kern_rwlock.c revision 164159
1154941Sjhb/*- 2154941Sjhb * Copyright (c) 2006 John Baldwin <jhb@FreeBSD.org> 3154941Sjhb * All rights reserved. 4154941Sjhb * 5154941Sjhb * Redistribution and use in source and binary forms, with or without 6154941Sjhb * modification, are permitted provided that the following conditions 7154941Sjhb * are met: 8154941Sjhb * 1. Redistributions of source code must retain the above copyright 9154941Sjhb * notice, this list of conditions and the following disclaimer. 10154941Sjhb * 2. Redistributions in binary form must reproduce the above copyright 11154941Sjhb * notice, this list of conditions and the following disclaimer in the 12154941Sjhb * documentation and/or other materials provided with the distribution. 13154941Sjhb * 3. Neither the name of the author nor the names of any co-contributors 14154941Sjhb * may be used to endorse or promote products derived from this software 15154941Sjhb * without specific prior written permission. 16154941Sjhb * 17154941Sjhb * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND 18154941Sjhb * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 19154941Sjhb * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 20154941Sjhb * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE 21154941Sjhb * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 22154941Sjhb * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 23154941Sjhb * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 24154941Sjhb * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 25154941Sjhb * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 26154941Sjhb * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 27154941Sjhb * SUCH DAMAGE. 28154941Sjhb */ 29154941Sjhb 30154941Sjhb/* 31154941Sjhb * Machine independent bits of reader/writer lock implementation. 32154941Sjhb */ 33154941Sjhb 34154941Sjhb#include <sys/cdefs.h> 35154941Sjhb__FBSDID("$FreeBSD: head/sys/kern/kern_rwlock.c 164159 2006-11-11 03:18:07Z kmacy $"); 36154941Sjhb 37154941Sjhb#include "opt_ddb.h" 38154941Sjhb 39154941Sjhb#include <sys/param.h> 40154941Sjhb#include <sys/ktr.h> 41154941Sjhb#include <sys/lock.h> 42154941Sjhb#include <sys/mutex.h> 43154941Sjhb#include <sys/proc.h> 44154941Sjhb#include <sys/rwlock.h> 45154941Sjhb#include <sys/systm.h> 46154941Sjhb#include <sys/turnstile.h> 47164159Skmacy#include <sys/lock_profile.h> 48154941Sjhb#include <machine/cpu.h> 49154941Sjhb 50154941Sjhb#ifdef DDB 51154941Sjhb#include <ddb/ddb.h> 52154941Sjhb 53154941Sjhbstatic void db_show_rwlock(struct lock_object *lock); 54154941Sjhb#endif 55154941Sjhb 56154941Sjhbstruct lock_class lock_class_rw = { 57154941Sjhb "rw", 58157882Sjhb LC_SLEEPLOCK | LC_RECURSABLE | LC_UPGRADABLE, 59154941Sjhb#ifdef DDB 60154941Sjhb db_show_rwlock 61154941Sjhb#endif 62154941Sjhb}; 63154941Sjhb 64157826Sjhb/* 65157826Sjhb * Return a pointer to the owning thread if the lock is write-locked or 66157826Sjhb * NULL if the lock is unlocked or read-locked. 67157826Sjhb */ 68157826Sjhb#define rw_wowner(rw) \ 69154941Sjhb ((rw)->rw_lock & RW_LOCK_READ ? NULL : \ 70154941Sjhb (struct thread *)RW_OWNER((rw)->rw_lock)) 71154941Sjhb 72157826Sjhb/* 73157826Sjhb * Return a pointer to the owning thread for this lock who should receive 74157826Sjhb * any priority lent by threads that block on this lock. Currently this 75157826Sjhb * is identical to rw_wowner(). 76157826Sjhb */ 77157826Sjhb#define rw_owner(rw) rw_wowner(rw) 78157826Sjhb 79154941Sjhb#ifndef INVARIANTS 80154941Sjhb#define _rw_assert(rw, what, file, line) 81154941Sjhb#endif 82154941Sjhb 83154941Sjhbvoid 84154941Sjhbrw_init(struct rwlock *rw, const char *name) 85154941Sjhb{ 86154941Sjhb 87154941Sjhb rw->rw_lock = RW_UNLOCKED; 88154941Sjhb 89164159Skmacy lock_profile_object_init(&rw->rw_object, name); 90154941Sjhb lock_init(&rw->rw_object, &lock_class_rw, name, NULL, LO_WITNESS | 91157882Sjhb LO_RECURSABLE | LO_UPGRADABLE); 92154941Sjhb} 93154941Sjhb 94154941Sjhbvoid 95154941Sjhbrw_destroy(struct rwlock *rw) 96154941Sjhb{ 97154941Sjhb 98154941Sjhb KASSERT(rw->rw_lock == RW_UNLOCKED, ("rw lock not unlocked")); 99164159Skmacy lock_profile_object_destroy(&rw->rw_object); 100154941Sjhb lock_destroy(&rw->rw_object); 101154941Sjhb} 102154941Sjhb 103154941Sjhbvoid 104154941Sjhbrw_sysinit(void *arg) 105154941Sjhb{ 106154941Sjhb struct