kern_rwlock.c revision 227588
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 227588 2011-11-16 21:51:17Z pjd $"); 36154941Sjhb 37154941Sjhb#include "opt_ddb.h" 38154941Sjhb#include "opt_kdtrace.h" 39154941Sjhb#include "opt_no_adaptive_rwlocks.h" 40154941Sjhb 41154941Sjhb#include <sys/param.h> 42154941Sjhb#include <sys/ktr.h> 43154941Sjhb#include <sys/kernel.h> 44154941Sjhb#include <sys/lock.h> 45154941Sjhb#include <sys/mutex.h> 46154941Sjhb#include <sys/proc.h> 47164159Skmacy#include <sys/rwlock.h> 48154941Sjhb#include <sys/sysctl.h> 49154941Sjhb#include <sys/systm.h> 50154941Sjhb#include <sys/turnstile.h> 51154941Sjhb 52154941Sjhb#include <machine/cpu.h> 53154941Sjhb 54154941Sjhb#if defined(SMP) && !defined(NO_ADAPTIVE_RWLOCKS) 55167368Sjhb#define ADAPTIVE_RWLOCKS 56167368Sjhb#endif 57154941Sjhb 58154941Sjhb#ifdef ADAPTIVE_RWLOCKS 59167365Sjhbstatic int rowner_retries = 10; 60167365Sjhbstatic int rowner_loops = 10000; 61154941Sjhbstatic SYSCTL_NODE(_debug, OID_AUTO, rwlock, CTLFLAG_RD, NULL, 62167365Sjhb "rwlock debugging"); 63154941SjhbSYSCTL_INT(_debug_rwlock, OID_AUTO, retry, CTLFLAG_RW, &rowner_retries, 0, ""); 64167368SjhbSYSCTL_INT(_debug_rwlock, OID_AUTO, loops, CTLFLAG_RW, &rowner_loops, 0, ""); 65167368Sjhb#endif 66154941Sjhb 67154941Sjhb#ifdef DDB 68157826Sjhb#include <ddb/ddb.h> 69157826Sjhb 70157826Sjhbstatic void db_show_rwlock(const struct lock_object *lock); 71157826Sjhb#endif 72157826Sjhbstatic void assert_rw(const struct lock_object *lock, int what); 73154941Sjhbstatic void lock_rw(struct lock_object *lock, int how); 74154941Sjhb#ifdef KDTRACE_HOOKS 75154941Sjhbstatic int owner_rw(const struct lock_object *lock, struct thread **owner); 76157826Sjhb#endif 77157826Sjhbstatic int unlock_rw(struct lock_object *lock); 78157826Sjhb 79157826Sjhbstruct lock_class lock_class_rw = { 80157826Sjhb .lc_name = "rw", 81157826Sjhb .lc_flags = LC_SLEEPLOCK | LC_RECURSABLE | LC_UPGRADABLE, 82157826Sjhb .lc_assert = assert_rw, 83154941Sjhb#ifdef DDB 84154941Sjhb .lc_ddb_show = db_show_rwlock, 85154941Sjhb#endif 86154941Sjhb .lc_lock = lock_rw, 87154941Sjhb .lc_unlock = unlock_rw, 88167368Sjhb#ifdef KDTRACE_HOOKS 89167368Sjhb .lc_owner = owner_rw, 90167368Sjhb#endif 91167368Sjhb}; 92167368Sjhb 93167368Sjhb/* 94167368Sjhb * Return a pointer to the owning thread if the lock is write-locked or 95167368Sjhb * NULL if the lock is unlocked or read-locked. 96167368Sjhb */ 97167368Sjhb#define rw_wowner(rw) \ 98167368Sjhb ((rw)->rw_lock & RW_LOCK_READ ? NULL : \ 99167368Sjhb (struct thread *)RW_OWNER((rw)->rw_lock)) 100167368Sjhb 101167368Sjhb/* 102167368Sjhb * Returns if a write owner is recursed. Write ownership is not assured 103167368Sjhb * here and should be previously checked. 104167368Sjhb */ 105167368Sjhb#define rw_recursed(rw) ((rw)->rw_recurse != 0) 106167368Sjhb 107167368Sjhb/* 108167368Sjhb * Return true if curthread helds the lock. 109167368Sjhb */ 110167368Sjhb#define rw_wlocked(rw) (rw_wowner((rw)) == curthread) 111167368Sjhb 112167368Sjhb/* 113167368Sjhb * Return a pointer to the owning thread for this lock who should receive 114167368Sjhb * any priority lent by threads that block on this lock. Currently this 115167368Sjhb * is identical to rw_wowner(). 116154941Sjhb */ 117154941Sjhb#define rw_owner(rw) rw_wowner(rw) 118154941Sjhb 119154941Sjhb#ifndef INVARIANTS 120154941Sjhb#define _rw_assert(rw, what, file, line) 121167787Sjhb#endif 122167787Sjhb 123157882Sjhbvoid 124154941Sjhbassert_rw(const struct lock_object *lock, int what) 125154941Sjhb{ 126154941Sjhb 127154941Sjhb rw_assert((const struct rwlock *)lock, what); 128154941Sjhb} 129154941Sjhb 130154941Sjhbvoid 131167787Sjhblock_rw(struct lock_object *lock, int how) 132167787Sjhb{ 133154941Sjhb struct rwlock *rw; 134154941Sjhb 135154941Sjhb rw = (struct rwlock *)lock; 136154941Sjhb if (how) 137154941Sjhb rw_wlock(rw); 138154941Sjhb else 139154941Sjhb rw_rlock(rw); 140154941Sjhb} 141154941Sjhb 142154941Sjhbint 143167024Srwatsonunlock_rw(struct lock_object *lock) 144167024Srwatson{ 145167024Srwatson struct rwlock *rw; 146167024Srwatson 147167024Srwatson rw = (struct rwlock *)lock; 148167024Srwatson rw_assert(rw, RA_LOCKED | LA_NOTRECURSED); 149167024Srwatson if (rw->rw_lock & RW_LOCK_READ) { 150154941Sjhb rw_runlock(rw); 151154941Sjhb return (0); 152154941Sjhb } else { 153154941Sjhb rw_wunlock(rw); 154154941Sjhb return (1); 155157826Sjhb } 156154941Sjhb} 157167787Sjhb 158167787Sjhb#ifdef KDTRACE_HOOKS 159154941Sjhbint 160154941Sjhbowner_rw(const struct lock_object *lock, struct thread **owner) 161167787Sjhb{ 162167787Sjhb const struct rwlock *rw = (const struct rwlock *)lock; 163160771Sjhb uintptr_t x = rw->rw_lock; 164154941Sjhb 165154941Sjhb *owner = rw_wowner(rw); 166154941Sjhb return ((x & RW_LOCK_READ) != 0 ? (RW_READERS(x) != 0) : 167154941Sjhb (*owner != NULL)); 168154941Sjhb} 169154941Sjhb#endif 170154941Sjhb 171154941Sjhbvoid 172160771Sjhbrw_init_flags(struct rwlock *rw, const char *name, int opts) 173167787Sjhb{ 174167787Sjhb int flags; 175167787Sjhb 176154941Sjhb MPASS((opts & ~(RW_DUPOK | RW_NOPROFILE | RW_NOWITNESS | RW_QUIET | 177154941Sjhb RW_RECURSE)) == 0); 178154941Sjhb ASSERT_ATOMIC_LOAD_PTR(rw->rw_lock, 179154941Sjhb ("%s: rw_lock not aligned for %s: %p", __func__, name, 180154941Sjhb &rw->rw_lock)); 181154941Sjhb 182157851Swkoszek flags = LO_UPGRADABLE; 183157846Sjhb if (opts & RW_DUPOK) 184157851Swkoszek flags |= LO_DUPOK; 185167307Sjhb if (opts & RW_NOPROFILE) 186167054Skmacy flags |= LO_NOPROFILE; 187154941Sjhb if (!