mdb.c

/** @file mdb.c
 *	@brief Lightning memory-mapped database library
 *
 *	A Btree-based database management library modeled loosely on the
 *	BerkeleyDB API, but much simplified.
 */
/*
 * Copyright 2011-2019 Howard Chu, Symas Corp.
 * All rights reserved.
 *
 * Redistribution and use in source and binary forms, with or without
 * modification, are permitted only as authorized by the OpenLDAP
 * Public License.
 *
 * A copy of this license is available in the file LICENSE in the
 * top-level directory of the distribution or, alternatively, at
 * <http://www.OpenLDAP.org/license.html>.
 *
 * This code is derived from btree.c written by Martin Hedenfalk.
 *
 * Copyright (c) 2009, 2010 Martin Hedenfalk <martin@bzero.se>
 *
 * Permission to use, copy, modify, and distribute this software for any
 * purpose with or without fee is hereby granted, provided that the above
 * copyright notice and this permission notice appear in all copies.
 *
 * THE SOFTWARE IS PROVIDED "AS IS" AND THE AUTHOR DISCLAIMS ALL WARRANTIES
 * WITH REGARD TO THIS SOFTWARE INCLUDING ALL IMPLIED WARRANTIES OF
 * MERCHANTABILITY AND FITNESS. IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR
 * ANY SPECIAL, DIRECT, INDIRECT, OR CONSEQUENTIAL DAMAGES OR ANY DAMAGES
 * WHATSOEVER RESULTING FROM LOSS OF USE, DATA OR PROFITS, WHETHER IN AN
 * ACTION OF CONTRACT, NEGLIGENCE OR OTHER TORTIOUS ACTION, ARISING OUT OF
 * OR IN CONNECTION WITH THE USE OR PERFORMANCE OF THIS SOFTWARE.
 */
#ifndef _GNU_SOURCE
#define _GNU_SOURCE 1
#endif
#if defined(MDB_VL32) || defined(__WIN64__)
#define _FILE_OFFSET_BITS	64
#endif
#ifdef _WIN32
#include <malloc.h>
#include <windows.h>
#include <wchar.h>				/* get wcscpy() */

/* We use native NT APIs to setup the memory map, so that we can
 * let the DB file grow incrementally instead of always preallocating
 * the full size. These APIs are defined in <wdm.h> and <ntifs.h>
 * but those headers are meant for driver-level development and
 * conflict with the regular user-level headers, so we explicitly
 * declare them here. We get pointers to these functions from
 * NTDLL.DLL at runtime, to avoid buildtime dependencies on any
 * NTDLL import libraries.
 */
typedef NTSTATUS (WINAPI NtCreateSectionFunc)
  (OUT PHANDLE sh, IN ACCESS_MASK acc,
  IN void * oa OPTIONAL,
  IN PLARGE_INTEGER ms OPTIONAL,
  IN ULONG pp, IN ULONG aa, IN HANDLE fh OPTIONAL);

static NtCreateSectionFunc *NtCreateSection;

typedef enum _SECTION_INHERIT {
	ViewShare = 1,
	ViewUnmap = 2
} SECTION_INHERIT;

typedef NTSTATUS (WINAPI NtMapViewOfSectionFunc)
  (IN PHANDLE sh, IN HANDLE ph,
  IN OUT PVOID *addr, IN ULONG_PTR zbits,
  IN SIZE_T cs, IN OUT PLARGE_INTEGER off OPTIONAL,
  IN OUT PSIZE_T vs, IN SECTION_INHERIT ih,
  IN ULONG at, IN ULONG pp);

static NtMapViewOfSectionFunc *NtMapViewOfSection;

typedef NTSTATUS (WINAPI NtCloseFunc)(HANDLE h);

static NtCloseFunc *NtClose;

/** getpid() returns int; MinGW defines pid_t but MinGW64 typedefs it
 *  as int64 which is wrong. MSVC doesn't define it at all, so just
 *  don't use it.
 */
#define MDB_PID_T	int
#define MDB_THR_T	DWORD
#include <sys/types.h>
#include <sys/stat.h>
#ifdef __GNUC__
# include <sys/param.h>
#else
# define LITTLE_ENDIAN	1234
# define BIG_ENDIAN	4321
# define BYTE_ORDER	LITTLE_ENDIAN
# ifndef SSIZE_MAX
#  define SSIZE_MAX	INT_MAX
# endif
#endif
#else
#include <sys/types.h>
#include <sys/stat.h>
#define MDB_PID_T	pid_t
#define MDB_THR_T	pthread_t
#include <sys/param.h>
#include <sys/uio.h>
#include <sys/mman.h>
#ifdef HAVE_SYS_FILE_H
#include <sys/file.h>
#endif
#include <fcntl.h>
#endif

#if defined(__mips) && defined(__linux)
/* MIPS has cache coherency issues, requires explicit cache control */
#include <asm/cachectl.h>
extern int cacheflush(char *addr, int nbytes, int cache);
#define CACHEFLUSH(addr, bytes, cache)	cacheflush(addr, bytes, cache)
#else
#define CACHEFLUSH(addr, bytes, cache)
#endif

#if defined(__linux) && !defined(MDB_FDATASYNC_WORKS)
/** fdatasync is broken on ext3/ext4fs on older kernels, see
 *	description in #mdb_env_open2 comments. You can safely
 *	define MDB_FDATASYNC_WORKS if this code will only be run
 *	on kernels 3.6 and newer.
 */
#define	BROKEN_FDATASYNC
#endif

#include <errno.h>
#include <limits.h>
#include <stddef.h>
#include <inttypes.h>
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <time.h>

#ifdef _MSC_VER
#include <io.h>
typedef SSIZE_T	ssize_t;
#else
#include <unistd.h>
#endif

#if defined(__sun) || defined(__ANDROID__)
/* Most platforms have posix_memalign, older may only have memalign */
#define HAVE_MEMALIGN	1
#include <malloc.h>
/* On Solaris, we need the POSIX sigwait function */
#if defined (__sun)
# define _POSIX_PTHREAD_SEMANTICS	1
#endif
#endif

#if !(defined(BYTE_ORDER) || defined(__BYTE_ORDER))
#include <netinet/in.h>
#include <resolv.h>	/* defines BYTE_ORDER on HPUX and Solaris */
#endif

#if defined(__APPLE__) || defined (BSD) || defined(__FreeBSD_kernel__)
# if !(defined(MDB_USE_POSIX_MUTEX) || defined(MDB_USE_POSIX_SEM))
# define MDB_USE_SYSV_SEM	1
# endif
# define MDB_FDATASYNC		fsync
#elif defined(__ANDROID__)
# define MDB_FDATASYNC		fsync
#endif

