kern_rwlock.c revision 196334
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 196334 2009-08-17 16:17:21Z 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) { 544154941Sjhb if (atomic_cmpset_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)); 562176017Sjeff if (atomic_cmpset_ptr(&rw->rw_lock, x, RW_UNLOCKED)) { 563167787Sjhb if (LOCK_LOG_TEST(&rw->lock_object, 0)) 564154941Sjhb CTR2(KTR_LOCK, "%s: %p last succeeded", 565154941Sjhb __func__, rw); 566154941Sjhb break; 567154941Sjhb } 568154941Sjhb continue; 569154941Sjhb } 570154941Sjhb /* 571176017Sjeff * Ok, we know we have waiters and we think we are the 572176017Sjeff * last reader, so grab the turnstile lock. 573154941Sjhb */ 574170295Sjeff turnstile_chain_lock(&rw->lock_object); 575176017Sjeff v = rw->rw_lock & (RW_LOCK_WAITERS | RW_LOCK_WRITE_SPINNER); 576176017Sjeff MPASS(v & RW_LOCK_WAITERS); 577154941Sjhb 578154941Sjhb /* 579154941Sjhb * Try to drop our lock leaving the lock in a unlocked 580154941Sjhb * state. 581154941Sjhb * 582154941Sjhb * If you wanted to do explicit lock handoff you'd have to 583154941Sjhb * do it here. You'd also want to use turnstile_signal() 584154941Sjhb * and you'd have to handle the race where a higher 585154941Sjhb * priority thread blocks on the write lock before the 586154941Sjhb * thread you wakeup actually runs and have the new thread 587154941Sjhb * "steal" the lock. For now it's a lot simpler to just 588154941Sjhb * wakeup all of the waiters. 589154941Sjhb * 590154941Sjhb * As above, if we fail, then another thread might have 591154941Sjhb * acquired a read lock, so drop the turnstile lock and 592154941Sjhb * restart. 593154941Sjhb */ 594176017Sjeff x = RW_UNLOCKED; 595176017Sjeff if (v & RW_LOCK_WRITE_WAITERS) { 596176017Sjeff queue = TS_EXCLUSIVE_QUEUE; 597176017Sjeff x |= (v & RW_LOCK_READ_WAITERS); 598176017Sjeff } else 599176017Sjeff queue = TS_SHARED_QUEUE; 600176017Sjeff if (!atomic_cmpset_ptr(&rw->rw_lock, RW_READERS_LOCK(1) | v, 601176017Sjeff x)) { 602170295Sjeff turnstile_chain_unlock(&rw->lock_object); 603154941Sjhb continue; 604154941Sjhb } 605167787Sjhb if (LOCK_LOG_TEST(&rw->lock_object, 0)) 606154941Sjhb CTR2(KTR_LOCK, "%s: %p last succeeded with waiters", 607154941Sjhb __func__, rw); 608154941Sjhb 609154941Sjhb /* 610154941Sjhb * Ok. The lock is released and all that's left is to 611154941Sjhb * wake up the waiters. Note that the lock might not be 612154941Sjhb * free anymore, but in that case the writers will just 613154941Sjhb * block again if they run before the new lock holder(s) 614154941Sjhb * release the lock. 615154941Sjhb */ 616167787Sjhb ts = turnstile_lookup(&rw->lock_object); 617157846Sjhb MPASS(ts != NULL); 618176017Sjeff turnstile_broadcast(ts, queue); 619154941Sjhb turnstile_unpend(ts, TS_SHARED_LOCK); 620170295Sjeff turnstile_chain_unlock(&rw->lock_object); 621154941Sjhb break; 622154941Sjhb } 623192853Ssson LOCKSTAT_PROFILE_RELEASE_LOCK(LS_RW_RUNLOCK_RELEASE, rw); 