kern_rwlock.c revision 193035
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 193035 2009-05-29 13:56:34Z jhb $"); 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 54167801SjhbCTASSERT((RW_RECURSE & LO_CLASSFLAGS) == RW_RECURSE); 55167801Sjhb 56167801Sjhb#if defined(SMP) && !defined(NO_ADAPTIVE_RWLOCKS) 57167801Sjhb#define ADAPTIVE_RWLOCKS 58177912Sjeff#endif 59177912Sjeff 60177912Sjeff#ifdef ADAPTIVE_RWLOCKS 61177912Sjeffstatic int rowner_retries = 10; 62177912Sjeffstatic int rowner_loops = 10000; 63177912SjeffSYSCTL_NODE(_debug, OID_AUTO, rwlock, CTLFLAG_RD, NULL, "rwlock debugging"); 64177912SjeffSYSCTL_INT(_debug_rwlock, OID_AUTO, retry, CTLFLAG_RW, &rowner_retries, 0, ""); 65177912SjeffSYSCTL_INT(_debug_rwlock, OID_AUTO, loops, CTLFLAG_RW, &rowner_loops, 0, ""); 66154941Sjhb#endif 67154941Sjhb 68154941Sjhb#ifdef DDB 69154941Sjhb#include <ddb/ddb.h> 70154941Sjhb 71173733Sattiliostatic void db_show_rwlock(struct lock_object *lock); 72167368Sjhb#endif 73192853Sssonstatic void assert_rw(struct lock_object *lock, int what); 74192853Sssonstatic void lock_rw(struct lock_object *lock, int how); 75192853Ssson#ifdef KDTRACE_HOOKS 76167368Sjhbstatic int owner_rw(struct lock_object *lock, struct thread **owner); 77154941Sjhb#endif 78154941Sjhbstatic int unlock_rw(struct lock_object *lock); 79167365Sjhb 80167365Sjhbstruct lock_class lock_class_rw = { 81173733Sattilio .lc_name = "rw", 82154941Sjhb .lc_flags = LC_SLEEPLOCK | LC_RECURSABLE | LC_UPGRADABLE, 83167365Sjhb .lc_assert = assert_rw, 84154941Sjhb#ifdef DDB 85167368Sjhb .lc_ddb_show = db_show_rwlock, 86167368Sjhb#endif 87192853Ssson .lc_lock = lock_rw, 88192853Ssson .lc_unlock = unlock_rw, 89192853Ssson#ifdef KDTRACE_HOOKS 90154941Sjhb .lc_owner = owner_rw, 91154941Sjhb#endif 92157826Sjhb}; 93157826Sjhb 94157826Sjhb/* 95157826Sjhb * Return a pointer to the owning thread if the lock is write-locked or 96157826Sjhb * NULL if the lock is unlocked or read-locked. 97154941Sjhb */ 98154941Sjhb#define rw_wowner(rw) \ 99154941Sjhb ((rw)->rw_lock & RW_LOCK_READ ? NULL : \ 100157826Sjhb (struct thread *)RW_OWNER((rw)->rw_lock)) 101171052Sattilio 102171052Sattilio/* 103171052Sattilio * Returns if a write owner is recursed. Write ownership is not assured 104171052Sattilio * here and should be previously checked. 105171052Sattilio */ 106171052Sattilio#define rw_recursed(rw) ((rw)->rw_recurse != 0) 107171052Sattilio 108171052Sattilio/* 109171052Sattilio * Return true if curthread helds the lock. 110171052Sattilio */ 111171052Sattilio#define rw_wlocked(rw) (rw_wowner((rw)) == curthread) 112157826Sjhb 113157826Sjhb/* 114157826Sjhb * Return a pointer to the owning thread for this lock who should receive 115157826Sjhb * any priority lent by threads that block on this lock. Currently this 116157826Sjhb * is identical to rw_wowner(). 117157826Sjhb */ 118154941Sjhb#define rw_owner(rw) rw_wowner(rw) 119154941Sjhb 120154941Sjhb#ifndef INVARIANTS 121154941Sjhb#define _rw_assert(rw, what, file, line) 122154941Sjhb#endif 123173733Sattilio 124173733Sattiliovoid 125173733Sattilioassert_rw(struct lock_object *lock, int what) 126173733Sattilio{ 127173733Sattilio 128173733Sattilio rw_assert((struct rwlock *)lock, what); 129173733Sattilio} 130167368Sjhb 131167368Sjhbvoid 132167368Sjhblock_rw(struct lock_object *lock, int how) 133167368Sjhb{ 134167368Sjhb struct rwlock *rw; 135167368Sjhb 136167368Sjhb rw = (struct rwlock *)lock; 137167368Sjhb if (how) 138167368Sjhb rw_wlock(rw); 139167368Sjhb else 140167368Sjhb rw_rlock(rw); 141167368Sjhb} 142167368Sjhb 143167368Sjhbint 144167368Sjhbunlock_rw(struct lock_object *lock) 145167368Sjhb{ 146167368Sjhb struct rwlock *rw; 147167368Sjhb 148167368Sjhb rw = (struct rwlock *)lock; 149167368Sjhb rw_assert(rw, RA_LOCKED | LA_NOTRECURSED); 150167368Sjhb if (rw->rw_lock & RW_LOCK_READ) { 151167368Sjhb rw_runlock(rw); 152167368Sjhb return (0); 153167368Sjhb } else { 154167368Sjhb rw_wunlock(rw); 155167368Sjhb return (1); 156167368Sjhb } 157192853Ssson} 158192853Ssson 159192853Ssson#ifdef KDTRACE_HOOKS 160192853Sssonint 161192853Sssonowner_rw(struct lock_object *lock, struct thread **owner) 162192853Ssson{ 163192853Ssson struct rwlock *rw = (struct rwlock *)lock; 164192853Ssson uintptr_t x = rw->rw_lock; 165192853Ssson 166192853Ssson *owner = rw_wowner(rw); 167192853Ssson return ((x & RW_LOCK_READ) != 0 ? (RW_READERS(x) != 0) : 168192853Ssson (*owner != NULL)); 169192853Ssson} 170167368Sjhb#endif 171171052Sattilio 172154941Sjhbvoid 173171052Sattiliorw_init_flags(struct rwlock *rw, const char *name, int opts) 174154941Sjhb{ 175171052Sattilio int flags; 176171052Sattilio 177196334Sattilio MPASS((opts & ~(RW_DUPOK | RW_NOPROFILE | RW_NOWITNESS | RW_QUIET | 178196334Sattilio RW_RECURSE)) == 0); 179196334Sattilio 180171052Sattilio flags = LO_UPGRADABLE | LO_RECURSABLE; 181193307Sattilio if (opts & RW_DUPOK) 182171052Sattilio flags |= LO_DUPOK; 183171052Sattilio if (opts & RW_NOPROFILE) 184171052Sattilio flags |= LO_NOPROFILE; 185171052Sattilio if (!