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 <fs/devfs/devfs_int.h> 71 72 #include <vm/vm.h> 73 #include <vm/vm_extern.h> 74 75 /* 76 * Force MUTEX_WAKE_ALL for now. 77 * single thread wakeup needs fixes to avoid race conditions with 78 * priority inheritance. 79 */ 80 #ifndef MUTEX_WAKE_ALL 81 #define MUTEX_WAKE_ALL 82 #endif 83 84 /* 85 * Internal utility macros. 86 */ 87 #define mtx_unowned(m) ((m)->mtx_lock == MTX_UNOWNED) 88 89 #define mtx_owner(m) (mtx_unowned((m)) ? NULL \ 90 : (struct thread *)((m)->mtx_lock & MTX_FLAGMASK)) 91 92 #ifdef DDB 93 static void db_show_mtx(struct lock_object *lock); 94 #endif 95 96 /* 97 * Lock classes for sleep and spin mutexes. 98 */ 99 struct lock_class lock_class_mtx_sleep = { 100 "sleep mutex", 101 LC_SLEEPLOCK | LC_RECURSABLE, 102 #ifdef DDB 103 db_show_mtx 104 #endif 105 }; 106 struct lock_class lock_class_mtx_spin = { 107 "spin mutex", 108 LC_SPINLOCK | LC_RECURSABLE, 109 #ifdef DDB 110 db_show_mtx 111 #endif 112 }; 113 114 /* 115 * System-wide mutexes 116 */ 117 struct mtx sched_lock; 118 struct mtx Giant; 119 120 #ifdef MUTEX_PROFILING 121 SYSCTL_NODE(_debug, OID_AUTO, mutex, CTLFLAG_RD, NULL, "mutex debugging"); 122 SYSCTL_NODE(_debug_mutex, OID_AUTO, prof, CTLFLAG_RD, NULL, "mutex profiling"); 123 static int mutex_prof_enable = 0; 124 SYSCTL_INT(_debug_mutex_prof, OID_AUTO, enable, CTLFLAG_RW, 125 &mutex_prof_enable, 0, "Enable tracing of mutex holdtime"); 126 127 struct mutex_prof { 128 const char *name; 129 const char *file; 130 int line; 131 uintmax_t cnt_max; 132 uintmax_t cnt_tot; 133 uintmax_t cnt_cur; 134 uintmax_t cnt_contest_holding; 135 uintmax_t cnt_contest_locking; 136 struct mutex_prof *next; 137 }; 138 139 /* 140 * mprof_buf is a static pool of profiling records to avoid possible 141 * reentrance of the memory allocation functions. 142 * 143 * Note: NUM_MPROF_BUFFERS must be smaller than MPROF_HASH_SIZE. 144 */ 145 #ifdef MPROF_BUFFERS 146 #define NUM_MPROF_BUFFERS MPROF_BUFFERS 147 #else 148 #define NUM_MPROF_BUFFERS 1000 149 #endif 150 static struct mutex_prof mprof_buf[NUM_MPROF_BUFFERS]; 151 static int first_free_mprof_buf; 152 #ifndef MPROF_HASH_SIZE 153 #define MPROF_HASH_SIZE 1009 154 #endif 155 #if NUM_MPROF_BUFFERS >= MPROF_HASH_SIZE 156 #error MPROF_BUFFERS must be larger than MPROF_HASH_SIZE 157 #endif 158 static struct mutex_prof *mprof_hash[MPROF_HASH_SIZE]; 159 /* SWAG: sbuf size = avg stat. line size * number of locks */ 160 #define MPROF_SBUF_SIZE 256 * 400 161 162 static int mutex_prof_acquisitions; 163 SYSCTL_INT(_debug_mutex_prof, OID_AUTO, acquisitions, CTLFLAG_RD, 164 &mutex_prof_acquisitions, 0, "Number of mutex acquistions recorded"); 165 static int mutex_prof_records; 166 SYSCTL_INT(_debug_mutex_prof, OID_AUTO, records, CTLFLAG_RD, 167 &mutex_prof_records, 0, "Number of profiling records"); 168 static int mutex_prof_maxrecords = NUM_MPROF_BUFFERS; 169 SYSCTL_INT(_debug_mutex_prof, OID_AUTO, maxrecords, CTLFLAG_RD, 170 &mutex_prof_maxrecords, 0, "Maximum number of profiling records"); 171 static int mutex_prof_rejected; 172 SYSCTL_INT(_debug_mutex_prof, OID_AUTO, rejected, CTLFLAG_RD, 173 &mutex_prof_rejected, 0, "Number