mdb.c

/** @file mdb.c
 *	@brief memory-mapped database library
 *
 *	A Btree-based database management library modeled loosely on the
 *	BerkeleyDB API, but much simplified.
 */
/*
 * Copyright 2011 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.
 */
#include <sys/types.h>
#include <sys/stat.h>
#include <sys/param.h>
#ifdef _WIN32
#include <windows.h>
#else
#include <sys/uio.h>
#include <sys/mman.h>
#ifdef HAVE_SYS_FILE_H
#include <sys/file.h>
#endif
#include <fcntl.h>
#endif

#include <assert.h>
#include <errno.h>
#include <stddef.h>
#include <stdint.h>
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <time.h>
#include <unistd.h>

#ifndef _WIN32
#include <pthread.h>
#endif

#include "mdb.h"
#include "midl.h"

/** @defgroup internal	MDB Internals
 *	@{
 */
/** @defgroup compat	Windows Compatibility Macros
 *	@{
 */
#ifdef _WIN32
#define pthread_t	DWORD
#define pthread_mutex_t	HANDLE
#define pthread_key_t	DWORD
#define pthread_self()	GetCurrentThreadId()
#define pthread_key_create(x,y)	*(x) = TlsAlloc()
#define pthread_key_delete(x)	TlsFree(x)
#define pthread_getspecific(x)	TlsGetValue(x)
#define pthread_setspecific(x,y)	TlsSetValue(x,y)
#define pthread_mutex_unlock(x)	ReleaseMutex(x)
#define pthread_mutex_lock(x)	WaitForSingleObject(x, INFINITE)
#define LOCK_MUTEX_R(env)	pthread_mutex_lock(env->me_rmutex)
#define UNLOCK_MUTEX_R(env)	pthread_mutex_unlock(env->me_rmutex)
#define LOCK_MUTEX_W(env)	pthread_mutex_lock(env->me_wmutex)
#define UNLOCK_MUTEX_W(env)	pthread_mutex_unlock(env->me_wmutex)
#define getpid()	GetCurrentProcessId()
#define	fdatasync(fd)	!FlushFileBuffers(fd)
#define	ErrCode()	GetLastError()
#define GetPageSize(x)	{SYSTEM_INFO si; GetSystemInfo(&si); (x) = si.dwPageSize;}
#define	close(fd)	CloseHandle(fd)
#define	munmap(ptr,len)	UnmapViewOfFile(ptr)
#else
#define LOCK_MUTEX_R(env)	pthread_mutex_lock(&env->me_txns->mti_mutex)
#define UNLOCK_MUTEX_R(env)	pthread_mutex_unlock(&env->me_txns->mti_mutex)
#define LOCK_MUTEX_W(env)	pthread_mutex_lock(&env->me_txns->mti_wmutex)
#define UNLOCK_MUTEX_W(env)	pthread_mutex_unlock(&env->me_txns->mti_wmutex)
#define	ErrCode()	errno
#define	HANDLE	int
#define INVALID_HANDLE_VALUE	-1
#define	GetPageSize(x)	(x) = sysconf(_SC_PAGE_SIZE)
#endif

/** @} */

#ifndef _WIN32
/* Note: If O_DSYNC is undefined but exists in /usr/include,
 * preferably set some compiler flag to get the definition.
 * Otherwise compile with the less efficient -DMDB_DSYNC=O_SYNC.
 */
#ifndef MDB_DSYNC
# define MDB_DSYNC	O_DSYNC
#endif
#endif

typedef ULONG		pgno_t;

#ifndef DEBUG
#define DEBUG 0
#endif

#if !(__STDC_VERSION__ >= 199901L || defined(__GNUC__))
# define DPRINTF	(void)	/* Vararg macros may be unsupported */
#elif DEBUG
# define DPRINTF(fmt, ...)	/* Requires 2 or more args */ \
	fprintf(stderr, "%s:%d:(%p) " fmt "\n", __func__, __LINE__, pthread_self(), __VA_ARGS__)
#else
# define DPRINTF(fmt, ...)	((void) 0)
#endif
#define DPUTS(arg)	DPRINTF("%s", arg)

#define PAGESIZE	 4096
#define MDB_MINKEYS	 2
#define MDB_MAGIC	 0xBEEFC0DE
#define MDB_VERSION	 1
#define MAXKEYSIZE	 511
#if DEBUG
#define	KBUF	(MAXKEYSIZE*2+1)
#define DKBUF	char kbuf[KBUF]
#define	DKEY(x)	mdb_dkey(x, kbuf)
#else
#define	DKBUF
#define DKEY(x)
#endif

