kern_mutex.c revision 154098
1238730Sdelphij/*- 2294286Sdelphij * Copyright (c) 1998 Berkeley Software Design, Inc. All rights reserved. 3238730Sdelphij * 4238730Sdelphij * Redistribution and use in source and binary forms, with or without 5238730Sdelphij * modification, are permitted provided that the following conditions 6238730Sdelphij * are met: 7238730Sdelphij * 1. Redistributions of source code must retain the above copyright 8238730Sdelphij * notice, this list of conditions and the following disclaimer. 960786Sps * 2. Redistributions in binary form must reproduce the above copyright 1060786Sps * notice, this list of conditions and the following disclaimer in the 1160786Sps * documentation and/or other materials provided with the distribution. 1260786Sps * 3. Berkeley Software Design Inc's name may not be used to endorse or 1360786Sps * promote products derived from this software without specific prior 1460786Sps * written permission. 1560786Sps * 1660786Sps * THIS SOFTWARE IS PROVIDED BY BERKELEY SOFTWARE DESIGN INC ``AS IS'' AND 1760786Sps * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 1860786Sps * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 1960786Sps * ARE DISCLAIMED. IN NO EVENT SHALL BERKELEY SOFTWARE DESIGN INC BE LIABLE 2060786Sps * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 2160786Sps * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 2260786Sps * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 23161475Sdelphij * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 2460786Sps * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 2560786Sps * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 2660786Sps * SUCH DAMAGE. 2760786Sps * 2860786Sps * from BSDI $Id: mutex_witness.c,v 1.1.2.20 2000/04/27 03:10:27 cp Exp $ 2960786Sps * and BSDI $Id: synch_machdep.c,v 2.3.2.39 2000/04/27 03:10:25 cp Exp $ 3060786Sps */ 3160786Sps 32172468Sdelphij/* 3360786Sps * Machine independent bits of mutex implementation. 3460786Sps */ 3560786Sps 3660786Sps#include <sys/cdefs.h> 3760786Sps__FBSDID("$FreeBSD: head/sys/kern/kern_mutex.c 154098 2006-01-07 14:03:15Z jhb $"); 3860786Sps 39191930Sdelphij#include "opt_adaptive_mutexes.h" 40191930Sdelphij#include "opt_ddb.h" 41191930Sdelphij#include "opt_mprof.h" 42191930Sdelphij#include "opt_mutex_wake_all.h" 43191930Sdelphij#include "opt_sched.h" 44191930Sdelphij 4560786Sps#include <sys/param.h> 4660786Sps#include <sys/systm.h> 4760786Sps#include <sys/bus.h> 4860786Sps#include <sys/conf.h> 4960786Sps#include <sys/kdb.h> 50172468Sdelphij#include <sys/kernel.h> 51172468Sdelphij#include <sys/ktr.h> 52172468Sdelphij#include <sys/lock.h> 5360786Sps#include <sys/malloc.h> 5460786Sps#include <sys/mutex.h> 5560786Sps#include <sys/proc.h> 56172468Sdelphij#include <sys/resourcevar.h> 5760786Sps#include <sys/sched.h> 58172468Sdelphij#include <sys/sbuf.h> 59172468Sdelphij#include <sys/sysctl.h> 60172468Sdelphij#include <sys/turnstile.h> 6160786Sps#include <sys/vmmeter.h> 6260786Sps 6360786Sps#include <machine/atomic.h> 64294286Sdelphij#include <machine/bus.h> 65294286Sdelphij#include <machine/clock.h> 66294286Sdelphij#include <machine/cpu.h> 67294286Sdelphij 68294286Sdelphij#include <ddb/ddb.h> 69294286Sdelphij 70294286Sdelphij#include <fs/devfs/devfs_int.h> 71294286Sdelphij 72294286Sdelphij#include <vm/vm.h> 73294286Sdelphij#include <vm/vm_extern.h> 74294286Sdelphij 75294286Sdelphij/* 76294286Sdelphij * Force MUTEX_WAKE_ALL for now. 77294286Sdelphij * single thread wakeup needs fixes to avoid race conditions with 78294286Sdelphij * priority inheritance. 79294286Sdelphij */ 80294286Sdelphij#ifndef MUTEX_WAKE_ALL 81294286Sdelphij#define MUTEX_WAKE_ALL 8260786Sps#endif 8360786Sps 8460786Sps/* 85128345Stjr * Internal utility macros. 86128345Stjr */ 8760786Sps#define mtx_unowned(m) ((m)->mtx_lock == MTX_UNOWNED) 8860786Sps 8960786Sps#define mtx_owner(m) (mtx_unowned((m)) ? NULL \ 9060786Sps : (struct thread *)((m)->mtx_lock & MTX_FLAGMASK)) 9160786Sps 9260786Sps#ifdef DDB 9360786Spsstatic void db_show_mtx(struct lock_object *lock); 9460786Sps#endif 9560786Sps 9660786Sps/* 97128345Stjr * Lock classes for sleep and spin mutexes. 