1 /* $NetBSD: kern_mutex.c,v 1.100 2022/10/26 23:21:19 riastradh 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 1.100 2022/10/26 23:21:19 riastradh 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_ACQUIRE() 173 #define MUTEX_MEMBAR_RELEASE() 174 #else 175 #define MUTEX_MEMBAR_ENTER() membar_enter() 176 #define MUTEX_MEMBAR_ACQUIRE() membar_acquire() 177 #define MUTEX_MEMBAR_RELEASE() membar_release() 178 #endif 179 180 /* 181 * For architectures that provide 'simple' mutexes: they provide a 182 * CAS function that is either MP-safe, or does not need to be MP 183 * safe. Adaptive mutexes on these architectures do not require an 184 * additional interlock. 185 */ 186 187 #ifdef __HAVE_SIMPLE_MUTEXES 188 189 #define MUTEX_OWNER(owner) \ 190 (owner & MUTEX_THREAD) 191 #define MUTEX_HAS_WAITERS(mtx) \ 192 (((int)(mtx)->mtx_owner & MUTEX_BIT_WAITERS) != 0) 193 194 #define MUTEX_INITIALIZE_ADAPTIVE(mtx, dodebug) \ 195 do { \ 196 if (!dodebug) \ 197 (mtx)->mtx_owner |= MUTEX_BIT_NODEBUG; \ 198 } while (/* CONSTCOND */ 0) 199 200 #define MUTEX_INITIALIZE_SPIN(mtx, dodebug, ipl) \ 201 do { \ 202 (mtx)->mtx_owner = MUTEX_BIT_SPIN; \ 203 if (!dodebug) \ 204 (mtx)->mtx_owner |= MUTEX_BIT_NODEBUG; \ 205 MUTEX_INITIALIZE_SPIN_IPL((mtx), (ipl)); \ 206 MUTEX_SPINBIT_LOCK_INIT((mtx)); \ 207 } while (/* CONSTCOND */ 0) 208 209 #define MUTEX_DESTROY(mtx) \ 210 do { \ 211 (mtx)->mtx_owner = MUTEX_THREAD; \ 212 } while (/* CONSTCOND */ 0) 213 214 #define MUTEX_SPIN_P(owner) \ 215 (((owner) & MUTEX_BIT_SPIN) != 0) 216 #define MUTEX_ADAPTIVE_P(owner) \ 217 (((owner) & MUTEX_BIT_SPIN) == 0) 218 219 #ifndef MUTEX_CAS 220 #define MUTEX_CAS(p, o, n) \ 221 (atomic_cas_ulong((volatile unsigned long *)(p), (o), (n)) == (o)) 222 #endif /* MUTEX_CAS */ 223 224 #define MUTEX_DEBUG_P(mtx) (((mtx)->mtx_owner & MUTEX_BIT_NODEBUG) == 0) 225 #if defined(LOCKDEBUG) 226 #define MUTEX_OWNED(owner) (((owner) & ~MUTEX_BIT_NODEBUG) != 0) 227 #define MUTEX_INHERITDEBUG(n, o) (n) |= (o) & MUTEX_BIT_NODEBUG 228 #else /* defined(LOCKDEBUG) */ 229 #define MUTEX_OWNED(owner) ((owner) != 0) 230 #define MUTEX_INHERITDEBUG(n, o) /* nothing */ 231 #endif /* defined(LOCKDEBUG) */ 232 233 static inline int 234 MUTEX_ACQUIRE(kmutex_t *mtx, uintptr_t curthread) 235 { 236 int rv; 237 uintptr_t oldown = 0; 238 uintptr_t newown = curthread; 239 240 MUTEX_INHERITDEBUG(oldown, mtx->mtx_owner); 241 MUTEX_INHERITDEBUG(newown, oldown); 242 rv = MUTEX_CAS(&mtx->mtx_owner, oldown, newown); 243 MUTEX_MEMBAR_ACQUIRE(); 244 return rv; 245 } 246 247 static inline int 248 MUTEX_SET_WAITERS(kmutex_t *mtx, uintptr_t owner) 249 { 250 int rv; 251 rv = MUTEX_CAS(&mtx->mtx_owner, owner, owner | MUTEX_BIT_WAITERS); 252 MUTEX_MEMBAR_ENTER(); 253 return rv; 254 } 255 256 static inline void 257 MUTEX_RELEASE(kmutex_t *mtx) 258 { 259 uintptr_t newown; 260 261 MUTEX_MEMBAR_RELEASE(); 262 newown = 0; 263 MUTEX_INHERITDEBUG(newown, mtx->mtx_owner); 264 mtx->mtx_owner = newown; 265 } 266 #endif /* __HAVE_SIMPLE_MUTEXES */ 267 268 /* 269 * Patch in stubs via strong alias where they are not available. 