLEEPABORT) { 562155741Sdavidxu td->td_flags &= ~TDF_SLEEPABORT; 563155741Sdavidxu return (td->td_intrval); 564155741Sdavidxu } 565155741Sdavidxu 566126324Sjhb return (0); 567126324Sjhb} 568126324Sjhb 569126324Sjhb/* 570126324Sjhb * Block the current thread until it is awakened from its sleep queue. 571126324Sjhb */ 572126324Sjhbvoid 573177085Sjeffsleepq_wait(void *wchan, int pri) 574126324Sjhb{ 575170294Sjeff struct thread *td; 576126324Sjhb 577170294Sjeff td = curthread; 578170294Sjeff MPASS(!(td->td_flags & TDF_SINTR)); 579170294Sjeff thread_lock(td); 580177085Sjeff sleepq_switch(wchan, pri); 581170294Sjeff thread_unlock(td); 582126324Sjhb} 583126324Sjhb 584126324Sjhb/* 585126324Sjhb * Block the current thread until it is awakened from its sleep queue 586126324Sjhb * or it is interrupted by a signal. 587126324Sjhb */ 588126324Sjhbint 589177085Sjeffsleepq_wait_sig(void *wchan, int pri) 590126324Sjhb{ 591155741Sdavidxu int rcatch; 592126324Sjhb int rval; 593126324Sjhb 594177085Sjeff rcatch = sleepq_catch_signals(wchan, pri); 595126324Sjhb rval = sleepq_check_signals(); 596170294Sjeff thread_unlock(curthread); 597155741Sdavidxu if (rcatch) 598155741Sdavidxu return (rcatch); 599126324Sjhb return (rval); 600126324Sjhb} 601126324Sjhb 602126324Sjhb/* 603126324Sjhb * Block the current thread until it is awakened from its sleep queue 604126324Sjhb * or it times out while waiting. 605126324Sjhb */ 606126324Sjhbint 607177085Sjeffsleepq_timedwait(void *wchan, int pri) 608126324Sjhb{ 609170294Sjeff struct thread *td; 610126324Sjhb int rval; 611126324Sjhb 612170294Sjeff td = curthread; 613170294Sjeff MPASS(!(td->td_flags & TDF_SINTR)); 614170294Sjeff thread_lock(td); 615177085Sjeff sleepq_switch(wchan, pri); 616126324Sjhb rval = sleepq_check_timeout(); 617170294Sjeff thread_unlock(td); 618170294Sjeff 619131249Sjhb return (rval); 620126324Sjhb} 621126324Sjhb 622126324Sjhb/* 623126324Sjhb * Block the current thread until it is awakened from its sleep queue, 624126324Sjhb * it is interrupted by a signal, or it times out waiting to be awakened. 625126324Sjhb */ 626126324Sjhbint 627177085Sjeffsleepq_timedwait_sig(void *wchan, int pri) 628126324Sjhb{ 629155741Sdavidxu int rcatch, rvalt, rvals; 630126324Sjhb 631177085Sjeff rcatch = sleepq_catch_signals(wchan, pri); 632126324Sjhb rvalt = sleepq_check_timeout(); 633126324Sjhb rvals = sleepq_check_signals(); 634170294Sjeff thread_unlock(curthread); 635155741Sdavidxu if (rcatch) 636155741Sdavidxu return (rcatch); 637155741Sdavidxu if (rvals) 638126324Sjhb return (rvals); 639155741Sdavidxu return (rvalt); 640126324Sjhb} 641126324Sjhb 642126324Sjhb/* 643145056Sjhb * Removes a thread from a sleep queue and makes it 644145056Sjhb * runnable. 