of rejected profiling records"); 174 static int mutex_prof_hashsize = MPROF_HASH_SIZE; 175 SYSCTL_INT(_debug_mutex_prof, OID_AUTO, hashsize, CTLFLAG_RD, 176 &mutex_prof_hashsize, 0, "Hash size"); 177 static int mutex_prof_collisions = 0; 178 SYSCTL_INT(_debug_mutex_prof, OID_AUTO, collisions, CTLFLAG_RD, 179 &mutex_prof_collisions, 0, "Number of hash collisions"); 180 181 /* 182 * mprof_mtx protects the profiling buffers and the hash. 183 */ 184 static struct mtx mprof_mtx; 185 MTX_SYSINIT(mprof, &mprof_mtx, "mutex profiling lock", MTX_SPIN | MTX_QUIET); 186 187 static u_int64_t 188 nanoseconds(void) 189 { 190 struct timespec tv; 191 192 nanotime(&tv); 193 return (tv.tv_sec * (u_int64_t)1000000000 + tv.tv_nsec); 194 } 195 196 static int 197 dump_mutex_prof_stats(SYSCTL_HANDLER_ARGS) 198 { 199 struct sbuf *sb; 200 int error, i; 201 static int multiplier = 1; 202 203 if (first_free_mprof_buf == 0) 204 return (SYSCTL_OUT(req, "No locking recorded", 205 sizeof("No locking recorded"))); 206 207 retry_sbufops: 208 sb = sbuf_new(NULL, NULL, MPROF_SBUF_SIZE * multiplier, SBUF_FIXEDLEN); 209 sbuf_printf(sb, "\n%6s %12s %11s %5s %12s %12s %s\n", 210 "max", "total", "count", "avg", "cnt_hold", "cnt_lock", "name"); 211 /* 212 * XXX this spinlock seems to be by far the largest perpetrator 213 * of spinlock latency (1.6 msec on an Athlon1600 was recorded 214 * even before I pessimized it further by moving the average 215 * computation here). 216 */ 217 mtx_lock_spin(&mprof_mtx); 218 for (i = 0; i < first_free_mprof_buf; ++i) { 219 sbuf_printf(sb, "%6ju %12ju %11ju %5ju %12ju %12ju %s:%d (%s)\n", 220 mprof_buf[i].cnt_max / 1000, 221 mprof_buf[i].cnt_tot / 1000, 222 mprof_buf[i].cnt_cur, 223 mprof_buf[i].cnt_cur == 0 ? (uintmax_t)0 : 224 mprof_buf[i].cnt_tot / (mprof_buf[i].cnt_cur * 1000), 225 mprof_buf[i].cnt_contest_holding, 226 mprof_buf[i].cnt_contest_locking, 227 mprof_buf[i].file, mprof_buf[i].line, mprof_buf[i].name); 228 if (sbuf_overflowed(sb)) { 229 mtx_unlock_spin(&mprof_mtx); 230 sbuf_delete(sb); 231 multiplier++; 232 goto retry_sbufops; 233 } 234 } 235 mtx_unlock_spin(&mprof_mtx); 236 sbuf_finish(sb); 237 error = SYSCTL_OUT(req, sbuf_data(sb), sbuf_len(sb) + 1); 238 sbuf_delete(sb); 239 return (error); 240 } 241 SYSCTL_PROC(_debug_mutex_prof, OID_AUTO, stats, CTLTYPE_STRING | CTLFLAG_RD, 242 NULL, 0, dump_mutex_prof_stats, "A", "Mutex profiling statistics"); 243 244 static int 245 reset_mutex_prof_stats(SYSCTL_HANDLER_ARGS) 246 { 247 int error, v; 248 249 if (first_free_mprof_buf == 0) 250 return (0); 251 252 v = 0; 253 error = sysctl_handle_int(oidp, &v, 0, req); 254 if (error) 255 return (error); 256 if (req->newptr == NULL) 257 return (error); 258 if (v == 0) 259 return (0); 260 261 mtx_lock_spin(&mprof_mtx); 262 bzero(mprof_buf, sizeof(*mprof_buf) * first_free_mprof_buf); 263 bzero(mprof_hash, sizeof(struct mtx *) * MPROF_HASH_SIZE); 264 first_free_mprof_buf = 0; 265 mtx_unlock_spin(&mprof_mtx); 266 return (0); 267 } 268 SYSCTL_PROC(_debug_mutex_prof, OID_AUTO, reset, CTLTYPE_INT | CTLFLAG_RW, 269 NULL, 0, reset_mutex_prof_stats, "I", "Reset mutex profiling statistics"); 270 #endif 271 272 /* 273 * Function versions of the inlined __mtx_* macros. These are used by 274 * modules and can also be called from assembly language if needed. 