270 */ 271 272 #if defined(LOCKDEBUG) 273 #undef __HAVE_MUTEX_STUBS 274 #undef __HAVE_SPIN_MUTEX_STUBS 275 #endif 276 277 #ifndef __HAVE_MUTEX_STUBS 278 __strong_alias(mutex_enter,mutex_vector_enter); 279 __strong_alias(mutex_exit,mutex_vector_exit); 280 #endif 281 282 #ifndef __HAVE_SPIN_MUTEX_STUBS 283 __strong_alias(mutex_spin_enter,mutex_vector_enter); 284 __strong_alias(mutex_spin_exit,mutex_vector_exit); 285 #endif 286 287 static void mutex_abort(const char *, size_t, const kmutex_t *, 288 const char *); 289 static void mutex_dump(const volatile void *, lockop_printer_t); 290 291 lockops_t mutex_spin_lockops = { 292 .lo_name = "Mutex", 293 .lo_type = LOCKOPS_SPIN, 294 .lo_dump = mutex_dump, 295 }; 296 297 lockops_t mutex_adaptive_lockops = { 298 .lo_name = "Mutex", 299 .lo_type = LOCKOPS_SLEEP, 300 .lo_dump = mutex_dump, 301 }; 302 303 syncobj_t mutex_syncobj = { 304 .sobj_flag = SOBJ_SLEEPQ_SORTED, 305 .sobj_unsleep = turnstile_unsleep, 306 .sobj_changepri = turnstile_changepri, 307 .sobj_lendpri = sleepq_lendpri, 308 .sobj_owner = (void *)mutex_owner, 309 }; 310 311 /* 312 * mutex_dump: 313 * 314 * Dump the contents of a mutex structure. 315 */ 316 static void 317 mutex_dump(const volatile void *cookie, lockop_printer_t pr) 318 { 319 const volatile kmutex_t *mtx = cookie; 320 uintptr_t owner = mtx->mtx_owner; 321 322 pr("owner field : %#018lx wait/spin: %16d/%d\n", 323 (long)MUTEX_OWNER(owner), MUTEX_HAS_WAITERS(mtx), 324 MUTEX_SPIN_P(owner)); 325 } 326 327 /* 328 * mutex_abort: 329 * 330 * Dump information about an error and panic the system. This 331 * generates a lot of machine code in the DIAGNOSTIC case, so 332 * we ask the compiler to not inline it. 333 */ 334 static void __noinline 335 mutex_abort(const char *func, size_t line, const kmutex_t *mtx, const char *msg) 336 { 337 338 LOCKDEBUG_ABORT(func, line, mtx, (MUTEX_SPIN_P(mtx->mtx_owner) ? 339 &mutex_spin_lockops : &mutex_adaptive_lockops), msg); 340 } 341 342 /* 343 * mutex_init: 344 * 345 * Initialize a mutex for use. Note that adaptive mutexes are in 346 * essence spin mutexes that can sleep to avoid deadlock and wasting 347 * CPU time. We can't easily provide a type of mutex that always 348 * sleeps - see comments in mutex_vector_enter() about releasing 349 * mutexes unlocked. 