X-Git-Url: https://git.sur5r.net/?a=blobdiff_plain;f=libraries%2Fliblmdb%2Fmdb.c;h=fa0c9e5b9ccd1f2ba6df9daba40bcf3258eee3aa;hb=ec32e90022f43af86db5330ed5a28a38644f1692;hp=410981e989ca1df6f7dd4e1ccb8e20e4c27a00f1;hpb=a36f7a7fc2128b3a6c9ac5b029c111904d28f49e;p=openldap

diff --git a/libraries/liblmdb/mdb.c b/libraries/liblmdb/mdb.c
index 410981e989..fa0c9e5b9c 100644
--- a/libraries/liblmdb/mdb.c
+++ b/libraries/liblmdb/mdb.c
@@ -96,7 +96,13 @@ extern int cacheflush(char *addr, int nbytes, int cache);
 #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 */
@@ -232,7 +238,34 @@ extern int cacheflush(char *addr, int nbytes, int cache);
 #define MDB_OWNERDEAD	EOWNERDEAD	/**< #LOCK_MUTEX0() result if dead owner */
 #endif
 
-#ifdef MDB_OWNERDEAD
+#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
+/* glibc < 2.12 only provided _np API */
+#  if defined(__GLIBC__) && GLIBC_VER < 0x02000c
+#   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
+#endif /* MDB_USE_ROBUST */
+
+#if defined(MDB_OWNERDEAD) && MDB_USE_ROBUST
 #define MDB_ROBUST_SUPPORTED	1
 #endif
 
@@ -370,10 +403,13 @@ static int mdb_mutex_failed(MDB_env *env, mdb_mutexref_t mutex, int rc);
  *
  *	@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
+# ifdef O_DSYNC
 # define MDB_DSYNC	O_DSYNC
+# else
+# define MDB_DSYNC	O_SYNC
+# endif
 #endif
 #endif
 
@@ -943,6 +979,11 @@ typedef struct MDB_db {
 #define	FREE_DBI	0
 	/** Handle for the default DB. */
 #define	MAIN_DBI	1
+	/** Number of DBs in metapage (free and main) - also hardcoded elsewhere */
+#define CORE_DBS	2
+
+	/** Number of meta pages - also hardcoded elsewhere */
+#define NUM_METAS	2
 
 	/** Meta page content.
 	 *	A meta page is the start point for accessing a database snapshot.
@@ -956,11 +997,11 @@ typedef struct MDB_meta {
 	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 */
+	MDB_db		mm_dbs[CORE_DBS];	/**< first is free space, 2nd is main db */
 	/** The size of pages used in this DB */
-#define	mm_psize	mm_dbs[0].md_pad
+#define	mm_psize	mm_dbs[FREE_DBI].md_pad
 	/** Any persistent environment flags. @ref mdb_env */
-#define	mm_flags	mm_dbs[0].md_flags
+#define	mm_flags	mm_dbs[FREE_DBI].md_flags
 	pgno_t		mm_last_pg;			/**< last used page in file */
 	volatile txnid_t	mm_txnid;	/**< txnid that committed this page */
 } MDB_meta;
@@ -995,7 +1036,8 @@ typedef struct MDB_dbx {
 	 */
 struct MDB_txn {
 	MDB_txn		*mt_parent;		/**< parent of a nested txn */
-	MDB_txn		*mt_child;		/**< nested txn under this txn */
+	/** Nested txn under this txn, set together with flag #MDB_TXN_HAS_CHILD */
+	MDB_txn		*mt_child;
 	pgno_t		mt_next_pgno;	/**< next unallocated page */
 	/** The ID of this transaction. IDs are integers incrementing from 1.
 	 *	Only committed write transactions increment the ID. If a transaction
@@ -1037,13 +1079,15 @@ struct MDB_txn {
 #define DB_STALE	0x02		/**< Named-DB record is older than txnID */
 #define DB_NEW		0x04		/**< Named-DB handle opened in this txn */
 #define DB_VALID	0x08		/**< DB handle is valid, see also #MDB_VALID */
+#define DB_USRVALID	0x10		/**< As #DB_VALID, but not set for #FREE_DBI */
 /** @} */
 	/** In write txns, array of cursors for each DB */
 	MDB_cursor	**mt_cursors;
 	/** Array of flags for each DB */
 	unsigned char	*mt_dbflags;
-	/**	Number of DB records in use. This number only ever increments;
-	 *	we don't decrement it when individual DB handles are closed.
+	/**	Number of DB records in use, or 0 when the txn is finished.
+	 *	This number only ever increments until the txn finishes; we
+	 *	don't decrement it when individual DB handles are closed.
 	 */
 	MDB_dbi		mt_numdbs;
 
@@ -1056,9 +1100,13 @@ struct MDB_txn {
 #define MDB_TXN_RDONLY		MDB_RDONLY	/**< read-only transaction */
 	/* internal txn flags */
 #define MDB_TXN_WRITEMAP	MDB_WRITEMAP	/**< copy of #MDB_env flag in writers */
+#define MDB_TXN_FINISHED	0x01		/**< txn is finished or never began */
 #define MDB_TXN_ERROR		0x02		/**< txn is unusable after an error */
 #define MDB_TXN_DIRTY		0x04		/**< must write, even if dirty list is empty */
 #define MDB_TXN_SPILLS		0x08		/**< txn or a parent has spilled pages */
+#define MDB_TXN_HAS_CHILD	0x10		/**< txn has an #MDB_txn.%mt_child */
+	/** most operations on the txn are currently illegal */
+#define MDB_TXN_BLOCKED		(MDB_TXN_FINISHED|MDB_TXN_ERROR|MDB_TXN_HAS_CHILD)
 /** @} */
 	unsigned int	mt_flags;		/**< @ref mdb_txn */
 	/** #dirty_list room: Array size - \#dirty pages visible to this txn.
@@ -1113,7 +1161,6 @@ struct MDB_cursor {
 #define C_EOF	0x02			/**< No more data */
 #define C_SUB	0x04			/**< Cursor is a sub-cursor */
 #define C_DEL	0x08			/**< last op was a cursor_del */
-#define C_SPLITTING	0x20		/**< Cursor is in page_split */
 #define C_UNTRACK	0x40		/**< Un-track cursor when closing */
 /** @} */
 	unsigned int	mc_flags;	/**< @ref mdb_cursor */
@@ -1168,7 +1215,7 @@ struct MDB_env {
 	char		*me_path;		/**< path to the DB files */
 	char		*me_map;		/**< the memory map of the data file */
 	MDB_txninfo	*me_txns;		/**< the memory map of the lock file or NULL */
-	MDB_meta	*me_metas[2];	/**< pointers to the two meta pages */
+	MDB_meta	*me_metas[NUM_METAS];	/**< pointers to the two meta pages */
 	void		*me_pbuf;		/**< scratch area for DUPSORT put() */
 	MDB_txn		*me_txn;		/**< current write transaction */
 	MDB_txn		*me_txn0;		/**< prealloc'd write transaction */
@@ -1227,8 +1274,8 @@ typedef struct MDB_ntxn {
 #define MAX_WRITE		(0x80000000U >> (sizeof(ssize_t) == 4))
 
 	/** Check \b txn and \b dbi arguments to a function */
-#define TXN_DBI_EXIST(txn, dbi) \
-	((txn) && (dbi) < (txn)->mt_numdbs && ((txn)->mt_dbflags[dbi] & DB_VALID))
+#define TXN_DBI_EXIST(txn, dbi, validity) \
+	((txn) && (dbi)<(txn)->mt_numdbs && ((txn)->mt_dbflags[dbi] & (validity)))
 
 	/** Check for misused \b dbi handles */
 #define TXN_DBI_CHANGED(txn, dbi) \
@@ -1238,6 +1285,19 @@ static int  mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp);
 static int  mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp);
 static int  mdb_page_touch(MDB_cursor *mc);
 
+#define MDB_END_NAMES {"committed", "empty-commit", "abort", "reset", \
+	"reset-tmp", "fail-begin", "fail-beginchild"}
+enum {
+	/* mdb_txn_end operation number, for logging */
+	MDB_END_COMMITTED, MDB_END_EMPTY_COMMIT, MDB_END_ABORT, MDB_END_RESET,
+	MDB_END_RESET_TMP, MDB_END_FAIL_BEGIN, MDB_END_FAIL_BEGINCHILD
+};
+#define MDB_END_OPMASK	0x0F	/**< mask for #mdb_txn_end() operation number */
+#define MDB_END_UPDATE	0x10	/**< update env state (DBIs) */
+#define MDB_END_FREE	0x20	/**< free txn unless it is #MDB_env.%me_txn0 */
+#define MDB_END_SLOT MDB_NOTLS	/**< release any reader slot if #MDB_NOTLS */
+static void mdb_txn_end(MDB_txn *txn, unsigned mode);
+
 static int  mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **mp, int *lvl);
 static int  mdb_page_search_root(MDB_cursor *mc,
 			    MDB_val *key, int modify);
@@ -1254,7 +1314,7 @@ static int	mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata,
 				pgno_t newpgno, unsigned int nflags);
 
 static int  mdb_env_read_header(MDB_env *env, MDB_meta *meta);
-static int  mdb_env_pick_meta(const MDB_env *env);
+static MDB_meta *mdb_env_pick_meta(const MDB_env *env);
 static int  mdb_env_write_meta(MDB_txn *txn);
 #ifdef MDB_USE_POSIX_MUTEX /* Drop unused excl arg */
 # define mdb_env_close0(env, excl) mdb_env_close1(env)
@@ -1266,7 +1326,7 @@ static int  mdb_node_add(MDB_cursor *mc, indx_t indx,
 			    MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags);
 static void mdb_node_del(MDB_cursor *mc, int ksize);
 static void mdb_node_shrink(MDB_page *mp, indx_t indx);
-static int	mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst);
+static int	mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst, int fromleft);
 static int  mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data);
 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);
@@ -1311,10 +1371,12 @@ static MDB_cmp_func	mdb_cmp_memn, mdb_cmp_memnr, mdb_cmp_int, mdb_cmp_cint, mdb_
 static SECURITY_DESCRIPTOR mdb_null_sd;
 static SECURITY_ATTRIBUTES mdb_all_sa;
 static int mdb_sec_inited;
+
+static int utf8_to_utf16(const char *src, int srcsize, wchar_t **dst, int *dstsize);
 #endif
 
 /** Return the library version info. */
-char *
+char * ESECT
 mdb_version(int *major, int *minor, int *patch)
 {
 	if (major) *major = MDB_VERSION_MAJOR;
@@ -1342,7 +1404,7 @@ static char *const mdb_errstr[] = {
 	"MDB_MAP_RESIZED: Database contents grew beyond environment mapsize",
 	"MDB_INCOMPATIBLE: Operation and DB incompatible, or DB flags changed",
 	"MDB_BAD_RSLOT: Invalid reuse of reader locktable slot",
-	"MDB_BAD_TXN: Transaction cannot recover - it must be aborted",
+	"MDB_BAD_TXN: Transaction must abort, has a child, or is invalid",
 	"MDB_BAD_VALSIZE: Unsupported size of key/DB name/data, or wrong DUPFIXED size",
 	"MDB_BAD_DBI: The specified DBI handle was closed/changed unexpectedly",
 };
@@ -1386,7 +1448,7 @@ mdb_strerror(int err)
 		;
 	}
 	buf[0] = 0;
-	FormatMessage(FORMAT_MESSAGE_FROM_SYSTEM |
+	FormatMessageA(FORMAT_MESSAGE_FROM_SYSTEM |
 		FORMAT_MESSAGE_IGNORE_INSERTS,
 		NULL, err, 0, ptr, sizeof(buf), (va_list *)pad);
 	return ptr;
@@ -1398,7 +1460,7 @@ mdb_strerror(int err)
 /** assert(3) variant in cursor context */
 #define mdb_cassert(mc, expr)	mdb_assert0((mc)->mc_txn->mt_env, expr, #expr)
 /** assert(3) variant in transaction context */
-#define mdb_tassert(mc, expr)	mdb_assert0((txn)->mt_env, expr, #expr)
+#define mdb_tassert(txn, expr)	mdb_assert0((txn)->mt_env, expr, #expr)
 /** assert(3) variant in environment context */
 #define mdb_eassert(env, expr)	mdb_assert0(env, expr, #expr)
 
@@ -1406,7 +1468,7 @@ mdb_strerror(int err)
 # define mdb_assert0(env, expr, expr_txt) ((expr) ? (void)0 : \
 		mdb_assert_fail(env, expr_txt, mdb_func_, __FILE__, __LINE__))
 
-static void
+static void ESECT
 mdb_assert_fail(MDB_env *env, const char *expr_txt,
 	const char *func, const char *file, int line)
 {
@@ -1542,7 +1604,7 @@ mdb_cursor_chk(MDB_cursor *mc)
 	MDB_node *node;
 	MDB_page *mp;
 
-	if (!mc->mc_snum && !(mc->mc_flags & C_INITIALIZED)) return;
+	if (!mc->mc_snum || !(mc->mc_flags & C_INITIALIZED)) return;
 	for (i=0; i<mc->mc_top; i++) {
 		mp = mc->mc_pg[i];
 		node = NODEPTR(mp, mc->mc_ki[i]);
@@ -1551,6 +1613,13 @@ mdb_cursor_chk(MDB_cursor *mc)
 	}
 	if (mc->mc_ki[i] >= NUMKEYS(mc->mc_pg[i]))
 		printf("ack!\n");
+	if (mc->mc_xcursor && (mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
+		node = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
+		if (((node->mn_flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA) &&
+			mc->mc_xcursor->mx_cursor.mc_pg[0] != NODEDATA(node)) {
+			printf("blah!\n");
+		}
+	}
 }
 #endif
 
