kern_mutex.c revision 132698
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 132698 2004-07-27 16:34:48Z rwatson $"); 38 39#include "opt_adaptive_mutexes.h" 40#include "opt_ddb.h" 41#include "opt_mutex_wake_all.h" 42 43#include <sys/param.h> 44#include <sys/systm.h> 45#include <sys/bus.h> 46#include <sys/kdb.h> 47#include <sys/kernel.h> 48#include <sys/ktr.h> 49#include <sys/lock.h> 50#include <sys/malloc.h> 51#include <sys/mutex.h> 52#include <sys/proc.h> 53#include <sys/resourcevar.h> 54#include <sys/sched.h> 55#include <sys/sbuf.h> 56#include <sys/sysctl.h> 57#include <sys/turnstile.h> 58#include <sys/vmmeter.h> 59 60#include <machine/atomic.h> 61#include <machine/bus.h> 62#include <machine/clock.h> 63#include <machine/cpu.h> 64 65#include <ddb/ddb.h> 66 67#include <vm/vm.h> 68#include <vm/vm_extern.h> 69 70/* 71 * Internal utility macros. 72 */ 73#define mtx_unowned(m) ((m)->mtx_lock == MTX_UNOWNED) 74 75#define mtx_owner(m) (mtx_unowned((m)) ? NULL \ 76 : (struct thread *)((m)->mtx_lock & MTX_FLAGMASK)) 77 78/* 79 * Lock classes for sleep and spin mutexes. 80 */ 81struct lock_class lock_class_mtx_sleep = { 82 "sleep mutex", 83 LC_SLEEPLOCK | LC_RECURSABLE 84}; 85struct lock_class lock_class_mtx_spin = { 86 "spin mutex", 87 LC_SPINLOCK | LC_RECURSABLE 88}; 89 90/* 91 * System-wide mutexes 92 */ 93struct mtx sched_lock; 94struct mtx Giant; 95 96#ifdef MUTEX_PROFILING 97SYSCTL_NODE(_debug, OID_AUTO, mutex, CTLFLAG_RD, NULL, "mutex debugging"); 98SYSCTL_NODE(_debug_mutex, OID_AUTO, prof, CTLFLAG_RD, NULL, "mutex profiling"); 99static int mutex_prof_enable = 0; 100SYSCTL_INT(_debug_mutex_prof, OID_AUTO, enable, CTLFLAG_RW, 101 &mutex_prof_enable, 0, "Enable tracing of mutex holdtime"); 102 103struct mutex_prof { 104 const char *name; 105 const char *file; 106 int line; 107 uintmax_t cnt_max; 108 uintmax_t cnt_tot; 109 uintmax_t cnt_cur; 110 uintmax_t cnt_contest_holding; 111 uintmax_t cnt_contest_locking; 112 struct mutex_prof *next; 113}; 114 115/* 116 * mprof_buf is a static pool of profiling records to avoid possible 117 * reentrance of the memory allocation functions. 118 * 119 * Note: NUM_MPROF_BUFFERS must be smaller than MPROF_HASH_SIZE. 120 */ 121#define NUM_MPROF_BUFFERS 1000 122static struct mutex_prof mprof_buf[NUM_MPROF_BUFFERS]; 123static int first_free_mprof_buf; 124#define MPROF_HASH_SIZE 1009 125static struct mutex_prof *mprof_hash[MPROF_HASH_SIZE]; 126/* SWAG: sbuf size = avg stat. line size * number of locks */ 127#define MPROF_SBUF_SIZE 256 * 400 128 129static int mutex_prof_acquisitions; 130SYSCTL_INT(_debug_mutex_prof, OID_AUTO, acquisitions, CTLFLAG_RD, 131 &mutex_prof_acquisitions, 0, "Number of mutex acquistions recorded"); 132static int mutex_prof_records; 133SYSCTL_INT(_debug_mutex_prof, OID_AUTO, records, CTLFLAG_RD, 134 &mutex_prof_records, 0, "Number of profiling records"); 135static int mutex_prof_maxrecords = NUM_MPROF_BUFFERS; 136SYSCTL_INT(_debug_mutex_prof, OID_AUTO, maxrecords, CTLFLAG_RD, 137 &mutex_prof_maxrecords, 0, "Maximum number of profiling records"); 138static int mutex_prof_rejected; 