[BACK]Return to kern_mutex.c CVS log [TXT][DIR] Up to [cvs.NetBSD.org] / src / sys / kern

Annotation of src/sys/kern/kern_mutex.c, Revision 1.21

1.21    ! pooka       1: /*     $NetBSD: kern_mutex.c,v 1.20 2007/10/19 12:16:42 ad Exp $       */
1.2       ad          2:
                      3: /*-
                      4:  * Copyright (c) 2002, 2006, 2007 The NetBSD Foundation, Inc.
                      5:  * All rights reserved.
                      6:  *
                      7:  * This code is derived from software contributed to The NetBSD Foundation
                      8:  * by Jason R. Thorpe and Andrew Doran.
                      9:  *
                     10:  * Redistribution and use in source and binary forms, with or without
                     11:  * modification, are permitted provided that the following conditions
                     12:  * are met:
                     13:  * 1. Redistributions of source code must retain the above copyright
                     14:  *    notice, this list of conditions and the following disclaimer.
                     15:  * 2. Redistributions in binary form must reproduce the above copyright
                     16:  *    notice, this list of conditions and the following disclaimer in the
                     17:  *    documentation and/or other materials provided with the distribution.
                     18:  * 3. All advertising materials mentioning features or use of this software
                     19:  *    must display the following acknowledgement:
                     20:  *     This product includes software developed by the NetBSD
                     21:  *     Foundation, Inc. and its contributors.
                     22:  * 4. Neither the name of The NetBSD Foundation nor the names of its
                     23:  *    contributors may be used to endorse or promote products derived
                     24:  *    from this software without specific prior written permission.
                     25:  *
                     26:  * THIS SOFTWARE IS PROVIDED BY THE NETBSD FOUNDATION, INC. AND CONTRIBUTORS
                     27:  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED
                     28:  * TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
                     29:  * PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE FOUNDATION OR CONTRIBUTORS
                     30:  * BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
                     31:  * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
                     32:  * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
                     33:  * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
                     34:  * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
                     35:  * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
                     36:  * POSSIBILITY OF SUCH DAMAGE.
                     37:  */
                     38:
                     39: /*
                     40:  * Kernel mutex implementation, modeled after those found in Solaris,
                     41:  * a description of which can be found in:
                     42:  *
                     43:  *     Solaris Internals: Core Kernel Architecture, Jim Mauro and
                     44:  *         Richard McDougall.
                     45:  */
                     46:
                     47: #define        __MUTEX_PRIVATE
                     48:
                     49: #include <sys/cdefs.h>
1.21    ! pooka      50: __KERNEL_RCSID(0, "$NetBSD: kern_mutex.c,v 1.20 2007/10/19 12:16:42 ad Exp $");
1.18      dsl        51:
                     52: #include "opt_multiprocessor.h"
1.2       ad         53:
                     54: #include <sys/param.h>
                     55: #include <sys/proc.h>
                     56: #include <sys/mutex.h>
                     57: #include <sys/sched.h>
                     58: #include <sys/sleepq.h>
                     59: #include <sys/systm.h>
                     60: #include <sys/lockdebug.h>
                     61: #include <sys/kernel.h>
                     62:
                     63: #include <dev/lockstat.h>
                     64:
1.20      ad         65: #include <sys/intr.h>
1.2       ad         66:
                     67: /*
                     68:  * When not running a debug kernel, spin mutexes are not much
                     69:  * more than an splraiseipl() and splx() pair.
                     70:  */
                     71:
                     72: #if defined(DIAGNOSTIC) || defined(MULTIPROCESSOR) || defined(LOCKDEBUG)
                     73: #define        FULL
                     74: #endif
                     75:
                     76: /*
                     77:  * Debugging support.
                     78:  */
                     79:
                     80: #define        MUTEX_WANTLOCK(mtx)                                     \
                     81:     LOCKDEBUG_WANTLOCK(MUTEX_GETID(mtx),                       \
                     82:         (uintptr_t)__builtin_return_address(0), 0)
                     83: #define        MUTEX_LOCKED(mtx)                                       \
                     84:     LOCKDEBUG_LOCKED(MUTEX_GETID(mtx),                         \
                     85:         (uintptr_t)__builtin_return_address(0), 0)
                     86: #define        MUTEX_UNLOCKED(mtx)                                     \
                     87:     LOCKDEBUG_UNLOCKED(MUTEX_GETID(mtx),                       \
                     88:         (uintptr_t)__builtin_return_address(0), 0)
                     89: #define        MUTEX_ABORT(mtx, msg)                                   \
1.17      ad         90:     mutex_abort(mtx, __func__, msg)
1.2       ad         91:
                     92: #if defined(LOCKDEBUG)
                     93:
                     94: #define        MUTEX_DASSERT(mtx, cond)                                \
                     95: do {                                                           \
                     96:        if (!(cond))                                            \
                     97:                MUTEX_ABORT(mtx, "assertion failed: " #cond);   \
                     98: } while (/* CONSTCOND */ 0);
                     99:
                    100: #else  /* LOCKDEBUG */
                    101:
                    102: #define        MUTEX_DASSERT(mtx, cond)        /* nothing */
                    103:
                    104: #endif /* LOCKDEBUG */
                    105:
                    106: #if defined(DIAGNOSTIC)
                    107:
                    108: #define        MUTEX_ASSERT(mtx, cond)                                 \
                    109: do {                                                           \
                    110:        if (!(cond))                                            \
                    111:                MUTEX_ABORT(mtx, "assertion failed: " #cond);   \
                    112: } while (/* CONSTCOND */ 0)
                    113:
                    114: #else  /* DIAGNOSTIC */
                    115:
                    116: #define        MUTEX_ASSERT(mtx, cond) /* nothing */
                    117:
                    118: #endif /* DIAGNOSTIC */
                    119:
                    120: /*
                    121:  * Spin mutex SPL save / restore.
