subr_sleepqueue.c revision 165272
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 * 3. Neither the name of the author nor the names of any co-contributors 14 * may be used to endorse or promote products derived from this software 15 * without specific prior written permission. 16 * 17 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND 18 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 19 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 20 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE 21 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 22 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 23 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 24 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 25 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 26 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 27 * SUCH DAMAGE. 28 */ 29 30/* 31 * Implementation of sleep queues used to hold queue of threads blocked on 32 * a wait channel. Sleep queues different from turnstiles in that wait 33 * channels are not owned by anyone, so there is no priority propagation. 34 * Sleep queues can also provide a timeout and can also be interrupted by 35 * signals. That said, there are several similarities between the turnstile 36 * and sleep queue implementations. (Note: turnstiles were implemented 37 * first.) For example, both use a hash table of the same size where each 38 * bucket is referred to as a "chain" that contains both a spin lock and 39 * a linked list of queues. An individual queue is located by using a hash 40 * to pick a chain, locking the chain, and then walking the chain searching 41 * for the queue. This means that a wait channel object does not need to 42 * embed it's queue head just as locks do not embed their turnstile queue 43 * head. Threads also carry around a sleep queue that they lend to the 44 * wait channel when blocking. Just as in turnstiles, the queue includes 45 * a free list of the sleep queues of other threads blocked on the same 46 * wait channel in the case of multiple waiters. 47 * 48 * Some additional functionality provided by sleep queues include the 49 * ability to set a timeout. The timeout is managed using a per-thread 50 * callout that resumes a thread if it is asleep. A thread may also 51 * catch signals while it is asleep (aka an interruptible sleep). The 52 * signal code uses sleepq_abort() to interrupt a sleeping thread. Finally, 53 * sleep queues also provide some extra assertions. One is not allowed to 54 * mix the sleep/wakeup and cv APIs for a given wait channel. Also, one 55 * must consistently use the same lock to synchronize with a wait channel, 56 * though this check is currently only a warning for sleep/wakeup due to 57 * pre-existing abuse of that API. The same lock must also be held when 58 * awakening threads, though that is currently only enforced for condition 59 * variables. 60 */ 61 62#include <sys/cdefs.h> 63__FBSDID("$FreeBSD: head/sys/kern/subr_sleepqueue.c 165272 2006-12-16 06:54:09Z kmacy $"); 64 65#include "opt_sleepqueue_profiling.h" 66#include "opt_ddb.h" 67 68#include <sys/param.h> 69#include <sys/systm.h> 70#include <sys/lock.h> 71#include <sys/kernel.h> 72#include <sys/ktr.h> 73#include <sys/malloc.h> 74#include <sys/mutex.h> 75#include <sys/proc.h> 76#include <sys/sched.h> 77#include <sys/signalvar.h> 78#include <sys/sleepqueue.h> 79#include <sys/sysctl.h> 80 81#ifdef DDB 82#include <ddb/ddb.h> 83#endif 84 85/* 86 * Constants for the hash table of sleep queue chains. These constants are 87 * the same ones that 4BSD (and possibly earlier versions of BSD) used. 88 * Basically, we ignore the lower 8 bits of the address since most wait 89 * channel pointers are aligned and only look at the next 7 bits for the 90 * hash. SC_TABLESIZE must be a power of two for SC_MASK to work properly. 