kern_mutex.c revision 125160
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 125160 2004-01-28 20:39:57Z 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 uintmax_t cnt_contest_holding; 109 uintmax_t cnt_contest_locking; 110 struct mutex_prof *next; 111}; 112 113/* 114 * mprof_buf is a static pool of profiling records to avoid possible 115 * reentrance of the memory allocation functions. 116 * 117 * Note: NUM_MPROF_BUFFERS must be smaller than MPROF_HASH_SIZE. 118 */ 119#define NUM_MPROF_BUFFERS 1000 120static struct mutex_prof mprof_buf[NUM_MPROF_BUFFERS]; 121static int first_free_mprof_buf; 122#define MPROF_HASH_SIZE 1009 123static struct mutex_prof *mprof_hash[MPROF_HASH_SIZE]; 124/* SWAG: sbuf size = avg stat. line size * number of locks */ 125#define MPROF_SBUF_SIZE 256 * 400 126 127static int mutex_prof_acquisitions; 128SYSCTL_INT(_debug_mutex_prof, OID_AUTO, acquisitions, CTLFLAG_RD, 129 &mutex_prof_acquisitions, 0, "Number of mutex acquistions recorded"); 130static int mutex_prof_records; 131SYSCTL_INT(_debug_mutex_prof, OID_AUTO, records, CTLFLAG_RD, 132 &mutex_prof_records, 0, "Number of profiling records"); 133static int mutex_prof_maxrecords = NUM_MPROF_BUFFERS; 134SYSCTL_INT(_debug_mutex_prof, OID_AUTO, maxrecords, CTLFLAG_RD, 135 &mutex_prof_maxrecords, 0, "Maximum number of profiling records"); 136static int mutex_prof_rejected; 137SYSCTL_INT(_debug_mutex_prof, OID_AUTO, rejected, CTLFLAG_RD, 138 &mutex_prof_rejected, 0, "Number of rejected profiling records"); 139static int mutex_prof_hashsize = MPROF_HASH_SIZE; 140SYSCTL_INT(_debug_mutex_prof, OID_AUTO, hashsize, CTLFLAG_RD, 141 &mutex_prof_hashsize, 0, "Hash size"); 142static int mutex_prof_collisions = 0; 143SYSCTL_INT(_debug_mutex_prof, OID_AUTO, collisions, CTLFLAG_RD, 144 &mutex_prof_collisions, 0, "Number of hash collisions"); 145 146/* 147 * mprof_mtx protects the profiling buffers and the hash. 148 */ 149static struct mtx mprof_mtx; 150MTX_SYSINIT(mprof, &mprof_mtx, "mutex profiling lock", MTX_SPIN | MTX_QUIET); 151 152static u_int64_t 153nanoseconds(void) 154{ 155 struct timespec tv; 156 157 nanotime(&tv); 158 return (tv.tv_sec * (u_int64_t)1000000000 + tv.tv_nsec); 159} 160 161static int 162dump_mutex_prof_stats(SYSCTL_HANDLER_ARGS) 163{ 164 struct sbuf *sb; 165 int error, i; 166 static int multiplier = 1; 167 168 if (first_free_mprof_buf == 0) 169 return (SYSCTL_OUT(req, "No locking recorded", 170 sizeof("No locking recorded"))); 171 172retry_sbufops: 173 sb = sbuf_new(NULL, NULL, MPROF_SBUF_SIZE * multiplier, SBUF_FIXEDLEN); 174 sbuf_printf(sb, "%6s %12s %11s %5s %12s %12s %s\n", 175 "max", "total", "count", "avg", "cnt_hold", "cnt_lock", "name"); 176 /* 177 * XXX this spinlock seems to be by far the largest perpetrator 178 * of spinlock latency (1.6 msec on an Athlon1600 was recorded 179 * even before I pessimized it further by moving the average 180 * computation here). 