X-Git-Url: https://pd.if.org/git/?p=btree;a=blobdiff_plain;f=threads2i.c;h=f1d0269b89776928b209bef823aa65c40c4f91ae;hp=25a26fdfa0cf83c56cb86106b5d95b8ceeceeabf;hb=HEAD;hpb=a1084ff92d2982f1225ad2db3ccb4ba1051bd5f8

diff --git a/threads2i.c b/threads2i.c
index 25a26fd..f1d0269 100644
--- a/threads2i.c
+++ b/threads2i.c
@@ -1,5 +1,6 @@
-// btree version threads2i linux futex concurrency version
-// 09 JAN 2014
+// btree version threads2i sched_yield version
+//	with reworked bt_deletekey code
+// 17 FEB 2014
 
 // author: karl malbrain, malbrain@cal.berkeley.edu
 
@@ -25,8 +26,6 @@ REDISTRIBUTION OF THIS SOFTWARE.
 
 #ifdef linux
 #define _GNU_SOURCE
-#include <linux/futex.h>
-#define SYS_futex 202
 #endif
 
 #ifdef unix
@@ -38,7 +37,6 @@ REDISTRIBUTION OF THIS SOFTWARE.
 #include <sys/mman.h>
 #include <errno.h>
 #include <pthread.h>
-#include <limits.h>
 #else
 #define WIN32_LEAN_AND_MEAN
 #include <windows.h>
@@ -52,6 +50,7 @@ REDISTRIBUTION OF THIS SOFTWARE.
 
 #include <memory.h>
 #include <string.h>
+#include <stddef.h>
 
 typedef unsigned long long	uid;
 
@@ -61,6 +60,8 @@ typedef unsigned short		ushort;
 typedef unsigned int		uint;
 #endif
 
+#define BT_latchtable	128					// number of latch manager slots
+
 #define BT_ro 0x6f72	// ro
 #define BT_rw 0x7772	// rw
 
@@ -86,36 +87,38 @@ typedef enum{
 	BtLockParent
 } BtLock;
 
-//	mode & definition for latch implementation
-
-enum {
-	Mutex = 1 << 0,		// the mutex bit
-	Write = 1 << 1, 	// the writers bit
-	Share = 1 << 2,		// reader count
-	PendRd = 1 << 12,	// reader contended count
-	PendWr = 1 << 22	// writer contended count
-} LockMode;
-
-enum {
-	QueRd = 1,	// reader queue
-	QueWr = 2	// writer queue
-} RWQueue;
+//	definition for latch implementation
 
+// exclusive is set for write access
 // share is count of read accessors
 // grant write lock when share == 0
 
+volatile typedef struct {
+	unsigned char mutex[1];
+	unsigned char exclusive:1;
+	unsigned char pending:1;
+	ushort share;
+} BtSpinLatch;
+
+//  hash table entries
+
 typedef struct {
-	volatile uint mutex:1;		// 1 = busy
-	volatile uint write:1;		// 1 = exclusive
-	volatile uint share:10;		// count of readers holding locks
-	volatile uint readwait:10;	// count of readers waiting
-	volatile uint writewait:10;	// count of writers waiting
-} BtLatch;
+	BtSpinLatch latch[1];
+	volatile ushort slot;		// Latch table entry at head of chain
+} BtHashEntry;
+
+//	latch manager table structure
 
 typedef struct {
-	BtLatch readwr[1];		// read/write page lock
-	BtLatch access[1];		// Access Intent/Page delete
-	BtLatch parent[1];		// Parent modification
+	BtSpinLatch readwr[1];		// read/write page lock
+	BtSpinLatch access[1];		// Access Intent/Page delete
+	BtSpinLatch parent[1];		// Posting of fence key in parent
+	BtSpinLatch busy[1];		// slot is being moved between chains
+	volatile ushort next;		// next entry in hash table chain
+	volatile ushort prev;		// prev entry in hash table chain
+	volatile ushort pin;		// number of outstanding locks
+	volatile ushort hash;		// hash slot entry is under
+	volatile uid page_no;		// latch set page number
 } BtLatchSet;
 
 //	Define the length of the page and key pointers
@@ -141,7 +144,7 @@ typedef struct {
 } BtSlot;
 
 //	The key structure occupies space at the upper end of
-//	each page.  It's a length byte followed by the value
+//	each page.  It's a length byte followed by the key
 //	bytes.
 
 typedef struct {
@@ -153,12 +156,12 @@ typedef struct {
 //	It is immediately followed
 //	by the BtSlot array of keys.
 
-typedef struct Page {
-	BtLatchSet latch[1];		// Set of three latches
+typedef struct BtPage_ {
 	uint cnt;					// count of keys in page
 	uint act;					// count of active keys
 	uint min;					// next key offset
-	unsigned char bits;			// page size in bits
+	unsigned char bits:7;		// page size in bits
+	unsigned char free:1;		// page is on free chain
 	unsigned char lvl:6;		// level of page
 	unsigned char kill:1;		// page is being deleted
 	unsigned char dirty:1;		// page has deleted keys
@@ -168,18 +171,41 @@ typedef struct Page {
 //	The memory mapping pool table buffer manager entry
 
 typedef struct {
-	unsigned long long int lru;	// number of times accessed
 	uid  basepage;				// mapped base page number
 	char *map;					// mapped memory pointer
-	uint slot;					// slot index in this array
-	volatile uint pin;			// mapped page pin counter
+	ushort slot;				// slot index in this array
+	ushort pin;					// mapped page pin counter
 	void *hashprev;				// previous pool entry for the same hash idx
 	void *hashnext;				// next pool entry for the same hash idx
 #ifndef unix
-	HANDLE hmap;
+	HANDLE hmap;				// Windows memory mapping handle
 #endif
 } BtPool;
 
+#define CLOCK_bit 0x8000		// bit in pool->pin
+
+//  The loadpage interface object
+
+typedef struct {
+	uid page_no;		// current page number
+	BtPage page;		// current page pointer
+	BtPool *pool;		// current page pool
+	BtLatchSet *latch;	// current page latch set
+} BtPageSet;
+
+//	structure for latch manager on ALLOC_page
+
+typedef struct {
+	struct BtPage_ alloc[2];	// next & free page_nos in right ptr
+	BtSpinLatch lock[1];		// allocation area lite latch
+	ushort latchdeployed;		// highest number of latch entries deployed
+	ushort nlatchpage;			// number of latch pages at BT_latch
+	ushort latchtotal;			// number of page latch entries
+	ushort latchhash;			// number of latch hash table slots
+	ushort latchvictim;			// next latch entry to examine
+	BtHashEntry table[0];		// the hash table
+} BtLatchMgr;
+
 //	The object structure for Btree access
 
 typedef struct {
@@ -188,32 +214,33 @@ typedef struct {
 	uint seg_bits;				// seg size in pages in bits
 	uint mode;					// read-write mode
 #ifdef unix
-	char *pooladvise;			// bit maps for pool page advisements
 	int idx;
 #else
 	HANDLE idx;
 #endif
-	volatile uint poolcnt;		// highest page pool node in use
+	ushort poolcnt;				// highest page pool node in use
+	ushort poolmax;				// highest page pool node allocated
+	ushort poolmask;			// total number of pages in mmap segment - 1
+	ushort hashsize;			// size of Hash Table for pool entries
 	volatile uint evicted;		// last evicted hash table slot
-	uint poolmax;				// highest page pool node allocated
-	uint poolmask;				// total size of pages in mmap segment - 1
-	uint hashsize;				// size of Hash Table for pool entries
 	ushort *hash;				// pool index for hash entries
-	BtLatch *latch;				// latches for hash table slots
+	BtSpinLatch *latch;			// latches for hash table slots
+	BtLatchMgr *latchmgr;		// mapped latch page from allocation page
+	BtLatchSet *latchsets;		// mapped latch set from latch pages
 	BtPool *pool;				// memory pool page segments
+#ifndef unix
+	HANDLE halloc;				// allocation and latch table handle
+#endif
 } BtMgr;
 
 typedef struct {
 	BtMgr *mgr;			// buffer manager for thread
-	BtPage temp;		// temporary frame buffer (memory mapped/file IO)
-	BtPage alloc;		// frame buffer for alloc page ( mapped to page 0 )
 	BtPage cursor;		// cached frame for start/next (never mapped)
 	BtPage frame;		// spare frame for the page split (never mapped)
-	BtPage zero;		// page of zeroes to extend the file (never mapped)
-	BtPage page;		// current page mapped from file
-	uid page_no;		// current page number	
+	BtPage zero;		// page frame for zeroes at end of file
 	uid cursor_page;	// current cursor page number	
 	unsigned char *mem;	// frame, cursor, page memory buffer
+	int found;			// last delete or insert was found
 	int err;			// last error
 } BtDb;
 
@@ -230,7 +257,7 @@ typedef enum {
 // B-Tree functions
 extern void bt_close (BtDb *bt);
 extern BtDb *bt_open (BtMgr *mgr);
-extern BTERR  bt_insertkey (BtDb *bt, unsigned char *key, uint len, uint lvl, uid id, uint tod);
+extern BTERR bt_insertkey (BtDb *bt, unsigned char *key, uint len, uint lvl, uid id, uint tod);
 extern BTERR  bt_deletekey (BtDb *bt, unsigned char *key, uint len, uint lvl);
 extern uid bt_findkey    (BtDb *bt, unsigned char *key, uint len);
 extern uint bt_startkey  (BtDb *bt, unsigned char *key, uint len);
@@ -247,9 +274,10 @@ extern uid bt_uid (BtDb *bt, uint slot);
 extern uint bt_tod (BtDb *bt, uint slot);
 
 //  BTree page number constants
-#define ALLOC_page		0
-#define ROOT_page		1
-#define LEAF_page		2
+#define ALLOC_page		0	// allocation & lock manager hash table
+#define ROOT_page		1	// root of the btree
+#define LEAF_page		2	// first page of leaves
+#define LATCH_page		3	// pages for lock manager
 
 //	Number of levels to create in a new BTree
 
@@ -264,7 +292,8 @@ extern uint bt_tod (BtDb *bt, uint slot);
 //  A key consists of a length byte, two bytes of
 //  index number (0 - 65534), and up to 253 bytes
 //  of key value.  Duplicate keys are discarded.
-//  Associated with each key is a 48 bit row-id.
+//  Associated with each key is a 48 bit row-id,
+//	or any other value desired.
 
 //  The b-tree root is always located at page 1.
 //	The first leaf page of level zero is always
@@ -279,8 +308,8 @@ extern uint bt_tod (BtDb *bt, uint slot);
 //	one with two keys.
 
 //	Deleted keys are marked with a dead bit until
-//	page cleanup The fence key for a node is always
-//	present, even after deletion and cleanup.
+//	page cleanup. The fence key for a node is
+//	always present
 
 //  Groups of pages called segments from the btree are optionally
 //  cached with a memory mapped pool. A hash table is used to keep
@@ -295,12 +324,13 @@ extern uint bt_tod (BtDb *bt, uint slot);
 //  Page 0 is dedicated to lock for new page extensions,
 //	and chains empty pages together for reuse.
 
-//	The ParentModification lock on a node is obtained to prevent resplitting
-//	or deleting a node before its fence is posted into its upper level.
+//	The ParentModification lock on a node is obtained to serialize posting
+//	or changing the fence key for a node.
 
 //	Empty pages are chained together through the ALLOC page and reused.
 
 //	Access macros to address slot and key values from the page
+//	Page slots use 1 based indexing.
 
 #define slotptr(page, slot) (((BtSlot *)(page+1)) + (slot-1))
 #define keyptr(page, slot) ((BtKey)((unsigned char*)(page) + slotptr(page, slot)->off))
@@ -324,6 +354,315 @@ int i;
 	return id;
 }
 