(opts & RW_NOWITNESS)) 188154941Sjhb flags |= LO_WITNESS; 189157826Sjhb if (opts & RW_RECURSE) 190154941Sjhb flags |= LO_RECURSABLE; 191167787Sjhb if (opts & RW_QUIET) 192167787Sjhb flags |= LO_QUIET; 193154941Sjhb 194154941Sjhb rw->rw_lock = RW_UNLOCKED; 195154941Sjhb rw->rw_recurse = 0; 196154941Sjhb lock_init(&rw->lock_object, &lock_class_rw, name, NULL, flags); 197154941Sjhb} 198154941Sjhb 199154941Sjhbvoid 200154941Sjhbrw_destroy(struct rwlock *rw) 201154941Sjhb{ 202154941Sjhb 203154941Sjhb KASSERT(rw->rw_lock == RW_UNLOCKED, ("rw lock %p not unlocked", rw)); 204154941Sjhb KASSERT(rw->rw_recurse == 0, ("rw lock %p still recursed", rw)); 205154941Sjhb rw->rw_lock = RW_DESTROYED; 206154941Sjhb lock_destroy(&rw->lock_object); 207154941Sjhb} 208154941Sjhb 209154941Sjhbvoid 210154941Sjhbrw_sysinit(void *arg) 211154941Sjhb{ 212154941Sjhb struct rw_args *args = arg; 213154941Sjhb 214154941Sjhb rw_init(args->ra_rw, args->ra_desc); 215154941Sjhb} 216154941Sjhb 217154941Sjhbvoid 218154941Sjhbrw_sysinit_flags(void *arg) 219154941Sjhb{ 220154941Sjhb struct rw_args_flags *args = arg; 221154941Sjhb 222154941Sjhb rw_init_flags(args->ra_rw, args->ra_desc, args->ra_flags); 223154941Sjhb} 224154941Sjhb 225154941Sjhbint 226154941Sjhbrw_wowned(const struct rwlock *rw) 227167787Sjhb{ 228154941Sjhb 229154941Sjhb return (rw_wowner(rw) == curthread); 230154941Sjhb} 231154941Sjhb 232167307Sjhbvoid 233167307Sjhb_rw_wlock(struct rwlock *rw, const char *file, int line) 234167787Sjhb{ 235167307Sjhb 236154941Sjhb MPASS(curthread != NULL); 237154941Sjhb KASSERT(rw->rw_lock != RW_DESTROYED, 238157846Sjhb ("rw_wlock() of destroyed rwlock @ %s:%d", file, line)); 239154941Sjhb WITNESS_CHECKORDER(&rw->lock_object, LOP_NEWORDER | LOP_EXCLUSIVE, file, 240154941Sjhb line, NULL); 241167787Sjhb __rw_wlock(rw, curthread, file, line); 242167307Sjhb LOCK_LOG_LOCK("WLOCK", &rw->lock_object, 0, rw->rw_recurse, file, line); 243154941Sjhb WITNESS_LOCK(&rw->lock_object, LOP_EXCLUSIVE, file, line); 244154941Sjhb curthread->td_locks++; 245154941Sjhb} 246154941Sjhb 247154941Sjhbint 248154941Sjhb_rw_try_wlock(struct rwlock *rw, const char *file, int line) 249167787Sjhb{ 250154941Sjhb int rval; 251154941Sjhb 252154941Sjhb KASSERT(rw->rw_lock != RW_DESTROYED, 253154941Sjhb ("rw_try_wlock() of destroyed rwlock @ %s:%d", file, line)); 254154941Sjhb 255154941Sjhb if (rw_wlocked(rw) && 256154941Sjhb (rw->lock_object.lo_flags & LO_RECURSABLE) != 0) { 257154941Sjhb rw->rw_recurse++; 258167787Sjhb rval = 1; 259157846Sjhb } else 260154941Sjhb rval = atomic_cmpset_acq_ptr(&rw->rw_lock, RW_UNLOCKED, 261154941Sjhb (uintptr_t)curthread); 262154941Sjhb 263154941Sjhb LOCK_LOG_TRY("WLOCK", &rw->lock_object, 0, rval, file, line); 264154941Sjhb if (rval) { 265154941Sjhb WITNESS_LOCK(&rw->lock_object, LOP_EXCLUSIVE | LOP_TRYLOCK, 266154941Sjhb file, line); 267154941Sjhb curthread->td_locks++; 268154941Sjhb } 269157826Sjhb return (rval); 270157826Sjhb} 271157826Sjhb 272167787Sjhbvoid 273157826Sjhb_rw_wunlock(struct rwlock *rw, const char *file, int line) 274157826Sjhb{ 275157826Sjhb 276167787Sjhb MPASS(curthread != NULL); 277157826Sjhb KASSERT(rw->rw_lock != RW_DESTROYED, 278157826Sjhb ("rw_wunlock() of destroyed rwlock @ %s:%d", file, line)); 279154941Sjhb _rw_assert(rw, RA_WLOCKED, file, line); 280154941Sjhb curthread->td_locks--; 281157846Sjhb WITNESS_UNLOCK(&rw->lock_object, LOP_EXCLUSIVE, file, line); 282154941Sjhb LOCK_LOG_LOCK("WUNLOCK", &rw->lock_object, 0, rw->rw_recurse, file, 283157846Sjhb line); 284157846Sjhb if (!rw_recursed(rw)) 285157846Sjhb LOCKSTAT_PROFILE_RELEASE_LOCK(LS_RW_WUNLOCK_RELEASE, rw); 286157846Sjhb __rw_wunlock(rw, curthread, file, line); 287157846Sjhb} 288157846Sjhb/* 289167787Sjhb * Determines whether a new reader can acquire a lock. Succeeds if the 290167787Sjhb * reader already owns a read lock and the lock is locked for read to 291157846Sjhb * prevent deadlock from reader recursion. Also succeeds if the lock 292157846Sjhb * is unlocked and has no writer waiters or spinners. Failing otherwise 293157846Sjhb * prioritizes writers before readers. 