X-Git-Url: https://git.sur5r.net/?a=blobdiff_plain;f=libraries%2Fliblmdb%2Fmdb.c;h=19930fcfc76780ab63ea8269241ef265b03c1944;hb=1c2a5888cfffdfb3976145634ee5663e4b621675;hp=31cf1884719cdecdda6fb24e3cfc19feaa821054;hpb=828107eb634dbf9ee13f6966e2c78cb27259927f;p=openldap

diff --git a/libraries/liblmdb/mdb.c b/libraries/liblmdb/mdb.c
index 31cf188471..19930fcfc7 100644
--- a/libraries/liblmdb/mdb.c
+++ b/libraries/liblmdb/mdb.c
@@ -35,9 +35,42 @@
 #ifndef _GNU_SOURCE
 #define _GNU_SOURCE 1
 #endif
+#ifdef MDB_VL32
+#define _FILE_OFFSET_BITS	64
+#endif
 #ifdef _WIN32
 #include <malloc.h>
 #include <windows.h>
+
+/* We use native NT APIs to setup the memory map, so that we can
+ * let the DB file grow incrementally instead of always preallocating
+ * the full size. These APIs are defined in <wdm.h> and <ntifs.h>
+ * but those headers are meant for driver-level development and
+ * conflict with the regular user-level headers, so we explicitly
+ * declare them here. Using these APIs also means we must link to
+ * ntdll.dll, which is not linked by default in user code.
+ */
+NTSTATUS WINAPI
+NtCreateSection(OUT PHANDLE sh, IN ACCESS_MASK acc,
+  IN void * oa OPTIONAL,
+  IN PLARGE_INTEGER ms OPTIONAL,
+  IN ULONG pp, IN ULONG aa, IN HANDLE fh OPTIONAL);
+
+typedef enum _SECTION_INHERIT {
+	ViewShare = 1,
+	ViewUnmap = 2
+} SECTION_INHERIT;
+
+NTSTATUS WINAPI
+NtMapViewOfSection(IN PHANDLE sh, IN HANDLE ph,
+  IN OUT PVOID *addr, IN ULONG_PTR zbits,
+  IN SIZE_T cs, IN OUT PLARGE_INTEGER off OPTIONAL,
+  IN OUT PSIZE_T vs, IN SECTION_INHERIT ih,
+  IN ULONG at, IN ULONG pp);
+
+NTSTATUS WINAPI
+NtClose(HANDLE h);
+
 /** getpid() returns int; MinGW defines pid_t but MinGW64 typedefs it
  *  as int64 which is wrong. MSVC doesn't define it at all, so just
  *  don't use it.
@@ -210,7 +243,7 @@ union semun {
 #define ESECT
 #endif
 
-#ifdef _MSC_VER
+#ifdef _WIN32
 #define CALL_CONV WINAPI
 #else
 #define CALL_CONV
@@ -270,8 +303,8 @@ union semun {
 #  define MDB_USE_ROBUST	0
 # else
 #  define MDB_USE_ROBUST	1
-/* glibc < 2.10 only provided _np API */
-#  if defined(__GLIBC__) && GLIBC_VER < 0x02000a
+/* 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)
@@ -422,6 +455,16 @@ typedef pthread_mutex_t mdb_mutex_t[1], *mdb_mutexref_t;
 #define	GET_PAGESIZE(x)	((x) = sysconf(_SC_PAGE_SIZE))
 #endif
 
+#ifdef MDB_VL32
+#ifdef _WIN32
+#define	Y	"I64"
+#else
+#define	Y	"ll"
+#endif
+#else
+#define Y	Z
+#endif
+
 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
 #define MNAME_LEN	32
 #elif defined(MDB_USE_SYSV_SEM)
@@ -1032,7 +1075,7 @@ typedef struct MDB_db {
 	pgno_t		md_branch_pages;	/**< number of internal pages */
 	pgno_t		md_leaf_pages;		/**< number of leaf pages */
 	pgno_t		md_overflow_pages;	/**< number of overflow pages */
-	size_t		md_entries;		/**< number of data items */
+	pgno_t		md_entries;		/**< number of data items */
 	pgno_t		md_root;		/**< the root page of this tree */
 } MDB_db;
 
@@ -1062,8 +1105,16 @@ typedef struct MDB_meta {
 	uint32_t	mm_magic;
 		/** Version number of this file. Must be set to #MDB_DATA_VERSION. */
 	uint32_t	mm_version;
+#ifdef MDB_VL32
+	union {		/* always zero since we don't support fixed mapping in MDB_VL32 */
+		MDB_ID	mmun_ull;
+		void *mmun_address;
+	} mm_un;
+#define	mm_address mm_un.mmun_address
+#else
 	void		*mm_address;		/**< address for fixed mapping */
-	size_t		mm_mapsize;			/**< size of mmap region */
+#endif
+	pgno_t		mm_mapsize;			/**< size of mmap region */
 	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[FREE_DBI].md_pad
@@ -1106,6 +1157,9 @@ struct MDB_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 */
+#ifdef MDB_VL32
+	pgno_t		mt_last_pgno;	/**< last written page */
+#endif
 	/** The ID of this transaction. IDs are integers incrementing from 1.
 	 *	Only committed write transactions increment the ID. If a transaction
 	 *	aborts, the ID may be re-used by the next writer.
@@ -1152,6 +1206,19 @@ struct MDB_txn {
 	MDB_cursor	**mt_cursors;
 	/** Array of flags for each DB */
 	unsigned char	*mt_dbflags;
+#ifdef MDB_VL32
+	/** List of read-only pages (actually chunks) */
+	MDB_ID3L	mt_rpages;
+	/** We map chunks of 16 pages. Even though Windows uses 4KB pages, all
+	 * mappings must begin on 64KB boundaries. So we round off all pgnos to
+	 * a chunk boundary. We do the same on Linux for symmetry, and also to
+	 * reduce the frequency of mmap/munmap calls.
+	 */
+#define MDB_RPAGE_CHUNK	16
+#define MDB_TRPAGE_SIZE	4096	/**< size of #mt_rpages array of chunks */
+#define MDB_TRPAGE_MAX	(MDB_TRPAGE_SIZE-1)	/**< maximum chunk index */
+	unsigned int mt_rpcheck;	/**< threshold for reclaiming unref'd chunks */
+#endif
 	/**	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.
@@ -1231,10 +1298,19 @@ struct MDB_cursor {
 #define C_SUB	0x04			/**< Cursor is a sub-cursor */
 #define C_DEL	0x08			/**< last op was a cursor_del */
 #define C_UNTRACK	0x40		/**< Un-track cursor when closing */
+#define C_WRITEMAP	MDB_TXN_WRITEMAP /**< Copy of txn flag */
+/** Read-only cursor into the txn's original snapshot in the map.
+ *	Set for read-only txns, and in #mdb_page_alloc() for #FREE_DBI when
+ *	#MDB_DEVEL & 2. Only implements code which is necessary for this.
+ */
+#define C_ORIG_RDONLY	MDB_TXN_RDONLY
 /** @} */
 	unsigned int	mc_flags;	/**< @ref mdb_cursor */
 	MDB_page	*mc_pg[CURSOR_STACK];	/**< stack of pushed pages */
 	indx_t		mc_ki[CURSOR_STACK];	/**< stack of page indices */
+#ifdef MDB_VL32
+	MDB_page	*mc_ovpg;		/**< a referenced overflow page */
+#endif
 };
 
 	/** Context for sorted-dup records.
@@ -1264,6 +1340,9 @@ struct MDB_env {
 	HANDLE		me_fd;		/**< The main data file */
 	HANDLE		me_lfd;		/**< The lock file */
 	HANDLE		me_mfd;			/**< just for writing the meta pages */
+#if defined(MDB_VL32) && defined(_WIN32)
+	HANDLE		me_fmh;		/**< File Mapping handle */
+#endif
 	/** Failed to update the meta page. Probably an I/O error. */
 #define	MDB_FATAL_ERROR	0x80000000U
 	/** Some fields are initialized. */
@@ -1288,7 +1367,7 @@ struct MDB_env {
 	void		*me_pbuf;		/**< scratch area for DUPSORT put() */
 	MDB_txn		*me_txn;		/**< current write transaction */
 	MDB_txn		*me_txn0;		/**< prealloc'd write transaction */
-	size_t		me_mapsize;		/**< size of the data memory map */
+	mdb_size_t	me_mapsize;		/**< size of the data memory map */
 	off_t		me_size;		/**< current file size */
 	pgno_t		me_maxpg;		/**< me_mapsize / me_psize */
 	MDB_dbx		*me_dbxs;		/**< array of static DB info */
@@ -1321,6 +1400,13 @@ struct MDB_env {
 #else
 	mdb_mutex_t	me_rmutex;
 	mdb_mutex_t	me_wmutex;
+#endif
+#ifdef MDB_VL32
+	MDB_ID3L	me_rpages;	/**< like #mt_rpages, but global to env */
+	pthread_mutex_t	me_rpmutex;	/**< control access to #me_rpages */
+#define MDB_ERPAGE_SIZE	16384
+#define MDB_ERPAGE_MAX	(MDB_ERPAGE_SIZE-1)
+	unsigned int me_rpcheck;
 #endif
 	void		*me_userctx;	 /**< User-settable context */
 	MDB_assert_func *me_assert_func; /**< Callback for assertion failures */
@@ -1367,7 +1453,7 @@ enum {
 #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_get(MDB_cursor *mc, pgno_t pgno, MDB_page **mp, int *lvl);
 static int  mdb_page_search_root(MDB_cursor *mc,
 			    MDB_val *key, int modify);
 #define MDB_PS_MODIFY	1
@@ -1396,7 +1482,7 @@ static int  mdb_node_add(MDB_cursor *mc, indx_t indx,
 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, int fromleft);
-static int  mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data);
+static int  mdb_node_read(MDB_cursor *mc, 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);
 
@@ -1619,20 +1705,20 @@ mdb_page_list(MDB_page *mp)
 	case P_LEAF|P_LEAF2:        type = "LEAF2 page";		break;
 	case P_LEAF|P_LEAF2|P_SUBP: type = "LEAF2 sub-page";	break;
 	case P_OVERFLOW:
-		fprintf(stderr, "Overflow page %"Z"u pages %u%s\n",
+		fprintf(stderr, "Overflow page %"Y"u pages %u%s\n",
 			pgno, mp->mp_pages, state);
 		return;
 	case P_META:
-		fprintf(stderr, "Meta-page %"Z"u txnid %"Z"u\n",
+		fprintf(stderr, "Meta-page %"Y"u txnid %"Y"u\n",
 			pgno, ((MDB_meta *)METADATA(mp))->mm_txnid);
 		return;
 	default:
-		fprintf(stderr, "Bad page %"Z"u flags 0x%u\n", pgno, mp->mp_flags);
+		fprintf(stderr, "Bad page %"Y"u flags 0x%u\n", pgno, mp->mp_flags);
 		return;
 	}
 
 	nkeys = NUMKEYS(mp);
-	fprintf(stderr, "%s %"Z"u numkeys %d%s\n", type, pgno, nkeys, state);
+	fprintf(stderr, "%s %"Y"u numkeys %d%s\n", type, pgno, nkeys, state);
 
 	for (i=0; i<nkeys; i++) {
 		if (IS_LEAF2(mp)) {	/* LEAF2 pages have no mp_ptrs[] or node headers */
@@ -1647,7 +1733,7 @@ mdb_page_list(MDB_page *mp)
 		key.mv_data = node->mn_data;
 		nsize = NODESIZE + key.mv_size;
 		if (IS_BRANCH(mp)) {
-			fprintf(stderr, "key %d: page %"Z"u, %s\n", i, NODEPGNO(node),
+			fprintf(stderr, "key %d: page %"Y"u, %s\n", i, NODEPGNO(node),
 				DKEY(&key));
 			total += nsize;
 		} else {
@@ -1682,6 +1768,13 @@ 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
 
@@ -1736,7 +1829,7 @@ static void mdb_audit(MDB_txn *txn)
 		}
 	}
 	if (freecount + count + NUM_METAS != txn->mt_next_pgno) {
-		fprintf(stderr, "audit: %lu freecount: %lu count: %lu total: %lu next_pgno: %lu\n",
+		fprintf(stderr, "audit: %"Y"u freecount: %"Y"u count: %"Y"u total: %"Y"u next_pgno: %"Y"u\n",
 			txn->mt_txnid, freecount, count+NUM_METAS,
 			freecount+count+NUM_METAS, txn->mt_next_pgno);
 	}
@@ -1753,8 +1846,8 @@ int
 mdb_dcmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
 {
 	MDB_cmp_func *dcmp = txn->mt_dbxs[dbi].md_dcmp;
-#if UINT_MAX < SIZE_MAX
-	if (dcmp == mdb_cmp_int && a->mv_size == sizeof(size_t))
+#if UINT_MAX < SIZE_MAX || defined(MDB_VL32)
+	if (dcmp == mdb_cmp_int && a->mv_size == sizeof(mdb_size_t))
 		dcmp = mdb_cmp_clong;
 #endif
 	return dcmp(a, b);
@@ -1836,6 +1929,45 @@ mdb_dlist_free(MDB_txn *txn)
 	dl[0].mid = 0;
 }
 