350 */ 351 void 352 _mutex_init(kmutex_t *mtx, kmutex_type_t type, int ipl, 353 uintptr_t return_address) 354 { 355 lockops_t *lockops __unused; 356 bool dodebug; 357 358 memset(mtx, 0, sizeof(*mtx)); 359 360 if (ipl == IPL_NONE || ipl == IPL_SOFTCLOCK || 361 ipl == IPL_SOFTBIO || ipl == IPL_SOFTNET || 362 ipl == IPL_SOFTSERIAL) { 363 lockops = (type == MUTEX_NODEBUG ? 364 NULL : &mutex_adaptive_lockops); 365 dodebug = LOCKDEBUG_ALLOC(mtx, lockops, return_address); 366 MUTEX_INITIALIZE_ADAPTIVE(mtx, dodebug); 367 } else { 368 lockops = (type == MUTEX_NODEBUG ? 369 NULL : &mutex_spin_lockops); 370 dodebug = LOCKDEBUG_ALLOC(mtx, lockops, return_address); 371 MUTEX_INITIALIZE_SPIN(mtx, dodebug, ipl); 372 } 373 } 374 375 void 376 mutex_init(kmutex_t *mtx, kmutex_type_t type, int ipl) 377 { 378 379 _mutex_init(mtx, type, ipl, (uintptr_t)__builtin_return_address(0)); 380 } 381 382 /* 383 * mutex_destroy: 384 * 385 * Tear down a mutex. 386 */ 387 void 388 mutex_destroy(kmutex_t *mtx) 389 { 390 uintptr_t owner = mtx->mtx_owner; 391 392 if (MUTEX_ADAPTIVE_P(owner)) { 393 MUTEX_ASSERT(mtx, !MUTEX_OWNED(owner)); 394 MUTEX_ASSERT(mtx, !MUTEX_HAS_WAITERS(mtx)); 395 } else { 396 MUTEX_ASSERT(mtx, !MUTEX_SPINBIT_LOCKED_P(mtx)); 397 } 398 399 LOCKDEBUG_FREE(MUTEX_DEBUG_P(mtx), mtx); 400 MUTEX_DESTROY(mtx); 401 } 402 403 #ifdef MULTIPROCESSOR 404 /* 405 * mutex_oncpu: 406 * 407 * Return true if an adaptive mutex owner is running on a CPU in the 408 * system. If the target is waiting on the kernel big lock, then we 409 * must release it. This is necessary to avoid deadlock. 410 */ 411 static bool 412 mutex_oncpu(uintptr_t owner) 413 { 414 struct cpu_info *ci; 415 lwp_t *l; 416 417 KASSERT(kpreempt_disabled()); 418 419 if (!MUTEX_OWNED(owner)) { 420 return false; 421 } 422 423 /* 424 * See lwp_dtor() why dereference of the LWP pointer is safe. 425 * We must have kernel preemption disabled for that. 426 */ 427 l = (lwp_t *)MUTEX_OWNER(owner); 428 ci = l->l_cpu; 429 430 if (ci && ci->ci_curlwp == l) { 431 /* Target is running; do we need to block? */ 432 return (ci->ci_biglock_wanted != l); 433 } 434 435 /* Not running. It may be safe to block now. */ 436 return false; 437 } 438 #endif /* MULTIPROCESSOR */ 439 440 /* 441 * mutex_vector_enter: 442 * 443 * Support routine for mutex_enter() that must handle all cases. In 444 * the LOCKDEBUG case, mutex_enter() is always aliased here, even if 445 * fast-path stubs are available. If a mutex_spin_enter() stub is 446 * not available, then it is also aliased directly here. 447 */ 448 void 449 mutex_vector_enter(kmutex_t *mtx) 450 { 451 uintptr_t owner, curthread; 452 turnstile_t *ts; 453 #ifdef MULTIPROCESSOR 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 KPREEMPT_DISABLE(curlwp); 466 owner = mtx->mtx_owner; 467 if (MUTEX_SPIN_P(owner)) { 468 #if defined(LOCKDEBUG) && defined(MULTIPROCESSOR) 469 u_int spins = 0; 470 #endif 471 KPREEMPT_ENABLE(curlwp); 472 MUTEX_SPIN_SPLRAISE(mtx); 