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/cpu.h> 66 67 #include <ddb/ddb.h> 68 69 #include <fs/devfs/devfs_int.h> 70 71 #include <vm/vm.h> 72 #include <vm/vm_extern.h> 73 74 /* 75 * Force MUTEX_WAKE_ALL for now. 76 * single thread wakeup needs fixes to avoid race conditions with 77 * priority inheritance. 78 */ 79 #ifndef MUTEX_WAKE_ALL 80 #define MUTEX_WAKE_ALL 81 #endif 82 83 /* 84 * Internal utility macros. 85 */ 86 #define mtx_unowned(m) ((m)->mtx_lock == MTX_UNOWNED) 87 88 #define mtx_owner(m) ((struct thread *)((m)->mtx_lock & ~MTX_FLAGMASK)) 89 90 #ifdef DDB 91 static void db_show_mtx(struct lock_object *lock); 92 #endif 93 94 /* 95 * Lock classes for sleep and spin mutexes. 96 */ 97 struct lock_class lock_class_mtx_sleep = { 98 "sleep mutex", 99 LC_SLEEPLOCK | LC_RECURSABLE, 100 #ifdef DDB 101 db_show_mtx 102 #endif 103 }; 104 struct lock_class lock_class_mtx_spin = { 105 "spin mutex", 106 LC_SPINLOCK | LC_RECURSABLE, 107 #ifdef DDB 108 db_show_mtx 109 #endif 110 }; 111 112 /* 113 * System-wide mutexes 114 */ 115 struct mtx sched_lock; 116 struct mtx Giant; 117 118 #ifdef MUTEX_PROFILING 119 SYSCTL_NODE(_debug, OID_AUTO, mutex, CTLFLAG_RD, NULL, "mutex debugging"); 120 SYSCTL_NODE(_debug_mutex, OID_AUTO, prof, CTLFLAG_RD, NULL, "mutex profiling"); 121 static int mutex_prof_enable = 0; 122 SYSCTL_INT(_debug_mutex_prof, OID_AUTO, enable, CTLFLAG_RW, 123 &mutex_prof_enable, 0, "Enable tracing of mutex holdtime"); 124 125 struct mutex_prof { 126 const char *name; 127 const char *file; 128 int line; 129 uintmax_t cnt_max; 130 uintmax_t cnt_tot; 131 uintmax_t cnt_cur; 132 uintmax_t cnt_contest_holding; 133 uintmax_t cnt_contest_locking; 134 struct mutex_prof *next; 135 }; 136 137 /* 138 * mprof_buf is a static pool of profiling records to avoid possible 139 * reentrance of the memory allocation functions. 140 * 141 * Note: NUM_MPROF_BUFFERS must be smaller than MPROF_HASH_SIZE. 142 */ 143 #ifdef MPROF_BUFFERS 144 #define NUM_MPROF_BUFFERS MPROF_BUFFERS 145 #else 146 #define NUM_MPROF_BUFFERS 1000 147 #endif 148 static struct mutex_prof mprof_buf[NUM_MPROF_BUFFERS]; 149 static int first_free_mprof_buf; 150 #ifndef MPROF_HASH_SIZE 151 #define MPROF_HASH_SIZE 1009 152 #endif 153 #if NUM_MPROF_BUFFERS >= MPROF_HASH_SIZE 154 #error MPROF_BUFFERS must be larger than MPROF_HASH_SIZE 155 #endif 156 static struct mutex_prof *mprof_hash[MPROF_HASH_SIZE]; 157 /* SWAG: sbuf size = avg stat. line size * number of locks */ 158 #define MPROF_SBUF_SIZE 256 * 400 159 160 static int mutex_prof_acquisitions; 161 SYSCTL_INT(_debug_mutex_prof, OID_AUTO, acquisitions, CTLFLAG_RD, 162 &mutex_prof_acquisitions, 0, "Number of mutex acquistions recorded"); 163 static int mutex_prof_records; 164 SYSCTL_INT(_debug_mutex_prof, OID_AUTO, records, CTLFLAG_RD, 165 &mutex_prof_records, 0, "Number of profiling records"); 166 static int mutex_prof_maxrecords = NUM_MPROF_BUFFERS; 167 SYSCTL_INT(_debug_mutex_prof, OID_AUTO, maxrecords, CTLFLAG_RD, 168 &mutex_prof_maxrecords, 0, "Maximum number of profiling records"); 169 static int mutex_prof_rejected; 170 SYSCTL_INT(_debug_mutex_prof, OID_AUTO, rejected, CTLFLAG_RD, 171 &mutex_prof_rejected, 0, "Number of rejected profiling records"); 172 static int mutex_prof_hashsize = MPROF_HASH_SIZE; 173 