+//	Latch Manager
+
+//	wait until write lock mode is clear
+//	and add 1 to the share count
+
+void bt_spinreadlock(BtSpinLatch *latch)
+{
+ushort prev;
+
+  do {
+	//	obtain latch mutex
+#ifdef unix
+	if( __sync_lock_test_and_set(latch->mutex, 1) )
+		continue;
+#else
+	if( _InterlockedExchange8(latch->mutex, 1) )
+		continue;
+#endif
+	//  see if exclusive request is granted or pending
+
+	if( prev = !(latch->exclusive | latch->pending) )
+		latch->share++;
+
+#ifdef unix
+	*latch->mutex = 0;
+#else
+	_InterlockedExchange8(latch->mutex, 0);
+#endif
+
+	if( prev )
+		return;
+
+#ifdef  unix
+  } while( sched_yield(), 1 );
+#else
+  } while( SwitchToThread(), 1 );
+#endif
+}
+
+//	wait for other read and write latches to relinquish
+
+void bt_spinwritelock(BtSpinLatch *latch)
+{
+uint prev;
+
+  do {
+#ifdef  unix
+	if( __sync_lock_test_and_set(latch->mutex, 1) )
+		continue;
+#else
+	if( _InterlockedExchange8(latch->mutex, 1) )
+		continue;
+#endif
+	if( prev = !(latch->share | latch->exclusive) )
+		latch->exclusive = 1, latch->pending = 0;
+	else
+		latch->pending = 1;
+#ifdef unix
+	*latch->mutex = 0;
+#else
+	_InterlockedExchange8(latch->mutex, 0);
+#endif
+	if( prev )
+		return;
+#ifdef  unix
+  } while( sched_yield(), 1 );
+#else
+  } while( SwitchToThread(), 1 );
+#endif
+}
+
+//	try to obtain write lock
+
+//	return 1 if obtained,
+//		0 otherwise
+
+int bt_spinwritetry(BtSpinLatch *latch)
+{
+uint prev;
+
+#ifdef unix
+	if( __sync_lock_test_and_set(latch->mutex, 1) )
+		return 0;
+#else
+	if( _InterlockedExchange8(latch->mutex, 1) )
+		return 0;
+#endif
+	//	take write access if all bits are clear
+
+	if( prev = !(latch->exclusive | latch->share) )
+		latch->exclusive = 1;
+
+#ifdef unix
+	*latch->mutex = 0;
+#else
+	_InterlockedExchange8(latch->mutex, 0);
+#endif
+	return prev;
+}
+
+//	clear write mode
+
+void bt_spinreleasewrite(BtSpinLatch *latch)
+{
+#ifdef unix
+	while( __sync_lock_test_and_set(latch->mutex, 1) )
+		sched_yield();
+#else
+	while( _InterlockedExchange8(latch->mutex, 1) )
+		SwitchToThread();
+#endif
+	latch->exclusive = 0;
+#ifdef unix
+	*latch->mutex = 0;
+#else
+	_InterlockedExchange8(latch->mutex, 0);
+#endif
+}
+
+//	decrement reader count
+
+void bt_spinreleaseread(BtSpinLatch *latch)
+{
+#ifdef unix
+	while( __sync_lock_test_and_set(latch->mutex, 1) )
+		sched_yield();
+#else
+	while( _InterlockedExchange8(latch->mutex, 1) )
+		SwitchToThread();
+#endif
+	latch->share--;
+#ifdef unix
+	*latch->mutex = 0;
+#else
+	_InterlockedExchange8(latch->mutex, 0);
+#endif
+}
+
+//	link latch table entry into latch hash table
+
+void bt_latchlink (BtDb *bt, ushort hashidx, ushort victim, uid page_no)
+{
+BtLatchSet *set = bt->mgr->latchsets + victim;
+
+	if( set->next = bt->mgr->latchmgr->table[hashidx].slot )
+		bt->mgr->latchsets[set->next].prev = victim;
+
+	bt->mgr->latchmgr->table[hashidx].slot = victim;
+	set->page_no = page_no;
+	set->hash = hashidx;
+	set->prev = 0;
+}
+
+//	release latch pin
+
+void bt_unpinlatch (BtLatchSet *set)
+{
+#ifdef unix
+	__sync_fetch_and_add(&set->pin, -1);
+#else
+	_InterlockedDecrement16 (&set->pin);
+#endif
+}
+
+//	find existing latchset or inspire new one
+//	return with latchset pinned
+
+BtLatchSet *bt_pinlatch (BtDb *bt, uid page_no)
+{
+ushort hashidx = page_no % bt->mgr->latchmgr->latchhash;
+ushort slot, avail = 0, victim, idx;
+BtLatchSet *set;
+
+	//  obtain read lock on hash table entry
+
+	bt_spinreadlock(bt->mgr->latchmgr->table[hashidx].latch);
+
+	if( slot = bt->mgr->latchmgr->table[hashidx].slot ) do
+	{
+		set = bt->mgr->latchsets + slot;
+		if( page_no == set->page_no )
+			break;
+	} while( slot = set->next );
+
+	if( slot ) {
+#ifdef unix
+		__sync_fetch_and_add(&set->pin, 1);
+#else
+		_InterlockedIncrement16 (&set->pin);
+#endif
+	}
+
+    bt_spinreleaseread (bt->mgr->latchmgr->table[hashidx].latch);
+
+	if( slot )
+		return set;
+
+  //  try again, this time with write lock
+
+  bt_spinwritelock(bt->mgr->latchmgr->table[hashidx].latch);
+
+  if( slot = bt->mgr->latchmgr->table[hashidx].slot ) do
+  {
+	set = bt->mgr->latchsets + slot;
+	if( page_no == set->page_no )
+		break;
+	if( !set->pin && !avail )
+		avail = slot;
+  } while( slot = set->next );
+
+  //  found our entry, or take over an unpinned one
+
+  if( slot || (slot = avail) ) {
+	set = bt->mgr->latchsets + slot;
+#ifdef unix
+	__sync_fetch_and_add(&set->pin, 1);
+#else
+	_InterlockedIncrement16 (&set->pin);
+#endif
+	set->page_no = page_no;
+	bt_spinreleasewrite(bt->mgr->latchmgr->table[hashidx].latch);
+	return set;
+  }
+
+	//  see if there are any unused entries
+#ifdef unix
+	victim = __sync_fetch_and_add (&bt->mgr->latchmgr->latchdeployed, 1) + 1;
+#else
+	victim = _InterlockedIncrement16 (&bt->mgr->latchmgr->latchdeployed);
+#endif
+
+	if( victim < bt->mgr->latchmgr->latchtotal ) {
+		set = bt->mgr->latchsets + victim;
+#ifdef unix
+		__sync_fetch_and_add(&set->pin, 1);
+#else
+		_InterlockedIncrement16 (&set->pin);
+#endif
+		bt_latchlink (bt, hashidx, victim, page_no);
+		bt_spinreleasewrite (bt->mgr->latchmgr->table[hashidx].latch);
+		return set;
+	}
+
+#ifdef unix
+	victim = __sync_fetch_and_add (&bt->mgr->latchmgr->latchdeployed, -1);
+#else
+	victim = _InterlockedDecrement16 (&bt->mgr->latchmgr->latchdeployed);
+#endif
+  //  find and reuse previous lock entry
+
+  while( 1 ) {
+#ifdef unix
+	victim = __sync_fetch_and_add(&bt->mgr->latchmgr->latchvictim, 1);
+#else
+	victim = _InterlockedIncrement16 (&bt->mgr->latchmgr->latchvictim) - 1;
+#endif
+	//	we don't use slot zero
+
+	if( victim %= bt->mgr->latchmgr->latchtotal )
+		set = bt->mgr->latchsets + victim;
+	else
+		continue;
+
+	//	take control of our slot
+	//	from other threads
+
+	if( set->pin || !bt_spinwritetry (set->busy) )
+		continue;
+
+	idx = set->hash;
+
+	// try to get write lock on hash chain
+	//	skip entry if not obtained
+	//	or has outstanding locks
+
+	if( !bt_spinwritetry (bt->mgr->latchmgr->table[idx].latch) ) {
+		bt_spinreleasewrite (set->busy);
+		continue;
+	}
+
+	if( set->pin ) {
+		bt_spinreleasewrite (set->busy);
+		bt_spinreleasewrite (bt->mgr->latchmgr->table[idx].latch);
+		continue;
+	}
+
+	//  unlink our available victim from its hash chain
+
+	if( set->prev )
+		bt->mgr->latchsets[set->prev].next = set->next;
+	else
+		bt->mgr->latchmgr->table[idx].slot = set->next;
+
+	if( set->next )
+		bt->mgr->latchsets[set->next].prev = set->prev;
+
+	bt_spinreleasewrite (bt->mgr->latchmgr->table[idx].latch);
+#ifdef unix
+	__sync_fetch_and_add(&set->pin, 1);
+#else
+	_InterlockedIncrement16 (&set->pin);
+#endif
+	bt_latchlink (bt, hashidx, victim, page_no);
+	bt_spinreleasewrite (bt->mgr->latchmgr->table[hashidx].latch);
+	bt_spinreleasewrite (set->busy);
+	return set;
+  }
+}
+
 void bt_mgrclose (BtMgr *mgr)
 {
 BtPool *pool;
@@ -346,19 +685,26 @@ uint slot;
 #endif
 	}
 
+#ifdef unix
+	munmap (mgr->latchsets, mgr->latchmgr->nlatchpage * mgr->page_size);
+	munmap (mgr->latchmgr, mgr->page_size);
+#else
+	FlushViewOfFile(mgr->latchmgr, 0);
+	UnmapViewOfFile(mgr->latchmgr);
+	CloseHandle(mgr->halloc);
+#endif
 #ifdef unix
 	close (mgr->idx);
 	free (mgr->pool);
 	free (mgr->hash);
-	free (mgr->latch);
-	free (mgr->pooladvise);
+	free ((void *)mgr->latch);
 	free (mgr);
 #else
 	FlushFileBuffers(mgr->idx);
 	CloseHandle(mgr->idx);
 	GlobalFree (mgr->pool);
 	GlobalFree (mgr->hash);
-	GlobalFree (mgr->latch);
+	GlobalFree ((void *)mgr->latch);
 	GlobalFree (mgr);
 #endif
 }
@@ -368,10 +714,10 @@ uint slot;
 void bt_close (BtDb *bt)
 {
 #ifdef unix
-	if ( bt->mem )
+	if( bt->mem )
 		free (bt->mem);
 #else
-	if ( bt->mem)
+	if( bt->mem)
 		VirtualFree (bt->mem, 0, MEM_RELEASE);
 #endif
 	free (bt);
@@ -384,13 +730,14 @@ void bt_close (BtDb *bt)
 
 BtMgr *bt_mgr (char *name, uint mode, uint bits, uint poolmax, uint segsize, uint hashsize)
 {
-uint lvl, attr, cacheblk, last;
-BtPage alloc;
-int lockmode;
+uint lvl, attr, cacheblk, last, slot, idx;
+uint nlatchpage, latchhash;
+BtLatchMgr *latchmgr;
 off64_t size;
 uint amt[1];
 BtMgr* mgr;
 BtKey key;
+int flag;
 
 #ifndef unix
 SYSTEM_INFO sysinfo[1];
@@ -409,19 +756,8 @@ SYSTEM_INFO sysinfo[1];
 #ifdef unix
 	mgr = calloc (1, sizeof(BtMgr));
 
-	switch (mode & 0x7fff)
-	{
-	case BT_rw:
-		mgr->idx = open ((char*)name, O_RDWR | O_CREAT, 0666);
-		lockmode = 1;
-		break;
+	mgr->idx = open ((char*)name, O_RDWR | O_CREAT, 0666);
 
-	case BT_ro:
-	default:
-		mgr->idx = open ((char*)name, O_RDONLY);
-		lockmode = 0;
-		break;
-	}
 	if( mgr->idx == -1 )
 		return free(mgr), NULL;
 	
@@ -430,19 +766,8 @@ SYSTEM_INFO sysinfo[1];
 #else
 	mgr = GlobalAlloc (GMEM_FIXED|GMEM_ZEROINIT, sizeof(BtMgr));
 	attr = FILE_ATTRIBUTE_NORMAL;
-	switch (mode & 0x7fff)
-	{
-	case BT_rw:
-		mgr->idx = CreateFile(name, GENERIC_READ| GENERIC_WRITE, FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, OPEN_ALWAYS, attr, NULL);
-		lockmode = 1;
-		break;
+	mgr->idx = CreateFile(name, GENERIC_READ| GENERIC_WRITE, FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, OPEN_ALWAYS, attr, NULL);
 
-	case BT_ro:
-	default:
-		mgr->idx = CreateFile(name, GENERIC_READ, FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, OPEN_EXISTING, attr, NULL);
-		lockmode = 0;
-		break;
-	}
 	if( mgr->idx == INVALID_HANDLE_VALUE )
 		return GlobalFree(mgr), NULL;
 
@@ -452,26 +777,26 @@ SYSTEM_INFO sysinfo[1];
 #endif
 
 #ifdef unix
-	alloc = malloc (BT_maxpage);
+	latchmgr = malloc (BT_maxpage);
 	*amt = 0;
 
 	// read minimum page size to get root info
 
 	if( size = lseek (mgr->idx, 0L, 2) ) {
-		if( pread(mgr->idx, alloc, BT_minpage, 0) == BT_minpage )
-			bits = alloc->bits;
+		if( pread(mgr->idx, latchmgr, BT_minpage, 0) == BT_minpage )
+			bits = latchmgr->alloc->bits;
 		else
-			return free(mgr), free(alloc), NULL;
+			return free(mgr), free(latchmgr), NULL;
 	} else if( mode == BT_ro )
-		return bt_mgrclose (mgr), NULL;
+		return free(latchmgr), bt_mgrclose (mgr), NULL;
 #else
-	alloc = VirtualAlloc(NULL, BT_maxpage, MEM_COMMIT, PAGE_READWRITE);
+	latchmgr = VirtualAlloc(NULL, BT_maxpage, MEM_COMMIT, PAGE_READWRITE);
 	size = GetFileSize(mgr->idx, amt);
 
 	if( size || *amt ) {
-		if( !ReadFile(mgr->idx, (char *)alloc, BT_minpage, amt, NULL) )
+		if( !ReadFile(mgr->idx, (char *)latchmgr, BT_minpage, amt, NULL) )
 			return bt_mgrclose (mgr), NULL;
-		bits = alloc->bits;
+		bits = latchmgr->alloc->bits;
 	} else if( mode == BT_ro )
 		return bt_mgrclose (mgr), NULL;
 #endif
@@ -504,53 +829,68 @@ SYSTEM_INFO sysinfo[1];
 #ifdef unix
 	mgr->pool = calloc (poolmax, sizeof(BtPool));
 	mgr->hash = calloc (hashsize, sizeof(ushort));
-	mgr->latch = calloc (hashsize, sizeof(BtLatch));
-	mgr->pooladvise = calloc (poolmax, (mgr->poolmask + 8) / 8);
+	mgr->latch = calloc (hashsize, sizeof(BtSpinLatch));
 #else
 	mgr->pool = GlobalAlloc (GMEM_FIXED|GMEM_ZEROINIT, poolmax * sizeof(BtPool));
 	mgr->hash = GlobalAlloc (GMEM_FIXED|GMEM_ZEROINIT, hashsize * sizeof(ushort));
-	mgr->latch = GlobalAlloc (GMEM_FIXED|GMEM_ZEROINIT, hashsize * sizeof(BtLatch));
+	mgr->latch = GlobalAlloc (GMEM_FIXED|GMEM_ZEROINIT, hashsize * sizeof(BtSpinLatch));
 #endif
 
 	if( size || *amt )
-		goto mgrxit;
+		goto mgrlatch;
+
+	// initialize an empty b-tree with latch page, root page, page of leaves
+	// and page(s) of latches
+
+	memset (latchmgr, 0, 1 << bits);
+	nlatchpage = BT_latchtable / (mgr->page_size / sizeof(BtLatchSet)) + 1; 
+	bt_putid(latchmgr->alloc->right, MIN_lvl+1+nlatchpage);
+	latchmgr->alloc->bits = mgr->page_bits;
+
+	latchmgr->nlatchpage = nlatchpage;
+	latchmgr->latchtotal = nlatchpage * (mgr->page_size / sizeof(BtLatchSet));
 
