kern_rwlock.c revision 192853
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 192853 2009-05-26 20:28:22Z sson $"); 36154941Sjhb 37154941Sjhb#include "opt_ddb.h" 38192853Ssson#include "opt_kdtrace.h" 39167801Sjhb#include "opt_no_adaptive_rwlocks.h" 40154941Sjhb 41154941Sjhb#include <sys/param.h> 42154941Sjhb#include <sys/ktr.h> 43177912Sjeff#include <sys/kernel.h> 44154941Sjhb#include <sys/lock.h> 45154941Sjhb#include <sys/mutex.h> 46154941Sjhb#include <sys/proc.h> 47154941Sjhb#include <sys/rwlock.h> 48177912Sjeff#include <sys/sysctl.h> 49154941Sjhb#include <sys/systm.h> 50154941Sjhb#include <sys/turnstile.h> 51171516Sattilio 52154941Sjhb#include <machine/cpu.h> 53154941Sjhb 54171052SattilioCTASSERT((RW_RECURSE & LO_CLASSFLAGS) == RW_RECURSE); 55171052Sattilio 56167801Sjhb#if defined(SMP) && !defined(NO_ADAPTIVE_RWLOCKS) 57167801Sjhb#define ADAPTIVE_RWLOCKS 58167801Sjhb#endif 59167801Sjhb 60177912Sjeff#ifdef ADAPTIVE_RWLOCKS 61177912Sjeffstatic int rowner_retries = 10; 62177912Sjeffstatic int rowner_loops = 10000; 63177912SjeffSYSCTL_NODE(_debug, OID_AUTO, rwlock, CTLFLAG_RD, NULL, "rwlock debugging"); 64177912SjeffSYSCTL_INT(_debug_rwlock, OID_AUTO, retry, CTLFLAG_RW, &rowner_retries, 0, ""); 65177912SjeffSYSCTL_INT(_debug_rwlock, OID_AUTO, loops, CTLFLAG_RW, &rowner_loops, 0, ""); 66177912Sjeff#endif 67177912Sjeff 68154941Sjhb#ifdef DDB 69154941Sjhb#include <ddb/ddb.h> 70154941Sjhb 71154941Sjhbstatic void db_show_rwlock(struct lock_object *lock); 72154941Sjhb#endif 73173733Sattiliostatic void assert_rw(struct lock_object *lock, int what); 74167368Sjhbstatic void lock_rw(struct lock_object *lock, int how); 75192853Ssson#ifdef KDTRACE_HOOKS 76192853Sssonstatic int owner_rw(struct lock_object *lock, struct thread **owner); 77192853Ssson#endif 78167368Sjhbstatic int unlock_rw(struct lock_object *lock); 79154941Sjhb 80154941Sjhbstruct lock_class lock_class_rw = { 81167365Sjhb .lc_name = "rw", 82167365Sjhb .lc_flags = LC_SLEEPLOCK | LC_RECURSABLE | LC_UPGRADABLE, 83173733Sattilio .lc_assert = assert_rw, 84154941Sjhb#ifdef DDB 85167365Sjhb .lc_ddb_show = db_show_rwlock, 86154941Sjhb#endif 87167368Sjhb .lc_lock = lock_rw, 88167368Sjhb .lc_unlock = unlock_rw, 89192853Ssson#ifdef KDTRACE_HOOKS 90192853Ssson .lc_owner = owner_rw, 91192853Ssson#endif 92154941Sjhb}; 93154941Sjhb 94157826Sjhb/* 95157826Sjhb * Return a pointer to the owning thread if the lock is write-locked or 96157826Sjhb * NULL if the lock is unlocked or read-locked. 97157826Sjhb */ 98157826Sjhb#define rw_wowner(rw) \ 99154941Sjhb ((rw)->rw_lock & RW_LOCK_READ ? NULL : \ 100154941Sjhb (struct thread *)RW_OWNER((rw)->rw_lock)) 101154941Sjhb 102157826Sjhb/* 103171052Sattilio * Returns if a write owner is recursed. Write ownership is not assured 104171052Sattilio * here and should be previously checked. 105171052Sattilio */ 106171052Sattilio#define rw_recursed(rw) ((rw)->rw_recurse != 0) 107171052Sattilio 108171052Sattilio/* 109171052Sattilio * Return true if curthread helds the lock. 110171052Sattilio */ 111171052Sattilio#define rw_wlocked(rw) (rw_wowner((rw)) == curthread) 112171052Sattilio 113171052Sattilio/* 114157826Sjhb * Return a pointer to the owning thread for this lock who should receive 115157826Sjhb * any priority lent by threads that block on this lock. Currently this 116157826Sjhb * is identical to rw_wowner(). 117157826Sjhb */ 118157826Sjhb#define rw_owner(rw) rw_wowner(rw) 119157826Sjhb 120154941Sjhb#ifndef INVARIANTS 121154941Sjhb#define _rw_assert(rw, what, file, line) 122154941Sjhb#endif 123154941Sjhb 124154941Sjhbvoid 125173733Sattilioassert_rw(struct lock_object *lock, int what) 126173733Sattilio{ 127173733Sattilio 128173733Sattilio rw_assert((struct rwlock *)lock, what); 129173733Sattilio} 130173733Sattilio 131173733Sattiliovoid 132167368Sjhblock_rw(struct lock_object *lock, int how) 133167368Sjhb{ 134167368Sjhb struct rwlock *rw; 135167368Sjhb 136167368Sjhb rw = (struct rwlock *)lock; 137167368Sjhb if (how) 138167368Sjhb rw_wlock(rw); 139167368Sjhb else 140167368Sjhb rw_rlock(rw); 141167368Sjhb} 142167368Sjhb 143167368Sjhbint 144167368Sjhbunlock_rw(struct lock_object *lock) 145167368Sjhb{ 146167368Sjhb struct rwlock *rw; 147167368Sjhb 148167368Sjhb rw = (struct rwlock *)lock; 149167368Sjhb rw_assert(rw, RA_LOCKED | LA_NOTRECURSED); 150167368Sjhb if (rw->rw_lock & RW_LOCK_READ) { 151167368Sjhb rw_runlock(rw); 152167368Sjhb return (0); 153167368Sjhb } else { 154167368Sjhb rw_wunlock(rw); 155167368Sjhb return (1); 156167368Sjhb } 157167368Sjhb} 158167368Sjhb 159192853Ssson#ifdef KDTRACE_HOOKS 160192853Sssonint 161192853Sssonowner_rw(struct lock_object *lock, struct thread **owner) 162192853Ssson{ 163192853Ssson struct rwlock *rw = (struct rwlock *)lock; 164192853Ssson uintptr_t x = rw->rw_lock; 165192853Ssson 166192853Ssson *owner = rw_wowner(rw); 167192853Ssson return ((x & RW_LOCK_READ) != 0 ? (RW_READERS(x) != 0) : 168192853Ssson (*owner != NULL)); 169192853Ssson} 170192853Ssson#endif 171192853Ssson 172167368Sjhbvoid 173171052Sattiliorw_init_flags(struct rwlock *rw, const char *name, int opts) 174154941Sjhb{ 175171052Sattilio int flags; 176154941Sjhb 177171052Sattilio MPASS((opts & ~(RW_DUPOK | RW_NOPROFILE | RW_NOWITNESS | RW_QUIET | 178171052Sattilio RW_RECURSE)) == 0); 179171052Sattilio 180171052Sattilio flags = LO_UPGRADABLE | LO_RECURSABLE; 181171052Sattilio if (opts & RW_DUPOK) 182171052Sattilio flags |= LO_DUPOK; 183171052Sattilio if (opts & RW_NOPROFILE) 184171052Sattilio flags |= LO_NOPROFILE; 185171052Sattilio if (!