kern_rwlock.c revision 192853
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 192853 2009-05-26 20:28:22Z sson $"); 36 37#include "opt_ddb.h" 38#include "opt_kdtrace.h" 39#include "opt_no_adaptive_rwlocks.h" 40 41#include <sys/param.h> 42#include <sys/ktr.h> 43#include <sys/kernel.h> 44#include <sys/lock.h> 45#include <sys/mutex.h> 46#include <sys/proc.h> 47#include <sys/rwlock.h> 48#include <sys/sysctl.h> 49#include <sys/systm.h> 50#include <sys/turnstile.h> 51 52#include <machine/cpu.h> 53 54CTASSERT((RW_RECURSE & LO_CLASSFLAGS) == RW_RECURSE); 55 56#if defined(SMP) && !defined(NO_ADAPTIVE_RWLOCKS) 57#define ADAPTIVE_RWLOCKS 58#endif 59 60#ifdef ADAPTIVE_RWLOCKS 61static int rowner_retries = 10; 62static int rowner_loops = 10000; 63SYSCTL_NODE(_debug, OID_AUTO, rwlock, CTLFLAG_RD, NULL, "rwlock debugging"); 64SYSCTL_INT(_debug_rwlock, OID_AUTO, retry, CTLFLAG_RW, &rowner_retries, 0, ""); 65SYSCTL_INT(_debug_rwlock, OID_AUTO, loops, CTLFLAG_RW, &rowner_loops, 0, ""); 66#endif 67 68#ifdef DDB 69#include <ddb/ddb.h> 70 71static void db_show_rwlock(struct lock_object *lock); 72#endif 73static void assert_rw(struct lock_object *lock, int what); 74static void lock_rw(struct lock_object *lock, int how); 75#ifdef KDTRACE_HOOKS 76static int owner_rw(struct lock_object *lock, struct thread **owner); 77#endif 78static int unlock_rw(struct lock_object *lock); 79 80struct lock_class lock_class_rw = { 81 .lc_name = "rw", 82 .lc_flags = LC_SLEEPLOCK | LC_RECURSABLE | LC_UPGRADABLE, 83 .lc_assert = assert_rw, 84#ifdef DDB 85 .lc_ddb_show = db_show_rwlock, 86#endif 87 .lc_lock = lock_rw, 88 .lc_unlock = unlock_rw, 89#ifdef KDTRACE_HOOKS 90 .lc_owner = owner_rw, 91#endif 92}; 93 94/* 95 * Return a pointer to the owning thread if the lock is write-locked or 96 * NULL if the lock is unlocked or read-locked. 97 */ 98#define rw_wowner(rw) \ 99 ((rw)->rw_lock & RW_LOCK_READ ? NULL : \ 100 (struct thread *)RW_OWNER((rw)->rw_lock)) 101 102/* 103 * Returns if a write owner is recursed. Write ownership is not assured 104 * here and should be previously checked. 105 */ 106#define rw_recursed(rw) ((rw)->rw_recurse != 0) 107 108/* 109 * Return true if curthread helds the lock. 110 */ 111#define rw_wlocked(rw) (rw_wowner((rw)) == curthread) 112 113/* 114 * Return a pointer to the owning thread for this lock who should receive 115 * any priority lent by threads that block on this lock. Currently this 116 * is identical to rw_wowner(). 117 */ 118#define rw_owner(rw) rw_wowner(rw) 119 120#ifndef INVARIANTS 121#define _rw_assert(rw, what, file, line) 122#endif 123 124void 125assert_rw(struct lock_object *lock, int what) 126{ 127 128 rw_assert((struct rwlock *)lock, what); 129} 130 131void 132lock_rw(struct lock_object *lock, int how) 133{ 134 struct rwlock *rw; 135 136 rw = (struct rwlock *)lock; 137 if (how) 138 rw_wlock(rw); 139 else 140 rw_rlock(rw); 141} 142 143int 144unlock_rw(struct lock_object *lock) 145{ 146 struct rwlock *rw; 147 148 rw = (struct rwlock *)lock; 149 rw_assert(rw, RA_LOCKED | LA_NOTRECURSED); 150 if (rw->rw_lock & RW_LOCK_READ) { 151 rw_runlock(rw); 152 return (0); 153 } else { 154 rw_wunlock(rw); 155 return (1); 156 } 157} 158 159#ifdef KDTRACE_HOOKS 160int 161owner_rw(struct lock_object *lock, struct thread **owner) 162{ 163 struct rwlock *rw = (struct rwlock *)lock; 164 uintptr_t x = rw->rw_lock; 165 166 *owner = rw_wowner(rw); 167 return ((x & RW_LOCK_READ) != 0 ? (RW_READERS(x) != 0) : 168 (*owner != NULL)); 169} 170#endif 171 172void 173rw_init_flags(struct rwlock *rw, const char *name, int opts) 174{ 175 int flags; 176 177 MPASS((opts & ~(RW_DUPOK | RW_NOPROFILE | RW_NOWITNESS | RW_QUIET | 178 RW_RECURSE)) == 0); 179 180 flags = LO_UPGRADABLE | LO_RECURSABLE; 181 if (opts & RW_DUPOK) 182 flags |= LO_DUPOK; 183 if (opts & RW_NOPROFILE) 184 flags |= LO_NOPROFILE; 185 if (!