1 /* $NetBSD: kern_rwlock.c,v 1.37 2011/03/20 23:19:16 rmind Exp $ */ 2 3 /*- 4 * Copyright (c) 2002, 2006, 2007, 2008, 2009 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 reader/writer lock implementation, modeled after those 34 * found in Solaris, a description of which can be found in: 35 * 36 * Solaris Internals: Core Kernel Architecture, Jim Mauro and 37 * Richard McDougall. 38 */ 39 40 #include <sys/cdefs.h> 41 __KERNEL_RCSID(0, "$NetBSD: kern_rwlock.c,v 1.37 2011/03/20 23:19:16 rmind Exp $"); 42 43 #define __RWLOCK_PRIVATE 44 45 #include <sys/param.h> 46 #include <sys/proc.h> 47 #include <sys/rwlock.h> 48 #include <sys/sched.h> 49 #include <sys/sleepq.h> 50 #include <sys/systm.h> 51 #include <sys/lockdebug.h> 52 #include <sys/cpu.h> 53 #include <sys/atomic.h> 54 #include <sys/lock.h> 55 56 #include <dev/lockstat.h> 57 58 /* 59 * LOCKDEBUG 60 */ 61 62 #if defined(LOCKDEBUG) 63 64 #define RW_WANTLOCK(rw, op, t) \ 65 LOCKDEBUG_WANTLOCK(RW_DEBUG_P(rw), (rw), \ 66 (uintptr_t)__builtin_return_address(0), op == RW_READER, t); 67 #define RW_LOCKED(rw, op) \ 68 LOCKDEBUG_LOCKED(RW_DEBUG_P(rw), (rw), NULL, \ 69 (uintptr_t)__builtin_return_address(0), op == RW_READER); 70 #define RW_UNLOCKED(rw, op) \ 71 LOCKDEBUG_UNLOCKED(RW_DEBUG_P(rw), (rw), \ 72 (uintptr_t)__builtin_return_address(0), op == RW_READER); 73 #define RW_DASSERT(rw, cond) \ 74 do { \ 75 if (!(cond)) \ 76 rw_abort(rw, __func__, "assertion failed: " #cond); \ 77 } while (/* CONSTCOND */ 0); 78 79 #else /* LOCKDEBUG */ 80 81 #define RW_WANTLOCK(rw, op, t) /* nothing */ 82 #define RW_LOCKED(rw, op) /* nothing */ 83 #define RW_UNLOCKED(rw, op) /* nothing */ 84 #define RW_DASSERT(rw, cond) /* nothing */ 85 86 #endif /* LOCKDEBUG */ 87 88 /* 89 * DIAGNOSTIC 90 */ 91 92 #if defined(DIAGNOSTIC) 93 94 #define RW_ASSERT(rw, cond) \ 95 do { \ 96 if (!(cond)) \ 97 rw_abort(rw, __func__, "assertion failed: " #cond); \ 98 } while (/* CONSTCOND */ 0) 99 100 #else 101 102 #define RW_ASSERT(rw, cond) /* nothing */ 103 104 #endif /* DIAGNOSTIC */ 105 106 #define RW_SETDEBUG(rw, on) ((rw)->rw_owner |= (on) ? 0 : RW_NODEBUG) 107 #define RW_DEBUG_P(rw) (((rw)->rw_owner & RW_NODEBUG) == 0) 108 #if defined(LOCKDEBUG) 109 #define RW_INHERITDEBUG(new, old) (new) |= (old) & RW_NODEBUG 110 #else /* defined(LOCKDEBUG) */ 111 #define RW_INHERITDEBUG(new, old) /* nothing */ 112 #endif /* defined(LOCKDEBUG) */ 113 114 static void rw_abort(krwlock_t *, const char *, const char *); 115 static void rw_dump(volatile void *); 116 static lwp_t *rw_owner(wchan_t); 117 118 static inline uintptr_t 119 rw_cas(krwlock_t *rw, uintptr_t o, uintptr_t n) 120 { 121 122 RW_INHERITDEBUG(n, o); 123 return (uintptr_t)atomic_cas_ptr((volatile void *)&rw->rw_owner, 124 (void *)o, (void *)n); 