subr_turnstile.c revision 118272
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 118272 2003-07-31 18:52:18Z 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), ( 123 "sleeping thread (pid %d) owns a mutex", 124 td->td_proc->p_pid)); 125 if (td->td_priority <= pri) /* lower is higher priority */ 126 return; 127 128 129 /* 130 * If lock holder is actually running, just bump priority. 131 */ 132 if (TD_IS_RUNNING(td)) { 133 td->td_priority = pri; 134 return; 135 } 136 137#ifndef SMP 138 /* 139 * For UP, we check to see if td is curthread (this shouldn't 140 * ever happen however as it would mean we are in a deadlock.) 141 */ 142 KASSERT(td != curthread, ("Deadlock detected")); 143#endif 144 145 /* 146 * If on run queue move to new run queue, and quit. 147 * XXXKSE this gets a lot more complicated under threads 148 * but try anyhow. 149 */ 150 if (TD_ON_RUNQ(td)) { 151 MPASS(td->td_blocked == NULL); 152 sched_prio(td, pri); 153 return; 154 } 155 /* 156 * Adjust for any other cases. 157 */ 158 td->td_priority = pri; 159 160 /* 161 * If we aren't blocked on a mutex, we should be. 162 */ 163 KASSERT(TD_ON_LOCK(td), ( 164 "process %d(%s):%d holds %s but isn't blocked on a mutex\n", 165 td->td_proc->p_pid, td->td_proc->p_comm, td->td_state, 166 m->mtx_object.lo_name)); 167 168 /* 169 * Pick up the mutex that td is blocked on. 170 */ 171 m = td->td_blocked; 172 MPASS(m != NULL); 173 174 /* 175 * Check if the thread needs to be moved up on 176 * the blocked chain 177 */ 178 if (td == TAILQ_FIRST(&m->mtx_blocked)) { 179 continue; 180 } 181 182 td1 = TAILQ_PREV(td, threadqueue, td_lockq); 183 if (td1->td_priority <= pri) { 184 continue; 185 } 186 187 /* 188 * Remove thread from blocked chain and determine where 189 * it should be moved up to. Since we know that td1 has 190 * a lower priority than td, we know that at least one 191 * thread in the chain has a lower priority and that 192 * td1 will thus not be NULL after the loop. 193 */ 194 TAILQ_REMOVE(&m->mtx_blocked, td, td_lockq); 195 TAILQ_FOREACH(td1, &m->mtx_blocked, td_lockq) { 196 MPASS(td1->td_proc->p_magic == P_MAGIC); 197 if (td1->td_priority > pri) 198 break; 199 } 200 201 MPASS(td1 != NULL); 202 TAILQ_INSERT_BEFORE(td1, td, td_lockq); 203 CTR4(KTR_LOCK, 204 "propagate_priority: p %p moved before %p on [%p] %s", 205 td, td1, m, m->mtx_object.lo_name); 206 } 207} 208 209#ifdef MUTEX_PROFILING 210SYSCTL_NODE(_debug, OID_AUTO, mutex, CTLFLAG_RD, NULL, "mutex debugging"); 211SYSCTL_NODE(_debug_mutex, OID_AUTO, prof, CTLFLAG_RD, NULL, "mutex profiling"); 212static int mutex_prof_enable = 0; 213SYSCTL_INT(_debug_mutex_prof, OID_AUTO, enable, CTLFLAG_RW, 214 &mutex_prof_enable, 0, "Enable tracing of mutex holdtime"); 215 216struct mutex_prof { 217 const char *name; 218 const char *file; 219 int line; 220 uintmax_t cnt_max; 221 uintmax_t cnt_tot; 222 uintmax_t cnt_cur; 223 struct mutex_prof *next; 224}; 225 226/* 227 * mprof_buf is a static pool of profiling records to avoid possible 228 * reentrance of the memory allocation functions. 