kern_mutex.c revision 100754
1151497Sru/*- 275584Sru * Copyright (c) 1998 Berkeley Software Design, Inc. All rights reserved. 375584Sru * 475584Sru * Redistribution and use in source and binary forms, with or without 575584Sru * modification, are permitted provided that the following conditions 675584Sru * are met: 775584Sru * 1. Redistributions of source code must retain the above copyright 875584Sru * notice, this list of conditions and the following disclaimer. 975584Sru * 2. Redistributions in binary form must reproduce the above copyright 1075584Sru * notice, this list of conditions and the following disclaimer in the 1175584Sru * documentation and/or other materials provided with the distribution. 1275584Sru * 3. Berkeley Software Design Inc's name may not be used to endorse or 1375584Sru * promote products derived from this software without specific prior 1475584Sru * written permission. 1575584Sru * 1675584Sru * THIS SOFTWARE IS PROVIDED BY BERKELEY SOFTWARE DESIGN INC ``AS IS'' AND 1775584Sru * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 18151497Sru * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 1975584Sru * ARE DISCLAIMED. IN NO EVENT SHALL BERKELEY SOFTWARE DESIGN INC BE LIABLE 2075584Sru * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 2175584Sru * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 2275584Sru * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 2375584Sru * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 2475584Sru * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 2575584Sru * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 2675584Sru * SUCH DAMAGE. 2775584Sru * 2875584Sru * from BSDI $Id: mutex_witness.c,v 1.1.2.20 2000/04/27 03:10:27 cp Exp $ 2975584Sru * and BSDI $Id: synch_machdep.c,v 2.3.2.39 2000/04/27 03:10:25 cp Exp $ 3075584Sru * $FreeBSD: head/sys/kern/kern_mutex.c 100754 2002-07-27 16:54:23Z jhb $ 3175584Sru */ 3275584Sru 3375584Sru/* 3475584Sru * Machine independent bits of mutex implementation. 3575584Sru */ 3675584Sru 3775584Sru#include "opt_adaptive_mutexes.h" 3875584Sru#include "opt_ddb.h" 3975584Sru 40114402Sru#include <sys/param.h> 4175584Sru#include <sys/systm.h> 4275584Sru#include <sys/bus.h> 4375584Sru#include <sys/kernel.h> 4475584Sru#include <sys/ktr.h> 45114402Sru#include <sys/lock.h> 4675584Sru#include <sys/malloc.h> 4775584Sru#include <sys/mutex.h> 4875584Sru#include <sys/proc.h> 4975584Sru#include <sys/resourcevar.h> 5075584Sru#include <sys/sbuf.h> 5175584Sru#include <sys/stdint.h> 5275584Sru#include <sys/sysctl.h> 5375584Sru#include <sys/vmmeter.h> 5475584Sru 55104862Sru#include <machine/atomic.h> 5675584Sru#include <machine/bus.h> 5775584Sru#include <machine/clock.h> 5875584Sru#include <machine/cpu.h> 59104862Sru 6075584Sru#include <ddb/ddb.h> 6175584Sru 6275584Sru#include <vm/vm.h> 6375584Sru#include <vm/vm_extern.h> 6475584Sru 6575584Sru/* 6675584Sru * Internal utility macros. 6775584Sru */ 6875584Sru#define mtx_unowned(m) ((m)->mtx_lock == MTX_UNOWNED) 6975584Sru 7075584Sru#define mtx_owner(m) (mtx_unowned((m)) ? NULL \ 71104862Sru : (struct thread *)((m)->mtx_lock & MTX_FLAGMASK)) 7275584Sru 7375584Sru/* XXXKSE This test will change. */ 7475584Sru#define thread_running(td) \ 75104862Sru ((td)->td_kse != NULL && (td)->td_kse->ke_oncpu != NOCPU) 7675584Sru 7775584Sru/* 7875584Sru * Lock classes for sleep and spin mutexes. 7975584Sru */ 8075584Srustruct lock_class lock_class_mtx_sleep = { 8175584Sru "sleep mutex", 8275584Sru LC_SLEEPLOCK | LC_RECURSABLE 8375584Sru}; 8475584Srustruct lock_class lock_class_mtx_spin = { 8575584Sru "spin mutex", 8675584Sru LC_SPINLOCK | LC_RECURSABLE 87104862Sru}; 8875584Sru 8975584Sru/* 9075584Sru * System-wide mutexes 91104862Sru */ 9275584Srustruct mtx sched_lock; 9375584Srustruct mtx Giant; 9475584Sru 9575584Sru/* 9675584Sru * Prototypes for non-exported routines. 