subr_sleepqueue.c revision 304905
1/*- 2 * Copyright (c) 2004 John Baldwin <jhb@FreeBSD.org> 3 * All rights reserved. 4 * 5 * Redistribution and use in source and binary forms, with or without 6 * modification, are permitted provided that the following conditions 7 * are met: 8 * 1. Redistributions of source code must retain the above copyright 9 * notice, this list of conditions and the following disclaimer. 10 * 2. Redistributions in binary form must reproduce the above copyright 11 * notice, this list of conditions and the following disclaimer in the 12 * documentation and/or other materials provided with the distribution. 13 * 14 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND 15 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 16 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 17 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE 18 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 19 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 20 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 21 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 22 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 23 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 24 * SUCH DAMAGE. 25 */ 26 27/* 28 * Implementation of sleep queues used to hold queue of threads blocked on 29 * a wait channel. Sleep queues different from turnstiles in that wait 30 * channels are not owned by anyone, so there is no priority propagation. 31 * Sleep queues can also provide a timeout and can also be interrupted by 32 * signals. That said, there are several similarities between the turnstile 33 * and sleep queue implementations. (Note: turnstiles were implemented 34 * first.) For example, both use a hash table of the same size where each 35 * bucket is referred to as a "chain" that contains both a spin lock and 36 * a linked list of queues. An individual queue is located by using a hash 37 * to pick a chain, locking the chain, and then walking the chain searching 38 * for the queue. This means that a wait channel object does not need to 39 * embed it's queue head just as locks do not embed their turnstile queue 40 * head. Threads also carry around a sleep queue that they lend to the 41 * wait channel when blocking. Just as in turnstiles, the queue includes 42 * a free list of the sleep queues of other threads blocked on the same 43 * wait channel in the case of multiple waiters. 44 * 45 * Some additional functionality provided by sleep queues include the 46 * ability to set a timeout. The timeout is managed using a per-thread 47 * callout that resumes a thread if it is asleep. A thread may also 48 * catch signals while it is asleep (aka an interruptible sleep). The 49 * signal code uses sleepq_abort() to interrupt a sleeping thread. Finally, 50 * sleep queues also provide some extra assertions. One is not allowed to 51 * mix the sleep/wakeup and cv APIs for a given wait channel. Also, one 52 * must consistently use the same lock to synchronize with a wait channel, 53 * though this check is currently only a warning for sleep/wakeup due to 54 * pre-existing abuse of that API. The same lock must also be held when 55 * awakening threads, though that is currently only enforced for condition 56 * variables. 57 */ 58 59#include <sys/cdefs.h> 60__FBSDID("$FreeBSD: stable/10/sys/kern/subr_sleepqueue.c 304905 2016-08-27 11:45:05Z kib $"); 61 62#include "opt_sleepqueue_profiling.h" 63#include "opt_ddb.h" 64#include "opt_kdtrace.h" 65#include "opt_sched.h" 66 67#include <sys/param.h> 68#include <sys/systm.h> 69#include <sys/lock.h> 70#include <sys/kernel.h> 71#include <sys/ktr.h> 72#include <sys/mutex.h> 73#include <sys/proc.h> 74#include <sys/sbuf.h> 75#include <sys/sched.h> 76#include <sys/sdt.h> 77#include <sys/signalvar.h> 78#include <sys/sleepqueue.h> 79#include <sys/sysctl.h> 80 81#include <vm/uma.h> 82 83#ifdef DDB 84#include <ddb/ddb.h> 85#endif 86 87/* 88 * Constants for the hash table of sleep queue chains. 89 * SC_TABLESIZE must be a power of two for SC_MASK to work properly. 90 */ 91#define SC_TABLESIZE 256 /* Must be power of 2. */ 92#define SC_MASK (SC_TABLESIZE - 1) 93#define SC_SHIFT 8 94#define SC_HASH(wc) ((((uintptr_t)(wc) >> SC_SHIFT) ^ (uintptr_t)(wc)) & \ 95 SC_MASK) 96#define SC_LOOKUP(wc) &sleepq_chains[SC_HASH(wc)] 97#define NR_SLEEPQS 2 98/* 99 * There two different lists of sleep queues. Both lists are connected 100 * via the sq_hash entries. The first list is the sleep queue chain list 101 * that a sleep queue is on when it is attached to a wait channel. The 102 * second list is the free list hung off of a sleep queue that is attached 103 * to a wait channel. 