1/*- 2 * Copyright (c) 2007 Attilio Rao <attilio@freebsd.org> 3 * Copyright (c) 2001 Jason Evans <jasone@freebsd.org> 4 * All rights reserved. 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(s), this list of conditions and the following disclaimer as 11 * the first lines of this file unmodified other than the possible 12 * addition of one or more copyright notices. 13 * 2. Redistributions in binary form must reproduce the above copyright 14 * notice(s), this list of conditions and the following disclaimer in the 15 * documentation and/or other materials provided with the distribution. 16 * 17 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDER(S) ``AS IS'' AND ANY 18 * EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED 19 * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE 20 * DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER(S) BE LIABLE FOR ANY 21 * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES 22 * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR 23 * SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER 24 * 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 SUCH 27 * DAMAGE. 28 */ 29 30/* 31 * Shared/exclusive locks. This implementation attempts to ensure 32 * deterministic lock granting behavior, so that slocks and xlocks are 33 * interleaved. 34 * 35 * Priority propagation will not generally raise the priority of lock holders, 36 * so should not be relied upon in combination with sx locks. 37 */ 38 39#include "opt_ddb.h" 40#include "opt_hwpmc_hooks.h" 41#include "opt_kdtrace.h" 42#include "opt_no_adaptive_sx.h" 43 44#include <sys/cdefs.h> 45__FBSDID("$FreeBSD$"); 46 47#include <sys/param.h> 48#include <sys/systm.h> 49#include <sys/kdb.h> 50#include <sys/ktr.h> 51#include <sys/lock.h> 52#include <sys/mutex.h> 53#include <sys/proc.h> 54#include <sys/sleepqueue.h> 55#include <sys/sx.h> 56#include <sys/sysctl.h> 57 58#if defined(SMP) && !defined(NO_ADAPTIVE_SX) 59#include <machine/cpu.h> 60#endif 61 62#ifdef DDB 63#include <ddb/ddb.h> 64#endif 65 66#if defined(SMP) && !defined(NO_ADAPTIVE_SX) 67#define ADAPTIVE_SX 68#endif 69 70CTASSERT((SX_NOADAPTIVE & LO_CLASSFLAGS) == SX_NOADAPTIVE); 71 72#ifdef HWPMC_HOOKS 73#include <sys/pmckern.h> 74PMC_SOFT_DECLARE( , , lock, failed); 75#endif 76 77/* Handy macros for sleep queues. */ 78#define SQ_EXCLUSIVE_QUEUE 0 79#define SQ_SHARED_QUEUE 1 80 81#ifdef ADAPTIVE_SX 82#define ASX_RETRIES 10 83#define ASX_LOOPS 10000 84#endif 85 86/* 87 * Variations on DROP_GIANT()/PICKUP_GIANT() for use in this file. We 88 * drop Giant anytime we have to sleep or if we adaptively spin. 89 */ 90#define GIANT_DECLARE \ 91 int _giantcnt = 0; \ 92 WITNESS_SAVE_DECL(Giant) \ 93 94#define GIANT_SAVE() do { \ 95 if (mtx_owned(&Giant)) { \ 96 WITNESS_SAVE(&Giant.lock_object, Giant); \ 97 while (mtx_owned(&Giant)) { \ 98 _giantcnt++; \ 99 mtx_unlock(&Giant); \ 100 } \ 101 } \ 102} while (0) 103 104#define GIANT_RESTORE() do { \ 105 if (_giantcnt > 0) { \ 106 mtx_assert(&Giant, MA_NOTOWNED); \ 107 while (_giantcnt--) \ 108 mtx_lock(&Giant); \ 109 WITNESS_RESTORE(&Giant.lock_object, Giant); \ 110 } \ 111} while (0) 112 113/* 114 * Returns true if an exclusive lock is recursed. It assumes 115 * curthread currently has an exclusive lock. 116 */ 117#define sx_recurse lock_object.lo_data 118#define sx_recursed(sx) ((sx)->sx_recurse != 0) 119 120static void assert_sx(struct lock_object *lock, int what); 121#ifdef DDB 122static void db_show_sx(struct lock_object *lock); 123#endif 124static void lock_sx(struct lock_object *lock, int how); 125#ifdef KDTRACE_HOOKS 126static int owner_sx(struct lock_object *lock, struct thread **owner); 127#endif 128static int unlock_sx(struct lock_object *lock); 129 130struct lock_class lock_class_sx = { 131 .lc_name = "sx", 132 .lc_flags = LC_SLEEPLOCK | LC_SLEEPABLE | LC_RECURSABLE | LC_UPGRADABLE, 133 .lc_assert = assert_sx, 134#ifdef DDB 135 .lc_ddb_show = db_show_sx, 136#endif 137 .lc_lock = lock_sx, 138 .lc_unlock = unlock_sx, 139#ifdef KDTRACE_HOOKS 140 .lc_owner = owner_sx, 141#endif 142}; 143 144#ifndef INVARIANTS 145#define _sx_assert(sx, what, file, line) 146#endif 147 148void 149assert_sx(struct lock_object *lock, int what) 150{ 151 152 sx_assert((struct sx *)lock, what); 153} 154 155void 156lock_sx(struct lock_object *lock, int how) 157{ 158 struct sx *sx; 159 160 sx = (struct sx *)lock; 161 if (how) 162 sx_xlock(sx); 163 else 164 sx_slock(sx); 165} 166 167int 168unlock_sx(struct lock_object *lock) 169{ 170 struct sx *sx; 171 172 sx = (struct sx *)lock; 173 sx_assert(sx, SA_LOCKED | SA_NOTRECURSED); 174 if (sx_xlocked(sx)) { 175 sx_xunlock(sx); 176 return (1); 177 } else { 178 sx_sunlock(sx); 179 return (0); 180 } 181} 182 183#ifdef KDTRACE_HOOKS 184int 185owner_sx(struct lock_object *lock, struct thread **owner) 186{ 187 struct sx *sx = (struct sx *)lock; 188 uintptr_t x = sx->sx_lock; 189 190 *owner = (struct thread *)SX_OWNER(x); 191 return ((x & SX_LOCK_SHARED) != 0 ? (SX_SHARERS(x) != 0) : 192 (*owner != NULL)); 193} 194#endif 195 196void 197sx_sysinit(void *arg) 198{ 199 struct sx_args *sargs = arg; 200 201 sx_init_flags(sargs->sa_sx, sargs->sa_desc, sargs->sa_flags); 202} 203 204void 205sx_init_flags(struct sx *sx, const char *description, int opts) 206{ 207 int flags; 208 209 MPASS((opts & ~(SX_QUIET | SX_RECURSE | SX_NOWITNESS | SX_DUPOK | 210 SX_NOPROFILE | SX_NOADAPTIVE)) == 0); 211 ASSERT_ATOMIC_LOAD_PTR(sx->sx_lock, 212 ("%s: sx_lock not aligned for %s: %p", __func__, description, 213 &sx->sx_lock)); 214 215 flags = LO_SLEEPABLE | LO_UPGRADABLE; 216 if (opts & SX_DUPOK) 217 flags |= LO_DUPOK; 218 if (opts & SX_NOPROFILE) 219 flags |= LO_NOPROFILE; 220 if (!(opts & SX_NOWITNESS)) 221 flags |= LO_WITNESS; 222 if (opts & SX_RECURSE) 223 flags |= LO_RECURSABLE; 224 if (opts & SX_QUIET) 225 flags |= LO_QUIET; 226 227 flags |= opts & SX_NOADAPTIVE; 228 sx->sx_lock = SX_LOCK_UNLOCKED; 229 sx->sx_recurse = 0; 230 lock_init(&sx->lock_object, &lock_class_sx, description, NULL, flags); 231} 232 233void 234sx_destroy(struct sx *sx) 235{ 236 237 KASSERT(sx->sx_lock == SX_LOCK_UNLOCKED, ("sx lock still held")); 238 KASSERT(sx->sx_recurse == 0, ("sx lock still recursed")); 239 sx->sx_lock = SX_LOCK_DESTROYED; 240 lock_destroy(&sx->lock_object); 241} 242 243int 244_sx_slock(struct sx *sx, int opts, const char *file, int line) 245{ 246 int error = 0; 247 248 if (SCHEDULER_STOPPED()) 249 return (0); 250 KASSERT(kdb_active != 0 || !TD_IS_IDLETHREAD(curthread), 251 ("sx_slock() by idle thread %p on sx %s @ %s:%d", 252 curthread, sx->lock_object.lo_name, file, line)); 253 KASSERT(sx->sx_lock != SX_LOCK_DESTROYED, 254 ("sx_slock() of destroyed sx @ %s:%d", file, line)); 255 WITNESS_CHECKORDER(&sx->lock_object, LOP_NEWORDER, file, line, NULL); 256 error = __sx_slock(sx, opts, file, line); 257 if (!error) { 258 LOCK_LOG_LOCK("SLOCK", &sx->lock_object, 0, 0, file, line); 259 WITNESS_LOCK(&sx->lock_object, 0, file, line); 260 curthread->td_locks++; 261 } 262 263 return (error); 264} 265 266int 267_sx_try_slock(struct sx *sx, const char *file, int line) 268{ 269 uintptr_t x; 270 271 if (SCHEDULER_STOPPED()) 272 return (1); 273 274 KASSERT(kdb_active != 0 || !TD_IS_IDLETHREAD(curthread), 275 ("sx_try_slock() by idle thread %p on sx %s @ %s:%d", 276 curthread, sx->lock_object.lo_name, file, line)); 277 278 for (;;) { 279 x = sx->sx_lock; 280 KASSERT(x != SX_LOCK_DESTROYED, 281 ("sx_try_slock() of destroyed sx @ %s:%d", file, line)); 282 if (!(x & SX_LOCK_SHARED)) 283 break; 284 if (atomic_cmpset_acq_ptr(&sx->sx_lock, x, x + SX_ONE_SHARER)) { 285 LOCK_LOG_TRY("SLOCK", &sx->lock_object, 0, 1, file, line); 286 WITNESS_LOCK(&sx->lock_object, LOP_TRYLOCK, file, line); 287 curthread->td_locks++; 288 return (1); 289 } 290 } 291 292 LOCK_LOG_TRY("SLOCK", &sx->lock_object, 0, 0, file, line); 293 return (0); 294} 295 296int 297_sx_xlock(struct sx *sx, int opts, const char *file, int line) 298{ 299 int error = 0; 300 301 if (SCHEDULER_STOPPED()) 302 return (0); 303 KASSERT(kdb_active != 0 || !TD_IS_IDLETHREAD(curthread), 304 ("sx_xlock() by idle thread %p on sx %s @ %s:%d", 305 curthread, sx->lock_object.lo_name, file, line)); 306 KASSERT(sx->sx_lock != SX_LOCK_DESTROYED, 307 ("sx_xlock() of destroyed sx @ %s:%d", file, line)); 308 WITNESS_CHECKORDER(&sx->lock_object, LOP_NEWORDER | LOP_EXCLUSIVE, file, 309 line, NULL); 310 error = __sx_xlock(sx, curthread, opts, file, line); 311 if (!error) { 312 LOCK_LOG_LOCK("XLOCK", &sx->lock_object, 0, sx->sx_recurse, 313 file, line); 314 WITNESS_LOCK(&sx->lock_object, LOP_EXCLUSIVE, file, line); 315 curthread->td_locks++; 316 } 317 318 return (error); 319} 320 321int 322_sx_try_xlock(struct