+#ifdef MDB_VL32
+static void
+mdb_page_unref(MDB_txn *txn, MDB_page *mp)
+{
+	pgno_t pgno;
+	MDB_ID3L tl = txn->mt_rpages;
+	unsigned x, rem;
+	if (mp->mp_flags & (P_SUBP|P_DIRTY))
+		return;
+	rem = mp->mp_pgno & (MDB_RPAGE_CHUNK-1);
+	pgno = mp->mp_pgno ^ rem;
+	x = mdb_mid3l_search(tl, pgno);
+	if (x != tl[0].mid && tl[x+1].mid == mp->mp_pgno)
+		x++;
+	if (tl[x].mref)
+		tl[x].mref--;
+}
+#define MDB_PAGE_UNREF(txn, mp)	mdb_page_unref(txn, mp)
+
+static void
+mdb_cursor_unref(MDB_cursor *mc)
+{
+	int i;
+	if (!mc->mc_pg[0] || IS_SUBP(mc->mc_pg[0]))
+		return;
+	for (i=0; i<mc->mc_snum; i++)
+		mdb_page_unref(mc->mc_txn, mc->mc_pg[i]);
+	if (mc->mc_ovpg) {
+		mdb_page_unref(mc->mc_txn, mc->mc_ovpg);
+		mc->mc_ovpg = 0;
+	}
+	mc->mc_snum = mc->mc_top = 0;
+	mc->mc_pg[0] = NULL;
+	mc->mc_flags &= ~C_INITIALIZED;
+}
+#else
+#define MDB_PAGE_UNREF(txn, mp)
+#endif /* MDB_VL32 */
+
 /** Loosen or free a single page.
  * Saves single pages to a list for future reuse
  * in this same txn. It has been pulled from the freeDB
@@ -1877,7 +2009,7 @@ mdb_page_loose(MDB_cursor *mc, MDB_page *mp)
 		}
 	}
 	if (loose) {
-		DPRINTF(("loosen db %d page %"Z"u", DDBI(mc),
+		DPRINTF(("loosen db %d page %"Y"u", DDBI(mc),
 			mp->mp_pgno));
 		NEXT_LOOSE_PAGE(mp) = txn->mt_loose_pgs;
 		txn->mt_loose_pgs = mp;
@@ -1904,7 +2036,7 @@ mdb_pages_xkeep(MDB_cursor *mc, unsigned pflags, int all)
 {
 	enum { Mask = P_SUBP|P_DIRTY|P_LOOSE|P_KEEP };
 	MDB_txn *txn = mc->mc_txn;
-	MDB_cursor *m3;
+	MDB_cursor *m3, *m0 = mc;
 	MDB_xcursor *mx;
 	MDB_page *dp, *mp;
 	MDB_node *leaf;
@@ -1947,7 +2079,7 @@ mdb_pages_xkeep(MDB_cursor *mc, unsigned pflags, int all)
 				pgno_t pgno = txn->mt_dbs[i].md_root;
 				if (pgno == P_INVALID)
 					continue;
-				if ((rc = mdb_page_get(txn, pgno, &dp, &level)) != MDB_SUCCESS)
+				if ((rc = mdb_page_get(m0, pgno, &dp, &level)) != MDB_SUCCESS)
 					break;
 				if ((dp->mp_flags & Mask) == pflags && level <= 1)
 					dp->mp_flags ^= P_KEEP;
@@ -2136,6 +2268,8 @@ mdb_page_dirty(MDB_txn *txn, MDB_page *mp)
  * Do not modify the freedB, just merge freeDB records into me_pghead[]
  * and move me_pglast to say which records were consumed.  Only this
  * function can create me_pghead and move me_pglast/mt_next_pgno.
+ * When #MDB_DEVEL & 2, it is not affected by #mdb_freelist_save(): it
+ * then uses the transaction's original snapshot of the freeDB.
  * @param[in] mc cursor A cursor handle identifying the transaction and
  *	database for which we are allocating.
  * @param[in] num the number of pages to allocate.
@@ -2173,7 +2307,7 @@ mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp)
 		np = txn->mt_loose_pgs;
 		txn->mt_loose_pgs = NEXT_LOOSE_PAGE(np);
 		txn->mt_loose_count--;
-		DPRINTF(("db %d use loose page %"Z"u", DDBI(mc),
+		DPRINTF(("db %d use loose page %"Y"u", DDBI(mc),
 				np->mp_pgno));
 		*mp = np;
 		return MDB_SUCCESS;
@@ -2211,6 +2345,14 @@ mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp)
 			last = env->me_pglast;
 			oldest = env->me_pgoldest;
 			mdb_cursor_init(&m2, txn, FREE_DBI, NULL);
+#if (MDB_DEVEL) & 2	/* "& 2" so MDB_DEVEL=1 won't hide bugs breaking freeDB */
+			/* Use original snapshot. TODO: Should need less care in code
+			 * which modifies the database. Maybe we can delete some code?
+			 */
+			m2.mc_flags |= C_ORIG_RDONLY;
+			m2.mc_db = &env->me_metas[(txn->mt_txnid-1) & 1]->mm_dbs[FREE_DBI];
+			m2.mc_dbflag = (unsigned char *)""; /* probably unnecessary */
+#endif
 			if (last) {
 				op = MDB_SET_RANGE;
 				key.mv_data = &last; /* will look up last+1 */
@@ -2251,7 +2393,7 @@ mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp)
 		}
 		np = m2.mc_pg[m2.mc_top];
 		leaf = NODEPTR(np, m2.mc_ki[m2.mc_top]);
-		if ((rc = mdb_node_read(txn, leaf, &data)) != MDB_SUCCESS)
+		if ((rc = mdb_node_read(&m2, leaf, &data)) != MDB_SUCCESS)
 			return rc;
 
 		idl = (MDB_ID *) data.mv_data;
@@ -2268,10 +2410,10 @@ mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp)
 		}
 		env->me_pglast = last;
 #if (MDB_DEBUG) > 1
-		DPRINTF(("IDL read txn %"Z"u root %"Z"u num %u",
+		DPRINTF(("IDL read txn %"Y"u root %"Y"u num %u",
 			last, txn->mt_dbs[FREE_DBI].md_root, i));
 		for (j = i; j; j--)
-			DPRINTF(("IDL %"Z"u", idl[j]));
+			DPRINTF(("IDL %"Y"u", idl[j]));
 #endif
 		/* Merge in descending sorted order */
 		mdb_midl_xmerge(mop, idl);
@@ -2286,6 +2428,20 @@ mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp)
 			rc = MDB_MAP_FULL;
 			goto fail;
 	}
+#ifdef _WIN32
+	if (env->me_flags & MDB_WRITEMAP) {
+		void *p;
+		p = (MDB_page *)(env->me_map + env->me_psize * pgno);
+		p = VirtualAlloc(p, env->me_psize * num, MEM_COMMIT,
+			(env->me_flags & MDB_WRITEMAP) ? PAGE_READWRITE:
+			PAGE_READONLY);
+		if (!p) {
+			DPUTS("VirtualAlloc failed");
+			rc = ErrCode();
+			goto fail;
+		}
+	}
+#endif
 
 search_done:
 	if (env->me_flags & MDB_WRITEMAP) {
@@ -2427,7 +2583,7 @@ mdb_page_touch(MDB_cursor *mc)
 			(rc = mdb_page_alloc(mc, 1, &np)))
 			goto fail;
 		pgno = np->mp_pgno;
-		DPRINTF(("touched db %d page %"Z"u -> %"Z"u", DDBI(mc),
+		DPRINTF(("touched db %d page %"Y"u -> %"Y"u", DDBI(mc),
 			mp->mp_pgno, pgno));
 		mdb_cassert(mc, mp->mp_pgno != pgno);
 		mdb_midl_xappend(txn->mt_free_pgs, mp->mp_pgno);
@@ -2487,19 +2643,21 @@ 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);
 				}
 			}
 		}
 	}
+	MDB_PAGE_UNREF(mc->mc_txn, mp);
 	return 0;
 
 fail:
@@ -2508,7 +2666,7 @@ fail:
 }
 
 int
-mdb_env_sync(MDB_env *env, int force)
+mdb_env_sync0(MDB_env *env, int force, pgno_t numpgs)
 {
 	int rc = 0;
 	if (env->me_flags & MDB_RDONLY)
@@ -2517,7 +2675,7 @@ mdb_env_sync(MDB_env *env, int force)
 		if (env->me_flags & MDB_WRITEMAP) {
 			int flags = ((env->me_flags & MDB_MAPASYNC) && !force)
 				? MS_ASYNC : MS_SYNC;
-			if (MDB_MSYNC(env->me_map, env->me_mapsize, flags))
+			if (MDB_MSYNC(env->me_map, env->me_psize * numpgs, flags))
 				rc = ErrCode();
 #ifdef _WIN32
 			else if (flags == MS_SYNC && MDB_FDATASYNC(env->me_fd))
@@ -2537,6 +2695,13 @@ mdb_env_sync(MDB_env *env, int force)
 	return rc;
 }
 
+int
+mdb_env_sync(MDB_env *env, int force)
+{
+	MDB_meta *m = mdb_env_pick_meta(env);
+	return mdb_env_sync0(env, force, m->mm_last_pg+1);
+}
+
 /** Back up parent txn's cursors, then grab the originals for tracking */
 static int
 mdb_cursor_shadow(MDB_txn *src, MDB_txn *dst)
@@ -2781,6 +2946,9 @@ mdb_txn_renew0(MDB_txn *txn)
 
 	/* Moved to here to avoid a data race in read TXNs */
 	txn->mt_next_pgno = meta->mm_last_pg+1;
+#ifdef MDB_VL32
+	txn->mt_last_pgno = txn->mt_next_pgno - 1;
+#endif
 
 	txn->mt_flags = flags;
 
@@ -2816,7 +2984,7 @@ mdb_txn_renew(MDB_txn *txn)
 
 	rc = mdb_txn_renew0(txn);
 	if (rc == MDB_SUCCESS) {
-		DPRINTF(("renew txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
+		DPRINTF(("renew txn %"Y"u%c %p on mdbenv %p, root page %"Y"u",
 			txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
 			(void *)txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root));
 	}
@@ -2859,6 +3027,17 @@ mdb_txn_begin(MDB_env *env, MDB_txn *parent, unsigned int flags, MDB_txn **ret)
 		DPRINTF(("calloc: %s", strerror(errno)));
 		return ENOMEM;
 	}
+#ifdef MDB_VL32
+	if (!parent) {
+		txn->mt_rpages = malloc(MDB_TRPAGE_SIZE * sizeof(MDB_ID3));
+		if (!txn->mt_rpages) {
+			free(txn);
+			return ENOMEM;
+		}
+		txn->mt_rpages[0].mid = 0;
+		txn->mt_rpcheck = MDB_TRPAGE_SIZE/2;
+	}
+#endif
 	txn->mt_dbxs = env->me_dbxs;	/* static */
 	txn->mt_dbs = (MDB_db *) ((char *)txn + tsize);
 	txn->mt_dbflags = (unsigned char *)txn + size - env->me_maxdbs;
@@ -2886,6 +3065,9 @@ mdb_txn_begin(MDB_env *env, MDB_txn *parent, unsigned int flags, MDB_txn **ret)
 		parent->mt_child = txn;
 		txn->mt_parent = parent;
 		txn->mt_numdbs = parent->mt_numdbs;
+#ifdef MDB_VL32
+		txn->mt_rpages = parent->mt_rpages;
+#endif
 		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++)
@@ -2911,12 +3093,16 @@ renew:
 		rc = mdb_txn_renew0(txn);
 	}
 	if (rc) {
-		if (txn != env->me_txn0)
+		if (txn != env->me_txn0) {
+#ifdef MDB_VL32
+			free(txn->mt_rpages);
+#endif
 			free(txn);
+		}
 	} else {
 		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",
+		DPRINTF(("begin txn %"Y"u%c %p on mdbenv %p, root page %"Y"u",
 			txn->mt_txnid, (flags & MDB_RDONLY) ? 'r' : 'w',
 			(void *) txn, (void *) env, txn->mt_dbs[MAIN_DBI].md_root));
 	}
@@ -2931,7 +3117,7 @@ mdb_txn_env(MDB_txn *txn)
 	return txn->mt_env;
 }
 