-	// initializes an empty b-tree with root page and page of leaves
+	//  initialize latch manager
 
-	memset (alloc, 0, 1 << bits);
-	bt_putid(alloc->right, MIN_lvl+1);
-	alloc->bits = mgr->page_bits;
+	latchhash = (mgr->page_size - sizeof(BtLatchMgr)) / sizeof(BtHashEntry);
+
+	//	size of hash table = total number of latchsets
+
+	if( latchhash > latchmgr->latchtotal )
+		latchhash = latchmgr->latchtotal;
+
+	latchmgr->latchhash = latchhash;
 
 #ifdef unix
-	if( write (mgr->idx, alloc, mgr->page_size) < mgr->page_size )
+	if( write (mgr->idx, latchmgr, mgr->page_size) < mgr->page_size )
 		return bt_mgrclose (mgr), NULL;
 #else
-	if( !WriteFile (mgr->idx, (char *)alloc, mgr->page_size, amt, NULL) )
+	if( !WriteFile (mgr->idx, (char *)latchmgr, mgr->page_size, amt, NULL) )
 		return bt_mgrclose (mgr), NULL;
 
 	if( *amt < mgr->page_size )
 		return bt_mgrclose (mgr), NULL;
 #endif
 
-	memset (alloc, 0, 1 << bits);
-	alloc->bits = mgr->page_bits;
+	memset (latchmgr, 0, 1 << bits);
+	latchmgr->alloc->bits = mgr->page_bits;
 
 	for( lvl=MIN_lvl; lvl--; ) {
-		slotptr(alloc, 1)->off = mgr->page_size - 3;
-		bt_putid(slotptr(alloc, 1)->id, lvl ? MIN_lvl - lvl + 1 : 0);		// next(lower) page number
-		key = keyptr(alloc, 1);
-		key->len = 2;			// create stopper key
+		slotptr(latchmgr->alloc, 1)->off = mgr->page_size - 3;
+		bt_putid(slotptr(latchmgr->alloc, 1)->id, lvl ? MIN_lvl - lvl + 1 : 0);		// next(lower) page number
+		key = keyptr(latchmgr->alloc, 1);
+		key->len = 2;		// create stopper key
 		key->key[0] = 0xff;
 		key->key[1] = 0xff;
-		alloc->min = mgr->page_size - 3;
-		alloc->lvl = lvl;
-		alloc->cnt = 1;
-		alloc->act = 1;
+		latchmgr->alloc->min = mgr->page_size - 3;
+		latchmgr->alloc->lvl = lvl;
+		latchmgr->alloc->cnt = 1;
+		latchmgr->alloc->act = 1;
 #ifdef unix
-		if( write (mgr->idx, alloc, mgr->page_size) < mgr->page_size )
+		if( write (mgr->idx, latchmgr, mgr->page_size) < mgr->page_size )
 			return bt_mgrclose (mgr), NULL;
 #else
-		if( !WriteFile (mgr->idx, (char *)alloc, mgr->page_size, amt, NULL) )
+		if( !WriteFile (mgr->idx, (char *)latchmgr, mgr->page_size, amt, NULL) )
 			return bt_mgrclose (mgr), NULL;
 
 		if( *amt < mgr->page_size )
@@ -558,32 +898,52 @@ SYSTEM_INFO sysinfo[1];
 #endif
 	}
 
-	// create empty page area by writing last page of first
-	// segment area (other pages are zeroed by O/S)
-
-	if( mgr->poolmask ) {
-		memset(alloc, 0, mgr->page_size);
-		last = mgr->poolmask;
+	// clear out latch manager locks
+	//	and rest of pages to round out segment
 
-		while( last < MIN_lvl + 1 )
-			last += mgr->poolmask + 1;
+	memset(latchmgr, 0, mgr->page_size);
+	last = MIN_lvl + 1;
 
+	while( last <= ((MIN_lvl + 1 + nlatchpage) | mgr->poolmask) ) {
 #ifdef unix
-		pwrite(mgr->idx, alloc, mgr->page_size, last << mgr->page_bits);
+		pwrite(mgr->idx, latchmgr, mgr->page_size, last << mgr->page_bits);
 #else
 		SetFilePointer (mgr->idx, last << mgr->page_bits, NULL, FILE_BEGIN);
-		if( !WriteFile (mgr->idx, (char *)alloc, mgr->page_size, amt, NULL) )
+		if( !WriteFile (mgr->idx, (char *)latchmgr, mgr->page_size, amt, NULL) )
 			return bt_mgrclose (mgr), NULL;
 		if( *amt < mgr->page_size )
 			return bt_mgrclose (mgr), NULL;
 #endif
+		last++;
 	}
 
-mgrxit:
+mgrlatch:
+#ifdef unix
+	flag = PROT_READ | PROT_WRITE;
+	mgr->latchmgr = mmap (0, mgr->page_size, flag, MAP_SHARED, mgr->idx, ALLOC_page * mgr->page_size);
+	if( mgr->latchmgr == MAP_FAILED )
+		return bt_mgrclose (mgr), NULL;
+	mgr->latchsets = (BtLatchSet *)mmap (0, mgr->latchmgr->nlatchpage * mgr->page_size, flag, MAP_SHARED, mgr->idx, LATCH_page * mgr->page_size);
+	if( mgr->latchsets == MAP_FAILED )
+		return bt_mgrclose (mgr), NULL;
+#else
+	flag = PAGE_READWRITE;
+	mgr->halloc = CreateFileMapping(mgr->idx, NULL, flag, 0, (BT_latchtable / (mgr->page_size / sizeof(BtLatchSet)) + 1 + LATCH_page) * mgr->page_size, NULL);
+	if( !mgr->halloc )
+		return bt_mgrclose (mgr), NULL;
+
+	flag = FILE_MAP_WRITE;
+	mgr->latchmgr = MapViewOfFile(mgr->halloc, flag, 0, 0, (BT_latchtable / (mgr->page_size / sizeof(BtLatchSet)) + 1 + LATCH_page) * mgr->page_size);
+	if( !mgr->latchmgr )
+		return GetLastError(), bt_mgrclose (mgr), NULL;
+
+	mgr->latchsets = (void *)((char *)mgr->latchmgr + LATCH_page * mgr->page_size);
+#endif
+
 #ifdef unix
-	free (alloc);
+	free (latchmgr);
 #else
-	VirtualFree (alloc, 0, MEM_RELEASE);
+	VirtualFree (latchmgr, 0, MEM_RELEASE);
 #endif
 	return mgr;
 }
@@ -629,167 +989,6 @@ int ans;
 	return 0;
 }
 
-//	Latch Manager
-
-int sys_futex(void *addr1, int op, int val1, struct timespec *timeout, void *addr2, int val3)
-{
-	return syscall(SYS_futex, addr1, op, val1, timeout, addr2, val3);
-}
-
-//	wait until write lock mode is clear
-//	and add 1 to the share count
-
-void bt_readlock(BtLatch *latch, int private)
-{
-uint prev;
-
-  if( private )
-	private = FUTEX_PRIVATE_FLAG;
-
-  while( 1 ) {
-	//	obtain latch mutex
-	if( __sync_fetch_and_or((uint *)latch, Mutex) & Mutex ) {
-		sched_yield();
-		continue;
-	}
-
-	//  wait for writers to clear
-	//	increment read waiters and wait
-
-	if( latch->write || latch->writewait ) {
-		__sync_fetch_and_add ((uint *)latch, PendRd);
-		prev = __sync_fetch_and_and ((uint *)latch, ~Mutex) & ~Mutex;
-		sys_futex( (uint *)latch, FUTEX_WAIT_BITSET | private, prev, NULL, NULL, QueRd );
-		__sync_fetch_and_sub ((uint *)latch, PendRd);
-		continue;
-	}
-	
-	// increment reader lock count
-	// and release latch mutex
-
-	__sync_fetch_and_add ((uint *)latch, Share);
-	__sync_fetch_and_and ((uint *)latch, ~Mutex);
-	return;
-  }
-}
-
-//	wait for other read and write latches to relinquish
-
-void bt_writelock(BtLatch *latch, int private)
-{
-uint prev;
-
-  if( private )
-	private = FUTEX_PRIVATE_FLAG;
-
-  while( 1 ) {
-	//	obtain latch mutex
-	if( __sync_fetch_and_or((uint *)latch, Mutex) & Mutex ) {
-		sched_yield();
-		continue;
-	}
-
-	//	wait for write and reader count to clear
-
-	if( latch->write || latch->share ) {
-		__sync_fetch_and_add ((uint *)latch, PendWr);
-		prev = __sync_fetch_and_and ((uint *)latch, ~Mutex) & ~Mutex;
-		sys_futex( (uint *)latch, FUTEX_WAIT_BITSET | private, prev, NULL, NULL, QueWr );
-		__sync_fetch_and_sub ((uint *)latch, PendWr);
-		continue;
-	}
-	
-	//	take write mutex
-	//	release latch mutex
-
-	__sync_fetch_and_or ((uint *)latch, Write);
-	__sync_fetch_and_and ((uint *)latch, ~Mutex);
-	return;
-  }
-}
-
-//	try to obtain write lock
-
-//	return 1 if obtained,
-//		0 otherwise
-
-int bt_writetry(BtLatch *latch)
-{
-int ans;
-
-	//	try for mutex,
-	//	abandon request if not taken
-
-	if( __sync_fetch_and_or((uint *)latch, Mutex) & Mutex )
-		return 0;
-
-	//	see if write mode is available
-
-	if( !latch->write && !latch->share ) {
-		__sync_fetch_and_or ((uint *)latch, Write);
-		ans = 1;
-	} else
-		ans = 0;
-
-	// release latch mutex
-
-	__sync_fetch_and_and ((uint *)latch, ~Mutex);
-	return ans;
-}
-
-//	clear write lock
-
-void bt_releasewrite(BtLatch *latch, int private)
-{
-  if( private )
-	private = FUTEX_PRIVATE_FLAG;
-
-	//	obtain latch mutex
-
-	while( __sync_fetch_and_or((uint *)latch, Mutex) & Mutex )
-		sched_yield();
-
-	__sync_fetch_and_and ((uint *)latch, ~Write);
-
-	// favor writers
-
-	if( latch->writewait )
-	  if( sys_futex( (uint *)latch, FUTEX_WAKE_BITSET | private, 1, NULL, NULL, QueWr ) )
-		goto wakexit;
-
-	if( latch->readwait )
-		sys_futex( (uint *)latch, FUTEX_WAKE_BITSET | private, INT_MAX, NULL, NULL, QueRd );
-
-	// release latch mutex
-
-wakexit:
-	__sync_fetch_and_and ((uint *)latch, ~Mutex);
-}
-
-//	decrement reader count
-
-void bt_releaseread(BtLatch *latch, int private)
-{
-  if( private )
-	private = FUTEX_PRIVATE_FLAG;
-
-	//	obtain latch mutex
-
-	while( __sync_fetch_and_or((uint *)latch, Mutex) & Mutex )
-		sched_yield();
-
-	__sync_fetch_and_sub ((uint *)latch, Share);
-
-	// wake waiting writers
-
-	if( !latch->share && latch->writewait )
-		sys_futex( (uint *)latch, FUTEX_WAKE_BITSET | private, 1, NULL, NULL, QueWr );
-
-	// release latch mutex
-
-	__sync_fetch_and_and ((uint *)latch, ~Mutex);
-}
-
 //	Buffer Pool mgr
 
 // find segment in pool
@@ -829,7 +1028,7 @@ uint slot;
 
 	pool->hashprev = pool->hashnext = NULL;
 	pool->basepage = page_no & ~bt->mgr->poolmask;
-	pool->lru = 1;
+	pool->pin = CLOCK_bit + 1;
 
 	if( slot = bt->mgr->hash[idx] ) {
 		node = bt->mgr->pool + slot;
@@ -840,32 +1039,6 @@ uint slot;
 	bt->mgr->hash[idx] = pool->slot;
 }
 
-//	find best segment to evict from buffer pool
-
-BtPool *bt_findlru (BtDb *bt, uint hashslot)
-{
-unsigned long long int target = ~0LL;
-BtPool *pool = NULL, *node;
-
-	if( !hashslot )
-		return NULL;
-
-	node = bt->mgr->pool + hashslot;
-
-	//  scan pool entries under hash table slot
-
-	do {
-	  if( node->pin )
-		continue;
-	  if( node->lru > target )
-		continue;
-	  target = node->lru;
-	  pool = node;
-	} while( node = node->hashnext );
-
-	return pool;
-}
-
 //  map new buffer pool segment to virtual memory
 
