kern_rwlock.c revision 285704
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 * 14154941Sjhb * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND 15154941Sjhb * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 16154941Sjhb * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 17154941Sjhb * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE 18154941Sjhb * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 19154941Sjhb * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 20154941Sjhb * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 21154941Sjhb * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 22154941Sjhb * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 23154941Sjhb * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 24154941Sjhb * SUCH DAMAGE. 25154941Sjhb */ 26154941Sjhb 27154941Sjhb/* 28154941Sjhb * Machine independent bits of reader/writer lock implementation. 29154941Sjhb */ 30154941Sjhb 31154941Sjhb#include <sys/cdefs.h> 32154941Sjhb__FBSDID("$FreeBSD: head/sys/kern/kern_rwlock.c 285704 2015-07-19 22:24:33Z markj $"); 33154941Sjhb 34154941Sjhb#include "opt_ddb.h" 35233628Sfabient#include "opt_hwpmc_hooks.h" 36167801Sjhb#include "opt_no_adaptive_rwlocks.h" 37154941Sjhb 38154941Sjhb#include <sys/param.h> 39244582Sattilio#include <sys/kdb.h> 40154941Sjhb#include <sys/ktr.h> 41177912Sjeff#include <sys/kernel.h> 42154941Sjhb#include <sys/lock.h> 43154941Sjhb#include <sys/mutex.h> 44154941Sjhb#include <sys/proc.h> 45154941Sjhb#include <sys/rwlock.h> 46274092Sjhb#include <sys/sched.h> 47177912Sjeff#include <sys/sysctl.h> 48154941Sjhb#include <sys/systm.h> 49154941Sjhb#include <sys/turnstile.h> 50171516Sattilio 51154941Sjhb#include <machine/cpu.h> 52154941Sjhb 53167801Sjhb#if defined(SMP) && !defined(NO_ADAPTIVE_RWLOCKS) 54167801Sjhb#define ADAPTIVE_RWLOCKS 55167801Sjhb#endif 56167801Sjhb 57233628Sfabient#ifdef HWPMC_HOOKS 58233628Sfabient#include <sys/pmckern.h> 59233628SfabientPMC_SOFT_DECLARE( , , lock, failed); 60233628Sfabient#endif 61233628Sfabient 62242515Sattilio/* 63242515Sattilio * Return the rwlock address when the lock cookie address is provided. 64242515Sattilio * This functionality assumes that struct rwlock* have a member named rw_lock. 65242515Sattilio */ 66242515Sattilio#define rwlock2rw(c) (__containerof(c, struct rwlock, rw_lock)) 67242515Sattilio 68177912Sjeff#ifdef ADAPTIVE_RWLOCKS 69177912Sjeffstatic int rowner_retries = 10; 70177912Sjeffstatic int rowner_loops = 10000; 71227309Sedstatic SYSCTL_NODE(_debug, OID_AUTO, rwlock, CTLFLAG_RD, NULL, 72227309Sed "rwlock debugging"); 73177912SjeffSYSCTL_INT(_debug_rwlock, OID_AUTO, retry, CTLFLAG_RW, &rowner_retries, 0, ""); 74177912SjeffSYSCTL_INT(_debug_rwlock, OID_AUTO, loops, CTLFLAG_RW, &rowner_loops, 0, ""); 75177912Sjeff#endif 76177912Sjeff 77154941Sjhb#ifdef DDB 78154941Sjhb#include <ddb/ddb.h> 79154941Sjhb 80227588Spjdstatic void db_show_rwlock(const struct lock_object *lock); 81154941Sjhb#endif 82227588Spjdstatic void assert_rw(const struct lock_object *lock, int what); 83255745Sdavidestatic void lock_rw(struct lock_object *lock, uintptr_t how); 84192853Ssson#ifdef KDTRACE_HOOKS 85227588Spjdstatic int owner_rw(const struct lock_object *lock, struct thread **owner); 86192853Ssson#endif 87255745Sdavidestatic uintptr_t unlock_rw(struct lock_object *lock); 88154941Sjhb 89154941Sjhbstruct lock_class lock_class_rw = { 90167365Sjhb .lc_name = "rw", 91167365Sjhb .lc_flags = LC_SLEEPLOCK | LC_RECURSABLE | LC_UPGRADABLE, 92173733Sattilio .lc_assert = assert_rw, 93154941Sjhb#ifdef DDB 94167365Sjhb .lc_ddb_show = db_show_rwlock, 95154941Sjhb#endif 96167368Sjhb .lc_lock = lock_rw, 97167368Sjhb .lc_unlock = unlock_rw, 98192853Ssson#ifdef KDTRACE_HOOKS 99192853Ssson .lc_owner = owner_rw, 100192853Ssson#endif 101154941Sjhb}; 102154941Sjhb 103157826Sjhb/* 104157826Sjhb * Return a pointer to the owning thread if the lock is write-locked or 105157826Sjhb * NULL if the lock is unlocked or read-locked. 106157826Sjhb */ 107157826Sjhb#define rw_wowner(rw) \ 108154941Sjhb ((rw)->rw_lock & RW_LOCK_READ ? NULL : \ 109154941Sjhb (struct thread *)RW_OWNER((rw)->rw_lock)) 110154941Sjhb 111157826Sjhb/* 112171052Sattilio * Returns if a write owner is recursed. Write ownership is not assured 113171052Sattilio * here and should be previously checked. 114171052Sattilio */ 115171052Sattilio#define rw_recursed(rw) ((rw)->rw_recurse != 0) 116171052Sattilio 117171052Sattilio/* 118171052Sattilio * Return true if curthread helds the lock. 119171052Sattilio */ 120171052Sattilio#define rw_wlocked(rw) (rw_wowner((rw)) == curthread) 121171052Sattilio 122171052Sattilio/* 123157826Sjhb * Return a pointer to the owning thread for this lock who should receive 124157826Sjhb * any priority lent by threads that block on this lock. Currently this 125157826Sjhb * is identical to rw_wowner(). 126157826Sjhb */ 127157826Sjhb#define rw_owner(rw) rw_wowner(rw) 128157826Sjhb 129154941Sjhb#ifndef INVARIANTS 130242515Sattilio#define __rw_assert(c, what, file, line) 131154941Sjhb#endif 132154941Sjhb 133154941Sjhbvoid 134227588Spjdassert_rw(const struct lock_object *lock, int what) 135173733Sattilio{ 136173733Sattilio 137227588Spjd rw_assert((const struct rwlock *)lock, what); 138173733Sattilio} 139173733Sattilio 140173733Sattiliovoid 141255745Sdavidelock_rw(struct lock_object *lock, uintptr_t how) 142167368Sjhb{ 143167368Sjhb struct rwlock *rw; 144167368Sjhb 145167368Sjhb rw = (struct rwlock *)lock; 146167368Sjhb if (how) 147255788Sdavide rw_rlock(rw); 148255788Sdavide else 149167368Sjhb rw_wlock(rw); 150167368Sjhb} 151167368Sjhb 152255745Sdavideuintptr_t 153167368Sjhbunlock_rw(struct lock_object *lock) 154167368Sjhb{ 155167368Sjhb struct rwlock *rw; 156167368Sjhb 157167368Sjhb rw = (struct rwlock *)lock; 158167368Sjhb rw_assert(rw, RA_LOCKED | LA_NOTRECURSED); 159167368Sjhb if (rw->rw_lock & RW_LOCK_READ) { 160167368Sjhb rw_runlock(rw); 161255788Sdavide return (1); 162167368Sjhb } else { 163167368Sjhb rw_wunlock(rw); 164255788Sdavide return (0); 165167368Sjhb } 166167368Sjhb} 167167368Sjhb 168192853Ssson#ifdef KDTRACE_HOOKS 169192853Sssonint 170227588Spjdowner_rw(const struct lock_object *lock, struct thread **owner) 171192853Ssson{ 172227588Spjd const struct rwlock *rw = (const struct rwlock *)lock; 173192853Ssson uintptr_t x = rw->rw_lock; 174192853Ssson 175192853Ssson *owner = rw_wowner(rw); 176192853Ssson return ((x & RW_LOCK_READ) != 0 ? (RW_READERS(x) != 0) : 177192853Ssson (*owner != NULL)); 178192853Ssson} 179192853Ssson#endif 180192853Ssson 181167368Sjhbvoid 182242515Sattilio_rw_init_flags(volatile uintptr_t *c, const char *name, int opts) 183154941Sjhb{ 184242515Sattilio struct rwlock *rw; 185171052Sattilio int flags; 186154941Sjhb 187242515Sattilio rw = rwlock2rw(c); 188242515Sattilio 189171052Sattilio MPASS((opts & ~(RW_DUPOK | RW_NOPROFILE | RW_NOWITNESS | RW_QUIET | 190275751Sdchagin RW_RECURSE | RW_NEW)) == 0); 191196334Sattilio ASSERT_ATOMIC_LOAD_PTR(rw->rw_lock, 192196334Sattilio ("%s: rw_lock not aligned for %s: %p", __func__, name, 193196334Sattilio &rw->rw_lock)); 194171052Sattilio 195193307Sattilio flags = LO_UPGRADABLE; 196171052Sattilio if (opts & RW_DUPOK) 197171052Sattilio flags |= LO_DUPOK; 198171052Sattilio if (opts & RW_NOPROFILE) 199171052Sattilio flags |= LO_NOPROFILE; 200171052Sattilio if (!(opts & RW_NOWITNESS)) 201171052Sattilio flags |= LO_WITNESS; 202193307Sattilio if (opts & RW_RECURSE) 203193307Sattilio flags |= LO_RECURSABLE; 204171052Sattilio if (opts & RW_QUIET) 205171052Sattilio flags |= LO_QUIET; 206275751Sdchagin if (opts & RW_NEW) 207275751Sdchagin flags |= LO_NEW; 208171052Sattilio 209252212Sjhb lock_init(&rw->lock_object, &lock_class_rw, name, NULL, flags); 210154941Sjhb rw->rw_lock = RW_UNLOCKED; 211171052Sattilio rw->rw_recurse = 0; 212154941Sjhb} 213154941Sjhb 214154941Sjhbvoid 215242515Sattilio_rw_destroy(volatile uintptr_t *c) 216154941Sjhb{ 217242515Sattilio struct rwlock *rw; 218154941Sjhb 219242515Sattilio rw = rwlock2rw(c); 220242515Sattilio 221205626Sbz KASSERT(rw->rw_lock == RW_UNLOCKED, ("rw lock %p not unlocked", rw)); 222205626Sbz KASSERT(rw->rw_recurse == 0, ("rw lock %p still recursed", rw)); 223169394Sjhb rw->rw_lock = RW_DESTROYED; 224167787Sjhb lock_destroy(&rw->lock_object); 225154941Sjhb} 226154941Sjhb 227154941Sjhbvoid 228154941Sjhbrw_sysinit(void *arg) 229154941Sjhb{ 230154941Sjhb struct rw_args *args = arg; 231154941Sjhb 232242515Sattilio rw_init((struct rwlock *)args->ra_rw, args->ra_desc); 233154941Sjhb} 234154941Sjhb 235185778Skmacyvoid 236185778Skmacyrw_sysinit_flags(void *arg) 237185778Skmacy{ 238185778Skmacy struct rw_args_flags *args = arg; 239185778Skmacy 240242515Sattilio rw_init_flags((struct rwlock *)args->ra_rw, args->ra_desc, 241242515Sattilio args->ra_flags); 242185778Skmacy} 243185778Skmacy 244167024Srwatsonint 245242515Sattilio_rw_wowned(const volatile uintptr_t *c) 246167024Srwatson{ 247167024Srwatson 248242515Sattilio return (rw_wowner(rwlock2rw(c)) == curthread); 249167024Srwatson} 250167024Srwatson 251154941Sjhbvoid 252242515Sattilio_rw_wlock_cookie(volatile uintptr_t *c, const char *file, int line) 253154941Sjhb{ 254242515Sattilio struct rwlock *rw; 255154941Sjhb 256228424Savg if (SCHEDULER_STOPPED()) 257228424Savg return; 258242515Sattilio 259242515Sattilio rw = rwlock2rw(c); 260242515Sattilio 261244582Sattilio KASSERT(kdb_active != 0 || !TD_IS_IDLETHREAD(curthread), 262240424Sattilio ("rw_wlock() by idle thread %p on rwlock %s @ %s:%d", 263240424Sattilio curthread, rw->lock_object.lo_name, file, line)); 264169394Sjhb KASSERT(rw->rw_lock != RW_DESTROYED, 265169394Sjhb ("rw_wlock() of destroyed rwlock @ %s:%d", file, line)); 266167787Sjhb WITNESS_CHECKORDER(&rw->lock_object, LOP_NEWORDER | LOP_EXCLUSIVE, file, 267182914Sjhb line, NULL); 268154941Sjhb __rw_wlock(rw, curthread, file, line); 269171052Sattilio LOCK_LOG_LOCK("WLOCK", &rw->lock_object, 0, rw->rw_recurse, file, line); 270167787Sjhb WITNESS_LOCK(&rw->lock_object, LOP_EXCLUSIVE, file, line); 271160771Sjhb curthread->td_locks++; 272154941Sjhb} 273154941Sjhb 274177843Sattilioint 275242515Sattilio__rw_try_wlock(volatile uintptr_t *c, const char *file, int line) 276177843Sattilio{ 277242515Sattilio struct rwlock *rw; 278177843Sattilio int rval; 279177843Sattilio 280228424Savg if (SCHEDULER_STOPPED()) 281228424Savg return (1); 282228424Savg 283242515Sattilio rw = rwlock2rw(c); 284242515Sattilio 285244582Sattilio KASSERT(kdb_active != 0 || !TD_IS_IDLETHREAD(curthread), 286240424Sattilio ("rw_try_wlock() by idle thread %p on rwlock %s @ %s:%d", 287240424Sattilio curthread, rw->lock_object.lo_name, file, line)); 288177843Sattilio KASSERT(rw->rw_lock != RW_DESTROYED, 289177843Sattilio ("rw_try_wlock() of destroyed rwlock @ %s:%d", file, line)); 290177843Sattilio 291193307Sattilio if (rw_wlocked(rw) && 292193307Sattilio (rw->lock_object.lo_flags & LO_RECURSABLE) != 0) { 293177843Sattilio rw->rw_recurse++; 294177843Sattilio rval = 1; 295177843Sattilio } else 296177843Sattilio rval = atomic_cmpset_acq_ptr(&rw->rw_lock, RW_UNLOCKED, 297177843Sattilio (uintptr_t)curthread); 298177843Sattilio 299177843Sattilio LOCK_LOG_TRY("WLOCK", &rw->lock_object, 0, rval, file, line); 300177843Sattilio if (rval) { 301177843Sattilio WITNESS_LOCK(&rw->lock_object, LOP_EXCLUSIVE | LOP_TRYLOCK, 302177843Sattilio file, line); 303284297Savg if (!rw_recursed(rw)) 304285704Smarkj LOCKSTAT_PROFILE_OBTAIN_RWLOCK_SUCCESS(rw__acquire, 305285704Smarkj rw, 0, 0, file, line, LOCKSTAT_WRITER); 306177843Sattilio curthread->td_locks++; 307177843Sattilio } 308177843Sattilio return (rval); 309177843Sattilio} 310177843Sattilio 311154941Sjhbvoid 312242515Sattilio_rw_wunlock_cookie(volatile uintptr_t *c, const char *file, int line) 313154941Sjhb{ 314242515Sattilio struct rwlock *rw; 315154941Sjhb 316228424Savg if (SCHEDULER_STOPPED()) 317228424Savg return; 318242515Sattilio 319242515Sattilio rw = rwlock2rw(c); 320242515Sattilio 321169394Sjhb KASSERT(rw->rw_lock != RW_DESTROYED, 322169394Sjhb ("rw_wunlock() of destroyed rwlock @ %s:%d", file, line)); 323242515Sattilio __rw_assert(c, RA_WLOCKED, file, line); 324167787Sjhb WITNESS_UNLOCK(&rw->lock_object, LOP_EXCLUSIVE, file, line); 325171052Sattilio LOCK_LOG_LOCK("WUNLOCK", &rw->lock_object, 0, rw->rw_recurse, file, 326171052Sattilio line); 327154941Sjhb __rw_wunlock(rw, curthread, file, line); 328252212Sjhb curthread->td_locks--; 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 358284297Savg uintptr_t state; 359192853Ssson uint64_t spin_cnt = 0; 360192853Ssson uint64_t sleep_cnt = 0; 361192853Ssson int64_t sleep_time = 0; 362284297Savg int64_t all_time = 0; 363192853Ssson#endif 364154941Sjhb 365228424Savg if (SCHEDULER_STOPPED()) 366228424Savg return; 367228424Savg 368242515Sattilio rw = rwlock2rw(c); 369242515Sattilio 370244582Sattilio KASSERT(kdb_active != 0 || !TD_IS_IDLETHREAD(curthread), 371240424Sattilio ("rw_rlock() by idle thread %p on rwlock %s @ %s:%d", 372240424Sattilio curthread, rw->lock_object.lo_name, file, line)); 373169394Sjhb KASSERT(rw->rw_lock != RW_DESTROYED, 374169394Sjhb ("rw_rlock() of destroyed rwlock @ %s:%d", file, line)); 375157826Sjhb KASSERT(rw_wowner(rw) != curthread, 376251323Sjhb ("rw_rlock: wlock already held for %s @ %s:%d", 377167787Sjhb rw->lock_object.lo_name, file, line)); 378182914Sjhb WITNESS_CHECKORDER(&rw->lock_object, LOP_NEWORDER, file, line, NULL); 379154941Sjhb 380284297Savg#ifdef KDTRACE_HOOKS 381285664Smarkj all_time -= lockstat_nsecs(&rw->lock_object); 382284297Savg state = rw->rw_lock; 383284297Savg#endif 384154941Sjhb for (;;) { 385192853Ssson#ifdef KDTRACE_HOOKS 386192853Ssson spin_cnt++; 387192853Ssson#endif 388154941Sjhb /* 389154941Sjhb * Handle the easy case. If no other thread has a write 390154941Sjhb * lock, then try to bump up the count of read locks. Note 391154941Sjhb * that we have to preserve the current state of the 392154941Sjhb * RW_LOCK_WRITE_WAITERS flag. If we fail to acquire a 393154941Sjhb * read lock, then rw_lock must have changed, so restart 394154941Sjhb * the loop. Note that this handles the case of a 395154941Sjhb * completely unlocked rwlock since such a lock is encoded 396154941Sjhb * as a read lock with no waiters. 397154941Sjhb */ 398176017Sjeff v = rw->rw_lock; 399176017Sjeff if (RW_CAN_READ(v)) { 400154941Sjhb /* 401154941Sjhb * The RW_LOCK_READ_WAITERS flag should only be set 402176017Sjeff * if the lock has been unlocked and write waiters 403176017Sjeff * were present. 404154941Sjhb */ 405176017Sjeff if (atomic_cmpset_acq_ptr(&rw->rw_lock, v, 406176017Sjeff v + RW_ONE_READER)) { 407167787Sjhb if (LOCK_LOG_TEST(&rw->lock_object, 0)) 408154941Sjhb CTR4(KTR_LOCK, 409154941Sjhb "%s: %p succeed %p -> %p", __func__, 410176017Sjeff rw, (void *)v, 411176017Sjeff (void *)(v + RW_ONE_READER)); 412154941Sjhb break; 413154941Sjhb } 414154941Sjhb continue; 415154941Sjhb } 416233628Sfabient#ifdef HWPMC_HOOKS 417233628Sfabient PMC_SOFT_CALL( , , lock, failed); 418233628Sfabient#endif 419174629Sjeff lock_profile_obtain_lock_failed(&rw->lock_object, 420174629Sjeff &contested, &waittime); 421154941Sjhb 422173960Sattilio#ifdef ADAPTIVE_RWLOCKS 423154941Sjhb /* 424173960Sattilio * If the owner is running on another CPU, spin until 425173960Sattilio * the owner stops running or the state of the lock 426173960Sattilio * changes. 427173960Sattilio */ 428176017Sjeff if ((v & RW_LOCK_READ) == 0) { 429176017Sjeff owner = (struct thread *)RW_OWNER(v); 430176017Sjeff if (TD_IS_RUNNING(owner)) { 431176017Sjeff if (LOCK_LOG_TEST(&rw->lock_object, 0)) 432176017Sjeff CTR3(KTR_LOCK, 433176017Sjeff "%s: spinning on %p held by %p", 434176017Sjeff __func__, rw, owner); 435274092Sjhb KTR_STATE1(KTR_SCHED, "thread", 436274092Sjhb sched_tdname(curthread), "spinning", 437274092Sjhb "lockname:\"%s\"", rw->lock_object.lo_name); 438176017Sjeff while ((struct thread*)RW_OWNER(rw->rw_lock) == 439192853Ssson owner && TD_IS_RUNNING(owner)) { 440176017Sjeff cpu_spinwait(); 441192853Ssson#ifdef KDTRACE_HOOKS 442192853Ssson spin_cnt++; 443192853Ssson#endif 444192853Ssson } 445274092Sjhb KTR_STATE0(KTR_SCHED, "thread", 446274092Sjhb sched_tdname(curthread), "running"); 447176017Sjeff continue; 448176017Sjeff } 449177912Sjeff } else if (spintries < rowner_retries) { 450177912Sjeff spintries++; 451274092Sjhb KTR_STATE1(KTR_SCHED, "thread", sched_tdname(curthread), 452274092Sjhb "spinning", "lockname:\"%s\"", 453274092Sjhb rw->lock_object.lo_name); 454177912Sjeff for (i = 0; i < rowner_loops; i++) { 455177912Sjeff v = rw->rw_lock; 456177912Sjeff if ((v & RW_LOCK_READ) == 0 || RW_CAN_READ(v)) 457177912Sjeff break; 458177912Sjeff cpu_spinwait(); 459177912Sjeff } 460259509Sattilio#ifdef KDTRACE_HOOKS 461259509Sattilio spin_cnt += rowner_loops - i; 462259509Sattilio#endif 463274092Sjhb KTR_STATE0(KTR_SCHED, "thread", sched_tdname(curthread), 464274092Sjhb "running"); 465177912Sjeff if (i != rowner_loops) 466177912Sjeff continue; 467173960Sattilio } 468173960Sattilio#endif 469173960Sattilio 470173960Sattilio /* 471154941Sjhb * Okay, now it's the hard case. Some other thread already 472176017Sjeff * has a write lock or there are write waiters present, 473176017Sjeff * acquire the turnstile lock so we can begin the process 474176017Sjeff * of blocking. 475154941Sjhb */ 476170295Sjeff ts = turnstile_trywait(&rw->lock_object); 477154941Sjhb 478154941Sjhb /* 479154941Sjhb * The lock might have been released while we spun, so 480176017Sjeff * recheck its state and restart the loop if needed. 481154941Sjhb */ 482176017Sjeff v = rw->rw_lock; 483176017Sjeff if (RW_CAN_READ(v)) { 484170295Sjeff turnstile_cancel(ts); 485154941Sjhb continue; 486154941Sjhb } 487154941Sjhb 488173960Sattilio#ifdef ADAPTIVE_RWLOCKS 489154941Sjhb /* 490193035Sjhb * The current lock owner might have started executing 491193035Sjhb * on another CPU (or the lock could have changed 492193035Sjhb * owners) while we were waiting on the turnstile 493193035Sjhb * chain lock. If so, drop the turnstile lock and try 494193035Sjhb * again. 495173960Sattilio */ 496176017Sjeff if ((v & RW_LOCK_READ) == 0) { 497176017Sjeff owner = (struct thread *)RW_OWNER(v); 498176017Sjeff if (TD_IS_RUNNING(owner)) { 499176017Sjeff turnstile_cancel(ts); 500176017Sjeff continue; 501176017Sjeff } 502173960Sattilio } 503173960Sattilio#endif 504173960Sattilio 505173960Sattilio /* 506176017Sjeff * The lock is held in write mode or it already has waiters. 507154941Sjhb */ 508176017Sjeff MPASS(!RW_CAN_READ(v)); 509176017Sjeff 510176017Sjeff /* 511176017Sjeff * If the RW_LOCK_READ_WAITERS flag is already set, then 512176017Sjeff * we can go ahead and block. If it is not set then try 513176017Sjeff * to set it. If we fail to set it drop the turnstile 514176017Sjeff * lock and restart the loop. 515176017Sjeff */ 516176017Sjeff if (!