181 */ 182 mtx_lock_spin(&mprof_mtx); 183 for (i = 0; i < first_free_mprof_buf; ++i) { 184 sbuf_printf(sb, "%6ju %12ju %11ju %5ju %12ju %12ju %s:%d (%s)\n", 185 mprof_buf[i].cnt_max / 1000, 186 mprof_buf[i].cnt_tot / 1000, 187 mprof_buf[i].cnt_cur, 188 mprof_buf[i].cnt_cur == 0 ? (uintmax_t)0 : 189 mprof_buf[i].cnt_tot / (mprof_buf[i].cnt_cur * 1000), 190 mprof_buf[i].cnt_contest_holding, 191 mprof_buf[i].cnt_contest_locking, 192 mprof_buf[i].file, mprof_buf[i].line, mprof_buf[i].name); 193 if (sbuf_overflowed(sb)) { 194 mtx_unlock_spin(&mprof_mtx); 195 sbuf_delete(sb); 196 multiplier++; 197 goto retry_sbufops; 198 } 199 } 200 mtx_unlock_spin(&mprof_mtx); 201 sbuf_finish(sb); 202 error = SYSCTL_OUT(req, sbuf_data(sb), sbuf_len(sb) + 1); 203 sbuf_delete(sb); 204 return (error); 205} 206SYSCTL_PROC(_debug_mutex_prof, OID_AUTO, stats, CTLTYPE_STRING | CTLFLAG_RD, 207 NULL, 0, dump_mutex_prof_stats, "A", "Mutex profiling statistics"); 208#endif 209 210/* 211 * Function versions of the inlined __mtx_* macros. These are used by 212 * modules and can also be called from assembly language if needed. 213 */ 214void 215_mtx_lock_flags(struct mtx *m, int opts, const char *file, int line) 216{ 217 218 MPASS(curthread != NULL); 219 KASSERT(m->mtx_object.lo_class == &lock_class_mtx_sleep, 220 ("mtx_lock() of spin mutex %s @ %s:%d", m->mtx_object.lo_name, 221 file, line)); 222 WITNESS_CHECKORDER(&m->mtx_object, opts | LOP_NEWORDER | LOP_EXCLUSIVE, 223 file, line); 224 _get_sleep_lock(m, curthread, opts, file, line); 225 LOCK_LOG_LOCK("LOCK", &m->mtx_object, opts, m->mtx_recurse, file, 226 line); 227 WITNESS_LOCK(&m->mtx_object, opts | LOP_EXCLUSIVE, file, line); 228#ifdef MUTEX_PROFILING 229 /* don't reset the timer when/if recursing */ 230 if (m->mtx_acqtime == 0) { 231 m->mtx_filename = file; 232 m->mtx_lineno = line; 233 m->mtx_acqtime = mutex_prof_enable ? nanoseconds() : 0; 234 ++mutex_prof_acquisitions; 235 } 236#endif 237} 238 239void 240_mtx_unlock_flags(struct mtx *m, int opts, const char *file, int line) 241{ 242 243 MPASS(curthread != NULL); 244 KASSERT(m->mtx_object.lo_class == &lock_class_mtx_sleep, 245 ("mtx_unlock() of spin mutex %s @ %s:%d", m->mtx_object.lo_name, 246 file, line)); 247 WITNESS_UNLOCK(&m->mtx_object, opts | LOP_EXCLUSIVE, file, line); 248 LOCK_LOG_LOCK("UNLOCK", &m->mtx_object, opts, m->mtx_recurse, file, 249 line); 250 mtx_assert(m, MA_OWNED); 251#ifdef MUTEX_PROFILING 252 if (m->mtx_acqtime != 0) { 253 static const char *unknown = "(unknown)"; 254 struct mutex_prof *mpp; 255 u_int64_t acqtime, now; 256 const char *p, *q; 257 volatile u_int hash; 258 259 now = nanoseconds(); 260 acqtime = m->mtx_acqtime; 261 m->mtx_acqtime = 0; 262 if (now <= acqtime) 263 goto out; 264 for (p = m->mtx_filename; 265 p != NULL && strncmp(p, "../", 3) == 0; p += 3) 266 /* nothing */ ; 267 if (p == NULL || *p == '\0') 268 p = unknown; 269 for (hash = m->mtx_lineno, q = p; *q != '\0'; ++q) 270 hash = (hash * 2 + *q) % MPROF_HASH_SIZE; 271 mtx_lock_spin(&mprof_mtx); 272 for (mpp = mprof_hash[hash]; mpp != NULL; mpp = mpp->next) 273 if (mpp->line == m->mtx_lineno && 274 strcmp(mpp->file, p) == 0) 275 break; 276 if (mpp == NULL) { 277 /* Just exit if we cannot get a trace buffer */ 278 if (first_free_mprof_buf >= NUM_MPROF_BUFFERS) { 279 ++mutex_prof_rejected; 280 goto unlock; 281 } 282 mpp = &mprof_buf[first_free_mprof_buf++]; 283 mpp->name = mtx_name(m); 284 mpp->file = p; 285 mpp->line = m->mtx_lineno; 286 mpp->next = mprof_hash[hash]; 287 if (mprof_hash[hash] != NULL) 288 ++mutex_prof_collisions; 289 mprof_hash[hash] = mpp; 290 ++mutex_prof_records; 291 } 292 /* 293 * Record if the mutex has been held longer now than ever 294 * before. 