(opts & RW_NOWITNESS)) 186171052Sattilio flags |= LO_WITNESS; 187171052Sattilio if (opts & RW_QUIET) 188193307Sattilio flags |= LO_QUIET; 189193307Sattilio flags |= opts & RW_RECURSE; 190171052Sattilio 191171052Sattilio rw->rw_lock = RW_UNLOCKED; 192171052Sattilio rw->rw_recurse = 0; 193154941Sjhb lock_init(&rw->lock_object, &lock_class_rw, name, NULL, flags); 194171052Sattilio} 195171052Sattilio 196154941Sjhbvoid 197154941Sjhbrw_destroy(struct rwlock *rw) 198154941Sjhb{ 199154941Sjhb 200154941Sjhb KASSERT(rw->rw_lock == RW_UNLOCKED, ("rw lock not unlocked")); 201154941Sjhb KASSERT(rw->rw_recurse == 0, ("rw lock still recursed")); 202154941Sjhb rw->rw_lock = RW_DESTROYED; 203171052Sattilio lock_destroy(&rw->lock_object); 204169394Sjhb} 205167787Sjhb 206154941Sjhbvoid 207154941Sjhbrw_sysinit(void *arg) 208154941Sjhb{ 209154941Sjhb struct rw_args *args = arg; 210154941Sjhb 211154941Sjhb rw_init(args->ra_rw, args->ra_desc); 212154941Sjhb} 213154941Sjhb 214154941Sjhbvoid 215154941Sjhbrw_sysinit_flags(void *arg) 216185778Skmacy{ 217185778Skmacy struct rw_args_flags *args = arg; 218185778Skmacy 219185778Skmacy rw_init_flags(args->ra_rw, args->ra_desc, args->ra_flags); 220185778Skmacy} 221185778Skmacy 222185778Skmacyint 223185778Skmacyrw_wowned(struct rwlock *rw) 224167024Srwatson{ 225167024Srwatson 226167024Srwatson return (rw_wowner(rw) == curthread); 227167024Srwatson} 228167024Srwatson 229167024Srwatsonvoid 230167024Srwatson_rw_wlock(struct rwlock *rw, const char *file, int line) 231154941Sjhb{ 232154941Sjhb 233154941Sjhb MPASS(curthread != NULL); 234154941Sjhb KASSERT(rw->rw_lock != RW_DESTROYED, 235154941Sjhb ("rw_wlock() of destroyed rwlock @ %s:%d", file, line)); 236169394Sjhb WITNESS_CHECKORDER(&rw->lock_object, LOP_NEWORDER | LOP_EXCLUSIVE, file, 237169394Sjhb line, NULL); 238167787Sjhb __rw_wlock(rw, curthread, file, line); 239182914Sjhb LOCK_LOG_LOCK("WLOCK", &rw->lock_object, 0, rw->rw_recurse, file, line); 240154941Sjhb WITNESS_LOCK(&rw->lock_object, LOP_EXCLUSIVE, file, line); 241171052Sattilio curthread->td_locks++; 242167787Sjhb} 243160771Sjhb 244154941Sjhbint 245154941Sjhb_rw_try_wlock(struct rwlock *rw, const char *file, int line) 246177843Sattilio{ 247177843Sattilio int rval; 248177843Sattilio 249177843Sattilio KASSERT(rw->rw_lock != RW_DESTROYED, 250177843Sattilio ("rw_try_wlock() of destroyed rwlock @ %s:%d", file, line)); 251177843Sattilio 252177843Sattilio if (rw_wlocked(rw) && (rw->lock_object.lo_flags & RW_RECURSE) != 0) { 253177843Sattilio rw->rw_recurse++; 254193307Sattilio rval = 1; 255193307Sattilio } else 256177843Sattilio rval = atomic_cmpset_acq_ptr(&rw->rw_lock, RW_UNLOCKED, 257177843Sattilio (uintptr_t)curthread); 258177843Sattilio 259177843Sattilio LOCK_LOG_TRY("WLOCK", &rw->lock_object, 0, rval, file, line); 260177843Sattilio if (rval) { 261177843Sattilio WITNESS_LOCK(&rw->lock_object, LOP_EXCLUSIVE | LOP_TRYLOCK, 262177843Sattilio file, line); 263177843Sattilio curthread->td_locks++; 264177843Sattilio } 265177843Sattilio return (rval); 266177843Sattilio} 267177843Sattilio 268177843Sattiliovoid 269177843Sattilio_rw_wunlock(struct rwlock *rw, const char *file, int line) 270177843Sattilio{ 271154941Sjhb 272154941Sjhb MPASS(curthread != NULL); 273154941Sjhb KASSERT(rw->rw_lock != RW_DESTROYED, 274154941Sjhb ("rw_wunlock() of destroyed rwlock @ %s:%d", file, line)); 275154941Sjhb _rw_assert(rw, RA_WLOCKED, file, line); 276169394Sjhb curthread->td_locks--; 277169394Sjhb WITNESS_UNLOCK(&rw->lock_object, LOP_EXCLUSIVE, file, line); 278154941Sjhb LOCK_LOG_LOCK("WUNLOCK", &rw->lock_object, 0, rw->rw_recurse, file, 279160771Sjhb line); 280167787Sjhb if (!rw_recursed(rw)) 281171052Sattilio LOCKSTAT_PROFILE_RELEASE_LOCK(LS_RW_WUNLOCK_RELEASE, rw); 282171052Sattilio __rw_wunlock(rw, curthread, file, line); 283171052Sattilio} 284192853Ssson/* 285154941Sjhb * Determines whether a new reader can acquire a lock. Succeeds if the 286154941Sjhb * reader already owns a read lock and the lock is locked for read to 287176017Sjeff * prevent deadlock from reader recursion. Also succeeds if the lock 288176017Sjeff * is unlocked and has no writer waiters or spinners. Failing otherwise 289176017Sjeff * prioritizes writers before readers. 290176017Sjeff */ 291176017Sjeff#define RW_CAN_READ(_rw) \ 292176017Sjeff ((curthread->td_rw_rlocks && (_rw) & RW_LOCK_READ) || ((_rw) & \ 293176017Sjeff (RW_LOCK_READ | RW_LOCK_WRITE_WAITERS | RW_LOCK_WRITE_SPINNER)) == \ 294176017Sjeff RW_LOCK_READ) 295176017Sjeff 296176017Sjeffvoid 297176017Sjeff_rw_rlock(struct rwlock *rw, const char *file, int line) 298154941Sjhb{ 299154941Sjhb struct turnstile *ts; 300154941Sjhb#ifdef ADAPTIVE_RWLOCKS 301154941Sjhb volatile struct thread *owner; 302170295Sjeff int spintries = 0; 303167801Sjhb int i; 304157846Sjhb#endif 305177912Sjeff#ifdef LOCK_PROFILING 306177912Sjeff uint64_t waittime = 0; 307157851Swkoszek int contested = 0; 308189846Sjeff#endif 309167307Sjhb uintptr_t v; 310167054Skmacy#ifdef KDTRACE_HOOKS 311189846Sjeff uint64_t spin_cnt = 0; 312176017Sjeff uint64_t sleep_cnt = 0; 313192853Ssson int64_t sleep_time = 0; 314192853Ssson#endif 315192853Ssson 316192853Ssson KASSERT(rw->rw_lock != RW_DESTROYED, 317192853Ssson ("rw_rlock() of destroyed rwlock @ %s:%d", file, line)); 318154941Sjhb KASSERT(rw_wowner(rw) != curthread, 319169394Sjhb ("%s (%s): wlock already held @ %s:%d", __func__, 320169394Sjhb rw->lock_object.lo_name, file, line)); 321157826Sjhb WITNESS_CHECKORDER(&rw->lock_object, LOP_NEWORDER, file, line, NULL); 322154941Sjhb 323167787Sjhb for (;;) { 324182914Sjhb#ifdef KDTRACE_HOOKS 325154941Sjhb spin_cnt++; 326154941Sjhb#endif 327192853Ssson /* 328192853Ssson * Handle the easy case. If no other thread has a write 329192853Ssson * lock, then try to bump up the count of read locks. Note 330154941Sjhb * that we have to preserve the current state of the 331154941Sjhb * RW_LOCK_WRITE_WAITERS flag. If we fail to acquire a 332154941Sjhb * read lock, then rw_lock must have changed, so restart 333154941Sjhb * the loop. Note that this handles the case of a 334154941Sjhb * completely unlocked rwlock since such a lock is encoded 335154941Sjhb * as a read lock with no waiters. 