kern_rwlock.c revision 171052
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 171052 2007-06-26 21:31:56Z 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#include <sys/lock_profile.h> 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 lock_rw(struct lock_object *lock, int how); 63static int unlock_rw(struct lock_object *lock); 64 65struct lock_class lock_class_rw = { 66 .lc_name = "rw", 67 .lc_flags = LC_SLEEPLOCK | LC_RECURSABLE | LC_UPGRADABLE, 68#ifdef DDB 69 .lc_ddb_show = db_show_rwlock, 70#endif 71 .lc_lock = lock_rw, 72 .lc_unlock = unlock_rw, 73}; 74 75/* 76 * Return a pointer to the owning thread if the lock is write-locked or 77 * NULL if the lock is unlocked or read-locked. 78 */ 79#define rw_wowner(rw) \ 80 ((rw)->rw_lock & RW_LOCK_READ ? NULL : \ 81 (struct thread *)RW_OWNER((rw)->rw_lock)) 82 83/* 84 * Returns if a write owner is recursed. Write ownership is not assured 85 * here and should be previously checked. 86 */ 87#define rw_recursed(rw) ((rw)->rw_recurse != 0) 88 89/* 90 * Return true if curthread helds the lock. 91 */ 92#define rw_wlocked(rw) (rw_wowner((rw)) == curthread) 93 94/* 95 * Return a pointer to the owning thread for this lock who should receive 96 * any priority lent by threads that block on this lock. Currently this 97 * is identical to rw_wowner(). 98 */ 99#define rw_owner(rw) rw_wowner(rw) 100 101#ifndef INVARIANTS 102#define _rw_assert(rw, what, file, line) 103#endif 104 105void 106lock_rw(struct lock_object *lock, int how) 107{ 108 struct rwlock *rw; 109 110 rw = (struct rwlock *)lock; 111 if (how) 112 rw_wlock(rw); 113 else 114 rw_rlock(rw); 115} 116 117int 118unlock_rw(struct lock_object *lock) 119{ 120 struct rwlock *rw; 121 122 rw = (struct rwlock *)lock; 123 rw_assert(rw, RA_LOCKED | LA_NOTRECURSED); 124 if (rw->rw_lock & RW_LOCK_READ) { 125 rw_runlock(rw); 126 return (0); 127 } else { 128 rw_wunlock(rw); 129 return (1); 130 } 131} 132 133void 134rw_init_flags(struct rwlock *rw, const char *name, int opts) 135{ 136 int flags; 137 138 MPASS((opts & ~(RW_DUPOK | RW_NOPROFILE | RW_NOWITNESS | RW_QUIET | 139 RW_RECURSE)) == 0); 140 141 flags = LO_UPGRADABLE | LO_RECURSABLE; 142 if (opts & RW_DUPOK) 143 flags |= LO_DUPOK; 144 if (opts & RW_NOPROFILE) 145 flags |= LO_NOPROFILE; 146 if (!(opts & RW_NOWITNESS)) 147 flags |= LO_WITNESS; 148 if (opts & RW_QUIET) 149 flags |= LO_QUIET; 150 flags |= opts & RW_RECURSE; 151 152 rw->rw_lock = RW_UNLOCKED; 153 rw->rw_recurse = 0; 154 lock_init(&rw->lock_object, &lock_class_rw, name, NULL, flags); 155} 156 157void 158rw_destroy(struct rwlock *rw) 159{ 160 161 KASSERT(rw->rw_lock == RW_UNLOCKED, ("rw lock not unlocked")); 162 KASSERT(rw->rw_recurse == 0, ("rw lock still recursed")); 163 rw->rw_lock = RW_DESTROYED; 164 lock_destroy(&rw->lock_object); 165} 166 167void 168rw_sysinit(void *arg) 169{ 170 struct rw_args *args = arg; 171 172 rw_init(args->ra_rw, args->ra_desc); 173} 174 175int 176rw_wowned(struct rwlock *rw) 177{ 178 179 return (rw_wowner(rw) == curthread); 180} 181 182void 183_rw_wlock(struct rwlock *rw, const char *file, int line) 184{ 185 186 MPASS(curthread != NULL); 187 KASSERT(rw->rw_lock != RW_DESTROYED, 188 ("rw_wlock() of destroyed rwlock @ %s:%d", file, line)); 189 KASSERT(rw_wowner(rw) != curthread, 190 ("%s (%s): wlock already held @ %s:%d", __func__, 191 rw->lock_object.lo_name, file, line)); 192 WITNESS_CHECKORDER(&rw->lock_object, LOP_NEWORDER | LOP_EXCLUSIVE, file, 193 line); 194 __rw_wlock(rw, curthread, file, line); 195 