1 /*- 2 * Copyright (c) 2004 John Baldwin <jhb@FreeBSD.org> 3 * All rights reserved. 4 * 5 * Redistribution and use in source and binary forms, with or without 6 * modification, are permitted provided that the following conditions 7 * are met: 8 * 1. Redistributions of source code must retain the above copyright 9 * notice, this list of conditions and the following disclaimer. 10 * 2. Redistributions in binary form must reproduce the above copyright 11 * notice, this list of conditions and the following disclaimer in the 12 * documentation and/or other materials provided with the distribution. 13 * 3. Neither the name of the author nor the names of any co-contributors 14 * may be used to endorse or promote products derived from this software 15 * without specific prior written permission. 16 * 17 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND 18 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 19 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 20 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE 21 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 22 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 23 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 24 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 25 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 26 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 27 * SUCH DAMAGE. 28 */ 29 30 /* 31 * Implementation of sleep queues used to hold queue of threads blocked on 32 * a wait channel. Sleep queues different from turnstiles in that wait 33 * channels are not owned by anyone, so there is no priority propagation. 34 * Sleep queues can also provide a timeout and can also be interrupted by 35 * signals. That said, there are several similarities between the turnstile 36 * and sleep queue implementations. (Note: turnstiles were implemented 37 * first.) For example, both use a hash table of the same size where each 38 * bucket is referred to as a "chain" that contains both a spin lock and 39 * a linked list of queues. An individual queue is located by using a hash 40 * to pick a chain, locking the chain, and then walking the chain searching 41 * for the queue. This means that a wait channel object does not need to 42 * embed it's queue head just as locks do not embed their turnstile queue 43 * head. Threads also carry around a sleep queue that they lend to the 44 * wait channel when blocking. Just as in turnstiles, the queue includes 45 * a free list of the sleep queues of other threads blocked on the same 46 * wait channel in the case of multiple waiters. 47 * 48 * Some additional functionality provided by sleep queues include the 49 * ability to set a timeout. The timeout is managed using a per-thread 50 * callout that resumes a thread if it is asleep. A thread may also 51 * catch signals while it is asleep (aka an interruptible sleep). The 52 * signal code uses sleepq_abort() to interrupt a sleeping thread. Finally, 53 * sleep queues also provide some extra assertions. One is not allowed to 54 * mix the sleep/wakeup and cv APIs for a given wait channel. Also, one 55 * must consistently use the same lock to synchronize with a wait channel, 56 * though this check is currently only a warning for sleep/wakeup due to 57 * pre-existing abuse of that API. The same lock must also be held when 58 * awakening threads, though that is currently only enforced for condition 59 * variables. 60 */ 61 62 #include <sys/cdefs.h> 63 __FBSDID("$FreeBSD$"); 64 65 #include "opt_sleepqueue_profiling.h" 66 #include "opt_ddb.h" 67 #include "opt_kdtrace.h" 68 #include "opt_sched.h" 69 70 #include <sys/param.h> 71 #include <sys/systm.h> 72 #include <sys/lock.h> 73 #include <sys/kernel.h> 74 #include <sys/ktr.h> 75 #include <sys/mutex.h> 76 #include <sys/proc.h> 77 #include <sys/sbuf.h> 78 #include <sys/sched.h> 79 #include <sys/sdt.h> 80 #include <sys/signalvar.h> 81 #include <sys/sleepqueue.h> 82 #include <sys/sysctl.h> 83 84 #include <vm/uma.h> 85 86 #ifdef DDB 87 #include <ddb/ddb.h> 88 #endif 89 90 /* 91 * Constants for the hash table of sleep queue chains. These constants are 92 * the same ones that 4BSD (and possibly earlier versions of BSD) used. 93 * Basically, we ignore the lower 8 bits of the address since most wait 94 * channel pointers are aligned and only look at the next 7 bits for the 95 * hash. SC_TABLESIZE must be a power of two for SC_MASK to work properly. 