kern_rwlock.c revision 197643
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 197643 2009-09-30 13:26:31Z attilio $"); 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; 61177912SjeffSYSCTL_NODE(_debug, OID_AUTO, rwlock, CTLFLAG_RD, NULL, "rwlock debugging"); 62177912SjeffSYSCTL_INT(_debug_rwlock, OID_AUTO, retry, CTLFLAG_RW, &rowner_retries, 0, ""); 63177912SjeffSYSCTL_INT(_debug_rwlock, OID_AUTO, loops, CTLFLAG_RW, &rowner_loops, 0, ""); 64177912Sjeff#endif 65177912Sjeff 66154941Sjhb#ifdef DDB 67154941Sjhb#include <ddb/ddb.h> 68154941Sjhb 69154941Sjhbstatic void db_show_rwlock(struct lock_object *lock); 70154941Sjhb#endif 71173733Sattiliostatic void assert_rw(struct lock_object *lock, int what); 72167368Sjhbstatic void lock_rw(struct lock_object *lock, int how); 73192853Ssson#ifdef KDTRACE_HOOKS 74192853Sssonstatic int owner_rw(struct lock_object *lock, struct thread **owner); 75192853Ssson#endif 76167368Sjhbstatic int unlock_rw(struct lock_object *lock); 77154941Sjhb 78154941Sjhbstruct lock_class lock_class_rw = { 79167365Sjhb .lc_name = "rw", 80167365Sjhb .lc_flags = LC_SLEEPLOCK | LC_RECURSABLE | LC_UPGRADABLE, 81173733Sattilio .lc_assert = assert_rw, 82154941Sjhb#ifdef DDB 83167365Sjhb .lc_ddb_show = db_show_rwlock, 84154941Sjhb#endif 85167368Sjhb .lc_lock = lock_rw, 86167368Sjhb .lc_unlock = unlock_rw, 87192853Ssson#ifdef KDTRACE_HOOKS 88192853Ssson .lc_owner = owner_rw, 89192853Ssson#endif 90154941Sjhb}; 91154941Sjhb 92157826Sjhb/* 93157826Sjhb * Return a pointer to the owning thread if the lock is write-locked or 94157826Sjhb * NULL if the lock is unlocked or read-locked. 95157826Sjhb */ 96157826Sjhb#define rw_wowner(rw) \ 97154941Sjhb ((rw)->rw_lock & RW_LOCK_READ ? NULL : \ 98154941Sjhb (struct thread *)RW_OWNER((rw)->rw_lock)) 99154941Sjhb 100157826Sjhb/* 101171052Sattilio * Returns if a write owner is recursed. Write ownership is not assured 102171052Sattilio * here and should be previously checked. 103171052Sattilio */ 104171052Sattilio#define rw_recursed(rw) ((rw)->rw_recurse != 0) 105171052Sattilio 106171052Sattilio/* 107171052Sattilio * Return true if curthread helds the lock. 108171052Sattilio */ 109171052Sattilio#define rw_wlocked(rw) (rw_wowner((rw)) == curthread) 110171052Sattilio 111171052Sattilio/* 112157826Sjhb * Return a pointer to the owning thread for this lock who should receive 113157826Sjhb * any priority lent by threads that block on this lock. Currently this 114157826Sjhb * is identical to rw_wowner(). 115157826Sjhb */ 116157826Sjhb#define rw_owner(rw) rw_wowner(rw) 117157826Sjhb 118154941Sjhb#ifndef INVARIANTS 119154941Sjhb#define _rw_assert(rw, what, file, line) 120154941Sjhb#endif 121154941Sjhb 122154941Sjhbvoid 123173733Sattilioassert_rw(struct lock_object *lock, int what) 124173733Sattilio{ 125173733Sattilio 126173733Sattilio rw_assert((struct rwlock *)lock, what); 127173733Sattilio} 128173733Sattilio 129173733Sattiliovoid 130167368Sjhblock_rw(struct lock_object *lock, int how) 131167368Sjhb{ 132167368Sjhb struct rwlock *rw; 133167368Sjhb 134167368Sjhb rw = (struct rwlock *)lock; 135167368Sjhb if (how) 136167368Sjhb rw_wlock(rw); 137167368Sjhb else 138167368Sjhb rw_rlock(rw); 139167368Sjhb} 140167368Sjhb 141167368Sjhbint 142167368Sjhbunlock_rw(struct lock_object *lock) 143167368Sjhb{ 144167368Sjhb struct rwlock *rw; 145167368Sjhb 146167368Sjhb rw = (struct rwlock *)lock; 147167368Sjhb rw_assert(rw, RA_LOCKED | LA_NOTRECURSED); 148167368Sjhb if (rw->rw_lock & RW_LOCK_READ) { 149167368Sjhb rw_runlock(rw); 150167368Sjhb return (0); 151167368Sjhb } else { 152167368Sjhb rw_wunlock(rw); 153167368Sjhb return (1); 154167368Sjhb } 155167368Sjhb} 156167368Sjhb 157192853Ssson#ifdef KDTRACE_HOOKS 158192853Sssonint 159192853Sssonowner_rw(struct lock_object *lock, struct thread **owner) 160192853Ssson{ 161192853Ssson struct rwlock *rw = (struct rwlock *)lock; 162192853Ssson uintptr_t x = rw->rw_lock; 163192853Ssson 164192853Ssson *owner = rw_wowner(rw); 165192853Ssson return ((x & RW_LOCK_READ) != 0 ? (RW_READERS(x) != 0) : 166192853Ssson (*owner != NULL)); 167192853Ssson} 168192853Ssson#endif 169192853Ssson 170167368Sjhbvoid 171171052Sattiliorw_init_flags(struct rwlock *rw, const char *name, int opts) 172154941Sjhb{ 173171052Sattilio int flags; 174154941Sjhb 175171052Sattilio MPASS((opts & ~(RW_DUPOK | RW_NOPROFILE | RW_NOWITNESS | RW_QUIET | 176171052Sattilio RW_RECURSE)) == 0); 177196334Sattilio ASSERT_ATOMIC_LOAD_PTR(rw->rw_lock, 178196334Sattilio ("%s: rw_lock not aligned for %s: %p", __func__, name, 179196334Sattilio &rw->rw_lock)); 180171052Sattilio 181193307Sattilio flags = LO_UPGRADABLE; 182171052Sattilio if (opts & RW_DUPOK) 183171052Sattilio flags |= LO_DUPOK; 184171052Sattilio if (opts & RW_NOPROFILE) 185171052Sattilio flags |= LO_NOPROFILE; 186171052Sattilio if (!