(opts & RW_NOWITNESS)) 186171052Sattilio flags |= LO_WITNESS; 187171052Sattilio if (opts & RW_QUIET) 188171052Sattilio flags |= LO_QUIET; 189171052Sattilio flags |= opts & RW_RECURSE; 190171052Sattilio 191154941Sjhb rw->rw_lock = RW_UNLOCKED; 192171052Sattilio rw->rw_recurse = 0; 193171052Sattilio lock_init(&rw->lock_object, &lock_class_rw, name, NULL, flags); 194154941Sjhb} 195154941Sjhb 196154941Sjhbvoid 197154941Sjhbrw_destroy(struct rwlock *rw) 198154941Sjhb{ 199154941Sjhb 200154941Sjhb KASSERT(rw->rw_lock == RW_UNLOCKED, ("rw lock not unlocked")); 201171052Sattilio KASSERT(rw->rw_recurse == 0, ("rw lock still recursed")); 202169394Sjhb rw->rw_lock = RW_DESTROYED; 203167787Sjhb lock_destroy(&rw->lock_object); 204154941Sjhb} 205154941Sjhb 206154941Sjhbvoid 207154941Sjhbrw_sysinit(void *arg) 208154941Sjhb{ 209154941Sjhb struct rw_args *args = arg; 210154941Sjhb 211154941Sjhb rw_init(args->ra_rw, args->ra_desc); 212154941Sjhb} 213154941Sjhb 214185778Skmacyvoid 215185778Skmacyrw_sysinit_flags(void *arg) 216185778Skmacy{ 217185778Skmacy struct rw_args_flags *args = arg; 218185778Skmacy 219185778Skmacy rw_init_flags(args->ra_rw, args->ra_desc, args->ra_flags); 220185778Skmacy} 221185778Skmacy 222167024Srwatsonint 223167024Srwatsonrw_wowned(struct rwlock *rw) 224167024Srwatson{ 225167024Srwatson 226167024Srwatson return (rw_wowner(rw) == curthread); 227167024Srwatson} 228167024Srwatson 229154941Sjhbvoid 230154941Sjhb_rw_wlock(struct rwlock *rw, const char *file, int line) 231154941Sjhb{ 232154941Sjhb 233154941Sjhb MPASS(curthread != NULL); 234169394Sjhb KASSERT(rw->rw_lock != RW_DESTROYED, 235169394Sjhb ("rw_wlock() of destroyed rwlock @ %s:%d", file, line)); 236167787Sjhb WITNESS_CHECKORDER(&rw->lock_object, LOP_NEWORDER | LOP_EXCLUSIVE, file, 237182914Sjhb line, NULL); 238154941Sjhb __rw_wlock(rw, curthread, file, line); 239171052Sattilio LOCK_LOG_LOCK("WLOCK", &rw->lock_object, 0, rw->rw_recurse, file, line); 240167787Sjhb WITNESS_LOCK(&rw->lock_object, LOP_EXCLUSIVE, file, line); 241160771Sjhb curthread->td_locks++; 242154941Sjhb} 243154941Sjhb 244177843Sattilioint 245177843Sattilio_rw_try_wlock(struct rwlock *rw, const char *file, int line) 246177843Sattilio{ 247177843Sattilio int rval; 248177843Sattilio 249177843Sattilio KASSERT(rw->rw_lock != RW_DESTROYED, 250177843Sattilio ("rw_try_wlock() of destroyed rwlock @ %s:%d", file, line)); 251177843Sattilio 252177843Sattilio if (rw_wlocked(rw) && (rw->lock_object.lo_flags & RW_RECURSE) != 0) { 253177843Sattilio rw->rw_recurse++; 254177843Sattilio rval = 1; 255177843Sattilio } else 256177843Sattilio rval = atomic_cmpset_acq_ptr(&rw->rw_lock, RW_UNLOCKED, 257177843Sattilio (uintptr_t)curthread); 258177843Sattilio 259177843Sattilio LOCK_LOG_TRY("WLOCK", &rw->lock_object, 0, rval, file, line); 260177843Sattilio if (rval) { 261177843Sattilio WITNESS_LOCK(&rw->lock_object, LOP_EXCLUSIVE | LOP_TRYLOCK, 262177843Sattilio file, line); 263177843Sattilio curthread->td_locks++; 264177843Sattilio } 265177843Sattilio return (rval); 266177843Sattilio} 267177843Sattilio 268154941Sjhbvoid 269154941Sjhb_rw_wunlock(struct rwlock *rw, const char *file, int line) 270154941Sjhb{ 271154941Sjhb 272154941Sjhb MPASS(curthread != NULL); 273169394Sjhb KASSERT(rw->rw_lock != RW_DESTROYED, 274169394Sjhb ("rw_wunlock() of destroyed rwlock @ %s:%d", file, line)); 275154941Sjhb _rw_assert(rw, RA_WLOCKED, file, line); 276160771Sjhb curthread->td_locks--; 277167787Sjhb WITNESS_UNLOCK(&rw->lock_object, LOP_EXCLUSIVE, file, line); 278171052Sattilio LOCK_LOG_LOCK("WUNLOCK", &rw->lock_object, 0, rw->rw_recurse, file, 279171052Sattilio line); 280171052Sattilio if (!rw_recursed(rw)) 281192853Ssson LOCKSTAT_PROFILE_RELEASE_LOCK(LS_RW_WUNLOCK_RELEASE, rw); 282154941Sjhb __rw_wunlock(rw, curthread, file, line); 283154941Sjhb} 284176017Sjeff/* 285176017Sjeff * Determines whether a new reader can acquire a lock. Succeeds if the 286176017Sjeff * reader already owns a read lock and the lock is locked for read to 287176017Sjeff * prevent deadlock from reader recursion. Also succeeds if the lock 288176017Sjeff * is unlocked and has no writer waiters or spinners. Failing otherwise 289176017Sjeff * prioritizes writers before readers. 290176017Sjeff */ 291176017Sjeff#define RW_CAN_READ(_rw) \ 292176017Sjeff ((curthread->td_rw_rlocks && (_rw) & RW_LOCK_READ) || ((_rw) & \ 293176017Sjeff (RW_LOCK_READ | RW_LOCK_WRITE_WAITERS | RW_LOCK_WRITE_SPINNER)) == \ 294176017Sjeff RW_LOCK_READ) 295154941Sjhb 296154941Sjhbvoid 297154941Sjhb_rw_rlock(struct rwlock *rw, const char *file, int line) 298154941Sjhb{ 299170295Sjeff struct turnstile *ts; 300167801Sjhb#ifdef ADAPTIVE_RWLOCKS 301157846Sjhb volatile struct thread *owner; 302177912Sjeff int spintries = 0; 303177912Sjeff int i; 304157851Swkoszek#endif 305189846Sjeff#ifdef LOCK_PROFILING 306167307Sjhb uint64_t waittime = 0; 307167054Skmacy int contested = 0; 308189846Sjeff#endif 309176017Sjeff uintptr_t v; 310192853Ssson#ifdef KDTRACE_HOOKS 311192853Ssson uint64_t spin_cnt = 0; 312192853Ssson uint64_t sleep_cnt = 0; 313192853Ssson int64_t sleep_time = 0; 314192853Ssson#endif 315154941Sjhb 316169394Sjhb KASSERT(rw->rw_lock != RW_DESTROYED, 317169394Sjhb ("rw_rlock() of destroyed rwlock @ %s:%d", file, line)); 318157826Sjhb KASSERT(rw_wowner(rw) != curthread, 319154941Sjhb ("%s (%s): wlock already held @ %s:%d", __func__, 320167787Sjhb rw->lock_object.lo_name, file, line)); 321182914Sjhb WITNESS_CHECKORDER(&rw->lock_object, LOP_NEWORDER, file, line, NULL); 322154941Sjhb 323154941Sjhb for (;;) { 324192853Ssson#ifdef KDTRACE_HOOKS 325192853Ssson spin_cnt++; 326192853Ssson#endif 327154941Sjhb /* 328154941Sjhb * Handle the easy case. If no other thread has a write 329154941Sjhb * lock, then try to bump up the count of read locks. Note 330154941Sjhb * that we have to preserve the current state of the 331154941Sjhb * RW_LOCK_WRITE_WAITERS flag. If we fail to acquire a 332154941Sjhb * read lock, then rw_lock must have changed, so restart 333154941Sjhb * the loop. Note that this handles the case of a 334154941Sjhb * completely unlocked rwlock since such a lock is encoded 335154941Sjhb * as a read lock with no waiters. 