 BTERR bt_mapsegment(BtDb *bt, BtPool *pool, uid page_no)
@@ -880,8 +1053,6 @@ int flag;
 	if( pool->map == MAP_FAILED )
 		return bt->err = BTERR_map;
 
-	// clear out madvise issued bits
-	memset (bt->mgr->pooladvise + pool->slot * ((bt->mgr->poolmask + 8) / 8), 0, (bt->mgr->poolmask + 8)/8);
 #else
 	flag = ( bt->mgr->mode == BT_ro ? PAGE_READONLY : PAGE_READWRITE );
 	pool->hmap = CreateFileMapping(bt->mgr->idx, NULL, flag, (DWORD)(limit >> 32), (DWORD)limit, NULL);
@@ -896,47 +1067,52 @@ int flag;
  	return bt->err = 0;
 }
 
-//	find or place requested page in segment-pool
-//	return pool table entry, incrementing pin
+//	calculate page within pool
 
-BtPool *bt_pinpage(BtDb *bt, uid page_no)
+BtPage bt_page (BtDb *bt, BtPool *pool, uid page_no)
 {
-BtPool *pool, *node, *next;
-uint slot, idx, victim;
-
-	//	lock hash table chain
+uint subpage = (uint)(page_no & bt->mgr->poolmask); // page within mapping
+BtPage page;
 
-	idx = (uint)(page_no >> bt->mgr->seg_bits) % bt->mgr->hashsize;
-	bt_readlock (&bt->mgr->latch[idx], 1);
+	page = (BtPage)(pool->map + (subpage << bt->mgr->page_bits));
+	return page;
+}
 
-	//	look up in hash table
+//  release pool pin
 
-	if( pool = bt_findpool(bt, page_no, idx) ) {
+void bt_unpinpool (BtPool *pool)
+{
 #ifdef unix
-		__sync_fetch_and_add(&pool->pin, 1);
+	__sync_fetch_and_add(&pool->pin, -1);
 #else
-		_InterlockedIncrement (&pool->pin);
+	_InterlockedDecrement16 (&pool->pin);
 #endif
-		bt_releaseread (&bt->mgr->latch[idx], 1);
-		pool->lru++;
-		return pool;
-	}
+}
 
-	//	upgrade to write lock
+//	find or place requested page in segment-pool
+//	return pool table entry, incrementing pin
 
-	bt_releaseread (&bt->mgr->latch[idx], 1);
-	bt_writelock (&bt->mgr->latch[idx], 1);
+BtPool *bt_pinpool(BtDb *bt, uid page_no)
+{
+uint slot, hashidx, idx, victim;
+BtPool *pool, *node, *next;
 
-	// try to find page in pool with write lock
+	//	lock hash table chain
+
+	hashidx = (uint)(page_no >> bt->mgr->seg_bits) % bt->mgr->hashsize;
+	bt_spinwritelock (&bt->mgr->latch[hashidx]);
+
+	//	look up in hash table
 
-	if( pool = bt_findpool(bt, page_no, idx) ) {
+	if( pool = bt_findpool(bt, page_no, hashidx) ) {
 #ifdef unix
+		__sync_fetch_and_or(&pool->pin, CLOCK_bit);
 		__sync_fetch_and_add(&pool->pin, 1);
 #else
-		_InterlockedIncrement (&pool->pin);
+		_InterlockedOr16 (&pool->pin, CLOCK_bit);
+		_InterlockedIncrement16 (&pool->pin);
 #endif
-		bt_releasewrite (&bt->mgr->latch[idx], 1);
-		pool->lru++;
+		bt_spinreleasewrite (&bt->mgr->latch[hashidx]);
 		return pool;
 	}
 
@@ -946,7 +1122,7 @@ uint slot, idx, victim;
 #ifdef unix
 	slot = __sync_fetch_and_add(&bt->mgr->poolcnt, 1);
 #else
-	slot = _InterlockedIncrement (&bt->mgr->poolcnt) - 1;
+	slot = _InterlockedIncrement16 (&bt->mgr->poolcnt) - 1;
 #endif
 
 	if( ++slot < bt->mgr->poolmax ) {
@@ -956,13 +1132,8 @@ uint slot, idx, victim;
 		if( bt_mapsegment(bt, pool, page_no) )
 			return NULL;
 
-		bt_linkhash(bt, pool, page_no, idx);
-#ifdef unix
-		__sync_fetch_and_add(&pool->pin, 1);
-#else
-		_InterlockedIncrement (&pool->pin);
-#endif
-		bt_releasewrite (&bt->mgr->latch[idx], 1);
+		bt_linkhash(bt, pool, page_no, hashidx);
+		bt_spinreleasewrite (&bt->mgr->latch[hashidx]);
 		return pool;
 	}
 
@@ -972,7 +1143,7 @@ uint slot, idx, victim;
 #ifdef unix
 	__sync_fetch_and_add(&bt->mgr->poolcnt, -1);
 #else
-	_InterlockedDecrement (&bt->mgr->poolcnt);
+	_InterlockedDecrement16 (&bt->mgr->poolcnt);
 #endif
 
 	while( 1 ) {
@@ -981,20 +1152,30 @@ uint slot, idx, victim;
 #else
 		victim = _InterlockedIncrement (&bt->mgr->evicted) - 1;
 #endif
-		victim %= bt->mgr->hashsize;
+		victim %= bt->mgr->poolmax;
+		pool = bt->mgr->pool + victim;
+		idx = (uint)(pool->basepage >> bt->mgr->seg_bits) % bt->mgr->hashsize;
+
+		if( !victim )
+			continue;
 
 		// try to get write lock
 		//	skip entry if not obtained
 
-		if( !bt_writetry (&bt->mgr->latch[victim]) )
+		if( !bt_spinwritetry (&bt->mgr->latch[idx]) )
 			continue;
 
-		//  if pool entry is empty
-		//	or any pages are pinned
-		//	skip this entry
+		//	skip this entry if
+		//	page is pinned
+		//  or clock bit is set
 
-		if( !(pool = bt_findlru(bt, bt->mgr->hash[victim])) ) {
-			bt_releasewrite (&bt->mgr->latch[victim], 1);
+		if( pool->pin ) {
+#ifdef unix
+			__sync_fetch_and_and(&pool->pin, ~CLOCK_bit);
+#else
+			_InterlockedAnd16 (&pool->pin, ~CLOCK_bit);
+#endif
+			bt_spinreleasewrite (&bt->mgr->latch[idx]);
 			continue;
 		}
 
@@ -1003,14 +1184,14 @@ uint slot, idx, victim;
 		if( node = pool->hashprev )
 			node->hashnext = pool->hashnext;
 		else if( node = pool->hashnext )
-			bt->mgr->hash[victim] = node->slot;
+			bt->mgr->hash[idx] = node->slot;
 		else
-			bt->mgr->hash[victim] = 0;
+			bt->mgr->hash[idx] = 0;
 
 		if( node = pool->hashnext )
 			node->hashprev = pool->hashprev;
 
-		bt_releasewrite (&bt->mgr->latch[victim], 1);
+		bt_spinreleasewrite (&bt->mgr->latch[idx]);
 
 		//	remove old file mapping
 #ifdef unix
@@ -1028,244 +1209,140 @@ uint slot, idx, victim;
 		if( bt_mapsegment(bt, pool, page_no) )
 			return NULL;
 
-		bt_linkhash(bt, pool, page_no, idx);
-#ifdef unix
-		__sync_fetch_and_add(&pool->pin, 1);
-#else
-		_InterlockedIncrement (&pool->pin);
-#endif
-		bt_releasewrite (&bt->mgr->latch[idx], 1);
+		bt_linkhash(bt, pool, page_no, hashidx);
+		bt_spinreleasewrite (&bt->mgr->latch[hashidx]);
 		return pool;
 	}
 }
 
 // place write, read, or parent lock on requested page_no.
-//	pin to buffer pool and return page pointer
 
-BTERR bt_lockpage(BtDb *bt, uid page_no, BtLock mode, BtPage *pageptr)
+void bt_lockpage(BtLock mode, BtLatchSet *set)
 {
-uint subpage;
-BtPool *pool;
-BtPage page;
-
-	//	find/create maping in pool table
-	//	  and pin our pool slot
-
-	if( pool = bt_pinpage(bt, page_no) )
-		subpage = (uint)(page_no & bt->mgr->poolmask); // page within mapping
-	else
-		return bt->err;
-
-	page = (BtPage)(pool->map + (subpage << bt->mgr->page_bits));
-#ifdef unix
-	{
-	uint idx = subpage / 8;
-	uint bit = subpage % 8;
-
-		if( ~((bt->mgr->pooladvise + pool->slot * ((bt->mgr->poolmask + 8)/8))[idx] >> bit) & 1 ) {
-		  madvise (page, bt->mgr->page_size, MADV_WILLNEED);
-		  (bt->mgr->pooladvise + pool->slot * ((bt->mgr->poolmask + 8)/8))[idx] |= 1 << bit;
-		}
-	}
-#endif
-
 	switch( mode ) {
 	case BtLockRead:
-		bt_readlock (page->latch->readwr, 0);
+		bt_spinreadlock (set->readwr);
 		break;
 	case BtLockWrite:
-		bt_writelock (page->latch->readwr, 0);
+		bt_spinwritelock (set->readwr);
 		break;
 	case BtLockAccess:
-		bt_readlock (page->latch->access, 0);
+		bt_spinreadlock (set->access);
 		break;
 	case BtLockDelete:
-		bt_writelock (page->latch->access, 0);
+		bt_spinwritelock (set->access);
 		break;
 	case BtLockParent:
-		bt_writelock (page->latch->parent, 0);
-		break;
-	default:
-		return bt->err = BTERR_lock;
-	}
-
-	if( pageptr )
-		*pageptr = page;
-	return bt->err = 0;
-}
-
-// remove write, read, or parent lock on requested page
-
-BTERR bt_unlockpage(BtDb *bt, uid page_no, BtLock mode)
-{
-uint subpage, idx;
-BtPool *pool;
-BtPage page;
-
-	//	since page is pinned
-	//	it should still be in the buffer pool
-	//	and is in no danger of being a victim for reuse
-
-	idx = (uint)(page_no >> bt->mgr->seg_bits) % bt->mgr->hashsize;
-	bt_readlock (&bt->mgr->latch[idx], 1);
-
-	if( pool = bt_findpool(bt, page_no, idx) )
-		subpage = (uint)(page_no & bt->mgr->poolmask);
-	else
-		return bt->err = BTERR_hash;
-
-	bt_releaseread (&bt->mgr->latch[idx], 1);
-	page = (BtPage)(pool->map + (subpage << bt->mgr->page_bits));
-
-	switch( mode ) {
-	case BtLockRead:
-		bt_releaseread (page->latch->readwr, 0);
-		break;
-	case BtLockWrite:
-		bt_releasewrite (page->latch->readwr, 0);
-		break;
-	case BtLockAccess:
-		bt_releaseread (page->latch->access, 0);
-		break;
-	case BtLockDelete:
-		bt_releasewrite (page->latch->access, 0);
-		break;
-	case BtLockParent:
-		bt_releasewrite (page->latch->parent, 0);
-		break;
-	default:
-		return bt->err = BTERR_lock;
-	}
-
-#ifdef  unix
-	__sync_fetch_and_add(&pool->pin, -1);
-#else
-	_InterlockedDecrement (&pool->pin);
-#endif
-	return bt->err = 0;
-}
-
-//	deallocate a deleted page 
-//	place on free chain out of allocator page
-
-BTERR bt_freepage(BtDb *bt, uid page_no)
-{
-	//  obtain delete lock on deleted page
-
-	if( bt_lockpage(bt, page_no, BtLockDelete, NULL) )
-		return bt->err;
-
-	//  obtain write lock on deleted page
-
-	if( bt_lockpage(bt, page_no, BtLockWrite, &bt->temp) )
-		return bt->err;
-
-	//	lock allocation page
-
-	if ( bt_lockpage(bt, ALLOC_page, BtLockWrite, &bt->alloc) )
-		return bt->err;
-
-	//	store chain in second right
-	bt_putid(bt->temp->right, bt_getid(bt->alloc[1].right));
-	bt_putid(bt->alloc[1].right, page_no);
-
-	// unlock page zero 
-
-	if( bt_unlockpage(bt, ALLOC_page, BtLockWrite) )
-		return bt->err;
-
-	//  remove write lock on deleted node
-
-	if( bt_unlockpage(bt, page_no, BtLockWrite) )
-		return bt->err;
-
-	//  remove delete lock on deleted node
+		bt_spinwritelock (set->parent);
+		break;
+	}
+}
 
-	if( bt_unlockpage(bt, page_no, BtLockDelete) )
-		return bt->err;
+// remove write, read, or parent lock on requested page
 
-	return 0;
+void bt_unlockpage(BtLock mode, BtLatchSet *set)
+{
+	switch( mode ) {
+	case BtLockRead:
+		bt_spinreleaseread (set->readwr);
+		break;
+	case BtLockWrite:
+		bt_spinreleasewrite (set->readwr);
+		break;
+	case BtLockAccess:
+		bt_spinreleaseread (set->access);
+		break;
+	case BtLockDelete:
+		bt_spinreleasewrite (set->access);
+		break;
+	case BtLockParent:
+		bt_spinreleasewrite (set->parent);
+		break;
+	}
 }
 
