kern_rwlock.c revision 177843
1/*- 2 * Copyright (c) 2006 John Baldwin <jhb@FreeBSD.org> 3 * All rights reserved. 4 * 5 * Redistribution and use in source and binary forms, with or without 6 * modification, are permitted provided that the following conditions 7 * are met: 8 * 1. Redistributions of source code must retain the above copyright 9 * notice, this list of conditions and the following disclaimer. 10 * 2. Redistributions in binary form must reproduce the above copyright 11 * notice, this list of conditions and the following disclaimer in the 12 * documentation and/or other materials provided with the distribution. 13 * 3. Neither the name of the author nor the names of any co-contributors 14 * may be used to endorse or promote products derived from this software 15 * without specific prior written permission. 16 * 17 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND 18 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 19 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 20 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE 21 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 22 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 23 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 24 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 25 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 26 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 27 * SUCH DAMAGE. 28 */ 29 30/* 31 * Machine independent bits of reader/writer lock implementation. 32 */ 33 34#include <sys/cdefs.h> 35__FBSDID("$FreeBSD: head/sys/kern/kern_rwlock.c 177843 2008-04-01 20:31:55Z attilio $"); 36 37#include "opt_ddb.h" 38#include "opt_no_adaptive_rwlocks.h" 39 40#include <sys/param.h> 41#include <sys/ktr.h> 42#include <sys/lock.h> 43#include <sys/mutex.h> 44#include <sys/proc.h> 45#include <sys/rwlock.h> 46#include <sys/systm.h> 47#include <sys/turnstile.h> 48 49#include <machine/cpu.h> 50 51CTASSERT((RW_RECURSE & LO_CLASSFLAGS) == RW_RECURSE); 52 53#if defined(SMP) && !defined(NO_ADAPTIVE_RWLOCKS) 54#define ADAPTIVE_RWLOCKS 55#endif 56 57#ifdef DDB 58#include <ddb/ddb.h> 59 60static void db_show_rwlock(struct lock_object *lock); 61#endif 62static void assert_rw(struct lock_object *lock, int what); 63static void lock_rw(struct lock_object *lock, int how); 64static int unlock_rw(struct lock_object *lock); 65 66struct lock_class lock_class_rw = { 67 .lc_name = "rw", 68 .lc_flags = LC_SLEEPLOCK | LC_RECURSABLE | LC_UPGRADABLE, 69 .lc_assert = assert_rw, 70#ifdef DDB 71 .lc_ddb_show = db_show_rwlock, 72#endif 73 .lc_lock = lock_rw, 74 .lc_unlock = unlock_rw, 75}; 76 77/* 78 * Return a pointer to the owning thread if the lock is write-locked or 79 * NULL if the lock is unlocked or read-locked. 80 */ 81#define rw_wowner(rw) \ 82 ((rw)->rw_lock & RW_LOCK_READ ? NULL : \ 83 (struct thread *)RW_OWNER((rw)->rw_lock)) 84 85/* 86 * Returns if a write owner is recursed. Write ownership is not assured 87 * here and should be previously checked. 88 */ 89#define rw_recursed(rw) ((rw)->rw_recurse != 0) 90 91/* 92 * Return true if curthread helds the lock. 93 */ 94#define rw_wlocked(rw) (rw_wowner((rw)) == curthread) 95 96/* 97 * Return a pointer to the owning thread for this lock who should receive 98 * any priority lent by threads that block on this lock. Currently this 99 * is identical to rw_wowner(). 