[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 1.3

1.3     ! ad          1: /*     $NetBSD: kern_mutex.c,v 1.2 2007/02/09 21:55:30 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: #include "opt_multiprocessor.h"
                     48:
                     49: #define        __MUTEX_PRIVATE
                     50:
                     51: #include <sys/cdefs.h>
1.3     ! ad         52: __KERNEL_RCSID(0, "$NetBSD: kern_mutex.c,v 1.2 2007/02/09 21:55:30 ad Exp $");
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:
                     65: #include <machine/intr.h>
                     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)                                   \
                     90:     mutex_abort(mtx, __FUNCTION__, msg)
                     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:  */
                    123:
                    124: #define        MUTEX_SPIN_SPLRAISE(mtx)                                        \
                    125: do {                                                                   \
                    126:        struct cpu_info *x__ci = curcpu();                              \
                    127:        int x__cnt, s;                                                  \
                    128:        x__cnt = x__ci->ci_mtx_count--;                                 \
                    129:        s = splraiseipl(mtx->mtx_ipl);                                  \
                    130:        if (x__cnt == 0)                                                \
                    131:                x__ci->ci_mtx_oldspl = (s);                             \
                    132: } while (/* CONSTCOND */ 0)
                    133:
                    134: #define        MUTEX_SPIN_SPLRESTORE(mtx)                                      \
                    135: do {                                                                   \
                    136:        struct cpu_info *x__ci = curcpu();                              \
                    137:        int s = x__ci->ci_mtx_oldspl;                                   \
                    138:        __insn_barrier();                                               \
                    139:        if (++(x__ci->ci_mtx_count) == 0)                               \
                    140:                splx(s);                                                \
                    141: } while (/* CONSTCOND */ 0)
                    142:
                    143: /*
                    144:  * For architectures that provide 'simple' mutexes: they provide a
                    145:  * CAS function that is either MP-safe, or does not need to be MP
                    146:  * safe.  Adaptive mutexes on these architectures do not require an
                    147:  * additional interlock.
                    148:  */
                    149:
                    150: #ifdef __HAVE_SIMPLE_MUTEXES
                    151:
                    152: #define        MUTEX_OWNER(owner)                                              \
                    153:        (owner & MUTEX_THREAD)
                    154: #define        MUTEX_OWNED(owner)                                              \
                    155:        (owner != 0)
                    156: #define        MUTEX_HAS_WAITERS(mtx)                                          \
                    157:        (((int)(mtx)->mtx_owner & MUTEX_BIT_WAITERS) != 0)
                    158:
                    159: #define        MUTEX_INITIALIZE_ADAPTIVE(mtx, id)                              \
                    160: do {                                                                   \
                    161:        (mtx)->mtx_id = (id);                                           \
                    162: } while (/* CONSTCOND */ 0);
                    163:
                    164: #define        MUTEX_INITIALIZE_SPIN(mtx, id, ipl)                             \
                    165: do {                                                                   \
                    166:        (mtx)->mtx_owner = MUTEX_BIT_SPIN;                              \
                    167:        (mtx)->mtx_ipl = makeiplcookie((ipl));                          \
                    168:        (mtx)->mtx_id = (id);                                           \
                    169:        __cpu_simple_lock_init(&(mtx)->mtx_lock);                       \
                    170: } while (/* CONSTCOND */ 0)
                    171:
                    172: #define        MUTEX_DESTROY(mtx)                                              \
                    173: do {                                                                   \
                    174:        (mtx)->mtx_owner = MUTEX_THREAD;                                \
                    175:        (mtx)->mtx_id = -1;                                             \
                    176: } while (/* CONSTCOND */ 0);
                    177:
                    178: #define        MUTEX_SPIN_P(mtx)               \
                    179:     (((mtx)->mtx_owner & MUTEX_BIT_SPIN) != 0)
                    180: #define        MUTEX_ADAPTIVE_P(mtx)           \
                    181:     (((mtx)->mtx_owner & MUTEX_BIT_SPIN) == 0)
                    182:
                    183: #define        MUTEX_GETID(mtx)                ((mtx)->mtx_id)
                    184:
                    185: static inline int
                    186: MUTEX_ACQUIRE(kmutex_t *mtx, uintptr_t curthread)
                    187: {
                    188:        int rv;
                    189:        rv = MUTEX_CAS(&mtx->mtx_owner, 0UL, curthread);
                    190:        MUTEX_RECEIVE();
                    191:        return rv;
                    192: }
                    193:
                    194: static inline int
                    195: MUTEX_SET_WAITERS(kmutex_t *mtx, uintptr_t owner)
                    196: {
                    197:        int rv;
                    198:        rv = MUTEX_CAS(&mtx->mtx_owner, owner, owner | MUTEX_BIT_WAITERS);
                    199:        MUTEX_RECEIVE();
                    200:        return rv;
                    201: }
                    202:
                    203: static inline void
                    204: MUTEX_RELEASE(kmutex_t *mtx)
                    205: {
                    206:        MUTEX_GIVE();
                    207:        mtx->mtx_owner = 0;
                    208: }
                    209: #endif /* __HAVE_SIMPLE_MUTEXES */
                    210:
                    211: /*
                    212:  * Patch in stubs via strong alias where they are not available.
