kern_rwlock.c revision 167365
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 167365 2007-03-09 16:04:44Z 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> 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 = { 57167365Sjhb .lc_name = "rw", 58167365Sjhb .lc_flags = LC_SLEEPLOCK | LC_RECURSABLE | LC_UPGRADABLE, 59154941Sjhb#ifdef DDB 60167365Sjhb .lc_ddb_show = 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 89164246Skmacy lock_profile_object_init(&rw->rw_object, &lock_class_rw, 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 111167024Srwatsonint 112167024Srwatsonrw_wowned(struct rwlock *rw) 113167024Srwatson{ 114167024Srwatson 115167024Srwatson return (rw_wowner(rw) == curthread); 116167024Srwatson} 117167024Srwatson 118154941Sjhbvoid 119154941Sjhb_rw_wlock(struct rwlock *rw, const char *file, int line) 120154941Sjhb{ 121154941Sjhb 122154941Sjhb MPASS(curthread != NULL); 123157826Sjhb KASSERT(rw_wowner(rw) != curthread, 124154941Sjhb ("%s (%s): wlock already held @ %s:%d", __func__, 125154941Sjhb rw->rw_object.lo_name, file, line)); 126154941Sjhb WITNESS_CHECKORDER(&rw->rw_object, LOP_NEWORDER | LOP_EXCLUSIVE, file, 127154941Sjhb line); 128154941Sjhb __rw_wlock(rw, curthread, file, line); 129154941Sjhb LOCK_LOG_LOCK("WLOCK", &rw->rw_object, 0, 0, file, line); 130154941Sjhb WITNESS_LOCK(&rw->rw_object, LOP_EXCLUSIVE, file, line); 131160771Sjhb curthread->td_locks++; 132154941Sjhb} 133154941Sjhb 134154941Sjhbvoid 135154941Sjhb_rw_wunlock(struct rwlock *rw, const char *file, int line) 136154941Sjhb{ 137154941Sjhb 138154941Sjhb MPASS(curthread != NULL); 139154941Sjhb _rw_assert(rw, RA_WLOCKED, file, line); 140160771Sjhb curthread->td_locks--; 141154941Sjhb WITNESS_UNLOCK(&rw->rw_object, LOP_EXCLUSIVE, file, line); 142154941Sjhb LOCK_LOG_LOCK("WUNLOCK", &rw->rw_object, 0, 0, file, line); 143164159Skmacy lock_profile_release_lock(&rw->rw_object); 144154941Sjhb __rw_wunlock(rw, curthread, file, line); 145154941Sjhb} 146154941Sjhb 147154941Sjhbvoid 148154941Sjhb_rw_rlock(struct rwlock *rw, const char *file, int line) 149154941Sjhb{ 150157851Swkoszek#ifdef SMP 151157846Sjhb volatile struct thread *owner; 152157851Swkoszek#endif 153167307Sjhb uint64_t waittime = 0; 154167054Skmacy int contested = 0; 155154941Sjhb uintptr_t x; 156154941Sjhb 157157826Sjhb KASSERT(rw_wowner(rw) != curthread, 158154941Sjhb ("%s (%s): wlock already held @ %s:%d", __func__, 159154941Sjhb rw->rw_object.lo_name, file, line)); 160154941Sjhb WITNESS_CHECKORDER(&rw->rw_object, LOP_NEWORDER, file, line); 161154941Sjhb 162154941Sjhb /* 163154941Sjhb * Note that we don't make any attempt to try to block read 164154941Sjhb * locks once a writer has blocked on the lock. The reason is 165154941Sjhb * that we currently allow for read locks to recurse and we 166154941Sjhb * don't keep track of all the holders of read locks. Thus, if 167154941Sjhb * we were to block readers once a writer blocked and a reader 168154941Sjhb * tried to recurse on their reader lock after a writer had 169154941Sjhb * blocked we would end up in a deadlock since the reader would 170154941Sjhb * be blocked on the writer, and the writer would be blocked 171154941Sjhb * waiting for the reader to release its original read lock. 172154941Sjhb */ 173154941Sjhb for (;;) { 174154941Sjhb /* 175154941Sjhb * Handle the easy case. If no other thread has a write 176154941Sjhb * lock, then try to bump up the count of read locks. Note 177154941Sjhb * that we have to preserve the current state of the 178154941Sjhb * RW_LOCK_WRITE_WAITERS flag. If we fail to acquire a 179154941Sjhb * read lock, then rw_lock must have changed, so restart 180154941Sjhb * the loop. Note that this handles the case of a 181154941Sjhb * completely unlocked rwlock since such a lock is encoded 182154941Sjhb * as a read lock with no waiters. 