LOCK_LOG_LOCK("WLOCK", &rw->lock_object, 0, rw->rw_recurse, file, line); 196 WITNESS_LOCK(&rw->lock_object, LOP_EXCLUSIVE, file, line); 197 curthread->td_locks++; 198} 199 200void 201_rw_wunlock(struct rwlock *rw, const char *file, int line) 202{ 203 204 MPASS(curthread != NULL); 205 KASSERT(rw->rw_lock != RW_DESTROYED, 206 ("rw_wunlock() of destroyed rwlock @ %s:%d", file, line)); 207 _rw_assert(rw, RA_WLOCKED, file, line); 208 curthread->td_locks--; 209 WITNESS_UNLOCK(&rw->lock_object, LOP_EXCLUSIVE, file, line); 210 LOCK_LOG_LOCK("WUNLOCK", &rw->lock_object, 0, rw->rw_recurse, file, 211 line); 212 if (!rw_recursed(rw)) 213 lock_profile_release_lock(&rw->lock_object); 214 __rw_wunlock(rw, curthread, file, line); 215} 216 217void 218_rw_rlock(struct rwlock *rw, const char *file, int line) 219{ 220 struct turnstile *ts; 221#ifdef ADAPTIVE_RWLOCKS 222 volatile struct thread *owner; 223#endif 224 uint64_t waittime = 0; 225 int contested = 0; 226 uintptr_t x; 227 228 KASSERT(rw->rw_lock != RW_DESTROYED, 229 ("rw_rlock() of destroyed rwlock @ %s:%d", file, line)); 230 KASSERT(rw_wowner(rw) != curthread, 231 ("%s (%s): wlock already held @ %s:%d", __func__, 232 rw->lock_object.lo_name, file, line)); 233 WITNESS_CHECKORDER(&rw->lock_object, LOP_NEWORDER, file, line); 234 235 /* 236 * Note that we don't make any attempt to try to block read 237 * locks once a writer has blocked on the lock. The reason is 238 * that we currently allow for read locks to recurse and we 239 * don't keep track of all the holders of read locks. Thus, if 240 * we were to block readers once a writer blocked and a reader 241 * tried to recurse on their reader lock after a writer had 242 * blocked we would end up in a deadlock since the reader would 243 * be blocked on the writer, and the writer would be blocked 244 * waiting for the reader to release its original read lock. 245 */ 246 for (;;) { 247 /* 248 * Handle the easy case. If no other thread has a write 249 * lock, then try to bump up the count of read locks. Note 250 * that we have to preserve the current state of the 251 * RW_LOCK_WRITE_WAITERS flag. If we fail to acquire a 252 * read lock, then rw_lock must have changed, so restart 253 * the loop. Note that this handles the case of a 254 * completely unlocked rwlock since such a lock is encoded 255 * as a read lock with no waiters. 256 */ 257 x = rw->rw_lock; 258 if (x & RW_LOCK_READ) { 259 260 /* 261 * The RW_LOCK_READ_WAITERS flag should only be set 262 * if another thread currently holds a write lock, 263 * and in that case RW_LOCK_READ should be clear. 264 */ 265 MPASS((x & RW_LOCK_READ_WAITERS) == 0); 266 if (atomic_cmpset_acq_ptr(&rw->rw_lock, x, 267 x + RW_ONE_READER)) { 268 if (LOCK_LOG_TEST(&rw->lock_object, 0)) 269 CTR4(KTR_LOCK, 270 "%s: %p succeed %p -> %p", __func__, 271 rw, (void *)x, 272 (void *)(x + RW_ONE_READER)); 273 if (RW_READERS(x) == 0) 274 lock_profile_obtain_lock_success( 275 &rw->lock_object, contested, waittime, 276 file, line); 277 break; 278 } 279 cpu_spinwait(); 280 continue; 281 } 282 lock_profile_obtain_lock_failed(&rw->lock_object, &contested, 283 &waittime); 284 285 /* 286 * Okay, now it's the hard case. Some other thread already 287 * has a write lock, so acquire the turnstile lock so we can 288 * begin the process of blocking. 