 //	allocate a new page and write page into it
 
 uid bt_newpage(BtDb *bt, BtPage page)
 {
+BtPageSet set[1];
 uid new_page;
-BtPage pmap;
-int subpage;
 int reuse;
 
-	// lock page zero
+	//	lock allocation page
 
-	if( bt_lockpage(bt, ALLOC_page, BtLockWrite, &bt->alloc) )
-		return 0;
+	bt_spinwritelock(bt->mgr->latchmgr->lock);
 
 	// use empty chain first
 	// else allocate empty page
 
-	if( new_page = bt_getid(bt->alloc[1].right) ) {
-		if( bt_lockpage (bt, new_page, BtLockWrite, &bt->temp) )
-			return 0;
-		bt_putid(bt->alloc[1].right, bt_getid(bt->temp->right));
-		if( bt_unlockpage (bt, new_page, BtLockWrite) )
+	if( new_page = bt_getid(bt->mgr->latchmgr->alloc[1].right) ) {
+		if( set->pool = bt_pinpool (bt, new_page) )
+			set->page = bt_page (bt, set->pool, new_page);
+		else
 			return 0;
+
+		bt_putid(bt->mgr->latchmgr->alloc[1].right, bt_getid(set->page->right));
+		bt_unpinpool (set->pool);
 		reuse = 1;
 	} else {
-		new_page = bt_getid(bt->alloc->right);
-		bt_putid(bt->alloc->right, new_page+1);
+		new_page = bt_getid(bt->mgr->latchmgr->alloc->right);
+		bt_putid(bt->mgr->latchmgr->alloc->right, new_page+1);
 		reuse = 0;
 	}
-
 #ifdef unix
-	// if writing first page of pool block
-	//	expand file thru last page in the block
-
-	if( !reuse && (new_page & bt->mgr->poolmask) == 0 )
-	  if( pwrite(bt->mgr->idx, bt->zero, bt->mgr->page_size, (new_page | bt->mgr->poolmask) << bt->mgr->page_bits) < bt->mgr->page_size )
+	if( pwrite(bt->mgr->idx, page, bt->mgr->page_size, new_page << bt->mgr->page_bits) < bt->mgr->page_size )
 		return bt->err = BTERR_wrt, 0;
-#endif
-	// unlock page allocation page 
 
-	if( bt_unlockpage(bt, ALLOC_page, BtLockWrite) )
-		return 0;
+	// if writing first page of pool block, zero last page in the block
 
+	if( !reuse && bt->mgr->poolmask > 0 && (new_page & bt->mgr->poolmask) == 0 )
+	{
+		// use zero buffer to write zeros
+		if( pwrite(bt->mgr->idx,bt->zero, bt->mgr->page_size, (new_page | bt->mgr->poolmask) << bt->mgr->page_bits) < bt->mgr->page_size )
+			return bt->err = BTERR_wrt, 0;
+	}
+#else
 	//	bring new page into pool and copy page.
-	//	on Windows, this will extend the file into the new page.
+	//	this will extend the file into the new pages.
 
-	if( bt_lockpage(bt, new_page, BtLockWrite, &pmap) )
+	if( set->pool = bt_pinpool (bt, new_page) )
+		set->page = bt_page (bt, set->pool, new_page);
+	else
 		return 0;
 
-	//	copy source page but leave latch area intact
-
-	memcpy((char *)pmap + sizeof(BtLatchSet), (char *)page + sizeof(BtLatchSet), bt->mgr->page_size - sizeof(BtLatchSet));
-
-	if( bt_unlockpage (bt, new_page, BtLockWrite) )
-		return 0;
+	memcpy(set->page, page, bt->mgr->page_size);
+	bt_unpinpool (set->pool);
+#endif
+	// unlock allocation latch and return new page no
 
+	bt_spinreleasewrite(bt->mgr->latchmgr->lock);
 	return new_page;
 }
 
 //  find slot in page for given key at a given level
 
-int bt_findslot (BtDb *bt, unsigned char *key, uint len)
+int bt_findslot (BtPageSet *set, unsigned char *key, uint len)
 {
-uint diff, higher = bt->page->cnt, low = 1, slot;
+uint diff, higher = set->page->cnt, low = 1, slot;
 uint good = 0;
 
-	//	make stopper key an infinite fence value
+	//	  make stopper key an infinite fence value
 
-	if( bt_getid (bt->page->right) )
+	if( bt_getid (set->page->right) )
 		higher++;
 	else
 		good++;
 
-	//	low is the next candidate, higher is already
-	//	tested as .ge. the given key, loop ends when they meet
+	//	low is the lowest candidate.
+	//  loop ends when they meet
+
+	//  higher is already
+	//	tested as .ge. the passed key.
 
 	while( diff = higher - low ) {
 		slot = low + ( diff >> 1 );
-		if( keycmp (keyptr(bt->page, slot), key, len) < 0 )
+		if( keycmp (keyptr(set->page, slot), key, len) < 0 )
 			low = slot + 1;
 		else
 			higher = slot, good++;
@@ -1273,92 +1350,105 @@ uint good = 0;
 
 	//	return zero if key is on right link page
 
- 	return good ? higher : 0;
+	return good ? higher : 0;
 }
 
 //  find and load page at given level for given key
 //	leave page rd or wr locked as requested
 
-int bt_loadpage (BtDb *bt, unsigned char *key, uint len, uint lvl, uint lock)
+int bt_loadpage (BtDb *bt, BtPageSet *set, unsigned char *key, uint len, uint lvl, BtLock lock)
 {
 uid page_no = ROOT_page, prevpage = 0;
 uint drill = 0xff, slot;
+BtLatchSet *prevlatch;
 uint mode, prevmode;
+BtPool *prevpool;
 
   //  start at root of btree and drill down
 
   do {
 	// determine lock mode of drill level
-	mode = (lock == BtLockWrite) && (drill == lvl) ? BtLockWrite : BtLockRead; 
+	mode = (drill == lvl) ? lock : BtLockRead; 
+
+	set->latch = bt_pinlatch (bt, page_no);
+	set->page_no = page_no;
 
-	bt->page_no = page_no;
+	// pin page contents
+
+	if( set->pool = bt_pinpool (bt, page_no) )
+		set->page = bt_page (bt, set->pool, page_no);
+	else
+		return 0;
 
  	// obtain access lock using lock chaining with Access mode
 
 	if( page_no > ROOT_page )
-	  if( bt_lockpage(bt, page_no, BtLockAccess, NULL) )
-		return 0;									
+	  bt_lockpage(BtLockAccess, set->latch);
 
-	if( prevpage )
-	  if( bt_unlockpage(bt, prevpage, prevmode) )
-		return 0;
+	//	release & unpin parent page
+
+	if( prevpage ) {
+	  bt_unlockpage(prevmode, prevlatch);
+	  bt_unpinlatch (prevlatch);
+	  bt_unpinpool (prevpool);
+	  prevpage = 0;
+	}
 
  	// obtain read lock using lock chaining
-	// and pin page contents
 
-	if( bt_lockpage(bt, page_no, mode, &bt->page) )
-		return 0;									
+	bt_lockpage(mode, set->latch);
+
+	if( set->page->free )
+		return bt->err = BTERR_struct, 0;
 
 	if( page_no > ROOT_page )
-	  if( bt_unlockpage(bt, page_no, BtLockAccess) )
-		return 0;									
+	  bt_unlockpage(BtLockAccess, set->latch);
 
 	// re-read and re-lock root after determining actual level of root
 
-	if( bt->page->lvl != drill) {
-		if ( bt->page_no != ROOT_page )
+	if( set->page->lvl != drill) {
+		if( set->page_no != ROOT_page )
 			return bt->err = BTERR_struct, 0;
 			
-		drill = bt->page->lvl;
+		drill = set->page->lvl;
 
-		if( lock == BtLockWrite && drill == lvl )
-		  if( bt_unlockpage(bt, page_no, mode) )
-			return 0;
-		  else
-			continue;
+		if( lock != BtLockRead && drill == lvl ) {
+		  bt_unlockpage(mode, set->latch);
+		  bt_unpinlatch (set->latch);
+		  bt_unpinpool (set->pool);
+		  continue;
+		}
 	}
 
+	prevpage = set->page_no;
+	prevlatch = set->latch;
+	prevpool = set->pool;
+	prevmode = mode;
+
 	//  find key on page at this level
 	//  and descend to requested level
 
-	if( !bt->page->kill && (slot = bt_findslot (bt, key, len)) ) {
+	if( !set->page->kill )
+	 if( slot = bt_findslot (set, key, len) ) {
 	  if( drill == lvl )
 		return slot;
 
-	  while( slotptr(bt->page, slot)->dead )
-		if( slot++ < bt->page->cnt )
+	  while( slotptr(set->page, slot)->dead )
+		if( slot++ < set->page->cnt )
 			continue;
-		else {
-			page_no = bt_getid(bt->page->right);
+		else
 			goto slideright;
-		}
 
-	  page_no = bt_getid(slotptr(bt->page, slot)->id);
+	  page_no = bt_getid(slotptr(set->page, slot)->id);
 	  drill--;
-	}
+	  continue;
+	 }
 
 	//  or slide right into next page
-	//  (slide left from deleted page)
-
-	else
-		page_no = bt_getid(bt->page->right);
-
-	//  continue down / right using overlapping locks
-	//  to protect pages being killed or split.
 
 slideright:
-	prevpage = bt->page_no;
-	prevmode = mode;
+	page_no = bt_getid(set->page->right);
+
   } while( page_no );
 
   // return error on end of right chain
@@ -1367,96 +1457,235 @@ slideright:
   return 0;	// return error
 }
 
+//	return page to free list
+//	page must be delete & write locked
+
+void bt_freepage (BtDb *bt, BtPageSet *set)
+{
+	//	lock allocation page
+
+	bt_spinwritelock (bt->mgr->latchmgr->lock);
+
+	//	store chain in second right
+	bt_putid(set->page->right, bt_getid(bt->mgr->latchmgr->alloc[1].right));
+	bt_putid(bt->mgr->latchmgr->alloc[1].right, set->page_no);
+	set->page->free = 1;
+
+	// unlock released page
+
+	bt_unlockpage (BtLockDelete, set->latch);
+	bt_unlockpage (BtLockWrite, set->latch);
+	bt_unpinlatch (set->latch);
+	bt_unpinpool (set->pool);
+
+	// unlock allocation page
+
+	bt_spinreleasewrite (bt->mgr->latchmgr->lock);
+}
+
+//	a fence key was deleted from a page
+//	push new fence value upwards
+
+BTERR bt_fixfence (BtDb *bt, BtPageSet *set, uint lvl)
+{
+unsigned char leftkey[256], rightkey[256];
+uid page_no;
+BtKey ptr;
+
+	//	remove the old fence value
+
+	ptr = keyptr(set->page, set->page->cnt);
+	memcpy (rightkey, ptr, ptr->len + 1);
+
+	memset (slotptr(set->page, set->page->cnt--), 0, sizeof(BtSlot));
+	set->page->dirty = 1;
+
+	ptr = keyptr(set->page, set->page->cnt);
+	memcpy (leftkey, ptr, ptr->len + 1);
+	page_no = set->page_no;
+
+	bt_lockpage (BtLockParent, set->latch);
+	bt_unlockpage (BtLockWrite, set->latch);
+
+	//	insert new (now smaller) fence key
+
+	if( bt_insertkey (bt, leftkey+1, *leftkey, lvl+1, page_no, time(NULL)) )
+	  return bt->err;
+
+	//	now delete old fence key
+
+	if( bt_deletekey (bt, rightkey+1, *rightkey, lvl+1) )
+		return bt->err;
+
+	bt_unlockpage (BtLockParent, set->latch);
+	bt_unpinlatch(set->latch);
+	bt_unpinpool (set->pool);
+	return 0;
+}
+
+//	root has a single child
+//	collapse a level from the tree
+
+BTERR bt_collapseroot (BtDb *bt, BtPageSet *root)
+{
+BtPageSet child[1];
+uint idx;
+
+  // find the child entry and promote as new root contents
+
+  do {
+	for( idx = 0; idx++ < root->page->cnt; )
+	  if( !slotptr(root->page, idx)->dead )
+		break;
+
+	child->page_no = bt_getid (slotptr(root->page, idx)->id);
+
+	child->latch = bt_pinlatch (bt, child->page_no);
+	bt_lockpage (BtLockDelete, child->latch);
+	bt_lockpage (BtLockWrite, child->latch);
+
+	if( child->pool = bt_pinpool (bt, child->page_no) )
+		child->page = bt_page (bt, child->pool, child->page_no);
+	else
+		return bt->err;
+
+	memcpy (root->page, child->page, bt->mgr->page_size);
+	bt_freepage (bt, child);
+
+  } while( root->page->lvl > 1 && root->page->act == 1 );
+
+  bt_unlockpage (BtLockWrite, root->latch);
+  bt_unpinlatch (root->latch);
+  bt_unpinpool (root->pool);
+  return 0;
+}
+
 //  find and delete key on page by marking delete flag bit
-//  when page becomes empty, delete it
+//  if page becomes empty, delete it from the btree
 
 BTERR bt_deletekey (BtDb *bt, unsigned char *key, uint len, uint lvl)
 {
-unsigned char lowerkey[256], higherkey[256];
-uid page_no, right;
-uint slot, tod;
+unsigned char lowerfence[256], higherfence[256];
+uint slot, idx, dirty = 0, fence, found;
+BtPageSet set[1], right[1];
 BtKey ptr;
 
-	if( slot = bt_loadpage (bt, key, len, lvl, BtLockWrite) )
-		ptr = keyptr(bt->page, slot);
+	if( slot = bt_loadpage (bt, set, key, len, lvl, BtLockWrite) )
+		ptr = keyptr(set->page, slot);
 	else
 		return bt->err;
 
+	//	are we deleting a fence slot?
+
+	fence = slot == set->page->cnt;
+
 	// if key is found delete it, otherwise ignore request
 
-	if( !keycmp (ptr, key, len) )
-		if( slotptr(bt->page, slot)->dead == 0 ) {
-			slotptr(bt->page,slot)->dead = 1;
-			if( slot < bt->page->cnt )
-				bt->page->dirty = 1;
-			bt->page->act--;
-		}
+	if( found = !keycmp (ptr, key, len) )
+	  if( found = slotptr(set->page, slot)->dead == 0 ) {
+		dirty = slotptr(set->page, slot)->dead = 1;
+ 		set->page->dirty = 1;
+ 		set->page->act--;
 
-	// return if page is not empty, or it has no right sibling
+		// collapse empty slots
 
-	right = bt_getid(bt->page->right);
-	page_no = bt->page_no;
+		while( idx = set->page->cnt - 1 )
+		  if( slotptr(set->page, idx)->dead ) {
+			*slotptr(set->page, idx) = *slotptr(set->page, idx + 1);
+			memset (slotptr(set->page, set->page->cnt--), 0, sizeof(BtSlot));
+		  } else
+			break;
+		  }
 
-	if( !right || bt->page->act )
-		return bt_unlockpage(bt, page_no, BtLockWrite);
+	//	did we delete a fence key in an upper level?
+
+	if( dirty && lvl && set->page->act && fence )
+	  if( bt_fixfence (bt, set, lvl) )
+		return bt->err;
+	  else
+		return bt->found = found, 0;
 
-	// obtain Parent lock over write lock
+	//	is this a collapsed root?
 