336154941Sjhb */ 337176017Sjeff v = rw->rw_lock; 338176017Sjeff if (RW_CAN_READ(v)) { 339154941Sjhb /* 340154941Sjhb * The RW_LOCK_READ_WAITERS flag should only be set 341176017Sjeff * if the lock has been unlocked and write waiters 342176017Sjeff * were present. 343154941Sjhb */ 344176017Sjeff if (atomic_cmpset_acq_ptr(&rw->rw_lock, v, 345176017Sjeff v + RW_ONE_READER)) { 346167787Sjhb if (LOCK_LOG_TEST(&rw->lock_object, 0)) 347154941Sjhb CTR4(KTR_LOCK, 348154941Sjhb "%s: %p succeed %p -> %p", __func__, 349176017Sjeff rw, (void *)v, 350176017Sjeff (void *)(v + RW_ONE_READER)); 351154941Sjhb break; 352154941Sjhb } 353157846Sjhb cpu_spinwait(); 354154941Sjhb continue; 355154941Sjhb } 356174629Sjeff lock_profile_obtain_lock_failed(&rw->lock_object, 357174629Sjeff &contested, &waittime); 358154941Sjhb 359173960Sattilio#ifdef ADAPTIVE_RWLOCKS 360154941Sjhb /* 361173960Sattilio * If the owner is running on another CPU, spin until 362173960Sattilio * the owner stops running or the state of the lock 363173960Sattilio * changes. 364173960Sattilio */ 365176017Sjeff if ((v & RW_LOCK_READ) == 0) { 366176017Sjeff owner = (struct thread *)RW_OWNER(v); 367176017Sjeff if (TD_IS_RUNNING(owner)) { 368176017Sjeff if (LOCK_LOG_TEST(&rw->lock_object, 0)) 369176017Sjeff CTR3(KTR_LOCK, 370176017Sjeff "%s: spinning on %p held by %p", 371176017Sjeff __func__, rw, owner); 372176017Sjeff while ((struct thread*)RW_OWNER(rw->rw_lock) == 373192853Ssson owner && TD_IS_RUNNING(owner)) { 374176017Sjeff cpu_spinwait(); 375192853Ssson#ifdef KDTRACE_HOOKS 376192853Ssson spin_cnt++; 377192853Ssson#endif 378192853Ssson } 379176017Sjeff continue; 380176017Sjeff } 381177912Sjeff } else if (spintries < rowner_retries) { 382177912Sjeff spintries++; 383177912Sjeff for (i = 0; i < rowner_loops; i++) { 384177912Sjeff v = rw->rw_lock; 385177912Sjeff if ((v & RW_LOCK_READ) == 0 || RW_CAN_READ(v)) 386177912Sjeff break; 387177912Sjeff cpu_spinwait(); 388177912Sjeff } 389177912Sjeff if (i != rowner_loops) 390177912Sjeff continue; 391173960Sattilio } 392173960Sattilio#endif 393173960Sattilio 394173960Sattilio /* 395154941Sjhb * Okay, now it's the hard case. Some other thread already 396176017Sjeff * has a write lock or there are write waiters present, 397176017Sjeff * acquire the turnstile lock so we can begin the process 398176017Sjeff * of blocking. 399154941Sjhb */ 400170295Sjeff ts = turnstile_trywait(&rw->lock_object); 401154941Sjhb 402154941Sjhb /* 403154941Sjhb * The lock might have been released while we spun, so 404176017Sjeff * recheck its state and restart the loop if needed. 405154941Sjhb */ 406176017Sjeff v = rw->rw_lock; 407176017Sjeff if (RW_CAN_READ(v)) { 408170295Sjeff turnstile_cancel(ts); 409157846Sjhb cpu_spinwait(); 410154941Sjhb continue; 411154941Sjhb } 412154941Sjhb 413173960Sattilio#ifdef ADAPTIVE_RWLOCKS 414154941Sjhb /* 415173960Sattilio * If the current owner of the lock is executing on another 416173960Sattilio * CPU quit the hard path and try to spin. 417173960Sattilio */ 418176017Sjeff if ((v & RW_LOCK_READ) == 0) { 419176017Sjeff owner = (struct thread *)RW_OWNER(v); 420176017Sjeff if (TD_IS_RUNNING(owner)) { 421176017Sjeff turnstile_cancel(ts); 422176017Sjeff cpu_spinwait(); 423176017Sjeff continue; 424176017Sjeff } 425173960Sattilio } 426173960Sattilio#endif 427173960Sattilio 428173960Sattilio /* 429176017Sjeff * The lock is held in write mode or it already has waiters. 430154941Sjhb */ 431176017Sjeff MPASS(!RW_CAN_READ(v)); 432176017Sjeff 433176017Sjeff /* 434176017Sjeff * If the RW_LOCK_READ_WAITERS flag is already set, then 435176017Sjeff * we can go ahead and block. If it is not set then try 436176017Sjeff * to set it. If we fail to set it drop the turnstile 437176017Sjeff * lock and restart the loop. 438176017Sjeff */ 439176017Sjeff if (!(v & RW_LOCK_READ_WAITERS)) { 440176017Sjeff if (!atomic_cmpset_ptr(&rw->rw_lock, v, 441176017Sjeff v | RW_LOCK_READ_WAITERS)) { 442170295Sjeff turnstile_cancel(ts); 443157826Sjhb cpu_spinwait(); 444157826Sjhb continue; 445157826Sjhb } 446167787Sjhb if (LOCK_LOG_TEST(&rw->lock_object, 0)) 447157826Sjhb CTR2(KTR_LOCK, "%s: %p set read waiters flag", 448157826Sjhb __func__, rw); 449154941Sjhb } 450154941Sjhb 451154941Sjhb /* 452154941Sjhb * We were unable to acquire the lock and the read waiters 453154941Sjhb * flag is set, so we must block on the turnstile. 