624154941Sjhb} 625154941Sjhb 626154941Sjhb/* 627154941Sjhb * This function is called when we are unable to obtain a write lock on the 628154941Sjhb * first try. This means that at least one other thread holds either a 629154941Sjhb * read or write lock. 630154941Sjhb */ 631154941Sjhbvoid 632154941Sjhb_rw_wlock_hard(struct rwlock *rw, uintptr_t tid, const char *file, int line) 633154941Sjhb{ 634170295Sjeff struct turnstile *ts; 635167801Sjhb#ifdef ADAPTIVE_RWLOCKS 636157846Sjhb volatile struct thread *owner; 637176017Sjeff int spintries = 0; 638176017Sjeff int i; 639157851Swkoszek#endif 640189846Sjeff uintptr_t v, x; 641189846Sjeff#ifdef LOCK_PROFILING 642171516Sattilio uint64_t waittime = 0; 643171516Sattilio int contested = 0; 644189846Sjeff#endif 645192853Ssson#ifdef KDTRACE_HOOKS 646192853Ssson uint64_t spin_cnt = 0; 647192853Ssson uint64_t sleep_cnt = 0; 648192853Ssson int64_t sleep_time = 0; 649192853Ssson#endif 650154941Sjhb 651171052Sattilio if (rw_wlocked(rw)) { 652193307Sattilio KASSERT(rw->lock_object.lo_flags & LO_RECURSABLE, 653171052Sattilio ("%s: recursing but non-recursive rw %s @ %s:%d\n", 654171052Sattilio __func__, rw->lock_object.lo_name, file, line)); 655171052Sattilio rw->rw_recurse++; 656171052Sattilio if (LOCK_LOG_TEST(&rw->lock_object, 0)) 657171052Sattilio CTR2(KTR_LOCK, "%s: %p recursing", __func__, rw); 658171052Sattilio return; 659171052Sattilio } 660171052Sattilio 661167787Sjhb if (LOCK_LOG_TEST(&rw->lock_object, 0)) 662154941Sjhb CTR5(KTR_LOCK, "%s: %s contested (lock=%p) at %s:%d", __func__, 663167787Sjhb rw->lock_object.lo_name, (void *)rw->rw_lock, file, line); 664154941Sjhb 665154941Sjhb while (!_rw_write_lock(rw, tid)) { 666192853Ssson#ifdef KDTRACE_HOOKS 667192853Ssson spin_cnt++; 668192853Ssson#endif 669174629Sjeff lock_profile_obtain_lock_failed(&rw->lock_object, 670174629Sjeff &contested, &waittime); 671173960Sattilio#ifdef ADAPTIVE_RWLOCKS 672173960Sattilio /* 673173960Sattilio * If the lock is write locked and the owner is 674173960Sattilio * running on another CPU, spin until the owner stops 675173960Sattilio * running or the state of the lock changes. 676173960Sattilio */ 677173960Sattilio v = rw->rw_lock; 678173960Sattilio owner = (struct thread *)RW_OWNER(v); 679173960Sattilio if (!(v & RW_LOCK_READ) && TD_IS_RUNNING(owner)) { 680173960Sattilio if (LOCK_LOG_TEST(&rw->lock_object, 0)) 681173960Sattilio CTR3(KTR_LOCK, "%s: spinning on %p held by %p", 682173960Sattilio __func__, rw, owner); 683173960Sattilio while ((struct thread*)RW_OWNER(rw->rw_lock) == owner && 684192853Ssson TD_IS_RUNNING(owner)) { 685173960Sattilio cpu_spinwait(); 686192853Ssson#ifdef KDTRACE_HOOKS 687192853Ssson spin_cnt++; 688192853Ssson#endif 689192853Ssson } 690173960Sattilio continue; 691173960Sattilio } 692177912Sjeff if ((v & RW_LOCK_READ) && RW_READERS(v) && 693177912Sjeff spintries < rowner_retries) { 694176017Sjeff if (!