1 /*- 2 * SPDX-License-Identifier: BSD-2-Clause-FreeBSD 3 * 4 * Copyright (c) 2004 John Baldwin <jhb@FreeBSD.org> 5 * 6 * Redistribution and use in source and binary forms, with or without 7 * modification, are permitted provided that the following conditions 8 * are met: 9 * 1. Redistributions of source code must retain the above copyright 10 * notice, this list of conditions and the following disclaimer. 11 * 2. Redistributions in binary form must reproduce the above copyright 12 * notice, this list of conditions and the following disclaimer in the 13 * documentation and/or other materials provided with the distribution. 14 * 15 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND 16 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 17 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 18 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE 19 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 20 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 21 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 22 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 23 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 24 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 25 * SUCH DAMAGE. 26 */ 27 28 /* 29 * Implementation of sleep queues used to hold queue of threads blocked on 30 * a wait channel. Sleep queues are different from turnstiles in that wait 31 * channels are not owned by anyone, so there is no priority propagation. 32 * Sleep queues can also provide a timeout and can also be interrupted by 33 * signals. That said, there are several similarities between the turnstile 34 * and sleep queue implementations. (Note: turnstiles were implemented 35 * first.) For example, both use a hash table of the same size where each 36 * bucket is referred to as a "chain" that contains both a spin lock and 37 * a linked list of queues. An individual queue is located by using a hash 38 * to pick a chain, locking the chain, and then walking the chain searching 39 * for the queue. This means that a wait channel object does not need to 40 * embed its queue head just as locks do not embed their turnstile queue 41 * head. Threads also carry around a sleep queue that they lend to the 42 * wait channel when blocking. Just as in turnstiles, the queue includes 43 * a free list of the sleep queues of other threads blocked on the same 44 * wait channel in the case of multiple waiters. 45 * 46 * Some additional functionality provided by sleep queues include the 47 * ability to set a timeout. The timeout is managed using a per-thread 48 * callout that resumes a thread if it is asleep. A thread may also 49 * catch signals while it is asleep (aka an interruptible sleep). The 50 * signal code uses sleepq_abort() to interrupt a sleeping thread. Finally, 51 * sleep queues also provide some extra assertions. One is not allowed to 52 * mix the sleep/wakeup and cv APIs for a given wait channel. Also, one 53 * must consistently use the same lock to synchronize with a wait channel, 54 * though this check is currently only a warning for sleep/wakeup due to 55 * pre-existing abuse of that API. The same lock must also be held when 56 * awakening threads, though that is currently only enforced for condition 57 * variables. 58 */ 59 60 #include <sys/cdefs.h> 61 __FBSDID("$FreeBSD$"); 62 63 #include "opt_sleepqueue_profiling.h" 64 #include "opt_ddb.h" 65 #include "opt_sched.h" 66 #include "opt_stack.h" 67 68 #include <sys/param.h> 69 #include <sys/systm.h> 70 #include <sys/lock.h> 71 #include <sys/kernel.h> 72 #include <sys/ktr.h> 73 #include <sys/mutex.h> 74 #include <sys/proc.h> 75 #include <sys/sbuf.h> 76 #include <sys/sched.h> 77 #include <sys/sdt.h> 78 #include <sys/signalvar.h> 79 #include <sys/sleepqueue.h> 80 #include <sys/stack.h> 81 #include <sys/sysctl.h> 82 #include <sys/time.h> 83 #ifdef EPOCH_TRACE 84 #include <sys/epoch.h> 85 #endif 86 87 #include <machine/atomic.h> 88 89 #include <vm/uma.h> 90 91 #ifdef DDB 92 #include <ddb/ddb.h> 93 #endif 94 95 /* 96 * Constants for the hash table of sleep queue chains. 97 * SC_TABLESIZE must be a power of two for SC_MASK to work properly. 98 */ 99 #ifndef SC_TABLESIZE 100 #define SC_TABLESIZE 256 101 #endif 102 CTASSERT(powerof2(SC_TABLESIZE)); 103 #define SC_MASK (SC_TABLESIZE - 1) 104 #define SC_SHIFT 8 105 #define SC_HASH(wc) ((((uintptr_t)(wc) >> SC_SHIFT) ^ (uintptr_t)(wc)) & \ 106 SC_MASK) 107 #define SC_LOOKUP(wc) &sleepq_chains[SC_HASH(wc)] 108 #define NR_SLEEPQS 2 109 /* 110 * There are two different lists of sleep queues. Both lists are connected 111 * via the sq_hash entries. The first list is the sleep queue chain list 112 * that a sleep queue is on when it is attached to a wait channel. The 113 * second list is the free list hung off of a sleep queue that is attached 114 * to a wait channel. 115 * 116 * Each sleep queue also contains the wait channel it is attached to, the 117 * list of threads blocked on that wait channel, flags specific to the 118 * wait channel, and the lock used to synchronize with a wait channel. 119 * The flags are used to catch mismatches between the various consumers 120 * of the sleep queue API (e.g. sleep/wakeup and condition variables). 121 * The lock pointer is only used when invariants are enabled for various 122 * debugging checks. 