From 0c15ab64c54d597779331be5002e2e0ba839a749 Mon Sep 17 00:00:00 2001 From: unknown Date: Thu, 23 Jan 2014 12:45:30 -0800 Subject: [PATCH] foster btree final code release --- fosterbtreea.c | 2075 ++++++++++++++++++++++++++++++++++++++ fosterbtreeb.c | 2192 ++++++++++++++++++++++++++++++++++++++++ fosterbtreec.c | 2166 ++++++++++++++++++++++++++++++++++++++++ fosterbtreee.c | 853 ++++++++-------- fosterbtreee1.c | 2529 ++++++++++++++++++++++++++++++++++++++++++++++ fosterbtreee2.c | 2554 +++++++++++++++++++++++++++++++++++++++++++++++ fosterbtreef.c | 902 +++++++++-------- fosterbtreef1.c | 2447 +++++++++++++++++++++++++++++++++++++++++++++ fosterbtreef2.c | 2509 ++++++++++++++++++++++++++++++++++++++++++++++ 9 files changed, 17402 insertions(+), 825 deletions(-) create mode 100644 fosterbtreea.c create mode 100644 fosterbtreeb.c create mode 100644 fosterbtreec.c create mode 100644 fosterbtreee1.c create mode 100644 fosterbtreee2.c create mode 100644 fosterbtreef1.c create mode 100644 fosterbtreef2.c diff --git a/fosterbtreea.c b/fosterbtreea.c new file mode 100644 index 0000000..4a96ff3 --- /dev/null +++ b/fosterbtreea.c @@ -0,0 +1,2075 @@ +// foster btree version a +// 16 DEC 2013 + +// author: karl malbrain, malbrain@cal.berkeley.edu + +/* +This work, including the source code, documentation +and related data, is placed into the public domain. + +The orginal author is Karl Malbrain. + +THIS SOFTWARE IS PROVIDED AS-IS WITHOUT WARRANTY +OF ANY KIND, NOT EVEN THE IMPLIED WARRANTY OF +MERCHANTABILITY. THE AUTHOR OF THIS SOFTWARE, +ASSUMES _NO_ RESPONSIBILITY FOR ANY CONSEQUENCE +RESULTING FROM THE USE, MODIFICATION, OR +REDISTRIBUTION OF THIS SOFTWARE. +*/ + +// Please see the project home page for documentation +// code.google.com/p/high-concurrency-btree + +#define _FILE_OFFSET_BITS 64 +#define _LARGEFILE64_SOURCE + +#ifdef linux +#define _GNU_SOURCE +#endif + +#ifdef unix +#include +#include +#include +#include +#include +#include +#include +#include +#else +#define WIN32_LEAN_AND_MEAN +#include +#include +#include +#include +#include +#include +#endif + +#include +#include + +typedef unsigned long long uid; + +#ifndef unix +typedef unsigned long long off64_t; +typedef unsigned short ushort; +typedef unsigned int uint; +#endif + +#define BT_ro 0x6f72 // ro +#define BT_rw 0x7772 // rw + +#define BT_maxbits 24 // maximum page size in bits +#define BT_minbits 9 // minimum page size in bits +#define BT_minpage (1 << BT_minbits) // minimum page size +#define BT_maxpage (1 << BT_maxbits) // maximum page size + +/* +There are five lock types for each node in three independent sets: +1. (set 1) AccessIntent: Sharable. Going to Read the node. Incompatible with NodeDelete. +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. +*/ + +typedef enum{ + BtLockAccess, + BtLockDelete, + BtLockRead, + BtLockWrite, + BtLockParent +}BtLock; + +// Define the length of the page and key pointers + +#define BtId 6 + +// Page key slot definition. + +// If BT_maxbits is 15 or less, you can save 4 bytes +// for each key stored by making the first two uints +// into ushorts. You can also save 4 bytes by removing +// the tod field from the key. + +// Keys are marked dead, but remain on the page until +// it cleanup is called. The fence key (highest key) for +// the page is always present, even after cleanup. + +typedef struct { + uint off:BT_maxbits; // page offset for key start + uint dead:1; // set for deleted key + uint tod; // time-stamp for key + unsigned char id[BtId]; // id associated with key +} BtSlot; + +// The key structure occupies space at the upper end of +// each page. It's a length byte followed by the value +// bytes. + +typedef struct { + unsigned char len; + unsigned char key[1]; +} *BtKey; + +// The first part of an index page. +// It is immediately followed +// by the BtSlot array of keys. + +typedef struct Page { + uint cnt; // count of keys in page + uint act; // count of active keys + uint min; // next key offset + uint foster; // count of foster children + unsigned char bits:6; // page size in bits + unsigned char dirty:1; // page needs to be cleaned + unsigned char kill:1; // page is being deleted + unsigned char lvl; // level of page + unsigned char right[BtId]; // page number to right +} *BtPage; + +// mode & definition for latch table implementation + +enum { + Write = 1, + Share = 2 +} LockMode; + +// latch table lock structure + +// mode is set for write access +// share is count of read accessors +// grant write lock when share == 0 + +typedef struct { + int mode:1; + int share:31; +} BtLatch; + +typedef struct { + BtLatch readwr[1]; // read/write page lock + BtLatch access[1]; // Access Intent/Page delete + BtLatch parent[1]; // adoption of foster children +} BtLatchSet; + +// The memory mapping hash 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 pin; // mapped page pin counter + uint slot; // slot index in this array + void *hashprev; // previous cache block for the same hash idx + void *hashnext; // next cache block for the same hash idx +#ifndef unix + HANDLE hmap; +#endif +// array of page latch sets, one for each page in map segment + BtLatchSet pagelatch[0]; +} BtHash; + +// The object