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("$FreeBSD$"); 38 39 #include "opt_adaptive_mutexes.h" 40 #include "opt_ddb.h" 41 #include "opt_mprof.h" 42 #include "opt_mutex_wake_all.h" 43 #include "opt_sched.h" 44 45 #include <sys/param.h> 46 #include <sys/systm.h> 47 #include <sys/bus.h> 48 #include <sys/kdb.h> 49 #include <sys/kernel.h> 50 #include <sys/ktr.h> 51 #include <sys/lock.h> 52 #include <sys/malloc.h> 53 #include <sys/mutex.h> 54 #include <sys/proc.h> 55 #include <sys/resourcevar.h> 56 #include <sys/sched.h> 57 #include <sys/sbuf.h> 58 #include <sys/sysctl.h> 59 #include <sys/turnstile.h> 60 #include <sys/vmmeter.h> 61 62 #include <machine/atomic.h> 63 #include <machine/bus.h> 64 #include <machine/clock.h> 65 #include <machine/cpu.h> 66 67 #include <ddb/ddb.h> 68 69 #include <vm/vm.h> 70 #include <vm/vm_extern.h> 71 72 /* 73 * Force MUTEX_WAKE_ALL for now. 74 * single thread wakeup needs fixes to avoid race conditions with 75 * priority inheritance. 76 */ 77 #ifndef MUTEX_WAKE_ALL 78 #define MUTEX_WAKE_ALL 79 #endif 80 81 /* 82 * Internal utility macros. 83 */ 84 #define mtx_unowned(m) ((m)->mtx_lock == MTX_UNOWNED) 85 86 #define mtx_owner(m) (mtx_unowned((m)) ? NULL \ 87 : (struct thread *)((m)->mtx_lock & MTX_FLAGMASK)) 88 89 /* 90 * Lock classes for sleep and spin mutexes. 91 */ 92 struct lock_class lock_class_mtx_sleep = { 93 "sleep mutex", 94 LC_SLEEPLOCK | LC_RECURSABLE 95 }; 96 struct lock_class lock_class_mtx_spin = { 97 "spin mutex", 98 LC_SPINLOCK | LC_RECURSABLE 99 }; 100 101 /* 102 * System-wide mutexes 103 */ 104 struct mtx sched_lock; 105 struct mtx Giant; 106 107 #ifdef MUTEX_PROFILING 108 SYSCTL_NODE(_debug, OID_AUTO, mutex, CTLFLAG_RD, NULL, "mutex debugging"); 109 SYSCTL_NODE(_debug_mutex, OID_AUTO, prof, CTLFLAG_RD, NULL, "mutex profiling"); 110 static int mutex_prof_enable = 0; 111 SYSCTL_INT(_debug_mutex_prof, OID_AUTO, enable, CTLFLAG_RW, 112 &mutex_prof_enable, 0, "Enable tracing of mutex holdtime"); 113 114 struct mutex_prof { 115 const char *name; 116 const char *file; 117 int line; 118 uintmax_t cnt_max; 119 uintmax_t cnt_tot; 120 uintmax_t cnt_cur; 121 uintmax_t cnt_contest_holding; 122 uintmax_t cnt_contest_locking; 123 struct mutex_prof *next; 124 }; 125 126 /* 127 * mprof_buf is a static pool of profiling records to avoid possible 128 * reentrance of the memory allocation functions. 129 * 130 * Note: NUM_MPROF_BUFFERS must be smaller than MPROF_HASH_SIZE. 131 */ 132 #ifdef MPROF_BUFFERS 133 #define NUM_MPROF_BUFFERS MPROF_BUFFERS 134 #else 135 #define NUM_MPROF_BUFFERS 1000 136 #endif 137 static struct mutex_prof mprof_buf[NUM_MPROF_BUFFERS]; 138 static int first_free_mprof_buf; 139 #ifndef MPROF_HASH_SIZE 140 #define MPROF_HASH_SIZE 1009 141 #endif 142 #if NUM_MPROF_BUFFERS >= MPROF_HASH_SIZE 143 #error MPROF_BUFFERS must be larger than MPROF_HASH_SIZE 144 #endif 145 static struct mutex_prof *mprof_hash[MPROF_HASH_SIZE]; 146 /* SWAG: sbuf size = avg stat. line size * number of locks */ 147 #define MPROF_SBUF_SIZE 256 * 400 148 149 static int mutex_prof_acquisitions; 150 SYSCTL_INT(_debug_mutex_prof, OID_AUTO, acquisitions, CTLFLAG_RD, 