SYSCTL_INT(_debug_mutex_prof, OID_AUTO, hashsize, CTLFLAG_RD, 174 &mutex_prof_hashsize, 0, "Hash size"); 175 static int mutex_prof_collisions = 0; 176 SYSCTL_INT(_debug_mutex_prof, OID_AUTO, collisions, CTLFLAG_RD, 177 &mutex_prof_collisions, 0, "Number of hash collisions"); 178 179 /* 180 * mprof_mtx protects the profiling buffers and the hash. 181 */ 182 static struct mtx mprof_mtx; 183 MTX_SYSINIT(mprof, &mprof_mtx, "mutex profiling lock", MTX_SPIN | MTX_QUIET); 184 185 static u_int64_t 186 nanoseconds(void) 187 { 188 struct timespec tv; 189 190 nanotime(&tv); 191 return (tv.tv_sec * (u_int64_t)1000000000 + tv.tv_nsec); 192 } 193 194 static int 195 dump_mutex_prof_stats(SYSCTL_HANDLER_ARGS) 196 { 197 struct sbuf *sb; 198 int error, i; 199 static int multiplier = 1; 200 201 if (first_free_mprof_buf == 0) 202 return (SYSCTL_OUT(req, "No locking recorded", 203 sizeof("No locking recorded"))); 204 205 retry_sbufops: 206 sb = sbuf_new(NULL, NULL, MPROF_SBUF_SIZE * multiplier, SBUF_FIXEDLEN); 207 sbuf_printf(sb, "\n%6s %12s %11s %5s %12s %12s %s\n", 208 "max", "total", "count", "avg", "cnt_hold", "cnt_lock", "name"); 209 /* 210 * XXX this spinlock seems to be by far the largest perpetrator 211 * of spinlock latency (1.6 msec on an Athlon1600 was recorded 212 * even before I pessimized it further by moving the average 213 * computation here). 214 */ 215 mtx_lock_spin(&mprof_mtx); 216 for (i = 0; i < first_free_mprof_buf; ++i) { 217 sbuf_printf(sb, "%6ju %12ju %11ju %5ju %12ju %12ju %s:%d (%s)\n", 218 mprof_buf[i].cnt_max / 1000, 219 mprof_buf[i].cnt_tot / 1000, 220 mprof_buf[i].cnt_cur, 221 mprof_buf[i].cnt_cur == 0 ? (uintmax_t)0 : 222 mprof_buf[i].cnt_tot / (mprof_buf[i].cnt_cur * 1000), 223 mprof_buf[i].cnt_contest_holding, 224 mprof_buf[i].cnt_contest_locking, 225 mprof_buf[i].file, mprof_buf[i].line, mprof_buf[i].name); 226 if (sbuf_overflowed(sb)) { 227 mtx_unlock_spin(&mprof_mtx); 228 sbuf_delete(sb); 229 multiplier++; 230 goto retry_sbufops; 231 } 232 } 233 mtx_unlock_spin(&mprof_mtx); 234 sbuf_finish(sb); 235 error = SYSCTL_OUT(req, sbuf_data(sb), sbuf_len(sb) + 1); 236 sbuf_delete(sb); 237 return (error); 238 } 239 SYSCTL_PROC(_debug_mutex_prof, OID_AUTO, stats, CTLTYPE_STRING | CTLFLAG_RD, 240 NULL, 0, dump_mutex_prof_stats, "A", "Mutex profiling statistics"); 241 242 static int 243 reset_mutex_prof_stats(SYSCTL_HANDLER_ARGS) 244 { 245 int error, v; 246 247 if (first_free_mprof_buf == 0) 248 return (0); 249 250 v = 0; 251 error = sysctl_handle_int(oidp, &v, 0, req); 252 if (error) 253 return (error); 254 if (req->newptr == NULL) 255 return (error); 256 if (v == 0) 257 return (0); 258 259 mtx_lock_spin(&mprof_mtx); 260 bzero(mprof_buf, sizeof(*mprof_buf) * first_free_mprof_buf); 261 bzero(mprof_hash, sizeof(struct mtx *) * MPROF_HASH_SIZE); 262 first_free_mprof_buf = 0; 263 mtx_unlock_spin(&mprof_mtx); 264 return (0); 265 } 266 SYSCTL_PROC(_debug_mutex_prof, OID_AUTO, reset, CTLTYPE_INT | CTLFLAG_RW, 267 NULL, 0, reset_mutex_prof_stats, "I", "Reset mutex profiling statistics"); 268 #endif 269 270 /* 271 * Function versions of the inlined __mtx_* macros. These are used by 272 * modules and can also be called from assembly language if needed. 