                    122:  */
1.12      matt      123: #ifndef MUTEX_COUNT_BIAS
                    124: #define        MUTEX_COUNT_BIAS        0
                    125: #endif
1.2       ad        126:
                    127: #define        MUTEX_SPIN_SPLRAISE(mtx)                                        \
                    128: do {                                                                   \
                    129:        struct cpu_info *x__ci = curcpu();                              \
                    130:        int x__cnt, s;                                                  \
                    131:        x__cnt = x__ci->ci_mtx_count--;                                 \
                    132:        s = splraiseipl(mtx->mtx_ipl);                                  \
1.12      matt      133:        if (x__cnt == MUTEX_COUNT_BIAS)                                 \
1.2       ad        134:                x__ci->ci_mtx_oldspl = (s);                             \
                    135: } while (/* CONSTCOND */ 0)
                    136:
                    137: #define        MUTEX_SPIN_SPLRESTORE(mtx)                                      \
                    138: do {                                                                   \
                    139:        struct cpu_info *x__ci = curcpu();                              \
                    140:        int s = x__ci->ci_mtx_oldspl;                                   \
                    141:        __insn_barrier();                                               \
1.12      matt      142:        if (++(x__ci->ci_mtx_count) == MUTEX_COUNT_BIAS)                \
1.2       ad        143:                splx(s);                                                \
                    144: } while (/* CONSTCOND */ 0)
                    145:
                    146: /*
                    147:  * For architectures that provide 'simple' mutexes: they provide a
                    148:  * CAS function that is either MP-safe, or does not need to be MP
                    149:  * safe.  Adaptive mutexes on these architectures do not require an
                    150:  * additional interlock.
                    151:  */
                    152:
                    153: #ifdef __HAVE_SIMPLE_MUTEXES
                    154:
                    155: #define        MUTEX_OWNER(owner)                                              \
                    156:        (owner & MUTEX_THREAD)
                    157: #define        MUTEX_OWNED(owner)                                              \
                    158:        (owner != 0)
                    159: #define        MUTEX_HAS_WAITERS(mtx)                                          \
                    160:        (((int)(mtx)->mtx_owner & MUTEX_BIT_WAITERS) != 0)
                    161:
                    162: #define        MUTEX_INITIALIZE_ADAPTIVE(mtx, id)                              \
                    163: do {                                                                   \
                    164:        (mtx)->mtx_id = (id);                                           \
                    165: } while (/* CONSTCOND */ 0);
                    166:
                    167: #define        MUTEX_INITIALIZE_SPIN(mtx, id, ipl)                             \
                    168: do {                                                                   \
                    169:        (mtx)->mtx_owner = MUTEX_BIT_SPIN;                              \
                    170:        (mtx)->mtx_ipl = makeiplcookie((ipl));                          \
                    171:        (mtx)->mtx_id = (id);                                           \
                    172:        __cpu_simple_lock_init(&(mtx)->mtx_lock);                       \
                    173: } while (/* CONSTCOND */ 0)
                    174:
                    175: #define        MUTEX_DESTROY(mtx)                                              \
                    176: do {                                                                   \
                    177:        (mtx)->mtx_owner = MUTEX_THREAD;                                \
                    178:        (mtx)->mtx_id = -1;                                             \
                    179: } while (/* CONSTCOND */ 0);
                    180:
                    181: #define        MUTEX_SPIN_P(mtx)               \
                    182:     (((mtx)->mtx_owner & MUTEX_BIT_SPIN) != 0)
                    183: #define        MUTEX_ADAPTIVE_P(mtx)           \
                    184:     (((mtx)->mtx_owner & MUTEX_BIT_SPIN) == 0)
                    185:
                    186: #define        MUTEX_GETID(mtx)                ((mtx)->mtx_id)
                    187:
                    188: static inline int
                    189: MUTEX_ACQUIRE(kmutex_t *mtx, uintptr_t curthread)
                    190: {
                    191:        int rv;
                    192:        rv = MUTEX_CAS(&mtx->mtx_owner, 0UL, curthread);
1.7       itohy     193:        MUTEX_RECEIVE(mtx);
1.2       ad        194:        return rv;
                    195: }
                    196:
                    197: static inline int
                    198: MUTEX_SET_WAITERS(kmutex_t *mtx, uintptr_t owner)
                    199: {
                    200:        int rv;
                    201:        rv = MUTEX_CAS(&mtx->mtx_owner, owner, owner | MUTEX_BIT_WAITERS);
1.7       itohy     202:        MUTEX_RECEIVE(mtx);
1.2       ad        203:        return rv;
                    204: }
                    205:
                    206: static inline void
                    207: MUTEX_RELEASE(kmutex_t *mtx)
                    208: {
1.7       itohy     209:        MUTEX_GIVE(mtx);
1.2       ad        210:        mtx->mtx_owner = 0;
                    211: }
1.4       ad        212:
                    213: static inline void
                    214: MUTEX_CLEAR_WAITERS(kmutex_t *mtx)
                    215: {
                    216:        /* nothing */
                    217: }
1.2       ad        218: #endif /* __HAVE_SIMPLE_MUTEXES */
                    219:
                    220: /*
                    221:  * Patch in stubs via strong alias where they are not available.
