kern_mutex.c revision 105644
11558Srgrimes/*- 21558Srgrimes * Copyright (c) 1998 Berkeley Software Design, Inc. All rights reserved. 31558Srgrimes * 41558Srgrimes * Redistribution and use in source and binary forms, with or without 51558Srgrimes * modification, are permitted provided that the following conditions 61558Srgrimes * are met: 71558Srgrimes * 1. Redistributions of source code must retain the above copyright 81558Srgrimes * notice, this list of conditions and the following disclaimer. 91558Srgrimes * 2. Redistributions in binary form must reproduce the above copyright 101558Srgrimes * notice, this list of conditions and the following disclaimer in the 111558Srgrimes * documentation and/or other materials provided with the distribution. 121558Srgrimes * 3. Berkeley Software Design Inc's name may not be used to endorse or 131558Srgrimes * promote products derived from this software without specific prior 141558Srgrimes * written permission. 151558Srgrimes * 161558Srgrimes * THIS SOFTWARE IS PROVIDED BY BERKELEY SOFTWARE DESIGN INC ``AS IS'' AND 171558Srgrimes * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 181558Srgrimes * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 191558Srgrimes * ARE DISCLAIMED. IN NO EVENT SHALL BERKELEY SOFTWARE DESIGN INC BE LIABLE 201558Srgrimes * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 211558Srgrimes * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 221558Srgrimes * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 231558Srgrimes * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 241558Srgrimes * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 251558Srgrimes * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 261558Srgrimes * SUCH DAMAGE. 271558Srgrimes * 281558Srgrimes * from BSDI $Id: mutex_witness.c,v 1.1.2.20 2000/04/27 03:10:27 cp Exp $ 291558Srgrimes * and BSDI $Id: synch_machdep.c,v 2.3.2.39 2000/04/27 03:10:25 cp Exp $ 301558Srgrimes * $FreeBSD: head/sys/kern/kern_mutex.c 105644 2002-10-21 18:48:28Z des $ 311558Srgrimes */ 321558Srgrimes 331558Srgrimes/* 34114589Sobrien * Machine independent bits of mutex implementation. 351558Srgrimes */ 3638040Scharnier 371558Srgrimes#include "opt_adaptive_mutexes.h" 381558Srgrimes#include "opt_ddb.h" 391558Srgrimes 401558Srgrimes#include <sys/param.h> 411558Srgrimes#include <sys/systm.h> 421558Srgrimes#include <sys/bus.h> 43114589Sobrien#include <sys/kernel.h> 4438040Scharnier#include <sys/ktr.h> 45114589Sobrien#include <sys/lock.h> 46114589Sobrien#include <sys/malloc.h> 471558Srgrimes#include <sys/mutex.h> 481558Srgrimes#include <sys/proc.h> 49102231Strhodes#include <sys/resourcevar.h> 501558Srgrimes#include <sys/sched.h> 511558Srgrimes#include <sys/sbuf.h> 5242873Sluoqi#include <sys/stdint.h> 5396478Sphk#include <sys/sysctl.h> 541558Srgrimes#include <sys/vmmeter.h> 551558Srgrimes 5698542Smckusick#include <machine/atomic.h> 5798542Smckusick#include <machine/bus.h> 581558Srgrimes#include <machine/clock.h> 591558Srgrimes#include <machine/cpu.h> 60110174Sgordon 611558Srgrimes#include <ddb/ddb.h> 621558Srgrimes 631558Srgrimes#include <vm/vm.h> 64109597Sjmallett#include <vm/vm_extern.h> 6538040Scharnier 661558Srgrimes/* 671558Srgrimes * Internal utility macros. 6858047Ssheldonh */ 691558Srgrimes#define mtx_unowned(m) ((m)->mtx_lock == MTX_UNOWNED) 701558Srgrimes 711558Srgrimes#define mtx_owner(m) (mtx_unowned((m)) ? NULL \ 721558Srgrimes : (struct thread *)((m)->mtx_lock & MTX_FLAGMASK)) 731558Srgrimes 74109597Sjmallett/* XXXKSE This test will change. */ 75109597Sjmallett#define thread_running(td) \ 761558Srgrimes ((td)->td_kse != NULL && (td)->td_kse->ke_oncpu != NOCPU) 7792883Simp 7892883Simp/* 791558Srgrimes * Lock classes for sleep and spin mutexes. 801558Srgrimes */ 81109597Sjmallettstruct lock_class lock_class_mtx_sleep = { 821558Srgrimes "sleep mutex", 83109963Sjmallett LC_SLEEPLOCK | LC_RECURSABLE 8479750Sdd}; 85110174Sgordonstruct lock_class lock_class_mtx_spin = { 86105120Srwatson "spin mutex", 8775377Smckusick LC_SPINLOCK | LC_RECURSABLE 88103005Sphk}; 8975377Smckusick 90127441Sbde/* 91110174Sgordon * System-wide mutexes 9279750Sdd */ 9342873Sluoqistruct mtx sched_lock; 9442873Sluoqistruct mtx Giant; 951558Srgrimes 96127441Sbde/* 97127441Sbde * Prototypes for non-exported routines. 