kern_mutex.c revision 124944
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 124944 2004-01-25 03:54:52Z jeff $"); 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 _get_sleep_lock(m, curthread, opts, file, line); 223 LOCK_LOG_LOCK("LOCK", &m->mtx_object, opts, m->mtx_recurse, file, 224 line); 225 WITNESS_LOCK(&m->mtx_object, opts | LOP_EXCLUSIVE, file, line); 226#ifdef MUTEX_PROFILING 227 /* don't reset the timer when/if recursing */ 228 if (m->mtx_acqtime == 0) { 229 m->mtx_filename = file; 230 m->mtx_lineno = line; 231 m->mtx_acqtime = mutex_prof_enable ? nanoseconds() : 0; 232 ++mutex_prof_acquisitions; 233 } 234#endif 235} 236 237void 238_mtx_unlock_flags(struct mtx *m, int opts, const char *file, int line) 239{ 240 241 MPASS(curthread != NULL); 242 KASSERT(m->mtx_object.lo_class == &lock_class_mtx_sleep, 243 ("mtx_unlock() of spin mutex %s @ %s:%d", m->mtx_object.lo_name, 244 file, line)); 245 WITNESS_UNLOCK(&m->mtx_object, opts | LOP_EXCLUSIVE, file, line); 246 LOCK_LOG_LOCK("UNLOCK", &m->mtx_object, opts, m->mtx_recurse, file, 247 line); 248 mtx_assert(m, MA_OWNED); 249#ifdef MUTEX_PROFILING 250 if (m->mtx_acqtime != 0) { 251 static const char *unknown = "(unknown)"; 252 struct mutex_prof *mpp; 253 u_int64_t acqtime, now; 254 const char *p, *q; 255 volatile u_int hash; 256 257 now = nanoseconds(); 258 acqtime = m->mtx_acqtime; 259 m->mtx_acqtime = 0; 260 if (now <= acqtime) 261 goto out; 262 for (p = m->mtx_filename; 263 p != NULL && strncmp(p, "../", 3) == 0; p += 3) 264 /* nothing */ ; 265 if (p == NULL || *p == '\0') 266 p = unknown; 267 for (hash = m->mtx_lineno, q = p; *q != '\0'; ++q) 268 hash = (hash * 2 + *q) % MPROF_HASH_SIZE; 269 mtx_lock_spin(&mprof_mtx); 270 for (mpp = mprof_hash[hash]; mpp != NULL; mpp = mpp->next) 271 if (mpp->line == m->mtx_lineno && 272 strcmp(mpp->file, p) == 0) 273 break; 274 if (mpp == NULL) { 275 /* Just exit if we cannot get a trace buffer */ 276 if (first_free_mprof_buf >= NUM_MPROF_BUFFERS) { 277 ++mutex_prof_rejected; 278 goto unlock; 279 } 280 mpp = &mprof_buf[first_free_mprof_buf++]; 281 mpp->name = mtx_name(m); 282 mpp->file = p; 283 mpp->line = m->mtx_lineno; 284 mpp->next = mprof_hash[hash]; 285 if (mprof_hash[hash] != NULL) 286 ++mutex_prof_collisions; 287 mprof_hash[hash] = mpp; 288 ++mutex_prof_records; 289 } 290 /* 291 * Record if the mutex has been held longer now than ever 292 * before. 293 */ 294 if (now - acqtime > mpp->cnt_max) 295 mpp->cnt_max = now - acqtime; 296 mpp->cnt_tot += now - acqtime; 297 mpp->cnt_cur++; 298 /* 299 * There's a small race, really we should cmpxchg 300 * 0 with the current value, but that would bill 301 * the contention to the wrong lock instance if 302 * it followed this also. 303 */ 304 mpp->cnt_contest_holding += m->mtx_contest_holding; 305 m->mtx_contest_holding = 0; 306 mpp->cnt_contest_locking += m->mtx_contest_locking; 307 m->mtx_contest_locking = 0; 308unlock: 309 mtx_unlock_spin(&mprof_mtx); 310 } 311out: 312#endif 313 _rel_sleep_lock(m, curthread, opts, file, line); 314} 315 316void 317_mtx_lock_spin_flags(struct mtx *m, int opts, const char *file, int line) 318{ 319 320 MPASS(curthread != NULL); 321 KASSERT(m->mtx_object.lo_class == &lock_class_mtx_spin, 322 ("mtx_lock_spin() of sleep mutex %s @ %s:%d", 323 m->mtx_object.lo_name, file, line)); 324#if defined(SMP) || LOCK_DEBUG > 0 || 1 325 _get_spin_lock(m, curthread, opts, file, line); 326#else 327 critical_enter(); 328#endif 329 LOCK_LOG_LOCK("LOCK", &m->mtx_object, opts, m->mtx_recurse, file, 