-size_t
+mdb_size_t
 mdb_txn_id(MDB_txn *txn)
 {
     if(!txn) return 0;
@@ -2983,7 +3169,7 @@ mdb_txn_end(MDB_txn *txn, unsigned mode)
 	/* 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",
+	DPRINTF(("%s txn %"Y"u%c %p on mdbenv %p, root page %"Y"u",
 		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));
@@ -3037,7 +3223,31 @@ mdb_txn_end(MDB_txn *txn, unsigned mode)
 
 		mdb_midl_free(pghead);
 	}
-
+#ifdef MDB_VL32
+	if (!txn->mt_parent) {
+		MDB_ID3L el = env->me_rpages, tl = txn->mt_rpages;
+		unsigned i, x, n = tl[0].mid;
+		pthread_mutex_lock(&env->me_rpmutex);
+		for (i = 1; i <= n; i++) {
+			if (tl[i].mid & (MDB_RPAGE_CHUNK-1)) {
+				/* tmp overflow pages that we didn't share in env */
+				munmap(tl[i].mptr, tl[i].mcnt * env->me_psize);
+			} else {
+				x = mdb_mid3l_search(el, tl[i].mid);
+				if (tl[i].mptr == el[x].mptr) {
+					el[x].mref--;
+				} else {
+					/* another tmp overflow page */
+					munmap(tl[i].mptr, tl[i].mcnt * env->me_psize);
+				}
+			}
+		}
+		pthread_mutex_unlock(&env->me_rpmutex);
+		tl[0].mid = 0;
+		if (mode & MDB_END_FREE)
+			free(tl);
+	}
+#endif
 	if (mode & MDB_END_FREE)
 		free(txn);
 }
@@ -3069,6 +3279,9 @@ mdb_txn_abort(MDB_txn *txn)
 
 /** Save the freelist as of this transaction to the freeDB.
  * This changes the freelist. Keep trying until it stabilizes.
+ *
+ * When (MDB_DEVEL) & 2, the changes do not affect #mdb_page_alloc(),
+ * it then uses the transaction's original snapshot of the freeDB.
  */
 static int
 mdb_freelist_save(MDB_txn *txn)
@@ -3157,10 +3370,10 @@ mdb_freelist_save(MDB_txn *txn)
 #if (MDB_DEBUG) > 1
 			{
 				unsigned int i = free_pgs[0];
-				DPRINTF(("IDL write txn %"Z"u root %"Z"u num %u",
+				DPRINTF(("IDL write txn %"Y"u root %"Y"u num %u",
 					txn->mt_txnid, txn->mt_dbs[FREE_DBI].md_root, i));
 				for (; i; i--)
-					DPRINTF(("IDL %"Z"u", free_pgs[i]));
+					DPRINTF(("IDL %"Y"u", free_pgs[i]));
 			}
 #endif
 			continue;
@@ -3271,15 +3484,16 @@ mdb_page_flush(MDB_txn *txn, int keep)
 	MDB_ID2L	dl = txn->mt_u.dirty_list;
 	unsigned	psize = env->me_psize, j;
 	int			i, pagecount = dl[0].mid, rc;
-	size_t		size = 0, pos = 0;
+	size_t		size = 0;
+	off_t		pos = 0;
 	pgno_t		pgno = 0;
 	MDB_page	*dp = NULL;
 #ifdef _WIN32
 	OVERLAPPED	ov;
 #else
 	struct iovec iov[MDB_COMMIT_PAGES];
-	ssize_t		wpos = 0, wsize = 0, wres;
-	size_t		next_pos = 1; /* impossible pos, so pos != next_pos */
+	ssize_t		wsize = 0, wres;
+	off_t		wpos = 0, next_pos = 1; /* impossible pos, so pos != next_pos */
 	int			n = 0;
 #endif
 
@@ -3378,7 +3592,7 @@ retry_seek:
 			wpos = pos;
 			wsize = 0;
 		}
-		DPRINTF(("committing page %"Z"u", pgno));
+		DPRINTF(("committing page %"Y"u", pgno));
 		next_pos = pos + size;
 		iov[n].iov_len = size;
 		iov[n].iov_base = (char *)dp;
@@ -3386,6 +3600,10 @@ retry_seek:
 		n++;
 #endif	/* _WIN32 */
 	}
+#ifdef MDB_VL32
+	if (pgno > txn->mt_last_pgno)
+		txn->mt_last_pgno = pgno;
+#endif
 
 	/* MIPS has cache coherency issues, this is a no-op everywhere else
 	 * Note: for any size >= on-chip cache size, entire on-chip cache is
@@ -3500,6 +3718,25 @@ 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 */
@@ -3568,7 +3805,7 @@ mdb_txn_commit(MDB_txn *txn)
 		!(txn->mt_flags & (MDB_TXN_DIRTY|MDB_TXN_SPILLS)))
 		goto done;
 
-	DPRINTF(("committing txn %"Z"u %p on mdbenv %p, root page %"Z"u",
+	DPRINTF(("committing txn %"Y"u %p on mdbenv %p, root page %"Y"u",
 	    txn->mt_txnid, (void*)txn, (void*)env, txn->mt_dbs[MAIN_DBI].md_root));
 
 	/* Update DB root pointers */
@@ -3609,7 +3846,7 @@ mdb_txn_commit(MDB_txn *txn)
 	if ((rc = mdb_page_flush(txn, 0)))
 		goto fail;
 	if (!F_ISSET(txn->mt_flags, MDB_TXN_NOSYNC) &&
-		(rc = mdb_env_sync(env, 0)))
+		(rc = mdb_env_sync0(env, 0, txn->mt_next_pgno)))
 		goto fail;
 	if ((rc = mdb_env_write_meta(txn)))
 		goto fail;
@@ -3666,7 +3903,7 @@ mdb_env_read_header(MDB_env *env, MDB_meta *meta)
 		p = (MDB_page *)&pbuf;
 
 		if (!F_ISSET(p->mp_flags, P_META)) {
-			DPRINTF(("page %"Z"u not a meta page", p->mp_pgno));
+			DPRINTF(("page %"Y"u not a meta page", p->mp_pgno));
 			return MDB_INVALID;
 		}
 
@@ -3766,7 +4003,7 @@ mdb_env_write_meta(MDB_txn *txn)
 	MDB_env *env;
 	MDB_meta	meta, metab, *mp;
 	unsigned flags;
-	size_t mapsize;
+	mdb_size_t mapsize;
 	off_t off;
 	int rc, len, toggle;
 	char *ptr;
@@ -3778,7 +4015,7 @@ mdb_env_write_meta(MDB_txn *txn)
 #endif
 
 	toggle = txn->mt_txnid & 1;
-	DPRINTF(("writing meta page %d for root page %"Z"u",
+	DPRINTF(("writing meta page %d for root page %"Y"u",
 		toggle, txn->mt_dbs[MAIN_DBI].md_root));
 
 	env = txn->mt_env;
@@ -3929,42 +4166,51 @@ mdb_env_map(MDB_env *env, void *addr)
 	unsigned int flags = env->me_flags;
 #ifdef _WIN32
 	int rc;
+	int access = SECTION_MAP_READ;
 	HANDLE mh;
-	LONG sizelo, sizehi;
-	size_t msize;
+	void *map;
+	SIZE_T msize;
+	ULONG pageprot = PAGE_READONLY, secprot, alloctype;
 
+	if (flags & MDB_WRITEMAP) {
+		access |= SECTION_MAP_WRITE;
+		pageprot = PAGE_READWRITE;
+	}
 	if (flags & MDB_RDONLY) {
-		/* Don't set explicit map size, use whatever exists */
+		secprot = PAGE_READONLY;
 		msize = 0;
-		sizelo = 0;
-		sizehi = 0;
+		alloctype = 0;
 	} else {
+		secprot = PAGE_READWRITE;
 		msize = env->me_mapsize;
-		sizelo = msize & 0xffffffff;
-		sizehi = msize >> 16 >> 16; /* only needed on Win64 */
-
-		/* Windows won't create mappings for zero length files.
-		 * and won't map more than the file size.
-		 * Just set the maxsize right now.
-		 */
-		if (SetFilePointer(env->me_fd, sizelo, &sizehi, 0) != (DWORD)sizelo
-			|| !SetEndOfFile(env->me_fd)
-			|| SetFilePointer(env->me_fd, 0, NULL, 0) != 0)
-			return ErrCode();
+		alloctype = MEM_RESERVE;
 	}
 
-	mh = CreateFileMapping(env->me_fd, NULL, flags & MDB_WRITEMAP ?
-		PAGE_READWRITE : PAGE_READONLY,
-		sizehi, sizelo, NULL);
-	if (!mh)
-		return ErrCode();
-	env->me_map = MapViewOfFileEx(mh, flags & MDB_WRITEMAP ?
-		FILE_MAP_WRITE : FILE_MAP_READ,
-		0, 0, msize, addr);
-	rc = env->me_map ? 0 : ErrCode();
-	CloseHandle(mh);
+	rc = NtCreateSection(&mh, access, NULL, NULL, secprot, SEC_RESERVE, env->me_fd);
 	if (rc)
 		return rc;
+	map = addr;
+#ifdef MDB_VL32
+	msize = NUM_METAS * env->me_psize;
+#endif
+	rc = NtMapViewOfSection(mh, GetCurrentProcess(), &map, 0, 0, NULL, &msize, ViewUnmap, alloctype, pageprot);
+#ifdef MDB_VL32
+	env->me_fmh = mh;
+#else
+	NtClose(mh);
+#endif
+	if (rc)
+		return rc;
+	env->me_map = map;
+#else
+#ifdef MDB_VL32
+	(void) flags;
+	env->me_map = mmap(addr, NUM_METAS * env->me_psize, PROT_READ, MAP_SHARED,
+		env->me_fd, 0);
+	if (env->me_map == MAP_FAILED) {
+		env->me_map = NULL;
+		return ErrCode();
+	}
 #else
 	int prot = PROT_READ;
 	if (flags & MDB_WRITEMAP) {
@@ -3998,6 +4244,7 @@ mdb_env_map(MDB_env *env, void *addr)
 	 */
 	if (addr && env->me_map != addr)
 		return EBUSY;	/* TODO: Make a new MDB_* error code? */
+#endif
 
 	p = (MDB_page *)env->me_map;
 	env->me_metas[0] = METADATA(p);
@@ -4007,15 +4254,17 @@ mdb_env_map(MDB_env *env, void *addr)
 }
 
 int ESECT
-mdb_env_set_mapsize(MDB_env *env, size_t size)
+mdb_env_set_mapsize(MDB_env *env, mdb_size_t size)
 {
 	/* If env is already open, caller is responsible for making
 	 * sure there are no active txns.
 	 */
 	if (env->me_map) {
-		int rc;
 		MDB_meta *meta;
+#ifndef MDB_VL32
 		void *old;
+		int rc;
+#endif
 		if (env->me_txn)
 			return EINVAL;
 		meta = mdb_env_pick_meta(env);
@@ -4027,12 +4276,17 @@ mdb_env_set_mapsize(MDB_env *env, size_t size)
 			if (size < minsize)
 				size = minsize;
 		}
+#ifndef MDB_VL32
+		/* For MDB_VL32 this bit is a noop since we dynamically remap
+		 * chunks of the DB anyway.
+		 */
 		munmap(env->me_map, env->me_mapsize);
 		env->me_mapsize = size;
 		old = (env->me_flags & MDB_FIXEDMAP) ? env->me_map : NULL;
 		rc = mdb_env_map(env, old);
 		if (rc)
 			return rc;
+#endif /* !MDB_VL32 */
 	}
 	env->me_mapsize = size;
 	if (env->me_psize)
@@ -4068,7 +4322,7 @@ mdb_env_get_maxreaders(MDB_env *env, unsigned int *readers)
 }
 
 static int ESECT
-mdb_fsize(HANDLE fd, size_t *size)
+mdb_fsize(HANDLE fd, mdb_size_t *size)
 {
 #ifdef _WIN32
 	LARGE_INTEGER fsize;
@@ -4182,7 +4436,7 @@ mdb_env_open2(MDB_env *env)
 		/* Make sure mapsize >= committed data size.  Even when using
 		 * mm_mapsize, which could be broken in old files (ITS#7789).
 		 */
-		size_t minsize = (meta.mm_last_pg + 1) * meta.mm_psize;
+		mdb_size_t minsize = (meta.mm_last_pg + 1) * meta.mm_psize;
 		if (env->me_mapsize < minsize)
 			env->me_mapsize = minsize;
 	}
@@ -4201,6 +4455,18 @@ mdb_env_open2(MDB_env *env)
 			return rc;
 		newenv = 0;
 	}
+#ifdef _WIN32
+	/* For FIXEDMAP, make sure the file is non-empty before we attempt to map it */
+	if (newenv) {
+		char dummy = 0;
+		DWORD len;
+		rc = WriteFile(env->me_fd, &dummy, 1, &len, NULL);
+		if (!rc) {
+			rc = ErrCode();
+			return rc;
+		}
+	}
+#endif
 
 	rc = mdb_env_map(env, (flags & MDB_FIXEDMAP) ? meta.mm_address : NULL);
 	if (rc)
@@ -4232,11 +4498,11 @@ mdb_env_open2(MDB_env *env)
 			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));
-		DPRINTF(("leaf pages: %"Z"u",     db->md_leaf_pages));
-		DPRINTF(("overflow pages: %"Z"u", db->md_overflow_pages));
-		DPRINTF(("root: %"Z"u",           db->md_root));
+		DPRINTF(("entries: %"Y"u",        db->md_entries));
+		DPRINTF(("branch pages: %"Y"u",   db->md_branch_pages));
+		DPRINTF(("leaf pages: %"Y"u",     db->md_leaf_pages));
+		DPRINTF(("overflow pages: %"Y"u", db->md_overflow_pages));
+		DPRINTF(("root: %"Y"u",           db->md_root));
 	}
 #endif
 
