kern_rwlock.c revision 157882
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 157882 2006-04-19 21:06:52Z jhb $"); 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> 47154941Sjhb 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 89154941Sjhb lock_init(&rw->rw_object, &lock_class_rw, name, NULL, LO_WITNESS | 90157882Sjhb LO_RECURSABLE | LO_UPGRADABLE); 91154941Sjhb} 92154941Sjhb 93154941Sjhbvoid 94154941Sjhbrw_destroy(struct rwlock *rw) 95154941Sjhb{ 96154941Sjhb 97154941Sjhb KASSERT(rw->rw_lock == RW_UNLOCKED, ("rw lock not unlocked")); 98154941Sjhb lock_destroy(&rw->rw_object); 99154941Sjhb} 100154941Sjhb 101154941Sjhbvoid 102154941Sjhbrw_sysinit(void *arg) 103154941Sjhb{ 104154941Sjhb struct rw_args *args = arg; 105154941Sjhb 106154941Sjhb rw_init(args->ra_rw, args->ra_desc); 107154941Sjhb} 108154941Sjhb 109154941Sjhbvoid 110154941Sjhb_rw_wlock(struct rwlock *rw, const char *file, int line) 111154941Sjhb{ 112154941Sjhb 113154941Sjhb MPASS(curthread != NULL); 114157826Sjhb KASSERT(rw_wowner(rw) != curthread, 115154941Sjhb ("%s (%s): wlock already held @ %s:%d", __func__, 116154941Sjhb rw->rw_object.lo_name, file, line)); 117154941Sjhb WITNESS_CHECKORDER(&rw->rw_object, LOP_NEWORDER | LOP_EXCLUSIVE, file, 118154941Sjhb line); 119154941Sjhb __rw_wlock(rw, curthread, file, line); 120154941Sjhb LOCK_LOG_LOCK("WLOCK", &rw->rw_object, 0, 0, file, line); 121154941Sjhb WITNESS_LOCK(&rw->rw_object, LOP_EXCLUSIVE, file, line); 122154941Sjhb} 123154941Sjhb 124154941Sjhbvoid 125154941Sjhb_rw_wunlock(struct rwlock *rw, const char *file, int line) 126154941Sjhb{ 127154941Sjhb 128154941Sjhb MPASS(curthread != NULL); 129154941Sjhb _rw_assert(rw, RA_WLOCKED, file, line); 130154941Sjhb WITNESS_UNLOCK(&rw->rw_object, LOP_EXCLUSIVE, file, line); 131154941Sjhb LOCK_LOG_LOCK("WUNLOCK", &rw->rw_object, 0, 0, file, line); 132154941Sjhb __rw_wunlock(rw, curthread, file, line); 133154941Sjhb} 134154941Sjhb 135154941Sjhbvoid 136154941Sjhb_rw_rlock(struct rwlock *rw, const char *file, int line) 137154941Sjhb{ 138157851Swkoszek#ifdef SMP 139157846Sjhb volatile struct thread *owner; 140157851Swkoszek#endif 141154941Sjhb uintptr_t x; 142154941Sjhb 143157826Sjhb KASSERT(rw_wowner(rw) != curthread, 144154941Sjhb ("%s (%s): wlock already held @ %s:%d", __func__, 145154941Sjhb rw->rw_object.lo_name, file, line)); 146154941Sjhb WITNESS_CHECKORDER(&rw->rw_object, LOP_NEWORDER, file, line); 147154941Sjhb 148154941Sjhb /* 149154941Sjhb * Note that we don't make any attempt to try to block read 150154941Sjhb * locks once a writer has blocked on the lock. The reason is 151154941Sjhb * that we currently allow for read locks to recurse and we 152154941Sjhb * don't keep track of all the holders of read locks. Thus, if 153154941Sjhb * we were to block readers once a writer blocked and a reader 154154941Sjhb * tried to recurse on their reader lock after a writer had 155154941Sjhb * blocked we would end up in a deadlock since the reader would 156154941Sjhb * be blocked on the writer, and the writer would be blocked 157154941Sjhb * waiting for the reader to release its original read lock. 158154941Sjhb */ 159154941Sjhb for (;;) { 160154941Sjhb /* 161154941Sjhb * Handle the easy case. If no other thread has a write 162154941Sjhb * lock, then try to bump up the count of read locks. Note 163154941Sjhb * that we have to preserve the current state of the 164154941Sjhb * RW_LOCK_WRITE_WAITERS flag. If we fail to acquire a 165154941Sjhb * read lock, then rw_lock must have changed, so restart 166154941Sjhb * the loop. Note that this handles the case of a 167154941Sjhb * completely unlocked rwlock since such a lock is encoded 168154941Sjhb * as a read lock with no waiters. 