kern_rwlock.c revision 301157
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 301157 2016-06-01 18:32:20Z mjg $"); 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); 271286166Smarkj TD_LOCKS_INC(curthread); 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); 306286166Smarkj TD_LOCKS_INC(curthread); 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); 328286166Smarkj TD_LOCKS_DEC(curthread); 329154941Sjhb} 330286166Smarkj 331176017Sjeff/* 332176017Sjeff * Determines whether a new reader can acquire a lock. Succeeds if the 333176017Sjeff * reader already owns a read lock and the lock is locked for read to 334176017Sjeff * prevent deadlock from reader recursion. Also succeeds if the lock 335176017Sjeff * is unlocked and has no writer waiters or spinners. Failing otherwise 336176017Sjeff * prioritizes writers before readers. 337176017Sjeff */ 338176017Sjeff#define RW_CAN_READ(_rw) \ 339176017Sjeff ((curthread->td_rw_rlocks && (_rw) & RW_LOCK_READ) || ((_rw) & \ 340176017Sjeff (RW_LOCK_READ | RW_LOCK_WRITE_WAITERS | RW_LOCK_WRITE_SPINNER)) == \ 341176017Sjeff RW_LOCK_READ) 342154941Sjhb 343154941Sjhbvoid 344242515Sattilio__rw_rlock(volatile uintptr_t *c, const char *file, int line) 345154941Sjhb{ 346242515Sattilio struct rwlock *rw; 347170295Sjeff struct turnstile *ts; 348167801Sjhb#ifdef ADAPTIVE_RWLOCKS 349157846Sjhb volatile struct thread *owner; 350177912Sjeff int spintries = 0; 351177912Sjeff int i; 352157851Swkoszek#endif 353189846Sjeff#ifdef LOCK_PROFILING 354167307Sjhb uint64_t waittime = 0; 355167054Skmacy int contested = 0; 356189846Sjeff#endif 357176017Sjeff uintptr_t v; 358192853Ssson#ifdef KDTRACE_HOOKS 359284297Savg uintptr_t state; 360192853Ssson uint64_t spin_cnt = 0; 361192853Ssson uint64_t sleep_cnt = 0; 362192853Ssson int64_t sleep_time = 0; 363284297Savg int64_t all_time = 0; 364192853Ssson#endif 365154941Sjhb 366228424Savg if (SCHEDULER_STOPPED()) 367228424Savg return; 368228424Savg 369242515Sattilio rw = rwlock2rw(c); 370242515Sattilio 371244582Sattilio KASSERT(kdb_active != 0 || !TD_IS_IDLETHREAD(curthread), 372240424Sattilio ("rw_rlock() by idle thread %p on rwlock %s @ %s:%d", 373240424Sattilio curthread, rw->lock_object.lo_name, file, line)); 374169394Sjhb KASSERT(rw->rw_lock != RW_DESTROYED, 375169394Sjhb ("rw_rlock() of destroyed rwlock @ %s:%d", file, line)); 376157826Sjhb KASSERT(rw_wowner(rw) != curthread, 377251323Sjhb ("rw_rlock: wlock already held for %s @ %s:%d", 378167787Sjhb rw->lock_object.lo_name, file, line)); 379182914Sjhb WITNESS_CHECKORDER(&rw->lock_object, LOP_NEWORDER, file, line, NULL); 380154941Sjhb 381284297Savg#ifdef KDTRACE_HOOKS 382285664Smarkj all_time -= lockstat_nsecs(&rw->lock_object); 383284297Savg state = rw->rw_lock; 384284297Savg#endif 385154941Sjhb for (;;) { 386154941Sjhb /* 387154941Sjhb * Handle the easy case. If no other thread has a write 388154941Sjhb * lock, then try to bump up the count of read locks. Note 389154941Sjhb * that we have to preserve the current state of the 390154941Sjhb * RW_LOCK_WRITE_WAITERS flag. If we fail to acquire a 391154941Sjhb * read lock, then rw_lock must have changed, so restart 392154941Sjhb * the loop. Note that this handles the case of a 393154941Sjhb * completely unlocked rwlock since such a lock is encoded 394154941Sjhb * as a read lock with no waiters. 395154941Sjhb */ 396176017Sjeff v = rw->rw_lock; 397176017Sjeff if (RW_CAN_READ(v)) { 398154941Sjhb /* 399154941Sjhb * The RW_LOCK_READ_WAITERS flag should only be set 400176017Sjeff * if the lock has been unlocked and write waiters 401176017Sjeff * were present. 402154941Sjhb */ 403176017Sjeff if (atomic_cmpset_acq_ptr(&rw->rw_lock, v, 404176017Sjeff v + RW_ONE_READER)) { 405167787Sjhb if (LOCK_LOG_TEST(&rw->lock_object, 0)) 406154941Sjhb CTR4(KTR_LOCK, 407154941Sjhb "%s: %p succeed %p -> %p", __func__, 408176017Sjeff rw, (void *)v, 409176017Sjeff (void *)(v + RW_ONE_READER)); 410154941Sjhb break; 411154941Sjhb } 412154941Sjhb continue; 413154941Sjhb } 414285706Smarkj#ifdef KDTRACE_HOOKS 415285706Smarkj spin_cnt++; 416285706Smarkj#endif 417233628Sfabient#ifdef HWPMC_HOOKS 418233628Sfabient PMC_SOFT_CALL( , , lock, failed); 419233628Sfabient#endif 420174629Sjeff lock_profile_obtain_lock_failed(&rw->lock_object, 421174629Sjeff &contested, &waittime); 422154941Sjhb 423173960Sattilio#ifdef ADAPTIVE_RWLOCKS 424154941Sjhb /* 425173960Sattilio * If the owner is running on another CPU, spin until 426173960Sattilio * the owner stops running or the state of the lock 427173960Sattilio * changes. 