1 /* $NetBSD: kern_mutex.c,v 1.45 2009/01/25 04:45:14 rmind 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 1.45 2009/01/25 04:45:14 rmind 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 #include "opt_sa.h" 63 64 /* 65 * When not running a debug kernel, spin mutexes are not much 66 * more than an splraiseipl() and splx() pair. 67 */ 68 69 #if defined(DIAGNOSTIC) || defined(MULTIPROCESSOR) || defined(LOCKDEBUG) 70 #define FULL 71 #endif 72 73 /* 74 * Debugging support. 75 */ 76 77 #define MUTEX_WANTLOCK(mtx) \ 78 LOCKDEBUG_WANTLOCK(MUTEX_DEBUG_P(mtx), (mtx), \ 79 (uintptr_t)__builtin_return_address(0), false, false) 80 #define MUTEX_LOCKED(mtx) \ 81 LOCKDEBUG_LOCKED(MUTEX_DEBUG_P(mtx), (mtx), NULL, \ 82 (uintptr_t)__builtin_return_address(0), 0) 83 #define MUTEX_UNLOCKED(mtx) \ 84 LOCKDEBUG_UNLOCKED(MUTEX_DEBUG_P(mtx), (mtx), \ 85 (uintptr_t)__builtin_return_address(0), 0) 86 #define MUTEX_ABORT(mtx, msg) \ 87 mutex_abort(mtx, __func__, msg) 88 89 #if defined(LOCKDEBUG) 90 91 #define MUTEX_DASSERT(mtx, cond) \ 92 do { \ 93 if (!(cond)) \ 94 MUTEX_ABORT(mtx, "assertion failed: " #cond); \ 95 } while (/* CONSTCOND */ 0); 96 97 #else /* LOCKDEBUG */ 98 99 #define MUTEX_DASSERT(mtx, cond) /* nothing */ 100 101 #endif /* LOCKDEBUG */ 102 103 #if defined(DIAGNOSTIC) 104 105 #define MUTEX_ASSERT(mtx, cond) \ 106 do { \ 107 if (!(cond)) \ 108 MUTEX_ABORT(mtx, "assertion failed: " #cond); \ 109 } while (/* CONSTCOND */ 0) 110 111 #else /* DIAGNOSTIC */ 112 113 #define MUTEX_ASSERT(mtx, cond) /* nothing */ 114 115 #endif /* DIAGNOSTIC */ 116 117 /* 118 * Spin mutex SPL save / restore. 119 */ 120 #ifndef MUTEX_COUNT_BIAS 121 #define MUTEX_COUNT_BIAS 0 122 #endif 123 124 #define MUTEX_SPIN_SPLRAISE(mtx) \ 125 do { \ 126 struct cpu_info *x__ci; \ 127 int x__cnt, s; \ 128 s = splraiseipl(mtx->mtx_ipl); \ 129 x__ci = curcpu(); \ 130 x__cnt = x__ci->ci_mtx_count--; \ 131 __insn_barrier(); \ 132 if (x__cnt == MUTEX_COUNT_BIAS) \ 133 x__ci->ci_mtx_oldspl = (s); \ 134 } while (/* CONSTCOND */ 0) 135 136 #define MUTEX_SPIN_SPLRESTORE(mtx) \ 137 do { \ 138 struct cpu_info *x__ci = curcpu(); \ 139 int s = x__ci->ci_mtx_oldspl; \ 140 __insn_barrier(); \ 141 if (++(x__ci->ci_mtx_count) == MUTEX_COUNT_BIAS) \ 142 splx(s); \ 143 } while (/* CONSTCOND */ 0) 144 145 /* 146 * For architectures that provide 'simple' mutexes: they provide a 147 * CAS function that is either MP-safe, or does not need to be MP 148 * safe. Adaptive mutexes on these architectures do not require an 149 * additional interlock. 