139SYSCTL_INT(_debug_mutex_prof, OID_AUTO, rejected, CTLFLAG_RD, 140 &mutex_prof_rejected, 0, "Number of rejected profiling records"); 141static int mutex_prof_hashsize = MPROF_HASH_SIZE; 142SYSCTL_INT(_debug_mutex_prof, OID_AUTO, hashsize, CTLFLAG_RD, 143 &mutex_prof_hashsize, 0, "Hash size"); 144static int mutex_prof_collisions = 0; 145SYSCTL_INT(_debug_mutex_prof, OID_AUTO, collisions, CTLFLAG_RD, 146 &mutex_prof_collisions, 0, "Number of hash collisions"); 147 148/* 149 * mprof_mtx protects the profiling buffers and the hash. 150 */ 151static struct mtx mprof_mtx; 152MTX_SYSINIT(mprof, &mprof_mtx, "mutex profiling lock", MTX_SPIN | MTX_QUIET); 153 154static u_int64_t 155nanoseconds(void) 156{ 157 struct timespec tv; 158 159 nanotime(&tv); 160 return (tv.tv_sec * (u_int64_t)1000000000 + tv.tv_nsec); 161} 162 163static int 164dump_mutex_prof_stats(SYSCTL_HANDLER_ARGS) 165{ 166 struct sbuf *sb; 167 int error, i; 168 static int multiplier = 1; 169 170 if (first_free_mprof_buf == 0) 171 return (SYSCTL_OUT(req, "No locking recorded", 172 sizeof("No locking recorded"))); 173 174retry_sbufops: 175 sb = sbuf_new(NULL, NULL, MPROF_SBUF_SIZE * multiplier, SBUF_FIXEDLEN); 176 sbuf_printf(sb, "%6s %12s %11s %5s %12s %12s %s\n", 177 "max", "total", "count", "avg", "cnt_hold", "cnt_lock", "name"); 178 /* 179 * XXX this spinlock seems to be by far the largest perpetrator 180 * of spinlock latency (1.6 msec on an Athlon1600 was recorded 181 * even before I pessimized it further by moving the average 182 * computation here). 183 */ 184 mtx_lock_spin(&mprof_mtx); 185 for (i = 0; i < first_free_mprof_buf; ++i) { 186 sbuf_printf(sb, "%6ju %12ju %11ju %5ju %12ju %12ju %s:%d (%s)\n", 187 mprof_buf[i].cnt_max / 1000, 188 mprof_buf[i].cnt_tot / 1000, 189 mprof_buf[i].cnt_cur, 190 mprof_buf[i].cnt_cur == 0 ? (uintmax_t)0 : 191 mprof_buf[i].cnt_tot / (mprof_buf[i].cnt_cur * 1000), 192 mprof_buf[i].cnt_contest_holding, 193 mprof_buf[i].cnt_contest_locking, 194 mprof_buf[i].file, mprof_buf[i].line, mprof_buf[i].name); 195 if (sbuf_overflowed(sb)) { 196 mtx_unlock_spin(&mprof_mtx); 197 sbuf_delete(sb); 198 multiplier++; 199 goto retry_sbufops; 200 } 201 } 202 mtx_unlock_spin(&mprof_mtx); 203 sbuf_finish(sb); 204 error = SYSCTL_OUT(req, sbuf_data(sb), sbuf_len(sb) + 1); 205 sbuf_delete(sb); 206 return (error); 207} 208SYSCTL_PROC(_debug_mutex_prof, OID_AUTO, stats, CTLTYPE_STRING | CTLFLAG_RD, 209 NULL, 0, dump_mutex_prof_stats, "A", "Mutex profiling statistics"); 210 211static int 212reset_mutex_prof_stats(SYSCTL_HANDLER_ARGS) 213{ 214 int error, v; 215 216 if (first_free_mprof_buf == 0) 217 return (0); 218 219 v = 0; 220 error = sysctl_handle_int(oidp, &v, 0, req); 221 if (error) 222 return (error); 223 if (req->newptr == NULL) 224 return (error); 225 if (v == 0) 226 return (0); 227 228 mtx_lock_spin(&mprof_mtx); 229 bzero(mprof_buf, sizeof(*mprof_buf) * first_free_mprof_buf); 230 bzero(mprof_hash, sizeof(struct mtx *) * MPROF_HASH_SIZE); 231 first_free_mprof_buf = 0; 232 mtx_unlock_spin(&mprof_mtx); 233 return (0); 234} 235SYSCTL_PROC(_debug_mutex_prof, OID_AUTO, reset, CTLTYPE_INT | CTLFLAG_RW, 236 NULL, 0, reset_mutex_prof_stats, "I", "Reset mutex profiling statistics"); 237#endif 238 239/* 240 * Function versions of the inlined __mtx_* macros. These are used by 241 * modules and can also be called from assembly language if needed. 