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 /* 97 * Constants for the hash table of sleep queue chains. 98 * SC_TABLESIZE must be a power of two for SC_MASK to work properly. 99 */ 100 #ifndef SC_TABLESIZE 101 #define SC_TABLESIZE 256 102 #endif 103 CTASSERT(powerof2(SC_TABLESIZE)); 104 #define SC_MASK (SC_TABLESIZE - 1) 105 #define SC_SHIFT 8 106 #define SC_HASH(wc) ((((uintptr_t)(wc) >> SC_SHIFT) ^ (uintptr_t)(wc)) & \ 107 SC_MASK) 108 #define SC_LOOKUP(wc) &sleepq_chains[SC_HASH(wc)] 109 #define NR_SLEEPQS 2 110 /* 111 * There are two different lists of sleep queues. Both lists are connected 112 * via the sq_hash entries. The first list is the sleep queue chain list 113 * that a sleep queue is on when it is attached to a wait channel. The 114 * second list is the free list hung off of a sleep queue that is attached 115 * to a wait channel. 116 * 117 * Each sleep queue also contains the wait channel it is attached to, the 118 * list of threads blocked on that wait channel, flags specific to the 119 * wait channel, and the lock used to synchronize with a wait channel. 120 * The flags are used to catch mismatches between the various consumers 121 * of the sleep queue API (e.g. sleep/wakeup and condition variables). 122 * The lock pointer is only used when invariants are enabled for various 123 * debugging checks. 124 * 125 * Locking key: 126 * c - sleep queue chain lock 127 */ 128 struct sleepqueue { 129 struct threadqueue sq_blocked[NR_SLEEPQS]; /* (c) Blocked threads. */ 130 u_int sq_blockedcnt[NR_SLEEPQS]; /* (c) N. of blocked threads. */ 131 LIST_ENTRY(sleepqueue) sq_hash; /* (c) Chain and free list. */ 132 LIST_HEAD(, sleepqueue) sq_free; /* (c) Free queues. */ 133 const void *sq_wchan; /* (c) Wait channel. */ 134 int sq_type; /* (c) Queue type. */ 135 #ifdef INVARIANTS 136 struct lock_object *sq_lock; /* (c) Associated lock. */ 137 #endif 138 }; 139 140 struct sleepqueue_chain { 141 LIST_HEAD(, sleepqueue) sc_queues; /* List of sleep queues. */ 142 struct mtx sc_lock; /* Spin lock for this chain. */ 143 #ifdef SLEEPQUEUE_PROFILING 144 u_int sc_depth; /* Length of sc_queues. */ 145 u_int sc_max_depth; /* Max length of sc_queues. */ 146 #endif 147 } __aligned(CACHE_LINE_SIZE); 148 149 #ifdef SLEEPQUEUE_PROFILING 150 u_int sleepq_max_depth; 151 static SYSCTL_NODE(_debug, OID_AUTO, sleepq, CTLFLAG_RD, 0, "sleepq profiling"); 152 static SYSCTL_NODE(_debug_sleepq, OID_AUTO, chains, CTLFLAG_RD, 0, 153 "sleepq chain stats"); 154 SYSCTL_UINT(_debug_sleepq, OID_AUTO, max_depth, CTLFLAG_RD, &sleepq_max_depth, 155 0, "maxmimum depth achieved of a single chain"); 156 157 static void sleepq_profile(const char *wmesg); 158 static int prof_enabled; 159 #endif 160 static struct sleepqueue_chain sleepq_chains[SC_TABLESIZE]; 161 static uma_zone_t sleepq_zone; 162 163 /* 164 * Prototypes for non-exported routines. 165 */ 166 static int sleepq_catch_signals(const void *wchan, int pri); 167 static inline int sleepq_check_signals(void); 168 static inline int sleepq_check_timeout(void); 169 #ifdef INVARIANTS 170 static void sleepq_dtor(void *mem, int size, void *arg); 171 #endif 172 static int sleepq_init(void *mem, int size, int flags); 173 static int sleepq_resume_thread(struct sleepqueue *sq, struct thread *td, 174 int pri, int srqflags); 175 static void sleepq_remove_thread(struct sleepqueue *sq, struct thread *td); 176 static void sleepq_switch(const void *wchan, int pri); 177 static void sleepq_timeout(void *arg); 178 179 SDT_PROBE_DECLARE(sched, , , sleep); 180 SDT_PROBE_DECLARE(sched, , , wakeup); 181 182 /* 183 * Initialize SLEEPQUEUE_PROFILING specific sysctl nodes. 184 * Note that it must happen after sleepinit() has been fully executed, so 185 * it must happen after SI_SUB_KMEM SYSINIT() subsystem setup. 186 */ 187 #ifdef SLEEPQUEUE_PROFILING 188 static void 189 init_sleepqueue_profiling(void) 190 { 191 char chain_name[10]; 192 struct sysctl_oid *chain_oid; 193 u_int i; 194 195 for (i = 0; i < SC_TABLESIZE; i++) { 196 snprintf(chain_name, sizeof(chain_name), "%u", i); 197 chain_oid = SYSCTL_ADD_NODE(NULL, 198 SYSCTL_STATIC_CHILDREN(_debug_sleepq_chains), OID_AUTO, 199 chain_name, CTLFLAG_RD, NULL, "sleepq chain stats"); 200 SYSCTL_ADD_UINT(NULL, SYSCTL_CHILDREN(chain_oid), OID_AUTO, 201 "depth", CTLFLAG_RD, &sleepq_chains[i].sc_depth, 0, NULL); 202 SYSCTL_ADD_UINT(NULL, SYSCTL_CHILDREN(chain_oid), OID_AUTO, 203 "max_depth", CTLFLAG_RD, &sleepq_chains[i].sc_max_depth, 0, 204 NULL); 205 } 206 } 207 208 SYSINIT(sleepqueue_profiling, SI_SUB_LOCK, SI_ORDER_ANY, 209 init_sleepqueue_profiling, NULL); 210 #endif 211 212 /* 213 * Early initialization of sleep queues that is called from the sleepinit() 214 * SYSINIT. 215 */ 216 void 217 init_sleepqueues(void) 218 { 219 int i; 220 221 for (i = 0; i < SC_TABLESIZE; i++) { 222 LIST_INIT(&sleepq_chains[i].sc_queues); 223 mtx_init(&sleepq_chains[i].sc_lock, "sleepq chain", NULL, 224 MTX_SPIN); 225 } 226 sleepq_zone = uma_zcreate("SLEEPQUEUE", sizeof(struct sleepqueue), 227 #ifdef INVARIANTS 228 NULL, sleepq_dtor, sleepq_init, NULL, UMA_ALIGN_CACHE, 0); 229 #else 230 NULL, NULL, sleepq_init, NULL, UMA_ALIGN_CACHE, 0); 231 #endif 232 233 thread0.td_sleepqueue = sleepq_alloc(); 234 } 235 236 /* 237 * Get a sleep queue for a new thread. 