100 */ 101#define rw_owner(rw) rw_wowner(rw) 102 103#ifndef INVARIANTS 104#define _rw_assert(rw, what, file, line) 105#endif 106 107void 108assert_rw(struct lock_object *lock, int what) 109{ 110 111 rw_assert((struct rwlock *)lock, what); 112} 113 114void 115lock_rw(struct lock_object *lock, int how) 116{ 117 struct rwlock *rw; 118 119 rw = (struct rwlock *)lock; 120 if (how) 121 rw_wlock(rw); 122 else 123 rw_rlock(rw); 124} 125 126int 127unlock_rw(struct lock_object *lock) 128{ 129 struct rwlock *rw; 130 131 rw = (struct rwlock *)lock; 132 rw_assert(rw, RA_LOCKED | LA_NOTRECURSED); 133 if (rw->rw_lock & RW_LOCK_READ) { 134 rw_runlock(rw); 135 return (0); 136 } else { 137 rw_wunlock(rw); 138 return (1); 139 } 140} 141 142void 143rw_init_flags(struct rwlock *rw, const char *name, int opts) 144{ 145 int flags; 146 147 MPASS((opts & ~(RW_DUPOK | RW_NOPROFILE | RW_NOWITNESS | RW_QUIET | 148 RW_RECURSE)) == 0); 149 150 flags = LO_UPGRADABLE | LO_RECURSABLE; 151 if (opts & RW_DUPOK) 152 flags |= LO_DUPOK; 153 if (opts & RW_NOPROFILE) 154 flags |= LO_NOPROFILE; 155 if (!(opts & RW_NOWITNESS)) 156 flags |= LO_WITNESS; 157 if (opts & RW_QUIET) 158 flags |= LO_QUIET; 159 flags |= opts & RW_RECURSE; 160 161 rw->rw_lock = RW_UNLOCKED; 162 rw->rw_recurse = 0; 163 lock_init(&rw->lock_object, &lock_class_rw, name, NULL, flags); 164} 165 166void 167rw_destroy(struct rwlock *rw) 168{ 169 170 KASSERT(rw->rw_lock == RW_UNLOCKED, ("rw lock not unlocked")); 171 KASSERT(rw->rw_recurse == 0, ("rw lock still recursed")); 172 rw->rw_lock = RW_DESTROYED; 173 lock_destroy(&rw->lock_object); 174} 175 176void 177rw_sysinit(void *arg) 178{ 179 struct rw_args *args = arg; 180 181 rw_init(args->ra_rw, args->ra_desc); 182} 183 184int 185rw_wowned(struct rwlock *rw) 186{ 187 188 return (rw_wowner(rw) == curthread); 189} 190 191void 192_rw_wlock(struct rwlock *rw, const char *file, int line) 193{ 194 195 MPASS(curthread != NULL); 196 KASSERT(rw->rw_lock != RW_DESTROYED, 197 ("rw_wlock() of destroyed rwlock @ %s:%d", file, line)); 198 WITNESS_CHECKORDER(&rw->lock_object, LOP_NEWORDER | LOP_EXCLUSIVE, file, 199 line); 200 __rw_wlock(rw, curthread, file, line); 201 LOCK_LOG_LOCK("WLOCK", &rw->lock_object, 0, rw->rw_recurse, file, line); 202 WITNESS_LOCK(&rw->lock_object, LOP_EXCLUSIVE, file, line); 203 curthread->td_locks++; 204} 205 206int 207_rw_try_wlock(struct rwlock *rw, const char *file, int line) 208{ 209 int rval; 210 211 KASSERT(rw->rw_lock != RW_DESTROYED, 212 ("rw_try_wlock() of destroyed rwlock @ %s:%d", file, line)); 213 214 if (rw_wlocked(rw) && (rw->lock_object.lo_flags & RW_RECURSE) != 0) { 215 rw->rw_recurse++; 216 rval = 1; 217 } else 218 rval = atomic_cmpset_acq_ptr(&rw->rw_lock, RW_UNLOCKED, 219 (uintptr_t)curthread); 220 221 LOCK_LOG_TRY("WLOCK", &rw->lock_object, 0, rval, file, line); 222 if (rval) { 223 WITNESS_LOCK(&rw->lock_object, LOP_EXCLUSIVE | LOP_TRYLOCK, 224 file, line); 225 curthread->td_locks++; 226 } 227 return (rval); 228} 229 230void 231_rw_wunlock(struct rwlock *rw, const char *file, int line) 232{ 233 234 MPASS(curthread != NULL); 235 KASSERT(rw->rw_lock != RW_DESTROYED, 236 ("rw_wunlock() of destroyed rwlock @ %s:%d", file, line)); 237 _rw_assert(rw, RA_WLOCKED, file, line); 238 curthread->td_locks--; 239 WITNESS_UNLOCK(&rw->lock_object, LOP_EXCLUSIVE, file, line); 240 LOCK_LOG_LOCK("WUNLOCK", &rw->lock_object, 0, rw->rw_recurse, file, 241 line); 242 if (!rw_recursed(rw)) 243 lock_profile_release_lock(&rw->lock_object); 244 __rw_wunlock(rw, curthread, file, line); 245} 246/* 247 * Determines whether a new reader can acquire a lock. Succeeds if the 248 * reader already owns a read lock and the lock is locked for read to 249 * prevent deadlock from reader recursion. Also succeeds if the lock 250 * is unlocked and has no writer waiters or spinners. Failing otherwise 251 * prioritizes writers before readers. 