kern_mutex.c revision 122514
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 122514 2003-11-11 22:07:29Z jhb $"); 38 39#include "opt_adaptive_mutexes.h" 40#include "opt_ddb.h" 41 42#include <sys/param.h> 43#include <sys/systm.h> 44#include <sys/bus.h> 45#include <sys/kernel.h> 46#include <sys/ktr.h> 47#include <sys/lock.h> 48#include <sys/malloc.h> 49#include <sys/mutex.h> 50#include <sys/proc.h> 51#include <sys/resourcevar.h> 52#include <sys/sched.h> 53#include <sys/sbuf.h> 54#include <sys/sysctl.h> 55#include <sys/turnstile.h> 56#include <sys/vmmeter.h> 57 58#include <machine/atomic.h> 59#include <machine/bus.h> 60#include <machine/clock.h> 61#include <machine/cpu.h> 62 63#include <ddb/ddb.h> 64 65#include <vm/vm.h> 66#include <vm/vm_extern.h> 67 68/* 69 * Internal utility macros. 70 */ 71#define mtx_unowned(m) ((m)->mtx_lock == MTX_UNOWNED) 72 73#define mtx_owner(m) (mtx_unowned((m)) ? NULL \ 74 : (struct thread *)((m)->mtx_lock & MTX_FLAGMASK)) 75 76/* 77 * Lock classes for sleep and spin mutexes. 78 */ 79struct lock_class lock_class_mtx_sleep = { 80 "sleep mutex", 81 LC_SLEEPLOCK | LC_RECURSABLE 82}; 83struct lock_class lock_class_mtx_spin = { 84 "spin mutex", 85 LC_SPINLOCK | LC_RECURSABLE 86}; 87 88/* 89 * System-wide mutexes 90 */ 91struct mtx sched_lock; 92struct mtx Giant; 93 94#ifdef MUTEX_PROFILING 95SYSCTL_NODE(_debug, OID_AUTO, mutex, CTLFLAG_RD, NULL, "mutex debugging"); 96SYSCTL_NODE(_debug_mutex, OID_AUTO, prof, CTLFLAG_RD, NULL, "mutex profiling"); 97static int mutex_prof_enable = 0; 98SYSCTL_INT(_debug_mutex_prof, OID_AUTO, enable, CTLFLAG_RW, 99 &mutex_prof_enable, 0, "Enable tracing of mutex holdtime"); 100 101struct mutex_prof { 102 const char *name; 103 const char *file; 104 int line; 105 uintmax_t cnt_max; 106 uintmax_t cnt_tot; 107 uintmax_t cnt_cur; 108 struct mutex_prof *next; 109}; 110 111/* 112 * mprof_buf is a static pool of profiling records to avoid possible 113 * reentrance of the memory allocation functions. 114 * 115 * Note: NUM_MPROF_BUFFERS must be smaller than MPROF_HASH_SIZE. 116 */ 117#define NUM_MPROF_BUFFERS 1000 118static struct mutex_prof mprof_buf[NUM_MPROF_BUFFERS]; 119static int first_free_mprof_buf; 120#define MPROF_HASH_SIZE 1009 121static struct mutex_prof *mprof_hash[MPROF_HASH_SIZE]; 122/* SWAG: sbuf size = avg stat. line size * number of locks */ 123#define MPROF_SBUF_SIZE 256 * 400 124 125static int mutex_prof_acquisitions; 126SYSCTL_INT(_debug_mutex_prof, OID_AUTO, acquisitions, CTLFLAG_RD, 127 &mutex_prof_acquisitions, 0, "Number of mutex acquistions recorded"); 128static int mutex_prof_records; 129SYSCTL_INT(_debug_mutex_prof, OID_AUTO, records, CTLFLAG_RD, 130 &mutex_prof_records, 0, "Number of profiling records"); 131static int mutex_prof_maxrecords = NUM_MPROF_BUFFERS; 132SYSCTL_INT(_debug_mutex_prof, OID_AUTO, maxrecords, CTLFLAG_RD, 133 &mutex_prof_maxrecords, 0, "Maximum number of profiling records"); 134static int mutex_prof_rejected; 135SYSCTL_INT(_debug_mutex_prof, OID_AUTO, rejected, CTLFLAG_RD, 136 &mutex_prof_rejected, 0, "Number of rejected profiling records"); 137static int mutex_prof_hashsize = MPROF_HASH_SIZE; 