kern_rwlock.c revision 244582
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 244582 2012-12-22 09:37:34Z attilio $"); 36154941Sjhb 37154941Sjhb#include "opt_ddb.h" 38233628Sfabient#include "opt_hwpmc_hooks.h" 39192853Ssson#include "opt_kdtrace.h" 40167801Sjhb#include "opt_no_adaptive_rwlocks.h" 41154941Sjhb 42154941Sjhb#include <sys/param.h> 43244582Sattilio#include <sys/kdb.h> 44154941Sjhb#include <sys/ktr.h> 45177912Sjeff#include <sys/kernel.h> 46154941Sjhb#include <sys/lock.h> 47154941Sjhb#include <sys/mutex.h> 48154941Sjhb#include <sys/proc.h> 49154941Sjhb#include <sys/rwlock.h> 50177912Sjeff#include <sys/sysctl.h> 51154941Sjhb#include <sys/systm.h> 52154941Sjhb#include <sys/turnstile.h> 53171516Sattilio 54154941Sjhb#include <machine/cpu.h> 55154941Sjhb 56167801Sjhb#if defined(SMP) && !defined(NO_ADAPTIVE_RWLOCKS) 57167801Sjhb#define ADAPTIVE_RWLOCKS 58167801Sjhb#endif 59167801Sjhb 60233628Sfabient#ifdef HWPMC_HOOKS 61233628Sfabient#include <sys/pmckern.h> 62233628SfabientPMC_SOFT_DECLARE( , , lock, failed); 63233628Sfabient#endif 64233628Sfabient 65242515Sattilio/* 66242515Sattilio * Return the rwlock address when the lock cookie address is provided. 67242515Sattilio * This functionality assumes that struct rwlock* have a member named rw_lock. 68242515Sattilio */ 69242515Sattilio#define rwlock2rw(c) (__containerof(c, struct rwlock, rw_lock)) 70242515Sattilio 71177912Sjeff#ifdef ADAPTIVE_RWLOCKS 72177912Sjeffstatic int rowner_retries = 10; 73177912Sjeffstatic int rowner_loops = 10000; 74227309Sedstatic SYSCTL_NODE(_debug, OID_AUTO, rwlock, CTLFLAG_RD, NULL, 75227309Sed "rwlock debugging"); 76177912SjeffSYSCTL_INT(_debug_rwlock, OID_AUTO, retry, CTLFLAG_RW, &rowner_retries, 0, ""); 77177912SjeffSYSCTL_INT(_debug_rwlock, OID_AUTO, loops, CTLFLAG_RW, &rowner_loops, 0, ""); 78177912Sjeff#endif 79177912Sjeff 80154941Sjhb#ifdef DDB 81154941Sjhb#include <ddb/ddb.h> 82154941Sjhb 83227588Spjdstatic void db_show_rwlock(const struct lock_object *lock); 84154941Sjhb#endif 85227588Spjdstatic void assert_rw(const struct lock_object *lock, int what); 86167368Sjhbstatic void lock_rw(struct lock_object *lock, int how); 87192853Ssson#ifdef KDTRACE_HOOKS 88227588Spjdstatic int owner_rw(const struct lock_object *lock, struct thread **owner); 89192853Ssson#endif 90167368Sjhbstatic int unlock_rw(struct lock_object *lock); 91154941Sjhb 92154941Sjhbstruct lock_class lock_class_rw = { 93167365Sjhb .lc_name = "rw", 94167365Sjhb .lc_flags = LC_SLEEPLOCK | LC_RECURSABLE | LC_UPGRADABLE, 95173733Sattilio .lc_assert = assert_rw, 96154941Sjhb#ifdef DDB 97167365Sjhb .lc_ddb_show = db_show_rwlock, 98154941Sjhb#endif 99167368Sjhb .lc_lock = lock_rw, 100167368Sjhb .lc_unlock = unlock_rw, 101192853Ssson#ifdef KDTRACE_HOOKS 102192853Ssson .lc_owner = owner_rw, 103192853Ssson#endif 104154941Sjhb}; 105154941Sjhb 106157826Sjhb/* 107157826Sjhb * Return a pointer to the owning thread if the lock is write-locked or 108157826Sjhb * NULL if the lock is unlocked or read-locked. 109157826Sjhb */ 110157826Sjhb#define rw_wowner(rw) \ 111154941Sjhb ((rw)->rw_lock & RW_LOCK_READ ? NULL : \ 112154941Sjhb (struct thread *)RW_OWNER((rw)->rw_lock)) 113154941Sjhb 114157826Sjhb/* 115171052Sattilio * Returns if a write owner is recursed. Write ownership is not assured 116171052Sattilio * here and should be previously checked. 117171052Sattilio */ 118171052Sattilio#define rw_recursed(rw) ((rw)->rw_recurse != 0) 119171052Sattilio 120171052Sattilio/* 121171052Sattilio * Return true if curthread helds the lock. 122171052Sattilio */ 123171052Sattilio#define rw_wlocked(rw) (rw_wowner((rw)) == curthread) 124171052Sattilio 125171052Sattilio/* 126157826Sjhb * Return a pointer to the owning thread for this lock who should receive 127157826Sjhb * any priority lent by threads that block on this lock. Currently this 128157826Sjhb * is identical to rw_wowner(). 129157826Sjhb */ 130157826Sjhb#define rw_owner(rw) rw_wowner(rw) 131157826Sjhb 132154941Sjhb#ifndef INVARIANTS 133242515Sattilio#define __rw_assert(c, what, file, line) 134154941Sjhb#endif 135154941Sjhb 136154941Sjhbvoid 137227588Spjdassert_rw(const struct lock_object *lock, int what) 138173733Sattilio{ 139173733Sattilio 140227588Spjd rw_assert((const struct rwlock *)lock, what); 141173733Sattilio} 142173733Sattilio 143173733Sattiliovoid 144167368Sjhblock_rw(struct lock_object *lock, int how) 145167368Sjhb{ 146167368Sjhb struct rwlock *rw; 147167368Sjhb 148167368Sjhb rw = (struct rwlock *)lock; 149167368Sjhb if (how) 150167368Sjhb rw_wlock(rw); 151167368Sjhb else 152167368Sjhb rw_rlock(rw); 153167368Sjhb} 154167368Sjhb 155167368Sjhbint 156167368Sjhbunlock_rw(struct lock_object *lock) 157167368Sjhb{ 158167368Sjhb struct rwlock *rw; 159167368Sjhb 160167368Sjhb rw = (struct rwlock *)lock; 161167368Sjhb rw_assert(rw, RA_LOCKED | LA_NOTRECURSED); 162167368Sjhb if (rw->rw_lock & RW_LOCK_READ) { 163167368Sjhb rw_runlock(rw); 164167368Sjhb return (0); 165167368Sjhb } else { 166167368Sjhb rw_wunlock(rw); 167167368Sjhb return (1); 168167368Sjhb } 169167368Sjhb} 170167368Sjhb 171192853Ssson#ifdef KDTRACE_HOOKS 172192853Sssonint 173227588Spjdowner_rw(const struct lock_object *lock, struct thread **owner) 174192853Ssson{ 175227588Spjd const struct rwlock *rw = (const struct rwlock *)lock; 176192853Ssson uintptr_t x = rw->rw_lock; 177192853Ssson 178192853Ssson *owner = rw_wowner(rw); 179192853Ssson return ((x & RW_LOCK_READ) != 0 ? (RW_READERS(x) != 0) : 180192853Ssson (*owner != NULL)); 181192853Ssson} 182192853Ssson#endif 183192853Ssson 184167368Sjhbvoid 185242515Sattilio_rw_init_flags(volatile uintptr_t *c, const char *name, int opts) 186154941Sjhb{ 187242515Sattilio struct rwlock *rw; 188171052Sattilio int flags; 189154941Sjhb 190242515Sattilio rw = rwlock2rw(c); 191242515Sattilio 192171052Sattilio MPASS((opts & ~(RW_DUPOK | RW_NOPROFILE | RW_NOWITNESS | RW_QUIET | 193171052Sattilio RW_RECURSE)) == 0); 194196334Sattilio ASSERT_ATOMIC_LOAD_PTR(rw->rw_lock, 195196334Sattilio ("%s: rw_lock not aligned for %s: %p", __func__, name, 196196334Sattilio &rw->rw_lock)); 197171052Sattilio 198193307Sattilio flags = LO_UPGRADABLE; 199171052Sattilio if (opts & RW_DUPOK) 200171052Sattilio flags |= LO_DUPOK; 201171052Sattilio if (opts & RW_NOPROFILE) 202171052Sattilio flags |= LO_NOPROFILE; 203171052Sattilio if (!