@@ -1604,9 +1673,10 @@ static void mdb_audit(MDB_txn *txn)
 			mdb_tassert(txn, rc == MDB_NOTFOUND);
 		}
 	}
-	if (freecount + count + 2 /* metapages */ != txn->mt_next_pgno) {
+	if (freecount + count + NUM_METAS != txn->mt_next_pgno) {
 		fprintf(stderr, "audit: %lu freecount: %lu count: %lu total: %lu next_pgno: %lu\n",
-			txn->mt_txnid, freecount, count+2, freecount+count+2, txn->mt_next_pgno);
+			txn->mt_txnid, freecount, count+NUM_METAS,
+			freecount+count+NUM_METAS, txn->mt_next_pgno);
 	}
 }
 #endif
@@ -1875,7 +1945,7 @@ mdb_page_spill(MDB_cursor *m0, MDB_val *key, MDB_val *data)
 	/* Estimate how much space this op will take */
 	i = m0->mc_db->md_depth;
 	/* Named DBs also dirty the main DB */
-	if (m0->mc_dbi > MAIN_DBI)
+	if (m0->mc_dbi >= CORE_DBS)
 		i += txn->mt_dbs[MAIN_DBI].md_depth;
 	/* For puts, roughly factor in the key+data size */
 	if (key)
@@ -2355,14 +2425,15 @@ done:
 	} else {
 		for (; m2; m2=m2->mc_next) {
 			if (m2->mc_snum < mc->mc_snum) continue;
+			if (m2 == mc) continue;
 			if (m2->mc_pg[mc->mc_top] == mp) {
 				m2->mc_pg[mc->mc_top] = np;
 				if ((mc->mc_db->md_flags & MDB_DUPSORT) &&
 					IS_LEAF(np) &&
-					m2->mc_ki[mc->mc_top] == mc->mc_ki[mc->mc_top])
+					(m2->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED))
 				{
-					MDB_node *leaf = NODEPTR(np, mc->mc_ki[mc->mc_top]);
-					if (!(leaf->mn_flags & F_SUBDATA))
+					MDB_node *leaf = NODEPTR(np, m2->mc_ki[mc->mc_top]);
+					if ((leaf->mn_flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA)
 						m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
 				}
 			}
@@ -2426,14 +2497,15 @@ mdb_cursor_shadow(MDB_txn *src, MDB_txn *dst)
 				*bk = *mc;
 				mc->mc_backup = bk;
 				mc->mc_db = &dst->mt_dbs[i];
-				/* Kill pointers into src - and dst to reduce abuse: The
-				 * user may not use mc until dst ends. Otherwise we'd...
+				/* Kill pointers into src to reduce abuse: The
+				 * user may not use mc until dst ends. But we need a valid
+				 * txn pointer here for cursor fixups to keep working.
 				 */
-				mc->mc_txn    = NULL;	/* ...set this to dst */
-				mc->mc_dbflag = NULL;	/* ...and &dst->mt_dbflags[i] */
+				mc->mc_txn    = dst;
+				mc->mc_dbflag = &dst->mt_dbflags[i];
 				if ((mx = mc->mc_xcursor) != NULL) {
 					*(MDB_xcursor *)(bk+1) = *mx;
-					mx->mx_cursor.mc_txn = NULL; /* ...and dst. */
+					mx->mx_cursor.mc_txn = dst;
 				}
 				mc->mc_next = dst->mt_cursors[i];
 				dst->mt_cursors[i] = mc;
@@ -2483,12 +2555,6 @@ mdb_cursors_close(MDB_txn *txn, unsigned merge)
 	}
 }
 
-#if !(MDB_DEBUG)
-#define mdb_txn_reset0(txn, act) mdb_txn_reset0(txn)
-#endif
-static void
-mdb_txn_reset0(MDB_txn *txn, const char *act);
-
 #if !(MDB_PIDLOCK)		/* Currently the same as defined(_WIN32) */
 enum Pidlock_op {
 	Pidset, Pidcheck
@@ -2559,7 +2625,7 @@ mdb_txn_renew0(MDB_txn *txn)
 
 	if ((flags &= MDB_TXN_RDONLY) != 0) {
 		if (!ti) {
-			meta = env->me_metas[ mdb_env_pick_meta(env) ];
+			meta = mdb_env_pick_meta(env);
 			txn->mt_txnid = meta->mm_txnid;
 			txn->mt_u.reader = NULL;
 		} else {
@@ -2619,7 +2685,7 @@ mdb_txn_renew0(MDB_txn *txn)
 			txn->mt_u.reader = r;
 			meta = env->me_metas[txn->mt_txnid & 1];
 		}
-		txn->mt_dbxs = env->me_dbxs;	/* mostly static anyway */
+
 	} else {
 		/* Not yet touching txn == env->me_txn0, it may be active */
 		if (ti) {
@@ -2628,7 +2694,7 @@ mdb_txn_renew0(MDB_txn *txn)
 			txn->mt_txnid = ti->mti_txnid;
 			meta = env->me_metas[txn->mt_txnid & 1];
 		} else {
-			meta = env->me_metas[ mdb_env_pick_meta(env) ];
+			meta = mdb_env_pick_meta(env);
 			txn->mt_txnid = meta->mm_txnid;
 		}
 		txn->mt_txnid++;
@@ -2650,7 +2716,7 @@ mdb_txn_renew0(MDB_txn *txn)
 	}
 
 	/* Copy the DB info and flags */
-	memcpy(txn->mt_dbs, meta->mm_dbs, 2 * sizeof(MDB_db));
+	memcpy(txn->mt_dbs, meta->mm_dbs, CORE_DBS * sizeof(MDB_db));
 
 	/* Moved to here to avoid a data race in read TXNs */
 	txn->mt_next_pgno = meta->mm_last_pg+1;
@@ -2659,23 +2725,24 @@ mdb_txn_renew0(MDB_txn *txn)
 
 	/* Setup db info */
 	txn->mt_numdbs = env->me_numdbs;
-	for (i=2; i<txn->mt_numdbs; i++) {
+	for (i=CORE_DBS; i<txn->mt_numdbs; i++) {
 		x = env->me_dbflags[i];
 		txn->mt_dbs[i].md_flags = x & PERSISTENT_FLAGS;
-		txn->mt_dbflags[i] = (x & MDB_VALID) ? DB_VALID|DB_STALE : 0;
+		txn->mt_dbflags[i] = (x & MDB_VALID) ? DB_VALID|DB_USRVALID|DB_STALE : 0;
 	}
-	txn->mt_dbflags[0] = txn->mt_dbflags[1] = DB_VALID;
+	txn->mt_dbflags[MAIN_DBI] = DB_VALID|DB_USRVALID;
+	txn->mt_dbflags[FREE_DBI] = DB_VALID;
 
-	if (env->me_maxpg < txn->mt_next_pgno) {
-		mdb_txn_reset0(txn, "renew0-mapfail");
-		if (new_notls) {
-			txn->mt_u.reader->mr_pid = 0;
-			txn->mt_u.reader = NULL;
-		}
-		return MDB_MAP_RESIZED;
+	if (env->me_flags & MDB_FATAL_ERROR) {
+		DPUTS("environment had fatal error, must shutdown!");
+		rc = MDB_PANIC;
+	} else if (env->me_maxpg < txn->mt_next_pgno) {
+		rc = MDB_MAP_RESIZED;
+	} else {
+		return MDB_SUCCESS;
 	}
-
-	return MDB_SUCCESS;
+	mdb_txn_end(txn, new_notls /*0 or MDB_END_SLOT*/ | MDB_END_FAIL_BEGIN);
+	return rc;
 }
 
 int
@@ -2683,14 +2750,9 @@ mdb_txn_renew(MDB_txn *txn)
 {
 	int rc;
 
-	if (!txn || txn->mt_dbxs)	/* A reset txn has mt_dbxs==NULL */
+	if (!txn || !F_ISSET(txn->mt_flags, MDB_TXN_RDONLY|MDB_TXN_FINISHED))
 		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 %"Z"u%c %p on mdbenv %p, root page %"Z"u",
@@ -2710,57 +2772,42 @@ mdb_txn_begin(MDB_env *env, MDB_txn *parent, unsigned int flags, MDB_txn **ret)
 	flags &= MDB_TXN_BEGIN_FLAGS;
 	flags |= env->me_flags & MDB_WRITEMAP;
 
-	if (env->me_flags & MDB_FATAL_ERROR) {
-		DPUTS("environment had fatal error, must shutdown!");
-		return MDB_PANIC;
-	}
 	if (env->me_flags & MDB_RDONLY & ~flags) /* write txn in RDONLY env */
 		return EACCES;
 
-	size = tsize = sizeof(MDB_txn);
 	if (parent) {
 		/* Nested transactions: Max 1 child, write txns only, no writemap */
 		flags |= parent->mt_flags;
-		if (parent->mt_child ||
-			(flags & (MDB_RDONLY|MDB_WRITEMAP|MDB_TXN_ERROR)))
-		{
+		if (flags & (MDB_RDONLY|MDB_WRITEMAP|MDB_TXN_BLOCKED)) {
 			return (parent->mt_flags & MDB_TXN_RDONLY) ? EINVAL : MDB_BAD_TXN;
 		}
 		/* Child txns save MDB_pgstate and use own copy of cursors */
-		size = tsize = sizeof(MDB_ntxn);
-		size += env->me_maxdbs * sizeof(MDB_cursor *);
-	} else if (!(flags & MDB_RDONLY)) {
+		size = env->me_maxdbs * (sizeof(MDB_db)+sizeof(MDB_cursor *)+1);
+		size += tsize = sizeof(MDB_ntxn);
+	} else if (flags & MDB_RDONLY) {
+		size = env->me_maxdbs * (sizeof(MDB_db)+1);
+		size += tsize = sizeof(MDB_txn);
+	} else {
 		/* Reuse preallocated write txn. However, do not touch it until
 		 * mdb_txn_renew0() succeeds, since it currently may be active.
 		 */
 		txn = env->me_txn0;
 		goto renew;
 	}
-	size += env->me_maxdbs * (sizeof(MDB_db)+1);
-
 	if ((txn = calloc(1, size)) == NULL) {
 		DPRINTF(("calloc: %s", strerror(errno)));
 		return ENOMEM;
 	}
+	txn->mt_dbxs = env->me_dbxs;	/* static */
 	txn->mt_dbs = (MDB_db *) ((char *)txn + tsize);
-	if (flags & MDB_RDONLY) {
-		txn->mt_dbflags = (unsigned char *)(txn->mt_dbs + env->me_maxdbs);
-		txn->mt_dbiseqs = env->me_dbiseqs;
-	} else {
-		txn->mt_cursors = (MDB_cursor **)(txn->mt_dbs + env->me_maxdbs);
-		if (parent) {
-			txn->mt_dbiseqs = parent->mt_dbiseqs;
-			txn->mt_dbflags = (unsigned char *)(txn->mt_cursors + env->me_maxdbs);
-		} else {
-			txn->mt_dbiseqs = (unsigned int *)(txn->mt_cursors + env->me_maxdbs);
-			txn->mt_dbflags = (unsigned char *)(txn->mt_dbiseqs + env->me_maxdbs);
-		}
-	}
+	txn->mt_dbflags = (unsigned char *)txn + size - env->me_maxdbs;
 	txn->mt_flags = flags;
 	txn->mt_env = env;
 
 	if (parent) {
 		unsigned int i;
+		txn->mt_cursors = (MDB_cursor **)(txn->mt_dbs + env->me_maxdbs);
+		txn->mt_dbiseqs = parent->mt_dbiseqs;
 		txn->mt_u.dirty_list = malloc(sizeof(MDB_ID2)*MDB_IDL_UM_SIZE);
 		if (!txn->mt_u.dirty_list ||
 			!(txn->mt_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)))
@@ -2774,10 +2821,10 @@ mdb_txn_begin(MDB_env *env, MDB_txn *parent, unsigned int flags, MDB_txn **ret)
 		txn->mt_u.dirty_list[0].mid = 0;
 		txn->mt_spill_pgs = NULL;
 		txn->mt_next_pgno = parent->mt_next_pgno;
+		parent->mt_flags |= MDB_TXN_HAS_CHILD;
 		parent->mt_child = txn;
 		txn->mt_parent = parent;
 		txn->mt_numdbs = parent->mt_numdbs;
-		txn->mt_dbxs = parent->mt_dbxs;
 		memcpy(txn->mt_dbs, parent->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
 		/* Copy parent's mt_dbflags, but clear DB_NEW */
 		for (i=0; i<txn->mt_numdbs; i++)
@@ -2796,8 +2843,9 @@ mdb_txn_begin(MDB_env *env, MDB_txn *parent, unsigned int flags, MDB_txn **ret)
 		if (!rc)
 			rc = mdb_cursor_shadow(parent, txn);
 		if (rc)
-			mdb_txn_reset0(txn, "beginchild-fail");
-	} else {
+			mdb_txn_end(txn, MDB_END_FAIL_BEGINCHILD);
+	} else { /* MDB_RDONLY */
+		txn->mt_dbiseqs = env->me_dbiseqs;
 renew:
 		rc = mdb_txn_renew0(txn);
 	}
@@ -2805,7 +2853,7 @@ renew:
 		if (txn != env->me_txn0)
 			free(txn);
 	} else {
-		txn->mt_flags |= flags;	/* for txn==me_txn0, no effect otherwise */
+		txn->mt_flags |= flags;	/* could not change txn=me_txn0 earlier */
 		*ret = txn;
 		DPRINTF(("begin txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
 			txn->mt_txnid, (flags & MDB_RDONLY) ? 'r' : 'w',
@@ -2838,7 +2886,7 @@ mdb_dbis_update(MDB_txn *txn, int keep)
 	MDB_env *env = txn->mt_env;
 	unsigned char *tdbflags = txn->mt_dbflags;
 
-	for (i = n; --i >= 2;) {
+	for (i = n; --i >= CORE_DBS;) {
 		if (tdbflags[i] & DB_NEW) {
 			if (keep) {
 				env->me_dbflags[i] = txn->mt_dbs[i].md_flags | MDB_VALID;
@@ -2858,39 +2906,52 @@ mdb_dbis_update(MDB_txn *txn, int keep)
 		env->me_numdbs = n;
 }
 
-/** Common code for #mdb_txn_reset() and #mdb_txn_abort().
+/** End a transaction, except successful commit of a nested transaction.
  * May be called twice for readonly txns: First reset it, then abort.
- * @param[in] txn the transaction handle to reset
- * @param[in] act why the transaction is being reset
+ * @param[in] txn the transaction handle to end
+ * @param[in] mode why and how to end the transaction
  */
 static void
-mdb_txn_reset0(MDB_txn *txn, const char *act)
+mdb_txn_end(MDB_txn *txn, unsigned mode)
 {
 	MDB_env	*env = txn->mt_env;
+#if MDB_DEBUG
+	static const char *const names[] = MDB_END_NAMES;
+#endif
 
