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 42 #include <sys/param.h> 43 #include <sys/systm.h> 44 #include <sys/bus.h> 45 #include <sys/kernel.h> 46 #include <sys/ktr.h> 47 #include <sys/lock.h> 48 #include <sys/malloc.h> 49 #include <sys/mutex.h> 50 #include <sys/proc.h> 51 #include <sys/resourcevar.h> 52 #include <sys/sched.h> 53 #include <sys/sbuf.h> 54 #include <sys/sysctl.h> 55 #include <sys/turnstile.h> 56 #include <sys/vmmeter.h> 57 58 #include <machine/atomic.h> 59 #include <machine/bus.h> 60 #include <machine/clock.h> 61 #include <machine/cpu.h> 62 63 #include <ddb/ddb.h> 64 65 #include <vm/vm.h> 66 #include <vm/vm_extern.h> 67 68 /* 69 * Internal utility macros. 70 */ 71 #define mtx_unowned(m) ((m)->mtx_lock == MTX_UNOWNED) 72 73 #define mtx_owner(m) (mtx_unowned((m)) ? NULL \ 74 : (struct thread *)((m)->mtx_lock & MTX_FLAGMASK)) 75 76 /* 77 * Lock classes for sleep and spin mutexes. 78 */ 79 struct lock_class lock_class_mtx_sleep = { 80 "sleep mutex", 81 LC_SLEEPLOCK | LC_RECURSABLE 82 }; 83 struct lock_class lock_class_mtx_spin = { 84 "spin mutex", 85 LC_SPINLOCK | LC_RECURSABLE 86 }; 87 88 /* 89 * System-wide mutexes 90 */ 91 struct mtx sched_lock; 92 struct mtx Giant; 93 94 #ifdef MUTEX_PROFILING 95 SYSCTL_NODE(_debug, OID_AUTO, mutex, CTLFLAG_RD, NULL, "mutex debugging"); 96 SYSCTL_NODE(_debug_mutex, OID_AUTO, prof, CTLFLAG_RD, NULL, "mutex profiling"); 97 static int mutex_prof_enable = 0; 98 SYSCTL_INT(_debug_mutex_prof, OID_AUTO, enable, CTLFLAG_RW, 99 &mutex_prof_enable, 0, "Enable tracing of mutex holdtime"); 100 101 struct mutex_prof { 102 const char *name; 103 const char *file; 104 int line; 105 uintmax_t cnt_max; 106 uintmax_t cnt_tot; 107 uintmax_t cnt_cur; 108 struct mutex_prof *next; 109 }; 110 111 /* 112 * mprof_buf is a static pool of profiling records to avoid possible 113 * reentrance of the memory allocation functions. 114 * 115 * Note: NUM_MPROF_BUFFERS must be smaller than MPROF_HASH_SIZE. 116 */ 117 #define NUM_MPROF_BUFFERS 1000 118 static struct mutex_prof mprof_buf[NUM_MPROF_BUFFERS]; 119 static int first_free_mprof_buf; 120 #define MPROF_HASH_SIZE 1009 121 static struct mutex_prof *mprof_hash[MPROF_HASH_SIZE]; 122 /* SWAG: sbuf size = avg stat. line size * number of locks */ 123 #define MPROF_SBUF_SIZE 256 * 400 124 125 static int mutex_prof_acquisitions; 126 SYSCTL_INT(_debug_mutex_prof, OID_AUTO, acquisitions, CTLFLAG_RD, 127 &mutex_prof_acquisitions, 0, "Number of mutex acquistions recorded"); 128 static int mutex_prof_records; 129 SYSCTL_INT(_debug_mutex_prof, OID_AUTO, records, CTLFLAG_RD, 130 &mutex_prof_records, 0, "Number of profiling records"); 131 static int mutex_prof_maxrecords = NUM_MPROF_BUFFERS; 132 SYSCTL_INT(_debug_mutex_prof, OID_AUTO, maxrecords, CTLFLAG_RD, 133 &mutex_prof_maxrecords, 0, "Maximum number of profiling records"); 134 static int mutex_prof_rejected; 135 SYSCTL_INT(_debug_mutex_prof, OID_AUTO, rejected, CTLFLAG_RD, 136 &mutex_prof_rejected, 0, "Number of rejected profiling records"); 137 static int mutex_prof_hashsize = MPROF_HASH_SIZE; 138 SYSCTL_INT(_debug_mutex_prof, OID_AUTO, hashsize, CTLFLAG_RD, 139 &mutex_prof_hashsize, 0, "Hash size"); 140 static int mutex_prof_collisions = 0; 141 SYSCTL_INT(_debug_mutex_prof, OID_AUTO, collisions, CTLFLAG_RD, 142 &mutex_prof_collisions, 0, "Number of hash collisions"); 143 144 /* 145 * mprof_mtx protects the profiling buffers and the hash. 146 */ 147 static struct mtx mprof_mtx; 148 MTX_SYSINIT(mprof, &mprof_mtx, "mutex profiling lock", MTX_SPIN | MTX_QUIET); 149 150 static u_int64_t 151 nanoseconds(void) 152 { 153 struct timespec tv; 154 155 nanotime(&tv); 156 return (tv.tv_sec * (u_int64_t)1000000000 + tv.tv_nsec); 157 } 158 159 static int 160 dump_mutex_prof_stats(SYSCTL_HANDLER_ARGS) 161 { 162 struct sbuf *sb; 163 int error, i; 164 static int multiplier = 1; 165 166 if (first_free_mprof_buf == 0) 167 return (SYSCTL_OUT(req, "No locking recorded", 168 sizeof("No locking recorded"))); 169 170 retry_sbufops: 171 sb = sbuf_new(NULL, NULL, MPROF_SBUF_SIZE * multiplier, SBUF_FIXEDLEN); 172 sbuf_printf(sb, "%6s %12s %11s %5s %s\n", 173 "max", "total", "count", "avg", "name"); 174 /* 175 * XXX this spinlock seems to be by far the largest perpetrator 176 * of spinlock latency (1.6 msec on an Athlon1600 was recorded 177 * even before I pessimized it further by moving the average 178 * computation here). 179 */ 180 mtx_lock_spin(&mprof_mtx); 181 for (i = 0; i < first_free_mprof_buf; ++i) { 182 sbuf_printf(sb, "%6ju %12ju %11ju %5ju %s:%d (%s)\n", 183 mprof_buf[i].cnt_max / 1000, 184 mprof_buf[i].cnt_tot / 1000, 185 mprof_buf[i].cnt_cur, 186 mprof_buf[i].cnt_cur == 0 ? (uintmax_t)0 : 187 mprof_buf[i].cnt_tot / (mprof_buf[i].cnt_cur * 1000), 188 mprof_buf[i].file, mprof_buf[i].line, mprof_buf[i].name); 189 if (sbuf_overflowed(sb)) { 190 mtx_unlock_spin(&mprof_mtx); 191 sbuf_delete(sb); 192 multiplier++; 193 goto retry_sbufops; 194 } 195 } 196 mtx_unlock_spin(&mprof_mtx); 197 sbuf_finish(sb); 198 error = SYSCTL_OUT(req, sbuf_data(sb), sbuf_len(sb) + 1); 199 sbuf_delete(sb); 200 return (error); 201 } 202 SYSCTL_PROC(_debug_mutex_prof, OID_AUTO, stats, CTLTYPE_STRING | CTLFLAG_RD, 203 NULL, 0, dump_mutex_prof_stats, "A", "Mutex profiling statistics"); 204 #endif 205 206 /* 207 * Function versions of the inlined __mtx_* macros. These are used by 208 * modules and can also be called from assembly language if needed. 209 */ 210 void 211 _mtx_lock_flags(struct mtx *m, int opts, const char *file, int line) 212 { 213 214 MPASS(curthread != NULL); 215 KASSERT(m->mtx_object.lo_class == &lock_class_mtx_sleep, 216 ("mtx_lock() of spin mutex %s @ %s:%d", m->mtx_object.lo_name, 217 file, line)); 218 _get_sleep_lock(m, curthread, opts, file, line); 219 LOCK_LOG_LOCK("LOCK", &m->mtx_object, opts, m->mtx_recurse, file, 220 line); 221 WITNESS_LOCK(&m->mtx_object, opts | LOP_EXCLUSIVE, file, line); 222 #ifdef MUTEX_PROFILING 223 /* don't reset the timer when/if recursing */ 224 if (m->mtx_acqtime == 0) { 225 m->mtx_filename = file; 226 m->mtx_lineno = line; 227 m->mtx_acqtime = mutex_prof_enable ? nanoseconds() : 0; 228 ++mutex_prof_acquisitions; 229 } 230 #endif 231 } 232 233 void 234 _mtx_unlock_flags(struct mtx *m, int opts, const char *file, int line) 235 { 236 237 MPASS(curthread != NULL); 238 KASSERT(m->mtx_object.lo_class == &lock_class_mtx_sleep, 239 ("mtx_unlock() of spin mutex %s @ %s:%d", m->mtx_object.lo_name, 240 file, line)); 241 WITNESS_UNLOCK(&m->mtx_object, opts | LOP_EXCLUSIVE, file, line); 242 LOCK_LOG_LOCK("UNLOCK", &m->mtx_object, opts, m->mtx_recurse, file, 243 line); 244 mtx_assert(m, MA_OWNED); 245 #ifdef MUTEX_PROFILING 246 if (m->mtx_acqtime != 0) { 247 static const char *unknown = "(unknown)"; 248 struct mutex_prof *mpp; 249 u_int64_t acqtime, now; 250 const char *p, *q; 251 volatile u_int hash; 252 253 now = nanoseconds(); 254 acqtime = m->mtx_acqtime; 255 m->mtx_acqtime = 0; 256 if (now <= acqtime) 257 goto out; 258 for (p = m->mtx_filename; 259 p != NULL && strncmp(p, "../", 3) == 0; p += 3) 260 /* nothing */ ; 261 if (p == NULL || *p == '\0') 262 p = unknown; 263 for (hash = m->mtx_lineno, q = p; *q != '\0'; ++q) 264 hash = (hash * 2 + *q) % MPROF_HASH_SIZE; 265 mtx_lock_spin(&mprof_mtx); 266 for (mpp = mprof_hash[hash]; mpp != NULL; mpp = mpp->next) 267 if (mpp->line == m->mtx_lineno && 268 strcmp(mpp->file, p) == 0) 269 break; 270 if (mpp == NULL) { 271 /* Just exit if we cannot get a trace buffer */ 272 if (first_free_mprof_buf >= NUM_MPROF_BUFFERS) { 273 ++mutex_prof_rejected; 274 goto unlock; 275 } 276 mpp = &mprof_buf[first_free_mprof_buf++]; 277 mpp->name = mtx_name(m); 278 mpp->file = p; 279 mpp->line = m->mtx_lineno; 280 mpp->next = mprof_hash[hash]; 281 if (mprof_hash[hash] != NULL) 282 ++mutex_prof_collisions; 283 mprof_hash[hash] = mpp; 284 ++mutex_prof_records; 285 } 286 /* 287 * Record if the mutex has been held longer now than ever 288 * before. 289 */ 290 if (now - acqtime > mpp->cnt_max) 291 mpp->cnt_max = now - acqtime; 292 mpp->cnt_tot += now - acqtime; 293 mpp->cnt_cur++; 294 unlock: 295 mtx_unlock_spin(&mprof_mtx); 296 } 297 out: 298 #endif 299 _rel_sleep_lock(m, curthread, opts, file, line); 300 } 301 302 void 303 _mtx_lock_spin_flags(struct mtx *m, int opts, const char *file, int line) 304 { 305 306 MPASS(curthread != NULL); 307 KASSERT(m->mtx_object.lo_class == &lock_class_mtx_spin, 308 ("mtx_lock_spin() of sleep mutex %s @ %s:%d", 309 m->mtx_object.lo_name, file, line)); 310 #if defined(SMP) || LOCK_DEBUG > 0 || 1 311 _get_spin_lock(m, curthread, opts, file, line); 312 #else 313 critical_enter(); 314 #endif 315 LOCK_LOG_LOCK("LOCK", &m->mtx_object, opts, m->mtx_recurse, file, 316 line); 317 WITNESS_LOCK(&m->mtx_object, opts | LOP_EXCLUSIVE, file, line); 318 } 319 320 void 321 _mtx_unlock_spin_flags(struct mtx *m, int opts, const char *file, int line) 322 { 323 324 MPASS(curthread != NULL); 325 KASSERT(m->mtx_object.lo_class == &lock_class_mtx_spin, 326 ("mtx_unlock_spin() of sleep mutex %s @ %s:%d", 327 m->mtx_object.lo_name, file, line)); 328 WITNESS_UNLOCK(&m->mtx_object, opts | LOP_EXCLUSIVE, file, line); 329 LOCK_LOG_LOCK("UNLOCK", &m->mtx_object, opts, m->mtx_recurse, file, 330 line); 331 mtx_assert(m, MA_OWNED); 332 #if defined(SMP) || LOCK_DEBUG > 0 || 1 333 _rel_spin_lock(m); 334 #else 335 critical_exit(); 336 #endif 337 } 338 339 /* 340 * The important part of mtx_trylock{,_flags}() 341 * Tries to acquire lock `m.' If this function is called on a mutex that 342 * is already owned, it will recursively acquire the lock. 343 */ 344 int 345 _mtx_trylock(struct mtx *m, int opts, const char *file, int line) 346 { 347 int rval; 348 349 MPASS(curthread != NULL); 350 351 if (mtx_owned(m) && (m->mtx_object.lo_flags & LO_RECURSABLE) != 0) { 352 m->mtx_recurse++; 353 atomic_set_ptr(&m->mtx_lock, MTX_RECURSED); 354 rval = 1; 355 } else 356 rval = _obtain_lock(m, curthread); 357 358 LOCK_LOG_TRY("LOCK", &m->mtx_object, opts, rval, file, line); 359 if (rval) 360 WITNESS_LOCK(&m->mtx_object, opts | LOP_EXCLUSIVE | LOP_TRYLOCK, 361 file, line); 362 363 return (rval); 364 } 365 366 /* 367 * _mtx_lock_sleep: the tougher part of acquiring an MTX_DEF lock. 368 * 369 * We call this if the lock is either contested (i.e. we need to go to 370 * sleep waiting for it), or if we need to recurse on it. 371 */ 372 void 373 _mtx_lock_sleep(struct mtx *m, int opts, const char *file, int line) 374 { 375 struct turnstile *ts; 376 struct thread *td = curthread; 377 #if defined(SMP) && defined(ADAPTIVE_MUTEXES) 378 struct thread *owner; 379 #endif 380 uintptr_t v; 381 #ifdef KTR 382 int cont_logged = 0; 383 #endif 384 385 if (mtx_owned(m)) { 386 KASSERT((m->mtx_object.lo_flags & LO_RECURSABLE) != 0, 387 ("_mtx_lock_sleep: recursed on non-recursive mutex %s @ %s:%d\n", 388 m->mtx_object.lo_name, file, line)); 389 m->mtx_recurse++; 390 atomic_set_ptr(&m->mtx_lock, MTX_RECURSED); 391 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 392 CTR1(KTR_LOCK, "_mtx_lock_sleep: %p recursing", m); 393 return; 394 } 395 396 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 397 CTR4(KTR_LOCK, 398 "_mtx_lock_sleep: %s contested (lock=%p) at %s:%d", 399 m->mtx_object.lo_name, (void *)m->mtx_lock, file, line); 400 401 while (!_obtain_lock(m, td)) { 402 403 ts = turnstile_lookup(&m->mtx_object); 404 v = m->mtx_lock; 405 406 /* 407 * Check if the lock has been released while spinning for 408 * the turnstile chain lock. 409 */ 410 if (v == MTX_UNOWNED) { 411 turnstile_release(&m->mtx_object); 412 #ifdef __i386__ 413 ia32_pause(); 414 #endif 415 continue; 416 } 417 418 /* 419 * The mutex was marked contested on release. This means that 420 * there are other threads blocked on it. Grab ownership of 421 * it and propagate its priority to the current thread if 422 * necessary. 423 */ 424 if (v == MTX_CONTESTED) { 425 MPASS(ts != NULL); 426 m->mtx_lock = (uintptr_t)td | MTX_CONTESTED; 427 turnstile_claim(ts); 428 return; 429 } 430 431 /* 432 * If the mutex isn't already contested and a failure occurs 433 * setting the contested bit, the mutex was either released 434 * or the state of the MTX_RECURSED bit changed. 435 */ 436 if ((v & MTX_CONTESTED) == 0 && 437 !atomic_cmpset_ptr(&m->mtx_lock, (void *)v, 438 (void *)(v | MTX_CONTESTED))) { 439 turnstile_release(&m->mtx_object); 440 #ifdef __i386__ 441 ia32_pause(); 442 #endif 443 continue; 444 } 445 446 #if defined(SMP) && defined(ADAPTIVE_MUTEXES) 447 /* 448 * If the current owner of the lock is executing on another 449 * CPU, spin instead of blocking. 