454154941Sjhb */ 455167787Sjhb if (LOCK_LOG_TEST(&rw->lock_object, 0)) 456154941Sjhb CTR2(KTR_LOCK, "%s: %p blocking on turnstile", __func__, 457154941Sjhb rw); 458192853Ssson#ifdef KDTRACE_HOOKS 459192853Ssson sleep_time -= lockstat_nsecs(); 460192853Ssson#endif 461170295Sjeff turnstile_wait(ts, rw_owner(rw), TS_SHARED_QUEUE); 462192853Ssson#ifdef KDTRACE_HOOKS 463192853Ssson sleep_time += lockstat_nsecs(); 464192853Ssson sleep_cnt++; 465192853Ssson#endif 466167787Sjhb if (LOCK_LOG_TEST(&rw->lock_object, 0)) 467154941Sjhb CTR2(KTR_LOCK, "%s: %p resuming from turnstile", 468154941Sjhb __func__, rw); 469154941Sjhb } 470154941Sjhb 471154941Sjhb /* 472154941Sjhb * TODO: acquire "owner of record" here. Here be turnstile dragons 473154941Sjhb * however. turnstiles don't like owners changing between calls to 474154941Sjhb * turnstile_wait() currently. 475154941Sjhb */ 476192853Ssson LOCKSTAT_PROFILE_OBTAIN_LOCK_SUCCESS(LS_RW_RLOCK_ACQUIRE, rw, contested, 477174629Sjeff waittime, file, line); 478167787Sjhb LOCK_LOG_LOCK("RLOCK", &rw->lock_object, 0, 0, file, line); 479167787Sjhb WITNESS_LOCK(&rw->lock_object, 0, file, line); 480160771Sjhb curthread->td_locks++; 481176017Sjeff curthread->td_rw_rlocks++; 482192853Ssson#ifdef KDTRACE_HOOKS 483192853Ssson if (sleep_time) 484192853Ssson LOCKSTAT_RECORD1(LS_RW_RLOCK_BLOCK, rw, sleep_time); 485192853Ssson 486192853Ssson /* 487192853Ssson * Record only the loops spinning and not sleeping. 488192853Ssson */ 489192853Ssson if (spin_cnt > sleep_cnt) 490192853Ssson LOCKSTAT_RECORD1(LS_RW_RLOCK_SPIN, rw, (spin_cnt - sleep_cnt)); 491192853Ssson#endif 492154941Sjhb} 493154941Sjhb 494177843Sattilioint 495177843Sattilio_rw_try_rlock(struct rwlock *rw, const char *file, int line) 496177843Sattilio{ 497177843Sattilio uintptr_t x; 498177843Sattilio 499177843Sattilio for (;;) { 500177843Sattilio x = rw->rw_lock; 501177843Sattilio KASSERT(rw->rw_lock != RW_DESTROYED, 502177843Sattilio ("rw_try_rlock() of destroyed rwlock @ %s:%d", file, line)); 503177843Sattilio if (!(x & RW_LOCK_READ)) 504177843Sattilio break; 505177843Sattilio if (atomic_cmpset_acq_ptr(&rw->rw_lock, x, x + RW_ONE_READER)) { 506177843Sattilio LOCK_LOG_TRY("RLOCK", &rw->lock_object, 0, 1, file, 507177843Sattilio line); 508177843Sattilio WITNESS_LOCK(&rw->lock_object, LOP_TRYLOCK, file, line); 509177843Sattilio curthread->td_locks++; 510177843Sattilio curthread->td_rw_rlocks++; 511177843Sattilio return (1); 512177843Sattilio } 513177843Sattilio } 514177843Sattilio 515177843Sattilio LOCK_LOG_TRY("RLOCK", &rw->lock_object, 0, 0, file, line); 516177843Sattilio return (0); 517177843Sattilio} 518177843Sattilio 519154941Sjhbvoid 520154941Sjhb_rw_runlock(struct rwlock *rw, const char *file, int line) 521154941Sjhb{ 522154941Sjhb struct turnstile *ts; 523176017Sjeff uintptr_t x, v, queue; 524154941Sjhb 525169394Sjhb KASSERT(rw->rw_lock != RW_DESTROYED, 526169394Sjhb ("rw_runlock() of destroyed rwlock @ %s:%d", file, line)); 527154941Sjhb _rw_assert(rw, RA_RLOCKED, file, line); 528160771Sjhb curthread->td_locks--; 529176017Sjeff curthread->td_rw_rlocks--; 530167787Sjhb WITNESS_UNLOCK(&rw->lock_object, 0, file, line); 531167787Sjhb LOCK_LOG_LOCK("RUNLOCK", &rw->lock_object, 0, 0, file, line); 532154941Sjhb 533154941Sjhb /* TODO: drop "owner of record" here. */ 534154941Sjhb 535154941Sjhb for (;;) { 536154941Sjhb /* 537154941Sjhb * See if there is more than one read lock held. If so, 538154941Sjhb * just drop one and return. 539154941Sjhb */ 540154941Sjhb x = rw->rw_lock; 541154941Sjhb if (RW_READERS(x) > 1) { 542154941Sjhb if (atomic_cmpset_ptr(&rw->rw_lock, x, 543154941Sjhb x - RW_ONE_READER)) { 544167787Sjhb if (LOCK_LOG_TEST(&rw->lock_object, 0)) 545154941Sjhb CTR4(KTR_LOCK, 546154941Sjhb "%s: %p succeeded %p -> %p", 547154941Sjhb __func__, rw, (void *)x, 548154941Sjhb (void *)(x - RW_ONE_READER)); 549154941Sjhb break; 550154941Sjhb } 551154941Sjhb continue; 552167307Sjhb } 553154941Sjhb /* 554154941Sjhb * If there aren't any waiters for a write lock, then try 555154941Sjhb * to drop it quickly. 556154941Sjhb */ 557176017Sjeff if (!(x & RW_LOCK_WAITERS)) { 558176017Sjeff MPASS((x & ~RW_LOCK_WRITE_SPINNER) == 559176017Sjeff RW_READERS_LOCK(1)); 560176017Sjeff if (atomic_cmpset_ptr(&rw->rw_lock, x, RW_UNLOCKED)) { 561167787Sjhb if (LOCK_LOG_TEST(&rw->lock_object, 0)) 562154941Sjhb CTR2(KTR_LOCK, "%s: %p last succeeded", 563154941Sjhb __func__, rw); 564154941Sjhb break; 565154941Sjhb } 566154941Sjhb continue; 567154941Sjhb } 568154941Sjhb /* 569176017Sjeff * Ok, we know we have waiters and we think we are the 570176017Sjeff * last reader, so grab the turnstile lock. 571154941Sjhb */ 572170295Sjeff turnstile_chain_lock(&rw->lock_object); 573176017Sjeff v = rw->rw_lock & (RW_LOCK_WAITERS | RW_LOCK_WRITE_SPINNER); 574176017Sjeff MPASS(v & RW_LOCK_WAITERS); 575154941Sjhb 576154941Sjhb /* 577154941Sjhb * Try to drop our lock leaving the lock in a unlocked 578154941Sjhb * state. 579154941Sjhb * 580154941Sjhb * If you wanted to do explicit lock handoff you'd have to 581154941Sjhb * do it here. You'd also want to use turnstile_signal() 582154941Sjhb * and you'd have to handle the race where a higher 583154941Sjhb * priority thread blocks on the write lock before the 584154941Sjhb * thread you wakeup actually runs and have the new thread 585154941Sjhb * "steal" the lock. For now it's a lot simpler to just 586154941Sjhb * wakeup all of the waiters. 587154941Sjhb * 588154941Sjhb * As above, if we fail, then another thread might have 589154941Sjhb * acquired a read lock, so drop the turnstile lock and 590154941Sjhb * restart. 