150 */ 151 152 #ifdef __HAVE_SIMPLE_MUTEXES 153 154 #define MUTEX_OWNER(owner) \ 155 (owner & MUTEX_THREAD) 156 #define MUTEX_HAS_WAITERS(mtx) \ 157 (((int)(mtx)->mtx_owner & MUTEX_BIT_WAITERS) != 0) 158 159 #define MUTEX_INITIALIZE_ADAPTIVE(mtx, dodebug) \ 160 do { \ 161 if (dodebug) \ 162 (mtx)->mtx_owner |= MUTEX_BIT_DEBUG; \ 163 } while (/* CONSTCOND */ 0); 164 165 #define MUTEX_INITIALIZE_SPIN(mtx, dodebug, ipl) \ 166 do { \ 167 (mtx)->mtx_owner = MUTEX_BIT_SPIN; \ 168 if (dodebug) \ 169 (mtx)->mtx_owner |= MUTEX_BIT_DEBUG; \ 170 (mtx)->mtx_ipl = makeiplcookie((ipl)); \ 171 __cpu_simple_lock_init(&(mtx)->mtx_lock); \ 172 } while (/* CONSTCOND */ 0) 173 174 #define MUTEX_DESTROY(mtx) \ 175 do { \ 176 (mtx)->mtx_owner = MUTEX_THREAD; \ 177 } while (/* CONSTCOND */ 0); 178 179 #define MUTEX_SPIN_P(mtx) \ 180 (((mtx)->mtx_owner & MUTEX_BIT_SPIN) != 0) 181 #define MUTEX_ADAPTIVE_P(mtx) \ 182 (((mtx)->mtx_owner & MUTEX_BIT_SPIN) == 0) 183 184 #define MUTEX_DEBUG_P(mtx) (((mtx)->mtx_owner & MUTEX_BIT_DEBUG) != 0) 185 #if defined(LOCKDEBUG) 186 #define MUTEX_OWNED(owner) (((owner) & ~MUTEX_BIT_DEBUG) != 0) 187 #define MUTEX_INHERITDEBUG(new, old) (new) |= (old) & MUTEX_BIT_DEBUG 188 #else /* defined(LOCKDEBUG) */ 189 #define MUTEX_OWNED(owner) ((owner) != 0) 190 #define MUTEX_INHERITDEBUG(new, old) /* nothing */ 191 #endif /* defined(LOCKDEBUG) */ 192 193 static inline int 194 MUTEX_ACQUIRE(kmutex_t *mtx, uintptr_t curthread) 195 { 196 int rv; 197 uintptr_t old = 0; 198 uintptr_t new = curthread; 199 200 MUTEX_INHERITDEBUG(old, mtx->mtx_owner); 201 MUTEX_INHERITDEBUG(new, old); 202 rv = MUTEX_CAS(&mtx->mtx_owner, old, new); 203 MUTEX_RECEIVE(mtx); 204 return rv; 205 } 206 207 static inline int 208 MUTEX_SET_WAITERS(kmutex_t *mtx, uintptr_t owner) 209 { 210 int rv; 211 rv = MUTEX_CAS(&mtx->mtx_owner, owner, owner | MUTEX_BIT_WAITERS); 212 MUTEX_RECEIVE(mtx); 213 return rv; 214 } 215 216 static inline void 217 MUTEX_RELEASE(kmutex_t *mtx) 218 { 219 uintptr_t new; 220 221 MUTEX_GIVE(mtx); 222 new = 0; 223 MUTEX_INHERITDEBUG(new, mtx->mtx_owner); 224 mtx->mtx_owner = new; 225 } 226 227 static inline void 228 MUTEX_CLEAR_WAITERS(kmutex_t *mtx) 229 { 230 /* nothing */ 231 } 232 #endif /* __HAVE_SIMPLE_MUTEXES */ 233 234 /* 235 * Patch in stubs via strong alias where they are not available. 236 */ 237 238 #if defined(LOCKDEBUG) 239 #undef __HAVE_MUTEX_STUBS 240 #undef __HAVE_SPIN_MUTEX_STUBS 241 #endif 242 243 #ifndef __HAVE_MUTEX_STUBS 244 __strong_alias(mutex_enter,mutex_vector_enter); 245 __strong_alias(mutex_exit,mutex_vector_exit); 246 #endif 247 248 #ifndef __HAVE_SPIN_MUTEX_STUBS 249 __strong_alias(mutex_spin_enter,mutex_vector_enter); 250 __strong_alias(mutex_spin_exit,mutex_vector_exit); 251 #endif 252 253 