125 } 126 127 static inline void 128 rw_swap(krwlock_t *rw, uintptr_t o, uintptr_t n) 129 { 130 131 RW_INHERITDEBUG(n, o); 132 n = (uintptr_t)atomic_swap_ptr((volatile void *)&rw->rw_owner, 133 (void *)n); 134 RW_DASSERT(rw, n == o); 135 } 136 137 /* 138 * For platforms that do not provide stubs, or for the LOCKDEBUG case. 139 */ 140 #ifdef LOCKDEBUG 141 #undef __HAVE_RW_STUBS 142 #endif 143 144 #ifndef __HAVE_RW_STUBS 145 __strong_alias(rw_enter,rw_vector_enter); 146 __strong_alias(rw_exit,rw_vector_exit); 147 __strong_alias(rw_tryenter,rw_vector_tryenter); 148 #endif 149 150 lockops_t rwlock_lockops = { 151 "Reader / writer lock", 152 LOCKOPS_SLEEP, 153 rw_dump 154 }; 155 156 syncobj_t rw_syncobj = { 157 SOBJ_SLEEPQ_SORTED, 158 turnstile_unsleep, 159 turnstile_changepri, 160 sleepq_lendpri, 161 rw_owner, 162 }; 163 164 /* 165 * rw_dump: 166 * 167 * Dump the contents of a rwlock structure. 168 */ 169 static void 170 rw_dump(volatile void *cookie) 171 { 172 volatile krwlock_t *rw = cookie; 173 174 printf_nolog("owner/count : %#018lx flags : %#018x\n", 175 (long)RW_OWNER(rw), (int)RW_FLAGS(rw)); 176 } 177 178 /* 179 * rw_abort: 180 * 181 * Dump information about an error and panic the system. This 182 * generates a lot of machine code in the DIAGNOSTIC case, so 183 * we ask the compiler to not inline it. 184 */ 185 static void __noinline 186 rw_abort(krwlock_t *rw, const char *func, const char *msg) 187 { 188 189 if (panicstr != NULL) 190 return; 191 192 LOCKDEBUG_ABORT(rw, &rwlock_lockops, func, msg); 193 } 194 195 /* 196 * rw_init: 197 * 198 * Initialize a rwlock for use. 199 */ 200 void 201 rw_init(krwlock_t *rw) 202 { 203 bool dodebug; 204 205 memset(rw, 0, sizeof(*rw)); 206 207 dodebug = LOCKDEBUG_ALLOC(rw, &rwlock_lockops, 208 (uintptr_t)__builtin_return_address(0)); 209 RW_SETDEBUG(rw, dodebug); 210 } 211 212 /* 213 * rw_destroy: 214 * 215 * Tear down a rwlock. 216 */ 217 void 218 rw_destroy(krwlock_t *rw) 219 { 220 221 RW_ASSERT(rw, (rw->rw_owner & ~RW_NODEBUG) == 0); 222 LOCKDEBUG_FREE(RW_DEBUG_P(rw), rw); 223 } 224 225 /* 226 * rw_oncpu: 227 * 228 * Return true if an rwlock owner is running on a CPU in the system. 229 * If the target is waiting on the kernel big lock, then we must 230 * release it. This is necessary to avoid deadlock. 231 */ 232 static bool 233 rw_oncpu(uintptr_t owner) 234 { 235 #ifdef MULTIPROCESSOR 236 struct cpu_info *ci; 237 lwp_t *l; 238 239 KASSERT(kpreempt_disabled()); 240 241 if ((owner & (RW_WRITE_LOCKED|RW_HAS_WAITERS)) != RW_WRITE_LOCKED) { 242 return false; 243 } 244 245 /* 246 * See lwp_dtor() why dereference of the LWP pointer is safe. 247 * We must have kernel preemption disabled for that. 248 */ 249 l = (lwp_t *)(owner & RW_THREAD); 250 ci = l->l_cpu; 251 252 if (ci && ci->ci_curlwp == l) { 253 /* Target is running; do we need to block? */ 254 return (ci->ci_biglock_wanted != l); 255 } 256 #endif 257 /* Not running. It may be safe to block now. */ 258 return false; 259 } 260 261 /* 262 * rw_vector_enter: 263 * 264 * Acquire a rwlock. 