-	if( bt_lockpage(bt, page_no, BtLockParent, NULL) )
+	if( lvl > 1 && set->page_no == ROOT_page && set->page->act == 1 )
+	  if( bt_collapseroot (bt, set) )
 		return bt->err;
+	  else
+		return bt->found = found, 0;
 
-	// keep copy of key to delete
+	//	return if page is not empty
 
-	ptr = keyptr(bt->page, bt->page->cnt);
-	memcpy(lowerkey, ptr, ptr->len + 1);
+ 	if( set->page->act ) {
+		bt_unlockpage(BtLockWrite, set->latch);
+		bt_unpinlatch (set->latch);
+		bt_unpinpool (set->pool);
+		return bt->found = found, 0;
+	}
 
-	// lock and map right page
+	//	cache copy of fence key
+	//	to post in parent
 
-	if ( bt_lockpage(bt, right, BtLockWrite, &bt->temp) )
-		return bt->err;
+	ptr = keyptr(set->page, set->page->cnt);
+	memcpy (lowerfence, ptr, ptr->len + 1);
 
-	// pull contents of next page into current empty page 
-	memcpy((char *)bt->page + sizeof(BtLatchSet), (char *)bt->temp + sizeof(BtLatchSet), bt->mgr->page_size - sizeof(BtLatchSet));
+	//	obtain lock on right page
 
-	//	keep copy of key to update
-	ptr = keyptr(bt->temp, bt->temp->cnt);
-	memcpy(higherkey, ptr, ptr->len + 1);
+	right->page_no = bt_getid(set->page->right);
+	right->latch = bt_pinlatch (bt, right->page_no);
+	bt_lockpage (BtLockWrite, right->latch);
 
-	//  Mark right page as deleted and point it to left page
-	//	until we can post updates at higher level.
+	// pin page contents
 
-	bt_putid(bt->temp->right, page_no);
-	bt->temp->kill = 1;
-	bt->temp->cnt = 0;
+	if( right->pool = bt_pinpool (bt, right->page_no) )
+		right->page = bt_page (bt, right->pool, right->page_no);
+	else
+		return 0;
 
-	if( bt_unlockpage(bt, right, BtLockWrite) )
-		return bt->err;
-	if( bt_unlockpage(bt, page_no, BtLockWrite) )
-		return bt->err;
+	if( right->page->kill )
+		return bt->err = BTERR_struct;
 
-	//  delete old lower key to consolidated node
+	// pull contents of right peer into our empty page
 
-	if( bt_deletekey (bt, lowerkey + 1, *lowerkey, lvl + 1) )
-		return bt->err;
+	memcpy (set->page, right->page, bt->mgr->page_size);
 
-	//  redirect higher key directly to consolidated node
+	// cache copy of key to update
 
-	tod = (uint)time(NULL);
+	ptr = keyptr(right->page, right->page->cnt);
+	memcpy (higherfence, ptr, ptr->len + 1);
 
-	if( bt_insertkey (bt, higherkey+1, *higherkey, lvl + 1, page_no, tod) )
-		return bt->err;
+	// mark right page deleted and point it to left page
+	//	until we can post parent updates
 
-	//	obtain write lock and
-	//	add right block to free chain
+	bt_putid (right->page->right, set->page_no);
+	right->page->kill = 1;
 
-	if( bt_freepage (bt, right) )
-		return bt->err;
+	bt_lockpage (BtLockParent, right->latch);
+	bt_unlockpage (BtLockWrite, right->latch);
 
-	// 	remove ParentModify lock
+	bt_lockpage (BtLockParent, set->latch);
+	bt_unlockpage (BtLockWrite, set->latch);
 
-	if( bt_unlockpage(bt, page_no, BtLockParent) )
-		return bt->err;
-	
+	// redirect higher key directly to our new node contents
+
+	if( bt_insertkey (bt, higherfence+1, *higherfence, lvl+1, set->page_no, time(NULL)) )
+	  return bt->err;
+
+	//	delete old lower key to our node
+
+	if( bt_deletekey (bt, lowerfence+1, *lowerfence, lvl+1) )
+	  return bt->err;
+
+	//	obtain delete and write locks to right node
+
+	bt_unlockpage (BtLockParent, right->latch);
+	bt_lockpage (BtLockDelete, right->latch);
+	bt_lockpage (BtLockWrite, right->latch);
+	bt_freepage (bt, right);
+
+	bt_unlockpage (BtLockParent, set->latch);
+	bt_unpinlatch (set->latch);
+	bt_unpinpool (set->pool);
+	bt->found = found;
 	return 0;
 }
 
@@ -1464,44 +1693,44 @@ BtKey ptr;
 
 uid bt_findkey (BtDb *bt, unsigned char *key, uint len)
 {
+BtPageSet set[1];
 uint  slot;
+uid id = 0;
 BtKey ptr;
-uid id;
 
-	if( slot = bt_loadpage (bt, key, len, 0, BtLockRead) )
-		ptr = keyptr(bt->page, slot);
+	if( slot = bt_loadpage (bt, set, key, len, 0, BtLockRead) )
+		ptr = keyptr(set->page, slot);
 	else
 		return 0;
 
 	// if key exists, return row-id
 	//	otherwise return 0
 
-	if( ptr->len == len && !memcmp (ptr->key, key, len) )
-		id = bt_getid(slotptr(bt->page,slot)->id);
-	else
-		id = 0;
-
-	if ( bt_unlockpage(bt, bt->page_no, BtLockRead) )
-		return 0;
+	if( slot <= set->page->cnt )
+	  if( !keycmp (ptr, key, len) )
+		id = bt_getid(slotptr(set->page,slot)->id);
 
+	bt_unlockpage (BtLockRead, set->latch);
+	bt_unpinlatch (set->latch);
+	bt_unpinpool (set->pool);
 	return id;
 }
 
 //	check page for space available,
 //	clean if necessary and return
 //	0 - page needs splitting
-//	1 - go ahead
+//	>0  new slot value
 
-uint bt_cleanpage(BtDb *bt, uint amt)
+uint bt_cleanpage(BtDb *bt, BtPage page, uint amt, uint slot)
 {
 uint nxt = bt->mgr->page_size;
-BtPage page = bt->page;
 uint cnt = 0, idx = 0;
 uint max = page->cnt;
+uint newslot = max;
 BtKey key;
 
-	if( page->min >= (page->cnt+1) * sizeof(BtSlot) + sizeof(*page) + amt + 1 )
-		return 1;
+	if( page->min >= (max+1) * sizeof(BtSlot) + sizeof(*page) + amt + 1 )
+		return slot;
 
 	//	skip cleanup if nothing to reclaim
 
@@ -1516,116 +1745,124 @@ BtKey key;
 	page->dirty = 0;
 	page->act = 0;
 
+	// try cleaning up page first
+	// by removing deleted keys
+
 	while( cnt++ < max ) {
-		// always leave fence key in list
+		if( cnt == slot )
+			newslot = idx + 1;
 		if( cnt < max && slotptr(bt->frame,cnt)->dead )
 			continue;
 
-		// copy key
+		// copy the key across
+
 		key = keyptr(bt->frame, cnt);
 		nxt -= key->len + 1;
 		memcpy ((unsigned char *)page + nxt, key, key->len + 1);
 
 		// copy slot
+
 		memcpy(slotptr(page, ++idx)->id, slotptr(bt->frame, cnt)->id, BtId);
 		if( !(slotptr(page, idx)->dead = slotptr(bt->frame, cnt)->dead) )
 			page->act++;
 		slotptr(page, idx)->tod = slotptr(bt->frame, cnt)->tod;
 		slotptr(page, idx)->off = nxt;
 	}
+
 	page->min = nxt;
 	page->cnt = idx;
 
-	if( page->min >= (page->cnt+1) * sizeof(BtSlot) + sizeof(*page) + amt + 1 )
-		return 1;
+	//	see if page has enough space now, or does it need splitting?
+
+	if( page->min >= (idx+1) * sizeof(BtSlot) + sizeof(*page) + amt + 1 )
+		return newslot;
 
 	return 0;
 }
 
 // split the root and raise the height of the btree
 
-BTERR bt_splitroot(BtDb *bt,  unsigned char *newkey, unsigned char *oldkey, uid page_no2)
+BTERR bt_splitroot(BtDb *bt, BtPageSet *root, unsigned char *leftkey, uid page_no2)
 {
 uint nxt = bt->mgr->page_size;
-BtPage root = bt->page;
-uid new_page;
+uid left;
 
 	//  Obtain an empty page to use, and copy the current
-	//  root contents into it which is the lower half of
-	//	the old root.
+	//  root contents into it, e.g. lower keys
 
-	if( !(new_page = bt_newpage(bt, root)) )
+	if( !(left = bt_newpage(bt, root->page)) )
 		return bt->err;
 
 	// preserve the page info at the bottom
-	// and set rest to zero
+	// of higher keys and set rest to zero
 
-	memset(root+1, 0, bt->mgr->page_size - sizeof(*root));
+	memset(root->page+1, 0, bt->mgr->page_size - sizeof(*root->page));
 
-	// insert first key on newroot page
+	// insert lower keys page fence key on newroot page as first key
 
-	nxt -= *newkey + 1;
-	memcpy ((unsigned char *)root + nxt, newkey, *newkey + 1);
-	bt_putid(slotptr(root, 1)->id, new_page);
-	slotptr(root, 1)->off = nxt;
+	nxt -= *leftkey + 1;
+	memcpy ((unsigned char *)root->page + nxt, leftkey, *leftkey + 1);
+	bt_putid(slotptr(root->page, 1)->id, left);
+	slotptr(root->page, 1)->off = nxt;
 	
-	// insert second key on newroot page
+	// insert stopper key on newroot page
 	// and increase the root height
 
-	nxt -= *oldkey + 1;
-	memcpy ((unsigned char *)root + nxt, oldkey, *oldkey + 1);
-	bt_putid(slotptr(root, 2)->id, page_no2);
-	slotptr(root, 2)->off = nxt;
+	nxt -= 3;
+	((unsigned char *)root->page)[nxt] = 2;
+	((unsigned char *)root->page)[nxt+1] = 0xff;
+	((unsigned char *)root->page)[nxt+2] = 0xff;
+	bt_putid(slotptr(root->page, 2)->id, page_no2);
+	slotptr(root->page, 2)->off = nxt;
 
-	bt_putid(root->right, 0);
-	root->min = nxt;		// reset lowest used offset and key count
-	root->cnt = 2;
-	root->act = 2;
-	root->lvl++;
+	bt_putid(root->page->right, 0);
+	root->page->min = nxt;		// reset lowest used offset and key count
+	root->page->cnt = 2;
+	root->page->act = 2;
+	root->page->lvl++;
 
-	// release root (bt->page)
+	// release and unpin root
 
-	return bt_unlockpage(bt, bt->page_no, BtLockWrite);
+	bt_unlockpage(BtLockWrite, root->latch);
+	bt_unpinlatch (root->latch);
+	bt_unpinpool (root->pool);
+	return 0;
 }
 
 //  split already locked full node
 //	return unlocked.
 