                    222:  */
                    223:
                    224: #if defined(LOCKDEBUG)
                    225: #undef __HAVE_MUTEX_STUBS
                    226: #undef __HAVE_SPIN_MUTEX_STUBS
                    227: #endif
                    228:
                    229: #ifndef __HAVE_MUTEX_STUBS
1.8       itohy     230: __strong_alias(mutex_enter,mutex_vector_enter);
                    231: __strong_alias(mutex_exit,mutex_vector_exit);
1.2       ad        232: #endif
                    233:
                    234: #ifndef __HAVE_SPIN_MUTEX_STUBS
1.8       itohy     235: __strong_alias(mutex_spin_enter,mutex_vector_enter);
                    236: __strong_alias(mutex_spin_exit,mutex_vector_exit);
1.2       ad        237: #endif
                    238:
                    239: void   mutex_abort(kmutex_t *, const char *, const char *);
                    240: void   mutex_dump(volatile void *);
                    241: int    mutex_onproc(uintptr_t, struct cpu_info **);
1.6       ad        242: static struct lwp *mutex_owner(wchan_t);
1.2       ad        243:
                    244: lockops_t mutex_spin_lockops = {
                    245:        "Mutex",
                    246:        0,
                    247:        mutex_dump
                    248: };
                    249:
                    250: lockops_t mutex_adaptive_lockops = {
                    251:        "Mutex",
                    252:        1,
                    253:        mutex_dump
                    254: };
                    255:
1.5       yamt      256: syncobj_t mutex_syncobj = {
                    257:        SOBJ_SLEEPQ_SORTED,
                    258:        turnstile_unsleep,
                    259:        turnstile_changepri,
                    260:        sleepq_lendpri,
1.6       ad        261:        mutex_owner,
1.5       yamt      262: };
                    263:
1.2       ad        264: /*
                    265:  * mutex_dump:
                    266:  *
                    267:  *     Dump the contents of a mutex structure.
                    268:  */
                    269: void
                    270: mutex_dump(volatile void *cookie)
                    271: {
                    272:        volatile kmutex_t *mtx = cookie;
                    273:
                    274:        printf_nolog("owner field  : %#018lx wait/spin: %16d/%d\n",
                    275:            (long)MUTEX_OWNER(mtx->mtx_owner), MUTEX_HAS_WAITERS(mtx),
                    276:            MUTEX_SPIN_P(mtx));
                    277: }
                    278:
                    279: /*
                    280:  * mutex_abort:
                    281:  *
1.3       ad        282:  *     Dump information about an error and panic the system.  This
                    283:  *     generates a lot of machine code in the DIAGNOSTIC case, so
                    284:  *     we ask the compiler to not inline it.
1.2       ad        285:  */
1.8       itohy     286:
                    287: #if __GNUC_PREREQ__(3, 0)
                    288: __attribute ((noinline)) __attribute ((noreturn))
                    289: #endif
                    290: void
1.2       ad        291: mutex_abort(kmutex_t *mtx, const char *func, const char *msg)
                    292: {
                    293:
                    294:        LOCKDEBUG_ABORT(MUTEX_GETID(mtx), mtx, (MUTEX_SPIN_P(mtx) ?
1.3       ad        295:            &mutex_spin_lockops : &mutex_adaptive_lockops), func, msg);
1.2       ad        296:        /* NOTREACHED */
                    297: }
                    298:
                    299: /*
                    300:  * mutex_init:
                    301:  *
                    302:  *     Initialize a mutex for use.  Note that adaptive mutexes are in
                    303:  *     essence spin mutexes that can sleep to avoid deadlock and wasting
                    304:  *     CPU time.  We can't easily provide a type of mutex that always
                    305:  *     sleeps - see comments in mutex_vector_enter() about releasing
                    306:  *     mutexes unlocked.
                    307:  */
                    308: void
                    309: mutex_init(kmutex_t *mtx, kmutex_type_t type, int ipl)
                    310: {
                    311:        u_int id;
                    312:
                    313:        memset(mtx, 0, sizeof(*mtx));
                    314:
1.15      ad        315:        switch (type) {
                    316:        case MUTEX_ADAPTIVE:
                    317:        case MUTEX_DEFAULT:
                    318:                KASSERT(ipl == IPL_NONE);
                    319:                break;
                    320:        case MUTEX_DRIVER:
1.2       ad        321:                type = (ipl == IPL_NONE ? MUTEX_ADAPTIVE : MUTEX_SPIN);
1.15      ad        322:                break;
                    323:        default:
                    324:                break;
                    325:        }
1.2       ad        326:
                    327:        switch (type) {
1.11      ad        328:        case MUTEX_NODEBUG:
1.19      ad        329:                id = LOCKDEBUG_ALLOC(mtx, NULL,
                    330:                    (uintptr_t)__builtin_return_address(0));
1.15      ad        331:                MUTEX_INITIALIZE_SPIN(mtx, id, ipl);
1.11      ad        332:                break;
1.2       ad        333:        case MUTEX_ADAPTIVE:
                    334:        case MUTEX_DEFAULT:
1.19      ad        335:                id = LOCKDEBUG_ALLOC(mtx, &mutex_adaptive_lockops,
                    336:                    (uintptr_t)__builtin_return_address(0));
1.2       ad        337:                MUTEX_INITIALIZE_ADAPTIVE(mtx, id);
                    338:                break;
                    339:        case MUTEX_SPIN:
1.19      ad        340:                id = LOCKDEBUG_ALLOC(mtx, &mutex_spin_lockops,
                    341:                    (uintptr_t)__builtin_return_address(0));
1.2       ad        342:                MUTEX_INITIALIZE_SPIN(mtx, id, ipl);
                    343:                break;
                    344:        default:
                    345:                panic("mutex_init: impossible type");
                    346:                break;
                    347:        }
                    348: }
                    349:
                    350: /*
                    351:  * mutex_destroy:
                    352:  *
                    353:  *     Tear down a mutex.