330 line); 331 WITNESS_LOCK(&m->mtx_object, opts | LOP_EXCLUSIVE, file, line); 332} 333 334void 335_mtx_unlock_spin_flags(struct mtx *m, int opts, const char *file, int line) 336{ 337 338 MPASS(curthread != NULL); 339 KASSERT(m->mtx_object.lo_class == &lock_class_mtx_spin, 340 ("mtx_unlock_spin() of sleep mutex %s @ %s:%d", 341 m->mtx_object.lo_name, file, line)); 342 WITNESS_UNLOCK(&m->mtx_object, opts | LOP_EXCLUSIVE, file, line); 343 LOCK_LOG_LOCK("UNLOCK", &m->mtx_object, opts, m->mtx_recurse, file, 344 line); 345 mtx_assert(m, MA_OWNED); 346#if defined(SMP) || LOCK_DEBUG > 0 || 1 347 _rel_spin_lock(m); 348#else 349 critical_exit(); 350#endif 351} 352 353/* 354 * The important part of mtx_trylock{,_flags}() 355 * Tries to acquire lock `m.' If this function is called on a mutex that 356 * is already owned, it will recursively acquire the lock. 357 */ 358int 359_mtx_trylock(struct mtx *m, int opts, const char *file, int line) 360{ 361 int rval; 362 363 MPASS(curthread != NULL); 364 365 if (mtx_owned(m) && (m->mtx_object.lo_flags & LO_RECURSABLE) != 0) { 366 m->mtx_recurse++; 367 atomic_set_ptr(&m->mtx_lock, MTX_RECURSED); 368 rval = 1; 369 } else 370 rval = _obtain_lock(m, curthread); 371 372 LOCK_LOG_TRY("LOCK", &m->mtx_object, opts, rval, file, line); 373 if (rval) 374 WITNESS_LOCK(&m->mtx_object, opts | LOP_EXCLUSIVE | LOP_TRYLOCK, 375 file, line); 376 377 return (rval); 378} 379 380/* 381 * _mtx_lock_sleep: the tougher part of acquiring an MTX_DEF lock. 382 * 383 * We call this if the lock is either contested (i.e. we need to go to 384 * sleep waiting for it), or if we need to recurse on it. 385 */ 386void 387_mtx_lock_sleep(struct mtx *m, int opts, const char *file, int line) 388{ 389 struct turnstile *ts; 390 struct thread *td = curthread; 391#if defined(SMP) && defined(ADAPTIVE_MUTEXES) 392 struct thread *owner; 393#endif 394 uintptr_t v; 395#ifdef KTR 396 int cont_logged = 0; 397#endif 398#ifdef MUTEX_PROFILING 399 int contested; 400#endif 401 402 if (mtx_owned(m)) { 403 KASSERT((m->mtx_object.lo_flags & LO_RECURSABLE) != 0, 404 ("_mtx_lock_sleep: recursed on non-recursive mutex %s @ %s:%d\n", 405 m->mtx_object.lo_name, file, line)); 406 m->mtx_recurse++; 407 atomic_set_ptr(&m->mtx_lock, MTX_RECURSED); 408 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 409 CTR1(KTR_LOCK, "_mtx_lock_sleep: %p recursing", m); 410 return; 411 } 412 413 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 414 CTR4(KTR_LOCK, 415 "_mtx_lock_sleep: %s contested (lock=%p) at %s:%d", 416 m->mtx_object.lo_name, (void *)m->mtx_lock, file, line); 417 418#ifdef MUTEX_PROFILING 419 contested = 0; 420#endif 421 while (!_obtain_lock(m, td)) { 422#ifdef MUTEX_PROFILING 423 contested = 1; 424 atomic_add_int(&m->mtx_contest_holding, 1); 425#endif 426 ts = turnstile_lookup(&m->mtx_object); 427 v = m->mtx_lock; 428 429 /* 430 * Check if the lock has been released while spinning for 431 * the turnstile chain lock. 432 */ 433 if (v == MTX_UNOWNED) { 434 turnstile_release(&m->mtx_object); 435#ifdef __i386__ 436 ia32_pause(); 437#endif 438 continue; 439 } 440 441 /* 442 * The mutex was marked contested on release. This means that 443 * there are other threads blocked on it. Grab ownership of 444 * it and propagate its priority to the current thread if 445 * necessary. 