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 if (flags & SLEEPQ_STOP_ON_BDRY) 356 td->td_flags |= TDF_SBDRY; 357 } 358 thread_unlock(td); 359 } 360 361 /* 362 * Sets a timeout that will remove the current thread from the specified 363 * sleep queue after timo ticks if the thread has not already been awakened. 364 */ 365 void 366 sleepq_set_timeout(void *wchan, int timo) 367 { 368 struct sleepqueue_chain *sc; 369 struct thread *td; 370 371 td = curthread; 372 sc = SC_LOOKUP(wchan); 373 mtx_assert(&sc->sc_lock, MA_OWNED); 374 MPASS(TD_ON_SLEEPQ(td)); 375 MPASS(td->td_sleepqueue == NULL); 376 MPASS(wchan != NULL); 377 callout_reset_curcpu(&td->td_slpcallout, timo, sleepq_timeout, td); 378 } 379 380 /* 381 * Return the number of actual sleepers for the specified queue. 382 */ 383 u_int 384 sleepq_sleepcnt(void *wchan, int queue) 385 { 386 struct sleepqueue *sq; 387 388 KASSERT(wchan != NULL, ("%s: invalid NULL wait channel", __func__)); 389 MPASS((queue >= 0) && (queue < NR_SLEEPQS)); 390 sq = sleepq_lookup(wchan); 391 if (sq == NULL) 392 return (0); 393 return (sq->sq_blockedcnt[queue]); 394 } 395 396 /* 397 * Marks the pending sleep of the current thread as interruptible and 398 * makes an initial check for pending signals before putting a thread 399 * to sleep. Enters and exits with the thread lock held. Thread lock 400 * may have transitioned from the sleepq lock to a run lock. 401 */ 402 static int 403 sleepq_catch_signals(void *wchan, int pri) 404 { 405 struct sleepqueue_chain *sc; 406 struct sleepqueue *sq; 407 struct thread *td; 408 struct proc *p; 409 struct sigacts *ps; 410 int sig, ret, stop_allowed; 411 412 td = curthread; 413 p = curproc; 414 sc = SC_LOOKUP(wchan); 415 mtx_assert(&sc->sc_lock, MA_OWNED); 416 MPASS(wchan != NULL); 417 if ((td->td_pflags & TDP_WAKEUP) != 0) { 418 td->td_pflags &= ~TDP_WAKEUP; 419 ret = EINTR; 420 thread_lock(td); 421 goto out; 422 } 423 424 /* 425 * See if there are any pending signals for this thread. If not 426 * we can switch immediately. Otherwise do the signal processing 427 * directly. 428 */ 429 thread_lock(td); 430 if ((td->td_flags & (TDF_NEEDSIGCHK | TDF_NEEDSUSPCHK)) == 0) { 431 sleepq_switch(wchan, pri); 432 return (0); 433 } 434 stop_allowed = (td->td_flags & TDF_SBDRY) ? SIG_STOP_NOT_ALLOWED : 435 SIG_STOP_ALLOWED; 436 thread_unlock(td); 437 mtx_unlock_spin(&sc->sc_lock); 438 CTR3(KTR_PROC, "sleepq catching signals: thread %p (pid %ld, %s)", 439 (void *)td, (long)p->p_pid, td->td_name); 440 PROC_LOCK(p); 441 ps = p->p_sigacts; 442 mtx_lock(&ps->ps_mtx); 443 sig = cursig(td, stop_allowed); 444 if (sig == 0) { 445 mtx_unlock(&ps->ps_mtx); 446 ret = thread_suspend_check(1); 447 MPASS(ret == 0 || ret == EINTR || ret == ERESTART); 448 } else { 449 if (SIGISMEMBER(ps->ps_sigintr, sig)) 450 ret = EINTR; 451 else 452 ret = ERESTART; 453 mtx_unlock(&ps->ps_mtx); 454 } 455 /* 456 * Lock the per-process spinlock prior to dropping the PROC_LOCK 457 * to avoid a signal delivery race. PROC_LOCK, PROC_SLOCK, and 458 * thread_lock() are currently held in tdsendsignal(). 