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_sched.h" 68 69 #include <sys/param.h> 70 #include <sys/systm.h> 71 #include <sys/lock.h> 72 #include <sys/kernel.h> 73 #include <sys/ktr.h> 74 #include <sys/mutex.h> 75 #include <sys/proc.h> 76 #include <sys/sbuf.h> 77 #include <sys/sched.h> 78 #include <sys/signalvar.h> 79 #include <sys/sleepqueue.h> 80 #include <sys/sysctl.h> 81 82 #include <vm/uma.h> 83 84 #ifdef DDB 85 #include <ddb/ddb.h> 86 #endif 87 88 /* 89 * Constants for the hash table of sleep queue chains. These constants are 90 * the same ones that 4BSD (and possibly earlier versions of BSD) used. 91 * Basically, we ignore the lower 8 bits of the address since most wait 92 * channel pointers are aligned and only look at the next 7 bits for the 93 * hash. SC_TABLESIZE must be a power of two for SC_MASK to work properly. 94 */ 95 #define SC_TABLESIZE 128 /* Must be power of 2. */ 96 #define SC_MASK (SC_TABLESIZE - 1) 97 #define SC_SHIFT 8 98 #define SC_HASH(wc) (((uintptr_t)(wc) >> SC_SHIFT) & 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 SYSCTL_NODE(_debug, OID_AUTO, sleepq, CTLFLAG_RD, 0, "sleepq profiling"); 143 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 /* 170 * Early initialization of sleep queues that is called from the sleepinit() 171 * SYSINIT. 172 */ 173 void 174 init_sleepqueues(void) 175 { 176 #ifdef SLEEPQUEUE_PROFILING 177 struct sysctl_oid *chain_oid; 178 char chain_name[10]; 179 #endif 180 int i; 181 182 for (i = 0; i < SC_TABLESIZE; i++) { 183 LIST_INIT(&sleepq_chains[i].sc_queues); 184 mtx_init(&sleepq_chains[i].sc_lock, "sleepq chain", NULL, 185 MTX_SPIN | MTX_RECURSE); 186 #ifdef SLEEPQUEUE_PROFILING 187 snprintf(chain_name, sizeof(chain_name), "%d", i); 188 chain_oid = SYSCTL_ADD_NODE(NULL, 189 SYSCTL_STATIC_CHILDREN(_debug_sleepq_chains), OID_AUTO, 190 chain_name, CTLFLAG_RD, NULL, "sleepq chain stats"); 191 SYSCTL_ADD_UINT(NULL, SYSCTL_CHILDREN(chain_oid), OID_AUTO, 192 "depth", CTLFLAG_RD, &sleepq_chains[i].sc_depth, 0, NULL); 193 SYSCTL_ADD_UINT(NULL, SYSCTL_CHILDREN(chain_oid), OID_AUTO, 194 "max_depth", CTLFLAG_RD, &sleepq_chains[i].sc_max_depth, 0, 195 NULL); 196 #endif 197 } 198 sleepq_zone = uma_zcreate("SLEEPQUEUE", sizeof(struct sleepqueue), 199 #ifdef INVARIANTS 200 NULL, sleepq_dtor, sleepq_init, NULL, UMA_ALIGN_CACHE, 0); 201 #else 202 NULL, NULL, sleepq_init, NULL, UMA_ALIGN_CACHE, 0); 203 #endif 204 205 thread0.td_sleepqueue = sleepq_alloc(); 206 } 207 208 /* 209 * Get a sleep queue for a new thread. 210 */ 211 struct sleepqueue * 212 sleepq_alloc(void) 213 { 214 215 return (uma_zalloc(sleepq_zone, M_WAITOK)); 216 } 217 218 /* 219 * Free a sleep queue when a thread is destroyed. 220 */ 221 void 222 sleepq_free(struct sleepqueue *sq) 223 { 224 225 uma_zfree(sleepq_zone, sq); 226 } 227 228 /* 229 * Lock the sleep queue chain associated with the specified wait channel. 230 */ 231 void 232 sleepq_lock(void *wchan) 233 { 234 struct sleepqueue_chain *sc; 235 236 sc = SC_LOOKUP(wchan); 237 mtx_lock_spin(&sc->sc_lock); 238 } 239 240 /* 241 * Look up the sleep queue associated with a given wait channel in the hash 242 * table locking the associated sleep queue chain. If no queue is found in 243 * the table, NULL is returned. 