kern_mutex.c revision 278694
1181834Sroberto/*- 2181834Sroberto * Copyright (c) 1998 Berkeley Software Design, Inc. All rights reserved. 3181834Sroberto * 4181834Sroberto * Redistribution and use in source and binary forms, with or without 5181834Sroberto * modification, are permitted provided that the following conditions 6181834Sroberto * are met: 7181834Sroberto * 1. Redistributions of source code must retain the above copyright 8181834Sroberto * notice, this list of conditions and the following disclaimer. 9181834Sroberto * 2. Redistributions in binary form must reproduce the above copyright 10181834Sroberto * notice, this list of conditions and the following disclaimer in the 11181834Sroberto * documentation and/or other materials provided with the distribution. 12181834Sroberto * 3. Berkeley Software Design Inc's name may not be used to endorse or 13181834Sroberto * promote products derived from this software without specific prior 14181834Sroberto * written permission. 15181834Sroberto * 16181834Sroberto * THIS SOFTWARE IS PROVIDED BY BERKELEY SOFTWARE DESIGN INC ``AS IS'' AND 17181834Sroberto * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 18181834Sroberto * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 19181834Sroberto * ARE DISCLAIMED. IN NO EVENT SHALL BERKELEY SOFTWARE DESIGN INC BE LIABLE 20181834Sroberto * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 21181834Sroberto * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 22181834Sroberto * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 23181834Sroberto * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 24181834Sroberto * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 25181834Sroberto * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 26181834Sroberto * SUCH DAMAGE. 27181834Sroberto * 28181834Sroberto * from BSDI $Id: mutex_witness.c,v 1.1.2.20 2000/04/27 03:10:27 cp Exp $ 29181834Sroberto * and BSDI $Id: synch_machdep.c,v 2.3.2.39 2000/04/27 03:10:25 cp Exp $ 30181834Sroberto */ 31181834Sroberto 32181834Sroberto/* 33181834Sroberto * Machine independent bits of mutex implementation. 34181834Sroberto */ 35181834Sroberto 36181834Sroberto#include <sys/cdefs.h> 37181834Sroberto__FBSDID("$FreeBSD: stable/10/sys/kern/kern_mutex.c 278694 2015-02-13 19:06:22Z sbruno $"); 38181834Sroberto 39181834Sroberto#include "opt_adaptive_mutexes.h" 40181834Sroberto#include "opt_ddb.h" 41181834Sroberto#include "opt_global.h" 42285612Sdelphij#include "opt_hwpmc_hooks.h" 43181834Sroberto#include "opt_kdtrace.h" 44181834Sroberto#include "opt_sched.h" 45181834Sroberto 46181834Sroberto#include <sys/param.h> 47181834Sroberto#include <sys/systm.h> 48285612Sdelphij#include <sys/bus.h> 49181834Sroberto#include <sys/conf.h> 50181834Sroberto#include <sys/kdb.h> 51181834Sroberto#include <sys/kernel.h> 52181834Sroberto#include <sys/ktr.h> 53181834Sroberto#include <sys/lock.h> 54181834Sroberto#include <sys/malloc.h> 55181834Sroberto#include <sys/mutex.h> 56181834Sroberto#include <sys/proc.h> 57181834Sroberto#include <sys/resourcevar.h> 58181834Sroberto#include <sys/sched.h> 59181834Sroberto#include <sys/sbuf.h> 60181834Sroberto#include <sys/sysctl.h> 61181834Sroberto#include <sys/turnstile.h> 62181834Sroberto#include <sys/vmmeter.h> 63181834Sroberto#include <sys/lock_profile.h> 64181834Sroberto 65181834Sroberto#include <machine/atomic.h> 66181834Sroberto#include <machine/bus.h> 67181834Sroberto#include <machine/cpu.h> 68181834Sroberto 69181834Sroberto#include <ddb/ddb.h> 70181834Sroberto 71181834Sroberto#include <fs/devfs/devfs_int.h> 72181834Sroberto 73181834Sroberto#include <vm/vm.h> 74181834Sroberto#include <vm/vm_extern.h> 75181834Sroberto 76181834Sroberto#if defined(SMP) && !defined(NO_ADAPTIVE_MUTEXES) 77181834Sroberto#define ADAPTIVE_MUTEXES 78181834Sroberto#endif 79181834Sroberto 80181834Sroberto#ifdef HWPMC_HOOKS 81181834Sroberto#include <sys/pmckern.h> 82181834SrobertoPMC_SOFT_DEFINE( , , lock, failed); 83181834Sroberto#endif 84181834Sroberto 85181834Sroberto/* 86181834Sroberto * Return the mutex address when the lock cookie address is provided. 87181834Sroberto * This functionality assumes that struct mtx* have a member named mtx_lock. 88181834Sroberto */ 89181834Sroberto#define mtxlock2mtx(c) (__containerof(c, struct mtx, mtx_lock)) 90181834Sroberto 91181834Sroberto/* 92181834Sroberto * Internal utility macros. 