242 */ 243void 244_mtx_lock_flags(struct mtx *m, int opts, const char *file, int line) 245{ 246 247 MPASS(curthread != NULL); 248 KASSERT(m->mtx_object.lo_class == &lock_class_mtx_sleep, 249 ("mtx_lock() of spin mutex %s @ %s:%d", m->mtx_object.lo_name, 250 file, line)); 251 WITNESS_CHECKORDER(&m->mtx_object, opts | LOP_NEWORDER | LOP_EXCLUSIVE, 252 file, line); 253 _get_sleep_lock(m, curthread, opts, file, line); 254 LOCK_LOG_LOCK("LOCK", &m->mtx_object, opts, m->mtx_recurse, file, 255 line); 256 WITNESS_LOCK(&m->mtx_object, opts | LOP_EXCLUSIVE, file, line); 257#ifdef MUTEX_PROFILING 258 /* don't reset the timer when/if recursing */ 259 if (m->mtx_acqtime == 0) { 260 m->mtx_filename = file; 261 m->mtx_lineno = line; 262 m->mtx_acqtime = mutex_prof_enable ? nanoseconds() : 0; 263 ++mutex_prof_acquisitions; 264 } 265#endif 266} 267 268void 269_mtx_unlock_flags(struct mtx *m, int opts, const char *file, int line) 270{ 271 272 MPASS(curthread != NULL); 273 KASSERT(m->mtx_object.lo_class == &lock_class_mtx_sleep, 274 ("mtx_unlock() of spin mutex %s @ %s:%d", m->mtx_object.lo_name, 275 file, line)); 276 WITNESS_UNLOCK(&m->mtx_object, opts | LOP_EXCLUSIVE, file, line); 277 LOCK_LOG_LOCK("UNLOCK", &m->mtx_object, opts, m->mtx_recurse, file, 278 line); 279 mtx_assert(m, MA_OWNED); 280#ifdef MUTEX_PROFILING 281 if (m->mtx_acqtime != 0) { 282 static const char *unknown = "(unknown)"; 283 struct mutex_prof *mpp; 284 u_int64_t acqtime, now; 285 const char *p, *q; 286 volatile u_int hash; 287 288 now = nanoseconds(); 289 acqtime = m->mtx_acqtime; 290 m->mtx_acqtime = 0; 291 if (now <= acqtime) 292 goto out; 293 for (p = m->mtx_filename; 294 p != NULL && strncmp(p, "../", 3) == 0; p += 3) 295 /* nothing */ ; 296 if (p == NULL || *p == '\0') 297 p = unknown; 298 for (hash = m->mtx_lineno, q = p; *q != '\0'; ++q) 299 hash = (hash * 2 + *q) % MPROF_HASH_SIZE; 300 mtx_lock_spin(&mprof_mtx); 301 for (mpp = mprof_hash[hash]; mpp != NULL; mpp = mpp->next) 302 if (mpp->line == m->mtx_lineno && 303 strcmp(mpp->file, p) == 0) 304 break; 305 if (mpp == NULL) { 306 /* Just exit if we cannot get a trace buffer */ 307 if (first_free_mprof_buf >= NUM_MPROF_BUFFERS) { 308 ++mutex_prof_rejected; 309 goto unlock; 310 } 311 mpp = &mprof_buf[first_free_mprof_buf++]; 312 mpp->name = mtx_name(m); 313 mpp->file = p; 314 mpp->line = m->mtx_lineno; 315 mpp->next = mprof_hash[hash]; 316 if (mprof_hash[hash] != NULL) 317 ++mutex_prof_collisions; 318 mprof_hash[hash] = mpp; 319 ++mutex_prof_records; 320 } 321 /* 322 * Record if the mutex has been held longer now than ever 323 * before. 