mariadb/storage/xtradb/include/os0sync.h

/*****************************************************************************

Copyright (c) 1995, 2016, Oracle and/or its affiliates. All Rights Reserved.
Copyright (c) 2008, Google Inc.
Copyright (c) 2017, MariaDB Corporation.

Portions of this file contain modifications contributed and copyrighted by
Google, Inc. Those modifications are gratefully acknowledged and are described
briefly in the InnoDB documentation. The contributions by Google are
incorporated with their permission, and subject to the conditions contained in
the file COPYING.Google.

This program is free software; you can redistribute it and/or modify it under
the terms of the GNU General Public License as published by the Free Software
Foundation; version 2 of the License.

This program is distributed in the hope that it will be useful, but WITHOUT
ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS
FOR A PARTICULAR PURPOSE. See the GNU General Public License for more details.

You should have received a copy of the GNU General Public License along with
this program; if not, write to the Free Software Foundation, Inc.,
51 Franklin Street, Suite 500, Boston, MA 02110-1335 USA

*****************************************************************************/

/**************************************************//**
@file include/os0sync.h
The interface to the operating system
synchronization primitives.

Created 9/6/1995 Heikki Tuuri
*******************************************************/

#ifndef os0sync_h
#define os0sync_h

#include "univ.i"
#include "ut0lst.h"
#include "sync0types.h"

#ifdef CPU_LEVEL1_DCACHE_LINESIZE
# define CACHE_LINE_SIZE              CPU_LEVEL1_DCACHE_LINESIZE
#else
# error CPU_LEVEL1_DCACHE_LINESIZE is undefined
#endif /* CPU_LEVEL1_DCACHE_LINESIZE */

#ifdef HAVE_WINDOWS_ATOMICS
typedef LONG lock_word_t;	/*!< On Windows, InterlockedExchange operates
				on LONG variable */
#elif defined(HAVE_ATOMIC_BUILTINS) && !defined(HAVE_ATOMIC_BUILTINS_BYTE)
typedef ulint lock_word_t;
#else

#define IB_LOCK_WORD_IS_BYTE

typedef byte lock_word_t;

#endif /* HAVE_WINDOWS_ATOMICS */

#ifdef __WIN__
/** Native event (slow)*/
typedef HANDLE			os_native_event_t;
/** Native mutex */
typedef CRITICAL_SECTION	fast_mutex_t;
/** Native condition variable. */
typedef CONDITION_VARIABLE	os_cond_t;
#else
/** Native mutex */
typedef pthread_mutex_t		fast_mutex_t;
/** Native condition variable */
typedef pthread_cond_t		os_cond_t;
#endif

/** Structure that includes Performance Schema Probe pfs_psi
in the os_fast_mutex structure if UNIV_PFS_MUTEX is defined */
struct os_fast_mutex_t {
	fast_mutex_t		mutex;	/*!< os_fast_mutex */
#ifdef UNIV_PFS_MUTEX
	struct PSI_mutex*	pfs_psi;/*!< The performance schema
					instrumentation hook */
#endif
};

/** Operating system event handle */
typedef struct os_event*	os_event_t;

/** An asynchronous signal sent between threads */
struct os_event {
#ifdef __WIN__
	HANDLE		handle;		/*!< kernel event object, slow,
					used on older Windows */
#endif
	os_fast_mutex_t	os_mutex;	/*!< this mutex protects the next
					fields */
private:
	/** Masks for the event signal count and set flag in the count_and_set
	field */
	static const ib_uint64_t count_mask = 0x7fffffffffffffffULL;
	static const ib_uint64_t set_mask   = 0x8000000000000000ULL;

	/** The MSB is set whenever when the event is in the signaled state,
	i.e. a thread does not stop if it tries to wait for this event. Lower
	bits are incremented each time the event becomes signaled. */
	ib_uint64_t	count_and_set;
public:
	os_cond_t	cond_var;	/*!< condition variable is used in
					waiting for the event */

	/** Initialise count_and_set field */
	void init_count_and_set(void)
	{
		/* We return this value in os_event_reset(), which can then be
		be used to pass to the os_event_wait_low(). The value of zero
		is reserved in os_event_wait_low() for the case when the
		caller does not want to pass any signal_count value. To
		distinguish between the two cases we initialize signal_count
		to 1 here. */
		count_and_set = 1;
	}

	/** Mark this event as set */
	void set(void)
	{
		count_and_set |= set_mask;
	}

	/** Unmark this event as set */
	void reset(void)
	{
		count_and_set &= count_mask;
	}

	/** Return true if this event is set */
	bool is_set(void) const
	{
		return count_and_set & set_mask;
	}