446 */ 447 if (v == MTX_CONTESTED) { 448 MPASS(ts != NULL); 449 m->mtx_lock = (uintptr_t)td | MTX_CONTESTED; 450 turnstile_claim(ts); 451 break; 452 } 453 454 /* 455 * If the mutex isn't already contested and a failure occurs 456 * setting the contested bit, the mutex was either released 457 * or the state of the MTX_RECURSED bit changed. 458 */ 459 if ((v & MTX_CONTESTED) == 0 && 460 !atomic_cmpset_ptr(&m->mtx_lock, (void *)v, 461 (void *)(v | MTX_CONTESTED))) { 462 turnstile_release(&m->mtx_object); 463#ifdef __i386__ 464 ia32_pause(); 465#endif 466 continue; 467 } 468 469#if defined(SMP) && defined(ADAPTIVE_MUTEXES) 470 /* 471 * If the current owner of the lock is executing on another 472 * CPU, spin instead of blocking. 473 */ 474 owner = (struct thread *)(v & MTX_FLAGMASK); 475 if (m != &Giant && TD_IS_RUNNING(owner)) { 476 turnstile_release(&m->mtx_object); 477 while (mtx_owner(m) == owner && TD_IS_RUNNING(owner)) { 478#ifdef __i386__ 479 ia32_pause(); 480#endif 481 } 482 continue; 483 } 484#endif /* SMP && ADAPTIVE_MUTEXES */ 485 486 /* 487 * We definitely must sleep for this lock. 488 */ 489 mtx_assert(m, MA_NOTOWNED); 490 491#ifdef KTR 492 if (!cont_logged) { 493 CTR6(KTR_CONTENTION, 494 "contention: %p at %s:%d wants %s, taken by %s:%d", 495 td, file, line, m->mtx_object.lo_name, 496 WITNESS_FILE(&m->mtx_object), 497 WITNESS_LINE(&m->mtx_object)); 498 cont_logged = 1; 499 } 500#endif 501 502 /* 503 * Block on the turnstile. 504 */ 505 turnstile_wait(ts, &m->mtx_object, mtx_owner(m)); 506 } 507 508#ifdef KTR 509 if (cont_logged) { 510 CTR4(KTR_CONTENTION, 511 "contention end: %s acquired by %p at %s:%d", 512 m->mtx_object.lo_name, td, file, line); 513 } 514#endif 515#ifdef MUTEX_PROFILING 516 if (contested) 517 m->mtx_contest_locking++; 518 m->mtx_contest_holding = 0; 519#endif 520 return; 521} 522 523/* 524 * _mtx_lock_spin: the tougher part of acquiring an MTX_SPIN lock. 525 * 526 * This is only called if we need to actually spin for the lock. Recursion 527 * is handled inline. 528 */ 529void 530_mtx_lock_spin(struct mtx *m, int opts, const char *file, int line) 531{ 532 int i = 0; 533 534 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 535 CTR1(KTR_LOCK, "_mtx_lock_spin: %p spinning", m); 536 537 for (;;) { 538 if (_obtain_lock(m, curthread)) 539 break; 540 541 /* Give interrupts a chance while we spin. */ 542 critical_exit(); 543 while (m->mtx_lock != MTX_UNOWNED) { 544 if (i++ < 10000000) { 545#ifdef __i386__ 546 ia32_pause(); 547#endif 548 continue; 549 } 550 if (i < 60000000) 551 DELAY(1); 552#ifdef DDB 553 else if (!db_active) { 554#else 555 else { 556#endif 557 printf("spin lock %s held by %p for > 5 seconds\n", 558 m->mtx_object.lo_name, (void *)m->mtx_lock); 559#ifdef WITNESS 560 witness_display_spinlock(&m->mtx_object, 561 mtx_owner(m)); 562#endif 563 panic("spin lock held too long"); 564 } 565#ifdef __i386__ 566 ia32_pause(); 567#endif 568 } 569 critical_enter(); 570 } 571 572 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 573 CTR1(KTR_LOCK, "_mtx_lock_spin: %p spin done", m); 574 575 return; 576} 577 578/* 579 * _mtx_unlock_sleep: the tougher part of releasing an MTX_DEF lock. 580 * 581 * We are only called here if the lock is recursed or contested (i.e. we 582 * need to wake up a blocked thread). 