459 */ 460 PROC_SLOCK(p); 461 mtx_lock_spin(&sc->sc_lock); 462 PROC_UNLOCK(p); 463 thread_lock(td); 464 PROC_SUNLOCK(p); 465 if (ret == 0) { 466 sleepq_switch(wchan, pri); 467 return (0); 468 } 469 out: 470 /* 471 * There were pending signals and this thread is still 472 * on the sleep queue, remove it from the sleep queue. 473 */ 474 if (TD_ON_SLEEPQ(td)) { 475 sq = sleepq_lookup(wchan); 476 if (sleepq_resume_thread(sq, td, 0)) { 477 #ifdef INVARIANTS 478 /* 479 * This thread hasn't gone to sleep yet, so it 480 * should not be swapped out. 481 */ 482 panic("not waking up swapper"); 483 #endif 484 } 485 } 486 mtx_unlock_spin(&sc->sc_lock); 487 MPASS(td->td_lock != &sc->sc_lock); 488 return (ret); 489 } 490 491 /* 492 * Switches to another thread if we are still asleep on a sleep queue. 493 * Returns with thread lock. 494 */ 495 static void 496 sleepq_switch(void *wchan, int pri) 497 { 498 struct sleepqueue_chain *sc; 499 struct sleepqueue *sq; 500 struct thread *td; 501 502 td = curthread; 503 sc = SC_LOOKUP(wchan); 504 mtx_assert(&sc->sc_lock, MA_OWNED); 505 THREAD_LOCK_ASSERT(td, MA_OWNED); 506 507 /* 508 * If we have a sleep queue, then we've already been woken up, so 509 * just return. 510 */ 511 if (td->td_sleepqueue != NULL) { 512 mtx_unlock_spin(&sc->sc_lock); 513 return; 514 } 515 516 /* 517 * If TDF_TIMEOUT is set, then our sleep has been timed out 518 * already but we are still on the sleep queue, so dequeue the 519 * thread and return. 520 */ 521 if (td->td_flags & TDF_TIMEOUT) { 522 MPASS(TD_ON_SLEEPQ(td)); 523 sq = sleepq_lookup(wchan); 524 if (sleepq_resume_thread(sq, td, 0)) { 525 #ifdef INVARIANTS 526 /* 527 * This thread hasn't gone to sleep yet, so it 528 * should not be swapped out. 529 */ 530 panic("not waking up swapper"); 531 #endif 532 } 533 mtx_unlock_spin(&sc->sc_lock); 534 return; 535 } 536 #ifdef SLEEPQUEUE_PROFILING 537 if (prof_enabled) 538 sleepq_profile(td->td_wmesg); 539 #endif 540 MPASS(td->td_sleepqueue == NULL); 541 sched_sleep(td, pri); 542 thread_lock_set(td, &sc->sc_lock); 543 SDT_PROBE0(sched, , , sleep); 544 TD_SET_SLEEPING(td); 545 mi_switch(SW_VOL | SWT_SLEEPQ, NULL); 546 KASSERT(TD_IS_RUNNING(td), ("running but not TDS_RUNNING")); 547 CTR3(KTR_PROC, "sleepq resume: thread %p (pid %ld, %s)", 548 (void *)td, (long)td->td_proc->p_pid, (void *)td->td_name); 549 } 550 551 /* 552 * Check to see if we timed out. 553 */ 554 static int 555 sleepq_check_timeout(void) 556 { 557 struct thread *td; 558 559 td = curthread; 560 THREAD_LOCK_ASSERT(td, MA_OWNED); 561 562 /* 563 * If TDF_TIMEOUT is set, we timed out. 564 */ 565 if (td->td_flags & TDF_TIMEOUT) { 566 td->td_flags &= ~TDF_TIMEOUT; 567 return (EWOULDBLOCK); 568 } 569 570 /* 571 * If TDF_TIMOFAIL is set, the timeout ran after we had 572 * already been woken up. 573 */ 574 if (td->td_flags & TDF_TIMOFAIL) 575 td->td_flags &= ~TDF_TIMOFAIL; 576 577 /* 578 * If callout_stop() fails, then the timeout is running on 579 * another CPU, so synchronize with it to avoid having it 580 * accidentally wake up a subsequent sleep. 