kern_mutex.c revision 97837
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 * $FreeBSD: head/sys/kern/kern_mutex.c 97837 2002-06-04 21:53:48Z jhb $ 31 */ 32 33/* 34 * Machine independent bits of mutex implementation. 35 */ 36 37#include "opt_adaptive_mutexes.h" 38#include "opt_ddb.h" 39 40#include <sys/param.h> 41#include <sys/systm.h> 42#include <sys/bus.h> 43#include <sys/kernel.h> 44#include <sys/ktr.h> 45#include <sys/lock.h> 46#include <sys/malloc.h> 47#include <sys/mutex.h> 48#include <sys/proc.h> 49#include <sys/resourcevar.h> 50#include <sys/sbuf.h> 51#include <sys/stdint.h> 52#include <sys/sysctl.h> 53#include <sys/vmmeter.h> 54 55#include <machine/atomic.h> 56#include <machine/bus.h> 57#include <machine/clock.h> 58#include <machine/cpu.h> 59 60#include <ddb/ddb.h> 61 62#include <vm/vm.h> 63#include <vm/vm_extern.h> 64 65/* 66 * Internal utility macros. 67 */ 68#define mtx_unowned(m) ((m)->mtx_lock == MTX_UNOWNED) 69 70#define mtx_owner(m) (mtx_unowned((m)) ? NULL \ 71 : (struct thread *)((m)->mtx_lock & MTX_FLAGMASK)) 72 73#define thread_runnable(td) \ 74 ((td)->td_kse != NULL && (td)->td_kse->ke_oncpu != NOCPU) 75 76/* 77 * Lock classes for sleep and spin mutexes. 78 */ 79struct lock_class lock_class_mtx_sleep = { 80 "sleep mutex", 81 LC_SLEEPLOCK | LC_RECURSABLE 82}; 83struct lock_class lock_class_mtx_spin = { 84 "spin mutex", 85 LC_SPINLOCK | LC_RECURSABLE 86}; 87 88/* 89 * System-wide mutexes 90 */ 91struct mtx sched_lock; 92struct mtx Giant; 93 94/* 95 * Prototypes for non-exported routines. 96 */ 97static void propagate_priority(struct thread *); 98 99static void 100propagate_priority(struct thread *td) 101{ 102 int pri = td->td_priority; 103 struct mtx *m = td->td_blocked; 104 105 mtx_assert(&sched_lock, MA_OWNED); 106 for (;;) { 107 struct thread *td1; 108 109 td = mtx_owner(m); 110 111 if (td == NULL) { 112 /* 113 * This really isn't quite right. Really 114 * ought to bump priority of thread that 115 * next acquires the mutex. 116 */ 117 MPASS(m->mtx_lock == MTX_CONTESTED); 118 return; 119 } 120 121 MPASS(td->td_proc->p_magic == P_MAGIC); 122 KASSERT(td->td_proc->p_stat != SSLEEP, ("sleeping thread owns a mutex")); 123 if (td->td_priority <= pri) /* lower is higher priority */ 124 return; 125 126 /* 127 * Bump this thread's priority. 128 */ 129 td->td_priority = pri; 130 131 /* 132 * If lock holder is actually running, just bump priority. 133 */ 134 /* XXXKSE this test is not sufficient */ 135 if (thread_runnable(td)) { 136 MPASS(td->td_proc->p_stat == SRUN 137 || td->td_proc->p_stat == SZOMB 138 || td->td_proc->p_stat == SSTOP); 139 return; 140 } 141 142#ifndef SMP 143 /* 144 * For UP, we check to see if td is curthread (this shouldn't 145 * ever happen however as it would mean we are in a deadlock.) 146 */ 147 KASSERT(td != curthread, ("Deadlock detected")); 148#endif 149 150 /* 151 * If on run queue move to new run queue, and quit. 152 * XXXKSE this gets a lot more complicated under threads 153 * but try anyhow. 