96 */ 97 #define SC_TABLESIZE 128 /* Must be power of 2. */ 98 #define SC_MASK (SC_TABLESIZE - 1) 99 #define SC_SHIFT 8 100 #define SC_HASH(wc) (((uintptr_t)(wc) >> SC_SHIFT) & SC_MASK) 101 #define SC_LOOKUP(wc) &sleepq_chains[SC_HASH(wc)] 102 #define NR_SLEEPQS 2 103 /* 104 * There two different lists of sleep queues. Both lists are connected 105 * via the sq_hash entries. The first list is the sleep queue chain list 106 * that a sleep queue is on when it is attached to a wait channel. The 107 * second list is the free list hung off of a sleep queue that is attached 108 * to a wait channel. 109 * 110 * Each sleep queue also contains the wait channel it is attached to, the 111 * list of threads blocked on that wait channel, flags specific to the 112 * wait channel, and the lock used to synchronize with a wait channel. 113 * The flags are used to catch mismatches between the various consumers 114 * of the sleep queue API (e.g. sleep/wakeup and condition variables). 115 * The lock pointer is only used when invariants are enabled for various 116 * debugging checks. 117 * 118 * Locking key: 119 * c - sleep queue chain lock 120 */ 121 struct sleepqueue { 122 TAILQ_HEAD(, thread) sq_blocked[NR_SLEEPQS]; /* (c) Blocked threads. */ 123 u_int sq_blockedcnt[NR_SLEEPQS]; /* (c) N. of blocked threads. */ 124 LIST_ENTRY(sleepqueue) sq_hash; /* (c) Chain and free list. */ 125 LIST_HEAD(, sleepqueue) sq_free; /* (c) Free queues. */ 126 void *sq_wchan; /* (c) Wait channel. */ 127 int sq_type; /* (c) Queue type. */ 128 #ifdef INVARIANTS 129 struct lock_object *sq_lock; /* (c) Associated lock. */ 130 #endif 131 }; 132 133 struct sleepqueue_chain { 134 LIST_HEAD(, sleepqueue) sc_queues; /* List of sleep queues. */ 135 struct mtx sc_lock; /* Spin lock for this chain. */ 136 #ifdef SLEEPQUEUE_PROFILING 137 u_int sc_depth; /* Length of sc_queues. */ 138 u_int sc_max_depth; /* Max length of sc_queues. */ 139 #endif 140 }; 141 142 #ifdef SLEEPQUEUE_PROFILING 143 u_int sleepq_max_depth; 144 static SYSCTL_NODE(_debug, OID_AUTO, sleepq, CTLFLAG_RD, 0, "sleepq profiling"); 145 static SYSCTL_NODE(_debug_sleepq, OID_AUTO, chains, CTLFLAG_RD, 0, 146 "sleepq chain stats"); 147 SYSCTL_UINT(_debug_sleepq, OID_AUTO, max_depth, CTLFLAG_RD, &sleepq_max_depth, 148 0, "maxmimum depth achieved of a single chain"); 149 150 static void sleepq_profile(const char *wmesg); 151 static int prof_enabled; 152 #endif 153 static struct sleepqueue_chain sleepq_chains[SC_TABLESIZE]; 154 static uma_zone_t sleepq_zone; 155 156 /* 157 * Prototypes for non-exported routines. 158 */ 159 static int sleepq_catch_signals(void *wchan, int pri); 160 static int sleepq_check_signals(void); 161 static int sleepq_check_timeout(void); 162 #ifdef INVARIANTS 163 static void sleepq_dtor(void *mem, int size, void *arg); 164 #endif 165 static int sleepq_init(void *mem, int size, int flags); 166 static int sleepq_resume_thread(struct sleepqueue *sq, struct thread *td, 167 int pri); 168 static void sleepq_switch(void *wchan, int pri); 169 static void sleepq_timeout(void *arg); 170 171 SDT_PROBE_DECLARE(sched, , , sleep); 172 SDT_PROBE_DECLARE(sched, , , wakeup); 173 174 /* 175 * Early initialization of sleep queues that is called from the sleepinit() 176 * SYSINIT. 177 */ 178 void 179 init_sleepqueues(void) 180 { 181 #ifdef SLEEPQUEUE_PROFILING 182 struct sysctl_oid *chain_oid; 183 char chain_name[10]; 184 #endif 185 int i; 186 187 for (i = 0; i < SC_TABLESIZE; i++) { 188 LIST_INIT(&sleepq_chains[i].sc_queues); 189 mtx_init(&sleepq_chains[i].sc_lock, "sleepq chain", NULL, 190 MTX_SPIN | MTX_RECURSE); 191 #ifdef SLEEPQUEUE_PROFILING 192 snprintf(chain_name, sizeof(chain_name), "%d", i); 193 chain_oid = SYSCTL_ADD_NODE(NULL, 194 SYSCTL_STATIC_CHILDREN(_debug_sleepq_chains), OID_AUTO, 195 chain_name, CTLFLAG_RD, NULL, "sleepq chain stats"); 196 SYSCTL_ADD_UINT(NULL, SYSCTL_CHILDREN(chain_oid), OID_AUTO, 197 "depth", CTLFLAG_RD, &sleepq_chains[i].sc_depth, 0, NULL); 198 SYSCTL_ADD_UINT(NULL, SYSCTL_CHILDREN(chain_oid), OID_AUTO, 199 "max_depth", CTLFLAG_RD, &sleepq_chains[i].sc_max_depth, 0, 200 NULL); 201 #endif 202 } 203 sleepq_zone = uma_zcreate("SLEEPQUEUE", sizeof(struct sleepqueue), 204 #ifdef INVARIANTS 205 NULL, sleepq_dtor, sleepq_init, NULL, UMA_ALIGN_CACHE, 0); 206 #else 207 NULL, NULL, sleepq_init, NULL, UMA_ALIGN_CACHE, 0); 208 #endif 209 210 thread0.td_sleepqueue = sleepq_alloc(); 211 } 212 213 /* 214 * Get a sleep queue for a new thread. 215 */ 216 struct sleepqueue * 217 sleepq_alloc(void) 218 { 219 220 return (uma_zalloc(sleepq_zone, M_WAITOK)); 221 } 222 223 /* 224 * Free a sleep queue when a thread is destroyed. 