#ifndef _WIN32
#include <pthread.h>
#include <signal.h>
#ifdef MDB_USE_POSIX_SEM
# define MDB_USE_HASH		1
#include <semaphore.h>
#elif defined(MDB_USE_SYSV_SEM)
#include <sys/ipc.h>
#include <sys/sem.h>
#ifdef _SEM_SEMUN_UNDEFINED
union semun {
	int val;
	struct semid_ds *buf;
	unsigned short *array;
};
#endif /* _SEM_SEMUN_UNDEFINED */
#else
#define MDB_USE_POSIX_MUTEX	1
#endif /* MDB_USE_POSIX_SEM */
#endif /* !_WIN32 */

#if defined(_WIN32) + defined(MDB_USE_POSIX_SEM) + defined(MDB_USE_SYSV_SEM) \
	+ defined(MDB_USE_POSIX_MUTEX) != 1
# error "Ambiguous shared-lock implementation"
#endif

#ifdef USE_VALGRIND
#include <valgrind/memcheck.h>
#define VGMEMP_CREATE(h,r,z)    VALGRIND_CREATE_MEMPOOL(h,r,z)
#define VGMEMP_ALLOC(h,a,s) VALGRIND_MEMPOOL_ALLOC(h,a,s)
#define VGMEMP_FREE(h,a) VALGRIND_MEMPOOL_FREE(h,a)
#define VGMEMP_DESTROY(h)	VALGRIND_DESTROY_MEMPOOL(h)
#define VGMEMP_DEFINED(a,s)	VALGRIND_MAKE_MEM_DEFINED(a,s)
#else
#define VGMEMP_CREATE(h,r,z)
#define VGMEMP_ALLOC(h,a,s)
#define VGMEMP_FREE(h,a)
#define VGMEMP_DESTROY(h)
#define VGMEMP_DEFINED(a,s)
#endif

#ifndef BYTE_ORDER
# if (defined(_LITTLE_ENDIAN) || defined(_BIG_ENDIAN)) && !(defined(_LITTLE_ENDIAN) && defined(_BIG_ENDIAN))
/* Solaris just defines one or the other */
#  define LITTLE_ENDIAN	1234
#  define BIG_ENDIAN	4321
#  ifdef _LITTLE_ENDIAN
#   define BYTE_ORDER  LITTLE_ENDIAN
#  else
#   define BYTE_ORDER  BIG_ENDIAN
#  endif
# else
#  define BYTE_ORDER   __BYTE_ORDER
# endif
#endif

#ifndef LITTLE_ENDIAN
#define LITTLE_ENDIAN	__LITTLE_ENDIAN
#endif
#ifndef BIG_ENDIAN
#define BIG_ENDIAN	__BIG_ENDIAN
#endif

#if defined(__i386) || defined(__x86_64) || defined(_M_IX86)
#define MISALIGNED_OK	1
#endif

#include "lmdb.h"
#include "midl.h"

#if (BYTE_ORDER == LITTLE_ENDIAN) == (BYTE_ORDER == BIG_ENDIAN)
# error "Unknown or unsupported endianness (BYTE_ORDER)"
#elif (-6 & 5) || CHAR_BIT!=8 || UINT_MAX!=0xffffffff || MDB_SIZE_MAX%UINT_MAX
# error "Two's complement, reasonably sized integer types, please"
#endif

#ifdef __GNUC__
/** Put infrequently used env functions in separate section */
# ifdef __APPLE__
#  define	ESECT	__attribute__ ((section("__TEXT,text_env")))
# else
#  define	ESECT	__attribute__ ((section("text_env")))
# endif
#else
#define ESECT
#endif

#ifdef _WIN32
#define CALL_CONV WINAPI
#else
#define CALL_CONV
#endif

/** @defgroup internal	LMDB Internals
 *	@{
 */
/** @defgroup compat	Compatibility Macros
 *	A bunch of macros to minimize the amount of platform-specific ifdefs
 *	needed throughout the rest of the code. When the features this library
 *	needs are similar enough to POSIX to be hidden in a one-or-two line
 *	replacement, this macro approach is used.
 *	@{
 */

	/** Features under development */
#ifndef MDB_DEVEL
#define MDB_DEVEL 0
#endif

	/** Wrapper around __func__, which is a C99 feature */
#if __STDC_VERSION__ >= 199901L
# define mdb_func_	__func__
#elif __GNUC__ >= 2 || _MSC_VER >= 1300
# define mdb_func_	__FUNCTION__
#else
/* If a debug message says <mdb_unknown>(), update the #if statements above */
# define mdb_func_	"<mdb_unknown>"
#endif

/* Internal error codes, not exposed outside liblmdb */
#define	MDB_NO_ROOT		(MDB_LAST_ERRCODE + 10)
#ifdef _WIN32
#define MDB_OWNERDEAD	((int) WAIT_ABANDONED)
#elif defined MDB_USE_SYSV_SEM
#define MDB_OWNERDEAD	(MDB_LAST_ERRCODE + 11)
#elif defined(MDB_USE_POSIX_MUTEX) && defined(EOWNERDEAD)
#define MDB_OWNERDEAD	EOWNERDEAD	/**< #LOCK_MUTEX0() result if dead owner */
#endif

#ifdef __GLIBC__
#define	GLIBC_VER	((__GLIBC__ << 16 )| __GLIBC_MINOR__)
#endif
/** Some platforms define the EOWNERDEAD error code
 * even though they don't support Robust Mutexes.
 * Compile with -DMDB_USE_ROBUST=0, or use some other
 * mechanism like -DMDB_USE_SYSV_SEM instead of
 * -DMDB_USE_POSIX_MUTEX. (SysV semaphores are
 * also Robust, but some systems don't support them
 * either.)
 */
#ifndef MDB_USE_ROBUST
/* Android currently lacks Robust Mutex support. So does glibc < 2.4. */
# if defined(MDB_USE_POSIX_MUTEX) && (defined(__ANDROID__) || \
	(defined(__GLIBC__) && GLIBC_VER < 0x020004))
#  define MDB_USE_ROBUST	0
# else
#  define MDB_USE_ROBUST	1
# endif
#endif /* !MDB_USE_ROBUST */

#if defined(MDB_USE_POSIX_MUTEX) && (MDB_USE_ROBUST)
/* glibc < 2.12 only provided _np API */
#  if (defined(__GLIBC__) && GLIBC_VER < 0x02000c) || \
	(defined(PTHREAD_MUTEX_ROBUST_NP) && !defined(PTHREAD_MUTEX_ROBUST))
#   define PTHREAD_MUTEX_ROBUST	PTHREAD_MUTEX_ROBUST_NP
#   define pthread_mutexattr_setrobust(attr, flag)	pthread_mutexattr_setrobust_np(attr, flag)
#   define pthread_mutex_consistent(mutex)	pthread_mutex_consistent_np(mutex)
#  endif
#endif /* MDB_USE_POSIX_MUTEX && MDB_USE_ROBUST */