(v & RW_LOCK_READ_WAITERS)) { 517176017Sjeff if (!atomic_cmpset_ptr(&rw->rw_lock, v, 518176017Sjeff v | RW_LOCK_READ_WAITERS)) { 519170295Sjeff turnstile_cancel(ts); 520157826Sjhb continue; 521157826Sjhb } 522167787Sjhb if (LOCK_LOG_TEST(&rw->lock_object, 0)) 523157826Sjhb CTR2(KTR_LOCK, "%s: %p set read waiters flag", 524157826Sjhb __func__, rw); 525154941Sjhb } 526154941Sjhb 527154941Sjhb /* 528154941Sjhb * We were unable to acquire the lock and the read waiters 529154941Sjhb * flag is set, so we must block on the turnstile. 530154941Sjhb */ 531167787Sjhb if (LOCK_LOG_TEST(&rw->lock_object, 0)) 532154941Sjhb CTR2(KTR_LOCK, "%s: %p blocking on turnstile", __func__, 533154941Sjhb rw); 534192853Ssson#ifdef KDTRACE_HOOKS 535285664Smarkj sleep_time -= lockstat_nsecs(&rw->lock_object); 536192853Ssson#endif 537170295Sjeff turnstile_wait(ts, rw_owner(rw), TS_SHARED_QUEUE); 538192853Ssson#ifdef KDTRACE_HOOKS 539285664Smarkj sleep_time += lockstat_nsecs(&rw->lock_object); 540192853Ssson sleep_cnt++; 541192853Ssson#endif 542167787Sjhb if (LOCK_LOG_TEST(&rw->lock_object, 0)) 543154941Sjhb CTR2(KTR_LOCK, "%s: %p resuming from turnstile", 544154941Sjhb __func__, rw); 545154941Sjhb } 546284297Savg#ifdef KDTRACE_HOOKS 547285664Smarkj all_time += lockstat_nsecs(&rw->lock_object); 548284297Savg if (sleep_time) 549285703Smarkj LOCKSTAT_RECORD4(rw__block, rw, sleep_time, 550284297Savg LOCKSTAT_READER, (state & RW_LOCK_READ) == 0, 551284297Savg (state & RW_LOCK_READ) == 0 ? 0 : RW_READERS(state)); 552154941Sjhb 553284297Savg /* Record only the loops spinning and not sleeping. */ 554284297Savg if (spin_cnt > sleep_cnt) 555285703Smarkj LOCKSTAT_RECORD4(rw__spin, rw, all_time - sleep_time, 556284297Savg LOCKSTAT_READER, (state & RW_LOCK_READ) == 0, 557284297Savg (state & RW_LOCK_READ) == 0 ? 0 : RW_READERS(state)); 558284297Savg#endif 559154941Sjhb /* 560154941Sjhb * TODO: acquire "owner of record" here. Here be turnstile dragons 561154941Sjhb * however. turnstiles don't like owners changing between calls to 562154941Sjhb * turnstile_wait() currently. 563154941Sjhb */ 564285704Smarkj LOCKSTAT_PROFILE_OBTAIN_RWLOCK_SUCCESS(rw__acquire, rw, contested, 565285704Smarkj waittime, file, line, LOCKSTAT_READER); 566167787Sjhb LOCK_LOG_LOCK("RLOCK", &rw->lock_object, 0, 0, file, line); 567167787Sjhb WITNESS_LOCK(&rw->lock_object, 0, file, line); 568160771Sjhb curthread->td_locks++; 569176017Sjeff curthread->td_rw_rlocks++; 570154941Sjhb} 571154941Sjhb 572177843Sattilioint 573242515Sattilio__rw_try_rlock(volatile uintptr_t *c, const char *file, int line) 574177843Sattilio{ 575242515Sattilio struct rwlock *rw; 576177843Sattilio uintptr_t x; 577177843Sattilio 578228424Savg if (SCHEDULER_STOPPED()) 579228424Savg return (1); 580228424Savg 581242515Sattilio rw = rwlock2rw(c); 582242515Sattilio 583244582Sattilio KASSERT(kdb_active != 0 || !TD_IS_IDLETHREAD(curthread), 584240424Sattilio ("rw_try_rlock() by idle thread %p on rwlock %s @ %s:%d", 585240424Sattilio curthread, rw->lock_object.lo_name, file, line)); 586240424Sattilio 587177843Sattilio for (;;) { 588177843Sattilio x = rw->rw_lock; 589177843Sattilio KASSERT(rw->rw_lock != RW_DESTROYED, 590177843Sattilio ("rw_try_rlock() of destroyed rwlock @ %s:%d", file, line)); 591177843Sattilio if (!(x & RW_LOCK_READ)) 592177843Sattilio break; 593177843Sattilio if (atomic_cmpset_acq_ptr(&rw->rw_lock, x, x + RW_ONE_READER)) { 594177843Sattilio LOCK_LOG_TRY("RLOCK", &rw->lock_object, 0, 1, file, 595177843Sattilio line); 596177843Sattilio WITNESS_LOCK(&rw->lock_object, LOP_TRYLOCK, file, line); 597285704Smarkj LOCKSTAT_PROFILE_OBTAIN_RWLOCK_SUCCESS(rw__acquire, 598285704Smarkj rw, 0, 0, file, line, LOCKSTAT_READER); 599177843Sattilio curthread->td_locks++; 600177843Sattilio curthread->td_rw_rlocks++; 601177843Sattilio return (1); 602177843Sattilio } 603177843Sattilio } 604177843Sattilio 605177843Sattilio LOCK_LOG_TRY("RLOCK", &rw->lock_object, 0, 0, file, line); 606177843Sattilio return (0); 607177843Sattilio} 608177843Sattilio 609154941Sjhbvoid 610242515Sattilio_rw_runlock_cookie(volatile uintptr_t *c, const char *file, int line) 611154941Sjhb{ 612242515Sattilio struct rwlock *rw; 613154941Sjhb struct turnstile *ts; 614176017Sjeff uintptr_t x, v, queue; 615154941Sjhb 616228424Savg if (SCHEDULER_STOPPED()) 617228424Savg return; 618228424Savg 619242515Sattilio rw = rwlock2rw(c); 620242515Sattilio 621169394Sjhb KASSERT(rw->rw_lock != RW_DESTROYED, 622169394Sjhb ("rw_runlock() of destroyed rwlock @ %s:%d", file, line)); 623242515Sattilio __rw_assert(c, RA_RLOCKED, file, line); 624167787Sjhb WITNESS_UNLOCK(&rw->lock_object, 0, file, line); 625167787Sjhb LOCK_LOG_LOCK("RUNLOCK", &rw->lock_object, 0, 0, file, line); 626154941Sjhb 627154941Sjhb /* TODO: drop "owner of record" here. */ 628154941Sjhb 629154941Sjhb for (;;) { 630154941Sjhb /* 631154941Sjhb * See if there is more than one read lock held. If so, 632154941Sjhb * just drop one and return. 633154941Sjhb */ 634154941Sjhb x = rw->rw_lock; 635154941Sjhb if (RW_READERS(x) > 1) { 636197643Sattilio if (atomic_cmpset_rel_ptr(&rw->rw_lock, x, 637154941Sjhb x - RW_ONE_READER)) { 638167787Sjhb if (LOCK_LOG_TEST(&rw->lock_object, 0)) 639154941Sjhb CTR4(KTR_LOCK, 640154941Sjhb "%s: %p succeeded %p -> %p", 641154941Sjhb __func__, rw, (void *)x, 642154941Sjhb (void *)(x - RW_ONE_READER)); 643154941Sjhb break; 644154941Sjhb } 645154941Sjhb continue; 646167307Sjhb } 647154941Sjhb /* 648154941Sjhb * If there aren't any waiters for a write lock, then try 649154941Sjhb * to drop it quickly. 650154941Sjhb */ 651176017Sjeff if (!(x & RW_LOCK_WAITERS)) { 652176017Sjeff MPASS((x & ~RW_LOCK_WRITE_SPINNER) == 653176017Sjeff RW_READERS_LOCK(1)); 654197643Sattilio if (atomic_cmpset_rel_ptr(&rw->rw_lock, x, 655197643Sattilio RW_UNLOCKED)) { 656167787Sjhb if (LOCK_LOG_TEST(&rw->lock_object, 0)) 657154941Sjhb CTR2(KTR_LOCK, "%s: %p last succeeded", 658154941Sjhb __func__, rw); 659154941Sjhb break; 660154941Sjhb } 661154941Sjhb continue; 662154941Sjhb } 663154941Sjhb /* 664176017Sjeff * Ok, we know we have waiters and we think we are the 665176017Sjeff * last reader, so grab the turnstile lock. 666154941Sjhb */ 667170295Sjeff turnstile_chain_lock(&rw->lock_object); 668176017Sjeff v = rw->rw_lock & (RW_LOCK_WAITERS | RW_LOCK_WRITE_SPINNER); 669176017Sjeff MPASS(v & RW_LOCK_WAITERS); 670154941Sjhb 671154941Sjhb /* 672154941Sjhb * Try to drop our lock leaving the lock in a unlocked 673154941Sjhb * state. 