9860786Sps */ 9960786Spsstruct lock_class lock_class_mtx_sleep = { 10060786Sps "sleep mutex", 10160786Sps LC_SLEEPLOCK | LC_RECURSABLE, 10260786Sps#ifdef DDB 103128345Stjr db_show_mtx 10460786Sps#endif 10560786Sps}; 10660786Spsstruct lock_class lock_class_mtx_spin = { 10760786Sps "spin mutex", 10860786Sps LC_SPINLOCK | LC_RECURSABLE, 109128345Stjr#ifdef DDB 110128345Stjr db_show_mtx 11160786Sps#endif 11260786Sps}; 11360786Sps 11460786Sps/* 11560786Sps * System-wide mutexes 11660786Sps */ 11760786Spsstruct mtx sched_lock; 11860786Spsstruct mtx Giant; 11960786Sps 12060786Sps#ifdef MUTEX_PROFILING 12160786SpsSYSCTL_NODE(_debug, OID_AUTO, mutex, CTLFLAG_RD, NULL, "mutex debugging"); 12260786SpsSYSCTL_NODE(_debug_mutex, OID_AUTO, prof, CTLFLAG_RD, NULL, "mutex profiling"); 12360786Spsstatic int mutex_prof_enable = 0; 12460786SpsSYSCTL_INT(_debug_mutex_prof, OID_AUTO, enable, CTLFLAG_RW, 12560786Sps &mutex_prof_enable, 0, "Enable tracing of mutex holdtime"); 12660786Sps 12760786Spsstruct mutex_prof { 12860786Sps const char *name; 12960786Sps const char *file; 13060786Sps int line; 13160786Sps uintmax_t cnt_max; 13260786Sps uintmax_t cnt_tot; 13360786Sps uintmax_t cnt_cur; 134170256Sdelphij uintmax_t cnt_contest_holding; 13560786Sps uintmax_t cnt_contest_locking; 136170256Sdelphij struct mutex_prof *next; 13760786Sps}; 13860786Sps 13960786Sps/* 14060786Sps * mprof_buf is a static pool of profiling records to avoid possible 14160786Sps * reentrance of the memory allocation functions. 14260786Sps * 14360786Sps * Note: NUM_MPROF_BUFFERS must be smaller than MPROF_HASH_SIZE. 14460786Sps */ 14560786Sps#ifdef MPROF_BUFFERS 14660786Sps#define NUM_MPROF_BUFFERS MPROF_BUFFERS 14760786Sps#else 14860786Sps#define NUM_MPROF_BUFFERS 1000 14960786Sps#endif 15060786Spsstatic struct mutex_prof mprof_buf[NUM_MPROF_BUFFERS]; 15160786Spsstatic int first_free_mprof_buf; 15260786Sps#ifndef MPROF_HASH_SIZE 15360786Sps#define MPROF_HASH_SIZE 1009 154170256Sdelphij#endif 15560786Sps#if NUM_MPROF_BUFFERS >= MPROF_HASH_SIZE 15660786Sps#error MPROF_BUFFERS must be larger than MPROF_HASH_SIZE 15760786Sps#endif 15860786Spsstatic struct mutex_prof *mprof_hash[MPROF_HASH_SIZE]; 15960786Sps/* SWAG: sbuf size = avg stat. line size * number of locks */ 16060786Sps#define MPROF_SBUF_SIZE 256 * 400 16160786Sps 16260786Spsstatic int mutex_prof_acquisitions; 16360786SpsSYSCTL_INT(_debug_mutex_prof, OID_AUTO, acquisitions, CTLFLAG_RD, 16460786Sps &mutex_prof_acquisitions, 0, "Number of mutex acquistions recorded"); 16560786Spsstatic int mutex_prof_records; 16660786SpsSYSCTL_INT(_debug_mutex_prof, OID_AUTO, records, CTLFLAG_RD, 16760786Sps &mutex_prof_records, 0, "Number of profiling records"); 16860786Spsstatic int mutex_prof_maxrecords = NUM_MPROF_BUFFERS; 16960786SpsSYSCTL_INT(_debug_mutex_prof, OID_AUTO, maxrecords, CTLFLAG_RD, 17060786Sps &mutex_prof_maxrecords, 0, "Maximum number of profiling records"); 17160786Spsstatic int mutex_prof_rejected; 17260786SpsSYSCTL_INT(_debug_mutex_prof, OID_AUTO, rejected, CTLFLAG_RD, 17360786Sps &mutex_prof_rejected, 0, "Number of rejected profiling records"); 17460786Spsstatic int mutex_prof_hashsize = MPROF_HASH_SIZE; 17560786SpsSYSCTL_INT(_debug_mutex_prof, OID_AUTO, hashsize, CTLFLAG_RD, 17660786Sps &mutex_prof_hashsize, 0, "Hash size"); 17760786Spsstatic int mutex_prof_collisions = 0; 17860786SpsSYSCTL_INT(_debug_mutex_prof, OID_AUTO, collisions, CTLFLAG_RD, 17960786Sps &mutex_prof_collisions, 0, "Number of hash collisions"); 18060786Sps 18160786Sps/* 18260786Sps * mprof_mtx protects the profiling buffers and the hash. 18360786Sps */ 18460786Spsstatic struct mtx mprof_mtx; 18560786SpsMTX_SYSINIT(mprof, &mprof_mtx, "mutex profiling lock", MTX_SPIN | MTX_QUIET); 18660786Sps 18760786Spsstatic u_int64_t 18860786Spsnanoseconds(void) 