(opts & RW_NOWITNESS)) 204171052Sattilio flags |= LO_WITNESS; 205193307Sattilio if (opts & RW_RECURSE) 206193307Sattilio flags |= LO_RECURSABLE; 207171052Sattilio if (opts & RW_QUIET) 208171052Sattilio flags |= LO_QUIET; 209171052Sattilio 210154941Sjhb rw->rw_lock = RW_UNLOCKED; 211171052Sattilio rw->rw_recurse = 0; 212171052Sattilio lock_init(&rw->lock_object, &lock_class_rw, name, NULL, flags); 213154941Sjhb} 214154941Sjhb 215154941Sjhbvoid 216242515Sattilio_rw_destroy(volatile uintptr_t *c) 217154941Sjhb{ 218242515Sattilio struct rwlock *rw; 219154941Sjhb 220242515Sattilio rw = rwlock2rw(c); 221242515Sattilio 222205626Sbz KASSERT(rw->rw_lock == RW_UNLOCKED, ("rw lock %p not unlocked", rw)); 223205626Sbz KASSERT(rw->rw_recurse == 0, ("rw lock %p still recursed", rw)); 224169394Sjhb rw->rw_lock = RW_DESTROYED; 225167787Sjhb lock_destroy(&rw->lock_object); 226154941Sjhb} 227154941Sjhb 228154941Sjhbvoid 229154941Sjhbrw_sysinit(void *arg) 230154941Sjhb{ 231154941Sjhb struct rw_args *args = arg; 232154941Sjhb 233242515Sattilio rw_init((struct rwlock *)args->ra_rw, args->ra_desc); 234154941Sjhb} 235154941Sjhb 236185778Skmacyvoid 237185778Skmacyrw_sysinit_flags(void *arg) 238185778Skmacy{ 239185778Skmacy struct rw_args_flags *args = arg; 240185778Skmacy 241242515Sattilio rw_init_flags((struct rwlock *)args->ra_rw, args->ra_desc, 242242515Sattilio args->ra_flags); 243185778Skmacy} 244185778Skmacy 245167024Srwatsonint 246242515Sattilio_rw_wowned(const volatile uintptr_t *c) 247167024Srwatson{ 248167024Srwatson 249242515Sattilio return (rw_wowner(rwlock2rw(c)) == curthread); 250167024Srwatson} 251167024Srwatson 252154941Sjhbvoid 253242515Sattilio_rw_wlock_cookie(volatile uintptr_t *c, const char *file, int line) 254154941Sjhb{ 255242515Sattilio struct rwlock *rw; 256154941Sjhb 257228424Savg if (SCHEDULER_STOPPED()) 258228424Savg return; 259242515Sattilio 260242515Sattilio rw = rwlock2rw(c); 261242515Sattilio 262244582Sattilio KASSERT(kdb_active != 0 || !TD_IS_IDLETHREAD(curthread), 263240424Sattilio ("rw_wlock() by idle thread %p on rwlock %s @ %s:%d", 264240424Sattilio curthread, rw->lock_object.lo_name, file, line)); 265169394Sjhb KASSERT(rw->rw_lock != RW_DESTROYED, 266169394Sjhb ("rw_wlock() of destroyed rwlock @ %s:%d", file, line)); 267167787Sjhb WITNESS_CHECKORDER(&rw->lock_object, LOP_NEWORDER | LOP_EXCLUSIVE, file, 268182914Sjhb line, NULL); 269154941Sjhb __rw_wlock(rw, curthread, file, line); 270171052Sattilio LOCK_LOG_LOCK("WLOCK", &rw->lock_object, 0, rw->rw_recurse, file, line); 271167787Sjhb WITNESS_LOCK(&rw->lock_object, LOP_EXCLUSIVE, file, line); 272160771Sjhb curthread->td_locks++; 273154941Sjhb} 274154941Sjhb 275177843Sattilioint 276242515Sattilio__rw_try_wlock(volatile uintptr_t *c, const char *file, int line) 277177843Sattilio{ 278242515Sattilio struct rwlock *rw; 279177843Sattilio int rval; 280177843Sattilio 281228424Savg if (SCHEDULER_STOPPED()) 282228424Savg return (1); 283228424Savg 284242515Sattilio rw = rwlock2rw(c); 285242515Sattilio 286244582Sattilio KASSERT(kdb_active != 0 || !TD_IS_IDLETHREAD(curthread), 287240424Sattilio ("rw_try_wlock() by idle thread %p on rwlock %s @ %s:%d", 288240424Sattilio curthread, rw->lock_object.lo_name, file, line)); 289177843Sattilio KASSERT(rw->rw_lock != RW_DESTROYED, 290177843Sattilio ("rw_try_wlock() of destroyed rwlock @ %s:%d", file, line)); 291177843Sattilio 292193307Sattilio if (rw_wlocked(rw) && 293193307Sattilio (rw->lock_object.lo_flags & LO_RECURSABLE) != 0) { 294177843Sattilio rw->rw_recurse++; 295177843Sattilio rval = 1; 296177843Sattilio } else 297177843Sattilio rval = atomic_cmpset_acq_ptr(&rw->rw_lock, RW_UNLOCKED, 298177843Sattilio (uintptr_t)curthread); 299177843Sattilio 300177843Sattilio LOCK_LOG_TRY("WLOCK", &rw->lock_object, 0, rval, file, line); 301177843Sattilio if (rval) { 302177843Sattilio WITNESS_LOCK(&rw->lock_object, LOP_EXCLUSIVE | LOP_TRYLOCK, 303177843Sattilio file, line); 304177843Sattilio curthread->td_locks++; 305177843Sattilio } 306177843Sattilio return (rval); 307177843Sattilio} 308177843Sattilio 309154941Sjhbvoid 310242515Sattilio_rw_wunlock_cookie(volatile uintptr_t *c, const char *file, int line) 311154941Sjhb{ 312242515Sattilio struct rwlock *rw; 313154941Sjhb 314228424Savg if (SCHEDULER_STOPPED()) 315228424Savg return; 316242515Sattilio 317242515Sattilio rw = rwlock2rw(c); 318242515Sattilio 319169394Sjhb KASSERT(rw->rw_lock != RW_DESTROYED, 320169394Sjhb ("rw_wunlock() of destroyed rwlock @ %s:%d", file, line)); 321242515Sattilio __rw_assert(c, RA_WLOCKED, file, line); 322160771Sjhb curthread->td_locks--; 323167787Sjhb WITNESS_UNLOCK(&rw->lock_object, LOP_EXCLUSIVE, file, line); 324171052Sattilio LOCK_LOG_LOCK("WUNLOCK", &rw->lock_object, 0, rw->rw_recurse, file, 325171052Sattilio line); 326171052Sattilio if (!rw_recursed(rw)) 327192853Ssson LOCKSTAT_PROFILE_RELEASE_LOCK(LS_RW_WUNLOCK_RELEASE, rw); 328154941Sjhb __rw_wunlock(rw, curthread, file, line); 329154941Sjhb} 330176017Sjeff/* 331176017Sjeff * Determines whether a new reader can acquire a lock. Succeeds if the 332176017Sjeff * reader already owns a read lock and the lock is locked for read to 333176017Sjeff * prevent deadlock from reader recursion. Also succeeds if the lock 334176017Sjeff * is unlocked and has no writer waiters or spinners. Failing otherwise 335176017Sjeff * prioritizes writers before readers. 