238 */ 239 struct sleepqueue * 240 sleepq_alloc(void) 241 { 242 243 return (uma_zalloc(sleepq_zone, M_WAITOK)); 244 } 245 246 /* 247 * Free a sleep queue when a thread is destroyed. 248 */ 249 void 250 sleepq_free(struct sleepqueue *sq) 251 { 252 253 uma_zfree(sleepq_zone, sq); 254 } 255 256 /* 257 * Lock the sleep queue chain associated with the specified wait channel. 258 */ 259 void 260 sleepq_lock(const void *wchan) 261 { 262 struct sleepqueue_chain *sc; 263 264 sc = SC_LOOKUP(wchan); 265 mtx_lock_spin(&sc->sc_lock); 266 } 267 268 /* 269 * Look up the sleep queue associated with a given wait channel in the hash 270 * table locking the associated sleep queue chain. If no queue is found in 271 * the table, NULL is returned. 272 */ 273 struct sleepqueue * 274 sleepq_lookup(const void *wchan) 275 { 276 struct sleepqueue_chain *sc; 277 struct sleepqueue *sq; 278 279 KASSERT(wchan != NULL, ("%s: invalid NULL wait channel", __func__)); 280 sc = SC_LOOKUP(wchan); 281 mtx_assert(&sc->sc_lock, MA_OWNED); 282 LIST_FOREACH(sq, &sc->sc_queues, sq_hash) 283 if (sq->sq_wchan == wchan) 284 return (sq); 285 return (NULL); 286 } 287 288 /* 289 * Unlock the sleep queue chain associated with a given wait channel. 290 */ 291 void 292 sleepq_release(const void *wchan) 293 { 294 struct sleepqueue_chain *sc; 295 296 sc = SC_LOOKUP(wchan); 297 mtx_unlock_spin(&sc->sc_lock); 298 } 299 300 /* 301 * Places the current thread on the sleep queue for the specified wait 302 * channel. If INVARIANTS is enabled, then it associates the passed in 303 * lock with the sleepq to make sure it is held when that sleep queue is 304 * woken up. 305 */ 306 void 307 sleepq_add(const void *wchan, struct lock_object *lock, const char *wmesg, 308 int flags, int queue) 309 { 310 struct sleepqueue_chain *sc; 311 struct sleepqueue *sq; 312 struct thread *td; 313 314 td = curthread; 315 sc = SC_LOOKUP(wchan); 316 mtx_assert(&sc->sc_lock, MA_OWNED); 317 MPASS(td->td_sleepqueue != NULL); 318 MPASS(wchan != NULL); 319 MPASS((queue >= 0) && (queue < NR_SLEEPQS)); 320 321 /* If this thread is not allowed to sleep, die a horrible death. */ 322 if (__predict_false(!THREAD_CAN_SLEEP())) { 323 #ifdef EPOCH_TRACE 324 epoch_trace_list(curthread); 325 #endif 326 KASSERT(1, 327 ("%s: td %p to sleep on wchan %p with sleeping prohibited", 328 __func__, td, wchan)); 329 } 330 331 /* Look up the sleep queue associated with the wait channel 'wchan'. */ 332 sq = sleepq_lookup(wchan); 333 334 /* 335 * If the wait channel does not already have a sleep queue, use 336 * this thread's sleep queue. Otherwise, insert the current thread 337 * into the sleep queue already in use by this wait channel. 338 */ 339 if (sq == NULL) { 340 #ifdef INVARIANTS 341 int i; 342 343 sq = td->td_sleepqueue; 344 for (i = 0; i < NR_SLEEPQS; i++) { 345 KASSERT(TAILQ_EMPTY(&sq->sq_blocked[i]), 346 ("thread's sleep queue %d is not empty", i)); 347 KASSERT(sq->sq_blockedcnt[i] == 0, 348 ("thread's sleep queue %d count mismatches", i)); 349 } 350 KASSERT(LIST_EMPTY(&sq->sq_free), 351 ("thread's sleep queue has a non-empty free list")); 352 KASSERT(sq->sq_wchan == NULL, ("stale sq_wchan pointer")); 353 sq->sq_lock = lock; 354 #endif 355 #ifdef SLEEPQUEUE_PROFILING 356 sc->sc_depth++; 357 if (sc->sc_depth > sc->sc_max_depth) { 358 sc->sc_max_depth = sc->sc_depth; 359 if (sc->sc_max_depth > sleepq_max_depth) 360 sleepq_max_depth = sc->sc_max_depth; 361 } 362 #endif 363 sq = td->td_sleepqueue; 364 LIST_INSERT_HEAD(&sc->sc_queues, sq, sq_hash); 365 sq->sq_wchan = wchan; 366 sq->sq_type = flags & SLEEPQ_TYPE; 367 } else { 368 MPASS(wchan == sq->sq_wchan); 369 MPASS(lock == sq->sq_lock); 370 MPASS((flags & SLEEPQ_TYPE) == sq->sq_type); 371 LIST_INSERT_HEAD(&sq->sq_free, td->td_sleepqueue, sq_hash); 372 } 373 thread_lock(td); 374 TAILQ_INSERT_TAIL(&sq->sq_blocked[queue], td, td_slpq); 375 sq->sq_blockedcnt[queue]++; 376 td->td_sleepqueue = NULL; 377 td->td_sqqueue = queue; 378 td->td_wchan = wchan; 379 td->td_wmesg = wmesg; 380 if (flags & SLEEPQ_INTERRUPTIBLE) { 381 td->td_intrval = 0; 382 td->td_flags |= TDF_SINTR; 383 } 384 td->td_flags &= ~TDF_TIMEOUT; 385 thread_unlock(td); 386 } 387 388 /* 389 * Sets a timeout that will remove the current thread from the specified 390 * sleep queue after timo ticks if the thread has not already been awakened. 391 */ 392 void 393 sleepq_set_timeout_sbt(const void *wchan, sbintime_t sbt, sbintime_t pr, 394 int flags) 395 { 396 struct sleepqueue_chain *sc __unused; 397 struct thread *td; 398 sbintime_t pr1; 399 400 td = curthread; 401 sc = SC_LOOKUP(wchan); 402 mtx_assert(&sc->sc_lock, MA_OWNED); 403 MPASS(TD_ON_SLEEPQ(td)); 404 MPASS(td->td_sleepqueue == NULL); 405 MPASS(wchan != NULL); 406 if (cold && td == &thread0) 407 panic("timed sleep before timers are working"); 408 KASSERT(td->td_sleeptimo == 0, ("td %d %p td_sleeptimo %jx", 409 td->td_tid, td, (uintmax_t)td->td_sleeptimo)); 410 thread_lock(td); 411 callout_when(sbt, pr, flags, &td->td_sleeptimo, &pr1); 412 thread_unlock(td); 413 callout_reset_sbt_on(&td->td_slpcallout, td->td_sleeptimo, pr1, 414 sleepq_timeout, td, PCPU_GET(cpuid), flags | C_PRECALC | 415 C_DIRECT_EXEC); 416 } 417 418 /* 419 * Return the number of actual sleepers for the specified queue. 