289 */ 290 ts = turnstile_trywait(&rw->lock_object); 291 292 /* 293 * The lock might have been released while we spun, so 294 * recheck its state and restart the loop if there is no 295 * longer a write lock. 296 */ 297 x = rw->rw_lock; 298 if (x & RW_LOCK_READ) { 299 turnstile_cancel(ts); 300 cpu_spinwait(); 301 continue; 302 } 303 304 /* 305 * Ok, it's still a write lock. If the RW_LOCK_READ_WAITERS 306 * flag is already set, then we can go ahead and block. If 307 * it is not set then try to set it. If we fail to set it 308 * drop the turnstile lock and restart the loop. 309 */ 310 if (!(x & RW_LOCK_READ_WAITERS)) { 311 if (!atomic_cmpset_ptr(&rw->rw_lock, x, 312 x | RW_LOCK_READ_WAITERS)) { 313 turnstile_cancel(ts); 314 cpu_spinwait(); 315 continue; 316 } 317 if (LOCK_LOG_TEST(&rw->lock_object, 0)) 318 CTR2(KTR_LOCK, "%s: %p set read waiters flag", 319 __func__, rw); 320 } 321 322#ifdef ADAPTIVE_RWLOCKS 323 /* 324 * If the owner is running on another CPU, spin until 325 * the owner stops running or the state of the lock 326 * changes. 327 */ 328 owner = (struct thread *)RW_OWNER(x); 329 if (TD_IS_RUNNING(owner)) { 330 turnstile_cancel(ts); 331 if (LOCK_LOG_TEST(&rw->lock_object, 0)) 332 CTR3(KTR_LOCK, "%s: spinning on %p held by %p", 333 __func__, rw, owner); 334 while ((struct thread*)RW_OWNER(rw->rw_lock)== owner && 335 TD_IS_RUNNING(owner)) 336 cpu_spinwait(); 337 continue; 338 } 339#endif 340 341 /* 342 * We were unable to acquire the lock and the read waiters 343 * flag is set, so we must block on the turnstile. 344 */ 345 if (LOCK_LOG_TEST(&rw->lock_object, 0)) 346 CTR2(KTR_LOCK, "%s: %p blocking on turnstile", __func__, 347 rw); 348 turnstile_wait(ts, rw_owner(rw), TS_SHARED_QUEUE); 349 if (LOCK_LOG_TEST(&rw->lock_object, 0)) 350 CTR2(KTR_LOCK, "%s: %p resuming from turnstile", 351 __func__, rw); 352 } 353 354 /* 355 * TODO: acquire "owner of record" here. Here be turnstile dragons 356 * however. turnstiles don't like owners changing between calls to 357 * turnstile_wait() currently. 358 */ 359 360 LOCK_LOG_LOCK("RLOCK", &rw->lock_object, 0, 0, file, line); 361 WITNESS_LOCK(&rw->lock_object, 0, file, line); 362 curthread->td_locks++; 363} 364 365void 366_rw_runlock(struct rwlock *rw, const char *file, int line) 367{ 368 struct turnstile *ts; 369 uintptr_t x; 370 371 KASSERT(rw->rw_lock != RW_DESTROYED, 372 ("rw_runlock() of destroyed rwlock @ %s:%d", file, line)); 373 _rw_assert(rw, RA_RLOCKED, file, line); 374 curthread->td_locks--; 375 WITNESS_UNLOCK(&rw->lock_object, 0, file, line); 376 LOCK_LOG_LOCK("RUNLOCK", &rw->lock_object, 0, 0, file, line); 377 378 /* TODO: drop "owner of record" here. */ 379 380 for (;;) { 381 /* 382 * See if there is more than one read lock held. If so, 383 * just drop one and return. 384 */ 385 x = rw->rw_lock; 386 if (RW_READERS(x) > 1) { 387 if (atomic_cmpset_ptr(&rw->rw_lock, x, 388 x - RW_ONE_READER)) { 389 if (LOCK_LOG_TEST(&rw->lock_object, 0)) 390 CTR4(KTR_LOCK, 391 "%s: %p succeeded %p -> %p", 392 __func__, rw, (void *)x, 393 (void *)(x - RW_ONE_READER)); 394 break; 395 } 396 continue; 397 } 398 399 400 /* 401 * We should never have read waiters while at least one 402 * thread holds a read lock. (See note above) 403 */ 404 KASSERT(!(x & RW_LOCK_READ_WAITERS), 405 ("%s: waiting readers", __func__)); 406 407 /* 408 * If there aren't any waiters for a write lock, then try 409 * to drop it quickly. 410 */ 411 if (!