18960786Sps{ 19060786Sps struct timespec tv; 19160786Sps 19260786Sps nanotime(&tv); 19360786Sps return (tv.tv_sec * (u_int64_t)1000000000 + tv.tv_nsec); 19460786Sps} 19560786Sps 19660786Spsstatic int 19760786Spsdump_mutex_prof_stats(SYSCTL_HANDLER_ARGS) 19860786Sps{ 19960786Sps struct sbuf *sb; 20060786Sps int error, i; 20160786Sps static int multiplier = 1; 20260786Sps 20360786Sps if (first_free_mprof_buf == 0) 20460786Sps return (SYSCTL_OUT(req, "No locking recorded", 20560786Sps sizeof("No locking recorded"))); 20660786Sps 20760786Spsretry_sbufops: 20860786Sps sb = sbuf_new(NULL, NULL, MPROF_SBUF_SIZE * multiplier, SBUF_FIXEDLEN); 20960786Sps sbuf_printf(sb, "\n%6s %12s %11s %5s %12s %12s %s\n", 21060786Sps "max", "total", "count", "avg", "cnt_hold", "cnt_lock", "name"); 21160786Sps /* 21260786Sps * XXX this spinlock seems to be by far the largest perpetrator 21360786Sps * of spinlock latency (1.6 msec on an Athlon1600 was recorded 21460786Sps * even before I pessimized it further by moving the average 21560786Sps * computation here). 21660786Sps */ 21760786Sps mtx_lock_spin(&mprof_mtx); 21860786Sps for (i = 0; i < first_free_mprof_buf; ++i) { 219191930Sdelphij sbuf_printf(sb, "%6ju %12ju %11ju %5ju %12ju %12ju %s:%d (%s)\n", 22060786Sps mprof_buf[i].cnt_max / 1000, 22160786Sps mprof_buf[i].cnt_tot / 1000, 222191930Sdelphij mprof_buf[i].cnt_cur, 22360786Sps mprof_buf[i].cnt_cur == 0 ? (uintmax_t)0 : 22460786Sps mprof_buf[i].cnt_tot / (mprof_buf[i].cnt_cur * 1000), 22560786Sps mprof_buf[i].cnt_contest_holding, 22660786Sps mprof_buf[i].cnt_contest_locking, 22760786Sps mprof_buf[i].file, mprof_buf[i].line, mprof_buf[i].name); 22860786Sps if (sbuf_overflowed(sb)) { 22960786Sps mtx_unlock_spin(&mprof_mtx); 23060786Sps sbuf_delete(sb); 23160786Sps multiplier++; 23260786Sps goto retry_sbufops; 23360786Sps } 23460786Sps } 23560786Sps mtx_unlock_spin(&mprof_mtx); 23660786Sps sbuf_finish(sb); 23760786Sps error = SYSCTL_OUT(req, sbuf_data(sb), sbuf_len(sb) + 1); 23860786Sps sbuf_delete(sb); 23960786Sps return (error); 24060786Sps} 24160786SpsSYSCTL_PROC(_debug_mutex_prof, OID_AUTO, stats, CTLTYPE_STRING | CTLFLAG_RD, 24260786Sps NULL, 0, dump_mutex_prof_stats, "A", "Mutex profiling statistics"); 24360786Sps 24460786Spsstatic int 24560786Spsreset_mutex_prof_stats(SYSCTL_HANDLER_ARGS) 24660786Sps{ 24760786Sps int error, v; 24860786Sps 24960786Sps if (first_free_mprof_buf == 0) 25060786Sps return (0); 25160786Sps 25260786Sps v = 0; 25360786Sps error = sysctl_handle_int(oidp, &v, 0, req); 25460786Sps if (error) 25560786Sps return (error); 25660786Sps if (req->newptr == NULL) 25760786Sps return (error); 25860786Sps if (v == 0) 25960786Sps return (0); 260191930Sdelphij 26160786Sps mtx_lock_spin(&mprof_mtx); 26260786Sps bzero(mprof_buf, sizeof(*mprof_buf) * first_free_mprof_buf); 263191930Sdelphij bzero(mprof_hash, sizeof(struct mtx *) * MPROF_HASH_SIZE); 26460786Sps first_free_mprof_buf = 0; 26560786Sps mtx_unlock_spin(&mprof_mtx); 26660786Sps return (0); 26760786Sps} 26860786SpsSYSCTL_PROC(_debug_mutex_prof, OID_AUTO, reset, CTLTYPE_INT | CTLFLAG_RW, 26960786Sps NULL, 0, reset_mutex_prof_stats, "I", "Reset mutex profiling statistics"); 27060786Sps#endif 27160786Sps 27260786Sps/* 27360786Sps * Function versions of the inlined __mtx_* macros. These are used by 27460786Sps * modules and can also be called from assembly language if needed. 