473 MUTEX_WANTLOCK(mtx); 474 #ifdef FULL 475 if (MUTEX_SPINBIT_LOCK_TRY(mtx)) { 476 MUTEX_LOCKED(mtx); 477 return; 478 } 479 #if !defined(MULTIPROCESSOR) 480 MUTEX_ABORT(mtx, "locking against myself"); 481 #else /* !MULTIPROCESSOR */ 482 483 LOCKSTAT_ENTER(lsflag); 484 LOCKSTAT_START_TIMER(lsflag, spintime); 485 count = SPINLOCK_BACKOFF_MIN; 486 487 /* 488 * Spin testing the lock word and do exponential backoff 489 * to reduce cache line ping-ponging between CPUs. 490 */ 491 do { 492 while (MUTEX_SPINBIT_LOCKED_P(mtx)) { 493 SPINLOCK_SPIN_HOOK; 494 SPINLOCK_BACKOFF(count); 495 #ifdef LOCKDEBUG 496 if (SPINLOCK_SPINOUT(spins)) 497 MUTEX_ABORT(mtx, "spinout"); 498 #endif /* LOCKDEBUG */ 499 } 500 } while (!MUTEX_SPINBIT_LOCK_TRY(mtx)); 501 502 if (count != SPINLOCK_BACKOFF_MIN) { 503 LOCKSTAT_STOP_TIMER(lsflag, spintime); 504 LOCKSTAT_EVENT(lsflag, mtx, 505 LB_SPIN_MUTEX | LB_SPIN, 1, spintime); 506 } 507 LOCKSTAT_EXIT(lsflag); 508 #endif /* !MULTIPROCESSOR */ 509 #endif /* FULL */ 510 MUTEX_LOCKED(mtx); 511 return; 512 } 513 514 curthread = (uintptr_t)curlwp; 515 516 MUTEX_DASSERT(mtx, MUTEX_ADAPTIVE_P(owner)); 517 MUTEX_ASSERT(mtx, curthread != 0); 518 MUTEX_ASSERT(mtx, !cpu_intr_p()); 519 MUTEX_WANTLOCK(mtx); 520 521 if (panicstr == NULL) { 522 KDASSERT(pserialize_not_in_read_section()); 523 LOCKDEBUG_BARRIER(&kernel_lock, 1); 524 } 525 526 LOCKSTAT_ENTER(lsflag); 527 528 /* 529 * Adaptive mutex; spin trying to acquire the mutex. If we 530 * determine that the owner is not running on a processor, 531 * then we stop spinning, and sleep instead. 532 */ 533 for (;;) { 534 if (!MUTEX_OWNED(owner)) { 535 /* 536 * Mutex owner clear could mean two things: 537 * 538 * * The mutex has been released. 539 * * The owner field hasn't been set yet. 540 * 541 * Try to acquire it again. If that fails, 542 * we'll just loop again. 543 */ 544 if (MUTEX_ACQUIRE(mtx, curthread)) 545 break; 546 owner = mtx->mtx_owner; 547 continue; 548 } 549 if (__predict_false(MUTEX_OWNER(owner) == curthread)) { 550 MUTEX_ABORT(mtx, "locking against myself"); 551 } 552 #ifdef MULTIPROCESSOR 553 /* 554 * Check to see if the owner is running on a processor. 555 * If so, then we should just spin, as the owner will 556 * likely release the lock very soon. 557 */ 558 if (mutex_oncpu(owner)) { 559 LOCKSTAT_START_TIMER(lsflag, spintime); 560 count = SPINLOCK_BACKOFF_MIN; 561 do { 562 KPREEMPT_ENABLE(curlwp); 563 SPINLOCK_BACKOFF(count); 564 KPREEMPT_DISABLE(curlwp); 565 owner = mtx->mtx_owner; 566 } while (mutex_oncpu(owner)); 567 LOCKSTAT_STOP_TIMER(lsflag, spintime); 568 LOCKSTAT_COUNT(spincnt, 1); 569 if (!MUTEX_OWNED(owner)) 570 continue; 571 } 572 #endif 573 574 ts = turnstile_lookup(mtx); 575 576 /* 577 * Once we have the turnstile chain interlock, mark the 578 * mutex as having waiters. If that fails, spin again: 579 * chances are that the mutex has been released. 