	/** Bump signal count for this event */
	void inc_signal_count(void)
	{
		ut_ad(static_cast<ib_uint64_t>(signal_count()) < count_mask);
		count_and_set++;
	}

	/** Return how many times this event has been signalled */
	ib_int64_t signal_count(void) const
	{
		return (count_and_set & count_mask);
	}
};

/** Denotes an infinite delay for os_event_wait_time() */
#define OS_SYNC_INFINITE_TIME   ULINT_UNDEFINED

/** Return value of os_event_wait_time() when the time is exceeded */
#define OS_SYNC_TIME_EXCEEDED   1

/** Operating system mutex handle */
typedef struct os_mutex_t*	os_ib_mutex_t;

// All the os_*_count variables are accessed atomically

/** This is incremented by 1 in os_thread_create and decremented by 1 in
os_thread_exit */
extern ulint		os_thread_count;

extern ulint		os_event_count;
extern ulint		os_mutex_count;
extern ulint		os_fast_mutex_count;

/*********************************************************//**
Initializes global event and OS 'slow' mutex lists. */
UNIV_INTERN
void
os_sync_init(void);
/*==============*/

/** Create an event semaphore, i.e., a semaphore which may just have two
states: signaled and nonsignaled. The created event is manual reset: it must be
reset explicitly by calling sync_os_reset_event.
@param[in,out]	event	memory block where to create the event */
UNIV_INTERN
void
os_event_create(os_event_t event);

/*********************************************************//**
Creates an event semaphore, i.e., a semaphore which may just have two states:
signaled and nonsignaled. The created event is manual reset: it must be reset
explicitly by calling sync_os_reset_event.
@return	the event handle */
UNIV_INTERN
os_event_t
os_event_create(void);
/*==================*/
/**********************************************************//**
Sets an event semaphore to the signaled state: lets waiting threads
proceed. */
UNIV_INTERN
void
os_event_set(
/*=========*/
	os_event_t	event);	/*!< in: event to set */
/**********************************************************//**
Resets an event semaphore to the nonsignaled state. Waiting threads will
stop to wait for the event.
The return value should be passed to os_even_wait_low() if it is desired
that this thread should not wait in case of an intervening call to
os_event_set() between this os_event_reset() and the
os_event_wait_low() call. See comments for os_event_wait_low(). */
UNIV_INTERN
ib_int64_t
os_event_reset(
/*===========*/
	os_event_t	event);	/*!< in: event to reset */
/**********************************************************//**
Frees an event object. */
UNIV_INTERN
void
os_event_free(
/*==========*/
	os_event_t	event,	/*!< in: event to free */
	bool		free_memory = true);
				/*!< in: if true, deallocate the event memory
				block too */

/**********************************************************//**
Waits for an event object until it is in the signaled state.

Typically, if the event has been signalled after the os_event_reset()
we'll return immediately because event->is_set == TRUE.
There are, however, situations (e.g.: sync_array code) where we may
lose this information. For example:

thread A calls os_event_reset()
thread B calls os_event_set()   [event->is_set == TRUE]
thread C calls os_event_reset() [event->is_set == FALSE]
thread A calls os_event_wait()  [infinite wait!]
thread C calls os_event_wait()  [infinite wait!]

Where such a scenario is possible, to avoid infinite wait, the
value returned by os_event_reset() should be passed in as
reset_sig_count. */
UNIV_INTERN
void
os_event_wait_low(
/*==============*/
	os_event_t	event,		/*!< in: event to wait */
	ib_int64_t	reset_sig_count);/*!< in: zero or the value
					returned by previous call of
					os_event_reset(). */

#define os_event_wait(event) os_event_wait_low(event, 0)
#define os_event_wait_time(event, t) os_event_wait_time_low(event, t, 0)

