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 70 CTASSERT((SX_NOADAPTIVE & LO_CLASSFLAGS) == SX_NOADAPTIVE); 71 72 #ifdef HWPMC_HOOKS 73 #include <sys/pmckern.h> 74 PMC_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 /* 82 * Variations on DROP_GIANT()/PICKUP_GIANT() for use in this file. We 83 * drop Giant anytime we have to sleep or if we adaptively spin. 84 */ 85 #define GIANT_DECLARE \ 86 int _giantcnt = 0; \ 87 WITNESS_SAVE_DECL(Giant) \ 88 89 #define GIANT_SAVE() do { \ 90 if (mtx_owned(&Giant)) { \ 91 WITNESS_SAVE(&Giant.lock_object, Giant); \ 92 while (mtx_owned(&Giant)) { \ 93 _giantcnt++; \ 94 mtx_unlock(&Giant); \ 95 } \ 96 } \ 97 } while (0) 98 99 #define GIANT_RESTORE() do { \ 100 if (_giantcnt > 0) { \ 101 mtx_assert(&Giant, MA_NOTOWNED); \ 102 while (_giantcnt--) \ 103 mtx_lock(&Giant); \ 104 WITNESS_RESTORE(&Giant.lock_object, Giant); \ 105 } \ 106 } while (0) 107 108 /* 109 * Returns true if an exclusive lock is recursed. It assumes 110 * curthread currently has an exclusive lock. 111 */ 112 #define sx_recurse lock_object.lo_data 113 #define sx_recursed(sx) ((sx)->sx_recurse != 0) 114 115 static void assert_sx(const struct lock_object *lock, int what); 116 #ifdef DDB 117 static void db_show_sx(const struct lock_object *lock); 118 #endif 119 static void lock_sx(struct lock_object *lock, int how); 120 #ifdef KDTRACE_HOOKS 121 static int owner_sx(const struct lock_object *lock, struct thread **owner); 122 #endif 123 static int unlock_sx(struct lock_object *lock); 124 125 struct lock_class lock_class_sx = { 126 .lc_name = "sx", 127 .lc_flags = LC_SLEEPLOCK | LC_SLEEPABLE | LC_RECURSABLE | LC_UPGRADABLE, 128 .lc_assert = assert_sx, 129 #ifdef DDB 130 .lc_ddb_show = db_show_sx, 131 #endif 132 .lc_lock = lock_sx, 133 .lc_unlock = unlock_sx, 134 #ifdef KDTRACE_HOOKS 135 .lc_owner = owner_sx, 136 #endif 137 }; 138 139 #ifndef INVARIANTS 140 #define _sx_assert(sx, what, file, line) 141 #endif 142 143 #ifdef ADAPTIVE_SX 144 static u_int asx_retries = 10; 145 static u_int asx_loops = 10000; 146 static SYSCTL_NODE(_debug, OID_AUTO, sx, CTLFLAG_RD, NULL, "sxlock debugging"); 147 SYSCTL_UINT(_debug_sx, OID_AUTO, retries, CTLFLAG_RW, &asx_retries, 0, ""); 148 SYSCTL_UINT(_debug_sx, OID_AUTO, loops, CTLFLAG_RW, &asx_loops, 0, ""); 149 #endif 150 151 void 152 assert_sx(const struct lock_object *lock, int what) 153 { 154 155 sx_assert((const struct sx *)lock, what); 156 } 157 158 void 159 lock_sx(struct lock_object *lock, int how) 160 { 161 struct sx *sx; 162 163 sx = (struct sx *)lock; 164 if (how) 165 sx_xlock(sx); 166 else 167 sx_slock(sx); 168 } 169 170 int 171 unlock_sx(struct lock_object *lock) 172 { 173 struct sx *sx; 174 175 sx = (struct sx *)lock; 176 sx_assert(sx, SA_LOCKED | SA_NOTRECURSED); 177 if (sx_xlocked(sx)) { 178 sx_xunlock(sx); 179 return (1); 180 } else { 181 sx_sunlock(sx); 182 return (0); 183 } 184 } 185 186 #ifdef KDTRACE_HOOKS 187 int 188 owner_sx(const struct lock_object *lock, struct thread **owner) 189 { 190 const struct sx *sx = (const struct sx *)lock; 191 uintptr_t x = sx->sx_lock; 192 193 *owner = (struct thread *)SX_OWNER(x); 194 return ((x & SX_LOCK_SHARED) != 0 ? (SX_SHARERS(x) != 0) : 195 (*owner != NULL)); 196 } 197 #endif 198 199 void 200 sx_sysinit(void *arg) 201 { 202 struct sx_args *sargs = arg; 203 204 sx_init_flags(sargs->sa_sx, sargs->sa_desc, sargs->sa_flags); 205 } 206 207 void 208 sx_init_flags(struct sx *sx, const char *description, int opts) 209 { 210 int flags; 211 212 MPASS((opts & ~(SX_QUIET | SX_RECURSE | SX_NOWITNESS | SX_DUPOK | 213 SX_NOPROFILE | SX_NOADAPTIVE)) == 0); 214 ASSERT_ATOMIC_LOAD_PTR(sx->sx_lock, 215 ("%s: sx_lock not aligned for %s: %p", __func__, description, 216 &sx->sx_lock)); 217 218 flags = LO_SLEEPABLE | LO_UPGRADABLE; 219 if (opts & SX_DUPOK) 220 flags |= LO_DUPOK; 221 if (opts & SX_NOPROFILE) 222 flags |= LO_NOPROFILE; 223 if (!