336176017Sjeff */ 337176017Sjeff#define RW_CAN_READ(_rw) \ 338176017Sjeff ((curthread->td_rw_rlocks && (_rw) & RW_LOCK_READ) || ((_rw) & \ 339176017Sjeff (RW_LOCK_READ | RW_LOCK_WRITE_WAITERS | RW_LOCK_WRITE_SPINNER)) == \ 340176017Sjeff RW_LOCK_READ) 341154941Sjhb 342154941Sjhbvoid 343242515Sattilio__rw_rlock(volatile uintptr_t *c, const char *file, int line) 344154941Sjhb{ 345242515Sattilio struct rwlock *rw; 346170295Sjeff struct turnstile *ts; 347167801Sjhb#ifdef ADAPTIVE_RWLOCKS 348157846Sjhb volatile struct thread *owner; 349177912Sjeff int spintries = 0; 350177912Sjeff int i; 351157851Swkoszek#endif 352189846Sjeff#ifdef LOCK_PROFILING 353167307Sjhb uint64_t waittime = 0; 354167054Skmacy int contested = 0; 355189846Sjeff#endif 356176017Sjeff uintptr_t v; 357192853Ssson#ifdef KDTRACE_HOOKS 358192853Ssson uint64_t spin_cnt = 0; 359192853Ssson uint64_t sleep_cnt = 0; 360192853Ssson int64_t sleep_time = 0; 361192853Ssson#endif 362154941Sjhb 363228424Savg if (SCHEDULER_STOPPED()) 364228424Savg return; 365228424Savg 366242515Sattilio rw = rwlock2rw(c); 367242515Sattilio 368244582Sattilio KASSERT(kdb_active != 0 || !TD_IS_IDLETHREAD(curthread), 369240424Sattilio ("rw_rlock() by idle thread %p on rwlock %s @ %s:%d", 370240424Sattilio curthread, rw->lock_object.lo_name, file, line)); 371169394Sjhb KASSERT(rw->rw_lock != RW_DESTROYED, 372169394Sjhb ("rw_rlock() of destroyed rwlock @ %s:%d", file, line)); 373157826Sjhb KASSERT(rw_wowner(rw) != curthread, 374154941Sjhb ("%s (%s): wlock already held @ %s:%d", __func__, 375167787Sjhb rw->lock_object.lo_name, file, line)); 376182914Sjhb WITNESS_CHECKORDER(&rw->lock_object, LOP_NEWORDER, file, line, NULL); 377154941Sjhb 378154941Sjhb for (;;) { 379192853Ssson#ifdef KDTRACE_HOOKS 380192853Ssson spin_cnt++; 381192853Ssson#endif 382154941Sjhb /* 383154941Sjhb * Handle the easy case. If no other thread has a write 384154941Sjhb * lock, then try to bump up the count of read locks. Note 385154941Sjhb * that we have to preserve the current state of the 386154941Sjhb * RW_LOCK_WRITE_WAITERS flag. If we fail to acquire a 387154941Sjhb * read lock, then rw_lock must have changed, so restart 388154941Sjhb * the loop. Note that this handles the case of a 389154941Sjhb * completely unlocked rwlock since such a lock is encoded 390154941Sjhb * as a read lock with no waiters. 391154941Sjhb */ 392176017Sjeff v = rw->rw_lock; 393176017Sjeff if (RW_CAN_READ(v)) { 394154941Sjhb /* 395154941Sjhb * The RW_LOCK_READ_WAITERS flag should only be set 396176017Sjeff * if the lock has been unlocked and write waiters 397176017Sjeff * were present. 398154941Sjhb */ 399176017Sjeff if (atomic_cmpset_acq_ptr(&rw->rw_lock, v, 400176017Sjeff v + RW_ONE_READER)) { 401167787Sjhb if (LOCK_LOG_TEST(&rw->lock_object, 0)) 402154941Sjhb CTR4(KTR_LOCK, 403154941Sjhb "%s: %p succeed %p -> %p", __func__, 404176017Sjeff rw, (void *)v, 405176017Sjeff (void *)(v + RW_ONE_READER)); 406154941Sjhb break; 407154941Sjhb } 408154941Sjhb continue; 409154941Sjhb } 410233628Sfabient#ifdef HWPMC_HOOKS 411233628Sfabient PMC_SOFT_CALL( , , lock, failed); 412233628Sfabient#endif 413174629Sjeff lock_profile_obtain_lock_failed(&rw->lock_object, 414174629Sjeff &contested, &waittime); 415154941Sjhb 416173960Sattilio#ifdef ADAPTIVE_RWLOCKS 417154941Sjhb /* 418173960Sattilio * If the owner is running on another CPU, spin until 419173960Sattilio * the owner stops running or the state of the lock 420173960Sattilio * changes. 421173960Sattilio */ 422176017Sjeff if ((v & RW_LOCK_READ) == 0) { 423176017Sjeff owner = (struct thread *)RW_OWNER(v); 424176017Sjeff if (TD_IS_RUNNING(owner)) { 425176017Sjeff if (LOCK_LOG_TEST(&rw->lock_object, 0)) 426176017Sjeff CTR3(KTR_LOCK, 427176017Sjeff "%s: spinning on %p held by %p", 428176017Sjeff __func__, rw, owner); 429176017Sjeff while ((struct thread*)RW_OWNER(rw->rw_lock) == 430192853Ssson owner && TD_IS_RUNNING(owner)) { 431176017Sjeff cpu_spinwait(); 432192853Ssson#ifdef KDTRACE_HOOKS 433192853Ssson spin_cnt++; 434192853Ssson#endif 435192853Ssson } 436176017Sjeff continue; 437176017Sjeff } 438177912Sjeff } else if (spintries < rowner_retries) { 439177912Sjeff spintries++; 440177912Sjeff for (i = 0; i < rowner_loops; i++) { 441177912Sjeff v = rw->rw_lock; 442177912Sjeff if ((v & RW_LOCK_READ) == 0 || RW_CAN_READ(v)) 443177912Sjeff break; 444177912Sjeff cpu_spinwait(); 445177912Sjeff } 446177912Sjeff if (i != rowner_loops) 447177912Sjeff continue; 448173960Sattilio } 449173960Sattilio#endif 450173960Sattilio 451173960Sattilio /* 452154941Sjhb * Okay, now it's the hard case. Some other thread already 453176017Sjeff * has a write lock or there are write waiters present, 454176017Sjeff * acquire the turnstile lock so we can begin the process 455176017Sjeff * of blocking. 456154941Sjhb */ 457170295Sjeff ts = turnstile_trywait(&rw->lock_object); 458154941Sjhb 459154941Sjhb /* 460154941Sjhb * The lock might have been released while we spun, so 461176017Sjeff * recheck its state and restart the loop if needed. 462154941Sjhb */ 463176017Sjeff v = rw->rw_lock; 464176017Sjeff if (RW_CAN_READ(v)) { 465170295Sjeff turnstile_cancel(ts); 466154941Sjhb continue; 467154941Sjhb } 468154941Sjhb 469173960Sattilio#ifdef ADAPTIVE_RWLOCKS 470154941Sjhb /* 471193035Sjhb * The current lock owner might have started executing 472193035Sjhb * on another CPU (or the lock could have changed 473193035Sjhb * owners) while we were waiting on the turnstile 474193035Sjhb * chain lock. If so, drop the turnstile lock and try 475193035Sjhb * again. 476173960Sattilio */ 477176017Sjeff if ((v & RW_LOCK_READ) == 0) { 478176017Sjeff owner = (struct thread *)RW_OWNER(v); 479176017Sjeff if (TD_IS_RUNNING(owner)) { 480176017Sjeff turnstile_cancel(ts); 481176017Sjeff continue; 482176017Sjeff } 483173960Sattilio } 484173960Sattilio#endif 485173960Sattilio 486173960Sattilio /* 487176017Sjeff * The lock is held in write mode or it already has waiters. 488154941Sjhb */ 489176017Sjeff MPASS(!RW_CAN_READ(v)); 490176017Sjeff 491176017Sjeff /* 492176017Sjeff * If the RW_LOCK_READ_WAITERS flag is already set, then 493176017Sjeff * we can go ahead and block. If it is not set then try 494176017Sjeff * to set it. If we fail to set it drop the turnstile 495176017Sjeff * lock and restart the loop. 496176017Sjeff */ 497176017Sjeff if (!