                    354:  */
                    355: void
                    356: mutex_destroy(kmutex_t *mtx)
                    357: {
                    358:
                    359:        if (MUTEX_ADAPTIVE_P(mtx)) {
                    360:                MUTEX_ASSERT(mtx, !MUTEX_OWNED(mtx->mtx_owner) &&
                    361:                    !MUTEX_HAS_WAITERS(mtx));
                    362:        } else {
1.16      skrll     363:                MUTEX_ASSERT(mtx, !__SIMPLELOCK_LOCKED_P(&mtx->mtx_lock));
1.2       ad        364:        }
                    365:
                    366:        LOCKDEBUG_FREE(mtx, MUTEX_GETID(mtx));
                    367:        MUTEX_DESTROY(mtx);
                    368: }
                    369:
                    370: /*
                    371:  * mutex_onproc:
                    372:  *
                    373:  *     Return true if an adaptive mutex owner is running on a CPU in the
                    374:  *     system.  If the target is waiting on the kernel big lock, then we
1.15      ad        375:  *     must release it.  This is necessary to avoid deadlock.
1.2       ad        376:  *
                    377:  *     Note that we can't use the mutex owner field as an LWP pointer.  We
                    378:  *     don't have full control over the timing of our execution, and so the
                    379:  *     pointer could be completely invalid by the time we dereference it.
                    380:  */
                    381: #ifdef MULTIPROCESSOR
                    382: int
                    383: mutex_onproc(uintptr_t owner, struct cpu_info **cip)
                    384: {
                    385:        CPU_INFO_ITERATOR cii;
                    386:        struct cpu_info *ci;
                    387:        struct lwp *l;
                    388:
                    389:        if (!MUTEX_OWNED(owner))
                    390:                return 0;
                    391:        l = (struct lwp *)MUTEX_OWNER(owner);
                    392:
1.15      ad        393:        /* See if the target is running on a CPU somewhere. */
1.10      ad        394:        if ((ci = *cip) != NULL && ci->ci_curlwp == l)
1.15      ad        395:                goto run;
                    396:        for (CPU_INFO_FOREACH(cii, ci))
                    397:                if (ci->ci_curlwp == l)
                    398:                        goto run;
1.2       ad        399:
1.15      ad        400:        /* No: it may be safe to block now. */
1.2       ad        401:        *cip = NULL;
                    402:        return 0;
1.15      ad        403:
                    404:  run:
                    405:        /* Target is running; do we need to block? */
                    406:        *cip = ci;
                    407:        return ci->ci_biglock_wanted != l;
1.2       ad        408: }
1.15      ad        409: #endif /* MULTIPROCESSOR */
1.2       ad        410:
                    411: /*
                    412:  * mutex_vector_enter:
                    413:  *
                    414:  *     Support routine for mutex_enter() that must handles all cases.  In
                    415:  *     the LOCKDEBUG case, mutex_enter() is always aliased here, even if
                    416:  *     fast-path stubs are available.  If an mutex_spin_enter() stub is
                    417:  *     not available, then it is also aliased directly here.
                    418:  */
                    419: void
                    420: mutex_vector_enter(kmutex_t *mtx)
                    421: {
                    422:        uintptr_t owner, curthread;
                    423:        turnstile_t *ts;
                    424: #ifdef MULTIPROCESSOR
                    425:        struct cpu_info *ci = NULL;
                    426:        u_int count;
                    427: #endif
                    428:        LOCKSTAT_COUNTER(spincnt);
                    429:        LOCKSTAT_COUNTER(slpcnt);
                    430:        LOCKSTAT_TIMER(spintime);
                    431:        LOCKSTAT_TIMER(slptime);
                    432:        LOCKSTAT_FLAG(lsflag);
                    433:
                    434:        /*
                    435:         * Handle spin mutexes.