294157846Sjhb */ 295157846Sjhb#define RW_CAN_READ(_rw) \ 296157846Sjhb ((curthread->td_rw_rlocks && (_rw) & RW_LOCK_READ) || ((_rw) & \ 297157846Sjhb (RW_LOCK_READ | RW_LOCK_WRITE_WAITERS | RW_LOCK_WRITE_SPINNER)) == \ 298157846Sjhb RW_LOCK_READ) 299157846Sjhb 300157846Sjhbvoid 301154941Sjhb_rw_rlock(struct rwlock *rw, const char *file, int line) 302154941Sjhb{ 303154941Sjhb struct turnstile *ts; 304167787Sjhb#ifdef ADAPTIVE_RWLOCKS 305154941Sjhb volatile struct thread *owner; 306154941Sjhb int spintries = 0; 307167787Sjhb int i; 308167787Sjhb#endif 309154941Sjhb#ifdef LOCK_PROFILING 310154941Sjhb uint64_t waittime = 0; 311154941Sjhb int contested = 0; 312154941Sjhb#endif 313154941Sjhb uintptr_t v; 314154941Sjhb#ifdef KDTRACE_HOOKS 315154941Sjhb uint64_t spin_cnt = 0; 316154941Sjhb uint64_t sleep_cnt = 0; 317154941Sjhb int64_t sleep_time = 0; 318154941Sjhb#endif 319167787Sjhb 320167787Sjhb KASSERT(rw->rw_lock != RW_DESTROYED, 321160771Sjhb ("rw_rlock() of destroyed rwlock @ %s:%d", file, line)); 322154941Sjhb KASSERT(rw_wowner(rw) != curthread, 323154941Sjhb ("%s (%s): wlock already held @ %s:%d", __func__, 324154941Sjhb rw->lock_object.lo_name, file, line)); 325154941Sjhb WITNESS_CHECKORDER(&rw->lock_object, LOP_NEWORDER, file, line, NULL); 326154941Sjhb 327154941Sjhb for (;;) { 328154941Sjhb#ifdef KDTRACE_HOOKS 329154941Sjhb spin_cnt++; 330154941Sjhb#endif 331160771Sjhb /* 332167787Sjhb * Handle the easy case. If no other thread has a write 333167787Sjhb * lock, then try to bump up the count of read locks. Note 334154941Sjhb * that we have to preserve the current state of the 335154941Sjhb * RW_LOCK_WRITE_WAITERS flag. If we fail to acquire a 336154941Sjhb * read lock, then rw_lock must have changed, so restart 337154941Sjhb * the loop. Note that this handles the case of a 338154941Sjhb * completely unlocked rwlock since such a lock is encoded 339154941Sjhb * as a read lock with no waiters. 340154941Sjhb */ 341154941Sjhb v = rw->rw_lock; 342154941Sjhb if (RW_CAN_READ(v)) { 343154941Sjhb /* 344154941Sjhb * The RW_LOCK_READ_WAITERS flag should only be set 345154941Sjhb * if the lock has been unlocked and write waiters 346167787Sjhb * were present. 347154941Sjhb */ 348154941Sjhb if (atomic_cmpset_acq_ptr(&rw->rw_lock, v, 349154941Sjhb v + RW_ONE_READER)) { 350154941Sjhb if (LOCK_LOG_TEST(&rw->lock_object, 0)) 351154941Sjhb CTR4(KTR_LOCK, 352154941Sjhb "%s: %p succeed %p -> %p", __func__, 353154941Sjhb rw, (void *)v, 354167307Sjhb (void *)(v + RW_ONE_READER)); 355154941Sjhb break; 356164159Skmacy } 357154941Sjhb continue; 358154941Sjhb } 359154941Sjhb lock_profile_obtain_lock_failed(&rw->lock_object, 360154941Sjhb &contested, &waittime); 361154941Sjhb 362154941Sjhb#ifdef ADAPTIVE_RWLOCKS 363154941Sjhb /* 364154941Sjhb * If the owner is running on another CPU, spin until 365154941Sjhb * the owner stops running or the state of the lock 366154941Sjhb * changes. 367154941Sjhb */ 368154941Sjhb if ((v & RW_LOCK_READ) == 0) { 369154941Sjhb owner = (struct thread *)RW_OWNER(v); 370154941Sjhb if (TD_IS_RUNNING(owner)) { 371154941Sjhb if (LOCK_LOG_TEST(&rw->lock_object, 0)) 372154941Sjhb CTR3(KTR_LOCK, 373154941Sjhb "%s: spinning on %p held by %p", 374154941Sjhb __func__, rw, owner); 375154941Sjhb while ((struct thread*)RW_OWNER(rw->rw_lock) == 376154941Sjhb owner && TD_IS_RUNNING(owner)) { 377154941Sjhb cpu_spinwait(); 378154941Sjhb#ifdef KDTRACE_HOOKS 379154941Sjhb spin_cnt++; 380167787Sjhb#endif 381154941Sjhb } 382154941Sjhb continue; 383154941Sjhb } 384154941Sjhb } else if (spintries < rowner_retries) { 385154941Sjhb spintries++; 386154941Sjhb for (i = 0; i < rowner_loops; i++) { 387154941Sjhb v = rw->rw_lock; 388154941Sjhb if ((v & RW_LOCK_READ) == 0 || RW_CAN_READ(v)) 389154941Sjhb break; 390154941Sjhb cpu_spinwait(); 391154941Sjhb } 392154941Sjhb if (i != rowner_loops) 393154941Sjhb continue; 394154941Sjhb } 395154941Sjhb#endif 396154941Sjhb 397154941Sjhb /* 398167787Sjhb * Okay, now it's the hard case. Some other thread already 399154941Sjhb * has a write lock or there are write waiters present, 400154941Sjhb * acquire the turnstile lock so we can begin the process 401154941Sjhb * of blocking. 402154941Sjhb */ 403154941Sjhb ts = turnstile_trywait(&rw->lock_object); 404154941Sjhb 405154941Sjhb /* 406154941Sjhb * The lock might have been released while we spun, so 407154941Sjhb * recheck its state and restart the loop if needed. 408154941Sjhb */ 409154941Sjhb v = rw->rw_lock; 410154941Sjhb if (RW_CAN_READ(v)) { 411154941Sjhb turnstile_cancel(ts); 412154941Sjhb continue; 413154941Sjhb } 414154941Sjhb 415154941Sjhb#ifdef ADAPTIVE_RWLOCKS 416154941Sjhb /* 417154941Sjhb * The current lock owner might have started executing 418167787Sjhb * on another CPU (or the lock could have changed 419154941Sjhb * owners) while we were waiting on the turnstile 420154941Sjhb * chain lock. If so, drop the turnstile lock and try 421167787Sjhb * again. 422154941Sjhb */ 423154941Sjhb if ((v & RW_LOCK_READ) == 0) { 424154941Sjhb owner = (struct thread *)RW_OWNER(v); 425154941Sjhb if (TD_IS_RUNNING(owner)) { 426154941Sjhb turnstile_cancel(ts); 427154941Sjhb continue; 428154941Sjhb } 429154941Sjhb } 430154941Sjhb#endif 431154941Sjhb 432167787Sjhb /* 433157846Sjhb * The lock is held in write mode or it already has waiters. 