273 */ 274 void 275 _mtx_lock_flags(struct mtx *m, int opts, const char *file, int line) 276 { 277 278 MPASS(curthread != NULL); 279 KASSERT(m->mtx_lock != MTX_DESTROYED, 280 ("mtx_lock() of destroyed mutex @ %s:%d", file, line)); 281 KASSERT(LOCK_CLASS(&m->mtx_object) == &lock_class_mtx_sleep, 282 ("mtx_lock() of spin mutex %s @ %s:%d", m->mtx_object.lo_name, 283 file, line)); 284 WITNESS_CHECKORDER(&m->mtx_object, opts | LOP_NEWORDER | LOP_EXCLUSIVE, 285 file, line); 286 _get_sleep_lock(m, curthread, opts, file, line); 287 LOCK_LOG_LOCK("LOCK", &m->mtx_object, opts, m->mtx_recurse, file, 288 line); 289 WITNESS_LOCK(&m->mtx_object, opts | LOP_EXCLUSIVE, file, line); 290 curthread->td_locks++; 291 #ifdef MUTEX_PROFILING 292 /* don't reset the timer when/if recursing */ 293 if (m->mtx_acqtime == 0) { 294 m->mtx_filename = file; 295 m->mtx_lineno = line; 296 m->mtx_acqtime = mutex_prof_enable ? nanoseconds() : 0; 297 ++mutex_prof_acquisitions; 298 } 299 #endif 300 } 301 302 void 303 _mtx_unlock_flags(struct mtx *m, int opts, const char *file, int line) 304 { 305 306 MPASS(curthread != NULL); 307 KASSERT(m->mtx_lock != MTX_DESTROYED, 308 ("mtx_unlock() of destroyed mutex @ %s:%d", file, line)); 309 KASSERT(LOCK_CLASS(&m->mtx_object) == &lock_class_mtx_sleep, 310 ("mtx_unlock() of spin mutex %s @ %s:%d", m->mtx_object.lo_name, 311 file, line)); 312 curthread->td_locks--; 313 WITNESS_UNLOCK(&m->mtx_object, opts | LOP_EXCLUSIVE, file, line); 314 LOCK_LOG_LOCK("UNLOCK", &m->mtx_object, opts, m->mtx_recurse, file, 315 line); 316 mtx_assert(m, MA_OWNED); 317 #ifdef MUTEX_PROFILING 318 if (m->mtx_acqtime != 0) { 319 static const char *unknown = "(unknown)"; 320 struct mutex_prof *mpp; 321 u_int64_t acqtime, now; 322 const char *p, *q; 323 volatile u_int hash; 324 325 now = nanoseconds(); 326 acqtime = m->mtx_acqtime; 327 m->mtx_acqtime = 0; 328 if (now <= acqtime) 329 goto out; 330 for (p = m->mtx_filename; 331 p != NULL && strncmp(p, "../", 3) == 0; p += 3) 332 /* nothing */ ; 333 if (p == NULL || *p == '\0') 334 p = unknown; 335 for (hash = m->mtx_lineno, q = p; *q != '\0'; ++q) 336 hash = (hash * 2 + *q) % MPROF_HASH_SIZE; 337 mtx_lock_spin(&mprof_mtx); 338 for (mpp = mprof_hash[hash]; mpp != NULL; mpp = mpp->next) 339 if (mpp->line == m->mtx_lineno && 340 strcmp(mpp->file, p) == 0) 341 break; 342 if (mpp == NULL) { 343 /* Just exit if we cannot get a trace buffer */ 344 if (first_free_mprof_buf >= NUM_MPROF_BUFFERS) { 345 ++mutex_prof_rejected; 346 goto unlock; 347 } 348 mpp = &mprof_buf[first_free_mprof_buf++]; 349 mpp->name = mtx_name(m); 350 mpp->file = p; 351 mpp->line = m->mtx_lineno; 352 mpp->next = mprof_hash[hash]; 353 if (mprof_hash[hash] != NULL) 354 ++mutex_prof_collisions; 355 mprof_hash[hash] = mpp; 356 ++mutex_prof_records; 357 } 358 /* 359 * Record if the mutex has been held longer now than ever 360 * before. 361 */ 362 if (now - acqtime > mpp->cnt_max) 363 mpp->cnt_max = now - acqtime; 364 mpp->cnt_tot += now - acqtime; 365 mpp->cnt_cur++; 366 /* 367 * There's a small race, really we should cmpxchg 368 * 0 with the current value, but that would bill 369 * the contention to the wrong lock instance if 370 * it followed this also. 371 */ 372 mpp->cnt_contest_holding += m->mtx_contest_holding; 373 m->mtx_contest_holding = 0; 374 mpp->cnt_contest_locking += m->mtx_contest_locking; 375 m->mtx_contest_locking = 0; 376 unlock: 377 mtx_unlock_spin(&mprof_mtx); 