kern_mutex.c revision 151450
1/*- 2 * Copyright (c) 1998 Berkeley Software Design, Inc. All rights reserved. 3 * 4 * Redistribution and use in source and binary forms, with or without 5 * modification, are permitted provided that the following conditions 6 * are met: 7 * 1. Redistributions of source code must retain the above copyright 8 * notice, this list of conditions and the following disclaimer. 9 * 2. Redistributions in binary form must reproduce the above copyright 10 * notice, this list of conditions and the following disclaimer in the 11 * documentation and/or other materials provided with the distribution. 12 * 3. Berkeley Software Design Inc's name may not be used to endorse or 13 * promote products derived from this software without specific prior 14 * written permission. 15 * 16 * THIS SOFTWARE IS PROVIDED BY BERKELEY SOFTWARE DESIGN INC ``AS IS'' AND 17 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 18 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 19 * ARE DISCLAIMED. IN NO EVENT SHALL BERKELEY SOFTWARE DESIGN INC BE LIABLE 20 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 21 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 22 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 23 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 24 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 25 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 26 * SUCH DAMAGE. 27 * 28 * from BSDI $Id: mutex_witness.c,v 1.1.2.20 2000/04/27 03:10:27 cp Exp $ 29 * and BSDI $Id: synch_machdep.c,v 2.3.2.39 2000/04/27 03:10:25 cp Exp $ 30 */ 31 32/* 33 * Machine independent bits of mutex implementation. 34 */ 35 36#include <sys/cdefs.h> 37__FBSDID("$FreeBSD: head/sys/kern/kern_mutex.c 151450 2005-10-18 18:27:44Z jhb $"); 38 39#include "opt_adaptive_mutexes.h" 40#include "opt_ddb.h" 41#include "opt_mprof.h" 42#include "opt_mutex_wake_all.h" 43#include "opt_sched.h" 44 45#include <sys/param.h> 46#include <sys/systm.h> 47#include <sys/bus.h> 48#include <sys/conf.h> 49#include <sys/kdb.h> 50#include <sys/kernel.h> 51#include <sys/ktr.h> 52#include <sys/lock.h> 53#include <sys/malloc.h> 54#include <sys/mutex.h> 55#include <sys/proc.h> 56#include <sys/resourcevar.h> 57#include <sys/sched.h> 58#include <sys/sbuf.h> 59#include <sys/sysctl.h> 60#include <sys/turnstile.h> 61#include <sys/vmmeter.h> 62 63#include <machine/atomic.h> 64#include <machine/bus.h> 65#include <machine/clock.h> 66#include <machine/cpu.h> 67 68#include <ddb/ddb.h> 69 70#include <fs/devfs/devfs_int.h> 71 72#include <vm/vm.h> 73#include <vm/vm_extern.h> 74 75/* 76 * Force MUTEX_WAKE_ALL for now. 77 * single thread wakeup needs fixes to avoid race conditions with 78 * priority inheritance. 79 */ 80#ifndef MUTEX_WAKE_ALL 81#define MUTEX_WAKE_ALL 82#endif 83 84/* 85 * Internal utility macros. 86 */ 87#define mtx_unowned(m) ((m)->mtx_lock == MTX_UNOWNED) 88 89#define mtx_owner(m) (mtx_unowned((m)) ? NULL \ 90 : (struct thread *)((m)->mtx_lock & MTX_FLAGMASK)) 91 92/* 93 * Lock classes for sleep and spin mutexes. 