1 /* $NetBSD: kern_mutex.c,v 1.41 2008/05/19 17:06:02 ad Exp $ */ 2 3 /*- 4 * Copyright (c) 2002, 2006, 2007, 2008 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 * 19 * THIS SOFTWARE IS PROVIDED BY THE NETBSD FOUNDATION, INC. AND CONTRIBUTORS 20 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED 21 * TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR 22 * PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE FOUNDATION OR CONTRIBUTORS 23 * BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR 24 * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF 25 * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS 26 * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN 27 * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) 28 * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE 29 * POSSIBILITY OF SUCH DAMAGE. 30 */ 31 32 /* 33 * Kernel mutex implementation, modeled after those found in Solaris, 34 * a description of which can be found in: 35 * 36 * Solaris Internals: Core Kernel Architecture, Jim Mauro and 37 * Richard McDougall. 38 */ 39 40 #define __MUTEX_PRIVATE 41 42 #include <sys/cdefs.h> 43 __KERNEL_RCSID(0, "$NetBSD: kern_mutex.c,v 1.41 2008/05/19 17:06:02 ad Exp $"); 44 45 #include <sys/param.h> 46 #include <sys/proc.h> 47 #include <sys/mutex.h> 48 #include <sys/sched.h> 49 #include <sys/sleepq.h> 50 #include <sys/systm.h> 51 #include <sys/lockdebug.h> 52 #include <sys/kernel.h> 53 #include <sys/atomic.h> 54 #include <sys/intr.h> 55 #include <sys/lock.h> 56 #include <sys/pool.h> 57 58 #include <dev/lockstat.h> 59 60 #include <machine/lock.h> 61 62 /* 63 * When not running a debug kernel, spin mutexes are not much 64 * more than an splraiseipl() and splx() pair. 65 */ 66 67 #if defined(DIAGNOSTIC) || defined(MULTIPROCESSOR) || defined(LOCKDEBUG) 68 #define FULL 69 #endif 70 71 /* 72 * Debugging support. 73 */ 74 75 #define MUTEX_WANTLOCK(mtx) \ 76 LOCKDEBUG_WANTLOCK(MUTEX_DEBUG_P(mtx), (mtx), \ 77 (uintptr_t)__builtin_return_address(0), false, false) 78 #define MUTEX_LOCKED(mtx) \ 79 LOCKDEBUG_LOCKED(MUTEX_DEBUG_P(mtx), (mtx), \ 80 (uintptr_t)__builtin_return_address(0), 0) 81 #define MUTEX_UNLOCKED(mtx) \ 82 LOCKDEBUG_UNLOCKED(MUTEX_DEBUG_P(mtx), (mtx), \ 83 (uintptr_t)__builtin_return_address(0), 0) 84 #define MUTEX_ABORT(mtx, msg) \ 85 mutex_abort(mtx, __func__, msg) 86 87 #if defined(LOCKDEBUG) 88 89 #define MUTEX_DASSERT(mtx, cond) \ 90 do { \ 91 if (!(cond)) \ 92 MUTEX_ABORT(mtx, "assertion failed: " #cond); \ 93 } while (/* CONSTCOND */ 0); 94 95 #else /* LOCKDEBUG */ 96 97 #define MUTEX_DASSERT(mtx, cond) /* nothing */ 98 99 #endif /* LOCKDEBUG */ 100 101 #if defined(DIAGNOSTIC) 102 103 #define MUTEX_ASSERT(mtx, cond) \ 104 do { \ 105 if (!(cond)) \ 106 MUTEX_ABORT(mtx, "assertion failed: " #cond); \ 107 } while (/* CONSTCOND */ 0) 108 109 #else /* DIAGNOSTIC */ 110 111 #define MUTEX_ASSERT(mtx, cond) /* nothing */ 112 113 #endif /* DIAGNOSTIC */ 114 115 /* 116 * Spin mutex SPL save / restore. 