/* The DB view is always consistent because all writes are wrapped in
 * the wmutex. Finer-grained locks aren't necessary.
 */
#ifndef	LAZY_LOCKS
#define	LAZY_LOCKS	1
#endif
#if	LAZY_LOCKS
#define	LAZY_MUTEX_LOCK(x)
#define	LAZY_MUTEX_UNLOCK(x)
#define	LAZY_RWLOCK_UNLOCK(x)
#define	LAZY_RWLOCK_WRLOCK(x)
#define	LAZY_RWLOCK_RDLOCK(x)
#define	LAZY_RWLOCK_DEF(x)
#define	LAZY_RWLOCK_INIT(x,y)
#define	LAZY_RWLOCK_DESTROY(x)
#else
#define	LAZY_MUTEX_LOCK(x)		pthread_mutex_lock(x)
#define	LAZY_MUTEX_UNLOCK(x)	pthread_mutex_unlock(x)
#define	LAZY_RWLOCK_UNLOCK(x)	pthread_rwlock_unlock(x)
#define	LAZY_RWLOCK_WRLOCK(x)	pthread_rwlock_wrlock(x)
#define	LAZY_RWLOCK_RDLOCK(x)	pthread_rwlock_rdlock(x)
#define	LAZY_RWLOCK_DEF(x)		pthread_rwlock_t	x
#define	LAZY_RWLOCK_INIT(x,y)	pthread_rwlock_init(x,y)
#define	LAZY_RWLOCK_DESTROY(x)	pthread_rwlock_destroy(x)
#endif

#define P_INVALID	 (~0UL)

#define F_ISSET(w, f)	 (((w) & (f)) == (f))

typedef uint16_t	 indx_t;

#define DEFAULT_READERS	126
#define DEFAULT_MAPSIZE	1048576

/* Lock descriptor stuff */
#ifndef CACHELINE
#define CACHELINE	64	/* most CPUs. Itanium uses 128 */
#endif

typedef struct MDB_rxbody {
	ULONG		mrb_txnid;
	pid_t		mrb_pid;
	pthread_t	mrb_tid;
} MDB_rxbody;

typedef struct MDB_reader {
	union {
		MDB_rxbody mrx;
#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;

typedef struct MDB_txbody {
	uint32_t	mtb_magic;
	uint32_t	mtb_version;
/* For POSIX the actual mutexes reside in shared memory.
 * On Windows, mutexes are allocated by the kernel; we store
 * the name in shared memory so that other processes can
 * grab them.
 */
#ifdef _WIN32
	char	mtb_rmname[32];
#else
	pthread_mutex_t	mtb_mutex;
#endif
	ULONG		mtb_txnid;
	uint32_t	mtb_numreaders;
	uint32_t	mtb_me_toggle;
} MDB_txbody;

typedef struct MDB_txninfo {
	union {
		MDB_txbody mtb;
#define mti_magic	mt1.mtb.mtb_magic
#define mti_version	mt1.mtb.mtb_version
#define mti_mutex	mt1.mtb.mtb_mutex
#define mti_rmname	mt1.mtb.mtb_rmname
#define mti_txnid	mt1.mtb.mtb_txnid
#define mti_numreaders	mt1.mtb.mtb_numreaders
#define mti_me_toggle	mt1.mtb.mtb_me_toggle
		char pad[(sizeof(MDB_txbody)+CACHELINE-1) & ~(CACHELINE-1)];
	} mt1;
	union {
#ifdef _WIN32
		char mt2_wmname[32];
#define	mti_wmname	mt2.mt2_wmname
#else
		pthread_mutex_t	mt2_wmutex;
#define mti_wmutex	mt2.mt2_wmutex
#endif
		char pad[(sizeof(pthread_mutex_t)+CACHELINE-1) & ~(CACHELINE-1)];
	} mt2;
	MDB_reader	mti_readers[1];
} MDB_txninfo;

/* Common header for all page types. Overflow pages
 * occupy a number of contiguous pages with no
 * headers on any page after the first.
 */
typedef struct MDB_page {		/* represents a page of storage */
#define	mp_pgno		mp_p.p_pgno
#define	mp_next		mp_p.p_align
	union padded {
		pgno_t		p_pgno;		/* page number */
		void *		p_align;	/* for IL32P64 */
	} mp_p;
#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_DIRTY		 0x10		/* dirty page */
#define	P_LEAF2		 0x20		/* DB with small, fixed size keys and no data */
	uint32_t	mp_flags;
#define mp_lower	mp_pb.pb.pb_lower
#define mp_upper	mp_pb.pb.pb_upper
#define mp_pages	mp_pb.pb_pages
	union page_bounds {
		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 */
	} mp_pb;
	indx_t		mp_ptrs[1];		/* dynamic size */
} MDB_page;

#define PAGEHDRSZ	 ((unsigned) offsetof(MDB_page, mp_ptrs))

#define NUMKEYS(p)	 (((p)->mp_lower - PAGEHDRSZ) >> 1)
#define SIZELEFT(p)	 (indx_t)((p)->mp_upper - (p)->mp_lower)
#define PAGEFILL(env, p) (1000L * ((env)->me_psize - PAGEHDRSZ - SIZELEFT(p)) / \
				((env)->me_psize - PAGEHDRSZ))
#define IS_LEAF(p)	 F_ISSET((p)->mp_flags, P_LEAF)
#define IS_LEAF2(p)	 F_ISSET((p)->mp_flags, P_LEAF2)
#define IS_BRANCH(p)	 F_ISSET((p)->mp_flags, P_BRANCH)
#define IS_OVERFLOW(p)	 F_ISSET((p)->mp_flags, P_OVERFLOW)

