1 /* $NetBSD: kern_mutex.c,v 1.23 2007/11/21 10:19:10 yamt Exp $ */ 2 3 /*- 4 * Copyright (c) 2002, 2006, 2007 The NetBSD Foundation, Inc. 5 * All rights reserved. 6 * 7 * This code is derived from software contributed to The NetBSD Foundation 8 * by Jason R. Thorpe and Andrew Doran. 9 * 10 * Redistribution and use in source and binary forms, with or without 11 * modification, are permitted provided that the following conditions 12 * are met: 13 * 1. Redistributions of source code must retain the above copyright 14 * notice, this list of conditions and the following disclaimer. 15 * 2. Redistributions in binary form must reproduce the above copyright 16 * notice, this list of conditions and the following disclaimer in the 17 * documentation and/or other materials provided with the distribution. 18 * 3. All advertising materials mentioning features or use of this software 19 * must display the following acknowledgement: 20 * This product includes software developed by the NetBSD 21 * Foundation, Inc. and its contributors. 22 * 4. Neither the name of The NetBSD Foundation nor the names of its 23 * contributors may be used to endorse or promote products derived 24 * from this software without specific prior written permission. 25 * 26 * THIS SOFTWARE IS PROVIDED BY THE NETBSD FOUNDATION, INC. AND CONTRIBUTORS 27 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED 28 * TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR 29 * PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE FOUNDATION OR CONTRIBUTORS 30 * BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR 31 * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF 32 * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS 33 * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN 34 * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) 35 * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE 36 * POSSIBILITY OF SUCH DAMAGE. 37 */ 38 39 /* 40 * Kernel mutex implementation, modeled after those found in Solaris, 41 * a description of which can be found in: 42 * 43 * Solaris Internals: Core Kernel Architecture, Jim Mauro and 44 * Richard McDougall. 45 */ 46 47 #define __MUTEX_PRIVATE 48 49 #include <sys/cdefs.h> 50 __KERNEL_RCSID(0, "$NetBSD: kern_mutex.c,v 1.23 2007/11/21 10:19:10 yamt Exp $"); 51 52 #include "opt_multiprocessor.h" 53 54 #include <sys/param.h> 55 #include <sys/proc.h> 56 #include <sys/mutex.h> 57 #include <sys/sched.h> 58 #include <sys/sleepq.h> 59 #include <sys/systm.h> 60 #include <sys/lockdebug.h> 61 #include <sys/kernel.h> 62 63 #include <dev/lockstat.h> 64 65 #include <sys/intr.h> 66 67 /* 68 * When not running a debug kernel, spin mutexes are not much 69 * more than an splraiseipl() and splx() pair. 70 */ 71 72 #if defined(DIAGNOSTIC) || defined(MULTIPROCESSOR) || defined(LOCKDEBUG) 73 #define FULL 74 #endif 75 76 /* 77 * Debugging support. 78 */ 79 80 #define MUTEX_WANTLOCK(mtx) \ 81 LOCKDEBUG_WANTLOCK(MUTEX_DEBUG_P(mtx), (mtx), \ 82 (uintptr_t)__builtin_return_address(0), 0) 83 #define MUTEX_LOCKED(mtx) \ 84 LOCKDEBUG_LOCKED(MUTEX_DEBUG_P(mtx), (mtx), \ 85 (uintptr_t)__builtin_return_address(0), 0) 86 #define MUTEX_UNLOCKED(mtx) \ 87 LOCKDEBUG_UNLOCKED(MUTEX_DEBUG_P(mtx), (mtx), \ 88 (uintptr_t)__builtin_return_address(0), 0) 89 #define MUTEX_ABORT(mtx, msg) \ 90 mutex_abort(mtx, __func__, msg) 91 92 #if defined(LOCKDEBUG) 93 94 #define MUTEX_DASSERT(mtx, cond) \ 95 do { \ 96 if (!