sx *sx, const char *file, int line) 323{ 324 int rval; 325 326 if (SCHEDULER_STOPPED()) 327 return (1); 328 329 KASSERT(kdb_active != 0 || !TD_IS_IDLETHREAD(curthread), 330 ("sx_try_xlock() by idle thread %p on sx %s @ %s:%d", 331 curthread, sx->lock_object.lo_name, file, line)); 332 KASSERT(sx->sx_lock != SX_LOCK_DESTROYED, 333 ("sx_try_xlock() of destroyed sx @ %s:%d", file, line)); 334 335 if (sx_xlocked(sx) && 336 (sx->lock_object.lo_flags & LO_RECURSABLE) != 0) { 337 sx->sx_recurse++; 338 atomic_set_ptr(&sx->sx_lock, SX_LOCK_RECURSED); 339 rval = 1; 340 } else 341 rval = atomic_cmpset_acq_ptr(&sx->sx_lock, SX_LOCK_UNLOCKED, 342 (uintptr_t)curthread); 343 LOCK_LOG_TRY("XLOCK", &sx->lock_object, 0, rval, file, line); 344 if (rval) { 345 WITNESS_LOCK(&sx->lock_object, LOP_EXCLUSIVE | LOP_TRYLOCK, 346 file, line); 347 curthread->td_locks++; 348 } 349 350 return (rval); 351} 352 353void 354_sx_sunlock(struct sx *sx, const char *file, int line) 355{ 356 357 if (SCHEDULER_STOPPED()) 358 return; 359 KASSERT(sx->sx_lock != SX_LOCK_DESTROYED, 360 ("sx_sunlock() of destroyed sx @ %s:%d", file, line)); 361 _sx_assert(sx, SA_SLOCKED, file, line); 362 curthread->td_locks--; 363 WITNESS_UNLOCK(&sx->lock_object, 0, file, line); 364 LOCK_LOG_LOCK("SUNLOCK", &sx->lock_object, 0, 0, file, line); 365 __sx_sunlock(sx, file, line); 366 LOCKSTAT_PROFILE_RELEASE_LOCK(LS_SX_SUNLOCK_RELEASE, sx); 367} 368 369void 370_sx_xunlock(struct sx *sx, const char *file, int line) 371{ 372 373 if (SCHEDULER_STOPPED()) 374 return; 375 KASSERT(sx->sx_lock != SX_LOCK_DESTROYED, 376 ("sx_xunlock() of destroyed sx @ %s:%d", file, line)); 377 _sx_assert(sx, SA_XLOCKED, file, line); 378 curthread->td_locks--; 379 WITNESS_UNLOCK(&sx->lock_object, LOP_EXCLUSIVE, file, line); 380 LOCK_LOG_LOCK("XUNLOCK", &sx->lock_object, 0, sx->sx_recurse, file, 381 line); 382 if (!sx_recursed(sx)) 383 LOCKSTAT_PROFILE_RELEASE_LOCK(LS_SX_XUNLOCK_RELEASE, sx); 384 __sx_xunlock(sx, curthread, file, line); 385} 386 387/* 388 * Try to do a non-blocking upgrade from a shared lock to an exclusive lock. 389 * This will only succeed if this thread holds a single shared lock. 390 * Return 1 if if the upgrade succeed, 0 otherwise. 391 */ 392int 393_sx_try_upgrade(struct sx *sx, const char *file, int line) 394{ 395 uintptr_t x; 396 int success; 397 398 if (SCHEDULER_STOPPED()) 399 return (1); 400 401 KASSERT(sx->sx_lock != SX_LOCK_DESTROYED, 402 ("sx_try_upgrade() of destroyed sx @ %s:%d", file, line)); 403 _sx_assert(sx, SA_SLOCKED, file, line); 404 405 /* 406 * Try to switch from one shared lock to an exclusive lock. We need 407 * to maintain the SX_LOCK_EXCLUSIVE_WAITERS flag if set so that 408 * we will wake up the exclusive waiters when we drop the lock. 409 */ 410 x = sx->sx_lock & SX_LOCK_EXCLUSIVE_WAITERS; 411 success = atomic_cmpset_ptr(&sx->sx_lock, SX_SHARERS_LOCK(1) | x, 412 (uintptr_t)curthread | x); 413 LOCK_LOG_TRY("XUPGRADE", &sx->lock_object, 0, success, file, line); 414 if (success) { 415 WITNESS_UPGRADE(&sx->lock_object, LOP_EXCLUSIVE | LOP_TRYLOCK, 416 file, line); 417 LOCKSTAT_RECORD0(LS_SX_TRYUPGRADE_UPGRADE, sx); 418 } 419 return (success); 420} 421 422/* 423 * Downgrade an unrecursed exclusive lock into a single shared lock. 424 */ 425void 426_sx_downgrade(struct sx *sx, const char *file, int line) 427{ 428 uintptr_t x; 429 int wakeup_swapper; 430 431 if (SCHEDULER_STOPPED()) 432 return; 433 434 KASSERT(sx->sx_lock != SX_LOCK_DESTROYED, 435 ("sx_downgrade() of destroyed sx @ %s:%d", file, line)); 436 _sx_assert(sx, SA_XLOCKED | SA_NOTRECURSED, file, line); 437#ifndef INVARIANTS 438 if (sx_recursed(sx)) 439 panic("downgrade of a recursed lock"); 440#endif 441 442 WITNESS_DOWNGRADE(&sx->lock_object, 0, file, line); 443 444 /* 445 * Try to switch from an exclusive lock with no shared waiters 446 * to one sharer with no shared waiters. If there are 447 * exclusive waiters, we don't need to lock the sleep queue so 448 * long as we preserve the flag. We do one quick try and if 449 * that fails we grab the sleepq lock to keep the flags from 450 * changing and do it the slow way. 451 * 452 * We have to lock the sleep queue if there are shared waiters 453 * so we can wake them up. 454 */ 455 x = sx->sx_lock; 456 if (!