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: releng/11.0/sys/kern/kern_mutex.c 303953 2016-08-11 09:28:49Z mjg $"); 38 39#include "opt_adaptive_mutexes.h" 40#include "opt_ddb.h" 41#include "opt_hwpmc_hooks.h" 42#include "opt_sched.h" 43 44#include <sys/param.h> 45#include <sys/systm.h> 46#include <sys/bus.h> 47#include <sys/conf.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/smp.h> 59#include <sys/sysctl.h> 60#include <sys/turnstile.h> 61#include <sys/vmmeter.h> 62#include <sys/lock_profile.h> 63 64#include <machine/atomic.h> 65#include <machine/bus.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#if defined(SMP) && !defined(NO_ADAPTIVE_MUTEXES) 76#define ADAPTIVE_MUTEXES 77#endif 78 79#ifdef HWPMC_HOOKS 80#include <sys/pmckern.h> 81PMC_SOFT_DEFINE( , , lock, failed); 82#endif 83 84/* 85 * Return the mutex address when the lock cookie address is provided. 86 * This functionality assumes that struct mtx* have a member named mtx_lock. 87 */ 88#define mtxlock2mtx(c) (__containerof(c, struct mtx, mtx_lock)) 89 90/* 91 * Internal utility macros. 92 */ 93#define mtx_unowned(m) ((m)->mtx_lock == MTX_UNOWNED) 94 95#define mtx_destroyed(m) ((m)->mtx_lock == MTX_DESTROYED) 96 97#define mtx_owner(m) ((struct thread *)((m)->mtx_lock & ~MTX_FLAGMASK)) 98 99static void assert_mtx(const struct lock_object *lock, int what); 100#ifdef DDB 101static void db_show_mtx(const struct lock_object *lock); 102#endif 103static void lock_mtx(struct lock_object *lock, uintptr_t how); 104static void lock_spin(struct lock_object *lock, uintptr_t how); 105#ifdef KDTRACE_HOOKS 106static int owner_mtx(const struct lock_object *lock, 107 struct thread **owner); 108#endif 109static uintptr_t unlock_mtx(struct lock_object *lock); 110static uintptr_t unlock_spin(struct lock_object *lock); 111 112/* 113 * Lock classes for sleep and spin mutexes. 114 */ 115struct lock_class lock_class_mtx_sleep = { 116 .lc_name = "sleep mutex", 117 .lc_flags = LC_SLEEPLOCK | LC_RECURSABLE, 118 .lc_assert = assert_mtx, 119#ifdef DDB 120 .lc_ddb_show = db_show_mtx, 121#endif 122 .lc_lock = lock_mtx, 123 .lc_unlock = unlock_mtx, 124#ifdef KDTRACE_HOOKS 125 .lc_owner = owner_mtx, 126#endif 127}; 128struct lock_class lock_class_mtx_spin = { 129 .lc_name = "spin mutex", 130 .lc_flags = LC_SPINLOCK | LC_RECURSABLE, 131 .lc_assert = assert_mtx, 132#ifdef DDB 133 .lc_ddb_show = db_show_mtx, 134#endif 135 .lc_lock = lock_spin, 136 .lc_unlock = unlock_spin, 137#ifdef KDTRACE_HOOKS 138 .lc_owner = owner_mtx, 139#endif 140}; 141 142#ifdef ADAPTIVE_MUTEXES 143static SYSCTL_NODE(_debug, OID_AUTO, mtx, CTLFLAG_RD, NULL, "mtx debugging"); 144 145static struct lock_delay_config mtx_delay = { 146 .initial = 1000, 147 .step = 500, 148 .min = 100, 149 .max = 5000, 150}; 151 152SYSCTL_INT(_debug_mtx, OID_AUTO, delay_initial, CTLFLAG_RW, &mtx_delay.initial, 153 0, ""); 154SYSCTL_INT(_debug_mtx, OID_AUTO, delay_step, CTLFLAG_RW, &mtx_delay.step, 155 0, ""); 156SYSCTL_INT(_debug_mtx, OID_AUTO, delay_min, CTLFLAG_RW, &mtx_delay.min, 157 0, ""); 158SYSCTL_INT(_debug_mtx, OID_AUTO, delay_max, CTLFLAG_RW, &mtx_delay.max, 159 0, ""); 160 161static void 162mtx_delay_sysinit(void *dummy) 163{ 164 165 mtx_delay.initial = mp_ncpus * 25; 166 mtx_delay.step = (mp_ncpus * 25) / 2; 167 mtx_delay.min = mp_ncpus * 5; 168 mtx_delay.max = mp_ncpus * 25 * 10; 169} 170LOCK_DELAY_SYSINIT(mtx_delay_sysinit); 171#endif 172 173/* 174 * System-wide mutexes 175 */ 176struct mtx blocked_lock; 177struct mtx Giant; 178 179void 180assert_mtx(const struct lock_object *lock, int what) 181{ 182 183 