93181834Sroberto */ 94181834Sroberto#define mtx_unowned(m) ((m)->mtx_lock == MTX_UNOWNED) 95181834Sroberto 96181834Sroberto#define mtx_destroyed(m) ((m)->mtx_lock == MTX_DESTROYED) 97181834Sroberto 98181834Sroberto#define mtx_owner(m) ((struct thread *)((m)->mtx_lock & ~MTX_FLAGMASK)) 99181834Sroberto 100181834Srobertostatic void assert_mtx(const struct lock_object *lock, int what); 101181834Sroberto#ifdef DDB 102181834Srobertostatic void db_show_mtx(const struct lock_object *lock); 103181834Sroberto#endif 104181834Srobertostatic void lock_mtx(struct lock_object *lock, uintptr_t how); 105181834Srobertostatic void lock_spin(struct lock_object *lock, uintptr_t how); 106181834Sroberto#ifdef KDTRACE_HOOKS 107181834Srobertostatic int owner_mtx(const struct lock_object *lock, 108181834Sroberto struct thread **owner); 109181834Sroberto#endif 110181834Srobertostatic uintptr_t unlock_mtx(struct lock_object *lock); 111181834Srobertostatic uintptr_t unlock_spin(struct lock_object *lock); 112181834Sroberto 113181834Sroberto/* 114181834Sroberto * Lock classes for sleep and spin mutexes. 115181834Sroberto */ 116181834Srobertostruct lock_class lock_class_mtx_sleep = { 117181834Sroberto .lc_name = "sleep mutex", 118181834Sroberto .lc_flags = LC_SLEEPLOCK | LC_RECURSABLE, 119181834Sroberto .lc_assert = assert_mtx, 120181834Sroberto#ifdef DDB 121181834Sroberto .lc_ddb_show = db_show_mtx, 122181834Sroberto#endif 123181834Sroberto .lc_lock = lock_mtx, 124181834Sroberto .lc_unlock = unlock_mtx, 125181834Sroberto#ifdef KDTRACE_HOOKS 126181834Sroberto .lc_owner = owner_mtx, 127181834Sroberto#endif 128181834Sroberto}; 129181834Srobertostruct lock_class lock_class_mtx_spin = { 130181834Sroberto .lc_name = "spin mutex", 131181834Sroberto .lc_flags = LC_SPINLOCK | LC_RECURSABLE, 132181834Sroberto .lc_assert = assert_mtx, 133181834Sroberto#ifdef DDB 134181834Sroberto .lc_ddb_show = db_show_mtx, 135181834Sroberto#endif 136181834Sroberto .lc_lock = lock_spin, 137181834Sroberto .lc_unlock = unlock_spin, 138181834Sroberto#ifdef KDTRACE_HOOKS 139181834Sroberto .lc_owner = owner_mtx, 140181834Sroberto#endif 141181834Sroberto}; 142181834Sroberto 143181834Sroberto/* 144181834Sroberto * System-wide mutexes 145181834Sroberto */ 146181834Srobertostruct mtx blocked_lock; 147181834Srobertostruct mtx Giant; 148181834Sroberto 149181834Srobertovoid 150181834Srobertoassert_mtx(const struct lock_object *lock, int what) 151181834Sroberto{ 152181834Sroberto 153181834Sroberto mtx_assert((const struct mtx *)lock, what); 154181834Sroberto} 155181834Sroberto 156181834Srobertovoid 157181834Srobertolock_mtx(struct lock_object *lock, uintptr_t how) 158181834Sroberto{ 159285612Sdelphij 160181834Sroberto mtx_lock((struct mtx *)lock); 161181834Sroberto} 162181834Sroberto 163181834Srobertovoid 164181834Srobertolock_spin(struct lock_object *lock, uintptr_t how) 165181834Sroberto{ 166181834Sroberto 167181834Sroberto panic("spin locks can only use msleep_spin"); 168181834Sroberto} 169181834Sroberto 170181834Srobertouintptr_t 171181834Srobertounlock_mtx(struct lock_object *lock) 172181834Sroberto{ 173181834Sroberto struct mtx *m; 174181834Sroberto 175181834Sroberto m = (struct mtx *)lock; 176181834Sroberto mtx_assert(m, MA_OWNED | MA_NOTRECURSED); 177181834Sroberto mtx_unlock(m); 178181834Sroberto return (0); 179181834Sroberto} 180181834Sroberto 181181834Srobertouintptr_t 182181834Srobertounlock_spin(struct lock_object *lock) 183181834Sroberto{ 184181834Sroberto 185181834Sroberto panic("spin locks can only use msleep_spin"); 186181834Sroberto} 187181834Sroberto 188181834Sroberto#ifdef KDTRACE_HOOKS 189181834Srobertoint 190181834Srobertoowner_mtx(const struct lock_object *lock, struct thread **owner) 191181834Sroberto{ 192181834Sroberto const struct mtx *m = (const struct mtx *)lock; 193181834Sroberto 194285612Sdelphij *owner = mtx_owner(m); 195285612Sdelphij return (mtx_unowned(m) == 0); 196181834Sroberto} 197181834Sroberto#endif 198181834Sroberto 199181834Sroberto/* 200181834Sroberto * Function versions of the inlined __mtx_* macros. These are used by 201181834Sroberto * modules and can also be called from assembly language if needed. 