183154941Sjhb */ 184154941Sjhb x = rw->rw_lock; 185154941Sjhb if (x & RW_LOCK_READ) { 186154941Sjhb 187154941Sjhb /* 188154941Sjhb * The RW_LOCK_READ_WAITERS flag should only be set 189154941Sjhb * if another thread currently holds a write lock, 190154941Sjhb * and in that case RW_LOCK_READ should be clear. 191154941Sjhb */ 192154941Sjhb MPASS((x & RW_LOCK_READ_WAITERS) == 0); 193154941Sjhb if (atomic_cmpset_acq_ptr(&rw->rw_lock, x, 194154941Sjhb x + RW_ONE_READER)) { 195154941Sjhb if (LOCK_LOG_TEST(&rw->rw_object, 0)) 196154941Sjhb CTR4(KTR_LOCK, 197154941Sjhb "%s: %p succeed %p -> %p", __func__, 198154941Sjhb rw, (void *)x, 199154941Sjhb (void *)(x + RW_ONE_READER)); 200167307Sjhb if (RW_READERS(x) == 0) 201167307Sjhb lock_profile_obtain_lock_success( 202167307Sjhb &rw->rw_object, contested, waittime, 203167307Sjhb file, line); 204154941Sjhb break; 205154941Sjhb } 206157846Sjhb cpu_spinwait(); 207154941Sjhb continue; 208154941Sjhb } 209167307Sjhb lock_profile_obtain_lock_failed(&rw->rw_object, &contested, 210167307Sjhb &waittime); 211154941Sjhb 212154941Sjhb /* 213154941Sjhb * Okay, now it's the hard case. Some other thread already 214154941Sjhb * has a write lock, so acquire the turnstile lock so we can 215154941Sjhb * begin the process of blocking. 216154941Sjhb */ 217154941Sjhb turnstile_lock(&rw->rw_object); 218154941Sjhb 219154941Sjhb /* 220154941Sjhb * The lock might have been released while we spun, so 221154941Sjhb * recheck its state and restart the loop if there is no 222154941Sjhb * longer a write lock. 223154941Sjhb */ 224154941Sjhb x = rw->rw_lock; 225154941Sjhb if (x & RW_LOCK_READ) { 226154941Sjhb turnstile_release(&rw->rw_object); 227157846Sjhb cpu_spinwait(); 228154941Sjhb continue; 229154941Sjhb } 230154941Sjhb 231154941Sjhb /* 232154941Sjhb * Ok, it's still a write lock. If the RW_LOCK_READ_WAITERS 233154941Sjhb * flag is already set, then we can go ahead and block. If 234154941Sjhb * it is not set then try to set it. If we fail to set it 235154941Sjhb * drop the turnstile lock and restart the loop. 236154941Sjhb */ 237157826Sjhb if (!(x & RW_LOCK_READ_WAITERS)) { 238157826Sjhb if (!atomic_cmpset_ptr(&rw->rw_lock, x, 239157826Sjhb x | RW_LOCK_READ_WAITERS)) { 240157826Sjhb turnstile_release(&rw->rw_object); 241157826Sjhb cpu_spinwait(); 242157826Sjhb continue; 243157826Sjhb } 244157826Sjhb if (LOCK_LOG_TEST(&rw->rw_object, 0)) 245157826Sjhb CTR2(KTR_LOCK, "%s: %p set read waiters flag", 246157826Sjhb __func__, rw); 247154941Sjhb } 248154941Sjhb 249157846Sjhb#ifdef SMP 250154941Sjhb /* 251157846Sjhb * If the owner is running on another CPU, spin until 252157846Sjhb * the owner stops running or the state of the lock 253157846Sjhb * changes. 254157846Sjhb */ 255157846Sjhb owner = (struct thread *)RW_OWNER(x); 256157846Sjhb if (TD_IS_RUNNING(owner)) { 257157846Sjhb turnstile_release(&rw->rw_object); 258157846Sjhb if (LOCK_LOG_TEST(&rw->rw_object, 0)) 259157846Sjhb CTR3(KTR_LOCK, "%s: spinning on %p held by %p", 260157846Sjhb __func__, rw, owner); 261157846Sjhb while ((struct thread*)RW_OWNER(rw->rw_lock)== owner && 262157846Sjhb TD_IS_RUNNING(owner)) 263157846Sjhb cpu_spinwait(); 264157846Sjhb continue; 265157846Sjhb } 266157846Sjhb#endif 267157846Sjhb 268157846Sjhb /* 269154941Sjhb * We were unable to acquire the lock and the read waiters 270154941Sjhb * flag is set, so we must block on the turnstile. 271154941Sjhb */ 272154941Sjhb if (LOCK_LOG_TEST(&rw->rw_object, 0)) 273154941Sjhb CTR2(KTR_LOCK, "%s: %p blocking on turnstile", __func__, 274154941Sjhb rw); 275154941Sjhb turnstile_wait(&rw->rw_object, rw_owner(rw), TS_SHARED_QUEUE); 276154941Sjhb if (LOCK_LOG_TEST(&rw->rw_object, 0)) 277154941Sjhb CTR2(KTR_LOCK, "%s: %p resuming from turnstile", 278154941Sjhb __func__, rw); 279154941Sjhb } 280154941Sjhb 281154941Sjhb /* 282154941Sjhb * TODO: acquire "owner of record" here. Here be turnstile dragons 283154941Sjhb * however. turnstiles don't like owners changing between calls to 284154941Sjhb * turnstile_wait() currently. 