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("$FreeBSD$"); 61 62 #include "opt_sleepqueue_profiling.h" 63 #include "opt_ddb.h" 64 #include "opt_sched.h" 65 66 #include <sys/param.h> 67 #include <sys/systm.h> 68 #include <sys/lock.h> 69 #include <sys/kernel.h> 70 #include <sys/ktr.h> 71 #include <sys/mutex.h> 72 #include <sys/proc.h> 73 #include <sys/sbuf.h> 74 #include <sys/sched.h> 75 #include <sys/sdt.h> 76 #include <sys/signalvar.h> 77 #include <sys/sleepqueue.h> 78 #include <sys/sysctl.h> 79 80 #include <vm/uma.h> 81 82 #ifdef DDB 83 #include <ddb/ddb.h> 84 #endif 85 86 /* 87 * Constants for the hash table of sleep queue chains. 88 * SC_TABLESIZE must be a power of two for SC_MASK to work properly. 89 */ 90 #define SC_TABLESIZE 256 /* Must be power of 2. */ 91 #define SC_MASK (SC_TABLESIZE - 1) 92 #define SC_SHIFT 8 93 #define SC_HASH(wc) ((((uintptr_t)(wc) >> SC_SHIFT) ^ (uintptr_t)(wc)) & \ 94 SC_MASK) 95 #define SC_LOOKUP(wc) &sleepq_chains[SC_HASH(wc)] 96 #define NR_SLEEPQS 2 97 /* 98 * There two different lists of sleep queues. Both lists are connected 99 * via the sq_hash entries. The first list is the sleep queue chain list 100 * that a sleep queue is on when it is attached to a wait channel. The 101 * second list is the free list hung off of a sleep queue that is attached 102 * to a wait channel. 103 * 104 * Each sleep queue also contains the wait channel it is attached to, the 105 * list of threads blocked on that wait channel, flags specific to the 106 * wait channel, and the lock used to synchronize with a wait channel. 107 * The flags are used to catch mismatches between the various consumers 108 * of the sleep queue API (e.g. sleep/wakeup and condition variables). 109 * The lock pointer is only used when invariants are enabled for various 110 * debugging checks. 111 * 112 * Locking key: 113 * c - sleep queue chain lock 114 */ 115 struct sleepqueue { 116 TAILQ_HEAD(, thread) sq_blocked[NR_SLEEPQS]; /* (c) Blocked threads. */ 117 u_int sq_blockedcnt[NR_SLEEPQS]; /* (c) N. of 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 int sq_type; /* (c) Queue type. */ 122 #ifdef INVARIANTS 123 struct lock_object *sq_lock; /* (c) Associated lock. */ 124 #endif 125 }; 126 127 struct 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 137 u_int sleepq_max_depth; 138 static SYSCTL_NODE(_debug, OID_AUTO, sleepq, CTLFLAG_RD, 0, "sleepq profiling"); 139 static SYSCTL_NODE(_debug_sleepq, OID_AUTO, chains, CTLFLAG_RD, 0, 140 "sleepq chain stats"); 141 SYSCTL_UINT(_debug_sleepq, OID_AUTO, max_depth, CTLFLAG_RD, &sleepq_max_depth, 142 0, "maxmimum depth achieved of a single chain"); 143 144 static void sleepq_profile(const char *wmesg); 145 static int prof_enabled; 146 #endif 147 static struct sleepqueue_chain sleepq_chains[SC_TABLESIZE]; 148 static uma_zone_t sleepq_zone; 149 150 /* 151 * Prototypes for non-exported routines. 152 */ 153 static int sleepq_catch_signals(void *wchan, int pri); 154 static int sleepq_check_signals(void); 155 static int sleepq_check_timeout(void); 156 #ifdef INVARIANTS 157 static void sleepq_dtor(void *mem, int size, void *arg); 158 #endif 159 static int sleepq_init(void *mem, int size, int flags); 160 static int sleepq_resume_thread(struct sleepqueue *sq, struct thread *td, 161 int pri); 162 static void sleepq_switch(void *wchan, int pri); 163 static void sleepq_timeout(void *arg); 164 165 SDT_PROBE_DECLARE(sched, , , sleep); 166 SDT_PROBE_DECLARE(sched, , , wakeup); 167 168 /* 169 * Early initialization of sleep queues that is called from the sleepinit() 170 * SYSINIT. 