252 */ 253#define RW_CAN_READ(_rw) \ 254 ((curthread->td_rw_rlocks && (_rw) & RW_LOCK_READ) || ((_rw) & \ 255 (RW_LOCK_READ | RW_LOCK_WRITE_WAITERS | RW_LOCK_WRITE_SPINNER)) == \ 256 RW_LOCK_READ) 257 258void 259_rw_rlock(struct rwlock *rw, const char *file, int line) 260{ 261 struct turnstile *ts; 262#ifdef ADAPTIVE_RWLOCKS 263 volatile struct thread *owner; 264#endif 265 uint64_t waittime = 0; 266 int contested = 0; 267 uintptr_t v; 268 269 KASSERT(rw->rw_lock != RW_DESTROYED, 270 ("rw_rlock() of destroyed rwlock @ %s:%d", file, line)); 271 KASSERT(rw_wowner(rw) != curthread, 272 ("%s (%s): wlock already held @ %s:%d", __func__, 273 rw->lock_object.lo_name, file, line)); 274 WITNESS_CHECKORDER(&rw->lock_object, LOP_NEWORDER, file, line); 275 276 for (;;) { 277 /* 278 * Handle the easy case. If no other thread has a write 279 * lock, then try to bump up the count of read locks. Note 280 * that we have to preserve the current state of the 281 * RW_LOCK_WRITE_WAITERS flag. If we fail to acquire a 282 * read lock, then rw_lock must have changed, so restart 283 * the loop. Note that this handles the case of a 284 * completely unlocked rwlock since such a lock is encoded 285 * as a read lock with no waiters. 286 */ 287 v = rw->rw_lock; 288 if (RW_CAN_READ(v)) { 289 /* 290 * The RW_LOCK_READ_WAITERS flag should only be set 291 * if the lock has been unlocked and write waiters 292 * were present. 293 */ 294 if (atomic_cmpset_acq_ptr(&rw->rw_lock, v, 295 v + RW_ONE_READER)) { 296 if (LOCK_LOG_TEST(&rw->lock_object, 0)) 297 CTR4(KTR_LOCK, 298 "%s: %p succeed %p -> %p", __func__, 299 rw, (void *)v, 300 (void *)(v + RW_ONE_READER)); 301 break; 302 } 303 cpu_spinwait(); 304 continue; 305 } 306 lock_profile_obtain_lock_failed(&rw->lock_object, 307 &contested, &waittime); 308 309#ifdef ADAPTIVE_RWLOCKS 310 /* 311 * If the owner is running on another CPU, spin until 312 * the owner stops running or the state of the lock 313 * changes. 314 */ 315 if ((v & RW_LOCK_READ) == 0) { 316 owner = (struct thread *)RW_OWNER(v); 317 if (TD_IS_RUNNING(owner)) { 318 if (LOCK_LOG_TEST(&rw->lock_object, 0)) 319 CTR3(KTR_LOCK, 320 "%s: spinning on %p held by %p", 321 __func__, rw, owner); 322 while ((struct thread*)RW_OWNER(rw->rw_lock) == 323 owner && TD_IS_RUNNING(owner)) 324 cpu_spinwait(); 325 continue; 326 } 327 } 328#endif 329 330 /* 331 * Okay, now it's the hard case. Some other thread already 332 * has a write lock or there are write waiters present, 333 * acquire the turnstile lock so we can begin the process 334 * of blocking. 335 */ 336 ts = turnstile_trywait(&rw->lock_object); 337 338 /* 339 * The lock might have been released while we spun, so 340 * recheck its state and restart the loop if needed. 341 */ 342 v = rw->rw_lock; 343 if (RW_CAN_READ(v)) { 344 turnstile_cancel(ts); 345 cpu_spinwait(); 346 continue; 347 } 348 349#ifdef ADAPTIVE_RWLOCKS 350 /* 351 * If the current owner of the lock is executing on another 352 * CPU quit the hard path and try to spin. 353 */ 354 if ((v & RW_LOCK_READ) == 0) { 355 owner = (struct thread *)RW_OWNER(v); 356 if (TD_IS_RUNNING(owner)) { 357 turnstile_cancel(ts); 358 cpu_spinwait(); 359 continue; 360 } 361 } 362#endif 363 364 /* 365 * The lock is held in write mode or it already has waiters. 366 */ 367 MPASS(!RW_CAN_READ(v)); 368 369 /* 370 * If the RW_LOCK_READ_WAITERS flag is already set, then 371 * we can go ahead and block. If it is not set then try 372 * to set it. If we fail to set it drop the turnstile 373 * lock and restart the loop. 374 */ 375 if (!(v & RW_LOCK_READ_WAITERS)) { 376 if (!atomic_cmpset_ptr(&rw->rw_lock, v, 377 v | RW_LOCK_READ_WAITERS)) { 378 turnstile_cancel(ts); 379 cpu_spinwait(); 380 continue; 381 } 382 if (LOCK_LOG_TEST(&rw->lock_object, 0)) 383 CTR2(KTR_LOCK, "%s: %p set read waiters flag", 384 __func__, rw); 385 } 386 387 /* 388 * We were unable to acquire the lock and the read waiters 389 * flag is set, so we must block on the turnstile. 