(opts & SX_NOWITNESS)) 224 flags |= LO_WITNESS; 225 if (opts & SX_RECURSE) 226 flags |= LO_RECURSABLE; 227 if (opts & SX_QUIET) 228 flags |= LO_QUIET; 229 230 flags |= opts & SX_NOADAPTIVE; 231 lock_init(&sx->lock_object, &lock_class_sx, description, NULL, flags); 232 sx->sx_lock = SX_LOCK_UNLOCKED; 233 sx->sx_recurse = 0; 234 } 235 236 void 237 sx_destroy(struct sx *sx) 238 { 239 240 KASSERT(sx->sx_lock == SX_LOCK_UNLOCKED, ("sx lock still held")); 241 KASSERT(sx->sx_recurse == 0, ("sx lock still recursed")); 242 sx->sx_lock = SX_LOCK_DESTROYED; 243 lock_destroy(&sx->lock_object); 244 } 245 246 int 247 _sx_slock(struct sx *sx, int opts, const char *file, int line) 248 { 249 int error = 0; 250 251 if (SCHEDULER_STOPPED()) 252 return (0); 253 KASSERT(kdb_active != 0 || !TD_IS_IDLETHREAD(curthread), 254 ("sx_slock() by idle thread %p on sx %s @ %s:%d", 255 curthread, sx->lock_object.lo_name, file, line)); 256 KASSERT(sx->sx_lock != SX_LOCK_DESTROYED, 257 ("sx_slock() of destroyed sx @ %s:%d", file, line)); 258 WITNESS_CHECKORDER(&sx->lock_object, LOP_NEWORDER, file, line, NULL); 259 error = __sx_slock(sx, opts, file, line); 260 if (!error) { 261 LOCK_LOG_LOCK("SLOCK", &sx->lock_object, 0, 0, file, line); 262 WITNESS_LOCK(&sx->lock_object, 0, file, line); 263 curthread->td_locks++; 264 } 265 266 return (error); 267 } 268 269 int 270 sx_try_slock_(struct sx *sx, const char *file, int line) 271 { 272 uintptr_t x; 273 274 if (SCHEDULER_STOPPED()) 275 return (1); 276 277 KASSERT(kdb_active != 0 || !TD_IS_IDLETHREAD(curthread), 278 ("sx_try_slock() by idle thread %p on sx %s @ %s:%d", 279 curthread, sx->lock_object.lo_name, file, line)); 280 281 for (;;) { 282 x = sx->sx_lock; 283 KASSERT(x != SX_LOCK_DESTROYED, 284 ("sx_try_slock() of destroyed sx @ %s:%d", file, line)); 285 if (!(x & SX_LOCK_SHARED)) 286 break; 287 if (atomic_cmpset_acq_ptr(&sx->sx_lock, x, x + SX_ONE_SHARER)) { 288 LOCK_LOG_TRY("SLOCK", &sx->lock_object, 0, 1, file, line); 289 WITNESS_LOCK(&sx->lock_object, LOP_TRYLOCK, file, line); 290 curthread->td_locks++; 291 return (1); 292 } 293 } 294 295 LOCK_LOG_TRY("SLOCK", &sx->lock_object, 0, 0, file, line); 296 return (0); 297 } 298 299 int 300 _sx_xlock(struct sx *sx, int opts, const char *file, int line) 301 { 302 int error = 0; 303 304 if (SCHEDULER_STOPPED()) 305 return (0); 306 KASSERT(kdb_active != 0 || !TD_IS_IDLETHREAD(curthread), 307 ("sx_xlock() by idle thread %p on sx %s @ %s:%d", 308 curthread, sx->lock_object.lo_name, file, line)); 309 KASSERT(sx->sx_lock != SX_LOCK_DESTROYED, 310 ("sx_xlock() of destroyed sx @ %s:%d", file, line)); 311 WITNESS_CHECKORDER(&sx->lock_object, LOP_NEWORDER | LOP_EXCLUSIVE, file, 312 line, NULL); 313 error = __sx_xlock(sx, curthread, opts, file, line); 314 if (!error) { 315 LOCK_LOG_LOCK("XLOCK", &sx->lock_object, 0, sx->sx_recurse, 316 file, line); 317 WITNESS_LOCK(&sx->lock_object, LOP_EXCLUSIVE, file, line); 318 curthread->td_locks++; 319 } 320 321 return (error); 322 } 323 324 int 325 sx_try_xlock_(struct sx *sx, const char *file, int line) 326 { 327 int rval; 328 329 if (SCHEDULER_STOPPED()) 330 return (1); 331 332 KASSERT(kdb_active != 0 || !TD_IS_IDLETHREAD(curthread), 333 ("sx_try_xlock() by idle thread %p on sx %s @ %s:%d", 334 curthread, sx->lock_object.lo_name, file, line)); 335 KASSERT(sx->sx_lock != SX_LOCK_DESTROYED, 336 ("sx_try_xlock() of destroyed sx @ %s:%d", file, line)); 337 338 if (sx_xlocked(sx) && 339 (sx->lock_object.lo_flags & LO_RECURSABLE) != 0) { 340 sx->sx_recurse++; 341 atomic_set_ptr(&sx->sx_lock, SX_LOCK_RECURSED); 342 rval = 1; 343 } else 344 rval = atomic_cmpset_acq_ptr(&sx->sx_lock, SX_LOCK_UNLOCKED, 345 (uintptr_t)curthread); 346 LOCK_LOG_TRY("XLOCK", &sx->lock_object, 0, rval, file, line); 347 if (rval) { 348 WITNESS_LOCK(&sx->lock_object, LOP_EXCLUSIVE | LOP_TRYLOCK, 349 file, line); 350 curthread->td_locks++; 351 } 352 353 return (rval); 354 } 355 356 void 357 _sx_sunlock(struct sx *sx, const char *file, int line) 358 { 359 360 if (SCHEDULER_STOPPED()) 361 return; 362 KASSERT(sx->sx_lock != SX_LOCK_DESTROYED, 363 ("sx_sunlock() of destroyed sx @ %s:%d", file, line)); 364 _sx_assert(sx, SA_SLOCKED, file, line); 365 WITNESS_UNLOCK(&sx->lock_object, 0, file, line); 366 LOCK_LOG_LOCK("SUNLOCK", &sx->lock_object, 0, 0, file, line); 367 __sx_sunlock(sx, file, line); 368 LOCKSTAT_PROFILE_RELEASE_LOCK(LS_SX_SUNLOCK_RELEASE, sx); 