                    213:  */
                    214:
                    215: #if defined(LOCKDEBUG)
                    216: #undef __HAVE_MUTEX_STUBS
                    217: #undef __HAVE_SPIN_MUTEX_STUBS
                    218: #endif
                    219:
                    220: #ifndef __HAVE_MUTEX_STUBS
                    221: __strong_alias(mutex_enter, mutex_vector_enter);
                    222: __strong_alias(mutex_exit, mutex_vector_exit);
                    223: #endif
                    224:
                    225: #ifndef __HAVE_SPIN_MUTEX_STUBS
                    226: __strong_alias(mutex_spin_enter, mutex_vector_enter);
                    227: __strong_alias(mutex_spin_exit, mutex_vector_exit);
                    228: #endif
                    229:
                    230: void   mutex_abort(kmutex_t *, const char *, const char *);
                    231: void   mutex_dump(volatile void *);
                    232: int    mutex_onproc(uintptr_t, struct cpu_info **);
                    233:
                    234: lockops_t mutex_spin_lockops = {
                    235:        "Mutex",
                    236:        0,
                    237:        mutex_dump
                    238: };
                    239:
                    240: lockops_t mutex_adaptive_lockops = {
                    241:        "Mutex",
                    242:        1,
                    243:        mutex_dump
                    244: };
                    245:
                    246: /*
                    247:  * mutex_dump:
                    248:  *
                    249:  *     Dump the contents of a mutex structure.
                    250:  */
                    251: void
                    252: mutex_dump(volatile void *cookie)
                    253: {
                    254:        volatile kmutex_t *mtx = cookie;
                    255:
                    256:        printf_nolog("owner field  : %#018lx wait/spin: %16d/%d\n",
                    257:            (long)MUTEX_OWNER(mtx->mtx_owner), MUTEX_HAS_WAITERS(mtx),
                    258:            MUTEX_SPIN_P(mtx));
                    259: }
                    260:
                    261: /*
                    262:  * mutex_abort:
                    263:  *
1.3     ! ad        264:  *     Dump information about an error and panic the system.  This
        !           265:  *     generates a lot of machine code in the DIAGNOSTIC case, so
        !           266:  *     we ask the compiler to not inline it.
1.2       ad        267:  */
                    268: __attribute ((noinline)) __attribute ((noreturn)) void
                    269: mutex_abort(kmutex_t *mtx, const char *func, const char *msg)
                    270: {
                    271:
                    272:        LOCKDEBUG_ABORT(MUTEX_GETID(mtx), mtx, (MUTEX_SPIN_P(mtx) ?
1.3     ! ad        273:            &mutex_spin_lockops : &mutex_adaptive_lockops), func, msg);
1.2       ad        274:        /* NOTREACHED */
                    275: }
                    276:
                    277: /*
                    278:  * mutex_init:
                    279:  *
                    280:  *     Initialize a mutex for use.  Note that adaptive mutexes are in
                    281:  *     essence spin mutexes that can sleep to avoid deadlock and wasting
                    282:  *     CPU time.  We can't easily provide a type of mutex that always
                    283:  *     sleeps - see comments in mutex_vector_enter() about releasing
                    284:  *     mutexes unlocked.