(opts & RW_NOWITNESS)) 186 flags |= LO_WITNESS; 187 if (opts & RW_QUIET) 188 flags |= LO_QUIET; 189 flags |= opts & RW_RECURSE; 190 191 rw->rw_lock = RW_UNLOCKED; 192 rw->rw_recurse = 0; 193 lock_init(&rw->lock_object, &lock_class_rw, name, NULL, flags); 194} 195 196void 197rw_destroy(struct rwlock *rw) 198{ 199 200 KASSERT(rw->rw_lock == RW_UNLOCKED, ("rw lock not unlocked")); 201 KASSERT(rw->rw_recurse == 0, ("rw lock still recursed")); 202 rw->rw_lock = RW_DESTROYED; 203 lock_destroy(&rw->lock_object); 204} 205 206void 207rw_sysinit(void *arg) 208{ 209 struct rw_args *args = arg; 210 211 rw_init(args->ra_rw, args->ra_desc); 212} 213 214void 215rw_sysinit_flags(void *arg) 216{ 217 struct rw_args_flags *args = arg; 218 219 rw_init_flags(args->ra_rw, args->ra_desc, args->ra_flags); 220} 221 222int 223rw_wowned(struct rwlock *rw) 224{ 225 226 return (rw_wowner(rw) == curthread); 227} 228 229void 230_rw_wlock(struct rwlock *rw, const char *file, int line) 231{ 232 233 MPASS(curthread != NULL); 234 KASSERT(rw->rw_lock != RW_DESTROYED, 235 ("rw_wlock() of destroyed rwlock @ %s:%d", file, line)); 236 WITNESS_CHECKORDER(&rw->lock_object, LOP_NEWORDER | LOP_EXCLUSIVE, file, 237 line, NULL); 238 __rw_wlock(rw, curthread, file, line); 239 LOCK_LOG_LOCK("WLOCK", &rw->lock_object, 0, rw->rw_recurse, file, line); 240 WITNESS_LOCK(&rw->lock_object, LOP_EXCLUSIVE, file, line); 241 curthread->td_locks++; 242} 243 244int 245_rw_try_wlock(struct rwlock *rw, const char *file, int line) 246{ 247 int rval; 248 249 KASSERT(rw->rw_lock != RW_DESTROYED, 250 ("rw_try_wlock() of destroyed rwlock @ %s:%d", file, line)); 251 252 if (rw_wlocked(rw) && (rw->lock_object.lo_flags & RW_RECURSE) != 0) { 253 rw->rw_recurse++; 254 rval = 1; 255 } else 256 rval = atomic_cmpset_acq_ptr(&rw->rw_lock, RW_UNLOCKED, 257 (uintptr_t)curthread); 258 259 LOCK_LOG_TRY("WLOCK", &rw->lock_object, 0, rval, file, line); 260 if (rval) { 261 WITNESS_LOCK(&rw->lock_object, LOP_EXCLUSIVE | LOP_TRYLOCK, 262 file, line); 263 curthread->td_locks++; 264 } 265 return (rval); 266} 267 268void 269_rw_wunlock(struct rwlock *rw, const char *file, int line) 270{ 271 272 MPASS(curthread != NULL); 273 KASSERT(rw->rw_lock != RW_DESTROYED, 274 ("rw_wunlock() of destroyed rwlock @ %s:%d", file, line)); 275 _rw_assert(rw, RA_WLOCKED, file, line); 276 curthread->td_locks--; 277 WITNESS_UNLOCK(&rw->lock_object, LOP_EXCLUSIVE, file, line); 278 LOCK_LOG_LOCK("WUNLOCK", &rw->lock_object, 0, rw->rw_recurse, file, 279 line); 280 if (!rw_recursed(rw)) 281 LOCKSTAT_PROFILE_RELEASE_LOCK(LS_RW_WUNLOCK_RELEASE, rw); 282 __rw_wunlock(rw, curthread, file, line); 283} 284/* 285 * Determines whether a new reader can acquire a lock. Succeeds if the 286 * reader already owns a read lock and the lock is locked for read to 287 * prevent deadlock from reader recursion. Also succeeds if the lock 288 * is unlocked and has no writer waiters or spinners. Failing otherwise 289 * prioritizes writers before readers. 290 */ 291#define RW_CAN_READ(_rw) \ 292 ((curthread->td_rw_rlocks && (_rw) & RW_LOCK_READ) || ((_rw) & \ 293 (RW_LOCK_READ | RW_LOCK_WRITE_WAITERS | RW_LOCK_WRITE_SPINNER)) == \ 294 RW_LOCK_READ) 295 296void 297_rw_rlock(struct rwlock *rw, const char *file, int line) 298{ 299 struct turnstile *ts; 300#ifdef ADAPTIVE_RWLOCKS 301 volatile struct thread *owner; 302 int spintries = 0; 303 int i; 304#endif 305#ifdef LOCK_PROFILING 306 uint64_t waittime = 0; 307 int contested = 0; 308#endif 309 uintptr_t v; 310#ifdef KDTRACE_HOOKS 311 uint64_t spin_cnt = 0; 312 uint64_t sleep_cnt = 0; 313 int64_t sleep_time = 0; 314#endif 315 316 KASSERT(rw->rw_lock != RW_DESTROYED, 317 ("rw_rlock() of destroyed rwlock @ %s:%d", file, line)); 318 KASSERT(rw_wowner(rw) != curthread, 319 ("%s (%s): wlock already held @ %s:%d", __func__, 320 rw->lock_object.lo_name, file, line)); 321 WITNESS_CHECKORDER(&rw->lock_object, LOP_NEWORDER, file, line, NULL); 322 323 for (;;) { 324#ifdef KDTRACE_HOOKS 325 spin_cnt++; 326#endif 327 /* 328 * Handle the easy case. If no other thread has a write 329 * lock, then try to bump up the count of read locks. Note 330 * that we have to preserve the current state of the 331 * RW_LOCK_WRITE_WAITERS flag. If we fail to acquire a 332 * read lock, then rw_lock must have changed, so restart 333 * the loop. Note that this handles the case of a 334 * completely unlocked rwlock since such a lock is encoded 335 * as a read lock with no waiters. 