324 */ 325 if (now - acqtime > mpp->cnt_max) 326 mpp->cnt_max = now - acqtime; 327 mpp->cnt_tot += now - acqtime; 328 mpp->cnt_cur++; 329 /* 330 * There's a small race, really we should cmpxchg 331 * 0 with the current value, but that would bill 332 * the contention to the wrong lock instance if 333 * it followed this also. 334 */ 335 mpp->cnt_contest_holding += m->mtx_contest_holding; 336 m->mtx_contest_holding = 0; 337 mpp->cnt_contest_locking += m->mtx_contest_locking; 338 m->mtx_contest_locking = 0; 339unlock: 340 mtx_unlock_spin(&mprof_mtx); 341 } 342out: 343#endif 344 _rel_sleep_lock(m, curthread, opts, file, line); 345} 346 347void 348_mtx_lock_spin_flags(struct mtx *m, int opts, const char *file, int line) 349{ 350 351 MPASS(curthread != NULL); 352 KASSERT(m->mtx_object.lo_class == &lock_class_mtx_spin, 353 ("mtx_lock_spin() of sleep mutex %s @ %s:%d", 354 m->mtx_object.lo_name, file, line)); 355 WITNESS_CHECKORDER(&m->mtx_object, opts | LOP_NEWORDER | LOP_EXCLUSIVE, 356 file, line); 357#if defined(SMP) || LOCK_DEBUG > 0 || 1 358 _get_spin_lock(m, curthread, opts, file, line); 359#else 360 critical_enter(); 361#endif 362 LOCK_LOG_LOCK("LOCK", &m->mtx_object, opts, m->mtx_recurse, file, 363 line); 364 WITNESS_LOCK(&m->mtx_object, opts | LOP_EXCLUSIVE, file, line); 365} 366 367void 368_mtx_unlock_spin_flags(struct mtx *m, int opts, const char *file, int line) 369{ 370 371 MPASS(curthread != NULL); 372 KASSERT(m->mtx_object.lo_class == &lock_class_mtx_spin, 373 ("mtx_unlock_spin() of sleep mutex %s @ %s:%d", 374 m->mtx_object.lo_name, file, line)); 375 WITNESS_UNLOCK(&m->mtx_object, opts | LOP_EXCLUSIVE, file, line); 376 LOCK_LOG_LOCK("UNLOCK", &m->mtx_object, opts, m->mtx_recurse, file, 377 line); 378 mtx_assert(m, MA_OWNED); 379#if defined(SMP) || LOCK_DEBUG > 0 || 1 380 _rel_spin_lock(m); 381#else 382 critical_exit(); 383#endif 384} 385 386/* 387 * The important part of mtx_trylock{,_flags}() 388 * Tries to acquire lock `m.' If this function is called on a mutex that 389 * is already owned, it will recursively acquire the lock. 390 */ 391int 392_mtx_trylock(struct mtx *m, int opts, const char *file, int line) 393{ 394 int rval; 395 396 MPASS(curthread != NULL); 397 398 if (mtx_owned(m) && (m->mtx_object.lo_flags & LO_RECURSABLE) != 0) { 399 m->mtx_recurse++; 400 atomic_set_ptr(&m->mtx_lock, MTX_RECURSED); 401 rval = 1; 402 } else 403 rval = _obtain_lock(m, curthread); 404 405 LOCK_LOG_TRY("LOCK", &m->mtx_object, opts, rval, file, line); 406 if (rval) 407 WITNESS_LOCK(&m->mtx_object, opts | LOP_EXCLUSIVE | LOP_TRYLOCK, 408 file, line); 409 410 return (rval); 411} 412 413/* 414 * _mtx_lock_sleep: the tougher part of acquiring an MTX_DEF lock. 415 * 416 * We call this if the lock is either contested (i.e. we need to go to 417 * sleep waiting for it), or if we need to recurse on it. 418 */ 419void 420_mtx_lock_sleep(struct mtx *m, int opts, const char *file, int line) 421{ 422 struct turnstile *ts; 423 struct thread *td = curthread; 424#if defined(SMP) && !defined(NO_ADAPTIVE_MUTEXES) 425 struct thread *owner; 426#endif 427 uintptr_t v; 428#ifdef KTR 429 int cont_logged = 0; 430#endif 431#ifdef MUTEX_PROFILING 432 int contested; 433#endif 434 435 if (mtx_owned(m)) { 436 KASSERT((m->mtx_object.lo_flags & LO_RECURSABLE) != 0, 