subr_sleepqueue.c revision 177471
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: head/sys/kern/subr_sleepqueue.c 177471 2008-03-21 08:23:25Z jeff $"); 64 65#include "opt_sleepqueue_profiling.h" 66#include "opt_ddb.h" 67#include "opt_sched.h" 68 69#include <sys/param.h> 70#include <sys/systm.h> 71#include <sys/lock.h> 72#include <sys/kernel.h> 73#include <sys/ktr.h> 74#include <sys/mutex.h> 75#include <sys/proc.h> 76#include <sys/sbuf.h> 77#include <sys/sched.h> 78#include <sys/signalvar.h> 79#include <sys/sleepqueue.h> 80#include <sys/sysctl.h> 81 82#include <vm/uma.h> 83 84#ifdef DDB 85#include <ddb/ddb.h> 86#endif 87 88/* 89 * Constants for the hash table of sleep queue chains. These constants are 90 * the same ones that 4BSD (and possibly earlier versions of BSD) used. 91 * Basically, we ignore the lower 8 bits of the address since most wait 92 * channel pointers are aligned and only look at the next 7 bits for the 93 * hash. SC_TABLESIZE must be a power of two for SC_MASK to work properly. 94 */ 95#define SC_TABLESIZE 128 /* Must be power of 2. */ 96#define SC_MASK (SC_TABLESIZE - 1) 97#define SC_SHIFT 8 98#define SC_HASH(wc) (((uintptr_t)(wc) >> SC_SHIFT) & SC_MASK) 99#define SC_LOOKUP(wc) &sleepq_chains[SC_HASH(wc)] 100#define NR_SLEEPQS 2 101/* 102 * There two different lists of sleep queues. Both lists are connected 103 * via the sq_hash entries. The first list is the sleep queue chain list 104 * that a sleep queue is on when it is attached to a wait channel. The 105 * second list is the free list hung off of a sleep queue that is attached 106 * to a wait channel. 107 * 108 * Each sleep queue also contains the wait channel it is attached to, the 109 * list of threads blocked on that wait channel, flags specific to the 110 * wait channel, and the lock used to synchronize with a wait channel. 111 * The flags are used to catch mismatches between the various consumers 112 * of the sleep queue API (e.g. sleep/wakeup and condition variables). 113 * The lock pointer is only used when invariants are enabled for various 114 * debugging checks. 115 * 116 * Locking key: 117 * c - sleep queue chain lock 118 */ 119struct sleepqueue { 120 TAILQ_HEAD(, thread) sq_blocked[NR_SLEEPQS]; /* (c) Blocked threads. */ 121 LIST_ENTRY(sleepqueue) sq_hash; /* (c) Chain and free list. */ 122 LIST_HEAD(, sleepqueue) sq_free; /* (c) Free queues. */ 123 void *sq_wchan; /* (c) Wait channel. */ 124#ifdef INVARIANTS 125 int sq_type; /* (c) Queue type. */ 126 struct lock_object *sq_lock; /* (c) Associated lock. */ 127#endif 128}; 129 130struct sleepqueue_chain { 131 LIST_HEAD(, sleepqueue) sc_queues; /* List of sleep queues. */ 132 struct mtx sc_lock; /* Spin lock for this chain. */ 133#ifdef SLEEPQUEUE_PROFILING 134 u_int sc_depth; /* Length of sc_queues. */ 135 u_int sc_max_depth; /* Max length of sc_queues. */ 136#endif 137}; 138 139#ifdef SLEEPQUEUE_PROFILING 140u_int sleepq_max_depth; 141SYSCTL_NODE(_debug, OID_AUTO, sleepq, CTLFLAG_RD, 0, "sleepq profiling"); 142SYSCTL_NODE(_debug_sleepq, OID_AUTO, chains, CTLFLAG_RD, 0, 143 "sleepq chain stats"); 144SYSCTL_UINT(_debug_sleepq, OID_AUTO, max_depth, CTLFLAG_RD, &sleepq_max_depth, 145 0, "maxmimum depth achieved of a single chain"); 146 147static void sleepq_profile(const char *wmesg); 148static int prof_enabled; 149#endif 150static struct sleepqueue_chain sleepq_chains[SC_TABLESIZE]; 151static uma_zone_t sleepq_zone; 152 153/* 154 * Prototypes for non-exported routines. 155 */ 156static int sleepq_catch_signals(void *wchan, int pri); 157static int sleepq_check_signals(void); 158static int sleepq_check_timeout(void); 159#ifdef INVARIANTS 160static void sleepq_dtor(void *mem, int size, void *arg); 161#endif 162static int sleepq_init(void *mem, int size, int flags); 163static void sleepq_resume_thread(struct sleepqueue *sq, struct thread *td, 164 int pri); 165static void sleepq_switch(void *wchan, int pri); 166static void sleepq_timeout(void *arg); 167 168/* 169 * Early initialization of sleep queues that is called from the sleepinit() 170 * SYSINIT. 