336 */ 337 v = rw->rw_lock; 338 if (RW_CAN_READ(v)) { 339 /* 340 * The RW_LOCK_READ_WAITERS flag should only be set 341 * if the lock has been unlocked and write waiters 342 * were present. 343 */ 344 if (atomic_cmpset_acq_ptr(&rw->rw_lock, v, 345 v + RW_ONE_READER)) { 346 if (LOCK_LOG_TEST(&rw->lock_object, 0)) 347 CTR4(KTR_LOCK, 348 "%s: %p succeed %p -> %p", __func__, 349 rw, (void *)v, 350 (void *)(v + RW_ONE_READER)); 351 break; 352 } 353 cpu_spinwait(); 354 continue; 355 } 356 lock_profile_obtain_lock_failed(&rw->lock_object, 357 &contested, &waittime); 358 359#ifdef ADAPTIVE_RWLOCKS 360 /* 361 * If the owner is running on another CPU, spin until 362 * the owner stops running or the state of the lock 363 * changes. 364 */ 365 if ((v & RW_LOCK_READ) == 0) { 366 owner = (struct thread *)RW_OWNER(v); 367 if (TD_IS_RUNNING(owner)) { 368 if (LOCK_LOG_TEST(&rw->lock_object, 0)) 369 CTR3(KTR_LOCK, 370 "%s: spinning on %p held by %p", 371 __func__, rw, owner); 372 while ((struct thread*)RW_OWNER(rw->rw_lock) == 373 owner && TD_IS_RUNNING(owner)) { 374 cpu_spinwait(); 375#ifdef KDTRACE_HOOKS 376 spin_cnt++; 377#endif 378 } 379 continue; 380 } 381 } else if (spintries < rowner_retries) { 382 spintries++; 383 for (i = 0; i < rowner_loops; i++) { 384 v = rw->rw_lock; 385 if ((v & RW_LOCK_READ) == 0 || RW_CAN_READ(v)) 386 break; 387 cpu_spinwait(); 388 } 389 if (i != rowner_loops) 390 continue; 391 } 392#endif 393 394 /* 395 * Okay, now it's the hard case. Some other thread already 396 * has a write lock or there are write waiters present, 397 * acquire the turnstile lock so we can begin the process 398 * of blocking. 399 */ 400 ts = turnstile_trywait(&rw->lock_object); 401 402 /* 403 * The lock might have been released while we spun, so 404 * recheck its state and restart the loop if needed. 405 */ 406 v = rw->rw_lock; 407 if (RW_CAN_READ(v)) { 408 turnstile_cancel(ts); 409 cpu_spinwait(); 410 continue; 411 } 412 413#ifdef ADAPTIVE_RWLOCKS 414 /* 415 * If the current owner of the lock is executing on another 416 * CPU quit the hard path and try to spin. 417 */ 418 if ((v & RW_LOCK_READ) == 0) { 419 owner = (struct thread *)RW_OWNER(v); 420 if (TD_IS_RUNNING(owner)) { 421 turnstile_cancel(ts); 422 cpu_spinwait(); 423 continue; 424 } 425 } 426#endif 427 428 /* 429 * The lock is held in write mode or it already has waiters. 430 */ 431 MPASS(!RW_CAN_READ(v)); 432 433 /* 434 * If the RW_LOCK_READ_WAITERS flag is already set, then 435 * we can go ahead and block. If it is not set then try 436 * to set it. If we fail to set it drop the turnstile 437 * lock and restart the loop. 438 */ 439 if (!(v & RW_LOCK_READ_WAITERS)) { 440 if (!atomic_cmpset_ptr(&rw->rw_lock, v, 441 v | RW_LOCK_READ_WAITERS)) { 442 turnstile_cancel(ts); 443 cpu_spinwait(); 444 continue; 445 } 446 if (LOCK_LOG_TEST(&rw->lock_object, 0)) 447 CTR2(KTR_LOCK, "%s: %p set read waiters flag", 448 __func__, rw); 449 } 450 451 /* 452 * We were unable to acquire the lock and the read waiters 453 * flag is set, so we must block on the turnstile. 454 */ 455 if (LOCK_LOG_TEST(&rw->lock_object, 0)) 456 CTR2(KTR_LOCK, "%s: %p blocking on turnstile", __func__, 457 rw); 458#ifdef KDTRACE_HOOKS 459 sleep_time -= lockstat_nsecs(); 460#endif 461 turnstile_wait(ts, rw_owner(rw), TS_SHARED_QUEUE); 462#ifdef KDTRACE_HOOKS 463 sleep_time += lockstat_nsecs(); 464 sleep_cnt++; 465#endif 466 if (LOCK_LOG_TEST(&rw->lock_object, 0)) 467 CTR2(KTR_LOCK, "%s: %p resuming from turnstile", 468 __func__, rw); 469 } 470 471 /* 472 * TODO: acquire "owner of record" here. Here be turnstile dragons 473 * however. turnstiles don't like owners changing between calls to 474 * turnstile_wait() currently. 