138SYSCTL_INT(_debug_mutex_prof, OID_AUTO, hashsize, CTLFLAG_RD, 139 &mutex_prof_hashsize, 0, "Hash size"); 140static int mutex_prof_collisions = 0; 141SYSCTL_INT(_debug_mutex_prof, OID_AUTO, collisions, CTLFLAG_RD, 142 &mutex_prof_collisions, 0, "Number of hash collisions"); 143 144/* 145 * mprof_mtx protects the profiling buffers and the hash. 146 */ 147static struct mtx mprof_mtx; 148MTX_SYSINIT(mprof, &mprof_mtx, "mutex profiling lock", MTX_SPIN | MTX_QUIET); 149 150static u_int64_t 151nanoseconds(void) 152{ 153 struct timespec tv; 154 155 nanotime(&tv); 156 return (tv.tv_sec * (u_int64_t)1000000000 + tv.tv_nsec); 157} 158 159static int 160dump_mutex_prof_stats(SYSCTL_HANDLER_ARGS) 161{ 162 struct sbuf *sb; 163 int error, i; 164 static int multiplier = 1; 165 166 if (first_free_mprof_buf == 0) 167 return (SYSCTL_OUT(req, "No locking recorded", 168 sizeof("No locking recorded"))); 169 170retry_sbufops: 171 sb = sbuf_new(NULL, NULL, MPROF_SBUF_SIZE * multiplier, SBUF_FIXEDLEN); 172 sbuf_printf(sb, "%6s %12s %11s %5s %s\n", 173 "max", "total", "count", "avg", "name"); 174 /* 175 * XXX this spinlock seems to be by far the largest perpetrator 176 * of spinlock latency (1.6 msec on an Athlon1600 was recorded 177 * even before I pessimized it further by moving the average 178 * computation here). 179 */ 180 mtx_lock_spin(&mprof_mtx); 181 for (i = 0; i < first_free_mprof_buf; ++i) { 182 sbuf_printf(sb, "%6ju %12ju %11ju %5ju %s:%d (%s)\n", 183 mprof_buf[i].cnt_max / 1000, 184 mprof_buf[i].cnt_tot / 1000, 185 mprof_buf[i].cnt_cur, 186 mprof_buf[i].cnt_cur == 0 ? (uintmax_t)0 : 187 mprof_buf[i].cnt_tot / (mprof_buf[i].cnt_cur * 1000), 188 mprof_buf[i].file, mprof_buf[i].line, mprof_buf[i].name); 189 if (sbuf_overflowed(sb)) { 190 mtx_unlock_spin(&mprof_mtx); 191 sbuf_delete(sb); 192 multiplier++; 193 goto retry_sbufops; 194 } 195 } 196 mtx_unlock_spin(&mprof_mtx); 197 sbuf_finish(sb); 198 error = SYSCTL_OUT(req, sbuf_data(sb), sbuf_len(sb) + 1); 199 sbuf_delete(sb); 200 return (error); 201} 202SYSCTL_PROC(_debug_mutex_prof, OID_AUTO, stats, CTLTYPE_STRING | CTLFLAG_RD, 203 NULL, 0, dump_mutex_prof_stats, "A", "Mutex profiling statistics"); 204#endif 205 206/* 207 * Function versions of the inlined __mtx_* macros. These are used by 208 * modules and can also be called from assembly language if needed. 209 */ 210void 211_mtx_lock_flags(struct mtx *m, int opts, const char *file, int line) 212{ 213 214 MPASS(curthread != NULL); 215 KASSERT(m->mtx_object.lo_class == &lock_class_mtx_sleep, 216 ("mtx_lock() of spin mutex %s @ %s:%d", m->mtx_object.lo_name, 217 file, line)); 218 _get_sleep_lock(m, curthread, opts, file, line); 219 LOCK_LOG_LOCK("LOCK", &m->mtx_object, opts, m->mtx_recurse, file, 220 line); 221 WITNESS_LOCK(&m->mtx_object, opts | LOP_EXCLUSIVE, file, line); 222#ifdef MUTEX_PROFILING 223 /* don't reset the timer when/if recursing */ 224 if (m->mtx_acqtime == 0) { 225 m->mtx_filename = file; 226 m->mtx_lineno = line; 227 m->mtx_acqtime = mutex_prof_enable ? nanoseconds() : 0; 228 ++mutex_prof_acquisitions; 229 } 230#endif 231} 232 233void 234_mtx_unlock_flags(struct mtx *m, int opts, const char *file, int line) 235{ 236 237 