117 */ 118 #ifndef MUTEX_COUNT_BIAS 119 #define MUTEX_COUNT_BIAS 0 120 #endif 121 122 #define MUTEX_SPIN_SPLRAISE(mtx) \ 123 do { \ 124 struct cpu_info *x__ci; \ 125 int x__cnt, s; \ 126 s = splraiseipl(mtx->mtx_ipl); \ 127 x__ci = curcpu(); \ 128 x__cnt = x__ci->ci_mtx_count--; \ 129 __insn_barrier(); \ 130 if (x__cnt == MUTEX_COUNT_BIAS) \ 131 x__ci->ci_mtx_oldspl = (s); \ 132 } while (/* CONSTCOND */ 0) 133 134 #define MUTEX_SPIN_SPLRESTORE(mtx) \ 135 do { \ 136 struct cpu_info *x__ci = curcpu(); \ 137 int s = x__ci->ci_mtx_oldspl; \ 138 __insn_barrier(); \ 139 if (++(x__ci->ci_mtx_count) == MUTEX_COUNT_BIAS) \ 140 splx(s); \ 141 } while (/* CONSTCOND */ 0) 142 143 /* 144 * For architectures that provide 'simple' mutexes: they provide a 145 * CAS function that is either MP-safe, or does not need to be MP 146 * safe. Adaptive mutexes on these architectures do not require an 147 * additional interlock. 148 */ 149 150 #ifdef __HAVE_SIMPLE_MUTEXES 151 152 #define MUTEX_OWNER(owner) \ 153 (owner & MUTEX_THREAD) 154 #define MUTEX_HAS_WAITERS(mtx) \ 155 (((int)(mtx)->mtx_owner & MUTEX_BIT_WAITERS) != 0) 156 157 #define MUTEX_INITIALIZE_ADAPTIVE(mtx, dodebug) \ 158 do { \ 159 if (dodebug) \ 160 (mtx)->mtx_owner |= MUTEX_BIT_DEBUG; \ 161 } while (/* CONSTCOND */ 0); 162 163 #define MUTEX_INITIALIZE_SPIN(mtx, dodebug, ipl) \ 164 do { \ 165 (mtx)->mtx_owner = MUTEX_BIT_SPIN; \ 166 if (dodebug) \ 167 (mtx)->mtx_owner |= MUTEX_BIT_DEBUG; \ 168 (mtx)->mtx_ipl = makeiplcookie((ipl)); \ 169 __cpu_simple_lock_init(&(mtx)->mtx_lock); \ 170 } while (/* CONSTCOND */ 0) 171 172 #define MUTEX_DESTROY(mtx) \ 173 do { \ 174 (mtx)->mtx_owner = MUTEX_THREAD; \ 175 } while (/* CONSTCOND */ 0); 176 177 #define MUTEX_SPIN_P(mtx) \ 178 (((mtx)->mtx_owner & MUTEX_BIT_SPIN) != 0) 179 #define MUTEX_ADAPTIVE_P(mtx) \ 180 (((mtx)->mtx_owner & MUTEX_BIT_SPIN) == 0) 181 182 #define MUTEX_DEBUG_P(mtx) (((mtx)->mtx_owner & MUTEX_BIT_DEBUG) != 0) 183 #if defined(LOCKDEBUG) 184 #define MUTEX_OWNED(owner) (((owner) & ~MUTEX_BIT_DEBUG) != 0) 185 #define MUTEX_INHERITDEBUG(new, old) (new) |= (old) & MUTEX_BIT_DEBUG 186 #else /* defined(LOCKDEBUG) */ 187 #define MUTEX_OWNED(owner) ((owner) != 0) 188 #define MUTEX_INHERITDEBUG(new, old) /* nothing */ 189 #endif /* defined(LOCKDEBUG) */ 190 191 static inline int 192 MUTEX_ACQUIRE(kmutex_t *mtx, uintptr_t curthread) 193 { 194 int rv; 195 uintptr_t old = 0; 196 uintptr_t new = curthread; 197 198 MUTEX_INHERITDEBUG(old, mtx->mtx_owner); 199 MUTEX_INHERITDEBUG(new, old); 200 rv = MUTEX_CAS(&mtx->mtx_owner, old, new); 201 MUTEX_RECEIVE(mtx); 202 return rv; 203 } 204 205 static inline int 206 MUTEX_SET_WAITERS(kmutex_t *mtx, uintptr_t owner) 207 { 208 int rv; 209 rv = MUTEX_CAS(&mtx->mtx_owner, owner, owner | MUTEX_BIT_WAITERS); 210 MUTEX_RECEIVE(mtx); 211 return rv; 212 } 213 214 static inline void 215 MUTEX_RELEASE(kmutex_t *mtx) 216 { 217 uintptr_t new; 218 219 MUTEX_GIVE(mtx); 220 new = 0; 221 MUTEX_INHERITDEBUG(new, mtx->mtx_owner); 222 mtx->mtx_owner = new; 223 } 224 225 static inline void 226 MUTEX_CLEAR_WAITERS(kmutex_t *mtx) 227 { 228 /* nothing */ 229 } 230 #endif /* __HAVE_SIMPLE_MUTEXES */ 231 232 /* 233 * Patch in stubs via strong alias where they are not available. 