-	/* Close any DBI handles opened in this txn */
-	mdb_dbis_update(txn, 0);
+	/* Export or close DBI handles opened in this txn */
+	mdb_dbis_update(txn, mode & MDB_END_UPDATE);
 
 	DPRINTF(("%s txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
-		act, txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
+		names[mode & MDB_END_OPMASK],
+		txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
 		(void *) txn, (void *)env, txn->mt_dbs[MAIN_DBI].md_root));
 
 	if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
 		if (txn->mt_u.reader) {
 			txn->mt_u.reader->mr_txnid = (txnid_t)-1;
-			if (!(env->me_flags & MDB_NOTLS))
+			if (!(env->me_flags & MDB_NOTLS)) {
 				txn->mt_u.reader = NULL; /* txn does not own reader */
+			} else if (mode & MDB_END_SLOT) {
+				txn->mt_u.reader->mr_pid = 0;
+				txn->mt_u.reader = NULL;
+			} /* else txn owns the slot until it does MDB_END_SLOT */
 		}
-		txn->mt_numdbs = 0;		/* close nothing if called again */
-		txn->mt_dbxs = NULL;	/* mark txn as reset */
-	} else {
+		txn->mt_numdbs = 0;		/* prevent further DBI activity */
+		txn->mt_flags |= MDB_TXN_FINISHED;
+
+	} else if (!F_ISSET(txn->mt_flags, MDB_TXN_FINISHED)) {
 		pgno_t *pghead = env->me_pghead;
 
-		mdb_cursors_close(txn, 0);
+		if (!(mode & MDB_END_UPDATE)) /* !(already closed cursors) */
+			mdb_cursors_close(txn, 0);
 		if (!(env->me_flags & MDB_WRITEMAP)) {
 			mdb_dlist_free(txn);
 		}
 
+		txn->mt_numdbs = 0;
+		txn->mt_flags = MDB_TXN_FINISHED;
+
 		if (!txn->mt_parent) {
 			mdb_midl_shrink(&txn->mt_free_pgs);
 			env->me_free_pgs = txn->mt_free_pgs;
@@ -2899,11 +2960,14 @@ mdb_txn_reset0(MDB_txn *txn, const char *act)
 			env->me_pglast = 0;
 
 			env->me_txn = NULL;
+			mode = 0;	/* txn == env->me_txn0, do not free() it */
+
 			/* The writer mutex was locked in mdb_txn_begin. */
 			if (env->me_txns)
 				UNLOCK_MUTEX(env->me_wmutex);
 		} else {
 			txn->mt_parent->mt_child = NULL;
+			txn->mt_parent->mt_flags &= ~MDB_TXN_HAS_CHILD;
 			env->me_pgstate = ((MDB_ntxn *)txn)->mnt_pgstate;
 			mdb_midl_free(txn->mt_free_pgs);
 			mdb_midl_free(txn->mt_spill_pgs);
@@ -2912,6 +2976,9 @@ mdb_txn_reset0(MDB_txn *txn, const char *act)
 
 		mdb_midl_free(pghead);
 	}
+
+	if (mode & MDB_END_FREE)
+		free(txn);
 }
 
 void
@@ -2924,7 +2991,7 @@ mdb_txn_reset(MDB_txn *txn)
 	if (!(txn->mt_flags & MDB_TXN_RDONLY))
 		return;
 
-	mdb_txn_reset0(txn, "reset");
+	mdb_txn_end(txn, MDB_END_RESET);
 }
 
 void
@@ -2936,13 +3003,7 @@ mdb_txn_abort(MDB_txn *txn)
 	if (txn->mt_child)
 		mdb_txn_abort(txn->mt_child);
 
-	mdb_txn_reset0(txn, "abort");
-	/* Free reader slot tied to this txn (if MDB_NOTLS && writable FS) */
-	if ((txn->mt_flags & MDB_TXN_RDONLY) && txn->mt_u.reader)
-		txn->mt_u.reader->mr_pid = 0;
-
-	if (txn != txn->mt_env->me_txn0)
-		free(txn);
+	mdb_txn_end(txn, MDB_END_ABORT|MDB_END_SLOT|MDB_END_FREE);
 }
 
 /** Save the freelist as of this transaction to the freeDB.
@@ -3293,15 +3354,17 @@ int
 mdb_txn_commit(MDB_txn *txn)
 {
 	int		rc;
-	unsigned int i;
+	unsigned int i, end_mode;
 	MDB_env	*env;
 
-	if (txn == NULL || txn->mt_env == NULL)
+	if (txn == NULL)
 		return EINVAL;
 
+	/* mdb_txn_end() mode for a commit which writes nothing */
+	end_mode = MDB_END_EMPTY_COMMIT|MDB_END_UPDATE|MDB_END_SLOT|MDB_END_FREE;
+
 	if (txn->mt_child) {
 		rc = mdb_txn_commit(txn->mt_child);
-		txn->mt_child = NULL;
 		if (rc)
 			goto fail;
 	}
@@ -3309,14 +3372,11 @@ mdb_txn_commit(MDB_txn *txn)
 	env = txn->mt_env;
 
 	if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
-		mdb_dbis_update(txn, 1);
-		txn->mt_numdbs = 2; /* so txn_abort() doesn't close any new handles */
-		mdb_txn_abort(txn);
-		return MDB_SUCCESS;
+		goto done;
 	}
 
-	if (F_ISSET(txn->mt_flags, MDB_TXN_ERROR)) {
-		DPUTS("error flag is set, can't commit");
+	if (txn->mt_flags & (MDB_TXN_FINISHED|MDB_TXN_ERROR)) {
+		DPUTS("txn has failed/finished, can't commit");
 		if (txn->mt_parent)
 			txn->mt_parent->mt_flags |= MDB_TXN_ERROR;
 		rc = MDB_BAD_TXN;
@@ -3348,9 +3408,9 @@ mdb_txn_commit(MDB_txn *txn)
 		/* Update parent's DB table. */
 		memcpy(parent->mt_dbs, txn->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
 		parent->mt_numdbs = txn->mt_numdbs;
-		parent->mt_dbflags[0] = txn->mt_dbflags[0];
-		parent->mt_dbflags[1] = txn->mt_dbflags[1];
-		for (i=2; i<txn->mt_numdbs; i++) {
+		parent->mt_dbflags[FREE_DBI] = txn->mt_dbflags[FREE_DBI];
+		parent->mt_dbflags[MAIN_DBI] = txn->mt_dbflags[MAIN_DBI];
+		for (i=CORE_DBS; i<txn->mt_numdbs; i++) {
 			/* preserve parent's DB_NEW status */
 			x = parent->mt_dbflags[i] & DB_NEW;
 			parent->mt_dbflags[i] = txn->mt_dbflags[i] | x;
@@ -3379,6 +3439,25 @@ mdb_txn_commit(MDB_txn *txn)
 			pspill[0] = y;
 		}
 
+		/* Remove anything in our spill list from parent's dirty list */
+		if (txn->mt_spill_pgs && txn->mt_spill_pgs[0]) {
+			for (i=1; i<=txn->mt_spill_pgs[0]; i++) {
+				MDB_ID pn = txn->mt_spill_pgs[i];
+				if (pn & 1)
+					continue;	/* deleted spillpg */
+				pn >>= 1;
+				y = mdb_mid2l_search(dst, pn);
+				if (y <= dst[0].mid && dst[y].mid == pn) {
+					free(dst[y].mptr);
+					while (y < dst[0].mid) {
+						dst[y] = dst[y+1];
+						y++;
+					}
+					dst[0].mid--;
+				}
+			}
+		}
+
 		/* Find len = length of merging our dirty list with parent's */
 		x = dst[0].mid;
 		dst[0].mid = 0;		/* simplify loops */
@@ -3424,7 +3503,7 @@ mdb_txn_commit(MDB_txn *txn)
 		}
 
 		/* Append our loose page list to parent's */
-		for (lp = &parent->mt_loose_pgs; *lp; lp = &NEXT_LOOSE_PAGE(lp))
+		for (lp = &parent->mt_loose_pgs; *lp; lp = &NEXT_LOOSE_PAGE(*lp))
 			;
 		*lp = txn->mt_loose_pgs;
 		parent->mt_loose_count += txn->mt_loose_count;
@@ -3451,14 +3530,14 @@ mdb_txn_commit(MDB_txn *txn)
 	    txn->mt_txnid, (void*)txn, (void*)env, txn->mt_dbs[MAIN_DBI].md_root));
 
 	/* Update DB root pointers */
-	if (txn->mt_numdbs > 2) {
+	if (txn->mt_numdbs > CORE_DBS) {
 		MDB_cursor mc;
 		MDB_dbi i;
 		MDB_val data;
 		data.mv_size = sizeof(MDB_db);
 
 		mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
-		for (i = 2; i < txn->mt_numdbs; i++) {
+		for (i = CORE_DBS; i < txn->mt_numdbs; i++) {
 			if (txn->mt_dbflags[i] & DB_DIRTY) {
 				if (TXN_DBI_CHANGED(txn, i)) {
 					rc = MDB_BAD_DBI;
@@ -3480,7 +3559,6 @@ mdb_txn_commit(MDB_txn *txn)
 	mdb_midl_free(env->me_pghead);
 	env->me_pghead = NULL;
 	mdb_midl_shrink(&txn->mt_free_pgs);
-	env->me_free_pgs = txn->mt_free_pgs;
 
 #if (MDB_DEBUG) > 2
 	mdb_audit(txn);
@@ -3490,21 +3568,10 @@ mdb_txn_commit(MDB_txn *txn)
 		(rc = mdb_env_sync(env, 0)) ||
 		(rc = mdb_env_write_meta(txn)))
 		goto fail;
-
-	/* Free P_LOOSE pages left behind in dirty_list */
-	if (!(env->me_flags & MDB_WRITEMAP))
-		mdb_dlist_free(txn);
+	end_mode = MDB_END_COMMITTED|MDB_END_UPDATE;
 
 done:
-	env->me_pglast = 0;
-	env->me_txn = NULL;
-	mdb_dbis_update(txn, 1);
-
-	if (env->me_txns)
-		UNLOCK_MUTEX(env->me_wmutex);
-	if (txn != env->me_txn0)
-		free(txn);
-
+	mdb_txn_end(txn, end_mode);
 	return MDB_SUCCESS;
 
 fail:
@@ -3531,7 +3598,7 @@ mdb_env_read_header(MDB_env *env, MDB_meta *meta)
 	 * Read both meta pages so we can use the latest one.
 	 */
 
-	for (i=off=0; i<2; i++, off = meta->mm_psize) {
+	for (i=off=0; i<NUM_METAS; i++, off += meta->mm_psize) {
 #ifdef _WIN32
 		DWORD len;
 		OVERLAPPED ov;
@@ -3584,11 +3651,11 @@ mdb_env_init_meta0(MDB_env *env, MDB_meta *meta)
 	meta->mm_version = MDB_DATA_VERSION;
 	meta->mm_mapsize = env->me_mapsize;
 	meta->mm_psize = env->me_psize;
-	meta->mm_last_pg = 1;
+	meta->mm_last_pg = NUM_METAS-1;
 	meta->mm_flags = env->me_flags & 0xffff;
-	meta->mm_flags |= MDB_INTEGERKEY;
-	meta->mm_dbs[0].md_root = P_INVALID;
-	meta->mm_dbs[1].md_root = P_INVALID;
+	meta->mm_flags |= MDB_INTEGERKEY; /* this is mm_dbs[FREE_DBI].md_flags */
+	meta->mm_dbs[FREE_DBI].md_root = P_INVALID;
+	meta->mm_dbs[MAIN_DBI].md_root = P_INVALID;
 }
 
 /** Write the environment parameters of a freshly created DB environment.
@@ -3621,7 +3688,7 @@ mdb_env_init_meta(MDB_env *env, MDB_meta *meta)
 
 	psize = env->me_psize;
 
-	p = calloc(2, psize);
+	p = calloc(NUM_METAS, psize);
 	if (!p)
 		return ENOMEM;
 
@@ -3634,10 +3701,10 @@ mdb_env_init_meta(MDB_env *env, MDB_meta *meta)
 	q->mp_flags = P_META;
 	*(MDB_meta *)METADATA(q) = *meta;
 
-	DO_PWRITE(rc, env->me_fd, p, psize * 2, len, 0);
+	DO_PWRITE(rc, env->me_fd, p, psize * NUM_METAS, len, 0);
 	if (!rc)
 		rc = ErrCode();
-	else if ((unsigned) len == psize * 2)
+	else if ((unsigned) len == psize * NUM_METAS)
 		rc = MDB_SUCCESS;
 	else
 		rc = ENOSPC;
@@ -3680,8 +3747,8 @@ mdb_env_write_meta(MDB_txn *txn)
 
 	if (flags & MDB_WRITEMAP) {
 		mp->mm_mapsize = mapsize;
-		mp->mm_dbs[0] = txn->mt_dbs[0];
-		mp->mm_dbs[1] = txn->mt_dbs[1];
+		mp->mm_dbs[FREE_DBI] = txn->mt_dbs[FREE_DBI];
+		mp->mm_dbs[MAIN_DBI] = txn->mt_dbs[MAIN_DBI];
 		mp->mm_last_pg = txn->mt_next_pgno - 1;
 #if (__GNUC__ * 100 + __GNUC_MINOR__ >= 404) && /* TODO: portability */	\
 	!(defined(__i386__) || defined(__x86_64__))
@@ -3692,15 +3759,12 @@ mdb_env_write_meta(MDB_txn *txn)
 		if (!(flags & (MDB_NOMETASYNC|MDB_NOSYNC))) {
 			unsigned meta_size = env->me_psize;
 			rc = (env->me_flags & MDB_MAPASYNC) ? MS_ASYNC : MS_SYNC;
-			ptr = env->me_map;
-			if (toggle) {
+			ptr = (char *)mp - PAGEHDRSZ;
 #ifndef _WIN32	/* POSIX msync() requires ptr = start of OS page */
-				if (meta_size < env->me_os_psize)
-					meta_size += meta_size;
-				else
+			r2 = (ptr - env->me_map) & (env->me_os_psize - 1);
+			ptr -= r2;
+			meta_size += r2;
 #endif
-					ptr += meta_size;
-			}
 			if (MDB_MSYNC(ptr, meta_size, rc)) {
 				rc = ErrCode();
 				goto fail;
@@ -3708,21 +3772,19 @@ mdb_env_write_meta(MDB_txn *txn)
 		}
 		goto done;
 	}
-	metab.mm_txnid = env->me_metas[toggle]->mm_txnid;
-	metab.mm_last_pg = env->me_metas[toggle]->mm_last_pg;
+	metab.mm_txnid = mp->mm_txnid;
+	metab.mm_last_pg = mp->mm_last_pg;
 
 	meta.mm_mapsize = mapsize;
-	meta.mm_dbs[0] = txn->mt_dbs[0];
-	meta.mm_dbs[1] = txn->mt_dbs[1];
+	meta.mm_dbs[FREE_DBI] = txn->mt_dbs[FREE_DBI];
+	meta.mm_dbs[MAIN_DBI] = txn->mt_dbs[MAIN_DBI];
 	meta.mm_last_pg = txn->mt_next_pgno - 1;
 	meta.mm_txnid = txn->mt_txnid;
 
 	off = offsetof(MDB_meta, mm_mapsize);
 	ptr = (char *)&meta + off;
 	len = sizeof(MDB_meta) - off;
-	if (toggle)
-		off += env->me_psize;
-	off += PAGEHDRSZ;
+	off += (char *)mp - env->me_map;
 