(cond)) \ 97 MUTEX_ABORT(mtx, "assertion failed: " #cond); \ 98 } while (/* CONSTCOND */ 0); 99 100 #else /* LOCKDEBUG */ 101 102 #define MUTEX_DASSERT(mtx, cond) /* nothing */ 103 104 #endif /* LOCKDEBUG */ 105 106 #if defined(DIAGNOSTIC) 107 108 #define MUTEX_ASSERT(mtx, cond) \ 109 do { \ 110 if (!(cond)) \ 111 MUTEX_ABORT(mtx, "assertion failed: " #cond); \ 112 } while (/* CONSTCOND */ 0) 113 114 #else /* DIAGNOSTIC */ 115 116 #define MUTEX_ASSERT(mtx, cond) /* nothing */ 117 118 #endif /* DIAGNOSTIC */ 119 120 /* 121 * Spin mutex SPL save / restore. 122 */ 123 #ifndef MUTEX_COUNT_BIAS 124 #define MUTEX_COUNT_BIAS 0 125 #endif 126 127 #define MUTEX_SPIN_SPLRAISE(mtx) \ 128 do { \ 129 struct cpu_info *x__ci = curcpu(); \ 130 int x__cnt, s; \ 131 x__cnt = x__ci->ci_mtx_count--; \ 132 s = splraiseipl(mtx->mtx_ipl); \ 133 if (x__cnt == MUTEX_COUNT_BIAS) \ 134 x__ci->ci_mtx_oldspl = (s); \ 135 } while (/* CONSTCOND */ 0) 136 137 #define MUTEX_SPIN_SPLRESTORE(mtx) \ 138 do { \ 139 struct cpu_info *x__ci = curcpu(); \ 140 int s = x__ci->ci_mtx_oldspl; \ 141 __insn_barrier(); \ 142 if (++(x__ci->ci_mtx_count) == MUTEX_COUNT_BIAS) \ 143 splx(s); \ 144 } while (/* CONSTCOND */ 0) 145 146 /* 147 * For architectures that provide 'simple' mutexes: they provide a 148 * CAS function that is either MP-safe, or does not need to be MP 149 * safe. Adaptive mutexes on these architectures do not require an 150 * additional interlock. 151 */ 152 153 #ifdef __HAVE_SIMPLE_MUTEXES 154 155 #define MUTEX_OWNER(owner) \ 156 (owner & MUTEX_THREAD) 157 #define MUTEX_HAS_WAITERS(mtx) \ 158 (((int)(mtx)->mtx_owner & MUTEX_BIT_WAITERS) != 0) 159 160 #define MUTEX_INITIALIZE_ADAPTIVE(mtx, dodebug) \ 161 do { \ 162 if (dodebug) \ 163 (mtx)->mtx_owner |= MUTEX_BIT_DEBUG; \ 164 } while (/* CONSTCOND */ 0); 165 166 #define MUTEX_INITIALIZE_SPIN(mtx, dodebug, ipl) \ 167 do { \ 168 (mtx)->mtx_owner = MUTEX_BIT_SPIN; \ 169 if (dodebug) \ 170 (mtx)->mtx_owner |= MUTEX_BIT_DEBUG; \ 171 (mtx)->mtx_ipl = makeiplcookie((ipl)); \ 172 __cpu_simple_lock_init(&(mtx)->mtx_lock); \ 173 } while (/* CONSTCOND */ 0) 174 175 #define MUTEX_DESTROY(mtx) \ 176 do { \ 177 (mtx)->mtx_owner = MUTEX_THREAD; \ 178 } while (/* CONSTCOND */ 0); 179 180 #define MUTEX_SPIN_P(mtx) \ 181 (((mtx)->mtx_owner & MUTEX_BIT_SPIN) != 0) 182 #define MUTEX_ADAPTIVE_P(mtx) \ 183 (((mtx)->mtx_owner & MUTEX_BIT_SPIN) == 0) 184 185 #define MUTEX_DEBUG_P(mtx) (((mtx)->mtx_owner & MUTEX_BIT_DEBUG) != 0) 186 #if defined(LOCKDEBUG) 187 #define MUTEX_OWNED(owner) (((owner) & ~MUTEX_BIT_DEBUG) != 0) 188 #define MUTEX_INHERITDEBUG(new, old) (new) |= (old) & MUTEX_BIT_DEBUG 189 #else /* defined(LOCKDEBUG) */ 190 #define MUTEX_OWNED(owner) ((owner) != 0) 191 #define MUTEX_INHERITDEBUG(new, old) /* nothing */ 192 #endif /* defined(LOCKDEBUG) */ 193 194 static inline int 195 MUTEX_ACQUIRE(kmutex_t *mtx, uintptr_t curthread) 196 { 197 int rv; 198 uintptr_t