591154941Sjhb */ 592176017Sjeff x = RW_UNLOCKED; 593176017Sjeff if (v & RW_LOCK_WRITE_WAITERS) { 594176017Sjeff queue = TS_EXCLUSIVE_QUEUE; 595176017Sjeff x |= (v & RW_LOCK_READ_WAITERS); 596176017Sjeff } else 597176017Sjeff queue = TS_SHARED_QUEUE; 598176017Sjeff if (!atomic_cmpset_ptr(&rw->rw_lock, RW_READERS_LOCK(1) | v, 599176017Sjeff x)) { 600170295Sjeff turnstile_chain_unlock(&rw->lock_object); 601154941Sjhb continue; 602154941Sjhb } 603167787Sjhb if (LOCK_LOG_TEST(&rw->lock_object, 0)) 604154941Sjhb CTR2(KTR_LOCK, "%s: %p last succeeded with waiters", 605154941Sjhb __func__, rw); 606154941Sjhb 607154941Sjhb /* 608154941Sjhb * Ok. The lock is released and all that's left is to 609154941Sjhb * wake up the waiters. Note that the lock might not be 610154941Sjhb * free anymore, but in that case the writers will just 611154941Sjhb * block again if they run before the new lock holder(s) 612154941Sjhb * release the lock. 613154941Sjhb */ 614167787Sjhb ts = turnstile_lookup(&rw->lock_object); 615157846Sjhb MPASS(ts != NULL); 616176017Sjeff turnstile_broadcast(ts, queue); 617154941Sjhb turnstile_unpend(ts, TS_SHARED_LOCK); 618170295Sjeff turnstile_chain_unlock(&rw->lock_object); 619154941Sjhb break; 620154941Sjhb } 621192853Ssson LOCKSTAT_PROFILE_RELEASE_LOCK(LS_RW_RUNLOCK_RELEASE, rw); 622154941Sjhb} 623154941Sjhb 624154941Sjhb/* 625154941Sjhb * This function is called when we are unable to obtain a write lock on the 626154941Sjhb * first try. This means that at least one other thread holds either a 627154941Sjhb * read or write lock. 628154941Sjhb */ 629154941Sjhbvoid 630154941Sjhb_rw_wlock_hard(struct rwlock *rw, uintptr_t tid, const char *file, int line) 631154941Sjhb{ 632170295Sjeff struct turnstile *ts; 633167801Sjhb#ifdef ADAPTIVE_RWLOCKS 634157846Sjhb volatile struct thread *owner; 635176017Sjeff int spintries = 0; 636176017Sjeff int i; 637157851Swkoszek#endif 638189846Sjeff uintptr_t v, x; 639189846Sjeff#ifdef LOCK_PROFILING 640171516Sattilio uint64_t waittime = 0; 641171516Sattilio int contested = 0; 642189846Sjeff#endif 643192853Ssson#ifdef KDTRACE_HOOKS 644192853Ssson uint64_t spin_cnt = 0; 645192853Ssson uint64_t sleep_cnt = 0; 646192853Ssson int64_t sleep_time = 0; 647192853Ssson#endif 648154941Sjhb 649171052Sattilio if (rw_wlocked(rw)) { 650171052Sattilio KASSERT(rw->lock_object.lo_flags & RW_RECURSE, 651171052Sattilio ("%s: recursing but non-recursive rw %s @ %s:%d\n", 652171052Sattilio __func__, rw->lock_object.lo_name, file, line)); 653171052Sattilio rw->rw_recurse++; 654171052Sattilio if (LOCK_LOG_TEST(&rw->lock_object, 0)) 655171052Sattilio CTR2(KTR_LOCK, "%s: %p recursing", __func__, rw); 656171052Sattilio return; 657171052Sattilio } 658171052Sattilio 659167787Sjhb if (LOCK_LOG_TEST(&rw->lock_object, 0)) 660154941Sjhb CTR5(KTR_LOCK, "%s: %s contested (lock=%p) at %s:%d", __func__, 661167787Sjhb rw->lock_object.lo_name, (void *)rw->rw_lock, file, line); 662154941Sjhb 663154941Sjhb while (!_rw_write_lock(rw, tid)) { 664192853Ssson#ifdef KDTRACE_HOOKS 665192853Ssson spin_cnt++; 666192853Ssson#endif 667174629Sjeff lock_profile_obtain_lock_failed(&rw->lock_object, 668174629Sjeff &contested, &waittime); 669173960Sattilio#ifdef ADAPTIVE_RWLOCKS 670173960Sattilio /* 671173960Sattilio * If the lock is write locked and the owner is 672173960Sattilio * running on another CPU, spin until the owner stops 673173960Sattilio * running or the state of the lock changes. 674173960Sattilio */ 675173960Sattilio v = rw->rw_lock; 676173960Sattilio owner = (struct thread *)RW_OWNER(v); 677173960Sattilio if (!(v & RW_LOCK_READ) && TD_IS_RUNNING(owner)) { 678173960Sattilio if (LOCK_LOG_TEST(&rw->lock_object, 0)) 679173960Sattilio CTR3(KTR_LOCK, "%s: spinning on %p held by %p", 680173960Sattilio __func__, rw, owner); 681173960Sattilio while ((struct thread*)RW_OWNER(rw->rw_lock) == owner && 682192853Ssson TD_IS_RUNNING(owner)) { 683173960Sattilio cpu_spinwait(); 684192853Ssson#ifdef KDTRACE_HOOKS 685192853Ssson spin_cnt++; 686192853Ssson#endif 687192853Ssson } 688173960Sattilio continue; 689173960Sattilio } 690177912Sjeff if ((v & RW_LOCK_READ) && RW_READERS(v) && 691177912Sjeff spintries < rowner_retries) { 692176017Sjeff if (!(v & RW_LOCK_WRITE_SPINNER)) { 693176017Sjeff if (!atomic_cmpset_ptr(&rw->rw_lock, v, 694176017Sjeff v | RW_LOCK_WRITE_SPINNER)) { 695176017Sjeff cpu_spinwait(); 696176017Sjeff continue; 697176017Sjeff } 698176017Sjeff } 699176017Sjeff spintries++; 700177912Sjeff for (i = 0; i < rowner_loops; i++) { 701176017Sjeff if ((rw->rw_lock & RW_LOCK_WRITE_SPINNER) == 0) 702176017Sjeff break; 703176017Sjeff cpu_spinwait(); 704176017Sjeff } 705192853Ssson#ifdef KDTRACE_HOOKS 706192853Ssson spin_cnt += rowner_loops - i; 707192853Ssson#endif 708177912Sjeff if (i != rowner_loops) 709176017Sjeff continue; 710176017Sjeff } 711173960Sattilio#endif 712170295Sjeff ts = turnstile_trywait(&rw->lock_object); 713154941Sjhb v = rw->rw_lock; 714154941Sjhb 715173960Sattilio#ifdef ADAPTIVE_RWLOCKS 716154941Sjhb /* 717173960Sattilio * If the current owner of the lock is executing on another 718173960Sattilio * CPU quit the hard path and try to spin. 719173960Sattilio */ 720173960Sattilio if (!(v & RW_LOCK_READ)) { 721173960Sattilio owner = (struct thread *)RW_OWNER(v); 722173960Sattilio if (TD_IS_RUNNING(owner)) { 723173960Sattilio turnstile_cancel(ts); 724173960Sattilio cpu_spinwait(); 725173960Sattilio continue; 726173960Sattilio } 727173960Sattilio } 728173960Sattilio#endif 729173960Sattilio /* 730179334Sattilio * Check for the waiters flags about this rwlock. 