428173960Sattilio */ 429176017Sjeff if ((v & RW_LOCK_READ) == 0) { 430176017Sjeff owner = (struct thread *)RW_OWNER(v); 431176017Sjeff if (TD_IS_RUNNING(owner)) { 432176017Sjeff if (LOCK_LOG_TEST(&rw->lock_object, 0)) 433176017Sjeff CTR3(KTR_LOCK, 434176017Sjeff "%s: spinning on %p held by %p", 435176017Sjeff __func__, rw, owner); 436274092Sjhb KTR_STATE1(KTR_SCHED, "thread", 437274092Sjhb sched_tdname(curthread), "spinning", 438274092Sjhb "lockname:\"%s\"", rw->lock_object.lo_name); 439176017Sjeff while ((struct thread*)RW_OWNER(rw->rw_lock) == 440192853Ssson owner && TD_IS_RUNNING(owner)) { 441176017Sjeff cpu_spinwait(); 442192853Ssson#ifdef KDTRACE_HOOKS 443192853Ssson spin_cnt++; 444192853Ssson#endif 445192853Ssson } 446274092Sjhb KTR_STATE0(KTR_SCHED, "thread", 447274092Sjhb sched_tdname(curthread), "running"); 448176017Sjeff continue; 449176017Sjeff } 450177912Sjeff } else if (spintries < rowner_retries) { 451177912Sjeff spintries++; 452274092Sjhb KTR_STATE1(KTR_SCHED, "thread", sched_tdname(curthread), 453274092Sjhb "spinning", "lockname:\"%s\"", 454274092Sjhb rw->lock_object.lo_name); 455177912Sjeff for (i = 0; i < rowner_loops; i++) { 456177912Sjeff v = rw->rw_lock; 457177912Sjeff if ((v & RW_LOCK_READ) == 0 || RW_CAN_READ(v)) 458177912Sjeff break; 459177912Sjeff cpu_spinwait(); 460177912Sjeff } 461259509Sattilio#ifdef KDTRACE_HOOKS 462259509Sattilio spin_cnt += rowner_loops - i; 463259509Sattilio#endif 464274092Sjhb KTR_STATE0(KTR_SCHED, "thread", sched_tdname(curthread), 465274092Sjhb "running"); 466177912Sjeff if (i != rowner_loops) 467177912Sjeff continue; 468173960Sattilio } 469173960Sattilio#endif 470173960Sattilio 471173960Sattilio /* 472154941Sjhb * Okay, now it's the hard case. Some other thread already 473176017Sjeff * has a write lock or there are write waiters present, 474176017Sjeff * acquire the turnstile lock so we can begin the process 475176017Sjeff * of blocking. 476154941Sjhb */ 477170295Sjeff ts = turnstile_trywait(&rw->lock_object); 478154941Sjhb 479154941Sjhb /* 480154941Sjhb * The lock might have been released while we spun, so 481176017Sjeff * recheck its state and restart the loop if needed. 482154941Sjhb */ 483176017Sjeff v = rw->rw_lock; 484176017Sjeff if (RW_CAN_READ(v)) { 485170295Sjeff turnstile_cancel(ts); 486154941Sjhb continue; 487154941Sjhb } 488154941Sjhb 489173960Sattilio#ifdef ADAPTIVE_RWLOCKS 490154941Sjhb /* 491193035Sjhb * The current lock owner might have started executing 492193035Sjhb * on another CPU (or the lock could have changed 493193035Sjhb * owners) while we were waiting on the turnstile 494193035Sjhb * chain lock. If so, drop the turnstile lock and try 495193035Sjhb * again. 496173960Sattilio */ 497176017Sjeff if ((v & RW_LOCK_READ) == 0) { 498176017Sjeff owner = (struct thread *)RW_OWNER(v); 499176017Sjeff if (TD_IS_RUNNING(owner)) { 500176017Sjeff turnstile_cancel(ts); 501176017Sjeff continue; 502176017Sjeff } 503173960Sattilio } 504173960Sattilio#endif 505173960Sattilio 506173960Sattilio /* 507176017Sjeff * The lock is held in write mode or it already has waiters. 508154941Sjhb */ 509176017Sjeff MPASS(!RW_CAN_READ(v)); 510176017Sjeff 511176017Sjeff /* 512176017Sjeff * If the RW_LOCK_READ_WAITERS flag is already set, then 513176017Sjeff * we can go ahead and block. If it is not set then try 514176017Sjeff * to set it. If we fail to set it drop the turnstile 515176017Sjeff * lock and restart the loop. 516176017Sjeff */ 517176017Sjeff if (!(v & RW_LOCK_READ_WAITERS)) { 518176017Sjeff if (!atomic_cmpset_ptr(&rw->rw_lock, v, 519176017Sjeff v | RW_LOCK_READ_WAITERS)) { 520170295Sjeff turnstile_cancel(ts); 521157826Sjhb continue; 522157826Sjhb } 523167787Sjhb if (LOCK_LOG_TEST(&rw->lock_object, 0)) 524157826Sjhb CTR2(KTR_LOCK, "%s: %p set read waiters flag", 525157826Sjhb __func__, rw); 526154941Sjhb } 527154941Sjhb 528154941Sjhb /* 529154941Sjhb * We were unable to acquire the lock and the read waiters 530154941Sjhb * flag is set, so we must block on the turnstile. 531154941Sjhb */ 532167787Sjhb if (LOCK_LOG_TEST(&rw->lock_object, 0)) 533154941Sjhb CTR2(KTR_LOCK, "%s: %p blocking on turnstile", __func__, 534154941Sjhb rw); 535192853Ssson#ifdef KDTRACE_HOOKS 536285664Smarkj sleep_time -= lockstat_nsecs(&rw->lock_object); 537192853Ssson#endif 538170295Sjeff turnstile_wait(ts, rw_owner(rw), TS_SHARED_QUEUE); 539192853Ssson#ifdef KDTRACE_HOOKS 540285664Smarkj sleep_time += lockstat_nsecs(&rw->lock_object); 541192853Ssson sleep_cnt++; 542192853Ssson#endif 543167787Sjhb if (LOCK_LOG_TEST(&rw->lock_object, 0)) 544154941Sjhb CTR2(KTR_LOCK, "%s: %p resuming from turnstile", 545154941Sjhb __func__, rw); 546154941Sjhb } 547284297Savg#ifdef KDTRACE_HOOKS 548285664Smarkj all_time += lockstat_nsecs(&rw->lock_object); 549284297Savg if (sleep_time) 550285703Smarkj LOCKSTAT_RECORD4(rw__block, rw, sleep_time, 551284297Savg LOCKSTAT_READER, (state & RW_LOCK_READ) == 0, 552284297Savg (state & RW_LOCK_READ) == 0 ? 0 : RW_READERS(state)); 553154941Sjhb 554284297Savg /* Record only the loops spinning and not sleeping. */ 555284297Savg if (spin_cnt > sleep_cnt) 556285703Smarkj LOCKSTAT_RECORD4(rw__spin, rw, all_time - sleep_time, 557284297Savg LOCKSTAT_READER, (state & RW_LOCK_READ) == 0, 558284297Savg (state & RW_LOCK_READ) == 0 ? 