void mutex_abort(kmutex_t *, const char *, const char *); 254 void mutex_dump(volatile void *); 255 int mutex_onproc(uintptr_t, struct cpu_info **); 256 257 lockops_t mutex_spin_lockops = { 258 "Mutex", 259 LOCKOPS_SPIN, 260 mutex_dump 261 }; 262 263 lockops_t mutex_adaptive_lockops = { 264 "Mutex", 265 LOCKOPS_SLEEP, 266 mutex_dump 267 }; 268 269 syncobj_t mutex_syncobj = { 270 SOBJ_SLEEPQ_SORTED, 271 turnstile_unsleep, 272 turnstile_changepri, 273 sleepq_lendpri, 274 (void *)mutex_owner, 275 }; 276 277 /* Mutex cache */ 278 #define MUTEX_OBJ_MAGIC 0x5aa3c85d 279 struct kmutexobj { 280 kmutex_t mo_lock; 281 u_int mo_magic; 282 u_int mo_refcnt; 283 }; 284 285 static int mutex_obj_ctor(void *, void *, int); 286 287 static pool_cache_t mutex_obj_cache; 288 289 /* 290 * mutex_dump: 291 * 292 * Dump the contents of a mutex structure. 293 */ 294 void 295 mutex_dump(volatile void *cookie) 296 { 297 volatile kmutex_t *mtx = cookie; 298 299 printf_nolog("owner field : %#018lx wait/spin: %16d/%d\n", 300 (long)MUTEX_OWNER(mtx->mtx_owner), MUTEX_HAS_WAITERS(mtx), 301 MUTEX_SPIN_P(mtx)); 302 } 303 304 /* 305 * mutex_abort: 306 * 307 * Dump information about an error and panic the system. This 308 * generates a lot of machine code in the DIAGNOSTIC case, so 309 * we ask the compiler to not inline it. 310 */ 311 void __noinline 312 mutex_abort(kmutex_t *mtx, const char *func, const char *msg) 313 { 314 315 LOCKDEBUG_ABORT(mtx, (MUTEX_SPIN_P(mtx) ? 316 &mutex_spin_lockops : &mutex_adaptive_lockops), func, msg); 317 } 318 319 /* 320 * mutex_init: 321 * 322 * Initialize a mutex for use. Note that adaptive mutexes are in 323 * essence spin mutexes that can sleep to avoid deadlock and wasting 324 * CPU time. We can't easily provide a type of mutex that always 325 * sleeps - see comments in mutex_vector_enter() about releasing 326 * mutexes unlocked. 327 */ 328 void 329 mutex_init(kmutex_t *mtx, kmutex_type_t type, int ipl) 330 { 331 bool dodebug; 332 333 memset(mtx, 0, sizeof(*mtx)); 334 335 switch (type) { 336 case MUTEX_ADAPTIVE: 337 KASSERT(ipl == IPL_NONE); 338 break; 339 case MUTEX_DEFAULT: 340 case MUTEX_DRIVER: 341 if (ipl == IPL_NONE || ipl == IPL_SOFTCLOCK || 342 ipl == IPL_SOFTBIO || ipl == IPL_SOFTNET || 343 ipl == IPL_SOFTSERIAL) { 344 type = MUTEX_ADAPTIVE; 345 } else { 346 type = MUTEX_SPIN; 347 } 348 break; 349 default: 350 break; 351 } 352 353 switch (type) { 354 case MUTEX_NODEBUG: 355 dodebug = LOCKDEBUG_ALLOC(mtx, NULL, 356 (uintptr_t)__builtin_return_address(0)); 357 MUTEX_INITIALIZE_SPIN(mtx, dodebug, ipl); 358 break; 359 case MUTEX_ADAPTIVE: 360 dodebug = LOCKDEBUG_ALLOC(mtx, &mutex_adaptive_lockops, 