123 * 124 * Locking key: 125 * c - sleep queue chain lock 126 */ 127 struct sleepqueue { 128 struct threadqueue sq_blocked[NR_SLEEPQS]; /* (c) Blocked threads. */ 129 u_int sq_blockedcnt[NR_SLEEPQS]; /* (c) N. of blocked threads. */ 130 LIST_ENTRY(sleepqueue) sq_hash; /* (c) Chain and free list. */ 131 LIST_HEAD(, sleepqueue) sq_free; /* (c) Free queues. */ 132 const void *sq_wchan; /* (c) Wait channel. */ 133 int sq_type; /* (c) Queue type. */ 134 #ifdef INVARIANTS 135 struct lock_object *sq_lock; /* (c) Associated lock. */ 136 #endif 137 }; 138 139 struct sleepqueue_chain { 140 LIST_HEAD(, sleepqueue) sc_queues; /* List of sleep queues. */ 141 struct mtx sc_lock; /* Spin lock for this chain. */ 142 #ifdef SLEEPQUEUE_PROFILING 143 u_int sc_depth; /* Length of sc_queues. */ 144 u_int sc_max_depth; /* Max length of sc_queues. */ 145 #endif 146 } __aligned(CACHE_LINE_SIZE); 147 148 #ifdef SLEEPQUEUE_PROFILING 149 u_int sleepq_max_depth; 150 static SYSCTL_NODE(_debug, OID_AUTO, sleepq, CTLFLAG_RD, 0, "sleepq profiling"); 151 static SYSCTL_NODE(_debug_sleepq, OID_AUTO, chains, CTLFLAG_RD, 0, 152 "sleepq chain stats"); 153 SYSCTL_UINT(_debug_sleepq, OID_AUTO, max_depth, CTLFLAG_RD, &sleepq_max_depth, 154 0, "maxmimum depth achieved of a single chain"); 155 156 static void sleepq_profile(const char *wmesg); 157 static int prof_enabled; 158 #endif 159 static struct sleepqueue_chain sleepq_chains[SC_TABLESIZE]; 160 static uma_zone_t sleepq_zone; 161 162 /* 163 * Prototypes for non-exported routines. 164 */ 165 static int sleepq_catch_signals(const void *wchan, int pri); 166 static inline int sleepq_check_signals(void); 167 static inline int sleepq_check_timeout(void); 168 #ifdef INVARIANTS 169 static void sleepq_dtor(void *mem, int size, void *arg); 170 #endif 171 static int sleepq_init(void *mem, int size, int flags); 172 static int sleepq_resume_thread(struct sleepqueue *sq, struct thread *td, 173 int pri, int srqflags); 174 static void sleepq_remove_thread(struct sleepqueue *sq, struct thread *td); 175 static void sleepq_switch(const void *wchan, int pri); 176 static void sleepq_timeout(void *arg); 177 178 SDT_PROBE_DECLARE(sched, , , sleep); 179 SDT_PROBE_DECLARE(sched, , , wakeup); 180 181 /* 182 * Initialize SLEEPQUEUE_PROFILING specific sysctl nodes. 183 * Note that it must happen after sleepinit() has been fully executed, so 184 * it must happen after SI_SUB_KMEM SYSINIT() subsystem setup. 185 */ 186 #ifdef SLEEPQUEUE_PROFILING 187 static void 188 init_sleepqueue_profiling(void) 189 { 190 char chain_name[10]; 191 struct sysctl_oid *chain_oid; 192 u_int i; 193 194 for (i = 0; i < SC_TABLESIZE; i++) { 195 snprintf(chain_name, sizeof(chain_name), "%u", i); 196 chain_oid = SYSCTL_ADD_NODE(NULL, 197 SYSCTL_STATIC_CHILDREN(_debug_sleepq_chains), OID_AUTO, 198 chain_name, CTLFLAG_RD, NULL, "sleepq chain stats"); 199 SYSCTL_ADD_UINT(NULL, SYSCTL_CHILDREN(chain_oid), OID_AUTO, 200 "depth", CTLFLAG_RD, &sleepq_chains[i].sc_depth, 0, NULL); 201 SYSCTL_ADD_UINT(NULL, SYSCTL_CHILDREN(chain_oid), OID_AUTO, 202 "max_depth", CTLFLAG_RD, &sleepq_chains[i].sc_max_depth, 0, 203 NULL); 204 } 205 } 206 207 SYSINIT(sleepqueue_profiling, SI_SUB_LOCK, SI_ORDER_ANY, 208 init_sleepqueue_profiling, NULL); 209 #endif 210 211 /* 212 * Early initialization of sleep queues that is called from the sleepinit() 213 * SYSINIT. 214 */ 215 void 216 init_sleepqueues(void) 217 { 218 int i; 219 220 for (i = 0; i < SC_TABLESIZE; i++) { 221 LIST_INIT(&sleepq_chains[i].sc_queues); 222 mtx_init(&sleepq_chains[i].sc_lock, "sleepq chain", NULL, 223 MTX_SPIN); 224 } 225 sleepq_zone = uma_zcreate("SLEEPQUEUE", sizeof(struct sleepqueue), 226 #ifdef INVARIANTS 227 NULL, sleepq_dtor, sleepq_init, NULL, UMA_ALIGN_CACHE, 0); 228 #else 229 NULL, NULL, sleepq_init, NULL, UMA_ALIGN_CACHE, 0); 230 #endif 231 232 thread0.td_sleepqueue = sleepq_alloc(); 233 } 234 235 /* 236 * Get a sleep queue for a new thread. 237 */ 238 struct sleepqueue * 239 sleepq_alloc(void) 240 { 241 242 return (uma_zalloc(sleepq_zone, M_WAITOK)); 243 } 244 245 /* 246 * Free a sleep queue when a thread is destroyed. 247 */ 248 void 249 sleepq_free(struct sleepqueue *sq) 250 { 251 252 uma_zfree(sleepq_zone, sq); 253 } 254 255 /* 256 * Lock the sleep queue chain associated with the specified wait channel. 257 */ 258 void 259 sleepq_lock(const void *wchan) 260 { 261 struct sleepqueue_chain *sc; 262 263 sc = SC_LOOKUP(wchan); 264 mtx_lock_spin(&sc->sc_lock); 265 } 266 267 /* 268 * Look up the sleep queue associated with a given wait channel in the hash 269 * table locking the associated sleep queue chain. If no queue is found in 270 * the table, NULL is returned. 271 */ 272 struct sleepqueue * 273 sleepq_lookup(const void *wchan) 274 { 275 struct sleepqueue_chain *sc; 276 struct sleepqueue *sq; 277 278 KASSERT(wchan != NULL, ("%s: invalid NULL wait channel", __func__)); 279 sc = SC_LOOKUP(wchan); 280 mtx_assert(&sc->sc_lock, MA_OWNED); 281 LIST_FOREACH(sq, &sc->sc_queues, sq_hash) 282 if (sq->sq_wchan == wchan) 283 return (sq); 284 return (NULL); 285 } 286 287 /* 288 * Unlock the sleep queue chain associated with a given wait channel. 289 */ 290 void 291 sleepq_release(const void *wchan) 292 { 293 struct sleepqueue_chain *sc; 294 295 sc = SC_LOOKUP(wchan); 296 mtx_unlock_spin(&sc->sc_lock); 297 } 298 299 /* 300 * Places the current thread on the sleep queue for the specified wait 301 * channel. If INVARIANTS is enabled, then it associates the passed in 302 * lock with the sleepq to make sure it is held when that sleep queue is 303 * woken up. 304 */ 305 void 306 sleepq_add(const void *wchan, struct lock_object *lock, const char *wmesg, 307 int flags, int queue) 308 { 309 struct sleepqueue_chain *sc; 310 struct sleepqueue *sq; 311 struct thread *td; 312 313 td = curthread; 314 sc = SC_LOOKUP(wchan); 315 mtx_assert(&sc->sc_lock, MA_OWNED); 316 MPASS(td->td_sleepqueue != NULL); 317 MPASS(wchan != NULL); 318 MPASS((queue >= 0) && (queue < NR_SLEEPQS)); 319 320 /* If this thread is not allowed to sleep, die a horrible death. */ 321 if (__predict_false(!THREAD_CAN_SLEEP())) { 322 #ifdef EPOCH_TRACE 323 epoch_trace_list(curthread); 324 #endif 325 KASSERT(1, 326 ("%s: td %p to sleep on wchan %p with sleeping prohibited", 327 __func__, td, wchan)); 328 } 329 330 /* Look up the sleep queue associated with the wait channel 'wchan'. */ 331 sq = sleepq_lookup(wchan); 332 333 /* 334 * If the wait channel does not already have a sleep queue, use 335 * this thread's sleep queue. Otherwise, insert the current thread 336 * into the sleep queue already in use by this wait channel. 