kern_mutex.c revision 131473
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 131473 2004-07-02 19:09:50Z jhb $"); 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/kernel.h> 47#include <sys/ktr.h> 48#include <sys/lock.h> 49#include <sys/malloc.h> 50#include <sys/mutex.h> 51#include <sys/proc.h> 52#include <sys/resourcevar.h> 53#include <sys/sched.h> 54#include <sys/sbuf.h> 55#include <sys/sysctl.h> 56#include <sys/turnstile.h> 57#include <sys/vmmeter.h> 58 59#include <machine/atomic.h> 60#include <machine/bus.h> 61#include <machine/clock.h> 62#include <machine/cpu.h> 63 64#include <ddb/ddb.h> 65 66#include <vm/vm.h> 67#include <vm/vm_extern.h> 68 69/* 70 * Internal utility macros. 71 */ 72#define mtx_unowned(m) ((m)->mtx_lock == MTX_UNOWNED) 73 74#define mtx_owner(m) (mtx_unowned((m)) ? NULL \ 75 : (struct thread *)((m)->mtx_lock & MTX_FLAGMASK)) 76 77/* 78 * Lock classes for sleep and spin mutexes. 79 */ 80struct lock_class lock_class_mtx_sleep = { 81 "sleep mutex", 82 LC_SLEEPLOCK | LC_RECURSABLE 83}; 84struct lock_class lock_class_mtx_spin = { 85 "spin mutex", 86 LC_SPINLOCK | LC_RECURSABLE 87}; 88 89/* 90 * System-wide mutexes 91 */ 92struct mtx sched_lock; 93struct mtx Giant; 94 95#ifdef MUTEX_PROFILING 96SYSCTL_NODE(_debug, OID_AUTO, mutex, CTLFLAG_RD, NULL, "mutex debugging"); 97SYSCTL_NODE(_debug_mutex, OID_AUTO, prof, CTLFLAG_RD, NULL, "mutex profiling"); 98static int mutex_prof_enable = 0; 99SYSCTL_INT(_debug_mutex_prof, OID_AUTO, enable, CTLFLAG_RW, 100 &mutex_prof_enable, 0, "Enable tracing of mutex holdtime"); 101 102struct mutex_prof { 103 const char *name; 104 const char *file; 105 int line; 106 uintmax_t cnt_max; 107 uintmax_t cnt_tot; 108 uintmax_t cnt_cur; 109 uintmax_t cnt_contest_holding; 110 uintmax_t cnt_contest_locking; 111 struct mutex_prof *next; 112}; 113 114/* 115 * mprof_buf is a static pool of profiling records to avoid possible 116 * reentrance of the memory allocation functions. 117 * 118 * Note: NUM_MPROF_BUFFERS must be smaller than MPROF_HASH_SIZE. 119 */ 120#define NUM_MPROF_BUFFERS 1000 121static struct mutex_prof mprof_buf[NUM_MPROF_BUFFERS]; 122static int first_free_mprof_buf; 123#define MPROF_HASH_SIZE 1009 124static struct mutex_prof *mprof_hash[MPROF_HASH_SIZE]; 125/* SWAG: sbuf size = avg stat. line size * number of locks */ 126#define MPROF_SBUF_SIZE 256 * 400 127 128static int mutex_prof_acquisitions; 129SYSCTL_INT(_debug_mutex_prof, OID_AUTO, acquisitions, CTLFLAG_RD, 130 &mutex_prof_acquisitions, 0, "Number of mutex acquistions recorded"); 131static int mutex_prof_records; 132SYSCTL_INT(_debug_mutex_prof, OID_AUTO, records, CTLFLAG_RD, 133 &mutex_prof_records, 0, "Number of profiling records"); 134static int mutex_prof_maxrecords = NUM_MPROF_BUFFERS; 135SYSCTL_INT(_debug_mutex_prof, OID_AUTO, maxrecords, CTLFLAG_RD, 136 &mutex_prof_maxrecords, 0, "Maximum number of profiling records"); 137static int mutex_prof_rejected; 138SYSCTL_INT(_debug_mutex_prof, OID_AUTO, rejected, CTLFLAG_RD, 139 &mutex_prof_rejected, 0, "Number