old = 0; 199 uintptr_t new = curthread; 200 201 MUTEX_INHERITDEBUG(old, mtx->mtx_owner); 202 MUTEX_INHERITDEBUG(new, old); 203 rv = MUTEX_CAS(&mtx->mtx_owner, old, new); 204 MUTEX_RECEIVE(mtx); 205 return rv; 206 } 207 208 static inline int 209 MUTEX_SET_WAITERS(kmutex_t *mtx, uintptr_t owner) 210 { 211 int rv; 212 rv = MUTEX_CAS(&mtx->mtx_owner, owner, owner | MUTEX_BIT_WAITERS); 213 MUTEX_RECEIVE(mtx); 214 return rv; 215 } 216 217 static inline void 218 MUTEX_RELEASE(kmutex_t *mtx) 219 { 220 uintptr_t new; 221 222 MUTEX_GIVE(mtx); 223 new = 0; 224 MUTEX_INHERITDEBUG(new, mtx->mtx_owner); 225 mtx->mtx_owner = new; 226 } 227 228 static inline void 229 MUTEX_CLEAR_WAITERS(kmutex_t *mtx) 230 { 231 /* nothing */ 232 } 233 #endif /* __HAVE_SIMPLE_MUTEXES */ 234 235 /* 236 * Patch in stubs via strong alias where they are not available. 237 */ 238 239 #if defined(LOCKDEBUG) 240 #undef __HAVE_MUTEX_STUBS 241 #undef __HAVE_SPIN_MUTEX_STUBS 242 #endif 243 244 #ifndef __HAVE_MUTEX_STUBS 245 __strong_alias(mutex_enter,mutex_vector_enter); 246 __strong_alias(mutex_exit,mutex_vector_exit); 247 #endif 248 249 #ifndef __HAVE_SPIN_MUTEX_STUBS 250 __strong_alias(mutex_spin_enter,mutex_vector_enter); 251 __strong_alias(mutex_spin_exit,mutex_vector_exit); 252 #endif 253 254 void mutex_abort(kmutex_t *, const char *, const char *); 255 void mutex_dump(volatile void *); 256 int mutex_onproc(uintptr_t, struct cpu_info **); 257 static struct lwp *mutex_owner(wchan_t); 258 259 lockops_t mutex_spin_lockops = { 260 "Mutex", 261 0, 262 mutex_dump 263 }; 264 265 lockops_t mutex_adaptive_lockops = { 266 "Mutex", 267 1, 268 mutex_dump 269 }; 270 271 syncobj_t mutex_syncobj = { 272 SOBJ_SLEEPQ_SORTED, 273 turnstile_unsleep, 274 turnstile_changepri, 275 sleepq_lendpri, 276 mutex_owner, 277 }; 278 279 /* 280 * mutex_dump: 281 * 282 * Dump the contents of a mutex structure. 283 */ 284 void 285 mutex_dump(volatile void *cookie) 286 { 287 volatile kmutex_t *mtx = cookie; 288 289 printf_nolog("owner field : %#018lx wait/spin: %16d/%d\n", 290 (long)MUTEX_OWNER(mtx->mtx_owner), MUTEX_HAS_WAITERS(mtx), 291 MUTEX_SPIN_P(mtx)); 292 } 293 294 /* 295 * mutex_abort: 296 * 297 * Dump information about an error and panic the system. This 298 * generates a lot of machine code in the DIAGNOSTIC case, so 299 * we ask the compiler to not inline it. 300 */ 301 302 #if __GNUC_PREREQ__(3, 0) 303 __attribute ((noinline)) __attribute ((noreturn)) 304 #endif 305 void 306 mutex_abort(kmutex_t *mtx, const char *func, const char *msg) 307 { 308 309 LOCKDEBUG_ABORT(mtx, (MUTEX_SPIN_P(mtx) ? 310 &mutex_spin_lockops : &mutex_adaptive_lockops), func, msg); 311 /* NOTREACHED */ 312 } 313 314 /* 315 * mutex_init: 316 * 317 * Initialize a mutex for use. Note that adaptive mutexes are in 318 * essence spin mutexes that can sleep to avoid deadlock and wasting 319 * CPU time. We can't easily provide a type of mutex that always 320 * sleeps - see comments in mutex_vector_enter() about releasing 321 * mutexes unlocked. 