337 */ 338 if (sq == NULL) { 339 #ifdef INVARIANTS 340 int i; 341 342 sq = td->td_sleepqueue; 343 for (i = 0; i < NR_SLEEPQS; i++) { 344 KASSERT(TAILQ_EMPTY(&sq->sq_blocked[i]), 345 ("thread's sleep queue %d is not empty", i)); 346 KASSERT(sq->sq_blockedcnt[i] == 0, 347 ("thread's sleep queue %d count mismatches", i)); 348 } 349 KASSERT(LIST_EMPTY(&sq->sq_free), 350 ("thread's sleep queue has a non-empty free list")); 351 KASSERT(sq->sq_wchan == NULL, ("stale sq_wchan pointer")); 352 sq->sq_lock = lock; 353 #endif 354 #ifdef SLEEPQUEUE_PROFILING 355 sc->sc_depth++; 356 if (sc->sc_depth > sc->sc_max_depth) { 357 sc->sc_max_depth = sc->sc_depth; 358 if (sc->sc_max_depth > sleepq_max_depth) 359 sleepq_max_depth = sc->sc_max_depth; 360 } 361 #endif 362 sq = td->td_sleepqueue; 363 LIST_INSERT_HEAD(&sc->sc_queues, sq, sq_hash); 364 sq->sq_wchan = wchan; 365 sq->sq_type = flags & SLEEPQ_TYPE; 366 } else { 367 MPASS(wchan == sq->sq_wchan); 368 MPASS(lock == sq->sq_lock); 369 MPASS((flags & SLEEPQ_TYPE) == sq->sq_type); 370 LIST_INSERT_HEAD(&sq->sq_free, td->td_sleepqueue, sq_hash); 371 } 372 thread_lock(td); 373 TAILQ_INSERT_TAIL(&sq->sq_blocked[queue], td, td_slpq); 374 sq->sq_blockedcnt[queue]++; 375 td->td_sleepqueue = NULL; 376 td->td_sqqueue = queue; 377 td->td_wchan = wchan; 378 td->td_wmesg = wmesg; 379 if (flags & SLEEPQ_INTERRUPTIBLE) { 380 td->td_intrval = 0; 381 td->td_flags |= TDF_SINTR; 382 } 383 td->td_flags &= ~TDF_TIMEOUT; 384 thread_unlock(td); 385 } 386 387 /* 388 * Sets a timeout that will remove the current thread from the specified 389 * sleep queue after timo ticks if the thread has not already been awakened. 390 */ 391 void 392 sleepq_set_timeout_sbt(const void *wchan, sbintime_t sbt, sbintime_t pr, 393 int flags) 394 { 395 struct sleepqueue_chain *sc __unused; 396 struct thread *td; 397 sbintime_t pr1; 398 399 td = curthread; 400 sc = SC_LOOKUP(wchan); 401 mtx_assert(&sc->sc_lock, MA_OWNED); 402 MPASS(TD_ON_SLEEPQ(td)); 403 MPASS(td->td_sleepqueue == NULL); 404 MPASS(wchan != NULL); 405 if (cold && td == &thread0) 406 panic("timed sleep before timers are working"); 407 KASSERT(td->td_sleeptimo == 0, ("td %d %p td_sleeptimo %jx", 408 td->td_tid, td, (uintmax_t)td->td_sleeptimo)); 409 thread_lock(td); 410 callout_when(sbt, pr, flags, &td->td_sleeptimo, &pr1); 411 thread_unlock(td); 412 callout_reset_sbt_on(&td->td_slpcallout, td->td_sleeptimo, pr1, 413 sleepq_timeout, td, PCPU_GET(cpuid), flags | C_PRECALC | 414 C_DIRECT_EXEC); 415 } 416 417 /* 418 * Return the number of actual sleepers for the specified queue. 419 */ 420 u_int 421 sleepq_sleepcnt(const void *wchan, int queue) 422 { 423 struct sleepqueue *sq; 424 425 KASSERT(wchan != NULL, ("%s: invalid NULL wait channel", __func__)); 426 MPASS((queue >= 0) && (queue < NR_SLEEPQS)); 427 sq = sleepq_lookup(wchan); 428 if (sq == NULL) 429 return (0); 430 return (sq->sq_blockedcnt[queue]); 431 } 432 433 /* 434 * Marks the pending sleep of the current thread as interruptible and 435 * makes an initial check for pending signals before putting a thread 436 * to sleep. Enters and exits with the thread lock held. Thread lock 437 * may have transitioned from the sleepq lock to a run lock. 438 */ 439 static int 440 sleepq_catch_signals(const void *wchan, int pri) 441 { 442 struct sleepqueue_chain *sc; 443 struct sleepqueue *sq; 444 struct thread *td; 445 struct proc *p; 446 struct sigacts *ps; 447 int sig, ret; 448 449 ret = 0; 450 td = curthread; 451 p = curproc; 452 sc = SC_LOOKUP(wchan); 453 mtx_assert(&sc->sc_lock, MA_OWNED); 454 MPASS(wchan != NULL); 455 if ((td->td_pflags & TDP_WAKEUP) != 0) { 456 td->td_pflags &= ~TDP_WAKEUP; 457 ret = EINTR; 458 thread_lock(td); 459 goto out; 460 } 461 462 /* 463 * See if there are any pending signals or suspension requests for this 464 * thread. If not, we can switch immediately. 465 */ 466 thread_lock(td); 467 if ((td->td_flags & (TDF_NEEDSIGCHK | TDF_NEEDSUSPCHK)) != 0) { 468 thread_unlock(td); 469 mtx_unlock_spin(&sc->sc_lock); 470 CTR3(KTR_PROC, "sleepq catching signals: thread %p (pid %ld, %s)", 471 (void *)td, (long)p->p_pid, td->td_name); 472 PROC_LOCK(p); 473 /* 474 * Check for suspension first. Checking for signals and then 475 * suspending could result in a missed signal, since a signal 476 * can be delivered while this thread is suspended. 477 */ 478 if ((td->td_flags & TDF_NEEDSUSPCHK) != 0) { 479 ret = thread_suspend_check(1); 480 MPASS(ret == 0 || ret == EINTR || ret == ERESTART); 481 if (ret != 0) { 482 PROC_UNLOCK(p); 483 mtx_lock_spin(&sc->sc_lock); 484 thread_lock(td); 485 goto out; 486 } 487 } 488 if ((td->td_flags & TDF_NEEDSIGCHK) != 0) { 489 ps = p->p_sigacts; 490 mtx_lock(&ps->ps_mtx); 491 sig = cursig(td); 492 if (sig == -1) { 493 mtx_unlock(&ps->ps_mtx); 494 KASSERT((td->td_flags & TDF_SBDRY) != 0, 495 ("lost TDF_SBDRY")); 496 KASSERT(TD_SBDRY_INTR(td), 497 ("lost TDF_SERESTART of TDF_SEINTR")); 498 KASSERT((td->td_flags & 499 (TDF_SEINTR | TDF_SERESTART)) != 500 (TDF_SEINTR | TDF_SERESTART), 501 ("both TDF_SEINTR and TDF_SERESTART")); 502 ret = TD_SBDRY_ERRNO(td); 503 } else if (sig != 0) { 504 ret = SIGISMEMBER(ps->ps_sigintr, sig) ? 