27560786Sps */ 27660786Spsvoid 27760786Sps_mtx_lock_flags(struct mtx *m, int opts, const char *file, int line) 27860786Sps{ 27960786Sps 28060786Sps MPASS(curthread != NULL); 28160786Sps KASSERT(LO_CLASSINDEX(&m->mtx_object) == LOCK_CLASS_SLEEP_MUTEX, 28260786Sps ("mtx_lock() of spin mutex %s @ %s:%d", m->mtx_object.lo_name, 28360786Sps file, line)); 28460786Sps WITNESS_CHECKORDER(&m->mtx_object, opts | LOP_NEWORDER | LOP_EXCLUSIVE, 28560786Sps file, line); 28660786Sps _get_sleep_lock(m, curthread, opts, file, line); 28760786Sps LOCK_LOG_LOCK("LOCK", &m->mtx_object, opts, m->mtx_recurse, file, 28860786Sps line); 28960786Sps WITNESS_LOCK(&m->mtx_object, opts | LOP_EXCLUSIVE, file, line); 29060786Sps#ifdef MUTEX_PROFILING 29160786Sps /* don't reset the timer when/if recursing */ 29260786Sps if (m->mtx_acqtime == 0) { 29360786Sps m->mtx_filename = file; 29460786Sps m->mtx_lineno = line; 29560786Sps m->mtx_acqtime = mutex_prof_enable ? nanoseconds() : 0; 29660786Sps ++mutex_prof_acquisitions; 29760786Sps } 29860786Sps#endif 29960786Sps} 30060786Sps 30160786Spsvoid 302294286Sdelphij_mtx_unlock_flags(struct mtx *m, int opts, const char *file, int line) 303294286Sdelphij{ 304294286Sdelphij 30560786Sps MPASS(curthread != NULL); 30660786Sps KASSERT(LO_CLASSINDEX(&m->mtx_object) == LOCK_CLASS_SLEEP_MUTEX, 30760786Sps ("mtx_unlock() of spin mutex %s @ %s:%d", m->mtx_object.lo_name, 30860786Sps file, line)); 30960786Sps WITNESS_UNLOCK(&m->mtx_object, opts | LOP_EXCLUSIVE, file, line); 31060786Sps LOCK_LOG_LOCK("UNLOCK", &m->mtx_object, opts, m->mtx_recurse, file, 31160786Sps line); 31260786Sps mtx_assert(m, MA_OWNED); 313191930Sdelphij#ifdef MUTEX_PROFILING 31460786Sps if (m->mtx_acqtime != 0) { 31560786Sps static const char *unknown = "(unknown)"; 316191930Sdelphij struct mutex_prof *mpp; 31760786Sps u_int64_t acqtime, now; 31860786Sps const char *p, *q; 31960786Sps volatile u_int hash; 32060786Sps 321170256Sdelphij now = nanoseconds(); 322161475Sdelphij acqtime = m->mtx_acqtime; 323161475Sdelphij m->mtx_acqtime = 0; 324161475Sdelphij if (now <= acqtime) 32560786Sps goto out; 32660786Sps for (p = m->mtx_filename; 32760786Sps p != NULL && strncmp(p, "../", 3) == 0; p += 3) 32860786Sps /* nothing */ ; 32960786Sps if (p == NULL || *p == '\0') 330 p = unknown; 331 for (hash = m->mtx_lineno, q = p; *q != '\0'; ++q) 332 hash = (hash * 2 + *q) % MPROF_HASH_SIZE; 333 mtx_lock_spin(&mprof_mtx); 334 for (mpp = mprof_hash[hash]; mpp != NULL; mpp = mpp->next) 335 if (mpp->line == m->mtx_lineno && 336 strcmp(mpp->file, p) == 0) 337 break; 338 if (mpp == NULL) { 339 /* Just exit if we cannot get a trace buffer */ 340 if (first_free_mprof_buf >= NUM_MPROF_BUFFERS) { 341 ++mutex_prof_rejected; 342 goto unlock; 343 } 344 mpp = &mprof_buf[first_free_mprof_buf++]; 345 mpp->name = mtx_name(m); 346 mpp->file = p; 347 mpp->line = m->mtx_lineno; 348 mpp->next = mprof_hash[hash]; 349 if (mprof_hash[hash] != NULL) 350 ++mutex_prof_collisions; 351 mprof_hash[hash] = mpp; 352 ++mutex_prof_records; 353 } 354 /* 355 * Record if the mutex has been held longer now than ever 356 * before. 357 */ 358 if (now - acqtime > mpp->cnt_max) 359 mpp->cnt_max = now - acqtime; 360 mpp->cnt_tot += now - acqtime; 361 mpp->cnt_cur++; 362 /* 363 * There's a small race, really we should cmpxchg 364 * 0 with the current value, but that would bill 365 * the contention to the wrong lock instance if 366 * it followed this also. 367 */ 368 mpp->cnt_contest_holding += m->mtx_contest_holding; 369 m->mtx_contest_holding = 0; 370 mpp->cnt_contest_locking += m->mtx_contest_locking; 371 m->mtx_contest_locking = 0; 372unlock: 373 mtx_unlock_spin(&mprof_mtx); 374 } 375out: 376#endif 377 _rel_sleep_lock(m, curthread, opts, file, line); 378} 379 380void 381_mtx_lock_spin_flags(struct mtx *m, int opts, const char *file, int line) 382{ 383 384 MPASS(curthread != NULL); 385 KASSERT(LO_CLASSINDEX(&m->mtx_object) == LOCK_CLASS_SPIN_MUTEX, 386 ("mtx_lock_spin() of sleep mutex %s @ %s:%d", 387 m->mtx_object.lo_name, file, line)); 388 WITNESS_CHECKORDER(&m->mtx_object, opts | LOP_NEWORDER | LOP_EXCLUSIVE, 389 file, line); 390 _get_spin_lock(m, curthread, opts, file, line); 391 LOCK_LOG_LOCK("LOCK", &m->mtx_object, opts, m->mtx_recurse, file, 392 line); 393 WITNESS_LOCK(&m->mtx_object, opts | LOP_EXCLUSIVE, file, line); 394} 395 396void 397_mtx_unlock_spin_flags(struct mtx *m, int opts, const char *file, int line) 398{ 399 400 MPASS(curthread != NULL); 401 KASSERT(LO_CLASSINDEX(&m->mtx_object) == LOCK_CLASS_SPIN_MUTEX, 402 ("mtx_unlock_spin() of