434154941Sjhb */ 435154941Sjhb MPASS(!RW_CAN_READ(v)); 436154941Sjhb 437154941Sjhb /* 438167787Sjhb * If the RW_LOCK_READ_WAITERS flag is already set, then 439154941Sjhb * we can go ahead and block. If it is not set then try 440154941Sjhb * to set it. If we fail to set it drop the turnstile 441154941Sjhb * lock and restart the loop. 442154941Sjhb */ 443154941Sjhb if (!(v & RW_LOCK_READ_WAITERS)) { 444154941Sjhb if (!atomic_cmpset_ptr(&rw->rw_lock, v, 445154941Sjhb v | RW_LOCK_READ_WAITERS)) { 446154941Sjhb turnstile_cancel(ts); 447154941Sjhb continue; 448154941Sjhb } 449157851Swkoszek if (LOCK_LOG_TEST(&rw->lock_object, 0)) 450157846Sjhb CTR2(KTR_LOCK, "%s: %p set read waiters flag", 451157851Swkoszek __func__, rw); 452154941Sjhb } 453154941Sjhb 454167787Sjhb /* 455154941Sjhb * We were unable to acquire the lock and the read waiters 456167787Sjhb * flag is set, so we must block on the turnstile. 457154941Sjhb */ 458154941Sjhb if (LOCK_LOG_TEST(&rw->lock_object, 0)) 459167787Sjhb CTR2(KTR_LOCK, "%s: %p blocking on turnstile", __func__, 460154941Sjhb rw); 461154941Sjhb#ifdef KDTRACE_HOOKS 462154941Sjhb sleep_time -= lockstat_nsecs(); 463154941Sjhb#endif 464154941Sjhb turnstile_wait(ts, rw_owner(rw), TS_SHARED_QUEUE); 465154941Sjhb#ifdef KDTRACE_HOOKS 466154941Sjhb sleep_time += lockstat_nsecs(); 467167787Sjhb sleep_cnt++; 468154941Sjhb#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. 478154941Sjhb */ 479154941Sjhb LOCKSTAT_PROFILE_OBTAIN_LOCK_SUCCESS(LS_RW_RLOCK_ACQUIRE, rw, contested, 480154941Sjhb waittime, file, line); 481154941Sjhb LOCK_LOG_LOCK("RLOCK", &rw->lock_object, 0, 0, file, line); 482154941Sjhb WITNESS_LOCK(&rw->lock_object, 0, file, line); 483154941Sjhb curthread->td_locks++; 484154941Sjhb curthread->td_rw_rlocks++; 485154941Sjhb#ifdef KDTRACE_HOOKS 486167787Sjhb if (sleep_time) 487154941Sjhb LOCKSTAT_RECORD1(LS_RW_RLOCK_BLOCK, rw, sleep_time); 488154941Sjhb 489154941Sjhb /* 490154941Sjhb * Record only the loops spinning and not sleeping. 491167787Sjhb */ 492154941Sjhb if (spin_cnt > sleep_cnt) 493154941Sjhb LOCKSTAT_RECORD1(LS_RW_RLOCK_SPIN, rw, (spin_cnt - sleep_cnt)); 494154941Sjhb#endif 495154941Sjhb} 496154941Sjhb 497154941Sjhbint 498154941Sjhb_rw_try_rlock(struct rwlock *rw, const char *file, int line) 499154941Sjhb{ 500154941Sjhb uintptr_t x; 501157826Sjhb 502157826Sjhb for (;;) { 503157826Sjhb x = rw->rw_lock; 504167787Sjhb KASSERT(rw->rw_lock != RW_DESTROYED, 505157826Sjhb ("rw_try_rlock() of destroyed rwlock @ %s:%d", file, line)); 506157826Sjhb if (!(x & RW_LOCK_READ)) 507157826Sjhb break; 508167787Sjhb if (atomic_cmpset_acq_ptr(&rw->rw_lock, x, x + RW_ONE_READER)) { 509157826Sjhb LOCK_LOG_TRY("RLOCK", &rw->lock_object, 0, 1, file, 510157826Sjhb line); 511154941Sjhb WITNESS_LOCK(&rw->lock_object, LOP_TRYLOCK, file, line); 512154941Sjhb curthread->td_locks++; 513157846Sjhb curthread->td_rw_rlocks++; 514157846Sjhb return (1); 515157846Sjhb } 516157846Sjhb } 517157846Sjhb 518157846Sjhb LOCK_LOG_TRY("RLOCK", &rw->lock_object, 0, 0, file, line); 519157846Sjhb return (0); 520157846Sjhb} 521167787Sjhb 522167787Sjhbvoid 523157846Sjhb_rw_runlock(struct rwlock *rw, const char *file, int line) 524157846Sjhb{ 525157846Sjhb struct turnstile *ts; 526157846Sjhb uintptr_t x, v, queue; 527157846Sjhb 528157846Sjhb KASSERT(rw->rw_lock != RW_DESTROYED, 529157846Sjhb ("rw_runlock() of destroyed rwlock @ %s:%d", file, line)); 530157846Sjhb _rw_assert(rw, RA_RLOCKED, file, line); 531154941Sjhb curthread->td_locks--; 532154941Sjhb curthread->td_rw_rlocks--; 533154941Sjhb WITNESS_UNLOCK(&rw->lock_object, 0, file, line); 534154941Sjhb LOCK_LOG_LOCK("RUNLOCK", &rw->lock_object, 0, 0, file, line); 535154941Sjhb 536167787Sjhb /* TODO: drop "owner of record" here. */ 537154941Sjhb 538154941Sjhb for (;;) { 539167787Sjhb /* 540154941Sjhb * See if there is more than one read lock held. If so, 541167787Sjhb * just drop one and return. 542154941Sjhb */ 543154941Sjhb x = rw->rw_lock; 544154941Sjhb if (RW_READERS(x) > 1) { 545154941Sjhb if (atomic_cmpset_rel_ptr(&rw->rw_lock, x, 546154941Sjhb 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 } 554154941Sjhb continue; 555154941Sjhb } 556154941Sjhb /* 557154941Sjhb * If there aren't any waiters for a write lock, then try 558154941Sjhb * to drop it quickly. 559154941Sjhb */ 560154941Sjhb if (!(x & RW_LOCK_WAITERS)) { 561154941Sjhb MPASS((x & ~RW_LOCK_WRITE_SPINNER) == 562167787Sjhb RW_READERS_LOCK(1)); 563154941Sjhb if (atomic_cmpset_rel_ptr(&rw->rw_lock, x, 564154941Sjhb RW_UNLOCKED)) { 565167787Sjhb if (LOCK_LOG_TEST(&rw->lock_object, 0)) 566167787Sjhb CTR2(KTR_LOCK, "%s: %p last succeeded", 567154941Sjhb __func__, rw); 568157846Sjhb break; 569157846Sjhb } 570157846Sjhb continue; 571157846Sjhb } 572157846Sjhb /* 573157846Sjhb * Ok, we know we have waiters and we think we are the 574157846Sjhb * last reader, so grab the turnstile lock. 575157846Sjhb */ 576167787Sjhb turnstile_chain_lock(&rw->lock_object); 577157846Sjhb v = rw->rw_lock & (RW_LOCK_WAITERS | RW_LOCK_WRITE_SPINNER); 578167787Sjhb MPASS(v & RW_LOCK_WAITERS); 579157846Sjhb 580157846Sjhb /* 581157846Sjhb * Try to drop our lock leaving the lock in a unlocked 582154941Sjhb * state. 