 	/* Write to the SYNC fd */
 	mfd = (flags & (MDB_NOSYNC|MDB_NOMETASYNC)) ? env->me_fd : env->me_mfd;
@@ -3779,12 +3841,13 @@ done:
 
 /** Check both meta pages to see which one is newer.
  * @param[in] env the environment handle
- * @return meta toggle (0 or 1).
+ * @return newest #MDB_meta.
  */
-static int
+static MDB_meta *
 mdb_env_pick_meta(const MDB_env *env)
 {
-	return (env->me_metas[0]->mm_txnid < env->me_metas[1]->mm_txnid);
+	MDB_meta *const *metas = env->me_metas;
+	return metas[ metas[0]->mm_txnid < metas[1]->mm_txnid ];
 }
 
 int ESECT
@@ -3797,7 +3860,7 @@ mdb_env_create(MDB_env **env)
 		return ENOMEM;
 
 	e->me_maxreaders = DEFAULT_READERS;
-	e->me_maxdbs = e->me_numdbs = 2;
+	e->me_maxdbs = e->me_numdbs = CORE_DBS;
 	e->me_fd = INVALID_HANDLE_VALUE;
 	e->me_lfd = INVALID_HANDLE_VALUE;
 	e->me_mfd = INVALID_HANDLE_VALUE;
@@ -3908,7 +3971,7 @@ mdb_env_set_mapsize(MDB_env *env, size_t size)
 		void *old;
 		if (env->me_txn)
 			return EINVAL;
-		meta = env->me_metas[mdb_env_pick_meta(env)];
+		meta = mdb_env_pick_meta(env);
 		if (!size)
 			size = meta->mm_mapsize;
 		{
@@ -3935,7 +3998,7 @@ mdb_env_set_maxdbs(MDB_env *env, MDB_dbi dbs)
 {
 	if (env->me_map)
 		return EINVAL;
-	env->me_maxdbs = dbs + 2; /* Named databases + main and free DB */
+	env->me_maxdbs = dbs + CORE_DBS;
 	return MDB_SUCCESS;
 }
 
@@ -4115,12 +4178,12 @@ mdb_env_open2(MDB_env *env)
 
 #if MDB_DEBUG
 	{
-		int toggle = mdb_env_pick_meta(env);
-		MDB_db *db = &env->me_metas[toggle]->mm_dbs[MAIN_DBI];
+		MDB_meta *meta = mdb_env_pick_meta(env);
+		MDB_db *db = &meta->mm_dbs[MAIN_DBI];
 
 		DPRINTF(("opened database version %u, pagesize %u",
-			env->me_metas[0]->mm_version, env->me_psize));
-		DPRINTF(("using meta page %d",    toggle));
+			meta->mm_version, env->me_psize));
+		DPRINTF(("using meta page %d",    (int) (meta->mm_txnid & 1)));
 		DPRINTF(("depth: %u",             db->md_depth));
 		DPRINTF(("entries: %"Z"u",        db->md_entries));
 		DPRINTF(("branch pages: %"Z"u",   db->md_branch_pages));
@@ -4207,9 +4270,10 @@ PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
 static int ESECT
 mdb_env_share_locks(MDB_env *env, int *excl)
 {
-	int rc = 0, toggle = mdb_env_pick_meta(env);
+	int rc = 0;
+	MDB_meta *meta = mdb_env_pick_meta(env);
 
-	env->me_txns->mti_txnid = env->me_metas[toggle]->mm_txnid;
+	env->me_txns->mti_txnid = meta->mm_txnid;
 
 #ifdef _WIN32
 	{
@@ -4358,7 +4422,7 @@ mdb_hash_val(MDB_val *val, mdb_hash_t hval)
  */
 static const char mdb_a85[]= "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz!#$%&()*+-;<=>?@^_`{|}~";
 
-static void
+static void ESECT
 mdb_pack85(unsigned long l, char *out)
 {
 	int i;
@@ -4369,7 +4433,7 @@ mdb_pack85(unsigned long l, char *out)
 	}
 }
 
-static void
+static void ESECT
 mdb_hash_enc(MDB_val *val, char *encbuf)
 {
 	mdb_hash_t h = mdb_hash_val(val, MDB_HASH_INIT);
@@ -4405,9 +4469,14 @@ mdb_env_setup_locks(MDB_env *env, char *lpath, int mode, int *excl)
 	off_t size, rsize;
 
 #ifdef _WIN32
-	env->me_lfd = CreateFile(lpath, GENERIC_READ|GENERIC_WRITE,
+	wchar_t *wlpath;
+	rc = utf8_to_utf16(lpath, -1, &wlpath, NULL);
+	if (rc)
+		return rc;
+	env->me_lfd = CreateFileW(wlpath, GENERIC_READ|GENERIC_WRITE,
 		FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, OPEN_ALWAYS,
 		FILE_ATTRIBUTE_NORMAL, NULL);
+	free(wlpath);
 #else
 	env->me_lfd = open(lpath, O_RDWR|O_CREAT|MDB_CLOEXEC, mode);
 #endif
@@ -4509,9 +4578,9 @@ mdb_env_setup_locks(MDB_env *env, char *lpath, int mode, int *excl)
 		mdb_hash_enc(&val, encbuf);
 		sprintf(env->me_txns->mti_rmname, "Global\\MDBr%s", encbuf);
 		sprintf(env->me_txns->mti_wmname, "Global\\MDBw%s", encbuf);
-		env->me_rmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_rmname);
+		env->me_rmutex = CreateMutexA(&mdb_all_sa, FALSE, env->me_txns->mti_rmname);
 		if (!env->me_rmutex) goto fail_errno;
-		env->me_wmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_wmname);
+		env->me_wmutex = CreateMutexA(&mdb_all_sa, FALSE, env->me_txns->mti_wmname);
 		if (!env->me_wmutex) goto fail_errno;
 #elif defined(MDB_USE_POSIX_SEM)
 		struct stat stbuf;
@@ -4583,9 +4652,9 @@ mdb_env_setup_locks(MDB_env *env, char *lpath, int mode, int *excl)
 			goto fail;
 		}
 #ifdef _WIN32
-		env->me_rmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_rmname);
+		env->me_rmutex = OpenMutexA(SYNCHRONIZE, FALSE, env->me_txns->mti_rmname);
 		if (!env->me_rmutex) goto fail_errno;
-		env->me_wmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_wmname);
+		env->me_wmutex = OpenMutexA(SYNCHRONIZE, FALSE, env->me_txns->mti_wmname);
 		if (!env->me_wmutex) goto fail_errno;
 #elif defined(MDB_USE_POSIX_SEM)
 		env->me_rmutex = sem_open(env->me_txns->mti_rmname, 0);
@@ -4625,6 +4694,9 @@ mdb_env_open(MDB_env *env, const char *path, unsigned int flags, mdb_mode_t mode
 {
 	int		oflags, rc, len, excl = -1;
 	char *lpath, *dpath;
+#ifdef _WIN32
+	wchar_t *wpath;
+#endif
 
 	if (env->me_fd!=INVALID_HANDLE_VALUE || (flags & ~(CHANGEABLE|CHANGELESS)))
 		return EINVAL;
@@ -4670,6 +4742,7 @@ mdb_env_open(MDB_env *env, const char *path, unsigned int flags, mdb_mode_t mode
 		rc = ENOMEM;
 		goto leave;
 	}
+	env->me_dbxs[FREE_DBI].md_cmp = mdb_cmp_long; /* aligned MDB_INTEGERKEY */
 
 	/* For RDONLY, get lockfile after we know datafile exists */
 	if (!(flags & (MDB_RDONLY|MDB_NOLOCK))) {
@@ -4687,8 +4760,12 @@ mdb_env_open(MDB_env *env, const char *path, unsigned int flags, mdb_mode_t mode
 		len = OPEN_ALWAYS;
 	}
 	mode = FILE_ATTRIBUTE_NORMAL;
-	env->me_fd = CreateFile(dpath, oflags, FILE_SHARE_READ|FILE_SHARE_WRITE,
+	rc = utf8_to_utf16(dpath, -1, &wpath, NULL);
+	if (rc)
+		goto leave;
+	env->me_fd = CreateFileW(wpath, oflags, FILE_SHARE_READ|FILE_SHARE_WRITE,
 		NULL, len, mode, NULL);
+	free(wpath);
 #else
 	if (F_ISSET(flags, MDB_RDONLY))
 		oflags = O_RDONLY;
@@ -4717,9 +4794,13 @@ mdb_env_open(MDB_env *env, const char *path, unsigned int flags, mdb_mode_t mode
 			 */
 #ifdef _WIN32
 			len = OPEN_EXISTING;
-			env->me_mfd = CreateFile(dpath, oflags,
+			rc = utf8_to_utf16(dpath, -1, &wpath, NULL);
+			if (rc)
+				goto leave;
+			env->me_mfd = CreateFileW(wpath, oflags,
 				FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, len,
 				mode | FILE_FLAG_WRITE_THROUGH, NULL);
+			free(wpath);
 #else
 			oflags &= ~O_CREAT;
 			env->me_mfd = open(dpath, oflags | MDB_DSYNC, mode);
@@ -4748,6 +4829,7 @@ mdb_env_open(MDB_env *env, const char *path, unsigned int flags, mdb_mode_t mode
 				txn->mt_dbflags = (unsigned char *)(txn->mt_dbiseqs + env->me_maxdbs);
 				txn->mt_env = env;
 				txn->mt_dbxs = env->me_dbxs;
+				txn->mt_flags = MDB_TXN_FINISHED;
 				env->me_txn0 = txn;
 			} else {
 				rc = ENOMEM;
@@ -4774,7 +4856,7 @@ mdb_env_close0(MDB_env *env, int excl)
 
 	/* Doing this here since me_dbxs may not exist during mdb_env_close */
 	if (env->me_dbxs) {
-		for (i = env->me_maxdbs; --i > MAIN_DBI; )
+		for (i = env->me_maxdbs; --i >= CORE_DBS; )
 			free(env->me_dbxs[i].md_name.mv_data);
 		free(env->me_dbxs);
 	}
@@ -5091,15 +5173,15 @@ static void
 mdb_cursor_pop(MDB_cursor *mc)
 {
 	if (mc->mc_snum) {
-#if MDB_DEBUG
-		MDB_page	*top = mc->mc_pg[mc->mc_top];
-#endif
+		DPRINTF(("popping page %"Z"u off db %d cursor %p",
+			mc->mc_pg[mc->mc_top]->mp_pgno, DDBI(mc), (void *) mc));
+
 		mc->mc_snum--;
-		if (mc->mc_snum)
+		if (mc->mc_snum) {
 			mc->mc_top--;
-
-		DPRINTF(("popped page %"Z"u off db %d cursor %p", top->mp_pgno,
-			DDBI(mc), (void *) mc));
+		} else {
+			mc->mc_flags &= ~C_INITIALIZED;
+		}
 	}
 }
 
@@ -5293,8 +5375,8 @@ mdb_page_search(MDB_cursor *mc, MDB_val *key, int flags)
 	/* Make sure the txn is still viable, then find the root from
 	 * the txn's db table and set it as the root of the cursor's stack.
 	 */
-	if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_ERROR)) {
-		DPUTS("transaction has failed, must abort");
+	if (mc->mc_txn->mt_flags & MDB_TXN_BLOCKED) {
+		DPUTS("transaction may not be used now");
 		return MDB_BAD_TXN;
 	} else {
 		/* Make sure we're using an up-to-date root */
@@ -5415,6 +5497,7 @@ mdb_ovpage_free(MDB_cursor *mc, MDB_page *mp)
 				return MDB_CORRUPTED;
 			}
 		}
+		txn->mt_dirty_room++;
 		if (!(env->me_flags & MDB_WRITEMAP))
 			mdb_dpage_free(env, mp);
 release:
@@ -5478,10 +5561,10 @@ mdb_get(MDB_txn *txn, MDB_dbi dbi,
 
 	DPRINTF(("===> get db %u key [%s]", dbi, DKEY(key)));
 
-	if (!key || !data || dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
+	if (!key || !data || !TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
 		return EINVAL;
 
-	if (txn->mt_flags & MDB_TXN_ERROR)
+	if (txn->mt_flags & MDB_TXN_BLOCKED)
 		return MDB_BAD_TXN;
 
 	mdb_cursor_init(&mc, txn, dbi, &mx);
@@ -5809,6 +5892,8 @@ mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data,
 			} else
 				return MDB_NOTFOUND;
 		}
+	} else {
+		mc->mc_pg[0] = 0;
 	}
 
 	rc = mdb_page_search(mc, key, 0);
@@ -6002,7 +6087,7 @@ mdb_cursor_get(MDB_cursor *mc, MDB_val *key, MDB_val *data,
 	if (mc == NULL)
 		return EINVAL;
 