(opts & RW_NOWITNESS)) 187171052Sattilio flags |= LO_WITNESS; 188193307Sattilio if (opts & RW_RECURSE) 189193307Sattilio flags |= LO_RECURSABLE; 190171052Sattilio if (opts & RW_QUIET) 191171052Sattilio flags |= LO_QUIET; 192171052Sattilio 193154941Sjhb rw->rw_lock = RW_UNLOCKED; 194171052Sattilio rw->rw_recurse = 0; 195171052Sattilio lock_init(&rw->lock_object, &lock_class_rw, name, NULL, flags); 196154941Sjhb} 197154941Sjhb 198154941Sjhbvoid 199154941Sjhbrw_destroy(struct rwlock *rw) 200154941Sjhb{ 201154941Sjhb 202154941Sjhb KASSERT(rw->rw_lock == RW_UNLOCKED, ("rw lock not unlocked")); 203171052Sattilio KASSERT(rw->rw_recurse == 0, ("rw lock still recursed")); 204169394Sjhb rw->rw_lock = RW_DESTROYED; 205167787Sjhb lock_destroy(&rw->lock_object); 206154941Sjhb} 207154941Sjhb 208154941Sjhbvoid 209154941Sjhbrw_sysinit(void *arg) 210154941Sjhb{ 211154941Sjhb struct rw_args *args = arg; 212154941Sjhb 213154941Sjhb rw_init(args->ra_rw, args->ra_desc); 214154941Sjhb} 215154941Sjhb 216185778Skmacyvoid 217185778Skmacyrw_sysinit_flags(void *arg) 218185778Skmacy{ 219185778Skmacy struct rw_args_flags *args = arg; 220185778Skmacy 221185778Skmacy rw_init_flags(args->ra_rw, args->ra_desc, args->ra_flags); 222185778Skmacy} 223185778Skmacy 224167024Srwatsonint 225167024Srwatsonrw_wowned(struct rwlock *rw) 226167024Srwatson{ 227167024Srwatson 228167024Srwatson return (rw_wowner(rw) == curthread); 229167024Srwatson} 230167024Srwatson 231154941Sjhbvoid 232154941Sjhb_rw_wlock(struct rwlock *rw, const char *file, int line) 233154941Sjhb{ 234154941Sjhb 235154941Sjhb MPASS(curthread != NULL); 236169394Sjhb KASSERT(rw->rw_lock != RW_DESTROYED, 237169394Sjhb ("rw_wlock() of destroyed rwlock @ %s:%d", file, line)); 238167787Sjhb WITNESS_CHECKORDER(&rw->lock_object, LOP_NEWORDER | LOP_EXCLUSIVE, file, 239182914Sjhb line, NULL); 240154941Sjhb __rw_wlock(rw, curthread, file, line); 241171052Sattilio LOCK_LOG_LOCK("WLOCK", &rw->lock_object, 0, rw->rw_recurse, file, line); 242167787Sjhb WITNESS_LOCK(&rw->lock_object, LOP_EXCLUSIVE, file, line); 243160771Sjhb curthread->td_locks++; 244154941Sjhb} 245154941Sjhb 246177843Sattilioint 247177843Sattilio_rw_try_wlock(struct rwlock *rw, const char *file, int line) 248177843Sattilio{ 249177843Sattilio int rval; 250177843Sattilio 251177843Sattilio KASSERT(rw->rw_lock != RW_DESTROYED, 252177843Sattilio ("rw_try_wlock() of destroyed rwlock @ %s:%d", file, line)); 253177843Sattilio 254193307Sattilio if (rw_wlocked(rw) && 255193307Sattilio (rw->lock_object.lo_flags & LO_RECURSABLE) != 0) { 256177843Sattilio rw->rw_recurse++; 257177843Sattilio rval = 1; 258177843Sattilio } else 259177843Sattilio rval = atomic_cmpset_acq_ptr(&rw->rw_lock, RW_UNLOCKED, 260177843Sattilio (uintptr_t)curthread); 261177843Sattilio 262177843Sattilio LOCK_LOG_TRY("WLOCK", &rw->lock_object, 0, rval, file, line); 263177843Sattilio if (rval) { 264177843Sattilio WITNESS_LOCK(&rw->lock_object, LOP_EXCLUSIVE | LOP_TRYLOCK, 265177843Sattilio file, line); 266177843Sattilio curthread->td_locks++; 267177843Sattilio } 268177843Sattilio return (rval); 269177843Sattilio} 270177843Sattilio 271154941Sjhbvoid 272154941Sjhb_rw_wunlock(struct rwlock *rw, const char *file, int line) 273154941Sjhb{ 274154941Sjhb 275154941Sjhb MPASS(curthread != NULL); 276169394Sjhb KASSERT(rw->rw_lock != RW_DESTROYED, 277169394Sjhb ("rw_wunlock() of destroyed rwlock @ %s:%d", file, line)); 278154941Sjhb _rw_assert(rw, RA_WLOCKED, file, line); 279160771Sjhb curthread->td_locks--; 280167787Sjhb WITNESS_UNLOCK(&rw->lock_object, LOP_EXCLUSIVE, file, line); 281171052Sattilio LOCK_LOG_LOCK("WUNLOCK", &rw->lock_object, 0, rw->rw_recurse, file, 282171052Sattilio line); 283171052Sattilio if (!rw_recursed(rw)) 284192853Ssson LOCKSTAT_PROFILE_RELEASE_LOCK(LS_RW_WUNLOCK_RELEASE, rw); 285154941Sjhb __rw_wunlock(rw, curthread, file, line); 286154941Sjhb} 287176017Sjeff/* 288176017Sjeff * Determines whether a new reader can acquire a lock. Succeeds if the 289176017Sjeff * reader already owns a read lock and the lock is locked for read to 290176017Sjeff * prevent deadlock from reader recursion. Also succeeds if the lock 291176017Sjeff * is unlocked and has no writer waiters or spinners. Failing otherwise 292176017Sjeff * prioritizes writers before readers. 293176017Sjeff */ 294176017Sjeff#define RW_CAN_READ(_rw) \ 295176017Sjeff ((curthread->td_rw_rlocks && (_rw) & RW_LOCK_READ) || ((_rw) & \ 296176017Sjeff (RW_LOCK_READ | RW_LOCK_WRITE_WAITERS | RW_LOCK_WRITE_SPINNER)) == \ 297176017Sjeff RW_LOCK_READ) 298154941Sjhb 299154941Sjhbvoid 300154941Sjhb_rw_rlock(struct rwlock *rw, const char *file, int line) 301154941Sjhb{ 302170295Sjeff struct turnstile *ts; 303167801Sjhb#ifdef ADAPTIVE_RWLOCKS 304157846Sjhb volatile struct thread *owner; 305177912Sjeff int spintries = 0; 306177912Sjeff int i; 307157851Swkoszek#endif 308189846Sjeff#ifdef LOCK_PROFILING 309167307Sjhb uint64_t waittime = 0; 310167054Skmacy int contested = 0; 311189846Sjeff#endif 312176017Sjeff uintptr_t v; 313192853Ssson#ifdef KDTRACE_HOOKS 