169154941Sjhb */ 170154941Sjhb x = rw->rw_lock; 171154941Sjhb if (x & RW_LOCK_READ) { 172154941Sjhb 173154941Sjhb /* 174154941Sjhb * The RW_LOCK_READ_WAITERS flag should only be set 175154941Sjhb * if another thread currently holds a write lock, 176154941Sjhb * and in that case RW_LOCK_READ should be clear. 177154941Sjhb */ 178154941Sjhb MPASS((x & RW_LOCK_READ_WAITERS) == 0); 179154941Sjhb if (atomic_cmpset_acq_ptr(&rw->rw_lock, x, 180154941Sjhb x + RW_ONE_READER)) { 181154941Sjhb if (LOCK_LOG_TEST(&rw->rw_object, 0)) 182154941Sjhb CTR4(KTR_LOCK, 183154941Sjhb "%s: %p succeed %p -> %p", __func__, 184154941Sjhb rw, (void *)x, 185154941Sjhb (void *)(x + RW_ONE_READER)); 186154941Sjhb break; 187154941Sjhb } 188157846Sjhb cpu_spinwait(); 189154941Sjhb continue; 190154941Sjhb } 191154941Sjhb 192154941Sjhb /* 193154941Sjhb * Okay, now it's the hard case. Some other thread already 194154941Sjhb * has a write lock, so acquire the turnstile lock so we can 195154941Sjhb * begin the process of blocking. 196154941Sjhb */ 197154941Sjhb turnstile_lock(&rw->rw_object); 198154941Sjhb 199154941Sjhb /* 200154941Sjhb * The lock might have been released while we spun, so 201154941Sjhb * recheck its state and restart the loop if there is no 202154941Sjhb * longer a write lock. 203154941Sjhb */ 204154941Sjhb x = rw->rw_lock; 205154941Sjhb if (x & RW_LOCK_READ) { 206154941Sjhb turnstile_release(&rw->rw_object); 207157846Sjhb cpu_spinwait(); 208154941Sjhb continue; 209154941Sjhb } 210154941Sjhb 211154941Sjhb /* 212154941Sjhb * Ok, it's still a write lock. If the RW_LOCK_READ_WAITERS 213154941Sjhb * flag is already set, then we can go ahead and block. If 214154941Sjhb * it is not set then try to set it. If we fail to set it 215154941Sjhb * drop the turnstile lock and restart the loop. 216154941Sjhb */ 217157826Sjhb if (!(x & RW_LOCK_READ_WAITERS)) { 218157826Sjhb if (!atomic_cmpset_ptr(&rw->rw_lock, x, 219157826Sjhb x | RW_LOCK_READ_WAITERS)) { 220157826Sjhb turnstile_release(&rw->rw_object); 221157826Sjhb cpu_spinwait(); 222157826Sjhb continue; 223157826Sjhb } 224157826Sjhb if (LOCK_LOG_TEST(&rw->rw_object, 0)) 225157826Sjhb CTR2(KTR_LOCK, "%s: %p set read waiters flag", 226157826Sjhb __func__, rw); 227154941Sjhb } 228154941Sjhb 229157846Sjhb#ifdef SMP 230154941Sjhb /* 231157846Sjhb * If the owner is running on another CPU, spin until 232157846Sjhb * the owner stops running or the state of the lock 233157846Sjhb * changes. 234157846Sjhb */ 235157846Sjhb owner = (struct thread *)RW_OWNER(x); 236157846Sjhb if (TD_IS_RUNNING(owner)) { 237157846Sjhb turnstile_release(&rw->rw_object); 238157846Sjhb if (LOCK_LOG_TEST(&rw->rw_object, 0)) 239157846Sjhb CTR3(KTR_LOCK, "%s: spinning on %p held by %p", 240157846Sjhb __func__, rw, owner); 241157846Sjhb while ((struct thread*)RW_OWNER(rw->rw_lock)== owner && 242157846Sjhb TD_IS_RUNNING(owner)) 243157846Sjhb cpu_spinwait(); 244157846Sjhb continue; 245157846Sjhb } 246157846Sjhb#endif 247157846Sjhb 248157846Sjhb /* 249154941Sjhb * We were unable to acquire the lock and the read waiters 250154941Sjhb * flag is set, so we must block on the turnstile. 251154941Sjhb */ 252154941Sjhb if (LOCK_LOG_TEST(&rw->rw_object, 0)) 253154941Sjhb CTR2(KTR_LOCK, "%s: %p blocking on turnstile", __func__, 254154941Sjhb rw); 255154941Sjhb turnstile_wait(&rw->rw_object, rw_owner(rw), TS_SHARED_QUEUE); 256154941Sjhb if (LOCK_LOG_TEST(&rw->rw_object, 0)) 257154941Sjhb CTR2(KTR_LOCK, "%s: %p resuming from turnstile", 258154941Sjhb __func__, rw); 259154941Sjhb } 260154941Sjhb 261154941Sjhb /* 262154941Sjhb * TODO: acquire "owner of record" here. Here be turnstile dragons 263154941Sjhb * however. turnstiles don't like owners changing between calls to 264154941Sjhb * turnstile_wait() currently. 