94 */ 95struct lock_class lock_class_mtx_sleep = { 96 "sleep mutex", 97 LC_SLEEPLOCK | LC_RECURSABLE 98}; 99struct lock_class lock_class_mtx_spin = { 100 "spin mutex", 101 LC_SPINLOCK | LC_RECURSABLE 102}; 103 104/* 105 * System-wide mutexes 106 */ 107struct mtx sched_lock; 108struct mtx Giant; 109 110#ifdef MUTEX_PROFILING 111SYSCTL_NODE(_debug, OID_AUTO, mutex, CTLFLAG_RD, NULL, "mutex debugging"); 112SYSCTL_NODE(_debug_mutex, OID_AUTO, prof, CTLFLAG_RD, NULL, "mutex profiling"); 113static int mutex_prof_enable = 0; 114SYSCTL_INT(_debug_mutex_prof, OID_AUTO, enable, CTLFLAG_RW, 115 &mutex_prof_enable, 0, "Enable tracing of mutex holdtime"); 116 117struct mutex_prof { 118 const char *name; 119 const char *file; 120 int line; 121 uintmax_t cnt_max; 122 uintmax_t cnt_tot; 123 uintmax_t cnt_cur; 124 uintmax_t cnt_contest_holding; 125 uintmax_t cnt_contest_locking; 126 struct mutex_prof *next; 127}; 128 129/* 130 * mprof_buf is a static pool of profiling records to avoid possible 131 * reentrance of the memory allocation functions. 132 * 133 * Note: NUM_MPROF_BUFFERS must be smaller than MPROF_HASH_SIZE. 134 */ 135#ifdef MPROF_BUFFERS 136#define NUM_MPROF_BUFFERS MPROF_BUFFERS 137#else 138#define NUM_MPROF_BUFFERS 1000 139#endif 140static struct mutex_prof mprof_buf[NUM_MPROF_BUFFERS]; 141static int first_free_mprof_buf; 142#ifndef MPROF_HASH_SIZE 143#define MPROF_HASH_SIZE 1009 144#endif 145#if NUM_MPROF_BUFFERS >= MPROF_HASH_SIZE 146#error MPROF_BUFFERS must be larger than MPROF_HASH_SIZE 147#endif 148static struct mutex_prof *mprof_hash[MPROF_HASH_SIZE]; 149/* SWAG: sbuf size = avg stat. line size * number of locks */ 150#define MPROF_SBUF_SIZE 256 * 400 151 152static int mutex_prof_acquisitions; 153SYSCTL_INT(_debug_mutex_prof, OID_AUTO, acquisitions, CTLFLAG_RD, 154 &mutex_prof_acquisitions, 0, "Number of mutex acquistions recorded"); 155static int mutex_prof_records; 156SYSCTL_INT(_debug_mutex_prof, OID_AUTO, records, CTLFLAG_RD, 157 &mutex_prof_records, 0, "Number of profiling records"); 158static int mutex_prof_maxrecords = NUM_MPROF_BUFFERS; 159SYSCTL_INT(_debug_mutex_prof, OID_AUTO, maxrecords, CTLFLAG_RD, 160 &mutex_prof_maxrecords, 0, "Maximum number of profiling records"); 161static int mutex_prof_rejected; 162SYSCTL_INT(_debug_mutex_prof, OID_AUTO, rejected, CTLFLAG_RD, 163 &mutex_prof_rejected, 0, "Number of rejected profiling records"); 164static int mutex_prof_hashsize = MPROF_HASH_SIZE; 165SYSCTL_INT(_debug_mutex_prof, OID_AUTO, hashsize, CTLFLAG_RD, 166 &mutex_prof_hashsize, 0, "Hash size"); 167static int mutex_prof_collisions = 0; 168SYSCTL_INT(_debug_mutex_prof, OID_AUTO, collisions, CTLFLAG_RD, 169 &mutex_prof_collisions, 0, "Number of hash collisions"); 170 171/* 172 * mprof_mtx protects the profiling buffers and the hash. 173 */ 174static struct mtx mprof_mtx; 175MTX_SYSINIT(mprof, &mprof_mtx, "mutex profiling lock", MTX_SPIN | MTX_QUIET); 176 177static u_int64_t 178nanoseconds(void) 179{ 180 struct timespec tv; 181 182 nanotime(&tv); 183 return (tv.tv_sec * (u_int64_t)1000000000 + tv.tv_nsec); 184} 185 186static int 187dump_mutex_prof_stats(SYSCTL_HANDLER_ARGS) 188{ 189 struct sbuf *sb; 190 int error, i; 191 static int multiplier = 1; 192 193 if (first_free_mprof_buf == 0) 194 return (SYSCTL_OUT(req, "No locking recorded", 195 sizeof("No locking recorded"))); 196 197retry_sbufops: 198 sb = sbuf_new(NULL, NULL, MPROF_SBUF_SIZE * multiplier, SBUF_FIXEDLEN); 199 sbuf_printf(sb, "\n%6s %12s %11s %5s %12s %12s %s\n", 200 "max", "total", "count", "avg", "cnt_hold", "cnt_lock", "name"); 201 /* 202 * XXX this spinlock seems to be by far the largest perpetrator 203 * of spinlock latency (1.6 msec on an Athlon1600 was recorded 204 * even before I pessimized it further by moving the average 205 * computation here). 