(x & SX_LOCK_SHARED_WAITERS) && 457 atomic_cmpset_rel_ptr(&sx->sx_lock, x, SX_SHARERS_LOCK(1) | 458 (x & SX_LOCK_EXCLUSIVE_WAITERS))) { 459 LOCK_LOG_LOCK("XDOWNGRADE", &sx->lock_object, 0, 0, file, line); 460 return; 461 } 462 463 /* 464 * Lock the sleep queue so we can read the waiters bits 465 * without any races and wakeup any shared waiters. 466 */ 467 sleepq_lock(&sx->lock_object); 468 469 /* 470 * Preserve SX_LOCK_EXCLUSIVE_WAITERS while downgraded to a single 471 * shared lock. If there are any shared waiters, wake them up. 472 */ 473 wakeup_swapper = 0; 474 x = sx->sx_lock; 475 atomic_store_rel_ptr(&sx->sx_lock, SX_SHARERS_LOCK(1) | 476 (x & SX_LOCK_EXCLUSIVE_WAITERS)); 477 if (x & SX_LOCK_SHARED_WAITERS) 478 wakeup_swapper = sleepq_broadcast(&sx->lock_object, SLEEPQ_SX, 479 0, SQ_SHARED_QUEUE); 480 sleepq_release(&sx->lock_object); 481 482 LOCK_LOG_LOCK("XDOWNGRADE", &sx->lock_object, 0, 0, file, line); 483 LOCKSTAT_RECORD0(LS_SX_DOWNGRADE_DOWNGRADE, sx); 484 485 if (wakeup_swapper) 486 kick_proc0(); 487} 488 489/* 490 * This function represents the so-called 'hard case' for sx_xlock 491 * operation. All 'easy case' failures are redirected to this. Note 492 * that ideally this would be a static function, but it needs to be 493 * accessible from at least sx.h. 494 */ 495int 496_sx_xlock_hard(struct sx *sx, uintptr_t tid, int opts, const char *file, 497 int line) 498{ 499 GIANT_DECLARE; 500#ifdef ADAPTIVE_SX 501 volatile struct thread *owner; 502 u_int i, spintries = 0; 503#endif 504 uintptr_t x; 505#ifdef LOCK_PROFILING 506 uint64_t waittime = 0; 507 int contested = 0; 508#endif 509 int error = 0; 510#ifdef KDTRACE_HOOKS 511 uint64_t spin_cnt = 0; 512 uint64_t sleep_cnt = 0; 513 int64_t sleep_time = 0; 514#endif 515 516 if (SCHEDULER_STOPPED()) 517 return (0); 518 519 /* If we already hold an exclusive lock, then recurse. */ 520 if (sx_xlocked(sx)) { 521 KASSERT((sx->lock_object.lo_flags & LO_RECURSABLE) != 0, 522 ("_sx_xlock_hard: recursed on non-recursive sx %s @ %s:%d\n", 523 sx->lock_object.lo_name, file, line)); 524 sx->sx_recurse++; 525 atomic_set_ptr(&sx->sx_lock, SX_LOCK_RECURSED); 526 if (LOCK_LOG_TEST(&sx->lock_object, 0)) 527 CTR2(KTR_LOCK, "%s: %p recursing", __func__, sx); 528 return (0); 529 } 530 531 if (LOCK_LOG_TEST(&sx->lock_object, 0)) 532 CTR5(KTR_LOCK, "%s: %s contested (lock=%p) at %s:%d", __func__, 533 sx->lock_object.lo_name, (void *)sx->sx_lock, file, line); 534 535 while (!atomic_cmpset_acq_ptr(&sx->sx_lock, SX_LOCK_UNLOCKED, tid)) { 536#ifdef KDTRACE_HOOKS 537 spin_cnt++; 538#endif 539#ifdef HWPMC_HOOKS 540 PMC_SOFT_CALL( , , lock, failed); 541#endif 542 lock_profile_obtain_lock_failed(&sx->lock_object, &contested, 543 &waittime); 544#ifdef ADAPTIVE_SX 545 /* 546 * If the lock is write locked and the owner is 547 * running on another CPU, spin until the owner stops 548 * running or the state of the lock changes. 549 */ 550 x = sx->sx_lock; 551 if ((sx->lock_object.lo_flags & SX_NOADAPTIVE) == 0) { 552 if ((x & SX_LOCK_SHARED) == 0) { 553 x = SX_OWNER(x); 554 owner = (struct thread *)x; 555 if (TD_IS_RUNNING(owner)) { 556 if (LOCK_LOG_TEST(&sx->lock_object, 0)) 557 CTR3(KTR_LOCK, 558 "%s: spinning on %p held by %p", 559 __func__, sx, owner); 560 GIANT_SAVE(); 561 while (SX_OWNER(sx->sx_lock) == x && 562 TD_IS_RUNNING(owner)) { 563 cpu_spinwait(); 564#ifdef KDTRACE_HOOKS 565 spin_cnt++; 566#endif 567 } 568 continue; 569 } 570 } else if (SX_SHARERS(x) && spintries < ASX_RETRIES) { 571 GIANT_SAVE(); 572 spintries++; 573 for (i = 0; i < ASX_LOOPS; i++) { 574 if (LOCK_LOG_TEST(&sx->lock_object, 0)) 575 CTR4(KTR_LOCK, 576 "%s: shared spinning on %p with %u and %u", 577 __func__, sx, spintries, i); 578 x = sx->sx_lock; 579 if ((x & SX_LOCK_SHARED) == 0 || 580 SX_SHARERS(x) == 0) 581 break; 582 cpu_spinwait(); 583#ifdef KDTRACE_HOOKS 584 spin_cnt++; 585#endif 586 } 587 if (i != ASX_LOOPS) 588 continue; 589 } 590 } 591#endif 592 593 sleepq_lock(&sx->lock_object); 594 x = sx->sx_lock; 595 596 /* 597 * If the lock was released while spinning on the 598 * sleep queue chain lock, try again. 599 */ 600 if (x == SX_LOCK_UNLOCKED) { 601 sleepq_release(&sx->lock_object); 602 continue; 603 } 604 605#ifdef ADAPTIVE_SX 606 /* 607 * The current lock owner might have started executing 608 * on another CPU (or the lock could have changed 609 * owners) while we were waiting on the sleep queue 610 * chain lock. If so, drop the sleep queue lock and try 611 * again. 