265154941Sjhb */ 266154941Sjhb 267154941Sjhb LOCK_LOG_LOCK("RLOCK", &rw->rw_object, 0, 0, file, line); 268154941Sjhb WITNESS_LOCK(&rw->rw_object, 0, file, line); 269154941Sjhb} 270154941Sjhb 271154941Sjhbvoid 272154941Sjhb_rw_runlock(struct rwlock *rw, const char *file, int line) 273154941Sjhb{ 274154941Sjhb struct turnstile *ts; 275154941Sjhb uintptr_t x; 276154941Sjhb 277154941Sjhb _rw_assert(rw, RA_RLOCKED, file, line); 278154941Sjhb WITNESS_UNLOCK(&rw->rw_object, 0, file, line); 279154941Sjhb LOCK_LOG_LOCK("RUNLOCK", &rw->rw_object, 0, 0, file, line); 280154941Sjhb 281154941Sjhb /* TODO: drop "owner of record" here. */ 282154941Sjhb 283154941Sjhb for (;;) { 284154941Sjhb /* 285154941Sjhb * See if there is more than one read lock held. If so, 286154941Sjhb * just drop one and return. 287154941Sjhb */ 288154941Sjhb x = rw->rw_lock; 289154941Sjhb if (RW_READERS(x) > 1) { 290154941Sjhb if (atomic_cmpset_ptr(&rw->rw_lock, x, 291154941Sjhb x - RW_ONE_READER)) { 292154941Sjhb if (LOCK_LOG_TEST(&rw->rw_object, 0)) 293154941Sjhb CTR4(KTR_LOCK, 294154941Sjhb "%s: %p succeeded %p -> %p", 295154941Sjhb __func__, rw, (void *)x, 296154941Sjhb (void *)(x - RW_ONE_READER)); 297154941Sjhb break; 298154941Sjhb } 299154941Sjhb continue; 300154941Sjhb } 301154941Sjhb 302154941Sjhb /* 303154941Sjhb * We should never have read waiters while at least one 304154941Sjhb * thread holds a read lock. (See note above) 305154941Sjhb */ 306154941Sjhb KASSERT(!(x & RW_LOCK_READ_WAITERS), 307154941Sjhb ("%s: waiting readers", __func__)); 308154941Sjhb 309154941Sjhb /* 310154941Sjhb * If there aren't any waiters for a write lock, then try 311154941Sjhb * to drop it quickly. 312154941Sjhb */ 313154941Sjhb if (!(x & RW_LOCK_WRITE_WAITERS)) { 314154941Sjhb 315154941Sjhb /* 316154941Sjhb * There shouldn't be any flags set and we should 317154941Sjhb * be the only read lock. If we fail to release 318154941Sjhb * the single read lock, then another thread might 319154941Sjhb * have just acquired a read lock, so go back up 320154941Sjhb * to the multiple read locks case. 321154941Sjhb */ 322154941Sjhb MPASS(x == RW_READERS_LOCK(1)); 323154941Sjhb if (atomic_cmpset_ptr(&rw->rw_lock, RW_READERS_LOCK(1), 324154941Sjhb RW_UNLOCKED)) { 325154941Sjhb if (LOCK_LOG_TEST(&rw->rw_object, 0)) 326154941Sjhb CTR2(KTR_LOCK, "%s: %p last succeeded", 327154941Sjhb __func__, rw); 328154941Sjhb break; 329154941Sjhb } 330154941Sjhb continue; 331154941Sjhb } 332154941Sjhb 333154941Sjhb /* 334154941Sjhb * There should just be one reader with one or more 335154941Sjhb * writers waiting. 336154941Sjhb */ 337154941Sjhb MPASS(x == (RW_READERS_LOCK(1) | RW_LOCK_WRITE_WAITERS)); 338154941Sjhb 339154941Sjhb /* 340154941Sjhb * Ok, we know we have a waiting writer and we think we 341154941Sjhb * are the last reader, so grab the turnstile lock. 342154941Sjhb */ 343154941Sjhb turnstile_lock(&rw->rw_object); 344154941Sjhb 345154941Sjhb /* 346154941Sjhb * Try to drop our lock leaving the lock in a unlocked 347154941Sjhb * state. 348154941Sjhb * 349154941Sjhb * If you wanted to do explicit lock handoff you'd have to 350154941Sjhb * do it here. You'd also want to use turnstile_signal() 351154941Sjhb * and you'd have to handle the race where a higher 352154941Sjhb * priority thread blocks on the write lock before the 353154941Sjhb * thread you wakeup actually runs and have the new thread 354154941Sjhb * "steal" the lock. For now it's a lot simpler to just 355154941Sjhb * wakeup all of the waiters. 356154941Sjhb * 357154941Sjhb * As above, if we fail, then another thread might have 358154941Sjhb * acquired a read lock, so drop the turnstile lock and 359154941Sjhb * restart. 