kern_rwlock.c revision 228424
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 228424 2011-12-11 21:02:01Z avg $"); 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 54167801Sjhb#if defined(SMP) && !defined(NO_ADAPTIVE_RWLOCKS) 55167801Sjhb#define ADAPTIVE_RWLOCKS 56167801Sjhb#endif 57167801Sjhb 58177912Sjeff#ifdef ADAPTIVE_RWLOCKS 59177912Sjeffstatic int rowner_retries = 10; 60177912Sjeffstatic int rowner_loops = 10000; 61227309Sedstatic SYSCTL_NODE(_debug, OID_AUTO, rwlock, CTLFLAG_RD, NULL, 62227309Sed "rwlock debugging"); 63177912SjeffSYSCTL_INT(_debug_rwlock, OID_AUTO, retry, CTLFLAG_RW, &rowner_retries, 0, ""); 64177912SjeffSYSCTL_INT(_debug_rwlock, OID_AUTO, loops, CTLFLAG_RW, &rowner_loops, 0, ""); 65177912Sjeff#endif 66177912Sjeff 67154941Sjhb#ifdef DDB 68154941Sjhb#include <ddb/ddb.h> 69154941Sjhb 70227588Spjdstatic void db_show_rwlock(const struct lock_object *lock); 71154941Sjhb#endif 72227588Spjdstatic void assert_rw(const struct lock_object *lock, int what); 73167368Sjhbstatic void lock_rw(struct lock_object *lock, int how); 74192853Ssson#ifdef KDTRACE_HOOKS 75227588Spjdstatic int owner_rw(const 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 124227588Spjdassert_rw(const struct lock_object *lock, int what) 125173733Sattilio{ 126173733Sattilio 127227588Spjd rw_assert((const 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 160227588Spjdowner_rw(const struct lock_object *lock, struct thread **owner) 161192853Ssson{ 162227588Spjd const struct rwlock *rw = (const 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 226227588Spjdrw_wowned(const 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 236228424Savg if (SCHEDULER_STOPPED()) 237228424Savg return; 238154941Sjhb MPASS(curthread != NULL); 239169394Sjhb KASSERT(rw->rw_lock != RW_DESTROYED, 240169394Sjhb ("rw_wlock() of destroyed rwlock @ %s:%d", file, line)); 241167787Sjhb WITNESS_CHECKORDER(&rw->lock_object, LOP_NEWORDER | LOP_EXCLUSIVE, file, 242182914Sjhb line, NULL); 243154941Sjhb __rw_wlock(rw, curthread, file, line); 244171052Sattilio LOCK_LOG_LOCK("WLOCK", &rw->lock_object, 0, rw->rw_recurse, file, line); 245167787Sjhb WITNESS_LOCK(&rw->lock_object, LOP_EXCLUSIVE, file, line); 246160771Sjhb curthread->td_locks++; 247154941Sjhb} 248154941Sjhb 249177843Sattilioint 250177843Sattilio_rw_try_wlock(struct rwlock *rw, const char *file, int line) 251177843Sattilio{ 252177843Sattilio int rval; 253177843Sattilio 254228424Savg if (SCHEDULER_STOPPED()) 255228424Savg return (1); 256228424Savg 257177843Sattilio KASSERT(rw->rw_lock != RW_DESTROYED, 258177843Sattilio ("rw_try_wlock() of destroyed rwlock @ %s:%d", file, line)); 259177843Sattilio 260193307Sattilio if (rw_wlocked(rw) && 261193307Sattilio (rw->lock_object.lo_flags & LO_RECURSABLE) != 0) { 262177843Sattilio rw->rw_recurse++; 263177843Sattilio rval = 1; 264177843Sattilio } else 265177843Sattilio rval = atomic_cmpset_acq_ptr(&rw->rw_lock, RW_UNLOCKED, 266177843Sattilio (uintptr_t)curthread); 267177843Sattilio 268177843Sattilio LOCK_LOG_TRY("WLOCK", &rw->lock_object, 0, rval, file, line); 269177843Sattilio if (rval) { 270177843Sattilio WITNESS_LOCK(&rw->lock_object, LOP_EXCLUSIVE | LOP_TRYLOCK, 271177843Sattilio file, line); 272177843Sattilio curthread->td_locks++; 273177843Sattilio } 274177843Sattilio return (rval); 275177843Sattilio} 276177843Sattilio 277154941Sjhbvoid 278154941Sjhb_rw_wunlock(struct rwlock *rw, const char *file, int line) 279154941Sjhb{ 280154941Sjhb 281228424Savg if (SCHEDULER_STOPPED()) 282228424Savg return; 283154941Sjhb MPASS(curthread != NULL); 284169394Sjhb KASSERT(rw->rw_lock != RW_DESTROYED, 285169394Sjhb ("rw_wunlock() of destroyed rwlock @ %s:%d", file, line)); 286154941Sjhb _rw_assert(rw, RA_WLOCKED, file, line); 287160771Sjhb curthread->td_locks--; 288167787Sjhb WITNESS_UNLOCK(&rw->lock_object, LOP_EXCLUSIVE, file, line); 289171052Sattilio LOCK_LOG_LOCK("WUNLOCK", &rw->lock_object, 0, rw->rw_recurse, file, 290171052Sattilio line); 291171052Sattilio if (!rw_recursed(rw)) 292192853Ssson LOCKSTAT_PROFILE_RELEASE_LOCK(LS_RW_WUNLOCK_RELEASE, rw); 293154941Sjhb __rw_wunlock(rw, curthread, file, line); 294154941Sjhb} 295176017Sjeff/* 296176017Sjeff * Determines whether a new reader can acquire a lock. Succeeds if the 297176017Sjeff * reader already owns a read lock and the lock is locked for read to 298176017Sjeff * prevent deadlock from reader recursion. Also succeeds if the lock 299176017Sjeff * is unlocked and has no writer waiters or spinners. Failing otherwise 300176017Sjeff * prioritizes writers before readers. 