9775584Sru */ 9875584Srustatic void propagate_priority(struct thread *); 9975584Sru 10075584Srustatic void 10175584Srupropagate_priority(struct thread *td) 10275584Sru{ 103104862Sru int pri = td->td_priority; 10475584Sru struct mtx *m = td->td_blocked; 10575584Sru 10675584Sru mtx_assert(&sched_lock, MA_OWNED); 107104862Sru for (;;) { 10875584Sru struct thread *td1; 10975584Sru 11075584Sru td = mtx_owner(m); 11175584Sru 11275584Sru if (td == NULL) { 11375584Sru /* 11475584Sru * This really isn't quite right. Really 11575584Sru * ought to bump priority of thread that 11675584Sru * next acquires the mutex. 11775584Sru */ 11875584Sru MPASS(m->mtx_lock == MTX_CONTESTED); 119104862Sru return; 12075584Sru } 12175584Sru 12275584Sru KASSERT(td->td_state != TDS_SURPLUS, ("Mutex owner SURPLUS")); 123104862Sru MPASS(td->td_proc != NULL); 12475584Sru MPASS(td->td_proc->p_magic == P_MAGIC); 12575584Sru KASSERT(td->td_state != TDS_SLP, 12675584Sru ("sleeping thread owns a mutex")); 12775584Sru if (td->td_priority <= pri) /* lower is higher priority */ 12875584Sru return; 12975584Sru 13075584Sru 13175584Sru /* 13275584Sru * If lock holder is actually running, just bump priority. 13375584Sru */ 13475584Sru if (td->td_state == TDS_RUNNING) { 135104862Sru td->td_priority = pri; 13675584Sru return; 13775584Sru } 13875584Sru 139104862Sru#ifndef SMP 14075584Sru /* 14175584Sru * For UP, we check to see if td is curthread (this shouldn't 14275584Sru * ever happen however as it would mean we are in a deadlock.) 14375584Sru */ 14475584Sru KASSERT(td != curthread, ("Deadlock detected")); 14575584Sru#endif 14675584Sru 14775584Sru /* 14875584Sru * If on run queue move to new run queue, and quit. 14975584Sru * XXXKSE this gets a lot more complicated under threads 15075584Sru * but try anyhow. 151104862Sru * We should have a special call to do this more efficiently. 15275584Sru */ 15375584Sru if (td->td_state == TDS_RUNQ) { 15475584Sru MPASS(td->td_blocked == NULL); 155104862Sru remrunqueue(td); 15675584Sru td->td_priority = pri; 15775584Sru setrunqueue(td); 15875584Sru return; 15975584Sru } 16075584Sru /* 16175584Sru * Adjust for any other cases. 16275584Sru */ 16375584Sru td->td_priority = pri; 16475584Sru 16575584Sru /* 16675584Sru * If we aren't blocked on a mutex, we should be. 167104862Sru */ 16875584Sru KASSERT(td->td_state == TDS_MTX, ( 16975584Sru "process %d(%s):%d holds %s but isn't blocked on a mutex\n", 17075584Sru td->td_proc->p_pid, td->td_proc->p_comm, td->td_state, 171104862Sru m->mtx_object.lo_name)); 17275584Sru 17375584Sru /* 17475584Sru * Pick up the mutex that td is blocked on. 17575584Sru */ 17675584Sru m = td->td_blocked; 17775584Sru MPASS(m != NULL); 17875584Sru 17975584Sru /* 18075584Sru * Check if the thread needs to be moved up on 18175584Sru * the blocked chain 18275584Sru */ 18375584Sru if (td == TAILQ_FIRST(&m->mtx_blocked)) { 18475584Sru continue; 18575584Sru } 18675584Sru 187104862Sru td1 = TAILQ_PREV(td, threadqueue, td_blkq); 18875584Sru if (td1->td_priority <= pri) { 18975584Sru continue; 19075584Sru } 191151497Sru 192151497Sru /* 19375584Sru * Remove thread from blocked chain and determine where 19475584Sru * it should be moved up to. Since we know that td1 has 195104862Sru * a lower priority than td, we know that at least one 19675584Sru * thread in the chain has a lower priority and that 19775584Sru * td1 will thus not be NULL after the loop. 198104862Sru */ 19975584Sru TAILQ_REMOVE(&m->mtx_blocked, td, td_blkq); 200114402Sru TAILQ_FOREACH(td1, &m->mtx_blocked, td_blkq) { 201114402Sru MPASS(td1->td_proc->p_magic == P_MAGIC); 202114402Sru if (td1->td_priority > pri) 203114402Sru break; 204114402Sru } 205114402Sru 206114402Sru MPASS(td1 != NULL); 20775584Sru