of rejected profiling records"); 140static int mutex_prof_hashsize = MPROF_HASH_SIZE; 141SYSCTL_INT(_debug_mutex_prof, OID_AUTO, hashsize, CTLFLAG_RD, 142 &mutex_prof_hashsize, 0, "Hash size"); 143static int mutex_prof_collisions = 0; 144SYSCTL_INT(_debug_mutex_prof, OID_AUTO, collisions, CTLFLAG_RD, 145 &mutex_prof_collisions, 0, "Number of hash collisions"); 146 147/* 148 * mprof_mtx protects the profiling buffers and the hash. 149 */ 150static struct mtx mprof_mtx; 151MTX_SYSINIT(mprof, &mprof_mtx, "mutex profiling lock", MTX_SPIN | MTX_QUIET); 152 153static u_int64_t 154nanoseconds(void) 155{ 156 struct timespec tv; 157 158 nanotime(&tv); 159 return (tv.tv_sec * (u_int64_t)1000000000 + tv.tv_nsec); 160} 161 162static int 163dump_mutex_prof_stats(SYSCTL_HANDLER_ARGS) 164{ 165 struct sbuf *sb; 166 int error, i; 167 static int multiplier = 1; 168 169 if (first_free_mprof_buf == 0) 170 return (SYSCTL_OUT(req, "No locking recorded", 171 sizeof("No locking recorded"))); 172 173retry_sbufops: 174 sb = sbuf_new(NULL, NULL, MPROF_SBUF_SIZE * multiplier, SBUF_FIXEDLEN); 175 sbuf_printf(sb, "%6s %12s %11s %5s %12s %12s %s\n", 176 "max", "total", "count", "avg", "cnt_hold", "cnt_lock", "name"); 177 /* 178 * XXX this spinlock seems to be by far the largest perpetrator 179 * of spinlock latency (1.6 msec on an Athlon1600 was recorded 180 * even before I pessimized it further by moving the average 181 * computation here). 182 */ 183 mtx_lock_spin(&mprof_mtx); 184 for (i = 0; i < first_free_mprof_buf; ++i) { 185 sbuf_printf(sb, "%6ju %12ju %11ju %5ju %12ju %12ju %s:%d (%s)\n", 186 mprof_buf[i].cnt_max / 1000, 187 mprof_buf[i].cnt_tot / 1000, 188 mprof_buf[i].cnt_cur, 189 mprof_buf[i].cnt_cur == 0 ? (uintmax_t)0 : 190 mprof_buf[i].cnt_tot / (mprof_buf[i].cnt_cur * 1000), 191 mprof_buf[i].cnt_contest_holding, 192 mprof_buf[i].cnt_contest_locking, 193 mprof_buf[i].file, mprof_buf[i].line, mprof_buf[i].name); 194 if (sbuf_overflowed(sb)) { 195 mtx_unlock_spin(&mprof_mtx); 196 sbuf_delete(sb); 197 multiplier++; 198 goto retry_sbufops; 199 } 200 } 201 mtx_unlock_spin(&mprof_mtx); 202 sbuf_finish(sb); 203 error = SYSCTL_OUT(req, sbuf_data(sb), sbuf_len(sb) + 1); 204 sbuf_delete(sb); 205 return (error); 206} 207SYSCTL_PROC(_debug_mutex_prof, OID_AUTO, stats, CTLTYPE_STRING | CTLFLAG_RD, 208 NULL, 0, dump_mutex_prof_stats, "A", "Mutex profiling statistics"); 209 210static int 211reset_mutex_prof_stats(SYSCTL_HANDLER_ARGS) 212{ 213 int error, v; 214 215 if (first_free_mprof_buf == 0) 216 return (0); 217 218 v = 0; 219 error = sysctl_handle_int(oidp, &v, 0, req); 220 if (error) 221 return (error); 222 if (req->newptr == NULL) 223 return (error); 224 if (v == 0) 225 return (0); 226 227 mtx_lock_spin(&mprof_mtx); 228 bzero(mprof_buf, sizeof(*mprof_buf) * first_free_mprof_buf); 229 bzero(mprof_hash, sizeof(struct mtx *) * MPROF_HASH_SIZE); 230 first_free_mprof_buf = 0; 231 mtx_unlock_spin(&mprof_mtx); 232 return (0); 233} 234SYSCTL_PROC(_debug_mutex_prof, OID_AUTO, reset, CTLTYPE_INT | CTLFLAG_RW, 235 NULL, 0, reset_mutex_prof_stats, "I", "Reset mutex profiling statistics"); 236#endif 237 238/* 239 * Function versions of the inlined __mtx_* macros. These are used by 240 * modules and can also be called from assembly language if needed. 