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_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> 81 PMC_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 99 static void assert_mtx(const struct lock_object *lock, int what); 100 #ifdef DDB 101 static void db_show_mtx(const struct lock_object *lock); 102 #endif 103 static void lock_mtx(struct lock_object *lock, uintptr_t how); 104 static void lock_spin(struct lock_object *lock, uintptr_t how); 105 #ifdef KDTRACE_HOOKS 106 static int owner_mtx(const struct lock_object *lock, 107 struct thread **owner); 108 #endif 109 static uintptr_t unlock_mtx(struct lock_object *lock); 110 static uintptr_t unlock_spin(struct lock_object *lock); 111 112 /* 113 * Lock classes for sleep and spin mutexes. 114 */ 115 struct 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 }; 128 struct 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 143 static SYSCTL_NODE(_debug, OID_AUTO, mtx, CTLFLAG_RD, NULL, "mtx debugging"); 144 145 static struct lock_delay_config mtx_delay = { 146 .initial = 1000, 147 .step = 500, 148 .min = 100, 149 .max = 5000, 150 }; 151 152 SYSCTL_INT(_debug_mtx, OID_AUTO, delay_initial, CTLFLAG_RW, &mtx_delay.initial, 153 0, ""); 154 SYSCTL_INT(_debug_mtx, OID_AUTO, delay_step, CTLFLAG_RW, &mtx_delay.step, 155 0, ""); 156 SYSCTL_INT(_debug_mtx, OID_AUTO, delay_min, CTLFLAG_RW, &mtx_delay.min, 157 0, ""); 158 SYSCTL_INT(_debug_mtx, OID_AUTO, delay_max, CTLFLAG_RW, &mtx_delay.max, 159 0, ""); 160 161 static void 162 mtx_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 } 170 LOCK_DELAY_SYSINIT(mtx_delay_sysinit); 171 #endif 172 173 static SYSCTL_NODE(_debug, OID_AUTO, mtx_spin, CTLFLAG_RD, NULL, 174 "mtx spin debugging"); 175 176 static struct lock_delay_config mtx_spin_delay = { 177 .initial = 1000, 178 .step = 500, 179 .min = 100, 180 .max = 5000, 181 }; 182 183 SYSCTL_INT(_debug_mtx_spin, OID_AUTO, delay_initial, CTLFLAG_RW, 184 &mtx_spin_delay.initial, 0, ""); 185 SYSCTL_INT(_debug_mtx_spin, OID_AUTO, delay_step, CTLFLAG_RW, &mtx_spin_delay.step, 186 0, ""); 187 SYSCTL_INT(_debug_mtx_spin, OID_AUTO, delay_min, CTLFLAG_RW, &mtx_spin_delay.min, 188 0, ""); 189 SYSCTL_INT(_debug_mtx_spin, OID_AUTO, delay_max, CTLFLAG_RW, &mtx_spin_delay.max, 190 0, ""); 191 192 static void 193 mtx_spin_delay_sysinit(void *dummy) 194 { 195 196 mtx_spin_delay.initial = mp_ncpus * 25; 197 mtx_spin_delay.step = (mp_ncpus * 25) / 2; 198 mtx_spin_delay.min = mp_ncpus * 5; 199 mtx_spin_delay.max = mp_ncpus * 25 * 10; 200 } 201 LOCK_DELAY_SYSINIT(mtx_spin_delay_sysinit); 202 203 /* 204 * System-wide mutexes 205 */ 206 struct mtx blocked_lock; 207 struct mtx Giant; 208 209 void 210 assert_mtx(const struct lock_object *lock, int what) 211 { 212 213 mtx_assert((const struct mtx *)lock, what); 214 } 215 216 void 217 lock_mtx(struct lock_object *lock, uintptr_t how) 218 { 219 220 mtx_lock((struct mtx *)lock); 221 } 222 223 void 224 lock_spin(struct lock_object *lock, uintptr_t how) 225 { 226 227 panic("spin locks can only use msleep_spin"); 228 } 229 230 uintptr_t 231 unlock_mtx(struct lock_object *lock) 232 { 233 struct mtx *m; 234 235 m = (struct mtx *)lock; 236 mtx_assert(m, MA_OWNED | MA_NOTRECURSED); 237 mtx_unlock(m); 238 return (0); 239 } 240 241 uintptr_t 242 unlock_spin(struct lock_object *lock) 243 { 244 245 panic("spin locks can only use msleep_spin"); 246 } 247 248 #ifdef KDTRACE_HOOKS 249 int 250 owner_mtx(const struct lock_object *lock, struct thread **owner) 251 { 252 const struct mtx *m; 253 uintptr_t x; 254 255 m = (const struct mtx *)lock; 256 x = m->mtx_lock; 257 *owner = (struct thread *)(x & ~MTX_FLAGMASK); 258 return (x != MTX_UNOWNED); 259 } 260 #endif 261 262 /* 263 * Function versions of the inlined __mtx_* macros. These are used by 264 * modules and can also be called from assembly language if needed. 