@@ -4521,7 +4787,9 @@ mdb_env_setup_locks(MDB_env *env, char *lpath, int mode, int *excl)
 
 #ifdef _WIN32
 	wchar_t *wlpath;
-	utf8_to_utf16(lpath, -1, &wlpath, NULL);
+	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);
@@ -4743,6 +5011,13 @@ mdb_env_setup_locks(MDB_env *env, char *lpath, int mode, int *excl)
 	env->me_wmutex->semnum = 1;
 	env->me_rmutex->locked = &env->me_txns->mti_rlocked;
 	env->me_wmutex->locked = &env->me_txns->mti_wlocked;
+#endif
+#ifdef MDB_VL32
+#ifdef _WIN32
+	env->me_rpmutex = CreateMutex(NULL, FALSE, NULL);
+#else
+	pthread_mutex_init(&env->me_rpmutex, NULL);
+#endif
 #endif
 
 	return MDB_SUCCESS;
@@ -4783,6 +5058,17 @@ mdb_env_open(MDB_env *env, const char *path, unsigned int flags, mdb_mode_t mode
 	if (env->me_fd!=INVALID_HANDLE_VALUE || (flags & ~(CHANGEABLE|CHANGELESS)))
 		return EINVAL;
 
+#ifdef MDB_VL32
+	if (flags & MDB_WRITEMAP) {
+		/* silently ignore WRITEMAP in 32 bit mode */
+		flags ^= MDB_WRITEMAP;
+	}
+	if (flags & MDB_FIXEDMAP) {
+		/* cannot support FIXEDMAP */
+		return EINVAL;
+	}
+#endif
+
 	len = strlen(path);
 	if (flags & MDB_NOSUBDIR) {
 		rc = len + sizeof(LOCKSUFF) + len + 1;
@@ -4812,6 +5098,17 @@ mdb_env_open(MDB_env *env, const char *path, unsigned int flags, mdb_mode_t mode
 			  (env->me_dirty_list = calloc(MDB_IDL_UM_SIZE, sizeof(MDB_ID2)))))
 			rc = ENOMEM;
 	}
+#ifdef MDB_VL32
+	if (!rc) {
+		env->me_rpages = malloc(MDB_ERPAGE_SIZE * sizeof(MDB_ID3));
+		if (!env->me_rpages) {
+			rc = ENOMEM;
+			goto leave;
+		}
+		env->me_rpages[0].mid = 0;
+		env->me_rpcheck = MDB_ERPAGE_SIZE/2;
+	}
+#endif
 	env->me_flags = flags |= MDB_ENV_ACTIVE;
 	if (rc)
 		goto leave;
@@ -4842,7 +5139,9 @@ mdb_env_open(MDB_env *env, const char *path, unsigned int flags, mdb_mode_t mode
 		len = OPEN_ALWAYS;
 	}
 	mode = FILE_ATTRIBUTE_NORMAL;
-	utf8_to_utf16(dpath, -1, &wpath, NULL);
+	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);
@@ -4874,7 +5173,9 @@ mdb_env_open(MDB_env *env, const char *path, unsigned int flags, mdb_mode_t mode
 			 */
 #ifdef _WIN32
 			len = OPEN_EXISTING;
-			utf8_to_utf16(dpath, -1, &wpath, NULL);
+			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);
@@ -4906,6 +5207,16 @@ mdb_env_open(MDB_env *env, const char *path, unsigned int flags, mdb_mode_t mode
 				txn->mt_dbiseqs = (unsigned int *)(txn->mt_cursors + env->me_maxdbs);
 				txn->mt_dbflags = (unsigned char *)(txn->mt_dbiseqs + env->me_maxdbs);
 				txn->mt_env = env;
+#ifdef MDB_VL32
+				txn->mt_rpages = malloc(MDB_TRPAGE_SIZE * sizeof(MDB_ID3));
+				if (!txn->mt_rpages) {
+					free(txn);
+					rc = ENOMEM;
+					goto leave;
+				}
+				txn->mt_rpages[0].mid = 0;
+				txn->mt_rpcheck = MDB_TRPAGE_SIZE/2;
+#endif
 				txn->mt_dbxs = env->me_dbxs;
 				txn->mt_flags = MDB_TXN_FINISHED;
 				env->me_txn0 = txn;
@@ -4945,6 +5256,13 @@ mdb_env_close0(MDB_env *env, int excl)
 	free(env->me_path);
 	free(env->me_dirty_list);
 	free(env->me_txn0);
+#ifdef MDB_VL32
+	{ unsigned int x;
+		for (x=1; x<=env->me_rpages[0].mid; x++)
+		munmap(env->me_rpages[x].mptr, env->me_rpages[x].mcnt * env->me_psize);
+	}
+	free(env->me_rpages);
+#endif
 	mdb_midl_free(env->me_free_pgs);
 
 	if (env->me_flags & MDB_ENV_TXKEY) {
@@ -4961,7 +5279,11 @@ mdb_env_close0(MDB_env *env, int excl)
 	}
 
 	if (env->me_map) {
+#ifdef MDB_VL32
+		munmap(env->me_map, NUM_METAS*env->me_psize);
+#else
 		munmap(env->me_map, env->me_mapsize);
+#endif
 	}
 	if (env->me_mfd != env->me_fd && env->me_mfd != INVALID_HANDLE_VALUE)
 		(void) close(env->me_mfd);
@@ -5026,6 +5348,14 @@ mdb_env_close0(MDB_env *env, int excl)
 #endif
 		(void) close(env->me_lfd);
 	}
+#ifdef MDB_VL32
+#ifdef _WIN32
+	if (env->me_fmh) CloseHandle(env->me_fmh);
+	if (env->me_rpmutex) CloseHandle(env->me_rpmutex);
+#else
+	pthread_mutex_destroy(&env->me_rpmutex);
+#endif
+#endif
 
 	env->me_flags &= ~(MDB_ENV_ACTIVE|MDB_ENV_TXKEY);
 }
@@ -5049,18 +5379,18 @@ mdb_env_close(MDB_env *env)
 	free(env);
 }
 
-/** Compare two items pointing at aligned size_t's */
+/** Compare two items pointing at aligned mdb_size_t's */
 static int
 mdb_cmp_long(const MDB_val *a, const MDB_val *b)
 {
-	return (*(size_t *)a->mv_data < *(size_t *)b->mv_data) ? -1 :
-		*(size_t *)a->mv_data > *(size_t *)b->mv_data;
+	return (*(mdb_size_t *)a->mv_data < *(mdb_size_t *)b->mv_data) ? -1 :
+		*(mdb_size_t *)a->mv_data > *(mdb_size_t *)b->mv_data;
 }
 
 /** Compare two items pointing at aligned unsigned int's.
  *
  *	This is also set as #MDB_INTEGERDUP|#MDB_DUPFIXED's #MDB_dbx.%md_dcmp,
- *	but #mdb_cmp_clong() is called instead if the data type is size_t.
+ *	but #mdb_cmp_clong() is called instead if the data type is mdb_size_t.
  */
 static int
 mdb_cmp_int(const MDB_val *a, const MDB_val *b)
@@ -5165,7 +5495,7 @@ mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp)
 
 	nkeys = NUMKEYS(mp);
 
-	DPRINTF(("searching %u keys in %s %spage %"Z"u",
+	DPRINTF(("searching %u keys in %s %spage %"Y"u",
 	    nkeys, IS_LEAF(mp) ? "leaf" : "branch", IS_SUBP(mp) ? "sub-" : "",
 	    mdb_dbg_pgno(mp)));
 
@@ -5177,7 +5507,7 @@ mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp)
 	 * alignment is guaranteed. Use faster mdb_cmp_int.
 	 */
 	if (cmp == mdb_cmp_cint && IS_BRANCH(mp)) {
-		if (NODEPTR(mp, 1)->mn_ksize == sizeof(size_t))
+		if (NODEPTR(mp, 1)->mn_ksize == sizeof(mdb_size_t))
 			cmp = mdb_cmp_long;
 		else
 			cmp = mdb_cmp_int;
@@ -5213,7 +5543,7 @@ mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp)
 				DPRINTF(("found leaf index %u [%s], rc = %i",
 				    i, DKEY(&nodekey), rc));
 			else
-				DPRINTF(("found branch index %u [%s -> %"Z"u], rc = %i",
+				DPRINTF(("found branch index %u [%s -> %"Y"u], rc = %i",
 				    i, DKEY(&nodekey), NODEPGNO(node), rc));
 #endif
 			if (rc == 0)
@@ -5261,12 +5591,15 @@ static void
 mdb_cursor_pop(MDB_cursor *mc)
 {
 	if (mc->mc_snum) {
-		DPRINTF(("popping page %"Z"u off db %d cursor %p",
+		DPRINTF(("popping page %"Y"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--;
+		} else {
+			mc->mc_flags &= ~C_INITIALIZED;
+		}
 	}
 }
 
@@ -5274,7 +5607,7 @@ mdb_cursor_pop(MDB_cursor *mc)
 static int
 mdb_cursor_push(MDB_cursor *mc, MDB_page *mp)
 {
-	DPRINTF(("pushing page %"Z"u on db %d cursor %p", mp->mp_pgno,
+	DPRINTF(("pushing page %"Y"u on db %d cursor %p", mp->mp_pgno,
 		DDBI(mc), (void *) mc));
 
 	if (mc->mc_snum >= CURSOR_STACK) {
@@ -5289,21 +5622,332 @@ mdb_cursor_push(MDB_cursor *mc, MDB_page *mp)
 	return MDB_SUCCESS;
 }
 