171 */ 172void 173init_sleepqueues(void) 174{ 175#ifdef SLEEPQUEUE_PROFILING 176 struct sysctl_oid *chain_oid; 177 char chain_name[10]; 178#endif 179 int i; 180 181 for (i = 0; i < SC_TABLESIZE; i++) { 182 LIST_INIT(&sleepq_chains[i].sc_queues); 183 mtx_init(&sleepq_chains[i].sc_lock, "sleepq chain", NULL, 184 MTX_SPIN | MTX_RECURSE); 185#ifdef SLEEPQUEUE_PROFILING 186 snprintf(chain_name, sizeof(chain_name), "%d", i); 187 chain_oid = SYSCTL_ADD_NODE(NULL, 188 SYSCTL_STATIC_CHILDREN(_debug_sleepq_chains), OID_AUTO, 189 chain_name, CTLFLAG_RD, NULL, "sleepq chain stats"); 190 SYSCTL_ADD_UINT(NULL, SYSCTL_CHILDREN(chain_oid), OID_AUTO, 191 "depth", CTLFLAG_RD, &sleepq_chains[i].sc_depth, 0, NULL); 192 SYSCTL_ADD_UINT(NULL, SYSCTL_CHILDREN(chain_oid), OID_AUTO, 193 "max_depth", CTLFLAG_RD, &sleepq_chains[i].sc_max_depth, 0, 194 NULL); 195#endif 196 } 197 sleepq_zone = uma_zcreate("SLEEPQUEUE", sizeof(struct sleepqueue), 198#ifdef INVARIANTS 199 NULL, sleepq_dtor, sleepq_init, NULL, UMA_ALIGN_CACHE, 0); 200#else 201 NULL, NULL, sleepq_init, NULL, UMA_ALIGN_CACHE, 0); 202#endif 203 204 thread0.td_sleepqueue = sleepq_alloc(); 205} 206 207/* 208 * Get a sleep queue for a new thread. 209 */ 210struct sleepqueue * 211sleepq_alloc(void) 212{ 213 214 return (uma_zalloc(sleepq_zone, M_WAITOK)); 215} 216 217/* 218 * Free a sleep queue when a thread is destroyed. 219 */ 220void 221sleepq_free(struct sleepqueue *sq) 222{ 223 224 uma_zfree(sleepq_zone, sq); 225} 226 227/* 228 * Lock the sleep queue chain associated with the specified wait channel. 229 */ 230void 231sleepq_lock(void *wchan) 232{ 233 struct sleepqueue_chain *sc; 234 235 sc = SC_LOOKUP(wchan); 236 mtx_lock_spin(&sc->sc_lock); 237} 238 239/* 240 * Look up the sleep queue associated with a given wait channel in the hash 241 * table locking the associated sleep queue chain. If no queue is found in 242 * the table, NULL is returned. 243 */ 244struct sleepqueue * 245sleepq_lookup(void *wchan) 246{ 247 struct sleepqueue_chain *sc; 248 struct sleepqueue *sq; 249 250 KASSERT(wchan != NULL, ("%s: invalid NULL wait channel", __func__)); 251 sc = SC_LOOKUP(wchan); 252 mtx_assert(&sc->sc_lock, MA_OWNED); 253 LIST_FOREACH(sq, &sc->sc_queues, sq_hash) 254 if (sq->sq_wchan == wchan) 255 return (sq); 256 return (NULL); 257} 258 259/* 260 * Unlock the sleep queue chain associated with a given wait channel. 261 */ 262void 263sleepq_release(void *wchan) 264{ 265 struct sleepqueue_chain *sc; 266 267 sc = SC_LOOKUP(wchan); 268 mtx_unlock_spin(&sc->sc_lock); 269} 270 271/* 272 * Places the current thread on the sleep queue for the specified wait 273 * channel. If INVARIANTS is enabled, then it associates the passed in 274 * lock with the sleepq to make sure it is held when that sleep queue is 275 * woken up. 276 */ 277void 278sleepq_add(void *wchan, struct lock_object *lock, const char *wmesg, int flags, 279 int queue) 280{ 281 struct sleepqueue_chain *sc; 282 struct sleepqueue *sq; 283 struct thread *td; 284 285 td = curthread; 286 sc = SC_LOOKUP(wchan); 287 mtx_assert(&sc->sc_lock, MA_OWNED); 288 MPASS(td->td_sleepqueue != NULL); 289 MPASS(wchan != NULL); 290 MPASS((queue >= 0) && (queue < NR_SLEEPQS)); 291 292 /* If this thread is not allowed to sleep, die a horrible death. */ 293 KASSERT(!(td->td_pflags & TDP_NOSLEEPING), 294 ("Trying sleep, but thread marked as sleeping prohibited")); 295 296 /* Look up the sleep queue associated with the wait channel 'wchan'. */ 297 sq = sleepq_lookup(wchan); 298 299 /* 300 * If the wait channel does not already have a sleep queue, use 301 * this thread's sleep queue. Otherwise, insert the current thread 302 * into the sleep queue already in use by this wait channel. 