229 * 230 * Note: NUM_MPROF_BUFFERS must be smaller than MPROF_HASH_SIZE. 231 */ 232#define NUM_MPROF_BUFFERS 1000 233static struct mutex_prof mprof_buf[NUM_MPROF_BUFFERS]; 234static int first_free_mprof_buf; 235#define MPROF_HASH_SIZE 1009 236static struct mutex_prof *mprof_hash[MPROF_HASH_SIZE]; 237/* SWAG: sbuf size = avg stat. line size * number of locks */ 238#define MPROF_SBUF_SIZE 256 * 400 239 240static int mutex_prof_acquisitions; 241SYSCTL_INT(_debug_mutex_prof, OID_AUTO, acquisitions, CTLFLAG_RD, 242 &mutex_prof_acquisitions, 0, "Number of mutex acquistions recorded"); 243static int mutex_prof_records; 244SYSCTL_INT(_debug_mutex_prof, OID_AUTO, records, CTLFLAG_RD, 245 &mutex_prof_records, 0, "Number of profiling records"); 246static int mutex_prof_maxrecords = NUM_MPROF_BUFFERS; 247SYSCTL_INT(_debug_mutex_prof, OID_AUTO, maxrecords, CTLFLAG_RD, 248 &mutex_prof_maxrecords, 0, "Maximum number of profiling records"); 249static int mutex_prof_rejected; 250SYSCTL_INT(_debug_mutex_prof, OID_AUTO, rejected, CTLFLAG_RD, 251 &mutex_prof_rejected, 0, "Number of rejected profiling records"); 252static int mutex_prof_hashsize = MPROF_HASH_SIZE; 253SYSCTL_INT(_debug_mutex_prof, OID_AUTO, hashsize, CTLFLAG_RD, 254 &mutex_prof_hashsize, 0, "Hash size"); 255static int mutex_prof_collisions = 0; 256SYSCTL_INT(_debug_mutex_prof, OID_AUTO, collisions, CTLFLAG_RD, 257 &mutex_prof_collisions, 0, "Number of hash collisions"); 258 259/* 260 * mprof_mtx protects the profiling buffers and the hash. 261 */ 262static struct mtx mprof_mtx; 263MTX_SYSINIT(mprof, &mprof_mtx, "mutex profiling lock", MTX_SPIN | MTX_QUIET); 264 265static u_int64_t 266nanoseconds(void) 267{ 268 struct timespec tv; 269 270 nanotime(&tv); 271 return (tv.tv_sec * (u_int64_t)1000000000 + tv.tv_nsec); 272} 273 274static int 275dump_mutex_prof_stats(SYSCTL_HANDLER_ARGS) 276{ 277 struct sbuf *sb; 278 int error, i; 279 static int multiplier = 1; 280 281 if (first_free_mprof_buf == 0) 282 return (SYSCTL_OUT(req, "No locking recorded", 283 sizeof("No locking recorded"))); 284 285retry_sbufops: 286 sb = sbuf_new(NULL, NULL, MPROF_SBUF_SIZE * multiplier, SBUF_FIXEDLEN); 287 sbuf_printf(sb, "%6s %12s %11s %5s %s\n", 288 "max", "total", "count", "avg", "name"); 289 /* 290 * XXX this spinlock seems to be by far the largest perpetrator 291 * of spinlock latency (1.6 msec on an Athlon1600 was recorded 292 * even before I pessimized it further by moving the average 293 * computation here). 294 */ 295 mtx_lock_spin(&mprof_mtx); 296 for (i = 0; i < first_free_mprof_buf; ++i) { 297 sbuf_printf(sb, "%6ju %12ju %11ju %5ju %s:%d (%s)\n", 298 mprof_buf[i].cnt_max / 1000, 299 mprof_buf[i].cnt_tot / 1000, 300 mprof_buf[i].cnt_cur, 301 mprof_buf[i].cnt_cur == 0 ? (uintmax_t)0 : 302 mprof_buf[i].cnt_tot / (mprof_buf[i].cnt_cur * 1000), 303 mprof_buf[i].file, mprof_buf[i].line, mprof_buf[i].name); 304 if (sbuf_overflowed(sb)) { 305 mtx_unlock_spin(&mprof_mtx); 306 sbuf_delete(sb); 307 multiplier++; 308 goto retry_sbufops; 309 } 310 } 311 mtx_unlock_spin(&mprof_mtx); 312 sbuf_finish(sb); 313 error = SYSCTL_OUT(req, sbuf_data(sb), sbuf_len(sb) + 1); 314 sbuf_delete(sb); 315 return (error); 316} 317SYSCTL_PROC(_debug_mutex_prof, OID_AUTO, stats, CTLTYPE_STRING | CTLFLAG_RD, 318 NULL, 0, dump_mutex_prof_stats, "A", "Mutex profiling statistics"); 319#endif 320 321/* 322 * Function versions of the inlined __mtx_* macros. These are used by 323 * modules and can also be called from assembly language if needed. 