361 (uintptr_t)__builtin_return_address(0)); 362 MUTEX_INITIALIZE_ADAPTIVE(mtx, dodebug); 363 break; 364 case MUTEX_SPIN: 365 dodebug = LOCKDEBUG_ALLOC(mtx, &mutex_spin_lockops, 366 (uintptr_t)__builtin_return_address(0)); 367 MUTEX_INITIALIZE_SPIN(mtx, dodebug, ipl); 368 break; 369 default: 370 panic("mutex_init: impossible type"); 371 break; 372 } 373 } 374 375 /* 376 * mutex_destroy: 377 * 378 * Tear down a mutex. 379 */ 380 void 381 mutex_destroy(kmutex_t *mtx) 382 { 383 384 if (MUTEX_ADAPTIVE_P(mtx)) { 385 MUTEX_ASSERT(mtx, !MUTEX_OWNED(mtx->mtx_owner) && 386 !MUTEX_HAS_WAITERS(mtx)); 387 } else { 388 MUTEX_ASSERT(mtx, !__SIMPLELOCK_LOCKED_P(&mtx->mtx_lock)); 389 } 390 391 LOCKDEBUG_FREE(MUTEX_DEBUG_P(mtx), mtx); 392 MUTEX_DESTROY(mtx); 393 } 394 395 /* 396 * mutex_onproc: 397 * 398 * Return true if an adaptive mutex owner is running on a CPU in the 399 * system. If the target is waiting on the kernel big lock, then we 400 * must release it. This is necessary to avoid deadlock. 401 * 402 * Note that we can't use the mutex owner field as an LWP pointer. We 403 * don't have full control over the timing of our execution, and so the 404 * pointer could be completely invalid by the time we dereference it. 405 */ 406 #ifdef MULTIPROCESSOR 407 int 408 mutex_onproc(uintptr_t owner, struct cpu_info **cip) 409 { 410 CPU_INFO_ITERATOR cii; 411 struct cpu_info *ci; 412 struct lwp *l; 413 414 if (!MUTEX_OWNED(owner)) 415 return 0; 416 l = (struct lwp *)MUTEX_OWNER(owner); 417 418 /* See if the target is running on a CPU somewhere. */ 419 if ((ci = *cip) != NULL && ci->ci_curlwp == l) 420 goto run; 421 for (CPU_INFO_FOREACH(cii, ci)) 422 if (ci->ci_curlwp == l) 423 goto run; 424 425 /* No: it may be safe to block now. */ 426 *cip = NULL; 427 return 0; 428 429 run: 430 /* Target is running; do we need to block? */ 431 *cip = ci; 432 return ci->ci_biglock_wanted != l; 433 } 434 #endif /* MULTIPROCESSOR */ 435 436 /* 437 * mutex_vector_enter: 438 * 439 * Support routine for mutex_enter() that must handle all cases. In 440 * the LOCKDEBUG case, mutex_enter() is always aliased here, even if 441 * fast-path stubs are available. If an mutex_spin_enter() stub is 442 * not available, then it is also aliased directly here. 443 */ 444 void 445 mutex_vector_enter(kmutex_t *mtx) 446 { 447 uintptr_t owner, curthread; 448 turnstile_t *ts; 449 #ifdef MULTIPROCESSOR 450 struct cpu_info *ci = NULL; 451 u_int count; 452 #endif 453 #ifdef KERN_SA 454 int f; 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 (__predict_false(panicstr != NULL)) 547 return; 548 if (__predict_false(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 #ifdef KERN_SA 689 /* 690 * Sleeping for a mutex should not generate an upcall. 