295 */ 296 if (now - acqtime > mpp->cnt_max) 297 mpp->cnt_max = now - acqtime; 298 mpp->cnt_tot += now - acqtime; 299 mpp->cnt_cur++; 300 /* 301 * There's a small race, really we should cmpxchg 302 * 0 with the current value, but that would bill 303 * the contention to the wrong lock instance if 304 * it followed this also. 305 */ 306 mpp->cnt_contest_holding += m->mtx_contest_holding; 307 m->mtx_contest_holding = 0; 308 mpp->cnt_contest_locking += m->mtx_contest_locking; 309 m->mtx_contest_locking = 0; 310unlock: 311 mtx_unlock_spin(&mprof_mtx); 312 } 313out: 314#endif 315 _rel_sleep_lock(m, curthread, opts, file, line); 316} 317 318void 319_mtx_lock_spin_flags(struct mtx *m, int opts, const char *file, int line) 320{ 321 322 MPASS(curthread != NULL); 323 KASSERT(m->mtx_object.lo_class == &lock_class_mtx_spin, 324 ("mtx_lock_spin() of sleep mutex %s @ %s:%d", 325 m->mtx_object.lo_name, file, line)); 326 WITNESS_CHECKORDER(&m->mtx_object, opts | LOP_NEWORDER | LOP_EXCLUSIVE, 327 file, line); 328#if defined(SMP) || LOCK_DEBUG > 0 || 1 329 _get_spin_lock(m, curthread, opts, file, line); 330#else 331 critical_enter(); 332#endif 333 LOCK_LOG_LOCK("LOCK", &m->mtx_object, opts, m->mtx_recurse, file, 334 line); 335 WITNESS_LOCK(&m->mtx_object, opts | LOP_EXCLUSIVE, file, line); 336} 337 338void 339_mtx_unlock_spin_flags(struct mtx *m, int opts, const char *file, int line) 340{ 341 342 MPASS(curthread != NULL); 343 KASSERT(m->mtx_object.lo_class == &lock_class_mtx_spin, 344 ("mtx_unlock_spin() of sleep mutex %s @ %s:%d", 345 m->mtx_object.lo_name, file, line)); 346 WITNESS_UNLOCK(&m->mtx_object, opts | LOP_EXCLUSIVE, file, line); 347 LOCK_LOG_LOCK("UNLOCK", &m->mtx_object, opts, m->mtx_recurse, file, 348 line); 349 mtx_assert(m, MA_OWNED); 350#if defined(SMP) || LOCK_DEBUG > 0 || 1 351 _rel_spin_lock(m); 352#else 353 critical_exit(); 354#endif 355} 356 357/* 358 * The important part of mtx_trylock{,_flags}() 359 * Tries to acquire lock `m.' If this function is called on a mutex that 360 * is already owned, it will recursively acquire the lock. 361 */ 362int 363_mtx_trylock(struct mtx *m, int opts, const char *file, int line) 364{ 365 int rval; 366 367 MPASS(curthread != NULL); 368 369 if (mtx_owned(m) && (m->mtx_object.lo_flags & LO_RECURSABLE) != 0) { 370 m->mtx_recurse++; 371 atomic_set_ptr(&m->mtx_lock, MTX_RECURSED); 372 rval = 1; 373 } else 374 rval = _obtain_lock(m, curthread); 375 376 LOCK_LOG_TRY("LOCK", &m->mtx_object, opts, rval, file, line); 377 if (rval) 378 WITNESS_LOCK(&m->mtx_object, opts | LOP_EXCLUSIVE | LOP_TRYLOCK, 379 file, line); 380 381 return (rval); 382} 383 384/* 385 * _mtx_lock_sleep: the tougher part of acquiring an MTX_DEF lock. 386 * 387 * We call