301176017Sjeff */ 302176017Sjeff#define RW_CAN_READ(_rw) \ 303176017Sjeff ((curthread->td_rw_rlocks && (_rw) & RW_LOCK_READ) || ((_rw) & \ 304176017Sjeff (RW_LOCK_READ | RW_LOCK_WRITE_WAITERS | RW_LOCK_WRITE_SPINNER)) == \ 305176017Sjeff RW_LOCK_READ) 306154941Sjhb 307154941Sjhbvoid 308154941Sjhb_rw_rlock(struct rwlock *rw, const char *file, int line) 309154941Sjhb{ 310170295Sjeff struct turnstile *ts; 311167801Sjhb#ifdef ADAPTIVE_RWLOCKS 312157846Sjhb volatile struct thread *owner; 313177912Sjeff int spintries = 0; 314177912Sjeff int i; 315157851Swkoszek#endif 316189846Sjeff#ifdef LOCK_PROFILING 317167307Sjhb uint64_t waittime = 0; 318167054Skmacy int contested = 0; 319189846Sjeff#endif 320176017Sjeff uintptr_t v; 321192853Ssson#ifdef KDTRACE_HOOKS 322192853Ssson uint64_t spin_cnt = 0; 323192853Ssson uint64_t sleep_cnt = 0; 324192853Ssson int64_t sleep_time = 0; 325192853Ssson#endif 326154941Sjhb 327228424Savg if (SCHEDULER_STOPPED()) 328228424Savg return; 329228424Savg 330169394Sjhb KASSERT(rw->rw_lock != RW_DESTROYED, 331169394Sjhb ("rw_rlock() of destroyed rwlock @ %s:%d", file, line)); 332157826Sjhb KASSERT(rw_wowner(rw) != curthread, 333154941Sjhb ("%s (%s): wlock already held @ %s:%d", __func__, 334167787Sjhb rw->lock_object.lo_name, file, line)); 335182914Sjhb WITNESS_CHECKORDER(&rw->lock_object, LOP_NEWORDER, file, line, NULL); 336154941Sjhb 337154941Sjhb for (;;) { 338192853Ssson#ifdef KDTRACE_HOOKS 339192853Ssson spin_cnt++; 340192853Ssson#endif 341154941Sjhb /* 342154941Sjhb * Handle the easy case. If no other thread has a write 343154941Sjhb * lock, then try to bump up the count of read locks. Note 344154941Sjhb * that we have to preserve the current state of the 345154941Sjhb * RW_LOCK_WRITE_WAITERS flag. If we fail to acquire a 346154941Sjhb * read lock, then rw_lock must have changed, so restart 347154941Sjhb * the loop. Note that this handles the case of a 348154941Sjhb * completely unlocked rwlock since such a lock is encoded 349154941Sjhb * as a read lock with no waiters. 350154941Sjhb */ 351176017Sjeff v = rw->rw_lock; 352176017Sjeff if (RW_CAN_READ(v)) { 353154941Sjhb /* 354154941Sjhb * The RW_LOCK_READ_WAITERS flag should only be set 355176017Sjeff * if the lock has been unlocked and write waiters 356176017Sjeff * were present. 357154941Sjhb */ 358176017Sjeff if (atomic_cmpset_acq_ptr(&rw->rw_lock, v, 359176017Sjeff v + RW_ONE_READER)) { 360167787Sjhb if (LOCK_LOG_TEST(&rw->lock_object, 0)) 361154941Sjhb CTR4(KTR_LOCK, 362154941Sjhb "%s: %p succeed %p -> %p", __func__, 363176017Sjeff rw, (void *)v, 364176017Sjeff (void *)(v + RW_ONE_READER)); 365154941Sjhb break; 366154941Sjhb } 367154941Sjhb continue; 368154941Sjhb } 369174629Sjeff lock_profile_obtain_lock_failed(&rw->lock_object, 370174629Sjeff &contested, &waittime); 371154941Sjhb 372173960Sattilio#ifdef ADAPTIVE_RWLOCKS 373154941Sjhb /* 374173960Sattilio * If the owner is running on another CPU, spin until 375173960Sattilio * the owner stops running or the state of the lock 376173960Sattilio * changes. 377173960Sattilio */ 378176017Sjeff if ((v & RW_LOCK_READ) == 0) { 379176017Sjeff owner = (struct thread *)RW_OWNER(v); 380176017Sjeff if (TD_IS_RUNNING(owner)) { 381176017Sjeff if (LOCK_LOG_TEST(&rw->lock_object, 0)) 382176017Sjeff CTR3(KTR_LOCK, 383176017Sjeff "%s: spinning on %p held by %p", 384176017Sjeff __func__, rw, owner); 385176017Sjeff while ((struct thread*)RW_OWNER(rw->rw_lock) == 386192853Ssson owner && TD_IS_RUNNING(owner)) { 387176017Sjeff cpu_spinwait(); 388192853Ssson#ifdef KDTRACE_HOOKS 389192853Ssson spin_cnt++; 390192853Ssson#endif 391192853Ssson } 392176017Sjeff continue; 393176017Sjeff } 394177912Sjeff } else if (spintries < rowner_retries) { 395177912Sjeff spintries++; 396177912Sjeff for (i = 0; i < rowner_loops; i++) { 397177912Sjeff v = rw->rw_lock; 398177912Sjeff if ((v & RW_LOCK_READ) == 0 || RW_CAN_READ(v)) 399177912Sjeff break; 400177912Sjeff cpu_spinwait(); 401177912Sjeff } 402177912Sjeff if (i != rowner_loops) 403177912Sjeff continue; 404173960Sattilio } 405173960Sattilio#endif 406173960Sattilio 407173960Sattilio /* 408154941Sjhb * Okay, now it's the hard case. Some other thread already 409176017Sjeff * has a write lock or there are write waiters present, 410176017Sjeff * acquire the turnstile lock so we can begin the process 411176017Sjeff * of blocking. 412154941Sjhb */ 413170295Sjeff ts = turnstile_trywait(&rw->lock_object); 414154941Sjhb 415154941Sjhb /* 416154941Sjhb * The lock might have been released while we spun, so 417176017Sjeff * recheck its state and restart the loop if needed. 418154941Sjhb */ 419176017Sjeff v = rw->rw_lock; 420176017Sjeff if (RW_CAN_READ(v)) { 421170295Sjeff turnstile_cancel(ts); 422154941Sjhb continue; 423154941Sjhb } 424154941Sjhb 425173960Sattilio#ifdef ADAPTIVE_RWLOCKS 426154941Sjhb /* 427193035Sjhb * The current lock owner might have started executing 428193035Sjhb * on another CPU (or the lock could have changed 429193035Sjhb * owners) while we were waiting on the turnstile 430193035Sjhb * chain lock. If so, drop the turnstile lock and try 431193035Sjhb * again. 