360154941Sjhb */ 361154941Sjhb if (!atomic_cmpset_ptr(&rw->rw_lock, 362154941Sjhb RW_READERS_LOCK(1) | RW_LOCK_WRITE_WAITERS, RW_UNLOCKED)) { 363154941Sjhb turnstile_release(&rw->rw_object); 364154941Sjhb continue; 365154941Sjhb } 366154941Sjhb if (LOCK_LOG_TEST(&rw->rw_object, 0)) 367154941Sjhb CTR2(KTR_LOCK, "%s: %p last succeeded with waiters", 368154941Sjhb __func__, rw); 369154941Sjhb 370154941Sjhb /* 371154941Sjhb * Ok. The lock is released and all that's left is to 372154941Sjhb * wake up the waiters. Note that the lock might not be 373154941Sjhb * free anymore, but in that case the writers will just 374154941Sjhb * block again if they run before the new lock holder(s) 375154941Sjhb * release the lock. 376154941Sjhb */ 377154941Sjhb ts = turnstile_lookup(&rw->rw_object); 378157846Sjhb MPASS(ts != NULL); 379154941Sjhb turnstile_broadcast(ts, TS_EXCLUSIVE_QUEUE); 380154941Sjhb turnstile_unpend(ts, TS_SHARED_LOCK); 381154941Sjhb break; 382154941Sjhb } 383154941Sjhb} 384154941Sjhb 385154941Sjhb/* 386154941Sjhb * This function is called when we are unable to obtain a write lock on the 387154941Sjhb * first try. This means that at least one other thread holds either a 388154941Sjhb * read or write lock. 389154941Sjhb */ 390154941Sjhbvoid 391154941Sjhb_rw_wlock_hard(struct rwlock *rw, uintptr_t tid, const char *file, int line) 392154941Sjhb{ 393157851Swkoszek#ifdef SMP 394157846Sjhb volatile struct thread *owner; 395157851Swkoszek#endif 396154941Sjhb uintptr_t v; 397154941Sjhb 398154941Sjhb if (LOCK_LOG_TEST(&rw->rw_object, 0)) 399154941Sjhb CTR5(KTR_LOCK, "%s: %s contested (lock=%p) at %s:%d", __func__, 400154941Sjhb rw->rw_object.lo_name, (void *)rw->rw_lock, file, line); 401154941Sjhb 402154941Sjhb while (!_rw_write_lock(rw, tid)) { 403154941Sjhb turnstile_lock(&rw->rw_object); 404154941Sjhb v = rw->rw_lock; 405154941Sjhb 406154941Sjhb /* 407154941Sjhb * If the lock was released while spinning on the 408154941Sjhb * turnstile chain lock, try again. 409154941Sjhb */ 410154941Sjhb if (v == RW_UNLOCKED) { 411154941Sjhb turnstile_release(&rw->rw_object); 412154941Sjhb cpu_spinwait(); 413154941Sjhb continue; 414154941Sjhb } 415154941Sjhb 416154941Sjhb /* 417154941Sjhb * If the lock was released by a writer with both readers 418154941Sjhb * and writers waiting and a reader hasn't woken up and 419154941Sjhb * acquired the lock yet, rw_lock will be set to the 420154941Sjhb * value RW_UNLOCKED | RW_LOCK_WRITE_WAITERS. If we see 421154941Sjhb * that value, try to acquire it once. Note that we have 422154941Sjhb * to preserve the RW_LOCK_WRITE_WAITERS flag as there are 423154941Sjhb * other writers waiting still. If we fail, restart the 424154941Sjhb * loop. 425154941Sjhb */ 426154941Sjhb if (v == (RW_UNLOCKED | RW_LOCK_WRITE_WAITERS)) { 427154941Sjhb if (atomic_cmpset_acq_ptr(&rw->rw_lock, 428154941Sjhb RW_UNLOCKED | RW_LOCK_WRITE_WAITERS, 429154941Sjhb tid | RW_LOCK_WRITE_WAITERS)) { 430154941Sjhb turnstile_claim(&rw->rw_object); 431154941Sjhb CTR2(KTR_LOCK, "%s: %p claimed by new writer", 432154941Sjhb __func__, rw); 433154941Sjhb break; 434154941Sjhb } 435154941Sjhb turnstile_release(&rw->rw_object); 436154941Sjhb cpu_spinwait(); 437154941Sjhb continue; 438154941Sjhb } 439154941Sjhb 440154941Sjhb /* 441154941Sjhb * If the RW_LOCK_WRITE_WAITERS flag isn't set, then try to 442154941Sjhb * set it. If we fail to set it, then loop back and try 443154941Sjhb * again. 444154941Sjhb */ 445157826Sjhb if (!(v & RW_LOCK_WRITE_WAITERS)) { 446157826Sjhb if (!atomic_cmpset_ptr(&rw->rw_lock, v, 447157826Sjhb v | RW_LOCK_WRITE_WAITERS)) { 448157826Sjhb turnstile_release(&rw->rw_object); 449157826Sjhb cpu_spinwait(); 450157826Sjhb continue; 451157826Sjhb } 452157826Sjhb if (LOCK_LOG_TEST(&rw->rw_object, 0)) 453157826Sjhb CTR2(KTR_LOCK, "%s: %p set write waiters flag", 454157826Sjhb __func__, rw); 455154941Sjhb } 456154941Sjhb 457157846Sjhb#ifdef SMP 458157846Sjhb /* 459157846Sjhb * If the lock is write locked and the owner is 460157846Sjhb * running on another CPU, spin until the owner stops 461157846Sjhb * running or the state of the lock changes. 