322 */ 323 void 324 mutex_init(kmutex_t *mtx, kmutex_type_t type, int ipl) 325 { 326 bool dodebug; 327 328 memset(mtx, 0, sizeof(*mtx)); 329 330 switch (type) { 331 case MUTEX_ADAPTIVE: 332 KASSERT(ipl == IPL_NONE); 333 break; 334 case MUTEX_DEFAULT: 335 case MUTEX_DRIVER: 336 switch (ipl) { 337 case IPL_NONE: 338 type = MUTEX_ADAPTIVE; 339 break; 340 default: 341 type = MUTEX_SPIN; 342 break; 343 } 344 break; 345 default: 346 break; 347 } 348 349 switch (type) { 350 case MUTEX_NODEBUG: 351 dodebug = LOCKDEBUG_ALLOC(mtx, NULL, 352 (uintptr_t)__builtin_return_address(0)); 353 MUTEX_INITIALIZE_SPIN(mtx, dodebug, ipl); 354 break; 355 case MUTEX_ADAPTIVE: 356 dodebug = LOCKDEBUG_ALLOC(mtx, &mutex_adaptive_lockops, 357 (uintptr_t)__builtin_return_address(0)); 358 MUTEX_INITIALIZE_ADAPTIVE(mtx, dodebug); 359 break; 360 case MUTEX_SPIN: 361 dodebug = LOCKDEBUG_ALLOC(mtx, &mutex_spin_lockops, 362 (uintptr_t)__builtin_return_address(0)); 363 MUTEX_INITIALIZE_SPIN(mtx, dodebug, ipl); 364 break; 365 default: 366 panic("mutex_init: impossible type"); 367 break; 368 } 369 } 370 371 /* 372 * mutex_destroy: 373 * 374 * Tear down a mutex. 375 */ 376 void 377 mutex_destroy(kmutex_t *mtx) 378 { 379 380 if (MUTEX_ADAPTIVE_P(mtx)) { 381 MUTEX_ASSERT(mtx, !MUTEX_OWNED(mtx->mtx_owner) && 382 !MUTEX_HAS_WAITERS(mtx)); 383 } else { 384 MUTEX_ASSERT(mtx, !__SIMPLELOCK_LOCKED_P(&mtx->mtx_lock)); 385 } 386 387 LOCKDEBUG_FREE(MUTEX_DEBUG_P(mtx), mtx); 388 MUTEX_DESTROY(mtx); 389 } 390 391 /* 392 * mutex_onproc: 393 * 394 * Return true if an adaptive mutex owner is running on a CPU in the 395 * system. If the target is waiting on the kernel big lock, then we 396 * must release it. This is necessary to avoid deadlock. 397 * 398 * Note that we can't use the mutex owner field as an LWP pointer. We 399 * don't have full control over the timing of our execution, and so the 400 * pointer could be completely invalid by the time we dereference it. 401 */ 402 #ifdef MULTIPROCESSOR 403 int 404 mutex_onproc(uintptr_t owner, struct cpu_info **cip) 405 { 406 CPU_INFO_ITERATOR cii; 407 struct cpu_info *ci; 408 struct lwp *l; 409 410 if (!MUTEX_OWNED(owner)) 411 return 0; 412 l = (struct lwp *)MUTEX_OWNER(owner); 413 414 /* See if the target is running on a CPU somewhere. */ 415 if ((ci = *cip) != NULL && ci->ci_curlwp == l) 416 goto run; 417 for (CPU_INFO_FOREACH(cii, ci)) 418 if (ci->ci_curlwp == l) 419 goto run; 420 421 /* No: it may be safe to block now. */ 422 *cip = NULL; 423 return 0; 424 425 run: 426 /* Target is running; do we need to block? */ 427 *cip = ci; 428 return ci->ci_biglock_wanted != l; 429 } 430 #endif /* MULTIPROCESSOR */ 431 432 /* 433 * mutex_vector_enter: 434 * 435 * Support routine for mutex_enter() that must handles all cases. In 436 * the LOCKDEBUG case, mutex_enter() is always aliased here, even if 437 * fast-path stubs are available. If an mutex_spin_enter() stub is 438 * not available, then it is also aliased directly here. 439 */ 440 void 441 mutex_vector_enter(kmutex_t *mtx) 442 { 443 uintptr_t owner, curthread; 444 turnstile_t *ts; 445 #ifdef MULTIPROCESSOR 446 struct cpu_info *ci = NULL; 447 u_int count; 448 #endif 449 LOCKSTAT_COUNTER(spincnt); 450 LOCKSTAT_COUNTER(slpcnt); 451 LOCKSTAT_TIMER(spintime); 452 LOCKSTAT_TIMER(slptime); 453 LOCKSTAT_FLAG(lsflag); 454 455 /* 456 * Handle spin mutexes. 