                    285:  */
                    286: void
                    287: mutex_init(kmutex_t *mtx, kmutex_type_t type, int ipl)
                    288: {
                    289:        u_int id;
                    290:
                    291:        memset(mtx, 0, sizeof(*mtx));
                    292:
                    293:        if (type == MUTEX_DRIVER)
                    294:                type = (ipl == IPL_NONE ? MUTEX_ADAPTIVE : MUTEX_SPIN);
                    295:
                    296:        switch (type) {
                    297:        case MUTEX_ADAPTIVE:
                    298:        case MUTEX_DEFAULT:
                    299:                KASSERT(ipl == IPL_NONE);
                    300:                id = LOCKDEBUG_ALLOC(mtx, &mutex_adaptive_lockops);
                    301:                MUTEX_INITIALIZE_ADAPTIVE(mtx, id);
                    302:                break;
                    303:        case MUTEX_SPIN:
                    304:                id = LOCKDEBUG_ALLOC(mtx, &mutex_spin_lockops);
                    305:                MUTEX_INITIALIZE_SPIN(mtx, id, ipl);
                    306:                break;
                    307:        default:
                    308:                panic("mutex_init: impossible type");
                    309:                break;
                    310:        }
                    311: }
                    312:
                    313: /*
                    314:  * mutex_destroy:
                    315:  *
                    316:  *     Tear down a mutex.
                    317:  */
                    318: void
                    319: mutex_destroy(kmutex_t *mtx)
                    320: {
                    321:
                    322:        if (MUTEX_ADAPTIVE_P(mtx)) {
                    323:                MUTEX_ASSERT(mtx, !MUTEX_OWNED(mtx->mtx_owner) &&
                    324:                    !MUTEX_HAS_WAITERS(mtx));
                    325:        } else {
                    326:                MUTEX_ASSERT(mtx, mtx->mtx_lock != __SIMPLELOCK_LOCKED);
                    327:        }
                    328:
                    329:        LOCKDEBUG_FREE(mtx, MUTEX_GETID(mtx));
                    330:        MUTEX_DESTROY(mtx);
                    331: }
                    332:
                    333: /*
                    334:  * mutex_onproc:
                    335:  *
                    336:  *     Return true if an adaptive mutex owner is running on a CPU in the
                    337:  *     system.  If the target is waiting on the kernel big lock, then we
                    338:  *     return false immediately.  This is necessary to avoid deadlock
                    339:  *     against the big lock.
                    340:  *
                    341:  *     Note that we can't use the mutex owner field as an LWP pointer.  We
                    342:  *     don't have full control over the timing of our execution, and so the
                    343:  *     pointer could be completely invalid by the time we dereference it.
1.3     ! ad        344:  *
        !           345:  *     XXX This should be optimised further to reduce potential cache line
        !           346:  *     ping-ponging and skewing of the spin time while busy waiting.
1.2       ad        347:  */
                    348: #ifdef MULTIPROCESSOR
                    349: int
                    350: mutex_onproc(uintptr_t owner, struct cpu_info **cip)
                    351: {
                    352:        CPU_INFO_ITERATOR cii;
                    353:        struct cpu_info *ci;
                    354:        struct lwp *l;
                    355:
                    356:        if (!MUTEX_OWNED(owner))
                    357:                return 0;
                    358:        l = (struct lwp *)MUTEX_OWNER(owner);
                    359:
                    360:        if ((ci = *cip) != NULL && ci->ci_curlwp == l) {
1.3     ! ad        361:                mb_read(); /* XXXSMP Very expensive, necessary? */
1.2       ad        362:                return ci->ci_biglock_wanted != l;
                    363:        }
                    364:
                    365:        for (CPU_INFO_FOREACH(cii, ci)) {
                    366:                if (ci->ci_curlwp == l) {
                    367:                        *cip = ci;
1.3     ! ad        368:                        mb_read(); /* XXXSMP Very expensive, necessary? */
1.2       ad        369:                        return ci->ci_biglock_wanted != l;
                    370:                }
                    371:        }
                    372:
                    373:        *cip = NULL;
                    374:        return 0;
                    375: }
                    376: #endif
                    377:
                    378: /*
                    379:  * mutex_vector_enter:
                    380:  *
                    381:  *     Support routine for mutex_enter() that must handles all cases.  In
                    382:  *     the LOCKDEBUG case, mutex_enter() is always aliased here, even if
                    383:  *     fast-path stubs are available.  If an mutex_spin_enter() stub is
                    384:  *     not available, then it is also aliased directly here.