265 */ 266 void 267 rw_vector_enter(krwlock_t *rw, const krw_t op) 268 { 269 uintptr_t owner, incr, need_wait, set_wait, curthread, next; 270 turnstile_t *ts; 271 int queue; 272 lwp_t *l; 273 LOCKSTAT_TIMER(slptime); 274 LOCKSTAT_TIMER(slpcnt); 275 LOCKSTAT_TIMER(spintime); 276 LOCKSTAT_COUNTER(spincnt); 277 LOCKSTAT_FLAG(lsflag); 278 279 l = curlwp; 280 curthread = (uintptr_t)l; 281 282 RW_ASSERT(rw, !cpu_intr_p()); 283 RW_ASSERT(rw, curthread != 0); 284 RW_WANTLOCK(rw, op, false); 285 286 if (panicstr == NULL) { 287 LOCKDEBUG_BARRIER(&kernel_lock, 1); 288 } 289 290 /* 291 * We play a slight trick here. If we're a reader, we want 292 * increment the read count. If we're a writer, we want to 293 * set the owner field and whe WRITE_LOCKED bit. 294 * 295 * In the latter case, we expect those bits to be zero, 296 * therefore we can use an add operation to set them, which 297 * means an add operation for both cases. 298 */ 299 if (__predict_true(op == RW_READER)) { 300 incr = RW_READ_INCR; 301 set_wait = RW_HAS_WAITERS; 302 need_wait = RW_WRITE_LOCKED | RW_WRITE_WANTED; 303 queue = TS_READER_Q; 304 } else { 305 RW_DASSERT(rw, op == RW_WRITER); 306 incr = curthread | RW_WRITE_LOCKED; 307 set_wait = RW_HAS_WAITERS | RW_WRITE_WANTED; 308 need_wait = RW_WRITE_LOCKED | RW_THREAD; 309 queue = TS_WRITER_Q; 310 } 311 312 LOCKSTAT_ENTER(lsflag); 313 314 KPREEMPT_DISABLE(curlwp); 315 for (owner = rw->rw_owner; ;) { 316 /* 317 * Read the lock owner field. If the need-to-wait 318 * indicator is clear, then try to acquire the lock. 319 */ 320 if ((owner & need_wait) == 0) { 321 next = rw_cas(rw, owner, (owner + incr) & 322 ~RW_WRITE_WANTED); 323 if (__predict_true(next == owner)) { 324 /* Got it! */ 325 membar_enter(); 326 break; 327 } 328 329 /* 330 * Didn't get it -- spin around again (we'll 331 * probably sleep on the next iteration). 332 */ 333 owner = next; 334 continue; 335 } 336 if (__predict_false(panicstr != NULL)) { 337 kpreempt_enable(); 338 return; 339 } 340 if (__predict_false(RW_OWNER(rw) == curthread)) { 341 rw_abort(rw, __func__, "locking against myself"); 342 } 343 /* 344 * If the lock owner is running on another CPU, and 345 * there are no existing waiters, then spin. 346 */ 347 if (rw_oncpu(owner)) { 348 LOCKSTAT_START_TIMER(lsflag, spintime); 349 u_int count = SPINLOCK_BACKOFF_MIN; 350 do { 351 kpreempt_enable(); 352 SPINLOCK_BACKOFF(count); 353 kpreempt_disable(); 354 owner = rw->rw_owner; 355 } while (rw_oncpu(owner)); 356 LOCKSTAT_STOP_TIMER(lsflag, spintime); 357 LOCKSTAT_COUNT(spincnt, 1); 358 if ((owner & need_wait) == 0) 359 continue; 360 } 361 362 /* 363 * Grab the turnstile chain lock. Once we have that, we 364 * can adjust the waiter bits and sleep queue. 