583 */ 584void 585_mtx_unlock_sleep(struct mtx *m, int opts, const char *file, int line) 586{ 587 struct turnstile *ts; 588 struct thread *td, *td1; 589 590 if (mtx_recursed(m)) { 591 if (--(m->mtx_recurse) == 0) 592 atomic_clear_ptr(&m->mtx_lock, MTX_RECURSED); 593 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 594 CTR1(KTR_LOCK, "_mtx_unlock_sleep: %p unrecurse", m); 595 return; 596 } 597 598 ts = turnstile_lookup(&m->mtx_object); 599 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 600 CTR1(KTR_LOCK, "_mtx_unlock_sleep: %p contested", m); 601 602#if defined(SMP) && defined(ADAPTIVE_MUTEXES) 603 if (ts == NULL) { 604 _release_lock_quick(m); 605 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 606 CTR1(KTR_LOCK, "_mtx_unlock_sleep: %p no sleepers", m); 607 turnstile_release(&m->mtx_object); 608 return; 609 } 610#else 611 MPASS(ts != NULL); 612#endif 613 /* XXX */ 614 td1 = turnstile_head(ts); 615 if (turnstile_signal(ts)) { 616 _release_lock_quick(m); 617 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 618 CTR1(KTR_LOCK, "_mtx_unlock_sleep: %p not held", m); 619 } else { 620 m->mtx_lock = MTX_CONTESTED; 621 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 622 CTR1(KTR_LOCK, "_mtx_unlock_sleep: %p still contested", 623 m); 624 } 625 turnstile_unpend(ts); 626 627 /* 628 * XXX: This is just a hack until preemption is done. However, 629 * once preemption is done we need to either wrap the 630 * turnstile_signal() and release of the actual lock in an 631 * extra critical section or change the preemption code to 632 * always just set a flag and never do instant-preempts. 633 */ 634 td = curthread; 635 if (td->td_critnest > 0 || td1->td_priority >= td->td_priority) 636 return; 637 mtx_lock_spin(&sched_lock); 638 if (!TD_IS_RUNNING(td1)) { 639#ifdef notyet 640 if (td->td_ithd != NULL) { 641 struct ithd *it = td->td_ithd; 642 643 if (it->it_interrupted) { 644 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 645 CTR2(KTR_LOCK, 646 "_mtx_unlock_sleep: %p interrupted %p", 647 it, it->it_interrupted); 648 intr_thd_fixup(it); 649 } 650 } 651#endif 652 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 653 CTR2(KTR_LOCK, 654 "_mtx_unlock_sleep: %p switching out lock=%p", m, 655 (void *)m->mtx_lock); 656 657 mi_switch(SW_INVOL); 658 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 659 CTR2(KTR_LOCK, "_mtx_unlock_sleep: %p resuming lock=%p", 660 m, (void *)m->mtx_lock); 661 } 662 mtx_unlock_spin(&sched_lock); 663 664 return; 665} 666 667/* 668 * All the unlocking of MTX_SPIN locks is done inline. 669 * See the _rel_spin_lock() macro for the details. 670 */ 671 672/* 673 * The backing function for the INVARIANTS-enabled mtx_assert() 674 */ 675#ifdef INVARIANT_SUPPORT 676void 677_mtx_assert(struct mtx *m, int what, const char *file, int line) 678{ 679 680 if (panicstr != NULL) 681 return; 682 switch (what) { 683 case MA_OWNED: 684 case MA_OWNED | MA_RECURSED: 685 case MA_OWNED | MA_NOTRECURSED: 686 if (!mtx_owned(m)) 687 panic("mutex %s not owned at %s:%d", 688 m->mtx_object.lo_name, file, line); 689 if (mtx_recursed(m)) { 690 if ((what & MA_NOTRECURSED) != 0) 691 panic("mutex %s recursed at %s:%d", 692 m->mtx_object.lo_name, file, line); 693 } else if ((what & MA_RECURSED) != 0) { 694 panic("mutex %s unrecursed at %s:%d", 695 m->mtx_object.lo_name, file, line); 696 } 697 break; 698 case MA_NOTOWNED: 699 if (mtx_owned(m)) 700 panic("mutex %s owned at %s:%d", 701 m->mtx_object.lo_name, file, line); 702 break; 703 default: 704 panic("unknown mtx_assert at %s:%d", file, line); 705 } 706} 707#endif 708 709/* 710 * The MUTEX_DEBUG-enabled mtx_validate() 711 * 712 * Most of these checks have been moved off into the LO_INITIALIZED flag 713 * maintained by the witness code. 714 */ 715#ifdef MUTEX_DEBUG 716 717void mtx_validate(struct mtx *); 718 719void 720mtx_validate(struct mtx *m) 721{ 722 723/* 724 * XXX: When kernacc() does not require Giant we can reenable this check 725 */ 726#ifdef notyet 727/* 728 * XXX - When kernacc() is fixed on the alpha to handle K0_SEG memory properly 729 * we can re-enable the kernacc() checks. 