mutex.h

File system using stacked.

typedef unsigned char tsl_t;

#ifdef LOAD_ACTUAL_MUTEX_CODE
/*
 * For gcc/ia64, 0 is clear, 1 is set.
 */
#define	MUTEX_SET(tsl) ({						\
	register tsl_t *__l = (tsl);					\
	long __r;							\
	asm volatile("xchg1 %0=%1,%3" : "=r"(__r), "=m"(*__l) : "1"(*__l), "r"(1));\
	__r ^ 1;							\
})

/*
 * Store through a "volatile" pointer so we get a store with "release"
 * semantics.
 */
#define	MUTEX_UNSET(tsl)	(*(volatile unsigned char *)(tsl) = 0)
#define	MUTEX_INIT(tsl)		MUTEX_UNSET(tsl)
#endif
#endif

/*********************************************************************
 * PowerPC/gcc assembly.
 *********************************************************************/
#if defined(HAVE_MUTEX_PPC_GENERIC_GCC_ASSEMBLY) ||			\
    (HAVE_MUTEX_PPC_APPLE_GCC_ASSEMBLY)
typedef u_int32_t tsl_t;

#ifdef LOAD_ACTUAL_MUTEX_CODE
/*
 * The PowerPC does a sort of pseudo-atomic locking.  You set up a
 * 'reservation' on a chunk of memory containing a mutex by loading the
 * mutex value with LWARX.  If the mutex has an 'unlocked' (arbitrary)
 * value, you then try storing into it with STWCX.  If no other process or
 * thread broke your 'reservation' by modifying the memory containing the
 * mutex, then the STCWX succeeds; otherwise it fails and you try to get
 * a reservation again.
 *
 * While mutexes are explicitly 4 bytes, a 'reservation' applies to an
 * entire cache line, normally 32 bytes, aligned naturally.  If the mutex
 * lives near data that gets changed a lot, there's a chance that you'll
 * see more broken reservations than you might otherwise.  The only
 * situation in which this might be a problem is if one processor is
 * beating on a variable in the same cache block as the mutex while another
 * processor tries to acquire the mutex.  That's bad news regardless
 * because of the way it bashes caches, but if you can't guarantee that a
 * mutex will reside in a relatively quiescent cache line, you might
 * consider padding the mutex to force it to live in a cache line by
 * itself.  No, you aren't guaranteed that cache lines are 32 bytes.  Some
 * embedded processors use 16-byte cache lines, while some 64-bit
 * processors use 128-bit cache lines.  But assuming a 32-byte cache line
 * won't get you into trouble for now.
 *
 * If mutex locking is a bottleneck, then you can speed it up by adding a
 * regular LWZ load before the LWARX load, so that you can test for the
 * common case of a locked mutex without wasting cycles making a reservation.
 *
 * 'set' mutexes have the value 1, like on Intel; the returned value from
 * MUTEX_SET() is 1 if the mutex previously had its low bit clear, 0 otherwise.
 *
 * Mutexes on Mac OS X work the same way as the standard PowerPC version, but
 * the assembler syntax is subtly different -- the standard PowerPC version
 * assembles but doesn't work correctly.  This version makes (unnecessary?)
 * use of a stupid linker trick: __db_mutex_tas_dummy is never called, but the
 * ___db_mutex_set label is used as a function name.
 */
#ifdef HAVE_MUTEX_PPC_APPLE_GCC_ASSEMBLY
extern int __db_mutex_set __P((volatile tsl_t *));
void
__db_mutex_tas_dummy()
{
	__asm__	__volatile__("		\n\
	.globl 	___db_mutex_set		\n\
___db_mutex_set:			\n\
	lwarx	r5,0,r3			\n\
	cmpwi 	r5,0			\n\
	bne 	fail			\n\
	addi 	r5,r5,1			\n\
	stwcx. 	r5,0,r3			\n\
	beq 	success			\n\
fail:					\n\
	li 	r3,0			\n\
	blr 				\n\
success:				\n\
	li 	r3,1			\n\
	blr");	
}
#define	MUTEX_SET(tsl)  __db_mutex_set(tsl)
#endif
#ifdef HAVE_MUTEX_PPC_GENERIC_GCC_ASSEMBLY
#define	MUTEX_SET(tsl)	({		\
	int __one = 1;			\
	int __r;			\
	tsl_t *__l = (tsl);		\
	asm volatile ("			\
0:					\
	lwarx %0,0,%1;			\
	cmpwi %0,0;			\
	bne 1f;				\
	stwcx. %2,0,%1;			\
	bne- 0b;			\
1:"					\
	: "=&r" (__r)			\
	: "r" (__l), "r" (__one));	\
	!(__r & 1);			\
})
#endif
#define	MUTEX_UNSET(tsl)	(*(tsl) = 0)
#define	MUTEX_INIT(tsl)		MUTEX_UNSET(tsl)
#endif
#endif

