subr_turnstile.c revision 93609
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/subr_turnstile.c 93609 2002-04-02 00:01:49Z des $ 31 */ 32 33/* 34 * Machine independent bits of mutex implementation. 35 */ 36 37#include "opt_ddb.h" 38 39#include <sys/param.h> 40#include <sys/systm.h> 41#include <sys/bus.h> 42#include <sys/kernel.h> 43#include <sys/ktr.h> 44#include <sys/lock.h> 45#include <sys/malloc.h> 46#include <sys/mutex.h> 47#include <sys/proc.h> 48#include <sys/resourcevar.h> 49#include <sys/sbuf.h> 50#include <sys/sysctl.h> 51#include <sys/vmmeter.h> 52 53#include <machine/atomic.h> 54#include <machine/bus.h> 55#include <machine/clock.h> 56#include <machine/cpu.h> 57 58#include <ddb/ddb.h> 59 60#include <vm/vm.h> 61#include <vm/vm_extern.h> 62 63/* 64 * Internal utility macros. 65 */ 66#define mtx_unowned(m) ((m)->mtx_lock == MTX_UNOWNED) 67 68#define mtx_owner(m) (mtx_unowned((m)) ? NULL \ 69 : (struct thread *)((m)->mtx_lock & MTX_FLAGMASK)) 70 71/* 72 * Lock classes for sleep and spin mutexes. 73 */ 74struct lock_class lock_class_mtx_sleep = { 75 "sleep mutex", 76 LC_SLEEPLOCK | LC_RECURSABLE 77}; 78struct lock_class lock_class_mtx_spin = { 79 "spin mutex", 80 LC_SPINLOCK | LC_RECURSABLE 81}; 82 83/* 84 * Prototypes for non-exported routines. 85 */ 86static void propagate_priority(struct thread *); 87 88static void 89propagate_priority(struct thread *td) 90{ 91 int pri = td->td_priority; 92 struct mtx *m = td->td_blocked; 93 94 mtx_assert(&sched_lock, MA_OWNED); 95 for (;;) { 96 struct thread *td1; 97 98 td = mtx_owner(m); 99 100 if (td == NULL) { 101 /* 102 * This really isn't quite right. Really 103 * ought to bump priority of thread that 104 * next acquires the mutex. 105 */ 106 MPASS(m->mtx_lock == MTX_CONTESTED); 107 return; 108 } 109 110 MPASS(td->td_proc->p_magic == P_MAGIC); 111 KASSERT(td->td_proc->p_stat != SSLEEP, ("sleeping thread owns a mutex")); 112 if (td->td_priority <= pri) /* lower is higher priority */ 113 return; 114 115 /* 116 * Bump this thread's priority. 117 */ 118 td->td_priority = pri; 119 120 /* 121 * If lock holder is actually running, just bump priority. 122 */ 123 /* XXXKSE this test is not sufficient */ 124 if (td->td_kse && (td->td_kse->ke_oncpu != NOCPU)) { 125 MPASS(td->td_proc->p_stat == SRUN 126 || td->td_proc->p_stat == SZOMB 127 || td->td_proc->p_stat == SSTOP); 128 return; 129 } 130 131#ifndef SMP 132 /* 133 * For UP, we check to see if td is curthread (this shouldn't 134 * ever happen however as it would mean we are in a deadlock.) 135 */ 136 KASSERT(td != curthread, ("Deadlock detected")); 137#endif 138 139 /* 140 * If on run queue move to new run queue, and quit. 141 * XXXKSE this gets a lot more complicated under threads 142 * but try anyhow. 