MPASS(curthread != NULL); 238 KASSERT(m->mtx_object.lo_class == &lock_class_mtx_sleep, 239 ("mtx_unlock() of spin mutex %s @ %s:%d", m->mtx_object.lo_name, 240 file, line)); 241 WITNESS_UNLOCK(&m->mtx_object, opts | LOP_EXCLUSIVE, file, line); 242 LOCK_LOG_LOCK("UNLOCK", &m->mtx_object, opts, m->mtx_recurse, file, 243 line); 244 mtx_assert(m, MA_OWNED); 245#ifdef MUTEX_PROFILING 246 if (m->mtx_acqtime != 0) { 247 static const char *unknown = "(unknown)"; 248 struct mutex_prof *mpp; 249 u_int64_t acqtime, now; 250 const char *p, *q; 251 volatile u_int hash; 252 253 now = nanoseconds(); 254 acqtime = m->mtx_acqtime; 255 m->mtx_acqtime = 0; 256 if (now <= acqtime) 257 goto out; 258 for (p = m->mtx_filename; 259 p != NULL && strncmp(p, "../", 3) == 0; p += 3) 260 /* nothing */ ; 261 if (p == NULL || *p == '\0') 262 p = unknown; 263 for (hash = m->mtx_lineno, q = p; *q != '\0'; ++q) 264 hash = (hash * 2 + *q) % MPROF_HASH_SIZE; 265 mtx_lock_spin(&mprof_mtx); 266 for (mpp = mprof_hash[hash]; mpp != NULL; mpp = mpp->next) 267 if (mpp->line == m->mtx_lineno && 268 strcmp(mpp->file, p) == 0) 269 break; 270 if (mpp == NULL) { 271 /* Just exit if we cannot get a trace buffer */ 272 if (first_free_mprof_buf >= NUM_MPROF_BUFFERS) { 273 ++mutex_prof_rejected; 274 goto unlock; 275 } 276 mpp = &mprof_buf[first_free_mprof_buf++]; 277 mpp->name = mtx_name(m); 278 mpp->file = p; 279 mpp->line = m->mtx_lineno; 280 mpp->next = mprof_hash[hash]; 281 if (mprof_hash[hash] != NULL) 282 ++mutex_prof_collisions; 283 mprof_hash[hash] = mpp; 284 ++mutex_prof_records; 285 } 286 /* 287 * Record if the mutex has been held longer now than ever 288 * before. 289 */ 290 if (now - acqtime > mpp->cnt_max) 291 mpp->cnt_max = now - acqtime; 292 mpp->cnt_tot += now - acqtime; 293 mpp->cnt_cur++; 294unlock: 295 mtx_unlock_spin(&mprof_mtx); 296 } 297out: 298#endif 299 _rel_sleep_lock(m, curthread, opts, file, line); 300} 301 302void 303_mtx_lock_spin_flags(struct mtx *m, int opts, const char *file, int line) 304{ 305 306 MPASS(curthread != NULL); 307 KASSERT(m->mtx_object.lo_class == &lock_class_mtx_spin, 308 ("mtx_lock_spin() of sleep mutex %s @ %s:%d", 309 m->mtx_object.lo_name, file, line)); 310#if defined(SMP) || LOCK_DEBUG > 0 || 1 311 _get_spin_lock(m, curthread, opts, file, line); 312#else 313 critical_enter(); 314#endif 315 LOCK_LOG_LOCK("LOCK", &m->mtx_object, opts, m->mtx_recurse, file, 316 line); 317 WITNESS_LOCK(&m->mtx_object, opts | LOP_EXCLUSIVE, file, line); 318} 319 320void 321_mtx_unlock_spin_flags(struct mtx *m, int opts, const char *file, int line) 322{ 323 324 MPASS(curthread != NULL); 325 KASSERT(m->mtx_object.lo_class == &lock_class_mtx_spin, 326 ("mtx_unlock_spin() of sleep mutex %s @ %s:%d", 327 m->mtx_object.lo_name, file, line)); 328 WITNESS_UNLOCK(&m->mtx_object, opts | LOP_EXCLUSIVE, file, line); 329 LOCK_LOG_LOCK("UNLOCK", &m->mtx_object, opts, m->mtx_recurse, file, 330 line); 331 mtx_assert(m, MA_OWNED); 332#if defined(SMP) || LOCK_DEBUG > 0 || 1 333 _rel_spin_lock(m); 334#else 335 critical_exit(); 336#endif 337} 338 339/* 340 * The important part of mtx_trylock{,_flags}() 341 * Tries to acquire lock `m.' We do NOT handle recursion here. If this 342 * function is called on a recursed mutex, it will return failure and 343 * will not recursively acquire the lock. You are expected to know what 344 * you are doing. 