580 */ 581 if (!MUTEX_SET_WAITERS(mtx, owner)) { 582 turnstile_exit(mtx); 583 owner = mtx->mtx_owner; 584 continue; 585 } 586 587 #ifdef MULTIPROCESSOR 588 /* 589 * mutex_exit() is permitted to release the mutex without 590 * any interlocking instructions, and the following can 591 * occur as a result: 592 * 593 * CPU 1: MUTEX_SET_WAITERS() CPU2: mutex_exit() 594 * ---------------------------- ---------------------------- 595 * .. acquire cache line 596 * .. test for waiters 597 * acquire cache line <- lose cache line 598 * lock cache line .. 599 * verify mutex is held .. 600 * set waiters .. 601 * unlock cache line .. 602 * lose cache line -> acquire cache line 603 * .. clear lock word, waiters 604 * return success 605 * 606 * There is another race that can occur: a third CPU could 607 * acquire the mutex as soon as it is released. Since 608 * adaptive mutexes are primarily spin mutexes, this is not 609 * something that we need to worry about too much. What we 610 * do need to ensure is that the waiters bit gets set. 611 * 612 * To allow the unlocked release, we need to make some 613 * assumptions here: 614 * 615 * o Release is the only non-atomic/unlocked operation 616 * that can be performed on the mutex. (It must still 617 * be atomic on the local CPU, e.g. in case interrupted 618 * or preempted). 619 * 620 * o At any given time, MUTEX_SET_WAITERS() can only ever 621 * be in progress on one CPU in the system - guaranteed 622 * by the turnstile chain lock. 623 * 624 * o No other operations other than MUTEX_SET_WAITERS() 625 * and release can modify a mutex with a non-zero 626 * owner field. 627 * 628 * o The result of a successful MUTEX_SET_WAITERS() call 629 * is an unbuffered write that is immediately visible 630 * to all other processors in the system. 631 * 632 * o If the holding LWP switches away, it posts a store 633 * fence before changing curlwp, ensuring that any 634 * overwrite of the mutex waiters flag by mutex_exit() 635 * completes before the modification of curlwp becomes 636 * visible to this CPU. 637 * 638 * o cpu_switchto() posts a store fence after setting curlwp 639 * and before resuming execution of an LWP. 640 * 641 * o _kernel_lock() posts a store fence before setting 642 * curcpu()->ci_biglock_wanted, and after clearing it. 643 * This ensures that any overwrite of the mutex waiters 644 * flag by mutex_exit() completes before the modification 645 * of ci_biglock_wanted becomes visible. 646 * 647 * We now post a read memory barrier (after setting the 648 * waiters field) and check the lock holder's status again. 