674154941Sjhb * 675154941Sjhb * If you wanted to do explicit lock handoff you'd have to 676154941Sjhb * do it here. You'd also want to use turnstile_signal() 677154941Sjhb * and you'd have to handle the race where a higher 678154941Sjhb * priority thread blocks on the write lock before the 679154941Sjhb * thread you wakeup actually runs and have the new thread 680154941Sjhb * "steal" the lock. For now it's a lot simpler to just 681154941Sjhb * wakeup all of the waiters. 682154941Sjhb * 683154941Sjhb * As above, if we fail, then another thread might have 684154941Sjhb * acquired a read lock, so drop the turnstile lock and 685154941Sjhb * restart. 686154941Sjhb */ 687176017Sjeff x = RW_UNLOCKED; 688176017Sjeff if (v & RW_LOCK_WRITE_WAITERS) { 689176017Sjeff queue = TS_EXCLUSIVE_QUEUE; 690176017Sjeff x |= (v & RW_LOCK_READ_WAITERS); 691176017Sjeff } else 692176017Sjeff queue = TS_SHARED_QUEUE; 693197643Sattilio if (!atomic_cmpset_rel_ptr(&rw->rw_lock, RW_READERS_LOCK(1) | v, 694176017Sjeff x)) { 695170295Sjeff turnstile_chain_unlock(&rw->lock_object); 696154941Sjhb continue; 697154941Sjhb } 698167787Sjhb if (LOCK_LOG_TEST(&rw->lock_object, 0)) 699154941Sjhb CTR2(KTR_LOCK, "%s: %p last succeeded with waiters", 700154941Sjhb __func__, rw); 701154941Sjhb 702154941Sjhb /* 703154941Sjhb * Ok. The lock is released and all that's left is to 704154941Sjhb * wake up the waiters. Note that the lock might not be 705154941Sjhb * free anymore, but in that case the writers will just 706154941Sjhb * block again if they run before the new lock holder(s) 707154941Sjhb * release the lock. 708154941Sjhb */ 709167787Sjhb ts = turnstile_lookup(&rw->lock_object); 710157846Sjhb MPASS(ts != NULL); 711176017Sjeff turnstile_broadcast(ts, queue); 712154941Sjhb turnstile_unpend(ts, TS_SHARED_LOCK); 713170295Sjeff turnstile_chain_unlock(&rw->lock_object); 714154941Sjhb break; 715154941Sjhb } 716285704Smarkj LOCKSTAT_PROFILE_RELEASE_RWLOCK(rw__release, rw, LOCKSTAT_READER); 717252212Sjhb curthread->td_locks--; 718252212Sjhb curthread->td_rw_rlocks--; 719154941Sjhb} 720154941Sjhb 721154941Sjhb/* 722154941Sjhb * This function is called when we are unable to obtain a write lock on the 723154941Sjhb * first try. This means that at least one other thread holds either a 724154941Sjhb * read or write lock. 725154941Sjhb */ 726154941Sjhbvoid 727242515Sattilio__rw_wlock_hard(volatile uintptr_t *c, uintptr_t tid, const char *file, 728242515Sattilio int line) 729154941Sjhb{ 730242515Sattilio struct rwlock *rw; 731170295Sjeff struct turnstile *ts; 732167801Sjhb#ifdef ADAPTIVE_RWLOCKS 733157846Sjhb volatile struct thread *owner; 734176017Sjeff int spintries = 0; 735176017Sjeff int i; 736157851Swkoszek#endif 737189846Sjeff uintptr_t v, x; 738189846Sjeff#ifdef LOCK_PROFILING 739171516Sattilio uint64_t waittime = 0; 740171516Sattilio int contested = 0; 741189846Sjeff#endif 742192853Ssson#ifdef KDTRACE_HOOKS 743284297Savg uintptr_t state; 744192853Ssson uint64_t spin_cnt = 0; 745192853Ssson uint64_t sleep_cnt = 0; 746192853Ssson int64_t sleep_time = 0; 747284297Savg int64_t all_time = 0; 748192853Ssson#endif 749154941Sjhb 750228424Savg if (SCHEDULER_STOPPED()) 751228424Savg return; 752228424Savg 753242515Sattilio rw = rwlock2rw(c); 754242515Sattilio 755171052Sattilio if (rw_wlocked(rw)) { 756193307Sattilio KASSERT(rw->lock_object.lo_flags & LO_RECURSABLE, 757171052Sattilio ("%s: recursing but non-recursive rw %s @ %s:%d\n", 758171052Sattilio __func__, rw->lock_object.lo_name, file, line)); 759171052Sattilio rw->rw_recurse++; 760171052Sattilio if (LOCK_LOG_TEST(&rw->lock_object, 0)) 761171052Sattilio CTR2(KTR_LOCK, "%s: %p recursing", __func__, rw); 762171052Sattilio return; 763171052Sattilio } 764171052Sattilio 765167787Sjhb if (LOCK_LOG_TEST(&rw->lock_object, 0)) 766154941Sjhb CTR5(KTR_LOCK, "%s: %s contested (lock=%p) at %s:%d", __func__, 767167787Sjhb rw->lock_object.lo_name, (void *)rw->rw_lock, file, line); 768154941Sjhb 769284297Savg#ifdef KDTRACE_HOOKS 770285664Smarkj all_time -= lockstat_nsecs(&rw->lock_object); 771284297Savg state = rw->rw_lock; 772284297Savg#endif 773154941Sjhb while (!_rw_write_lock(rw, tid)) { 774192853Ssson#ifdef KDTRACE_HOOKS 775192853Ssson spin_cnt++; 776192853Ssson#endif 777233628Sfabient#ifdef HWPMC_HOOKS 778233628Sfabient PMC_SOFT_CALL( , , lock, failed); 779233628Sfabient#endif 780174629Sjeff lock_profile_obtain_lock_failed(&rw->lock_object, 781174629Sjeff &contested, &waittime); 782173960Sattilio#ifdef ADAPTIVE_RWLOCKS 783173960Sattilio /* 784173960Sattilio * If the lock is write locked and the owner is 785173960Sattilio * running on another CPU, spin until the owner stops 786173960Sattilio * running or the state of the lock changes. 787173960Sattilio */ 788173960Sattilio v = rw->rw_lock; 789173960Sattilio owner = (struct thread *)RW_OWNER(v); 790173960Sattilio if (!(v & RW_LOCK_READ) && TD_IS_RUNNING(owner)) { 791173960Sattilio if (LOCK_LOG_TEST(&rw->lock_object, 0)) 792173960Sattilio CTR3(KTR_LOCK, "%s: spinning on %p held by %p", 793173960Sattilio __func__, rw, owner); 794274092Sjhb KTR_STATE1(KTR_SCHED, "thread", sched_tdname(curthread), 795274092Sjhb "spinning", "lockname:\"%s\"", 796274092Sjhb rw->lock_object.lo_name); 797173960Sattilio while ((struct thread*)RW_OWNER(rw->rw_lock) == owner && 798192853Ssson TD_IS_RUNNING(owner)) { 799173960Sattilio cpu_spinwait(); 800192853Ssson#ifdef KDTRACE_HOOKS 801192853Ssson spin_cnt++; 802192853Ssson#endif 803192853Ssson } 804274092Sjhb KTR_STATE0(KTR_SCHED, "thread", sched_tdname(curthread), 805274092Sjhb "running"); 806173960Sattilio continue; 807173960Sattilio } 808177912Sjeff if ((v & RW_LOCK_READ) && RW_READERS(v) && 809177912Sjeff spintries < rowner_retries) { 810176017Sjeff if (!