241 */ 242void 243_mtx_lock_flags(struct mtx *m, int opts, const char *file, int line) 244{ 245 246 MPASS(curthread != NULL); 247 KASSERT(m->mtx_object.lo_class == &lock_class_mtx_sleep, 248 ("mtx_lock() of spin mutex %s @ %s:%d", m->mtx_object.lo_name, 249 file, line)); 250 WITNESS_CHECKORDER(&m->mtx_object, opts | LOP_NEWORDER | LOP_EXCLUSIVE, 251 file, line); 252 _get_sleep_lock(m, curthread, opts, file, line); 253 LOCK_LOG_LOCK("LOCK", &m->mtx_object, opts, m->mtx_recurse, file, 254 line); 255 WITNESS_LOCK(&m->mtx_object, opts | LOP_EXCLUSIVE, file, line); 256#ifdef MUTEX_PROFILING 257 /* don't reset the timer when/if recursing */ 258 if (m->mtx_acqtime == 0) { 259 m->mtx_filename = file; 260 m->mtx_lineno = line; 261 m->mtx_acqtime = mutex_prof_enable ? nanoseconds() : 0; 262 ++mutex_prof_acquisitions; 263 } 264#endif 265} 266 267void 268_mtx_unlock_flags(struct mtx *m, int opts, const char *file, int line) 269{ 270 271 MPASS(curthread != NULL); 272 KASSERT(m->mtx_object.lo_class == &lock_class_mtx_sleep, 273 ("mtx_unlock() of spin mutex %s @ %s:%d", m->mtx_object.lo_name, 274 file, line)); 275 WITNESS_UNLOCK(&m->mtx_object, opts | LOP_EXCLUSIVE, file, line); 276 LOCK_LOG_LOCK("UNLOCK", &m->mtx_object, opts, m->mtx_recurse, file, 277 line); 278 mtx_assert(m, MA_OWNED); 279#ifdef MUTEX_PROFILING 280 if (m->mtx_acqtime != 0) { 281 static const char *unknown = "(unknown)"; 282 struct mutex_prof *mpp; 283 u_int64_t acqtime, now; 284 const char *p, *q; 285 volatile u_int hash; 286 287 now = nanoseconds(); 288 acqtime = m->mtx_acqtime; 289 m->mtx_acqtime = 0; 290 if (now <= acqtime) 291 goto out; 292 for (p = m->mtx_filename; 293 p != NULL && strncmp(p, "../", 3) == 0; p += 3) 294 /* nothing */ ; 295 if (p == NULL || *p == '\0') 296 p = unknown; 297 for (hash = m->mtx_lineno, q = p; *q != '\0'; ++q) 298 hash = (hash * 2 + *q) % MPROF_HASH_SIZE; 299 mtx_lock_spin(&mprof_mtx); 300 for (mpp = mprof_hash[hash]; mpp != NULL; mpp = mpp->next) 301 if (mpp->line == m->mtx_lineno && 302 strcmp(mpp->file, p) == 0) 303 break; 304 if (mpp == NULL) { 305 /* Just exit if we cannot get a trace buffer */ 306 if (first_free_mprof_buf >= NUM_MPROF_BUFFERS) { 307 ++mutex_prof_rejected; 308 goto unlock; 309 } 310 mpp = &mprof_buf[first_free_mprof_buf++]; 311 mpp->name = mtx_name(m); 312 mpp->file = p; 313 mpp->line = m->mtx_lineno; 314 mpp->next = mprof_hash[hash]; 315 if (mprof_hash[hash] != NULL) 316 ++mutex_prof_collisions; 317 mprof_hash[hash] = mpp; 318 ++mutex_prof_records; 319 } 320 /* 321 * Record if the mutex has been held longer now than ever 322 * before. 