-BTERR bt_splitpage (BtDb *bt)
+BTERR bt_splitpage (BtDb *bt, BtPageSet *set)
 {
 uint cnt = 0, idx = 0, max, nxt = bt->mgr->page_size;
-unsigned char oldkey[256], lowerkey[256];
-uid page_no = bt->page_no, right;
-BtPage page = bt->page;
-uint lvl = page->lvl;
-uid new_page;
+unsigned char fencekey[256], rightkey[256];
+uint lvl = set->page->lvl;
+BtPageSet right[1];
+uint prev;
 BtKey key;
-uint tod;
 
 	//  split higher half of keys to bt->frame
-	//	the last key (fence key) might be dead
-
-	tod = (uint)time(NULL);
 
 	memset (bt->frame, 0, bt->mgr->page_size);
-	max = (int)page->cnt;
+	max = set->page->cnt;
 	cnt = max / 2;
 	idx = 0;
 
 	while( cnt++ < max ) {
-		key = keyptr(page, cnt);
+		key = keyptr(set->page, cnt);
 		nxt -= key->len + 1;
 		memcpy ((unsigned char *)bt->frame + nxt, key, key->len + 1);
-		memcpy(slotptr(bt->frame,++idx)->id, slotptr(page,cnt)->id, BtId);
-		if( !(slotptr(bt->frame, idx)->dead = slotptr(page, cnt)->dead) )
+
+		memcpy(slotptr(bt->frame,++idx)->id, slotptr(set->page,cnt)->id, BtId);
+		if( !(slotptr(bt->frame, idx)->dead = slotptr(set->page, cnt)->dead) )
 			bt->frame->act++;
-		slotptr(bt->frame, idx)->tod = slotptr(page, cnt)->tod;
+		slotptr(bt->frame, idx)->tod = slotptr(set->page, cnt)->tod;
 		slotptr(bt->frame, idx)->off = nxt;
 	}
 
 	// remember existing fence key for new page to the right
 
-	memcpy (oldkey, key, key->len + 1);
+	memcpy (rightkey, key, key->len + 1);
 
 	bt->frame->bits = bt->mgr->page_bits;
 	bt->frame->min = nxt;
@@ -1634,167 +1871,166 @@ uint tod;
 
 	// link right node
 
-	if( page_no > ROOT_page ) {
-		right = bt_getid (page->right);
-		bt_putid(bt->frame->right, right);
-	}
+	if( set->page_no > ROOT_page )
+		memcpy (bt->frame->right, set->page->right, BtId);
 
-	//	get new free page and write frame to it.
+	//	get new free page and write higher keys to it.
 
-	if( !(new_page = bt_newpage(bt, bt->frame)) )
+	if( !(right->page_no = bt_newpage(bt, bt->frame)) )
 		return bt->err;
 
 	//	update lower keys to continue in old page
 
-	memcpy (bt->frame, page, bt->mgr->page_size);
-	memset (page+1, 0, bt->mgr->page_size - sizeof(*page));
+	memcpy (bt->frame, set->page, bt->mgr->page_size);
+	memset (set->page+1, 0, bt->mgr->page_size - sizeof(*set->page));
 	nxt = bt->mgr->page_size;
-	page->act = 0;
+	set->page->dirty = 0;
+	set->page->act = 0;
 	cnt = 0;
 	idx = 0;
 
 	//  assemble page of smaller keys
-	//	(they're all active keys)
 
 	while( cnt++ < max / 2 ) {
 		key = keyptr(bt->frame, cnt);
 		nxt -= key->len + 1;
-		memcpy ((unsigned char *)page + nxt, key, key->len + 1);
-		memcpy(slotptr(page,++idx)->id, slotptr(bt->frame,cnt)->id, BtId);
-		slotptr(page, idx)->tod = slotptr(bt->frame, cnt)->tod;
-		slotptr(page, idx)->off = nxt;
-		page->act++;
+		memcpy ((unsigned char *)set->page + nxt, key, key->len + 1);
+		memcpy(slotptr(set->page,++idx)->id, slotptr(bt->frame,cnt)->id, BtId);
+		slotptr(set->page, idx)->tod = slotptr(bt->frame, cnt)->tod;
+		slotptr(set->page, idx)->off = nxt;
+		set->page->act++;
 	}
 
-	// remember fence key for old page
+	// remember fence key for smaller page
 
-	memcpy(lowerkey, key, key->len + 1);
-	bt_putid(page->right, new_page);
-	page->min = nxt;
-	page->cnt = idx;
-
-	// if current page is the root page, split it
+	memcpy(fencekey, key, key->len + 1);
 
-	if( page_no == ROOT_page )
-		return bt_splitroot (bt, lowerkey, oldkey, new_page);
+	bt_putid(set->page->right, right->page_no);
+	set->page->min = nxt;
+	set->page->cnt = idx;
 
-	// obtain Parent/Write locks
-	// for left and right node pages
-
-	if( bt_lockpage (bt, new_page, BtLockParent, NULL) )
-		return bt->err;
-
-	if( bt_lockpage (bt, page_no, BtLockParent, NULL) )
-		return bt->err;
+	// if current page is the root page, split it
 
-	//  release wr lock on left page
+	if( set->page_no == ROOT_page )
+		return bt_splitroot (bt, set, fencekey, right->page_no);
 
-	if( bt_unlockpage (bt, page_no, BtLockWrite) )
-		return bt->err;
+	// insert new fences in their parent pages
 
-	// insert new fence for reformulated left block
+	right->latch = bt_pinlatch (bt, right->page_no);
+	bt_lockpage (BtLockParent, right->latch);
 
-	if( bt_insertkey (bt, lowerkey+1, *lowerkey, lvl + 1, page_no, tod) )
-		return bt->err;
+	bt_lockpage (BtLockParent, set->latch);
+	bt_unlockpage (BtLockWrite, set->latch);
 
-	// fix old fence for newly allocated right block page
+	// insert new fence for reformulated left block of smaller keys
 
-	if( bt_insertkey (bt, oldkey+1, *oldkey, lvl + 1, new_page, tod) )
+	if( bt_insertkey (bt, fencekey+1, *fencekey, lvl+1, set->page_no, time(NULL)) )
 		return bt->err;
 
-	// release Parent & Write locks
+	// switch fence for right block of larger keys to new right page
 
-	if( bt_unlockpage (bt, new_page, BtLockParent) )
+	if( bt_insertkey (bt, rightkey+1, *rightkey, lvl+1, right->page_no, time(NULL)) )
 		return bt->err;
 
-	if( bt_unlockpage (bt, page_no, BtLockParent) )
-		return bt->err;
+	bt_unlockpage (BtLockParent, set->latch);
+	bt_unpinlatch (set->latch);
+	bt_unpinpool (set->pool);
 
+	bt_unlockpage (BtLockParent, right->latch);
+	bt_unpinlatch (right->latch);
 	return 0;
 }
-
-//  Insert new key into the btree at requested level.
-//  Level zero pages are leaf pages and are unlocked at exit.
-//	Interior pages remain locked.
+//  Insert new key into the btree at given level.
 
 BTERR bt_insertkey (BtDb *bt, unsigned char *key, uint len, uint lvl, uid id, uint tod)
 {
+BtPageSet set[1];
 uint slot, idx;
-BtPage page;
 BtKey ptr;
 
-  while( 1 ) {
-	if( slot = bt_loadpage (bt, key, len, lvl, BtLockWrite) )
-		ptr = keyptr(bt->page, slot);
-	else
-	{
-		if ( !bt->err )
-			bt->err = BTERR_ovflw;
-		return bt->err;
-	}
-
-	// if key already exists, update id and return
-
-	page = bt->page;
+	while( 1 ) {
+		if( slot = bt_loadpage (bt, set, key, len, lvl, BtLockWrite) )
+			ptr = keyptr(set->page, slot);
+		else
+		{
+			if( !bt->err )
+				bt->err = BTERR_ovflw;
+			return bt->err;
+		}
 
-	if( !keycmp (ptr, key, len) ) {
-		slotptr(page, slot)->dead = 0;
-		slotptr(page, slot)->tod = tod;
-		bt_putid(slotptr(page,slot)->id, id);
-		return bt_unlockpage(bt, bt->page_no, BtLockWrite);
-	}
+		// if key already exists, update id and return
+
+		if( !keycmp (ptr, key, len) ) {
+			if( slotptr(set->page, slot)->dead )
+				set->page->act++;
+			slotptr(set->page, slot)->dead = 0;
+			slotptr(set->page, slot)->tod = tod;
+			bt_putid(slotptr(set->page,slot)->id, id);
+			bt_unlockpage(BtLockWrite, set->latch);
+			bt_unpinlatch (set->latch);
+			bt_unpinpool (set->pool);
+			return 0;
+		}
 
-	// check if page has enough space
+		// check if page has enough space
 
-	if( bt_cleanpage (bt, len) )
-		break;
+ 		if( slot = bt_cleanpage (bt, set->page, len, slot) )
+			break;
 
-	if( bt_splitpage (bt) )
-		return bt->err;
-  }
+		if( bt_splitpage (bt, set) )
+			return bt->err;
+	}
 
-  // calculate next available slot and copy key into page
+	// calculate next available slot and copy key into page
 
-  page->min -= len + 1; // reset lowest used offset
-  ((unsigned char *)page)[page->min] = len;
-  memcpy ((unsigned char *)page + page->min +1, key, len );
+	set->page->min -= len + 1; // reset lowest used offset
+	((unsigned char *)set->page)[set->page->min] = len;
+	memcpy ((unsigned char *)set->page + set->page->min +1, key, len );
 
-  for( idx = slot; idx < page->cnt; idx++ )
-	if( slotptr(page, idx)->dead )
+	for( idx = slot; idx < set->page->cnt; idx++ )
+	  if( slotptr(set->page, idx)->dead )
 		break;
 
-  // now insert key into array before slot
-  // preserving the fence slot
+	// now insert key into array before slot
 
-  if( idx == page->cnt )
-	idx++, page->cnt++;
+	if( idx == set->page->cnt )
+		idx++, set->page->cnt++;
 
-  page->act++;
+	set->page->act++;
 
-  while( idx > slot )
-	*slotptr(page, idx) = *slotptr(page, idx -1), idx--;
+	while( idx > slot )
+		*slotptr(set->page, idx) = *slotptr(set->page, idx -1), idx--;
 
-  bt_putid(slotptr(page,slot)->id, id);
-  slotptr(page, slot)->off = page->min;
-  slotptr(page, slot)->tod = tod;
-  slotptr(page, slot)->dead = 0;
+	bt_putid(slotptr(set->page,slot)->id, id);
+	slotptr(set->page, slot)->off = set->page->min;
+	slotptr(set->page, slot)->tod = tod;
+	slotptr(set->page, slot)->dead = 0;
 
-  return bt_unlockpage(bt, bt->page_no, BtLockWrite);
+	bt_unlockpage (BtLockWrite, set->latch);
+	bt_unpinlatch (set->latch);
+	bt_unpinpool (set->pool);
+	return 0;
 }
 
 //  cache page of keys into cursor and return starting slot for given key
 
 uint bt_startkey (BtDb *bt, unsigned char *key, uint len)
 {
+BtPageSet set[1];
 uint slot;
 
 	// cache page for retrieval
-	if( slot = bt_loadpage (bt, key, len, 0, BtLockRead) )
-		memcpy (bt->cursor, bt->page, bt->mgr->page_size);
-	bt->cursor_page = bt->page_no;
-	if ( bt_unlockpage(bt, bt->page_no, BtLockRead) )
-		return 0;
 
+	if( slot = bt_loadpage (bt, set, key, len, 0, BtLockRead) )
+	  memcpy (bt->cursor, set->page, bt->mgr->page_size);
+	else
+	  return 0;
+
+	bt->cursor_page = set->page_no;
+
+	bt_unlockpage(BtLockRead, set->latch);
+	bt_unpinlatch (set->latch);
+	bt_unpinpool (set->pool);
 	return slot;
 }
 
@@ -1803,14 +2039,16 @@ uint slot;
 
 uint bt_nextkey (BtDb *bt, uint slot)
 {
-off64_t right;
+BtPageSet set[1];
+uid right;
 
   do {
 	right = bt_getid(bt->cursor->right);
+
 	while( slot++ < bt->cursor->cnt )
 	  if( slotptr(bt->cursor,slot)->dead )
 		continue;
-	  else if( right || (slot < bt->cursor->cnt))
+	  else if( right || (slot < bt->cursor->cnt) ) // skip infinite stopper
 		return slot;
 	  else
 		break;
@@ -1820,15 +2058,21 @@ off64_t right;
 
 	bt->cursor_page = right;
 
-    if( bt_lockpage(bt, right, BtLockRead, &bt->page) )
+	if( set->pool = bt_pinpool (bt, right) )
+		set->page = bt_page (bt, set->pool, right);
+	else
 		return 0;
 
-	memcpy (bt->cursor, bt->page, bt->mgr->page_size);
+	set->latch = bt_pinlatch (bt, right);
+    bt_lockpage(BtLockRead, set->latch);
 
-	if ( bt_unlockpage(bt, right, BtLockRead) )
-		return 0;
+	memcpy (bt->cursor, set->page, bt->mgr->page_size);
 
+	bt_unlockpage(BtLockRead, set->latch);
+	bt_unpinlatch (set->latch);
+	bt_unpinpool (set->pool);
 	slot = 0;
+
   } while( 1 );
 
   return bt->err = 0;
@@ -1849,9 +2093,138 @@ uint bt_tod(BtDb *bt, uint slot)
 	return slotptr(bt->cursor,slot)->tod;
 }
 