structure for Btree access + +typedef struct { + uint page_size; // page size + uint page_bits; // page size in bits + uint seg_bits; // seg size in pages in bits + uint mode; // read-write mode +#ifdef unix + int idx; +#else + HANDLE idx; +#endif + uint nodecnt; // highest page cache node in use + uint nodemax; // highest page cache node allocated + uint hashmask; // number of pages in mmap segment + uint hashsize; // size of Hash Table + uint evicted; // last evicted hash slot + ushort *cache; // hash index for memory pool + BtLatch *latch; // latches for hash table slots + char *nodes; // memory pool page hash nodes +} 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 ( page 0 ) + 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 + uid page_no; // current page number + uid cursor_page; // current cursor page number + unsigned char *mem; // frame, cursor, page memory buffer + int err; // last error +} BtDb; + +typedef enum { + BTERR_ok = 0, + BTERR_again, + BTERR_struct, + BTERR_ovflw, + BTERR_lock, + BTERR_map, + BTERR_wrt, + BTERR_hash +} BTERR; + +// 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); +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); + +// manager functions +extern BtMgr *bt_mgr (char *name, uint mode, uint bits, uint cacheblk, uint segsize, uint hashsize); +void bt_mgrclose (BtMgr *mgr); + +// Helper functions to return cursor slot values + +extern BtKey bt_key (BtDb *bt, uint slot); +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 + +// Number of levels to create in a new BTree + +#define MIN_lvl 2 + +// The page is allocated from low and hi ends. +// The key offsets and row-id's are allocated +// from the bottom, while the text of the key +// is allocated from the top. When the two +// areas meet, the page is split into two. + +// 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. + +// The b-tree root is always located at page 1. +// The first leaf page of level zero is always +// located on page 2. + +// 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. + +// 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. + +// Groups of pages called segments from the btree are +// cached with memory mapping. 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. + +// Page 0 is dedicated to lock for new page extensions, +// and chains empty pages together for reuse. + +// Empty pages are chained together through the ALLOC page and reused. + +// Access macros to address slot and key values from the page + +#define slotptr(page, slot) (((BtSlot *)(page+1)) + (slot-1)) +#define keyptr(page, slot) ((BtKey)((unsigned char*)(page) + slotptr(page, slot)->off)) + +void bt_putid(unsigned char *dest, uid id) +{ +int i = BtId; + + while( i-- ) + dest[i] = (unsigned char)id, id >>= 8; +} + +uid bt_getid(unsigned char *src) +{ +uid id = 0; +int i; + + for( i = 0; i < BtId; i++ ) + id <<= 8, id |= *src++; + + return id; +} + +void bt_mgrclose (BtMgr *mgr) +{ +BtHash *hash; +uint slot; + + // release mapped pages + + for( slot = 0; slot < mgr->nodemax; slot++ ) { + hash = (BtHash *)(mgr->nodes + slot * (sizeof(BtHash) + (mgr->hashmask + 1) * sizeof(BtLatchSet))); + if( hash->slot ) +#ifdef unix + munmap (hash->map, (mgr->hashmask+1) << mgr->page_bits); +#else + { + FlushViewOfFile(hash->map, 0); + UnmapViewOfFile(hash->map); + CloseHandle(hash->hmap); + } +#endif + } + +#ifdef unix + close (mgr->idx); + free (mgr->nodes); + free (mgr->cache); + free (mgr->latch); +#else + FlushFileBuffers(mgr->idx); + CloseHandle(mgr->idx); + GlobalFree (mgr->nodes); + GlobalFree (mgr->cache); + GlobalFree (mgr->latch); +#endif +} + +// close and release memory + +void bt_close (BtDb *bt) +{ +#ifdef unix + if ( bt->mem ) + free (bt->mem); + free (bt); +#else + if ( bt->mem) + VirtualFree (bt->mem, 0, MEM_RELEASE); + GlobalFree (bt); +#endif +} + +// open/create new btree buffer manager + +// call with file_name, BT_openmode, bits in page size (e.g. 16), +// size of mapped page cache (e.g. 8192) + +BtMgr *bt_mgr (char *name, uint mode, uint bits, uint nodemax, uint segsize, uint hashsize) +{ +uint lvl, attr, cacheblk, last; +BtPage alloc; +int lockmode; +off64_t size; +uint amt[1]; +BtMgr* mgr; +BtKey key; + +#ifndef unix +SYSTEM_INFO sysinfo[1]; +#endif + + // determine sanity of page size and buffer pool + + if( bits > BT_maxbits ) + bits = BT_maxbits; + else if( bits < BT_minbits ) + bits = BT_minbits; + + if( !nodemax ) + return NULL; // must have buffer pool + +#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; + + case BT_ro: + default: + mgr->idx = open ((char*)name, O_RDONLY); + lockmode = 0; + break; + } + if( mgr->idx == -1 ) + return free(mgr), NULL; + + cacheblk = 4096; // minimum mmap segment size for unix + +#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; + + 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; + + // normalize cacheblk to multiple of sysinfo->dwAllocationGranularity + GetSystemInfo(sysinfo); + cacheblk = sysinfo->dwAllocationGranularity; +#endif + +#ifdef unix + alloc = 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; + else + return free(mgr), free(alloc), NULL; + } else if( mode == BT_ro ) + return bt_mgrclose (mgr), NULL; +#else + alloc = 