151 &mutex_prof_acquisitions, 0, "Number of mutex acquistions recorded"); 152 static int mutex_prof_records; 153 SYSCTL_INT(_debug_mutex_prof, OID_AUTO, records, CTLFLAG_RD, 154 &mutex_prof_records, 0, "Number of profiling records"); 155 static int mutex_prof_maxrecords = NUM_MPROF_BUFFERS; 156 SYSCTL_INT(_debug_mutex_prof, OID_AUTO, maxrecords, CTLFLAG_RD, 157 &mutex_prof_maxrecords, 0, "Maximum number of profiling records"); 158 static int mutex_prof_rejected; 159 SYSCTL_INT(_debug_mutex_prof, OID_AUTO, rejected, CTLFLAG_RD, 160 &mutex_prof_rejected, 0, "Number of rejected profiling records"); 161 static int mutex_prof_hashsize = MPROF_HASH_SIZE; 162 SYSCTL_INT(_debug_mutex_prof, OID_AUTO, hashsize, CTLFLAG_RD, 163 &mutex_prof_hashsize, 0, "Hash size"); 164 static int mutex_prof_collisions = 0; 165 SYSCTL_INT(_debug_mutex_prof, OID_AUTO, collisions, CTLFLAG_RD, 166 &mutex_prof_collisions, 0, "Number of hash collisions"); 167 168 /* 169 * mprof_mtx protects the profiling buffers and the hash. 170 */ 171 static struct mtx mprof_mtx; 172 MTX_SYSINIT(mprof, &mprof_mtx, "mutex profiling lock", MTX_SPIN | MTX_QUIET); 173 174 static u_int64_t 175 nanoseconds(void) 176 { 177 struct timespec tv; 178 179 nanotime(&tv); 180 return (tv.tv_sec * (u_int64_t)1000000000 + tv.tv_nsec); 181 } 182 183 static int 184 dump_mutex_prof_stats(SYSCTL_HANDLER_ARGS) 185 { 186 struct sbuf *sb; 187 int error, i; 188 static int multiplier = 1; 189 190 if (first_free_mprof_buf == 0) 191 return (SYSCTL_OUT(req, "No locking recorded", 192 sizeof("No locking recorded"))); 193 194 retry_sbufops: 195 sb = sbuf_new(NULL, NULL, MPROF_SBUF_SIZE * multiplier, SBUF_FIXEDLEN); 196 sbuf_printf(sb, "%6s %12s %11s %5s %12s %12s %s\n", 197 "max", "total", "count", "avg", "cnt_hold", "cnt_lock", "name"); 198 /* 199 * XXX this spinlock seems to be by far the largest perpetrator 200 * of spinlock latency (1.6 msec on an Athlon1600 was recorded 201 * even before I pessimized it further by moving the average 202 * computation here). 203 */ 204 mtx_lock_spin(&mprof_mtx); 205 for (i = 0; i < first_free_mprof_buf; ++i) { 206 sbuf_printf(sb, "%6ju %12ju %11ju %5ju %12ju %12ju %s:%d (%s)\n", 207 mprof_buf[i].cnt_max / 1000, 208 mprof_buf[i].cnt_tot / 1000, 209 mprof_buf[i].cnt_cur, 210 mprof_buf[i].cnt_cur == 0 ? (uintmax_t)0 : 211 mprof_buf[i].cnt_tot / (mprof_buf[i].cnt_cur * 1000), 212 mprof_buf[i].cnt_contest_holding, 213 mprof_buf[i].cnt_contest_locking, 214 mprof_buf[i].file, mprof_buf[i].line, mprof_buf[i].name); 215 if (sbuf_overflowed(sb)) { 216 mtx_unlock_spin(&mprof_mtx); 217 sbuf_delete(sb); 218 multiplier++; 219 goto retry_sbufops; 220 } 221 } 222 mtx_unlock_spin(&mprof_mtx); 223 sbuf_finish(sb); 224 error = SYSCTL_OUT(req, sbuf_data(sb), sbuf_len(sb) + 1); 225 sbuf_delete(sb); 226 return (error); 227 } 228 SYSCTL_PROC(_debug_mutex_prof, OID_AUTO, stats, CTLTYPE_STRING | CTLFLAG_RD, 229 NULL, 0, dump_mutex_prof_stats, "A", "Mutex profiling statistics"); 230 231 static int 232 reset_mutex_prof_stats(SYSCTL_HANDLER_ARGS) 233 { 234 int error, v; 235 236 if (first_free_mprof_buf == 0) 237 return (0); 238 239 v = 0; 