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 /* 412 * See if there are any pending signals for this thread. If not 413 * we can switch immediately. Otherwise do the signal processing 414 * directly. 415 */ 416 thread_lock(td); 417 if ((td->td_flags & (TDF_NEEDSIGCHK | TDF_NEEDSUSPCHK)) == 0) { 418 sleepq_switch(wchan, pri); 419 return (0); 420 } 421 stop_allowed = (td->td_flags & TDF_SBDRY) ? SIG_STOP_NOT_ALLOWED : 422 SIG_STOP_ALLOWED; 423 thread_unlock(td); 424 mtx_unlock_spin(&sc->sc_lock); 425 CTR3(KTR_PROC, "sleepq catching signals: thread %p (pid %ld, %s)", 426 (void *)td, (long)p->p_pid, td->td_name); 427 PROC_LOCK(p); 428 ps = p->p_sigacts; 429 mtx_lock(&ps->ps_mtx); 430 sig = cursig(td, stop_allowed); 431 if (sig == 0) { 432 mtx_unlock(&ps->ps_mtx); 433 ret = thread_suspend_check(1); 434 MPASS(ret == 0 || ret == EINTR || ret == ERESTART); 435 } else { 436 if (SIGISMEMBER(ps->ps_sigintr, sig)) 437 ret = EINTR; 438 else 439 ret = ERESTART; 440 mtx_unlock(&ps->ps_mtx); 441 } 442 /* 443 * Lock the per-process spinlock prior to dropping the PROC_LOCK 444 * to avoid a signal delivery race. PROC_LOCK, PROC_SLOCK, and 445 * thread_lock() are currently held in tdsendsignal(). 446 */ 447 PROC_SLOCK(p); 448 mtx_lock_spin(&sc->sc_lock); 449 PROC_UNLOCK(p); 450 thread_lock(td); 451 PROC_SUNLOCK(p); 452 if (ret == 0) { 453 sleepq_switch(wchan, pri); 454 return (0); 455 } 456 /* 457 * There were pending signals and this thread is still 458 * on the sleep queue, remove it from the sleep queue. 459 */ 460 if (TD_ON_SLEEPQ(td)) { 461 sq = sleepq_lookup(wchan); 462 if (sleepq_resume_thread(sq, td, 0)) { 463 #ifdef INVARIANTS 464 /* 465 * This thread hasn't gone to sleep yet, so it 466 * should not be swapped out. 467 */ 468 panic("not waking up swapper"); 469 #endif 470 } 471 } 472 mtx_unlock_spin(&sc->sc_lock); 473 MPASS(td->td_lock != &sc->sc_lock); 474 return (ret); 475 } 476 477 /* 478 * Switches to another thread if we are still asleep on a sleep queue. 479 * Returns with thread lock. 480 */ 481 static void 482 sleepq_switch(void *wchan, int pri) 483 { 484 struct sleepqueue_chain *sc; 485 struct sleepqueue *sq; 486 struct thread *td; 487 488 td = curthread; 489 sc = SC_LOOKUP(wchan); 490 mtx_assert(&sc->sc_lock, MA_OWNED); 491 THREAD_LOCK_ASSERT(td, MA_OWNED); 492 493 /* 494 * If we have a sleep queue, then we've already been woken up, so 495 * just return. 496 */ 497 if (td->td_sleepqueue != NULL) { 498 mtx_unlock_spin(&sc->sc_lock); 499 return; 500 } 501 502 /* 503 * If TDF_TIMEOUT is set, then our sleep has been timed out 504 * already but we are still on the sleep queue, so dequeue the 505 * thread and return. 506 */ 507 if (td->td_flags & TDF_TIMEOUT) { 508 MPASS(TD_ON_SLEEPQ(td)); 509 sq = sleepq_lookup(wchan); 510 if (sleepq_resume_thread(sq, td, 0)) { 511 #ifdef INVARIANTS 512 /* 513 * This thread hasn't gone to sleep yet, so it 514 * should not be swapped out. 515 */ 516 panic("not waking up swapper"); 517 #endif 518 } 519 mtx_unlock_spin(&sc->sc_lock); 520 return; 521 } 522 #ifdef SLEEPQUEUE_PROFILING 523 if (prof_enabled) 524 sleepq_profile(td->td_wmesg); 525 #endif 526 