rw_args *args = arg; 107154941Sjhb 108154941Sjhb rw_init(args->ra_rw, args->ra_desc); 109154941Sjhb} 110154941Sjhb 111154941Sjhbvoid 112154941Sjhb_rw_wlock(struct rwlock *rw, const char *file, int line) 113154941Sjhb{ 114164159Skmacy uint64_t waitstart; 115154941Sjhb 116154941Sjhb MPASS(curthread != NULL); 117157826Sjhb KASSERT(rw_wowner(rw) != curthread, 118154941Sjhb ("%s (%s): wlock already held @ %s:%d", __func__, 119154941Sjhb rw->rw_object.lo_name, file, line)); 120154941Sjhb WITNESS_CHECKORDER(&rw->rw_object, LOP_NEWORDER | LOP_EXCLUSIVE, file, 121154941Sjhb line); 122164159Skmacy lock_profile_waitstart(&waitstart); 123154941Sjhb __rw_wlock(rw, curthread, file, line); 124164159Skmacy lock_profile_obtain_lock_success(&rw->rw_object, waitstart, file, line); 125154941Sjhb LOCK_LOG_LOCK("WLOCK", &rw->rw_object, 0, 0, file, line); 126154941Sjhb WITNESS_LOCK(&rw->rw_object, LOP_EXCLUSIVE, file, line); 127160771Sjhb curthread->td_locks++; 128154941Sjhb} 129154941Sjhb 130154941Sjhbvoid 131154941Sjhb_rw_wunlock(struct rwlock *rw, const char *file, int line) 132154941Sjhb{ 133154941Sjhb 134154941Sjhb MPASS(curthread != NULL); 135154941Sjhb _rw_assert(rw, RA_WLOCKED, file, line); 136160771Sjhb curthread->td_locks--; 137154941Sjhb WITNESS_UNLOCK(&rw->rw_object, LOP_EXCLUSIVE, file, line); 138154941Sjhb LOCK_LOG_LOCK("WUNLOCK", &rw->rw_object, 0, 0, file, line); 139164159Skmacy lock_profile_release_lock(&rw->rw_object); 140154941Sjhb __rw_wunlock(rw, curthread, file, line); 141154941Sjhb} 142154941Sjhb 143154941Sjhbvoid 144154941Sjhb_rw_rlock(struct rwlock *rw, const char *file, int line) 145154941Sjhb{ 146157851Swkoszek#ifdef SMP 147157846Sjhb volatile struct thread *owner; 148157851Swkoszek#endif 149164159Skmacy uint64_t waitstart; 150164159Skmacy int contested; 151154941Sjhb uintptr_t x; 152154941Sjhb 153157826Sjhb KASSERT(rw_wowner(rw) != curthread, 154154941Sjhb ("%s (%s): wlock already held @ %s:%d", __func__, 155154941Sjhb rw->rw_object.lo_name, file, line)); 156154941Sjhb WITNESS_CHECKORDER(&rw->rw_object, LOP_NEWORDER, file, line); 157154941Sjhb 158154941Sjhb /* 159154941Sjhb * Note that we don't make any attempt to try to block read 160154941Sjhb * locks once a writer has blocked on the lock. The reason is 161154941Sjhb * that we currently allow for read locks to recurse and we 162154941Sjhb * don't keep track of all the holders of read locks. Thus, if 163154941Sjhb * we were to block readers once a writer blocked and a reader 164154941Sjhb * tried to recurse on their reader lock after a writer had 165154941Sjhb * blocked we would end up in a deadlock since the reader would 166154941Sjhb * be blocked on the writer, and the writer would be blocked 167154941Sjhb * waiting for the reader to release its original read lock. 168154941Sjhb */ 169164159Skmacy lock_profile_waitstart(&waitstart); 170154941Sjhb for (;;) { 171154941Sjhb /* 172154941Sjhb * Handle the easy case. If no other thread has a write 173154941Sjhb * lock, then try to bump up the count of read locks. Note 174154941Sjhb * that we have to preserve the current state of the 175154941Sjhb * RW_LOCK_WRITE_WAITERS flag. If we fail to acquire a 176154941Sjhb * read lock, then rw_lock must have changed, so restart 177154941Sjhb * the loop. Note that this handles the case of a 178154941Sjhb * completely unlocked rwlock since such a lock is encoded 179154941Sjhb * as a read lock with no waiters. 180154941Sjhb */ 181154941Sjhb x = rw->rw_lock; 182154941Sjhb if (x & RW_LOCK_READ) { 183154941Sjhb 184154941Sjhb /* 185154941Sjhb * The RW_LOCK_READ_WAITERS flag should only be set 186154941Sjhb * if another thread currently holds a write lock, 187154941Sjhb * and in that case RW_LOCK_READ should be clear. 188154941Sjhb */ 189154941Sjhb MPASS((x & RW_LOCK_READ_WAITERS) == 0); 190154941Sjhb if (atomic_cmpset_acq_ptr(&rw->rw_lock, x, 191154941Sjhb x + RW_ONE_READER)) { 192164159Skmacy lock_profile_obtain_lock_success(&rw->rw_object, waitstart, file, line); 193154941Sjhb if (LOCK_LOG_TEST(&rw->rw_object, 0)) 194154941Sjhb CTR4(KTR_LOCK, 195154941Sjhb "%s: %p succeed %p -> %p", __func__, 196154941Sjhb rw, (void *)x, 197154941Sjhb (void *)(x + RW_ONE_READER)); 198154941Sjhb break; 199154941Sjhb } 200157846Sjhb cpu_spinwait(); 201164159Skmacy lock_profile_obtain_lock_failed(&rw->rw_object, &contested); 202154941Sjhb continue; 203154941Sjhb } 204154941Sjhb 205154941Sjhb /* 206154941Sjhb * Okay, now it's the hard case. Some other thread already 207154941Sjhb * has a write lock, so acquire the turnstile lock so we can 208154941Sjhb * begin the process of blocking. 