265 */ 266 void 267 __mtx_lock_flags(volatile uintptr_t *c, int opts, const char *file, int line) 268 { 269 struct mtx *m; 270 271 if (SCHEDULER_STOPPED()) 272 return; 273 274 m = mtxlock2mtx(c); 275 276 KASSERT(kdb_active != 0 || !TD_IS_IDLETHREAD(curthread), 277 ("mtx_lock() by idle thread %p on sleep mutex %s @ %s:%d", 278 curthread, m->lock_object.lo_name, file, line)); 279 KASSERT(m->mtx_lock != MTX_DESTROYED, 280 ("mtx_lock() of destroyed mutex @ %s:%d", file, line)); 281 KASSERT(LOCK_CLASS(&m->lock_object) == &lock_class_mtx_sleep, 282 ("mtx_lock() of spin mutex %s @ %s:%d", m->lock_object.lo_name, 283 file, line)); 284 WITNESS_CHECKORDER(&m->lock_object, (opts & ~MTX_RECURSE) | 285 LOP_NEWORDER | LOP_EXCLUSIVE, file, line, NULL); 286 287 __mtx_lock(m, curthread, opts, file, line); 288 LOCK_LOG_LOCK("LOCK", &m->lock_object, opts, m->mtx_recurse, file, 289 line); 290 WITNESS_LOCK(&m->lock_object, (opts & ~MTX_RECURSE) | LOP_EXCLUSIVE, 291 file, line); 292 TD_LOCKS_INC(curthread); 293 } 294 295 void 296 __mtx_unlock_flags(volatile uintptr_t *c, int opts, const char *file, int line) 297 { 298 struct mtx *m; 299 300 if (SCHEDULER_STOPPED()) 301 return; 302 303 m = mtxlock2mtx(c); 304 305 KASSERT(m->mtx_lock != MTX_DESTROYED, 306 ("mtx_unlock() of destroyed mutex @ %s:%d", file, line)); 307 KASSERT(LOCK_CLASS(&m->lock_object) == &lock_class_mtx_sleep, 308 ("mtx_unlock() of spin mutex %s @ %s:%d", m->lock_object.lo_name, 309 file, line)); 310 WITNESS_UNLOCK(&m->lock_object, opts | LOP_EXCLUSIVE, file, line); 311 LOCK_LOG_LOCK("UNLOCK", &m->lock_object, opts, m->mtx_recurse, file, 312 line); 313 mtx_assert(m, MA_OWNED); 314 315 __mtx_unlock(m, curthread, opts, file, line); 316 TD_LOCKS_DEC(curthread); 317 } 318 319 void 320 __mtx_lock_spin_flags(volatile uintptr_t *c, int opts, const char *file, 321 int line) 322 { 323 struct mtx *m; 324 325 if (SCHEDULER_STOPPED()) 326 return; 327 328 m = mtxlock2mtx(c); 329 330 KASSERT(m->mtx_lock != MTX_DESTROYED, 331 ("mtx_lock_spin() of destroyed mutex @ %s:%d", file, line)); 332 KASSERT(LOCK_CLASS(&m->lock_object) == &lock_class_mtx_spin, 333 ("mtx_lock_spin() of sleep mutex %s @ %s:%d", 334 m->lock_object.lo_name, file, line)); 335 if (mtx_owned(m)) 336 KASSERT((m->lock_object.lo_flags & LO_RECURSABLE) != 0 || 337 (opts & MTX_RECURSE) != 0, 338 ("mtx_lock_spin: recursed on non-recursive mutex %s @ %s:%d\n", 339 m->lock_object.lo_name, file, line)); 340 opts &= ~MTX_RECURSE; 341 WITNESS_CHECKORDER(&m->lock_object, opts | LOP_NEWORDER | LOP_EXCLUSIVE, 342 file, line, NULL); 343 __mtx_lock_spin(m, curthread, opts, file, line); 344 LOCK_LOG_LOCK("LOCK", &m->lock_object, opts, m->mtx_recurse, file, 345 line); 346 WITNESS_LOCK(&m->lock_object, opts | LOP_EXCLUSIVE, file, line); 347 } 348 349 int 350 __mtx_trylock_spin_flags(volatile uintptr_t *c, int opts, const char *file, 351 int line) 352 { 353 struct mtx *m; 354 355 if (SCHEDULER_STOPPED()) 356 return (1); 357 358 m = mtxlock2mtx(c); 359 360 KASSERT(m->mtx_lock != MTX_DESTROYED, 361 ("mtx_trylock_spin() of destroyed mutex @ %s:%d", file, line)); 362 KASSERT(LOCK_CLASS(&m->lock_object) == &lock_class_mtx_spin, 363 ("mtx_trylock_spin() of sleep