1 /* $NetBSD: kern_mutex.c,v 1.92 2020/05/12 21:56:17 ad Exp $ */ 2 3 /*- 4 * Copyright (c) 2002, 2006, 2007, 2008, 2019 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 1.92 2020/05/12 21:56:17 ad Exp $"); 44 45 #include <sys/param.h> 46 #include <sys/atomic.h> 47 #include <sys/proc.h> 48 #include <sys/mutex.h> 49 #include <sys/sched.h> 50 #include <sys/sleepq.h> 51 #include <sys/systm.h> 52 #include <sys/lockdebug.h> 53 #include <sys/kernel.h> 54 #include <sys/intr.h> 55 #include <sys/lock.h> 56 #include <sys/types.h> 57 #include <sys/cpu.h> 58 #include <sys/pserialize.h> 59 60 #include <dev/lockstat.h> 61 62 #include <machine/lock.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), 0) 80 #define MUTEX_TESTLOCK(mtx) \ 81 LOCKDEBUG_WANTLOCK(MUTEX_DEBUG_P(mtx), (mtx), \ 82 (uintptr_t)__builtin_return_address(0), -1) 83 #define MUTEX_LOCKED(mtx) \ 84 LOCKDEBUG_LOCKED(MUTEX_DEBUG_P(mtx), (mtx), NULL, \ 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(__func__, __LINE__, mtx, msg) 91 92 #if defined(LOCKDEBUG) 93 94 #define MUTEX_DASSERT(mtx, cond) \ 95 do { \ 96 if (__predict_false(!(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 (__predict_false(!(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 * Some architectures can't use __cpu_simple_lock as is so allow a way 122 * for them to use an alternate definition. 123 */ 124 #ifndef MUTEX_SPINBIT_LOCK_INIT 125 #define MUTEX_SPINBIT_LOCK_INIT(mtx) __cpu_simple_lock_init(&(mtx)->mtx_lock) 126 #endif 127 #ifndef MUTEX_SPINBIT_LOCKED_P 128 #define MUTEX_SPINBIT_LOCKED_P(mtx) __SIMPLELOCK_LOCKED_P(&(mtx)->mtx_lock) 129 #endif 130 #ifndef MUTEX_SPINBIT_LOCK_TRY 131 #define MUTEX_SPINBIT_LOCK_TRY(mtx) __cpu_simple_lock_try(&(mtx)->mtx_lock) 132 #endif 133 #ifndef MUTEX_SPINBIT_LOCK_UNLOCK 134 #define MUTEX_SPINBIT_LOCK_UNLOCK(mtx) __cpu_simple_unlock(&(mtx)->mtx_lock) 135 #endif 136 137 #ifndef MUTEX_INITIALIZE_SPIN_IPL 138 #define MUTEX_INITIALIZE_SPIN_IPL(mtx, ipl) \ 139 ((mtx)->mtx_ipl = makeiplcookie((ipl))) 140 #endif 141 142 /* 143 * Spin mutex SPL save / restore. 144 */ 145 146 #define MUTEX_SPIN_SPLRAISE(mtx) \ 147 do { \ 148 struct cpu_info *x__ci; \ 149 int x__cnt, s; \ 150 s = splraiseipl(MUTEX_SPIN_IPL(mtx)); \ 151 x__ci = curcpu(); \ 152 x__cnt = x__ci->ci_mtx_count--; \ 153 __insn_barrier(); \ 154 if (x__cnt == 0) \ 155 x__ci->ci_mtx_oldspl = (s); \ 156 } while (/* CONSTCOND */ 0) 157 158 #define MUTEX_SPIN_SPLRESTORE(mtx) \ 159 do { \ 160 struct cpu_info *x__ci = curcpu(); \ 161 int s = x__ci->ci_mtx_oldspl; \ 162 __insn_barrier(); \ 163 if (++(x__ci->ci_mtx_count) == 0) \ 164 splx(s); \ 165 } while (/* CONSTCOND */ 0) 166 167 /* 168 * Memory barriers. 169 */ 170 #ifdef __HAVE_ATOMIC_AS_MEMBAR 171 #define MUTEX_MEMBAR_ENTER() 172 #define MUTEX_MEMBAR_EXIT() 173 #else 174 #define MUTEX_MEMBAR_ENTER() membar_enter() 175 #define MUTEX_MEMBAR_EXIT() membar_exit() 176 #endif 177 178 /* 179 * For architectures that provide 'simple' mutexes: they provide a 180 * CAS function that is either MP-safe, or does not need to be MP 181 * safe. Adaptive mutexes on these architectures do not require an 182 * additional interlock. 