98127441Sbde */ 99111287Srustatic void propagate_priority(struct thread *); 100127441Sbde 101127441Sbdestatic void 102127441Sbdepropagate_priority(struct thread *td) 103127441Sbde{ 104127441Sbde int pri = td->td_priority; 105127441Sbde struct mtx *m = td->td_blocked; 106127441Sbde 107127441Sbde mtx_assert(&sched_lock, MA_OWNED); 108127441Sbde for (;;) { 109127441Sbde struct thread *td1; 110127441Sbde 111127441Sbde td = mtx_owner(m); 112127441Sbde 113127441Sbde if (td == NULL) { 114127441Sbde /* 115127441Sbde * This really isn't quite right. Really 116127441Sbde * ought to bump priority of thread that 117127441Sbde * next acquires the mutex. 118127441Sbde */ 119127441Sbde MPASS(m->mtx_lock == MTX_CONTESTED); 120127441Sbde return; 121127441Sbde } 122127441Sbde 123127441Sbde MPASS(td->td_proc != NULL); 124127441Sbde MPASS(td->td_proc->p_magic == P_MAGIC); 125127441Sbde KASSERT(!TD_IS_SLEEPING(td), ("sleeping thread owns a mutex")); 126127441Sbde if (td->td_priority <= pri) /* lower is higher priority */ 127127441Sbde return; 128127441Sbde 129127441Sbde 130127441Sbde /* 131127441Sbde * If lock holder is actually running, just bump priority. 132127441Sbde */ 133127441Sbde if (TD_IS_RUNNING(td)) { 134127441Sbde td->td_priority = pri; 135127441Sbde return; 136127441Sbde } 137127441Sbde 138127441Sbde#ifndef SMP 139127441Sbde /* 140127441Sbde * For UP, we check to see if td is curthread (this shouldn't 141127441Sbde * ever happen however as it would mean we are in a deadlock.) 142127441Sbde */ 143127441Sbde KASSERT(td != curthread, ("Deadlock detected")); 144127441Sbde#endif 145127441Sbde 146127441Sbde /* 147127441Sbde * If on run queue move to new run queue, and quit. 148127441Sbde * XXXKSE this gets a lot more complicated under threads 149127441Sbde * but try anyhow. 150127441Sbde */ 151127441Sbde if (TD_ON_RUNQ(td)) { 152127441Sbde MPASS(td->td_blocked == NULL); 153127441Sbde sched_prio(td, pri); 154127441Sbde return; 155127441Sbde } 156127441Sbde /* 157127441Sbde * Adjust for any other cases. 158127441Sbde */ 159127441Sbde td->td_priority = pri; 160127441Sbde 161127441Sbde /* 162127441Sbde * If we aren't blocked on a mutex, we should be. 163127441Sbde */ 164127441Sbde KASSERT(TD_ON_LOCK(td), ( 165127441Sbde "process %d(%s):%d holds %s but isn't blocked on a mutex\n", 166127441Sbde td->td_proc->p_pid, td->td_proc->p_comm, td->td_state, 167127441Sbde m->mtx_object.lo_name)); 168127441Sbde 169127441Sbde /* 170127441Sbde * Pick up the mutex that td is blocked on. 171127441Sbde */ 172127441Sbde m = td->td_blocked; 173127441Sbde MPASS(m != NULL); 174127441Sbde 175127441Sbde /* 176127441Sbde * Check if the thread needs to be moved up on 177127441Sbde * the blocked chain 178127441Sbde */ 179127441Sbde if (td == TAILQ_FIRST(&m->mtx_blocked)) { 180127441Sbde continue; 181127441Sbde } 182127441Sbde 183127441Sbde td1 = TAILQ_PREV(td, threadqueue, td_lockq); 184127441Sbde if (td1->td_priority <= pri) { 185127441Sbde continue; 186127441Sbde } 187127441Sbde 188127441Sbde /* 189127441Sbde * Remove thread from blocked chain and determine where 190127441Sbde * it should be moved up to. Since we know that td1 has 191127441Sbde * a lower priority than td, we know that at least one 192127441Sbde * thread in the chain has a lower priority and that 193127441Sbde * td1 will thus not be NULL after the loop. 194127441Sbde */ 195127441Sbde TAILQ_REMOVE(&m->mtx_blocked, td, td_lockq); 196127441Sbde TAILQ_FOREACH(td1, &m->mtx_blocked, td_lockq) { 197127441Sbde MPASS(td1->td_proc->p_magic == P_MAGIC); 198127441Sbde if (td1->td_priority > pri) 199127441Sbde break; 200127441Sbde } 201127441Sbde 202127441Sbde MPASS(td1 != NULL); 203127441Sbde TAILQ_INSERT_BEFORE(td1, td, td_lockq); 204127441Sbde CTR4(KTR_LOCK, 205127441Sbde "propagate_priority: p %p moved before %p on [%p] %s", 206127441Sbde td, td1, m, m->mtx_object.lo_name); 207127441Sbde } 208127441Sbde} 209127441Sbde 210127441Sbde#ifdef MUTEX_PROFILING 211127441SbdeSYSCTL_NODE(_debug, OID_AUTO, mutex, CTLFLAG_RD, NULL, "mutex debugging"); 212127441SbdeSYSCTL_NODE(_debug_mutex, OID_AUTO, prof, CTLFLAG_RD, NULL, "mutex profiling"); 213127441Sbdestatic int mutex_prof_enable = 0; 214127441SbdeSYSCTL_INT(_debug_mutex_prof, OID_AUTO, enable, CTLFLAG_RW, 215127441Sbde &mutex_prof_enable, 0, "Enable tracing of mutex holdtime"); 216127441Sbde 217127441Sbdestruct mutex_prof { 218127441Sbde const char *name; 219127441Sbde const char *file; 220127441Sbde int line; 221127441Sbde /* 222127441Sbde * XXX should use specialized struct members instead of an array 223105120Srwatson * and these silly #defines. 