275 */ 276 void 277 _mtx_lock_flags(struct mtx *m, int opts, const char *file, int line) 278 { 279 280 MPASS(curthread != NULL); 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 #ifdef MUTEX_PROFILING 291 /* don't reset the timer when/if recursing */ 292 if (m->mtx_acqtime == 0) { 293 m->mtx_filename = file; 294 m->mtx_lineno = line; 295 m->mtx_acqtime = mutex_prof_enable ? nanoseconds() : 0; 296 ++mutex_prof_acquisitions; 297 } 298 #endif 299 } 300 301 void 302 _mtx_unlock_flags(struct mtx *m, int opts, const char *file, int line) 303 { 304 305 MPASS(curthread != NULL); 306 KASSERT(LOCK_CLASS(&m->mtx_object) == &lock_class_mtx_sleep, 307 ("mtx_unlock() of spin mutex %s @ %s:%d", m->mtx_object.lo_name, 308 file, line)); 309 WITNESS_UNLOCK(&m->mtx_object, opts | LOP_EXCLUSIVE, file, line); 310 LOCK_LOG_LOCK("UNLOCK", &m->mtx_object, opts, m->mtx_recurse, file, 311 line); 312 mtx_assert(m, MA_OWNED); 313 #ifdef MUTEX_PROFILING 314 if (m->mtx_acqtime != 0) { 315 static const char *unknown = "(unknown)"; 316 struct mutex_prof *mpp; 317 u_int64_t acqtime, now; 318 const char *p, *q; 319 volatile u_int hash; 320 321 now = nanoseconds(); 322 acqtime = m->mtx_acqtime; 323 m->mtx_acqtime = 0; 324 if (now <= acqtime) 325 goto out; 326 for (p = m->mtx_filename; 327 p != NULL && strncmp(p, "../", 3) == 0; p += 3) 328 /* nothing */ ; 329 if (p == NULL || *p == '\0') 330 p = unknown; 331 for (hash = m->mtx_lineno, q = p; *q != '\0'; ++q) 332 hash = (hash * 2 + *q) % MPROF_HASH_SIZE; 333 mtx_lock_spin(&mprof_mtx); 334 for (mpp = mprof_hash[hash]; mpp != NULL; mpp = mpp->next) 335 if (mpp->line == m->mtx_lineno && 336 strcmp(mpp->file, p) == 0) 337 break; 338 if (mpp == NULL) { 339 /* Just exit if we cannot get a trace buffer */ 340 if (first_free_mprof_buf >= NUM_MPROF_BUFFERS) { 341 ++mutex_prof_rejected; 342 goto unlock; 343 } 344 mpp = &mprof_buf[first_free_mprof_buf++]; 345 mpp->name = mtx_name(m); 346 mpp->file = p; 347 mpp->line = m->mtx_lineno; 348 mpp->next = mprof_hash[hash]; 349 if (mprof_hash[hash] != NULL) 350 ++mutex_prof_collisions; 351 mprof_hash[hash] = mpp; 352 ++mutex_prof_records; 353 } 354 /* 355 * Record if the mutex has been held longer now than ever 356 * before. 357 */ 358 if (now - acqtime > mpp->cnt_max) 359 mpp->cnt_max = now - acqtime; 360 mpp->cnt_tot += now - acqtime; 361 mpp->cnt_cur++; 362 /* 363 * There's a small race, really we should cmpxchg 364 * 0 with the current value, but that would bill 365 * the contention to the wrong lock instance if 366 * it followed this also. 367 */ 368 mpp->cnt_contest_holding += m->mtx_contest_holding; 369 m->mtx_contest_holding = 0; 370 mpp->cnt_contest_locking += m->mtx_contest_locking; 371 m->mtx_contest_locking = 0; 372 unlock: 373 mtx_unlock_spin(&mprof_mtx); 374 } 375 out: 376 #endif 377 _rel_sleep_lock(m, curthread, opts, file, line); 378 } 379 380 void 381 _mtx_lock_spin_flags(struct mtx *m, int opts, const char *file, int line) 382 { 383 384 MPASS(curthread != NULL); 385 KASSERT(LOCK_CLASS(&m->mtx_object) == &lock_class_mtx_spin, 