234 */ 235 236 #if defined(LOCKDEBUG) 237 #undef __HAVE_MUTEX_STUBS 238 #undef __HAVE_SPIN_MUTEX_STUBS 239 #endif 240 241 #ifndef __HAVE_MUTEX_STUBS 242 __strong_alias(mutex_enter,mutex_vector_enter); 243 __strong_alias(mutex_exit,mutex_vector_exit); 244 #endif 245 246 #ifndef __HAVE_SPIN_MUTEX_STUBS 247 __strong_alias(mutex_spin_enter,mutex_vector_enter); 248 __strong_alias(mutex_spin_exit,mutex_vector_exit); 249 #endif 250 251 void mutex_abort(kmutex_t *, const char *, const char *); 252 void mutex_dump(volatile void *); 253 int mutex_onproc(uintptr_t, struct cpu_info **); 254 255 lockops_t mutex_spin_lockops = { 256 "Mutex", 257 0, 258 mutex_dump 259 }; 260 261 lockops_t mutex_adaptive_lockops = { 262 "Mutex", 263 1, 264 mutex_dump 265 }; 266 267 syncobj_t mutex_syncobj = { 268 SOBJ_SLEEPQ_SORTED, 269 turnstile_unsleep, 270 turnstile_changepri, 271 sleepq_lendpri, 272 (void *)mutex_owner, 273 }; 274 275 /* Mutex cache */ 276 #define MUTEX_OBJ_MAGIC 0x5aa3c85d 277 struct kmutexobj { 278 kmutex_t mo_lock; 279 u_int mo_magic; 280 u_int mo_refcnt; 281 }; 282 283 static int mutex_obj_ctor(void *, void *, int); 284 285 static pool_cache_t mutex_obj_cache; 286 287 /* 288 * mutex_dump: 289 * 290 * Dump the contents of a mutex structure. 291 */ 292 void 293 mutex_dump(volatile void *cookie) 294 { 295 volatile kmutex_t *mtx = cookie; 296 297 printf_nolog("owner field : %#018lx wait/spin: %16d/%d\n", 298 (long)MUTEX_OWNER(mtx->mtx_owner), MUTEX_HAS_WAITERS(mtx), 299 MUTEX_SPIN_P(mtx)); 300 } 301 302 /* 303 * mutex_abort: 304 * 305 * Dump information about an error and panic the system. This 306 * generates a lot of machine code in the DIAGNOSTIC case, so 307 * we ask the compiler to not inline it. 308 */ 309 310 #if __GNUC_PREREQ__(3, 0) 311 __attribute ((noinline)) __attribute ((noreturn)) 312 #endif 313 void 314 mutex_abort(kmutex_t *mtx, const char *func, const char *msg) 315 { 316 317 LOCKDEBUG_ABORT(mtx, (MUTEX_SPIN_P(mtx) ? 318 &mutex_spin_lockops : &mutex_adaptive_lockops), func, msg); 319 /* NOTREACHED */ 320 } 321 322 /* 323 * mutex_init: 324 * 325 * Initialize a mutex for use. Note that adaptive mutexes are in 326 * essence spin mutexes that can sleep to avoid deadlock and wasting 327 * CPU time. We can't easily provide a type of mutex that always 328 * sleeps - see comments in mutex_vector_enter() about releasing 329 * mutexes unlocked. 330 */ 331 void 332 mutex_init(kmutex_t *mtx, kmutex_type_t type, int ipl) 333 { 334 bool dodebug; 335 336 memset(mtx, 0, sizeof(*mtx)); 337 338 switch (type) { 339 case MUTEX_ADAPTIVE: 340 KASSERT(ipl == IPL_NONE); 341 break; 342 case MUTEX_DEFAULT: 343 case MUTEX_DRIVER: 344 if (ipl == IPL_NONE || ipl == IPL_SOFTCLOCK || 345 ipl == IPL_SOFTBIO || ipl == IPL_SOFTNET || 346 ipl == IPL_SOFTSERIAL) { 347 type = MUTEX_ADAPTIVE; 348 } else { 349 type = MUTEX_SPIN; 350 } 351 break; 352 default: 353 break; 354 } 355 356 switch (type) { 357 case MUTEX_NODEBUG: 358 dodebug = LOCKDEBUG_ALLOC(mtx, NULL, 359 (uintptr_t)__builtin_return_address(0)); 360 MUTEX_INITIALIZE_SPIN(mtx, dodebug, ipl); 361 break; 362 case MUTEX_ADAPTIVE: 363 dodebug = LOCKDEBUG_ALLOC(mtx, &mutex_adaptive_lockops, 364 (uintptr_t)__builtin_return_address(0)); 365 MUTEX_INITIALIZE_ADAPTIVE(mtx, dodebug); 366 break; 367 case MUTEX_SPIN: 368 dodebug = LOCKDEBUG_ALLOC(mtx, &mutex_spin_lockops, 369 (uintptr_t)__builtin_return_address(0)); 370 MUTEX_INITIALIZE_SPIN(mtx, dodebug, ipl); 371 break; 372 default: 373 panic("mutex_init: impossible type"); 374 break; 375 } 376 } 377 378 /* 379 * mutex_destroy: 380 * 381 * Tear down a mutex. 