183 */ 184 185 #ifdef __HAVE_SIMPLE_MUTEXES 186 187 #define MUTEX_OWNER(owner) \ 188 (owner & MUTEX_THREAD) 189 #define MUTEX_HAS_WAITERS(mtx) \ 190 (((int)(mtx)->mtx_owner & MUTEX_BIT_WAITERS) != 0) 191 192 #define MUTEX_INITIALIZE_ADAPTIVE(mtx, dodebug) \ 193 do { \ 194 if (!dodebug) \ 195 (mtx)->mtx_owner |= MUTEX_BIT_NODEBUG; \ 196 } while (/* CONSTCOND */ 0) 197 198 #define MUTEX_INITIALIZE_SPIN(mtx, dodebug, ipl) \ 199 do { \ 200 (mtx)->mtx_owner = MUTEX_BIT_SPIN; \ 201 if (!dodebug) \ 202 (mtx)->mtx_owner |= MUTEX_BIT_NODEBUG; \ 203 MUTEX_INITIALIZE_SPIN_IPL((mtx), (ipl)); \ 204 MUTEX_SPINBIT_LOCK_INIT((mtx)); \ 205 } while (/* CONSTCOND */ 0) 206 207 #define MUTEX_DESTROY(mtx) \ 208 do { \ 209 (mtx)->mtx_owner = MUTEX_THREAD; \ 210 } while (/* CONSTCOND */ 0) 211 212 #define MUTEX_SPIN_P(owner) \ 213 (((owner) & MUTEX_BIT_SPIN) != 0) 214 #define MUTEX_ADAPTIVE_P(owner) \ 215 (((owner) & MUTEX_BIT_SPIN) == 0) 216 217 #define MUTEX_DEBUG_P(mtx) (((mtx)->mtx_owner & MUTEX_BIT_NODEBUG) == 0) 218 #if defined(LOCKDEBUG) 219 #define MUTEX_OWNED(owner) (((owner) & ~MUTEX_BIT_NODEBUG) != 0) 220 #define MUTEX_INHERITDEBUG(n, o) (n) |= (o) & MUTEX_BIT_NODEBUG 221 #else /* defined(LOCKDEBUG) */ 222 #define MUTEX_OWNED(owner) ((owner) != 0) 223 #define MUTEX_INHERITDEBUG(n, o) /* nothing */ 224 #endif /* defined(LOCKDEBUG) */ 225 226 static inline int 227 MUTEX_ACQUIRE(kmutex_t *mtx, uintptr_t curthread) 228 { 229 int rv; 230 uintptr_t oldown = 0; 231 uintptr_t newown = curthread; 232 233 MUTEX_INHERITDEBUG(oldown, mtx->mtx_owner); 234 MUTEX_INHERITDEBUG(newown, oldown); 235 rv = MUTEX_CAS(&mtx->mtx_owner, oldown, newown); 236 MUTEX_MEMBAR_ENTER(); 237 return rv; 238 } 239 240 static inline int 241 MUTEX_SET_WAITERS(kmutex_t *mtx, uintptr_t owner) 242 { 243 int rv; 244 rv = MUTEX_CAS(&mtx->mtx_owner, owner, owner | MUTEX_BIT_WAITERS); 245 MUTEX_MEMBAR_ENTER(); 246 return rv; 247 } 248 249 static inline void 250 MUTEX_RELEASE(kmutex_t *mtx) 251 { 252 uintptr_t newown; 253 254 MUTEX_MEMBAR_EXIT(); 255 newown = 0; 256 MUTEX_INHERITDEBUG(newown, mtx->mtx_owner); 257 mtx->mtx_owner = newown; 258 } 259 #endif /* __HAVE_SIMPLE_MUTEXES */ 260 261 /* 262 * Patch in stubs via strong alias where they are not available. 263 */ 264 265 #if defined(LOCKDEBUG) 266 #undef __HAVE_MUTEX_STUBS 267 #undef __HAVE_SPIN_MUTEX_STUBS 268 #endif 269 270 #ifndef __HAVE_MUTEX_STUBS 271 __strong_alias(mutex_enter,mutex_vector_enter); 272 __strong_alias(mutex_exit,mutex_vector_exit); 273 #endif 274 275 #ifndef __HAVE_SPIN_MUTEX_STUBS 276 __strong_alias(mutex_spin_enter,mutex_vector_enter); 277 __strong_alias(mutex_spin_exit,mutex_vector_exit); 278 #endif 279 280 static void mutex_abort(const char *, size_t, const kmutex_t *, 281 const char *); 282 static void mutex_dump(const volatile void *, lockop_printer_t); 283 284 lockops_t mutex_spin_lockops = { 285 .lo_name = "Mutex", 286 .lo_type = LOCKOPS_SPIN, 287 .lo_dump = mutex_dump, 288 }; 289 290 lockops_t mutex_adaptive_lockops = { 291 .lo_name = "Mutex", 292 .lo_type = LOCKOPS_SLEEP, 293 .lo_dump = mutex_dump, 294 }; 295 296 syncobj_t mutex_syncobj = { 297 .sobj_flag = SOBJ_SLEEPQ_SORTED, 298 .sobj_unsleep = turnstile_unsleep, 299 .sobj_changepri = turnstile_changepri, 300 .sobj_lendpri = sleepq_lendpri, 301 .sobj_owner = (void *)mutex_owner, 302 }; 303 304 /* 305 * mutex_dump: 306 * 307 * Dump the contents of a mutex structure. 308 */ 309 static void 310 mutex_dump(const volatile void *cookie, lockop_printer_t pr) 311 { 312 const volatile kmutex_t *mtx = cookie; 313 uintptr_t owner = mtx->mtx_owner; 314 315 pr("owner field : %#018lx wait/spin: %16d/%d\n", 316 (long)MUTEX_OWNER(owner), MUTEX_HAS_WAITERS(mtx), 317 MUTEX_SPIN_P(owner)); 318 } 319 320 /* 321 * mutex_abort: 322 * 323 * Dump information about an error and panic the system. This 324 * generates a lot of machine code in the DIAGNOSTIC case, so 325 * we ask the compiler to not inline it. 326 */ 327 static void __noinline 328 mutex_abort(const char *func, size_t line, const kmutex_t *mtx, const char *msg) 329 { 330 331 LOCKDEBUG_ABORT(func, line, mtx, (MUTEX_SPIN_P(mtx->mtx_owner) ? 332 &mutex_spin_lockops : &mutex_adaptive_lockops), msg); 333 } 334 335 /* 336 * mutex_init: 337 * 338 * Initialize a mutex for use. Note that adaptive mutexes are in 339 * essence spin mutexes that can sleep to avoid deadlock and wasting 340 * CPU time. We can't easily provide a type of mutex that always 341 * sleeps - see comments in mutex_vector_enter() about releasing 342 * mutexes unlocked. 343 */ 344 void _mutex_init(kmutex_t *, kmutex_type_t, int, uintptr_t); 345 void 346 _mutex_init(kmutex_t *mtx, kmutex_type_t type, int ipl, 347 uintptr_t return_address) 348 { 349 lockops_t *lockops __unused; 350 bool dodebug; 351 352 memset(mtx, 0, sizeof(*mtx)); 353 354 if (ipl == IPL_NONE || ipl == IPL_SOFTCLOCK || 355 ipl == IPL_SOFTBIO || ipl == IPL_SOFTNET || 356 ipl == IPL_SOFTSERIAL) { 357 lockops = (type == MUTEX_NODEBUG ? 358 NULL : &mutex_adaptive_lockops); 359 dodebug = LOCKDEBUG_ALLOC(mtx, lockops, return_address); 360 MUTEX_INITIALIZE_ADAPTIVE(mtx, dodebug); 361 } else { 362 lockops = (type == MUTEX_NODEBUG ? 363 NULL : &mutex_spin_lockops); 364 dodebug = LOCKDEBUG_ALLOC(mtx, lockops, return_address); 365 MUTEX_INITIALIZE_SPIN(mtx, dodebug, ipl); 366 } 367 } 368 369 void 370 mutex_init(kmutex_t *mtx, kmutex_type_t type, int ipl) 371 { 372 373 _mutex_init(mtx, type, ipl, (uintptr_t)__builtin_return_address(0)); 374 } 375 376 /* 377 * mutex_destroy: 378 * 379 * Tear down a mutex. 