mtx_assert((const struct mtx *)lock, what); 184} 185 186void 187lock_mtx(struct lock_object *lock, uintptr_t how) 188{ 189 190 mtx_lock((struct mtx *)lock); 191} 192 193void 194lock_spin(struct lock_object *lock, uintptr_t how) 195{ 196 197 panic("spin locks can only use msleep_spin"); 198} 199 200uintptr_t 201unlock_mtx(struct lock_object *lock) 202{ 203 struct mtx *m; 204 205 m = (struct mtx *)lock; 206 mtx_assert(m, MA_OWNED | MA_NOTRECURSED); 207 mtx_unlock(m); 208 return (0); 209} 210 211uintptr_t 212unlock_spin(struct lock_object *lock) 213{ 214 215 panic("spin locks can only use msleep_spin"); 216} 217 218#ifdef KDTRACE_HOOKS 219int 220owner_mtx(const struct lock_object *lock, struct thread **owner) 221{ 222 const struct mtx *m = (const struct mtx *)lock; 223 224 *owner = mtx_owner(m); 225 return (mtx_unowned(m) == 0); 226} 227#endif 228 229/* 230 * Function versions of the inlined __mtx_* macros. These are used by 231 * modules and can also be called from assembly language if needed. 232 */ 233void 234__mtx_lock_flags(volatile uintptr_t *c, int opts, const char *file, int line) 235{ 236 struct mtx *m; 237 238 if (SCHEDULER_STOPPED()) 239 return; 240 241 m = mtxlock2mtx(c); 242 243 KASSERT(kdb_active != 0 || !TD_IS_IDLETHREAD(curthread), 244 ("mtx_lock() by idle thread %p on sleep mutex %s @ %s:%d", 245 curthread, m->lock_object.lo_name, file, line)); 246 KASSERT(m->mtx_lock != MTX_DESTROYED, 247 ("mtx_lock() of destroyed mutex @ %s:%d", file, line)); 248 KASSERT(LOCK_CLASS(&m->lock_object) == &lock_class_mtx_sleep, 249 ("mtx_lock() of spin mutex %s @ %s:%d", m->lock_object.lo_name, 250 file, line)); 251 WITNESS_CHECKORDER(&m->lock_object, (opts & ~MTX_RECURSE) | 252 LOP_NEWORDER | LOP_EXCLUSIVE, file, line, NULL); 253 254 __mtx_lock(m, curthread, opts, file, line); 255 LOCK_LOG_LOCK("LOCK", &m->lock_object, opts, m->mtx_recurse, file, 256 line); 257 WITNESS_LOCK(&m->lock_object, (opts & ~MTX_RECURSE) | LOP_EXCLUSIVE, 258 file, line); 259 TD_LOCKS_INC(curthread); 260} 261 262void 263__mtx_unlock_flags(volatile uintptr_t *c, int opts, const char *file, int line) 264{ 265 struct mtx *m; 266 267 if (SCHEDULER_STOPPED()) 268 return; 269 270 m = mtxlock2mtx(c); 271 272 KASSERT(m->mtx_lock != MTX_DESTROYED, 273 ("mtx_unlock() of destroyed mutex @ %s:%d", file, line)); 274 KASSERT(LOCK_CLASS(&m->lock_object) == &lock_class_mtx_sleep, 275 ("mtx_unlock() of spin mutex %s @ %s:%d", m->lock_object.lo_name, 276 file, line)); 277 WITNESS_UNLOCK(&m->lock_object, opts | LOP_EXCLUSIVE, file, line); 278 LOCK_LOG_LOCK("UNLOCK", &m->lock_object, opts, m->mtx_recurse, file, 279 line); 280 mtx_assert(m, MA_OWNED); 281 282 __mtx_unlock(m, curthread, opts, file, line); 283 TD_LOCKS_DEC(curthread); 284} 285 286void 287__mtx_lock_spin_flags(volatile uintptr_t *c, int opts, const char *file, 288 int line) 289{ 290 struct mtx *m; 291 292 if (SCHEDULER_STOPPED()) 293 return; 294 295 m = mtxlock2mtx(c); 296 297 KASSERT(m->mtx_lock != MTX_DESTROYED, 298 ("mtx_lock_spin() of destroyed mutex @ %s:%d", file, line)); 299 KASSERT(LOCK_CLASS(&m->lock_object) == &lock_class_mtx_spin, 300 ("mtx_lock_spin() of sleep mutex %s @ %s:%d", 301 m->lock_object.lo_name, file, line)); 302 if (mtx_owned(m)) 303 KASSERT((m->lock_object.lo_flags & LO_RECURSABLE) != 0 || 304 (opts & MTX_RECURSE) != 0, 305 ("mtx_lock_spin: recursed on non-recursive mutex %s @ %s:%d\n", 306 m->lock_object.lo_name, file, line)); 307 opts &= ~MTX_RECURSE; 308 WITNESS_CHECKORDER(&m->lock_object, opts | LOP_NEWORDER | LOP_EXCLUSIVE, 309 file, line, NULL); 310 __mtx_lock_spin(m, curthread, opts, file, line); 