244 */ 245 struct sleepqueue * 246 sleepq_lookup(void *wchan) 247 { 248 struct sleepqueue_chain *sc; 249 struct sleepqueue *sq; 250 251 KASSERT(wchan != NULL, ("%s: invalid NULL wait channel", __func__)); 252 sc = SC_LOOKUP(wchan); 253 mtx_assert(&sc->sc_lock, MA_OWNED); 254 LIST_FOREACH(sq, &sc->sc_queues, sq_hash) 255 if (sq->sq_wchan == wchan) 256 return (sq); 257 return (NULL); 258 } 259 260 /* 261 * Unlock the sleep queue chain associated with a given wait channel. 262 */ 263 void 264 sleepq_release(void *wchan) 265 { 266 struct sleepqueue_chain *sc; 267 268 sc = SC_LOOKUP(wchan); 269 mtx_unlock_spin(&sc->sc_lock); 270 } 271 272 /* 273 * Places the current thread on the sleep queue for the specified wait 274 * channel. If INVARIANTS is enabled, then it associates the passed in 275 * lock with the sleepq to make sure it is held when that sleep queue is 276 * woken up. 277 */ 278 void 279 sleepq_add(void *wchan, struct lock_object *lock, const char *wmesg, int flags, 280 int queue) 281 { 282 struct sleepqueue_chain *sc; 283 struct sleepqueue *sq; 284 struct thread *td; 285 286 td = curthread; 287 sc = SC_LOOKUP(wchan); 288 mtx_assert(&sc->sc_lock, MA_OWNED); 289 MPASS(td->td_sleepqueue != NULL); 290 MPASS(wchan != NULL); 291 MPASS((queue >= 0) && (queue < NR_SLEEPQS)); 292 293 /* If this thread is not allowed to sleep, die a horrible death. */ 294 KASSERT(!(td->td_pflags & TDP_NOSLEEPING), 295 ("Trying sleep, but thread marked as sleeping prohibited")); 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 if (flags & SLEEPQ_STOP_ON_BDRY) 350 td->td_flags |= TDF_SBDRY; 351 } 352 thread_unlock(td); 353 } 354 355 /* 356 * Sets a timeout that will remove the current thread from the specified 357 * sleep queue after timo ticks if the thread has not already been awakened. 358 */ 359 void 360 sleepq_set_timeout(void *wchan, int timo) 361 { 362 struct sleepqueue_chain *sc; 363 struct thread *td; 364 365 td = curthread; 366 sc = SC_LOOKUP(wchan); 367 mtx_assert(&sc->sc_lock, MA_OWNED); 368 MPASS(TD_ON_SLEEPQ(td)); 369 MPASS(td->td_sleepqueue == NULL); 370 MPASS(wchan != NULL); 371 callout_reset_curcpu(&td->td_slpcallout, timo, sleepq_timeout, td); 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, stop_allowed; 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 stop_allowed = (td->td_flags & TDF_SBDRY) ? SIG_STOP_NOT_ALLOWED : 429 SIG_STOP_ALLOWED; 430 thread_unlock(td); 431 mtx_unlock_spin(&sc->sc_lock); 432 CTR3(KTR_PROC, "sleepq catching signals: thread %p (pid %ld, %s)", 433 (void *)td, (long)p->p_pid, td->td_name); 434 PROC_LOCK(p); 435 ps = p->p_sigacts; 436 mtx_lock(&ps->ps_mtx); 437 sig = cursig(td, stop_allowed); 438 if (sig == 0) { 439 mtx_unlock(&ps->ps_mtx); 440 ret = thread_suspend_check(1); 441 MPASS(ret == 0 || ret == EINTR || ret == ERESTART); 442 } else { 443 if (SIGISMEMBER(ps->ps_sigintr, sig)) 444 ret = EINTR; 445 else 446 ret = ERESTART; 447 mtx_unlock(&ps->ps_mtx); 448 } 449 /* 450 * Lock the per-process spinlock prior to dropping the PROC_LOCK 451 * to avoid a signal delivery race. PROC_LOCK, PROC_SLOCK, and 452 * thread_lock() are currently held in tdsendsignal(). 453 */ 454 PROC_SLOCK(p); 455 mtx_lock_spin(&sc->sc_lock); 456 PROC_UNLOCK(p); 457 thread_lock(td); 458 PROC_SUNLOCK(p); 459 if (ret == 0) { 460 sleepq_switch(wchan, pri); 461 return (0); 462 } 463 out: 464 /* 465 * There were pending signals and this thread is still 466 * on the sleep queue, remove it from the sleep queue. 