457 */ 458 if (MUTEX_SPIN_P(mtx)) { 459 #if defined(LOCKDEBUG) && defined(MULTIPROCESSOR) 460 u_int spins = 0; 461 #endif 462 MUTEX_SPIN_SPLRAISE(mtx); 463 MUTEX_WANTLOCK(mtx); 464 #ifdef FULL 465 if (__cpu_simple_lock_try(&mtx->mtx_lock)) { 466 MUTEX_LOCKED(mtx); 467 return; 468 } 469 #if !defined(MULTIPROCESSOR) 470 MUTEX_ABORT(mtx, "locking against myself"); 471 #else /* !MULTIPROCESSOR */ 472 473 LOCKSTAT_ENTER(lsflag); 474 LOCKSTAT_START_TIMER(lsflag, spintime); 475 count = SPINLOCK_BACKOFF_MIN; 476 477 /* 478 * Spin testing the lock word and do exponential backoff 479 * to reduce cache line ping-ponging between CPUs. 480 */ 481 do { 482 if (panicstr != NULL) 483 break; 484 while (__SIMPLELOCK_LOCKED_P(&mtx->mtx_lock)) { 485 SPINLOCK_BACKOFF(count); 486 #ifdef LOCKDEBUG 487 if (SPINLOCK_SPINOUT(spins)) 488 MUTEX_ABORT(mtx, "spinout"); 489 #endif /* LOCKDEBUG */ 490 } 491 } while (!__cpu_simple_lock_try(&mtx->mtx_lock)); 492 493 if (count != SPINLOCK_BACKOFF_MIN) { 494 LOCKSTAT_STOP_TIMER(lsflag, spintime); 495 LOCKSTAT_EVENT(lsflag, mtx, 496 LB_SPIN_MUTEX | LB_SPIN, 1, spintime); 497 } 498 LOCKSTAT_EXIT(lsflag); 499 #endif /* !MULTIPROCESSOR */ 500 #endif /* FULL */ 501 MUTEX_LOCKED(mtx); 502 return; 503 } 504 505 curthread = (uintptr_t)curlwp; 506 507 MUTEX_DASSERT(mtx, MUTEX_ADAPTIVE_P(mtx)); 508 MUTEX_ASSERT(mtx, curthread != 0); 509 MUTEX_WANTLOCK(mtx); 510 511 #ifdef LOCKDEBUG 512 if (panicstr == NULL) { 513 simple_lock_only_held(NULL, "mutex_enter"); 514 #ifdef MULTIPROCESSOR 515 LOCKDEBUG_BARRIER(&kernel_lock, 1); 516 #else 517 LOCKDEBUG_BARRIER(NULL, 1); 518 #endif 519 } 520 #endif 521 522 LOCKSTAT_ENTER(lsflag); 523 524 /* 525 * Adaptive mutex; spin trying to acquire the mutex. If we 526 * determine that the owner is not running on a processor, 527 * then we stop spinning, and sleep instead. 528 */ 529 for (;;) { 530 owner = mtx->mtx_owner; 531 if (!MUTEX_OWNED(owner)) { 532 /* 533 * Mutex owner clear could mean two things: 534 * 535 * * The mutex has been released. 536 * * The owner field hasn't been set yet. 537 * 538 * Try to acquire it again. If that fails, 539 * we'll just loop again. 540 */ 541 if (MUTEX_ACQUIRE(mtx, curthread)) 542 break; 543 continue; 544 } 545 546 if (panicstr != NULL) 547 return; 548 if (MUTEX_OWNER(owner) == curthread) 549 MUTEX_ABORT(mtx, "locking against myself"); 550 551 #ifdef MULTIPROCESSOR 552 /* 553 * Check to see if the owner is running on a processor. 554 * If so, then we should just spin, as the owner will 555 * likely release the lock very soon. 556 */ 557 if (mutex_onproc(owner, &ci)) { 558 LOCKSTAT_START_TIMER(lsflag, spintime); 559 count = SPINLOCK_BACKOFF_MIN; 560 for (;;) { 561 owner = mtx->mtx_owner; 562 if (!mutex_onproc(owner, &ci)) 563 break; 564 SPINLOCK_BACKOFF(count); 565 } 566 LOCKSTAT_STOP_TIMER(lsflag, spintime); 567 LOCKSTAT_COUNT(spincnt, 1); 568 if (!MUTEX_OWNED(owner)) 569 continue; 570 } 571 #endif 572 573 ts = turnstile_lookup(mtx); 574 575 /* 576 * Once we have the turnstile chain interlock, mark the 577 * mutex has having waiters. If that fails, spin again: 578 * chances are that the mutex has been released. 