462157846Sjhb */ 463157846Sjhb owner = (struct thread *)RW_OWNER(v); 464157846Sjhb if (!(v & RW_LOCK_READ) && TD_IS_RUNNING(owner)) { 465157846Sjhb turnstile_release(&rw->rw_object); 466157846Sjhb if (LOCK_LOG_TEST(&rw->rw_object, 0)) 467157846Sjhb CTR3(KTR_LOCK, "%s: spinning on %p held by %p", 468157846Sjhb __func__, rw, owner); 469157846Sjhb while ((struct thread*)RW_OWNER(rw->rw_lock)== owner && 470157846Sjhb TD_IS_RUNNING(owner)) 471157846Sjhb cpu_spinwait(); 472157846Sjhb continue; 473157846Sjhb } 474157846Sjhb#endif 475154941Sjhb 476154941Sjhb /* 477154941Sjhb * We were unable to acquire the lock and the write waiters 478154941Sjhb * flag is set, so we must block on the turnstile. 479154941Sjhb */ 480154941Sjhb if (LOCK_LOG_TEST(&rw->rw_object, 0)) 481154941Sjhb CTR2(KTR_LOCK, "%s: %p blocking on turnstile", __func__, 482154941Sjhb rw); 483154941Sjhb turnstile_wait(&rw->rw_object, rw_owner(rw), 484154941Sjhb TS_EXCLUSIVE_QUEUE); 485154941Sjhb if (LOCK_LOG_TEST(&rw->rw_object, 0)) 486154941Sjhb CTR2(KTR_LOCK, "%s: %p resuming from turnstile", 487154941Sjhb __func__, rw); 488154941Sjhb } 489154941Sjhb} 490154941Sjhb 491154941Sjhb/* 492154941Sjhb * This function is called if the first try at releasing a write lock failed. 493154941Sjhb * This means that one of the 2 waiter bits must be set indicating that at 494154941Sjhb * least one thread is waiting on this lock. 495154941Sjhb */ 496154941Sjhbvoid 497154941Sjhb_rw_wunlock_hard(struct rwlock *rw, uintptr_t tid, const char *file, int line) 498154941Sjhb{ 499154941Sjhb struct turnstile *ts; 500154941Sjhb uintptr_t v; 501154941Sjhb int queue; 502154941Sjhb 503154941Sjhb KASSERT(rw->rw_lock & (RW_LOCK_READ_WAITERS | RW_LOCK_WRITE_WAITERS), 504154941Sjhb ("%s: neither of the waiter flags are set", __func__)); 505154941Sjhb 506154941Sjhb if (LOCK_LOG_TEST(&rw->rw_object, 0)) 507154941Sjhb CTR2(KTR_LOCK, "%s: %p contested", __func__, rw); 508154941Sjhb 509154941Sjhb turnstile_lock(&rw->rw_object); 510154941Sjhb ts = turnstile_lookup(&rw->rw_object); 511154941Sjhb 512157846Sjhb#ifdef SMP 513157846Sjhb /* 514157846Sjhb * There might not be a turnstile for this lock if all of 515157846Sjhb * the waiters are adaptively spinning. In that case, just 516157846Sjhb * reset the lock to the unlocked state and return. 517157846Sjhb */ 518157846Sjhb if (ts == NULL) { 519157846Sjhb atomic_store_rel_ptr(&rw->rw_lock, RW_UNLOCKED); 520157846Sjhb if (LOCK_LOG_TEST(&rw->rw_object, 0)) 521157846Sjhb CTR2(KTR_LOCK, "%s: %p no sleepers", __func__, rw); 522157846Sjhb turnstile_release(&rw->rw_object); 523157846Sjhb return; 524157846Sjhb } 525157846Sjhb#else 526154941Sjhb MPASS(ts != NULL); 527157846Sjhb#endif 528154941Sjhb 529154941Sjhb /* 530154941Sjhb * Use the same algo as sx locks for now. Prefer waking up shared 531154941Sjhb * waiters if we have any over writers. This is probably not ideal. 532154941Sjhb * 533154941Sjhb * 'v' is the value we are going to write back to rw_lock. If we 534154941Sjhb * have waiters on both queues, we need to preserve the state of 535154941Sjhb * the waiter flag for the queue we don't wake up. For now this is 536154941Sjhb * hardcoded for the algorithm mentioned above. 537154941Sjhb * 538154941Sjhb * In the case of both readers and writers waiting we wakeup the 539154941Sjhb * readers but leave the RW_LOCK_WRITE_WAITERS flag set. If a 540154941Sjhb * new writer comes in before a reader it will claim the lock up 541154941Sjhb * above. There is probably a potential priority inversion in 542154941Sjhb * there that could be worked around either by waking both queues 543154941Sjhb * of waiters or doing some complicated lock handoff gymnastics. 