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 "opt_sleepqueue_profiling.h" 63 64 #include <sys/cdefs.h> 65 __FBSDID("$FreeBSD$"); 66 67 #include <sys/param.h> 68 #include <sys/systm.h> 69 #include <sys/lock.h> 70 #include <sys/kernel.h> 71 #include <sys/ktr.h> 72 #include <sys/malloc.h> 73 #include <sys/mutex.h> 74 #include <sys/proc.h> 75 #include <sys/sched.h> 76 #include <sys/signalvar.h> 77 #include <sys/sleepqueue.h> 78 #include <sys/sysctl.h> 79 80 /* 81 * Constants for the hash table of sleep queue chains. These constants are 82 * the same ones that 4BSD (and possibly earlier versions of BSD) used. 83 * Basically, we ignore the lower 8 bits of the address since most wait 84 * channel pointers are aligned and only look at the next 7 bits for the 85 * hash. SC_TABLESIZE must be a power of two for SC_MASK to work properly. 86 */ 87 #define SC_TABLESIZE 128 /* Must be power of 2. */ 88 #define SC_MASK (SC_TABLESIZE - 1) 89 #define SC_SHIFT 8 90 #define SC_HASH(wc) (((uintptr_t)(wc) >> SC_SHIFT) & SC_MASK) 91 #define SC_LOOKUP(wc) &sleepq_chains[SC_HASH(wc)] 92 93 /* 94 * There two different lists of sleep queues. Both lists are connected 95 * via the sq_hash entries. The first list is the sleep queue chain list 96 * that a sleep queue is on when it is attached to a wait channel. The 97 * second list is the free list hung off of a sleep queue that is attached 98 * to a wait channel. 99 * 100 * Each sleep queue also contains the wait channel it is attached to, the 101 * list of threads blocked on that wait channel, flags specific to the 102 * wait channel, and the lock used to synchronize with a wait channel. 103 * The flags are used to catch mismatches between the various consumers 104 * of the sleep queue API (e.g. sleep/wakeup and condition variables). 105 * The lock pointer is only used when invariants are enabled for various 106 * debugging checks. 107 * 108 * Locking key: 109 * c - sleep queue chain lock 110 */ 111 struct sleepqueue { 112 TAILQ_HEAD(, thread) sq_blocked; /* (c) Blocked threads. */ 113 LIST_ENTRY(sleepqueue) sq_hash; /* (c) Chain and free list. */ 114 LIST_HEAD(, sleepqueue) sq_free; /* (c) Free queues. */ 115 void *sq_wchan; /* (c) Wait channel. */ 116 int sq_flags; /* (c) Flags. */ 117 #ifdef INVARIANTS 118 struct mtx *sq_lock; /* (c) Associated lock. */ 119 #endif 120 }; 121 122 struct sleepqueue_chain { 123 LIST_HEAD(, sleepqueue) sc_queues; /* List of sleep queues. */ 124 struct mtx sc_lock; /* Spin lock for this chain. */ 125 #ifdef SLEEPQUEUE_PROFILING 126 u_int sc_depth; /* Length of sc_queues. */ 127 u_int sc_max_depth; /* Max length of sc_queues. */ 128 #endif 129 }; 130 131 #ifdef SLEEPQUEUE_PROFILING 132 u_int sleepq_max_depth; 133 SYSCTL_NODE(_debug, OID_AUTO, sleepq, CTLFLAG_RD, 0, "sleepq profiling"); 134 SYSCTL_NODE(_debug_sleepq, OID_AUTO, chains, CTLFLAG_RD, 0, 135 "sleepq chain stats"); 136 SYSCTL_UINT(_debug_sleepq, OID_AUTO, max_depth, CTLFLAG_RD, &sleepq_max_depth, 137 0, "maxmimum depth achieved of a single chain"); 138 #endif 139 static struct sleepqueue_chain sleepq_chains[SC_TABLESIZE]; 140 141 MALLOC_DEFINE(M_SLEEPQUEUE, "sleep queues", "sleep queues"); 142 143 /* 144 * Prototypes for non-exported routines. 