314192853Ssson uint64_t spin_cnt = 0; 315192853Ssson uint64_t sleep_cnt = 0; 316192853Ssson int64_t sleep_time = 0; 317192853Ssson#endif 318154941Sjhb 319169394Sjhb KASSERT(rw->rw_lock != RW_DESTROYED, 320169394Sjhb ("rw_rlock() of destroyed rwlock @ %s:%d", file, line)); 321157826Sjhb KASSERT(rw_wowner(rw) != curthread, 322154941Sjhb ("%s (%s): wlock already held @ %s:%d", __func__, 323167787Sjhb rw->lock_object.lo_name, file, line)); 324182914Sjhb WITNESS_CHECKORDER(&rw->lock_object, LOP_NEWORDER, file, line, NULL); 325154941Sjhb 326154941Sjhb for (;;) { 327192853Ssson#ifdef KDTRACE_HOOKS 328192853Ssson spin_cnt++; 329192853Ssson#endif 330154941Sjhb /* 331154941Sjhb * Handle the easy case. If no other thread has a write 332154941Sjhb * lock, then try to bump up the count of read locks. Note 333154941Sjhb * that we have to preserve the current state of the 334154941Sjhb * RW_LOCK_WRITE_WAITERS flag. If we fail to acquire a 335154941Sjhb * read lock, then rw_lock must have changed, so restart 336154941Sjhb * the loop. Note that this handles the case of a 337154941Sjhb * completely unlocked rwlock since such a lock is encoded 338154941Sjhb * as a read lock with no waiters. 339154941Sjhb */ 340176017Sjeff v = rw->rw_lock; 341176017Sjeff if (RW_CAN_READ(v)) { 342154941Sjhb /* 343154941Sjhb * The RW_LOCK_READ_WAITERS flag should only be set 344176017Sjeff * if the lock has been unlocked and write waiters 345176017Sjeff * were present. 346154941Sjhb */ 347176017Sjeff if (atomic_cmpset_acq_ptr(&rw->rw_lock, v, 348176017Sjeff v + RW_ONE_READER)) { 349167787Sjhb if (LOCK_LOG_TEST(&rw->lock_object, 0)) 350154941Sjhb CTR4(KTR_LOCK, 351154941Sjhb "%s: %p succeed %p -> %p", __func__, 352176017Sjeff rw, (void *)v, 353176017Sjeff (void *)(v + RW_ONE_READER)); 354154941Sjhb break; 355154941Sjhb } 356154941Sjhb continue; 357154941Sjhb } 358174629Sjeff lock_profile_obtain_lock_failed(&rw->lock_object, 359174629Sjeff &contested, &waittime); 360154941Sjhb 361173960Sattilio#ifdef ADAPTIVE_RWLOCKS 362154941Sjhb /* 363173960Sattilio * If the owner is running on another CPU, spin until 364173960Sattilio * the owner stops running or the state of the lock 365173960Sattilio * changes. 366173960Sattilio */ 367176017Sjeff if ((v & RW_LOCK_READ) == 0) { 368176017Sjeff owner = (struct thread *)RW_OWNER(v); 369176017Sjeff if (TD_IS_RUNNING(owner)) { 370176017Sjeff if (LOCK_LOG_TEST(&rw->lock_object, 0)) 371176017Sjeff CTR3(KTR_LOCK, 372176017Sjeff "%s: spinning on %p held by %p", 373176017Sjeff __func__, rw, owner); 374176017Sjeff while ((struct thread*)RW_OWNER(rw->rw_lock) == 375192853Ssson owner && TD_IS_RUNNING(owner)) { 376176017Sjeff cpu_spinwait(); 377192853Ssson#ifdef KDTRACE_HOOKS 378192853Ssson spin_cnt++; 379192853Ssson#endif 380192853Ssson } 381176017Sjeff continue; 382176017Sjeff } 383177912Sjeff } else if (spintries < rowner_retries) { 384177912Sjeff spintries++; 385177912Sjeff for (i = 0; i < rowner_loops; i++) { 386177912Sjeff v = rw->rw_lock; 387177912Sjeff if ((v & RW_LOCK_READ) == 0 || RW_CAN_READ(v)) 388177912Sjeff break; 389177912Sjeff cpu_spinwait(); 390177912Sjeff } 391177912Sjeff if (i != rowner_loops) 392177912Sjeff continue; 393173960Sattilio } 394173960Sattilio#endif 395173960Sattilio 396173960Sattilio /* 397154941Sjhb * Okay, now it's the hard case. Some other thread already 398176017Sjeff * has a write lock or there are write waiters present, 399176017Sjeff * acquire the turnstile lock so we can begin the process 400176017Sjeff * of blocking. 401154941Sjhb */ 402170295Sjeff ts = turnstile_trywait(&rw->lock_object); 403154941Sjhb 404154941Sjhb /* 405154941Sjhb * The lock might have been released while we spun, so 406176017Sjeff * recheck its state and restart the loop if needed. 407154941Sjhb */ 408176017Sjeff v = rw->rw_lock; 409176017Sjeff if (RW_CAN_READ(v)) { 410170295Sjeff turnstile_cancel(ts); 411154941Sjhb continue; 412154941Sjhb } 413154941Sjhb 414173960Sattilio#ifdef ADAPTIVE_RWLOCKS 415154941Sjhb /* 416193035Sjhb * The current lock owner might have started executing 417193035Sjhb * on another CPU (or the lock could have changed 418193035Sjhb * owners) while we were waiting on the turnstile 419193035Sjhb * chain lock. If so, drop the turnstile lock and try 420193035Sjhb * again. 421173960Sattilio */ 422176017Sjeff if ((v & RW_LOCK_READ) == 0) { 423176017Sjeff owner = (struct thread *)RW_OWNER(v); 424176017Sjeff if (TD_IS_RUNNING(owner)) { 425176017Sjeff turnstile_cancel(ts); 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 continue; 447157826Sjhb } 448167787Sjhb if (LOCK_LOG_TEST(&rw->lock_object, 0)) 449157826Sjhb CTR2(KTR_LOCK, "%s: %p set read waiters flag", 450157826Sjhb __func__, rw); 451154941Sjhb } 452154941Sjhb 453154941Sjhb /* 454154941Sjhb * We were unable to acquire the lock and the read waiters 455154941Sjhb * flag is set, so we must block on the turnstile. 