104 * 105 * Each sleep queue also contains the wait channel it is attached to, the 106 * list of threads blocked on that wait channel, flags specific to the 107 * wait channel, and the lock used to synchronize with a wait channel. 108 * The flags are used to catch mismatches between the various consumers 109 * of the sleep queue API (e.g. sleep/wakeup and condition variables). 110 * The lock pointer is only used when invariants are enabled for various 111 * debugging checks. 112 * 113 * Locking key: 114 * c - sleep queue chain lock 115 */ 116struct sleepqueue { 117 TAILQ_HEAD(, thread) sq_blocked[NR_SLEEPQS]; /* (c) Blocked threads. */ 118 u_int sq_blockedcnt[NR_SLEEPQS]; /* (c) N. of blocked threads. */ 119 LIST_ENTRY(sleepqueue) sq_hash; /* (c) Chain and free list. */ 120 LIST_HEAD(, sleepqueue) sq_free; /* (c) Free queues. */ 121 void *sq_wchan; /* (c) Wait channel. */ 122 int sq_type; /* (c) Queue type. */ 123#ifdef INVARIANTS 124 struct lock_object *sq_lock; /* (c) Associated lock. */ 125#endif 126}; 127 128struct sleepqueue_chain { 129 LIST_HEAD(, sleepqueue) sc_queues; /* List of sleep queues. */ 130 struct mtx sc_lock; /* Spin lock for this chain. */ 131#ifdef SLEEPQUEUE_PROFILING 132 u_int sc_depth; /* Length of sc_queues. */ 133 u_int sc_max_depth; /* Max length of sc_queues. */ 134#endif 135}; 136 137#ifdef SLEEPQUEUE_PROFILING 138u_int sleepq_max_depth; 139static SYSCTL_NODE(_debug, OID_AUTO, sleepq, CTLFLAG_RD, 0, "sleepq profiling"); 140static SYSCTL_NODE(_debug_sleepq, OID_AUTO, chains, CTLFLAG_RD, 0, 141 "sleepq chain stats"); 142SYSCTL_UINT(_debug_sleepq, OID_AUTO, max_depth, CTLFLAG_RD, &sleepq_max_depth, 143 0, "maxmimum depth achieved of a single chain"); 144 145static void sleepq_profile(const char *wmesg); 146static int prof_enabled; 147#endif 148static struct sleepqueue_chain sleepq_chains[SC_TABLESIZE]; 149static uma_zone_t sleepq_zone; 150 151/* 152 * Prototypes for non-exported routines. 153 */ 154static int sleepq_catch_signals(void *wchan, int pri); 155static int sleepq_check_signals(void); 156static int sleepq_check_timeout(void); 157#ifdef INVARIANTS 158static void sleepq_dtor(void *mem, int size, void *arg); 159#endif 160static int sleepq_init(void *mem, int size, int flags); 161static int sleepq_resume_thread(struct sleepqueue *sq, struct thread *td, 162 int pri); 163static void sleepq_switch(void *wchan, int pri); 164static void sleepq_timeout(void *arg); 165 166SDT_PROBE_DECLARE(sched, , , sleep); 167SDT_PROBE_DECLARE(sched, , , wakeup); 168 169/* 170 * Early initialization of sleep queues that is called from the sleepinit() 171 * SYSINIT. 172 */ 173void 174init_sleepqueues(void) 175{ 176#ifdef SLEEPQUEUE_PROFILING 177 struct sysctl_oid *chain_oid; 178 char chain_name[10]; 179#endif 180 int i; 181 182 for (i = 0; i < SC_TABLESIZE; i++) { 183 LIST_INIT(&sleepq_chains[i].sc_queues); 184 mtx_init(&sleepq_chains[i].sc_lock, "sleepq chain", NULL, 185 MTX_SPIN | MTX_RECURSE); 186#ifdef SLEEPQUEUE_PROFILING 187 snprintf(chain_name, sizeof(chain_name), "%d", i); 188 chain_oid = SYSCTL_ADD_NODE(NULL, 189 SYSCTL_STATIC_CHILDREN(_debug_sleepq_chains), OID_AUTO, 190 chain_name, CTLFLAG_RD, NULL, "sleepq chain stats"); 191 SYSCTL_ADD_UINT(NULL, SYSCTL_CHILDREN(chain_oid), OID_AUTO, 192 "depth", CTLFLAG_RD, &sleepq_chains[i].sc_depth, 0, NULL); 193 SYSCTL_ADD_UINT(NULL, SYSCTL_CHILDREN(chain_oid), OID_AUTO, 194 "max_depth", CTLFLAG_RD, &sleepq_chains[i].sc_max_depth, 0, 195 NULL); 196#endif 197 } 198 sleepq_zone = uma_zcreate("SLEEPQUEUE", sizeof(struct sleepqueue), 199#ifdef INVARIANTS 200 NULL, sleepq_dtor, sleepq_init, NULL, UMA_ALIGN_CACHE, 0); 201#else 202 NULL, NULL, sleepq_init, NULL, UMA_ALIGN_CACHE, 0); 203#endif 204 205 thread0.td_sleepqueue = sleepq_alloc(); 206} 207 208/* 209 * Get a sleep queue for a new thread. 210 */ 211struct sleepqueue * 212sleepq_alloc(void) 213{ 214 215 return (uma_zalloc(sleepq_zone, M_WAITOK)); 216} 217 218/* 219 * Free a sleep queue when a thread is destroyed. 220 */ 221void 222sleepq_free(struct sleepqueue *sq) 223{ 224 225 uma_zfree(sleepq_zone, sq); 226} 227 228/* 229 * Lock the sleep queue chain associated with the specified wait channel. 