this if the lock is either contested (i.e. we need to go to 388 * sleep waiting for it), or if we need to recurse on it. 389 */ 390void 391_mtx_lock_sleep(struct mtx *m, int opts, const char *file, int line) 392{ 393 struct turnstile *ts; 394 struct thread *td = curthread; 395#if defined(SMP) && defined(ADAPTIVE_MUTEXES) 396 struct thread *owner; 397#endif 398 uintptr_t v; 399#ifdef KTR 400 int cont_logged = 0; 401#endif 402#ifdef MUTEX_PROFILING 403 int contested; 404#endif 405 406 if (mtx_owned(m)) { 407 KASSERT((m->mtx_object.lo_flags & LO_RECURSABLE) != 0, 408 ("_mtx_lock_sleep: recursed on non-recursive mutex %s @ %s:%d\n", 409 m->mtx_object.lo_name, file, line)); 410 m->mtx_recurse++; 411 atomic_set_ptr(&m->mtx_lock, MTX_RECURSED); 412 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 413 CTR1(KTR_LOCK, "_mtx_lock_sleep: %p recursing", m); 414 return; 415 } 416 417 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 418 CTR4(KTR_LOCK, 419 "_mtx_lock_sleep: %s contested (lock=%p) at %s:%d", 420 m->mtx_object.lo_name, (void *)m->mtx_lock, file, line); 421 422#ifdef MUTEX_PROFILING 423 contested = 0; 424#endif 425 while (!_obtain_lock(m, td)) { 426#ifdef MUTEX_PROFILING 427 contested = 1; 428 atomic_add_int(&m->mtx_contest_holding, 1); 429#endif 430 ts = turnstile_lookup(&m->mtx_object); 431 v = m->mtx_lock; 432 433 /* 434 * Check if the lock has been released while spinning for 435 * the turnstile chain lock. 436 */ 437 if (v == MTX_UNOWNED) { 438 turnstile_release(&m->mtx_object); 439#ifdef __i386__ 440 ia32_pause(); 441#endif 442 continue; 443 } 444 445 /* 446 * The mutex was marked contested on release. This means that 447 * there are other threads blocked on it. Grab ownership of 448 * it and propagate its priority to the current thread if 449 * necessary. 450 */ 451 if (v == MTX_CONTESTED) { 452 MPASS(ts != NULL); 453 m->mtx_lock = (uintptr_t)td | MTX_CONTESTED; 454 turnstile_claim(ts); 455 break; 456 } 457 458 /* 459 * If the mutex isn't already contested and a failure occurs 460 * setting the contested bit, the mutex was either released 461 * or the state of the MTX_RECURSED bit changed. 462 */ 463 if ((v & MTX_CONTESTED) == 0 && 464 !atomic_cmpset_ptr(&m->mtx_lock, (void *)v, 465 (void *)(v | MTX_CONTESTED))) { 466 turnstile_release(&m->mtx_object); 467#ifdef __i386__ 468 ia32_pause(); 469#endif 470 continue; 471 } 472 473#if defined(SMP) && defined(ADAPTIVE_MUTEXES) 474 /* 475 * If the current owner of the lock is executing on another 476 * CPU, spin instead of blocking. 477 */ 478 owner = (struct thread *)(v & MTX_FLAGMASK); 479 if (m != &Giant && TD_IS_RUNNING(owner)) { 480 turnstile_release(&m->mtx_object); 481 while (mtx_owner(m) == owner && TD_IS_RUNNING(owner)) { 482#ifdef __i386__ 483 ia32_pause(); 484#endif 485 } 486 continue; 487 } 488#endif /* SMP && ADAPTIVE_MUTEXES */ 489 490 /* 491 * We definitely must sleep for this lock. 492 */ 493 mtx_assert(m, MA_NOTOWNED); 494 495#ifdef KTR 496 if (!cont_logged) { 497 CTR6(KTR_CONTENTION, 498 "contention: %p at %s:%d wants %s, taken by %s:%d", 499 td, file, line, m->mtx_object.lo_name, 500 WITNESS_FILE(&m->mtx_object), 501 WITNESS_LINE(&m->mtx_object)); 502 cont_logged = 1; 503 } 504#endif 505 506 /* 507 * Block on the turnstile. 