202181834Sroberto */ 203181834Srobertovoid 204181834Sroberto__mtx_lock_flags(volatile uintptr_t *c, int opts, const char *file, int line) 205181834Sroberto{ 206181834Sroberto struct mtx *m; 207285612Sdelphij 208181834Sroberto if (SCHEDULER_STOPPED()) 209181834Sroberto return; 210181834Sroberto 211285612Sdelphij m = mtxlock2mtx(c); 212181834Sroberto 213285612Sdelphij KASSERT(kdb_active != 0 || !TD_IS_IDLETHREAD(curthread), 214181834Sroberto ("mtx_lock() by idle thread %p on sleep mutex %s @ %s:%d", 215181834Sroberto curthread, m->lock_object.lo_name, file, line)); 216181834Sroberto KASSERT(m->mtx_lock != MTX_DESTROYED, 217181834Sroberto ("mtx_lock() of destroyed mutex @ %s:%d", file, line)); 218181834Sroberto KASSERT(LOCK_CLASS(&m->lock_object) == &lock_class_mtx_sleep, 219181834Sroberto ("mtx_lock() of spin mutex %s @ %s:%d", m->lock_object.lo_name, 220181834Sroberto file, line)); 221181834Sroberto WITNESS_CHECKORDER(&m->lock_object, (opts & ~MTX_RECURSE) | 222181834Sroberto LOP_NEWORDER | LOP_EXCLUSIVE, file, line, NULL); 223181834Sroberto 224181834Sroberto __mtx_lock(m, curthread, opts, file, line); 225181834Sroberto LOCK_LOG_LOCK("LOCK", &m->lock_object, opts, m->mtx_recurse, file, 226181834Sroberto line); 227181834Sroberto WITNESS_LOCK(&m->lock_object, (opts & ~MTX_RECURSE) | LOP_EXCLUSIVE, 228181834Sroberto file, line); 229181834Sroberto curthread->td_locks++; 230181834Sroberto} 231181834Sroberto 232181834Srobertovoid 233181834Sroberto__mtx_unlock_flags(volatile uintptr_t *c, int opts, const char *file, int line) 234181834Sroberto{ 235181834Sroberto struct mtx *m; 236181834Sroberto 237181834Sroberto if (SCHEDULER_STOPPED()) 238181834Sroberto return; 239181834Sroberto 240181834Sroberto m = mtxlock2mtx(c); 241181834Sroberto 242181834Sroberto KASSERT(m->mtx_lock != MTX_DESTROYED, 243181834Sroberto ("mtx_unlock() of destroyed mutex @ %s:%d", file, line)); 244181834Sroberto KASSERT(LOCK_CLASS(&m->lock_object) == &lock_class_mtx_sleep, 245181834Sroberto ("mtx_unlock() of spin mutex %s @ %s:%d", m->lock_object.lo_name, 246181834Sroberto file, line)); 247181834Sroberto WITNESS_UNLOCK(&m->lock_object, opts | LOP_EXCLUSIVE, file, line); 248181834Sroberto LOCK_LOG_LOCK("UNLOCK", &m->lock_object, opts, m->mtx_recurse, file, 249181834Sroberto line); 250181834Sroberto mtx_assert(m, MA_OWNED); 251181834Sroberto 252181834Sroberto if (m->mtx_recurse == 0) 253181834Sroberto LOCKSTAT_PROFILE_RELEASE_LOCK(LS_MTX_UNLOCK_RELEASE, m); 254181834Sroberto __mtx_unlock(m, curthread, opts, file, line); 255181834Sroberto curthread->td_locks--; 256181834Sroberto} 257181834Sroberto 258181834Srobertovoid 259181834Sroberto__mtx_lock_spin_flags(volatile uintptr_t *c, int opts, const char *file, 260181834Sroberto int line) 261181834Sroberto{ 262181834Sroberto struct mtx *m; 263181834Sroberto 264285612Sdelphij if (SCHEDULER_STOPPED()) 265181834Sroberto return; 266285612Sdelphij 267181834Sroberto m = mtxlock2mtx(c); 268285612Sdelphij 269181834Sroberto KASSERT(m->mtx_lock != MTX_DESTROYED, 270181834Sroberto ("mtx_lock_spin() of destroyed mutex @ %s:%d", file, line)); 271181834Sroberto KASSERT(LOCK_CLASS(&m->lock_object) == &lock_class_mtx_spin, 272181834Sroberto ("mtx_lock_spin() of sleep mutex %s @ %s:%d", 273181834Sroberto m->lock_object.lo_name, file, line)); 274181834Sroberto if (mtx_owned(m)) 275285612Sdelphij KASSERT((m->lock_object.lo_flags & LO_RECURSABLE) != 0 || 276181834Sroberto (opts & MTX_RECURSE) != 0, 277285612Sdelphij ("mtx_lock_spin: recursed on non-recursive mutex %s @ %s:%d\n", 278285612Sdelphij m->lock_object.lo_name, file, line)); 279285612Sdelphij opts &= ~MTX_RECURSE; 280181834Sroberto WITNESS_CHECKORDER(&m->lock_object, opts | LOP_NEWORDER | LOP_EXCLUSIVE, 281181834Sroberto file, line, NULL); 282181834Sroberto __mtx_lock_spin(m, curthread, opts, file, line); 283181834Sroberto LOCK_LOG_LOCK("LOCK", &m->lock_object, opts, m->mtx_recurse, file, 284181834Sroberto line); 285181834Sroberto WITNESS_LOCK(&m->lock_object, opts | LOP_EXCLUSIVE, file, line); 286181834Sroberto} 287181834Sroberto 288181834Srobertovoid 289181834Sroberto__mtx_unlock_spin_flags(volatile uintptr_t *c, int opts, const char *file, 290181834Sroberto int line) 291181834Sroberto{ 292181834Sroberto struct mtx *m; 293181834Sroberto 294181834Sroberto if (SCHEDULER_STOPPED()) 295181834Sroberto return; 296181834Sroberto 297181834Sroberto m = mtxlock2mtx(c); 298181834Sroberto 299181834Sroberto KASSERT(m->mtx_lock != MTX_DESTROYED, 300181834Sroberto ("mtx_unlock_spin() of destroyed mutex @ %s:%d", file, line)); 301181834Sroberto KASSERT(LOCK_CLASS(&m->lock_object) == &lock_class_mtx_spin, 302181834Sroberto ("mtx_unlock_spin() of sleep mutex %s @ %s:%d", 303181834Sroberto