225 */ 226 void 227 sleepq_free(struct sleepqueue *sq) 228 { 229 230 uma_zfree(sleepq_zone, sq); 231 } 232 233 /* 234 * Lock the sleep queue chain associated with the specified wait channel. 235 */ 236 void 237 sleepq_lock(void *wchan) 238 { 239 struct sleepqueue_chain *sc; 240 241 sc = SC_LOOKUP(wchan); 242 mtx_lock_spin(&sc->sc_lock); 243 } 244 245 /* 246 * Look up the sleep queue associated with a given wait channel in the hash 247 * table locking the associated sleep queue chain. If no queue is found in 248 * the table, NULL is returned. 249 */ 250 struct sleepqueue * 251 sleepq_lookup(void *wchan) 252 { 253 struct sleepqueue_chain *sc; 254 struct sleepqueue *sq; 255 256 KASSERT(wchan != NULL, ("%s: invalid NULL wait channel", __func__)); 257 sc = SC_LOOKUP(wchan); 258 mtx_assert(&sc->sc_lock, MA_OWNED); 259 LIST_FOREACH(sq, &sc->sc_queues, sq_hash) 260 if (sq->sq_wchan == wchan) 261 return (sq); 262 return (NULL); 263 } 264 265 /* 266 * Unlock the sleep queue chain associated with a given wait channel. 267 */ 268 void 269 sleepq_release(void *wchan) 270 { 271 struct sleepqueue_chain *sc; 272 273 sc = SC_LOOKUP(wchan); 274 mtx_unlock_spin(&sc->sc_lock); 275 } 276 277 /* 278 * Places the current thread on the sleep queue for the specified wait 279 * channel. If INVARIANTS is enabled, then it associates the passed in 280 * lock with the sleepq to make sure it is held when that sleep queue is 281 * woken up. 282 */ 283 void 284 sleepq_add(void *wchan, struct lock_object *lock, const char *wmesg, int flags, 285 int queue) 286 { 287 struct sleepqueue_chain *sc; 288 struct sleepqueue *sq; 289 struct thread *td; 290 291 td = curthread; 292 sc = SC_LOOKUP(wchan); 293 mtx_assert(&sc->sc_lock, MA_OWNED); 294 MPASS(td->td_sleepqueue != NULL); 295 MPASS(wchan != NULL); 296 MPASS((queue >= 0) && (queue < NR_SLEEPQS)); 297 298 /* If this thread is not allowed to sleep, die a horrible death. */ 299 KASSERT(!(td->td_pflags & TDP_NOSLEEPING), 300 ("%s: td %p to sleep on wchan %p with TDP_NOSLEEPING on", 301 __func__, td, wchan)); 302 303 /* Look up the sleep queue associated with the wait channel 'wchan'. */ 304 sq = sleepq_lookup(wchan); 305 306 /* 307 * If the wait channel does not already have a sleep queue, use 308 * this thread's sleep queue. Otherwise, insert the current thread 309 * into the sleep queue already in use by this wait channel. 310 */ 311 if (sq == NULL) { 312 #ifdef INVARIANTS 313 int i; 314 315 sq = td->td_sleepqueue; 316 for (i = 0; i < NR_SLEEPQS; i++) { 317 KASSERT(TAILQ_EMPTY(&sq->sq_blocked[i]), 318 ("thread's sleep queue %d is not empty", i)); 319 KASSERT(sq->sq_blockedcnt[i] == 0, 320 ("thread's sleep queue %d count mismatches", i)); 321 } 322 KASSERT(LIST_EMPTY(&sq->sq_free), 323 ("thread's sleep queue has a non-empty free list")); 324 KASSERT(sq->sq_wchan == NULL, ("stale sq_wchan pointer")); 325 sq->sq_lock = lock; 326 #endif 327 #ifdef SLEEPQUEUE_PROFILING 328 sc->sc_depth++; 329 if (sc->sc_depth > sc->sc_max_depth) { 330 sc->sc_max_depth = sc->sc_depth; 331 if (sc->sc_max_depth > sleepq_max_depth) 332 sleepq_max_depth = sc->sc_max_depth; 333 } 334 #endif 335 sq = td->td_sleepqueue; 336 LIST_INSERT_HEAD(&sc->sc_queues, sq, sq_hash); 337 sq->sq_wchan = wchan; 338 sq->sq_type = flags & SLEEPQ_TYPE; 339 } else { 340 MPASS(wchan == sq->sq_wchan); 341 MPASS(lock == sq->sq_lock); 342 MPASS((flags & SLEEPQ_TYPE) == sq->sq_type); 343 LIST_INSERT_HEAD(&sq->sq_free, td->td_sleepqueue, sq_hash); 344 } 345 thread_lock(td); 346 TAILQ_INSERT_TAIL(&sq->sq_blocked[queue], td, td_slpq); 347 sq->sq_blockedcnt[queue]++; 348 td->td_sleepqueue = NULL; 349 td->td_sqqueue = queue; 350 td->td_wchan = wchan; 351 td->td_wmesg = wmesg; 352 if (flags & SLEEPQ_INTERRUPTIBLE) { 353 td->td_flags |= TDF_SINTR; 354 td->td_flags &= ~TDF_SLEEPABORT; 355 } 356 thread_unlock(td); 357 } 358 359 /* 360 * Sets a timeout that will remove the current thread from the specified 361 * sleep queue after timo ticks if the thread has not already been awakened. 362 */ 363 void 364 sleepq_set_timeout(void *wchan, int timo) 365 { 366 struct sleepqueue_chain *sc; 367 struct thread *td; 368 369 td = curthread; 370 sc = SC_LOOKUP(wchan); 371 mtx_assert(&sc->sc_lock, MA_OWNED); 372 MPASS(TD_ON_SLEEPQ(td)); 373 MPASS(td->td_sleepqueue == NULL); 374 MPASS(wchan != NULL); 375 callout_reset_curcpu(&td->td_slpcallout, timo, sleepq_timeout, td); 376 } 377 378 /* 379 * Return the number of actual sleepers for the specified queue. 