-	if (mc->mc_txn->mt_flags & MDB_TXN_ERROR)
+	if (mc->mc_txn->mt_flags & MDB_TXN_BLOCKED)
 		return MDB_BAD_TXN;
 
 	switch (op) {
@@ -6026,8 +6111,6 @@ mdb_cursor_get(MDB_cursor *mc, MDB_val *key, MDB_val *data,
 				MDB_GET_KEY(leaf, key);
 				if (data) {
 					if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
-						if (mc->mc_flags & C_DEL)
-							mdb_xcursor_init1(mc, leaf);
 						rc = mdb_cursor_get(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_GET_CURRENT);
 					} else {
 						rc = mdb_node_read(mc->mc_txn, leaf, data);
@@ -6173,7 +6256,7 @@ mdb_cursor_touch(MDB_cursor *mc)
 {
 	int rc = MDB_SUCCESS;
 
-	if (mc->mc_dbi > MAIN_DBI && !(*mc->mc_dbflag & DB_DIRTY)) {
+	if (mc->mc_dbi >= CORE_DBS && !(*mc->mc_dbflag & DB_DIRTY)) {
 		MDB_cursor mc2;
 		MDB_xcursor mcx;
 		if (TXN_DBI_CHANGED(mc->mc_txn, mc->mc_dbi))
@@ -6232,7 +6315,7 @@ mdb_cursor_put(MDB_cursor *mc, MDB_val *key, MDB_val *data,
 	nospill = flags & MDB_NOSPILL;
 	flags &= ~MDB_NOSPILL;
 
-	if (mc->mc_txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR))
+	if (mc->mc_txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_BLOCKED))
 		return (mc->mc_txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
 
 	if (key->mv_size-1 >= ENV_MAXKEY(env))
@@ -6357,16 +6440,18 @@ fix_parent:
 			 * update branch key if there is a parent page
 			 */
 			if (mc->mc_top && !mc->mc_ki[mc->mc_top]) {
-				unsigned short top = mc->mc_top;
+				unsigned short dtop = 1;
 				mc->mc_top--;
 				/* slot 0 is always an empty key, find real slot */
-				while (mc->mc_top && !mc->mc_ki[mc->mc_top])
+				while (mc->mc_top && !mc->mc_ki[mc->mc_top]) {
 					mc->mc_top--;
+					dtop++;
+				}
 				if (mc->mc_ki[mc->mc_top])
 					rc2 = mdb_update_key(mc, key);
 				else
 					rc2 = MDB_SUCCESS;
-				mc->mc_top = top;
+				mc->mc_top += dtop;
 				if (rc2)
 					return rc2;
 			}
@@ -6547,6 +6632,7 @@ current:
 						return ENOMEM;
 					id2.mid = pg;
 					id2.mptr = np;
+					/* Note - this page is already counted in parent's dirty_room */
 					rc2 = mdb_mid2l_insert(mc->mc_txn->mt_u.dirty_list, &id2);
 					mdb_cassert(mc, rc2 == 0);
 					if (!(flags & MDB_RESERVE)) {
@@ -6603,7 +6689,7 @@ new_sub:
 	} else {
 		/* There is room already in this leaf page. */
 		rc = mdb_node_add(mc, mc->mc_ki[mc->mc_top], key, rdata, 0, nflags);
-		if (rc == 0 && insert_key) {
+		if (rc == 0) {
 			/* Adjust other cursors pointing to mp */
 			MDB_cursor *m2, *m3;
 			MDB_dbi dbi = mc->mc_dbi;
@@ -6615,10 +6701,15 @@ new_sub:
 					m3 = &m2->mc_xcursor->mx_cursor;
 				else
 					m3 = m2;
-				if (m3 == mc || m3->mc_snum < mc->mc_snum) continue;
-				if (m3->mc_pg[i] == mp && m3->mc_ki[i] >= mc->mc_ki[i]) {
+				if (m3 == mc || m3->mc_snum < mc->mc_snum || m3->mc_pg[i] != mp) continue;
+				if (m3->mc_ki[i] >= mc->mc_ki[i] && insert_key) {
 					m3->mc_ki[i]++;
 				}
+				if (m3->mc_xcursor && (m3->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
+					MDB_node *n2 = NODEPTR(mp, m3->mc_ki[i]);
+					if ((n2->mn_flags & (F_SUBDATA|F_DUPDATA)) == F_DUPDATA)
+						m3->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(n2);
+				}
 			}
 		}
 	}
@@ -6660,6 +6751,7 @@ put_sub:
 				MDB_xcursor *mx = mc->mc_xcursor;
 				unsigned i = mc->mc_top;
 				MDB_page *mp = mc->mc_pg[i];
+				int nkeys = NUMKEYS(mp);
 
 				for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
 					if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
@@ -6667,9 +6759,9 @@ put_sub:
 					if (m2->mc_pg[i] == mp) {
 						if (m2->mc_ki[i] == mc->mc_ki[i]) {
 							mdb_xcursor_init2(m2, mx, new_dupdata);
-						} else if (!insert_key) {
+						} else if (!insert_key && m2->mc_ki[i] < nkeys) {
 							MDB_node *n2 = NODEPTR(mp, m2->mc_ki[i]);
-							if (!(n2->mn_flags & F_SUBDATA))
+							if ((n2->mn_flags & (F_SUBDATA|F_DUPDATA)) == F_DUPDATA)
 								m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(n2);
 						}
 					}
@@ -6725,7 +6817,7 @@ mdb_cursor_del(MDB_cursor *mc, unsigned int flags)
 	MDB_page	*mp;
 	int rc;
 
-	if (mc->mc_txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR))
+	if (mc->mc_txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_BLOCKED))
 		return (mc->mc_txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
 
 	if (!(mc->mc_flags & C_INITIALIZED))
@@ -6750,6 +6842,7 @@ mdb_cursor_del(MDB_cursor *mc, unsigned int flags)
 		if (flags & MDB_NODUPDATA) {
 			/* mdb_cursor_del0() will subtract the final entry */
 			mc->mc_db->md_entries -= mc->mc_xcursor->mx_db.md_entries - 1;
+			mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
 		} else {
 			if (!F_ISSET(leaf->mn_flags, F_SUBDATA)) {
 				mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
@@ -6785,8 +6878,9 @@ mdb_cursor_del(MDB_cursor *mc, unsigned int flags)
 					}
 				}
 				mc->mc_db->md_entries--;
-				mc->mc_flags |= C_DEL;
 				return rc;
+			} else {
+				mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
 			}
 			/* otherwise fall thru and delete the sub-DB */
 		}
@@ -6934,6 +7028,7 @@ mdb_node_add(MDB_cursor *mc, indx_t indx,
 	MDB_node	*node;
 	MDB_page	*mp = mc->mc_pg[mc->mc_top];
 	MDB_page	*ofp = NULL;		/* overflow page */
+	void		*ndata;
 	DKBUF;
 
 	mdb_cassert(mc, mp->mp_upper >= mp->mp_lower);
@@ -6964,7 +7059,7 @@ mdb_node_add(MDB_cursor *mc, indx_t indx,
 	if (key != NULL)
 		node_size += key->mv_size;
 	if (IS_LEAF(mp)) {
-		mdb_cassert(mc, data);
+		mdb_cassert(mc, key && data);
 		if (F_ISSET(flags, F_BIGDATA)) {
 			/* Data already on overflow page. */
 			node_size += sizeof(pgno_t);
@@ -7015,23 +7110,21 @@ update:
 		memcpy(NODEKEY(node), key->mv_data, key->mv_size);
 
 	if (IS_LEAF(mp)) {
-		mdb_cassert(mc, key);
+		ndata = NODEDATA(node);
 		if (ofp == NULL) {
 			if (F_ISSET(flags, F_BIGDATA))
-				memcpy(node->mn_data + key->mv_size, data->mv_data,
-				    sizeof(pgno_t));
+				memcpy(ndata, data->mv_data, sizeof(pgno_t));
 			else if (F_ISSET(flags, MDB_RESERVE))
-				data->mv_data = node->mn_data + key->mv_size;
+				data->mv_data = ndata;
 			else
-				memcpy(node->mn_data + key->mv_size, data->mv_data,
-				    data->mv_size);
+				memcpy(ndata, data->mv_data, data->mv_size);
 		} else {
-			memcpy(node->mn_data + key->mv_size, &ofp->mp_pgno,
-			    sizeof(pgno_t));
+			memcpy(ndata, &ofp->mp_pgno, sizeof(pgno_t));
+			ndata = METADATA(ofp);
 			if (F_ISSET(flags, MDB_RESERVE))
-				data->mv_data = METADATA(ofp);
+				data->mv_data = ndata;
 			else
-				memcpy(METADATA(ofp), data->mv_data, data->mv_size);
+				memcpy(ndata, data->mv_data, data->mv_size);
 		}
 	}
 
@@ -7113,45 +7206,38 @@ mdb_node_shrink(MDB_page *mp, indx_t indx)
 	MDB_node *node;
 	MDB_page *sp, *xp;
 	char *base;
-	int nsize, delta;
-	indx_t		 i, numkeys, ptr;
+	indx_t delta, nsize, len, ptr;
+	int i;
 
 	node = NODEPTR(mp, indx);
 	sp = (MDB_page *)NODEDATA(node);
 	delta = SIZELEFT(sp);
-	xp = (MDB_page *)((char *)sp + delta);
+	nsize = NODEDSZ(node) - delta;
 
-	/* shift subpage upward */
+	/* Prepare to shift upward, set len = length(subpage part to shift) */
 	if (IS_LEAF2(sp)) {
-		nsize = NUMKEYS(sp) * sp->mp_pad;
+		len = nsize;
 		if (nsize & 1)
 			return;		/* do not make the node uneven-sized */
-		memmove(METADATA(xp), METADATA(sp), nsize);
 	} else {
-		int i;
-		numkeys = NUMKEYS(sp);
-		for (i=numkeys-1; i>=0; i--)
+		xp = (MDB_page *)((char *)sp + delta); /* destination subpage */
+		for (i = NUMKEYS(sp); --i >= 0; )
 			xp->mp_ptrs[i] = sp->mp_ptrs[i] - delta;
+		len = PAGEHDRSZ;
 	}
-	xp->mp_upper = sp->mp_lower;
-	xp->mp_lower = sp->mp_lower;
-	xp->mp_flags = sp->mp_flags;
-	xp->mp_pad = sp->mp_pad;
-	COPY_PGNO(xp->mp_pgno, mp->mp_pgno);
-
-	nsize = NODEDSZ(node) - delta;
+	sp->mp_upper = sp->mp_lower;
+	COPY_PGNO(sp->mp_pgno, mp->mp_pgno);
 	SETDSZ(node, nsize);
 
-	/* shift lower nodes upward */
+	/* Shift <lower nodes...initial part of subpage> upward */
+	base = (char *)mp + mp->mp_upper + PAGEBASE;
+	memmove(base + delta, base, (char *)sp + len - base);
+
 	ptr = mp->mp_ptrs[indx];
-	numkeys = NUMKEYS(mp);
-	for (i = 0; i < numkeys; i++) {
+	for (i = NUMKEYS(mp); --i >= 0; ) {
 		if (mp->mp_ptrs[i] <= ptr)
 			mp->mp_ptrs[i] += delta;
 	}
-
-	base = (char *)mp + mp->mp_upper + PAGEBASE;
-	memmove(base + delta, base, ptr - mp->mp_upper + NODESIZE + NODEKSZ(node));
 	mp->mp_upper += delta;
 }
 
@@ -7226,7 +7312,7 @@ mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node)
 	}
 	DPRINTF(("Sub-db -%u root page %"Z"u", mx->mx_cursor.mc_dbi,
 		mx->mx_db.md_root));
-	mx->mx_dbflag = DB_VALID|DB_DIRTY; /* DB_DIRTY guides mdb_cursor_touch */
+	mx->mx_dbflag = DB_VALID|DB_USRVALID|DB_DIRTY; /* DB_DIRTY guides mdb_cursor_touch */
 #if UINT_MAX < SIZE_MAX
 	if (mx->mx_dbx.md_cmp == mdb_cmp_int && mx->mx_db.md_pad == sizeof(size_t))
 		mx->mx_dbx.md_cmp = mdb_cmp_clong;
@@ -7299,14 +7385,13 @@ mdb_cursor_open(MDB_txn *txn, MDB_dbi dbi, MDB_cursor **ret)
 	MDB_cursor	*mc;
 	size_t size = sizeof(MDB_cursor);
 
-	if (!ret || !TXN_DBI_EXIST(txn, dbi))
+	if (!ret || !TXN_DBI_EXIST(txn, dbi, DB_VALID))
 		return EINVAL;
 
-	if (txn->mt_flags & MDB_TXN_ERROR)
+	if (txn->mt_flags & MDB_TXN_BLOCKED)
 		return MDB_BAD_TXN;
 
-	/* Allow read access to the freelist */
-	if (!dbi && !F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
+	if (dbi == FREE_DBI && !F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
 		return EINVAL;
 
 	if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT)
@@ -7331,13 +7416,13 @@ mdb_cursor_open(MDB_txn *txn, MDB_dbi dbi, MDB_cursor **ret)
 int
 mdb_cursor_renew(MDB_txn *txn, MDB_cursor *mc)
 {
-	if (!mc || !TXN_DBI_EXIST(txn, mc->mc_dbi))
+	if (!mc || !TXN_DBI_EXIST(txn, mc->mc_dbi, DB_VALID))
 		return EINVAL;
 
 	if ((mc->mc_flags & C_UNTRACK) || txn->mt_cursors)
 		return EINVAL;
 
-	if (txn->mt_flags & MDB_TXN_ERROR)
+	if (txn->mt_flags & MDB_TXN_BLOCKED)
 		return MDB_BAD_TXN;
 
 	mdb_cursor_init(mc, txn, mc->mc_dbi, mc->mc_xcursor);
@@ -7356,7 +7441,7 @@ mdb_cursor_count(MDB_cursor *mc, size_t *countp)
 	if (mc->mc_xcursor == NULL)
 		return MDB_INCOMPATIBLE;
 
-	if (mc->mc_txn->mt_flags & MDB_TXN_ERROR)
+	if (mc->mc_txn->mt_flags & MDB_TXN_BLOCKED)
 		return MDB_BAD_TXN;
 
 	if (!(mc->mc_flags & C_INITIALIZED))
@@ -7482,10 +7567,26 @@ mdb_update_key(MDB_cursor *mc, MDB_val *key)
 static void
 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst);
 
