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 * 14154941Sjhb * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND 15154941Sjhb * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 16154941Sjhb * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 17154941Sjhb * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE 18154941Sjhb * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 19154941Sjhb * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 20154941Sjhb * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 21154941Sjhb * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 22154941Sjhb * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 23154941Sjhb * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 24154941Sjhb * SUCH DAMAGE. 25154941Sjhb */ 26154941Sjhb 27154941Sjhb/* 28154941Sjhb * Machine independent bits of reader/writer lock implementation. 29154941Sjhb */ 30154941Sjhb 31154941Sjhb#include <sys/cdefs.h> 32154941Sjhb__FBSDID("$FreeBSD$"); 33154941Sjhb 34154941Sjhb#include "opt_ddb.h" 35236238Sfabient#include "opt_hwpmc_hooks.h" 36192853Ssson#include "opt_kdtrace.h" 37167801Sjhb#include "opt_no_adaptive_rwlocks.h" 38154941Sjhb 39154941Sjhb#include <sys/param.h> 40255862Sjhb#include <sys/kdb.h> 41154941Sjhb#include <sys/ktr.h> 42177912Sjeff#include <sys/kernel.h> 43154941Sjhb#include <sys/lock.h> 44154941Sjhb#include <sys/mutex.h> 45154941Sjhb#include <sys/proc.h> 46154941Sjhb#include <sys/rwlock.h> 47177912Sjeff#include <sys/sysctl.h> 48154941Sjhb#include <sys/systm.h> 49154941Sjhb#include <sys/turnstile.h> 50171516Sattilio 51154941Sjhb#include <machine/cpu.h> 52154941Sjhb 53167801Sjhb#if defined(SMP) && !defined(NO_ADAPTIVE_RWLOCKS) 54167801Sjhb#define ADAPTIVE_RWLOCKS 55167801Sjhb#endif 56167801Sjhb 57236238Sfabient#ifdef HWPMC_HOOKS 58236238Sfabient#include <sys/pmckern.h> 59236238SfabientPMC_SOFT_DECLARE( , , lock, failed); 60236238Sfabient#endif 61236238Sfabient 62177912Sjeff#ifdef ADAPTIVE_RWLOCKS 63226255Sattilio#define ROWNER_RETRIES 10 64226255Sattilio#define ROWNER_LOOPS 10000 65177912Sjeff#endif 66177912Sjeff 67154941Sjhb#ifdef DDB 68154941Sjhb#include <ddb/ddb.h> 69154941Sjhb 70154941Sjhbstatic void db_show_rwlock(struct lock_object *lock); 71154941Sjhb#endif 72173733Sattiliostatic void assert_rw(struct lock_object *lock, int what); 73167368Sjhbstatic void lock_rw(struct lock_object *lock, int how); 74192853Ssson#ifdef KDTRACE_HOOKS 75192853Sssonstatic int owner_rw(struct lock_object *lock, struct thread **owner); 76192853Ssson#endif 77167368Sjhbstatic int unlock_rw(struct lock_object *lock); 78154941Sjhb 79154941Sjhbstruct lock_class lock_class_rw = { 80167365Sjhb .lc_name = "rw", 81167365Sjhb .lc_flags = LC_SLEEPLOCK | LC_RECURSABLE | LC_UPGRADABLE, 82173733Sattilio .lc_assert = assert_rw, 83154941Sjhb#ifdef DDB 84167365Sjhb .lc_ddb_show = db_show_rwlock, 85154941Sjhb#endif 86167368Sjhb .lc_lock = lock_rw, 87167368Sjhb .lc_unlock = unlock_rw, 88192853Ssson#ifdef KDTRACE_HOOKS 89192853Ssson .lc_owner = owner_rw, 90192853Ssson#endif 91154941Sjhb}; 92154941Sjhb 93157826Sjhb/* 94157826Sjhb * Return a pointer to the owning thread if the lock is write-locked or 95157826Sjhb * NULL if the lock is unlocked or read-locked. 96157826Sjhb */ 97157826Sjhb#define rw_wowner(rw) \ 98154941Sjhb ((rw)->rw_lock & RW_LOCK_READ ? NULL : \ 99154941Sjhb (struct thread *)RW_OWNER((rw)->rw_lock)) 100154941Sjhb 101157826Sjhb/* 102171052Sattilio * Returns if a write owner is recursed. Write ownership is not assured 103171052Sattilio * here and should be previously checked. 104171052Sattilio */ 105171052Sattilio#define rw_recursed(rw) ((rw)->rw_recurse != 0) 106171052Sattilio 107171052Sattilio/* 108171052Sattilio * Return true if curthread helds the lock. 109171052Sattilio */ 110171052Sattilio#define rw_wlocked(rw) (rw_wowner((rw)) == curthread) 111171052Sattilio 112171052Sattilio/* 113157826Sjhb * Return a pointer to the owning thread for this lock who should receive 114157826Sjhb * any priority lent by threads that block on this lock. Currently this 115157826Sjhb * is identical to rw_wowner(). 116157826Sjhb */ 117157826Sjhb#define rw_owner(rw) rw_wowner(rw) 118157826Sjhb 119154941Sjhb#ifndef INVARIANTS 120154941Sjhb#define _rw_assert(rw, what, file, line) 121154941Sjhb#endif 122154941Sjhb 123154941Sjhbvoid 124173733Sattilioassert_rw(struct lock_object *lock, int what) 125173733Sattilio{ 126173733Sattilio 127173733Sattilio rw_assert((struct rwlock *)lock, what); 128173733Sattilio} 129173733Sattilio 130173733Sattiliovoid 131167368Sjhblock_rw(struct lock_object *lock, int how) 132167368Sjhb{ 133167368Sjhb struct rwlock *rw; 134167368Sjhb 135167368Sjhb rw = (struct rwlock *)lock; 136167368Sjhb if (how) 137167368Sjhb rw_wlock(rw); 138167368Sjhb else 139167368Sjhb rw_rlock(rw); 140167368Sjhb} 141167368Sjhb 142167368Sjhbint 143167368Sjhbunlock_rw(struct lock_object *lock) 144167368Sjhb{ 145167368Sjhb struct rwlock *rw; 146167368Sjhb 147167368Sjhb rw = (struct rwlock *)lock; 148167368Sjhb rw_assert(rw, RA_LOCKED | LA_NOTRECURSED); 149167368Sjhb if (rw->rw_lock & RW_LOCK_READ) { 150167368Sjhb rw_runlock(rw); 151167368Sjhb return (0); 152167368Sjhb } else { 153167368Sjhb rw_wunlock(rw); 154167368Sjhb return (1); 155167368Sjhb } 156167368Sjhb} 157167368Sjhb 158192853Ssson#ifdef KDTRACE_HOOKS 159192853Sssonint 160192853Sssonowner_rw(struct lock_object *lock, struct thread **owner) 161192853Ssson{ 162192853Ssson struct rwlock *rw = (struct rwlock *)lock; 163192853Ssson uintptr_t x = rw->rw_lock; 164192853Ssson 165192853Ssson *owner = rw_wowner(rw); 166192853Ssson return ((x & RW_LOCK_READ) != 0 ? (RW_READERS(x) != 0) : 167192853Ssson (*owner != NULL)); 168192853Ssson} 169192853Ssson#endif 170192853Ssson 171167368Sjhbvoid 172171052Sattiliorw_init_flags(struct rwlock *rw, const char *name, int opts) 173154941Sjhb{ 174171052Sattilio int flags; 175154941Sjhb 176171052Sattilio MPASS((opts & ~(RW_DUPOK | RW_NOPROFILE | RW_NOWITNESS | RW_QUIET | 177171052Sattilio RW_RECURSE)) == 0); 178196334Sattilio ASSERT_ATOMIC_LOAD_PTR(rw->rw_lock, 179196334Sattilio ("%s: rw_lock not aligned for %s: %p", __func__, name, 180196334Sattilio &rw->rw_lock)); 181171052Sattilio 182193307Sattilio flags = LO_UPGRADABLE; 183171052Sattilio if (opts & RW_DUPOK) 184171052Sattilio flags |= LO_DUPOK; 185171052Sattilio if (opts & RW_NOPROFILE) 186171052Sattilio flags |= LO_NOPROFILE; 187171052Sattilio if (!