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) ) + return bt_mgrclose (mgr), NULL; + bits = alloc->bits; + } else if( mode == BT_ro ) + return bt_mgrclose (mgr), NULL; +#endif + + mgr->page_size = 1 << bits; + mgr->page_bits = bits; + + mgr->nodemax = nodemax; + mgr->mode = mode; + + if( cacheblk < mgr->page_size ) + cacheblk = mgr->page_size; + + // mask for partial memmaps + + mgr->hashmask = (cacheblk >> bits) - 1; + + // see if requested number of pages per memmap is greater + + if( (1 << segsize) > mgr->hashmask ) + mgr->hashmask = (1 << segsize) - 1; + + mgr->seg_bits = 0; + + while( (1 << mgr->seg_bits) <= mgr->hashmask ) + mgr->seg_bits++; + + mgr->hashsize = hashsize; + +#ifdef unix + mgr->nodes = calloc (nodemax, (sizeof(BtHash) + (mgr->hashmask + 1) * sizeof(BtLatchSet))); + mgr->cache = calloc (hashsize, sizeof(ushort)); + mgr->latch = calloc (hashsize, sizeof(BtLatch)); +#else + mgr->nodes = GlobalAlloc (GMEM_FIXED|GMEM_ZEROINIT, cacheblk * (sizeof(BtHash) + (mgr->hashmask + 1) * sizeof(BtLatchSet))); + mgr->cache = GlobalAlloc (GMEM_FIXED|GMEM_ZEROINIT, hashsize * sizeof(ushort)); + mgr->latch = GlobalAlloc (GMEM_FIXED|GMEM_ZEROINIT, hashsize * sizeof(BtLatch)); +#endif + + if( size || *amt ) + goto mgrxit; + + // initializes an empty b-tree with root page and page of leaves + + memset (alloc, 0, 1 << bits); + bt_putid(slotptr(alloc, 2)->id, MIN_lvl+1); + alloc->bits = mgr->page_bits; + +#ifdef unix + if( write (mgr->idx, alloc, mgr->page_size) < mgr->page_size ) + return bt_mgrclose (mgr), NULL; +#else + if( !WriteFile (mgr->idx, (char *)alloc, 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; + + 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 + key->key[0] = 0xff; + key->key[1] = 0xff; + alloc->min = mgr->page_size - 3; + alloc->lvl = lvl; + alloc->cnt = 1; + alloc->act = 1; +#ifdef unix + if( write (mgr->idx, alloc, mgr->page_size) < mgr->page_size ) + return bt_mgrclose (mgr), NULL; +#else + if( !WriteFile (mgr->idx, (char *)alloc, mgr->page_size, amt, NULL) ) + return bt_mgrclose (mgr), NULL; + + if( *amt < mgr->page_size ) + return bt_mgrclose (mgr), NULL; +#endif + } + + // create empty page area by writing last page of first + // cache area (other pages are zeroed by O/S) + + if( mgr->hashmask ) { + memset(alloc, 0, mgr->page_size); + last = mgr->hashmask; + + while( last < MIN_lvl + 1 ) + last += mgr->hashmask + 1; + +#ifdef unix + pwrite(mgr->idx, alloc, 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) ) + return bt_mgrclose (mgr), NULL; + if( *amt < mgr->page_size ) + return bt_mgrclose (mgr), NULL; +#endif + } + +mgrxit: +#ifdef unix + free (alloc); +#else + VirtualFree (alloc, 0, MEM_RELEASE); +#endif + return mgr; +} + +// open BTree access method +// based on buffer manager + +BtDb *bt_open (BtMgr *mgr) +{ +BtDb *bt = malloc (sizeof(*bt)); + + memset (bt, 0, sizeof(*bt)); + bt->mgr = mgr; +#ifdef unix + bt->mem = malloc (3 *mgr->page_size); +#else + bt->mem = VirtualAlloc(NULL, 3 * mgr->page_size, MEM_COMMIT, PAGE_READWRITE); +#endif + bt->frame = (BtPage)bt->mem; + bt->zero = (BtPage)(bt->mem + 1 * mgr->page_size); + bt->cursor = (BtPage)(bt->mem + 2 * mgr->page_size); + return bt; +} + +// compare two keys, returning > 0, = 0, or < 0 +// as the comparison value + +int keycmp (BtKey key1, unsigned char *key2, uint len2) +{ +uint len1 = key1->len; +int ans; + + if( ans = memcmp (key1->key, key2, len1 > len2 ? len2 : len1) ) + return ans; + + if( len1 > len2 ) + return 1; + if( len1 < len2 ) + return -1; + + return 0; +} + +// Latch Manager + +// wait until write lock mode is clear +// and add 1 to the share count + +void bt_readlock(BtLatch *latch) +{ + do { + // add one to counter, check write bit + +#ifdef unix + if( ~__sync_fetch_and_add((int *)latch, Share) & Write ) + return; +#else + if( ~InterlockedAdd((int *)latch, Share) & Write ) + return; +#endif + // didn't get latch, reset counter by one + +#ifdef unix + __sync_fetch_and_add((int *)latch, -Share); +#else + InterlockedAdd ((int *)latch, -Share); +#endif + + // and yield +#ifdef unix + sched_yield(); +#else + SwitchToThread(); +#endif + } while( 1 ); +} + +// wait for other read and write latches to relinquish + +void bt_writelock(BtLatch *latch) +{ +int prev, ours = 0; + + do { + // see if we can get write access + // with no readers +#ifdef unix + prev = __sync_fetch_and_or((int *)latch, Write); +#else + prev = InterlockedOr((int *)latch, Write); +#endif + + if( ~prev & 1 ) + ours++; // it's ours + + if( !(prev >> 1) && ours ) + return; + + // otherwise yield + +#ifdef unix + sched_yield(); +#else + SwitchToThread(); +#endif + } while( 1 ); +} + +// try to obtain write lock + +// return 1 if obtained, +// 0 if already write locked + +int bt_writetry(BtLatch *latch) +{ +int prev, ours = 0; + + do { + // see if we can get write access + // with no readers +#ifdef unix + prev = __sync_fetch_and_or((int *)latch, Write); +#else + prev = InterlockedOr((int *)latch, Write); +#endif + + if( ~prev & 1 ) + ours++; // it's ours + + if( !ours ) + return 0; + + if( !