432173960Sattilio */ 433176017Sjeff if ((v & RW_LOCK_READ) == 0) { 434176017Sjeff owner = (struct thread *)RW_OWNER(v); 435176017Sjeff if (TD_IS_RUNNING(owner)) { 436176017Sjeff turnstile_cancel(ts); 437176017Sjeff continue; 438176017Sjeff } 439173960Sattilio } 440173960Sattilio#endif 441173960Sattilio 442173960Sattilio /* 443176017Sjeff * The lock is held in write mode or it already has waiters. 444154941Sjhb */ 445176017Sjeff MPASS(!RW_CAN_READ(v)); 446176017Sjeff 447176017Sjeff /* 448176017Sjeff * If the RW_LOCK_READ_WAITERS flag is already set, then 449176017Sjeff * we can go ahead and block. If it is not set then try 450176017Sjeff * to set it. If we fail to set it drop the turnstile 451176017Sjeff * lock and restart the loop. 452176017Sjeff */ 453176017Sjeff if (!(v & RW_LOCK_READ_WAITERS)) { 454176017Sjeff if (!atomic_cmpset_ptr(&rw->rw_lock, v, 455176017Sjeff v | RW_LOCK_READ_WAITERS)) { 456170295Sjeff turnstile_cancel(ts); 457157826Sjhb continue; 458157826Sjhb } 459167787Sjhb if (LOCK_LOG_TEST(&rw->lock_object, 0)) 460157826Sjhb CTR2(KTR_LOCK, "%s: %p set read waiters flag", 461157826Sjhb __func__, rw); 462154941Sjhb } 463154941Sjhb 464154941Sjhb /* 465154941Sjhb * We were unable to acquire the lock and the read waiters 466154941Sjhb * flag is set, so we must block on the turnstile. 467154941Sjhb */ 468167787Sjhb if (LOCK_LOG_TEST(&rw->lock_object, 0)) 469154941Sjhb CTR2(KTR_LOCK, "%s: %p blocking on turnstile", __func__, 470154941Sjhb rw); 471192853Ssson#ifdef KDTRACE_HOOKS 472192853Ssson sleep_time -= lockstat_nsecs(); 473192853Ssson#endif 474170295Sjeff turnstile_wait(ts, rw_owner(rw), TS_SHARED_QUEUE); 475192853Ssson#ifdef KDTRACE_HOOKS 476192853Ssson sleep_time += lockstat_nsecs(); 477192853Ssson sleep_cnt++; 478192853Ssson#endif 479167787Sjhb if (LOCK_LOG_TEST(&rw->lock_object, 0)) 480154941Sjhb CTR2(KTR_LOCK, "%s: %p resuming from turnstile", 481154941Sjhb __func__, rw); 482154941Sjhb } 483154941Sjhb 484154941Sjhb /* 485154941Sjhb * TODO: acquire "owner of record" here. Here be turnstile dragons 486154941Sjhb * however. turnstiles don't like owners changing between calls to 487154941Sjhb * turnstile_wait() currently. 488154941Sjhb */ 489192853Ssson LOCKSTAT_PROFILE_OBTAIN_LOCK_SUCCESS(LS_RW_RLOCK_ACQUIRE, rw, contested, 490174629Sjeff waittime, file, line); 491167787Sjhb LOCK_LOG_LOCK("RLOCK", &rw->lock_object, 0, 0, file, line); 492167787Sjhb WITNESS_LOCK(&rw->lock_object, 0, file, line); 493160771Sjhb curthread->td_locks++; 494176017Sjeff curthread->td_rw_rlocks++; 495192853Ssson#ifdef KDTRACE_HOOKS 496192853Ssson if (sleep_time) 497192853Ssson LOCKSTAT_RECORD1(LS_RW_RLOCK_BLOCK, rw, sleep_time); 498192853Ssson 499192853Ssson /* 500192853Ssson * Record only the loops spinning and not sleeping. 501192853Ssson */ 502192853Ssson if (spin_cnt > sleep_cnt) 503192853Ssson LOCKSTAT_RECORD1(LS_RW_RLOCK_SPIN, rw, (spin_cnt - sleep_cnt)); 504192853Ssson#endif 505154941Sjhb} 506154941Sjhb 507177843Sattilioint 508177843Sattilio_rw_try_rlock(struct rwlock *rw, const char *file, int line) 509177843Sattilio{ 510177843Sattilio uintptr_t x; 511177843Sattilio 512228424Savg if (SCHEDULER_STOPPED()) 513228424Savg return (1); 514228424Savg 515177843Sattilio for (;;) { 516177843Sattilio x = rw->rw_lock; 517177843Sattilio KASSERT(rw->rw_lock != RW_DESTROYED, 518177843Sattilio ("rw_try_rlock() of destroyed rwlock @ %s:%d", file, line)); 519177843Sattilio if (!(x & RW_LOCK_READ)) 520177843Sattilio break; 521177843Sattilio if (atomic_cmpset_acq_ptr(&rw->rw_lock, x, x + RW_ONE_READER)) { 522177843Sattilio LOCK_LOG_TRY("RLOCK", &rw->lock_object, 0, 1, file, 523177843Sattilio line); 524177843Sattilio WITNESS_LOCK(&rw->lock_object, LOP_TRYLOCK, file, line); 525177843Sattilio curthread->td_locks++; 526177843Sattilio curthread->td_rw_rlocks++; 527177843Sattilio return (1); 528177843Sattilio } 529177843Sattilio } 530177843Sattilio 531177843Sattilio LOCK_LOG_TRY("RLOCK", &rw->lock_object, 0, 0, file, line); 532177843Sattilio return (0); 533177843Sattilio} 534177843Sattilio 535154941Sjhbvoid 536154941Sjhb_rw_runlock(struct rwlock *rw, const char *file, int line) 537154941Sjhb{ 538154941Sjhb struct turnstile *ts; 539176017Sjeff uintptr_t x, v, queue; 540154941Sjhb 541228424Savg if (SCHEDULER_STOPPED()) 542228424Savg return; 543228424Savg 544169394Sjhb KASSERT(rw->rw_lock != RW_DESTROYED, 545169394Sjhb ("rw_runlock() of destroyed rwlock @ %s:%d", file, line)); 546154941Sjhb _rw_assert(rw, RA_RLOCKED, file, line); 547160771Sjhb curthread->td_locks--; 548176017Sjeff curthread->td_rw_rlocks--; 549167787Sjhb WITNESS_UNLOCK(&rw->lock_object, 0, file, line); 550167787Sjhb LOCK_LOG_LOCK("RUNLOCK", &rw->lock_object, 0, 0, file, line); 551154941Sjhb 552154941Sjhb /* TODO: drop "owner of record" here. */ 553154941Sjhb 554154941Sjhb for (;;) { 555154941Sjhb /* 556154941Sjhb * See if there is more than one read lock held. If so, 557154941Sjhb * just drop one and return. 