579 */ 580 if (!MUTEX_SET_WAITERS(mtx, owner)) { 581 turnstile_exit(mtx); 582 continue; 583 } 584 585 #ifdef MULTIPROCESSOR 586 /* 587 * mutex_exit() is permitted to release the mutex without 588 * any interlocking instructions, and the following can 589 * occur as a result: 590 * 591 * CPU 1: MUTEX_SET_WAITERS() CPU2: mutex_exit() 592 * ---------------------------- ---------------------------- 593 * .. acquire cache line 594 * .. test for waiters 595 * acquire cache line <- lose cache line 596 * lock cache line .. 597 * verify mutex is held .. 598 * set waiters .. 599 * unlock cache line .. 600 * lose cache line -> acquire cache line 601 * .. clear lock word, waiters 602 * return success 603 * 604 * There is a another race that can occur: a third CPU could 605 * acquire the mutex as soon as it is released. Since 606 * adaptive mutexes are primarily spin mutexes, this is not 607 * something that we need to worry about too much. What we 608 * do need to ensure is that the waiters bit gets set. 609 * 610 * To allow the unlocked release, we need to make some 611 * assumptions here: 612 * 613 * o Release is the only non-atomic/unlocked operation 614 * that can be performed on the mutex. (It must still 615 * be atomic on the local CPU, e.g. in case interrupted 616 * or preempted). 617 * 618 * o At any given time, MUTEX_SET_WAITERS() can only ever 619 * be in progress on one CPU in the system - guaranteed 620 * by the turnstile chain lock. 621 * 622 * o No other operations other than MUTEX_SET_WAITERS() 623 * and release can modify a mutex with a non-zero 624 * owner field. 625 * 626 * o The result of a successful MUTEX_SET_WAITERS() call 627 * is an unbuffered write that is immediately visible 628 * to all other processors in the system. 629 * 630 * o If the holding LWP switches away, it posts a store 631 * fence before changing curlwp, ensuring that any 632 * overwrite of the mutex waiters flag by mutex_exit() 633 * completes before the modification of curlwp becomes 634 * visible to this CPU. 635 * 636 * o mi_switch() posts a store fence before setting curlwp 637 * and before resuming execution of an LWP. 638 * 639 * o _kernel_lock() posts a store fence before setting 640 * curcpu()->ci_biglock_wanted, and after clearing it. 641 * This ensures that any overwrite of the mutex waiters 642 * flag by mutex_exit() completes before the modification 643 * of ci_biglock_wanted becomes visible. 644 * 645 * We now post a read memory barrier (after setting the 646 * waiters field) and check the lock holder's status again. 647 * Some of the possible outcomes (not an exhaustive list): 648 * 649 * 1. The onproc check returns true: the holding LWP is 650 * running again. The lock may be released soon and 651 * we should spin. Importantly, we can't trust the 652 * value of the waiters flag. 653 * 654 * 2. The onproc check returns false: the holding LWP is 655 * not running. We now have the oppertunity to check 656 * if mutex_exit() has blatted the modifications made 657 * by MUTEX_SET_WAITERS(). 658 * 659 * 3. The onproc check returns false: the holding LWP may 660 * or may not be running. It has context switched at 661 * some point during our check. Again, we have the 662 * chance to see if the waiters bit is still set or 663 * has been overwritten. 