380 */ 381 void 382 mutex_destroy(kmutex_t *mtx) 383 { 384 uintptr_t owner = mtx->mtx_owner; 385 386 if (MUTEX_ADAPTIVE_P(owner)) { 387 MUTEX_ASSERT(mtx, !MUTEX_OWNED(owner)); 388 MUTEX_ASSERT(mtx, !MUTEX_HAS_WAITERS(mtx)); 389 } else { 390 MUTEX_ASSERT(mtx, !MUTEX_SPINBIT_LOCKED_P(mtx)); 391 } 392 393 LOCKDEBUG_FREE(MUTEX_DEBUG_P(mtx), mtx); 394 MUTEX_DESTROY(mtx); 395 } 396 397 #ifdef MULTIPROCESSOR 398 /* 399 * mutex_oncpu: 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 static bool 406 mutex_oncpu(uintptr_t owner) 407 { 408 struct cpu_info *ci; 409 lwp_t *l; 410 411 KASSERT(kpreempt_disabled()); 412 413 if (!MUTEX_OWNED(owner)) { 414 return false; 415 } 416 417 /* 418 * See lwp_dtor() why dereference of the LWP pointer is safe. 419 * We must have kernel preemption disabled for that. 420 */ 421 l = (lwp_t *)MUTEX_OWNER(owner); 422 ci = l->l_cpu; 423 424 if (ci && ci->ci_curlwp == l) { 425 /* Target is running; do we need to block? */ 426 return (ci->ci_biglock_wanted != l); 427 } 428 429 /* Not running. It may be safe to block now. */ 430 return false; 431 } 432 #endif /* MULTIPROCESSOR */ 433 434 /* 435 * mutex_vector_enter: 436 * 437 * Support routine for mutex_enter() that must handle all cases. In 438 * the LOCKDEBUG case, mutex_enter() is always aliased here, even if 439 * fast-path stubs are available. If a mutex_spin_enter() stub is 440 * not available, then it is also aliased directly here. 441 */ 442 void 443 mutex_vector_enter(kmutex_t *mtx) 444 { 445 uintptr_t owner, curthread; 446 turnstile_t *ts; 447 #ifdef MULTIPROCESSOR 448 u_int count; 449 #endif 450 LOCKSTAT_COUNTER(spincnt); 451 LOCKSTAT_COUNTER(slpcnt); 452 LOCKSTAT_TIMER(spintime); 453 LOCKSTAT_TIMER(slptime); 454 LOCKSTAT_FLAG(lsflag); 455 456 /* 457 * Handle spin mutexes. 458 */ 459 KPREEMPT_DISABLE(curlwp); 460 owner = mtx->mtx_owner; 461 if (MUTEX_SPIN_P(owner)) { 462 #if defined(LOCKDEBUG) && defined(MULTIPROCESSOR) 463 u_int spins = 0; 464 #endif 465 KPREEMPT_ENABLE(curlwp); 466 MUTEX_SPIN_SPLRAISE(mtx); 467 MUTEX_WANTLOCK(mtx); 468 #ifdef FULL 469 if (MUTEX_SPINBIT_LOCK_TRY(mtx)) { 470 MUTEX_LOCKED(mtx); 471 return; 472 } 473 #if !defined(MULTIPROCESSOR) 474 MUTEX_ABORT(mtx, "locking against myself"); 475 #else /* !MULTIPROCESSOR */ 476 477 LOCKSTAT_ENTER(lsflag); 478 LOCKSTAT_START_TIMER(lsflag, spintime); 479 count = SPINLOCK_BACKOFF_MIN; 480 481 /* 482 * Spin testing the lock word and do exponential backoff 483 * to reduce cache line ping-ponging between CPUs. 484 */ 485 do { 486 while (MUTEX_SPINBIT_LOCKED_P(mtx)) { 487 SPINLOCK_BACKOFF(count); 488 #ifdef LOCKDEBUG 489 if (SPINLOCK_SPINOUT(spins)) 490 MUTEX_ABORT(mtx, "spinout"); 491 #endif /* LOCKDEBUG */ 492 } 493 } while (!MUTEX_SPINBIT_LOCK_TRY(mtx)); 494 495 if (count != SPINLOCK_BACKOFF_MIN) { 496 LOCKSTAT_STOP_TIMER(lsflag, spintime); 497 LOCKSTAT_EVENT(lsflag, mtx, 498 LB_SPIN_MUTEX | LB_SPIN, 1, spintime); 499 } 500 LOCKSTAT_EXIT(lsflag); 501 #endif /* !MULTIPROCESSOR */ 502 #endif /* FULL */ 503 MUTEX_LOCKED(mtx); 504 return; 505 } 506 507 curthread = (uintptr_t)curlwp; 508 509 MUTEX_DASSERT(mtx, MUTEX_ADAPTIVE_P(owner)); 510 MUTEX_ASSERT(mtx, curthread != 0); 511 MUTEX_ASSERT(mtx, !cpu_intr_p()); 512 MUTEX_WANTLOCK(mtx); 513 514 if (panicstr == NULL) { 515 KDASSERT(pserialize_not_in_read_section()); 516 LOCKDEBUG_BARRIER(&kernel_lock, 1); 517 } 518 519 LOCKSTAT_ENTER(lsflag); 520 521 /* 522 * Adaptive mutex; spin trying to acquire the mutex. If we 523 * determine that the owner is not running on a processor, 524 * then we stop spinning, and sleep instead. 