#if defined(MDB_OWNERDEAD) && (MDB_USE_ROBUST)
#define MDB_ROBUST_SUPPORTED	1
#endif

#ifdef _WIN32
#define MDB_USE_HASH	1
#define MDB_PIDLOCK	0
#define THREAD_RET	DWORD
#define pthread_t	HANDLE
#define pthread_mutex_t	HANDLE
#define pthread_cond_t	HANDLE
typedef HANDLE mdb_mutex_t, mdb_mutexref_t;
#define pthread_key_t	DWORD
#define pthread_self()	GetCurrentThreadId()
#define pthread_key_create(x,y)	\
	((*(x) = TlsAlloc()) == TLS_OUT_OF_INDEXES ? ErrCode() : 0)
#define pthread_key_delete(x)	TlsFree(x)
#define pthread_getspecific(x)	TlsGetValue(x)
#define pthread_setspecific(x,y)	(TlsSetValue(x,y) ? 0 : ErrCode())
#define pthread_mutex_unlock(x)	ReleaseMutex(*x)
#define pthread_mutex_lock(x)	WaitForSingleObject(*x, INFINITE)
#define pthread_cond_signal(x)	SetEvent(*x)
#define pthread_cond_wait(cond,mutex)	do{SignalObjectAndWait(*mutex, *cond, INFINITE, FALSE); WaitForSingleObject(*mutex, INFINITE);}while(0)
#define THREAD_CREATE(thr,start,arg) \
	(((thr) = CreateThread(NULL, 0, start, arg, 0, NULL)) ? 0 : ErrCode())
#define THREAD_FINISH(thr) \
	(WaitForSingleObject(thr, INFINITE) ? ErrCode() : 0)
#define LOCK_MUTEX0(mutex)		WaitForSingleObject(mutex, INFINITE)
#define UNLOCK_MUTEX(mutex)		ReleaseMutex(mutex)
#define mdb_mutex_consistent(mutex)	0
#define getpid()	GetCurrentProcessId()
#define	MDB_FDATASYNC(fd)	(!FlushFileBuffers(fd))
#define	MDB_MSYNC(addr,len,flags)	(!FlushViewOfFile(addr,len))
#define	ErrCode()	GetLastError()
#define GET_PAGESIZE(x) {SYSTEM_INFO si; GetSystemInfo(&si); (x) = si.dwPageSize;}
#define	close(fd)	(CloseHandle(fd) ? 0 : -1)
#define	munmap(ptr,len)	UnmapViewOfFile(ptr)
#ifdef PROCESS_QUERY_LIMITED_INFORMATION
#define MDB_PROCESS_QUERY_LIMITED_INFORMATION PROCESS_QUERY_LIMITED_INFORMATION
#else
#define MDB_PROCESS_QUERY_LIMITED_INFORMATION 0x1000
#endif
#else
#define THREAD_RET	void *
#define THREAD_CREATE(thr,start,arg)	pthread_create(&thr,NULL,start,arg)
#define THREAD_FINISH(thr)	pthread_join(thr,NULL)

	/** For MDB_LOCK_FORMAT: True if readers take a pid lock in the lockfile */
#define MDB_PIDLOCK			1

#ifdef MDB_USE_POSIX_SEM

typedef sem_t *mdb_mutex_t, *mdb_mutexref_t;
#define LOCK_MUTEX0(mutex)		mdb_sem_wait(mutex)
#define UNLOCK_MUTEX(mutex)		sem_post(mutex)

static int
mdb_sem_wait(sem_t *sem)
{
   int rc;
   while ((rc = sem_wait(sem)) && (rc = errno) == EINTR) ;
   return rc;
}

#elif defined MDB_USE_SYSV_SEM

typedef struct mdb_mutex {
	int semid;
	int semnum;
	int *locked;
} mdb_mutex_t[1], *mdb_mutexref_t;

#define LOCK_MUTEX0(mutex)		mdb_sem_wait(mutex)
#define UNLOCK_MUTEX(mutex)		do { \
	struct sembuf sb = { 0, 1, SEM_UNDO }; \
	sb.sem_num = (mutex)->semnum; \
	*(mutex)->locked = 0; \
	semop((mutex)->semid, &sb, 1); \
} while(0)

static int
mdb_sem_wait(mdb_mutexref_t sem)
{
	int rc, *locked = sem->locked;
	struct sembuf sb = { 0, -1, SEM_UNDO };
	sb.sem_num = sem->semnum;
	do {
		if (!semop(sem->semid, &sb, 1)) {
			rc = *locked ? MDB_OWNERDEAD : MDB_SUCCESS;
			*locked = 1;
			break;
		}
	} while ((rc = errno) == EINTR);
	return rc;
}

#define mdb_mutex_consistent(mutex)	0

#else	/* MDB_USE_POSIX_MUTEX: */
	/** Shared mutex/semaphore as the original is stored.
	 *
	 *	Not for copies.  Instead it can be assigned to an #mdb_mutexref_t.
	 *	When mdb_mutexref_t is a pointer and mdb_mutex_t is not, then it
	 *	is array[size 1] so it can be assigned to the pointer.
	 */
typedef pthread_mutex_t mdb_mutex_t[1];
	/** Reference to an #mdb_mutex_t */
typedef pthread_mutex_t *mdb_mutexref_t;
	/** Lock the reader or writer mutex.
	 *	Returns 0 or a code to give #mdb_mutex_failed(), as in #LOCK_MUTEX().
	 */
#define LOCK_MUTEX0(mutex)	pthread_mutex_lock(mutex)
	/** Unlock the reader or writer mutex.
	 */
#define UNLOCK_MUTEX(mutex)	pthread_mutex_unlock(mutex)
	/** Mark mutex-protected data as repaired, after death of previous owner.
	 */
#define mdb_mutex_consistent(mutex)	pthread_mutex_consistent(mutex)
#endif	/* MDB_USE_POSIX_SEM || MDB_USE_SYSV_SEM */

	/** Get the error code for the last failed system function.
	 */
#define	ErrCode()	errno

	/** An abstraction for a file handle.
	 *	On POSIX systems file handles are small integers. On Windows
	 *	they're opaque pointers.
	 */
#define	HANDLE	int

	/**	A value for an invalid file handle.
	 *	Mainly used to initialize file variables and signify that they are
	 *	unused.
	 */
#define INVALID_HANDLE_VALUE	(-1)