sleep mutex %s @ %s:%d", 403 m->mtx_object.lo_name, file, line)); 404 WITNESS_UNLOCK(&m->mtx_object, opts | LOP_EXCLUSIVE, file, line); 405 LOCK_LOG_LOCK("UNLOCK", &m->mtx_object, opts, m->mtx_recurse, file, 406 line); 407 mtx_assert(m, MA_OWNED); 408 _rel_spin_lock(m); 409} 410 411/* 412 * The important part of mtx_trylock{,_flags}() 413 * Tries to acquire lock `m.' If this function is called on a mutex that 414 * is already owned, it will recursively acquire the lock. 415 */ 416int 417_mtx_trylock(struct mtx *m, int opts, const char *file, int line) 418{ 419 int rval; 420 421 MPASS(curthread != NULL); 422 KASSERT(LO_CLASSINDEX(&m->mtx_object) == LOCK_CLASS_SLEEP_MUTEX, 423 ("mtx_trylock() of spin mutex %s @ %s:%d", m->mtx_object.lo_name, 424 file, line)); 425 426 if (mtx_owned(m) && (m->mtx_object.lo_flags & LO_RECURSABLE) != 0) { 427 m->mtx_recurse++; 428 atomic_set_ptr(&m->mtx_lock, MTX_RECURSED); 429 rval = 1; 430 } else 431 rval = _obtain_lock(m, (uintptr_t)curthread); 432 433 LOCK_LOG_TRY("LOCK", &m->mtx_object, opts, rval, file, line); 434 if (rval) 435 WITNESS_LOCK(&m->mtx_object, opts | LOP_EXCLUSIVE | LOP_TRYLOCK, 436 file, line); 437 438 return (rval); 439} 440 441/* 442 * _mtx_lock_sleep: the tougher part of acquiring an MTX_DEF lock. 443 * 444 * We call this if the lock is either contested (i.e. we need to go to 445 * sleep waiting for it), or if we need to recurse on it. 446 */ 447void 448_mtx_lock_sleep(struct mtx *m, uintptr_t tid, int opts, const char *file, 449 int line) 450{ 451#if defined(SMP) && !defined(NO_ADAPTIVE_MUTEXES) 452 struct thread *owner; 453#endif 454 uintptr_t v; 455#ifdef KTR 456 int cont_logged = 0; 457#endif 458#ifdef MUTEX_PROFILING 459 int contested; 460#endif 461 462 if (mtx_owned(m)) { 463 KASSERT((m->mtx_object.lo_flags & LO_RECURSABLE) != 0, 464 ("_mtx_lock_sleep: recursed on non-recursive mutex %s @ %s:%d\n", 465 m->mtx_object.lo_name, file, line)); 466 m->mtx_recurse++; 467 atomic_set_ptr(&m->mtx_lock, MTX_RECURSED); 468 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 469 CTR1(KTR_LOCK, "_mtx_lock_sleep: %p recursing", m); 470 return; 471 } 472 473 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 474 CTR4(KTR_LOCK, 475 "_mtx_lock_sleep: %s contested (lock=%p) at %s:%d", 476 m->mtx_object.lo_name, (void *)m->mtx_lock, file, line); 477 478#ifdef MUTEX_PROFILING 479 contested = 0; 480#endif 481 while (!_obtain_lock(m, tid)) { 482#ifdef MUTEX_PROFILING 483 contested = 1; 484 atomic_add_int(&m->mtx_contest_holding, 1); 485#endif 486 turnstile_lock(&m->mtx_object); 487 v = m->mtx_lock; 488 489 /* 490 * Check if the lock has been released while spinning for 491 * the turnstile chain lock. 492 */ 493 if (v == MTX_UNOWNED) { 494 turnstile_release(&m->mtx_object); 495 cpu_spinwait(); 496 continue; 497 } 498 499#ifdef MUTEX_WAKE_ALL 500 MPASS(v != MTX_CONTESTED); 501#else 502 /* 503 * The mutex was marked contested on release. This means that 504 * there are other threads blocked on it. Grab ownership of 505 * it and propagate its priority to the current thread if 506 * necessary. 507 */ 508 if (v == MTX_CONTESTED) { 509 m->mtx_lock = tid | MTX_CONTESTED; 510 turnstile_claim(&m->mtx_object); 511 break; 512 } 513#endif 514 515 /* 516 * If the mutex isn't already contested and a failure occurs 517 * setting the contested bit, the mutex was either released 518 * or the state of the MTX_RECURSED bit changed. 519 */ 520 if ((v & MTX_CONTESTED) == 0 && 521 !atomic_cmpset_ptr(&m->mtx_lock, v, v | MTX_CONTESTED)) { 522 turnstile_release(&m->mtx_object); 523 cpu_spinwait(); 524 continue; 525 } 526 527#if defined(SMP) && !defined(NO_ADAPTIVE_MUTEXES) 528 /* 529 * If the current owner of the lock is executing on another 530 * CPU, spin instead of blocking. 