645126324Sjhb */ 646181334Sjhbstatic int 647145056Sjhbsleepq_resume_thread(struct sleepqueue *sq, struct thread *td, int pri) 648126324Sjhb{ 649126324Sjhb struct sleepqueue_chain *sc; 650126324Sjhb 651126324Sjhb MPASS(td != NULL); 652126324Sjhb MPASS(sq->sq_wchan != NULL); 653126324Sjhb MPASS(td->td_wchan == sq->sq_wchan); 654165272Skmacy MPASS(td->td_sqqueue < NR_SLEEPQS && td->td_sqqueue >= 0); 655170294Sjeff THREAD_LOCK_ASSERT(td, MA_OWNED); 656126324Sjhb sc = SC_LOOKUP(sq->sq_wchan); 657126324Sjhb mtx_assert(&sc->sc_lock, MA_OWNED); 658126324Sjhb 659126324Sjhb /* Remove the thread from the queue. */ 660165272Skmacy TAILQ_REMOVE(&sq->sq_blocked[td->td_sqqueue], td, td_slpq); 661126324Sjhb 662126324Sjhb /* 663126324Sjhb * Get a sleep queue for this thread. If this is the last waiter, 664126324Sjhb * use the queue itself and take it out of the chain, otherwise, 665126324Sjhb * remove a queue from the free list. 666126324Sjhb */ 667126324Sjhb if (LIST_EMPTY(&sq->sq_free)) { 668126324Sjhb td->td_sleepqueue = sq; 669126324Sjhb#ifdef INVARIANTS 670126324Sjhb sq->sq_wchan = NULL; 671126324Sjhb#endif 672131259Sjhb#ifdef SLEEPQUEUE_PROFILING 673131259Sjhb sc->sc_depth--; 674131259Sjhb#endif 675126324Sjhb } else 676126324Sjhb td->td_sleepqueue = LIST_FIRST(&sq->sq_free); 677126324Sjhb LIST_REMOVE(td->td_sleepqueue, sq_hash); 678126324Sjhb 679129188Sjhb td->td_wmesg = NULL; 680129188Sjhb td->td_wchan = NULL; 681157743Sdavidxu td->td_flags &= ~TDF_SINTR; 682129188Sjhb 683129241Sbde CTR3(KTR_PROC, "sleepq_wakeup: thread %p (pid %ld, %s)", 684173600Sjulian (void *)td, (long)td->td_proc->p_pid, td->td_name); 685126324Sjhb 686126324Sjhb /* Adjust priority if requested. */ 687177085Sjeff MPASS(pri == 0 || (pri >= PRI_MIN && pri <= PRI_MAX)); 688177085Sjeff if (pri != 0 && td->td_priority > pri) 689136439Sups sched_prio(td, pri); 690184653Sjhb 691184653Sjhb /* 692184653Sjhb * Note that thread td might not be sleeping if it is running 693184653Sjhb * sleepq_catch_signals() on another CPU or is blocked on its 694184653Sjhb * proc lock to check signals. There's no need to mark the 695184653Sjhb * thread runnable in that case. 696184653Sjhb */ 697184653Sjhb if (TD_IS_SLEEPING(td)) { 698184653Sjhb TD_CLR_SLEEPING(td); 699184653Sjhb return (setrunnable(td)); 700184653Sjhb } 701184653Sjhb return (0); 702126324Sjhb} 703126324Sjhb 704169666Sjeff#ifdef INVARIANTS 705126324Sjhb/* 706169666Sjeff * UMA zone item deallocator. 707169666Sjeff */ 708169666Sjeffstatic void 709169666Sjeffsleepq_dtor(void *mem, int size, void *arg) 710169666Sjeff{ 711169666Sjeff struct sleepqueue *sq; 712169666Sjeff int i; 713169666Sjeff 714169666Sjeff sq = mem; 715169666Sjeff for (i = 0; i < NR_SLEEPQS; i++) 716169666Sjeff MPASS(TAILQ_EMPTY(&sq->sq_blocked[i])); 717169666Sjeff} 718169666Sjeff#endif 719169666Sjeff 720169666Sjeff/* 721169666Sjeff * UMA zone item initializer. 