583157846Sjhb * 584154941Sjhb * If you wanted to do explicit lock handoff you'd have to 585154941Sjhb * do it here. You'd also want to use turnstile_signal() 586154941Sjhb * and you'd have to handle the race where a higher 587154941Sjhb * priority thread blocks on the write lock before the 588154941Sjhb * thread you wakeup actually runs and have the new thread 589154941Sjhb * "steal" the lock. For now it's a lot simpler to just 590154941Sjhb * wakeup all of the waiters. 591154941Sjhb * 592154941Sjhb * As above, if we fail, then another thread might have 593154941Sjhb * acquired a read lock, so drop the turnstile lock and 594154941Sjhb * restart. 595154941Sjhb */ 596154941Sjhb x = RW_UNLOCKED; 597154941Sjhb if (v & RW_LOCK_WRITE_WAITERS) { 598154941Sjhb queue = TS_EXCLUSIVE_QUEUE; 599154941Sjhb x |= (v & RW_LOCK_READ_WAITERS); 600157846Sjhb } else 601157846Sjhb queue = TS_SHARED_QUEUE; 602157846Sjhb if (!atomic_cmpset_rel_ptr(&rw->rw_lock, RW_READERS_LOCK(1) | v, 603157846Sjhb x)) { 604157846Sjhb turnstile_chain_unlock(&rw->lock_object); 605157846Sjhb continue; 606154941Sjhb } 607157846Sjhb if (LOCK_LOG_TEST(&rw->lock_object, 0)) 608154941Sjhb CTR2(KTR_LOCK, "%s: %p last succeeded with waiters", 609154941Sjhb __func__, rw); 610157846Sjhb 611157846Sjhb /* 612157846Sjhb * Ok. The lock is released and all that's left is to 613157846Sjhb * wake up the waiters. Note that the lock might not be 614157846Sjhb * free anymore, but in that case the writers will just 615157846Sjhb * block again if they run before the new lock holder(s) 616157846Sjhb * release the lock. 617157846Sjhb */ 618154941Sjhb ts = turnstile_lookup(&rw->lock_object); 619157846Sjhb MPASS(ts != NULL); 620157846Sjhb turnstile_broadcast(ts, queue); 621157846Sjhb turnstile_unpend(ts, TS_SHARED_LOCK); 622157846Sjhb turnstile_chain_unlock(&rw->lock_object); 623157846Sjhb break; 624157846Sjhb } 625157846Sjhb LOCKSTAT_PROFILE_RELEASE_LOCK(LS_RW_RUNLOCK_RELEASE, rw); 626157846Sjhb} 627167787Sjhb 628157846Sjhb/* 629157846Sjhb * This function is called when we are unable to obtain a write lock on the 630157846Sjhb * first try. This means that at least one other thread holds either a 631157846Sjhb * read or write lock. 632154941Sjhb */ 633157846Sjhbvoid 634157846Sjhb_rw_wlock_hard(struct rwlock *rw, uintptr_t tid, const char *file, int line) 635157846Sjhb{ 636167787Sjhb struct turnstile *ts; 637154941Sjhb#ifdef ADAPTIVE_RWLOCKS 638154941Sjhb volatile struct thread *owner; 639154941Sjhb int spintries = 0; 640154941Sjhb int i; 641154941Sjhb#endif 642154941Sjhb uintptr_t v, x; 643154941Sjhb#ifdef LOCK_PROFILING 644157882Sjhb uint64_t waittime = 0; 645157882Sjhb int contested = 0; 646157882Sjhb#endif 647157882Sjhb#ifdef KDTRACE_HOOKS 648157882Sjhb uint64_t spin_cnt = 0; 649157882Sjhb uint64_t sleep_cnt = 0; 650157882Sjhb int64_t sleep_time = 0; 651157882Sjhb#endif 652157882Sjhb 653157882Sjhb if (rw_wlocked(rw)) { 654157882Sjhb KASSERT(rw->lock_object.lo_flags & LO_RECURSABLE, 655157882Sjhb ("%s: recursing but non-recursive rw %s @ %s:%d\n", 656157882Sjhb __func__, rw->lock_object.lo_name, file, line)); 657157882Sjhb rw->rw_recurse++; 658157882Sjhb if (LOCK_LOG_TEST(&rw->lock_object, 0)) 659157882Sjhb CTR2(KTR_LOCK, "%s: %p recursing", __func__, rw); 660157882Sjhb return; 661157882Sjhb } 662157882Sjhb 663157882Sjhb if (LOCK_LOG_TEST(&rw->lock_object, 0)) 664157882Sjhb CTR5(KTR_LOCK, "%s: %s contested (lock=%p) at %s:%d", __func__, 665157882Sjhb rw->lock_object.lo_name, (void *)rw->rw_lock, file, line); 666157882Sjhb 667157882Sjhb while (!_rw_write_lock(rw, tid)) { 668157882Sjhb#ifdef KDTRACE_HOOKS 669157882Sjhb spin_cnt++; 670157882Sjhb#endif 671157882Sjhb lock_profile_obtain_lock_failed(&rw->lock_object, 672157882Sjhb &contested, &waittime); 673157882Sjhb#ifdef ADAPTIVE_RWLOCKS 674157882Sjhb /* 675167787Sjhb * If the lock is write locked and the owner is 676157882Sjhb * running on another CPU, spin until the owner stops 677157882Sjhb * running or the state of the lock changes. 678157882Sjhb */ 679157882Sjhb v = rw->rw_lock; 680157882Sjhb owner = (struct thread *)RW_OWNER(v); 681157882Sjhb if (!(v & RW_LOCK_READ) && TD_IS_RUNNING(owner)) { 682157882Sjhb if (LOCK_LOG_TEST(&rw->lock_object, 0)) 683157882Sjhb CTR3(KTR_LOCK, "%s: spinning on %p held by %p", 684157882Sjhb __func__, rw, owner); 685157882Sjhb while ((struct thread*)RW_OWNER(rw->rw_lock) == owner && 686157882Sjhb TD_IS_RUNNING(owner)) { 687157882Sjhb cpu_spinwait(); 688157882Sjhb#ifdef KDTRACE_HOOKS 689167787Sjhb spin_cnt++; 690157882Sjhb#endif 691157882Sjhb } 692157882Sjhb continue; 693167787Sjhb } 694157882Sjhb if ((v & RW_LOCK_READ) && RW_READERS(v) && 695167787Sjhb spintries < rowner_retries) { 696157882Sjhb if (!