531 */ 532 owner = (struct thread *)(v & MTX_FLAGMASK); 533#ifdef ADAPTIVE_GIANT 534 if (TD_IS_RUNNING(owner)) { 535#else 536 if (m != &Giant && TD_IS_RUNNING(owner)) { 537#endif 538 turnstile_release(&m->mtx_object); 539 while (mtx_owner(m) == owner && TD_IS_RUNNING(owner)) { 540 cpu_spinwait(); 541 } 542 continue; 543 } 544#endif /* SMP && !NO_ADAPTIVE_MUTEXES */ 545 546 /* 547 * We definitely must sleep for this lock. 548 */ 549 mtx_assert(m, MA_NOTOWNED); 550 551#ifdef KTR 552 if (!cont_logged) { 553 CTR6(KTR_CONTENTION, 554 "contention: %p at %s:%d wants %s, taken by %s:%d", 555 (void *)tid, file, line, m->mtx_object.lo_name, 556 WITNESS_FILE(&m->mtx_object), 557 WITNESS_LINE(&m->mtx_object)); 558 cont_logged = 1; 559 } 560#endif 561 562 /* 563 * Block on the turnstile. 564 */ 565 turnstile_wait(&m->mtx_object, mtx_owner(m)); 566 } 567 568#ifdef KTR 569 if (cont_logged) { 570 CTR4(KTR_CONTENTION, 571 "contention end: %s acquired by %p at %s:%d", 572 m->mtx_object.lo_name, (void *)tid, file, line); 573 } 574#endif 575#ifdef MUTEX_PROFILING 576 if (contested) 577 m->mtx_contest_locking++; 578 m->mtx_contest_holding = 0; 579#endif 580 return; 581} 582 583#ifdef SMP 584/* 585 * _mtx_lock_spin: the tougher part of acquiring an MTX_SPIN lock. 586 * 587 * This is only called if we need to actually spin for the lock. Recursion 588 * is handled inline. 589 */ 590void 591_mtx_lock_spin(struct mtx *m, uintptr_t tid, int opts, const char *file, 592 int line) 593{ 594 int i = 0; 595 596 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 597 CTR1(KTR_LOCK, "_mtx_lock_spin: %p spinning", m); 598 599 for (;;) { 600 if (_obtain_lock(m, tid)) 601 break; 602 603 /* Give interrupts a chance while we spin. */ 604 spinlock_exit(); 605 while (m->mtx_lock != MTX_UNOWNED) { 606 if (i++ < 10000000) { 607 cpu_spinwait(); 608 continue; 609 } 610 if (i < 60000000) 611 DELAY(1); 612 else if (!kdb_active && !panicstr) { 613 printf("spin lock %s held by %p for > 5 seconds\n", 614 m->mtx_object.lo_name, (void *)m->mtx_lock); 615#ifdef WITNESS 616 witness_display_spinlock(&m->mtx_object, 617 mtx_owner(m)); 618#endif 619 panic("spin lock held too long"); 620 } 621 cpu_spinwait(); 622 } 623 spinlock_enter(); 624 } 625 626 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 627 CTR1(KTR_LOCK, "_mtx_lock_spin: %p spin done", m); 628 629 return; 630} 631#endif /* SMP */ 632 633/* 634 * _mtx_unlock_sleep: the tougher part of releasing an MTX_DEF lock. 635 * 636 * We are only called here if the lock is recursed or contested (i.e. we 637 * need to wake up a blocked thread). 638 */ 639void 640_mtx_unlock_sleep(struct mtx *m, int opts, const char *file, int line) 641{ 642 struct turnstile *ts; 643#ifndef PREEMPTION 644 struct thread *td, *td1; 645#endif 646 647 if (mtx_recursed(m)) { 648 if (--(m->mtx_recurse) == 0) 649 atomic_clear_ptr(&m->mtx_lock, MTX_RECURSED); 650 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 651 CTR1(KTR_LOCK, "_mtx_unlock_sleep: %p unrecurse", m); 652 return; 653 } 654 655 turnstile_lock(&m->mtx_object); 656 ts = turnstile_lookup(&m->mtx_object); 657 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 658 CTR1(KTR_LOCK, "_mtx_unlock_sleep: %p contested", m); 659 660#if defined(SMP) && !defined(NO_ADAPTIVE_MUTEXES) 661 if (ts == NULL) { 662 _release_lock_quick(m); 663 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 664 CTR1(KTR_LOCK, "_mtx_unlock_sleep: %p no sleepers", m); 665 turnstile_release(&m->mtx_object); 666 return; 667 } 668#else 669 MPASS(ts != NULL); 670#endif 671#ifndef PREEMPTION 672 /* XXX */ 673 td1 = turnstile_head(ts); 674#endif 675#ifdef MUTEX_WAKE_ALL 676 turnstile_broadcast(ts); 677 _release_lock_quick(m); 678#else 679 if (turnstile_signal(ts)) { 680 _release_lock_quick(m); 681 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 682 CTR1(KTR_LOCK, "_mtx_unlock_sleep: %p not held", m); 683 } else { 684 m->mtx_lock = MTX_CONTESTED; 685 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 686 CTR1(KTR_LOCK, "_mtx_unlock_sleep: %p still contested", 687 m); 688 } 689#endif 690 turnstile_unpend(ts); 691 692#ifndef PREEMPTION 693 /* 694 * XXX: This is just a hack until preemption is done. However, 695 * once preemption is done we need to either wrap the 696 * turnstile_signal() and release of the actual lock in an 697 * extra critical section or change the preemption code to 698 * always just set a flag and never do instant-preempts. 