390 */ 391 if (LOCK_LOG_TEST(&rw->lock_object, 0)) 392 CTR2(KTR_LOCK, "%s: %p blocking on turnstile", __func__, 393 rw); 394 turnstile_wait(ts, rw_owner(rw), TS_SHARED_QUEUE); 395 if (LOCK_LOG_TEST(&rw->lock_object, 0)) 396 CTR2(KTR_LOCK, "%s: %p resuming from turnstile", 397 __func__, rw); 398 } 399 400 /* 401 * TODO: acquire "owner of record" here. Here be turnstile dragons 402 * however. turnstiles don't like owners changing between calls to 403 * turnstile_wait() currently. 404 */ 405 lock_profile_obtain_lock_success( &rw->lock_object, contested, 406 waittime, file, line); 407 LOCK_LOG_LOCK("RLOCK", &rw->lock_object, 0, 0, file, line); 408 WITNESS_LOCK(&rw->lock_object, 0, file, line); 409 curthread->td_locks++; 410 curthread->td_rw_rlocks++; 411} 412 413int 414_rw_try_rlock(struct rwlock *rw, const char *file, int line) 415{ 416 uintptr_t x; 417 418 for (;;) { 419 x = rw->rw_lock; 420 KASSERT(rw->rw_lock != RW_DESTROYED, 421 ("rw_try_rlock() of destroyed rwlock @ %s:%d", file, line)); 422 if (!(x & RW_LOCK_READ)) 423 break; 424 if (atomic_cmpset_acq_ptr(&rw->rw_lock, x, x + RW_ONE_READER)) { 425 LOCK_LOG_TRY("RLOCK", &rw->lock_object, 0, 1, file, 426 line); 427 WITNESS_LOCK(&rw->lock_object, LOP_TRYLOCK, file, line); 428 curthread->td_locks++; 429 curthread->td_rw_rlocks++; 430 return (1); 431 } 432 } 433 434 LOCK_LOG_TRY("RLOCK", &rw->lock_object, 0, 0, file, line); 435 return (0); 436} 437 438void 439_rw_runlock(struct rwlock *rw, const char *file, int line) 440{ 441 struct turnstile *ts; 442 uintptr_t x, v, queue; 443 444 KASSERT(rw->rw_lock != RW_DESTROYED, 445 ("rw_runlock() of destroyed rwlock @ %s:%d", file, line)); 446 _rw_assert(rw, RA_RLOCKED, file, line); 447 curthread->td_locks--; 448 curthread->td_rw_rlocks--; 449 WITNESS_UNLOCK(&rw->lock_object, 0, file, line); 450 LOCK_LOG_LOCK("RUNLOCK", &rw->lock_object, 0, 0, file, line); 451 452 /* TODO: drop "owner of record" here. */ 453 454 for (;;) { 455 /* 456 * See if there is more than one read lock held. If so, 457 * just drop one and return. 458 */ 459 x = rw->rw_lock; 460 if (RW_READERS(x) > 1) { 461 if (atomic_cmpset_ptr(&rw->rw_lock, x, 462 x - RW_ONE_READER)) { 463 if (LOCK_LOG_TEST(&rw->lock_object, 0)) 464 CTR4(KTR_LOCK, 465 "%s: %p succeeded %p -> %p", 466 __func__, rw, (void *)x, 467 (void *)(x - RW_ONE_READER)); 468 break; 469 } 470 continue; 471 } 472 /* 473 * If there aren't any waiters for a write lock, then try 474 * to drop it quickly. 475 */ 476 if (!(x & RW_LOCK_WAITERS)) { 477 MPASS((x & ~RW_LOCK_WRITE_SPINNER) == 478 RW_READERS_LOCK(1)); 479 if (atomic_cmpset_ptr(&rw->rw_lock, x, RW_UNLOCKED)) { 480 if (LOCK_LOG_TEST(&rw->lock_object, 0)) 481 CTR2(KTR_LOCK, "%s: %p last succeeded", 482 __func__, rw); 483 break; 484 } 485 continue; 486 } 487 /* 488 * Ok, we know we have waiters and we think we are the 489 * last reader, so grab the turnstile lock. 490 */ 491 turnstile_chain_lock(&rw->lock_object); 492 v = rw->rw_lock & (RW_LOCK_WAITERS | RW_LOCK_WRITE_SPINNER); 493 MPASS(v & RW_LOCK_WAITERS); 494 495 /* 496 * Try to drop our lock leaving the lock in a unlocked 497 * state. 498 * 499 * If you wanted to do explicit lock handoff you'd have to 500 * do it here. You'd also want to use turnstile_signal() 501 * and you'd have to handle the race where a higher 502 * priority thread blocks on the write lock before the 503 * thread you wakeup actually runs and have the new thread 504 * "steal" the lock. For now it's a lot simpler to just 505 * wakeup all of the waiters. 506 * 507 * As above, if we fail, then another thread might have 508 * acquired a read lock, so drop the turnstile lock and 509 * restart. 