(v & RW_LOCK_WRITE_SPINNER)) { 697167787Sjhb if (!atomic_cmpset_ptr(&rw->rw_lock, v, 698157882Sjhb v | RW_LOCK_WRITE_SPINNER)) { 699167787Sjhb continue; 700157882Sjhb } 701157882Sjhb } 702157882Sjhb spintries++; 703157882Sjhb for (i = 0; i < rowner_loops; i++) { 704157882Sjhb if ((rw->rw_lock & RW_LOCK_WRITE_SPINNER) == 0) 705157882Sjhb break; 706157882Sjhb cpu_spinwait(); 707157882Sjhb } 708157882Sjhb#ifdef KDTRACE_HOOKS 709157882Sjhb spin_cnt += rowner_loops - i; 710157882Sjhb#endif 711157882Sjhb if (i != rowner_loops) 712157882Sjhb continue; 713157882Sjhb } 714157882Sjhb#endif 715167787Sjhb ts = turnstile_trywait(&rw->lock_object); 716157882Sjhb v = rw->rw_lock; 717157882Sjhb 718157882Sjhb#ifdef ADAPTIVE_RWLOCKS 719157882Sjhb /* 720157882Sjhb * The current lock owner might have started executing 721157882Sjhb * on another CPU (or the lock could have changed 722157882Sjhb * owners) while we were waiting on the turnstile 723157882Sjhb * chain lock. If so, drop the turnstile lock and try 724157882Sjhb * again. 725157882Sjhb */ 726157882Sjhb if (!(v & RW_LOCK_READ)) { 727157882Sjhb owner = (struct thread *)RW_OWNER(v); 728157882Sjhb if (TD_IS_RUNNING(owner)) { 729157882Sjhb turnstile_cancel(ts); 730157882Sjhb continue; 731167787Sjhb } 732157882Sjhb } 733157882Sjhb#endif 734157882Sjhb /* 735157882Sjhb * Check for the waiters flags about this rwlock. 736157882Sjhb * If the lock was released, without maintain any pending 737157882Sjhb * waiters queue, simply try to acquire it. 738157882Sjhb * If a pending waiters queue is present, claim the lock 739157882Sjhb * ownership and maintain the pending queue. 740157882Sjhb */ 741157882Sjhb x = v & (RW_LOCK_WAITERS | RW_LOCK_WRITE_SPINNER); 742157882Sjhb if ((v & ~x) == RW_UNLOCKED) { 743157882Sjhb x &= ~RW_LOCK_WRITE_SPINNER; 744157882Sjhb if (atomic_cmpset_acq_ptr(&rw->rw_lock, v, tid | x)) { 745157882Sjhb if (x) 746167787Sjhb turnstile_claim(ts); 747157882Sjhb else 748157882Sjhb turnstile_cancel(ts); 749157882Sjhb break; 750157882Sjhb } 751157882Sjhb turnstile_cancel(ts); 752157882Sjhb continue; 753157882Sjhb } 754157882Sjhb /* 755157882Sjhb * If the RW_LOCK_WRITE_WAITERS flag isn't set, then try to 756157882Sjhb * set it. If we fail to set it, then loop back and try 757157882Sjhb * again. 758157882Sjhb */ 759157882Sjhb if (!(v & RW_LOCK_WRITE_WAITERS)) { 760157882Sjhb if (!atomic_cmpset_ptr(&rw->rw_lock, v, 761157882Sjhb v | RW_LOCK_WRITE_WAITERS)) { 762157882Sjhb turnstile_cancel(ts); 763157882Sjhb continue; 764157882Sjhb } 765157882Sjhb if (LOCK_LOG_TEST(&rw->lock_object, 0)) 766157882Sjhb CTR2(KTR_LOCK, "%s: %p set write waiters flag", 767167787Sjhb __func__, rw); 768157882Sjhb } 769157882Sjhb /* 770157882Sjhb * We were unable to acquire the lock and the write waiters 771157882Sjhb * flag is set, so we must block on the turnstile. 772167787Sjhb */ 773157882Sjhb if (LOCK_LOG_TEST(&rw->lock_object, 0)) 774157882Sjhb CTR2(KTR_LOCK, "%s: %p blocking on turnstile", __func__, 775154941Sjhb rw); 776155162Sscottl#ifdef KDTRACE_HOOKS 777154941Sjhb sleep_time -= lockstat_nsecs(); 778154941Sjhb#endif 779154941Sjhb turnstile_wait(ts, rw_owner(rw), TS_EXCLUSIVE_QUEUE); 780154941Sjhb#ifdef KDTRACE_HOOKS 781154941Sjhb sleep_time += lockstat_nsecs(); 782154941Sjhb sleep_cnt++; 783154941Sjhb#endif 784154941Sjhb if (LOCK_LOG_TEST(&rw->lock_object, 0)) 785154941Sjhb CTR2(KTR_LOCK, "%s: %p resuming from turnstile", 786154941Sjhb __func__, rw); 787154941Sjhb#ifdef ADAPTIVE_RWLOCKS 788154941Sjhb spintries = 0; 789154941Sjhb#endif 790154941Sjhb } 791154941Sjhb LOCKSTAT_PROFILE_OBTAIN_LOCK_SUCCESS(LS_RW_WLOCK_ACQUIRE, rw, contested, 792154941Sjhb waittime, file, line); 793167368Sjhb#ifdef KDTRACE_HOOKS 794154941Sjhb if (sleep_time) 795154941Sjhb LOCKSTAT_RECORD1(LS_RW_WLOCK_BLOCK, rw, sleep_time); 796167787Sjhb 797154941Sjhb /* 798154941Sjhb * Record only the loops spinning and not sleeping. 799154941Sjhb */ 800154941Sjhb if (spin_cnt > sleep_cnt) 801154941Sjhb LOCKSTAT_RECORD1(LS_RW_WLOCK_SPIN, rw, (spin_cnt - sleep_cnt)); 802154941Sjhb#endif 803155061Sscottl} 804155061Sscottl 805157826Sjhb/* 806154941Sjhb * This function is called if the first try at releasing a write lock failed. 807167787Sjhb * This means that one of the 2 waiter bits must be set indicating that at 808154941Sjhb * least one thread is waiting on this lock. 809154941Sjhb */ 810154941Sjhbvoid 811154941Sjhb_rw_wunlock_hard(struct rwlock *rw, uintptr_t tid, const char *file, int line) 812157826Sjhb{ 813154941Sjhb struct turnstile *ts; 814167787Sjhb uintptr_t v; 815154941Sjhb int queue; 816154941Sjhb 817154941Sjhb if (rw_wlocked(rw) && rw_recursed(rw)) { 818167787Sjhb rw->rw_recurse--; 819154941Sjhb if (LOCK_LOG_TEST(&rw->lock_object, 0)) 820154941Sjhb CTR2(KTR_LOCK, "%s: %p unrecursing", __func__, rw); 821154941Sjhb return; 822154941Sjhb } 823154941Sjhb 824157826Sjhb KASSERT(rw->rw_lock & (RW_LOCK_READ_WAITERS | RW_LOCK_WRITE_WAITERS), 825154941Sjhb ("%s: neither of the waiter flags are set", __func__)); 826167787Sjhb 827154941Sjhb if (LOCK_LOG_TEST(&rw->lock_object, 0)) 828154941Sjhb CTR2(KTR_LOCK, "%s: %p contested", __func__, rw); 829154941Sjhb 830154941Sjhb turnstile_chain_lock(&rw->lock_object); 831154941Sjhb ts = turnstile_lookup(&rw->lock_object); 832154941Sjhb MPASS(ts != NULL); 833154941Sjhb 834154941Sjhb /* 835154941Sjhb * Use the same algo as sx locks for now. Prefer waking up shared 836154941Sjhb * waiters if we have any over writers. This is probably not ideal. 