324 */ 325void 326_mtx_lock_flags(struct mtx *m, int opts, const char *file, int line) 327{ 328 329 MPASS(curthread != NULL); 330 KASSERT(m->mtx_object.lo_class == &lock_class_mtx_sleep, 331 ("mtx_lock() of spin mutex %s @ %s:%d", m->mtx_object.lo_name, 332 file, line)); 333 _get_sleep_lock(m, curthread, opts, file, line); 334 LOCK_LOG_LOCK("LOCK", &m->mtx_object, opts, m->mtx_recurse, file, 335 line); 336 WITNESS_LOCK(&m->mtx_object, opts | LOP_EXCLUSIVE, file, line); 337#ifdef MUTEX_PROFILING 338 /* don't reset the timer when/if recursing */ 339 if (m->mtx_acqtime == 0) { 340 m->mtx_filename = file; 341 m->mtx_lineno = line; 342 m->mtx_acqtime = mutex_prof_enable ? nanoseconds() : 0; 343 ++mutex_prof_acquisitions; 344 } 345#endif 346} 347 348void 349_mtx_unlock_flags(struct mtx *m, int opts, const char *file, int line) 350{ 351 352 MPASS(curthread != NULL); 353 KASSERT(m->mtx_object.lo_class == &lock_class_mtx_sleep, 354 ("mtx_unlock() of spin mutex %s @ %s:%d", m->mtx_object.lo_name, 355 file, line)); 356 WITNESS_UNLOCK(&m->mtx_object, opts | LOP_EXCLUSIVE, file, line); 357 LOCK_LOG_LOCK("UNLOCK", &m->mtx_object, opts, m->mtx_recurse, file, 358 line); 359 mtx_assert(m, MA_OWNED); 360#ifdef MUTEX_PROFILING 361 if (m->mtx_acqtime != 0) { 362 static const char *unknown = "(unknown)"; 363 struct mutex_prof *mpp; 364 u_int64_t acqtime, now; 365 const char *p, *q; 366 volatile u_int hash; 367 368 now = nanoseconds(); 369 acqtime = m->mtx_acqtime; 370 m->mtx_acqtime = 0; 371 if (now <= acqtime) 372 goto out; 373 for (p = m->mtx_filename; 374 p != NULL && strncmp(p, "../", 3) == 0; p += 3) 375 /* nothing */ ; 376 if (p == NULL || *p == '\0') 377 p = unknown; 378 for (hash = m->mtx_lineno, q = p; *q != '\0'; ++q) 379 hash = (hash * 2 + *q) % MPROF_HASH_SIZE; 380 mtx_lock_spin(&mprof_mtx); 381 for (mpp = mprof_hash[hash]; mpp != NULL; mpp = mpp->next) 382 if (mpp->line == m->mtx_lineno && 383 strcmp(mpp->file, p) == 0) 384 break; 385 if (mpp == NULL) { 386 /* Just exit if we cannot get a trace buffer */ 387 if (first_free_mprof_buf >= NUM_MPROF_BUFFERS) { 388 ++mutex_prof_rejected; 389 goto unlock; 390 } 391 mpp = &mprof_buf[first_free_mprof_buf++]; 392 mpp->name = mtx_name(m); 393 mpp->file = p; 394 mpp->line = m->mtx_lineno; 395 mpp->next = mprof_hash[hash]; 396 if (mprof_hash[hash] != NULL) 397 ++mutex_prof_collisions; 398 mprof_hash[hash] = mpp; 399 ++mutex_prof_records; 400 } 401 /* 402 * Record if the mutex has been held longer now than ever 403 * before. 