-/** Find the address of the page corresponding to a given page number.
+#ifdef MDB_VL32
+/** Map a read-only page.
+ * There are two levels of tracking in use, a per-txn list and a per-env list.
+ * ref'ing and unref'ing the per-txn list is faster since it requires no
+ * locking. Pages are cached in the per-env list for global reuse, and a lock
+ * is required. Pages are not immediately unmapped when their refcnt goes to
+ * zero; they hang around in case they will be reused again soon.
+ *
+ * When the per-txn list gets full, all pages with refcnt=0 are purged from the
+ * list and their refcnts in the per-env list are decremented.
+ *
+ * When the per-env list gets full, all pages with refcnt=0 are purged from the
+ * list and their pages are unmapped.
+ *
+ * @note "full" means the list has reached its respective rpcheck threshold.
+ * This threshold slowly raises if no pages could be purged on a given check,
+ * and returns to its original value when enough pages were purged.
+ *
+ * If purging doesn't free any slots, filling the per-txn list will return
+ * MDB_TXN_FULL, and filling the per-env list returns MDB_MAP_FULL.
+ *
+ * Reference tracking in a txn is imperfect, pages can linger with non-zero
+ * refcnt even without active references. It was deemed to be too invasive
+ * to add unrefs in every required location. However, all pages are unref'd
+ * at the end of the transaction. This guarantees that no stale references
+ * linger in the per-env list.
+ *
+ * Usually we map chunks of 16 pages at a time, but if an overflow page begins
+ * at the tail of the chunk we extend the chunk to include the entire overflow
+ * page. Unfortunately, pages can be turned into overflow pages after their
+ * chunk was already mapped. In that case we must remap the chunk if the
+ * overflow page is referenced. If the chunk's refcnt is 0 we can just remap
+ * it, otherwise we temporarily map a new chunk just for the overflow page.
+ *
+ * @note this chunk handling means we cannot guarantee that a data item
+ * returned from the DB will stay alive for the duration of the transaction:
+ *   We unref pages as soon as a cursor moves away from the page
+ *   A subsequent op may cause a purge, which may unmap any unref'd chunks
+ * The caller must copy the data if it must be used later in the same txn.
+ *
+ * Also - our reference counting revolves around cursors, but overflow pages
+ * aren't pointed to by a cursor's page stack. We have to remember them
+ * explicitly, in the added mc_ovpg field. A single cursor can only hold a
+ * reference to one overflow page at a time.
+ *
  * @param[in] txn the transaction for this access.
  * @param[in] pgno the page number for the page to retrieve.
  * @param[out] ret address of a pointer where the page's address will be stored.
+ * @return 0 on success, non-zero on failure.
+ */
+static int
+mdb_rpage_get(MDB_txn *txn, pgno_t pg0, MDB_page **ret)
+{
+	MDB_env *env = txn->mt_env;
+	MDB_page *p;
+	MDB_ID3L tl = txn->mt_rpages;
+	MDB_ID3L el = env->me_rpages;
+	MDB_ID3 id3;
+	unsigned x, rem;
+	pgno_t pgno;
+	int rc, retries = 1;
+#ifdef _WIN32
+	LARGE_INTEGER off;
+	SIZE_T len;
+#define SET_OFF(off,val)	off.QuadPart = val
+#define MAP(rc,env,addr,len,off)	\
+	addr = NULL; \
+	rc = NtMapViewOfSection(env->me_fmh, GetCurrentProcess(), &addr, 0, \
+		len, &off, &len, ViewUnmap, (env->me_flags & MDB_RDONLY) ? 0 : MEM_RESERVE, PAGE_READONLY)
+#else
+	off_t off;
+	size_t len;
+#define SET_OFF(off,val)	off = val
+#define MAP(rc,env,addr,len,off)	\
+	addr = mmap(NULL, len, PROT_READ, MAP_SHARED, env->me_fd, off); \
+	rc = (addr == MAP_FAILED) ? errno : 0
+#endif
+
+	/* remember the offset of the actual page number, so we can
+	 * return the correct pointer at the end.
+	 */
+	rem = pg0 & (MDB_RPAGE_CHUNK-1);
+	pgno = pg0 ^ rem;
+
+	id3.mid = 0;
+	x = mdb_mid3l_search(tl, pgno);
+	if (x <= tl[0].mid && tl[x].mid == pgno) {
+		if (x != tl[0].mid && tl[x+1].mid == pg0)
+			x++;
+		/* check for overflow size */
+		p = (MDB_page *)((char *)tl[x].mptr + rem * env->me_psize);
+		if (IS_OVERFLOW(p) && p->mp_pages + rem > tl[x].mcnt) {
+			id3.mcnt = p->mp_pages + rem;
+			len = id3.mcnt * env->me_psize;
+			SET_OFF(off, pgno * env->me_psize);
+			MAP(rc, env, id3.mptr, len, off);
+			if (rc)
+				return rc;
+			/* check for local-only page */
+			if (rem) {
+				mdb_tassert(txn, tl[x].mid != pg0);
+				/* hope there's room to insert this locally.
+				 * setting mid here tells later code to just insert
+				 * this id3 instead of searching for a match.
+				 */
+				id3.mid = pg0;
+				goto notlocal;
+			} else {
+				/* ignore the mapping we got from env, use new one */
+				tl[x].mptr = id3.mptr;
+				tl[x].mcnt = id3.mcnt;
+				/* if no active ref, see if we can replace in env */
+				if (!tl[x].mref) {
+					unsigned i;
+					pthread_mutex_lock(&env->me_rpmutex);
+					i = mdb_mid3l_search(el, tl[x].mid);
+					if (el[i].mref == 1) {
+						/* just us, replace it */
+						munmap(el[i].mptr, el[i].mcnt * env->me_psize);
+						el[i].mptr = tl[x].mptr;
+						el[i].mcnt = tl[x].mcnt;
+					} else {
+						/* there are others, remove ourself */
+						el[i].mref--;
+					}
+					pthread_mutex_unlock(&env->me_rpmutex);
+				}
+			}
+		}
+		id3.mptr = tl[x].mptr;
+		id3.mcnt = tl[x].mcnt;
+		tl[x].mref++;
+		goto ok;
+	}
+
+notlocal:
+	if (tl[0].mid >= MDB_TRPAGE_MAX - txn->mt_rpcheck) {
+		unsigned i, y;
+		/* purge unref'd pages from our list and unref in env */
+		pthread_mutex_lock(&env->me_rpmutex);
+retry:
+		y = 0;
+		for (i=1; i<tl[0].mid; i++) {
+			if (!tl[i].mref) {
+				if (!y) y = i;
+				/* tmp overflow pages don't go to env */
+				if (tl[i].mid & (MDB_RPAGE_CHUNK-1)) {
+					munmap(tl[i].mptr, tl[i].mcnt * env->me_psize);
+					continue;
+				}
+				x = mdb_mid3l_search(el, tl[i].mid);
+				el[x].mref--;
+			}
+		}
+		pthread_mutex_unlock(&env->me_rpmutex);
+		if (!y) {
+			/* we didn't find any unref'd chunks.
+			 * if we're out of room, fail.
+			 */
+			if (tl[0].mid >= MDB_TRPAGE_MAX)
+				return MDB_TXN_FULL;
+			/* otherwise, raise threshold for next time around
+			 * and let this go.
+			 */
+			txn->mt_rpcheck /= 2;
+		} else {
+			/* we found some unused; consolidate the list */
+			for (i=y+1; i<= tl[0].mid; i++)
+				if (tl[i].mref)
+					tl[y++] = tl[i];
+			tl[0].mid = y-1;
+			/* decrease the check threshold toward its original value */
+			if (!txn->mt_rpcheck)
+				txn->mt_rpcheck = 1;
+			while (txn->mt_rpcheck < tl[0].mid && txn->mt_rpcheck < MDB_TRPAGE_SIZE/2)
+				txn->mt_rpcheck *= 2;
+		}
+	}
+	if (tl[0].mid < MDB_TRPAGE_SIZE) {
+		id3.mref = 1;
+		if (id3.mid)
+			goto found;
+		/* don't map past last written page in read-only envs */
+		if ((env->me_flags & MDB_RDONLY) && pgno + MDB_RPAGE_CHUNK-1 > txn->mt_last_pgno)
+			id3.mcnt = txn->mt_last_pgno + 1 - pgno;
+		else
+			id3.mcnt = MDB_RPAGE_CHUNK;
+		len = id3.mcnt * env->me_psize;
+		id3.mid = pgno;
+
+		/* search for page in env */
+		pthread_mutex_lock(&env->me_rpmutex);
+		x = mdb_mid3l_search(el, pgno);
+		if (x <= el[0].mid && el[x].mid == pgno) {
+			id3.mptr = el[x].mptr;
+			id3.mcnt = el[x].mcnt;
+			/* check for overflow size */
+			p = (MDB_page *)((char *)id3.mptr + rem * env->me_psize);
+			if (IS_OVERFLOW(p) && p->mp_pages + rem > id3.mcnt) {
+				id3.mcnt = p->mp_pages + rem;
+				len = id3.mcnt * env->me_psize;
+				SET_OFF(off, pgno * env->me_psize);
+				MAP(rc, env, id3.mptr, len, off);
+				if (rc)
+					goto fail;
+				if (!el[x].mref) {
+					munmap(el[x].mptr, el[x].mcnt);
+					el[x].mptr = id3.mptr;
+					el[x].mcnt = id3.mcnt;
+				} else {
+					id3.mid = pg0;
+					pthread_mutex_unlock(&env->me_rpmutex);
+					goto found;
+				}
+			}
+			el[x].mref++;
+			pthread_mutex_unlock(&env->me_rpmutex);
+			goto found;
+		}
+		if (el[0].mid >= MDB_ERPAGE_MAX - env->me_rpcheck) {
+			/* purge unref'd pages */
+			unsigned i, y = 0;
+			for (i=1; i<el[0].mid; i++) {
+				if (!el[i].mref) {
+					if (!y) y = i;
+					munmap(el[i].mptr, env->me_psize * el[i].mcnt);
+				}
+			}
+			if (!y) {
+				if (retries) {
+					/* see if we can unref some local pages */
+					retries--;
+					id3.mid = 0;
+					goto retry;
+				}
+				if (el[0].mid >= MDB_ERPAGE_MAX) {
+					pthread_mutex_unlock(&env->me_rpmutex);
+					return MDB_MAP_FULL;
+				}
+				env->me_rpcheck /= 2;
+			} else {
+				for (i=y+1; i<= el[0].mid; i++)
+					if (el[i].mref)
+						el[y++] = el[i];
+				el[0].mid = y-1;
+				if (!env->me_rpcheck)
+					env->me_rpcheck = 1;
+				while (env->me_rpcheck < el[0].mid && env->me_rpcheck < MDB_ERPAGE_SIZE/2)
+					env->me_rpcheck *= 2;
+			}
+		}
+		SET_OFF(off, pgno * env->me_psize);
+		MAP(rc, env, id3.mptr, len, off);
+		if (rc) {
+fail:
+			pthread_mutex_unlock(&env->me_rpmutex);
+			return rc;
+		}
+		/* If this page is far enough from the end of the env, scan for
+		 * any overflow pages that would spill onto another block.
+		 * Note we must compare against mt_last_pgno, the last written
+		 * page in the environment. Not mt_next_pgno, which increases
+		 * for every newly allocated (but not yet written) page. If
+		 * we scanned beyond the last written page we'd get a bus error.
+		 */
+		if (pgno + MDB_RPAGE_CHUNK <= txn->mt_last_pgno) {
+			int i;
+			char *cp = (char *)id3.mptr + rem * env->me_psize;
+			for (i=rem; i<MDB_RPAGE_CHUNK;) {
+				p = (MDB_page *)cp;
+				if (IS_OVERFLOW(p)) {
+					int nop = p->mp_pages;
+					if (nop + i > MDB_RPAGE_CHUNK) {
+						munmap(id3.mptr, len);
+						id3.mcnt = nop + i;
+						len = id3.mcnt * env->me_psize;
+						MAP(rc, env, id3.mptr, len, off);
+						if (rc)
+							goto fail;
+						break;
+					}
+					i += nop;
+					cp += nop * env->me_psize;
+				} else {
+					i++;
+					cp += env->me_psize;
+				}
+			}
+		}
+		mdb_mid3l_insert(el, &id3);
+		pthread_mutex_unlock(&env->me_rpmutex);
+found:
+		mdb_mid3l_insert(tl, &id3);
+	} else {
+		return MDB_TXN_FULL;
+	}
+ok:
+	p = (MDB_page *)((char *)id3.mptr + rem * env->me_psize);
+#if MDB_DEBUG	/* we don't need this check any more */
+	if (IS_OVERFLOW(p)) {
+		mdb_tassert(txn, p->mp_pages + rem <= id3.mcnt);
+	}
+#endif
+	*ret = p;
+	return MDB_SUCCESS;
+}
+#endif
+
+/** Find the address of the page corresponding to a given page number.
+ * @param[in] mc the cursor accessing the page.
+ * @param[in] pgno the page number for the page to retrieve.
+ * @param[out] ret address of a pointer where the page's address will be stored.
  * @param[out] lvl dirty_list inheritance level of found page. 1=current txn, 0=mapped page.
  * @return 0 on success, non-zero on failure.
  */
 static int
-mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **ret, int *lvl)
+mdb_page_get(MDB_cursor *mc, pgno_t pgno, MDB_page **ret, int *lvl)
 {
+	MDB_txn *txn = mc->mc_txn;
+#ifndef MDB_VL32
 	MDB_env *env = txn->mt_env;
+#endif
 	MDB_page *p = NULL;
 	int level;
 
-	if (! (txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_WRITEMAP))) {
+	if (! (mc->mc_flags & (C_ORIG_RDONLY|C_WRITEMAP))) {
 		MDB_txn *tx2 = txn;
 		level = 1;
 		do {
@@ -5318,7 +5962,13 @@ mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **ret, int *lvl)
 				MDB_ID pn = pgno << 1;
 				x = mdb_midl_search(tx2->mt_spill_pgs, pn);
 				if (x <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[x] == pn) {
+#ifdef MDB_VL32
+					int rc = mdb_rpage_get(txn, pgno, &p);
+					if (rc)
+						return rc;
+#else
 					p = (MDB_page *)(env->me_map + env->me_psize * pgno);
+#endif
 					goto done;
 				}
 			}
@@ -5335,9 +5985,17 @@ mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **ret, int *lvl)
 
 	if (pgno < txn->mt_next_pgno) {
 		level = 0;
+#ifdef MDB_VL32
+		{
+			int rc = mdb_rpage_get(txn, pgno, &p);
+			if (rc)
+				return rc;
+		}
+#else
 		p = (MDB_page *)(env->me_map + env->me_psize * pgno);
+#endif
 	} else {
-		DPRINTF(("page %"Z"u not found", pgno));
+		DPRINTF(("page %"Y"u not found", pgno));
 		txn->mt_flags |= MDB_TXN_ERROR;
 		return MDB_PAGE_NOTFOUND;
 	}
@@ -5363,9 +6021,13 @@ mdb_page_search_root(MDB_cursor *mc, MDB_val *key, int flags)
 		MDB_node	*node;
 		indx_t		i;
 
-		DPRINTF(("branch page %"Z"u has %u keys", mp->mp_pgno, NUMKEYS(mp)));
-		mdb_cassert(mc, NUMKEYS(mp) > 1);
-		DPRINTF(("found index 0 to page %"Z"u", NODEPGNO(NODEPTR(mp, 0))));
+		DPRINTF(("branch page %"Y"u has %u keys", mp->mp_pgno, NUMKEYS(mp)));
+		/* Don't assert on branch pages in the FreeDB. We can get here
+		 * while in the process of rebalancing a FreeDB branch page; we must
+		 * let that proceed. ITS#8336
+		 */
+		mdb_cassert(mc, !mc->mc_dbi || NUMKEYS(mp) > 1);
+		DPRINTF(("found index 0 to page %"Y"u", NODEPGNO(NODEPTR(mp, 0))));
 
 		if (flags & (MDB_PS_FIRST|MDB_PS_LAST)) {
 			i = 0;
@@ -5389,7 +6051,7 @@ mdb_page_search_root(MDB_cursor *mc, MDB_val *key, int flags)
 		mdb_cassert(mc, i < NUMKEYS(mp));
 		node = NODEPTR(mp, i);
 
-		if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp, NULL)) != 0)
+		if ((rc = mdb_page_get(mc, NODEPGNO(node), &mp, NULL)) != 0)
 			return rc;
 
 		mc->mc_ki[mc->mc_top] = i;
@@ -5410,7 +6072,7 @@ mdb_page_search_root(MDB_cursor *mc, MDB_val *key, int flags)
 		return MDB_CORRUPTED;
 	}
 