336154941Sjhb */ 337154941Sjhb v = rw->rw_lock; 338154941Sjhb if (RW_CAN_READ(v)) { 339154941Sjhb /* 340176017Sjeff * The RW_LOCK_READ_WAITERS flag should only be set 341176017Sjeff * if the lock has been unlocked and write waiters 342154941Sjhb * were present. 343154941Sjhb */ 344176017Sjeff if (atomic_cmpset_acq_ptr(&rw->rw_lock, v, 345176017Sjeff v + RW_ONE_READER)) { 346154941Sjhb if (LOCK_LOG_TEST(&rw->lock_object, 0)) 347176017Sjeff CTR4(KTR_LOCK, 348176017Sjeff "%s: %p succeed %p -> %p", __func__, 349167787Sjhb rw, (void *)v, 350154941Sjhb (void *)(v + RW_ONE_READER)); 351154941Sjhb break; 352176017Sjeff } 353176017Sjeff cpu_spinwait(); 354154941Sjhb continue; 355154941Sjhb } 356154941Sjhb lock_profile_obtain_lock_failed(&rw->lock_object, 357154941Sjhb &contested, &waittime); 358174629Sjeff 359174629Sjeff#ifdef ADAPTIVE_RWLOCKS 360154941Sjhb /* 361173960Sattilio * If the owner is running on another CPU, spin until 362154941Sjhb * the owner stops running or the state of the lock 363173960Sattilio * changes. 364173960Sattilio */ 365173960Sattilio if ((v & RW_LOCK_READ) == 0) { 366173960Sattilio owner = (struct thread *)RW_OWNER(v); 367176017Sjeff if (TD_IS_RUNNING(owner)) { 368176017Sjeff if (LOCK_LOG_TEST(&rw->lock_object, 0)) 369176017Sjeff CTR3(KTR_LOCK, 370176017Sjeff "%s: spinning on %p held by %p", 371176017Sjeff __func__, rw, owner); 372176017Sjeff while ((struct thread*)RW_OWNER(rw->rw_lock) == 373176017Sjeff owner && TD_IS_RUNNING(owner)) { 374176017Sjeff cpu_spinwait(); 375192853Ssson#ifdef KDTRACE_HOOKS 376176017Sjeff spin_cnt++; 377192853Ssson#endif 378192853Ssson } 379192853Ssson continue; 380192853Ssson } 381176017Sjeff } else if (spintries < rowner_retries) { 382176017Sjeff spintries++; 383177912Sjeff for (i = 0; i < rowner_loops; i++) { 384177912Sjeff v = rw->rw_lock; 385177912Sjeff if ((v & RW_LOCK_READ) == 0 || RW_CAN_READ(v)) 386177912Sjeff break; 387177912Sjeff cpu_spinwait(); 388177912Sjeff } 389177912Sjeff if (i != rowner_loops) 390177912Sjeff continue; 391177912Sjeff } 392177912Sjeff#endif 393173960Sattilio 394173960Sattilio /* 395173960Sattilio * Okay, now it's the hard case. Some other thread already 396173960Sattilio * has a write lock or there are write waiters present, 397154941Sjhb * acquire the turnstile lock so we can begin the process 398176017Sjeff * of blocking. 399176017Sjeff */ 400176017Sjeff ts = turnstile_trywait(&rw->lock_object); 401154941Sjhb 402170295Sjeff /* 403154941Sjhb * The lock might have been released while we spun, so 404154941Sjhb * recheck its state and restart the loop if needed. 405154941Sjhb */ 406176017Sjeff v = rw->rw_lock; 407154941Sjhb if (RW_CAN_READ(v)) { 408176017Sjeff turnstile_cancel(ts); 409176017Sjeff cpu_spinwait(); 410170295Sjeff continue; 411154941Sjhb } 412154941Sjhb 413154941Sjhb#ifdef ADAPTIVE_RWLOCKS 414173960Sattilio /* 415154941Sjhb * The current lock owner might have started executing 416193035Sjhb * on another CPU (or the lock could have changed 417193035Sjhb * owners) while we were waiting on the turnstile 418193035Sjhb * chain lock. If so, drop the turnstile lock and try 419193035Sjhb * again. 420193035Sjhb */ 421173960Sattilio if ((v & RW_LOCK_READ) == 0) { 422176017Sjeff owner = (struct thread *)RW_OWNER(v); 423176017Sjeff if (TD_IS_RUNNING(owner)) { 424176017Sjeff turnstile_cancel(ts); 425176017Sjeff cpu_spinwait(); 426176017Sjeff continue; 427176017Sjeff } 428173960Sattilio } 429173960Sattilio#endif 430173960Sattilio 431173960Sattilio /* 432176017Sjeff * The lock is held in write mode or it already has waiters. 433154941Sjhb */ 434176017Sjeff MPASS(!RW_CAN_READ(v)); 435176017Sjeff 436176017Sjeff /* 437176017Sjeff * If the RW_LOCK_READ_WAITERS flag is already set, then 438176017Sjeff * we can go ahead and block. If it is not set then try 439176017Sjeff * to set it. If we fail to set it drop the turnstile 440176017Sjeff * lock and restart the loop. 441176017Sjeff */ 442176017Sjeff if (!(v & RW_LOCK_READ_WAITERS)) { 443176017Sjeff if (!atomic_cmpset_ptr(&rw->rw_lock, v, 444176017Sjeff v | RW_LOCK_READ_WAITERS)) { 445170295Sjeff turnstile_cancel(ts); 446157826Sjhb cpu_spinwait(); 447157826Sjhb continue; 448167787Sjhb } 449157826Sjhb if (LOCK_LOG_TEST(&rw->lock_object, 0)) 450157826Sjhb CTR2(KTR_LOCK, "%s: %p set read waiters flag", 451154941Sjhb __func__, rw); 452154941Sjhb } 453154941Sjhb 454154941Sjhb /* 455154941Sjhb * We were unable to acquire the lock and the read waiters 456154941Sjhb * flag is set, so we must block on the turnstile. 457167787Sjhb */ 458154941Sjhb if (LOCK_LOG_TEST(&rw->lock_object, 0)) 459154941Sjhb CTR2(KTR_LOCK, "%s: %p blocking on turnstile", __func__, 460192853Ssson rw); 461192853Ssson#ifdef KDTRACE_HOOKS 462192853Ssson sleep_time -= lockstat_nsecs(); 463170295Sjeff#endif 464192853Ssson turnstile_wait(ts, rw_owner(rw), TS_SHARED_QUEUE); 465192853Ssson#ifdef KDTRACE_HOOKS 466192853Ssson sleep_time += lockstat_nsecs(); 467192853Ssson sleep_cnt++; 468167787Sjhb#endif 469154941Sjhb if (LOCK_LOG_TEST(&rw->lock_object, 0)) 470154941Sjhb CTR2(KTR_LOCK, "%s: %p resuming from turnstile", 471154941Sjhb __func__, rw); 472154941Sjhb } 473154941Sjhb 474154941Sjhb /* 475154941Sjhb * TODO: acquire "owner of record" here. Here be turnstile dragons 476154941Sjhb * however. turnstiles don't like owners changing between calls to 477154941Sjhb * turnstile_wait() currently. 