91 */ 92#define SC_TABLESIZE 128 /* Must be power of 2. */ 93#define SC_MASK (SC_TABLESIZE - 1) 94#define SC_SHIFT 8 95#define SC_HASH(wc) (((uintptr_t)(wc) >> SC_SHIFT) & 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 LIST_ENTRY(sleepqueue) sq_hash; /* (c) Chain and free list. */ 119 LIST_HEAD(, sleepqueue) sq_free; /* (c) Free queues. */ 120 void *sq_wchan; /* (c) Wait channel. */ 121#ifdef INVARIANTS 122 int sq_type; /* (c) Queue type. */ 123 struct lock_object *sq_lock; /* (c) Associated lock. */ 124#endif 125}; 126 127struct sleepqueue_chain { 128 LIST_HEAD(, sleepqueue) sc_queues; /* List of sleep queues. */ 129 struct mtx sc_lock; /* Spin lock for this chain. */ 130#ifdef SLEEPQUEUE_PROFILING 131 u_int sc_depth; /* Length of sc_queues. */ 132 u_int sc_max_depth; /* Max length of sc_queues. */ 133#endif 134}; 135 136#ifdef SLEEPQUEUE_PROFILING 137u_int sleepq_max_depth; 138SYSCTL_NODE(_debug, OID_AUTO, sleepq, CTLFLAG_RD, 0, "sleepq profiling"); 139SYSCTL_NODE(_debug_sleepq, OID_AUTO, chains, CTLFLAG_RD, 0, 140 "sleepq chain stats"); 141SYSCTL_UINT(_debug_sleepq, OID_AUTO, max_depth, CTLFLAG_RD, &sleepq_max_depth, 142 0, "maxmimum depth achieved of a single chain"); 143#endif 144static struct sleepqueue_chain sleepq_chains[SC_TABLESIZE]; 145 146static MALLOC_DEFINE(M_SLEEPQUEUE, "sleepqueue", "sleep queues"); 147 148/* 149 * Prototypes for non-exported routines. 150 */ 151static int sleepq_catch_signals(void *wchan); 152static int sleepq_check_signals(void); 153static int sleepq_check_timeout(void); 154static void sleepq_switch(void *wchan); 155static void sleepq_timeout(void *arg); 156static void sleepq_resume_thread(struct sleepqueue *sq, struct thread *td, int pri); 157 158/* 159 * Early initialization of sleep queues that is called from the sleepinit() 160 * SYSINIT. 161 */ 162void 163init_sleepqueues(void) 164{ 165#ifdef SLEEPQUEUE_PROFILING 166 struct sysctl_oid *chain_oid; 167 char chain_name[10]; 168#endif 169 int i; 170 171 for (i = 0; i < SC_TABLESIZE; i++) { 172 LIST_INIT(&sleepq_chains[i].sc_queues); 173 mtx_init(&sleepq_chains[i].sc_lock, "sleepq chain", NULL, 174 MTX_SPIN); 175#ifdef SLEEPQUEUE_PROFILING 176 snprintf(chain_name, sizeof(chain_name), "%d", i); 177 chain_oid = SYSCTL_ADD_NODE(NULL, 178 SYSCTL_STATIC_CHILDREN(_debug_sleepq_chains), OID_AUTO, 179 chain_name, CTLFLAG_RD, NULL, "sleepq chain stats"); 180 SYSCTL_ADD_UINT(NULL, SYSCTL_CHILDREN(chain_oid), OID_AUTO, 181 "depth", CTLFLAG_RD, &sleepq_chains[i].sc_depth, 0, NULL); 182 SYSCTL_ADD_UINT(NULL, SYSCTL_CHILDREN(chain_oid), OID_AUTO, 183 "max_depth", CTLFLAG_RD, &sleepq_chains[i].sc_max_depth, 0, 184 NULL); 185#endif 186 } 187 thread0.td_sleepqueue = sleepq_alloc(); 188} 189 190/* 191 * Malloc and initialize a new sleep queue for a new thread. 