691 * So set LP_SA_NOBLOCK to indicate this. 692 * f indicates if we should clear LP_SA_NOBLOCK when done. 693 */ 694 f = ~curlwp->l_pflag & LP_SA_NOBLOCK; 695 curlwp->l_pflag |= LP_SA_NOBLOCK; 696 #endif /* KERN_SA */ 697 698 LOCKSTAT_START_TIMER(lsflag, slptime); 699 700 turnstile_block(ts, TS_WRITER_Q, mtx, &mutex_syncobj); 701 702 LOCKSTAT_STOP_TIMER(lsflag, slptime); 703 LOCKSTAT_COUNT(slpcnt, 1); 704 705 #ifdef KERN_SA 706 curlwp->l_pflag ^= f; 707 #endif /* KERN_SA */ 708 709 owner = mtx->mtx_owner; 710 } 711 712 LOCKSTAT_EVENT(lsflag, mtx, LB_ADAPTIVE_MUTEX | LB_SLEEP1, 713 slpcnt, slptime); 714 LOCKSTAT_EVENT(lsflag, mtx, LB_ADAPTIVE_MUTEX | LB_SPIN, 715 spincnt, spintime); 716 LOCKSTAT_EXIT(lsflag); 717 718 MUTEX_DASSERT(mtx, MUTEX_OWNER(mtx->mtx_owner) == curthread); 719 MUTEX_LOCKED(mtx); 720 } 721 722 /* 723 * mutex_vector_exit: 724 * 725 * Support routine for mutex_exit() that handles all cases. 726 */ 727 void 728 mutex_vector_exit(kmutex_t *mtx) 729 { 730 turnstile_t *ts; 731 uintptr_t curthread; 732 733 if (MUTEX_SPIN_P(mtx)) { 734 #ifdef FULL 735 if (__predict_false(!__SIMPLELOCK_LOCKED_P(&mtx->mtx_lock))) { 736 if (panicstr != NULL) 737 return; 738 MUTEX_ABORT(mtx, "exiting unheld spin mutex"); 739 } 740 MUTEX_UNLOCKED(mtx); 741 __cpu_simple_unlock(&mtx->mtx_lock); 742 #endif 743 MUTEX_SPIN_SPLRESTORE(mtx); 744 return; 745 } 746 747 if (__predict_false((uintptr_t)panicstr | cold)) { 748 MUTEX_UNLOCKED(mtx); 749 MUTEX_RELEASE(mtx); 750 return; 751 } 752 753 curthread = (uintptr_t)curlwp; 754 MUTEX_DASSERT(mtx, curthread != 0); 755 MUTEX_ASSERT(mtx, MUTEX_OWNER(mtx->mtx_owner) == curthread); 756 MUTEX_UNLOCKED(mtx); 757 758 #ifdef LOCKDEBUG 759 /* 760 * Avoid having to take the turnstile chain lock every time 761 * around. Raise the priority level to splhigh() in order 762 * to disable preemption and so make the following atomic. 763 */ 764 { 765 int s = splhigh(); 766 if (!MUTEX_HAS_WAITERS(mtx)) { 767 MUTEX_RELEASE(mtx); 768 splx(s); 769 return; 770 } 771 splx(s); 772 } 773 #endif 774 775 /* 776 * Get this lock's turnstile. This gets the interlock on 777 * the sleep queue. Once we have that, we can clear the 778 * lock. If there was no turnstile for the lock, there 779 * were no waiters remaining. 780 */ 781 ts = turnstile_lookup(mtx); 782 783 if (ts == NULL) { 784 MUTEX_RELEASE(mtx); 785 turnstile_exit(mtx); 786 } else { 787 MUTEX_RELEASE(mtx); 788 turnstile_wakeup(ts, TS_WRITER_Q, 789 TS_WAITERS(ts, TS_WRITER_Q), NULL); 790 } 791 } 792 793 #ifndef __HAVE_SIMPLE_MUTEXES 794 /* 795 * mutex_wakeup: 796 * 797 * Support routine for mutex_exit() that wakes up all waiters. 