420 */ 421 u_int 422 sleepq_sleepcnt(const void *wchan, int queue) 423 { 424 struct sleepqueue *sq; 425 426 KASSERT(wchan != NULL, ("%s: invalid NULL wait channel", __func__)); 427 MPASS((queue >= 0) && (queue < NR_SLEEPQS)); 428 sq = sleepq_lookup(wchan); 429 if (sq == NULL) 430 return (0); 431 return (sq->sq_blockedcnt[queue]); 432 } 433 434 /* 435 * Marks the pending sleep of the current thread as interruptible and 436 * makes an initial check for pending signals before putting a thread 437 * to sleep. Enters and exits with the thread lock held. Thread lock 438 * may have transitioned from the sleepq lock to a run lock. 439 */ 440 static int 441 sleepq_catch_signals(const void *wchan, int pri) 442 { 443 struct sleepqueue_chain *sc; 444 struct sleepqueue *sq; 445 struct thread *td; 446 struct proc *p; 447 struct sigacts *ps; 448 int sig, ret; 449 450 ret = 0; 451 td = curthread; 452 p = curproc; 453 sc = SC_LOOKUP(wchan); 454 mtx_assert(&sc->sc_lock, MA_OWNED); 455 MPASS(wchan != NULL); 456 if ((td->td_pflags & TDP_WAKEUP) != 0) { 457 td->td_pflags &= ~TDP_WAKEUP; 458 ret = EINTR; 459 thread_lock(td); 460 goto out; 461 } 462 463 /* 464 * See if there are any pending signals or suspension requests for this 465 * thread. If not, we can switch immediately. 466 */ 467 thread_lock(td); 468 if ((td->td_flags & (TDF_NEEDSIGCHK | TDF_NEEDSUSPCHK)) != 0) { 469 thread_unlock(td); 470 mtx_unlock_spin(&sc->sc_lock); 471 CTR3(KTR_PROC, "sleepq catching signals: thread %p (pid %ld, %s)", 472 (void *)td, (long)p->p_pid, td->td_name); 473 PROC_LOCK(p); 474 /* 475 * Check for suspension first. Checking for signals and then 476 * suspending could result in a missed signal, since a signal 477 * can be delivered while this thread is suspended. 478 */ 479 if ((td->td_flags & TDF_NEEDSUSPCHK) != 0) { 480 ret = thread_suspend_check(1); 481 MPASS(ret == 0 || ret == EINTR || ret == ERESTART); 482 if (ret != 0) { 483 PROC_UNLOCK(p); 484 mtx_lock_spin(&sc->sc_lock); 485 thread_lock(td); 486 goto out; 487 } 488 } 489 if ((td->td_flags & TDF_NEEDSIGCHK) != 0) { 490 ps = p->p_sigacts; 491 mtx_lock(&ps->ps_mtx); 492 sig = cursig(td); 493 if (sig == -1) { 494 mtx_unlock(&ps->ps_mtx); 495 KASSERT((td->td_flags & TDF_SBDRY) != 0, 496 ("lost TDF_SBDRY")); 497 KASSERT(TD_SBDRY_INTR(td), 498 ("lost TDF_SERESTART of TDF_SEINTR")); 499 KASSERT((td->td_flags & 500 (TDF_SEINTR | TDF_SERESTART)) != 501 (TDF_SEINTR | TDF_SERESTART), 502 ("both TDF_SEINTR and TDF_SERESTART")); 503 ret = TD_SBDRY_ERRNO(td); 504 } else if (sig != 0) { 505 ret = SIGISMEMBER(ps->ps_sigintr, sig) ? 506 EINTR : ERESTART; 507 mtx_unlock(&ps->ps_mtx); 508 } else { 509 mtx_unlock(&ps->ps_mtx); 510 } 511 512 /* 513 * Do not go into sleep if this thread was the 514 * ptrace(2) attach leader. cursig() consumed 515 * SIGSTOP from PT_ATTACH, but we usually act 516 * on the signal by interrupting sleep, and 517 * should do that here as well. 518 */ 519 if ((td->td_dbgflags & TDB_FSTP) != 0) { 520 if (ret == 0) 521 ret = EINTR; 522 td->td_dbgflags &= ~TDB_FSTP; 523 } 524 } 525 /* 526 * Lock the per-process spinlock prior to dropping the PROC_LOCK 527 * to avoid a signal delivery race. PROC_LOCK, PROC_SLOCK, and 528 * thread_lock() are currently held in tdsendsignal(). 529 */ 530 PROC_SLOCK(p); 531 mtx_lock_spin(&sc->sc_lock); 532 PROC_UNLOCK(p); 533 thread_lock(td); 534 PROC_SUNLOCK(p); 535 } 536 if (ret == 0) { 537 sleepq_switch(wchan, pri); 538 return (0); 539 } 540 out: 541 /* 542 * There were pending signals and this thread is still 543 * on the sleep queue, remove it from the sleep queue. 544 */ 545 if (TD_ON_SLEEPQ(td)) { 546 sq = sleepq_lookup(wchan); 547 sleepq_remove_thread(sq, td); 548 } 549 MPASS(td->td_lock != &sc->sc_lock); 550 mtx_unlock_spin(&sc->sc_lock); 551 thread_unlock(td); 552 553 return (ret); 554 } 555 556 /* 557 * Switches to another thread if we are still asleep on a sleep queue. 558 * Returns with thread lock. 559 */ 560 static void 561 sleepq_switch(const void *wchan, int pri) 562 { 563 struct sleepqueue_chain *sc; 564 struct sleepqueue *sq; 565 struct thread *td; 566 bool rtc_changed; 567 568 td = curthread; 569 sc = SC_LOOKUP(wchan); 570 mtx_assert(&sc->sc_lock, MA_OWNED); 571 THREAD_LOCK_ASSERT(td, MA_OWNED); 572 573 /* 574 * If we have a sleep queue, then we've already been woken up, so 575 * just return. 576 */ 577 if (td->td_sleepqueue != NULL) { 578 mtx_unlock_spin(&sc->sc_lock); 579 thread_unlock(td); 580 return; 581 } 582 583 /* 584 * If TDF_TIMEOUT is set, then our sleep has been timed out 585 * already but we are still on the sleep queue, so dequeue the 586 * thread and return. 587 * 588 * Do the same if the real-time clock has been adjusted since this 589 * thread calculated its timeout based on that clock. This handles 590 * the following race: 591 * - The Ts thread needs to sleep until an absolute real-clock time. 592 * It copies the global rtc_generation into curthread->td_rtcgen, 593 * reads the RTC, and calculates a sleep duration based on that time. 594 * See umtxq_sleep() for an example. 