192 */ 193struct sleepqueue * 194sleepq_alloc(void) 195{ 196 struct sleepqueue *sq; 197 int i; 198 199 sq = malloc(sizeof(struct sleepqueue), M_SLEEPQUEUE, M_WAITOK | M_ZERO); 200 for (i = 0; i < NR_SLEEPQS; i++) 201 TAILQ_INIT(&sq->sq_blocked[i]); 202 LIST_INIT(&sq->sq_free); 203 return (sq); 204} 205 206/* 207 * Free a sleep queue when a thread is destroyed. 208 */ 209void 210sleepq_free(struct sleepqueue *sq) 211{ 212 int i; 213 214 MPASS(sq != NULL); 215 for (i = 0; i < NR_SLEEPQS; i++) 216 MPASS(TAILQ_EMPTY(&sq->sq_blocked[i])); 217 free(sq, M_SLEEPQUEUE); 218} 219 220/* 221 * Lock the sleep queue chain associated with the specified wait channel. 222 */ 223void 224sleepq_lock(void *wchan) 225{ 226 struct sleepqueue_chain *sc; 227 228 sc = SC_LOOKUP(wchan); 229 mtx_lock_spin(&sc->sc_lock); 230} 231 232/* 233 * Look up the sleep queue associated with a given wait channel in the hash 234 * table locking the associated sleep queue chain. If no queue is found in 235 * the table, NULL is returned. 236 */ 237struct sleepqueue * 238sleepq_lookup(void *wchan) 239{ 240 struct sleepqueue_chain *sc; 241 struct sleepqueue *sq; 242 243 KASSERT(wchan != NULL, ("%s: invalid NULL wait channel", __func__)); 244 sc = SC_LOOKUP(wchan); 245 mtx_assert(&sc->sc_lock, MA_OWNED); 246 LIST_FOREACH(sq, &sc->sc_queues, sq_hash) 247 if (sq->sq_wchan == wchan) 248 return (sq); 249 return (NULL); 250} 251 252/* 253 * Unlock the sleep queue chain associated with a given wait channel. 254 */ 255void 256sleepq_release(void *wchan) 257{ 258 struct sleepqueue_chain *sc; 259 260 sc = SC_LOOKUP(wchan); 261 mtx_unlock_spin(&sc->sc_lock); 262} 263 264/* 265 * Places the current thread on the sleep queue for the specified wait 266 * channel. If INVARIANTS is enabled, then it associates the passed in 267 * lock with the sleepq to make sure it is held when that sleep queue is 268 * woken up. 269 */ 270void 271sleepq_add(void *wchan, struct lock_object *lock, const char *wmesg, int flags, 272 int queue) 273{ 274 struct sleepqueue_chain *sc; 275 struct sleepqueue *sq; 276 struct thread *td; 277 278 td = curthread; 279 sc = SC_LOOKUP(wchan); 280 mtx_assert(&sc->sc_lock, MA_OWNED); 281 MPASS(td->td_sleepqueue != NULL); 282 MPASS(wchan != NULL); 283 MPASS((queue >= 0) && (queue < NR_SLEEPQS)); 284 285 /* If this thread is not allowed to sleep, die a horrible death. */ 286 KASSERT(!(td->td_pflags & TDP_NOSLEEPING), 287 ("Trying sleep, but thread marked as sleeping prohibited")); 288 289 /* Look up the sleep queue associated with the wait channel 'wchan'. */ 290 sq = sleepq_lookup(wchan); 291 292 /* 293 * If the wait channel does not already have a sleep queue, use 294 * this thread's sleep queue. Otherwise, insert the current thread 295 * into the sleep queue already in use by this wait channel. 296 */ 297 if (sq == NULL) { 298#ifdef INVARIANTS 299 int i; 300 for (i = 0; i < NR_SLEEP_QUEUEUS; i++) 301 KASSERT(TAILQ_EMPTY(&sq->sq_blocked[i]), 302 ("thread's sleep queue %d is not empty", i)); 303 304 KASSERT(LIST_EMPTY(&sq->sq_free), 305 ("thread's sleep queue has a non-empty free list")); 306 KASSERT(sq->sq_wchan == NULL, ("stale sq_wchan pointer")); 307#endif 308#ifdef SLEEPQUEUE_PROFILING 309 sc->sc_depth++; 310 if (sc->sc_depth > sc->sc_max_depth) { 311 sc->sc_max_depth = sc->sc_depth; 312 if (sc->sc_max_depth > sleepq_max_depth) 313 sleepq_max_depth = sc->sc_max_depth; 314 } 315#endif 316 sq = td->td_sleepqueue; 317 LIST_INSERT_HEAD(&sc->sc_queues, sq, sq_hash); 318 sq->sq_wchan = wchan; 319#ifdef INVARIANTS 320 sq->sq_lock = lock; 321 sq->sq_type = flags & SLEEPQ_TYPE; 322#endif 323 } else { 324 MPASS(wchan == sq->sq_wchan); 325 MPASS(lock == sq->sq_lock); 326 MPASS((flags & SLEEPQ_TYPE) == sq->sq_type); 327 LIST_INSERT_HEAD(&sq->sq_free, td->td_sleepqueue, sq_hash); 328 } 329 TAILQ_INSERT_TAIL(&sq->sq_blocked[queue], td, td_slpq); 330 td->td_sleepqueue = NULL; 331 mtx_lock_spin(&sched_lock); 332 td->td_sqqueue = queue; 333 td->td_wchan = wchan; 334 td->td_wmesg = wmesg; 335 if (flags & SLEEPQ_INTERRUPTIBLE) { 336 td->td_flags |= TDF_SINTR; 337 td->td_flags &= ~TDF_SLEEPABORT; 338 } 339 mtx_unlock_spin(&sched_lock); 340} 341 342/* 343 * Sets a timeout that will remove the current thread from the specified 344 * sleep queue after timo ticks if the thread has not already been awakened. 345 */ 346void 347sleepq_set_timeout(void *wchan, int timo) 348{ 349 struct sleepqueue_chain *sc; 350 struct thread *td; 351 352 td = curthread; 353 sc = SC_LOOKUP(wchan); 354 mtx_assert(&sc->sc_lock, MA_OWNED); 355 MPASS(TD_ON_SLEEPQ(td)); 356 MPASS(td->td_sleepqueue == NULL); 357 MPASS(wchan != NULL); 358 callout_reset(&td->td_slpcallout, timo, sleepq_timeout, td); 359} 360 361/* 362 * Marks the pending sleep of the current thread as interruptible and 363 * makes an initial check for pending signals before putting a thread 364 * to sleep. Return with sleep queue and scheduler lock held. 365 */ 366static int 367sleepq_catch_signals(void *wchan) 368{ 369 struct sleepqueue_chain *sc; 370 struct sleepqueue *sq; 371 struct thread *td; 372 struct proc *p; 373 struct sigacts *ps; 374 int sig, ret; 375 376 td = curthread; 377 p = curproc; 378 sc = SC_LOOKUP(wchan); 379 mtx_assert(&sc->sc_lock, MA_OWNED); 380 MPASS(wchan != NULL); 381 CTR3(KTR_PROC, "sleepq catching signals: thread %p (pid %ld, %s)", 382 (void *)td, (long)p->p_pid, p->p_comm); 383 384 MPASS(td->td_flags & TDF_SINTR); 385 mtx_unlock_spin(&sc->sc_lock); 386 387 /* See if there are any pending signals for this thread. */ 388 PROC_LOCK(p); 389 ps = p->p_sigacts; 390 mtx_lock(&ps->ps_mtx); 391 sig = cursig(td); 392 if (sig == 0) { 393 mtx_unlock(&ps->ps_mtx); 394 ret = thread_suspend_check(1); 395 MPASS(ret == 0 || ret == EINTR || ret == ERESTART); 396 } else { 397 if (SIGISMEMBER(ps->ps_sigintr, sig)) 398 ret = EINTR; 399 else 400 ret = ERESTART; 401 mtx_unlock(&ps->ps_mtx); 402 } 403 404 if (ret == 0) { 405 mtx_lock_spin(&sc->sc_lock); 406 /* 407 * Lock sched_lock before unlocking proc lock, 408 * without this, we could lose a race. 409 */ 410 mtx_lock_spin(&sched_lock); 411 PROC_UNLOCK(p); 412 if (!(td->td_flags & TDF_INTERRUPT)) 413 return (0); 414 /* KSE threads tried unblocking us. */ 415 ret = td->td_intrval; 416 mtx_unlock_spin(&sched_lock); 417 MPASS(ret == EINTR || ret == ERESTART); 418 } else { 419 PROC_UNLOCK(p); 420 mtx_lock_spin(&sc->sc_lock); 421 } 422 /* 423 * There were pending signals and this thread is still 424 * on the sleep queue, remove it from the sleep queue. 425 */ 426 sq = sleepq_lookup(wchan); 427 mtx_lock_spin(&sched_lock); 428 if (TD_ON_SLEEPQ(td)) 429 sleepq_resume_thread(sq, td, -1); 430 return (ret); 431} 432 433/* 434 * Switches to another thread if we are still asleep on a sleep queue and 435 * drop the lock on the sleep queue chain. Returns with sched_lock held. 436 */ 437static void 438sleepq_switch(void *wchan) 439{ 440 struct sleepqueue_chain *sc; 441 struct thread *td; 442 443 td = curthread; 444 sc = SC_LOOKUP(wchan); 445 mtx_assert(&sc->sc_lock, MA_OWNED); 446 mtx_assert(&sched_lock, MA_OWNED); 447 448 /* 449 * If we have a sleep queue, then we've already been woken up, so 450 * just return. 