456154941Sjhb */ 457167787Sjhb if (LOCK_LOG_TEST(&rw->lock_object, 0)) 458154941Sjhb CTR2(KTR_LOCK, "%s: %p blocking on turnstile", __func__, 459154941Sjhb rw); 460192853Ssson#ifdef KDTRACE_HOOKS 461192853Ssson sleep_time -= lockstat_nsecs(); 462192853Ssson#endif 463170295Sjeff turnstile_wait(ts, rw_owner(rw), TS_SHARED_QUEUE); 464192853Ssson#ifdef KDTRACE_HOOKS 465192853Ssson sleep_time += lockstat_nsecs(); 466192853Ssson sleep_cnt++; 467192853Ssson#endif 468167787Sjhb if (LOCK_LOG_TEST(&rw->lock_object, 0)) 469154941Sjhb CTR2(KTR_LOCK, "%s: %p resuming from turnstile", 470154941Sjhb __func__, rw); 471154941Sjhb } 472154941Sjhb 473154941Sjhb /* 474154941Sjhb * TODO: acquire "owner of record" here. Here be turnstile dragons 475154941Sjhb * however. turnstiles don't like owners changing between calls to 476154941Sjhb * turnstile_wait() currently. 477154941Sjhb */ 478192853Ssson LOCKSTAT_PROFILE_OBTAIN_LOCK_SUCCESS(LS_RW_RLOCK_ACQUIRE, rw, contested, 479174629Sjeff waittime, file, line); 480167787Sjhb LOCK_LOG_LOCK("RLOCK", &rw->lock_object, 0, 0, file, line); 481167787Sjhb WITNESS_LOCK(&rw->lock_object, 0, file, line); 482160771Sjhb curthread->td_locks++; 483176017Sjeff curthread->td_rw_rlocks++; 484192853Ssson#ifdef KDTRACE_HOOKS 485192853Ssson if (sleep_time) 486192853Ssson LOCKSTAT_RECORD1(LS_RW_RLOCK_BLOCK, rw, sleep_time); 487192853Ssson 488192853Ssson /* 489192853Ssson * Record only the loops spinning and not sleeping. 490192853Ssson */ 491192853Ssson if (spin_cnt > sleep_cnt) 492192853Ssson LOCKSTAT_RECORD1(LS_RW_RLOCK_SPIN, rw, (spin_cnt - sleep_cnt)); 493192853Ssson#endif 494154941Sjhb} 495154941Sjhb 496177843Sattilioint 497177843Sattilio_rw_try_rlock(struct rwlock *rw, const char *file, int line) 498177843Sattilio{ 499177843Sattilio uintptr_t x; 500177843Sattilio 501177843Sattilio for (;;) { 502177843Sattilio x = rw->rw_lock; 503177843Sattilio KASSERT(rw->rw_lock != RW_DESTROYED, 504177843Sattilio ("rw_try_rlock() of destroyed rwlock @ %s:%d", file, line)); 505177843Sattilio if (!(x & RW_LOCK_READ)) 506177843Sattilio break; 507177843Sattilio if (atomic_cmpset_acq_ptr(&rw->rw_lock, x, x + RW_ONE_READER)) { 508177843Sattilio LOCK_LOG_TRY("RLOCK", &rw->lock_object, 0, 1, file, 509177843Sattilio line); 510177843Sattilio WITNESS_LOCK(&rw->lock_object, LOP_TRYLOCK, file, line); 511177843Sattilio curthread->td_locks++; 512177843Sattilio curthread->td_rw_rlocks++; 513177843Sattilio return (1); 514177843Sattilio } 515177843Sattilio } 516177843Sattilio 517177843Sattilio LOCK_LOG_TRY("RLOCK", &rw->lock_object, 0, 0, file, line); 518177843Sattilio return (0); 519177843Sattilio} 520177843Sattilio 521154941Sjhbvoid 522154941Sjhb_rw_runlock(struct rwlock *rw, const char *file, int line) 523154941Sjhb{ 524154941Sjhb struct turnstile *ts; 525176017Sjeff uintptr_t x, v, queue; 526154941Sjhb 527169394Sjhb KASSERT(rw->rw_lock != RW_DESTROYED, 528169394Sjhb ("rw_runlock() of destroyed rwlock @ %s:%d", file, line)); 529154941Sjhb _rw_assert(rw, RA_RLOCKED, file, line); 530160771Sjhb curthread->td_locks--; 531176017Sjeff curthread->td_rw_rlocks--; 532167787Sjhb WITNESS_UNLOCK(&rw->lock_object, 0, file, line); 533167787Sjhb LOCK_LOG_LOCK("RUNLOCK", &rw->lock_object, 0, 0, file, line); 534154941Sjhb 535154941Sjhb /* TODO: drop "owner of record" here. */ 536154941Sjhb 537154941Sjhb for (;;) { 538154941Sjhb /* 539154941Sjhb * See if there is more than one read lock held. If so, 540154941Sjhb * just drop one and return. 541154941Sjhb */ 542154941Sjhb x = rw->rw_lock; 543154941Sjhb if (RW_READERS(x) > 1) { 544197643Sattilio if (atomic_cmpset_rel_ptr(&rw->rw_lock, x, 545154941Sjhb x - RW_ONE_READER)) { 546167787Sjhb if (LOCK_LOG_TEST(&rw->lock_object, 0)) 547154941Sjhb CTR4(KTR_LOCK, 548154941Sjhb "%s: %p succeeded %p -> %p", 549154941Sjhb __func__, rw, (void *)x, 550154941Sjhb (void *)(x - RW_ONE_READER)); 551154941Sjhb break; 552154941Sjhb } 553154941Sjhb continue; 554167307Sjhb } 555154941Sjhb /* 556154941Sjhb * If there aren't any waiters for a write lock, then try 557154941Sjhb * to drop it quickly. 558154941Sjhb */ 559176017Sjeff if (!(x & RW_LOCK_WAITERS)) { 560176017Sjeff MPASS((x & ~RW_LOCK_WRITE_SPINNER) == 561176017Sjeff RW_READERS_LOCK(1)); 562197643Sattilio if (atomic_cmpset_rel_ptr(&rw->rw_lock, x, 563197643Sattilio RW_UNLOCKED)) { 564167787Sjhb 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 } 571154941Sjhb /* 572176017Sjeff * Ok, we know we have waiters and we think we are the 573176017Sjeff * last reader, so grab the turnstile lock. 574154941Sjhb */ 575170295Sjeff turnstile_chain_lock(&rw->lock_object); 576176017Sjeff v = rw->rw_lock & (RW_LOCK_WAITERS | RW_LOCK_WRITE_SPINNER); 577176017Sjeff 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. 594154941Sjhb */ 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; 601197643Sattilio if (!atomic_cmpset_rel_ptr(&rw->rw_lock, RW_READERS_LOCK(1) | v, 602176017Sjeff x)) { 603170295Sjeff turnstile_chain_unlock(&rw->lock_object); 604154941Sjhb continue; 605154941Sjhb } 606167787Sjhb 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. 