664 * 665 * 4. The onproc check returns false: the holding LWP is 666 * running on a CPU, but wants the big lock. It's OK 667 * to check the waiters field in this case. 668 * 669 * 5. The has-waiters check fails: the mutex has been 670 * released, the waiters flag cleared and another LWP 671 * now owns the mutex. 672 * 673 * 6. The has-waiters check fails: the mutex has been 674 * released. 675 * 676 * If the waiters bit is not set it's unsafe to go asleep, 677 * as we might never be awoken. 678 */ 679 if ((mb_read(), mutex_onproc(owner, &ci)) || 680 (mb_read(), !MUTEX_HAS_WAITERS(mtx))) { 681 turnstile_exit(mtx); 682 continue; 683 } 684 #endif /* MULTIPROCESSOR */ 685 686 LOCKSTAT_START_TIMER(lsflag, slptime); 687 688 turnstile_block(ts, TS_WRITER_Q, mtx, &mutex_syncobj); 689 690 LOCKSTAT_STOP_TIMER(lsflag, slptime); 691 LOCKSTAT_COUNT(slpcnt, 1); 692 } 693 694 LOCKSTAT_EVENT(lsflag, mtx, LB_ADAPTIVE_MUTEX | LB_SLEEP1, 695 slpcnt, slptime); 696 LOCKSTAT_EVENT(lsflag, mtx, LB_ADAPTIVE_MUTEX | LB_SPIN, 697 spincnt, spintime); 698 LOCKSTAT_EXIT(lsflag); 699 700 MUTEX_DASSERT(mtx, MUTEX_OWNER(mtx->mtx_owner) == curthread); 701 MUTEX_LOCKED(mtx); 702 } 703 704 /* 705 * mutex_vector_exit: 706 * 707 * Support routine for mutex_exit() that handles all cases. 708 */ 709 void 710 mutex_vector_exit(kmutex_t *mtx) 711 { 712 turnstile_t *ts; 713 uintptr_t curthread; 714 715 if (MUTEX_SPIN_P(mtx)) { 716 #ifdef FULL 717 if (!__SIMPLELOCK_LOCKED_P(&mtx->mtx_lock)) 718 MUTEX_ABORT(mtx, "exiting unheld spin mutex"); 719 MUTEX_UNLOCKED(mtx); 720 __cpu_simple_unlock(&mtx->mtx_lock); 721 #endif 722 MUTEX_SPIN_SPLRESTORE(mtx); 723 return; 724 } 725 726 if (__predict_false((uintptr_t)panicstr | cold)) { 727 MUTEX_UNLOCKED(mtx); 728 MUTEX_RELEASE(mtx); 729 return; 730 } 731 732 curthread = (uintptr_t)curlwp; 733 MUTEX_DASSERT(mtx, curthread != 0); 734 MUTEX_ASSERT(mtx, MUTEX_OWNER(mtx->mtx_owner) == curthread); 735 MUTEX_UNLOCKED(mtx); 736 737 #ifdef LOCKDEBUG 738 /* 739 * Avoid having to take the turnstile chain lock every time 740 * around. Raise the priority level to splhigh() in order 741 * to disable preemption and so make the following atomic. 742 */ 743 { 744 int s = splhigh(); 745 if (!MUTEX_HAS_WAITERS(mtx)) { 746 MUTEX_RELEASE(mtx); 747 splx(s); 748 return; 749 } 750 splx(s); 751 } 752 #endif 753 754 /* 755 * Get this lock's turnstile. This gets the interlock on 756 * the sleep queue. Once we have that, we can clear the 757 * lock. If there was no turnstile for the lock, there 758 * were no waiters remaining. 759 */ 760 ts = turnstile_lookup(mtx); 761 762 if (ts == NULL) { 763 MUTEX_RELEASE(mtx); 764 turnstile_exit(mtx); 765 } else { 766 MUTEX_RELEASE(mtx); 767 turnstile_wakeup(ts, TS_WRITER_Q, 768 TS_WAITERS(ts, TS_WRITER_Q), NULL); 769 } 770 } 771 772 #ifndef __HAVE_SIMPLE_MUTEXES 773 /* 774 * mutex_wakeup: 775 * 776 * Support routine for mutex_exit() that wakes up all waiters. 