subr_turnstile.c revision 117168
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: head/sys/kern/subr_turnstile.c 117168 2003-07-02 16:14:09Z jhb $"); 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/vmmeter.h> 56 57#include <machine/atomic.h> 58#include <machine/bus.h> 59#include <machine/clock.h> 60#include <machine/cpu.h> 61 62#include <ddb/ddb.h> 63 64#include <vm/vm.h> 65#include <vm/vm_extern.h> 66 67/* 68 * Internal utility macros. 69 */ 70#define mtx_unowned(m) ((m)->mtx_lock == MTX_UNOWNED) 71 72#define mtx_owner(m) (mtx_unowned((m)) ? NULL \ 73 : (struct thread *)((m)->mtx_lock & MTX_FLAGMASK)) 74 75/* 76 * Lock classes for sleep and spin mutexes. 77 */ 78struct lock_class lock_class_mtx_sleep = { 79 "sleep mutex", 80 LC_SLEEPLOCK | LC_RECURSABLE 81}; 82struct lock_class lock_class_mtx_spin = { 83 "spin mutex", 84 LC_SPINLOCK | LC_RECURSABLE 85}; 86 87/* 88 * System-wide mutexes 89 */ 90struct mtx sched_lock; 91struct mtx Giant; 92 93/* 94 * Prototypes for non-exported routines. 95 */ 96static void propagate_priority(struct thread *); 97 98static void 99propagate_priority(struct thread *td) 100{ 101 int pri = td->td_priority; 102 struct mtx *m = td->td_blocked; 103 104 mtx_assert(&sched_lock, MA_OWNED); 105 for (;;) { 106 struct thread *td1; 107 108 td = mtx_owner(m); 109 110 if (td == NULL) { 111 /* 112 * This really isn't quite right. Really 113 * ought to bump priority of thread that 114 * next acquires the mutex. 115 */ 116 MPASS(m->mtx_lock == MTX_CONTESTED); 117 return; 118 } 119 120 MPASS(td->td_proc != NULL); 121 MPASS(td->td_proc->p_magic == P_MAGIC); 122 KASSERT(!TD_IS_SLEEPING(td), ("sleeping thread owns a mutex")); 123 if (td->td_priority <= pri) /* lower is higher priority */ 124 return; 125 126 127 /* 128 * If lock holder is actually running, just bump priority. 129 */ 130 if (TD_IS_RUNNING(td)) { 131 td->td_priority = pri; 132 return; 133 } 134 135#ifndef SMP 136 /* 137 * For UP, we check to see if td is curthread (this shouldn't 138 * ever happen however as it would mean we are in a deadlock.) 139 */ 140 KASSERT(td != curthread, ("Deadlock detected")); 141#endif 142 143 /* 144 * If on run queue move to new run queue, and quit. 145 * XXXKSE this gets a lot more complicated under threads 146 * but try anyhow. 147 */ 148 if (TD_ON_RUNQ(td)) { 149 MPASS(td->td_blocked == NULL); 150 sched_prio(td, pri); 151 return; 152 } 153 /* 154 * Adjust for any other cases. 155 */ 156 td->td_priority = pri; 157 158 /* 159 * If we aren't blocked on a mutex, we should be. 160 */ 161 KASSERT(TD_ON_LOCK(td), ( 162 "process %d(%s):%d holds %s but isn't blocked on a mutex\n", 163 td->td_proc->p_pid, td->td_proc->p_comm, td->td_state, 164 m->mtx_object.lo_name)); 165 166 /* 167 * Pick up the mutex that td is blocked on. 168 */ 169 m = td->td_blocked; 170 MPASS(m != NULL); 171 172 /* 173 * Check if the thread needs to be moved up on 174 * the blocked chain 175 */ 176 if (td == TAILQ_FIRST(&m->mtx_blocked)) { 177 continue; 178 } 179 180 td1 = TAILQ_PREV(td, threadqueue, td_lockq); 181 if (td1->td_priority <= pri) { 182 continue; 183 } 184 185 /* 186 * Remove thread from blocked chain and determine where 187 * it should be moved up to. Since we know that td1 has 188 * a lower priority than td, we know that at least one 