206 */ 207 mtx_lock_spin(&mprof_mtx); 208 for (i = 0; i < first_free_mprof_buf; ++i) { 209 sbuf_printf(sb, "%6ju %12ju %11ju %5ju %12ju %12ju %s:%d (%s)\n", 210 mprof_buf[i].cnt_max / 1000, 211 mprof_buf[i].cnt_tot / 1000, 212 mprof_buf[i].cnt_cur, 213 mprof_buf[i].cnt_cur == 0 ? (uintmax_t)0 : 214 mprof_buf[i].cnt_tot / (mprof_buf[i].cnt_cur * 1000), 215 mprof_buf[i].cnt_contest_holding, 216 mprof_buf[i].cnt_contest_locking, 217 mprof_buf[i].file, mprof_buf[i].line, mprof_buf[i].name); 218 if (sbuf_overflowed(sb)) { 219 mtx_unlock_spin(&mprof_mtx); 220 sbuf_delete(sb); 221 multiplier++; 222 goto retry_sbufops; 223 } 224 } 225 mtx_unlock_spin(&mprof_mtx); 226 sbuf_finish(sb); 227 error = SYSCTL_OUT(req, sbuf_data(sb), sbuf_len(sb) + 1); 228 sbuf_delete(sb); 229 return (error); 230} 231SYSCTL_PROC(_debug_mutex_prof, OID_AUTO, stats, CTLTYPE_STRING | CTLFLAG_RD, 232 NULL, 0, dump_mutex_prof_stats, "A", "Mutex profiling statistics"); 233 234static int 235reset_mutex_prof_stats(SYSCTL_HANDLER_ARGS) 236{ 237 int error, v; 238 239 if (first_free_mprof_buf == 0) 240 return (0); 241 242 v = 0; 243 error = sysctl_handle_int(oidp, &v, 0, req); 244 if (error) 245 return (error); 246 if (req->newptr == NULL) 247 return (error); 248 if (v == 0) 249 return (0); 250 251 mtx_lock_spin(&mprof_mtx); 252 bzero(mprof_buf, sizeof(*mprof_buf) * first_free_mprof_buf); 253 bzero(mprof_hash, sizeof(struct mtx *) * MPROF_HASH_SIZE); 254 first_free_mprof_buf = 0; 255 mtx_unlock_spin(&mprof_mtx); 256 return (0); 257} 258SYSCTL_PROC(_debug_mutex_prof, OID_AUTO, reset, CTLTYPE_INT | CTLFLAG_RW, 259 NULL, 0, reset_mutex_prof_stats, "I", "Reset mutex profiling statistics"); 260#endif 261 262/* 263 * Function versions of the inlined __mtx_* macros. These are used by 264 * modules and can also be called from assembly language if needed. 265 */ 266void 267_mtx_lock_flags(struct mtx *m, int opts, const char *file, int line) 268{ 269 270 MPASS(curthread != NULL); 271 KASSERT(m->mtx_object.lo_class == &lock_class_mtx_sleep, 272 ("mtx_lock() of spin mutex %s @ %s:%d", m->mtx_object.lo_name, 273 file, line)); 274 WITNESS_CHECKORDER(&m->mtx_object, opts | LOP_NEWORDER | LOP_EXCLUSIVE, 275 file, line); 276 _get_sleep_lock(m, curthread, opts, file, line); 277 LOCK_LOG_LOCK("LOCK", &m->mtx_object, opts, m->mtx_recurse, file, 278 line); 279 WITNESS_LOCK(&m->mtx_object, opts | LOP_EXCLUSIVE, file, line); 280#ifdef MUTEX_PROFILING 281 /* don't reset the timer when/if recursing */ 282 if (m->mtx_acqtime == 0) { 283 m->mtx_filename = file; 284 m->mtx_lineno = line; 285 m->mtx_acqtime = mutex_prof_enable ? nanoseconds() : 0; 286 ++mutex_prof_acquisitions; 287 } 288#endif 289} 290 291void 292_mtx_unlock_flags(struct mtx *m, int opts, const char *file, int line) 293{ 294 295 MPASS(curthread != NULL); 296 KASSERT(m->mtx_object.lo_class == &lock_class_mtx_sleep, 297 ("mtx_unlock() of spin mutex %s @ %s:%d", m->mtx_object.lo_name, 298 file, line)); 299 WITNESS_UNLOCK(&m->mtx_object, opts | LOP_EXCLUSIVE, file, line); 300 LOCK_LOG_LOCK("UNLOCK", &m->mtx_object, opts, m->mtx_recurse, file, 301 line); 302 mtx_assert(m, MA_OWNED); 303#ifdef MUTEX_PROFILING 304 if (m->mtx_acqtime != 0) { 305 static const char *unknown = "(unknown)"; 306 struct mutex_prof *mpp; 307 u_int64_t acqtime, now; 308 const char *p, *q; 309 volatile u_int hash; 310 311 now = nanoseconds(); 312 acqtime = m->mtx_acqtime; 313 m->mtx_acqtime = 0; 314 if (now <= acqtime) 315 goto out; 316 for (p = m->mtx_filename; 317 p != NULL && strncmp(p, "../", 3) == 0; p += 3) 318 /* nothing */ ; 319 if (p == NULL || *p == '\0') 320 p = unknown; 321 for (hash = m->mtx_lineno, q = p; *q != '\0'; ++q) 322 hash = (hash * 2 + *q) % MPROF_HASH_SIZE; 323 mtx_lock_spin(&mprof_mtx); 324 for (mpp = mprof_hash[hash]; mpp != NULL; mpp = mpp->next) 325 if (mpp->line == m->mtx_lineno && 326 strcmp(mpp->file, p) == 0) 327 break; 328 if (mpp == NULL) { 329 /* Just exit if we cannot get a trace buffer */ 330 if (first_free_mprof_buf >= NUM_MPROF_BUFFERS) { 331 ++mutex_prof_rejected; 332 goto unlock; 333 } 334 mpp = &mprof_buf[first_free_mprof_buf++]; 335 mpp->name = mtx_name(m); 336 mpp->file = p; 337 mpp->line = m->mtx_lineno; 338 mpp->next = mprof_hash[hash]; 339 if (mprof_hash[hash] != NULL) 340 ++mutex_prof_collisions; 341 mprof_hash[hash] = mpp; 342 ++mutex_prof_records; 343 } 344 /* 345 * Record if the mutex has been held longer now than ever 346 * before. 347 */ 348 if (now - acqtime > mpp->cnt_max) 349 mpp->cnt_max = now - acqtime; 350 mpp->cnt_tot += now - acqtime; 351 mpp->cnt_cur++; 352 /* 353 * There's a small race, really we should cmpxchg 354 * 0 with the current value, but that would bill 355 * the contention to the wrong lock instance if 356 * it followed this also. 357 */ 358 mpp->cnt_contest_holding += m->mtx_contest_holding; 359 m->mtx_contest_holding = 0; 360 mpp->cnt_contest_locking += m->mtx_contest_locking; 361 m->mtx_contest_locking = 0; 362unlock: 363 mtx_unlock_spin(&mprof_mtx); 364 } 365out: 366#endif 367 _rel_sleep_lock(m, curthread, opts, file, line); 368} 369 370void 371_mtx_lock_spin_flags(struct mtx *m, int opts, const char *file, int line) 372{ 373 374 MPASS(curthread != NULL); 375 KASSERT(m->mtx_object.lo_class == &lock_class_mtx_spin, 376 ("mtx_lock_spin() of sleep mutex %s @ %s:%d", 377 m->mtx_object.lo_name, file, line)); 378 WITNESS_CHECKORDER(&m->mtx_object, opts | LOP_NEWORDER | LOP_EXCLUSIVE, 379 file, line); 380 _get_spin_lock(m, curthread, opts, file, line); 381 LOCK_LOG_LOCK("LOCK", &m->mtx_object, opts, m->mtx_recurse, file, 382 line); 383 WITNESS_LOCK(&m->mtx_object, opts | LOP_EXCLUSIVE, file, line); 384} 385 386void 387_mtx_unlock_spin_flags(struct mtx *m, int opts, const char *file, int line) 388{ 389 390 MPASS(curthread != NULL); 391 KASSERT(m->mtx_object.lo_class == &lock_class_mtx_spin, 392 ("mtx_unlock_spin() of sleep mutex %s @ %s:%d", 393 m->mtx_object.lo_name, file, line)); 394 WITNESS_UNLOCK(&m->mtx_object, opts | LOP_EXCLUSIVE, file, line); 395 LOCK_LOG_LOCK("UNLOCK", &m->mtx_object, opts, m->mtx_recurse, file, 396 line); 397 mtx_assert(m, MA_OWNED); 398 _rel_spin_lock(m); 399} 400 401/* 402 * The important part of mtx_trylock{,_flags}() 403 * Tries to acquire lock `m.' If this function is called on a mutex that 404 * is already owned, it will recursively acquire the lock. 405 */ 406int 407_mtx_trylock(struct mtx *m, int opts, const char *file, int line) 408{ 409 int rval; 410 411 MPASS(curthread != NULL); 412 KASSERT(m->mtx_object.lo_class == &lock_class_mtx_sleep, 413 ("mtx_trylock() of spin mutex %s @ %s:%d", m->mtx_object.lo_name, 414 file, line)); 415 416 if (mtx_owned(m) && (m->mtx_object.lo_flags & LO_RECURSABLE) != 0) { 417 m->mtx_recurse++; 418 atomic_set_ptr(&m->mtx_lock, MTX_RECURSED); 419 rval = 1; 420 } else 421 rval = _obtain_lock(m, (uintptr_t)curthread); 422 423 LOCK_LOG_TRY("LOCK", &m->mtx_object, opts, rval, file, line); 424 if (rval) 425 WITNESS_LOCK(&m->mtx_object, opts | LOP_EXCLUSIVE | LOP_TRYLOCK, 426 file, line); 427 428 return (rval); 429} 430 431/* 432 * _mtx_lock_sleep: the tougher part of acquiring an MTX_DEF lock. 433 * 434 * We call this if the lock is either contested (i.e. we need to go to 435 * sleep waiting for it), or if we need to recurse on it. 436 */ 437void 438_mtx_lock_sleep(struct mtx *m, uintptr_t tid, int opts, const char *file, 439 int line) 440{ 441#if defined(SMP) && !defined(NO_ADAPTIVE_MUTEXES) 442 struct thread *owner; 443#endif 444 uintptr_t v; 445#ifdef KTR 446 int cont_logged = 0; 447#endif 448#ifdef MUTEX_PROFILING 449 int contested; 450#endif 451 452 if (mtx_owned(m)) { 453 KASSERT((m->mtx_object.lo_flags & LO_RECURSABLE) != 0, 454 ("_mtx_lock_sleep: recursed on non-recursive mutex %s @ %s:%d\n", 455 m->mtx_object.lo_name, file, line)); 456 m->mtx_recurse++; 457 atomic_set_ptr(&m->mtx_lock, MTX_RECURSED); 458 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 459 CTR1(KTR_LOCK, "_mtx_lock_sleep: %p recursing", m); 460 return; 461 } 462 463 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 464 CTR4(KTR_LOCK, 465 "_mtx_lock_sleep: %s contested (lock=%p) at %s:%d", 466 m->mtx_object.lo_name, (void *)m->mtx_lock, file, line); 467 468#ifdef MUTEX_PROFILING 469 contested = 0; 470#endif 471 while (!_obtain_lock(m, tid)) { 472#ifdef MUTEX_PROFILING 473 contested = 1; 474 atomic_add_int(&m->mtx_contest_holding, 1); 475#endif 476 turnstile_lock(&m->mtx_object); 477 v = m->mtx_lock; 478 479 /* 480 * Check if the lock has been released while spinning for 481 * the turnstile chain lock. 482 */ 483 if (v == MTX_UNOWNED) { 484 turnstile_release(&m->mtx_object); 485 cpu_spinwait(); 486 continue; 487 } 488 489#ifdef MUTEX_WAKE_ALL 490 MPASS(v != MTX_CONTESTED); 491#else 492 /* 493 * The mutex was marked contested on release. This means that 494 * there are other threads blocked on it. Grab ownership of 495 * it and propagate its priority to the current thread if 496 * necessary. 497 */ 498 if (v == MTX_CONTESTED) { 499 m->mtx_lock = tid | MTX_CONTESTED; 500 turnstile_claim(&m->mtx_object); 501 break; 502 } 503#endif 504 505 /* 506 * If the mutex isn't already contested and a failure occurs 507 * setting the contested bit, the mutex was either released 508 * or the state of the MTX_RECURSED bit changed. 509 */ 510 if ((v & MTX_CONTESTED) == 0 && 511 !atomic_cmpset_ptr(&m->mtx_lock, v, v | MTX_CONTESTED)) { 512 turnstile_release(&m->mtx_object); 513 cpu_spinwait(); 514 continue; 515 } 516 517#if defined(SMP) && !defined(NO_ADAPTIVE_MUTEXES) 518 /* 519 * If the current owner of the lock is executing on another 520 * CPU, spin instead of blocking. 