378 } 379 out: 380 #endif 381 _rel_sleep_lock(m, curthread, opts, file, line); 382 } 383 384 void 385 _mtx_lock_spin_flags(struct mtx *m, int opts, const char *file, int line) 386 { 387 388 MPASS(curthread != NULL); 389 KASSERT(m->mtx_lock != MTX_DESTROYED, 390 ("mtx_lock_spin() of destroyed mutex @ %s:%d", file, line)); 391 KASSERT(LOCK_CLASS(&m->mtx_object) == &lock_class_mtx_spin, 392 ("mtx_lock_spin() of sleep mutex %s @ %s:%d", 393 m->mtx_object.lo_name, file, line)); 394 WITNESS_CHECKORDER(&m->mtx_object, opts | LOP_NEWORDER | LOP_EXCLUSIVE, 395 file, line); 396 _get_spin_lock(m, curthread, opts, file, line); 397 LOCK_LOG_LOCK("LOCK", &m->mtx_object, opts, m->mtx_recurse, file, 398 line); 399 WITNESS_LOCK(&m->mtx_object, opts | LOP_EXCLUSIVE, file, line); 400 } 401 402 void 403 _mtx_unlock_spin_flags(struct mtx *m, int opts, const char *file, int line) 404 { 405 406 MPASS(curthread != NULL); 407 KASSERT(m->mtx_lock != MTX_DESTROYED, 408 ("mtx_unlock_spin() of destroyed mutex @ %s:%d", file, line)); 409 KASSERT(LOCK_CLASS(&m->mtx_object) == &lock_class_mtx_spin, 410 ("mtx_unlock_spin() of sleep mutex %s @ %s:%d", 411 m->mtx_object.lo_name, file, line)); 412 WITNESS_UNLOCK(&m->mtx_object, opts | LOP_EXCLUSIVE, file, line); 413 LOCK_LOG_LOCK("UNLOCK", &m->mtx_object, opts, m->mtx_recurse, file, 414 line); 415 mtx_assert(m, MA_OWNED); 416 _rel_spin_lock(m); 417 } 418 419 /* 420 * The important part of mtx_trylock{,_flags}() 421 * Tries to acquire lock `m.' If this function is called on a mutex that 422 * is already owned, it will recursively acquire the lock. 423 */ 424 int 425 _mtx_trylock(struct mtx *m, int opts, const char *file, int line) 426 { 427 int rval; 428 429 MPASS(curthread != NULL); 430 KASSERT(m->mtx_lock != MTX_DESTROYED, 431 ("mtx_trylock() of destroyed mutex @ %s:%d", file, line)); 432 KASSERT(LOCK_CLASS(&m->mtx_object) == &lock_class_mtx_sleep, 433 ("mtx_trylock() of spin mutex %s @ %s:%d", m->mtx_object.lo_name, 434 file, line)); 435 436 if (mtx_owned(m) && (m->mtx_object.lo_flags & LO_RECURSABLE) != 0) { 437 m->mtx_recurse++; 438 atomic_set_ptr(&m->mtx_lock, MTX_RECURSED); 439 rval = 1; 440 } else 441 rval = _obtain_lock(m, (uintptr_t)curthread); 442 443 LOCK_LOG_TRY("LOCK", &m->mtx_object, opts, rval, file, line); 444 if (rval) { 445 WITNESS_LOCK(&m->mtx_object, opts | LOP_EXCLUSIVE | LOP_TRYLOCK, 446 file, line); 447 curthread->td_locks++; 448 } 449 450 return (rval); 451 } 452 453 /* 454 * _mtx_lock_sleep: the tougher part of acquiring an MTX_DEF lock. 455 * 456 * We call this if the lock is either contested (i.e. we need to go to 457 * sleep waiting for it), or if we need to recurse on it. 458 */ 459 void 460 _mtx_lock_sleep(struct mtx *m, uintptr_t tid, int opts, const char *file, 461 int line) 462 { 463 #if defined(SMP) && !defined(NO_ADAPTIVE_MUTEXES) 464 volatile struct thread *owner; 465 #endif 466 uintptr_t v; 467 #ifdef KTR 468 int cont_logged = 0; 469 #endif 470 #ifdef MUTEX_PROFILING 471 int contested; 472 #endif 473 474 if (mtx_owned(m)) { 475 KASSERT((m->mtx_object.lo_flags & LO_RECURSABLE) != 0, 476 ("_mtx_lock_sleep: recursed on non-recursive mutex %s @ %s:%d\n", 477 m->mtx_object.lo_name, file, line)); 478 m->mtx_recurse++; 479 atomic_set_ptr(&m->mtx_lock, MTX_RECURSED); 480 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 481 CTR1(KTR_LOCK, "_mtx_lock_sleep: %p recursing", m); 482 return; 483 } 484 485 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 486 CTR4(KTR_LOCK, 487 "_mtx_lock_sleep: %s contested (lock=%p) at %s:%d", 488 m->mtx_object.lo_name, (void *)m->mtx_lock, file, line); 489 490 #ifdef MUTEX_PROFILING 491 contested = 0; 492 #endif 493 while (!