subr_sleepqueue.c revision 184216
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 184216 2008-10-24 01:03:31Z davidxu $"); 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 int 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_curcpu(&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 } 398 399 thread_unlock(td); 400 mtx_unlock_spin(&sc->sc_lock); 401 CTR3(KTR_PROC, "sleepq catching signals: thread %p (pid %ld, %s)", 402 (void *)td, (long)p->p_pid, td->td_name); 403 PROC_LOCK(p); 404 ps = p->p_sigacts; 405 mtx_lock(&ps->ps_mtx); 406 sig = cursig(td); 407 if (sig == 0) { 408 mtx_unlock(&ps->ps_mtx); 409 ret = thread_suspend_check(1); 410 MPASS(ret == 0 || ret == EINTR || ret == ERESTART); 411 } else { 412 if (SIGISMEMBER(ps->ps_sigintr, sig)) 413 ret = EINTR; 414 else 415 ret = ERESTART; 416 mtx_unlock(&ps->ps_mtx); 417 } 418 419 mtx_lock_spin(&sc->sc_lock); 420 thread_lock(td); 421 PROC_UNLOCK(p); 422 if (ret == 0) { 423 sleepq_switch(wchan, pri); 424 return (0); 425 } 426 427 /* 428 * There were pending signals and this thread is still 429 * on the sleep queue, remove it from the sleep queue. 430 */ 431 if (TD_ON_SLEEPQ(td)) { 432 sq = sleepq_lookup(wchan); 433 if (sleepq_resume_thread(sq, td, 0)) { 434#ifdef INVARIANTS 435 /* 436 * This thread hasn't gone to sleep yet, so it 437 * should not be swapped out. 438 */ 439 panic("not waking up swapper"); 440#endif 441 } 442 } 443 mtx_unlock_spin(&sc->sc_lock); 444 MPASS(td->td_lock != &sc->sc_lock); 445 return (ret); 446} 447 448/* 449 * Switches to another thread if we are still asleep on a sleep queue. 450 * Returns with thread lock. 451 */ 452static void 453sleepq_switch(void *wchan, int pri) 454{ 455 struct sleepqueue_chain *sc; 456 struct sleepqueue *sq; 457 struct thread *td; 458 459 td = curthread; 460 sc = SC_LOOKUP(wchan); 461 mtx_assert(&sc->sc_lock, MA_OWNED); 462 THREAD_LOCK_ASSERT(td, MA_OWNED); 463 464 /* 465 * If we have a sleep queue, then we've already been woken up, so 466 * just return. 467 */ 468 if (td->td_sleepqueue != NULL) { 469 mtx_unlock_spin(&sc->sc_lock); 470 return; 471 } 472 473 /* 474 * If TDF_TIMEOUT is set, then our sleep has been timed out 475 * already but we are still on the sleep queue, so dequeue the 476 * thread and return. 477 */ 478 if (td->td_flags & TDF_TIMEOUT) { 479 MPASS(TD_ON_SLEEPQ(td)); 480 sq = sleepq_lookup(wchan); 481 if (sleepq_resume_thread(sq, td, 0)) { 482#ifdef INVARIANTS 483 /* 484 * This thread hasn't gone to sleep yet, so it 485 * should not be swapped out. 486 */ 487 panic("not waking up swapper"); 488#endif 489 } 490 mtx_unlock_spin(&sc->sc_lock); 491 return; 492 } 493#ifdef SLEEPQUEUE_PROFILING 494 if (prof_enabled) 495 sleepq_profile(td->td_wmesg); 496#endif 497 MPASS(td->td_sleepqueue == NULL); 498 sched_sleep(td, pri); 499 thread_lock_set(td, &sc->sc_lock); 500 TD_SET_SLEEPING(td); 501 mi_switch(SW_VOL | SWT_SLEEPQ, NULL); 502 KASSERT(TD_IS_RUNNING(td), ("running but not TDS_RUNNING")); 503 