616154941Sjhb */ 617167787Sjhb ts = turnstile_lookup(&rw->lock_object); 618157846Sjhb MPASS(ts != NULL); 619176017Sjeff turnstile_broadcast(ts, queue); 620154941Sjhb turnstile_unpend(ts, TS_SHARED_LOCK); 621170295Sjeff turnstile_chain_unlock(&rw->lock_object); 622154941Sjhb break; 623154941Sjhb } 624192853Ssson 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) 634154941Sjhb{ 635170295Sjeff struct turnstile *ts; 636167801Sjhb#ifdef ADAPTIVE_RWLOCKS 637157846Sjhb volatile struct thread *owner; 638176017Sjeff int spintries = 0; 639176017Sjeff int i; 640157851Swkoszek#endif 641189846Sjeff uintptr_t v, x; 642189846Sjeff#ifdef LOCK_PROFILING 643171516Sattilio uint64_t waittime = 0; 644171516Sattilio int contested = 0; 645189846Sjeff#endif 646192853Ssson#ifdef KDTRACE_HOOKS 647192853Ssson uint64_t spin_cnt = 0; 648192853Ssson uint64_t sleep_cnt = 0; 649192853Ssson int64_t sleep_time = 0; 650192853Ssson#endif 651154941Sjhb 652171052Sattilio if (rw_wlocked(rw)) { 653193307Sattilio KASSERT(rw->lock_object.lo_flags & LO_RECURSABLE, 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 } 661171052Sattilio 662167787Sjhb if (LOCK_LOG_TEST(&rw->lock_object, 0)) 663154941Sjhb CTR5(KTR_LOCK, "%s: %s contested (lock=%p) at %s:%d", __func__, 664167787Sjhb rw->lock_object.lo_name, (void *)rw->rw_lock, file, line); 665154941Sjhb 666154941Sjhb while (!_rw_write_lock(rw, tid)) { 667192853Ssson#ifdef KDTRACE_HOOKS 668192853Ssson spin_cnt++; 669192853Ssson#endif 670174629Sjeff lock_profile_obtain_lock_failed(&rw->lock_object, 671174629Sjeff &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); 684173960Sattilio while ((struct thread*)RW_OWNER(rw->rw_lock) == owner && 685192853Ssson TD_IS_RUNNING(owner)) { 686173960Sattilio cpu_spinwait(); 687192853Ssson#ifdef KDTRACE_HOOKS 688192853Ssson spin_cnt++; 689192853Ssson#endif 690192853Ssson } 691173960Sattilio continue; 692173960Sattilio } 693177912Sjeff if ((v & RW_LOCK_READ) && RW_READERS(v) && 694177912Sjeff 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 continue; 699176017Sjeff } 700176017Sjeff } 701176017Sjeff spintries++; 702177912Sjeff for (i = 0; i < rowner_loops; i++) { 703176017Sjeff if ((rw->rw_lock & RW_LOCK_WRITE_SPINNER) == 0) 704176017Sjeff break; 705176017Sjeff cpu_spinwait(); 706176017Sjeff } 707192853Ssson#ifdef KDTRACE_HOOKS 708192853Ssson spin_cnt += rowner_loops - i; 709192853Ssson#endif 710177912Sjeff if (i != rowner_loops) 711176017Sjeff continue; 712176017Sjeff } 713173960Sattilio#endif 714170295Sjeff ts = turnstile_trywait(&rw->lock_object); 715154941Sjhb v = rw->rw_lock; 716154941Sjhb 717173960Sattilio#ifdef ADAPTIVE_RWLOCKS 718154941Sjhb /* 719193035Sjhb * The current lock owner might have started executing 720193035Sjhb * on another CPU (or the lock could have changed 721193035Sjhb * owners) while we were waiting on the turnstile 722193035Sjhb * chain lock. If so, drop the turnstile lock and try 723193035Sjhb * again. 724173960Sattilio */ 725173960Sattilio if (!(v & RW_LOCK_READ)) { 726173960Sattilio owner = (struct thread *)RW_OWNER(v); 727173960Sattilio if (TD_IS_RUNNING(owner)) { 728173960Sattilio turnstile_cancel(ts); 729173960Sattilio continue; 730173960Sattilio } 731173960Sattilio } 732173960Sattilio#endif 733173960Sattilio /* 734179334Sattilio * Check for the waiters flags about this rwlock. 735179334Sattilio * If the lock was released, without maintain any pending 736179334Sattilio * waiters queue, simply try to acquire it. 737179334Sattilio * If a pending waiters queue is present, claim the lock 738179334Sattilio * ownership and maintain the pending queue. 739154941Sjhb */ 740176017Sjeff x = v & (RW_LOCK_WAITERS | RW_LOCK_WRITE_SPINNER); 741176017Sjeff if ((v & ~x) == RW_UNLOCKED) { 742176017Sjeff x &= ~RW_LOCK_WRITE_SPINNER; 743176017Sjeff if (atomic_cmpset_acq_ptr(&rw->rw_lock, v, tid | x)) { 744176017Sjeff if (x) 745176017Sjeff turnstile_claim(ts); 746176017Sjeff else 747176017Sjeff turnstile_cancel(ts); 748154941Sjhb break; 749154941Sjhb } 750170295Sjeff turnstile_cancel(ts); 751154941Sjhb continue; 752154941Sjhb } 753154941Sjhb /* 754154941Sjhb * If the RW_LOCK_WRITE_WAITERS flag isn't set, then try to 755154941Sjhb * set it. If we fail to set it, then loop back and try 756154941Sjhb * again. 757154941Sjhb */ 758157826Sjhb if (!(v & RW_LOCK_WRITE_WAITERS)) { 759157826Sjhb if (!atomic_cmpset_ptr(&rw->rw_lock, v, 760157826Sjhb v | RW_LOCK_WRITE_WAITERS)) { 761170295Sjeff turnstile_cancel(ts); 762157826Sjhb continue; 763157826Sjhb } 764167787Sjhb if (LOCK_LOG_TEST(&rw->lock_object, 0)) 765157826Sjhb CTR2(KTR_LOCK, "%s: %p set write waiters flag", 766157826Sjhb __func__, rw); 767154941Sjhb } 768157846Sjhb /* 769154941Sjhb * We were unable to acquire the lock and the write waiters 770154941Sjhb * flag is set, so we must block on the turnstile. 