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 static void assert_mtx(const struct lock_object *lock, int what); 98 #ifdef DDB 99 static void db_show_mtx(const struct lock_object *lock); 100 #endif 101 static void lock_mtx(struct lock_object *lock, uintptr_t how); 102 static void lock_spin(struct lock_object *lock, uintptr_t how); 103 #ifdef KDTRACE_HOOKS 104 static int owner_mtx(const struct lock_object *lock, 105 struct thread **owner); 106 #endif 107 static uintptr_t unlock_mtx(struct lock_object *lock); 108 static uintptr_t unlock_spin(struct lock_object *lock); 109 110 /* 111 * Lock classes for sleep and spin mutexes. 112 */ 113 struct lock_class lock_class_mtx_sleep = { 114 .lc_name = "sleep mutex", 115 .lc_flags = LC_SLEEPLOCK | LC_RECURSABLE, 116 .lc_assert = assert_mtx, 117 #ifdef DDB 118 .lc_ddb_show = db_show_mtx, 119 #endif 120 .lc_lock = lock_mtx, 121 .lc_unlock = unlock_mtx, 122 #ifdef KDTRACE_HOOKS 123 .lc_owner = owner_mtx, 124 #endif 125 }; 126 struct lock_class lock_class_mtx_spin = { 127 .lc_name = "spin mutex", 128 .lc_flags = LC_SPINLOCK | LC_RECURSABLE, 129 .lc_assert = assert_mtx, 130 #ifdef DDB 131 .lc_ddb_show = db_show_mtx, 132 #endif 133 .lc_lock = lock_spin, 134 .lc_unlock = unlock_spin, 135 #ifdef KDTRACE_HOOKS 136 .lc_owner = owner_mtx, 137 #endif 138 }; 139 140 #ifdef ADAPTIVE_MUTEXES 141 static SYSCTL_NODE(_debug, OID_AUTO, mtx, CTLFLAG_RD, NULL, "mtx debugging"); 142 143 static struct lock_delay_config __read_mostly mtx_delay = { 144 .initial = 1000, 145 .step = 500, 146 .min = 100, 147 .max = 5000, 148 }; 149 150 SYSCTL_INT(_debug_mtx, OID_AUTO, delay_initial, CTLFLAG_RW, &mtx_delay.initial, 151 0, ""); 152 SYSCTL_INT(_debug_mtx, OID_AUTO, delay_step, CTLFLAG_RW, &mtx_delay.step, 153 0, ""); 154 SYSCTL_INT(_debug_mtx, OID_AUTO, delay_min, CTLFLAG_RW, &mtx_delay.min, 155 0, ""); 156 SYSCTL_INT(_debug_mtx, OID_AUTO, delay_max, CTLFLAG_RW, &mtx_delay.max, 157 0, ""); 158 159 static void 160 mtx_delay_sysinit(void *dummy) 161 { 162 163 mtx_delay.initial = mp_ncpus * 25; 164 mtx_delay.step = (mp_ncpus * 25) / 2; 165 mtx_delay.min = mp_ncpus * 5; 166 mtx_delay.max = mp_ncpus * 25 * 10; 167 } 168 LOCK_DELAY_SYSINIT(mtx_delay_sysinit); 169 #endif 170 171 static SYSCTL_NODE(_debug, OID_AUTO, mtx_spin, CTLFLAG_RD, NULL, 172 "mtx spin debugging"); 173 174 static struct lock_delay_config __read_mostly mtx_spin_delay = { 175 .initial = 1000, 176 .step = 500, 177 .min = 100, 178 .max = 5000, 179 }; 180 181 SYSCTL_INT(_debug_mtx_spin, OID_AUTO, delay_initial, CTLFLAG_RW, 182 &mtx_spin_delay.initial, 0, ""); 183 SYSCTL_INT(_debug_mtx_spin, OID_AUTO, delay_step, CTLFLAG_RW, &mtx_spin_delay.step, 184 0, ""); 185 SYSCTL_INT(_debug_mtx_spin, OID_AUTO, delay_min, CTLFLAG_RW, &mtx_spin_delay.min, 186 0, ""); 187 SYSCTL_INT(_debug_mtx_spin, OID_AUTO, delay_max, CTLFLAG_RW, &mtx_spin_delay.max, 188 0, ""); 189 190 static void 191 mtx_spin_delay_sysinit(void *dummy) 192 { 193 194 mtx_spin_delay.initial = mp_ncpus * 25; 195 mtx_spin_delay.step = (mp_ncpus * 25) / 2; 196 mtx_spin_delay.min = mp_ncpus * 5; 197 mtx_spin_delay.max = mp_ncpus * 25 * 10; 198 } 199 LOCK_DELAY_SYSINIT(mtx_spin_delay_sysinit); 200 201 /* 202 * System-wide mutexes 203 */ 204 struct mtx blocked_lock; 205 struct mtx Giant; 206 207 void 208 assert_mtx(const struct lock_object *lock, int what) 209 { 210 211 mtx_assert((const struct mtx *)lock, what); 212 } 213 214 void 215 lock_mtx(struct lock_object *lock, uintptr_t how) 216 { 217 218 mtx_lock((struct mtx *)lock); 219 } 220 221 void 222 lock_spin(struct lock_object *lock, uintptr_t how) 223 { 224 225 panic("spin locks can only use msleep_spin"); 226 } 227 228 uintptr_t 229 unlock_mtx(struct