145 */ 146 static int sleepq_check_timeout(void); 147 static void sleepq_switch(void *wchan); 148 static void sleepq_timeout(void *arg); 149 static void sleepq_remove_thread(struct sleepqueue *sq, struct thread *td); 150 static void sleepq_resume_thread(struct thread *td, int pri); 151 152 /* 153 * Early initialization of sleep queues that is called from the sleepinit() 154 * SYSINIT. 155 */ 156 void 157 init_sleepqueues(void) 158 { 159 #ifdef SLEEPQUEUE_PROFILING 160 struct sysctl_oid *chain_oid; 161 char chain_name[10]; 162 #endif 163 int i; 164 165 for (i = 0; i < SC_TABLESIZE; i++) { 166 LIST_INIT(&sleepq_chains[i].sc_queues); 167 mtx_init(&sleepq_chains[i].sc_lock, "sleepq chain", NULL, 168 MTX_SPIN); 169 #ifdef SLEEPQUEUE_PROFILING 170 snprintf(chain_name, sizeof(chain_name), "%d", i); 171 chain_oid = SYSCTL_ADD_NODE(NULL, 172 SYSCTL_STATIC_CHILDREN(_debug_sleepq_chains), OID_AUTO, 173 chain_name, CTLFLAG_RD, NULL, "sleepq chain stats"); 174 SYSCTL_ADD_UINT(NULL, SYSCTL_CHILDREN(chain_oid), OID_AUTO, 175 "depth", CTLFLAG_RD, &sleepq_chains[i].sc_depth, 0, NULL); 176 SYSCTL_ADD_UINT(NULL, SYSCTL_CHILDREN(chain_oid), OID_AUTO, 177 "max_depth", CTLFLAG_RD, &sleepq_chains[i].sc_max_depth, 0, 178 NULL); 179 #endif 180 } 181 thread0.td_sleepqueue = sleepq_alloc(); 182 } 183 184 /* 185 * Malloc and initialize a new sleep queue for a new thread. 186 */ 187 struct sleepqueue * 188 sleepq_alloc(void) 189 { 190 struct sleepqueue *sq; 191 192 sq = malloc(sizeof(struct sleepqueue), M_SLEEPQUEUE, M_WAITOK | M_ZERO); 193 TAILQ_INIT(&sq->sq_blocked); 194 LIST_INIT(&sq->sq_free); 195 return (sq); 196 } 197 198 /* 199 * Free a sleep queue when a thread is destroyed. 200 */ 201 void 202 sleepq_free(struct sleepqueue *sq) 203 { 204 205 MPASS(sq != NULL); 206 MPASS(TAILQ_EMPTY(&sq->sq_blocked)); 207 free(sq, M_SLEEPQUEUE); 208 } 209 210 /* 211 * Look up the sleep queue associated with a given wait channel in the hash 212 * table locking the associated sleep queue chain. Return holdind the sleep 213 * queue chain lock. If no queue is found in the table, NULL is returned. 214 */ 215 struct sleepqueue * 216 sleepq_lookup(void *wchan) 217 { 218 struct sleepqueue_chain *sc; 219 struct sleepqueue *sq; 220 221 KASSERT(wchan != NULL, ("%s: invalid NULL wait channel", __func__)); 222 sc = SC_LOOKUP(wchan); 223 mtx_lock_spin(&sc->sc_lock); 224 LIST_FOREACH(sq, &sc->sc_queues, sq_hash) 225 if (sq->sq_wchan == wchan) 226 return (sq); 227 return (NULL); 228 } 229 230 /* 231 * Unlock the sleep queue chain associated with a given wait channel. 232 */ 233 void 234 sleepq_release(void *wchan) 235 { 236 struct sleepqueue_chain *sc; 237 238 sc = SC_LOOKUP(wchan); 239 mtx_unlock_spin(&sc->sc_lock); 240 } 241 242 /* 243 * Places the current thread on the sleepqueue for the specified wait 244 * channel. If INVARIANTS is enabled, then it associates the passed in 245 * lock with the sleepq to make sure it is held when that sleep queue is 246 * woken up. 247 */ 248 void 249 sleepq_add(struct sleepqueue *sq, void *wchan, struct mtx *lock, 250 const char *wmesg, int flags) 251 { 252 struct sleepqueue_chain *sc; 253 struct thread *td, *td1; 254 255 td = curthread; 256 sc = SC_LOOKUP(wchan); 257 