240 error = sysctl_handle_int(oidp, &v, 0, req); 241 if (error) 242 return (error); 243 if (req->newptr == NULL) 244 return (error); 245 if (v == 0) 246 return (0); 247 248 mtx_lock_spin(&mprof_mtx); 249 bzero(mprof_buf, sizeof(*mprof_buf) * first_free_mprof_buf); 250 bzero(mprof_hash, sizeof(struct mtx *) * MPROF_HASH_SIZE); 251 first_free_mprof_buf = 0; 252 mtx_unlock_spin(&mprof_mtx); 253 return (0); 254 } 255 SYSCTL_PROC(_debug_mutex_prof, OID_AUTO, reset, CTLTYPE_INT | CTLFLAG_RW, 256 NULL, 0, reset_mutex_prof_stats, "I", "Reset mutex profiling statistics"); 257 #endif 258 259 /* 260 * Function versions of the inlined __mtx_* macros. These are used by 261 * modules and can also be called from assembly language if needed. 262 */ 263 void 264 _mtx_lock_flags(struct mtx *m, int opts, const char *file, int line) 265 { 266 267 MPASS(curthread != NULL); 268 KASSERT(m->mtx_object.lo_class == &lock_class_mtx_sleep, 269 ("mtx_lock() of spin mutex %s @ %s:%d", m->mtx_object.lo_name, 270 file, line)); 271 WITNESS_CHECKORDER(&m->mtx_object, opts | LOP_NEWORDER | LOP_EXCLUSIVE, 272 file, line); 273 _get_sleep_lock(m, curthread, opts, file, line); 274 LOCK_LOG_LOCK("LOCK", &m->mtx_object, opts, m->mtx_recurse, file, 275 line); 276 WITNESS_LOCK(&m->mtx_object, opts | LOP_EXCLUSIVE, file, line); 277 #ifdef MUTEX_PROFILING 278 /* don't reset the timer when/if recursing */ 279 if (m->mtx_acqtime == 0) { 280 m->mtx_filename = file; 281 m->mtx_lineno = line; 282 m->mtx_acqtime = mutex_prof_enable ? nanoseconds() : 0; 283 ++mutex_prof_acquisitions; 284 } 285 #endif 286 } 287 288 void 289 _mtx_unlock_flags(struct mtx *m, int opts, const char *file, int line) 290 { 291 292 MPASS(curthread != NULL); 293 KASSERT(m->mtx_object.lo_class == &lock_class_mtx_sleep, 294 ("mtx_unlock() of spin mutex %s @ %s:%d", m->mtx_object.lo_name, 295 file, line)); 296 WITNESS_UNLOCK(&m->mtx_object, opts | LOP_EXCLUSIVE, file, line); 297 LOCK_LOG_LOCK("UNLOCK", &m->mtx_object, opts, m->mtx_recurse, file, 298 line); 299 mtx_assert(m, MA_OWNED); 300 #ifdef MUTEX_PROFILING 301 if (m->mtx_acqtime != 0) { 302 static const char *unknown = "(unknown)"; 303 struct mutex_prof *mpp; 304 u_int64_t acqtime, now; 305 const char *p, *q; 306 volatile u_int hash; 307 308 now = nanoseconds(); 309 acqtime = m->mtx_acqtime; 310 m->mtx_acqtime = 0; 311 if (now <= acqtime) 312 goto out; 313 for (p = m->mtx_filename; 314 p != NULL && strncmp(p, "../", 3) == 0; p += 3) 315 /* nothing */ ; 316 if (p == NULL || *p == '\0') 317 p = unknown; 318 for (hash = m->mtx_lineno, q = p; *q != '\0'; ++q) 319 hash = (hash * 2 + *q) % MPROF_HASH_SIZE; 320 mtx_lock_spin(&mprof_mtx); 321 for (mpp = mprof_hash[hash]; mpp != NULL; mpp = mpp->next) 322 if (mpp->line == m->mtx_lineno && 323 strcmp(mpp->file, p) == 0) 324 break; 325 if (mpp == NULL) { 326 /* Just exit if we cannot get a trace buffer */ 327 if (first_free_mprof_buf >= NUM_MPROF_BUFFERS) { 328 ++mutex_prof_rejected; 329 goto unlock; 330 } 331 mpp = &mprof_buf[first_free_mprof_buf++]; 332 mpp->name = mtx_name(m); 333 mpp->file = p; 334 mpp->line = m->mtx_lineno; 335 mpp->next = mprof_hash[hash]; 336 if (mprof_hash[hash] != NULL) 337 ++mutex_prof_collisions; 338 mprof_hash[hash] = mpp; 339 ++mutex_prof_records; 340 } 341 /* 342 * Record if the mutex has been held longer now than ever 343 * before. 