(v & RW_LOCK_WRITE_SPINNER)) { 695176017Sjeff if (!atomic_cmpset_ptr(&rw->rw_lock, v, 696176017Sjeff v | RW_LOCK_WRITE_SPINNER)) { 697176017Sjeff continue; 698176017Sjeff } 699176017Sjeff } 700176017Sjeff spintries++; 701177912Sjeff for (i = 0; i < rowner_loops; i++) { 702176017Sjeff if ((rw->rw_lock & RW_LOCK_WRITE_SPINNER) == 0) 703176017Sjeff break; 704176017Sjeff cpu_spinwait(); 705176017Sjeff } 706192853Ssson#ifdef KDTRACE_HOOKS 707192853Ssson spin_cnt += rowner_loops - i; 708192853Ssson#endif 709177912Sjeff if (i != rowner_loops) 710176017Sjeff continue; 711176017Sjeff } 712173960Sattilio#endif 713170295Sjeff ts = turnstile_trywait(&rw->lock_object); 714154941Sjhb v = rw->rw_lock; 715154941Sjhb 716173960Sattilio#ifdef ADAPTIVE_RWLOCKS 717154941Sjhb /* 718193035Sjhb * The current lock owner might have started executing 719193035Sjhb * on another CPU (or the lock could have changed 720193035Sjhb * owners) while we were waiting on the turnstile 721193035Sjhb * chain lock. If so, drop the turnstile lock and try 722193035Sjhb * again. 723173960Sattilio */ 724173960Sattilio if (!(v & RW_LOCK_READ)) { 725173960Sattilio owner = (struct thread *)RW_OWNER(v); 726173960Sattilio if (TD_IS_RUNNING(owner)) { 727173960Sattilio turnstile_cancel(ts); 728173960Sattilio continue; 729173960Sattilio } 730173960Sattilio } 731173960Sattilio#endif 732173960Sattilio /* 733179334Sattilio * Check for the waiters flags about this rwlock. 734179334Sattilio * If the lock was released, without maintain any pending 735179334Sattilio * waiters queue, simply try to acquire it. 736179334Sattilio * If a pending waiters queue is present, claim the lock 737179334Sattilio * ownership and maintain the pending queue. 738154941Sjhb */ 739176017Sjeff x = v & (RW_LOCK_WAITERS | RW_LOCK_WRITE_SPINNER); 740176017Sjeff if ((v & ~x) == RW_UNLOCKED) { 741176017Sjeff x &= ~RW_LOCK_WRITE_SPINNER; 742176017Sjeff if (atomic_cmpset_acq_ptr(&rw->rw_lock, v, tid | x)) { 743176017Sjeff if (x) 744176017Sjeff turnstile_claim(ts); 745176017Sjeff else 746176017Sjeff turnstile_cancel(ts); 747154941Sjhb break; 748154941Sjhb } 749170295Sjeff turnstile_cancel(ts); 750154941Sjhb continue; 751154941Sjhb } 752154941Sjhb /* 753154941Sjhb * If the RW_LOCK_WRITE_WAITERS flag isn't set, then try to 754154941Sjhb * set it. If we fail to set it, then loop back and try 755154941Sjhb * again. 756154941Sjhb */ 757157826Sjhb if (!(v & RW_LOCK_WRITE_WAITERS)) { 758157826Sjhb if (!atomic_cmpset_ptr(&rw->rw_lock, v, 759157826Sjhb v | RW_LOCK_WRITE_WAITERS)) { 760170295Sjeff turnstile_cancel(ts); 761157826Sjhb continue; 762157826Sjhb } 763167787Sjhb if (LOCK_LOG_TEST(&rw->lock_object, 0)) 764157826Sjhb CTR2(KTR_LOCK, "%s: %p set write waiters flag", 765157826Sjhb __func__, rw); 766154941Sjhb } 767157846Sjhb /* 768154941Sjhb * We were unable to acquire the lock and the write waiters 769154941Sjhb * flag is set, so we must block on the turnstile. 