558154941Sjhb */ 559154941Sjhb x = rw->rw_lock; 560154941Sjhb if (RW_READERS(x) > 1) { 561197643Sattilio if (atomic_cmpset_rel_ptr(&rw->rw_lock, x, 562154941Sjhb x - RW_ONE_READER)) { 563167787Sjhb if (LOCK_LOG_TEST(&rw->lock_object, 0)) 564154941Sjhb CTR4(KTR_LOCK, 565154941Sjhb "%s: %p succeeded %p -> %p", 566154941Sjhb __func__, rw, (void *)x, 567154941Sjhb (void *)(x - RW_ONE_READER)); 568154941Sjhb break; 569154941Sjhb } 570154941Sjhb continue; 571167307Sjhb } 572154941Sjhb /* 573154941Sjhb * If there aren't any waiters for a write lock, then try 574154941Sjhb * to drop it quickly. 575154941Sjhb */ 576176017Sjeff if (!(x & RW_LOCK_WAITERS)) { 577176017Sjeff MPASS((x & ~RW_LOCK_WRITE_SPINNER) == 578176017Sjeff RW_READERS_LOCK(1)); 579197643Sattilio if (atomic_cmpset_rel_ptr(&rw->rw_lock, x, 580197643Sattilio RW_UNLOCKED)) { 581167787Sjhb if (LOCK_LOG_TEST(&rw->lock_object, 0)) 582154941Sjhb CTR2(KTR_LOCK, "%s: %p last succeeded", 583154941Sjhb __func__, rw); 584154941Sjhb break; 585154941Sjhb } 586154941Sjhb continue; 587154941Sjhb } 588154941Sjhb /* 589176017Sjeff * Ok, we know we have waiters and we think we are the 590176017Sjeff * last reader, so grab the turnstile lock. 591154941Sjhb */ 592170295Sjeff turnstile_chain_lock(&rw->lock_object); 593176017Sjeff v = rw->rw_lock & (RW_LOCK_WAITERS | RW_LOCK_WRITE_SPINNER); 594176017Sjeff MPASS(v & RW_LOCK_WAITERS); 595154941Sjhb 596154941Sjhb /* 597154941Sjhb * Try to drop our lock leaving the lock in a unlocked 598154941Sjhb * state. 599154941Sjhb * 600154941Sjhb * If you wanted to do explicit lock handoff you'd have to 601154941Sjhb * do it here. You'd also want to use turnstile_signal() 602154941Sjhb * and you'd have to handle the race where a higher 603154941Sjhb * priority thread blocks on the write lock before the 604154941Sjhb * thread you wakeup actually runs and have the new thread 605154941Sjhb * "steal" the lock. For now it's a lot simpler to just 606154941Sjhb * wakeup all of the waiters. 607154941Sjhb * 608154941Sjhb * As above, if we fail, then another thread might have 609154941Sjhb * acquired a read lock, so drop the turnstile lock and 610154941Sjhb * restart. 611154941Sjhb */ 612176017Sjeff x = RW_UNLOCKED; 613176017Sjeff if (v & RW_LOCK_WRITE_WAITERS) { 614176017Sjeff queue = TS_EXCLUSIVE_QUEUE; 615176017Sjeff x |= (v & RW_LOCK_READ_WAITERS); 616176017Sjeff } else 617176017Sjeff queue = TS_SHARED_QUEUE; 618197643Sattilio if (!atomic_cmpset_rel_ptr(&rw->rw_lock, RW_READERS_LOCK(1) | v, 619176017Sjeff x)) { 620170295Sjeff turnstile_chain_unlock(&rw->lock_object); 621154941Sjhb continue; 622154941Sjhb } 623167787Sjhb if (LOCK_LOG_TEST(&rw->lock_object, 0)) 624154941Sjhb CTR2(KTR_LOCK, "%s: %p last succeeded with waiters", 625154941Sjhb __func__, rw); 626154941Sjhb 627154941Sjhb /* 628154941Sjhb * Ok. The lock is released and all that's left is to 629154941Sjhb * wake up the waiters. Note that the lock might not be 630154941Sjhb * free anymore, but in that case the writers will just 631154941Sjhb * block again if they run before the new lock holder(s) 632154941Sjhb * release the lock. 633154941Sjhb */ 634167787Sjhb ts = turnstile_lookup(&rw->lock_object); 635157846Sjhb MPASS(ts != NULL); 636176017Sjeff turnstile_broadcast(ts, queue); 637154941Sjhb turnstile_unpend(ts, TS_SHARED_LOCK); 638170295Sjeff turnstile_chain_unlock(&rw->lock_object); 639154941Sjhb break; 640154941Sjhb } 641192853Ssson LOCKSTAT_PROFILE_RELEASE_LOCK(LS_RW_RUNLOCK_RELEASE, rw); 642154941Sjhb} 643154941Sjhb 644154941Sjhb/* 645154941Sjhb * This function is called when we are unable to obtain a write lock on the 646154941Sjhb * first try. This means that at least one other thread holds either a 647154941Sjhb * read or write lock. 648154941Sjhb */ 649154941Sjhbvoid 650154941Sjhb_rw_wlock_hard(struct rwlock *rw, uintptr_t tid, const char *file, int line) 651154941Sjhb{ 652170295Sjeff struct turnstile *ts; 653167801Sjhb#ifdef ADAPTIVE_RWLOCKS 654157846Sjhb volatile struct thread *owner; 655176017Sjeff int spintries = 0; 656176017Sjeff int i; 657157851Swkoszek#endif 658189846Sjeff uintptr_t v, x; 659189846Sjeff#ifdef LOCK_PROFILING 660171516Sattilio uint64_t waittime = 0; 661171516Sattilio int contested = 0; 662189846Sjeff#endif 663192853Ssson#ifdef KDTRACE_HOOKS 664192853Ssson uint64_t spin_cnt = 0; 665192853Ssson uint64_t sleep_cnt = 0; 666192853Ssson int64_t sleep_time = 0; 667192853Ssson#endif 668154941Sjhb 669228424Savg if (SCHEDULER_STOPPED()) 670228424Savg return; 671228424Savg 672171052Sattilio if (rw_wlocked(rw)) { 673193307Sattilio KASSERT(rw->lock_object.lo_flags & LO_RECURSABLE, 674171052Sattilio ("%s: recursing but non-recursive rw %s @ %s:%d\n", 675171052Sattilio __func__, rw->lock_object.lo_name, file, line)); 676171052Sattilio rw->rw_recurse++; 677171052Sattilio if (LOCK_LOG_TEST(&rw->lock_object, 0)) 678171052Sattilio CTR2(KTR_LOCK, "%s: %p recursing", __func__, rw); 679171052Sattilio return; 680171052Sattilio } 681171052Sattilio 682167787Sjhb if (LOCK_LOG_TEST(&rw->lock_object, 0)) 683154941Sjhb CTR5(KTR_LOCK, "%s: %s contested (lock=%p) at %s:%d", __func__, 684167787Sjhb rw->lock_object.lo_name, (void *)rw->rw_lock, file, line); 685154941Sjhb 686154941Sjhb while (!