(x & RW_LOCK_WRITE_WAITERS)) { 412 413 /* 414 * There shouldn't be any flags set and we should 415 * be the only read lock. If we fail to release 416 * the single read lock, then another thread might 417 * have just acquired a read lock, so go back up 418 * to the multiple read locks case. 419 */ 420 MPASS(x == RW_READERS_LOCK(1)); 421 if (atomic_cmpset_ptr(&rw->rw_lock, RW_READERS_LOCK(1), 422 RW_UNLOCKED)) { 423 if (LOCK_LOG_TEST(&rw->lock_object, 0)) 424 CTR2(KTR_LOCK, "%s: %p last succeeded", 425 __func__, rw); 426 break; 427 } 428 continue; 429 } 430 431 /* 432 * There should just be one reader with one or more 433 * writers waiting. 434 */ 435 MPASS(x == (RW_READERS_LOCK(1) | RW_LOCK_WRITE_WAITERS)); 436 437 /* 438 * Ok, we know we have a waiting writer and we think we 439 * are the last reader, so grab the turnstile lock. 440 */ 441 turnstile_chain_lock(&rw->lock_object); 442 443 /* 444 * Try to drop our lock leaving the lock in a unlocked 445 * state. 446 * 447 * If you wanted to do explicit lock handoff you'd have to 448 * do it here. You'd also want to use turnstile_signal() 449 * and you'd have to handle the race where a higher 450 * priority thread blocks on the write lock before the 451 * thread you wakeup actually runs and have the new thread 452 * "steal" the lock. For now it's a lot simpler to just 453 * wakeup all of the waiters. 454 * 455 * As above, if we fail, then another thread might have 456 * acquired a read lock, so drop the turnstile lock and 457 * restart. 458 */ 459 if (!atomic_cmpset_ptr(&rw->rw_lock, 460 RW_READERS_LOCK(1) | RW_LOCK_WRITE_WAITERS, RW_UNLOCKED)) { 461 turnstile_chain_unlock(&rw->lock_object); 462 continue; 463 } 464 if (LOCK_LOG_TEST(&rw->lock_object, 0)) 465 CTR2(KTR_LOCK, "%s: %p last succeeded with waiters", 466 __func__, rw); 467 468 /* 469 * Ok. The lock is released and all that's left is to 470 * wake up the waiters. Note that the lock might not be 471 * free anymore, but in that case the writers will just 472 * block again if they run before the new lock holder(s) 473 * release the lock. 474 */ 475 ts = turnstile_lookup(&rw->lock_object); 476 MPASS(ts != NULL); 477 turnstile_broadcast(ts, TS_EXCLUSIVE_QUEUE); 478 turnstile_unpend(ts, TS_SHARED_LOCK); 479 turnstile_chain_unlock(&rw->lock_object); 480 break; 481 } 482 lock_profile_release_lock(&rw->lock_object); 483} 484 485/* 486 * This function is called when we are unable to obtain a write lock on the 487 * first try. This means that at least one other thread holds either a 488 * read or write lock. 489 */ 490void 491_rw_wlock_hard(struct rwlock *rw, uintptr_t tid, const char *file, int line) 492{ 493 struct turnstile *ts; 494#ifdef ADAPTIVE_RWLOCKS 495 volatile struct thread *owner; 496#endif 497 uintptr_t v; 498 499 if (rw_wlocked(rw)) { 500 KASSERT(rw->lock_object.lo_flags & RW_RECURSE, 501 ("%s: recursing but non-recursive rw %s @ %s:%d\n", 502 __func__, rw->lock_object.lo_name, file, line)); 503 rw->rw_recurse++; 504 atomic_set_ptr(&rw->rw_lock, RW_LOCK_RECURSED); 505 if (LOCK_LOG_TEST(&rw->lock_object, 0)) 506 CTR2(KTR_LOCK, "%s: %p recursing", __func__, rw); 507 return; 508 } 509 510 if (LOCK_LOG_TEST(&rw->lock_object, 0)) 511 CTR5(KTR_LOCK, "%s: %s contested (lock=%p) at %s:%d", __func__, 512 rw->lock_object.lo_name, (void *)rw->rw_lock, file, line); 513 514 while (!_rw_write_lock(rw, tid)) { 515 ts = turnstile_trywait(&rw->lock_object); 516 v = rw->rw_lock; 517 518 /* 519 * If the lock was released while spinning on the 520 * turnstile chain lock, try again. 