+/** Perform \b act while tracking temporary cursor \b mn */
+#define WITH_CURSOR_TRACKING(mn, act) do { \
+	MDB_cursor dummy, *tracked, **tp = &(mn).mc_txn->mt_cursors[mn.mc_dbi]; \
+	if ((mn).mc_flags & C_SUB) { \
+		dummy.mc_flags =  C_INITIALIZED; \
+		dummy.mc_xcursor = (MDB_xcursor *)&(mn);	\
+		tracked = &dummy; \
+	} else { \
+		tracked = &(mn); \
+	} \
+	tracked->mc_next = *tp; \
+	*tp = tracked; \
+	{ act; } \
+	*tp = tracked->mc_next; \
+} while (0)
+
 /** Move a node from csrc to cdst.
  */
 static int
-mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst)
+mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst, int fromleft)
 {
 	MDB_node		*srcnode;
 	MDB_val		 key, data;
@@ -7537,6 +7638,7 @@ mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst)
 		data.mv_size = NODEDSZ(srcnode);
 		data.mv_data = NODEDATA(srcnode);
 	}
+	mn.mc_xcursor = NULL;
 	if (IS_BRANCH(cdst->mc_pg[cdst->mc_top]) && cdst->mc_ki[cdst->mc_top] == 0) {
 		unsigned int snum = cdst->mc_snum;
 		MDB_node *s2;
@@ -7583,32 +7685,64 @@ mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst)
 		/* Adjust other cursors pointing to mp */
 		MDB_cursor *m2, *m3;
 		MDB_dbi dbi = csrc->mc_dbi;
-		MDB_page *mp;
-
-		mp = cdst->mc_pg[csrc->mc_top];
-		for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
-			if (csrc->mc_flags & C_SUB)
-				m3 = &m2->mc_xcursor->mx_cursor;
-			else
-				m3 = m2;
-			if (m3 == cdst) continue;
-			if (m3->mc_pg[csrc->mc_top] == mp && m3->mc_ki[csrc->mc_top] >=
-				cdst->mc_ki[csrc->mc_top]) {
-				m3->mc_ki[csrc->mc_top]++;
+		MDB_page *mpd, *mps;
+
+		mps = csrc->mc_pg[csrc->mc_top];
+		/* If we're adding on the left, bump others up */
+		if (fromleft) {
+			mpd = cdst->mc_pg[csrc->mc_top];
+			for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
+				if (csrc->mc_flags & C_SUB)
+					m3 = &m2->mc_xcursor->mx_cursor;
+				else
+					m3 = m2;
+				if (!(m3->mc_flags & C_INITIALIZED) || m3->mc_top < csrc->mc_top)
+					continue;
+				if (m3 != cdst &&
+					m3->mc_pg[csrc->mc_top] == mpd &&
+					m3->mc_ki[csrc->mc_top] >= cdst->mc_ki[csrc->mc_top]) {
+					m3->mc_ki[csrc->mc_top]++;
+				}
+				if (m3 !=csrc &&
+					m3->mc_pg[csrc->mc_top] == mps &&
+					m3->mc_ki[csrc->mc_top] == csrc->mc_ki[csrc->mc_top]) {
+					m3->mc_pg[csrc->mc_top] = cdst->mc_pg[cdst->mc_top];
+					m3->mc_ki[csrc->mc_top] = cdst->mc_ki[cdst->mc_top];
+					m3->mc_ki[csrc->mc_top-1]++;
+				}
+				if (m3->mc_xcursor && (m3->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) &&
+					IS_LEAF(mps)) {
+					MDB_node *node = NODEPTR(m3->mc_pg[csrc->mc_top], m3->mc_ki[csrc->mc_top]);
+					if ((node->mn_flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA)
+						m3->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(node);
+				}
 			}
-		}
-
-		mp = csrc->mc_pg[csrc->mc_top];
-		for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
-			if (csrc->mc_flags & C_SUB)
-				m3 = &m2->mc_xcursor->mx_cursor;
-			else
-				m3 = m2;
-			if (m3 == csrc) continue;
-			if (m3->mc_pg[csrc->mc_top] == mp && m3->mc_ki[csrc->mc_top] ==
-				csrc->mc_ki[csrc->mc_top]) {
-				m3->mc_pg[csrc->mc_top] = cdst->mc_pg[cdst->mc_top];
-				m3->mc_ki[csrc->mc_top] = cdst->mc_ki[cdst->mc_top];
+		} else
+		/* Adding on the right, bump others down */
+		{
+			for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
+				if (csrc->mc_flags & C_SUB)
+					m3 = &m2->mc_xcursor->mx_cursor;
+				else
+					m3 = m2;
+				if (m3 == csrc) continue;
+				if (!(m3->mc_flags & C_INITIALIZED) || m3->mc_top < csrc->mc_top)
+					continue;
+				if (m3->mc_pg[csrc->mc_top] == mps) {
+					if (!m3->mc_ki[csrc->mc_top]) {
+						m3->mc_pg[csrc->mc_top] = cdst->mc_pg[cdst->mc_top];
+						m3->mc_ki[csrc->mc_top] = cdst->mc_ki[cdst->mc_top];
+						m3->mc_ki[csrc->mc_top-1]--;
+					} else {
+						m3->mc_ki[csrc->mc_top]--;
+					}
+					if (m3->mc_xcursor && (m3->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) &&
+						IS_LEAF(mps)) {
+						MDB_node *node = NODEPTR(m3->mc_pg[csrc->mc_top], m3->mc_ki[csrc->mc_top]);
+						if ((node->mn_flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA)
+							m3->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(node);
+					}
+				}
 			}
 		}
 	}
@@ -7629,7 +7763,10 @@ mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst)
 			mdb_cursor_copy(csrc, &mn);
 			mn.mc_snum--;
 			mn.mc_top--;
-			if ((rc = mdb_update_key(&mn, &key)) != MDB_SUCCESS)
+			/* We want mdb_rebalance to find mn when doing fixups */
+			WITH_CURSOR_TRACKING(mn,
+				rc = mdb_update_key(&mn, &key));
+			if (rc)
 				return rc;
 		}
 		if (IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
@@ -7657,7 +7794,10 @@ mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst)
 			mdb_cursor_copy(cdst, &mn);
 			mn.mc_snum--;
 			mn.mc_top--;
-			if ((rc = mdb_update_key(&mn, &key)) != MDB_SUCCESS)
+			/* We want mdb_rebalance to find mn when doing fixups */
+			WITH_CURSOR_TRACKING(mn,
+				rc = mdb_update_key(&mn, &key));
+			if (rc)
 				return rc;
 		}
 		if (IS_BRANCH(cdst->mc_pg[cdst->mc_top])) {
@@ -7704,6 +7844,9 @@ mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst)
 	if ((rc = mdb_page_touch(cdst)))
 		return rc;
 
+	/* get dst page again now that we've touched it. */
+	pdst = cdst->mc_pg[cdst->mc_top];
+
 	/* Move all nodes from src to dst.
 	 */
 	j = nkeys = NUMKEYS(pdst);
@@ -7723,6 +7866,7 @@ mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst)
 				MDB_cursor mn;
 				MDB_node *s2;
 				mdb_cursor_copy(csrc, &mn);
+				mn.mc_xcursor = NULL;
 				/* must find the lowest key below src */
 				rc = mdb_page_search_lowest(&mn);
 				if (rc)
@@ -7781,6 +7925,7 @@ mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst)
 		/* Adjust other cursors pointing to mp */
 		MDB_cursor *m2, *m3;
 		MDB_dbi dbi = csrc->mc_dbi;
+		unsigned int top = csrc->mc_top;
 
 		for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
 			if (csrc->mc_flags & C_SUB)
@@ -7789,9 +7934,19 @@ mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst)
 				m3 = m2;
 			if (m3 == csrc) continue;
 			if (m3->mc_snum < csrc->mc_snum) continue;
-			if (m3->mc_pg[csrc->mc_top] == psrc) {
-				m3->mc_pg[csrc->mc_top] = pdst;
-				m3->mc_ki[csrc->mc_top] += nkeys;
+			if (m3->mc_pg[top] == psrc) {
+				m3->mc_pg[top] = pdst;
+				m3->mc_ki[top] += nkeys;
+				m3->mc_ki[top-1] = cdst->mc_ki[top-1];
+			} else if (m3->mc_pg[top-1] == csrc->mc_pg[top-1] &&
+				m3->mc_ki[top-1] > csrc->mc_ki[top-1]) {
+				m3->mc_ki[top-1]--;
+			}
+			if (m3->mc_xcursor && (m3->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) &&
+				IS_LEAF(psrc)) {
+				MDB_node *node = NODEPTR(m3->mc_pg[top], m3->mc_ki[top]);
+				if ((node->mn_flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA)
+					m3->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(node);
 			}
 		}
 	}
@@ -7841,7 +7996,7 @@ static int
 mdb_rebalance(MDB_cursor *mc)
 {
 	MDB_node	*node;
-	int rc;
+	int rc, fromleft;
 	unsigned int ptop, minkeys, thresh;
 	MDB_cursor	mn;
 	indx_t oldki;
@@ -7892,7 +8047,8 @@ mdb_rebalance(MDB_cursor *mc)
 						m3 = &m2->mc_xcursor->mx_cursor;
 					else
 						m3 = m2;
-					if (m3->mc_snum < mc->mc_snum) continue;
+					if (!(m3->mc_flags & C_INITIALIZED) || (m3->mc_snum < mc->mc_snum))
+						continue;
 					if (m3->mc_pg[0] == mp) {
 						m3->mc_snum = 0;
 						m3->mc_top = 0;
@@ -7927,9 +8083,11 @@ mdb_rebalance(MDB_cursor *mc)
 						m3 = &m2->mc_xcursor->mx_cursor;
 					else
 						m3 = m2;
-					if (m3 == mc || m3->mc_snum < mc->mc_snum) continue;
+					if (m3 == mc) continue;
+					if (!(m3->mc_flags & C_INITIALIZED))
+						continue;
 					if (m3->mc_pg[0] == mp) {
-						for (i=0; i<m3->mc_snum; i++) {
+						for (i=0; i<mc->mc_db->md_depth; i++) {
 							m3->mc_pg[i] = m3->mc_pg[i+1];
 							m3->mc_ki[i] = m3->mc_ki[i+1];
 						}
@@ -7971,6 +8129,7 @@ mdb_rebalance(MDB_cursor *mc)
 			return rc;
 		mn.mc_ki[mn.mc_top] = 0;
 		mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]);
+		fromleft = 0;
 	} else {
 		/* There is at least one neighbor to the left.
 		 */
@@ -7982,6 +8141,7 @@ mdb_rebalance(MDB_cursor *mc)
 			return rc;
 		mn.mc_ki[mn.mc_top] = NUMKEYS(mn.mc_pg[mn.mc_top]) - 1;
 		mc->mc_ki[mc->mc_top] = 0;
+		fromleft = 1;
 	}
 
 	DPRINTF(("found neighbor page %"Z"u (%u keys, %.1f%% full)",
@@ -7993,31 +8153,20 @@ mdb_rebalance(MDB_cursor *mc)
 	 * (A branch page must never have less than 2 keys.)
 	 */
 	if (PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) >= thresh && NUMKEYS(mn.mc_pg[mn.mc_top]) > minkeys) {
-		rc = mdb_node_move(&mn, mc);
-		if (mc->mc_ki[mc->mc_top-1]) {
+		rc = mdb_node_move(&mn, mc, fromleft);
+		if (fromleft) {
+			/* if we inserted on left, bump position up */
 			oldki++;
 		}
 	} else {
-		if (mc->mc_ki[ptop] == 0) {
+		if (!fromleft) {
 			rc = mdb_page_merge(&mn, mc);
 		} else {
-			MDB_cursor dummy;
 			oldki += NUMKEYS(mn.mc_pg[mn.mc_top]);
 			mn.mc_ki[mn.mc_top] += mc->mc_ki[mn.mc_top] + 1;
 			/* We want mdb_rebalance to find mn when doing fixups */
-			if (mc->mc_flags & C_SUB) {
-				dummy.mc_next = mc->mc_txn->mt_cursors[mc->mc_dbi];
-				mc->mc_txn->mt_cursors[mc->mc_dbi] = &dummy;
-				dummy.mc_xcursor = (MDB_xcursor *)&mn;
-			} else {
-				mn.mc_next = mc->mc_txn->mt_cursors[mc->mc_dbi];
-				mc->mc_txn->mt_cursors[mc->mc_dbi] = &mn;
-			}
-			rc = mdb_page_merge(mc, &mn);
-			if (mc->mc_flags & C_SUB)
-				mc->mc_txn->mt_cursors[mc->mc_dbi] = dummy.mc_next;
-			else
-				mc->mc_txn->mt_cursors[mc->mc_dbi] = mn.mc_next;
+			WITH_CURSOR_TRACKING(mn,
+				rc = mdb_page_merge(mc, &mn));
 			mdb_cursor_copy(&mn, mc);
 		}
 		mc->mc_flags &= ~C_EOF;
@@ -8050,12 +8199,17 @@ mdb_cursor_del0(MDB_cursor *mc)
 			if (m3 == mc || m3->mc_snum < mc->mc_snum)
 				continue;
 			if (m3->mc_pg[mc->mc_top] == mp) {
-				if (m3->mc_ki[mc->mc_top] >= ki) {
+				if (m3->mc_ki[mc->mc_top] == ki) {
 					m3->mc_flags |= C_DEL;
-					if (m3->mc_ki[mc->mc_top] > ki)
-						m3->mc_ki[mc->mc_top]--;
-					else if (mc->mc_db->md_flags & MDB_DUPSORT)
-						m3->mc_xcursor->mx_cursor.mc_flags |= C_EOF;
+					if (mc->mc_db->md_flags & MDB_DUPSORT)
+						m3->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
+				} else if (m3->mc_ki[mc->mc_top] > ki) {
+					m3->mc_ki[mc->mc_top]--;
+				}
+				if (m3->mc_xcursor && (m3->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
+					MDB_node *node = NODEPTR(m3->mc_pg[mc->mc_top], m3->mc_ki[mc->mc_top]);
+					if ((node->mn_flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA)
+						m3->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(node);
 				}
 			}
 		}
@@ -8103,10 +8257,10 @@ int
 mdb_del(MDB_txn *txn, MDB_dbi dbi,
     MDB_val *key, MDB_val *data)
 {
-	if (!key || dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
+	if (!key || !TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
 		return EINVAL;
 