lock_object *lock) 230 { 231 struct mtx *m; 232 233 m = (struct mtx *)lock; 234 mtx_assert(m, MA_OWNED | MA_NOTRECURSED); 235 mtx_unlock(m); 236 return (0); 237 } 238 239 uintptr_t 240 unlock_spin(struct lock_object *lock) 241 { 242 243 panic("spin locks can only use msleep_spin"); 244 } 245 246 #ifdef KDTRACE_HOOKS 247 int 248 owner_mtx(const struct lock_object *lock, struct thread **owner) 249 { 250 const struct mtx *m; 251 uintptr_t x; 252 253 m = (const struct mtx *)lock; 254 x = m->mtx_lock; 255 *owner = (struct thread *)(x & ~MTX_FLAGMASK); 256 return (x != MTX_UNOWNED); 257 } 258 #endif 259 260 /* 261 * Function versions of the inlined __mtx_* macros. These are used by 262 * modules and can also be called from assembly language if needed. 263 */ 264 void 265 __mtx_lock_flags(volatile uintptr_t *c, int opts, const char *file, int line) 266 { 267 struct mtx *m; 268 uintptr_t tid, v; 269 270 if (SCHEDULER_STOPPED()) 271 return; 272 273 m = mtxlock2mtx(c); 274 275 KASSERT(kdb_active != 0 || !TD_IS_IDLETHREAD(curthread), 276 ("mtx_lock() by idle thread %p on sleep mutex %s @ %s:%d", 277 curthread, m->lock_object.lo_name, file, line)); 278 KASSERT(m->mtx_lock != MTX_DESTROYED, 279 ("mtx_lock() of destroyed mutex @ %s:%d", file, line)); 280 KASSERT(LOCK_CLASS(&m->lock_object) == &lock_class_mtx_sleep, 281 ("mtx_lock() of spin mutex %s @ %s:%d", m->lock_object.lo_name, 282 file, line)); 283 WITNESS_CHECKORDER(&m->lock_object, (opts & ~MTX_RECURSE) | 284 LOP_NEWORDER | LOP_EXCLUSIVE, file, line, NULL); 285 286 tid = (uintptr_t)curthread; 287 v = MTX_UNOWNED; 288 if (!_mtx_obtain_lock_fetch(m, &v, tid)) 289 _mtx_lock_sleep(m, v, tid, opts, file, line); 290 else 291 LOCKSTAT_PROFILE_OBTAIN_LOCK_SUCCESS(adaptive__acquire, 292 m, 0, 0, file, line); 293 LOCK_LOG_LOCK("LOCK", &m->lock_object, opts, m->mtx_recurse, file, 294 line); 295 WITNESS_LOCK(&m->lock_object, (opts & ~MTX_RECURSE) | LOP_EXCLUSIVE, 296 file, line); 297 TD_LOCKS_INC(curthread); 298 } 299 300 void 301 __mtx_unlock_flags(volatile uintptr_t *c, int opts, const char *file, int line) 302 { 303 struct mtx *m; 304 305 if (SCHEDULER_STOPPED()) 306 return; 307 308 m = mtxlock2mtx(c); 309 310 KASSERT(m->mtx_lock != MTX_DESTROYED, 311 ("mtx_unlock() of destroyed mutex @ %s:%d", file, line)); 312 KASSERT(LOCK_CLASS(&m->lock_object) == &lock_class_mtx_sleep, 313 ("mtx_unlock() of spin mutex %s @ %s:%d", m->lock_object.lo_name, 314 file, line)); 315 WITNESS_UNLOCK(&m->lock_object, opts | LOP_EXCLUSIVE, file, line); 316 LOCK_LOG_LOCK("UNLOCK", &m->lock_object, opts, m->mtx_recurse, file, 317 line); 318 mtx_assert(m, MA_OWNED); 319 320 __mtx_unlock_sleep(c, opts, file, line); 321 TD_LOCKS_DEC(curthread); 322 } 323 324 void 325 __mtx_lock_spin_flags(volatile uintptr_t *c, int opts, const char *file, 326 int line) 327 { 328 struct mtx *m; 329 330 if (SCHEDULER_STOPPED()) 331 return; 332 333 m = mtxlock2mtx(c); 334 335 KASSERT(m->mtx_lock != MTX_DESTROYED, 336 ("mtx_lock_spin() of destroyed mutex @ %s:%d", file, line)); 337 KASSERT(LOCK_CLASS(&m->lock_object) == &lock_class_mtx_spin, 338 ("mtx_lock_spin() of sleep mutex %s @ %s:%d", 339 m->lock_object.lo_name, file, line)); 340 if (mtx_owned(m)) 341 KASSERT((m->lock_object.lo_flags & LO_RECURSABLE) != 0 || 342 (opts & MTX_RECURSE) != 0, 343 ("mtx_lock_spin: recursed on non-recursive mutex %s @ %s:%d\n", 344 m->lock_object.lo_name, file, line)); 345 opts &= ~MTX_RECURSE; 346 WITNESS_CHECKORDER(&m->lock_object, opts | LOP_NEWORDER | LOP_EXCLUSIVE, 347 file, line, NULL); 348 __mtx_lock_spin(m, curthread, opts, file, line); 349 LOCK_LOG_LOCK("LOCK", &m->lock_object, opts, m->mtx_recurse, file, 350 line); 351 WITNESS_LOCK(&m->lock_object, opts | LOP_EXCLUSIVE, file, line); 352 } 353 354 int 355 __mtx_trylock_spin_flags(volatile uintptr_t *c, int opts, const char *file, 356 int line) 357 { 358 struct mtx *m; 359 360 if (SCHEDULER_STOPPED()) 361 return (1); 362 363 m = mtxlock2mtx(c); 364 365 KASSERT(m->mtx_lock != MTX_DESTROYED, 366 ("mtx_trylock_spin() of destroyed mutex @ %s:%d", file, line)); 367 KASSERT(LOCK_CLASS(&m->lock_object) == &lock_class_mtx_spin, 368 ("mtx_trylock_spin() of sleep mutex %s @ %s:%d", 369 m->lock_object.lo_name, file, line)); 370 KASSERT((opts & MTX_RECURSE) == 0, 371 ("mtx_trylock_spin: unsupp. opt MTX_RECURSE on mutex %s @ %s:%d\n", 372 m->lock_object.lo_name, file, line)); 373 if (__mtx_trylock_spin(m, curthread, opts, file, line)) { 374 LOCK_LOG_TRY("LOCK", &m->lock_object, opts, 1, file, line); 375 WITNESS_LOCK(&m->lock_object, opts | LOP_EXCLUSIVE, file, line); 376 return (1); 377 } 378 LOCK_LOG_TRY("LOCK", &m->lock_object, opts, 0, file, line); 379 return (0); 380 } 381 382 void 383 __mtx_unlock_spin_flags(volatile uintptr_t *c, int opts, const char *file, 384 int line) 385 { 386 struct mtx *m; 387 388 if (SCHEDULER_STOPPED()) 389 return; 390 391 m = mtxlock2mtx(c); 392 393 KASSERT(m->mtx_lock != MTX_DESTROYED, 394 ("mtx_unlock_spin() of destroyed mutex @ %s:%d", file, line)); 395 KASSERT(LOCK_CLASS(&m->lock_object) == &lock_class_mtx_spin, 396 ("mtx_unlock_spin() of sleep mutex %s @ %s:%d", 397 m->lock_object.lo_name, file, line)); 398 WITNESS_UNLOCK(&m->lock_object, opts | LOP_EXCLUSIVE, file, line); 399 LOCK_LOG_LOCK("UNLOCK", &m->lock_object, opts, m->mtx_recurse, file, 400 line); 401 mtx_assert(m, MA_OWNED); 402 403 __mtx_unlock_spin(m); 404 } 405 406 /* 407 * The important part of mtx_trylock{,_flags}() 408 * Tries to acquire lock `m.' If this function is called on a mutex that 409 * is already owned, it will recursively acquire the lock. 410 */ 411 int 412 _mtx_trylock_flags_(volatile uintptr_t *c, int opts, const char *file, int line) 413 { 414 struct mtx *m; 415 #ifdef LOCK_PROFILING 416 uint64_t waittime = 0; 417 int contested = 0; 418 #endif 419 int rval; 420 421 if (SCHEDULER_STOPPED()) 422 return (1); 423 424 m = mtxlock2mtx(c); 425 426 KASSERT(kdb_active != 0 || !TD_IS_IDLETHREAD(curthread), 427 ("mtx_trylock() by idle thread %p on sleep mutex %s @ %s:%d", 428 curthread, m->lock_object.lo_name, file, line)); 429 KASSERT(m->mtx_lock != MTX_DESTROYED, 430 ("mtx_trylock() of destroyed mutex @ %s:%d", file, line)); 431 KASSERT(LOCK_CLASS(&m->lock_object) == &lock_class_mtx_sleep, 432 ("mtx_trylock() of spin mutex %s @ %s:%d", m->lock_object.lo_name, 433 file, line)); 434 435 if (mtx_owned(m) && ((m->lock_object.lo_flags & LO_RECURSABLE) != 0 || 436 (opts & MTX_RECURSE) != 0)) { 437 m->mtx_recurse++; 438 atomic_set_ptr(&m->mtx_lock, MTX_RECURSED); 439 rval = 1; 440 } else 441 rval = _mtx_obtain_lock(m, (uintptr_t)curthread); 442 opts &= ~MTX_RECURSE; 443 444 LOCK_LOG_TRY("LOCK", &m->lock_object, opts, rval, file, line); 445 if (rval) { 446 WITNESS_LOCK(&m->lock_object, opts | LOP_EXCLUSIVE | LOP_TRYLOCK, 447 file, line); 448 TD_LOCKS_INC(curthread); 449 if (m->mtx_recurse == 0) 450 LOCKSTAT_PROFILE_OBTAIN_LOCK_SUCCESS(adaptive__acquire, 451 m, contested, waittime, file, line); 452 453 } 454 455 return (rval); 456 } 457 458 /* 459 * __mtx_lock_sleep: the tougher part of acquiring an MTX_DEF lock. 460 * 461 * We call this if the lock is either contested (i.e. we need to go to 462 * sleep waiting for it), or if we need to recurse on it. 463 */ 464 void 465 __mtx_lock_sleep(volatile uintptr_t *c, uintptr_t v, uintptr_t tid, int opts, 466 const char *file, int line) 467 { 468 struct mtx *m; 469 struct turnstile *ts; 