544157846Sjhb * 545157846Sjhb * Note that in the SMP case, if both flags are set, there might 546157846Sjhb * not be any actual writers on the turnstile as they might all 547157846Sjhb * be spinning. In that case, we don't want to preserve the 548157846Sjhb * RW_LOCK_WRITE_WAITERS flag as the turnstile is going to go 549157846Sjhb * away once we wakeup all the readers. 550154941Sjhb */ 551157846Sjhb v = RW_UNLOCKED; 552154941Sjhb if (rw->rw_lock & RW_LOCK_READ_WAITERS) { 553154941Sjhb queue = TS_SHARED_QUEUE; 554157846Sjhb#ifdef SMP 555157846Sjhb if (rw->rw_lock & RW_LOCK_WRITE_WAITERS && 556157846Sjhb !turnstile_empty(ts, TS_EXCLUSIVE_QUEUE)) 557157846Sjhb v |= RW_LOCK_WRITE_WAITERS; 558157846Sjhb#else 559157846Sjhb v |= (rw->rw_lock & RW_LOCK_WRITE_WAITERS); 560157846Sjhb#endif 561157846Sjhb } else 562154941Sjhb queue = TS_EXCLUSIVE_QUEUE; 563157846Sjhb 564157846Sjhb#ifdef SMP 565157846Sjhb /* 566157846Sjhb * We have to make sure that we actually have waiters to 567157846Sjhb * wakeup. If they are all spinning, then we just need to 568157846Sjhb * disown the turnstile and return. 569157846Sjhb */ 570157846Sjhb if (turnstile_empty(ts, queue)) { 571157846Sjhb if (LOCK_LOG_TEST(&rw->rw_object, 0)) 572157846Sjhb CTR2(KTR_LOCK, "%s: %p no sleepers 2", __func__, rw); 573157846Sjhb atomic_store_rel_ptr(&rw->rw_lock, v); 574157846Sjhb turnstile_disown(ts); 575157846Sjhb return; 576154941Sjhb } 577157846Sjhb#endif 578157846Sjhb 579157846Sjhb /* Wake up all waiters for the specific queue. */ 580154941Sjhb if (LOCK_LOG_TEST(&rw->rw_object, 0)) 581154941Sjhb CTR3(KTR_LOCK, "%s: %p waking up %s waiters", __func__, rw, 582154941Sjhb queue == TS_SHARED_QUEUE ? "read" : "write"); 583154941Sjhb turnstile_broadcast(ts, queue); 584154941Sjhb atomic_store_rel_ptr(&rw->rw_lock, v); 585154941Sjhb turnstile_unpend(ts, TS_EXCLUSIVE_LOCK); 586154941Sjhb} 587154941Sjhb 588157882Sjhb/* 589157882Sjhb * Attempt to do a non-blocking upgrade from a read lock to a write 590157882Sjhb * lock. This will only succeed if this thread holds a single read 591157882Sjhb * lock. Returns true if the upgrade succeeded and false otherwise. 592157882Sjhb */ 593157882Sjhbint 594157882Sjhb_rw_try_upgrade(struct rwlock *rw, const char *file, int line) 595157882Sjhb{ 596157882Sjhb uintptr_t v, tid; 597157882Sjhb int success; 598157882Sjhb 599157882Sjhb _rw_assert(rw, RA_RLOCKED, file, line); 600157882Sjhb 601157882Sjhb /* 602157882Sjhb * Attempt to switch from one reader to a writer. If there 603157882Sjhb * are any write waiters, then we will have to lock the 604157882Sjhb * turnstile first to prevent races with another writer 605157882Sjhb * calling turnstile_wait() before we have claimed this 606157882Sjhb * turnstile. So, do the simple case of no waiters first. 607157882Sjhb */ 608157882Sjhb tid = (uintptr_t)curthread; 609157882Sjhb if (!(rw->rw_lock & RW_LOCK_WRITE_WAITERS)) { 610157882Sjhb success = atomic_cmpset_acq_ptr(&rw->rw_lock, 611157882Sjhb RW_READERS_LOCK(1), tid); 612157882Sjhb goto out; 613157882Sjhb } 614157882Sjhb 615157882Sjhb /* 616157882Sjhb * Ok, we think we have write waiters, so lock the 617157882Sjhb * turnstile. 618157882Sjhb */ 619157882Sjhb turnstile_lock(&rw->rw_object); 620157882Sjhb 621157882Sjhb /* 622157882Sjhb * Try to switch from one reader to a writer again. This time 623157882Sjhb * we honor the current state of the RW_LOCK_WRITE_WAITERS 624157882Sjhb * flag. If we obtain the lock with the flag set, then claim 625157882Sjhb * ownership of the turnstile. In the SMP case it is possible 626157882Sjhb * for there to not be an associated turnstile even though there 627157882Sjhb * are waiters if all of the waiters are spinning. 