(opts & RW_NOWITNESS)) 188171052Sattilio flags |= LO_WITNESS; 189193307Sattilio if (opts & RW_RECURSE) 190193307Sattilio flags |= LO_RECURSABLE; 191171052Sattilio if (opts & RW_QUIET) 192171052Sattilio flags |= LO_QUIET; 193171052Sattilio 194154941Sjhb rw->rw_lock = RW_UNLOCKED; 195171052Sattilio rw->rw_recurse = 0; 196171052Sattilio lock_init(&rw->lock_object, &lock_class_rw, name, NULL, flags); 197154941Sjhb} 198154941Sjhb 199154941Sjhbvoid 200154941Sjhbrw_destroy(struct rwlock *rw) 201154941Sjhb{ 202154941Sjhb 203205626Sbz KASSERT(rw->rw_lock == RW_UNLOCKED, ("rw lock %p not unlocked", rw)); 204205626Sbz KASSERT(rw->rw_recurse == 0, ("rw lock %p still recursed", rw)); 205169394Sjhb rw->rw_lock = RW_DESTROYED; 206167787Sjhb lock_destroy(&rw->lock_object); 207154941Sjhb} 208154941Sjhb 209154941Sjhbvoid 210154941Sjhbrw_sysinit(void *arg) 211154941Sjhb{ 212154941Sjhb struct rw_args *args = arg; 213154941Sjhb 214154941Sjhb rw_init(args->ra_rw, args->ra_desc); 215154941Sjhb} 216154941Sjhb 217185778Skmacyvoid 218185778Skmacyrw_sysinit_flags(void *arg) 219185778Skmacy{ 220185778Skmacy struct rw_args_flags *args = arg; 221185778Skmacy 222185778Skmacy rw_init_flags(args->ra_rw, args->ra_desc, args->ra_flags); 223185778Skmacy} 224185778Skmacy 225167024Srwatsonint 226167024Srwatsonrw_wowned(struct rwlock *rw) 227167024Srwatson{ 228167024Srwatson 229167024Srwatson return (rw_wowner(rw) == curthread); 230167024Srwatson} 231167024Srwatson 232154941Sjhbvoid 233154941Sjhb_rw_wlock(struct rwlock *rw, const char *file, int line) 234154941Sjhb{ 235154941Sjhb 236235404Savg if (SCHEDULER_STOPPED()) 237235404Savg return; 238255862Sjhb KASSERT(kdb_active != 0 || !TD_IS_IDLETHREAD(curthread), 239255862Sjhb ("rw_wlock() by idle thread %p on rwlock %s @ %s:%d", 240255862Sjhb curthread, rw->lock_object.lo_name, file, line)); 241169394Sjhb KASSERT(rw->rw_lock != RW_DESTROYED, 242169394Sjhb ("rw_wlock() of destroyed rwlock @ %s:%d", file, line)); 243167787Sjhb WITNESS_CHECKORDER(&rw->lock_object, LOP_NEWORDER | LOP_EXCLUSIVE, file, 244182914Sjhb line, NULL); 245154941Sjhb __rw_wlock(rw, curthread, file, line); 246171052Sattilio LOCK_LOG_LOCK("WLOCK", &rw->lock_object, 0, rw->rw_recurse, file, line); 247167787Sjhb WITNESS_LOCK(&rw->lock_object, LOP_EXCLUSIVE, file, line); 248160771Sjhb curthread->td_locks++; 249154941Sjhb} 250154941Sjhb 251177843Sattilioint 252177843Sattilio_rw_try_wlock(struct rwlock *rw, const char *file, int line) 253177843Sattilio{ 254177843Sattilio int rval; 255177843Sattilio 256235404Savg if (SCHEDULER_STOPPED()) 257235404Savg return (1); 258235404Savg 259255862Sjhb KASSERT(kdb_active != 0 || !TD_IS_IDLETHREAD(curthread), 260255862Sjhb ("rw_try_wlock() by idle thread %p on rwlock %s @ %s:%d", 261255862Sjhb curthread, rw->lock_object.lo_name, file, line)); 262177843Sattilio KASSERT(rw->rw_lock != RW_DESTROYED, 263177843Sattilio ("rw_try_wlock() of destroyed rwlock @ %s:%d", file, line)); 264177843Sattilio 265193307Sattilio if (rw_wlocked(rw) && 266193307Sattilio (rw->lock_object.lo_flags & LO_RECURSABLE) != 0) { 267177843Sattilio rw->rw_recurse++; 268177843Sattilio rval = 1; 269177843Sattilio } else 270177843Sattilio rval = atomic_cmpset_acq_ptr(&rw->rw_lock, RW_UNLOCKED, 271177843Sattilio (uintptr_t)curthread); 272177843Sattilio 273177843Sattilio LOCK_LOG_TRY("WLOCK", &rw->lock_object, 0, rval, file, line); 274177843Sattilio if (rval) { 275177843Sattilio WITNESS_LOCK(&rw->lock_object, LOP_EXCLUSIVE | LOP_TRYLOCK, 276177843Sattilio file, line); 277177843Sattilio curthread->td_locks++; 278177843Sattilio } 279177843Sattilio return (rval); 280177843Sattilio} 281177843Sattilio 282154941Sjhbvoid 283154941Sjhb_rw_wunlock(struct rwlock *rw, const char *file, int line) 284154941Sjhb{ 285154941Sjhb 286235404Savg if (SCHEDULER_STOPPED()) 287235404Savg return; 288169394Sjhb KASSERT(rw->rw_lock != RW_DESTROYED, 289169394Sjhb ("rw_wunlock() of destroyed rwlock @ %s:%d", file, line)); 290154941Sjhb _rw_assert(rw, RA_WLOCKED, file, line); 291160771Sjhb curthread->td_locks--; 292167787Sjhb WITNESS_UNLOCK(&rw->lock_object, LOP_EXCLUSIVE, file, line); 293171052Sattilio LOCK_LOG_LOCK("WUNLOCK", &rw->lock_object, 0, rw->rw_recurse, file, 294171052Sattilio line); 295171052Sattilio if (!rw_recursed(rw)) 296192853Ssson LOCKSTAT_PROFILE_RELEASE_LOCK(LS_RW_WUNLOCK_RELEASE, rw); 297154941Sjhb __rw_wunlock(rw, curthread, file, line); 298154941Sjhb} 299176017Sjeff/* 300176017Sjeff * Determines whether a new reader can acquire a lock. Succeeds if the 301176017Sjeff * reader already owns a read lock and the lock is locked for read to 302176017Sjeff * prevent deadlock from reader recursion. Also succeeds if the lock 303176017Sjeff * is unlocked and has no writer waiters or spinners. Failing otherwise 304176017Sjeff * prioritizes writers before readers. 