mtx_assert(&sc->sc_lock, MA_OWNED); 258 MPASS(td->td_sleepqueue != NULL); 259 MPASS(wchan != NULL); 260 261 /* If the passed in sleep queue is NULL, use this thread's queue. */ 262 if (sq == NULL) { 263 #ifdef SLEEPQUEUE_PROFILING 264 sc->sc_depth++; 265 if (sc->sc_depth > sc->sc_max_depth) { 266 sc->sc_max_depth = sc->sc_depth; 267 if (sc->sc_max_depth > sleepq_max_depth) 268 sleepq_max_depth = sc->sc_max_depth; 269 } 270 #endif 271 sq = td->td_sleepqueue; 272 LIST_INSERT_HEAD(&sc->sc_queues, sq, sq_hash); 273 KASSERT(TAILQ_EMPTY(&sq->sq_blocked), 274 ("thread's sleep queue has a non-empty queue")); 275 KASSERT(LIST_EMPTY(&sq->sq_free), 276 ("thread's sleep queue has a non-empty free list")); 277 KASSERT(sq->sq_wchan == NULL, ("stale sq_wchan pointer")); 278 sq->sq_wchan = wchan; 279 #ifdef INVARIANTS 280 sq->sq_lock = lock; 281 #endif 282 sq->sq_flags = flags; 283 TAILQ_INSERT_TAIL(&sq->sq_blocked, td, td_slpq); 284 } else { 285 MPASS(wchan == sq->sq_wchan); 286 MPASS(lock == sq->sq_lock); 287 TAILQ_FOREACH(td1, &sq->sq_blocked, td_slpq) 288 if (td1->td_priority > td->td_priority) 289 break; 290 if (td1 != NULL) 291 TAILQ_INSERT_BEFORE(td1, td, td_slpq); 292 else 293 TAILQ_INSERT_TAIL(&sq->sq_blocked, td, td_slpq); 294 LIST_INSERT_HEAD(&sq->sq_free, td->td_sleepqueue, sq_hash); 295 } 296 td->td_sleepqueue = NULL; 297 mtx_lock_spin(&sched_lock); 298 td->td_wchan = wchan; 299 td->td_wmesg = wmesg; 300 mtx_unlock_spin(&sched_lock); 301 } 302 303 /* 304 * Sets a timeout that will remove the current thread from the specified 305 * sleep queue after timo ticks if the thread has not already been awakened. 306 */ 307 void 308 sleepq_set_timeout(void *wchan, int timo) 309 { 310 struct sleepqueue_chain *sc; 311 struct thread *td; 312 313 td = curthread; 314 sc = SC_LOOKUP(wchan); 315 mtx_assert(&sc->sc_lock, MA_OWNED); 316 MPASS(TD_ON_SLEEPQ(td)); 317 MPASS(td->td_sleepqueue == NULL); 318 MPASS(wchan != NULL); 319 callout_reset(&td->td_slpcallout, timo, sleepq_timeout, td); 320 } 321 322 /* 323 * Marks the pending sleep of the current thread as interruptible and 324 * makes an initial check for pending signals before putting a thread 325 * to sleep. 326 */ 327 int 328 sleepq_catch_signals(void *wchan) 329 { 330 struct sleepqueue_chain *sc; 331 struct sleepqueue *sq; 332 struct thread *td; 333 struct proc *p; 334 int do_upcall; 335 int sig; 336 337 do_upcall = 0; 338 td = curthread; 339 p = td->td_proc; 340 sc = SC_LOOKUP(wchan); 341 mtx_assert(&sc->sc_lock, MA_OWNED); 342 MPASS(td->td_sleepqueue == NULL); 343 MPASS(wchan != NULL); 344 CTR3(KTR_PROC, "sleepq catching signals: thread %p (pid %ld, %s)", 345 (void *)td, (long)p->p_pid, p->p_comm); 346 347 /* Mark thread as being in an interruptible sleep. */ 348 mtx_lock_spin(&sched_lock); 349 MPASS(TD_ON_SLEEPQ(td)); 350 td->td_flags |= TDF_SINTR; 351 mtx_unlock_spin(&sched_lock); 352 sleepq_release(wchan); 353 354 /* See if there are any pending signals for this thread. */ 355 PROC_LOCK(p); 356 mtx_lock(&p->p_sigacts->ps_mtx); 357 sig = cursig(td); 358 mtx_unlock(&p->p_sigacts->ps_mtx); 359 if (sig == 0 && thread_suspend_check(1)) 360 sig = SIGSTOP; 361 else 362 do_upcall = thread_upcall_check(td); 363 PROC_UNLOCK(p); 364 365 /* 366 * If there were pending signals and this thread is still on 367 * the sleep queue, remove it from the sleep queue. 