475 */ 476 LOCKSTAT_PROFILE_OBTAIN_LOCK_SUCCESS(LS_RW_RLOCK_ACQUIRE, rw, contested, 477 waittime, file, line); 478 LOCK_LOG_LOCK("RLOCK", &rw->lock_object, 0, 0, file, line); 479 WITNESS_LOCK(&rw->lock_object, 0, file, line); 480 curthread->td_locks++; 481 curthread->td_rw_rlocks++; 482#ifdef KDTRACE_HOOKS 483 if (sleep_time) 484 LOCKSTAT_RECORD1(LS_RW_RLOCK_BLOCK, rw, sleep_time); 485 486 /* 487 * Record only the loops spinning and not sleeping. 488 */ 489 if (spin_cnt > sleep_cnt) 490 LOCKSTAT_RECORD1(LS_RW_RLOCK_SPIN, rw, (spin_cnt - sleep_cnt)); 491#endif 492} 493 494int 495_rw_try_rlock(struct rwlock *rw, const char *file, int line) 496{ 497 uintptr_t x; 498 499 for (;;) { 500 x = rw->rw_lock; 501 KASSERT(rw->rw_lock != RW_DESTROYED, 502 ("rw_try_rlock() of destroyed rwlock @ %s:%d", file, line)); 503 if (!(x & RW_LOCK_READ)) 504 break; 505 if (atomic_cmpset_acq_ptr(&rw->rw_lock, x, x + RW_ONE_READER)) { 506 LOCK_LOG_TRY("RLOCK", &rw->lock_object, 0, 1, file, 507 line); 508 WITNESS_LOCK(&rw->lock_object, LOP_TRYLOCK, file, line); 509 curthread->td_locks++; 510 curthread->td_rw_rlocks++; 511 return (1); 512 } 513 } 514 515 LOCK_LOG_TRY("RLOCK", &rw->lock_object, 0, 0, file, line); 516 return (0); 517} 518 519void 520_rw_runlock(struct rwlock *rw, const char *file, int line) 521{ 522 struct turnstile *ts; 523 uintptr_t x, v, queue; 524 525 KASSERT(rw->rw_lock != RW_DESTROYED, 526 ("rw_runlock() of destroyed rwlock @ %s:%d", file, line)); 527 _rw_assert(rw, RA_RLOCKED, file, line); 528 curthread->td_locks--; 529 curthread->td_rw_rlocks--; 530 WITNESS_UNLOCK(&rw->lock_object, 0, file, line); 531 LOCK_LOG_LOCK("RUNLOCK", &rw->lock_object, 0, 0, file, line); 532 533 /* TODO: drop "owner of record" here. */ 534 535 for (;;) { 536 /* 537 * See if there is more than one read lock held. If so, 538 * just drop one and return. 539 */ 540 x = rw->rw_lock; 541 if (RW_READERS(x) > 1) { 542 if (atomic_cmpset_ptr(&rw->rw_lock, x, 543 x - RW_ONE_READER)) { 544 if (LOCK_LOG_TEST(&rw->lock_object, 0)) 545 CTR4(KTR_LOCK, 546 "%s: %p succeeded %p -> %p", 547 __func__, rw, (void *)x, 548 (void *)(x - RW_ONE_READER)); 549 break; 550 } 551 continue; 552 } 553 /* 554 * If there aren't any waiters for a write lock, then try 555 * to drop it quickly. 556 */ 557 if (!(x & RW_LOCK_WAITERS)) { 558 MPASS((x & ~RW_LOCK_WRITE_SPINNER) == 559 RW_READERS_LOCK(1)); 560 if (atomic_cmpset_ptr(&rw->rw_lock, x, RW_UNLOCKED)) { 561 if (LOCK_LOG_TEST(&rw->lock_object, 0)) 562 CTR2(KTR_LOCK, "%s: %p last succeeded", 563 __func__, rw); 564 break; 565 } 566 continue; 567 } 568 /* 569 * Ok, we know we have waiters and we think we are the 570 * last reader, so grab the turnstile lock. 571 */ 572 turnstile_chain_lock(&rw->lock_object); 573 v = rw->rw_lock & (RW_LOCK_WAITERS | RW_LOCK_WRITE_SPINNER); 574 MPASS(v & RW_LOCK_WAITERS); 575 576 /* 577 * Try to drop our lock leaving the lock in a unlocked 578 * state. 579 * 580 * If you wanted to do explicit lock handoff you'd have to 581 * do it here. You'd also want to use turnstile_signal() 582 * and you'd have to handle the race where a higher 583 * priority thread blocks on the write lock before the 584 * thread you wakeup actually runs and have the new thread 585 * "steal" the lock. For now it's a lot simpler to just 586 * wakeup all of the waiters. 587 * 588 * As above, if we fail, then another thread might have 589 * acquired a read lock, so drop the turnstile lock and 590 * restart. 