722169666Sjeff */ 723169666Sjeffstatic int 724169666Sjeffsleepq_init(void *mem, int size, int flags) 725169666Sjeff{ 726169666Sjeff struct sleepqueue *sq; 727169666Sjeff int i; 728169666Sjeff 729169666Sjeff bzero(mem, size); 730169666Sjeff sq = mem; 731169666Sjeff for (i = 0; i < NR_SLEEPQS; i++) 732169666Sjeff TAILQ_INIT(&sq->sq_blocked[i]); 733169666Sjeff LIST_INIT(&sq->sq_free); 734169666Sjeff return (0); 735169666Sjeff} 736169666Sjeff 737169666Sjeff/* 738126324Sjhb * Find the highest priority thread sleeping on a wait channel and resume it. 739126324Sjhb */ 740181334Sjhbint 741165272Skmacysleepq_signal(void *wchan, int flags, int pri, int queue) 742126324Sjhb{ 743126324Sjhb struct sleepqueue *sq; 744137277Sjhb struct thread *td, *besttd; 745181334Sjhb int wakeup_swapper; 746126324Sjhb 747126324Sjhb CTR2(KTR_PROC, "sleepq_signal(%p, %d)", wchan, flags); 748126324Sjhb KASSERT(wchan != NULL, ("%s: invalid NULL wait channel", __func__)); 749165272Skmacy MPASS((queue >= 0) && (queue < NR_SLEEPQS)); 750126324Sjhb sq = sleepq_lookup(wchan); 751170294Sjeff if (sq == NULL) 752181334Sjhb return (0); 753134013Sjhb KASSERT(sq->sq_type == (flags & SLEEPQ_TYPE), 754126324Sjhb ("%s: mismatch between sleep/wakeup and cv_*", __func__)); 755129188Sjhb 756137277Sjhb /* 757137277Sjhb * Find the highest priority thread on the queue. If there is a 758137277Sjhb * tie, use the thread that first appears in the queue as it has 759137277Sjhb * been sleeping the longest since threads are always added to 760137277Sjhb * the tail of sleep queues. 761137277Sjhb */ 762137277Sjhb besttd = NULL; 763165272Skmacy TAILQ_FOREACH(td, &sq->sq_blocked[queue], td_slpq) { 764137277Sjhb if (besttd == NULL || td->td_priority < besttd->td_priority) 765137277Sjhb besttd = td; 766137277Sjhb } 767137277Sjhb MPASS(besttd != NULL); 768170294Sjeff thread_lock(besttd); 769181334Sjhb wakeup_swapper = sleepq_resume_thread(sq, besttd, pri); 770170294Sjeff thread_unlock(besttd); 771181334Sjhb return (wakeup_swapper); 772126324Sjhb} 773126324Sjhb 774126324Sjhb/* 775126324Sjhb * Resume all threads sleeping on a specified wait channel. 776126324Sjhb */ 777181334Sjhbint 778165272Skmacysleepq_broadcast(void *wchan, int flags, int pri, int queue) 779126324Sjhb{ 780126324Sjhb struct sleepqueue *sq; 781182875Sjhb struct thread *td, *tdn; 782181334Sjhb int wakeup_swapper; 783126324Sjhb 784126324Sjhb CTR2(KTR_PROC, "sleepq_broadcast(%p, %d)", wchan, flags); 785126324Sjhb KASSERT(wchan != NULL, ("%s: invalid NULL wait channel", __func__)); 786165272Skmacy MPASS((queue >= 0) && (queue < NR_SLEEPQS)); 787126324Sjhb sq = sleepq_lookup(wchan); 788177085Sjeff if (sq == NULL) 789181334Sjhb return (0); 790134013Sjhb KASSERT(sq->sq_type == (flags & SLEEPQ_TYPE), 791126324Sjhb ("%s: mismatch between sleep/wakeup and cv_*", __func__)); 792129188Sjhb 793145056Sjhb /* Resume all blocked threads on the sleep queue. */ 794181334Sjhb wakeup_swapper = 0; 795182875Sjhb TAILQ_FOREACH_SAFE(td, &sq->sq_blocked[queue], td_slpq, tdn) { 