777 * We assume that the mutex has been released, but it need not 778 * be. 779 */ 780 void 781 mutex_wakeup(kmutex_t *mtx) 782 { 783 turnstile_t *ts; 784 785 ts = turnstile_lookup(mtx); 786 if (ts == NULL) { 787 turnstile_exit(mtx); 788 return; 789 } 790 MUTEX_CLEAR_WAITERS(mtx); 791 turnstile_wakeup(ts, TS_WRITER_Q, TS_WAITERS(ts, TS_WRITER_Q), NULL); 792 } 793 #endif /* !__HAVE_SIMPLE_MUTEXES */ 794 795 /* 796 * mutex_owned: 797 * 798 * Return true if the current LWP (adaptive) or CPU (spin) 799 * holds the mutex. 800 */ 801 int 802 mutex_owned(kmutex_t *mtx) 803 { 804 805 if (MUTEX_ADAPTIVE_P(mtx)) 806 return MUTEX_OWNER(mtx->mtx_owner) == (uintptr_t)curlwp; 807 #ifdef FULL 808 return __SIMPLELOCK_LOCKED_P(&mtx->mtx_lock); 809 #else 810 return 1; 811 #endif 812 } 813 814 /* 815 * mutex_owner: 816 * 817 * Return the current owner of an adaptive mutex. Used for 818 * priority inheritance. 819 */ 820 static struct lwp * 821 mutex_owner(wchan_t obj) 822 { 823 kmutex_t *mtx = (void *)(uintptr_t)obj; /* discard qualifiers */ 824 825 MUTEX_ASSERT(mtx, MUTEX_ADAPTIVE_P(mtx)); 826 return (struct lwp *)MUTEX_OWNER(mtx->mtx_owner); 827 } 828 829 /* 830 * mutex_tryenter: 831 * 832 * Try to acquire the mutex; return non-zero if we did. 833 */ 834 int 835 mutex_tryenter(kmutex_t *mtx) 836 { 837 uintptr_t curthread; 838 839 /* 840 * Handle spin mutexes. 841 */ 842 if (MUTEX_SPIN_P(mtx)) { 843 MUTEX_SPIN_SPLRAISE(mtx); 844 #ifdef FULL 845 if (__cpu_simple_lock_try(&mtx->mtx_lock)) { 846 MUTEX_WANTLOCK(mtx); 847 MUTEX_LOCKED(mtx); 848 return 1; 849 } 850 MUTEX_SPIN_SPLRESTORE(mtx); 851 #else 852 MUTEX_WANTLOCK(mtx); 853 MUTEX_LOCKED(mtx); 854 return 1; 855 #endif 856 } else { 857 curthread = (uintptr_t)curlwp; 858 MUTEX_ASSERT(mtx, curthread != 0); 859 if (MUTEX_ACQUIRE(mtx, curthread)) { 860 MUTEX_WANTLOCK(mtx); 861 MUTEX_LOCKED(mtx); 862 MUTEX_DASSERT(mtx, 863 MUTEX_OWNER(mtx->mtx_owner) == curthread); 864 return 1; 865 } 866 } 867 868 return 0; 869 } 870 871 #if defined(__HAVE_SPIN_MUTEX_STUBS) || defined(FULL) 872 /* 873 * mutex_spin_retry: 874 * 875 * Support routine for mutex_spin_enter(). Assumes that the caller 876 * has already raised the SPL, and adjusted counters. 877 */ 878 void 879 mutex_spin_retry(kmutex_t *mtx) 880 { 881 #ifdef MULTIPROCESSOR 882 u_int count; 883 LOCKSTAT_TIMER(spintime); 884 LOCKSTAT_FLAG(lsflag); 885 #ifdef LOCKDEBUG 886 u_int spins = 0; 887 #endif /* LOCKDEBUG */ 888 889 MUTEX_WANTLOCK(mtx); 890 891 LOCKSTAT_ENTER(lsflag); 892 LOCKSTAT_START_TIMER(lsflag, spintime); 893 count = SPINLOCK_BACKOFF_MIN; 894 895 /* 896 * Spin testing the lock word and do exponential backoff 897 * to reduce cache line ping-ponging between CPUs. 898 */ 899 do { 900 if (panicstr != NULL) 901 break; 902 while (__SIMPLELOCK_LOCKED_P(&mtx->mtx_lock)) { 903 SPINLOCK_BACKOFF(count); 904 #ifdef LOCKDEBUG 905 if (SPINLOCK_SPINOUT(spins)) 906 MUTEX_ABORT(mtx, "spinout"); 907 #endif /* LOCKDEBUG */ 908 } 909 } while (!__cpu_simple_lock_try(&mtx->mtx_lock)); 910 911 LOCKSTAT_STOP_TIMER(lsflag, spintime); 912 LOCKSTAT_EVENT(lsflag, mtx, LB_SPIN_MUTEX | LB_SPIN, 1, spintime); 913 LOCKSTAT_EXIT(lsflag); 914 915 MUTEX_LOCKED(mtx); 916 #else /* MULTIPROCESSOR */ 917 MUTEX_ABORT(mtx, "locking against myself"); 918 #endif /* MULTIPROCESSOR */ 919 } 920 #endif /* defined(__HAVE_SPIN_MUTEX_STUBS) || defined(FULL) */ 921