311 LOCK_LOG_LOCK("LOCK", &m->lock_object, opts, m->mtx_recurse, file, 312 line); 313 WITNESS_LOCK(&m->lock_object, opts | LOP_EXCLUSIVE, file, line); 314} 315 316int 317__mtx_trylock_spin_flags(volatile uintptr_t *c, int opts, const char *file, 318 int line) 319{ 320 struct mtx *m; 321 322 if (SCHEDULER_STOPPED()) 323 return (1); 324 325 m = mtxlock2mtx(c); 326 327 KASSERT(m->mtx_lock != MTX_DESTROYED, 328 ("mtx_trylock_spin() of destroyed mutex @ %s:%d", file, line)); 329 KASSERT(LOCK_CLASS(&m->lock_object) == &lock_class_mtx_spin, 330 ("mtx_trylock_spin() of sleep mutex %s @ %s:%d", 331 m->lock_object.lo_name, file, line)); 332 KASSERT((opts & MTX_RECURSE) == 0, 333 ("mtx_trylock_spin: unsupp. opt MTX_RECURSE on mutex %s @ %s:%d\n", 334 m->lock_object.lo_name, file, line)); 335 if (__mtx_trylock_spin(m, curthread, opts, file, line)) { 336 LOCK_LOG_TRY("LOCK", &m->lock_object, opts, 1, file, line); 337 WITNESS_LOCK(&m->lock_object, opts | LOP_EXCLUSIVE, file, line); 338 return (1); 339 } 340 LOCK_LOG_TRY("LOCK", &m->lock_object, opts, 0, file, line); 341 return (0); 342} 343 344void 345__mtx_unlock_spin_flags(volatile uintptr_t *c, int opts, const char *file, 346 int line) 347{ 348 struct mtx *m; 349 350 if (SCHEDULER_STOPPED()) 351 return; 352 353 m = mtxlock2mtx(c); 354 355 KASSERT(m->mtx_lock != MTX_DESTROYED, 356 ("mtx_unlock_spin() of destroyed mutex @ %s:%d", file, line)); 357 KASSERT(LOCK_CLASS(&m->lock_object) == &lock_class_mtx_spin, 358 ("mtx_unlock_spin() of sleep mutex %s @ %s:%d", 359 m->lock_object.lo_name, file, line)); 360 WITNESS_UNLOCK(&m->lock_object, opts | LOP_EXCLUSIVE, file, line); 361 LOCK_LOG_LOCK("UNLOCK", &m->lock_object, opts, m->mtx_recurse, file, 362 line); 363 mtx_assert(m, MA_OWNED); 364 365 __mtx_unlock_spin(m); 366} 367 368/* 369 * The important part of mtx_trylock{,_flags}() 370 * Tries to acquire lock `m.' If this function is called on a mutex that 371 * is already owned, it will recursively acquire the lock. 372 */ 373int 374_mtx_trylock_flags_(volatile uintptr_t *c, int opts, const char *file, int line) 375{ 376 struct mtx *m; 377#ifdef LOCK_PROFILING 378 uint64_t waittime = 0; 379 int contested = 0; 380#endif 381 int rval; 382 383 if (SCHEDULER_STOPPED()) 384 return (1); 385 386 m = mtxlock2mtx(c); 387 388 KASSERT(kdb_active != 0 || !TD_IS_IDLETHREAD(curthread), 389 ("mtx_trylock() by idle thread %p on sleep mutex %s @ %s:%d", 390 curthread, m->lock_object.lo_name, file, line)); 391 KASSERT(m->mtx_lock != MTX_DESTROYED, 392 ("mtx_trylock() of destroyed mutex @ %s:%d", file, line)); 393 KASSERT(LOCK_CLASS(&m->lock_object) == &lock_class_mtx_sleep, 394 ("mtx_trylock() of spin mutex %s @ %s:%d", m->lock_object.lo_name, 395 file, line)); 396 397 if (mtx_owned(m) && ((m->lock_object.lo_flags & LO_RECURSABLE) != 0 || 398 (opts & MTX_RECURSE) != 0)) { 399 m->mtx_recurse++; 400 atomic_set_ptr(&m->mtx_lock, MTX_RECURSED); 401 rval = 1; 402 } else 403 rval = _mtx_obtain_lock(m, (uintptr_t)curthread); 404 opts &= ~MTX_RECURSE; 405 406 LOCK_LOG_TRY("LOCK", &m->lock_object, opts, rval, file, line); 407 if (rval) { 408 WITNESS_LOCK(&m->lock_object, opts | LOP_EXCLUSIVE | LOP_TRYLOCK, 409 file, line); 410 TD_LOCKS_INC(curthread); 411 if (m->mtx_recurse == 0) 412 LOCKSTAT_PROFILE_OBTAIN_LOCK_SUCCESS(adaptive__acquire, 413 m, contested, waittime, file, line); 414 415 } 416 417 return (rval); 418} 419 420/* 421 * __mtx_lock_sleep: the tougher part of acquiring an MTX_DEF lock. 422 * 423 * We call this if the lock is either contested (i.e. we need to go to 424 * sleep