171 */ 172 void 173 init_sleepqueues(void) 174 { 175 #ifdef SLEEPQUEUE_PROFILING 176 struct sysctl_oid *chain_oid; 177 char chain_name[10]; 178 #endif 179 int i; 180 181 for (i = 0; i < SC_TABLESIZE; i++) { 182 LIST_INIT(&sleepq_chains[i].sc_queues); 183 mtx_init(&sleepq_chains[i].sc_lock, "sleepq chain", NULL, 184 MTX_SPIN | MTX_RECURSE); 185 #ifdef SLEEPQUEUE_PROFILING 186 snprintf(chain_name, sizeof(chain_name), "%d", i); 187 chain_oid = SYSCTL_ADD_NODE(NULL, 188 SYSCTL_STATIC_CHILDREN(_debug_sleepq_chains), OID_AUTO, 189 chain_name, CTLFLAG_RD, NULL, "sleepq chain stats"); 190 SYSCTL_ADD_UINT(NULL, SYSCTL_CHILDREN(chain_oid), OID_AUTO, 191 "depth", CTLFLAG_RD, &sleepq_chains[i].sc_depth, 0, NULL); 192 SYSCTL_ADD_UINT(NULL, SYSCTL_CHILDREN(chain_oid), OID_AUTO, 193 "max_depth", CTLFLAG_RD, &sleepq_chains[i].sc_max_depth, 0, 194 NULL); 195 #endif 196 } 197 sleepq_zone = uma_zcreate("SLEEPQUEUE", sizeof(struct sleepqueue), 198 #ifdef INVARIANTS 199 NULL, sleepq_dtor, sleepq_init, NULL, UMA_ALIGN_CACHE, 0); 200 #else 201 NULL, NULL, sleepq_init, NULL, UMA_ALIGN_CACHE, 0); 202 #endif 203 204 thread0.td_sleepqueue = sleepq_alloc(); 205 } 206 207 /* 208 * Get a sleep queue for a new thread. 209 */ 210 struct sleepqueue * 211 sleepq_alloc(void) 212 { 213 214 return (uma_zalloc(sleepq_zone, M_WAITOK)); 215 } 216 217 /* 218 * Free a sleep queue when a thread is destroyed. 219 */ 220 void 221 sleepq_free(struct sleepqueue *sq) 222 { 223 224 uma_zfree(sleepq_zone, sq); 225 } 226 227 /* 228 * Lock the sleep queue chain associated with the specified wait channel. 229 */ 230 void 231 sleepq_lock(void *wchan) 232 { 233 struct sleepqueue_chain *sc; 234 235 sc = SC_LOOKUP(wchan); 236 mtx_lock_spin(&sc->sc_lock); 237 } 238 239 /* 240 * Look up the sleep queue associated with a given wait channel in the hash 241 * table locking the associated sleep queue chain. If no queue is found in 242 * the table, NULL is returned. 243 */ 244 struct sleepqueue * 245 sleepq_lookup(void *wchan) 246 { 247 struct sleepqueue_chain *sc; 248 struct sleepqueue *sq; 249 250 KASSERT(wchan != NULL, ("%s: invalid NULL wait channel", __func__)); 251 sc = SC_LOOKUP(wchan); 252 mtx_assert(&sc->sc_lock, MA_OWNED); 253 LIST_FOREACH(sq, &sc->sc_queues, sq_hash) 254 if (sq->sq_wchan == wchan) 255 return (sq); 256 return (NULL); 257 } 258 259 /* 260 * Unlock the sleep queue chain associated with a given wait channel. 261 */ 262 void 263 sleepq_release(void *wchan) 264 { 265 struct sleepqueue_chain *sc; 266 267 sc = SC_LOOKUP(wchan); 268 mtx_unlock_spin(&sc->sc_lock); 269 } 270 271 /* 272 * Places the current thread on the sleep queue for the specified wait 273 * channel. If INVARIANTS is enabled, then it associates the passed in 274 * lock with the sleepq to make sure it is held when that sleep queue is 275 * woken up. 276 */ 277 void 278 sleepq_add(void *wchan, struct lock_object *lock, const char *wmesg, int flags, 279 int queue) 280 { 281 struct sleepqueue_chain *sc; 282 struct sleepqueue *sq; 283 struct thread *td; 284 285 td = curthread; 286 sc = SC_LOOKUP(wchan); 287 mtx_assert(&sc->sc_lock, MA_OWNED); 288 MPASS(td->td_sleepqueue != NULL); 289 MPASS(wchan != NULL); 290 MPASS((queue >= 0) && (queue < NR_SLEEPQS)); 291 292 /* If this thread is not allowed to sleep, die a horrible death. */ 293 KASSERT(td->td_no_sleeping == 0, 294 ("%s: td %p to sleep on wchan %p with sleeping prohibited", 295 __func__, td, wchan)); 296 297 /* Look up the sleep queue associated with the wait channel 'wchan'. */ 298 sq = sleepq_lookup(wchan); 299 300 /* 301 * If the wait channel does not already have a sleep queue, use 302 * this thread's sleep queue. Otherwise, insert the current thread 303 * into the sleep queue already in use by this wait channel. 304 */ 305 if (sq == NULL) { 306 #ifdef INVARIANTS 307 int i; 308 309 sq = td->td_sleepqueue; 310 for (i = 0; i < NR_SLEEPQS; i++) { 311 KASSERT(TAILQ_EMPTY(&sq->sq_blocked[i]), 312 ("thread's sleep queue %d is not empty", i)); 313 KASSERT(sq->sq_blockedcnt[i] == 0, 314 ("thread's sleep queue %d count mismatches", i)); 315 } 316 KASSERT(LIST_EMPTY(&sq->sq_free), 317 ("thread's sleep queue has a non-empty free list")); 318 KASSERT(sq->sq_wchan == NULL, ("stale sq_wchan pointer")); 319 sq->sq_lock = lock; 320 #endif 321 #ifdef SLEEPQUEUE_PROFILING 322 sc->sc_depth++; 323 if (sc->sc_depth > sc->sc_max_depth) { 324 sc->sc_max_depth = sc->sc_depth; 325 if (sc->sc_max_depth > sleepq_max_depth) 326 sleepq_max_depth = sc->sc_max_depth; 327 } 328 #endif 329 sq = td->td_sleepqueue; 330 