MPASS(td->td_sleepqueue == NULL); 527 sched_sleep(td, pri); 528 thread_lock_set(td, &sc->sc_lock); 529 TD_SET_SLEEPING(td); 530 mi_switch(SW_VOL | SWT_SLEEPQ, NULL); 531 KASSERT(TD_IS_RUNNING(td), ("running but not TDS_RUNNING")); 532 CTR3(KTR_PROC, "sleepq resume: thread %p (pid %ld, %s)", 533 (void *)td, (long)td->td_proc->p_pid, (void *)td->td_name); 534 } 535 536 /* 537 * Check to see if we timed out. 538 */ 539 static int 540 sleepq_check_timeout(void) 541 { 542 struct thread *td; 543 544 td = curthread; 545 THREAD_LOCK_ASSERT(td, MA_OWNED); 546 547 /* 548 * If TDF_TIMEOUT is set, we timed out. 549 */ 550 if (td->td_flags & TDF_TIMEOUT) { 551 td->td_flags &= ~TDF_TIMEOUT; 552 return (EWOULDBLOCK); 553 } 554 555 /* 556 * If TDF_TIMOFAIL is set, the timeout ran after we had 557 * already been woken up. 558 */ 559 if (td->td_flags & TDF_TIMOFAIL) 560 td->td_flags &= ~TDF_TIMOFAIL; 561 562 /* 563 * If callout_stop() fails, then the timeout is running on 564 * another CPU, so synchronize with it to avoid having it 565 * accidentally wake up a subsequent sleep. 566 */ 567 else if (callout_stop(&td->td_slpcallout) == 0) { 568 td->td_flags |= TDF_TIMEOUT; 569 TD_SET_SLEEPING(td); 570 mi_switch(SW_INVOL | SWT_SLEEPQTIMO, NULL); 571 } 572 return (0); 573 } 574 575 /* 576 * Check to see if we were awoken by a signal. 577 */ 578 static int 579 sleepq_check_signals(void) 580 { 581 struct thread *td; 582 583 td = curthread; 584 THREAD_LOCK_ASSERT(td, MA_OWNED); 585 586 /* We are no longer in an interruptible sleep. */ 587 if (td->td_flags & TDF_SINTR) 588 td->td_flags &= ~(TDF_SINTR | TDF_SBDRY); 589 590 if (td->td_flags & TDF_SLEEPABORT) { 591 td->td_flags &= ~TDF_SLEEPABORT; 592 return (td->td_intrval); 593 } 594 595 return (0); 596 } 597 598 /* 599 * Block the current thread until it is awakened from its sleep queue. 600 */ 601 void 602 sleepq_wait(void *wchan, int pri) 603 { 604 struct thread *td; 605 606 td = curthread; 607 MPASS(!(td->td_flags & TDF_SINTR)); 608 thread_lock(td); 609 sleepq_switch(wchan, pri); 610 thread_unlock(td); 611 } 612 613 /* 614 * Block the current thread until it is awakened from its sleep queue 615 * or it is interrupted by a signal. 616 */ 617 int 618 sleepq_wait_sig(void *wchan, int pri) 619 { 620 int rcatch; 621 int rval; 622 623 rcatch = sleepq_catch_signals(wchan, pri); 624 rval = sleepq_check_signals(); 625 thread_unlock(curthread); 626 if (rcatch) 627 return (rcatch); 628 return (rval); 629 } 630 631 /* 632 * Block the current thread until it is awakened from its sleep queue 633 * or it times out while waiting. 634 */ 635 int 636 sleepq_timedwait(void *wchan, int pri) 637 { 638 struct thread *td; 639 int rval; 640 641 td = curthread; 642 MPASS(!(td->td_flags & TDF_SINTR)); 643 thread_lock(td); 644 sleepq_switch(wchan, pri); 645 rval = sleepq_check_timeout(); 646 thread_unlock(td); 647 648 return (rval); 649 } 650 651 /* 652 * Block the current thread until it is awakened from its sleep queue, 653 * it is interrupted by a signal, or it times out waiting to be awakened. 