699 */ 700 td = curthread; 701 if (td->td_critnest > 0 || td1->td_priority >= td->td_priority) 702 return; 703 mtx_lock_spin(&sched_lock); 704 if (!TD_IS_RUNNING(td1)) { 705#ifdef notyet 706 if (td->td_ithd != NULL) { 707 struct ithd *it = td->td_ithd; 708 709 if (it->it_interrupted) { 710 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 711 CTR2(KTR_LOCK, 712 "_mtx_unlock_sleep: %p interrupted %p", 713 it, it->it_interrupted); 714 intr_thd_fixup(it); 715 } 716 } 717#endif 718 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 719 CTR2(KTR_LOCK, 720 "_mtx_unlock_sleep: %p switching out lock=%p", m, 721 (void *)m->mtx_lock); 722 723 mi_switch(SW_INVOL, NULL); 724 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 725 CTR2(KTR_LOCK, "_mtx_unlock_sleep: %p resuming lock=%p", 726 m, (void *)m->mtx_lock); 727 } 728 mtx_unlock_spin(&sched_lock); 729#endif 730 731 return; 732} 733 734/* 735 * All the unlocking of MTX_SPIN locks is done inline. 736 * See the _rel_spin_lock() macro for the details. 737 */ 738 739/* 740 * The backing function for the INVARIANTS-enabled mtx_assert() 741 */ 742#ifdef INVARIANT_SUPPORT 743void 744_mtx_assert(struct mtx *m, int what, const char *file, int line) 745{ 746 747 if (panicstr != NULL || dumping) 748 return; 749 switch (what) { 750 case MA_OWNED: 751 case MA_OWNED | MA_RECURSED: 752 case MA_OWNED | MA_NOTRECURSED: 753 if (!mtx_owned(m)) 754 panic("mutex %s not owned at %s:%d", 755 m->mtx_object.lo_name, file, line); 756 if (mtx_recursed(m)) { 757 if ((what & MA_NOTRECURSED) != 0) 758 panic("mutex %s recursed at %s:%d", 759 m->mtx_object.lo_name, file, line); 760 } else if ((what & MA_RECURSED) != 0) { 761 panic("mutex %s unrecursed at %s:%d", 762 m->mtx_object.lo_name, file, line); 763 } 764 break; 765 case MA_NOTOWNED: 766 if (mtx_owned(m)) 767 panic("mutex %s owned at %s:%d", 768 m->mtx_object.lo_name, file, line); 769 break; 770 default: 771 panic("unknown mtx_assert at %s:%d", file, line); 772 } 773} 774#endif 775 776/* 777 * The MUTEX_DEBUG-enabled mtx_validate() 778 * 779 * Most of these checks have been moved off into the LO_INITIALIZED flag 780 * maintained by the witness code. 781 */ 782#ifdef MUTEX_DEBUG 783 784void mtx_validate(struct mtx *); 785 786void 787mtx_validate(struct mtx *m) 788{ 789 790/* 791 * XXX: When kernacc() does not require Giant we can reenable this check 792 */ 793#ifdef notyet 794/* 795 * XXX - When kernacc() is fixed on the alpha to handle K0_SEG memory properly 796 * we can re-enable the kernacc() checks. 797 */ 798#ifndef __alpha__ 799 /* 800 * Can't call kernacc() from early init386(), especially when 801 * initializing Giant mutex, because some stuff in kernacc() 802 * requires Giant itself. 803 */ 804 if (!cold) 805 if (!kernacc((caddr_t)m, sizeof(m), 806 VM_PROT_READ | VM_PROT_WRITE)) 807 panic("Can't read and write to mutex %p", m); 808#endif 809#endif 810} 811#endif 812 813/* 814 * General init routine used by the MTX_SYSINIT() macro. 815 */ 816void 817mtx_sysinit(void *arg) 818{ 819 struct mtx_args *margs = arg; 820 821 mtx_init(margs->ma_mtx, margs->ma_desc, NULL, margs->ma_opts); 822} 823 824/* 825 * Mutex initialization routine; initialize lock `m' of type contained in 826 * `opts' with options contained in `opts' and name `name.' The optional 827 * lock type `type' is used as a general lock category name for use with 828 * witness. 