467 */ 468 if (TD_ON_SLEEPQ(td)) { 469 sq = sleepq_lookup(wchan); 470 if (sleepq_resume_thread(sq, td, 0)) { 471 #ifdef INVARIANTS 472 /* 473 * This thread hasn't gone to sleep yet, so it 474 * should not be swapped out. 475 */ 476 panic("not waking up swapper"); 477 #endif 478 } 479 } 480 mtx_unlock_spin(&sc->sc_lock); 481 MPASS(td->td_lock != &sc->sc_lock); 482 return (ret); 483 } 484 485 /* 486 * Switches to another thread if we are still asleep on a sleep queue. 487 * Returns with thread lock. 488 */ 489 static void 490 sleepq_switch(void *wchan, int pri) 491 { 492 struct sleepqueue_chain *sc; 493 struct sleepqueue *sq; 494 struct thread *td; 495 496 td = curthread; 497 sc = SC_LOOKUP(wchan); 498 mtx_assert(&sc->sc_lock, MA_OWNED); 499 THREAD_LOCK_ASSERT(td, MA_OWNED); 500 501 /* 502 * If we have a sleep queue, then we've already been woken up, so 503 * just return. 504 */ 505 if (td->td_sleepqueue != NULL) { 506 mtx_unlock_spin(&sc->sc_lock); 507 return; 508 } 509 510 /* 511 * If TDF_TIMEOUT is set, then our sleep has been timed out 512 * already but we are still on the sleep queue, so dequeue the 513 * thread and return. 514 */ 515 if (td->td_flags & TDF_TIMEOUT) { 516 MPASS(TD_ON_SLEEPQ(td)); 517 sq = sleepq_lookup(wchan); 518 if (sleepq_resume_thread(sq, td, 0)) { 519 #ifdef INVARIANTS 520 /* 521 * This thread hasn't gone to sleep yet, so it 522 * should not be swapped out. 523 */ 524 panic("not waking up swapper"); 525 #endif 526 } 527 mtx_unlock_spin(&sc->sc_lock); 528 return; 529 } 530 #ifdef SLEEPQUEUE_PROFILING 531 if (prof_enabled) 532 sleepq_profile(td->td_wmesg); 533 #endif 534 MPASS(td->td_sleepqueue == NULL); 535 sched_sleep(td, pri); 536 thread_lock_set(td, &sc->sc_lock); 537 TD_SET_SLEEPING(td); 538 mi_switch(SW_VOL | SWT_SLEEPQ, NULL); 539 KASSERT(TD_IS_RUNNING(td), ("running but not TDS_RUNNING")); 540 CTR3(KTR_PROC, "sleepq resume: thread %p (pid %ld, %s)", 541 (void *)td, (long)td->td_proc->p_pid, (void *)td->td_name); 542 } 543 544 /* 545 * Check to see if we timed out. 546 */ 547 static int 548 sleepq_check_timeout(void) 549 { 550 struct thread *td; 551 552 td = curthread; 553 THREAD_LOCK_ASSERT(td, MA_OWNED); 554 555 /* 556 * If TDF_TIMEOUT is set, we timed out. 557 */ 558 if (td->td_flags & TDF_TIMEOUT) { 559 td->td_flags &= ~TDF_TIMEOUT; 560 return (EWOULDBLOCK); 561 } 562 563 /* 564 * If TDF_TIMOFAIL is set, the timeout ran after we had 565 * already been woken up. 566 */ 567 if (td->td_flags & TDF_TIMOFAIL) 568 td->td_flags &= ~TDF_TIMOFAIL; 569 570 /* 571 * If callout_stop() fails, then the timeout is running on 572 * another CPU, so synchronize with it to avoid having it 573 * accidentally wake up a subsequent sleep. 574 */ 575 else if (callout_stop(&td->td_slpcallout) == 0) { 576 td->td_flags |= TDF_TIMEOUT; 577 TD_SET_SLEEPING(td); 578 mi_switch(SW_INVOL | SWT_SLEEPQTIMO, NULL); 579 } 580 return (0); 581 } 582 583 /* 584 * Check to see if we were awoken by a signal. 585 */ 586 static int 587 sleepq_check_signals(void) 588 { 589 struct thread *td; 590 591 td = curthread; 592 THREAD_LOCK_ASSERT(td, MA_OWNED); 593 594 /* We are no longer in an interruptible sleep. */ 595 if (td->td_flags & TDF_SINTR) 596 td->td_flags &= ~(TDF_SINTR | TDF_SBDRY); 597 598 if (td->td_flags & TDF_SLEEPABORT) { 599 td->td_flags &= ~TDF_SLEEPABORT; 600 return (td->td_intrval); 601 } 602 603 return (0); 604 } 605 606 /* 607 * Block the current thread until it is awakened from its sleep queue. 