LIST_INSERT_HEAD(&sc->sc_queues, sq, sq_hash); 331 sq->sq_wchan = wchan; 332 sq->sq_type = flags & SLEEPQ_TYPE; 333 } else { 334 MPASS(wchan == sq->sq_wchan); 335 MPASS(lock == sq->sq_lock); 336 MPASS((flags & SLEEPQ_TYPE) == sq->sq_type); 337 LIST_INSERT_HEAD(&sq->sq_free, td->td_sleepqueue, sq_hash); 338 } 339 thread_lock(td); 340 TAILQ_INSERT_TAIL(&sq->sq_blocked[queue], td, td_slpq); 341 sq->sq_blockedcnt[queue]++; 342 td->td_sleepqueue = NULL; 343 td->td_sqqueue = queue; 344 td->td_wchan = wchan; 345 td->td_wmesg = wmesg; 346 if (flags & SLEEPQ_INTERRUPTIBLE) { 347 td->td_flags |= TDF_SINTR; 348 td->td_flags &= ~TDF_SLEEPABORT; 349 } 350 thread_unlock(td); 351 } 352 353 /* 354 * Sets a timeout that will remove the current thread from the specified 355 * sleep queue after timo ticks if the thread has not already been awakened. 356 */ 357 void 358 sleepq_set_timeout_sbt(void *wchan, sbintime_t sbt, sbintime_t pr, 359 int flags) 360 { 361 struct sleepqueue_chain *sc; 362 struct thread *td; 363 364 td = curthread; 365 sc = SC_LOOKUP(wchan); 366 mtx_assert(&sc->sc_lock, MA_OWNED); 367 MPASS(TD_ON_SLEEPQ(td)); 368 MPASS(td->td_sleepqueue == NULL); 369 MPASS(wchan != NULL); 370 callout_reset_sbt_on(&td->td_slpcallout, sbt, pr, 371 sleepq_timeout, td, PCPU_GET(cpuid), flags | C_DIRECT_EXEC); 372 } 373 374 /* 375 * Return the number of actual sleepers for the specified queue. 376 */ 377 u_int 378 sleepq_sleepcnt(void *wchan, int queue) 379 { 380 struct sleepqueue *sq; 381 382 KASSERT(wchan != NULL, ("%s: invalid NULL wait channel", __func__)); 383 MPASS((queue >= 0) && (queue < NR_SLEEPQS)); 384 sq = sleepq_lookup(wchan); 385 if (sq == NULL) 386 return (0); 387 return (sq->sq_blockedcnt[queue]); 388 } 389 390 /* 391 * Marks the pending sleep of the current thread as interruptible and 392 * makes an initial check for pending signals before putting a thread 393 * to sleep. Enters and exits with the thread lock held. Thread lock 394 * may have transitioned from the sleepq lock to a run lock. 395 */ 396 static int 397 sleepq_catch_signals(void *wchan, int pri) 398 { 399 struct sleepqueue_chain *sc; 400 struct sleepqueue *sq; 401 struct thread *td; 402 struct proc *p; 403 struct sigacts *ps; 404 int sig, ret; 405 406 td = curthread; 407 p = curproc; 408 sc = SC_LOOKUP(wchan); 409 mtx_assert(&sc->sc_lock, MA_OWNED); 410 MPASS(wchan != NULL); 411 if ((td->td_pflags & TDP_WAKEUP) != 0) { 412 td->td_pflags &= ~TDP_WAKEUP; 413 ret = EINTR; 414 thread_lock(td); 415 goto out; 416 } 417 418 /* 419 * See if there are any pending signals for this thread. If not 420 * we can switch immediately. Otherwise do the signal processing 421 * directly. 422 */ 423 thread_lock(td); 424 if ((td->td_flags & (TDF_NEEDSIGCHK | TDF_NEEDSUSPCHK)) == 0) { 425 sleepq_switch(wchan, pri); 426 return (0); 427 } 428 thread_unlock(td); 429 mtx_unlock_spin(&sc->sc_lock); 430 CTR3(KTR_PROC, "sleepq catching signals: thread %p (pid %ld, %s)", 431 (void *)td, (long)p->p_pid, td->td_name); 432 PROC_LOCK(p); 433 ps = p->p_sigacts; 434 mtx_lock(&ps->ps_mtx); 435 sig = cursig(td); 436 if (sig == 0) { 437 mtx_unlock(&ps->ps_mtx); 438 ret = thread_suspend_check(1); 439 MPASS(ret == 0 || ret == EINTR || ret == ERESTART); 440 } else { 441 if (SIGISMEMBER(ps->ps_sigintr, sig)) 442 ret = EINTR; 443 else 444 ret = ERESTART; 445 mtx_unlock(&ps->ps_mtx); 446 } 447 /* 448 * Lock the per-process spinlock prior to dropping the PROC_LOCK 449 * to avoid a signal delivery race. PROC_LOCK, PROC_SLOCK, and 450 * thread_lock() are currently held in tdsendsignal(). 451 */ 452 PROC_SLOCK(p); 453 mtx_lock_spin(&sc->sc_lock); 454 PROC_UNLOCK(p); 455 thread_lock(td); 456 PROC_SUNLOCK(p); 457 if (ret == 0) { 458 sleepq_switch(wchan, pri); 459 return (0); 460 } 461 out: 462 /* 463 * There were pending signals and this thread is still 464 * on the sleep queue, remove it from the sleep queue. 