154 */ 155 if (td->td_proc->p_stat == SRUN) { 156 MPASS(td->td_blocked == NULL); 157 remrunqueue(td); 158 setrunqueue(td); 159 return; 160 } 161 162 /* 163 * If we aren't blocked on a mutex, we should be. 164 */ 165 KASSERT(td->td_proc->p_stat == SMTX, ( 166 "process %d(%s):%d holds %s but isn't blocked on a mutex\n", 167 td->td_proc->p_pid, td->td_proc->p_comm, td->td_proc->p_stat, 168 m->mtx_object.lo_name)); 169 170 /* 171 * Pick up the mutex that td is blocked on. 172 */ 173 m = td->td_blocked; 174 MPASS(m != NULL); 175 176 /* 177 * Check if the thread needs to be moved up on 178 * the blocked chain 179 */ 180 if (td == TAILQ_FIRST(&m->mtx_blocked)) { 181 continue; 182 } 183 184 td1 = TAILQ_PREV(td, threadqueue, td_blkq); 185 if (td1->td_priority <= pri) { 186 continue; 187 } 188 189 /* 190 * Remove thread from blocked chain and determine where 191 * it should be moved up to. Since we know that td1 has 192 * a lower priority than td, we know that at least one 193 * thread in the chain has a lower priority and that 194 * td1 will thus not be NULL after the loop. 195 */ 196 TAILQ_REMOVE(&m->mtx_blocked, td, td_blkq); 197 TAILQ_FOREACH(td1, &m->mtx_blocked, td_blkq) { 198 MPASS(td1->td_proc->p_magic == P_MAGIC); 199 if (td1->td_priority > pri) 200 break; 201 } 202 203 MPASS(td1 != NULL); 204 TAILQ_INSERT_BEFORE(td1, td, td_blkq); 205 CTR4(KTR_LOCK, 206 "propagate_priority: p %p moved before %p on [%p] %s", 207 td, td1, m, m->mtx_object.lo_name); 208 } 209} 210 211#ifdef MUTEX_PROFILING 212SYSCTL_NODE(_debug, OID_AUTO, mutex, CTLFLAG_RD, NULL, "mutex debugging"); 213SYSCTL_NODE(_debug_mutex, OID_AUTO, prof, CTLFLAG_RD, NULL, "mutex profiling"); 214static int mutex_prof_enable = 0; 215SYSCTL_INT(_debug_mutex_prof, OID_AUTO, enable, CTLFLAG_RW, 216 &mutex_prof_enable, 0, "Enable tracing of mutex holdtime"); 217 218struct mutex_prof { 219 const char *name; 220 const char *file; 221 int line; 222#define MPROF_MAX 0 223#define MPROF_TOT 1 224#define MPROF_CNT 2 225#define MPROF_AVG 3 226 uintmax_t counter[4]; 227 struct mutex_prof *next; 228}; 229 230/* 231 * mprof_buf is a static pool of profiling records to avoid possible 232 * reentrance of the memory allocation functions. 233 * 234 * Note: NUM_MPROF_BUFFERS must be smaller than MPROF_HASH_SIZE. 235 */ 236#define NUM_MPROF_BUFFERS 1000 237static struct mutex_prof mprof_buf[NUM_MPROF_BUFFERS]; 238static int first_free_mprof_buf; 239#define MPROF_HASH_SIZE 1009 240static struct mutex_prof *mprof_hash[MPROF_HASH_SIZE]; 241 242static int mutex_prof_acquisitions; 243SYSCTL_INT(_debug_mutex_prof, OID_AUTO, acquisitions, CTLFLAG_RD, 244 &mutex_prof_acquisitions, 0, "Number of mutex acquistions recorded"); 245static int mutex_prof_records; 246SYSCTL_INT(_debug_mutex_prof, OID_AUTO, records, CTLFLAG_RD, 247 &mutex_prof_records, 0, "Number of profiling records"); 248static int mutex_prof_maxrecords = NUM_MPROF_BUFFERS; 249SYSCTL_INT(_debug_mutex_prof, OID_AUTO, maxrecords, CTLFLAG_RD, 250 &mutex_prof_maxrecords, 0, "Maximum number of profiling records"); 251static int mutex_prof_rejected; 252SYSCTL_INT(_debug_mutex_prof, OID_AUTO, rejected, CTLFLAG_RD, 253 &mutex_prof_rejected, 0, "Number of rejected profiling records"); 254static int mutex_prof_hashsize = MPROF_HASH_SIZE; 255SYSCTL_INT(_debug_mutex_prof, OID_AUTO, hashsize, CTLFLAG_RD, 256 &mutex_prof_hashsize, 0, "Hash size"); 257static int mutex_prof_collisions = 0; 258SYSCTL_INT(_debug_mutex_prof, OID_AUTO, collisions, CTLFLAG_RD, 259 &mutex_prof_collisions, 0, "Number of hash collisions"); 260 261/* 262 * mprof_mtx protects the profiling buffers and the hash. 263 */ 264static struct mtx mprof_mtx; 265MTX_SYSINIT(mprof, &mprof_mtx, "mutex profiling lock", MTX_SPIN | MTX_QUIET); 266 267static u_int64_t 268nanoseconds(void) 269{ 270 struct timespec tv; 271 272 nanotime(&tv); 273 return (tv.tv_sec * (u_int64_t)1000000000 + tv.tv_nsec); 274} 275 276static int 277dump_mutex_prof_stats(SYSCTL_HANDLER_ARGS) 278{ 279 struct sbuf *sb; 280 int error, i; 281 282 if (first_free_mprof_buf == 0) 283 return SYSCTL_OUT(req, "No locking recorded", 284 sizeof("No locking recorded")); 285 286 sb = sbuf_new(NULL, NULL, 1024, SBUF_AUTOEXTEND); 287 sbuf_printf(sb, "%12s %12s %12s %12s %s\n", 288 "max", "total", "count", "average", "name"); 289 mtx_lock_spin(&mprof_mtx); 290 for (i = 0; i < first_free_mprof_buf; ++i) 291 sbuf_printf(sb, "%12ju %12ju %12ju %12ju %s:%d (%s)\n", 292 mprof_buf[i].counter[MPROF_MAX] / 1000, 293 mprof_buf[i].counter[MPROF_TOT] / 1000, 294 mprof_buf[i].counter[MPROF_CNT], 295 mprof_buf[i].counter[MPROF_AVG] / 1000, 296 mprof_buf[i].file, mprof_buf[i].line, mprof_buf[i].name); 297 mtx_unlock_spin(&mprof_mtx); 298 sbuf_finish(sb); 299 error = SYSCTL_OUT(req, sbuf_data(sb), sbuf_len(sb) + 1); 300 sbuf_delete(sb); 301 return (error); 302} 303SYSCTL_PROC(_debug_mutex_prof, OID_AUTO, stats, CTLTYPE_STRING|CTLFLAG_RD, 304 NULL, 0, dump_mutex_prof_stats, "A", "Mutex profiling statistics"); 305#endif 306 307/* 308 * Function versions of the inlined __mtx_* macros. These are used by 309 * modules and can also be called from assembly language if needed. 310 */ 311void 312_mtx_lock_flags(struct mtx *m, int opts, const char *file, int line) 313{ 314 315 MPASS(curthread != NULL); 316 _get_sleep_lock(m, curthread, opts, file, line); 317 LOCK_LOG_LOCK("LOCK", &m->mtx_object, opts, m->mtx_recurse, file, 318 line); 319 WITNESS_LOCK(&m->mtx_object, opts | LOP_EXCLUSIVE, file, line); 320#ifdef MUTEX_PROFILING 321 /* don't reset the timer when/if recursing */ 322 if (m->acqtime == 0) { 323 m->file = file; 324 m->line = line; 325 m->acqtime = mutex_prof_enable ? nanoseconds() : 0; 326 ++mutex_prof_acquisitions; 327 } 328#endif 329} 330 331void 332_mtx_unlock_flags(struct mtx *m, int opts, const char *file, int line) 333{ 334 335 MPASS(curthread != NULL); 336 mtx_assert(m, MA_OWNED); 337#ifdef MUTEX_PROFILING 338 if (m->acqtime != 0) { 339 static const char *unknown = "(unknown)"; 340 struct mutex_prof *mpp; 341 u_int64_t acqtime, now; 342 const char *p, *q; 343 volatile u_int hash; 344 345 now = nanoseconds(); 346 acqtime = m->acqtime; 347 m->acqtime = 0; 348 if (now <= acqtime) 349 goto out; 350 for (p = file; strncmp(p, "../", 3) == 0; p += 3) 351 /* nothing */ ; 352 if (p == NULL || *p == '\0') 353 p = unknown; 354 for (hash = line, q = p; *q != '\0'; ++q) 355 hash = (hash * 2 + *q) % MPROF_HASH_SIZE; 356 mtx_lock_spin(&mprof_mtx); 357 for (mpp = mprof_hash[hash]; mpp != NULL; mpp = mpp->next) 358 if (mpp->line == line && strcmp(mpp->file, p) == 0) 359 break; 360 if (mpp == NULL) { 361 /* Just exit if we cannot get a trace buffer */ 362 if (first_free_mprof_buf >= NUM_MPROF_BUFFERS) { 363 ++mutex_prof_rejected; 364 goto unlock; 365 } 366 mpp = &mprof_buf[first_free_mprof_buf++]; 367 mpp->name = mtx_name(m); 368 mpp->file = p; 369 mpp->line = line; 370 mpp->next = mprof_hash[hash]; 371 if (mprof_hash[hash] != NULL) 372 ++mutex_prof_collisions; 373 mprof_hash[hash] = mpp; 374 ++mutex_prof_records; 375 } 376 /* 377 * Record if the mutex has been held longer now than ever 378 * before 379 */ 380 if ((now - acqtime) > mpp->counter[MPROF_MAX]) 381 mpp->counter[MPROF_MAX] = now - acqtime; 382 mpp->counter[MPROF_TOT] += now - acqtime; 383 mpp->counter[MPROF_CNT] += 1; 384 mpp->counter[MPROF_AVG] = 385 mpp->counter[MPROF_TOT] / mpp->counter[MPROF_CNT]; 386unlock: 387 mtx_unlock_spin(&mprof_mtx); 388 } 389out: 390#endif 391 WITNESS_UNLOCK(&m->mtx_object, opts | LOP_EXCLUSIVE, file, line); 392 LOCK_LOG_LOCK("UNLOCK", &m->mtx_object, opts, m->mtx_recurse, file, 393 line); 394 _rel_sleep_lock(m, curthread, opts, file, line); 395} 396 397void 398_mtx_lock_spin_flags(struct mtx *m, int opts, const char *file, int line) 399{ 400 401 MPASS(curthread != NULL); 402#if defined(SMP) || LOCK_DEBUG > 0 403 _get_spin_lock(m, curthread, opts, file, line); 404#else 405 critical_enter(); 406#endif 407 LOCK_LOG_LOCK("LOCK", &m->mtx_object, opts, m->mtx_recurse, file, 408 line); 409 WITNESS_LOCK(&m->mtx_object, opts | LOP_EXCLUSIVE, file, line); 410} 411 412void 413_mtx_unlock_spin_flags(struct mtx *m, int opts, const char *file, int line) 414{ 415 416 MPASS(curthread != NULL); 417 mtx_assert(m, MA_OWNED); 418 WITNESS_UNLOCK(&m->mtx_object, opts | LOP_EXCLUSIVE, file, line); 419 LOCK_LOG_LOCK("UNLOCK", &m->mtx_object, opts, m->mtx_recurse, file, 420 line); 421#if defined(SMP) || LOCK_DEBUG > 0 422 _rel_spin_lock(m); 423#else 424 critical_exit(); 425#endif 426} 427 428/* 429 * The important part of mtx_trylock{,_flags}() 430 * Tries to acquire lock `m.' We do NOT handle recursion here; we assume that 431 * if we're called, it's because we know we don't already own this lock. 432 */ 433int 434_mtx_trylock(struct mtx *m, int opts, const char *file, int line) 435{ 436 int rval; 437 438 MPASS(curthread != NULL); 439 440 rval = _obtain_lock(m, curthread); 441 442 LOCK_LOG_TRY("LOCK", &m->mtx_object, opts, rval, file, line); 443 if (rval) { 444 /* 445 * We do not handle recursion in _mtx_trylock; see the 446 * note at the top of the routine. 447 */ 448 KASSERT(!mtx_recursed(m), 449 ("mtx_trylock() called on a recursed mutex")); 450 WITNESS_LOCK(&m->mtx_object, opts | LOP_EXCLUSIVE | LOP_TRYLOCK, 451 file, line); 452 } 453 454 return (rval); 455} 456 457/* 458 * _mtx_lock_sleep: the tougher part of acquiring an MTX_DEF lock. 459 * 460 * We call this if the lock is either contested (i.e. we need to go to 461 * sleep waiting for it), or if we need to recurse on it. 462 */ 463void 464_mtx_lock_sleep(struct mtx *m, int opts, const char *file, int line) 465{ 466 struct thread *td = curthread; 467#if defined(SMP) && defined(ADAPTIVE_MUTEXES) 468 struct thread *owner; 469#endif 470 471 if ((m->mtx_lock & MTX_FLAGMASK) == (uintptr_t)td) { 472 m->mtx_recurse++; 473 atomic_set_ptr(&m->mtx_lock, MTX_RECURSED); 474 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 475 CTR1(KTR_LOCK, "_mtx_lock_sleep: %p recursing", m); 476 return; 477 } 478 479 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 480 CTR4(KTR_LOCK, 481 "_mtx_lock_sleep: %s contested (lock=%p) at %s:%d", 482 m->mtx_object.lo_name, (void *)m->mtx_lock, file, line); 483 484 while (!