(v & RW_LOCK_READ_WAITERS)) { 498176017Sjeff if (!atomic_cmpset_ptr(&rw->rw_lock, v, 499176017Sjeff v | RW_LOCK_READ_WAITERS)) { 500170295Sjeff turnstile_cancel(ts); 501157826Sjhb continue; 502157826Sjhb } 503167787Sjhb if (LOCK_LOG_TEST(&rw->lock_object, 0)) 504157826Sjhb CTR2(KTR_LOCK, "%s: %p set read waiters flag", 505157826Sjhb __func__, rw); 506154941Sjhb } 507154941Sjhb 508154941Sjhb /* 509154941Sjhb * We were unable to acquire the lock and the read waiters 510154941Sjhb * flag is set, so we must block on the turnstile. 511154941Sjhb */ 512167787Sjhb if (LOCK_LOG_TEST(&rw->lock_object, 0)) 513154941Sjhb CTR2(KTR_LOCK, "%s: %p blocking on turnstile", __func__, 514154941Sjhb rw); 515192853Ssson#ifdef KDTRACE_HOOKS 516192853Ssson sleep_time -= lockstat_nsecs(); 517192853Ssson#endif 518170295Sjeff turnstile_wait(ts, rw_owner(rw), TS_SHARED_QUEUE); 519192853Ssson#ifdef KDTRACE_HOOKS 520192853Ssson sleep_time += lockstat_nsecs(); 521192853Ssson sleep_cnt++; 522192853Ssson#endif 523167787Sjhb if (LOCK_LOG_TEST(&rw->lock_object, 0)) 524154941Sjhb CTR2(KTR_LOCK, "%s: %p resuming from turnstile", 525154941Sjhb __func__, rw); 526154941Sjhb } 527154941Sjhb 528154941Sjhb /* 529154941Sjhb * TODO: acquire "owner of record" here. Here be turnstile dragons 530154941Sjhb * however. turnstiles don't like owners changing between calls to 531154941Sjhb * turnstile_wait() currently. 532154941Sjhb */ 533192853Ssson LOCKSTAT_PROFILE_OBTAIN_LOCK_SUCCESS(LS_RW_RLOCK_ACQUIRE, rw, contested, 534174629Sjeff waittime, file, line); 535167787Sjhb LOCK_LOG_LOCK("RLOCK", &rw->lock_object, 0, 0, file, line); 536167787Sjhb WITNESS_LOCK(&rw->lock_object, 0, file, line); 537160771Sjhb curthread->td_locks++; 538176017Sjeff curthread->td_rw_rlocks++; 539192853Ssson#ifdef KDTRACE_HOOKS 540192853Ssson if (sleep_time) 541192853Ssson LOCKSTAT_RECORD1(LS_RW_RLOCK_BLOCK, rw, sleep_time); 542192853Ssson 543192853Ssson /* 544192853Ssson * Record only the loops spinning and not sleeping. 545192853Ssson */ 546192853Ssson if (spin_cnt > sleep_cnt) 547192853Ssson LOCKSTAT_RECORD1(LS_RW_RLOCK_SPIN, rw, (spin_cnt - sleep_cnt)); 548192853Ssson#endif 549154941Sjhb} 550154941Sjhb 551177843Sattilioint 552242515Sattilio__rw_try_rlock(volatile uintptr_t *c, const char *file, int line) 553177843Sattilio{ 554242515Sattilio struct rwlock *rw; 555177843Sattilio uintptr_t x; 556177843Sattilio 557228424Savg if (SCHEDULER_STOPPED()) 558228424Savg return (1); 559228424Savg 560242515Sattilio rw = rwlock2rw(c); 561242515Sattilio 562244582Sattilio KASSERT(kdb_active != 0 || !TD_IS_IDLETHREAD(curthread), 563240424Sattilio ("rw_try_rlock() by idle thread %p on rwlock %s @ %s:%d", 564240424Sattilio curthread, rw->lock_object.lo_name, file, line)); 565240424Sattilio 566177843Sattilio for (;;) { 567177843Sattilio x = rw->rw_lock; 568177843Sattilio KASSERT(rw->rw_lock != RW_DESTROYED, 569177843Sattilio ("rw_try_rlock() of destroyed rwlock @ %s:%d", file, line)); 570177843Sattilio if (!(x & RW_LOCK_READ)) 571177843Sattilio break; 572177843Sattilio if (atomic_cmpset_acq_ptr(&rw->rw_lock, x, x + RW_ONE_READER)) { 573177843Sattilio LOCK_LOG_TRY("RLOCK", &rw->lock_object, 0, 1, file, 574177843Sattilio line); 575177843Sattilio WITNESS_LOCK(&rw->lock_object, LOP_TRYLOCK, file, line); 576177843Sattilio curthread->td_locks++; 577177843Sattilio curthread->td_rw_rlocks++; 578177843Sattilio return (1); 579177843Sattilio } 580177843Sattilio } 581177843Sattilio 582177843Sattilio LOCK_LOG_TRY("RLOCK", &rw->lock_object, 0, 0, file, line); 583177843Sattilio return (0); 584177843Sattilio} 585177843Sattilio 586154941Sjhbvoid 587242515Sattilio_rw_runlock_cookie(volatile uintptr_t *c, const char *file, int line) 588154941Sjhb{ 589242515Sattilio struct rwlock *rw; 590154941Sjhb struct turnstile *ts; 591176017Sjeff uintptr_t x, v, queue; 592154941Sjhb 593228424Savg if (SCHEDULER_STOPPED()) 594228424Savg return; 595228424Savg 596242515Sattilio rw = rwlock2rw(c); 597242515Sattilio 598169394Sjhb KASSERT(rw->rw_lock != RW_DESTROYED, 599169394Sjhb ("rw_runlock() of destroyed rwlock @ %s:%d", file, line)); 600242515Sattilio __rw_assert(c, RA_RLOCKED, file, line); 601160771Sjhb curthread->td_locks--; 602176017Sjeff curthread->td_rw_rlocks--; 603167787Sjhb WITNESS_UNLOCK(&rw->lock_object, 0, file, line); 604167787Sjhb LOCK_LOG_LOCK("RUNLOCK", &rw->lock_object, 0, 0, file, line); 605154941Sjhb 606154941Sjhb /* TODO: drop "owner of record" here. */ 607154941Sjhb 608154941Sjhb for (;;) { 609154941Sjhb /* 610154941Sjhb * See if there is more than one read lock held. If so, 611154941Sjhb * just drop one and return. 612154941Sjhb */ 613154941Sjhb x = rw->rw_lock; 614154941Sjhb if (RW_READERS(x) > 1) { 615197643Sattilio if (atomic_cmpset_rel_ptr(&rw->rw_lock, x, 616154941Sjhb x - RW_ONE_READER)) { 617167787Sjhb if (LOCK_LOG_TEST(&rw->lock_object, 0)) 618154941Sjhb CTR4(KTR_LOCK, 619154941Sjhb "%s: %p succeeded %p -> %p", 620154941Sjhb __func__, rw, (void *)x, 621154941Sjhb (void *)(x - RW_ONE_READER)); 622154941Sjhb break; 623154941Sjhb } 624154941Sjhb continue; 625167307Sjhb } 626154941Sjhb /* 627154941Sjhb * If there aren't any waiters for a write lock, then try 628154941Sjhb * to drop it quickly. 629154941Sjhb */ 630176017Sjeff if (!(x & RW_LOCK_WAITERS)) { 631176017Sjeff MPASS((x & ~RW_LOCK_WRITE_SPINNER) == 632176017Sjeff RW_READERS_LOCK(1)); 633197643Sattilio if (atomic_cmpset_rel_ptr(&rw->rw_lock, x, 634197643Sattilio RW_UNLOCKED)) { 635167787Sjhb if (LOCK_LOG_TEST(&rw->lock_object, 0)) 636154941Sjhb CTR2(KTR_LOCK, "%s: %p last succeeded", 637154941Sjhb __func__, rw); 638154941Sjhb break; 639154941Sjhb } 640154941Sjhb continue; 641154941Sjhb } 642154941Sjhb /* 643176017Sjeff * Ok, we know we have waiters and we think we are the 644176017Sjeff * last reader, so grab the turnstile lock. 645154941Sjhb */ 646170295Sjeff turnstile_chain_lock(&rw->lock_object); 647176017Sjeff v = rw->rw_lock & (RW_LOCK_WAITERS | RW_LOCK_WRITE_SPINNER); 648176017Sjeff MPASS(v & RW_LOCK_WAITERS); 649154941Sjhb 650154941Sjhb /* 651154941Sjhb * Try to drop our lock leaving the lock in a unlocked 652154941Sjhb * state. 