209154941Sjhb */ 210154941Sjhb turnstile_lock(&rw->rw_object); 211154941Sjhb 212154941Sjhb /* 213154941Sjhb * The lock might have been released while we spun, so 214154941Sjhb * recheck its state and restart the loop if there is no 215154941Sjhb * longer a write lock. 216154941Sjhb */ 217154941Sjhb x = rw->rw_lock; 218154941Sjhb if (x & RW_LOCK_READ) { 219154941Sjhb turnstile_release(&rw->rw_object); 220157846Sjhb cpu_spinwait(); 221154941Sjhb continue; 222154941Sjhb } 223154941Sjhb 224154941Sjhb /* 225154941Sjhb * Ok, it's still a write lock. If the RW_LOCK_READ_WAITERS 226154941Sjhb * flag is already set, then we can go ahead and block. If 227154941Sjhb * it is not set then try to set it. If we fail to set it 228154941Sjhb * drop the turnstile lock and restart the loop. 229154941Sjhb */ 230157826Sjhb if (!(x & RW_LOCK_READ_WAITERS)) { 231157826Sjhb if (!atomic_cmpset_ptr(&rw->rw_lock, x, 232157826Sjhb x | RW_LOCK_READ_WAITERS)) { 233157826Sjhb turnstile_release(&rw->rw_object); 234157826Sjhb cpu_spinwait(); 235157826Sjhb continue; 236157826Sjhb } 237157826Sjhb if (LOCK_LOG_TEST(&rw->rw_object, 0)) 238157826Sjhb CTR2(KTR_LOCK, "%s: %p set read waiters flag", 239157826Sjhb __func__, rw); 240154941Sjhb } 241154941Sjhb 242157846Sjhb#ifdef SMP 243154941Sjhb /* 244157846Sjhb * If the owner is running on another CPU, spin until 245157846Sjhb * the owner stops running or the state of the lock 246157846Sjhb * changes. 247157846Sjhb */ 248157846Sjhb owner = (struct thread *)RW_OWNER(x); 249157846Sjhb if (TD_IS_RUNNING(owner)) { 250164159Skmacy lock_profile_obtain_lock_failed(&rw->rw_object, &contested); 251157846Sjhb turnstile_release(&rw->rw_object); 252157846Sjhb if (LOCK_LOG_TEST(&rw->rw_object, 0)) 253157846Sjhb CTR3(KTR_LOCK, "%s: spinning on %p held by %p", 254157846Sjhb __func__, rw, owner); 255157846Sjhb while ((struct thread*)RW_OWNER(rw->rw_lock)== owner && 256157846Sjhb TD_IS_RUNNING(owner)) 257157846Sjhb cpu_spinwait(); 258157846Sjhb continue; 259157846Sjhb } 260157846Sjhb#endif 261157846Sjhb 262157846Sjhb /* 263154941Sjhb * We were unable to acquire the lock and the read waiters 264154941Sjhb * flag is set, so we must block on the turnstile. 265154941Sjhb */ 266154941Sjhb if (LOCK_LOG_TEST(&rw->rw_object, 0)) 267154941Sjhb CTR2(KTR_LOCK, "%s: %p blocking on turnstile", __func__, 268154941Sjhb rw); 269154941Sjhb turnstile_wait(&rw->rw_object, rw_owner(rw), TS_SHARED_QUEUE); 270154941Sjhb if (LOCK_LOG_TEST(&rw->rw_object, 0)) 271154941Sjhb CTR2(KTR_LOCK, "%s: %p resuming from turnstile", 272154941Sjhb __func__, rw); 273154941Sjhb } 274154941Sjhb 275154941Sjhb /* 276154941Sjhb * TODO: acquire "owner of record" here. Here be turnstile dragons 277154941Sjhb * however. turnstiles don't like owners changing between calls to 278154941Sjhb * turnstile_wait() currently. 279154941Sjhb */ 280154941Sjhb 281154941Sjhb LOCK_LOG_LOCK("RLOCK", &rw->rw_object, 0, 0, file, line); 282154941Sjhb WITNESS_LOCK(&rw->rw_object, 0, file, line); 283160771Sjhb curthread->td_locks++; 284154941Sjhb} 285154941Sjhb 286154941Sjhbvoid 287154941Sjhb_rw_runlock(struct rwlock *rw, const char *file, int line) 288154941Sjhb{ 289154941Sjhb struct turnstile *ts; 290154941Sjhb uintptr_t x; 291154941Sjhb 292154941Sjhb _rw_assert(rw, RA_RLOCKED, file, line); 293160771Sjhb curthread->td_locks--; 294154941Sjhb WITNESS_UNLOCK(&rw->rw_object, 0, file, line); 295154941Sjhb LOCK_LOG_LOCK("RUNLOCK", &rw->rw_object, 0, 0, file, line); 296154941Sjhb 297154941Sjhb /* TODO: drop "owner of record" here. */ 298154941Sjhb 299154941Sjhb for (;;) { 300154941Sjhb /* 301154941Sjhb * See if there is more than one read lock held. If so, 302154941Sjhb * just drop one and return. 