505 EINTR : ERESTART; 506 mtx_unlock(&ps->ps_mtx); 507 } else { 508 mtx_unlock(&ps->ps_mtx); 509 } 510 511 /* 512 * Do not go into sleep if this thread was the 513 * ptrace(2) attach leader. cursig() consumed 514 * SIGSTOP from PT_ATTACH, but we usually act 515 * on the signal by interrupting sleep, and 516 * should do that here as well. 517 */ 518 if ((td->td_dbgflags & TDB_FSTP) != 0) { 519 if (ret == 0) 520 ret = EINTR; 521 td->td_dbgflags &= ~TDB_FSTP; 522 } 523 } 524 /* 525 * Lock the per-process spinlock prior to dropping the PROC_LOCK 526 * to avoid a signal delivery race. PROC_LOCK, PROC_SLOCK, and 527 * thread_lock() are currently held in tdsendsignal(). 528 */ 529 PROC_SLOCK(p); 530 mtx_lock_spin(&sc->sc_lock); 531 PROC_UNLOCK(p); 532 thread_lock(td); 533 PROC_SUNLOCK(p); 534 } 535 if (ret == 0) { 536 sleepq_switch(wchan, pri); 537 return (0); 538 } 539 out: 540 /* 541 * There were pending signals and this thread is still 542 * on the sleep queue, remove it from the sleep queue. 543 */ 544 if (TD_ON_SLEEPQ(td)) { 545 sq = sleepq_lookup(wchan); 546 sleepq_remove_thread(sq, td); 547 } 548 MPASS(td->td_lock != &sc->sc_lock); 549 mtx_unlock_spin(&sc->sc_lock); 550 thread_unlock(td); 551 552 return (ret); 553 } 554 555 /* 556 * Switches to another thread if we are still asleep on a sleep queue. 557 * Returns with thread lock. 558 */ 559 static void 560 sleepq_switch(const void *wchan, int pri) 561 { 562 struct sleepqueue_chain *sc; 563 struct sleepqueue *sq; 564 struct thread *td; 565 bool rtc_changed; 566 567 td = curthread; 568 sc = SC_LOOKUP(wchan); 569 mtx_assert(&sc->sc_lock, MA_OWNED); 570 THREAD_LOCK_ASSERT(td, MA_OWNED); 571 572 /* 573 * If we have a sleep queue, then we've already been woken up, so 574 * just return. 575 */ 576 if (td->td_sleepqueue != NULL) { 577 mtx_unlock_spin(&sc->sc_lock); 578 thread_unlock(td); 579 return; 580 } 581 582 /* 583 * If TDF_TIMEOUT is set, then our sleep has been timed out 584 * already but we are still on the sleep queue, so dequeue the 585 * thread and return. 586 * 587 * Do the same if the real-time clock has been adjusted since this 588 * thread calculated its timeout based on that clock. This handles 589 * the following race: 590 * - The Ts thread needs to sleep until an absolute real-clock time. 591 * It copies the global rtc_generation into curthread->td_rtcgen, 592 * reads the RTC, and calculates a sleep duration based on that time. 593 * See umtxq_sleep() for an example. 594 * - The Tc thread adjusts the RTC, bumps rtc_generation, and wakes 595 * threads that are sleeping until an absolute real-clock time. 596 * See tc_setclock() and the POSIX specification of clock_settime(). 597 * - Ts reaches the code below. It holds the sleepqueue chain lock, 598 * so Tc has finished waking, so this thread must test td_rtcgen. 599 * (The declaration of td_rtcgen refers to this comment.) 600 */ 601 rtc_changed = td->td_rtcgen != 0 && td->td_rtcgen != rtc_generation; 602 if ((td->td_flags & TDF_TIMEOUT) || rtc_changed) { 603 if (rtc_changed) { 604 td->td_rtcgen = 0; 605 } 606 MPASS(TD_ON_SLEEPQ(td)); 607 sq = sleepq_lookup(wchan); 608 sleepq_remove_thread(sq, td); 609 mtx_unlock_spin(&sc->sc_lock); 610 thread_unlock(td); 611 return; 612 } 613 #ifdef SLEEPQUEUE_PROFILING 614 if (prof_enabled) 615 sleepq_profile(td->td_wmesg); 616 #endif 617 MPASS(td->td_sleepqueue == NULL); 618 sched_sleep(td, pri); 619 thread_lock_set(td, &sc->sc_lock); 620 SDT_PROBE0(sched, , , sleep); 621 TD_SET_SLEEPING(td); 622 mi_switch(SW_VOL | SWT_SLEEPQ); 623 KASSERT(TD_IS_RUNNING(td), ("running but not TDS_RUNNING")); 624 CTR3(KTR_PROC, "sleepq resume: thread %p (pid %ld, %s)", 625 (void *)td, (long)td->td_proc->p_pid, (void *)td->td_name); 626 } 627 628 /* 629 * Check to see if we timed out. 630 */ 631 static inline int 632 sleepq_check_timeout(void) 633 { 634 struct thread *td; 635 int res; 636 637 res = 0; 638 td = curthread; 639 if (td->td_sleeptimo != 0) { 640 if (td->td_sleeptimo <= sbinuptime()) 641 res = EWOULDBLOCK; 642 td->td_sleeptimo = 0; 643 } 644 return (res); 645 } 646 647 /* 648 * Check to see if we were awoken by a signal. 649 */ 650 static inline int 651 sleepq_check_signals(void) 652 { 653 struct thread *td; 654 655 td = curthread; 656 KASSERT((td->td_flags & TDF_SINTR) == 0, 657 ("thread %p still in interruptible sleep?", td)); 658 659 return (td->td_intrval); 660 } 661 662 /* 663 * Block the current thread until it is awakened from its sleep queue. 664 */ 665 void 666 sleepq_wait(const void *wchan, int pri) 667 { 668 struct thread *td; 669 670 td = curthread; 671 MPASS(!(td->td_flags & TDF_SINTR)); 672 thread_lock(td); 673 sleepq_switch(wchan, pri); 674 } 675 676 /* 677 * Block the current thread until it is awakened from its sleep queue 678 * or it is interrupted by a signal. 679 */ 680 int 681 sleepq_wait_sig(const void *wchan, int pri) 682 { 683 int rcatch; 684 685 rcatch = sleepq_catch_signals(wchan, pri); 686 if (rcatch) 687 return (rcatch); 688 return (sleepq_check_signals()); 689 } 690 691 /* 692 * Block the current thread until it is awakened from its sleep queue 693 * or it times out while waiting. 694 */ 695 int 696 sleepq_timedwait(const void *wchan, int pri) 697 { 698 struct thread *td; 699 700 td = curthread; 701 MPASS(!(td->td_flags & TDF_SINTR)); 702 703 thread_lock(td); 704 sleepq_switch(wchan, pri); 705 706 return (sleepq_check_timeout()); 707 } 708 709 /* 710 * Block the current thread until it is awakened from its sleep queue, 711 * it is interrupted by a signal, or it times out waiting to be awakened. 