612 */ 613 if (!(x & SX_LOCK_SHARED) && 614 (sx->lock_object.lo_flags & SX_NOADAPTIVE) == 0) { 615 owner = (struct thread *)SX_OWNER(x); 616 if (TD_IS_RUNNING(owner)) { 617 sleepq_release(&sx->lock_object); 618 continue; 619 } 620 } 621#endif 622 623 /* 624 * If an exclusive lock was released with both shared 625 * and exclusive waiters and a shared waiter hasn't 626 * woken up and acquired the lock yet, sx_lock will be 627 * set to SX_LOCK_UNLOCKED | SX_LOCK_EXCLUSIVE_WAITERS. 628 * If we see that value, try to acquire it once. Note 629 * that we have to preserve SX_LOCK_EXCLUSIVE_WAITERS 630 * as there are other exclusive waiters still. If we 631 * fail, restart the loop. 632 */ 633 if (x == (SX_LOCK_UNLOCKED | SX_LOCK_EXCLUSIVE_WAITERS)) { 634 if (atomic_cmpset_acq_ptr(&sx->sx_lock, 635 SX_LOCK_UNLOCKED | SX_LOCK_EXCLUSIVE_WAITERS, 636 tid | SX_LOCK_EXCLUSIVE_WAITERS)) { 637 sleepq_release(&sx->lock_object); 638 CTR2(KTR_LOCK, "%s: %p claimed by new writer", 639 __func__, sx); 640 break; 641 } 642 sleepq_release(&sx->lock_object); 643 continue; 644 } 645 646 /* 647 * Try to set the SX_LOCK_EXCLUSIVE_WAITERS. If we fail, 648 * than loop back and retry. 649 */ 650 if (!(x & SX_LOCK_EXCLUSIVE_WAITERS)) { 651 if (!atomic_cmpset_ptr(&sx->sx_lock, x, 652 x | SX_LOCK_EXCLUSIVE_WAITERS)) { 653 sleepq_release(&sx->lock_object); 654 continue; 655 } 656 if (LOCK_LOG_TEST(&sx->lock_object, 0)) 657 CTR2(KTR_LOCK, "%s: %p set excl waiters flag", 658 __func__, sx); 659 } 660 661 /* 662 * Since we have been unable to acquire the exclusive 663 * lock and the exclusive waiters flag is set, we have 664 * to sleep. 665 */ 666 if (LOCK_LOG_TEST(&sx->lock_object, 0)) 667 CTR2(KTR_LOCK, "%s: %p blocking on sleep queue", 668 __func__, sx); 669 670#ifdef KDTRACE_HOOKS 671 sleep_time -= lockstat_nsecs(); 672#endif 673 GIANT_SAVE(); 674 sleepq_add(&sx->lock_object, NULL, sx->lock_object.lo_name, 675 SLEEPQ_SX | ((opts & SX_INTERRUPTIBLE) ? 676 SLEEPQ_INTERRUPTIBLE : 0), SQ_EXCLUSIVE_QUEUE); 677 if (!(opts & SX_INTERRUPTIBLE)) 678 sleepq_wait(&sx->lock_object, 0); 679 else 680 error = sleepq_wait_sig(&sx->lock_object, 0); 681#ifdef KDTRACE_HOOKS 682 sleep_time += lockstat_nsecs(); 683 sleep_cnt++; 684#endif 685 if (error) { 686 if (LOCK_LOG_TEST(&sx->lock_object, 0)) 687 CTR2(KTR_LOCK, 688 "%s: interruptible sleep by %p suspended by signal", 689 __func__, sx); 690 break; 691 } 692 if (LOCK_LOG_TEST(&sx->lock_object, 0)) 693 CTR2(KTR_LOCK, "%s: %p resuming from sleep queue", 694 __func__, sx); 695 } 696 697 GIANT_RESTORE(); 698 if (!error) 699 LOCKSTAT_PROFILE_OBTAIN_LOCK_SUCCESS(LS_SX_XLOCK_ACQUIRE, sx, 700 contested, waittime, file, line); 701#ifdef KDTRACE_HOOKS 702 if (sleep_time) 703 LOCKSTAT_RECORD1(LS_SX_XLOCK_BLOCK, sx, sleep_time); 704 if (spin_cnt > sleep_cnt) 705 LOCKSTAT_RECORD1(LS_SX_XLOCK_SPIN, sx, (spin_cnt - sleep_cnt)); 706#endif 707 return (error); 708} 709 710/* 711 * This function represents the so-called 'hard case' for sx_xunlock 712 * operation. All 'easy case' failures are redirected to this. Note 713 * that ideally this would be a static function, but it needs to be 714 * accessible from at least sx.h. 715 */ 716void 717_sx_xunlock_hard(struct sx *sx, uintptr_t tid, const char *file, int line) 718{ 719 uintptr_t x; 720 int queue, wakeup_swapper; 721 722 if (SCHEDULER_STOPPED()) 723 return; 724 725 MPASS(!(sx->sx_lock & SX_LOCK_SHARED)); 726 727 /* If the lock is recursed, then unrecurse one level. */ 728 if (sx_xlocked(sx) && sx_recursed(sx)) { 729 if ((--sx->sx_recurse) == 0) 730 atomic_clear_ptr(&sx->sx_lock, SX_LOCK_RECURSED); 731 if (LOCK_LOG_TEST(&sx->lock_object, 0)) 732 CTR2(KTR_LOCK, "%s: %p unrecursing", __func__, sx); 733 return; 734 } 735 MPASS(sx->sx_lock & (SX_LOCK_SHARED_WAITERS | 736 SX_LOCK_EXCLUSIVE_WAITERS)); 737 if (LOCK_LOG_TEST(&sx->lock_object, 0)) 738 CTR2(KTR_LOCK, "%s: %p contested", __func__, sx); 739 740 sleepq_lock(&sx->lock_object); 741 x = SX_LOCK_UNLOCKED; 742 743 /* 744 * The wake up algorithm here is quite simple and probably not 745 * ideal. It gives precedence to shared waiters if they are 746 * present. For this condition, we have to preserve the 747 * state of the exclusive waiters flag. 748 * If interruptible sleeps left the shared queue empty avoid a 749 * starvation for the threads sleeping on the exclusive queue by giving 750 * them precedence and cleaning up the shared waiters bit anyway. 