#define OVPAGES(size, psize)	((PAGEHDRSZ-1 + (size)) / (psize) + 1)

typedef struct MDB_db {
	uint32_t	md_pad;		/* also ksize for LEAF2 pages */
	uint16_t	md_flags;
	uint16_t	md_depth;
	ULONG		md_branch_pages;
	ULONG		md_leaf_pages;
	ULONG		md_overflow_pages;
	ULONG		md_entries;
	pgno_t		md_root;
} MDB_db;

#define	FREE_DBI	0
#define	MAIN_DBI	1

typedef struct MDB_meta {			/* meta (footer) page content */
	uint32_t	mm_magic;
	uint32_t	mm_version;
	void		*mm_address;		/* address for fixed mapping */
	size_t		mm_mapsize;			/* size of mmap region */
	MDB_db		mm_dbs[2];			/* first is free space, 2nd is main db */
#define	mm_psize	mm_dbs[0].md_pad
#define	mm_flags	mm_dbs[0].md_flags
	pgno_t		mm_last_pg;			/* last used page in file */
	ULONG		mm_txnid;			/* txnid that committed this page */
} MDB_meta;

typedef struct MDB_oldpages {
	struct MDB_oldpages *mo_next;
	ULONG		mo_txnid;
	pgno_t		mo_pages[1];	/* dynamic */
} MDB_oldpages;

static MDB_page *mdb_alloc_page(MDB_cursor *mc, int num);
static int 		mdb_touch(MDB_cursor *mc);

/* Enough space for 2^32 nodes with minimum of 2 keys per node. I.e., plenty.
 * At 4 keys per node, enough for 2^64 nodes, so there's probably no need to
 * raise this on a 64 bit machine.
 */
#define CURSOR_STACK		 32

struct MDB_xcursor;

struct MDB_cursor {
	struct MDB_xcursor	*mc_xcursor;
	MDB_txn		*mc_txn;
	MDB_dbi		mc_dbi;
	unsigned short 	mc_snum;		/* number of pushed pages */
	unsigned short	mc_top;		/* index of top page, mc_snum-1 */
	unsigned int	mc_flags;
#define C_INITIALIZED	0x01
#define C_EOF	0x02
#define C_XDIRTY	0x04
	MDB_page	*mc_pg[CURSOR_STACK];	/* stack of pushed pages */
	indx_t		mc_ki[CURSOR_STACK];	/* stack of page indices */
};

#define METADATA(p)	 ((void *)((char *)(p) + PAGEHDRSZ))

/* We guarantee 2-byte alignment for nodes */
typedef struct MDB_node {
	/* lo and hi are used for data size on leaf nodes and for
	 * child pgno on branch nodes. On 64 bit platforms, flags
	 * is also used for pgno. (branch nodes ignore flags)
	 */
	unsigned short	mn_lo;
	unsigned short	mn_hi;
	unsigned short	mn_flags;
#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 */
	unsigned short	mn_ksize;				/* key size */
	char		mn_data[1];
} MDB_node;

typedef struct MDB_dbx {
	MDB_val		md_name;
	MDB_cmp_func	*md_cmp;		/* user compare function */
	MDB_cmp_func	*md_dcmp;		/* user dupsort function */
	MDB_rel_func	*md_rel;		/* user relocate function */
	MDB_dbi	md_parent;
	unsigned int	md_dirty;
} MDB_dbx;

struct MDB_txn {
	pgno_t		mt_next_pgno;	/* next unallocated page */
	ULONG		mt_txnid;
	MDB_env		*mt_env;	
	pgno_t		*mt_free_pgs;	/* this is an IDL */
	union {
		ID2L	dirty_list;	/* modified pages */
		MDB_reader	*reader;
	} mt_u;
	MDB_dbx		*mt_dbxs;		/* array */
	MDB_db		*mt_dbs;
	unsigned int	mt_numdbs;

#define MDB_TXN_RDONLY		0x01		/* read-only transaction */
#define MDB_TXN_ERROR		0x02		/* an error has occurred */
	unsigned int	mt_flags;
	unsigned int	mt_toggle;
};

/* Context for sorted-dup records */
typedef struct MDB_xcursor {
	MDB_cursor mx_cursor;
	MDB_txn mx_txn;
	MDB_dbx	mx_dbxs[4];
	MDB_db	mx_dbs[4];
} MDB_xcursor;

struct MDB_env {
	HANDLE		me_fd;
	HANDLE		me_lfd;
	HANDLE		me_mfd;			/* just for writing the meta pages */
#define	MDB_FATAL_ERROR	0x80000000U
	uint32_t 	me_flags;
	uint32_t	me_extrapad;	/* unused for now */
	unsigned int	me_maxreaders;
	unsigned int	me_numdbs;
	unsigned int	me_maxdbs;
	char		*me_path;
	char		*me_map;
	MDB_txninfo	*me_txns;
	MDB_meta	*me_metas[2];
	MDB_txn		*me_txn;		/* current write transaction */
	size_t		me_mapsize;
	off_t		me_size;		/* current file size */
	pgno_t		me_maxpg;		/* me_mapsize / me_psize */
	unsigned int	me_psize;
	unsigned int	me_db_toggle;
	MDB_dbx		*me_dbxs;		/* array */
	MDB_db		*me_dbs[2];
	MDB_oldpages *me_pghead;
	pthread_key_t	me_txkey;	/* thread-key for readers */
	MDB_page	*me_dpages;
	pgno_t		me_free_pgs[MDB_IDL_UM_SIZE];
	ID2			me_dirty_list[MDB_IDL_UM_SIZE];
	LAZY_RWLOCK_DEF(me_dblock);
#ifdef _WIN32
	HANDLE		me_rmutex;		/* Windows mutexes don't reside in shared mem */
	HANDLE		me_wmutex;
#endif
};