303154941Sjhb */ 304154941Sjhb x = rw->rw_lock; 305154941Sjhb if (RW_READERS(x) > 1) { 306154941Sjhb if (atomic_cmpset_ptr(&rw->rw_lock, x, 307154941Sjhb x - RW_ONE_READER)) { 308154941Sjhb if (LOCK_LOG_TEST(&rw->rw_object, 0)) 309154941Sjhb CTR4(KTR_LOCK, 310154941Sjhb "%s: %p succeeded %p -> %p", 311154941Sjhb __func__, rw, (void *)x, 312154941Sjhb (void *)(x - RW_ONE_READER)); 313154941Sjhb break; 314154941Sjhb } 315154941Sjhb continue; 316164159Skmacy } else 317164159Skmacy lock_profile_release_lock(&rw->rw_object); 318154941Sjhb 319164159Skmacy 320154941Sjhb /* 321154941Sjhb * We should never have read waiters while at least one 322154941Sjhb * thread holds a read lock. (See note above) 323154941Sjhb */ 324154941Sjhb KASSERT(!(x & RW_LOCK_READ_WAITERS), 325154941Sjhb ("%s: waiting readers", __func__)); 326154941Sjhb 327154941Sjhb /* 328154941Sjhb * If there aren't any waiters for a write lock, then try 329154941Sjhb * to drop it quickly. 330154941Sjhb */ 331154941Sjhb if (!(x & RW_LOCK_WRITE_WAITERS)) { 332154941Sjhb 333154941Sjhb /* 334154941Sjhb * There shouldn't be any flags set and we should 335154941Sjhb * be the only read lock. If we fail to release 336154941Sjhb * the single read lock, then another thread might 337154941Sjhb * have just acquired a read lock, so go back up 338154941Sjhb * to the multiple read locks case. 339154941Sjhb */ 340154941Sjhb MPASS(x == RW_READERS_LOCK(1)); 341154941Sjhb if (atomic_cmpset_ptr(&rw->rw_lock, RW_READERS_LOCK(1), 342154941Sjhb RW_UNLOCKED)) { 343154941Sjhb if (LOCK_LOG_TEST(&rw->rw_object, 0)) 344154941Sjhb CTR2(KTR_LOCK, "%s: %p last succeeded", 345154941Sjhb __func__, rw); 346154941Sjhb break; 347154941Sjhb } 348154941Sjhb continue; 349154941Sjhb } 350154941Sjhb 351154941Sjhb /* 352154941Sjhb * There should just be one reader with one or more 353154941Sjhb * writers waiting. 354154941Sjhb */ 355154941Sjhb MPASS(x == (RW_READERS_LOCK(1) | RW_LOCK_WRITE_WAITERS)); 356154941Sjhb 357154941Sjhb /* 358154941Sjhb * Ok, we know we have a waiting writer and we think we 359154941Sjhb * are the last reader, so grab the turnstile lock. 360154941Sjhb */ 361154941Sjhb turnstile_lock(&rw->rw_object); 362154941Sjhb 363154941Sjhb /* 364154941Sjhb * Try to drop our lock leaving the lock in a unlocked 365154941Sjhb * state. 366154941Sjhb * 367154941Sjhb * If you wanted to do explicit lock handoff you'd have to 368154941Sjhb * do it here. You'd also want to use turnstile_signal() 369154941Sjhb * and you'd have to handle the race where a higher 370154941Sjhb * priority thread blocks on the write lock before the 371154941Sjhb * thread you wakeup actually runs and have the new thread 372154941Sjhb * "steal" the lock. For now it's a lot simpler to just 373154941Sjhb * wakeup all of the waiters. 374154941Sjhb * 375154941Sjhb * As above, if we fail, then another thread might have 376154941Sjhb * acquired a read lock, so drop the turnstile lock and 377154941Sjhb * restart. 378154941Sjhb */ 379154941Sjhb if (!atomic_cmpset_ptr(&rw->rw_lock, 380154941Sjhb RW_READERS_LOCK(1) | RW_LOCK_WRITE_WAITERS, RW_UNLOCKED)) { 381154941Sjhb turnstile_release(&rw->rw_object); 382154941Sjhb continue; 383154941Sjhb } 384154941Sjhb if (LOCK_LOG_TEST(&rw->rw_object, 0)) 385154941Sjhb CTR2(KTR_LOCK, "%s: %p last succeeded with waiters", 386154941Sjhb __func__, rw); 387154941Sjhb 388154941Sjhb /* 389154941Sjhb * Ok. The lock is released and all that's left is to 390154941Sjhb * wake up the waiters. Note that the lock might not be 391154941Sjhb * free anymore, but in that case the writers will just 392154941Sjhb * block again if they run before the new lock holder(s) 393154941Sjhb * release the lock. 394154941Sjhb */ 395154941Sjhb ts = turnstile_lookup(&rw->rw_object); 396157846Sjhb MPASS(ts != NULL); 397154941Sjhb turnstile_broadcast(ts, TS_EXCLUSIVE_QUEUE); 398154941Sjhb turnstile_unpend(ts, TS_SHARED_LOCK); 399154941Sjhb break; 400154941Sjhb } 401154941Sjhb} 402154941Sjhb 403154941Sjhb/* 404154941Sjhb * This function is called when we are unable to obtain a write lock on the 405154941Sjhb * first try. This means that at least one other thread holds either a 406154941Sjhb * read or write lock. 407154941Sjhb */ 408154941Sjhbvoid 409154941Sjhb_rw_wlock_hard(struct rwlock *rw, uintptr_t tid, const char *file, int line) 410154941Sjhb{ 411157851Swkoszek#ifdef SMP 412157846Sjhb volatile struct thread *owner; 413157851Swkoszek#endif 414164159Skmacy int contested; 415154941Sjhb uintptr_t v; 416154941Sjhb 417154941Sjhb if (LOCK_LOG_TEST(&rw->rw_object, 0)) 418154941Sjhb CTR5(KTR_LOCK, "%s: %s contested (lock=%p) at %s:%d", __func__, 419154941Sjhb rw->rw_object.lo_name, (void *)rw->rw_lock, file, line); 420154941Sjhb 421154941Sjhb while (!