712 */ 713 int 714 sleepq_timedwait_sig(const void *wchan, int pri) 715 { 716 int rcatch, rvalt, rvals; 717 718 rcatch = sleepq_catch_signals(wchan, pri); 719 /* We must always call check_timeout() to clear sleeptimo. */ 720 rvalt = sleepq_check_timeout(); 721 rvals = sleepq_check_signals(); 722 if (rcatch) 723 return (rcatch); 724 if (rvals) 725 return (rvals); 726 return (rvalt); 727 } 728 729 /* 730 * Returns the type of sleepqueue given a waitchannel. 731 */ 732 int 733 sleepq_type(const void *wchan) 734 { 735 struct sleepqueue *sq; 736 int type; 737 738 MPASS(wchan != NULL); 739 740 sq = sleepq_lookup(wchan); 741 if (sq == NULL) 742 return (-1); 743 type = sq->sq_type; 744 745 return (type); 746 } 747 748 /* 749 * Removes a thread from a sleep queue and makes it 750 * runnable. 751 * 752 * Requires the sc chain locked on entry. If SRQ_HOLD is specified it will 753 * be locked on return. Returns without the thread lock held. 754 */ 755 static int 756 sleepq_resume_thread(struct sleepqueue *sq, struct thread *td, int pri, 757 int srqflags) 758 { 759 struct sleepqueue_chain *sc; 760 bool drop; 761 762 MPASS(td != NULL); 763 MPASS(sq->sq_wchan != NULL); 764 MPASS(td->td_wchan == sq->sq_wchan); 765 766 sc = SC_LOOKUP(sq->sq_wchan); 767 mtx_assert(&sc->sc_lock, MA_OWNED); 768 769 /* 770 * Avoid recursing on the chain lock. If the locks don't match we 771 * need to acquire the thread lock which setrunnable will drop for 772 * us. In this case we need to drop the chain lock afterwards. 773 * 774 * There is no race that will make td_lock equal to sc_lock because 775 * we hold sc_lock. 776 */ 777 drop = false; 778 if (!TD_IS_SLEEPING(td)) { 779 thread_lock(td); 780 drop = true; 781 } else 782 thread_lock_block_wait(td); 783 784 /* Remove thread from the sleepq. */ 785 sleepq_remove_thread(sq, td); 786 787 /* If we're done with the sleepqueue release it. */ 788 if ((srqflags & SRQ_HOLD) == 0 && drop) 789 mtx_unlock_spin(&sc->sc_lock); 790 791 /* Adjust priority if requested. */ 792 MPASS(pri == 0 || (pri >= PRI_MIN && pri <= PRI_MAX)); 793 if (pri != 0 && td->td_priority > pri && 794 PRI_BASE(td->td_pri_class) == PRI_TIMESHARE) 795 sched_prio(td, pri); 796 797 /* 798 * Note that thread td might not be sleeping if it is running 799 * sleepq_catch_signals() on another CPU or is blocked on its 800 * proc lock to check signals. There's no need to mark the 801 * thread runnable in that case. 802 */ 803 if (TD_IS_SLEEPING(td)) { 804 MPASS(!drop); 805 TD_CLR_SLEEPING(td); 806 return (setrunnable(td, srqflags)); 807 } 808 MPASS(drop); 809 thread_unlock(td); 810 811 return (0); 812 } 813 814 static void 815 sleepq_remove_thread(struct sleepqueue *sq, struct thread *td) 816 { 817 struct sleepqueue_chain *sc __unused; 818 819 MPASS(td != NULL); 820 MPASS(sq->sq_wchan != NULL); 821 MPASS(td->td_wchan == sq->sq_wchan); 822 MPASS(td->td_sqqueue < NR_SLEEPQS && td->td_sqqueue >= 0); 823 THREAD_LOCK_ASSERT(td, MA_OWNED); 824 sc = SC_LOOKUP(sq->sq_wchan); 825 mtx_assert(&sc->sc_lock, MA_OWNED); 826 827 SDT_PROBE2(sched, , , wakeup, td, td->td_proc); 828 829 /* Remove the thread from the queue. */ 830 sq->sq_blockedcnt[td->td_sqqueue]--; 831 TAILQ_REMOVE(&sq->sq_blocked[td->td_sqqueue], td, td_slpq); 832 833 /* 834 * Get a sleep queue for this thread. If this is the last waiter, 835 * use the queue itself and take it out of the chain, otherwise, 836 * remove a queue from the free list. 837 */ 838 if (LIST_EMPTY(&sq->sq_free)) { 839 td->td_sleepqueue = sq; 840 #ifdef INVARIANTS 841 sq->sq_wchan = NULL; 842 #endif 843 #ifdef SLEEPQUEUE_PROFILING 844 sc->sc_depth--; 845 #endif 846 } else 847 td->td_sleepqueue = LIST_FIRST(&sq->sq_free); 848 LIST_REMOVE(td->td_sleepqueue, sq_hash); 849 850 if ((td->td_flags & TDF_TIMEOUT) == 0 && td->td_sleeptimo != 0) 851 /* 852 * We ignore the situation where timeout subsystem was 853 * unable to stop our callout. The struct thread is 854 * type-stable, the callout will use the correct 855 * memory when running. The checks of the 856 * td_sleeptimo value in this function and in 857 * sleepq_timeout() ensure that the thread does not 858 * get spurious wakeups, even if the callout was reset 859 * or thread reused. 860 */ 861 callout_stop(&td->td_slpcallout); 862 863 td->td_wmesg = NULL; 864 td->td_wchan = NULL; 865 td->td_flags &= ~(TDF_SINTR | TDF_TIMEOUT); 866 867 CTR3(KTR_PROC, "sleepq_wakeup: thread %p (pid %ld, %s)", 868 (void *)td, (long)td->td_proc->p_pid, td->td_name); 869 } 870 871 #ifdef INVARIANTS 872 /* 873 * UMA zone item deallocator. 874 */ 875 static void 876 sleepq_dtor(void *mem, int size, void *arg) 877 { 878 struct sleepqueue *sq; 879 int i; 880 881 sq = mem; 882 for (i = 0; i < NR_SLEEPQS; i++) { 883 MPASS(TAILQ_EMPTY(&sq->sq_blocked[i])); 884 MPASS(sq->sq_blockedcnt[i] == 0); 885 } 886 } 887 #endif 888 889 /* 890 * UMA zone item initializer. 