369 curthread->td_locks--; 370 } 371 372 void 373 _sx_xunlock(struct sx *sx, const char *file, int line) 374 { 375 376 if (SCHEDULER_STOPPED()) 377 return; 378 KASSERT(sx->sx_lock != SX_LOCK_DESTROYED, 379 ("sx_xunlock() of destroyed sx @ %s:%d", file, line)); 380 _sx_assert(sx, SA_XLOCKED, file, line); 381 WITNESS_UNLOCK(&sx->lock_object, LOP_EXCLUSIVE, file, line); 382 LOCK_LOG_LOCK("XUNLOCK", &sx->lock_object, 0, sx->sx_recurse, file, 383 line); 384 if (!sx_recursed(sx)) 385 LOCKSTAT_PROFILE_RELEASE_LOCK(LS_SX_XUNLOCK_RELEASE, sx); 386 __sx_xunlock(sx, curthread, file, line); 387 curthread->td_locks--; 388 } 389 390 /* 391 * Try to do a non-blocking upgrade from a shared lock to an exclusive lock. 392 * This will only succeed if this thread holds a single shared lock. 393 * Return 1 if if the upgrade succeed, 0 otherwise. 394 */ 395 int 396 sx_try_upgrade_(struct sx *sx, const char *file, int line) 397 { 398 uintptr_t x; 399 int success; 400 401 if (SCHEDULER_STOPPED()) 402 return (1); 403 404 KASSERT(sx->sx_lock != SX_LOCK_DESTROYED, 405 ("sx_try_upgrade() of destroyed sx @ %s:%d", file, line)); 406 _sx_assert(sx, SA_SLOCKED, file, line); 407 408 /* 409 * Try to switch from one shared lock to an exclusive lock. We need 410 * to maintain the SX_LOCK_EXCLUSIVE_WAITERS flag if set so that 411 * we will wake up the exclusive waiters when we drop the lock. 412 */ 413 x = sx->sx_lock & SX_LOCK_EXCLUSIVE_WAITERS; 414 success = atomic_cmpset_ptr(&sx->sx_lock, SX_SHARERS_LOCK(1) | x, 415 (uintptr_t)curthread | x); 416 LOCK_LOG_TRY("XUPGRADE", &sx->lock_object, 0, success, file, line); 417 if (success) { 418 WITNESS_UPGRADE(&sx->lock_object, LOP_EXCLUSIVE | LOP_TRYLOCK, 419 file, line); 420 LOCKSTAT_RECORD0(LS_SX_TRYUPGRADE_UPGRADE, sx); 421 } 422 return (success); 423 } 424 425 /* 426 * Downgrade an unrecursed exclusive lock into a single shared lock. 427 */ 428 void 429 sx_downgrade_(struct sx *sx, const char *file, int line) 430 { 431 uintptr_t x; 432 int wakeup_swapper; 433 434 if (SCHEDULER_STOPPED()) 435 return; 436 437 KASSERT(sx->sx_lock != SX_LOCK_DESTROYED, 438 ("sx_downgrade() of destroyed sx @ %s:%d", file, line)); 439 _sx_assert(sx, SA_XLOCKED | SA_NOTRECURSED, file, line); 440 #ifndef INVARIANTS 441 if (sx_recursed(sx)) 442 panic("downgrade of a recursed lock"); 443 #endif 444 445 WITNESS_DOWNGRADE(&sx->lock_object, 0, file, line); 446 447 /* 448 * Try to switch from an exclusive lock with no shared waiters 449 * to one sharer with no shared waiters. If there are 450 * exclusive waiters, we don't need to lock the sleep queue so 451 * long as we preserve the flag. We do one quick try and if 452 * that fails we grab the sleepq lock to keep the flags from 453 * changing and do it the slow way. 454 * 455 * We have to lock the sleep queue if there are shared waiters 456 * so we can wake them up. 457 */ 458 x = sx->sx_lock; 459 if (!(x & SX_LOCK_SHARED_WAITERS) && 460 atomic_cmpset_rel_ptr(&sx->sx_lock, x, SX_SHARERS_LOCK(1) | 461 (x & SX_LOCK_EXCLUSIVE_WAITERS))) { 462 LOCK_LOG_LOCK("XDOWNGRADE", &sx->lock_object, 0, 0, file, line); 463 return; 464 } 465 466 /* 467 * Lock the sleep queue so we can read the waiters bits 468 * without any races and wakeup any shared waiters. 469 */ 470 sleepq_lock(&sx->lock_object); 471 472 /* 473 * Preserve SX_LOCK_EXCLUSIVE_WAITERS while downgraded to a single 474 * shared lock. If there are any shared waiters, wake them up. 475 */ 476 wakeup_swapper = 0; 477 x = sx->sx_lock; 478 atomic_store_rel_ptr(&sx->sx_lock, SX_SHARERS_LOCK(1) | 479 (x & SX_LOCK_EXCLUSIVE_WAITERS)); 480 if (x & SX_LOCK_SHARED_WAITERS) 481 wakeup_swapper = sleepq_broadcast(&sx->lock_object, SLEEPQ_SX, 482 0, SQ_SHARED_QUEUE); 483 sleepq_release(&sx->lock_object); 484 485 LOCK_LOG_LOCK("XDOWNGRADE", &sx->lock_object, 0, 0, file, line); 486 LOCKSTAT_RECORD0(LS_SX_DOWNGRADE_DOWNGRADE, sx); 487 488 if (wakeup_swapper) 489 kick_proc0(); 490 } 491 492 /* 493 * This function represents the so-called 'hard case' for sx_xlock 494 * operation. All 'easy case' failures are redirected to this. Note 495 * that ideally this would be a static function, but