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("$FreeBSD$"); 64 65 #include "opt_sleepqueue_profiling.h" 66 #include "opt_ddb.h" 67 #include "opt_kdtrace.h" 68 #include "opt_sched.h" 69 70 #include <sys/param.h> 71 #include <sys/systm.h> 72 #include <sys/lock.h> 73 #include <sys/kernel.h> 74 #include <sys/ktr.h> 75 #include <sys/mutex.h> 76 #include <sys/proc.h> 77 #include <sys/sbuf.h> 78 #include <sys/sched.h> 79 #include <sys/sdt.h> 80 #include <sys/signalvar.h> 81 #include <sys/sleepqueue.h> 82 #include <sys/sysctl.h> 83 84 #include <vm/uma.h> 85 86 #ifdef DDB 87 #include <ddb/ddb.h> 88 #endif 89 90 /* 91 * Constants for the hash table of sleep queue chains. 92 * SC_TABLESIZE must be a power of two for SC_MASK to work properly. 93 */ 94 #define SC_TABLESIZE 256 /* Must be power of 2. */ 95 #define SC_MASK (SC_TABLESIZE - 1) 96 #define SC_SHIFT 8 97 #define SC_HASH(wc) ((((uintptr_t)(wc) >> SC_SHIFT) ^ (uintptr_t)(wc)) & \ 98 SC_MASK) 99 #define SC_LOOKUP(wc) &sleepq_chains[SC_HASH(wc)] 100 #define NR_SLEEPQS 2 101 /* 102 * There two different lists of sleep queues. Both lists are connected 103 * via the sq_hash entries. The first list is the sleep queue chain list 104 * that a sleep queue is on when it is attached to a wait channel. The 105 * second list is the free list hung off of a sleep queue that is attached 106 * to a wait channel. 107 * 108 * Each sleep queue also contains the wait channel it is attached to, the 109 * list of threads blocked on that wait channel, flags specific to the 110 * wait channel, and the lock used to synchronize with a wait channel. 111 * The flags are used to catch mismatches between the various consumers 112 * of the sleep queue API (e.g. sleep/wakeup and condition variables). 113 * The lock pointer is only used when invariants are enabled for various 114 * debugging checks. 115 * 116 * Locking key: 117 * c - sleep queue chain lock 118 */ 119 struct sleepqueue { 120 TAILQ_HEAD(, thread) sq_blocked[NR_SLEEPQS]; /* (c) Blocked threads. */ 121 u_int sq_blockedcnt[NR_SLEEPQS]; /* (c) N. of blocked threads. */ 122 LIST_ENTRY(sleepqueue) sq_hash; /* (c) Chain and free list. */ 123 LIST_HEAD(, sleepqueue) sq_free; /* (c) Free queues. */ 124 void *sq_wchan; /* (c) Wait channel. */ 125 int sq_type; /* (c) Queue type. */ 126 #ifdef INVARIANTS 127 struct lock_object *sq_lock; /* (c) Associated lock. */ 128 #endif 129 }; 130 131 struct sleepqueue_chain { 132 LIST_HEAD(, sleepqueue) sc_queues; /* List of sleep queues. */ 133 struct mtx sc_lock; /* Spin lock for this chain. */ 134 #ifdef SLEEPQUEUE_PROFILING 135 u_int sc_depth; /* Length of sc_queues. */ 136 u_int sc_max_depth; /* Max length of sc_queues. */ 137 #endif 138 }; 139 140 #ifdef SLEEPQUEUE_PROFILING 141 u_int sleepq_max_depth; 142 static SYSCTL_NODE(_debug, OID_AUTO, sleepq, CTLFLAG_RD, 0, "sleepq profiling"); 143 static SYSCTL_NODE(_debug_sleepq, OID_AUTO, chains, CTLFLAG_RD, 0, 144 "sleepq chain stats"); 145 SYSCTL_UINT(_debug_sleepq, OID_AUTO, max_depth, CTLFLAG_RD, &sleepq_max_depth, 146 0, "maxmimum depth achieved of a single chain"); 147 148 static void sleepq_profile(const char *wmesg); 149 static int prof_enabled; 150 #endif 151 static struct sleepqueue_chain sleepq_chains[SC_TABLESIZE]; 152 static uma_zone_t sleepq_zone; 153 154 /* 155 * Prototypes for non-exported routines. 156 */ 157 static int sleepq_catch_signals(void *wchan, int pri); 158 static int sleepq_check_signals(void); 159 static int sleepq_check_timeout(void); 160 #ifdef INVARIANTS 161 static void sleepq_dtor(void *mem, int size, void *arg); 162 #endif 163 static int sleepq_init(void *mem, int size, int flags); 164 static int sleepq_resume_thread(struct sleepqueue *sq, struct thread *td, 165 int pri); 166 static void sleepq_switch(void *wchan, int pri); 167 static void sleepq_timeout(void *arg); 168 169 SDT_PROBE_DECLARE(sched, , , sleep); 170 SDT_PROBE_DECLARE(sched, , , wakeup); 171 172 /* 173 * Early initialization of sleep queues that is called from the sleepinit() 174 * SYSINIT. 