386 ("mtx_lock_spin() of sleep mutex %s @ %s:%d", 387 m->mtx_object.lo_name, file, line)); 388 WITNESS_CHECKORDER(&m->mtx_object, opts | LOP_NEWORDER | LOP_EXCLUSIVE, 389 file, line); 390 _get_spin_lock(m, curthread, opts, file, line); 391 LOCK_LOG_LOCK("LOCK", &m->mtx_object, opts, m->mtx_recurse, file, 392 line); 393 WITNESS_LOCK(&m->mtx_object, opts | LOP_EXCLUSIVE, file, line); 394 } 395 396 void 397 _mtx_unlock_spin_flags(struct mtx *m, int opts, const char *file, int line) 398 { 399 400 MPASS(curthread != NULL); 401 KASSERT(LOCK_CLASS(&m->mtx_object) == &lock_class_mtx_spin, 402 ("mtx_unlock_spin() of sleep mutex %s @ %s:%d", 403 m->mtx_object.lo_name, file, line)); 404 WITNESS_UNLOCK(&m->mtx_object, opts | LOP_EXCLUSIVE, file, line); 405 LOCK_LOG_LOCK("UNLOCK", &m->mtx_object, opts, m->mtx_recurse, file, 406 line); 407 mtx_assert(m, MA_OWNED); 408 _rel_spin_lock(m); 409 } 410 411 /* 412 * The important part of mtx_trylock{,_flags}() 413 * Tries to acquire lock `m.' If this function is called on a mutex that 414 * is already owned, it will recursively acquire the lock. 415 */ 416 int 417 _mtx_trylock(struct mtx *m, int opts, const char *file, int line) 418 { 419 int rval; 420 421 MPASS(curthread != NULL); 422 KASSERT(LOCK_CLASS(&m->mtx_object) == &lock_class_mtx_sleep, 423 ("mtx_trylock() of spin mutex %s @ %s:%d", m->mtx_object.lo_name, 424 file, line)); 425 426 if (mtx_owned(m) && (m->mtx_object.lo_flags & LO_RECURSABLE) != 0) { 427 m->mtx_recurse++; 428 atomic_set_ptr(&m->mtx_lock, MTX_RECURSED); 429 rval = 1; 430 } else 431 rval = _obtain_lock(m, (uintptr_t)curthread); 432 433 LOCK_LOG_TRY("LOCK", &m->mtx_object, opts, rval, file, line); 434 if (rval) 435 WITNESS_LOCK(&m->mtx_object, opts | LOP_EXCLUSIVE | LOP_TRYLOCK, 436 file, line); 437 438 return (rval); 439 } 440 441 /* 442 * _mtx_lock_sleep: the tougher part of acquiring an MTX_DEF lock. 443 * 444 * We call this if the lock is either contested (i.e. we need to go to 445 * sleep waiting for it), or if we need to recurse on it. 446 */ 447 void 448 _mtx_lock_sleep(struct mtx *m, uintptr_t tid, int opts, const char *file, 449 int line) 450 { 451 #if defined(SMP) && !defined(NO_ADAPTIVE_MUTEXES) 452 struct thread *owner; 453 #endif 454 uintptr_t v; 455 #ifdef KTR 456 int cont_logged = 0; 457 #endif 458 #ifdef MUTEX_PROFILING 459 int contested; 460 #endif 461 462 if (mtx_owned(m)) { 463 KASSERT((m->mtx_object.lo_flags & LO_RECURSABLE) != 0, 464 ("_mtx_lock_sleep: recursed on non-recursive mutex %s @ %s:%d\n", 465 m->mtx_object.lo_name, file, line)); 466 m->mtx_recurse++; 467 atomic_set_ptr(&m->mtx_lock, MTX_RECURSED); 468 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 469 CTR1(KTR_LOCK, "_mtx_lock_sleep: %p recursing", m); 470 return; 471 } 472 473 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 474 CTR4(KTR_LOCK, 475 "_mtx_lock_sleep: %s contested (lock=%p) at %s:%d", 476 m->mtx_object.lo_name, (void *)m->mtx_lock, file, line); 477 478 #ifdef MUTEX_PROFILING 479 contested = 0; 480 #endif 481 while (!