796170294Sjeff thread_lock(td); 797181334Sjhb if (sleepq_resume_thread(sq, td, pri)) 798181334Sjhb wakeup_swapper = 1; 799170294Sjeff thread_unlock(td); 800170294Sjeff } 801181334Sjhb return (wakeup_swapper); 802126324Sjhb} 803126324Sjhb 804126324Sjhb/* 805126324Sjhb * Time sleeping threads out. When the timeout expires, the thread is 806126324Sjhb * removed from the sleep queue and made runnable if it is still asleep. 807126324Sjhb */ 808126324Sjhbstatic void 809126324Sjhbsleepq_timeout(void *arg) 810126324Sjhb{ 811170294Sjeff struct sleepqueue_chain *sc; 812126324Sjhb struct sleepqueue *sq; 813126324Sjhb struct thread *td; 814126324Sjhb void *wchan; 815181334Sjhb int wakeup_swapper; 816126324Sjhb 817129241Sbde td = arg; 818181334Sjhb wakeup_swapper = 0; 819129241Sbde CTR3(KTR_PROC, "sleepq_timeout: thread %p (pid %ld, %s)", 820173600Sjulian (void *)td, (long)td->td_proc->p_pid, (void *)td->td_name); 821126324Sjhb 822126324Sjhb /* 823126324Sjhb * First, see if the thread is asleep and get the wait channel if 824126324Sjhb * it is. 825126324Sjhb */ 826170294Sjeff thread_lock(td); 827170294Sjeff if (TD_IS_SLEEPING(td) && TD_ON_SLEEPQ(td)) { 828126324Sjhb wchan = td->td_wchan; 829170294Sjeff sc = SC_LOOKUP(wchan); 830176078Sjeff THREAD_LOCKPTR_ASSERT(td, &sc->sc_lock); 831126324Sjhb sq = sleepq_lookup(wchan); 832170294Sjeff MPASS(sq != NULL); 833170294Sjeff td->td_flags |= TDF_TIMEOUT; 834181334Sjhb wakeup_swapper = sleepq_resume_thread(sq, td, 0); 835170294Sjeff thread_unlock(td); 836181334Sjhb if (wakeup_swapper) 837181334Sjhb kick_proc0(); 838170294Sjeff return; 839126324Sjhb } 840175654Sjhb 841126324Sjhb /* 842175654Sjhb * If the thread is on the SLEEPQ but isn't sleeping yet, it 843175654Sjhb * can either be on another CPU in between sleepq_add() and 844175654Sjhb * one of the sleepq_*wait*() routines or it can be in 845175654Sjhb * sleepq_catch_signals(). 846126324Sjhb */ 847126324Sjhb if (TD_ON_SLEEPQ(td)) { 848175664Sjhb td->td_flags |= TDF_TIMEOUT; 849170294Sjeff thread_unlock(td); 850126324Sjhb return; 851170294Sjeff } 852126324Sjhb 853126324Sjhb /* 854126324Sjhb * Now check for the edge cases. First, if TDF_TIMEOUT is set, 855126324Sjhb * then the other thread has already yielded to us, so clear 856126324Sjhb * the flag and resume it. If TDF_TIMEOUT is not set, then the 857126324Sjhb * we know that the other thread is not on a sleep queue, but it 858126324Sjhb * hasn't resumed execution yet. In that case, set TDF_TIMOFAIL 859126324Sjhb * to let it know that the timeout has already run and doesn't 860126324Sjhb * need to be canceled. 