(prev >> 1) && ours ) + return 1; + + // otherwise yield +#ifdef unix + sched_yield(); +#else + SwitchToThread(); +#endif + } while( 1 ); +} + +// clear write mode + +void bt_releasewrite(BtLatch *latch) +{ +#ifdef unix + __sync_fetch_and_and((int *)latch, ~Write); +#else + InterlockedAnd ((int *)latch, ~Write); +#endif +} + +// decrement reader count + +void bt_releaseread(BtLatch *latch) +{ +#ifdef unix + __sync_fetch_and_add((int *)latch, -Share); +#else + InterlockedAdd((int *)latch, -Share); +#endif +} + +// Buffer Pool mgr + +// find segment in cache +// return NULL if not there +// otherwise return node + +BtHash *bt_findhash(BtDb *bt, uid page_no, uint idx) +{ +BtHash *hash; +uint slot; + + // compute cache block first page and hash idx + + if( slot = bt->mgr->cache[idx] ) + hash = (BtHash *)(bt->mgr->nodes + slot * (sizeof(BtHash) + (bt->mgr->hashmask + 1) * sizeof(BtLatchSet))); + else + return NULL; + + page_no &= ~bt->mgr->hashmask; + + while( hash->basepage != page_no ) + if( hash = hash->hashnext ) + continue; + else + return NULL; + + return hash; +} + +// add segment to hash table + +void bt_linkhash(BtDb *bt, BtHash *hash, uid page_no, int idx) +{ +BtHash *node; +uint slot; + + hash->hashprev = hash->hashnext = NULL; + hash->basepage = page_no & ~bt->mgr->hashmask; + hash->pin = 1; + hash->lru = 1; + + if( slot = bt->mgr->cache[idx] ) { + node = (BtHash *)(bt->mgr->nodes + slot * (sizeof(BtHash) + (bt->mgr->hashmask + 1) * sizeof(BtLatchSet))); + hash->hashnext = node; + node->hashprev = hash; + } + + bt->mgr->cache[idx] = hash->slot; +} + +// find best segment to evict from buffer pool + +BtHash *bt_findlru (BtDb *bt, uint slot) +{ +unsigned long long int target = ~0LL; +BtHash *hash = NULL, *node; + + if( !slot ) + return NULL; + + node = (BtHash *)(bt->mgr->nodes + slot * (sizeof(BtHash) + (bt->mgr->hashmask + 1) * sizeof(BtLatchSet))); + + do { + if( node->pin ) + continue; + if( node->lru > target ) + continue; + target = node->lru; + hash = node; + } while( node = node->hashnext ); + + return hash; +} + +// map new segment to virtual memory + +BTERR bt_mapsegment(BtDb *bt, BtHash *hash, uid page_no) +{ +off64_t off = (page_no & ~bt->mgr->hashmask) << bt->mgr->page_bits; +off64_t limit = off + ((bt->mgr->hashmask+1) << bt->mgr->page_bits); +int flag; + +#ifdef unix + flag = PROT_READ | ( bt->mgr->mode == BT_ro ? 0 : PROT_WRITE ); + hash->map = mmap (0, (bt->mgr->hashmask+1) << bt->mgr->page_bits, flag, MAP_SHARED, bt->mgr->idx, off); + if( hash->map == MAP_FAILED ) + return bt->err = BTERR_map; +#else + flag = ( bt->mgr->mode == BT_ro ? PAGE_READONLY : PAGE_READWRITE ); + hash->hmap = CreateFileMapping(bt->mgr->idx, NULL, flag, (DWORD)(limit >> 32), (DWORD)limit, NULL); + if( !hash->hmap ) + return bt->err = BTERR_map; + + flag = ( bt->mgr->mode == BT_ro ? FILE_MAP_READ : FILE_MAP_WRITE ); + hash->map = MapViewOfFile(hash->hmap, flag, (DWORD)(off >> 32), (DWORD)off, (bt->mgr->hashmask+1) << bt->mgr->page_bits); + if( !hash->map ) + return bt->err = BTERR_map; +#endif + return bt->err = 0; +} + +// find or place requested page in segment-cache +// return hash table entry + +BtHash *bt_hashpage(BtDb *bt, uid page_no) +{ +BtHash *hash, *node, *next; +uint slot, idx, victim; +BtLatchSet *set; + + // lock hash table chain + + idx = (uint)(page_no >> bt->mgr->seg_bits) % bt->mgr->hashsize; + bt_readlock (&bt->mgr->latch[idx]); + + // look up in hash table + + if( hash = bt_findhash(bt, page_no, idx) ) { +#ifdef unix + __sync_fetch_and_add(&hash->pin, 1); +#else + InterlockedIncrement (&hash->pin); +#endif + bt_releaseread (&bt->mgr->latch[idx]); + hash->lru++; + return hash; + } + + // upgrade to write lock + + bt_releaseread (&bt->mgr->latch[idx]); + bt_writelock (&bt->mgr->latch[idx]); + + // try to find page in cache with write lock + + if( hash = bt_findhash(bt, page_no, idx) ) { +#ifdef unix + __sync_fetch_and_add(&hash->pin, 1); +#else + InterlockedIncrement (&hash->pin); +#endif + bt_releasewrite (&bt->mgr->latch[idx]); + hash->lru++; + return hash; + } + + // allocate a new hash node + // and add to hash table + +#ifdef unix + slot = __sync_fetch_and_add(&bt->mgr->nodecnt, 1); +#else + slot = InterlockedIncrement (&bt->mgr->nodecnt) - 1; +#endif + + if( ++slot < bt->mgr->nodemax ) { + hash = (BtHash *)(bt->mgr->nodes + slot * (sizeof(BtHash) + (bt->mgr->hashmask + 1) * sizeof(BtLatchSet))); + hash->slot = slot; + + if( bt_mapsegment(bt, hash, page_no) ) + return NULL; + + bt_linkhash(bt, hash, page_no, idx); + bt_releasewrite (&bt->mgr->latch[idx]); + return hash; + } + + // hash table is full + // find best cache entry to evict + +#ifdef unix + __sync_fetch_and_add(&bt->mgr->nodecnt, -1); +#else + InterlockedDecrement (&bt->mgr->nodecnt); +#endif + + while( 1 ) { +#ifdef unix + victim = __sync_fetch_and_add(&bt->mgr->evicted, 1); +#else + victim = InterlockedIncrement (&bt->mgr->evicted) - 1; +#endif + victim %= bt->mgr->hashsize; + + // try to get write lock + // skip entry if not obtained + + if( !bt_writetry (&bt->mgr->latch[victim]) ) + continue; + + // if cache entry is empty + // or no slots are unpinned + // skip this entry + + if( !