345 */ 346int 347_mtx_trylock(struct mtx *m, int opts, const char *file, int line) 348{ 349 int rval; 350 351 MPASS(curthread != NULL); 352 353 rval = _obtain_lock(m, curthread); 354 355 LOCK_LOG_TRY("LOCK", &m->mtx_object, opts, rval, file, line); 356 if (rval) 357 WITNESS_LOCK(&m->mtx_object, opts | LOP_EXCLUSIVE | LOP_TRYLOCK, 358 file, line); 359 360 return (rval); 361} 362 363/* 364 * _mtx_lock_sleep: the tougher part of acquiring an MTX_DEF lock. 365 * 366 * We call this if the lock is either contested (i.e. we need to go to 367 * sleep waiting for it), or if we need to recurse on it. 368 */ 369void 370_mtx_lock_sleep(struct mtx *m, int opts, const char *file, int line) 371{ 372 struct turnstile *ts; 373 struct thread *td = curthread; 374#if defined(SMP) && defined(ADAPTIVE_MUTEXES) 375 struct thread *owner; 376#endif 377 uintptr_t v; 378#ifdef KTR 379 int cont_logged = 0; 380#endif 381 382 if (mtx_owned(m)) { 383 m->mtx_recurse++; 384 atomic_set_ptr(&m->mtx_lock, MTX_RECURSED); 385 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 386 CTR1(KTR_LOCK, "_mtx_lock_sleep: %p recursing", m); 387 return; 388 } 389 390 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 391 CTR4(KTR_LOCK, 392 "_mtx_lock_sleep: %s contested (lock=%p) at %s:%d", 393 m->mtx_object.lo_name, (void *)m->mtx_lock, file, line); 394 395 while (!_obtain_lock(m, td)) { 396 397 ts = turnstile_lookup(&m->mtx_object); 398 v = m->mtx_lock; 399 400 /* 401 * Check if the lock has been released while spinning for 402 * the turnstile chain lock. 403 */ 404 if (v == MTX_UNOWNED) { 405 turnstile_release(&m->mtx_object); 406#ifdef __i386__ 407 ia32_pause(); 408#endif 409 continue; 410 } 411 412 /* 413 * The mutex was marked contested on release. This means that 414 * there are other threads blocked on it. Grab ownership of 415 * it and propagate its priority to the current thread if 416 * necessary. 417 */ 418 if (v == MTX_CONTESTED) { 419 MPASS(ts != NULL); 420 m->mtx_lock = (uintptr_t)td | MTX_CONTESTED; 421 turnstile_claim(ts); 422 return; 423 } 424 425 /* 426 * If the mutex isn't already contested and a failure occurs 427 * setting the contested bit, the mutex was either released 428 * or the state of the MTX_RECURSED bit changed. 429 */ 430 if ((v & MTX_CONTESTED) == 0 && 431 !atomic_cmpset_ptr(&m->mtx_lock, (void *)v, 432 (void *)(v | MTX_CONTESTED))) { 433 turnstile_release(&m->mtx_object); 434#ifdef __i386__ 435 ia32_pause(); 436#endif 437 continue; 438 } 439 440#if defined(SMP) && defined(ADAPTIVE_MUTEXES) 441 /* 442 * If the current owner of the lock is executing on another 443 * CPU, spin instead of blocking. 