22469314Scharnier */ 22569314Scharnier#define MPROF_MAX 0 22669314Scharnier#define MPROF_TOT 1 227127441Sbde#define MPROF_CNT 2 22869314Scharnier uintmax_t counter[3]; 229109963Sjmallett struct mutex_prof *next; 230109963Sjmallett}; 231109963Sjmallett 232109963Sjmallett/* 233109963Sjmallett * mprof_buf is a static pool of profiling records to avoid possible 23469829Scharnier * reentrance of the memory allocation functions. 235109963Sjmallett * 23669829Scharnier * Note: NUM_MPROF_BUFFERS must be smaller than MPROF_HASH_SIZE. 23769829Scharnier */ 23869829Scharnier#define NUM_MPROF_BUFFERS 1000 23969829Scharnierstatic struct mutex_prof mprof_buf[NUM_MPROF_BUFFERS]; 24069829Scharnierstatic int first_free_mprof_buf; 24169829Scharnier#define MPROF_HASH_SIZE 1009 24269829Scharnierstatic struct mutex_prof *mprof_hash[MPROF_HASH_SIZE]; 243110174Sgordon 244110174Sgordonstatic int mutex_prof_acquisitions; 245110174SgordonSYSCTL_INT(_debug_mutex_prof, OID_AUTO, acquisitions, CTLFLAG_RD, 246110174Sgordon &mutex_prof_acquisitions, 0, "Number of mutex acquistions recorded"); 247105120Srwatsonstatic int mutex_prof_records; 248105120SrwatsonSYSCTL_INT(_debug_mutex_prof, OID_AUTO, records, CTLFLAG_RD, 249105120Srwatson &mutex_prof_records, 0, "Number of profiling records"); 250105120Srwatsonstatic int mutex_prof_maxrecords = NUM_MPROF_BUFFERS; 251105120SrwatsonSYSCTL_INT(_debug_mutex_prof, OID_AUTO, maxrecords, CTLFLAG_RD, 252105120Srwatson &mutex_prof_maxrecords, 0, "Maximum number of profiling records"); 253105120Srwatsonstatic int mutex_prof_rejected; 254105120SrwatsonSYSCTL_INT(_debug_mutex_prof, OID_AUTO, rejected, CTLFLAG_RD, 255105120Srwatson &mutex_prof_rejected, 0, "Number of rejected profiling records"); 256105120Srwatsonstatic int mutex_prof_hashsize = MPROF_HASH_SIZE; 257105120SrwatsonSYSCTL_INT(_debug_mutex_prof, OID_AUTO, hashsize, CTLFLAG_RD, 258105120Srwatson &mutex_prof_hashsize, 0, "Hash size"); 259105120Srwatsonstatic int mutex_prof_collisions = 0; 260105120SrwatsonSYSCTL_INT(_debug_mutex_prof, OID_AUTO, collisions, CTLFLAG_RD, 261105120Srwatson &mutex_prof_collisions, 0, "Number of hash collisions"); 262105120Srwatson 263105206Srwatson/* 264105120Srwatson * mprof_mtx protects the profiling buffers and the hash. 265105120Srwatson */ 266105120Srwatsonstatic struct mtx mprof_mtx; 26769829ScharnierMTX_SYSINIT(mprof, &mprof_mtx, "mutex profiling lock", MTX_SPIN | MTX_QUIET); 26869829Scharnier 26969829Scharnierstatic u_int64_t 27069829Scharniernanoseconds(void) 27169829Scharnier{ 27269829Scharnier struct timespec tv; 27369829Scharnier 27469829Scharnier nanotime(&tv); 27569829Scharnier return (tv.tv_sec * (u_int64_t)1000000000 + tv.tv_nsec); 27669829Scharnier} 27769829Scharnier 27875377Smckusickstatic int 27975377Smckusickdump_mutex_prof_stats(SYSCTL_HANDLER_ARGS) 28075377Smckusick{ 28175377Smckusick struct sbuf *sb; 28275377Smckusick int error, i; 28375377Smckusick 28475377Smckusick if (first_free_mprof_buf == 0) 28575377Smckusick return (SYSCTL_OUT(req, "No locking recorded", 28675377Smckusick sizeof("No locking recorded"))); 28775377Smckusick 28875377Smckusick sb = sbuf_new(NULL, NULL, 1024, SBUF_AUTOEXTEND); 289105120Srwatson sbuf_printf(sb, "%6s %12s %11s %5s %s\n", 290105120Srwatson "max", "total", "count", "avg", "name"); 291105120Srwatson /* 292105120Srwatson * XXX this spinlock seems to be by far the largest perpetrator 293105120Srwatson * of spinlock latency (1.6 msec on an Athlon1600 was recorded 294105120Srwatson * even before I pessimized it further by moving the average 295105120Srwatson * computation here). 296105120Srwatson */ 297105120Srwatson mtx_lock_spin(&mprof_mtx); 298105120Srwatson for (i = 0; i < first_free_mprof_buf; ++i) 299105120Srwatson sbuf_printf(sb, "%6ju %12ju %11ju %5ju %s:%d (%s)\n", 300105120Srwatson mprof_buf[i].counter[MPROF_MAX] / 1000, 301105120Srwatson mprof_buf[i].counter[MPROF_TOT] / 1000, 302105120Srwatson mprof_buf[i].counter[MPROF_CNT], 303105120Srwatson mprof_buf[i].counter[MPROF_CNT] == 0 ? (uintmax_t)0 : 304105120Srwatson mprof_buf[i].counter[MPROF_TOT] / 305105206Srwatson (mprof_buf[i].counter[MPROF_CNT] * 1000), 306105120Srwatson mprof_buf[i].file, mprof_buf[i].line, mprof_buf[i].name); 307105120Srwatson mtx_unlock_spin(&mprof_mtx); 308105120Srwatson sbuf_finish(sb); 30969829Scharnier error = SYSCTL_OUT(req, sbuf_data(sb), sbuf_len(sb) + 1); 31069829Scharnier sbuf_delete(sb); 31169829Scharnier return (error); 31269829Scharnier} 31369829ScharnierSYSCTL_PROC(_debug_mutex_prof, OID_AUTO, stats, CTLTYPE_STRING | CTLFLAG_RD, 31469829Scharnier NULL, 0, dump_mutex_prof_stats, "A", "Mutex profiling statistics"); 31569829Scharnier#endif 31669829Scharnier 31769829Scharnier/* 31869829Scharnier * Function versions of the inlined __mtx_* macros. These are used by 31969829Scharnier * modules and can also be called from assembly language if needed. 