(hash = bt_findlru(bt, bt->mgr->cache[victim])) ) { + bt_releasewrite (&bt->mgr->latch[victim]); + continue; + } + + // unlink victim hash node from hash table + + if( node = hash->hashprev ) + node->hashnext = hash->hashnext; + else if( node = hash->hashnext ) + bt->mgr->cache[victim] = node->slot; + else + bt->mgr->cache[victim] = 0; + + if( node = hash->hashnext ) + node->hashprev = hash->hashprev; + + // remove old file mapping +#ifdef unix + munmap (hash->map, (bt->mgr->hashmask+1) << bt->mgr->page_bits); +#else + FlushViewOfFile(hash->map, 0); + UnmapViewOfFile(hash->map); + CloseHandle(hash->hmap); +#endif + hash->map = NULL; + bt_releasewrite (&bt->mgr->latch[victim]); + + // create new file mapping + // and link into hash table + + if( bt_mapsegment(bt, hash, page_no) ) + return NULL; + + bt_linkhash(bt, hash, page_no, idx); + bt_releasewrite (&bt->mgr->latch[idx]); + return hash; + } +} + +// place write, read, or parent lock on requested page_no. +// pin to buffer pool + +BTERR bt_lockpage(BtDb *bt, uid page_no, BtLock mode, BtPage *page) +{ +BtLatchSet *set; +BtHash *hash; +uint subpage; + + // find/create maping in hash table + + if( hash = bt_hashpage(bt, page_no) ) + subpage = (uint)(page_no & bt->mgr->hashmask); // page within mapping + else + return bt->err; + + set = hash->pagelatch + subpage; + + switch( mode ) { + case BtLockRead: + bt_readlock (set->readwr); + break; + case BtLockWrite: + bt_writelock (set->readwr); + break; + case BtLockAccess: + bt_readlock (set->access); + break; + case BtLockDelete: + bt_writelock (set->access); + break; + case BtLockParent: + bt_writelock (set->parent); + break; + default: + return bt->err = BTERR_lock; + } + + if( page ) + *page = (BtPage)(hash->map + (subpage << bt->mgr->page_bits)); + + return bt->err = 0; +} + +// remove write, read, or parent lock on requested page_no. + +BTERR bt_unlockpage(BtDb *bt, uid page_no, BtLock mode) +{ +uint subpage, idx; +BtLatchSet *set; +BtHash *hash; + + // since page is pinned + // it should still be in the buffer pool + + idx = (uint)(page_no >> bt->mgr->seg_bits) % bt->mgr->hashsize; + bt_readlock (&bt->mgr->latch[idx]); + + if( hash = bt_findhash(bt, page_no, idx) ) + subpage = (uint)(page_no & bt->mgr->hashmask); + else + return bt->err = BTERR_hash; + + bt_releaseread (&bt->mgr->latch[idx]); + set = hash->pagelatch + subpage; + + switch( mode ) { + case BtLockRead: + bt_releaseread (set->readwr); + break; + case BtLockWrite: + bt_releasewrite (set->readwr); + break; + case BtLockAccess: + bt_releaseread (set->access); + break; + case BtLockDelete: + bt_releasewrite (set->access); + break; + case BtLockParent: + bt_releasewrite (set->parent); + break; + default: + return bt->err = BTERR_lock; + } + +#ifdef unix + __sync_fetch_and_add(&hash->pin, -1); +#else + InterlockedDecrement (&hash->pin); +#endif + return bt->err = 0; +} + +// deallocate a deleted page that has no tree pointers +// 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 first key + bt_putid(slotptr(bt->temp, 1)->id, bt_getid(slotptr(bt->alloc, 1)->id)); + bt_putid(slotptr(bt->alloc, 1)->id, 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 + + if( bt_unlockpage(bt, page_no, BtLockDelete) ) + return bt->err; + + return 0; +} + +// allocate a new page and write page into it + +uid bt_newpage(BtDb *bt, BtPage page) +{ +uid new_page; +BtPage pmap; +int reuse; + + // lock page zero + + if ( bt_lockpage(bt, ALLOC_page, BtLockWrite, &bt->alloc) ) + return 0; + + // use empty chain first + // else allocate empty page + + if( new_page = bt_getid(slotptr(bt->alloc, 1)->id) ) { + if( bt_lockpage (bt, new_page, BtLockWrite, &bt->temp) ) + return 0; + bt_putid(slotptr(bt->alloc, 1)->id, bt_getid(slotptr(bt->temp, 1)->id)); + if( bt_unlockpage (bt, new_page, BtLockWrite) ) + return 0; + reuse = 1; + } else { + new_page = bt_getid(slotptr(bt->alloc, 2)->id); + bt_putid(slotptr(bt->alloc, 2)->id, new_page+1); + 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 hash block, zero last page in the block + + if ( !reuse && bt->mgr->hashmask > 0 && (new_page & bt->mgr->hashmask) == 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->hashmask) << bt->mgr->page_bits) < bt->mgr->page_size ) + return bt->err = BTERR_wrt, 0; + } +#else + // bring new page into page-cache and copy page. + // this will extend the file into the new pages. + + if( bt_lockpage(bt, new_page, BtLockWrite, &pmap) ) + return 0; + + memcpy(pmap, page, bt->mgr->page_size); + + if( bt_unlockpage (bt, new_page, BtLockWrite) ) + return 0; +#endif + // unlock page zero + + if ( bt_unlockpage(bt, ALLOC_page, BtLockWrite) ) + return 0; + + return new_page; +} + +// find slot in page for given key at a given level + +int bt_findslot (BtDb *bt, unsigned char *key, uint len) +{ +uint diff, higher = bt->page->cnt, low = 1, slot; + + // low is the lowest candidate, higher is already + // tested as .ge. the given key, loop ends when they meet + + while( diff = higher - low ) { + slot = low + ( diff >> 1 ); + if( keycmp (keyptr(bt->page, slot), key, len) < 0 ) + low = slot + 1; + else + higher = slot; + } + + return higher; +} + +// 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) +{ +uid page_no = ROOT_page, prevpage = 0; +uint drill = 0xff, slot; +uint mode, prevmode; + + // start at root of btree and drill down + + do { + // determine lock mode of drill level + mode = (lock == BtLockWrite) && (drill == lvl) ? BtLockWrite : BtLockRead; + + bt->page_no = page_no; + + // obtain access lock using lock chaining with Access mode + + if( page_no > ROOT_page ) + if( bt_lockpage(bt, page_no, BtLockAccess, NULL) ) + return 0; + + if( prevpage ) + if( bt_unlockpage(bt, prevpage, prevmode) ) + return 0; + + // obtain read lock using lock chaining + // and pin page contents + + if( bt_lockpage(bt, page_no, mode, &bt->page) ) + return 0; + + if( page_no > ROOT_page ) + if( bt_unlockpage(bt, page_no, BtLockAccess) ) + return 0; + + // re-read and re-lock root after determining actual level of root + + if( bt->page_no == ROOT_page ) + if( bt->page->lvl != drill) { + drill = bt->page->lvl; + + if( lock == BtLockWrite && drill == lvl ) + if( bt_unlockpage(bt, page_no, mode) ) + return 0; + else + continue; + } + + // if page is being deleted, + // move back to preceeding page + + if( bt->page->kill ) { + page_no = bt_getid (bt->page->right); + continue; + } + + // find key on page at this level + // and descend to requested level + + slot = bt_findslot (bt, key, len); + + // is this slot a foster child? + + if( slot <= bt->page->cnt - bt->page->foster ) + if( drill == lvl ) + return slot; + else + drill--; + + while( slotptr(bt->page, slot)->dead ) + if( slot++ < bt->page->cnt ) + continue; + else + return bt->err = BTERR_struct, 0; + + // continue down / right using overlapping locks + // to protect pages being killed or split. + + prevmode = mode; + prevpage = bt->page_no; + page_no = bt_getid(slotptr(bt->page, slot)->id); + } while( page_no ); + + // return error on end of chain + + bt->err = BTERR_struct; + return 0; // return error +} + +// find and delete key on page by marking delete flag bit +// when page becomes empty, delete it from the btree + +BTERR bt_deletekey (BtDb *bt, unsigned char *key, uint len, uint lvl) +{ +unsigned char leftkey[256], rightkey[256]; +uid page_no, right; +uint slot, tod; +BtKey ptr; + + 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 + + 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--; + } + + // return if page is not empty, or it has no right sibling + + right = bt_getid(bt->page->right); + page_no = bt->page_no; + + if( !right || bt->page->act ) + return bt_unlockpage(bt, page_no, BtLockWrite); + + // obtain Parent lock over write lock + + if( bt_lockpage(bt, page_no, BtLockParent, NULL) ) + return bt->err; + + // cache copy of key to delete + + ptr = keyptr(bt->page, bt->page->cnt); + memcpy(leftkey, ptr, ptr->len + 1); + + // lock and map right page + + if ( bt_lockpage(bt, right, BtLockWrite, &bt->temp) ) + return bt->err; + + // pull contents of next page into current empty page + memcpy (bt->page, bt->temp, bt->mgr->page_size); + + // cache copy of key to update + ptr = keyptr(bt->temp, bt->temp->cnt); + memcpy(rightkey, ptr, ptr->len + 1); + + // Mark right page as deleted and point it to left page + // until we can post updates at higher level. + + bt_putid(bt->temp->right, page_no); + bt->temp->kill = 1; + bt->temp->cnt = 0; + + if( bt_unlockpage(bt, right, BtLockWrite) ) + return bt->err; + if( bt_unlockpage(bt, page_no, BtLockWrite) ) + return bt->err; + + // delete old lower key to consolidated node + + if( bt_deletekey (bt, leftkey + 1, *leftkey, lvl + 1) ) + return bt->err; + + // redirect higher key directly to consolidated node + + if( slot = bt_loadpage (bt, rightkey+1, *rightkey, lvl+1, BtLockWrite) ) + ptr = keyptr(bt->page, slot); + else + return bt->err; + + // since key already exists, update id + + if( keycmp (ptr, rightkey+1, *rightkey) ) + return bt->err = BTERR_struct; + + slotptr(bt->page, slot)->dead = 0; + bt_putid(slotptr(bt->page,slot)->id, page_no); + bt_unlockpage(bt, bt->page_no, BtLockWrite); + + // obtain write lock and + // add right block to free chain + + if( bt_freepage (bt, right) ) + return bt->err; + + // remove ParentModify lock + + if( bt_unlockpage(bt, page_no, BtLockParent) ) + return bt->err; + + return 0; +} + +// find key in leaf level and return row-id + +uid bt_findkey (BtDb *bt, unsigned char *key, uint len) +{ +uint slot; +BtKey ptr; +uid id; + + if( slot = bt_loadpage (bt, key, len, 0, BtLockRead) ) + ptr = keyptr(bt->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; + + return id; +} + +// check page for space available, +// clean if necessary and return +// 0 - page needs splitting +// 1 - go ahead + +uint bt_cleanpage(BtDb *bt, uint amt) +{ +uint nxt = bt->mgr->page_size; +BtPage page = bt->page; +uint cnt = 0, idx = 0; +uint max = page->cnt; +BtKey key; + + if( page->min >= (max+1) * sizeof(BtSlot) + sizeof(*page) + amt + 1 ) + return 1; + + // skip cleanup if nothing to reclaim + + if( !page->dirty ) + return 0; + + memcpy (bt->frame, page, bt->mgr->page_size); + + // skip page info and set rest of page to zero + + memset (page+1, 0, bt->mgr->page_size - sizeof(*page)); + page->dirty = 0; + page->act = 0; + + // try cleaning up page first + + while( cnt++ < max ) { + // always leave fence key in list + if( cnt < max && slotptr(bt->frame,cnt)->dead ) + continue; + + // copy key + 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 >= (idx+1) * sizeof(BtSlot) + sizeof(*page) + amt + 1 ) + return 1; + + return 0; +} + +// add key to page +// return with page unlocked + +BTERR bt_addkeytopage (BtDb *bt, uint slot, unsigned char *key, uint len, uid id, uint tod) +{ +BtPage page = bt->page; +uint idx; + + // calculate next available slot and copy key into page + + page->min -= len + 1; + ((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; + + return bt_unlockpage(bt, bt->page_no, BtLockWrite); +} + +// split the root and raise the height of the btree + +BTERR bt_splitroot(BtDb *bt, uid right) +{ +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. Strip foster child key. + // Save left fence key. + + bt->page->act--; + bt->page->cnt--; + bt->page->foster--; + key = keyptr(bt->page, bt->page->cnt); + memcpy (fencekey, key, key->len + 1); + + if( !