mutex %s @ %s:%d", 364 m->lock_object.lo_name, file, line)); 365 KASSERT((opts & MTX_RECURSE) == 0, 366 ("mtx_trylock_spin: unsupp. opt MTX_RECURSE on mutex %s @ %s:%d\n", 367 m->lock_object.lo_name, file, line)); 368 if (__mtx_trylock_spin(m, curthread, opts, file, line)) { 369 LOCK_LOG_TRY("LOCK", &m->lock_object, opts, 1, file, line); 370 WITNESS_LOCK(&m->lock_object, opts | LOP_EXCLUSIVE, file, line); 371 return (1); 372 } 373 LOCK_LOG_TRY("LOCK", &m->lock_object, opts, 0, file, line); 374 return (0); 375 } 376 377 void 378 __mtx_unlock_spin_flags(volatile uintptr_t *c, int opts, const char *file, 379 int line) 380 { 381 struct mtx *m; 382 383 if (SCHEDULER_STOPPED()) 384 return; 385 386 m = mtxlock2mtx(c); 387 388 KASSERT(m->mtx_lock != MTX_DESTROYED, 389 ("mtx_unlock_spin() of destroyed mutex @ %s:%d", file, line)); 390 KASSERT(LOCK_CLASS(&m->lock_object) == &lock_class_mtx_spin, 391 ("mtx_unlock_spin() of sleep mutex %s @ %s:%d", 392 m->lock_object.lo_name, file, line)); 393 WITNESS_UNLOCK(&m->lock_object, opts | LOP_EXCLUSIVE, file, line); 394 LOCK_LOG_LOCK("UNLOCK", &m->lock_object, opts, m->mtx_recurse, file, 395 line); 396 mtx_assert(m, MA_OWNED); 397 398 __mtx_unlock_spin(m); 399 } 400 401 /* 402 * The important part of mtx_trylock{,_flags}() 403 * Tries to acquire lock `m.' If this function is called on a mutex that 404 * is already owned, it will recursively acquire the lock. 405 */ 406 int 407 _mtx_trylock_flags_(volatile uintptr_t *c, int opts, const char *file, int line) 408 { 409 struct mtx *m; 410 #ifdef LOCK_PROFILING 411 uint64_t waittime = 0; 412 int contested = 0; 413 #endif 414 int rval; 415 416 if (SCHEDULER_STOPPED()) 417 return (1); 418 419 m = mtxlock2mtx(c); 420 421 KASSERT(kdb_active != 0 || !TD_IS_IDLETHREAD(curthread), 422 ("mtx_trylock() by idle thread %p on sleep mutex %s @ %s:%d", 423 curthread, m->lock_object.lo_name, file, line)); 424 KASSERT(m->mtx_lock != MTX_DESTROYED, 425 ("mtx_trylock() of destroyed mutex @ %s:%d", file, line)); 426 KASSERT(LOCK_CLASS(&m->lock_object) == &lock_class_mtx_sleep, 427 ("mtx_trylock() of spin mutex %s @ %s:%d", m->lock_object.lo_name, 428 file, line)); 429 430 if (mtx_owned(m) && ((m->lock_object.lo_flags & LO_RECURSABLE) != 0 || 431 (opts & MTX_RECURSE) != 0)) { 432 m->mtx_recurse++; 433 atomic_set_ptr(&m->mtx_lock, MTX_RECURSED); 434 rval = 1; 435 } else 436 rval = _mtx_obtain_lock(m, (uintptr_t)curthread); 437 opts &= ~MTX_RECURSE; 438 439 LOCK_LOG_TRY("LOCK", &m->lock_object, opts, rval, file, line); 440 if (rval) { 441 WITNESS_LOCK(&m->lock_object, opts | LOP_EXCLUSIVE | LOP_TRYLOCK, 442 file, line); 443 TD_LOCKS_INC(curthread); 444 if (m->mtx_recurse == 0) 445 LOCKSTAT_PROFILE_OBTAIN_LOCK_SUCCESS(adaptive__acquire, 446 m, contested, waittime, file, line); 447 448 } 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(volatile uintptr_t *c, uintptr_t tid, int opts, 461 const char *file, int line) 462 { 463 struct mtx *m; 464 struct turnstile *ts; 465 uintptr_t v; 466 #ifdef ADAPTIVE_MUTEXES 467 volatile struct thread *owner; 468 #endif 469 #ifdef KTR 470 int cont_logged = 0; 471 #endif 472 #ifdef LOCK_PROFILING 473 int contested = 0; 474 uint64_t waittime = 0; 475 #endif 476 #if defined(ADAPTIVE_MUTEXES) || defined(KDTRACE_HOOKS) 477 struct lock_delay_arg lda; 478 #endif 479 #ifdef KDTRACE_HOOKS 480 u_int sleep_cnt = 0; 481 int64_t sleep_time = 0; 482 int64_t all_time = 0; 483 #endif 484 485 if (SCHEDULER_STOPPED()) 486 return; 487 488 #if defined(ADAPTIVE_MUTEXES) 489 