891 */ 892 static int 893 sleepq_init(void *mem, int size, int flags) 894 { 895 struct sleepqueue *sq; 896 int i; 897 898 bzero(mem, size); 899 sq = mem; 900 for (i = 0; i < NR_SLEEPQS; i++) { 901 TAILQ_INIT(&sq->sq_blocked[i]); 902 sq->sq_blockedcnt[i] = 0; 903 } 904 LIST_INIT(&sq->sq_free); 905 return (0); 906 } 907 908 /* 909 * Find thread sleeping on a wait channel and resume it. 910 */ 911 int 912 sleepq_signal(const void *wchan, int flags, int pri, int queue) 913 { 914 struct sleepqueue_chain *sc; 915 struct sleepqueue *sq; 916 struct threadqueue *head; 917 struct thread *td, *besttd; 918 int wakeup_swapper; 919 920 CTR2(KTR_PROC, "sleepq_signal(%p, %d)", wchan, flags); 921 KASSERT(wchan != NULL, ("%s: invalid NULL wait channel", __func__)); 922 MPASS((queue >= 0) && (queue < NR_SLEEPQS)); 923 sq = sleepq_lookup(wchan); 924 if (sq == NULL) 925 return (0); 926 KASSERT(sq->sq_type == (flags & SLEEPQ_TYPE), 927 ("%s: mismatch between sleep/wakeup and cv_*", __func__)); 928 929 head = &sq->sq_blocked[queue]; 930 if (flags & SLEEPQ_UNFAIR) { 931 /* 932 * Find the most recently sleeping thread, but try to 933 * skip threads still in process of context switch to 934 * avoid spinning on the thread lock. 935 */ 936 sc = SC_LOOKUP(wchan); 937 besttd = TAILQ_LAST_FAST(head, thread, td_slpq); 938 while (besttd->td_lock != &sc->sc_lock) { 939 td = TAILQ_PREV_FAST(besttd, head, thread, td_slpq); 940 if (td == NULL) 941 break; 942 besttd = td; 943 } 944 } else { 945 /* 946 * Find the highest priority thread on the queue. If there 947 * is a tie, use the thread that first appears in the queue 948 * as it has been sleeping the longest since threads are 949 * always added to the tail of sleep queues. 950 */ 951 besttd = td = TAILQ_FIRST(head); 952 while ((td = TAILQ_NEXT(td, td_slpq)) != NULL) { 953 if (td->td_priority < besttd->td_priority) 954 besttd = td; 955 } 956 } 957 MPASS(besttd != NULL); 958 wakeup_swapper = sleepq_resume_thread(sq, besttd, pri, SRQ_HOLD); 959 return (wakeup_swapper); 960 } 961 962 static bool 963 match_any(struct thread *td __unused) 964 { 965 966 return (true); 967 } 968 969 /* 970 * Resume all threads sleeping on a specified wait channel. 971 */ 972 int 973 sleepq_broadcast(const void *wchan, int flags, int pri, int queue) 974 { 975 struct sleepqueue *sq; 976 977 CTR2(KTR_PROC, "sleepq_broadcast(%p, %d)", wchan, flags); 978 KASSERT(wchan != NULL, ("%s: invalid NULL wait channel", __func__)); 979 MPASS((queue >= 0) && (queue < NR_SLEEPQS)); 980 sq = sleepq_lookup(wchan); 981 if (sq == NULL) 982 return (0); 983 KASSERT(sq->sq_type == (flags & SLEEPQ_TYPE), 984 ("%s: mismatch between sleep/wakeup and cv_*", __func__)); 985 986 return (sleepq_remove_matching(sq, queue, match_any, pri)); 987 } 988 989 /* 990 * Resume threads on the sleep queue that match the given predicate. 991 */ 992 int 993 sleepq_remove_matching(struct sleepqueue *sq, int queue, 994 bool (*matches)(struct thread *), int pri) 995 { 996 struct thread *td, *tdn; 997 int wakeup_swapper; 998 999 /* 1000 * The last thread will be given ownership of sq and may 1001 * re-enqueue itself before sleepq_resume_thread() returns, 1002 * so we must cache the "next" queue item at the beginning 1003 * of the final iteration. 1004 */ 1005 wakeup_swapper = 0; 1006 TAILQ_FOREACH_SAFE(td, &sq->sq_blocked[queue], td_slpq, tdn) { 1007 if (matches(td)) 1008 wakeup_swapper |= sleepq_resume_thread(sq, td, pri, 1009 SRQ_HOLD); 1010 } 1011 1012 return (wakeup_swapper); 1013 } 1014 1015 /* 1016 * Time sleeping threads out. When the timeout expires, the thread is 1017 * removed from the sleep queue and made runnable if it is still asleep. 1018 */ 1019 static void 1020 sleepq_timeout(void *arg) 1021 { 1022 struct sleepqueue_chain *sc __unused; 1023 struct sleepqueue *sq; 1024 struct thread *td; 1025 const void *wchan; 1026 int wakeup_swapper; 1027 1028 td = arg; 1029 CTR3(KTR_PROC, "sleepq_timeout: thread %p (pid %ld, %s)", 1030 (void *)td, (long)td->td_proc->p_pid, (void *)td->td_name); 1031 1032 thread_lock(td); 1033 if (td->td_sleeptimo == 0 || td->td_sleeptimo > sbinuptime()) { 1034 /* 1035 * The thread does not want a timeout (yet). 1036 */ 1037 } else if (TD_IS_SLEEPING(td) && TD_ON_SLEEPQ(td)) { 1038 /* 1039 * See if the thread is asleep and get the wait 1040 * channel if it is. 1041 */ 1042 wchan = td->td_wchan; 1043 sc = SC_LOOKUP(wchan); 1044 THREAD_LOCKPTR_ASSERT(td, &sc->sc_lock); 1045 sq = sleepq_lookup(wchan); 1046 MPASS(sq != NULL); 1047 td->td_flags |= TDF_TIMEOUT; 1048 wakeup_swapper = sleepq_resume_thread(sq, td, 0, 0); 1049 if (wakeup_swapper) 1050 kick_proc0(); 1051 return; 1052 } else if (TD_ON_SLEEPQ(td)) { 1053 /* 1054 * If the thread is on the SLEEPQ but isn't sleeping 1055 * yet, it can either be on another CPU in between 1056 * sleepq_add() and one of the sleepq_*wait*() 1057 * routines or it can be in sleepq_catch_signals(). 1058 */ 1059 td->td_flags |= TDF_TIMEOUT; 1060 } 1061 thread_unlock(td); 1062 } 1063 1064 /* 1065 * Resumes a specific thread from the sleep queue associated with a specific 1066 * wait channel if it is on that queue. 1067 */ 1068 void 1069 sleepq_remove(struct thread *td, const void *wchan) 1070 { 1071 struct sleepqueue_chain *sc; 1072 struct sleepqueue *sq; 1073 int wakeup_swapper; 1074 1075 /* 1076 * Look up the sleep queue for this wait channel, then re-check 1077 * that the thread is asleep on that channel, if it is not, then 1078 * bail. 1079 */ 1080 MPASS(wchan != NULL); 1081 sc = SC_LOOKUP(wchan); 1082 mtx_lock_spin(&sc->sc_lock); 1083 /* 1084 * We can not lock the thread here as it may be sleeping on a 1085 * different sleepq. However, holding the sleepq lock for this 1086 * wchan can guarantee that we do not miss a wakeup for this 1087 * channel. The asserts below will catch any false positives. 1088 */ 1089 if (!TD_ON_SLEEPQ(td) || td->td_wchan != wchan) { 1090 mtx_unlock_spin(&sc->sc_lock); 1091 return; 1092 } 1093 1094 /* Thread is asleep on sleep queue sq, so wake it up. */ 1095 sq = sleepq_lookup(wchan); 1096 MPASS(sq != NULL); 1097 MPASS(td->td_wchan == wchan); 1098 wakeup_swapper = sleepq_resume_thread(sq, td, 0, 0); 1099 if (wakeup_swapper) 1100 kick_proc0(); 1101 } 1102 1103 /* 1104 * Abort a thread as if an interrupt had occurred. Only abort 1105 * interruptible waits (unfortunately it isn't safe to abort others). 