_rw_write_lock(rw, tid)) { 687192853Ssson#ifdef KDTRACE_HOOKS 688192853Ssson spin_cnt++; 689192853Ssson#endif 690174629Sjeff lock_profile_obtain_lock_failed(&rw->lock_object, 691174629Sjeff &contested, &waittime); 692173960Sattilio#ifdef ADAPTIVE_RWLOCKS 693173960Sattilio /* 694173960Sattilio * If the lock is write locked and the owner is 695173960Sattilio * running on another CPU, spin until the owner stops 696173960Sattilio * running or the state of the lock changes. 697173960Sattilio */ 698173960Sattilio v = rw->rw_lock; 699173960Sattilio owner = (struct thread *)RW_OWNER(v); 700173960Sattilio if (!(v & RW_LOCK_READ) && TD_IS_RUNNING(owner)) { 701173960Sattilio if (LOCK_LOG_TEST(&rw->lock_object, 0)) 702173960Sattilio CTR3(KTR_LOCK, "%s: spinning on %p held by %p", 703173960Sattilio __func__, rw, owner); 704173960Sattilio while ((struct thread*)RW_OWNER(rw->rw_lock) == owner && 705192853Ssson TD_IS_RUNNING(owner)) { 706173960Sattilio cpu_spinwait(); 707192853Ssson#ifdef KDTRACE_HOOKS 708192853Ssson spin_cnt++; 709192853Ssson#endif 710192853Ssson } 711173960Sattilio continue; 712173960Sattilio } 713177912Sjeff if ((v & RW_LOCK_READ) && RW_READERS(v) && 714177912Sjeff spintries < rowner_retries) { 715176017Sjeff if (!(v & RW_LOCK_WRITE_SPINNER)) { 716176017Sjeff if (!atomic_cmpset_ptr(&rw->rw_lock, v, 717176017Sjeff v | RW_LOCK_WRITE_SPINNER)) { 718176017Sjeff continue; 719176017Sjeff } 720176017Sjeff } 721176017Sjeff spintries++; 722177912Sjeff for (i = 0; i < rowner_loops; i++) { 723176017Sjeff if ((rw->rw_lock & RW_LOCK_WRITE_SPINNER) == 0) 724176017Sjeff break; 725176017Sjeff cpu_spinwait(); 726176017Sjeff } 727192853Ssson#ifdef KDTRACE_HOOKS 728192853Ssson spin_cnt += rowner_loops - i; 729192853Ssson#endif 730177912Sjeff if (i != rowner_loops) 731176017Sjeff continue; 732176017Sjeff } 733173960Sattilio#endif 734170295Sjeff ts = turnstile_trywait(&rw->lock_object); 735154941Sjhb v = rw->rw_lock; 736154941Sjhb 737173960Sattilio#ifdef ADAPTIVE_RWLOCKS 738154941Sjhb /* 739193035Sjhb * The current lock owner might have started executing 740193035Sjhb * on another CPU (or the lock could have changed 741193035Sjhb * owners) while we were waiting on the turnstile 742193035Sjhb * chain lock. If so, drop the turnstile lock and try 743193035Sjhb * again. 744173960Sattilio */ 745173960Sattilio if (!(v & RW_LOCK_READ)) { 746173960Sattilio owner = (struct thread *)RW_OWNER(v); 747173960Sattilio if (TD_IS_RUNNING(owner)) { 748173960Sattilio turnstile_cancel(ts); 749173960Sattilio continue; 750173960Sattilio } 751173960Sattilio } 752173960Sattilio#endif 753173960Sattilio /* 754179334Sattilio * Check for the waiters flags about this rwlock. 755179334Sattilio * If the lock was released, without maintain any pending 756179334Sattilio * waiters queue, simply try to acquire it. 757179334Sattilio * If a pending waiters queue is present, claim the lock 758179334Sattilio * ownership and maintain the pending queue. 759154941Sjhb */ 760176017Sjeff x = v & (RW_LOCK_WAITERS | RW_LOCK_WRITE_SPINNER); 761176017Sjeff if ((v & ~x) == RW_UNLOCKED) { 762176017Sjeff x &= ~RW_LOCK_WRITE_SPINNER; 763176017Sjeff if (atomic_cmpset_acq_ptr(&rw->rw_lock, v, tid | x)) { 764176017Sjeff if (x) 765176017Sjeff turnstile_claim(ts); 766176017Sjeff else 767176017Sjeff turnstile_cancel(ts); 768154941Sjhb break; 769154941Sjhb } 770170295Sjeff turnstile_cancel(ts); 771154941Sjhb continue; 772154941Sjhb } 773154941Sjhb /* 774154941Sjhb * If the RW_LOCK_WRITE_WAITERS flag isn't set, then try to 775154941Sjhb * set it. If we fail to set it, then loop back and try 776154941Sjhb * again. 777154941Sjhb */ 778157826Sjhb if (!(v & RW_LOCK_WRITE_WAITERS)) { 779157826Sjhb if (!atomic_cmpset_ptr(&rw->rw_lock, v, 780157826Sjhb v | RW_LOCK_WRITE_WAITERS)) { 781170295Sjeff turnstile_cancel(ts); 782157826Sjhb continue; 783157826Sjhb } 784167787Sjhb if (LOCK_LOG_TEST(&rw->lock_object, 0)) 785157826Sjhb CTR2(KTR_LOCK, "%s: %p set write waiters flag", 786157826Sjhb __func__, rw); 787154941Sjhb } 788157846Sjhb /* 789154941Sjhb * We were unable to acquire the lock and the write waiters 790154941Sjhb * flag is set, so we must block on the turnstile. 