595 * - The Tc thread adjusts the RTC, bumps rtc_generation, and wakes 596 * threads that are sleeping until an absolute real-clock time. 597 * See tc_setclock() and the POSIX specification of clock_settime(). 598 * - Ts reaches the code below. It holds the sleepqueue chain lock, 599 * so Tc has finished waking, so this thread must test td_rtcgen. 600 * (The declaration of td_rtcgen refers to this comment.) 601 */ 602 rtc_changed = td->td_rtcgen != 0 && td->td_rtcgen != rtc_generation; 603 if ((td->td_flags & TDF_TIMEOUT) || rtc_changed) { 604 if (rtc_changed) { 605 td->td_rtcgen = 0; 606 } 607 MPASS(TD_ON_SLEEPQ(td)); 608 sq = sleepq_lookup(wchan); 609 sleepq_remove_thread(sq, td); 610 mtx_unlock_spin(&sc->sc_lock); 611 thread_unlock(td); 612 return; 613 } 614 #ifdef SLEEPQUEUE_PROFILING 615 if (prof_enabled) 616 sleepq_profile(td->td_wmesg); 617 #endif 618 MPASS(td->td_sleepqueue == NULL); 619 sched_sleep(td, pri); 620 thread_lock_set(td, &sc->sc_lock); 621 SDT_PROBE0(sched, , , sleep); 622 TD_SET_SLEEPING(td); 623 mi_switch(SW_VOL | SWT_SLEEPQ); 624 KASSERT(TD_IS_RUNNING(td), ("running but not TDS_RUNNING")); 625 CTR3(KTR_PROC, "sleepq resume: thread %p (pid %ld, %s)", 626 (void *)td, (long)td->td_proc->p_pid, (void *)td->td_name); 627 } 628 629 /* 630 * Check to see if we timed out. 631 */ 632 static inline int 633 sleepq_check_timeout(void) 634 { 635 struct thread *td; 636 int res; 637 638 res = 0; 639 td = curthread; 640 if (td->td_sleeptimo != 0) { 641 if (td->td_sleeptimo <= sbinuptime()) 642 res = EWOULDBLOCK; 643 td->td_sleeptimo = 0; 644 } 645 return (res); 646 } 647 648 /* 649 * Check to see if we were awoken by a signal. 650 */ 651 static inline int 652 sleepq_check_signals(void) 653 { 654 struct thread *td; 655 656 td = curthread; 657 KASSERT((td->td_flags & TDF_SINTR) == 0, 658 ("thread %p still in interruptible sleep?", td)); 659 660 return (td->td_intrval); 661 } 662 663 /* 664 * Block the current thread until it is awakened from its sleep queue. 665 */ 666 void 667 sleepq_wait(const void *wchan, int pri) 668 { 669 struct thread *td; 670 671 td = curthread; 672 MPASS(!(td->td_flags & TDF_SINTR)); 673 thread_lock(td); 674 sleepq_switch(wchan, pri); 675 } 676 677 /* 678 * Block the current thread until it is awakened from its sleep queue 679 * or it is interrupted by a signal. 680 */ 681 int 682 sleepq_wait_sig(const void *wchan, int pri) 683 { 684 int rcatch; 685 686 rcatch = sleepq_catch_signals(wchan, pri); 687 if (rcatch) 688 return (rcatch); 689 return (sleepq_check_signals()); 690 } 691 692 /* 693 * Block the current thread until it is awakened from its sleep queue 694 * or it times out while waiting. 695 */ 696 int 697 sleepq_timedwait(const void *wchan, int pri) 698 { 699 struct thread *td; 700 701 td = curthread; 702 MPASS(!(td->td_flags & TDF_SINTR)); 703 704 thread_lock(td); 705 sleepq_switch(wchan, pri); 706 707 return (sleepq_check_timeout()); 708 } 709 710 /* 711 * Block the current thread until it is awakened from its sleep queue, 712 * it is interrupted by a signal, or it times out waiting to be awakened. 713 */ 714 int 715 sleepq_timedwait_sig(const void *wchan, int pri) 716 { 717 int rcatch, rvalt, rvals; 718 719 rcatch = sleepq_catch_signals(wchan, pri); 720 /* We must always call check_timeout() to clear sleeptimo. */ 721 rvalt = sleepq_check_timeout(); 722 rvals = sleepq_check_signals(); 723 if (rcatch) 724 return (rcatch); 725 if (rvals) 726 return (rvals); 727 return (rvalt); 728 } 729 730 /* 731 * Returns the type of sleepqueue given a waitchannel. 732 */ 733 int 734 sleepq_type(const void *wchan) 735 { 736 struct sleepqueue *sq; 737 int type; 738 739 MPASS(wchan != NULL); 740 741 sq = sleepq_lookup(wchan); 742 if (sq == NULL) 743 return (-1); 744 type = sq->sq_type; 745 746 return (type); 747 } 748 749 /* 750 * Removes a thread from a sleep queue and makes it 751 * runnable. 752 * 753 * Requires the sc chain locked on entry. If SRQ_HOLD is specified it will 754 * be locked on return. Returns without the thread lock held. 755 */ 756 static int 757 sleepq_resume_thread(struct sleepqueue *sq, struct thread *td, int pri, 758 int srqflags) 759 { 760 struct sleepqueue_chain *sc; 761 bool drop; 762 763 MPASS(td != NULL); 764 MPASS(sq->sq_wchan != NULL); 765 MPASS(td->td_wchan == sq->sq_wchan); 766 767 sc = SC_LOOKUP(sq->sq_wchan); 768 mtx_assert(&sc->sc_lock, MA_OWNED); 769 770 /* 771 * Avoid recursing on the chain lock. If the locks don't match we 772 * need to acquire the thread lock which setrunnable will drop for 773 * us. In this case we need to drop the chain lock afterwards. 