TAILQ_INSERT_BEFORE(td1, td, td_blkq); 20875584Sru CTR4(KTR_LOCK, 20975584Sru "propagate_priority: p %p moved before %p on [%p] %s", 21075584Sru td, td1, m, m->mtx_object.lo_name); 21175584Sru } 21275584Sru} 21375584Sru 21475584Sru#ifdef MUTEX_PROFILING 21575584SruSYSCTL_NODE(_debug, OID_AUTO, mutex, CTLFLAG_RD, NULL, "mutex debugging"); 21675584SruSYSCTL_NODE(_debug_mutex, OID_AUTO, prof, CTLFLAG_RD, NULL, "mutex profiling"); 21775584Srustatic int mutex_prof_enable = 0; 21875584SruSYSCTL_INT(_debug_mutex_prof, OID_AUTO, enable, CTLFLAG_RW, 21975584Sru &mutex_prof_enable, 0, "Enable tracing of mutex holdtime"); 22075584Sru 22175584Srustruct mutex_prof { 22275584Sru const char *name; 22375584Sru const char *file; 224104862Sru int line; 225104862Sru#define MPROF_MAX 0 226104862Sru#define MPROF_TOT 1 227104862Sru#define MPROF_CNT 2 22875584Sru#define MPROF_AVG 3 22975584Sru uintmax_t counter[4]; 23075584Sru struct mutex_prof *next; 23175584Sru}; 23275584Sru 23375584Sru/* 23475584Sru * mprof_buf is a static pool of profiling records to avoid possible 23575584Sru * reentrance of the memory allocation functions. 23675584Sru * 23775584Sru * Note: NUM_MPROF_BUFFERS must be smaller than MPROF_HASH_SIZE. 23875584Sru */ 23975584Sru#define NUM_MPROF_BUFFERS 1000 240static struct mutex_prof mprof_buf[NUM_MPROF_BUFFERS]; 241static int first_free_mprof_buf; 242#define MPROF_HASH_SIZE 1009 243static struct mutex_prof *mprof_hash[MPROF_HASH_SIZE]; 244 245static int mutex_prof_acquisitions; 246SYSCTL_INT(_debug_mutex_prof, OID_AUTO, acquisitions, CTLFLAG_RD, 247 &mutex_prof_acquisitions, 0, "Number of mutex acquistions recorded"); 248static int mutex_prof_records; 249SYSCTL_INT(_debug_mutex_prof, OID_AUTO, records, CTLFLAG_RD, 250 &mutex_prof_records, 0, "Number of profiling records"); 251static int mutex_prof_maxrecords = NUM_MPROF_BUFFERS; 252SYSCTL_INT(_debug_mutex_prof, OID_AUTO, maxrecords, CTLFLAG_RD, 253 &mutex_prof_maxrecords, 0, "Maximum number of profiling records"); 254static int mutex_prof_rejected; 255SYSCTL_INT(_debug_mutex_prof, OID_AUTO, rejected, CTLFLAG_RD, 256 &mutex_prof_rejected, 0, "Number of rejected profiling records"); 257static int mutex_prof_hashsize = MPROF_HASH_SIZE; 258SYSCTL_INT(_debug_mutex_prof, OID_AUTO, hashsize, CTLFLAG_RD, 259 &mutex_prof_hashsize, 0, "Hash size"); 260static int mutex_prof_collisions = 0; 261SYSCTL_INT(_debug_mutex_prof, OID_AUTO, collisions, CTLFLAG_RD, 262 &mutex_prof_collisions, 0, "Number of hash collisions"); 263 264/* 265 * mprof_mtx protects the profiling buffers and the hash. 266 */ 267static struct mtx mprof_mtx; 268MTX_SYSINIT(mprof, &mprof_mtx, "mutex profiling lock", MTX_SPIN | MTX_QUIET); 269 270static u_int64_t 271nanoseconds(void) 272{ 273 struct timespec tv; 274 275 nanotime(&tv); 276 return (tv.tv_sec * (u_int64_t)1000000000 + tv.tv_nsec); 277} 278 279static int 280dump_mutex_prof_stats(SYSCTL_HANDLER_ARGS) 281{ 282 struct sbuf *sb; 283 int error, i; 284 285 if (first_free_mprof_buf == 0) 286 return SYSCTL_OUT(req, "No locking recorded", 287 sizeof("No locking recorded")); 288 289 sb = sbuf_new(NULL, NULL, 1024, SBUF_AUTOEXTEND); 290 sbuf_printf(sb, "%12s %12s %12s %12s %s\n", 291 "max", "total", "count", "average", "name"); 292 mtx_lock_spin(&mprof_mtx); 293 for (i = 0; i < first_free_mprof_buf; ++i) 294 sbuf_printf(sb, "%12ju %12ju %12ju %12ju %s:%d (%s)\n", 295 mprof_buf[i].counter[MPROF_MAX] / 1000, 296 mprof_buf[i].counter[MPROF_TOT] / 1000, 297 mprof_buf[i].counter[MPROF_CNT], 298 mprof_buf[i].counter[MPROF_AVG] / 1000, 299 mprof_buf[i].file, mprof_buf[i].line, mprof_buf[i].name); 300 mtx_unlock_spin(&mprof_mtx); 301 sbuf_finish(sb); 302 error = SYSCTL_OUT(req, sbuf_data(sb), sbuf_len(sb) + 1); 303 sbuf_delete(sb); 304 return (error); 305} 306SYSCTL_PROC(_debug_mutex_prof, OID_AUTO, stats, CTLTYPE_STRING|CTLFLAG_RD, 307 NULL, 0, dump_mutex_prof_stats, "A", "Mutex profiling statistics"); 308#endif 309 310/* 311 * Function versions of the inlined __mtx_* macros. These are used by 312 * modules and can also be called from assembly language if needed. 