591 */ 592 x = RW_UNLOCKED; 593 if (v & RW_LOCK_WRITE_WAITERS) { 594 queue = TS_EXCLUSIVE_QUEUE; 595 x |= (v & RW_LOCK_READ_WAITERS); 596 } else 597 queue = TS_SHARED_QUEUE; 598 if (!atomic_cmpset_ptr(&rw->rw_lock, RW_READERS_LOCK(1) | v, 599 x)) { 600 turnstile_chain_unlock(&rw->lock_object); 601 continue; 602 } 603 if (LOCK_LOG_TEST(&rw->lock_object, 0)) 604 CTR2(KTR_LOCK, "%s: %p last succeeded with waiters", 605 __func__, rw); 606 607 /* 608 * Ok. The lock is released and all that's left is to 609 * wake up the waiters. Note that the lock might not be 610 * free anymore, but in that case the writers will just 611 * block again if they run before the new lock holder(s) 612 * release the lock. 613 */ 614 ts = turnstile_lookup(&rw->lock_object); 615 MPASS(ts != NULL); 616 turnstile_broadcast(ts, queue); 617 turnstile_unpend(ts, TS_SHARED_LOCK); 618 turnstile_chain_unlock(&rw->lock_object); 619 break; 620 } 621 LOCKSTAT_PROFILE_RELEASE_LOCK(LS_RW_RUNLOCK_RELEASE, rw); 622} 623 624/* 625 * This function is called when we are unable to obtain a write lock on the 626 * first try. This means that at least one other thread holds either a 627 * read or write lock. 628 */ 629void 630_rw_wlock_hard(struct rwlock *rw, uintptr_t tid, const char *file, int line) 631{ 632 struct turnstile *ts; 633#ifdef ADAPTIVE_RWLOCKS 634 volatile struct thread *owner; 635 int spintries = 0; 636 int i; 637#endif 638 uintptr_t v, x; 639#ifdef LOCK_PROFILING 640 uint64_t waittime = 0; 641 int contested = 0; 642#endif 643#ifdef KDTRACE_HOOKS 644 uint64_t spin_cnt = 0; 645 uint64_t sleep_cnt = 0; 646 int64_t sleep_time = 0; 647#endif 648 649 if (rw_wlocked(rw)) { 650 KASSERT(rw->lock_object.lo_flags & RW_RECURSE, 651 ("%s: recursing but non-recursive rw %s @ %s:%d\n", 652 __func__, rw->lock_object.lo_name, file, line)); 653 rw->rw_recurse++; 654 if (LOCK_LOG_TEST(&rw->lock_object, 0)) 655 CTR2(KTR_LOCK, "%s: %p recursing", __func__, rw); 656 return; 657 } 658 659 if (LOCK_LOG_TEST(&rw->lock_object, 0)) 660 CTR5(KTR_LOCK, "%s: %s contested (lock=%p) at %s:%d", __func__, 661 rw->lock_object.lo_name, (void *)rw->rw_lock, file, line); 662 663 while (!_rw_write_lock(rw, tid)) { 664#ifdef KDTRACE_HOOKS 665 spin_cnt++; 666#endif 667 lock_profile_obtain_lock_failed(&rw->lock_object, 668 &contested, &waittime); 669#ifdef ADAPTIVE_RWLOCKS 670 /* 671 * If the lock is write locked and the owner is 672 * running on another CPU, spin until the owner stops 673 * running or the state of the lock changes. 674 */ 675 v = rw->rw_lock; 676 owner = (struct thread *)RW_OWNER(v); 677 if (!(v & RW_LOCK_READ) && TD_IS_RUNNING(owner)) { 678 if (LOCK_LOG_TEST(&rw->lock_object, 0)) 679 CTR3(KTR_LOCK, "%s: spinning on %p held by %p", 680 __func__, rw, owner); 681 while ((struct thread*)RW_OWNER(rw->rw_lock) == owner && 682 TD_IS_RUNNING(owner)) { 683 cpu_spinwait(); 684#ifdef KDTRACE_HOOKS 685 spin_cnt++; 686#endif 687 } 688 continue; 689 } 690 if ((v & RW_LOCK_READ) && RW_READERS(v) && 691 spintries < rowner_retries) { 692 if (!(v & RW_LOCK_WRITE_SPINNER)) { 693 if (!atomic_cmpset_ptr(&rw->rw_lock, v, 694 v | RW_LOCK_WRITE_SPINNER)) { 695 cpu_spinwait(); 696 continue; 697 } 698 } 699 spintries++; 700 for (i = 0; i < rowner_loops; i++) { 701 if ((rw->rw_lock & RW_LOCK_WRITE_SPINNER) == 0) 702 break; 703 cpu_spinwait(); 704 } 705#ifdef KDTRACE_HOOKS 706 spin_cnt += rowner_loops - i; 707#endif 708 if (i != rowner_loops) 709 continue; 710 } 711#endif 712 ts = turnstile_trywait(&rw->lock_object); 713 v = rw->rw_lock; 714 715#ifdef ADAPTIVE_RWLOCKS 716 /* 717 * If the current owner of the lock is executing on another 718 * CPU quit the hard path and try to spin. 719 */ 720 if (!(v & RW_LOCK_READ)) { 721 owner = (struct thread *)RW_OWNER(v); 722 if (TD_IS_RUNNING(owner)) { 723 turnstile_cancel(ts); 724 cpu_spinwait(); 725 continue; 726 } 727 } 728#endif 729 /* 730 * Check for the waiters flags about this rwlock. 731 * If the lock was released, without maintain any pending 732 * waiters queue, simply try to acquire it. 733 * If a pending waiters queue is present, claim the lock 734 * ownership and maintain the pending queue. 