32069829Scharnier */ 32169829Scharniervoid 32269829Scharnier_mtx_lock_flags(struct mtx *m, int opts, const char *file, int line) 32369829Scharnier{ 32469829Scharnier 32569829Scharnier MPASS(curthread != NULL); 32669829Scharnier KASSERT(m->mtx_object.lo_class == &lock_class_mtx_sleep, 32775498Smckusick ("mtx_lock() of spin mutex %s @ %s:%d", m->mtx_object.lo_name, 32869829Scharnier file, line)); 32975498Smckusick _get_sleep_lock(m, curthread, opts, file, line); 33075498Smckusick LOCK_LOG_LOCK("LOCK", &m->mtx_object, opts, m->mtx_recurse, file, 33175498Smckusick line); 33269829Scharnier WITNESS_LOCK(&m->mtx_object, opts | LOP_EXCLUSIVE, file, line); 33369829Scharnier#ifdef MUTEX_PROFILING 33469829Scharnier /* don't reset the timer when/if recursing */ 33569829Scharnier if (m->mtx_acqtime == 0) { 33669829Scharnier m->mtx_filename = file; 33769829Scharnier m->mtx_lineno = line; 33869829Scharnier m->mtx_acqtime = mutex_prof_enable ? nanoseconds() : 0; 33969829Scharnier ++mutex_prof_acquisitions; 34069829Scharnier } 34169829Scharnier#endif 34269829Scharnier} 34369829Scharnier 34469829Scharniervoid 34569829Scharnier_mtx_unlock_flags(struct mtx *m, int opts, const char *file, int line) 34669829Scharnier{ 34769829Scharnier 34869829Scharnier MPASS(curthread != NULL); 34969829Scharnier KASSERT(m->mtx_object.lo_class == &lock_class_mtx_sleep, 35069829Scharnier ("mtx_unlock() of spin mutex %s @ %s:%d", m->mtx_object.lo_name, 35169829Scharnier file, line)); 35269829Scharnier WITNESS_UNLOCK(&m->mtx_object, opts | LOP_EXCLUSIVE, file, line); 35369829Scharnier LOCK_LOG_LOCK("UNLOCK", &m->mtx_object, opts, m->mtx_recurse, file, 35469829Scharnier line); 35569829Scharnier mtx_assert(m, MA_OWNED); 35669829Scharnier#ifdef MUTEX_PROFILING 35769829Scharnier if (m->mtx_acqtime != 0) { 35869829Scharnier static const char *unknown = "(unknown)"; 35969829Scharnier struct mutex_prof *mpp; 36069829Scharnier u_int64_t acqtime, now; 36169829Scharnier const char *p, *q; 36269829Scharnier volatile u_int hash; 36369829Scharnier 36475377Smckusick now = nanoseconds(); 36575377Smckusick acqtime = m->mtx_acqtime; 36675377Smckusick m->mtx_acqtime = 0; 36775377Smckusick if (now <= acqtime) 36875377Smckusick goto out; 36975377Smckusick for (p = m->mtx_filename; strncmp(p, "../", 3) == 0; p += 3) 37075377Smckusick /* nothing */ ; 37175377Smckusick if (p == NULL || *p == '\0') 37275377Smckusick p = unknown; 37375377Smckusick for (hash = m->mtx_lineno, q = p; *q != '\0'; ++q) 37475377Smckusick hash = (hash * 2 + *q) % MPROF_HASH_SIZE; 37569829Scharnier mtx_lock_spin(&mprof_mtx); 376109963Sjmallett for (mpp = mprof_hash[hash]; mpp != NULL; mpp = mpp->next) 377109963Sjmallett if (mpp->line == m->mtx_lineno && 378109963Sjmallett strcmp(mpp->file, p) == 0) 37942873Sluoqi break; 38042873Sluoqi if (mpp == NULL) { 381109963Sjmallett /* Just exit if we cannot get a trace buffer */ 38242873Sluoqi if (first_free_mprof_buf >= NUM_MPROF_BUFFERS) { 38342873Sluoqi ++mutex_prof_rejected; 384102231Strhodes goto unlock; 38542873Sluoqi } 3861558Srgrimes mpp = &mprof_buf[first_free_mprof_buf++]; 387109963Sjmallett mpp->name = mtx_name(m); 388109963Sjmallett mpp->file = p; 389109963Sjmallett mpp->line = m->mtx_lineno; 390109963Sjmallett mpp->next = mprof_hash[hash]; 391109963Sjmallett if (mprof_hash[hash] != NULL) 3921558Srgrimes ++mutex_prof_collisions; 3931558Srgrimes mprof_hash[hash] = mpp; 3941558Srgrimes ++mutex_prof_records; 395109597Sjmallett } 3961558Srgrimes /* 397110174Sgordon * Record if the mutex has been held longer now than ever 398111287Sru * before. 