774 * 775 * There is no race that will make td_lock equal to sc_lock because 776 * we hold sc_lock. 777 */ 778 drop = false; 779 if (!TD_IS_SLEEPING(td)) { 780 thread_lock(td); 781 drop = true; 782 } else 783 thread_lock_block_wait(td); 784 785 /* Remove thread from the sleepq. */ 786 sleepq_remove_thread(sq, td); 787 788 /* If we're done with the sleepqueue release it. */ 789 if ((srqflags & SRQ_HOLD) == 0 && drop) 790 mtx_unlock_spin(&sc->sc_lock); 791 792 /* Adjust priority if requested. */ 793 MPASS(pri == 0 || (pri >= PRI_MIN && pri <= PRI_MAX)); 794 if (pri != 0 && td->td_priority > pri && 795 PRI_BASE(td->td_pri_class) == PRI_TIMESHARE) 796 sched_prio(td, pri); 797 798 /* 799 * Note that thread td might not be sleeping if it is running 800 * sleepq_catch_signals() on another CPU or is blocked on its 801 * proc lock to check signals. There's no need to mark the 802 * thread runnable in that case. 803 */ 804 if (TD_IS_SLEEPING(td)) { 805 MPASS(!drop); 806 TD_CLR_SLEEPING(td); 807 return (setrunnable(td, srqflags)); 808 } 809 MPASS(drop); 810 thread_unlock(td); 811 812 return (0); 813 } 814 815 static void 816 sleepq_remove_thread(struct sleepqueue *sq, struct thread *td) 817 { 818 struct sleepqueue_chain *sc __unused; 819 820 MPASS(td != NULL); 821 MPASS(sq->sq_wchan != NULL); 822 MPASS(td->td_wchan == sq->sq_wchan); 823 MPASS(td->td_sqqueue < NR_SLEEPQS && td->td_sqqueue >= 0); 824 THREAD_LOCK_ASSERT(td, MA_OWNED); 825 sc = SC_LOOKUP(sq->sq_wchan); 826 mtx_assert(&sc->sc_lock, MA_OWNED); 827 828 SDT_PROBE2(sched, , , wakeup, td, td->td_proc); 829 830 /* Remove the thread from the queue. */ 831 sq->sq_blockedcnt[td->td_sqqueue]--; 832 TAILQ_REMOVE(&sq->sq_blocked[td->td_sqqueue], td, td_slpq); 833 834 /* 835 * Get a sleep queue for this thread. If this is the last waiter, 836 * use the queue itself and take it out of the chain, otherwise, 837 * remove a queue from the free list. 838 */ 839 if (LIST_EMPTY(&sq->sq_free)) { 840 td->td_sleepqueue = sq; 841 #ifdef INVARIANTS 842 sq->sq_wchan = NULL; 843 #endif 844 #ifdef SLEEPQUEUE_PROFILING 845 sc->sc_depth--; 846 #endif 847 } else 848 td->td_sleepqueue = LIST_FIRST(&sq->sq_free); 849 LIST_REMOVE(td->td_sleepqueue, sq_hash); 850 851 if ((td->td_flags & TDF_TIMEOUT) == 0 && td->td_sleeptimo != 0) 852 /* 853 * We ignore the situation where timeout subsystem was 854 * unable to stop our callout. The struct thread is 855 * type-stable, the callout will use the correct 856 * memory when running. The checks of the 857 * td_sleeptimo value in this function and in 858 * sleepq_timeout() ensure that the thread does not 859 * get spurious wakeups, even if the callout was reset 860 * or thread reused. 861 */ 862 callout_stop(&td->td_slpcallout); 863 864 td->td_wmesg = NULL; 865 td->td_wchan = NULL; 866 td->td_flags &= ~(TDF_SINTR | TDF_TIMEOUT); 867 868 CTR3(KTR_PROC, "sleepq_wakeup: thread %p (pid %ld, %s)", 869 (void *)td, (long)td->td_proc->p_pid, td->td_name); 870 } 871 872 #ifdef INVARIANTS 873 /* 874 * UMA zone item deallocator. 875 */ 876 static void 877 sleepq_dtor(void *mem, int size, void *arg) 878 { 879 struct sleepqueue *sq; 880 int i; 881 882 sq = mem; 883 for (i = 0; i < NR_SLEEPQS; i++) { 884 MPASS(TAILQ_EMPTY(&sq->sq_blocked[i])); 885 MPASS(sq->sq_blockedcnt[i] == 0); 886 } 887 } 888 #endif 889 890 /* 891 * UMA zone item initializer. 892 */ 893 static int 894 sleepq_init(void *mem, int size, int flags) 895 { 896 struct sleepqueue *sq; 897 int i; 898 899 bzero(mem, size); 900 sq = mem; 901 for (i = 0; i < NR_SLEEPQS; i++) { 902 TAILQ_INIT(&sq->sq_blocked[i]); 903 sq->sq_blockedcnt[i] = 0; 904 } 905 LIST_INIT(&sq->sq_free); 906 return (0); 907 } 908 909 /* 910 * Find thread sleeping on a wait channel and resume it. 911 */ 912 int 913 sleepq_signal(const void *wchan, int flags, int pri, int queue) 914 { 915 struct sleepqueue_chain *sc; 916 struct sleepqueue *sq; 917 struct threadqueue *head; 918 struct thread *td, *besttd; 919 int wakeup_swapper; 920 921 CTR2(KTR_PROC, "sleepq_signal(%p, %d)", wchan, flags); 922 KASSERT(wchan != NULL, ("%s: invalid NULL wait channel", __func__)); 923 MPASS((queue >= 0) && (queue < NR_SLEEPQS)); 924 sq = sleepq_lookup(wchan); 925 if (sq == NULL) 926 return (0); 927 KASSERT(sq->sq_type == (flags & SLEEPQ_TYPE), 928 ("%s: mismatch between sleep/wakeup and cv_*", __func__)); 929 930 head = &sq->sq_blocked[queue]; 931 if (flags & SLEEPQ_UNFAIR) { 932 /* 933 * Find the most recently sleeping thread, but try to 934 * skip threads still in process of context switch to 935 * avoid spinning on the thread lock. 936 */ 937 sc = SC_LOOKUP(wchan); 938 besttd = TAILQ_LAST_FAST(head, thread, td_slpq); 939 while (besttd->td_lock != &sc->sc_lock) { 940 td = TAILQ_PREV_FAST(besttd, head, thread, td_slpq); 941 if (td == NULL) 942 break; 943 besttd = td; 944 } 945 } else { 946 /* 947 * Find the highest priority thread on the queue. If there 948 * is a tie, use the thread that first appears in the queue 949 * as it has been sleeping the longest since threads are 950 * always added to the tail of sleep queues. 951 */ 952 besttd = td = TAILQ_FIRST(head); 953 while ((td = TAILQ_NEXT(td, td_slpq)) != NULL) { 954 if (td->td_priority < besttd->td_priority) 955 besttd = td; 956 } 957 } 958 MPASS(besttd != NULL); 959 wakeup_swapper = sleepq_resume_thread(sq, besttd, pri, SRQ_HOLD); 960 return (wakeup_swapper); 961 } 962 963 static bool 964 match_any(struct thread *td __unused) 965 { 966 967 return (true); 968 } 969 970 /* 971 * Resume all threads sleeping on a specified wait channel. 