735 */ 736 x = v & (RW_LOCK_WAITERS | RW_LOCK_WRITE_SPINNER); 737 if ((v & ~x) == RW_UNLOCKED) { 738 x &= ~RW_LOCK_WRITE_SPINNER; 739 if (atomic_cmpset_acq_ptr(&rw->rw_lock, v, tid | x)) { 740 if (x) 741 turnstile_claim(ts); 742 else 743 turnstile_cancel(ts); 744 break; 745 } 746 turnstile_cancel(ts); 747 cpu_spinwait(); 748 continue; 749 } 750 /* 751 * If the RW_LOCK_WRITE_WAITERS flag isn't set, then try to 752 * set it. If we fail to set it, then loop back and try 753 * again. 754 */ 755 if (!(v & RW_LOCK_WRITE_WAITERS)) { 756 if (!atomic_cmpset_ptr(&rw->rw_lock, v, 757 v | RW_LOCK_WRITE_WAITERS)) { 758 turnstile_cancel(ts); 759 cpu_spinwait(); 760 continue; 761 } 762 if (LOCK_LOG_TEST(&rw->lock_object, 0)) 763 CTR2(KTR_LOCK, "%s: %p set write waiters flag", 764 __func__, rw); 765 } 766 /* 767 * We were unable to acquire the lock and the write waiters 768 * flag is set, so we must block on the turnstile. 769 */ 770 if (LOCK_LOG_TEST(&rw->lock_object, 0)) 771 CTR2(KTR_LOCK, "%s: %p blocking on turnstile", __func__, 772 rw); 773#ifdef KDTRACE_HOOKS 774 sleep_time -= lockstat_nsecs(); 775#endif 776 turnstile_wait(ts, rw_owner(rw), TS_EXCLUSIVE_QUEUE); 777#ifdef KDTRACE_HOOKS 778 sleep_time += lockstat_nsecs(); 779 sleep_cnt++; 780#endif 781 if (LOCK_LOG_TEST(&rw->lock_object, 0)) 782 CTR2(KTR_LOCK, "%s: %p resuming from turnstile", 783 __func__, rw); 784#ifdef ADAPTIVE_RWLOCKS 785 spintries = 0; 786#endif 787 } 788 LOCKSTAT_PROFILE_OBTAIN_LOCK_SUCCESS(LS_RW_WLOCK_ACQUIRE, rw, contested, 789 waittime, file, line); 790#ifdef KDTRACE_HOOKS 791 if (sleep_time) 792 LOCKSTAT_RECORD1(LS_RW_WLOCK_BLOCK, rw, sleep_time); 793 794 /* 795 * Record only the loops spinning and not sleeping. 796 */ 797 if (spin_cnt > sleep_cnt) 798 LOCKSTAT_RECORD1(LS_RW_WLOCK_SPIN, rw, (spin_cnt - sleep_cnt)); 799#endif 800} 801 802/* 803 * This function is called if the first try at releasing a write lock failed. 804 * This means that one of the 2 waiter bits must be set indicating that at 805 * least one thread is waiting on this lock. 806 */ 807void 808_rw_wunlock_hard(struct rwlock *rw, uintptr_t tid, const char *file, int line) 809{ 810 struct turnstile *ts; 811 uintptr_t v; 812 int queue; 813 814 if (rw_wlocked(rw) && rw_recursed(rw)) { 815 rw->rw_recurse--; 816 if (LOCK_LOG_TEST(&rw->lock_object, 0)) 817 CTR2(KTR_LOCK, "%s: %p unrecursing", __func__, rw); 818 return; 819 } 820 821 KASSERT(rw->rw_lock & (RW_LOCK_READ_WAITERS | RW_LOCK_WRITE_WAITERS), 822 ("%s: neither of the waiter flags are set", __func__)); 823 824 if (LOCK_LOG_TEST(&rw->lock_object, 0)) 825 CTR2(KTR_LOCK, "%s: %p contested", __func__, rw); 826 827 turnstile_chain_lock(&rw->lock_object); 828 ts = turnstile_lookup(&rw->lock_object); 829 MPASS(ts != NULL); 830 831 /* 832 * Use the same algo as sx locks for now. Prefer waking up shared 833 * waiters if we have any over writers. This is probably not ideal. 834 * 835 * 'v' is the value we are going to write back to rw_lock. If we 836 * have waiters on both queues, we need to preserve the state of 837 * the waiter flag for the queue we don't wake up. For now this is 838 * hardcoded for the algorithm mentioned above. 839 * 840 * In the case of both readers and writers waiting we wakeup the 841 * readers but leave the RW_LOCK_WRITE_WAITERS flag set. If a 842 * new writer comes in before a reader it will claim the lock up 843 * above. There is probably a potential priority inversion in 844 * there that could be worked around either by waking both queues 845 * of waiters or doing some complicated lock handoff gymnastics. 846 */ 847 v = RW_UNLOCKED; 848 if (rw->rw_lock & RW_LOCK_WRITE_WAITERS) { 849 queue = TS_EXCLUSIVE_QUEUE; 850 v |= (rw->rw_lock & RW_LOCK_READ_WAITERS); 851 } else 852 queue = TS_SHARED_QUEUE; 853 854 /* Wake up all waiters for the specific queue. */ 855 if (LOCK_LOG_TEST(&rw->lock_object, 0)) 856 CTR3(KTR_LOCK, "%s: %p waking up %s waiters", __func__, rw, 857 queue == TS_SHARED_QUEUE ? "read" : "write"); 858 turnstile_broadcast(ts, queue); 859 atomic_store_rel_ptr(&rw->rw_lock, v); 860 turnstile_unpend(ts, TS_EXCLUSIVE_LOCK); 861 turnstile_chain_unlock(&rw->lock_object); 862} 863 864/* 865 * Attempt to do a non-blocking upgrade from a read lock to a write 866 * lock. This will only succeed if this thread holds a single read 867 * lock. Returns true if the upgrade succeeded and false otherwise. 