_rw_write_lock(rw, tid)) { 422154941Sjhb turnstile_lock(&rw->rw_object); 423154941Sjhb v = rw->rw_lock; 424154941Sjhb 425154941Sjhb /* 426154941Sjhb * If the lock was released while spinning on the 427154941Sjhb * turnstile chain lock, try again. 428154941Sjhb */ 429154941Sjhb if (v == RW_UNLOCKED) { 430154941Sjhb turnstile_release(&rw->rw_object); 431154941Sjhb cpu_spinwait(); 432154941Sjhb continue; 433154941Sjhb } 434154941Sjhb 435154941Sjhb /* 436154941Sjhb * If the lock was released by a writer with both readers 437154941Sjhb * and writers waiting and a reader hasn't woken up and 438154941Sjhb * acquired the lock yet, rw_lock will be set to the 439154941Sjhb * value RW_UNLOCKED | RW_LOCK_WRITE_WAITERS. If we see 440154941Sjhb * that value, try to acquire it once. Note that we have 441154941Sjhb * to preserve the RW_LOCK_WRITE_WAITERS flag as there are 442154941Sjhb * other writers waiting still. If we fail, restart the 443154941Sjhb * loop. 444154941Sjhb */ 445154941Sjhb if (v == (RW_UNLOCKED | RW_LOCK_WRITE_WAITERS)) { 446154941Sjhb if (atomic_cmpset_acq_ptr(&rw->rw_lock, 447154941Sjhb RW_UNLOCKED | RW_LOCK_WRITE_WAITERS, 448154941Sjhb tid | RW_LOCK_WRITE_WAITERS)) { 449154941Sjhb turnstile_claim(&rw->rw_object); 450154941Sjhb CTR2(KTR_LOCK, "%s: %p claimed by new writer", 451154941Sjhb __func__, rw); 452154941Sjhb break; 453154941Sjhb } 454154941Sjhb turnstile_release(&rw->rw_object); 455154941Sjhb cpu_spinwait(); 456164159Skmacy lock_profile_obtain_lock_failed(&rw->rw_object, &contested); 457154941Sjhb continue; 458154941Sjhb } 459154941Sjhb 460154941Sjhb /* 461154941Sjhb * If the RW_LOCK_WRITE_WAITERS flag isn't set, then try to 462154941Sjhb * set it. If we fail to set it, then loop back and try 463154941Sjhb * again. 464154941Sjhb */ 465157826Sjhb if (!(v & RW_LOCK_WRITE_WAITERS)) { 466157826Sjhb if (!atomic_cmpset_ptr(&rw->rw_lock, v, 467157826Sjhb v | RW_LOCK_WRITE_WAITERS)) { 468157826Sjhb turnstile_release(&rw->rw_object); 469157826Sjhb cpu_spinwait(); 470164159Skmacy lock_profile_obtain_lock_failed(&rw->rw_object, &contested); 471157826Sjhb continue; 472157826Sjhb } 473157826Sjhb if (LOCK_LOG_TEST(&rw->rw_object, 0)) 474157826Sjhb CTR2(KTR_LOCK, "%s: %p set write waiters flag", 475157826Sjhb __func__, rw); 476154941Sjhb } 477154941Sjhb 478157846Sjhb#ifdef SMP 479157846Sjhb /* 480157846Sjhb * If the lock is write locked and the owner is 481157846Sjhb * running on another CPU, spin until the owner stops 482157846Sjhb * running or the state of the lock changes. 483157846Sjhb */ 484157846Sjhb owner = (struct thread *)RW_OWNER(v); 485157846Sjhb if (!(v & RW_LOCK_READ) && TD_IS_RUNNING(owner)) { 486164159Skmacy lock_profile_obtain_lock_failed(&rw->rw_object, &contested); 487157846Sjhb turnstile_release(&rw->rw_object); 488157846Sjhb if (LOCK_LOG_TEST(&rw->rw_object, 0)) 489157846Sjhb CTR3(KTR_LOCK, "%s: spinning on %p held by %p", 490157846Sjhb __func__, rw, owner); 491157846Sjhb while ((struct thread*)RW_OWNER(rw->rw_lock)== owner && 492157846Sjhb TD_IS_RUNNING(owner)) 493157846Sjhb cpu_spinwait(); 494157846Sjhb continue; 495157846Sjhb } 496157846Sjhb#endif 497154941Sjhb 498154941Sjhb /* 499154941Sjhb * We were unable to acquire the lock and the write waiters 500154941Sjhb * flag is set, so we must block on the turnstile. 501154941Sjhb */ 502154941Sjhb if (LOCK_LOG_TEST(&rw->rw_object, 0)) 503154941Sjhb CTR2(KTR_LOCK, "%s: %p blocking on turnstile", __func__, 504154941Sjhb rw); 505154941Sjhb turnstile_wait(&rw->rw_object, rw_owner(rw), 506154941Sjhb TS_EXCLUSIVE_QUEUE); 507154941Sjhb if (LOCK_LOG_TEST(&rw->rw_object, 0)) 508154941Sjhb CTR2(KTR_LOCK, "%s: %p resuming from turnstile", 509154941Sjhb __func__, rw); 510154941Sjhb } 511154941Sjhb} 512154941Sjhb 513154941Sjhb/* 514154941Sjhb * This function is called if the first try at releasing a write lock failed. 