653154941Sjhb * 654154941Sjhb * If you wanted to do explicit lock handoff you'd have to 655154941Sjhb * do it here. You'd also want to use turnstile_signal() 656154941Sjhb * and you'd have to handle the race where a higher 657154941Sjhb * priority thread blocks on the write lock before the 658154941Sjhb * thread you wakeup actually runs and have the new thread 659154941Sjhb * "steal" the lock. For now it's a lot simpler to just 660154941Sjhb * wakeup all of the waiters. 661154941Sjhb * 662154941Sjhb * As above, if we fail, then another thread might have 663154941Sjhb * acquired a read lock, so drop the turnstile lock and 664154941Sjhb * restart. 665154941Sjhb */ 666176017Sjeff x = RW_UNLOCKED; 667176017Sjeff if (v & RW_LOCK_WRITE_WAITERS) { 668176017Sjeff queue = TS_EXCLUSIVE_QUEUE; 669176017Sjeff x |= (v & RW_LOCK_READ_WAITERS); 670176017Sjeff } else 671176017Sjeff queue = TS_SHARED_QUEUE; 672197643Sattilio if (!atomic_cmpset_rel_ptr(&rw->rw_lock, RW_READERS_LOCK(1) | v, 673176017Sjeff x)) { 674170295Sjeff turnstile_chain_unlock(&rw->lock_object); 675154941Sjhb continue; 676154941Sjhb } 677167787Sjhb if (LOCK_LOG_TEST(&rw->lock_object, 0)) 678154941Sjhb CTR2(KTR_LOCK, "%s: %p last succeeded with waiters", 679154941Sjhb __func__, rw); 680154941Sjhb 681154941Sjhb /* 682154941Sjhb * Ok. The lock is released and all that's left is to 683154941Sjhb * wake up the waiters. Note that the lock might not be 684154941Sjhb * free anymore, but in that case the writers will just 685154941Sjhb * block again if they run before the new lock holder(s) 686154941Sjhb * release the lock. 687154941Sjhb */ 688167787Sjhb ts = turnstile_lookup(&rw->lock_object); 689157846Sjhb MPASS(ts != NULL); 690176017Sjeff turnstile_broadcast(ts, queue); 691154941Sjhb turnstile_unpend(ts, TS_SHARED_LOCK); 692170295Sjeff turnstile_chain_unlock(&rw->lock_object); 693154941Sjhb break; 694154941Sjhb } 695192853Ssson LOCKSTAT_PROFILE_RELEASE_LOCK(LS_RW_RUNLOCK_RELEASE, rw); 696154941Sjhb} 697154941Sjhb 698154941Sjhb/* 699154941Sjhb * This function is called when we are unable to obtain a write lock on the 700154941Sjhb * first try. This means that at least one other thread holds either a 701154941Sjhb * read or write lock. 702154941Sjhb */ 703154941Sjhbvoid 704242515Sattilio__rw_wlock_hard(volatile uintptr_t *c, uintptr_t tid, const char *file, 705242515Sattilio int line) 706154941Sjhb{ 707242515Sattilio struct rwlock *rw; 708170295Sjeff struct turnstile *ts; 709167801Sjhb#ifdef ADAPTIVE_RWLOCKS 710157846Sjhb volatile struct thread *owner; 711176017Sjeff int spintries = 0; 712176017Sjeff int i; 713157851Swkoszek#endif 714189846Sjeff uintptr_t v, x; 715189846Sjeff#ifdef LOCK_PROFILING 716171516Sattilio uint64_t waittime = 0; 717171516Sattilio int contested = 0; 718189846Sjeff#endif 719192853Ssson#ifdef KDTRACE_HOOKS 720192853Ssson uint64_t spin_cnt = 0; 721192853Ssson uint64_t sleep_cnt = 0; 722192853Ssson int64_t sleep_time = 0; 723192853Ssson#endif 724154941Sjhb 725228424Savg if (SCHEDULER_STOPPED()) 726228424Savg return; 727228424Savg 728242515Sattilio rw = rwlock2rw(c); 729242515Sattilio 730171052Sattilio if (rw_wlocked(rw)) { 731193307Sattilio KASSERT(rw->lock_object.lo_flags & LO_RECURSABLE, 732171052Sattilio ("%s: recursing but non-recursive rw %s @ %s:%d\n", 733171052Sattilio __func__, rw->lock_object.lo_name, file, line)); 734171052Sattilio rw->rw_recurse++; 735171052Sattilio if (LOCK_LOG_TEST(&rw->lock_object, 0)) 736171052Sattilio CTR2(KTR_LOCK, "%s: %p recursing", __func__, rw); 737171052Sattilio return; 738171052Sattilio } 739171052Sattilio 740167787Sjhb if (LOCK_LOG_TEST(&rw->lock_object, 0)) 741154941Sjhb CTR5(KTR_LOCK, "%s: %s contested (lock=%p) at %s:%d", __func__, 742167787Sjhb rw->lock_object.lo_name, (void *)rw->rw_lock, file, line); 743154941Sjhb 744154941Sjhb while (!_rw_write_lock(rw, tid)) { 745192853Ssson#ifdef KDTRACE_HOOKS 746192853Ssson spin_cnt++; 747192853Ssson#endif 748233628Sfabient#ifdef HWPMC_HOOKS 749233628Sfabient PMC_SOFT_CALL( , , lock, failed); 750233628Sfabient#endif 751174629Sjeff lock_profile_obtain_lock_failed(&rw->lock_object, 752174629Sjeff &contested, &waittime); 753173960Sattilio#ifdef ADAPTIVE_RWLOCKS 754173960Sattilio /* 755173960Sattilio * If the lock is write locked and the owner is 756173960Sattilio * running on another CPU, spin until the owner stops 757173960Sattilio * running or the state of the lock changes. 758173960Sattilio */ 759173960Sattilio v = rw->rw_lock; 760173960Sattilio owner = (struct thread *)RW_OWNER(v); 761173960Sattilio if (!(v & RW_LOCK_READ) && TD_IS_RUNNING(owner)) { 762173960Sattilio if (LOCK_LOG_TEST(&rw->lock_object, 0)) 763173960Sattilio CTR3(KTR_LOCK, "%s: spinning on %p held by %p", 764173960Sattilio __func__, rw, owner); 765173960Sattilio while ((struct thread*)RW_OWNER(rw->rw_lock) == owner && 766192853Ssson TD_IS_RUNNING(owner)) { 767173960Sattilio cpu_spinwait(); 768192853Ssson#ifdef KDTRACE_HOOKS 769192853Ssson spin_cnt++; 770192853Ssson#endif 771192853Ssson } 772173960Sattilio continue; 773173960Sattilio } 774177912Sjeff if ((v & RW_LOCK_READ) && RW_READERS(v) && 775177912Sjeff spintries < rowner_retries) { 776176017Sjeff if (!(v & RW_LOCK_WRITE_SPINNER)) { 777176017Sjeff if (!atomic_cmpset_ptr(&rw->rw_lock, v, 778176017Sjeff v | RW_LOCK_WRITE_SPINNER)) { 779176017Sjeff continue; 780176017Sjeff } 781176017Sjeff } 782176017Sjeff spintries++; 783177912Sjeff for (i = 0; i < rowner_loops; i++) { 784176017Sjeff if ((rw->rw_lock & RW_LOCK_WRITE_SPINNER) == 0) 785176017Sjeff break; 786176017Sjeff cpu_spinwait(); 787176017Sjeff } 788192853Ssson#ifdef KDTRACE_HOOKS 789192853Ssson spin_cnt += rowner_loops - i; 790192853Ssson#endif 791177912Sjeff if (i != rowner_loops) 792176017Sjeff continue; 793176017Sjeff } 794173960Sattilio#endif 795170295Sjeff ts = turnstile_trywait(&rw->lock_object); 796154941Sjhb v = rw->rw_lock; 797154941Sjhb 798173960Sattilio#ifdef ADAPTIVE_RWLOCKS 799154941Sjhb /* 800193035Sjhb * The current lock owner might have started executing 801193035Sjhb * on another CPU (or the lock could have changed 802193035Sjhb * owners) while we were waiting on the turnstile 803193035Sjhb * chain lock. If so, drop the turnstile lock and try 804193035Sjhb * again. 