868 */ 869int 870_rw_try_upgrade(struct rwlock *rw, const char *file, int line) 871{ 872 uintptr_t v, x, tid; 873 struct turnstile *ts; 874 int success; 875 876 KASSERT(rw->rw_lock != RW_DESTROYED, 877 ("rw_try_upgrade() of destroyed rwlock @ %s:%d", file, line)); 878 _rw_assert(rw, RA_RLOCKED, file, line); 879 880 /* 881 * Attempt to switch from one reader to a writer. If there 882 * are any write waiters, then we will have to lock the 883 * turnstile first to prevent races with another writer 884 * calling turnstile_wait() before we have claimed this 885 * turnstile. So, do the simple case of no waiters first. 886 */ 887 tid = (uintptr_t)curthread; 888 success = 0; 889 for (;;) { 890 v = rw->rw_lock; 891 if (RW_READERS(v) > 1) 892 break; 893 if (!(v & RW_LOCK_WAITERS)) { 894 success = atomic_cmpset_ptr(&rw->rw_lock, v, tid); 895 if (!success) 896 continue; 897 break; 898 } 899 900 /* 901 * Ok, we think we have waiters, so lock the turnstile. 902 */ 903 ts = turnstile_trywait(&rw->lock_object); 904 v = rw->rw_lock; 905 if (RW_READERS(v) > 1) { 906 turnstile_cancel(ts); 907 break; 908 } 909 /* 910 * Try to switch from one reader to a writer again. This time 911 * we honor the current state of the waiters flags. 912 * If we obtain the lock with the flags set, then claim 913 * ownership of the turnstile. 914 */ 915 x = rw->rw_lock & RW_LOCK_WAITERS; 916 success = atomic_cmpset_ptr(&rw->rw_lock, v, tid | x); 917 if (success) { 918 if (x) 919 turnstile_claim(ts); 920 else 921 turnstile_cancel(ts); 922 break; 923 } 924 turnstile_cancel(ts); 925 } 926 LOCK_LOG_TRY("WUPGRADE", &rw->lock_object, 0, success, file, line); 927 if (success) { 928 curthread->td_rw_rlocks--; 929 WITNESS_UPGRADE(&rw->lock_object, LOP_EXCLUSIVE | LOP_TRYLOCK, 930 file, line); 931 LOCKSTAT_RECORD0(LS_RW_TRYUPGRADE_UPGRADE, rw); 932 } 933 return (success); 934} 935 936/* 937 * Downgrade a write lock into a single read lock. 938 */ 939void 940_rw_downgrade(struct rwlock *rw, const char *file, int line) 941{ 942 struct turnstile *ts; 943 uintptr_t tid, v; 944 int rwait, wwait; 945 946 KASSERT(rw->rw_lock != RW_DESTROYED, 947 ("rw_downgrade() of destroyed rwlock @ %s:%d", file, line)); 948 _rw_assert(rw, RA_WLOCKED | RA_NOTRECURSED, file, line); 949#ifndef INVARIANTS 950 if (rw_recursed(rw)) 951 panic("downgrade of a recursed lock"); 952#endif 953 954 WITNESS_DOWNGRADE(&rw->lock_object, 0, file, line); 955 956 /* 957 * Convert from a writer to a single reader. First we handle 958 * the easy case with no waiters. If there are any waiters, we 959 * lock the turnstile and "disown" the lock. 960 */ 961 tid = (uintptr_t)curthread; 962 if (atomic_cmpset_rel_ptr(&rw->rw_lock, tid, RW_READERS_LOCK(1))) 963 goto out; 964 965 /* 966 * Ok, we think we have waiters, so lock the turnstile so we can 967 * read the waiter flags without any races. 968 */ 969 turnstile_chain_lock(&rw->lock_object); 970 v = rw->rw_lock & RW_LOCK_WAITERS; 971 rwait = v & RW_LOCK_READ_WAITERS; 972 wwait = v & RW_LOCK_WRITE_WAITERS; 973 MPASS(rwait | wwait); 974 975 /* 976 * Downgrade from a write lock while preserving waiters flag 977 * and give up ownership of the turnstile. 978 */ 979 ts = turnstile_lookup(&rw->lock_object); 980 MPASS(ts != NULL); 981 if (!wwait) 982 v &= ~RW_LOCK_READ_WAITERS; 983 atomic_store_rel_ptr(&rw->rw_lock, RW_READERS_LOCK(1) | v); 984 /* 985 * Wake other readers if there are no writers pending. Otherwise they 986 * won't be able to acquire the lock anyway. 