	/** Get the size of a memory page for the system.
	 *	This is the basic size that the platform's memory manager uses, and is
	 *	fundamental to the use of memory-mapped files.
	 */
#define	GET_PAGESIZE(x)	((x) = sysconf(_SC_PAGE_SIZE))
#endif

#define	Z	MDB_FMT_Z	/**< printf/scanf format modifier for size_t */
#define	Yu	MDB_PRIy(u)	/**< printf format for #mdb_size_t */
#define	Yd	MDB_PRIy(d)	/**< printf format for 'signed #mdb_size_t' */

#ifdef MDB_USE_SYSV_SEM
#define MNAME_LEN	(sizeof(int))
#else
#define MNAME_LEN	(sizeof(pthread_mutex_t))
#endif

/** Initial part of #MDB_env.%me_mutexname[].
 *	Changes to this code must be reflected in #MDB_LOCK_FORMAT.
 */
#ifdef _WIN32
#define MUTEXNAME_PREFIX		"Global\\MDB"
#elif defined MDB_USE_POSIX_SEM
#define MUTEXNAME_PREFIX		"/MDB"
#endif

/** @} */

#ifdef MDB_ROBUST_SUPPORTED
	/** Lock mutex, handle any error, set rc = result.
	 *	Return 0 on success, nonzero (not rc) on error.
	 */
#define LOCK_MUTEX(rc, env, mutex) \
	(((rc) = LOCK_MUTEX0(mutex)) && \
	 ((rc) = mdb_mutex_failed(env, mutex, rc)))
static int mdb_mutex_failed(MDB_env *env, mdb_mutexref_t mutex, int rc);
#else
#define LOCK_MUTEX(rc, env, mutex) ((rc) = LOCK_MUTEX0(mutex))
#define mdb_mutex_failed(env, mutex, rc) (rc)
#endif

#ifndef _WIN32
/**	A flag for opening a file and requesting synchronous data writes.
 *	This is only used when writing a meta page. It's not strictly needed;
 *	we could just do a normal write and then immediately perform a flush.
 *	But if this flag is available it saves us an extra system call.
 *
 *	@note If O_DSYNC is undefined but exists in /usr/include,
 * preferably set some compiler flag to get the definition.
 */
#ifndef MDB_DSYNC
# ifdef O_DSYNC
# define MDB_DSYNC	O_DSYNC
# else
# define MDB_DSYNC	O_SYNC
# endif
#endif
#endif

/** Function for flushing the data of a file. Define this to fsync
 *	if fdatasync() is not supported.
 */
#ifndef MDB_FDATASYNC
# define MDB_FDATASYNC	fdatasync
#endif

#ifndef MDB_MSYNC
# define MDB_MSYNC(addr,len,flags)	msync(addr,len,flags)
#endif

#ifndef MS_SYNC
#define	MS_SYNC	1
#endif

#ifndef MS_ASYNC
#define	MS_ASYNC	0
#endif

	/** A page number in the database.
	 *	Note that 64 bit page numbers are overkill, since pages themselves
	 *	already represent 12-13 bits of addressable memory, and the OS will
	 *	always limit applications to a maximum of 63 bits of address space.
	 *
	 *	@note In the #MDB_node structure, we only store 48 bits of this value,
	 *	which thus limits us to only 60 bits of addressable data.
	 */
typedef MDB_ID	pgno_t;

	/** A transaction ID.
	 *	See struct MDB_txn.mt_txnid for details.
	 */
typedef MDB_ID	txnid_t;

/** @defgroup debug	Debug Macros
 *	@{
 */
#ifndef MDB_DEBUG
	/**	Enable debug output.  Needs variable argument macros (a C99 feature).
	 *	Set this to 1 for copious tracing. Set to 2 to add dumps of all IDLs
	 *	read from and written to the database (used for free space management).
	 */
#define MDB_DEBUG 0
#endif

#if MDB_DEBUG
static int mdb_debug;
static txnid_t mdb_debug_start;

	/**	Print a debug message with printf formatting.
	 *	Requires double parenthesis around 2 or more args.
	 */
# define DPRINTF(args) ((void) ((mdb_debug) && DPRINTF0 args))
# define DPRINTF0(fmt, ...) \
	fprintf(stderr, "%s:%d " fmt "\n", mdb_func_, __LINE__, __VA_ARGS__)
#else
# define DPRINTF(args)	((void) 0)
#endif
	/**	Print a debug string.
	 *	The string is printed literally, with no format processing.
	 */
#define DPUTS(arg)	DPRINTF(("%s", arg))
	/** Debuging output value of a cursor DBI: Negative in a sub-cursor. */
#define DDBI(mc) \
	(((mc)->mc_flags & C_SUB) ? -(int)(mc)->mc_dbi : (int)(mc)->mc_dbi)
/** @} */

	/**	@brief The maximum size of a database page.
	 *
	 *	It is 32k or 64k, since value-PAGEBASE must fit in
	 *	#MDB_page.%mp_upper.
	 *
	 *	LMDB will use database pages < OS pages if needed.
	 *	That causes more I/O in write transactions: The OS must
	 *	know (read) the whole page before writing a partial page.
	 *
	 *	Note that we don't currently support Huge pages. On Linux,
	 *	regular data files cannot use Huge pages, and in general
	 *	Huge pages aren't actually pageable. We rely on the OS
	 *	demand-pager to read our data and page it out when memory
	 *	pressure from other processes is high. So until OSs have
	 *	actual paging support for Huge pages, they're not viable.
	 */
#define MAX_PAGESIZE	 (PAGEBASE ? 0x10000 : 0x8000)

	/** The minimum number of keys required in a database page.
	 *	Setting this to a larger value will place a smaller bound on the
	 *	maximum size of a data item. Data items larger than this size will
	 *	be pushed into overflow pages instead of being stored directly in
	 *	the B-tree node. This value used to default to 4. With a page size
	 *	of 4096 bytes that meant that any item larger than 1024 bytes would
	 *	go into an overflow page. That also meant that on average 2-3KB of
	 *	each overflow page was wasted space. The value cannot be lower than
	 *	2 because then there would no longer be a tree structure. With this
	 *	value, items larger than 2KB will go into overflow pages, and on
	 *	average only 1KB will be wasted.
	 */
#define MDB_MINKEYS	 2

	/**	A stamp that identifies a file as an LMDB file.
	 *	There's nothing special about this value other than that it is easily
	 *	recognizable, and it will reflect any byte order mismatches.
	 */
#define MDB_MAGIC	 0xBEEFC0DE

	/**	The version number for a database's datafile format. */
#define MDB_DATA_VERSION	 ((MDB_DEVEL) ? 999 : 2)
	/**	The version number for a database's lockfile format. */
#define MDB_LOCK_VERSION	 ((MDB_DEVEL) ? 999 : 2)
	/** Number of bits representing #MDB_LOCK_VERSION in #MDB_LOCK_FORMAT.
	 *	The remaining bits must leave room for #MDB_lock_desc.
	 */
#define MDB_LOCK_VERSION_BITS 12