143 */ 144 if (td->td_proc->p_stat == SRUN) { 145 MPASS(td->td_blocked == NULL); 146 remrunqueue(td); 147 setrunqueue(td); 148 return; 149 } 150 151 /* 152 * If we aren't blocked on a mutex, we should be. 153 */ 154 KASSERT(td->td_proc->p_stat == SMTX, ( 155 "process %d(%s):%d holds %s but isn't blocked on a mutex\n", 156 td->td_proc->p_pid, td->td_proc->p_comm, td->td_proc->p_stat, 157 m->mtx_object.lo_name)); 158 159 /* 160 * Pick up the mutex that td is blocked on. 161 */ 162 m = td->td_blocked; 163 MPASS(m != NULL); 164 165 /* 166 * Check if the thread needs to be moved up on 167 * the blocked chain 168 */ 169 if (td == TAILQ_FIRST(&m->mtx_blocked)) { 170 continue; 171 } 172 173 td1 = TAILQ_PREV(td, threadqueue, td_blkq); 174 if (td1->td_priority <= pri) { 175 continue; 176 } 177 178 /* 179 * Remove thread from blocked chain and determine where 180 * it should be moved up to. Since we know that td1 has 181 * a lower priority than td, we know that at least one 182 * thread in the chain has a lower priority and that 183 * td1 will thus not be NULL after the loop. 184 */ 185 TAILQ_REMOVE(&m->mtx_blocked, td, td_blkq); 186 TAILQ_FOREACH(td1, &m->mtx_blocked, td_blkq) { 187 MPASS(td1->td_proc->p_magic == P_MAGIC); 188 if (td1->td_priority > pri) 189 break; 190 } 191 192 MPASS(td1 != NULL); 193 TAILQ_INSERT_BEFORE(td1, td, td_blkq); 194 CTR4(KTR_LOCK, 195 "propagate_priority: p %p moved before %p on [%p] %s", 196 td, td1, m, m->mtx_object.lo_name); 197 } 198} 199 200#ifdef MUTEX_PROFILING 201SYSCTL_NODE(_debug, OID_AUTO, mutex, CTLFLAG_RD, NULL, "mutex debugging"); 202SYSCTL_NODE(_debug_mutex, OID_AUTO, prof, CTLFLAG_RD, NULL, "mutex profiling"); 203static int mutex_prof_enable = 0; 204SYSCTL_INT(_debug_mutex_prof, OID_AUTO, enable, CTLFLAG_RW, 205 &mutex_prof_enable, 0, "Enable tracing of mutex holdtime"); 206 207struct mutex_prof { 208 const char *name; 209 const char *file; 210 int line; 211#define MPROF_MAX 0 212#define MPROF_TOT 1 213#define MPROF_CNT 2 214#define MPROF_AVG 3 215 u_int64_t cycles[4]; 216}; 217 218/* 219 * mprof_buf is a static pool of profiling records to avoid possible 220 * reentrance of the memory allocation functions. 221 * 222 * Note: NUM_MPROF_BUFFERS must be smaller than MPROF_HASH_SIZE. 223 */ 224#define NUM_MPROF_BUFFERS 4096 225static struct mutex_prof mprof_buf[NUM_MPROF_BUFFERS]; 226static int first_free_mprof_buf; 227#define MPROF_HASH_SIZE 32771 228static struct mutex_prof *mprof_hash[MPROF_HASH_SIZE]; 229 230static int mutex_prof_acquisitions; 231SYSCTL_INT(_debug_mutex_prof, OID_AUTO, acquisitions, CTLFLAG_RD, 232 &mutex_prof_acquisitions, 0, "Number of mutex acquistions recorded"); 233static int mutex_prof_records; 234SYSCTL_INT(_debug_mutex_prof, OID_AUTO, records, CTLFLAG_RD, 235 &mutex_prof_records, 0, "Number of profiling records"); 236static int mutex_prof_maxrecords = NUM_MPROF_BUFFERS; 237SYSCTL_INT(_debug_mutex_prof, OID_AUTO, maxrecords, CTLFLAG_RD, 238 &mutex_prof_maxrecords, 0, "Maximum number of profiling records"); 239static int mutex_prof_rejected; 240SYSCTL_INT(_debug_mutex_prof, OID_AUTO, rejected, CTLFLAG_RD, 241 &mutex_prof_rejected, 0, "Number of rejected profiling records"); 242static int mutex_prof_hashsize = MPROF_HASH_SIZE; 243SYSCTL_INT(_debug_mutex_prof, OID_AUTO, hashsize, CTLFLAG_RD, 244 &mutex_prof_hashsize, 0, "Hash