189 * thread in the chain has a lower priority and that 190 * td1 will thus not be NULL after the loop. 191 */ 192 TAILQ_REMOVE(&m->mtx_blocked, td, td_lockq); 193 TAILQ_FOREACH(td1, &m->mtx_blocked, td_lockq) { 194 MPASS(td1->td_proc->p_magic == P_MAGIC); 195 if (td1->td_priority > pri) 196 break; 197 } 198 199 MPASS(td1 != NULL); 200 TAILQ_INSERT_BEFORE(td1, td, td_lockq); 201 CTR4(KTR_LOCK, 202 "propagate_priority: p %p moved before %p on [%p] %s", 203 td, td1, m, m->mtx_object.lo_name); 204 } 205} 206 207#ifdef MUTEX_PROFILING 208SYSCTL_NODE(_debug, OID_AUTO, mutex, CTLFLAG_RD, NULL, "mutex debugging"); 209SYSCTL_NODE(_debug_mutex, OID_AUTO, prof, CTLFLAG_RD, NULL, "mutex profiling"); 210static int mutex_prof_enable = 0; 211SYSCTL_INT(_debug_mutex_prof, OID_AUTO, enable, CTLFLAG_RW, 212 &mutex_prof_enable, 0, "Enable tracing of mutex holdtime"); 213 214struct mutex_prof { 215 const char *name; 216 const char *file; 217 int line; 218 uintmax_t cnt_max; 219 uintmax_t cnt_tot; 220 uintmax_t cnt_cur; 221 struct mutex_prof *next; 222}; 223 224/* 225 * mprof_buf is a static pool of profiling records to avoid possible 226 * reentrance of the memory allocation functions. 227 * 228 * Note: NUM_MPROF_BUFFERS must be smaller than MPROF_HASH_SIZE. 229 */ 230#define NUM_MPROF_BUFFERS 1000 231static struct mutex_prof mprof_buf[NUM_MPROF_BUFFERS]; 232static int first_free_mprof_buf; 233#define MPROF_HASH_SIZE 1009 234static struct mutex_prof *mprof_hash[MPROF_HASH_SIZE]; 235/* SWAG: sbuf size = avg stat. line size * number of locks */ 236#define MPROF_SBUF_SIZE 256 * 400 237 238static int mutex_prof_acquisitions; 239SYSCTL_INT(_debug_mutex_prof, OID_AUTO, acquisitions, CTLFLAG_RD, 240 &mutex_prof_acquisitions, 0, "Number of mutex acquistions recorded"); 241static int mutex_prof_records; 242SYSCTL_INT(_debug_mutex_prof, OID_AUTO, records, CTLFLAG_RD, 243 &mutex_prof_records, 0, "Number of profiling records"); 244static int mutex_prof_maxrecords = NUM_MPROF_BUFFERS; 245SYSCTL_INT(_debug_mutex_prof, OID_AUTO, maxrecords, CTLFLAG_RD, 246 &mutex_prof_maxrecords, 0, "Maximum number of profiling records"); 247static int mutex_prof_rejected; 248SYSCTL_INT(_debug_mutex_prof, OID_AUTO, rejected, CTLFLAG_RD, 249 &mutex_prof_rejected, 0, "Number of rejected profiling records"); 250static int mutex_prof_hashsize = MPROF_HASH_SIZE; 251SYSCTL_INT(_debug_mutex_prof, OID_AUTO, hashsize, CTLFLAG_RD, 252 &mutex_prof_hashsize, 0, "Hash size"); 253static int mutex_prof_collisions = 0; 254SYSCTL_INT(_debug_mutex_prof, OID_AUTO, collisions, CTLFLAG_RD, 255 &mutex_prof_collisions, 0, "Number of hash collisions"); 256 257/* 258 * mprof_mtx protects the profiling buffers and the hash. 259 */ 260static struct mtx mprof_mtx; 261MTX_SYSINIT(mprof, &mprof_mtx, "mutex profiling lock", MTX_SPIN | MTX_QUIET); 262 263static u_int64_t 264nanoseconds(void) 265{ 266 struct timespec tv; 267 268 nanotime(&tv); 269 return (tv.tv_sec * (u_int64_t)1000000000 + tv.tv_nsec); 270} 271 272static int 273dump_mutex_prof_stats(SYSCTL_HANDLER_ARGS) 274{ 275 struct sbuf *sb; 276 int error, i; 277 static int multiplier = 1; 278 279 if (first_free_mprof_buf == 0) 280 return (SYSCTL_OUT(req, "No locking recorded", 281 sizeof("No locking recorded"))); 282 283retry_sbufops: 284 sb = sbuf_new(NULL, NULL, MPROF_SBUF_SIZE * multiplier, SBUF_FIXEDLEN); 285 sbuf_printf(sb, "%6s %12s %11s %5s %s\n", 286 "max", "total", "count", "avg", "name"); 287 /* 288 * XXX this spinlock seems to be by far the largest perpetrator 289 * of spinlock latency (1.6 msec on an Athlon1600 was recorded 290 * even before I pessimized it further by moving the average 291 * computation here). 