X-Git-Url: https://pd.if.org/git/?a=blobdiff_plain;f=threads2j.c;fp=fosterbtreef2.c;h=d7cb99498cb37d383ed6494aa049337fb6522866;hb=42c2e785f24492e30f4262e7332bd41c7c672a56;hp=e4592d2d2611a5911fa4006ba67c0532d41db766;hpb=fe2ddf2193540d99fa9fb6ecf93beed5b66b8808;p=btree

diff --git a/fosterbtreef2.c b/threads2j.c
similarity index 73%
rename from fosterbtreef2.c
rename to threads2j.c
index e4592d2..d7cb994 100644
--- a/fosterbtreef2.c
+++ b/threads2j.c
@@ -1,5 +1,5 @@
-// foster btree version f2
-// 18 JAN 2014
+// btree version threads2j linux futex concurrency version
+// 24 JAN 2014
 
 // author: karl malbrain, malbrain@cal.berkeley.edu
 
@@ -25,6 +25,8 @@ REDISTRIBUTION OF THIS SOFTWARE.
 
 #ifdef linux
 #define _GNU_SOURCE
+#include <linux/futex.h>
+#define SYS_futex 202
 #endif
 
 #ifdef unix
@@ -36,6 +38,7 @@ 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>
@@ -74,7 +77,7 @@ There are five lock types for each node in three independent sets:
 2. (set 1) NodeDelete: Exclusive. About to release the node. Incompatible with AccessIntent. 
 3. (set 2) ReadLock: Sharable. Read the node. Incompatible with WriteLock. 
 4. (set 2) WriteLock: Exclusive. Modify the node. Incompatible with ReadLock and other WriteLocks. 
-5. (set 3) ParentLock: Exclusive. Have parent adopt/delete maximum foster child from the node.
+5. (set 3) ParentModification: Exclusive. Change the node's parent keys. Incompatible with another ParentModification. 
 */
 
 typedef enum{
@@ -83,7 +86,33 @@ typedef enum{
 	BtLockRead,
 	BtLockWrite,
 	BtLockParent
-}BtLock;
+} 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;
+
+// share is count of read accessors
+// grant write lock when share == 0
+
+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;
 
 //	Define the length of the page and key pointers
 
@@ -121,50 +150,30 @@ typedef struct {
 //	by the BtSlot array of keys.
 
 typedef struct Page {
-	volatile uint cnt;			// count of keys in page
-	volatile uint act;			// count of active keys
-	volatile uint min;			// next key offset
-	volatile uint foster;		// count of foster children
+	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 lvl:7;		// level of page
- 	unsigned char dirty:1;		// page needs to be cleaned
+	unsigned char lvl:6;		// level of page
+	unsigned char kill:1;		// page is being deleted
+	unsigned char dirty:1;		// page has deleted keys
 	unsigned char right[BtId];	// page number to right
 } *BtPage;
 
-//	mode & definition for hash latch implementation
-
-enum {
-	Mutex = 1,
-	Write = 2,
-	Pending = 4,
-	Share = 8
-} LockMode;
-
-// mutex locks the other fields
-// exclusive is set for write access
-// share is count of read accessors
-
-typedef struct {
-	volatile ushort mutex:1;
-	volatile ushort exclusive:1;
-	volatile ushort pending:1;
-	volatile ushort share:13;
-} BtSpinLatch;
-
 //  hash table entries
 
 typedef struct {
-	BtSpinLatch latch[1];
+	BtLatch latch[1];
 	volatile ushort slot;		// Latch table entry at head of chain
 } BtHashEntry;
 
 //	latch manager table structure
 
 typedef struct {
-	BtSpinLatch readwr[1];		// read/write page lock
-	BtSpinLatch access[1];		// Access Intent/Page delete
-	BtSpinLatch parent[1];		// adoption of foster children
-	BtSpinLatch busy[1];		// slot is being moved between chains
+	BtLatch readwr[1];			// read/write page lock
+	BtLatch access[1];			// Access Intent/Page delete
+	BtLatch parent[1];			// adoption of foster children
+	BtLatch 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
@@ -178,8 +187,8 @@ typedef struct {
 	unsigned long long int lru;	// number of times accessed
 	uid  basepage;				// mapped base page number
 	char *map;					// mapped memory pointer
-	ushort 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
@@ -191,7 +200,7 @@ typedef struct {
 
 typedef struct {
 	struct Page alloc[2];		// next & free page_nos in right ptr
-	BtSpinLatch lock[1];		// allocation area lite latch
+	BtLatch 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
@@ -208,19 +217,19 @@ typedef struct {
 	uint seg_bits;				// seg size in pages in bits
 	uint mode;					// read-write mode
 #ifdef unix
-	int idx;
 	char *pooladvise;			// bit maps for pool page advisements
+	int idx;
 #else
 	HANDLE idx;
 #endif
 	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
 	ushort evicted;				// last evicted hash table slot
-	ushort *hash;				// hash table of pool entries
+	ushort hashsize;			// size of Hash Table for pool entries
+	ushort *hash;				// pool index for hash entries
 	BtPool *pool;				// memory pool page segments
-	BtSpinLatch *latch;			// latches for pool hash slots
+	BtLatch *latch;				// latches for pool hash slots
 	BtLatchMgr *latchmgr;		// mapped latch page from allocation page
 	BtLatchSet *latchsets;		// mapped latch set from latch pages
 #ifndef unix
@@ -232,11 +241,11 @@ typedef struct {
 	BtMgr *mgr;			// buffer manager for thread
 	BtPage cursor;		// cached frame for start/next (never mapped)
 	BtPage frame;		// spare frame for the page split (never mapped)
-	BtPage zero;		// page frame for zeroes at end of file
-	BtPage page;		// current page
+	BtPage zero;		// page of zeroes to extend the file (never mapped)
+	BtPage page;		// current page mapped from file
 	uid page_no;		// current page number	
 	uid cursor_page;	// current cursor page number	
-	BtLatchSet *set;	// current page latch set
+	BtLatchSet *set;	// current page latchset
 	BtPool *pool;		// current page pool
 	unsigned char *mem;	// frame, cursor, page memory buffer
 	int found;			// last delete was found
@@ -257,8 +266,8 @@ 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, uid id, uint tod, uint lvl);
-extern BTERR  bt_deletekey (BtDb *bt, unsigned char *key, uint len);
+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);
 extern uint bt_nextkey   (BtDb *bt, uint slot);
@@ -267,7 +276,7 @@ extern uint bt_nextkey   (BtDb *bt, uint slot);
 extern BtMgr *bt_mgr (char *name, uint mode, uint bits, uint poolsize, uint segsize, uint hashsize);
 void bt_mgrclose (BtMgr *mgr);
 
-//  Helper functions to return cursor slot values
+//  Helper functions to return slot values
 
 extern BtKey bt_key (BtDb *bt, uint slot);
 extern uid bt_uid (BtDb *bt, uint slot);
@@ -298,6 +307,10 @@ extern uint bt_tod (BtDb *bt, uint slot);
 //	The first leaf page of level zero is always
 //	located on page 2.
 
+//	The b-tree pages are linked with next
+//	pointers to facilitate enumerators,
+//	and provide for concurrency.
+
 //	When to root page fills, it is split in two and
 //	the tree height is raised by a new root at page
 //	one with two keys.
@@ -306,20 +319,22 @@ extern uint bt_tod (BtDb *bt, uint slot);
 //	page cleanup The fence key for a node is always
 //	present, even after deletion and cleanup.
 