it needs to be 496 * accessible from at least sx.h. 497 */ 498 int 499 _sx_xlock_hard(struct sx *sx, uintptr_t tid, int opts, const char *file, 500 int line) 501 { 502 GIANT_DECLARE; 503 #ifdef ADAPTIVE_SX 504 volatile struct thread *owner; 505 u_int i, spintries = 0; 506 #endif 507 uintptr_t x; 508 #ifdef LOCK_PROFILING 509 uint64_t waittime = 0; 510 int contested = 0; 511 #endif 512 int error = 0; 513 #ifdef KDTRACE_HOOKS 514 uint64_t spin_cnt = 0; 515 uint64_t sleep_cnt = 0; 516 int64_t sleep_time = 0; 517 #endif 518 519 if (SCHEDULER_STOPPED()) 520 return (0); 521 522 /* If we already hold an exclusive lock, then recurse. */ 523 if (sx_xlocked(sx)) { 524 KASSERT((sx->lock_object.lo_flags & LO_RECURSABLE) != 0, 525 ("_sx_xlock_hard: recursed on non-recursive sx %s @ %s:%d\n", 526 sx->lock_object.lo_name, file, line)); 527 sx->sx_recurse++; 528 atomic_set_ptr(&sx->sx_lock, SX_LOCK_RECURSED); 529 if (LOCK_LOG_TEST(&sx->lock_object, 0)) 530 CTR2(KTR_LOCK, "%s: %p recursing", __func__, sx); 531 return (0); 532 } 533 534 if (LOCK_LOG_TEST(&sx->lock_object, 0)) 535 CTR5(KTR_LOCK, "%s: %s contested (lock=%p) at %s:%d", __func__, 536 sx->lock_object.lo_name, (void *)sx->sx_lock, file, line); 537 538 while (!atomic_cmpset_acq_ptr(&sx->sx_lock, SX_LOCK_UNLOCKED, tid)) { 539 #ifdef KDTRACE_HOOKS 540 spin_cnt++; 541 #endif 542 #ifdef HWPMC_HOOKS 543 PMC_SOFT_CALL( , , lock, failed); 544 #endif 545 lock_profile_obtain_lock_failed(&sx->lock_object, &contested, 546 &waittime); 547 #ifdef ADAPTIVE_SX 548 /* 549 * If the lock is write locked and the owner is 550 * running on another CPU, spin until the owner stops 551 * running or the state of the lock changes. 552 */ 553 x = sx->sx_lock; 554 if ((sx->lock_object.lo_flags & SX_NOADAPTIVE) == 0) { 555 if ((x & SX_LOCK_SHARED) == 0) { 556 x = SX_OWNER(x); 557 owner = (struct thread *)x; 558 if (TD_IS_RUNNING(owner)) { 559 if (LOCK_LOG_TEST(&sx->lock_object, 0)) 560 CTR3(KTR_LOCK, 561 "%s: spinning on %p held by %p", 562 __func__, sx, owner); 563 GIANT_SAVE(); 564 while (SX_OWNER(sx->sx_lock) == x && 565 TD_IS_RUNNING(owner)) { 566 cpu_spinwait(); 567 #ifdef KDTRACE_HOOKS 568 spin_cnt++; 569 #endif 570 } 571 continue; 572 } 573 } else if (SX_SHARERS(x) && spintries < asx_retries) { 574 GIANT_SAVE(); 575 spintries++; 576 for (i = 0; i < asx_loops; i++) { 577 if (LOCK_LOG_TEST(&sx->lock_object, 0)) 578 CTR4(KTR_LOCK, 579 "%s: shared spinning on %p with %u and %u", 580 __func__, sx, spintries, i); 581 x = sx->sx_lock; 582 if ((x & SX_LOCK_SHARED) == 0 || 583 SX_SHARERS(x) == 0) 584 break; 585 cpu_spinwait(); 586 #ifdef KDTRACE_HOOKS 587 spin_cnt++; 588 #endif 589 } 590 if (i != asx_loops) 591 continue; 592 } 593 } 594 #endif 595 596 sleepq_lock(&sx->lock_object); 597 x = sx->sx_lock; 598 599 /* 600 * If the lock was released while spinning on the 601 * sleep queue chain lock, try again. 602 */ 603 if (x == SX_LOCK_UNLOCKED) { 604 sleepq_release(&sx->lock_object); 605 continue; 606 } 607 608 #ifdef ADAPTIVE_SX 609 /* 610 * The current lock owner might have started executing 611 * on another CPU (or the lock could have changed 612 * owners) while we were waiting on the sleep queue 613 * chain lock. If so, drop the sleep queue lock and try 614 * again. 615 */ 616 if (!(x & SX_LOCK_SHARED) && 617 (sx->lock_object.lo_flags & SX_NOADAPTIVE) == 0) { 618 owner = (struct thread *)SX_OWNER(x); 619 if (TD_IS_RUNNING(owner)) { 620 sleepq_release(&sx->lock_object); 621 continue; 622 } 623 } 624 #endif 625 626 /* 627 * If an exclusive lock was released with both shared 628 * and exclusive waiters and a shared waiter hasn't 629 * woken up and acquired the lock yet, sx_lock will be 630 * set to SX_LOCK_UNLOCKED | SX_LOCK_EXCLUSIVE_WAITERS. 631 * If we see that value, try to acquire it once. Note 632 * that we have to preserve SX_LOCK_EXCLUSIVE_WAITERS 633 * as there are other exclusive waiters still. If we 634 * fail, restart the loop. 635 */ 636 if (x == (SX_LOCK_UNLOCKED | SX_LOCK_EXCLUSIVE_WAITERS)) { 637 if (atomic_cmpset_acq_ptr(&sx->sx_lock, 638 SX_LOCK_UNLOCKED | SX_LOCK_EXCLUSIVE_WAITERS, 639 tid | SX_LOCK_EXCLUSIVE_WAITERS)) { 640 sleepq_release(&sx->lock_object); 641 CTR2(KTR_LOCK, "%s: %p claimed by new writer", 642 __func__, sx); 643 break; 644 } 645 sleepq_release(&sx->lock_object); 646 continue; 647 } 648 649 /* 650 * Try to set the SX_LOCK_EXCLUSIVE_WAITERS. If we fail, 651 * than loop back and retry. 