837154941Sjhb * 838154941Sjhb * 'v' is the value we are going to write back to rw_lock. If we 839154941Sjhb * have waiters on both queues, we need to preserve the state of 840154941Sjhb * the waiter flag for the queue we don't wake up. For now this is 841154941Sjhb * hardcoded for the algorithm mentioned above. 842154941Sjhb * 843154941Sjhb * In the case of both readers and writers waiting we wakeup the 844154941Sjhb * readers but leave the RW_LOCK_WRITE_WAITERS flag set. If a 845154941Sjhb * new writer comes in before a reader it will claim the lock up 846154941Sjhb * above. There is probably a potential priority inversion in 847154941Sjhb * there that could be worked around either by waking both queues 848154941Sjhb * of waiters or doing some complicated lock handoff gymnastics. 849167504Sjhb */ 850167504Sjhb v = RW_UNLOCKED; 851154941Sjhb if (rw->rw_lock & RW_LOCK_WRITE_WAITERS) { 852157826Sjhb queue = TS_EXCLUSIVE_QUEUE; 853154941Sjhb v |= (rw->rw_lock & RW_LOCK_READ_WAITERS); 854154941Sjhb } else 855154941Sjhb queue = TS_SHARED_QUEUE; 856154941Sjhb 857154941Sjhb /* Wake up all waiters for the specific queue. */ 858154941Sjhb if (LOCK_LOG_TEST(&rw->lock_object, 0)) 859154941Sjhb CTR3(KTR_LOCK, "%s: %p waking up %s waiters", __func__, rw, 860154941Sjhb queue == TS_SHARED_QUEUE ? "read" : "write"); 861154941Sjhb turnstile_broadcast(ts, queue); 862154941Sjhb atomic_store_rel_ptr(&rw->rw_lock, v); 863154941Sjhb turnstile_unpend(ts, TS_EXCLUSIVE_LOCK); 864154941Sjhb turnstile_chain_unlock(&rw->lock_object); 865167492Sjhb} 866154941Sjhb 867154941Sjhb/* 868154941Sjhb * Attempt to do a non-blocking upgrade from a read lock to a write 869154941Sjhb * lock. This will only succeed if this thread holds a single read 870154941Sjhb * lock. Returns true if the upgrade succeeded and false otherwise. 871154941Sjhb */ 872154941Sjhbint 873154941Sjhb_rw_try_upgrade(struct rwlock *rw, const char *file, int line) 874{ 875 uintptr_t v, x, tid; 876 struct turnstile *ts; 877 int success; 878 879 KASSERT(rw->rw_lock != RW_DESTROYED, 880 ("rw_try_upgrade() of destroyed rwlock @ %s:%d", file, line)); 881 _rw_assert(rw, RA_RLOCKED, file, line); 882 883 /* 884 * Attempt to switch from one reader to a writer. If there 885 * are any write waiters, then we will have to lock the 886 * turnstile first to prevent races with another writer 887 * calling turnstile_wait() before we have claimed this 888 * turnstile. So, do the simple case of no waiters first. 889 */ 890 tid = (uintptr_t)curthread; 891 success = 0; 892 for (;;) { 893 v = rw->rw_lock; 894 if (RW_READERS(v) > 1) 895 break; 896 if (!(v & RW_LOCK_WAITERS)) { 897 success = atomic_cmpset_ptr(&rw->rw_lock, v, tid); 898 if (!success) 899 continue; 900 break; 901 } 902 903 /* 904 * Ok, we think we have waiters, so lock the turnstile. 905 */ 906 ts = turnstile_trywait(&rw->lock_object); 907 v = rw->rw_lock; 908 if (RW_READERS(v) > 1) { 909 turnstile_cancel(ts); 910 break; 911 } 912 /* 913 * Try to switch from one reader to a writer again. This time 914 * we honor the current state of the waiters flags. 915 * If we obtain the lock with the flags set, then claim 916 * ownership of the turnstile. 917 */ 918 x = rw->rw_lock & RW_LOCK_WAITERS; 919 success = atomic_cmpset_ptr(&rw->rw_lock, v, tid | x); 920 if (success) { 921 if (x) 922 turnstile_claim(ts); 923 else 924 turnstile_cancel(ts); 925 break; 926 } 927 turnstile_cancel(ts); 928 } 929 LOCK_LOG_TRY("WUPGRADE", &rw->lock_object, 0, success, file, line); 930 if (success) { 931 curthread->td_rw_rlocks--; 932 WITNESS_UPGRADE(&rw->lock_object, LOP_EXCLUSIVE | LOP_TRYLOCK, 933 file, line); 934 LOCKSTAT_RECORD0(LS_RW_TRYUPGRADE_UPGRADE, rw); 935 } 936 return (success); 937} 938 939/* 940 * Downgrade a write lock into a single read lock. 941 */ 942void 943_rw_downgrade(struct rwlock *rw, const char *file, int line) 944{ 945 struct turnstile *ts; 946 uintptr_t tid, v; 947 int rwait, wwait; 948 949 KASSERT(rw->rw_lock != RW_DESTROYED, 950 ("rw_downgrade() of destroyed rwlock @ %s:%d", file, line)); 951 _rw_assert(rw, RA_WLOCKED | RA_NOTRECURSED, file, line); 952#ifndef INVARIANTS 953 if (rw_recursed(rw)) 954 panic("downgrade of a recursed lock"); 955#endif 956 957 WITNESS_DOWNGRADE(&rw->lock_object, 0, file, line); 958 959 /* 960 * Convert from a writer to a single reader. First we handle 961 * the easy case with no waiters. If there are any waiters, we 962 * lock the turnstile and "disown" the lock. 963 */ 964 tid = (uintptr_t)curthread; 965 if (atomic_cmpset_rel_ptr(&rw->rw_lock, tid, RW_READERS_LOCK(1))) 966 goto out; 967 968 /* 969 * Ok, we think we have waiters, so lock the turnstile so we can 970 * read the waiter flags without any races. 