608 */ 609 void 610 sleepq_wait(void *wchan, int pri) 611 { 612 struct thread *td; 613 614 td = curthread; 615 MPASS(!(td->td_flags & TDF_SINTR)); 616 thread_lock(td); 617 sleepq_switch(wchan, pri); 618 thread_unlock(td); 619 } 620 621 /* 622 * Block the current thread until it is awakened from its sleep queue 623 * or it is interrupted by a signal. 624 */ 625 int 626 sleepq_wait_sig(void *wchan, int pri) 627 { 628 int rcatch; 629 int rval; 630 631 rcatch = sleepq_catch_signals(wchan, pri); 632 rval = sleepq_check_signals(); 633 thread_unlock(curthread); 634 if (rcatch) 635 return (rcatch); 636 return (rval); 637 } 638 639 /* 640 * Block the current thread until it is awakened from its sleep queue 641 * or it times out while waiting. 642 */ 643 int 644 sleepq_timedwait(void *wchan, int pri) 645 { 646 struct thread *td; 647 int rval; 648 649 td = curthread; 650 MPASS(!(td->td_flags & TDF_SINTR)); 651 thread_lock(td); 652 sleepq_switch(wchan, pri); 653 rval = sleepq_check_timeout(); 654 thread_unlock(td); 655 656 return (rval); 657 } 658 659 /* 660 * Block the current thread until it is awakened from its sleep queue, 661 * it is interrupted by a signal, or it times out waiting to be awakened. 662 */ 663 int 664 sleepq_timedwait_sig(void *wchan, int pri) 665 { 666 int rcatch, rvalt, rvals; 667 668 rcatch = sleepq_catch_signals(wchan, pri); 669 rvalt = sleepq_check_timeout(); 670 rvals = sleepq_check_signals(); 671 thread_unlock(curthread); 672 if (rcatch) 673 return (rcatch); 674 if (rvals) 675 return (rvals); 676 return (rvalt); 677 } 678 679 /* 680 * Returns the type of sleepqueue given a waitchannel. 681 */ 682 int 683 sleepq_type(void *wchan) 684 { 685 struct sleepqueue *sq; 686 int type; 687 688 MPASS(wchan != NULL); 689 690 sleepq_lock(wchan); 691 sq = sleepq_lookup(wchan); 692 if (sq == NULL) { 693 sleepq_release(wchan); 694 return (-1); 695 } 696 type = sq->sq_type; 697 sleepq_release(wchan); 698 return (type); 699 } 700 701 /* 702 * Removes a thread from a sleep queue and makes it 703 * runnable. 704 */ 705 static int 706 sleepq_resume_thread(struct sleepqueue *sq, struct thread *td, int pri) 707 { 708 struct sleepqueue_chain *sc; 709 710 MPASS(td != NULL); 711 MPASS(sq->sq_wchan != NULL); 712 MPASS(td->td_wchan == sq->sq_wchan); 713 MPASS(td->td_sqqueue < NR_SLEEPQS && td->td_sqqueue >= 0); 714 THREAD_LOCK_ASSERT(td, MA_OWNED); 715 sc = SC_LOOKUP(sq->sq_wchan); 716 mtx_assert(&sc->sc_lock, MA_OWNED); 717 718 /* Remove the thread from the queue. */ 719 sq->sq_blockedcnt[td->td_sqqueue]--; 720 TAILQ_REMOVE(&sq->sq_blocked[td->td_sqqueue], td, td_slpq); 721 722 /* 723 * Get a sleep queue for this thread. If this is the last waiter, 724 * use the queue itself and take it out of the chain, otherwise, 725 * remove a queue from the free list. 726 */ 727 if (LIST_EMPTY(&sq->sq_free)) { 728 td->td_sleepqueue = sq; 729 #ifdef INVARIANTS 730 sq->sq_wchan = NULL; 731 #endif 732 #ifdef SLEEPQUEUE_PROFILING 733 sc->sc_depth--; 734 #endif 735 } else 736 td->td_sleepqueue = LIST_FIRST(&sq->sq_free); 737 LIST_REMOVE(td->td_sleepqueue, sq_hash); 738 739 td->td_wmesg = NULL; 740 td->td_wchan = NULL; 741 td->td_flags &= ~(TDF_SINTR | TDF_SBDRY); 742 743 CTR3(KTR_PROC, "sleepq_wakeup: thread %p (pid %ld, %s)", 744 (void *)td, (long)td->td_proc->p_pid, td->td_name); 745 746 /* Adjust priority if requested. */ 747 MPASS(pri == 0 || (pri >= PRI_MIN && pri <= PRI_MAX)); 748 if (pri != 0 && td->td_priority > pri && 749 PRI_BASE(td->td_pri_class) == PRI_TIMESHARE) 750 sched_prio(td, pri); 751 752 /* 753 * Note that thread td might not be sleeping if it is running 754 * sleepq_catch_signals() on another CPU or is blocked on its 755 * proc lock to check signals. There's no need to mark the 756 * thread runnable in that case. 757 */ 758 if (TD_IS_SLEEPING(td)) { 759 TD_CLR_SLEEPING(td); 760 return (setrunnable(td)); 761 } 762 return (0); 763 } 764 765 #ifdef INVARIANTS 766 /* 767 * UMA zone item deallocator. 768 */ 769 static void 770 sleepq_dtor(void *mem, int size, void *arg) 771 { 772 struct sleepqueue *sq; 773 int i; 774 775 sq = mem; 776 for (i = 0; i < NR_SLEEPQS; i++) { 777 MPASS(TAILQ_EMPTY(&sq->sq_blocked[i])); 778 MPASS(sq->sq_blockedcnt[i] == 0); 779 } 780 } 781 #endif 782 783 /* 784 * UMA zone item initializer. 785 */ 786 static int 787 sleepq_init(void *mem, int size, int flags) 788 { 789 struct sleepqueue *sq; 790 int i; 791 792 bzero(mem, size); 793 sq = mem; 794 for (i = 0; i < NR_SLEEPQS; i++) { 795 TAILQ_INIT(&sq->sq_blocked[i]); 796 sq->sq_blockedcnt[i] = 0; 797 } 798 LIST_INIT(&sq->sq_free); 799 return (0); 800 } 801 802 /* 803 * Find the highest priority thread sleeping on a wait channel and resume it. 804 */ 805 int 806 sleepq_signal(void *wchan, int flags, int pri, int queue) 807 { 808 struct sleepqueue *sq; 809 struct thread *td, *besttd; 810 int wakeup_swapper; 811 812 CTR2(KTR_PROC, "sleepq_signal(%p, %d)", wchan, flags); 813 KASSERT(wchan != NULL, ("%s: invalid NULL wait channel", __func__)); 814 MPASS((queue >= 0) && (queue < NR_SLEEPQS)); 815 sq = sleepq_lookup(wchan); 816 if (sq == NULL) 817 return (0); 818 KASSERT(sq->sq_type == (flags & SLEEPQ_TYPE), 819 ("%s: mismatch between sleep/wakeup and cv_*", __func__)); 820 821 /* 822 * Find the highest priority thread on the queue. If there is a 823 * tie, use the thread that first appears in the queue as it has 824 * been sleeping the longest since threads are always added to 825 * the tail of sleep queues. 826 */ 827 besttd = NULL; 828 TAILQ_FOREACH(td, &sq->sq_blocked[queue], td_slpq) { 829 if (besttd == NULL || td->td_priority < besttd->td_priority) 830 besttd = td; 831 } 832 MPASS(besttd != NULL); 833 thread_lock(besttd); 834 wakeup_swapper = sleepq_resume_thread(sq, besttd, pri); 835 thread_unlock(besttd); 836 return (wakeup_swapper); 837 } 838 839 /* 840 * Resume all threads sleeping on a specified wait channel. 841 */ 842 int 843 sleepq_broadcast(void *wchan, int flags, int pri, int queue) 844 { 845 struct sleepqueue *sq; 846 struct thread *td, *tdn; 847 int wakeup_swapper; 848 849 CTR2(KTR_PROC, "sleepq_broadcast(%p, %d)", wchan, flags); 850 KASSERT(wchan != NULL, ("%s: invalid NULL wait channel", __func__)); 851 MPASS((queue >= 0) && (queue < NR_SLEEPQS)); 852 sq = sleepq_lookup(wchan); 853 if (sq == NULL) 854 return (0); 855 KASSERT(sq->sq_type == (flags & SLEEPQ_TYPE), 856 ("%s: mismatch between sleep/wakeup and cv_*", __func__)); 857 858 /* Resume all blocked threads on the sleep queue. */ 859 wakeup_swapper = 0; 860 TAILQ_FOREACH_SAFE(td, &sq->sq_blocked[queue], td_slpq, tdn) { 861 thread_lock(td); 862 if (sleepq_resume_thread(sq, td, pri)) 863 wakeup_swapper = 1; 864 thread_unlock(td); 865 } 866 return (wakeup_swapper); 867 } 868 869 /* 870 * Time sleeping threads out. When the timeout expires, the thread is 871 * removed from the sleep queue and made runnable if it is still asleep. 