525 */ 526 for (;;) { 527 if (!MUTEX_OWNED(owner)) { 528 /* 529 * Mutex owner clear could mean two things: 530 * 531 * * The mutex has been released. 532 * * The owner field hasn't been set yet. 533 * 534 * Try to acquire it again. If that fails, 535 * we'll just loop again. 536 */ 537 if (MUTEX_ACQUIRE(mtx, curthread)) 538 break; 539 owner = mtx->mtx_owner; 540 continue; 541 } 542 if (__predict_false(MUTEX_OWNER(owner) == curthread)) { 543 MUTEX_ABORT(mtx, "locking against myself"); 544 } 545 #ifdef MULTIPROCESSOR 546 /* 547 * Check to see if the owner is running on a processor. 548 * If so, then we should just spin, as the owner will 549 * likely release the lock very soon. 550 */ 551 if (mutex_oncpu(owner)) { 552 LOCKSTAT_START_TIMER(lsflag, spintime); 553 count = SPINLOCK_BACKOFF_MIN; 554 do { 555 KPREEMPT_ENABLE(curlwp); 556 SPINLOCK_BACKOFF(count); 557 KPREEMPT_DISABLE(curlwp); 558 owner = mtx->mtx_owner; 559 } while (mutex_oncpu(owner)); 560 LOCKSTAT_STOP_TIMER(lsflag, spintime); 561 LOCKSTAT_COUNT(spincnt, 1); 562 if (!MUTEX_OWNED(owner)) 563 continue; 564 } 565 #endif 566 567 ts = turnstile_lookup(mtx); 568 569 /* 570 * Once we have the turnstile chain interlock, mark the 571 * mutex as having waiters. If that fails, spin again: 572 * chances are that the mutex has been released. 573 */ 574 if (!MUTEX_SET_WAITERS(mtx, owner)) { 575 turnstile_exit(mtx); 576 owner = mtx->mtx_owner; 577 continue; 578 } 579 580 #ifdef MULTIPROCESSOR 581 /* 582 * mutex_exit() is permitted to release the mutex without 583 * any interlocking instructions, and the following can 584 * occur as a result: 585 * 586 * CPU 1: MUTEX_SET_WAITERS() CPU2: mutex_exit() 587 * ---------------------------- ---------------------------- 588 * .. acquire cache line 589 * .. test for waiters 590 * acquire cache line <- lose cache line 591 * lock cache line .. 592 * verify mutex is held .. 593 * set waiters .. 594 * unlock cache line .. 595 * lose cache line -> acquire cache line 596 * .. clear lock word, waiters 597 * return success 598 * 599 * There is another race that can occur: a third CPU could 600 * acquire the mutex as soon as it is released. Since 601 * adaptive mutexes are primarily spin mutexes, this is not 602 * something that we need to worry about too much. What we 603 * do need to ensure is that the waiters bit gets set. 604 * 605 * To allow the unlocked release, we need to make some 606 * assumptions here: 607 * 608 * o Release is the only non-atomic/unlocked operation 609 * that can be performed on the mutex. (It must still 610 * be atomic on the local CPU, e.g. in case interrupted 611 * or preempted). 612 * 613 * o At any given time, MUTEX_SET_WAITERS() can only ever 614 * be in progress on one CPU in the system - guaranteed 615 * by the turnstile chain lock. 616 * 617 * o No other operations other than MUTEX_SET_WAITERS() 618 * and release can modify a mutex with a non-zero 619 * owner field. 