_obtain_lock(m, td)) { 485 uintptr_t v; 486 struct thread *td1; 487 488 mtx_lock_spin(&sched_lock); 489 /* 490 * Check if the lock has been released while spinning for 491 * the sched_lock. 492 */ 493 if ((v = m->mtx_lock) == MTX_UNOWNED) { 494 mtx_unlock_spin(&sched_lock); 495#ifdef __i386__ 496 ia32_pause(); 497#endif 498 continue; 499 } 500 501 /* 502 * The mutex was marked contested on release. This means that 503 * there are threads blocked on it. 504 */ 505 if (v == MTX_CONTESTED) { 506 td1 = TAILQ_FIRST(&m->mtx_blocked); 507 MPASS(td1 != NULL); 508 m->mtx_lock = (uintptr_t)td | MTX_CONTESTED; 509 510 if (td1->td_priority < td->td_priority) 511 td->td_priority = td1->td_priority; 512 mtx_unlock_spin(&sched_lock); 513 return; 514 } 515 516 /* 517 * If the mutex isn't already contested and a failure occurs 518 * setting the contested bit, the mutex was either released 519 * or the state of the MTX_RECURSED bit changed. 520 */ 521 if ((v & MTX_CONTESTED) == 0 && 522 !atomic_cmpset_ptr(&m->mtx_lock, (void *)v, 523 (void *)(v | MTX_CONTESTED))) { 524 mtx_unlock_spin(&sched_lock); 525#ifdef __i386__ 526 ia32_pause(); 527#endif 528 continue; 529 } 530 531#if defined(SMP) && defined(ADAPTIVE_MUTEXES) 532 /* 533 * If the current owner of the lock is executing on another 534 * CPU, spin instead of blocking. 535 */ 536 owner = (struct thread *)(v & MTX_FLAGMASK); 537 if (m != &Giant && thread_runnable(owner)) { 538 mtx_unlock_spin(&sched_lock); 539 while (mtx_owner(m) == owner && 540 thread_runnable(owner)) { 541#ifdef __i386__ 542 ia32_pause(); 543#endif 544 } 545 continue; 546 } 547#endif /* SMP && ADAPTIVE_MUTEXES */ 548 549 /* 550 * We definitely must sleep for this lock. 551 */ 552 mtx_assert(m, MA_NOTOWNED); 553 554#ifdef notyet 555 /* 556 * If we're borrowing an interrupted thread's VM context, we 557 * must clean up before going to sleep. 558 */ 559 if (td->td_ithd != NULL) { 560 struct ithd *it = td->td_ithd; 561 562 if (it->it_interrupted) { 563 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 564 CTR2(KTR_LOCK, 565 "_mtx_lock_sleep: %p interrupted %p", 566 it, it->it_interrupted); 567 intr_thd_fixup(it); 568 } 569 } 570#endif 571 572 /* 573 * Put us on the list of threads blocked on this mutex. 574 */ 575 if (TAILQ_EMPTY(&m->mtx_blocked)) { 576 td1 = mtx_owner(m); 577 LIST_INSERT_HEAD(&td1->td_contested, m, mtx_contested); 578 TAILQ_INSERT_TAIL(&m->mtx_blocked, td, td_blkq); 579 } else { 580 TAILQ_FOREACH(td1, &m->mtx_blocked, td_blkq) 581 if (td1->td_priority > td->td_priority) 582 break; 583 if (td1) 584 TAILQ_INSERT_BEFORE(td1, td, td_blkq); 585 else 586 TAILQ_INSERT_TAIL(&m->mtx_blocked, td, td_blkq); 587 } 588 589 /* 590 * Save who we're blocked on. 591 */ 592 td->td_blocked = m; 593 td->td_mtxname = m->mtx_object.lo_name; 594 td->td_proc->p_stat = SMTX; 595 propagate_priority(td); 596 597 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 598 CTR3(KTR_LOCK, 599 "_mtx_lock_sleep: p %p blocked on [%p] %s", td, m, 600 m->mtx_object.lo_name); 601 602 td->td_proc->p_stats->p_ru.ru_nvcsw++; 603 mi_switch(); 604 605 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 606 CTR3(KTR_LOCK, 607 "_mtx_lock_sleep: p %p free from blocked on [%p] %s", 608 td, m, m->mtx_object.lo_name); 609 610 mtx_unlock_spin(&sched_lock); 611 } 612 613 return; 614} 615 616/* 617 * _mtx_lock_spin: the tougher part of acquiring an MTX_SPIN lock. 618 * 619 * This is only called if we need to actually spin for the lock. Recursion 620 * is handled inline. 