872 */ 873 static void 874 sleepq_timeout(void *arg) 875 { 876 struct sleepqueue_chain *sc; 877 struct sleepqueue *sq; 878 struct thread *td; 879 void *wchan; 880 int wakeup_swapper; 881 882 td = arg; 883 wakeup_swapper = 0; 884 CTR3(KTR_PROC, "sleepq_timeout: thread %p (pid %ld, %s)", 885 (void *)td, (long)td->td_proc->p_pid, (void *)td->td_name); 886 887 /* 888 * First, see if the thread is asleep and get the wait channel if 889 * it is. 890 */ 891 thread_lock(td); 892 if (TD_IS_SLEEPING(td) && TD_ON_SLEEPQ(td)) { 893 wchan = td->td_wchan; 894 sc = SC_LOOKUP(wchan); 895 THREAD_LOCKPTR_ASSERT(td, &sc->sc_lock); 896 sq = sleepq_lookup(wchan); 897 MPASS(sq != NULL); 898 td->td_flags |= TDF_TIMEOUT; 899 wakeup_swapper = sleepq_resume_thread(sq, td, 0); 900 thread_unlock(td); 901 if (wakeup_swapper) 902 kick_proc0(); 903 return; 904 } 905 906 /* 907 * If the thread is on the SLEEPQ but isn't sleeping yet, it 908 * can either be on another CPU in between sleepq_add() and 909 * one of the sleepq_*wait*() routines or it can be in 910 * sleepq_catch_signals(). 911 */ 912 if (TD_ON_SLEEPQ(td)) { 913 td->td_flags |= TDF_TIMEOUT; 914 thread_unlock(td); 915 return; 916 } 917 918 /* 919 * Now check for the edge cases. First, if TDF_TIMEOUT is set, 920 * then the other thread has already yielded to us, so clear 921 * the flag and resume it. If TDF_TIMEOUT is not set, then the 922 * we know that the other thread is not on a sleep queue, but it 923 * hasn't resumed execution yet. In that case, set TDF_TIMOFAIL 924 * to let it know that the timeout has already run and doesn't 925 * need to be canceled. 926 */ 927 if (td->td_flags & TDF_TIMEOUT) { 928 MPASS(TD_IS_SLEEPING(td)); 929 td->td_flags &= ~TDF_TIMEOUT; 930 TD_CLR_SLEEPING(td); 931 wakeup_swapper = setrunnable(td); 932 } else 933 td->td_flags |= TDF_TIMOFAIL; 934 thread_unlock(td); 935 if (wakeup_swapper) 936 kick_proc0(); 937 } 938 939 /* 940 * Resumes a specific thread from the sleep queue associated with a specific 941 * wait channel if it is on that queue. 942 */ 943 void 944 sleepq_remove(struct thread *td, void *wchan) 945 { 946 struct sleepqueue *sq; 947 int wakeup_swapper; 948 949 /* 950 * Look up the sleep queue for this wait channel, then re-check 951 * that the thread is asleep on that channel, if it is not, then 952 * bail. 953 */ 954 MPASS(wchan != NULL); 955 sleepq_lock(wchan); 956 sq = sleepq_lookup(wchan); 957 /* 958 * We can not lock the thread here as it may be sleeping on a 959 * different sleepq. However, holding the sleepq lock for this 960 * wchan can guarantee that we do not miss a wakeup for this 961 * channel. The asserts below will catch any false positives. 962 */ 963 if (!TD_ON_SLEEPQ(td) || td->td_wchan != wchan) { 964 sleepq_release(wchan); 965 return; 966 } 967 /* Thread is asleep on sleep queue sq, so wake it up. */ 968 thread_lock(td); 969 MPASS(sq != NULL); 970 MPASS(td->td_wchan == wchan); 971 wakeup_swapper = sleepq_resume_thread(sq, td, 0); 972 thread_unlock(td); 973 sleepq_release(wchan); 974 if (wakeup_swapper) 975 kick_proc0(); 976 } 977 978 /* 979 * Abort a thread as if an interrupt had occurred. Only abort 980 * interruptible waits (unfortunately it isn't safe to abort others). 