0 : RW_READERS(state)); 559284297Savg#endif 560154941Sjhb /* 561154941Sjhb * TODO: acquire "owner of record" here. Here be turnstile dragons 562154941Sjhb * however. turnstiles don't like owners changing between calls to 563154941Sjhb * turnstile_wait() currently. 564154941Sjhb */ 565285704Smarkj LOCKSTAT_PROFILE_OBTAIN_RWLOCK_SUCCESS(rw__acquire, rw, contested, 566285704Smarkj waittime, file, line, LOCKSTAT_READER); 567167787Sjhb LOCK_LOG_LOCK("RLOCK", &rw->lock_object, 0, 0, file, line); 568167787Sjhb WITNESS_LOCK(&rw->lock_object, 0, file, line); 569286166Smarkj TD_LOCKS_INC(curthread); 570176017Sjeff curthread->td_rw_rlocks++; 571154941Sjhb} 572154941Sjhb 573177843Sattilioint 574242515Sattilio__rw_try_rlock(volatile uintptr_t *c, const char *file, int line) 575177843Sattilio{ 576242515Sattilio struct rwlock *rw; 577177843Sattilio uintptr_t x; 578177843Sattilio 579228424Savg if (SCHEDULER_STOPPED()) 580228424Savg return (1); 581228424Savg 582242515Sattilio rw = rwlock2rw(c); 583242515Sattilio 584244582Sattilio KASSERT(kdb_active != 0 || !TD_IS_IDLETHREAD(curthread), 585240424Sattilio ("rw_try_rlock() by idle thread %p on rwlock %s @ %s:%d", 586240424Sattilio curthread, rw->lock_object.lo_name, file, line)); 587240424Sattilio 588177843Sattilio for (;;) { 589177843Sattilio x = rw->rw_lock; 590177843Sattilio KASSERT(rw->rw_lock != RW_DESTROYED, 591177843Sattilio ("rw_try_rlock() of destroyed rwlock @ %s:%d", file, line)); 592177843Sattilio if (!(x & RW_LOCK_READ)) 593177843Sattilio break; 594177843Sattilio if (atomic_cmpset_acq_ptr(&rw->rw_lock, x, x + RW_ONE_READER)) { 595177843Sattilio LOCK_LOG_TRY("RLOCK", &rw->lock_object, 0, 1, file, 596177843Sattilio line); 597177843Sattilio WITNESS_LOCK(&rw->lock_object, LOP_TRYLOCK, file, line); 598285704Smarkj LOCKSTAT_PROFILE_OBTAIN_RWLOCK_SUCCESS(rw__acquire, 599285704Smarkj rw, 0, 0, file, line, LOCKSTAT_READER); 600286166Smarkj TD_LOCKS_INC(curthread); 601177843Sattilio curthread->td_rw_rlocks++; 602177843Sattilio return (1); 603177843Sattilio } 604177843Sattilio } 605177843Sattilio 606177843Sattilio LOCK_LOG_TRY("RLOCK", &rw->lock_object, 0, 0, file, line); 607177843Sattilio return (0); 608177843Sattilio} 609177843Sattilio 610154941Sjhbvoid 611242515Sattilio_rw_runlock_cookie(volatile uintptr_t *c, const char *file, int line) 612154941Sjhb{ 613242515Sattilio struct rwlock *rw; 614154941Sjhb struct turnstile *ts; 615176017Sjeff uintptr_t x, v, queue; 616154941Sjhb 617228424Savg if (SCHEDULER_STOPPED()) 618228424Savg return; 619228424Savg 620242515Sattilio rw = rwlock2rw(c); 621242515Sattilio 622169394Sjhb KASSERT(rw->rw_lock != RW_DESTROYED, 623169394Sjhb ("rw_runlock() of destroyed rwlock @ %s:%d", file, line)); 624242515Sattilio __rw_assert(c, RA_RLOCKED, file, line); 625167787Sjhb WITNESS_UNLOCK(&rw->lock_object, 0, file, line); 626167787Sjhb LOCK_LOG_LOCK("RUNLOCK", &rw->lock_object, 0, 0, file, line); 627154941Sjhb 628154941Sjhb /* TODO: drop "owner of record" here. */ 629154941Sjhb 630154941Sjhb for (;;) { 631154941Sjhb /* 632154941Sjhb * See if there is more than one read lock held. If so, 633154941Sjhb * just drop one and return. 634154941Sjhb */ 635154941Sjhb x = rw->rw_lock; 636154941Sjhb if (RW_READERS(x) > 1) { 637197643Sattilio if (atomic_cmpset_rel_ptr(&rw->rw_lock, x, 638154941Sjhb x - RW_ONE_READER)) { 639167787Sjhb if (LOCK_LOG_TEST(&rw->lock_object, 0)) 640154941Sjhb CTR4(KTR_LOCK, 641154941Sjhb "%s: %p succeeded %p -> %p", 642154941Sjhb __func__, rw, (void *)x, 643154941Sjhb (void *)(x - RW_ONE_READER)); 644154941Sjhb break; 645154941Sjhb } 646154941Sjhb continue; 647167307Sjhb } 648154941Sjhb /* 649154941Sjhb * If there aren't any waiters for a write lock, then try 650154941Sjhb * to drop it quickly. 651154941Sjhb */ 652176017Sjeff if (!(x & RW_LOCK_WAITERS)) { 653176017Sjeff MPASS((x & ~RW_LOCK_WRITE_SPINNER) == 654176017Sjeff RW_READERS_LOCK(1)); 655197643Sattilio if (atomic_cmpset_rel_ptr(&rw->rw_lock, x, 656197643Sattilio RW_UNLOCKED)) { 657167787Sjhb if (LOCK_LOG_TEST(&rw->lock_object, 0)) 658154941Sjhb CTR2(KTR_LOCK, "%s: %p last succeeded", 659154941Sjhb __func__, rw); 660154941Sjhb break; 661154941Sjhb } 662154941Sjhb continue; 663154941Sjhb } 664154941Sjhb /* 665176017Sjeff * Ok, we know we have waiters and we think we are the 666176017Sjeff * last reader, so grab the turnstile lock. 667154941Sjhb */ 668170295Sjeff turnstile_chain_lock(&rw->lock_object); 669176017Sjeff v = rw->rw_lock & (RW_LOCK_WAITERS | RW_LOCK_WRITE_SPINNER); 670176017Sjeff MPASS(v & RW_LOCK_WAITERS); 671154941Sjhb 672154941Sjhb /* 673154941Sjhb * Try to drop our lock leaving the lock in a unlocked 674154941Sjhb * state. 675154941Sjhb * 676154941Sjhb * If you wanted to do explicit lock handoff you'd have to 677154941Sjhb * do it here. You'd also want to use turnstile_signal() 678154941Sjhb * and you'd have to handle the race where a higher 679154941Sjhb * priority thread blocks on the write lock before the 680154941Sjhb * thread you wakeup actually runs and have the new thread 681154941Sjhb * "steal" the lock. For now it's a lot simpler to just 682154941Sjhb * wakeup all of the waiters. 