/**********************************************************//**
Waits for an event object until it is in the signaled state or
a timeout is exceeded.
@return	0 if success, OS_SYNC_TIME_EXCEEDED if timeout was exceeded */
UNIV_INTERN
ulint
os_event_wait_time_low(
/*===================*/
	os_event_t	event,			/*!< in: event to wait */
	ulint		time_in_usec,		/*!< in: timeout in
						microseconds, or
						OS_SYNC_INFINITE_TIME */
	ib_int64_t	reset_sig_count);	/*!< in: zero or the value
						returned by previous call of
						os_event_reset(). */
/*********************************************************//**
Creates an operating system mutex semaphore. Because these are slow, the
mutex semaphore of InnoDB itself (ib_mutex_t) should be used where possible.
@return	the mutex handle */
UNIV_INTERN
os_ib_mutex_t
os_mutex_create(void);
/*=================*/
/**********************************************************//**
Acquires ownership of a mutex semaphore. */
UNIV_INTERN
void
os_mutex_enter(
/*===========*/
	os_ib_mutex_t	mutex);	/*!< in: mutex to acquire */
/**********************************************************//**
Releases ownership of a mutex. */
UNIV_INTERN
void
os_mutex_exit(
/*==========*/
	os_ib_mutex_t	mutex);	/*!< in: mutex to release */
/**********************************************************//**
Frees an mutex object. */
UNIV_INTERN
void
os_mutex_free(
/*==========*/
	os_ib_mutex_t	mutex);	/*!< in: mutex to free */
/**********************************************************//**
Acquires ownership of a fast mutex. Currently in Windows this is the same
as os_fast_mutex_lock!
@return	0 if success, != 0 if was reserved by another thread */
UNIV_INLINE
ulint
os_fast_mutex_trylock(
/*==================*/
	os_fast_mutex_t*	fast_mutex);	/*!< in: mutex to acquire */

/**********************************************************************
Following os_fast_ mutex APIs would be performance schema instrumented:

os_fast_mutex_init
os_fast_mutex_lock
os_fast_mutex_unlock
os_fast_mutex_free

These mutex APIs will point to corresponding wrapper functions that contain
the performance schema instrumentation.

NOTE! The following macro should be used in mutex operation, not the
corresponding function. */

#ifdef UNIV_PFS_MUTEX
# define os_fast_mutex_init(K, M)			\
	pfs_os_fast_mutex_init(K, M)

# define os_fast_mutex_lock(M)				\
	pfs_os_fast_mutex_lock(M, __FILE__, __LINE__)

# define os_fast_mutex_unlock(M)	pfs_os_fast_mutex_unlock(M)

# define os_fast_mutex_free(M)		pfs_os_fast_mutex_free(M)

/*********************************************************//**
NOTE! Please use the corresponding macro os_fast_mutex_init(), not directly
this function!
A wrapper function for os_fast_mutex_init_func(). Initializes an operating
system fast mutex semaphore. */
UNIV_INLINE
void
pfs_os_fast_mutex_init(
/*===================*/
	PSI_mutex_key		key,		/*!< in: Performance Schema
						key */
	os_fast_mutex_t*	fast_mutex);	/*!< out: fast mutex */
/**********************************************************//**
NOTE! Please use the corresponding macro os_fast_mutex_free(), not directly
this function!
Wrapper function for pfs_os_fast_mutex_free(). Also destroys the performance
schema probes when freeing the mutex */
UNIV_INLINE
void
pfs_os_fast_mutex_free(
/*===================*/
	os_fast_mutex_t*	fast_mutex);	/*!< in/out: mutex to free */
/**********************************************************//**
NOTE! Please use the corresponding macro os_fast_mutex_lock, not directly
this function!
Wrapper function of os_fast_mutex_lock. Acquires ownership of a fast mutex. */
UNIV_INLINE
void
pfs_os_fast_mutex_lock(
/*===================*/
	os_fast_mutex_t*	fast_mutex,	/*!< in/out: mutex to acquire */
	const char*		file_name,	/*!< in: file name where
						 locked */
	ulint			line);		/*!< in: line where locked */
/**********************************************************//**
NOTE! Please use the corresponding macro os_fast_mutex_unlock, not directly
this function!
Wrapper function of os_fast_mutex_unlock. Releases ownership of a fast mutex. */
UNIV_INLINE
void
pfs_os_fast_mutex_unlock(
/*=====================*/
	os_fast_mutex_t*	fast_mutex);	/*!< in/out: mutex to release */

#else /* UNIV_PFS_MUTEX */

# define os_fast_mutex_init(K, M)			\
	os_fast_mutex_init_func(&((os_fast_mutex_t*)(M))->mutex)

# define os_fast_mutex_lock(M)				\
	os_fast_mutex_lock_func(&((os_fast_mutex_t*)(M))->mutex)

# define os_fast_mutex_unlock(M)			\
	os_fast_mutex_unlock_func(&((os_fast_mutex_t*)(M))->mutex)

# define os_fast_mutex_free(M)				\
	os_fast_mutex_free_func(&((os_fast_mutex_t*)(M))->mutex)
#endif /* UNIV_PFS_MUTEX */