654 */ 655 int 656 sleepq_timedwait_sig(void *wchan, int pri) 657 { 658 int rcatch, rvalt, rvals; 659 660 rcatch = sleepq_catch_signals(wchan, pri); 661 rvalt = sleepq_check_timeout(); 662 rvals = sleepq_check_signals(); 663 thread_unlock(curthread); 664 if (rcatch) 665 return (rcatch); 666 if (rvals) 667 return (rvals); 668 return (rvalt); 669 } 670 671 /* 672 * Returns the type of sleepqueue given a waitchannel. 673 */ 674 int 675 sleepq_type(void *wchan) 676 { 677 struct sleepqueue *sq; 678 int type; 679 680 MPASS(wchan != NULL); 681 682 sleepq_lock(wchan); 683 sq = sleepq_lookup(wchan); 684 if (sq == NULL) { 685 sleepq_release(wchan); 686 return (-1); 687 } 688 type = sq->sq_type; 689 sleepq_release(wchan); 690 return (type); 691 } 692 693 /* 694 * Removes a thread from a sleep queue and makes it 695 * runnable. 696 */ 697 static int 698 sleepq_resume_thread(struct sleepqueue *sq, struct thread *td, int pri) 699 { 700 struct sleepqueue_chain *sc; 701 702 MPASS(td != NULL); 703 MPASS(sq->sq_wchan != NULL); 704 MPASS(td->td_wchan == sq->sq_wchan); 705 MPASS(td->td_sqqueue < NR_SLEEPQS && td->td_sqqueue >= 0); 706 THREAD_LOCK_ASSERT(td, MA_OWNED); 707 sc = SC_LOOKUP(sq->sq_wchan); 708 mtx_assert(&sc->sc_lock, MA_OWNED); 709 710 /* Remove the thread from the queue. */ 711 sq->sq_blockedcnt[td->td_sqqueue]--; 712 TAILQ_REMOVE(&sq->sq_blocked[td->td_sqqueue], td, td_slpq); 713 714 /* 715 * Get a sleep queue for this thread. If this is the last waiter, 716 * use the queue itself and take it out of the chain, otherwise, 717 * remove a queue from the free list. 718 */ 719 if (LIST_EMPTY(&sq->sq_free)) { 720 td->td_sleepqueue = sq; 721 #ifdef INVARIANTS 722 sq->sq_wchan = NULL; 723 #endif 724 #ifdef SLEEPQUEUE_PROFILING 725 sc->sc_depth--; 726 #endif 727 } else 728 td->td_sleepqueue = LIST_FIRST(&sq->sq_free); 729 LIST_REMOVE(td->td_sleepqueue, sq_hash); 730 731 td->td_wmesg = NULL; 732 td->td_wchan = NULL; 733 td->td_flags &= ~(TDF_SINTR | TDF_SBDRY); 734 735 CTR3(KTR_PROC, "sleepq_wakeup: thread %p (pid %ld, %s)", 736 (void *)td, (long)td->td_proc->p_pid, td->td_name); 737 738 /* Adjust priority if requested. */ 739 MPASS(pri == 0 || (pri >= PRI_MIN && pri <= PRI_MAX)); 740 if (pri != 0 && td->td_priority > pri) 741 sched_prio(td, pri); 742 743 /* 744 * Note that thread td might not be sleeping if it is running 745 * sleepq_catch_signals() on another CPU or is blocked on its 746 * proc lock to check signals. There's no need to mark the 747 * thread runnable in that case. 748 */ 749 if (TD_IS_SLEEPING(td)) { 750 TD_CLR_SLEEPING(td); 751 return (setrunnable(td)); 752 } 753 return (0); 754 } 755 756 #ifdef INVARIANTS 757 /* 758 * UMA zone item deallocator. 759 */ 760 static void 761 sleepq_dtor(void *mem, int size, void *arg) 762 { 763 struct sleepqueue *sq; 764 int i; 765 766 sq = mem; 767 for (i = 0; i < NR_SLEEPQS; i++) { 768 MPASS(TAILQ_EMPTY(&sq->sq_blocked[i])); 769 MPASS(sq->sq_blockedcnt[i] == 0); 770 } 771 } 772 #endif 773 774 /* 775 * UMA zone item initializer. 