(v & RW_LOCK_WRITE_SPINNER)) { 811176017Sjeff if (!atomic_cmpset_ptr(&rw->rw_lock, v, 812176017Sjeff v | RW_LOCK_WRITE_SPINNER)) { 813176017Sjeff continue; 814176017Sjeff } 815176017Sjeff } 816176017Sjeff spintries++; 817274092Sjhb KTR_STATE1(KTR_SCHED, "thread", sched_tdname(curthread), 818274092Sjhb "spinning", "lockname:\"%s\"", 819274092Sjhb rw->lock_object.lo_name); 820177912Sjeff for (i = 0; i < rowner_loops; i++) { 821176017Sjeff if ((rw->rw_lock & RW_LOCK_WRITE_SPINNER) == 0) 822176017Sjeff break; 823176017Sjeff cpu_spinwait(); 824176017Sjeff } 825274092Sjhb KTR_STATE0(KTR_SCHED, "thread", sched_tdname(curthread), 826274092Sjhb "running"); 827192853Ssson#ifdef KDTRACE_HOOKS 828192853Ssson spin_cnt += rowner_loops - i; 829192853Ssson#endif 830177912Sjeff if (i != rowner_loops) 831176017Sjeff continue; 832176017Sjeff } 833173960Sattilio#endif 834170295Sjeff ts = turnstile_trywait(&rw->lock_object); 835154941Sjhb v = rw->rw_lock; 836154941Sjhb 837173960Sattilio#ifdef ADAPTIVE_RWLOCKS 838154941Sjhb /* 839193035Sjhb * The current lock owner might have started executing 840193035Sjhb * on another CPU (or the lock could have changed 841193035Sjhb * owners) while we were waiting on the turnstile 842193035Sjhb * chain lock. If so, drop the turnstile lock and try 843193035Sjhb * again. 844173960Sattilio */ 845173960Sattilio if (!(v & RW_LOCK_READ)) { 846173960Sattilio owner = (struct thread *)RW_OWNER(v); 847173960Sattilio if (TD_IS_RUNNING(owner)) { 848173960Sattilio turnstile_cancel(ts); 849173960Sattilio continue; 850173960Sattilio } 851173960Sattilio } 852173960Sattilio#endif 853173960Sattilio /* 854179334Sattilio * Check for the waiters flags about this rwlock. 855179334Sattilio * If the lock was released, without maintain any pending 856179334Sattilio * waiters queue, simply try to acquire it. 857179334Sattilio * If a pending waiters queue is present, claim the lock 858179334Sattilio * ownership and maintain the pending queue. 859154941Sjhb */ 860176017Sjeff x = v & (RW_LOCK_WAITERS | RW_LOCK_WRITE_SPINNER); 861176017Sjeff if ((v & ~x) == RW_UNLOCKED) { 862176017Sjeff x &= ~RW_LOCK_WRITE_SPINNER; 863176017Sjeff if (atomic_cmpset_acq_ptr(&rw->rw_lock, v, tid | x)) { 864176017Sjeff if (x) 865176017Sjeff turnstile_claim(ts); 866176017Sjeff else 867176017Sjeff turnstile_cancel(ts); 868154941Sjhb break; 869154941Sjhb } 870170295Sjeff turnstile_cancel(ts); 871154941Sjhb continue; 872154941Sjhb } 873154941Sjhb /* 874154941Sjhb * If the RW_LOCK_WRITE_WAITERS flag isn't set, then try to 875154941Sjhb * set it. If we fail to set it, then loop back and try 876154941Sjhb * again. 877154941Sjhb */ 878157826Sjhb if (!(v & RW_LOCK_WRITE_WAITERS)) { 879157826Sjhb if (!atomic_cmpset_ptr(&rw->rw_lock, v, 880157826Sjhb v | RW_LOCK_WRITE_WAITERS)) { 881170295Sjeff turnstile_cancel(ts); 882157826Sjhb continue; 883157826Sjhb } 884167787Sjhb if (LOCK_LOG_TEST(&rw->lock_object, 0)) 885157826Sjhb CTR2(KTR_LOCK, "%s: %p set write waiters flag", 886157826Sjhb __func__, rw); 887154941Sjhb } 888157846Sjhb /* 889154941Sjhb * We were unable to acquire the lock and the write waiters 890154941Sjhb * flag is set, so we must block on the turnstile. 891154941Sjhb */ 892167787Sjhb if (LOCK_LOG_TEST(&rw->lock_object, 0)) 893154941Sjhb CTR2(KTR_LOCK, "%s: %p blocking on turnstile", __func__, 894154941Sjhb rw); 895192853Ssson#ifdef KDTRACE_HOOKS 896285664Smarkj sleep_time -= lockstat_nsecs(&rw->lock_object); 897192853Ssson#endif 898170295Sjeff turnstile_wait(ts, rw_owner(rw), TS_EXCLUSIVE_QUEUE); 899192853Ssson#ifdef KDTRACE_HOOKS 900285664Smarkj sleep_time += lockstat_nsecs(&rw->lock_object); 901192853Ssson sleep_cnt++; 902192853Ssson#endif 903167787Sjhb if (LOCK_LOG_TEST(&rw->lock_object, 0)) 904154941Sjhb CTR2(KTR_LOCK, "%s: %p resuming from turnstile", 905154941Sjhb __func__, rw); 906176017Sjeff#ifdef ADAPTIVE_RWLOCKS 907176017Sjeff spintries = 0; 908176017Sjeff#endif 909154941Sjhb } 910192853Ssson#ifdef KDTRACE_HOOKS 911285664Smarkj all_time += lockstat_nsecs(&rw->lock_object); 912192853Ssson if (sleep_time) 913285703Smarkj LOCKSTAT_RECORD4(rw__block, rw, sleep_time, 914284297Savg LOCKSTAT_WRITER, (state & RW_LOCK_READ) == 0, 915284297Savg (state & RW_LOCK_READ) == 0 ? 0 : RW_READERS(state)); 916192853Ssson 917284297Savg /* Record only the loops spinning and not sleeping. */ 918192853Ssson if (spin_cnt > sleep_cnt) 919285703Smarkj LOCKSTAT_RECORD4(rw__spin, rw, all_time - sleep_time, 920284297Savg LOCKSTAT_READER, (state & RW_LOCK_READ) == 0, 921284297Savg (state & RW_LOCK_READ) == 0 ? 0 : RW_READERS(state)); 922192853Ssson#endif 923285704Smarkj LOCKSTAT_PROFILE_OBTAIN_RWLOCK_SUCCESS(rw__acquire, rw, contested, 924285704Smarkj waittime, file, line, LOCKSTAT_WRITER); 925154941Sjhb} 926154941Sjhb 927154941Sjhb/* 928154941Sjhb * This function is called if the first try at releasing a write lock failed. 929154941Sjhb * This means that one of the 2 waiter bits must be set indicating that at 930154941Sjhb * least one thread is waiting on this lock. 931154941Sjhb */ 932154941Sjhbvoid 933242515Sattilio__rw_wunlock_hard(volatile uintptr_t *c, uintptr_t tid, const char *file, 934242515Sattilio int line) 935154941Sjhb{ 936242515Sattilio struct rwlock *rw; 937154941Sjhb struct turnstile *ts; 938154941Sjhb uintptr_t v; 939154941Sjhb int queue; 940154941Sjhb 941228424Savg if (SCHEDULER_STOPPED()) 942228424Savg return; 943228424Savg 944242515Sattilio rw = rwlock2rw(c); 945242515Sattilio 946171052Sattilio if (rw_wlocked(rw) && rw_recursed(rw)) { 947176017Sjeff rw->rw_recurse--; 948171052Sattilio if (LOCK_LOG_TEST(&rw->lock_object, 0)) 949171052Sattilio CTR2(KTR_LOCK, "%s: %p unrecursing", __func__, rw); 950171052Sattilio return; 951171052Sattilio } 952171052Sattilio 953154941Sjhb KASSERT(rw->rw_lock & (RW_LOCK_READ_WAITERS | RW_LOCK_WRITE_WAITERS), 954154941Sjhb ("%s: neither of the waiter flags are set", __func__)); 955154941Sjhb 956167787Sjhb if (LOCK_LOG_TEST(&rw->lock_object, 0)) 957154941Sjhb CTR2(KTR_LOCK, "%s: %p contested", __func__, rw); 958154941Sjhb 959170295Sjeff turnstile_chain_lock(&rw->lock_object); 960167787Sjhb ts = turnstile_lookup(&rw->lock_object); 961154941Sjhb MPASS(ts != NULL); 962154941Sjhb 963154941Sjhb /* 964154941Sjhb * Use the same algo as sx locks for now. Prefer waking up shared 965154941Sjhb * waiters if we have any over writers. This is probably not ideal. 966154941Sjhb * 967154941Sjhb * 'v' is the value we are going to write back to rw_lock. If we 968154941Sjhb * have waiters on both queues, we need to preserve the state of 969154941Sjhb * the waiter flag for the queue we don't wake up. For now this is 970154941Sjhb * hardcoded for the algorithm mentioned above. 