771154941Sjhb */ 772167787Sjhb if (LOCK_LOG_TEST(&rw->lock_object, 0)) 773154941Sjhb CTR2(KTR_LOCK, "%s: %p blocking on turnstile", __func__, 774154941Sjhb rw); 775192853Ssson#ifdef KDTRACE_HOOKS 776192853Ssson sleep_time -= lockstat_nsecs(); 777192853Ssson#endif 778170295Sjeff turnstile_wait(ts, rw_owner(rw), TS_EXCLUSIVE_QUEUE); 779192853Ssson#ifdef KDTRACE_HOOKS 780192853Ssson sleep_time += lockstat_nsecs(); 781192853Ssson sleep_cnt++; 782192853Ssson#endif 783167787Sjhb if (LOCK_LOG_TEST(&rw->lock_object, 0)) 784154941Sjhb CTR2(KTR_LOCK, "%s: %p resuming from turnstile", 785154941Sjhb __func__, rw); 786176017Sjeff#ifdef ADAPTIVE_RWLOCKS 787176017Sjeff spintries = 0; 788176017Sjeff#endif 789154941Sjhb } 790192853Ssson LOCKSTAT_PROFILE_OBTAIN_LOCK_SUCCESS(LS_RW_WLOCK_ACQUIRE, rw, contested, 791192853Ssson waittime, file, line); 792192853Ssson#ifdef KDTRACE_HOOKS 793192853Ssson if (sleep_time) 794192853Ssson LOCKSTAT_RECORD1(LS_RW_WLOCK_BLOCK, rw, sleep_time); 795192853Ssson 796192853Ssson /* 797192853Ssson * Record only the loops spinning and not sleeping. 798192853Ssson */ 799192853Ssson if (spin_cnt > sleep_cnt) 800192853Ssson LOCKSTAT_RECORD1(LS_RW_WLOCK_SPIN, rw, (spin_cnt - sleep_cnt)); 801192853Ssson#endif 802154941Sjhb} 803154941Sjhb 804154941Sjhb/* 805154941Sjhb * This function is called if the first try at releasing a write lock failed. 806154941Sjhb * This means that one of the 2 waiter bits must be set indicating that at 807154941Sjhb * least one thread is waiting on this lock. 808154941Sjhb */ 809154941Sjhbvoid 810154941Sjhb_rw_wunlock_hard(struct rwlock *rw, uintptr_t tid, const char *file, int line) 811154941Sjhb{ 812154941Sjhb struct turnstile *ts; 813154941Sjhb uintptr_t v; 814154941Sjhb int queue; 815154941Sjhb 816171052Sattilio if (rw_wlocked(rw) && rw_recursed(rw)) { 817176017Sjeff rw->rw_recurse--; 818171052Sattilio if (LOCK_LOG_TEST(&rw->lock_object, 0)) 819171052Sattilio CTR2(KTR_LOCK, "%s: %p unrecursing", __func__, rw); 820171052Sattilio return; 821171052Sattilio } 822171052Sattilio 823154941Sjhb KASSERT(rw->rw_lock & (RW_LOCK_READ_WAITERS | RW_LOCK_WRITE_WAITERS), 824154941Sjhb ("%s: neither of the waiter flags are set", __func__)); 825154941Sjhb 826167787Sjhb if (LOCK_LOG_TEST(&rw->lock_object, 0)) 827154941Sjhb CTR2(KTR_LOCK, "%s: %p contested", __func__, rw); 828154941Sjhb 829170295Sjeff turnstile_chain_lock(&rw->lock_object); 830167787Sjhb ts = turnstile_lookup(&rw->lock_object); 831154941Sjhb MPASS(ts != NULL); 832154941Sjhb 833154941Sjhb /* 834154941Sjhb * Use the same algo as sx locks for now. Prefer waking up shared 835154941Sjhb * waiters if we have any over writers. This is probably not ideal. 836154941Sjhb * 837154941Sjhb * 'v' is the value we are going to write back to rw_lock. If we 838154941Sjhb * have waiters on both queues, we need to preserve the state of 839154941Sjhb * the waiter flag for the queue we don't wake up. For now this is 840154941Sjhb * hardcoded for the algorithm mentioned above. 841154941Sjhb * 842154941Sjhb * In the case of both readers and writers waiting we wakeup the 843154941Sjhb * readers but leave the RW_LOCK_WRITE_WAITERS flag set. If a 844154941Sjhb * new writer comes in before a reader it will claim the lock up 845154941Sjhb * above. There is probably a potential priority inversion in 846154941Sjhb * there that could be worked around either by waking both queues 847154941Sjhb * of waiters or doing some complicated lock handoff gymnastics. 848154941Sjhb */ 849157846Sjhb v = RW_UNLOCKED; 850176076Sjeff if (rw->rw_lock & RW_LOCK_WRITE_WAITERS) { 851176076Sjeff queue = TS_EXCLUSIVE_QUEUE; 852176076Sjeff v |= (rw->rw_lock & RW_LOCK_READ_WAITERS); 853176076Sjeff } else 854154941Sjhb queue = TS_SHARED_QUEUE; 855157846Sjhb 856157846Sjhb /* Wake up all waiters for the specific queue. */ 857167787Sjhb if (LOCK_LOG_TEST(&rw->lock_object, 0)) 858154941Sjhb CTR3(KTR_LOCK, "%s: %p waking up %s waiters", __func__, rw, 859154941Sjhb queue == TS_SHARED_QUEUE ? "read" : "write"); 860154941Sjhb turnstile_broadcast(ts, queue); 861154941Sjhb atomic_store_rel_ptr(&rw->rw_lock, v); 862154941Sjhb turnstile_unpend(ts, TS_EXCLUSIVE_LOCK); 863170295Sjeff turnstile_chain_unlock(&rw->lock_object); 864154941Sjhb} 865154941Sjhb 866157882Sjhb/* 867157882Sjhb * Attempt to do a non-blocking upgrade from a read lock to a write 868157882Sjhb * lock. This will only succeed if this thread holds a single read 869157882Sjhb * lock. Returns true if the upgrade succeeded and false otherwise. 870157882Sjhb */ 871157882Sjhbint 872157882Sjhb_rw_try_upgrade(struct rwlock *rw, const char *file, int line) 873157882Sjhb{ 874176017Sjeff uintptr_t v, x, tid; 875170295Sjeff struct turnstile *ts; 876157882Sjhb int success; 877157882Sjhb 878169394Sjhb KASSERT(rw->rw_lock != RW_DESTROYED, 879169394Sjhb ("rw_try_upgrade() of destroyed rwlock @ %s:%d", file, line)); 880157882Sjhb _rw_assert(rw, RA_RLOCKED, file, line); 881157882Sjhb 882157882Sjhb /* 883157882Sjhb * Attempt to switch from one reader to a writer. If there 884157882Sjhb * are any write waiters, then we will have to lock the 885157882Sjhb * turnstile first to prevent races with another writer 886157882Sjhb * calling turnstile_wait() before we have claimed this 887157882Sjhb * turnstile. So, do the simple case of no waiters first. 