404 */ 405 if (now - acqtime > mpp->cnt_max) 406 mpp->cnt_max = now - acqtime; 407 mpp->cnt_tot += now - acqtime; 408 mpp->cnt_cur++; 409unlock: 410 mtx_unlock_spin(&mprof_mtx); 411 } 412out: 413#endif 414 _rel_sleep_lock(m, curthread, opts, file, line); 415} 416 417void 418_mtx_lock_spin_flags(struct mtx *m, int opts, const char *file, int line) 419{ 420 421 MPASS(curthread != NULL); 422 KASSERT(m->mtx_object.lo_class == &lock_class_mtx_spin, 423 ("mtx_lock_spin() of sleep mutex %s @ %s:%d", 424 m->mtx_object.lo_name, file, line)); 425#if defined(SMP) || LOCK_DEBUG > 0 || 1 426 _get_spin_lock(m, curthread, opts, file, line); 427#else 428 critical_enter(); 429#endif 430 LOCK_LOG_LOCK("LOCK", &m->mtx_object, opts, m->mtx_recurse, file, 431 line); 432 WITNESS_LOCK(&m->mtx_object, opts | LOP_EXCLUSIVE, file, line); 433} 434 435void 436_mtx_unlock_spin_flags(struct mtx *m, int opts, const char *file, int line) 437{ 438 439 MPASS(curthread != NULL); 440 KASSERT(m->mtx_object.lo_class == &lock_class_mtx_spin, 441 ("mtx_unlock_spin() of sleep mutex %s @ %s:%d", 442 m->mtx_object.lo_name, file, line)); 443 WITNESS_UNLOCK(&m->mtx_object, opts | LOP_EXCLUSIVE, file, line); 444 LOCK_LOG_LOCK("UNLOCK", &m->mtx_object, opts, m->mtx_recurse, file, 445 line); 446 mtx_assert(m, MA_OWNED); 447#if defined(SMP) || LOCK_DEBUG > 0 || 1 448 _rel_spin_lock(m); 449#else 450 critical_exit(); 451#endif 452} 453 454/* 455 * The important part of mtx_trylock{,_flags}() 456 * Tries to acquire lock `m.' We do NOT handle recursion here. If this 457 * function is called on a recursed mutex, it will return failure and 458 * will not recursively acquire the lock. You are expected to know what 459 * you are doing. 460 */ 461int 462_mtx_trylock(struct mtx *m, int opts, const char *file, int line) 463{ 464 int rval; 465 466 MPASS(curthread != NULL); 467 468 rval = _obtain_lock(m, curthread); 469 470 LOCK_LOG_TRY("LOCK", &m->mtx_object, opts, rval, file, line); 471 if (rval) 472 WITNESS_LOCK(&m->mtx_object, opts | LOP_EXCLUSIVE | LOP_TRYLOCK, 473 file, line); 474 475 return (rval); 476} 477 478/* 479 * _mtx_lock_sleep: the tougher part of acquiring an MTX_DEF lock. 480 * 481 * We call this if the lock is either contested (i.e. we need to go to 482 * sleep waiting for it), or if we need to recurse on it. 483 */ 484void 485_mtx_lock_sleep(struct mtx *m, int opts, const char *file, int line) 486{ 487 struct thread *td = curthread; 488 struct thread *td1; 489#if defined(SMP) && defined(ADAPTIVE_MUTEXES) 490 struct thread *owner; 491#endif 492 uintptr_t v; 493#ifdef KTR 494 int cont_logged = 0; 495#endif 496 497 if (mtx_owned(m)) { 498 m->mtx_recurse++; 499 atomic_set_ptr(&m->mtx_lock, MTX_RECURSED); 500 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 501 CTR1(KTR_LOCK, "_mtx_lock_sleep: %p recursing", m); 502 return; 503 } 504 505 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 506 CTR4(KTR_LOCK, 507 "_mtx_lock_sleep: %s contested (lock=%p) at %s:%d", 508 m->mtx_object.lo_name, (void *)m->mtx_lock, file, line); 509 510 while (!_obtain_lock(m, td)) { 511 512 mtx_lock_spin(&sched_lock); 513 v = m->mtx_lock; 514 515 /* 516 * Check if the lock has been released while spinning for 517 * the sched_lock. 518 */ 519 if (v == MTX_UNOWNED) { 520 mtx_unlock_spin(&sched_lock); 521#ifdef __i386__ 522 ia32_pause(); 523#endif 524 continue; 525 } 526 527 /* 528 * The mutex was marked contested on release. This means that 529 * there are other threads blocked on it. Grab ownership of 530 * it and propagate its priority to the current thread if 531 * necessary. 