731179334Sattilio * If the lock was released, without maintain any pending 732179334Sattilio * waiters queue, simply try to acquire it. 733179334Sattilio * If a pending waiters queue is present, claim the lock 734179334Sattilio * ownership and maintain the pending queue. 735154941Sjhb */ 736176017Sjeff x = v & (RW_LOCK_WAITERS | RW_LOCK_WRITE_SPINNER); 737176017Sjeff if ((v & ~x) == RW_UNLOCKED) { 738176017Sjeff x &= ~RW_LOCK_WRITE_SPINNER; 739176017Sjeff if (atomic_cmpset_acq_ptr(&rw->rw_lock, v, tid | x)) { 740176017Sjeff if (x) 741176017Sjeff turnstile_claim(ts); 742176017Sjeff else 743176017Sjeff turnstile_cancel(ts); 744154941Sjhb break; 745154941Sjhb } 746170295Sjeff turnstile_cancel(ts); 747154941Sjhb cpu_spinwait(); 748154941Sjhb continue; 749154941Sjhb } 750154941Sjhb /* 751154941Sjhb * If the RW_LOCK_WRITE_WAITERS flag isn't set, then try to 752154941Sjhb * set it. If we fail to set it, then loop back and try 753154941Sjhb * again. 754154941Sjhb */ 755157826Sjhb if (!(v & RW_LOCK_WRITE_WAITERS)) { 756157826Sjhb if (!atomic_cmpset_ptr(&rw->rw_lock, v, 757157826Sjhb v | RW_LOCK_WRITE_WAITERS)) { 758170295Sjeff turnstile_cancel(ts); 759157826Sjhb cpu_spinwait(); 760157826Sjhb continue; 761157826Sjhb } 762167787Sjhb if (LOCK_LOG_TEST(&rw->lock_object, 0)) 763157826Sjhb CTR2(KTR_LOCK, "%s: %p set write waiters flag", 764157826Sjhb __func__, rw); 765154941Sjhb } 766157846Sjhb /* 767154941Sjhb * We were unable to acquire the lock and the write waiters 768154941Sjhb * flag is set, so we must block on the turnstile. 769154941Sjhb */ 770167787Sjhb if (LOCK_LOG_TEST(&rw->lock_object, 0)) 771154941Sjhb CTR2(KTR_LOCK, "%s: %p blocking on turnstile", __func__, 772154941Sjhb rw); 773192853Ssson#ifdef KDTRACE_HOOKS 774192853Ssson sleep_time -= lockstat_nsecs(); 775192853Ssson#endif 776170295Sjeff turnstile_wait(ts, rw_owner(rw), TS_EXCLUSIVE_QUEUE); 777192853Ssson#ifdef KDTRACE_HOOKS 778192853Ssson sleep_time += lockstat_nsecs(); 779192853Ssson sleep_cnt++; 780192853Ssson#endif 781167787Sjhb if (LOCK_LOG_TEST(&rw->lock_object, 0)) 782154941Sjhb CTR2(KTR_LOCK, "%s: %p resuming from turnstile", 783154941Sjhb __func__, rw); 784176017Sjeff#ifdef ADAPTIVE_RWLOCKS 785176017Sjeff spintries = 0; 786176017Sjeff#endif 787154941Sjhb } 788192853Ssson LOCKSTAT_PROFILE_OBTAIN_LOCK_SUCCESS(LS_RW_WLOCK_ACQUIRE, rw, contested, 789192853Ssson waittime, file, line); 790192853Ssson#ifdef KDTRACE_HOOKS 791192853Ssson if (sleep_time) 792192853Ssson LOCKSTAT_RECORD1(LS_RW_WLOCK_BLOCK, rw, sleep_time); 793192853Ssson 794192853Ssson /* 795192853Ssson * Record only the loops spinning and not sleeping. 796192853Ssson */ 797192853Ssson if (spin_cnt > sleep_cnt) 798192853Ssson LOCKSTAT_RECORD1(LS_RW_WLOCK_SPIN, rw, (spin_cnt - sleep_cnt)); 799192853Ssson#endif 800154941Sjhb} 801154941Sjhb 802154941Sjhb/* 803154941Sjhb * This function is called if the first try at releasing a write lock failed. 804154941Sjhb * This means that one of the 2 waiter bits must be set indicating that at 805154941Sjhb * least one thread is waiting on this lock. 806154941Sjhb */ 807154941Sjhbvoid 808154941Sjhb_rw_wunlock_hard(struct rwlock *rw, uintptr_t tid, const char *file, int line) 809154941Sjhb{ 810154941Sjhb struct turnstile *ts; 811154941Sjhb uintptr_t v; 812154941Sjhb int queue; 813154941Sjhb 814171052Sattilio if (rw_wlocked(rw) && rw_recursed(rw)) { 815176017Sjeff rw->rw_recurse--; 816171052Sattilio if (LOCK_LOG_TEST(&rw->lock_object, 0)) 817171052Sattilio CTR2(KTR_LOCK, "%s: %p unrecursing", __func__, rw); 818171052Sattilio return; 819171052Sattilio } 820171052Sattilio 821154941Sjhb KASSERT(rw->rw_lock & (RW_LOCK_READ_WAITERS | RW_LOCK_WRITE_WAITERS), 822154941Sjhb ("%s: neither of the waiter flags are set", __func__)); 823154941Sjhb 824167787Sjhb if (LOCK_LOG_TEST(&rw->lock_object, 0)) 825154941Sjhb CTR2(KTR_LOCK, "%s: %p contested", __func__, rw); 826154941Sjhb 827170295Sjeff turnstile_chain_lock(&rw->lock_object); 828167787Sjhb ts = turnstile_lookup(&rw->lock_object); 829154941Sjhb MPASS(ts != NULL); 830154941Sjhb 831154941Sjhb /* 832154941Sjhb * Use the same algo as sx locks for now. Prefer waking up shared 833154941Sjhb * waiters if we have any over writers. This is probably not ideal. 834154941Sjhb * 835154941Sjhb * 'v' is the value we are going to write back to rw_lock. If we 836154941Sjhb * have waiters on both queues, we need to preserve the state of 837154941Sjhb * the waiter flag for the queue we don't wake up. For now this is 838154941Sjhb * hardcoded for the algorithm mentioned above. 839154941Sjhb * 840154941Sjhb * In the case of both readers and writers waiting we wakeup the 841154941Sjhb * readers but leave the RW_LOCK_WRITE_WAITERS flag set. If a 842154941Sjhb * new writer comes in before a reader it will claim the lock up 843154941Sjhb * above. There is probably a potential priority inversion in 844154941Sjhb * there that could be worked around either by waking both queues 845154941Sjhb * of waiters or doing some complicated lock handoff gymnastics. 