/*********************************************************************
 * S/390 32-bit assembly.
 *********************************************************************/
#ifdef HAVE_MUTEX_S390_GCC_ASSEMBLY
typedef int tsl_t;

#ifdef LOAD_ACTUAL_MUTEX_CODE
/*
 * For gcc/S390, 0 is clear, 1 is set.
 */
static inline int
MUTEX_SET(tsl_t *tsl) {							\
	register tsl_t *__l = (tsl);					\
	int __r;							\
  asm volatile(								\
       "    la    1,%1\n"						\
       "    lhi   0,1\n"						\
       "    l     %0,%1\n"						\
       "0:  cs    %0,0,0(1)\n"						\
       "    jl    0b"							\
       : "=&d" (__r), "+m" (*__l)					\
       : : "0", "1", "cc");						\
	return !__r;							\
}

#define	MUTEX_UNSET(tsl)	(*(tsl) = 0)
#define	MUTEX_INIT(tsl)		MUTEX_UNSET(tsl)
#endif
#endif

/*********************************************************************
 * SCO/cc assembly.
 *********************************************************************/
#ifdef HAVE_MUTEX_SCO_X86_CC_ASSEMBLY
typedef unsigned char tsl_t;

#ifdef LOAD_ACTUAL_MUTEX_CODE
/*
 * UnixWare has threads in libthread, but OpenServer doesn't (yet).
 *
 * For cc/x86, 0 is clear, 1 is set.
 */

#if defined(__USLC__)
asm int
_tsl_set(void *tsl)
{
%mem tsl
	movl	tsl, %ecx
	movl	$1, %eax
	lock
	xchgb	(%ecx),%al
	xorl	$1,%eax
}
#endif

#define	MUTEX_SET(tsl)		_tsl_set(tsl)
#define	MUTEX_UNSET(tsl)	(*(tsl) = 0)
#define	MUTEX_INIT(tsl)		MUTEX_UNSET(tsl)
#endif
#endif

/*********************************************************************
 * Sparc/gcc assembly.
 *********************************************************************/
#ifdef HAVE_MUTEX_SPARC_GCC_ASSEMBLY
typedef unsigned char tsl_t;

#ifdef LOAD_ACTUAL_MUTEX_CODE
/*
 *
 * The ldstub instruction takes the location specified by its first argument
 * (a register containing a memory address) and loads its contents into its
 * second argument (a register) and atomically sets the contents the location
 * specified by its first argument to a byte of 1s.  (The value in the second
 * argument is never read, but only overwritten.)
 *
 * The stbar is needed for v8, and is implemented as membar #sync on v9,
 * so is functional there as well.  For v7, stbar may generate an illegal
 * instruction and we have no way to tell what we're running on.  Some
 * operating systems notice and skip this instruction in the fault handler.
 *
 * For gcc/sparc, 0 is clear, 1 is set.
 */
#define	MUTEX_SET(tsl) ({						\
	register tsl_t *__l = (tsl);					\
	register tsl_t __r;						\
	__asm__ volatile						\
	    ("ldstub [%1],%0; stbar"					\
	    : "=r"( __r) : "r" (__l));					\
	!__r;								\
})