751 */ 752 if ((sx->sx_lock & SX_LOCK_SHARED_WAITERS) != 0 && 753 sleepq_sleepcnt(&sx->lock_object, SQ_SHARED_QUEUE) != 0) { 754 queue = SQ_SHARED_QUEUE; 755 x |= (sx->sx_lock & SX_LOCK_EXCLUSIVE_WAITERS); 756 } else 757 queue = SQ_EXCLUSIVE_QUEUE; 758 759 /* Wake up all the waiters for the specific queue. */ 760 if (LOCK_LOG_TEST(&sx->lock_object, 0)) 761 CTR3(KTR_LOCK, "%s: %p waking up all threads on %s queue", 762 __func__, sx, queue == SQ_SHARED_QUEUE ? "shared" : 763 "exclusive"); 764 atomic_store_rel_ptr(&sx->sx_lock, x); 765 wakeup_swapper = sleepq_broadcast(&sx->lock_object, SLEEPQ_SX, 0, 766 queue); 767 sleepq_release(&sx->lock_object); 768 if (wakeup_swapper) 769 kick_proc0(); 770} 771 772/* 773 * This function represents the so-called 'hard case' for sx_slock 774 * operation. All 'easy case' failures are redirected to this. Note 775 * that ideally this would be a static function, but it needs to be 776 * accessible from at least sx.h. 777 */ 778int 779_sx_slock_hard(struct sx *sx, int opts, const char *file, int line) 780{ 781 GIANT_DECLARE; 782#ifdef ADAPTIVE_SX 783 volatile struct thread *owner; 784#endif 785#ifdef LOCK_PROFILING 786 uint64_t waittime = 0; 787 int contested = 0; 788#endif 789 uintptr_t x; 790 int error = 0; 791#ifdef KDTRACE_HOOKS 792 uint64_t spin_cnt = 0; 793 uint64_t sleep_cnt = 0; 794 int64_t sleep_time = 0; 795#endif 796 797 if (SCHEDULER_STOPPED()) 798 return (0); 799 800 /* 801 * As with rwlocks, we don't make any attempt to try to block 802 * shared locks once there is an exclusive waiter. 803 */ 804 for (;;) { 805#ifdef KDTRACE_HOOKS 806 spin_cnt++; 807#endif 808 x = sx->sx_lock; 809 810 /* 811 * If no other thread has an exclusive lock then try to bump up 812 * the count of sharers. Since we have to preserve the state 813 * of SX_LOCK_EXCLUSIVE_WAITERS, if we fail to acquire the 814 * shared lock loop back and retry. 815 */ 816 if (x & SX_LOCK_SHARED) { 817 MPASS(!(x & SX_LOCK_SHARED_WAITERS)); 818 if (atomic_cmpset_acq_ptr(&sx->sx_lock, x, 819 x + SX_ONE_SHARER)) { 820 if (LOCK_LOG_TEST(&sx->lock_object, 0)) 821 CTR4(KTR_LOCK, 822 "%s: %p succeed %p -> %p", __func__, 823 sx, (void *)x, 824 (void *)(x + SX_ONE_SHARER)); 825 break; 826 } 827 continue; 828 } 829#ifdef HWPMC_HOOKS 830 PMC_SOFT_CALL( , , lock, failed); 831#endif 832 lock_profile_obtain_lock_failed(&sx->lock_object, &contested, 833 &waittime); 834 835#ifdef ADAPTIVE_SX 836 /* 837 * If the owner is running on another CPU, spin until 838 * the owner stops running or the state of the lock 839 * changes. 840 */ 841 if ((sx->lock_object.lo_flags & SX_NOADAPTIVE) == 0) { 842 x = SX_OWNER(x); 843 owner = (struct thread *)x; 844 if (TD_IS_RUNNING(owner)) { 845 if (LOCK_LOG_TEST(&sx->lock_object, 0)) 846 CTR3(KTR_LOCK, 847 "%s: spinning on %p held by %p", 848 __func__, sx, owner); 849 GIANT_SAVE(); 850 while (SX_OWNER(sx->sx_lock) == x && 851 TD_IS_RUNNING(owner)) { 852#ifdef KDTRACE_HOOKS 853 spin_cnt++; 854#endif 855 cpu_spinwait(); 856 } 857 continue; 858 } 859 } 860#endif 861 862 /* 863 * Some other thread already has an exclusive lock, so 864 * start the process of blocking. 865 */ 866 sleepq_lock(&sx->lock_object); 867 x = sx->sx_lock; 868 869 /* 870 * The lock could have been released while we spun. 871 * In this case loop back and retry. 872 */ 873 if (x & SX_LOCK_SHARED) { 874 sleepq_release(&sx->lock_object); 875 continue; 876 } 877 878#ifdef ADAPTIVE_SX 879 /* 880 * If the owner is running on another CPU, spin until 881 * the owner stops running or the state of the lock 882 * changes. 883 */ 884 if (!(x & SX_LOCK_SHARED) && 885 (sx->lock_object.lo_flags & SX_NOADAPTIVE) == 0) { 886 owner = (struct thread *)SX_OWNER(x); 887 if (TD_IS_RUNNING(owner)) { 888 sleepq_release(&sx->lock_object); 889 continue; 890 } 891 } 892#endif 893 894 /* 895 * Try to set the SX_LOCK_SHARED_WAITERS flag. If we 896 * fail to set it drop the sleep queue lock and loop 897 * back. 898 */ 899 if (!(x & SX_LOCK_SHARED_WAITERS)) { 900 if (!atomic_cmpset_ptr(&sx->sx_lock, x, 901 x | SX_LOCK_SHARED_WAITERS)) { 902 sleepq_release(&sx->lock_object); 903 continue; 904 } 905 if (LOCK_LOG_TEST(&sx->lock_object, 0)) 906 CTR2(KTR_LOCK, "%s: %p set shared waiters flag", 907 __func__, sx); 908 } 909 910 /* 911 * Since we have been unable to acquire the shared lock, 912 * we have to sleep. 