305176017Sjeff */ 306176017Sjeff#define RW_CAN_READ(_rw) \ 307176017Sjeff ((curthread->td_rw_rlocks && (_rw) & RW_LOCK_READ) || ((_rw) & \ 308176017Sjeff (RW_LOCK_READ | RW_LOCK_WRITE_WAITERS | RW_LOCK_WRITE_SPINNER)) == \ 309176017Sjeff RW_LOCK_READ) 310154941Sjhb 311154941Sjhbvoid 312154941Sjhb_rw_rlock(struct rwlock *rw, const char *file, int line) 313154941Sjhb{ 314170295Sjeff struct turnstile *ts; 315167801Sjhb#ifdef ADAPTIVE_RWLOCKS 316157846Sjhb volatile struct thread *owner; 317177912Sjeff int spintries = 0; 318177912Sjeff int i; 319157851Swkoszek#endif 320189846Sjeff#ifdef LOCK_PROFILING 321167307Sjhb uint64_t waittime = 0; 322167054Skmacy int contested = 0; 323189846Sjeff#endif 324176017Sjeff uintptr_t v; 325192853Ssson#ifdef KDTRACE_HOOKS 326192853Ssson uint64_t spin_cnt = 0; 327192853Ssson uint64_t sleep_cnt = 0; 328192853Ssson int64_t sleep_time = 0; 329192853Ssson#endif 330154941Sjhb 331235404Savg if (SCHEDULER_STOPPED()) 332235404Savg return; 333235404Savg 334255862Sjhb KASSERT(kdb_active != 0 || !TD_IS_IDLETHREAD(curthread), 335255862Sjhb ("rw_rlock() by idle thread %p on rwlock %s @ %s:%d", 336255862Sjhb curthread, rw->lock_object.lo_name, file, line)); 337169394Sjhb KASSERT(rw->rw_lock != RW_DESTROYED, 338169394Sjhb ("rw_rlock() of destroyed rwlock @ %s:%d", file, line)); 339157826Sjhb KASSERT(rw_wowner(rw) != curthread, 340252172Sjhb ("rw_rlock: wlock already held for %s @ %s:%d", 341167787Sjhb rw->lock_object.lo_name, file, line)); 342182914Sjhb WITNESS_CHECKORDER(&rw->lock_object, LOP_NEWORDER, file, line, NULL); 343154941Sjhb 344154941Sjhb for (;;) { 345192853Ssson#ifdef KDTRACE_HOOKS 346192853Ssson spin_cnt++; 347192853Ssson#endif 348154941Sjhb /* 349154941Sjhb * Handle the easy case. If no other thread has a write 350154941Sjhb * lock, then try to bump up the count of read locks. Note 351154941Sjhb * that we have to preserve the current state of the 352154941Sjhb * RW_LOCK_WRITE_WAITERS flag. If we fail to acquire a 353154941Sjhb * read lock, then rw_lock must have changed, so restart 354154941Sjhb * the loop. Note that this handles the case of a 355154941Sjhb * completely unlocked rwlock since such a lock is encoded 356154941Sjhb * as a read lock with no waiters. 357154941Sjhb */ 358176017Sjeff v = rw->rw_lock; 359176017Sjeff if (RW_CAN_READ(v)) { 360154941Sjhb /* 361154941Sjhb * The RW_LOCK_READ_WAITERS flag should only be set 362176017Sjeff * if the lock has been unlocked and write waiters 363176017Sjeff * were present. 364154941Sjhb */ 365176017Sjeff if (atomic_cmpset_acq_ptr(&rw->rw_lock, v, 366176017Sjeff v + RW_ONE_READER)) { 367167787Sjhb if (LOCK_LOG_TEST(&rw->lock_object, 0)) 368154941Sjhb CTR4(KTR_LOCK, 369154941Sjhb "%s: %p succeed %p -> %p", __func__, 370176017Sjeff rw, (void *)v, 371176017Sjeff (void *)(v + RW_ONE_READER)); 372154941Sjhb break; 373154941Sjhb } 374154941Sjhb continue; 375154941Sjhb } 376236238Sfabient#ifdef HWPMC_HOOKS 377236238Sfabient PMC_SOFT_CALL( , , lock, failed); 378236238Sfabient#endif 379174629Sjeff lock_profile_obtain_lock_failed(&rw->lock_object, 380174629Sjeff &contested, &waittime); 381154941Sjhb 382173960Sattilio#ifdef ADAPTIVE_RWLOCKS 383154941Sjhb /* 384173960Sattilio * If the owner is running on another CPU, spin until 385173960Sattilio * the owner stops running or the state of the lock 386173960Sattilio * changes. 387173960Sattilio */ 388176017Sjeff if ((v & RW_LOCK_READ) == 0) { 389176017Sjeff owner = (struct thread *)RW_OWNER(v); 390176017Sjeff if (TD_IS_RUNNING(owner)) { 391176017Sjeff if (LOCK_LOG_TEST(&rw->lock_object, 0)) 392176017Sjeff CTR3(KTR_LOCK, 393176017Sjeff "%s: spinning on %p held by %p", 394176017Sjeff __func__, rw, owner); 395176017Sjeff while ((struct thread*)RW_OWNER(rw->rw_lock) == 396192853Ssson owner && TD_IS_RUNNING(owner)) { 397176017Sjeff cpu_spinwait(); 398192853Ssson#ifdef KDTRACE_HOOKS 399192853Ssson spin_cnt++; 400192853Ssson#endif 401192853Ssson } 402176017Sjeff continue; 403176017Sjeff } 404226255Sattilio } else if (spintries < ROWNER_RETRIES) { 405177912Sjeff spintries++; 406226255Sattilio for (i = 0; i < ROWNER_LOOPS; i++) { 407177912Sjeff v = rw->rw_lock; 408177912Sjeff if ((v & RW_LOCK_READ) == 0 || RW_CAN_READ(v)) 409177912Sjeff break; 410177912Sjeff cpu_spinwait(); 411177912Sjeff } 412226255Sattilio if (i != ROWNER_LOOPS) 413177912Sjeff continue; 414173960Sattilio } 415173960Sattilio#endif 416173960Sattilio 417173960Sattilio /* 418154941Sjhb * Okay, now it's the hard case. Some other thread already 419176017Sjeff * has a write lock or there are write waiters present, 420176017Sjeff * acquire the turnstile lock so we can begin the process 421176017Sjeff * of blocking. 422154941Sjhb */ 423170295Sjeff ts = turnstile_trywait(&rw->lock_object); 424154941Sjhb 425154941Sjhb /* 426154941Sjhb * The lock might have been released while we spun, so 427176017Sjeff * recheck its state and restart the loop if needed. 428154941Sjhb */ 429176017Sjeff v = rw->rw_lock; 430176017Sjeff if (RW_CAN_READ(v)) { 431170295Sjeff turnstile_cancel(ts); 432154941Sjhb continue; 433154941Sjhb } 434154941Sjhb 435173960Sattilio#ifdef ADAPTIVE_RWLOCKS 436154941Sjhb /* 437193035Sjhb * The current lock owner might have started executing 438193035Sjhb * on another CPU (or the lock could have changed 439193035Sjhb * owners) while we were waiting on the turnstile 440193035Sjhb * chain lock. If so, drop the turnstile lock and try 441193035Sjhb * again. 442173960Sattilio */ 443176017Sjeff if ((v & RW_LOCK_READ) == 0) { 444176017Sjeff owner = (struct thread *)RW_OWNER(v); 445176017Sjeff if (TD_IS_RUNNING(owner)) { 446176017Sjeff turnstile_cancel(ts); 447176017Sjeff continue; 448176017Sjeff } 449173960Sattilio } 450173960Sattilio#endif 451173960Sattilio 452173960Sattilio /* 453176017Sjeff * The lock is held in write mode or it already has waiters. 454154941Sjhb */ 455176017Sjeff MPASS(!RW_CAN_READ(v)); 456176017Sjeff 457176017Sjeff /* 458176017Sjeff * If the RW_LOCK_READ_WAITERS flag is already set, then 459176017Sjeff * we can go ahead and block. If it is not set then try 460176017Sjeff * to set it. If we fail to set it drop the turnstile 461176017Sjeff * lock and restart the loop. 462176017Sjeff */ 463176017Sjeff if (!