776 */ 777 static int 778 sleepq_init(void *mem, int size, int flags) 779 { 780 struct sleepqueue *sq; 781 int i; 782 783 bzero(mem, size); 784 sq = mem; 785 for (i = 0; i < NR_SLEEPQS; i++) { 786 TAILQ_INIT(&sq->sq_blocked[i]); 787 sq->sq_blockedcnt[i] = 0; 788 } 789 LIST_INIT(&sq->sq_free); 790 return (0); 791 } 792 793 /* 794 * Find the highest priority thread sleeping on a wait channel and resume it. 795 */ 796 int 797 sleepq_signal(void *wchan, int flags, int pri, int queue) 798 { 799 struct sleepqueue *sq; 800 struct thread *td, *besttd; 801 int wakeup_swapper; 802 803 CTR2(KTR_PROC, "sleepq_signal(%p, %d)", wchan, flags); 804 KASSERT(wchan != NULL, ("%s: invalid NULL wait channel", __func__)); 805 MPASS((queue >= 0) && (queue < NR_SLEEPQS)); 806 sq = sleepq_lookup(wchan); 807 if (sq == NULL) 808 return (0); 809 KASSERT(sq->sq_type == (flags & SLEEPQ_TYPE), 810 ("%s: mismatch between sleep/wakeup and cv_*", __func__)); 811 812 /* 813 * Find the highest priority thread on the queue. If there is a 814 * tie, use the thread that first appears in the queue as it has 815 * been sleeping the longest since threads are always added to 816 * the tail of sleep queues. 817 */ 818 besttd = NULL; 819 TAILQ_FOREACH(td, &sq->sq_blocked[queue], td_slpq) { 820 if (besttd == NULL || td->td_priority < besttd->td_priority) 821 besttd = td; 822 } 823 MPASS(besttd != NULL); 824 thread_lock(besttd); 825 wakeup_swapper = sleepq_resume_thread(sq, besttd, pri); 826 thread_unlock(besttd); 827 return (wakeup_swapper); 828 } 829 830 /* 831 * Resume all threads sleeping on a specified wait channel. 832 */ 833 int 834 sleepq_broadcast(void *wchan, int flags, int pri, int queue) 835 { 836 struct sleepqueue *sq; 837 struct thread *td, *tdn; 838 int wakeup_swapper; 839 840 CTR2(KTR_PROC, "sleepq_broadcast(%p, %d)", wchan, flags); 841 KASSERT(wchan != NULL, ("%s: invalid NULL wait channel", __func__)); 842 MPASS((queue >= 0) && (queue < NR_SLEEPQS)); 843 sq = sleepq_lookup(wchan); 844 if (sq == NULL) 845 return (0); 846 KASSERT(sq->sq_type == (flags & SLEEPQ_TYPE), 847 ("%s: mismatch between sleep/wakeup and cv_*", __func__)); 848 849 /* Resume all blocked threads on the sleep queue. */ 850 wakeup_swapper = 0; 851 TAILQ_FOREACH_SAFE(td, &sq->sq_blocked[queue], td_slpq, tdn) { 852 thread_lock(td); 853 if (sleepq_resume_thread(sq, td, pri)) 854 wakeup_swapper = 1; 855 thread_unlock(td); 856 } 857 return (wakeup_swapper); 858 } 859 860 /* 861 * Time sleeping threads out. When the timeout expires, the thread is 862 * removed from the sleep queue and made runnable if it is still asleep. 863 */ 864 static void 865 sleepq_timeout(void *arg) 866 { 867 struct sleepqueue_chain *sc; 868 struct sleepqueue *sq; 869 struct thread *td; 870 void *wchan; 871 int wakeup_swapper; 872 873 td = arg; 874 wakeup_swapper = 0; 875 CTR3(KTR_PROC, "sleepq_timeout: thread %p (pid %ld, %s)", 876 (void *)td, (long)td->td_proc->p_pid, (void *)td->td_name); 877 878 /* 879 * First, see if the thread is asleep and get the wait channel if 880 * it is. 881 */ 882 thread_lock(td); 883 if (TD_IS_SLEEPING(td) && TD_ON_SLEEPQ(td)) { 884 wchan = td->td_wchan; 885 sc = SC_LOOKUP(wchan); 886 THREAD_LOCKPTR_ASSERT(td, &sc->sc_lock); 887 sq = sleepq_lookup(wchan); 888 MPASS(sq != NULL); 889 td->td_flags |= TDF_TIMEOUT; 890 wakeup_swapper = sleepq_resume_thread(sq, td, 0); 891 thread_unlock(td); 892 if (wakeup_swapper) 893 kick_proc0(); 894 return; 895 } 896 897 /* 898 * If the thread is on the SLEEPQ but isn't sleeping yet, it 899 * can either be on another CPU in between sleepq_add() and 900 * one of the sleepq_*wait*() routines or it can be in 901 * sleepq_catch_signals(). 