621 */ 622void 623_mtx_lock_spin(struct mtx *m, int opts, const char *file, int line) 624{ 625 int i = 0; 626 627 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 628 CTR1(KTR_LOCK, "_mtx_lock_spin: %p spinning", m); 629 630 for (;;) { 631 if (_obtain_lock(m, curthread)) 632 break; 633 634 /* Give interrupts a chance while we spin. */ 635 critical_exit(); 636 while (m->mtx_lock != MTX_UNOWNED) { 637 if (i++ < 10000000) { 638#ifdef __i386__ 639 ia32_pause(); 640#endif 641 continue; 642 } 643 if (i < 60000000) 644 DELAY(1); 645#ifdef DDB 646 else if (!db_active) 647#else 648 else 649#endif 650 panic("spin lock %s held by %p for > 5 seconds", 651 m->mtx_object.lo_name, (void *)m->mtx_lock); 652#ifdef __i386__ 653 ia32_pause(); 654#endif 655 } 656 critical_enter(); 657 } 658 659 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 660 CTR1(KTR_LOCK, "_mtx_lock_spin: %p spin done", m); 661 662 return; 663} 664 665/* 666 * _mtx_unlock_sleep: the tougher part of releasing an MTX_DEF lock. 667 * 668 * We are only called here if the lock is recursed or contested (i.e. we 669 * need to wake up a blocked thread). 670 */ 671void 672_mtx_unlock_sleep(struct mtx *m, int opts, const char *file, int line) 673{ 674 struct thread *td, *td1; 675 struct mtx *m1; 676 int pri; 677 678 td = curthread; 679 680 if (mtx_recursed(m)) { 681 if (--(m->mtx_recurse) == 0) 682 atomic_clear_ptr(&m->mtx_lock, MTX_RECURSED); 683 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 684 CTR1(KTR_LOCK, "_mtx_unlock_sleep: %p unrecurse", m); 685 return; 686 } 687 688 mtx_lock_spin(&sched_lock); 689 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 690 CTR1(KTR_LOCK, "_mtx_unlock_sleep: %p contested", m); 691 692 td1 = TAILQ_FIRST(&m->mtx_blocked); 693#if defined(SMP) && defined(ADAPTIVE_MUTEXES) 694 if (td1 == NULL) { 695 _release_lock_quick(m); 696 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 697 CTR1(KTR_LOCK, "_mtx_unlock_sleep: %p no sleepers", m); 698 mtx_unlock_spin(&sched_lock); 699 return; 700 } 701#endif 702 MPASS(td->td_proc->p_magic == P_MAGIC); 703 MPASS(td1->td_proc->p_magic == P_MAGIC); 704 705 TAILQ_REMOVE(&m->mtx_blocked, td1, td_blkq); 706 707 if (TAILQ_EMPTY(&m->mtx_blocked)) { 708 LIST_REMOVE(m, mtx_contested); 709 _release_lock_quick(m); 710 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 711 CTR1(KTR_LOCK, "_mtx_unlock_sleep: %p not held", m); 712 } else 713 atomic_store_rel_ptr(&m->mtx_lock, (void *)MTX_CONTESTED); 714 715 pri = PRI_MAX; 716 LIST_FOREACH(m1, &td->td_contested, mtx_contested) { 717 int cp = TAILQ_FIRST(&m1->mtx_blocked)->td_priority; 718 if (cp < pri) 719 pri = cp; 720 } 721 722 if (pri > td->td_base_pri) 723 pri = td->td_base_pri; 724 td->td_priority = pri; 725 726 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 727 CTR2(KTR_LOCK, "_mtx_unlock_sleep: %p contested setrunqueue %p", 728 m, td1); 729 730 td1->td_blocked = NULL; 731 td1->td_proc->p_stat = SRUN; 732 setrunqueue(td1); 733 734 if (td->td_critnest == 1 && td1->td_priority < pri) { 735#ifdef notyet 736 if (td->td_ithd != NULL) { 737 struct ithd *it = td->td_ithd; 738 739 if (it->it_interrupted) { 740 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 741 CTR2(KTR_LOCK, 742 "_mtx_unlock_sleep: %p interrupted %p", 743 it, it->it_interrupted); 744 intr_thd_fixup(it); 745 } 746 } 747#endif 748 setrunqueue(td); 749 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 750 CTR2(KTR_LOCK, 751 "_mtx_unlock_sleep: %p switching out lock=%p", m, 752 (void *)m->mtx_lock); 753 754 td->td_proc->p_stats->p_ru.ru_nivcsw++; 755 mi_switch(); 756 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 757 CTR2(KTR_LOCK, "_mtx_unlock_sleep: %p resuming lock=%p", 758 m, (void *)m->mtx_lock); 759 } 760 761 mtx_unlock_spin(&sched_lock); 762 763 return; 764} 765 766/* 767 * All the unlocking of MTX_SPIN locks is done inline. 