-//  Groups of pages called segments from the btree are
-//  cached with memory mapping. A hash table is used to keep
+//  Groups of pages called segments from the btree are optionally
+//  cached with a memory mapped pool. A hash table is used to keep
 //  track of the cached segments.  This behaviour is controlled
 //  by the cache block size parameter to bt_open.
 
 //  To achieve maximum concurrency one page is locked at a time
-//  as the tree is traversed to find leaf key in question.
-
-//	An adoption traversal leaves the parent node locked as the
-//	tree is traversed to the level in quesiton.
+//  as the tree is traversed to find leaf key in question. The right
+//  page numbers are used in cases where the page is being split,
+//	or consolidated.
 
 //  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.
+
 //	Empty pages are chained together through the ALLOC page and reused.
 
 //	Access macros to address slot and key values from the page
@@ -346,76 +361,83 @@ int i;
 	return id;
 }
 
+//	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_spinreadlock(BtSpinLatch *latch)
+void bt_spinreadlock(BtLatch *latch, int private)
 {
-ushort prev;
+uint prev;
 
-  do {
-#ifdef unix
-	while( __sync_fetch_and_or((ushort *)latch, Mutex) & Mutex )
+  if( private )
+	private = FUTEX_PRIVATE_FLAG;
+
+  while( 1 ) {
+	//	obtain latch mutex
+	if( __sync_fetch_and_or((uint *)latch, Mutex) & Mutex ) {
 		sched_yield();
-#else
-	while( _InterlockedOr16((ushort *)latch, Mutex) & Mutex )
-		SwitchToThread();
-#endif
+		continue;
+	}
 
-	//  see if exclusive request is granted or pending
+	//  wait for writers to clear
+	//	increment read waiters and wait
 
-	if( prev = !(latch->exclusive | latch->pending) )
-#ifdef unix
-		__sync_fetch_and_add((ushort *)latch, Share);
-#else
-		_InterlockedExchangeAdd16 ((ushort *)latch, Share);
-#endif
+	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
 
-#ifdef unix
-	__sync_fetch_and_and ((ushort *)latch, ~Mutex);
-#else
-	_InterlockedAnd16((ushort *)latch, ~Mutex);
-#endif
-	if( prev )
-		return;
-#ifdef  unix
-  } while( sched_yield(), 1 );
-#else
-  } while( SwitchToThread(), 1 );
-#endif
+	__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_spinwritelock(BtSpinLatch *latch)
+void bt_spinwritelock(BtLatch *latch, int private)
 {
-  do {
-#ifdef  unix
-	while( __sync_fetch_and_or((ushort *)latch, Mutex | Pending) & Mutex )
+uint prev;
+
+  if( private )
+	private = FUTEX_PRIVATE_FLAG;
+
+  while( 1 ) {
+	//	obtain latch mutex
+	if( __sync_fetch_and_or((uint *)latch, Mutex) & Mutex ) {
 		sched_yield();
-#else
-	while( _InterlockedOr16((ushort *)latch, Mutex | Pending) & Mutex )
-		SwitchToThread();
-#endif
-	if( !(latch->share | latch->exclusive) ) {
-#ifdef unix
-		__sync_fetch_and_or((ushort *)latch, Write);
-		__sync_fetch_and_and ((ushort *)latch, ~(Mutex | Pending));
-#else
-		_InterlockedOr16((ushort *)latch, Write);
-		_InterlockedAnd16((ushort *)latch, ~(Mutex | Pending));
-#endif
-		return;
+		continue;
 	}
 
-#ifdef unix
-	__sync_fetch_and_and ((ushort *)latch, ~Mutex);
-	sched_yield();
-#else
-	_InterlockedAnd16((ushort *)latch, ~Mutex);
-	SwitchToThread();
-#endif
-  } while( 1 );
+	//	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
@@ -423,54 +445,81 @@ void bt_spinwritelock(BtSpinLatch *latch)
 //	return 1 if obtained,
 //		0 otherwise
 
-int bt_spinwritetry(BtSpinLatch *latch)
+int bt_spinwritetry(BtLatch *latch)
 {
-ushort prev;
+int ans;
 
-#ifdef unix
-	if( prev = __sync_fetch_and_or((ushort *)latch, Mutex), prev & Mutex )
-		return 0;
-#else
-	if( prev = _InterlockedOr16((ushort *)latch, Mutex), prev & Mutex )
+	//	try for mutex,
+	//	abandon request if not taken
+
+	if( __sync_fetch_and_or((uint *)latch, Mutex) & Mutex )
 		return 0;
-#endif
-	//	take write access if all bits are clear
 
-	if( !prev )
-#ifdef unix
-		__sync_fetch_and_or ((ushort *)latch, Write);
-#else
-		_InterlockedOr16((ushort *)latch, Write);
-#endif
+	//	see if write mode is available
 
-#ifdef unix
-	__sync_fetch_and_and ((ushort *)latch, ~Mutex);
-#else
-	_InterlockedAnd16((ushort *)latch, ~Mutex);
-#endif
-	return !prev;
+	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 mode
+//	clear write lock
 
-void bt_spinreleasewrite(BtSpinLatch *latch)
+void bt_spinreleasewrite(BtLatch *latch, int private)
 {
-#ifdef unix
-	__sync_fetch_and_and ((ushort *)latch, ~Write);
-#else
-	_InterlockedAnd16((ushort *)latch, ~Write);
-#endif
+  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_spinreleaseread(BtSpinLatch *latch)
+void bt_spinreleaseread(BtLatch *latch, int private)
 {
-#ifdef unix
-	__sync_fetch_and_add((ushort *)latch, -Share);
-#else
-	_InterlockedExchangeAdd16 ((ushort *)latch, -Share);
-#endif
+  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);
 }
 
 //	link latch table entry into latch hash table
@@ -510,7 +559,7 @@ BtLatchSet *set;
 
 	//  obtain read lock on hash table entry
 
-	bt_spinreadlock(bt->mgr->latchmgr->table[hashidx].latch);
+	bt_spinreadlock(bt->mgr->latchmgr->table[hashidx].latch, 0);
 
 	if( slot = bt->mgr->latchmgr->table[hashidx].slot ) do
 	{
@@ -527,14 +576,14 @@ BtLatchSet *set;
 #endif
 	}
 
-    bt_spinreleaseread (bt->mgr->latchmgr->table[hashidx].latch);
+    bt_spinreleaseread (bt->mgr->latchmgr->table[hashidx].latch, 0);
 
 	if( slot )
 		return set;
 
   //  try again, this time with write lock
 
-  bt_spinwritelock(bt->mgr->latchmgr->table[hashidx].latch);
+  bt_spinwritelock(bt->mgr->latchmgr->table[hashidx].latch, 0);
 
   if( slot = bt->mgr->latchmgr->table[hashidx].slot ) do
   {
@@ -555,7 +604,7 @@ BtLatchSet *set;
 	_InterlockedIncrement16 (&set->pin);
 #endif
 	set->page_no = page_no;
-	bt_spinreleasewrite(bt->mgr->latchmgr->table[hashidx].latch);
+	bt_spinreleasewrite(bt->mgr->latchmgr->table[hashidx].latch, 0);
 	return set;
   }
 