303 */ 304 if (sq == NULL) { 305#ifdef INVARIANTS 306 int i; 307 308 sq = td->td_sleepqueue; 309 for (i = 0; i < NR_SLEEPQS; i++) 310 KASSERT(TAILQ_EMPTY(&sq->sq_blocked[i]), 311 ("thread's sleep queue %d is not empty", i)); 312 KASSERT(LIST_EMPTY(&sq->sq_free), 313 ("thread's sleep queue has a non-empty free list")); 314 KASSERT(sq->sq_wchan == NULL, ("stale sq_wchan pointer")); 315 sq->sq_lock = lock; 316 sq->sq_type = flags & SLEEPQ_TYPE; 317#endif 318#ifdef SLEEPQUEUE_PROFILING 319 sc->sc_depth++; 320 if (sc->sc_depth > sc->sc_max_depth) { 321 sc->sc_max_depth = sc->sc_depth; 322 if (sc->sc_max_depth > sleepq_max_depth) 323 sleepq_max_depth = sc->sc_max_depth; 324 } 325#endif 326 sq = td->td_sleepqueue; 327 LIST_INSERT_HEAD(&sc->sc_queues, sq, sq_hash); 328 sq->sq_wchan = wchan; 329 } else { 330 MPASS(wchan == sq->sq_wchan); 331 MPASS(lock == sq->sq_lock); 332 MPASS((flags & SLEEPQ_TYPE) == sq->sq_type); 333 LIST_INSERT_HEAD(&sq->sq_free, td->td_sleepqueue, sq_hash); 334 } 335 thread_lock(td); 336 TAILQ_INSERT_TAIL(&sq->sq_blocked[queue], td, td_slpq); 337 td->td_sleepqueue = NULL; 338 td->td_sqqueue = queue; 339 td->td_wchan = wchan; 340 td->td_wmesg = wmesg; 341 if (flags & SLEEPQ_INTERRUPTIBLE) { 342 td->td_flags |= TDF_SINTR; 343 td->td_flags &= ~TDF_SLEEPABORT; 344 } 345 thread_unlock(td); 346} 347 348/* 349 * Sets a timeout that will remove the current thread from the specified 350 * sleep queue after timo ticks if the thread has not already been awakened. 351 */ 352void 353sleepq_set_timeout(void *wchan, int timo) 354{ 355 struct sleepqueue_chain *sc; 356 struct thread *td; 357 358 td = curthread; 359 sc = SC_LOOKUP(wchan); 360 mtx_assert(&sc->sc_lock, MA_OWNED); 361 MPASS(TD_ON_SLEEPQ(td)); 362 MPASS(td->td_sleepqueue == NULL); 363 MPASS(wchan != NULL); 364 callout_reset(&td->td_slpcallout, timo, sleepq_timeout, td); 365} 366 367/* 368 * Marks the pending sleep of the current thread as interruptible and 369 * makes an initial check for pending signals before putting a thread 370 * to sleep. Enters and exits with the thread lock held. Thread lock 371 * may have transitioned from the sleepq lock to a run lock. 372 */ 373static int 374sleepq_catch_signals(void *wchan, int pri) 375{ 376 struct sleepqueue_chain *sc; 377 struct sleepqueue *sq; 378 struct thread *td; 379 struct proc *p; 380 struct sigacts *ps; 381 int sig, ret; 382 383 td = curthread; 384 p = curproc; 385 sc = SC_LOOKUP(wchan); 386 mtx_assert(&sc->sc_lock, MA_OWNED); 387 MPASS(wchan != NULL); 388 /* 389 * See if there are any pending signals for this thread. If not 390 * we can switch immediately. Otherwise do the signal processing 391 * directly. 392 */ 393 thread_lock(td); 394 if ((td->td_flags & (TDF_NEEDSIGCHK | TDF_NEEDSUSPCHK)) == 0) { 395 sleepq_switch(wchan, pri); 396 return (0); 397 } 398 thread_unlock(td); 399 mtx_unlock_spin(&sc->sc_lock); 400 CTR3(KTR_PROC, "sleepq catching signals: thread %p (pid %ld, %s)", 401 (void *)td, (long)p->p_pid, td->td_name); 402 PROC_LOCK(p); 403 ps = p->p_sigacts; 404 mtx_lock(&ps->ps_mtx); 405 sig = cursig(td); 406 if (sig == 0) { 407 mtx_unlock(&ps->ps_mtx); 408 ret = thread_suspend_check(1); 409 MPASS(ret == 0 || ret == EINTR || ret == ERESTART); 410 } else { 411 if (SIGISMEMBER(ps->ps_sigintr, sig)) 412 ret = EINTR; 413 else 414 ret = ERESTART; 415 mtx_unlock(&ps->ps_mtx); 416 } 417 /* 418 * Lock the per-process spinlock prior to dropping the PROC_LOCK 419 * to avoid a signal delivery race. PROC_LOCK, PROC_SLOCK, and 420 * thread_lock() are currently held in tdsignal(). 421 */ 422 PROC_SLOCK(p); 423 mtx_lock_spin(&sc->sc_lock); 424 PROC_UNLOCK(p); 425 thread_lock(td); 426 PROC_SUNLOCK(p); 427 if (ret == 0) { 428 sleepq_switch(wchan, pri); 429 return (0); 430 } 431 /* 432 * There were pending signals and this thread is still 433 * on the sleep queue, remove it from the sleep queue. 