437 ("_mtx_lock_sleep: recursed on non-recursive mutex %s @ %s:%d\n", 438 m->mtx_object.lo_name, file, line)); 439 m->mtx_recurse++; 440 atomic_set_ptr(&m->mtx_lock, MTX_RECURSED); 441 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 442 CTR1(KTR_LOCK, "_mtx_lock_sleep: %p recursing", m); 443 return; 444 } 445 446 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 447 CTR4(KTR_LOCK, 448 "_mtx_lock_sleep: %s contested (lock=%p) at %s:%d", 449 m->mtx_object.lo_name, (void *)m->mtx_lock, file, line); 450 451#ifdef MUTEX_PROFILING 452 contested = 0; 453#endif 454 while (!_obtain_lock(m, td)) { 455#ifdef MUTEX_PROFILING 456 contested = 1; 457 atomic_add_int(&m->mtx_contest_holding, 1); 458#endif 459 ts = turnstile_lookup(&m->mtx_object); 460 v = m->mtx_lock; 461 462 /* 463 * Check if the lock has been released while spinning for 464 * the turnstile chain lock. 465 */ 466 if (v == MTX_UNOWNED) { 467 turnstile_release(&m->mtx_object); 468#if defined(__i386__) || defined(__amd64__) 469 ia32_pause(); 470#endif 471 continue; 472 } 473 474#ifdef MUTEX_WAKE_ALL 475 MPASS(v != MTX_CONTESTED); 476#else 477 /* 478 * The mutex was marked contested on release. This means that 479 * there are other threads blocked on it. Grab ownership of 480 * it and propagate its priority to the current thread if 481 * necessary. 482 */ 483 if (v == MTX_CONTESTED) { 484 MPASS(ts != NULL); 485 m->mtx_lock = (uintptr_t)td | MTX_CONTESTED; 486 turnstile_claim(ts); 487 break; 488 } 489#endif 490 491 /* 492 * If the mutex isn't already contested and a failure occurs 493 * setting the contested bit, the mutex was either released 494 * or the state of the MTX_RECURSED bit changed. 495 */ 496 if ((v & MTX_CONTESTED) == 0 && 497 !atomic_cmpset_ptr(&m->mtx_lock, (void *)v, 498 (void *)(v | MTX_CONTESTED))) { 499 turnstile_release(&m->mtx_object); 500#if defined(__i386__) || defined(__amd64__) 501 ia32_pause(); 502#endif 503 continue; 504 } 505 506#if defined(SMP) && !defined(NO_ADAPTIVE_MUTEXES) 507 /* 508 * If the current owner of the lock is executing on another 509 * CPU, spin instead of blocking. 510 */ 511 owner = (struct thread *)(v & MTX_FLAGMASK); 512#ifdef ADAPTIVE_GIANT 513 if (TD_IS_RUNNING(owner)) { 514#else 515 if (m != &Giant && TD_IS_RUNNING(owner)) { 516#endif 517 turnstile_release(&m->mtx_object); 518 while (mtx_owner(m) == owner && TD_IS_RUNNING(owner)) { 519#if defined(__i386__) || defined(__amd64__) 520 ia32_pause(); 521#endif 522 } 523 continue; 524 } 525#endif /* SMP && !NO_ADAPTIVE_MUTEXES */ 526 527 /* 528 * We definitely must sleep for this lock. 529 */ 530 mtx_assert(m, MA_NOTOWNED); 531 532#ifdef KTR 533 if (!cont_logged) { 534 CTR6(KTR_CONTENTION, 535 "contention: %p at %s:%d wants %s, taken by %s:%d", 536 td, file, line, m->mtx_object.lo_name, 537 WITNESS_FILE(&m->mtx_object), 538 WITNESS_LINE(&m->mtx_object)); 539 cont_logged = 1; 540 } 541#endif 542 543 /* 544 * Block on the turnstile. 