lock_delay_arg_init(&lda, &mtx_delay); 490 #elif defined(KDTRACE_HOOKS) 491 lock_delay_arg_init(&lda, NULL); 492 #endif 493 m = mtxlock2mtx(c); 494 495 if (mtx_owned(m)) { 496 KASSERT((m->lock_object.lo_flags & LO_RECURSABLE) != 0 || 497 (opts & MTX_RECURSE) != 0, 498 ("_mtx_lock_sleep: recursed on non-recursive mutex %s @ %s:%d\n", 499 m->lock_object.lo_name, file, line)); 500 opts &= ~MTX_RECURSE; 501 m->mtx_recurse++; 502 atomic_set_ptr(&m->mtx_lock, MTX_RECURSED); 503 if (LOCK_LOG_TEST(&m->lock_object, opts)) 504 CTR1(KTR_LOCK, "_mtx_lock_sleep: %p recursing", m); 505 return; 506 } 507 opts &= ~MTX_RECURSE; 508 509 #ifdef HWPMC_HOOKS 510 PMC_SOFT_CALL( , , lock, failed); 511 #endif 512 lock_profile_obtain_lock_failed(&m->lock_object, 513 &contested, &waittime); 514 if (LOCK_LOG_TEST(&m->lock_object, opts)) 515 CTR4(KTR_LOCK, 516 "_mtx_lock_sleep: %s contested (lock=%p) at %s:%d", 517 m->lock_object.lo_name, (void *)m->mtx_lock, file, line); 518 #ifdef KDTRACE_HOOKS 519 all_time -= lockstat_nsecs(&m->lock_object); 520 #endif 521 522 for (;;) { 523 if (m->mtx_lock == MTX_UNOWNED && _mtx_obtain_lock(m, tid)) 524 break; 525 #ifdef KDTRACE_HOOKS 526 lda.spin_cnt++; 527 #endif 528 #ifdef ADAPTIVE_MUTEXES 529 /* 530 * If the owner is running on another CPU, spin until the 531 * owner stops running or the state of the lock changes. 532 */ 533 v = m->mtx_lock; 534 if (v != MTX_UNOWNED) { 535 owner = (struct thread *)(v & ~MTX_FLAGMASK); 536 if (TD_IS_RUNNING(owner)) { 537 if (LOCK_LOG_TEST(&m->lock_object, 0)) 538 CTR3(KTR_LOCK, 539 "%s: spinning on %p held by %p", 540 __func__, m, owner); 541 KTR_STATE1(KTR_SCHED, "thread", 542 sched_tdname((struct thread *)tid), 543 "spinning", "lockname:\"%s\"", 544 m->lock_object.lo_name); 545 while (mtx_owner(m) == owner && 546 TD_IS_RUNNING(owner)) 547 lock_delay(&lda); 548 KTR_STATE0(KTR_SCHED, "thread", 549 sched_tdname((struct thread *)tid), 550 "running"); 551 continue; 552 } 553 } 554 #endif 555 556 ts = turnstile_trywait(&m->lock_object); 557 v = m->mtx_lock; 558 559 /* 560 * Check if the lock has been released while spinning for 561 * the turnstile chain lock. 562 */ 563 if (v == MTX_UNOWNED) { 564 turnstile_cancel(ts); 565 continue; 566 } 567 568 #ifdef ADAPTIVE_MUTEXES 569 /* 570 * The current lock owner might have started executing 571 * on another CPU (or the lock could have changed 572 * owners) while we were waiting on the turnstile 573 * chain lock. If so, drop the turnstile lock and try 574 * again. 575 */ 576 owner = (struct thread *)(v & ~MTX_FLAGMASK); 577 if (TD_IS_RUNNING(owner)) { 578 turnstile_cancel(ts); 579 continue; 580 } 581 #endif 582 583 /* 584 * If the mutex isn't already contested and a failure occurs 585 * setting the contested bit, the mutex was either released 586 * or the state of the MTX_RECURSED bit changed. 587 */ 588 if ((v & MTX_CONTESTED) == 0 && 589 !atomic_cmpset_ptr(&m->mtx_lock, v, v | MTX_CONTESTED)) { 590 turnstile_cancel(ts); 591 continue; 592 } 593 594 /* 595 * We definitely must sleep for this lock. 596 */ 597 mtx_assert(m, MA_NOTOWNED); 598 599 #ifdef KTR 600 if (!cont_logged) { 601 CTR6(KTR_CONTENTION, 602 "contention: %p at %s:%d wants %s, taken by %s:%d", 603 (void *)tid, file, line, m->lock_object.lo_name, 604 WITNESS_FILE(&m->lock_object), 605 WITNESS_LINE(&m->lock_object)); 606 cont_logged = 1; 607 } 608 #endif 609 610 /* 611 * Block on the turnstile. 