175 */ 176 void 177 init_sleepqueues(void) 178 { 179 #ifdef SLEEPQUEUE_PROFILING 180 struct sysctl_oid *chain_oid; 181 char chain_name[10]; 182 #endif 183 int i; 184 185 for (i = 0; i < SC_TABLESIZE; i++) { 186 LIST_INIT(&sleepq_chains[i].sc_queues); 187 mtx_init(&sleepq_chains[i].sc_lock, "sleepq chain", NULL, 188 MTX_SPIN | MTX_RECURSE); 189 #ifdef SLEEPQUEUE_PROFILING 190 snprintf(chain_name, sizeof(chain_name), "%d", i); 191 chain_oid = SYSCTL_ADD_NODE(NULL, 192 SYSCTL_STATIC_CHILDREN(_debug_sleepq_chains), OID_AUTO, 193 chain_name, CTLFLAG_RD, NULL, "sleepq chain stats"); 194 SYSCTL_ADD_UINT(NULL, SYSCTL_CHILDREN(chain_oid), OID_AUTO, 195 "depth", CTLFLAG_RD, &sleepq_chains[i].sc_depth, 0, NULL); 196 SYSCTL_ADD_UINT(NULL, SYSCTL_CHILDREN(chain_oid), OID_AUTO, 197 "max_depth", CTLFLAG_RD, &sleepq_chains[i].sc_max_depth, 0, 198 NULL); 199 #endif 200 } 201 sleepq_zone = uma_zcreate("SLEEPQUEUE", sizeof(struct sleepqueue), 202 #ifdef INVARIANTS 203 NULL, sleepq_dtor, sleepq_init, NULL, UMA_ALIGN_CACHE, 0); 204 #else 205 NULL, NULL, sleepq_init, NULL, UMA_ALIGN_CACHE, 0); 206 #endif 207 208 thread0.td_sleepqueue = sleepq_alloc(); 209 } 210 211 /* 212 * Get a sleep queue for a new thread. 213 */ 214 struct sleepqueue * 215 sleepq_alloc(void) 216 { 217 218 return (uma_zalloc(sleepq_zone, M_WAITOK)); 219 } 220 221 /* 222 * Free a sleep queue when a thread is destroyed. 223 */ 224 void 225 sleepq_free(struct sleepqueue *sq) 226 { 227 228 uma_zfree(sleepq_zone, sq); 229 } 230 231 /* 232 * Lock the sleep queue chain associated with the specified wait channel. 233 */ 234 void 235 sleepq_lock(void *wchan) 236 { 237 struct sleepqueue_chain *sc; 238 239 sc = SC_LOOKUP(wchan); 240 mtx_lock_spin(&sc->sc_lock); 241 } 242 243 /* 244 * Look up the sleep queue associated with a given wait channel in the hash 245 * table locking the associated sleep queue chain. If no queue is found in 246 * the table, NULL is returned. 247 */ 248 struct sleepqueue * 249 sleepq_lookup(void *wchan) 250 { 251 struct sleepqueue_chain *sc; 252 struct sleepqueue *sq; 253 254 KASSERT(wchan != NULL, ("%s: invalid NULL wait channel", __func__)); 255 sc = SC_LOOKUP(wchan); 256 mtx_assert(&sc->sc_lock, MA_OWNED); 257 LIST_FOREACH(sq, &sc->sc_queues, sq_hash) 258 if (sq->sq_wchan == wchan) 259 return (sq); 260 return (NULL); 261 } 262 263 /* 264 * Unlock the sleep queue chain associated with a given wait channel. 265 */ 266 void 267 sleepq_release(void *wchan) 268 { 269 struct sleepqueue_chain *sc; 270 271 sc = SC_LOOKUP(wchan); 272 mtx_unlock_spin(&sc->sc_lock); 273 } 274 275 /* 276 * Places the current thread on the sleep queue for the specified wait 277 * channel. If INVARIANTS is enabled, then it associates the passed in 278 * lock with the sleepq to make sure it is held when that sleep queue is 279 * woken up. 280 */ 281 void 282 sleepq_add(void *wchan, struct lock_object *lock, const char *wmesg, int flags, 283 int queue) 284 { 285 struct sleepqueue_chain *sc; 286 struct sleepqueue *sq; 287 struct thread *td; 288 289 td = curthread; 290 sc = SC_LOOKUP(wchan); 291 mtx_assert(&sc->sc_lock, MA_OWNED); 292 MPASS(td->td_sleepqueue != NULL); 293 MPASS(wchan != NULL); 294 MPASS((queue >= 0) && (queue < NR_SLEEPQS)); 295 296 /* If this thread is not allowed to sleep, die a horrible death. */ 297 KASSERT(td->td_no_sleeping == 0, 298 ("%s: td %p to sleep on wchan %p with sleeping prohibited", 299 __func__, td, wchan)); 300 301 /* Look up the sleep queue associated with the wait channel 'wchan'. */ 302 sq = sleepq_lookup(wchan); 303 304 /* 305 * If the wait channel does not already have a sleep queue, use 306 * this thread's sleep queue. Otherwise, insert the current thread 307 * into the sleep queue already in use by this wait channel. 