649 * Some of the possible outcomes (not an exhaustive list): 650 * 651 * 1. The on-CPU check returns true: the holding LWP is 652 * running again. The lock may be released soon and 653 * we should spin. Importantly, we can't trust the 654 * value of the waiters flag. 655 * 656 * 2. The on-CPU check returns false: the holding LWP is 657 * not running. We now have the opportunity to check 658 * if mutex_exit() has blatted the modifications made 659 * by MUTEX_SET_WAITERS(). 660 * 661 * 3. The on-CPU check returns false: the holding LWP may 662 * or may not be running. It has context switched at 663 * some point during our check. Again, we have the 664 * chance to see if the waiters bit is still set or 665 * has been overwritten. 666 * 667 * 4. The on-CPU check returns false: the holding LWP is 668 * running on a CPU, but wants the big lock. It's OK 669 * to check the waiters field in this case. 670 * 671 * 5. The has-waiters check fails: the mutex has been 672 * released, the waiters flag cleared and another LWP 673 * now owns the mutex. 674 * 675 * 6. The has-waiters check fails: the mutex has been 676 * released. 677 * 678 * If the waiters bit is not set it's unsafe to go asleep, 679 * as we might never be awoken. 680 */ 681 membar_consumer(); 682 if (mutex_oncpu(owner)) { 683 turnstile_exit(mtx); 684 owner = mtx->mtx_owner; 685 continue; 686 } 687 membar_consumer(); 688 if (!MUTEX_HAS_WAITERS(mtx)) { 689 turnstile_exit(mtx); 690 owner = mtx->mtx_owner; 691 continue; 692 } 693 #endif /* MULTIPROCESSOR */ 694 695 LOCKSTAT_START_TIMER(lsflag, slptime); 696 697 turnstile_block(ts, TS_WRITER_Q, mtx, &mutex_syncobj); 698 699 LOCKSTAT_STOP_TIMER(lsflag, slptime); 700 LOCKSTAT_COUNT(slpcnt, 1); 701 702 owner = mtx->mtx_owner; 703 } 704 KPREEMPT_ENABLE(curlwp); 705 706 LOCKSTAT_EVENT(lsflag, mtx, LB_ADAPTIVE_MUTEX | LB_SLEEP1, 707 slpcnt, slptime); 708 LOCKSTAT_EVENT(lsflag, mtx, LB_ADAPTIVE_MUTEX | LB_SPIN, 709 spincnt, spintime); 710 LOCKSTAT_EXIT(lsflag); 711 712 MUTEX_DASSERT(mtx, MUTEX_OWNER(mtx->mtx_owner) == curthread); 713 MUTEX_LOCKED(mtx); 714 } 715 716 /* 717 * mutex_vector_exit: 718 * 719 * Support routine for mutex_exit() that handles all cases. 720 */ 721 void 722 mutex_vector_exit(kmutex_t *mtx) 723 { 724 turnstile_t *ts; 725 uintptr_t curthread; 726 727 if (MUTEX_SPIN_P(mtx->mtx_owner)) { 728 #ifdef FULL 729 if (__predict_false(!MUTEX_SPINBIT_LOCKED_P(mtx))) { 730 MUTEX_ABORT(mtx, "exiting unheld spin mutex"); 731 } 732 MUTEX_UNLOCKED(mtx); 733 MUTEX_SPINBIT_LOCK_UNLOCK(mtx); 734 #endif 735 MUTEX_SPIN_SPLRESTORE(mtx); 736 return; 737 } 738 739 #ifndef __HAVE_MUTEX_STUBS 740 /* 741 * On some architectures without mutex stubs, we can enter here to 742 * release mutexes before interrupts and whatnot are up and running. 743 * We need this hack to keep them sweet. 