521 */ 522 owner = (struct thread *)(v & MTX_FLAGMASK); 523#ifdef ADAPTIVE_GIANT 524 if (TD_IS_RUNNING(owner)) { 525#else 526 if (m != &Giant && TD_IS_RUNNING(owner)) { 527#endif 528 turnstile_release(&m->mtx_object); 529 while (mtx_owner(m) == owner && TD_IS_RUNNING(owner)) { 530 cpu_spinwait(); 531 } 532 continue; 533 } 534#endif /* SMP && !NO_ADAPTIVE_MUTEXES */ 535 536 /* 537 * We definitely must sleep for this lock. 538 */ 539 mtx_assert(m, MA_NOTOWNED); 540 541#ifdef KTR 542 if (!cont_logged) { 543 CTR6(KTR_CONTENTION, 544 "contention: %p at %s:%d wants %s, taken by %s:%d", 545 (void *)tid, file, line, m->mtx_object.lo_name, 546 WITNESS_FILE(&m->mtx_object), 547 WITNESS_LINE(&m->mtx_object)); 548 cont_logged = 1; 549 } 550#endif 551 552 /* 553 * Block on the turnstile. 554 */ 555 turnstile_wait(&m->mtx_object, mtx_owner(m)); 556 } 557 558#ifdef KTR 559 if (cont_logged) { 560 CTR4(KTR_CONTENTION, 561 "contention end: %s acquired by %p at %s:%d", 562 m->mtx_object.lo_name, (void *)tid, file, line); 563 } 564#endif 565#ifdef MUTEX_PROFILING 566 if (contested) 567 m->mtx_contest_locking++; 568 m->mtx_contest_holding = 0; 569#endif 570 return; 571} 572 573#ifdef SMP 574/* 575 * _mtx_lock_spin: the tougher part of acquiring an MTX_SPIN lock. 576 * 577 * This is only called if we need to actually spin for the lock. Recursion 578 * is handled inline. 579 */ 580void 581_mtx_lock_spin(struct mtx *m, uintptr_t tid, int opts, const char *file, 582 int line) 583{ 584 int i = 0; 585 586 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 587 CTR1(KTR_LOCK, "_mtx_lock_spin: %p spinning", m); 588 589 for (;;) { 590 if (_obtain_lock(m, tid)) 591 break; 592 593 /* Give interrupts a chance while we spin. */ 594 spinlock_exit(); 595 while (m->mtx_lock != MTX_UNOWNED) { 596 if (i++ < 10000000) { 597 cpu_spinwait(); 598 continue; 599 } 600 if (i < 60000000) 601 DELAY(1); 602 else if (!kdb_active && !panicstr) { 603 printf("spin lock %s held by %p for > 5 seconds\n", 604 m->mtx_object.lo_name, (void *)m->mtx_lock); 605#ifdef WITNESS 606 witness_display_spinlock(&m->mtx_object, 607 mtx_owner(m)); 608#endif 609 panic("spin lock held too long"); 610 } 611 cpu_spinwait(); 612 } 613 spinlock_enter(); 614 } 615 616 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 617 CTR1(KTR_LOCK, "_mtx_lock_spin: %p spin done", m); 618 619 return; 620} 621#endif /* SMP */ 622 623/* 624 * _mtx_unlock_sleep: the tougher part of releasing an MTX_DEF lock. 625 * 626 * We are only called here if the lock is recursed or contested (i.e. we 627 * need to wake up a blocked thread). 628 */ 629void 630_mtx_unlock_sleep(struct mtx *m, int opts, const char *file, int line) 631{ 632 struct turnstile *ts; 633#ifndef PREEMPTION 634 struct thread *td, *td1; 635#endif 636 637 if (mtx_recursed(m)) { 638 if (--(m->mtx_recurse) == 0) 639 atomic_clear_ptr(&m->mtx_lock, MTX_RECURSED); 640 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 641 CTR1(KTR_LOCK, "_mtx_unlock_sleep: %p unrecurse", m); 642 return; 643 } 644 645 turnstile_lock(&m->mtx_object); 646 ts = turnstile_lookup(&m->mtx_object); 647 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 648 CTR1(KTR_LOCK, "_mtx_unlock_sleep: %p contested", m); 649 650#if defined(SMP) && !defined(NO_ADAPTIVE_MUTEXES) 651 if (ts == NULL) { 652 _release_lock_quick(m); 653 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 654 CTR1(KTR_LOCK, "_mtx_unlock_sleep: %p no sleepers", m); 655 turnstile_release(&m->mtx_object); 656 return; 657 } 658#else 659 MPASS(ts != NULL); 660#endif 661#ifndef PREEMPTION 662 /* XXX */ 663 td1 = turnstile_head(ts); 664#endif 665#ifdef MUTEX_WAKE_ALL 666 turnstile_broadcast(ts); 667 _release_lock_quick(m); 668#else 669 if (turnstile_signal(ts)) { 670 _release_lock_quick(m); 671 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 672 CTR1(KTR_LOCK, "_mtx_unlock_sleep: %p not held", m); 673 } else { 674 m->mtx_lock = MTX_CONTESTED; 675 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 676 CTR1(KTR_LOCK, "_mtx_unlock_sleep: %p still contested", 677 m); 678 } 679#endif 680 turnstile_unpend(ts); 681 682#ifndef PREEMPTION 683 /* 684 * XXX: This is just a hack until preemption is done. However, 685 * once preemption is done we need to either wrap the 686 * turnstile_signal() and release of the actual lock in an 687 * extra critical section or change the preemption code to 688 * always just set a flag and never do instant-preempts. 689 */ 690 td = curthread; 691 if (td->td_critnest > 0 || td1->td_priority >= td->td_priority) 692 return; 693 mtx_lock_spin(&sched_lock); 694 if (!TD_IS_RUNNING(td1)) { 695#ifdef notyet 696 if (td->td_ithd != NULL) { 697 struct ithd *it = td->td_ithd; 698 699 if (it->it_interrupted) { 700 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 701 CTR2(KTR_LOCK, 702 "_mtx_unlock_sleep: %p interrupted %p", 703 it, it->it_interrupted); 704 intr_thd_fixup(it); 705 } 706 } 707#endif 708 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 709 CTR2(KTR_LOCK, 710 "_mtx_unlock_sleep: %p switching out lock=%p", m, 711 (void *)m->mtx_lock); 712 713 mi_switch(SW_INVOL, NULL); 714 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 715 CTR2(KTR_LOCK, "_mtx_unlock_sleep: %p resuming lock=%p", 716 m, (void *)m->mtx_lock); 717 } 718 mtx_unlock_spin(&sched_lock); 719#endif 720 721 return; 722} 723 724/* 725 * All the unlocking of MTX_SPIN locks is done inline. 726 * See the _rel_spin_lock() macro for the details. 727 */ 728 729/* 730 * The backing function for the INVARIANTS-enabled mtx_assert() 731 */ 732#ifdef INVARIANT_SUPPORT 733void 734_mtx_assert(struct mtx *m, int what, const char *file, int line) 735{ 736 737 if (panicstr != NULL || dumping) 738 return; 739 switch (what) { 740 case MA_OWNED: 741 case MA_OWNED | MA_RECURSED: 742 case MA_OWNED | MA_NOTRECURSED: 743 if (!mtx_owned(m)) 744 panic("mutex %s not owned at %s:%d", 745 m->mtx_object.lo_name, file, line); 746 if (mtx_recursed(m)) { 747 if ((what & MA_NOTRECURSED) != 0) 748 panic("mutex %s recursed at %s:%d", 749 m->mtx_object.lo_name, file, line); 750 } else if ((what & MA_RECURSED) != 0) { 751 panic("mutex %s unrecursed at %s:%d", 752 m->mtx_object.lo_name, file, line); 753 } 754 break; 755 case MA_NOTOWNED: 756 if (mtx_owned(m)) 757 panic("mutex %s owned at %s:%d", 758 m->mtx_object.lo_name, file, line); 759 break; 760 default: 761 panic("unknown mtx_assert at %s:%d", file, line); 762 } 763} 764#endif 765 766/* 767 * The MUTEX_DEBUG-enabled mtx_validate() 768 * 769 * Most of these checks have been moved off into the LO_INITIALIZED flag 770 * maintained by the witness code. 771 */ 772#ifdef MUTEX_DEBUG 773 774void mtx_validate(struct mtx *); 775 776void 777mtx_validate(struct mtx *m) 778{ 779 780/* 781 * XXX: When kernacc() does not require Giant we can reenable this check 782 */ 783#ifdef notyet 784/* 785 * XXX - When kernacc() is fixed on the alpha to handle K0_SEG memory properly 786 * we can re-enable the kernacc() checks. 