730 */ 731#ifndef __alpha__ 732 /* 733 * Can't call kernacc() from early init386(), especially when 734 * initializing Giant mutex, because some stuff in kernacc() 735 * requires Giant itself. 736 */ 737 if (!cold) 738 if (!kernacc((caddr_t)m, sizeof(m), 739 VM_PROT_READ | VM_PROT_WRITE)) 740 panic("Can't read and write to mutex %p", m); 741#endif 742#endif 743} 744#endif 745 746/* 747 * General init routine used by the MTX_SYSINIT() macro. 748 */ 749void 750mtx_sysinit(void *arg) 751{ 752 struct mtx_args *margs = arg; 753 754 mtx_init(margs->ma_mtx, margs->ma_desc, NULL, margs->ma_opts); 755} 756 757/* 758 * Mutex initialization routine; initialize lock `m' of type contained in 759 * `opts' with options contained in `opts' and name `name.' The optional 760 * lock type `type' is used as a general lock category name for use with 761 * witness. 762 */ 763void 764mtx_init(struct mtx *m, const char *name, const char *type, int opts) 765{ 766 struct lock_object *lock; 767 768 MPASS((opts & ~(MTX_SPIN | MTX_QUIET | MTX_RECURSE | 769 MTX_NOWITNESS | MTX_DUPOK)) == 0); 770 771#ifdef MUTEX_DEBUG 772 /* Diagnostic and error correction */ 773 mtx_validate(m); 774#endif 775 776 lock = &m->mtx_object; 777 KASSERT((lock->lo_flags & LO_INITIALIZED) == 0, 778 ("mutex \"%s\" %p already initialized", name, m)); 779 bzero(m, sizeof(*m)); 780 if (opts & MTX_SPIN) 781 lock->lo_class = &lock_class_mtx_spin; 782 else 783 lock->lo_class = &lock_class_mtx_sleep; 784 lock->lo_name = name; 785 lock->lo_type = type != NULL ? type : name; 786 if (opts & MTX_QUIET) 787 lock->lo_flags = LO_QUIET; 788 if (opts & MTX_RECURSE) 789 lock->lo_flags |= LO_RECURSABLE; 790 if ((opts & MTX_NOWITNESS) == 0) 791 lock->lo_flags |= LO_WITNESS; 792 if (opts & MTX_DUPOK) 793 lock->lo_flags |= LO_DUPOK; 794 795 m->mtx_lock = MTX_UNOWNED; 796 797 LOCK_LOG_INIT(lock, opts); 798 799 WITNESS_INIT(lock); 800} 801 802/* 803 * Remove lock `m' from all_mtx queue. We don't allow MTX_QUIET to be 804 * passed in as a flag here because if the corresponding mtx_init() was 805 * called with MTX_QUIET set, then it will already be set in the mutex's 806 * flags. 807 */ 808void 809mtx_destroy(struct mtx *m) 810{ 811 812 LOCK_LOG_DESTROY(&m->mtx_object, 0); 813 814 if (!mtx_owned(m)) 815 MPASS(mtx_unowned(m)); 816 else { 817 MPASS((m->mtx_lock & (MTX_RECURSED|MTX_CONTESTED)) == 0); 818 819 /* Tell witness this isn't locked to make it happy. */ 820 WITNESS_UNLOCK(&m->mtx_object, LOP_EXCLUSIVE, __FILE__, 821 __LINE__); 822 } 823 824 WITNESS_DESTROY(&m->mtx_object); 825} 826 827/* 828 * Intialize the mutex code and system mutexes. This is called from the MD 829 * startup code prior to mi_startup(). The per-CPU data space needs to be 830 * setup before this is called. 831 */ 832void 833mutex_init(void) 834{ 835 836 /* Setup thread0 so that mutexes work. */ 837 LIST_INIT(&thread0.td_contested); 838 839 /* Setup turnstiles so that sleep mutexes work. */ 840 init_turnstiles(); 841 842 /* 843 * Initialize mutexes. 844 */ 845 mtx_init(&Giant, "Giant", NULL, MTX_DEF | MTX_RECURSE); 846 mtx_init(&sched_lock, "sched lock", NULL, MTX_SPIN | MTX_RECURSE); 847 mtx_init(&proc0.p_mtx, "process lock", NULL, MTX_DEF | MTX_DUPOK); 848 mtx_lock(&Giant); 849} 850