380 */ 381 u_int 382 sleepq_sleepcnt(void *wchan, int queue) 383 { 384 struct sleepqueue *sq; 385 386 KASSERT(wchan != NULL, ("%s: invalid NULL wait channel", __func__)); 387 MPASS((queue >= 0) && (queue < NR_SLEEPQS)); 388 sq = sleepq_lookup(wchan); 389 if (sq == NULL) 390 return (0); 391 return (sq->sq_blockedcnt[queue]); 392 } 393 394 /* 395 * Marks the pending sleep of the current thread as interruptible and 396 * makes an initial check for pending signals before putting a thread 397 * to sleep. Enters and exits with the thread lock held. Thread lock 398 * may have transitioned from the sleepq lock to a run lock. 399 */ 400 static int 401 sleepq_catch_signals(void *wchan, int pri) 402 { 403 struct sleepqueue_chain *sc; 404 struct sleepqueue *sq; 405 struct thread *td; 406 struct proc *p; 407 struct sigacts *ps; 408 int sig, ret, stop_allowed; 409 410 td = curthread; 411 p = curproc; 412 sc = SC_LOOKUP(wchan); 413 mtx_assert(&sc->sc_lock, MA_OWNED); 414 MPASS(wchan != NULL); 415 if ((td->td_pflags & TDP_WAKEUP) != 0) { 416 td->td_pflags &= ~TDP_WAKEUP; 417 ret = EINTR; 418 thread_lock(td); 419 goto out; 420 } 421 422 /* 423 * See if there are any pending signals for this thread. If not 424 * we can switch immediately. Otherwise do the signal processing 425 * directly. 426 */ 427 thread_lock(td); 428 if ((td->td_flags & (TDF_NEEDSIGCHK | TDF_NEEDSUSPCHK)) == 0) { 429 sleepq_switch(wchan, pri); 430 return (0); 431 } 432 stop_allowed = (td->td_flags & TDF_SBDRY) ? SIG_STOP_NOT_ALLOWED : 433 SIG_STOP_ALLOWED; 434 thread_unlock(td); 435 mtx_unlock_spin(&sc->sc_lock); 436 CTR3(KTR_PROC, "sleepq catching signals: thread %p (pid %ld, %s)", 437 (void *)td, (long)p->p_pid, td->td_name); 438 PROC_LOCK(p); 439 ps = p->p_sigacts; 440 mtx_lock(&ps->ps_mtx); 441 sig = cursig(td, stop_allowed); 442 if (sig == 0) { 443 mtx_unlock(&ps->ps_mtx); 444 ret = thread_suspend_check(1); 445 MPASS(ret == 0 || ret == EINTR || ret == ERESTART); 446 } else { 447 if (SIGISMEMBER(ps->ps_sigintr, sig)) 448 ret = EINTR; 449 else 450 ret = ERESTART; 451 mtx_unlock(&ps->ps_mtx); 452 } 453 /* 454 * Lock the per-process spinlock prior to dropping the PROC_LOCK 455 * to avoid a signal delivery race. PROC_LOCK, PROC_SLOCK, and 456 * thread_lock() are currently held in tdsendsignal(). 457 */ 458 PROC_SLOCK(p); 459 mtx_lock_spin(&sc->sc_lock); 460 PROC_UNLOCK(p); 461 thread_lock(td); 462 PROC_SUNLOCK(p); 463 if (ret == 0) { 464 sleepq_switch(wchan, pri); 465 return (0); 466 } 467 out: 468 /* 469 * There were pending signals and this thread is still 470 * on the sleep queue, remove it from the sleep queue. 471 */ 472 if (TD_ON_SLEEPQ(td)) { 473 sq = sleepq_lookup(wchan); 474 if (sleepq_resume_thread(sq, td, 0)) { 475 #ifdef INVARIANTS 476 /* 477 * This thread hasn't gone to sleep yet, so it 478 * should not be swapped out. 479 */ 480 panic("not waking up swapper"); 481 #endif 482 } 483 } 484 mtx_unlock_spin(&sc->sc_lock); 485 MPASS(td->td_lock != &sc->sc_lock); 486 return (ret); 487 } 488 489 /* 490 * Switches to another thread if we are still asleep on a sleep queue. 491 * Returns with thread lock. 492 */ 493 static void 494 sleepq_switch(void *wchan, int pri) 495 { 496 struct sleepqueue_chain *sc; 497 struct sleepqueue *sq; 498 struct thread *td; 499 500 td = curthread; 501 sc = SC_LOOKUP(wchan); 502 mtx_assert(&sc->sc_lock, MA_OWNED); 503 THREAD_LOCK_ASSERT(td, MA_OWNED); 504 505 /* 506 * If we have a sleep queue, then we've already been woken up, so 507 * just return. 508 */ 509 if (td->td_sleepqueue != NULL) { 510 mtx_unlock_spin(&sc->sc_lock); 511 return; 512 } 513 514 /* 515 * If TDF_TIMEOUT is set, then our sleep has been timed out 516 * already but we are still on the sleep queue, so dequeue the 517 * thread and return. 518 */ 519 if (td->td_flags & TDF_TIMEOUT) { 520 MPASS(TD_ON_SLEEPQ(td)); 521 sq = sleepq_lookup(wchan); 522 if (sleepq_resume_thread(sq, td, 0)) { 523 #ifdef INVARIANTS 524 /* 525 * This thread hasn't gone to sleep yet, so it 526 * should not be swapped out. 527 */ 528 panic("not waking up swapper"); 529 #endif 530 } 531 mtx_unlock_spin(&sc->sc_lock); 532 return; 533 } 534 #ifdef SLEEPQUEUE_PROFILING 535 if (prof_enabled) 536 sleepq_profile(td->td_wmesg); 537 #endif 538 MPASS(td->td_sleepqueue == NULL); 539 sched_sleep(td, pri); 540 thread_lock_set(td, &sc->sc_lock); 541 SDT_PROBE0(sched, , , sleep); 542 TD_SET_SLEEPING(td); 543 mi_switch(SW_VOL | SWT_SLEEPQ, NULL); 544 KASSERT(TD_IS_RUNNING(td), ("running but not TDS_RUNNING")); 545 CTR3(KTR_PROC, "sleepq resume: thread %p (pid %ld, %s)", 546 (void *)td, (long)td->td_proc->p_pid, (void *)td->td_name); 547 } 548 549 /* 550 * Check to see if we timed out. 551 */ 552 static int 553 sleepq_check_timeout(void) 554 { 555 struct thread *td; 556 557 td = curthread; 558 THREAD_LOCK_ASSERT(td, MA_OWNED); 559 560 /* 561 * If TDF_TIMEOUT is set, we timed out. 562 */ 563 if (td->td_flags & TDF_TIMEOUT) { 564 td->td_flags &= ~TDF_TIMEOUT; 565 return (EWOULDBLOCK); 566 } 567 568 /* 569 * If TDF_TIMOFAIL is set, the timeout ran after we had 570 * already been woken up. 571 */ 572 if (td->td_flags & TDF_TIMOFAIL) 573 td->td_flags &= ~TDF_TIMOFAIL; 574 575 /* 576 * If callout_stop() fails, then the timeout is running on 577 * another CPU, so synchronize with it to avoid having it 578 * accidentally wake up a subsequent sleep. 