CTR3(KTR_PROC, "sleepq resume: thread %p (pid %ld, %s)", 504 (void *)td, (long)td->td_proc->p_pid, (void *)td->td_name); 505} 506 507/* 508 * Check to see if we timed out. 509 */ 510static int 511sleepq_check_timeout(void) 512{ 513 struct thread *td; 514 515 td = curthread; 516 THREAD_LOCK_ASSERT(td, MA_OWNED); 517 518 /* 519 * If TDF_TIMEOUT is set, we timed out. 520 */ 521 if (td->td_flags & TDF_TIMEOUT) { 522 td->td_flags &= ~TDF_TIMEOUT; 523 return (EWOULDBLOCK); 524 } 525 526 /* 527 * If TDF_TIMOFAIL is set, the timeout ran after we had 528 * already been woken up. 529 */ 530 if (td->td_flags & TDF_TIMOFAIL) 531 td->td_flags &= ~TDF_TIMOFAIL; 532 533 /* 534 * If callout_stop() fails, then the timeout is running on 535 * another CPU, so synchronize with it to avoid having it 536 * accidentally wake up a subsequent sleep. 537 */ 538 else if (callout_stop(&td->td_slpcallout) == 0) { 539 td->td_flags |= TDF_TIMEOUT; 540 TD_SET_SLEEPING(td); 541 mi_switch(SW_INVOL | SWT_SLEEPQTIMO, NULL); 542 } 543 return (0); 544} 545 546/* 547 * Check to see if we were awoken by a signal. 548 */ 549static int 550sleepq_check_signals(void) 551{ 552 struct thread *td; 553 554 td = curthread; 555 THREAD_LOCK_ASSERT(td, MA_OWNED); 556 557 /* We are no longer in an interruptible sleep. */ 558 if (td->td_flags & TDF_SINTR) 559 td->td_flags &= ~TDF_SINTR; 560 561 if (td->td_flags & TDF_SLEEPABORT) { 562 td->td_flags &= ~TDF_SLEEPABORT; 563 return (td->td_intrval); 564 } 565 566 return (0); 567} 568 569/* 570 * Block the current thread until it is awakened from its sleep queue. 571 */ 572void 573sleepq_wait(void *wchan, int pri) 574{ 575 struct thread *td; 576 577 td = curthread; 578 MPASS(!(td->td_flags & TDF_SINTR)); 579 thread_lock(td); 580 sleepq_switch(wchan, pri); 581 thread_unlock(td); 582} 583 584/* 585 * Block the current thread until it is awakened from its sleep queue 586 * or it is interrupted by a signal. 587 */ 588int 589sleepq_wait_sig(void *wchan, int pri) 590{ 591 int rcatch; 592 int rval; 593 594 rcatch = sleepq_catch_signals(wchan, pri); 595 rval = sleepq_check_signals(); 596 thread_unlock(curthread); 597 if (rcatch) 598 return (rcatch); 599 return (rval); 600} 601 602/* 603 * Block the current thread until it is awakened from its sleep queue 604 * or it times out while waiting. 605 */ 606int 607sleepq_timedwait(void *wchan, int pri) 608{ 609 struct thread *td; 610 int rval; 611 612 td = curthread; 613 MPASS(!(td->td_flags & TDF_SINTR)); 614 thread_lock(td); 615 sleepq_switch(wchan, pri); 616 rval = sleepq_check_timeout(); 617 thread_unlock(td); 618 619 return (rval); 620} 621 622/* 623 * Block the current thread until it is awakened from its sleep queue, 624 * it is interrupted by a signal, or it times out waiting to be awakened. 