434 */ 435 if (TD_ON_SLEEPQ(td)) { 436 sq = sleepq_lookup(wchan); 437 sleepq_resume_thread(sq, td, 0); 438 } 439 mtx_unlock_spin(&sc->sc_lock); 440 MPASS(td->td_lock != &sc->sc_lock); 441 return (ret); 442} 443 444/* 445 * Switches to another thread if we are still asleep on a sleep queue. 446 * Returns with thread lock. 447 */ 448static void 449sleepq_switch(void *wchan, int pri) 450{ 451 struct sleepqueue_chain *sc; 452 struct sleepqueue *sq; 453 struct thread *td; 454 455 td = curthread; 456 sc = SC_LOOKUP(wchan); 457 mtx_assert(&sc->sc_lock, MA_OWNED); 458 THREAD_LOCK_ASSERT(td, MA_OWNED); 459 460 /* 461 * If we have a sleep queue, then we've already been woken up, so 462 * just return. 463 */ 464 if (td->td_sleepqueue != NULL) { 465 mtx_unlock_spin(&sc->sc_lock); 466 return; 467 } 468 469 /* 470 * If TDF_TIMEOUT is set, then our sleep has been timed out 471 * already but we are still on the sleep queue, so dequeue the 472 * thread and return. 473 */ 474 if (td->td_flags & TDF_TIMEOUT) { 475 MPASS(TD_ON_SLEEPQ(td)); 476 sq = sleepq_lookup(wchan); 477 sleepq_resume_thread(sq, td, 0); 478 mtx_unlock_spin(&sc->sc_lock); 479 return; 480 } 481#ifdef SLEEPQUEUE_PROFILING 482 if (prof_enabled) 483 sleepq_profile(td->td_wmesg); 484#endif 485 MPASS(td->td_sleepqueue == NULL); 486 sched_sleep(td, pri); 487 thread_lock_set(td, &sc->sc_lock); 488 TD_SET_SLEEPING(td); 489 SCHED_STAT_INC(switch_sleepq); 490 mi_switch(SW_VOL, NULL); 491 KASSERT(TD_IS_RUNNING(td), ("running but not TDS_RUNNING")); 492 CTR3(KTR_PROC, "sleepq resume: thread %p (pid %ld, %s)", 493 (void *)td, (long)td->td_proc->p_pid, (void *)td->td_name); 494} 495 496/* 497 * Check to see if we timed out. 498 */ 499static int 500sleepq_check_timeout(void) 501{ 502 struct thread *td; 503 504 td = curthread; 505 THREAD_LOCK_ASSERT(td, MA_OWNED); 506 507 /* 508 * If TDF_TIMEOUT is set, we timed out. 509 */ 510 if (td->td_flags & TDF_TIMEOUT) { 511 td->td_flags &= ~TDF_TIMEOUT; 512 return (EWOULDBLOCK); 513 } 514 515 /* 516 * If TDF_TIMOFAIL is set, the timeout ran after we had 517 * already been woken up. 518 */ 519 if (td->td_flags & TDF_TIMOFAIL) 520 td->td_flags &= ~TDF_TIMOFAIL; 521 522 /* 523 * If callout_stop() fails, then the timeout is running on 524 * another CPU, so synchronize with it to avoid having it 525 * accidentally wake up a subsequent sleep. 526 */ 527 else if (callout_stop(&td->td_slpcallout) == 0) { 528 td->td_flags |= TDF_TIMEOUT; 529 TD_SET_SLEEPING(td); 530 SCHED_STAT_INC(switch_sleepqtimo); 531 mi_switch(SW_INVOL, NULL); 532 } 533 return (0); 534} 535 536/* 537 * Check to see if we were awoken by a signal. 538 */ 539static int 540sleepq_check_signals(void) 541{ 542 struct thread *td; 543 544 td = curthread; 545 THREAD_LOCK_ASSERT(td, MA_OWNED); 546 547 /* We are no longer in an interruptible sleep. */ 548 if (td->td_flags & TDF_SINTR) 549 td->td_flags &= ~TDF_SINTR; 550 551 if (td->td_flags & TDF_SLEEPABORT) { 552 td->td_flags &= ~TDF_SLEEPABORT; 553 return (td->td_intrval); 554 } 555 556 return (0); 557} 558 559/* 560 * Block the current thread until it is awakened from its sleep queue. 561 */ 562void 563sleepq_wait(void *wchan, int pri) 564{ 565 struct thread *td; 566 567 td = curthread; 568 MPASS(!(td->td_flags & TDF_SINTR)); 569 thread_lock(td); 570 sleepq_switch(wchan, pri); 571 thread_unlock(td); 572} 573 574/* 575 * Block the current thread until it is awakened from its sleep queue 576 * or it is interrupted by a signal. 577 */ 578int 579sleepq_wait_sig(void *wchan, int pri) 580{ 581 int rcatch; 582 int rval; 583 584 rcatch = sleepq_catch_signals(wchan, pri); 585 rval = sleepq_check_signals(); 586 thread_unlock(curthread); 587 if (rcatch) 588 return (rcatch); 589 return (rval); 590} 591 592/* 593 * Block the current thread until it is awakened from its sleep queue 594 * or it times out while waiting. 