(v & RW_LOCK_READ_WAITERS)) { 464176017Sjeff if (!atomic_cmpset_ptr(&rw->rw_lock, v, 465176017Sjeff v | RW_LOCK_READ_WAITERS)) { 466170295Sjeff turnstile_cancel(ts); 467157826Sjhb continue; 468157826Sjhb } 469167787Sjhb if (LOCK_LOG_TEST(&rw->lock_object, 0)) 470157826Sjhb CTR2(KTR_LOCK, "%s: %p set read waiters flag", 471157826Sjhb __func__, rw); 472154941Sjhb } 473154941Sjhb 474154941Sjhb /* 475154941Sjhb * We were unable to acquire the lock and the read waiters 476154941Sjhb * flag is set, so we must block on the turnstile. 477154941Sjhb */ 478167787Sjhb if (LOCK_LOG_TEST(&rw->lock_object, 0)) 479154941Sjhb CTR2(KTR_LOCK, "%s: %p blocking on turnstile", __func__, 480154941Sjhb rw); 481192853Ssson#ifdef KDTRACE_HOOKS 482192853Ssson sleep_time -= lockstat_nsecs(); 483192853Ssson#endif 484170295Sjeff turnstile_wait(ts, rw_owner(rw), TS_SHARED_QUEUE); 485192853Ssson#ifdef KDTRACE_HOOKS 486192853Ssson sleep_time += lockstat_nsecs(); 487192853Ssson sleep_cnt++; 488192853Ssson#endif 489167787Sjhb if (LOCK_LOG_TEST(&rw->lock_object, 0)) 490154941Sjhb CTR2(KTR_LOCK, "%s: %p resuming from turnstile", 491154941Sjhb __func__, rw); 492154941Sjhb } 493154941Sjhb 494154941Sjhb /* 495154941Sjhb * TODO: acquire "owner of record" here. Here be turnstile dragons 496154941Sjhb * however. turnstiles don't like owners changing between calls to 497154941Sjhb * turnstile_wait() currently. 498154941Sjhb */ 499192853Ssson LOCKSTAT_PROFILE_OBTAIN_LOCK_SUCCESS(LS_RW_RLOCK_ACQUIRE, rw, contested, 500174629Sjeff waittime, file, line); 501167787Sjhb LOCK_LOG_LOCK("RLOCK", &rw->lock_object, 0, 0, file, line); 502167787Sjhb WITNESS_LOCK(&rw->lock_object, 0, file, line); 503160771Sjhb curthread->td_locks++; 504176017Sjeff curthread->td_rw_rlocks++; 505192853Ssson#ifdef KDTRACE_HOOKS 506192853Ssson if (sleep_time) 507192853Ssson LOCKSTAT_RECORD1(LS_RW_RLOCK_BLOCK, rw, sleep_time); 508192853Ssson 509192853Ssson /* 510192853Ssson * Record only the loops spinning and not sleeping. 511192853Ssson */ 512192853Ssson if (spin_cnt > sleep_cnt) 513192853Ssson LOCKSTAT_RECORD1(LS_RW_RLOCK_SPIN, rw, (spin_cnt - sleep_cnt)); 514192853Ssson#endif 515154941Sjhb} 516154941Sjhb 517177843Sattilioint 518177843Sattilio_rw_try_rlock(struct rwlock *rw, const char *file, int line) 519177843Sattilio{ 520177843Sattilio uintptr_t x; 521177843Sattilio 522235404Savg if (SCHEDULER_STOPPED()) 523235404Savg return (1); 524235404Savg 525255862Sjhb KASSERT(kdb_active != 0 || !TD_IS_IDLETHREAD(curthread), 526255862Sjhb ("rw_try_rlock() by idle thread %p on rwlock %s @ %s:%d", 527255862Sjhb curthread, rw->lock_object.lo_name, file, line)); 528255862Sjhb 529177843Sattilio for (;;) { 530177843Sattilio x = rw->rw_lock; 531177843Sattilio KASSERT(rw->rw_lock != RW_DESTROYED, 532177843Sattilio ("rw_try_rlock() of destroyed rwlock @ %s:%d", file, line)); 533177843Sattilio if (!(x & RW_LOCK_READ)) 534177843Sattilio break; 535177843Sattilio if (atomic_cmpset_acq_ptr(&rw->rw_lock, x, x + RW_ONE_READER)) { 536177843Sattilio LOCK_LOG_TRY("RLOCK", &rw->lock_object, 0, 1, file, 537177843Sattilio line); 538177843Sattilio WITNESS_LOCK(&rw->lock_object, LOP_TRYLOCK, file, line); 539177843Sattilio curthread->td_locks++; 540177843Sattilio curthread->td_rw_rlocks++; 541177843Sattilio return (1); 542177843Sattilio } 543177843Sattilio } 544177843Sattilio 545177843Sattilio LOCK_LOG_TRY("RLOCK", &rw->lock_object, 0, 0, file, line); 546177843Sattilio return (0); 547177843Sattilio} 548177843Sattilio 549154941Sjhbvoid 550154941Sjhb_rw_runlock(struct rwlock *rw, const char *file, int line) 551154941Sjhb{ 552154941Sjhb struct turnstile *ts; 553176017Sjeff uintptr_t x, v, queue; 554154941Sjhb 555235404Savg if (SCHEDULER_STOPPED()) 556235404Savg return; 557235404Savg 558169394Sjhb KASSERT(rw->rw_lock != RW_DESTROYED, 559169394Sjhb ("rw_runlock() of destroyed rwlock @ %s:%d", file, line)); 560154941Sjhb _rw_assert(rw, RA_RLOCKED, file, line); 561160771Sjhb curthread->td_locks--; 562176017Sjeff curthread->td_rw_rlocks--; 563167787Sjhb WITNESS_UNLOCK(&rw->lock_object, 0, file, line); 564167787Sjhb LOCK_LOG_LOCK("RUNLOCK", &rw->lock_object, 0, 0, file, line); 565154941Sjhb 566154941Sjhb /* TODO: drop "owner of record" here. */ 567154941Sjhb 568154941Sjhb for (;;) { 569154941Sjhb /* 570154941Sjhb * See if there is more than one read lock held. If so, 571154941Sjhb * just drop one and return. 572154941Sjhb */ 573154941Sjhb x = rw->rw_lock; 574154941Sjhb if (RW_READERS(x) > 1) { 575197643Sattilio if (atomic_cmpset_rel_ptr(&rw->rw_lock, x, 576154941Sjhb x - RW_ONE_READER)) { 577167787Sjhb if (LOCK_LOG_TEST(&rw->lock_object, 0)) 578154941Sjhb CTR4(KTR_LOCK, 579154941Sjhb "%s: %p succeeded %p -> %p", 580154941Sjhb __func__, rw, (void *)x, 581154941Sjhb (void *)(x - RW_ONE_READER)); 582154941Sjhb break; 583154941Sjhb } 584154941Sjhb continue; 585167307Sjhb } 586154941Sjhb /* 587154941Sjhb * If there aren't any waiters for a write lock, then try 588154941Sjhb * to drop it quickly. 589154941Sjhb */ 590176017Sjeff if (!(x & RW_LOCK_WAITERS)) { 591176017Sjeff MPASS((x & ~RW_LOCK_WRITE_SPINNER) == 592176017Sjeff RW_READERS_LOCK(1)); 593197643Sattilio if (atomic_cmpset_rel_ptr(&rw->rw_lock, x, 594197643Sattilio RW_UNLOCKED)) { 595167787Sjhb if (LOCK_LOG_TEST(&rw->lock_object, 0)) 596154941Sjhb CTR2(KTR_LOCK, "%s: %p last succeeded", 597154941Sjhb __func__, rw); 598154941Sjhb break; 599154941Sjhb } 600154941Sjhb continue; 601154941Sjhb } 602154941Sjhb /* 603176017Sjeff * Ok, we know we have waiters and we think we are the 604176017Sjeff * last reader, so grab the turnstile lock. 605154941Sjhb */ 606170295Sjeff turnstile_chain_lock(&rw->lock_object); 607176017Sjeff v = rw->rw_lock & (RW_LOCK_WAITERS | RW_LOCK_WRITE_SPINNER); 608176017Sjeff MPASS(v & RW_LOCK_WAITERS); 609154941Sjhb 610154941Sjhb /* 611154941Sjhb * Try to drop our lock leaving the lock in a unlocked 612154941Sjhb * state. 