382 */ 383 void 384 mutex_destroy(kmutex_t *mtx) 385 { 386 387 if (MUTEX_ADAPTIVE_P(mtx)) { 388 MUTEX_ASSERT(mtx, !MUTEX_OWNED(mtx->mtx_owner) && 389 !MUTEX_HAS_WAITERS(mtx)); 390 } else { 391 MUTEX_ASSERT(mtx, !__SIMPLELOCK_LOCKED_P(&mtx->mtx_lock)); 392 } 393 394 LOCKDEBUG_FREE(MUTEX_DEBUG_P(mtx), mtx); 395 MUTEX_DESTROY(mtx); 396 } 397 398 /* 399 * mutex_onproc: 400 * 401 * Return true if an adaptive mutex owner is running on a CPU in the 402 * system. If the target is waiting on the kernel big lock, then we 403 * must release it. This is necessary to avoid deadlock. 404 * 405 * Note that we can't use the mutex owner field as an LWP pointer. We 406 * don't have full control over the timing of our execution, and so the 407 * pointer could be completely invalid by the time we dereference it. 408 */ 409 #ifdef MULTIPROCESSOR 410 int 411 mutex_onproc(uintptr_t owner, struct cpu_info **cip) 412 { 413 CPU_INFO_ITERATOR cii; 414 struct cpu_info *ci; 415 struct lwp *l; 416 417 if (!MUTEX_OWNED(owner)) 418 return 0; 419 l = (struct lwp *)MUTEX_OWNER(owner); 420 421 /* See if the target is running on a CPU somewhere. */ 422 if ((ci = *cip) != NULL && ci->ci_curlwp == l) 423 goto run; 424 for (CPU_INFO_FOREACH(cii, ci)) 425 if (ci->ci_curlwp == l) 426 goto run; 427 428 /* No: it may be safe to block now. */ 429 *cip = NULL; 430 return 0; 431 432 run: 433 /* Target is running; do we need to block? */ 434 *cip = ci; 435 return ci->ci_biglock_wanted != l; 436 } 437 #endif /* MULTIPROCESSOR */ 438 439 /* 440 * mutex_vector_enter: 441 * 442 * Support routine for mutex_enter() that must handles all cases. In 443 * the LOCKDEBUG case, mutex_enter() is always aliased here, even if 444 * fast-path stubs are available. If an mutex_spin_enter() stub is 445 * not available, then it is also aliased directly here. 446 */ 447 void 448 mutex_vector_enter(kmutex_t *mtx) 449 { 450 uintptr_t owner, curthread; 451 turnstile_t *ts; 452 #ifdef MULTIPROCESSOR 453 struct cpu_info *ci = NULL; 454 u_int count; 455 #endif 456 LOCKSTAT_COUNTER(spincnt); 457 LOCKSTAT_COUNTER(slpcnt); 458 LOCKSTAT_TIMER(spintime); 459 LOCKSTAT_TIMER(slptime); 460 LOCKSTAT_FLAG(lsflag); 461 462 /* 463 * Handle spin mutexes. 464 */ 465 if (MUTEX_SPIN_P(mtx)) { 466 #if defined(LOCKDEBUG) && defined(MULTIPROCESSOR) 467 u_int spins = 0; 468 #endif 469 MUTEX_SPIN_SPLRAISE(mtx); 470 MUTEX_WANTLOCK(mtx); 471 #ifdef FULL 472 if (__cpu_simple_lock_try(&mtx->mtx_lock)) { 473 MUTEX_LOCKED(mtx); 474 return; 475 } 476 #if !defined(MULTIPROCESSOR) 477 MUTEX_ABORT(mtx, "locking against myself"); 478 #else /* !MULTIPROCESSOR */ 479 480 LOCKSTAT_ENTER(lsflag); 481 LOCKSTAT_START_TIMER(lsflag, spintime); 482 count = SPINLOCK_BACKOFF_MIN; 483 484 /* 485 * Spin testing the lock word and do exponential backoff 486 * to reduce cache line ping-ponging between CPUs. 487 */ 488 do { 489 if (panicstr != NULL) 490 break; 491 while (__SIMPLELOCK_LOCKED_P(&mtx->mtx_lock)) { 492 SPINLOCK_BACKOFF(count); 493 #ifdef LOCKDEBUG 494 if (SPINLOCK_SPINOUT(spins)) 495 MUTEX_ABORT(mtx, "spinout"); 496 #endif /* LOCKDEBUG */ 497 } 498 } while (!