-	if (txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR))
+	if (txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_BLOCKED))
 		return (txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
 
 	if (!F_ISSET(txn->mt_dbs[dbi].md_flags, MDB_DUPSORT)) {
@@ -8202,12 +8356,19 @@ mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata, pgno_t newpgno
 	rp->mp_pad = mp->mp_pad;
 	DPRINTF(("new right sibling: page %"Z"u", rp->mp_pgno));
 
-	if (mc->mc_snum < 2) {
+	/* Usually when splitting the root page, the cursor
+	 * height is 1. But when called from mdb_update_key,
+	 * the cursor height may be greater because it walks
+	 * up the stack while finding the branch slot to update.
+	 */
+	if (mc->mc_top < 1) {
 		if ((rc = mdb_page_new(mc, P_BRANCH, 1, &pp)))
 			goto done;
 		/* shift current top to make room for new parent */
-		mc->mc_pg[1] = mc->mc_pg[0];
-		mc->mc_ki[1] = mc->mc_ki[0];
+		for (i=mc->mc_snum; i>0; i--) {
+			mc->mc_pg[i] = mc->mc_pg[i-1];
+			mc->mc_ki[i] = mc->mc_ki[i-1];
+		}
 		mc->mc_pg[0] = pp;
 		mc->mc_ki[0] = 0;
 		mc->mc_db->md_root = pp->mp_pgno;
@@ -8223,16 +8384,16 @@ mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata, pgno_t newpgno
 			mc->mc_db->md_depth--;
 			goto done;
 		}
-		mc->mc_snum = 2;
-		mc->mc_top = 1;
+		mc->mc_snum++;
+		mc->mc_top++;
 		ptop = 0;
 	} else {
 		ptop = mc->mc_top-1;
 		DPRINTF(("parent branch page is %"Z"u", mc->mc_pg[ptop]->mp_pgno));
 	}
 
-	mc->mc_flags |= C_SPLITTING;
 	mdb_cursor_copy(mc, &mn);
+	mn.mc_xcursor = NULL;
 	mn.mc_pg[mn.mc_top] = rp;
 	mn.mc_ki[ptop] = mc->mc_ki[ptop]+1;
 
@@ -8282,7 +8443,6 @@ mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata, pgno_t newpgno
 				rp->mp_lower += sizeof(indx_t);
 				rp->mp_upper -= ksize - sizeof(indx_t);
 				mc->mc_ki[mc->mc_top] = x;
-				mc->mc_pg[mc->mc_top] = rp;
 			}
 		} else {
 			int psize, nsize, k;
@@ -8375,21 +8535,18 @@ mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata, pgno_t newpgno
 	/* Copy separator key to the parent.
 	 */
 	if (SIZELEFT(mn.mc_pg[ptop]) < mdb_branch_size(env, &sepkey)) {
+		int snum = mc->mc_snum;
 		mn.mc_snum--;
 		mn.mc_top--;
 		did_split = 1;
-		rc = mdb_page_split(&mn, &sepkey, NULL, rp->mp_pgno, 0);
+		/* We want other splits to find mn when doing fixups */
+		WITH_CURSOR_TRACKING(mn,
+			rc = mdb_page_split(&mn, &sepkey, NULL, rp->mp_pgno, 0));
 		if (rc)
 			goto done;
 
 		/* root split? */
-		if (mn.mc_snum == mc->mc_snum) {
-			mc->mc_pg[mc->mc_snum] = mc->mc_pg[mc->mc_top];
-			mc->mc_ki[mc->mc_snum] = mc->mc_ki[mc->mc_top];
-			mc->mc_pg[mc->mc_top] = mc->mc_pg[ptop];
-			mc->mc_ki[mc->mc_top] = mc->mc_ki[ptop];
-			mc->mc_snum++;
-			mc->mc_top++;
+		if (mc->mc_snum > snum) {
 			ptop++;
 		}
 		/* Right page might now have changed parent.
@@ -8415,7 +8572,6 @@ mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata, pgno_t newpgno
 		rc = mdb_node_add(&mn, mn.mc_ki[ptop], &sepkey, NULL, rp->mp_pgno, 0);
 		mn.mc_top++;
 	}
-	mc->mc_flags ^= C_SPLITTING;
 	if (rc != MDB_SUCCESS) {
 		goto done;
 	}
@@ -8485,11 +8641,6 @@ mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata, pgno_t newpgno
 		/* reset back to original page */
 		if (newindx < split_indx) {
 			mc->mc_pg[mc->mc_top] = mp;
-			if (nflags & MDB_RESERVE) {
-				node = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
-				if (!(node->mn_flags & F_BIGDATA))
-					newdata->mv_data = NODEDATA(node);
-			}
 		} else {
 			mc->mc_pg[mc->mc_top] = rp;
 			mc->mc_ki[ptop]++;
@@ -8503,13 +8654,32 @@ mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata, pgno_t newpgno
 				}
 			}
 		}
+		if (nflags & MDB_RESERVE) {
+			node = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
+			if (!(node->mn_flags & F_BIGDATA))
+				newdata->mv_data = NODEDATA(node);
+		}
+	} else {
+		if (newindx >= split_indx) {
+			mc->mc_pg[mc->mc_top] = rp;
+			mc->mc_ki[ptop]++;
+			/* Make sure mc_ki is still valid.
+			 */
+			if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
+				mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
+				for (i=0; i<=ptop; i++) {
+					mc->mc_pg[i] = mn.mc_pg[i];
+					mc->mc_ki[i] = mn.mc_ki[i];
+				}
+			}
+		}
 	}
 
 	{
 		/* Adjust other cursors pointing to mp */
 		MDB_cursor *m2, *m3;
 		MDB_dbi dbi = mc->mc_dbi;
-		int fixup = NUMKEYS(mp);
+		nkeys = NUMKEYS(mp);
 
 		for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
 			if (mc->mc_flags & C_SUB)
@@ -8520,16 +8690,17 @@ mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata, pgno_t newpgno
 				continue;
 			if (!(m2->mc_flags & m3->mc_flags & C_INITIALIZED))
 				continue;
-			if (m3->mc_flags & C_SPLITTING)
-				continue;
 			if (new_root) {
 				int k;
+				/* sub cursors may be on different DB */
+				if (m3->mc_pg[0] != mp)
+					continue;
 				/* root split */
 				for (k=new_root; k>=0; k--) {
 					m3->mc_ki[k+1] = m3->mc_ki[k];
 					m3->mc_pg[k+1] = m3->mc_pg[k];
 				}
-				if (m3->mc_ki[0] >= split_indx) {
+				if (m3->mc_ki[0] >= nkeys) {
 					m3->mc_ki[0] = 1;
 				} else {
 					m3->mc_ki[0] = 0;
@@ -8541,15 +8712,24 @@ mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata, pgno_t newpgno
 			if (m3->mc_top >= mc->mc_top && m3->mc_pg[mc->mc_top] == mp) {
 				if (m3->mc_ki[mc->mc_top] >= newindx && !(nflags & MDB_SPLIT_REPLACE))
 					m3->mc_ki[mc->mc_top]++;
-				if (m3->mc_ki[mc->mc_top] >= fixup) {
+				if (m3->mc_ki[mc->mc_top] >= nkeys) {
 					m3->mc_pg[mc->mc_top] = rp;
-					m3->mc_ki[mc->mc_top] -= fixup;
-					m3->mc_ki[ptop] = mn.mc_ki[ptop];
+					m3->mc_ki[mc->mc_top] -= nkeys;
+					for (i=0; i<mc->mc_top; i++) {
+						m3->mc_ki[i] = mn.mc_ki[i];
+						m3->mc_pg[i] = mn.mc_pg[i];
+					}
 				}
 			} else if (!did_split && m3->mc_top >= ptop && m3->mc_pg[ptop] == mc->mc_pg[ptop] &&
 				m3->mc_ki[ptop] >= mc->mc_ki[ptop]) {
 				m3->mc_ki[ptop]++;
 			}
+			if (m3->mc_xcursor && (m3->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) &&
+				IS_LEAF(mp)) {
+				MDB_node *node = NODEPTR(m3->mc_pg[mc->mc_top], m3->mc_ki[mc->mc_top]);
+				if ((node->mn_flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA)
+					m3->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(node);
+			}
 		}
 	}
 	DPRINTF(("mp left: %d, rp left: %d", SIZELEFT(mp), SIZELEFT(rp)));
@@ -8568,15 +8748,23 @@ mdb_put(MDB_txn *txn, MDB_dbi dbi,
 {
 	MDB_cursor mc;
 	MDB_xcursor mx;
+	int rc;
 
-	if (!key || !data || dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
+	if (!key || !data || !TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
 		return EINVAL;
 
-	if ((flags & (MDB_NOOVERWRITE|MDB_NODUPDATA|MDB_RESERVE|MDB_APPEND|MDB_APPENDDUP)) != flags)
+	if (flags & ~(MDB_NOOVERWRITE|MDB_NODUPDATA|MDB_RESERVE|MDB_APPEND|MDB_APPENDDUP))
 		return EINVAL;
 
+	if (txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_BLOCKED))
+		return (txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
+
 	mdb_cursor_init(&mc, txn, dbi, &mx);
-	return mdb_cursor_put(&mc, key, data, flags);
+	mc.mc_next = txn->mt_cursors[dbi];
+	txn->mt_cursors[dbi] = &mc;
+	rc = mdb_cursor_put(&mc, key, data, flags);
+	txn->mt_cursors[dbi] = mc.mc_next;
+	return rc;
 }
 
 #ifndef MDB_WBUF
@@ -8879,7 +9067,7 @@ mdb_env_copyfd1(MDB_env *env, HANDLE fd)
 	my.mc_wlen[1] = 0;
 	my.mc_olen[0] = 0;
 	my.mc_olen[1] = 0;
-	my.mc_next_pgno = 2;
+	my.mc_next_pgno = NUM_METAS;
 	my.mc_status = 0;
 	my.mc_new = 1;
 	my.mc_toggle = 0;
@@ -8892,7 +9080,7 @@ mdb_env_copyfd1(MDB_env *env, HANDLE fd)
 		return rc;
 
 	mp = (MDB_page *)my.mc_wbuf[0];
-	memset(mp, 0, 2*env->me_psize);
+	memset(mp, 0, NUM_METAS * env->me_psize);
 	mp->mp_pgno = 0;
 	mp->mp_flags = P_META;
 	mm = (MDB_meta *)METADATA(mp);
@@ -8915,27 +9103,27 @@ mdb_env_copyfd1(MDB_env *env, HANDLE fd)
 		mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
 		while ((rc = mdb_cursor_get(&mc, &key, &data, MDB_NEXT)) == 0)
 			freecount += *(MDB_ID *)data.mv_data;
-		freecount += txn->mt_dbs[0].md_branch_pages +
-			txn->mt_dbs[0].md_leaf_pages +
-			txn->mt_dbs[0].md_overflow_pages;
+		freecount += txn->mt_dbs[FREE_DBI].md_branch_pages +
+			txn->mt_dbs[FREE_DBI].md_leaf_pages +
+			txn->mt_dbs[FREE_DBI].md_overflow_pages;
 
 		/* Set metapage 1 */
 		mm->mm_last_pg = txn->mt_next_pgno - freecount - 1;
-		mm->mm_dbs[1] = txn->mt_dbs[1];
-		if (mm->mm_last_pg > 1) {
-			mm->mm_dbs[1].md_root = mm->mm_last_pg;
+		mm->mm_dbs[MAIN_DBI] = txn->mt_dbs[MAIN_DBI];
+		if (mm->mm_last_pg > NUM_METAS-1) {
+			mm->mm_dbs[MAIN_DBI].md_root = mm->mm_last_pg;
 			mm->mm_txnid = 1;
 		} else {
-			mm->mm_dbs[1].md_root = P_INVALID;
+			mm->mm_dbs[MAIN_DBI].md_root = P_INVALID;
 		}
 	}
-	my.mc_wlen[0] = env->me_psize * 2;
+	my.mc_wlen[0] = env->me_psize * NUM_METAS;
 	my.mc_txn = txn;
 	pthread_mutex_lock(&my.mc_mutex);
 	while(my.mc_new)
 		pthread_cond_wait(&my.mc_cond, &my.mc_mutex);
 	pthread_mutex_unlock(&my.mc_mutex);
-	rc = mdb_env_cwalk(&my, &txn->mt_dbs[1].md_root, 0);
+	rc = mdb_env_cwalk(&my, &txn->mt_dbs[MAIN_DBI].md_root, 0);
 	if (rc == MDB_SUCCESS && my.mc_wlen[my.mc_toggle])
 		rc = mdb_env_cthr_toggle(&my, 1);
 	mdb_env_cthr_toggle(&my, -1);
@@ -8985,7 +9173,7 @@ mdb_env_copyfd0(MDB_env *env, HANDLE fd)
 
 	if (env->me_txns) {
 		/* We must start the actual read txn after blocking writers */
-		mdb_txn_reset0(txn, "reset-stage1");
+		mdb_txn_end(txn, MDB_END_RESET_TMP);
 
 		/* Temporarily block writers until we snapshot the meta pages */
 		wmutex = env->me_wmutex;
@@ -8999,7 +9187,7 @@ mdb_env_copyfd0(MDB_env *env, HANDLE fd)
 		}
 	}
 