770154941Sjhb */ 771167787Sjhb if (LOCK_LOG_TEST(&rw->lock_object, 0)) 772154941Sjhb CTR2(KTR_LOCK, "%s: %p blocking on turnstile", __func__, 773154941Sjhb rw); 774192853Ssson#ifdef KDTRACE_HOOKS 775192853Ssson sleep_time -= lockstat_nsecs(); 776192853Ssson#endif 777170295Sjeff turnstile_wait(ts, rw_owner(rw), TS_EXCLUSIVE_QUEUE); 778192853Ssson#ifdef KDTRACE_HOOKS 779192853Ssson sleep_time += lockstat_nsecs(); 780192853Ssson sleep_cnt++; 781192853Ssson#endif 782167787Sjhb if (LOCK_LOG_TEST(&rw->lock_object, 0)) 783154941Sjhb CTR2(KTR_LOCK, "%s: %p resuming from turnstile", 784154941Sjhb __func__, rw); 785176017Sjeff#ifdef ADAPTIVE_RWLOCKS 786176017Sjeff spintries = 0; 787176017Sjeff#endif 788154941Sjhb } 789192853Ssson LOCKSTAT_PROFILE_OBTAIN_LOCK_SUCCESS(LS_RW_WLOCK_ACQUIRE, rw, contested, 790192853Ssson waittime, file, line); 791192853Ssson#ifdef KDTRACE_HOOKS 792192853Ssson if (sleep_time) 793192853Ssson LOCKSTAT_RECORD1(LS_RW_WLOCK_BLOCK, rw, sleep_time); 794192853Ssson 795192853Ssson /* 796192853Ssson * Record only the loops spinning and not sleeping. 797192853Ssson */ 798192853Ssson if (spin_cnt > sleep_cnt) 799192853Ssson LOCKSTAT_RECORD1(LS_RW_WLOCK_SPIN, rw, (spin_cnt - sleep_cnt)); 800192853Ssson#endif 801154941Sjhb} 802154941Sjhb 803154941Sjhb/* 804154941Sjhb * This function is called if the first try at releasing a write lock failed. 805154941Sjhb * This means that one of the 2 waiter bits must be set indicating that at 806154941Sjhb * least one thread is waiting on this lock. 807154941Sjhb */ 808154941Sjhbvoid 809154941Sjhb_rw_wunlock_hard(struct rwlock *rw, uintptr_t tid, const char *file, int line) 810154941Sjhb{ 811154941Sjhb struct turnstile *ts; 812154941Sjhb uintptr_t v; 813154941Sjhb int queue; 814154941Sjhb 815171052Sattilio if (rw_wlocked(rw) && rw_recursed(rw)) { 816176017Sjeff rw->rw_recurse--; 817171052Sattilio if (LOCK_LOG_TEST(&rw->lock_object, 0)) 818171052Sattilio CTR2(KTR_LOCK, "%s: %p unrecursing", __func__, rw); 819171052Sattilio return; 820171052Sattilio } 821171052Sattilio 822154941Sjhb KASSERT(rw->rw_lock & (RW_LOCK_READ_WAITERS | RW_LOCK_WRITE_WAITERS), 823154941Sjhb ("%s: neither of the waiter flags are set", __func__)); 824154941Sjhb 825167787Sjhb if (LOCK_LOG_TEST(&rw->lock_object, 0)) 826154941Sjhb CTR2(KTR_LOCK, "%s: %p contested", __func__, rw); 827154941Sjhb 828170295Sjeff turnstile_chain_lock(&rw->lock_object); 829167787Sjhb ts = turnstile_lookup(&rw->lock_object); 830154941Sjhb MPASS(ts != NULL); 831154941Sjhb 832154941Sjhb /* 833154941Sjhb * Use the same algo as sx locks for now. Prefer waking up shared 834154941Sjhb * waiters if we have any over writers. This is probably not ideal. 835154941Sjhb * 836154941Sjhb * 'v' is the value we are going to write back to rw_lock. If we 837154941Sjhb * have waiters on both queues, we need to preserve the state of 838154941Sjhb * the waiter flag for the queue we don't wake up. For now this is 839154941Sjhb * hardcoded for the algorithm mentioned above. 