313 */ 314void 315_mtx_lock_flags(struct mtx *m, int opts, const char *file, int line) 316{ 317 318 MPASS(curthread != NULL); 319 _get_sleep_lock(m, curthread, opts, file, line); 320 LOCK_LOG_LOCK("LOCK", &m->mtx_object, opts, m->mtx_recurse, file, 321 line); 322 WITNESS_LOCK(&m->mtx_object, opts | LOP_EXCLUSIVE, file, line); 323#ifdef MUTEX_PROFILING 324 /* don't reset the timer when/if recursing */ 325 if (m->mtx_acqtime == 0) { 326 m->mtx_filename = file; 327 m->mtx_lineno = line; 328 m->mtx_acqtime = mutex_prof_enable ? nanoseconds() : 0; 329 ++mutex_prof_acquisitions; 330 } 331#endif 332} 333 334void 335_mtx_unlock_flags(struct mtx *m, int opts, const char *file, int line) 336{ 337 338 MPASS(curthread != NULL); 339 mtx_assert(m, MA_OWNED); 340#ifdef MUTEX_PROFILING 341 if (m->mtx_acqtime != 0) { 342 static const char *unknown = "(unknown)"; 343 struct mutex_prof *mpp; 344 u_int64_t acqtime, now; 345 const char *p, *q; 346 volatile u_int hash; 347 348 now = nanoseconds(); 349 acqtime = m->mtx_acqtime; 350 m->mtx_acqtime = 0; 351 if (now <= acqtime) 352 goto out; 353 for (p = m->mtx_filename; strncmp(p, "../", 3) == 0; p += 3) 354 /* nothing */ ; 355 if (p == NULL || *p == '\0') 356 p = unknown; 357 for (hash = m->mtx_lineno, q = p; *q != '\0'; ++q) 358 hash = (hash * 2 + *q) % MPROF_HASH_SIZE; 359 mtx_lock_spin(&mprof_mtx); 360 for (mpp = mprof_hash[hash]; mpp != NULL; mpp = mpp->next) 361 if (mpp->line == m->mtx_lineno && 362 strcmp(mpp->file, p) == 0) 363 break; 364 if (mpp == NULL) { 365 /* Just exit if we cannot get a trace buffer */ 366 if (first_free_mprof_buf >= NUM_MPROF_BUFFERS) { 367 ++mutex_prof_rejected; 368 goto unlock; 369 } 370 mpp = &mprof_buf[first_free_mprof_buf++]; 371 mpp->name = mtx_name(m); 372 mpp->file = p; 373 mpp->line = m->mtx_lineno; 374 mpp->next = mprof_hash[hash]; 375 if (mprof_hash[hash] != NULL) 376 ++mutex_prof_collisions; 377 mprof_hash[hash] = mpp; 378 ++mutex_prof_records; 379 } 380 /* 381 * Record if the mutex has been held longer now than ever 382 * before 383 */ 384 if ((now - acqtime) > mpp->counter[MPROF_MAX]) 385 mpp->counter[MPROF_MAX] = now - acqtime; 386 mpp->counter[MPROF_TOT] += now - acqtime; 387 mpp->counter[MPROF_CNT] += 1; 388 mpp->counter[MPROF_AVG] = 389 mpp->counter[MPROF_TOT] / mpp->counter[MPROF_CNT]; 390unlock: 391 mtx_unlock_spin(&mprof_mtx); 392 } 393out: 394#endif 395 WITNESS_UNLOCK(&m->mtx_object, opts | LOP_EXCLUSIVE, file, line); 396 LOCK_LOG_LOCK("UNLOCK", &m->mtx_object, opts, m->mtx_recurse, file, 397 line); 398 _rel_sleep_lock(m, curthread, opts, file, line); 399} 400 401void 402_mtx_lock_spin_flags(struct mtx *m, int opts, const char *file, int line) 403{ 404 405 MPASS(curthread != NULL); 406#if defined(SMP) || LOCK_DEBUG > 0 || 1 407 _get_spin_lock(m, curthread, opts, file, line); 408#else 409 critical_enter(); 410#endif 411 LOCK_LOG_LOCK("LOCK", &m->mtx_object, opts, m->mtx_recurse, file, 412 line); 413 WITNESS_LOCK(&m->mtx_object, opts | LOP_EXCLUSIVE, file, line); 414} 415 416void 417_mtx_unlock_spin_flags(struct mtx *m, int opts, const char *file, int line) 418{ 419 420 MPASS(curthread != NULL); 421 mtx_assert(m, MA_OWNED); 422 WITNESS_UNLOCK(&m->mtx_object, opts | LOP_EXCLUSIVE, file, line); 423 LOCK_LOG_LOCK("UNLOCK", &m->mtx_object, opts, m->mtx_recurse, file, 424 line); 425#if defined(SMP) || LOCK_DEBUG > 0 || 1 426 _rel_spin_lock(m); 427#else 428 critical_exit(); 429#endif 430} 431 432/* 433 * The important part of mtx_trylock{,_flags}() 434 * Tries to acquire lock `m.' We do NOT handle recursion here; we assume that 435 * if we're called, it's because we know we don't already own this lock. 436 */ 437int 438_mtx_trylock(struct mtx *m, int opts, const char *file, int line) 439{ 440 int rval; 441 442 MPASS(curthread != NULL); 443 444 rval = _obtain_lock(m, curthread); 445 446 LOCK_LOG_TRY("LOCK", &m->mtx_object, opts, rval, file, line); 447 if (rval) { 448 /* 449 * We do not handle recursion in _mtx_trylock; see the 450 * note at the top of the routine. 