waiting for it), or if we need to recurse on it. 425 */ 426void 427__mtx_lock_sleep(volatile uintptr_t *c, uintptr_t tid, int opts, 428 const char *file, int line) 429{ 430 struct mtx *m; 431 struct turnstile *ts; 432 uintptr_t v; 433#ifdef ADAPTIVE_MUTEXES 434 volatile struct thread *owner; 435#endif 436#ifdef KTR 437 int cont_logged = 0; 438#endif 439#ifdef LOCK_PROFILING 440 int contested = 0; 441 uint64_t waittime = 0; 442#endif 443#if defined(ADAPTIVE_MUTEXES) || defined(KDTRACE_HOOKS) 444 struct lock_delay_arg lda; 445#endif 446#ifdef KDTRACE_HOOKS 447 u_int sleep_cnt = 0; 448 int64_t sleep_time = 0; 449 int64_t all_time = 0; 450#endif 451 452 if (SCHEDULER_STOPPED()) 453 return; 454 455#if defined(ADAPTIVE_MUTEXES) 456 lock_delay_arg_init(&lda, &mtx_delay); 457#elif defined(KDTRACE_HOOKS) 458 lock_delay_arg_init(&lda, NULL); 459#endif 460 m = mtxlock2mtx(c); 461 462 if (mtx_owned(m)) { 463 KASSERT((m->lock_object.lo_flags & LO_RECURSABLE) != 0 || 464 (opts & MTX_RECURSE) != 0, 465 ("_mtx_lock_sleep: recursed on non-recursive mutex %s @ %s:%d\n", 466 m->lock_object.lo_name, file, line)); 467 opts &= ~MTX_RECURSE; 468 m->mtx_recurse++; 469 atomic_set_ptr(&m->mtx_lock, MTX_RECURSED); 470 if (LOCK_LOG_TEST(&m->lock_object, opts)) 471 CTR1(KTR_LOCK, "_mtx_lock_sleep: %p recursing", m); 472 return; 473 } 474 opts &= ~MTX_RECURSE; 475 476#ifdef HWPMC_HOOKS 477 PMC_SOFT_CALL( , , lock, failed); 478#endif 479 lock_profile_obtain_lock_failed(&m->lock_object, 480 &contested, &waittime); 481 if (LOCK_LOG_TEST(&m->lock_object, opts)) 482 CTR4(KTR_LOCK, 483 "_mtx_lock_sleep: %s contested (lock=%p) at %s:%d", 484 m->lock_object.lo_name, (void *)m->mtx_lock, file, line); 485#ifdef KDTRACE_HOOKS 486 all_time -= lockstat_nsecs(&m->lock_object); 487#endif 488 489 for (;;) { 490 if (m->mtx_lock == MTX_UNOWNED && _mtx_obtain_lock(m, tid)) 491 break; 492#ifdef KDTRACE_HOOKS 493 lda.spin_cnt++; 494#endif 495#ifdef ADAPTIVE_MUTEXES 496 /* 497 * If the owner is running on another CPU, spin until the 498 * owner stops running or the state of the lock changes. 499 */ 500 v = m->mtx_lock; 501 if (v != MTX_UNOWNED) { 502 owner = (struct thread *)(v & ~MTX_FLAGMASK); 503 if (TD_IS_RUNNING(owner)) { 504 if (LOCK_LOG_TEST(&m->lock_object, 0)) 505 CTR3(KTR_LOCK, 506 "%s: spinning on %p held by %p", 507 __func__, m, owner); 508 KTR_STATE1(KTR_SCHED, "thread", 509 sched_tdname((struct thread *)tid), 510 "spinning", "lockname:\"%s\"", 511 m->lock_object.lo_name); 512 while (mtx_owner(m) == owner && 513 TD_IS_RUNNING(owner)) 514 lock_delay(&lda); 515 KTR_STATE0(KTR_SCHED, "thread", 516 sched_tdname((struct thread *)tid), 517 "running"); 518 continue; 519 } 520 } 521#endif 522 523 ts = turnstile_trywait(&m->lock_object); 524 v = m->mtx_lock; 525 526 /* 527 * Check if the lock has been released while spinning for 528 * the turnstile chain lock. 529 */ 530 if (v == MTX_UNOWNED) { 531 turnstile_cancel(ts); 532 continue; 533 } 534 535#ifdef ADAPTIVE_MUTEXES 536 /* 537 * The current lock owner might have started executing 538 * on another CPU (or the lock could have changed 539 * owners) while we were waiting on the turnstile 540 * chain lock. If so, drop the turnstile lock and try 541 * again. 542 */ 543 owner = (struct thread *)(v & ~MTX_FLAGMASK); 544 if (TD_IS_RUNNING(owner)) { 545 turnstile_cancel(ts); 546 continue; 547 } 548#endif 549 550 /* 551 * If the mutex isn't already contested and a failure occurs 552 * setting the contested bit, the mutex was either released 553 * or the state of the MTX_RECURSED bit changed. 