m->lock_object.lo_name, file, line)); 304181834Sroberto WITNESS_UNLOCK(&m->lock_object, opts | LOP_EXCLUSIVE, file, line); 305181834Sroberto LOCK_LOG_LOCK("UNLOCK", &m->lock_object, opts, m->mtx_recurse, file, 306285612Sdelphij line); 307181834Sroberto mtx_assert(m, MA_OWNED); 308285612Sdelphij 309285612Sdelphij __mtx_unlock_spin(m); 310285612Sdelphij} 311285612Sdelphij 312285612Sdelphij/* 313285612Sdelphij * The important part of mtx_trylock{,_flags}() 314285612Sdelphij * Tries to acquire lock `m.' If this function is called on a mutex that 315285612Sdelphij * is already owned, it will recursively acquire the lock. 316285612Sdelphij */ 317285612Sdelphijint 318285612Sdelphij_mtx_trylock_flags_(volatile uintptr_t *c, int opts, const char *file, int line) 319285612Sdelphij{ 320285612Sdelphij struct mtx *m; 321285612Sdelphij#ifdef LOCK_PROFILING 322285612Sdelphij uint64_t waittime = 0; 323285612Sdelphij int contested = 0; 324285612Sdelphij#endif 325181834Sroberto int rval; 326181834Sroberto 327181834Sroberto if (SCHEDULER_STOPPED()) 328181834Sroberto return (1); 329181834Sroberto 330181834Sroberto m = mtxlock2mtx(c); 331181834Sroberto 332181834Sroberto KASSERT(kdb_active != 0 || !TD_IS_IDLETHREAD(curthread), 333181834Sroberto ("mtx_trylock() by idle thread %p on sleep mutex %s @ %s:%d", 334181834Sroberto curthread, m->lock_object.lo_name, file, line)); 335181834Sroberto KASSERT(m->mtx_lock != MTX_DESTROYED, 336181834Sroberto ("mtx_trylock() of destroyed mutex @ %s:%d", file, line)); 337181834Sroberto KASSERT(LOCK_CLASS(&m->lock_object) == &lock_class_mtx_sleep, 338181834Sroberto ("mtx_trylock() of spin mutex %s @ %s:%d", m->lock_object.lo_name, 339181834Sroberto file, line)); 340181834Sroberto 341181834Sroberto if (mtx_owned(m) && ((m->lock_object.lo_flags & LO_RECURSABLE) != 0 || 342181834Sroberto (opts & MTX_RECURSE) != 0)) { 343181834Sroberto m->mtx_recurse++; 344181834Sroberto atomic_set_ptr(&m->mtx_lock, MTX_RECURSED); 345181834Sroberto rval = 1; 346181834Sroberto } else 347181834Sroberto rval = _mtx_obtain_lock(m, (uintptr_t)curthread); 348181834Sroberto opts &= ~MTX_RECURSE; 349181834Sroberto 350181834Sroberto LOCK_LOG_TRY("LOCK", &m->lock_object, opts, rval, file, line); 351181834Sroberto if (rval) { 352181834Sroberto WITNESS_LOCK(&m->lock_object, opts | LOP_EXCLUSIVE | LOP_TRYLOCK, 353181834Sroberto file, line); 354181834Sroberto curthread->td_locks++; 355181834Sroberto if (m->mtx_recurse == 0) 356181834Sroberto LOCKSTAT_PROFILE_OBTAIN_LOCK_SUCCESS(LS_MTX_LOCK_ACQUIRE, 357181834Sroberto m, contested, waittime, file, line); 358181834Sroberto 359181834Sroberto } 360181834Sroberto 361181834Sroberto return (rval); 362181834Sroberto} 363181834Sroberto 364181834Sroberto/* 365181834Sroberto * __mtx_lock_sleep: the tougher part of acquiring an MTX_DEF lock. 366181834Sroberto * 367181834Sroberto * We call this if the lock is either contested (i.e. we need to go to 368181834Sroberto * sleep waiting for it), or if we need to recurse on it. 369181834Sroberto */ 370181834Srobertovoid 371181834Sroberto__mtx_lock_sleep(volatile uintptr_t *c, uintptr_t tid, int opts, 372181834Sroberto const char *file, int line) 373181834Sroberto{ 374181834Sroberto struct mtx *m; 375181834Sroberto struct turnstile *ts; 376181834Sroberto uintptr_t v; 377181834Sroberto#ifdef ADAPTIVE_MUTEXES 378181834Sroberto volatile struct thread *owner; 379181834Sroberto#endif 380181834Sroberto#ifdef KTR 381181834Sroberto int cont_logged = 0; 382181834Sroberto#endif 383181834Sroberto#ifdef LOCK_PROFILING 384181834Sroberto int contested = 0; 385181834Sroberto uint64_t waittime = 0; 386181834Sroberto#endif 387285612Sdelphij#ifdef KDTRACE_HOOKS 388181834Sroberto uint64_t spin_cnt = 0; 389181834Sroberto uint64_t sleep_cnt = 0; 390181834Sroberto int64_t sleep_time = 0; 391181834Sroberto#endif 392181834Sroberto 393181834Sroberto if (SCHEDULER_STOPPED()) 394181834Sroberto return; 395181834Sroberto 396181834Sroberto m = mtxlock2mtx(c); 397181834Sroberto 398181834Sroberto if (mtx_owned(m)) { 399181834Sroberto KASSERT((m->lock_object.lo_flags & LO_RECURSABLE) != 0 || 400181834Sroberto (opts & MTX_RECURSE) != 0, 401181834Sroberto ("_mtx_lock_sleep: recursed on non-recursive mutex %s @ %s:%d\n", 402181834Sroberto m->lock_object.lo_name, file, line)); 403181834Sroberto opts &= ~MTX_RECURSE; 404181834Sroberto m->mtx_recurse++; 405181834Sroberto atomic_set_ptr(&m->mtx_lock, MTX_RECURSED); 406181834Sroberto