451 */ 452 if (td->td_sleepqueue != NULL) { 453 MPASS(!TD_ON_SLEEPQ(td)); 454 mtx_unlock_spin(&sc->sc_lock); 455 return; 456 } 457 458 /* 459 * Otherwise, actually go to sleep. 460 */ 461 mtx_unlock_spin(&sc->sc_lock); 462 sched_sleep(td); 463 TD_SET_SLEEPING(td); 464 mi_switch(SW_VOL, NULL); 465 KASSERT(TD_IS_RUNNING(td), ("running but not TDS_RUNNING")); 466 CTR3(KTR_PROC, "sleepq resume: thread %p (pid %ld, %s)", 467 (void *)td, (long)td->td_proc->p_pid, (void *)td->td_proc->p_comm); 468} 469 470/* 471 * Check to see if we timed out. 472 */ 473static int 474sleepq_check_timeout(void) 475{ 476 struct thread *td; 477 478 mtx_assert(&sched_lock, MA_OWNED); 479 td = curthread; 480 481 /* 482 * If TDF_TIMEOUT is set, we timed out. 483 */ 484 if (td->td_flags & TDF_TIMEOUT) { 485 td->td_flags &= ~TDF_TIMEOUT; 486 return (EWOULDBLOCK); 487 } 488 489 /* 490 * If TDF_TIMOFAIL is set, the timeout ran after we had 491 * already been woken up. 492 */ 493 if (td->td_flags & TDF_TIMOFAIL) 494 td->td_flags &= ~TDF_TIMOFAIL; 495 496 /* 497 * If callout_stop() fails, then the timeout is running on 498 * another CPU, so synchronize with it to avoid having it 499 * accidentally wake up a subsequent sleep. 500 */ 501 else if (callout_stop(&td->td_slpcallout) == 0) { 502 td->td_flags |= TDF_TIMEOUT; 503 TD_SET_SLEEPING(td); 504 mi_switch(SW_INVOL, NULL); 505 } 506 return (0); 507} 508 509/* 510 * Check to see if we were awoken by a signal. 511 */ 512static int 513sleepq_check_signals(void) 514{ 515 struct thread *td; 516 517 mtx_assert(&sched_lock, MA_OWNED); 518 td = curthread; 519 520 /* We are no longer in an interruptible sleep. */ 521 if (td->td_flags & TDF_SINTR) 522 td->td_flags &= ~TDF_SINTR; 523 524 if (td->td_flags & TDF_SLEEPABORT) { 525 td->td_flags &= ~TDF_SLEEPABORT; 526 return (td->td_intrval); 527 } 528 529 if (td->td_flags & TDF_INTERRUPT) 530 return (td->td_intrval); 531 532 return (0); 533} 534 535/* 536 * Block the current thread until it is awakened from its sleep queue. 537 */ 538void 539sleepq_wait(void *wchan) 540{ 541 542 MPASS(!(curthread->td_flags & TDF_SINTR)); 543 mtx_lock_spin(&sched_lock); 544 sleepq_switch(wchan); 545 mtx_unlock_spin(&sched_lock); 546} 547 548/* 549 * Block the current thread until it is awakened from its sleep queue 550 * or it is interrupted by a signal. 551 */ 552int 553sleepq_wait_sig(void *wchan) 554{ 555 int rcatch; 556 int rval; 557 558 rcatch = sleepq_catch_signals(wchan); 559 if (rcatch == 0) 560 sleepq_switch(wchan); 561 else 562 sleepq_release(wchan); 563 rval = sleepq_check_signals(); 564 mtx_unlock_spin(&sched_lock); 565 if (rcatch) 566 return (rcatch); 567 return (rval); 568} 569 570/* 571 * Block the current thread until it is awakened from its sleep queue 572 * or it times out while waiting. 573 */ 574int 575sleepq_timedwait(void *wchan) 576{ 577 int rval; 578 579 MPASS(!(curthread->td_flags & TDF_SINTR)); 580 mtx_lock_spin(&sched_lock); 581 sleepq_switch(wchan); 582 rval = sleepq_check_timeout(); 583 mtx_unlock_spin(&sched_lock); 584 return (rval); 585} 586 587/* 588 * Block the current thread until it is awakened from its sleep queue, 589 * it is interrupted by a signal, or it times out waiting to be awakened. 