-
 #ifdef STANDALONE
 
+#ifndef unix
+double getCpuTime(int type)
+{
+FILETIME crtime[1];
+FILETIME xittime[1];
+FILETIME systime[1];
+FILETIME usrtime[1];
+SYSTEMTIME timeconv[1];
+double ans = 0;
+
+	memset (timeconv, 0, sizeof(SYSTEMTIME));
+
+	switch( type ) {
+	case 0:
+		GetSystemTimeAsFileTime (xittime);
+		FileTimeToSystemTime (xittime, timeconv);
+		ans = (double)timeconv->wDayOfWeek * 3600 * 24;
+		break;
+	case 1:
+		GetProcessTimes (GetCurrentProcess(), crtime, xittime, systime, usrtime);
+		FileTimeToSystemTime (usrtime, timeconv);
+		break;
+	case 2:
+		GetProcessTimes (GetCurrentProcess(), crtime, xittime, systime, usrtime);
+		FileTimeToSystemTime (systime, timeconv);
+		break;
+	}
+
+	ans += (double)timeconv->wHour * 3600;
+	ans += (double)timeconv->wMinute * 60;
+	ans += (double)timeconv->wSecond;
+	ans += (double)timeconv->wMilliseconds / 1000;
+	return ans;
+}
+#else
+#include <time.h>
+#include <sys/resource.h>
+
+double getCpuTime(int type)
+{
+struct rusage used[1];
+struct timeval tv[1];
+
+	switch( type ) {
+	case 0:
+		gettimeofday(tv, NULL);
+		return (double)tv->tv_sec + (double)tv->tv_usec / 1000000;
+
+	case 1:
+		getrusage(RUSAGE_SELF, used);
+		return (double)used->ru_utime.tv_sec + (double)used->ru_utime.tv_usec / 1000000;
+
+	case 2:
+		getrusage(RUSAGE_SELF, used);
+		return (double)used->ru_stime.tv_sec + (double)used->ru_stime.tv_usec / 1000000;
+	}
+
+	return 0;
+}
+#endif
+
+void bt_latchaudit (BtDb *bt)
+{
+ushort idx, hashidx;
+uid next, page_no;
+BtLatchSet *latch;
+BtKey ptr;
+
+#ifdef unix
+	if( *(uint *)(bt->mgr->latchmgr->lock) )
+		fprintf(stderr, "Alloc page locked\n");
+	*(uint *)(bt->mgr->latchmgr->lock) = 0;
+
+	for( idx = 1; idx <= bt->mgr->latchmgr->latchdeployed; idx++ ) {
+		latch = bt->mgr->latchsets + idx;
+		if( *(uint *)latch->readwr )
+			fprintf(stderr, "latchset %d rwlocked for page %.8x\n", idx, latch->page_no);
+		*(uint *)latch->readwr = 0;
+
+		if( *(uint *)latch->access )
+			fprintf(stderr, "latchset %d accesslocked for page %.8x\n", idx, latch->page_no);
+		*(uint *)latch->access = 0;
+
+		if( *(uint *)latch->parent )
+			fprintf(stderr, "latchset %d parentlocked for page %.8x\n", idx, latch->page_no);
+		*(uint *)latch->parent = 0;
+
+		if( latch->pin ) {
+			fprintf(stderr, "latchset %d pinned for page %.8x\n", idx, latch->page_no);
+			latch->pin = 0;
+		}
+	}
+
+	for( hashidx = 0; hashidx < bt->mgr->latchmgr->latchhash; hashidx++ ) {
+	  if( *(uint *)(bt->mgr->latchmgr->table[hashidx].latch) )
+			fprintf(stderr, "hash entry %d locked\n", hashidx);
+
+	  *(uint *)(bt->mgr->latchmgr->table[hashidx].latch) = 0;
+
+	  if( idx = bt->mgr->latchmgr->table[hashidx].slot ) do {
+		latch = bt->mgr->latchsets + idx;
+		if( *(uint *)latch->busy )
+			fprintf(stderr, "latchset %d busylocked for page %.8x\n", idx, latch->page_no);
+		*(uint *)latch->busy = 0;
+		if( latch->hash != hashidx )
+			fprintf(stderr, "latchset %d wrong hashidx\n", idx);
+		if( latch->pin )
+			fprintf(stderr, "latchset %d pinned for page %.8x\n", idx, latch->page_no);
+	  } while( idx = latch->next );
+	}
+
+	next = bt->mgr->latchmgr->nlatchpage + LATCH_page;
+	page_no = LEAF_page;
+
+	while( page_no < bt_getid(bt->mgr->latchmgr->alloc->right) ) {
+		pread (bt->mgr->idx, bt->frame, bt->mgr->page_size, page_no << bt->mgr->page_bits);
+		if( !bt->frame->free )
+		 for( idx = 0; idx++ < bt->frame->cnt - 1; ) {
+		  ptr = keyptr(bt->frame, idx+1);
+		  if( keycmp (keyptr(bt->frame, idx), ptr->key, ptr->len) >= 0 )
+			fprintf(stderr, "page %.8x idx %.2x out of order\n", page_no, idx);
+		 }
+
+		if( page_no > LEAF_page )
+			next = page_no + 1;
+		page_no = next;
+	}
+#endif
+}
+
 typedef struct {
 	char type, idx;
 	char *infile;
@@ -1873,8 +2246,8 @@ uid next, page_no = LEAF_page;	// start on first page of leaves
 unsigned char key[256];
 ThreadArg *args = arg;
 int ch, len = 0, slot;
+BtPageSet set[1];
 time_t tod[1];
-BtPage page;
 BtKey ptr;
 BtDb *bt;
 FILE *in;
@@ -1884,6 +2257,12 @@ FILE *in;
 
 	switch(args->type | 0x20)
 	{
+	case 'a':
+		fprintf(stderr, "started latch mgr audit\n");
+		bt_latchaudit (bt);
+		fprintf(stderr, "finished latch mgr audit\n");
+		break;
+
 	case 'w':
 		fprintf(stderr, "started indexing for %s\n", args->infile);
 		if( in = fopen (args->infile, "rb") )
@@ -1894,8 +2273,9 @@ FILE *in;
 
 			  if( args->num == 1 )
 		  		sprintf((char *)key+len, "%.9d", 1000000000 - line), len += 9;
+
 			  else if( args->num )
-		  		sprintf((char *)key+len, "%.9d", line+args->idx * args->num), len += 9;
+		  		sprintf((char *)key+len, "%.9d", line + args->idx * args->num), len += 9;
 
 			  if( bt_insertkey (bt, key, len, 0, line, *tod) )
 				fprintf(stderr, "Error %d Line: %d\n", bt->err, line), exit(0);
@@ -1915,8 +2295,9 @@ FILE *in;
 			  line++;
 			  if( args->num == 1 )
 		  		sprintf((char *)key+len, "%.9d", 1000000000 - line), len += 9;
+
 			  else if( args->num )
-		  		sprintf((char *)key+len, "%.9d", line+args->idx * args->num), len += 9;
+		  		sprintf((char *)key+len, "%.9d", line + args->idx * args->num), len += 9;
 
 			  if( bt_deletekey (bt, key, len, 0) )
 				fprintf(stderr, "Error %d Line: %d\n", bt->err, line), exit(0);
@@ -1936,8 +2317,9 @@ FILE *in;
 			  line++;
 			  if( args->num == 1 )
 		  		sprintf((char *)key+len, "%.9d", 1000000000 - line), len += 9;
+
 			  else if( args->num )
-		  		sprintf((char *)key+len, "%.9d", line+args->idx * args->num), len += 9;
+		  		sprintf((char *)key+len, "%.9d", line + args->idx * args->num), len += 9;
 
 			  if( bt_findkey (bt, key, len) )
 				found++;
@@ -1951,33 +2333,61 @@ FILE *in;
 		break;
 
 	case 's':
-		len = key[0] = 0;
-
-		fprintf(stderr, "started reading\n");
-
-		if( slot = bt_startkey (bt, key, len) )
-		  slot--;
-		else
-		  fprintf(stderr, "Error %d in StartKey. Syserror: %d\n", bt->err, errno), exit(0);
-
-		while( slot = bt_nextkey (bt, slot) ) {
-			ptr = bt_key(bt, slot);
-			fwrite (ptr->key, ptr->len, 1, stdout);
-			fputc ('\n', stdout);
-		}
+		fprintf(stderr, "started scanning\n");
+	  	do {
+			if( set->pool = bt_pinpool (bt, page_no) )
+				set->page = bt_page (bt, set->pool, page_no);
+			else
+				break;
+			set->latch = bt_pinlatch (bt, page_no);
+			bt_lockpage (BtLockRead, set->latch);
+			next = bt_getid (set->page->right);
+			cnt += set->page->act;
+
+			for( slot = 0; slot++ < set->page->cnt; )
+			 if( next || slot < set->page->cnt )
+			  if( !slotptr(set->page, slot)->dead ) {
+				ptr = keyptr(set->page, slot);
+				fwrite (ptr->key, ptr->len, 1, stdout);
+				fputc ('\n', stdout);
+			  }
+
+			bt_unlockpage (BtLockRead, set->latch);
+			bt_unpinlatch (set->latch);
+			bt_unpinpool (set->pool);
+	  	} while( page_no = next );
 
+	  	cnt--;	// remove stopper key
+		fprintf(stderr, " Total keys read %d\n", cnt);
 		break;
 
 	case 'c':
-		fprintf(stderr, "started reading\n");
+		fprintf(stderr, "started counting\n");
+		next = bt->mgr->latchmgr->nlatchpage + LATCH_page;
+		page_no = LEAF_page;
 
-	  	do {
-			bt_lockpage (bt, page_no, BtLockRead, &page);
-			cnt += page->act;
-			next = bt_getid (page->right);
-			bt_unlockpage (bt, page_no, BtLockRead);
-	  	} while( page_no = next );
+		while( page_no < bt_getid(bt->mgr->latchmgr->alloc->right) ) {
+		uid off = page_no << bt->mgr->page_bits;
+#ifdef unix
+		  pread (bt->mgr->idx, bt->frame, bt->mgr->page_size, off);
+#else
+		DWORD amt[1];
 
+		  SetFilePointer (bt->mgr->idx, (long)off, (long*)(&off)+1, FILE_BEGIN);
+
+		  if( !ReadFile(bt->mgr->idx, bt->frame, bt->mgr->page_size, amt, NULL))
+			return bt->err = BTERR_map;
+
+		  if( *amt <  bt->mgr->page_size )
+			return bt->err = BTERR_map;
+#endif
+			if( !bt->frame->free && !bt->frame->lvl )
+				cnt += bt->frame->act;
+			if( page_no > LEAF_page )
+				next = page_no + 1;
+			page_no = next;
+		}
+		
 	  	cnt--;	// remove stopper key
 		fprintf(stderr, " Total keys read %d\n", cnt);
 		break;
@@ -1997,16 +2407,15 @@ int main (int argc, char **argv)
 {
 int idx, cnt, len, slot, err;
 int segsize, bits = 16;
+double start, stop;
 #ifdef unix
 pthread_t *threads;
-timer start, stop;
 #else
-time_t start[1], stop[1];
 HANDLE *threads;
 #endif
-double real_time;
 ThreadArg *args;
 uint poolsize = 0;
+float elapsed;
 int num = 0;
 char key[1];
 BtMgr *mgr;
@@ -2018,16 +2427,12 @@ BtDb *bt;
 		fprintf (stderr, "  where page_bits is the page size in bits\n");
 		fprintf (stderr, "  mapped_segments is the number of mmap segments in buffer pool\n");
 		fprintf (stderr, "  seg_bits is the size of individual segments in buffer pool in pages in bits\n");
-		fprintf (stderr, "  line_numbers set to 1 to append line numbers to input lines\n");
+		fprintf (stderr, "  line_numbers = 1 to append line numbers to keys\n");
 		fprintf (stderr, "  src_file1 thru src_filen are files of keys separated by newline\n");
 		exit(0);
 	}
 
-#ifdef unix
-	gettimeofday(&start, NULL);
-#else
-	time(start);
-#endif
+	start = getCpuTime(0);
 
 	if( argc > 3 )
 		bits = atoi(argv[3]);
@@ -2036,7 +2441,7 @@ BtDb *bt;
 		poolsize = atoi(argv[4]);
 
 	if( !poolsize )
-		fprintf (stderr, "Warning: mapped_pool has no segments\n");
+		fprintf (stderr, "Warning: no mapped_pool\n");
 
 	if( poolsize > 65535 )
 		fprintf (stderr, "Warning: mapped_pool > 65535 segments\n");
@@ -2085,18 +2490,20 @@ BtDb *bt;
 #ifdef unix
 	for( idx = 0; idx < cnt; idx++ )
 		pthread_join (threads[idx], NULL);
-	gettimeofday(&stop, NULL);
-	real_time = 1000.0 * ( stop.tv_sec - start.tv_sec ) + 0.001 * (stop.tv_usec - start.tv_usec );
 #else
 	WaitForMultipleObjects (cnt, threads, TRUE, INFINITE);
 
 	for( idx = 0; idx < cnt; idx++ )
 		CloseHandle(threads[idx]);
 
-	time (stop);
-	real_time = 1000 * (*stop - *start);
 #endif
-	fprintf(stderr, " Time to complete: %.2f seconds\n", real_time/1000);
+	elapsed = getCpuTime(0) - start;
+	fprintf(stderr, " real %dm%.3fs\n", (int)(elapsed/60), elapsed - (int)(elapsed/60)*60);
+	elapsed = getCpuTime(1);
+	fprintf(stderr, " user %dm%.3fs\n", (int)(elapsed/60), elapsed - (int)(elapsed/60)*60);
+	elapsed = getCpuTime(2);
+	fprintf(stderr, " sys  %dm%.3fs\n", (int)(elapsed/60), elapsed - (int)(elapsed/60)*60);
+
 	bt_mgrclose (mgr);
 }