768 * See the _rel_spin_lock() macro for the details. 769 */ 770 771/* 772 * The backing function for the INVARIANTS-enabled mtx_assert() 773 */ 774#ifdef INVARIANT_SUPPORT 775void 776_mtx_assert(struct mtx *m, int what, const char *file, int line) 777{ 778 779 if (panicstr != NULL) 780 return; 781 switch (what) { 782 case MA_OWNED: 783 case MA_OWNED | MA_RECURSED: 784 case MA_OWNED | MA_NOTRECURSED: 785 if (!mtx_owned(m)) 786 panic("mutex %s not owned at %s:%d", 787 m->mtx_object.lo_name, file, line); 788 if (mtx_recursed(m)) { 789 if ((what & MA_NOTRECURSED) != 0) 790 panic("mutex %s recursed at %s:%d", 791 m->mtx_object.lo_name, file, line); 792 } else if ((what & MA_RECURSED) != 0) { 793 panic("mutex %s unrecursed at %s:%d", 794 m->mtx_object.lo_name, file, line); 795 } 796 break; 797 case MA_NOTOWNED: 798 if (mtx_owned(m)) 799 panic("mutex %s owned at %s:%d", 800 m->mtx_object.lo_name, file, line); 801 break; 802 default: 803 panic("unknown mtx_assert at %s:%d", file, line); 804 } 805} 806#endif 807 808/* 809 * The MUTEX_DEBUG-enabled mtx_validate() 810 * 811 * Most of these checks have been moved off into the LO_INITIALIZED flag 812 * maintained by the witness code. 813 */ 814#ifdef MUTEX_DEBUG 815 816void mtx_validate(struct mtx *); 817 818void 819mtx_validate(struct mtx *m) 820{ 821 822/* 823 * XXX - When kernacc() is fixed on the alpha to handle K0_SEG memory properly 824 * we can re-enable the kernacc() checks. 825 */ 826#ifndef __alpha__ 827 /* 828 * Can't call kernacc() from early init386(), especially when 829 * initializing Giant mutex, because some stuff in kernacc() 830 * requires Giant itself. 831 */ 832 if (!cold) 833 if (!kernacc((caddr_t)m, sizeof(m), 834 VM_PROT_READ | VM_PROT_WRITE)) 835 panic("Can't read and write to mutex %p", m); 836#endif 837} 838#endif 839 840/* 841 * General init routine used by the MTX_SYSINIT() macro. 842 */ 843void 844mtx_sysinit(void *arg) 845{ 846 struct mtx_args *margs = arg; 847 848 mtx_init(margs->ma_mtx, margs->ma_desc, NULL, margs->ma_opts); 849} 850 851/* 852 * Mutex initialization routine; initialize lock `m' of type contained in 853 * `opts' with options contained in `opts' and name `name.' The optional 854 * lock type `type' is used as a general lock category name for use with 855 * witness. 856 */ 857void 858mtx_init(struct mtx *m, const char *name, const char *type, int opts) 859{ 860 struct lock_object *lock; 861 862 MPASS((opts & ~(MTX_SPIN | MTX_QUIET | MTX_RECURSE | 863 MTX_SLEEPABLE | MTX_NOWITNESS | MTX_DUPOK)) == 0); 864 865#ifdef MUTEX_DEBUG 866 /* Diagnostic and error correction */ 867 mtx_validate(m); 868#endif 869 870 lock = &m->mtx_object; 871 KASSERT((lock->lo_flags & LO_INITIALIZED) == 0, 872 ("mutex %s %p already initialized", name, m)); 873 bzero(m, sizeof(*m)); 874 if (opts & MTX_SPIN) 875 lock->lo_class = &lock_class_mtx_spin; 876 else 877 lock->lo_class = &lock_class_mtx_sleep; 878 lock->lo_name = name; 879 lock->lo_type = type != NULL ? type : name; 880 