581 */ 582 else if (callout_stop(&td->td_slpcallout) == 0) { 583 td->td_flags |= TDF_TIMEOUT; 584 TD_SET_SLEEPING(td); 585 mi_switch(SW_INVOL | SWT_SLEEPQTIMO, NULL); 586 } 587 return (0); 588 } 589 590 /* 591 * Check to see if we were awoken by a signal. 592 */ 593 static int 594 sleepq_check_signals(void) 595 { 596 struct thread *td; 597 598 td = curthread; 599 THREAD_LOCK_ASSERT(td, MA_OWNED); 600 601 /* We are no longer in an interruptible sleep. */ 602 if (td->td_flags & TDF_SINTR) 603 td->td_flags &= ~(TDF_SINTR | TDF_SBDRY); 604 605 if (td->td_flags & TDF_SLEEPABORT) { 606 td->td_flags &= ~TDF_SLEEPABORT; 607 return (td->td_intrval); 608 } 609 610 return (0); 611 } 612 613 /* 614 * Block the current thread until it is awakened from its sleep queue. 615 */ 616 void 617 sleepq_wait(void *wchan, int pri) 618 { 619 struct thread *td; 620 621 td = curthread; 622 MPASS(!(td->td_flags & TDF_SINTR)); 623 thread_lock(td); 624 sleepq_switch(wchan, pri); 625 thread_unlock(td); 626 } 627 628 /* 629 * Block the current thread until it is awakened from its sleep queue 630 * or it is interrupted by a signal. 631 */ 632 int 633 sleepq_wait_sig(void *wchan, int pri) 634 { 635 int rcatch; 636 int rval; 637 638 rcatch = sleepq_catch_signals(wchan, pri); 639 rval = sleepq_check_signals(); 640 thread_unlock(curthread); 641 if (rcatch) 642 return (rcatch); 643 return (rval); 644 } 645 646 /* 647 * Block the current thread until it is awakened from its sleep queue 648 * or it times out while waiting. 649 */ 650 int 651 sleepq_timedwait(void *wchan, int pri) 652 { 653 struct thread *td; 654 int rval; 655 656 td = curthread; 657 MPASS(!(td->td_flags & TDF_SINTR)); 658 thread_lock(td); 659 sleepq_switch(wchan, pri); 660 rval = sleepq_check_timeout(); 661 thread_unlock(td); 662 663 return (rval); 664 } 665 666 /* 667 * Block the current thread until it is awakened from its sleep queue, 668 * it is interrupted by a signal, or it times out waiting to be awakened. 669 */ 670 int 671 sleepq_timedwait_sig(void *wchan, int pri) 672 { 673 int rcatch, rvalt, rvals; 674 675 rcatch = sleepq_catch_signals(wchan, pri); 676 rvalt = sleepq_check_timeout(); 677 rvals = sleepq_check_signals(); 678 thread_unlock(curthread); 679 if (rcatch) 680 return (rcatch); 681 if (rvals) 682 return (rvals); 683 return (rvalt); 684 } 685 686 /* 687 * Returns the type of sleepqueue given a waitchannel. 688 */ 689 int 690 sleepq_type(void *wchan) 691 { 692 struct sleepqueue *sq; 693 int type; 694 695 MPASS(wchan != NULL); 696 697 sleepq_lock(wchan); 698 sq = sleepq_lookup(wchan); 699 if (sq == NULL) { 700 sleepq_release(wchan); 701 return (-1); 702 } 703 type = sq->sq_type; 704 sleepq_release(wchan); 705 return (type); 706 } 707 708 /* 709 * Removes a thread from a sleep queue and makes it 710 * runnable. 