451 */ 452 KASSERT(!mtx_recursed(m), 453 ("mtx_trylock() called on a recursed mutex")); 454 WITNESS_LOCK(&m->mtx_object, opts | LOP_EXCLUSIVE | LOP_TRYLOCK, 455 file, line); 456 } 457 458 return (rval); 459} 460 461/* 462 * _mtx_lock_sleep: the tougher part of acquiring an MTX_DEF lock. 463 * 464 * We call this if the lock is either contested (i.e. we need to go to 465 * sleep waiting for it), or if we need to recurse on it. 466 */ 467void 468_mtx_lock_sleep(struct mtx *m, int opts, const char *file, int line) 469{ 470 struct thread *td = curthread; 471#if defined(SMP) && defined(ADAPTIVE_MUTEXES) 472 struct thread *owner; 473#endif 474 475 if ((m->mtx_lock & MTX_FLAGMASK) == (uintptr_t)td) { 476 m->mtx_recurse++; 477 atomic_set_ptr(&m->mtx_lock, MTX_RECURSED); 478 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 479 CTR1(KTR_LOCK, "_mtx_lock_sleep: %p recursing", m); 480 return; 481 } 482 483 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 484 CTR4(KTR_LOCK, 485 "_mtx_lock_sleep: %s contested (lock=%p) at %s:%d", 486 m->mtx_object.lo_name, (void *)m->mtx_lock, file, line); 487 488 while (!_obtain_lock(m, td)) { 489 uintptr_t v; 490 struct thread *td1; 491 492 mtx_lock_spin(&sched_lock); 493 /* 494 * Check if the lock has been released while spinning for 495 * the sched_lock. 496 */ 497 if ((v = m->mtx_lock) == MTX_UNOWNED) { 498 mtx_unlock_spin(&sched_lock); 499#ifdef __i386__ 500 ia32_pause(); 501#endif 502 continue; 503 } 504 505 /* 506 * The mutex was marked contested on release. This means that 507 * there are threads blocked on it. 508 */ 509 if (v == MTX_CONTESTED) { 510 td1 = TAILQ_FIRST(&m->mtx_blocked); 511 MPASS(td1 != NULL); 512 m->mtx_lock = (uintptr_t)td | MTX_CONTESTED; 513 514 if (td1->td_priority < td->td_priority) 515 td->td_priority = td1->td_priority; 516 mtx_unlock_spin(&sched_lock); 517 return; 518 } 519 520 /* 521 * If the mutex isn't already contested and a failure occurs 522 * setting the contested bit, the mutex was either released 523 * or the state of the MTX_RECURSED bit changed. 524 */ 525 if ((v & MTX_CONTESTED) == 0 && 526 !atomic_cmpset_ptr(&m->mtx_lock, (void *)v, 527 (void *)(v | MTX_CONTESTED))) { 528 mtx_unlock_spin(&sched_lock); 529#ifdef __i386__ 530 ia32_pause(); 531#endif 532 continue; 533 } 534 535#if defined(SMP) && defined(ADAPTIVE_MUTEXES) 536 /* 537 * If the current owner of the lock is executing on another 538 * CPU, spin instead of blocking. 539 */ 540 owner = (struct thread *)(v & MTX_FLAGMASK); 541 if (m != &Giant && thread_running(owner)) { 542 mtx_unlock_spin(&sched_lock); 543 while (mtx_owner(m) == owner && thread_running(owner)) { 544#ifdef __i386__ 545 ia32_pause(); 546#endif 547 } 548 continue; 549 } 550#endif /* SMP && ADAPTIVE_MUTEXES */ 551 552 /* 553 * We definitely must sleep for this lock. 554 */ 555 mtx_assert(m, MA_NOTOWNED); 556 557#ifdef notyet 558 /* 559 * If we're borrowing an interrupted thread's VM context, we 560 * must clean up before going to sleep. 561 */ 562 if (td->td_ithd != NULL) { 563 struct ithd *it = td->td_ithd; 564 565 if (it->it_interrupted) { 566 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 567 CTR2(KTR_LOCK, 568 "_mtx_lock_sleep: %p interrupted %p", 569 it, it->it_interrupted); 570 intr_thd_fixup(it); 571 } 572 } 573#endif 574 575 /* 576 * Put us on the list of threads blocked on this mutex. 