@@ -574,7 +623,7 @@ BtLatchSet *set;
 		_InterlockedIncrement16 (&set->pin);
 #endif
 		bt_latchlink (bt, hashidx, victim, page_no);
-		bt_spinreleasewrite (bt->mgr->latchmgr->table[hashidx].latch);
+		bt_spinreleasewrite (bt->mgr->latchmgr->table[hashidx].latch, 0);
 		return set;
 	}
 
@@ -611,13 +660,13 @@ BtLatchSet *set;
 	//	or has outstanding locks
 
 	if( !bt_spinwritetry (bt->mgr->latchmgr->table[idx].latch) ) {
-		bt_spinreleasewrite (set->busy);
+		bt_spinreleasewrite (set->busy, 0);
 		continue;
 	}
 
 	if( set->pin ) {
-		bt_spinreleasewrite (set->busy);
-		bt_spinreleasewrite (bt->mgr->latchmgr->table[idx].latch);
+		bt_spinreleasewrite (set->busy, 0);
+		bt_spinreleasewrite (bt->mgr->latchmgr->table[idx].latch, 0);
 		continue;
 	}
 
@@ -631,15 +680,15 @@ BtLatchSet *set;
 	if( set->next )
 		bt->mgr->latchsets[set->next].prev = set->prev;
 
-	bt_spinreleasewrite (bt->mgr->latchmgr->table[idx].latch);
+	bt_spinreleasewrite (bt->mgr->latchmgr->table[idx].latch, 0);
 #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);
+	bt_spinreleasewrite (bt->mgr->latchmgr->table[hashidx].latch, 0);
+	bt_spinreleasewrite (set->busy, 0);
 	return set;
   }
 }
@@ -668,7 +717,7 @@ uint slot;
 
 #ifdef unix
 	munmap (mgr->latchsets, mgr->latchmgr->nlatchpage * mgr->page_size);
-	munmap (mgr->latchmgr, mgr->page_size);
+ 	munmap (mgr->latchmgr, mgr->page_size);
 #else
 	FlushViewOfFile(mgr->latchmgr, 0);
 	UnmapViewOfFile(mgr->latchmgr);
@@ -720,7 +769,6 @@ uint amt[1];
 BtMgr* mgr;
 BtKey key;
 int flag;
-
 #ifndef unix
 SYSTEM_INFO sysinfo[1];
 #endif
@@ -737,7 +785,6 @@ SYSTEM_INFO sysinfo[1];
 
 #ifdef unix
 	mgr = calloc (1, sizeof(BtMgr));
-
 	mgr->idx = open ((char*)name, O_RDWR | O_CREAT, 0666);
 
 	if( mgr->idx == -1 )
@@ -770,7 +817,7 @@ SYSTEM_INFO sysinfo[1];
 		else
 			return free(mgr), free(latchmgr), NULL;
 	} else if( mode == BT_ro )
-		return free(latchmgr), bt_mgrclose (mgr), NULL;
+		return free(latchmgr), free (mgr), NULL;
 #else
 	latchmgr = VirtualAlloc(NULL, BT_maxpage, MEM_COMMIT, PAGE_READWRITE);
 	size = GetFileSize(mgr->idx, amt);
@@ -811,12 +858,12 @@ SYSTEM_INFO sysinfo[1];
 #ifdef unix
 	mgr->pool = calloc (poolmax, sizeof(BtPool));
 	mgr->hash = calloc (hashsize, sizeof(ushort));
-	mgr->latch = calloc (hashsize, sizeof(BtSpinLatch));
+	mgr->latch = calloc (hashsize, sizeof(BtLatch));
 	mgr->pooladvise = calloc (poolmax, (mgr->poolmask + 8) / 8);
 #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(BtSpinLatch));
+	mgr->latch = GlobalAlloc (GMEM_FIXED|GMEM_ZEROINIT, hashsize * sizeof(BtLatch));
 #endif
 
 	if( size || *amt )
@@ -900,6 +947,27 @@ SYSTEM_INFO sysinfo[1];
 		last++;
 	}
 
+	// create empty page area by writing last page of first
+	// segment area (other pages are zeroed by O/S)
+
+	if( mgr->poolmask ) {
+		memset(latchmgr, 0, mgr->page_size);
+		last = mgr->poolmask;
+
+		while( last < MIN_lvl + 1 )
+			last += mgr->poolmask + 1;
+
+#ifdef unix
+		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 *)latchmgr, mgr->page_size, amt, NULL) )
+			return bt_mgrclose (mgr), NULL;
+		if( *amt < mgr->page_size )
+			return bt_mgrclose (mgr), NULL;
+#endif
+	}
+
 mgrlatch:
 #ifdef unix
 	flag = PROT_READ | PROT_WRITE;
@@ -949,7 +1017,7 @@ BtDb *bt = malloc (sizeof(*bt));
 	bt->zero = (BtPage)(bt->mem + 1 * mgr->page_size);
 	bt->cursor = (BtPage)(bt->mem + 2 * mgr->page_size);
 
-	memset(bt->zero, 0, mgr->page_size);
+	memset (bt->zero, 0, mgr->page_size);
 	return bt;
 }
 
@@ -1034,7 +1102,7 @@ BtPool *pool = NULL, *node;
 
 	node = bt->mgr->pool + hashslot;
 
-	//	scan pool entries under hash table slot
+	//  scan pool entries under hash table slot
 
 	do {
 	  if( node->pin )
@@ -1061,6 +1129,7 @@ int flag;
 	pool->map = mmap (0, (bt->mgr->poolmask+1) << bt->mgr->page_bits, flag, MAP_SHARED, bt->mgr->idx, off);
 	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
@@ -1117,12 +1186,11 @@ BtPool *bt_pinpool(BtDb *bt, uid page_no)
 {
 BtPool *pool, *node, *next;
 uint slot, idx, victim;
-BtLatchSet *set;
 
 	//	lock hash table chain
 
 	idx = (uint)(page_no >> bt->mgr->seg_bits) % bt->mgr->hashsize;
-	bt_spinreadlock (&bt->mgr->latch[idx]);
+	bt_spinreadlock (&bt->mgr->latch[idx], 1);
 
 	//	look up in hash table
 
@@ -1132,15 +1200,15 @@ BtLatchSet *set;
 #else
 		_InterlockedIncrement16 (&pool->pin);
 #endif
-		bt_spinreleaseread (&bt->mgr->latch[idx]);
+		bt_spinreleaseread (&bt->mgr->latch[idx], 1);
 		pool->lru++;
 		return pool;
 	}
 
 	//	upgrade to write lock
 
-	bt_spinreleaseread (&bt->mgr->latch[idx]);
-	bt_spinwritelock (&bt->mgr->latch[idx]);
+	bt_spinreleaseread (&bt->mgr->latch[idx], 1);
+	bt_spinwritelock (&bt->mgr->latch[idx], 1);
 
 	// try to find page in pool with write lock
 
@@ -1150,7 +1218,7 @@ BtLatchSet *set;
 #else
 		_InterlockedIncrement16 (&pool->pin);
 #endif
-		bt_spinreleasewrite (&bt->mgr->latch[idx]);
+		bt_spinreleasewrite (&bt->mgr->latch[idx], 1);
 		pool->lru++;
 		return pool;
 	}
@@ -1177,7 +1245,7 @@ BtLatchSet *set;
 #else
 		_InterlockedIncrement16 (&pool->pin);
 #endif
-		bt_spinreleasewrite (&bt->mgr->latch[idx]);
+		bt_spinreleasewrite (&bt->mgr->latch[idx], 1);
 		return pool;
 	}
 