	/**	@brief The max size of a key we can write, or 0 for computed max.
	 *
	 *	This macro should normally be left alone or set to 0.
	 *	Note that a database with big keys or dupsort data cannot be
	 *	reliably modified by a liblmdb which uses a smaller max.
	 *	The default is 511 for backwards compat, or 0 when #MDB_DEVEL.
	 *
	 *	Other values are allowed, for backwards compat.  However:
	 *	A value bigger than the computed max can break if you do not
	 *	know what you are doing, and liblmdb <= 0.9.10 can break when
	 *	modifying a DB with keys/dupsort data bigger than its max.
	 *
	 *	Data items in an #MDB_DUPSORT database are also limited to
	 *	this size, since they're actually keys of a sub-DB.  Keys and
	 *	#MDB_DUPSORT data items must fit on a node in a regular page.
	 */
#ifndef MDB_MAXKEYSIZE
#define MDB_MAXKEYSIZE	 ((MDB_DEVEL) ? 0 : 511)
#endif

	/**	The maximum size of a key we can write to the environment. */
#if MDB_MAXKEYSIZE
#define ENV_MAXKEY(env)	(MDB_MAXKEYSIZE)
#else
#define ENV_MAXKEY(env)	((env)->me_maxkey)
#endif

	/**	@brief The maximum size of a data item.
	 *
	 *	We only store a 32 bit value for node sizes.
	 */
#define MAXDATASIZE	0xffffffffUL

#if MDB_DEBUG
	/**	Key size which fits in a #DKBUF.
	 *	@ingroup debug
	 */
#define DKBUF_MAXKEYSIZE ((MDB_MAXKEYSIZE) > 0 ? (MDB_MAXKEYSIZE) : 511)
	/**	A key buffer.
	 *	@ingroup debug
	 *	This is used for printing a hex dump of a key's contents.
	 */
#define DKBUF	char kbuf[DKBUF_MAXKEYSIZE*2+1]
	/**	Display a key in hex.
	 *	@ingroup debug
	 *	Invoke a function to display a key in hex.
	 */
#define	DKEY(x)	mdb_dkey(x, kbuf)
#else
#define	DKBUF
#define DKEY(x)	0
#endif

	/** An invalid page number.
	 *	Mainly used to denote an empty tree.
	 */
#define P_INVALID	 (~(pgno_t)0)

	/** Test if the flags \b f are set in a flag word \b w. */
#define F_ISSET(w, f)	 (((w) & (f)) == (f))

	/** Round \b n up to an even number. */
#define EVEN(n)		(((n) + 1U) & -2) /* sign-extending -2 to match n+1U */

	/** Least significant 1-bit of \b n.  n must be of an unsigned type. */
#define LOW_BIT(n)		((n) & (-(n)))

	/** (log2(\b p2) % \b n), for p2 = power of 2 and 0 < n < 8. */
#define LOG2_MOD(p2, n)	(7 - 86 / ((p2) % ((1U<<(n))-1) + 11))
	/* Explanation: Let p2 = 2**(n*y + x), x<n and M = (1U<<n)-1. Now p2 =
	 * (M+1)**y * 2**x = 2**x (mod M). Finally "/" "happens" to return 7-x.
	 */

	/** Should be alignment of \b type. Ensure it is a power of 2. */
#define ALIGNOF2(type) \
	LOW_BIT(offsetof(struct { char ch_; type align_; }, align_))

	/**	Used for offsets within a single page.
	 *	Since memory pages are typically 4 or 8KB in size, 12-13 bits,
	 *	this is plenty.
	 */
typedef uint16_t	 indx_t;

typedef unsigned long long	mdb_hash_t;

	/**	Default size of memory map.
	 *	This is certainly too small for any actual applications. Apps should always set
	 *	the size explicitly using #mdb_env_set_mapsize().
	 */
#define DEFAULT_MAPSIZE	1048576

/**	@defgroup readers	Reader Lock Table
 *	Readers don't acquire any locks for their data access. Instead, they
 *	simply record their transaction ID in the reader table. The reader
 *	mutex is needed just to find an empty slot in the reader table. The
 *	slot's address is saved in thread-specific data so that subsequent read
 *	transactions started by the same thread need no further locking to proceed.
 *
 *	If #MDB_NOTLS is set, the slot address is not saved in thread-specific data.
 *
 *	No reader table is used if the database is on a read-only filesystem, or
 *	if #MDB_NOLOCK is set.
 *
 *	Since the database uses multi-version concurrency control, readers don't
 *	actually need any locking. This table is used to keep track of which
 *	readers are using data from which old transactions, so that we'll know
 *	when a particular old transaction is no longer in use. Old transactions
 *	that have discarded any data pages can then have those pages reclaimed
 *	for use by a later write transaction.
 *
 *	The lock table is constructed such that reader slots are aligned with the
 *	processor's cache line size. Any slot is only ever used by one thread.
 *	This alignment guarantees that there will be no contention or cache
 *	thrashing as threads update their own slot info, and also eliminates
 *	any need for locking when accessing a slot.
 *
 *	A writer thread will scan every slot in the table to determine the oldest
 *	outstanding reader transaction. Any freed pages older than this will be
 *	reclaimed by the writer. The writer doesn't use any locks when scanning
 *	this table. This means that there's no guarantee that the writer will
 *	see the most up-to-date reader info, but that's not required for correct
 *	operation - all we need is to know the upper bound on the oldest reader,
 *	we don't care at all about the newest reader. So the only consequence of
 *	reading stale information here is that old pages might hang around a
 *	while longer before being reclaimed. That's actually good anyway, because
 *	the longer we delay reclaiming old pages, the more likely it is that a
 *	string of contiguous pages can be found after coalescing old pages from
 *	many old transactions together.
 *	@{
 */
	/**	Number of slots in the reader table.
	 *	This value was chosen somewhat arbitrarily. 126 readers plus a
	 *	couple mutexes fit exactly into 8KB on my development machine.
	 *	Applications should set the table size using #mdb_env_set_maxreaders().
	 */
#define DEFAULT_READERS	126

	/**	The size of a CPU cache line in bytes. We want our lock structures
	 *	aligned to this size to avoid false cache line sharing in the
	 *	lock table.
	 *	This value works for most CPUs. For Itanium this should be 128.
	 */
#ifndef CACHELINE
#define CACHELINE	64
#endif