805173960Sattilio */ 806173960Sattilio if (!(v & RW_LOCK_READ)) { 807173960Sattilio owner = (struct thread *)RW_OWNER(v); 808173960Sattilio if (TD_IS_RUNNING(owner)) { 809173960Sattilio turnstile_cancel(ts); 810173960Sattilio continue; 811173960Sattilio } 812173960Sattilio } 813173960Sattilio#endif 814173960Sattilio /* 815179334Sattilio * Check for the waiters flags about this rwlock. 816179334Sattilio * If the lock was released, without maintain any pending 817179334Sattilio * waiters queue, simply try to acquire it. 818179334Sattilio * If a pending waiters queue is present, claim the lock 819179334Sattilio * ownership and maintain the pending queue. 820154941Sjhb */ 821176017Sjeff x = v & (RW_LOCK_WAITERS | RW_LOCK_WRITE_SPINNER); 822176017Sjeff if ((v & ~x) == RW_UNLOCKED) { 823176017Sjeff x &= ~RW_LOCK_WRITE_SPINNER; 824176017Sjeff if (atomic_cmpset_acq_ptr(&rw->rw_lock, v, tid | x)) { 825176017Sjeff if (x) 826176017Sjeff turnstile_claim(ts); 827176017Sjeff else 828176017Sjeff turnstile_cancel(ts); 829154941Sjhb break; 830154941Sjhb } 831170295Sjeff turnstile_cancel(ts); 832154941Sjhb continue; 833154941Sjhb } 834154941Sjhb /* 835154941Sjhb * If the RW_LOCK_WRITE_WAITERS flag isn't set, then try to 836154941Sjhb * set it. If we fail to set it, then loop back and try 837154941Sjhb * again. 838154941Sjhb */ 839157826Sjhb if (!(v & RW_LOCK_WRITE_WAITERS)) { 840157826Sjhb if (!atomic_cmpset_ptr(&rw->rw_lock, v, 841157826Sjhb v | RW_LOCK_WRITE_WAITERS)) { 842170295Sjeff turnstile_cancel(ts); 843157826Sjhb continue; 844157826Sjhb } 845167787Sjhb if (LOCK_LOG_TEST(&rw->lock_object, 0)) 846157826Sjhb CTR2(KTR_LOCK, "%s: %p set write waiters flag", 847157826Sjhb __func__, rw); 848154941Sjhb } 849157846Sjhb /* 850154941Sjhb * We were unable to acquire the lock and the write waiters 851154941Sjhb * flag is set, so we must block on the turnstile. 852154941Sjhb */ 853167787Sjhb if (LOCK_LOG_TEST(&rw->lock_object, 0)) 854154941Sjhb CTR2(KTR_LOCK, "%s: %p blocking on turnstile", __func__, 855154941Sjhb rw); 856192853Ssson#ifdef KDTRACE_HOOKS 857192853Ssson sleep_time -= lockstat_nsecs(); 858192853Ssson#endif 859170295Sjeff turnstile_wait(ts, rw_owner(rw), TS_EXCLUSIVE_QUEUE); 860192853Ssson#ifdef KDTRACE_HOOKS 861192853Ssson sleep_time += lockstat_nsecs(); 862192853Ssson sleep_cnt++; 863192853Ssson#endif 864167787Sjhb if (LOCK_LOG_TEST(&rw->lock_object, 0)) 865154941Sjhb CTR2(KTR_LOCK, "%s: %p resuming from turnstile", 866154941Sjhb __func__, rw); 867176017Sjeff#ifdef ADAPTIVE_RWLOCKS 868176017Sjeff spintries = 0; 869176017Sjeff#endif 870154941Sjhb } 871192853Ssson LOCKSTAT_PROFILE_OBTAIN_LOCK_SUCCESS(LS_RW_WLOCK_ACQUIRE, rw, contested, 872192853Ssson waittime, file, line); 873192853Ssson#ifdef KDTRACE_HOOKS 874192853Ssson if (sleep_time) 875192853Ssson LOCKSTAT_RECORD1(LS_RW_WLOCK_BLOCK, rw, sleep_time); 876192853Ssson 877192853Ssson /* 878192853Ssson * Record only the loops spinning and not sleeping. 879192853Ssson */ 880192853Ssson if (spin_cnt > sleep_cnt) 881192853Ssson LOCKSTAT_RECORD1(LS_RW_WLOCK_SPIN, rw, (spin_cnt - sleep_cnt)); 882192853Ssson#endif 883154941Sjhb} 884154941Sjhb 885154941Sjhb/* 886154941Sjhb * This function is called if the first try at releasing a write lock failed. 887154941Sjhb * This means that one of the 2 waiter bits must be set indicating that at 888154941Sjhb * least one thread is waiting on this lock. 889154941Sjhb */ 890154941Sjhbvoid 891242515Sattilio__rw_wunlock_hard(volatile uintptr_t *c, uintptr_t tid, const char *file, 892242515Sattilio int line) 893154941Sjhb{ 894242515Sattilio struct rwlock *rw; 895154941Sjhb struct turnstile *ts; 896154941Sjhb uintptr_t v; 897154941Sjhb int queue; 898154941Sjhb 899228424Savg if (SCHEDULER_STOPPED()) 900228424Savg return; 901228424Savg 902242515Sattilio rw = rwlock2rw(c); 903242515Sattilio 904171052Sattilio if (rw_wlocked(rw) && rw_recursed(rw)) { 905176017Sjeff rw->rw_recurse--; 906171052Sattilio if (LOCK_LOG_TEST(&rw->lock_object, 0)) 907171052Sattilio CTR2(KTR_LOCK, "%s: %p unrecursing", __func__, rw); 908171052Sattilio return; 909171052Sattilio } 910171052Sattilio 911154941Sjhb KASSERT(rw->rw_lock & (RW_LOCK_READ_WAITERS | RW_LOCK_WRITE_WAITERS), 912154941Sjhb ("%s: neither of the waiter flags are set", __func__)); 913154941Sjhb 914167787Sjhb if (LOCK_LOG_TEST(&rw->lock_object, 0)) 915154941Sjhb CTR2(KTR_LOCK, "%s: %p contested", __func__, rw); 916154941Sjhb 917170295Sjeff turnstile_chain_lock(&rw->lock_object); 918167787Sjhb ts = turnstile_lookup(&rw->lock_object); 919154941Sjhb MPASS(ts != NULL); 920154941Sjhb 921154941Sjhb /* 922154941Sjhb * Use the same algo as sx locks for now. Prefer waking up shared 923154941Sjhb * waiters if we have any over writers. This is probably not ideal. 924154941Sjhb * 925154941Sjhb * 'v' is the value we are going to write back to rw_lock. If we 926154941Sjhb * have waiters on both queues, we need to preserve the state of 927154941Sjhb * the waiter flag for the queue we don't wake up. For now this is 928154941Sjhb * hardcoded for the algorithm mentioned above. 929154941Sjhb * 930154941Sjhb * In the case of both readers and writers waiting we wakeup the 931154941Sjhb * readers but leave the RW_LOCK_WRITE_WAITERS flag set. If a 932154941Sjhb * new writer comes in before a reader it will claim the lock up 933154941Sjhb * above. There is probably a potential priority inversion in 934154941Sjhb * there that could be worked around either by waking both queues 935154941Sjhb * of waiters or doing some complicated lock handoff gymnastics. 