515154941Sjhb * This means that one of the 2 waiter bits must be set indicating that at 516154941Sjhb * least one thread is waiting on this lock. 517154941Sjhb */ 518154941Sjhbvoid 519154941Sjhb_rw_wunlock_hard(struct rwlock *rw, uintptr_t tid, const char *file, int line) 520154941Sjhb{ 521154941Sjhb struct turnstile *ts; 522154941Sjhb uintptr_t v; 523154941Sjhb int queue; 524154941Sjhb 525154941Sjhb KASSERT(rw->rw_lock & (RW_LOCK_READ_WAITERS | RW_LOCK_WRITE_WAITERS), 526154941Sjhb ("%s: neither of the waiter flags are set", __func__)); 527154941Sjhb 528154941Sjhb if (LOCK_LOG_TEST(&rw->rw_object, 0)) 529154941Sjhb CTR2(KTR_LOCK, "%s: %p contested", __func__, rw); 530154941Sjhb 531154941Sjhb turnstile_lock(&rw->rw_object); 532154941Sjhb ts = turnstile_lookup(&rw->rw_object); 533154941Sjhb 534157846Sjhb#ifdef SMP 535157846Sjhb /* 536157846Sjhb * There might not be a turnstile for this lock if all of 537157846Sjhb * the waiters are adaptively spinning. In that case, just 538157846Sjhb * reset the lock to the unlocked state and return. 539157846Sjhb */ 540157846Sjhb if (ts == NULL) { 541157846Sjhb atomic_store_rel_ptr(&rw->rw_lock, RW_UNLOCKED); 542157846Sjhb if (LOCK_LOG_TEST(&rw->rw_object, 0)) 543157846Sjhb CTR2(KTR_LOCK, "%s: %p no sleepers", __func__, rw); 544157846Sjhb turnstile_release(&rw->rw_object); 545157846Sjhb return; 546157846Sjhb } 547157846Sjhb#else 548154941Sjhb MPASS(ts != NULL); 549157846Sjhb#endif 550154941Sjhb 551154941Sjhb /* 552154941Sjhb * Use the same algo as sx locks for now. Prefer waking up shared 553154941Sjhb * waiters if we have any over writers. This is probably not ideal. 554154941Sjhb * 555154941Sjhb * 'v' is the value we are going to write back to rw_lock. If we 556154941Sjhb * have waiters on both queues, we need to preserve the state of 557154941Sjhb * the waiter flag for the queue we don't wake up. For now this is 558154941Sjhb * hardcoded for the algorithm mentioned above. 559154941Sjhb * 560154941Sjhb * In the case of both readers and writers waiting we wakeup the 561154941Sjhb * readers but leave the RW_LOCK_WRITE_WAITERS flag set. If a 562154941Sjhb * new writer comes in before a reader it will claim the lock up 563154941Sjhb * above. There is probably a potential priority inversion in 564154941Sjhb * there that could be worked around either by waking both queues 565154941Sjhb * of waiters or doing some complicated lock handoff gymnastics. 566157846Sjhb * 567157846Sjhb * Note that in the SMP case, if both flags are set, there might 568157846Sjhb * not be any actual writers on the turnstile as they might all 569157846Sjhb * be spinning. In that case, we don't want to preserve the 570157846Sjhb * RW_LOCK_WRITE_WAITERS flag as the turnstile is going to go 571157846Sjhb * away once we wakeup all the readers. 572154941Sjhb */ 573157846Sjhb v = RW_UNLOCKED; 574154941Sjhb if (rw->rw_lock & RW_LOCK_READ_WAITERS) { 575154941Sjhb queue = TS_SHARED_QUEUE; 576157846Sjhb#ifdef SMP 577157846Sjhb if (rw->rw_lock & RW_LOCK_WRITE_WAITERS && 578157846Sjhb !turnstile_empty(ts, TS_EXCLUSIVE_QUEUE)) 579157846Sjhb v |= RW_LOCK_WRITE_WAITERS; 580157846Sjhb#else 581157846Sjhb v |= (rw->rw_lock & RW_LOCK_WRITE_WAITERS); 582157846Sjhb#endif 583157846Sjhb } else 584154941Sjhb queue = TS_EXCLUSIVE_QUEUE; 585157846Sjhb 586157846Sjhb#ifdef SMP 587157846Sjhb /* 588157846Sjhb * We have to make sure that we actually have waiters to 589157846Sjhb * wakeup. If they are all spinning, then we just need to 590157846Sjhb * disown the turnstile and return. 591157846Sjhb */ 592157846Sjhb if (turnstile_empty(ts, queue)) { 593157846Sjhb if (LOCK_LOG_TEST(&rw->rw_object, 0)) 594157846Sjhb CTR2(KTR_LOCK, "%s: %p no sleepers 2", __func__, rw); 595157846Sjhb atomic_store_rel_ptr(&rw->rw_lock, v); 596157846Sjhb turnstile_disown(ts); 597157846Sjhb return; 598154941Sjhb } 599157846Sjhb#endif 600157846Sjhb 601157846Sjhb /* Wake up all waiters for the specific queue. */ 602154941Sjhb if (LOCK_LOG_TEST(&rw->rw_object, 0)) 603154941Sjhb CTR3(KTR_LOCK, "%s: %p waking up %s waiters", __func__, rw, 604154941Sjhb queue == TS_SHARED_QUEUE ? "read" : "write"); 605154941Sjhb turnstile_broadcast(ts, queue); 606154941Sjhb atomic_store_rel_ptr(&rw->rw_lock, v); 607154941Sjhb turnstile_unpend(ts, TS_EXCLUSIVE_LOCK); 608154941Sjhb} 609154941Sjhb 610157882Sjhb/* 611157882Sjhb * Attempt to do a non-blocking upgrade from a read lock to a write 612157882Sjhb * lock. This will only succeed if this thread holds a single read 613157882Sjhb * lock. Returns true if the upgrade succeeded and false otherwise. 