652 */ 653 if (!(x & SX_LOCK_EXCLUSIVE_WAITERS)) { 654 if (!atomic_cmpset_ptr(&sx->sx_lock, x, 655 x | SX_LOCK_EXCLUSIVE_WAITERS)) { 656 sleepq_release(&sx->lock_object); 657 continue; 658 } 659 if (LOCK_LOG_TEST(&sx->lock_object, 0)) 660 CTR2(KTR_LOCK, "%s: %p set excl waiters flag", 661 __func__, sx); 662 } 663 664 /* 665 * Since we have been unable to acquire the exclusive 666 * lock and the exclusive waiters flag is set, we have 667 * to sleep. 668 */ 669 if (LOCK_LOG_TEST(&sx->lock_object, 0)) 670 CTR2(KTR_LOCK, "%s: %p blocking on sleep queue", 671 __func__, sx); 672 673 #ifdef KDTRACE_HOOKS 674 sleep_time -= lockstat_nsecs(); 675 #endif 676 GIANT_SAVE(); 677 sleepq_add(&sx->lock_object, NULL, sx->lock_object.lo_name, 678 SLEEPQ_SX | ((opts & SX_INTERRUPTIBLE) ? 679 SLEEPQ_INTERRUPTIBLE : 0), SQ_EXCLUSIVE_QUEUE); 680 if (!(opts & SX_INTERRUPTIBLE)) 681 sleepq_wait(&sx->lock_object, 0); 682 else 683 error = sleepq_wait_sig(&sx->lock_object, 0); 684 #ifdef KDTRACE_HOOKS 685 sleep_time += lockstat_nsecs(); 686 sleep_cnt++; 687 #endif 688 if (error) { 689 if (LOCK_LOG_TEST(&sx->lock_object, 0)) 690 CTR2(KTR_LOCK, 691 "%s: interruptible sleep by %p suspended by signal", 692 __func__, sx); 693 break; 694 } 695 if (LOCK_LOG_TEST(&sx->lock_object, 0)) 696 CTR2(KTR_LOCK, "%s: %p resuming from sleep queue", 697 __func__, sx); 698 } 699 700 GIANT_RESTORE(); 701 if (!error) 702 LOCKSTAT_PROFILE_OBTAIN_LOCK_SUCCESS(LS_SX_XLOCK_ACQUIRE, sx, 703 contested, waittime, file, line); 704 #ifdef KDTRACE_HOOKS 705 if (sleep_time) 706 LOCKSTAT_RECORD1(LS_SX_XLOCK_BLOCK, sx, sleep_time); 707 if (spin_cnt > sleep_cnt) 708 LOCKSTAT_RECORD1(LS_SX_XLOCK_SPIN, sx, (spin_cnt - sleep_cnt)); 709 #endif 710 return (error); 711 } 712 713 /* 714 * This function represents the so-called 'hard case' for sx_xunlock 715 * operation. All 'easy case' failures are redirected to this. Note 716 * that ideally this would be a static function, but it needs to be 717 * accessible from at least sx.h. 718 */ 719 void 720 _sx_xunlock_hard(struct sx *sx, uintptr_t tid, const char *file, int line) 721 { 722 uintptr_t x; 723 int queue, wakeup_swapper; 724 725 if (SCHEDULER_STOPPED()) 726 return; 727 728 MPASS(!(sx->sx_lock & SX_LOCK_SHARED)); 729 730 /* If the lock is recursed, then unrecurse one level. */ 731 if (sx_xlocked(sx) && sx_recursed(sx)) { 732 if ((--sx->sx_recurse) == 0) 733 atomic_clear_ptr(&sx->sx_lock, SX_LOCK_RECURSED); 734 if (LOCK_LOG_TEST(&sx->lock_object, 0)) 735 CTR2(KTR_LOCK, "%s: %p unrecursing", __func__, sx); 736 return; 737 } 738 MPASS(sx->sx_lock & (SX_LOCK_SHARED_WAITERS | 739 SX_LOCK_EXCLUSIVE_WAITERS)); 740 if (LOCK_LOG_TEST(&sx->lock_object, 0)) 741 CTR2(KTR_LOCK, "%s: %p contested", __func__, sx); 742 743 sleepq_lock(&sx->lock_object); 744 x = SX_LOCK_UNLOCKED; 745 746 /* 747 * The wake up algorithm here is quite simple and probably not 748 * ideal. It gives precedence to shared waiters if they are 749 * present. For this condition, we have to preserve the 750 * state of the exclusive waiters flag. 751 * If interruptible sleeps left the shared queue empty avoid a 752 * starvation for the threads sleeping on the exclusive queue by giving 753 * them precedence and cleaning up the shared waiters bit anyway. 