936154941Sjhb */ 937157846Sjhb v = RW_UNLOCKED; 938176076Sjeff if (rw->rw_lock & RW_LOCK_WRITE_WAITERS) { 939176076Sjeff queue = TS_EXCLUSIVE_QUEUE; 940176076Sjeff v |= (rw->rw_lock & RW_LOCK_READ_WAITERS); 941176076Sjeff } else 942154941Sjhb queue = TS_SHARED_QUEUE; 943157846Sjhb 944157846Sjhb /* Wake up all waiters for the specific queue. */ 945167787Sjhb if (LOCK_LOG_TEST(&rw->lock_object, 0)) 946154941Sjhb CTR3(KTR_LOCK, "%s: %p waking up %s waiters", __func__, rw, 947154941Sjhb queue == TS_SHARED_QUEUE ? "read" : "write"); 948154941Sjhb turnstile_broadcast(ts, queue); 949154941Sjhb atomic_store_rel_ptr(&rw->rw_lock, v); 950154941Sjhb turnstile_unpend(ts, TS_EXCLUSIVE_LOCK); 951170295Sjeff turnstile_chain_unlock(&rw->lock_object); 952154941Sjhb} 953154941Sjhb 954157882Sjhb/* 955157882Sjhb * Attempt to do a non-blocking upgrade from a read lock to a write 956157882Sjhb * lock. This will only succeed if this thread holds a single read 957157882Sjhb * lock. Returns true if the upgrade succeeded and false otherwise. 958157882Sjhb */ 959157882Sjhbint 960242515Sattilio__rw_try_upgrade(volatile uintptr_t *c, const char *file, int line) 961157882Sjhb{ 962242515Sattilio struct rwlock *rw; 963176017Sjeff uintptr_t v, x, tid; 964170295Sjeff struct turnstile *ts; 965157882Sjhb int success; 966157882Sjhb 967228424Savg if (SCHEDULER_STOPPED()) 968228424Savg return (1); 969228424Savg 970242515Sattilio rw = rwlock2rw(c); 971242515Sattilio 972169394Sjhb KASSERT(rw->rw_lock != RW_DESTROYED, 973169394Sjhb ("rw_try_upgrade() of destroyed rwlock @ %s:%d", file, line)); 974242515Sattilio __rw_assert(c, RA_RLOCKED, file, line); 975157882Sjhb 976157882Sjhb /* 977157882Sjhb * Attempt to switch from one reader to a writer. If there 978157882Sjhb * are any write waiters, then we will have to lock the 979157882Sjhb * turnstile first to prevent races with another writer 980157882Sjhb * calling turnstile_wait() before we have claimed this 981157882Sjhb * turnstile. So, do the simple case of no waiters first. 982157882Sjhb */ 983157882Sjhb tid = (uintptr_t)curthread; 984176017Sjeff success = 0; 985176017Sjeff for (;;) { 986176017Sjeff v = rw->rw_lock; 987176017Sjeff if (RW_READERS(v) > 1) 988176017Sjeff break; 989176017Sjeff if (!(v & RW_LOCK_WAITERS)) { 990176017Sjeff success = atomic_cmpset_ptr(&rw->rw_lock, v, tid); 991176017Sjeff if (!success) 992176017Sjeff continue; 993176017Sjeff break; 994176017Sjeff } 995157882Sjhb 996176017Sjeff /* 997176017Sjeff * Ok, we think we have waiters, so lock the turnstile. 998176017Sjeff */ 999176017Sjeff ts = turnstile_trywait(&rw->lock_object); 1000176017Sjeff v = rw->rw_lock; 1001176017Sjeff if (RW_READERS(v) > 1) { 1002176017Sjeff turnstile_cancel(ts); 1003176017Sjeff break; 1004176017Sjeff } 1005176017Sjeff /* 1006176017Sjeff * Try to switch from one reader to a writer again. This time 1007176017Sjeff * we honor the current state of the waiters flags. 1008176017Sjeff * If we obtain the lock with the flags set, then claim 1009176017Sjeff * ownership of the turnstile. 1010176017Sjeff */ 1011176017Sjeff x = rw->rw_lock & RW_LOCK_WAITERS; 1012176017Sjeff success = atomic_cmpset_ptr(&rw->rw_lock, v, tid | x); 1013176017Sjeff if (success) { 1014176017Sjeff if (x) 1015176017Sjeff turnstile_claim(ts); 1016176017Sjeff else 1017176017Sjeff turnstile_cancel(ts); 1018176017Sjeff break; 1019176017Sjeff } 1020170295Sjeff turnstile_cancel(ts); 1021176017Sjeff } 1022167787Sjhb LOCK_LOG_TRY("WUPGRADE", &rw->lock_object, 0, success, file, line); 1023176017Sjeff if (success) { 1024176017Sjeff curthread->td_rw_rlocks--; 1025167787Sjhb WITNESS_UPGRADE(&rw->lock_object, LOP_EXCLUSIVE | LOP_TRYLOCK, 1026157882Sjhb file, line); 1027192853Ssson LOCKSTAT_RECORD0(LS_RW_TRYUPGRADE_UPGRADE, rw); 1028176017Sjeff } 1029157882Sjhb return (success); 1030157882Sjhb} 1031157882Sjhb 1032157882Sjhb/* 1033157882Sjhb * Downgrade a write lock into a single read lock. 1034157882Sjhb */ 1035157882Sjhbvoid 1036242515Sattilio__rw_downgrade(volatile uintptr_t *c, const char *file, int line) 1037157882Sjhb{ 1038242515Sattilio struct rwlock *rw; 1039157882Sjhb struct turnstile *ts; 1040157882Sjhb uintptr_t tid, v; 1041176017Sjeff int rwait, wwait; 1042157882Sjhb 1043228424Savg if (SCHEDULER_STOPPED()) 1044228424Savg return; 1045228424Savg 1046242515Sattilio rw = rwlock2rw(c); 1047242515Sattilio 1048169394Sjhb KASSERT(rw->rw_lock != RW_DESTROYED, 1049169394Sjhb ("rw_downgrade() of destroyed rwlock @ %s:%d", file, line)); 1050242515Sattilio __rw_assert(c, RA_WLOCKED | RA_NOTRECURSED, file, line); 1051171052Sattilio#ifndef INVARIANTS 1052171052Sattilio if (rw_recursed(rw)) 1053171052Sattilio panic("downgrade of a recursed lock"); 1054171052Sattilio#endif 1055157882Sjhb 1056167787Sjhb WITNESS_DOWNGRADE(&rw->lock_object, 0, file, line); 1057157882Sjhb 1058157882Sjhb /* 1059157882Sjhb * Convert from a writer to a single reader. First we handle 1060157882Sjhb * the easy case with no waiters. If there are any waiters, we 1061176017Sjeff * lock the turnstile and "disown" the lock. 1062157882Sjhb */ 1063157882Sjhb tid = (uintptr_t)curthread; 1064157882Sjhb if (atomic_cmpset_rel_ptr(&rw->rw_lock, tid, RW_READERS_LOCK(1))) 1065157882Sjhb goto out; 1066157882Sjhb 1067157882Sjhb /* 1068157882Sjhb * Ok, we think we have waiters, so lock the turnstile so we can 1069157882Sjhb * read the waiter flags without any races. 1070157882Sjhb */ 1071170295Sjeff turnstile_chain_lock(&rw->lock_object); 1072176017Sjeff v = rw->rw_lock & RW_LOCK_WAITERS; 1073176017Sjeff rwait = v & RW_LOCK_READ_WAITERS; 1074176017Sjeff wwait = v & RW_LOCK_WRITE_WAITERS; 1075176017Sjeff MPASS(rwait | wwait); 1076157882Sjhb 1077157882Sjhb /* 1078176017Sjeff * Downgrade from a write lock while preserving waiters flag 1079176017Sjeff * and give up ownership of the turnstile. 