545 */ 546 turnstile_wait(ts, &m->mtx_object, mtx_owner(m)); 547 } 548 549#ifdef KTR 550 if (cont_logged) { 551 CTR4(KTR_CONTENTION, 552 "contention end: %s acquired by %p at %s:%d", 553 m->mtx_object.lo_name, td, file, line); 554 } 555#endif 556#ifdef MUTEX_PROFILING 557 if (contested) 558 m->mtx_contest_locking++; 559 m->mtx_contest_holding = 0; 560#endif 561 return; 562} 563 564/* 565 * _mtx_lock_spin: the tougher part of acquiring an MTX_SPIN lock. 566 * 567 * This is only called if we need to actually spin for the lock. Recursion 568 * is handled inline. 569 */ 570void 571_mtx_lock_spin(struct mtx *m, int opts, const char *file, int line) 572{ 573 int i = 0; 574 575 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 576 CTR1(KTR_LOCK, "_mtx_lock_spin: %p spinning", m); 577 578 for (;;) { 579 if (_obtain_lock(m, curthread)) 580 break; 581 582 /* Give interrupts a chance while we spin. */ 583 critical_exit(); 584 while (m->mtx_lock != MTX_UNOWNED) { 585 if (i++ < 10000000) { 586#if defined(__i386__) || defined(__amd64__) 587 ia32_pause(); 588#endif 589 continue; 590 } 591 if (i < 60000000) 592 DELAY(1); 593 else if (!kdb_active) { 594 printf("spin lock %s held by %p for > 5 seconds\n", 595 m->mtx_object.lo_name, (void *)m->mtx_lock); 596#ifdef WITNESS 597 witness_display_spinlock(&m->mtx_object, 598 mtx_owner(m)); 599#endif 600 panic("spin lock held too long"); 601 } 602#if defined(__i386__) || defined(__amd64__) 603 ia32_pause(); 604#endif 605 } 606 critical_enter(); 607 } 608 609 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 610 CTR1(KTR_LOCK, "_mtx_lock_spin: %p spin done", m); 611 612 return; 613} 614 615/* 616 * _mtx_unlock_sleep: the tougher part of releasing an MTX_DEF lock. 617 * 618 * We are only called here if the lock is recursed or contested (i.e. we 619 * need to wake up a blocked thread). 620 */ 621void 622_mtx_unlock_sleep(struct mtx *m, int opts, const char *file, int line) 623{ 624 struct turnstile *ts; 625#ifndef PREEMPTION 626 struct thread *td, *td1; 627#endif 628 629 if (mtx_recursed(m)) { 630 if (--(m->mtx_recurse) == 0) 631 atomic_clear_ptr(&m->mtx_lock, MTX_RECURSED); 632 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 633 CTR1(KTR_LOCK, "_mtx_unlock_sleep: %p unrecurse", m); 634 return; 635 } 636 637 ts = turnstile_lookup(&m->mtx_object); 638 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 639 CTR1(KTR_LOCK, "_mtx_unlock_sleep: %p contested", m); 640 641#if defined(SMP) && !defined(NO_ADAPTIVE_MUTEXES) 642 if (ts == NULL) { 643 _release_lock_quick(m); 644 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 645 CTR1(KTR_LOCK, "_mtx_unlock_sleep: %p no sleepers", m); 646 turnstile_release(&m->mtx_object); 647 return; 648 } 649#else 650 MPASS(ts != NULL); 651#endif 652#ifndef PREEMPTION 653 /* XXX */ 654 td1 = turnstile_head(ts); 655#endif 656#ifdef MUTEX_WAKE_ALL 657 turnstile_broadcast(ts); 658 _release_lock_quick(m); 659#else 660 if (turnstile_signal(ts)) { 661 _release_lock_quick(m); 662 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 663 CTR1(KTR_LOCK, "_mtx_unlock_sleep: %p not held", m); 664 } else { 665 m->mtx_lock = MTX_CONTESTED; 666 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 667 CTR1(KTR_LOCK, "_mtx_unlock_sleep: %p still contested", 668 m); 669 } 670#endif 671 turnstile_unpend(ts); 672 673#ifndef PREEMPTION 674 /* 675 * XXX: This is just a hack until preemption is done. However, 676 * once preemption is done we need to either wrap the 677 * turnstile_signal() and release of the actual lock in an 678 * extra critical section or change the preemption code to 679 * always just set a flag and never do instant-preempts. 