532 */ 533 if (v == MTX_CONTESTED) { 534 td1 = TAILQ_FIRST(&m->mtx_blocked); 535 MPASS(td1 != NULL); 536 m->mtx_lock = (uintptr_t)td | MTX_CONTESTED; 537 LIST_INSERT_HEAD(&td->td_contested, m, mtx_contested); 538 539 if (td1->td_priority < td->td_priority) 540 td->td_priority = td1->td_priority; 541 mtx_unlock_spin(&sched_lock); 542 return; 543 } 544 545 /* 546 * If the mutex isn't already contested and a failure occurs 547 * setting the contested bit, the mutex was either released 548 * or the state of the MTX_RECURSED bit changed. 549 */ 550 if ((v & MTX_CONTESTED) == 0 && 551 !atomic_cmpset_ptr(&m->mtx_lock, (void *)v, 552 (void *)(v | MTX_CONTESTED))) { 553 mtx_unlock_spin(&sched_lock); 554#ifdef __i386__ 555 ia32_pause(); 556#endif 557 continue; 558 } 559 560#if defined(SMP) && defined(ADAPTIVE_MUTEXES) 561 /* 562 * If the current owner of the lock is executing on another 563 * CPU, spin instead of blocking. 564 */ 565 owner = (struct thread *)(v & MTX_FLAGMASK); 566 if (m != &Giant && TD_IS_RUNNING(owner)) { 567 mtx_unlock_spin(&sched_lock); 568 while (mtx_owner(m) == owner && TD_IS_RUNNING(owner)) { 569#ifdef __i386__ 570 ia32_pause(); 571#endif 572 } 573 continue; 574 } 575#endif /* SMP && ADAPTIVE_MUTEXES */ 576 577 /* 578 * We definitely must sleep for this lock. 579 */ 580 mtx_assert(m, MA_NOTOWNED); 581 582#ifdef notyet 583 /* 584 * If we're borrowing an interrupted thread's VM context, we 585 * must clean up before going to sleep. 586 */ 587 if (td->td_ithd != NULL) { 588 struct ithd *it = td->td_ithd; 589 590 if (it->it_interrupted) { 591 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 592 CTR2(KTR_LOCK, 593 "_mtx_lock_sleep: %p interrupted %p", 594 it, it->it_interrupted); 595 intr_thd_fixup(it); 596 } 597 } 598#endif 599 600 /* 601 * Put us on the list of threads blocked on this mutex 602 * and add this mutex to the owning thread's list of 603 * contested mutexes if needed. 604 */ 605 if (TAILQ_EMPTY(&m->mtx_blocked)) { 606 td1 = mtx_owner(m); 607 LIST_INSERT_HEAD(&td1->td_contested, m, mtx_contested); 608 TAILQ_INSERT_TAIL(&m->mtx_blocked, td, td_lockq); 609 } else { 610 TAILQ_FOREACH(td1, &m->mtx_blocked, td_lockq) 611 if (td1->td_priority > td->td_priority) 612 break; 613 if (td1) 614 TAILQ_INSERT_BEFORE(td1, td, td_lockq); 615 else 616 TAILQ_INSERT_TAIL(&m->mtx_blocked, td, td_lockq); 617 } 618#ifdef KTR 619 if (!cont_logged) { 620 CTR6(KTR_CONTENTION, 621 "contention: %p at %s:%d wants %s, taken by %s:%d", 622 td, file, line, m->mtx_object.lo_name, 623 WITNESS_FILE(&m->mtx_object), 624 WITNESS_LINE(&m->mtx_object)); 625 cont_logged = 1; 626 } 627#endif 628 629 /* 630 * Save who we're blocked on. 631 */ 632 td->td_blocked = m; 633 td->td_lockname = m->mtx_object.lo_name; 634 TD_SET_LOCK(td); 635 propagate_priority(td); 636 637 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 638 CTR3(KTR_LOCK, 639 "_mtx_lock_sleep: p %p blocked on [%p] %s", td, m, 640 m->mtx_object.lo_name); 641 642 td->td_proc->p_stats->p_ru.ru_nvcsw++; 643 