521 */ 522 if (v == RW_UNLOCKED) { 523 turnstile_cancel(ts); 524 cpu_spinwait(); 525 continue; 526 } 527 528 /* 529 * If the lock was released by a writer with both readers 530 * and writers waiting and a reader hasn't woken up and 531 * acquired the lock yet, rw_lock will be set to the 532 * value RW_UNLOCKED | RW_LOCK_WRITE_WAITERS. If we see 533 * that value, try to acquire it once. Note that we have 534 * to preserve the RW_LOCK_WRITE_WAITERS flag as there are 535 * other writers waiting still. If we fail, restart the 536 * loop. 537 */ 538 if (v == (RW_UNLOCKED | RW_LOCK_WRITE_WAITERS)) { 539 if (atomic_cmpset_acq_ptr(&rw->rw_lock, 540 RW_UNLOCKED | RW_LOCK_WRITE_WAITERS, 541 tid | RW_LOCK_WRITE_WAITERS)) { 542 turnstile_claim(ts); 543 CTR2(KTR_LOCK, "%s: %p claimed by new writer", 544 __func__, rw); 545 break; 546 } 547 turnstile_cancel(ts); 548 cpu_spinwait(); 549 continue; 550 } 551 552 /* 553 * If the RW_LOCK_WRITE_WAITERS flag isn't set, then try to 554 * set it. If we fail to set it, then loop back and try 555 * again. 556 */ 557 if (!(v & RW_LOCK_WRITE_WAITERS)) { 558 if (!atomic_cmpset_ptr(&rw->rw_lock, v, 559 v | RW_LOCK_WRITE_WAITERS)) { 560 turnstile_cancel(ts); 561 cpu_spinwait(); 562 continue; 563 } 564 if (LOCK_LOG_TEST(&rw->lock_object, 0)) 565 CTR2(KTR_LOCK, "%s: %p set write waiters flag", 566 __func__, rw); 567 } 568 569#ifdef ADAPTIVE_RWLOCKS 570 /* 571 * If the lock is write locked and the owner is 572 * running on another CPU, spin until the owner stops 573 * running or the state of the lock changes. 574 */ 575 owner = (struct thread *)RW_OWNER(v); 576 if (!(v & RW_LOCK_READ) && TD_IS_RUNNING(owner)) { 577 turnstile_cancel(ts); 578 if (LOCK_LOG_TEST(&rw->lock_object, 0)) 579 CTR3(KTR_LOCK, "%s: spinning on %p held by %p", 580 __func__, rw, owner); 581 while ((struct thread*)RW_OWNER(rw->rw_lock)== owner && 582 TD_IS_RUNNING(owner)) 583 cpu_spinwait(); 584 continue; 585 } 586#endif 587 588 /* 589 * We were unable to acquire the lock and the write waiters 590 * flag is set, so we must block on the turnstile. 591 */ 592 if (LOCK_LOG_TEST(&rw->lock_object, 0)) 593 CTR2(KTR_LOCK, "%s: %p blocking on turnstile", __func__, 594 rw); 595 turnstile_wait(ts, rw_owner(rw), TS_EXCLUSIVE_QUEUE); 596 if (LOCK_LOG_TEST(&rw->lock_object, 0)) 597 CTR2(KTR_LOCK, "%s: %p resuming from turnstile", 598 __func__, rw); 599 } 600} 601 602/* 603 * This function is called if the first try at releasing a write lock failed. 604 * This means that one of the 2 waiter bits must be set indicating that at 605 * least one thread is waiting on this lock. 606 */ 607void 608_rw_wunlock_hard(struct rwlock *rw, uintptr_t tid, const char *file, int line) 609{ 610 struct turnstile *ts; 611 uintptr_t v; 612 int queue; 613 614 if (rw_wlocked(rw) && rw_recursed(rw)) { 615 if ((--rw->rw_recurse) == 0) 616 atomic_clear_ptr(&rw->rw_lock, RW_LOCK_RECURSED); 617 if (LOCK_LOG_TEST(&rw->lock_object, 0)) 618 CTR2(KTR_LOCK, "%s: %p unrecursing", __func__, rw); 619 return; 620 } 621 622 KASSERT(rw->rw_lock & (RW_LOCK_READ_WAITERS | RW_LOCK_WRITE_WAITERS), 623 ("%s: neither of the waiter flags are set", __func__)); 624 625 if (LOCK_LOG_TEST(&rw->lock_object, 0)) 626 CTR2(KTR_LOCK, "%s: %p contested", __func__, rw); 627 628 turnstile_chain_lock(&rw->lock_object); 629 ts = turnstile_lookup(&rw->lock_object); 630 631#ifdef ADAPTIVE_RWLOCKS 632 /* 633 * There might not be a turnstile for this lock if all of 634 * the waiters are adaptively spinning. In that case, just 635 * reset the lock to the unlocked state and return. 636 */ 637 if (ts == NULL) { 638 atomic_store_rel_ptr(&rw->rw_lock, RW_UNLOCKED); 639 if (LOCK_LOG_TEST(&rw->lock_object, 0)) 640 CTR2(KTR_LOCK, "%s: %p no sleepers", __func__, rw); 641 turnstile_chain_unlock(&rw->lock_object); 642 return; 643 } 644#else 645 MPASS(ts != NULL); 646#endif 647 648 /* 649 * Use the same algo as sx locks for now. Prefer waking up shared 650 * waiters if we have any over writers. This is probably not ideal. 