579 */ 580 else if (callout_stop(&td->td_slpcallout) == 0) { 581 td->td_flags |= TDF_TIMEOUT; 582 TD_SET_SLEEPING(td); 583 mi_switch(SW_INVOL | SWT_SLEEPQTIMO, NULL); 584 } 585 return (0); 586 } 587 588 /* 589 * Check to see if we were awoken by a signal. 590 */ 591 static int 592 sleepq_check_signals(void) 593 { 594 struct thread *td; 595 596 td = curthread; 597 THREAD_LOCK_ASSERT(td, MA_OWNED); 598 599 /* We are no longer in an interruptible sleep. */ 600 if (td->td_flags & TDF_SINTR) 601 td->td_flags &= ~TDF_SINTR; 602 603 if (td->td_flags & TDF_SLEEPABORT) { 604 td->td_flags &= ~TDF_SLEEPABORT; 605 return (td->td_intrval); 606 } 607 608 return (0); 609 } 610 611 /* 612 * Block the current thread until it is awakened from its sleep queue. 613 */ 614 void 615 sleepq_wait(void *wchan, int pri) 616 { 617 struct thread *td; 618 619 td = curthread; 620 MPASS(!(td->td_flags & TDF_SINTR)); 621 thread_lock(td); 622 sleepq_switch(wchan, pri); 623 thread_unlock(td); 624 } 625 626 /* 627 * Block the current thread until it is awakened from its sleep queue 628 * or it is interrupted by a signal. 629 */ 630 int 631 sleepq_wait_sig(void *wchan, int pri) 632 { 633 int rcatch; 634 int rval; 635 636 rcatch = sleepq_catch_signals(wchan, pri); 637 rval = sleepq_check_signals(); 638 thread_unlock(curthread); 639 if (rcatch) 640 return (rcatch); 641 return (rval); 642 } 643 644 /* 645 * Block the current thread until it is awakened from its sleep queue 646 * or it times out while waiting. 647 */ 648 int 649 sleepq_timedwait(void *wchan, int pri) 650 { 651 struct thread *td; 652 int rval; 653 654 td = curthread; 655 MPASS(!(td->td_flags & TDF_SINTR)); 656 thread_lock(td); 657 sleepq_switch(wchan, pri); 658 rval = sleepq_check_timeout(); 659 thread_unlock(td); 660 661 return (rval); 662 } 663 664 /* 665 * Block the current thread until it is awakened from its sleep queue, 666 * it is interrupted by a signal, or it times out waiting to be awakened. 667 */ 668 int 669 sleepq_timedwait_sig(void *wchan, int pri) 670 { 671 int rcatch, rvalt, rvals; 672 673 rcatch = sleepq_catch_signals(wchan, pri); 674 rvalt = sleepq_check_timeout(); 675 rvals = sleepq_check_signals(); 676 thread_unlock(curthread); 677 if (rcatch) 678 return (rcatch); 679 if (rvals) 680 return (rvals); 681 return (rvalt); 682 } 683 684 /* 685 * Returns the type of sleepqueue given a waitchannel. 686 */ 687 int 688 sleepq_type(void *wchan) 689 { 690 struct sleepqueue *sq; 691 int type; 692 693 MPASS(wchan != NULL); 694 695 sleepq_lock(wchan); 696 sq = sleepq_lookup(wchan); 697 if (sq == NULL) { 698 sleepq_release(wchan); 699 return (-1); 700 } 701 type = sq->sq_type; 702 sleepq_release(wchan); 703 return (type); 704 } 705 706 /* 707 * Removes a thread from a sleep queue and makes it 708 * runnable. 709 */ 710 static int 711 sleepq_resume_thread(struct sleepqueue *sq, struct thread *td, int pri) 712 { 713 struct sleepqueue_chain *sc; 714 715 MPASS(td != NULL); 716 MPASS(sq->sq_wchan != NULL); 717 MPASS(td->td_wchan == sq->sq_wchan); 718 MPASS(td->td_sqqueue < NR_SLEEPQS && td->td_sqqueue >= 0); 719 THREAD_LOCK_ASSERT(td, MA_OWNED); 720 sc = SC_LOOKUP(sq->sq_wchan); 721 mtx_assert(&sc->sc_lock, MA_OWNED); 722 723 SDT_PROBE2(sched, , , wakeup, td, td->td_proc); 724 725 /* Remove the thread from the queue. */ 726 sq->sq_blockedcnt[td->td_sqqueue]--; 727 TAILQ_REMOVE(&sq->sq_blocked[td->td_sqqueue], td, td_slpq); 728 729 /* 730 * Get a sleep queue for this thread. If this is the last waiter, 731 * use the queue itself and take it out of the chain, otherwise, 732 * remove a queue from the free list. 733 */ 734 if (LIST_EMPTY(&sq->sq_free)) { 735 td->td_sleepqueue = sq; 736 #ifdef INVARIANTS 737 sq->sq_wchan = NULL; 738 #endif 739 #ifdef SLEEPQUEUE_PROFILING 740 sc->sc_depth--; 741 #endif 742 } else 743 td->td_sleepqueue = LIST_FIRST(&sq->sq_free); 744 LIST_REMOVE(td->td_sleepqueue, sq_hash); 745 746 td->td_wmesg = NULL; 747 td->td_wchan = NULL; 748 td->td_flags &= ~TDF_SINTR; 749 750 CTR3(KTR_PROC, "sleepq_wakeup: thread %p (pid %ld, %s)", 751 (void *)td, (long)td->td_proc->p_pid, td->td_name); 752 753 /* Adjust priority if requested. */ 754 MPASS(pri == 0 || (pri >= PRI_MIN && pri <= PRI_MAX)); 755 if (pri != 0 && td->td_priority > pri && 756 PRI_BASE(td->td_pri_class) == PRI_TIMESHARE) 757 sched_prio(td, pri); 758 759 /* 760 * Note that thread td might not be sleeping if it is running 761 * sleepq_catch_signals() on another CPU or is blocked on its 762 * proc lock to check signals. There's no need to mark the 763 * thread runnable in that case. 764 */ 765 if (TD_IS_SLEEPING(td)) { 766 TD_CLR_SLEEPING(td); 767 return (setrunnable(td)); 768 } 769 return (0); 770 } 771 772 #ifdef INVARIANTS 773 /* 774 * UMA zone item deallocator. 775 */ 776 static void 777 sleepq_dtor(void *mem, int size, void *arg) 778 { 779 struct sleepqueue *sq; 780 int i; 781 782 sq = mem; 783 for (i = 0; i < NR_SLEEPQS; i++) { 784 MPASS(TAILQ_EMPTY(&sq->sq_blocked[i])); 785 MPASS(sq->sq_blockedcnt[i] == 0); 786 } 787 } 788 #endif 789 790 /* 791 * UMA zone item initializer. 792 */ 793 static int 794 sleepq_init(void *mem, int size, int flags) 795 { 796 struct sleepqueue *sq; 797 int i; 798 799 bzero(mem, size); 800 sq = mem; 801 for (i = 0; i < NR_SLEEPQS; i++) { 802 TAILQ_INIT(&sq->sq_blocked[i]); 803 sq->sq_blockedcnt[i] = 0; 804 } 805 LIST_INIT(&sq->sq_free); 806 return (0); 807 } 808 809 /* 810 * Find the highest priority thread sleeping on a wait channel and resume it. 811 */ 812 int 813 sleepq_signal(void *wchan, int flags, int pri, int queue) 814 { 815 struct sleepqueue *sq; 816 struct thread *td, *besttd; 817 int wakeup_swapper; 818 819 CTR2(KTR_PROC, "sleepq_signal(%p, %d)", wchan, flags); 820 KASSERT(wchan != NULL, ("%s: invalid NULL wait channel", __func__)); 821 MPASS((queue >= 0) && (queue < NR_SLEEPQS)); 822 sq = sleepq_lookup(wchan); 823 if (sq == NULL) 824 return (0); 825 KASSERT(sq->sq_type == (flags & SLEEPQ_TYPE), 826 ("%s: mismatch between sleep/wakeup and cv_*", __func__)); 827 828 /* 829 * Find the highest priority thread on the queue. If there is a 830 * tie, use the thread that first appears in the queue as it has 831 * been sleeping the longest since threads are always added to 832 * the tail of sleep queues. 833 */ 834 besttd = NULL; 835 TAILQ_FOREACH(td, &sq->sq_blocked[queue], td_slpq) { 836 if (besttd == NULL || td->td_priority < besttd->td_priority) 837 besttd = td; 838 } 839 MPASS(besttd != NULL); 840 thread_lock(besttd); 841 wakeup_swapper = sleepq_resume_thread(sq, besttd, pri); 842 thread_unlock(besttd); 843 return (wakeup_swapper); 844 } 845 846 /* 847 * Resume all threads sleeping on a specified wait channel. 848 */ 849 int 850 sleepq_broadcast(void *wchan, int flags, int pri, int queue) 851 { 852 struct sleepqueue *sq; 853 struct thread *td, *tdn; 854 int wakeup_swapper; 855 856 CTR2(KTR_PROC, "sleepq_broadcast(%p, %d)", wchan, flags); 857 KASSERT(wchan != NULL, ("%s: invalid NULL wait channel", __func__)); 858 MPASS((queue >= 0) && (queue < NR_SLEEPQS)); 859 sq = sleepq_lookup(wchan); 860 if (sq == NULL) 861 return (0); 862 KASSERT(sq->sq_type == (flags & SLEEPQ_TYPE), 863 ("%s: mismatch between sleep/wakeup and cv_*", __func__)); 864 865 /* Resume all blocked threads on the sleep queue. */ 866 wakeup_swapper = 0; 867 TAILQ_FOREACH_SAFE(td, &sq->sq_blocked[queue], td_slpq, tdn) { 868 thread_lock(td); 869 if (sleepq_resume_thread(sq, td, pri)) 870 wakeup_swapper = 1; 871 thread_unlock(td); 872 } 873 return (wakeup_swapper); 874 } 875 876 /* 877 * Time sleeping threads out. When the timeout expires, the thread is 878 * removed from the sleep queue and made runnable if it is still asleep. 879 */ 880 static void 881 sleepq_timeout(void *arg) 882 { 883 struct sleepqueue_chain *sc; 884 struct sleepqueue *sq; 885 struct thread *td; 886 void *wchan; 887 int wakeup_swapper; 888 889 td = arg; 890 wakeup_swapper = 0; 891 CTR3(KTR_PROC, "sleepq_timeout: thread %p (pid %ld, %s)", 892 (void *)td, (long)td->td_proc->p_pid, (void *)td->td_name); 893 894 /* 895 * First, see if the thread is asleep and get the wait channel if 896 * it is. 897 */ 898 thread_lock(td); 899 if (TD_IS_SLEEPING(td) && TD_ON_SLEEPQ(td)) { 900 wchan = td->td_wchan; 901 sc = SC_LOOKUP(wchan); 902 THREAD_LOCKPTR_ASSERT(td, &sc->sc_lock); 903 sq = sleepq_lookup(wchan); 904 MPASS(sq != NULL); 905 td->td_flags |= TDF_TIMEOUT; 906 wakeup_swapper = sleepq_resume_thread(sq, td, 0); 907 thread_unlock(td); 908 if (wakeup_swapper) 909 kick_proc0(); 910 return; 911 } 912 913 /* 914 * If the thread is on the SLEEPQ but isn't sleeping yet, it 915 * can either be on another CPU in between sleepq_add() and 916 * one of the sleepq_*wait*() routines or it can be in 917 * sleepq_catch_signals(). 