_obtain_lock(m, tid)) { 482 #ifdef MUTEX_PROFILING 483 contested = 1; 484 atomic_add_int(&m->mtx_contest_holding, 1); 485 #endif 486 turnstile_lock(&m->mtx_object); 487 v = m->mtx_lock; 488 489 /* 490 * Check if the lock has been released while spinning for 491 * the turnstile chain lock. 492 */ 493 if (v == MTX_UNOWNED) { 494 turnstile_release(&m->mtx_object); 495 cpu_spinwait(); 496 continue; 497 } 498 499 #ifdef MUTEX_WAKE_ALL 500 MPASS(v != MTX_CONTESTED); 501 #else 502 /* 503 * The mutex was marked contested on release. This means that 504 * there are other threads blocked on it. Grab ownership of 505 * it and propagate its priority to the current thread if 506 * necessary. 507 */ 508 if (v == MTX_CONTESTED) { 509 m->mtx_lock = tid | MTX_CONTESTED; 510 turnstile_claim(&m->mtx_object); 511 break; 512 } 513 #endif 514 515 /* 516 * If the mutex isn't already contested and a failure occurs 517 * setting the contested bit, the mutex was either released 518 * or the state of the MTX_RECURSED bit changed. 519 */ 520 if ((v & MTX_CONTESTED) == 0 && 521 !atomic_cmpset_ptr(&m->mtx_lock, v, v | MTX_CONTESTED)) { 522 turnstile_release(&m->mtx_object); 523 cpu_spinwait(); 524 continue; 525 } 526 527 #if defined(SMP) && !defined(NO_ADAPTIVE_MUTEXES) 528 /* 529 * If the current owner of the lock is executing on another 530 * CPU, spin instead of blocking. 531 */ 532 owner = (struct thread *)(v & MTX_FLAGMASK); 533 #ifdef ADAPTIVE_GIANT 534 if (TD_IS_RUNNING(owner)) { 535 #else 536 if (m != &Giant && TD_IS_RUNNING(owner)) { 537 #endif 538 turnstile_release(&m->mtx_object); 539 while (mtx_owner(m) == owner && TD_IS_RUNNING(owner)) { 540 cpu_spinwait(); 541 } 542 continue; 543 } 544 #endif /* SMP && !NO_ADAPTIVE_MUTEXES */ 545 546 /* 547 * We definitely must sleep for this lock. 548 */ 549 mtx_assert(m, MA_NOTOWNED); 550 551 #ifdef KTR 552 if (!cont_logged) { 553 CTR6(KTR_CONTENTION, 554 "contention: %p at %s:%d wants %s, taken by %s:%d", 555 (void *)tid, file, line, m->mtx_object.lo_name, 556 WITNESS_FILE(&m->mtx_object), 557 WITNESS_LINE(&m->mtx_object)); 558 cont_logged = 1; 559 } 560 #endif 561 562 /* 563 * Block on the turnstile. 564 */ 565 turnstile_wait(&m->mtx_object, mtx_owner(m), 566 TS_EXCLUSIVE_QUEUE); 567 } 568 569 #ifdef KTR 570 if (cont_logged) { 571 CTR4(KTR_CONTENTION, 572 "contention end: %s acquired by %p at %s:%d", 573 m->mtx_object.lo_name, (void *)tid, file, line); 574 } 575 #endif 576 #ifdef MUTEX_PROFILING 577 if (contested) 578 m->mtx_contest_locking++; 579 m->mtx_contest_holding = 0; 580 #endif 581 return; 582 } 583 584 #ifdef SMP 585 /* 586 * _mtx_lock_spin: the tougher part of acquiring an MTX_SPIN lock. 587 * 588 * This is only called if we need to actually spin for the lock. Recursion 589 * is handled inline. 590 */ 591 void 592 _mtx_lock_spin(struct mtx *m, uintptr_t tid, int opts, const char *file, 593 int line) 594 { 595 int i = 0; 596 597 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 598 CTR1(KTR_LOCK, "_mtx_lock_spin: %p spinning", m); 599 600 for (;;) { 601 if (_obtain_lock(m, tid)) 602 break; 603 604 /* Give interrupts a chance while we spin. */ 605 spinlock_exit(); 606 while (m->mtx_lock != MTX_UNOWNED) { 607 if (i++ < 10000000) { 608 cpu_spinwait(); 609 continue; 610 } 611 if (i < 60000000) 612 DELAY(1); 613 else