577 */ 578 if (TAILQ_EMPTY(&m->mtx_blocked)) { 579 td1 = mtx_owner(m); 580 LIST_INSERT_HEAD(&td1->td_contested, m, mtx_contested); 581 TAILQ_INSERT_TAIL(&m->mtx_blocked, td, td_blkq); 582 } else { 583 TAILQ_FOREACH(td1, &m->mtx_blocked, td_blkq) 584 if (td1->td_priority > td->td_priority) 585 break; 586 if (td1) 587 TAILQ_INSERT_BEFORE(td1, td, td_blkq); 588 else 589 TAILQ_INSERT_TAIL(&m->mtx_blocked, td, td_blkq); 590 } 591 592 /* 593 * Save who we're blocked on. 594 */ 595 td->td_blocked = m; 596 td->td_mtxname = m->mtx_object.lo_name; 597 td->td_state = TDS_MTX; 598 propagate_priority(td); 599 600 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 601 CTR3(KTR_LOCK, 602 "_mtx_lock_sleep: p %p blocked on [%p] %s", td, m, 603 m->mtx_object.lo_name); 604 605 td->td_proc->p_stats->p_ru.ru_nvcsw++; 606 mi_switch(); 607 608 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 609 CTR3(KTR_LOCK, 610 "_mtx_lock_sleep: p %p free from blocked on [%p] %s", 611 td, m, m->mtx_object.lo_name); 612 613 mtx_unlock_spin(&sched_lock); 614 } 615 616 return; 617} 618 619/* 620 * _mtx_lock_spin: the tougher part of acquiring an MTX_SPIN lock. 621 * 622 * This is only called if we need to actually spin for the lock. Recursion 623 * is handled inline. 624 */ 625void 626_mtx_lock_spin(struct mtx *m, int opts, const char *file, int line) 627{ 628 int i = 0; 629 630 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 631 CTR1(KTR_LOCK, "_mtx_lock_spin: %p spinning", m); 632 633 for (;;) { 634 if (_obtain_lock(m, curthread)) 635 break; 636 637 /* Give interrupts a chance while we spin. */ 638 critical_exit(); 639 while (m->mtx_lock != MTX_UNOWNED) { 640 if (i++ < 10000000) { 641#ifdef __i386__ 642 ia32_pause(); 643#endif 644 continue; 645 } 646 if (i < 60000000) 647 DELAY(1); 648#ifdef DDB 649 else if (!db_active) 650#else 651 else 652#endif 653 panic("spin lock %s held by %p for > 5 seconds", 654 m->mtx_object.lo_name, (void *)m->mtx_lock); 655#ifdef __i386__ 656 ia32_pause(); 657#endif 658 } 659 critical_enter(); 660 } 661 662 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 663 CTR1(KTR_LOCK, "_mtx_lock_spin: %p spin done", m); 664 665 return; 666} 667 668/* 669 * _mtx_unlock_sleep: the tougher part of releasing an MTX_DEF lock. 670 * 671 * We are only called here if the lock is recursed or contested (i.e. we 672 * need to wake up a blocked thread). 673 */ 674void 675_mtx_unlock_sleep(struct mtx *m, int opts, const char *file, int line) 676{ 677 struct thread *td, *td1; 678 struct mtx *m1; 679 int pri; 680 681 td = curthread; 682 683 if (mtx_recursed(m)) { 684 if (--(m->mtx_recurse) == 0) 685 atomic_clear_ptr(&m->mtx_lock, MTX_RECURSED); 686 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 687 CTR1(KTR_LOCK, "_mtx_unlock_sleep: %p unrecurse", m); 688 return; 689 } 690 691 mtx_lock_spin(&sched_lock); 692 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 693 CTR1(KTR_LOCK, "_mtx_unlock_sleep: %p contested", m); 694 695 td1 = TAILQ_FIRST(&m->mtx_blocked); 696#if defined(SMP) && defined(ADAPTIVE_MUTEXES) 697 if (td1 == NULL) { 698 _release_lock_quick(m); 699 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 700 CTR1(KTR_LOCK, "_mtx_unlock_sleep: %p no sleepers", m); 701 mtx_unlock_spin(&sched_lock); 702 return; 703 } 704#endif 705 MPASS(td->td_proc->p_magic == P_MAGIC); 706 MPASS(td1->td_proc->p_magic == P_MAGIC); 707 708 TAILQ_REMOVE(&m->mtx_blocked, td1, td_blkq); 709 710 if (TAILQ_EMPTY(&m->mtx_blocked)) { 711 LIST_REMOVE(m, mtx_contested); 712 _release_lock_quick(m); 713 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 714 CTR1(KTR_LOCK, "_mtx_unlock_sleep: %p not held", m); 715 } else 716 atomic_store_rel_ptr(&m->mtx_lock, (void *)MTX_CONTESTED); 717 718 pri = PRI_MAX; 719 LIST_FOREACH(m1, &td->td_contested, mtx_contested) { 