399111287Sru */ 400110174Sgordon if (now - acqtime > mpp->counter[MPROF_MAX]) 401110174Sgordon mpp->counter[MPROF_MAX] = now - acqtime; 4021558Srgrimes mpp->counter[MPROF_TOT] += now - acqtime; 4031558Srgrimes mpp->counter[MPROF_CNT]++; 4041558Srgrimesunlock: 4051558Srgrimes mtx_unlock_spin(&mprof_mtx); 406109597Sjmallett } 4071558Srgrimesout: 408105162Srwatson#endif 409105162Srwatson _rel_sleep_lock(m, curthread, opts, file, line); 410105162Srwatson} 411105162Srwatson 412109468Sjmallettvoid 41334266Sjulian_mtx_lock_spin_flags(struct mtx *m, int opts, const char *file, int line) 4149315Sjoerg{ 4159315Sjoerg 41675377Smckusick MPASS(curthread != NULL); 41775377Smckusick KASSERT(m->mtx_object.lo_class == &lock_class_mtx_spin, 41875377Smckusick ("mtx_lock_spin() of sleep mutex %s @ %s:%d", 41975377Smckusick m->mtx_object.lo_name, file, line)); 4209315Sjoerg#if defined(SMP) || LOCK_DEBUG > 0 || 1 4219315Sjoerg _get_spin_lock(m, curthread, opts, file, line); 4229315Sjoerg#else 4239315Sjoerg critical_enter(); 4249315Sjoerg#endif 4259315Sjoerg LOCK_LOG_LOCK("LOCK", &m->mtx_object, opts, m->mtx_recurse, file, 4269315Sjoerg line); 4279315Sjoerg WITNESS_LOCK(&m->mtx_object, opts | LOP_EXCLUSIVE, file, line); 4289315Sjoerg} 4299315Sjoerg 430110174Sgordonvoid 431110174Sgordon_mtx_unlock_spin_flags(struct mtx *m, int opts, const char *file, int line) 4329315Sjoerg{ 433 434 MPASS(curthread != NULL); 435 KASSERT(m->mtx_object.lo_class == &lock_class_mtx_spin, 436 ("mtx_unlock_spin() of sleep mutex %s @ %s:%d", 437 m->mtx_object.lo_name, file, line)); 438 WITNESS_UNLOCK(&m->mtx_object, opts | LOP_EXCLUSIVE, file, line); 439 LOCK_LOG_LOCK("UNLOCK", &m->mtx_object, opts, m->mtx_recurse, file, 440 line); 441 mtx_assert(m, MA_OWNED); 442#if defined(SMP) || LOCK_DEBUG > 0 || 1 443 _rel_spin_lock(m); 444#else 445 critical_exit(); 446#endif 447} 448 449/* 450 * The important part of mtx_trylock{,_flags}() 451 * Tries to acquire lock `m.' We do NOT handle recursion here; we assume that 452 * if we're called, it's because we know we don't already own this lock. 453 */ 454int 455_mtx_trylock(struct mtx *m, int opts, const char *file, int line) 456{ 457 int rval; 458 459 MPASS(curthread != NULL); 460 461 rval = _obtain_lock(m, curthread); 462 463 LOCK_LOG_TRY("LOCK", &m->mtx_object, opts, rval, file, line); 464 if (rval) { 465 /* 466 * We do not handle recursion in _mtx_trylock; see the 467 * note at the top of the routine. 468 */ 469 KASSERT(!mtx_recursed(m), 470 ("mtx_trylock() called on a recursed mutex")); 471 WITNESS_LOCK(&m->mtx_object, opts | LOP_EXCLUSIVE | LOP_TRYLOCK, 472 file, line); 473 } 474 475 return (rval); 476} 477 478/* 479 * _mtx_lock_sleep: the tougher part of acquiring an MTX_DEF lock. 480 * 481 * We call this if the lock is either contested (i.e. we need to go to 482 * sleep waiting for it), or if we need to recurse on it. 483 */ 484void 485_mtx_lock_sleep(struct mtx *m, int opts, const char *file, int line) 486{ 487 struct thread *td = curthread; 488#if defined(SMP) && defined(ADAPTIVE_MUTEXES) 489 struct thread *owner; 490#endif 491#ifdef KTR 492 int cont_logged = 0; 493#endif 494 495 if ((m->mtx_lock & MTX_FLAGMASK) == (uintptr_t)td) { 496 m->mtx_recurse++; 497 atomic_set_ptr(&m->mtx_lock, MTX_RECURSED); 498 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 499 CTR1(KTR_LOCK, "_mtx_lock_sleep: %p recursing", m); 500 return; 501 } 502 503 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 504 CTR4(KTR_LOCK, 505 "_mtx_lock_sleep: %s contested (lock=%p) at %s:%d", 506 m->mtx_object.lo_name, (void *)m->mtx_lock, file, line); 507 508 while (!_obtain_lock(m, td)) { 509 uintptr_t v; 510 struct thread *td1; 511 512 mtx_lock_spin(&sched_lock); 513 /* 514 * Check if the lock has been released while spinning for 515 * the sched_lock. 516 */ 517 if ((v = m->mtx_lock) == MTX_UNOWNED) { 518 mtx_unlock_spin(&sched_lock); 519#ifdef __i386__ 520 ia32_pause(); 521#endif 522 continue; 523 } 524 525 /* 526 * The mutex was marked contested on release. This means that 527 * there are threads blocked on it. 528 */ 529 if (v == MTX_CONTESTED) { 530 td1 = TAILQ_FIRST(&m->mtx_blocked); 531 MPASS(td1 != NULL); 532 m->mtx_lock = (uintptr_t)td | MTX_CONTESTED; 533 534 if (td1->td_priority < td->td_priority) 535 td->td_priority = td1->td_priority; 536 mtx_unlock_spin(&sched_lock); 537 return; 538 } 539 540 /* 541 * If the mutex isn't already contested and a failure occurs 542 * setting the contested bit, the mutex was either released 543 * or the state of the MTX_RECURSED bit changed. 