613154941Sjhb * 614154941Sjhb * If you wanted to do explicit lock handoff you'd have to 615154941Sjhb * do it here. You'd also want to use turnstile_signal() 616154941Sjhb * and you'd have to handle the race where a higher 617154941Sjhb * priority thread blocks on the write lock before the 618154941Sjhb * thread you wakeup actually runs and have the new thread 619154941Sjhb * "steal" the lock. For now it's a lot simpler to just 620154941Sjhb * wakeup all of the waiters. 621154941Sjhb * 622154941Sjhb * As above, if we fail, then another thread might have 623154941Sjhb * acquired a read lock, so drop the turnstile lock and 624154941Sjhb * restart. 625154941Sjhb */ 626176017Sjeff x = RW_UNLOCKED; 627176017Sjeff if (v & RW_LOCK_WRITE_WAITERS) { 628176017Sjeff queue = TS_EXCLUSIVE_QUEUE; 629176017Sjeff x |= (v & RW_LOCK_READ_WAITERS); 630176017Sjeff } else 631176017Sjeff queue = TS_SHARED_QUEUE; 632197643Sattilio if (!atomic_cmpset_rel_ptr(&rw->rw_lock, RW_READERS_LOCK(1) | v, 633176017Sjeff x)) { 634170295Sjeff turnstile_chain_unlock(&rw->lock_object); 635154941Sjhb continue; 636154941Sjhb } 637167787Sjhb if (LOCK_LOG_TEST(&rw->lock_object, 0)) 638154941Sjhb CTR2(KTR_LOCK, "%s: %p last succeeded with waiters", 639154941Sjhb __func__, rw); 640154941Sjhb 641154941Sjhb /* 642154941Sjhb * Ok. The lock is released and all that's left is to 643154941Sjhb * wake up the waiters. Note that the lock might not be 644154941Sjhb * free anymore, but in that case the writers will just 645154941Sjhb * block again if they run before the new lock holder(s) 646154941Sjhb * release the lock. 647154941Sjhb */ 648167787Sjhb ts = turnstile_lookup(&rw->lock_object); 649157846Sjhb MPASS(ts != NULL); 650176017Sjeff turnstile_broadcast(ts, queue); 651154941Sjhb turnstile_unpend(ts, TS_SHARED_LOCK); 652170295Sjeff turnstile_chain_unlock(&rw->lock_object); 653154941Sjhb break; 654154941Sjhb } 655192853Ssson LOCKSTAT_PROFILE_RELEASE_LOCK(LS_RW_RUNLOCK_RELEASE, rw); 656154941Sjhb} 657154941Sjhb 658154941Sjhb/* 659154941Sjhb * This function is called when we are unable to obtain a write lock on the 660154941Sjhb * first try. This means that at least one other thread holds either a 661154941Sjhb * read or write lock. 662154941Sjhb */ 663154941Sjhbvoid 664154941Sjhb_rw_wlock_hard(struct rwlock *rw, uintptr_t tid, const char *file, int line) 665154941Sjhb{ 666170295Sjeff struct turnstile *ts; 667167801Sjhb#ifdef ADAPTIVE_RWLOCKS 668157846Sjhb volatile struct thread *owner; 669176017Sjeff int spintries = 0; 670176017Sjeff int i; 671157851Swkoszek#endif 672189846Sjeff uintptr_t v, x; 673189846Sjeff#ifdef LOCK_PROFILING 674171516Sattilio uint64_t waittime = 0; 675171516Sattilio int contested = 0; 676189846Sjeff#endif 677192853Ssson#ifdef KDTRACE_HOOKS 678192853Ssson uint64_t spin_cnt = 0; 679192853Ssson uint64_t sleep_cnt = 0; 680192853Ssson int64_t sleep_time = 0; 681192853Ssson#endif 682154941Sjhb 683235404Savg if (SCHEDULER_STOPPED()) 684235404Savg return; 685235404Savg 686171052Sattilio if (rw_wlocked(rw)) { 687193307Sattilio KASSERT(rw->lock_object.lo_flags & LO_RECURSABLE, 688171052Sattilio ("%s: recursing but non-recursive rw %s @ %s:%d\n", 689171052Sattilio __func__, rw->lock_object.lo_name, file, line)); 690171052Sattilio rw->rw_recurse++; 691171052Sattilio if (LOCK_LOG_TEST(&rw->lock_object, 0)) 692171052Sattilio CTR2(KTR_LOCK, "%s: %p recursing", __func__, rw); 693171052Sattilio return; 694171052Sattilio } 695171052Sattilio 696167787Sjhb if (LOCK_LOG_TEST(&rw->lock_object, 0)) 697154941Sjhb CTR5(KTR_LOCK, "%s: %s contested (lock=%p) at %s:%d", __func__, 698167787Sjhb rw->lock_object.lo_name, (void *)rw->rw_lock, file, line); 699154941Sjhb 700154941Sjhb while (!_rw_write_lock(rw, tid)) { 701192853Ssson#ifdef KDTRACE_HOOKS 702192853Ssson spin_cnt++; 703192853Ssson#endif 704236238Sfabient#ifdef HWPMC_HOOKS 705236238Sfabient PMC_SOFT_CALL( , , lock, failed); 706236238Sfabient#endif 707174629Sjeff lock_profile_obtain_lock_failed(&rw->lock_object, 708174629Sjeff &contested, &waittime); 709173960Sattilio#ifdef ADAPTIVE_RWLOCKS 710173960Sattilio /* 711173960Sattilio * If the lock is write locked and the owner is 712173960Sattilio * running on another CPU, spin until the owner stops 713173960Sattilio * running or the state of the lock changes. 714173960Sattilio */ 715173960Sattilio v = rw->rw_lock; 716173960Sattilio owner = (struct thread *)RW_OWNER(v); 717173960Sattilio if (!(v & RW_LOCK_READ) && TD_IS_RUNNING(owner)) { 718173960Sattilio if (LOCK_LOG_TEST(&rw->lock_object, 0)) 719173960Sattilio CTR3(KTR_LOCK, "%s: spinning on %p held by %p", 720173960Sattilio __func__, rw, owner); 721173960Sattilio while ((struct thread*)RW_OWNER(rw->rw_lock) == owner && 722192853Ssson TD_IS_RUNNING(owner)) { 723173960Sattilio cpu_spinwait(); 724192853Ssson#ifdef KDTRACE_HOOKS 725192853Ssson spin_cnt++; 726192853Ssson#endif 727192853Ssson } 728173960Sattilio continue; 729173960Sattilio } 730177912Sjeff if ((v & RW_LOCK_READ) && RW_READERS(v) && 731226255Sattilio spintries < ROWNER_RETRIES) { 732176017Sjeff if (!(v & RW_LOCK_WRITE_SPINNER)) { 733176017Sjeff if (!atomic_cmpset_ptr(&rw->rw_lock, v, 734176017Sjeff v | RW_LOCK_WRITE_SPINNER)) { 735176017Sjeff continue; 736176017Sjeff } 737176017Sjeff } 738176017Sjeff spintries++; 739226255Sattilio for (i = 0; i < ROWNER_LOOPS; i++) { 740176017Sjeff if ((rw->rw_lock & RW_LOCK_WRITE_SPINNER) == 0) 741176017Sjeff break; 742176017Sjeff cpu_spinwait(); 743176017Sjeff } 744192853Ssson#ifdef KDTRACE_HOOKS 745226255Sattilio spin_cnt += ROWNER_LOOPS - i; 746192853Ssson#endif 747226255Sattilio if (i != ROWNER_LOOPS) 748176017Sjeff continue; 749176017Sjeff } 750173960Sattilio#endif 751170295Sjeff ts = turnstile_trywait(&rw->lock_object); 752154941Sjhb v = rw->rw_lock; 753154941Sjhb 754173960Sattilio#ifdef ADAPTIVE_RWLOCKS 755154941Sjhb /* 756193035Sjhb * The current lock owner might have started executing 757193035Sjhb * on another CPU (or the lock could have changed 758193035Sjhb * owners) while we were waiting on the turnstile 759193035Sjhb * chain lock. If so, drop the turnstile lock and try 760193035Sjhb * again. 