478192853Ssson */ 479174629Sjeff LOCKSTAT_PROFILE_OBTAIN_LOCK_SUCCESS(LS_RW_RLOCK_ACQUIRE, rw, contested, 480167787Sjhb waittime, file, line); 481167787Sjhb LOCK_LOG_LOCK("RLOCK", &rw->lock_object, 0, 0, file, line); 482160771Sjhb WITNESS_LOCK(&rw->lock_object, 0, file, line); 483176017Sjeff curthread->td_locks++; 484192853Ssson curthread->td_rw_rlocks++; 485192853Ssson#ifdef KDTRACE_HOOKS 486192853Ssson if (sleep_time) 487192853Ssson LOCKSTAT_RECORD1(LS_RW_RLOCK_BLOCK, rw, sleep_time); 488192853Ssson 489192853Ssson /* 490192853Ssson * Record only the loops spinning and not sleeping. 491192853Ssson */ 492192853Ssson if (spin_cnt > sleep_cnt) 493192853Ssson LOCKSTAT_RECORD1(LS_RW_RLOCK_SPIN, rw, (spin_cnt - sleep_cnt)); 494154941Sjhb#endif 495154941Sjhb} 496177843Sattilio 497177843Sattilioint 498177843Sattilio_rw_try_rlock(struct rwlock *rw, const char *file, int line) 499177843Sattilio{ 500177843Sattilio uintptr_t x; 501177843Sattilio 502177843Sattilio for (;;) { 503177843Sattilio x = rw->rw_lock; 504177843Sattilio KASSERT(rw->rw_lock != RW_DESTROYED, 505177843Sattilio ("rw_try_rlock() of destroyed rwlock @ %s:%d", file, line)); 506177843Sattilio if (!(x & RW_LOCK_READ)) 507177843Sattilio break; 508177843Sattilio if (atomic_cmpset_acq_ptr(&rw->rw_lock, x, x + RW_ONE_READER)) { 509177843Sattilio LOCK_LOG_TRY("RLOCK", &rw->lock_object, 0, 1, file, 510177843Sattilio line); 511177843Sattilio WITNESS_LOCK(&rw->lock_object, LOP_TRYLOCK, file, line); 512177843Sattilio curthread->td_locks++; 513177843Sattilio curthread->td_rw_rlocks++; 514177843Sattilio return (1); 515177843Sattilio } 516177843Sattilio } 517177843Sattilio 518177843Sattilio LOCK_LOG_TRY("RLOCK", &rw->lock_object, 0, 0, file, line); 519177843Sattilio return (0); 520177843Sattilio} 521154941Sjhb 522154941Sjhbvoid 523154941Sjhb_rw_runlock(struct rwlock *rw, const char *file, int line) 524154941Sjhb{ 525176017Sjeff struct turnstile *ts; 526154941Sjhb uintptr_t x, v, queue; 527169394Sjhb 528169394Sjhb KASSERT(rw->rw_lock != RW_DESTROYED, 529154941Sjhb ("rw_runlock() of destroyed rwlock @ %s:%d", file, line)); 530160771Sjhb _rw_assert(rw, RA_RLOCKED, file, line); 531176017Sjeff curthread->td_locks--; 532167787Sjhb curthread->td_rw_rlocks--; 533167787Sjhb WITNESS_UNLOCK(&rw->lock_object, 0, file, line); 534154941Sjhb LOCK_LOG_LOCK("RUNLOCK", &rw->lock_object, 0, 0, file, line); 535154941Sjhb 536154941Sjhb /* TODO: drop "owner of record" here. */ 537154941Sjhb 538154941Sjhb for (;;) { 539154941Sjhb /* 540154941Sjhb * See if there is more than one read lock held. If so, 541154941Sjhb * just drop one and return. 542154941Sjhb */ 543154941Sjhb x = rw->rw_lock; 544154941Sjhb if (RW_READERS(x) > 1) { 545154941Sjhb if (atomic_cmpset_ptr(&rw->rw_lock, x, 546167787Sjhb x - RW_ONE_READER)) { 547154941Sjhb if (LOCK_LOG_TEST(&rw->lock_object, 0)) 548154941Sjhb CTR4(KTR_LOCK, 549154941Sjhb "%s: %p succeeded %p -> %p", 550154941Sjhb __func__, rw, (void *)x, 551154941Sjhb (void *)(x - RW_ONE_READER)); 552154941Sjhb break; 553154941Sjhb } 554167307Sjhb continue; 555154941Sjhb } 556154941Sjhb /* 557154941Sjhb * If there aren't any waiters for a write lock, then try 558154941Sjhb * to drop it quickly. 559176017Sjeff */ 560176017Sjeff if (!(x & RW_LOCK_WAITERS)) { 561176017Sjeff MPASS((x & ~RW_LOCK_WRITE_SPINNER) == 562176017Sjeff RW_READERS_LOCK(1)); 563167787Sjhb if (atomic_cmpset_ptr(&rw->rw_lock, x, RW_UNLOCKED)) { 564154941Sjhb if (LOCK_LOG_TEST(&rw->lock_object, 0)) 565154941Sjhb CTR2(KTR_LOCK, "%s: %p last succeeded", 566154941Sjhb __func__, rw); 567154941Sjhb break; 568154941Sjhb } 569154941Sjhb continue; 570154941Sjhb } 571176017Sjeff /* 572176017Sjeff * Ok, we know we have waiters and we think we are the 573154941Sjhb * last reader, so grab the turnstile lock. 574170295Sjeff */ 575176017Sjeff turnstile_chain_lock(&rw->lock_object); 576176017Sjeff v = rw->rw_lock & (RW_LOCK_WAITERS | RW_LOCK_WRITE_SPINNER); 577154941Sjhb MPASS(v & RW_LOCK_WAITERS); 578154941Sjhb 579154941Sjhb /* 580154941Sjhb * Try to drop our lock leaving the lock in a unlocked 581154941Sjhb * state. 582154941Sjhb * 583154941Sjhb * If you wanted to do explicit lock handoff you'd have to 584154941Sjhb * do it here. You'd also want to use turnstile_signal() 585154941Sjhb * and you'd have to handle the race where a higher 586154941Sjhb * priority thread blocks on the write lock before the 587154941Sjhb * thread you wakeup actually runs and have the new thread 588154941Sjhb * "steal" the lock. For now it's a lot simpler to just 589154941Sjhb * wakeup all of the waiters. 590154941Sjhb * 591154941Sjhb * As above, if we fail, then another thread might have 592154941Sjhb * acquired a read lock, so drop the turnstile lock and 593154941Sjhb * restart. 