470 #ifdef ADAPTIVE_MUTEXES 471 volatile struct thread *owner; 472 #endif 473 #ifdef KTR 474 int cont_logged = 0; 475 #endif 476 #ifdef LOCK_PROFILING 477 int contested = 0; 478 uint64_t waittime = 0; 479 #endif 480 #if defined(ADAPTIVE_MUTEXES) || defined(KDTRACE_HOOKS) 481 struct lock_delay_arg lda; 482 #endif 483 #ifdef KDTRACE_HOOKS 484 u_int sleep_cnt = 0; 485 int64_t sleep_time = 0; 486 int64_t all_time = 0; 487 #endif 488 489 if (SCHEDULER_STOPPED()) 490 return; 491 492 #if defined(ADAPTIVE_MUTEXES) 493 lock_delay_arg_init(&lda, &mtx_delay); 494 #elif defined(KDTRACE_HOOKS) 495 lock_delay_arg_init(&lda, NULL); 496 #endif 497 m = mtxlock2mtx(c); 498 499 if (__predict_false(lv_mtx_owner(v) == (struct thread *)tid)) { 500 KASSERT((m->lock_object.lo_flags & LO_RECURSABLE) != 0 || 501 (opts & MTX_RECURSE) != 0, 502 ("_mtx_lock_sleep: recursed on non-recursive mutex %s @ %s:%d\n", 503 m->lock_object.lo_name, file, line)); 504 opts &= ~MTX_RECURSE; 505 m->mtx_recurse++; 506 atomic_set_ptr(&m->mtx_lock, MTX_RECURSED); 507 if (LOCK_LOG_TEST(&m->lock_object, opts)) 508 CTR1(KTR_LOCK, "_mtx_lock_sleep: %p recursing", m); 509 return; 510 } 511 opts &= ~MTX_RECURSE; 512 513 #ifdef HWPMC_HOOKS 514 PMC_SOFT_CALL( , , lock, failed); 515 #endif 516 lock_profile_obtain_lock_failed(&m->lock_object, 517 &contested, &waittime); 518 if (LOCK_LOG_TEST(&m->lock_object, opts)) 519 CTR4(KTR_LOCK, 520 "_mtx_lock_sleep: %s contested (lock=%p) at %s:%d", 521 m->lock_object.lo_name, (void *)m->mtx_lock, file, line); 522 #ifdef KDTRACE_HOOKS 523 all_time -= lockstat_nsecs(&m->lock_object); 524 #endif 525 526 for (;;) { 527 if (v == MTX_UNOWNED) { 528 if (_mtx_obtain_lock_fetch(m, &v, tid)) 529 break; 530 continue; 531 } 532 #ifdef KDTRACE_HOOKS 533 lda.spin_cnt++; 534 #endif 535 #ifdef ADAPTIVE_MUTEXES 536 /* 537 * If the owner is running on another CPU, spin until the 538 * owner stops running or the state of the lock changes. 539 */ 540 owner = lv_mtx_owner(v); 541 if (TD_IS_RUNNING(owner)) { 542 if (LOCK_LOG_TEST(&m->lock_object, 0)) 543 CTR3(KTR_LOCK, 544 "%s: spinning on %p held by %p", 545 __func__, m, owner); 546 KTR_STATE1(KTR_SCHED, "thread", 547 sched_tdname((struct thread *)tid), 548 "spinning", "lockname:\"%s\"", 549 m->lock_object.lo_name); 550 do { 551 lock_delay(&lda); 552 v = MTX_READ_VALUE(m); 553 owner = lv_mtx_owner(v); 554 } while (v != MTX_UNOWNED && TD_IS_RUNNING(owner)); 555 KTR_STATE0(KTR_SCHED, "thread", 556 sched_tdname((struct thread *)tid), 557 "running"); 558 continue; 559 } 560 #endif 561 562 ts = turnstile_trywait(&m->lock_object); 563 v = MTX_READ_VALUE(m); 564 565 /* 566 * Check if the lock has been released while spinning for 567 * the turnstile chain lock. 568 */ 569 if (v == MTX_UNOWNED) { 570 turnstile_cancel(ts); 571 continue; 572 } 573 574 #ifdef ADAPTIVE_MUTEXES 575 /* 576 * The current lock owner might have started executing 577 * on another CPU (or the lock could have changed 578 * owners) while we were waiting on the turnstile 579 * chain lock. If so, drop the turnstile lock and try 580 * again. 581 */ 582 owner = lv_mtx_owner(v); 583 if (TD_IS_RUNNING(owner)) { 584 turnstile_cancel(ts); 585 continue; 586 } 587 #endif 588 589 /* 590 * If the mutex isn't already contested and a failure occurs 591 * setting the contested bit, the mutex was either released 592 * or the state of the MTX_RECURSED bit changed. 593 */ 594 if ((v & MTX_CONTESTED) == 0 && 595 !atomic_cmpset_ptr(&m->mtx_lock, v, v | MTX_CONTESTED)) { 596 turnstile_cancel(ts); 597 v = MTX_READ_VALUE(m); 598 continue; 599 } 600 601 /* 602 * We definitely must sleep for this lock. 603 */ 604 mtx_assert(m, MA_NOTOWNED); 605 606 #ifdef KTR 607 if (!cont_logged) { 608 CTR6(KTR_CONTENTION, 609 "contention: %p at %s:%d wants %s, taken by %s:%d", 610 (void *)tid, file, line, m->lock_object.lo_name, 611 WITNESS_FILE(&m->lock_object), 612 WITNESS_LINE(&m->lock_object)); 613 cont_logged = 1; 614 } 615 #endif 616 617 /* 618 * Block on the turnstile. 