1106 * 1107 * Requires thread lock on entry, releases on return. 1108 */ 1109 int 1110 sleepq_abort(struct thread *td, int intrval) 1111 { 1112 struct sleepqueue *sq; 1113 const void *wchan; 1114 1115 THREAD_LOCK_ASSERT(td, MA_OWNED); 1116 MPASS(TD_ON_SLEEPQ(td)); 1117 MPASS(td->td_flags & TDF_SINTR); 1118 MPASS(intrval == EINTR || intrval == ERESTART); 1119 1120 /* 1121 * If the TDF_TIMEOUT flag is set, just leave. A 1122 * timeout is scheduled anyhow. 1123 */ 1124 if (td->td_flags & TDF_TIMEOUT) { 1125 thread_unlock(td); 1126 return (0); 1127 } 1128 1129 CTR3(KTR_PROC, "sleepq_abort: thread %p (pid %ld, %s)", 1130 (void *)td, (long)td->td_proc->p_pid, (void *)td->td_name); 1131 td->td_intrval = intrval; 1132 1133 /* 1134 * If the thread has not slept yet it will find the signal in 1135 * sleepq_catch_signals() and call sleepq_resume_thread. Otherwise 1136 * we have to do it here. 1137 */ 1138 if (!TD_IS_SLEEPING(td)) { 1139 thread_unlock(td); 1140 return (0); 1141 } 1142 wchan = td->td_wchan; 1143 MPASS(wchan != NULL); 1144 sq = sleepq_lookup(wchan); 1145 MPASS(sq != NULL); 1146 1147 /* Thread is asleep on sleep queue sq, so wake it up. */ 1148 return (sleepq_resume_thread(sq, td, 0, 0)); 1149 } 1150 1151 void 1152 sleepq_chains_remove_matching(bool (*matches)(struct thread *)) 1153 { 1154 struct sleepqueue_chain *sc; 1155 struct sleepqueue *sq, *sq1; 1156 int i, wakeup_swapper; 1157 1158 wakeup_swapper = 0; 1159 for (sc = &sleepq_chains[0]; sc < sleepq_chains + SC_TABLESIZE; ++sc) { 1160 if (LIST_EMPTY(&sc->sc_queues)) { 1161 continue; 1162 } 1163 mtx_lock_spin(&sc->sc_lock); 1164 LIST_FOREACH_SAFE(sq, &sc->sc_queues, sq_hash, sq1) { 1165 for (i = 0; i < NR_SLEEPQS; ++i) { 1166 wakeup_swapper |= sleepq_remove_matching(sq, i, 1167 matches, 0); 1168 } 1169 } 1170 mtx_unlock_spin(&sc->sc_lock); 1171 } 1172 if (wakeup_swapper) { 1173 kick_proc0(); 1174 } 1175 } 1176 1177 /* 1178 * Prints the stacks of all threads presently sleeping on wchan/queue to 1179 * the sbuf sb. Sets count_stacks_printed to the number of stacks actually 1180 * printed. Typically, this will equal the number of threads sleeping on the 1181 * queue, but may be less if sb overflowed before all stacks were printed. 1182 */ 1183 #ifdef STACK 1184 int 1185 sleepq_sbuf_print_stacks(struct sbuf *sb, const void *wchan, int queue, 1186 int *count_stacks_printed) 1187 { 1188 struct thread *td, *td_next; 1189 struct sleepqueue *sq; 1190 struct stack **st; 1191 struct sbuf **td_infos; 1192 int i, stack_idx, error, stacks_to_allocate; 1193 bool finished; 1194 1195 error = 0; 1196 finished = false; 1197 1198 KASSERT(wchan != NULL, ("%s: invalid NULL wait channel", __func__)); 1199 MPASS((queue >= 0) && (queue < NR_SLEEPQS)); 1200 1201 stacks_to_allocate = 10; 1202 for (i = 0; i < 3 && !finished ; i++) { 1203 /* We cannot malloc while holding the queue's spinlock, so 1204 * we do our mallocs now, and hope it is enough. If it 1205 * isn't, we will free these, drop the lock, malloc more, 1206 * and try again, up to a point. After that point we will 1207 * give up and report ENOMEM. We also cannot write to sb 1208 * during this time since the client may have set the 1209 * SBUF_AUTOEXTEND flag on their sbuf, which could cause a 1210 * malloc as we print to it. So we defer actually printing 1211 * to sb until after we drop the spinlock. 1212 */ 1213 1214 /* Where we will store the stacks. */ 1215 st = malloc(sizeof(struct stack *) * stacks_to_allocate, 1216 M_TEMP, M_WAITOK); 1217 for (stack_idx = 0; stack_idx < stacks_to_allocate; 1218 stack_idx++) 1219 st[stack_idx] = stack_create(M_WAITOK); 1220 1221 /* Where we will store the td name, tid, etc. */ 1222 td_infos = malloc(sizeof(struct sbuf *) * stacks_to_allocate, 1223 M_TEMP, M_WAITOK); 1224 for (stack_idx = 0; stack_idx < stacks_to_allocate; 1225 stack_idx++) 1226 td_infos[stack_idx] = sbuf_new(NULL, NULL, 1227 MAXCOMLEN + sizeof(struct thread *) * 2 + 40, 1228 SBUF_FIXEDLEN); 1229 1230 sleepq_lock(wchan); 1231 sq = sleepq_lookup(wchan); 1232 if (sq == NULL) { 1233 /* This sleepq does not exist; exit and return ENOENT. */ 1234 error = ENOENT; 1235 finished = true; 1236 sleepq_release(wchan); 1237 goto loop_end; 1238 } 1239 1240 stack_idx = 0; 1241 /* Save thread info */ 1242 TAILQ_FOREACH_SAFE(td, &sq->sq_blocked[queue], td_slpq, 1243 td_next) { 1244 if (stack_idx >= stacks_to_allocate) 1245 goto loop_end; 1246 1247 /* Note the td_lock is equal to the sleepq_lock here. */ 1248 (void)stack_save_td(st[stack_idx], td); 1249 1250 sbuf_printf(td_infos[stack_idx], "%d: %s %p", 1251 td->td_tid, td->td_name, td); 1252 1253 ++stack_idx; 1254 } 1255 1256 finished = true; 1257 sleepq_release(wchan); 1258 1259 /* Print the stacks */ 1260 for (i = 0; i < stack_idx; i++) { 1261 sbuf_finish(td_infos[i]); 1262 sbuf_printf(sb, "--- thread %s: ---\n", sbuf_data(td_infos[i])); 1263 stack_sbuf_print(sb, st[i]); 1264 sbuf_printf(sb, "\n"); 1265 1266 error = sbuf_error(sb); 1267 if (error == 0) 1268 *count_stacks_printed = stack_idx; 1269 } 1270 1271 loop_end: 1272 if (!finished) 1273 sleepq_release(wchan); 1274 for (stack_idx = 0; stack_idx < stacks_to_allocate; 1275 stack_idx++) 1276 stack_destroy(st[stack_idx]); 1277 for (stack_idx = 0; stack_idx < stacks_to_allocate; 1278 stack_idx++) 1279 sbuf_delete(td_infos[stack_idx]); 1280 free(st, M_TEMP); 1281 free(td_infos, M_TEMP); 1282 stacks_to_allocate *= 10; 1283 } 1284 1285 if (!finished && error == 0) 1286 error = ENOMEM; 1287 1288 return (error); 1289 } 1290 #endif 1291 1292 #ifdef SLEEPQUEUE_PROFILING 1293 #define SLEEPQ_PROF_LOCATIONS 1024 1294 #define SLEEPQ_SBUFSIZE 512 1295 