846154941Sjhb */ 847157846Sjhb v = RW_UNLOCKED; 848176076Sjeff if (rw->rw_lock & RW_LOCK_WRITE_WAITERS) { 849176076Sjeff queue = TS_EXCLUSIVE_QUEUE; 850176076Sjeff v |= (rw->rw_lock & RW_LOCK_READ_WAITERS); 851176076Sjeff } else 852154941Sjhb queue = TS_SHARED_QUEUE; 853157846Sjhb 854157846Sjhb /* Wake up all waiters for the specific queue. */ 855167787Sjhb if (LOCK_LOG_TEST(&rw->lock_object, 0)) 856154941Sjhb CTR3(KTR_LOCK, "%s: %p waking up %s waiters", __func__, rw, 857154941Sjhb queue == TS_SHARED_QUEUE ? "read" : "write"); 858154941Sjhb turnstile_broadcast(ts, queue); 859154941Sjhb atomic_store_rel_ptr(&rw->rw_lock, v); 860154941Sjhb turnstile_unpend(ts, TS_EXCLUSIVE_LOCK); 861170295Sjeff turnstile_chain_unlock(&rw->lock_object); 862154941Sjhb} 863154941Sjhb 864157882Sjhb/* 865157882Sjhb * Attempt to do a non-blocking upgrade from a read lock to a write 866157882Sjhb * lock. This will only succeed if this thread holds a single read 867157882Sjhb * lock. Returns true if the upgrade succeeded and false otherwise. 868157882Sjhb */ 869157882Sjhbint 870157882Sjhb_rw_try_upgrade(struct rwlock *rw, const char *file, int line) 871157882Sjhb{ 872176017Sjeff uintptr_t v, x, tid; 873170295Sjeff struct turnstile *ts; 874157882Sjhb int success; 875157882Sjhb 876169394Sjhb KASSERT(rw->rw_lock != RW_DESTROYED, 877169394Sjhb ("rw_try_upgrade() of destroyed rwlock @ %s:%d", file, line)); 878157882Sjhb _rw_assert(rw, RA_RLOCKED, file, line); 879157882Sjhb 880157882Sjhb /* 881157882Sjhb * Attempt to switch from one reader to a writer. If there 882157882Sjhb * are any write waiters, then we will have to lock the 883157882Sjhb * turnstile first to prevent races with another writer 884157882Sjhb * calling turnstile_wait() before we have claimed this 885157882Sjhb * turnstile. So, do the simple case of no waiters first. 886157882Sjhb */ 887157882Sjhb tid = (uintptr_t)curthread; 888176017Sjeff success = 0; 889176017Sjeff for (;;) { 890176017Sjeff v = rw->rw_lock; 891176017Sjeff if (RW_READERS(v) > 1) 892176017Sjeff break; 893176017Sjeff if (!(v & RW_LOCK_WAITERS)) { 894176017Sjeff success = atomic_cmpset_ptr(&rw->rw_lock, v, tid); 895176017Sjeff if (!success) 896176017Sjeff continue; 897176017Sjeff break; 898176017Sjeff } 899157882Sjhb 900176017Sjeff /* 901176017Sjeff * Ok, we think we have waiters, so lock the turnstile. 902176017Sjeff */ 903176017Sjeff ts = turnstile_trywait(&rw->lock_object); 904176017Sjeff v = rw->rw_lock; 905176017Sjeff if (RW_READERS(v) > 1) { 906176017Sjeff turnstile_cancel(ts); 907176017Sjeff break; 908176017Sjeff } 909176017Sjeff /* 910176017Sjeff * Try to switch from one reader to a writer again. This time 911176017Sjeff * we honor the current state of the waiters flags. 912176017Sjeff * If we obtain the lock with the flags set, then claim 913176017Sjeff * ownership of the turnstile. 914176017Sjeff */ 915176017Sjeff x = rw->rw_lock & RW_LOCK_WAITERS; 916176017Sjeff success = atomic_cmpset_ptr(&rw->rw_lock, v, tid | x); 917176017Sjeff if (success) { 918176017Sjeff if (x) 919176017Sjeff turnstile_claim(ts); 920176017Sjeff else 921176017Sjeff turnstile_cancel(ts); 922176017Sjeff break; 923176017Sjeff } 924170295Sjeff turnstile_cancel(ts); 925176017Sjeff } 926167787Sjhb LOCK_LOG_TRY("WUPGRADE", &rw->lock_object, 0, success, file, line); 927176017Sjeff if (success) { 928176017Sjeff curthread->td_rw_rlocks--; 929167787Sjhb WITNESS_UPGRADE(&rw->lock_object, LOP_EXCLUSIVE | LOP_TRYLOCK, 930157882Sjhb file, line); 931192853Ssson LOCKSTAT_RECORD0(LS_RW_TRYUPGRADE_UPGRADE, rw); 932176017Sjeff } 933157882Sjhb return (success); 934157882Sjhb} 935157882Sjhb 936157882Sjhb/* 937157882Sjhb * Downgrade a write lock into a single read lock. 938157882Sjhb */ 939157882Sjhbvoid 940157882Sjhb_rw_downgrade(struct rwlock *rw, const char *file, int line) 941157882Sjhb{ 942157882Sjhb struct turnstile *ts; 943157882Sjhb uintptr_t tid, v; 944176017Sjeff int rwait, wwait; 945157882Sjhb 946169394Sjhb KASSERT(rw->rw_lock != RW_DESTROYED, 947169394Sjhb ("rw_downgrade() of destroyed rwlock @ %s:%d", file, line)); 948171052Sattilio _rw_assert(rw, RA_WLOCKED | RA_NOTRECURSED, file, line); 949171052Sattilio#ifndef INVARIANTS 950171052Sattilio if (rw_recursed(rw)) 951171052Sattilio panic("downgrade of a recursed lock"); 952171052Sattilio#endif 953157882Sjhb 954167787Sjhb WITNESS_DOWNGRADE(&rw->lock_object, 0, file, line); 955157882Sjhb 956157882Sjhb /* 957157882Sjhb * Convert from a writer to a single reader. First we handle 958157882Sjhb * the easy case with no waiters. If there are any waiters, we 959176017Sjeff * lock the turnstile and "disown" the lock. 960157882Sjhb */ 961157882Sjhb tid = (uintptr_t)curthread; 962157882Sjhb if (atomic_cmpset_rel_ptr(&rw->rw_lock, tid, RW_READERS_LOCK(1))) 963157882Sjhb goto out; 964157882Sjhb 965157882Sjhb /* 966157882Sjhb * Ok, we think we have waiters, so lock the turnstile so we can 967157882Sjhb * read the waiter flags without any races. 968157882Sjhb */ 969170295Sjeff turnstile_chain_lock(&rw->lock_object); 970176017Sjeff v = rw->rw_lock & RW_LOCK_WAITERS; 971176017Sjeff rwait = v & RW_LOCK_READ_WAITERS; 972176017Sjeff wwait = v & RW_LOCK_WRITE_WAITERS; 973176017Sjeff MPASS(rwait | wwait); 974157882Sjhb 975157882Sjhb /* 976176017Sjeff * Downgrade from a write lock while preserving waiters flag 977176017Sjeff * and give up ownership of the turnstile. 978157882Sjhb */ 979167787Sjhb ts = turnstile_lookup(&rw->lock_object); 980157882Sjhb MPASS(ts != NULL); 981176017Sjeff if (!wwait) 982176017Sjeff v &= ~RW_LOCK_READ_WAITERS; 983176017Sjeff atomic_store_rel_ptr(&rw->rw_lock, RW_READERS_LOCK(1) | v); 984176017Sjeff /* 985176017Sjeff * Wake other readers if there are no writers pending. Otherwise they 986176017Sjeff * won't be able to acquire the lock anyway. 