861126324Sjhb */ 862126324Sjhb if (td->td_flags & TDF_TIMEOUT) { 863127085Sjhb MPASS(TD_IS_SLEEPING(td)); 864126324Sjhb td->td_flags &= ~TDF_TIMEOUT; 865126324Sjhb TD_CLR_SLEEPING(td); 866181334Sjhb wakeup_swapper = setrunnable(td); 867126324Sjhb } else 868126324Sjhb td->td_flags |= TDF_TIMOFAIL; 869170294Sjeff thread_unlock(td); 870181334Sjhb if (wakeup_swapper) 871181334Sjhb kick_proc0(); 872126324Sjhb} 873126324Sjhb 874126324Sjhb/* 875126324Sjhb * Resumes a specific thread from the sleep queue associated with a specific 876126324Sjhb * wait channel if it is on that queue. 877126324Sjhb */ 878126324Sjhbvoid 879126324Sjhbsleepq_remove(struct thread *td, void *wchan) 880126324Sjhb{ 881126324Sjhb struct sleepqueue *sq; 882181334Sjhb int wakeup_swapper; 883126324Sjhb 884126324Sjhb /* 885126324Sjhb * Look up the sleep queue for this wait channel, then re-check 886126324Sjhb * that the thread is asleep on that channel, if it is not, then 887126324Sjhb * bail. 888126324Sjhb */ 889126324Sjhb MPASS(wchan != NULL); 890136445Sjhb sleepq_lock(wchan); 891126324Sjhb sq = sleepq_lookup(wchan); 892170294Sjeff /* 893170294Sjeff * We can not lock the thread here as it may be sleeping on a 894170294Sjeff * different sleepq. However, holding the sleepq lock for this 895170294Sjeff * wchan can guarantee that we do not miss a wakeup for this 896170294Sjeff * channel. The asserts below will catch any false positives. 897170294Sjeff */ 898126324Sjhb if (!TD_ON_SLEEPQ(td) || td->td_wchan != wchan) { 899126324Sjhb sleepq_release(wchan); 900126324Sjhb return; 901126324Sjhb } 902170294Sjeff /* Thread is asleep on sleep queue sq, so wake it up. */ 903170294Sjeff thread_lock(td); 904126324Sjhb MPASS(sq != NULL); 905170294Sjeff MPASS(td->td_wchan == wchan); 906181334Sjhb wakeup_swapper = sleepq_resume_thread(sq, td, 0); 907170294Sjeff thread_unlock(td); 908126324Sjhb sleepq_release(wchan); 909181334Sjhb if (wakeup_swapper) 910181334Sjhb kick_proc0(); 911126324Sjhb} 912126324Sjhb 913126324Sjhb/* 914129241Sbde * Abort a thread as if an interrupt had occurred. Only abort 915129241Sbde * interruptible waits (unfortunately it isn't safe to abort others). 916126324Sjhb */ 917181334Sjhbint 918155741Sdavidxusleepq_abort(struct thread *td, int intrval) 919126324Sjhb{ 920170294Sjeff struct sleepqueue *sq; 921126324Sjhb void *wchan; 922126324Sjhb 923170294Sjeff THREAD_LOCK_ASSERT(td, MA_OWNED); 924126324Sjhb MPASS(TD_ON_SLEEPQ(td)); 925126324Sjhb MPASS(td->td_flags & TDF_SINTR); 926155741Sdavidxu MPASS(intrval == EINTR || intrval == ERESTART); 927126324Sjhb 928126324Sjhb /* 929126324Sjhb * If the TDF_TIMEOUT flag is set, just leave. A 930126324Sjhb * timeout is scheduled anyhow. 931126324Sjhb */ 932126324Sjhb if (td->td_flags & TDF_TIMEOUT) 933181334Sjhb return (0); 934126324Sjhb 935129241Sbde CTR3(KTR_PROC, "sleepq_abort: thread %p (pid %ld, %s)", 936173600Sjulian (void *)td, (long)td->td_proc->p_pid, (void *)td->td_name); 937170294Sjeff td->td_intrval = intrval; 938170294Sjeff td->td_flags |= TDF_SLEEPABORT; 939170294Sjeff /* 940170294Sjeff * If the thread has not slept yet it will find the signal in 941170294Sjeff * sleepq_catch_signals() and call sleepq_resume_thread. Otherwise 942170294Sjeff * we have to do it here. 943170294Sjeff */ 944170294Sjeff if (!TD_IS_SLEEPING(td)) 945181334Sjhb return (0); 946126324Sjhb wchan = td->td_wchan; 947170294Sjeff MPASS(wchan != NULL); 948170294Sjeff sq = sleepq_lookup(wchan); 949170294Sjeff MPASS(sq != NULL); 950170294Sjeff 951170294Sjeff /* Thread is asleep on sleep queue sq, so wake it up. */ 952181334Sjhb return (sleepq_resume_thread(sq, td, 0)); 953126324Sjhb} 954154936Sjhb 955177372Sjeff#ifdef SLEEPQUEUE_PROFILING 956177372Sjeff#define SLEEPQ_PROF_LOCATIONS 1024 957177372Sjeff#define SLEEPQ_SBUFSIZE (40 * 512) 958177372Sjeffstruct sleepq_prof { 959177372Sjeff LIST_ENTRY(sleepq_prof) sp_link; 960177372Sjeff const char *sp_wmesg; 961177372Sjeff long sp_count; 962177372Sjeff}; 963177372Sjeff 964177372SjeffLIST_HEAD(sqphead, sleepq_prof); 965177372Sjeff 966177372Sjeffstruct sqphead sleepq_prof_free; 967177372Sjeffstruct sqphead sleepq_hash[SC_TABLESIZE]; 968177372Sjeffstatic struct sleepq_prof sleepq_profent[SLEEPQ_PROF_LOCATIONS]; 969177372Sjeffstatic struct mtx sleepq_prof_lock; 970177372SjeffMTX_SYSINIT(sleepq_prof_lock, &sleepq_prof_lock, "sleepq_prof", MTX_SPIN); 971177372Sjeff 972177372Sjeffstatic void 973177372Sjeffsleepq_profile(const char *wmesg) 974177372Sjeff{ 975177372Sjeff struct sleepq_prof *sp; 976177372Sjeff 977177372Sjeff mtx_lock_spin(&sleepq_prof_lock); 978177372Sjeff if (prof_enabled == 0) 979177372Sjeff goto unlock; 980177372Sjeff LIST_FOREACH(sp, &sleepq_hash[SC_HASH(wmesg)], sp_link) 981177372Sjeff if (sp->sp_wmesg == wmesg) 982177372Sjeff goto done; 983177372Sjeff sp = LIST_FIRST(&sleepq_prof_free); 984177372Sjeff if (sp == NULL) 985177372Sjeff goto unlock; 986177372Sjeff sp->sp_wmesg = wmesg; 987177372Sjeff LIST_REMOVE(sp, sp_link); 988177372Sjeff LIST_INSERT_HEAD(&sleepq_hash[SC_HASH(wmesg)], sp, sp_link); 989177372Sjeffdone: 990177372Sjeff sp->sp_count++; 991177372Sjeffunlock: 992177372Sjeff mtx_unlock_spin(&sleepq_prof_lock); 993177372Sjeff return; 994177372Sjeff} 995177372Sjeff 996177372Sjeffstatic void 997177372Sjeffsleepq_prof_reset(void) 998177372Sjeff{ 999177372Sjeff struct sleepq_prof *sp; 1000177372Sjeff int enabled; 1001177372Sjeff int i; 1002177372Sjeff 1003177372Sjeff mtx_lock_spin(&sleepq_prof_lock); 1004177372Sjeff enabled = prof_enabled; 1005177372Sjeff prof_enabled = 0; 1006177372Sjeff for (i = 0; i < SC_TABLESIZE; i++) 1007177372Sjeff LIST_INIT(&sleepq_hash[i]); 1008177372Sjeff LIST_INIT(&sleepq_prof_free); 1009177372Sjeff for (i = 0; i < SLEEPQ_PROF_LOCATIONS; i++) { 1010177372Sjeff sp = &sleepq_profent[i]; 1011177372Sjeff sp->sp_wmesg = NULL; 1012177372Sjeff sp->sp_count = 0; 1013177372Sjeff LIST_INSERT_HEAD(&sleepq_prof_free, sp, sp_link); 1014177372Sjeff } 1015177372Sjeff prof_enabled = enabled; 1016177372Sjeff mtx_unlock_spin(&sleepq_prof_lock); 