__cpu_simple_lock_try(&mtx->mtx_lock)); 499 500 if (count != SPINLOCK_BACKOFF_MIN) { 501 LOCKSTAT_STOP_TIMER(lsflag, spintime); 502 LOCKSTAT_EVENT(lsflag, mtx, 503 LB_SPIN_MUTEX | LB_SPIN, 1, spintime); 504 } 505 LOCKSTAT_EXIT(lsflag); 506 #endif /* !MULTIPROCESSOR */ 507 #endif /* FULL */ 508 MUTEX_LOCKED(mtx); 509 return; 510 } 511 512 curthread = (uintptr_t)curlwp; 513 514 MUTEX_DASSERT(mtx, MUTEX_ADAPTIVE_P(mtx)); 515 MUTEX_ASSERT(mtx, curthread != 0); 516 MUTEX_WANTLOCK(mtx); 517 518 if (panicstr == NULL) { 519 LOCKDEBUG_BARRIER(&kernel_lock, 1); 520 } 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 (owner = mtx->mtx_owner;;) { 530 if (!MUTEX_OWNED(owner)) { 531 /* 532 * Mutex owner clear could mean two things: 533 * 534 * * The mutex has been released. 535 * * The owner field hasn't been set yet. 536 * 537 * Try to acquire it again. If that fails, 538 * we'll just loop again. 539 */ 540 if (MUTEX_ACQUIRE(mtx, curthread)) 541 break; 542 owner = mtx->mtx_owner; 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 SPINLOCK_BACKOFF(count); 562 owner = mtx->mtx_owner; 563 if (!mutex_onproc(owner, &ci)) 564 break; 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 owner = mtx->mtx_owner; 583 continue; 584 } 585 586 #ifdef MULTIPROCESSOR 587 /* 588 * mutex_exit() is permitted to release the mutex without 589 * any interlocking instructions, and the following can 590 * occur as a result: 591 * 592 * CPU 1: MUTEX_SET_WAITERS() CPU2: mutex_exit() 593 * ---------------------------- ---------------------------- 594 * .. acquire cache line 595 * .. test for waiters 596 * acquire cache line <- lose cache line 597 * lock cache line .. 598 * verify mutex is held .. 599 * set waiters .. 600 * unlock cache line .. 601 * lose cache line -> acquire cache line 602 * .. clear lock word, waiters 603 * return success 604 * 605 * There is a another race that can occur: a third CPU could 606 * acquire the mutex as soon as it is released. Since 607 * adaptive mutexes are primarily spin mutexes, this is not 608 * something that we need to worry about too much. What we 609 * do need to ensure is that the waiters bit gets set. 610 * 611 * To allow the unlocked release, we need to make some 612 * assumptions here: 613 * 614 * o Release is the only non-atomic/unlocked operation 615 * that can be performed on the mutex. (It must still 616 * be atomic on the local CPU, e.g. in case interrupted 617 * or preempted). 618 * 619 * o At any given time, MUTEX_SET_WAITERS() can only ever 620 * be in progress on one CPU in the system - guaranteed 621 * by the turnstile chain lock. 622 * 623 * o No other operations other than MUTEX_SET_WAITERS() 624 * and release can modify a mutex with a non-zero 625 * owner field. 626 * 627 * o The result of a successful MUTEX_SET_WAITERS() call 628 * is an unbuffered write that is immediately visible 629 * to all other processors in the system. 