510 */ 511 x = RW_UNLOCKED; 512 if (v & RW_LOCK_WRITE_WAITERS) { 513 queue = TS_EXCLUSIVE_QUEUE; 514 x |= (v & RW_LOCK_READ_WAITERS); 515 } else 516 queue = TS_SHARED_QUEUE; 517 if (!atomic_cmpset_ptr(&rw->rw_lock, RW_READERS_LOCK(1) | v, 518 x)) { 519 turnstile_chain_unlock(&rw->lock_object); 520 continue; 521 } 522 if (LOCK_LOG_TEST(&rw->lock_object, 0)) 523 CTR2(KTR_LOCK, "%s: %p last succeeded with waiters", 524 __func__, rw); 525 526 /* 527 * Ok. The lock is released and all that's left is to 528 * wake up the waiters. Note that the lock might not be 529 * free anymore, but in that case the writers will just 530 * block again if they run before the new lock holder(s) 531 * release the lock. 532 */ 533 ts = turnstile_lookup(&rw->lock_object); 534 MPASS(ts != NULL); 535 turnstile_broadcast(ts, queue); 536 turnstile_unpend(ts, TS_SHARED_LOCK); 537 turnstile_chain_unlock(&rw->lock_object); 538 break; 539 } 540 lock_profile_release_lock(&rw->lock_object); 541} 542 543/* 544 * This function is called when we are unable to obtain a write lock on the 545 * first try. This means that at least one other thread holds either a 546 * read or write lock. 547 */ 548void 549_rw_wlock_hard(struct rwlock *rw, uintptr_t tid, const char *file, int line) 550{ 551 struct turnstile *ts; 552#ifdef ADAPTIVE_RWLOCKS 553 volatile struct thread *owner; 554 int spintries = 0; 555 int i; 556#endif 557 uint64_t waittime = 0; 558 uintptr_t v, x; 559 int contested = 0; 560 561 if (rw_wlocked(rw)) { 562 KASSERT(rw->lock_object.lo_flags & RW_RECURSE, 563 ("%s: recursing but non-recursive rw %s @ %s:%d\n", 564 __func__, rw->lock_object.lo_name, file, line)); 565 rw->rw_recurse++; 566 if (LOCK_LOG_TEST(&rw->lock_object, 0)) 567 CTR2(KTR_LOCK, "%s: %p recursing", __func__, rw); 568 return; 569 } 570 571 if (LOCK_LOG_TEST(&rw->lock_object, 0)) 572 CTR5(KTR_LOCK, "%s: %s contested (lock=%p) at %s:%d", __func__, 573 rw->lock_object.lo_name, (void *)rw->rw_lock, file, line); 574 575 while (!_rw_write_lock(rw, tid)) { 576 lock_profile_obtain_lock_failed(&rw->lock_object, 577 &contested, &waittime); 578#ifdef ADAPTIVE_RWLOCKS 579 /* 580 * If the lock is write locked and the owner is 581 * running on another CPU, spin until the owner stops 582 * running or the state of the lock changes. 583 */ 584 v = rw->rw_lock; 585 owner = (struct thread *)RW_OWNER(v); 586 if (!(v & RW_LOCK_READ) && TD_IS_RUNNING(owner)) { 587 if (LOCK_LOG_TEST(&rw->lock_object, 0)) 588 CTR3(KTR_LOCK, "%s: spinning on %p held by %p", 589 __func__, rw, owner); 590 while ((struct thread*)RW_OWNER(rw->rw_lock) == owner && 591 TD_IS_RUNNING(owner)) 592 cpu_spinwait(); 593 continue; 594 } 595 if ((v & RW_LOCK_READ) && RW_READERS(v) && spintries < 100) { 596 if (!(v & RW_LOCK_WRITE_SPINNER)) { 597 if (!atomic_cmpset_ptr(&rw->rw_lock, v, 598 v | RW_LOCK_WRITE_SPINNER)) { 599 cpu_spinwait(); 600 continue; 601 } 602 } 603 spintries++; 604 for (i = 100000; i > 0; i--) { 605 if ((rw->rw_lock & RW_LOCK_WRITE_SPINNER) == 0) 606 break; 607 cpu_spinwait(); 608 } 609 if (i) 610 continue; 611 } 612#endif 613 ts = turnstile_trywait(&rw->lock_object); 614 v = rw->rw_lock; 615 616#ifdef ADAPTIVE_RWLOCKS 617 /* 618 * If the current owner of the lock is executing on another 619 * CPU quit the hard path and try to spin. 620 */ 621 if (!(v & RW_LOCK_READ)) { 622 owner = (struct thread *)RW_OWNER(v); 623 if (TD_IS_RUNNING(owner)) { 624 turnstile_cancel(ts); 625 cpu_spinwait(); 626 continue; 627 } 628 } 629#endif 630 /* 631 * If the lock was released while waiting for the turnstile 632 * chain lock retry. 