754 */ 755 if ((sx->sx_lock & SX_LOCK_SHARED_WAITERS) != 0 && 756 sleepq_sleepcnt(&sx->lock_object, SQ_SHARED_QUEUE) != 0) { 757 queue = SQ_SHARED_QUEUE; 758 x |= (sx->sx_lock & SX_LOCK_EXCLUSIVE_WAITERS); 759 } else 760 queue = SQ_EXCLUSIVE_QUEUE; 761 762 /* Wake up all the waiters for the specific queue. */ 763 if (LOCK_LOG_TEST(&sx->lock_object, 0)) 764 CTR3(KTR_LOCK, "%s: %p waking up all threads on %s queue", 765 __func__, sx, queue == SQ_SHARED_QUEUE ? "shared" : 766 "exclusive"); 767 atomic_store_rel_ptr(&sx->sx_lock, x); 768 wakeup_swapper = sleepq_broadcast(&sx->lock_object, SLEEPQ_SX, 0, 769 queue); 770 sleepq_release(&sx->lock_object); 771 if (wakeup_swapper) 772 kick_proc0(); 773 } 774 775 /* 776 * This function represents the so-called 'hard case' for sx_slock 777 * operation. All 'easy case' failures are redirected to this. Note 778 * that ideally this would be a static function, but it needs to be 779 * accessible from at least sx.h. 780 */ 781 int 782 _sx_slock_hard(struct sx *sx, int opts, const char *file, int line) 783 { 784 GIANT_DECLARE; 785 #ifdef ADAPTIVE_SX 786 volatile struct thread *owner; 787 #endif 788 #ifdef LOCK_PROFILING 789 uint64_t waittime = 0; 790 int contested = 0; 791 #endif 792 uintptr_t x; 793 int error = 0; 794 #ifdef KDTRACE_HOOKS 795 uint64_t spin_cnt = 0; 796 uint64_t sleep_cnt = 0; 797 int64_t sleep_time = 0; 798 #endif 799 800 if (SCHEDULER_STOPPED()) 801 return (0); 802 803 /* 804 * As with rwlocks, we don't make any attempt to try to block 805 * shared locks once there is an exclusive waiter. 806 */ 807 for (;;) { 808 #ifdef KDTRACE_HOOKS 809 spin_cnt++; 810 #endif 811 x = sx->sx_lock; 812 813 /* 814 * If no other thread has an exclusive lock then try to bump up 815 * the count of sharers. Since we have to preserve the state 816 * of SX_LOCK_EXCLUSIVE_WAITERS, if we fail to acquire the 817 * shared lock loop back and retry. 818 */ 819 if (x & SX_LOCK_SHARED) { 820 MPASS(!(x & SX_LOCK_SHARED_WAITERS)); 821 if (atomic_cmpset_acq_ptr(&sx->sx_lock, x, 822 x + SX_ONE_SHARER)) { 823 if (LOCK_LOG_TEST(&sx->lock_object, 0)) 824 CTR4(KTR_LOCK, 825 "%s: %p succeed %p -> %p", __func__, 826 sx, (void *)x, 827 (void *)(x + SX_ONE_SHARER)); 828 break; 829 } 830 continue; 831 } 832 #ifdef HWPMC_HOOKS 833 PMC_SOFT_CALL( , , lock, failed); 834 #endif 835 lock_profile_obtain_lock_failed(&sx->lock_object, &contested, 836 &waittime); 837 838 #ifdef ADAPTIVE_SX 839 /* 840 * If the owner is running on another CPU, spin until 841 * the owner stops running or the state of the lock 842 * changes. 843 */ 844 if ((sx->lock_object.lo_flags & SX_NOADAPTIVE) == 0) { 845 x = SX_OWNER(x); 846 owner = (struct thread *)x; 847 if (TD_IS_RUNNING(owner)) { 848 if (LOCK_LOG_TEST(&sx->lock_object, 0)) 849 CTR3(KTR_LOCK, 850 "%s: spinning on %p held by %p", 851 __func__, sx, owner); 852 GIANT_SAVE(); 853 while (SX_OWNER(sx->sx_lock) == x && 854 TD_IS_RUNNING(owner)) { 855 #ifdef KDTRACE_HOOKS 856 spin_cnt++; 857 #endif 858 cpu_spinwait(); 859 } 860 continue; 861 } 862 } 863 #endif 864 865 /* 866 * Some other thread already has an exclusive lock, so 867 * start the process of blocking. 868 */ 869 sleepq_lock(&sx->lock_object); 870 x = sx->sx_lock; 871 872 /* 873 * The lock could have been released while we spun. 874 * In this case loop back and retry. 875 */ 876 if (x & SX_LOCK_SHARED) { 877 sleepq_release(&sx->lock_object); 878 continue; 879 } 880 881 #ifdef ADAPTIVE_SX 882 /* 883 * If the owner is running on another CPU, spin until 884 * the owner stops running or the state of the lock 885 * changes. 886 */ 887 if (!(x & SX_LOCK_SHARED) && 888 (sx->lock_object.lo_flags & SX_NOADAPTIVE) == 0) { 889 owner = (struct thread *)SX_OWNER(x); 890 if (TD_IS_RUNNING(owner)) { 891 sleepq_release(&sx->lock_object); 892 continue; 893 } 894 } 895 #endif 896 897 /* 898 * Try to set the SX_LOCK_SHARED_WAITERS flag. If we 899 * fail to set it drop the sleep queue lock and loop 900 * back. 901 */ 902 if (!(x & SX_LOCK_SHARED_WAITERS)) { 903 if (!atomic_cmpset_ptr(&sx->sx_lock, x, 904 x | SX_LOCK_SHARED_WAITERS)) { 905 sleepq_release(&sx->lock_object); 906 continue; 907 } 908 if (LOCK_LOG_TEST(&sx->lock_object, 0)) 909 CTR2(KTR_LOCK, "%s: %p set shared waiters flag", 910 __func__, sx); 911 } 912 913 /* 914 * Since we have been unable to acquire the shared lock, 915 * we have to sleep. 916 */ 917 if (LOCK_LOG_TEST(&sx->lock_object, 0)) 918 CTR2(KTR_LOCK, "%s: %p blocking on sleep queue", 919 __func__, sx); 920 921 #ifdef KDTRACE_HOOKS 922 sleep_time -= lockstat_nsecs(); 923 #endif 924 GIANT_SAVE(); 925 sleepq_add(&sx->lock_object, NULL, sx->lock_object.lo_name, 926 SLEEPQ_SX | ((opts & SX_INTERRUPTIBLE) ? 927 SLEEPQ_INTERRUPTIBLE : 0), SQ_SHARED_QUEUE); 928 if (!