1017177372Sjeff} 1018177372Sjeff 1019177372Sjeffstatic int 1020177372Sjeffenable_sleepq_prof(SYSCTL_HANDLER_ARGS) 1021177372Sjeff{ 1022177372Sjeff int error, v; 1023177372Sjeff 1024177372Sjeff v = prof_enabled; 1025177372Sjeff error = sysctl_handle_int(oidp, &v, v, req); 1026177372Sjeff if (error) 1027177372Sjeff return (error); 1028177372Sjeff if (req->newptr == NULL) 1029177372Sjeff return (error); 1030177372Sjeff if (v == prof_enabled) 1031177372Sjeff return (0); 1032177372Sjeff if (v == 1) 1033177372Sjeff sleepq_prof_reset(); 1034177372Sjeff mtx_lock_spin(&sleepq_prof_lock); 1035177372Sjeff prof_enabled = !!v; 1036177372Sjeff mtx_unlock_spin(&sleepq_prof_lock); 1037177372Sjeff 1038177372Sjeff return (0); 1039177372Sjeff} 1040177372Sjeff 1041177372Sjeffstatic int 1042177372Sjeffreset_sleepq_prof_stats(SYSCTL_HANDLER_ARGS) 1043177372Sjeff{ 1044177372Sjeff int error, v; 1045177372Sjeff 1046177372Sjeff v = 0; 1047177372Sjeff error = sysctl_handle_int(oidp, &v, 0, req); 1048177372Sjeff if (error) 1049177372Sjeff return (error); 1050177372Sjeff if (req->newptr == NULL) 1051177372Sjeff return (error); 1052177372Sjeff if (v == 0) 1053177372Sjeff return (0); 1054177372Sjeff sleepq_prof_reset(); 1055177372Sjeff 1056177372Sjeff return (0); 1057177372Sjeff} 1058177372Sjeff 1059177372Sjeffstatic int 1060177372Sjeffdump_sleepq_prof_stats(SYSCTL_HANDLER_ARGS) 1061177372Sjeff{ 1062177372Sjeff static int multiplier = 1; 1063177372Sjeff struct sleepq_prof *sp; 1064177372Sjeff struct sbuf *sb; 1065177372Sjeff int enabled; 1066177372Sjeff int error; 1067177372Sjeff int i; 1068177372Sjeff 1069177372Sjeffretry_sbufops: 1070177372Sjeff sb = sbuf_new(NULL, NULL, SLEEPQ_SBUFSIZE * multiplier, SBUF_FIXEDLEN); 1071177372Sjeff sbuf_printf(sb, "\nwmesg\tcount\n"); 1072177372Sjeff enabled = prof_enabled; 1073177372Sjeff mtx_lock_spin(&sleepq_prof_lock); 1074177372Sjeff prof_enabled = 0; 1075177372Sjeff mtx_unlock_spin(&sleepq_prof_lock); 1076177372Sjeff for (i = 0; i < SC_TABLESIZE; i++) { 1077177372Sjeff LIST_FOREACH(sp, &sleepq_hash[i], sp_link) { 1078177372Sjeff sbuf_printf(sb, "%s\t%ld\n", 1079177372Sjeff sp->sp_wmesg, sp->sp_count); 1080177372Sjeff if (sbuf_overflowed(sb)) { 1081177372Sjeff sbuf_delete(sb); 1082177372Sjeff multiplier++; 1083177372Sjeff goto retry_sbufops; 1084177372Sjeff } 1085177372Sjeff } 1086177372Sjeff } 1087177372Sjeff mtx_lock_spin(&sleepq_prof_lock); 1088177372Sjeff prof_enabled = enabled; 1089177372Sjeff mtx_unlock_spin(&sleepq_prof_lock); 1090177372Sjeff 1091177372Sjeff sbuf_finish(sb); 1092177372Sjeff error = SYSCTL_OUT(req, sbuf_data(sb), sbuf_len(sb) + 1); 1093177372Sjeff sbuf_delete(sb); 1094177372Sjeff return (error); 1095177372Sjeff} 1096177372Sjeff 1097177372SjeffSYSCTL_PROC(_debug_sleepq, OID_AUTO, stats, CTLTYPE_STRING | CTLFLAG_RD, 1098177372Sjeff NULL, 0, dump_sleepq_prof_stats, "A", "Sleepqueue