633 */ 634 x = v & (RW_LOCK_WAITERS | RW_LOCK_WRITE_SPINNER); 635 if ((v & ~x) == RW_UNLOCKED) { 636 x &= ~RW_LOCK_WRITE_SPINNER; 637 if (atomic_cmpset_acq_ptr(&rw->rw_lock, v, tid | x)) { 638 if (x) 639 turnstile_claim(ts); 640 else 641 turnstile_cancel(ts); 642 break; 643 } 644 turnstile_cancel(ts); 645 cpu_spinwait(); 646 continue; 647 } 648 /* 649 * If the RW_LOCK_WRITE_WAITERS flag isn't set, then try to 650 * set it. If we fail to set it, then loop back and try 651 * again. 652 */ 653 if (!(v & RW_LOCK_WRITE_WAITERS)) { 654 if (!atomic_cmpset_ptr(&rw->rw_lock, v, 655 v | RW_LOCK_WRITE_WAITERS)) { 656 turnstile_cancel(ts); 657 cpu_spinwait(); 658 continue; 659 } 660 if (LOCK_LOG_TEST(&rw->lock_object, 0)) 661 CTR2(KTR_LOCK, "%s: %p set write waiters flag", 662 __func__, rw); 663 } 664 /* 665 * We were unable to acquire the lock and the write waiters 666 * flag is set, so we must block on the turnstile. 667 */ 668 if (LOCK_LOG_TEST(&rw->lock_object, 0)) 669 CTR2(KTR_LOCK, "%s: %p blocking on turnstile", __func__, 670 rw); 671 turnstile_wait(ts, rw_owner(rw), TS_EXCLUSIVE_QUEUE); 672 if (LOCK_LOG_TEST(&rw->lock_object, 0)) 673 CTR2(KTR_LOCK, "%s: %p resuming from turnstile", 674 __func__, rw); 675#ifdef ADAPTIVE_RWLOCKS 676 spintries = 0; 677#endif 678 } 679 lock_profile_obtain_lock_success(&rw->lock_object, contested, waittime, 680 file, line); 681} 682 683/* 684 * This function is called if the first try at releasing a write lock failed. 685 * This means that one of the 2 waiter bits must be set indicating that at 686 * least one thread is waiting on this lock. 687 */ 688void 689_rw_wunlock_hard(struct rwlock *rw, uintptr_t tid, const char *file, int line) 690{ 691 struct turnstile *ts; 692 uintptr_t v; 693 int queue; 694 695 if (rw_wlocked(rw) && rw_recursed(rw)) { 696 rw->rw_recurse--; 697 if (LOCK_LOG_TEST(&rw->lock_object, 0)) 698 CTR2(KTR_LOCK, "%s: %p unrecursing", __func__, rw); 699 return; 700 } 701 v = rw->rw_lock; 702 703 KASSERT(rw->rw_lock & (RW_LOCK_READ_WAITERS | RW_LOCK_WRITE_WAITERS), 704 ("%s: neither of the waiter flags are set", __func__)); 705 706 if (LOCK_LOG_TEST(&rw->lock_object, 0)) 707 CTR2(KTR_LOCK, "%s: %p contested", __func__, rw); 708 709 turnstile_chain_lock(&rw->lock_object); 710 ts = turnstile_lookup(&rw->lock_object); 711 712 MPASS(ts != NULL); 713 714 /* 715 * Use the same algo as sx locks for now. Prefer waking up shared 716 * waiters if we have any over writers. This is probably not ideal. 717 * 718 * 'v' is the value we are going to write back to rw_lock. If we 719 * have waiters on both queues, we need to preserve the state of 720 * the waiter flag for the queue we don't wake up. For now this is 721 * hardcoded for the algorithm mentioned above. 722 * 723 * In the case of both readers and writers waiting we wakeup the 724 * readers but leave the RW_LOCK_WRITE_WAITERS flag set. If a 725 * new writer comes in before a reader it will claim the lock up 726 * above. There is probably a potential priority inversion in 727 * there that could be worked around either by waking both queues 728 * of waiters or doing some complicated lock handoff gymnastics. 729 */ 730 v = RW_UNLOCKED; 731 if (rw->rw_lock & RW_LOCK_WRITE_WAITERS) { 732 queue = TS_EXCLUSIVE_QUEUE; 733 v |= (rw->rw_lock & RW_LOCK_READ_WAITERS); 734 } else 735 queue = TS_SHARED_QUEUE; 736 737 /* Wake up all waiters for the specific queue. */ 738 if (LOCK_LOG_TEST(&rw->lock_object, 0)) 739 CTR3(KTR_LOCK, "%s: %p waking up %s waiters", __func__, rw, 740 queue == TS_SHARED_QUEUE ? "read" : "write"); 741 turnstile_broadcast(ts, queue); 742 atomic_store_rel_ptr(&rw->rw_lock, v); 743 turnstile_unpend(ts, TS_EXCLUSIVE_LOCK); 744 turnstile_chain_unlock(&rw->lock_object); 745} 746 747/* 748 * Attempt to do a non-blocking upgrade from a read lock to a write 749 * lock. This will only succeed if this thread holds a single read 750 * lock. Returns true if the upgrade succeeded and false otherwise. 