368 */ 369 sq = sleepq_lookup(wchan); 370 mtx_lock_spin(&sched_lock); 371 if (TD_ON_SLEEPQ(td) && (sig != 0 || do_upcall != 0)) { 372 mtx_unlock_spin(&sched_lock); 373 sleepq_remove_thread(sq, td); 374 } else 375 mtx_unlock_spin(&sched_lock); 376 return (sig); 377 } 378 379 /* 380 * Switches to another thread if we are still asleep on a sleep queue and 381 * drop the lock on the sleepqueue chain. Returns with sched_lock held. 382 */ 383 static void 384 sleepq_switch(void *wchan) 385 { 386 struct sleepqueue_chain *sc; 387 struct thread *td; 388 389 td = curthread; 390 sc = SC_LOOKUP(wchan); 391 mtx_assert(&sc->sc_lock, MA_OWNED); 392 393 /* 394 * If we have a sleep queue, then we've already been woken up, so 395 * just return. 396 */ 397 if (td->td_sleepqueue != NULL) { 398 MPASS(!TD_ON_SLEEPQ(td)); 399 mtx_unlock_spin(&sc->sc_lock); 400 mtx_lock_spin(&sched_lock); 401 return; 402 } 403 404 /* 405 * Otherwise, actually go to sleep. 406 */ 407 mtx_lock_spin(&sched_lock); 408 mtx_unlock_spin(&sc->sc_lock); 409 410 sched_sleep(td); 411 TD_SET_SLEEPING(td); 412 mi_switch(SW_VOL, NULL); 413 KASSERT(TD_IS_RUNNING(td), ("running but not TDS_RUNNING")); 414 CTR3(KTR_PROC, "sleepq resume: thread %p (pid %ld, %s)", 415 (void *)td, (long)td->td_proc->p_pid, (void *)td->td_proc->p_comm); 416 } 417 418 /* 419 * Check to see if we timed out. 420 */ 421 static int 422 sleepq_check_timeout(void) 423 { 424 struct thread *td; 425 426 mtx_assert(&sched_lock, MA_OWNED); 427 td = curthread; 428 429 /* 430 * If TDF_TIMEOUT is set, we timed out. 431 */ 432 if (td->td_flags & TDF_TIMEOUT) { 433 td->td_flags &= ~TDF_TIMEOUT; 434 return (EWOULDBLOCK); 435 } 436 437 /* 438 * If TDF_TIMOFAIL is set, the timeout ran after we had 439 * already been woken up. 440 */ 441 if (td->td_flags & TDF_TIMOFAIL) 442 td->td_flags &= ~TDF_TIMOFAIL; 443 444 /* 445 * If callout_stop() fails, then the timeout is running on 446 * another CPU, so synchronize with it to avoid having it 447 * accidentally wake up a subsequent sleep. 448 */ 449 else if (callout_stop(&td->td_slpcallout) == 0) { 450 td->td_flags |= TDF_TIMEOUT; 451 TD_SET_SLEEPING(td); 452 mi_switch(SW_INVOL, NULL); 453 } 454 return (0); 455 } 456 457 /* 458 * Check to see if we were awoken by a signal. 459 */ 460 static int 461 sleepq_check_signals(void) 462 { 463 struct thread *td; 464 465 mtx_assert(&sched_lock, MA_OWNED); 466 td = curthread; 467 468 /* We are no longer in an interruptible sleep. */ 469 td->td_flags &= ~TDF_SINTR; 470 471 if (td->td_flags & TDF_INTERRUPT) 472 return (td->td_intrval); 473 return (0); 474 } 475 476 /* 477 * If we were in an interruptible sleep and we weren't interrupted and 478 * didn't timeout, check to see if there are any pending signals and 479 * which return value we should use if so. The return value from an 480 * earlier call to sleepq_catch_signals() should be passed in as the 481 * argument. 