902 */ 903 if (TD_ON_SLEEPQ(td)) { 904 td->td_flags |= TDF_TIMEOUT; 905 thread_unlock(td); 906 return; 907 } 908 909 /* 910 * Now check for the edge cases. First, if TDF_TIMEOUT is set, 911 * then the other thread has already yielded to us, so clear 912 * the flag and resume it. If TDF_TIMEOUT is not set, then the 913 * we know that the other thread is not on a sleep queue, but it 914 * hasn't resumed execution yet. In that case, set TDF_TIMOFAIL 915 * to let it know that the timeout has already run and doesn't 916 * need to be canceled. 917 */ 918 if (td->td_flags & TDF_TIMEOUT) { 919 MPASS(TD_IS_SLEEPING(td)); 920 td->td_flags &= ~TDF_TIMEOUT; 921 TD_CLR_SLEEPING(td); 922 wakeup_swapper = setrunnable(td); 923 } else 924 td->td_flags |= TDF_TIMOFAIL; 925 thread_unlock(td); 926 if (wakeup_swapper) 927 kick_proc0(); 928 } 929 930 /* 931 * Resumes a specific thread from the sleep queue associated with a specific 932 * wait channel if it is on that queue. 933 */ 934 void 935 sleepq_remove(struct thread *td, void *wchan) 936 { 937 struct sleepqueue *sq; 938 int wakeup_swapper; 939 940 /* 941 * Look up the sleep queue for this wait channel, then re-check 942 * that the thread is asleep on that channel, if it is not, then 943 * bail. 944 */ 945 MPASS(wchan != NULL); 946 sleepq_lock(wchan); 947 sq = sleepq_lookup(wchan); 948 /* 949 * We can not lock the thread here as it may be sleeping on a 950 * different sleepq. However, holding the sleepq lock for this 951 * wchan can guarantee that we do not miss a wakeup for this 952 * channel. The asserts below will catch any false positives. 953 */ 954 if (!TD_ON_SLEEPQ(td) || td->td_wchan != wchan) { 955 sleepq_release(wchan); 956 return; 957 } 958 /* Thread is asleep on sleep queue sq, so wake it up. */ 959 thread_lock(td); 960 MPASS(sq != NULL); 961 MPASS(td->td_wchan == wchan); 962 wakeup_swapper = sleepq_resume_thread(sq, td, 0); 963 thread_unlock(td); 964 sleepq_release(wchan); 965 if (wakeup_swapper) 966 kick_proc0(); 967 } 968 969 /* 970 * Abort a thread as if an interrupt had occurred. Only abort 971 * interruptible waits (unfortunately it isn't safe to abort others). 972 */ 973 int 974 sleepq_abort(struct thread *td, int intrval) 975 { 976 struct sleepqueue *sq; 977 void *wchan; 978 979 THREAD_LOCK_ASSERT(td, MA_OWNED); 980 MPASS(TD_ON_SLEEPQ(td)); 981 MPASS(td->td_flags & TDF_SINTR); 982 MPASS(intrval == EINTR || intrval == ERESTART); 983 984 /* 985 * If the TDF_TIMEOUT flag is set, just leave. A 986 * timeout is scheduled anyhow. 