if (LOCK_LOG_TEST(&m->lock_object, opts)) 407181834Sroberto CTR1(KTR_LOCK, "_mtx_lock_sleep: %p recursing", m); 408181834Sroberto return; 409181834Sroberto } 410181834Sroberto opts &= ~MTX_RECURSE; 411181834Sroberto 412181834Sroberto#ifdef HWPMC_HOOKS 413181834Sroberto PMC_SOFT_CALL( , , lock, failed); 414181834Sroberto#endif 415181834Sroberto lock_profile_obtain_lock_failed(&m->lock_object, 416181834Sroberto &contested, &waittime); 417181834Sroberto if (LOCK_LOG_TEST(&m->lock_object, opts)) 418181834Sroberto CTR4(KTR_LOCK, 419181834Sroberto "_mtx_lock_sleep: %s contested (lock=%p) at %s:%d", 420181834Sroberto m->lock_object.lo_name, (void *)m->mtx_lock, file, line); 421285612Sdelphij 422181834Sroberto while (!_mtx_obtain_lock(m, tid)) { 423181834Sroberto#ifdef KDTRACE_HOOKS 424181834Sroberto spin_cnt++; 425181834Sroberto#endif 426181834Sroberto#ifdef ADAPTIVE_MUTEXES 427181834Sroberto /* 428181834Sroberto * If the owner is running on another CPU, spin until the 429181834Sroberto * owner stops running or the state of the lock changes. 430181834Sroberto */ 431181834Sroberto v = m->mtx_lock; 432181834Sroberto if (v != MTX_UNOWNED) { 433181834Sroberto owner = (struct thread *)(v & ~MTX_FLAGMASK); 434181834Sroberto if (TD_IS_RUNNING(owner)) { 435181834Sroberto if (LOCK_LOG_TEST(&m->lock_object, 0)) 436181834Sroberto CTR3(KTR_LOCK, 437181834Sroberto "%s: spinning on %p held by %p", 438181834Sroberto __func__, m, owner); 439181834Sroberto KTR_STATE1(KTR_SCHED, "thread", 440181834Sroberto sched_tdname((struct thread *)tid), 441181834Sroberto "spinning", "lockname:\"%s\"", 442181834Sroberto m->lock_object.lo_name); 443181834Sroberto while (mtx_owner(m) == owner && 444285612Sdelphij TD_IS_RUNNING(owner)) { 445181834Sroberto cpu_spinwait(); 446181834Sroberto#ifdef KDTRACE_HOOKS 447181834Sroberto spin_cnt++; 448181834Sroberto#endif 449181834Sroberto } 450181834Sroberto KTR_STATE0(KTR_SCHED, "thread", 451181834Sroberto sched_tdname((struct thread *)tid), 452285612Sdelphij "running"); 453181834Sroberto continue; 454285612Sdelphij } 455181834Sroberto } 456181834Sroberto#endif 457181834Sroberto 458181834Sroberto ts = turnstile_trywait(&m->lock_object); 459181834Sroberto v = m->mtx_lock; 460181834Sroberto 461181834Sroberto /* 462181834Sroberto * Check if the lock has been released while spinning for 463181834Sroberto * the turnstile chain lock. 464181834Sroberto */ 465181834Sroberto if (v == MTX_UNOWNED) { 466181834Sroberto turnstile_cancel(ts); 467181834Sroberto continue; 468181834Sroberto } 469181834Sroberto 470181834Sroberto#ifdef ADAPTIVE_MUTEXES 471181834Sroberto /* 472181834Sroberto * The current lock owner might have started executing 473181834Sroberto * on another CPU (or the lock could have changed 474181834Sroberto * owners) while we were waiting on the turnstile 475181834Sroberto * chain lock. If so, drop the turnstile lock and try 476285612Sdelphij * again. 477181834Sroberto */ 478181834Sroberto owner = (struct thread *)(v & ~MTX_FLAGMASK); 479181834Sroberto if (TD_IS_RUNNING(owner)) { 480181834Sroberto turnstile_cancel(ts); 481181834Sroberto continue; 482181834Sroberto } 483181834Sroberto#endif 484181834Sroberto 485181834Sroberto /* 486181834Sroberto * If the mutex isn't already contested and a failure occurs 487181834Sroberto * setting the contested bit, the mutex was either released 488181834Sroberto * or the state of the MTX_RECURSED bit changed. 489181834Sroberto */ 490181834Sroberto if ((v & MTX_CONTESTED) == 0 && 491181834Sroberto !atomic_cmpset_ptr(&m->mtx_lock, v, v | MTX_CONTESTED)) { 492181834Sroberto turnstile_cancel(ts); 493181834Sroberto continue; 494181834Sroberto } 495181834Sroberto 496181834Sroberto /* 497181834Sroberto * We definitely must sleep for this lock. 498181834Sroberto */ 499181834Sroberto mtx_assert(m, MA_NOTOWNED); 500 501#ifdef KTR 502 if (!cont_logged) { 503 CTR6(KTR_CONTENTION, 504 "contention: %p at %s:%d wants %s, taken by %s:%d", 505 (void *)tid, file, line, m->lock_object.lo_name, 506 WITNESS_FILE(&m->lock_object), 507 WITNESS_LINE(&m->lock_object)); 508 cont_logged = 1; 509 } 510#endif 511 512 /* 513 * Block on the turnstile. 