683154941Sjhb * 684154941Sjhb * As above, if we fail, then another thread might have 685154941Sjhb * acquired a read lock, so drop the turnstile lock and 686154941Sjhb * restart. 687154941Sjhb */ 688176017Sjeff x = RW_UNLOCKED; 689176017Sjeff if (v & RW_LOCK_WRITE_WAITERS) { 690176017Sjeff queue = TS_EXCLUSIVE_QUEUE; 691176017Sjeff x |= (v & RW_LOCK_READ_WAITERS); 692176017Sjeff } else 693176017Sjeff queue = TS_SHARED_QUEUE; 694197643Sattilio if (!atomic_cmpset_rel_ptr(&rw->rw_lock, RW_READERS_LOCK(1) | v, 695176017Sjeff x)) { 696170295Sjeff turnstile_chain_unlock(&rw->lock_object); 697154941Sjhb continue; 698154941Sjhb } 699167787Sjhb if (LOCK_LOG_TEST(&rw->lock_object, 0)) 700154941Sjhb CTR2(KTR_LOCK, "%s: %p last succeeded with waiters", 701154941Sjhb __func__, rw); 702154941Sjhb 703154941Sjhb /* 704154941Sjhb * Ok. The lock is released and all that's left is to 705154941Sjhb * wake up the waiters. Note that the lock might not be 706154941Sjhb * free anymore, but in that case the writers will just 707154941Sjhb * block again if they run before the new lock holder(s) 708154941Sjhb * release the lock. 709154941Sjhb */ 710167787Sjhb ts = turnstile_lookup(&rw->lock_object); 711157846Sjhb MPASS(ts != NULL); 712176017Sjeff turnstile_broadcast(ts, queue); 713154941Sjhb turnstile_unpend(ts, TS_SHARED_LOCK); 714170295Sjeff turnstile_chain_unlock(&rw->lock_object); 715154941Sjhb break; 716154941Sjhb } 717285704Smarkj LOCKSTAT_PROFILE_RELEASE_RWLOCK(rw__release, rw, LOCKSTAT_READER); 718286166Smarkj TD_LOCKS_DEC(curthread); 719252212Sjhb curthread->td_rw_rlocks--; 720154941Sjhb} 721154941Sjhb 722154941Sjhb/* 723154941Sjhb * This function is called when we are unable to obtain a write lock on the 724154941Sjhb * first try. This means that at least one other thread holds either a 725154941Sjhb * read or write lock. 726154941Sjhb */ 727154941Sjhbvoid 728242515Sattilio__rw_wlock_hard(volatile uintptr_t *c, uintptr_t tid, const char *file, 729242515Sattilio int line) 730154941Sjhb{ 731242515Sattilio struct rwlock *rw; 732170295Sjeff struct turnstile *ts; 733167801Sjhb#ifdef ADAPTIVE_RWLOCKS 734157846Sjhb volatile struct thread *owner; 735176017Sjeff int spintries = 0; 736176017Sjeff int i; 737157851Swkoszek#endif 738189846Sjeff uintptr_t v, x; 739189846Sjeff#ifdef LOCK_PROFILING 740171516Sattilio uint64_t waittime = 0; 741171516Sattilio int contested = 0; 742189846Sjeff#endif 743192853Ssson#ifdef KDTRACE_HOOKS 744284297Savg uintptr_t state; 745192853Ssson uint64_t spin_cnt = 0; 746192853Ssson uint64_t sleep_cnt = 0; 747192853Ssson int64_t sleep_time = 0; 748284297Savg int64_t all_time = 0; 749192853Ssson#endif 750154941Sjhb 751228424Savg if (SCHEDULER_STOPPED()) 752228424Savg return; 753228424Savg 754242515Sattilio rw = rwlock2rw(c); 755242515Sattilio 756171052Sattilio if (rw_wlocked(rw)) { 757193307Sattilio KASSERT(rw->lock_object.lo_flags & LO_RECURSABLE, 758171052Sattilio ("%s: recursing but non-recursive rw %s @ %s:%d\n", 759171052Sattilio __func__, rw->lock_object.lo_name, file, line)); 760171052Sattilio rw->rw_recurse++; 761171052Sattilio if (LOCK_LOG_TEST(&rw->lock_object, 0)) 762171052Sattilio CTR2(KTR_LOCK, "%s: %p recursing", __func__, rw); 763171052Sattilio return; 764171052Sattilio } 765171052Sattilio 766167787Sjhb if (LOCK_LOG_TEST(&rw->lock_object, 0)) 767154941Sjhb CTR5(KTR_LOCK, "%s: %s contested (lock=%p) at %s:%d", __func__, 768167787Sjhb rw->lock_object.lo_name, (void *)rw->rw_lock, file, line); 769154941Sjhb 770284297Savg#ifdef KDTRACE_HOOKS 771285664Smarkj all_time -= lockstat_nsecs(&rw->lock_object); 772284297Savg state = rw->rw_lock; 773284297Savg#endif 774301157Smjg for (;;) { 775301157Smjg if (rw->rw_lock == RW_UNLOCKED && _rw_write_lock(rw, tid)) 776301157Smjg break; 777192853Ssson#ifdef KDTRACE_HOOKS 778192853Ssson spin_cnt++; 779192853Ssson#endif 780233628Sfabient#ifdef HWPMC_HOOKS 781233628Sfabient PMC_SOFT_CALL( , , lock, failed); 782233628Sfabient#endif 783174629Sjeff lock_profile_obtain_lock_failed(&rw->lock_object, 784174629Sjeff &contested, &waittime); 785173960Sattilio#ifdef ADAPTIVE_RWLOCKS 786173960Sattilio /* 787173960Sattilio * If the lock is write locked and the owner is 788173960Sattilio * running on another CPU, spin until the owner stops 789173960Sattilio * running or the state of the lock changes. 790173960Sattilio */ 791173960Sattilio v = rw->rw_lock; 792173960Sattilio owner = (struct thread *)RW_OWNER(v); 793173960Sattilio if (!