972 */ 973 int 974 sleepq_broadcast(const void *wchan, int flags, int pri, int queue) 975 { 976 struct sleepqueue *sq; 977 978 CTR2(KTR_PROC, "sleepq_broadcast(%p, %d)", wchan, flags); 979 KASSERT(wchan != NULL, ("%s: invalid NULL wait channel", __func__)); 980 MPASS((queue >= 0) && (queue < NR_SLEEPQS)); 981 sq = sleepq_lookup(wchan); 982 if (sq == NULL) 983 return (0); 984 KASSERT(sq->sq_type == (flags & SLEEPQ_TYPE), 985 ("%s: mismatch between sleep/wakeup and cv_*", __func__)); 986 987 return (sleepq_remove_matching(sq, queue, match_any, pri)); 988 } 989 990 /* 991 * Resume threads on the sleep queue that match the given predicate. 992 */ 993 int 994 sleepq_remove_matching(struct sleepqueue *sq, int queue, 995 bool (*matches)(struct thread *), int pri) 996 { 997 struct thread *td, *tdn; 998 int wakeup_swapper; 999 1000 /* 1001 * The last thread will be given ownership of sq and may 1002 * re-enqueue itself before sleepq_resume_thread() returns, 1003 * so we must cache the "next" queue item at the beginning 1004 * of the final iteration. 1005 */ 1006 wakeup_swapper = 0; 1007 TAILQ_FOREACH_SAFE(td, &sq->sq_blocked[queue], td_slpq, tdn) { 1008 if (matches(td)) 1009 wakeup_swapper |= sleepq_resume_thread(sq, td, pri, 1010 SRQ_HOLD); 1011 } 1012 1013 return (wakeup_swapper); 1014 } 1015 1016 /* 1017 * Time sleeping threads out. When the timeout expires, the thread is 1018 * removed from the sleep queue and made runnable if it is still asleep. 1019 */ 1020 static void 1021 sleepq_timeout(void *arg) 1022 { 1023 struct sleepqueue_chain *sc __unused; 1024 struct sleepqueue *sq; 1025 struct thread *td; 1026 const void *wchan; 1027 int wakeup_swapper; 1028 1029 td = arg; 1030 CTR3(KTR_PROC, "sleepq_timeout: thread %p (pid %ld, %s)", 1031 (void *)td, (long)td->td_proc->p_pid, (void *)td->td_name); 1032 1033 thread_lock(td); 1034 if (td->td_sleeptimo == 0 || td->td_sleeptimo > sbinuptime()) { 1035 /* 1036 * The thread does not want a timeout (yet). 1037 */ 1038 } else if (TD_IS_SLEEPING(td) && TD_ON_SLEEPQ(td)) { 1039 /* 1040 * See if the thread is asleep and get the wait 1041 * channel if it is. 1042 */ 1043 wchan = td->td_wchan; 1044 sc = SC_LOOKUP(wchan); 1045 THREAD_LOCKPTR_ASSERT(td, &sc->sc_lock); 1046 sq = sleepq_lookup(wchan); 1047 MPASS(sq != NULL); 1048 td->td_flags |= TDF_TIMEOUT; 1049 wakeup_swapper = sleepq_resume_thread(sq, td, 0, 0); 1050 if (wakeup_swapper) 1051 kick_proc0(); 1052 return; 1053 } else if (TD_ON_SLEEPQ(td)) { 1054 /* 1055 * If the thread is on the SLEEPQ but isn't sleeping 1056 * yet, it can either be on another CPU in between 1057 * sleepq_add() and one of the sleepq_*wait*() 1058 * routines or it can be in sleepq_catch_signals(). 1059 */ 1060 td->td_flags |= TDF_TIMEOUT; 1061 } 1062 thread_unlock(td); 1063 } 1064 1065 /* 1066 * Resumes a specific thread from the sleep queue associated with a specific 1067 * wait channel if it is on that queue. 1068 */ 1069 void 1070 sleepq_remove(struct thread *td, const void *wchan) 1071 { 1072 struct sleepqueue_chain *sc; 1073 struct sleepqueue *sq; 1074 int wakeup_swapper; 1075 1076 /* 1077 * Look up the sleep queue for this wait channel, then re-check 1078 * that the thread is asleep on that channel, if it is not, then 1079 * bail. 1080 */ 1081 MPASS(wchan != NULL); 1082 sc = SC_LOOKUP(wchan); 1083 mtx_lock_spin(&sc->sc_lock); 1084 /* 1085 * We can not lock the thread here as it may be sleeping on a 1086 * different sleepq. However, holding the sleepq lock for this 1087 * wchan can guarantee that we do not miss a wakeup for this 1088 * channel. The asserts below will catch any false positives. 1089 */ 1090 if (!TD_ON_SLEEPQ(td) || td->td_wchan != wchan) { 1091 mtx_unlock_spin(&sc->sc_lock); 1092 return; 1093 } 1094 1095 /* Thread is asleep on sleep queue sq, so wake it up. */ 1096 sq = sleepq_lookup(wchan); 1097 MPASS(sq != NULL); 1098 MPASS(td->td_wchan == wchan); 1099 wakeup_swapper = sleepq_resume_thread(sq, td, 0, 0); 1100 if (wakeup_swapper) 1101 kick_proc0(); 1102 } 1103 1104 /* 1105 * Abort a thread as if an interrupt had occurred. Only abort 1106 * interruptible waits (unfortunately it isn't safe to abort others). 1107 * 1108 * Requires thread lock on entry, releases on return. 1109 */ 1110 int 1111 sleepq_abort(struct thread *td, int intrval) 1112 { 1113 struct sleepqueue *sq; 1114 const void *wchan; 1115 1116 THREAD_LOCK_ASSERT(td, MA_OWNED); 1117 MPASS(TD_ON_SLEEPQ(td)); 1118 MPASS(td->td_flags & TDF_SINTR); 1119 MPASS(intrval == EINTR || intrval == ERESTART); 1120 1121 /* 1122 * If the TDF_TIMEOUT flag is set, just leave. A 1123 * timeout is scheduled anyhow. 