323 */ 324 if (now - acqtime > mpp->cnt_max) 325 mpp->cnt_max = now - acqtime; 326 mpp->cnt_tot += now - acqtime; 327 mpp->cnt_cur++; 328 /* 329 * There's a small race, really we should cmpxchg 330 * 0 with the current value, but that would bill 331 * the contention to the wrong lock instance if 332 * it followed this also. 333 */ 334 mpp->cnt_contest_holding += m->mtx_contest_holding; 335 m->mtx_contest_holding = 0; 336 mpp->cnt_contest_locking += m->mtx_contest_locking; 337 m->mtx_contest_locking = 0; 338unlock: 339 mtx_unlock_spin(&mprof_mtx); 340 } 341out: 342#endif 343 _rel_sleep_lock(m, curthread, opts, file, line); 344} 345 346void 347_mtx_lock_spin_flags(struct mtx *m, int opts, const char *file, int line) 348{ 349 350 MPASS(curthread != NULL); 351 KASSERT(m->mtx_object.lo_class == &lock_class_mtx_spin, 352 ("mtx_lock_spin() of sleep mutex %s @ %s:%d", 353 m->mtx_object.lo_name, file, line)); 354 WITNESS_CHECKORDER(&m->mtx_object, opts | LOP_NEWORDER | LOP_EXCLUSIVE, 355 file, line); 356#if defined(SMP) || LOCK_DEBUG > 0 || 1 357 _get_spin_lock(m, curthread, opts, file, line); 358#else 359 critical_enter(); 360#endif 361 LOCK_LOG_LOCK("LOCK", &m->mtx_object, opts, m->mtx_recurse, file, 362 line); 363 WITNESS_LOCK(&m->mtx_object, opts | LOP_EXCLUSIVE, file, line); 364} 365 366void 367_mtx_unlock_spin_flags(struct mtx *m, int opts, const char *file, int line) 368{ 369 370 MPASS(curthread != NULL); 371 KASSERT(m->mtx_object.lo_class == &lock_class_mtx_spin, 372 ("mtx_unlock_spin() of sleep mutex %s @ %s:%d", 373 m->mtx_object.lo_name, file, line)); 374 WITNESS_UNLOCK(&m->mtx_object, opts | LOP_EXCLUSIVE, file, line); 375 LOCK_LOG_LOCK("UNLOCK", &m->mtx_object, opts, m->mtx_recurse, file, 376 line); 377 mtx_assert(m, MA_OWNED); 378#if defined(SMP) || LOCK_DEBUG > 0 || 1 379 _rel_spin_lock(m); 380#else 381 critical_exit(); 382#endif 383} 384 385/* 386 * The important part of mtx_trylock{,_flags}() 387 * Tries to acquire lock `m.' If this function is called on a mutex that 388 * is already owned, it will recursively acquire the lock. 389 */ 390int 391_mtx_trylock(struct mtx *m, int opts, const char *file, int line) 392{ 393 int rval; 394 395 MPASS(curthread != NULL); 396 397 if (mtx_owned(m) && (m->mtx_object.lo_flags & LO_RECURSABLE) != 0) { 398 m->mtx_recurse++; 399 atomic_set_ptr(&m->mtx_lock, MTX_RECURSED); 400 rval = 1; 401 } else 402 rval = _obtain_lock(m, curthread); 403 404 LOCK_LOG_TRY("LOCK", &m->mtx_object, opts, rval, file, line); 405 if (rval) 406 WITNESS_LOCK(&m->mtx_object, opts | LOP_EXCLUSIVE | LOP_TRYLOCK, 407 file, line); 408 409 return (rval); 410} 411 412/* 413 * _mtx_lock_sleep: the tougher part of acquiring an MTX_DEF lock. 414 * 415 * We call this if the lock is either contested (i.e. we need to go to 416 * sleep waiting for it), or if we need to recurse on it. 417 */ 418void 419_mtx_lock_sleep(struct mtx *m, int opts, const char *file, int line) 420{ 421 struct turnstile *ts; 422 struct thread *td = curthread; 423#if defined(SMP) && defined(ADAPTIVE_MUTEXES) 424 struct thread *owner; 425#endif 426 uintptr_t v; 427#ifdef KTR 428 int cont_logged = 0; 429#endif 430#ifdef MUTEX_PROFILING 431 int contested; 432#endif 433 434 if (mtx_owned(m)) { 