size"); 245static int mutex_prof_collisions = 0; 246SYSCTL_INT(_debug_mutex_prof, OID_AUTO, collisions, CTLFLAG_RD, 247 &mutex_prof_collisions, 0, "Number of hash collisions"); 248 249/* 250 * mprof_mtx protects the profiling buffers and the hash. 251 */ 252static struct mtx mprof_mtx; 253 254static void 255mprof_init(void *arg __unused) 256{ 257 mtx_init(&mprof_mtx, "mutex profiling lock", MTX_SPIN | MTX_QUIET); 258} 259SYSINIT(mprofinit, SI_SUB_LOCK, SI_ORDER_ANY, mprof_init, NULL); 260 261static int 262dump_mutex_prof_stats(SYSCTL_HANDLER_ARGS) 263{ 264 struct sbuf *sb; 265 int error, i; 266 267 if (first_free_mprof_buf == 0) 268 return SYSCTL_OUT(req, "No locking recorded", 269 sizeof("No locking recorded")); 270 271 sb = sbuf_new(NULL, NULL, 1024, SBUF_AUTOEXTEND); 272 sbuf_printf(sb, "%12s %12s %12s %12s %s\n", 273 "max", "total", "count", "average", "name"); 274 mtx_lock_spin(&mprof_mtx); 275 for (i = 0; i < first_free_mprof_buf; ++i) 276 sbuf_printf(sb, "%12llu %12llu %12llu %12llu %s:%d (%s)\n", 277 mprof_buf[i].cycles[MPROF_MAX], 278 mprof_buf[i].cycles[MPROF_TOT], 279 mprof_buf[i].cycles[MPROF_CNT], 280 mprof_buf[i].cycles[MPROF_AVG], 281 mprof_buf[i].file, mprof_buf[i].line, mprof_buf[i].name); 282 mtx_unlock_spin(&mprof_mtx); 283 sbuf_finish(sb); 284 error = SYSCTL_OUT(req, sbuf_data(sb), sbuf_len(sb) + 1); 285 sbuf_delete(sb); 286 return (error); 287} 288SYSCTL_PROC(_debug_mutex_prof, OID_AUTO, stats, CTLTYPE_STRING|CTLFLAG_RD, 289 NULL, 0, dump_mutex_prof_stats, "A", "Mutex profiling statistics"); 290#endif 291 292/* 293 * Function versions of the inlined __mtx_* macros. These are used by 294 * modules and can also be called from assembly language if needed. 295 */ 296void 297_mtx_lock_flags(struct mtx *m, int opts, const char *file, int line) 298{ 299 300 MPASS(curthread != NULL); 301 _get_sleep_lock(m, curthread, opts, file, line); 302 LOCK_LOG_LOCK("LOCK", &m->mtx_object, opts, m->mtx_recurse, file, 303 line); 304 WITNESS_LOCK(&m->mtx_object, opts | LOP_EXCLUSIVE, file, line); 305#ifdef MUTEX_PROFILING 306 /* don't reset the timer when/if recursing */ 307 if (m->cycles == 0) { 308 m->file = file; 309 m->line = line; 310 m->cycles = mutex_prof_enable ? get_cyclecount() : 0; 311 ++mutex_prof_acquisitions; 312 } 313#endif 314} 315 316void 317_mtx_unlock_flags(struct mtx *m, int opts, const char *file, int line) 318{ 319 320 MPASS(curthread != NULL); 321 mtx_assert(m, MA_OWNED); 322#ifdef MUTEX_PROFILING 323 if (m->cycles != 0) { 324 static const char *unknown = "(unknown)"; 325 struct mutex_prof *mpp; 326 u_int64_t cycles, mcycles; 327 const char *p, *q; 328 volatile u_int hash, n; 329 330 cycles = get_cyclecount(); 331 mcycles = m->cycles; 332 m->cycles = 0; 333 if (cycles <= mcycles) 334 goto out; 335 for (p = file; strncmp(p, "../", 3) == 0; p += 3) 336 /* nothing */ ; 337 if (p == NULL || *p == '\0') 338 p = unknown; 339 for (hash = line, q = p; *q != '\0'; ++q) 340 hash = (hash * 2 + *q) % MPROF_HASH_SIZE; 341 mtx_lock_spin(&mprof_mtx); 342 n = hash; 343 while ((mpp = mprof_hash[n]) != NULL) { 344 if (mpp->line == line && strcmp(mpp->file, p) == 0) 345 break; 346 n = (n + 1) % MPROF_HASH_SIZE; 347 } 348 if (mpp == NULL) { 349 /* Just exit if we cannot get a trace buffer */ 350 if (first_free_mprof_buf >= NUM_MPROF_BUFFERS) { 351 ++mutex_prof_rejected; 352 goto unlock; 353 } 354 mpp = &mprof_buf[first_free_mprof_buf++]; 355 mpp->name = mtx_name(m); 356 mpp->file = p; 357 mpp->line = line; 358 mutex_prof_collisions += n - hash; 359 ++mutex_prof_records; 360 mprof_hash[hash] = mpp; 361 } 362 /* 363 * Record if the mutex has been held longer now than ever 364 * before 365 */ 366 if ((cycles - mcycles) > mpp->cycles[MPROF_MAX]) 367 mpp->cycles[MPROF_MAX] = cycles - mcycles; 368 mpp->cycles[MPROF_TOT] += cycles - mcycles; 369 mpp->cycles[MPROF_CNT] += 1; 370 mpp->cycles[MPROF_AVG] = 371 mpp->cycles[MPROF_TOT] / mpp->cycles[MPROF_CNT]; 372unlock: 373 mtx_unlock_spin(&mprof_mtx); 374 } 375out: 376#endif 377 WITNESS_UNLOCK(&m->mtx_object, opts | LOP_EXCLUSIVE, file, line); 378 LOCK_LOG_LOCK("UNLOCK", &m->mtx_object, opts, m->mtx_recurse, file, 379 line); 380 _rel_sleep_lock(m, curthread, opts, file, line); 381} 382 383void 384_mtx_lock_spin_flags(struct mtx *m, int opts, const char *file, int line) 385{ 386 387 MPASS(curthread != NULL); 388 _get_spin_lock(m, curthread, opts, file, line); 389 LOCK_LOG_LOCK("LOCK", &m->mtx_object, opts, m->mtx_recurse, file, 390 line); 391 WITNESS_LOCK(&m->mtx_object, opts | LOP_EXCLUSIVE, file, line); 392} 393 394void 395_mtx_unlock_spin_flags(struct mtx *m, int opts, const char *file, int line) 396{ 397 398 MPASS(curthread != NULL); 399 mtx_assert(m, MA_OWNED); 400 WITNESS_UNLOCK(&m->mtx_object, opts | LOP_EXCLUSIVE, file, line); 401 LOCK_LOG_LOCK("UNLOCK", &m->mtx_object, opts, m->mtx_recurse, file, 402 line); 403 _rel_spin_lock(m); 404} 405 406/* 407 * The important part of mtx_trylock{,_flags}() 408 * Tries to acquire lock `m.' We do NOT handle recursion here; we assume that 409 * if we're called, it's because we know we don't already own this lock. 410 */ 411int 412_mtx_trylock(struct mtx *m, int opts, const char *file, int line) 413{ 414 int rval; 415 416 MPASS(curthread != NULL); 417 418 rval = _obtain_lock(m, curthread); 419 420 LOCK_LOG_TRY("LOCK", &m->mtx_object, opts, rval, file, line); 421 if (rval) { 422 /* 423 * We do not handle recursion in _mtx_trylock; see the 424 * note at the top of the routine. 425 */ 426 KASSERT(!mtx_recursed(m), 427 ("mtx_trylock() called on a recursed mutex")); 428 WITNESS_LOCK(&m->mtx_object, opts | LOP_EXCLUSIVE | LOP_TRYLOCK, 429 file, line); 430 } 431 432 return (rval); 433} 434 435/* 436 * _mtx_lock_sleep: the tougher part of acquiring an MTX_DEF lock. 437 * 438 * We call this if the lock is either contested (i.e. we need to go to 439 * sleep waiting for it), or if we need to recurse on it. 440 */ 441void 442_mtx_lock_sleep(struct mtx *m, int opts, const char *file, int line) 443{ 444 struct thread *td = curthread; 445 446 if ((m->mtx_lock & MTX_FLAGMASK) == (uintptr_t)td) { 447 m->mtx_recurse++; 448 atomic_set_ptr(&m->mtx_lock, MTX_RECURSED); 449 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 450 CTR1(KTR_LOCK, "_mtx_lock_sleep: %p recursing", m); 451 return; 452 } 453 454 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 455 CTR4(KTR_LOCK, 456 "_mtx_lock_sleep: %s contested (lock=%p) at %s:%d", 457 m->mtx_object.lo_name, (void *)m->mtx_lock, file, line); 458 459 while (!