787 */ 788#ifndef __alpha__ 789 /* 790 * Can't call kernacc() from early init386(), especially when 791 * initializing Giant mutex, because some stuff in kernacc() 792 * requires Giant itself. 793 */ 794 if (!cold) 795 if (!kernacc((caddr_t)m, sizeof(m), 796 VM_PROT_READ | VM_PROT_WRITE)) 797 panic("Can't read and write to mutex %p", m); 798#endif 799#endif 800} 801#endif 802 803/* 804 * General init routine used by the MTX_SYSINIT() macro. 805 */ 806void 807mtx_sysinit(void *arg) 808{ 809 struct mtx_args *margs = arg; 810 811 mtx_init(margs->ma_mtx, margs->ma_desc, NULL, margs->ma_opts); 812} 813 814/* 815 * Mutex initialization routine; initialize lock `m' of type contained in 816 * `opts' with options contained in `opts' and name `name.' The optional 817 * lock type `type' is used as a general lock category name for use with 818 * witness. 819 */ 820void 821mtx_init(struct mtx *m, const char *name, const char *type, int opts) 822{ 823 struct lock_object *lock; 824 825 MPASS((opts & ~(MTX_SPIN | MTX_QUIET | MTX_RECURSE | 826 MTX_NOWITNESS | MTX_DUPOK)) == 0); 827 828#ifdef MUTEX_DEBUG 829 /* Diagnostic and error correction */ 830 mtx_validate(m); 831#endif 832 833 lock = &m->mtx_object; 834 KASSERT((lock->lo_flags & LO_INITIALIZED) == 0, 835 ("mutex \"%s\" %p already initialized", name, m)); 836 bzero(m, sizeof(*m)); 837 if (opts & MTX_SPIN) 838 lock->lo_class = &lock_class_mtx_spin; 839 else 840 lock->lo_class = &lock_class_mtx_sleep; 841 lock->lo_name = name; 842 lock->lo_type = type != NULL ? type : name; 843 if (opts & MTX_QUIET) 844 lock->lo_flags = LO_QUIET; 845 if (opts & MTX_RECURSE) 846 lock->lo_flags |= LO_RECURSABLE; 847 if ((opts & MTX_NOWITNESS) == 0) 848 lock->lo_flags |= LO_WITNESS; 849 if (opts & MTX_DUPOK) 850 lock->lo_flags |= LO_DUPOK; 851 852 m->mtx_lock = MTX_UNOWNED; 853 854 LOCK_LOG_INIT(lock, opts); 855 856 WITNESS_INIT(lock); 857} 858 859/* 860 * Remove lock `m' from all_mtx queue. We don't allow MTX_QUIET to be 861 * passed in as a flag here because if the corresponding mtx_init() was 862 * called with MTX_QUIET set, then it will already be set in the mutex's 863 * flags. 864 */ 865void 866mtx_destroy(struct mtx *m) 867{ 868 869 LOCK_LOG_DESTROY(&m->mtx_object, 0); 870 871 if (!mtx_owned(m)) 872 MPASS(mtx_unowned(m)); 873 else { 874 MPASS((m->mtx_lock & (MTX_RECURSED|MTX_CONTESTED)) == 0); 875 876 /* Tell witness this isn't locked to make it happy. */ 877 WITNESS_UNLOCK(&m->mtx_object, LOP_EXCLUSIVE, __FILE__, 878 __LINE__); 879 } 880 881 WITNESS_DESTROY(&m->mtx_object); 882} 883 884/* 885 * Intialize the mutex code and system mutexes. This is called from the MD 886 * startup code prior to mi_startup(). The per-CPU data space needs to be 887 * setup before this is called. 888 */ 889void 890mutex_init(void) 891{ 892 893 /* Setup thread0 so that mutexes work. */ 894 LIST_INIT(&thread0.td_contested); 895 896 /* Setup turnstiles so that sleep mutexes work. */ 897 init_turnstiles(); 898 899 /* 900 * Initialize mutexes. 901 */ 902 mtx_init(&Giant, "Giant", NULL, MTX_DEF | MTX_RECURSE); 903 mtx_init(&sched_lock, "sched lock", NULL, MTX_SPIN | MTX_RECURSE); 904 mtx_init(&proc0.p_mtx, "process lock", NULL, MTX_DEF | MTX_DUPOK); 905 mtx_init(&devmtx, "cdev", NULL, MTX_DEF); 906 mtx_lock(&Giant); 907} 908