365 */ 366 ts = turnstile_lookup(rw); 367 368 /* 369 * Mark the rwlock as having waiters. If the set fails, 370 * then we may not need to sleep and should spin again. 371 * Reload rw_owner because turnstile_lookup() may have 372 * spun on the turnstile chain lock. 373 */ 374 owner = rw->rw_owner; 375 if ((owner & need_wait) == 0 || rw_oncpu(owner)) { 376 turnstile_exit(rw); 377 continue; 378 } 379 next = rw_cas(rw, owner, owner | set_wait); 380 if (__predict_false(next != owner)) { 381 turnstile_exit(rw); 382 owner = next; 383 continue; 384 } 385 386 LOCKSTAT_START_TIMER(lsflag, slptime); 387 turnstile_block(ts, queue, rw, &rw_syncobj); 388 LOCKSTAT_STOP_TIMER(lsflag, slptime); 389 LOCKSTAT_COUNT(slpcnt, 1); 390 391 /* 392 * No need for a memory barrier because of context switch. 393 * If not handed the lock, then spin again. 394 */ 395 if (op == RW_READER || (rw->rw_owner & RW_THREAD) == curthread) 396 break; 397 } 398 KPREEMPT_ENABLE(curlwp); 399 400 LOCKSTAT_EVENT(lsflag, rw, LB_RWLOCK | 401 (op == RW_WRITER ? LB_SLEEP1 : LB_SLEEP2), slpcnt, slptime); 402 LOCKSTAT_EVENT(lsflag, rw, LB_RWLOCK | LB_SPIN, spincnt, spintime); 403 LOCKSTAT_EXIT(lsflag); 404 405 RW_DASSERT(rw, (op != RW_READER && RW_OWNER(rw) == curthread) || 406 (op == RW_READER && RW_COUNT(rw) != 0)); 407 RW_LOCKED(rw, op); 408 } 409 410 /* 411 * rw_vector_exit: 412 * 413 * Release a rwlock. 414 */ 415 void 416 rw_vector_exit(krwlock_t *rw) 417 { 418 uintptr_t curthread, owner, decr, new, next; 419 turnstile_t *ts; 420 int rcnt, wcnt; 421 lwp_t *l; 422 423 curthread = (uintptr_t)curlwp; 424 RW_ASSERT(rw, curthread != 0); 425 426 if (__predict_false(panicstr != NULL)) 427 return; 428 429 /* 430 * Again, we use a trick. Since we used an add operation to 431 * set the required lock bits, we can use a subtract to clear 432 * them, which makes the read-release and write-release path 433 * the same. 434 */ 435 owner = rw->rw_owner; 436 if (__predict_false((owner & RW_WRITE_LOCKED) != 0)) { 437 RW_UNLOCKED(rw, RW_WRITER); 438 RW_ASSERT(rw, RW_OWNER(rw) == curthread); 439 decr = curthread | RW_WRITE_LOCKED; 440 } else { 441 RW_UNLOCKED(rw, RW_READER); 442 RW_ASSERT(rw, RW_COUNT(rw) != 0); 443 decr = RW_READ_INCR; 444 } 445 446 /* 447 * Compute what we expect the new value of the lock to be. Only 448 * proceed to do direct handoff if there are waiters, and if the 449 * lock would become unowned. 