751 */ 752int 753_rw_try_upgrade(struct rwlock *rw, const char *file, int line) 754{ 755 uintptr_t v, x, tid; 756 struct turnstile *ts; 757 int success; 758 759 KASSERT(rw->rw_lock != RW_DESTROYED, 760 ("rw_try_upgrade() of destroyed rwlock @ %s:%d", file, line)); 761 _rw_assert(rw, RA_RLOCKED, file, line); 762 763 /* 764 * Attempt to switch from one reader to a writer. If there 765 * are any write waiters, then we will have to lock the 766 * turnstile first to prevent races with another writer 767 * calling turnstile_wait() before we have claimed this 768 * turnstile. So, do the simple case of no waiters first. 769 */ 770 tid = (uintptr_t)curthread; 771 success = 0; 772 for (;;) { 773 v = rw->rw_lock; 774 if (RW_READERS(v) > 1) 775 break; 776 if (!(v & RW_LOCK_WAITERS)) { 777 success = atomic_cmpset_ptr(&rw->rw_lock, v, tid); 778 if (!success) 779 continue; 780 break; 781 } 782 783 /* 784 * Ok, we think we have waiters, so lock the turnstile. 785 */ 786 ts = turnstile_trywait(&rw->lock_object); 787 v = rw->rw_lock; 788 if (RW_READERS(v) > 1) { 789 turnstile_cancel(ts); 790 break; 791 } 792 /* 793 * Try to switch from one reader to a writer again. This time 794 * we honor the current state of the waiters flags. 795 * If we obtain the lock with the flags set, then claim 796 * ownership of the turnstile. 797 */ 798 x = rw->rw_lock & RW_LOCK_WAITERS; 799 success = atomic_cmpset_ptr(&rw->rw_lock, v, tid | x); 800 if (success) { 801 if (x) 802 turnstile_claim(ts); 803 else 804 turnstile_cancel(ts); 805 break; 806 } 807 turnstile_cancel(ts); 808 } 809 LOCK_LOG_TRY("WUPGRADE", &rw->lock_object, 0, success, file, line); 810 if (success) { 811 curthread->td_rw_rlocks--; 812 WITNESS_UPGRADE(&rw->lock_object, LOP_EXCLUSIVE | LOP_TRYLOCK, 813 file, line); 814 } 815 return (success); 816} 817 818/* 819 * Downgrade a write lock into a single read lock. 820 */ 821void 822_rw_downgrade(struct rwlock *rw, const char *file, int line) 823{ 824 struct turnstile *ts; 825 uintptr_t tid, v; 826 int rwait, wwait; 827 828 KASSERT(rw->rw_lock != RW_DESTROYED, 829 ("rw_downgrade() of destroyed rwlock @ %s:%d", file, line)); 830 _rw_assert(rw, RA_WLOCKED | RA_NOTRECURSED, file, line); 831#ifndef INVARIANTS 832 if (rw_recursed(rw)) 833 panic("downgrade of a recursed lock"); 834#endif 835 836 WITNESS_DOWNGRADE(&rw->lock_object, 0, file, line); 837 838 /* 839 * Convert from a writer to a single reader. First we handle 840 * the easy case with no waiters. If there are any waiters, we 841 * lock the turnstile and "disown" the lock. 842 */ 843 tid = (uintptr_t)curthread; 844 if (atomic_cmpset_rel_ptr(&rw->rw_lock, tid, RW_READERS_LOCK(1))) 845 goto out; 846 847 /* 848 * Ok, we think we have waiters, so lock the turnstile so we can 849 * read the waiter flags without any races. 850 */ 851 turnstile_chain_lock(&rw->lock_object); 852 v = rw->rw_lock & RW_LOCK_WAITERS; 853 rwait = v & RW_LOCK_READ_WAITERS; 854 wwait = v & RW_LOCK_WRITE_WAITERS; 855 MPASS(rwait | wwait); 856 857 /* 858 * Downgrade from a write lock while preserving waiters flag 859 * and give up ownership of the turnstile. 860 */ 861 ts = turnstile_lookup(&rw->lock_object); 862 MPASS(ts != NULL); 863 if (!wwait) 864 v &= ~RW_LOCK_READ_WAITERS; 865 atomic_store_rel_ptr(&rw->rw_lock, RW_READERS_LOCK(1) | v); 866 /* 867 * Wake other readers if there are no writers pending. Otherwise they 868 * won't be able to acquire the lock anyway. 