913 */ 914 if (LOCK_LOG_TEST(&sx->lock_object, 0)) 915 CTR2(KTR_LOCK, "%s: %p blocking on sleep queue", 916 __func__, sx); 917 918#ifdef KDTRACE_HOOKS 919 sleep_time -= lockstat_nsecs(); 920#endif 921 GIANT_SAVE(); 922 sleepq_add(&sx->lock_object, NULL, sx->lock_object.lo_name, 923 SLEEPQ_SX | ((opts & SX_INTERRUPTIBLE) ? 924 SLEEPQ_INTERRUPTIBLE : 0), SQ_SHARED_QUEUE); 925 if (!(opts & SX_INTERRUPTIBLE)) 926 sleepq_wait(&sx->lock_object, 0); 927 else 928 error = sleepq_wait_sig(&sx->lock_object, 0); 929#ifdef KDTRACE_HOOKS 930 sleep_time += lockstat_nsecs(); 931 sleep_cnt++; 932#endif 933 if (error) { 934 if (LOCK_LOG_TEST(&sx->lock_object, 0)) 935 CTR2(KTR_LOCK, 936 "%s: interruptible sleep by %p suspended by signal", 937 __func__, sx); 938 break; 939 } 940 if (LOCK_LOG_TEST(&sx->lock_object, 0)) 941 CTR2(KTR_LOCK, "%s: %p resuming from sleep queue", 942 __func__, sx); 943 } 944 if (error == 0) 945 LOCKSTAT_PROFILE_OBTAIN_LOCK_SUCCESS(LS_SX_SLOCK_ACQUIRE, sx, 946 contested, waittime, file, line); 947#ifdef KDTRACE_HOOKS 948 if (sleep_time) 949 LOCKSTAT_RECORD1(LS_SX_XLOCK_BLOCK, sx, sleep_time); 950 if (spin_cnt > sleep_cnt) 951 LOCKSTAT_RECORD1(LS_SX_XLOCK_SPIN, sx, (spin_cnt - sleep_cnt)); 952#endif 953 GIANT_RESTORE(); 954 return (error); 955} 956 957/* 958 * This function represents the so-called 'hard case' for sx_sunlock 959 * operation. All 'easy case' failures are redirected to this. Note 960 * that ideally this would be a static function, but it needs to be 961 * accessible from at least sx.h. 962 */ 963void 964_sx_sunlock_hard(struct sx *sx, const char *file, int line) 965{ 966 uintptr_t x; 967 int wakeup_swapper; 968 969 if (SCHEDULER_STOPPED()) 970 return; 971 972 for (;;) { 973 x = sx->sx_lock; 974 975 /* 976 * We should never have sharers while at least one thread 977 * holds a shared lock. 978 */ 979 KASSERT(!(x & SX_LOCK_SHARED_WAITERS), 980 ("%s: waiting sharers", __func__)); 981 982 /* 983 * See if there is more than one shared lock held. If 984 * so, just drop one and return. 985 */ 986 if (SX_SHARERS(x) > 1) { 987 if (atomic_cmpset_rel_ptr(&sx->sx_lock, x, 988 x - SX_ONE_SHARER)) { 989 if (LOCK_LOG_TEST(&sx->lock_object, 0)) 990 CTR4(KTR_LOCK, 991 "%s: %p succeeded %p -> %p", 992 __func__, sx, (void *)x, 993 (void *)(x - SX_ONE_SHARER)); 994 break; 995 } 996 continue; 997 } 998 999 /* 1000 * If there aren't any waiters for an exclusive lock, 1001 * then try to drop it quickly. 1002 */ 1003 if (!(x & SX_LOCK_EXCLUSIVE_WAITERS)) { 1004 MPASS(x == SX_SHARERS_LOCK(1)); 1005 if (atomic_cmpset_rel_ptr(&sx->sx_lock, 1006 SX_SHARERS_LOCK(1), SX_LOCK_UNLOCKED)) { 1007 if (LOCK_LOG_TEST(&sx->lock_object, 0)) 1008 CTR2(KTR_LOCK, "%s: %p last succeeded", 1009 __func__, sx); 1010 break; 1011 } 1012 continue; 1013 } 1014 1015 /* 1016 * At this point, there should just be one sharer with 1017 * exclusive waiters. 1018 */ 1019 MPASS(x == (SX_SHARERS_LOCK(1) | SX_LOCK_EXCLUSIVE_WAITERS)); 1020 1021 sleepq_lock(&sx->lock_object); 1022 1023 /* 1024 * Wake up semantic here is quite simple: 1025 * Just wake up all the exclusive waiters. 1026 * Note that the state of the lock could have changed, 1027 * so if it fails loop back and retry. 1028 */ 1029 if (!atomic_cmpset_rel_ptr(&sx->sx_lock, 1030 SX_SHARERS_LOCK(1) | SX_LOCK_EXCLUSIVE_WAITERS, 1031 SX_LOCK_UNLOCKED)) { 1032 sleepq_release(&sx->lock_object); 1033 continue; 1034 } 1035 if (LOCK_LOG_TEST(&sx->lock_object, 0)) 1036 CTR2(KTR_LOCK, "%s: %p waking up all thread on" 1037 "exclusive queue", __func__, sx); 1038 wakeup_swapper = sleepq_broadcast(&sx->lock_object, SLEEPQ_SX, 1039 0, SQ_EXCLUSIVE_QUEUE); 1040 sleepq_release(&sx->lock_object); 1041 if (wakeup_swapper) 1042 kick_proc0(); 1043 break; 1044 } 1045} 1046 1047#ifdef INVARIANT_SUPPORT 1048#ifndef INVARIANTS 1049#undef _sx_assert 1050#endif 1051 1052/* 1053 * In the non-WITNESS case, sx_assert() can only detect that at least 1054 * *some* thread owns an slock, but it cannot guarantee that *this* 1055 * thread owns an slock. 