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