840154941Sjhb * 841154941Sjhb * In the case of both readers and writers waiting we wakeup the 842154941Sjhb * readers but leave the RW_LOCK_WRITE_WAITERS flag set. If a 843154941Sjhb * new writer comes in before a reader it will claim the lock up 844154941Sjhb * above. There is probably a potential priority inversion in 845154941Sjhb * there that could be worked around either by waking both queues 846154941Sjhb * of waiters or doing some complicated lock handoff gymnastics. 847154941Sjhb */ 848157846Sjhb v = RW_UNLOCKED; 849176076Sjeff if (rw->rw_lock & RW_LOCK_WRITE_WAITERS) { 850176076Sjeff queue = TS_EXCLUSIVE_QUEUE; 851176076Sjeff v |= (rw->rw_lock & RW_LOCK_READ_WAITERS); 852176076Sjeff } else 853154941Sjhb queue = TS_SHARED_QUEUE; 854157846Sjhb 855157846Sjhb /* Wake up all waiters for the specific queue. */ 856167787Sjhb if (LOCK_LOG_TEST(&rw->lock_object, 0)) 857154941Sjhb CTR3(KTR_LOCK, "%s: %p waking up %s waiters", __func__, rw, 858154941Sjhb queue == TS_SHARED_QUEUE ? "read" : "write"); 859154941Sjhb turnstile_broadcast(ts, queue); 860154941Sjhb atomic_store_rel_ptr(&rw->rw_lock, v); 861154941Sjhb turnstile_unpend(ts, TS_EXCLUSIVE_LOCK); 862170295Sjeff turnstile_chain_unlock(&rw->lock_object); 863154941Sjhb} 864154941Sjhb 865157882Sjhb/* 866157882Sjhb * Attempt to do a non-blocking upgrade from a read lock to a write 867157882Sjhb * lock. This will only succeed if this thread holds a single read 868157882Sjhb * lock. Returns true if the upgrade succeeded and false otherwise. 869157882Sjhb */ 870157882Sjhbint 871157882Sjhb_rw_try_upgrade(struct rwlock *rw, const char *file, int line) 872157882Sjhb{ 873176017Sjeff uintptr_t v, x, tid; 874170295Sjeff struct turnstile *ts; 875157882Sjhb int success; 876157882Sjhb 877169394Sjhb KASSERT(rw->rw_lock != RW_DESTROYED, 878169394Sjhb ("rw_try_upgrade() of destroyed rwlock @ %s:%d", file, line)); 879157882Sjhb _rw_assert(rw, RA_RLOCKED, file, line); 880157882Sjhb 881157882Sjhb /* 882157882Sjhb * Attempt to switch from one reader to a writer. If there 883157882Sjhb * are any write waiters, then we will have to lock the 884157882Sjhb * turnstile first to prevent races with another writer 885157882Sjhb * calling turnstile_wait() before we have claimed this 886157882Sjhb * turnstile. So, do the simple case of no waiters first. 887157882Sjhb */ 888157882Sjhb tid = (uintptr_t)curthread; 889176017Sjeff success = 0; 890176017Sjeff for (;;) { 891176017Sjeff v = rw->rw_lock; 892176017Sjeff if (RW_READERS(v) > 1) 893176017Sjeff break; 894176017Sjeff if (!(v & RW_LOCK_WAITERS)) { 895176017Sjeff success = atomic_cmpset_ptr(&rw->rw_lock, v, tid); 896176017Sjeff if (!success) 897176017Sjeff continue; 898176017Sjeff break; 899176017Sjeff } 900157882Sjhb 901176017Sjeff /* 902176017Sjeff * Ok, we think we have waiters, so lock the turnstile. 