@@ -1204,12 +1272,12 @@ BtLatchSet *set;
 		if( !bt_spinwritetry (&bt->mgr->latch[victim]) )
 			continue;
 
-		//  if cache entry is empty
-		//	or no slots are unpinned
+		//  if pool entry is empty
+		//	or any pages are pinned
 		//	skip this entry
 
 		if( !(pool = bt_findlru(bt, bt->mgr->hash[victim])) ) {
-			bt_spinreleasewrite (&bt->mgr->latch[victim]);
+			bt_spinreleasewrite (&bt->mgr->latch[victim], 1);
 			continue;
 		}
 
@@ -1225,7 +1293,7 @@ BtLatchSet *set;
 		if( node = pool->hashnext )
 			node->hashprev = pool->hashprev;
 
-		bt_spinreleasewrite (&bt->mgr->latch[victim]);
+		bt_spinreleasewrite (&bt->mgr->latch[victim], 1);
 
 		//	remove old file mapping
 #ifdef unix
@@ -1249,54 +1317,54 @@ BtLatchSet *set;
 #else
 		_InterlockedIncrement16 (&pool->pin);
 #endif
-		bt_spinreleasewrite (&bt->mgr->latch[idx]);
+		bt_spinreleasewrite (&bt->mgr->latch[idx], 1);
 		return pool;
 	}
 }
 
 // place write, read, or parent lock on requested page_no.
-//	pin to buffer pool and return latchset pointer
+//	pin to buffer pool and return page pointer
 
 void bt_lockpage(BtLock mode, BtLatchSet *set)
 {
 	switch( mode ) {
 	case BtLockRead:
-		bt_spinreadlock (set->readwr);
+		bt_spinreadlock (set->readwr, 0);
 		break;
 	case BtLockWrite:
-		bt_spinwritelock (set->readwr);
+		bt_spinwritelock (set->readwr, 0);
 		break;
 	case BtLockAccess:
-		bt_spinreadlock (set->access);
+		bt_spinreadlock (set->access, 0);
 		break;
 	case BtLockDelete:
-		bt_spinwritelock (set->access);
+		bt_spinwritelock (set->access, 0);
 		break;
 	case BtLockParent:
-		bt_spinwritelock (set->parent);
+		bt_spinwritelock (set->parent, 0);
 		break;
 	}
 }
 
-// remove write, read, or parent lock on requested page_no.
+// remove write, read, or parent lock on requested page
 
 void bt_unlockpage(BtLock mode, BtLatchSet *set)
 {
 	switch( mode ) {
 	case BtLockRead:
-		bt_spinreleaseread (set->readwr);
+		bt_spinreleaseread (set->readwr, 0);
 		break;
 	case BtLockWrite:
-		bt_spinreleasewrite (set->readwr);
+		bt_spinreleasewrite (set->readwr, 0);
 		break;
 	case BtLockAccess:
-		bt_spinreleaseread (set->access);
+		bt_spinreleaseread (set->access, 0);
 		break;
 	case BtLockDelete:
-		bt_spinreleasewrite (set->access);
+		bt_spinreleasewrite (set->access, 0);
 		break;
 	case BtLockParent:
-		bt_spinreleasewrite (set->parent);
+		bt_spinreleasewrite (set->parent, 0);
 		break;
 	}
 }
@@ -1313,7 +1381,7 @@ int reuse;
 
 	//	lock allocation page
 
-	bt_spinwritelock(bt->mgr->latchmgr->lock);
+	bt_spinwritelock(bt->mgr->latchmgr->lock, 0);
 
 	// use empty chain first
 	// else allocate empty page
@@ -1332,30 +1400,21 @@ int reuse;
 		reuse = 0;
 	}
 #ifdef unix
+	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;
+
 	// 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
+		memset(bt->zero, 0, bt->mgr->page_size);
 		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;
 	}
-
-	// unlock allocation latch
-
-	bt_spinreleasewrite(bt->mgr->latchmgr->lock);
-
-	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;
-
 #else
-	// unlock allocation latch
-
-	bt_spinreleasewrite(bt->mgr->latchmgr->lock);
-
 	//	bring new page into pool and copy page.
 	//	this will extend the file into the new pages.
-	//	NB -- no latch required
 
 	if( pool = bt_pinpool (bt, new_page) )
 		pmap = bt_page (bt, pool, new_page);
@@ -1365,6 +1424,9 @@ int reuse;
 	memcpy(pmap, page, bt->mgr->page_size);
 	bt_unpinpool (pool);
 #endif
+	// unlock allocation latch and return new page no
+
+	bt_spinreleasewrite(bt->mgr->latchmgr->lock, 0);
 	return new_page;
 }
 
@@ -1373,8 +1435,16 @@ int reuse;
 int bt_findslot (BtDb *bt, unsigned char *key, uint len)
 {
 uint diff, higher = bt->page->cnt, low = 1, slot;
+uint good = 0;
+
+	//	make stopper key an infinite fence value
 
-	//	low is the lowest candidate, higher is already
+	if( bt_getid (bt->page->right) )
+		higher++;
+	else
+		good++;
+
+	//	low is the next candidate, higher is already
 	//	tested as .ge. the given key, loop ends when they meet
 
 	while( diff = higher - low ) {
@@ -1382,16 +1452,18 @@ uint diff, higher = bt->page->cnt, low = 1, slot;
 		if( keycmp (keyptr(bt->page, slot), key, len) < 0 )
 			low = slot + 1;
 		else
-			higher = slot;
+			higher = slot, good++;
 	}
 
-	return higher;
+	//	return zero if key is on right link page
+
+ 	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, BtLock lock)
+int bt_loadpage (BtDb *bt, unsigned char *key, uint len, uint lvl, uint lock)
 {
 uid page_no = ROOT_page, prevpage = 0;
 BtLatchSet *set, *prevset;
@@ -1401,12 +1473,12 @@ BtPool *prevpool;
 
   //  start at root of btree and drill down
 
+  bt->set = NULL;
+
   do {
 	// determine lock mode of drill level
 	mode = (lock == BtLockWrite) && (drill == lvl) ? BtLockWrite : BtLockRead; 
 
-	//	obtain latch set for this page
-
 	bt->set = bt_pinlatch (bt, page_no);
 	bt->page_no = page_no;
 
@@ -1422,7 +1494,7 @@ BtPool *prevpool;
 	if( page_no > ROOT_page )
 	  bt_lockpage(BtLockAccess, bt->set);
 
-	//  now unlock and unpin our (possibly foster) parent
+	//	release & unpin parent page
 
 	if( prevpage ) {
 	  bt_unlockpage(prevmode, prevset);
@@ -1440,184 +1512,113 @@ BtPool *prevpool;
 
 	// re-read and re-lock root after determining actual level of root
 
-	if( page_no == ROOT_page )
-	  if( bt->page->lvl != drill) {
+	if( bt->page->lvl != drill) {
+		if ( bt->page_no != ROOT_page )
+			return bt->err = BTERR_struct, 0;
+			
 		drill = bt->page->lvl;
 
-	    if( lock == BtLockWrite && drill == lvl ) {
+		if( lock == BtLockWrite && drill == lvl ) {
 		  bt_unlockpage(mode, bt->set);
 		  bt_unpinlatch (bt->set);
 		  bt_unpinpool (bt->pool);
 		  continue;
 		}
-	  }
-
-	prevpage = bt->page_no;
-	prevpool = bt->pool;
-	prevset = bt->set;
-	prevmode = mode;
+	}
 