465 */ 466 if (TD_ON_SLEEPQ(td)) { 467 sq = sleepq_lookup(wchan); 468 if (sleepq_resume_thread(sq, td, 0)) { 469 #ifdef INVARIANTS 470 /* 471 * This thread hasn't gone to sleep yet, so it 472 * should not be swapped out. 473 */ 474 panic("not waking up swapper"); 475 #endif 476 } 477 } 478 mtx_unlock_spin(&sc->sc_lock); 479 MPASS(td->td_lock != &sc->sc_lock); 480 return (ret); 481 } 482 483 /* 484 * Switches to another thread if we are still asleep on a sleep queue. 485 * Returns with thread lock. 486 */ 487 static void 488 sleepq_switch(void *wchan, int pri) 489 { 490 struct sleepqueue_chain *sc; 491 struct sleepqueue *sq; 492 struct thread *td; 493 494 td = curthread; 495 sc = SC_LOOKUP(wchan); 496 mtx_assert(&sc->sc_lock, MA_OWNED); 497 THREAD_LOCK_ASSERT(td, MA_OWNED); 498 499 /* 500 * If we have a sleep queue, then we've already been woken up, so 501 * just return. 502 */ 503 if (td->td_sleepqueue != NULL) { 504 mtx_unlock_spin(&sc->sc_lock); 505 return; 506 } 507 508 /* 509 * If TDF_TIMEOUT is set, then our sleep has been timed out 510 * already but we are still on the sleep queue, so dequeue the 511 * thread and return. 512 */ 513 if (td->td_flags & TDF_TIMEOUT) { 514 MPASS(TD_ON_SLEEPQ(td)); 515 sq = sleepq_lookup(wchan); 516 if (sleepq_resume_thread(sq, td, 0)) { 517 #ifdef INVARIANTS 518 /* 519 * This thread hasn't gone to sleep yet, so it 520 * should not be swapped out. 521 */ 522 panic("not waking up swapper"); 523 #endif 524 } 525 mtx_unlock_spin(&sc->sc_lock); 526 return; 527 } 528 #ifdef SLEEPQUEUE_PROFILING 529 if (prof_enabled) 530 sleepq_profile(td->td_wmesg); 531 #endif 532 MPASS(td->td_sleepqueue == NULL); 533 sched_sleep(td, pri); 534 thread_lock_set(td, &sc->sc_lock); 535 SDT_PROBE0(sched, , , sleep); 536 TD_SET_SLEEPING(td); 537 mi_switch(SW_VOL | SWT_SLEEPQ, NULL); 538 KASSERT(TD_IS_RUNNING(td), ("running but not TDS_RUNNING")); 539 CTR3(KTR_PROC, "sleepq resume: thread %p (pid %ld, %s)", 540 (void *)td, (long)td->td_proc->p_pid, (void *)td->td_name); 541 } 542 543 /* 544 * Check to see if we timed out. 545 */ 546 static int 547 sleepq_check_timeout(void) 548 { 549 struct thread *td; 550 551 td = curthread; 552 THREAD_LOCK_ASSERT(td, MA_OWNED); 553 554 /* 555 * If TDF_TIMEOUT is set, we timed out. 556 */ 557 if (td->td_flags & TDF_TIMEOUT) { 558 td->td_flags &= ~TDF_TIMEOUT; 559 return (EWOULDBLOCK); 560 } 561 562 /* 563 * If TDF_TIMOFAIL is set, the timeout ran after we had 564 * already been woken up. 565 */ 566 if (td->td_flags & TDF_TIMOFAIL) 567 td->td_flags &= ~TDF_TIMOFAIL; 568 569 /* 570 * If callout_stop() fails, then the timeout is running on 571 * another CPU, so synchronize with it to avoid having it 572 * accidentally wake up a subsequent sleep. 573 */ 574 else if (callout_stop(&td->td_slpcallout) == 0) { 575 td->td_flags |= TDF_TIMEOUT; 576 TD_SET_SLEEPING(td); 577 mi_switch(SW_INVOL | SWT_SLEEPQTIMO, NULL); 578 } 579 return (0); 580 } 581 582 /* 583 * Check to see if we were awoken by a signal. 584 */ 585 static int 586 sleepq_check_signals(void) 587 { 588 struct thread *td; 589 590 td = curthread; 591 THREAD_LOCK_ASSERT(td, MA_OWNED); 592 593 /* We are no longer in an interruptible sleep. */ 594 if (td->td_flags & TDF_SINTR) 595 td->td_flags &= ~TDF_SINTR; 596 597 if (td->td_flags & TDF_SLEEPABORT) { 598 td->td_flags &= ~TDF_SLEEPABORT; 599 return (td->td_intrval); 600 } 601 602 return (0); 603 } 604 605 /* 606 * Block the current thread until it is awakened from its sleep queue. 607 */ 608 void 609 sleepq_wait(void *wchan, int pri) 610 { 611 struct thread *td; 612 613 td = curthread; 614 MPASS(!(td->td_flags & TDF_SINTR)); 615 thread_lock(td); 616 sleepq_switch(wchan, pri); 617 thread_unlock(td); 618 } 619 620 /* 621 * Block the current thread until it is awakened from its sleep queue 622 * or it is interrupted by a signal. 623 */ 624 int 625 sleepq_wait_sig(void *wchan, int pri) 626 { 627 int rcatch; 628 int rval; 629 630 rcatch = sleepq_catch_signals(wchan, pri); 631 rval = sleepq_check_signals(); 632 thread_unlock(curthread); 633 if (rcatch) 634 return (rcatch); 635 return (rval); 636 } 637 638 /* 639 * Block the current thread until it is awakened from its sleep queue 640 * or it times out while waiting. 