#define	MUTEX_UNSET(tsl)	(*(tsl) = 0)
#define	MUTEX_INIT(tsl)		MUTEX_UNSET(tsl)
#endif
#endif

/*********************************************************************
 * UTS/cc assembly.
 *********************************************************************/
#ifdef HAVE_MUTEX_UTS_CC_ASSEMBLY
typedef int tsl_t;

#define	MUTEX_ALIGN	sizeof(int)
#ifdef LOAD_ACTUAL_MUTEX_CODE
#define	MUTEX_INIT(x)	0
#define	MUTEX_SET(x)	(!uts_lock(x, 1))
#define	MUTEX_UNSET(x)	(*(x) = 0)
#endif
#endif




#endif  /* __KERNEL__ */






/*********************************************************************
 * x86/gcc assembly.
 *********************************************************************/
#ifdef HAVE_MUTEX_X86_GCC_ASSEMBLY
typedef unsigned char tsl_t;

#ifdef LOAD_ACTUAL_MUTEX_CODE
/*
 * For gcc/x86, 0 is clear, 1 is set.
 */
#define	MUTEX_SET(tsl) ({						\
	register tsl_t *__l = (tsl);					\
	int __r;							\
	asm volatile("movl $1,%%eax; lock; xchgb %1,%%al; xorl $1,%%eax"\
	    : "=&a" (__r), "=m" (*__l)					\
	    : "1" (*__l)						\
	    );								\
	__r & 1;							\
})

#define	MUTEX_UNSET(tsl)	(*(tsl) = 0)
#define	MUTEX_INIT(tsl)		MUTEX_UNSET(tsl)
#endif
#endif

/*
 * Mutex alignment defaults to one byte.
 *
 * !!!
 * Various systems require different alignments for mutexes (the worst we've
 * seen so far is 16-bytes on some HP architectures).  Malloc(3) is assumed
 * to return reasonable alignment, all other mutex users must ensure proper
 * alignment locally.
 */
#ifndef	MUTEX_ALIGN
#define	MUTEX_ALIGN	1
#endif

/*
 * Mutex destruction defaults to a no-op.
 */
#ifdef LOAD_ACTUAL_MUTEX_CODE
#ifndef	MUTEX_DESTROY
#define	MUTEX_DESTROY(x)
#endif
#endif

/*
 * !!!
 * These defines are separated into the u_int8_t flags stored in the
 * mutex below, and the 32 bit flags passed to __db_mutex_setup.
 * But they must co-exist and not overlap.  Flags to __db_mutex_setup are:
 *
 * MUTEX_ALLOC - Use when the mutex to initialize needs to be allocated.
 *    The 'ptr' arg to __db_mutex_setup should be a DB_MUTEX ** whenever
 *    you use this flag.  If this flag is not set, the 'ptr' arg is
 *    a DB_MUTEX *.
 * MUTEX_NO_RECORD - Explicitly do not record the mutex in the region.
 *    Otherwise the mutex will be recorded by default.  If you set
 *    this you need to understand why you don't need it recorded.  The
 *    *only* ones not recorded are those that are part of region structures
 *    that only get destroyed when the regions are destroyed.
 * MUTEX_NO_RLOCK - Explicitly do not lock the given region otherwise
 *    the region will be locked by default.
 * MUTEX_SELF_BLOCK - Set if self blocking mutex.
 * MUTEX_THREAD - Set if mutex is a thread-only mutex.
 */
#define	MUTEX_IGNORE		0x001	/* Ignore, no lock required. */
#define	MUTEX_INITED		0x002	/* Mutex is successfully initialized */
#define	MUTEX_MPOOL		0x004	/* Allocated from mpool. */
#define	MUTEX_SELF_BLOCK	0x008	/* Must block self. */
/* Flags only, may be larger than 0xff. */
#define	MUTEX_ALLOC		0x00000100 /* Allocate and init a mutex */
#define	MUTEX_NO_RECORD		0x00000200 /* Do not record lock */
#define	MUTEX_NO_RLOCK		0x00000400 /* Do not acquire region lock */
#define	MUTEX_THREAD		0x00000800 /* Thread-only mutex. */