if (opts & MTX_QUIET) 881 lock->lo_flags = LO_QUIET; 882 if (opts & MTX_RECURSE) 883 lock->lo_flags |= LO_RECURSABLE; 884 if (opts & MTX_SLEEPABLE) 885 lock->lo_flags |= LO_SLEEPABLE; 886 if ((opts & MTX_NOWITNESS) == 0) 887 lock->lo_flags |= LO_WITNESS; 888 if (opts & MTX_DUPOK) 889 lock->lo_flags |= LO_DUPOK; 890 891 m->mtx_lock = MTX_UNOWNED; 892 TAILQ_INIT(&m->mtx_blocked); 893 894 LOCK_LOG_INIT(lock, opts); 895 896 WITNESS_INIT(lock); 897} 898 899/* 900 * Remove lock `m' from all_mtx queue. We don't allow MTX_QUIET to be 901 * passed in as a flag here because if the corresponding mtx_init() was 902 * called with MTX_QUIET set, then it will already be set in the mutex's 903 * flags. 904 */ 905void 906mtx_destroy(struct mtx *m) 907{ 908 909 LOCK_LOG_DESTROY(&m->mtx_object, 0); 910 911 if (!mtx_owned(m)) 912 MPASS(mtx_unowned(m)); 913 else { 914 MPASS((m->mtx_lock & (MTX_RECURSED|MTX_CONTESTED)) == 0); 915 916 /* Tell witness this isn't locked to make it happy. */ 917 WITNESS_UNLOCK(&m->mtx_object, LOP_EXCLUSIVE, __FILE__, 918 __LINE__); 919 } 920 921 WITNESS_DESTROY(&m->mtx_object); 922} 923 924/* 925 * Intialize the mutex code and system mutexes. This is called from the MD 926 * startup code prior to mi_startup(). The per-CPU data space needs to be 927 * setup before this is called. 928 */ 929void 930mutex_init(void) 931{ 932 933 /* Setup thread0 so that mutexes work. */ 934 LIST_INIT(&thread0.td_contested); 935 936 /* 937 * Initialize mutexes. 938 */ 939 mtx_init(&Giant, "Giant", NULL, MTX_DEF | MTX_RECURSE); 940 mtx_init(&sched_lock, "sched lock", NULL, MTX_SPIN | MTX_RECURSE); 941 mtx_init(&proc0.p_mtx, "process lock", NULL, MTX_DEF | MTX_DUPOK); 942 mtx_lock(&Giant); 943} 944 945/* 946 * Encapsulated Giant mutex routines. These routines provide encapsulation 947 * control for the Giant mutex, allowing sysctls to be used to turn on and 948 * off Giant around certain subsystems. The default value for the sysctls 949 * are set to what developers believe is stable and working in regards to 950 * the Giant pushdown. Developers should not turn off Giant via these 951 * sysctls unless they know what they are doing. 952 * 953 * Callers of mtx_lock_giant() are expected to pass the return value to an 954 * accompanying mtx_unlock_giant() later on. If multiple subsystems are 955 * effected by a Giant wrap, all related sysctl variables must be zero for 956 * the subsystem call to operate without Giant (as determined by the caller). 957 */ 958 959SYSCTL_NODE(_kern, OID_AUTO, giant, CTLFLAG_RD, NULL, "Giant mutex manipulation"); 960 961static int kern_giant_all = 0; 962SYSCTL_INT(_kern_giant, OID_AUTO, all, CTLFLAG_RW, &kern_giant_all, 0, ""); 963 964int kern_giant_proc = 1; /* Giant around PROC locks */ 965int kern_giant_file = 1; /* Giant around struct file & filedesc */ 966int kern_giant_ucred = 1; /* Giant around ucred */ 967SYSCTL_INT(_kern_giant, OID_AUTO, proc, CTLFLAG_RW, &kern_giant_proc, 0, ""); 968SYSCTL_INT(_kern_giant, OID_AUTO, file, CTLFLAG_RW, &kern_giant_file, 0, ""); 969SYSCTL_INT(_kern_giant, OID_AUTO, ucred, CTLFLAG_RW, &kern_giant_ucred, 0, ""); 970 971int 972mtx_lock_giant(int sysctlvar) 973{ 974 if (sysctlvar || kern_giant_all) { 975 mtx_lock(&Giant); 976 return(1); 977 } 978 return(0); 979} 980 981void 982mtx_unlock_giant(int s) 983{ 984 if (s) 985 mtx_unlock(&Giant); 986} 987 988