_obtain_lock(m, tid)) { 494 #ifdef MUTEX_PROFILING 495 contested = 1; 496 atomic_add_int(&m->mtx_contest_holding, 1); 497 #endif 498 turnstile_lock(&m->mtx_object); 499 v = m->mtx_lock; 500 501 /* 502 * Check if the lock has been released while spinning for 503 * the turnstile chain lock. 504 */ 505 if (v == MTX_UNOWNED) { 506 turnstile_release(&m->mtx_object); 507 cpu_spinwait(); 508 continue; 509 } 510 511 #ifdef MUTEX_WAKE_ALL 512 MPASS(v != MTX_CONTESTED); 513 #else 514 /* 515 * The mutex was marked contested on release. This means that 516 * there are other threads blocked on it. Grab ownership of 517 * it and propagate its priority to the current thread if 518 * necessary. 519 */ 520 if (v == MTX_CONTESTED) { 521 m->mtx_lock = tid | MTX_CONTESTED; 522 turnstile_claim(&m->mtx_object); 523 break; 524 } 525 #endif 526 527 /* 528 * If the mutex isn't already contested and a failure occurs 529 * setting the contested bit, the mutex was either released 530 * or the state of the MTX_RECURSED bit changed. 531 */ 532 if ((v & MTX_CONTESTED) == 0 && 533 !atomic_cmpset_ptr(&m->mtx_lock, v, v | MTX_CONTESTED)) { 534 turnstile_release(&m->mtx_object); 535 cpu_spinwait(); 536 continue; 537 } 538 539 #if defined(SMP) && !defined(NO_ADAPTIVE_MUTEXES) 540 /* 541 * If the current owner of the lock is executing on another 542 * CPU, spin instead of blocking. 543 */ 544 owner = (struct thread *)(v & ~MTX_FLAGMASK); 545 #ifdef ADAPTIVE_GIANT 546 if (TD_IS_RUNNING(owner)) { 547 #else 548 if (m != &Giant && TD_IS_RUNNING(owner)) { 549 #endif 550 turnstile_release(&m->mtx_object); 551 while (mtx_owner(m) == owner && TD_IS_RUNNING(owner)) { 552 cpu_spinwait(); 553 } 554 continue; 555 } 556 #endif /* SMP && !NO_ADAPTIVE_MUTEXES */ 557 558 /* 559 * We definitely must sleep for this lock. 560 */ 561 mtx_assert(m, MA_NOTOWNED); 562 563 #ifdef KTR 564 if (!cont_logged) { 565 CTR6(KTR_CONTENTION, 566 "contention: %p at %s:%d wants %s, taken by %s:%d", 567 (void *)tid, file, line, m->mtx_object.lo_name, 568 WITNESS_FILE(&m->mtx_object), 569 WITNESS_LINE(&m->mtx_object)); 570 cont_logged = 1; 571 } 572 #endif 573 574 /* 575 * Block on the turnstile. 576 */ 577 turnstile_wait(&m->mtx_object, mtx_owner(m), 578 TS_EXCLUSIVE_QUEUE); 579 } 580 581 #ifdef KTR 582 if (cont_logged) { 583 CTR4(KTR_CONTENTION, 584 "contention end: %s acquired by %p at %s:%d", 585 m->mtx_object.lo_name, (void *)tid, file, line); 586 } 587 #endif 588 #ifdef MUTEX_PROFILING 589 if (contested) 590 m->mtx_contest_locking++; 591 m->mtx_contest_holding = 0; 592 #endif 593 return; 594 } 595 596 #ifdef SMP 597 /* 598 * _mtx_lock_spin: the tougher part of acquiring an MTX_SPIN lock. 599 * 600 * This is only called if we need to actually spin for the lock. Recursion 601 * is handled inline. 602 */ 603 void 604 _mtx_lock_spin(struct mtx *m, uintptr_t tid, int opts, const char *file, 605 int line) 606 { 607 struct thread *td; 608 int i = 0; 609 610 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 611 CTR1(KTR_LOCK, "_mtx_lock_spin: %p spinning", m); 612 613 while (!