1080157882Sjhb */ 1081167787Sjhb ts = turnstile_lookup(&rw->lock_object); 1082157882Sjhb MPASS(ts != NULL); 1083176017Sjeff if (!wwait) 1084176017Sjeff v &= ~RW_LOCK_READ_WAITERS; 1085176017Sjeff atomic_store_rel_ptr(&rw->rw_lock, RW_READERS_LOCK(1) | v); 1086176017Sjeff /* 1087176017Sjeff * Wake other readers if there are no writers pending. Otherwise they 1088176017Sjeff * won't be able to acquire the lock anyway. 1089176017Sjeff */ 1090176017Sjeff if (rwait && !wwait) { 1091157882Sjhb turnstile_broadcast(ts, TS_SHARED_QUEUE); 1092157882Sjhb turnstile_unpend(ts, TS_EXCLUSIVE_LOCK); 1093176017Sjeff } else 1094157882Sjhb turnstile_disown(ts); 1095170295Sjeff turnstile_chain_unlock(&rw->lock_object); 1096157882Sjhbout: 1097176017Sjeff curthread->td_rw_rlocks++; 1098167787Sjhb LOCK_LOG_LOCK("WDOWNGRADE", &rw->lock_object, 0, 0, file, line); 1099192853Ssson LOCKSTAT_RECORD0(LS_RW_DOWNGRADE_DOWNGRADE, rw); 1100157882Sjhb} 1101157882Sjhb 1102154941Sjhb#ifdef INVARIANT_SUPPORT 1103155162Sscottl#ifndef INVARIANTS 1104242515Sattilio#undef __rw_assert 1105154941Sjhb#endif 1106154941Sjhb 1107154941Sjhb/* 1108154941Sjhb * In the non-WITNESS case, rw_assert() can only detect that at least 1109154941Sjhb * *some* thread owns an rlock, but it cannot guarantee that *this* 1110154941Sjhb * thread owns an rlock. 1111154941Sjhb */ 1112154941Sjhbvoid 1113242515Sattilio__rw_assert(const volatile uintptr_t *c, int what, const char *file, int line) 1114154941Sjhb{ 1115242515Sattilio const struct rwlock *rw; 1116154941Sjhb 1117154941Sjhb if (panicstr != NULL) 1118154941Sjhb return; 1119242515Sattilio 1120242515Sattilio rw = rwlock2rw(c); 1121242515Sattilio 1122154941Sjhb switch (what) { 1123154941Sjhb case RA_LOCKED: 1124171052Sattilio case RA_LOCKED | RA_RECURSED: 1125171052Sattilio case RA_LOCKED | RA_NOTRECURSED: 1126154941Sjhb case RA_RLOCKED: 1127154941Sjhb#ifdef WITNESS 1128167787Sjhb witness_assert(&rw->lock_object, what, file, line); 1129154941Sjhb#else 1130154941Sjhb /* 1131154941Sjhb * If some other thread has a write lock or we have one 1132154941Sjhb * and are asserting a read lock, fail. Also, if no one 1133154941Sjhb * has a lock at all, fail. 1134154941Sjhb */ 1135155061Sscottl if (rw->rw_lock == RW_UNLOCKED || 1136155061Sscottl (!(rw->rw_lock & RW_LOCK_READ) && (what == RA_RLOCKED || 1137157826Sjhb rw_wowner(rw) != curthread))) 1138154941Sjhb panic("Lock %s not %slocked @ %s:%d\n", 1139167787Sjhb rw->lock_object.lo_name, (what == RA_RLOCKED) ? 1140154941Sjhb "read " : "", file, line); 1141171052Sattilio 1142171052Sattilio if (!(rw->rw_lock & RW_LOCK_READ)) { 1143171052Sattilio if (rw_recursed(rw)) { 1144171052Sattilio if (what & RA_NOTRECURSED) 1145171052Sattilio panic("Lock %s recursed @ %s:%d\n", 1146171052Sattilio rw->lock_object.lo_name, file, 1147171052Sattilio line); 1148171052Sattilio } else if (what & RA_RECURSED) 1149171052Sattilio panic("Lock %s not recursed @ %s:%d\n", 1150171052Sattilio rw->lock_object.lo_name, file, line); 1151171052Sattilio } 1152154941Sjhb#endif 1153154941Sjhb break; 1154154941Sjhb case RA_WLOCKED: 1155171052Sattilio case RA_WLOCKED | RA_RECURSED: 1156171052Sattilio case RA_WLOCKED | RA_NOTRECURSED: 1157157826Sjhb if (rw_wowner(rw) != curthread) 1158154941Sjhb panic("Lock %s not exclusively locked @ %s:%d\n", 1159167787Sjhb rw->lock_object.lo_name, file, line); 1160171052Sattilio if (rw_recursed(rw)) { 1161171052Sattilio if (what & RA_NOTRECURSED) 1162171052Sattilio panic("Lock %s recursed @ %s:%d\n", 1163171052Sattilio rw->lock_object.lo_name, file, line); 1164171052Sattilio } else if (what & RA_RECURSED) 1165171052Sattilio panic("Lock %s not recursed @ %s:%d\n", 1166171052Sattilio rw->lock_object.lo_name, file, line); 1167154941Sjhb break; 1168154941Sjhb case RA_UNLOCKED: 1169154941Sjhb#ifdef WITNESS 1170167787Sjhb witness_assert(&rw->lock_object, what, file, line); 1171154941Sjhb#else 1172154941Sjhb /* 1173154941Sjhb * If we hold a write lock fail. We can't reliably check 1174154941Sjhb * to see if we hold a read lock or not. 1175154941Sjhb */ 1176157826Sjhb if (rw_wowner(rw) == curthread) 1177154941Sjhb panic("Lock %s exclusively locked @ %s:%d\n", 1178167787Sjhb rw->lock_object.lo_name, file, line); 1179154941Sjhb#endif 1180154941Sjhb break; 1181154941Sjhb default: 1182154941Sjhb panic("Unknown rw lock assertion: %d @ %s:%d", what, file, 1183154941Sjhb line); 1184154941Sjhb } 1185154941Sjhb} 1186154941Sjhb#endif /* INVARIANT_SUPPORT */ 1187154941Sjhb 1188154941Sjhb#ifdef DDB 1189154941Sjhbvoid 1190227588Spjddb_show_rwlock(const struct lock_object *lock) 1191154941Sjhb{ 1192227588Spjd const struct rwlock *rw; 1193154941Sjhb struct thread *td; 1194154941Sjhb 1195227588Spjd rw = (const struct rwlock *)lock; 1196154941Sjhb 1197154941Sjhb db_printf(" state: "); 1198154941Sjhb if (rw->rw_lock == RW_UNLOCKED) 1199154941Sjhb db_printf("UNLOCKED\n"); 1200169394Sjhb else if (rw->rw_lock == RW_DESTROYED) { 1201169394Sjhb db_printf("DESTROYED\n"); 1202169394Sjhb return; 1203169394Sjhb } else if (rw->rw_lock & RW_LOCK_READ) 1204167504Sjhb db_printf("RLOCK: %ju locks\n", 1205167504Sjhb (uintmax_t)(RW_READERS(rw->rw_lock))); 1206154941Sjhb else { 1207157826Sjhb td = rw_wowner(rw); 1208154941Sjhb db_printf("WLOCK: %p (tid %d, pid %d, \"%s\")\n", td, 1209173600Sjulian td->td_tid, td->td_proc->p_pid, td->td_name); 1210171052Sattilio if (rw_recursed(rw)) 1211171052Sattilio db_printf(" recursed: %u\n", rw->rw_recurse); 1212154941Sjhb } 1213154941Sjhb db_printf(" waiters: "); 1214154941Sjhb switch (rw->rw_lock & (RW_LOCK_READ_WAITERS | RW_LOCK_WRITE_WAITERS)) { 1215154941Sjhb case RW_LOCK_READ_WAITERS: 1216154941Sjhb db_printf("readers\n"); 1217154941Sjhb break; 1218154941Sjhb case RW_LOCK_WRITE_WAITERS: 1219154941Sjhb db_printf("writers\n"); 1220154941Sjhb break; 1221154941Sjhb case RW_LOCK_READ_WAITERS | RW_LOCK_WRITE_WAITERS: 1222167492Sjhb db_printf("readers and writers\n"); 1223154941Sjhb break; 1224154941Sjhb default: 1225154941Sjhb db_printf("none\n"); 1226154941Sjhb break; 1227154941Sjhb } 1228154941Sjhb} 1229154941Sjhb 1230154941Sjhb#endif 1231