mi_switch(); 644 645 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 646 CTR3(KTR_LOCK, 647 "_mtx_lock_sleep: p %p free from blocked on [%p] %s", 648 td, m, m->mtx_object.lo_name); 649 650 mtx_unlock_spin(&sched_lock); 651 } 652 653#ifdef KTR 654 if (cont_logged) { 655 CTR4(KTR_CONTENTION, 656 "contention end: %s acquired by %p at %s:%d", 657 m->mtx_object.lo_name, td, file, line); 658 } 659#endif 660 return; 661} 662 663/* 664 * _mtx_lock_spin: the tougher part of acquiring an MTX_SPIN lock. 665 * 666 * This is only called if we need to actually spin for the lock. Recursion 667 * is handled inline. 668 */ 669void 670_mtx_lock_spin(struct mtx *m, int opts, const char *file, int line) 671{ 672 int i = 0; 673 674 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 675 CTR1(KTR_LOCK, "_mtx_lock_spin: %p spinning", m); 676 677 for (;;) { 678 if (_obtain_lock(m, curthread)) 679 break; 680 681 /* Give interrupts a chance while we spin. */ 682 critical_exit(); 683 while (m->mtx_lock != MTX_UNOWNED) { 684 if (i++ < 10000000) { 685#ifdef __i386__ 686 ia32_pause(); 687#endif 688 continue; 689 } 690 if (i < 60000000) 691 DELAY(1); 692#ifdef DDB 693 else if (!db_active) { 694#else 695 else { 696#endif 697 printf("spin lock %s held by %p for > 5 seconds\n", 698 m->mtx_object.lo_name, (void *)m->mtx_lock); 699#ifdef WITNESS 700 witness_display_spinlock(&m->mtx_object, 701 mtx_owner(m)); 702#endif 703 panic("spin lock held too long"); 704 } 705#ifdef __i386__ 706 ia32_pause(); 707#endif 708 } 709 critical_enter(); 710 } 711 712 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 713 CTR1(KTR_LOCK, "_mtx_lock_spin: %p spin done", m); 714 715 return; 716} 717 718/* 719 * _mtx_unlock_sleep: the tougher part of releasing an MTX_DEF lock. 720 * 721 * We are only called here if the lock is recursed or contested (i.e. we 722 * need to wake up a blocked thread). 723 */ 724void 725_mtx_unlock_sleep(struct mtx *m, int opts, const char *file, int line) 726{ 727 struct thread *td, *td1; 728 struct mtx *m1; 729 int pri; 730 731 td = curthread; 732 733 if (mtx_recursed(m)) { 734 if (--(m->mtx_recurse) == 0) 735 atomic_clear_ptr(&m->mtx_lock, MTX_RECURSED); 736 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 737 CTR1(KTR_LOCK, "_mtx_unlock_sleep: %p unrecurse", m); 738 return; 739 } 740 741 mtx_lock_spin(&sched_lock); 742 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 743 CTR1(KTR_LOCK, "_mtx_unlock_sleep: %p contested", m); 744 745 td1 = TAILQ_FIRST(&m->mtx_blocked); 746#if defined(SMP) && defined(ADAPTIVE_MUTEXES) 747 if (td1 == NULL) { 748 _release_lock_quick(m); 749 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 750 CTR1(KTR_LOCK, "_mtx_unlock_sleep: %p no sleepers", m); 751 mtx_unlock_spin(&sched_lock); 752 return; 753 } 754#endif 755 MPASS(td->td_proc->p_magic == P_MAGIC); 756 MPASS(td1->td_proc->p_magic == P_MAGIC); 757 758 TAILQ_REMOVE(&m->mtx_blocked, td1, td_lockq); 759 760 LIST_REMOVE(m, mtx_contested); 761 if (TAILQ_EMPTY(&m->mtx_blocked)) { 762 _release_lock_quick(m); 763 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 764 CTR1(KTR_LOCK, "_mtx_unlock_sleep: %p not held", m); 765 } else 766 m->mtx_lock = MTX_CONTESTED; 767 768 pri = PRI_MAX; 769 LIST_FOREACH(m1, &td->td_contested, mtx_contested) { 770 int cp = TAILQ_FIRST(&m1->mtx_blocked)->td_priority; 771 if (cp < pri) 772 pri = cp; 773 } 774 775 if (pri > td->td_base_pri) 776 pri = td->td_base_pri; 777 td->td_priority = pri; 778 779 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 780 CTR2(KTR_LOCK, "_mtx_unlock_sleep: %p contested setrunqueue %p", 781 m, td1); 782 783 td1->td_blocked = NULL; 784 TD_CLR_LOCK(td1); 785 if (!TD_CAN_RUN(td1)) { 786 mtx_unlock_spin(&sched_lock); 787 return; 788 } 789 setrunqueue(td1); 790 791 if (td->td_critnest == 1 && td1->td_priority < pri) { 792#ifdef notyet 793 if (td->td_ithd != NULL) { 794 struct ithd *it = td->td_ithd; 795 796 if (it->it_interrupted) { 797 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 798 CTR2(KTR_LOCK, 799 "_mtx_unlock_sleep: %p interrupted %p", 800 it, it->it_interrupted); 801 intr_thd_fixup(it); 802 } 803 } 804#endif 805 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 806 CTR2(KTR_LOCK, 807 "_mtx_unlock_sleep: %p switching out lock=%p", m, 808 (void *)m->mtx_lock); 809 810 td->td_proc->p_stats->p_ru.ru_nivcsw++; 811 mi_switch(); 812 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 813 CTR2(KTR_LOCK, "_mtx_unlock_sleep: %p resuming lock=%p", 814 m, (void *)m->mtx_lock); 815 } 816 817 mtx_unlock_spin(&sched_lock); 818 819 return; 820} 821 822/* 823 * All the unlocking of MTX_SPIN locks is done inline. 824 * See the _rel_spin_lock() macro for the details. 825 */ 826 827/* 828 * The backing function for the INVARIANTS-enabled mtx_assert() 829 */ 830#ifdef INVARIANT_SUPPORT 831void 832_mtx_assert(struct mtx *m, int what, const char *file, int line) 833{ 834 835 if (panicstr != NULL) 836 return; 837 switch (what) { 838 case MA_OWNED: 839 case MA_OWNED | MA_RECURSED: 840 case MA_OWNED | MA_NOTRECURSED: 841 if (!mtx_owned(m)) 842 panic("mutex %s not owned at %s:%d", 843 m->mtx_object.lo_name, file, line); 844 if (mtx_recursed(m)) { 845 if ((what & MA_NOTRECURSED) != 0) 846 panic("mutex %s recursed at %s:%d", 847 m->mtx_object.lo_name, file, line); 848 } else if ((what & MA_RECURSED) != 0) { 849 panic("mutex %s unrecursed at %s:%d", 850 m->mtx_object.lo_name, file, line); 851 } 852 break; 853 case MA_NOTOWNED: 854 if (mtx_owned(m)) 855 panic("mutex %s owned at %s:%d", 856 m->mtx_object.lo_name, file, line); 857 break; 858 default: 859 panic("unknown mtx_assert at %s:%d", file, line); 860 } 861} 862#endif 863 864/* 865 * The MUTEX_DEBUG-enabled mtx_validate() 866 * 867 * Most of these checks have been moved off into the LO_INITIALIZED flag 868 * maintained by the witness code. 869 */ 870#ifdef MUTEX_DEBUG 871 872void mtx_validate(struct mtx *); 873 874void 875mtx_validate(struct mtx *m) 876{ 877 878/* 879 * XXX: When kernacc() does not require Giant we can reenable this check 880 */ 881#ifdef notyet 882/* 883 * XXX - When kernacc() is fixed on the alpha to handle K0_SEG memory properly 884 * we can re-enable the kernacc() checks. 