(new_page = bt_newpage(bt, bt->page)) ) + return bt->err; + + // preserve the page info at the bottom + // and set rest to zero + + memset (root+1, 0, bt->mgr->page_size - sizeof(*root)); + + // insert left fence key on newroot page + + nxt -= *fencekey + 1; + memcpy ((unsigned char *)root + nxt, fencekey, *fencekey + 1); + bt_putid(slotptr(root, 1)->id, new_page); + slotptr(root, 1)->off = nxt; + + // insert stopper 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); + slotptr(root, 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++; + + // release root (bt->page) + + return bt_unlockpage(bt, bt->page_no, BtLockWrite); +} + +// split already locked full node +// 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; +BtPage page = bt->page; +uint tod = time(NULL); +uint lvl = page->lvl; +uid new_page, right; +BtKey key; + + // initialize frame buffer + + memset (bt->frame, 0, bt->mgr->page_size); + max = page->cnt - page->foster; + tod = (uint)time(NULL); + cnt = max / 2; + idx = 0; + + // split higher half of keys to bt->frame + // leaving foster children in the left node. + + while( cnt++ < max ) { + key = keyptr(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); + slotptr(bt->frame, idx)->tod = slotptr(page, cnt)->tod; + slotptr(bt->frame, idx)->off = nxt; + bt->frame->act++; + } + + // transfer right link node + + if( page_no > ROOT_page ) { + right = bt_getid (page->right); + bt_putid(bt->frame->right, right); + } + + 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 frame to it. + + if( !(new_page = bt_newpage(bt, bt->frame)) ) + return bt->err; + + // update lower keys and foster children 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->act = 0; + cnt = 0; + idx = 0; + + // assemble page of smaller keys + // to remain in the old page + + 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++; + } + + // assemble old foster child keys + // add new foster child fence + + cnt = bt->frame->cnt - bt->frame->foster - 1; + + 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); + slotptr(page, idx)->tod = slotptr(bt->frame, cnt)->tod; + slotptr(page, idx)->off = nxt; + page->act++; + } + + // link new right page + + bt_putid (page->right, new_page); + + // put new page as smallest foster child key + + page->min = nxt; + page->cnt = idx; + cnt = page->cnt - page->foster++; + bt_putid (slotptr(page,cnt)->id, new_page); + + // if current page is the root page, split it + + if( page_no == ROOT_page ) + return bt_splitroot (bt, new_page); + + // release wr lock on page + + if( bt_unlockpage (bt, page_no, BtLockWrite) ) + return bt->err; + + // obtain ParentModification lock for current page + // to fix highest foster child on page + + if( bt_lockpage (bt, page_no, BtLockParent, NULL) ) + return bt->err; + + if( bt_lockpage (bt, page_no, BtLockRead, &page) ) + return bt->err; + + // get our old fence key + + key = keyptr(page, page->cnt); + memcpy (fencekey, key, key->len+1); + + if( bt_unlockpage (bt, page_no, BtLockRead) ) + return bt->err; + + do { + slot = bt_loadpage (bt, fencekey + 1, *fencekey, lvl + 1, BtLockWrite); + + if( !slot ) + return bt->err; + + // check if parent page has enough space for largest possible key + + if( bt_cleanpage (bt, 256) ) + break; + + if( bt_splitpage (bt) ) + return bt->err; + } while( 1 ); + + // wait until readers from parent get their locks + + if( bt_lockpage (bt, page_no, BtLockDelete, NULL) ) + return bt->err; + + if( bt_lockpage (bt, page_no, BtLockWrite, &page) ) + return bt->err; + + // switch parent fence key to foster child + + if( slotptr(page, page->cnt)->dead ) + slotptr(bt->page, slot)->dead = 1; + else + bt_putid (slotptr(bt->page, slot)->id, bt_getid(slotptr(page, page->cnt)->id)); + + // remove foster child from our page + // add our new fence key to parent + + page->cnt--; + page->act--; + page->foster--; + page->dirty = 1; + key = keyptr(page, page->cnt); + + if( bt_addkeytopage (bt, slot, key->key, key->len, page_no, tod) ) + return bt->err; + + if( bt_unlockpage (bt, page_no, BtLockDelete) ) + return bt->err; + + if( bt_unlockpage (bt, page_no, BtLockParent) ) + return bt->err; + + return bt_unlockpage (bt, page_no, BtLockWrite); +} + +// Insert new key into the btree at leaf level. + +BTERR bt_insertkey (BtDb *bt, unsigned char *key, uint len, uid id, uint tod) +{ +uint slot, idx; +BtPage page; +BtKey ptr; + + while( 1 ) { + if( slot = bt_loadpage (bt, key, len, 0, 