(opts & SX_INTERRUPTIBLE)) 929 sleepq_wait(&sx->lock_object, 0); 930 else 931 error = sleepq_wait_sig(&sx->lock_object, 0); 932 #ifdef KDTRACE_HOOKS 933 sleep_time += lockstat_nsecs(); 934 sleep_cnt++; 935 #endif 936 if (error) { 937 if (LOCK_LOG_TEST(&sx->lock_object, 0)) 938 CTR2(KTR_LOCK, 939 "%s: interruptible sleep by %p suspended by signal", 940 __func__, sx); 941 break; 942 } 943 if (LOCK_LOG_TEST(&sx->lock_object, 0)) 944 CTR2(KTR_LOCK, "%s: %p resuming from sleep queue", 945 __func__, sx); 946 } 947 if (error == 0) 948 LOCKSTAT_PROFILE_OBTAIN_LOCK_SUCCESS(LS_SX_SLOCK_ACQUIRE, sx, 949 contested, waittime, file, line); 950 #ifdef KDTRACE_HOOKS 951 if (sleep_time) 952 LOCKSTAT_RECORD1(LS_SX_XLOCK_BLOCK, sx, sleep_time); 953 if (spin_cnt > sleep_cnt) 954 LOCKSTAT_RECORD1(LS_SX_XLOCK_SPIN, sx, (spin_cnt - sleep_cnt)); 955 #endif 956 GIANT_RESTORE(); 957 return (error); 958 } 959 960 /* 961 * This function represents the so-called 'hard case' for sx_sunlock 962 * operation. All 'easy case' failures are redirected to this. Note 963 * that ideally this would be a static function, but it needs to be 964 * accessible from at least sx.h. 965 */ 966 void 967 _sx_sunlock_hard(struct sx *sx, const char *file, int line) 968 { 969 uintptr_t x; 970 int wakeup_swapper; 971 972 if (SCHEDULER_STOPPED()) 973 return; 974 975 for (;;) { 976 x = sx->sx_lock; 977 978 /* 979 * We should never have sharers while at least one thread 980 * holds a shared lock. 981 */ 982 KASSERT(!(x & SX_LOCK_SHARED_WAITERS), 983 ("%s: waiting sharers", __func__)); 984 985 /* 986 * See if there is more than one shared lock held. If 987 * so, just drop one and return. 988 */ 989 if (SX_SHARERS(x) > 1) { 990 if (atomic_cmpset_rel_ptr(&sx->sx_lock, x, 991 x - SX_ONE_SHARER)) { 992 if (LOCK_LOG_TEST(&sx->lock_object, 0)) 993 CTR4(KTR_LOCK, 994 "%s: %p succeeded %p -> %p", 995 __func__, sx, (void *)x, 996 (void *)(x - SX_ONE_SHARER)); 997 break; 998 } 999 continue; 1000 } 1001 1002 /* 1003 * If there aren't any waiters for an exclusive lock, 1004 * then try to drop it quickly. 1005 */ 1006 if (!(x & SX_LOCK_EXCLUSIVE_WAITERS)) { 1007 MPASS(x == SX_SHARERS_LOCK(1)); 1008 if (atomic_cmpset_rel_ptr(&sx->sx_lock, 1009 SX_SHARERS_LOCK(1), SX_LOCK_UNLOCKED)) { 1010 if (LOCK_LOG_TEST(&sx->lock_object, 0)) 1011 CTR2(KTR_LOCK, "%s: %p last succeeded", 1012 __func__, sx); 1013 break; 1014 } 1015 continue; 1016 } 1017 1018 /* 1019 * At this point, there should just be one sharer with 1020 * exclusive waiters. 1021 */ 1022 MPASS(x == (SX_SHARERS_LOCK(1) | SX_LOCK_EXCLUSIVE_WAITERS)); 1023 1024 sleepq_lock(&sx->lock_object); 1025 1026 /* 1027 * Wake up semantic here is quite simple: 1028 * Just wake up all the exclusive waiters. 1029 * Note that the state of the lock could have changed, 1030 * so if it fails loop back and retry. 1031 */ 1032 if (!atomic_cmpset_rel_ptr(&sx->sx_lock, 1033 SX_SHARERS_LOCK(1) | SX_LOCK_EXCLUSIVE_WAITERS, 1034 SX_LOCK_UNLOCKED)) { 1035 sleepq_release(&sx->lock_object); 1036 continue; 1037 } 1038 if (LOCK_LOG_TEST(&sx->lock_object, 0)) 1039 CTR2(KTR_LOCK, "%s: %p waking up all thread on" 1040 "exclusive queue", __func__, sx); 1041 wakeup_swapper = sleepq_broadcast(&sx->lock_object, SLEEPQ_SX, 1042 0, SQ_EXCLUSIVE_QUEUE); 1043 sleepq_release(&sx->lock_object); 1044 if (wakeup_swapper) 1045 kick_proc0(); 1046 break; 1047 } 1048 } 1049 1050 #ifdef INVARIANT_SUPPORT 1051 #ifndef INVARIANTS 1052 #undef _sx_assert 1053 #endif 1054 1055 /* 1056 * In the non-WITNESS case, sx_assert() can only detect that at least 1057 * *some* thread owns an slock, but it cannot guarantee that *this* 1058 * thread owns an slock. 1059 */ 1060 void 1061 _sx_assert(const struct sx *sx, int what, const char *file, int line) 1062 { 1063 #ifndef WITNESS 1064 int slocked = 0; 1065 #endif 1066 1067 if (panicstr != NULL) 1068 return; 1069 switch (what) { 1070 case SA_SLOCKED: 1071 case SA_SLOCKED | SA_NOTRECURSED: 1072 case SA_SLOCKED | SA_RECURSED: 1073 #ifndef WITNESS 1074 slocked = 1; 1075 /* FALLTHROUGH */ 1076 #endif 1077 case SA_LOCKED: 1078 case SA_LOCKED | SA_NOTRECURSED: 1079 case SA_LOCKED | SA_RECURSED: 1080 #ifdef WITNESS 1081 witness_assert(&sx->lock_object, what, file, line); 1082 #else 1083 /* 1084 * If some other thread has an exclusive lock or we 1085 * have one and are asserting a shared lock, fail. 1086 * Also, if no one has a lock at all, fail. 1087 */ 1088 if (sx->sx_lock == SX_LOCK_UNLOCKED || 1089 (!