720 int cp = TAILQ_FIRST(&m1->mtx_blocked)->td_priority; 721 if (cp < pri) 722 pri = cp; 723 } 724 725 if (pri > td->td_base_pri) 726 pri = td->td_base_pri; 727 td->td_priority = pri; 728 729 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 730 CTR2(KTR_LOCK, "_mtx_unlock_sleep: %p contested setrunqueue %p", 731 m, td1); 732 733 td1->td_blocked = NULL; 734 setrunqueue(td1); 735 736 if (td->td_critnest == 1 && td1->td_priority < pri) { 737#ifdef notyet 738 if (td->td_ithd != NULL) { 739 struct ithd *it = td->td_ithd; 740 741 if (it->it_interrupted) { 742 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 743 CTR2(KTR_LOCK, 744 "_mtx_unlock_sleep: %p interrupted %p", 745 it, it->it_interrupted); 746 intr_thd_fixup(it); 747 } 748 } 749#endif 750 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 751 CTR2(KTR_LOCK, 752 "_mtx_unlock_sleep: %p switching out lock=%p", m, 753 (void *)m->mtx_lock); 754 755 td->td_proc->p_stats->p_ru.ru_nivcsw++; 756 mi_switch(); 757 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 758 CTR2(KTR_LOCK, "_mtx_unlock_sleep: %p resuming lock=%p", 759 m, (void *)m->mtx_lock); 760 } 761 762 mtx_unlock_spin(&sched_lock); 763 764 return; 765} 766 767/* 768 * All the unlocking of MTX_SPIN locks is done inline. 769 * See the _rel_spin_lock() macro for the details. 770 */ 771 772/* 773 * The backing function for the INVARIANTS-enabled mtx_assert() 774 */ 775#ifdef INVARIANT_SUPPORT 776void 777_mtx_assert(struct mtx *m, int what, const char *file, int line) 778{ 779 780 if (panicstr != NULL) 781 return; 782 switch (what) { 783 case MA_OWNED: 784 case MA_OWNED | MA_RECURSED: 785 case MA_OWNED | MA_NOTRECURSED: 786 if (!mtx_owned(m)) 787 panic("mutex %s not owned at %s:%d", 788 m->mtx_object.lo_name, file, line); 789 if (mtx_recursed(m)) { 790 if ((what & MA_NOTRECURSED) != 0) 791 panic("mutex %s recursed at %s:%d", 792 m->mtx_object.lo_name, file, line); 793 } else if ((what & MA_RECURSED) != 0) { 794 panic("mutex %s unrecursed at %s:%d", 795 m->mtx_object.lo_name, file, line); 796 } 797 break; 798 case MA_NOTOWNED: 799 if (mtx_owned(m)) 800 panic("mutex %s owned at %s:%d", 801 m->mtx_object.lo_name, file, line); 802 break; 803 default: 804 panic("unknown mtx_assert at %s:%d", file, line); 805 } 806} 807#endif 808 809/* 810 * The MUTEX_DEBUG-enabled mtx_validate() 811 * 812 * Most of these checks have been moved off into the LO_INITIALIZED flag 813 * maintained by the witness code. 814 */ 815#ifdef MUTEX_DEBUG 816 817void mtx_validate(struct mtx *); 818 819void 820mtx_validate(struct mtx *m) 821{ 822 823/* 824 * XXX - When kernacc() is fixed on the alpha to handle K0_SEG memory properly 825 * we can re-enable the kernacc() checks. 826 */ 827#ifndef __alpha__ 828 /* 829 * Can't call kernacc() from early init386(), especially when 830 * initializing Giant mutex, because some stuff in kernacc() 831 * requires Giant itself. 832 */ 833 if (!cold) 834 if (!kernacc((caddr_t)m, sizeof(m), 835 VM_PROT_READ | VM_PROT_WRITE)) 836 panic("Can't read and write to mutex %p", m); 837#endif 838} 839#endif 840 841/* 842 * General init routine used by the MTX_SYSINIT() macro. 843 */ 844void 845mtx_sysinit(void *arg) 846{ 847 struct mtx_args *margs = arg; 848 849 mtx_init(margs->ma_mtx, margs->ma_desc, NULL, margs->ma_opts); 850} 851 852/* 853 * Mutex initialization routine; initialize lock `m' of type contained in 854 * `opts' with options contained in `opts' and name `name.' The optional 855 * lock type `type' is used as a general lock category name for use with 856 * witness. 