#define NODESIZE	 offsetof(MDB_node, mn_data)

#define INDXSIZE(k)	 (NODESIZE + ((k) == NULL ? 0 : (k)->mv_size))
#define LEAFSIZE(k, d)	 (NODESIZE + (k)->mv_size + (d)->mv_size)
#define NODEPTR(p, i)	 ((MDB_node *)((char *)(p) + (p)->mp_ptrs[i]))
#define NODEKEY(node)	 (void *)((node)->mn_data)
#define NODEDATA(node)	 (void *)((char *)(node)->mn_data + (node)->mn_ksize)
#if LONG_MAX == 0x7fffffff
#define NODEPGNO(node)	 ((node)->mn_lo | ((node)->mn_hi << 16))
#define SETPGNO(node,pgno)	do { \
	(node)->mn_lo = (pgno) & 0xffff; (node)->mn_hi = (pgno) >> 16;} while(0)
#else
#define NODEPGNO(node)	 ((node)->mn_lo | ((node)->mn_hi << 16) | ((unsigned long)(node)->mn_flags << 32))
#define SETPGNO(node,pgno)	do { \
	(node)->mn_lo = (pgno) & 0xffff; (node)->mn_hi = (pgno) >> 16; \
	(node)->mn_flags = (pgno) >> 32; } while(0)
#endif
#define NODEDSZ(node)	 ((node)->mn_lo | ((unsigned)(node)->mn_hi << 16))
#define SETDSZ(node,size)	do { \
	(node)->mn_lo = (size) & 0xffff; (node)->mn_hi = (size) >> 16;} while(0)
#define NODEKSZ(node)	 ((node)->mn_ksize)
#define LEAF2KEY(p, i, ks)	((char *)(p) + PAGEHDRSZ + ((i)*(ks)))

#define MDB_SET_KEY(node, key)	if (key!=NULL) {(key)->mv_size = NODEKSZ(node); (key)->mv_data = NODEKEY(node);}

#define MDB_COMMIT_PAGES	 64	/* max number of pages to write in one commit */

static int  mdb_search_page_root(MDB_cursor *mc,
			    MDB_val *key, int modify);
static int  mdb_search_page(MDB_cursor *mc,
			    MDB_val *key, int modify);

static int  mdb_env_read_header(MDB_env *env, MDB_meta *meta);
static int  mdb_env_read_meta(MDB_env *env, int *which);
static int  mdb_env_write_meta(MDB_txn *txn);
static int  mdb_get_page(MDB_txn *txn, pgno_t pgno, MDB_page **mp);

static MDB_node *mdb_search_node(MDB_cursor *mc, MDB_val *key, int *exactp);
static int  mdb_add_node(MDB_cursor *mc, indx_t indx,
			    MDB_val *key, MDB_val *data, pgno_t pgno, uint8_t flags);
static void mdb_del_node(MDB_page *mp, indx_t indx, int ksize);
static int mdb_del0(MDB_cursor *mc, MDB_node *leaf);
static int  mdb_read_data(MDB_txn *txn, MDB_node *leaf, MDB_val *data);

static int	mdb_rebalance(MDB_cursor *mc);
static int	mdb_update_key(MDB_page *mp, indx_t indx, MDB_val *key);
static int	mdb_move_node(MDB_cursor *csrcrc, MDB_cursor *cdstst);
static int	mdb_merge(MDB_cursor *csrcrc, MDB_cursor *cdstst);
static int	mdb_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata,
				pgno_t newpgno);
static MDB_page *mdb_new_page(MDB_cursor *mc, uint32_t flags, int num);

static void	cursor_pop_page(MDB_cursor *mc);
static int	cursor_push_page(MDB_cursor *mc, MDB_page *mp);

static int	mdb_sibling(MDB_cursor *mc, int move_right);
static int	mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
static int	mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
static int	mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op,
				int *exactp);
static int	mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data);
static int	mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data);

static void	mdb_xcursor_init0(MDB_cursor *mc);
static void	mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node);
static void	mdb_xcursor_init2(MDB_cursor *mc);
static void	mdb_xcursor_fini(MDB_cursor *mc);

static size_t	mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data);
static size_t	mdb_branch_size(MDB_env *env, MDB_val *key);

static void mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi);

static MDB_cmp_func	memncmp, memnrcmp, intcmp, cintcmp;