761173960Sattilio */ 762173960Sattilio if (!(v & RW_LOCK_READ)) { 763173960Sattilio owner = (struct thread *)RW_OWNER(v); 764173960Sattilio if (TD_IS_RUNNING(owner)) { 765173960Sattilio turnstile_cancel(ts); 766173960Sattilio continue; 767173960Sattilio } 768173960Sattilio } 769173960Sattilio#endif 770173960Sattilio /* 771179334Sattilio * Check for the waiters flags about this rwlock. 772179334Sattilio * If the lock was released, without maintain any pending 773179334Sattilio * waiters queue, simply try to acquire it. 774179334Sattilio * If a pending waiters queue is present, claim the lock 775179334Sattilio * ownership and maintain the pending queue. 776154941Sjhb */ 777176017Sjeff x = v & (RW_LOCK_WAITERS | RW_LOCK_WRITE_SPINNER); 778176017Sjeff if ((v & ~x) == RW_UNLOCKED) { 779176017Sjeff x &= ~RW_LOCK_WRITE_SPINNER; 780176017Sjeff if (atomic_cmpset_acq_ptr(&rw->rw_lock, v, tid | x)) { 781176017Sjeff if (x) 782176017Sjeff turnstile_claim(ts); 783176017Sjeff else 784176017Sjeff turnstile_cancel(ts); 785154941Sjhb break; 786154941Sjhb } 787170295Sjeff turnstile_cancel(ts); 788154941Sjhb continue; 789154941Sjhb } 790154941Sjhb /* 791154941Sjhb * If the RW_LOCK_WRITE_WAITERS flag isn't set, then try to 792154941Sjhb * set it. If we fail to set it, then loop back and try 793154941Sjhb * again. 794154941Sjhb */ 795157826Sjhb if (!(v & RW_LOCK_WRITE_WAITERS)) { 796157826Sjhb if (!atomic_cmpset_ptr(&rw->rw_lock, v, 797157826Sjhb v | RW_LOCK_WRITE_WAITERS)) { 798170295Sjeff turnstile_cancel(ts); 799157826Sjhb continue; 800157826Sjhb } 801167787Sjhb if (LOCK_LOG_TEST(&rw->lock_object, 0)) 802157826Sjhb CTR2(KTR_LOCK, "%s: %p set write waiters flag", 803157826Sjhb __func__, rw); 804154941Sjhb } 805157846Sjhb /* 806154941Sjhb * We were unable to acquire the lock and the write waiters 807154941Sjhb * flag is set, so we must block on the turnstile. 808154941Sjhb */ 809167787Sjhb if (LOCK_LOG_TEST(&rw->lock_object, 0)) 810154941Sjhb CTR2(KTR_LOCK, "%s: %p blocking on turnstile", __func__, 811154941Sjhb rw); 812192853Ssson#ifdef KDTRACE_HOOKS 813192853Ssson sleep_time -= lockstat_nsecs(); 814192853Ssson#endif 815170295Sjeff turnstile_wait(ts, rw_owner(rw), TS_EXCLUSIVE_QUEUE); 816192853Ssson#ifdef KDTRACE_HOOKS 817192853Ssson sleep_time += lockstat_nsecs(); 818192853Ssson sleep_cnt++; 819192853Ssson#endif 820167787Sjhb if (LOCK_LOG_TEST(&rw->lock_object, 0)) 821154941Sjhb CTR2(KTR_LOCK, "%s: %p resuming from turnstile", 822154941Sjhb __func__, rw); 823176017Sjeff#ifdef ADAPTIVE_RWLOCKS 824176017Sjeff spintries = 0; 825176017Sjeff#endif 826154941Sjhb } 827192853Ssson LOCKSTAT_PROFILE_OBTAIN_LOCK_SUCCESS(LS_RW_WLOCK_ACQUIRE, rw, contested, 828192853Ssson waittime, file, line); 829192853Ssson#ifdef KDTRACE_HOOKS 830192853Ssson if (sleep_time) 831192853Ssson LOCKSTAT_RECORD1(LS_RW_WLOCK_BLOCK, rw, sleep_time); 832192853Ssson 833192853Ssson /* 834192853Ssson * Record only the loops spinning and not sleeping. 835192853Ssson */ 836192853Ssson if (spin_cnt > sleep_cnt) 837192853Ssson LOCKSTAT_RECORD1(LS_RW_WLOCK_SPIN, rw, (spin_cnt - sleep_cnt)); 838192853Ssson#endif 839154941Sjhb} 840154941Sjhb 841154941Sjhb/* 842154941Sjhb * This function is called if the first try at releasing a write lock failed. 843154941Sjhb * This means that one of the 2 waiter bits must be set indicating that at 844154941Sjhb * least one thread is waiting on this lock. 845154941Sjhb */ 846154941Sjhbvoid 847154941Sjhb_rw_wunlock_hard(struct rwlock *rw, uintptr_t tid, const char *file, int line) 848154941Sjhb{ 849154941Sjhb struct turnstile *ts; 850154941Sjhb uintptr_t v; 851154941Sjhb int queue; 852154941Sjhb 853235404Savg if (SCHEDULER_STOPPED()) 854235404Savg return; 855235404Savg 856171052Sattilio if (rw_wlocked(rw) && rw_recursed(rw)) { 857176017Sjeff rw->rw_recurse--; 858171052Sattilio if (LOCK_LOG_TEST(&rw->lock_object, 0)) 859171052Sattilio CTR2(KTR_LOCK, "%s: %p unrecursing", __func__, rw); 860171052Sattilio return; 861171052Sattilio } 862171052Sattilio 863154941Sjhb KASSERT(rw->rw_lock & (RW_LOCK_READ_WAITERS | RW_LOCK_WRITE_WAITERS), 864154941Sjhb ("%s: neither of the waiter flags are set", __func__)); 865154941Sjhb 866167787Sjhb if (LOCK_LOG_TEST(&rw->lock_object, 0)) 867154941Sjhb CTR2(KTR_LOCK, "%s: %p contested", __func__, rw); 868154941Sjhb 869170295Sjeff turnstile_chain_lock(&rw->lock_object); 870167787Sjhb ts = turnstile_lookup(&rw->lock_object); 871154941Sjhb MPASS(ts != NULL); 872154941Sjhb 873154941Sjhb /* 874154941Sjhb * Use the same algo as sx locks for now. Prefer waking up shared 875154941Sjhb * waiters if we have any over writers. This is probably not ideal. 876154941Sjhb * 877154941Sjhb * 'v' is the value we are going to write back to rw_lock. If we 878154941Sjhb * have waiters on both queues, we need to preserve the state of 879154941Sjhb * the waiter flag for the queue we don't wake up. For now this is 880154941Sjhb * hardcoded for the algorithm mentioned above. 881154941Sjhb * 882154941Sjhb * In the case of both readers and writers waiting we wakeup the 883154941Sjhb * readers but leave the RW_LOCK_WRITE_WAITERS flag set. If a 884154941Sjhb * new writer comes in before a reader it will claim the lock up 885154941Sjhb * above. There is probably a potential priority inversion in 886154941Sjhb * there that could be worked around either by waking both queues 887154941Sjhb * of waiters or doing some complicated lock handoff gymnastics. 