554 */ 555 if ((v & MTX_CONTESTED) == 0 && 556 !atomic_cmpset_ptr(&m->mtx_lock, v, v | MTX_CONTESTED)) { 557 turnstile_cancel(ts); 558 continue; 559 } 560 561 /* 562 * We definitely must sleep for this lock. 563 */ 564 mtx_assert(m, MA_NOTOWNED); 565 566#ifdef KTR 567 if (!cont_logged) { 568 CTR6(KTR_CONTENTION, 569 "contention: %p at %s:%d wants %s, taken by %s:%d", 570 (void *)tid, file, line, m->lock_object.lo_name, 571 WITNESS_FILE(&m->lock_object), 572 WITNESS_LINE(&m->lock_object)); 573 cont_logged = 1; 574 } 575#endif 576 577 /* 578 * Block on the turnstile. 579 */ 580#ifdef KDTRACE_HOOKS 581 sleep_time -= lockstat_nsecs(&m->lock_object); 582#endif 583 turnstile_wait(ts, mtx_owner(m), TS_EXCLUSIVE_QUEUE); 584#ifdef KDTRACE_HOOKS 585 sleep_time += lockstat_nsecs(&m->lock_object); 586 sleep_cnt++; 587#endif 588 } 589#ifdef KDTRACE_HOOKS 590 all_time += lockstat_nsecs(&m->lock_object); 591#endif 592#ifdef KTR 593 if (cont_logged) { 594 CTR4(KTR_CONTENTION, 595 "contention end: %s acquired by %p at %s:%d", 596 m->lock_object.lo_name, (void *)tid, file, line); 597 } 598#endif 599 LOCKSTAT_PROFILE_OBTAIN_LOCK_SUCCESS(adaptive__acquire, m, contested, 600 waittime, file, line); 601#ifdef KDTRACE_HOOKS 602 if (sleep_time) 603 LOCKSTAT_RECORD1(adaptive__block, m, sleep_time); 604 605 /* 606 * Only record the loops spinning and not sleeping. 607 */ 608 if (lda.spin_cnt > sleep_cnt) 609 LOCKSTAT_RECORD1(adaptive__spin, m, all_time - sleep_time); 610#endif 611} 612 613static void 614_mtx_lock_spin_failed(struct mtx *m) 615{ 616 struct thread *td; 617 618 td = mtx_owner(m); 619 620 /* If the mutex is unlocked, try again. */ 621 if (td == NULL) 622 return; 623 624 printf( "spin lock %p (%s) held by %p (tid %d) too long\n", 625 m, m->lock_object.lo_name, td, td->td_tid); 626#ifdef WITNESS 627 witness_display_spinlock(&m->lock_object, td, printf); 628#endif 629 panic("spin lock held too long"); 630} 631 632#ifdef SMP 633/* 634 * _mtx_lock_spin_cookie: the tougher part of acquiring an MTX_SPIN lock. 635 * 636 * This is only called if we need to actually spin for the lock. Recursion 637 * is handled inline. 638 */ 639void 640_mtx_lock_spin_cookie(volatile uintptr_t *c, uintptr_t tid, int opts, 641 const char *file, int line) 642{ 643 struct mtx *m; 644 int i = 0; 645#ifdef LOCK_PROFILING 646 int contested = 0; 647 uint64_t waittime = 0; 648#endif 649#ifdef KDTRACE_HOOKS 650 int64_t spin_time = 0; 651#endif 652 653 if (SCHEDULER_STOPPED()) 654 return; 655 656 m = mtxlock2mtx(c); 657 658 if (LOCK_LOG_TEST(&m->lock_object, opts)) 659 CTR1(KTR_LOCK, "_mtx_lock_spin: %p spinning", m); 660 KTR_STATE1(KTR_SCHED, "thread", sched_tdname((struct thread *)tid), 661 "spinning", "lockname:\"%s\"", m->lock_object.lo_name); 662 663#ifdef HWPMC_HOOKS 664 PMC_SOFT_CALL( , , lock, failed); 665#endif 666 lock_profile_obtain_lock_failed(&m->lock_object, &contested, &waittime); 667#ifdef KDTRACE_HOOKS 668 spin_time -= lockstat_nsecs(&m->lock_object); 669#endif 670 for (;;) { 671 if (m->mtx_lock == MTX_UNOWNED && _mtx_obtain_lock(m, tid)) 672 break; 673 /* Give interrupts a chance while we spin. */ 674 spinlock_exit(); 675 while (m->mtx_lock != MTX_UNOWNED) { 676 if (i++ < 10000000) { 677 cpu_spinwait(); 678 continue; 679 } 680 if (i < 60000000 || kdb_active || panicstr != NULL) 681 DELAY(1); 682 else 683 _mtx_lock_spin_failed(m); 684 cpu_spinwait(); 685 } 686 spinlock_enter(); 687 } 688#ifdef KDTRACE_HOOKS 689 spin_time += lockstat_nsecs(&m->lock_object); 690#endif 691 692 if (LOCK_LOG_TEST(&m->lock_object, opts)) 693 CTR1(KTR_LOCK, "_mtx_lock_spin: %p spin done", m); 694 