308 */ 309 if (sq == NULL) { 310 #ifdef INVARIANTS 311 int i; 312 313 sq = td->td_sleepqueue; 314 for (i = 0; i < NR_SLEEPQS; i++) { 315 KASSERT(TAILQ_EMPTY(&sq->sq_blocked[i]), 316 ("thread's sleep queue %d is not empty", i)); 317 KASSERT(sq->sq_blockedcnt[i] == 0, 318 ("thread's sleep queue %d count mismatches", i)); 319 } 320 KASSERT(LIST_EMPTY(&sq->sq_free), 321 ("thread's sleep queue has a non-empty free list")); 322 KASSERT(sq->sq_wchan == NULL, ("stale sq_wchan pointer")); 323 sq->sq_lock = lock; 324 #endif 325 #ifdef SLEEPQUEUE_PROFILING 326 sc->sc_depth++; 327 if (sc->sc_depth > sc->sc_max_depth) { 328 sc->sc_max_depth = sc->sc_depth; 329 if (sc->sc_max_depth > sleepq_max_depth) 330 sleepq_max_depth = sc->sc_max_depth; 331 } 332 #endif 333 sq = td->td_sleepqueue; 334 LIST_INSERT_HEAD(&sc->sc_queues, sq, sq_hash); 335 sq->sq_wchan = wchan; 336 sq->sq_type = flags & SLEEPQ_TYPE; 337 } else { 338 MPASS(wchan == sq->sq_wchan); 339 MPASS(lock == sq->sq_lock); 340 MPASS((flags & SLEEPQ_TYPE) == sq->sq_type); 341 LIST_INSERT_HEAD(&sq->sq_free, td->td_sleepqueue, sq_hash); 342 } 343 thread_lock(td); 344 TAILQ_INSERT_TAIL(&sq->sq_blocked[queue], td, td_slpq); 345 sq->sq_blockedcnt[queue]++; 346 td->td_sleepqueue = NULL; 347 td->td_sqqueue = queue; 348 td->td_wchan = wchan; 349 td->td_wmesg = wmesg; 350 if (flags & SLEEPQ_INTERRUPTIBLE) { 351 td->td_flags |= TDF_SINTR; 352 td->td_flags &= ~TDF_SLEEPABORT; 353 } 354 thread_unlock(td); 355 } 356 357 /* 358 * Sets a timeout that will remove the current thread from the specified 359 * sleep queue after timo ticks if the thread has not already been awakened. 360 */ 361 void 362 sleepq_set_timeout_sbt(void *wchan, sbintime_t sbt, sbintime_t pr, 363 int flags) 364 { 365 struct sleepqueue_chain *sc; 366 struct thread *td; 367 368 td = curthread; 369 sc = SC_LOOKUP(wchan); 370 mtx_assert(&sc->sc_lock, MA_OWNED); 371 MPASS(TD_ON_SLEEPQ(td)); 372 MPASS(td->td_sleepqueue == NULL); 373 MPASS(wchan != NULL); 374 callout_reset_sbt_on(&td->td_slpcallout, sbt, pr, 375 sleepq_timeout, td, PCPU_GET(cpuid), flags | C_DIRECT_EXEC); 376 } 377 378 /* 379 * Return the number of actual sleepers for the specified queue. 380 */ 381 u_int 382 sleepq_sleepcnt(void *wchan, int queue) 383 { 384 struct sleepqueue *sq; 385 386 KASSERT(wchan != NULL, ("%s: invalid NULL wait channel", __func__)); 387 MPASS((queue >= 0) && (queue < NR_SLEEPQS)); 388 sq = sleepq_lookup(wchan); 389 if (sq == NULL) 390 return (0); 391 return (sq->sq_blockedcnt[queue]); 392 } 393 394 /* 395 * Marks the pending sleep of the current thread as interruptible and 396 * makes an initial check for pending signals before putting a thread 397 * to sleep. Enters and exits with the thread lock held. Thread lock 398 * may have transitioned from the sleepq lock to a run lock. 399 */ 400 static int 401 sleepq_catch_signals(void *wchan, int pri) 402 { 403 struct sleepqueue_chain *sc; 404 struct sleepqueue *sq; 405 struct thread *td; 406 struct proc *p; 407 struct sigacts *ps; 408 int sig, ret; 409 410 td = curthread; 411 p = curproc; 412 sc = SC_LOOKUP(wchan); 413 mtx_assert(&sc->sc_lock, MA_OWNED); 414 MPASS(wchan != NULL); 415 if ((td->td_pflags & TDP_WAKEUP) != 0) { 416 td->td_pflags &= ~TDP_WAKEUP; 417 ret = EINTR; 418 thread_lock(td); 419 goto out; 420 } 421 422 /* 423 * See if there are any pending signals for this thread. If not 424 * we can switch immediately. Otherwise do the signal processing 425 * directly. 426 */ 427 thread_lock(td); 428 if ((td->td_flags & (TDF_NEEDSIGCHK | TDF_NEEDSUSPCHK)) == 0) { 429 sleepq_switch(wchan, pri); 430 return (0); 431 } 432 thread_unlock(td); 433 mtx_unlock_spin(&sc->sc_lock); 434 CTR3(KTR_PROC, "sleepq catching signals: thread %p (pid %ld, %s)", 435 (void *)td, (long)p->p_pid, td->td_name); 436 PROC_LOCK(p); 437 ps = p->p_sigacts; 438 mtx_lock(&ps->ps_mtx); 439 sig = cursig(td); 440 if (sig == 0) { 441 mtx_unlock(&ps->ps_mtx); 442 ret = thread_suspend_check(1); 443 MPASS(ret == 0 || ret == EINTR || ret == ERESTART); 444 } else { 445 if (SIGISMEMBER(ps->ps_sigintr, sig)) 446 ret = EINTR; 447 else 448 ret = ERESTART; 449 mtx_unlock(&ps->ps_mtx); 450 } 451 /* 452 * Lock the per-process spinlock prior to dropping the PROC_LOCK 453 * to avoid a signal delivery race. PROC_LOCK, PROC_SLOCK, and 454 * thread_lock() are currently held in tdsendsignal(). 455 */ 456 PROC_SLOCK(p); 457 mtx_lock_spin(&sc->sc_lock); 458 PROC_UNLOCK(p); 459 thread_lock(td); 460 PROC_SUNLOCK(p); 461 if (ret == 0) { 462 sleepq_switch(wchan, pri); 463 return (0); 464 } 465 out: 466 /* 467 * There were pending signals and this thread is still 468 * on the sleep queue, remove it from the sleep queue. 