	/**	The information we store in a single slot of the reader table.
	 *	In addition to a transaction ID, we also record the process and
	 *	thread ID that owns a slot, so that we can detect stale information,
	 *	e.g. threads or processes that went away without cleaning up.
	 *	@note We currently don't check for stale records. We simply re-init
	 *	the table when we know that we're the only process opening the
	 *	lock file.
	 */
typedef struct MDB_rxbody {
	/**	Current Transaction ID when this transaction began, or (txnid_t)-1.
	 *	Multiple readers that start at the same time will probably have the
	 *	same ID here. Again, it's not important to exclude them from
	 *	anything; all we need to know is which version of the DB they
	 *	started from so we can avoid overwriting any data used in that
	 *	particular version.
	 */
	volatile txnid_t		mrb_txnid;
	/** The process ID of the process owning this reader txn. */
	volatile MDB_PID_T	mrb_pid;
	/** The thread ID of the thread owning this txn. */
	volatile MDB_THR_T	mrb_tid;
} MDB_rxbody;

	/** The actual reader record, with cacheline padding. */
typedef struct MDB_reader {
	union {
		MDB_rxbody mrx;
		/** shorthand for mrb_txnid */
#define	mr_txnid	mru.mrx.mrb_txnid
#define	mr_pid	mru.mrx.mrb_pid
#define	mr_tid	mru.mrx.mrb_tid
		/** cache line alignment */
		char pad[(sizeof(MDB_rxbody)+CACHELINE-1) & ~(CACHELINE-1)];
	} mru;
} MDB_reader;

	/** The header for the reader table.
	 *	The table resides in a memory-mapped file. (This is a different file
	 *	than is used for the main database.)
	 *
	 *	For POSIX the actual mutexes reside in the shared memory of this
	 *	mapped file. On Windows, mutexes are named objects allocated by the
	 *	kernel; we store the mutex names in this mapped file so that other
	 *	processes can grab them. This same approach is also used on
	 *	MacOSX/Darwin (using named semaphores) since MacOSX doesn't support
	 *	process-shared POSIX mutexes. For these cases where a named object
	 *	is used, the object name is derived from a 64 bit FNV hash of the
	 *	environment pathname. As such, naming collisions are extremely
	 *	unlikely. If a collision occurs, the results are unpredictable.
	 */
typedef struct MDB_txbody {
		/** Stamp identifying this as an LMDB file. It must be set
		 *	to #MDB_MAGIC. */
	uint32_t	mtb_magic;
		/** Format of this lock file. Must be set to #MDB_LOCK_FORMAT. */
	uint32_t	mtb_format;
		/**	The ID of the last transaction committed to the database.
		 *	This is recorded here only for convenience; the value can always
		 *	be determined by reading the main database meta pages.
		 */
	volatile txnid_t		mtb_txnid;
		/** The number of slots that have been used in the reader table.
		 *	This always records the maximum count, it is not decremented
		 *	when readers release their slots.
		 */
	volatile unsigned	mtb_numreaders;
#if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
		/** Binary form of names of the reader/writer locks */
	mdb_hash_t			mtb_mutexid;
#elif defined(MDB_USE_SYSV_SEM)
	int 	mtb_semid;
	int		mtb_rlocked;
#else
		/** Mutex protecting access to this table.
		 *	This is the reader table lock used with LOCK_MUTEX().
		 */
	mdb_mutex_t	mtb_rmutex;
#endif
} MDB_txbody;

	/** The actual reader table definition. */
typedef struct MDB_txninfo {
	union {
		MDB_txbody mtb;
#define mti_magic	mt1.mtb.mtb_magic
#define mti_format	mt1.mtb.mtb_format
#define mti_rmutex	mt1.mtb.mtb_rmutex
#define mti_txnid	mt1.mtb.mtb_txnid
#define mti_numreaders	mt1.mtb.mtb_numreaders
#define mti_mutexid	mt1.mtb.mtb_mutexid
#ifdef MDB_USE_SYSV_SEM
#define	mti_semid	mt1.mtb.mtb_semid
#define	mti_rlocked	mt1.mtb.mtb_rlocked
#endif
		char pad[(sizeof(MDB_txbody)+CACHELINE-1) & ~(CACHELINE-1)];
	} mt1;
#if !(defined(_WIN32) || defined(MDB_USE_POSIX_SEM))
	union {
#ifdef MDB_USE_SYSV_SEM
		int mt2_wlocked;
#define mti_wlocked	mt2.mt2_wlocked
#else
		mdb_mutex_t	mt2_wmutex;
#define mti_wmutex	mt2.mt2_wmutex
#endif
		char pad[(MNAME_LEN+CACHELINE-1) & ~(CACHELINE-1)];
	} mt2;
#endif
	MDB_reader	mti_readers[1];
} MDB_txninfo;

	/** Lockfile format signature: version, features and field layout */
#define MDB_LOCK_FORMAT \
	((uint32_t)         \
	 (((MDB_LOCK_VERSION) % (1U << MDB_LOCK_VERSION_BITS)) \
	  + MDB_lock_desc     * (1U << MDB_LOCK_VERSION_BITS)))

	/** Lock type and layout. Values 0-119. _WIN32 implies #MDB_PIDLOCK.
	 *	Some low values are reserved for future tweaks.
	 */
#ifdef _WIN32
# define MDB_LOCK_TYPE	(0 + ALIGNOF2(mdb_hash_t)/8 % 2)
#elif defined MDB_USE_POSIX_SEM
# define MDB_LOCK_TYPE	(4 + ALIGNOF2(mdb_hash_t)/8 % 2)
#elif defined MDB_USE_SYSV_SEM
# define MDB_LOCK_TYPE	(8)
#elif defined MDB_USE_POSIX_MUTEX
/* We do not know the inside of a POSIX mutex and how to check if mutexes
 * used by two executables are compatible. Just check alignment and size.
 */
# define MDB_LOCK_TYPE	(10 + \
		LOG2_MOD(ALIGNOF2(pthread_mutex_t), 5) + \
		sizeof(pthread_mutex_t) / 4U % 22 * 5)
#endif

enum {
	/** Magic number for lockfile layout and features.
	 *
	 *  This *attempts* to stop liblmdb variants compiled with conflicting
	 *	options from using the lockfile at the same time and thus breaking
	 *	it.  It describes locking types, and sizes and sometimes alignment
	 *	of the various lockfile items.
	 *
	 *	The detected ranges are mostly guesswork, or based simply on how
	 *	big they could be without using more bits.  So we can tweak them
	 *	in good conscience when updating #MDB_LOCK_VERSION.
	 */
	MDB_lock_desc =
	/* Default CACHELINE=64 vs. other values (have seen mention of 32-256) */
	(CACHELINE==64 ? 0 : 1 + LOG2_MOD(CACHELINE >> (CACHELINE>64), 5))
	+ 6  * (sizeof(MDB_PID_T)/4 % 3)    /* legacy(2) to word(4/8)? */
	+ 18 * (sizeof(pthread_t)/4 % 5)    /* can be struct{id, active data} */
	+ 90 * (sizeof(MDB_txbody) / CACHELINE % 3)
	+ 270 * (MDB_LOCK_TYPE % 120)
	/* The above is < 270*120 < 2**15 */
	+ ((sizeof(txnid_t) == 8) << 15)    /* 32bit/64bit */
	+ ((sizeof(MDB_reader) > CACHELINE) << 16)
	/* Not really needed - implied by MDB_LOCK_TYPE != (_WIN32 locking) */
	+ (((MDB_PIDLOCK) != 0)   << 17)
	/* 18 bits total: Must be <= (32 - MDB_LOCK_VERSION_BITS). */
};
/** @} */