344 */ 345 if (now - acqtime > mpp->cnt_max) 346 mpp->cnt_max = now - acqtime; 347 mpp->cnt_tot += now - acqtime; 348 mpp->cnt_cur++; 349 /* 350 * There's a small race, really we should cmpxchg 351 * 0 with the current value, but that would bill 352 * the contention to the wrong lock instance if 353 * it followed this also. 354 */ 355 mpp->cnt_contest_holding += m->mtx_contest_holding; 356 m->mtx_contest_holding = 0; 357 mpp->cnt_contest_locking += m->mtx_contest_locking; 358 m->mtx_contest_locking = 0; 359 unlock: 360 mtx_unlock_spin(&mprof_mtx); 361 } 362 out: 363 #endif 364 _rel_sleep_lock(m, curthread, opts, file, line); 365 } 366 367 void 368 _mtx_lock_spin_flags(struct mtx *m, int opts, const char *file, int line) 369 { 370 371 MPASS(curthread != NULL); 372 KASSERT(m->mtx_object.lo_class == &lock_class_mtx_spin, 373 ("mtx_lock_spin() of sleep mutex %s @ %s:%d", 374 m->mtx_object.lo_name, file, line)); 375 WITNESS_CHECKORDER(&m->mtx_object, opts | LOP_NEWORDER | LOP_EXCLUSIVE, 376 file, line); 377 #if defined(SMP) || LOCK_DEBUG > 0 || 1 378 _get_spin_lock(m, curthread, opts, file, line); 379 #else 380 critical_enter(); 381 #endif 382 LOCK_LOG_LOCK("LOCK", &m->mtx_object, opts, m->mtx_recurse, file, 383 line); 384 WITNESS_LOCK(&m->mtx_object, opts | LOP_EXCLUSIVE, file, line); 385 } 386 387 void 388 _mtx_unlock_spin_flags(struct mtx *m, int opts, const char *file, int line) 389 { 390 391 MPASS(curthread != NULL); 392 KASSERT(m->mtx_object.lo_class == &lock_class_mtx_spin, 393 ("mtx_unlock_spin() of sleep mutex %s @ %s:%d", 394 m->mtx_object.lo_name, file, line)); 395 WITNESS_UNLOCK(&m->mtx_object, opts | LOP_EXCLUSIVE, file, line); 396 LOCK_LOG_LOCK("UNLOCK", &m->mtx_object, opts, m->mtx_recurse, file, 397 line); 398 mtx_assert(m, MA_OWNED); 399 #if defined(SMP) || LOCK_DEBUG > 0 || 1 400 _rel_spin_lock(m); 401 #else 402 critical_exit(); 403 #endif 404 } 405 406 /* 407 * The important part of mtx_trylock{,_flags}() 408 * Tries to acquire lock `m.' If this function is called on a mutex that 409 * is already owned, it will recursively acquire the lock. 410 */ 411 int 412 _mtx_trylock(struct mtx *m, int opts, const char *file, int line) 413 { 414 int rval; 415 416 MPASS(curthread != NULL); 417 418 if (mtx_owned(m) && (m->mtx_object.lo_flags & LO_RECURSABLE) != 0) { 419 m->mtx_recurse++; 420 atomic_set_ptr(&m->mtx_lock, MTX_RECURSED); 421 rval = 1; 422 } else 423 rval = _obtain_lock(m, curthread); 424 425 LOCK_LOG_TRY("LOCK", &m->mtx_object, opts, rval, file, line); 426 if (rval) 427 WITNESS_LOCK(&m->mtx_object, opts | LOP_EXCLUSIVE | LOP_TRYLOCK, 428 file, line); 429 430 return (rval); 431 } 432 433 /* 434 * _mtx_lock_sleep: the tougher part of acquiring an MTX_DEF lock. 435 * 436 * We call this if the lock is either contested (i.e. we need to go to 437 * sleep waiting for it), or if we need to recurse on it. 438 */ 439 void 440 _mtx_lock_sleep(struct mtx *m, struct thread *td, int opts, const char *file, 441 int line) 442 { 443 #if defined(SMP) && !defined(NO_ADAPTIVE_MUTEXES) 444 struct thread *owner; 445 #endif 446 uintptr_t v; 447 #ifdef KTR 448 int cont_logged = 0; 449 #endif 450 #ifdef MUTEX_PROFILING 451 int contested; 452 #endif 453 454 if (mtx_owned(m)) { 455 KASSERT((m->mtx_object.lo_flags & LO_RECURSABLE) != 0, 456 ("_mtx_lock_sleep: recursed on non-recursive mutex %s @ %s:%d\n", 457 m->mtx_object.lo_name, file, line)); 458 m->mtx_recurse++; 459 atomic_set_ptr(&m->mtx_lock, MTX_RECURSED); 460 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 461 CTR1(KTR_LOCK, "_mtx_lock_sleep: %p recursing", m); 462 return; 463 } 464 465 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 466 CTR4(KTR_LOCK, 467 "_mtx_lock_sleep: %s contested (lock=%p) at %s:%d", 468 m->mtx_object.lo_name, (void *)m->mtx_lock, file, line); 469 470 #ifdef MUTEX_PROFILING 471 contested = 0; 472 #endif 473 while (!_obtain_lock(m, td)) { 474 #ifdef MUTEX_PROFILING 475 contested = 1; 476 atomic_add_int(&m->mtx_contest_holding, 1); 477 #endif 478 turnstile_lock(&m->mtx_object); 479 v = m->mtx_lock; 480 481 /* 482 * Check if the lock has been released while spinning for 483 * the turnstile chain lock. 484 */ 485 if (v == MTX_UNOWNED) { 486 turnstile_release(&m->mtx_object); 487 cpu_spinwait(); 488 continue; 489 } 490 491 #ifdef MUTEX_WAKE_ALL 492 MPASS(v != MTX_CONTESTED); 493 #else 494 /* 495 * The mutex was marked contested on release. This means that 496 * there are other threads blocked on it. Grab ownership of 497 * it and propagate its priority to the current thread if 498 * necessary. 499 */ 500 if (v == MTX_CONTESTED) { 501 m->mtx_lock = (uintptr_t)td | MTX_CONTESTED; 502 turnstile_claim(&m->mtx_object); 503 break; 504 } 505 #endif 506 507 /* 508 * If the mutex isn't already contested and a failure occurs 509 * setting the contested bit, the mutex was either released 510 * or the state of the MTX_RECURSED bit changed. 511 */ 512 if ((v & MTX_CONTESTED) == 0 && 513 !atomic_cmpset_ptr(&m->mtx_lock, (void *)v, 514 (void *)(v | MTX_CONTESTED))) { 515 turnstile_release(&m->mtx_object); 516 cpu_spinwait(); 517 continue; 518 } 519 520 #if defined(SMP) && !defined(NO_ADAPTIVE_MUTEXES) 521 /* 522 * If the current owner of the lock is executing on another 523 * CPU, spin instead of blocking. 524 */ 525 owner = (struct thread *)(v & MTX_FLAGMASK); 526 #ifdef ADAPTIVE_GIANT 527 if (TD_IS_RUNNING(owner)) { 528 #else 529 if (m != &Giant && TD_IS_RUNNING(owner)) { 530 #endif 531 turnstile_release(&m->mtx_object); 532 while (mtx_owner(m) == owner && TD_IS_RUNNING(owner)) { 533 cpu_spinwait(); 534 } 535 continue; 536 } 537 #endif /* SMP && !NO_ADAPTIVE_MUTEXES */ 538 539 /* 540 * We definitely must sleep for this lock. 541 */ 542 mtx_assert(m, MA_NOTOWNED); 543 544 #ifdef KTR 545 if (!cont_logged) { 546 CTR6(KTR_CONTENTION, 547 "contention: %p at %s:%d wants %s, taken by %s:%d", 548 td, file, line, m->mtx_object.lo_name, 549 WITNESS_FILE(&m->mtx_object), 550 WITNESS_LINE(&m->mtx_object)); 551 cont_logged = 1; 552 } 553 #endif 554 555 /* 556 * Block on the turnstile. 