508 */ 509 turnstile_wait(ts, &m->mtx_object, mtx_owner(m)); 510 } 511 512#ifdef KTR 513 if (cont_logged) { 514 CTR4(KTR_CONTENTION, 515 "contention end: %s acquired by %p at %s:%d", 516 m->mtx_object.lo_name, td, file, line); 517 } 518#endif 519#ifdef MUTEX_PROFILING 520 if (contested) 521 m->mtx_contest_locking++; 522 m->mtx_contest_holding = 0; 523#endif 524 return; 525} 526 527/* 528 * _mtx_lock_spin: the tougher part of acquiring an MTX_SPIN lock. 529 * 530 * This is only called if we need to actually spin for the lock. Recursion 531 * is handled inline. 532 */ 533void 534_mtx_lock_spin(struct mtx *m, int opts, const char *file, int line) 535{ 536 int i = 0; 537 538 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 539 CTR1(KTR_LOCK, "_mtx_lock_spin: %p spinning", m); 540 541 for (;;) { 542 if (_obtain_lock(m, curthread)) 543 break; 544 545 /* Give interrupts a chance while we spin. */ 546 critical_exit(); 547 while (m->mtx_lock != MTX_UNOWNED) { 548 if (i++ < 10000000) { 549#ifdef __i386__ 550 ia32_pause(); 551#endif 552 continue; 553 } 554 if (i < 60000000) 555 DELAY(1); 556#ifdef DDB 557 else if (!db_active) { 558#else 559 else { 560#endif 561 printf("spin lock %s held by %p for > 5 seconds\n", 562 m->mtx_object.lo_name, (void *)m->mtx_lock); 563#ifdef WITNESS 564 witness_display_spinlock(&m->mtx_object, 565 mtx_owner(m)); 566#endif 567 panic("spin lock held too long"); 568 } 569#ifdef __i386__ 570 ia32_pause(); 571#endif 572 } 573 critical_enter(); 574 } 575 576 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 577 CTR1(KTR_LOCK, "_mtx_lock_spin: %p spin done", m); 578 579 return; 580} 581 582/* 583 * _mtx_unlock_sleep: the tougher part of releasing an MTX_DEF lock. 584 * 585 * We are only called here if the lock is recursed or contested (i.e. we 586 * need to wake up a blocked thread). 587 */ 588void 589_mtx_unlock_sleep(struct mtx *m, int opts, const char *file, int line) 590{ 591 struct turnstile *ts; 592 struct thread *td, *td1; 593 594 if (mtx_recursed(m)) { 595 if (--(m->mtx_recurse) == 0) 596 atomic_clear_ptr(&m->mtx_lock, MTX_RECURSED); 597 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 598 CTR1(KTR_LOCK, "_mtx_unlock_sleep: %p unrecurse", m); 599 return; 600 } 601 602 ts = turnstile_lookup(&m->mtx_object); 603 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 604 CTR1(KTR_LOCK, "_mtx_unlock_sleep: %p contested", m); 605 606#if defined(SMP) && defined(ADAPTIVE_MUTEXES) 607 if (ts == NULL) { 608 _release_lock_quick(m); 609 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 610 CTR1(KTR_LOCK, "_mtx_unlock_sleep: %p no sleepers", m); 611 turnstile_release(&m->mtx_object); 612 return; 613 } 614#else 615 MPASS(ts != NULL); 616#endif 617 /* XXX */ 618 td1 = turnstile_head(ts); 619 if (turnstile_signal(ts)) { 620 _release_lock_quick(m); 621 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 622 CTR1(KTR_LOCK, "_mtx_unlock_sleep: %p not held", m); 623 } else { 624 m->mtx_lock = MTX_CONTESTED; 625 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 626 CTR1(KTR_LOCK, "_mtx_unlock_sleep: %p still contested", 627 m); 628 } 629 turnstile_unpend(ts); 630 631 /* 632 * XXX: This is just a hack until preemption is done. However, 633 * once preemption is done we need to either wrap the 634 * turnstile_signal() and release of the actual lock in an 635 * extra critical section or change the preemption code to 636 * always just set a flag and never do instant-preempts. 637 */ 638 td = curthread; 639 if (td->td_critnest > 0 || td1->td_priority >= td->td_priority) 640 return; 641 mtx_lock_spin(&sched_lock); 642 if (!TD_IS_RUNNING(td1)) { 643#ifdef notyet 644 if (td->td_ithd != NULL) { 645 struct ithd *it = td->td_ithd; 646 647 if (it->it_interrupted) { 648 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 649 CTR2(KTR_LOCK, 650 "_mtx_unlock_sleep: %p interrupted %p", 651 it, it->it_interrupted); 652 intr_thd_fixup(it); 653 } 654 } 655#endif 656 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 657 CTR2(KTR_LOCK, 658 "_mtx_unlock_sleep: %p switching out lock=%p", m, 659 (void *)m->mtx_lock); 660 661 mi_switch(SW_INVOL); 662 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 663 CTR2(KTR_LOCK, "_mtx_unlock_sleep: %p resuming lock=%p", 664 m, (void *)m->mtx_lock); 665 } 666 mtx_unlock_spin(&sched_lock); 667 668 return; 669} 670 671/* 672 * All the unlocking of MTX_SPIN locks is done inline. 673 * See the _rel_spin_lock() macro for the details. 674 */ 675 676/* 677 * The backing function for the INVARIANTS-enabled mtx_assert() 678 */ 679#ifdef INVARIANT_SUPPORT 680void 681_mtx_assert(struct mtx *m, int what, const char *file, int line) 682{ 683 684 if (panicstr != NULL) 685 return; 686 switch (what) { 687 case MA_OWNED: 688 case MA_OWNED | MA_RECURSED: 689 case MA_OWNED | MA_NOTRECURSED: 690 if (!mtx_owned(m)) 691 panic("mutex %s not owned at %s:%d", 692 m->mtx_object.lo_name, file, line); 693 if (mtx_recursed(m)) { 694 if ((what & MA_NOTRECURSED) != 0) 695 panic("mutex %s recursed at %s:%d", 696 m->mtx_object.lo_name, file, line); 697 } else if ((what & MA_RECURSED) != 0) { 698 panic("mutex %s unrecursed at %s:%d", 699 m->mtx_object.lo_name, file, line); 700 } 701 break; 702 case MA_NOTOWNED: 703 if (mtx_owned(m)) 704 panic("mutex %s owned at %s:%d", 705 m->mtx_object.lo_name, file, line); 706 break; 707 default: 708 panic("unknown mtx_assert at %s:%d", file, line); 709 } 710} 711#endif 712 713/* 714 * The MUTEX_DEBUG-enabled mtx_validate() 715 * 716 * Most of these checks have been moved off into the LO_INITIALIZED flag 717 * maintained by the witness code. 718 */ 719#ifdef MUTEX_DEBUG 720 721void mtx_validate(struct mtx *); 722 723void 724mtx_validate(struct mtx *m) 725{ 726 727/* 728 * XXX: When kernacc() does not require Giant we can reenable this check 729 */ 730#ifdef notyet 731/* 732 * XXX - When kernacc() is fixed on the alpha to handle K0_SEG memory properly 733 * we can re-enable the kernacc() checks. 734 */ 735#ifndef __alpha__ 736 /* 737 * Can't call kernacc() from early init386(), especially when 738 * initializing Giant mutex, because some stuff in kernacc() 739 * requires Giant itself. 