/**********************************************************//**
Acquires ownership of a fast mutex. Implies a full memory barrier even on
platforms such as PowerPC where this is not normally required.
@return	0 if success, != 0 if was reserved by another thread */
UNIV_INLINE
ulint
os_fast_mutex_trylock_full_barrier(
/*==================*/
	os_fast_mutex_t*	fast_mutex);	/*!< in: mutex to acquire */
/**********************************************************//**
Releases ownership of a fast mutex. */
UNIV_INTERN
void
os_fast_mutex_unlock_func(
/*======================*/
	fast_mutex_t*		fast_mutex);	/*!< in: mutex to release */
/**********************************************************//**
Releases ownership of a fast mutex. Implies a full memory barrier even on
platforms such as PowerPC where this is not normally required. */
UNIV_INTERN
void
os_fast_mutex_unlock_full_barrier(
/*=================*/
	os_fast_mutex_t*	fast_mutex);	/*!< in: mutex to release */
/*********************************************************//**
Initializes an operating system fast mutex semaphore. */
UNIV_INTERN
void
os_fast_mutex_init_func(
/*====================*/
	fast_mutex_t*		fast_mutex);	/*!< in: fast mutex */
/**********************************************************//**
Acquires ownership of a fast mutex. */
UNIV_INTERN
void
os_fast_mutex_lock_func(
/*====================*/
	fast_mutex_t*		fast_mutex);	/*!< in: mutex to acquire */
/**********************************************************//**
Frees an mutex object. */
UNIV_INTERN
void
os_fast_mutex_free_func(
/*====================*/
	fast_mutex_t*		fast_mutex);	/*!< in: mutex to free */

/**********************************************************//**
Atomic compare-and-swap and increment for InnoDB. */

#if defined(HAVE_IB_GCC_ATOMIC_BUILTINS)

# define HAVE_ATOMIC_BUILTINS

# ifdef HAVE_IB_GCC_ATOMIC_BUILTINS_BYTE
#  define HAVE_ATOMIC_BUILTINS_BYTE
# endif

# ifdef HAVE_IB_GCC_ATOMIC_BUILTINS_64
#  define HAVE_ATOMIC_BUILTINS_64
# endif

/**********************************************************//**
Returns true if swapped, ptr is pointer to target, old_val is value to
compare to, new_val is the value to swap in. */

# define os_compare_and_swap(ptr, old_val, new_val) \
	__sync_bool_compare_and_swap(ptr, old_val, new_val)

# define os_compare_and_swap_ulint(ptr, old_val, new_val) \
	os_compare_and_swap(ptr, old_val, new_val)

# define os_compare_and_swap_lint(ptr, old_val, new_val) \
	os_compare_and_swap(ptr, old_val, new_val)

#  define os_compare_and_swap_uint32(ptr, old_val, new_val) \
	os_compare_and_swap(ptr, old_val, new_val)

# ifdef HAVE_IB_ATOMIC_PTHREAD_T_GCC
#  define os_compare_and_swap_thread_id(ptr, old_val, new_val) \
	os_compare_and_swap(ptr, old_val, new_val)
#  define INNODB_RW_LOCKS_USE_ATOMICS
#  define IB_ATOMICS_STARTUP_MSG \
	"Mutexes and rw_locks use GCC atomic builtins"
# else /* HAVE_IB_ATOMIC_PTHREAD_T_GCC */
#  define IB_ATOMICS_STARTUP_MSG \
	"Mutexes use GCC atomic builtins, rw_locks do not"
# endif /* HAVE_IB_ATOMIC_PTHREAD_T_GCC */

/**********************************************************//**
Returns the resulting value, ptr is pointer to target, amount is the
amount of increment. */

# define os_atomic_increment(ptr, amount) \
	__sync_add_and_fetch(ptr, amount)

# define os_atomic_increment_lint(ptr, amount) \
	os_atomic_increment(ptr, amount)

# define os_atomic_increment_uint32(ptr, amount ) \
	os_atomic_increment(ptr, amount)

# define os_atomic_increment_ulint(ptr, amount) \
	os_atomic_increment(ptr, amount)

# define os_atomic_increment_uint64(ptr, amount) \
	os_atomic_increment(ptr, amount)

/* Returns the resulting value, ptr is pointer to target, amount is the
amount to decrement. */

# define os_atomic_decrement(ptr, amount) \
	__sync_sub_and_fetch(ptr, amount)

# define os_atomic_decrement_uint32(ptr, amount) \
	os_atomic_decrement(ptr, amount)

# define os_atomic_decrement_lint(ptr, amount) \
	os_atomic_decrement(ptr, amount)

# define os_atomic_decrement_ulint(ptr, amount) \
	os_atomic_decrement(ptr, amount)

# define os_atomic_decrement_uint64(ptr, amount) \
	os_atomic_decrement(ptr, amount)

# if defined(HAVE_ATOMIC_BUILTINS)

/** Do an atomic test and set.
@param[in,out]	ptr		Memory location to set to non-zero
@return the previous value */
inline
lock_word_t
os_atomic_test_and_set(volatile lock_word_t* ptr)
{
	return(__sync_lock_test_and_set(ptr, 1));
}