-	DPRINTF(("found leaf page %"Z"u for key [%s]", mp->mp_pgno,
+	DPRINTF(("found leaf page %"Y"u for key [%s]", mp->mp_pgno,
 	    key ? DKEY(key) : "null"));
 	mc->mc_flags |= C_INITIALIZED;
 	mc->mc_flags &= ~C_EOF;
@@ -5431,7 +6093,7 @@ mdb_page_search_lowest(MDB_cursor *mc)
 	MDB_node	*node = NODEPTR(mp, 0);
 	int rc;
 
-	if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp, NULL)) != 0)
+	if ((rc = mdb_page_get(mc, NODEPGNO(node), &mp, NULL)) != 0)
 		return rc;
 
 	mc->mc_ki[mc->mc_top] = 0;
@@ -5483,7 +6145,7 @@ mdb_page_search(MDB_cursor *mc, MDB_val *key, int flags)
 						return MDB_NOTFOUND;
 					if ((leaf->mn_flags & (F_DUPDATA|F_SUBDATA)) != F_SUBDATA)
 						return MDB_INCOMPATIBLE; /* not a named DB */
-					rc = mdb_node_read(mc->mc_txn, leaf, &data);
+					rc = mdb_node_read(&mc2, leaf, &data);
 					if (rc)
 						return rc;
 					memcpy(&flags, ((char *) data.mv_data + offsetof(MDB_db, md_flags)),
@@ -5506,14 +6168,26 @@ mdb_page_search(MDB_cursor *mc, MDB_val *key, int flags)
 	}
 
 	mdb_cassert(mc, root > 1);
-	if (!mc->mc_pg[0] || mc->mc_pg[0]->mp_pgno != root)
-		if ((rc = mdb_page_get(mc->mc_txn, root, &mc->mc_pg[0], NULL)) != 0)
+	if (!mc->mc_pg[0] || mc->mc_pg[0]->mp_pgno != root) {
+#ifdef MDB_VL32
+		if (mc->mc_pg[0])
+			MDB_PAGE_UNREF(mc->mc_txn, mc->mc_pg[0]);
+#endif
+		if ((rc = mdb_page_get(mc, root, &mc->mc_pg[0], NULL)) != 0)
 			return rc;
+	}
 
+#ifdef MDB_VL32
+	{
+		int i;
+		for (i=1; i<mc->mc_snum; i++)
+			MDB_PAGE_UNREF(mc->mc_txn, mc->mc_pg[i]);
+	}
+#endif
 	mc->mc_snum = 1;
 	mc->mc_top = 0;
 
-	DPRINTF(("db %d root page %"Z"u has flags 0x%X",
+	DPRINTF(("db %d root page %"Y"u has flags 0x%X",
 		DDBI(mc), root, mc->mc_pg[0]->mp_flags));
 
 	if (flags & MDB_PS_MODIFY) {
@@ -5538,7 +6212,7 @@ mdb_ovpage_free(MDB_cursor *mc, MDB_page *mp)
 	MDB_ID pn = pg << 1;
 	int rc;
 
-	DPRINTF(("free ov page %"Z"u (%d)", pg, ovpages));
+	DPRINTF(("free ov page %"Y"u (%d)", pg, ovpages));
 	/* If the page is dirty or on the spill list we just acquired it,
 	 * so we should give it back to our current free list, if any.
 	 * Otherwise put it onto the list of pages we freed in this txn.
@@ -5604,18 +6278,24 @@ release:
 }
 
 /** Return the data associated with a given node.
- * @param[in] txn The transaction for this operation.
+ * @param[in] mc The cursor for this operation.
  * @param[in] leaf The node being read.
  * @param[out] data Updated to point to the node's data.
  * @return 0 on success, non-zero on failure.
  */
 static int
-mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data)
+mdb_node_read(MDB_cursor *mc, MDB_node *leaf, MDB_val *data)
 {
 	MDB_page	*omp;		/* overflow page */
 	pgno_t		 pgno;
 	int rc;
 
+#ifdef MDB_VL32
+	if (mc->mc_ovpg) {
+		MDB_PAGE_UNREF(mc->mc_txn, mc->mc_ovpg);
+		mc->mc_ovpg = 0;
+	}
+#endif
 	if (!F_ISSET(leaf->mn_flags, F_BIGDATA)) {
 		data->mv_size = NODEDSZ(leaf);
 		data->mv_data = NODEDATA(leaf);
@@ -5626,11 +6306,14 @@ mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data)
 	 */
 	data->mv_size = NODEDSZ(leaf);
 	memcpy(&pgno, NODEDATA(leaf), sizeof(pgno));
-	if ((rc = mdb_page_get(txn, pgno, &omp, NULL)) != 0) {
-		DPRINTF(("read overflow page %"Z"u failed", pgno));
+	if ((rc = mdb_page_get(mc, pgno, &omp, NULL)) != 0) {
+		DPRINTF(("read overflow page %"Y"u failed", pgno));
 		return rc;
 	}
 	data->mv_data = METADATA(omp);
+#ifdef MDB_VL32
+	mc->mc_ovpg = omp;
+#endif
 
 	return MDB_SUCCESS;
 }
@@ -5641,7 +6324,7 @@ mdb_get(MDB_txn *txn, MDB_dbi dbi,
 {
 	MDB_cursor	mc;
 	MDB_xcursor	mx;
-	int exact = 0;
+	int exact = 0, rc;
 	DKBUF;
 
 	DPRINTF(("===> get db %u key [%s]", dbi, DKEY(key)));
@@ -5653,7 +6336,16 @@ mdb_get(MDB_txn *txn, MDB_dbi dbi,
 		return MDB_BAD_TXN;
 
 	mdb_cursor_init(&mc, txn, dbi, &mx);
-	return mdb_cursor_set(&mc, key, data, MDB_SET, &exact);
+	rc = mdb_cursor_set(&mc, key, data, MDB_SET, &exact);
+#ifdef MDB_VL32
+	{
+		/* unref all the pages - caller must copy the data
+		 * before doing anything else
+		 */
+		mdb_cursor_unref(&mc);
+	}
+#endif
+	return rc;
 }
 
 /** Find a sibling for a page.
@@ -5670,13 +6362,19 @@ mdb_cursor_sibling(MDB_cursor *mc, int move_right)
 	int		 rc;
 	MDB_node	*indx;
 	MDB_page	*mp;
+#ifdef MDB_VL32
+	MDB_page	*op;
+#endif
 
 	if (mc->mc_snum < 2) {
 		return MDB_NOTFOUND;		/* root has no siblings */
 	}
 
+#ifdef MDB_VL32
+	op = mc->mc_pg[mc->mc_top];
+#endif
 	mdb_cursor_pop(mc);
-	DPRINTF(("parent page is page %"Z"u, index %u",
+	DPRINTF(("parent page is page %"Y"u, index %u",
 		mc->mc_pg[mc->mc_top]->mp_pgno, mc->mc_ki[mc->mc_top]));
 
 	if (move_right ? (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mc->mc_pg[mc->mc_top]))
@@ -5699,8 +6397,10 @@ mdb_cursor_sibling(MDB_cursor *mc, int move_right)
 	}
 	mdb_cassert(mc, IS_BRANCH(mc->mc_pg[mc->mc_top]));
 
+	MDB_PAGE_UNREF(mc->mc_txn, op);
+
 	indx = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
-	if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(indx), &mp, NULL)) != 0) {
+	if ((rc = mdb_page_get(mc, NODEPGNO(indx), &mp, NULL)) != 0) {
 		/* mc will be inconsistent if caller does mc_snum++ as above */
 		mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
 		return rc;
@@ -5740,6 +6440,13 @@ mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
 					return rc;
 				}
 			}
+#ifdef MDB_VL32
+			else {
+				if (mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) {
+					mdb_cursor_unref(&mc->mc_xcursor->mx_cursor);
+				}
+			}
+#endif
 		} else {
 			mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
 			if (op == MDB_NEXT_DUP)
@@ -5747,7 +6454,7 @@ mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
 		}
 	}
 
-	DPRINTF(("cursor_next: top page is %"Z"u in cursor %p",
+	DPRINTF(("cursor_next: top page is %"Y"u in cursor %p",
 		mdb_dbg_pgno(mp), (void *) mc));
 	if (mc->mc_flags & C_DEL)
 		goto skip;
@@ -5759,12 +6466,12 @@ mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
 			return rc;
 		}
 		mp = mc->mc_pg[mc->mc_top];
-		DPRINTF(("next page is %"Z"u, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]));
+		DPRINTF(("next page is %"Y"u, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]));
 	} else
 		mc->mc_ki[mc->mc_top]++;
 
 skip:
-	DPRINTF(("==> cursor points to page %"Z"u with %u keys, key index %u",
+	DPRINTF(("==> cursor points to page %"Y"u with %u keys, key index %u",
 	    mdb_dbg_pgno(mp), NUMKEYS(mp), mc->mc_ki[mc->mc_top]));
 
 	if (IS_LEAF2(mp)) {
@@ -5780,7 +6487,7 @@ skip:
 		mdb_xcursor_init1(mc, leaf);
 	}
 	if (data) {
-		if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
+		if ((rc = mdb_node_read(mc, leaf, data)) != MDB_SUCCESS)
 			return rc;
 
 		if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
@@ -5819,6 +6526,13 @@ mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
 					return rc;
 				}
 			}
+#ifdef MDB_VL32
+			else {
+				if (mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) {
+					mdb_cursor_unref(&mc->mc_xcursor->mx_cursor);
+				}
+			}
+#endif
 		} else {
 			mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
 			if (op == MDB_PREV_DUP)
@@ -5826,7 +6540,7 @@ mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
 		}
 	}
 
-	DPRINTF(("cursor_prev: top page is %"Z"u in cursor %p",
+	DPRINTF(("cursor_prev: top page is %"Y"u in cursor %p",
 		mdb_dbg_pgno(mp), (void *) mc));
 
 	if (mc->mc_ki[mc->mc_top] == 0)  {
@@ -5836,13 +6550,13 @@ mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
 		}
 		mp = mc->mc_pg[mc->mc_top];
 		mc->mc_ki[mc->mc_top] = NUMKEYS(mp) - 1;
-		DPRINTF(("prev page is %"Z"u, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]));
+		DPRINTF(("prev page is %"Y"u, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]));
 	} else
 		mc->mc_ki[mc->mc_top]--;
 
 	mc->mc_flags &= ~C_EOF;
 
-	DPRINTF(("==> cursor points to page %"Z"u with %u keys, key index %u",
+	DPRINTF(("==> cursor points to page %"Y"u with %u keys, key index %u",
 	    mdb_dbg_pgno(mp), NUMKEYS(mp), mc->mc_ki[mc->mc_top]));
 
 	if (IS_LEAF2(mp)) {
@@ -5858,7 +6572,7 @@ mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
 		mdb_xcursor_init1(mc, leaf);
 	}
 	if (data) {
-		if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
+		if ((rc = mdb_node_read(mc, leaf, data)) != MDB_SUCCESS)
 			return rc;
 
 		if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
@@ -5977,6 +6691,8 @@ 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);
@@ -6016,6 +6732,11 @@ set1:
 		return MDB_SUCCESS;
 	}
 
+#ifdef MDB_VL32
+	if (mc->mc_xcursor && mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) {
+		mdb_cursor_unref(&mc->mc_xcursor->mx_cursor);
+	}
+#endif
 	if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
 		mdb_xcursor_init1(mc, leaf);
 	}
@@ -6038,11 +6759,11 @@ set1:
 		} else if (op == MDB_GET_BOTH || op == MDB_GET_BOTH_RANGE) {
 			MDB_val olddata;
 			MDB_cmp_func *dcmp;
-			if ((rc = mdb_node_read(mc->mc_txn, leaf, &olddata)) != MDB_SUCCESS)
+			if ((rc = mdb_node_read(mc, leaf, &olddata)) != MDB_SUCCESS)
 				return rc;
 			dcmp = mc->mc_dbx->md_dcmp;
-#if UINT_MAX < SIZE_MAX
-			if (dcmp == mdb_cmp_int && olddata.mv_size == sizeof(size_t))
+#if UINT_MAX < SIZE_MAX || defined(MDB_VL32)
+			if (dcmp == mdb_cmp_int && olddata.mv_size == sizeof(mdb_size_t))
 				dcmp = mdb_cmp_clong;
 #endif
 			rc = dcmp(data, &olddata);
@@ -6056,7 +6777,7 @@ set1:
 		} else {
 			if (mc->mc_xcursor)
 				mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
-			if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
+			if ((rc = mdb_node_read(mc, leaf, data)) != MDB_SUCCESS)
 				return rc;
 		}
 	}
@@ -6076,8 +6797,14 @@ mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data)
 	int		 rc;
 	MDB_node	*leaf;
 
-	if (mc->mc_xcursor)
+	if (mc->mc_xcursor) {
+#ifdef MDB_VL32
+		if (mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) {
+			mdb_cursor_unref(&mc->mc_xcursor->mx_cursor);
+		}
+#endif
 		mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
+	}
 
 	if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
 		rc = mdb_page_search(mc, NULL, MDB_PS_FIRST);
@@ -6105,7 +6832,7 @@ mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data)
 			if (rc)
 				return rc;
 		} else {
-			if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
+			if ((rc = mdb_node_read(mc, leaf, data)) != MDB_SUCCESS)
 				return rc;
 		}
 	}
@@ -6120,8 +6847,14 @@ mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data)
 	int		 rc;
 	MDB_node	*leaf;
 