                    385:  */
                    386: void
                    387: mutex_vector_enter(kmutex_t *mtx)
                    388: {
                    389:        uintptr_t owner, curthread;
                    390:        turnstile_t *ts;
                    391: #ifdef MULTIPROCESSOR
                    392:        struct cpu_info *ci = NULL;
                    393:        u_int count;
                    394: #endif
                    395:        LOCKSTAT_COUNTER(spincnt);
                    396:        LOCKSTAT_COUNTER(slpcnt);
                    397:        LOCKSTAT_TIMER(spintime);
                    398:        LOCKSTAT_TIMER(slptime);
                    399:        LOCKSTAT_FLAG(lsflag);
                    400:
                    401:        /*
                    402:         * Handle spin mutexes.
                    403:         */
                    404:        if (MUTEX_SPIN_P(mtx)) {
                    405: #if defined(LOCKDEBUG) && defined(MULTIPROCESSOR)
                    406:                u_int spins = 0;
                    407: #endif
                    408:                MUTEX_SPIN_SPLRAISE(mtx);
                    409:                MUTEX_WANTLOCK(mtx);
                    410: #ifdef FULL
                    411:                if (__cpu_simple_lock_try(&mtx->mtx_lock)) {
                    412:                        MUTEX_LOCKED(mtx);
                    413:                        return;
                    414:                }
                    415: #if !defined(MULTIPROCESSOR)
                    416:                MUTEX_ABORT(mtx, "locking against myself");
                    417: #else /* !MULTIPROCESSOR */
                    418:
                    419:                LOCKSTAT_ENTER(lsflag);
                    420:                LOCKSTAT_START_TIMER(lsflag, spintime);
                    421:                count = SPINLOCK_BACKOFF_MIN;
                    422:
                    423:                /*
                    424:                 * Spin testing the lock word and do exponential backoff
                    425:                 * to reduce cache line ping-ponging between CPUs.
                    426:                 */
                    427:                do {
                    428:                        if (panicstr != NULL)
                    429:                                break;
                    430:                        while (mtx->mtx_lock == __SIMPLELOCK_LOCKED) {
                    431:                                SPINLOCK_BACKOFF(count);
                    432: #ifdef LOCKDEBUG
                    433:                                if (SPINLOCK_SPINOUT(spins))
                    434:                                        MUTEX_ABORT(mtx, "spinout");
                    435: #endif /* LOCKDEBUG */
                    436:                        }
                    437:                } while (!__cpu_simple_lock_try(&mtx->mtx_lock));
                    438:
                    439:                if (count != SPINLOCK_BACKOFF_MIN) {
                    440:                        LOCKSTAT_STOP_TIMER(lsflag, spintime);
                    441:                        LOCKSTAT_EVENT(lsflag, mtx,
                    442:                            LB_SPIN_MUTEX | LB_SPIN, 1, spintime);
                    443:                }
                    444:                LOCKSTAT_EXIT(lsflag);
                    445: #endif /* !MULTIPROCESSOR */
                    446: #endif /* FULL */
                    447:                MUTEX_LOCKED(mtx);
                    448:                return;
                    449:        }
                    450:
                    451:        curthread = (uintptr_t)curlwp;
                    452:
                    453:        MUTEX_DASSERT(mtx, MUTEX_ADAPTIVE_P(mtx));
                    454:        MUTEX_ASSERT(mtx, curthread != 0);
                    455:        MUTEX_WANTLOCK(mtx);
                    456:
                    457: #ifdef LOCKDEBUG
                    458:        if (panicstr == NULL) {
                    459:                simple_lock_only_held(NULL, "mutex_enter");
                    460: #ifdef MULTIPROCESSOR
                    461:                LOCKDEBUG_BARRIER(&kernel_lock, 1);
                    462: #else
                    463:                LOCKDEBUG_BARRIER(NULL, 1);
                    464: #endif
                    465:        }
                    466: #endif
                    467:
                    468:        LOCKSTAT_ENTER(lsflag);
                    469:
                    470:        /*
                    471:         * Adaptive mutex; spin trying to acquire the mutex.  If we
                    472:         * determine that the owner is not running on a processor,
                    473:         * then we stop spinning, and sleep instead.
                    474:         */
                    475:        for (;;) {
                    476:                owner = mtx->mtx_owner;
                    477:                if (!MUTEX_OWNED(owner)) {
                    478:                        /*
                    479:                         * Mutex owner clear could mean two things:
                    480:                         *
                    481:                         *      * The mutex has been released.