711 */ 712 static int 713 sleepq_resume_thread(struct sleepqueue *sq, struct thread *td, int pri) 714 { 715 struct sleepqueue_chain *sc; 716 717 MPASS(td != NULL); 718 MPASS(sq->sq_wchan != NULL); 719 MPASS(td->td_wchan == sq->sq_wchan); 720 MPASS(td->td_sqqueue < NR_SLEEPQS && td->td_sqqueue >= 0); 721 THREAD_LOCK_ASSERT(td, MA_OWNED); 722 sc = SC_LOOKUP(sq->sq_wchan); 723 mtx_assert(&sc->sc_lock, MA_OWNED); 724 725 SDT_PROBE2(sched, , , wakeup, td, td->td_proc); 726 727 /* Remove the thread from the queue. */ 728 sq->sq_blockedcnt[td->td_sqqueue]--; 729 TAILQ_REMOVE(&sq->sq_blocked[td->td_sqqueue], td, td_slpq); 730 731 /* 732 * Get a sleep queue for this thread. If this is the last waiter, 733 * use the queue itself and take it out of the chain, otherwise, 734 * remove a queue from the free list. 735 */ 736 if (LIST_EMPTY(&sq->sq_free)) { 737 td->td_sleepqueue = sq; 738 #ifdef INVARIANTS 739 sq->sq_wchan = NULL; 740 #endif 741 #ifdef SLEEPQUEUE_PROFILING 742 sc->sc_depth--; 743 #endif 744 } else 745 td->td_sleepqueue = LIST_FIRST(&sq->sq_free); 746 LIST_REMOVE(td->td_sleepqueue, sq_hash); 747 748 td->td_wmesg = NULL; 749 td->td_wchan = NULL; 750 td->td_flags &= ~(TDF_SINTR | TDF_SBDRY); 751 752 CTR3(KTR_PROC, "sleepq_wakeup: thread %p (pid %ld, %s)", 753 (void *)td, (long)td->td_proc->p_pid, td->td_name); 754 755 /* Adjust priority if requested. */ 756 MPASS(pri == 0 || (pri >= PRI_MIN && pri <= PRI_MAX)); 757 if (pri != 0 && td->td_priority > pri && 758 PRI_BASE(td->td_pri_class) == PRI_TIMESHARE) 759 sched_prio(td, pri); 760 761 /* 762 * Note that thread td might not be sleeping if it is running 763 * sleepq_catch_signals() on another CPU or is blocked on its 764 * proc lock to check signals. There's no need to mark the 765 * thread runnable in that case. 766 */ 767 if (TD_IS_SLEEPING(td)) { 768 TD_CLR_SLEEPING(td); 769 return (setrunnable(td)); 770 } 771 return (0); 772 } 773 774 #ifdef INVARIANTS 775 /* 776 * UMA zone item deallocator. 777 */ 778 static void 779 sleepq_dtor(void *mem, int size, void *arg) 780 { 781 struct sleepqueue *sq; 782 int i; 783 784 sq = mem; 785 for (i = 0; i < NR_SLEEPQS; i++) { 786 MPASS(TAILQ_EMPTY(&sq->sq_blocked[i])); 787 MPASS(sq->sq_blockedcnt[i] == 0); 788 } 789 } 790 #endif 791 792 /* 793 * UMA zone item initializer. 794 */ 795 static int 796 sleepq_init(void *mem, int size, int flags) 797 { 798 struct sleepqueue *sq; 799 int i; 800 801 bzero(mem, size); 802 sq = mem; 803 for (i = 0; i < NR_SLEEPQS; i++) { 804 TAILQ_INIT(&sq->sq_blocked[i]); 805 sq->sq_blockedcnt[i] = 0; 806 } 807 LIST_INIT(&sq->sq_free); 808 return (0); 809 } 810 811 /* 812 * Find the highest priority thread sleeping on a wait channel and resume it. 813 */ 814 int 815 sleepq_signal(void *wchan, int flags, int pri, int queue) 816 { 817 struct sleepqueue *sq; 818 struct thread *td, *besttd; 819 int wakeup_swapper; 820 821 CTR2(KTR_PROC, "sleepq_signal(%p, %d)", wchan, flags); 822 KASSERT(wchan != NULL, ("%s: invalid NULL wait channel", __func__)); 823 