/* Mutex. */
struct __mutex_t {
#ifndef __KERNEL__
#ifdef	HAVE_MUTEX_THREADS
#ifdef	MUTEX_FIELDS
	MUTEX_FIELDS
#else
	tsl_t	tas;			/* Test and set. */

#endif
	u_int32_t spins;		/* Spins before block. */
	u_int32_t locked;		/* !0 if locked. */
#else
	u_int32_t off;			/* Byte offset to lock. */
	u_int32_t pid;			/* Lock holder: 0 or process pid. */
#endif
	u_int32_t mutex_set_wait;	/* Granted after wait. */
	u_int32_t mutex_set_nowait;	/* Granted without waiting. */
	u_int32_t mutex_set_spin;	/* Granted without spinning. */
	u_int32_t mutex_set_spins;	/* Total number of spins. */
#ifdef HAVE_MUTEX_SYSTEM_RESOURCES
	roff_t	  reg_off;		/* Shared lock info offset. */
#endif
#else
#ifdef	MUTEX_FIELDS
	MUTEX_FIELDS
	u_int32_t spins;		/* Spins before block. */
	u_int32_t locked;		/* !0 if locked. */
	u_int32_t mutex_set_wait;	/* Granted after wait. */
	u_int32_t mutex_set_nowait;	/* Granted without waiting. */
#endif

#include <asm/semaphore.h>
	struct semaphore tas;		/* Mutex. */	

#endif /* __KERNEL__ */

	u_int8_t  flags;		/* MUTEX_XXX */
};

/* Redirect calls to the correct functions. */
#ifdef __KERNEL__

#define	__db_mutex_init_int(a, b, c, d)	__db_tas_mutex_init(a, b, d)
#define	__db_mutex_lock(a, b)		__db_tas_mutex_lock(a, b)
#define	__db_mutex_unlock(a, b)		__db_tas_mutex_unlock(a, b)
#define	__db_mutex_destroy(a)		__db_tas_mutex_destroy(a)

#else

#ifdef HAVE_MUTEX_THREADS
#if defined(HAVE_MUTEX_PTHREADS) ||					\
    defined(HAVE_MUTEX_SOLARIS_LWP) ||					\
    defined(HAVE_MUTEX_UI_THREADS)
#define	__db_mutex_init_int(a, b, c, d)	__db_pthread_mutex_init(a, b, d)
#define	__db_mutex_lock(a, b)		__db_pthread_mutex_lock(a, b)
#define	__db_mutex_unlock(a, b)		__db_pthread_mutex_unlock(a, b)
#define	__db_mutex_destroy(a)		__db_pthread_mutex_destroy(a)
#elif defined(HAVE_MUTEX_WIN32)
#define	__db_mutex_init_int(a, b, c, d)	__db_win32_mutex_init(a, b, d)
#define	__db_mutex_lock(a, b)		__db_win32_mutex_lock(a, b)
#define	__db_mutex_unlock(a, b)		__db_win32_mutex_unlock(a, b)
#define	__db_mutex_destroy(a)		__db_win32_mutex_destroy(a)
#else
#define	__db_mutex_init_int(a, b, c, d)	__db_tas_mutex_init(a, b, d)
#define	__db_mutex_lock(a, b)		__db_tas_mutex_lock(a, b)
#define	__db_mutex_unlock(a, b)		__db_tas_mutex_unlock(a, b)
#define	__db_mutex_destroy(a)		__db_tas_mutex_destroy(a)
#endif
#else
#define	__db_mutex_init_int(a, b, c, d)	__db_fcntl_mutex_init(a, b, c)
#define	__db_mutex_lock(a, b)		__db_fcntl_mutex_lock(a, b)
#define	__db_mutex_unlock(a, b)		__db_fcntl_mutex_unlock(a, b)
#define	__db_mutex_destroy(a)		__db_fcntl_mutex_destroy(a)
#endif