444 */ 445 owner = (struct thread *)(v & MTX_FLAGMASK); 446 if (m != &Giant && TD_IS_RUNNING(owner)) { 447 turnstile_release(&m->mtx_object); 448 while (mtx_owner(m) == owner && TD_IS_RUNNING(owner)) { 449#ifdef __i386__ 450 ia32_pause(); 451#endif 452 } 453 continue; 454 } 455#endif /* SMP && ADAPTIVE_MUTEXES */ 456 457 /* 458 * We definitely must sleep for this lock. 459 */ 460 mtx_assert(m, MA_NOTOWNED); 461 462#ifdef KTR 463 if (!cont_logged) { 464 CTR6(KTR_CONTENTION, 465 "contention: %p at %s:%d wants %s, taken by %s:%d", 466 td, file, line, m->mtx_object.lo_name, 467 WITNESS_FILE(&m->mtx_object), 468 WITNESS_LINE(&m->mtx_object)); 469 cont_logged = 1; 470 } 471#endif 472 473 /* 474 * Block on the turnstile. 475 */ 476 turnstile_wait(ts, &m->mtx_object, mtx_owner(m)); 477 } 478 479#ifdef KTR 480 if (cont_logged) { 481 CTR4(KTR_CONTENTION, 482 "contention end: %s acquired by %p at %s:%d", 483 m->mtx_object.lo_name, td, file, line); 484 } 485#endif 486 return; 487} 488 489/* 490 * _mtx_lock_spin: the tougher part of acquiring an MTX_SPIN lock. 491 * 492 * This is only called if we need to actually spin for the lock. Recursion 493 * is handled inline. 494 */ 495void 496_mtx_lock_spin(struct mtx *m, int opts, const char *file, int line) 497{ 498 int i = 0; 499 500 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 501 CTR1(KTR_LOCK, "_mtx_lock_spin: %p spinning", m); 502 503 for (;;) { 504 if (_obtain_lock(m, curthread)) 505 break; 506 507 /* Give interrupts a chance while we spin. */ 508 critical_exit(); 509 while (m->mtx_lock != MTX_UNOWNED) { 510 if (i++ < 10000000) { 511#ifdef __i386__ 512 ia32_pause(); 513#endif 514 continue; 515 } 516 if (i < 60000000) 517 DELAY(1); 518#ifdef DDB 519 else if (!db_active) { 520#else 521 else { 522#endif 523 printf("spin lock %s held by %p for > 5 seconds\n", 524 m->mtx_object.lo_name, (void *)m->mtx_lock); 525#ifdef WITNESS 526 witness_display_spinlock(&m->mtx_object, 527 mtx_owner(m)); 528#endif 529 panic("spin lock held too long"); 530 } 531#ifdef __i386__ 532 ia32_pause(); 533#endif 534 } 535 critical_enter(); 536 } 537 538 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 539 CTR1(KTR_LOCK, "_mtx_lock_spin: %p spin done", m); 540 541 return; 542} 543 544/* 545 * _mtx_unlock_sleep: the tougher part of releasing an MTX_DEF lock. 546 * 547 * We are only called here if the lock is recursed or contested (i.e. we 548 * need to wake up a blocked thread). 549 */ 550void 551_mtx_unlock_sleep(struct mtx *m, int opts, const char *file, int line) 552{ 553 struct turnstile *ts; 554 struct thread *td, *td1; 555 556 if (mtx_recursed(m)) { 557 if (--(m->mtx_recurse) == 0) 558 atomic_clear_ptr(&m->mtx_lock, MTX_RECURSED); 559 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 560 CTR1(KTR_LOCK, "_mtx_unlock_sleep: %p unrecurse", m); 561 return; 562 } 563 564 ts = turnstile_lookup(&m->mtx_object); 565 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 566 CTR1(KTR_LOCK, "_mtx_unlock_sleep: %p contested", m); 567 568#if defined(SMP) && defined(ADAPTIVE_MUTEXES) 569 if (ts == NULL) { 570 _release_lock_quick(m); 571 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 572 CTR1(KTR_LOCK, "_mtx_unlock_sleep: %p no sleepers", m); 573 turnstile_release(&m->mtx_object); 574 return; 575 } 576#else 577 MPASS(ts != NULL); 578#endif 579 /* XXX */ 580 td1 = turnstile_head(ts); 581 if (turnstile_signal(ts)) { 582 _release_lock_quick(m); 583 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 584 