/** Do an atomic release.
@param[in,out]	ptr		Memory location to write to
@return the previous value */
inline
void
os_atomic_clear(volatile lock_word_t* ptr)
{
	__sync_lock_release(ptr);
}

# elif defined(HAVE_IB_GCC_ATOMIC_TEST_AND_SET)

/** Do an atomic test-and-set.
@param[in,out]	ptr		Memory location to set to non-zero
@return the previous value */
inline
lock_word_t
os_atomic_test_and_set(volatile lock_word_t* ptr)
{
       return(__atomic_test_and_set(ptr, __ATOMIC_ACQUIRE));
}

/** Do an atomic clear.
@param[in,out]	ptr		Memory location to set to zero */
inline
void
os_atomic_clear(volatile lock_word_t* ptr)
{
	__atomic_clear(ptr, __ATOMIC_RELEASE);
}

# else

#  error "Unsupported platform"

# endif /* HAVE_IB_GCC_ATOMIC_TEST_AND_SET */

#if defined(__powerpc__) || defined(__aarch64__)
/*
  os_atomic_test_and_set_byte_release() should imply a release barrier before
  setting, and a full barrier after. But __sync_lock_test_and_set() is only
  documented as an aquire barrier. So on PowerPC we need to add the full
  barrier explicitly.  */
# define os_atomic_test_and_set_byte_release(ptr, new_val) \
        do { __sync_lock_release(ptr); \
		__sync_synchronize(); } while (0)
#else
/*
  On x86, __sync_lock_test_and_set() happens to be full barrier, due to
  LOCK prefix.
*/
# define os_atomic_test_and_set_byte_release(ptr, new_val) \
	__sync_lock_test_and_set(ptr, (byte) new_val)
#endif
/*
  os_atomic_test_and_set_byte_acquire() is a full memory barrier on x86. But
  in general, just an aquire barrier should be sufficient. */
# define os_atomic_test_and_set_byte_acquire(ptr, new_val) \
	__sync_lock_test_and_set(ptr, (byte) new_val)

#elif defined(HAVE_IB_SOLARIS_ATOMICS)

# define HAVE_ATOMIC_BUILTINS
# define HAVE_ATOMIC_BUILTINS_BYTE
# define HAVE_ATOMIC_BUILTINS_64

/* If not compiling with GCC or GCC doesn't support the atomic
intrinsics and running on Solaris >= 10 use Solaris atomics */

# include <atomic.h>

/**********************************************************//**
Returns true if swapped, ptr is pointer to target, old_val is value to
compare to, new_val is the value to swap in. */

# define os_compare_and_swap_uint32(ptr, old_val, new_val) \
	(atomic_cas_32(ptr, old_val, new_val) == old_val)

# define os_compare_and_swap_ulint(ptr, old_val, new_val) \
	(atomic_cas_ulong(ptr, old_val, new_val) == old_val)

# define os_compare_and_swap_lint(ptr, old_val, new_val) \
	((lint) atomic_cas_ulong((ulong_t*) ptr, old_val, new_val) == old_val)

# ifdef HAVE_IB_ATOMIC_PTHREAD_T_SOLARIS
#  if SIZEOF_PTHREAD_T == 4
#   define os_compare_and_swap_thread_id(ptr, old_val, new_val) \
	((pthread_t) atomic_cas_32(ptr, old_val, new_val) == old_val)
#  elif SIZEOF_PTHREAD_T == 8
#   define os_compare_and_swap_thread_id(ptr, old_val, new_val) \
	((pthread_t) atomic_cas_64(ptr, old_val, new_val) == old_val)
#  else
#   error "SIZEOF_PTHREAD_T != 4 or 8"
#  endif /* SIZEOF_PTHREAD_T CHECK */
#  define INNODB_RW_LOCKS_USE_ATOMICS
#  define IB_ATOMICS_STARTUP_MSG \
	"Mutexes and rw_locks use Solaris atomic functions"
# else /* HAVE_IB_ATOMIC_PTHREAD_T_SOLARIS */
#  define IB_ATOMICS_STARTUP_MSG \
	"Mutexes use Solaris atomic functions, rw_locks do not"
# endif /* HAVE_IB_ATOMIC_PTHREAD_T_SOLARIS */