595 */ 596int 597sleepq_timedwait(void *wchan, int pri) 598{ 599 struct thread *td; 600 int rval; 601 602 td = curthread; 603 MPASS(!(td->td_flags & TDF_SINTR)); 604 thread_lock(td); 605 sleepq_switch(wchan, pri); 606 rval = sleepq_check_timeout(); 607 thread_unlock(td); 608 609 return (rval); 610} 611 612/* 613 * Block the current thread until it is awakened from its sleep queue, 614 * it is interrupted by a signal, or it times out waiting to be awakened. 615 */ 616int 617sleepq_timedwait_sig(void *wchan, int pri) 618{ 619 int rcatch, rvalt, rvals; 620 621 rcatch = sleepq_catch_signals(wchan, pri); 622 rvalt = sleepq_check_timeout(); 623 rvals = sleepq_check_signals(); 624 thread_unlock(curthread); 625 if (rcatch) 626 return (rcatch); 627 if (rvals) 628 return (rvals); 629 return (rvalt); 630} 631 632/* 633 * Removes a thread from a sleep queue and makes it 634 * runnable. 635 */ 636static void 637sleepq_resume_thread(struct sleepqueue *sq, struct thread *td, int pri) 638{ 639 struct sleepqueue_chain *sc; 640 641 MPASS(td != NULL); 642 MPASS(sq->sq_wchan != NULL); 643 MPASS(td->td_wchan == sq->sq_wchan); 644 MPASS(td->td_sqqueue < NR_SLEEPQS && td->td_sqqueue >= 0); 645 THREAD_LOCK_ASSERT(td, MA_OWNED); 646 sc = SC_LOOKUP(sq->sq_wchan); 647 mtx_assert(&sc->sc_lock, MA_OWNED); 648 649 /* Remove the thread from the queue. */ 650 TAILQ_REMOVE(&sq->sq_blocked[td->td_sqqueue], td, td_slpq); 651 652 /* 653 * Get a sleep queue for this thread. If this is the last waiter, 654 * use the queue itself and take it out of the chain, otherwise, 655 * remove a queue from the free list. 656 */ 657 if (LIST_EMPTY(&sq->sq_free)) { 658 td->td_sleepqueue = sq; 659#ifdef INVARIANTS 660 sq->sq_wchan = NULL; 661#endif 662#ifdef SLEEPQUEUE_PROFILING 663 sc->sc_depth--; 664#endif 665 } else 666 td->td_sleepqueue = LIST_FIRST(&sq->sq_free); 667 LIST_REMOVE(td->td_sleepqueue, sq_hash); 668 669 td->td_wmesg = NULL; 670 td->td_wchan = NULL; 671 td->td_flags &= ~TDF_SINTR; 672 673 /* 674 * Note that thread td might not be sleeping if it is running 675 * sleepq_catch_signals() on another CPU or is blocked on 676 * its proc lock to check signals. It doesn't hurt to clear 677 * the sleeping flag if it isn't set though, so we just always 678 * do it. However, we can't assert that it is set. 679 */ 680 CTR3(KTR_PROC, "sleepq_wakeup: thread %p (pid %ld, %s)", 681 (void *)td, (long)td->td_proc->p_pid, td->td_name); 682 TD_CLR_SLEEPING(td); 683 684 /* Adjust priority if requested. */ 685 MPASS(pri == 0 || (pri >= PRI_MIN && pri <= PRI_MAX)); 686 if (pri != 0 && td->td_priority > pri) 687 sched_prio(td, pri); 688 setrunnable(td); 689} 690 691#ifdef INVARIANTS 692/* 693 * UMA zone item deallocator. 694 */ 695static void 696sleepq_dtor(void *mem, int size, void *arg) 697{ 698 struct sleepqueue *sq; 699 int i; 700 701 sq = mem; 702 for (i = 0; i < NR_SLEEPQS; i++) 703 MPASS(TAILQ_EMPTY(&sq->sq_blocked[i])); 704} 705#endif 706 707/* 708 * UMA zone item initializer. 709 */ 710static int 711sleepq_init(void *mem, int size, int flags) 712{ 713 struct sleepqueue *sq; 714 int i; 715 716 bzero(mem, size); 717 sq = mem; 718 for (i = 0; i < NR_SLEEPQS; i++) 719 TAILQ_INIT(&sq->sq_blocked[i]); 720 LIST_INIT(&sq->sq_free); 721 return (0); 722} 723 724/* 725 * Find the highest priority thread sleeping on a wait channel and resume it. 726 */ 727void 728sleepq_signal(void *wchan, int flags, int pri, int queue) 729{ 730 struct sleepqueue *sq; 731 struct thread *td, *besttd; 732 733 CTR2(KTR_PROC, "sleepq_signal(%p, %d)", wchan, flags); 734 KASSERT(wchan != NULL, ("%s: invalid NULL wait channel", __func__)); 735 MPASS((queue >= 0) && (queue < NR_SLEEPQS)); 736 sq = sleepq_lookup(wchan); 737 if (sq == NULL) 738 return; 739 KASSERT(sq->sq_type == (flags & SLEEPQ_TYPE), 740 ("%s: mismatch between sleep/wakeup and cv_*", __func__)); 741 742 /* 743 * Find the highest priority thread on the queue. If there is a 744 * tie, use the thread that first appears in the queue as it has 745 * been sleeping the longest since threads are always added to 746 * the tail of sleep queues. 