557 */ 558 turnstile_wait(&m->mtx_object, mtx_owner(m)); 559 } 560 561 #ifdef KTR 562 if (cont_logged) { 563 CTR4(KTR_CONTENTION, 564 "contention end: %s acquired by %p at %s:%d", 565 m->mtx_object.lo_name, td, file, line); 566 } 567 #endif 568 #ifdef MUTEX_PROFILING 569 if (contested) 570 m->mtx_contest_locking++; 571 m->mtx_contest_holding = 0; 572 #endif 573 return; 574 } 575 576 /* 577 * _mtx_lock_spin: the tougher part of acquiring an MTX_SPIN lock. 578 * 579 * This is only called if we need to actually spin for the lock. Recursion 580 * is handled inline. 581 */ 582 void 583 _mtx_lock_spin(struct mtx *m, struct thread *td, int opts, const char *file, 584 int line) 585 { 586 int i = 0; 587 588 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 589 CTR1(KTR_LOCK, "_mtx_lock_spin: %p spinning", m); 590 591 for (;;) { 592 if (_obtain_lock(m, td)) 593 break; 594 595 /* Give interrupts a chance while we spin. */ 596 critical_exit(); 597 while (m->mtx_lock != MTX_UNOWNED) { 598 if (i++ < 10000000) { 599 cpu_spinwait(); 600 continue; 601 } 602 if (i < 60000000) 603 DELAY(1); 604 else if (!kdb_active) { 605 printf("spin lock %s held by %p for > 5 seconds\n", 606 m->mtx_object.lo_name, (void *)m->mtx_lock); 607 #ifdef WITNESS 608 witness_display_spinlock(&m->mtx_object, 609 mtx_owner(m)); 610 #endif 611 panic("spin lock held too long"); 612 } 613 cpu_spinwait(); 614 } 615 critical_enter(); 616 } 617 618 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 619 CTR1(KTR_LOCK, "_mtx_lock_spin: %p spin done", m); 620 621 return; 622 } 623 624 /* 625 * _mtx_unlock_sleep: the tougher part of releasing an MTX_DEF lock. 626 * 627 * We are only called here if the lock is recursed or contested (i.e. we 628 * need to wake up a blocked thread). 629 */ 630 void 631 _mtx_unlock_sleep(struct mtx *m, int opts, const char *file, int line) 632 { 633 struct turnstile *ts; 634 #ifndef PREEMPTION 635 struct thread *td, *td1; 636 #endif 637 638 if (mtx_recursed(m)) { 639 if (--(m->mtx_recurse) == 0) 640 atomic_clear_ptr(&m->mtx_lock, MTX_RECURSED); 641 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 642 CTR1(KTR_LOCK, "_mtx_unlock_sleep: %p unrecurse", m); 643 return; 644 } 645 646 turnstile_lock(&m->mtx_object); 647 ts = turnstile_lookup(&m->mtx_object); 648 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 649 CTR1(KTR_LOCK, "_mtx_unlock_sleep: %p contested", m); 650 651 #if defined(SMP) && !defined(NO_ADAPTIVE_MUTEXES) 652 if (ts == NULL) { 653 _release_lock_quick(m); 654 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 655 CTR1(KTR_LOCK, "_mtx_unlock_sleep: %p no sleepers", m); 656 turnstile_release(&m->mtx_object); 657 return; 658 } 659 #else 660 MPASS(ts != NULL); 661 #endif 662 #ifndef PREEMPTION 663 /* XXX */ 664 td1 = turnstile_head(ts); 665 #endif 666 #ifdef MUTEX_WAKE_ALL 667 turnstile_broadcast(ts); 668 _release_lock_quick(m); 669 #else 670 if (turnstile_signal(ts)) { 671 _release_lock_quick(m); 672 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 673 CTR1(KTR_LOCK, "_mtx_unlock_sleep: %p not held", m); 674 } else { 675 m->mtx_lock = MTX_CONTESTED; 676 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 