230 */ 231void 232sleepq_lock(void *wchan) 233{ 234 struct sleepqueue_chain *sc; 235 236 sc = SC_LOOKUP(wchan); 237 mtx_lock_spin(&sc->sc_lock); 238} 239 240/* 241 * Look up the sleep queue associated with a given wait channel in the hash 242 * table locking the associated sleep queue chain. If no queue is found in 243 * the table, NULL is returned. 244 */ 245struct sleepqueue * 246sleepq_lookup(void *wchan) 247{ 248 struct sleepqueue_chain *sc; 249 struct sleepqueue *sq; 250 251 KASSERT(wchan != NULL, ("%s: invalid NULL wait channel", __func__)); 252 sc = SC_LOOKUP(wchan); 253 mtx_assert(&sc->sc_lock, MA_OWNED); 254 LIST_FOREACH(sq, &sc->sc_queues, sq_hash) 255 if (sq->sq_wchan == wchan) 256 return (sq); 257 return (NULL); 258} 259 260/* 261 * Unlock the sleep queue chain associated with a given wait channel. 262 */ 263void 264sleepq_release(void *wchan) 265{ 266 struct sleepqueue_chain *sc; 267 268 sc = SC_LOOKUP(wchan); 269 mtx_unlock_spin(&sc->sc_lock); 270} 271 272/* 273 * Places the current thread on the sleep queue for the specified wait 274 * channel. If INVARIANTS is enabled, then it associates the passed in 275 * lock with the sleepq to make sure it is held when that sleep queue is 276 * woken up. 277 */ 278void 279sleepq_add(void *wchan, struct lock_object *lock, const char *wmesg, int flags, 280 int queue) 281{ 282 struct sleepqueue_chain *sc; 283 struct sleepqueue *sq; 284 struct thread *td; 285 286 td = curthread; 287 sc = SC_LOOKUP(wchan); 288 mtx_assert(&sc->sc_lock, MA_OWNED); 289 MPASS(td->td_sleepqueue != NULL); 290 MPASS(wchan != NULL); 291 MPASS((queue >= 0) && (queue < NR_SLEEPQS)); 292 293 /* If this thread is not allowed to sleep, die a horrible death. */ 294 KASSERT(td->td_no_sleeping == 0, 295 ("%s: td %p to sleep on wchan %p with sleeping prohibited", 296 __func__, td, wchan)); 297 298 /* Look up the sleep queue associated with the wait channel 'wchan'. */ 299 sq = sleepq_lookup(wchan); 300 301 /* 302 * If the wait channel does not already have a sleep queue, use 303 * this thread's sleep queue. Otherwise, insert the current thread 304 * into the sleep queue already in use by this wait channel. 305 */ 306 if (sq == NULL) { 307#ifdef INVARIANTS 308 int i; 309 310 sq = td->td_sleepqueue; 311 for (i = 0; i < NR_SLEEPQS; i++) { 312 KASSERT(TAILQ_EMPTY(&sq->sq_blocked[i]), 313 ("thread's sleep queue %d is not empty", i)); 314 KASSERT(sq->sq_blockedcnt[i] == 0, 315 ("thread's sleep queue %d count mismatches", i)); 316 } 317 KASSERT(LIST_EMPTY(&sq->sq_free), 318 ("thread's sleep queue has a non-empty free list")); 319 KASSERT(sq->sq_wchan == NULL, ("stale sq_wchan pointer")); 320 sq->sq_lock = lock; 321#endif 322#ifdef SLEEPQUEUE_PROFILING 323 sc->sc_depth++; 324 if (sc->sc_depth > sc->sc_max_depth) { 325 sc->sc_max_depth = sc->sc_depth; 326 if (sc->sc_max_depth > sleepq_max_depth) 327 sleepq_max_depth = sc->sc_max_depth; 328 } 329#endif 330 sq = td->td_sleepqueue; 331 LIST_INSERT_HEAD(&sc->sc_queues, sq, sq_hash); 332 sq->sq_wchan = wchan; 333 sq->sq_type = flags & SLEEPQ_TYPE; 334 } else { 335 MPASS(wchan == sq->sq_wchan); 336 MPASS(lock == sq->sq_lock); 337 MPASS((flags & SLEEPQ_TYPE) == sq->sq_type); 338 LIST_INSERT_HEAD(&sq->sq_free, td->td_sleepqueue, sq_hash); 339 } 340 thread_lock(td); 341 TAILQ_INSERT_TAIL(&sq->sq_blocked[queue], td, td_slpq); 342 sq->sq_blockedcnt[queue]++; 343 td->td_sleepqueue = NULL; 344 td->td_sqqueue = queue; 345 td->td_wchan = wchan; 346 td->td_wmesg = wmesg; 347 if (flags & SLEEPQ_INTERRUPTIBLE) { 348 td->td_flags |= TDF_SINTR; 349 td->td_flags &= ~TDF_SLEEPABORT; 350 } 351 thread_unlock(td); 352} 353 354/* 355 * Sets a timeout that will remove the current thread from the specified 356 * sleep queue after timo ticks if the thread has not already been awakened. 357 */ 358void 359sleepq_set_timeout_sbt(void *wchan, sbintime_t sbt, sbintime_t pr, 360 int flags) 361{ 362 struct sleepqueue_chain *sc; 363 struct thread *td; 364 sbintime_t pr1; 365 366 td = curthread; 367 sc = SC_LOOKUP(wchan); 368 mtx_assert(&sc->sc_lock, MA_OWNED); 369 MPASS(TD_ON_SLEEPQ(td)); 370 MPASS(td->td_sleepqueue == NULL); 371 MPASS(wchan != NULL); 372 KASSERT(td->td_sleeptimo == 0, ("td %d %p td_sleeptimo %jx", 373 td->td_tid, td, (uintmax_t)td->td_sleeptimo)); 374 thread_lock(td); 375 callout_when(sbt, pr, flags, &td->td_sleeptimo, &pr1); 376 thread_unlock(td); 377 callout_reset_sbt_on(&td->td_slpcallout, td->td_sleeptimo, pr1, 378 sleepq_timeout, td, PCPU_GET(cpuid), flags | C_PRECALC | 379 C_DIRECT_EXEC); 380} 381 382/* 383 * Return the number of actual sleepers for the specified queue. 