435 KASSERT((m->mtx_object.lo_flags & LO_RECURSABLE) != 0, 436 ("_mtx_lock_sleep: recursed on non-recursive mutex %s @ %s:%d\n", 437 m->mtx_object.lo_name, file, line)); 438 m->mtx_recurse++; 439 atomic_set_ptr(&m->mtx_lock, MTX_RECURSED); 440 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 441 CTR1(KTR_LOCK, "_mtx_lock_sleep: %p recursing", m); 442 return; 443 } 444 445 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 446 CTR4(KTR_LOCK, 447 "_mtx_lock_sleep: %s contested (lock=%p) at %s:%d", 448 m->mtx_object.lo_name, (void *)m->mtx_lock, file, line); 449 450#ifdef MUTEX_PROFILING 451 contested = 0; 452#endif 453 while (!_obtain_lock(m, td)) { 454#ifdef MUTEX_PROFILING 455 contested = 1; 456 atomic_add_int(&m->mtx_contest_holding, 1); 457#endif 458 ts = turnstile_lookup(&m->mtx_object); 459 v = m->mtx_lock; 460 461 /* 462 * Check if the lock has been released while spinning for 463 * the turnstile chain lock. 464 */ 465 if (v == MTX_UNOWNED) { 466 turnstile_release(&m->mtx_object); 467#ifdef __i386__ 468 ia32_pause(); 469#endif 470 continue; 471 } 472 473#ifdef MUTEX_WAKE_ALL 474 MPASS(v != MTX_CONTESTED); 475#else 476 /* 477 * The mutex was marked contested on release. This means that 478 * there are other threads blocked on it. Grab ownership of 479 * it and propagate its priority to the current thread if 480 * necessary. 481 */ 482 if (v == MTX_CONTESTED) { 483 MPASS(ts != NULL); 484 m->mtx_lock = (uintptr_t)td | MTX_CONTESTED; 485 turnstile_claim(ts); 486 break; 487 } 488#endif 489 490 /* 491 * If the mutex isn't already contested and a failure occurs 492 * setting the contested bit, the mutex was either released 493 * or the state of the MTX_RECURSED bit changed. 494 */ 495 if ((v & MTX_CONTESTED) == 0 && 496 !atomic_cmpset_ptr(&m->mtx_lock, (void *)v, 497 (void *)(v | MTX_CONTESTED))) { 498 turnstile_release(&m->mtx_object); 499#ifdef __i386__ 500 ia32_pause(); 501#endif 502 continue; 503 } 504 505#if defined(SMP) && defined(ADAPTIVE_MUTEXES) 506 /* 507 * If the current owner of the lock is executing on another 508 * CPU, spin instead of blocking. 509 */ 510 owner = (struct thread *)(v & MTX_FLAGMASK); 511 if (m != &Giant && TD_IS_RUNNING(owner)) { 512 turnstile_release(&m->mtx_object); 513 while (mtx_owner(m) == owner && TD_IS_RUNNING(owner)) { 514#ifdef __i386__ 515 ia32_pause(); 516#endif 517 } 518 continue; 519 } 520#endif /* SMP && ADAPTIVE_MUTEXES */ 521 522 /* 523 * We definitely must sleep for this lock. 524 */ 525 mtx_assert(m, MA_NOTOWNED); 526 527#ifdef KTR 528 if (!cont_logged) { 529 CTR6(KTR_CONTENTION, 530 "contention: %p at %s:%d wants %s, taken by %s:%d", 531 td, file, line, m->mtx_object.lo_name, 532 WITNESS_FILE(&m->mtx_object), 533 WITNESS_LINE(&m->mtx_object)); 534 cont_logged = 1; 535 } 536#endif 537 538 /* 539 * Block on the turnstile. 