641 */ 642 int 643 sleepq_timedwait(void *wchan, int pri) 644 { 645 struct thread *td; 646 int rval; 647 648 td = curthread; 649 MPASS(!(td->td_flags & TDF_SINTR)); 650 thread_lock(td); 651 sleepq_switch(wchan, pri); 652 rval = sleepq_check_timeout(); 653 thread_unlock(td); 654 655 return (rval); 656 } 657 658 /* 659 * Block the current thread until it is awakened from its sleep queue, 660 * it is interrupted by a signal, or it times out waiting to be awakened. 661 */ 662 int 663 sleepq_timedwait_sig(void *wchan, int pri) 664 { 665 int rcatch, rvalt, rvals; 666 667 rcatch = sleepq_catch_signals(wchan, pri); 668 rvalt = sleepq_check_timeout(); 669 rvals = sleepq_check_signals(); 670 thread_unlock(curthread); 671 if (rcatch) 672 return (rcatch); 673 if (rvals) 674 return (rvals); 675 return (rvalt); 676 } 677 678 /* 679 * Returns the type of sleepqueue given a waitchannel. 680 */ 681 int 682 sleepq_type(void *wchan) 683 { 684 struct sleepqueue *sq; 685 int type; 686 687 MPASS(wchan != NULL); 688 689 sleepq_lock(wchan); 690 sq = sleepq_lookup(wchan); 691 if (sq == NULL) { 692 sleepq_release(wchan); 693 return (-1); 694 } 695 type = sq->sq_type; 696 sleepq_release(wchan); 697 return (type); 698 } 699 700 /* 701 * Removes a thread from a sleep queue and makes it 702 * runnable. 703 */ 704 static int 705 sleepq_resume_thread(struct sleepqueue *sq, struct thread *td, int pri) 706 { 707 struct sleepqueue_chain *sc; 708 709 MPASS(td != NULL); 710 MPASS(sq->sq_wchan != NULL); 711 MPASS(td->td_wchan == sq->sq_wchan); 712 MPASS(td->td_sqqueue < NR_SLEEPQS && td->td_sqqueue >= 0); 713 THREAD_LOCK_ASSERT(td, MA_OWNED); 714 sc = SC_LOOKUP(sq->sq_wchan); 715 mtx_assert(&sc->sc_lock, MA_OWNED); 716 717 SDT_PROBE2(sched, , , wakeup, td, td->td_proc); 718 719 /* Remove the thread from the queue. */ 720 sq->sq_blockedcnt[td->td_sqqueue]--; 721 TAILQ_REMOVE(&sq->sq_blocked[td->td_sqqueue], td, td_slpq); 722 723 /* 724 * Get a sleep queue for this thread. If this is the last waiter, 725 * use the queue itself and take it out of the chain, otherwise, 726 * remove a queue from the free list. 727 */ 728 if (LIST_EMPTY(&sq->sq_free)) { 729 td->td_sleepqueue = sq; 730 #ifdef INVARIANTS 731 sq->sq_wchan = NULL; 732 #endif 733 #ifdef SLEEPQUEUE_PROFILING 734 sc->sc_depth--; 735 #endif 736 } else 737 td->td_sleepqueue = LIST_FIRST(&sq->sq_free); 738 LIST_REMOVE(td->td_sleepqueue, sq_hash); 739 740 td->td_wmesg = NULL; 741 td->td_wchan = NULL; 742 td->td_flags &= ~TDF_SINTR; 743 744 CTR3(KTR_PROC, "sleepq_wakeup: thread %p (pid %ld, %s)", 745 (void *)td, (long)td->td_proc->p_pid, td->td_name); 746 747 /* Adjust priority if requested. */ 748 MPASS(pri == 0 || (pri >= PRI_MIN && pri <= PRI_MAX)); 749 if (pri != 0 && td->td_priority > pri && 750 PRI_BASE(td->td_pri_class) == PRI_TIMESHARE) 751 sched_prio(td, pri); 752 753 /* 754 * Note that thread td might not be sleeping if it is running 755 * sleepq_catch_signals() on another CPU or is blocked on its 756 * proc lock to check signals. There's no need to mark the 757 * thread runnable in that case. 758 */ 759 if (TD_IS_SLEEPING(td)) { 760 TD_CLR_SLEEPING(td); 761 return (setrunnable(td)); 762 } 763 return (0); 764 } 765 766 #ifdef INVARIANTS 767 /* 768 * UMA zone item deallocator. 769 */ 770 static void 771 sleepq_dtor(void *mem, int size, void *arg) 772 { 773 struct sleepqueue *sq; 774 int i; 775 776 sq = mem; 777 for (i = 0; i < NR_SLEEPQS; i++) { 778 MPASS(TAILQ_EMPTY(&sq->sq_blocked[i])); 779 MPASS(sq->sq_blockedcnt[i] == 0); 780 } 781 } 782 #endif 783 784 /* 785 * UMA zone item initializer. 