                    436:         */
                    437:        if (MUTEX_SPIN_P(mtx)) {
                    438: #if defined(LOCKDEBUG) && defined(MULTIPROCESSOR)
                    439:                u_int spins = 0;
                    440: #endif
                    441:                MUTEX_SPIN_SPLRAISE(mtx);
                    442:                MUTEX_WANTLOCK(mtx);
                    443: #ifdef FULL
                    444:                if (__cpu_simple_lock_try(&mtx->mtx_lock)) {
                    445:                        MUTEX_LOCKED(mtx);
                    446:                        return;
                    447:                }
                    448: #if !defined(MULTIPROCESSOR)
                    449:                MUTEX_ABORT(mtx, "locking against myself");
                    450: #else /* !MULTIPROCESSOR */
                    451:
                    452:                LOCKSTAT_ENTER(lsflag);
                    453:                LOCKSTAT_START_TIMER(lsflag, spintime);
                    454:                count = SPINLOCK_BACKOFF_MIN;
                    455:
                    456:                /*
                    457:                 * Spin testing the lock word and do exponential backoff
                    458:                 * to reduce cache line ping-ponging between CPUs.
                    459:                 */
                    460:                do {
                    461:                        if (panicstr != NULL)
                    462:                                break;
1.16      skrll     463:                        while (__SIMPLELOCK_LOCKED_P(&mtx->mtx_lock)) {
1.2       ad        464:                                SPINLOCK_BACKOFF(count);
                    465: #ifdef LOCKDEBUG
                    466:                                if (SPINLOCK_SPINOUT(spins))
                    467:                                        MUTEX_ABORT(mtx, "spinout");
                    468: #endif /* LOCKDEBUG */
                    469:                        }
                    470:                } while (!__cpu_simple_lock_try(&mtx->mtx_lock));
                    471:
                    472:                if (count != SPINLOCK_BACKOFF_MIN) {
                    473:                        LOCKSTAT_STOP_TIMER(lsflag, spintime);
                    474:                        LOCKSTAT_EVENT(lsflag, mtx,
                    475:                            LB_SPIN_MUTEX | LB_SPIN, 1, spintime);
                    476:                }
                    477:                LOCKSTAT_EXIT(lsflag);
                    478: #endif /* !MULTIPROCESSOR */
                    479: #endif /* FULL */
                    480:                MUTEX_LOCKED(mtx);
                    481:                return;
                    482:        }
                    483:
                    484:        curthread = (uintptr_t)curlwp;
                    485:
                    486:        MUTEX_DASSERT(mtx, MUTEX_ADAPTIVE_P(mtx));
                    487:        MUTEX_ASSERT(mtx, curthread != 0);
                    488:        MUTEX_WANTLOCK(mtx);
                    489:
                    490: #ifdef LOCKDEBUG
                    491:        if (panicstr == NULL) {
                    492:                simple_lock_only_held(NULL, "mutex_enter");
                    493: #ifdef MULTIPROCESSOR
                    494:                LOCKDEBUG_BARRIER(&kernel_lock, 1);
                    495: #else
                    496:                LOCKDEBUG_BARRIER(NULL, 1);
                    497: #endif
                    498:        }
                    499: #endif
                    500:
                    501:        LOCKSTAT_ENTER(lsflag);
                    502:
                    503:        /*
                    504:         * Adaptive mutex; spin trying to acquire the mutex.  If we
                    505:         * determine that the owner is not running on a processor,
                    506:         * then we stop spinning, and sleep instead.
                    507:         */
                    508:        for (;;) {
                    509:                owner = mtx->mtx_owner;
                    510:                if (!MUTEX_OWNED(owner)) {
                    511:                        /*
                    512:                         * Mutex owner clear could mean two things:
                    513:                         *
                    514:                         *      * The mutex has been released.
                    515:                         *      * The owner field hasn't been set yet.
                    516:                         *
                    517:                         * Try to acquire it again.  If that fails,
                    518:                         * we'll just loop again.
                    519:                         */
                    520:                        if (MUTEX_ACQUIRE(mtx, curthread))
                    521:                                break;
                    522:                        continue;
                    523:                }
                    524:
                    525:                if (panicstr != NULL)
                    526:                        return;
                    527:                if (MUTEX_OWNER(owner) == curthread)
                    528:                        MUTEX_ABORT(mtx, "locking against myself");
                    529:
                    530: #ifdef MULTIPROCESSOR
                    531:                /*
                    532:                 * Check to see if the owner is running on a processor.
                    533:                 * If so, then we should just spin, as the owner will
                    534:                 * likely release the lock very soon.
                    535:                 */
                    536:                if (mutex_onproc(owner, &ci)) {
                    537:                        LOCKSTAT_START_TIMER(lsflag, spintime);
                    538:                        count = SPINLOCK_BACKOFF_MIN;
                    539:                        for (;;) {
                    540:                                owner = mtx->mtx_owner;
                    541:                                if (!mutex_onproc(owner, &ci))
                    542:                                        break;
                    543:                                SPINLOCK_BACKOFF(count);
                    544:                        }
                    545:                        LOCKSTAT_STOP_TIMER(lsflag, spintime);
                    546:                        LOCKSTAT_COUNT(spincnt, 1);
                    547:                        if (!MUTEX_OWNED(owner))
                    548:                                continue;
                    549:                }
                    550: #endif
                    551:
                    552:                ts = turnstile_lookup(mtx);
                    553:
                    554:                /*
                    555:                 * Once we have the turnstile chain interlock, mark the
                    556:                 * mutex has having waiters.  If that fails, spin again:
                    557:                 * chances are that the mutex has been released.