struct sleepq_prof { 1296 LIST_ENTRY(sleepq_prof) sp_link; 1297 const char *sp_wmesg; 1298 long sp_count; 1299 }; 1300 1301 LIST_HEAD(sqphead, sleepq_prof); 1302 1303 struct sqphead sleepq_prof_free; 1304 struct sqphead sleepq_hash[SC_TABLESIZE]; 1305 static struct sleepq_prof sleepq_profent[SLEEPQ_PROF_LOCATIONS]; 1306 static struct mtx sleepq_prof_lock; 1307 MTX_SYSINIT(sleepq_prof_lock, &sleepq_prof_lock, "sleepq_prof", MTX_SPIN); 1308 1309 static void 1310 sleepq_profile(const char *wmesg) 1311 { 1312 struct sleepq_prof *sp; 1313 1314 mtx_lock_spin(&sleepq_prof_lock); 1315 if (prof_enabled == 0) 1316 goto unlock; 1317 LIST_FOREACH(sp, &sleepq_hash[SC_HASH(wmesg)], sp_link) 1318 if (sp->sp_wmesg == wmesg) 1319 goto done; 1320 sp = LIST_FIRST(&sleepq_prof_free); 1321 if (sp == NULL) 1322 goto unlock; 1323 sp->sp_wmesg = wmesg; 1324 LIST_REMOVE(sp, sp_link); 1325 LIST_INSERT_HEAD(&sleepq_hash[SC_HASH(wmesg)], sp, sp_link); 1326 done: 1327 sp->sp_count++; 1328 unlock: 1329 mtx_unlock_spin(&sleepq_prof_lock); 1330 return; 1331 } 1332 1333 static void 1334 sleepq_prof_reset(void) 1335 { 1336 struct sleepq_prof *sp; 1337 int enabled; 1338 int i; 1339 1340 mtx_lock_spin(&sleepq_prof_lock); 1341 enabled = prof_enabled; 1342 prof_enabled = 0; 1343 for (i = 0; i < SC_TABLESIZE; i++) 1344 LIST_INIT(&sleepq_hash[i]); 1345 LIST_INIT(&sleepq_prof_free); 1346 for (i = 0; i < SLEEPQ_PROF_LOCATIONS; i++) { 1347 sp = &sleepq_profent[i]; 1348 sp->sp_wmesg = NULL; 1349 sp->sp_count = 0; 1350 LIST_INSERT_HEAD(&sleepq_prof_free, sp, sp_link); 1351 } 1352 prof_enabled = enabled; 1353 mtx_unlock_spin(&sleepq_prof_lock); 1354 } 1355 1356 static int 1357 enable_sleepq_prof(SYSCTL_HANDLER_ARGS) 1358 { 1359 int error, v; 1360 1361 v = prof_enabled; 1362 error = sysctl_handle_int(oidp, &v, v, req); 1363 if (error) 1364 return (error); 1365 if (req->newptr == NULL) 1366 return (error); 1367 if (v == prof_enabled) 1368 return (0); 1369 if (v == 1) 1370 sleepq_prof_reset(); 1371 mtx_lock_spin(&sleepq_prof_lock); 1372 prof_enabled = !!v; 1373 mtx_unlock_spin(&sleepq_prof_lock); 1374 1375 return (0); 1376 } 1377 1378 static int 1379 reset_sleepq_prof_stats(SYSCTL_HANDLER_ARGS) 1380 { 1381 int error, v; 1382 1383 v = 0; 1384 error = sysctl_handle_int(oidp, &v, 0, req); 1385 if (error) 1386 return (error); 1387 if (req->newptr == NULL) 1388 return (error); 1389 if (v == 0) 1390 return (0); 1391 sleepq_prof_reset(); 1392 1393 return (0); 1394 } 1395 1396 static int 1397 dump_sleepq_prof_stats(SYSCTL_HANDLER_ARGS) 1398 { 1399 struct sleepq_prof *sp; 1400 struct sbuf *sb; 1401 int enabled; 1402 int error; 1403 int i; 1404 1405 error = sysctl_wire_old_buffer(req, 0); 1406 if (error != 0) 1407 return (error); 1408 sb = sbuf_new_for_sysctl(NULL, NULL, SLEEPQ_SBUFSIZE, req); 1409 sbuf_printf(sb, "\nwmesg\tcount\n"); 1410 enabled = prof_enabled; 1411 mtx_lock_spin(&sleepq_prof_lock); 1412 prof_enabled = 0; 1413 mtx_unlock_spin(&sleepq_prof_lock); 1414 for (i = 0; i < SC_TABLESIZE; i++) { 1415 LIST_FOREACH(sp, &sleepq_hash[i], sp_link) { 1416 sbuf_printf(sb, "%s\t%ld\n", 1417 sp->sp_wmesg, sp->sp_count); 1418 } 1419 } 1420 mtx_lock_spin(&sleepq_prof_lock); 1421 prof_enabled = enabled; 1422 mtx_unlock_spin(&sleepq_prof_lock); 1423 1424 error = sbuf_finish(sb); 1425 sbuf_delete(sb); 1426 return (error); 1427 } 1428 1429 SYSCTL_PROC(_debug_sleepq, OID_AUTO, stats, CTLTYPE_STRING | CTLFLAG_RD, 1430 NULL, 0, dump_sleepq_prof_stats, "A", "Sleepqueue profiling statistics"); 1431 SYSCTL_PROC(_debug_sleepq, OID_AUTO, reset, CTLTYPE_INT | CTLFLAG_RW, 1432 NULL, 0, reset_sleepq_prof_stats, "I", 1433 "Reset sleepqueue profiling statistics"); 1434 SYSCTL_PROC(_debug_sleepq, OID_AUTO, enable, CTLTYPE_INT | CTLFLAG_RW, 1435 NULL, 0, enable_sleepq_prof, "I", "Enable sleepqueue profiling"); 1436 #endif 1437 1438 #ifdef DDB 1439 DB_SHOW_COMMAND(sleepq, db_show_sleepqueue) 1440 { 1441 struct sleepqueue_chain *sc; 1442 struct sleepqueue *sq; 1443 #ifdef INVARIANTS 1444 struct lock_object *lock; 1445 #endif 1446 struct thread *td; 1447 void *wchan; 1448 int i; 1449 1450 if (!have_addr) 1451 return; 1452 1453 /* 1454 * First, see if there is an active sleep queue for the wait channel 1455 * indicated by the address. 1456 */ 1457 wchan = (void *)addr; 1458 sc = SC_LOOKUP(wchan); 1459 LIST_FOREACH(sq, &sc->sc_queues, sq_hash) 1460 if (sq->sq_wchan == wchan) 1461 goto found; 1462 1463 /* 1464 * Second, see if there is an active sleep queue at the address 1465 * indicated. 1466 */ 1467 for (i = 0; i < SC_TABLESIZE; i++) 1468 LIST_FOREACH(sq, &sleepq_chains[i].sc_queues, sq_hash) { 1469 if (sq == (struct sleepqueue *)addr) 1470 goto found; 1471 } 1472 1473 db_printf("Unable to locate a sleep queue via %p\n", (void *)addr); 1474 return; 1475 found: 1476 db_printf("Wait channel: %p\n", sq->sq_wchan); 1477 db_printf("Queue type: %d\n", sq->sq_type); 1478 #ifdef INVARIANTS 1479 if (sq->sq_lock) { 1480 lock = sq->sq_lock; 1481 db_printf("Associated Interlock: %p - (%s) %s\n", lock, 1482 LOCK_CLASS(lock)->lc_name, lock->lo_name); 1483 } 1484 #endif 1485 db_printf("Blocked threads:\n"); 1486 for (i = 0; i < NR_SLEEPQS; i++) { 1487 db_printf("\nQueue[%d]:\n", i); 1488 if (TAILQ_EMPTY(&sq->sq_blocked[i])) 1489 db_printf("\tempty\n"); 1490 else 1491 TAILQ_FOREACH(td, &sq->sq_blocked[i], 1492 td_slpq) { 1493 db_printf("\t%p (tid %d, pid %d, \"%s\")\n", td, 1494 td->td_tid, td->td_proc->p_pid, 1495 td->td_name); 1496 } 1497 db_printf("(expected: %u)\n", sq->sq_blockedcnt[i]); 1498 } 1499 } 1500 1501 /* Alias 'show sleepqueue' to 'show sleepq'. */ 1502 DB_SHOW_ALIAS(sleepqueue, db_show_sleepqueue); 1503 #endif 1504