540 */ 541 turnstile_wait(ts, &m->mtx_object, mtx_owner(m)); 542 } 543 544#ifdef KTR 545 if (cont_logged) { 546 CTR4(KTR_CONTENTION, 547 "contention end: %s acquired by %p at %s:%d", 548 m->mtx_object.lo_name, td, file, line); 549 } 550#endif 551#ifdef MUTEX_PROFILING 552 if (contested) 553 m->mtx_contest_locking++; 554 m->mtx_contest_holding = 0; 555#endif 556 return; 557} 558 559/* 560 * _mtx_lock_spin: the tougher part of acquiring an MTX_SPIN lock. 561 * 562 * This is only called if we need to actually spin for the lock. Recursion 563 * is handled inline. 564 */ 565void 566_mtx_lock_spin(struct mtx *m, int opts, const char *file, int line) 567{ 568 int i = 0; 569 570 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 571 CTR1(KTR_LOCK, "_mtx_lock_spin: %p spinning", m); 572 573 for (;;) { 574 if (_obtain_lock(m, curthread)) 575 break; 576 577 /* Give interrupts a chance while we spin. */ 578 critical_exit(); 579 while (m->mtx_lock != MTX_UNOWNED) { 580 if (i++ < 10000000) { 581#ifdef __i386__ 582 ia32_pause(); 583#endif 584 continue; 585 } 586 if (i < 60000000) 587 DELAY(1); 588#ifdef DDB 589 else if (!db_active) { 590#else 591 else { 592#endif 593 printf("spin lock %s held by %p for > 5 seconds\n", 594 m->mtx_object.lo_name, (void *)m->mtx_lock); 595#ifdef WITNESS 596 witness_display_spinlock(&m->mtx_object, 597 mtx_owner(m)); 598#endif 599 panic("spin lock held too long"); 600 } 601#ifdef __i386__ 602 ia32_pause(); 603#endif 604 } 605 critical_enter(); 606 } 607 608 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 609 CTR1(KTR_LOCK, "_mtx_lock_spin: %p spin done", m); 610 611 return; 612} 613 614/* 615 * _mtx_unlock_sleep: the tougher part of releasing an MTX_DEF lock. 616 * 617 * We are only called here if the lock is recursed or contested (i.e. we 618 * need to wake up a blocked thread). 619 */ 620void 621_mtx_unlock_sleep(struct mtx *m, int opts, const char *file, int line) 622{ 623 struct turnstile *ts; 624 struct thread *td, *td1; 625 626 if (mtx_recursed(m)) { 627 if (--(m->mtx_recurse) == 0) 628 atomic_clear_ptr(&m->mtx_lock, MTX_RECURSED); 629 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 630 CTR1(KTR_LOCK, "_mtx_unlock_sleep: %p unrecurse", m); 631 return; 632 } 633 634 ts = turnstile_lookup(&m->mtx_object); 635 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 636 CTR1(KTR_LOCK, "_mtx_unlock_sleep: %p contested", m); 637 638#if defined(SMP) && defined(ADAPTIVE_MUTEXES) 639 if (ts == NULL) { 640 _release_lock_quick(m); 641 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 642 CTR1(KTR_LOCK, "_mtx_unlock_sleep: %p no sleepers", m); 643 turnstile_release(&m->mtx_object); 644 return; 645 } 646#else 647 MPASS(ts != NULL); 648#endif 649 /* XXX */ 650 td1 = turnstile_head(ts); 651#ifdef MUTEX_WAKE_ALL 652 turnstile_broadcast(ts); 653 _release_lock_quick(m); 654#else 655 if (turnstile_signal(ts)) { 656 _release_lock_quick(m); 657 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 658 CTR1(KTR_LOCK, "_mtx_unlock_sleep: %p not held", m); 659 } else { 660 m->mtx_lock = MTX_CONTESTED; 661 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 662 CTR1(KTR_LOCK, "_mtx_unlock_sleep: %p still contested", 663 m); 664 } 665#endif 666 turnstile_unpend(ts); 667 668 /* 669 * XXX: This is just a hack until preemption is done. However, 670 * once preemption is done we need to either wrap the 671 * turnstile_signal() and release of the actual lock in an 672 * extra critical section or change the preemption code to 673 * always just set a flag and never do instant-preempts. 