(v & RW_LOCK_READ) && TD_IS_RUNNING(owner)) { 794173960Sattilio if (LOCK_LOG_TEST(&rw->lock_object, 0)) 795173960Sattilio CTR3(KTR_LOCK, "%s: spinning on %p held by %p", 796173960Sattilio __func__, rw, owner); 797274092Sjhb KTR_STATE1(KTR_SCHED, "thread", sched_tdname(curthread), 798274092Sjhb "spinning", "lockname:\"%s\"", 799274092Sjhb rw->lock_object.lo_name); 800173960Sattilio while ((struct thread*)RW_OWNER(rw->rw_lock) == owner && 801192853Ssson TD_IS_RUNNING(owner)) { 802173960Sattilio cpu_spinwait(); 803192853Ssson#ifdef KDTRACE_HOOKS 804192853Ssson spin_cnt++; 805192853Ssson#endif 806192853Ssson } 807274092Sjhb KTR_STATE0(KTR_SCHED, "thread", sched_tdname(curthread), 808274092Sjhb "running"); 809173960Sattilio continue; 810173960Sattilio } 811177912Sjeff if ((v & RW_LOCK_READ) && RW_READERS(v) && 812177912Sjeff spintries < rowner_retries) { 813176017Sjeff if (!(v & RW_LOCK_WRITE_SPINNER)) { 814176017Sjeff if (!atomic_cmpset_ptr(&rw->rw_lock, v, 815176017Sjeff v | RW_LOCK_WRITE_SPINNER)) { 816176017Sjeff continue; 817176017Sjeff } 818176017Sjeff } 819176017Sjeff spintries++; 820274092Sjhb KTR_STATE1(KTR_SCHED, "thread", sched_tdname(curthread), 821274092Sjhb "spinning", "lockname:\"%s\"", 822274092Sjhb rw->lock_object.lo_name); 823177912Sjeff for (i = 0; i < rowner_loops; i++) { 824176017Sjeff if ((rw->rw_lock & RW_LOCK_WRITE_SPINNER) == 0) 825176017Sjeff break; 826176017Sjeff cpu_spinwait(); 827176017Sjeff } 828274092Sjhb KTR_STATE0(KTR_SCHED, "thread", sched_tdname(curthread), 829274092Sjhb "running"); 830192853Ssson#ifdef KDTRACE_HOOKS 831192853Ssson spin_cnt += rowner_loops - i; 832192853Ssson#endif 833177912Sjeff if (i != rowner_loops) 834176017Sjeff continue; 835176017Sjeff } 836173960Sattilio#endif 837170295Sjeff ts = turnstile_trywait(&rw->lock_object); 838154941Sjhb v = rw->rw_lock; 839154941Sjhb 840173960Sattilio#ifdef ADAPTIVE_RWLOCKS 841154941Sjhb /* 842193035Sjhb * The current lock owner might have started executing 843193035Sjhb * on another CPU (or the lock could have changed 844193035Sjhb * owners) while we were waiting on the turnstile 845193035Sjhb * chain lock. If so, drop the turnstile lock and try 846193035Sjhb * again. 847173960Sattilio */ 848173960Sattilio if (!(v & RW_LOCK_READ)) { 849173960Sattilio owner = (struct thread *)RW_OWNER(v); 850173960Sattilio if (TD_IS_RUNNING(owner)) { 851173960Sattilio turnstile_cancel(ts); 852173960Sattilio continue; 853173960Sattilio } 854173960Sattilio } 855173960Sattilio#endif 856173960Sattilio /* 857179334Sattilio * Check for the waiters flags about this rwlock. 858179334Sattilio * If the lock was released, without maintain any pending 859179334Sattilio * waiters queue, simply try to acquire it. 860179334Sattilio * If a pending waiters queue is present, claim the lock 861179334Sattilio * ownership and maintain the pending queue. 862154941Sjhb */ 863176017Sjeff x = v & (RW_LOCK_WAITERS | RW_LOCK_WRITE_SPINNER); 864176017Sjeff if ((v & ~x) == RW_UNLOCKED) { 865176017Sjeff x &= ~RW_LOCK_WRITE_SPINNER; 866176017Sjeff if (atomic_cmpset_acq_ptr(&rw->rw_lock, v, tid | x)) { 867176017Sjeff if (x) 868176017Sjeff turnstile_claim(ts); 869176017Sjeff else 870176017Sjeff turnstile_cancel(ts); 871154941Sjhb break; 872154941Sjhb } 873170295Sjeff turnstile_cancel(ts); 874154941Sjhb continue; 875154941Sjhb } 876154941Sjhb /* 877154941Sjhb * If the RW_LOCK_WRITE_WAITERS flag isn't set, then try to 878154941Sjhb * set it. If we fail to set it, then loop back and try 879154941Sjhb * again. 880154941Sjhb */ 881157826Sjhb if (!(v & RW_LOCK_WRITE_WAITERS)) { 882157826Sjhb if (!atomic_cmpset_ptr(&rw->rw_lock, v, 883157826Sjhb v | RW_LOCK_WRITE_WAITERS)) { 884170295Sjeff turnstile_cancel(ts); 885157826Sjhb continue; 886157826Sjhb } 887167787Sjhb if (LOCK_LOG_TEST(&rw->lock_object, 0)) 888157826Sjhb CTR2(KTR_LOCK, "%s: %p set write waiters flag", 889157826Sjhb __func__, rw); 890154941Sjhb } 891157846Sjhb /* 892154941Sjhb * We were unable to acquire the lock and the write waiters 893154941Sjhb * flag is set, so we must block on the turnstile. 894154941Sjhb */ 895167787Sjhb if (LOCK_LOG_TEST(&rw->lock_object, 0)) 896154941Sjhb CTR2(KTR_LOCK, "%s: %p blocking on turnstile", __func__, 897154941Sjhb rw); 898192853Ssson#ifdef KDTRACE_HOOKS 899285664Smarkj sleep_time -= lockstat_nsecs(&rw->lock_object); 900192853Ssson#endif 901170295Sjeff turnstile_wait(ts, rw_owner(rw), TS_EXCLUSIVE_QUEUE); 902192853Ssson#ifdef KDTRACE_HOOKS 903285664Smarkj sleep_time += lockstat_nsecs(&rw->lock_object); 904192853Ssson sleep_cnt++; 905192853Ssson#endif 906167787Sjhb if (LOCK_LOG_TEST(&rw->lock_object, 0)) 907154941Sjhb CTR2(KTR_LOCK, "%s: %p resuming from turnstile", 908154941Sjhb __func__, rw); 909176017Sjeff#ifdef ADAPTIVE_RWLOCKS 910176017Sjeff spintries = 0; 911176017Sjeff#endif 912154941Sjhb } 913192853Ssson#ifdef KDTRACE_HOOKS 914285664Smarkj all_time += lockstat_nsecs(&rw->lock_object); 915192853Ssson if (sleep_time) 916285703Smarkj LOCKSTAT_RECORD4(rw__block, rw, sleep_time, 917284297Savg LOCKSTAT_WRITER, (state & RW_LOCK_READ) == 0, 918284297Savg (state & RW_LOCK_READ) == 0 ? 0 : RW_READERS(state)); 919192853Ssson 920284297Savg /* Record only the loops spinning and not sleeping. */ 921192853Ssson if (spin_cnt > sleep_cnt) 922285703Smarkj LOCKSTAT_RECORD4(rw__spin, rw, all_time - sleep_time, 923284297Savg LOCKSTAT_READER, (state & RW_LOCK_READ) == 0, 924284297Savg (state & RW_LOCK_READ) == 0 ? 