450 */ 451 membar_exit(); 452 for (;;) { 453 new = (owner - decr); 454 if ((new & (RW_THREAD | RW_HAS_WAITERS)) == RW_HAS_WAITERS) 455 break; 456 next = rw_cas(rw, owner, new); 457 if (__predict_true(next == owner)) 458 return; 459 owner = next; 460 } 461 462 /* 463 * Grab the turnstile chain lock. This gets the interlock 464 * on the sleep queue. Once we have that, we can adjust the 465 * waiter bits. 466 */ 467 ts = turnstile_lookup(rw); 468 owner = rw->rw_owner; 469 RW_DASSERT(rw, ts != NULL); 470 RW_DASSERT(rw, (owner & RW_HAS_WAITERS) != 0); 471 472 wcnt = TS_WAITERS(ts, TS_WRITER_Q); 473 rcnt = TS_WAITERS(ts, TS_READER_Q); 474 475 /* 476 * Give the lock away. 477 * 478 * If we are releasing a write lock, then prefer to wake all 479 * outstanding readers. Otherwise, wake one writer if there 480 * are outstanding readers, or all writers if there are no 481 * pending readers. If waking one specific writer, the writer 482 * is handed the lock here. If waking multiple writers, we 483 * set WRITE_WANTED to block out new readers, and let them 484 * do the work of acquring the lock in rw_vector_enter(). 485 */ 486 if (rcnt == 0 || decr == RW_READ_INCR) { 487 RW_DASSERT(rw, wcnt != 0); 488 RW_DASSERT(rw, (owner & RW_WRITE_WANTED) != 0); 489 490 if (rcnt != 0) { 491 /* Give the lock to the longest waiting writer. */ 492 l = TS_FIRST(ts, TS_WRITER_Q); 493 new = (uintptr_t)l | RW_WRITE_LOCKED | RW_HAS_WAITERS; 494 if (wcnt > 1) 495 new |= RW_WRITE_WANTED; 496 rw_swap(rw, owner, new); 497 turnstile_wakeup(ts, TS_WRITER_Q, 1, l); 498 } else { 499 /* Wake all writers and let them fight it out. */ 500 rw_swap(rw, owner, RW_WRITE_WANTED); 501 turnstile_wakeup(ts, TS_WRITER_Q, wcnt, NULL); 502 } 503 } else { 504 RW_DASSERT(rw, rcnt != 0); 505 506 /* 507 * Give the lock to all blocked readers. If there 508 * is a writer waiting, new readers that arrive 509 * after the release will be blocked out. 510 */ 511 new = rcnt << RW_READ_COUNT_SHIFT; 512 if (wcnt != 0) 513 new |= RW_HAS_WAITERS | RW_WRITE_WANTED; 514 515 /* Wake up all sleeping readers. */ 516 rw_swap(rw, owner, new); 517 turnstile_wakeup(ts, TS_READER_Q, rcnt, NULL); 518 } 519 } 520 521 /* 522 * rw_vector_tryenter: 523 * 524 * Try to acquire a rwlock. 525 */ 526 int 527 rw_vector_tryenter(krwlock_t *rw, const krw_t op) 528 { 529 uintptr_t curthread, owner, incr, need_wait, next; 530 531 curthread = (uintptr_t)curlwp; 532 533 RW_ASSERT(rw, curthread != 0); 534 535 if (op == RW_READER) { 536 incr = RW_READ_INCR; 537 need_wait = RW_WRITE_LOCKED | RW_WRITE_WANTED; 538 } else { 539 RW_DASSERT(rw, op == RW_WRITER); 540 incr = curthread | RW_WRITE_LOCKED; 541 need_wait = RW_WRITE_LOCKED | RW_THREAD; 542 } 543 544 for (owner = rw->rw_owner;; owner = next) { 545 owner = rw->rw_owner; 546 if (__predict_false((owner & need_wait) != 0)) 547 return 0; 548 next = rw_cas(rw, owner, owner + incr); 549 if (__predict_true(next == owner)) { 550 /* Got it! */ 551 membar_enter(); 552 break; 553 } 554 } 555 556 RW_WANTLOCK(rw, op, true); 557 RW_LOCKED(rw, op); 558 RW_DASSERT(rw, (op != RW_READER && RW_OWNER(rw) == curthread) || 559 (op == RW_READER && RW_COUNT(rw) != 0)); 560 561 return 1; 562 } 563 564 /* 565 * rw_downgrade: 566 * 567 * Downgrade a write lock to a read lock. 568 */ 569 void 570 rw_downgrade(krwlock_t *rw) 571 { 572 uintptr_t owner, curthread, new, next; 573 turnstile_t *ts; 574 int rcnt, wcnt; 575 576 curthread = (uintptr_t)curlwp; 577 RW_ASSERT(rw, curthread != 0); 578 RW_DASSERT(rw, (rw->rw_owner & RW_WRITE_LOCKED) != 0); 579 RW_ASSERT(rw, RW_OWNER(rw) == curthread); 580 RW_UNLOCKED(rw, RW_WRITER); 581 582 membar_producer(); 583 owner = rw->rw_owner; 584 if ((owner & RW_HAS_WAITERS) == 0) { 585 /* 586 * There are no waiters, so we can do this the easy way. 587 * Try swapping us down to one read hold. If it fails, the 588 * lock condition has changed and we most likely now have 589 * waiters. 590 */ 591 next = rw_cas(rw, owner, RW_READ_INCR); 592 if (__predict_true(next == owner)) { 593 RW_LOCKED(rw, RW_READER); 594 RW_DASSERT(rw, (rw->rw_owner & RW_WRITE_LOCKED) == 0); 595 RW_DASSERT(rw, RW_COUNT(rw) != 0); 596 return; 597 } 598 owner = next; 599 } 600 601 /* 602 * Grab the turnstile chain lock. This gets the interlock 603 * on the sleep queue. Once we have that, we can adjust the 604 * waiter bits. 605 */ 606 for (;; owner = next) { 607 ts = turnstile_lookup(rw); 608 RW_DASSERT(rw, ts != NULL); 609 610 rcnt = TS_WAITERS(ts, TS_READER_Q); 611 wcnt = TS_WAITERS(ts, TS_WRITER_Q); 612 613 /* 614 * If there are no readers, just preserve the waiters 615 * bits, swap us down to one read hold and return. 616 */ 617 if (rcnt == 0) { 618 RW_DASSERT(rw, wcnt != 0); 619 RW_DASSERT(rw, (rw->rw_owner & RW_WRITE_WANTED) != 0); 620 RW_DASSERT(rw, (rw->rw_owner & RW_HAS_WAITERS) != 0); 621 622 new = RW_READ_INCR | RW_HAS_WAITERS | RW_WRITE_WANTED; 623 next = rw_cas(rw, owner, new); 624 turnstile_exit(rw); 625 if (__predict_true(next == owner)) 626 break; 627 } else { 628 /* 629 * Give the lock to all blocked readers. We may 630 * retain one read hold if downgrading. If there 631 * is a writer waiting, new readers will be blocked 632 * out. 633 */ 634 new = (rcnt << RW_READ_COUNT_SHIFT) + RW_READ_INCR; 635 if (wcnt != 0) 636 new |= RW_HAS_WAITERS | RW_WRITE_WANTED; 637 638 next = rw_cas(rw, owner, new); 639 if (__predict_true(next == owner)) { 640 /* Wake up all sleeping readers. */ 641 turnstile_wakeup(ts, TS_READER_Q, rcnt, NULL); 642 break; 643 } 644 turnstile_exit(rw); 645 } 646 } 647 648 RW_WANTLOCK(rw, RW_READER, false); 649 RW_LOCKED(rw, RW_READER); 650 RW_DASSERT(rw, (rw->rw_owner & RW_WRITE_LOCKED) == 0); 651 RW_DASSERT(rw, RW_COUNT(rw) != 0); 652 } 653 654 /* 655 * rw_tryupgrade: 656 * 657 * Try to upgrade a read lock to a write lock. We must be the 658 * only reader. 