888154941Sjhb */ 889157846Sjhb v = RW_UNLOCKED; 890176076Sjeff if (rw->rw_lock & RW_LOCK_WRITE_WAITERS) { 891176076Sjeff queue = TS_EXCLUSIVE_QUEUE; 892176076Sjeff v |= (rw->rw_lock & RW_LOCK_READ_WAITERS); 893176076Sjeff } else 894154941Sjhb queue = TS_SHARED_QUEUE; 895157846Sjhb 896157846Sjhb /* Wake up all waiters for the specific queue. */ 897167787Sjhb if (LOCK_LOG_TEST(&rw->lock_object, 0)) 898154941Sjhb CTR3(KTR_LOCK, "%s: %p waking up %s waiters", __func__, rw, 899154941Sjhb queue == TS_SHARED_QUEUE ? "read" : "write"); 900154941Sjhb turnstile_broadcast(ts, queue); 901154941Sjhb atomic_store_rel_ptr(&rw->rw_lock, v); 902154941Sjhb turnstile_unpend(ts, TS_EXCLUSIVE_LOCK); 903170295Sjeff turnstile_chain_unlock(&rw->lock_object); 904154941Sjhb} 905154941Sjhb 906157882Sjhb/* 907157882Sjhb * Attempt to do a non-blocking upgrade from a read lock to a write 908157882Sjhb * lock. This will only succeed if this thread holds a single read 909157882Sjhb * lock. Returns true if the upgrade succeeded and false otherwise. 910157882Sjhb */ 911157882Sjhbint 912157882Sjhb_rw_try_upgrade(struct rwlock *rw, const char *file, int line) 913157882Sjhb{ 914176017Sjeff uintptr_t v, x, tid; 915170295Sjeff struct turnstile *ts; 916157882Sjhb int success; 917157882Sjhb 918235404Savg if (SCHEDULER_STOPPED()) 919235404Savg return (1); 920235404Savg 921169394Sjhb KASSERT(rw->rw_lock != RW_DESTROYED, 922169394Sjhb ("rw_try_upgrade() of destroyed rwlock @ %s:%d", file, line)); 923157882Sjhb _rw_assert(rw, RA_RLOCKED, file, line); 924157882Sjhb 925157882Sjhb /* 926157882Sjhb * Attempt to switch from one reader to a writer. If there 927157882Sjhb * are any write waiters, then we will have to lock the 928157882Sjhb * turnstile first to prevent races with another writer 929157882Sjhb * calling turnstile_wait() before we have claimed this 930157882Sjhb * turnstile. So, do the simple case of no waiters first. 931157882Sjhb */ 932157882Sjhb tid = (uintptr_t)curthread; 933176017Sjeff success = 0; 934176017Sjeff for (;;) { 935176017Sjeff v = rw->rw_lock; 936176017Sjeff if (RW_READERS(v) > 1) 937176017Sjeff break; 938176017Sjeff if (!(v & RW_LOCK_WAITERS)) { 939176017Sjeff success = atomic_cmpset_ptr(&rw->rw_lock, v, tid); 940176017Sjeff if (!success) 941176017Sjeff continue; 942176017Sjeff break; 943176017Sjeff } 944157882Sjhb 945176017Sjeff /* 946176017Sjeff * Ok, we think we have waiters, so lock the turnstile. 947176017Sjeff */ 948176017Sjeff ts = turnstile_trywait(&rw->lock_object); 949176017Sjeff v = rw->rw_lock; 950176017Sjeff if (RW_READERS(v) > 1) { 951176017Sjeff turnstile_cancel(ts); 952176017Sjeff break; 953176017Sjeff } 954176017Sjeff /* 955176017Sjeff * Try to switch from one reader to a writer again. This time 956176017Sjeff * we honor the current state of the waiters flags. 957176017Sjeff * If we obtain the lock with the flags set, then claim 958176017Sjeff * ownership of the turnstile. 959176017Sjeff */ 960176017Sjeff x = rw->rw_lock & RW_LOCK_WAITERS; 961176017Sjeff success = atomic_cmpset_ptr(&rw->rw_lock, v, tid | x); 962176017Sjeff if (success) { 963176017Sjeff if (x) 964176017Sjeff turnstile_claim(ts); 965176017Sjeff else 966176017Sjeff turnstile_cancel(ts); 967176017Sjeff break; 968176017Sjeff } 969170295Sjeff turnstile_cancel(ts); 970176017Sjeff } 971167787Sjhb LOCK_LOG_TRY("WUPGRADE", &rw->lock_object, 0, success, file, line); 972176017Sjeff if (success) { 973176017Sjeff curthread->td_rw_rlocks--; 974167787Sjhb WITNESS_UPGRADE(&rw->lock_object, LOP_EXCLUSIVE | LOP_TRYLOCK, 975157882Sjhb file, line); 976192853Ssson LOCKSTAT_RECORD0(LS_RW_TRYUPGRADE_UPGRADE, rw); 977176017Sjeff } 978157882Sjhb return (success); 979157882Sjhb} 980157882Sjhb 981157882Sjhb/* 982157882Sjhb * Downgrade a write lock into a single read lock. 983157882Sjhb */ 984157882Sjhbvoid 985157882Sjhb_rw_downgrade(struct rwlock *rw, const char *file, int line) 986157882Sjhb{ 987157882Sjhb struct turnstile *ts; 988157882Sjhb uintptr_t tid, v; 989176017Sjeff int rwait, wwait; 990157882Sjhb 991235404Savg if (SCHEDULER_STOPPED()) 992235404Savg return; 993235404Savg 994169394Sjhb KASSERT(rw->rw_lock != RW_DESTROYED, 995169394Sjhb ("rw_downgrade() of destroyed rwlock @ %s:%d", file, line)); 996171052Sattilio _rw_assert(rw, RA_WLOCKED | RA_NOTRECURSED, file, line); 997171052Sattilio#ifndef INVARIANTS 998171052Sattilio if (rw_recursed(rw)) 999171052Sattilio panic("downgrade of a recursed lock"); 1000171052Sattilio#endif 1001157882Sjhb 1002167787Sjhb WITNESS_DOWNGRADE(&rw->lock_object, 0, file, line); 1003157882Sjhb 1004157882Sjhb /* 1005157882Sjhb * Convert from a writer to a single reader. First we handle 1006157882Sjhb * the easy case with no waiters. If there are any waiters, we 1007176017Sjeff * lock the turnstile and "disown" the lock. 1008157882Sjhb */ 1009157882Sjhb tid = (uintptr_t)curthread; 1010157882Sjhb if (atomic_cmpset_rel_ptr(&rw->rw_lock, tid, RW_READERS_LOCK(1))) 1011157882Sjhb goto out; 1012157882Sjhb 1013157882Sjhb /* 1014157882Sjhb * Ok, we think we have waiters, so lock the turnstile so we can 1015157882Sjhb * read the waiter flags without any races. 1016157882Sjhb */ 1017170295Sjeff turnstile_chain_lock(&rw->lock_object); 1018176017Sjeff v = rw->rw_lock & RW_LOCK_WAITERS; 1019176017Sjeff rwait = v & RW_LOCK_READ_WAITERS; 1020176017Sjeff wwait = v & RW_LOCK_WRITE_WAITERS; 1021176017Sjeff MPASS(rwait | wwait); 1022157882Sjhb 1023157882Sjhb /* 1024176017Sjeff * Downgrade from a write lock while preserving waiters flag 1025176017Sjeff * and give up ownership of the turnstile. 1026157882Sjhb */ 1027167787Sjhb ts = turnstile_lookup(&rw->lock_object); 1028157882Sjhb MPASS(ts != NULL); 1029176017Sjeff if (!wwait) 1030176017Sjeff v &= ~RW_LOCK_READ_WAITERS; 1031176017Sjeff atomic_store_rel_ptr(&rw->rw_lock, RW_READERS_LOCK(1) | v); 1032176017Sjeff /* 1033176017Sjeff * Wake other readers if there are no writers pending. Otherwise they 1034176017Sjeff * won't be able to acquire the lock anyway. 