#ifdef _WIN32
static SECURITY_DESCRIPTOR mdb_null_sd;
static SECURITY_ATTRIBUTES mdb_all_sa;
static int mdb_sec_inited;
#endif

char *
mdb_version(int *maj, int *min, int *pat)
{
	if (maj) *maj = MDB_VERSION_MAJOR;
	if (min) *min = MDB_VERSION_MINOR;
	if (pat) *pat = MDB_VERSION_PATCH;
	return MDB_VERSION_STRING;
}

static char *const mdb_errstr[] = {
	"MDB_KEYEXIST: Key/data pair already exists",
	"MDB_NOTFOUND: No matching key/data pair found",
	"MDB_PAGE_NOTFOUND: Requested page not found",
	"MDB_CORRUPTED: Located page was wrong type",
	"MDB_PANIC: Update of meta page failed",
	"MDB_VERSION_MISMATCH: Database environment version mismatch"
};

char *
mdb_strerror(int err)
{
	if (!err)
		return ("Successful return: 0");

	if (err >= MDB_KEYEXIST && err <= MDB_VERSION_MISMATCH)
		return mdb_errstr[err - MDB_KEYEXIST];

	return strerror(err);
}

#if DEBUG
static char *
mdb_dkey(MDB_val *key, char *buf)
{
	char *ptr = buf;
	unsigned char *c = key->mv_data;
	unsigned int i;
	if (key->mv_size > MAXKEYSIZE)
		return "MAXKEYSIZE";
#if 1
	for (i=0; i<key->mv_size; i++)
		ptr += sprintf(ptr, "%02x", *c++);
#else
	sprintf(buf, "%.*s", key->mv_size, key->mv_data);
#endif
	return buf;
}
#endif

int
mdb_cmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
{
	return txn->mt_dbxs[dbi].md_cmp(a, b);
}

int
mdb_dcmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
{
	if (txn->mt_dbxs[dbi].md_dcmp)
		return txn->mt_dbxs[dbi].md_dcmp(a, b);
	else
		return EINVAL;	/* too bad you can't distinguish this from a valid result */
}

/* Allocate new page(s) for writing */
static MDB_page *
mdb_alloc_page(MDB_cursor *mc, int num)
{
	MDB_txn *txn = mc->mc_txn;
	MDB_page *np;
	pgno_t pgno = P_INVALID;
	ID2 mid;

	if (txn->mt_txnid > 2) {

		if (!txn->mt_env->me_pghead && mc->mc_dbi != FREE_DBI &&
			txn->mt_dbs[FREE_DBI].md_root != P_INVALID) {
			/* See if there's anything in the free DB */
			MDB_cursor m2;
			MDB_node *leaf;
			ULONG *kptr, oldest;

			m2.mc_txn = txn;
			m2.mc_dbi = FREE_DBI;
			m2.mc_snum = 0;
			m2.mc_flags = 0;
			mdb_search_page(&m2, NULL, 0);
			leaf = NODEPTR(m2.mc_pg[m2.mc_top], 0);
			kptr = (ULONG *)NODEKEY(leaf);

			{
				unsigned int i;
				oldest = txn->mt_txnid - 1;
				for (i=0; i<txn->mt_env->me_txns->mti_numreaders; i++) {
					ULONG mr = txn->mt_env->me_txns->mti_readers[i].mr_txnid;
					if (mr && mr < oldest)
						oldest = mr;
				}
			}

			if (oldest > *kptr) {
				/* It's usable, grab it.
				 */
				MDB_oldpages *mop;
				MDB_val data;
				pgno_t *idl;

				mdb_read_data(txn, leaf, &data);
				idl = (ULONG *)data.mv_data;
				mop = malloc(sizeof(MDB_oldpages) + MDB_IDL_SIZEOF(idl) - sizeof(pgno_t));
				mop->mo_next = txn->mt_env->me_pghead;
				mop->mo_txnid = *kptr;
				txn->mt_env->me_pghead = mop;
				memcpy(mop->mo_pages, idl, MDB_IDL_SIZEOF(idl));

#if DEBUG > 1
				{
					unsigned int i;
					DPRINTF("IDL read txn %lu root %lu num %lu",
						mop->mo_txnid, txn->mt_dbs[FREE_DBI].md_root, idl[0]);
					for (i=0; i<idl[0]; i++) {
						DPRINTF("IDL %lu", idl[i+1]);
					}
				}
#endif
				/* drop this IDL from the DB */
				m2.mc_ki[m2.mc_top] = 0;
				m2.mc_flags = C_INITIALIZED;
				mdb_cursor_del(&m2, 0);
			}
		}
		if (txn->mt_env->me_pghead) {
			MDB_oldpages *mop = txn->mt_env->me_pghead;
			if (num > 1) {
				/* FIXME: For now, always use fresh pages. We
				 * really ought to search the free list for a
				 * contiguous range.
				 */
				;
			} else {
				/* peel pages off tail, so we only have to truncate the list */
				pgno = MDB_IDL_LAST(mop->mo_pages);
				if (MDB_IDL_IS_RANGE(mop->mo_pages)) {
					mop->mo_pages[2]++;
					if (mop->mo_pages[2] > mop->mo_pages[1])
						mop->mo_pages[0] = 0;
				} else {
					mop->mo_pages[0]--;
				}
				if (MDB_IDL_IS_ZERO(mop->mo_pages)) {
					txn->mt_env->me_pghead = mop->mo_next;
					free(mop);
				}
			}
		}
	}