619 */ 620 #ifdef KDTRACE_HOOKS 621 sleep_time -= lockstat_nsecs(&m->lock_object); 622 #endif 623 turnstile_wait(ts, mtx_owner(m), TS_EXCLUSIVE_QUEUE); 624 #ifdef KDTRACE_HOOKS 625 sleep_time += lockstat_nsecs(&m->lock_object); 626 sleep_cnt++; 627 #endif 628 v = MTX_READ_VALUE(m); 629 } 630 #ifdef KDTRACE_HOOKS 631 all_time += lockstat_nsecs(&m->lock_object); 632 #endif 633 #ifdef KTR 634 if (cont_logged) { 635 CTR4(KTR_CONTENTION, 636 "contention end: %s acquired by %p at %s:%d", 637 m->lock_object.lo_name, (void *)tid, file, line); 638 } 639 #endif 640 LOCKSTAT_PROFILE_OBTAIN_LOCK_SUCCESS(adaptive__acquire, m, contested, 641 waittime, file, line); 642 #ifdef KDTRACE_HOOKS 643 if (sleep_time) 644 LOCKSTAT_RECORD1(adaptive__block, m, sleep_time); 645 646 /* 647 * Only record the loops spinning and not sleeping. 648 */ 649 if (lda.spin_cnt > sleep_cnt) 650 LOCKSTAT_RECORD1(adaptive__spin, m, all_time - sleep_time); 651 #endif 652 } 653 654 static void 655 _mtx_lock_spin_failed(struct mtx *m) 656 { 657 struct thread *td; 658 659 td = mtx_owner(m); 660 661 /* If the mutex is unlocked, try again. */ 662 if (td == NULL) 663 return; 664 665 printf( "spin lock %p (%s) held by %p (tid %d) too long\n", 666 m, m->lock_object.lo_name, td, td->td_tid); 667 #ifdef WITNESS 668 witness_display_spinlock(&m->lock_object, td, printf); 669 #endif 670 panic("spin lock held too long"); 671 } 672 673 #ifdef SMP 674 /* 675 * _mtx_lock_spin_cookie: the tougher part of acquiring an MTX_SPIN lock. 676 * 677 * This is only called if we need to actually spin for the lock. Recursion 678 * is handled inline. 679 */ 680 void 681 _mtx_lock_spin_cookie(volatile uintptr_t *c, uintptr_t v, uintptr_t tid, 682 int opts, const char *file, int line) 683 { 684 struct mtx *m; 685 struct lock_delay_arg lda; 686 #ifdef LOCK_PROFILING 687 int contested = 0; 688 uint64_t waittime = 0; 689 #endif 690 #ifdef KDTRACE_HOOKS 691 int64_t spin_time = 0; 692 #endif 693 694 if (SCHEDULER_STOPPED()) 695 return; 696 697 lock_delay_arg_init(&lda, &mtx_spin_delay); 698 m = mtxlock2mtx(c); 699 700 if (LOCK_LOG_TEST(&m->lock_object, opts)) 701 CTR1(KTR_LOCK, "_mtx_lock_spin: %p spinning", m); 702 KTR_STATE1(KTR_SCHED, "thread", sched_tdname((struct thread *)tid), 703 "spinning", "lockname:\"%s\"", m->lock_object.lo_name); 704 705 #ifdef HWPMC_HOOKS 706 PMC_SOFT_CALL( , , lock, failed); 707 #endif 708 lock_profile_obtain_lock_failed(&m->lock_object, &contested, &waittime); 709 #ifdef KDTRACE_HOOKS 710 spin_time -= lockstat_nsecs(&m->lock_object); 711 #endif 712 for (;;) { 713 if (v == MTX_UNOWNED) { 714 if (_mtx_obtain_lock_fetch(m, &v, tid)) 715 break; 716 continue; 717 } 718 /* Give interrupts a chance while we spin. */ 719 spinlock_exit(); 720 do { 721 if (lda.spin_cnt < 10000000) { 722 lock_delay(&lda); 723 } else { 724 lda.spin_cnt++; 725 if (lda.spin_cnt < 60000000 || kdb_active || 726 panicstr != NULL) 727 DELAY(1); 728 else 729 _mtx_lock_spin_failed(m); 730 cpu_spinwait(); 731 } 732 v = MTX_READ_VALUE(m); 733 } while (v != MTX_UNOWNED); 734 spinlock_enter(); 735 } 736 #ifdef KDTRACE_HOOKS 737 spin_time += lockstat_nsecs(&m->lock_object); 738 #endif 739 740 if (LOCK_LOG_TEST(&m->lock_object, opts)) 741 CTR1(KTR_LOCK, "_mtx_lock_spin: %p spin done", m); 742 KTR_STATE0(KTR_SCHED, "thread", sched_tdname((struct thread *)tid), 743 "running"); 744 745 #ifdef KDTRACE_HOOKS 746 