KTR_STATE0(KTR_SCHED, "thread", sched_tdname((struct thread *)tid), 695 "running"); 696 697#ifdef KDTRACE_HOOKS 698 LOCKSTAT_PROFILE_OBTAIN_LOCK_SUCCESS(spin__acquire, m, 699 contested, waittime, file, line); 700 if (spin_time != 0) 701 LOCKSTAT_RECORD1(spin__spin, m, spin_time); 702#endif 703} 704#endif /* SMP */ 705 706void 707thread_lock_flags_(struct thread *td, int opts, const char *file, int line) 708{ 709 struct mtx *m; 710 uintptr_t tid; 711 int i; 712#ifdef LOCK_PROFILING 713 int contested = 0; 714 uint64_t waittime = 0; 715#endif 716#ifdef KDTRACE_HOOKS 717 int64_t spin_time = 0; 718#endif 719 720 i = 0; 721 tid = (uintptr_t)curthread; 722 723 if (SCHEDULER_STOPPED()) { 724 /* 725 * Ensure that spinlock sections are balanced even when the 726 * scheduler is stopped, since we may otherwise inadvertently 727 * re-enable interrupts while dumping core. 728 */ 729 spinlock_enter(); 730 return; 731 } 732 733#ifdef KDTRACE_HOOKS 734 spin_time -= lockstat_nsecs(&td->td_lock->lock_object); 735#endif 736 for (;;) { 737retry: 738 spinlock_enter(); 739 m = td->td_lock; 740 KASSERT(m->mtx_lock != MTX_DESTROYED, 741 ("thread_lock() of destroyed mutex @ %s:%d", file, line)); 742 KASSERT(LOCK_CLASS(&m->lock_object) == &lock_class_mtx_spin, 743 ("thread_lock() of sleep mutex %s @ %s:%d", 744 m->lock_object.lo_name, file, line)); 745 if (mtx_owned(m)) 746 KASSERT((m->lock_object.lo_flags & LO_RECURSABLE) != 0, 747 ("thread_lock: recursed on non-recursive mutex %s @ %s:%d\n", 748 m->lock_object.lo_name, file, line)); 749 WITNESS_CHECKORDER(&m->lock_object, 750 opts | LOP_NEWORDER | LOP_EXCLUSIVE, file, line, NULL); 751 for (;;) { 752 if (m->mtx_lock == MTX_UNOWNED && _mtx_obtain_lock(m, tid)) 753 break; 754 if (m->mtx_lock == tid) { 755 m->mtx_recurse++; 756 break; 757 } 758#ifdef HWPMC_HOOKS 759 PMC_SOFT_CALL( , , lock, failed); 760#endif 761 lock_profile_obtain_lock_failed(&m->lock_object, 762 &contested, &waittime); 763 /* Give interrupts a chance while we spin. */ 764 spinlock_exit(); 765 while (m->mtx_lock != MTX_UNOWNED) { 766 if (i++ < 10000000) 767 cpu_spinwait(); 768 else if (i < 60000000 || 769 kdb_active || panicstr != NULL) 770 DELAY(1); 771 else 772 _mtx_lock_spin_failed(m); 773 cpu_spinwait(); 774 if (m != td->td_lock) 775 goto retry; 776 } 777 spinlock_enter(); 778 } 779 if (m == td->td_lock) 780 break; 781 __mtx_unlock_spin(m); /* does spinlock_exit() */ 782 } 783#ifdef KDTRACE_HOOKS 784 spin_time += lockstat_nsecs(&m->lock_object); 785#endif 786 if (m->mtx_recurse == 0) 787 LOCKSTAT_PROFILE_OBTAIN_LOCK_SUCCESS(spin__acquire, m, 788 contested, waittime, file, line); 789 LOCK_LOG_LOCK("LOCK", &m->lock_object, opts, m->mtx_recurse, file, 790 line); 791 WITNESS_LOCK(&m->lock_object, opts | LOP_EXCLUSIVE, file, line); 792#ifdef KDTRACE_HOOKS 793 if (spin_time != 0) 794 LOCKSTAT_RECORD1(thread__spin, m, spin_time); 795#endif 796} 797 798struct mtx * 799thread_lock_block(struct thread *td) 800{ 801 struct mtx *lock; 802 803 THREAD_LOCK_ASSERT(td, MA_OWNED); 804 lock = td->td_lock; 805 td->td_lock = &blocked_lock; 806 mtx_unlock_spin(lock); 807 808 return (lock); 809} 810 811void 812thread_lock_unblock(struct thread *td, struct mtx *new) 813{ 814 mtx_assert(new, MA_OWNED); 815 MPASS(td->td_lock == &blocked_lock); 816 atomic_store_rel_ptr((volatile void *)&td->td_lock, (uintptr_t)new); 817} 818 819void 820thread_lock_set(struct thread *td, struct mtx *new) 821{ 822 struct mtx *lock; 823 824 mtx_assert(new, MA_OWNED); 825 THREAD_LOCK_ASSERT(td, MA_OWNED); 826 lock = td->td_lock; 827 td->td_lock = new; 828 mtx_unlock_spin(lock); 829} 830 831/* 832 * __mtx_unlock_sleep: the tougher part of releasing an MTX_DEF lock. 833 * 834 * We are only called here if the lock is recursed or contested (i.e. we 835 * need to wake up a blocked thread). 