MPASS((queue >= 0) && (queue < NR_SLEEPQS)); 824 sq = sleepq_lookup(wchan); 825 if (sq == NULL) 826 return (0); 827 KASSERT(sq->sq_type == (flags & SLEEPQ_TYPE), 828 ("%s: mismatch between sleep/wakeup and cv_*", __func__)); 829 830 /* 831 * Find the highest priority thread on the queue. If there is a 832 * tie, use the thread that first appears in the queue as it has 833 * been sleeping the longest since threads are always added to 834 * the tail of sleep queues. 835 */ 836 besttd = NULL; 837 TAILQ_FOREACH(td, &sq->sq_blocked[queue], td_slpq) { 838 if (besttd == NULL || td->td_priority < besttd->td_priority) 839 besttd = td; 840 } 841 MPASS(besttd != NULL); 842 thread_lock(besttd); 843 wakeup_swapper = sleepq_resume_thread(sq, besttd, pri); 844 thread_unlock(besttd); 845 return (wakeup_swapper); 846 } 847 848 /* 849 * Resume all threads sleeping on a specified wait channel. 850 */ 851 int 852 sleepq_broadcast(void *wchan, int flags, int pri, int queue) 853 { 854 struct sleepqueue *sq; 855 struct thread *td, *tdn; 856 int wakeup_swapper; 857 858 CTR2(KTR_PROC, "sleepq_broadcast(%p, %d)", wchan, flags); 859 KASSERT(wchan != NULL, ("%s: invalid NULL wait channel", __func__)); 860 MPASS((queue >= 0) && (queue < NR_SLEEPQS)); 861 sq = sleepq_lookup(wchan); 862 if (sq == NULL) 863 return (0); 864 KASSERT(sq->sq_type == (flags & SLEEPQ_TYPE), 865 ("%s: mismatch between sleep/wakeup and cv_*", __func__)); 866 867 /* Resume all blocked threads on the sleep queue. */ 868 wakeup_swapper = 0; 869 TAILQ_FOREACH_SAFE(td, &sq->sq_blocked[queue], td_slpq, tdn) { 870 thread_lock(td); 871 if (sleepq_resume_thread(sq, td, pri)) 872 wakeup_swapper = 1; 873 thread_unlock(td); 874 } 875 return (wakeup_swapper); 876 } 877 878 /* 879 * Time sleeping threads out. When the timeout expires, the thread is 880 * removed from the sleep queue and made runnable if it is still asleep. 881 */ 882 static void 883 sleepq_timeout(void *arg) 884 { 885 struct sleepqueue_chain *sc; 886 struct sleepqueue *sq; 887 struct thread *td; 888 void *wchan; 889 int wakeup_swapper; 890 891 td = arg; 892 wakeup_swapper = 0; 893 CTR3(KTR_PROC, "sleepq_timeout: thread %p (pid %ld, %s)", 894 (void *)td, (long)td->td_proc->p_pid, (void *)td->td_name); 895 896 /* 897 * First, see if the thread is asleep and get the wait channel if 898 * it is. 899 */ 900 thread_lock(td); 901 if (TD_IS_SLEEPING(td) && TD_ON_SLEEPQ(td)) { 902 wchan = td->td_wchan; 903 sc = SC_LOOKUP(wchan); 904 THREAD_LOCKPTR_ASSERT(td, &sc->sc_lock); 905 sq = sleepq_lookup(wchan); 906 MPASS(sq != NULL); 907 td->td_flags |= TDF_TIMEOUT; 908 wakeup_swapper = sleepq_resume_thread(sq, td, 0); 909 thread_unlock(td); 910 if (wakeup_swapper) 911 kick_proc0(); 912 return; 913 } 914 915 /* 916 * If the thread is on the SLEEPQ but isn't sleeping yet, it 917 * can either be on another CPU in between sleepq_add() and 918 * one of the sleepq_*wait*() routines or it can be in 919 * sleepq_catch_signals(). 