LOCKSTAT_PROFILE_OBTAIN_LOCK_SUCCESS(spin__acquire, m, 747 contested, waittime, file, line); 748 if (spin_time != 0) 749 LOCKSTAT_RECORD1(spin__spin, m, spin_time); 750 #endif 751 } 752 #endif /* SMP */ 753 754 void 755 thread_lock_flags_(struct thread *td, int opts, const char *file, int line) 756 { 757 struct mtx *m; 758 uintptr_t tid, v; 759 struct lock_delay_arg lda; 760 #ifdef LOCK_PROFILING 761 int contested = 0; 762 uint64_t waittime = 0; 763 #endif 764 #ifdef KDTRACE_HOOKS 765 int64_t spin_time = 0; 766 #endif 767 768 tid = (uintptr_t)curthread; 769 770 if (SCHEDULER_STOPPED()) { 771 /* 772 * Ensure that spinlock sections are balanced even when the 773 * scheduler is stopped, since we may otherwise inadvertently 774 * re-enable interrupts while dumping core. 775 */ 776 spinlock_enter(); 777 return; 778 } 779 780 lock_delay_arg_init(&lda, &mtx_spin_delay); 781 782 #ifdef KDTRACE_HOOKS 783 spin_time -= lockstat_nsecs(&td->td_lock->lock_object); 784 #endif 785 for (;;) { 786 retry: 787 v = MTX_UNOWNED; 788 spinlock_enter(); 789 m = td->td_lock; 790 KASSERT(m->mtx_lock != MTX_DESTROYED, 791 ("thread_lock() of destroyed mutex @ %s:%d", file, line)); 792 KASSERT(LOCK_CLASS(&m->lock_object) == &lock_class_mtx_spin, 793 ("thread_lock() of sleep mutex %s @ %s:%d", 794 m->lock_object.lo_name, file, line)); 795 if (mtx_owned(m)) 796 KASSERT((m->lock_object.lo_flags & LO_RECURSABLE) != 0, 797 ("thread_lock: recursed on non-recursive mutex %s @ %s:%d\n", 798 m->lock_object.lo_name, file, line)); 799 WITNESS_CHECKORDER(&m->lock_object, 800 opts | LOP_NEWORDER | LOP_EXCLUSIVE, file, line, NULL); 801 for (;;) { 802 if (_mtx_obtain_lock_fetch(m, &v, tid)) 803 break; 804 if (v == MTX_UNOWNED) 805 continue; 806 if (v == tid) { 807 m->mtx_recurse++; 808 break; 809 } 810 #ifdef HWPMC_HOOKS 811 PMC_SOFT_CALL( , , lock, failed); 812 #endif 813 lock_profile_obtain_lock_failed(&m->lock_object, 814 &contested, &waittime); 815 /* Give interrupts a chance while we spin. */ 816 spinlock_exit(); 817 do { 818 if (lda.spin_cnt < 10000000) { 819 lock_delay(&lda); 820 } else { 821 lda.spin_cnt++; 822 if (lda.spin_cnt < 60000000 || 823 kdb_active || panicstr != NULL) 824 DELAY(1); 825 else 826 _mtx_lock_spin_failed(m); 827 cpu_spinwait(); 828 } 829 if (m != td->td_lock) 830 goto retry; 831 v = MTX_READ_VALUE(m); 832 } while (v != MTX_UNOWNED); 833 spinlock_enter(); 834 } 835 if (m == td->td_lock) 836 break; 837 __mtx_unlock_spin(m); /* does spinlock_exit() */ 838 } 839 #ifdef KDTRACE_HOOKS 840 spin_time += lockstat_nsecs(&m->lock_object); 841 #endif 842 if (m->mtx_recurse == 0) 843 LOCKSTAT_PROFILE_OBTAIN_LOCK_SUCCESS(spin__acquire, m, 844 contested, waittime, file, line); 845 LOCK_LOG_LOCK("LOCK", &m->lock_object, opts, m->mtx_recurse, file, 846 line); 847 WITNESS_LOCK(&m->lock_object, opts | LOP_EXCLUSIVE, file, line); 848 #ifdef KDTRACE_HOOKS 849 if (spin_time != 0) 850 LOCKSTAT_RECORD1(thread__spin, m, spin_time); 851 #endif 852 } 853 854 struct mtx * 855 thread_lock_block(struct thread *td) 856 { 857 struct mtx *lock; 858 859 THREAD_LOCK_ASSERT(td, MA_OWNED); 860 lock = td->td_lock; 861 td->td_lock = &blocked_lock; 862 mtx_unlock_spin(lock); 863 864 return (lock); 865 } 866 867 void 868 thread_lock_unblock(struct thread *td, struct mtx *new) 869 { 870 mtx_assert(new, MA_OWNED); 871 MPASS(td->td_lock == &blocked_lock); 872 atomic_store_rel_ptr((volatile void *)&td->td_lock, (uintptr_t)new); 873 } 874 875 void 876 thread_lock_set(struct thread *td, struct mtx *new) 877 { 878 struct mtx *lock; 879 880 mtx_assert(new, MA_OWNED); 881 THREAD_LOCK_ASSERT(td, MA_OWNED); 882 lock = td->td_lock; 883 td->td_lock = new; 884 mtx_unlock_spin(lock); 885 } 886 887 /* 888 * __mtx_unlock_sleep: the tougher part of releasing an MTX_DEF lock. 889 * 890 * We are only called here if the lock is recursed or contested (i.e. we 891 * need to wake up a blocked thread). 