	if (pgno == P_INVALID) {
		/* DB size is maxed out */
		if (txn->mt_next_pgno + num >= txn->mt_env->me_maxpg)
			return NULL;
	}
	if (txn->mt_env->me_dpages && num == 1) {
		np = txn->mt_env->me_dpages;
		txn->mt_env->me_dpages = np->mp_next;
	} else {
		if ((np = malloc(txn->mt_env->me_psize * num )) == NULL)
			return NULL;
	}
	if (pgno == P_INVALID) {
		np->mp_pgno = txn->mt_next_pgno;
		txn->mt_next_pgno += num;
	} else {
		np->mp_pgno = pgno;
	}
	mid.mid = np->mp_pgno;
	mid.mptr = np;
	mdb_mid2l_insert(txn->mt_u.dirty_list, &mid);

	return np;
}

/* Touch a page: make it dirty and re-insert into tree with updated pgno.
 */
static int
mdb_touch(MDB_cursor *mc)
{
	MDB_page *mp = mc->mc_pg[mc->mc_top];
	pgno_t	pgno;

	if (!F_ISSET(mp->mp_flags, P_DIRTY)) {
		MDB_page *np;
		if ((np = mdb_alloc_page(mc, 1)) == NULL)
			return ENOMEM;
		DPRINTF("touched db %u page %lu -> %lu", mc->mc_dbi, mp->mp_pgno, np->mp_pgno);
		assert(mp->mp_pgno != np->mp_pgno);
		mdb_midl_append(mc->mc_txn->mt_free_pgs, mp->mp_pgno);
		pgno = np->mp_pgno;
		memcpy(np, mp, mc->mc_txn->mt_env->me_psize);
		mp = np;
		mp->mp_pgno = pgno;
		mp->mp_flags |= P_DIRTY;

		mc->mc_pg[mc->mc_top] = mp;
		/* Update the page number to new touched page. */
		if (mc->mc_top)
			SETPGNO(NODEPTR(mc->mc_pg[mc->mc_top-1], mc->mc_ki[mc->mc_top-1]), mp->mp_pgno);
	}
	return 0;
}

int
mdb_env_sync(MDB_env *env, int force)
{
	int rc = 0;
	if (force || !F_ISSET(env->me_flags, MDB_NOSYNC)) {
		if (fdatasync(env->me_fd))
			rc = ErrCode();
	}
	return rc;
}

static inline void
mdb_txn_reset0(MDB_txn *txn);

static inline int
mdb_txn_renew0(MDB_txn *txn)
{
	MDB_env *env = txn->mt_env;

	if (txn->mt_flags & MDB_TXN_RDONLY) {
		MDB_reader *r = pthread_getspecific(env->me_txkey);
		if (!r) {
			unsigned int i;
			pid_t pid = getpid();
			pthread_t tid = pthread_self();

			LOCK_MUTEX_R(env);
			for (i=0; i<env->me_txns->mti_numreaders; i++)
				if (env->me_txns->mti_readers[i].mr_pid == 0)
					break;
			if (i == env->me_maxreaders) {
				UNLOCK_MUTEX_R(env);
				return ENOMEM;
			}
			env->me_txns->mti_readers[i].mr_pid = pid;
			env->me_txns->mti_readers[i].mr_tid = tid;
			if (i >= env->me_txns->mti_numreaders)
				env->me_txns->mti_numreaders = i+1;
			UNLOCK_MUTEX_R(env);
			r = &env->me_txns->mti_readers[i];
			pthread_setspecific(env->me_txkey, r);
		}
		txn->mt_txnid = env->me_txns->mti_txnid;
		txn->mt_toggle = env->me_txns->mti_me_toggle;
		r->mr_txnid = txn->mt_txnid;
		txn->mt_u.reader = r;
	} else {
		LOCK_MUTEX_W(env);

		txn->mt_txnid = env->me_txns->mti_txnid+1;
		txn->mt_toggle = env->me_txns->mti_me_toggle;
		txn->mt_u.dirty_list = env->me_dirty_list;
		txn->mt_u.dirty_list[0].mid = 0;
		txn->mt_free_pgs = env->me_free_pgs;
		txn->mt_free_pgs[0] = 0;
		txn->mt_next_pgno = env->me_metas[txn->mt_toggle]->mm_last_pg+1;
		env->me_txn = txn;
	}

	/* Copy the DB arrays */
	LAZY_RWLOCK_RDLOCK(&env->me_dblock);
	txn->mt_numdbs = env->me_numdbs;
	txn->mt_dbxs = env->me_dbxs;	/* mostly static anyway */
	memcpy(txn->mt_dbs, env->me_metas[txn->mt_toggle]->mm_dbs, 2 * sizeof(MDB_db));
	if (txn->mt_numdbs > 2)
		memcpy(txn->mt_dbs+2, env->me_dbs[env->me_db_toggle]+2,
			(txn->mt_numdbs - 2) * sizeof(MDB_db));
	LAZY_RWLOCK_UNLOCK(&env->me_dblock);

	return MDB_SUCCESS;
}

int
mdb_txn_renew(MDB_txn *txn)
{
	int rc;

	if (!txn)
		return EINVAL;

	if (txn->mt_env->me_flags & MDB_FATAL_ERROR) {
		DPUTS("environment had fatal error, must shutdown!");
		return MDB_PANIC;
	}

	rc = mdb_txn_renew0(txn);
	if (rc == MDB_SUCCESS) {
		DPRINTF("renew txn %lu%c %p on mdbenv %p, root page %lu",
			txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w', txn,
			(void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root);
	}
	return rc;
}