                    482:                         *      * The owner field hasn't been set yet.
                    483:                         *
                    484:                         * Try to acquire it again.  If that fails,
                    485:                         * we'll just loop again.
                    486:                         */
                    487:                        if (MUTEX_ACQUIRE(mtx, curthread))
                    488:                                break;
                    489:                        continue;
                    490:                }
                    491:
                    492:                if (panicstr != NULL)
                    493:                        return;
                    494:                if (MUTEX_OWNER(owner) == curthread)
                    495:                        MUTEX_ABORT(mtx, "locking against myself");
                    496:
                    497: #ifdef MULTIPROCESSOR
                    498:                /*
                    499:                 * Check to see if the owner is running on a processor.
                    500:                 * If so, then we should just spin, as the owner will
                    501:                 * likely release the lock very soon.
                    502:                 */
                    503:                if (mutex_onproc(owner, &ci)) {
                    504:                        LOCKSTAT_START_TIMER(lsflag, spintime);
                    505:                        count = SPINLOCK_BACKOFF_MIN;
                    506:                        for (;;) {
                    507:                                owner = mtx->mtx_owner;
                    508:                                if (!mutex_onproc(owner, &ci))
                    509:                                        break;
                    510:                                SPINLOCK_BACKOFF(count);
                    511:                        }
                    512:                        LOCKSTAT_STOP_TIMER(lsflag, spintime);
                    513:                        LOCKSTAT_COUNT(spincnt, 1);
                    514:                        if (!MUTEX_OWNED(owner))
                    515:                                continue;
                    516:                }
                    517: #endif
                    518:
                    519:                ts = turnstile_lookup(mtx);
                    520:
                    521:                /*
                    522:                 * Once we have the turnstile chain interlock, mark the
                    523:                 * mutex has having waiters.  If that fails, spin again:
                    524:                 * chances are that the mutex has been released.
                    525:                 */
                    526:                if (!MUTEX_SET_WAITERS(mtx, owner)) {
                    527:                        turnstile_exit(mtx);
                    528:                        continue;
                    529:                }
                    530:
                    531: #ifdef MULTIPROCESSOR
                    532:                /*
                    533:                 * mutex_exit() is permitted to release the mutex without
                    534:                 * any interlocking instructions, and the following can
                    535:                 * occur as a result:
                    536:                 *
                    537:                 *  CPU 1: MUTEX_SET_WAITERS()      CPU2: mutex_exit()
                    538:                 * ---------------------------- ----------------------------
                    539:                 *              ..                  acquire cache line
                    540:                 *              ..                   test for waiters
                    541:                 *      acquire cache line    <-      lose cache line
                    542:                 *       lock cache line                   ..
                    543:                 *     verify mutex is held                ..
                    544:                 *          set waiters                    ..
                    545:                 *       unlock cache line                 ..
                    546:                 *        lose cache line     ->    acquire cache line
                    547:                 *              ..                clear lock word, waiters
                    548:                 *        return success
                    549:                 *
                    550:                 * There is a another race that can occur: a third CPU could
                    551:                 * acquire the mutex as soon as it is released.  Since
                    552:                 * adaptive mutexes are primarily spin mutexes, this is not
                    553:                 * something that we need to worry about too much.  What we
                    554:                 * do need to ensure is that the waiters bit gets set.
                    555:                 *
                    556:                 * To allow the unlocked release, we need to make some
                    557:                 * assumptions here:
                    558:                 *
                    559:                 * o Release is the only non-atomic/unlocked operation
                    560:                 *   that can be performed on the mutex.  (It must still
                    561:                 *   be atomic on the local CPU, e.g. in case interrupted
                    562:                 *   or preempted).
                    563:                 *
                    564:                 * o At any given time, MUTEX_SET_WAITERS() can only ever
                    565:                 *   be in progress on one CPU in the system - guarenteed
                    566:                 *   by the turnstile chain lock.
                    567:                 *
                    568:                 * o No other operations other than MUTEX_SET_WAITERS()
                    569:                 *   and release can modify a mutex with a non-zero
                    570:                 *   owner field.
                    571:                 *
                    572:                 * o The result of a successful MUTEX_SET_WAITERS() call
                    573:                 *   is an unbuffered write that is immediately visible
                    574:                 *   to all other processors in the system.