CTR1(KTR_LOCK, "_mtx_unlock_sleep: %p not held", m); 585 } else { 586 m->mtx_lock = MTX_CONTESTED; 587 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 588 CTR1(KTR_LOCK, "_mtx_unlock_sleep: %p still contested", 589 m); 590 } 591 turnstile_unpend(ts); 592 593 /* 594 * XXX: This is just a hack until preemption is done. However, 595 * once preemption is done we need to either wrap the 596 * turnstile_signal() and release of the actual lock in an 597 * extra critical section or change the preemption code to 598 * always just set a flag and never do instant-preempts. 599 */ 600 td = curthread; 601 if (td->td_critnest > 0 || td1->td_priority >= td->td_priority) 602 return; 603 mtx_lock_spin(&sched_lock); 604 if (!TD_IS_RUNNING(td1)) { 605#ifdef notyet 606 if (td->td_ithd != NULL) { 607 struct ithd *it = td->td_ithd; 608 609 if (it->it_interrupted) { 610 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 611 CTR2(KTR_LOCK, 612 "_mtx_unlock_sleep: %p interrupted %p", 613 it, it->it_interrupted); 614 intr_thd_fixup(it); 615 } 616 } 617#endif 618 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 619 CTR2(KTR_LOCK, 620 "_mtx_unlock_sleep: %p switching out lock=%p", m, 621 (void *)m->mtx_lock); 622 623 td->td_proc->p_stats->p_ru.ru_nivcsw++; 624 mi_switch(); 625 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 626 CTR2(KTR_LOCK, "_mtx_unlock_sleep: %p resuming lock=%p", 627 m, (void *)m->mtx_lock); 628 } 629 mtx_unlock_spin(&sched_lock); 630 631 return; 632} 633 634/* 635 * All the unlocking of MTX_SPIN locks is done inline. 636 * See the _rel_spin_lock() macro for the details. 637 */ 638 639/* 640 * The backing function for the INVARIANTS-enabled mtx_assert() 641 */ 642#ifdef INVARIANT_SUPPORT 643void 644_mtx_assert(struct mtx *m, int what, const char *file, int line) 645{ 646 647 if (panicstr != NULL) 648 return; 649 switch (what) { 650 case MA_OWNED: 651 case MA_OWNED | MA_RECURSED: 652 case MA_OWNED | MA_NOTRECURSED: 653 if (!mtx_owned(m)) 654 panic("mutex %s not owned at %s:%d", 655 m->mtx_object.lo_name, file, line); 656 if (mtx_recursed(m)) { 657 if ((what & MA_NOTRECURSED) != 0) 658 panic("mutex %s recursed at %s:%d", 659 m->mtx_object.lo_name, file, line); 660 } else if ((what & MA_RECURSED) != 0) { 661 panic("mutex %s unrecursed at %s:%d", 662 m->mtx_object.lo_name, file, line); 663 } 664 break; 665 case MA_NOTOWNED: 666 if (mtx_owned(m)) 667 panic("mutex %s owned at %s:%d", 668 m->mtx_object.lo_name, file, line); 669 break; 670 default: 671 panic("unknown mtx_assert at %s:%d", file, line); 672 } 673} 674#endif 675 676/* 677 * The MUTEX_DEBUG-enabled mtx_validate() 678 * 679 * Most of these checks have been moved off into the LO_INITIALIZED flag 680 * maintained by the witness code. 681 */ 682#ifdef MUTEX_DEBUG 683 684void mtx_validate(struct mtx *); 685 686void 687mtx_validate(struct mtx *m) 688{ 689 690/* 691 * XXX: When kernacc() does not require Giant we can reenable this check 692 */ 693#ifdef notyet 694/* 695 * XXX - When kernacc() is fixed on the alpha to handle K0_SEG memory properly 696 * we can re-enable the kernacc() checks. 