594176017Sjeff */ 595176017Sjeff x = RW_UNLOCKED; 596176017Sjeff if (v & RW_LOCK_WRITE_WAITERS) { 597176017Sjeff queue = TS_EXCLUSIVE_QUEUE; 598176017Sjeff x |= (v & RW_LOCK_READ_WAITERS); 599176017Sjeff } else 600176017Sjeff queue = TS_SHARED_QUEUE; 601176017Sjeff if (!atomic_cmpset_ptr(&rw->rw_lock, RW_READERS_LOCK(1) | v, 602170295Sjeff x)) { 603154941Sjhb turnstile_chain_unlock(&rw->lock_object); 604154941Sjhb continue; 605167787Sjhb } 606154941Sjhb if (LOCK_LOG_TEST(&rw->lock_object, 0)) 607154941Sjhb CTR2(KTR_LOCK, "%s: %p last succeeded with waiters", 608154941Sjhb __func__, rw); 609154941Sjhb 610154941Sjhb /* 611154941Sjhb * Ok. The lock is released and all that's left is to 612154941Sjhb * wake up the waiters. Note that the lock might not be 613154941Sjhb * free anymore, but in that case the writers will just 614154941Sjhb * block again if they run before the new lock holder(s) 615154941Sjhb * release the lock. 616167787Sjhb */ 617157846Sjhb ts = turnstile_lookup(&rw->lock_object); 618176017Sjeff MPASS(ts != NULL); 619154941Sjhb turnstile_broadcast(ts, queue); 620170295Sjeff turnstile_unpend(ts, TS_SHARED_LOCK); 621154941Sjhb turnstile_chain_unlock(&rw->lock_object); 622154941Sjhb break; 623192853Ssson } 624154941Sjhb LOCKSTAT_PROFILE_RELEASE_LOCK(LS_RW_RUNLOCK_RELEASE, rw); 625154941Sjhb} 626154941Sjhb 627154941Sjhb/* 628154941Sjhb * This function is called when we are unable to obtain a write lock on the 629154941Sjhb * first try. This means that at least one other thread holds either a 630154941Sjhb * read or write lock. 631154941Sjhb */ 632154941Sjhbvoid 633154941Sjhb_rw_wlock_hard(struct rwlock *rw, uintptr_t tid, const char *file, int line) 634170295Sjeff{ 635167801Sjhb struct turnstile *ts; 636157846Sjhb#ifdef ADAPTIVE_RWLOCKS 637176017Sjeff volatile struct thread *owner; 638176017Sjeff int spintries = 0; 639157851Swkoszek int i; 640189846Sjeff#endif 641189846Sjeff uintptr_t v, x; 642171516Sattilio#ifdef LOCK_PROFILING 643171516Sattilio uint64_t waittime = 0; 644189846Sjeff int contested = 0; 645192853Ssson#endif 646192853Ssson#ifdef KDTRACE_HOOKS 647192853Ssson uint64_t spin_cnt = 0; 648192853Ssson uint64_t sleep_cnt = 0; 649192853Ssson int64_t sleep_time = 0; 650154941Sjhb#endif 651171052Sattilio 652193307Sattilio if (rw_wlocked(rw)) { 653171052Sattilio KASSERT(rw->lock_object.lo_flags & RW_RECURSE, 654171052Sattilio ("%s: recursing but non-recursive rw %s @ %s:%d\n", 655171052Sattilio __func__, rw->lock_object.lo_name, file, line)); 656171052Sattilio rw->rw_recurse++; 657171052Sattilio if (LOCK_LOG_TEST(&rw->lock_object, 0)) 658171052Sattilio CTR2(KTR_LOCK, "%s: %p recursing", __func__, rw); 659171052Sattilio return; 660171052Sattilio } 661167787Sjhb 662154941Sjhb if (LOCK_LOG_TEST(&rw->lock_object, 0)) 663167787Sjhb CTR5(KTR_LOCK, "%s: %s contested (lock=%p) at %s:%d", __func__, 664154941Sjhb rw->lock_object.lo_name, (void *)rw->rw_lock, file, line); 665154941Sjhb 666192853Ssson while (!_rw_write_lock(rw, tid)) { 667192853Ssson#ifdef KDTRACE_HOOKS 668192853Ssson spin_cnt++; 669174629Sjeff#endif 670174629Sjeff lock_profile_obtain_lock_failed(&rw->lock_object, 671173960Sattilio &contested, &waittime); 672173960Sattilio#ifdef ADAPTIVE_RWLOCKS 673173960Sattilio /* 674173960Sattilio * If the lock is write locked and the owner is 675173960Sattilio * running on another CPU, spin until the owner stops 676173960Sattilio * running or the state of the lock changes. 677173960Sattilio */ 678173960Sattilio v = rw->rw_lock; 679173960Sattilio owner = (struct thread *)RW_OWNER(v); 680173960Sattilio if (!(v & RW_LOCK_READ) && TD_IS_RUNNING(owner)) { 681173960Sattilio if (LOCK_LOG_TEST(&rw->lock_object, 0)) 682173960Sattilio CTR3(KTR_LOCK, "%s: spinning on %p held by %p", 683173960Sattilio __func__, rw, owner); 684192853Ssson while ((struct thread*)RW_OWNER(rw->rw_lock) == owner && 685173960Sattilio TD_IS_RUNNING(owner)) { 686192853Ssson cpu_spinwait(); 687192853Ssson#ifdef KDTRACE_HOOKS 688192853Ssson spin_cnt++; 689192853Ssson#endif 690173960Sattilio } 691173960Sattilio continue; 692177912Sjeff } 693177912Sjeff if ((v & RW_LOCK_READ) && RW_READERS(v) && 694176017Sjeff spintries < rowner_retries) { 695176017Sjeff if (!(v & RW_LOCK_WRITE_SPINNER)) { 696176017Sjeff if (!atomic_cmpset_ptr(&rw->rw_lock, v, 697176017Sjeff v | RW_LOCK_WRITE_SPINNER)) { 698176017Sjeff cpu_spinwait(); 699176017Sjeff continue; 700176017Sjeff } 701177912Sjeff } 702176017Sjeff spintries++; 703176017Sjeff for (i = 0; i < rowner_loops; i++) { 704176017Sjeff if ((rw->rw_lock & RW_LOCK_WRITE_SPINNER) == 0) 705176017Sjeff break; 706192853Ssson cpu_spinwait(); 707192853Ssson } 708192853Ssson#ifdef KDTRACE_HOOKS 709177912Sjeff spin_cnt += rowner_loops - i; 710176017Sjeff#endif 711176017Sjeff if (i != rowner_loops) 712173960Sattilio continue; 713170295Sjeff } 714154941Sjhb#endif 715154941Sjhb ts = turnstile_trywait(&rw->lock_object); 716173960Sattilio v = rw->rw_lock; 717154941Sjhb 718193035Sjhb#ifdef ADAPTIVE_RWLOCKS 719193035Sjhb /* 720193035Sjhb * The current lock owner might have started executing 721193035Sjhb * on another CPU (or the lock could have changed 722193035Sjhb * owners) while we were waiting on the turnstile 723173960Sattilio * chain lock. If so, drop the turnstile lock and try 724173960Sattilio * again. 725173960Sattilio */ 726173960Sattilio if (!(v & RW_LOCK_READ)) { 727173960Sattilio owner = (struct thread *)RW_OWNER(v); 728173960Sattilio if (TD_IS_RUNNING(owner)) { 729173960Sattilio turnstile_cancel(ts); 730173960Sattilio cpu_spinwait(); 731173960Sattilio continue; 732173960Sattilio } 733179334Sattilio } 734179334Sattilio#endif 735179334Sattilio /* 736179334Sattilio * Check for the waiters flags about this rwlock. 