677 CTR1(KTR_LOCK, "_mtx_unlock_sleep: %p still contested", 678 m); 679 } 680 #endif 681 turnstile_unpend(ts); 682 683 #ifndef PREEMPTION 684 /* 685 * XXX: This is just a hack until preemption is done. However, 686 * once preemption is done we need to either wrap the 687 * turnstile_signal() and release of the actual lock in an 688 * extra critical section or change the preemption code to 689 * always just set a flag and never do instant-preempts. 690 */ 691 td = curthread; 692 if (td->td_critnest > 0 || td1->td_priority >= td->td_priority) 693 return; 694 mtx_lock_spin(&sched_lock); 695 if (!TD_IS_RUNNING(td1)) { 696 #ifdef notyet 697 if (td->td_ithd != NULL) { 698 struct ithd *it = td->td_ithd; 699 700 if (it->it_interrupted) { 701 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 702 CTR2(KTR_LOCK, 703 "_mtx_unlock_sleep: %p interrupted %p", 704 it, it->it_interrupted); 705 intr_thd_fixup(it); 706 } 707 } 708 #endif 709 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 710 CTR2(KTR_LOCK, 711 "_mtx_unlock_sleep: %p switching out lock=%p", m, 712 (void *)m->mtx_lock); 713 714 mi_switch(SW_INVOL, NULL); 715 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 716 CTR2(KTR_LOCK, "_mtx_unlock_sleep: %p resuming lock=%p", 717 m, (void *)m->mtx_lock); 718 } 719 mtx_unlock_spin(&sched_lock); 720 #endif 721 722 return; 723 } 724 725 /* 726 * All the unlocking of MTX_SPIN locks is done inline. 727 * See the _rel_spin_lock() macro for the details. 728 */ 729 730 /* 731 * The backing function for the INVARIANTS-enabled mtx_assert() 732 */ 733 #ifdef INVARIANT_SUPPORT 734 void 735 _mtx_assert(struct mtx *m, int what, const char *file, int line) 736 { 737 738 if (panicstr != NULL) 739 return; 740 switch (what) { 741 case MA_OWNED: 742 case MA_OWNED | MA_RECURSED: 743 case MA_OWNED | MA_NOTRECURSED: 744 if (!mtx_owned(m)) 745 panic("mutex %s not owned at %s:%d", 746 m->mtx_object.lo_name, file, line); 747 if (mtx_recursed(m)) { 748 if ((what & MA_NOTRECURSED) != 0) 749 panic("mutex %s recursed at %s:%d", 750 m->mtx_object.lo_name, file, line); 751 } else if ((what & MA_RECURSED) != 0) { 752 panic("mutex %s unrecursed at %s:%d", 753 m->mtx_object.lo_name, file, line); 754 } 755 break; 756 case MA_NOTOWNED: 757 if (mtx_owned(m)) 758 panic("mutex %s owned at %s:%d", 759 m->mtx_object.lo_name, file, line); 760 break; 761 default: 762 panic("unknown mtx_assert at %s:%d", file, line); 763 } 764 } 765 #endif 766 767 /* 768 * The MUTEX_DEBUG-enabled mtx_validate() 769 * 770 * Most of these checks have been moved off into the LO_INITIALIZED flag 771 * maintained by the witness code. 772 */ 773 #ifdef MUTEX_DEBUG 774 775 void mtx_validate(struct mtx *); 776 777 void 778 mtx_validate(struct mtx *m) 779 { 780 781 /* 782 * XXX: When kernacc() does not require Giant we can reenable this check 783 */ 784 #ifdef notyet 785 /* 786 * XXX - When kernacc() is fixed on the alpha to handle K0_SEG memory properly 787 * we can re-enable the kernacc() checks. 788 */ 789 #ifndef __alpha__ 790 /* 791 * Can't call kernacc() from early init386(), especially when 792 * initializing Giant mutex, because some stuff in kernacc() 793 * requires Giant itself. 