1056 */ 1057void 1058_sx_assert(struct sx *sx, int what, const char *file, int line) 1059{ 1060#ifndef WITNESS 1061 int slocked = 0; 1062#endif 1063 1064 if (panicstr != NULL) 1065 return; 1066 switch (what) { 1067 case SA_SLOCKED: 1068 case SA_SLOCKED | SA_NOTRECURSED: 1069 case SA_SLOCKED | SA_RECURSED: 1070#ifndef WITNESS 1071 slocked = 1; 1072 /* FALLTHROUGH */ 1073#endif 1074 case SA_LOCKED: 1075 case SA_LOCKED | SA_NOTRECURSED: 1076 case SA_LOCKED | SA_RECURSED: 1077#ifdef WITNESS 1078 witness_assert(&sx->lock_object, what, file, line); 1079#else 1080 /* 1081 * If some other thread has an exclusive lock or we 1082 * have one and are asserting a shared lock, fail. 1083 * Also, if no one has a lock at all, fail. 1084 */ 1085 if (sx->sx_lock == SX_LOCK_UNLOCKED || 1086 (!(sx->sx_lock & SX_LOCK_SHARED) && (slocked || 1087 sx_xholder(sx) != curthread))) 1088 panic("Lock %s not %slocked @ %s:%d\n", 1089 sx->lock_object.lo_name, slocked ? "share " : "", 1090 file, line); 1091 1092 if (!(sx->sx_lock & SX_LOCK_SHARED)) { 1093 if (sx_recursed(sx)) { 1094 if (what & SA_NOTRECURSED) 1095 panic("Lock %s recursed @ %s:%d\n", 1096 sx->lock_object.lo_name, file, 1097 line); 1098 } else if (what & SA_RECURSED) 1099 panic("Lock %s not recursed @ %s:%d\n", 1100 sx->lock_object.lo_name, file, line); 1101 } 1102#endif 1103 break; 1104 case SA_XLOCKED: 1105 case SA_XLOCKED | SA_NOTRECURSED: 1106 case SA_XLOCKED | SA_RECURSED: 1107 if (sx_xholder(sx) != curthread) 1108 panic("Lock %s not exclusively locked @ %s:%d\n", 1109 sx->lock_object.lo_name, file, line); 1110 if (sx_recursed(sx)) { 1111 if (what & SA_NOTRECURSED) 1112 panic("Lock %s recursed @ %s:%d\n", 1113 sx->lock_object.lo_name, file, line); 1114 } else if (what & SA_RECURSED) 1115 panic("Lock %s not recursed @ %s:%d\n", 1116 sx->lock_object.lo_name, file, line); 1117 break; 1118 case SA_UNLOCKED: 1119#ifdef WITNESS 1120 witness_assert(&sx->lock_object, what, file, line); 1121#else 1122 /* 1123 * If we hold an exclusve lock fail. We can't 1124 * reliably check to see if we hold a shared lock or 1125 * not. 1126 */ 1127 if (sx_xholder(sx) == curthread) 1128 panic("Lock %s exclusively locked @ %s:%d\n", 1129 sx->lock_object.lo_name, file, line); 1130#endif 1131 break; 1132 default: 1133 panic("Unknown sx lock assertion: %d @ %s:%d", what, file, 1134 line); 1135 } 1136} 1137#endif /* INVARIANT_SUPPORT */ 1138 1139#ifdef DDB 1140static void 1141db_show_sx(struct lock_object *lock) 1142{ 1143 struct thread *td; 1144 struct sx *sx; 1145 1146 sx = (struct sx *)lock; 1147 1148 db_printf(" state: "); 1149 if (sx->sx_lock == SX_LOCK_UNLOCKED) 1150 db_printf("UNLOCKED\n"); 1151 else if (sx->sx_lock == SX_LOCK_DESTROYED) { 1152 db_printf("DESTROYED\n"); 1153 return; 1154 } else if (sx->sx_lock & SX_LOCK_SHARED) 1155 db_printf("SLOCK: %ju\n", (uintmax_t)SX_SHARERS(sx->sx_lock)); 1156 else { 1157 td = sx_xholder(sx); 1158 db_printf("XLOCK: %p (tid %d, pid %d, \"%s\")\n", td, 1159 td->td_tid, td->td_proc->p_pid, td->td_name); 1160 if (sx_recursed(sx)) 1161 db_printf(" recursed: %d\n", sx->sx_recurse); 1162 } 1163 1164 db_printf(" waiters: "); 1165 switch(sx->sx_lock & 1166 (SX_LOCK_SHARED_WAITERS | SX_LOCK_EXCLUSIVE_WAITERS)) { 1167 case SX_LOCK_SHARED_WAITERS: 1168 db_printf("shared\n"); 1169 break; 1170 case SX_LOCK_EXCLUSIVE_WAITERS: 1171 db_printf("exclusive\n"); 1172 break; 1173 case SX_LOCK_SHARED_WAITERS | SX_LOCK_EXCLUSIVE_WAITERS: 1174 db_printf("exclusive and shared\n"); 1175 break; 1176 default: 1177 db_printf("none\n"); 1178 } 1179} 1180 1181/* 1182 * Check to see if a thread that is blocked on a sleep queue is actually 1183 * blocked on an sx lock. If so, output some details and return true. 1184 * If the lock has an exclusive owner, return that in *ownerp. 1185 */ 1186int 1187sx_chain(struct thread *td, struct thread **ownerp) 1188{ 1189 struct sx *sx; 1190 1191 /* 1192 * Check to see if this thread is blocked on an sx lock. 1193 * First, we check the lock class. If that is ok, then we 1194 * compare the lock name against the wait message. 1195 */ 1196 sx = td->td_wchan; 1197 if (LOCK_CLASS(&sx->lock_object) != &lock_class_sx || 1198 sx->lock_object.lo_name != td->td_wmesg) 1199 return (0); 1200 1201 /* We think we have an sx lock, so output some details. */ 1202 db_printf("blocked on sx \"%s\" ", td->td_wmesg); 1203 *ownerp = sx_xholder(sx); 1204 if (sx->sx_lock & SX_LOCK_SHARED) 1205 db_printf("SLOCK (count %ju)\n", 1206 (uintmax_t)SX_SHARERS(sx->sx_lock)); 1207 else 1208 db_printf("XLOCK\n"); 1209 return (1); 1210} 1211#endif 1212