384 */ 385u_int 386sleepq_sleepcnt(void *wchan, int queue) 387{ 388 struct sleepqueue *sq; 389 390 KASSERT(wchan != NULL, ("%s: invalid NULL wait channel", __func__)); 391 MPASS((queue >= 0) && (queue < NR_SLEEPQS)); 392 sq = sleepq_lookup(wchan); 393 if (sq == NULL) 394 return (0); 395 return (sq->sq_blockedcnt[queue]); 396} 397 398/* 399 * Marks the pending sleep of the current thread as interruptible and 400 * makes an initial check for pending signals before putting a thread 401 * to sleep. Enters and exits with the thread lock held. Thread lock 402 * may have transitioned from the sleepq lock to a run lock. 403 */ 404static int 405sleepq_catch_signals(void *wchan, int pri) 406{ 407 struct sleepqueue_chain *sc; 408 struct sleepqueue *sq; 409 struct thread *td; 410 struct proc *p; 411 struct sigacts *ps; 412 int sig, ret; 413 414 td = curthread; 415 p = curproc; 416 sc = SC_LOOKUP(wchan); 417 mtx_assert(&sc->sc_lock, MA_OWNED); 418 MPASS(wchan != NULL); 419 if ((td->td_pflags & TDP_WAKEUP) != 0) { 420 td->td_pflags &= ~TDP_WAKEUP; 421 ret = EINTR; 422 thread_lock(td); 423 goto out; 424 } 425 426 /* 427 * See if there are any pending signals for this thread. If not 428 * we can switch immediately. Otherwise do the signal processing 429 * directly. 430 */ 431 thread_lock(td); 432 if ((td->td_flags & (TDF_NEEDSIGCHK | TDF_NEEDSUSPCHK)) == 0) { 433 sleepq_switch(wchan, pri); 434 return (0); 435 } 436 thread_unlock(td); 437 mtx_unlock_spin(&sc->sc_lock); 438 CTR3(KTR_PROC, "sleepq catching signals: thread %p (pid %ld, %s)", 439 (void *)td, (long)p->p_pid, td->td_name); 440 PROC_LOCK(p); 441 ps = p->p_sigacts; 442 mtx_lock(&ps->ps_mtx); 443 sig = cursig(td); 444 if (sig == 0) { 445 mtx_unlock(&ps->ps_mtx); 446 ret = thread_suspend_check(1); 447 MPASS(ret == 0 || ret == EINTR || ret == ERESTART); 448 } else { 449 if (SIGISMEMBER(ps->ps_sigintr, sig)) 450 ret = EINTR; 451 else 452 ret = ERESTART; 453 mtx_unlock(&ps->ps_mtx); 454 } 455 /* 456 * Lock the per-process spinlock prior to dropping the PROC_LOCK 457 * to avoid a signal delivery race. PROC_LOCK, PROC_SLOCK, and 458 * thread_lock() are currently held in tdsendsignal(). 459 */ 460 PROC_SLOCK(p); 461 mtx_lock_spin(&sc->sc_lock); 462 PROC_UNLOCK(p); 463 thread_lock(td); 464 PROC_SUNLOCK(p); 465 if (ret == 0) { 466 sleepq_switch(wchan, pri); 467 return (0); 468 } 469out: 470 /* 471 * There were pending signals and this thread is still 472 * on the sleep queue, remove it from the sleep queue. 473 */ 474 if (TD_ON_SLEEPQ(td)) { 475 sq = sleepq_lookup(wchan); 476 if (sleepq_resume_thread(sq, td, 0)) { 477#ifdef INVARIANTS 478 /* 479 * This thread hasn't gone to sleep yet, so it 480 * should not be swapped out. 481 */ 482 panic("not waking up swapper"); 483#endif 484 } 485 } 486 mtx_unlock_spin(&sc->sc_lock); 487 MPASS(td->td_lock != &sc->sc_lock); 488 return (ret); 489} 490 491/* 492 * Switches to another thread if we are still asleep on a sleep queue. 493 * Returns with thread lock. 494 */ 495static void 496sleepq_switch(void *wchan, int pri) 497{ 498 struct sleepqueue_chain *sc; 499 struct sleepqueue *sq; 500 struct thread *td; 501 502 td = curthread; 503 sc = SC_LOOKUP(wchan); 504 mtx_assert(&sc->sc_lock, MA_OWNED); 505 THREAD_LOCK_ASSERT(td, MA_OWNED); 506 507 /* 508 * If we have a sleep queue, then we've already been woken up, so 509 * just return. 510 */ 511 if (td->td_sleepqueue != NULL) { 512 mtx_unlock_spin(&sc->sc_lock); 513 return; 514 } 515 516 /* 517 * If TDF_TIMEOUT is set, then our sleep has been timed out 518 * already but we are still on the sleep queue, so dequeue the 519 * thread and return. 520 */ 521 if (td->td_flags & TDF_TIMEOUT) { 522 MPASS(TD_ON_SLEEPQ(td)); 523 sq = sleepq_lookup(wchan); 524 if (sleepq_resume_thread(sq, td, 0)) { 525#ifdef INVARIANTS 526 /* 527 * This thread hasn't gone to sleep yet, so it 528 * should not be swapped out. 529 */ 530 panic("not waking up swapper"); 531#endif 532 } 533 mtx_unlock_spin(&sc->sc_lock); 534 return; 535 } 536#ifdef SLEEPQUEUE_PROFILING 537 if (prof_enabled) 538 sleepq_profile(td->td_wmesg); 539#endif 540 MPASS(td->td_sleepqueue == NULL); 541 sched_sleep(td, pri); 542 thread_lock_set(td, &sc->sc_lock); 543 SDT_PROBE0(sched, , , sleep); 544 TD_SET_SLEEPING(td); 545 mi_switch(SW_VOL | SWT_SLEEPQ, NULL); 546 KASSERT(TD_IS_RUNNING(td), ("running but not TDS_RUNNING")); 547 CTR3(KTR_PROC, "sleepq resume: thread %p (pid %ld, %s)", 548 (void *)td, (long)td->td_proc->p_pid, (void *)td->td_name); 549} 550 551/* 552 * Check to see if we timed out. 553 */ 554static int 555sleepq_check_timeout(void) 556{ 557 struct thread *td; 558 int res; 559 560 td = curthread; 561 THREAD_LOCK_ASSERT(td, MA_OWNED); 562 563 /* 564 * If TDF_TIMEOUT is set, we timed out. But recheck 565 * td_sleeptimo anyway. 