514 */ 515#ifdef KDTRACE_HOOKS 516 sleep_time -= lockstat_nsecs(); 517#endif 518 turnstile_wait(ts, mtx_owner(m), TS_EXCLUSIVE_QUEUE); 519#ifdef KDTRACE_HOOKS 520 sleep_time += lockstat_nsecs(); 521 sleep_cnt++; 522#endif 523 } 524#ifdef KTR 525 if (cont_logged) { 526 CTR4(KTR_CONTENTION, 527 "contention end: %s acquired by %p at %s:%d", 528 m->lock_object.lo_name, (void *)tid, file, line); 529 } 530#endif 531 LOCKSTAT_PROFILE_OBTAIN_LOCK_SUCCESS(LS_MTX_LOCK_ACQUIRE, m, contested, 532 waittime, file, line); 533#ifdef KDTRACE_HOOKS 534 if (sleep_time) 535 LOCKSTAT_RECORD1(LS_MTX_LOCK_BLOCK, m, sleep_time); 536 537 /* 538 * Only record the loops spinning and not sleeping. 539 */ 540 if (spin_cnt > sleep_cnt) 541 LOCKSTAT_RECORD1(LS_MTX_LOCK_SPIN, m, (spin_cnt - sleep_cnt)); 542#endif 543} 544 545static void 546_mtx_lock_spin_failed(struct mtx *m) 547{ 548 struct thread *td; 549 550 td = mtx_owner(m); 551 552 /* If the mutex is unlocked, try again. */ 553 if (td == NULL) 554 return; 555 556 printf( "spin lock %p (%s) held by %p (tid %d) too long\n", 557 m, m->lock_object.lo_name, td, td->td_tid); 558#ifdef WITNESS 559 witness_display_spinlock(&m->lock_object, td, printf); 560#endif 561 panic("spin lock held too long"); 562} 563 564#ifdef SMP 565/* 566 * _mtx_lock_spin_cookie: the tougher part of acquiring an MTX_SPIN lock. 567 * 568 * This is only called if we need to actually spin for the lock. Recursion 569 * is handled inline. 570 */ 571void 572_mtx_lock_spin_cookie(volatile uintptr_t *c, uintptr_t tid, int opts, 573 const char *file, int line) 574{ 575 struct mtx *m; 576 int i = 0; 577#ifdef LOCK_PROFILING 578 int contested = 0; 579 uint64_t waittime = 0; 580#endif 581 582 if (SCHEDULER_STOPPED()) 583 return; 584 585 m = mtxlock2mtx(c); 586 587 if (LOCK_LOG_TEST(&m->lock_object, opts)) 588 CTR1(KTR_LOCK, "_mtx_lock_spin: %p spinning", m); 589 KTR_STATE1(KTR_SCHED, "thread", sched_tdname((struct thread *)tid), 590 "spinning", "lockname:\"%s\"", m->lock_object.lo_name); 591 592#ifdef HWPMC_HOOKS 593 PMC_SOFT_CALL( , , lock, failed); 594#endif 595 lock_profile_obtain_lock_failed(&m->lock_object, &contested, &waittime); 596 while (!_mtx_obtain_lock(m, tid)) { 597 598 /* Give interrupts a chance while we spin. */ 599 spinlock_exit(); 600 while (m->mtx_lock != MTX_UNOWNED) { 601 if (i++ < 10000000) { 602 cpu_spinwait(); 603 continue; 604 } 605 if (i < 60000000 || kdb_active || panicstr != NULL) 606 DELAY(1); 607 else 608 _mtx_lock_spin_failed(m); 609 cpu_spinwait(); 610 } 611 spinlock_enter(); 612 } 613 614 if (LOCK_LOG_TEST(&m->lock_object, opts)) 615 CTR1(KTR_LOCK, "_mtx_lock_spin: %p spin done", m); 616 KTR_STATE0(KTR_SCHED, "thread", sched_tdname((struct thread *)tid), 617 "running"); 618 619 LOCKSTAT_PROFILE_OBTAIN_LOCK_SUCCESS(LS_MTX_SPIN_LOCK_ACQUIRE, m, 620 contested, waittime, (file), (line)); 621 LOCKSTAT_RECORD1(LS_MTX_SPIN_LOCK_SPIN, m, i); 622} 623#endif /* SMP */ 624 625void 626thread_lock_flags_(struct thread *td, int opts, const char *file, int line) 627{ 628 struct mtx *m; 629 uintptr_t tid; 630 int i; 631#ifdef LOCK_PROFILING 632 int contested = 0; 633 uint64_t waittime = 0; 634#endif 635#ifdef KDTRACE_HOOKS 636 uint64_t spin_cnt = 0; 637#endif 638 639 i = 0; 640 tid = (uintptr_t)curthread; 641 642 if (SCHEDULER_STOPPED()) 643 return; 644 645 for (;;) { 646retry: 647 spinlock_enter(); 648 m = td->td_lock; 649 KASSERT(m->mtx_lock != MTX_DESTROYED, 650 ("thread_lock() of destroyed mutex @ %s:%d", file, line)); 651 KASSERT(LOCK_CLASS(&m->lock_object) == &lock_class_mtx_spin, 652 ("thread_lock() of sleep mutex %s @ %s:%d", 653 m->lock_object.lo_name, file, line)); 654 if (mtx_owned(m)) 655 KASSERT((m->lock_object.lo_flags & LO_RECURSABLE) != 0, 656 ("thread_lock: recursed on non-recursive mutex %s @ %s:%d\n", 657 m->lock_object.lo_name, file, line)); 658 WITNESS_CHECKORDER(&m->lock_object, 659 opts | LOP_NEWORDER | LOP_EXCLUSIVE, file, line, NULL); 660 while (!_mtx_obtain_lock(m, tid)) { 661#ifdef KDTRACE_HOOKS 662 spin_cnt++; 663#endif 664 if (m->mtx_lock == tid) { 665 m->mtx_recurse++; 666 break; 667 } 668#ifdef HWPMC_HOOKS 669 PMC_SOFT_CALL( , , lock, failed); 670#endif 671 lock_profile_obtain_lock_failed(&m->lock_object, 672 &contested, &waittime); 673 /* Give interrupts a chance while we spin. */ 674 spinlock_exit(); 675 while (m->mtx_lock != MTX_UNOWNED) { 676 if (i++ < 10000000) 677 cpu_spinwait(); 678 else if (i < 60000000 || 679 kdb_active || panicstr != NULL) 680 DELAY(1); 681 else 682 _mtx_lock_spin_failed(m); 683 cpu_spinwait(); 684 if (m != td->td_lock) 685 goto retry; 686 } 687 spinlock_enter(); 688 } 689 if (m == td->td_lock) 690 break; 691 __mtx_unlock_spin(m); /* does spinlock_exit() */ 