 	//  find key on page at this level
-	//  and either descend to requested level
-	//	or return key slot
-
-	slot = bt_findslot (bt, key, len);
+	//  and descend to requested level
 
-	//	is this slot < foster child area
-	//	on the requested level?
-
-	//	if so, return actual slot even if dead
-
-	if( slot <= bt->page->cnt - bt->page->foster )
+	if( !bt->page->kill && (slot = bt_findslot (bt, key, len)) ) {
 	  if( drill == lvl )
 		return slot;
 
-	//	find next active slot
-
-	//	note: foster children are never dead
-	//	nor fence keys for interiour nodes
-
-	while( slotptr(bt->page, slot)->dead )
-	  if( slot++ < bt->page->cnt )
-		continue;
-	  else
-		return bt->err = BTERR_struct, 0;	// last key shouldn't be deleted
+	  while( slotptr(bt->page, slot)->dead )
+		if( slot++ < bt->page->cnt )
+			continue;
+		else {
+			page_no = bt_getid(bt->page->right);
+			goto slideright;
+		}
 
-	//	is this slot < foster child area
-	//	if so, drill to next level
+	  page_no = bt_getid(slotptr(bt->page, slot)->id);
+	  drill--;
+	}
 
-	if( slot <= bt->page->cnt - bt->page->foster )
-		drill--;
+	//  or slide right into next page
+	//  (slide left from deleted page)
 
-	//  continue right onto foster child
-	//	or down to next level.
+	else
+		page_no = bt_getid(bt->page->right);
 
-	page_no = bt_getid(slotptr(bt->page, slot)->id);
+	//  continue down / right using overlapping locks
+	//  to protect pages being killed or split.
 
+slideright:
+	prevpage = bt->page_no;
+	prevpool = bt->pool;
+	prevset = bt->set;
+	prevmode = mode;
   } while( page_no );
 
-  // return error on end of chain
+  // return error on end of right chain
 
   bt->err = BTERR_struct;
   return 0;	// return error
 }
 
 //  find and delete key on page by marking delete flag bit
-//  when leaf page becomes empty, delete it from the btree
+//  when page becomes empty, delete it
 
-BTERR bt_deletekey (BtDb *bt, unsigned char *key, uint len)
+BTERR bt_deletekey (BtDb *bt, unsigned char *key, uint len, uint lvl)
 {
-unsigned char leftkey[256];
+unsigned char lowerkey[256], higherkey[256];
 BtLatchSet *rset, *set;
 BtPool *pool, *rpool;
-BtPage rpage, page;
 uid page_no, right;
 uint slot, tod;
+BtPage rpage;
 BtKey ptr;
 
-	if( slot = bt_loadpage (bt, key, len, 0, BtLockWrite) )
+	if( slot = bt_loadpage (bt, key, len, lvl, BtLockWrite) )
 		ptr = keyptr(bt->page, slot);
 	else
 		return bt->err;
 
 	// if key is found delete it, otherwise ignore request
-	// note that fence keys of interiour nodes are not deleted.
 
 	if( bt->found = !keycmp (ptr, key, len) )
 		if( bt->found = slotptr(bt->page, slot)->dead == 0 ) {
- 			slotptr(bt->page,slot)->dead = 1;
+			slotptr(bt->page,slot)->dead = 1;
 			if( slot < bt->page->cnt )
- 				bt->page->dirty = 1;
- 			bt->page->act--;
+				bt->page->dirty = 1;
+			bt->page->act--;
 		}
 
+	// return if page is not empty, or it has no right sibling
+
+	right = bt_getid(bt->page->right);
 	page_no = bt->page_no;
 	pool = bt->pool;
-	page = bt->page;
 	set = bt->set;
 
-	// return if page is not empty or not found
-
-	if( page->act || !bt->found ) {
+	if( !right || bt->page->act ) {
 		bt_unlockpage(BtLockWrite, set);
 		bt_unpinlatch (set);
 		bt_unpinpool (pool);
 		return bt->err;
 	}
 
-	// cache copy of fence key of empty node
-
-	ptr = keyptr(page, page->cnt);
-	memcpy(leftkey, ptr, ptr->len + 1);
+	// obtain Parent lock over write lock
 
-	//	release write lock on empty node
-	//	obtain Parent lock
-
-	bt_unlockpage(BtLockWrite, set);
 	bt_lockpage(BtLockParent, set);
 
-	//	load and lock parent to see
-	//  if delete of empty node is OK
-	//	ie, not a fence key of parent
-
-	while( 1 ) {
-	  if( slot = bt_loadpage (bt, leftkey+1, *leftkey, 1, BtLockWrite) )
-		ptr = keyptr(bt->page, slot);
-	  else
-		return bt->err;
-
-	  // does parent level contain our fence key yet?
-	  // and is it free of foster children?
-
-	  if( !bt->page->foster )
-	   if( !keycmp (ptr, leftkey+1, *leftkey) )
-		break;
-
-	  bt_unlockpage(BtLockWrite, bt->set);
-	  bt_unpinlatch (bt->set);
-	  bt_unpinpool (bt->pool);
-#ifdef unix
-	  sched_yield();
-#else
-	  SwitchToThread();
-#endif
-	}
-
-	//	find our left fence key
-
-	while( slotptr(bt->page, slot)->dead )
-	  if( slot++ < bt->page->cnt )
-		continue;
-	  else
-		return bt->err = BTERR_struct;	// last key shouldn't be deleted
-
-	//	now we have both parent and child
-
-	bt_lockpage(BtLockDelete, set);
-	bt_lockpage(BtLockWrite, set);
+	// keep copy of key to delete
 
-	// return if page has no right sibling within parent
-	//	 or if empty node is no longer empty
-
-	if( page->act || slot == bt->page->cnt ) {
-		// unpin parent
-		bt_unlockpage(BtLockWrite, bt->set);
-		bt_unpinlatch (bt->set);
-		bt_unpinpool (bt->pool);
-		// unpin empty node
-		bt_unlockpage(BtLockParent, set);
-		bt_unlockpage(BtLockDelete, set);
-		bt_unlockpage(BtLockWrite, set);
-		bt_unpinlatch (set);
-		bt_unpinpool (pool);
-		return bt->err;
-	}
+	ptr = keyptr(bt->page, bt->page->cnt);
+	memcpy(lowerkey, ptr, ptr->len + 1);
 
-	// lock and map our right page
-	// note that it cannot be our foster child
-	// since the our node is empty
-
-	right = bt_getid(page->right);
+	// lock and map right page
 
 	if( rpool = bt_pinpool (bt, right) )
 		rpage = bt_page (bt, rpool, right);
@@ -1626,37 +1627,42 @@ BtKey ptr;
 
 	rset = bt_pinlatch (bt, right);
 	bt_lockpage(BtLockWrite, rset);
-	bt_lockpage(BtLockDelete, rset);
-
-	// pull contents of right page into empty page 
 
-	memcpy (page, rpage, bt->mgr->page_size);
+	// pull contents of next page into current empty page 
 
-	//	delete left parent slot for old empty page
-	//	and redirect right parent slot to it
+	memcpy (bt->page, rpage, bt->mgr->page_size);
 
-	bt->page->act--;
-	bt->page->dirty = 1;
-	slotptr(bt->page, slot)->dead = 1;
+	//	keep copy of key to update
 
-	while( slot++ < bt->page->cnt )
-	  if( !slotptr(bt->page, slot)->dead )
-		break;
+	ptr = keyptr(rpage, rpage->cnt);
+	memcpy(higherkey, ptr, ptr->len + 1);
 
-	bt_putid(slotptr(bt->page,slot)->id, page_no);
+	//  Mark right page as deleted and point it to left page
+	//	until we can post updates at higher level.
 