987 */ 988 if (td->td_flags & TDF_TIMEOUT) 989 return (0); 990 991 CTR3(KTR_PROC, "sleepq_abort: thread %p (pid %ld, %s)", 992 (void *)td, (long)td->td_proc->p_pid, (void *)td->td_name); 993 td->td_intrval = intrval; 994 td->td_flags |= TDF_SLEEPABORT; 995 /* 996 * If the thread has not slept yet it will find the signal in 997 * sleepq_catch_signals() and call sleepq_resume_thread. Otherwise 998 * we have to do it here. 999 */ 1000 if (!TD_IS_SLEEPING(td)) 1001 return (0); 1002 wchan = td->td_wchan; 1003 MPASS(wchan != NULL); 1004 sq = sleepq_lookup(wchan); 1005 MPASS(sq != NULL); 1006 1007 /* Thread is asleep on sleep queue sq, so wake it up. */ 1008 return (sleepq_resume_thread(sq, td, 0)); 1009 } 1010 1011 #ifdef SLEEPQUEUE_PROFILING 1012 #define SLEEPQ_PROF_LOCATIONS 1024 1013 #define SLEEPQ_SBUFSIZE (40 * 512) 1014 struct sleepq_prof { 1015 LIST_ENTRY(sleepq_prof) sp_link; 1016 const char *sp_wmesg; 1017 long sp_count; 1018 }; 1019 1020 LIST_HEAD(sqphead, sleepq_prof); 1021 1022 struct sqphead sleepq_prof_free; 1023 struct sqphead sleepq_hash[SC_TABLESIZE]; 1024 static struct sleepq_prof sleepq_profent[SLEEPQ_PROF_LOCATIONS]; 1025 static struct mtx sleepq_prof_lock; 1026 MTX_SYSINIT(sleepq_prof_lock, &sleepq_prof_lock, "sleepq_prof", MTX_SPIN); 1027 1028 static void 1029 sleepq_profile(const char *wmesg) 1030 { 1031 struct sleepq_prof *sp; 1032 1033 mtx_lock_spin(&sleepq_prof_lock); 1034 if (prof_enabled == 0) 1035 goto unlock; 1036 LIST_FOREACH(sp, &sleepq_hash[SC_HASH(wmesg)], sp_link) 1037 if (sp->sp_wmesg == wmesg) 1038 goto done; 1039 sp = LIST_FIRST(&sleepq_prof_free); 1040 if (sp == NULL) 1041 goto unlock; 1042 sp->sp_wmesg = wmesg; 1043 LIST_REMOVE(sp, sp_link); 1044 LIST_INSERT_HEAD(&sleepq_hash[SC_HASH(wmesg)], sp, sp_link); 1045 done: 1046 sp->sp_count++; 1047 unlock: 1048 mtx_unlock_spin(&sleepq_prof_lock); 1049 return; 1050 } 1051 1052 static void 1053 sleepq_prof_reset(void) 1054 { 1055 struct sleepq_prof *sp; 1056 int enabled; 1057 int i; 1058 1059 mtx_lock_spin(&sleepq_prof_lock); 1060 enabled = prof_enabled; 1061 prof_enabled = 0; 1062 for (i = 0; i < SC_TABLESIZE; i++) 1063 LIST_INIT(&sleepq_hash[i]); 1064 LIST_INIT(&sleepq_prof_free); 1065 for (i = 0; i < SLEEPQ_PROF_LOCATIONS; i++) { 1066 sp = &sleepq_profent[i]; 1067 sp->sp_wmesg = NULL; 1068 sp->sp_count = 0; 1069 LIST_INSERT_HEAD(&sleepq_prof_free, sp, sp_link); 1070 } 1071 prof_enabled = enabled; 1072 mtx_unlock_spin(&sleepq_prof_lock); 1073 } 1074 1075 static int 1076 enable_sleepq_prof(SYSCTL_HANDLER_ARGS) 1077 { 1078 int error, v; 1079 1080 v = prof_enabled; 1081 error = sysctl_handle_int(oidp, &v, v, req); 1082 if (error) 1083 return (error); 1084 if (req->newptr == NULL) 1085 return (error); 1086 if (v == prof_enabled) 1087 return (0); 1088 if (v == 1) 1089 sleepq_prof_reset(); 1090 mtx_lock_spin(&sleepq_prof_lock); 1091 prof_enabled = !!v; 1092 mtx_unlock_spin(&sleepq_prof_lock); 1093 1094 return (0); 1095 } 1096 1097 static int 1098 reset_sleepq_prof_stats(SYSCTL_HANDLER_ARGS) 1099 { 1100 int error, v; 1101 1102 v = 0; 1103 error = sysctl_handle_int(oidp, &v, 0, req); 1104 if (error) 1105 return (error); 1106 if (req->newptr == NULL) 1107 return (error); 1108 if (v == 0) 1109 return (0); 1110 sleepq_prof_reset(); 1111 1112 return (0); 1113 } 1114 1115 static int 1116 dump_sleepq_prof_stats(SYSCTL_HANDLER_ARGS) 1117 { 1118 static int multiplier = 1; 1119 struct sleepq_prof *sp; 1120 struct sbuf *sb; 1121 int enabled; 