620 * 621 * o The result of a successful MUTEX_SET_WAITERS() call 622 * is an unbuffered write that is immediately visible 623 * to all other processors in the system. 624 * 625 * o If the holding LWP switches away, it posts a store 626 * fence before changing curlwp, ensuring that any 627 * overwrite of the mutex waiters flag by mutex_exit() 628 * completes before the modification of curlwp becomes 629 * visible to this CPU. 630 * 631 * o mi_switch() posts a store fence before setting curlwp 632 * and before resuming execution of an LWP. 633 * 634 * o _kernel_lock() posts a store fence before setting 635 * curcpu()->ci_biglock_wanted, and after clearing it. 636 * This ensures that any overwrite of the mutex waiters 637 * flag by mutex_exit() completes before the modification 638 * of ci_biglock_wanted becomes visible. 639 * 640 * We now post a read memory barrier (after setting the 641 * waiters field) and check the lock holder's status again. 642 * Some of the possible outcomes (not an exhaustive list): 643 * 644 * 1. The on-CPU check returns true: the holding LWP is 645 * running again. The lock may be released soon and 646 * we should spin. Importantly, we can't trust the 647 * value of the waiters flag. 648 * 649 * 2. The on-CPU check returns false: the holding LWP is 650 * not running. We now have the opportunity to check 651 * if mutex_exit() has blatted the modifications made 652 * by MUTEX_SET_WAITERS(). 653 * 654 * 3. The on-CPU check returns false: the holding LWP may 655 * or may not be running. It has context switched at 656 * some point during our check. Again, we have the 657 * chance to see if the waiters bit is still set or 658 * has been overwritten. 659 * 660 * 4. The on-CPU check returns false: the holding LWP is 661 * running on a CPU, but wants the big lock. It's OK 662 * to check the waiters field in this case. 663 * 664 * 5. The has-waiters check fails: the mutex has been 665 * released, the waiters flag cleared and another LWP 666 * now owns the mutex. 667 * 668 * 6. The has-waiters check fails: the mutex has been 669 * released. 670 * 671 * If the waiters bit is not set it's unsafe to go asleep, 672 * as we might never be awoken. 673 */ 674 membar_consumer(); 675 if (mutex_oncpu(owner)) { 676 turnstile_exit(mtx); 677 owner = mtx->mtx_owner; 678 continue; 679 } 680 membar_consumer(); 681 if (!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 KPREEMPT_ENABLE(curlwp); 698 699 LOCKSTAT_EVENT(lsflag, mtx, LB_ADAPTIVE_MUTEX | LB_SLEEP1, 700 slpcnt, slptime); 701 LOCKSTAT_EVENT(lsflag, mtx, LB_ADAPTIVE_MUTEX | LB_SPIN, 702 spincnt, spintime); 703 LOCKSTAT_EXIT(lsflag); 704 705 MUTEX_DASSERT(mtx, MUTEX_OWNER(mtx->mtx_owner) == curthread); 706 MUTEX_LOCKED(mtx); 707 } 708 709 /* 710 * mutex_vector_exit: 711 * 712 * Support routine for mutex_exit() that handles all cases. 