if (!kdb_active && !panicstr) { 614 printf("spin lock %s held by %p for > 5 seconds\n", 615 m->mtx_object.lo_name, (void *)m->mtx_lock); 616 #ifdef WITNESS 617 witness_display_spinlock(&m->mtx_object, 618 mtx_owner(m)); 619 #endif 620 panic("spin lock held too long"); 621 } 622 cpu_spinwait(); 623 } 624 spinlock_enter(); 625 } 626 627 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 628 CTR1(KTR_LOCK, "_mtx_lock_spin: %p spin done", m); 629 630 return; 631 } 632 #endif /* SMP */ 633 634 /* 635 * _mtx_unlock_sleep: the tougher part of releasing an MTX_DEF lock. 636 * 637 * We are only called here if the lock is recursed or contested (i.e. we 638 * need to wake up a blocked thread). 639 */ 640 void 641 _mtx_unlock_sleep(struct mtx *m, int opts, const char *file, int line) 642 { 643 struct turnstile *ts; 644 #ifndef PREEMPTION 645 struct thread *td, *td1; 646 #endif 647 648 if (mtx_recursed(m)) { 649 if (--(m->mtx_recurse) == 0) 650 atomic_clear_ptr(&m->mtx_lock, MTX_RECURSED); 651 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 652 CTR1(KTR_LOCK, "_mtx_unlock_sleep: %p unrecurse", m); 653 return; 654 } 655 656 turnstile_lock(&m->mtx_object); 657 ts = turnstile_lookup(&m->mtx_object); 658 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 659 CTR1(KTR_LOCK, "_mtx_unlock_sleep: %p contested", m); 660 661 #if defined(SMP) && !defined(NO_ADAPTIVE_MUTEXES) 662 if (ts == NULL) { 663 _release_lock_quick(m); 664 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 665 CTR1(KTR_LOCK, "_mtx_unlock_sleep: %p no sleepers", m); 666 turnstile_release(&m->mtx_object); 667 return; 668 } 669 #else 670 MPASS(ts != NULL); 671 #endif 672 #ifndef PREEMPTION 673 /* XXX */ 674 td1 = turnstile_head(ts, TS_EXCLUSIVE_QUEUE); 675 #endif 676 #ifdef MUTEX_WAKE_ALL 677 turnstile_broadcast(ts, TS_EXCLUSIVE_QUEUE); 678 _release_lock_quick(m); 679 #else 680 if (turnstile_signal(ts, TS_EXCLUSIVE_QUEUE)) { 681 _release_lock_quick(m); 682 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 683 CTR1(KTR_LOCK, "_mtx_unlock_sleep: %p not held", m); 684 } else { 685 m->mtx_lock = MTX_CONTESTED; 686 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 687 CTR1(KTR_LOCK, "_mtx_unlock_sleep: %p still contested", 688 m); 689 } 690 #endif 691 turnstile_unpend(ts, TS_EXCLUSIVE_LOCK); 692 693 #ifndef PREEMPTION 694 /* 695 * XXX: This is just a hack until preemption is done. However, 696 * once preemption is done we need to either wrap the 697 * turnstile_signal() and release of the actual lock in an 698 * extra critical section or change the preemption code to 699 * always just set a flag and never do instant-preempts. 700 */ 701 td = curthread; 702 if (td->td_critnest > 0 || td1->td_priority >= td->td_priority) 703 return; 704 mtx_lock_spin(&sched_lock); 705 if (!TD_IS_RUNNING(td1)) { 706 #ifdef notyet 707 if (td->td_ithd != NULL) { 708 struct ithd *it = td->td_ithd; 709 710 if (it->it_interrupted) { 711 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 712 CTR2(KTR_LOCK, 713 "_mtx_unlock_sleep: %p interrupted %p", 714 it, it->it_interrupted); 715 intr_thd_fixup(it); 716 } 717 } 718 #endif 719 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 720 CTR2(KTR_LOCK, 721 "_mtx_unlock_sleep: %p switching out lock=%p", m, 722 (void *)m->mtx_lock); 723 724 mi_switch(SW_INVOL, NULL); 725 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 726 CTR2(KTR_LOCK, "_mtx_unlock_sleep: %p resuming lock=%p", 727 m, (void *)m->mtx_lock); 728 } 729 mtx_unlock_spin(&sched_lock); 730 #endif 731 732 return; 733 } 734 735 /* 736 * All the unlocking of MTX_SPIN locks is done inline. 