680 */ 681 td = curthread; 682 if (td->td_critnest > 0 || td1->td_priority >= td->td_priority) 683 return; 684 mtx_lock_spin(&sched_lock); 685 if (!TD_IS_RUNNING(td1)) { 686#ifdef notyet 687 if (td->td_ithd != NULL) { 688 struct ithd *it = td->td_ithd; 689 690 if (it->it_interrupted) { 691 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 692 CTR2(KTR_LOCK, 693 "_mtx_unlock_sleep: %p interrupted %p", 694 it, it->it_interrupted); 695 intr_thd_fixup(it); 696 } 697 } 698#endif 699 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 700 CTR2(KTR_LOCK, 701 "_mtx_unlock_sleep: %p switching out lock=%p", m, 702 (void *)m->mtx_lock); 703 704 mi_switch(SW_INVOL, NULL); 705 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 706 CTR2(KTR_LOCK, "_mtx_unlock_sleep: %p resuming lock=%p", 707 m, (void *)m->mtx_lock); 708 } 709 mtx_unlock_spin(&sched_lock); 710#endif 711 712 return; 713} 714 715/* 716 * All the unlocking of MTX_SPIN locks is done inline. 717 * See the _rel_spin_lock() macro for the details. 718 */ 719 720/* 721 * The backing function for the INVARIANTS-enabled mtx_assert() 722 */ 723#ifdef INVARIANT_SUPPORT 724void 725_mtx_assert(struct mtx *m, int what, const char *file, int line) 726{ 727 728 if (panicstr != NULL) 729 return; 730 switch (what) { 731 case MA_OWNED: 732 case MA_OWNED | MA_RECURSED: 733 case MA_OWNED | MA_NOTRECURSED: 734 if (!mtx_owned(m)) 735 panic("mutex %s not owned at %s:%d", 736 m->mtx_object.lo_name, file, line); 737 if (mtx_recursed(m)) { 738 if ((what & MA_NOTRECURSED) != 0) 739 panic("mutex %s recursed at %s:%d", 740 m->mtx_object.lo_name, file, line); 741 } else if ((what & MA_RECURSED) != 0) { 742 panic("mutex %s unrecursed at %s:%d", 743 m->mtx_object.lo_name, file, line); 744 } 745 break; 746 case MA_NOTOWNED: 747 if (mtx_owned(m)) 748 panic("mutex %s owned at %s:%d", 749 m->mtx_object.lo_name, file, line); 750 break; 751 default: 752 panic("unknown mtx_assert at %s:%d", file, line); 753 } 754} 755#endif 756 757/* 758 * The MUTEX_DEBUG-enabled mtx_validate() 759 * 760 * Most of these checks have been moved off into the LO_INITIALIZED flag 761 * maintained by the witness code. 762 */ 763#ifdef MUTEX_DEBUG 764 765void mtx_validate(struct mtx *); 766 767void 768mtx_validate(struct mtx *m) 769{ 770 771/* 772 * XXX: When kernacc() does not require Giant we can reenable this check 773 */ 774#ifdef notyet 775/* 776 * XXX - When kernacc() is fixed on the alpha to handle K0_SEG memory properly 777 * we can re-enable the kernacc() checks. 778 */ 779#ifndef __alpha__ 780 /* 781 * Can't call kernacc() from early init386(), especially when 782 * initializing Giant mutex, because some stuff in kernacc() 783 * requires Giant itself. 