798 * We assume that the mutex has been released, but it need not 799 * be. 800 */ 801 void 802 mutex_wakeup(kmutex_t *mtx) 803 { 804 turnstile_t *ts; 805 806 ts = turnstile_lookup(mtx); 807 if (ts == NULL) { 808 turnstile_exit(mtx); 809 return; 810 } 811 MUTEX_CLEAR_WAITERS(mtx); 812 turnstile_wakeup(ts, TS_WRITER_Q, TS_WAITERS(ts, TS_WRITER_Q), NULL); 813 } 814 #endif /* !__HAVE_SIMPLE_MUTEXES */ 815 816 /* 817 * mutex_owned: 818 * 819 * Return true if the current LWP (adaptive) or CPU (spin) 820 * holds the mutex. 821 */ 822 int 823 mutex_owned(kmutex_t *mtx) 824 { 825 826 if (mtx == NULL) 827 return 0; 828 if (MUTEX_ADAPTIVE_P(mtx)) 829 return MUTEX_OWNER(mtx->mtx_owner) == (uintptr_t)curlwp; 830 #ifdef FULL 831 return __SIMPLELOCK_LOCKED_P(&mtx->mtx_lock); 832 #else 833 return 1; 834 #endif 835 } 836 837 /* 838 * mutex_owner: 839 * 840 * Return the current owner of an adaptive mutex. Used for 841 * priority inheritance. 842 */ 843 lwp_t * 844 mutex_owner(kmutex_t *mtx) 845 { 846 847 MUTEX_ASSERT(mtx, MUTEX_ADAPTIVE_P(mtx)); 848 return (struct lwp *)MUTEX_OWNER(mtx->mtx_owner); 849 } 850 851 /* 852 * mutex_tryenter: 853 * 854 * Try to acquire the mutex; return non-zero if we did. 855 */ 856 int 857 mutex_tryenter(kmutex_t *mtx) 858 { 859 uintptr_t curthread; 860 861 /* 862 * Handle spin mutexes. 863 */ 864 if (MUTEX_SPIN_P(mtx)) { 865 MUTEX_SPIN_SPLRAISE(mtx); 866 #ifdef FULL 867 if (__cpu_simple_lock_try(&mtx->mtx_lock)) { 868 MUTEX_WANTLOCK(mtx); 869 MUTEX_LOCKED(mtx); 870 return 1; 871 } 872 MUTEX_SPIN_SPLRESTORE(mtx); 873 #else 874 MUTEX_WANTLOCK(mtx); 875 MUTEX_LOCKED(mtx); 876 return 1; 877 #endif 878 } else { 879 curthread = (uintptr_t)curlwp; 880 MUTEX_ASSERT(mtx, curthread != 0); 881 if (MUTEX_ACQUIRE(mtx, curthread)) { 882 MUTEX_WANTLOCK(mtx); 883 MUTEX_LOCKED(mtx); 884 MUTEX_DASSERT(mtx, 885 MUTEX_OWNER(mtx->mtx_owner) == curthread); 886 return 1; 887 } 888 } 889 890 return 0; 891 } 892 893 #if defined(__HAVE_SPIN_MUTEX_STUBS) || defined(FULL) 894 /* 895 * mutex_spin_retry: 896 * 897 * Support routine for mutex_spin_enter(). Assumes that the caller 898 * has already raised the SPL, and adjusted counters. 899 */ 900 void 901 mutex_spin_retry(kmutex_t *mtx) 902 { 903 #ifdef MULTIPROCESSOR 904 u_int count; 905 LOCKSTAT_TIMER(spintime); 906 LOCKSTAT_FLAG(lsflag); 907 #ifdef LOCKDEBUG 908 u_int spins = 0; 909 #endif /* LOCKDEBUG */ 910 911 MUTEX_WANTLOCK(mtx); 912 913 LOCKSTAT_ENTER(lsflag); 914 LOCKSTAT_START_TIMER(lsflag, spintime); 915 count = SPINLOCK_BACKOFF_MIN; 916 917 /* 918 * Spin testing the lock word and do exponential backoff 919 * to reduce cache line ping-ponging between CPUs. 920 */ 921 do { 922 if (panicstr != NULL) 923 break; 924 while (__SIMPLELOCK_LOCKED_P(&mtx->mtx_lock)) { 925 SPINLOCK_BACKOFF(count); 926 #ifdef LOCKDEBUG 927 if (SPINLOCK_SPINOUT(spins)) 928 MUTEX_ABORT(mtx, "spinout"); 929 #endif /* LOCKDEBUG */ 930 } 931 } while (!