285154941Sjhb */ 286154941Sjhb 287154941Sjhb LOCK_LOG_LOCK("RLOCK", &rw->rw_object, 0, 0, file, line); 288154941Sjhb WITNESS_LOCK(&rw->rw_object, 0, file, line); 289160771Sjhb curthread->td_locks++; 290154941Sjhb} 291154941Sjhb 292154941Sjhbvoid 293154941Sjhb_rw_runlock(struct rwlock *rw, const char *file, int line) 294154941Sjhb{ 295154941Sjhb struct turnstile *ts; 296154941Sjhb uintptr_t x; 297154941Sjhb 298154941Sjhb _rw_assert(rw, RA_RLOCKED, file, line); 299160771Sjhb curthread->td_locks--; 300154941Sjhb WITNESS_UNLOCK(&rw->rw_object, 0, file, line); 301154941Sjhb LOCK_LOG_LOCK("RUNLOCK", &rw->rw_object, 0, 0, file, line); 302154941Sjhb 303154941Sjhb /* TODO: drop "owner of record" here. */ 304154941Sjhb 305154941Sjhb for (;;) { 306154941Sjhb /* 307154941Sjhb * See if there is more than one read lock held. If so, 308154941Sjhb * just drop one and return. 309154941Sjhb */ 310154941Sjhb x = rw->rw_lock; 311154941Sjhb if (RW_READERS(x) > 1) { 312154941Sjhb if (atomic_cmpset_ptr(&rw->rw_lock, x, 313154941Sjhb x - RW_ONE_READER)) { 314154941Sjhb if (LOCK_LOG_TEST(&rw->rw_object, 0)) 315154941Sjhb CTR4(KTR_LOCK, 316154941Sjhb "%s: %p succeeded %p -> %p", 317154941Sjhb __func__, rw, (void *)x, 318154941Sjhb (void *)(x - RW_ONE_READER)); 319154941Sjhb break; 320154941Sjhb } 321154941Sjhb continue; 322167307Sjhb } 323154941Sjhb 324164159Skmacy 325154941Sjhb /* 326154941Sjhb * We should never have read waiters while at least one 327154941Sjhb * thread holds a read lock. (See note above) 328154941Sjhb */ 329154941Sjhb KASSERT(!(x & RW_LOCK_READ_WAITERS), 330154941Sjhb ("%s: waiting readers", __func__)); 331154941Sjhb 332154941Sjhb /* 333154941Sjhb * If there aren't any waiters for a write lock, then try 334154941Sjhb * to drop it quickly. 335154941Sjhb */ 336154941Sjhb if (!(x & RW_LOCK_WRITE_WAITERS)) { 337154941Sjhb 338154941Sjhb /* 339154941Sjhb * There shouldn't be any flags set and we should 340154941Sjhb * be the only read lock. If we fail to release 341154941Sjhb * the single read lock, then another thread might 342154941Sjhb * have just acquired a read lock, so go back up 343154941Sjhb * to the multiple read locks case. 344154941Sjhb */ 345154941Sjhb MPASS(x == RW_READERS_LOCK(1)); 346154941Sjhb if (atomic_cmpset_ptr(&rw->rw_lock, RW_READERS_LOCK(1), 347154941Sjhb RW_UNLOCKED)) { 348154941Sjhb if (LOCK_LOG_TEST(&rw->rw_object, 0)) 349154941Sjhb CTR2(KTR_LOCK, "%s: %p last succeeded", 350154941Sjhb __func__, rw); 351154941Sjhb break; 352154941Sjhb } 353154941Sjhb continue; 354154941Sjhb } 355154941Sjhb 356154941Sjhb /* 357154941Sjhb * There should just be one reader with one or more 358154941Sjhb * writers waiting. 359154941Sjhb */ 360154941Sjhb MPASS(x == (RW_READERS_LOCK(1) | RW_LOCK_WRITE_WAITERS)); 361154941Sjhb 362154941Sjhb /* 363154941Sjhb * Ok, we know we have a waiting writer and we think we 364154941Sjhb * are the last reader, so grab the turnstile lock. 365154941Sjhb */ 366154941Sjhb turnstile_lock(&rw->rw_object); 367154941Sjhb 368154941Sjhb /* 369154941Sjhb * Try to drop our lock leaving the lock in a unlocked 370154941Sjhb * state. 371154941Sjhb * 372154941Sjhb * If you wanted to do explicit lock handoff you'd have to 373154941Sjhb * do it here. You'd also want to use turnstile_signal() 374154941Sjhb * and you'd have to handle the race where a higher 375154941Sjhb * priority thread blocks on the write lock before the 376154941Sjhb * thread you wakeup actually runs and have the new thread 377154941Sjhb * "steal" the lock. For now it's a lot simpler to just 378154941Sjhb * wakeup all of the waiters. 