_obtain_lock(m, tid)) { 614 615 /* Give interrupts a chance while we spin. */ 616 spinlock_exit(); 617 while (m->mtx_lock != MTX_UNOWNED) { 618 if (i++ < 10000000) { 619 cpu_spinwait(); 620 continue; 621 } 622 if (i < 60000000 || kdb_active || panicstr != NULL) 623 DELAY(1); 624 else { 625 td = mtx_owner(m); 626 627 /* If the mutex is unlocked, try again. */ 628 if (td == NULL) 629 continue; 630 printf( 631 "spin lock %p (%s) held by %p (tid %d) too long\n", 632 m, m->mtx_object.lo_name, td, td->td_tid); 633 #ifdef WITNESS 634 witness_display_spinlock(&m->mtx_object, td); 635 #endif 636 panic("spin lock held too long"); 637 } 638 cpu_spinwait(); 639 } 640 spinlock_enter(); 641 } 642 643 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 644 CTR1(KTR_LOCK, "_mtx_lock_spin: %p spin done", m); 645 646 return; 647 } 648 #endif /* SMP */ 649 650 /* 651 * _mtx_unlock_sleep: the tougher part of releasing an MTX_DEF lock. 652 * 653 * We are only called here if the lock is recursed or contested (i.e. we 654 * need to wake up a blocked thread). 655 */ 656 void 657 _mtx_unlock_sleep(struct mtx *m, int opts, const char *file, int line) 658 { 659 struct turnstile *ts; 660 #ifndef PREEMPTION 661 struct thread *td, *td1; 662 #endif 663 664 if (mtx_recursed(m)) { 665 if (--(m->mtx_recurse) == 0) 666 atomic_clear_ptr(&m->mtx_lock, MTX_RECURSED); 667 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 668 CTR1(KTR_LOCK, "_mtx_unlock_sleep: %p unrecurse", m); 669 return; 670 } 671 672 turnstile_lock(&m->mtx_object); 673 ts = turnstile_lookup(&m->mtx_object); 674 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 675 CTR1(KTR_LOCK, "_mtx_unlock_sleep: %p contested", m); 676 677 #if defined(SMP) && !defined(NO_ADAPTIVE_MUTEXES) 678 if (ts == NULL) { 679 _release_lock_quick(m); 680 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 681 CTR1(KTR_LOCK, "_mtx_unlock_sleep: %p no sleepers", m); 682 turnstile_release(&m->mtx_object); 683 return; 684 } 685 #else 686 MPASS(ts != NULL); 687 #endif 688 #ifndef PREEMPTION 689 /* XXX */ 690 td1 = turnstile_head(ts, TS_EXCLUSIVE_QUEUE); 691 #endif 692 #ifdef MUTEX_WAKE_ALL 693 turnstile_broadcast(ts, TS_EXCLUSIVE_QUEUE); 694 _release_lock_quick(m); 695 #else 696 if (turnstile_signal(ts, TS_EXCLUSIVE_QUEUE)) { 697 _release_lock_quick(m); 698 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 699 CTR1(KTR_LOCK, "_mtx_unlock_sleep: %p not held", m); 700 } else { 701 m->mtx_lock = MTX_CONTESTED; 702 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 703 CTR1(KTR_LOCK, "_mtx_unlock_sleep: %p still contested", 704 m); 705 } 706 #endif 707 turnstile_unpend(ts, TS_EXCLUSIVE_LOCK); 708 709 #ifndef PREEMPTION 710 /* 711 * XXX: This is just a hack until preemption is done. However, 712 * once preemption is done we need to either wrap the 713 * turnstile_signal() and release of the actual lock in an 714 * extra critical section or change the preemption code to 715 * always just set a flag and never do instant-preempts. 