590 */ 591int 592sleepq_timedwait_sig(void *wchan) 593{ 594 int rcatch, rvalt, rvals; 595 596 rcatch = sleepq_catch_signals(wchan); 597 if (rcatch == 0) 598 sleepq_switch(wchan); 599 else 600 sleepq_release(wchan); 601 rvalt = sleepq_check_timeout(); 602 rvals = sleepq_check_signals(); 603 mtx_unlock_spin(&sched_lock); 604 if (rcatch) 605 return (rcatch); 606 if (rvals) 607 return (rvals); 608 return (rvalt); 609} 610 611/* 612 * Removes a thread from a sleep queue and makes it 613 * runnable. 614 */ 615static void 616sleepq_resume_thread(struct sleepqueue *sq, struct thread *td, int pri) 617{ 618 struct sleepqueue_chain *sc; 619 620 MPASS(td != NULL); 621 MPASS(sq->sq_wchan != NULL); 622 MPASS(td->td_wchan == sq->sq_wchan); 623 MPASS(td->td_sqqueue < NR_SLEEPQS && td->td_sqqueue >= 0); 624 sc = SC_LOOKUP(sq->sq_wchan); 625 mtx_assert(&sc->sc_lock, MA_OWNED); 626 mtx_assert(&sched_lock, MA_OWNED); 627 628 /* Remove the thread from the queue. */ 629 TAILQ_REMOVE(&sq->sq_blocked[td->td_sqqueue], td, td_slpq); 630 631 /* 632 * Get a sleep queue for this thread. If this is the last waiter, 633 * use the queue itself and take it out of the chain, otherwise, 634 * remove a queue from the free list. 635 */ 636 if (LIST_EMPTY(&sq->sq_free)) { 637 td->td_sleepqueue = sq; 638#ifdef INVARIANTS 639 sq->sq_wchan = NULL; 640#endif 641#ifdef SLEEPQUEUE_PROFILING 642 sc->sc_depth--; 643#endif 644 } else 645 td->td_sleepqueue = LIST_FIRST(&sq->sq_free); 646 LIST_REMOVE(td->td_sleepqueue, sq_hash); 647 648 td->td_wmesg = NULL; 649 td->td_wchan = NULL; 650 td->td_flags &= ~TDF_SINTR; 651 652 /* 653 * Note that thread td might not be sleeping if it is running 654 * sleepq_catch_signals() on another CPU or is blocked on 655 * its proc lock to check signals. It doesn't hurt to clear 656 * the sleeping flag if it isn't set though, so we just always 657 * do it. However, we can't assert that it is set. 658 */ 659 CTR3(KTR_PROC, "sleepq_wakeup: thread %p (pid %ld, %s)", 660 (void *)td, (long)td->td_proc->p_pid, td->td_proc->p_comm); 661 TD_CLR_SLEEPING(td); 662 663 /* Adjust priority if requested. */ 664 MPASS(pri == -1 || (pri >= PRI_MIN && pri <= PRI_MAX)); 665 if (pri != -1 && td->td_priority > pri) 666 sched_prio(td, pri); 667 setrunnable(td); 668} 669 670/* 671 * Find the highest priority thread sleeping on a wait channel and resume it. 672 */ 673void 674sleepq_signal(void *wchan, int flags, int pri, int queue) 675{ 676 struct sleepqueue *sq; 677 struct thread *td, *besttd; 678 679 CTR2(KTR_PROC, "sleepq_signal(%p, %d)", wchan, flags); 680 KASSERT(wchan != NULL, ("%s: invalid NULL wait channel", __func__)); 681 MPASS((queue >= 0) && (queue < NR_SLEEPQS)); 682 sq = sleepq_lookup(wchan); 683 if (sq == NULL) { 684 sleepq_release(wchan); 685 return; 686 } 687 KASSERT(sq->sq_type == (flags & SLEEPQ_TYPE), 688 ("%s: mismatch between sleep/wakeup and cv_*", __func__)); 689 690 /* 691 * Find the highest priority thread on the queue. If there is a 692 * tie, use the thread that first appears in the queue as it has 693 * been sleeping the longest since threads are always added to 694 * the tail of sleep queues. 695 */ 696 besttd = NULL; 697 TAILQ_FOREACH(td, &sq->sq_blocked[queue], td_slpq) { 698 if (besttd == NULL || td->td_priority < besttd->td_priority) 699 besttd = td; 700 } 701 MPASS(besttd != NULL); 702 mtx_lock_spin(&sched_lock); 703 sleepq_resume_thread(sq, besttd, pri); 704 mtx_unlock_spin(&sched_lock); 705 sleepq_release(wchan); 706} 707 708/* 709 * Resume all threads sleeping on a specified wait channel. 