-	if (mc->mc_xcursor)
+	if (mc->mc_xcursor) {
+#ifdef MDB_VL32
+		if (mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) {
+			mdb_cursor_unref(&mc->mc_xcursor->mx_cursor);
+		}
+#endif
 		mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
+	}
 
 	if (!(mc->mc_flags & C_EOF)) {
 
@@ -6150,7 +6883,7 @@ mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data)
 			if (rc)
 				return rc;
 		} else {
-			if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
+			if ((rc = mdb_node_read(mc, leaf, data)) != MDB_SUCCESS)
 				return rc;
 		}
 	}
@@ -6194,11 +6927,9 @@ 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);
+						rc = mdb_node_read(mc, leaf, data);
 					}
 				}
 			}
@@ -6304,7 +7035,7 @@ fetchm:
 			MDB_node *leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
 			if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
 				MDB_GET_KEY(leaf, key);
-				rc = mdb_node_read(mc->mc_txn, leaf, data);
+				rc = mdb_node_read(mc, leaf, data);
 				break;
 			}
 		}
@@ -6566,13 +7297,13 @@ more:
 				if (flags == MDB_CURRENT)
 					goto current;
 				dcmp = mc->mc_dbx->md_dcmp;
-#if UINT_MAX < SIZE_MAX
-				if (dcmp == mdb_cmp_int && olddata.mv_size == sizeof(size_t))
+#if UINT_MAX < SIZE_MAX || defined(MDB_VL32)
+				if (dcmp == mdb_cmp_int && olddata.mv_size == sizeof(mdb_size_t))
 					dcmp = mdb_cmp_clong;
 #endif
 				/* does data match? */
 				if (!dcmp(data, &olddata)) {
-					if (flags & MDB_NODUPDATA)
+					if (flags & (MDB_NODUPDATA|MDB_APPENDDUP))
 						return MDB_KEYEXIST;
 					/* overwrite it */
 					goto current;
@@ -6688,7 +7419,7 @@ current:
 			int level, ovpages, dpages = OVPAGES(data->mv_size, env->me_psize);
 
 			memcpy(&pg, olddata.mv_data, sizeof(pg));
-			if ((rc2 = mdb_page_get(mc->mc_txn, pg, &omp, &level)) != 0)
+			if ((rc2 = mdb_page_get(mc, pg, &omp, &level)) != 0)
 				return rc2;
 			ovpages = omp->mp_pages;
 
@@ -6774,7 +7505,7 @@ 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;
@@ -6786,10 +7517,15 @@ 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);
+				}
 			}
 		}
 	}
@@ -6922,6 +7658,7 @@ 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);
@@ -6957,8 +7694,9 @@ 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 */
 		}
@@ -6982,7 +7720,7 @@ mdb_cursor_del(MDB_cursor *mc, unsigned int flags)
 		pgno_t pg;
 
 		memcpy(&pg, NODEDATA(leaf), sizeof(pg));
-		if ((rc = mdb_page_get(mc->mc_txn, pg, &omp, NULL)) ||
+		if ((rc = mdb_page_get(mc, pg, &omp, NULL)) ||
 			(rc = mdb_ovpage_free(mc, omp)))
 			goto fail;
 	}
@@ -7011,7 +7749,7 @@ mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp)
 
 	if ((rc = mdb_page_alloc(mc, num, &np)))
 		return rc;
-	DPRINTF(("allocated new mpage %"Z"u, page size %u",
+	DPRINTF(("allocated new mpage %"Y"u, page size %u",
 	    np->mp_pgno, mc->mc_txn->mt_env->me_psize));
 	np->mp_flags = flags | P_DIRTY;
 	np->mp_lower = (PAGEHDRSZ-PAGEBASE);
@@ -7111,7 +7849,7 @@ mdb_node_add(MDB_cursor *mc, indx_t indx,
 
 	mdb_cassert(mc, mp->mp_upper >= mp->mp_lower);
 
-	DPRINTF(("add to %s %spage %"Z"u index %i, data size %"Z"u key size %"Z"u [%s]",
+	DPRINTF(("add to %s %spage %"Y"u index %i, data size %"Z"u key size %"Z"u [%s]",
 	    IS_LEAF(mp) ? "leaf" : "branch",
 		IS_SUBP(mp) ? "sub-" : "",
 		mdb_dbg_pgno(mp), indx, data ? data->mv_size : 0,
@@ -7152,7 +7890,7 @@ mdb_node_add(MDB_cursor *mc, indx_t indx,
 				goto full;
 			if ((rc = mdb_page_new(mc, P_OVERFLOW, ovpages, &ofp)))
 				return rc;
-			DPRINTF(("allocated overflow page %"Z"u", ofp->mp_pgno));
+			DPRINTF(("allocated overflow page %"Y"u", ofp->mp_pgno));
 			flags |= F_BIGDATA;
 			goto update;
 		} else {
@@ -7209,7 +7947,7 @@ update:
 	return MDB_SUCCESS;
 
 full:
-	DPRINTF(("not enough room in page %"Z"u, got %u ptrs",
+	DPRINTF(("not enough room in page %"Y"u, got %u ptrs",
 		mdb_dbg_pgno(mp), NUMKEYS(mp)));
 	DPRINTF(("upper-lower = %u - %u = %"Z"d", mp->mp_upper,mp->mp_lower,room));
 	DPRINTF(("node size = %"Z"u", node_size));
@@ -7232,7 +7970,7 @@ mdb_node_del(MDB_cursor *mc, int ksize)
 	MDB_node	*node;
 	char		*base;
 
-	DPRINTF(("delete node %u on %s page %"Z"u", indx,
+	DPRINTF(("delete node %u on %s page %"Y"u", indx,
 	    IS_LEAF(mp) ? "leaf" : "branch", mdb_dbg_pgno(mp)));
 	numkeys = NUMKEYS(mp);
 	mdb_cassert(mc, indx < numkeys);
@@ -7341,7 +8079,10 @@ mdb_xcursor_init0(MDB_cursor *mc)
 	mx->mx_cursor.mc_dbflag = &mx->mx_dbflag;
 	mx->mx_cursor.mc_snum = 0;
 	mx->mx_cursor.mc_top = 0;
-	mx->mx_cursor.mc_flags = C_SUB;
+#ifdef MDB_VL32
+	mx->mx_cursor.mc_ovpg = 0;
+#endif
+	mx->mx_cursor.mc_flags = C_SUB | (mc->mc_flags & (C_ORIG_RDONLY|C_WRITEMAP));
 	mx->mx_dbx.md_name.mv_size = 0;
 	mx->mx_dbx.md_name.mv_data = NULL;
 	mx->mx_dbx.md_cmp = mc->mc_dbx->md_dcmp;
@@ -7360,12 +8101,12 @@ mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node)
 {
 	MDB_xcursor *mx = mc->mc_xcursor;
 
+	mx->mx_cursor.mc_flags &= C_SUB|C_ORIG_RDONLY|C_WRITEMAP;
 	if (node->mn_flags & F_SUBDATA) {
 		memcpy(&mx->mx_db, NODEDATA(node), sizeof(MDB_db));
 		mx->mx_cursor.mc_pg[0] = 0;
 		mx->mx_cursor.mc_snum = 0;
 		mx->mx_cursor.mc_top = 0;
-		mx->mx_cursor.mc_flags = C_SUB;
 	} else {
 		MDB_page *fp = NODEDATA(node);
 		mx->mx_db.md_pad = 0;
@@ -7378,7 +8119,7 @@ mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node)
 		COPY_PGNO(mx->mx_db.md_root, fp->mp_pgno);
 		mx->mx_cursor.mc_snum = 1;
 		mx->mx_cursor.mc_top = 0;
-		mx->mx_cursor.mc_flags = C_INITIALIZED|C_SUB;
+		mx->mx_cursor.mc_flags |= C_INITIALIZED;
 		mx->mx_cursor.mc_pg[0] = fp;
 		mx->mx_cursor.mc_ki[0] = 0;
 		if (mc->mc_db->md_flags & MDB_DUPFIXED) {
@@ -7388,11 +8129,11 @@ mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node)
 				mx->mx_db.md_flags |= MDB_INTEGERKEY;
 		}
 	}
-	DPRINTF(("Sub-db -%u root page %"Z"u", mx->mx_cursor.mc_dbi,
+	DPRINTF(("Sub-db -%u root page %"Y"u", mx->mx_cursor.mc_dbi,
 		mx->mx_db.md_root));
 	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))
+#if UINT_MAX < SIZE_MAX || defined(MDB_VL32)
+	if (mx->mx_dbx.md_cmp == mdb_cmp_int && mx->mx_db.md_pad == sizeof(mdb_size_t))
 		mx->mx_dbx.md_cmp = mdb_cmp_clong;
 #endif
 }
@@ -7425,7 +8166,7 @@ mdb_xcursor_init2(MDB_cursor *mc, MDB_xcursor *src_mx, int new_dupdata)
 	}
 	mx->mx_db = src_mx->mx_db;
 	mx->mx_cursor.mc_pg[0] = src_mx->mx_cursor.mc_pg[0];
-	DPRINTF(("Sub-db -%u root page %"Z"u", mx->mx_cursor.mc_dbi,
+	DPRINTF(("Sub-db -%u root page %"Y"u", mx->mx_cursor.mc_dbi,
 		mx->mx_db.md_root));
 }
 
@@ -7444,7 +8185,10 @@ mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx)
 	mc->mc_top = 0;
 	mc->mc_pg[0] = 0;
 	mc->mc_ki[0] = 0;
-	mc->mc_flags = 0;
+#ifdef MDB_VL32
+	mc->mc_ovpg = 0;
+#endif
+	mc->mc_flags = txn->mt_flags & (C_ORIG_RDONLY|C_WRITEMAP);
 	if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
 		mdb_tassert(txn, mx != NULL);
 		mc->mc_xcursor = mx;
@@ -7594,7 +8338,7 @@ mdb_update_key(MDB_cursor *mc, MDB_val *key)
 		char kbuf2[DKBUF_MAXKEYSIZE*2+1];
 		k2.mv_data = NODEKEY(node);
 		k2.mv_size = node->mn_ksize;
-		DPRINTF(("update key %u (ofs %u) [%s] to [%s] on page %"Z"u",
+		DPRINTF(("update key %u (ofs %u) [%s] to [%s] on page %"Y"u",
 			indx, ptr,
 			mdb_dkey(&k2, kbuf2),
 			DKEY(key),
@@ -7645,6 +8389,22 @@ 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
@@ -7700,6 +8460,7 @@ mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst, int fromleft)
 		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;
@@ -7725,7 +8486,7 @@ mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst, int fromleft)
 			return rc;
 	}
 
-	DPRINTF(("moving %s node %u [%s] on page %"Z"u to node %u on page %"Z"u",
+	DPRINTF(("moving %s node %u [%s] on page %"Y"u to node %u on page %"Y"u",
 	    IS_LEAF(csrc->mc_pg[csrc->mc_top]) ? "leaf" : "branch",
 	    csrc->mc_ki[csrc->mc_top],
 		DKEY(&key),
@@ -7771,6 +8532,12 @@ mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst, int fromleft)
 					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);
+				}
 			}
 		} else
 		/* Adding on the right, bump others down */
@@ -7791,6 +8558,12 @@ mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst, int fromleft)
 					} 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);
+					}
 				}
 			}
 		}
@@ -7807,12 +8580,15 @@ mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst, int fromleft)
 				key.mv_size = NODEKSZ(srcnode);
 				key.mv_data = NODEKEY(srcnode);
 			}
-			DPRINTF(("update separator for source page %"Z"u to [%s]",
+			DPRINTF(("update separator for source page %"Y"u to [%s]",
 				csrc->mc_pg[csrc->mc_top]->mp_pgno, DKEY(&key)));
 			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])) {
@@ -7835,12 +8611,15 @@ mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst, int fromleft)
 				key.mv_size = NODEKSZ(srcnode);
 				key.mv_data = NODEKEY(srcnode);
 			}
-			DPRINTF(("update separator for destination page %"Z"u to [%s]",
+			DPRINTF(("update separator for destination page %"Y"u to [%s]",
 				cdst->mc_pg[cdst->mc_top]->mp_pgno, DKEY(&key)));
 			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])) {
@@ -7878,7 +8657,7 @@ mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst)
 	psrc = csrc->mc_pg[csrc->mc_top];
 	pdst = cdst->mc_pg[cdst->mc_top];
 
-	DPRINTF(("merging page %"Z"u into %"Z"u", psrc->mp_pgno, pdst->mp_pgno));
+	DPRINTF(("merging page %"Y"u into %"Y"u", psrc->mp_pgno, pdst->mp_pgno));
 
 	mdb_cassert(csrc, csrc->mc_snum > 1);	/* can't merge root page */
 	mdb_cassert(csrc, cdst->mc_snum > 1);
@@ -7909,6 +8688,7 @@ 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)
@@ -7934,7 +8714,7 @@ mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst)
 		}
 	}
 