625 */ 626int 627sleepq_timedwait_sig(void *wchan, int pri) 628{ 629 int rcatch, rvalt, rvals; 630 631 rcatch = sleepq_catch_signals(wchan, pri); 632 rvalt = sleepq_check_timeout(); 633 rvals = sleepq_check_signals(); 634 thread_unlock(curthread); 635 if (rcatch) 636 return (rcatch); 637 if (rvals) 638 return (rvals); 639 return (rvalt); 640} 641 642/* 643 * Removes a thread from a sleep queue and makes it 644 * runnable. 645 */ 646static int 647sleepq_resume_thread(struct sleepqueue *sq, struct thread *td, int pri) 648{ 649 struct sleepqueue_chain *sc; 650 651 MPASS(td != NULL); 652 MPASS(sq->sq_wchan != NULL); 653 MPASS(td->td_wchan == sq->sq_wchan); 654 MPASS(td->td_sqqueue < NR_SLEEPQS && td->td_sqqueue >= 0); 655 THREAD_LOCK_ASSERT(td, MA_OWNED); 656 sc = SC_LOOKUP(sq->sq_wchan); 657 mtx_assert(&sc->sc_lock, MA_OWNED); 658 659 /* Remove the thread from the queue. */ 660 TAILQ_REMOVE(&sq->sq_blocked[td->td_sqqueue], td, td_slpq); 661 662 /* 663 * Get a sleep queue for this thread. If this is the last waiter, 664 * use the queue itself and take it out of the chain, otherwise, 665 * remove a queue from the free list. 666 */ 667 if (LIST_EMPTY(&sq->sq_free)) { 668 td->td_sleepqueue = sq; 669#ifdef INVARIANTS 670 sq->sq_wchan = NULL; 671#endif 672#ifdef SLEEPQUEUE_PROFILING 673 sc->sc_depth--; 674#endif 675 } else 676 td->td_sleepqueue = LIST_FIRST(&sq->sq_free); 677 LIST_REMOVE(td->td_sleepqueue, sq_hash); 678 679 td->td_wmesg = NULL; 680 td->td_wchan = NULL; 681 td->td_flags &= ~TDF_SINTR; 682 683 /* 684 * Note that thread td might not be sleeping if it is running 685 * sleepq_catch_signals() on another CPU or is blocked on 686 * its proc lock to check signals. It doesn't hurt to clear 687 * the sleeping flag if it isn't set though, so we just always 688 * do it. However, we can't assert that it is set. 689 */ 690 CTR3(KTR_PROC, "sleepq_wakeup: thread %p (pid %ld, %s)", 691 (void *)td, (long)td->td_proc->p_pid, td->td_name); 692 TD_CLR_SLEEPING(td); 693 694 /* Adjust priority if requested. */ 695 MPASS(pri == 0 || (pri >= PRI_MIN && pri <= PRI_MAX)); 696 if (pri != 0 && td->td_priority > pri) 697 sched_prio(td, pri); 698 return (setrunnable(td)); 699} 700 701#ifdef INVARIANTS 702/* 703 * UMA zone item deallocator. 704 */ 705static void 706sleepq_dtor(void *mem, int size, void *arg) 707{ 708 struct sleepqueue *sq; 709 int i; 710 711 sq = mem; 712 for (i = 0; i < NR_SLEEPQS; i++) 713 MPASS(TAILQ_EMPTY(&sq->sq_blocked[i])); 714} 715#endif 716 717/* 718 * UMA zone item initializer. 719 */ 720static int 721sleepq_init(void *mem, int size, int flags) 722{ 723 struct sleepqueue *sq; 724 int i; 725 726 bzero(mem, size); 727 sq = mem; 728 for (i = 0; i < NR_SLEEPQS; i++) 729 TAILQ_INIT(&sq->sq_blocked[i]); 730 LIST_INIT(&sq->sq_free); 731 return (0); 732} 733 734/* 735 * Find the highest priority thread sleeping on a wait channel and resume it. 736 */ 737int 738sleepq_signal(void *wchan, int flags, int pri, int queue) 739{ 740 struct sleepqueue *sq; 741 struct thread *td, *besttd; 742 int wakeup_swapper; 743 744 CTR2(KTR_PROC, "sleepq_signal(%p, %d)", wchan, flags); 745 KASSERT(wchan != NULL, ("%s: invalid NULL wait channel", __func__)); 