628157882Sjhb */ 629157882Sjhb v = rw->rw_lock & RW_LOCK_WRITE_WAITERS; 630157882Sjhb success = atomic_cmpset_acq_ptr(&rw->rw_lock, RW_READERS_LOCK(1) | v, 631157882Sjhb tid | v); 632157882Sjhb#ifdef SMP 633157882Sjhb if (success && v && turnstile_lookup(&rw->rw_object) != NULL) 634157882Sjhb#else 635157882Sjhb if (success && v) 636157882Sjhb#endif 637157882Sjhb turnstile_claim(&rw->rw_object); 638157882Sjhb else 639157882Sjhb turnstile_release(&rw->rw_object); 640157882Sjhbout: 641157882Sjhb LOCK_LOG_TRY("WUPGRADE", &rw->rw_object, 0, success, file, line); 642157882Sjhb if (success) 643157882Sjhb WITNESS_UPGRADE(&rw->rw_object, LOP_EXCLUSIVE | LOP_TRYLOCK, 644157882Sjhb file, line); 645157882Sjhb return (success); 646157882Sjhb} 647157882Sjhb 648157882Sjhb/* 649157882Sjhb * Downgrade a write lock into a single read lock. 650157882Sjhb */ 651157882Sjhbvoid 652157882Sjhb_rw_downgrade(struct rwlock *rw, const char *file, int line) 653157882Sjhb{ 654157882Sjhb struct turnstile *ts; 655157882Sjhb uintptr_t tid, v; 656157882Sjhb 657157882Sjhb _rw_assert(rw, RA_WLOCKED, file, line); 658157882Sjhb 659157882Sjhb WITNESS_DOWNGRADE(&rw->rw_object, 0, file, line); 660157882Sjhb 661157882Sjhb /* 662157882Sjhb * Convert from a writer to a single reader. First we handle 663157882Sjhb * the easy case with no waiters. If there are any waiters, we 664157882Sjhb * lock the turnstile, "disown" the lock, and awaken any read 665157882Sjhb * waiters. 666157882Sjhb */ 667157882Sjhb tid = (uintptr_t)curthread; 668157882Sjhb if (atomic_cmpset_rel_ptr(&rw->rw_lock, tid, RW_READERS_LOCK(1))) 669157882Sjhb goto out; 670157882Sjhb 671157882Sjhb /* 672157882Sjhb * Ok, we think we have waiters, so lock the turnstile so we can 673157882Sjhb * read the waiter flags without any races. 674157882Sjhb */ 675157882Sjhb turnstile_lock(&rw->rw_object); 676157882Sjhb v = rw->rw_lock; 677157882Sjhb MPASS(v & (RW_LOCK_READ_WAITERS | RW_LOCK_WRITE_WAITERS)); 678157882Sjhb 679157882Sjhb /* 680157882Sjhb * Downgrade from a write lock while preserving 681157882Sjhb * RW_LOCK_WRITE_WAITERS and give up ownership of the 682157882Sjhb * turnstile. If there are any read waiters, wake them up. 683157882Sjhb * 684157882Sjhb * For SMP, we have to allow for the fact that all of the 685157882Sjhb * read waiters might be spinning. In that case, act as if 686157882Sjhb * RW_LOCK_READ_WAITERS is not set. Also, only preserve 687157882Sjhb * the RW_LOCK_WRITE_WAITERS flag if at least one writer is 688157882Sjhb * blocked on the turnstile. 689157882Sjhb */ 690157882Sjhb ts = turnstile_lookup(&rw->rw_object); 691157882Sjhb#ifdef SMP 692157882Sjhb if (ts == NULL) 693157882Sjhb v &= ~(RW_LOCK_READ_WAITERS | RW_LOCK_WRITE_WAITERS); 694157882Sjhb else if (v & RW_LOCK_READ_WAITERS && 695157882Sjhb turnstile_empty(ts, TS_SHARED_QUEUE)) 696157882Sjhb v &= ~RW_LOCK_READ_WAITERS; 697157882Sjhb else if (v & RW_LOCK_WRITE_WAITERS && 698157882Sjhb turnstile_empty(ts, TS_EXCLUSIVE_QUEUE)) 699157882Sjhb v &= ~RW_LOCK_WRITE_WAITERS; 700157882Sjhb#else 701157882Sjhb MPASS(ts != NULL); 702157882Sjhb#endif 703157882Sjhb if (v & RW_LOCK_READ_WAITERS) 704157882Sjhb turnstile_broadcast(ts, TS_SHARED_QUEUE); 705157882Sjhb atomic_store_rel_ptr(&rw->rw_lock, RW_READERS_LOCK(1) | 706157882Sjhb (v & RW_LOCK_WRITE_WAITERS)); 707157882Sjhb if (v & RW_LOCK_READ_WAITERS) 708157882Sjhb turnstile_unpend(ts, TS_EXCLUSIVE_LOCK); 709157882Sjhb#ifdef SMP 710157882Sjhb else if (ts == NULL) 711157882Sjhb turnstile_release(&rw->rw_object); 712157882Sjhb#endif 713157882Sjhb else 714157882Sjhb turnstile_disown(ts); 715157882Sjhbout: 716157882Sjhb LOCK_LOG_LOCK("WDOWNGRADE", &rw->rw_object, 0, 0, file, line); 717157882Sjhb} 718157882Sjhb 719154941Sjhb#ifdef INVARIANT_SUPPORT 720155162Sscottl#ifndef INVARIANTS 721154941Sjhb#undef _rw_assert 722154941Sjhb#endif 723154941Sjhb 724154941Sjhb/* 725154941Sjhb * In the non-WITNESS case, rw_assert() can only detect that at least 726154941Sjhb * *some* thread owns an rlock, but it cannot guarantee that *this* 727154941Sjhb * thread owns an rlock. 