784 */ 785 if (!cold) 786 if (!kernacc((caddr_t)m, sizeof(m), 787 VM_PROT_READ | VM_PROT_WRITE)) 788 panic("Can't read and write to mutex %p", m); 789#endif 790#endif 791} 792#endif 793 794/* 795 * General init routine used by the MTX_SYSINIT() macro. 796 */ 797void 798mtx_sysinit(void *arg) 799{ 800 struct mtx_args *margs = arg; 801 802 mtx_init(margs->ma_mtx, margs->ma_desc, NULL, margs->ma_opts); 803} 804 805/* 806 * Mutex initialization routine; initialize lock `m' of type contained in 807 * `opts' with options contained in `opts' and name `name.' The optional 808 * lock type `type' is used as a general lock category name for use with 809 * witness. 810 */ 811void 812mtx_init(struct mtx *m, const char *name, const char *type, int opts) 813{ 814 struct lock_object *lock; 815 816 MPASS((opts & ~(MTX_SPIN | MTX_QUIET | MTX_RECURSE | 817 MTX_NOWITNESS | MTX_DUPOK)) == 0); 818 819#ifdef MUTEX_DEBUG 820 /* Diagnostic and error correction */ 821 mtx_validate(m); 822#endif 823 824 lock = &m->mtx_object; 825 KASSERT((lock->lo_flags & LO_INITIALIZED) == 0, 826 ("mutex \"%s\" %p already initialized", name, m)); 827 bzero(m, sizeof(*m)); 828 if (opts & MTX_SPIN) 829 lock->lo_class = &lock_class_mtx_spin; 830 else 831 lock->lo_class = &lock_class_mtx_sleep; 832 lock->lo_name = name; 833 lock->lo_type = type != NULL ? type : name; 834 if (opts & MTX_QUIET) 835 lock->lo_flags = LO_QUIET; 836 if (opts & MTX_RECURSE) 837 lock->lo_flags |= LO_RECURSABLE; 838 if ((opts & MTX_NOWITNESS) == 0) 839 lock->lo_flags |= LO_WITNESS; 840 if (opts & MTX_DUPOK) 841 lock->lo_flags |= LO_DUPOK; 842 843 m->mtx_lock = MTX_UNOWNED; 844 845 LOCK_LOG_INIT(lock, opts); 846 847 WITNESS_INIT(lock); 848} 849 850/* 851 * Remove lock `m' from all_mtx queue. We don't allow MTX_QUIET to be 852 * passed in as a flag here because if the corresponding mtx_init() was 853 * called with MTX_QUIET set, then it will already be set in the mutex's 854 * flags. 855 */ 856void 857mtx_destroy(struct mtx *m) 858{ 859 860 LOCK_LOG_DESTROY(&m->mtx_object, 0); 861 862 if (!mtx_owned(m)) 863 MPASS(mtx_unowned(m)); 864 else { 865 MPASS((m->mtx_lock & (MTX_RECURSED|MTX_CONTESTED)) == 0); 866 867 /* Tell witness this isn't locked to make it happy. */ 868 WITNESS_UNLOCK(&m->mtx_object, LOP_EXCLUSIVE, __FILE__, 869 __LINE__); 870 } 871 872 WITNESS_DESTROY(&m->mtx_object); 873} 874 875/* 876 * Intialize the mutex code and system mutexes. This is called from the MD 877 * startup code prior to mi_startup(). The per-CPU data space needs to be 878 * setup before this is called. 879 */ 880void 881mutex_init(void) 882{ 883 884 /* Setup thread0 so that mutexes work. */ 885 LIST_INIT(&thread0.td_contested); 886 887 /* Setup turnstiles so that sleep mutexes work. */ 888 init_turnstiles(); 889 890 /* 891 * Initialize mutexes. 892 */ 893 mtx_init(&Giant, "Giant", NULL, MTX_DEF | MTX_RECURSE); 894 mtx_init(&sched_lock, "sched lock", NULL, MTX_SPIN | MTX_RECURSE); 895 mtx_init(&proc0.p_mtx, "process lock", NULL, MTX_DEF | MTX_DUPOK); 896 mtx_lock(&Giant); 897} 898