profiling statistics"); 1099177372SjeffSYSCTL_PROC(_debug_sleepq, OID_AUTO, reset, CTLTYPE_INT | CTLFLAG_RW, 1100177372Sjeff NULL, 0, reset_sleepq_prof_stats, "I", 1101177372Sjeff "Reset sleepqueue profiling statistics"); 1102177372SjeffSYSCTL_PROC(_debug_sleepq, OID_AUTO, enable, CTLTYPE_INT | CTLFLAG_RW, 1103177372Sjeff NULL, 0, enable_sleepq_prof, "I", "Enable sleepqueue profiling"); 1104177372Sjeff#endif 1105177372Sjeff 1106154936Sjhb#ifdef DDB 1107154936SjhbDB_SHOW_COMMAND(sleepq, db_show_sleepqueue) 1108154936Sjhb{ 1109154936Sjhb struct sleepqueue_chain *sc; 1110154936Sjhb struct sleepqueue *sq; 1111154944Simp#ifdef INVARIANTS 1112154936Sjhb struct lock_object *lock; 1113154944Simp#endif 1114154936Sjhb struct thread *td; 1115154936Sjhb void *wchan; 1116154936Sjhb int i; 1117154936Sjhb 1118154936Sjhb if (!have_addr) 1119154936Sjhb return; 1120154936Sjhb 1121154936Sjhb /* 1122154936Sjhb * First, see if there is an active sleep queue for the wait channel 1123154936Sjhb * indicated by the address. 1124154936Sjhb */ 1125154936Sjhb wchan = (void *)addr; 1126154936Sjhb sc = SC_LOOKUP(wchan); 1127154936Sjhb LIST_FOREACH(sq, &sc->sc_queues, sq_hash) 1128154936Sjhb if (sq->sq_wchan == wchan) 1129154936Sjhb goto found; 1130154936Sjhb 1131154936Sjhb /* 1132154936Sjhb * Second, see if there is an active sleep queue at the address 1133154936Sjhb * indicated. 1134154936Sjhb */ 1135154936Sjhb for (i = 0; i < SC_TABLESIZE; i++) 1136154936Sjhb LIST_FOREACH(sq, &sleepq_chains[i].sc_queues, sq_hash) { 1137154936Sjhb if (sq == (struct sleepqueue *)addr) 1138154936Sjhb goto found; 1139154936Sjhb } 1140154936Sjhb 1141154936Sjhb db_printf("Unable to locate a sleep queue via %p\n", (void *)addr); 1142154936Sjhb return; 1143154936Sjhbfound: 1144154936Sjhb db_printf("Wait channel: %p\n", sq->sq_wchan); 1145154936Sjhb#ifdef INVARIANTS 1146154936Sjhb db_printf("Queue type: %d\n", sq->sq_type); 1147154936Sjhb if (sq->sq_lock) { 1148164325Spjd lock = sq->sq_lock; 1149154936Sjhb db_printf("Associated Interlock: %p - (%s) %s\n", lock, 1150154936Sjhb LOCK_CLASS(lock)->lc_name, lock->lo_name); 1151154936Sjhb } 1152154936Sjhb#endif 1153154936Sjhb db_printf("Blocked threads:\n"); 1154165272Skmacy for (i = 0; i < NR_SLEEPQS; i++) { 1155165272Skmacy db_printf("\nQueue[%d]:\n", i); 1156165272Skmacy if (TAILQ_EMPTY(&sq->sq_blocked[i])) 1157165272Skmacy db_printf("\tempty\n"); 1158165272Skmacy else 1159165272Skmacy TAILQ_FOREACH(td, &sq->sq_blocked[0], 1160165272Skmacy td_slpq) { 1161165272Skmacy db_printf("\t%p (tid %d, pid %d, \"%s\")\n", td, 1162165272Skmacy td->td_tid, td->td_proc->p_pid, 1163180930Sjhb td->td_name); 1164165272Skmacy } 1165165272Skmacy } 1166154936Sjhb} 1167157823Sjhb 1168157823Sjhb/* Alias 'show sleepqueue' to 'show sleepq'. */ 1169183054SsamDB_SHOW_ALIAS(sleepqueue, db_show_sleepqueue); 1170154936Sjhb#endif 1171