                    575:                 *
                    576:                 * o If the holding LWP switches away, it posts a store
                    577:                 *   fence before changing curlwp, ensuring that any
                    578:                 *   overwrite of the mutex waiters flag by mutex_exit()
                    579:                 *   completes before the modification of curlwp becomes
                    580:                 *   visible to this CPU.
                    581:                 *
                    582:                 * o cpu_switch() posts a store fence before setting curlwp
                    583:                 *   and before resuming execution of an LWP.
                    584:                 *
                    585:                 * o _kernel_lock() posts a store fence before setting
                    586:                 *   curcpu()->ci_biglock_wanted, and after clearing it.
                    587:                 *   This ensures that any overwrite of the mutex waiters
                    588:                 *   flag by mutex_exit() completes before the modification
                    589:                 *   of ci_biglock_wanted becomes visible.
                    590:                 *
                    591:                 * We now post a read memory barrier (after setting the
                    592:                 * waiters field) and check the lock holder's status again.
                    593:                 * Some of the possible outcomes (not an exhaustive list):
                    594:                 *
                    595:                 * 1. The onproc check returns true: the holding LWP is
                    596:                 *    running again.  The lock may be released soon and
                    597:                 *    we should spin.  Importantly, we can't trust the
                    598:                 *    value of the waiters flag.
                    599:                 *
                    600:                 * 2. The onproc check returns false: the holding LWP is
                    601:                 *    not running.  We now have the oppertunity to check
                    602:                 *    if mutex_exit() has blatted the modifications made
                    603:                 *    by MUTEX_SET_WAITERS().
                    604:                 *
                    605:                 * 3. The onproc check returns false: the holding LWP may
                    606:                 *    or may not be running.  It has context switched at
                    607:                 *    some point during our check.  Again, we have the
                    608:                 *    chance to see if the waiters bit is still set or
                    609:                 *    has been overwritten.
                    610:                 *
                    611:                 * 4. The onproc check returns false: the holding LWP is
                    612:                 *    running on a CPU, but wants the big lock.  It's OK
                    613:                 *    to check the waiters field in this case.
                    614:                 *
                    615:                 * 5. The has-waiters check fails: the mutex has been
                    616:                 *    released, the waiters flag cleared and another LWP
                    617:                 *    now owns the mutex.
                    618:                 *
                    619:                 * 6. The has-waiters check fails: the mutex has been
                    620:                 *    released.
                    621:                 *
                    622:                 * If the waiters bit is not set it's unsafe to go asleep,
                    623:                 * as we might never be awoken.
                    624:                 */
                    625:                mb_read();
                    626:                if (mutex_onproc(owner, &ci) || !MUTEX_HAS_WAITERS(mtx)) {
                    627:                        turnstile_exit(mtx);
                    628:                        continue;
                    629:                }
                    630: #endif /* MULTIPROCESSOR */
                    631:
                    632:                LOCKSTAT_START_TIMER(lsflag, slptime);
                    633:
                    634:                turnstile_block(ts, TS_WRITER_Q, mtx);
                    635:
                    636:                LOCKSTAT_STOP_TIMER(lsflag, slptime);
                    637:                LOCKSTAT_COUNT(slpcnt, 1);
                    638:
                    639:                turnstile_unblock();
                    640:        }
                    641:
                    642:        LOCKSTAT_EVENT(lsflag, mtx, LB_ADAPTIVE_MUTEX | LB_SLEEP1,
                    643:            slpcnt, slptime);
                    644:        LOCKSTAT_EVENT(lsflag, mtx, LB_ADAPTIVE_MUTEX | LB_SPIN,
                    645:            spincnt, spintime);
                    646:        LOCKSTAT_EXIT(lsflag);
                    647:
                    648:        MUTEX_DASSERT(mtx, MUTEX_OWNER(mtx->mtx_owner) == curthread);
                    649:        MUTEX_LOCKED(mtx);
                    650: }
                    651:
                    652: /*
                    653:  * mutex_vector_exit:
                    654:  *
                    655:  *     Support routine for mutex_exit() that handles all cases.