659 */ 660 int 661 rw_tryupgrade(krwlock_t *rw) 662 { 663 uintptr_t owner, curthread, new, next; 664 665 curthread = (uintptr_t)curlwp; 666 RW_ASSERT(rw, curthread != 0); 667 RW_ASSERT(rw, rw_read_held(rw)); 668 669 for (owner = rw->rw_owner;; owner = next) { 670 RW_ASSERT(rw, (owner & RW_WRITE_LOCKED) == 0); 671 if (__predict_false((owner & RW_THREAD) != RW_READ_INCR)) { 672 RW_ASSERT(rw, (owner & RW_THREAD) != 0); 673 return 0; 674 } 675 new = curthread | RW_WRITE_LOCKED | (owner & ~RW_THREAD); 676 next = rw_cas(rw, owner, new); 677 if (__predict_true(next == owner)) { 678 membar_producer(); 679 break; 680 } 681 } 682 683 RW_UNLOCKED(rw, RW_READER); 684 RW_WANTLOCK(rw, RW_WRITER, true); 685 RW_LOCKED(rw, RW_WRITER); 686 RW_DASSERT(rw, rw->rw_owner & RW_WRITE_LOCKED); 687 RW_DASSERT(rw, RW_OWNER(rw) == curthread); 688 689 return 1; 690 } 691 692 /* 693 * rw_read_held: 694 * 695 * Returns true if the rwlock is held for reading. Must only be 696 * used for diagnostic assertions, and never be used to make 697 * decisions about how to use a rwlock. 698 */ 699 int 700 rw_read_held(krwlock_t *rw) 701 { 702 uintptr_t owner; 703 704 if (panicstr != NULL) 705 return 1; 706 if (rw == NULL) 707 return 0; 708 owner = rw->rw_owner; 709 return (owner & RW_WRITE_LOCKED) == 0 && (owner & RW_THREAD) != 0; 710 } 711 712 /* 713 * rw_write_held: 714 * 715 * Returns true if the rwlock is held for writing. Must only be 716 * used for diagnostic assertions, and never be used to make 717 * decisions about how to use a rwlock. 718 */ 719 int 720 rw_write_held(krwlock_t *rw) 721 { 722 723 if (panicstr != NULL) 724 return 1; 725 if (rw == NULL) 726 return 0; 727 return (rw->rw_owner & (RW_WRITE_LOCKED | RW_THREAD)) == 728 (RW_WRITE_LOCKED | (uintptr_t)curlwp); 729 } 730 731 /* 732 * rw_lock_held: 733 * 734 * Returns true if the rwlock is held for reading or writing. Must 735 * only be used for diagnostic assertions, and never be used to make 736 * decisions about how to use a rwlock. 737 */ 738 int 739 rw_lock_held(krwlock_t *rw) 740 { 741 742 if (panicstr != NULL) 743 return 1; 744 if (rw == NULL) 745 return 0; 746 return (rw->rw_owner & RW_THREAD) != 0; 747 } 748 749 /* 750 * rw_owner: 751 * 752 * Return the current owner of an RW lock, but only if it is write 753 * held. Used for priority inheritance. 754 */ 755 static lwp_t * 756 rw_owner(wchan_t obj) 757 { 758 krwlock_t *rw = (void *)(uintptr_t)obj; /* discard qualifiers */ 759 uintptr_t owner = rw->rw_owner; 760 761 if ((owner & RW_WRITE_LOCKED) == 0) 762 return NULL; 763 764 return (void *)(owner & RW_THREAD); 765 } 766