987 */ 988 if (rwait && !wwait) { 989 turnstile_broadcast(ts, TS_SHARED_QUEUE); 990 turnstile_unpend(ts, TS_EXCLUSIVE_LOCK); 991 } else 992 turnstile_disown(ts); 993 turnstile_chain_unlock(&rw->lock_object); 994out: 995 curthread->td_rw_rlocks++; 996 LOCK_LOG_LOCK("WDOWNGRADE", &rw->lock_object, 0, 0, file, line); 997 LOCKSTAT_RECORD0(LS_RW_DOWNGRADE_DOWNGRADE, rw); 998} 999 1000#ifdef INVARIANT_SUPPORT 1001#ifndef INVARIANTS 1002#undef _rw_assert 1003#endif 1004 1005/* 1006 * In the non-WITNESS case, rw_assert() can only detect that at least 1007 * *some* thread owns an rlock, but it cannot guarantee that *this* 1008 * thread owns an rlock. 1009 */ 1010void 1011_rw_assert(struct rwlock *rw, int what, const char *file, int line) 1012{ 1013 1014 if (panicstr != NULL) 1015 return; 1016 switch (what) { 1017 case RA_LOCKED: 1018 case RA_LOCKED | RA_RECURSED: 1019 case RA_LOCKED | RA_NOTRECURSED: 1020 case RA_RLOCKED: 1021#ifdef WITNESS 1022 witness_assert(&rw->lock_object, what, file, line); 1023#else 1024 /* 1025 * If some other thread has a write lock or we have one 1026 * and are asserting a read lock, fail. Also, if no one 1027 * has a lock at all, fail. 1028 */ 1029 if (rw->rw_lock == RW_UNLOCKED || 1030 (!(rw->rw_lock & RW_LOCK_READ) && (what == RA_RLOCKED || 1031 rw_wowner(rw) != curthread))) 1032 panic("Lock %s not %slocked @ %s:%d\n", 1033 rw->lock_object.lo_name, (what == RA_RLOCKED) ? 1034 "read " : "", file, line); 1035 1036 if (!(rw->rw_lock & RW_LOCK_READ)) { 1037 if (rw_recursed(rw)) { 1038 if (what & RA_NOTRECURSED) 1039 panic("Lock %s recursed @ %s:%d\n", 1040 rw->lock_object.lo_name, file, 1041 line); 1042 } else if (what & RA_RECURSED) 1043 panic("Lock %s not recursed @ %s:%d\n", 1044 rw->lock_object.lo_name, file, line); 1045 } 1046#endif 1047 break; 1048 case RA_WLOCKED: 1049 case RA_WLOCKED | RA_RECURSED: 1050 case RA_WLOCKED | RA_NOTRECURSED: 1051 if (rw_wowner(rw) != curthread) 1052 panic("Lock %s not exclusively locked @ %s:%d\n", 1053 rw->lock_object.lo_name, file, line); 1054 if (rw_recursed(rw)) { 1055 if (what & RA_NOTRECURSED) 1056 panic("Lock %s recursed @ %s:%d\n", 1057 rw->lock_object.lo_name, file, line); 1058 } else if (what & RA_RECURSED) 1059 panic("Lock %s not recursed @ %s:%d\n", 1060 rw->lock_object.lo_name, file, line); 1061 break; 1062 case RA_UNLOCKED: 1063#ifdef WITNESS 1064 witness_assert(&rw->lock_object, what, file, line); 1065#else 1066 /* 1067 * If we hold a write lock fail. We can't reliably check 1068 * to see if we hold a read lock or not. 1069 */ 1070 if (rw_wowner(rw) == curthread) 1071 panic("Lock %s exclusively locked @ %s:%d\n", 1072 rw->lock_object.lo_name, file, line); 1073#endif 1074 break; 1075 default: 1076 panic("Unknown rw lock assertion: %d @ %s:%d", what, file, 1077 line); 1078 } 1079} 1080#endif /* INVARIANT_SUPPORT */ 1081 1082#ifdef DDB 1083void 1084db_show_rwlock(struct lock_object *lock) 1085{ 1086 struct rwlock *rw; 1087 struct thread *td; 1088 1089 rw = (struct rwlock *)lock; 1090 1091 db_printf(" state: "); 1092 if (rw->rw_lock == RW_UNLOCKED) 1093 db_printf("UNLOCKED\n"); 1094 else if (rw->rw_lock == RW_DESTROYED) { 1095 db_printf("DESTROYED\n"); 1096 return; 1097 } else if (rw->rw_lock & RW_LOCK_READ) 1098 db_printf("RLOCK: %ju locks\n", 1099 (uintmax_t)(RW_READERS(rw->rw_lock))); 1100 else { 1101 td = rw_wowner(rw); 1102 db_printf("WLOCK: %p (tid %d, pid %d, \"%s\")\n", td, 1103 td->td_tid, td->td_proc->p_pid, td->td_name); 1104 if (rw_recursed(rw)) 1105 db_printf(" recursed: %u\n", rw->rw_recurse); 1106 } 1107 db_printf(" waiters: "); 1108 switch (rw->rw_lock & (RW_LOCK_READ_WAITERS | RW_LOCK_WRITE_WAITERS)) { 1109 case RW_LOCK_READ_WAITERS: 1110 db_printf("readers\n"); 1111 break; 1112 case RW_LOCK_WRITE_WAITERS: 1113 db_printf("writers\n"); 1114 break; 1115 case RW_LOCK_READ_WAITERS | RW_LOCK_WRITE_WAITERS: 1116 db_printf("readers and writers\n"); 1117 break; 1118 default: 1119 db_printf("none\n"); 1120 break; 1121 } 1122} 1123 1124#endif 1125