651 * 652 * 'v' is the value we are going to write back to rw_lock. If we 653 * have waiters on both queues, we need to preserve the state of 654 * the waiter flag for the queue we don't wake up. For now this is 655 * hardcoded for the algorithm mentioned above. 656 * 657 * In the case of both readers and writers waiting we wakeup the 658 * readers but leave the RW_LOCK_WRITE_WAITERS flag set. If a 659 * new writer comes in before a reader it will claim the lock up 660 * above. There is probably a potential priority inversion in 661 * there that could be worked around either by waking both queues 662 * of waiters or doing some complicated lock handoff gymnastics. 663 * 664 * Note that in the ADAPTIVE_RWLOCKS case, if both flags are 665 * set, there might not be any actual writers on the turnstile 666 * as they might all be spinning. In that case, we don't want 667 * to preserve the RW_LOCK_WRITE_WAITERS flag as the turnstile 668 * is going to go away once we wakeup all the readers. 669 */ 670 v = RW_UNLOCKED; 671 if (rw->rw_lock & RW_LOCK_READ_WAITERS) { 672 queue = TS_SHARED_QUEUE; 673#ifdef ADAPTIVE_RWLOCKS 674 if (rw->rw_lock & RW_LOCK_WRITE_WAITERS && 675 !turnstile_empty(ts, TS_EXCLUSIVE_QUEUE)) 676 v |= RW_LOCK_WRITE_WAITERS; 677#else 678 v |= (rw->rw_lock & RW_LOCK_WRITE_WAITERS); 679#endif 680 } else 681 queue = TS_EXCLUSIVE_QUEUE; 682 683#ifdef ADAPTIVE_RWLOCKS 684 /* 685 * We have to make sure that we actually have waiters to 686 * wakeup. If they are all spinning, then we just need to 687 * disown the turnstile and return. 688 */ 689 if (turnstile_empty(ts, queue)) { 690 if (LOCK_LOG_TEST(&rw->lock_object, 0)) 691 CTR2(KTR_LOCK, "%s: %p no sleepers 2", __func__, rw); 692 atomic_store_rel_ptr(&rw->rw_lock, v); 693 turnstile_disown(ts); 694 turnstile_chain_unlock(&rw->lock_object); 695 return; 696 } 697#endif 698 699 /* Wake up all waiters for the specific queue. */ 700 if (LOCK_LOG_TEST(&rw->lock_object, 0)) 701 CTR3(KTR_LOCK, "%s: %p waking up %s waiters", __func__, rw, 702 queue == TS_SHARED_QUEUE ? "read" : "write"); 703 turnstile_broadcast(ts, queue); 704 atomic_store_rel_ptr(&rw->rw_lock, v); 705 turnstile_unpend(ts, TS_EXCLUSIVE_LOCK); 706 turnstile_chain_unlock(&rw->lock_object); 707} 708 709/* 710 * Attempt to do a non-blocking upgrade from a read lock to a write 711 * lock. This will only succeed if this thread holds a single read 712 * lock. Returns true if the upgrade succeeded and false otherwise. 713 */ 714int 715_rw_try_upgrade(struct rwlock *rw, const char *file, int line) 716{ 717 uintptr_t v, tid; 718 struct turnstile *ts; 719 int success; 720 721 KASSERT(rw->rw_lock != RW_DESTROYED, 722 ("rw_try_upgrade() of destroyed rwlock @ %s:%d", file, line)); 723 _rw_assert(rw, RA_RLOCKED, file, line); 724 725 /* 726 * Attempt to switch from one reader to a writer. If there 727 * are any write waiters, then we will have to lock the 728 * turnstile first to prevent races with another writer 729 * calling turnstile_wait() before we have claimed this 730 * turnstile. So, do the simple case of no waiters first. 731 */ 732 tid = (uintptr_t)curthread; 733 if (!(rw->rw_lock & RW_LOCK_WRITE_WAITERS)) { 734 success = atomic_cmpset_ptr(&rw->rw_lock, RW_READERS_LOCK(1), 735 tid); 736 goto out; 737 } 738 739 /* 740 * Ok, we think we have write waiters, so lock the 741 * turnstile. 