892 */ 893 void 894 __mtx_unlock_sleep(volatile uintptr_t *c, int opts, const char *file, int line) 895 { 896 struct mtx *m; 897 struct turnstile *ts; 898 899 if (SCHEDULER_STOPPED()) 900 return; 901 902 m = mtxlock2mtx(c); 903 904 if (!mtx_recursed(m)) { 905 LOCKSTAT_PROFILE_RELEASE_LOCK(adaptive__release, m); 906 if (_mtx_release_lock(m, (uintptr_t)curthread)) 907 return; 908 } else { 909 if (--(m->mtx_recurse) == 0) 910 atomic_clear_ptr(&m->mtx_lock, MTX_RECURSED); 911 if (LOCK_LOG_TEST(&m->lock_object, opts)) 912 CTR1(KTR_LOCK, "_mtx_unlock_sleep: %p unrecurse", m); 913 return; 914 } 915 916 /* 917 * We have to lock the chain before the turnstile so this turnstile 918 * can be removed from the hash list if it is empty. 919 */ 920 turnstile_chain_lock(&m->lock_object); 921 ts = turnstile_lookup(&m->lock_object); 922 if (LOCK_LOG_TEST(&m->lock_object, opts)) 923 CTR1(KTR_LOCK, "_mtx_unlock_sleep: %p contested", m); 924 MPASS(ts != NULL); 925 turnstile_broadcast(ts, TS_EXCLUSIVE_QUEUE); 926 _mtx_release_lock_quick(m); 927 928 /* 929 * This turnstile is now no longer associated with the mutex. We can 930 * unlock the chain lock so a new turnstile may take it's place. 931 */ 932 turnstile_unpend(ts, TS_EXCLUSIVE_LOCK); 933 turnstile_chain_unlock(&m->lock_object); 934 } 935 936 /* 937 * All the unlocking of MTX_SPIN locks is done inline. 938 * See the __mtx_unlock_spin() macro for the details. 939 */ 940 941 /* 942 * The backing function for the INVARIANTS-enabled mtx_assert() 943 */ 944 #ifdef INVARIANT_SUPPORT 945 void 946 __mtx_assert(const volatile uintptr_t *c, int what, const char *file, int line) 947 { 948 const struct mtx *m; 949 950 if (panicstr != NULL || dumping || SCHEDULER_STOPPED()) 951 return; 952 953 m = mtxlock2mtx(c); 954 955 switch (what) { 956 case MA_OWNED: 957 case MA_OWNED | MA_RECURSED: 958 case MA_OWNED | MA_NOTRECURSED: 959 if (!mtx_owned(m)) 960 panic("mutex %s not owned at %s:%d", 961 m->lock_object.lo_name, file, line); 962 if (mtx_recursed(m)) { 963 if ((what & MA_NOTRECURSED) != 0) 964 panic("mutex %s recursed at %s:%d", 965 m->lock_object.lo_name, file, line); 966 } else if ((what & MA_RECURSED) != 0) { 967 panic("mutex %s unrecursed at %s:%d", 968 m->lock_object.lo_name, file, line); 969 } 970 break; 971 case MA_NOTOWNED: 972 if (mtx_owned(m)) 973 panic("mutex %s owned at %s:%d", 974 m->lock_object.lo_name, file, line); 975 break; 976 default: 977 panic("unknown mtx_assert at %s:%d", file, line); 978 } 979 } 980 #endif 981 982 /* 983 * General init routine used by the MTX_SYSINIT() macro. 984 */ 985 void 986 mtx_sysinit(void *arg) 987 { 988 struct mtx_args *margs = arg; 989 990 mtx_init((struct mtx *)margs->ma_mtx, margs->ma_desc, NULL, 991 margs->ma_opts); 992 } 993 994 /* 995 * Mutex initialization routine; initialize lock `m' of type contained in 996 * `opts' with options contained in `opts' and name `name.' The optional 997 * lock type `type' is used as a general lock category name for use with 998 * witness. 