/** Common header for all page types. The page type depends on #mp_flags.
 *
 * #P_BRANCH and #P_LEAF pages have unsorted '#MDB_node's at the end, with
 * sorted #mp_ptrs[] entries referring to them. Exception: #P_LEAF2 pages
 * omit mp_ptrs and pack sorted #MDB_DUPFIXED values after the page header.
 *
 * #P_OVERFLOW records occupy one or more contiguous pages where only the
 * first has a page header. They hold the real data of #F_BIGDATA nodes,
 * and the node stores the pgno and number of pages used by the record.
 *
 * #P_SUBP sub-pages are small leaf "pages" with duplicate data.
 * A node with flag #F_DUPDATA but not #F_SUBDATA contains a sub-page.
 * (Duplicate data can also go in sub-databases, which use normal pages.)
 *
 * #P_META pages contain #MDB_meta, the start point of an LMDB snapshot.
 *
 * Each non-metapage up to #MDB_meta.%mm_last_pg is reachable exactly once
 * in the snapshot: Either used by a database or listed in a freeDB record.
 */
typedef struct MDB_page_header {
#define	mp_pgno	mp_p.p_pgno
#define	mp_next	mp_p.p_next
	union {
		pgno_t		p_pgno;	/**< page number */
		struct MDB_page *p_next; /**< for in-memory list of freed pages */
	} mh_p;

	/** If page is clean: snapshot txnid, dirty: txn workid, metapage: unused.
	 *
	 *	The value indicates which snapshot/#MDB_txn the page belongs to.
	 *	Tested with #IS_MUTABLE(), #IS_WRITABLE(), #IS_DIRTY_NW().
	 *
	 *	(clean page).mp_txnid == txnid of creator < txnid of later txns.
	 *	(dirty page).mp_txnid >= mt_workid of creator txn.
	 *	(dirty page).mt_txnid  < mt_workid of children of creator txn.
	 *
	 *	Thus an #MDB_txn can write to pages with mp_txnid >= txn.mt_workid.
	 *	A page with smaller mp_txnid is dirty in an ancestor txn or clean.
	 *
	 *	Non-#MDB_WRITEMAP sets txn.mt_workid > txn.mt_txnid, to tell apart
	 *	spilled and dirty pages.  WRITEMAP sets mt_workid = mt_txnid, since
	 *	it does not copy/spill pages.  Thus (page.mp_txnid == txn.mt_txnid)
	 *	says "spilled page" without WRITEMAP, "dirty page" with WRITEMAP.
	 *
	 *	Finally, ((dirty page).mp_txnid & #MDB_PGTXNID_FLAGMASK) can be used
	 *	for flags with non-WRITEMAP; it keeps low bits in workid = 0.
	 */
	txnid_t		mh_txnid;

	uint16_t	mh_pad;			/**< key size if this is a LEAF2 page */
/**	@defgroup mdb_page	Page Flags
 *	@ingroup internal
 *	Flags for the page headers.
 *	@{
 */
#define	P_BRANCH	 0x01		/**< branch page */
#define	P_LEAF		 0x02		/**< leaf page */
#define	P_OVERFLOW	 0x04		/**< overflow page */
#define	P_META		 0x08		/**< meta page */
#define	P_LEAF2		 0x20		/**< for #MDB_DUPFIXED records */
#define	P_SUBP		 0x40		/**< for #MDB_DUPSORT sub-pages */
#define	P_DIRTY_OVF	 0x2000		/**< page has dirty overflow nodes */
#define	P_LOOSE		 0x4000		/**< page was dirtied then freed, can be reused */
#define	P_KEEP		 0x8000		/**< leave this page alone during spill */
/** Persistent flags for page administration rather than page contents */
#define	P_ADM_FLAGS	 0 /* later... */
/** @} */
	uint16_t	mh_flags;		/**< @ref mdb_page */
#define mp_lower	mp_pb.pb.pb_lower
#define mp_upper	mp_pb.pb.pb_upper
#define mp_pages	mp_pb.pb_pages
	union {
		struct {
			indx_t		pb_lower;		/**< lower bound of free space */
			indx_t		pb_upper;		/**< upper bound of free space */
		} pb;
		uint32_t	pb_pages;	/**< number of overflow pages */
	} mh_pb;
} MDB_page_header;

typedef struct MDB_page {
	MDB_page_header	mp_hdr;
#define mp_p	mp_hdr.mh_p
#define mp_txnid	mp_hdr.mh_txnid
#define mp_pad		mp_hdr.mh_pad
#define mp_flags	mp_hdr.mh_flags
#define mp_pb		mp_hdr.mh_pb
	indx_t		mp_ptrs[1];		/**< dynamic size */
} MDB_page;

	/** Size of the page header, excluding dynamic data at the end */
#define PAGEHDRSZ	 ((unsigned)sizeof(MDB_page_header))

	/** Address of first usable data byte in a page, after the header */
#define METADATA(p)	 ((void *)((char *)(p) + PAGEHDRSZ))

	/** ITS#7713, change PAGEBASE to handle 65536 byte pages */
#define	PAGEBASE	((MDB_DEVEL) ? PAGEHDRSZ : 0)

	/** Number of nodes on a page */
#define NUMKEYS(p)	 (((p)->mp_lower - (PAGEHDRSZ-PAGEBASE)) >> 1)

	/** The amount of space remaining in the page */
#define SIZELEFT(p)	 (indx_t)((p)->mp_upper - (p)->mp_lower)

	/** The percentage of space used in the page, in tenths of a percent. */
#define PAGEFILL(env, p) (1000L * ((env)->me_psize - PAGEHDRSZ - SIZELEFT(p)) / \
				((env)->me_psize - PAGEHDRSZ))
	/** The minimum page fill factor, in tenths of a percent.
	 *	Pages emptier than this are candidates for merging.
	 */
#define FILL_THRESHOLD	 250