885 */ 886#ifndef __alpha__ 887 /* 888 * Can't call kernacc() from early init386(), especially when 889 * initializing Giant mutex, because some stuff in kernacc() 890 * requires Giant itself. 891 */ 892 if (!cold) 893 if (!kernacc((caddr_t)m, sizeof(m), 894 VM_PROT_READ | VM_PROT_WRITE)) 895 panic("Can't read and write to mutex %p", m); 896#endif 897#endif 898} 899#endif 900 901/* 902 * General init routine used by the MTX_SYSINIT() macro. 903 */ 904void 905mtx_sysinit(void *arg) 906{ 907 struct mtx_args *margs = arg; 908 909 mtx_init(margs->ma_mtx, margs->ma_desc, NULL, margs->ma_opts); 910} 911 912/* 913 * Mutex initialization routine; initialize lock `m' of type contained in 914 * `opts' with options contained in `opts' and name `name.' The optional 915 * lock type `type' is used as a general lock category name for use with 916 * witness. 917 */ 918void 919mtx_init(struct mtx *m, const char *name, const char *type, int opts) 920{ 921 struct lock_object *lock; 922 923 MPASS((opts & ~(MTX_SPIN | MTX_QUIET | MTX_RECURSE | 924 MTX_NOWITNESS | MTX_DUPOK)) == 0); 925 926#ifdef MUTEX_DEBUG 927 /* Diagnostic and error correction */ 928 mtx_validate(m); 929#endif 930 931 lock = &m->mtx_object; 932 KASSERT((lock->lo_flags & LO_INITIALIZED) == 0, 933 ("mutex \"%s\" %p already initialized", name, m)); 934 bzero(m, sizeof(*m)); 935 if (opts & MTX_SPIN) 936 lock->lo_class = &lock_class_mtx_spin; 937 else 938 lock->lo_class = &lock_class_mtx_sleep; 939 lock->lo_name = name; 940 lock->lo_type = type != NULL ? type : name; 941 if (opts & MTX_QUIET) 942 lock->lo_flags = LO_QUIET; 943 if (opts & MTX_RECURSE) 944 lock->lo_flags |= LO_RECURSABLE; 945 if ((opts & MTX_NOWITNESS) == 0) 946 lock->lo_flags |= LO_WITNESS; 947 if (opts & MTX_DUPOK) 948 lock->lo_flags |= LO_DUPOK; 949 950 m->mtx_lock = MTX_UNOWNED; 951 TAILQ_INIT(&m->mtx_blocked); 952 953 LOCK_LOG_INIT(lock, opts); 954 955 WITNESS_INIT(lock); 956} 957 958/* 959 * Remove lock `m' from all_mtx queue. We don't allow MTX_QUIET to be 960 * passed in as a flag here because if the corresponding mtx_init() was 961 * called with MTX_QUIET set, then it will already be set in the mutex's 962 * flags. 963 */ 964void 965mtx_destroy(struct mtx *m) 966{ 967 968 LOCK_LOG_DESTROY(&m->mtx_object, 0); 969 970 if (!mtx_owned(m)) 971 MPASS(mtx_unowned(m)); 972 else { 973 MPASS((m->mtx_lock & (MTX_RECURSED|MTX_CONTESTED)) == 0); 974 975 /* Tell witness this isn't locked to make it happy. */ 976 WITNESS_UNLOCK(&m->mtx_object, LOP_EXCLUSIVE, __FILE__, 977 __LINE__); 978 } 979 980 WITNESS_DESTROY(&m->mtx_object); 981} 982 983/* 984 * Intialize the mutex code and system mutexes. This is called from the MD 985 * startup code prior to mi_startup(). The per-CPU data space needs to be 986 * setup before this is called. 987 */ 988void 989mutex_init(void) 990{ 991 992 /* Setup thread0 so that mutexes work. */ 993 LIST_INIT(&thread0.td_contested); 994 995 /* 996 * Initialize mutexes. 997 */ 998 mtx_init(&Giant, "Giant", NULL, MTX_DEF | MTX_RECURSE); 999 mtx_init(&sched_lock, "sched lock", NULL, MTX_SPIN | MTX_RECURSE); 1000 mtx_init(&proc0.p_mtx, "process lock", NULL, MTX_DEF | MTX_DUPOK); 1001 mtx_lock(&Giant); 1002} 1003