612 */ 613 #ifdef KDTRACE_HOOKS 614 sleep_time -= lockstat_nsecs(&m->lock_object); 615 #endif 616 turnstile_wait(ts, mtx_owner(m), TS_EXCLUSIVE_QUEUE); 617 #ifdef KDTRACE_HOOKS 618 sleep_time += lockstat_nsecs(&m->lock_object); 619 sleep_cnt++; 620 #endif 621 } 622 #ifdef KDTRACE_HOOKS 623 all_time += lockstat_nsecs(&m->lock_object); 624 #endif 625 #ifdef KTR 626 if (cont_logged) { 627 CTR4(KTR_CONTENTION, 628 "contention end: %s acquired by %p at %s:%d", 629 m->lock_object.lo_name, (void *)tid, file, line); 630 } 631 #endif 632 LOCKSTAT_PROFILE_OBTAIN_LOCK_SUCCESS(adaptive__acquire, m, contested, 633 waittime, file, line); 634 #ifdef KDTRACE_HOOKS 635 if (sleep_time) 636 LOCKSTAT_RECORD1(adaptive__block, m, sleep_time); 637 638 /* 639 * Only record the loops spinning and not sleeping. 640 */ 641 if (lda.spin_cnt > sleep_cnt) 642 LOCKSTAT_RECORD1(adaptive__spin, m, all_time - sleep_time); 643 #endif 644 } 645 646 static void 647 _mtx_lock_spin_failed(struct mtx *m) 648 { 649 struct thread *td; 650 651 td = mtx_owner(m); 652 653 /* If the mutex is unlocked, try again. */ 654 if (td == NULL) 655 return; 656 657 printf( "spin lock %p (%s) held by %p (tid %d) too long\n", 658 m, m->lock_object.lo_name, td, td->td_tid); 659 #ifdef WITNESS 660 witness_display_spinlock(&m->lock_object, td, printf); 661 #endif 662 panic("spin lock held too long"); 663 } 664 665 #ifdef SMP 666 /* 667 * _mtx_lock_spin_cookie: the tougher part of acquiring an MTX_SPIN lock. 668 * 669 * This is only called if we need to actually spin for the lock. Recursion 670 * is handled inline. 671 */ 672 void 673 _mtx_lock_spin_cookie(volatile uintptr_t *c, uintptr_t tid, int opts, 674 const char *file, int line) 675 { 676 struct mtx *m; 677 struct lock_delay_arg lda; 678 #ifdef LOCK_PROFILING 679 int contested = 0; 680 uint64_t waittime = 0; 681 #endif 682 #ifdef KDTRACE_HOOKS 683 int64_t spin_time = 0; 684 #endif 685 686 if (SCHEDULER_STOPPED()) 687 return; 688 689 lock_delay_arg_init(&lda, &mtx_spin_delay); 690 m = mtxlock2mtx(c); 691 692 if (LOCK_LOG_TEST(&m->lock_object, opts)) 693 CTR1(KTR_LOCK, "_mtx_lock_spin: %p spinning", m); 694 KTR_STATE1(KTR_SCHED, "thread", sched_tdname((struct thread *)tid), 695 "spinning", "lockname:\"%s\"", m->lock_object.lo_name); 696 697 #ifdef HWPMC_HOOKS 698 PMC_SOFT_CALL( , , lock, failed); 699 #endif 700 lock_profile_obtain_lock_failed(&m->lock_object, &contested, &waittime); 701 #ifdef KDTRACE_HOOKS 702 spin_time -= lockstat_nsecs(&m->lock_object); 703 #endif 704 for (;;) { 705 if (m->mtx_lock == MTX_UNOWNED && _mtx_obtain_lock(m, tid)) 706 break; 707 /* Give interrupts a chance while we spin. */ 708 spinlock_exit(); 709 while (m->mtx_lock != MTX_UNOWNED) { 710 if (lda.spin_cnt < 10000000) { 711 lock_delay(&lda); 712 continue; 713 } 714 lda.spin_cnt++; 715 if (lda.spin_cnt < 60000000 || kdb_active || 716 panicstr != NULL) 717 DELAY(1); 718 else 719 _mtx_lock_spin_failed(m); 720 cpu_spinwait(); 721 } 722 spinlock_enter(); 723 } 724 #ifdef KDTRACE_HOOKS 725 spin_time += lockstat_nsecs(&m->lock_object); 726 #endif 727 728 if (LOCK_LOG_TEST(&m->lock_object, opts)) 729 CTR1(KTR_LOCK, "_mtx_lock_spin: %p spin done", m); 730 KTR_STATE0(KTR_SCHED, "thread", sched_tdname((struct thread *)tid), 731 "running"); 732 733 #ifdef KDTRACE_HOOKS 734 LOCKSTAT_PROFILE_OBTAIN_LOCK_SUCCESS(spin__acquire, m, 735 contested, waittime, file, line); 736 if (spin_time != 0) 737 LOCKSTAT_RECORD1(spin__spin, m, spin_time); 738 #endif 739 } 740 #endif /* SMP */ 741 742 void 743 thread_lock_flags_(struct thread *td, int opts, const char *file, int line) 744 { 745 struct mtx *m; 746 uintptr_t