/**********************************************************//**
Returns the resulting value, ptr is pointer to target, amount is the
amount of increment. */

# define os_atomic_increment_uint32(ptr, amount) \
	atomic_add_32_nv(ptr, amount)

# define os_atomic_increment_ulint(ptr, amount) \
	atomic_add_long_nv(ptr, amount)

# define os_atomic_increment_lint(ptr, amount) \
	os_atomic_increment_ulint((ulong_t*) ptr, amount)

# define os_atomic_increment_uint64(ptr, amount) \
	atomic_add_64_nv((uint64_t *) ptr, amount)

/* Returns the resulting value, ptr is pointer to target, amount is the
amount to decrement. */

# define os_atomic_decrement_uint32(ptr, amount) \
	os_atomic_increment_uint32(ptr, -(amount))

# define os_atomic_decrement_lint(ptr, amount) \
	os_atomic_increment_ulint((ulong_t*) ptr, -(amount))

# define os_atomic_decrement_ulint(ptr, amount) \
	os_atomic_increment_ulint(ptr, -(amount))

# define os_atomic_decrement_uint64(ptr, amount) \
	os_atomic_increment_uint64(ptr, -(amount))

# ifdef IB_LOCK_WORD_IS_BYTE

/** Do an atomic xchg and set to non-zero.
@param[in,out]	ptr		Memory location to set to non-zero
@return the previous value */
inline
lock_word_t
os_atomic_test_and_set(volatile lock_word_t* ptr)
{
	return(atomic_swap_uchar(ptr, 1));
}

/** Do an atomic xchg and set to zero.
@param[in,out]	ptr		Memory location to set to zero
@return the previous value */
inline
lock_word_t
os_atomic_clear(volatile lock_word_t* ptr)
{
	return(atomic_swap_uchar(ptr, 0));
}

# else

/** Do an atomic xchg and set to non-zero.
@param[in,out]	ptr		Memory location to set to non-zero
@return the previous value */
inline
lock_word_t
os_atomic_test_and_set(volatile lock_word_t* ptr)
{
	return(atomic_swap_ulong(ptr, 1));
}

/** Do an atomic xchg and set to zero.
@param[in,out]	ptr		Memory location to set to zero
@return the previous value */
inline
lock_word_t
os_atomic_clear(volatile lock_word_t* ptr)
{
	return(atomic_swap_ulong(ptr, 0));
}

# endif /* IB_LOCK_WORD_IS_BYTE */

# define os_atomic_test_and_set_byte_acquire(ptr, new_val) \
	atomic_swap_uchar(ptr, new_val)

# define os_atomic_test_and_set_byte_release(ptr, new_val) \
	atomic_swap_uchar(ptr, new_val)

#elif defined(HAVE_WINDOWS_ATOMICS)

# define HAVE_ATOMIC_BUILTINS
# define HAVE_ATOMIC_BUILTINS_BYTE
# define HAVE_ATOMIC_BUILTINS_64

/**********************************************************//**
Atomic compare and exchange of signed integers (both 32 and 64 bit).
@return value found before the exchange.
If it is not equal to old_value the exchange did not happen. */
UNIV_INLINE
lint
win_cmp_and_xchg_lint(
/*==================*/
	volatile lint*	ptr,		/*!< in/out: source/destination */
	lint		new_val,	/*!< in: exchange value */
	lint		old_val);	/*!< in: value to compare to */

/**********************************************************//**
Atomic addition of signed integers.
@return Initial value of the variable pointed to by ptr */
UNIV_INLINE
lint
win_xchg_and_add(
/*=============*/
	volatile lint*	ptr,	/*!< in/out: address of destination */
	lint		val);	/*!< in: number to be added */

/**********************************************************//**
Atomic compare and exchange of unsigned integers.
@return value found before the exchange.
If it is not equal to old_value the exchange did not happen. */
UNIV_INLINE
ulint
win_cmp_and_xchg_ulint(
/*===================*/
	volatile ulint*	ptr,		/*!< in/out: source/destination */
	ulint		new_val,	/*!< in: exchange value */
	ulint		old_val);	/*!< in: value to compare to */

/**********************************************************//**
Atomic compare and exchange of 32 bit unsigned integers.
@return value found before the exchange.
If it is not equal to old_value the exchange did not happen. */
UNIV_INLINE
DWORD
win_cmp_and_xchg_dword(
/*===================*/
	volatile DWORD*	ptr,		/*!< in/out: source/destination */
	DWORD		new_val,	/*!< in: exchange value */
	DWORD		old_val);	/*!< in: value to compare to */