971154941Sjhb * 972154941Sjhb * In the case of both readers and writers waiting we wakeup the 973154941Sjhb * readers but leave the RW_LOCK_WRITE_WAITERS flag set. If a 974154941Sjhb * new writer comes in before a reader it will claim the lock up 975154941Sjhb * above. There is probably a potential priority inversion in 976154941Sjhb * there that could be worked around either by waking both queues 977154941Sjhb * of waiters or doing some complicated lock handoff gymnastics. 978154941Sjhb */ 979157846Sjhb v = RW_UNLOCKED; 980176076Sjeff if (rw->rw_lock & RW_LOCK_WRITE_WAITERS) { 981176076Sjeff queue = TS_EXCLUSIVE_QUEUE; 982176076Sjeff v |= (rw->rw_lock & RW_LOCK_READ_WAITERS); 983176076Sjeff } else 984154941Sjhb queue = TS_SHARED_QUEUE; 985157846Sjhb 986157846Sjhb /* Wake up all waiters for the specific queue. */ 987167787Sjhb if (LOCK_LOG_TEST(&rw->lock_object, 0)) 988154941Sjhb CTR3(KTR_LOCK, "%s: %p waking up %s waiters", __func__, rw, 989154941Sjhb queue == TS_SHARED_QUEUE ? "read" : "write"); 990154941Sjhb turnstile_broadcast(ts, queue); 991154941Sjhb atomic_store_rel_ptr(&rw->rw_lock, v); 992154941Sjhb turnstile_unpend(ts, TS_EXCLUSIVE_LOCK); 993170295Sjeff turnstile_chain_unlock(&rw->lock_object); 994154941Sjhb} 995154941Sjhb 996157882Sjhb/* 997157882Sjhb * Attempt to do a non-blocking upgrade from a read lock to a write 998157882Sjhb * lock. This will only succeed if this thread holds a single read 999157882Sjhb * lock. Returns true if the upgrade succeeded and false otherwise. 1000157882Sjhb */ 1001157882Sjhbint 1002242515Sattilio__rw_try_upgrade(volatile uintptr_t *c, const char *file, int line) 1003157882Sjhb{ 1004242515Sattilio struct rwlock *rw; 1005176017Sjeff uintptr_t v, x, tid; 1006170295Sjeff struct turnstile *ts; 1007157882Sjhb int success; 1008157882Sjhb 1009228424Savg if (SCHEDULER_STOPPED()) 1010228424Savg return (1); 1011228424Savg 1012242515Sattilio rw = rwlock2rw(c); 1013242515Sattilio 1014169394Sjhb KASSERT(rw->rw_lock != RW_DESTROYED, 1015169394Sjhb ("rw_try_upgrade() of destroyed rwlock @ %s:%d", file, line)); 1016242515Sattilio __rw_assert(c, RA_RLOCKED, file, line); 1017157882Sjhb 1018157882Sjhb /* 1019157882Sjhb * Attempt to switch from one reader to a writer. If there 1020157882Sjhb * are any write waiters, then we will have to lock the 1021157882Sjhb * turnstile first to prevent races with another writer 1022157882Sjhb * calling turnstile_wait() before we have claimed this 1023157882Sjhb * turnstile. So, do the simple case of no waiters first. 1024157882Sjhb */ 1025157882Sjhb tid = (uintptr_t)curthread; 1026176017Sjeff success = 0; 1027176017Sjeff for (;;) { 1028176017Sjeff v = rw->rw_lock; 1029176017Sjeff if (RW_READERS(v) > 1) 1030176017Sjeff break; 1031176017Sjeff if (!(v & RW_LOCK_WAITERS)) { 1032176017Sjeff success = atomic_cmpset_ptr(&rw->rw_lock, v, tid); 1033176017Sjeff if (!success) 1034176017Sjeff continue; 1035176017Sjeff break; 1036176017Sjeff } 1037157882Sjhb 1038176017Sjeff /* 1039176017Sjeff * Ok, we think we have waiters, so lock the turnstile. 1040176017Sjeff */ 1041176017Sjeff ts = turnstile_trywait(&rw->lock_object); 1042176017Sjeff v = rw->rw_lock; 1043176017Sjeff if (RW_READERS(v) > 1) { 1044176017Sjeff turnstile_cancel(ts); 1045176017Sjeff break; 1046176017Sjeff } 1047176017Sjeff /* 1048176017Sjeff * Try to switch from one reader to a writer again. This time 1049176017Sjeff * we honor the current state of the waiters flags. 1050176017Sjeff * If we obtain the lock with the flags set, then claim 1051176017Sjeff * ownership of the turnstile. 1052176017Sjeff */ 1053176017Sjeff x = rw->rw_lock & RW_LOCK_WAITERS; 1054176017Sjeff success = atomic_cmpset_ptr(&rw->rw_lock, v, tid | x); 1055176017Sjeff if (success) { 1056176017Sjeff if (x) 1057176017Sjeff turnstile_claim(ts); 1058176017Sjeff else 1059176017Sjeff turnstile_cancel(ts); 1060176017Sjeff break; 1061176017Sjeff } 1062170295Sjeff turnstile_cancel(ts); 1063176017Sjeff } 1064167787Sjhb LOCK_LOG_TRY("WUPGRADE", &rw->lock_object, 0, success, file, line); 1065176017Sjeff if (success) { 1066176017Sjeff curthread->td_rw_rlocks--; 1067167787Sjhb WITNESS_UPGRADE(&rw->lock_object, LOP_EXCLUSIVE | LOP_TRYLOCK, 1068157882Sjhb file, line); 1069285703Smarkj LOCKSTAT_RECORD0(rw__upgrade, rw); 1070176017Sjeff } 1071157882Sjhb return (success); 1072157882Sjhb} 1073157882Sjhb 1074157882Sjhb/* 1075157882Sjhb * Downgrade a write lock into a single read lock. 1076157882Sjhb */ 1077157882Sjhbvoid 1078242515Sattilio__rw_downgrade(volatile uintptr_t *c, const char *file, int line) 1079157882Sjhb{ 1080242515Sattilio struct rwlock *rw; 1081157882Sjhb struct turnstile *ts; 1082157882Sjhb uintptr_t tid, v; 1083176017Sjeff int rwait, wwait; 1084157882Sjhb 1085228424Savg if (SCHEDULER_STOPPED()) 1086228424Savg return; 1087228424Savg 1088242515Sattilio rw = rwlock2rw(c); 1089242515Sattilio 1090169394Sjhb KASSERT(rw->rw_lock != RW_DESTROYED, 1091169394Sjhb ("rw_downgrade() of destroyed rwlock @ %s:%d", file, line)); 1092242515Sattilio __rw_assert(c, RA_WLOCKED | RA_NOTRECURSED, file, line); 1093171052Sattilio#ifndef INVARIANTS 1094171052Sattilio if (rw_recursed(rw)) 1095171052Sattilio panic("downgrade of a recursed lock"); 1096171052Sattilio#endif 1097157882Sjhb 1098167787Sjhb WITNESS_DOWNGRADE(&rw->lock_object, 0, file, line); 1099157882Sjhb 1100157882Sjhb /* 1101157882Sjhb * Convert from a writer to a single reader. First we handle 1102157882Sjhb * the easy case with no waiters. If there are any waiters, we 1103176017Sjeff * lock the turnstile and "disown" the lock. 1104157882Sjhb */ 1105157882Sjhb tid = (uintptr_t)curthread; 1106157882Sjhb if (atomic_cmpset_rel_ptr(&rw->rw_lock, tid, RW_READERS_LOCK(1))) 1107157882Sjhb goto out; 1108157882Sjhb 1109157882Sjhb /* 1110157882Sjhb * Ok, we think we have waiters, so lock the turnstile so we can 1111157882Sjhb * read the waiter flags without any races. 1112157882Sjhb */ 1113170295Sjeff turnstile_chain_lock(&rw->lock_object); 1114176017Sjeff v = rw->rw_lock & RW_LOCK_WAITERS; 1115176017Sjeff rwait = v & RW_LOCK_READ_WAITERS; 1116176017Sjeff wwait = v & RW_LOCK_WRITE_WAITERS; 1117176017Sjeff MPASS(rwait | wwait); 1118157882Sjhb 1119157882Sjhb /* 1120176017Sjeff * Downgrade from a write lock while preserving waiters flag 1121176017Sjeff * and give up ownership of the turnstile. 