794 */ 795 if (!cold) 796 if (!kernacc((caddr_t)m, sizeof(m), 797 VM_PROT_READ | VM_PROT_WRITE)) 798 panic("Can't read and write to mutex %p", m); 799 #endif 800 #endif 801 } 802 #endif 803 804 /* 805 * General init routine used by the MTX_SYSINIT() macro. 806 */ 807 void 808 mtx_sysinit(void *arg) 809 { 810 struct mtx_args *margs = arg; 811 812 mtx_init(margs->ma_mtx, margs->ma_desc, NULL, margs->ma_opts); 813 } 814 815 /* 816 * Mutex initialization routine; initialize lock `m' of type contained in 817 * `opts' with options contained in `opts' and name `name.' The optional 818 * lock type `type' is used as a general lock category name for use with 819 * witness. 820 */ 821 void 822 mtx_init(struct mtx *m, const char *name, const char *type, int opts) 823 { 824 struct lock_object *lock; 825 826 MPASS((opts & ~(MTX_SPIN | MTX_QUIET | MTX_RECURSE | 827 MTX_NOWITNESS | MTX_DUPOK)) == 0); 828 829 #ifdef MUTEX_DEBUG 830 /* Diagnostic and error correction */ 831 mtx_validate(m); 832 #endif 833 834 lock = &m->mtx_object; 835 KASSERT((lock->lo_flags & LO_INITIALIZED) == 0, 836 ("mutex \"%s\" %p already initialized", name, m)); 837 bzero(m, sizeof(*m)); 838 if (opts & MTX_SPIN) 839 lock->lo_class = &lock_class_mtx_spin; 840 else 841 lock->lo_class = &lock_class_mtx_sleep; 842 lock->lo_name = name; 843 lock->lo_type = type != NULL ? type : name; 844 if (opts & MTX_QUIET) 845 lock->lo_flags = LO_QUIET; 846 if (opts & MTX_RECURSE) 847 lock->lo_flags |= LO_RECURSABLE; 848 if ((opts & MTX_NOWITNESS) == 0) 849 lock->lo_flags |= LO_WITNESS; 850 if (opts & MTX_DUPOK) 851 lock->lo_flags |= LO_DUPOK; 852 853 m->mtx_lock = MTX_UNOWNED; 854 855 LOCK_LOG_INIT(lock, opts); 856 857 WITNESS_INIT(lock); 858 } 859 860 /* 861 * Remove lock `m' from all_mtx queue. We don't allow MTX_QUIET to be 862 * passed in as a flag here because if the corresponding mtx_init() was 863 * called with MTX_QUIET set, then it will already be set in the mutex's 864 * flags. 865 */ 866 void 867 mtx_destroy(struct mtx *m) 868 { 869 870 LOCK_LOG_DESTROY(&m->mtx_object, 0); 871 872 if (!mtx_owned(m)) 873 MPASS(mtx_unowned(m)); 874 else { 875 MPASS((m->mtx_lock & (MTX_RECURSED|MTX_CONTESTED)) == 0); 876 877 /* Tell witness this isn't locked to make it happy. */ 878 WITNESS_UNLOCK(&m->mtx_object, LOP_EXCLUSIVE, __FILE__, 879 __LINE__); 880 } 881 882 WITNESS_DESTROY(&m->mtx_object); 883 } 884 885 /* 886 * Intialize the mutex code and system mutexes. This is called from the MD 887 * startup code prior to mi_startup(). The per-CPU data space needs to be 888 * setup before this is called. 889 */ 890 void 891 mutex_init(void) 892 { 893 894 /* Setup thread0 so that mutexes work. */ 895 LIST_INIT(&thread0.td_contested); 896 897 /* Setup turnstiles so that sleep mutexes work. */ 898 init_turnstiles(); 899 900 /* 901 * Initialize mutexes. 902 */ 903 mtx_init(&Giant, "Giant", NULL, MTX_DEF | MTX_RECURSE); 904 mtx_init(&sched_lock, "sched lock", NULL, MTX_SPIN | MTX_RECURSE); 905 mtx_init(&proc0.p_mtx, "process lock", NULL, MTX_DEF | MTX_DUPOK); 906 mtx_lock(&Giant); 907 } 908