836 */ 837void 838__mtx_unlock_sleep(volatile uintptr_t *c, int opts, const char *file, int line) 839{ 840 struct mtx *m; 841 struct turnstile *ts; 842 843 if (SCHEDULER_STOPPED()) 844 return; 845 846 m = mtxlock2mtx(c); 847 848 if (mtx_recursed(m)) { 849 if (--(m->mtx_recurse) == 0) 850 atomic_clear_ptr(&m->mtx_lock, MTX_RECURSED); 851 if (LOCK_LOG_TEST(&m->lock_object, opts)) 852 CTR1(KTR_LOCK, "_mtx_unlock_sleep: %p unrecurse", m); 853 return; 854 } 855 856 /* 857 * We have to lock the chain before the turnstile so this turnstile 858 * can be removed from the hash list if it is empty. 859 */ 860 turnstile_chain_lock(&m->lock_object); 861 ts = turnstile_lookup(&m->lock_object); 862 if (LOCK_LOG_TEST(&m->lock_object, opts)) 863 CTR1(KTR_LOCK, "_mtx_unlock_sleep: %p contested", m); 864 MPASS(ts != NULL); 865 turnstile_broadcast(ts, TS_EXCLUSIVE_QUEUE); 866 _mtx_release_lock_quick(m); 867 868 /* 869 * This turnstile is now no longer associated with the mutex. We can 870 * unlock the chain lock so a new turnstile may take it's place. 871 */ 872 turnstile_unpend(ts, TS_EXCLUSIVE_LOCK); 873 turnstile_chain_unlock(&m->lock_object); 874} 875 876/* 877 * All the unlocking of MTX_SPIN locks is done inline. 878 * See the __mtx_unlock_spin() macro for the details. 879 */ 880 881/* 882 * The backing function for the INVARIANTS-enabled mtx_assert() 883 */ 884#ifdef INVARIANT_SUPPORT 885void 886__mtx_assert(const volatile uintptr_t *c, int what, const char *file, int line) 887{ 888 const struct mtx *m; 889 890 if (panicstr != NULL || dumping) 891 return; 892 893 m = mtxlock2mtx(c); 894 895 switch (what) { 896 case MA_OWNED: 897 case MA_OWNED | MA_RECURSED: 898 case MA_OWNED | MA_NOTRECURSED: 899 if (!mtx_owned(m)) 900 panic("mutex %s not owned at %s:%d", 901 m->lock_object.lo_name, file, line); 902 if (mtx_recursed(m)) { 903 if ((what & MA_NOTRECURSED) != 0) 904 panic("mutex %s recursed at %s:%d", 905 m->lock_object.lo_name, file, line); 906 } else if ((what & MA_RECURSED) != 0) { 907 panic("mutex %s unrecursed at %s:%d", 908 m->lock_object.lo_name, file, line); 909 } 910 break; 911 case MA_NOTOWNED: 912 if (mtx_owned(m)) 913 panic("mutex %s owned at %s:%d", 914 m->lock_object.lo_name, file, line); 915 break; 916 default: 917 panic("unknown mtx_assert at %s:%d", file, line); 918 } 919} 920#endif 921 922/* 923 * General init routine used by the MTX_SYSINIT() macro. 924 */ 925void 926mtx_sysinit(void *arg) 927{ 928 struct mtx_args *margs = arg; 929 930 mtx_init((struct mtx *)margs->ma_mtx, margs->ma_desc, NULL, 931 margs->ma_opts); 932} 933 934/* 935 * Mutex initialization routine; initialize lock `m' of type contained in 936 * `opts' with options contained in `opts' and name `name.' The optional 937 * lock type `type' is used as a general lock category name for use with 938 * witness. 