786 */ 787 static int 788 sleepq_init(void *mem, int size, int flags) 789 { 790 struct sleepqueue *sq; 791 int i; 792 793 bzero(mem, size); 794 sq = mem; 795 for (i = 0; i < NR_SLEEPQS; i++) { 796 TAILQ_INIT(&sq->sq_blocked[i]); 797 sq->sq_blockedcnt[i] = 0; 798 } 799 LIST_INIT(&sq->sq_free); 800 return (0); 801 } 802 803 /* 804 * Find the highest priority thread sleeping on a wait channel and resume it. 805 */ 806 int 807 sleepq_signal(void *wchan, int flags, int pri, int queue) 808 { 809 struct sleepqueue *sq; 810 struct thread *td, *besttd; 811 int wakeup_swapper; 812 813 CTR2(KTR_PROC, "sleepq_signal(%p, %d)", wchan, flags); 814 KASSERT(wchan != NULL, ("%s: invalid NULL wait channel", __func__)); 815 MPASS((queue >= 0) && (queue < NR_SLEEPQS)); 816 sq = sleepq_lookup(wchan); 817 if (sq == NULL) 818 return (0); 819 KASSERT(sq->sq_type == (flags & SLEEPQ_TYPE), 820 ("%s: mismatch between sleep/wakeup and cv_*", __func__)); 821 822 /* 823 * Find the highest priority thread on the queue. If there is a 824 * tie, use the thread that first appears in the queue as it has 825 * been sleeping the longest since threads are always added to 826 * the tail of sleep queues. 827 */ 828 besttd = NULL; 829 TAILQ_FOREACH(td, &sq->sq_blocked[queue], td_slpq) { 830 if (besttd == NULL || td->td_priority < besttd->td_priority) 831 besttd = td; 832 } 833 MPASS(besttd != NULL); 834 thread_lock(besttd); 835 wakeup_swapper = sleepq_resume_thread(sq, besttd, pri); 836 thread_unlock(besttd); 837 return (wakeup_swapper); 838 } 839 840 /* 841 * Resume all threads sleeping on a specified wait channel. 842 */ 843 int 844 sleepq_broadcast(void *wchan, int flags, int pri, int queue) 845 { 846 struct sleepqueue *sq; 847 struct thread *td, *tdn; 848 int wakeup_swapper; 849 850 CTR2(KTR_PROC, "sleepq_broadcast(%p, %d)", wchan, flags); 851 KASSERT(wchan != NULL, ("%s: invalid NULL wait channel", __func__)); 852 MPASS((queue >= 0) && (queue < NR_SLEEPQS)); 853 sq = sleepq_lookup(wchan); 854 if (sq == NULL) 855 return (0); 856 KASSERT(sq->sq_type == (flags & SLEEPQ_TYPE), 857 ("%s: mismatch between sleep/wakeup and cv_*", __func__)); 858 859 /* Resume all blocked threads on the sleep queue. */ 860 wakeup_swapper = 0; 861 TAILQ_FOREACH_SAFE(td, &sq->sq_blocked[queue], td_slpq, tdn) { 862 thread_lock(td); 863 if (sleepq_resume_thread(sq, td, pri)) 864 wakeup_swapper = 1; 865 thread_unlock(td); 866 } 867 return (wakeup_swapper); 868 } 869 870 /* 871 * Time sleeping threads out. When the timeout expires, the thread is 872 * removed from the sleep queue and made runnable if it is still asleep. 873 */ 874 static void 875 sleepq_timeout(void *arg) 876 { 877 struct sleepqueue_chain *sc; 878 struct sleepqueue *sq; 879 struct thread *td; 880 void *wchan; 881 int wakeup_swapper; 882 883 td = arg; 884 wakeup_swapper = 0; 885 CTR3(KTR_PROC, "sleepq_timeout: thread %p (pid %ld, %s)", 886 (void *)td, (long)td->td_proc->p_pid, (void *)td->td_name); 887 888 /* 889 * First, see if the thread is asleep and get the wait channel if 890 * it is. 891 */ 892 thread_lock(td); 893 if (TD_IS_SLEEPING(td) && TD_ON_SLEEPQ(td)) { 894 wchan = td->td_wchan; 895 sc = SC_LOOKUP(wchan); 896 THREAD_LOCKPTR_ASSERT(td, &sc->sc_lock); 897 sq = sleepq_lookup(wchan); 898 MPASS(sq != NULL); 899 td->td_flags |= TDF_TIMEOUT; 900 wakeup_swapper = sleepq_resume_thread(sq, td, 0); 901 thread_unlock(td); 902 if (wakeup_swapper) 903 kick_proc0(); 904 return; 905 } 906 907 /* 908 * If the thread is on the SLEEPQ but isn't sleeping yet, it 909 * can either be on another CPU in between sleepq_add() and 910 * one of the sleepq_*wait*() routines or it can be in 911 * sleepq_catch_signals(). 912 */ 913 if (TD_ON_SLEEPQ(td)) { 914 td->td_flags |= TDF_TIMEOUT; 915 thread_unlock(td); 916 return; 917 } 918 919 /* 920 * Now check for the edge cases. First, if TDF_TIMEOUT is set, 921 * then the other thread has already yielded to us, so clear 922 * the flag and resume it. If TDF_TIMEOUT is not set, then the 923 * we know that the other thread is not on a sleep queue, but it 924 * hasn't resumed execution yet. In that case, set TDF_TIMOFAIL 925 * to let it know that the timeout has already run and doesn't 926 * need to be canceled. 927 */ 928 if (td->td_flags & TDF_TIMEOUT) { 929 MPASS(TD_IS_SLEEPING(td)); 930 td->td_flags &= ~TDF_TIMEOUT; 931 TD_CLR_SLEEPING(td); 932 wakeup_swapper = setrunnable(td); 933 } else 934 td->td_flags |= TDF_TIMOFAIL; 935 thread_unlock(td); 936 if (wakeup_swapper) 937 kick_proc0(); 938 } 939 940 /* 941 * Resumes a specific thread from the sleep queue associated with a specific 942 * wait channel if it is on that queue. 