450 */ 451 owner = (struct thread *)(v & MTX_FLAGMASK); 452 if (m != &Giant && TD_IS_RUNNING(owner)) { 453 turnstile_release(&m->mtx_object); 454 while (mtx_owner(m) == owner && TD_IS_RUNNING(owner)) { 455 #ifdef __i386__ 456 ia32_pause(); 457 #endif 458 } 459 continue; 460 } 461 #endif /* SMP && ADAPTIVE_MUTEXES */ 462 463 /* 464 * We definitely must sleep for this lock. 465 */ 466 mtx_assert(m, MA_NOTOWNED); 467 468 #ifdef KTR 469 if (!cont_logged) { 470 CTR6(KTR_CONTENTION, 471 "contention: %p at %s:%d wants %s, taken by %s:%d", 472 td, file, line, m->mtx_object.lo_name, 473 WITNESS_FILE(&m->mtx_object), 474 WITNESS_LINE(&m->mtx_object)); 475 cont_logged = 1; 476 } 477 #endif 478 479 /* 480 * Block on the turnstile. 481 */ 482 turnstile_wait(ts, &m->mtx_object, mtx_owner(m)); 483 } 484 485 #ifdef KTR 486 if (cont_logged) { 487 CTR4(KTR_CONTENTION, 488 "contention end: %s acquired by %p at %s:%d", 489 m->mtx_object.lo_name, td, file, line); 490 } 491 #endif 492 return; 493 } 494 495 /* 496 * _mtx_lock_spin: the tougher part of acquiring an MTX_SPIN lock. 497 * 498 * This is only called if we need to actually spin for the lock. Recursion 499 * is handled inline. 500 */ 501 void 502 _mtx_lock_spin(struct mtx *m, int opts, const char *file, int line) 503 { 504 int i = 0; 505 506 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 507 CTR1(KTR_LOCK, "_mtx_lock_spin: %p spinning", m); 508 509 for (;;) { 510 if (_obtain_lock(m, curthread)) 511 break; 512 513 /* Give interrupts a chance while we spin. */ 514 critical_exit(); 515 while (m->mtx_lock != MTX_UNOWNED) { 516 if (i++ < 10000000) { 517 #ifdef __i386__ 518 ia32_pause(); 519 #endif 520 continue; 521 } 522 if (i < 60000000) 523 DELAY(1); 524 #ifdef DDB 525 else if (!db_active) { 526 #else 527 else { 528 #endif 529 printf("spin lock %s held by %p for > 5 seconds\n", 530 m->mtx_object.lo_name, (void *)m->mtx_lock); 531 #ifdef WITNESS 532 witness_display_spinlock(&m->mtx_object, 533 mtx_owner(m)); 534 #endif 535 panic("spin lock held too long"); 536 } 537 #ifdef __i386__ 538 ia32_pause(); 539 #endif 540 } 541 critical_enter(); 542 } 543 544 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 545 CTR1(KTR_LOCK, "_mtx_lock_spin: %p spin done", m); 546 547 return; 548 } 549 550 /* 551 * _mtx_unlock_sleep: the tougher part of releasing an MTX_DEF lock. 552 * 553 * We are only called here if the lock is recursed or contested (i.e. we 554 * need to wake up a blocked thread). 555 */ 556 void 557 _mtx_unlock_sleep(struct mtx *m, int opts, const char *file, int line) 558 { 559 struct turnstile *ts; 560 struct thread *td, *td1; 561 562 if (mtx_recursed(m)) { 563 if (--(m->mtx_recurse) == 0) 564 atomic_clear_ptr(&m->mtx_lock, MTX_RECURSED); 565 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 566 CTR1(KTR_LOCK, "_mtx_unlock_sleep: %p unrecurse", m); 567 return; 568 } 569 570 ts = turnstile_lookup(&m->mtx_object); 571 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 572 CTR1(KTR_LOCK, "_mtx_unlock_sleep: %p contested", m); 573 574 #if defined(SMP) && defined(ADAPTIVE_MUTEXES) 575 if (ts == NULL) { 576 _release_lock_quick(m); 577 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 578 CTR1(KTR_LOCK, "_mtx_unlock_sleep: %p no sleepers", m); 579 turnstile_release(&m->mtx_object); 580 return; 581 } 582 #else 583 MPASS(ts != NULL); 584 #endif 585 /* XXX */ 586 td1 = turnstile_head(ts); 587 if (turnstile_signal(ts)) { 588 _release_lock_quick(m); 589 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 