710 */ 711void 712sleepq_broadcast(void *wchan, int flags, int pri, int queue) 713{ 714 struct sleepqueue *sq; 715 716 CTR2(KTR_PROC, "sleepq_broadcast(%p, %d)", wchan, flags); 717 KASSERT(wchan != NULL, ("%s: invalid NULL wait channel", __func__)); 718 MPASS((queue >= 0) && (queue < NR_SLEEPQS)); 719 sq = sleepq_lookup(wchan); 720 if (sq == NULL) { 721 sleepq_release(wchan); 722 return; 723 } 724 KASSERT(sq->sq_type == (flags & SLEEPQ_TYPE), 725 ("%s: mismatch between sleep/wakeup and cv_*", __func__)); 726 727 /* Resume all blocked threads on the sleep queue. */ 728 mtx_lock_spin(&sched_lock); 729 while (!TAILQ_EMPTY(&sq->sq_blocked[queue])) 730 sleepq_resume_thread(sq, TAILQ_FIRST(&sq->sq_blocked[queue]), 731 pri); 732 mtx_unlock_spin(&sched_lock); 733 sleepq_release(wchan); 734} 735 736/* 737 * Time sleeping threads out. When the timeout expires, the thread is 738 * removed from the sleep queue and made runnable if it is still asleep. 739 */ 740static void 741sleepq_timeout(void *arg) 742{ 743 struct sleepqueue *sq; 744 struct thread *td; 745 void *wchan; 746 747 td = arg; 748 CTR3(KTR_PROC, "sleepq_timeout: thread %p (pid %ld, %s)", 749 (void *)td, (long)td->td_proc->p_pid, (void *)td->td_proc->p_comm); 750 751 /* 752 * First, see if the thread is asleep and get the wait channel if 753 * it is. 754 */ 755 mtx_lock_spin(&sched_lock); 756 if (TD_ON_SLEEPQ(td)) { 757 wchan = td->td_wchan; 758 mtx_unlock_spin(&sched_lock); 759 sleepq_lock(wchan); 760 sq = sleepq_lookup(wchan); 761 mtx_lock_spin(&sched_lock); 762 } else { 763 wchan = NULL; 764 sq = NULL; 765 } 766 767 /* 768 * At this point, if the thread is still on the sleep queue, 769 * we have that sleep queue locked as it cannot migrate sleep 770 * queues while we dropped sched_lock. If it had resumed and 771 * was on another CPU while the lock was dropped, it would have 772 * seen that TDF_TIMEOUT and TDF_TIMOFAIL are clear and the 773 * call to callout_stop() to stop this routine would have failed 774 * meaning that it would have already set TDF_TIMEOUT to 775 * synchronize with this function. 776 */ 777 if (TD_ON_SLEEPQ(td)) { 778 MPASS(td->td_wchan == wchan); 779 MPASS(sq != NULL); 780 td->td_flags |= TDF_TIMEOUT; 781 sleepq_resume_thread(sq, td, -1); 782 mtx_unlock_spin(&sched_lock); 783 sleepq_release(wchan); 784 return; 785 } else if (wchan != NULL) 786 sleepq_release(wchan); 787 788 /* 789 * Now check for the edge cases. First, if TDF_TIMEOUT is set, 790 * then the other thread has already yielded to us, so clear 791 * the flag and resume it. If TDF_TIMEOUT is not set, then the 792 * we know that the other thread is not on a sleep queue, but it 793 * hasn't resumed execution yet. In that case, set TDF_TIMOFAIL 794 * to let it know that the timeout has already run and doesn't 795 * need to be canceled. 796 */ 797 if (td->td_flags & TDF_TIMEOUT) { 798 MPASS(TD_IS_SLEEPING(td)); 799 td->td_flags &= ~TDF_TIMEOUT; 800 TD_CLR_SLEEPING(td); 801 setrunnable(td); 802 } else 803 td->td_flags |= TDF_TIMOFAIL; 804 mtx_unlock_spin(&sched_lock); 805} 806 807/* 808 * Resumes a specific thread from the sleep queue associated with a specific 809 * wait channel if it is on that queue. 