869 */ 870 if (rwait && !wwait) { 871 turnstile_broadcast(ts, TS_SHARED_QUEUE); 872 turnstile_unpend(ts, TS_EXCLUSIVE_LOCK); 873 } else 874 turnstile_disown(ts); 875 turnstile_chain_unlock(&rw->lock_object); 876out: 877 curthread->td_rw_rlocks++; 878 LOCK_LOG_LOCK("WDOWNGRADE", &rw->lock_object, 0, 0, file, line); 879} 880 881#ifdef INVARIANT_SUPPORT 882#ifndef INVARIANTS 883#undef _rw_assert 884#endif 885 886/* 887 * In the non-WITNESS case, rw_assert() can only detect that at least 888 * *some* thread owns an rlock, but it cannot guarantee that *this* 889 * thread owns an rlock. 890 */ 891void 892_rw_assert(struct rwlock *rw, int what, const char *file, int line) 893{ 894 895 if (panicstr != NULL) 896 return; 897 switch (what) { 898 case RA_LOCKED: 899 case RA_LOCKED | RA_RECURSED: 900 case RA_LOCKED | RA_NOTRECURSED: 901 case RA_RLOCKED: 902#ifdef WITNESS 903 witness_assert(&rw->lock_object, what, file, line); 904#else 905 /* 906 * If some other thread has a write lock or we have one 907 * and are asserting a read lock, fail. Also, if no one 908 * has a lock at all, fail. 909 */ 910 if (rw->rw_lock == RW_UNLOCKED || 911 (!(rw->rw_lock & RW_LOCK_READ) && (what == RA_RLOCKED || 912 rw_wowner(rw) != curthread))) 913 panic("Lock %s not %slocked @ %s:%d\n", 914 rw->lock_object.lo_name, (what == RA_RLOCKED) ? 915 "read " : "", file, line); 916 917 if (!(rw->rw_lock & RW_LOCK_READ)) { 918 if (rw_recursed(rw)) { 919 if (what & RA_NOTRECURSED) 920 panic("Lock %s recursed @ %s:%d\n", 921 rw->lock_object.lo_name, file, 922 line); 923 } else if (what & RA_RECURSED) 924 panic("Lock %s not recursed @ %s:%d\n", 925 rw->lock_object.lo_name, file, line); 926 } 927#endif 928 break; 929 case RA_WLOCKED: 930 case RA_WLOCKED | RA_RECURSED: 931 case RA_WLOCKED | RA_NOTRECURSED: 932 if (rw_wowner(rw) != curthread) 933 panic("Lock %s not exclusively locked @ %s:%d\n", 934 rw->lock_object.lo_name, file, line); 935 if (rw_recursed(rw)) { 936 if (what & RA_NOTRECURSED) 937 panic("Lock %s recursed @ %s:%d\n", 938 rw->lock_object.lo_name, file, line); 939 } else if (what & RA_RECURSED) 940 panic("Lock %s not recursed @ %s:%d\n", 941 rw->lock_object.lo_name, file, line); 942 break; 943 case RA_UNLOCKED: 944#ifdef WITNESS 945 witness_assert(&rw->lock_object, what, file, line); 946#else 947 /* 948 * If we hold a write lock fail. We can't reliably check 949 * to see if we hold a read lock or not. 950 */ 951 if (rw_wowner(rw) == curthread) 952 panic("Lock %s exclusively locked @ %s:%d\n", 953 rw->lock_object.lo_name, file, line); 954#endif 955 break; 956 default: 957 panic("Unknown rw lock assertion: %d @ %s:%d", what, file, 958 line); 959 } 960} 961#endif /* INVARIANT_SUPPORT */ 962 963#ifdef DDB 964void 965db_show_rwlock(struct lock_object *lock) 966{ 967 struct rwlock *rw; 968 struct thread *td; 969 970 rw = (struct rwlock *)lock; 971 972 db_printf(" state: "); 973 if (rw->rw_lock == RW_UNLOCKED) 974 db_printf("UNLOCKED\n"); 975 else if (rw->rw_lock == RW_DESTROYED) { 976 db_printf("DESTROYED\n"); 977 return; 978 } else if (rw->rw_lock & RW_LOCK_READ) 979 db_printf("RLOCK: %ju locks\n", 980 (uintmax_t)(RW_READERS(rw->rw_lock))); 981 else { 982 td = rw_wowner(rw); 983 db_printf("WLOCK: %p (tid %d, pid %d, \"%s\")\n", td, 984 td->td_tid, td->td_proc->p_pid, td->td_name); 985 if (rw_recursed(rw)) 986 db_printf(" recursed: %u\n", rw->rw_recurse); 987 } 988 db_printf(" waiters: "); 989 switch (rw->rw_lock & (RW_LOCK_READ_WAITERS | RW_LOCK_WRITE_WAITERS)) { 990 case RW_LOCK_READ_WAITERS: 991 db_printf("readers\n"); 992 break; 993 case RW_LOCK_WRITE_WAITERS: 994 db_printf("writers\n"); 995 break; 996 case RW_LOCK_READ_WAITERS | RW_LOCK_WRITE_WAITERS: 997 db_printf("readers and writers\n"); 998 break; 999 default: 1000 db_printf("none\n"); 1001 break; 1002 } 1003} 1004 1005#endif 1006