469 */ 470 if (TD_ON_SLEEPQ(td)) { 471 sq = sleepq_lookup(wchan); 472 if (sleepq_resume_thread(sq, td, 0)) { 473 #ifdef INVARIANTS 474 /* 475 * This thread hasn't gone to sleep yet, so it 476 * should not be swapped out. 477 */ 478 panic("not waking up swapper"); 479 #endif 480 } 481 } 482 mtx_unlock_spin(&sc->sc_lock); 483 MPASS(td->td_lock != &sc->sc_lock); 484 return (ret); 485 } 486 487 /* 488 * Switches to another thread if we are still asleep on a sleep queue. 489 * Returns with thread lock. 490 */ 491 static void 492 sleepq_switch(void *wchan, int pri) 493 { 494 struct sleepqueue_chain *sc; 495 struct sleepqueue *sq; 496 struct thread *td; 497 498 td = curthread; 499 sc = SC_LOOKUP(wchan); 500 mtx_assert(&sc->sc_lock, MA_OWNED); 501 THREAD_LOCK_ASSERT(td, MA_OWNED); 502 503 /* 504 * If we have a sleep queue, then we've already been woken up, so 505 * just return. 506 */ 507 if (td->td_sleepqueue != NULL) { 508 mtx_unlock_spin(&sc->sc_lock); 509 return; 510 } 511 512 /* 513 * If TDF_TIMEOUT is set, then our sleep has been timed out 514 * already but we are still on the sleep queue, so dequeue the 515 * thread and return. 516 */ 517 if (td->td_flags & TDF_TIMEOUT) { 518 MPASS(TD_ON_SLEEPQ(td)); 519 sq = sleepq_lookup(wchan); 520 if (sleepq_resume_thread(sq, td, 0)) { 521 #ifdef INVARIANTS 522 /* 523 * This thread hasn't gone to sleep yet, so it 524 * should not be swapped out. 525 */ 526 panic("not waking up swapper"); 527 #endif 528 } 529 mtx_unlock_spin(&sc->sc_lock); 530 return; 531 } 532 #ifdef SLEEPQUEUE_PROFILING 533 if (prof_enabled) 534 sleepq_profile(td->td_wmesg); 535 #endif 536 MPASS(td->td_sleepqueue == NULL); 537 sched_sleep(td, pri); 538 thread_lock_set(td, &sc->sc_lock); 539 SDT_PROBE0(sched, , , sleep); 540 TD_SET_SLEEPING(td); 541 mi_switch(SW_VOL | SWT_SLEEPQ, NULL); 542 KASSERT(TD_IS_RUNNING(td), ("running but not TDS_RUNNING")); 543 CTR3(KTR_PROC, "sleepq resume: thread %p (pid %ld, %s)", 544 (void *)td, (long)td->td_proc->p_pid, (void *)td->td_name); 545 } 546 547 /* 548 * Check to see if we timed out. 549 */ 550 static int 551 sleepq_check_timeout(void) 552 { 553 struct thread *td; 554 555 td = curthread; 556 THREAD_LOCK_ASSERT(td, MA_OWNED); 557 558 /* 559 * If TDF_TIMEOUT is set, we timed out. 560 */ 561 if (td->td_flags & TDF_TIMEOUT) { 562 td->td_flags &= ~TDF_TIMEOUT; 563 return (EWOULDBLOCK); 564 } 565 566 /* 567 * If TDF_TIMOFAIL is set, the timeout ran after we had 568 * already been woken up. 569 */ 570 if (td->td_flags & TDF_TIMOFAIL) 571 td->td_flags &= ~TDF_TIMOFAIL; 572 573 /* 574 * If callout_stop() fails, then the timeout is running on 575 * another CPU, so synchronize with it to avoid having it 576 * accidentally wake up a subsequent sleep. 577 */ 578 else if (callout_stop(&td->td_slpcallout) == 0) { 579 td->td_flags |= TDF_TIMEOUT; 580 TD_SET_SLEEPING(td); 581 mi_switch(SW_INVOL | SWT_SLEEPQTIMO, NULL); 582 } 583 return (0); 584 } 585 586 /* 587 * Check to see if we were awoken by a signal. 588 */ 589 static int 590 sleepq_check_signals(void) 591 { 592 struct thread *td; 593 594 td = curthread; 595 THREAD_LOCK_ASSERT(td, MA_OWNED); 596 597 /* We are no longer in an interruptible sleep. */ 598 if (td->td_flags & TDF_SINTR) 599 td->td_flags &= ~TDF_SINTR; 600 601 if (td->td_flags & TDF_SLEEPABORT) { 602 td->td_flags &= ~TDF_SLEEPABORT; 603 return (td->td_intrval); 604 } 605 606 return (0); 607 } 608 609 /* 610 * Block the current thread until it is awakened from its sleep queue. 611 */ 612 void 613 sleepq_wait(void *wchan, int pri) 614 { 615 struct thread *td; 616 617 td = curthread; 618 MPASS(!(td->td_flags & TDF_SINTR)); 619 thread_lock(td); 620 sleepq_switch(wchan, pri); 621 thread_unlock(td); 622 } 623 624 /* 625 * Block the current thread until it is awakened from its sleep queue 626 * or it is interrupted by a signal. 