292 */ 293 mtx_lock_spin(&mprof_mtx); 294 for (i = 0; i < first_free_mprof_buf; ++i) { 295 sbuf_printf(sb, "%6ju %12ju %11ju %5ju %s:%d (%s)\n", 296 mprof_buf[i].cnt_max / 1000, 297 mprof_buf[i].cnt_tot / 1000, 298 mprof_buf[i].cnt_cur, 299 mprof_buf[i].cnt_cur == 0 ? (uintmax_t)0 : 300 mprof_buf[i].cnt_tot / (mprof_buf[i].cnt_cur * 1000), 301 mprof_buf[i].file, mprof_buf[i].line, mprof_buf[i].name); 302 if (sbuf_overflowed(sb)) { 303 mtx_unlock_spin(&mprof_mtx); 304 sbuf_delete(sb); 305 multiplier++; 306 goto retry_sbufops; 307 } 308 } 309 mtx_unlock_spin(&mprof_mtx); 310 sbuf_finish(sb); 311 error = SYSCTL_OUT(req, sbuf_data(sb), sbuf_len(sb) + 1); 312 sbuf_delete(sb); 313 return (error); 314} 315SYSCTL_PROC(_debug_mutex_prof, OID_AUTO, stats, CTLTYPE_STRING | CTLFLAG_RD, 316 NULL, 0, dump_mutex_prof_stats, "A", "Mutex profiling statistics"); 317#endif 318 319/* 320 * Function versions of the inlined __mtx_* macros. These are used by 321 * modules and can also be called from assembly language if needed. 322 */ 323void 324_mtx_lock_flags(struct mtx *m, int opts, const char *file, int line) 325{ 326 327 MPASS(curthread != NULL); 328 KASSERT(m->mtx_object.lo_class == &lock_class_mtx_sleep, 329 ("mtx_lock() of spin mutex %s @ %s:%d", m->mtx_object.lo_name, 330 file, line)); 331 _get_sleep_lock(m, curthread, opts, file, line); 332 LOCK_LOG_LOCK("LOCK", &m->mtx_object, opts, m->mtx_recurse, file, 333 line); 334 WITNESS_LOCK(&m->mtx_object, opts | LOP_EXCLUSIVE, file, line); 335#ifdef MUTEX_PROFILING 336 /* don't reset the timer when/if recursing */ 337 if (m->mtx_acqtime == 0) { 338 m->mtx_filename = file; 339 m->mtx_lineno = line; 340 m->mtx_acqtime = mutex_prof_enable ? nanoseconds() : 0; 341 ++mutex_prof_acquisitions; 342 } 343#endif 344} 345 346void 347_mtx_unlock_flags(struct mtx *m, int opts, const char *file, int line) 348{ 349 350 MPASS(curthread != NULL); 351 KASSERT(m->mtx_object.lo_class == &lock_class_mtx_sleep, 352 ("mtx_unlock() of spin mutex %s @ %s:%d", m->mtx_object.lo_name, 353 file, line)); 354 WITNESS_UNLOCK(&m->mtx_object, opts | LOP_EXCLUSIVE, file, line); 355 LOCK_LOG_LOCK("UNLOCK", &m->mtx_object, opts, m->mtx_recurse, file, 356 line); 357 mtx_assert(m, MA_OWNED); 358#ifdef MUTEX_PROFILING 359 if (m->mtx_acqtime != 0) { 360 static const char *unknown = "(unknown)"; 361 struct mutex_prof *mpp; 362 u_int64_t acqtime, now; 363 const char *p, *q; 364 volatile u_int hash; 365 366 now = nanoseconds(); 367 acqtime = m->mtx_acqtime; 368 m->mtx_acqtime = 0; 369 if (now <= acqtime) 370 goto out; 371 for (p = m->mtx_filename; 372 p != NULL && strncmp(p, "../", 3) == 0; p += 3) 373 /* nothing */ ; 374 if (p == NULL || *p == '\0') 375 p = unknown; 376 for (hash = m->mtx_lineno, q = p; *q != '\0'; ++q) 377 hash = (hash * 2 + *q) % MPROF_HASH_SIZE; 378 mtx_lock_spin(&mprof_mtx); 379 for (mpp = mprof_hash[hash]; mpp != NULL; mpp = mpp->next) 380 if (mpp->line == m->mtx_lineno && 381 strcmp(mpp->file, p) == 0) 382 break; 383 if (mpp == NULL) { 384 /* Just exit if we cannot get a trace buffer */ 385 if (first_free_mprof_buf >= NUM_MPROF_BUFFERS) { 386 ++mutex_prof_rejected; 387 goto unlock; 388 } 389 mpp = &mprof_buf[first_free_mprof_buf++]; 390 mpp->name = mtx_name(m); 391 mpp->file = p; 392 mpp->line = m->mtx_lineno; 393 mpp->next = mprof_hash[hash]; 394 if (mprof_hash[hash] != NULL) 395 ++mutex_prof_collisions; 396 mprof_hash[hash] = mpp; 397 ++mutex_prof_records; 398 } 399 /* 400 * Record if the mutex has been held longer now than ever 401 * before. 402 */ 403 if (now - acqtime > mpp->cnt_max) 404 mpp->cnt_max = now - acqtime; 405 mpp->cnt_tot += now - acqtime; 406 mpp->cnt_cur++; 407unlock: 408 mtx_unlock_spin(&mprof_mtx); 409 } 410out: 411#endif 412 _rel_sleep_lock(m, curthread, opts, file, line); 413} 414 415void 416_mtx_lock_spin_flags(struct mtx *m, int opts, const char *file, int line) 417{ 418 419 MPASS(curthread != NULL); 420 KASSERT(m->mtx_object.lo_class == &lock_class_mtx_spin, 421 ("mtx_lock_spin() of sleep mutex %s @ %s:%d", 422 m->mtx_object.lo_name, file, line)); 423#if defined(SMP) || LOCK_DEBUG > 0 || 1 424 _get_spin_lock(m, curthread, opts, file, line); 425#else 426 critical_enter(); 427#endif 428 LOCK_LOG_LOCK("LOCK", &m->mtx_object, opts, m->mtx_recurse, file, 429 line); 430 WITNESS_LOCK(&m->mtx_object, opts | LOP_EXCLUSIVE, file, line); 431} 432 433void 434_mtx_unlock_spin_flags(struct mtx *m, int opts, const char *file, int line) 435{ 436 437 MPASS(curthread != NULL); 438 KASSERT(m->mtx_object.lo_class == &lock_class_mtx_spin, 439 ("mtx_unlock_spin() of sleep mutex %s @ %s:%d", 440 m->mtx_object.lo_name, file, line)); 441 WITNESS_UNLOCK(&m->mtx_object, opts | LOP_EXCLUSIVE, file, line); 442 LOCK_LOG_LOCK("UNLOCK", &m->mtx_object, opts, m->mtx_recurse, file, 443 line); 444 mtx_assert(m, MA_OWNED); 445#if defined(SMP) || LOCK_DEBUG > 0 || 1 446 _rel_spin_lock(m); 447#else 448 critical_exit(); 449#endif 450} 451 452/* 453 * The important part of mtx_trylock{,_flags}() 454 * Tries to acquire lock `m.' We do NOT handle recursion here. If this 455 * function is called on a recursed mutex, it will return failure and 456 * will not recursively acquire the lock. You are expected to know what 457 * you are doing. 458 */ 459int 460_mtx_trylock(struct mtx *m, int opts, const char *file, int line) 461{ 462 int rval; 463 464 MPASS(curthread != NULL); 465 466 rval = _obtain_lock(m, curthread); 467 468 LOCK_LOG_TRY("LOCK", &m->mtx_object, opts, rval, file, line); 469 if (rval) 470 WITNESS_LOCK(&m->mtx_object, opts | LOP_EXCLUSIVE | LOP_TRYLOCK, 471 file, line); 472 473 return (rval); 474} 475 476/* 477 * _mtx_lock_sleep: the tougher part of acquiring an MTX_DEF lock. 478 * 479 * We call this if the lock is either contested (i.e. we need to go to 480 * sleep waiting for it), or if we need to recurse on it. 481 */ 482void 483_mtx_lock_sleep(struct mtx *m, int opts, const char *file, int line) 484{ 485 struct thread *td = curthread; 486 struct thread *td1; 487#if defined(SMP) && defined(ADAPTIVE_MUTEXES) 488 struct thread *owner; 489#endif 490 uintptr_t v; 491#ifdef KTR 492 int cont_logged = 0; 493#endif 494 495 if (mtx_owned(m)) { 496 m->mtx_recurse++; 497 atomic_set_ptr(&m->mtx_lock, MTX_RECURSED); 498 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 499 CTR1(KTR_LOCK, "_mtx_lock_sleep: %p recursing", m); 500 return; 501 } 502 503 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 504 CTR4(KTR_LOCK, 505 "_mtx_lock_sleep: %s contested (lock=%p) at %s:%d", 506 m->mtx_object.lo_name, (void *)m->mtx_lock, file, line); 507 508 while (!