791154941Sjhb */ 792167787Sjhb if (LOCK_LOG_TEST(&rw->lock_object, 0)) 793154941Sjhb CTR2(KTR_LOCK, "%s: %p blocking on turnstile", __func__, 794154941Sjhb rw); 795192853Ssson#ifdef KDTRACE_HOOKS 796192853Ssson sleep_time -= lockstat_nsecs(); 797192853Ssson#endif 798170295Sjeff turnstile_wait(ts, rw_owner(rw), TS_EXCLUSIVE_QUEUE); 799192853Ssson#ifdef KDTRACE_HOOKS 800192853Ssson sleep_time += lockstat_nsecs(); 801192853Ssson sleep_cnt++; 802192853Ssson#endif 803167787Sjhb if (LOCK_LOG_TEST(&rw->lock_object, 0)) 804154941Sjhb CTR2(KTR_LOCK, "%s: %p resuming from turnstile", 805154941Sjhb __func__, rw); 806176017Sjeff#ifdef ADAPTIVE_RWLOCKS 807176017Sjeff spintries = 0; 808176017Sjeff#endif 809154941Sjhb } 810192853Ssson LOCKSTAT_PROFILE_OBTAIN_LOCK_SUCCESS(LS_RW_WLOCK_ACQUIRE, rw, contested, 811192853Ssson waittime, file, line); 812192853Ssson#ifdef KDTRACE_HOOKS 813192853Ssson if (sleep_time) 814192853Ssson LOCKSTAT_RECORD1(LS_RW_WLOCK_BLOCK, rw, sleep_time); 815192853Ssson 816192853Ssson /* 817192853Ssson * Record only the loops spinning and not sleeping. 818192853Ssson */ 819192853Ssson if (spin_cnt > sleep_cnt) 820192853Ssson LOCKSTAT_RECORD1(LS_RW_WLOCK_SPIN, rw, (spin_cnt - sleep_cnt)); 821192853Ssson#endif 822154941Sjhb} 823154941Sjhb 824154941Sjhb/* 825154941Sjhb * This function is called if the first try at releasing a write lock failed. 826154941Sjhb * This means that one of the 2 waiter bits must be set indicating that at 827154941Sjhb * least one thread is waiting on this lock. 828154941Sjhb */ 829154941Sjhbvoid 830154941Sjhb_rw_wunlock_hard(struct rwlock *rw, uintptr_t tid, const char *file, int line) 831154941Sjhb{ 832154941Sjhb struct turnstile *ts; 833154941Sjhb uintptr_t v; 834154941Sjhb int queue; 835154941Sjhb 836228424Savg if (SCHEDULER_STOPPED()) 837228424Savg return; 838228424Savg 839171052Sattilio if (rw_wlocked(rw) && rw_recursed(rw)) { 840176017Sjeff rw->rw_recurse--; 841171052Sattilio if (LOCK_LOG_TEST(&rw->lock_object, 0)) 842171052Sattilio CTR2(KTR_LOCK, "%s: %p unrecursing", __func__, rw); 843171052Sattilio return; 844171052Sattilio } 845171052Sattilio 846154941Sjhb KASSERT(rw->rw_lock & (RW_LOCK_READ_WAITERS | RW_LOCK_WRITE_WAITERS), 847154941Sjhb ("%s: neither of the waiter flags are set", __func__)); 848154941Sjhb 849167787Sjhb if (LOCK_LOG_TEST(&rw->lock_object, 0)) 850154941Sjhb CTR2(KTR_LOCK, "%s: %p contested", __func__, rw); 851154941Sjhb 852170295Sjeff turnstile_chain_lock(&rw->lock_object); 853167787Sjhb ts = turnstile_lookup(&rw->lock_object); 854154941Sjhb MPASS(ts != NULL); 855154941Sjhb 856154941Sjhb /* 857154941Sjhb * Use the same algo as sx locks for now. Prefer waking up shared 858154941Sjhb * waiters if we have any over writers. This is probably not ideal. 859154941Sjhb * 860154941Sjhb * 'v' is the value we are going to write back to rw_lock. If we 861154941Sjhb * have waiters on both queues, we need to preserve the state of 862154941Sjhb * the waiter flag for the queue we don't wake up. For now this is 863154941Sjhb * hardcoded for the algorithm mentioned above. 864154941Sjhb * 865154941Sjhb * In the case of both readers and writers waiting we wakeup the 866154941Sjhb * readers but leave the RW_LOCK_WRITE_WAITERS flag set. If a 867154941Sjhb * new writer comes in before a reader it will claim the lock up 868154941Sjhb * above. There is probably a potential priority inversion in 869154941Sjhb * there that could be worked around either by waking both queues 870154941Sjhb * of waiters or doing some complicated lock handoff gymnastics. 871154941Sjhb */ 872157846Sjhb v = RW_UNLOCKED; 873176076Sjeff if (rw->rw_lock & RW_LOCK_WRITE_WAITERS) { 874176076Sjeff queue = TS_EXCLUSIVE_QUEUE; 875176076Sjeff v |= (rw->rw_lock & RW_LOCK_READ_WAITERS); 876176076Sjeff } else 877154941Sjhb queue = TS_SHARED_QUEUE; 878157846Sjhb 879157846Sjhb /* Wake up all waiters for the specific queue. */ 880167787Sjhb if (LOCK_LOG_TEST(&rw->lock_object, 0)) 881154941Sjhb CTR3(KTR_LOCK, "%s: %p waking up %s waiters", __func__, rw, 882154941Sjhb queue == TS_SHARED_QUEUE ? "read" : "write"); 883154941Sjhb turnstile_broadcast(ts, queue); 884154941Sjhb atomic_store_rel_ptr(&rw->rw_lock, v); 885154941Sjhb turnstile_unpend(ts, TS_EXCLUSIVE_LOCK); 886170295Sjeff turnstile_chain_unlock(&rw->lock_object); 887154941Sjhb} 888154941Sjhb 889157882Sjhb/* 890157882Sjhb * Attempt to do a non-blocking upgrade from a read lock to a write 891157882Sjhb * lock. This will only succeed if this thread holds a single read 892157882Sjhb * lock. Returns true if the upgrade succeeded and false otherwise. 893157882Sjhb */ 894157882Sjhbint 895157882Sjhb_rw_try_upgrade(struct rwlock *rw, const char *file, int line) 896157882Sjhb{ 897176017Sjeff uintptr_t v, x, tid; 898170295Sjeff struct turnstile *ts; 899157882Sjhb int success; 900157882Sjhb 901228424Savg if (SCHEDULER_STOPPED()) 902228424Savg return (1); 903228424Savg 904169394Sjhb KASSERT(rw->rw_lock != RW_DESTROYED, 905169394Sjhb ("rw_try_upgrade() of destroyed rwlock @ %s:%d", file, line)); 906157882Sjhb _rw_assert(rw, RA_RLOCKED, file, line); 907157882Sjhb 908157882Sjhb /* 909157882Sjhb * Attempt to switch from one reader to a writer. If there 910157882Sjhb * are any write waiters, then we will have to lock the 911157882Sjhb * turnstile first to prevent races with another writer 912157882Sjhb * calling turnstile_wait() before we have claimed this 913157882Sjhb * turnstile. So, do the simple case of no waiters first. 