0 : RW_READERS(state)); 925192853Ssson#endif 926285704Smarkj LOCKSTAT_PROFILE_OBTAIN_RWLOCK_SUCCESS(rw__acquire, rw, contested, 927285704Smarkj waittime, file, line, LOCKSTAT_WRITER); 928154941Sjhb} 929154941Sjhb 930154941Sjhb/* 931154941Sjhb * This function is called if the first try at releasing a write lock failed. 932154941Sjhb * This means that one of the 2 waiter bits must be set indicating that at 933154941Sjhb * least one thread is waiting on this lock. 934154941Sjhb */ 935154941Sjhbvoid 936242515Sattilio__rw_wunlock_hard(volatile uintptr_t *c, uintptr_t tid, const char *file, 937242515Sattilio int line) 938154941Sjhb{ 939242515Sattilio struct rwlock *rw; 940154941Sjhb struct turnstile *ts; 941154941Sjhb uintptr_t v; 942154941Sjhb int queue; 943154941Sjhb 944228424Savg if (SCHEDULER_STOPPED()) 945228424Savg return; 946228424Savg 947242515Sattilio rw = rwlock2rw(c); 948242515Sattilio 949171052Sattilio if (rw_wlocked(rw) && rw_recursed(rw)) { 950176017Sjeff rw->rw_recurse--; 951171052Sattilio if (LOCK_LOG_TEST(&rw->lock_object, 0)) 952171052Sattilio CTR2(KTR_LOCK, "%s: %p unrecursing", __func__, rw); 953171052Sattilio return; 954171052Sattilio } 955171052Sattilio 956154941Sjhb KASSERT(rw->rw_lock & (RW_LOCK_READ_WAITERS | RW_LOCK_WRITE_WAITERS), 957154941Sjhb ("%s: neither of the waiter flags are set", __func__)); 958154941Sjhb 959167787Sjhb if (LOCK_LOG_TEST(&rw->lock_object, 0)) 960154941Sjhb CTR2(KTR_LOCK, "%s: %p contested", __func__, rw); 961154941Sjhb 962170295Sjeff turnstile_chain_lock(&rw->lock_object); 963167787Sjhb ts = turnstile_lookup(&rw->lock_object); 964154941Sjhb MPASS(ts != NULL); 965154941Sjhb 966154941Sjhb /* 967154941Sjhb * Use the same algo as sx locks for now. Prefer waking up shared 968154941Sjhb * waiters if we have any over writers. This is probably not ideal. 969154941Sjhb * 970154941Sjhb * 'v' is the value we are going to write back to rw_lock. If we 971154941Sjhb * have waiters on both queues, we need to preserve the state of 972154941Sjhb * the waiter flag for the queue we don't wake up. For now this is 973154941Sjhb * hardcoded for the algorithm mentioned above. 974154941Sjhb * 975154941Sjhb * In the case of both readers and writers waiting we wakeup the 976154941Sjhb * readers but leave the RW_LOCK_WRITE_WAITERS flag set. If a 977154941Sjhb * new writer comes in before a reader it will claim the lock up 978154941Sjhb * above. There is probably a potential priority inversion in 979154941Sjhb * there that could be worked around either by waking both queues 980154941Sjhb * of waiters or doing some complicated lock handoff gymnastics. 981154941Sjhb */ 982157846Sjhb v = RW_UNLOCKED; 983176076Sjeff if (rw->rw_lock & RW_LOCK_WRITE_WAITERS) { 984176076Sjeff queue = TS_EXCLUSIVE_QUEUE; 985176076Sjeff v |= (rw->rw_lock & RW_LOCK_READ_WAITERS); 986176076Sjeff } else 987154941Sjhb queue = TS_SHARED_QUEUE; 988157846Sjhb 989157846Sjhb /* Wake up all waiters for the specific queue. */ 990167787Sjhb if (LOCK_LOG_TEST(&rw->lock_object, 0)) 991154941Sjhb CTR3(KTR_LOCK, "%s: %p waking up %s waiters", __func__, rw, 992154941Sjhb queue == TS_SHARED_QUEUE ? "read" : "write"); 993154941Sjhb turnstile_broadcast(ts, queue); 994154941Sjhb atomic_store_rel_ptr(&rw->rw_lock, v); 995154941Sjhb turnstile_unpend(ts, TS_EXCLUSIVE_LOCK); 996170295Sjeff turnstile_chain_unlock(&rw->lock_object); 997154941Sjhb} 998154941Sjhb 999157882Sjhb/* 1000157882Sjhb * Attempt to do a non-blocking upgrade from a read lock to a write 1001157882Sjhb * lock. This will only succeed if this thread holds a single read 1002157882Sjhb * lock. Returns true if the upgrade succeeded and false otherwise. 1003157882Sjhb */ 1004157882Sjhbint 1005242515Sattilio__rw_try_upgrade(volatile uintptr_t *c, const char *file, int line) 1006157882Sjhb{ 1007242515Sattilio struct rwlock *rw; 1008176017Sjeff uintptr_t v, x, tid; 1009170295Sjeff struct turnstile *ts; 1010157882Sjhb int success; 1011157882Sjhb 1012228424Savg if (SCHEDULER_STOPPED()) 1013228424Savg return (1); 1014228424Savg 1015242515Sattilio rw = rwlock2rw(c); 1016242515Sattilio 1017169394Sjhb KASSERT(rw->rw_lock != RW_DESTROYED, 1018169394Sjhb ("rw_try_upgrade() of destroyed rwlock @ %s:%d", file, line)); 1019242515Sattilio __rw_assert(c, RA_RLOCKED, file, line); 1020157882Sjhb 1021157882Sjhb /* 1022157882Sjhb * Attempt to switch from one reader to a writer. If there 1023157882Sjhb * are any write waiters, then we will have to lock the 1024157882Sjhb * turnstile first to prevent races with another writer 1025157882Sjhb * calling turnstile_wait() before we have claimed this 1026157882Sjhb * turnstile. So, do the simple case of no waiters first. 