716 */ 717 td = curthread; 718 if (td->td_critnest > 0 || td1->td_priority >= td->td_priority) 719 return; 720 mtx_lock_spin(&sched_lock); 721 if (!TD_IS_RUNNING(td1)) { 722 #ifdef notyet 723 if (td->td_ithd != NULL) { 724 struct ithd *it = td->td_ithd; 725 726 if (it->it_interrupted) { 727 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 728 CTR2(KTR_LOCK, 729 "_mtx_unlock_sleep: %p interrupted %p", 730 it, it->it_interrupted); 731 intr_thd_fixup(it); 732 } 733 } 734 #endif 735 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 736 CTR2(KTR_LOCK, 737 "_mtx_unlock_sleep: %p switching out lock=%p", m, 738 (void *)m->mtx_lock); 739 740 mi_switch(SW_INVOL, NULL); 741 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 742 CTR2(KTR_LOCK, "_mtx_unlock_sleep: %p resuming lock=%p", 743 m, (void *)m->mtx_lock); 744 } 745 mtx_unlock_spin(&sched_lock); 746 #endif 747 748 return; 749 } 750 751 /* 752 * All the unlocking of MTX_SPIN locks is done inline. 753 * See the _rel_spin_lock() macro for the details. 754 */ 755 756 /* 757 * The backing function for the INVARIANTS-enabled mtx_assert() 758 */ 759 #ifdef INVARIANT_SUPPORT 760 void 761 _mtx_assert(struct mtx *m, int what, const char *file, int line) 762 { 763 764 if (panicstr != NULL || dumping) 765 return; 766 switch (what) { 767 case MA_OWNED: 768 case MA_OWNED | MA_RECURSED: 769 case MA_OWNED | MA_NOTRECURSED: 770 if (!mtx_owned(m)) 771 panic("mutex %s not owned at %s:%d", 772 m->mtx_object.lo_name, file, line); 773 if (mtx_recursed(m)) { 774 if ((what & MA_NOTRECURSED) != 0) 775 panic("mutex %s recursed at %s:%d", 776 m->mtx_object.lo_name, file, line); 777 } else if ((what & MA_RECURSED) != 0) { 778 panic("mutex %s unrecursed at %s:%d", 779 m->mtx_object.lo_name, file, line); 780 } 781 break; 782 case MA_NOTOWNED: 783 if (mtx_owned(m)) 784 panic("mutex %s owned at %s:%d", 785 m->mtx_object.lo_name, file, line); 786 break; 787 default: 788 panic("unknown mtx_assert at %s:%d", file, line); 789 } 790 } 791 #endif 792 793 /* 794 * The MUTEX_DEBUG-enabled mtx_validate() 795 * 796 * Most of these checks have been moved off into the LO_INITIALIZED flag 797 * maintained by the witness code. 798 */ 799 #ifdef MUTEX_DEBUG 800 801 void mtx_validate(struct mtx *); 802 803 void 804 mtx_validate(struct mtx *m) 805 { 806 807 /* 808 * XXX: When kernacc() does not require Giant we can reenable this check 809 */ 810 #ifdef notyet 811 /* 812 * Can't call kernacc() from early init386(), especially when 813 * initializing Giant mutex, because some stuff in kernacc() 814 * requires Giant itself. 815 */ 816 if (!cold) 817 if (!kernacc((caddr_t)m, sizeof(m), 818 VM_PROT_READ | VM_PROT_WRITE)) 819 panic("Can't read and write to mutex %p", m); 820 #endif 821 } 822 #endif 823 824 /* 825 * General init routine used by the MTX_SYSINIT() macro. 826 */ 827 void 828 mtx_sysinit(void *arg) 829 { 830 struct mtx_args *margs = arg; 831 832 mtx_init(margs->ma_mtx, margs->ma_desc, NULL, margs->ma_opts); 833 } 834 835 /* 836 * Mutex initialization routine; initialize lock `m' of type contained in 837 * `opts' with options contained in `opts' and name `name.' The optional 838 * lock type `type' is used as a general lock category name for use with 839 * witness. 