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