566 */ 567 res = 0; 568 if (td->td_sleeptimo != 0) { 569 if (td->td_sleeptimo <= sbinuptime()) 570 res = EWOULDBLOCK; 571 td->td_sleeptimo = 0; 572 } 573 if (td->td_flags & TDF_TIMEOUT) 574 td->td_flags &= ~TDF_TIMEOUT; 575 else 576 /* 577 * We ignore the situation where timeout subsystem was 578 * unable to stop our callout. The struct thread is 579 * type-stable, the callout will use the correct 580 * memory when running. The checks of the 581 * td_sleeptimo value in this function and in 582 * sleepq_timeout() ensure that the thread does not 583 * get spurious wakeups, even if the callout was reset 584 * or thread reused. 585 */ 586 callout_stop(&td->td_slpcallout); 587 return (res); 588} 589 590/* 591 * Check to see if we were awoken by a signal. 592 */ 593static int 594sleepq_check_signals(void) 595{ 596 struct thread *td; 597 598 td = curthread; 599 THREAD_LOCK_ASSERT(td, MA_OWNED); 600 601 /* We are no longer in an interruptible sleep. */ 602 if (td->td_flags & TDF_SINTR) 603 td->td_flags &= ~TDF_SINTR; 604 605 if (td->td_flags & TDF_SLEEPABORT) { 606 td->td_flags &= ~TDF_SLEEPABORT; 607 return (td->td_intrval); 608 } 609 610 return (0); 611} 612 613/* 614 * Block the current thread until it is awakened from its sleep queue. 615 */ 616void 617sleepq_wait(void *wchan, int pri) 618{ 619 struct thread *td; 620 621 td = curthread; 622 MPASS(!(td->td_flags & TDF_SINTR)); 623 thread_lock(td); 624 sleepq_switch(wchan, pri); 625 thread_unlock(td); 626} 627 628/* 629 * Block the current thread until it is awakened from its sleep queue 630 * or it is interrupted by a signal. 631 */ 632int 633sleepq_wait_sig(void *wchan, int pri) 634{ 635 int rcatch; 636 int rval; 637 638 rcatch = sleepq_catch_signals(wchan, pri); 639 rval = sleepq_check_signals(); 640 thread_unlock(curthread); 641 if (rcatch) 642 return (rcatch); 643 return (rval); 644} 645 646/* 647 * Block the current thread until it is awakened from its sleep queue 648 * or it times out while waiting. 649 */ 650int 651sleepq_timedwait(void *wchan, int pri) 652{ 653 struct thread *td; 654 int rval; 655 656 td = curthread; 657 MPASS(!(td->td_flags & TDF_SINTR)); 658 thread_lock(td); 659 sleepq_switch(wchan, pri); 660 rval = sleepq_check_timeout(); 661 thread_unlock(td); 662 663 return (rval); 664} 665 666/* 667 * Block the current thread until it is awakened from its sleep queue, 668 * it is interrupted by a signal, or it times out waiting to be awakened. 669 */ 670int 671sleepq_timedwait_sig(void *wchan, int pri) 672{ 673 int rcatch, rvalt, rvals; 674 675 rcatch = sleepq_catch_signals(wchan, pri); 676 rvalt = sleepq_check_timeout(); 677 rvals = sleepq_check_signals(); 678 thread_unlock(curthread); 679 if (rcatch) 680 return (rcatch); 681 if (rvals) 682 return (rvals); 683 return (rvalt); 684} 685 686/* 687 * Returns the type of sleepqueue given a waitchannel. 688 */ 689int 690sleepq_type(void *wchan) 691{ 692 struct sleepqueue *sq; 693 int type; 694 695 MPASS(wchan != NULL); 696 697 sleepq_lock(wchan); 698 sq = sleepq_lookup(wchan); 699 if (sq == NULL) { 700 sleepq_release(wchan); 701 return (-1); 702 } 703 type = sq->sq_type; 704 sleepq_release(wchan); 705 return (type); 706} 707 708/* 709 * Removes a thread from a sleep queue and makes it 710 * runnable. 711 */ 712static int 713sleepq_resume_thread(struct sleepqueue *sq, struct thread *td, int pri) 714{ 715 struct sleepqueue_chain *sc; 716 717 MPASS(td != NULL); 718 MPASS(sq->sq_wchan != NULL); 719 MPASS(td->td_wchan == sq->sq_wchan); 720 MPASS(td->td_sqqueue < NR_SLEEPQS && td->td_sqqueue >= 0); 721 THREAD_LOCK_ASSERT(td, MA_OWNED); 722 sc = SC_LOOKUP(sq->sq_wchan); 723 mtx_assert(&sc->sc_lock, MA_OWNED); 724 725 SDT_PROBE2(sched, , , wakeup, td, td->td_proc); 726 727 /* Remove the thread from the queue. */ 728 sq->sq_blockedcnt[td->td_sqqueue]--; 729 TAILQ_REMOVE(&sq->sq_blocked[td->td_sqqueue], td, td_slpq); 730 731 /* 732 * Get a sleep queue for this thread. If this is the last waiter, 733 * use the queue itself and take it out of the chain, otherwise, 734 * remove a queue from the free list. 735 */ 736 if (LIST_EMPTY(&sq->sq_free)) { 737 td->td_sleepqueue = sq; 738#ifdef INVARIANTS 739 sq->sq_wchan = NULL; 740#endif 741#ifdef SLEEPQUEUE_PROFILING 742 sc->sc_depth--; 743#endif 744 } else 745 td->td_sleepqueue = LIST_FIRST(&sq->sq_free); 746 LIST_REMOVE(td->td_sleepqueue, sq_hash); 747 748 td->td_wmesg = NULL; 749 td->td_wchan = NULL; 750 td->td_flags &= ~TDF_SINTR; 751 752 CTR3(KTR_PROC, "sleepq_wakeup: thread %p (pid %ld, %s)", 753 (void *)td, (long)td->td_proc->p_pid, td->td_name); 754 755 /* Adjust priority if requested. */ 756 MPASS(pri == 0 || (pri >= PRI_MIN && pri <= PRI_MAX)); 757 if (pri != 0 && td->td_priority > pri && 758 PRI_BASE(td->td_pri_class) == PRI_TIMESHARE) 759 sched_prio(td, pri); 760 761 /* 762 * Note that thread td might not be sleeping if it is running 763 * sleepq_catch_signals() on another CPU or is blocked on its 764 * proc lock to check signals. There's no need to mark the 765 * thread runnable in that case. 766 */ 767 if (TD_IS_SLEEPING(td)) { 768 TD_CLR_SLEEPING(td); 769 return (setrunnable(td)); 770 } 771 return (0); 772} 773 774#ifdef INVARIANTS 775/* 776 * UMA zone item deallocator. 777 */ 778static void 779sleepq_dtor(void *mem, int size, void *arg) 780{ 781 struct sleepqueue *sq; 782 int i; 783 784 sq = mem; 785 for (i = 0; i < NR_SLEEPQS; i++) { 786 MPASS(TAILQ_EMPTY(&sq->sq_blocked[i])); 787 MPASS(sq->sq_blockedcnt[i] == 0); 788 } 789} 790#endif 791 792/* 793 * UMA zone item initializer. 794 */ 795static int 796sleepq_init(void *mem, int size, int flags) 797{ 798 struct sleepqueue *sq; 799 int i; 800 801 bzero(mem, size); 802 sq = mem; 803 for (i = 0; i < NR_SLEEPQS; i++) { 804 TAILQ_INIT(&sq->sq_blocked[i]); 805 sq->sq_blockedcnt[i] = 0; 806 } 807 LIST_INIT(&sq->sq_free); 808 return (0); 809} 810 811/* 812 * Find the highest priority thread sleeping on a wait channel and resume it. 813 */ 814int 815sleepq_signal(void *wchan, int flags, int pri, int queue) 816{ 817 struct sleepqueue *sq; 818 struct thread *td, *besttd; 819 int wakeup_swapper; 820 821 CTR2(KTR_PROC, "sleepq_signal(%p, %d)", wchan, flags); 822 KASSERT(wchan != NULL, ("%s: invalid NULL wait channel", __func__)); 823 MPASS((queue >= 0) && (queue < NR_SLEEPQS)); 824 sq = sleepq_lookup(wchan); 825 if (sq == NULL) 826 return (0); 827 KASSERT(sq->sq_type == (flags & SLEEPQ_TYPE), 828 ("%s: mismatch between sleep/wakeup and cv_*", __func__)); 829 830 /* 831 * Find the highest priority thread on the queue. If there is a 832 * tie, use the thread that first appears in the queue as it has 833 * been sleeping the longest since threads are always added to 834 * the tail of sleep queues. 835 */ 836 besttd = NULL; 837 TAILQ_FOREACH(td, &sq->sq_blocked[queue], td_slpq) { 838 if (besttd == NULL || td->td_priority < besttd->td_priority) 839 besttd = td; 840 } 841 MPASS(besttd != NULL); 842 thread_lock(besttd); 843 wakeup_swapper = sleepq_resume_thread(sq, besttd, pri); 844 thread_unlock(besttd); 845 return (wakeup_swapper); 846} 847 848/* 849 * Resume all threads sleeping on a specified wait channel. 850 */ 851int 852sleepq_broadcast(void *wchan, int flags, int pri, int queue) 853{ 854 struct sleepqueue *sq; 855 struct thread *td, *tdn; 856 int wakeup_swapper; 857 858 CTR2(KTR_PROC, "sleepq_broadcast(%p, %d)", wchan, flags); 859 KASSERT(wchan != NULL, ("%s: invalid NULL wait channel", __func__)); 860 MPASS((queue >= 0) && (queue < NR_SLEEPQS)); 861 sq = sleepq_lookup(wchan); 862 if (sq == NULL) 863 return (0); 864 KASSERT(sq->sq_type == (flags & SLEEPQ_TYPE), 865 ("%s: mismatch between sleep/wakeup and cv_*", __func__)); 866 867 /* Resume all blocked threads on the sleep queue. */ 868 wakeup_swapper = 0; 869 TAILQ_FOREACH_SAFE(td, &sq->sq_blocked[queue], td_slpq, tdn) { 870 thread_lock(td); 871 if (sleepq_resume_thread(sq, td, pri)) 872 wakeup_swapper = 1; 873 thread_unlock(td); 874 } 875 return (wakeup_swapper); 876} 877 878/* 879 * Time sleeping threads out. When the timeout expires, the thread is 880 * removed from the sleep queue and made runnable if it is still asleep. 881 */ 882static void 883sleepq_timeout(void *arg) 884{ 885 struct sleepqueue_chain *sc; 886 struct sleepqueue *sq; 887 struct thread *td; 888 void *wchan; 889 int wakeup_swapper; 890 891 td = arg; 892 wakeup_swapper = 0; 893 CTR3(KTR_PROC, "sleepq_timeout: thread %p (pid %ld, %s)", 894 (void *)td, (long)td->td_proc->p_pid, (void *)td->td_name); 895 896 thread_lock(td); 897 898 if (td->td_sleeptimo > sbinuptime() || td->td_sleeptimo == 0) { 899 /* 900 * The thread does not want a timeout (yet). 