674 */ 675 td = curthread; 676 if (td->td_critnest > 0 || td1->td_priority >= td->td_priority) 677 return; 678 mtx_lock_spin(&sched_lock); 679 if (!TD_IS_RUNNING(td1)) { 680#ifdef notyet 681 if (td->td_ithd != NULL) { 682 struct ithd *it = td->td_ithd; 683 684 if (it->it_interrupted) { 685 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 686 CTR2(KTR_LOCK, 687 "_mtx_unlock_sleep: %p interrupted %p", 688 it, it->it_interrupted); 689 intr_thd_fixup(it); 690 } 691 } 692#endif 693 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 694 CTR2(KTR_LOCK, 695 "_mtx_unlock_sleep: %p switching out lock=%p", m, 696 (void *)m->mtx_lock); 697 698 mi_switch(SW_INVOL, NULL); 699 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 700 CTR2(KTR_LOCK, "_mtx_unlock_sleep: %p resuming lock=%p", 701 m, (void *)m->mtx_lock); 702 } 703 mtx_unlock_spin(&sched_lock); 704 705 return; 706} 707 708/* 709 * All the unlocking of MTX_SPIN locks is done inline. 710 * See the _rel_spin_lock() macro for the details. 711 */ 712 713/* 714 * The backing function for the INVARIANTS-enabled mtx_assert() 715 */ 716#ifdef INVARIANT_SUPPORT 717void 718_mtx_assert(struct mtx *m, int what, const char *file, int line) 719{ 720 721 if (panicstr != NULL) 722 return; 723 switch (what) { 724 case MA_OWNED: 725 case MA_OWNED | MA_RECURSED: 726 case MA_OWNED | MA_NOTRECURSED: 727 if (!mtx_owned(m)) 728 panic("mutex %s not owned at %s:%d", 729 m->mtx_object.lo_name, file, line); 730 if (mtx_recursed(m)) { 731 if ((what & MA_NOTRECURSED) != 0) 732 panic("mutex %s recursed at %s:%d", 733 m->mtx_object.lo_name, file, line); 734 } else if ((what & MA_RECURSED) != 0) { 735 panic("mutex %s unrecursed at %s:%d", 736 m->mtx_object.lo_name, file, line); 737 } 738 break; 739 case MA_NOTOWNED: 740 if (mtx_owned(m)) 741 panic("mutex %s owned at %s:%d", 742 m->mtx_object.lo_name, file, line); 743 break; 744 default: 745 panic("unknown mtx_assert at %s:%d", file, line); 746 } 747} 748#endif 749 750/* 751 * The MUTEX_DEBUG-enabled mtx_validate() 752 * 753 * Most of these checks have been moved off into the LO_INITIALIZED flag 754 * maintained by the witness code. 755 */ 756#ifdef MUTEX_DEBUG 757 758void mtx_validate(struct mtx *); 759 760void 761mtx_validate(struct mtx *m) 762{ 763 764/* 765 * XXX: When kernacc() does not require Giant we can reenable this check 766 */ 767#ifdef notyet 768/* 769 * XXX - When kernacc() is fixed on the alpha to handle K0_SEG memory properly 770 * we can re-enable the kernacc() checks. 771 */ 772#ifndef __alpha__ 773 /* 774 * Can't call kernacc() from early init386(), especially when 775 * initializing Giant mutex, because some stuff in kernacc() 776 * requires Giant itself. 