(sx->sx_lock & SX_LOCK_SHARED) && (slocked || 1090 sx_xholder(sx) != curthread))) 1091 panic("Lock %s not %slocked @ %s:%d\n", 1092 sx->lock_object.lo_name, slocked ? "share " : "", 1093 file, line); 1094 1095 if (!(sx->sx_lock & SX_LOCK_SHARED)) { 1096 if (sx_recursed(sx)) { 1097 if (what & SA_NOTRECURSED) 1098 panic("Lock %s recursed @ %s:%d\n", 1099 sx->lock_object.lo_name, file, 1100 line); 1101 } else if (what & SA_RECURSED) 1102 panic("Lock %s not recursed @ %s:%d\n", 1103 sx->lock_object.lo_name, file, line); 1104 } 1105 #endif 1106 break; 1107 case SA_XLOCKED: 1108 case SA_XLOCKED | SA_NOTRECURSED: 1109 case SA_XLOCKED | SA_RECURSED: 1110 if (sx_xholder(sx) != curthread) 1111 panic("Lock %s not exclusively locked @ %s:%d\n", 1112 sx->lock_object.lo_name, file, line); 1113 if (sx_recursed(sx)) { 1114 if (what & SA_NOTRECURSED) 1115 panic("Lock %s recursed @ %s:%d\n", 1116 sx->lock_object.lo_name, file, line); 1117 } else if (what & SA_RECURSED) 1118 panic("Lock %s not recursed @ %s:%d\n", 1119 sx->lock_object.lo_name, file, line); 1120 break; 1121 case SA_UNLOCKED: 1122 #ifdef WITNESS 1123 witness_assert(&sx->lock_object, what, file, line); 1124 #else 1125 /* 1126 * If we hold an exclusve lock fail. We can't 1127 * reliably check to see if we hold a shared lock or 1128 * not. 1129 */ 1130 if (sx_xholder(sx) == curthread) 1131 panic("Lock %s exclusively locked @ %s:%d\n", 1132 sx->lock_object.lo_name, file, line); 1133 #endif 1134 break; 1135 default: 1136 panic("Unknown sx lock assertion: %d @ %s:%d", what, file, 1137 line); 1138 } 1139 } 1140 #endif /* INVARIANT_SUPPORT */ 1141 1142 #ifdef DDB 1143 static void 1144 db_show_sx(const struct lock_object *lock) 1145 { 1146 struct thread *td; 1147 const struct sx *sx; 1148 1149 sx = (const struct sx *)lock; 1150 1151 db_printf(" state: "); 1152 if (sx->sx_lock == SX_LOCK_UNLOCKED) 1153 db_printf("UNLOCKED\n"); 1154 else if (sx->sx_lock == SX_LOCK_DESTROYED) { 1155 db_printf("DESTROYED\n"); 1156 return; 1157 } else if (sx->sx_lock & SX_LOCK_SHARED) 1158 db_printf("SLOCK: %ju\n", (uintmax_t)SX_SHARERS(sx->sx_lock)); 1159 else { 1160 td = sx_xholder(sx); 1161 db_printf("XLOCK: %p (tid %d, pid %d, \"%s\")\n", td, 1162 td->td_tid, td->td_proc->p_pid, td->td_name); 1163 if (sx_recursed(sx)) 1164 db_printf(" recursed: %d\n", sx->sx_recurse); 1165 } 1166 1167 db_printf(" waiters: "); 1168 switch(sx->sx_lock & 1169 (SX_LOCK_SHARED_WAITERS | SX_LOCK_EXCLUSIVE_WAITERS)) { 1170 case SX_LOCK_SHARED_WAITERS: 1171 db_printf("shared\n"); 1172 break; 1173 case SX_LOCK_EXCLUSIVE_WAITERS: 1174 db_printf("exclusive\n"); 1175 break; 1176 case SX_LOCK_SHARED_WAITERS | SX_LOCK_EXCLUSIVE_WAITERS: 1177 db_printf("exclusive and shared\n"); 1178 break; 1179 default: 1180 db_printf("none\n"); 1181 } 1182 } 1183 1184 /* 1185 * Check to see if a thread that is blocked on a sleep queue is actually 1186 * blocked on an sx lock. If so, output some details and return true. 1187 * If the lock has an exclusive owner, return that in *ownerp. 1188 */ 1189 int 1190 sx_chain(struct thread *td, struct thread **ownerp) 1191 { 1192 struct sx *sx; 1193 1194 /* 1195 * Check to see if this thread is blocked on an sx lock. 1196 * First, we check the lock class. If that is ok, then we 1197 * compare the lock name against the wait message. 1198 */ 1199 sx = td->td_wchan; 1200 if (LOCK_CLASS(&sx->lock_object) != &lock_class_sx || 1201 sx->lock_object.lo_name != td->td_wmesg) 1202 return (0); 1203 1204 /* We think we have an sx lock, so output some details. */ 1205 db_printf("blocked on sx \"%s\" ", td->td_wmesg); 1206 *ownerp = sx_xholder(sx); 1207 if (sx->sx_lock & SX_LOCK_SHARED) 1208 db_printf("SLOCK (count %ju)\n", 1209 (uintmax_t)SX_SHARERS(sx->sx_lock)); 1210 else 1211 db_printf("XLOCK\n"); 1212 return (1); 1213 } 1214 #endif 1215