1122157882Sjhb */ 1123167787Sjhb ts = turnstile_lookup(&rw->lock_object); 1124157882Sjhb MPASS(ts != NULL); 1125176017Sjeff if (!wwait) 1126176017Sjeff v &= ~RW_LOCK_READ_WAITERS; 1127176017Sjeff atomic_store_rel_ptr(&rw->rw_lock, RW_READERS_LOCK(1) | v); 1128176017Sjeff /* 1129176017Sjeff * Wake other readers if there are no writers pending. Otherwise they 1130176017Sjeff * won't be able to acquire the lock anyway. 1131176017Sjeff */ 1132176017Sjeff if (rwait && !wwait) { 1133157882Sjhb turnstile_broadcast(ts, TS_SHARED_QUEUE); 1134157882Sjhb turnstile_unpend(ts, TS_EXCLUSIVE_LOCK); 1135176017Sjeff } else 1136157882Sjhb turnstile_disown(ts); 1137170295Sjeff turnstile_chain_unlock(&rw->lock_object); 1138157882Sjhbout: 1139176017Sjeff curthread->td_rw_rlocks++; 1140167787Sjhb LOCK_LOG_LOCK("WDOWNGRADE", &rw->lock_object, 0, 0, file, line); 1141285703Smarkj LOCKSTAT_RECORD0(rw__downgrade, rw); 1142157882Sjhb} 1143157882Sjhb 1144154941Sjhb#ifdef INVARIANT_SUPPORT 1145155162Sscottl#ifndef INVARIANTS 1146242515Sattilio#undef __rw_assert 1147154941Sjhb#endif 1148154941Sjhb 1149154941Sjhb/* 1150154941Sjhb * In the non-WITNESS case, rw_assert() can only detect that at least 1151154941Sjhb * *some* thread owns an rlock, but it cannot guarantee that *this* 1152154941Sjhb * thread owns an rlock. 1153154941Sjhb */ 1154154941Sjhbvoid 1155242515Sattilio__rw_assert(const volatile uintptr_t *c, int what, const char *file, int line) 1156154941Sjhb{ 1157242515Sattilio const struct rwlock *rw; 1158154941Sjhb 1159154941Sjhb if (panicstr != NULL) 1160154941Sjhb return; 1161242515Sattilio 1162242515Sattilio rw = rwlock2rw(c); 1163242515Sattilio 1164154941Sjhb switch (what) { 1165154941Sjhb case RA_LOCKED: 1166171052Sattilio case RA_LOCKED | RA_RECURSED: 1167171052Sattilio case RA_LOCKED | RA_NOTRECURSED: 1168154941Sjhb case RA_RLOCKED: 1169251323Sjhb case RA_RLOCKED | RA_RECURSED: 1170251323Sjhb case RA_RLOCKED | RA_NOTRECURSED: 1171154941Sjhb#ifdef WITNESS 1172167787Sjhb witness_assert(&rw->lock_object, what, file, line); 1173154941Sjhb#else 1174154941Sjhb /* 1175154941Sjhb * If some other thread has a write lock or we have one 1176154941Sjhb * and are asserting a read lock, fail. Also, if no one 1177154941Sjhb * has a lock at all, fail. 1178154941Sjhb */ 1179155061Sscottl if (rw->rw_lock == RW_UNLOCKED || 1180251323Sjhb (!(rw->rw_lock & RW_LOCK_READ) && (what & RA_RLOCKED || 1181157826Sjhb rw_wowner(rw) != curthread))) 1182154941Sjhb panic("Lock %s not %slocked @ %s:%d\n", 1183251323Sjhb rw->lock_object.lo_name, (what & RA_RLOCKED) ? 1184154941Sjhb "read " : "", file, line); 1185171052Sattilio 1186251323Sjhb if (!(rw->rw_lock & RW_LOCK_READ) && !(what & RA_RLOCKED)) { 1187171052Sattilio if (rw_recursed(rw)) { 1188171052Sattilio if (what & RA_NOTRECURSED) 1189171052Sattilio panic("Lock %s recursed @ %s:%d\n", 1190171052Sattilio rw->lock_object.lo_name, file, 1191171052Sattilio line); 1192171052Sattilio } else if (what & RA_RECURSED) 1193171052Sattilio panic("Lock %s not recursed @ %s:%d\n", 1194171052Sattilio rw->lock_object.lo_name, file, line); 1195171052Sattilio } 1196154941Sjhb#endif 1197154941Sjhb break; 1198154941Sjhb case RA_WLOCKED: 1199171052Sattilio case RA_WLOCKED | RA_RECURSED: 1200171052Sattilio case RA_WLOCKED | RA_NOTRECURSED: 1201157826Sjhb if (rw_wowner(rw) != curthread) 1202154941Sjhb panic("Lock %s not exclusively locked @ %s:%d\n", 1203167787Sjhb rw->lock_object.lo_name, file, line); 1204171052Sattilio if (rw_recursed(rw)) { 1205171052Sattilio if (what & RA_NOTRECURSED) 1206171052Sattilio panic("Lock %s recursed @ %s:%d\n", 1207171052Sattilio rw->lock_object.lo_name, file, line); 1208171052Sattilio } else if (what & RA_RECURSED) 1209171052Sattilio panic("Lock %s not recursed @ %s:%d\n", 1210171052Sattilio rw->lock_object.lo_name, file, line); 1211154941Sjhb break; 1212154941Sjhb case RA_UNLOCKED: 1213154941Sjhb#ifdef WITNESS 1214167787Sjhb witness_assert(&rw->lock_object, what, file, line); 1215154941Sjhb#else 1216154941Sjhb /* 1217154941Sjhb * If we hold a write lock fail. We can't reliably check 1218154941Sjhb * to see if we hold a read lock or not. 1219154941Sjhb */ 1220157826Sjhb if (rw_wowner(rw) == curthread) 1221154941Sjhb panic("Lock %s exclusively locked @ %s:%d\n", 1222167787Sjhb rw->lock_object.lo_name, file, line); 1223154941Sjhb#endif 1224154941Sjhb break; 1225154941Sjhb default: 1226154941Sjhb panic("Unknown rw lock assertion: %d @ %s:%d", what, file, 1227154941Sjhb line); 1228154941Sjhb } 1229154941Sjhb} 1230154941Sjhb#endif /* INVARIANT_SUPPORT */ 1231154941Sjhb 1232154941Sjhb#ifdef DDB 1233154941Sjhbvoid 1234227588Spjddb_show_rwlock(const struct lock_object *lock) 1235154941Sjhb{ 1236227588Spjd const struct rwlock *rw; 1237154941Sjhb struct thread *td; 1238154941Sjhb 1239227588Spjd rw = (const struct rwlock *)lock; 1240154941Sjhb 1241154941Sjhb db_printf(" state: "); 1242154941Sjhb if (rw->rw_lock == RW_UNLOCKED) 1243154941Sjhb db_printf("UNLOCKED\n"); 1244169394Sjhb else if (rw->rw_lock == RW_DESTROYED) { 1245169394Sjhb db_printf("DESTROYED\n"); 1246169394Sjhb return; 1247169394Sjhb } else if (rw->rw_lock & RW_LOCK_READ) 1248167504Sjhb db_printf("RLOCK: %ju locks\n", 1249167504Sjhb (uintmax_t)(RW_READERS(rw->rw_lock))); 1250154941Sjhb else { 1251157826Sjhb td = rw_wowner(rw); 1252154941Sjhb db_printf("WLOCK: %p (tid %d, pid %d, \"%s\")\n", td, 1253173600Sjulian td->td_tid, td->td_proc->p_pid, td->td_name); 1254171052Sattilio if (rw_recursed(rw)) 1255171052Sattilio db_printf(" recursed: %u\n", rw->rw_recurse); 1256154941Sjhb } 1257154941Sjhb db_printf(" waiters: "); 1258154941Sjhb switch (rw->rw_lock & (RW_LOCK_READ_WAITERS | RW_LOCK_WRITE_WAITERS)) { 1259154941Sjhb case RW_LOCK_READ_WAITERS: 1260154941Sjhb db_printf("readers\n"); 1261154941Sjhb break; 1262154941Sjhb case RW_LOCK_WRITE_WAITERS: 1263154941Sjhb db_printf("writers\n"); 1264154941Sjhb break; 1265154941Sjhb case RW_LOCK_READ_WAITERS | RW_LOCK_WRITE_WAITERS: 1266167492Sjhb db_printf("readers and writers\n"); 1267154941Sjhb break; 1268154941Sjhb default: 1269154941Sjhb db_printf("none\n"); 1270154941Sjhb break; 1271154941Sjhb } 1272154941Sjhb} 1273154941Sjhb 1274154941Sjhb#endif 1275