713 */ 714 void 715 mutex_vector_exit(kmutex_t *mtx) 716 { 717 turnstile_t *ts; 718 uintptr_t curthread; 719 720 if (MUTEX_SPIN_P(mtx->mtx_owner)) { 721 #ifdef FULL 722 if (__predict_false(!MUTEX_SPINBIT_LOCKED_P(mtx))) { 723 MUTEX_ABORT(mtx, "exiting unheld spin mutex"); 724 } 725 MUTEX_UNLOCKED(mtx); 726 MUTEX_SPINBIT_LOCK_UNLOCK(mtx); 727 #endif 728 MUTEX_SPIN_SPLRESTORE(mtx); 729 return; 730 } 731 732 #ifndef __HAVE_MUTEX_STUBS 733 /* 734 * On some architectures without mutex stubs, we can enter here to 735 * release mutexes before interrupts and whatnot are up and running. 736 * We need this hack to keep them sweet. 737 */ 738 if (__predict_false(cold)) { 739 MUTEX_UNLOCKED(mtx); 740 MUTEX_RELEASE(mtx); 741 return; 742 } 743 #endif 744 745 curthread = (uintptr_t)curlwp; 746 MUTEX_DASSERT(mtx, curthread != 0); 747 MUTEX_ASSERT(mtx, MUTEX_OWNER(mtx->mtx_owner) == curthread); 748 MUTEX_UNLOCKED(mtx); 749 #if !defined(LOCKDEBUG) 750 __USE(curthread); 751 #endif 752 753 #ifdef LOCKDEBUG 754 /* 755 * Avoid having to take the turnstile chain lock every time 756 * around. Raise the priority level to splhigh() in order 757 * to disable preemption and so make the following atomic. 758 */ 759 { 760 int s = splhigh(); 761 if (!MUTEX_HAS_WAITERS(mtx)) { 762 MUTEX_RELEASE(mtx); 763 splx(s); 764 return; 765 } 766 splx(s); 767 } 768 #endif 769 770 /* 771 * Get this lock's turnstile. This gets the interlock on 772 * the sleep queue. Once we have that, we can clear the 773 * lock. If there was no turnstile for the lock, there 774 * were no waiters remaining. 775 */ 776 ts = turnstile_lookup(mtx); 777 778 if (ts == NULL) { 779 MUTEX_RELEASE(mtx); 780 turnstile_exit(mtx); 781 } else { 782 MUTEX_RELEASE(mtx); 783 turnstile_wakeup(ts, TS_WRITER_Q, 784 TS_WAITERS(ts, TS_WRITER_Q), NULL); 785 } 786 } 787 788 #ifndef __HAVE_SIMPLE_MUTEXES 789 /* 790 * mutex_wakeup: 791 * 792 * Support routine for mutex_exit() that wakes up all waiters. 793 * We assume that the mutex has been released, but it need not 794 * be. 795 */ 796 void 797 mutex_wakeup(kmutex_t *mtx) 798 { 799 turnstile_t *ts; 800 801 ts = turnstile_lookup(mtx); 802 if (ts == NULL) { 803 turnstile_exit(mtx); 804 return; 805 } 806 MUTEX_CLEAR_WAITERS(mtx); 807 turnstile_wakeup(ts, TS_WRITER_Q, TS_WAITERS(ts, TS_WRITER_Q), NULL); 808 } 809 #endif /* !__HAVE_SIMPLE_MUTEXES */ 810 811 /* 812 * mutex_owned: 813 * 814 * Return true if the current LWP (adaptive) or CPU (spin) 815 * holds the mutex. 816 */ 817 int 818 mutex_owned(const kmutex_t *mtx) 819 { 820 821 if (mtx == NULL) 822 return 0; 823 if (MUTEX_ADAPTIVE_P(mtx->mtx_owner)) 824 return MUTEX_OWNER(mtx->mtx_owner) == (uintptr_t)curlwp; 825 #ifdef FULL 826 return MUTEX_SPINBIT_LOCKED_P(mtx); 827 #else 828 return 1; 829 #endif 830 } 831 832 /* 833 * mutex_owner: 834 * 835 * Return the current owner of an adaptive mutex. Used for 836 * priority inheritance. 