                    558:                 */
                    559:                if (!MUTEX_SET_WAITERS(mtx, owner)) {
                    560:                        turnstile_exit(mtx);
                    561:                        continue;
                    562:                }
                    563:
                    564: #ifdef MULTIPROCESSOR
                    565:                /*
                    566:                 * mutex_exit() is permitted to release the mutex without
                    567:                 * any interlocking instructions, and the following can
                    568:                 * occur as a result:
                    569:                 *
                    570:                 *  CPU 1: MUTEX_SET_WAITERS()      CPU2: mutex_exit()
                    571:                 * ---------------------------- ----------------------------
                    572:                 *              ..                  acquire cache line
                    573:                 *              ..                   test for waiters
                    574:                 *      acquire cache line    <-      lose cache line
                    575:                 *       lock cache line                   ..
                    576:                 *     verify mutex is held                ..
                    577:                 *          set waiters                    ..
                    578:                 *       unlock cache line                 ..
                    579:                 *        lose cache line     ->    acquire cache line
                    580:                 *              ..                clear lock word, waiters
                    581:                 *        return success
                    582:                 *
                    583:                 * There is a another race that can occur: a third CPU could
                    584:                 * acquire the mutex as soon as it is released.  Since
                    585:                 * adaptive mutexes are primarily spin mutexes, this is not
                    586:                 * something that we need to worry about too much.  What we
                    587:                 * do need to ensure is that the waiters bit gets set.
                    588:                 *
                    589:                 * To allow the unlocked release, we need to make some
                    590:                 * assumptions here:
                    591:                 *
                    592:                 * o Release is the only non-atomic/unlocked operation
                    593:                 *   that can be performed on the mutex.  (It must still
                    594:                 *   be atomic on the local CPU, e.g. in case interrupted
                    595:                 *   or preempted).
                    596:                 *
                    597:                 * o At any given time, MUTEX_SET_WAITERS() can only ever
1.21    ! pooka     598:                 *   be in progress on one CPU in the system - guaranteed
1.2       ad        599:                 *   by the turnstile chain lock.
                    600:                 *
                    601:                 * o No other operations other than MUTEX_SET_WAITERS()
                    602:                 *   and release can modify a mutex with a non-zero
                    603:                 *   owner field.
                    604:                 *
                    605:                 * o The result of a successful MUTEX_SET_WAITERS() call
                    606:                 *   is an unbuffered write that is immediately visible
                    607:                 *   to all other processors in the system.
                    608:                 *
                    609:                 * o If the holding LWP switches away, it posts a store
                    610:                 *   fence before changing curlwp, ensuring that any
                    611:                 *   overwrite of the mutex waiters flag by mutex_exit()
                    612:                 *   completes before the modification of curlwp becomes
                    613:                 *   visible to this CPU.
                    614:                 *
1.14      yamt      615:                 * o mi_switch() posts a store fence before setting curlwp
1.2       ad        616:                 *   and before resuming execution of an LWP.
                    617:                 *
                    618:                 * o _kernel_lock() posts a store fence before setting
                    619:                 *   curcpu()->ci_biglock_wanted, and after clearing it.
                    620:                 *   This ensures that any overwrite of the mutex waiters
                    621:                 *   flag by mutex_exit() completes before the modification
                    622:                 *   of ci_biglock_wanted becomes visible.
                    623:                 *
                    624:                 * We now post a read memory barrier (after setting the
                    625:                 * waiters field) and check the lock holder's status again.
                    626:                 * Some of the possible outcomes (not an exhaustive list):
                    627:                 *
                    628:                 * 1. The onproc check returns true: the holding LWP is
                    629:                 *    running again.  The lock may be released soon and
                    630:                 *    we should spin.  Importantly, we can't trust the
                    631:                 *    value of the waiters flag.
                    632:                 *
                    633:                 * 2. The onproc check returns false: the holding LWP is
                    634:                 *    not running.  We now have the oppertunity to check
                    635:                 *    if mutex_exit() has blatted the modifications made
                    636:                 *    by MUTEX_SET_WAITERS().
                    637:                 *
                    638:                 * 3. The onproc check returns false: the holding LWP may
                    639:                 *    or may not be running.  It has context switched at
                    640:                 *    some point during our check.  Again, we have the
                    641:                 *    chance to see if the waiters bit is still set or
                    642:                 *    has been overwritten.
                    643:                 *
                    644:                 * 4. The onproc check returns false: the holding LWP is
                    645:                 *    running on a CPU, but wants the big lock.  It's OK
                    646:                 *    to check the waiters field in this case.
                    647:                 *
                    648:                 * 5. The has-waiters check fails: the mutex has been
                    649:                 *    released, the waiters flag cleared and another LWP
                    650:                 *    now owns the mutex.
                    651:                 *
                    652:                 * 6. The has-waiters check fails: the mutex has been
                    653:                 *    released.
                    654:                 *
                    655:                 * If the waiters bit is not set it's unsafe to go asleep,
                    656:                 * as we might never be awoken.