379154941Sjhb * 380154941Sjhb * As above, if we fail, then another thread might have 381154941Sjhb * acquired a read lock, so drop the turnstile lock and 382154941Sjhb * restart. 383154941Sjhb */ 384154941Sjhb if (!atomic_cmpset_ptr(&rw->rw_lock, 385154941Sjhb RW_READERS_LOCK(1) | RW_LOCK_WRITE_WAITERS, RW_UNLOCKED)) { 386154941Sjhb turnstile_release(&rw->rw_object); 387154941Sjhb continue; 388154941Sjhb } 389154941Sjhb if (LOCK_LOG_TEST(&rw->rw_object, 0)) 390154941Sjhb CTR2(KTR_LOCK, "%s: %p last succeeded with waiters", 391154941Sjhb __func__, rw); 392154941Sjhb 393154941Sjhb /* 394154941Sjhb * Ok. The lock is released and all that's left is to 395154941Sjhb * wake up the waiters. Note that the lock might not be 396154941Sjhb * free anymore, but in that case the writers will just 397154941Sjhb * block again if they run before the new lock holder(s) 398154941Sjhb * release the lock. 399154941Sjhb */ 400154941Sjhb ts = turnstile_lookup(&rw->rw_object); 401157846Sjhb MPASS(ts != NULL); 402154941Sjhb turnstile_broadcast(ts, TS_EXCLUSIVE_QUEUE); 403154941Sjhb turnstile_unpend(ts, TS_SHARED_LOCK); 404154941Sjhb break; 405154941Sjhb } 406167307Sjhb lock_profile_release_lock(&rw->rw_object); 407154941Sjhb} 408154941Sjhb 409154941Sjhb/* 410154941Sjhb * This function is called when we are unable to obtain a write lock on the 411154941Sjhb * first try. This means that at least one other thread holds either a 412154941Sjhb * read or write lock. 413154941Sjhb */ 414154941Sjhbvoid 415154941Sjhb_rw_wlock_hard(struct rwlock *rw, uintptr_t tid, const char *file, int line) 416154941Sjhb{ 417157851Swkoszek#ifdef SMP 418157846Sjhb volatile struct thread *owner; 419157851Swkoszek#endif 420154941Sjhb uintptr_t v; 421154941Sjhb 422154941Sjhb if (LOCK_LOG_TEST(&rw->rw_object, 0)) 423154941Sjhb CTR5(KTR_LOCK, "%s: %s contested (lock=%p) at %s:%d", __func__, 424154941Sjhb rw->rw_object.lo_name, (void *)rw->rw_lock, file, line); 425154941Sjhb 426154941Sjhb while (!_rw_write_lock(rw, tid)) { 427154941Sjhb turnstile_lock(&rw->rw_object); 428154941Sjhb v = rw->rw_lock; 429154941Sjhb 430154941Sjhb /* 431154941Sjhb * If the lock was released while spinning on the 432154941Sjhb * turnstile chain lock, try again. 433154941Sjhb */ 434154941Sjhb if (v == RW_UNLOCKED) { 435154941Sjhb turnstile_release(&rw->rw_object); 436154941Sjhb cpu_spinwait(); 437154941Sjhb continue; 438154941Sjhb } 439154941Sjhb 440154941Sjhb /* 441154941Sjhb * If the lock was released by a writer with both readers 442154941Sjhb * and writers waiting and a reader hasn't woken up and 443154941Sjhb * acquired the lock yet, rw_lock will be set to the 444154941Sjhb * value RW_UNLOCKED | RW_LOCK_WRITE_WAITERS. If we see 445154941Sjhb * that value, try to acquire it once. Note that we have 446154941Sjhb * to preserve the RW_LOCK_WRITE_WAITERS flag as there are 447154941Sjhb * other writers waiting still. If we fail, restart the 448154941Sjhb * loop. 449154941Sjhb */ 450154941Sjhb if (v == (RW_UNLOCKED | RW_LOCK_WRITE_WAITERS)) { 451154941Sjhb if (atomic_cmpset_acq_ptr(&rw->rw_lock, 452154941Sjhb RW_UNLOCKED | RW_LOCK_WRITE_WAITERS, 453154941Sjhb tid | RW_LOCK_WRITE_WAITERS)) { 454154941Sjhb turnstile_claim(&rw->rw_object); 455154941Sjhb CTR2(KTR_LOCK, "%s: %p claimed by new writer", 456154941Sjhb __func__, rw); 457154941Sjhb break; 458154941Sjhb } 459154941Sjhb turnstile_release(&rw->rw_object); 460154941Sjhb cpu_spinwait(); 461154941Sjhb continue; 462154941Sjhb } 463154941Sjhb 464154941Sjhb /* 465154941Sjhb * If the RW_LOCK_WRITE_WAITERS flag isn't set, then try to 466154941Sjhb * set it. If we fail to set it, then loop back and try 467154941Sjhb * again. 468154941Sjhb */ 469157826Sjhb if (!