-	wsize = env->me_psize * 2;
+	wsize = env->me_psize * NUM_METAS;
 	ptr = env->me_map;
 	w2 = wsize;
 	while (w2 > 0) {
@@ -9079,6 +9267,9 @@ mdb_env_copy2(MDB_env *env, const char *path, unsigned int flags)
 	int rc, len;
 	char *lpath;
 	HANDLE newfd = INVALID_HANDLE_VALUE;
+#ifdef _WIN32
+	wchar_t *wpath;
+#endif
 
 	if (env->me_flags & MDB_NOSUBDIR) {
 		lpath = (char *)path;
@@ -9096,8 +9287,12 @@ mdb_env_copy2(MDB_env *env, const char *path, unsigned int flags)
 	 * already in the OS cache.
 	 */
 #ifdef _WIN32
-	newfd = CreateFile(lpath, GENERIC_WRITE, 0, NULL, CREATE_NEW,
+	rc = utf8_to_utf16(lpath, -1, &wpath, NULL);
+	if (rc)
+		return rc;
+	newfd = CreateFileW(wpath, GENERIC_WRITE, 0, NULL, CREATE_NEW,
 				FILE_FLAG_NO_BUFFERING|FILE_FLAG_WRITE_THROUGH, NULL);
+	free(wpath);
 #else
 	newfd = open(lpath, O_WRONLY|O_CREAT|O_EXCL, 0666);
 #endif
@@ -9142,7 +9337,7 @@ mdb_env_copy(MDB_env *env, const char *path)
 int ESECT
 mdb_env_set_flags(MDB_env *env, unsigned int flag, int onoff)
 {
-	if ((flag & CHANGEABLE) != flag)
+	if (flag & ~CHANGEABLE)
 		return EINVAL;
 	if (onoff)
 		env->me_flags |= flag;
@@ -9229,32 +9424,32 @@ mdb_stat0(MDB_env *env, MDB_db *db, MDB_stat *arg)
 int ESECT
 mdb_env_stat(MDB_env *env, MDB_stat *arg)
 {
-	int toggle;
+	MDB_meta *meta;
 
 	if (env == NULL || arg == NULL)
 		return EINVAL;
 
-	toggle = mdb_env_pick_meta(env);
+	meta = mdb_env_pick_meta(env);
 
-	return mdb_stat0(env, &env->me_metas[toggle]->mm_dbs[MAIN_DBI], arg);
+	return mdb_stat0(env, &meta->mm_dbs[MAIN_DBI], arg);
 }
 
 int ESECT
 mdb_env_info(MDB_env *env, MDB_envinfo *arg)
 {
-	int toggle;
+	MDB_meta *meta;
 
 	if (env == NULL || arg == NULL)
 		return EINVAL;
 
-	toggle = mdb_env_pick_meta(env);
-	arg->me_mapaddr = env->me_metas[toggle]->mm_address;
+	meta = mdb_env_pick_meta(env);
+	arg->me_mapaddr = meta->mm_address;
+	arg->me_last_pgno = meta->mm_last_pg;
+	arg->me_last_txnid = meta->mm_txnid;
+
 	arg->me_mapsize = env->me_mapsize;
 	arg->me_maxreaders = env->me_maxreaders;
 	arg->me_numreaders = env->me_txns ? env->me_txns->mti_numreaders : 0;
-
-	arg->me_last_pgno = env->me_metas[toggle]->mm_last_pg;
-	arg->me_last_txnid = env->me_metas[toggle]->mm_txnid;
 	return MDB_SUCCESS;
 }
 
@@ -9289,15 +9484,12 @@ int mdb_dbi_open(MDB_txn *txn, const char *name, unsigned int flags, MDB_dbi *db
 	MDB_db dummy;
 	int rc, dbflag, exact;
 	unsigned int unused = 0, seq;
+	char *namedup;
 	size_t len;
 
-	if (txn->mt_dbxs[FREE_DBI].md_cmp == NULL) {
-		mdb_default_cmp(txn, FREE_DBI);
-	}
-
-	if ((flags & VALID_FLAGS) != flags)
+	if (flags & ~VALID_FLAGS)
 		return EINVAL;
-	if (txn->mt_flags & MDB_TXN_ERROR)
+	if (txn->mt_flags & MDB_TXN_BLOCKED)
 		return MDB_BAD_TXN;
 
 	/* main DB? */
@@ -9321,7 +9513,7 @@ int mdb_dbi_open(MDB_txn *txn, const char *name, unsigned int flags, MDB_dbi *db
 
 	/* Is the DB already open? */
 	len = strlen(name);
-	for (i=2; i<txn->mt_numdbs; i++) {
+	for (i=CORE_DBS; i<txn->mt_numdbs; i++) {
 		if (!txn->mt_dbxs[i].md_name.mv_size) {
 			/* Remember this free slot */
 			if (!unused) unused = i;
@@ -9343,7 +9535,7 @@ int mdb_dbi_open(MDB_txn *txn, const char *name, unsigned int flags, MDB_dbi *db
 		return (flags & MDB_CREATE) ? MDB_INCOMPATIBLE : MDB_NOTFOUND;
 
 	/* Find the DB info */
-	dbflag = DB_NEW|DB_VALID;
+	dbflag = DB_NEW|DB_VALID|DB_USRVALID;
 	exact = 0;
 	key.mv_size = len;
 	key.mv_data = (void *)name;
@@ -9354,8 +9546,16 @@ int mdb_dbi_open(MDB_txn *txn, const char *name, unsigned int flags, MDB_dbi *db
 		MDB_node *node = NODEPTR(mc.mc_pg[mc.mc_top], mc.mc_ki[mc.mc_top]);
 		if ((node->mn_flags & (F_DUPDATA|F_SUBDATA)) != F_SUBDATA)
 			return MDB_INCOMPATIBLE;
-	} else if (rc == MDB_NOTFOUND && (flags & MDB_CREATE)) {
-		/* Create if requested */
+	} else if (! (rc == MDB_NOTFOUND && (flags & MDB_CREATE))) {
+		return rc;
+	}
+
+	/* Done here so we cannot fail after creating a new DB */
+	if ((namedup = strdup(name)) == NULL)
+		return ENOMEM;
+
+	if (rc) {
+		/* MDB_NOTFOUND and MDB_CREATE: Create new DB */
 		data.mv_size = sizeof(MDB_db);
 		data.mv_data = &dummy;
 		memset(&dummy, 0, sizeof(dummy));
@@ -9365,10 +9565,12 @@ int mdb_dbi_open(MDB_txn *txn, const char *name, unsigned int flags, MDB_dbi *db
 		dbflag |= DB_DIRTY;
 	}
 
-	/* OK, got info, add to table */
-	if (rc == MDB_SUCCESS) {
+	if (rc) {
+		free(namedup);
+	} else {
+		/* Got info, register DBI in this txn */
 		unsigned int slot = unused ? unused : txn->mt_numdbs;
-		txn->mt_dbxs[slot].md_name.mv_data = strdup(name);
+		txn->mt_dbxs[slot].md_name.mv_data = namedup;
 		txn->mt_dbxs[slot].md_name.mv_size = len;
 		txn->mt_dbxs[slot].md_rel = NULL;
 		txn->mt_dbflags[slot] = dbflag;
@@ -9389,12 +9591,13 @@ int mdb_dbi_open(MDB_txn *txn, const char *name, unsigned int flags, MDB_dbi *db
 	return rc;
 }
 
-int mdb_stat(MDB_txn *txn, MDB_dbi dbi, MDB_stat *arg)
+int ESECT
+mdb_stat(MDB_txn *txn, MDB_dbi dbi, MDB_stat *arg)
 {
-	if (!arg || !TXN_DBI_EXIST(txn, dbi))
+	if (!arg || !TXN_DBI_EXIST(txn, dbi, DB_VALID))
 		return EINVAL;
 
-	if (txn->mt_flags & MDB_TXN_ERROR)
+	if (txn->mt_flags & MDB_TXN_BLOCKED)
 		return MDB_BAD_TXN;
 
 	if (txn->mt_dbflags[dbi] & DB_STALE) {
@@ -9409,7 +9612,7 @@ int mdb_stat(MDB_txn *txn, MDB_dbi dbi, MDB_stat *arg)
 void mdb_dbi_close(MDB_env *env, MDB_dbi dbi)
 {
 	char *ptr;
-	if (dbi <= MAIN_DBI || dbi >= env->me_maxdbs)
+	if (dbi < CORE_DBS || dbi >= env->me_maxdbs)
 		return;
 	ptr = env->me_dbxs[dbi].md_name.mv_data;
 	/* If there was no name, this was already closed */
@@ -9425,7 +9628,7 @@ void mdb_dbi_close(MDB_env *env, MDB_dbi dbi)
 int mdb_dbi_flags(MDB_txn *txn, MDB_dbi dbi, unsigned int *flags)
 {
 	/* We could return the flags for the FREE_DBI too but what's the point? */
-	if (dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
+	if (!TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
 		return EINVAL;
 	*flags = txn->mt_dbs[dbi].md_flags & PERSISTENT_FLAGS;
 	return MDB_SUCCESS;
@@ -9448,8 +9651,10 @@ mdb_drop0(MDB_cursor *mc, int subs)
 		MDB_cursor mx;
 		unsigned int i;
 
-		/* LEAF2 pages have no nodes, cannot have sub-DBs */
-		if (IS_LEAF2(mc->mc_pg[mc->mc_top]))
+		/* DUPSORT sub-DBs have no ovpages/DBs. Omit scanning leaves.
+		 * This also avoids any P_LEAF2 pages, which have no nodes.
+		 */
+		if (mc->mc_flags & C_SUB)
 			mdb_cursor_pop(mc);
 
 		mdb_cursor_copy(mc, &mx);
@@ -9515,6 +9720,7 @@ done:
 	} else if (rc == MDB_NOTFOUND) {
 		rc = MDB_SUCCESS;
 	}
+	mc->mc_flags &= ~C_INITIALIZED;
 	return rc;
 }
 
@@ -9523,13 +9729,13 @@ int mdb_drop(MDB_txn *txn, MDB_dbi dbi, int del)
 	MDB_cursor *mc, *m2;
 	int rc;
 
-	if ((unsigned)del > 1 || dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
+	if ((unsigned)del > 1 || !TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
 		return EINVAL;
 
 	if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
 		return EACCES;
 
-	if (dbi > MAIN_DBI && TXN_DBI_CHANGED(txn, dbi))
+	if (TXN_DBI_CHANGED(txn, dbi))
 		return MDB_BAD_DBI;
 
 	rc = mdb_cursor_open(txn, dbi, &mc);
@@ -9544,7 +9750,7 @@ int mdb_drop(MDB_txn *txn, MDB_dbi dbi, int del)
 		goto leave;
 
 	/* Can't delete the main DB */
-	if (del && dbi > MAIN_DBI) {
+	if (del && dbi >= CORE_DBS) {
 		rc = mdb_del0(txn, MAIN_DBI, &mc->mc_dbx->md_name, NULL, F_SUBDATA);
 		if (!rc) {
 			txn->mt_dbflags[dbi] = DB_STALE;
@@ -9571,7 +9777,7 @@ leave:
 
 int mdb_set_compare(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
 {
-	if (dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
+	if (!TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
 		return EINVAL;
 
 	txn->mt_dbxs[dbi].md_cmp = cmp;
@@ -9580,7 +9786,7 @@ int mdb_set_compare(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
 
 int mdb_set_dupsort(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
 {
-	if (dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
+	if (!TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
 		return EINVAL;
 
 	txn->mt_dbxs[dbi].md_dcmp = cmp;
@@ -9589,7 +9795,7 @@ int mdb_set_dupsort(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
 
 int mdb_set_relfunc(MDB_txn *txn, MDB_dbi dbi, MDB_rel_func *rel)
 {
-	if (dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
+	if (!TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
 		return EINVAL;
 
 	txn->mt_dbxs[dbi].md_rel = rel;
@@ -9598,7 +9804,7 @@ int mdb_set_relfunc(MDB_txn *txn, MDB_dbi dbi, MDB_rel_func *rel)
 
 int mdb_set_relctx(MDB_txn *txn, MDB_dbi dbi, void *ctx)
 {
-	if (dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
+	if (!TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
 		return EINVAL;
 
 	txn->mt_dbxs[dbi].md_relctx = ctx;
@@ -9700,7 +9906,8 @@ mdb_reader_check(MDB_env *env, int *dead)
 }
 
 /** As #mdb_reader_check(). rlocked = <caller locked the reader mutex>. */
-static int mdb_reader_check0(MDB_env *env, int rlocked, int *dead)
+static int ESECT
+mdb_reader_check0(MDB_env *env, int rlocked, int *dead)
 {
 	mdb_mutexref_t rmutex = rlocked ? NULL : env->me_rmutex;
 	unsigned int i, j, rdrs;
@@ -9759,9 +9966,11 @@ static int mdb_reader_check0(MDB_env *env, int rlocked, int *dead)
  * @param[in] rc	LOCK_MUTEX0() error (nonzero)
  * @return 0 on success with the mutex locked, or an error code on failure.
  */
-static int mdb_mutex_failed(MDB_env *env, mdb_mutexref_t mutex, int rc)
+static int ESECT
+mdb_mutex_failed(MDB_env *env, mdb_mutexref_t mutex, int rc)
 {
-	int toggle, rlocked, rc2;
+	int rlocked, rc2;
+	MDB_meta *meta;
 
 	if (rc == MDB_OWNERDEAD) {
 		/* We own the mutex. Clean up after dead previous owner. */
@@ -9771,8 +9980,8 @@ static int mdb_mutex_failed(MDB_env *env, mdb_mutexref_t mutex, int rc)
 			/* Keep mti_txnid updated, otherwise next writer can
 			 * overwrite data which latest meta page refers to.
 			 */
-			toggle = mdb_env_pick_meta(env);
-			env->me_txns->mti_txnid = env->me_metas[toggle]->mm_txnid;
+			meta = mdb_env_pick_meta(env);
+			env->me_txns->mti_txnid = meta->mm_txnid;
 			/* env is hosed if the dead thread was ours */
 			if (env->me_txn) {
 				env->me_flags |= MDB_FATAL_ERROR;
@@ -9800,3 +10009,24 @@ static int mdb_mutex_failed(MDB_env *env, mdb_mutexref_t mutex, int rc)
 }
 #endif	/* MDB_ROBUST_SUPPORTED */
 /** @} */
+
+#if defined(_WIN32)
+static int utf8_to_utf16(const char *src, int srcsize, wchar_t **dst, int *dstsize)
+{
+	int need;
+	wchar_t *result;
+	need = MultiByteToWideChar(CP_UTF8, 0, src, srcsize, NULL, 0);
+	if (need == 0xFFFD)
+		return EILSEQ;
+	if (need == 0)
+		return EINVAL;
+	result = malloc(sizeof(wchar_t) * need);
+	if (!result)
+		return ENOMEM;
+	MultiByteToWideChar(CP_UTF8, 0, src, srcsize, result, need);
+	if (dstsize)
+		*dstsize = need;
+	*dst = result;
+	return 0;
+}
+#endif /* defined(_WIN32) */