627 */ 628 int 629 sleepq_wait_sig(void *wchan, int pri) 630 { 631 int rcatch; 632 int rval; 633 634 rcatch = sleepq_catch_signals(wchan, pri); 635 rval = sleepq_check_signals(); 636 thread_unlock(curthread); 637 if (rcatch) 638 return (rcatch); 639 return (rval); 640 } 641 642 /* 643 * Block the current thread until it is awakened from its sleep queue 644 * or it times out while waiting. 645 */ 646 int 647 sleepq_timedwait(void *wchan, int pri) 648 { 649 struct thread *td; 650 int rval; 651 652 td = curthread; 653 MPASS(!(td->td_flags & TDF_SINTR)); 654 thread_lock(td); 655 sleepq_switch(wchan, pri); 656 rval = sleepq_check_timeout(); 657 thread_unlock(td); 658 659 return (rval); 660 } 661 662 /* 663 * Block the current thread until it is awakened from its sleep queue, 664 * it is interrupted by a signal, or it times out waiting to be awakened. 665 */ 666 int 667 sleepq_timedwait_sig(void *wchan, int pri) 668 { 669 int rcatch, rvalt, rvals; 670 671 rcatch = sleepq_catch_signals(wchan, pri); 672 rvalt = sleepq_check_timeout(); 673 rvals = sleepq_check_signals(); 674 thread_unlock(curthread); 675 if (rcatch) 676 return (rcatch); 677 if (rvals) 678 return (rvals); 679 return (rvalt); 680 } 681 682 /* 683 * Returns the type of sleepqueue given a waitchannel. 684 */ 685 int 686 sleepq_type(void *wchan) 687 { 688 struct sleepqueue *sq; 689 int type; 690 691 MPASS(wchan != NULL); 692 693 sleepq_lock(wchan); 694 sq = sleepq_lookup(wchan); 695 if (sq == NULL) { 696 sleepq_release(wchan); 697 return (-1); 698 } 699 type = sq->sq_type; 700 sleepq_release(wchan); 701 return (type); 702 } 703 704 /* 705 * Removes a thread from a sleep queue and makes it 706 * runnable. 707 */ 708 static int 709 sleepq_resume_thread(struct sleepqueue *sq, struct thread *td, int pri) 710 { 711 struct sleepqueue_chain *sc; 712 713 MPASS(td != NULL); 714 MPASS(sq->sq_wchan != NULL); 715 MPASS(td->td_wchan == sq->sq_wchan); 716 MPASS(td->td_sqqueue < NR_SLEEPQS && td->td_sqqueue >= 0); 717 THREAD_LOCK_ASSERT(td, MA_OWNED); 718 sc = SC_LOOKUP(sq->sq_wchan); 719 mtx_assert(&sc->sc_lock, MA_OWNED); 720 721 SDT_PROBE2(sched, , , wakeup, td, td->td_proc); 722 723 /* Remove the thread from the queue. */ 724 sq->sq_blockedcnt[td->td_sqqueue]--; 725 TAILQ_REMOVE(&sq->sq_blocked[td->td_sqqueue], td, td_slpq); 726 727 /* 728 * Get a sleep queue for this thread. If this is the last waiter, 729 * use the queue itself and take it out of the chain, otherwise, 730 * remove a queue from the free list. 731 */ 732 if (LIST_EMPTY(&sq->sq_free)) { 733 td->td_sleepqueue = sq; 734 #ifdef INVARIANTS 735 sq->sq_wchan = NULL; 736 #endif 737 #ifdef SLEEPQUEUE_PROFILING 738 sc->sc_depth--; 739 #endif 740 } else 741 td->td_sleepqueue = LIST_FIRST(&sq->sq_free); 742 LIST_REMOVE(td->td_sleepqueue, sq_hash); 743 744 td->td_wmesg = NULL; 745 td->td_wchan = NULL; 746 td->td_flags &= ~TDF_SINTR; 747 748 CTR3(KTR_PROC, "sleepq_wakeup: thread %p (pid %ld, %s)", 749 (void *)td, (long)td->td_proc->p_pid, td->td_name); 750 751 /* Adjust priority if requested. */ 752 MPASS(pri == 0 || (pri >= PRI_MIN && pri <= PRI_MAX)); 753 if (pri != 0 && td->td_priority > pri && 754 PRI_BASE(td->td_pri_class) == PRI_TIMESHARE) 755 sched_prio(td, pri); 756 757 /* 758 * Note that thread td might not be sleeping if it is running 759 * sleepq_catch_signals() on another CPU or is blocked on its 760 * proc lock to check signals. There's no need to mark the 761 * thread runnable in that case. 762 */ 763 if (TD_IS_SLEEPING(td)) { 764 TD_CLR_SLEEPING(td); 765 return (setrunnable(td)); 766 } 767 return (0); 768 } 769 770 #ifdef INVARIANTS 771 /* 772 * UMA zone item deallocator. 773 */ 774 static void 775 sleepq_dtor(void *mem, int size, void *arg) 776 { 777 struct sleepqueue *sq; 778 int i; 779 780 sq = mem; 781 for (i = 0; i < NR_SLEEPQS; i++) { 782 MPASS(TAILQ_EMPTY(&sq->sq_blocked[i])); 783 MPASS(sq->sq_blockedcnt[i] == 0); 784 } 785 } 786 #endif 787 788 /* 789 * UMA zone item initializer. 