737 * See the _rel_spin_lock() macro for the details. 738 */ 739 740 /* 741 * The backing function for the INVARIANTS-enabled mtx_assert() 742 */ 743 #ifdef INVARIANT_SUPPORT 744 void 745 _mtx_assert(struct mtx *m, int what, const char *file, int line) 746 { 747 748 if (panicstr != NULL || dumping) 749 return; 750 switch (what) { 751 case MA_OWNED: 752 case MA_OWNED | MA_RECURSED: 753 case MA_OWNED | MA_NOTRECURSED: 754 if (!mtx_owned(m)) 755 panic("mutex %s not owned at %s:%d", 756 m->mtx_object.lo_name, file, line); 757 if (mtx_recursed(m)) { 758 if ((what & MA_NOTRECURSED) != 0) 759 panic("mutex %s recursed at %s:%d", 760 m->mtx_object.lo_name, file, line); 761 } else if ((what & MA_RECURSED) != 0) { 762 panic("mutex %s unrecursed at %s:%d", 763 m->mtx_object.lo_name, file, line); 764 } 765 break; 766 case MA_NOTOWNED: 767 if (mtx_owned(m)) 768 panic("mutex %s owned at %s:%d", 769 m->mtx_object.lo_name, file, line); 770 break; 771 default: 772 panic("unknown mtx_assert at %s:%d", file, line); 773 } 774 } 775 #endif 776 777 /* 778 * The MUTEX_DEBUG-enabled mtx_validate() 779 * 780 * Most of these checks have been moved off into the LO_INITIALIZED flag 781 * maintained by the witness code. 782 */ 783 #ifdef MUTEX_DEBUG 784 785 void mtx_validate(struct mtx *); 786 787 void 788 mtx_validate(struct mtx *m) 789 { 790 791 /* 792 * XXX: When kernacc() does not require Giant we can reenable this check 793 */ 794 #ifdef notyet 795 /* 796 * XXX - When kernacc() is fixed on the alpha to handle K0_SEG memory properly 797 * we can re-enable the kernacc() checks. 798 */ 799 #ifndef __alpha__ 800 /* 801 * Can't call kernacc() from early init386(), especially when 802 * initializing Giant mutex, because some stuff in kernacc() 803 * requires Giant itself. 804 */ 805 if (!cold) 806 if (!kernacc((caddr_t)m, sizeof(m), 807 VM_PROT_READ | VM_PROT_WRITE)) 808 panic("Can't read and write to mutex %p", m); 809 #endif 810 #endif 811 } 812 #endif 813 814 /* 815 * General init routine used by the MTX_SYSINIT() macro. 816 */ 817 void 818 mtx_sysinit(void *arg) 819 { 820 struct mtx_args *margs = arg; 821 822 mtx_init(margs->ma_mtx, margs->ma_desc, NULL, margs->ma_opts); 823 } 824 825 /* 826 * Mutex initialization routine; initialize lock `m' of type contained in 827 * `opts' with options contained in `opts' and name `name.' The optional 828 * lock type `type' is used as a general lock category name for use with 829 * witness. 