630 * 631 * o If the holding LWP switches away, it posts a store 632 * fence before changing curlwp, ensuring that any 633 * overwrite of the mutex waiters flag by mutex_exit() 634 * completes before the modification of curlwp becomes 635 * visible to this CPU. 636 * 637 * o mi_switch() posts a store fence before setting curlwp 638 * and before resuming execution of an LWP. 639 * 640 * o _kernel_lock() posts a store fence before setting 641 * curcpu()->ci_biglock_wanted, and after clearing it. 642 * This ensures that any overwrite of the mutex waiters 643 * flag by mutex_exit() completes before the modification 644 * of ci_biglock_wanted becomes visible. 645 * 646 * We now post a read memory barrier (after setting the 647 * waiters field) and check the lock holder's status again. 648 * Some of the possible outcomes (not an exhaustive list): 649 * 650 * 1. The onproc check returns true: the holding LWP is 651 * running again. The lock may be released soon and 652 * we should spin. Importantly, we can't trust the 653 * value of the waiters flag. 654 * 655 * 2. The onproc check returns false: the holding LWP is 656 * not running. We now have the opportunity to check 657 * if mutex_exit() has blatted the modifications made 658 * by MUTEX_SET_WAITERS(). 659 * 660 * 3. The onproc check returns false: the holding LWP may 661 * or may not be running. It has context switched at 662 * some point during our check. Again, we have the 663 * chance to see if the waiters bit is still set or 664 * has been overwritten. 665 * 666 * 4. The onproc check returns false: the holding LWP is 667 * running on a CPU, but wants the big lock. It's OK 668 * to check the waiters field in this case. 669 * 670 * 5. The has-waiters check fails: the mutex has been 671 * released, the waiters flag cleared and another LWP 672 * now owns the mutex. 673 * 674 * 6. The has-waiters check fails: the mutex has been 675 * released. 676 * 677 * If the waiters bit is not set it's unsafe to go asleep, 678 * as we might never be awoken. 679 */ 680 if ((membar_consumer(), mutex_onproc(owner, &ci)) || 681 (membar_consumer(), !MUTEX_HAS_WAITERS(mtx))) { 682 turnstile_exit(mtx); 683 owner = mtx->mtx_owner; 684 continue; 685 } 686 #endif /* MULTIPROCESSOR */ 687 688 LOCKSTAT_START_TIMER(lsflag, slptime); 689 690 turnstile_block(ts, TS_WRITER_Q, mtx, &mutex_syncobj); 691 692 LOCKSTAT_STOP_TIMER(lsflag, slptime); 693 LOCKSTAT_COUNT(slpcnt, 1); 694 695 owner = mtx->mtx_owner; 696 } 697 698 LOCKSTAT_EVENT(lsflag, mtx, LB_ADAPTIVE_MUTEX | LB_SLEEP1, 699 slpcnt, slptime); 700 LOCKSTAT_EVENT(lsflag, mtx, LB_ADAPTIVE_MUTEX | LB_SPIN, 701 spincnt, spintime); 702 LOCKSTAT_EXIT(lsflag); 703 704 MUTEX_DASSERT(mtx, MUTEX_OWNER(mtx->mtx_owner) == curthread); 705 MUTEX_LOCKED(mtx); 706 } 707 708 /* 709 * mutex_vector_exit: 710 * 711 * Support routine for mutex_exit() that handles all cases. 712 */ 713 void 714 mutex_vector_exit(kmutex_t *mtx) 715 { 716 turnstile_t *ts; 717 uintptr_t curthread; 718 719 if (MUTEX_SPIN_P(mtx)) { 720 #ifdef FULL 721 if (__predict_false(!