544 */ 545 if ((v & MTX_CONTESTED) == 0 && 546 !atomic_cmpset_ptr(&m->mtx_lock, (void *)v, 547 (void *)(v | MTX_CONTESTED))) { 548 mtx_unlock_spin(&sched_lock); 549#ifdef __i386__ 550 ia32_pause(); 551#endif 552 continue; 553 } 554 555#if defined(SMP) && defined(ADAPTIVE_MUTEXES) 556 /* 557 * If the current owner of the lock is executing on another 558 * CPU, spin instead of blocking. 559 */ 560 owner = (struct thread *)(v & MTX_FLAGMASK); 561 if (m != &Giant && thread_running(owner)) { 562 mtx_unlock_spin(&sched_lock); 563 while (mtx_owner(m) == owner && thread_running(owner)) { 564#ifdef __i386__ 565 ia32_pause(); 566#endif 567 } 568 continue; 569 } 570#endif /* SMP && ADAPTIVE_MUTEXES */ 571 572 /* 573 * We definitely must sleep for this lock. 574 */ 575 mtx_assert(m, MA_NOTOWNED); 576 577#ifdef notyet 578 /* 579 * If we're borrowing an interrupted thread's VM context, we 580 * must clean up before going to sleep. 581 */ 582 if (td->td_ithd != NULL) { 583 struct ithd *it = td->td_ithd; 584 585 if (it->it_interrupted) { 586 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 587 CTR2(KTR_LOCK, 588 "_mtx_lock_sleep: %p interrupted %p", 589 it, it->it_interrupted); 590 intr_thd_fixup(it); 591 } 592 } 593#endif 594 595 /* 596 * Put us on the list of threads blocked on this mutex. 597 */ 598 if (TAILQ_EMPTY(&m->mtx_blocked)) { 599 td1 = mtx_owner(m); 600 LIST_INSERT_HEAD(&td1->td_contested, m, mtx_contested); 601 TAILQ_INSERT_TAIL(&m->mtx_blocked, td, td_lockq); 602 } else { 603 TAILQ_FOREACH(td1, &m->mtx_blocked, td_lockq) 604 if (td1->td_priority > td->td_priority) 605 break; 606 if (td1) 607 TAILQ_INSERT_BEFORE(td1, td, td_lockq); 608 else 609 TAILQ_INSERT_TAIL(&m->mtx_blocked, td, td_lockq); 610 } 611#ifdef KTR 612 if (!cont_logged) { 613 CTR6(KTR_CONTENTION, 614 "contention: %p at %s:%d wants %s, taken by %s:%d", 615 td, file, line, m->mtx_object.lo_name, 616 WITNESS_FILE(&m->mtx_object), 617 WITNESS_LINE(&m->mtx_object)); 618 cont_logged = 1; 619 } 620#endif 621 622 /* 623 * Save who we're blocked on. 624 */ 625 td->td_blocked = m; 626 td->td_lockname = m->mtx_object.lo_name; 627 TD_SET_LOCK(td); 628 propagate_priority(td); 629 630 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 631 CTR3(KTR_LOCK, 632 "_mtx_lock_sleep: p %p blocked on [%p] %s", td, m, 633 m->mtx_object.lo_name); 634 635 td->td_proc->p_stats->p_ru.ru_nvcsw++; 636 mi_switch(); 637 638 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 639 CTR3(KTR_LOCK, 640 "_mtx_lock_sleep: p %p free from blocked on [%p] %s", 641 td, m, m->mtx_object.lo_name); 642 643 mtx_unlock_spin(&sched_lock); 644 } 645 646#ifdef KTR 647 if (cont_logged) { 648 CTR4(KTR_CONTENTION, 649 "contention end: %s acquired by %p at %s:%d", 650 m->mtx_object.lo_name, td, file, line); 651 } 652#endif 653 return; 654} 655 656/* 657 * _mtx_lock_spin: the tougher part of acquiring an MTX_SPIN lock. 658 * 659 * This is only called if we need to actually spin for the lock. Recursion 660 * is handled inline. 661 */ 662void 663_mtx_lock_spin(struct mtx *m, int opts, const char *file, int line) 664{ 665 int i = 0; 666 667 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 668 CTR1(KTR_LOCK, "_mtx_lock_spin: %p spinning", m); 669 670 for (;;) { 671 if (_obtain_lock(m, curthread)) 672 break; 673 674 /* Give interrupts a chance while we spin. */ 675 critical_exit(); 676 while (m->mtx_lock != MTX_UNOWNED) { 677 if (i++ < 10000000) { 678#ifdef __i386__ 679 ia32_pause(); 680#endif 681 continue; 682 } 683 if (i < 60000000) 684 DELAY(1); 685#ifdef DDB 686 else if (!db_active) 687#else 688 else 689#endif 690 panic("spin lock %s held by %p for > 5 seconds", 691 m->mtx_object.lo_name, (void *)m->mtx_lock); 692#ifdef __i386__ 693 ia32_pause(); 694#endif 695 } 696 critical_enter(); 697 } 698 699 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 700 CTR1(KTR_LOCK, "_mtx_lock_spin: %p spin done", m); 701 702 return; 703} 704 705/* 706 * _mtx_unlock_sleep: the tougher part of releasing an MTX_DEF lock. 707 * 708 * We are only called here if the lock is recursed or contested (i.e. we 709 * need to wake up a blocked thread). 