746 MPASS((queue >= 0) && (queue < NR_SLEEPQS)); 747 sq = sleepq_lookup(wchan); 748 if (sq == NULL) 749 return (0); 750 KASSERT(sq->sq_type == (flags & SLEEPQ_TYPE), 751 ("%s: mismatch between sleep/wakeup and cv_*", __func__)); 752 753 /* 754 * Find the highest priority thread on the queue. If there is a 755 * tie, use the thread that first appears in the queue as it has 756 * been sleeping the longest since threads are always added to 757 * the tail of sleep queues. 758 */ 759 besttd = NULL; 760 TAILQ_FOREACH(td, &sq->sq_blocked[queue], td_slpq) { 761 if (besttd == NULL || td->td_priority < besttd->td_priority) 762 besttd = td; 763 } 764 MPASS(besttd != NULL); 765 thread_lock(besttd); 766 wakeup_swapper = sleepq_resume_thread(sq, besttd, pri); 767 thread_unlock(besttd); 768 return (wakeup_swapper); 769} 770 771/* 772 * Resume all threads sleeping on a specified wait channel. 773 */ 774int 775sleepq_broadcast(void *wchan, int flags, int pri, int queue) 776{ 777 struct sleepqueue *sq; 778 struct thread *td, *tdn; 779 int wakeup_swapper; 780 781 CTR2(KTR_PROC, "sleepq_broadcast(%p, %d)", wchan, flags); 782 KASSERT(wchan != NULL, ("%s: invalid NULL wait channel", __func__)); 783 MPASS((queue >= 0) && (queue < NR_SLEEPQS)); 784 sq = sleepq_lookup(wchan); 785 if (sq == NULL) 786 return (0); 787 KASSERT(sq->sq_type == (flags & SLEEPQ_TYPE), 788 ("%s: mismatch between sleep/wakeup and cv_*", __func__)); 789 790 /* Resume all blocked threads on the sleep queue. */ 791 wakeup_swapper = 0; 792 TAILQ_FOREACH_SAFE(td, &sq->sq_blocked[queue], td_slpq, tdn) { 793 thread_lock(td); 794 if (sleepq_resume_thread(sq, td, pri)) 795 wakeup_swapper = 1; 796 thread_unlock(td); 797 } 798 return (wakeup_swapper); 799} 800 801/* 802 * Time sleeping threads out. When the timeout expires, the thread is 803 * removed from the sleep queue and made runnable if it is still asleep. 804 */ 805static void 806sleepq_timeout(void *arg) 807{ 808 struct sleepqueue_chain *sc; 809 struct sleepqueue *sq; 810 struct thread *td; 811 void *wchan; 812 int wakeup_swapper; 813 814 td = arg; 815 wakeup_swapper = 0; 816 CTR3(KTR_PROC, "sleepq_timeout: thread %p (pid %ld, %s)", 817 (void *)td, (long)td->td_proc->p_pid, (void *)td->td_name); 818 819 /* 820 * First, see if the thread is asleep and get the wait channel if 821 * it is. 822 */ 823 thread_lock(td); 824 if (TD_IS_SLEEPING(td) && TD_ON_SLEEPQ(td)) { 825 wchan = td->td_wchan; 826 sc = SC_LOOKUP(wchan); 827 THREAD_LOCKPTR_ASSERT(td, &sc->sc_lock); 828 sq = sleepq_lookup(wchan); 829 MPASS(sq != NULL); 830 td->td_flags |= TDF_TIMEOUT; 831 wakeup_swapper = sleepq_resume_thread(sq, td, 0); 832 thread_unlock(td); 833 if (wakeup_swapper) 834 kick_proc0(); 835 return; 836 } 837 838 /* 839 * If the thread is on the SLEEPQ but isn't sleeping yet, it 840 * can either be on another CPU in between sleepq_add() and 841 * one of the sleepq_*wait*() routines or it can be in 842 * sleepq_catch_signals(). 