888157882Sjhb */ 889157882Sjhb tid = (uintptr_t)curthread; 890176017Sjeff success = 0; 891176017Sjeff for (;;) { 892176017Sjeff v = rw->rw_lock; 893176017Sjeff if (RW_READERS(v) > 1) 894176017Sjeff break; 895176017Sjeff if (!(v & RW_LOCK_WAITERS)) { 896176017Sjeff success = atomic_cmpset_ptr(&rw->rw_lock, v, tid); 897176017Sjeff if (!success) 898176017Sjeff continue; 899176017Sjeff break; 900176017Sjeff } 901157882Sjhb 902176017Sjeff /* 903176017Sjeff * Ok, we think we have waiters, so lock the turnstile. 904176017Sjeff */ 905176017Sjeff ts = turnstile_trywait(&rw->lock_object); 906176017Sjeff v = rw->rw_lock; 907176017Sjeff if (RW_READERS(v) > 1) { 908176017Sjeff turnstile_cancel(ts); 909176017Sjeff break; 910176017Sjeff } 911176017Sjeff /* 912176017Sjeff * Try to switch from one reader to a writer again. This time 913176017Sjeff * we honor the current state of the waiters flags. 914176017Sjeff * If we obtain the lock with the flags set, then claim 915176017Sjeff * ownership of the turnstile. 916176017Sjeff */ 917176017Sjeff x = rw->rw_lock & RW_LOCK_WAITERS; 918176017Sjeff success = atomic_cmpset_ptr(&rw->rw_lock, v, tid | x); 919176017Sjeff if (success) { 920176017Sjeff if (x) 921176017Sjeff turnstile_claim(ts); 922176017Sjeff else 923176017Sjeff turnstile_cancel(ts); 924176017Sjeff break; 925176017Sjeff } 926170295Sjeff turnstile_cancel(ts); 927176017Sjeff } 928167787Sjhb LOCK_LOG_TRY("WUPGRADE", &rw->lock_object, 0, success, file, line); 929176017Sjeff if (success) { 930176017Sjeff curthread->td_rw_rlocks--; 931167787Sjhb WITNESS_UPGRADE(&rw->lock_object, LOP_EXCLUSIVE | LOP_TRYLOCK, 932157882Sjhb file, line); 933192853Ssson LOCKSTAT_RECORD0(LS_RW_TRYUPGRADE_UPGRADE, rw); 934176017Sjeff } 935157882Sjhb return (success); 936157882Sjhb} 937157882Sjhb 938157882Sjhb/* 939157882Sjhb * Downgrade a write lock into a single read lock. 940157882Sjhb */ 941157882Sjhbvoid 942157882Sjhb_rw_downgrade(struct rwlock *rw, const char *file, int line) 943157882Sjhb{ 944157882Sjhb struct turnstile *ts; 945157882Sjhb uintptr_t tid, v; 946176017Sjeff int rwait, wwait; 947157882Sjhb 948169394Sjhb KASSERT(rw->rw_lock != RW_DESTROYED, 949169394Sjhb ("rw_downgrade() of destroyed rwlock @ %s:%d", file, line)); 950171052Sattilio _rw_assert(rw, RA_WLOCKED | RA_NOTRECURSED, file, line); 951171052Sattilio#ifndef INVARIANTS 952171052Sattilio if (rw_recursed(rw)) 953171052Sattilio panic("downgrade of a recursed lock"); 954171052Sattilio#endif 955157882Sjhb 956167787Sjhb WITNESS_DOWNGRADE(&rw->lock_object, 0, file, line); 957157882Sjhb 958157882Sjhb /* 959157882Sjhb * Convert from a writer to a single reader. First we handle 960157882Sjhb * the easy case with no waiters. If there are any waiters, we 961176017Sjeff * lock the turnstile and "disown" the lock. 962157882Sjhb */ 963157882Sjhb tid = (uintptr_t)curthread; 964157882Sjhb if (atomic_cmpset_rel_ptr(&rw->rw_lock, tid, RW_READERS_LOCK(1))) 965157882Sjhb goto out; 966157882Sjhb 967157882Sjhb /* 968157882Sjhb * Ok, we think we have waiters, so lock the turnstile so we can 969157882Sjhb * read the waiter flags without any races. 970157882Sjhb */ 971170295Sjeff turnstile_chain_lock(&rw->lock_object); 972176017Sjeff v = rw->rw_lock & RW_LOCK_WAITERS; 973176017Sjeff rwait = v & RW_LOCK_READ_WAITERS; 974176017Sjeff wwait = v & RW_LOCK_WRITE_WAITERS; 975176017Sjeff MPASS(rwait | wwait); 976157882Sjhb 977157882Sjhb /* 978176017Sjeff * Downgrade from a write lock while preserving waiters flag 979176017Sjeff * and give up ownership of the turnstile. 980157882Sjhb */ 981167787Sjhb ts = turnstile_lookup(&rw->lock_object); 982157882Sjhb MPASS(ts != NULL); 983176017Sjeff if (!wwait) 984176017Sjeff v &= ~RW_LOCK_READ_WAITERS; 985176017Sjeff atomic_store_rel_ptr(&rw->rw_lock, RW_READERS_LOCK(1) | v); 986176017Sjeff /* 987176017Sjeff * Wake other readers if there are no writers pending. Otherwise they 988176017Sjeff * won't be able to acquire the lock anyway. 989176017Sjeff */ 990176017Sjeff if (rwait && !wwait) { 991157882Sjhb turnstile_broadcast(ts, TS_SHARED_QUEUE); 992157882Sjhb turnstile_unpend(ts, TS_EXCLUSIVE_LOCK); 993176017Sjeff } else 994157882Sjhb turnstile_disown(ts); 995170295Sjeff turnstile_chain_unlock(&rw->lock_object); 996157882Sjhbout: 997176017Sjeff curthread->td_rw_rlocks++; 998167787Sjhb LOCK_LOG_LOCK("WDOWNGRADE", &rw->lock_object, 0, 0, file, line); 999192853Ssson LOCKSTAT_RECORD0(LS_RW_DOWNGRADE_DOWNGRADE, rw); 1000157882Sjhb} 1001157882Sjhb 1002154941Sjhb#ifdef INVARIANT_SUPPORT 1003155162Sscottl#ifndef INVARIANTS 1004154941Sjhb#undef _rw_assert 1005154941Sjhb#endif 1006154941Sjhb 1007154941Sjhb/* 1008154941Sjhb * In the non-WITNESS case, rw_assert() can only detect that at least 1009154941Sjhb * *some* thread owns an rlock, but it cannot guarantee that *this* 1010154941Sjhb * thread owns an rlock. 