830 */ 831 void 832 mtx_init(struct mtx *m, const char *name, const char *type, int opts) 833 { 834 struct lock_class *class; 835 int flags; 836 837 MPASS((opts & ~(MTX_SPIN | MTX_QUIET | MTX_RECURSE | 838 MTX_NOWITNESS | MTX_DUPOK)) == 0); 839 840 #ifdef MUTEX_DEBUG 841 /* Diagnostic and error correction */ 842 mtx_validate(m); 843 #endif 844 845 /* Determine lock class and lock flags. */ 846 if (opts & MTX_SPIN) 847 class = &lock_class_mtx_spin; 848 else 849 class = &lock_class_mtx_sleep; 850 flags = 0; 851 if (opts & MTX_QUIET) 852 flags |= LO_QUIET; 853 if (opts & MTX_RECURSE) 854 flags |= LO_RECURSABLE; 855 if ((opts & MTX_NOWITNESS) == 0) 856 flags |= LO_WITNESS; 857 if (opts & MTX_DUPOK) 858 flags |= LO_DUPOK; 859 860 /* Initialize mutex. */ 861 m->mtx_lock = MTX_UNOWNED; 862 m->mtx_recurse = 0; 863 #ifdef MUTEX_PROFILING 864 m->mtx_acqtime = 0; 865 m->mtx_filename = NULL; 866 m->mtx_lineno = 0; 867 m->mtx_contest_holding = 0; 868 m->mtx_contest_locking = 0; 869 #endif 870 871 lock_init(&m->mtx_object, class, name, type, flags); 872 } 873 874 /* 875 * Remove lock `m' from all_mtx queue. We don't allow MTX_QUIET to be 876 * passed in as a flag here because if the corresponding mtx_init() was 877 * called with MTX_QUIET set, then it will already be set in the mutex's 878 * flags. 879 */ 880 void 881 mtx_destroy(struct mtx *m) 882 { 883 884 if (!mtx_owned(m)) 885 MPASS(mtx_unowned(m)); 886 else { 887 MPASS((m->mtx_lock & (MTX_RECURSED|MTX_CONTESTED)) == 0); 888 889 /* Perform the non-mtx related part of mtx_unlock_spin(). */ 890 if (LOCK_CLASS(&m->mtx_object) == &lock_class_mtx_spin) 891 spinlock_exit(); 892 893 /* Tell witness this isn't locked to make it happy. */ 894 WITNESS_UNLOCK(&m->mtx_object, LOP_EXCLUSIVE, __FILE__, 895 __LINE__); 896 } 897 898 lock_destroy(&m->mtx_object); 899 } 900 901 /* 902 * Intialize the mutex code and system mutexes. This is called from the MD 903 * startup code prior to mi_startup(). The per-CPU data space needs to be 904 * setup before this is called. 905 */ 906 void 907 mutex_init(void) 908 { 909 910 /* Setup turnstiles so that sleep mutexes work. */ 911 init_turnstiles(); 912 913 /* 914 * Initialize mutexes. 915 */ 916 mtx_init(&Giant, "Giant", NULL, MTX_DEF | MTX_RECURSE); 917 mtx_init(&sched_lock, "sched lock", NULL, MTX_SPIN | MTX_RECURSE); 918 mtx_init(&proc0.p_mtx, "process lock", NULL, MTX_DEF | MTX_DUPOK); 919 mtx_init(&devmtx, "cdev", NULL, MTX_DEF); 920 mtx_lock(&Giant); 921 } 922 923 #ifdef DDB 924 void 925 db_show_mtx(struct lock_object *lock) 926 { 927 struct thread *td; 928 struct mtx *m; 929 930 m = (struct mtx *)lock; 931 932 db_printf(" flags: {"); 933 if (LOCK_CLASS(lock) == &lock_class_mtx_spin) 934 db_printf("SPIN"); 935 else 936 db_printf("DEF"); 937 if (m->mtx_object.lo_flags & LO_RECURSABLE) 938 db_printf(", RECURSE"); 939 if (m->mtx_object.lo_flags & LO_DUPOK) 940 db_printf(", DUPOK"); 941 db_printf("}\n"); 942 db_printf(" state: {"); 943 if (mtx_unowned(m)) 944 db_printf("UNOWNED"); 945 else { 946 db_printf("OWNED"); 947 if (m->mtx_lock & MTX_CONTESTED) 948 db_printf(", CONTESTED"); 949 if (m->mtx_lock & MTX_RECURSED) 950 db_printf(", RECURSED"); 951 } 952 db_printf("}\n"); 953 if (!mtx_unowned(m)) { 954 td = mtx_owner(m); 955 db_printf(" owner: %p (tid %d, pid %d, \"%s\")\n", td, 956 td->td_tid, td->td_proc->p_pid, td->td_proc->p_comm); 957 if (mtx_recursed(m)) 958 db_printf(" recursed: %d\n", m->mtx_recurse); 959 } 960 } 961 #endif 962