#endif /* __KERNEL__ */

/* Redirect system resource calls to correct functions */
#ifdef HAVE_MUTEX_SYSTEM_RESOURCES
#define	__db_maintinit(a, b, c)		__db_shreg_maintinit(a, b, c)
#define	__db_shlocks_clear(a, b, c)	__db_shreg_locks_clear(a, b, c)
#define	__db_shlocks_destroy(a, b)	__db_shreg_locks_destroy(a, b)
#define	__db_mutex_init(a, b, c, d, e, f)	\
    __db_shreg_mutex_init(a, b, c, d, e, f)
#else
#define	__db_maintinit(a, b, c)
#define	__db_shlocks_clear(a, b, c)
#define	__db_shlocks_destroy(a, b)
#define	__db_mutex_init(a, b, c, d, e, f)	__db_mutex_init_int(a, b, c, d)
#endif

/*
 * Lock/unlock a mutex.  If the mutex was marked as uninteresting, the thread
 * of control can proceed without it.
 *
 * If the lock is for threads-only, then it was optionally not allocated and
 * file handles aren't necessary, as threaded applications aren't supported by
 * fcntl(2) locking.
 */
#ifdef DIAGNOSTIC
	/*
	 * XXX
	 * We want to switch threads as often as possible.  Yield every time
	 * we get a mutex to ensure contention.
	 */
#define	MUTEX_LOCK(dbenv, mp)						\
	if (!F_ISSET((mp), MUTEX_IGNORE))				\
		DB_ASSERT(__db_mutex_lock(dbenv, mp) == 0);		\
	if (F_ISSET(dbenv, DB_ENV_YIELDCPU))				\
		__os_yield(NULL, 1);
#else
#define	MUTEX_LOCK(dbenv, mp)						\
	if (!F_ISSET((mp), MUTEX_IGNORE))				\
		(void)__db_mutex_lock(dbenv, mp);
#endif
#define	MUTEX_UNLOCK(dbenv, mp)						\
	if (!F_ISSET((mp), MUTEX_IGNORE))				\
		(void)__db_mutex_unlock(dbenv, mp);
#define	MUTEX_THREAD_LOCK(dbenv, mp)					\
	if (mp != NULL)							\
		MUTEX_LOCK(dbenv, mp)
#define	MUTEX_THREAD_UNLOCK(dbenv, mp)					\
	if (mp != NULL)							\
		MUTEX_UNLOCK(dbenv, mp)

/*
 * We use a single file descriptor for fcntl(2) locking, and (generally) the
 * object's offset in a shared region as the byte that we're locking.  So,
 * there's a (remote) possibility that two objects might have the same offsets
 * such that the locks could conflict, resulting in deadlock.  To avoid this
 * possibility, we offset the region offset by a small integer value, using a
 * different offset for each subsystem's locks.  Since all region objects are
 * suitably aligned, the offset guarantees that we don't collide with another
 * region's objects.
 */
#define	DB_FCNTL_OFF_GEN	0		/* Everything else. */
#define	DB_FCNTL_OFF_LOCK	1		/* Lock subsystem offset. */
#define	DB_FCNTL_OFF_MPOOL	2		/* Mpool subsystem offset. */

#ifdef HAVE_MUTEX_SYSTEM_RESOURCES
/*
 * When the underlying mutexes require library (most likely heap) or system
 * resources, we have to clean up when we discard mutexes (for the library
 * resources) and both when discarding mutexes and after application failure
 * (for the mutexes requiring system resources).  This violates the rule that
 * we never look at a shared region after application failure, but we've no
 * other choice.  In those cases, the #define HAVE_MUTEX_SYSTEM_RESOURCES is
 * set.
 *
 * To support mutex release after application failure, allocate thread-handle
 * mutexes in shared memory instead of in the heap.  The number of slots we
 * allocate for this purpose isn't configurable, but this tends to be an issue
 * only on embedded systems where we don't expect large server applications.
 */
#define	DB_MAX_HANDLES	100			/* Mutex slots for handles. */
#endif
#endif /* !_DB_MUTEX_H_ */