_obtain_lock(m, td)) { 509 510 mtx_lock_spin(&sched_lock); 511 v = m->mtx_lock; 512 513 /* 514 * Check if the lock has been released while spinning for 515 * the sched_lock. 516 */ 517 if (v == MTX_UNOWNED) { 518 mtx_unlock_spin(&sched_lock); 519#ifdef __i386__ 520 ia32_pause(); 521#endif 522 continue; 523 } 524 525 /* 526 * The mutex was marked contested on release. This means that 527 * there are other threads blocked on it. Grab ownership of 528 * it and propagate its priority to the current thread if 529 * necessary. 530 */ 531 if (v == MTX_CONTESTED) { 532 td1 = TAILQ_FIRST(&m->mtx_blocked); 533 MPASS(td1 != NULL); 534 m->mtx_lock = (uintptr_t)td | MTX_CONTESTED; 535 LIST_INSERT_HEAD(&td->td_contested, m, mtx_contested); 536 537 if (td1->td_priority < td->td_priority) 538 td->td_priority = td1->td_priority; 539 mtx_unlock_spin(&sched_lock); 540 return; 541 } 542 543 /* 544 * If the mutex isn't already contested and a failure occurs 545 * setting the contested bit, the mutex was either released 546 * or the state of the MTX_RECURSED bit changed. 547 */ 548 if ((v & MTX_CONTESTED) == 0 && 549 !atomic_cmpset_ptr(&m->mtx_lock, (void *)v, 550 (void *)(v | MTX_CONTESTED))) { 551 mtx_unlock_spin(&sched_lock); 552#ifdef __i386__ 553 ia32_pause(); 554#endif 555 continue; 556 } 557 558#if defined(SMP) && defined(ADAPTIVE_MUTEXES) 559 /* 560 * If the current owner of the lock is executing on another 561 * CPU, spin instead of blocking. 562 */ 563 owner = (struct thread *)(v & MTX_FLAGMASK); 564 if (m != &Giant && TD_IS_RUNNING(owner)) { 565 mtx_unlock_spin(&sched_lock); 566 while (mtx_owner(m) == owner && TD_IS_RUNNING(owner)) { 567#ifdef __i386__ 568 ia32_pause(); 569#endif 570 } 571 continue; 572 } 573#endif /* SMP && ADAPTIVE_MUTEXES */ 574 575 /* 576 * We definitely must sleep for this lock. 577 */ 578 mtx_assert(m, MA_NOTOWNED); 579 580#ifdef notyet 581 /* 582 * If we're borrowing an interrupted thread's VM context, we 583 * must clean up before going to sleep. 584 */ 585 if (td->td_ithd != NULL) { 586 struct ithd *it = td->td_ithd; 587 588 if (it->it_interrupted) { 589 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 590 CTR2(KTR_LOCK, 591 "_mtx_lock_sleep: %p interrupted %p", 592 it, it->it_interrupted); 593 intr_thd_fixup(it); 594 } 595 } 596#endif 597 598 /* 599 * Put us on the list of threads blocked on this mutex 600 * and add this mutex to the owning thread's list of 601 * contested mutexes if needed. 