614157882Sjhb */ 615157882Sjhbint 616157882Sjhb_rw_try_upgrade(struct rwlock *rw, const char *file, int line) 617157882Sjhb{ 618157882Sjhb uintptr_t v, tid; 619157882Sjhb int success; 620157882Sjhb 621157882Sjhb _rw_assert(rw, RA_RLOCKED, file, line); 622157882Sjhb 623157882Sjhb /* 624157882Sjhb * Attempt to switch from one reader to a writer. If there 625157882Sjhb * are any write waiters, then we will have to lock the 626157882Sjhb * turnstile first to prevent races with another writer 627157882Sjhb * calling turnstile_wait() before we have claimed this 628157882Sjhb * turnstile. So, do the simple case of no waiters first. 629157882Sjhb */ 630157882Sjhb tid = (uintptr_t)curthread; 631157882Sjhb if (!(rw->rw_lock & RW_LOCK_WRITE_WAITERS)) { 632157882Sjhb success = atomic_cmpset_acq_ptr(&rw->rw_lock, 633157882Sjhb RW_READERS_LOCK(1), tid); 634157882Sjhb goto out; 635157882Sjhb } 636157882Sjhb 637157882Sjhb /* 638157882Sjhb * Ok, we think we have write waiters, so lock the 639157882Sjhb * turnstile. 640157882Sjhb */ 641157882Sjhb turnstile_lock(&rw->rw_object); 642157882Sjhb 643157882Sjhb /* 644157882Sjhb * Try to switch from one reader to a writer again. This time 645157882Sjhb * we honor the current state of the RW_LOCK_WRITE_WAITERS 646157882Sjhb * flag. If we obtain the lock with the flag set, then claim 647157882Sjhb * ownership of the turnstile. In the SMP case it is possible 648157882Sjhb * for there to not be an associated turnstile even though there 649157882Sjhb * are waiters if all of the waiters are spinning. 650157882Sjhb */ 651157882Sjhb v = rw->rw_lock & RW_LOCK_WRITE_WAITERS; 652157882Sjhb success = atomic_cmpset_acq_ptr(&rw->rw_lock, RW_READERS_LOCK(1) | v, 653157882Sjhb tid | v); 654157882Sjhb#ifdef SMP 655157882Sjhb if (success && v && turnstile_lookup(&rw->rw_object) != NULL) 656157882Sjhb#else 657157882Sjhb if (success && v) 658157882Sjhb#endif 659157882Sjhb turnstile_claim(&rw->rw_object); 660157882Sjhb else 661157882Sjhb turnstile_release(&rw->rw_object); 662157882Sjhbout: 663157882Sjhb LOCK_LOG_TRY("WUPGRADE", &rw->rw_object, 0, success, file, line); 664157882Sjhb if (success) 665157882Sjhb WITNESS_UPGRADE(&rw->rw_object, LOP_EXCLUSIVE | LOP_TRYLOCK, 666157882Sjhb file, line); 667157882Sjhb return (success); 668157882Sjhb} 669157882Sjhb 670157882Sjhb/* 671157882Sjhb * Downgrade a write lock into a single read lock. 672157882Sjhb */ 673157882Sjhbvoid 674157882Sjhb_rw_downgrade(struct rwlock *rw, const char *file, int line) 675157882Sjhb{ 676157882Sjhb struct turnstile *ts; 677157882Sjhb uintptr_t tid, v; 678157882Sjhb 679157882Sjhb _rw_assert(rw, RA_WLOCKED, file, line); 680157882Sjhb 681157882Sjhb WITNESS_DOWNGRADE(&rw->rw_object, 0, file, line); 682157882Sjhb 683157882Sjhb /* 684157882Sjhb * Convert from a writer to a single reader. First we handle 685157882Sjhb * the easy case with no waiters. If there are any waiters, we 686157882Sjhb * lock the turnstile, "disown" the lock, and awaken any read 687157882Sjhb * waiters. 688157882Sjhb */ 689157882Sjhb tid = (uintptr_t)curthread; 690157882Sjhb if (atomic_cmpset_rel_ptr(&rw->rw_lock, tid, RW_READERS_LOCK(1))) 691157882Sjhb goto out; 692157882Sjhb 693157882Sjhb /* 694157882Sjhb * Ok, we think we have waiters, so lock the turnstile so we can 695157882Sjhb * read the waiter flags without any races. 696157882Sjhb */ 697157882Sjhb turnstile_lock(&rw->rw_object); 698157882Sjhb v = rw->rw_lock; 699157882Sjhb MPASS(v & (RW_LOCK_READ_WAITERS | RW_LOCK_WRITE_WAITERS)); 700157882Sjhb 701157882Sjhb /* 702157882Sjhb * Downgrade from a write lock while preserving 703157882Sjhb * RW_LOCK_WRITE_WAITERS and give up ownership of the 704157882Sjhb * turnstile. If there are any read waiters, wake them up. 705157882Sjhb * 706157882Sjhb * For SMP, we have to allow for the fact that all of the 707157882Sjhb * read waiters might be spinning. In that case, act as if 708157882Sjhb * RW_LOCK_READ_WAITERS is not set. Also, only preserve 709157882Sjhb * the RW_LOCK_WRITE_WAITERS flag if at least one writer is 710157882Sjhb * blocked on the turnstile. 