tid; 747 struct lock_delay_arg lda; 748 #ifdef LOCK_PROFILING 749 int contested = 0; 750 uint64_t waittime = 0; 751 #endif 752 #ifdef KDTRACE_HOOKS 753 int64_t spin_time = 0; 754 #endif 755 756 tid = (uintptr_t)curthread; 757 758 if (SCHEDULER_STOPPED()) { 759 /* 760 * Ensure that spinlock sections are balanced even when the 761 * scheduler is stopped, since we may otherwise inadvertently 762 * re-enable interrupts while dumping core. 763 */ 764 spinlock_enter(); 765 return; 766 } 767 768 lock_delay_arg_init(&lda, &mtx_spin_delay); 769 770 #ifdef KDTRACE_HOOKS 771 spin_time -= lockstat_nsecs(&td->td_lock->lock_object); 772 #endif 773 for (;;) { 774 retry: 775 spinlock_enter(); 776 m = td->td_lock; 777 KASSERT(m->mtx_lock != MTX_DESTROYED, 778 ("thread_lock() of destroyed mutex @ %s:%d", file, line)); 779 KASSERT(LOCK_CLASS(&m->lock_object) == &lock_class_mtx_spin, 780 ("thread_lock() of sleep mutex %s @ %s:%d", 781 m->lock_object.lo_name, file, line)); 782 if (mtx_owned(m)) 783 KASSERT((m->lock_object.lo_flags & LO_RECURSABLE) != 0, 784 ("thread_lock: recursed on non-recursive mutex %s @ %s:%d\n", 785 m->lock_object.lo_name, file, line)); 786 WITNESS_CHECKORDER(&m->lock_object, 787 opts | LOP_NEWORDER | LOP_EXCLUSIVE, file, line, NULL); 788 for (;;) { 789 if (m->mtx_lock == MTX_UNOWNED && _mtx_obtain_lock(m, tid)) 790 break; 791 if (m->mtx_lock == tid) { 792 m->mtx_recurse++; 793 break; 794 } 795 #ifdef HWPMC_HOOKS 796 PMC_SOFT_CALL( , , lock, failed); 797 #endif 798 lock_profile_obtain_lock_failed(&m->lock_object, 799 &contested, &waittime); 800 /* Give interrupts a chance while we spin. */ 801 spinlock_exit(); 802 while (m->mtx_lock != MTX_UNOWNED) { 803 if (lda.spin_cnt < 10000000) { 804 lock_delay(&lda); 805 } else { 806 lda.spin_cnt++; 807 if (lda.spin_cnt < 60000000 || 808 kdb_active || panicstr != NULL) 809 DELAY(1); 810 else 811 _mtx_lock_spin_failed(m); 812 cpu_spinwait(); 813 } 814 if (m != td->td_lock) 815 goto retry; 816 } 817 spinlock_enter(); 818 } 819 if (m == td->td_lock) 820 break; 821 __mtx_unlock_spin(m); /* does spinlock_exit() */ 822 } 823 #ifdef KDTRACE_HOOKS 824 spin_time += lockstat_nsecs(&m->lock_object); 825 #endif 826 if (m->mtx_recurse == 0) 827 LOCKSTAT_PROFILE_OBTAIN_LOCK_SUCCESS(spin__acquire, m, 828 contested, waittime, file, line); 829 LOCK_LOG_LOCK("LOCK", &m->lock_object, opts, m->mtx_recurse, file, 830 line); 831 WITNESS_LOCK(&m->lock_object, opts | LOP_EXCLUSIVE, file, line); 832 #ifdef KDTRACE_HOOKS 833 if (spin_time != 0) 834 LOCKSTAT_RECORD1(thread__spin, m, spin_time); 835 #endif 836 } 837 838 struct mtx * 839 thread_lock_block(struct thread *td) 840 { 841 struct mtx *lock; 842 843 THREAD_LOCK_ASSERT(td, MA_OWNED); 844 lock = td->td_lock; 845 td->td_lock = &blocked_lock; 846 mtx_unlock_spin(lock); 847 848 return (lock); 849 } 850 851 void 852 thread_lock_unblock(struct thread *td, struct mtx *new) 853 { 854 mtx_assert(new, MA_OWNED); 855 MPASS(td->td_lock == &blocked_lock); 856 atomic_store_rel_ptr((volatile void *)&td->td_lock, (uintptr_t)new); 857 } 858 859 void 860 thread_lock_set(struct thread *td, struct mtx *new) 861 { 862 struct mtx *lock; 863 864 mtx_assert(new, MA_OWNED); 865 THREAD_LOCK_ASSERT(td, MA_OWNED); 866 lock = td->td_lock; 867 td->td_lock = new; 868 mtx_unlock_spin(lock); 869 } 870 871 /* 872 * __mtx_unlock_sleep: the tougher part of releasing an MTX_DEF lock. 873 * 874 * We are only called here if the lock is recursed or contested (i.e. we 875 * need to wake up a blocked thread). 