(v & RW_LOCK_WRITE_WAITERS)) { 470157826Sjhb if (!atomic_cmpset_ptr(&rw->rw_lock, v, 471157826Sjhb v | RW_LOCK_WRITE_WAITERS)) { 472157826Sjhb turnstile_release(&rw->rw_object); 473157826Sjhb cpu_spinwait(); 474157826Sjhb continue; 475157826Sjhb } 476157826Sjhb if (LOCK_LOG_TEST(&rw->rw_object, 0)) 477157826Sjhb CTR2(KTR_LOCK, "%s: %p set write waiters flag", 478157826Sjhb __func__, rw); 479154941Sjhb } 480154941Sjhb 481157846Sjhb#ifdef SMP 482157846Sjhb /* 483157846Sjhb * If the lock is write locked and the owner is 484157846Sjhb * running on another CPU, spin until the owner stops 485157846Sjhb * running or the state of the lock changes. 486157846Sjhb */ 487157846Sjhb owner = (struct thread *)RW_OWNER(v); 488157846Sjhb if (!(v & RW_LOCK_READ) && TD_IS_RUNNING(owner)) { 489157846Sjhb turnstile_release(&rw->rw_object); 490157846Sjhb if (LOCK_LOG_TEST(&rw->rw_object, 0)) 491157846Sjhb CTR3(KTR_LOCK, "%s: spinning on %p held by %p", 492157846Sjhb __func__, rw, owner); 493157846Sjhb while ((struct thread*)RW_OWNER(rw->rw_lock)== owner && 494157846Sjhb TD_IS_RUNNING(owner)) 495157846Sjhb cpu_spinwait(); 496157846Sjhb continue; 497157846Sjhb } 498157846Sjhb#endif 499154941Sjhb 500154941Sjhb /* 501154941Sjhb * We were unable to acquire the lock and the write waiters 502154941Sjhb * flag is set, so we must block on the turnstile. 503154941Sjhb */ 504154941Sjhb if (LOCK_LOG_TEST(&rw->rw_object, 0)) 505154941Sjhb CTR2(KTR_LOCK, "%s: %p blocking on turnstile", __func__, 506154941Sjhb rw); 507154941Sjhb turnstile_wait(&rw->rw_object, rw_owner(rw), 508154941Sjhb TS_EXCLUSIVE_QUEUE); 509154941Sjhb if (LOCK_LOG_TEST(&rw->rw_object, 0)) 510154941Sjhb CTR2(KTR_LOCK, "%s: %p resuming from turnstile", 511154941Sjhb __func__, rw); 512154941Sjhb } 513154941Sjhb} 514154941Sjhb 515154941Sjhb/* 516154941Sjhb * This function is called if the first try at releasing a write lock failed. 517154941Sjhb * This means that one of the 2 waiter bits must be set indicating that at 518154941Sjhb * least one thread is waiting on this lock. 519154941Sjhb */ 520154941Sjhbvoid 521154941Sjhb_rw_wunlock_hard(struct rwlock *rw, uintptr_t tid, const char *file, int line) 522154941Sjhb{ 523154941Sjhb struct turnstile *ts; 524154941Sjhb uintptr_t v; 525154941Sjhb int queue; 526154941Sjhb 527154941Sjhb KASSERT(rw->rw_lock & (RW_LOCK_READ_WAITERS | RW_LOCK_WRITE_WAITERS), 528154941Sjhb ("%s: neither of the waiter flags are set", __func__)); 529154941Sjhb 530154941Sjhb if (LOCK_LOG_TEST(&rw->rw_object, 0)) 531154941Sjhb CTR2(KTR_LOCK, "%s: %p contested", __func__, rw); 532154941Sjhb 533154941Sjhb turnstile_lock(&rw->rw_object); 534154941Sjhb ts = turnstile_lookup(&rw->rw_object); 535154941Sjhb 536157846Sjhb#ifdef SMP 537157846Sjhb /* 538157846Sjhb * There might not be a turnstile for this lock if all of 539157846Sjhb * the waiters are adaptively spinning. In that case, just 540157846Sjhb * reset the lock to the unlocked state and return. 541157846Sjhb */ 542157846Sjhb if (ts == NULL) { 543157846Sjhb atomic_store_rel_ptr(&rw->rw_lock, RW_UNLOCKED); 544157846Sjhb if (LOCK_LOG_TEST(&rw->rw_object, 0)) 545157846Sjhb CTR2(KTR_LOCK, "%s: %p no sleepers", __func__, rw); 546157846Sjhb turnstile_release(&rw->rw_object); 547157846Sjhb return; 548157846Sjhb } 549157846Sjhb#else 550154941Sjhb MPASS(ts != NULL); 551157846Sjhb#endif 552154941Sjhb 553154941Sjhb /* 554154941Sjhb * Use the same algo as sx locks for now. Prefer waking up shared 555154941Sjhb * waiters if we have any over writers. This is probably not ideal. 556154941Sjhb * 557154941Sjhb * 'v' is the value we are going to write back to rw_lock. If we 558154941Sjhb * have waiters on both queues, we need to preserve the state of 559154941Sjhb * the waiter flag for the queue we don't wake up. For now this is 560154941Sjhb * hardcoded for the algorithm mentioned above. 