                    656:  */
                    657: void
                    658: mutex_vector_exit(kmutex_t *mtx)
                    659: {
                    660:        turnstile_t *ts;
                    661:        uintptr_t curthread;
                    662:
                    663:        if (MUTEX_SPIN_P(mtx)) {
                    664: #ifdef FULL
                    665:                if (mtx->mtx_lock != __SIMPLELOCK_LOCKED)
                    666:                        MUTEX_ABORT(mtx, "exiting unheld spin mutex");
                    667:                MUTEX_UNLOCKED(mtx);
                    668:                __cpu_simple_unlock(&mtx->mtx_lock);
                    669: #endif
                    670:                MUTEX_SPIN_SPLRESTORE(mtx);
                    671:                return;
                    672:        }
                    673:
                    674:        if (__predict_false(panicstr != NULL) || __predict_false(cold)) {
                    675:                MUTEX_UNLOCKED(mtx);
                    676:                MUTEX_RELEASE(mtx);
                    677:                return;
                    678:        }
                    679:
                    680:        curthread = (uintptr_t)curlwp;
                    681:        MUTEX_DASSERT(mtx, curthread != 0);
                    682:        MUTEX_ASSERT(mtx, MUTEX_OWNER(mtx->mtx_owner) == curthread);
                    683:        MUTEX_UNLOCKED(mtx);
                    684:
                    685:        /*
                    686:         * Get this lock's turnstile.  This gets the interlock on
                    687:         * the sleep queue.  Once we have that, we can clear the
                    688:         * lock.  If there was no turnstile for the lock, there
                    689:         * were no waiters remaining.
                    690:         */
                    691:        ts = turnstile_lookup(mtx);
                    692:
                    693:        if (ts == NULL) {
                    694:                MUTEX_RELEASE(mtx);
                    695:                turnstile_exit(mtx);
                    696:        } else {
                    697:                MUTEX_RELEASE(mtx);
                    698:                turnstile_wakeup(ts, TS_WRITER_Q,
                    699:                    TS_WAITERS(ts, TS_WRITER_Q), NULL);
                    700:        }
                    701: }
                    702:
                    703: /*
                    704:  * mutex_owned:
                    705:  *
1.3     ! ad        706:  *     Return true if the current LWP (adaptive) or CPU (spin)
        !           707:  *     holds the mutex.
1.2       ad        708:  */
                    709: int
                    710: mutex_owned(kmutex_t *mtx)
                    711: {
                    712:
                    713:        if (MUTEX_ADAPTIVE_P(mtx))
                    714:                return MUTEX_OWNER(mtx->mtx_owner) == (uintptr_t)curlwp;
                    715: #ifdef FULL
                    716:        return mtx->mtx_lock == __SIMPLELOCK_LOCKED;
                    717: #else
                    718:        return 1;
                    719: #endif
                    720: }
                    721:
                    722: /*
                    723:  * mutex_owner:
                    724:  *
                    725:  *     Return the current owner of an adaptive mutex.
                    726:  */
                    727: struct lwp *
                    728: mutex_owner(kmutex_t *mtx)
                    729: {
                    730:
                    731:        MUTEX_ASSERT(mtx, MUTEX_ADAPTIVE_P(mtx));
                    732:        return (struct lwp *)MUTEX_OWNER(mtx->mtx_owner);
                    733: }
                    734:
                    735: /*
                    736:  * mutex_tryenter:
                    737:  *
                    738:  *     Try to acquire the mutex; return non-zero if we did.
                    739:  */
                    740: int
                    741: mutex_tryenter(kmutex_t *mtx)
                    742: {
                    743:        uintptr_t curthread;
                    744:
                    745:        MUTEX_WANTLOCK(mtx);
                    746:
                    747:        /*
                    748:         * Handle spin mutexes.
                    749:         */
                    750:        if (MUTEX_SPIN_P(mtx)) {
                    751:                MUTEX_SPIN_SPLRAISE(mtx);
                    752: #ifdef FULL
                    753:                if (__cpu_simple_lock_try(&mtx->mtx_lock)) {
                    754:                        MUTEX_LOCKED(mtx);
                    755:                        return 1;
                    756:                }
                    757:                MUTEX_SPIN_SPLRESTORE(mtx);
                    758: #else
                    759:                MUTEX_LOCKED(mtx);
                    760:                return 1;
                    761: #endif
                    762:        } else {
                    763:                curthread = (uintptr_t)curlwp;
                    764:                MUTEX_ASSERT(mtx, curthread != 0);
                    765:                if (MUTEX_ACQUIRE(mtx, curthread)) {
                    766:                        MUTEX_LOCKED(mtx);
                    767:                        MUTEX_DASSERT(mtx,
                    768:                            MUTEX_OWNER(mtx->mtx_owner) == curthread);
                    769:                        return 1;
                    770:                }
                    771:        }
                    772:
                    773:        return 0;
                    774: }
                    775:
                    776: #if defined(__HAVE_SPIN_MUTEX_STUBS) || defined(FULL)
                    777: /*
                    778:  * mutex_spin_retry:
                    779:  *
                    780:  *     Support routine for mutex_spin_enter().  Assumes that the caller
                    781:  *     has already raised the SPL, and adjusted counters.