1124 */ 1125 if (td->td_flags & TDF_TIMEOUT) { 1126 thread_unlock(td); 1127 return (0); 1128 } 1129 1130 CTR3(KTR_PROC, "sleepq_abort: thread %p (pid %ld, %s)", 1131 (void *)td, (long)td->td_proc->p_pid, (void *)td->td_name); 1132 td->td_intrval = intrval; 1133 1134 /* 1135 * If the thread has not slept yet it will find the signal in 1136 * sleepq_catch_signals() and call sleepq_resume_thread. Otherwise 1137 * we have to do it here. 1138 */ 1139 if (!TD_IS_SLEEPING(td)) { 1140 thread_unlock(td); 1141 return (0); 1142 } 1143 wchan = td->td_wchan; 1144 MPASS(wchan != NULL); 1145 sq = sleepq_lookup(wchan); 1146 MPASS(sq != NULL); 1147 1148 /* Thread is asleep on sleep queue sq, so wake it up. */ 1149 return (sleepq_resume_thread(sq, td, 0, 0)); 1150 } 1151 1152 void 1153 sleepq_chains_remove_matching(bool (*matches)(struct thread *)) 1154 { 1155 struct sleepqueue_chain *sc; 1156 struct sleepqueue *sq, *sq1; 1157 int i, wakeup_swapper; 1158 1159 wakeup_swapper = 0; 1160 for (sc = &sleepq_chains[0]; sc < sleepq_chains + SC_TABLESIZE; ++sc) { 1161 if (LIST_EMPTY(&sc->sc_queues)) { 1162 continue; 1163 } 1164 mtx_lock_spin(&sc->sc_lock); 1165 LIST_FOREACH_SAFE(sq, &sc->sc_queues, sq_hash, sq1) { 1166 for (i = 0; i < NR_SLEEPQS; ++i) { 1167 wakeup_swapper |= sleepq_remove_matching(sq, i, 1168 matches, 0); 1169 } 1170 } 1171 mtx_unlock_spin(&sc->sc_lock); 1172 } 1173 if (wakeup_swapper) { 1174 kick_proc0(); 1175 } 1176 } 1177 1178 /* 1179 * Prints the stacks of all threads presently sleeping on wchan/queue to 1180 * the sbuf sb. Sets count_stacks_printed to the number of stacks actually 1181 * printed. Typically, this will equal the number of threads sleeping on the 1182 * queue, but may be less if sb overflowed before all stacks were printed. 1183 */ 1184 #ifdef STACK 1185 int 1186 sleepq_sbuf_print_stacks(struct sbuf *sb, const void *wchan, int queue, 1187 int *count_stacks_printed) 1188 { 1189 struct thread *td, *td_next; 1190 struct sleepqueue *sq; 1191 struct stack **st; 1192 struct sbuf **td_infos; 1193 int i, stack_idx, error, stacks_to_allocate; 1194 bool finished; 1195 1196 error = 0; 1197 finished = false; 1198 1199 KASSERT(wchan != NULL, ("%s: invalid NULL wait channel", __func__)); 1200 MPASS((queue >= 0) && (queue < NR_SLEEPQS)); 1201 1202 stacks_to_allocate = 10; 1203 for (i = 0; i < 3 && !finished ; i++) { 1204 /* We cannot malloc while holding the queue's spinlock, so 1205 * we do our mallocs now, and hope it is enough. If it 1206 * isn't, we will free these, drop the lock, malloc more, 1207 * and try again, up to a point. After that point we will 1208 * give up and report ENOMEM. We also cannot write to sb 1209 * during this time since the client may have set the 1210 * SBUF_AUTOEXTEND flag on their sbuf, which could cause a 1211 * malloc as we print to it. So we defer actually printing 1212 * to sb until after we drop the spinlock. 1213 */ 1214 1215 /* Where we will store the stacks. */ 1216 st = malloc(sizeof(struct stack *) * stacks_to_allocate, 1217 M_TEMP, M_WAITOK); 1218 for (stack_idx = 0; stack_idx < stacks_to_allocate; 1219 stack_idx++) 1220 st[stack_idx] = stack_create(M_WAITOK); 1221 1222 /* Where we will store the td name, tid, etc. */ 1223 td_infos = malloc(sizeof(struct sbuf *) * stacks_to_allocate, 1224 M_TEMP, M_WAITOK); 1225 for (stack_idx = 0; stack_idx < stacks_to_allocate; 1226 stack_idx++) 1227 td_infos[stack_idx] = sbuf_new(NULL, NULL, 1228 MAXCOMLEN + sizeof(struct thread *) * 2 + 40, 1229 SBUF_FIXEDLEN); 1230 1231 sleepq_lock(wchan); 1232 sq = sleepq_lookup(wchan); 1233 if (sq == NULL) { 1234 /* This sleepq does not exist; exit and return ENOENT. */ 1235 error = ENOENT; 1236 finished = true; 1237 sleepq_release(wchan); 1238 goto loop_end; 1239 } 1240 1241 stack_idx = 0; 1242 /* Save thread info */ 1243 TAILQ_FOREACH_SAFE(td, &sq->sq_blocked[queue], td_slpq, 1244 td_next) { 1245 if (stack_idx >= stacks_to_allocate) 1246 goto loop_end; 1247 1248 /* Note the td_lock is equal to the sleepq_lock here. */ 1249 stack_save_td(st[stack_idx], td); 1250 1251 sbuf_printf(td_infos[stack_idx], "%d: %s %p", 1252 td->td_tid, td->td_name, td); 1253 1254 ++stack_idx; 1255 } 1256 1257 finished = true; 1258 sleepq_release(wchan); 1259 1260 /* Print the stacks */ 1261 for (i = 0; i < stack_idx; i++) { 1262 sbuf_finish(td_infos[i]); 1263 sbuf_printf(sb, "--- thread %s: ---\n", sbuf_data(td_infos[i])); 1264 stack_sbuf_print(sb, st[i]); 1265 sbuf_printf(sb, "\n"); 1266 1267 error = sbuf_error(sb); 1268 if (error == 0) 1269 *count_stacks_printed = stack_idx; 1270 } 1271 1272 loop_end: 1273 if (!finished) 1274 sleepq_release(wchan); 1275 for (stack_idx = 0; stack_idx < stacks_to_allocate; 1276 stack_idx++) 1277 stack_destroy(st[stack_idx]); 1278 for (stack_idx = 0; stack_idx < stacks_to_allocate; 1279 stack_idx++) 1280 sbuf_delete(td_infos[stack_idx]); 1281 free(st, M_TEMP); 1282 free(td_infos, M_TEMP); 1283 stacks_to_allocate *= 10; 1284 } 1285 1286 if (!finished && error == 0) 1287 error = ENOMEM; 1288 1289 return (error); 1290 } 1291 #endif 1292 1293 #ifdef SLEEPQUEUE_PROFILING 1294 #define SLEEPQ_PROF_LOCATIONS 1024 1295 #define SLEEPQ_SBUFSIZE 512 1296 struct sleepq_prof { 1297 LIST_ENTRY(sleepq_prof) sp_link; 1298 const char *sp_wmesg; 1299 long sp_count; 1300 }; 1301 1302 LIST_HEAD(sqphead, sleepq_prof); 1303 1304 struct sqphead sleepq_prof_free; 1305 struct sqphead sleepq_hash[SC_TABLESIZE]; 1306 static struct sleepq_prof sleepq_profent[SLEEPQ_PROF_LOCATIONS]; 1307 static struct mtx sleepq_prof_lock; 1308 MTX_SYSINIT(sleepq_prof_lock, &sleepq_prof_lock, "sleepq_prof", MTX_SPIN); 1309 1310 static void 1311 sleepq_profile(const char *wmesg) 