_obtain_lock(m, td)) { 460 uintptr_t v; 461 struct thread *td1; 462 463 mtx_lock_spin(&sched_lock); 464 /* 465 * Check if the lock has been released while spinning for 466 * the sched_lock. 467 */ 468 if ((v = m->mtx_lock) == MTX_UNOWNED) { 469 mtx_unlock_spin(&sched_lock); 470 continue; 471 } 472 473 /* 474 * The mutex was marked contested on release. This means that 475 * there are threads blocked on it. 476 */ 477 if (v == MTX_CONTESTED) { 478 td1 = TAILQ_FIRST(&m->mtx_blocked); 479 MPASS(td1 != NULL); 480 m->mtx_lock = (uintptr_t)td | MTX_CONTESTED; 481 482 if (td1->td_priority < td->td_priority) 483 td->td_priority = td1->td_priority; 484 mtx_unlock_spin(&sched_lock); 485 return; 486 } 487 488 /* 489 * If the mutex isn't already contested and a failure occurs 490 * setting the contested bit, the mutex was either released 491 * or the state of the MTX_RECURSED bit changed. 492 */ 493 if ((v & MTX_CONTESTED) == 0 && 494 !atomic_cmpset_ptr(&m->mtx_lock, (void *)v, 495 (void *)(v | MTX_CONTESTED))) { 496 mtx_unlock_spin(&sched_lock); 497 continue; 498 } 499 500 /* 501 * We deffinately must sleep for this lock. 502 */ 503 mtx_assert(m, MA_NOTOWNED); 504 505#ifdef notyet 506 /* 507 * If we're borrowing an interrupted thread's VM context, we 508 * must clean up before going to sleep. 509 */ 510 if (td->td_ithd != NULL) { 511 struct ithd *it = td->td_ithd; 512 513 if (it->it_interrupted) { 514 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 515 CTR2(KTR_LOCK, 516 "_mtx_lock_sleep: %p interrupted %p", 517 it, it->it_interrupted); 518 intr_thd_fixup(it); 519 } 520 } 521#endif 522 523 /* 524 * Put us on the list of threads blocked on this mutex. 525 */ 526 if (TAILQ_EMPTY(&m->mtx_blocked)) { 527 td1 = mtx_owner(m); 528 LIST_INSERT_HEAD(&td1->td_contested, m, mtx_contested); 529 TAILQ_INSERT_TAIL(&m->mtx_blocked, td, td_blkq); 530 } else { 531 TAILQ_FOREACH(td1, &m->mtx_blocked, td_blkq) 532 if (td1->td_priority > td->td_priority) 533 break; 534 if (td1) 535 TAILQ_INSERT_BEFORE(td1, td, td_blkq); 536 else 537 TAILQ_INSERT_TAIL(&m->mtx_blocked, td, td_blkq); 538 } 539 540 /* 541 * Save who we're blocked on. 542 */ 543 td->td_blocked = m; 544 td->td_mtxname = m->mtx_object.lo_name; 545 td->td_proc->p_stat = SMTX; 546 propagate_priority(td); 547 548 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 549 CTR3(KTR_LOCK, 550 "_mtx_lock_sleep: p %p blocked on [%p] %s", td, m, 551 m->mtx_object.lo_name); 552 553 td->td_proc->p_stats->p_ru.ru_nvcsw++; 554 mi_switch(); 555 556 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 557 CTR3(KTR_LOCK, 558 "_mtx_lock_sleep: p %p free from blocked on [%p] %s", 559 td, m, m->mtx_object.lo_name); 560 561 mtx_unlock_spin(&sched_lock); 562 } 563 564 return; 565} 566 567/* 568 * _mtx_lock_spin: the tougher part of acquiring an MTX_SPIN lock. 569 * 570 * This is only called if we need to actually spin for the lock. Recursion 571 * is handled inline. 