1035176017Sjeff */ 1036176017Sjeff if (rwait && !wwait) { 1037157882Sjhb turnstile_broadcast(ts, TS_SHARED_QUEUE); 1038157882Sjhb turnstile_unpend(ts, TS_EXCLUSIVE_LOCK); 1039176017Sjeff } else 1040157882Sjhb turnstile_disown(ts); 1041170295Sjeff turnstile_chain_unlock(&rw->lock_object); 1042157882Sjhbout: 1043176017Sjeff curthread->td_rw_rlocks++; 1044167787Sjhb LOCK_LOG_LOCK("WDOWNGRADE", &rw->lock_object, 0, 0, file, line); 1045192853Ssson LOCKSTAT_RECORD0(LS_RW_DOWNGRADE_DOWNGRADE, rw); 1046157882Sjhb} 1047157882Sjhb 1048154941Sjhb#ifdef INVARIANT_SUPPORT 1049155162Sscottl#ifndef INVARIANTS 1050154941Sjhb#undef _rw_assert 1051154941Sjhb#endif 1052154941Sjhb 1053154941Sjhb/* 1054154941Sjhb * In the non-WITNESS case, rw_assert() can only detect that at least 1055154941Sjhb * *some* thread owns an rlock, but it cannot guarantee that *this* 1056154941Sjhb * thread owns an rlock. 1057154941Sjhb */ 1058154941Sjhbvoid 1059154941Sjhb_rw_assert(struct rwlock *rw, int what, const char *file, int line) 1060154941Sjhb{ 1061154941Sjhb 1062154941Sjhb if (panicstr != NULL) 1063154941Sjhb return; 1064154941Sjhb switch (what) { 1065154941Sjhb case RA_LOCKED: 1066171052Sattilio case RA_LOCKED | RA_RECURSED: 1067171052Sattilio case RA_LOCKED | RA_NOTRECURSED: 1068154941Sjhb case RA_RLOCKED: 1069252172Sjhb case RA_RLOCKED | RA_RECURSED: 1070252172Sjhb case RA_RLOCKED | RA_NOTRECURSED: 1071154941Sjhb#ifdef WITNESS 1072167787Sjhb witness_assert(&rw->lock_object, what, file, line); 1073154941Sjhb#else 1074154941Sjhb /* 1075154941Sjhb * If some other thread has a write lock or we have one 1076154941Sjhb * and are asserting a read lock, fail. Also, if no one 1077154941Sjhb * has a lock at all, fail. 1078154941Sjhb */ 1079155061Sscottl if (rw->rw_lock == RW_UNLOCKED || 1080252172Sjhb (!(rw->rw_lock & RW_LOCK_READ) && (what & RA_RLOCKED || 1081157826Sjhb rw_wowner(rw) != curthread))) 1082154941Sjhb panic("Lock %s not %slocked @ %s:%d\n", 1083252172Sjhb rw->lock_object.lo_name, (what & RA_RLOCKED) ? 1084154941Sjhb "read " : "", file, line); 1085171052Sattilio 1086252172Sjhb if (!(rw->rw_lock & RW_LOCK_READ) && !(what & RA_RLOCKED)) { 1087171052Sattilio if (rw_recursed(rw)) { 1088171052Sattilio if (what & RA_NOTRECURSED) 1089171052Sattilio panic("Lock %s recursed @ %s:%d\n", 1090171052Sattilio rw->lock_object.lo_name, file, 1091171052Sattilio line); 1092171052Sattilio } else if (what & RA_RECURSED) 1093171052Sattilio panic("Lock %s not recursed @ %s:%d\n", 1094171052Sattilio rw->lock_object.lo_name, file, line); 1095171052Sattilio } 1096154941Sjhb#endif 1097154941Sjhb break; 1098154941Sjhb case RA_WLOCKED: 1099171052Sattilio case RA_WLOCKED | RA_RECURSED: 1100171052Sattilio case RA_WLOCKED | RA_NOTRECURSED: 1101157826Sjhb if (rw_wowner(rw) != curthread) 1102154941Sjhb panic("Lock %s not exclusively locked @ %s:%d\n", 1103167787Sjhb rw->lock_object.lo_name, file, line); 1104171052Sattilio if (rw_recursed(rw)) { 1105171052Sattilio if (what & RA_NOTRECURSED) 1106171052Sattilio panic("Lock %s recursed @ %s:%d\n", 1107171052Sattilio rw->lock_object.lo_name, file, line); 1108171052Sattilio } else if (what & RA_RECURSED) 1109171052Sattilio panic("Lock %s not recursed @ %s:%d\n", 1110171052Sattilio rw->lock_object.lo_name, file, line); 1111154941Sjhb break; 1112154941Sjhb case RA_UNLOCKED: 1113154941Sjhb#ifdef WITNESS 1114167787Sjhb witness_assert(&rw->lock_object, what, file, line); 1115154941Sjhb#else 1116154941Sjhb /* 1117154941Sjhb * If we hold a write lock fail. We can't reliably check 1118154941Sjhb * to see if we hold a read lock or not. 1119154941Sjhb */ 1120157826Sjhb if (rw_wowner(rw) == curthread) 1121154941Sjhb panic("Lock %s exclusively locked @ %s:%d\n", 1122167787Sjhb rw->lock_object.lo_name, file, line); 1123154941Sjhb#endif 1124154941Sjhb break; 1125154941Sjhb default: 1126154941Sjhb panic("Unknown rw lock assertion: %d @ %s:%d", what, file, 1127154941Sjhb line); 1128154941Sjhb } 1129154941Sjhb} 1130154941Sjhb#endif /* INVARIANT_SUPPORT */ 1131154941Sjhb 1132154941Sjhb#ifdef DDB 1133154941Sjhbvoid 1134154941Sjhbdb_show_rwlock(struct lock_object *lock) 1135154941Sjhb{ 1136154941Sjhb struct rwlock *rw; 1137154941Sjhb struct thread *td; 1138154941Sjhb 1139154941Sjhb rw = (struct rwlock *)lock; 1140154941Sjhb 1141154941Sjhb db_printf(" state: "); 1142154941Sjhb if (rw->rw_lock == RW_UNLOCKED) 1143154941Sjhb db_printf("UNLOCKED\n"); 1144169394Sjhb else if (rw->rw_lock == RW_DESTROYED) { 1145169394Sjhb db_printf("DESTROYED\n"); 1146169394Sjhb return; 1147169394Sjhb } else if (rw->rw_lock & RW_LOCK_READ) 1148167504Sjhb db_printf("RLOCK: %ju locks\n", 1149167504Sjhb (uintmax_t)(RW_READERS(rw->rw_lock))); 1150154941Sjhb else { 1151157826Sjhb td = rw_wowner(rw); 1152154941Sjhb db_printf("WLOCK: %p (tid %d, pid %d, \"%s\")\n", td, 1153173600Sjulian td->td_tid, td->td_proc->p_pid, td->td_name); 1154171052Sattilio if (rw_recursed(rw)) 1155171052Sattilio db_printf(" recursed: %u\n", rw->rw_recurse); 1156154941Sjhb } 1157154941Sjhb db_printf(" waiters: "); 1158154941Sjhb switch (rw->rw_lock & (RW_LOCK_READ_WAITERS | RW_LOCK_WRITE_WAITERS)) { 1159154941Sjhb case RW_LOCK_READ_WAITERS: 1160154941Sjhb db_printf("readers\n"); 1161154941Sjhb break; 1162154941Sjhb case RW_LOCK_WRITE_WAITERS: 1163154941Sjhb db_printf("writers\n"); 1164154941Sjhb break; 1165154941Sjhb case RW_LOCK_READ_WAITERS | RW_LOCK_WRITE_WAITERS: 1166167492Sjhb db_printf("readers and writers\n"); 1167154941Sjhb break; 1168154941Sjhb default: 1169154941Sjhb db_printf("none\n"); 1170154941Sjhb break; 1171154941Sjhb } 1172154941Sjhb} 1173154941Sjhb 1174154941Sjhb#endif 1175