843 */ 844 if (TD_ON_SLEEPQ(td)) { 845 td->td_flags |= TDF_TIMEOUT; 846 thread_unlock(td); 847 return; 848 } 849 850 /* 851 * Now check for the edge cases. First, if TDF_TIMEOUT is set, 852 * then the other thread has already yielded to us, so clear 853 * the flag and resume it. If TDF_TIMEOUT is not set, then the 854 * we know that the other thread is not on a sleep queue, but it 855 * hasn't resumed execution yet. In that case, set TDF_TIMOFAIL 856 * to let it know that the timeout has already run and doesn't 857 * need to be canceled. 858 */ 859 if (td->td_flags & TDF_TIMEOUT) { 860 MPASS(TD_IS_SLEEPING(td)); 861 td->td_flags &= ~TDF_TIMEOUT; 862 TD_CLR_SLEEPING(td); 863 wakeup_swapper = setrunnable(td); 864 } else 865 td->td_flags |= TDF_TIMOFAIL; 866 thread_unlock(td); 867 if (wakeup_swapper) 868 kick_proc0(); 869} 870 871/* 872 * Resumes a specific thread from the sleep queue associated with a specific 873 * wait channel if it is on that queue. 874 */ 875void 876sleepq_remove(struct thread *td, void *wchan) 877{ 878 struct sleepqueue *sq; 879 int wakeup_swapper; 880 881 /* 882 * Look up the sleep queue for this wait channel, then re-check 883 * that the thread is asleep on that channel, if it is not, then 884 * bail. 885 */ 886 MPASS(wchan != NULL); 887 sleepq_lock(wchan); 888 sq = sleepq_lookup(wchan); 889 /* 890 * We can not lock the thread here as it may be sleeping on a 891 * different sleepq. However, holding the sleepq lock for this 892 * wchan can guarantee that we do not miss a wakeup for this 893 * channel. The asserts below will catch any false positives. 894 */ 895 if (!TD_ON_SLEEPQ(td) || td->td_wchan != wchan) { 896 sleepq_release(wchan); 897 return; 898 } 899 /* Thread is asleep on sleep queue sq, so wake it up. */ 900 thread_lock(td); 901 MPASS(sq != NULL); 902 MPASS(td->td_wchan == wchan); 903 wakeup_swapper = sleepq_resume_thread(sq, td, 0); 904 thread_unlock(td); 905 sleepq_release(wchan); 906 if (wakeup_swapper) 907 kick_proc0(); 908} 909 910/* 911 * Abort a thread as if an interrupt had occurred. Only abort 912 * interruptible waits (unfortunately it isn't safe to abort others). 913 */ 914int 915sleepq_abort(struct thread *td, int intrval) 916{ 917 struct sleepqueue *sq; 918 void *wchan; 919 920 THREAD_LOCK_ASSERT(td, MA_OWNED); 921 MPASS(TD_ON_SLEEPQ(td)); 922 MPASS(td->td_flags & TDF_SINTR); 923 MPASS(intrval == EINTR || intrval == ERESTART); 924 925 /* 926 * If the TDF_TIMEOUT flag is set, just leave. A 927 * timeout is scheduled anyhow. 928 */ 929 if (td->td_flags & TDF_TIMEOUT) 930 return (0); 931 932 CTR3(KTR_PROC, "sleepq_abort: thread %p (pid %ld, %s)", 933 (void *)td, (long)td->td_proc->p_pid, (void *)td->td_name); 934 td->td_intrval = intrval; 935 td->td_flags |= TDF_SLEEPABORT; 936 /* 937 * If the thread has not slept yet it will find the signal in 938 * sleepq_catch_signals() and call sleepq_resume_thread. Otherwise 939 * we have to do it here. 940 */ 941 if (!TD_IS_SLEEPING(td)) 942 return (0); 943 wchan = td->td_wchan; 944 MPASS(wchan != NULL); 945 sq = sleepq_lookup(wchan); 946 MPASS(sq != NULL); 947 948 /* Thread is asleep on sleep queue sq, so wake it up. */ 949 return (sleepq_resume_thread(sq, td, 0)); 950} 951 952#ifdef SLEEPQUEUE_PROFILING 953#define SLEEPQ_PROF_LOCATIONS 1024 954#define SLEEPQ_SBUFSIZE (40 * 512) 955struct sleepq_prof { 956 LIST_ENTRY(sleepq_prof) sp_link; 957 const char *sp_wmesg; 958 long sp_count; 959}; 960 961LIST_HEAD(sqphead, sleepq_prof); 962 963struct sqphead sleepq_prof_free; 964struct sqphead sleepq_hash[SC_TABLESIZE]; 965static struct sleepq_prof sleepq_profent[SLEEPQ_PROF_LOCATIONS]; 966static struct mtx sleepq_prof_lock; 967MTX_SYSINIT(sleepq_prof_lock, &sleepq_prof_lock, "sleepq_prof", MTX_SPIN); 968 969static void 970sleepq_profile(const char *wmesg) 971{ 972 struct sleepq_prof *sp; 973 974 mtx_lock_spin(&sleepq_prof_lock); 975 if (prof_enabled == 0) 976 goto unlock; 977 LIST_FOREACH(sp, &sleepq_hash[SC_HASH(wmesg)], sp_link) 978 if (sp->sp_wmesg == wmesg) 979 goto done; 980 sp = LIST_FIRST(&sleepq_prof_free); 981 if (sp == NULL) 982 goto unlock; 983 sp->sp_wmesg = wmesg; 984 LIST_REMOVE(sp, sp_link); 985 LIST_INSERT_HEAD(&sleepq_hash[SC_HASH(wmesg)], sp, sp_link); 986done: 987 sp->sp_count++; 988unlock: 989 mtx_unlock_spin(&sleepq_prof_lock); 990 return; 991} 992 993static void 994sleepq_prof_reset(void) 995{ 996 struct sleepq_prof *sp; 997 int enabled; 998 int i; 999 1000 mtx_lock_spin(&sleepq_prof_lock); 1001 enabled = prof_enabled; 1002 prof_enabled = 0; 1003 for (i = 0; i < SC_TABLESIZE; i++) 1004 LIST_INIT(&sleepq_hash[i]); 1005 LIST_INIT(&sleepq_prof_free); 1006 for (i = 0; i < SLEEPQ_PROF_LOCATIONS; i++) { 1007 sp = &sleepq_profent[i]; 1008 sp->sp_wmesg = NULL; 1009 sp->sp_count = 0; 1010 LIST_INSERT_HEAD(&sleepq_prof_free, sp, sp_link); 1011 } 1012 prof_enabled = enabled; 1013 mtx_unlock_spin(&sleepq_prof_lock); 1014} 1015 1016static int 1017enable_sleepq_prof(SYSCTL_HANDLER_ARGS) 1018{ 1019 int error, v; 1020 1021 v = prof_enabled; 1022 error = sysctl_handle_int(oidp, &v, v, req); 1023 if (error) 1024 return (error); 1025 if (req->newptr == NULL) 1026 return (error); 1027 if (v == prof_enabled) 1028 return (0); 1029 if (v == 1) 1030 sleepq_prof_reset(); 1031 mtx_lock_spin(&sleepq_prof_lock); 1032 prof_enabled = !!v; 1033 mtx_unlock_spin(&sleepq_prof_lock); 1034 1035 return (0); 1036} 1037 1038static int 1039reset_sleepq_prof_stats(SYSCTL_HANDLER_ARGS) 1040{ 1041 int error, v; 1042 1043 v = 0; 1044 error = sysctl_handle_int(oidp, &v, 0, req); 1045 if (error) 1046 return (error); 1047 if (req->newptr == NULL) 1048 return (error); 1049 if (v == 0) 1050 return (0); 1051 sleepq_prof_reset(); 1052 1053 return (0); 1054} 1055 1056static int 1057dump_sleepq_prof_stats(SYSCTL_HANDLER_ARGS) 1058{ 1059 static int multiplier = 1; 1060 struct sleepq_prof *sp; 1061 struct sbuf *sb; 1062 int enabled; 1063 int error; 1064 int i; 1065 1066retry_sbufops: 1067 sb = sbuf_new(NULL, NULL, SLEEPQ_SBUFSIZE * multiplier, SBUF_FIXEDLEN); 1068 sbuf_printf(sb, "\nwmesg\tcount\n"); 1069 enabled = prof_enabled; 1070 mtx_lock_spin(&sleepq_prof_lock); 1071 prof_enabled = 0; 1072 mtx_unlock_spin(&sleepq_prof_lock); 1073 for (i = 0; i < SC_TABLESIZE; i++) { 1074 LIST_FOREACH(sp, &sleepq_hash[i], sp_link) { 1075 sbuf_printf(sb, "%s\t%ld\n", 1076 sp->sp_wmesg, sp->sp_count); 1077 if (sbuf_overflowed(sb)) { 1078 sbuf_delete(sb); 1079 multiplier++; 1080 goto retry_sbufops; 1081 } 1082 } 1083 } 1084 mtx_lock_spin(&sleepq_prof_lock); 1085 prof_enabled = enabled; 1086 mtx_unlock_spin(&sleepq_prof_lock); 1087 1088 sbuf_finish(sb); 1089 error = SYSCTL_OUT(req, sbuf_data(sb), sbuf_len(sb) + 1); 1090 sbuf_delete(sb); 1091 return (error); 1092} 1093 1094SYSCTL_PROC(_debug_sleepq, OID_AUTO, stats, CTLTYPE_STRING | CTLFLAG_RD, 1095 NULL, 0, dump_sleepq_prof_stats, "A", "Sleepqueue profiling statistics"); 1096SYSCTL_PROC(_debug_sleepq, OID_AUTO, reset, CTLTYPE_INT | CTLFLAG_RW, 1097 NULL, 0, reset_sleepq_prof_stats, "I", 1098 "Reset sleepqueue profiling statistics"); 1099SYSCTL_PROC(_debug_sleepq, OID_AUTO, enable, CTLTYPE_INT | CTLFLAG_RW, 1100 NULL, 0, enable_sleepq_prof, "I", "Enable sleepqueue profiling"); 1101#endif 1102 1103#ifdef DDB 1104DB_SHOW_COMMAND(sleepq, db_show_sleepqueue) 1105{ 1106 struct sleepqueue_chain *sc; 1107 struct sleepqueue *sq; 1108#ifdef INVARIANTS 1109 struct lock_object *lock; 1110#endif 1111 struct thread *td; 1112 void *wchan; 1113 int i; 1114 1115 if (!have_addr) 1116 return; 1117 1118 /* 1119 * First, see if there is an active sleep queue for the wait channel 1120 * indicated by the address. 1121 */ 1122 wchan = (void *)addr; 1123 sc = SC_LOOKUP(wchan); 1124 LIST_FOREACH(sq, &sc->sc_queues, sq_hash) 1125 if (sq->sq_wchan == wchan) 1126 goto found; 1127 1128 /* 1129 * Second, see if there is an active sleep queue at the address 1130 * indicated. 1131 */ 1132 for (i = 0; i < SC_TABLESIZE; i++) 1133 LIST_FOREACH(sq, &sleepq_chains[i].sc_queues, sq_hash) { 1134 if (sq == (struct sleepqueue *)addr) 1135 goto found; 1136 } 1137 1138 db_printf("Unable to locate a sleep queue via %p\n", (void *)addr); 1139 return; 1140found: 1141 db_printf("Wait channel: %p\n", sq->sq_wchan); 1142#ifdef INVARIANTS 1143 db_printf("Queue type: %d\n", sq->sq_type); 1144 if (sq->sq_lock) { 1145 lock = sq->sq_lock; 1146 db_printf("Associated Interlock: %p - (%s) %s\n", lock, 1147 LOCK_CLASS(lock)->lc_name, lock->lo_name); 1148 } 1149#endif 1150 db_printf("Blocked threads:\n"); 1151 for (i = 0; i < NR_SLEEPQS; i++) { 1152 db_printf("\nQueue[%d]:\n", i); 1153 if (TAILQ_EMPTY(&sq->sq_blocked[i])) 1154 db_printf("\tempty\n"); 1155 else 1156 TAILQ_FOREACH(td, &sq->sq_blocked[0], 1157 td_slpq) { 1158 db_printf("\t%p (tid %d, pid %d, \"%s\")\n", td, 1159 td->td_tid, td->td_proc->p_pid, 1160 td->td_name); 1161 } 1162 } 1163} 1164 1165/* Alias 'show sleepqueue' to 'show sleepq'. */ 1166DB_SHOW_ALIAS(sleepqueue, db_show_sleepqueue); 1167#endif 1168