/**********************************************************//**
Returns true if swapped, ptr is pointer to target, old_val is value to
compare to, new_val is the value to swap in. */

# define os_compare_and_swap_uint32(ptr, old_val, new_val) \
	(InterlockedCompareExchange(reinterpret_cast<volatile long*>(ptr), \
				    new_val, old_val) == old_val)

# define os_compare_and_swap_ulint(ptr, old_val, new_val) \
	(win_cmp_and_xchg_ulint(ptr, new_val, old_val) == old_val)

# define os_compare_and_swap_lint(ptr, old_val, new_val) \
	(win_cmp_and_xchg_lint(ptr, new_val, old_val) == old_val)

/* windows thread objects can always be passed to windows atomic functions */
# define os_compare_and_swap_thread_id(ptr, old_val, new_val) \
	(win_cmp_and_xchg_dword(ptr, new_val, old_val) == old_val)

# define INNODB_RW_LOCKS_USE_ATOMICS
# define IB_ATOMICS_STARTUP_MSG \
	"Mutexes and rw_locks use Windows interlocked functions"

/**********************************************************//**
Returns the resulting value, ptr is pointer to target, amount is the
amount of increment. */

# define os_atomic_increment_lint(ptr, amount) \
	(win_xchg_and_add(ptr, amount) + amount)

# define os_atomic_increment_uint32(ptr, amount) \
	((ulint) InterlockedExchangeAdd((long*) ptr, amount))

# define os_atomic_increment_ulint(ptr, amount) \
	((ulint) (win_xchg_and_add((lint*) ptr, (lint) amount) + amount))

# define os_atomic_increment_uint64(ptr, amount)		\
	((ib_uint64_t) (InterlockedExchangeAdd64(		\
				(ib_int64_t*) ptr,		\
				(ib_int64_t) amount) + amount))

/**********************************************************//**
Returns the resulting value, ptr is pointer to target, amount is the
amount to decrement. There is no atomic substract function on Windows */

# define os_atomic_decrement_uint32(ptr, amount) \
	((ulint) InterlockedExchangeAdd((long*) ptr, (-amount)))

# define os_atomic_decrement_lint(ptr, amount) \
	(win_xchg_and_add(ptr, -(lint) amount) - amount)

# define os_atomic_decrement_ulint(ptr, amount) \
	((ulint) (win_xchg_and_add((lint*) ptr, -(lint) amount) - amount))

# define os_atomic_decrement_uint64(ptr, amount)		\
	((ib_uint64_t) (InterlockedExchangeAdd64(		\
				(ib_int64_t*) ptr,		\
				-(ib_int64_t) amount) - amount))

/** Do an atomic test and set.
InterlockedExchange() operates on LONG, and the LONG will be clobbered
@param[in,out]	ptr		Memory location to set to non-zero
@return the previous value */
inline
lock_word_t
os_atomic_test_and_set(volatile lock_word_t* ptr)
{
	return(InterlockedExchange(ptr, 1));
}

/** Do an atomic release.
InterlockedExchange() operates on LONG, and the LONG will be clobbered
@param[in,out]	ptr		Memory location to set to zero
@return the previous value */
inline
lock_word_t
os_atomic_clear(volatile lock_word_t* ptr)
{
	return(InterlockedExchange(ptr, 0));
}

# define os_atomic_lock_release_byte(ptr) \
	(void) InterlockedExchange(ptr, 0)

#else
# define IB_ATOMICS_STARTUP_MSG \
	"Mutexes and rw_locks use InnoDB's own implementation"
#endif
#ifdef HAVE_ATOMIC_BUILTINS
#define os_atomic_inc_ulint(m,v,d)	os_atomic_increment_ulint(v, d)
#define os_atomic_dec_ulint(m,v,d)	os_atomic_decrement_ulint(v, d)
#else
#define os_atomic_inc_ulint(m,v,d)	os_atomic_inc_ulint_func(m, v, d)
#define os_atomic_dec_ulint(m,v,d)	os_atomic_dec_ulint_func(m, v, d)
#endif /* HAVE_ATOMIC_BUILTINS */

/**********************************************************//**
Following macros are used to update specified counter atomically
if HAVE_ATOMIC_BUILTINS defined. Otherwise, use mutex passed in
for synchronization */
#ifdef HAVE_ATOMIC_BUILTINS
#define os_increment_counter_by_amount(mutex, counter, amount)	\
	(void) os_atomic_increment_ulint(&counter, amount)