901 */ 902 } else if (TD_IS_SLEEPING(td) && TD_ON_SLEEPQ(td)) { 903 /* 904 * See if the thread is asleep and get the wait 905 * channel if it is. 906 */ 907 wchan = td->td_wchan; 908 sc = SC_LOOKUP(wchan); 909 THREAD_LOCKPTR_ASSERT(td, &sc->sc_lock); 910 sq = sleepq_lookup(wchan); 911 MPASS(sq != NULL); 912 td->td_flags |= TDF_TIMEOUT; 913 wakeup_swapper = sleepq_resume_thread(sq, td, 0); 914 } else if (TD_ON_SLEEPQ(td)) { 915 /* 916 * If the thread is on the SLEEPQ but isn't sleeping 917 * yet, it can either be on another CPU in between 918 * sleepq_add() and one of the sleepq_*wait*() 919 * routines or it can be in sleepq_catch_signals(). 920 */ 921 td->td_flags |= TDF_TIMEOUT; 922 } 923 924 thread_unlock(td); 925 if (wakeup_swapper) 926 kick_proc0(); 927} 928 929/* 930 * Resumes a specific thread from the sleep queue associated with a specific 931 * wait channel if it is on that queue. 932 */ 933void 934sleepq_remove(struct thread *td, void *wchan) 935{ 936 struct sleepqueue *sq; 937 int wakeup_swapper; 938 939 /* 940 * Look up the sleep queue for this wait channel, then re-check 941 * that the thread is asleep on that channel, if it is not, then 942 * bail. 943 */ 944 MPASS(wchan != NULL); 945 sleepq_lock(wchan); 946 sq = sleepq_lookup(wchan); 947 /* 948 * We can not lock the thread here as it may be sleeping on a 949 * different sleepq. However, holding the sleepq lock for this 950 * wchan can guarantee that we do not miss a wakeup for this 951 * channel. The asserts below will catch any false positives. 952 */ 953 if (!TD_ON_SLEEPQ(td) || td->td_wchan != wchan) { 954 sleepq_release(wchan); 955 return; 956 } 957 /* Thread is asleep on sleep queue sq, so wake it up. */ 958 thread_lock(td); 959 MPASS(sq != NULL); 960 MPASS(td->td_wchan == wchan); 961 wakeup_swapper = sleepq_resume_thread(sq, td, 0); 962 thread_unlock(td); 963 sleepq_release(wchan); 964 if (wakeup_swapper) 965 kick_proc0(); 966} 967 968/* 969 * Abort a thread as if an interrupt had occurred. Only abort 970 * interruptible waits (unfortunately it isn't safe to abort others). 971 */ 972int 973sleepq_abort(struct thread *td, int intrval) 974{ 975 struct sleepqueue *sq; 976 void *wchan; 977 978 THREAD_LOCK_ASSERT(td, MA_OWNED); 979 MPASS(TD_ON_SLEEPQ(td)); 980 MPASS(td->td_flags & TDF_SINTR); 981 MPASS(intrval == EINTR || intrval == ERESTART); 982 983 /* 984 * If the TDF_TIMEOUT flag is set, just leave. A 985 * timeout is scheduled anyhow. 986 */ 987 if (td->td_flags & TDF_TIMEOUT) 988 return (0); 989 990 CTR3(KTR_PROC, "sleepq_abort: thread %p (pid %ld, %s)", 991 (void *)td, (long)td->td_proc->p_pid, (void *)td->td_name); 992 td->td_intrval = intrval; 993 td->td_flags |= TDF_SLEEPABORT; 994 /* 995 * If the thread has not slept yet it will find the signal in 996 * sleepq_catch_signals() and call sleepq_resume_thread. Otherwise 997 * we have to do it here. 998 */ 999 if (!TD_IS_SLEEPING(td)) 1000 return (0); 1001 wchan = td->td_wchan; 1002 MPASS(wchan != NULL); 1003 sq = sleepq_lookup(wchan); 1004 MPASS(sq != NULL); 1005 1006 /* Thread is asleep on sleep queue sq, so wake it up. */ 1007 return (sleepq_resume_thread(sq, td, 0)); 1008} 1009 1010#ifdef SLEEPQUEUE_PROFILING 1011#define SLEEPQ_PROF_LOCATIONS 1024 1012#define SLEEPQ_SBUFSIZE 512 1013struct sleepq_prof { 1014 LIST_ENTRY(sleepq_prof) sp_link; 1015 const char *sp_wmesg; 1016 long sp_count; 1017}; 1018 1019LIST_HEAD(sqphead, sleepq_prof); 1020 1021struct sqphead sleepq_prof_free; 1022struct sqphead sleepq_hash[SC_TABLESIZE]; 1023static struct sleepq_prof sleepq_profent[SLEEPQ_PROF_LOCATIONS]; 1024static struct mtx sleepq_prof_lock; 1025MTX_SYSINIT(sleepq_prof_lock, &sleepq_prof_lock, "sleepq_prof", MTX_SPIN); 1026 1027static void 1028sleepq_profile(const char *wmesg) 1029{ 1030 struct sleepq_prof *sp; 1031 1032 mtx_lock_spin(&sleepq_prof_lock); 1033 if (prof_enabled == 0) 1034 goto unlock; 1035 LIST_FOREACH(sp, &sleepq_hash[SC_HASH(wmesg)], sp_link) 1036 if (sp->sp_wmesg == wmesg) 1037 goto done; 1038 sp = LIST_FIRST(&sleepq_prof_free); 1039 if (sp == NULL) 1040 goto unlock; 1041 sp->sp_wmesg = wmesg; 1042 LIST_REMOVE(sp, sp_link); 1043 LIST_INSERT_HEAD(&sleepq_hash[SC_HASH(wmesg)], sp, sp_link); 1044done: 1045 sp->sp_count++; 1046unlock: 1047 mtx_unlock_spin(&sleepq_prof_lock); 1048 return; 1049} 1050 1051static void 1052sleepq_prof_reset(void) 1053{ 1054 struct sleepq_prof *sp; 1055 int enabled; 