747 */ 748 besttd = NULL; 749 TAILQ_FOREACH(td, &sq->sq_blocked[queue], td_slpq) { 750 if (besttd == NULL || td->td_priority < besttd->td_priority) 751 besttd = td; 752 } 753 MPASS(besttd != NULL); 754 thread_lock(besttd); 755 sleepq_resume_thread(sq, besttd, pri); 756 thread_unlock(besttd); 757} 758 759/* 760 * Resume all threads sleeping on a specified wait channel. 761 */ 762void 763sleepq_broadcast(void *wchan, int flags, int pri, int queue) 764{ 765 struct sleepqueue *sq; 766 struct thread *td; 767 768 CTR2(KTR_PROC, "sleepq_broadcast(%p, %d)", wchan, flags); 769 KASSERT(wchan != NULL, ("%s: invalid NULL wait channel", __func__)); 770 MPASS((queue >= 0) && (queue < NR_SLEEPQS)); 771 sq = sleepq_lookup(wchan); 772 if (sq == NULL) 773 return; 774 KASSERT(sq->sq_type == (flags & SLEEPQ_TYPE), 775 ("%s: mismatch between sleep/wakeup and cv_*", __func__)); 776 777 /* Resume all blocked threads on the sleep queue. */ 778 while (!TAILQ_EMPTY(&sq->sq_blocked[queue])) { 779 td = TAILQ_FIRST(&sq->sq_blocked[queue]); 780 thread_lock(td); 781 sleepq_resume_thread(sq, td, pri); 782 thread_unlock(td); 783 } 784} 785 786/* 787 * Time sleeping threads out. When the timeout expires, the thread is 788 * removed from the sleep queue and made runnable if it is still asleep. 789 */ 790static void 791sleepq_timeout(void *arg) 792{ 793 struct sleepqueue_chain *sc; 794 struct sleepqueue *sq; 795 struct thread *td; 796 void *wchan; 797 798 td = arg; 799 CTR3(KTR_PROC, "sleepq_timeout: thread %p (pid %ld, %s)", 800 (void *)td, (long)td->td_proc->p_pid, (void *)td->td_name); 801 802 /* 803 * First, see if the thread is asleep and get the wait channel if 804 * it is. 805 */ 806 thread_lock(td); 807 if (TD_IS_SLEEPING(td) && TD_ON_SLEEPQ(td)) { 808 wchan = td->td_wchan; 809 sc = SC_LOOKUP(wchan); 810 THREAD_LOCKPTR_ASSERT(td, &sc->sc_lock); 811 sq = sleepq_lookup(wchan); 812 MPASS(sq != NULL); 813 td->td_flags |= TDF_TIMEOUT; 814 sleepq_resume_thread(sq, td, 0); 815 thread_unlock(td); 816 return; 817 } 818 819 /* 820 * If the thread is on the SLEEPQ but isn't sleeping yet, it 821 * can either be on another CPU in between sleepq_add() and 822 * one of the sleepq_*wait*() routines or it can be in 823 * sleepq_catch_signals(). 824 */ 825 if (TD_ON_SLEEPQ(td)) { 826 td->td_flags |= TDF_TIMEOUT; 827 thread_unlock(td); 828 return; 829 } 830 831 /* 832 * Now check for the edge cases. First, if TDF_TIMEOUT is set, 833 * then the other thread has already yielded to us, so clear 834 * the flag and resume it. If TDF_TIMEOUT is not set, then the 835 * we know that the other thread is not on a sleep queue, but it 836 * hasn't resumed execution yet. In that case, set TDF_TIMOFAIL 837 * to let it know that the timeout has already run and doesn't 838 * need to be canceled. 839 */ 840 if (td->td_flags & TDF_TIMEOUT) { 841 MPASS(TD_IS_SLEEPING(td)); 842 td->td_flags &= ~TDF_TIMEOUT; 843 TD_CLR_SLEEPING(td); 844 setrunnable(td); 845 } else 846 td->td_flags |= TDF_TIMOFAIL; 847 thread_unlock(td); 848} 849 850/* 851 * Resumes a specific thread from the sleep queue associated with a specific 852 * wait channel if it is on that queue. 