1011154941Sjhb */ 1012154941Sjhbvoid 1013154941Sjhb_rw_assert(struct rwlock *rw, int what, const char *file, int line) 1014154941Sjhb{ 1015154941Sjhb 1016154941Sjhb if (panicstr != NULL) 1017154941Sjhb return; 1018154941Sjhb switch (what) { 1019154941Sjhb case RA_LOCKED: 1020171052Sattilio case RA_LOCKED | RA_RECURSED: 1021171052Sattilio case RA_LOCKED | RA_NOTRECURSED: 1022154941Sjhb case RA_RLOCKED: 1023154941Sjhb#ifdef WITNESS 1024167787Sjhb witness_assert(&rw->lock_object, what, file, line); 1025154941Sjhb#else 1026154941Sjhb /* 1027154941Sjhb * If some other thread has a write lock or we have one 1028154941Sjhb * and are asserting a read lock, fail. Also, if no one 1029154941Sjhb * has a lock at all, fail. 1030154941Sjhb */ 1031155061Sscottl if (rw->rw_lock == RW_UNLOCKED || 1032155061Sscottl (!(rw->rw_lock & RW_LOCK_READ) && (what == RA_RLOCKED || 1033157826Sjhb rw_wowner(rw) != curthread))) 1034154941Sjhb panic("Lock %s not %slocked @ %s:%d\n", 1035167787Sjhb rw->lock_object.lo_name, (what == RA_RLOCKED) ? 1036154941Sjhb "read " : "", file, line); 1037171052Sattilio 1038171052Sattilio if (!(rw->rw_lock & RW_LOCK_READ)) { 1039171052Sattilio if (rw_recursed(rw)) { 1040171052Sattilio if (what & RA_NOTRECURSED) 1041171052Sattilio panic("Lock %s recursed @ %s:%d\n", 1042171052Sattilio rw->lock_object.lo_name, file, 1043171052Sattilio line); 1044171052Sattilio } else if (what & RA_RECURSED) 1045171052Sattilio panic("Lock %s not recursed @ %s:%d\n", 1046171052Sattilio rw->lock_object.lo_name, file, line); 1047171052Sattilio } 1048154941Sjhb#endif 1049154941Sjhb break; 1050154941Sjhb case RA_WLOCKED: 1051171052Sattilio case RA_WLOCKED | RA_RECURSED: 1052171052Sattilio case RA_WLOCKED | RA_NOTRECURSED: 1053157826Sjhb if (rw_wowner(rw) != curthread) 1054154941Sjhb panic("Lock %s not exclusively locked @ %s:%d\n", 1055167787Sjhb rw->lock_object.lo_name, file, line); 1056171052Sattilio if (rw_recursed(rw)) { 1057171052Sattilio if (what & RA_NOTRECURSED) 1058171052Sattilio panic("Lock %s recursed @ %s:%d\n", 1059171052Sattilio rw->lock_object.lo_name, file, line); 1060171052Sattilio } else if (what & RA_RECURSED) 1061171052Sattilio panic("Lock %s not recursed @ %s:%d\n", 1062171052Sattilio rw->lock_object.lo_name, file, line); 1063154941Sjhb break; 1064154941Sjhb case RA_UNLOCKED: 1065154941Sjhb#ifdef WITNESS 1066167787Sjhb witness_assert(&rw->lock_object, what, file, line); 1067154941Sjhb#else 1068154941Sjhb /* 1069154941Sjhb * If we hold a write lock fail. We can't reliably check 1070154941Sjhb * to see if we hold a read lock or not. 1071154941Sjhb */ 1072157826Sjhb if (rw_wowner(rw) == curthread) 1073154941Sjhb panic("Lock %s exclusively locked @ %s:%d\n", 1074167787Sjhb rw->lock_object.lo_name, file, line); 1075154941Sjhb#endif 1076154941Sjhb break; 1077154941Sjhb default: 1078154941Sjhb panic("Unknown rw lock assertion: %d @ %s:%d", what, file, 1079154941Sjhb line); 1080154941Sjhb } 1081154941Sjhb} 1082154941Sjhb#endif /* INVARIANT_SUPPORT */ 1083154941Sjhb 1084154941Sjhb#ifdef DDB 1085154941Sjhbvoid 1086154941Sjhbdb_show_rwlock(struct lock_object *lock) 1087154941Sjhb{ 1088154941Sjhb struct rwlock *rw; 1089154941Sjhb struct thread *td; 1090154941Sjhb 1091154941Sjhb rw = (struct rwlock *)lock; 1092154941Sjhb 1093154941Sjhb db_printf(" state: "); 1094154941Sjhb if (rw->rw_lock == RW_UNLOCKED) 1095154941Sjhb db_printf("UNLOCKED\n"); 1096169394Sjhb else if (rw->rw_lock == RW_DESTROYED) { 1097169394Sjhb db_printf("DESTROYED\n"); 1098169394Sjhb return; 1099169394Sjhb } else if (rw->rw_lock & RW_LOCK_READ) 1100167504Sjhb db_printf("RLOCK: %ju locks\n", 1101167504Sjhb (uintmax_t)(RW_READERS(rw->rw_lock))); 1102154941Sjhb else { 1103157826Sjhb td = rw_wowner(rw); 1104154941Sjhb db_printf("WLOCK: %p (tid %d, pid %d, \"%s\")\n", td, 1105173600Sjulian td->td_tid, td->td_proc->p_pid, td->td_name); 1106171052Sattilio if (rw_recursed(rw)) 1107171052Sattilio db_printf(" recursed: %u\n", rw->rw_recurse); 1108154941Sjhb } 1109154941Sjhb db_printf(" waiters: "); 1110154941Sjhb switch (rw->rw_lock & (RW_LOCK_READ_WAITERS | RW_LOCK_WRITE_WAITERS)) { 1111154941Sjhb case RW_LOCK_READ_WAITERS: 1112154941Sjhb db_printf("readers\n"); 1113154941Sjhb break; 1114154941Sjhb case RW_LOCK_WRITE_WAITERS: 1115154941Sjhb db_printf("writers\n"); 1116154941Sjhb break; 1117154941Sjhb case RW_LOCK_READ_WAITERS | RW_LOCK_WRITE_WAITERS: 1118167492Sjhb db_printf("readers and writers\n"); 1119154941Sjhb break; 1120154941Sjhb default: 1121154941Sjhb db_printf("none\n"); 1122154941Sjhb break; 1123154941Sjhb } 1124154941Sjhb} 1125154941Sjhb 1126154941Sjhb#endif 1127