561154941Sjhb * 562154941Sjhb * In the case of both readers and writers waiting we wakeup the 563154941Sjhb * readers but leave the RW_LOCK_WRITE_WAITERS flag set. If a 564154941Sjhb * new writer comes in before a reader it will claim the lock up 565154941Sjhb * above. There is probably a potential priority inversion in 566154941Sjhb * there that could be worked around either by waking both queues 567154941Sjhb * of waiters or doing some complicated lock handoff gymnastics. 568157846Sjhb * 569157846Sjhb * Note that in the SMP case, if both flags are set, there might 570157846Sjhb * not be any actual writers on the turnstile as they might all 571157846Sjhb * be spinning. In that case, we don't want to preserve the 572157846Sjhb * RW_LOCK_WRITE_WAITERS flag as the turnstile is going to go 573157846Sjhb * away once we wakeup all the readers. 574154941Sjhb */ 575157846Sjhb v = RW_UNLOCKED; 576154941Sjhb if (rw->rw_lock & RW_LOCK_READ_WAITERS) { 577154941Sjhb queue = TS_SHARED_QUEUE; 578157846Sjhb#ifdef SMP 579157846Sjhb if (rw->rw_lock & RW_LOCK_WRITE_WAITERS && 580157846Sjhb !turnstile_empty(ts, TS_EXCLUSIVE_QUEUE)) 581157846Sjhb v |= RW_LOCK_WRITE_WAITERS; 582157846Sjhb#else 583157846Sjhb v |= (rw->rw_lock & RW_LOCK_WRITE_WAITERS); 584157846Sjhb#endif 585157846Sjhb } else 586154941Sjhb queue = TS_EXCLUSIVE_QUEUE; 587157846Sjhb 588157846Sjhb#ifdef SMP 589157846Sjhb /* 590157846Sjhb * We have to make sure that we actually have waiters to 591157846Sjhb * wakeup. If they are all spinning, then we just need to 592157846Sjhb * disown the turnstile and return. 593157846Sjhb */ 594157846Sjhb if (turnstile_empty(ts, queue)) { 595157846Sjhb if (LOCK_LOG_TEST(&rw->rw_object, 0)) 596157846Sjhb CTR2(KTR_LOCK, "%s: %p no sleepers 2", __func__, rw); 597157846Sjhb atomic_store_rel_ptr(&rw->rw_lock, v); 598157846Sjhb turnstile_disown(ts); 599157846Sjhb return; 600154941Sjhb } 601157846Sjhb#endif 602157846Sjhb 603157846Sjhb /* Wake up all waiters for the specific queue. */ 604154941Sjhb if (LOCK_LOG_TEST(&rw->rw_object, 0)) 605154941Sjhb CTR3(KTR_LOCK, "%s: %p waking up %s waiters", __func__, rw, 606154941Sjhb queue == TS_SHARED_QUEUE ? "read" : "write"); 607154941Sjhb turnstile_broadcast(ts, queue); 608154941Sjhb atomic_store_rel_ptr(&rw->rw_lock, v); 609154941Sjhb turnstile_unpend(ts, TS_EXCLUSIVE_LOCK); 610154941Sjhb} 611154941Sjhb 612157882Sjhb/* 613157882Sjhb * Attempt to do a non-blocking upgrade from a read lock to a write 614157882Sjhb * lock. This will only succeed if this thread holds a single read 615157882Sjhb * lock. Returns true if the upgrade succeeded and false otherwise. 616157882Sjhb */ 617157882Sjhbint 618157882Sjhb_rw_try_upgrade(struct rwlock *rw, const char *file, int line) 619157882Sjhb{ 620157882Sjhb uintptr_t v, tid; 621157882Sjhb int success; 622157882Sjhb 623157882Sjhb _rw_assert(rw, RA_RLOCKED, file, line); 624157882Sjhb 625157882Sjhb /* 626157882Sjhb * Attempt to switch from one reader to a writer. If there 627157882Sjhb * are any write waiters, then we will have to lock the 628157882Sjhb * turnstile first to prevent races with another writer 629157882Sjhb * calling turnstile_wait() before we have claimed this 630157882Sjhb * turnstile. So, do the simple case of no waiters first. 631157882Sjhb */ 632157882Sjhb tid = (uintptr_t)curthread; 633157882Sjhb if (!(rw->rw_lock & RW_LOCK_WRITE_WAITERS)) { 634157882Sjhb success = atomic_cmpset_acq_ptr(&rw->rw_lock, 635157882Sjhb RW_READERS_LOCK(1), tid); 636157882Sjhb goto out; 637157882Sjhb } 638157882Sjhb 639157882Sjhb /* 640157882Sjhb * Ok, we think we have write waiters, so lock the 641157882Sjhb * turnstile. 