740 */ 741 if (!cold) 742 if (!kernacc((caddr_t)m, sizeof(m), 743 VM_PROT_READ | VM_PROT_WRITE)) 744 panic("Can't read and write to mutex %p", m); 745#endif 746#endif 747} 748#endif 749 750/* 751 * General init routine used by the MTX_SYSINIT() macro. 752 */ 753void 754mtx_sysinit(void *arg) 755{ 756 struct mtx_args *margs = arg; 757 758 mtx_init(margs->ma_mtx, margs->ma_desc, NULL, margs->ma_opts); 759} 760 761/* 762 * Mutex initialization routine; initialize lock `m' of type contained in 763 * `opts' with options contained in `opts' and name `name.' The optional 764 * lock type `type' is used as a general lock category name for use with 765 * witness. 766 */ 767void 768mtx_init(struct mtx *m, const char *name, const char *type, int opts) 769{ 770 struct lock_object *lock; 771 772 MPASS((opts & ~(MTX_SPIN | MTX_QUIET | MTX_RECURSE | 773 MTX_NOWITNESS | MTX_DUPOK)) == 0); 774 775#ifdef MUTEX_DEBUG 776 /* Diagnostic and error correction */ 777 mtx_validate(m); 778#endif 779 780 lock = &m->mtx_object; 781 KASSERT((lock->lo_flags & LO_INITIALIZED) == 0, 782 ("mutex \"%s\" %p already initialized", name, m)); 783 bzero(m, sizeof(*m)); 784 if (opts & MTX_SPIN) 785 lock->lo_class = &lock_class_mtx_spin; 786 else 787 lock->lo_class = &lock_class_mtx_sleep; 788 lock->lo_name = name; 789 lock->lo_type = type != NULL ? type : name; 790 if (opts & MTX_QUIET) 791 lock->lo_flags = LO_QUIET; 792 if (opts & MTX_RECURSE) 793 lock->lo_flags |= LO_RECURSABLE; 794 if ((opts & MTX_NOWITNESS) == 0) 795 lock->lo_flags |= LO_WITNESS; 796 if (opts & MTX_DUPOK) 797 lock->lo_flags |= LO_DUPOK; 798 799 m->mtx_lock = MTX_UNOWNED; 800 801 LOCK_LOG_INIT(lock, opts); 802 803 WITNESS_INIT(lock); 804} 805 806/* 807 * Remove lock `m' from all_mtx queue. We don't allow MTX_QUIET to be 808 * passed in as a flag here because if the corresponding mtx_init() was 809 * called with MTX_QUIET set, then it will already be set in the mutex's 810 * flags. 811 */ 812void 813mtx_destroy(struct mtx *m) 814{ 815 816 LOCK_LOG_DESTROY(&m->mtx_object, 0); 817 818 if (!mtx_owned(m)) 819 MPASS(mtx_unowned(m)); 820 else { 821 MPASS((m->mtx_lock & (MTX_RECURSED|MTX_CONTESTED)) == 0); 822 823 /* Tell witness this isn't locked to make it happy. */ 824 WITNESS_UNLOCK(&m->mtx_object, LOP_EXCLUSIVE, __FILE__, 825 __LINE__); 826 } 827 828 WITNESS_DESTROY(&m->mtx_object); 829} 830 831/* 832 * Intialize the mutex code and system mutexes. This is called from the MD 833 * startup code prior to mi_startup(). The per-CPU data space needs to be 834 * setup before this is called. 835 */ 836void 837mutex_init(void) 838{ 839 840 /* Setup thread0 so that mutexes work. */ 841 LIST_INIT(&thread0.td_contested); 842 843 /* Setup turnstiles so that sleep mutexes work. */ 844 init_turnstiles(); 845 846 /* 847 * Initialize mutexes. 848 */ 849 mtx_init(&Giant, "Giant", NULL, MTX_DEF | MTX_RECURSE); 850 mtx_init(&sched_lock, "sched lock", NULL, MTX_SPIN | MTX_RECURSE); 851 mtx_init(&proc0.p_mtx, "process lock", NULL, MTX_DEF | MTX_DUPOK); 852 mtx_lock(&Giant); 853} 854