876 */ 877 void 878 __mtx_unlock_sleep(volatile uintptr_t *c, int opts, const char *file, int line) 879 { 880 struct mtx *m; 881 struct turnstile *ts; 882 883 if (SCHEDULER_STOPPED()) 884 return; 885 886 m = mtxlock2mtx(c); 887 888 if (mtx_recursed(m)) { 889 if (--(m->mtx_recurse) == 0) 890 atomic_clear_ptr(&m->mtx_lock, MTX_RECURSED); 891 if (LOCK_LOG_TEST(&m->lock_object, opts)) 892 CTR1(KTR_LOCK, "_mtx_unlock_sleep: %p unrecurse", m); 893 return; 894 } 895 896 /* 897 * We have to lock the chain before the turnstile so this turnstile 898 * can be removed from the hash list if it is empty. 899 */ 900 turnstile_chain_lock(&m->lock_object); 901 ts = turnstile_lookup(&m->lock_object); 902 if (LOCK_LOG_TEST(&m->lock_object, opts)) 903 CTR1(KTR_LOCK, "_mtx_unlock_sleep: %p contested", m); 904 MPASS(ts != NULL); 905 turnstile_broadcast(ts, TS_EXCLUSIVE_QUEUE); 906 _mtx_release_lock_quick(m); 907 908 /* 909 * This turnstile is now no longer associated with the mutex. We can 910 * unlock the chain lock so a new turnstile may take it's place. 911 */ 912 turnstile_unpend(ts, TS_EXCLUSIVE_LOCK); 913 turnstile_chain_unlock(&m->lock_object); 914 } 915 916 /* 917 * All the unlocking of MTX_SPIN locks is done inline. 918 * See the __mtx_unlock_spin() macro for the details. 919 */ 920 921 /* 922 * The backing function for the INVARIANTS-enabled mtx_assert() 923 */ 924 #ifdef INVARIANT_SUPPORT 925 void 926 __mtx_assert(const volatile uintptr_t *c, int what, const char *file, int line) 927 { 928 const struct mtx *m; 929 930 if (panicstr != NULL || dumping || SCHEDULER_STOPPED()) 931 return; 932 933 m = mtxlock2mtx(c); 934 935 switch (what) { 936 case MA_OWNED: 937 case MA_OWNED | MA_RECURSED: 938 case MA_OWNED | MA_NOTRECURSED: 939 if (!mtx_owned(m)) 940 panic("mutex %s not owned at %s:%d", 941 m->lock_object.lo_name, file, line); 942 if (mtx_recursed(m)) { 943 if ((what & MA_NOTRECURSED) != 0) 944 panic("mutex %s recursed at %s:%d", 945 m->lock_object.lo_name, file, line); 946 } else if ((what & MA_RECURSED) != 0) { 947 panic("mutex %s unrecursed at %s:%d", 948 m->lock_object.lo_name, file, line); 949 } 950 break; 951 case MA_NOTOWNED: 952 if (mtx_owned(m)) 953 panic("mutex %s owned at %s:%d", 954 m->lock_object.lo_name, file, line); 955 break; 956 default: 957 panic("unknown mtx_assert at %s:%d", file, line); 958 } 959 } 960 #endif 961 962 /* 963 * General init routine used by the MTX_SYSINIT() macro. 964 */ 965 void 966 mtx_sysinit(void *arg) 967 { 968 struct mtx_args *margs = arg; 969 970 mtx_init((struct mtx *)margs->ma_mtx, margs->ma_desc, NULL, 971 margs->ma_opts); 972 } 973 974 /* 975 * Mutex initialization routine; initialize lock `m' of type contained in 976 * `opts' with options contained in `opts' and name `name.' The optional 977 * lock type `type' is used as a general lock category name for use with 978 * witness. 979 */ 980 void 981 _mtx_init(volatile uintptr_t *c, const char *name, const char *type, int opts) 982 { 983 struct mtx *m; 984 struct lock_class *class; 985 int flags; 986 987 m = mtxlock2mtx(c); 988 989 MPASS((opts & ~(MTX_SPIN | MTX_QUIET | MTX_RECURSE | 990 MTX_NOWITNESS | MTX_DUPOK | MTX_NOPROFILE | MTX_NEW)) == 0); 991 ASSERT_ATOMIC_LOAD_PTR(m->mtx_lock, 992 ("%s: mtx_lock not aligned for %s: %p", __func__, name, 993 &m->mtx_lock)); 994 995 /* Determine lock class and lock flags. */ 996 if (opts & MTX_SPIN) 997 class = &lock_class_mtx_spin; 998 else 999 class = &lock_class_mtx_sleep; 1000 flags = 0; 1001 if (opts & MTX_QUIET) 1002 flags |= LO_QUIET; 