__SIMPLELOCK_LOCKED_P(&mtx->mtx_lock))) { 722 if (panicstr != NULL) 723 return; 724 MUTEX_ABORT(mtx, "exiting unheld spin mutex"); 725 } 726 MUTEX_UNLOCKED(mtx); 727 __cpu_simple_unlock(&mtx->mtx_lock); 728 #endif 729 MUTEX_SPIN_SPLRESTORE(mtx); 730 return; 731 } 732 733 if (__predict_false((uintptr_t)panicstr | cold)) { 734 MUTEX_UNLOCKED(mtx); 735 MUTEX_RELEASE(mtx); 736 return; 737 } 738 739 curthread = (uintptr_t)curlwp; 740 MUTEX_DASSERT(mtx, curthread != 0); 741 MUTEX_ASSERT(mtx, MUTEX_OWNER(mtx->mtx_owner) == curthread); 742 MUTEX_UNLOCKED(mtx); 743 744 #ifdef LOCKDEBUG 745 /* 746 * Avoid having to take the turnstile chain lock every time 747 * around. Raise the priority level to splhigh() in order 748 * to disable preemption and so make the following atomic. 749 */ 750 { 751 int s = splhigh(); 752 if (!MUTEX_HAS_WAITERS(mtx)) { 753 MUTEX_RELEASE(mtx); 754 splx(s); 755 return; 756 } 757 splx(s); 758 } 759 #endif 760 761 /* 762 * Get this lock's turnstile. This gets the interlock on 763 * the sleep queue. Once we have that, we can clear the 764 * lock. If there was no turnstile for the lock, there 765 * were no waiters remaining. 766 */ 767 ts = turnstile_lookup(mtx); 768 769 if (ts == NULL) { 770 MUTEX_RELEASE(mtx); 771 turnstile_exit(mtx); 772 } else { 773 MUTEX_RELEASE(mtx); 774 turnstile_wakeup(ts, TS_WRITER_Q, 775 TS_WAITERS(ts, TS_WRITER_Q), NULL); 776 } 777 } 778 779 #ifndef __HAVE_SIMPLE_MUTEXES 780 /* 781 * mutex_wakeup: 782 * 783 * Support routine for mutex_exit() that wakes up all waiters. 784 * We assume that the mutex has been released, but it need not 785 * be. 786 */ 787 void 788 mutex_wakeup(kmutex_t *mtx) 789 { 790 turnstile_t *ts; 791 792 ts = turnstile_lookup(mtx); 793 if (ts == NULL) { 794 turnstile_exit(mtx); 795 return; 796 } 797 MUTEX_CLEAR_WAITERS(mtx); 798 turnstile_wakeup(ts, TS_WRITER_Q, TS_WAITERS(ts, TS_WRITER_Q), NULL); 799 } 800 #endif /* !__HAVE_SIMPLE_MUTEXES */ 801 802 /* 803 * mutex_owned: 804 * 805 * Return true if the current LWP (adaptive) or CPU (spin) 806 * holds the mutex. 807 */ 808 int 809 mutex_owned(kmutex_t *mtx) 810 { 811 812 if (mtx == NULL) 813 return 0; 814 if (MUTEX_ADAPTIVE_P(mtx)) 815 return MUTEX_OWNER(mtx->mtx_owner) == (uintptr_t)curlwp; 816 #ifdef FULL 817 return __SIMPLELOCK_LOCKED_P(&mtx->mtx_lock); 818 #else 819 return 1; 820 #endif 821 } 822 823 /* 824 * mutex_owner: 825 * 826 * Return the current owner of an adaptive mutex. Used for 827 * priority inheritance. 828 */ 829 lwp_t * 830 mutex_owner(kmutex_t *mtx) 831 { 832 833 MUTEX_ASSERT(mtx, MUTEX_ADAPTIVE_P(mtx)); 834 return (struct lwp *)MUTEX_OWNER(mtx->mtx_owner); 835 } 836 837 /* 838 * mutex_tryenter: 839 * 840 * Try to acquire the mutex; return non-zero if we did. 841 */ 842 int 843 mutex_tryenter(kmutex_t *mtx) 844 { 845 uintptr_t curthread; 846 847 /* 848 * Handle spin mutexes. 849 */ 850 if (MUTEX_SPIN_P(mtx)) { 851 MUTEX_SPIN_SPLRAISE(mtx); 852 #ifdef FULL 853 if (__cpu_simple_lock_try(&mtx->mtx_lock)) { 854 MUTEX_WANTLOCK(mtx); 855 MUTEX_LOCKED(mtx); 856 return 1; 857 } 858 MUTEX_SPIN_SPLRESTORE(mtx); 859 #else 860 MUTEX_WANTLOCK(mtx); 861 MUTEX_LOCKED(mtx); 862 return 1; 863 #endif 864 } else { 865 curthread = (uintptr_t)curlwp; 866 MUTEX_ASSERT(mtx, curthread != 0); 867 if (MUTEX_ACQUIRE(mtx, curthread)) { 868 MUTEX_WANTLOCK(mtx); 869 MUTEX_LOCKED(mtx); 870 MUTEX_DASSERT(mtx, 871 MUTEX_OWNER(mtx->mtx_owner) == curthread); 872 return 1; 873 } 874 } 875 876 return 0; 877 } 878 879 #if defined(__HAVE_SPIN_MUTEX_STUBS) || defined(FULL) 880 /* 881 * mutex_spin_retry: 882 * 883 * Support routine for mutex_spin_enter(). Assumes that the caller 884 * has already raised the SPL, and adjusted counters. 