	/** Test if a page is a leaf page */
#define IS_LEAF(p)	 F_ISSET((p)->mp_flags, P_LEAF)
	/** Test if a page is a LEAF2 page */
#define IS_LEAF2(p)	 F_ISSET((p)->mp_flags, P_LEAF2)
	/** Test if a page is a branch page */
#define IS_BRANCH(p)	 F_ISSET((p)->mp_flags, P_BRANCH)
	/** Test if a page is an overflow page */
#define IS_OVERFLOW(p)	 F_ISSET((p)->mp_flags, P_OVERFLOW)
	/** Test if a page is a sub page */
#define IS_SUBP(p)	 F_ISSET((p)->mp_flags, P_SUBP)

	/** Test if (this non-sub page is dirty && env is non-#MDB_WRITEMAP) */
#define IS_DIRTY_NW(txn, p)	((p)->mp_txnid > (txn)->mt_txnid)
	/** Test if this non-sub page belongs to the current snapshot */
#define IS_MUTABLE(txn, p)	((p)->mp_txnid >= (txn)->mt_txnid)
	/** Test if this non-sub page is writable in this txn (not an ancestor) */
#define IS_WRITABLE(txn, p)	((p)->mp_txnid >= (txn)->mt_workid)

	/** Info about overflow page, stored in an F_BIGDATA node */
typedef struct MDB_ovpage {
	pgno_t	op_pgno;
	txnid_t	op_txnid;
	mdb_size_t	op_pages;
} MDB_ovpage;

#if OVERFLOW_NOTYET
	/** Header for a dirty overflow page in memory */
typedef struct MDB_dovpage {
	MDB_page_header mp_hdr;
	void	*mp_ptr;
} MDB_dovpage;
#endif

	/** The number of overflow pages needed to store the given size. */
#define OVPAGES(size, psize)	((PAGEHDRSZ-1 + (size)) / (psize) + 1)

	/** Link in #MDB_txn.%mt_loose_pgs list.
	 *  Kept outside the page header, which is needed when reusing the page.
	 */
#define NEXT_LOOSE_PAGE(p)		(*(MDB_page **)((p) + 2))

	/** Mark the page as writable by this txn */
#ifndef MDB_TEST
#define SET_PGTXNID(txn, mp)	((mp)->mp_txnid = (txn)->mt_workid)
#else
#define SET_PGTXNID(txn, mp) \
	((mp)->mp_txnid = (txn)->mt_workid \
	 /* random unused "flags" added when not WRITEMAP for debugging */ \
	 | (((txn)->mt_flags & MDB_TXN_WRITEMAP) ? 0 : \
		(MDB_RAND((size_t)(txn)) >> (32-MDB_PGTXNID_FLAGBITS))))
#define MDB_RAND(x) (mdb_rnd = (mdb_rnd + (unsigned)(x)) * 987654321 + 54321)
static volatile unsigned mdb_rnd;
#endif

	/** mp_txnid bits reserved in dirty pages for flags.
	 *	TODO: For future code with header-free ovpages, if we omit mp_flags
	 *	from the "header" kept elsewhere.  Otherwise, drop this code.
	 */
#define MDB_PGTXNID_FLAGBITS	4
#define MDB_PGTXNID_STEP		((txnid_t)1 << MDB_PGTXNID_FLAGBITS)
#define MDB_PGTXNID_FLAGMASK	(MDB_PGTXNID_STEP-1)

	/** Header for a single key/data pair within a page.
	 * Used in pages of type #P_BRANCH and #P_LEAF without #P_LEAF2.
	 * We guarantee 2-byte alignment for 'MDB_node's.
	 *
	 * #mn_lo and #mn_hi are used for data size on leaf nodes, and for child
	 * pgno on branch nodes.  On 64 bit platforms, #mn_flags is also used
	 * for pgno.  (Branch nodes have no flags).  Lo and hi are in host byte
	 * order in case some accesses can be optimized to 32-bit word access.
	 *
	 * Leaf node flags describe node contents.  #F_BIGDATA says the node's
	 * data part is an MDB_ovpage struct pointing to a page with actual data.
	 * #F_DUPDATA and #F_SUBDATA can be combined giving duplicate data in
	 * a sub-page/sub-database, and named databases (just #F_SUBDATA).
	 */
typedef struct MDB_node {
	/** part of data size or pgno
	 *	@{ */
#if BYTE_ORDER == LITTLE_ENDIAN
	unsigned short	mn_lo, mn_hi;
#else
	unsigned short	mn_hi, mn_lo;
#endif
	/** @} */
/** @defgroup mdb_node Node Flags
 *	@ingroup internal
 *	Flags for node headers.
 *	@{
 */
#define F_BIGDATA	 0x01			/**< data put on overflow page */
#define F_SUBDATA	 0x02			/**< data is a sub-database */
#define F_DUPDATA	 0x04			/**< data has duplicates */

/** valid flags for #mdb_node_add() */
#define	NODE_ADD_FLAGS	(F_DUPDATA|F_SUBDATA|MDB_RESERVE|MDB_APPEND)

/** @} */
	unsigned short	mn_flags;		/**< @ref mdb_node */
	unsigned short	mn_ksize;		/**< key size */
	char		mn_data[1];			/**< key and data are appended here */
} MDB_node;

	/** Size of the node header, excluding dynamic data at the end */
#define NODESIZE	 offsetof(MDB_node, mn_data)

	/** Bit position of top word in page number, for shifting mn_flags */
#define PGNO_TOPWORD ((pgno_t)-1 > 0xffffffffu ? 32 : 0)

	/** Size of a node in a branch page with a given key.
	 *	This is just the node header plus the key, there is no data.
	 */
#define INDXSIZE(k)	 (NODESIZE + ((k) == NULL ? 0 : (k)->mv_size))

	/** Size of a node in a leaf page with a given key and data.
	 *	This is node header plus key plus data size.
	 */
#define LEAFSIZE(k, d)	 (NODESIZE + (k)->mv_size + (d)->mv_size)

	/** Address of node \b i in page \b p */
#define NODEPTR(p, i)	 ((MDB_node *)((char *)(p) + (p)->mp_ptrs[i] + PAGEBASE))

	/** Address of the key for the node */
#define NODEKEY(node)	 (void *)((node)->mn_data)

	/** Address of the data for a node */
#define NODEDATA(node)	 (void *)((char *)(node)->mn_data + (node)->mn_ksize)

	/** Get the page number pointed to by a branch node */
#define NODEPGNO(node) \
	((node)->mn_lo | ((pgno_t) (node)->mn_hi << 16) | \
	 (PGNO_TOPWORD ? ((pgno_t) (node)->mn_flags << PGNO_TOPWORD) : 0))