1122 int error; 1123 int i; 1124 1125 retry_sbufops: 1126 sb = sbuf_new(NULL, NULL, SLEEPQ_SBUFSIZE * multiplier, SBUF_FIXEDLEN); 1127 sbuf_printf(sb, "\nwmesg\tcount\n"); 1128 enabled = prof_enabled; 1129 mtx_lock_spin(&sleepq_prof_lock); 1130 prof_enabled = 0; 1131 mtx_unlock_spin(&sleepq_prof_lock); 1132 for (i = 0; i < SC_TABLESIZE; i++) { 1133 LIST_FOREACH(sp, &sleepq_hash[i], sp_link) { 1134 sbuf_printf(sb, "%s\t%ld\n", 1135 sp->sp_wmesg, sp->sp_count); 1136 if (sbuf_overflowed(sb)) { 1137 sbuf_delete(sb); 1138 multiplier++; 1139 goto retry_sbufops; 1140 } 1141 } 1142 } 1143 mtx_lock_spin(&sleepq_prof_lock); 1144 prof_enabled = enabled; 1145 mtx_unlock_spin(&sleepq_prof_lock); 1146 1147 sbuf_finish(sb); 1148 error = SYSCTL_OUT(req, sbuf_data(sb), sbuf_len(sb) + 1); 1149 sbuf_delete(sb); 1150 return (error); 1151 } 1152 1153 SYSCTL_PROC(_debug_sleepq, OID_AUTO, stats, CTLTYPE_STRING | CTLFLAG_RD, 1154 NULL, 0, dump_sleepq_prof_stats, "A", "Sleepqueue profiling statistics"); 1155 SYSCTL_PROC(_debug_sleepq, OID_AUTO, reset, CTLTYPE_INT | CTLFLAG_RW, 1156 NULL, 0, reset_sleepq_prof_stats, "I", 1157 "Reset sleepqueue profiling statistics"); 1158 SYSCTL_PROC(_debug_sleepq, OID_AUTO, enable, CTLTYPE_INT | CTLFLAG_RW, 1159 NULL, 0, enable_sleepq_prof, "I", "Enable sleepqueue profiling"); 1160 #endif 1161 1162 #ifdef DDB 1163 DB_SHOW_COMMAND(sleepq, db_show_sleepqueue) 1164 { 1165 struct sleepqueue_chain *sc; 1166 struct sleepqueue *sq; 1167 #ifdef INVARIANTS 1168 struct lock_object *lock; 1169 #endif 1170 struct thread *td; 1171 void *wchan; 1172 int i; 1173 1174 if (!have_addr) 1175 return; 1176 1177 /* 1178 * First, see if there is an active sleep queue for the wait channel 1179 * indicated by the address. 1180 */ 1181 wchan = (void *)addr; 1182 sc = SC_LOOKUP(wchan); 1183 LIST_FOREACH(sq, &sc->sc_queues, sq_hash) 1184 if (sq->sq_wchan == wchan) 1185 goto found; 1186 1187 /* 1188 * Second, see if there is an active sleep queue at the address 1189 * indicated. 1190 */ 1191 for (i = 0; i < SC_TABLESIZE; i++) 1192 LIST_FOREACH(sq, &sleepq_chains[i].sc_queues, sq_hash) { 1193 if (sq == (struct sleepqueue *)addr) 1194 goto found; 1195 } 1196 1197 db_printf("Unable to locate a sleep queue via %p\n", (void *)addr); 1198 return; 1199 found: 1200 db_printf("Wait channel: %p\n", sq->sq_wchan); 1201 db_printf("Queue type: %d\n", sq->sq_type); 1202 #ifdef INVARIANTS 1203 if (sq->sq_lock) { 1204 lock = sq->sq_lock; 1205 db_printf("Associated Interlock: %p - (%s) %s\n", lock, 1206 LOCK_CLASS(lock)->lc_name, lock->lo_name); 1207 } 1208 #endif 1209 db_printf("Blocked threads:\n"); 1210 for (i = 0; i < NR_SLEEPQS; i++) { 1211 db_printf("\nQueue[%d]:\n", i); 1212 if (TAILQ_EMPTY(&sq->sq_blocked[i])) 1213 db_printf("\tempty\n"); 1214 else 1215 TAILQ_FOREACH(td, &sq->sq_blocked[0], 1216 td_slpq) { 1217 db_printf("\t%p (tid %d, pid %d, \"%s\")\n", td, 1218 td->td_tid, td->td_proc->p_pid, 1219 td->td_name); 1220 } 1221 db_printf("(expected: %u)\n", sq->sq_blockedcnt[i]); 1222 } 1223 } 1224 1225 /* Alias 'show sleepqueue' to 'show sleepq'. */ 1226 DB_SHOW_ALIAS(sleepqueue, db_show_sleepqueue); 1227 #endif 1228