710 */ 711void 712_mtx_unlock_sleep(struct mtx *m, int opts, const char *file, int line) 713{ 714 struct thread *td, *td1; 715 struct mtx *m1; 716 int pri; 717 718 td = curthread; 719 720 if (mtx_recursed(m)) { 721 if (--(m->mtx_recurse) == 0) 722 atomic_clear_ptr(&m->mtx_lock, MTX_RECURSED); 723 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 724 CTR1(KTR_LOCK, "_mtx_unlock_sleep: %p unrecurse", m); 725 return; 726 } 727 728 mtx_lock_spin(&sched_lock); 729 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 730 CTR1(KTR_LOCK, "_mtx_unlock_sleep: %p contested", m); 731 732 td1 = TAILQ_FIRST(&m->mtx_blocked); 733#if defined(SMP) && defined(ADAPTIVE_MUTEXES) 734 if (td1 == NULL) { 735 _release_lock_quick(m); 736 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 737 CTR1(KTR_LOCK, "_mtx_unlock_sleep: %p no sleepers", m); 738 mtx_unlock_spin(&sched_lock); 739 return; 740 } 741#endif 742 MPASS(td->td_proc->p_magic == P_MAGIC); 743 MPASS(td1->td_proc->p_magic == P_MAGIC); 744 745 TAILQ_REMOVE(&m->mtx_blocked, td1, td_lockq); 746 747 if (TAILQ_EMPTY(&m->mtx_blocked)) { 748 LIST_REMOVE(m, mtx_contested); 749 _release_lock_quick(m); 750 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 751 CTR1(KTR_LOCK, "_mtx_unlock_sleep: %p not held", m); 752 } else 753 atomic_store_rel_ptr(&m->mtx_lock, (void *)MTX_CONTESTED); 754 755 pri = PRI_MAX; 756 LIST_FOREACH(m1, &td->td_contested, mtx_contested) { 757 int cp = TAILQ_FIRST(&m1->mtx_blocked)->td_priority; 758 if (cp < pri) 759 pri = cp; 760 } 761 762 if (pri > td->td_base_pri) 763 pri = td->td_base_pri; 764 td->td_priority = pri; 765 766 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 767 CTR2(KTR_LOCK, "_mtx_unlock_sleep: %p contested setrunqueue %p", 768 m, td1); 769 770 td1->td_blocked = NULL; 771 TD_CLR_LOCK(td1); 772 if (!TD_CAN_RUN(td1)) { 773 mtx_unlock_spin(&sched_lock); 774 return; 775 } 776 setrunqueue(td1); 777 778 if (td->td_critnest == 1 && td1->td_priority < pri) { 779#ifdef notyet 780 if (td->td_ithd != NULL) { 781 struct ithd *it = td->td_ithd; 782 783 if (it->it_interrupted) { 784 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 785 CTR2(KTR_LOCK, 786 "_mtx_unlock_sleep: %p interrupted %p", 787 it, it->it_interrupted); 788 intr_thd_fixup(it); 789 } 790 } 791#endif 792 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 793 CTR2(KTR_LOCK, 794 "_mtx_unlock_sleep: %p switching out lock=%p", m, 795 (void *)m->mtx_lock); 796 797 td->td_proc->p_stats->p_ru.ru_nivcsw++; 798 mi_switch(); 799 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 800 CTR2(KTR_LOCK, "_mtx_unlock_sleep: %p resuming lock=%p", 801 m, (void *)m->mtx_lock); 802 } 803 804 mtx_unlock_spin(&sched_lock); 805 806 return; 807} 808 809/* 810 * All the unlocking of MTX_SPIN locks is done inline. 811 * See the _rel_spin_lock() macro for the details. 812 */ 813 814/* 815 * The backing function for the INVARIANTS-enabled mtx_assert() 816 */ 817#ifdef INVARIANT_SUPPORT 818void 819_mtx_assert(struct mtx *m, int what, const char *file, int line) 820{ 821 822 if (panicstr != NULL) 823 return; 824 switch (what) { 825 case MA_OWNED: 826 case MA_OWNED | MA_RECURSED: 827 case MA_OWNED | MA_NOTRECURSED: 828 if (!mtx_owned(m)) 829 panic("mutex %s not owned at %s:%d", 830 m->mtx_object.lo_name, file, line); 831 if (mtx_recursed(m)) { 832 if ((what & MA_NOTRECURSED) != 0) 833 panic("mutex %s recursed at %s:%d", 834 m->mtx_object.lo_name, file, line); 835 } else if ((what & MA_RECURSED) != 0) { 836 panic("mutex %s unrecursed at %s:%d", 837 m->mtx_object.lo_name, file, line); 838 } 839 break; 840 case MA_NOTOWNED: 841 if (mtx_owned(m)) 842 panic("mutex %s owned at %s:%d", 843 m->mtx_object.lo_name, file, line); 844 break; 845 default: 846 panic("unknown mtx_assert at %s:%d", file, line); 847 } 848} 849#endif 850 851/* 852 * The MUTEX_DEBUG-enabled mtx_validate() 853 * 854 * Most of these checks have been moved off into the LO_INITIALIZED flag 855 * maintained by the witness code. 856 */ 857#ifdef MUTEX_DEBUG 858 859void mtx_validate(struct mtx *); 860 861void 862mtx_validate(struct mtx *m) 863{ 864 865/* 866 * XXX - When kernacc() is fixed on the alpha to handle K0_SEG memory properly 867 * we can re-enable the kernacc() checks. 868 */ 869#ifndef __alpha__ 870 /* 871 * Can't call kernacc() from early init386(), especially when 872 * initializing Giant mutex, because some stuff in kernacc() 873 * requires Giant itself. 874 */ 875 if (!cold) 876 if (!kernacc((caddr_t)m, sizeof(m), 877 VM_PROT_READ | VM_PROT_WRITE)) 878 panic("Can't read and write to mutex %p", m); 879#endif 880} 881#endif 882 883/* 884 * General init routine used by the MTX_SYSINIT() macro. 