-	// release parent level lock
-	//	and our empty node lock
+	bt_putid(rpage->right, page_no);
+	rpage->kill = 1;
+	rpage->cnt = 0;
 
+	bt_unlockpage(BtLockWrite, rset);
 	bt_unlockpage(BtLockWrite, set);
-	bt_unlockpage(BtLockWrite, bt->set);
-	bt_unpinlatch (bt->set);
-	bt_unpinpool (bt->pool);
+
+	//  delete old lower key to consolidated node
+
+	if( bt_deletekey (bt, lowerkey + 1, *lowerkey, lvl + 1) )
+		return bt->err;
+
+	//  redirect higher key directly to consolidated node
+
+	tod = (uint)time(NULL);
+
+	if( bt_insertkey (bt, higherkey+1, *higherkey, lvl + 1, page_no, tod) )
+		return bt->err;
 
 	//	add killed right block to free chain
 	//	lock latch mgr
 
-	bt_spinwritelock(bt->mgr->latchmgr->lock);
+	bt_spinwritelock(bt->mgr->latchmgr->lock, 0);
 
 	//	store free chain in allocation page second right
 	bt_putid(rpage->right, bt_getid(bt->mgr->latchmgr->alloc[1].right));
@@ -1664,7 +1670,7 @@ BtKey ptr;
 
 	// unlock latch mgr and right page
 
-	bt_spinreleasewrite(bt->mgr->latchmgr->lock);
+	bt_spinreleasewrite(bt->mgr->latchmgr->lock, 0);
 
 	bt_unlockpage(BtLockWrite, rset);
 	bt_unlockpage(BtLockDelete, rset);
@@ -1674,11 +1680,10 @@ BtKey ptr;
 	// 	remove ParentModify lock
 
 	bt_unlockpage(BtLockParent, set);
-	bt_unlockpage(BtLockDelete, set);
 	bt_unpinlatch (set);
 	bt_unpinpool (pool);
 	return 0;
-} 
+}
 
 //	find key in leaf level and return row-id
 
@@ -1696,7 +1701,7 @@ uid id;
 	// if key exists, return row-id
 	//	otherwise return 0
 
-	if( slot <= bt->page->cnt && !keycmp (ptr, key, len) )
+	if( ptr->len == len && !memcmp (ptr->key, key, len) )
 		id = bt_getid(slotptr(bt->page,slot)->id);
 	else
 		id = 0;
@@ -1709,8 +1714,8 @@ uid id;
 
 //	check page for space available,
 //	clean if necessary and return
-//	0 - page needs splitting
-//	>0  new slot value
+//	=0 - page needs splitting
+//	>0 - go ahead at returned slot
 
 uint bt_cleanpage(BtDb *bt, uint amt, uint slot)
 {
@@ -1737,13 +1742,7 @@ BtKey key;
 	page->dirty = 0;
 	page->act = 0;
 
-	// try cleaning up page first
-
-	// always leave fence key in the array
-	// otherwise, remove deleted key
-
-	// note: foster children are never dead
-	//	nor are fence keys for interiour nodes
+	// always leave fence key in list
 
 	while( cnt++ < max ) {
 		if( cnt == slot )
@@ -1752,7 +1751,6 @@ BtKey key;
 			continue;
 
 		// copy key
-
 		key = keyptr(bt->frame, cnt);
 		nxt -= key->len + 1;
 		memcpy ((unsigned char *)page + nxt, key, key->len + 1);
@@ -1764,108 +1762,48 @@ BtKey key;
 		slotptr(page, idx)->tod = slotptr(bt->frame, cnt)->tod;
 		slotptr(page, idx)->off = nxt;
 	}
-
 	page->min = nxt;
 	page->cnt = idx;
 
-	//	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;
 }
 
-//	add key to current page
-//	page must already be writelocked
-
-void bt_addkeytopage (BtDb *bt, uint slot, unsigned char *key, uint len, uid id, uint tod)
-{
-BtPage page = bt->page;
-uint idx;
-
-	// find next available dead slot and copy key onto page
-	// note that foster children on the page are never dead
-
-	// look for next hole, but stay back from the fence key
-
-	for( idx = slot; idx < page->cnt; idx++ )
-	  if( slotptr(page, idx)->dead )
-		break;
-
-	if( idx == page->cnt )
-		idx++, page->cnt++;
-
-	page->act++;
-
-	// now insert key into array before slot
-
-	while( idx > slot )
-		*slotptr(page, idx) = *slotptr(page, idx -1), idx--;
-
-	page->min -= len + 1;
-	((unsigned char *)page)[page->min] = len;
-	memcpy ((unsigned char *)page + page->min +1, key, len );
-
-	bt_putid(slotptr(page,slot)->id, id);
-	slotptr(page, slot)->off = page->min;
-	slotptr(page, slot)->tod = tod;
-	slotptr(page, slot)->dead = 0;
-}
-
 // split the root and raise the height of the btree
-//	call with current page locked and page no of foster child
-//	return with current page (root) unlocked
 
-BTERR bt_splitroot(BtDb *bt, uid right)
+BTERR bt_splitroot(BtDb *bt,  unsigned char *newkey, unsigned char *oldkey, uid page_no2)
 {
 uint nxt = bt->mgr->page_size;
-unsigned char fencekey[256];
 BtPage root = bt->page;
 uid new_page;
-BtKey key;
 
-	//  Obtain an empty page to use, and copy the left page
-	//  contents into it from the root.  Strip foster child key.
-	//	(it's the stopper key)
-
-	memset (slotptr(root, root->cnt), 0, sizeof(BtSlot));
-	root->dirty = 1;
-	root->foster--;
-	root->act--;
-	root->cnt--;
-
-	//	Save left fence key.
-
-	key = keyptr(root, root->cnt);
-	memcpy (fencekey, key, key->len + 1);
-
-	//  copy the lower keys into a new left page
+	//  Obtain an empty page to use, and copy the current
+	//  root contents into it which is the lower half of
+	//	the old root.
 
 	if( !(new_page = bt_newpage(bt, root)) )
 		return bt->err;
 
 	// preserve the page info at the bottom
-	// and set rest of the root to zero
+	// and set rest to zero
 
-	memset (root+1, 0, bt->mgr->page_size - sizeof(*root));
+	memset(root+1, 0, bt->mgr->page_size - sizeof(*root));
 
-	// insert left fence key on empty newroot page
+	// insert first key on newroot page
 
-	nxt -= *fencekey + 1;
-	memcpy ((unsigned char *)root + nxt, fencekey, *fencekey + 1);
+	nxt -= *newkey + 1;
+	memcpy ((unsigned char *)root + nxt, newkey, *newkey + 1);
 	bt_putid(slotptr(root, 1)->id, new_page);
 	slotptr(root, 1)->off = nxt;
 	
-	// insert stopper key on newroot page
+	// insert second key on newroot page
 	// and increase the root height
 
-	nxt -= 3;
-	fencekey[0] = 2;
-	fencekey[1] = 0xff;
-	fencekey[2] = 0xff;
-	memcpy ((unsigned char *)root + nxt, fencekey, *fencekey + 1);
-	bt_putid(slotptr(root, 2)->id, right);
+	nxt -= *oldkey + 1;
+	memcpy ((unsigned char *)root + nxt, oldkey, *oldkey + 1);
+	bt_putid(slotptr(root, 2)->id, page_no2);
 	slotptr(root, 2)->off = nxt;
 
 	bt_putid(root->right, 0);
@@ -1883,34 +1821,31 @@ BtKey key;
 }
 
 //  split already locked full node
-//	in current page variables
-//	return unlocked and unpinned.
+//	return unlocked.
 