918 */ 919 if (TD_ON_SLEEPQ(td)) { 920 td->td_flags |= TDF_TIMEOUT; 921 thread_unlock(td); 922 return; 923 } 924 925 /* 926 * Now check for the edge cases. First, if TDF_TIMEOUT is set, 927 * then the other thread has already yielded to us, so clear 928 * the flag and resume it. If TDF_TIMEOUT is not set, then the 929 * we know that the other thread is not on a sleep queue, but it 930 * hasn't resumed execution yet. In that case, set TDF_TIMOFAIL 931 * to let it know that the timeout has already run and doesn't 932 * need to be canceled. 933 */ 934 if (td->td_flags & TDF_TIMEOUT) { 935 MPASS(TD_IS_SLEEPING(td)); 936 td->td_flags &= ~TDF_TIMEOUT; 937 TD_CLR_SLEEPING(td); 938 wakeup_swapper = setrunnable(td); 939 } else 940 td->td_flags |= TDF_TIMOFAIL; 941 thread_unlock(td); 942 if (wakeup_swapper) 943 kick_proc0(); 944 } 945 946 /* 947 * Resumes a specific thread from the sleep queue associated with a specific 948 * wait channel if it is on that queue. 949 */ 950 void 951 sleepq_remove(struct thread *td, void *wchan) 952 { 953 struct sleepqueue *sq; 954 int wakeup_swapper; 955 956 /* 957 * Look up the sleep queue for this wait channel, then re-check 958 * that the thread is asleep on that channel, if it is not, then 959 * bail. 960 */ 961 MPASS(wchan != NULL); 962 sleepq_lock(wchan); 963 sq = sleepq_lookup(wchan); 964 /* 965 * We can not lock the thread here as it may be sleeping on a 966 * different sleepq. However, holding the sleepq lock for this 967 * wchan can guarantee that we do not miss a wakeup for this 968 * channel. The asserts below will catch any false positives. 969 */ 970 if (!TD_ON_SLEEPQ(td) || td->td_wchan != wchan) { 971 sleepq_release(wchan); 972 return; 973 } 974 /* Thread is asleep on sleep queue sq, so wake it up. */ 975 thread_lock(td); 976 MPASS(sq != NULL); 977 MPASS(td->td_wchan == wchan); 978 wakeup_swapper = sleepq_resume_thread(sq, td, 0); 979 thread_unlock(td); 980 sleepq_release(wchan); 981 if (wakeup_swapper) 982 kick_proc0(); 983 } 984 985 /* 986 * Abort a thread as if an interrupt had occurred. Only abort 987 * interruptible waits (unfortunately it isn't safe to abort others). 988 */ 989 int 990 sleepq_abort(struct thread *td, int intrval) 991 { 992 struct sleepqueue *sq; 993 void *wchan; 994 995 THREAD_LOCK_ASSERT(td, MA_OWNED); 996 MPASS(TD_ON_SLEEPQ(td)); 997 MPASS(td->td_flags & TDF_SINTR); 998 MPASS(intrval == EINTR || intrval == ERESTART); 999 1000 /* 1001 * If the TDF_TIMEOUT flag is set, just leave. A 1002 * timeout is scheduled anyhow. 1003 */ 1004 if (td->td_flags & TDF_TIMEOUT) 1005 return (0); 1006 1007 CTR3(KTR_PROC, "sleepq_abort: thread %p (pid %ld, %s)", 1008 (void *)td, (long)td->td_proc->p_pid, (void *)td->td_name); 1009 td->td_intrval = intrval; 1010 td->td_flags |= TDF_SLEEPABORT; 1011 /* 1012 * If the thread has not slept yet it will find the signal in 1013 * sleepq_catch_signals() and call sleepq_resume_thread. Otherwise 1014 * we have to do it here. 1015 */ 1016 if (!TD_IS_SLEEPING(td)) 1017 return (0); 1018 wchan = td->td_wchan; 1019 MPASS(wchan != NULL); 1020 sq = sleepq_lookup(wchan); 1021 MPASS(sq != NULL); 1022 1023 /* Thread is asleep on sleep queue sq, so wake it up. */ 1024 return (sleepq_resume_thread(sq, td, 0)); 1025 } 1026 1027 #ifdef SLEEPQUEUE_PROFILING 1028 #define SLEEPQ_PROF_LOCATIONS 1024 1029 #define SLEEPQ_SBUFSIZE 512 1030 struct sleepq_prof { 1031 LIST_ENTRY(sleepq_prof) sp_link; 1032 const char *sp_wmesg; 1033 long sp_count; 1034 }; 1035 1036 LIST_HEAD(sqphead, sleepq_prof); 1037 1038 struct sqphead sleepq_prof_free; 1039 struct sqphead sleepq_hash[SC_TABLESIZE]; 1040 static struct sleepq_prof sleepq_profent[SLEEPQ_PROF_LOCATIONS]; 1041 static struct mtx sleepq_prof_lock; 1042 MTX_SYSINIT(sleepq_prof_lock, &sleepq_prof_lock, "sleepq_prof", MTX_SPIN); 1043 1044 static void 1045 sleepq_profile(const char *wmesg) 1046 { 1047 struct sleepq_prof *sp; 1048 1049 mtx_lock_spin(&sleepq_prof_lock); 1050 if (prof_enabled == 0) 1051 goto unlock; 1052 LIST_FOREACH(sp, &sleepq_hash[SC_HASH(wmesg)], sp_link) 1053 if (sp->sp_wmesg == wmesg) 1054 goto done; 1055 sp = LIST_FIRST(&sleepq_prof_free); 1056 if (sp == NULL) 1057 goto unlock; 1058 sp->sp_wmesg = wmesg; 1059 LIST_REMOVE(sp, sp_link); 1060 LIST_INSERT_HEAD(&sleepq_hash[SC_HASH(wmesg)], sp, sp_link); 1061 done: 1062 sp->sp_count++; 1063 unlock: 1064 mtx_unlock_spin(&sleepq_prof_lock); 1065 return; 1066 } 1067 1068 static