                    657:                 */
1.13      ad        658:                if ((mb_read(), mutex_onproc(owner, &ci)) ||
                    659:                    (mb_read(), !MUTEX_HAS_WAITERS(mtx))) {
1.2       ad        660:                        turnstile_exit(mtx);
                    661:                        continue;
                    662:                }
                    663: #endif /* MULTIPROCESSOR */
                    664:
                    665:                LOCKSTAT_START_TIMER(lsflag, slptime);
                    666:
1.5       yamt      667:                turnstile_block(ts, TS_WRITER_Q, mtx, &mutex_syncobj);
1.2       ad        668:
                    669:                LOCKSTAT_STOP_TIMER(lsflag, slptime);
                    670:                LOCKSTAT_COUNT(slpcnt, 1);
                    671:        }
                    672:
                    673:        LOCKSTAT_EVENT(lsflag, mtx, LB_ADAPTIVE_MUTEX | LB_SLEEP1,
                    674:            slpcnt, slptime);
                    675:        LOCKSTAT_EVENT(lsflag, mtx, LB_ADAPTIVE_MUTEX | LB_SPIN,
                    676:            spincnt, spintime);
                    677:        LOCKSTAT_EXIT(lsflag);
                    678:
                    679:        MUTEX_DASSERT(mtx, MUTEX_OWNER(mtx->mtx_owner) == curthread);
                    680:        MUTEX_LOCKED(mtx);
                    681: }
                    682:
                    683: /*
                    684:  * mutex_vector_exit:
                    685:  *
                    686:  *     Support routine for mutex_exit() that handles all cases.
                    687:  */
                    688: void
                    689: mutex_vector_exit(kmutex_t *mtx)
                    690: {
                    691:        turnstile_t *ts;
                    692:        uintptr_t curthread;
                    693:
                    694:        if (MUTEX_SPIN_P(mtx)) {
                    695: #ifdef FULL
1.16      skrll     696:                if (!__SIMPLELOCK_LOCKED_P(&mtx->mtx_lock))
1.2       ad        697:                        MUTEX_ABORT(mtx, "exiting unheld spin mutex");
                    698:                MUTEX_UNLOCKED(mtx);
                    699:                __cpu_simple_unlock(&mtx->mtx_lock);
                    700: #endif
                    701:                MUTEX_SPIN_SPLRESTORE(mtx);
                    702:                return;
                    703:        }
                    704:
1.11      ad        705:        if (__predict_false((uintptr_t)panicstr | cold)) {
1.2       ad        706:                MUTEX_UNLOCKED(mtx);
                    707:                MUTEX_RELEASE(mtx);
                    708:                return;
                    709:        }
                    710:
                    711:        curthread = (uintptr_t)curlwp;
                    712:        MUTEX_DASSERT(mtx, curthread != 0);
                    713:        MUTEX_ASSERT(mtx, MUTEX_OWNER(mtx->mtx_owner) == curthread);
                    714:        MUTEX_UNLOCKED(mtx);
                    715:
1.15      ad        716: #ifdef LOCKDEBUG
                    717:        /*
                    718:         * Avoid having to take the turnstile chain lock every time
                    719:         * around.  Raise the priority level to splhigh() in order
                    720:         * to disable preemption and so make the following atomic.
                    721:         */
                    722:        {
                    723:                int s = splhigh();
                    724:                if (!MUTEX_HAS_WAITERS(mtx)) {
                    725:                        MUTEX_RELEASE(mtx);
                    726:                        splx(s);
                    727:                        return;
                    728:                }
                    729:                splx(s);
                    730:        }
                    731: #endif
                    732:
1.2       ad        733:        /*
                    734:         * Get this lock's turnstile.  This gets the interlock on
                    735:         * the sleep queue.  Once we have that, we can clear the
                    736:         * lock.  If there was no turnstile for the lock, there
                    737:         * were no waiters remaining.
                    738:         */
                    739:        ts = turnstile_lookup(mtx);
                    740:
                    741:        if (ts == NULL) {
                    742:                MUTEX_RELEASE(mtx);
                    743:                turnstile_exit(mtx);
                    744:        } else {
                    745:                MUTEX_RELEASE(mtx);
                    746:                turnstile_wakeup(ts, TS_WRITER_Q,
                    747:                    TS_WAITERS(ts, TS_WRITER_Q), NULL);
                    748:        }
                    749: }
                    750:
1.4       ad        751: #ifndef __HAVE_SIMPLE_MUTEXES
                    752: /*
                    753:  * mutex_wakeup:
                    754:  *
                    755:  *     Support routine for mutex_exit() that wakes up all waiters.
                    756:  *     We assume that the mutex has been released, but it need not
                    757:  *     be.
                    758:  */
                    759: void
                    760: mutex_wakeup(kmutex_t *mtx)
                    761: {
                    762:        turnstile_t *ts;
                    763:
                    764:        ts = turnstile_lookup(mtx);
                    765:        if (ts == NULL) {
                    766:                turnstile_exit(mtx);
                    767:                return;
                    768:        }
                    769:        MUTEX_CLEAR_WAITERS(mtx);
                    770:        turnstile_wakeup(ts, TS_WRITER_Q, TS_WAITERS(ts, TS_WRITER_Q), NULL);
                    771: }
                    772: #endif /* !__HAVE_SIMPLE_MUTEXES */
                    773:
1.2       ad        774: /*
                    775:  * mutex_owned:
                    776:  *
1.3       ad        777:  *     Return true if the current LWP (adaptive) or CPU (spin)
                    778:  *     holds the mutex.