int
mdb_txn_begin(MDB_env *env, unsigned int flags, MDB_txn **ret)
{
	MDB_txn *txn;
	int rc;

	if (env->me_flags & MDB_FATAL_ERROR) {
		DPUTS("environment had fatal error, must shutdown!");
		return MDB_PANIC;
	}
	if ((txn = calloc(1, sizeof(MDB_txn) + env->me_maxdbs * sizeof(MDB_db))) == NULL) {
		DPRINTF("calloc: %s", strerror(ErrCode()));
		return ENOMEM;
	}
	txn->mt_dbs = (MDB_db *)(txn+1);
	if (flags & MDB_RDONLY) {
		txn->mt_flags |= MDB_TXN_RDONLY;
	}
	txn->mt_env = env;

	rc = mdb_txn_renew0(txn);
	if (rc)
		free(txn);
	else {
		*ret = txn;
		DPRINTF("begin txn %lu%c %p on mdbenv %p, root page %lu",
			txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w', txn,
			(void *) env, txn->mt_dbs[MAIN_DBI].md_root);
	}

	return rc;
}

static inline void
mdb_txn_reset0(MDB_txn *txn)
{
	MDB_env	*env = txn->mt_env;

	if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
		txn->mt_u.reader->mr_txnid = 0;
	} else {
		MDB_oldpages *mop;
		MDB_page *dp;
		unsigned int i;

		/* return all dirty pages to dpage list */
		for (i=1; i<=txn->mt_u.dirty_list[0].mid; i++) {
			dp = txn->mt_u.dirty_list[i].mptr;
			if (!IS_OVERFLOW(dp) || dp->mp_pages == 1) {
				dp->mp_next = txn->mt_env->me_dpages;
				txn->mt_env->me_dpages = dp;
			} else {
				/* large pages just get freed directly */
				free(dp);
			}
		}

		while ((mop = txn->mt_env->me_pghead)) {
			txn->mt_env->me_pghead = mop->mo_next;
			free(mop);
		}

		env->me_txn = NULL;
		for (i=2; i<env->me_numdbs; i++)
			env->me_dbxs[i].md_dirty = 0;
		UNLOCK_MUTEX_W(env);
	}
}

void
mdb_txn_reset(MDB_txn *txn)
{
	if (txn == NULL)
		return;

	DPRINTF("reset txn %lu%c %p on mdbenv %p, root page %lu",
		txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w', txn,
		(void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root);

	mdb_txn_reset0(txn);
}

void
mdb_txn_abort(MDB_txn *txn)
{
	if (txn == NULL)
		return;

	DPRINTF("abort txn %lu%c %p on mdbenv %p, root page %lu",
		txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w', txn,
		(void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root);

	mdb_txn_reset0(txn);
	free(txn);
}

int
mdb_txn_commit(MDB_txn *txn)
{
	int		 n, done;
	unsigned int i;
	ssize_t		 rc;
	off_t		 size;
	MDB_page	*dp;
	MDB_env	*env;
	pgno_t	next;
	MDB_cursor mc;

	assert(txn != NULL);
	assert(txn->mt_env != NULL);

	env = txn->mt_env;

	if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
		mdb_txn_abort(txn);
		return MDB_SUCCESS;
	}

	if (txn != env->me_txn) {
		DPUTS("attempt to commit unknown transaction");
		mdb_txn_abort(txn);
		return EINVAL;
	}

	if (F_ISSET(txn->mt_flags, MDB_TXN_ERROR)) {
		DPUTS("error flag is set, can't commit");
		mdb_txn_abort(txn);
		return EINVAL;
	}

	if (!txn->mt_u.dirty_list[0].mid)
		goto done;

	DPRINTF("committing txn %lu %p on mdbenv %p, root page %lu",
	    txn->mt_txnid, txn, (void *)env, txn->mt_dbs[MAIN_DBI].md_root);

	mc.mc_txn = txn;
	mc.mc_dbi = FREE_DBI;
	mc.mc_flags = 0;

	/* should only be one record now */
	if (env->me_pghead) {
		/* make sure first page of freeDB is touched and on freelist */
		mdb_search_page(&mc, NULL, 1);
	}
	/* save to free list */
	if (!MDB_IDL_IS_ZERO(txn->mt_free_pgs)) {
		MDB_val key, data;
		ULONG i;

		/* make sure last page of freeDB is touched and on freelist */
		key.mv_size = MAXKEYSIZE+1;
		key.mv_data = NULL;
		mdb_search_page(&mc, &key, 1);

		mdb_midl_sort(txn->mt_free_pgs);
#if DEBUG > 1
		{
			unsigned int i;
			ULONG *idl = txn->mt_free_pgs;
			DPRINTF("IDL write txn %lu root %lu num %lu",
				txn->mt_txnid, txn->mt_dbs[FREE_DBI].md_root, idl[0]);
			for (i=0; i<idl[0]; i++) {
				DPRINTF("IDL %lu", idl[i+1]);
			}
		}
#endif
		/* write to last page of freeDB */
		key.mv_size = sizeof(pgno_t);
		key.mv_data = (char *)&txn->mt_txnid;
		data.mv_data = txn->mt_free_pgs;
		/* The free list can still grow during this call,
		 * despite the pre-emptive touches above. So check
		 * and make sure the entire thing got written.
		 */
		do {
			i = txn->mt_free_pgs[0];
			data.mv_size = MDB_IDL_SIZEOF(txn->mt_free_pgs);
			rc = mdb_cursor_put(&mc, &key, &data, 0);
			if (rc) {
				mdb_txn_abort(txn);
				return rc;
			}
		} while (i != txn->mt_free_pgs[0]);
	}
	/* should only be one record now */
	if (env->me_pghead) {
		MDB_val key, data;
		MDB_oldpages *mop;