 BTERR bt_splitpage (BtDb *bt)
 {
-uint slot, cnt, idx, max, nxt = bt->mgr->page_size;
-unsigned char fencekey[256];
-uid page_no = bt->page_no;
-BtLatchSet *set = bt->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;
+BtLatchSet *nset, *set = bt->set;
 BtPool *pool = bt->pool;
 BtPage page = bt->page;
-uint tod = time(NULL);
 uint lvl = page->lvl;
-uid new_page, right;
+uid new_page;
 BtKey key;
+uint tod;
 
-	//	initialize frame buffer for right node
+	//  split higher half of keys to bt->frame
+	//	the last key (fence key) might be dead
 
-	memset (bt->frame, 0, bt->mgr->page_size);
-	max = page->cnt - page->foster;
 	tod = (uint)time(NULL);
+
+	memset (bt->frame, 0, bt->mgr->page_size);
+	max = (int)page->cnt;
 	cnt = max / 2;
 	idx = 0;
 
-	//  split higher half of keys to bt->frame
-	//	leaving old foster children in the left node,
-	//	and adding a new foster child there.
-
 	while( cnt++ < max ) {
 		key = keyptr(page, cnt);
 		nxt -= key->len + 1;
@@ -1922,196 +1857,165 @@ BtKey key;
 		slotptr(bt->frame, idx)->off = nxt;
 	}
 
-	// transfer right link node to new right node
+	// remember existing fence key for new page to the right
 
-	if( page_no > ROOT_page ) {
-		right = bt_getid (page->right);
-		bt_putid(bt->frame->right, right);
-	}
+	memcpy (oldkey, key, key->len + 1);
 
 	bt->frame->bits = bt->mgr->page_bits;
 	bt->frame->min = nxt;
 	bt->frame->cnt = idx;
 	bt->frame->lvl = lvl;
 
-	//	get new free page and write right frame to it.
+	// link right node
 
-	if( !(new_page = bt_newpage(bt, bt->frame)) )
-		return bt->err;
+	if( page_no > ROOT_page ) {
+		right = bt_getid (page->right);
+		bt_putid(bt->frame->right, right);
+	}
 
-	//	remember fence key for new right page to add
-	//	as foster child to the left node
+	//	get new free page and write frame to it.
 
-	key = keyptr(bt->frame, idx);
-	memcpy (fencekey, key, key->len + 1);
+	if( !(new_page = bt_newpage(bt, bt->frame)) )
+		return bt->err;
 
-	//	update lower keys and foster children to continue in old page
+	//	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));
 	nxt = bt->mgr->page_size;
-	page->dirty = 0;
 	page->act = 0;
 	cnt = 0;
 	idx = 0;
 
 	//  assemble page of smaller keys
-	//	to remain in the old page
+	//	(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);
-		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;
-	}
-
-	//	insert new foster child for right page in queue
-	//	before any of the current foster children
-
-	nxt -= *fencekey + 1;
-	memcpy ((unsigned char *)page + nxt, fencekey, *fencekey + 1);
-
-	bt_putid (slotptr(page,++idx)->id, new_page);
-	slotptr(page, idx)->tod = tod;
-	slotptr(page, idx)->off = nxt;
-	page->foster++;
-	page->act++;
-
-	//  continue with old foster child keys
-	//	note that none will be dead
-
-	cnt = bt->frame->cnt - bt->frame->foster;
-
-	while( cnt++ < bt->frame->cnt ) {
-		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);
+		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++;
 	}
 
+	// remember fence key for old page
+
+	memcpy(lowerkey, key, key->len + 1);
+	bt_putid(page->right, new_page);
 	page->min = nxt;
 	page->cnt = idx;
 
-	//	link new right page
-
-	bt_putid (page->right, new_page);
-
 	// if current page is the root page, split it
 
 	if( page_no == ROOT_page )
-		return bt_splitroot (bt, new_page);
+		return bt_splitroot (bt, lowerkey, oldkey, new_page);
 
-	//  release wr lock on our page
+	//  release wr lock on left page
 
 	bt_unlockpage (BtLockWrite, set);
 
-	//  obtain ParentModification lock for current page
-	//	to fix new fence key and oldest foster child on page
+	// obtain Parent/Write locks
+	// for left and right node pages
 
-	bt_lockpage (BtLockParent, set);
+	nset = bt_pinlatch (bt, new_page);
 
-	//  get our new fence key to insert in parent node
-
-	bt_lockpage (BtLockRead, set);
-
-	key = keyptr(page, page->cnt-1);
-	memcpy (fencekey, key, key->len+1);
+	bt_lockpage (BtLockParent, nset);
+	bt_lockpage (BtLockParent, set);
 
-	bt_unlockpage (BtLockRead, set);
+	// insert new fence for reformulated left block
 
-	if( bt_insertkey (bt, fencekey + 1, *fencekey, page_no, tod, lvl + 1) )
+	if( bt_insertkey (bt, lowerkey+1, *lowerkey, lvl + 1, page_no, tod) )
 		return bt->err;
 
-	//	lock our page for writing
-
-	bt_lockpage (BtLockRead, set);
-
-	//	switch old parent key from us to our oldest foster child
+	// fix old fence for newly allocated right block page
 
-	key = keyptr(page, page->cnt);
-	memcpy (fencekey, key, key->len+1);
-
-	new_page = bt_getid (slotptr(page, page->cnt)->id);
-	bt_unlockpage (BtLockRead, set);
-
-	if( bt_insertkey (bt, fencekey + 1, *fencekey, new_page, tod, lvl + 1) )
+	if( bt_insertkey (bt, oldkey+1, *oldkey, lvl + 1, new_page, tod) )
 		return bt->err;
 
-	//	now that it has its own parent pointer,
-	//	remove oldest foster child from our page
-
-	bt_lockpage (BtLockWrite, set);
-	memset (slotptr(page, page->cnt), 0, sizeof(BtSlot));
-	page->dirty = 1;
-	page->foster--;
-	page->cnt--;
-	page->act--;
+	// release Parent locks
 
-	//	unlock and unpin
-
-	bt_unlockpage (BtLockWrite, set);
+	bt_unlockpage (BtLockParent, nset);
 	bt_unlockpage (BtLockParent, set);
+	bt_unpinlatch (nset);
 	bt_unpinlatch (set);
 	bt_unpinpool (pool);
 	return 0;
 }
 
-//  Insert new key into the btree at leaf level.
+//  Insert new key into the btree at requested level.
+//  Level zero pages are leaf pages. Page is unlocked at exit.
 