572 */ 573void 574_mtx_lock_spin(struct mtx *m, int opts, const char *file, int line) 575{ 576 int i = 0; 577 578 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 579 CTR1(KTR_LOCK, "_mtx_lock_spin: %p spinning", m); 580 581 for (;;) { 582 if (_obtain_lock(m, curthread)) 583 break; 584 585 /* Give interrupts a chance while we spin. */ 586 critical_exit(); 587 while (m->mtx_lock != MTX_UNOWNED) { 588 if (i++ < 10000000) 589 continue; 590 if (i++ < 60000000) 591 DELAY(1); 592#ifdef DDB 593 else if (!db_active) 594#else 595 else 596#endif 597 panic("spin lock %s held by %p for > 5 seconds", 598 m->mtx_object.lo_name, (void *)m->mtx_lock); 599 } 600 critical_enter(); 601 } 602 603 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 604 CTR1(KTR_LOCK, "_mtx_lock_spin: %p spin done", m); 605 606 return; 607} 608 609/* 610 * _mtx_unlock_sleep: the tougher part of releasing an MTX_DEF lock. 611 * 612 * We are only called here if the lock is recursed or contested (i.e. we 613 * need to wake up a blocked thread). 614 */ 615void 616_mtx_unlock_sleep(struct mtx *m, int opts, const char *file, int line) 617{ 618 struct thread *td, *td1; 619 struct mtx *m1; 620 int pri; 621 622 td = curthread; 623 624 if (mtx_recursed(m)) { 625 if (--(m->mtx_recurse) == 0) 626 atomic_clear_ptr(&m->mtx_lock, MTX_RECURSED); 627 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 628 CTR1(KTR_LOCK, "_mtx_unlock_sleep: %p unrecurse", m); 629 return; 630 } 631 632 mtx_lock_spin(&sched_lock); 633 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 634 CTR1(KTR_LOCK, "_mtx_unlock_sleep: %p contested", m); 635 636 td1 = TAILQ_FIRST(&m->mtx_blocked); 637 MPASS(td->td_proc->p_magic == P_MAGIC); 638 MPASS(td1->td_proc->p_magic == P_MAGIC); 639 640 TAILQ_REMOVE(&m->mtx_blocked, td1, td_blkq); 641 642 if (TAILQ_EMPTY(&m->mtx_blocked)) { 643 LIST_REMOVE(m, mtx_contested); 644 _release_lock_quick(m); 645 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 646 CTR1(KTR_LOCK, "_mtx_unlock_sleep: %p not held", m); 647 } else 648 atomic_store_rel_ptr(&m->mtx_lock, (void *)MTX_CONTESTED); 649 650 pri = PRI_MAX; 651 LIST_FOREACH(m1, &td->td_contested, mtx_contested) { 652 int cp = TAILQ_FIRST(&m1->mtx_blocked)->td_priority; 653 if (cp < pri) 654 pri = cp; 655 } 656 657 if (pri > td->td_base_pri) 658 pri = td->td_base_pri; 659 td->td_priority = pri; 660 661 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 662 CTR2(KTR_LOCK, "_mtx_unlock_sleep: %p contested setrunqueue %p", 663 m, td1); 664 665 td1->td_blocked = NULL; 666 td1->td_proc->p_stat = SRUN; 667 setrunqueue(td1); 668 669 if (td->td_critnest == 1 && td1->td_priority < pri) { 670#ifdef notyet 671 if (td->td_ithd != NULL) { 672 struct ithd *it = td->td_ithd; 673 674 if (it->it_interrupted) { 