939 */ 940void 941_mtx_init(volatile uintptr_t *c, const char *name, const char *type, int opts) 942{ 943 struct mtx *m; 944 struct lock_class *class; 945 int flags; 946 947 m = mtxlock2mtx(c); 948 949 MPASS((opts & ~(MTX_SPIN | MTX_QUIET | MTX_RECURSE | 950 MTX_NOWITNESS | MTX_DUPOK | MTX_NOPROFILE | MTX_NEW)) == 0); 951 ASSERT_ATOMIC_LOAD_PTR(m->mtx_lock, 952 ("%s: mtx_lock not aligned for %s: %p", __func__, name, 953 &m->mtx_lock)); 954 955 /* Determine lock class and lock flags. */ 956 if (opts & MTX_SPIN) 957 class = &lock_class_mtx_spin; 958 else 959 class = &lock_class_mtx_sleep; 960 flags = 0; 961 if (opts & MTX_QUIET) 962 flags |= LO_QUIET; 963 if (opts & MTX_RECURSE) 964 flags |= LO_RECURSABLE; 965 if ((opts & MTX_NOWITNESS) == 0) 966 flags |= LO_WITNESS; 967 if (opts & MTX_DUPOK) 968 flags |= LO_DUPOK; 969 if (opts & MTX_NOPROFILE) 970 flags |= LO_NOPROFILE; 971 if (opts & MTX_NEW) 972 flags |= LO_NEW; 973 974 /* Initialize mutex. */ 975 lock_init(&m->lock_object, class, name, type, flags); 976 977 m->mtx_lock = MTX_UNOWNED; 978 m->mtx_recurse = 0; 979} 980 981/* 982 * Remove lock `m' from all_mtx queue. We don't allow MTX_QUIET to be 983 * passed in as a flag here because if the corresponding mtx_init() was 984 * called with MTX_QUIET set, then it will already be set in the mutex's 985 * flags. 986 */ 987void 988_mtx_destroy(volatile uintptr_t *c) 989{ 990 struct mtx *m; 991 992 m = mtxlock2mtx(c); 993 994 if (!mtx_owned(m)) 995 MPASS(mtx_unowned(m)); 996 else { 997 MPASS((m->mtx_lock & (MTX_RECURSED|MTX_CONTESTED)) == 0); 998 999 /* Perform the non-mtx related part of mtx_unlock_spin(). */ 1000 if (LOCK_CLASS(&m->lock_object) == &lock_class_mtx_spin) 1001 spinlock_exit(); 1002 else 1003 TD_LOCKS_DEC(curthread); 1004 1005 lock_profile_release_lock(&m->lock_object); 1006 /* Tell witness this isn't locked to make it happy. */ 1007 WITNESS_UNLOCK(&m->lock_object, LOP_EXCLUSIVE, __FILE__, 1008 __LINE__); 1009 } 1010 1011 m->mtx_lock = MTX_DESTROYED; 1012 lock_destroy(&m->lock_object); 1013} 1014 1015/* 1016 * Intialize the mutex code and system mutexes. This is called from the MD 1017 * startup code prior to mi_startup(). The per-CPU data space needs to be 1018 * setup before this is called. 1019 */ 1020void 1021mutex_init(void) 1022{ 1023 1024 /* Setup turnstiles so that sleep mutexes work. */ 1025 init_turnstiles(); 1026 1027 /* 1028 * Initialize mutexes. 1029 */ 1030 mtx_init(&Giant, "Giant", NULL, MTX_DEF | MTX_RECURSE); 1031 mtx_init(&blocked_lock, "blocked lock", NULL, MTX_SPIN); 1032 blocked_lock.mtx_lock = 0xdeadc0de; /* Always blocked. */ 1033 mtx_init(&proc0.p_mtx, "process lock", NULL, MTX_DEF | MTX_DUPOK); 1034 mtx_init(&proc0.p_slock, "process slock", NULL, MTX_SPIN); 1035 mtx_init(&proc0.p_statmtx, "pstatl", NULL, MTX_SPIN); 1036 mtx_init(&proc0.p_itimmtx, "pitiml", NULL, MTX_SPIN); 1037 mtx_init(&proc0.p_profmtx, "pprofl", NULL, MTX_SPIN); 1038 mtx_init(&devmtx, "cdev", NULL, MTX_DEF); 1039 mtx_lock(&Giant); 1040} 1041 1042#ifdef DDB 1043void 1044db_show_mtx(const struct lock_object *lock) 1045{ 1046 struct thread *td; 1047 const struct mtx *m; 1048 1049 m = (const struct mtx *)lock; 1050 1051 db_printf(" flags: {"); 1052 if (LOCK_CLASS(lock) == &lock_class_mtx_spin) 1053 db_printf("SPIN"); 1054 else 1055 db_printf("DEF"); 1056 if (m->lock_object.lo_flags & LO_RECURSABLE) 1057 db_printf(", RECURSE"); 1058 if (m->lock_object.lo_flags & LO_DUPOK) 1059 db_printf(", DUPOK"); 1060 db_printf("}\n"); 1061 db_printf(" state: {"); 1062 if (mtx_unowned(m)) 1063 db_printf("UNOWNED"); 1064 else if (mtx_destroyed(m)) 1065 db_printf("DESTROYED"); 1066 else { 1067 db_printf("OWNED"); 1068 if (m->mtx_lock & MTX_CONTESTED) 1069 db_printf(", CONTESTED"); 1070 if (m->mtx_lock & MTX_RECURSED) 1071 db_printf(", RECURSED"); 1072 } 1073 db_printf("}\n"); 1074 if (!mtx_unowned(m) && !mtx_destroyed(m)) { 1075 td = mtx_owner(m); 1076 db_printf(" owner: %p (tid %d, pid %d, \"%s\")\n", td, 1077 td->td_tid, td->td_proc->p_pid, td->td_name); 1078 if (mtx_recursed(m)) 1079 db_printf(" recursed: %d\n", m->mtx_recurse); 1080 } 1081} 1082#endif 1083