482 */ 483 int 484 sleepq_calc_signal_retval(int sig) 485 { 486 struct thread *td; 487 struct proc *p; 488 int rval; 489 490 td = curthread; 491 p = td->td_proc; 492 PROC_LOCK(p); 493 mtx_lock(&p->p_sigacts->ps_mtx); 494 /* XXX: Should we always be calling cursig()? */ 495 if (sig == 0) 496 sig = cursig(td); 497 if (sig != 0) { 498 if (SIGISMEMBER(p->p_sigacts->ps_sigintr, sig)) 499 rval = EINTR; 500 else 501 rval = ERESTART; 502 } else 503 rval = 0; 504 mtx_unlock(&p->p_sigacts->ps_mtx); 505 PROC_UNLOCK(p); 506 return (rval); 507 } 508 509 /* 510 * Block the current thread until it is awakened from its sleep queue. 511 */ 512 void 513 sleepq_wait(void *wchan) 514 { 515 516 sleepq_switch(wchan); 517 mtx_unlock_spin(&sched_lock); 518 } 519 520 /* 521 * Block the current thread until it is awakened from its sleep queue 522 * or it is interrupted by a signal. 523 */ 524 int 525 sleepq_wait_sig(void *wchan) 526 { 527 int rval; 528 529 sleepq_switch(wchan); 530 rval = sleepq_check_signals(); 531 mtx_unlock_spin(&sched_lock); 532 return (rval); 533 } 534 535 /* 536 * Block the current thread until it is awakened from its sleep queue 537 * or it times out while waiting. 538 */ 539 int 540 sleepq_timedwait(void *wchan) 541 { 542 int rval; 543 544 sleepq_switch(wchan); 545 rval = sleepq_check_timeout(); 546 mtx_unlock_spin(&sched_lock); 547 return (rval); 548 } 549 550 /* 551 * Block the current thread until it is awakened from its sleep queue, 552 * it is interrupted by a signal, or it times out waiting to be awakened. 553 */ 554 int 555 sleepq_timedwait_sig(void *wchan, int signal_caught) 556 { 557 int rvalt, rvals; 558 559 sleepq_switch(wchan); 560 rvalt = sleepq_check_timeout(); 561 rvals = sleepq_check_signals(); 562 mtx_unlock_spin(&sched_lock); 563 if (signal_caught || rvalt == 0) 564 return (rvals); 565 else 566 return (rvalt); 567 } 568 569 /* 570 * Removes a thread from a sleep queue. 571 */ 572 static void 573 sleepq_remove_thread(struct sleepqueue *sq, struct thread *td) 574 { 575 struct sleepqueue_chain *sc; 576 577 MPASS(td != NULL); 578 MPASS(sq->sq_wchan != NULL); 579 MPASS(td->td_wchan == sq->sq_wchan); 580 sc = SC_LOOKUP(sq->sq_wchan); 581 mtx_assert(&sc->sc_lock, MA_OWNED); 582 583 /* Remove the thread from the queue. */ 584 TAILQ_REMOVE(&sq->sq_blocked, td, td_slpq); 585 586 /* 587 * Get a sleep queue for this thread. If this is the last waiter, 588 * use the queue itself and take it out of the chain, otherwise, 589 * remove a queue from the free list. 590 */ 591 if (LIST_EMPTY(&sq->sq_free)) { 592 td->td_sleepqueue = sq; 593 #ifdef INVARIANTS 594 sq->sq_wchan = NULL; 595 #endif 596 #ifdef SLEEPQUEUE_PROFILING 597 sc->sc_depth--; 598 #endif 599 } else 600 td->td_sleepqueue = LIST_FIRST(&sq->sq_free); 601 LIST_REMOVE(td->td_sleepqueue, sq_hash); 602 603 mtx_lock_spin(&sched_lock); 604 td->td_wmesg = NULL; 605 td->td_wchan = NULL; 606 mtx_unlock_spin(&sched_lock); 607 } 608 609 /* 610 * Resumes a thread that was asleep on a queue. 