1056 int i; 1057 1058 mtx_lock_spin(&sleepq_prof_lock); 1059 enabled = prof_enabled; 1060 prof_enabled = 0; 1061 for (i = 0; i < SC_TABLESIZE; i++) 1062 LIST_INIT(&sleepq_hash[i]); 1063 LIST_INIT(&sleepq_prof_free); 1064 for (i = 0; i < SLEEPQ_PROF_LOCATIONS; i++) { 1065 sp = &sleepq_profent[i]; 1066 sp->sp_wmesg = NULL; 1067 sp->sp_count = 0; 1068 LIST_INSERT_HEAD(&sleepq_prof_free, sp, sp_link); 1069 } 1070 prof_enabled = enabled; 1071 mtx_unlock_spin(&sleepq_prof_lock); 1072} 1073 1074static int 1075enable_sleepq_prof(SYSCTL_HANDLER_ARGS) 1076{ 1077 int error, v; 1078 1079 v = prof_enabled; 1080 error = sysctl_handle_int(oidp, &v, v, req); 1081 if (error) 1082 return (error); 1083 if (req->newptr == NULL) 1084 return (error); 1085 if (v == prof_enabled) 1086 return (0); 1087 if (v == 1) 1088 sleepq_prof_reset(); 1089 mtx_lock_spin(&sleepq_prof_lock); 1090 prof_enabled = !!v; 1091 mtx_unlock_spin(&sleepq_prof_lock); 1092 1093 return (0); 1094} 1095 1096static int 1097reset_sleepq_prof_stats(SYSCTL_HANDLER_ARGS) 1098{ 1099 int error, v; 1100 1101 v = 0; 1102 error = sysctl_handle_int(oidp, &v, 0, req); 1103 if (error) 1104 return (error); 1105 if (req->newptr == NULL) 1106 return (error); 1107 if (v == 0) 1108 return (0); 1109 sleepq_prof_reset(); 1110 1111 return (0); 1112} 1113 1114static int 1115dump_sleepq_prof_stats(SYSCTL_HANDLER_ARGS) 1116{ 1117 struct sleepq_prof *sp; 1118 struct sbuf *sb; 1119 int enabled; 1120 int error; 1121 int i; 1122 1123 error = sysctl_wire_old_buffer(req, 0); 1124 if (error != 0) 1125 return (error); 1126 sb = sbuf_new_for_sysctl(NULL, NULL, SLEEPQ_SBUFSIZE, req); 1127 sbuf_printf(sb, "\nwmesg\tcount\n"); 1128 enabled = prof_enabled; 1129 mtx_lock_spin(&sleepq_prof_lock); 1130 prof_enabled = 0; 1131 mtx_unlock_spin(&sleepq_prof_lock); 1132 for (i = 0; i < SC_TABLESIZE; i++) { 1133 LIST_FOREACH(sp, &sleepq_hash[i], sp_link) { 1134 sbuf_printf(sb, "%s\t%ld\n", 1135 sp->sp_wmesg, sp->sp_count); 1136 } 1137 } 1138 mtx_lock_spin(&sleepq_prof_lock); 1139 prof_enabled = enabled; 1140 mtx_unlock_spin(&sleepq_prof_lock); 1141 1142 error = sbuf_finish(sb); 1143 sbuf_delete(sb); 1144 return (error); 1145} 1146 1147SYSCTL_PROC(_debug_sleepq, OID_AUTO, stats, CTLTYPE_STRING | CTLFLAG_RD, 1148 NULL, 0, dump_sleepq_prof_stats, "A", "Sleepqueue profiling statistics"); 1149SYSCTL_PROC(_debug_sleepq, OID_AUTO, reset, CTLTYPE_INT | CTLFLAG_RW, 1150 NULL, 0, reset_sleepq_prof_stats, "I", 1151 "Reset sleepqueue profiling statistics"); 1152SYSCTL_PROC(_debug_sleepq, OID_AUTO, enable, CTLTYPE_INT | CTLFLAG_RW, 1153 NULL, 0, enable_sleepq_prof, "I", "Enable sleepqueue profiling"); 1154#endif 1155 1156#ifdef DDB 1157DB_SHOW_COMMAND(sleepq, db_show_sleepqueue) 1158{ 1159 struct sleepqueue_chain *sc; 1160 struct sleepqueue *sq; 1161#ifdef INVARIANTS 1162 struct lock_object *lock; 1163#endif 1164 struct thread *td; 1165 void *wchan; 1166 int i; 1167 1168 if (!have_addr) 1169 return; 1170 1171 /* 1172 * First, see if there is an active sleep queue for the wait channel 1173 * indicated by the address. 1174 */ 1175 wchan = (void *)addr; 1176 sc = SC_LOOKUP(wchan); 1177 LIST_FOREACH(sq, &sc->sc_queues, sq_hash) 1178 if (sq->sq_wchan == wchan) 1179 goto found; 1180 1181 /* 1182 * Second, see if there is an active sleep queue at the address 1183 * indicated. 1184 */ 1185 for (i = 0; i < SC_TABLESIZE; i++) 1186 LIST_FOREACH(sq, &sleepq_chains[i].sc_queues, sq_hash) { 1187 if (sq == (struct sleepqueue *)addr) 1188 goto found; 1189 } 1190 1191 db_printf("Unable to locate a sleep queue via %p\n", (void *)addr); 1192 return; 1193found: 1194 db_printf("Wait channel: %p\n", sq->sq_wchan); 1195 db_printf("Queue type: %d\n", sq->sq_type); 1196#ifdef INVARIANTS 1197 if (sq->sq_lock) { 1198 lock = sq->sq_lock; 1199 db_printf("Associated Interlock: %p - (%s) %s\n", lock, 1200 LOCK_CLASS(lock)->lc_name, lock->lo_name); 1201 } 1202#endif 1203 db_printf("Blocked threads:\n"); 1204 for (i = 0; i < NR_SLEEPQS; i++) { 1205 db_printf("\nQueue[%d]:\n", i); 1206 if (TAILQ_EMPTY(&sq->sq_blocked[i])) 1207 db_printf("\tempty\n"); 1208 else 1209 TAILQ_FOREACH(td, &sq->sq_blocked[0], 1210 td_slpq) { 1211 db_printf("\t%p (tid %d, pid %d, \"%s\")\n", td, 1212 td->td_tid, td->td_proc->p_pid, 1213 td->td_name); 1214 } 1215 db_printf("(expected: %u)\n", sq->sq_blockedcnt[i]); 1216 } 1217} 1218 1219/* Alias 'show sleepqueue' to 'show sleepq'. */ 1220DB_SHOW_ALIAS(sleepqueue, db_show_sleepqueue); 1221#endif 1222