602 */ 603 if (TAILQ_EMPTY(&m->mtx_blocked)) { 604 td1 = mtx_owner(m); 605 LIST_INSERT_HEAD(&td1->td_contested, m, mtx_contested); 606 TAILQ_INSERT_TAIL(&m->mtx_blocked, td, td_lockq); 607 } else { 608 TAILQ_FOREACH(td1, &m->mtx_blocked, td_lockq) 609 if (td1->td_priority > td->td_priority) 610 break; 611 if (td1) 612 TAILQ_INSERT_BEFORE(td1, td, td_lockq); 613 else 614 TAILQ_INSERT_TAIL(&m->mtx_blocked, td, td_lockq); 615 } 616#ifdef KTR 617 if (!cont_logged) { 618 CTR6(KTR_CONTENTION, 619 "contention: %p at %s:%d wants %s, taken by %s:%d", 620 td, file, line, m->mtx_object.lo_name, 621 WITNESS_FILE(&m->mtx_object), 622 WITNESS_LINE(&m->mtx_object)); 623 cont_logged = 1; 624 } 625#endif 626 627 /* 628 * Save who we're blocked on. 629 */ 630 td->td_blocked = m; 631 td->td_lockname = m->mtx_object.lo_name; 632 TD_SET_LOCK(td); 633 propagate_priority(td); 634 635 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 636 CTR3(KTR_LOCK, 637 "_mtx_lock_sleep: p %p blocked on [%p] %s", td, m, 638 m->mtx_object.lo_name); 639 640 td->td_proc->p_stats->p_ru.ru_nvcsw++; 641 mi_switch(); 642 643 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 644 CTR3(KTR_LOCK, 645 "_mtx_lock_sleep: p %p free from blocked on [%p] %s", 646 td, m, m->mtx_object.lo_name); 647 648 mtx_unlock_spin(&sched_lock); 649 } 650 651#ifdef KTR 652 if (cont_logged) { 653 CTR4(KTR_CONTENTION, 654 "contention end: %s acquired by %p at %s:%d", 655 m->mtx_object.lo_name, td, file, line); 656 } 657#endif 658 return; 659} 660 661/* 662 * _mtx_lock_spin: the tougher part of acquiring an MTX_SPIN lock. 663 * 664 * This is only called if we need to actually spin for the lock. Recursion 665 * is handled inline. 666 */ 667void 668_mtx_lock_spin(struct mtx *m, int opts, const char *file, int line) 669{ 670 int i = 0; 671 672 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 673 CTR1(KTR_LOCK, "_mtx_lock_spin: %p spinning", m); 674 675 for (;;) { 676 if (_obtain_lock(m, curthread)) 677 break; 678 679 /* Give interrupts a chance while we spin. */ 680 critical_exit(); 681 while (m->mtx_lock != MTX_UNOWNED) { 682 if (i++ < 10000000) { 683#ifdef __i386__ 684 ia32_pause(); 685#endif 686 continue; 687 } 688 if (i < 60000000) 689 DELAY(1); 690#ifdef DDB 691 else if (!db_active) 692#else 693 else 694#endif 695 panic("spin lock %s held by %p for > 5 seconds", 696 m->mtx_object.lo_name, (void *)m->mtx_lock); 697#ifdef __i386__ 698 ia32_pause(); 699#endif 700 } 701 critical_enter(); 702 } 703 704 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 705 CTR1(KTR_LOCK, "_mtx_lock_spin: %p spin done", m); 706 707 return; 708} 709 710/* 711 * _mtx_unlock_sleep: the tougher part of releasing an MTX_DEF lock. 712 * 713 * We are only called here if the lock is recursed or contested (i.e. we 714 * need to wake up a blocked thread). 715 */ 716void 717_mtx_unlock_sleep(struct mtx *m, int opts, const char *file, int line) 718{ 719 struct thread *td, *td1; 720 struct mtx *m1; 721 int pri; 722 723 td = curthread; 724 725 if (mtx_recursed(m)) { 726 if (--(m->mtx_recurse) == 0) 727 atomic_clear_ptr(&m->mtx_lock, MTX_RECURSED); 728 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 729 CTR1(KTR_LOCK, "_mtx_unlock_sleep: %p unrecurse", m); 730 return; 731 } 732 733 mtx_lock_spin(&sched_lock); 734 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 735 CTR1(KTR_LOCK, "_mtx_unlock_sleep: %p contested", m); 736 737 td1 = TAILQ_FIRST(&m->mtx_blocked); 738#if defined(SMP) && defined(ADAPTIVE_MUTEXES) 739 if (td1 == NULL) { 740 _release_lock_quick(m); 741 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 742 CTR1(KTR_LOCK, "_mtx_unlock_sleep: %p no sleepers", m); 743 mtx_unlock_spin(&sched_lock); 744 return; 745 } 746#endif 747 