744 */ 745 if (__predict_false(cold)) { 746 MUTEX_UNLOCKED(mtx); 747 MUTEX_RELEASE(mtx); 748 return; 749 } 750 #endif 751 752 curthread = (uintptr_t)curlwp; 753 MUTEX_DASSERT(mtx, curthread != 0); 754 MUTEX_ASSERT(mtx, MUTEX_OWNER(mtx->mtx_owner) == curthread); 755 MUTEX_UNLOCKED(mtx); 756 #if !defined(LOCKDEBUG) 757 __USE(curthread); 758 #endif 759 760 #ifdef LOCKDEBUG 761 /* 762 * Avoid having to take the turnstile chain lock every time 763 * around. Raise the priority level to splhigh() in order 764 * to disable preemption and so make the following atomic. 765 */ 766 { 767 int s = splhigh(); 768 if (!MUTEX_HAS_WAITERS(mtx)) { 769 MUTEX_RELEASE(mtx); 770 splx(s); 771 return; 772 } 773 splx(s); 774 } 775 #endif 776 777 /* 778 * Get this lock's turnstile. This gets the interlock on 779 * the sleep queue. Once we have that, we can clear the 780 * lock. If there was no turnstile for the lock, there 781 * were no waiters remaining. 782 */ 783 ts = turnstile_lookup(mtx); 784 785 if (ts == NULL) { 786 MUTEX_RELEASE(mtx); 787 turnstile_exit(mtx); 788 } else { 789 MUTEX_RELEASE(mtx); 790 turnstile_wakeup(ts, TS_WRITER_Q, 791 TS_WAITERS(ts, TS_WRITER_Q), NULL); 792 } 793 } 794 795 #ifndef __HAVE_SIMPLE_MUTEXES 796 /* 797 * mutex_wakeup: 798 * 799 * Support routine for mutex_exit() that wakes up all waiters. 800 * We assume that the mutex has been released, but it need not 801 * be. 802 */ 803 void 804 mutex_wakeup(kmutex_t *mtx) 805 { 806 turnstile_t *ts; 807 808 ts = turnstile_lookup(mtx); 809 if (ts == NULL) { 810 turnstile_exit(mtx); 811 return; 812 } 813 MUTEX_CLEAR_WAITERS(mtx); 814 turnstile_wakeup(ts, TS_WRITER_Q, TS_WAITERS(ts, TS_WRITER_Q), NULL); 815 } 816 #endif /* !__HAVE_SIMPLE_MUTEXES */ 817 818 /* 819 * mutex_owned: 820 * 821 * Return true if the current LWP (adaptive) or CPU (spin) 822 * holds the mutex. 823 */ 824 int 825 mutex_owned(const kmutex_t *mtx) 826 { 827 828 if (mtx == NULL) 829 return 0; 830 if (MUTEX_ADAPTIVE_P(mtx->mtx_owner)) 831 return MUTEX_OWNER(mtx->mtx_owner) == (uintptr_t)curlwp; 832 #ifdef FULL 833 return MUTEX_SPINBIT_LOCKED_P(mtx); 834 #else 835 return 1; 836 #endif 837 } 838 839 /* 840 * mutex_owner: 841 * 842 * Return the current owner of an adaptive mutex. Used for 843 * priority inheritance. 844 */ 845 lwp_t * 846 mutex_owner(const kmutex_t *mtx) 847 { 848 849 MUTEX_ASSERT(mtx, MUTEX_ADAPTIVE_P(mtx->mtx_owner)); 850 return (struct lwp *)MUTEX_OWNER(mtx->mtx_owner); 851 } 852 853 /* 854 * mutex_owner_running: 855 * 856 * Return true if an adaptive mutex is unheld, or held and the owner is 857 * running on a CPU. For the pagedaemon only - do not document or use 858 * in other code. 859 */ 860 bool 861 mutex_owner_running(const kmutex_t *mtx) 862 { 863 #ifdef MULTIPROCESSOR 864 uintptr_t owner; 865 bool rv; 866 867 