-	DPRINTF(("dst page %"Z"u now has %u keys (%.1f%% filled)",
+	DPRINTF(("dst page %"Y"u now has %u keys (%.1f%% filled)",
 	    pdst->mp_pgno, NUMKEYS(pdst),
 		(float)PAGEFILL(cdst->mc_txn->mt_env, pdst) / 10));
 
@@ -7984,6 +8764,12 @@ mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst)
 				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);
+			}
 		}
 	}
 	{
@@ -8016,6 +8802,9 @@ mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst)
 	cdst->mc_snum = csrc->mc_snum;
 	cdst->mc_top = csrc->mc_top;
 	cdst->mc_flags = csrc->mc_flags;
+#ifdef MDB_VL32
+	cdst->mc_ovpg = csrc->mc_ovpg;
+#endif
 
 	for (i=0; i<csrc->mc_snum; i++) {
 		cdst->mc_pg[i] = csrc->mc_pg[i];
@@ -8044,14 +8833,14 @@ mdb_rebalance(MDB_cursor *mc)
 		minkeys = 1;
 		thresh = FILL_THRESHOLD;
 	}
-	DPRINTF(("rebalancing %s page %"Z"u (has %u keys, %.1f%% full)",
+	DPRINTF(("rebalancing %s page %"Y"u (has %u keys, %.1f%% full)",
 	    IS_LEAF(mc->mc_pg[mc->mc_top]) ? "leaf" : "branch",
 	    mdb_dbg_pgno(mc->mc_pg[mc->mc_top]), NUMKEYS(mc->mc_pg[mc->mc_top]),
 		(float)PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) / 10));
 
 	if (PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) >= thresh &&
 		NUMKEYS(mc->mc_pg[mc->mc_top]) >= minkeys) {
-		DPRINTF(("no need to rebalance page %"Z"u, above fill threshold",
+		DPRINTF(("no need to rebalance page %"Y"u, above fill threshold",
 		    mdb_dbg_pgno(mc->mc_pg[mc->mc_top])));
 		return MDB_SUCCESS;
 	}
@@ -8099,7 +8888,7 @@ mdb_rebalance(MDB_cursor *mc)
 			if (rc)
 				return rc;
 			mc->mc_db->md_root = NODEPGNO(NODEPTR(mp, 0));
-			rc = mdb_page_get(mc->mc_txn,mc->mc_db->md_root,&mc->mc_pg[0],NULL);
+			rc = mdb_page_get(mc, mc->mc_db->md_root, &mc->mc_pg[0], NULL);
 			if (rc)
 				return rc;
 			mc->mc_db->md_depth--;
@@ -8160,7 +8949,7 @@ mdb_rebalance(MDB_cursor *mc)
 		DPUTS("reading right neighbor");
 		mn.mc_ki[ptop]++;
 		node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
-		rc = mdb_page_get(mc->mc_txn,NODEPGNO(node),&mn.mc_pg[mn.mc_top],NULL);
+		rc = mdb_page_get(mc, NODEPGNO(node), &mn.mc_pg[mn.mc_top], NULL);
 		if (rc)
 			return rc;
 		mn.mc_ki[mn.mc_top] = 0;
@@ -8172,7 +8961,7 @@ mdb_rebalance(MDB_cursor *mc)
 		DPUTS("reading left neighbor");
 		mn.mc_ki[ptop]--;
 		node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
-		rc = mdb_page_get(mc->mc_txn,NODEPGNO(node),&mn.mc_pg[mn.mc_top],NULL);
+		rc = mdb_page_get(mc, NODEPGNO(node), &mn.mc_pg[mn.mc_top], NULL);
 		if (rc)
 			return rc;
 		mn.mc_ki[mn.mc_top] = NUMKEYS(mn.mc_pg[mn.mc_top]) - 1;
@@ -8180,7 +8969,7 @@ mdb_rebalance(MDB_cursor *mc)
 		fromleft = 1;
 	}
 
-	DPRINTF(("found neighbor page %"Z"u (%u keys, %.1f%% full)",
+	DPRINTF(("found neighbor page %"Y"u (%u keys, %.1f%% full)",
 	    mn.mc_pg[mn.mc_top]->mp_pgno, NUMKEYS(mn.mc_pg[mn.mc_top]),
 		(float)PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) / 10));
 
@@ -8198,24 +8987,11 @@ mdb_rebalance(MDB_cursor *mc)
 		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_flags = C_INITIALIZED;
-				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;
@@ -8248,12 +9024,17 @@ 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)
+					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);
 				}
 			}
 		}
@@ -8390,7 +9171,7 @@ mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata, pgno_t newpgno
 	newindx = mc->mc_ki[mc->mc_top];
 	nkeys = NUMKEYS(mp);
 
-	DPRINTF(("-----> splitting %s page %"Z"u and adding [%s] at index %i/%i",
+	DPRINTF(("-----> splitting %s page %"Y"u and adding [%s] at index %i/%i",
 	    IS_LEAF(mp) ? "leaf" : "branch", mp->mp_pgno,
 	    DKEY(newkey), mc->mc_ki[mc->mc_top], nkeys));
 
@@ -8398,7 +9179,7 @@ mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata, pgno_t newpgno
 	if ((rc = mdb_page_new(mc, mp->mp_flags, 1, &rp)))
 		return rc;
 	rp->mp_pad = mp->mp_pad;
-	DPRINTF(("new right sibling: page %"Z"u", rp->mp_pgno));
+	DPRINTF(("new right sibling: page %"Y"u", rp->mp_pgno));
 
 	/* Usually when splitting the root page, the cursor
 	 * height is 1. But when called from mdb_update_key,
@@ -8416,7 +9197,7 @@ mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata, pgno_t newpgno
 		mc->mc_pg[0] = pp;
 		mc->mc_ki[0] = 0;
 		mc->mc_db->md_root = pp->mp_pgno;
-		DPRINTF(("root split! new root = %"Z"u", pp->mp_pgno));
+		DPRINTF(("root split! new root = %"Y"u", pp->mp_pgno));
 		new_root = mc->mc_db->md_depth++;
 
 		/* Add left (implicit) pointer. */
@@ -8433,10 +9214,11 @@ mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata, pgno_t newpgno
 		ptop = 0;
 	} else {
 		ptop = mc->mc_top-1;
-		DPRINTF(("parent branch page is %"Z"u", mc->mc_pg[ptop]->mp_pgno));
+		DPRINTF(("parent branch page is %"Y"u", mc->mc_pg[ptop]->mp_pgno));
 	}
 
 	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;
 
@@ -8486,8 +9268,6 @@ 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;
-				mc->mc_ki[ptop]++;
 			}
 		} else {
 			int psize, nsize, k;
@@ -8584,7 +9364,9 @@ mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata, pgno_t newpgno
 		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;
 
@@ -8684,11 +9466,6 @@ 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]++;
@@ -8702,13 +9479,32 @@ 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)
@@ -8721,12 +9517,15 @@ mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata, pgno_t newpgno
 				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;
@@ -8738,15 +9537,24 @@ 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)));
@@ -8765,6 +9573,7 @@ mdb_put(MDB_txn *txn, MDB_dbi dbi,
 {
 	MDB_cursor mc;
 	MDB_xcursor mx;
+	int rc;
 
 	if (!key || !data || !TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
 		return EINVAL;
@@ -8776,7 +9585,11 @@ mdb_put(MDB_txn *txn, MDB_dbi dbi,
 		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
@@ -8890,7 +9703,6 @@ static int ESECT
 mdb_env_cwalk(mdb_copy *my, pgno_t *pg, int flags)
 {
 	MDB_cursor mc;
-	MDB_txn *txn = my->mc_txn;
 	MDB_node *ni;
 	MDB_page *mo, *mp, *leaf;
 	char *buf, *ptr;
@@ -8903,9 +9715,9 @@ mdb_env_cwalk(mdb_copy *my, pgno_t *pg, int flags)
 
 	mc.mc_snum = 1;
 	mc.mc_top = 0;
-	mc.mc_txn = txn;
+	mc.mc_txn = my->mc_txn;
 
-	rc = mdb_page_get(my->mc_txn, *pg, &mc.mc_pg[0], NULL);
+	rc = mdb_page_get(&mc, *pg, &mc.mc_pg[0], NULL);
 	if (rc)
 		return rc;
 	rc = mdb_page_search_root(&mc, NULL, MDB_PS_FIRST);
@@ -8949,7 +9761,7 @@ mdb_env_cwalk(mdb_copy *my, pgno_t *pg, int flags)
 						}
 
 						memcpy(&pg, NODEDATA(ni), sizeof(pg));
-						rc = mdb_page_get(txn, pg, &omp, NULL);
+						rc = mdb_page_get(&mc, pg, &omp, NULL);
 						if (rc)
 							goto done;
 						if (my->mc_wlen[toggle] >= MDB_WBUF) {
@@ -9000,7 +9812,7 @@ mdb_env_cwalk(mdb_copy *my, pgno_t *pg, int flags)
 again:
 				ni = NODEPTR(mp, mc.mc_ki[mc.mc_top]);
 				pg = NODEPGNO(ni);
-				rc = mdb_page_get(txn, pg, &mp, NULL);
+				rc = mdb_page_get(&mc, pg, &mp, NULL);
 				if (rc)
 					goto done;
 				mc.mc_top++;
@@ -9226,7 +10038,7 @@ mdb_env_copyfd0(MDB_env *env, HANDLE fd)
 
 	w2 = txn->mt_next_pgno * env->me_psize;
 	{
-		size_t fsize = 0;
+		mdb_size_t fsize = 0;
 		if ((rc = mdb_fsize(env->me_fd, &fsize)))
 			goto leave;
 		if (w2 > fsize)
@@ -9299,7 +10111,9 @@ mdb_env_copy2(MDB_env *env, const char *path, unsigned int flags)
 	 * already in the OS cache.
 	 */
 #ifdef _WIN32
-	utf8_to_utf16(lpath, -1, &wpath, NULL);
+	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);
@@ -9494,6 +10308,7 @@ 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 (flags & ~VALID_FLAGS)
@@ -9555,8 +10370,16 @@ 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));
@@ -9566,10 +10389,12 @@ 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;
@@ -9657,6 +10482,11 @@ mdb_drop0(MDB_cursor *mc, int subs)
 			mdb_cursor_pop(mc);
 
 		mdb_cursor_copy(mc, &mx);
+#ifdef MDB_VL32
+		/* bump refcount for mx's pages */
+		for (i=0; i<mc->mc_snum; i++)
+			mdb_page_get(&mx, mc->mc_pg[i]->mp_pgno, &mx.mc_pg[i], NULL);
+#endif
 		while (mc->mc_snum > 0) {
 			MDB_page *mp = mc->mc_pg[mc->mc_top];
 			unsigned n = NUMKEYS(mp);
@@ -9667,7 +10497,7 @@ mdb_drop0(MDB_cursor *mc, int subs)
 						MDB_page *omp;
 						pgno_t pg;
 						memcpy(&pg, NODEDATA(ni), sizeof(pg));
-						rc = mdb_page_get(txn, pg, &omp, NULL);
+						rc = mdb_page_get(mc, pg, &omp, NULL);
 						if (rc != 0)
 							goto done;
 						mdb_cassert(mc, IS_OVERFLOW(omp));
@@ -9716,9 +10546,14 @@ mdb_drop0(MDB_cursor *mc, int subs)
 done:
 		if (rc)
 			txn->mt_flags |= MDB_TXN_ERROR;
+#ifdef MDB_VL32
+		/* drop refcount for mx's pages */
+		mdb_cursor_unref(&mx);
+#endif
 	} else if (rc == MDB_NOTFOUND) {
 		rc = MDB_SUCCESS;
 	}
+	mc->mc_flags &= ~C_INITIALIZED;
 	return rc;
 }
 
@@ -9834,7 +10669,7 @@ mdb_reader_list(MDB_env *env, MDB_msg_func *func, void *ctx)
 		if (mr[i].mr_pid) {
 			txnid_t	txnid = mr[i].mr_txnid;
 			sprintf(buf, txnid == (txnid_t)-1 ?
-				"%10d %"Z"x -\n" : "%10d %"Z"x %"Z"u\n",
+				"%10d %"Z"x -\n" : "%10d %"Z"x %"Y"u\n",
 				(int)mr[i].mr_pid, (size_t)mr[i].mr_tid, txnid);
 			if (first) {
 				first = 0;
@@ -9939,7 +10774,7 @@ mdb_reader_check0(MDB_env *env, int rlocked, int *dead)
 					}
 					for (; j<rdrs; j++)
 							if (mr[j].mr_pid == pid) {
-								DPRINTF(("clear stale reader pid %u txn %"Z"d",
+								DPRINTF(("clear stale reader pid %u txn %"Y"d",
 									(unsigned) pid, mr[j].mr_txnid));
 								mr[j].mr_pid = 0;
 								count++;
@@ -10019,6 +10854,8 @@ static int utf8_to_utf16(const char *src, int srcsize, wchar_t **dst, int *dstsi
 	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;