903176017Sjeff */ 904176017Sjeff ts = turnstile_trywait(&rw->lock_object); 905176017Sjeff v = rw->rw_lock; 906176017Sjeff if (RW_READERS(v) > 1) { 907176017Sjeff turnstile_cancel(ts); 908176017Sjeff break; 909176017Sjeff } 910176017Sjeff /* 911176017Sjeff * Try to switch from one reader to a writer again. This time 912176017Sjeff * we honor the current state of the waiters flags. 913176017Sjeff * If we obtain the lock with the flags set, then claim 914176017Sjeff * ownership of the turnstile. 915176017Sjeff */ 916176017Sjeff x = rw->rw_lock & RW_LOCK_WAITERS; 917176017Sjeff success = atomic_cmpset_ptr(&rw->rw_lock, v, tid | x); 918176017Sjeff if (success) { 919176017Sjeff if (x) 920176017Sjeff turnstile_claim(ts); 921176017Sjeff else 922176017Sjeff turnstile_cancel(ts); 923176017Sjeff break; 924176017Sjeff } 925170295Sjeff turnstile_cancel(ts); 926176017Sjeff } 927167787Sjhb LOCK_LOG_TRY("WUPGRADE", &rw->lock_object, 0, success, file, line); 928176017Sjeff if (success) { 929176017Sjeff curthread->td_rw_rlocks--; 930167787Sjhb WITNESS_UPGRADE(&rw->lock_object, LOP_EXCLUSIVE | LOP_TRYLOCK, 931157882Sjhb file, line); 932192853Ssson LOCKSTAT_RECORD0(LS_RW_TRYUPGRADE_UPGRADE, rw); 933176017Sjeff } 934157882Sjhb return (success); 935157882Sjhb} 936157882Sjhb 937157882Sjhb/* 938157882Sjhb * Downgrade a write lock into a single read lock. 939157882Sjhb */ 940157882Sjhbvoid 941157882Sjhb_rw_downgrade(struct rwlock *rw, const char *file, int line) 942157882Sjhb{ 943157882Sjhb struct turnstile *ts; 944157882Sjhb uintptr_t tid, v; 945176017Sjeff int rwait, wwait; 946157882Sjhb 947169394Sjhb KASSERT(rw->rw_lock != RW_DESTROYED, 948169394Sjhb ("rw_downgrade() of destroyed rwlock @ %s:%d", file, line)); 949171052Sattilio _rw_assert(rw, RA_WLOCKED | RA_NOTRECURSED, file, line); 950171052Sattilio#ifndef INVARIANTS 951171052Sattilio if (rw_recursed(rw)) 952171052Sattilio panic("downgrade of a recursed lock"); 953171052Sattilio#endif 954157882Sjhb 955167787Sjhb WITNESS_DOWNGRADE(&rw->lock_object, 0, file, line); 956157882Sjhb 957157882Sjhb /* 958157882Sjhb * Convert from a writer to a single reader. First we handle 959157882Sjhb * the easy case with no waiters. If there are any waiters, we 960176017Sjeff * lock the turnstile and "disown" the lock. 961157882Sjhb */ 962157882Sjhb tid = (uintptr_t)curthread; 963157882Sjhb if (atomic_cmpset_rel_ptr(&rw->rw_lock, tid, RW_READERS_LOCK(1))) 964157882Sjhb goto out; 965157882Sjhb 966157882Sjhb /* 967157882Sjhb * Ok, we think we have waiters, so lock the turnstile so we can 968157882Sjhb * read the waiter flags without any races. 969157882Sjhb */ 970170295Sjeff turnstile_chain_lock(&rw->lock_object); 971176017Sjeff v = rw->rw_lock & RW_LOCK_WAITERS; 972176017Sjeff rwait = v & RW_LOCK_READ_WAITERS; 973176017Sjeff wwait = v & RW_LOCK_WRITE_WAITERS; 974176017Sjeff MPASS(rwait | wwait); 975157882Sjhb 976157882Sjhb /* 977176017Sjeff * Downgrade from a write lock while preserving waiters flag 978176017Sjeff * and give up ownership of the turnstile. 