1003 if (opts & MTX_RECURSE) 1004 flags |= LO_RECURSABLE; 1005 if ((opts & MTX_NOWITNESS) == 0) 1006 flags |= LO_WITNESS; 1007 if (opts & MTX_DUPOK) 1008 flags |= LO_DUPOK; 1009 if (opts & MTX_NOPROFILE) 1010 flags |= LO_NOPROFILE; 1011 if (opts & MTX_NEW) 1012 flags |= LO_NEW; 1013 1014 /* Initialize mutex. */ 1015 lock_init(&m->lock_object, class, name, type, flags); 1016 1017 m->mtx_lock = MTX_UNOWNED; 1018 m->mtx_recurse = 0; 1019 } 1020 1021 /* 1022 * Remove lock `m' from all_mtx queue. We don't allow MTX_QUIET to be 1023 * passed in as a flag here because if the corresponding mtx_init() was 1024 * called with MTX_QUIET set, then it will already be set in the mutex's 1025 * flags. 1026 */ 1027 void 1028 _mtx_destroy(volatile uintptr_t *c) 1029 { 1030 struct mtx *m; 1031 1032 m = mtxlock2mtx(c); 1033 1034 if (!mtx_owned(m)) 1035 MPASS(mtx_unowned(m)); 1036 else { 1037 MPASS((m->mtx_lock & (MTX_RECURSED|MTX_CONTESTED)) == 0); 1038 1039 /* Perform the non-mtx related part of mtx_unlock_spin(). */ 1040 if (LOCK_CLASS(&m->lock_object) == &lock_class_mtx_spin) 1041 spinlock_exit(); 1042 else 1043 TD_LOCKS_DEC(curthread); 1044 1045 lock_profile_release_lock(&m->lock_object); 1046 /* Tell witness this isn't locked to make it happy. */ 1047 WITNESS_UNLOCK(&m->lock_object, LOP_EXCLUSIVE, __FILE__, 1048 __LINE__); 1049 } 1050 1051 m->mtx_lock = MTX_DESTROYED; 1052 lock_destroy(&m->lock_object); 1053 } 1054 1055 /* 1056 * Intialize the mutex code and system mutexes. This is called from the MD 1057 * startup code prior to mi_startup(). The per-CPU data space needs to be 1058 * setup before this is called. 1059 */ 1060 void 1061 mutex_init(void) 1062 { 1063 1064 /* Setup turnstiles so that sleep mutexes work. */ 1065 init_turnstiles(); 1066 1067 /* 1068 * Initialize mutexes. 1069 */ 1070 mtx_init(&Giant, "Giant", NULL, MTX_DEF | MTX_RECURSE); 1071 mtx_init(&blocked_lock, "blocked lock", NULL, MTX_SPIN); 1072 blocked_lock.mtx_lock = 0xdeadc0de; /* Always blocked. */ 1073 mtx_init(&proc0.p_mtx, "process lock", NULL, MTX_DEF | MTX_DUPOK); 1074 mtx_init(&proc0.p_slock, "process slock", NULL, MTX_SPIN); 1075 mtx_init(&proc0.p_statmtx, "pstatl", NULL, MTX_SPIN); 1076 mtx_init(&proc0.p_itimmtx, "pitiml", NULL, MTX_SPIN); 1077 mtx_init(&proc0.p_profmtx, "pprofl", NULL, MTX_SPIN); 1078 mtx_init(&devmtx, "cdev", NULL, MTX_DEF); 1079 mtx_lock(&Giant); 1080 } 1081 1082 #ifdef DDB 1083 void 1084 db_show_mtx(const struct lock_object *lock) 1085 { 1086 struct thread *td; 1087 const struct mtx *m; 1088 1089 m = (const struct mtx *)lock; 1090 1091 db_printf(" flags: {"); 1092 if (LOCK_CLASS(lock) == &lock_class_mtx_spin) 1093 db_printf("SPIN"); 1094 else 1095 db_printf("DEF"); 1096 if (m->lock_object.lo_flags & LO_RECURSABLE) 1097 db_printf(", RECURSE"); 1098 if (m->lock_object.lo_flags & LO_DUPOK) 1099 db_printf(", DUPOK"); 1100 db_printf("}\n"); 1101 db_printf(" state: {"); 1102 if (mtx_unowned(m)) 1103 db_printf("UNOWNED"); 1104 else if (mtx_destroyed(m)) 1105 db_printf("DESTROYED"); 1106 else { 1107 db_printf("OWNED"); 1108 if (m->mtx_lock & MTX_CONTESTED) 1109 db_printf(", CONTESTED"); 1110 if (m->mtx_lock & MTX_RECURSED) 1111 db_printf(", RECURSED"); 1112 } 1113 db_printf("}\n"); 1114 if (!mtx_unowned(m) && !mtx_destroyed(m)) { 1115 td = mtx_owner(m); 1116 db_printf(" owner: %p (tid %d, pid %d, \"%s\")\n", td, 1117 td->td_tid, td->td_proc->p_pid, td->td_name); 1118 if (mtx_recursed(m)) 1119 db_printf(" recursed: %d\n", m->mtx_recurse); 1120 } 1121 } 1122 #endif 1123