590 CTR1(KTR_LOCK, "_mtx_unlock_sleep: %p not held", m); 591 } else { 592 m->mtx_lock = MTX_CONTESTED; 593 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 594 CTR1(KTR_LOCK, "_mtx_unlock_sleep: %p still contested", 595 m); 596 } 597 turnstile_unpend(ts); 598 599 /* 600 * XXX: This is just a hack until preemption is done. However, 601 * once preemption is done we need to either wrap the 602 * turnstile_signal() and release of the actual lock in an 603 * extra critical section or change the preemption code to 604 * always just set a flag and never do instant-preempts. 605 */ 606 td = curthread; 607 if (td->td_critnest > 0 || td1->td_priority >= td->td_priority) 608 return; 609 mtx_lock_spin(&sched_lock); 610 if (!TD_IS_RUNNING(td1)) { 611 #ifdef notyet 612 if (td->td_ithd != NULL) { 613 struct ithd *it = td->td_ithd; 614 615 if (it->it_interrupted) { 616 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 617 CTR2(KTR_LOCK, 618 "_mtx_unlock_sleep: %p interrupted %p", 619 it, it->it_interrupted); 620 intr_thd_fixup(it); 621 } 622 } 623 #endif 624 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 625 CTR2(KTR_LOCK, 626 "_mtx_unlock_sleep: %p switching out lock=%p", m, 627 (void *)m->mtx_lock); 628 629 td->td_proc->p_stats->p_ru.ru_nivcsw++; 630 mi_switch(); 631 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 632 CTR2(KTR_LOCK, "_mtx_unlock_sleep: %p resuming lock=%p", 633 m, (void *)m->mtx_lock); 634 } 635 mtx_unlock_spin(&sched_lock); 636 637 return; 638 } 639 640 /* 641 * All the unlocking of MTX_SPIN locks is done inline. 642 * See the _rel_spin_lock() macro for the details. 643 */ 644 645 /* 646 * The backing function for the INVARIANTS-enabled mtx_assert() 647 */ 648 #ifdef INVARIANT_SUPPORT 649 void 650 _mtx_assert(struct mtx *m, int what, const char *file, int line) 651 { 652 653 if (panicstr != NULL) 654 return; 655 switch (what) { 656 case MA_OWNED: 657 case MA_OWNED | MA_RECURSED: 658 case MA_OWNED | MA_NOTRECURSED: 659 if (!mtx_owned(m)) 660 panic("mutex %s not owned at %s:%d", 661 m->mtx_object.lo_name, file, line); 662 if (mtx_recursed(m)) { 663 if ((what & MA_NOTRECURSED) != 0) 664 panic("mutex %s recursed at %s:%d", 665 m->mtx_object.lo_name, file, line); 666 } else if ((what & MA_RECURSED) != 0) { 667 panic("mutex %s unrecursed at %s:%d", 668 m->mtx_object.lo_name, file, line); 669 } 670 break; 671 case MA_NOTOWNED: 672 if (mtx_owned(m)) 673 panic("mutex %s owned at %s:%d", 674 m->mtx_object.lo_name, file, line); 675 break; 676 default: 677 panic("unknown mtx_assert at %s:%d", file, line); 678 } 679 } 680 #endif 681 682 /* 683 * The MUTEX_DEBUG-enabled mtx_validate() 684 * 685 * Most of these checks have been moved off into the LO_INITIALIZED flag 686 * maintained by the witness code. 687 */ 688 #ifdef MUTEX_DEBUG 689 690 void mtx_validate(struct mtx *); 691 692 void 693 mtx_validate(struct mtx *m) 694 { 695 696 /* 697 * XXX: When kernacc() does not require Giant we can reenable this check 698 */ 699 #ifdef notyet 700 /* 701 * XXX - When kernacc() is fixed on the alpha to handle K0_SEG memory properly 702 * we can re-enable the kernacc() checks. 