885 */ 886 void 887 mutex_spin_retry(kmutex_t *mtx) 888 { 889 #ifdef MULTIPROCESSOR 890 u_int count; 891 LOCKSTAT_TIMER(spintime); 892 LOCKSTAT_FLAG(lsflag); 893 #ifdef LOCKDEBUG 894 u_int spins = 0; 895 #endif /* LOCKDEBUG */ 896 897 MUTEX_WANTLOCK(mtx); 898 899 LOCKSTAT_ENTER(lsflag); 900 LOCKSTAT_START_TIMER(lsflag, spintime); 901 count = SPINLOCK_BACKOFF_MIN; 902 903 /* 904 * Spin testing the lock word and do exponential backoff 905 * to reduce cache line ping-ponging between CPUs. 906 */ 907 do { 908 if (panicstr != NULL) 909 break; 910 while (__SIMPLELOCK_LOCKED_P(&mtx->mtx_lock)) { 911 SPINLOCK_BACKOFF(count); 912 #ifdef LOCKDEBUG 913 if (SPINLOCK_SPINOUT(spins)) 914 MUTEX_ABORT(mtx, "spinout"); 915 #endif /* LOCKDEBUG */ 916 } 917 } while (!__cpu_simple_lock_try(&mtx->mtx_lock)); 918 919 LOCKSTAT_STOP_TIMER(lsflag, spintime); 920 LOCKSTAT_EVENT(lsflag, mtx, LB_SPIN_MUTEX | LB_SPIN, 1, spintime); 921 LOCKSTAT_EXIT(lsflag); 922 923 MUTEX_LOCKED(mtx); 924 #else /* MULTIPROCESSOR */ 925 MUTEX_ABORT(mtx, "locking against myself"); 926 #endif /* MULTIPROCESSOR */ 927 } 928 #endif /* defined(__HAVE_SPIN_MUTEX_STUBS) || defined(FULL) */ 929 930 /* 931 * mutex_obj_init: 932 * 933 * Initialize the mutex object store. 934 */ 935 void 936 mutex_obj_init(void) 937 { 938 939 mutex_obj_cache = pool_cache_init(sizeof(struct kmutexobj), 940 coherency_unit, 0, 0, "mutex", NULL, IPL_NONE, mutex_obj_ctor, 941 NULL, NULL); 942 } 943 944 /* 945 * mutex_obj_ctor: 946 * 947 * Initialize a new lock for the cache. 948 */ 949 static int 950 mutex_obj_ctor(void *arg, void *obj, int flags) 951 { 952 struct kmutexobj * mo = obj; 953 954 mo->mo_magic = MUTEX_OBJ_MAGIC; 955 956 return 0; 957 } 958 959 /* 960 * mutex_obj_alloc: 961 * 962 * Allocate a single lock object. 963 */ 964 kmutex_t * 965 mutex_obj_alloc(kmutex_type_t type, int ipl) 966 { 967 struct kmutexobj *mo; 968 969 mo = pool_cache_get(mutex_obj_cache, PR_WAITOK); 970 mutex_init(&mo->mo_lock, type, ipl); 971 mo->mo_refcnt = 1; 972 973 return (kmutex_t *)mo; 974 } 975 976 /* 977 * mutex_obj_hold: 978 * 979 * Add a single reference to a lock object. A reference to the object 980 * must already be held, and must be held across this call. 981 */ 982 void 983 mutex_obj_hold(kmutex_t *lock) 984 { 985 struct kmutexobj *mo = (struct kmutexobj *)lock; 986 987 KASSERT(mo->mo_magic == MUTEX_OBJ_MAGIC); 988 KASSERT(mo->mo_refcnt > 0); 989 990 atomic_inc_uint(&mo->mo_refcnt); 991 } 992 993 /* 994 * mutex_obj_free: 995 * 996 * Drop a reference from a lock object. If the last reference is being 997 * dropped, free the object and return true. Otherwise, return false. 998 */ 999 bool 1000 mutex_obj_free(kmutex_t *lock) 1001 { 1002 struct kmutexobj *mo = (struct kmutexobj *)lock; 1003 1004 KASSERT(mo->mo_magic == MUTEX_OBJ_MAGIC); 1005 KASSERT(mo->mo_refcnt > 0); 1006 1007 if (atomic_dec_uint_nv(&mo->mo_refcnt) > 0) { 1008 return false; 1009 } 1010 mutex_destroy(&mo->mo_lock); 1011 pool_cache_put(mutex_obj_cache, mo); 1012 return true; 1013 } 1014