837 */ 838 lwp_t * 839 mutex_owner(const kmutex_t *mtx) 840 { 841 842 MUTEX_ASSERT(mtx, MUTEX_ADAPTIVE_P(mtx->mtx_owner)); 843 return (struct lwp *)MUTEX_OWNER(mtx->mtx_owner); 844 } 845 846 /* 847 * mutex_owner_running: 848 * 849 * Return true if an adaptive mutex is unheld, or held and the owner is 850 * running on a CPU. For the pagedaemon only - do not document or use 851 * in other code. 852 */ 853 bool 854 mutex_owner_running(const kmutex_t *mtx) 855 { 856 #ifdef MULTIPROCESSOR 857 uintptr_t owner; 858 bool rv; 859 860 MUTEX_ASSERT(mtx, MUTEX_ADAPTIVE_P(mtx->mtx_owner)); 861 kpreempt_disable(); 862 owner = mtx->mtx_owner; 863 rv = !MUTEX_OWNED(owner) || mutex_oncpu(MUTEX_OWNER(owner)); 864 kpreempt_enable(); 865 return rv; 866 #else 867 return mutex_owner(mtx) == curlwp; 868 #endif 869 } 870 871 /* 872 * mutex_ownable: 873 * 874 * When compiled with DEBUG and LOCKDEBUG defined, ensure that 875 * the mutex is available. We cannot use !mutex_owned() since 876 * that won't work correctly for spin mutexes. 877 */ 878 int 879 mutex_ownable(const kmutex_t *mtx) 880 { 881 882 #ifdef LOCKDEBUG 883 MUTEX_TESTLOCK(mtx); 884 #endif 885 return 1; 886 } 887 888 /* 889 * mutex_tryenter: 890 * 891 * Try to acquire the mutex; return non-zero if we did. 892 */ 893 int 894 mutex_tryenter(kmutex_t *mtx) 895 { 896 uintptr_t curthread; 897 898 /* 899 * Handle spin mutexes. 900 */ 901 if (MUTEX_SPIN_P(mtx->mtx_owner)) { 902 MUTEX_SPIN_SPLRAISE(mtx); 903 #ifdef FULL 904 if (MUTEX_SPINBIT_LOCK_TRY(mtx)) { 905 MUTEX_WANTLOCK(mtx); 906 MUTEX_LOCKED(mtx); 907 return 1; 908 } 909 MUTEX_SPIN_SPLRESTORE(mtx); 910 #else 911 MUTEX_WANTLOCK(mtx); 912 MUTEX_LOCKED(mtx); 913 return 1; 914 #endif 915 } else { 916 curthread = (uintptr_t)curlwp; 917 MUTEX_ASSERT(mtx, curthread != 0); 918 if (MUTEX_ACQUIRE(mtx, curthread)) { 919 MUTEX_WANTLOCK(mtx); 920 MUTEX_LOCKED(mtx); 921 MUTEX_DASSERT(mtx, 922 MUTEX_OWNER(mtx->mtx_owner) == curthread); 923 return 1; 924 } 925 } 926 927 return 0; 928 } 929 930 #if defined(__HAVE_SPIN_MUTEX_STUBS) || defined(FULL) 931 /* 932 * mutex_spin_retry: 933 * 934 * Support routine for mutex_spin_enter(). Assumes that the caller 935 * has already raised the SPL, and adjusted counters. 936 */ 937 void 938 mutex_spin_retry(kmutex_t *mtx) 939 { 940 #ifdef MULTIPROCESSOR 941 u_int count; 942 LOCKSTAT_TIMER(spintime); 943 LOCKSTAT_FLAG(lsflag); 944 #ifdef LOCKDEBUG 945 u_int spins = 0; 946 #endif /* LOCKDEBUG */ 947 948 MUTEX_WANTLOCK(mtx); 949 950 LOCKSTAT_ENTER(lsflag); 951 LOCKSTAT_START_TIMER(lsflag, spintime); 952 count = SPINLOCK_BACKOFF_MIN; 953 954 /* 955 * Spin testing the lock word and do exponential backoff 956 * to reduce cache line ping-ponging between CPUs. 957 */ 958 do { 959 while (MUTEX_SPINBIT_LOCKED_P(mtx)) { 960 SPINLOCK_BACKOFF(count); 961 #ifdef LOCKDEBUG 962 if (SPINLOCK_SPINOUT(spins)) 963 MUTEX_ABORT(mtx, "spinout"); 964 #endif /* LOCKDEBUG */ 965 } 966 } while (!MUTEX_SPINBIT_LOCK_TRY(mtx)); 967 968 LOCKSTAT_STOP_TIMER(lsflag, spintime); 969 LOCKSTAT_EVENT(lsflag, mtx, LB_SPIN_MUTEX | LB_SPIN, 1, spintime); 970 LOCKSTAT_EXIT(lsflag); 971 972 MUTEX_LOCKED(mtx); 973 #else /* MULTIPROCESSOR */ 974 MUTEX_ABORT(mtx, "locking against myself"); 975 #endif /* MULTIPROCESSOR */ 976 } 977 #endif /* defined(__HAVE_SPIN_MUTEX_STUBS) || defined(FULL) */ 978