void 1069 sleepq_prof_reset(void) 1070 { 1071 struct sleepq_prof *sp; 1072 int enabled; 1073 int i; 1074 1075 mtx_lock_spin(&sleepq_prof_lock); 1076 enabled = prof_enabled; 1077 prof_enabled = 0; 1078 for (i = 0; i < SC_TABLESIZE; i++) 1079 LIST_INIT(&sleepq_hash[i]); 1080 LIST_INIT(&sleepq_prof_free); 1081 for (i = 0; i < SLEEPQ_PROF_LOCATIONS; i++) { 1082 sp = &sleepq_profent[i]; 1083 sp->sp_wmesg = NULL; 1084 sp->sp_count = 0; 1085 LIST_INSERT_HEAD(&sleepq_prof_free, sp, sp_link); 1086 } 1087 prof_enabled = enabled; 1088 mtx_unlock_spin(&sleepq_prof_lock); 1089 } 1090 1091 static int 1092 enable_sleepq_prof(SYSCTL_HANDLER_ARGS) 1093 { 1094 int error, v; 1095 1096 v = prof_enabled; 1097 error = sysctl_handle_int(oidp, &v, v, req); 1098 if (error) 1099 return (error); 1100 if (req->newptr == NULL) 1101 return (error); 1102 if (v == prof_enabled) 1103 return (0); 1104 if (v == 1) 1105 sleepq_prof_reset(); 1106 mtx_lock_spin(&sleepq_prof_lock); 1107 prof_enabled = !!v; 1108 mtx_unlock_spin(&sleepq_prof_lock); 1109 1110 return (0); 1111 } 1112 1113 static int 1114 reset_sleepq_prof_stats(SYSCTL_HANDLER_ARGS) 1115 { 1116 int error, v; 1117 1118 v = 0; 1119 error = sysctl_handle_int(oidp, &v, 0, req); 1120 if (error) 1121 return (error); 1122 if (req->newptr == NULL) 1123 return (error); 1124 if (v == 0) 1125 return (0); 1126 sleepq_prof_reset(); 1127 1128 return (0); 1129 } 1130 1131 static int 1132 dump_sleepq_prof_stats(SYSCTL_HANDLER_ARGS) 1133 { 1134 struct sleepq_prof *sp; 1135 struct sbuf *sb; 1136 int enabled; 1137 int error; 1138 int i; 1139 1140 error = sysctl_wire_old_buffer(req, 0); 1141 if (error != 0) 1142 return (error); 1143 sb = sbuf_new_for_sysctl(NULL, NULL, SLEEPQ_SBUFSIZE, req); 1144 sbuf_printf(sb, "\nwmesg\tcount\n"); 1145 enabled = prof_enabled; 1146 mtx_lock_spin(&sleepq_prof_lock); 1147 prof_enabled = 0; 1148 mtx_unlock_spin(&sleepq_prof_lock); 1149 for (i = 0; i < SC_TABLESIZE; i++) { 1150 LIST_FOREACH(sp, &sleepq_hash[i], sp_link) { 1151 sbuf_printf(sb, "%s\t%ld\n", 1152 sp->sp_wmesg, sp->sp_count); 1153 } 1154 } 1155 mtx_lock_spin(&sleepq_prof_lock); 1156 prof_enabled = enabled; 1157 mtx_unlock_spin(&sleepq_prof_lock); 1158 1159 error = sbuf_finish(sb); 1160 sbuf_delete(sb); 1161 return (error); 1162 } 1163 1164 SYSCTL_PROC(_debug_sleepq, OID_AUTO, stats, CTLTYPE_STRING | CTLFLAG_RD, 1165 NULL, 0, dump_sleepq_prof_stats, "A", "Sleepqueue profiling statistics"); 1166 SYSCTL_PROC(_debug_sleepq, OID_AUTO, reset, CTLTYPE_INT | CTLFLAG_RW, 1167 NULL, 0, reset_sleepq_prof_stats, "I", 1168 "Reset sleepqueue profiling statistics"); 1169 SYSCTL_PROC(_debug_sleepq, OID_AUTO, enable, CTLTYPE_INT | CTLFLAG_RW, 1170 NULL, 0, enable_sleepq_prof, "I", "Enable sleepqueue profiling"); 1171 #endif 1172 1173 #ifdef DDB 1174 DB_SHOW_COMMAND(sleepq, db_show_sleepqueue) 1175 { 1176 struct sleepqueue_chain *sc; 1177 struct sleepqueue *sq; 1178 #ifdef INVARIANTS 1179 struct lock_object *lock; 1180 #endif 1181 struct thread *td; 1182 void *wchan; 1183 int i; 1184 1185 if (!have_addr) 1186 return; 1187 1188 /* 1189 * First, see if there is an active sleep queue for the wait channel 1190 * indicated by the address. 1191 */ 1192 wchan = (void *)addr; 1193 sc = SC_LOOKUP(wchan); 1194 LIST_FOREACH(sq, &sc->sc_queues, sq_hash) 1195 if (sq->sq_wchan == wchan) 1196 goto found; 1197 1198 /* 1199 * Second, see if there is an active sleep queue at the address 1200 * indicated. 1201 */ 1202 for (i = 0; i < SC_TABLESIZE; i++) 1203 LIST_FOREACH(sq, &sleepq_chains[i].sc_queues, sq_hash) { 1204 if (sq == (struct sleepqueue *)addr) 1205 goto found; 1206 } 1207 1208 db_printf("Unable to locate a sleep queue via %p\n", (void *)addr); 1209 return; 1210 found: 1211 db_printf("Wait channel: %p\n", sq->sq_wchan); 1212 db_printf("Queue type: %d\n", sq->sq_type); 1213 #ifdef INVARIANTS 1214 if (sq->sq_lock) { 1215 lock = sq->sq_lock; 1216 db_printf("Associated Interlock: %p - (%s) %s\n", lock, 1217 LOCK_CLASS(lock)->lc_name, lock->lo_name); 1218 } 1219 #endif 1220 db_printf("Blocked threads:\n"); 1221 for (i = 0; i < NR_SLEEPQS; i++) { 1222 db_printf("\nQueue[%d]:\n", i); 1223 if (TAILQ_EMPTY(&sq->sq_blocked[i])) 1224 db_printf("\tempty\n"); 1225 else 1226 TAILQ_FOREACH(td, &sq->sq_blocked[0], 1227 td_slpq) { 1228 db_printf("\t%p (tid %d, pid %d, \"%s\")\n", td, 1229 td->td_tid, td->td_proc->p_pid, 1230 td->td_name); 1231 } 1232 db_printf("(expected: %u)\n", sq->sq_blockedcnt[i]); 1233 } 1234 } 1235 1236 /* Alias 'show sleepqueue' to 'show sleepq'. */ 1237 DB_SHOW_ALIAS(sleepqueue, db_show_sleepqueue); 1238 #endif 1239