742 */ 743 ts = turnstile_trywait(&rw->lock_object); 744 745 /* 746 * Try to switch from one reader to a writer again. This time 747 * we honor the current state of the RW_LOCK_WRITE_WAITERS 748 * flag. If we obtain the lock with the flag set, then claim 749 * ownership of the turnstile. In the ADAPTIVE_RWLOCKS case 750 * it is possible for there to not be an associated turnstile 751 * even though there are waiters if all of the waiters are 752 * spinning. 753 */ 754 v = rw->rw_lock & RW_LOCK_WRITE_WAITERS; 755 success = atomic_cmpset_ptr(&rw->rw_lock, RW_READERS_LOCK(1) | v, 756 tid | v); 757#ifdef ADAPTIVE_RWLOCKS 758 if (success && v && turnstile_lookup(&rw->lock_object) != NULL) 759#else 760 if (success && v) 761#endif 762 turnstile_claim(ts); 763 else 764 turnstile_cancel(ts); 765out: 766 LOCK_LOG_TRY("WUPGRADE", &rw->lock_object, 0, success, file, line); 767 if (success) 768 WITNESS_UPGRADE(&rw->lock_object, LOP_EXCLUSIVE | LOP_TRYLOCK, 769 file, line); 770 return (success); 771} 772 773/* 774 * Downgrade a write lock into a single read lock. 775 */ 776void 777_rw_downgrade(struct rwlock *rw, const char *file, int line) 778{ 779 struct turnstile *ts; 780 uintptr_t tid, v; 781 782 KASSERT(rw->rw_lock != RW_DESTROYED, 783 ("rw_downgrade() of destroyed rwlock @ %s:%d", file, line)); 784 _rw_assert(rw, RA_WLOCKED | RA_NOTRECURSED, file, line); 785#ifndef INVARIANTS 786 if (rw_recursed(rw)) 787 panic("downgrade of a recursed lock"); 788#endif 789 790 WITNESS_DOWNGRADE(&rw->lock_object, 0, file, line); 791 792 /* 793 * Convert from a writer to a single reader. First we handle 794 * the easy case with no waiters. If there are any waiters, we 795 * lock the turnstile, "disown" the lock, and awaken any read 796 * waiters. 797 */ 798 tid = (uintptr_t)curthread; 799 if (atomic_cmpset_rel_ptr(&rw->rw_lock, tid, RW_READERS_LOCK(1))) 800 goto out; 801 802 /* 803 * Ok, we think we have waiters, so lock the turnstile so we can 804 * read the waiter flags without any races. 805 */ 806 turnstile_chain_lock(&rw->lock_object); 807 v = rw->rw_lock; 808 MPASS(v & (RW_LOCK_READ_WAITERS | RW_LOCK_WRITE_WAITERS)); 809 810 /* 811 * Downgrade from a write lock while preserving 812 * RW_LOCK_WRITE_WAITERS and give up ownership of the 813 * turnstile. If there are any read waiters, wake them up. 814 * 815 * For ADAPTIVE_RWLOCKS, we have to allow for the fact that 816 * all of the read waiters might be spinning. In that case, 817 * act as if RW_LOCK_READ_WAITERS is not set. Also, only 818 * preserve the RW_LOCK_WRITE_WAITERS flag if at least one 819 * writer is blocked on the turnstile. 820 */ 821 ts = turnstile_lookup(&rw->lock_object); 822#ifdef ADAPTIVE_RWLOCKS 823 if (ts == NULL) 824 v &= ~(RW_LOCK_READ_WAITERS | RW_LOCK_WRITE_WAITERS); 825 else if (v & RW_LOCK_READ_WAITERS && 826 turnstile_empty(ts, TS_SHARED_QUEUE)) 827 v &= ~RW_LOCK_READ_WAITERS; 828 else if (v & RW_LOCK_WRITE_WAITERS && 829 turnstile_empty(ts, TS_EXCLUSIVE_QUEUE)) 830 v &= ~RW_LOCK_WRITE_WAITERS; 831#else 832 MPASS(ts != NULL); 833#endif 834 if (v & RW_LOCK_READ_WAITERS) 835 turnstile_broadcast(ts, TS_SHARED_QUEUE); 836 atomic_store_rel_ptr(&rw->rw_lock, RW_READERS_LOCK(1) | 837 (v & RW_LOCK_WRITE_WAITERS)); 838 if (v & RW_LOCK_READ_WAITERS) 839 turnstile_unpend(ts, TS_EXCLUSIVE_LOCK); 840 else if (ts) 841 turnstile_disown(ts); 842 turnstile_chain_unlock(&rw->lock_object); 843out: 844 LOCK_LOG_LOCK("WDOWNGRADE", &rw->lock_object, 0, 0, file, line); 845} 846 847#ifdef INVARIANT_SUPPORT 848#ifndef INVARIANTS 849#undef _rw_assert 850#endif 851 852/* 853 * In the non-WITNESS case, rw_assert() can only detect that at least 854 * *some* thread owns an rlock, but it cannot guarantee that *this* 855 * thread owns an rlock. 