943 */ 944 void 945 sleepq_remove(struct thread *td, void *wchan) 946 { 947 struct sleepqueue *sq; 948 int wakeup_swapper; 949 950 /* 951 * Look up the sleep queue for this wait channel, then re-check 952 * that the thread is asleep on that channel, if it is not, then 953 * bail. 954 */ 955 MPASS(wchan != NULL); 956 sleepq_lock(wchan); 957 sq = sleepq_lookup(wchan); 958 /* 959 * We can not lock the thread here as it may be sleeping on a 960 * different sleepq. However, holding the sleepq lock for this 961 * wchan can guarantee that we do not miss a wakeup for this 962 * channel. The asserts below will catch any false positives. 963 */ 964 if (!TD_ON_SLEEPQ(td) || td->td_wchan != wchan) { 965 sleepq_release(wchan); 966 return; 967 } 968 /* Thread is asleep on sleep queue sq, so wake it up. */ 969 thread_lock(td); 970 MPASS(sq != NULL); 971 MPASS(td->td_wchan == wchan); 972 wakeup_swapper = sleepq_resume_thread(sq, td, 0); 973 thread_unlock(td); 974 sleepq_release(wchan); 975 if (wakeup_swapper) 976 kick_proc0(); 977 } 978 979 /* 980 * Abort a thread as if an interrupt had occurred. Only abort 981 * interruptible waits (unfortunately it isn't safe to abort others). 982 */ 983 int 984 sleepq_abort(struct thread *td, int intrval) 985 { 986 struct sleepqueue *sq; 987 void *wchan; 988 989 THREAD_LOCK_ASSERT(td, MA_OWNED); 990 MPASS(TD_ON_SLEEPQ(td)); 991 MPASS(td->td_flags & TDF_SINTR); 992 MPASS(intrval == EINTR || intrval == ERESTART); 993 994 /* 995 * If the TDF_TIMEOUT flag is set, just leave. A 996 * timeout is scheduled anyhow. 997 */ 998 if (td->td_flags & TDF_TIMEOUT) 999 return (0); 1000 1001 CTR3(KTR_PROC, "sleepq_abort: thread %p (pid %ld, %s)", 1002 (void *)td, (long)td->td_proc->p_pid, (void *)td->td_name); 1003 td->td_intrval = intrval; 1004 td->td_flags |= TDF_SLEEPABORT; 1005 /* 1006 * If the thread has not slept yet it will find the signal in 1007 * sleepq_catch_signals() and call sleepq_resume_thread. Otherwise 1008 * we have to do it here. 1009 */ 1010 if (!TD_IS_SLEEPING(td)) 1011 return (0); 1012 wchan = td->td_wchan; 1013 MPASS(wchan != NULL); 1014 sq = sleepq_lookup(wchan); 1015 MPASS(sq != NULL); 1016 1017 /* Thread is asleep on sleep queue sq, so wake it up. */ 1018 return (sleepq_resume_thread(sq, td, 0)); 1019 } 1020 1021 #ifdef SLEEPQUEUE_PROFILING 1022 #define SLEEPQ_PROF_LOCATIONS 1024 1023 #define SLEEPQ_SBUFSIZE 512 1024 struct sleepq_prof { 1025 LIST_ENTRY(sleepq_prof) sp_link; 1026 const char *sp_wmesg; 1027 long sp_count; 1028 }; 1029 1030 LIST_HEAD(sqphead, sleepq_prof); 1031 1032 struct sqphead sleepq_prof_free; 1033 struct sqphead sleepq_hash[SC_TABLESIZE]; 1034 static struct sleepq_prof sleepq_profent[SLEEPQ_PROF_LOCATIONS]; 1035 static struct mtx sleepq_prof_lock; 1036 MTX_SYSINIT(sleepq_prof_lock, &sleepq_prof_lock, "sleepq_prof", MTX_SPIN); 1037 1038 static void 1039 sleepq_profile(const char *wmesg) 1040 { 1041 struct sleepq_prof *sp; 1042 1043 mtx_lock_spin(&sleepq_prof_lock); 1044 if (prof_enabled == 0) 1045 goto unlock; 1046 LIST_FOREACH(sp, &sleepq_hash[SC_HASH(wmesg)], sp_link) 1047 if (sp->sp_wmesg == wmesg) 1048 goto done; 1049 sp = LIST_FIRST(&sleepq_prof_free); 1050 if (sp == NULL) 1051 goto unlock; 1052 sp->sp_wmesg = wmesg; 1053 LIST_REMOVE(sp, sp_link); 1054 LIST_INSERT_HEAD(&sleepq_hash[SC_HASH(wmesg)], sp, sp_link); 1055 done: 1056 sp->sp_count++; 1057 unlock: 1058 mtx_unlock_spin(&sleepq_prof_lock); 1059 return; 1060 } 1061 1062 static void 1063 sleepq_prof_reset(void) 1064 { 1065 struct sleepq_prof *sp; 1066 int enabled; 1067 int i; 1068 1069 mtx_lock_spin(&sleepq_prof_lock); 1070 enabled = prof_enabled; 1071 prof_enabled = 0; 1072 for (i = 0; i < SC_TABLESIZE; i++) 1073 LIST_INIT(&sleepq_hash[i]); 1074 LIST_INIT(&sleepq_prof_free); 1075 for (i = 0; i < SLEEPQ_PROF_LOCATIONS; i++) { 1076 sp = &sleepq_profent[i]; 1077 sp->sp_wmesg = NULL; 1078 sp->sp_count = 0; 1079 LIST_INSERT_HEAD(&sleepq_prof_free, sp, sp_link); 1080 } 1081 prof_enabled = enabled; 1082 mtx_unlock_spin(&sleepq_prof_lock); 1083 } 1084 1085 static int 1086 enable_sleepq_prof(SYSCTL_HANDLER_ARGS) 1087 { 1088 int error, v; 1089 1090 v = prof_enabled; 1091 error = sysctl_handle_int(oidp, &v, v, req); 1092 if (error) 1093 return (error); 1094 if (req->newptr == NULL) 1095 return (error); 1096 if (v == prof_enabled) 1097 return (0); 1098 if (v == 1) 1099 sleepq_prof_reset(); 1100 mtx_lock_spin(&sleepq_prof_lock); 1101 prof_enabled = !!v; 1102 mtx_unlock_spin(&sleepq_prof_lock); 1103 1104 return (0); 1105 } 1106 1107 static int 1108 reset_sleepq_prof_stats(SYSCTL_HANDLER_ARGS) 1109 { 1110 int error, v; 1111 1112 v = 0; 1113 error = sysctl_handle_int(oidp, &v, 0, req); 1114 if (error) 1115 return (error); 1116 if (req->newptr == NULL) 1117 return (error); 1118 if (v == 0) 1119 return (0); 1120 sleepq_prof_reset(); 1121 1122 return (0); 1123 } 1124 1125 static int 1126 dump_sleepq_prof_stats(SYSCTL_HANDLER_ARGS) 1127 { 1128 struct sleepq_prof *sp; 1129 struct sbuf *sb; 1130 int enabled; 1131 int error; 1132 int i; 1133 1134 error = sysctl_wire_old_buffer(req, 0); 1135 if (error != 0) 1136 return (error); 1137 sb = sbuf_new_for_sysctl(NULL, NULL, SLEEPQ_SBUFSIZE, req); 1138 sbuf_printf(sb, "\nwmesg\tcount\n"); 1139 enabled = prof_enabled; 1140 mtx_lock_spin(&sleepq_prof_lock); 1141 prof_enabled = 0; 1142 mtx_unlock_spin(&sleepq_prof_lock); 1143 for (i = 0; i < SC_TABLESIZE; i++) { 1144 LIST_FOREACH(sp, &sleepq_hash[i], sp_link) { 1145 sbuf_printf(sb, "%s\t%ld\n", 1146 sp->sp_wmesg, sp->sp_count); 1147 } 1148 } 1149 mtx_lock_spin(&sleepq_prof_lock); 1150 prof_enabled = enabled; 1151 mtx_unlock_spin(&sleepq_prof_lock); 1152 1153 error = sbuf_finish(sb); 1154 sbuf_delete(sb); 1155 return (error); 1156 } 1157 1158 SYSCTL_PROC(_debug_sleepq, OID_AUTO, stats, CTLTYPE_STRING | CTLFLAG_RD, 1159 NULL, 0, dump_sleepq_prof_stats, "A", "Sleepqueue profiling statistics"); 1160 SYSCTL_PROC(_debug_sleepq, OID_AUTO, reset, CTLTYPE_INT | CTLFLAG_RW, 1161 NULL, 0, reset_sleepq_prof_stats, "I", 1162 "Reset sleepqueue profiling statistics"); 1163 SYSCTL_PROC(_debug_sleepq, OID_AUTO, enable, CTLTYPE_INT | CTLFLAG_RW, 1164 NULL, 0, enable_sleepq_prof, "I", "Enable sleepqueue profiling"); 1165 #endif 1166 1167 #ifdef DDB 1168 DB_SHOW_COMMAND(sleepq, db_show_sleepqueue) 1169 { 1170 struct sleepqueue_chain *sc; 1171 struct sleepqueue *sq; 1172 #ifdef INVARIANTS 1173 struct lock_object *lock; 1174 #endif 1175 struct thread *td; 1176 void *wchan; 1177 int i; 1178 1179 if (!have_addr) 1180 return; 1181 1182 /* 1183 * First, see if there is an active sleep queue for the wait channel 1184 * indicated by the address. 1185 */ 1186 wchan = (void *)addr; 1187 sc = SC_LOOKUP(wchan); 1188 LIST_FOREACH(sq, &sc->sc_queues, sq_hash) 1189 if (sq->sq_wchan == wchan) 1190 goto found; 1191 1192 /* 1193 * Second, see if there is an active sleep queue at the address 1194 * indicated. 1195 */ 1196 for (i = 0; i < SC_TABLESIZE; i++) 1197 LIST_FOREACH(sq, &sleepq_chains[i].sc_queues, sq_hash) { 1198 if (sq == (struct sleepqueue *)addr) 1199 goto found; 1200 } 1201 1202 db_printf("Unable to locate a sleep queue via %p\n", (void *)addr); 1203 return; 1204 found: 1205 db_printf("Wait channel: %p\n", sq->sq_wchan); 1206 db_printf("Queue type: %d\n", sq->sq_type); 1207 #ifdef INVARIANTS 1208 if (sq->sq_lock) { 1209 lock = sq->sq_lock; 1210 db_printf("Associated Interlock: %p - (%s) %s\n", lock, 1211 LOCK_CLASS(lock)->lc_name, lock->lo_name); 1212 } 1213 #endif 1214 db_printf("Blocked threads:\n"); 1215 for (i = 0; i < NR_SLEEPQS; i++) { 1216 db_printf("\nQueue[%d]:\n", i); 1217 if (TAILQ_EMPTY(&sq->sq_blocked[i])) 1218 db_printf("\tempty\n"); 1219 else 1220 TAILQ_FOREACH(td, &sq->sq_blocked[0], 1221 td_slpq) { 1222 db_printf("\t%p (tid %d, pid %d, \"%s\")\n", td, 1223 td->td_tid, td->td_proc->p_pid, 1224 td->td_name); 1225 } 1226 db_printf("(expected: %u)\n", sq->sq_blockedcnt[i]); 1227 } 1228 } 1229 1230 /* Alias 'show sleepqueue' to 'show sleepq'. */ 1231 DB_SHOW_ALIAS(sleepqueue, db_show_sleepqueue); 1232 #endif 1233