__cpu_simple_lock_try(&mtx->mtx_lock)); 932 933 LOCKSTAT_STOP_TIMER(lsflag, spintime); 934 LOCKSTAT_EVENT(lsflag, mtx, LB_SPIN_MUTEX | LB_SPIN, 1, spintime); 935 LOCKSTAT_EXIT(lsflag); 936 937 MUTEX_LOCKED(mtx); 938 #else /* MULTIPROCESSOR */ 939 MUTEX_ABORT(mtx, "locking against myself"); 940 #endif /* MULTIPROCESSOR */ 941 } 942 #endif /* defined(__HAVE_SPIN_MUTEX_STUBS) || defined(FULL) */ 943 944 /* 945 * mutex_obj_init: 946 * 947 * Initialize the mutex object store. 948 */ 949 void 950 mutex_obj_init(void) 951 { 952 953 mutex_obj_cache = pool_cache_init(sizeof(struct kmutexobj), 954 coherency_unit, 0, 0, "mutex", NULL, IPL_NONE, mutex_obj_ctor, 955 NULL, NULL); 956 } 957 958 /* 959 * mutex_obj_ctor: 960 * 961 * Initialize a new lock for the cache. 962 */ 963 static int 964 mutex_obj_ctor(void *arg, void *obj, int flags) 965 { 966 struct kmutexobj * mo = obj; 967 968 mo->mo_magic = MUTEX_OBJ_MAGIC; 969 970 return 0; 971 } 972 973 /* 974 * mutex_obj_alloc: 975 * 976 * Allocate a single lock object. 977 */ 978 kmutex_t * 979 mutex_obj_alloc(kmutex_type_t type, int ipl) 980 { 981 struct kmutexobj *mo; 982 983 mo = pool_cache_get(mutex_obj_cache, PR_WAITOK); 984 mutex_init(&mo->mo_lock, type, ipl); 985 mo->mo_refcnt = 1; 986 987 return (kmutex_t *)mo; 988 } 989 990 /* 991 * mutex_obj_hold: 992 * 993 * Add a single reference to a lock object. A reference to the object 994 * must already be held, and must be held across this call. 995 */ 996 void 997 mutex_obj_hold(kmutex_t *lock) 998 { 999 struct kmutexobj *mo = (struct kmutexobj *)lock; 1000 1001 KASSERT(mo->mo_magic == MUTEX_OBJ_MAGIC); 1002 KASSERT(mo->mo_refcnt > 0); 1003 1004 atomic_inc_uint(&mo->mo_refcnt); 1005 } 1006 1007 /* 1008 * mutex_obj_free: 1009 * 1010 * Drop a reference from a lock object. If the last reference is being 1011 * dropped, free the object and return true. Otherwise, return false. 1012 */ 1013 bool 1014 mutex_obj_free(kmutex_t *lock) 1015 { 1016 struct kmutexobj *mo = (struct kmutexobj *)lock; 1017 1018 KASSERT(mo->mo_magic == MUTEX_OBJ_MAGIC); 1019 KASSERT(mo->mo_refcnt > 0); 1020 1021 if (atomic_dec_uint_nv(&mo->mo_refcnt) > 0) { 1022 return false; 1023 } 1024 mutex_destroy(&mo->mo_lock); 1025 pool_cache_put(mutex_obj_cache, mo); 1026 return true; 1027 } 1028