885 */ 886void 887mtx_sysinit(void *arg) 888{ 889 struct mtx_args *margs = arg; 890 891 mtx_init(margs->ma_mtx, margs->ma_desc, NULL, margs->ma_opts); 892} 893 894/* 895 * Mutex initialization routine; initialize lock `m' of type contained in 896 * `opts' with options contained in `opts' and name `name.' The optional 897 * lock type `type' is used as a general lock category name for use with 898 * witness. 899 */ 900void 901mtx_init(struct mtx *m, const char *name, const char *type, int opts) 902{ 903 struct lock_object *lock; 904 905 MPASS((opts & ~(MTX_SPIN | MTX_QUIET | MTX_RECURSE | 906 MTX_SLEEPABLE | MTX_NOWITNESS | MTX_DUPOK)) == 0); 907 908#ifdef MUTEX_DEBUG 909 /* Diagnostic and error correction */ 910 mtx_validate(m); 911#endif 912 913 lock = &m->mtx_object; 914 KASSERT((lock->lo_flags & LO_INITIALIZED) == 0, 915 ("mutex %s %p already initialized", name, m)); 916 bzero(m, sizeof(*m)); 917 if (opts & MTX_SPIN) 918 lock->lo_class = &lock_class_mtx_spin; 919 else 920 lock->lo_class = &lock_class_mtx_sleep; 921 lock->lo_name = name; 922 lock->lo_type = type != NULL ? type : name; 923 if (opts & MTX_QUIET) 924 lock->lo_flags = LO_QUIET; 925 if (opts & MTX_RECURSE) 926 lock->lo_flags |= LO_RECURSABLE; 927 if (opts & MTX_SLEEPABLE) 928 lock->lo_flags |= LO_SLEEPABLE; 929 if ((opts & MTX_NOWITNESS) == 0) 930 lock->lo_flags |= LO_WITNESS; 931 if (opts & MTX_DUPOK) 932 lock->lo_flags |= LO_DUPOK; 933 934 m->mtx_lock = MTX_UNOWNED; 935 TAILQ_INIT(&m->mtx_blocked); 936 937 LOCK_LOG_INIT(lock, opts); 938 939 WITNESS_INIT(lock); 940} 941 942/* 943 * Remove lock `m' from all_mtx queue. We don't allow MTX_QUIET to be 944 * passed in as a flag here because if the corresponding mtx_init() was 945 * called with MTX_QUIET set, then it will already be set in the mutex's 946 * flags. 947 */ 948void 949mtx_destroy(struct mtx *m) 950{ 951 952 LOCK_LOG_DESTROY(&m->mtx_object, 0); 953 954 if (!mtx_owned(m)) 955 MPASS(mtx_unowned(m)); 956 else { 957 MPASS((m->mtx_lock & (MTX_RECURSED|MTX_CONTESTED)) == 0); 958 959 /* Tell witness this isn't locked to make it happy. */ 960 WITNESS_UNLOCK(&m->mtx_object, LOP_EXCLUSIVE, __FILE__, 961 __LINE__); 962 } 963 964 WITNESS_DESTROY(&m->mtx_object); 965} 966 967/* 968 * Intialize the mutex code and system mutexes. This is called from the MD 969 * startup code prior to mi_startup(). The per-CPU data space needs to be 970 * setup before this is called. 971 */ 972void 973mutex_init(void) 974{ 975 976 /* Setup thread0 so that mutexes work. */ 977 LIST_INIT(&thread0.td_contested); 978 979 /* 980 * Initialize mutexes. 981 */ 982 mtx_init(&Giant, "Giant", NULL, MTX_DEF | MTX_RECURSE); 983 mtx_init(&sched_lock, "sched lock", NULL, MTX_SPIN | MTX_RECURSE); 984 mtx_init(&proc0.p_mtx, "process lock", NULL, MTX_DEF | MTX_DUPOK); 985 mtx_lock(&Giant); 986} 987 988/* 989 * Encapsulated Giant mutex routines. These routines provide encapsulation 990 * control for the Giant mutex, allowing sysctls to be used to turn on and 991 * off Giant around certain subsystems. The default value for the sysctls 992 * are set to what developers believe is stable and working in regards to 993 * the Giant pushdown. Developers should not turn off Giant via these 994 * sysctls unless they know what they are doing. 995 * 996 * Callers of mtx_lock_giant() are expected to pass the return value to an 997 * accompanying mtx_unlock_giant() later on. If multiple subsystems are 998 * effected by a Giant wrap, all related sysctl variables must be zero for 999 * the subsystem call to operate without Giant (as determined by the caller). 1000 */ 1001 1002SYSCTL_NODE(_kern, OID_AUTO, giant, CTLFLAG_RD, NULL, "Giant mutex manipulation"); 1003 1004static int kern_giant_all = 0; 1005SYSCTL_INT(_kern_giant, OID_AUTO, all, CTLFLAG_RW, &kern_giant_all, 0, ""); 1006 1007int kern_giant_proc = 1; /* Giant around PROC locks */ 1008int kern_giant_file = 1; /* Giant around struct file & filedesc */ 1009int kern_giant_ucred = 1; /* Giant around ucred */ 1010SYSCTL_INT(_kern_giant, OID_AUTO, proc, CTLFLAG_RW, &kern_giant_proc, 0, ""); 1011SYSCTL_INT(_kern_giant, OID_AUTO, file, CTLFLAG_RW, &kern_giant_file, 0, ""); 1012SYSCTL_INT(_kern_giant, OID_AUTO, ucred, CTLFLAG_RW, &kern_giant_ucred, 0, ""); 1013 1014int 1015mtx_lock_giant(int sysctlvar) 1016{ 1017 if (sysctlvar || kern_giant_all) { 1018 mtx_lock(&Giant); 1019 return(1); 1020 } 1021 return(0); 1022} 1023 1024void 1025mtx_unlock_giant(int s) 1026{ 1027 if (s) 1028 mtx_unlock(&Giant); 1029} 1030 1031