737179334Sattilio * If the lock was released, without maintain any pending 738154941Sjhb * waiters queue, simply try to acquire it. 739176017Sjeff * If a pending waiters queue is present, claim the lock 740176017Sjeff * ownership and maintain the pending queue. 741176017Sjeff */ 742176017Sjeff x = v & (RW_LOCK_WAITERS | RW_LOCK_WRITE_SPINNER); 743176017Sjeff if ((v & ~x) == RW_UNLOCKED) { 744176017Sjeff x &= ~RW_LOCK_WRITE_SPINNER; 745176017Sjeff if (atomic_cmpset_acq_ptr(&rw->rw_lock, v, tid | x)) { 746176017Sjeff if (x) 747154941Sjhb turnstile_claim(ts); 748154941Sjhb else 749170295Sjeff turnstile_cancel(ts); 750154941Sjhb break; 751154941Sjhb } 752154941Sjhb turnstile_cancel(ts); 753154941Sjhb cpu_spinwait(); 754154941Sjhb continue; 755154941Sjhb } 756154941Sjhb /* 757157826Sjhb * If the RW_LOCK_WRITE_WAITERS flag isn't set, then try to 758157826Sjhb * set it. If we fail to set it, then loop back and try 759157826Sjhb * again. 760170295Sjeff */ 761157826Sjhb if (!(v & RW_LOCK_WRITE_WAITERS)) { 762157826Sjhb if (!atomic_cmpset_ptr(&rw->rw_lock, v, 763167787Sjhb v | RW_LOCK_WRITE_WAITERS)) { 764157826Sjhb turnstile_cancel(ts); 765157826Sjhb cpu_spinwait(); 766154941Sjhb continue; 767157846Sjhb } 768154941Sjhb if (LOCK_LOG_TEST(&rw->lock_object, 0)) 769154941Sjhb CTR2(KTR_LOCK, "%s: %p set write waiters flag", 770154941Sjhb __func__, rw); 771167787Sjhb } 772154941Sjhb /* 773154941Sjhb * We were unable to acquire the lock and the write waiters 774192853Ssson * flag is set, so we must block on the turnstile. 775192853Ssson */ 776192853Ssson if (LOCK_LOG_TEST(&rw->lock_object, 0)) 777170295Sjeff CTR2(KTR_LOCK, "%s: %p blocking on turnstile", __func__, 778192853Ssson rw); 779192853Ssson#ifdef KDTRACE_HOOKS 780192853Ssson sleep_time -= lockstat_nsecs(); 781192853Ssson#endif 782167787Sjhb turnstile_wait(ts, rw_owner(rw), TS_EXCLUSIVE_QUEUE); 783154941Sjhb#ifdef KDTRACE_HOOKS 784154941Sjhb sleep_time += lockstat_nsecs(); 785176017Sjeff sleep_cnt++; 786176017Sjeff#endif 787176017Sjeff if (LOCK_LOG_TEST(&rw->lock_object, 0)) 788154941Sjhb CTR2(KTR_LOCK, "%s: %p resuming from turnstile", 789192853Ssson __func__, rw); 790192853Ssson#ifdef ADAPTIVE_RWLOCKS 791192853Ssson spintries = 0; 792192853Ssson#endif 793192853Ssson } 794192853Ssson LOCKSTAT_PROFILE_OBTAIN_LOCK_SUCCESS(LS_RW_WLOCK_ACQUIRE, rw, contested, 795192853Ssson waittime, file, line); 796192853Ssson#ifdef KDTRACE_HOOKS 797192853Ssson if (sleep_time) 798192853Ssson LOCKSTAT_RECORD1(LS_RW_WLOCK_BLOCK, rw, sleep_time); 799192853Ssson 800192853Ssson /* 801154941Sjhb * Record only the loops spinning and not sleeping. 802154941Sjhb */ 803154941Sjhb if (spin_cnt > sleep_cnt) 804154941Sjhb LOCKSTAT_RECORD1(LS_RW_WLOCK_SPIN, rw, (spin_cnt - sleep_cnt)); 805154941Sjhb#endif 806154941Sjhb} 807154941Sjhb 808154941Sjhb/* 809154941Sjhb * This function is called if the first try at releasing a write lock failed. 810154941Sjhb * This means that one of the 2 waiter bits must be set indicating that at 811154941Sjhb * least one thread is waiting on this lock. 812154941Sjhb */ 813154941Sjhbvoid 814154941Sjhb_rw_wunlock_hard(struct rwlock *rw, uintptr_t tid, const char *file, int line) 815171052Sattilio{ 816176017Sjeff struct turnstile *ts; 817171052Sattilio uintptr_t v; 818171052Sattilio int queue; 819171052Sattilio 820171052Sattilio if (rw_wlocked(rw) && rw_recursed(rw)) { 821171052Sattilio rw->rw_recurse--; 822154941Sjhb if (LOCK_LOG_TEST(&rw->lock_object, 0)) 823154941Sjhb CTR2(KTR_LOCK, "%s: %p unrecursing", __func__, rw); 824154941Sjhb return; 825167787Sjhb } 826154941Sjhb 827154941Sjhb KASSERT(rw->rw_lock & (RW_LOCK_READ_WAITERS | RW_LOCK_WRITE_WAITERS), 828170295Sjeff ("%s: neither of the waiter flags are set", __func__)); 829167787Sjhb 830154941Sjhb if (LOCK_LOG_TEST(&rw->lock_object, 0)) 831154941Sjhb CTR2(KTR_LOCK, "%s: %p contested", __func__, rw); 832154941Sjhb 833154941Sjhb turnstile_chain_lock(&rw->lock_object); 834154941Sjhb ts = turnstile_lookup(&rw->lock_object); 835154941Sjhb MPASS(ts != NULL); 836154941Sjhb 837154941Sjhb /* 838154941Sjhb * Use the same algo as sx locks for now. Prefer waking up shared 839154941Sjhb * waiters if we have any over writers. This is probably not ideal. 840154941Sjhb * 841154941Sjhb * 'v' is the value we are going to write back to rw_lock. If we 842154941Sjhb * have waiters on both queues, we need to preserve the state of 843154941Sjhb * the waiter flag for the queue we don't wake up. For now this is 844154941Sjhb * hardcoded for the algorithm mentioned above. 845154941Sjhb * 846154941Sjhb * In the case of both readers and writers waiting we wakeup the 847154941Sjhb * readers but leave the RW_LOCK_WRITE_WAITERS flag set. If a 848157846Sjhb * new writer comes in before a reader it will claim the lock up 849176076Sjeff * above. There is probably a potential priority inversion in 850176076Sjeff * there that could be worked around either by waking both queues 851176076Sjeff * of waiters or doing some complicated lock handoff gymnastics. 