1027157882Sjhb */ 1028157882Sjhb tid = (uintptr_t)curthread; 1029176017Sjeff success = 0; 1030176017Sjeff for (;;) { 1031176017Sjeff v = rw->rw_lock; 1032176017Sjeff if (RW_READERS(v) > 1) 1033176017Sjeff break; 1034176017Sjeff if (!(v & RW_LOCK_WAITERS)) { 1035176017Sjeff success = atomic_cmpset_ptr(&rw->rw_lock, v, tid); 1036176017Sjeff if (!success) 1037176017Sjeff continue; 1038176017Sjeff break; 1039176017Sjeff } 1040157882Sjhb 1041176017Sjeff /* 1042176017Sjeff * Ok, we think we have waiters, so lock the turnstile. 1043176017Sjeff */ 1044176017Sjeff ts = turnstile_trywait(&rw->lock_object); 1045176017Sjeff v = rw->rw_lock; 1046176017Sjeff if (RW_READERS(v) > 1) { 1047176017Sjeff turnstile_cancel(ts); 1048176017Sjeff break; 1049176017Sjeff } 1050176017Sjeff /* 1051176017Sjeff * Try to switch from one reader to a writer again. This time 1052176017Sjeff * we honor the current state of the waiters flags. 1053176017Sjeff * If we obtain the lock with the flags set, then claim 1054176017Sjeff * ownership of the turnstile. 1055176017Sjeff */ 1056176017Sjeff x = rw->rw_lock & RW_LOCK_WAITERS; 1057176017Sjeff success = atomic_cmpset_ptr(&rw->rw_lock, v, tid | x); 1058176017Sjeff if (success) { 1059176017Sjeff if (x) 1060176017Sjeff turnstile_claim(ts); 1061176017Sjeff else 1062176017Sjeff turnstile_cancel(ts); 1063176017Sjeff break; 1064176017Sjeff } 1065170295Sjeff turnstile_cancel(ts); 1066176017Sjeff } 1067167787Sjhb LOCK_LOG_TRY("WUPGRADE", &rw->lock_object, 0, success, file, line); 1068176017Sjeff if (success) { 1069176017Sjeff curthread->td_rw_rlocks--; 1070167787Sjhb WITNESS_UPGRADE(&rw->lock_object, LOP_EXCLUSIVE | LOP_TRYLOCK, 1071157882Sjhb file, line); 1072285703Smarkj LOCKSTAT_RECORD0(rw__upgrade, rw); 1073176017Sjeff } 1074157882Sjhb return (success); 1075157882Sjhb} 1076157882Sjhb 1077157882Sjhb/* 1078157882Sjhb * Downgrade a write lock into a single read lock. 1079157882Sjhb */ 1080157882Sjhbvoid 1081242515Sattilio__rw_downgrade(volatile uintptr_t *c, const char *file, int line) 1082157882Sjhb{ 1083242515Sattilio struct rwlock *rw; 1084157882Sjhb struct turnstile *ts; 1085157882Sjhb uintptr_t tid, v; 1086176017Sjeff int rwait, wwait; 1087157882Sjhb 1088228424Savg if (SCHEDULER_STOPPED()) 1089228424Savg return; 1090228424Savg 1091242515Sattilio rw = rwlock2rw(c); 1092242515Sattilio 1093169394Sjhb KASSERT(rw->rw_lock != RW_DESTROYED, 1094169394Sjhb ("rw_downgrade() of destroyed rwlock @ %s:%d", file, line)); 1095242515Sattilio __rw_assert(c, RA_WLOCKED | RA_NOTRECURSED, file, line); 1096171052Sattilio#ifndef INVARIANTS 1097171052Sattilio if (rw_recursed(rw)) 1098171052Sattilio panic("downgrade of a recursed lock"); 1099171052Sattilio#endif 1100157882Sjhb 1101167787Sjhb WITNESS_DOWNGRADE(&rw->lock_object, 0, file, line); 1102157882Sjhb 1103157882Sjhb /* 1104157882Sjhb * Convert from a writer to a single reader. First we handle 1105157882Sjhb * the easy case with no waiters. If there are any waiters, we 1106176017Sjeff * lock the turnstile and "disown" the lock. 1107157882Sjhb */ 1108157882Sjhb tid = (uintptr_t)curthread; 1109157882Sjhb if (atomic_cmpset_rel_ptr(&rw->rw_lock, tid, RW_READERS_LOCK(1))) 1110157882Sjhb goto out; 1111157882Sjhb 1112157882Sjhb /* 1113157882Sjhb * Ok, we think we have waiters, so lock the turnstile so we can 1114157882Sjhb * read the waiter flags without any races. 1115157882Sjhb */ 1116170295Sjeff turnstile_chain_lock(&rw->lock_object); 1117176017Sjeff v = rw->rw_lock & RW_LOCK_WAITERS; 1118176017Sjeff rwait = v & RW_LOCK_READ_WAITERS; 1119176017Sjeff wwait = v & RW_LOCK_WRITE_WAITERS; 1120176017Sjeff MPASS(rwait | wwait); 1121157882Sjhb 1122157882Sjhb /* 1123176017Sjeff * Downgrade from a write lock while preserving waiters flag 1124176017Sjeff * and give up ownership of the turnstile. 1125157882Sjhb */ 1126167787Sjhb ts = turnstile_lookup(&rw->lock_object); 1127157882Sjhb MPASS(ts != NULL); 1128176017Sjeff if (!wwait) 1129176017Sjeff v &= ~RW_LOCK_READ_WAITERS; 1130176017Sjeff atomic_store_rel_ptr(&rw->rw_lock, RW_READERS_LOCK(1) | v); 1131176017Sjeff /* 1132176017Sjeff * Wake other readers if there are no writers pending. Otherwise they 1133176017Sjeff * won't be able to acquire the lock anyway. 1134176017Sjeff */ 1135176017Sjeff if (rwait && !wwait) { 1136157882Sjhb turnstile_broadcast(ts, TS_SHARED_QUEUE); 1137157882Sjhb turnstile_unpend(ts, TS_EXCLUSIVE_LOCK); 1138176017Sjeff } else 1139157882Sjhb turnstile_disown(ts); 1140170295Sjeff turnstile_chain_unlock(&rw->lock_object); 1141157882Sjhbout: 1142176017Sjeff curthread->td_rw_rlocks++; 1143167787Sjhb LOCK_LOG_LOCK("WDOWNGRADE", &rw->lock_object, 0, 0, file, line); 1144285703Smarkj LOCKSTAT_RECORD0(rw__downgrade, rw); 1145157882Sjhb} 1146157882Sjhb 1147154941Sjhb#ifdef INVARIANT_SUPPORT 1148155162Sscottl#ifndef INVARIANTS 1149242515Sattilio#undef __rw_assert 1150154941Sjhb#endif 1151154941Sjhb 1152154941Sjhb/* 1153154941Sjhb * In the non-WITNESS case, rw_assert() can only detect that at least 1154154941Sjhb * *some* thread owns an rlock, but it cannot guarantee that *this* 1155154941Sjhb * thread owns an rlock. 