MPASS(td->td_proc->p_magic == P_MAGIC); 748 MPASS(td1->td_proc->p_magic == P_MAGIC); 749 750 TAILQ_REMOVE(&m->mtx_blocked, td1, td_lockq); 751 752 LIST_REMOVE(m, mtx_contested); 753 if (TAILQ_EMPTY(&m->mtx_blocked)) { 754 _release_lock_quick(m); 755 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 756 CTR1(KTR_LOCK, "_mtx_unlock_sleep: %p not held", m); 757 } else 758 m->mtx_lock = MTX_CONTESTED; 759 760 pri = PRI_MAX; 761 LIST_FOREACH(m1, &td->td_contested, mtx_contested) { 762 int cp = TAILQ_FIRST(&m1->mtx_blocked)->td_priority; 763 if (cp < pri) 764 pri = cp; 765 } 766 767 if (pri > td->td_base_pri) 768 pri = td->td_base_pri; 769 td->td_priority = pri; 770 771 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 772 CTR2(KTR_LOCK, "_mtx_unlock_sleep: %p contested setrunqueue %p", 773 m, td1); 774 775 td1->td_blocked = NULL; 776 TD_CLR_LOCK(td1); 777 if (!TD_CAN_RUN(td1)) { 778 mtx_unlock_spin(&sched_lock); 779 return; 780 } 781 setrunqueue(td1); 782 783 if (td->td_critnest == 1 && td1->td_priority < pri) { 784#ifdef notyet 785 if (td->td_ithd != NULL) { 786 struct ithd *it = td->td_ithd; 787 788 if (it->it_interrupted) { 789 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 790 CTR2(KTR_LOCK, 791 "_mtx_unlock_sleep: %p interrupted %p", 792 it, it->it_interrupted); 793 intr_thd_fixup(it); 794 } 795 } 796#endif 797 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 798 CTR2(KTR_LOCK, 799 "_mtx_unlock_sleep: %p switching out lock=%p", m, 800 (void *)m->mtx_lock); 801 802 td->td_proc->p_stats->p_ru.ru_nivcsw++; 803 mi_switch(); 804 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 805 CTR2(KTR_LOCK, "_mtx_unlock_sleep: %p resuming lock=%p", 806 m, (void *)m->mtx_lock); 807 } 808 809 mtx_unlock_spin(&sched_lock); 810 811 return; 812} 813 814/* 815 * All the unlocking of MTX_SPIN locks is done inline. 816 * See the _rel_spin_lock() macro for the details. 817 */ 818 819/* 820 * The backing function for the INVARIANTS-enabled mtx_assert() 821 */ 822#ifdef INVARIANT_SUPPORT 823void 824_mtx_assert(struct mtx *m, int what, const char *file, int line) 825{ 826 827 if (panicstr != NULL) 828 return; 829 switch (what) { 830 case MA_OWNED: 831 case MA_OWNED | MA_RECURSED: 832 case MA_OWNED | MA_NOTRECURSED: 833 if (!mtx_owned(m)) 834 panic("mutex %s not owned at %s:%d", 835 m->mtx_object.lo_name, file, line); 836 if (mtx_recursed(m)) { 837 if ((what & MA_NOTRECURSED) != 0) 838 panic("mutex %s recursed at %s:%d", 839 m->mtx_object.lo_name, file, line); 840 } else if ((what & MA_RECURSED) != 0) { 841 panic("mutex %s unrecursed at %s:%d", 842 m->mtx_object.lo_name, file, line); 843 } 844 break; 845 case MA_NOTOWNED: 846 if (mtx_owned(m)) 847 panic("mutex %s owned at %s:%d", 848 m->mtx_object.lo_name, file, line); 849 break; 850 default: 851 panic("unknown mtx_assert at %s:%d", file, line); 852 } 853} 854#endif 855 856/* 857 * The MUTEX_DEBUG-enabled mtx_validate() 858 * 859 * Most of these checks have been moved off into the LO_INITIALIZED flag 860 * maintained by the witness code. 861 */ 862#ifdef MUTEX_DEBUG 863 864void mtx_validate(struct mtx *); 865 866void 867mtx_validate(struct mtx *m) 868{ 869 870/* 871 * XXX: When kernacc() does not require Giant we can reenable this check 872 */ 873#ifdef notyet 874/* 875 * XXX - When kernacc() is fixed on the alpha to handle K0_SEG memory properly 876 * we can re-enable the kernacc() checks. 