856 */ 857void 858_rw_assert(struct rwlock *rw, int what, const char *file, int line) 859{ 860 861 if (panicstr != NULL) 862 return; 863 switch (what) { 864 case RA_LOCKED: 865 case RA_LOCKED | RA_RECURSED: 866 case RA_LOCKED | RA_NOTRECURSED: 867 case RA_RLOCKED: 868#ifdef WITNESS 869 witness_assert(&rw->lock_object, what, file, line); 870#else 871 /* 872 * If some other thread has a write lock or we have one 873 * and are asserting a read lock, fail. Also, if no one 874 * has a lock at all, fail. 875 */ 876 if (rw->rw_lock == RW_UNLOCKED || 877 (!(rw->rw_lock & RW_LOCK_READ) && (what == RA_RLOCKED || 878 rw_wowner(rw) != curthread))) 879 panic("Lock %s not %slocked @ %s:%d\n", 880 rw->lock_object.lo_name, (what == RA_RLOCKED) ? 881 "read " : "", file, line); 882 883 if (!(rw->rw_lock & RW_LOCK_READ)) { 884 if (rw_recursed(rw)) { 885 if (what & RA_NOTRECURSED) 886 panic("Lock %s recursed @ %s:%d\n", 887 rw->lock_object.lo_name, file, 888 line); 889 } else if (what & RA_RECURSED) 890 panic("Lock %s not recursed @ %s:%d\n", 891 rw->lock_object.lo_name, file, line); 892 } 893#endif 894 break; 895 case RA_WLOCKED: 896 case RA_WLOCKED | RA_RECURSED: 897 case RA_WLOCKED | RA_NOTRECURSED: 898 if (rw_wowner(rw) != curthread) 899 panic("Lock %s not exclusively locked @ %s:%d\n", 900 rw->lock_object.lo_name, file, line); 901 if (rw_recursed(rw)) { 902 if (what & RA_NOTRECURSED) 903 panic("Lock %s recursed @ %s:%d\n", 904 rw->lock_object.lo_name, file, line); 905 } else if (what & RA_RECURSED) 906 panic("Lock %s not recursed @ %s:%d\n", 907 rw->lock_object.lo_name, file, line); 908 break; 909 case RA_UNLOCKED: 910#ifdef WITNESS 911 witness_assert(&rw->lock_object, what, file, line); 912#else 913 /* 914 * If we hold a write lock fail. We can't reliably check 915 * to see if we hold a read lock or not. 916 */ 917 if (rw_wowner(rw) == curthread) 918 panic("Lock %s exclusively locked @ %s:%d\n", 919 rw->lock_object.lo_name, file, line); 920#endif 921 break; 922 default: 923 panic("Unknown rw lock assertion: %d @ %s:%d", what, file, 924 line); 925 } 926} 927#endif /* INVARIANT_SUPPORT */ 928 929#ifdef DDB 930void 931db_show_rwlock(struct lock_object *lock) 932{ 933 struct rwlock *rw; 934 struct thread *td; 935 936 rw = (struct rwlock *)lock; 937 938 db_printf(" state: "); 939 if (rw->rw_lock == RW_UNLOCKED) 940 db_printf("UNLOCKED\n"); 941 else if (rw->rw_lock == RW_DESTROYED) { 942 db_printf("DESTROYED\n"); 943 return; 944 } else if (rw->rw_lock & RW_LOCK_READ) 945 db_printf("RLOCK: %ju locks\n", 946 (uintmax_t)(RW_READERS(rw->rw_lock))); 947 else { 948 td = rw_wowner(rw); 949 db_printf("WLOCK: %p (tid %d, pid %d, \"%s\")\n", td, 950 td->td_tid, td->td_proc->p_pid, td->td_proc->p_comm); 951 if (rw_recursed(rw)) 952 db_printf(" recursed: %u\n", rw->rw_recurse); 953 } 954 db_printf(" waiters: "); 955 switch (rw->rw_lock & (RW_LOCK_READ_WAITERS | RW_LOCK_WRITE_WAITERS)) { 956 case RW_LOCK_READ_WAITERS: 957 db_printf("readers\n"); 958 break; 959 case RW_LOCK_WRITE_WAITERS: 960 db_printf("writers\n"); 961 break; 962 case RW_LOCK_READ_WAITERS | RW_LOCK_WRITE_WAITERS: 963 db_printf("readers and writers\n"); 964 break; 965 default: 966 db_printf("none\n"); 967 break; 968 } 969} 970 971#endif 972