1312 { 1313 struct sleepq_prof *sp; 1314 1315 mtx_lock_spin(&sleepq_prof_lock); 1316 if (prof_enabled == 0) 1317 goto unlock; 1318 LIST_FOREACH(sp, &sleepq_hash[SC_HASH(wmesg)], sp_link) 1319 if (sp->sp_wmesg == wmesg) 1320 goto done; 1321 sp = LIST_FIRST(&sleepq_prof_free); 1322 if (sp == NULL) 1323 goto unlock; 1324 sp->sp_wmesg = wmesg; 1325 LIST_REMOVE(sp, sp_link); 1326 LIST_INSERT_HEAD(&sleepq_hash[SC_HASH(wmesg)], sp, sp_link); 1327 done: 1328 sp->sp_count++; 1329 unlock: 1330 mtx_unlock_spin(&sleepq_prof_lock); 1331 return; 1332 } 1333 1334 static void 1335 sleepq_prof_reset(void) 1336 { 1337 struct sleepq_prof *sp; 1338 int enabled; 1339 int i; 1340 1341 mtx_lock_spin(&sleepq_prof_lock); 1342 enabled = prof_enabled; 1343 prof_enabled = 0; 1344 for (i = 0; i < SC_TABLESIZE; i++) 1345 LIST_INIT(&sleepq_hash[i]); 1346 LIST_INIT(&sleepq_prof_free); 1347 for (i = 0; i < SLEEPQ_PROF_LOCATIONS; i++) { 1348 sp = &sleepq_profent[i]; 1349 sp->sp_wmesg = NULL; 1350 sp->sp_count = 0; 1351 LIST_INSERT_HEAD(&sleepq_prof_free, sp, sp_link); 1352 } 1353 prof_enabled = enabled; 1354 mtx_unlock_spin(&sleepq_prof_lock); 1355 } 1356 1357 static int 1358 enable_sleepq_prof(SYSCTL_HANDLER_ARGS) 1359 { 1360 int error, v; 1361 1362 v = prof_enabled; 1363 error = sysctl_handle_int(oidp, &v, v, req); 1364 if (error) 1365 return (error); 1366 if (req->newptr == NULL) 1367 return (error); 1368 if (v == prof_enabled) 1369 return (0); 1370 if (v == 1) 1371 sleepq_prof_reset(); 1372 mtx_lock_spin(&sleepq_prof_lock); 1373 prof_enabled = !!v; 1374 mtx_unlock_spin(&sleepq_prof_lock); 1375 1376 return (0); 1377 } 1378 1379 static int 1380 reset_sleepq_prof_stats(SYSCTL_HANDLER_ARGS) 1381 { 1382 int error, v; 1383 1384 v = 0; 1385 error = sysctl_handle_int(oidp, &v, 0, req); 1386 if (error) 1387 return (error); 1388 if (req->newptr == NULL) 1389 return (error); 1390 if (v == 0) 1391 return (0); 1392 sleepq_prof_reset(); 1393 1394 return (0); 1395 } 1396 1397 static int 1398 dump_sleepq_prof_stats(SYSCTL_HANDLER_ARGS) 1399 { 1400 struct sleepq_prof *sp; 1401 struct sbuf *sb; 1402 int enabled; 1403 int error; 1404 int i; 1405 1406 error = sysctl_wire_old_buffer(req, 0); 1407 if (error != 0) 1408 return (error); 1409 sb = sbuf_new_for_sysctl(NULL, NULL, SLEEPQ_SBUFSIZE, req); 1410 sbuf_printf(sb, "\nwmesg\tcount\n"); 1411 enabled = prof_enabled; 1412 mtx_lock_spin(&sleepq_prof_lock); 1413 prof_enabled = 0; 1414 mtx_unlock_spin(&sleepq_prof_lock); 1415 for (i = 0; i < SC_TABLESIZE; i++) { 1416 LIST_FOREACH(sp, &sleepq_hash[i], sp_link) { 1417 sbuf_printf(sb, "%s\t%ld\n", 1418 sp->sp_wmesg, sp->sp_count); 1419 } 1420 } 1421 mtx_lock_spin(&sleepq_prof_lock); 1422 prof_enabled = enabled; 1423 mtx_unlock_spin(&sleepq_prof_lock); 1424 1425 error = sbuf_finish(sb); 1426 sbuf_delete(sb); 1427 return (error); 1428 } 1429 1430 SYSCTL_PROC(_debug_sleepq, OID_AUTO, stats, CTLTYPE_STRING | CTLFLAG_RD, 1431 NULL, 0, dump_sleepq_prof_stats, "A", "Sleepqueue profiling statistics"); 1432 SYSCTL_PROC(_debug_sleepq, OID_AUTO, reset, CTLTYPE_INT | CTLFLAG_RW, 1433 NULL, 0, reset_sleepq_prof_stats, "I", 1434 "Reset sleepqueue profiling statistics"); 1435 SYSCTL_PROC(_debug_sleepq, OID_AUTO, enable, CTLTYPE_INT | CTLFLAG_RW, 1436 NULL, 0, enable_sleepq_prof, "I", "Enable sleepqueue profiling"); 1437 #endif 1438 1439 #ifdef DDB 1440 DB_SHOW_COMMAND(sleepq, db_show_sleepqueue) 1441 { 1442 struct sleepqueue_chain *sc; 1443 struct sleepqueue *sq; 1444 #ifdef INVARIANTS 1445 struct lock_object *lock; 1446 #endif 1447 struct thread *td; 1448 void *wchan; 1449 int i; 1450 1451 if (!have_addr) 1452 return; 1453 1454 /* 1455 * First, see if there is an active sleep queue for the wait channel 1456 * indicated by the address. 1457 */ 1458 wchan = (void *)addr; 1459 sc = SC_LOOKUP(wchan); 1460 LIST_FOREACH(sq, &sc->sc_queues, sq_hash) 1461 if (sq->sq_wchan == wchan) 1462 goto found; 1463 1464 /* 1465 * Second, see if there is an active sleep queue at the address 1466 * indicated. 1467 */ 1468 for (i = 0; i < SC_TABLESIZE; i++) 1469 LIST_FOREACH(sq, &sleepq_chains[i].sc_queues, sq_hash) { 1470 if (sq == (struct sleepqueue *)addr) 1471 goto found; 1472 } 1473 1474 db_printf("Unable to locate a sleep queue via %p\n", (void *)addr); 1475 return; 1476 found: 1477 db_printf("Wait channel: %p\n", sq->sq_wchan); 1478 db_printf("Queue type: %d\n", sq->sq_type); 1479 #ifdef INVARIANTS 1480 if (sq->sq_lock) { 1481 lock = sq->sq_lock; 1482 db_printf("Associated Interlock: %p - (%s) %s\n", lock, 1483 LOCK_CLASS(lock)->lc_name, lock->lo_name); 1484 } 1485 #endif 1486 db_printf("Blocked threads:\n"); 1487 for (i = 0; i < NR_SLEEPQS; i++) { 1488 db_printf("\nQueue[%d]:\n", i); 1489 if (TAILQ_EMPTY(&sq->sq_blocked[i])) 1490 db_printf("\tempty\n"); 1491 else 1492 TAILQ_FOREACH(td, &sq->sq_blocked[i], 1493 td_slpq) { 1494 db_printf("\t%p (tid %d, pid %d, \"%s\")\n", td, 1495 td->td_tid, td->td_proc->p_pid, 1496 td->td_name); 1497 } 1498 db_printf("(expected: %u)\n", sq->sq_blockedcnt[i]); 1499 } 1500 } 1501 1502 /* Alias 'show sleepqueue' to 'show sleepq'. */ 1503 DB_SHOW_ALIAS(sleepqueue, db_show_sleepqueue); 1504 #endif 1505