877 */ 878#ifndef __alpha__ 879 /* 880 * Can't call kernacc() from early init386(), especially when 881 * initializing Giant mutex, because some stuff in kernacc() 882 * requires Giant itself. 883 */ 884 if (!cold) 885 if (!kernacc((caddr_t)m, sizeof(m), 886 VM_PROT_READ | VM_PROT_WRITE)) 887 panic("Can't read and write to mutex %p", m); 888#endif 889#endif 890} 891#endif 892 893/* 894 * General init routine used by the MTX_SYSINIT() macro. 895 */ 896void 897mtx_sysinit(void *arg) 898{ 899 struct mtx_args *margs = arg; 900 901 mtx_init(margs->ma_mtx, margs->ma_desc, NULL, margs->ma_opts); 902} 903 904/* 905 * Mutex initialization routine; initialize lock `m' of type contained in 906 * `opts' with options contained in `opts' and name `name.' The optional 907 * lock type `type' is used as a general lock category name for use with 908 * witness. 909 */ 910void 911mtx_init(struct mtx *m, const char *name, const char *type, int opts) 912{ 913 struct lock_object *lock; 914 915 MPASS((opts & ~(MTX_SPIN | MTX_QUIET | MTX_RECURSE | 916 MTX_NOWITNESS | MTX_DUPOK)) == 0); 917 918#ifdef MUTEX_DEBUG 919 /* Diagnostic and error correction */ 920 mtx_validate(m); 921#endif 922 923 lock = &m->mtx_object; 924 KASSERT((lock->lo_flags & LO_INITIALIZED) == 0, 925 ("mutex \"%s\" %p already initialized", name, m)); 926 bzero(m, sizeof(*m)); 927 if (opts & MTX_SPIN) 928 lock->lo_class = &lock_class_mtx_spin; 929 else 930 lock->lo_class = &lock_class_mtx_sleep; 931 lock->lo_name = name; 932 lock->lo_type = type != NULL ? type : name; 933 if (opts & MTX_QUIET) 934 lock->lo_flags = LO_QUIET; 935 if (opts & MTX_RECURSE) 936 lock->lo_flags |= LO_RECURSABLE; 937 if ((opts & MTX_NOWITNESS) == 0) 938 lock->lo_flags |= LO_WITNESS; 939 if (opts & MTX_DUPOK) 940 lock->lo_flags |= LO_DUPOK; 941 942 m->mtx_lock = MTX_UNOWNED; 943 TAILQ_INIT(&m->mtx_blocked); 944 945 LOCK_LOG_INIT(lock, opts); 946 947 WITNESS_INIT(lock); 948} 949 950/* 951 * Remove lock `m' from all_mtx queue. We don't allow MTX_QUIET to be 952 * passed in as a flag here because if the corresponding mtx_init() was 953 * called with MTX_QUIET set, then it will already be set in the mutex's 954 * flags. 955 */ 956void 957mtx_destroy(struct mtx *m) 958{ 959 960 LOCK_LOG_DESTROY(&m->mtx_object, 0); 961 962 if (!mtx_owned(m)) 963 MPASS(mtx_unowned(m)); 964 else { 965 MPASS((m->mtx_lock & (MTX_RECURSED|MTX_CONTESTED)) == 0); 966 967 /* Tell witness this isn't locked to make it happy. */ 968 WITNESS_UNLOCK(&m->mtx_object, LOP_EXCLUSIVE, __FILE__, 969 __LINE__); 970 } 971 972 WITNESS_DESTROY(&m->mtx_object); 973} 974 975/* 976 * Intialize the mutex code and system mutexes. This is called from the MD 977 * startup code prior to mi_startup(). The per-CPU data space needs to be 978 * setup before this is called. 979 */ 980void 981mutex_init(void) 982{ 983 984 /* Setup thread0 so that mutexes work. */ 985 LIST_INIT(&thread0.td_contested); 986 987 /* 988 * Initialize mutexes. 989 */ 990 mtx_init(&Giant, "Giant", NULL, MTX_DEF | MTX_RECURSE); 991 mtx_init(&sched_lock, "sched lock", NULL, MTX_SPIN | MTX_RECURSE); 992 mtx_init(&proc0.p_mtx, "process lock", NULL, MTX_DEF | MTX_DUPOK); 993 mtx_lock(&Giant); 994} 995