1156154941Sjhb */ 1157154941Sjhbvoid 1158242515Sattilio__rw_assert(const volatile uintptr_t *c, int what, const char *file, int line) 1159154941Sjhb{ 1160242515Sattilio const struct rwlock *rw; 1161154941Sjhb 1162154941Sjhb if (panicstr != NULL) 1163154941Sjhb return; 1164242515Sattilio 1165242515Sattilio rw = rwlock2rw(c); 1166242515Sattilio 1167154941Sjhb switch (what) { 1168154941Sjhb case RA_LOCKED: 1169171052Sattilio case RA_LOCKED | RA_RECURSED: 1170171052Sattilio case RA_LOCKED | RA_NOTRECURSED: 1171154941Sjhb case RA_RLOCKED: 1172251323Sjhb case RA_RLOCKED | RA_RECURSED: 1173251323Sjhb case RA_RLOCKED | RA_NOTRECURSED: 1174154941Sjhb#ifdef WITNESS 1175167787Sjhb witness_assert(&rw->lock_object, what, file, line); 1176154941Sjhb#else 1177154941Sjhb /* 1178154941Sjhb * If some other thread has a write lock or we have one 1179154941Sjhb * and are asserting a read lock, fail. Also, if no one 1180154941Sjhb * has a lock at all, fail. 1181154941Sjhb */ 1182155061Sscottl if (rw->rw_lock == RW_UNLOCKED || 1183251323Sjhb (!(rw->rw_lock & RW_LOCK_READ) && (what & RA_RLOCKED || 1184157826Sjhb rw_wowner(rw) != curthread))) 1185154941Sjhb panic("Lock %s not %slocked @ %s:%d\n", 1186251323Sjhb rw->lock_object.lo_name, (what & RA_RLOCKED) ? 1187154941Sjhb "read " : "", file, line); 1188171052Sattilio 1189251323Sjhb if (!(rw->rw_lock & RW_LOCK_READ) && !(what & RA_RLOCKED)) { 1190171052Sattilio if (rw_recursed(rw)) { 1191171052Sattilio if (what & RA_NOTRECURSED) 1192171052Sattilio panic("Lock %s recursed @ %s:%d\n", 1193171052Sattilio rw->lock_object.lo_name, file, 1194171052Sattilio line); 1195171052Sattilio } else if (what & RA_RECURSED) 1196171052Sattilio panic("Lock %s not recursed @ %s:%d\n", 1197171052Sattilio rw->lock_object.lo_name, file, line); 1198171052Sattilio } 1199154941Sjhb#endif 1200154941Sjhb break; 1201154941Sjhb case RA_WLOCKED: 1202171052Sattilio case RA_WLOCKED | RA_RECURSED: 1203171052Sattilio case RA_WLOCKED | RA_NOTRECURSED: 1204157826Sjhb if (rw_wowner(rw) != curthread) 1205154941Sjhb panic("Lock %s not exclusively locked @ %s:%d\n", 1206167787Sjhb rw->lock_object.lo_name, file, line); 1207171052Sattilio if (rw_recursed(rw)) { 1208171052Sattilio if (what & RA_NOTRECURSED) 1209171052Sattilio panic("Lock %s recursed @ %s:%d\n", 1210171052Sattilio rw->lock_object.lo_name, file, line); 1211171052Sattilio } else if (what & RA_RECURSED) 1212171052Sattilio panic("Lock %s not recursed @ %s:%d\n", 1213171052Sattilio rw->lock_object.lo_name, file, line); 1214154941Sjhb break; 1215154941Sjhb case RA_UNLOCKED: 1216154941Sjhb#ifdef WITNESS 1217167787Sjhb witness_assert(&rw->lock_object, what, file, line); 1218154941Sjhb#else 1219154941Sjhb /* 1220154941Sjhb * If we hold a write lock fail. We can't reliably check 1221154941Sjhb * to see if we hold a read lock or not. 1222154941Sjhb */ 1223157826Sjhb if (rw_wowner(rw) == curthread) 1224154941Sjhb panic("Lock %s exclusively locked @ %s:%d\n", 1225167787Sjhb rw->lock_object.lo_name, file, line); 1226154941Sjhb#endif 1227154941Sjhb break; 1228154941Sjhb default: 1229154941Sjhb panic("Unknown rw lock assertion: %d @ %s:%d", what, file, 1230154941Sjhb line); 1231154941Sjhb } 1232154941Sjhb} 1233154941Sjhb#endif /* INVARIANT_SUPPORT */ 1234154941Sjhb 1235154941Sjhb#ifdef DDB 1236154941Sjhbvoid 1237227588Spjddb_show_rwlock(const struct lock_object *lock) 1238154941Sjhb{ 1239227588Spjd const struct rwlock *rw; 1240154941Sjhb struct thread *td; 1241154941Sjhb 1242227588Spjd rw = (const struct rwlock *)lock; 1243154941Sjhb 1244154941Sjhb db_printf(" state: "); 1245154941Sjhb if (rw->rw_lock == RW_UNLOCKED) 1246154941Sjhb db_printf("UNLOCKED\n"); 1247169394Sjhb else if (rw->rw_lock == RW_DESTROYED) { 1248169394Sjhb db_printf("DESTROYED\n"); 1249169394Sjhb return; 1250169394Sjhb } else if (rw->rw_lock & RW_LOCK_READ) 1251167504Sjhb db_printf("RLOCK: %ju locks\n", 1252167504Sjhb (uintmax_t)(RW_READERS(rw->rw_lock))); 1253154941Sjhb else { 1254157826Sjhb td = rw_wowner(rw); 1255154941Sjhb db_printf("WLOCK: %p (tid %d, pid %d, \"%s\")\n", td, 1256173600Sjulian td->td_tid, td->td_proc->p_pid, td->td_name); 1257171052Sattilio if (rw_recursed(rw)) 1258171052Sattilio db_printf(" recursed: %u\n", rw->rw_recurse); 1259154941Sjhb } 1260154941Sjhb db_printf(" waiters: "); 1261154941Sjhb switch (rw->rw_lock & (RW_LOCK_READ_WAITERS | RW_LOCK_WRITE_WAITERS)) { 1262154941Sjhb case RW_LOCK_READ_WAITERS: 1263154941Sjhb db_printf("readers\n"); 1264154941Sjhb break; 1265154941Sjhb case RW_LOCK_WRITE_WAITERS: 1266154941Sjhb db_printf("writers\n"); 1267154941Sjhb break; 1268154941Sjhb case RW_LOCK_READ_WAITERS | RW_LOCK_WRITE_WAITERS: 1269167492Sjhb db_printf("readers and writers\n"); 1270154941Sjhb break; 1271154941Sjhb default: 1272154941Sjhb db_printf("none\n"); 1273154941Sjhb break; 1274154941Sjhb } 1275154941Sjhb} 1276154941Sjhb 1277154941Sjhb#endif 1278