852176076Sjeff */ 853154941Sjhb v = RW_UNLOCKED; 854157846Sjhb if (rw->rw_lock & RW_LOCK_WRITE_WAITERS) { 855157846Sjhb queue = TS_EXCLUSIVE_QUEUE; 856167787Sjhb v |= (rw->rw_lock & RW_LOCK_READ_WAITERS); 857154941Sjhb } else 858154941Sjhb queue = TS_SHARED_QUEUE; 859154941Sjhb 860154941Sjhb /* Wake up all waiters for the specific queue. */ 861154941Sjhb if (LOCK_LOG_TEST(&rw->lock_object, 0)) 862170295Sjeff CTR3(KTR_LOCK, "%s: %p waking up %s waiters", __func__, rw, 863154941Sjhb queue == TS_SHARED_QUEUE ? "read" : "write"); 864154941Sjhb turnstile_broadcast(ts, queue); 865157882Sjhb atomic_store_rel_ptr(&rw->rw_lock, v); 866157882Sjhb turnstile_unpend(ts, TS_EXCLUSIVE_LOCK); 867157882Sjhb turnstile_chain_unlock(&rw->lock_object); 868157882Sjhb} 869157882Sjhb 870157882Sjhb/* 871157882Sjhb * Attempt to do a non-blocking upgrade from a read lock to a write 872157882Sjhb * lock. This will only succeed if this thread holds a single read 873176017Sjeff * lock. Returns true if the upgrade succeeded and false otherwise. 874170295Sjeff */ 875157882Sjhbint 876157882Sjhb_rw_try_upgrade(struct rwlock *rw, const char *file, int line) 877169394Sjhb{ 878169394Sjhb uintptr_t v, x, tid; 879157882Sjhb struct turnstile *ts; 880157882Sjhb int success; 881157882Sjhb 882157882Sjhb KASSERT(rw->rw_lock != RW_DESTROYED, 883157882Sjhb ("rw_try_upgrade() of destroyed rwlock @ %s:%d", file, line)); 884157882Sjhb _rw_assert(rw, RA_RLOCKED, file, line); 885157882Sjhb 886157882Sjhb /* 887157882Sjhb * Attempt to switch from one reader to a writer. If there 888157882Sjhb * are any write waiters, then we will have to lock the 889176017Sjeff * turnstile first to prevent races with another writer 890176017Sjeff * calling turnstile_wait() before we have claimed this 891176017Sjeff * turnstile. So, do the simple case of no waiters first. 892176017Sjeff */ 893176017Sjeff tid = (uintptr_t)curthread; 894176017Sjeff success = 0; 895176017Sjeff for (;;) { 896176017Sjeff v = rw->rw_lock; 897176017Sjeff if (RW_READERS(v) > 1) 898176017Sjeff break; 899176017Sjeff if (!(v & RW_LOCK_WAITERS)) { 900157882Sjhb success = atomic_cmpset_ptr(&rw->rw_lock, v, tid); 901176017Sjeff if (!success) 902176017Sjeff continue; 903176017Sjeff break; 904176017Sjeff } 905176017Sjeff 906176017Sjeff /* 907176017Sjeff * Ok, we think we have waiters, so lock the turnstile. 908176017Sjeff */ 909176017Sjeff ts = turnstile_trywait(&rw->lock_object); 910176017Sjeff v = rw->rw_lock; 911176017Sjeff if (RW_READERS(v) > 1) { 912176017Sjeff turnstile_cancel(ts); 913176017Sjeff break; 914176017Sjeff } 915176017Sjeff /* 916176017Sjeff * Try to switch from one reader to a writer again. This time 917176017Sjeff * we honor the current state of the waiters flags. 918176017Sjeff * If we obtain the lock with the flags set, then claim 919176017Sjeff * ownership of the turnstile. 920176017Sjeff */ 921176017Sjeff x = rw->rw_lock & RW_LOCK_WAITERS; 922176017Sjeff success = atomic_cmpset_ptr(&rw->rw_lock, v, tid | x); 923176017Sjeff if (success) { 924176017Sjeff if (x) 925170295Sjeff turnstile_claim(ts); 926176017Sjeff else 927167787Sjhb turnstile_cancel(ts); 928176017Sjeff break; 929176017Sjeff } 930167787Sjhb turnstile_cancel(ts); 931157882Sjhb } 932192853Ssson LOCK_LOG_TRY("WUPGRADE", &rw->lock_object, 0, success, file, line); 933176017Sjeff if (success) { 934157882Sjhb curthread->td_rw_rlocks--; 935157882Sjhb WITNESS_UPGRADE(&rw->lock_object, LOP_EXCLUSIVE | LOP_TRYLOCK, 936157882Sjhb file, line); 937157882Sjhb LOCKSTAT_RECORD0(LS_RW_TRYUPGRADE_UPGRADE, rw); 938157882Sjhb } 939157882Sjhb return (success); 940157882Sjhb} 941157882Sjhb 942157882Sjhb/* 943157882Sjhb * Downgrade a write lock into a single read lock. 944157882Sjhb */ 945176017Sjeffvoid 946157882Sjhb_rw_downgrade(struct rwlock *rw, const char *file, int line) 947169394Sjhb{ 948169394Sjhb struct turnstile *ts; 949171052Sattilio uintptr_t tid, v; 950171052Sattilio int rwait, wwait; 951171052Sattilio 952171052Sattilio KASSERT(rw->rw_lock != RW_DESTROYED, 953171052Sattilio ("rw_downgrade() of destroyed rwlock @ %s:%d", file, line)); 954157882Sjhb _rw_assert(rw, RA_WLOCKED | RA_NOTRECURSED, file, line); 955167787Sjhb#ifndef INVARIANTS 956157882Sjhb if (rw_recursed(rw)) 957157882Sjhb panic("downgrade of a recursed lock"); 958157882Sjhb#endif 959157882Sjhb 960176017Sjeff WITNESS_DOWNGRADE(&rw->lock_object, 0, file, line); 961157882Sjhb 962157882Sjhb /* 963157882Sjhb * Convert from a writer to a single reader. First we handle 964157882Sjhb * the easy case with no waiters. If there are any waiters, we 965157882Sjhb * lock the turnstile and "disown" the lock. 966157882Sjhb */ 967157882Sjhb tid = (uintptr_t)curthread; 968157882Sjhb if (atomic_cmpset_rel_ptr(&rw->rw_lock, tid, RW_READERS_LOCK(1))) 969157882Sjhb goto out; 970170295Sjeff 971176017Sjeff /* 972176017Sjeff * Ok, we think we have waiters, so lock the turnstile so we can 973176017Sjeff * read the waiter flags without any races. 974176017Sjeff */ 975157882Sjhb turnstile_chain_lock(&rw->lock_object); 976157882Sjhb v = rw->rw_lock & RW_LOCK_WAITERS; 977176017Sjeff rwait = v & RW_LOCK_READ_WAITERS; 978176017Sjeff wwait = v & RW_LOCK_WRITE_WAITERS; 979157882Sjhb MPASS(rwait | wwait); 980167787Sjhb 981157882Sjhb /* 982176017Sjeff * Downgrade from a write lock while preserving waiters flag 983176017Sjeff * and give up ownership of the turnstile. 984176017Sjeff */ 985176017Sjeff ts = turnstile_lookup(&rw->lock_object); 986176017Sjeff MPASS(ts != NULL); 987176017Sjeff if (!wwait) 988176017Sjeff v &= ~RW_LOCK_READ_WAITERS; 989176017Sjeff atomic_store_rel_ptr(&rw->rw_lock, RW_READERS_LOCK(1) | v); 990157882Sjhb /* 991157882Sjhb * Wake other readers if there are no writers pending. Otherwise they 992176017Sjeff * won't be able to acquire the lock anyway. 