697 */ 698#ifndef __alpha__ 699 /* 700 * Can't call kernacc() from early init386(), especially when 701 * initializing Giant mutex, because some stuff in kernacc() 702 * requires Giant itself. 703 */ 704 if (!cold) 705 if (!kernacc((caddr_t)m, sizeof(m), 706 VM_PROT_READ | VM_PROT_WRITE)) 707 panic("Can't read and write to mutex %p", m); 708#endif 709#endif 710} 711#endif 712 713/* 714 * General init routine used by the MTX_SYSINIT() macro. 715 */ 716void 717mtx_sysinit(void *arg) 718{ 719 struct mtx_args *margs = arg; 720 721 mtx_init(margs->ma_mtx, margs->ma_desc, NULL, margs->ma_opts); 722} 723 724/* 725 * Mutex initialization routine; initialize lock `m' of type contained in 726 * `opts' with options contained in `opts' and name `name.' The optional 727 * lock type `type' is used as a general lock category name for use with 728 * witness. 729 */ 730void 731mtx_init(struct mtx *m, const char *name, const char *type, int opts) 732{ 733 struct lock_object *lock; 734 735 MPASS((opts & ~(MTX_SPIN | MTX_QUIET | MTX_RECURSE | 736 MTX_NOWITNESS | MTX_DUPOK)) == 0); 737 738#ifdef MUTEX_DEBUG 739 /* Diagnostic and error correction */ 740 mtx_validate(m); 741#endif 742 743 lock = &m->mtx_object; 744 KASSERT((lock->lo_flags & LO_INITIALIZED) == 0, 745 ("mutex \"%s\" %p already initialized", name, m)); 746 bzero(m, sizeof(*m)); 747 if (opts & MTX_SPIN) 748 lock->lo_class = &lock_class_mtx_spin; 749 else 750 lock->lo_class = &lock_class_mtx_sleep; 751 lock->lo_name = name; 752 lock->lo_type = type != NULL ? type : name; 753 if (opts & MTX_QUIET) 754 lock->lo_flags = LO_QUIET; 755 if (opts & MTX_RECURSE) 756 lock->lo_flags |= LO_RECURSABLE; 757 if ((opts & MTX_NOWITNESS) == 0) 758 lock->lo_flags |= LO_WITNESS; 759 if (opts & MTX_DUPOK) 760 lock->lo_flags |= LO_DUPOK; 761 762 m->mtx_lock = MTX_UNOWNED; 763 764 LOCK_LOG_INIT(lock, opts); 765 766 WITNESS_INIT(lock); 767} 768 769/* 770 * Remove lock `m' from all_mtx queue. We don't allow MTX_QUIET to be 771 * passed in as a flag here because if the corresponding mtx_init() was 772 * called with MTX_QUIET set, then it will already be set in the mutex's 773 * flags. 774 */ 775void 776mtx_destroy(struct mtx *m) 777{ 778 779 LOCK_LOG_DESTROY(&m->mtx_object, 0); 780 781 if (!mtx_owned(m)) 782 MPASS(mtx_unowned(m)); 783 else { 784 MPASS((m->mtx_lock & (MTX_RECURSED|MTX_CONTESTED)) == 0); 785 786 /* Tell witness this isn't locked to make it happy. */ 787 WITNESS_UNLOCK(&m->mtx_object, LOP_EXCLUSIVE, __FILE__, 788 __LINE__); 789 } 790 791 WITNESS_DESTROY(&m->mtx_object); 792} 793 794/* 795 * Intialize the mutex code and system mutexes. This is called from the MD 796 * startup code prior to mi_startup(). The per-CPU data space needs to be 797 * setup before this is called. 798 */ 799void 800mutex_init(void) 801{ 802 803 /* Setup thread0 so that mutexes work. */ 804 LIST_INIT(&thread0.td_contested); 805 806 /* Setup turnstiles so that sleep mutexes work. */ 807 init_turnstiles(); 808 809 /* 810 * Initialize mutexes. 811 */ 812 mtx_init(&Giant, "Giant", NULL, MTX_DEF | MTX_RECURSE); 813 mtx_init(&sched_lock, "sched lock", NULL, MTX_SPIN | MTX_RECURSE); 814 mtx_init(&proc0.p_mtx, "process lock", NULL, MTX_DEF | MTX_DUPOK); 815 mtx_lock(&Giant); 816} 817