829 */ 830void 831mtx_init(struct mtx *m, const char *name, const char *type, int opts) 832{ 833 struct lock_object *lock; 834 835 MPASS((opts & ~(MTX_SPIN | MTX_QUIET | MTX_RECURSE | 836 MTX_NOWITNESS | MTX_DUPOK)) == 0); 837 838#ifdef MUTEX_DEBUG 839 /* Diagnostic and error correction */ 840 mtx_validate(m); 841#endif 842 843 lock = &m->mtx_object; 844 KASSERT((lock->lo_flags & LO_INITIALIZED) == 0, 845 ("mutex \"%s\" %p already initialized", name, m)); 846 bzero(m, sizeof(*m)); 847 if (opts & MTX_SPIN) 848 lock->lo_flags = LOCK_CLASS_SPIN_MUTEX << LO_CLASSSHIFT; 849 else 850 lock->lo_flags = LOCK_CLASS_SLEEP_MUTEX << LO_CLASSSHIFT; 851 lock->lo_name = name; 852 lock->lo_type = type != NULL ? type : name; 853 if (opts & MTX_QUIET) 854 lock->lo_flags |= LO_QUIET; 855 if (opts & MTX_RECURSE) 856 lock->lo_flags |= LO_RECURSABLE; 857 if ((opts & MTX_NOWITNESS) == 0) 858 lock->lo_flags |= LO_WITNESS; 859 if (opts & MTX_DUPOK) 860 lock->lo_flags |= LO_DUPOK; 861 862 m->mtx_lock = MTX_UNOWNED; 863 864 LOCK_LOG_INIT(lock, opts); 865 866 WITNESS_INIT(lock); 867} 868 869/* 870 * Remove lock `m' from all_mtx queue. We don't allow MTX_QUIET to be 871 * passed in as a flag here because if the corresponding mtx_init() was 872 * called with MTX_QUIET set, then it will already be set in the mutex's 873 * flags. 874 */ 875void 876mtx_destroy(struct mtx *m) 877{ 878 879 LOCK_LOG_DESTROY(&m->mtx_object, 0); 880 881 if (!mtx_owned(m)) 882 MPASS(mtx_unowned(m)); 883 else { 884 MPASS((m->mtx_lock & (MTX_RECURSED|MTX_CONTESTED)) == 0); 885 886 /* Tell witness this isn't locked to make it happy. */ 887 WITNESS_UNLOCK(&m->mtx_object, LOP_EXCLUSIVE, __FILE__, 888 __LINE__); 889 } 890 891 WITNESS_DESTROY(&m->mtx_object); 892} 893 894/* 895 * Intialize the mutex code and system mutexes. This is called from the MD 896 * startup code prior to mi_startup(). The per-CPU data space needs to be 897 * setup before this is called. 898 */ 899void 900mutex_init(void) 901{ 902 903 /* Setup thread0 so that mutexes work. */ 904 LIST_INIT(&thread0.td_contested); 905 906 /* Setup turnstiles so that sleep mutexes work. */ 907 init_turnstiles(); 908 909 /* 910 * Initialize mutexes. 911 */ 912 mtx_init(&Giant, "Giant", NULL, MTX_DEF | MTX_RECURSE); 913 mtx_init(&sched_lock, "sched lock", NULL, MTX_SPIN | MTX_RECURSE); 914 mtx_init(&proc0.p_mtx, "process lock", NULL, MTX_DEF | MTX_DUPOK); 915 mtx_init(&devmtx, "cdev", NULL, MTX_DEF); 916 mtx_lock(&Giant); 917} 918 919#if LOCK_DEBUG > 0 || defined(DDB) 920/* XXX: This is not mutex-specific. */ 921struct lock_class *lock_classes[LOCK_CLASS_MAX + 1] = { 922 &lock_class_mtx_spin, 923 &lock_class_mtx_sleep, 924 &lock_class_sx, 925}; 926#endif 927 928#ifdef DDB 929/* XXX: This function is not mutex-specific. */ 930DB_SHOW_COMMAND(lock, db_show_lock) 931{ 932 struct lock_object *lock; 933 struct lock_class *class; 934 935 if (!have_addr) 936 return; 937 lock = (struct lock_object *)addr; 938 if (LO_CLASSINDEX(lock) > LOCK_CLASS_MAX) { 939 db_printf("Unknown lock class: %d\n", LO_CLASSINDEX(lock)); 940 return; 941 } 942 class = LOCK_CLASS(lock); 943 db_printf(" class: %s\n", class->lc_name); 944 db_printf(" name: %s\n", lock->lo_name); 945 if (lock->lo_type && lock->lo_type != lock->lo_name) 946 db_printf(" type: %s\n", lock->lo_type); 947 class->lc_ddb_show(lock); 948} 949 950void 951db_show_mtx(struct lock_object *lock) 952{ 953 struct thread *td; 954 struct mtx *m; 955 956 m = (struct mtx *)lock; 957 958 db_printf(" flags: {"); 959 if (LO_CLASSINDEX(lock) == LOCK_CLASS_SPIN_MUTEX) 960 db_printf("SPIN"); 961 else 962 db_printf("DEF"); 963 if (m->mtx_object.lo_flags & LO_RECURSABLE) 964 db_printf(", RECURSE"); 965 if (m->mtx_object.lo_flags & LO_DUPOK) 966 db_printf(", DUPOK"); 967 db_printf("}\n"); 968 db_printf(" state: {"); 969 if (mtx_unowned(m)) 970 db_printf("UNOWNED"); 971 else { 972 db_printf("OWNED"); 973 if (m->mtx_lock & MTX_CONTESTED) 974 db_printf(", CONTESTED"); 975 if (m->mtx_lock & MTX_RECURSED) 976 db_printf(", RECURSED"); 977 } 978 db_printf("}\n"); 979 if (!mtx_unowned(m)) { 980 td = mtx_owner(m); 981 db_printf(" owner: %p (tid %d, pid %d, \"%s\")\n", td, 982 td->td_tid, td->td_proc->p_pid, td->td_proc->p_comm); 983 if (mtx_recursed(m)) 984 db_printf(" recursed: %d\n", m->mtx_recurse); 985 } 986} 987#endif 988