728154941Sjhb */ 729154941Sjhbvoid 730154941Sjhb_rw_assert(struct rwlock *rw, int what, const char *file, int line) 731154941Sjhb{ 732154941Sjhb 733154941Sjhb if (panicstr != NULL) 734154941Sjhb return; 735154941Sjhb switch (what) { 736154941Sjhb case RA_LOCKED: 737154941Sjhb case RA_RLOCKED: 738154941Sjhb#ifdef WITNESS 739154941Sjhb witness_assert(&rw->rw_object, what, file, line); 740154941Sjhb#else 741154941Sjhb /* 742154941Sjhb * If some other thread has a write lock or we have one 743154941Sjhb * and are asserting a read lock, fail. Also, if no one 744154941Sjhb * has a lock at all, fail. 745154941Sjhb */ 746155061Sscottl if (rw->rw_lock == RW_UNLOCKED || 747155061Sscottl (!(rw->rw_lock & RW_LOCK_READ) && (what == RA_RLOCKED || 748157826Sjhb rw_wowner(rw) != curthread))) 749154941Sjhb panic("Lock %s not %slocked @ %s:%d\n", 750155012Sscottl rw->rw_object.lo_name, (what == RA_RLOCKED) ? 751154941Sjhb "read " : "", file, line); 752154941Sjhb#endif 753154941Sjhb break; 754154941Sjhb case RA_WLOCKED: 755157826Sjhb if (rw_wowner(rw) != curthread) 756154941Sjhb panic("Lock %s not exclusively locked @ %s:%d\n", 757154941Sjhb rw->rw_object.lo_name, file, line); 758154941Sjhb break; 759154941Sjhb case RA_UNLOCKED: 760154941Sjhb#ifdef WITNESS 761154941Sjhb witness_assert(&rw->rw_object, what, file, line); 762154941Sjhb#else 763154941Sjhb /* 764154941Sjhb * If we hold a write lock fail. We can't reliably check 765154941Sjhb * to see if we hold a read lock or not. 766154941Sjhb */ 767157826Sjhb if (rw_wowner(rw) == curthread) 768154941Sjhb panic("Lock %s exclusively locked @ %s:%d\n", 769154941Sjhb rw->rw_object.lo_name, file, line); 770154941Sjhb#endif 771154941Sjhb break; 772154941Sjhb default: 773154941Sjhb panic("Unknown rw lock assertion: %d @ %s:%d", what, file, 774154941Sjhb line); 775154941Sjhb } 776154941Sjhb} 777154941Sjhb#endif /* INVARIANT_SUPPORT */ 778154941Sjhb 779154941Sjhb#ifdef DDB 780154941Sjhbvoid 781154941Sjhbdb_show_rwlock(struct lock_object *lock) 782154941Sjhb{ 783154941Sjhb struct rwlock *rw; 784154941Sjhb struct thread *td; 785154941Sjhb 786154941Sjhb rw = (struct rwlock *)lock; 787154941Sjhb 788154941Sjhb db_printf(" state: "); 789154941Sjhb if (rw->rw_lock == RW_UNLOCKED) 790154941Sjhb db_printf("UNLOCKED\n"); 791154941Sjhb else if (rw->rw_lock & RW_LOCK_READ) 792154973Smlaier db_printf("RLOCK: %jd locks\n", 793154973Smlaier (intmax_t)(RW_READERS(rw->rw_lock))); 794154941Sjhb else { 795157826Sjhb td = rw_wowner(rw); 796154941Sjhb db_printf("WLOCK: %p (tid %d, pid %d, \"%s\")\n", td, 797154941Sjhb td->td_tid, td->td_proc->p_pid, td->td_proc->p_comm); 798154941Sjhb } 799154941Sjhb db_printf(" waiters: "); 800154941Sjhb switch (rw->rw_lock & (RW_LOCK_READ_WAITERS | RW_LOCK_WRITE_WAITERS)) { 801154941Sjhb case RW_LOCK_READ_WAITERS: 802154941Sjhb db_printf("readers\n"); 803154941Sjhb break; 804154941Sjhb case RW_LOCK_WRITE_WAITERS: 805154941Sjhb db_printf("writers\n"); 806154941Sjhb break; 807154941Sjhb case RW_LOCK_READ_WAITERS | RW_LOCK_WRITE_WAITERS: 808154941Sjhb db_printf("readers and waiters\n"); 809154941Sjhb break; 810154941Sjhb default: 811154941Sjhb db_printf("none\n"); 812154941Sjhb break; 813154941Sjhb } 814154941Sjhb} 815154941Sjhb 816154941Sjhb#endif 817