711157882Sjhb */ 712157882Sjhb ts = turnstile_lookup(&rw->rw_object); 713157882Sjhb#ifdef SMP 714157882Sjhb if (ts == NULL) 715157882Sjhb v &= ~(RW_LOCK_READ_WAITERS | RW_LOCK_WRITE_WAITERS); 716157882Sjhb else if (v & RW_LOCK_READ_WAITERS && 717157882Sjhb turnstile_empty(ts, TS_SHARED_QUEUE)) 718157882Sjhb v &= ~RW_LOCK_READ_WAITERS; 719157882Sjhb else if (v & RW_LOCK_WRITE_WAITERS && 720157882Sjhb turnstile_empty(ts, TS_EXCLUSIVE_QUEUE)) 721157882Sjhb v &= ~RW_LOCK_WRITE_WAITERS; 722157882Sjhb#else 723157882Sjhb MPASS(ts != NULL); 724157882Sjhb#endif 725157882Sjhb if (v & RW_LOCK_READ_WAITERS) 726157882Sjhb turnstile_broadcast(ts, TS_SHARED_QUEUE); 727157882Sjhb atomic_store_rel_ptr(&rw->rw_lock, RW_READERS_LOCK(1) | 728157882Sjhb (v & RW_LOCK_WRITE_WAITERS)); 729157882Sjhb if (v & RW_LOCK_READ_WAITERS) 730157882Sjhb turnstile_unpend(ts, TS_EXCLUSIVE_LOCK); 731157882Sjhb#ifdef SMP 732157882Sjhb else if (ts == NULL) 733157882Sjhb turnstile_release(&rw->rw_object); 734157882Sjhb#endif 735157882Sjhb else 736157882Sjhb turnstile_disown(ts); 737157882Sjhbout: 738157882Sjhb LOCK_LOG_LOCK("WDOWNGRADE", &rw->rw_object, 0, 0, file, line); 739157882Sjhb} 740157882Sjhb 741154941Sjhb#ifdef INVARIANT_SUPPORT 742155162Sscottl#ifndef INVARIANTS 743154941Sjhb#undef _rw_assert 744154941Sjhb#endif 745154941Sjhb 746154941Sjhb/* 747154941Sjhb * In the non-WITNESS case, rw_assert() can only detect that at least 748154941Sjhb * *some* thread owns an rlock, but it cannot guarantee that *this* 749154941Sjhb * thread owns an rlock. 750154941Sjhb */ 751154941Sjhbvoid 752154941Sjhb_rw_assert(struct rwlock *rw, int what, const char *file, int line) 753154941Sjhb{ 754154941Sjhb 755154941Sjhb if (panicstr != NULL) 756154941Sjhb return; 757154941Sjhb switch (what) { 758154941Sjhb case RA_LOCKED: 759154941Sjhb case RA_RLOCKED: 760154941Sjhb#ifdef WITNESS 761154941Sjhb witness_assert(&rw->rw_object, what, file, line); 762154941Sjhb#else 763154941Sjhb /* 764154941Sjhb * If some other thread has a write lock or we have one 765154941Sjhb * and are asserting a read lock, fail. Also, if no one 766154941Sjhb * has a lock at all, fail. 767154941Sjhb */ 768155061Sscottl if (rw->rw_lock == RW_UNLOCKED || 769155061Sscottl (!(rw->rw_lock & RW_LOCK_READ) && (what == RA_RLOCKED || 770157826Sjhb rw_wowner(rw) != curthread))) 771154941Sjhb panic("Lock %s not %slocked @ %s:%d\n", 772155012Sscottl rw->rw_object.lo_name, (what == RA_RLOCKED) ? 773154941Sjhb "read " : "", file, line); 774154941Sjhb#endif 775154941Sjhb break; 776154941Sjhb case RA_WLOCKED: 777157826Sjhb if (rw_wowner(rw) != curthread) 778154941Sjhb panic("Lock %s not exclusively locked @ %s:%d\n", 779154941Sjhb rw->rw_object.lo_name, file, line); 780154941Sjhb break; 781154941Sjhb case RA_UNLOCKED: 782154941Sjhb#ifdef WITNESS 783154941Sjhb witness_assert(&rw->rw_object, what, file, line); 784154941Sjhb#else 785154941Sjhb /* 786154941Sjhb * If we hold a write lock fail. We can't reliably check 787154941Sjhb * to see if we hold a read lock or not. 788154941Sjhb */ 789157826Sjhb if (rw_wowner(rw) == curthread) 790154941Sjhb panic("Lock %s exclusively locked @ %s:%d\n", 791154941Sjhb rw->rw_object.lo_name, file, line); 792154941Sjhb#endif 793154941Sjhb break; 794154941Sjhb default: 795154941Sjhb panic("Unknown rw lock assertion: %d @ %s:%d", what, file, 796154941Sjhb line); 797154941Sjhb } 798154941Sjhb} 799154941Sjhb#endif /* INVARIANT_SUPPORT */ 800154941Sjhb 801154941Sjhb#ifdef DDB 802154941Sjhbvoid 803154941Sjhbdb_show_rwlock(struct lock_object *lock) 804154941Sjhb{ 805154941Sjhb struct rwlock *rw; 806154941Sjhb struct thread *td; 807154941Sjhb 808154941Sjhb rw = (struct rwlock *)lock; 809154941Sjhb 810154941Sjhb db_printf(" state: "); 811154941Sjhb if (rw->rw_lock == RW_UNLOCKED) 812154941Sjhb db_printf("UNLOCKED\n"); 813154941Sjhb else if (rw->rw_lock & RW_LOCK_READ) 814154973Smlaier db_printf("RLOCK: %jd locks\n", 815154973Smlaier (intmax_t)(RW_READERS(rw->rw_lock))); 816154941Sjhb else { 817157826Sjhb td = rw_wowner(rw); 818154941Sjhb db_printf("WLOCK: %p (tid %d, pid %d, \"%s\")\n", td, 819154941Sjhb td->td_tid, td->td_proc->p_pid, td->td_proc->p_comm); 820154941Sjhb } 821154941Sjhb db_printf(" waiters: "); 822154941Sjhb switch (rw->rw_lock & (RW_LOCK_READ_WAITERS | RW_LOCK_WRITE_WAITERS)) { 823154941Sjhb case RW_LOCK_READ_WAITERS: 824154941Sjhb db_printf("readers\n"); 825154941Sjhb break; 826154941Sjhb case RW_LOCK_WRITE_WAITERS: 827154941Sjhb db_printf("writers\n"); 828154941Sjhb break; 829154941Sjhb case RW_LOCK_READ_WAITERS | RW_LOCK_WRITE_WAITERS: 830154941Sjhb db_printf("readers and waiters\n"); 831154941Sjhb break; 832154941Sjhb default: 833154941Sjhb db_printf("none\n"); 834154941Sjhb break; 835154941Sjhb } 836154941Sjhb} 837154941Sjhb 838154941Sjhb#endif 839