692#ifdef KDTRACE_HOOKS 693 spin_cnt++; 694#endif 695 } 696 if (m->mtx_recurse == 0) 697 LOCKSTAT_PROFILE_OBTAIN_LOCK_SUCCESS(LS_MTX_SPIN_LOCK_ACQUIRE, 698 m, contested, waittime, (file), (line)); 699 LOCK_LOG_LOCK("LOCK", &m->lock_object, opts, m->mtx_recurse, file, 700 line); 701 WITNESS_LOCK(&m->lock_object, opts | LOP_EXCLUSIVE, file, line); 702 LOCKSTAT_RECORD1(LS_THREAD_LOCK_SPIN, m, spin_cnt); 703} 704 705struct mtx * 706thread_lock_block(struct thread *td) 707{ 708 struct mtx *lock; 709 710 THREAD_LOCK_ASSERT(td, MA_OWNED); 711 lock = td->td_lock; 712 td->td_lock = &blocked_lock; 713 mtx_unlock_spin(lock); 714 715 return (lock); 716} 717 718void 719thread_lock_unblock(struct thread *td, struct mtx *new) 720{ 721 mtx_assert(new, MA_OWNED); 722 MPASS(td->td_lock == &blocked_lock); 723 atomic_store_rel_ptr((volatile void *)&td->td_lock, (uintptr_t)new); 724} 725 726void 727thread_lock_set(struct thread *td, struct mtx *new) 728{ 729 struct mtx *lock; 730 731 mtx_assert(new, MA_OWNED); 732 THREAD_LOCK_ASSERT(td, MA_OWNED); 733 lock = td->td_lock; 734 td->td_lock = new; 735 mtx_unlock_spin(lock); 736} 737 738/* 739 * __mtx_unlock_sleep: the tougher part of releasing an MTX_DEF lock. 740 * 741 * We are only called here if the lock is recursed or contested (i.e. we 742 * need to wake up a blocked thread). 743 */ 744void 745__mtx_unlock_sleep(volatile uintptr_t *c, int opts, const char *file, int line) 746{ 747 struct mtx *m; 748 struct turnstile *ts; 749 750 if (SCHEDULER_STOPPED()) 751 return; 752 753 m = mtxlock2mtx(c); 754 755 if (mtx_recursed(m)) { 756 if (--(m->mtx_recurse) == 0) 757 atomic_clear_ptr(&m->mtx_lock, MTX_RECURSED); 758 if (LOCK_LOG_TEST(&m->lock_object, opts)) 759 CTR1(KTR_LOCK, "_mtx_unlock_sleep: %p unrecurse", m); 760 return; 761 } 762 763 /* 764 * We have to lock the chain before the turnstile so this turnstile 765 * can be removed from the hash list if it is empty. 766 */ 767 turnstile_chain_lock(&m->lock_object); 768 ts = turnstile_lookup(&m->lock_object); 769 if (LOCK_LOG_TEST(&m->lock_object, opts)) 770 CTR1(KTR_LOCK, "_mtx_unlock_sleep: %p contested", m); 771 MPASS(ts != NULL); 772 turnstile_broadcast(ts, TS_EXCLUSIVE_QUEUE); 773 _mtx_release_lock_quick(m); 774 775 /* 776 * This turnstile is now no longer associated with the mutex. We can 777 * unlock the chain lock so a new turnstile may take it's place. 778 */ 779 turnstile_unpend(ts, TS_EXCLUSIVE_LOCK); 780 turnstile_chain_unlock(&m->lock_object); 781} 782 783/* 784 * All the unlocking of MTX_SPIN locks is done inline. 785 * See the __mtx_unlock_spin() macro for the details. 786 */ 787 788/* 789 * The backing function for the INVARIANTS-enabled mtx_assert() 790 */ 791#ifdef INVARIANT_SUPPORT 792void 793__mtx_assert(const volatile uintptr_t *c, int what, const char *file, int line) 794{ 795 const struct mtx *m; 796 797 if (panicstr != NULL || dumping) 798 return; 799 800 m = mtxlock2mtx(c); 801 802 switch (what) { 803 case MA_OWNED: 804 case MA_OWNED | MA_RECURSED: 805 case MA_OWNED | MA_NOTRECURSED: 806 if (!mtx_owned(m)) 807 panic("mutex %s not owned at %s:%d", 808 m->lock_object.lo_name, file, line); 809 if (mtx_recursed(m)) { 810 if ((what & MA_NOTRECURSED) != 0) 811 panic("mutex %s recursed at %s:%d", 812 m->lock_object.lo_name, file, line); 813 } else if ((what & MA_RECURSED) != 0) { 814 panic("mutex %s unrecursed at %s:%d", 815 m->lock_object.lo_name, file, line); 816 } 817 break; 818 case MA_NOTOWNED: 819 if (mtx_owned(m)) 820 panic("mutex %s owned at %s:%d", 821 m->lock_object.lo_name, file, line); 822 break; 823 default: 824 panic("unknown mtx_assert at %s:%d", file, line); 825 } 826} 827#endif 828 829/* 830 * The MUTEX_DEBUG-enabled mtx_validate() 831 * 832 * Most of these checks have been moved off into the LO_INITIALIZED flag 833 * maintained by the witness code. 834 */ 835#ifdef MUTEX_DEBUG 836 837void mtx_validate(struct mtx *); 838 839void 840mtx_validate(struct mtx *m) 841{ 842 843/* 844 * XXX: When kernacc() does not require Giant we can reenable this check 845 */ 846#ifdef notyet 847 /* 848 * Can't call kernacc() from early init386(), especially when 849 * initializing Giant mutex, because some stuff in kernacc() 850 * requires Giant itself. 851 */ 852 if (!cold) 853 if (!kernacc((caddr_t)m, sizeof(m), 854 VM_PROT_READ | VM_PROT_WRITE)) 855 panic("Can't read and write to mutex %p", m); 856#endif 857} 858#endif 859 860/* 861 * General init routine used by the MTX_SYSINIT() macro. 