790 */ 791 static int 792 sleepq_init(void *mem, int size, int flags) 793 { 794 struct sleepqueue *sq; 795 int i; 796 797 bzero(mem, size); 798 sq = mem; 799 for (i = 0; i < NR_SLEEPQS; i++) { 800 TAILQ_INIT(&sq->sq_blocked[i]); 801 sq->sq_blockedcnt[i] = 0; 802 } 803 LIST_INIT(&sq->sq_free); 804 return (0); 805 } 806 807 /* 808 * Find the highest priority thread sleeping on a wait channel and resume it. 809 */ 810 int 811 sleepq_signal(void *wchan, int flags, int pri, int queue) 812 { 813 struct sleepqueue *sq; 814 struct thread *td, *besttd; 815 int wakeup_swapper; 816 817 CTR2(KTR_PROC, "sleepq_signal(%p, %d)", wchan, flags); 818 KASSERT(wchan != NULL, ("%s: invalid NULL wait channel", __func__)); 819 MPASS((queue >= 0) && (queue < NR_SLEEPQS)); 820 sq = sleepq_lookup(wchan); 821 if (sq == NULL) 822 return (0); 823 KASSERT(sq->sq_type == (flags & SLEEPQ_TYPE), 824 ("%s: mismatch between sleep/wakeup and cv_*", __func__)); 825 826 /* 827 * Find the highest priority thread on the queue. If there is a 828 * tie, use the thread that first appears in the queue as it has 829 * been sleeping the longest since threads are always added to 830 * the tail of sleep queues. 831 */ 832 besttd = NULL; 833 TAILQ_FOREACH(td, &sq->sq_blocked[queue], td_slpq) { 834 if (besttd == NULL || td->td_priority < besttd->td_priority) 835 besttd = td; 836 } 837 MPASS(besttd != NULL); 838 thread_lock(besttd); 839 wakeup_swapper = sleepq_resume_thread(sq, besttd, pri); 840 thread_unlock(besttd); 841 return (wakeup_swapper); 842 } 843 844 /* 845 * Resume all threads sleeping on a specified wait channel. 846 */ 847 int 848 sleepq_broadcast(void *wchan, int flags, int pri, int queue) 849 { 850 struct sleepqueue *sq; 851 struct thread *td, *tdn; 852 int wakeup_swapper; 853 854 CTR2(KTR_PROC, "sleepq_broadcast(%p, %d)", wchan, flags); 855 KASSERT(wchan != NULL, ("%s: invalid NULL wait channel", __func__)); 856 MPASS((queue >= 0) && (queue < NR_SLEEPQS)); 857 sq = sleepq_lookup(wchan); 858 if (sq == NULL) 859 return (0); 860 KASSERT(sq->sq_type == (flags & SLEEPQ_TYPE), 861 ("%s: mismatch between sleep/wakeup and cv_*", __func__)); 862 863 /* Resume all blocked threads on the sleep queue. */ 864 wakeup_swapper = 0; 865 TAILQ_FOREACH_SAFE(td, &sq->sq_blocked[queue], td_slpq, tdn) { 866 thread_lock(td); 867 if (sleepq_resume_thread(sq, td, pri)) 868 wakeup_swapper = 1; 869 thread_unlock(td); 870 } 871 return (wakeup_swapper); 872 } 873 874 /* 875 * Time sleeping threads out. When the timeout expires, the thread is 876 * removed from the sleep queue and made runnable if it is still asleep. 877 */ 878 static void 879 sleepq_timeout(void *arg) 880 { 881 struct sleepqueue_chain *sc; 882 struct sleepqueue *sq; 883 struct thread *td; 884 void *wchan; 885 int wakeup_swapper; 886 887 td = arg; 888 wakeup_swapper = 0; 889 CTR3(KTR_PROC, "sleepq_timeout: thread %p (pid %ld, %s)", 890 (void *)td, (long)td->td_proc->p_pid, (void *)td->td_name); 891 892 /* 893 * First, see if the thread is asleep and get the wait channel if 894 * it is. 895 */ 896 thread_lock(td); 897 if (TD_IS_SLEEPING(td) && TD_ON_SLEEPQ(td)) { 898 wchan = td->td_wchan; 899 sc = SC_LOOKUP(wchan); 900 THREAD_LOCKPTR_ASSERT(td, &sc->sc_lock); 901 sq = sleepq_lookup(wchan); 902 MPASS(sq != NULL); 903 td->td_flags |= TDF_TIMEOUT; 904 wakeup_swapper = sleepq_resume_thread(sq, td, 0); 905 thread_unlock(td); 906 if (wakeup_swapper) 907 kick_proc0(); 908 return; 909 } 910 911 /* 912 * If the thread is on the SLEEPQ but isn't sleeping yet, it 913 * can either be on another CPU in between sleepq_add() and 914 * one of the sleepq_*wait*() routines or it can be in 915 * sleepq_catch_signals(). 916 */ 917 if (TD_ON_SLEEPQ(td)) { 918 td->td_flags |= TDF_TIMEOUT; 919 thread_unlock(td); 920 return; 921 } 922 923 /* 924 * Now check for the edge cases. First, if TDF_TIMEOUT is set, 925 * then the other thread has already yielded to us, so clear 926 * the flag and resume it. If TDF_TIMEOUT is not set, then the 927 * we know that the other thread is not on a sleep queue, but it 928 * hasn't resumed execution yet. In that case, set TDF_TIMOFAIL 929 * to let it know that the timeout has already run and doesn't 930 * need to be canceled. 931 */ 932 if (td->td_flags & TDF_TIMEOUT) { 933 MPASS(TD_IS_SLEEPING(td)); 934 td->td_flags &= ~TDF_TIMEOUT; 935 TD_CLR_SLEEPING(td); 936 wakeup_swapper = setrunnable(td); 937 } else 938 td->td_flags |= TDF_TIMOFAIL; 939 thread_unlock(td); 940 if (wakeup_swapper) 941 kick_proc0(); 942 } 943 944 /* 945 * Resumes a specific thread from the sleep queue associated with a specific 946 * wait channel if it is on that queue. 