999 */ 1000 void 1001 _mtx_init(volatile uintptr_t *c, const char *name, const char *type, int opts) 1002 { 1003 struct mtx *m; 1004 struct lock_class *class; 1005 int flags; 1006 1007 m = mtxlock2mtx(c); 1008 1009 MPASS((opts & ~(MTX_SPIN | MTX_QUIET | MTX_RECURSE | 1010 MTX_NOWITNESS | MTX_DUPOK | MTX_NOPROFILE | MTX_NEW)) == 0); 1011 ASSERT_ATOMIC_LOAD_PTR(m->mtx_lock, 1012 ("%s: mtx_lock not aligned for %s: %p", __func__, name, 1013 &m->mtx_lock)); 1014 1015 /* Determine lock class and lock flags. */ 1016 if (opts & MTX_SPIN) 1017 class = &lock_class_mtx_spin; 1018 else 1019 class = &lock_class_mtx_sleep; 1020 flags = 0; 1021 if (opts & MTX_QUIET) 1022 flags |= LO_QUIET; 1023 if (opts & MTX_RECURSE) 1024 flags |= LO_RECURSABLE; 1025 if ((opts & MTX_NOWITNESS) == 0) 1026 flags |= LO_WITNESS; 1027 if (opts & MTX_DUPOK) 1028 flags |= LO_DUPOK; 1029 if (opts & MTX_NOPROFILE) 1030 flags |= LO_NOPROFILE; 1031 if (opts & MTX_NEW) 1032 flags |= LO_NEW; 1033 1034 /* Initialize mutex. */ 1035 lock_init(&m->lock_object, class, name, type, flags); 1036 1037 m->mtx_lock = MTX_UNOWNED; 1038 m->mtx_recurse = 0; 1039 } 1040 1041 /* 1042 * Remove lock `m' from all_mtx queue. We don't allow MTX_QUIET to be 1043 * passed in as a flag here because if the corresponding mtx_init() was 1044 * called with MTX_QUIET set, then it will already be set in the mutex's 1045 * flags. 1046 */ 1047 void 1048 _mtx_destroy(volatile uintptr_t *c) 1049 { 1050 struct mtx *m; 1051 1052 m = mtxlock2mtx(c); 1053 1054 if (!mtx_owned(m)) 1055 MPASS(mtx_unowned(m)); 1056 else { 1057 MPASS((m->mtx_lock & (MTX_RECURSED|MTX_CONTESTED)) == 0); 1058 1059 /* Perform the non-mtx related part of mtx_unlock_spin(). */ 1060 if (LOCK_CLASS(&m->lock_object) == &lock_class_mtx_spin) 1061 spinlock_exit(); 1062 else 1063 TD_LOCKS_DEC(curthread); 1064 1065 lock_profile_release_lock(&m->lock_object); 1066 /* Tell witness this isn't locked to make it happy. */ 1067 WITNESS_UNLOCK(&m->lock_object, LOP_EXCLUSIVE, __FILE__, 1068 __LINE__); 1069 } 1070 1071 m->mtx_lock = MTX_DESTROYED; 1072 lock_destroy(&m->lock_object); 1073 } 1074 1075 /* 1076 * Intialize the mutex code and system mutexes. This is called from the MD 1077 * startup code prior to mi_startup(). The per-CPU data space needs to be 1078 * setup before this is called. 1079 */ 1080 void 1081 mutex_init(void) 1082 { 1083 1084 /* Setup turnstiles so that sleep mutexes work. */ 1085 init_turnstiles(); 1086 1087 /* 1088 * Initialize mutexes. 1089 */ 1090 mtx_init(&Giant, "Giant", NULL, MTX_DEF | MTX_RECURSE); 1091 mtx_init(&blocked_lock, "blocked lock", NULL, MTX_SPIN); 1092 blocked_lock.mtx_lock = 0xdeadc0de; /* Always blocked. */ 1093 mtx_init(&proc0.p_mtx, "process lock", NULL, MTX_DEF | MTX_DUPOK); 1094 mtx_init(&proc0.p_slock, "process slock", NULL, MTX_SPIN); 1095 mtx_init(&proc0.p_statmtx, "pstatl", NULL, MTX_SPIN); 1096 mtx_init(&proc0.p_itimmtx, "pitiml", NULL, MTX_SPIN); 1097 mtx_init(&proc0.p_profmtx, "pprofl", NULL, MTX_SPIN); 1098 mtx_init(&devmtx, "cdev", NULL, MTX_DEF); 1099 mtx_lock(&Giant); 1100 } 1101 1102 #ifdef DDB 1103 void 1104 db_show_mtx(const struct lock_object *lock) 1105 { 1106 struct thread *td; 1107 const struct mtx *m; 1108 1109 m = (const struct mtx *)lock; 1110 1111 db_printf(" flags: {"); 1112 if (LOCK_CLASS(lock) == &lock_class_mtx_spin) 1113 db_printf("SPIN"); 1114 else 1115 db_printf("DEF"); 1116 if (m->lock_object.lo_flags & LO_RECURSABLE) 1117 db_printf(", RECURSE"); 1118 if (m->lock_object.lo_flags & LO_DUPOK) 1119 db_printf(", DUPOK"); 1120 db_printf("}\n"); 1121 db_printf(" state: {"); 1122 if (mtx_unowned(m)) 1123 db_printf("UNOWNED"); 1124 else if (mtx_destroyed(m)) 1125 db_printf("DESTROYED"); 1126 else { 1127 db_printf("OWNED"); 1128 if (m->mtx_lock & MTX_CONTESTED) 1129 db_printf(", CONTESTED"); 1130 if (m->mtx_lock & MTX_RECURSED) 1131 db_printf(", RECURSED"); 1132 } 1133 db_printf("}\n"); 1134 if (!mtx_unowned(m) && !mtx_destroyed(m)) { 1135 td = mtx_owner(m); 1136 db_printf(" owner: %p (tid %d, pid %d, \"%s\")\n", td, 1137 td->td_tid, td->td_proc->p_pid, td->td_name); 1138 if (mtx_recursed(m)) 1139 db_printf(" recursed: %d\n", m->mtx_recurse); 1140 } 1141 } 1142 #endif 1143