862 */ 863void 864mtx_sysinit(void *arg) 865{ 866 struct mtx_args *margs = arg; 867 868 mtx_init((struct mtx *)margs->ma_mtx, margs->ma_desc, NULL, 869 margs->ma_opts); 870} 871 872/* 873 * Mutex initialization routine; initialize lock `m' of type contained in 874 * `opts' with options contained in `opts' and name `name.' The optional 875 * lock type `type' is used as a general lock category name for use with 876 * witness. 877 */ 878void 879_mtx_init(volatile uintptr_t *c, const char *name, const char *type, int opts) 880{ 881 struct mtx *m; 882 struct lock_class *class; 883 int flags; 884 885 m = mtxlock2mtx(c); 886 887 MPASS((opts & ~(MTX_SPIN | MTX_QUIET | MTX_RECURSE | 888 MTX_NOWITNESS | MTX_DUPOK | MTX_NOPROFILE)) == 0); 889 ASSERT_ATOMIC_LOAD_PTR(m->mtx_lock, 890 ("%s: mtx_lock not aligned for %s: %p", __func__, name, 891 &m->mtx_lock)); 892 893#ifdef MUTEX_DEBUG 894 /* Diagnostic and error correction */ 895 mtx_validate(m); 896#endif 897 898 /* Determine lock class and lock flags. */ 899 if (opts & MTX_SPIN) 900 class = &lock_class_mtx_spin; 901 else 902 class = &lock_class_mtx_sleep; 903 flags = 0; 904 if (opts & MTX_QUIET) 905 flags |= LO_QUIET; 906 if (opts & MTX_RECURSE) 907 flags |= LO_RECURSABLE; 908 if ((opts & MTX_NOWITNESS) == 0) 909 flags |= LO_WITNESS; 910 if (opts & MTX_DUPOK) 911 flags |= LO_DUPOK; 912 if (opts & MTX_NOPROFILE) 913 flags |= LO_NOPROFILE; 914 915 /* Initialize mutex. */ 916 lock_init(&m->lock_object, class, name, type, flags); 917 918 m->mtx_lock = MTX_UNOWNED; 919 m->mtx_recurse = 0; 920} 921 922/* 923 * Remove lock `m' from all_mtx queue. We don't allow MTX_QUIET to be 924 * passed in as a flag here because if the corresponding mtx_init() was 925 * called with MTX_QUIET set, then it will already be set in the mutex's 926 * flags. 927 */ 928void 929_mtx_destroy(volatile uintptr_t *c) 930{ 931 struct mtx *m; 932 933 m = mtxlock2mtx(c); 934 935 if (!mtx_owned(m)) 936 MPASS(mtx_unowned(m)); 937 else { 938 MPASS((m->mtx_lock & (MTX_RECURSED|MTX_CONTESTED)) == 0); 939 940 /* Perform the non-mtx related part of mtx_unlock_spin(). */ 941 if (LOCK_CLASS(&m->lock_object) == &lock_class_mtx_spin) 942 spinlock_exit(); 943 else 944 curthread->td_locks--; 945 946 lock_profile_release_lock(&m->lock_object); 947 /* Tell witness this isn't locked to make it happy. */ 948 WITNESS_UNLOCK(&m->lock_object, LOP_EXCLUSIVE, __FILE__, 949 __LINE__); 950 } 951 952 m->mtx_lock = MTX_DESTROYED; 953 lock_destroy(&m->lock_object); 954} 955 956/* 957 * Intialize the mutex code and system mutexes. This is called from the MD 958 * startup code prior to mi_startup(). The per-CPU data space needs to be 959 * setup before this is called. 960 */ 961void 962mutex_init(void) 963{ 964 965 /* Setup turnstiles so that sleep mutexes work. */ 966 init_turnstiles(); 967 968 /* 969 * Initialize mutexes. 970 */ 971 mtx_init(&Giant, "Giant", NULL, MTX_DEF | MTX_RECURSE); 972 mtx_init(&blocked_lock, "blocked lock", NULL, MTX_SPIN); 973 blocked_lock.mtx_lock = 0xdeadc0de; /* Always blocked. */ 974 mtx_init(&proc0.p_mtx, "process lock", NULL, MTX_DEF | MTX_DUPOK); 975 mtx_init(&proc0.p_slock, "process slock", NULL, MTX_SPIN | MTX_RECURSE); 976 mtx_init(&devmtx, "cdev", NULL, MTX_DEF); 977 mtx_lock(&Giant); 978} 979 980#ifdef DDB 981void 982db_show_mtx(const struct lock_object *lock) 983{ 984 struct thread *td; 985 const struct mtx *m; 986 987 m = (const struct mtx *)lock; 988 989 db_printf(" flags: {"); 990 if (LOCK_CLASS(lock) == &lock_class_mtx_spin) 991 db_printf("SPIN"); 992 else 993 db_printf("DEF"); 994 if (m->lock_object.lo_flags & LO_RECURSABLE) 995 db_printf(", RECURSE"); 996 if (m->lock_object.lo_flags & LO_DUPOK) 997 db_printf(", DUPOK"); 998 db_printf("}\n"); 999 db_printf(" state: {"); 1000 if (mtx_unowned(m)) 1001 db_printf("UNOWNED"); 1002 else if (mtx_destroyed(m)) 1003 db_printf("DESTROYED"); 1004 else { 1005 db_printf("OWNED"); 1006 if (m->mtx_lock & MTX_CONTESTED) 1007 db_printf(", CONTESTED"); 1008 if (m->mtx_lock & MTX_RECURSED) 1009 db_printf(", RECURSED"); 1010 } 1011 db_printf("}\n"); 1012 if (!mtx_unowned(m) && !mtx_destroyed(m)) { 1013 td = mtx_owner(m); 1014 db_printf(" owner: %p (tid %d, pid %d, \"%s\")\n", td, 1015 td->td_tid, td->td_proc->p_pid, td->td_name); 1016 if (mtx_recursed(m)) 1017 db_printf(" recursed: %d\n", m->mtx_recurse); 1018 } 1019} 1020#endif 1021