810 */ 811void 812sleepq_remove(struct thread *td, void *wchan) 813{ 814 struct sleepqueue *sq; 815 816 /* 817 * Look up the sleep queue for this wait channel, then re-check 818 * that the thread is asleep on that channel, if it is not, then 819 * bail. 820 */ 821 MPASS(wchan != NULL); 822 sleepq_lock(wchan); 823 sq = sleepq_lookup(wchan); 824 mtx_lock_spin(&sched_lock); 825 if (!TD_ON_SLEEPQ(td) || td->td_wchan != wchan) { 826 mtx_unlock_spin(&sched_lock); 827 sleepq_release(wchan); 828 return; 829 } 830 MPASS(sq != NULL); 831 832 /* Thread is asleep on sleep queue sq, so wake it up. */ 833 sleepq_resume_thread(sq, td, -1); 834 sleepq_release(wchan); 835 mtx_unlock_spin(&sched_lock); 836} 837 838/* 839 * Abort a thread as if an interrupt had occurred. Only abort 840 * interruptible waits (unfortunately it isn't safe to abort others). 841 * 842 * XXX: What in the world does the comment below mean? 843 * Also, whatever the signal code does... 844 */ 845void 846sleepq_abort(struct thread *td, int intrval) 847{ 848 void *wchan; 849 850 mtx_assert(&sched_lock, MA_OWNED); 851 MPASS(TD_ON_SLEEPQ(td)); 852 MPASS(td->td_flags & TDF_SINTR); 853 MPASS(intrval == EINTR || intrval == ERESTART); 854 855 /* 856 * If the TDF_TIMEOUT flag is set, just leave. A 857 * timeout is scheduled anyhow. 858 */ 859 if (td->td_flags & TDF_TIMEOUT) 860 return; 861 862 CTR3(KTR_PROC, "sleepq_abort: thread %p (pid %ld, %s)", 863 (void *)td, (long)td->td_proc->p_pid, (void *)td->td_proc->p_comm); 864 wchan = td->td_wchan; 865 if (wchan != NULL) { 866 td->td_intrval = intrval; 867 td->td_flags |= TDF_SLEEPABORT; 868 } 869 mtx_unlock_spin(&sched_lock); 870 sleepq_remove(td, wchan); 871 mtx_lock_spin(&sched_lock); 872} 873 874#ifdef DDB 875DB_SHOW_COMMAND(sleepq, db_show_sleepqueue) 876{ 877 struct sleepqueue_chain *sc; 878 struct sleepqueue *sq; 879#ifdef INVARIANTS 880 struct lock_object *lock; 881#endif 882 struct thread *td; 883 void *wchan; 884 int i; 885 886 if (!have_addr) 887 return; 888 889 /* 890 * First, see if there is an active sleep queue for the wait channel 891 * indicated by the address. 892 */ 893 wchan = (void *)addr; 894 sc = SC_LOOKUP(wchan); 895 LIST_FOREACH(sq, &sc->sc_queues, sq_hash) 896 if (sq->sq_wchan == wchan) 897 goto found; 898 899 /* 900 * Second, see if there is an active sleep queue at the address 901 * indicated. 902 */ 903 for (i = 0; i < SC_TABLESIZE; i++) 904 LIST_FOREACH(sq, &sleepq_chains[i].sc_queues, sq_hash) { 905 if (sq == (struct sleepqueue *)addr) 906 goto found; 907 } 908 909 db_printf("Unable to locate a sleep queue via %p\n", (void *)addr); 910 return; 911found: 912 db_printf("Wait channel: %p\n", sq->sq_wchan); 913#ifdef INVARIANTS 914 db_printf("Queue type: %d\n", sq->sq_type); 915 if (sq->sq_lock) { 916 lock = sq->sq_lock; 917 db_printf("Associated Interlock: %p - (%s) %s\n", lock, 918 LOCK_CLASS(lock)->lc_name, lock->lo_name); 919 } 920#endif 921 db_printf("Blocked threads:\n"); 922 for (i = 0; i < NR_SLEEPQS; i++) { 923 db_printf("\nQueue[%d]:\n", i); 924 if (TAILQ_EMPTY(&sq->sq_blocked[i])) 925 db_printf("\tempty\n"); 926 else 927 TAILQ_FOREACH(td, &sq->sq_blocked[0], 928 td_slpq) { 929 db_printf("\t%p (tid %d, pid %d, \"%s\")\n", td, 930 td->td_tid, td->td_proc->p_pid, 931 td->td_name[i] != '\0' ? td->td_name : 932 td->td_proc->p_comm); 933 } 934 } 935} 936 937/* Alias 'show sleepqueue' to 'show sleepq'. */ 938DB_SET(sleepqueue, db_show_sleepqueue, db_show_cmd_set, 0, NULL); 939#endif 940