1.2       ad        779:  */
                    780: int
                    781: mutex_owned(kmutex_t *mtx)
                    782: {
                    783:
                    784:        if (MUTEX_ADAPTIVE_P(mtx))
                    785:                return MUTEX_OWNER(mtx->mtx_owner) == (uintptr_t)curlwp;
                    786: #ifdef FULL
1.16      skrll     787:        return __SIMPLELOCK_LOCKED_P(&mtx->mtx_lock);
1.2       ad        788: #else
                    789:        return 1;
                    790: #endif
                    791: }
                    792:
                    793: /*
                    794:  * mutex_owner:
                    795:  *
1.6       ad        796:  *     Return the current owner of an adaptive mutex.  Used for
                    797:  *     priority inheritance.
1.2       ad        798:  */
1.6       ad        799: static struct lwp *
                    800: mutex_owner(wchan_t obj)
1.2       ad        801: {
1.6       ad        802:        kmutex_t *mtx = (void *)(uintptr_t)obj; /* discard qualifiers */
1.2       ad        803:
                    804:        MUTEX_ASSERT(mtx, MUTEX_ADAPTIVE_P(mtx));
                    805:        return (struct lwp *)MUTEX_OWNER(mtx->mtx_owner);
                    806: }
                    807:
                    808: /*
                    809:  * mutex_tryenter:
                    810:  *
                    811:  *     Try to acquire the mutex; return non-zero if we did.
                    812:  */
                    813: int
                    814: mutex_tryenter(kmutex_t *mtx)
                    815: {
                    816:        uintptr_t curthread;
                    817:
                    818:        /*
                    819:         * Handle spin mutexes.
                    820:         */
                    821:        if (MUTEX_SPIN_P(mtx)) {
                    822:                MUTEX_SPIN_SPLRAISE(mtx);
                    823: #ifdef FULL
                    824:                if (__cpu_simple_lock_try(&mtx->mtx_lock)) {
1.4       ad        825:                        MUTEX_WANTLOCK(mtx);
1.2       ad        826:                        MUTEX_LOCKED(mtx);
                    827:                        return 1;
                    828:                }
                    829:                MUTEX_SPIN_SPLRESTORE(mtx);
                    830: #else
1.4       ad        831:                MUTEX_WANTLOCK(mtx);
1.2       ad        832:                MUTEX_LOCKED(mtx);
                    833:                return 1;
                    834: #endif
                    835:        } else {
                    836:                curthread = (uintptr_t)curlwp;
                    837:                MUTEX_ASSERT(mtx, curthread != 0);
                    838:                if (MUTEX_ACQUIRE(mtx, curthread)) {
1.4       ad        839:                        MUTEX_WANTLOCK(mtx);
1.2       ad        840:                        MUTEX_LOCKED(mtx);
                    841:                        MUTEX_DASSERT(mtx,
                    842:                            MUTEX_OWNER(mtx->mtx_owner) == curthread);
                    843:                        return 1;
                    844:                }
                    845:        }
                    846:
                    847:        return 0;
                    848: }
                    849:
                    850: #if defined(__HAVE_SPIN_MUTEX_STUBS) || defined(FULL)
                    851: /*
                    852:  * mutex_spin_retry:
                    853:  *
                    854:  *     Support routine for mutex_spin_enter().  Assumes that the caller
                    855:  *     has already raised the SPL, and adjusted counters.
                    856:  */
                    857: void
                    858: mutex_spin_retry(kmutex_t *mtx)
                    859: {
                    860: #ifdef MULTIPROCESSOR
                    861:        u_int count;
                    862:        LOCKSTAT_TIMER(spintime);
                    863:        LOCKSTAT_FLAG(lsflag);
                    864: #ifdef LOCKDEBUG
                    865:        u_int spins = 0;
                    866: #endif /* LOCKDEBUG */
                    867:
                    868:        MUTEX_WANTLOCK(mtx);
                    869:
                    870:        LOCKSTAT_ENTER(lsflag);
                    871:        LOCKSTAT_START_TIMER(lsflag, spintime);
                    872:        count = SPINLOCK_BACKOFF_MIN;
                    873:
                    874:        /*
                    875:         * Spin testing the lock word and do exponential backoff
                    876:         * to reduce cache line ping-ponging between CPUs.
                    877:         */
                    878:        do {
                    879:                if (panicstr != NULL)
                    880:                        break;
1.16      skrll     881:                while (__SIMPLELOCK_LOCKED_P(&mtx->mtx_lock)) {
1.2       ad        882:                        SPINLOCK_BACKOFF(count);
                    883: #ifdef LOCKDEBUG
                    884:                        if (SPINLOCK_SPINOUT(spins))
                    885:                                MUTEX_ABORT(mtx, "spinout");
                    886: #endif /* LOCKDEBUG */
                    887:                }
                    888:        } while (!__cpu_simple_lock_try(&mtx->mtx_lock));
                    889:
                    890:        LOCKSTAT_STOP_TIMER(lsflag, spintime);
                    891:        LOCKSTAT_EVENT(lsflag, mtx, LB_SPIN_MUTEX | LB_SPIN, 1, spintime);
                    892:        LOCKSTAT_EXIT(lsflag);
                    893:
                    894:        MUTEX_LOCKED(mtx);
                    895: #else  /* MULTIPROCESSOR */
                    896:        MUTEX_ABORT(mtx, "locking against myself");
                    897: #endif /* MULTIPROCESSOR */
                    898: }
                    899: #endif /* defined(__HAVE_SPIN_MUTEX_STUBS) || defined(FULL) */

CVSweb <webmaster@jp.NetBSD.org>