-BTERR bt_insertkey (BtDb *bt, unsigned char *key, uint len, uid id, uint tod, uint lvl)
+BTERR bt_insertkey (BtDb *bt, unsigned char *key, uint len, uint lvl, uid id, uint tod)
 {
 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;
-		}
+  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
+	// if key already exists, update id and return
 
-		page = bt->page;
+	page = bt->page;
 
-		if( !keycmp (ptr, key, len) ) {
-			if( slotptr(page, slot)->dead )
-				page->act++;
-			slotptr(page, slot)->dead = 0;
-			slotptr(page, slot)->tod = tod;
-			bt_putid(slotptr(page,slot)->id, id);
-			bt_unlockpage(BtLockWrite, bt->set);
-			bt_unpinlatch (bt->set);
-			bt_unpinpool (bt->pool);
-			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);
+		bt_unlockpage(BtLockWrite, bt->set);
+		bt_unpinlatch(bt->set);
+		bt_unpinpool (bt->pool);
+		return bt->err;
+	}
 
-		// check if page has enough space
+	// check if page has enough space
 
- 		if( slot = bt_cleanpage (bt, len, slot) )
-			break;
+	if( slot = bt_cleanpage (bt, len, slot) )
+		break;
 
-		if( bt_splitpage (bt) )
-			return bt->err;
-	}
+	if( bt_splitpage (bt) )
+		return bt->err;
+  }
 
-  	bt_addkeytopage (bt, slot, key, len, id, tod);
+  // calculate next available slot and copy key into page
 
-	bt_unlockpage (BtLockWrite, bt->set);
-	bt_unpinlatch (bt->set);
-	bt_unpinpool (bt->pool);
-	return 0;
+  page->min -= len + 1; // reset lowest used offset
+  ((unsigned char *)page)[page->min] = len;
+  memcpy ((unsigned char *)page + page->min +1, key, len );
+
+  for( idx = slot; idx < page->cnt; idx++ )
+	if( slotptr(page, idx)->dead )
+		break;
+
+  // now insert key into array before slot
+  // preserving the fence slot
+
+  if( idx == page->cnt )
+	idx++, page->cnt++;
+
+  page->act++;
+
+  while( idx > slot )
+	*slotptr(page, idx) = *slotptr(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_unlockpage (BtLockWrite, bt->set);
+  bt_unpinlatch (bt->set);
+  bt_unpinpool (bt->pool);
+  return 0;
 }
 
 //  cache page of keys into cursor and return starting slot for given key
@@ -2123,9 +2027,7 @@ 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;
-
 	bt_unlockpage(BtLockRead, bt->set);
 	bt_unpinlatch (bt->set);
 	bt_unpinpool (bt->pool);
@@ -2137,17 +2039,16 @@ uint slot;
 
 uint bt_nextkey (BtDb *bt, uint slot)
 {
-BtLatchSet *set;
 BtPool *pool;
 BtPage page;
 uid right;
 
   do {
 	right = bt_getid(bt->cursor->right);
-	while( slot++ < bt->cursor->cnt - bt->cursor->foster )
+	while( slot++ < bt->cursor->cnt )
 	  if( slotptr(bt->cursor,slot)->dead )
 		continue;
-	  else if( right || (slot < bt->cursor->cnt - bt->cursor->foster) )
+	  else if( right || (slot < bt->cursor->cnt))
 		return slot;
 	  else
 		break;
@@ -2156,18 +2057,19 @@ uid right;
 		break;
 
 	bt->cursor_page = right;
+
 	if( pool = bt_pinpool (bt, right) )
 		page = bt_page (bt, pool, right);
 	else
 		return 0;
 
-	set = bt_pinlatch (bt, right);
-    bt_lockpage(BtLockRead, set);
+	bt->set = bt_pinlatch (bt, right);
+    bt_lockpage(BtLockRead, bt->set);
 
 	memcpy (bt->cursor, page, bt->mgr->page_size);
 
-	bt_unlockpage(BtLockRead, set);
-	bt_unpinlatch (set);
+	bt_unlockpage(BtLockRead, bt->set);
+	bt_unpinlatch (bt->set);
 	bt_unpinpool (pool);
 	slot = 0;
   } while( 1 );
@@ -2190,7 +2092,6 @@ uint bt_tod(BtDb *bt, uint slot)
 	return slotptr(bt->cursor,slot)->tod;
 }
 
-
 #ifdef STANDALONE
 
 void bt_latchaudit (BtDb *bt)
@@ -2217,11 +2118,11 @@ uid page_no;
 	}
 
 	for( hashidx = 0; hashidx < bt->mgr->latchmgr->latchhash; hashidx++ ) {
-	  if( *(ushort *)bt->mgr->latchmgr->table[hashidx].latch )
+	  if( *(uint *)bt->mgr->latchmgr->table[hashidx].latch )
 		fprintf(stderr, "latchmgr locked\n");
 	  if( idx = bt->mgr->latchmgr->table[hashidx].slot ) do {
 		set = bt->mgr->latchsets + idx;
-		if( *(ushort *)set->readwr || *(ushort *)set->access || *(ushort *)set->parent )
+		if( *(uint *)set->readwr || *(ushort *)set->access || *(ushort *)set->parent )
 			fprintf(stderr, "latchset %d locked\n", idx);
 		if( set->hash != hashidx )
 			fprintf(stderr, "latchset %d wrong hashidx\n", idx);
@@ -2262,7 +2163,6 @@ uid next, page_no = LEAF_page;	// start on first page of leaves
 unsigned char key[256];
 ThreadArg *args = arg;
 int ch, len = 0, slot;
-BtLatchSet *set;
 time_t tod[1];
 BtPool *pool;
 BtPage page;
@@ -2280,6 +2180,7 @@ FILE *in;
 		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") )
@@ -2294,7 +2195,7 @@ FILE *in;
 			  else if( args->num )
 		  		sprintf((char *)key+len, "%.9d", line + args->idx * args->num), len += 9;
 
-			  if( bt_insertkey (bt, key, len, line, *tod, 0) )
+			  if( bt_insertkey (bt, key, len, 0, line, *tod) )
 				fprintf(stderr, "Error %d Line: %d\n", bt->err, line), exit(0);
 			  len = 0;
 			}
@@ -2316,7 +2217,7 @@ FILE *in;
 			  else if( args->num )
 		  		sprintf((char *)key+len, "%.9d", line + args->idx * args->num), len += 9;
 
-			  if( bt_deletekey (bt, key, len) )
+			  if( bt_deletekey (bt, key, len, 0) )
 				fprintf(stderr, "Error %d Line: %d\n", bt->err, line), exit(0);
 			  len = 0;
 			}
@@ -2371,17 +2272,17 @@ FILE *in;
 		fprintf(stderr, "started reading\n");
 
 	  	do {
-			if( pool = bt_pinpool (bt, page_no) )
-				page = bt_page (bt, pool, page_no);
+			if( bt->pool = bt_pinpool (bt, page_no) )
+				page = bt_page (bt, bt->pool, page_no);
 			else
 				break;
-			set = bt_pinlatch (bt, page_no);
-			bt_lockpage (BtLockRead, set);
+			bt->set = bt_pinlatch (bt, page_no);
+			bt_lockpage (BtLockRead, bt->set);
 			cnt += page->act;
 			next = bt_getid (page->right);
-			bt_unlockpage (BtLockRead, set);
-			bt_unpinlatch (set);
-			bt_unpinpool (pool);
+			bt_unlockpage (BtLockRead, bt->set);
+			bt_unpinlatch (bt->set);
+			bt_unpinpool (bt->pool);
 	  	} while( page_no = next );
 
 	  	cnt--;	// remove stopper key