611 */ 612 static void 613 sleepq_resume_thread(struct thread *td, int pri) 614 { 615 616 /* 617 * Note that thread td might not be sleeping if it is running 618 * sleepq_catch_signals() on another CPU or is blocked on 619 * its proc lock to check signals. It doesn't hurt to clear 620 * the sleeping flag if it isn't set though, so we just always 621 * do it. However, we can't assert that it is set. 622 */ 623 mtx_lock_spin(&sched_lock); 624 CTR3(KTR_PROC, "sleepq_wakeup: thread %p (pid %ld, %s)", 625 (void *)td, (long)td->td_proc->p_pid, td->td_proc->p_comm); 626 TD_CLR_SLEEPING(td); 627 628 /* Adjust priority if requested. */ 629 MPASS(pri == -1 || (pri >= PRI_MIN && pri <= PRI_MAX)); 630 if (pri != -1 && td->td_priority > pri) 631 td->td_priority = pri; 632 setrunnable(td); 633 mtx_unlock_spin(&sched_lock); 634 } 635 636 /* 637 * Find the highest priority thread sleeping on a wait channel and resume it. 638 */ 639 void 640 sleepq_signal(void *wchan, int flags, int pri) 641 { 642 struct sleepqueue *sq; 643 struct thread *td; 644 645 CTR2(KTR_PROC, "sleepq_signal(%p, %d)", wchan, flags); 646 KASSERT(wchan != NULL, ("%s: invalid NULL wait channel", __func__)); 647 sq = sleepq_lookup(wchan); 648 if (sq == NULL) { 649 sleepq_release(wchan); 650 return; 651 } 652 KASSERT(sq->sq_flags == flags, 653 ("%s: mismatch between sleep/wakeup and cv_*", __func__)); 654 /* XXX: Do for all sleep queues eventually. */ 655 if (flags & SLEEPQ_CONDVAR) 656 mtx_assert(sq->sq_lock, MA_OWNED); 657 658 /* Remove first thread from queue and awaken it. */ 659 td = TAILQ_FIRST(&sq->sq_blocked); 660 sleepq_remove_thread(sq, td); 661 sleepq_release(wchan); 662 sleepq_resume_thread(td, pri); 663 } 664 665 /* 666 * Resume all threads sleeping on a specified wait channel. 667 */ 668 void 669 sleepq_broadcast(void *wchan, int flags, int pri) 670 { 671 TAILQ_HEAD(, thread) list; 672 struct sleepqueue *sq; 673 struct thread *td; 674 675 CTR2(KTR_PROC, "sleepq_broadcast(%p, %d)", wchan, flags); 676 KASSERT(wchan != NULL, ("%s: invalid NULL wait channel", __func__)); 677 sq = sleepq_lookup(wchan); 678 if (sq == NULL) { 679 sleepq_release(wchan); 680 return; 681 } 682 KASSERT(sq->sq_flags == flags, 683 ("%s: mismatch between sleep/wakeup and cv_*", __func__)); 684 /* XXX: Do for all sleep queues eventually. */ 685 if (flags & SLEEPQ_CONDVAR) 686 mtx_assert(sq->sq_lock, MA_OWNED); 687 688 /* Move blocked threads from the sleep queue to a temporary list. */ 689 TAILQ_INIT(&list); 690 while (!TAILQ_EMPTY(&sq->sq_blocked)) { 691 td = TAILQ_FIRST(&sq->sq_blocked); 692 sleepq_remove_thread(sq, td); 693 TAILQ_INSERT_TAIL(&list, td, td_slpq); 694 } 695 sleepq_release(wchan); 696 697 /* Resume all the threads on the temporary list. */ 698 while (!TAILQ_EMPTY(&list)) { 699 td = TAILQ_FIRST(&list); 700 TAILQ_REMOVE(&list, td, td_slpq); 701 sleepq_resume_thread(td, pri); 702 } 703 } 704 705 /* 706 * Time sleeping threads out. When the timeout expires, the thread is 707 * removed from the sleep queue and made runnable if it is still asleep. 708 */ 709 static void 710 sleepq_timeout(void *arg) 711 { 712 struct sleepqueue *sq; 713 struct thread *td; 714 void *wchan; 715 716 td = arg; 717 CTR3(KTR_PROC, "sleepq_timeout: thread %p (pid %ld, %s)", 718 (void *)td, (long)td->td_proc->p_pid, (void *)td->td_proc->p_comm); 719 720 /* 721 * First, see if the thread is asleep and get the wait channel if 722 * it is. 723 */ 724 mtx_lock_spin(&sched_lock); 725 if (TD_ON_SLEEPQ(td)) { 726 wchan = td->td_wchan; 727 mtx_unlock_spin(&sched_lock); 728 sq = sleepq_lookup(wchan); 729 mtx_lock_spin(&sched_lock); 730 } else { 731 wchan = NULL; 732 