853 */ 854void 855sleepq_remove(struct thread *td, void *wchan) 856{ 857 struct sleepqueue *sq; 858 859 /* 860 * Look up the sleep queue for this wait channel, then re-check 861 * that the thread is asleep on that channel, if it is not, then 862 * bail. 863 */ 864 MPASS(wchan != NULL); 865 sleepq_lock(wchan); 866 sq = sleepq_lookup(wchan); 867 /* 868 * We can not lock the thread here as it may be sleeping on a 869 * different sleepq. However, holding the sleepq lock for this 870 * wchan can guarantee that we do not miss a wakeup for this 871 * channel. The asserts below will catch any false positives. 872 */ 873 if (!TD_ON_SLEEPQ(td) || td->td_wchan != wchan) { 874 sleepq_release(wchan); 875 return; 876 } 877 /* Thread is asleep on sleep queue sq, so wake it up. */ 878 thread_lock(td); 879 MPASS(sq != NULL); 880 MPASS(td->td_wchan == wchan); 881 sleepq_resume_thread(sq, td, 0); 882 thread_unlock(td); 883 sleepq_release(wchan); 884} 885 886/* 887 * Abort a thread as if an interrupt had occurred. Only abort 888 * interruptible waits (unfortunately it isn't safe to abort others). 889 */ 890void 891sleepq_abort(struct thread *td, int intrval) 892{ 893 struct sleepqueue *sq; 894 void *wchan; 895 896 THREAD_LOCK_ASSERT(td, MA_OWNED); 897 MPASS(TD_ON_SLEEPQ(td)); 898 MPASS(td->td_flags & TDF_SINTR); 899 MPASS(intrval == EINTR || intrval == ERESTART); 900 901 /* 902 * If the TDF_TIMEOUT flag is set, just leave. A 903 * timeout is scheduled anyhow. 904 */ 905 if (td->td_flags & TDF_TIMEOUT) 906 return; 907 908 CTR3(KTR_PROC, "sleepq_abort: thread %p (pid %ld, %s)", 909 (void *)td, (long)td->td_proc->p_pid, (void *)td->td_name); 910 td->td_intrval = intrval; 911 td->td_flags |= TDF_SLEEPABORT; 912 /* 913 * If the thread has not slept yet it will find the signal in 914 * sleepq_catch_signals() and call sleepq_resume_thread. Otherwise 915 * we have to do it here. 916 */ 917 if (!TD_IS_SLEEPING(td)) 918 return; 919 wchan = td->td_wchan; 920 MPASS(wchan != NULL); 921 sq = sleepq_lookup(wchan); 922 MPASS(sq != NULL); 923 924 /* Thread is asleep on sleep queue sq, so wake it up. */ 925 sleepq_resume_thread(sq, td, 0); 926} 927 928#ifdef SLEEPQUEUE_PROFILING 929#define SLEEPQ_PROF_LOCATIONS 1024 930#define SLEEPQ_SBUFSIZE (40 * 512) 931struct sleepq_prof { 932 LIST_ENTRY(sleepq_prof) sp_link; 933 const char *sp_wmesg; 934 long sp_count; 935}; 936 937LIST_HEAD(sqphead, sleepq_prof); 938 939struct sqphead sleepq_prof_free; 940struct sqphead sleepq_hash[SC_TABLESIZE]; 941static struct sleepq_prof sleepq_profent[SLEEPQ_PROF_LOCATIONS]; 942static struct mtx sleepq_prof_lock; 943MTX_SYSINIT(sleepq_prof_lock, &sleepq_prof_lock, "sleepq_prof", MTX_SPIN); 944 945static void 946sleepq_profile(const char *wmesg) 947{ 948 struct sleepq_prof *sp; 949 950 mtx_lock_spin(&sleepq_prof_lock); 951 if (prof_enabled == 0) 952 goto unlock; 953 LIST_FOREACH(sp, &sleepq_hash[SC_HASH(wmesg)], sp_link) 954 if (sp->sp_wmesg == wmesg) 955 goto done; 956 sp = LIST_FIRST(&sleepq_prof_free); 957 if (sp == NULL) 958 goto unlock; 959 sp->sp_wmesg = wmesg; 960 LIST_REMOVE(sp, sp_link); 961 LIST_INSERT_HEAD(&sleepq_hash[SC_HASH(wmesg)], sp, sp_link); 962done: 963 sp->sp_count++; 964unlock: 965 mtx_unlock_spin(&sleepq_prof_lock); 966 return; 967} 968 969static void 970sleepq_prof_reset(void) 971{ 972 struct sleepq_prof *sp; 973 int enabled; 974 int i; 975 976 mtx_lock_spin(&sleepq_prof_lock); 977 enabled = prof_enabled; 978 prof_enabled = 0; 979 for (i = 0; i < SC_TABLESIZE; i++) 980 LIST_INIT(&sleepq_hash[i]); 981 LIST_INIT(&sleepq_prof_free); 982 for (i = 0; i < SLEEPQ_PROF_LOCATIONS; i++) { 983 sp = &sleepq_profent[i]; 984 sp->sp_wmesg = NULL; 985 sp->sp_count = 0; 986 LIST_INSERT_HEAD(&sleepq_prof_free, sp, sp_link); 987 } 988 prof_enabled = enabled; 989 mtx_unlock_spin(&sleepq_prof_lock); 990} 991 992static int 993enable_sleepq_prof(SYSCTL_HANDLER_ARGS) 994{ 995 int error, v; 996 997 v = prof_enabled; 998 error = sysctl_handle_int(oidp, &v, v, req); 999 if (error) 1000 return (error); 1001 if (req->newptr == NULL) 1002 return (error); 1003 if (v == prof_enabled) 1004 