642157882Sjhb */ 643157882Sjhb turnstile_lock(&rw->rw_object); 644157882Sjhb 645157882Sjhb /* 646157882Sjhb * Try to switch from one reader to a writer again. This time 647157882Sjhb * we honor the current state of the RW_LOCK_WRITE_WAITERS 648157882Sjhb * flag. If we obtain the lock with the flag set, then claim 649157882Sjhb * ownership of the turnstile. In the SMP case it is possible 650157882Sjhb * for there to not be an associated turnstile even though there 651157882Sjhb * are waiters if all of the waiters are spinning. 652157882Sjhb */ 653157882Sjhb v = rw->rw_lock & RW_LOCK_WRITE_WAITERS; 654157882Sjhb success = atomic_cmpset_acq_ptr(&rw->rw_lock, RW_READERS_LOCK(1) | v, 655157882Sjhb tid | v); 656157882Sjhb#ifdef SMP 657157882Sjhb if (success && v && turnstile_lookup(&rw->rw_object) != NULL) 658157882Sjhb#else 659157882Sjhb if (success && v) 660157882Sjhb#endif 661157882Sjhb turnstile_claim(&rw->rw_object); 662157882Sjhb else 663157882Sjhb turnstile_release(&rw->rw_object); 664157882Sjhbout: 665157882Sjhb LOCK_LOG_TRY("WUPGRADE", &rw->rw_object, 0, success, file, line); 666157882Sjhb if (success) 667157882Sjhb WITNESS_UPGRADE(&rw->rw_object, LOP_EXCLUSIVE | LOP_TRYLOCK, 668157882Sjhb file, line); 669157882Sjhb return (success); 670157882Sjhb} 671157882Sjhb 672157882Sjhb/* 673157882Sjhb * Downgrade a write lock into a single read lock. 674157882Sjhb */ 675157882Sjhbvoid 676157882Sjhb_rw_downgrade(struct rwlock *rw, const char *file, int line) 677157882Sjhb{ 678157882Sjhb struct turnstile *ts; 679157882Sjhb uintptr_t tid, v; 680157882Sjhb 681157882Sjhb _rw_assert(rw, RA_WLOCKED, file, line); 682157882Sjhb 683157882Sjhb WITNESS_DOWNGRADE(&rw->rw_object, 0, file, line); 684157882Sjhb 685157882Sjhb /* 686157882Sjhb * Convert from a writer to a single reader. First we handle 687157882Sjhb * the easy case with no waiters. If there are any waiters, we 688157882Sjhb * lock the turnstile, "disown" the lock, and awaken any read 689157882Sjhb * waiters. 690157882Sjhb */ 691157882Sjhb tid = (uintptr_t)curthread; 692157882Sjhb if (atomic_cmpset_rel_ptr(&rw->rw_lock, tid, RW_READERS_LOCK(1))) 693157882Sjhb goto out; 694157882Sjhb 695157882Sjhb /* 696157882Sjhb * Ok, we think we have waiters, so lock the turnstile so we can 697157882Sjhb * read the waiter flags without any races. 698157882Sjhb */ 699157882Sjhb turnstile_lock(&rw->rw_object); 700157882Sjhb v = rw->rw_lock; 701157882Sjhb MPASS(v & (RW_LOCK_READ_WAITERS | RW_LOCK_WRITE_WAITERS)); 702157882Sjhb 703157882Sjhb /* 704157882Sjhb * Downgrade from a write lock while preserving 705157882Sjhb * RW_LOCK_WRITE_WAITERS and give up ownership of the 706157882Sjhb * turnstile. If there are any read waiters, wake them up. 707157882Sjhb * 708157882Sjhb * For SMP, we have to allow for the fact that all of the 709157882Sjhb * read waiters might be spinning. In that case, act as if 710157882Sjhb * RW_LOCK_READ_WAITERS is not set. Also, only preserve 711157882Sjhb * the RW_LOCK_WRITE_WAITERS flag if at least one writer is 712157882Sjhb * blocked on the turnstile. 713157882Sjhb */ 714157882Sjhb ts = turnstile_lookup(&rw->rw_object); 715157882Sjhb#ifdef SMP 716157882Sjhb if (ts == NULL) 717157882Sjhb v &= ~(RW_LOCK_READ_WAITERS | RW_LOCK_WRITE_WAITERS); 718157882Sjhb else if (v & RW_LOCK_READ_WAITERS && 719157882Sjhb turnstile_empty(ts, TS_SHARED_QUEUE)) 720157882Sjhb v &= ~RW_LOCK_READ_WAITERS; 721157882Sjhb else if (v & RW_LOCK_WRITE_WAITERS && 722157882Sjhb turnstile_empty(ts, TS_EXCLUSIVE_QUEUE)) 723157882Sjhb v &= ~RW_LOCK_WRITE_WAITERS; 724157882Sjhb#else 725157882Sjhb MPASS(ts != NULL); 726157882Sjhb#endif 727157882Sjhb if (v & RW_LOCK_READ_WAITERS) 728157882Sjhb turnstile_broadcast(ts, TS_SHARED_QUEUE); 729157882Sjhb atomic_store_rel_ptr(&rw->rw_lock, RW_READERS_LOCK(1) | 730157882Sjhb (v & RW_LOCK_WRITE_WAITERS)); 731157882Sjhb if (v & RW_LOCK_READ_WAITERS) 732157882Sjhb turnstile_unpend(ts, TS_EXCLUSIVE_LOCK); 733157882Sjhb#ifdef SMP 734157882Sjhb else if (ts == NULL) 735157882Sjhb turnstile_release(&rw->rw_object); 736157882Sjhb#endif 737157882Sjhb else 738157882Sjhb turnstile_disown(ts); 739157882Sjhbout: 740157882Sjhb LOCK_LOG_LOCK("WDOWNGRADE", &rw->rw_object, 0, 0, file, line); 741157882Sjhb} 742157882Sjhb 743154941Sjhb#ifdef INVARIANT_SUPPORT 744155162Sscottl#ifndef INVARIANTS 745154941Sjhb#undef _rw_assert 746154941Sjhb#endif 747154941Sjhb 748154941Sjhb/* 749154941Sjhb * In the non-WITNESS case, rw_assert() can only detect that at least 750154941Sjhb * *some* thread owns an rlock, but it cannot guarantee that *this* 751154941Sjhb * thread owns an rlock. 