675 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 676 CTR2(KTR_LOCK, 677 "_mtx_unlock_sleep: %p interrupted %p", 678 it, it->it_interrupted); 679 intr_thd_fixup(it); 680 } 681 } 682#endif 683 setrunqueue(td); 684 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 685 CTR2(KTR_LOCK, 686 "_mtx_unlock_sleep: %p switching out lock=%p", m, 687 (void *)m->mtx_lock); 688 689 td->td_proc->p_stats->p_ru.ru_nivcsw++; 690 mi_switch(); 691 if (LOCK_LOG_TEST(&m->mtx_object, opts)) 692 CTR2(KTR_LOCK, "_mtx_unlock_sleep: %p resuming lock=%p", 693 m, (void *)m->mtx_lock); 694 } 695 696 mtx_unlock_spin(&sched_lock); 697 698 return; 699} 700 701/* 702 * All the unlocking of MTX_SPIN locks is done inline. 703 * See the _rel_spin_lock() macro for the details. 704 */ 705 706/* 707 * The backing function for the INVARIANTS-enabled mtx_assert() 708 */ 709#ifdef INVARIANT_SUPPORT 710void 711_mtx_assert(struct mtx *m, int what, const char *file, int line) 712{ 713 714 if (panicstr != NULL) 715 return; 716 switch (what) { 717 case MA_OWNED: 718 case MA_OWNED | MA_RECURSED: 719 case MA_OWNED | MA_NOTRECURSED: 720 if (!mtx_owned(m)) 721 panic("mutex %s not owned at %s:%d", 722 m->mtx_object.lo_name, file, line); 723 if (mtx_recursed(m)) { 724 if ((what & MA_NOTRECURSED) != 0) 725 panic("mutex %s recursed at %s:%d", 726 m->mtx_object.lo_name, file, line); 727 } else if ((what & MA_RECURSED) != 0) { 728 panic("mutex %s unrecursed at %s:%d", 729 m->mtx_object.lo_name, file, line); 730 } 731 break; 732 case MA_NOTOWNED: 733 if (mtx_owned(m)) 734 panic("mutex %s owned at %s:%d", 735 m->mtx_object.lo_name, file, line); 736 break; 737 default: 738 panic("unknown mtx_assert at %s:%d", file, line); 739 } 740} 741#endif 742 743/* 744 * The MUTEX_DEBUG-enabled mtx_validate() 745 * 746 * Most of these checks have been moved off into the LO_INITIALIZED flag 747 * maintained by the witness code. 748 */ 749#ifdef MUTEX_DEBUG 750 751void mtx_validate(struct mtx *); 752 753void 754mtx_validate(struct mtx *m) 755{ 756 757/* 758 * XXX - When kernacc() is fixed on the alpha to handle K0_SEG memory properly 759 * we can re-enable the kernacc() checks. 760 */ 761#ifndef __alpha__ 762 /* 763 * Can't call kernacc() from early init386(), especially when 764 * initializing Giant mutex, because some stuff in kernacc() 765 * requires Giant itself. 766 */ 767 if (!cold) 768 if (!kernacc((caddr_t)m, sizeof(m), 769 VM_PROT_READ | VM_PROT_WRITE)) 770 panic("Can't read and write to mutex %p", m); 771#endif 772} 773#endif 774 775/* 776 * Mutex initialization routine; initialize lock `m' of type contained in 777 * `opts' with options contained in `opts' and description `description.' 