920 */ 921 if (TD_ON_SLEEPQ(td)) { 922 td->td_flags |= TDF_TIMEOUT; 923 thread_unlock(td); 924 return; 925 } 926 927 /* 928 * Now check for the edge cases. First, if TDF_TIMEOUT is set, 929 * then the other thread has already yielded to us, so clear 930 * the flag and resume it. If TDF_TIMEOUT is not set, then the 931 * we know that the other thread is not on a sleep queue, but it 932 * hasn't resumed execution yet. In that case, set TDF_TIMOFAIL 933 * to let it know that the timeout has already run and doesn't 934 * need to be canceled. 935 */ 936 if (td->td_flags & TDF_TIMEOUT) { 937 MPASS(TD_IS_SLEEPING(td)); 938 td->td_flags &= ~TDF_TIMEOUT; 939 TD_CLR_SLEEPING(td); 940 wakeup_swapper = setrunnable(td); 941 } else 942 td->td_flags |= TDF_TIMOFAIL; 943 thread_unlock(td); 944 if (wakeup_swapper) 945 kick_proc0(); 946 } 947 948 /* 949 * Resumes a specific thread from the sleep queue associated with a specific 950 * wait channel if it is on that queue. 951 */ 952 void 953 sleepq_remove(struct thread *td, void *wchan) 954 { 955 struct sleepqueue *sq; 956 int wakeup_swapper; 957 958 /* 959 * Look up the sleep queue for this wait channel, then re-check 960 * that the thread is asleep on that channel, if it is not, then 961 * bail. 962 */ 963 MPASS(wchan != NULL); 964 sleepq_lock(wchan); 965 sq = sleepq_lookup(wchan); 966 /* 967 * We can not lock the thread here as it may be sleeping on a 968 * different sleepq. However, holding the sleepq lock for this 969 * wchan can guarantee that we do not miss a wakeup for this 970 * channel. The asserts below will catch any false positives. 971 */ 972 if (!TD_ON_SLEEPQ(td) || td->td_wchan != wchan) { 973 sleepq_release(wchan); 974 return; 975 } 976 /* Thread is asleep on sleep queue sq, so wake it up. */ 977 thread_lock(td); 978 MPASS(sq != NULL); 979 MPASS(td->td_wchan == wchan); 980 wakeup_swapper = sleepq_resume_thread(sq, td, 0); 981 thread_unlock(td); 982 sleepq_release(wchan); 983 if (wakeup_swapper) 984 kick_proc0(); 985 } 986 987 /* 988 * Abort a thread as if an interrupt had occurred. Only abort 989 * interruptible waits (unfortunately it isn't safe to abort others). 990 */ 991 int 992 sleepq_abort(struct thread *td, int intrval) 993 { 994 struct sleepqueue *sq; 995 void *wchan; 996 997 THREAD_LOCK_ASSERT(td, MA_OWNED); 998 MPASS(TD_ON_SLEEPQ(td)); 999 MPASS(td->td_flags & TDF_SINTR); 1000 MPASS(intrval == EINTR || intrval == ERESTART); 1001 1002 /* 1003 * If the TDF_TIMEOUT flag is set, just leave. A 1004 * timeout is scheduled anyhow. 1005 */ 1006 if (td->td_flags & TDF_TIMEOUT) 1007 return (0); 1008 1009 CTR3(KTR_PROC, "sleepq_abort: thread %p (pid %ld, %s)", 1010 (void *)td, (long)td->td_proc->p_pid, (void *)td->td_name); 1011 td->td_intrval = intrval; 1012 td->td_flags |= TDF_SLEEPABORT; 1013 /* 1014 * If the thread has not slept yet it will find the signal in 1015 * sleepq_catch_signals() and call sleepq_resume_thread. Otherwise 1016 * we have to do it here. 1017 */ 1018 if (!TD_IS_SLEEPING(td)) 1019 return (0); 1020 wchan = td->td_wchan; 1021 MPASS(wchan != NULL); 1022 sq = sleepq_lookup(wchan); 1023 MPASS(sq != NULL); 1024 1025 /* Thread is asleep on sleep queue sq, so wake it up. */ 1026 return (sleepq_resume_thread(sq, td, 0)); 1027 } 1028 1029 #ifdef SLEEPQUEUE_PROFILING 1030 #define SLEEPQ_PROF_LOCATIONS 1024 1031 #define SLEEPQ_SBUFSIZE 512 1032 struct sleepq_prof { 1033 LIST_ENTRY(sleepq_prof) sp_link; 1034 const char *sp_wmesg; 1035 long sp_count; 1036 }; 1037 1038 LIST_HEAD(sqphead, sleepq_prof); 1039 1040 struct sqphead sleepq_prof_free; 1041 struct sqphead sleepq_hash[SC_TABLESIZE]; 1042 static struct sleepq_prof sleepq_profent[SLEEPQ_PROF_LOCATIONS]; 1043 static struct mtx sleepq_prof_lock; 1044 MTX_SYSINIT(sleepq_prof_lock, &sleepq_prof_lock, "sleepq_prof", MTX_SPIN); 1045 1046 static void 1047 sleepq_profile(const char *wmesg) 1048 { 1049 struct sleepq_prof *sp; 1050 1051 mtx_lock_spin(&sleepq_prof_lock); 1052 if (prof_enabled == 0) 1053 goto unlock; 1054 LIST_FOREACH(sp, &sleepq_hash[SC_HASH(wmesg)], sp_link) 1055 if (sp->sp_wmesg == wmesg) 1056 goto done; 1057 sp = LIST_FIRST(&sleepq_prof_free); 1058 if (sp == NULL) 1059 goto unlock; 1060 sp->sp_wmesg = wmesg; 1061 LIST_REMOVE(sp, sp_link); 1062 LIST_INSERT_HEAD(&sleepq_hash[SC_HASH(wmesg)], sp, sp_link); 1063 done: 1064 sp->sp_count++; 1065 unlock: 1066 mtx_unlock_spin(&sleepq_prof_lock); 1067 return; 1068 } 1069 1070 static void 1071 sleepq_prof_reset(void) 1072 { 1073 struct sleepq_prof *sp; 1074 int enabled; 1075 int i; 1076 1077 mtx_lock_spin(&sleepq_prof_lock); 1078 enabled = prof_enabled; 1079 prof_enabled = 0; 1080 for (i = 0; i < SC_TABLESIZE; i++) 1081 LIST_INIT(&sleepq_hash[i]); 1082 LIST_INIT(&sleepq_prof_free); 1083 for (i = 0; i < SLEEPQ_PROF_LOCATIONS; i++) { 1084 sp = &sleepq_profent[i]; 1085 sp->sp_wmesg = NULL; 1086 sp->sp_count = 0; 1087 LIST_INSERT_HEAD(&sleepq_prof_free, sp, sp_link); 1088 } 1089 prof_enabled = enabled; 1090 mtx_unlock_spin(&sleepq_prof_lock); 1091 } 1092 1093 static int 1094 enable_sleepq_prof(SYSCTL_HANDLER_ARGS) 1095 { 1096 int error, v; 1097 1098 v = prof_enabled; 1099 error = sysctl_handle_int(oidp, &v, v, req); 1100 if (error) 1101 return (error); 1102 if (req->newptr == NULL) 1103 return (error); 1104 if (v == prof_enabled) 1105 return (0); 1106 if (v == 1) 1107 sleepq_prof_reset(); 1108 mtx_lock_spin(&sleepq_prof_lock); 1109 prof_enabled = !!v; 1110 mtx_unlock_spin(&sleepq_prof_lock); 1111 1112 return (0); 1113 } 1114 1115 static int 1116 reset_sleepq_prof_stats(SYSCTL_HANDLER_ARGS) 1117 { 1118 int error, v; 1119 1120 v = 0; 1121 error = sysctl_handle_int(oidp, &v, 0, req); 1122 if (error) 1123 return (error); 1124 if (req->newptr == NULL) 1125 return (error); 1126 if (v == 0) 1127 return (0); 1128 sleepq_prof_reset(); 1129 1130 return (0); 1131 } 1132 1133 static int 1134 dump_sleepq_prof_stats(SYSCTL_HANDLER_ARGS) 1135 { 1136 struct sleepq_prof *sp; 1137 struct sbuf *sb; 1138 int enabled; 1139 int error; 1140 int i; 1141 1142 error = sysctl_wire_old_buffer(req, 0); 1143 if (error != 0) 1144 return (error); 1145 sb = sbuf_new_for_sysctl(NULL, NULL, SLEEPQ_SBUFSIZE, req); 1146 sbuf_printf(sb, "\nwmesg\tcount\n"); 1147 enabled = prof_enabled; 1148 mtx_lock_spin(&sleepq_prof_lock); 1149 prof_enabled = 0; 1150 mtx_unlock_spin(&sleepq_prof_lock); 1151 for (i = 0; i < SC_TABLESIZE; i++) { 1152 LIST_FOREACH(sp, &sleepq_hash[i], sp_link) { 1153 sbuf_printf(sb, "%s\t%ld\n", 1154 sp->sp_wmesg, sp->sp_count); 1155 } 1156 } 1157 mtx_lock_spin(&sleepq_prof_lock); 1158 prof_enabled = enabled; 1159 mtx_unlock_spin(&sleepq_prof_lock); 1160 1161 error = sbuf_finish(sb); 1162 sbuf_delete(sb); 1163 return (error); 1164 } 1165 1166 SYSCTL_PROC(_debug_sleepq, OID_AUTO, stats, CTLTYPE_STRING | CTLFLAG_RD, 1167 NULL, 0, dump_sleepq_prof_stats, "A", "Sleepqueue profiling statistics"); 1168 SYSCTL_PROC(_debug_sleepq, OID_AUTO, reset, CTLTYPE_INT | CTLFLAG_RW, 1169 NULL, 0, reset_sleepq_prof_stats, "I", 1170 "Reset sleepqueue profiling statistics"); 1171 SYSCTL_PROC(_debug_sleepq, OID_AUTO, enable, CTLTYPE_INT | CTLFLAG_RW, 1172 NULL, 0, enable_sleepq_prof, "I", "Enable sleepqueue profiling"); 1173 #endif 1174 1175 #ifdef DDB 1176 DB_SHOW_COMMAND(sleepq, db_show_sleepqueue) 1177 { 1178 struct sleepqueue_chain *sc; 1179 struct sleepqueue *sq; 1180 #ifdef INVARIANTS 1181 struct lock_object *lock; 1182 #endif 1183 struct thread *td; 1184 void *wchan; 1185 int i; 1186 1187 if (!have_addr) 1188 return; 1189 1190 /* 1191 * First, see if there is an active sleep queue for the wait channel 1192 * indicated by the address. 1193 */ 1194 wchan = (void *)addr; 1195 sc = SC_LOOKUP(wchan); 1196 LIST_FOREACH(sq, &sc->sc_queues, sq_hash) 1197 if (sq->sq_wchan == wchan) 1198 goto found; 1199 1200 /* 1201 * Second, see if there is an active sleep queue at the address 1202 * indicated. 1203 */ 1204 for (i = 0; i < SC_TABLESIZE; i++) 1205 LIST_FOREACH(sq, &sleepq_chains[i].sc_queues, sq_hash) { 1206 if (sq == (struct sleepqueue *)addr) 1207 goto found; 1208 } 1209 1210 db_printf("Unable to locate a sleep queue via %p\n", (void *)addr); 1211 return; 1212 found: 1213 db_printf("Wait channel: %p\n", sq->sq_wchan); 1214 db_printf("Queue type: %d\n", sq->sq_type); 1215 #ifdef INVARIANTS 1216 if (sq->sq_lock) { 1217 lock = sq->sq_lock; 1218 db_printf("Associated Interlock: %p - (%s) %s\n", lock, 1219 LOCK_CLASS(lock)->lc_name, lock->lo_name); 1220 } 1221 #endif 1222 db_printf("Blocked threads:\n"); 1223 for (i = 0; i < NR_SLEEPQS; i++) { 1224 db_printf("\nQueue[%d]:\n", i); 1225 if (TAILQ_EMPTY(&sq->sq_blocked[i])) 1226 db_printf("\tempty\n"); 1227 else 1228 TAILQ_FOREACH(td, &sq->sq_blocked[0], 1229 td_slpq) { 1230 db_printf("\t%p (tid %d, pid %d, \"%s\")\n", td, 1231 td->td_tid, td->td_proc->p_pid, 1232 td->td_name); 1233 } 1234 db_printf("(expected: %u)\n", sq->sq_blockedcnt[i]); 1235 } 1236 } 1237 1238 /* Alias 'show sleepqueue' to 'show sleepq'. */ 1239 DB_SHOW_ALIAS(sleepqueue, db_show_sleepqueue); 1240 #endif 1241