979157882Sjhb */ 980167787Sjhb ts = turnstile_lookup(&rw->lock_object); 981157882Sjhb MPASS(ts != NULL); 982176017Sjeff if (!wwait) 983176017Sjeff v &= ~RW_LOCK_READ_WAITERS; 984176017Sjeff atomic_store_rel_ptr(&rw->rw_lock, RW_READERS_LOCK(1) | v); 985176017Sjeff /* 986176017Sjeff * Wake other readers if there are no writers pending. Otherwise they 987176017Sjeff * won't be able to acquire the lock anyway. 988176017Sjeff */ 989176017Sjeff if (rwait && !wwait) { 990157882Sjhb turnstile_broadcast(ts, TS_SHARED_QUEUE); 991157882Sjhb turnstile_unpend(ts, TS_EXCLUSIVE_LOCK); 992176017Sjeff } else 993157882Sjhb turnstile_disown(ts); 994170295Sjeff turnstile_chain_unlock(&rw->lock_object); 995157882Sjhbout: 996176017Sjeff curthread->td_rw_rlocks++; 997167787Sjhb LOCK_LOG_LOCK("WDOWNGRADE", &rw->lock_object, 0, 0, file, line); 998192853Ssson LOCKSTAT_RECORD0(LS_RW_DOWNGRADE_DOWNGRADE, rw); 999157882Sjhb} 1000157882Sjhb 1001154941Sjhb#ifdef INVARIANT_SUPPORT 1002155162Sscottl#ifndef INVARIANTS 1003154941Sjhb#undef _rw_assert 1004154941Sjhb#endif 1005154941Sjhb 1006154941Sjhb/* 1007154941Sjhb * In the non-WITNESS case, rw_assert() can only detect that at least 1008154941Sjhb * *some* thread owns an rlock, but it cannot guarantee that *this* 1009154941Sjhb * thread owns an rlock. 1010154941Sjhb */ 1011154941Sjhbvoid 1012154941Sjhb_rw_assert(struct rwlock *rw, int what, const char *file, int line) 1013154941Sjhb{ 1014154941Sjhb 1015154941Sjhb if (panicstr != NULL) 1016154941Sjhb return; 1017154941Sjhb switch (what) { 1018154941Sjhb case RA_LOCKED: 1019171052Sattilio case RA_LOCKED | RA_RECURSED: 1020171052Sattilio case RA_LOCKED | RA_NOTRECURSED: 1021154941Sjhb case RA_RLOCKED: 1022154941Sjhb#ifdef WITNESS 1023167787Sjhb witness_assert(&rw->lock_object, what, file, line); 1024154941Sjhb#else 1025154941Sjhb /* 1026154941Sjhb * If some other thread has a write lock or we have one 1027154941Sjhb * and are asserting a read lock, fail. Also, if no one 1028154941Sjhb * has a lock at all, fail. 1029154941Sjhb */ 1030155061Sscottl if (rw->rw_lock == RW_UNLOCKED || 1031155061Sscottl (!(rw->rw_lock & RW_LOCK_READ) && (what == RA_RLOCKED || 1032157826Sjhb rw_wowner(rw) != curthread))) 1033154941Sjhb panic("Lock %s not %slocked @ %s:%d\n", 1034167787Sjhb rw->lock_object.lo_name, (what == RA_RLOCKED) ? 1035154941Sjhb "read " : "", file, line); 1036171052Sattilio 1037171052Sattilio if (!(rw->rw_lock & RW_LOCK_READ)) { 1038171052Sattilio if (rw_recursed(rw)) { 1039171052Sattilio if (what & RA_NOTRECURSED) 1040171052Sattilio panic("Lock %s recursed @ %s:%d\n", 1041171052Sattilio rw->lock_object.lo_name, file, 1042171052Sattilio line); 1043171052Sattilio } else if (what & RA_RECURSED) 1044171052Sattilio panic("Lock %s not recursed @ %s:%d\n", 1045171052Sattilio rw->lock_object.lo_name, file, line); 1046171052Sattilio } 1047154941Sjhb#endif 1048154941Sjhb break; 1049154941Sjhb case RA_WLOCKED: 1050171052Sattilio case RA_WLOCKED | RA_RECURSED: 1051171052Sattilio case RA_WLOCKED | RA_NOTRECURSED: 1052157826Sjhb if (rw_wowner(rw) != curthread) 1053154941Sjhb panic("Lock %s not exclusively locked @ %s:%d\n", 1054167787Sjhb rw->lock_object.lo_name, file, line); 1055171052Sattilio if (rw_recursed(rw)) { 1056171052Sattilio if (what & RA_NOTRECURSED) 1057171052Sattilio panic("Lock %s recursed @ %s:%d\n", 1058171052Sattilio rw->lock_object.lo_name, file, line); 1059171052Sattilio } else if (what & RA_RECURSED) 1060171052Sattilio panic("Lock %s not recursed @ %s:%d\n", 1061171052Sattilio rw->lock_object.lo_name, file, line); 1062154941Sjhb break; 1063154941Sjhb case RA_UNLOCKED: 1064154941Sjhb#ifdef WITNESS 1065167787Sjhb witness_assert(&rw->lock_object, what, file, line); 1066154941Sjhb#else 1067154941Sjhb /* 1068154941Sjhb * If we hold a write lock fail. We can't reliably check 1069154941Sjhb * to see if we hold a read lock or not. 1070154941Sjhb */ 1071157826Sjhb if (rw_wowner(rw) == curthread) 1072154941Sjhb panic("Lock %s exclusively locked @ %s:%d\n", 1073167787Sjhb rw->lock_object.lo_name, file, line); 1074154941Sjhb#endif 1075154941Sjhb break; 1076154941Sjhb default: 1077154941Sjhb panic("Unknown rw lock assertion: %d @ %s:%d", what, file, 1078154941Sjhb line); 1079154941Sjhb } 1080154941Sjhb} 1081154941Sjhb#endif /* INVARIANT_SUPPORT */ 1082154941Sjhb 1083154941Sjhb#ifdef DDB 1084154941Sjhbvoid 1085154941Sjhbdb_show_rwlock(struct lock_object *lock) 1086154941Sjhb{ 1087154941Sjhb struct rwlock *rw; 1088154941Sjhb struct thread *td; 1089154941Sjhb 1090154941Sjhb rw = (struct rwlock *)lock; 1091154941Sjhb 1092154941Sjhb db_printf(" state: "); 1093154941Sjhb if (rw->rw_lock == RW_UNLOCKED) 1094154941Sjhb db_printf("UNLOCKED\n"); 1095169394Sjhb else if (rw->rw_lock == RW_DESTROYED) { 1096169394Sjhb db_printf("DESTROYED\n"); 1097169394Sjhb return; 1098169394Sjhb } else if (rw->rw_lock & RW_LOCK_READ) 1099167504Sjhb db_printf("RLOCK: %ju locks\n", 1100167504Sjhb (uintmax_t)(RW_READERS(rw->rw_lock))); 1101154941Sjhb else { 1102157826Sjhb td = rw_wowner(rw); 1103154941Sjhb db_printf("WLOCK: %p (tid %d, pid %d, \"%s\")\n", td, 1104173600Sjulian td->td_tid, td->td_proc->p_pid, td->td_name); 1105171052Sattilio if (rw_recursed(rw)) 1106171052Sattilio db_printf(" recursed: %u\n", rw->rw_recurse); 1107154941Sjhb } 1108154941Sjhb db_printf(" waiters: "); 1109154941Sjhb switch (rw->rw_lock & (RW_LOCK_READ_WAITERS | RW_LOCK_WRITE_WAITERS)) { 1110154941Sjhb case RW_LOCK_READ_WAITERS: 1111154941Sjhb db_printf("readers\n"); 1112154941Sjhb break; 1113154941Sjhb case RW_LOCK_WRITE_WAITERS: 1114154941Sjhb db_printf("writers\n"); 1115154941Sjhb break; 1116154941Sjhb case RW_LOCK_READ_WAITERS | RW_LOCK_WRITE_WAITERS: 1117167492Sjhb db_printf("readers and writers\n"); 1118154941Sjhb break; 1119154941Sjhb default: 1120154941Sjhb db_printf("none\n"); 1121154941Sjhb break; 1122154941Sjhb } 1123154941Sjhb} 1124154941Sjhb 1125154941Sjhb#endif 1126