947 */ 948 void 949 sleepq_remove(struct thread *td, void *wchan) 950 { 951 struct sleepqueue *sq; 952 int wakeup_swapper; 953 954 /* 955 * Look up the sleep queue for this wait channel, then re-check 956 * that the thread is asleep on that channel, if it is not, then 957 * bail. 958 */ 959 MPASS(wchan != NULL); 960 sleepq_lock(wchan); 961 sq = sleepq_lookup(wchan); 962 /* 963 * We can not lock the thread here as it may be sleeping on a 964 * different sleepq. However, holding the sleepq lock for this 965 * wchan can guarantee that we do not miss a wakeup for this 966 * channel. The asserts below will catch any false positives. 967 */ 968 if (!TD_ON_SLEEPQ(td) || td->td_wchan != wchan) { 969 sleepq_release(wchan); 970 return; 971 } 972 /* Thread is asleep on sleep queue sq, so wake it up. */ 973 thread_lock(td); 974 MPASS(sq != NULL); 975 MPASS(td->td_wchan == wchan); 976 wakeup_swapper = sleepq_resume_thread(sq, td, 0); 977 thread_unlock(td); 978 sleepq_release(wchan); 979 if (wakeup_swapper) 980 kick_proc0(); 981 } 982 983 /* 984 * Abort a thread as if an interrupt had occurred. Only abort 985 * interruptible waits (unfortunately it isn't safe to abort others). 986 */ 987 int 988 sleepq_abort(struct thread *td, int intrval) 989 { 990 struct sleepqueue *sq; 991 void *wchan; 992 993 THREAD_LOCK_ASSERT(td, MA_OWNED); 994 MPASS(TD_ON_SLEEPQ(td)); 995 MPASS(td->td_flags & TDF_SINTR); 996 MPASS(intrval == EINTR || intrval == ERESTART); 997 998 /* 999 * If the TDF_TIMEOUT flag is set, just leave. A 1000 * timeout is scheduled anyhow. 1001 */ 1002 if (td->td_flags & TDF_TIMEOUT) 1003 return (0); 1004 1005 CTR3(KTR_PROC, "sleepq_abort: thread %p (pid %ld, %s)", 1006 (void *)td, (long)td->td_proc->p_pid, (void *)td->td_name); 1007 td->td_intrval = intrval; 1008 td->td_flags |= TDF_SLEEPABORT; 1009 /* 1010 * If the thread has not slept yet it will find the signal in 1011 * sleepq_catch_signals() and call sleepq_resume_thread. Otherwise 1012 * we have to do it here. 1013 */ 1014 if (!TD_IS_SLEEPING(td)) 1015 return (0); 1016 wchan = td->td_wchan; 1017 MPASS(wchan != NULL); 1018 sq = sleepq_lookup(wchan); 1019 MPASS(sq != NULL); 1020 1021 /* Thread is asleep on sleep queue sq, so wake it up. */ 1022 return (sleepq_resume_thread(sq, td, 0)); 1023 } 1024 1025 #ifdef SLEEPQUEUE_PROFILING 1026 #define SLEEPQ_PROF_LOCATIONS 1024 1027 #define SLEEPQ_SBUFSIZE 512 1028 struct sleepq_prof { 1029 LIST_ENTRY(sleepq_prof) sp_link; 1030 const char *sp_wmesg; 1031 long sp_count; 1032 }; 1033 1034 LIST_HEAD(sqphead, sleepq_prof); 1035 1036 struct sqphead sleepq_prof_free; 1037 struct sqphead sleepq_hash[SC_TABLESIZE]; 1038 static struct sleepq_prof sleepq_profent[SLEEPQ_PROF_LOCATIONS]; 1039 static struct mtx sleepq_prof_lock; 1040 MTX_SYSINIT(sleepq_prof_lock, &sleepq_prof_lock, "sleepq_prof", MTX_SPIN); 1041 1042 static void 1043 sleepq_profile(const char *wmesg) 1044 { 1045 struct sleepq_prof *sp; 1046 1047 mtx_lock_spin(&sleepq_prof_lock); 1048 if (prof_enabled == 0) 1049 goto unlock; 1050 LIST_FOREACH(sp, &sleepq_hash[SC_HASH(wmesg)], sp_link) 1051 if (sp->sp_wmesg == wmesg) 1052 goto done; 1053 sp = LIST_FIRST(&sleepq_prof_free); 1054 if (sp == NULL) 1055 goto unlock; 1056 sp->sp_wmesg = wmesg; 1057 LIST_REMOVE(sp, sp_link); 1058 LIST_INSERT_HEAD(&sleepq_hash[SC_HASH(wmesg)], sp, sp_link); 1059 done: 1060 sp->sp_count++; 1061 unlock: 1062 mtx_unlock_spin(&sleepq_prof_lock); 1063 return; 1064 } 1065 1066 static void 1067 sleepq_prof_reset(void) 1068 { 1069 struct sleepq_prof *sp; 1070 int enabled; 1071 int i; 1072 1073 mtx_lock_spin(&sleepq_prof_lock); 1074 enabled = prof_enabled; 1075 prof_enabled = 0; 1076 for (i = 0; i < SC_TABLESIZE; i++) 1077 LIST_INIT(&sleepq_hash[i]); 1078 LIST_INIT(&sleepq_prof_free); 1079 for (i = 0; i < SLEEPQ_PROF_LOCATIONS; i++) { 1080 sp = &sleepq_profent[i]; 1081 sp->sp_wmesg = NULL; 1082 sp->sp_count = 0; 1083 