971 */ 972 turnstile_chain_lock(&rw->lock_object); 973 v = rw->rw_lock & RW_LOCK_WAITERS; 974 rwait = v & RW_LOCK_READ_WAITERS; 975 wwait = v & RW_LOCK_WRITE_WAITERS; 976 MPASS(rwait | wwait); 977 978 /* 979 * Downgrade from a write lock while preserving waiters flag 980 * and give up ownership of the turnstile. 981 */ 982 ts = turnstile_lookup(&rw->lock_object); 983 MPASS(ts != NULL); 984 if (!wwait) 985 v &= ~RW_LOCK_READ_WAITERS; 986 atomic_store_rel_ptr(&rw->rw_lock, RW_READERS_LOCK(1) | v); 987 /* 988 * Wake other readers if there are no writers pending. Otherwise they 989 * won't be able to acquire the lock anyway. 990 */ 991 if (rwait && !wwait) { 992 turnstile_broadcast(ts, TS_SHARED_QUEUE); 993 turnstile_unpend(ts, TS_EXCLUSIVE_LOCK); 994 } else 995 turnstile_disown(ts); 996 turnstile_chain_unlock(&rw->lock_object); 997out: 998 curthread->td_rw_rlocks++; 999 LOCK_LOG_LOCK("WDOWNGRADE", &rw->lock_object, 0, 0, file, line); 1000 LOCKSTAT_RECORD0(LS_RW_DOWNGRADE_DOWNGRADE, rw); 1001} 1002 1003#ifdef INVARIANT_SUPPORT 1004#ifndef INVARIANTS 1005#undef _rw_assert 1006#endif 1007 1008/* 1009 * In the non-WITNESS case, rw_assert() can only detect that at least 1010 * *some* thread owns an rlock, but it cannot guarantee that *this* 1011 * thread owns an rlock. 1012 */ 1013void 1014_rw_assert(const struct rwlock *rw, int what, const char *file, int line) 1015{ 1016 1017 if (panicstr != NULL) 1018 return; 1019 switch (what) { 1020 case RA_LOCKED: 1021 case RA_LOCKED | RA_RECURSED: 1022 case RA_LOCKED | RA_NOTRECURSED: 1023 case RA_RLOCKED: 1024#ifdef WITNESS 1025 witness_assert(&rw->lock_object, what, file, line); 1026#else 1027 /* 1028 * If some other thread has a write lock or we have one 1029 * and are asserting a read lock, fail. Also, if no one 1030 * has a lock at all, fail. 1031 */ 1032 if (rw->rw_lock == RW_UNLOCKED || 1033 (!(rw->rw_lock & RW_LOCK_READ) && (what == RA_RLOCKED || 1034 rw_wowner(rw) != curthread))) 1035 panic("Lock %s not %slocked @ %s:%d\n", 1036 rw->lock_object.lo_name, (what == RA_RLOCKED) ? 1037 "read " : "", file, line); 1038 1039 if (!(rw->rw_lock & RW_LOCK_READ)) { 1040 if (rw_recursed(rw)) { 1041 if (what & RA_NOTRECURSED) 1042 panic("Lock %s recursed @ %s:%d\n", 1043 rw->lock_object.lo_name, file, 1044 line); 1045 } else if (what & RA_RECURSED) 1046 panic("Lock %s not recursed @ %s:%d\n", 1047 rw->lock_object.lo_name, file, line); 1048 } 1049#endif 1050 break; 1051 case RA_WLOCKED: 1052 case RA_WLOCKED | RA_RECURSED: 1053 case RA_WLOCKED | RA_NOTRECURSED: 1054 if (rw_wowner(rw) != curthread) 1055 panic("Lock %s not exclusively locked @ %s:%d\n", 1056 rw->lock_object.lo_name, file, line); 1057 if (rw_recursed(rw)) { 1058 if (what & RA_NOTRECURSED) 1059 panic("Lock %s recursed @ %s:%d\n", 1060 rw->lock_object.lo_name, file, line); 1061 } else if (what & RA_RECURSED) 1062 panic("Lock %s not recursed @ %s:%d\n", 1063 rw->lock_object.lo_name, file, line); 1064 break; 1065 case RA_UNLOCKED: 1066#ifdef WITNESS 1067 witness_assert(&rw->lock_object, what, file, line); 1068#else 1069 /* 1070 * If we hold a write lock fail. We can't reliably check 1071 * to see if we hold a read lock or not. 1072 */ 1073 if (rw_wowner(rw) == curthread) 1074 panic("Lock %s exclusively locked @ %s:%d\n", 1075 rw->lock_object.lo_name, file, line); 1076#endif 1077 break; 1078 default: 1079 panic("Unknown rw lock assertion: %d @ %s:%d", what, file, 1080 line); 1081 } 1082} 1083#endif /* INVARIANT_SUPPORT */ 1084 1085#ifdef DDB 1086void 1087db_show_rwlock(const struct lock_object *lock) 1088{ 1089 const struct rwlock *rw; 1090 struct thread *td; 1091 1092 rw = (const struct rwlock *)lock; 1093 1094 db_printf(" state: "); 1095 if (rw->rw_lock == RW_UNLOCKED) 1096 db_printf("UNLOCKED\n"); 1097 else if (rw->rw_lock == RW_DESTROYED) { 1098 db_printf("DESTROYED\n"); 1099 return; 1100 } else if (rw->rw_lock & RW_LOCK_READ) 1101 db_printf("RLOCK: %ju locks\n", 1102 (uintmax_t)(RW_READERS(rw->rw_lock))); 1103 else { 1104 td = rw_wowner(rw); 1105 db_printf("WLOCK: %p (tid %d, pid %d, \"%s\")\n", td, 1106 td->td_tid, td->td_proc->p_pid, td->td_name); 1107 if (rw_recursed(rw)) 1108 db_printf(" recursed: %u\n", rw->rw_recurse); 1109 } 1110 db_printf(" waiters: "); 1111 switch (rw->rw_lock & (RW_LOCK_READ_WAITERS | RW_LOCK_WRITE_WAITERS)) { 1112 case RW_LOCK_READ_WAITERS: 1113 db_printf("readers\n"); 1114 break; 1115 case RW_LOCK_WRITE_WAITERS: 1116 db_printf("writers\n"); 1117 break; 1118 case RW_LOCK_READ_WAITERS | RW_LOCK_WRITE_WAITERS: 1119 db_printf("readers and writers\n"); 1120 break; 1121 default: 1122 db_printf("none\n"); 1123 break; 1124 } 1125} 1126 1127#endif 1128