993157882Sjhb */ 994170295Sjeff if (rwait && !wwait) { 995157882Sjhb turnstile_broadcast(ts, TS_SHARED_QUEUE); 996176017Sjeff turnstile_unpend(ts, TS_EXCLUSIVE_LOCK); 997167787Sjhb } else 998192853Ssson turnstile_disown(ts); 999157882Sjhb turnstile_chain_unlock(&rw->lock_object); 1000157882Sjhbout: 1001154941Sjhb curthread->td_rw_rlocks++; 1002155162Sscottl LOCK_LOG_LOCK("WDOWNGRADE", &rw->lock_object, 0, 0, file, line); 1003154941Sjhb LOCKSTAT_RECORD0(LS_RW_DOWNGRADE_DOWNGRADE, rw); 1004154941Sjhb} 1005154941Sjhb 1006154941Sjhb#ifdef INVARIANT_SUPPORT 1007154941Sjhb#ifndef INVARIANTS 1008154941Sjhb#undef _rw_assert 1009154941Sjhb#endif 1010154941Sjhb 1011154941Sjhb/* 1012154941Sjhb * In the non-WITNESS case, rw_assert() can only detect that at least 1013154941Sjhb * *some* thread owns an rlock, but it cannot guarantee that *this* 1014154941Sjhb * thread owns an rlock. 1015154941Sjhb */ 1016154941Sjhbvoid 1017154941Sjhb_rw_assert(struct rwlock *rw, int what, const char *file, int line) 1018154941Sjhb{ 1019171052Sattilio 1020171052Sattilio if (panicstr != NULL) 1021154941Sjhb return; 1022154941Sjhb switch (what) { 1023167787Sjhb case RA_LOCKED: 1024154941Sjhb case RA_LOCKED | RA_RECURSED: 1025154941Sjhb case RA_LOCKED | RA_NOTRECURSED: 1026154941Sjhb case RA_RLOCKED: 1027154941Sjhb#ifdef WITNESS 1028154941Sjhb witness_assert(&rw->lock_object, what, file, line); 1029154941Sjhb#else 1030155061Sscottl /* 1031155061Sscottl * If some other thread has a write lock or we have one 1032157826Sjhb * and are asserting a read lock, fail. Also, if no one 1033154941Sjhb * has a lock at all, fail. 1034167787Sjhb */ 1035154941Sjhb if (rw->rw_lock == RW_UNLOCKED || 1036171052Sattilio (!(rw->rw_lock & RW_LOCK_READ) && (what == RA_RLOCKED || 1037171052Sattilio rw_wowner(rw) != curthread))) 1038171052Sattilio panic("Lock %s not %slocked @ %s:%d\n", 1039171052Sattilio rw->lock_object.lo_name, (what == RA_RLOCKED) ? 1040171052Sattilio "read " : "", file, line); 1041171052Sattilio 1042171052Sattilio if (!(rw->rw_lock & RW_LOCK_READ)) { 1043171052Sattilio if (rw_recursed(rw)) { 1044171052Sattilio if (what & RA_NOTRECURSED) 1045171052Sattilio panic("Lock %s recursed @ %s:%d\n", 1046171052Sattilio rw->lock_object.lo_name, file, 1047154941Sjhb line); 1048154941Sjhb } else if (what & RA_RECURSED) 1049154941Sjhb panic("Lock %s not recursed @ %s:%d\n", 1050171052Sattilio rw->lock_object.lo_name, file, line); 1051171052Sattilio } 1052157826Sjhb#endif 1053154941Sjhb break; 1054167787Sjhb case RA_WLOCKED: 1055171052Sattilio case RA_WLOCKED | RA_RECURSED: 1056171052Sattilio case RA_WLOCKED | RA_NOTRECURSED: 1057171052Sattilio if (rw_wowner(rw) != curthread) 1058171052Sattilio panic("Lock %s not exclusively locked @ %s:%d\n", 1059171052Sattilio rw->lock_object.lo_name, file, line); 1060171052Sattilio if (rw_recursed(rw)) { 1061171052Sattilio if (what & RA_NOTRECURSED) 1062154941Sjhb panic("Lock %s recursed @ %s:%d\n", 1063154941Sjhb rw->lock_object.lo_name, file, line); 1064154941Sjhb } else if (what & RA_RECURSED) 1065167787Sjhb panic("Lock %s not recursed @ %s:%d\n", 1066154941Sjhb rw->lock_object.lo_name, file, line); 1067154941Sjhb break; 1068154941Sjhb case RA_UNLOCKED: 1069154941Sjhb#ifdef WITNESS 1070154941Sjhb witness_assert(&rw->lock_object, what, file, line); 1071157826Sjhb#else 1072154941Sjhb /* 1073167787Sjhb * If we hold a write lock fail. We can't reliably check 1074154941Sjhb * to see if we hold a read lock or not. 1075154941Sjhb */ 1076154941Sjhb if (rw_wowner(rw) == curthread) 1077154941Sjhb panic("Lock %s exclusively locked @ %s:%d\n", 1078154941Sjhb rw->lock_object.lo_name, file, line); 1079154941Sjhb#endif 1080154941Sjhb break; 1081154941Sjhb default: 1082154941Sjhb panic("Unknown rw lock assertion: %d @ %s:%d", what, file, 1083154941Sjhb line); 1084154941Sjhb } 1085154941Sjhb} 1086154941Sjhb#endif /* INVARIANT_SUPPORT */ 1087154941Sjhb 1088154941Sjhb#ifdef DDB 1089154941Sjhbvoid 1090154941Sjhbdb_show_rwlock(struct lock_object *lock) 1091154941Sjhb{ 1092154941Sjhb struct rwlock *rw; 1093154941Sjhb struct thread *td; 1094154941Sjhb 1095169394Sjhb rw = (struct rwlock *)lock; 1096169394Sjhb 1097169394Sjhb db_printf(" state: "); 1098169394Sjhb if (rw->rw_lock == RW_UNLOCKED) 1099167504Sjhb db_printf("UNLOCKED\n"); 1100167504Sjhb else if (rw->rw_lock == RW_DESTROYED) { 1101154941Sjhb db_printf("DESTROYED\n"); 1102157826Sjhb return; 1103154941Sjhb } else if (rw->rw_lock & RW_LOCK_READ) 1104173600Sjulian db_printf("RLOCK: %ju locks\n", 1105171052Sattilio (uintmax_t)(RW_READERS(rw->rw_lock))); 1106171052Sattilio else { 1107154941Sjhb td = rw_wowner(rw); 1108154941Sjhb db_printf("WLOCK: %p (tid %d, pid %d, \"%s\")\n", td, 1109154941Sjhb td->td_tid, td->td_proc->p_pid, td->td_name); 1110154941Sjhb if (rw_recursed(rw)) 1111154941Sjhb db_printf(" recursed: %u\n", rw->rw_recurse); 1112154941Sjhb } 1113154941Sjhb db_printf(" waiters: "); 1114154941Sjhb switch (rw->rw_lock & (RW_LOCK_READ_WAITERS | RW_LOCK_WRITE_WAITERS)) { 1115154941Sjhb case RW_LOCK_READ_WAITERS: 1116154941Sjhb db_printf("readers\n"); 1117167492Sjhb break; 1118154941Sjhb case RW_LOCK_WRITE_WAITERS: 1119154941Sjhb db_printf("writers\n"); 1120154941Sjhb break; 1121154941Sjhb case RW_LOCK_READ_WAITERS | RW_LOCK_WRITE_WAITERS: 1122154941Sjhb db_printf("readers and writers\n"); 1123154941Sjhb break; 1124154941Sjhb default: 1125154941Sjhb db_printf("none\n"); 1126 break; 1127 } 1128} 1129 1130#endif 1131