703 */ 704 #ifndef __alpha__ 705 /* 706 * Can't call kernacc() from early init386(), especially when 707 * initializing Giant mutex, because some stuff in kernacc() 708 * requires Giant itself. 709 */ 710 if (!cold) 711 if (!kernacc((caddr_t)m, sizeof(m), 712 VM_PROT_READ | VM_PROT_WRITE)) 713 panic("Can't read and write to mutex %p", m); 714 #endif 715 #endif 716 } 717 #endif 718 719 /* 720 * General init routine used by the MTX_SYSINIT() macro. 721 */ 722 void 723 mtx_sysinit(void *arg) 724 { 725 struct mtx_args *margs = arg; 726 727 mtx_init(margs->ma_mtx, margs->ma_desc, NULL, margs->ma_opts); 728 } 729 730 /* 731 * Mutex initialization routine; initialize lock `m' of type contained in 732 * `opts' with options contained in `opts' and name `name.' The optional 733 * lock type `type' is used as a general lock category name for use with 734 * witness. 735 */ 736 void 737 mtx_init(struct mtx *m, const char *name, const char *type, int opts) 738 { 739 struct lock_object *lock; 740 741 MPASS((opts & ~(MTX_SPIN | MTX_QUIET | MTX_RECURSE | 742 MTX_NOWITNESS | MTX_DUPOK)) == 0); 743 744 #ifdef MUTEX_DEBUG 745 /* Diagnostic and error correction */ 746 mtx_validate(m); 747 #endif 748 749 lock = &m->mtx_object; 750 KASSERT((lock->lo_flags & LO_INITIALIZED) == 0, 751 ("mutex \"%s\" %p already initialized", name, m)); 752 bzero(m, sizeof(*m)); 753 if (opts & MTX_SPIN) 754 lock->lo_class = &lock_class_mtx_spin; 755 else 756 lock->lo_class = &lock_class_mtx_sleep; 757 lock->lo_name = name; 758 lock->lo_type = type != NULL ? type : name; 759 if (opts & MTX_QUIET) 760 lock->lo_flags = LO_QUIET; 761 if (opts & MTX_RECURSE) 762 lock->lo_flags |= LO_RECURSABLE; 763 if ((opts & MTX_NOWITNESS) == 0) 764 lock->lo_flags |= LO_WITNESS; 765 if (opts & MTX_DUPOK) 766 lock->lo_flags |= LO_DUPOK; 767 768 m->mtx_lock = MTX_UNOWNED; 769 770 LOCK_LOG_INIT(lock, opts); 771 772 WITNESS_INIT(lock); 773 } 774 775 /* 776 * Remove lock `m' from all_mtx queue. We don't allow MTX_QUIET to be 777 * passed in as a flag here because if the corresponding mtx_init() was 778 * called with MTX_QUIET set, then it will already be set in the mutex's 779 * flags. 780 */ 781 void 782 mtx_destroy(struct mtx *m) 783 { 784 785 LOCK_LOG_DESTROY(&m->mtx_object, 0); 786 787 if (!mtx_owned(m)) 788 MPASS(mtx_unowned(m)); 789 else { 790 MPASS((m->mtx_lock & (MTX_RECURSED|MTX_CONTESTED)) == 0); 791 792 /* Tell witness this isn't locked to make it happy. */ 793 WITNESS_UNLOCK(&m->mtx_object, LOP_EXCLUSIVE, __FILE__, 794 __LINE__); 795 } 796 797 WITNESS_DESTROY(&m->mtx_object); 798 } 799 800 /* 801 * Intialize the mutex code and system mutexes. This is called from the MD 802 * startup code prior to mi_startup(). The per-CPU data space needs to be 803 * setup before this is called. 804 */ 805 void 806 mutex_init(void) 807 { 808 809 /* Setup thread0 so that mutexes work. */ 810 LIST_INIT(&thread0.td_contested); 811 812 /* Setup turnstiles so that sleep mutexes work. */ 813 init_turnstiles(); 814 815 /* 816 * Initialize mutexes. 817 */ 818 mtx_init(&Giant, "Giant", NULL, MTX_DEF | MTX_RECURSE); 819 mtx_init(&sched_lock, "sched lock", NULL, MTX_SPIN | MTX_RECURSE); 820 mtx_init(&proc0.p_mtx, "process lock", NULL, MTX_DEF | MTX_DUPOK); 821 mtx_lock(&Giant); 822 } 823