#define os_decrement_counter_by_amount(mutex, counter, amount)	\
	(void) os_atomic_increment_ulint(&counter, (-((lint) amount)))
#else
#define os_increment_counter_by_amount(mutex, counter, amount)	\
	do {							\
		mutex_enter(&(mutex));				\
		(counter) += (amount);				\
		mutex_exit(&(mutex));				\
	} while (0)

#define os_decrement_counter_by_amount(mutex, counter, amount)	\
	do {							\
		ut_a(counter >= amount);			\
		mutex_enter(&(mutex));				\
		(counter) -= (amount);				\
		mutex_exit(&(mutex));				\
	} while (0)
#endif  /* HAVE_ATOMIC_BUILTINS */

#define os_inc_counter(mutex, counter)				\
	os_increment_counter_by_amount(mutex, counter, 1)

#define os_dec_counter(mutex, counter)				\
	do {							\
		os_decrement_counter_by_amount(mutex, counter, 1);\
	} while (0);

/** barrier definitions for memory ordering */
#if defined(HAVE_IB_GCC_ATOMIC_THREAD_FENCE)
# define HAVE_MEMORY_BARRIER
# define os_rmb	__atomic_thread_fence(__ATOMIC_ACQUIRE)
# define os_wmb	__atomic_thread_fence(__ATOMIC_RELEASE)
# define os_mb __atomic_thread_fence(__ATOMIC_SEQ_CST)

# define IB_MEMORY_BARRIER_STARTUP_MSG \
	"GCC builtin __atomic_thread_fence() is used for memory barrier"

#elif defined(HAVE_IB_GCC_SYNC_SYNCHRONISE)
# define HAVE_MEMORY_BARRIER
# define os_rmb	__sync_synchronize()
# define os_wmb	__sync_synchronize()
# define os_mb	__sync_synchronize()
# define IB_MEMORY_BARRIER_STARTUP_MSG \
	"GCC builtin __sync_synchronize() is used for memory barrier"

#elif defined(HAVE_IB_MACHINE_BARRIER_SOLARIS)
# define HAVE_MEMORY_BARRIER
# include <mbarrier.h>
# define os_rmb	__machine_r_barrier()
# define os_wmb	__machine_w_barrier()
# define os_mb __machine_rw_barrier()
# define IB_MEMORY_BARRIER_STARTUP_MSG \
	"Solaris memory ordering functions are used for memory barrier"

#elif defined(HAVE_WINDOWS_MM_FENCE)
# define HAVE_MEMORY_BARRIER
# include <intrin.h>
# define os_rmb	_mm_lfence()
# define os_wmb	_mm_sfence()
# define os_mb	_mm_mfence()
# define IB_MEMORY_BARRIER_STARTUP_MSG \
	"_mm_lfence() and _mm_sfence() are used for memory barrier"

#else
# define os_rmb do { } while(0)
# define os_wmb do { } while(0)
# define os_mb do { } while(0)
# define IB_MEMORY_BARRIER_STARTUP_MSG \
	"Memory barrier is not used"
#endif


/** Simple counter aligned to CACHE_LINE_SIZE
@tparam	Type	the integer type of the counter
@tparam	atomic	whether to use atomic memory access */
template <typename Type = ulint, bool atomic = false>
struct MY_ALIGNED(CACHE_LINE_SIZE) simple_counter
{
	/** Increment the counter */
	Type inc() { return add(1); }
	/** Decrement the counter */
	Type dec() { return sub(1); }

	/** Add to the counter
	@param[in]	i	amount to be added
	@return	the value of the counter after adding */
	Type add(Type i)
	{
		compile_time_assert(!atomic || sizeof(Type) == sizeof(ulint));
		if (atomic) {
			/* GCC would perform a type check in this code
			also in case the template is instantiated with
			simple_counter<Type=not_ulint, atomic=false>.
			On Solaris, os_atomic_increment_ulint() maps
			to atomic_add_long_nv(), which expects the
			parameter to be correctly typed. */
			return os_atomic_increment_ulint(
				reinterpret_cast<ulint*>(&m_counter), i);
		} else {
			return m_counter += i;
		}
	}
	/** Subtract from the counter
	@param[in]	i	amount to be subtracted
	@return	the value of the counter after adding */
	Type sub(Type i)
	{
		compile_time_assert(!atomic || sizeof(Type) == sizeof(ulint));
		if (atomic) {
			return os_atomic_decrement_ulint(&m_counter, i);
		} else {
			return m_counter -= i;
		}
	}

	/** @return the value of the counter (non-atomic access)! */
	operator Type() const { return m_counter; }

private:
	/** The counter */
	Type	m_counter;
};

#ifndef UNIV_NONINL
#include "os0sync.ic"
#endif

#endif