840 */ 841 void 842 mtx_init(struct mtx *m, const char *name, const char *type, int opts) 843 { 844 struct lock_class *class; 845 int flags; 846 847 MPASS((opts & ~(MTX_SPIN | MTX_QUIET | MTX_RECURSE | 848 MTX_NOWITNESS | MTX_DUPOK)) == 0); 849 850 #ifdef MUTEX_DEBUG 851 /* Diagnostic and error correction */ 852 mtx_validate(m); 853 #endif 854 855 /* Determine lock class and lock flags. */ 856 if (opts & MTX_SPIN) 857 class = &lock_class_mtx_spin; 858 else 859 class = &lock_class_mtx_sleep; 860 flags = 0; 861 if (opts & MTX_QUIET) 862 flags |= LO_QUIET; 863 if (opts & MTX_RECURSE) 864 flags |= LO_RECURSABLE; 865 if ((opts & MTX_NOWITNESS) == 0) 866 flags |= LO_WITNESS; 867 if (opts & MTX_DUPOK) 868 flags |= LO_DUPOK; 869 870 /* Initialize mutex. */ 871 m->mtx_lock = MTX_UNOWNED; 872 m->mtx_recurse = 0; 873 #ifdef MUTEX_PROFILING 874 m->mtx_acqtime = 0; 875 m->mtx_filename = NULL; 876 m->mtx_lineno = 0; 877 m->mtx_contest_holding = 0; 878 m->mtx_contest_locking = 0; 879 #endif 880 881 lock_init(&m->mtx_object, class, name, type, flags); 882 } 883 884 /* 885 * Remove lock `m' from all_mtx queue. We don't allow MTX_QUIET to be 886 * passed in as a flag here because if the corresponding mtx_init() was 887 * called with MTX_QUIET set, then it will already be set in the mutex's 888 * flags. 889 */ 890 void 891 mtx_destroy(struct mtx *m) 892 { 893 894 if (!mtx_owned(m)) 895 MPASS(mtx_unowned(m)); 896 else { 897 MPASS((m->mtx_lock & (MTX_RECURSED|MTX_CONTESTED)) == 0); 898 899 /* Perform the non-mtx related part of mtx_unlock_spin(). */ 900 if (LOCK_CLASS(&m->mtx_object) == &lock_class_mtx_spin) 901 spinlock_exit(); 902 else 903 curthread->td_locks--; 904 905 /* Tell witness this isn't locked to make it happy. */ 906 WITNESS_UNLOCK(&m->mtx_object, LOP_EXCLUSIVE, __FILE__, 907 __LINE__); 908 } 909 910 m->mtx_lock = MTX_DESTROYED; 911 lock_destroy(&m->mtx_object); 912 } 913 914 /* 915 * Intialize the mutex code and system mutexes. This is called from the MD 916 * startup code prior to mi_startup(). The per-CPU data space needs to be 917 * setup before this is called. 918 */ 919 void 920 mutex_init(void) 921 { 922 923 /* Setup turnstiles so that sleep mutexes work. */ 924 init_turnstiles(); 925 926 /* 927 * Initialize mutexes. 928 */ 929 mtx_init(&Giant, "Giant", NULL, MTX_DEF | MTX_RECURSE); 930 mtx_init(&sched_lock, "sched lock", NULL, MTX_SPIN | MTX_RECURSE); 931 mtx_init(&proc0.p_mtx, "process lock", NULL, MTX_DEF | MTX_DUPOK); 932 mtx_init(&devmtx, "cdev", NULL, MTX_DEF); 933 mtx_lock(&Giant); 934 } 935 936 #ifdef DDB 937 void 938 db_show_mtx(struct lock_object *lock) 939 { 940 struct thread *td; 941 struct mtx *m; 942 943 m = (struct mtx *)lock; 944 945 db_printf(" flags: {"); 946 if (LOCK_CLASS(lock) == &lock_class_mtx_spin) 947 db_printf("SPIN"); 948 else 949 db_printf("DEF"); 950 if (m->mtx_object.lo_flags & LO_RECURSABLE) 951 db_printf(", RECURSE"); 952 if (m->mtx_object.lo_flags & LO_DUPOK) 953 db_printf(", DUPOK"); 954 db_printf("}\n"); 955 db_printf(" state: {"); 956 if (mtx_unowned(m)) 957 db_printf("UNOWNED"); 958 else { 959 db_printf("OWNED"); 960 if (m->mtx_lock & MTX_CONTESTED) 961 db_printf(", CONTESTED"); 962 if (m->mtx_lock & MTX_RECURSED) 963 db_printf(", RECURSED"); 964 } 965 db_printf("}\n"); 966 if (!mtx_unowned(m)) { 967 td = mtx_owner(m); 968 db_printf(" owner: %p (tid %d, pid %d, \"%s\")\n", td, 969 td->td_tid, td->td_proc->p_pid, td->td_proc->p_comm); 970 if (mtx_recursed(m)) 971 db_printf(" recursed: %d\n", m->mtx_recurse); 972 } 973 } 974 #endif 975