857 */ 858void 859mtx_init(struct mtx *m, const char *name, const char *type, int opts) 860{ 861 struct lock_object *lock; 862 863 MPASS((opts & ~(MTX_SPIN | MTX_QUIET | MTX_RECURSE | 864 MTX_SLEEPABLE | MTX_NOWITNESS | MTX_DUPOK)) == 0); 865 866#ifdef MUTEX_DEBUG 867 /* Diagnostic and error correction */ 868 mtx_validate(m); 869#endif 870 871 lock = &m->mtx_object; 872 KASSERT((lock->lo_flags & LO_INITIALIZED) == 0, 873 ("mutex %s %p already initialized", name, m)); 874 bzero(m, sizeof(*m)); 875 if (opts & MTX_SPIN) 876 lock->lo_class = &lock_class_mtx_spin; 877 else 878 lock->lo_class = &lock_class_mtx_sleep; 879 lock->lo_name = name; 880 lock->lo_type = type != NULL ? type : name; 881 if (opts & MTX_QUIET) 882 lock->lo_flags = LO_QUIET; 883 if (opts & MTX_RECURSE) 884 lock->lo_flags |= LO_RECURSABLE; 885 if (opts & MTX_SLEEPABLE) 886 lock->lo_flags |= LO_SLEEPABLE; 887 if ((opts & MTX_NOWITNESS) == 0) 888 lock->lo_flags |= LO_WITNESS; 889 if (opts & MTX_DUPOK) 890 lock->lo_flags |= LO_DUPOK; 891 892 m->mtx_lock = MTX_UNOWNED; 893 TAILQ_INIT(&m->mtx_blocked); 894 895 LOCK_LOG_INIT(lock, opts); 896 897 WITNESS_INIT(lock); 898} 899 900/* 901 * Remove lock `m' from all_mtx queue. We don't allow MTX_QUIET to be 902 * passed in as a flag here because if the corresponding mtx_init() was 903 * called with MTX_QUIET set, then it will already be set in the mutex's 904 * flags. 905 */ 906void 907mtx_destroy(struct mtx *m) 908{ 909 910 LOCK_LOG_DESTROY(&m->mtx_object, 0); 911 912 if (!mtx_owned(m)) 913 MPASS(mtx_unowned(m)); 914 else { 915 MPASS((m->mtx_lock & (MTX_RECURSED|MTX_CONTESTED)) == 0); 916 917 /* Tell witness this isn't locked to make it happy. */ 918 WITNESS_UNLOCK(&m->mtx_object, LOP_EXCLUSIVE, __FILE__, 919 __LINE__); 920 } 921 922 WITNESS_DESTROY(&m->mtx_object); 923} 924 925/* 926 * Intialize the mutex code and system mutexes. This is called from the MD 927 * startup code prior to mi_startup(). The per-CPU data space needs to be 928 * setup before this is called. 929 */ 930void 931mutex_init(void) 932{ 933 934 /* Setup thread0 so that mutexes work. */ 935 LIST_INIT(&thread0.td_contested); 936 937 /* 938 * Initialize mutexes. 939 */ 940 mtx_init(&Giant, "Giant", NULL, MTX_DEF | MTX_RECURSE); 941 mtx_init(&sched_lock, "sched lock", NULL, MTX_SPIN | MTX_RECURSE); 942 mtx_init(&proc0.p_mtx, "process lock", NULL, MTX_DEF | MTX_DUPOK); 943 mtx_lock(&Giant); 944} 945 946/* 947 * Encapsulated Giant mutex routines. These routines provide encapsulation 948 * control for the Giant mutex, allowing sysctls to be used to turn on and 949 * off Giant around certain subsystems. The default value for the sysctls 950 * are set to what developers believe is stable and working in regards to 951 * the Giant pushdown. Developers should not turn off Giant via these 952 * sysctls unless they know what they are doing. 953 * 954 * Callers of mtx_lock_giant() are expected to pass the return value to an 955 * accompanying mtx_unlock_giant() later on. If multiple subsystems are 956 * effected by a Giant wrap, all related sysctl variables must be zero for 957 * the subsystem call to operate without Giant (as determined by the caller). 958 */ 959 960SYSCTL_NODE(_kern, OID_AUTO, giant, CTLFLAG_RD, NULL, "Giant mutex manipulation"); 961 962static int kern_giant_all = 0; 963SYSCTL_INT(_kern_giant, OID_AUTO, all, CTLFLAG_RW, &kern_giant_all, 0, ""); 964 965int kern_giant_proc = 1; /* Giant around PROC locks */ 966int kern_giant_file = 1; /* Giant around struct file & filedesc */ 967int kern_giant_ucred = 1; /* Giant around ucred */ 968SYSCTL_INT(_kern_giant, OID_AUTO, proc, CTLFLAG_RW, &kern_giant_proc, 0, ""); 969SYSCTL_INT(_kern_giant, OID_AUTO, file, CTLFLAG_RW, &kern_giant_file, 0, ""); 970SYSCTL_INT(_kern_giant, OID_AUTO, ucred, CTLFLAG_RW, &kern_giant_ucred, 0, ""); 971 972int 973mtx_lock_giant(int sysctlvar) 974{ 975 if (sysctlvar || kern_giant_all) { 976 mtx_lock(&Giant); 977 return(1); 978 } 979 return(0); 980} 981 982void 983mtx_unlock_giant(int s) 984{ 985 if (s) 986 mtx_unlock(&Giant); 987} 988 989