LIST_INSERT_HEAD(&sleepq_prof_free, sp, sp_link); 1084 } 1085 prof_enabled = enabled; 1086 mtx_unlock_spin(&sleepq_prof_lock); 1087 } 1088 1089 static int 1090 enable_sleepq_prof(SYSCTL_HANDLER_ARGS) 1091 { 1092 int error, v; 1093 1094 v = prof_enabled; 1095 error = sysctl_handle_int(oidp, &v, v, req); 1096 if (error) 1097 return (error); 1098 if (req->newptr == NULL) 1099 return (error); 1100 if (v == prof_enabled) 1101 return (0); 1102 if (v == 1) 1103 sleepq_prof_reset(); 1104 mtx_lock_spin(&sleepq_prof_lock); 1105 prof_enabled = !!v; 1106 mtx_unlock_spin(&sleepq_prof_lock); 1107 1108 return (0); 1109 } 1110 1111 static int 1112 reset_sleepq_prof_stats(SYSCTL_HANDLER_ARGS) 1113 { 1114 int error, v; 1115 1116 v = 0; 1117 error = sysctl_handle_int(oidp, &v, 0, req); 1118 if (error) 1119 return (error); 1120 if (req->newptr == NULL) 1121 return (error); 1122 if (v == 0) 1123 return (0); 1124 sleepq_prof_reset(); 1125 1126 return (0); 1127 } 1128 1129 static int 1130 dump_sleepq_prof_stats(SYSCTL_HANDLER_ARGS) 1131 { 1132 struct sleepq_prof *sp; 1133 struct sbuf *sb; 1134 int enabled; 1135 int error; 1136 int i; 1137 1138 error = sysctl_wire_old_buffer(req, 0); 1139 if (error != 0) 1140 return (error); 1141 sb = sbuf_new_for_sysctl(NULL, NULL, SLEEPQ_SBUFSIZE, req); 1142 sbuf_printf(sb, "\nwmesg\tcount\n"); 1143 enabled = prof_enabled; 1144 mtx_lock_spin(&sleepq_prof_lock); 1145 prof_enabled = 0; 1146 mtx_unlock_spin(&sleepq_prof_lock); 1147 for (i = 0; i < SC_TABLESIZE; i++) { 1148 LIST_FOREACH(sp, &sleepq_hash[i], sp_link) { 1149 sbuf_printf(sb, "%s\t%ld\n", 1150 sp->sp_wmesg, sp->sp_count); 1151 } 1152 } 1153 mtx_lock_spin(&sleepq_prof_lock); 1154 prof_enabled = enabled; 1155 mtx_unlock_spin(&sleepq_prof_lock); 1156 1157 error = sbuf_finish(sb); 1158 sbuf_delete(sb); 1159 return (error); 1160 } 1161 1162 SYSCTL_PROC(_debug_sleepq, OID_AUTO, stats, CTLTYPE_STRING | CTLFLAG_RD, 1163 NULL, 0, dump_sleepq_prof_stats, "A", "Sleepqueue profiling statistics"); 1164 SYSCTL_PROC(_debug_sleepq, OID_AUTO, reset, CTLTYPE_INT | CTLFLAG_RW, 1165 NULL, 0, reset_sleepq_prof_stats, "I", 1166 "Reset sleepqueue profiling statistics"); 1167 SYSCTL_PROC(_debug_sleepq, OID_AUTO, enable, CTLTYPE_INT | CTLFLAG_RW, 1168 NULL, 0, enable_sleepq_prof, "I", "Enable sleepqueue profiling"); 1169 #endif 1170 1171 #ifdef DDB 1172 DB_SHOW_COMMAND(sleepq, db_show_sleepqueue) 1173 { 1174 struct sleepqueue_chain *sc; 1175 struct sleepqueue *sq; 1176 #ifdef INVARIANTS 1177 struct lock_object *lock; 1178 #endif 1179 struct thread *td; 1180 void *wchan; 1181 int i; 1182 1183 if (!have_addr) 1184 return; 1185 1186 /* 1187 * First, see if there is an active sleep queue for the wait channel 1188 * indicated by the address. 1189 */ 1190 wchan = (void *)addr; 1191 sc = SC_LOOKUP(wchan); 1192 LIST_FOREACH(sq, &sc->sc_queues, sq_hash) 1193 if (sq->sq_wchan == wchan) 1194 goto found; 1195 1196 /* 1197 * Second, see if there is an active sleep queue at the address 1198 * indicated. 1199 */ 1200 for (i = 0; i < SC_TABLESIZE; i++) 1201 LIST_FOREACH(sq, &sleepq_chains[i].sc_queues, sq_hash) { 1202 if (sq == (struct sleepqueue *)addr) 1203 goto found; 1204 } 1205 1206 db_printf("Unable to locate a sleep queue via %p\n", (void *)addr); 1207 return; 1208 found: 1209 db_printf("Wait channel: %p\n", sq->sq_wchan); 1210 db_printf("Queue type: %d\n", sq->sq_type); 1211 #ifdef INVARIANTS 1212 if (sq->sq_lock) { 1213 lock = sq->sq_lock; 1214 db_printf("Associated Interlock: %p - (%s) %s\n", lock, 1215 LOCK_CLASS(lock)->lc_name, lock->lo_name); 1216 } 1217 #endif 1218 db_printf("Blocked threads:\n"); 1219 for (i = 0; i < NR_SLEEPQS; i++) { 1220 db_printf("\nQueue[%d]:\n", i); 1221 if (TAILQ_EMPTY(&sq->sq_blocked[i])) 1222 db_printf("\tempty\n"); 1223 else 1224 TAILQ_FOREACH(td, &sq->sq_blocked[0], 1225 td_slpq) { 1226 db_printf("\t%p (tid %d, pid %d, \"%s\")\n", td, 1227 td->td_tid, td->td_proc->p_pid, 1228 td->td_name); 1229 } 1230 db_printf("(expected: %u)\n", sq->sq_blockedcnt[i]); 1231 } 1232 } 1233 1234 /* Alias 'show sleepqueue' to 'show sleepq'. */ 1235 DB_SHOW_ALIAS(sleepqueue, db_show_sleepqueue); 1236 #endif 1237