                    782:  */
                    783: void
                    784: mutex_spin_retry(kmutex_t *mtx)
                    785: {
                    786: #ifdef MULTIPROCESSOR
                    787:        u_int count;
                    788:        LOCKSTAT_TIMER(spintime);
                    789:        LOCKSTAT_FLAG(lsflag);
                    790: #ifdef LOCKDEBUG
                    791:        u_int spins = 0;
                    792: #endif /* LOCKDEBUG */
                    793:
                    794:        MUTEX_WANTLOCK(mtx);
                    795:
                    796:        LOCKSTAT_ENTER(lsflag);
                    797:        LOCKSTAT_START_TIMER(lsflag, spintime);
                    798:        count = SPINLOCK_BACKOFF_MIN;
                    799:
                    800:        /*
                    801:         * Spin testing the lock word and do exponential backoff
                    802:         * to reduce cache line ping-ponging between CPUs.
                    803:         */
                    804:        do {
                    805:                if (panicstr != NULL)
                    806:                        break;
                    807:                while (mtx->mtx_lock == __SIMPLELOCK_LOCKED) {
                    808:                        SPINLOCK_BACKOFF(count);
                    809: #ifdef LOCKDEBUG
                    810:                        if (SPINLOCK_SPINOUT(spins))
                    811:                                MUTEX_ABORT(mtx, "spinout");
                    812: #endif /* LOCKDEBUG */
                    813:                }
                    814:        } while (!__cpu_simple_lock_try(&mtx->mtx_lock));
                    815:
                    816:        LOCKSTAT_STOP_TIMER(lsflag, spintime);
                    817:        LOCKSTAT_EVENT(lsflag, mtx, LB_SPIN_MUTEX | LB_SPIN, 1, spintime);
                    818:        LOCKSTAT_EXIT(lsflag);
                    819:
                    820:        MUTEX_LOCKED(mtx);
                    821: #else  /* MULTIPROCESSOR */
                    822:        MUTEX_ABORT(mtx, "locking against myself");
                    823: #endif /* MULTIPROCESSOR */
                    824: }
                    825: #endif /* defined(__HAVE_SPIN_MUTEX_STUBS) || defined(FULL) */
                    826:
                    827: /*
                    828:  * sched_lock_idle:
                    829:  *
                    830:  *     XXX Ugly hack for cpu_switch().
                    831:  */
                    832: void
                    833: sched_lock_idle(void)
                    834: {
                    835: #ifdef FULL
                    836:        kmutex_t *mtx = &sched_mutex;
                    837:
                    838:        curcpu()->ci_mtx_count--;
                    839:
                    840:        if (!__cpu_simple_lock_try(&mtx->mtx_lock)) {
                    841:                mutex_spin_retry(mtx);
                    842:                return;
                    843:        }
                    844:
                    845:        MUTEX_LOCKED(mtx);
                    846: #else
                    847:        curcpu()->ci_mtx_count--;
                    848: #endif /* FULL */
                    849: }
                    850:
                    851: /*
                    852:  * sched_unlock_idle:
                    853:  *
                    854:  *     XXX Ugly hack for cpu_switch().
                    855:  */
                    856: void
                    857: sched_unlock_idle(void)
                    858: {
                    859: #ifdef FULL
                    860:        kmutex_t *mtx = &sched_mutex;
                    861:
                    862:        if (mtx->mtx_lock != __SIMPLELOCK_LOCKED)
                    863:                MUTEX_ABORT(mtx, "sched_unlock_idle");
                    864:
                    865:        MUTEX_UNLOCKED(mtx);
                    866:        __cpu_simple_unlock(&mtx->mtx_lock);
                    867: #endif /* FULL */
                    868:        curcpu()->ci_mtx_count++;
                    869: }

CVSweb <webmaster@jp.NetBSD.org>