914157882Sjhb */ 915157882Sjhb tid = (uintptr_t)curthread; 916176017Sjeff success = 0; 917176017Sjeff for (;;) { 918176017Sjeff v = rw->rw_lock; 919176017Sjeff if (RW_READERS(v) > 1) 920176017Sjeff break; 921176017Sjeff if (!(v & RW_LOCK_WAITERS)) { 922176017Sjeff success = atomic_cmpset_ptr(&rw->rw_lock, v, tid); 923176017Sjeff if (!success) 924176017Sjeff continue; 925176017Sjeff break; 926176017Sjeff } 927157882Sjhb 928176017Sjeff /* 929176017Sjeff * Ok, we think we have waiters, so lock the turnstile. 930176017Sjeff */ 931176017Sjeff ts = turnstile_trywait(&rw->lock_object); 932176017Sjeff v = rw->rw_lock; 933176017Sjeff if (RW_READERS(v) > 1) { 934176017Sjeff turnstile_cancel(ts); 935176017Sjeff break; 936176017Sjeff } 937176017Sjeff /* 938176017Sjeff * Try to switch from one reader to a writer again. This time 939176017Sjeff * we honor the current state of the waiters flags. 940176017Sjeff * If we obtain the lock with the flags set, then claim 941176017Sjeff * ownership of the turnstile. 942176017Sjeff */ 943176017Sjeff x = rw->rw_lock & RW_LOCK_WAITERS; 944176017Sjeff success = atomic_cmpset_ptr(&rw->rw_lock, v, tid | x); 945176017Sjeff if (success) { 946176017Sjeff if (x) 947176017Sjeff turnstile_claim(ts); 948176017Sjeff else 949176017Sjeff turnstile_cancel(ts); 950176017Sjeff break; 951176017Sjeff } 952170295Sjeff turnstile_cancel(ts); 953176017Sjeff } 954167787Sjhb LOCK_LOG_TRY("WUPGRADE", &rw->lock_object, 0, success, file, line); 955176017Sjeff if (success) { 956176017Sjeff curthread->td_rw_rlocks--; 957167787Sjhb WITNESS_UPGRADE(&rw->lock_object, LOP_EXCLUSIVE | LOP_TRYLOCK, 958157882Sjhb file, line); 959192853Ssson LOCKSTAT_RECORD0(LS_RW_TRYUPGRADE_UPGRADE, rw); 960176017Sjeff } 961157882Sjhb return (success); 962157882Sjhb} 963157882Sjhb 964157882Sjhb/* 965157882Sjhb * Downgrade a write lock into a single read lock. 966157882Sjhb */ 967157882Sjhbvoid 968157882Sjhb_rw_downgrade(struct rwlock *rw, const char *file, int line) 969157882Sjhb{ 970157882Sjhb struct turnstile *ts; 971157882Sjhb uintptr_t tid, v; 972176017Sjeff int rwait, wwait; 973157882Sjhb 974228424Savg if (SCHEDULER_STOPPED()) 975228424Savg return; 976228424Savg 977169394Sjhb KASSERT(rw->rw_lock != RW_DESTROYED, 978169394Sjhb ("rw_downgrade() of destroyed rwlock @ %s:%d", file, line)); 979171052Sattilio _rw_assert(rw, RA_WLOCKED | RA_NOTRECURSED, file, line); 980171052Sattilio#ifndef INVARIANTS 981171052Sattilio if (rw_recursed(rw)) 982171052Sattilio panic("downgrade of a recursed lock"); 983171052Sattilio#endif 984157882Sjhb 985167787Sjhb WITNESS_DOWNGRADE(&rw->lock_object, 0, file, line); 986157882Sjhb 987157882Sjhb /* 988157882Sjhb * Convert from a writer to a single reader. First we handle 989157882Sjhb * the easy case with no waiters. If there are any waiters, we 990176017Sjeff * lock the turnstile and "disown" the lock. 991157882Sjhb */ 992157882Sjhb tid = (uintptr_t)curthread; 993157882Sjhb if (atomic_cmpset_rel_ptr(&rw->rw_lock, tid, RW_READERS_LOCK(1))) 994157882Sjhb goto out; 995157882Sjhb 996157882Sjhb /* 997157882Sjhb * Ok, we think we have waiters, so lock the turnstile so we can 998157882Sjhb * read the waiter flags without any races. 999157882Sjhb */ 1000170295Sjeff turnstile_chain_lock(&rw->lock_object); 1001176017Sjeff v = rw->rw_lock & RW_LOCK_WAITERS; 1002176017Sjeff rwait = v & RW_LOCK_READ_WAITERS; 1003176017Sjeff wwait = v & RW_LOCK_WRITE_WAITERS; 1004176017Sjeff MPASS(rwait | wwait); 1005157882Sjhb 1006157882Sjhb /* 1007176017Sjeff * Downgrade from a write lock while preserving waiters flag 1008176017Sjeff * and give up ownership of the turnstile. 1009157882Sjhb */ 1010167787Sjhb ts = turnstile_lookup(&rw->lock_object); 1011157882Sjhb MPASS(ts != NULL); 1012176017Sjeff if (!wwait) 1013176017Sjeff v &= ~RW_LOCK_READ_WAITERS; 1014176017Sjeff atomic_store_rel_ptr(&rw->rw_lock, RW_READERS_LOCK(1) | v); 1015176017Sjeff /* 1016176017Sjeff * Wake other readers if there are no writers pending. Otherwise they 1017176017Sjeff * won't be able to acquire the lock anyway. 1018176017Sjeff */ 1019176017Sjeff if (rwait && !wwait) { 1020157882Sjhb turnstile_broadcast(ts, TS_SHARED_QUEUE); 1021157882Sjhb turnstile_unpend(ts, TS_EXCLUSIVE_LOCK); 1022176017Sjeff } else 1023157882Sjhb turnstile_disown(ts); 1024170295Sjeff turnstile_chain_unlock(&rw->lock_object); 1025157882Sjhbout: 1026176017Sjeff curthread->td_rw_rlocks++; 1027167787Sjhb LOCK_LOG_LOCK("WDOWNGRADE", &rw->lock_object, 0, 0, file, line); 1028192853Ssson LOCKSTAT_RECORD0(LS_RW_DOWNGRADE_DOWNGRADE, rw); 1029157882Sjhb} 1030157882Sjhb 1031154941Sjhb#ifdef INVARIANT_SUPPORT 1032155162Sscottl#ifndef INVARIANTS 1033154941Sjhb#undef _rw_assert 1034154941Sjhb#endif 1035154941Sjhb 1036154941Sjhb/* 1037154941Sjhb * In the non-WITNESS case, rw_assert() can only detect that at least 1038154941Sjhb * *some* thread owns an rlock, but it cannot guarantee that *this* 1039154941Sjhb * thread owns an rlock. 