		mop = env->me_pghead;
		key.mv_size = sizeof(pgno_t);
		key.mv_data = (char *)&mop->mo_txnid;
		data.mv_size = MDB_IDL_SIZEOF(mop->mo_pages);
		data.mv_data = mop->mo_pages;
		mdb_cursor_put(&mc, &key, &data, 0);
		free(env->me_pghead);
		env->me_pghead = NULL;
	}

	/* Update DB root pointers. Their pages have already been
	 * touched so this is all in-place and cannot fail.
	 */
	{
		MDB_val data;
		data.mv_size = sizeof(MDB_db);

		mc.mc_dbi = MAIN_DBI;
		mc.mc_flags = 0;
		for (i = 2; i < txn->mt_numdbs; i++) {
			if (txn->mt_dbxs[i].md_dirty) {
				data.mv_data = &txn->mt_dbs[i];
				mdb_cursor_put(&mc, &txn->mt_dbxs[i].md_name, &data, 0);
			}
		}
	}

	/* Commit up to MDB_COMMIT_PAGES dirty pages to disk until done.
	 */
	next = 0;
	i = 1;
	do {
#ifdef _WIN32
		/* Windows actually supports scatter/gather I/O, but only on
		 * unbuffered file handles. Since we're relying on the OS page
		 * cache for all our data, that's self-defeating. So we just
		 * write pages one at a time. We use the ov structure to set
		 * the write offset, to at least save the overhead of a Seek
		 * system call.
		 */
		OVERLAPPED ov;
		memset(&ov, 0, sizeof(ov));
		for (; i<=txn->mt_u.dirty_list[0].mid; i++) {
			size_t wsize;
			dp = txn->mt_u.dirty_list[i].mptr;
			DPRINTF("committing page %lu", dp->mp_pgno);
			size = dp->mp_pgno * env->me_psize;
			ov.Offset = size & 0xffffffff;
			ov.OffsetHigh = size >> 16;
			ov.OffsetHigh >>= 16;
			/* clear dirty flag */
			dp->mp_flags &= ~P_DIRTY;
			wsize = env->me_psize;
			if (IS_OVERFLOW(dp)) wsize *= dp->mp_pages;
			rc = WriteFile(env->me_fd, dp, wsize, NULL, &ov);
			if (!rc) {
				n = ErrCode();
				DPRINTF("WriteFile: %d", n);
				mdb_txn_abort(txn);
				return n;
			}
		}
		done = 1;;
#else
		struct iovec	 iov[MDB_COMMIT_PAGES];
		n = 0;
		done = 1;
		size = 0;
		for (; i<=txn->mt_u.dirty_list[0].mid; i++) {
			dp = txn->mt_u.dirty_list[i].mptr;
			if (dp->mp_pgno != next) {
				if (n) {
					DPRINTF("committing %u dirty pages", n);
					rc = writev(env->me_fd, iov, n);
					if (rc != size) {
						n = ErrCode();
						if (rc > 0)
							DPUTS("short write, filesystem full?");
						else
							DPRINTF("writev: %s", strerror(n));
						mdb_txn_abort(txn);
						return n;
					}
					n = 0;
					size = 0;
				}
				lseek(env->me_fd, dp->mp_pgno * env->me_psize, SEEK_SET);
				next = dp->mp_pgno;
			}
			DPRINTF("committing page %lu", dp->mp_pgno);
			iov[n].iov_len = env->me_psize;
			if (IS_OVERFLOW(dp)) iov[n].iov_len *= dp->mp_pages;
			iov[n].iov_base = dp;
			size += iov[n].iov_len;
			next = dp->mp_pgno + (IS_OVERFLOW(dp) ? dp->mp_pages : 1);
			/* clear dirty flag */
			dp->mp_flags &= ~P_DIRTY;
			if (++n >= MDB_COMMIT_PAGES) {
				done = 0;
				i++;
				break;
			}
		}

		if (n == 0)
			break;

		DPRINTF("committing %u dirty pages", n);
		rc = writev(env->me_fd, iov, n);
		if (rc != size) {
			n = ErrCode();
			if (rc > 0)
				DPUTS("short write, filesystem full?");
			else
				DPRINTF("writev: %s", strerror(n));
			mdb_txn_abort(txn);
			return n;
		}
#endif
	} while (!done);

	/* Drop the dirty pages.
	 */
	for (i=1; i<=txn->mt_u.dirty_list[0].mid; i++) {
		dp = txn->mt_u.dirty_list[i].mptr;
		if (!IS_OVERFLOW(dp) || dp->mp_pages == 1) {
			dp->mp_next = txn->mt_env->me_dpages;
			txn->mt_env->me_dpages = dp;