981 */ 982 int 983 sleepq_abort(struct thread *td, int intrval) 984 { 985 struct sleepqueue *sq; 986 void *wchan; 987 988 THREAD_LOCK_ASSERT(td, MA_OWNED); 989 MPASS(TD_ON_SLEEPQ(td)); 990 MPASS(td->td_flags & TDF_SINTR); 991 MPASS(intrval == EINTR || intrval == ERESTART); 992 993 /* 994 * If the TDF_TIMEOUT flag is set, just leave. A 995 * timeout is scheduled anyhow. 996 */ 997 if (td->td_flags & TDF_TIMEOUT) 998 return (0); 999 1000 CTR3(KTR_PROC, "sleepq_abort: thread %p (pid %ld, %s)", 1001 (void *)td, (long)td->td_proc->p_pid, (void *)td->td_name); 1002 td->td_intrval = intrval; 1003 td->td_flags |= TDF_SLEEPABORT; 1004 /* 1005 * If the thread has not slept yet it will find the signal in 1006 * sleepq_catch_signals() and call sleepq_resume_thread. Otherwise 1007 * we have to do it here. 1008 */ 1009 if (!TD_IS_SLEEPING(td)) 1010 return (0); 1011 wchan = td->td_wchan; 1012 MPASS(wchan != NULL); 1013 sq = sleepq_lookup(wchan); 1014 MPASS(sq != NULL); 1015 1016 /* Thread is asleep on sleep queue sq, so wake it up. */ 1017 return (sleepq_resume_thread(sq, td, 0)); 1018 } 1019 1020 #ifdef SLEEPQUEUE_PROFILING 1021 #define SLEEPQ_PROF_LOCATIONS 1024 1022 #define SLEEPQ_SBUFSIZE 512 1023 struct sleepq_prof { 1024 LIST_ENTRY(sleepq_prof) sp_link; 1025 const char *sp_wmesg; 1026 long sp_count; 1027 }; 1028 1029 LIST_HEAD(sqphead, sleepq_prof); 1030 1031 struct sqphead sleepq_prof_free; 1032 struct sqphead sleepq_hash[SC_TABLESIZE]; 1033 static struct sleepq_prof sleepq_profent[SLEEPQ_PROF_LOCATIONS]; 1034 static struct mtx sleepq_prof_lock; 1035 MTX_SYSINIT(sleepq_prof_lock, &sleepq_prof_lock, "sleepq_prof", MTX_SPIN); 1036 1037 static void 1038 sleepq_profile(const char *wmesg) 1039 { 1040 struct sleepq_prof *sp; 1041 1042 mtx_lock_spin(&sleepq_prof_lock); 1043 if (prof_enabled == 0) 1044 goto unlock; 1045 LIST_FOREACH(sp, &sleepq_hash[SC_HASH(wmesg)], sp_link) 1046 if (sp->sp_wmesg == wmesg) 1047 goto done; 1048 sp = LIST_FIRST(&sleepq_prof_free); 1049 if (sp == NULL) 1050 goto unlock; 1051 sp->sp_wmesg = wmesg; 1052 LIST_REMOVE(sp, sp_link); 1053 LIST_INSERT_HEAD(&sleepq_hash[SC_HASH(wmesg)], sp, sp_link); 1054 done: 1055 sp->sp_count++; 1056 unlock: 1057 mtx_unlock_spin(&sleepq_prof_lock); 1058 return; 1059 } 1060 1061 static void 1062 sleepq_prof_reset(void) 1063 { 1064 struct sleepq_prof *sp; 1065 int enabled; 1066 int i; 1067 1068 mtx_lock_spin(&sleepq_prof_lock); 1069 enabled = prof_enabled; 1070 prof_enabled = 0; 1071 for (i = 0; i < SC_TABLESIZE; i++) 1072 LIST_INIT(&sleepq_hash[i]); 1073 LIST_INIT(&sleepq_prof_free); 1074 for (i = 0; i < SLEEPQ_PROF_LOCATIONS; i++) { 1075 sp = &sleepq_profent[i]; 1076 sp->sp_wmesg = NULL; 1077 sp->sp_count = 0; 1078 LIST_INSERT_HEAD(&sleepq_prof_free, sp, sp_link); 1079 } 1080 prof_enabled = enabled; 1081 mtx_unlock_spin(&sleepq_prof_lock); 1082 } 1083 1084 static int 1085 enable_sleepq_prof(SYSCTL_HANDLER_ARGS) 1086 { 1087 int error, v; 1088 1089 v = prof_enabled; 1090 error = sysctl_handle_int(oidp, &v, v, req); 1091 if (error) 1092 return (error); 1093 if (req->newptr == NULL) 1094 return (error); 1095 if (v == prof_enabled) 1096 return (0); 1097 if (v == 1) 1098 sleepq_prof_reset(); 1099 mtx_lock_spin(&sleepq_prof_lock); 1100 prof_enabled = !!v; 1101 mtx_unlock_spin(&sleepq_prof_lock); 1102 1103 return (0); 1104 } 1105 1106 static int 1107 reset_sleepq_prof_stats(SYSCTL_HANDLER_ARGS) 1108 { 1109 int error, v; 1110 1111 v = 0; 1112 error = sysctl_handle_int(oidp, &v, 0, req); 1113 if (error) 1114 return (error); 1115 if (req->newptr == NULL) 1116 return (error); 1117 if (v == 0) 1118 return (0); 1119 sleepq_prof_reset(); 1120 1121 return (0); 1122 } 1123 1124 static int 1125 dump_sleepq_prof_stats(SYSCTL_HANDLER_ARGS) 1126 { 1127 struct sleepq_prof *sp; 1128 struct sbuf *sb; 1129 int enabled; 1130 int error; 1131 int i; 1132 1133 error = sysctl_wire_old_buffer(req, 0); 1134 if (error != 0) 1135 return (error); 1136 sb = sbuf_new_for_sysctl(NULL, NULL, SLEEPQ_SBUFSIZE, req); 1137 sbuf_printf(sb, "\nwmesg\tcount\n"); 1138 enabled = prof_enabled; 1139 mtx_lock_spin(&sleepq_prof_lock); 1140 prof_enabled = 0; 1141 mtx_unlock_spin(&sleepq_prof_lock); 1142 for (i = 0; i < SC_TABLESIZE; i++) { 1143 LIST_FOREACH(sp, &sleepq_hash[i], sp_link) { 1144 sbuf_printf(sb, "%s\t%ld\n", 1145 sp->sp_wmesg, sp->sp_count); 1146 } 1147 } 1148 mtx_lock_spin(&sleepq_prof_lock); 1149 prof_enabled = enabled; 1150 mtx_unlock_spin(&sleepq_prof_lock); 1151 1152 error = sbuf_finish(sb); 1153 sbuf_delete(sb); 1154 return (error); 1155 } 1156 1157 SYSCTL_PROC(_debug_sleepq, OID_AUTO, stats, CTLTYPE_STRING | CTLFLAG_RD, 1158 NULL, 0, dump_sleepq_prof_stats, "A", "Sleepqueue profiling statistics"); 1159 SYSCTL_PROC(_debug_sleepq, OID_AUTO, reset, CTLTYPE_INT | CTLFLAG_RW, 1160 NULL, 0, reset_sleepq_prof_stats, "I", 1161 "Reset sleepqueue profiling statistics"); 1162 SYSCTL_PROC(_debug_sleepq, OID_AUTO, enable, CTLTYPE_INT | CTLFLAG_RW, 1163 NULL, 0, enable_sleepq_prof, "I", "Enable sleepqueue profiling"); 1164 #endif 1165 1166 #ifdef DDB 1167 DB_SHOW_COMMAND(sleepq, db_show_sleepqueue) 1168 { 1169 struct sleepqueue_chain *sc; 1170 struct sleepqueue *sq; 1171 #ifdef INVARIANTS 1172 struct lock_object *lock; 1173 #endif 1174 struct thread *td; 1175 void *wchan; 1176 int i; 1177 1178 if (!have_addr) 1179 return; 1180 1181 /* 1182 * First, see if there is an active sleep queue for the wait channel 1183 * indicated by the address. 1184 */ 1185 wchan = (void *)addr; 1186 sc = SC_LOOKUP(wchan); 1187 LIST_FOREACH(sq, &sc->sc_queues, sq_hash) 1188 if (sq->sq_wchan == wchan) 1189 goto found; 1190 1191 /* 1192 * Second, see if there is an active sleep queue at the address 1193 * indicated. 1194 */ 1195 for (i = 0; i < SC_TABLESIZE; i++) 1196 LIST_FOREACH(sq, &sleepq_chains[i].sc_queues, sq_hash) { 1197 if (sq == (struct sleepqueue *)addr) 1198 goto found; 1199 } 1200 1201 db_printf("Unable to locate a sleep queue via %p\n", (void *)addr); 1202 return; 1203 found: 1204 db_printf("Wait channel: %p\n", sq->sq_wchan); 1205 db_printf("Queue type: %d\n", sq->sq_type); 1206 #ifdef INVARIANTS 1207 if (sq->sq_lock) { 1208 lock = sq->sq_lock; 1209 db_printf("Associated Interlock: %p - (%s) %s\n", lock, 1210 LOCK_CLASS(lock)->lc_name, lock->lo_name); 1211 } 1212 #endif 1213 db_printf("Blocked threads:\n"); 1214 for (i = 0; i < NR_SLEEPQS; i++) { 1215 db_printf("\nQueue[%d]:\n", i); 1216 if (TAILQ_EMPTY(&sq->sq_blocked[i])) 1217 db_printf("\tempty\n"); 1218 else 1219 TAILQ_FOREACH(td, &sq->sq_blocked[0], 1220 td_slpq) { 1221 db_printf("\t%p (tid %d, pid %d, \"%s\")\n", td, 1222 td->td_tid, td->td_proc->p_pid, 1223 td->td_name); 1224 } 1225 db_printf("(expected: %u)\n", sq->sq_blockedcnt[i]); 1226 } 1227 } 1228 1229 /* Alias 'show sleepqueue' to 'show sleepq'. */ 1230 DB_SHOW_ALIAS(sleepqueue, db_show_sleepqueue); 1231 #endif 1232