1040154941Sjhb */ 1041154941Sjhbvoid 1042227588Spjd_rw_assert(const struct rwlock *rw, int what, const char *file, int line) 1043154941Sjhb{ 1044154941Sjhb 1045154941Sjhb if (panicstr != NULL) 1046154941Sjhb return; 1047154941Sjhb switch (what) { 1048154941Sjhb case RA_LOCKED: 1049171052Sattilio case RA_LOCKED | RA_RECURSED: 1050171052Sattilio case RA_LOCKED | RA_NOTRECURSED: 1051154941Sjhb case RA_RLOCKED: 1052154941Sjhb#ifdef WITNESS 1053167787Sjhb witness_assert(&rw->lock_object, what, file, line); 1054154941Sjhb#else 1055154941Sjhb /* 1056154941Sjhb * If some other thread has a write lock or we have one 1057154941Sjhb * and are asserting a read lock, fail. Also, if no one 1058154941Sjhb * has a lock at all, fail. 1059154941Sjhb */ 1060155061Sscottl if (rw->rw_lock == RW_UNLOCKED || 1061155061Sscottl (!(rw->rw_lock & RW_LOCK_READ) && (what == RA_RLOCKED || 1062157826Sjhb rw_wowner(rw) != curthread))) 1063154941Sjhb panic("Lock %s not %slocked @ %s:%d\n", 1064167787Sjhb rw->lock_object.lo_name, (what == RA_RLOCKED) ? 1065154941Sjhb "read " : "", file, line); 1066171052Sattilio 1067171052Sattilio if (!(rw->rw_lock & RW_LOCK_READ)) { 1068171052Sattilio if (rw_recursed(rw)) { 1069171052Sattilio if (what & RA_NOTRECURSED) 1070171052Sattilio panic("Lock %s recursed @ %s:%d\n", 1071171052Sattilio rw->lock_object.lo_name, file, 1072171052Sattilio line); 1073171052Sattilio } else if (what & RA_RECURSED) 1074171052Sattilio panic("Lock %s not recursed @ %s:%d\n", 1075171052Sattilio rw->lock_object.lo_name, file, line); 1076171052Sattilio } 1077154941Sjhb#endif 1078154941Sjhb break; 1079154941Sjhb case RA_WLOCKED: 1080171052Sattilio case RA_WLOCKED | RA_RECURSED: 1081171052Sattilio case RA_WLOCKED | RA_NOTRECURSED: 1082157826Sjhb if (rw_wowner(rw) != curthread) 1083154941Sjhb panic("Lock %s not exclusively locked @ %s:%d\n", 1084167787Sjhb rw->lock_object.lo_name, file, line); 1085171052Sattilio if (rw_recursed(rw)) { 1086171052Sattilio if (what & RA_NOTRECURSED) 1087171052Sattilio panic("Lock %s recursed @ %s:%d\n", 1088171052Sattilio rw->lock_object.lo_name, file, line); 1089171052Sattilio } else if (what & RA_RECURSED) 1090171052Sattilio panic("Lock %s not recursed @ %s:%d\n", 1091171052Sattilio rw->lock_object.lo_name, file, line); 1092154941Sjhb break; 1093154941Sjhb case RA_UNLOCKED: 1094154941Sjhb#ifdef WITNESS 1095167787Sjhb witness_assert(&rw->lock_object, what, file, line); 1096154941Sjhb#else 1097154941Sjhb /* 1098154941Sjhb * If we hold a write lock fail. We can't reliably check 1099154941Sjhb * to see if we hold a read lock or not. 1100154941Sjhb */ 1101157826Sjhb if (rw_wowner(rw) == curthread) 1102154941Sjhb panic("Lock %s exclusively locked @ %s:%d\n", 1103167787Sjhb rw->lock_object.lo_name, file, line); 1104154941Sjhb#endif 1105154941Sjhb break; 1106154941Sjhb default: 1107154941Sjhb panic("Unknown rw lock assertion: %d @ %s:%d", what, file, 1108154941Sjhb line); 1109154941Sjhb } 1110154941Sjhb} 1111154941Sjhb#endif /* INVARIANT_SUPPORT */ 1112154941Sjhb 1113154941Sjhb#ifdef DDB 1114154941Sjhbvoid 1115227588Spjddb_show_rwlock(const struct lock_object *lock) 1116154941Sjhb{ 1117227588Spjd const struct rwlock *rw; 1118154941Sjhb struct thread *td; 1119154941Sjhb 1120227588Spjd rw = (const struct rwlock *)lock; 1121154941Sjhb 1122154941Sjhb db_printf(" state: "); 1123154941Sjhb if (rw->rw_lock == RW_UNLOCKED) 1124154941Sjhb db_printf("UNLOCKED\n"); 1125169394Sjhb else if (rw->rw_lock == RW_DESTROYED) { 1126169394Sjhb db_printf("DESTROYED\n"); 1127169394Sjhb return; 1128169394Sjhb } else if (rw->rw_lock & RW_LOCK_READ) 1129167504Sjhb db_printf("RLOCK: %ju locks\n", 1130167504Sjhb (uintmax_t)(RW_READERS(rw->rw_lock))); 1131154941Sjhb else { 1132157826Sjhb td = rw_wowner(rw); 1133154941Sjhb db_printf("WLOCK: %p (tid %d, pid %d, \"%s\")\n", td, 1134173600Sjulian td->td_tid, td->td_proc->p_pid, td->td_name); 1135171052Sattilio if (rw_recursed(rw)) 1136171052Sattilio db_printf(" recursed: %u\n", rw->rw_recurse); 1137154941Sjhb } 1138154941Sjhb db_printf(" waiters: "); 1139154941Sjhb switch (rw->rw_lock & (RW_LOCK_READ_WAITERS | RW_LOCK_WRITE_WAITERS)) { 1140154941Sjhb case RW_LOCK_READ_WAITERS: 1141154941Sjhb db_printf("readers\n"); 1142154941Sjhb break; 1143154941Sjhb case RW_LOCK_WRITE_WAITERS: 1144154941Sjhb db_printf("writers\n"); 1145154941Sjhb break; 1146154941Sjhb case RW_LOCK_READ_WAITERS | RW_LOCK_WRITE_WAITERS: 1147167492Sjhb db_printf("readers and writers\n"); 1148154941Sjhb break; 1149154941Sjhb default: 1150154941Sjhb db_printf("none\n"); 1151154941Sjhb break; 1152154941Sjhb } 1153154941Sjhb} 1154154941Sjhb 1155154941Sjhb#endif 1156