752154941Sjhb */ 753154941Sjhbvoid 754154941Sjhb_rw_assert(struct rwlock *rw, int what, const char *file, int line) 755154941Sjhb{ 756154941Sjhb 757154941Sjhb if (panicstr != NULL) 758154941Sjhb return; 759154941Sjhb switch (what) { 760154941Sjhb case RA_LOCKED: 761154941Sjhb case RA_RLOCKED: 762154941Sjhb#ifdef WITNESS 763154941Sjhb witness_assert(&rw->rw_object, what, file, line); 764154941Sjhb#else 765154941Sjhb /* 766154941Sjhb * If some other thread has a write lock or we have one 767154941Sjhb * and are asserting a read lock, fail. Also, if no one 768154941Sjhb * has a lock at all, fail. 769154941Sjhb */ 770155061Sscottl if (rw->rw_lock == RW_UNLOCKED || 771155061Sscottl (!(rw->rw_lock & RW_LOCK_READ) && (what == RA_RLOCKED || 772157826Sjhb rw_wowner(rw) != curthread))) 773154941Sjhb panic("Lock %s not %slocked @ %s:%d\n", 774155012Sscottl rw->rw_object.lo_name, (what == RA_RLOCKED) ? 775154941Sjhb "read " : "", file, line); 776154941Sjhb#endif 777154941Sjhb break; 778154941Sjhb case RA_WLOCKED: 779157826Sjhb if (rw_wowner(rw) != curthread) 780154941Sjhb panic("Lock %s not exclusively locked @ %s:%d\n", 781154941Sjhb rw->rw_object.lo_name, file, line); 782154941Sjhb break; 783154941Sjhb case RA_UNLOCKED: 784154941Sjhb#ifdef WITNESS 785154941Sjhb witness_assert(&rw->rw_object, what, file, line); 786154941Sjhb#else 787154941Sjhb /* 788154941Sjhb * If we hold a write lock fail. We can't reliably check 789154941Sjhb * to see if we hold a read lock or not. 790154941Sjhb */ 791157826Sjhb if (rw_wowner(rw) == curthread) 792154941Sjhb panic("Lock %s exclusively locked @ %s:%d\n", 793154941Sjhb rw->rw_object.lo_name, file, line); 794154941Sjhb#endif 795154941Sjhb break; 796154941Sjhb default: 797154941Sjhb panic("Unknown rw lock assertion: %d @ %s:%d", what, file, 798154941Sjhb line); 799154941Sjhb } 800154941Sjhb} 801154941Sjhb#endif /* INVARIANT_SUPPORT */ 802154941Sjhb 803154941Sjhb#ifdef DDB 804154941Sjhbvoid 805154941Sjhbdb_show_rwlock(struct lock_object *lock) 806154941Sjhb{ 807154941Sjhb struct rwlock *rw; 808154941Sjhb struct thread *td; 809154941Sjhb 810154941Sjhb rw = (struct rwlock *)lock; 811154941Sjhb 812154941Sjhb db_printf(" state: "); 813154941Sjhb if (rw->rw_lock == RW_UNLOCKED) 814154941Sjhb db_printf("UNLOCKED\n"); 815154941Sjhb else if (rw->rw_lock & RW_LOCK_READ) 816154973Smlaier db_printf("RLOCK: %jd locks\n", 817154973Smlaier (intmax_t)(RW_READERS(rw->rw_lock))); 818154941Sjhb else { 819157826Sjhb td = rw_wowner(rw); 820154941Sjhb db_printf("WLOCK: %p (tid %d, pid %d, \"%s\")\n", td, 821154941Sjhb td->td_tid, td->td_proc->p_pid, td->td_proc->p_comm); 822154941Sjhb } 823154941Sjhb db_printf(" waiters: "); 824154941Sjhb switch (rw->rw_lock & (RW_LOCK_READ_WAITERS | RW_LOCK_WRITE_WAITERS)) { 825154941Sjhb case RW_LOCK_READ_WAITERS: 826154941Sjhb db_printf("readers\n"); 827154941Sjhb break; 828154941Sjhb case RW_LOCK_WRITE_WAITERS: 829154941Sjhb db_printf("writers\n"); 830154941Sjhb break; 831154941Sjhb case RW_LOCK_READ_WAITERS | RW_LOCK_WRITE_WAITERS: 832154941Sjhb db_printf("readers and waiters\n"); 833154941Sjhb break; 834154941Sjhb default: 835154941Sjhb db_printf("none\n"); 836154941Sjhb break; 837154941Sjhb } 838154941Sjhb} 839154941Sjhb 840154941Sjhb#endif 841