777 */ 778 if (!cold) 779 if (!kernacc((caddr_t)m, sizeof(m), 780 VM_PROT_READ | VM_PROT_WRITE)) 781 panic("Can't read and write to mutex %p", m); 782#endif 783#endif 784} 785#endif 786 787/* 788 * General init routine used by the MTX_SYSINIT() macro. 789 */ 790void 791mtx_sysinit(void *arg) 792{ 793 struct mtx_args *margs = arg; 794 795 mtx_init(margs->ma_mtx, margs->ma_desc, NULL, margs->ma_opts); 796} 797 798/* 799 * Mutex initialization routine; initialize lock `m' of type contained in 800 * `opts' with options contained in `opts' and name `name.' The optional 801 * lock type `type' is used as a general lock category name for use with 802 * witness. 803 */ 804void 805mtx_init(struct mtx *m, const char *name, const char *type, int opts) 806{ 807 struct lock_object *lock; 808 809 MPASS((opts & ~(MTX_SPIN | MTX_QUIET | MTX_RECURSE | 810 MTX_NOWITNESS | MTX_DUPOK)) == 0); 811 812#ifdef MUTEX_DEBUG 813 /* Diagnostic and error correction */ 814 mtx_validate(m); 815#endif 816 817 lock = &m->mtx_object; 818 KASSERT((lock->lo_flags & LO_INITIALIZED) == 0, 819 ("mutex \"%s\" %p already initialized", name, m)); 820 bzero(m, sizeof(*m)); 821 if (opts & MTX_SPIN) 822 lock->lo_class = &lock_class_mtx_spin; 823 else 824 lock->lo_class = &lock_class_mtx_sleep; 825 lock->lo_name = name; 826 lock->lo_type = type != NULL ? type : name; 827 if (opts & MTX_QUIET) 828 lock->lo_flags = LO_QUIET; 829 if (opts & MTX_RECURSE) 830 lock->lo_flags |= LO_RECURSABLE; 831 if ((opts & MTX_NOWITNESS) == 0) 832 lock->lo_flags |= LO_WITNESS; 833 if (opts & MTX_DUPOK) 834 lock->lo_flags |= LO_DUPOK; 835 836 m->mtx_lock = MTX_UNOWNED; 837 838 LOCK_LOG_INIT(lock, opts); 839 840 WITNESS_INIT(lock); 841} 842 843/* 844 * Remove lock `m' from all_mtx queue. We don't allow MTX_QUIET to be 845 * passed in as a flag here because if the corresponding mtx_init() was 846 * called with MTX_QUIET set, then it will already be set in the mutex's 847 * flags. 848 */ 849void 850mtx_destroy(struct mtx *m) 851{ 852 853 LOCK_LOG_DESTROY(&m->mtx_object, 0); 854 855 if (!mtx_owned(m)) 856 MPASS(mtx_unowned(m)); 857 else { 858 MPASS((m->mtx_lock & (MTX_RECURSED|MTX_CONTESTED)) == 0); 859 860 /* Tell witness this isn't locked to make it happy. */ 861 WITNESS_UNLOCK(&m->mtx_object, LOP_EXCLUSIVE, __FILE__, 862 __LINE__); 863 } 864 865 WITNESS_DESTROY(&m->mtx_object); 866} 867 868/* 869 * Intialize the mutex code and system mutexes. This is called from the MD 870 * startup code prior to mi_startup(). The per-CPU data space needs to be 871 * setup before this is called. 872 */ 873void 874mutex_init(void) 875{ 876 877 /* Setup thread0 so that mutexes work. */ 878 LIST_INIT(&thread0.td_contested); 879 880 /* Setup turnstiles so that sleep mutexes work. */ 881 init_turnstiles(); 882 883 /* 884 * Initialize mutexes. 885 */ 886 mtx_init(&Giant, "Giant", NULL, MTX_DEF | MTX_RECURSE); 887 mtx_init(&sched_lock, "sched lock", NULL, MTX_SPIN | MTX_RECURSE); 888 mtx_init(&proc0.p_mtx, "process lock", NULL, MTX_DEF | MTX_DUPOK); 889 mtx_lock(&Giant); 890} 891