987176017Sjeff */ 988176017Sjeff if (rwait && !wwait) { 989157882Sjhb turnstile_broadcast(ts, TS_SHARED_QUEUE); 990157882Sjhb turnstile_unpend(ts, TS_EXCLUSIVE_LOCK); 991176017Sjeff } else 992157882Sjhb turnstile_disown(ts); 993170295Sjeff turnstile_chain_unlock(&rw->lock_object); 994157882Sjhbout: 995176017Sjeff curthread->td_rw_rlocks++; 996167787Sjhb LOCK_LOG_LOCK("WDOWNGRADE", &rw->lock_object, 0, 0, file, line); 997192853Ssson LOCKSTAT_RECORD0(LS_RW_DOWNGRADE_DOWNGRADE, rw); 998157882Sjhb} 999157882Sjhb 1000154941Sjhb#ifdef INVARIANT_SUPPORT 1001155162Sscottl#ifndef INVARIANTS 1002154941Sjhb#undef _rw_assert 1003154941Sjhb#endif 1004154941Sjhb 1005154941Sjhb/* 1006154941Sjhb * In the non-WITNESS case, rw_assert() can only detect that at least 1007154941Sjhb * *some* thread owns an rlock, but it cannot guarantee that *this* 1008154941Sjhb * thread owns an rlock. 1009154941Sjhb */ 1010154941Sjhbvoid 1011154941Sjhb_rw_assert(struct rwlock *rw, int what, const char *file, int line) 1012154941Sjhb{ 1013154941Sjhb 1014154941Sjhb if (panicstr != NULL) 1015154941Sjhb return; 1016154941Sjhb switch (what) { 1017154941Sjhb case RA_LOCKED: 1018171052Sattilio case RA_LOCKED | RA_RECURSED: 1019171052Sattilio case RA_LOCKED | RA_NOTRECURSED: 1020154941Sjhb case RA_RLOCKED: 1021154941Sjhb#ifdef WITNESS 1022167787Sjhb witness_assert(&rw->lock_object, what, file, line); 1023154941Sjhb#else 1024154941Sjhb /* 1025154941Sjhb * If some other thread has a write lock or we have one 1026154941Sjhb * and are asserting a read lock, fail. Also, if no one 1027154941Sjhb * has a lock at all, fail. 1028154941Sjhb */ 1029155061Sscottl if (rw->rw_lock == RW_UNLOCKED || 1030155061Sscottl (!(rw->rw_lock & RW_LOCK_READ) && (what == RA_RLOCKED || 1031157826Sjhb rw_wowner(rw) != curthread))) 1032154941Sjhb panic("Lock %s not %slocked @ %s:%d\n", 1033167787Sjhb rw->lock_object.lo_name, (what == RA_RLOCKED) ? 1034154941Sjhb "read " : "", file, line); 1035171052Sattilio 1036171052Sattilio if (!(rw->rw_lock & RW_LOCK_READ)) { 1037171052Sattilio if (rw_recursed(rw)) { 1038171052Sattilio if (what & RA_NOTRECURSED) 1039171052Sattilio panic("Lock %s recursed @ %s:%d\n", 1040171052Sattilio rw->lock_object.lo_name, file, 1041171052Sattilio line); 1042171052Sattilio } else if (what & RA_RECURSED) 1043171052Sattilio panic("Lock %s not recursed @ %s:%d\n", 1044171052Sattilio rw->lock_object.lo_name, file, line); 1045171052Sattilio } 1046154941Sjhb#endif 1047154941Sjhb break; 1048154941Sjhb case RA_WLOCKED: 1049171052Sattilio case RA_WLOCKED | RA_RECURSED: 1050171052Sattilio case RA_WLOCKED | RA_NOTRECURSED: 1051157826Sjhb if (rw_wowner(rw) != curthread) 1052154941Sjhb panic("Lock %s not exclusively locked @ %s:%d\n", 1053167787Sjhb rw->lock_object.lo_name, file, line); 1054171052Sattilio if (rw_recursed(rw)) { 1055171052Sattilio if (what & RA_NOTRECURSED) 1056171052Sattilio panic("Lock %s recursed @ %s:%d\n", 1057171052Sattilio rw->lock_object.lo_name, file, line); 1058171052Sattilio } else if (what & RA_RECURSED) 1059171052Sattilio panic("Lock %s not recursed @ %s:%d\n", 1060171052Sattilio rw->lock_object.lo_name, file, line); 1061154941Sjhb break; 1062154941Sjhb case RA_UNLOCKED: 1063154941Sjhb#ifdef WITNESS 1064167787Sjhb witness_assert(&rw->lock_object, what, file, line); 1065154941Sjhb#else 1066154941Sjhb /* 1067154941Sjhb * If we hold a write lock fail. We can't reliably check 1068154941Sjhb * to see if we hold a read lock or not. 1069154941Sjhb */ 1070157826Sjhb if (rw_wowner(rw) == curthread) 1071154941Sjhb panic("Lock %s exclusively locked @ %s:%d\n", 1072167787Sjhb rw->lock_object.lo_name, file, line); 1073154941Sjhb#endif 1074154941Sjhb break; 1075154941Sjhb default: 1076154941Sjhb panic("Unknown rw lock assertion: %d @ %s:%d", what, file, 1077154941Sjhb line); 1078154941Sjhb } 1079154941Sjhb} 1080154941Sjhb#endif /* INVARIANT_SUPPORT */ 1081154941Sjhb 1082154941Sjhb#ifdef DDB 1083154941Sjhbvoid 1084154941Sjhbdb_show_rwlock(struct lock_object *lock) 1085154941Sjhb{ 1086154941Sjhb struct rwlock *rw; 1087154941Sjhb struct thread *td; 1088154941Sjhb 1089154941Sjhb rw = (struct rwlock *)lock; 1090154941Sjhb 1091154941Sjhb db_printf(" state: "); 1092154941Sjhb if (rw->rw_lock == RW_UNLOCKED) 1093154941Sjhb db_printf("UNLOCKED\n"); 1094169394Sjhb else if (rw->rw_lock == RW_DESTROYED) { 1095169394Sjhb db_printf("DESTROYED\n"); 1096169394Sjhb return; 1097169394Sjhb } else if (rw->rw_lock & RW_LOCK_READ) 1098167504Sjhb db_printf("RLOCK: %ju locks\n", 1099167504Sjhb (uintmax_t)(RW_READERS(rw->rw_lock))); 1100154941Sjhb else { 1101157826Sjhb td = rw_wowner(rw); 1102154941Sjhb db_printf("WLOCK: %p (tid %d, pid %d, \"%s\")\n", td, 1103173600Sjulian td->td_tid, td->td_proc->p_pid, td->td_name); 1104171052Sattilio if (rw_recursed(rw)) 1105171052Sattilio db_printf(" recursed: %u\n", rw->rw_recurse); 1106154941Sjhb } 1107154941Sjhb db_printf(" waiters: "); 1108154941Sjhb switch (rw->rw_lock & (RW_LOCK_READ_WAITERS | RW_LOCK_WRITE_WAITERS)) { 1109154941Sjhb case RW_LOCK_READ_WAITERS: 1110154941Sjhb db_printf("readers\n"); 1111154941Sjhb break; 1112154941Sjhb case RW_LOCK_WRITE_WAITERS: 1113154941Sjhb db_printf("writers\n"); 1114154941Sjhb break; 1115154941Sjhb case RW_LOCK_READ_WAITERS | RW_LOCK_WRITE_WAITERS: 1116167492Sjhb db_printf("readers and writers\n"); 1117154941Sjhb break; 1118154941Sjhb default: 1119154941Sjhb db_printf("none\n"); 1120154941Sjhb break; 1121154941Sjhb } 1122154941Sjhb} 1123154941Sjhb 1124154941Sjhb#endif 1125