sq = NULL; 733 } 734 735 /* 736 * At this point, if the thread is still on the sleep queue, 737 * we have that sleep queue locked as it cannot migrate sleep 738 * queues while we dropped sched_lock. If it had resumed and 739 * was on another CPU while the lock was dropped, it would have 740 * seen that TDF_TIMEOUT and TDF_TIMOFAIL are clear and the 741 * call to callout_stop() to stop this routine would have failed 742 * meaning that it would have already set TDF_TIMEOUT to 743 * synchronize with this function. 744 */ 745 if (TD_ON_SLEEPQ(td)) { 746 MPASS(td->td_wchan == wchan); 747 MPASS(sq != NULL); 748 td->td_flags |= TDF_TIMEOUT; 749 mtx_unlock_spin(&sched_lock); 750 sleepq_remove_thread(sq, td); 751 sleepq_release(wchan); 752 sleepq_resume_thread(td, -1); 753 return; 754 } else if (wchan != NULL) 755 sleepq_release(wchan); 756 757 /* 758 * Now check for the edge cases. First, if TDF_TIMEOUT is set, 759 * then the other thread has already yielded to us, so clear 760 * the flag and resume it. If TDF_TIMEOUT is not set, then the 761 * we know that the other thread is not on a sleep queue, but it 762 * hasn't resumed execution yet. In that case, set TDF_TIMOFAIL 763 * to let it know that the timeout has already run and doesn't 764 * need to be canceled. 765 */ 766 if (td->td_flags & TDF_TIMEOUT) { 767 MPASS(TD_IS_SLEEPING(td)); 768 td->td_flags &= ~TDF_TIMEOUT; 769 TD_CLR_SLEEPING(td); 770 setrunnable(td); 771 } else 772 td->td_flags |= TDF_TIMOFAIL; 773 mtx_unlock_spin(&sched_lock); 774 } 775 776 /* 777 * Resumes a specific thread from the sleep queue associated with a specific 778 * wait channel if it is on that queue. 779 */ 780 void 781 sleepq_remove(struct thread *td, void *wchan) 782 { 783 struct sleepqueue *sq; 784 785 /* 786 * Look up the sleep queue for this wait channel, then re-check 787 * that the thread is asleep on that channel, if it is not, then 788 * bail. 789 */ 790 MPASS(wchan != NULL); 791 sq = sleepq_lookup(wchan); 792 mtx_lock_spin(&sched_lock); 793 if (!TD_ON_SLEEPQ(td) || td->td_wchan != wchan) { 794 mtx_unlock_spin(&sched_lock); 795 sleepq_release(wchan); 796 return; 797 } 798 mtx_unlock_spin(&sched_lock); 799 MPASS(sq != NULL); 800 801 /* Thread is asleep on sleep queue sq, so wake it up. */ 802 sleepq_remove_thread(sq, td); 803 sleepq_release(wchan); 804 sleepq_resume_thread(td, -1); 805 } 806 807 /* 808 * Abort a thread as if an interrupt had occurred. Only abort 809 * interruptible waits (unfortunately it isn't safe to abort others). 810 * 811 * XXX: What in the world does the comment below mean? 812 * Also, whatever the signal code does... 813 */ 814 void 815 sleepq_abort(struct thread *td) 816 { 817 void *wchan; 818 819 mtx_assert(&sched_lock, MA_OWNED); 820 MPASS(TD_ON_SLEEPQ(td)); 821 MPASS(td->td_flags & TDF_SINTR); 822 823 /* 824 * If the TDF_TIMEOUT flag is set, just leave. A 825 * timeout is scheduled anyhow. 826 */ 827 if (td->td_flags & TDF_TIMEOUT) 828 return; 829 830 CTR3(KTR_PROC, "sleepq_abort: thread %p (pid %ld, %s)", 831 (void *)td, (long)td->td_proc->p_pid, (void *)td->td_proc->p_comm); 832 wchan = td->td_wchan; 833 mtx_unlock_spin(&sched_lock); 834 sleepq_remove(td, wchan); 835 mtx_lock_spin(&sched_lock); 836 } 837