778 */ 779void 780mtx_init(struct mtx *m, const char *description, int opts) 781{ 782 struct lock_object *lock; 783 784 MPASS((opts & ~(MTX_SPIN | MTX_QUIET | MTX_RECURSE | 785 MTX_SLEEPABLE | MTX_NOWITNESS | MTX_DUPOK)) == 0); 786 787#ifdef MUTEX_DEBUG 788 /* Diagnostic and error correction */ 789 mtx_validate(m); 790#endif 791 792 lock = &m->mtx_object; 793 KASSERT((lock->lo_flags & LO_INITIALIZED) == 0, 794 ("mutex %s %p already initialized", description, m)); 795 bzero(m, sizeof(*m)); 796 if (opts & MTX_SPIN) 797 lock->lo_class = &lock_class_mtx_spin; 798 else 799 lock->lo_class = &lock_class_mtx_sleep; 800 lock->lo_name = description; 801 if (opts & MTX_QUIET) 802 lock->lo_flags = LO_QUIET; 803 if (opts & MTX_RECURSE) 804 lock->lo_flags |= LO_RECURSABLE; 805 if (opts & MTX_SLEEPABLE) 806 lock->lo_flags |= LO_SLEEPABLE; 807 if ((opts & MTX_NOWITNESS) == 0) 808 lock->lo_flags |= LO_WITNESS; 809 if (opts & MTX_DUPOK) 810 lock->lo_flags |= LO_DUPOK; 811 812 m->mtx_lock = MTX_UNOWNED; 813 TAILQ_INIT(&m->mtx_blocked); 814 815 LOCK_LOG_INIT(lock, opts); 816 817 WITNESS_INIT(lock); 818} 819 820/* 821 * Remove lock `m' from all_mtx queue. We don't allow MTX_QUIET to be 822 * passed in as a flag here because if the corresponding mtx_init() was 823 * called with MTX_QUIET set, then it will already be set in the mutex's 824 * flags. 825 */ 826void 827mtx_destroy(struct mtx *m) 828{ 829 830 LOCK_LOG_DESTROY(&m->mtx_object, 0); 831 832 if (!mtx_owned(m)) 833 MPASS(mtx_unowned(m)); 834 else { 835 MPASS((m->mtx_lock & (MTX_RECURSED|MTX_CONTESTED)) == 0); 836 837 /* Tell witness this isn't locked to make it happy. */ 838 WITNESS_UNLOCK(&m->mtx_object, LOP_EXCLUSIVE, __FILE__, 839 __LINE__); 840 } 841 842 WITNESS_DESTROY(&m->mtx_object); 843} 844 845/* 846 * Encapsulated Giant mutex routines. These routines provide encapsulation 847 * control for the Giant mutex, allowing sysctls to be used to turn on and 848 * off Giant around certain subsystems. The default value for the sysctls 849 * are set to what developers believe is stable and working in regards to 850 * the Giant pushdown. Developers should not turn off Giant via these 851 * sysctls unless they know what they are doing. 852 * 853 * Callers of mtx_lock_giant() are expected to pass the return value to an 854 * accompanying mtx_unlock_giant() later on. If multiple subsystems are 855 * effected by a Giant wrap, all related sysctl variables must be zero for 856 * the subsystem call to operate without Giant (as determined by the caller). 857 */ 858 859SYSCTL_NODE(_kern, OID_AUTO, giant, CTLFLAG_RD, NULL, "Giant mutex manipulation"); 860 861static int kern_giant_all = 0; 862SYSCTL_INT(_kern_giant, OID_AUTO, all, CTLFLAG_RW, &kern_giant_all, 0, ""); 863 864int kern_giant_proc = 1; /* Giant around PROC locks */ 865int kern_giant_file = 1; /* Giant around struct file & filedesc */ 866int kern_giant_ucred = 1; /* Giant around ucred */ 867SYSCTL_INT(_kern_giant, OID_AUTO, proc, CTLFLAG_RW, &kern_giant_proc, 0, ""); 868SYSCTL_INT(_kern_giant, OID_AUTO, file, CTLFLAG_RW, &kern_giant_file, 0, ""); 869SYSCTL_INT(_kern_giant, OID_AUTO, ucred, CTLFLAG_RW, &kern_giant_ucred, 0, ""); 870 871int 872mtx_lock_giant(int sysctlvar) 873{ 874 if (sysctlvar || kern_giant_all) { 875 mtx_lock(&Giant); 876 return(1); 877 } 878 return(0); 879} 880 881void 882mtx_unlock_giant(int s) 883{ 884 if (s) 885 mtx_unlock(&Giant); 886} 887 888