return (0); 1005 if (v == 1) 1006 sleepq_prof_reset(); 1007 mtx_lock_spin(&sleepq_prof_lock); 1008 prof_enabled = !!v; 1009 mtx_unlock_spin(&sleepq_prof_lock); 1010 1011 return (0); 1012} 1013 1014static int 1015reset_sleepq_prof_stats(SYSCTL_HANDLER_ARGS) 1016{ 1017 int error, v; 1018 1019 v = 0; 1020 error = sysctl_handle_int(oidp, &v, 0, req); 1021 if (error) 1022 return (error); 1023 if (req->newptr == NULL) 1024 return (error); 1025 if (v == 0) 1026 return (0); 1027 sleepq_prof_reset(); 1028 1029 return (0); 1030} 1031 1032static int 1033dump_sleepq_prof_stats(SYSCTL_HANDLER_ARGS) 1034{ 1035 static int multiplier = 1; 1036 struct sleepq_prof *sp; 1037 struct sbuf *sb; 1038 int enabled; 1039 int error; 1040 int i; 1041 1042retry_sbufops: 1043 sb = sbuf_new(NULL, NULL, SLEEPQ_SBUFSIZE * multiplier, SBUF_FIXEDLEN); 1044 sbuf_printf(sb, "\nwmesg\tcount\n"); 1045 enabled = prof_enabled; 1046 mtx_lock_spin(&sleepq_prof_lock); 1047 prof_enabled = 0; 1048 mtx_unlock_spin(&sleepq_prof_lock); 1049 for (i = 0; i < SC_TABLESIZE; i++) { 1050 LIST_FOREACH(sp, &sleepq_hash[i], sp_link) { 1051 sbuf_printf(sb, "%s\t%ld\n", 1052 sp->sp_wmesg, sp->sp_count); 1053 if (sbuf_overflowed(sb)) { 1054 sbuf_delete(sb); 1055 multiplier++; 1056 goto retry_sbufops; 1057 } 1058 } 1059 } 1060 mtx_lock_spin(&sleepq_prof_lock); 1061 prof_enabled = enabled; 1062 mtx_unlock_spin(&sleepq_prof_lock); 1063 1064 sbuf_finish(sb); 1065 error = SYSCTL_OUT(req, sbuf_data(sb), sbuf_len(sb) + 1); 1066 sbuf_delete(sb); 1067 return (error); 1068} 1069 1070SYSCTL_PROC(_debug_sleepq, OID_AUTO, stats, CTLTYPE_STRING | CTLFLAG_RD, 1071 NULL, 0, dump_sleepq_prof_stats, "A", "Sleepqueue profiling statistics"); 1072SYSCTL_PROC(_debug_sleepq, OID_AUTO, reset, CTLTYPE_INT | CTLFLAG_RW, 1073 NULL, 0, reset_sleepq_prof_stats, "I", 1074 "Reset sleepqueue profiling statistics"); 1075SYSCTL_PROC(_debug_sleepq, OID_AUTO, enable, CTLTYPE_INT | CTLFLAG_RW, 1076 NULL, 0, enable_sleepq_prof, "I", "Enable sleepqueue profiling"); 1077#endif 1078 1079#ifdef DDB 1080DB_SHOW_COMMAND(sleepq, db_show_sleepqueue) 1081{ 1082 struct sleepqueue_chain *sc; 1083 struct sleepqueue *sq; 1084#ifdef INVARIANTS 1085 struct lock_object *lock; 1086#endif 1087 struct thread *td; 1088 void *wchan; 1089 int i; 1090 1091 if (!have_addr) 1092 return; 1093 1094 /* 1095 * First, see if there is an active sleep queue for the wait channel 1096 * indicated by the address. 1097 */ 1098 wchan = (void *)addr; 1099 sc = SC_LOOKUP(wchan); 1100 LIST_FOREACH(sq, &sc->sc_queues, sq_hash) 1101 if (sq->sq_wchan == wchan) 1102 goto found; 1103 1104 /* 1105 * Second, see if there is an active sleep queue at the address 1106 * indicated. 1107 */ 1108 for (i = 0; i < SC_TABLESIZE; i++) 1109 LIST_FOREACH(sq, &sleepq_chains[i].sc_queues, sq_hash) { 1110 if (sq == (struct sleepqueue *)addr) 1111 goto found; 1112 } 1113 1114 db_printf("Unable to locate a sleep queue via %p\n", (void *)addr); 1115 return; 1116found: 1117 db_printf("Wait channel: %p\n", sq->sq_wchan); 1118#ifdef INVARIANTS 1119 db_printf("Queue type: %d\n", sq->sq_type); 1120 if (sq->sq_lock) { 1121 lock = sq->sq_lock; 1122 db_printf("Associated Interlock: %p - (%s) %s\n", lock, 1123 LOCK_CLASS(lock)->lc_name, lock->lo_name); 1124 } 1125#endif 1126 db_printf("Blocked threads:\n"); 1127 for (i = 0; i < NR_SLEEPQS; i++) { 1128 db_printf("\nQueue[%d]:\n", i); 1129 if (TAILQ_EMPTY(&sq->sq_blocked[i])) 1130 db_printf("\tempty\n"); 1131 else 1132 TAILQ_FOREACH(td, &sq->sq_blocked[0], 1133 td_slpq) { 1134 db_printf("\t%p (tid %d, pid %d, \"%s\")\n", td, 1135 td->td_tid, td->td_proc->p_pid, 1136 td->td_name[i] != '\0' ? td->td_name : 1137 td->td_name); 1138 } 1139 } 1140} 1141 1142/* Alias 'show sleepqueue' to 'show sleepq'. */ 1143DB_SET(sleepqueue, db_show_sleepqueue, db_show_cmd_set, 0, NULL); 1144#endif 1145