MUTEX_ASSERT(mtx, MUTEX_ADAPTIVE_P(mtx->mtx_owner)); 868 kpreempt_disable(); 869 owner = mtx->mtx_owner; 870 rv = !MUTEX_OWNED(owner) || mutex_oncpu(MUTEX_OWNER(owner)); 871 kpreempt_enable(); 872 return rv; 873 #else 874 return mutex_owner(mtx) == curlwp; 875 #endif 876 } 877 878 /* 879 * mutex_ownable: 880 * 881 * When compiled with DEBUG and LOCKDEBUG defined, ensure that 882 * the mutex is available. We cannot use !mutex_owned() since 883 * that won't work correctly for spin mutexes. 884 */ 885 int 886 mutex_ownable(const kmutex_t *mtx) 887 { 888 889 #ifdef LOCKDEBUG 890 MUTEX_TESTLOCK(mtx); 891 #endif 892 return 1; 893 } 894 895 /* 896 * mutex_tryenter: 897 * 898 * Try to acquire the mutex; return non-zero if we did. 899 */ 900 int 901 mutex_tryenter(kmutex_t *mtx) 902 { 903 uintptr_t curthread; 904 905 /* 906 * Handle spin mutexes. 907 */ 908 if (MUTEX_SPIN_P(mtx->mtx_owner)) { 909 MUTEX_SPIN_SPLRAISE(mtx); 910 #ifdef FULL 911 if (MUTEX_SPINBIT_LOCK_TRY(mtx)) { 912 MUTEX_WANTLOCK(mtx); 913 MUTEX_LOCKED(mtx); 914 return 1; 915 } 916 MUTEX_SPIN_SPLRESTORE(mtx); 917 #else 918 MUTEX_WANTLOCK(mtx); 919 MUTEX_LOCKED(mtx); 920 return 1; 921 #endif 922 } else { 923 curthread = (uintptr_t)curlwp; 924 MUTEX_ASSERT(mtx, curthread != 0); 925 if (MUTEX_ACQUIRE(mtx, curthread)) { 926 MUTEX_WANTLOCK(mtx); 927 MUTEX_LOCKED(mtx); 928 MUTEX_DASSERT(mtx, 929 MUTEX_OWNER(mtx->mtx_owner) == curthread); 930 return 1; 931 } 932 } 933 934 return 0; 935 } 936 937 #if defined(__HAVE_SPIN_MUTEX_STUBS) || defined(FULL) 938 /* 939 * mutex_spin_retry: 940 * 941 * Support routine for mutex_spin_enter(). Assumes that the caller 942 * has already raised the SPL, and adjusted counters. 943 */ 944 void 945 mutex_spin_retry(kmutex_t *mtx) 946 { 947 #ifdef MULTIPROCESSOR 948 u_int count; 949 LOCKSTAT_TIMER(spintime); 950 LOCKSTAT_FLAG(lsflag); 951 #ifdef LOCKDEBUG 952 u_int spins = 0; 953 #endif /* LOCKDEBUG */ 954 955 MUTEX_WANTLOCK(mtx); 956 957 LOCKSTAT_ENTER(lsflag); 958 LOCKSTAT_START_TIMER(lsflag, spintime); 959 count = SPINLOCK_BACKOFF_MIN; 960 961 /* 962 * Spin testing the lock word and do exponential backoff 963 * to reduce cache line ping-ponging between CPUs. 964 */ 965 do { 966 while (MUTEX_SPINBIT_LOCKED_P(mtx)) { 967 SPINLOCK_BACKOFF(count); 968 #ifdef LOCKDEBUG 969 if (SPINLOCK_SPINOUT(spins)) 970 MUTEX_ABORT(mtx, "spinout"); 971 #endif /* LOCKDEBUG */ 972 } 973 } while (!MUTEX_SPINBIT_LOCK_TRY(mtx)); 974 975 LOCKSTAT_STOP_TIMER(lsflag, spintime); 976 LOCKSTAT_EVENT(lsflag, mtx, LB_SPIN_MUTEX | LB_SPIN, 1, spintime); 977 LOCKSTAT_EXIT(lsflag); 978 979 MUTEX_LOCKED(mtx); 980 #else /* MULTIPROCESSOR */ 981 MUTEX_ABORT(mtx, "locking against myself"); 982 #endif /* MULTIPROCESSOR */ 983 } 984 #endif /* defined(__HAVE_SPIN_MUTEX_STUBS) || defined(FULL) */ 985