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; + + 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); + } + + // check if page has enough space + + if( bt_cleanpage (bt, len) ) + break; + + if( bt_splitpage (bt) ) + return bt->err; + } + + return bt_addkeytopage (bt, slot, key, len, id, tod); +} + +// cache page of keys into cursor and return starting slot for given key + +uint bt_startkey (BtDb *bt, unsigned char *key, uint len) +{ +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; + + return slot; +} + +// return next slot for cursor page +// or slide cursor right into next page + +uint bt_nextkey (BtDb *bt, uint slot) +{ +BtPage page; +uid right; + + do { + right = bt_getid(bt->cursor->right); + while( slot++ < bt->cursor->cnt - bt->cursor->foster ) + if( slotptr(bt->cursor,slot)->dead ) + continue; + else if( right || (slot < bt->cursor->cnt - bt->cursor->foster) ) + return slot; + else + break; + + if( !right ) + break; + + bt->cursor_page = right; + + if( bt_lockpage(bt, right, BtLockRead, &page) ) + return 0; + + memcpy (bt->cursor, page, bt->mgr->page_size); + + if ( bt_unlockpage(bt, right, BtLockRead) ) + return 0; + + slot = 0; + } while( 1 ); + + return bt->err = 0; +} + +BtKey bt_key(BtDb *bt, uint slot) +{ + return keyptr(bt->cursor, slot); +} + +uid bt_uid(BtDb *bt, uint slot) +{ + return bt_getid(slotptr(bt->cursor,slot)->id); +} + +uint bt_tod(BtDb *bt, uint slot) +{ + return slotptr(bt->cursor,slot)->tod; +} + + +#ifdef STANDALONE + +typedef struct { + char type, num; + char *infile; + BtMgr *mgr; +} ThreadArg; + +// standalone program to index file of keys +// then list them onto std-out + +#ifdef unix +void *index_file (void *arg) +#else +uint __stdcall index_file (void *arg) +#endif +{ +int line = 0, found = 0; +unsigned char key[256]; +ThreadArg *args = arg; +int ch, len = 0, slot; +time_t tod[1]; +BtKey ptr; +BtDb *bt; +FILE *in; + + bt = bt_open (args->mgr); + time (tod); + + switch(args->type | 0x20) + { + case 'w': + fprintf(stderr, "started indexing for %s\n", args->infile); + if( in = fopen (args->infile, "rb") ) + while( ch = getc(in), ch != EOF ) + if( ch == '\n' ) + { + line++; + + if( args->num ) + sprintf((char *)key+len, "%.9d", line), len += 9; + + if( bt_insertkey (bt, key, len, line, *tod) ) + fprintf(stderr, "Error %d Line: %d\n", bt->err, line), exit(0); + len = 0; + } + else if( len < 255 ) + key[len++] = ch; + fprintf(stderr, "finished %s for %d keys\n", args->infile, line); + break; + + case 'd': + fprintf(stderr, "started deleting keys for %s\n", args->infile); + if( in = fopen (args->infile, "rb") ) + while( ch = getc(in), ch != EOF ) + if( ch == '\n' ) + { + line++; + if( bt_deletekey (bt, key, len, 0) ) + fprintf(stderr, "Error %d Line: %d\n", bt->err, line), exit(0); + len = 0; + } + else if( len < 255 ) + key[len++] = ch; + fprintf(stderr, "finished %s for keys, %d \n", args->infile, line); + break; + + case 'f': + fprintf(stderr, "started finding keys for %s\n", args->infile); + if( in = fopen (args->infile, "rb") ) + while( ch = getc(in), ch != EOF ) + if( ch == '\n' ) + { + line++; + if( bt_findkey (bt, key, len) ) + found++; + else if( bt->err ) + fprintf(stderr, "Error %d Syserr %d Line: %d\n", bt->err, errno, line), exit(0); + len = 0; + } + else if( len < 255 ) + key[len++] = ch; + fprintf(stderr, "finished %s for %d keys, found %d\n", args->infile, line, found); + 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); + } + } + + bt_close (bt); +#ifdef unix + return NULL; +#else + return 0; +#endif +} + +typedef struct timeval timer; + +int main (int argc, char **argv) +{ +int idx, cnt, len, slot, err; +int segsize, bits = 16; +#ifdef unix +pthread_t *threads; +timer start, stop; +#else +time_t start[1], stop[1]; +HANDLE *threads; +#endif +double real_time; +ThreadArg *args; +uint map = 0; +int num = 0; +char key[1]; +BtMgr *mgr; +BtKey ptr; +BtDb *bt; + + if( argc < 3 ) { + fprintf (stderr, "Usage: %s idx_file Read/Write/Scan/Delete/Find [page_bits mapped_segments seg_bits line_numbers src_file1 src_file2 ... ]\n", argv[0]); + 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 = 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 + + if( argc > 3 ) + bits = atoi(argv[3]); + + if( argc > 4 ) + map = atoi(argv[4]); + + if( map > 65536 ) + fprintf (stderr, "Warning: mapped_pool > 65536 segments\n"); + + if( argc > 5 ) + segsize = atoi(argv[5]); + else + segsize = 4; // 16 pages per mmap segment + + if( argc > 6 ) + num = atoi(argv[6]); + + cnt = argc - 7; +#ifdef unix + threads = malloc (cnt * sizeof(pthread_t)); +#else + threads = GlobalAlloc (GMEM_FIXED|GMEM_ZEROINIT, cnt * sizeof(HANDLE)); +#endif + args = malloc (cnt * sizeof(ThreadArg)); + + mgr = bt_mgr ((argv[1]), BT_rw, bits, map, segsize, map / 8); + + if( !mgr ) { + fprintf(stderr, "Index Open Error %s\n", argv[1]); + exit (1); + } + + // fire off threads + + for( idx = 0; idx < cnt; idx++ ) { + args[idx].infile = argv[idx + 7]; + args[idx].type = argv[2][0]; + args[idx].mgr = mgr; + args[idx].num = num; +#ifdef unix + if( err = pthread_create (threads + idx, NULL, index_file, args + idx) ) + fprintf(stderr, "Error creating thread %d\n", err); +#else + threads[idx] = (HANDLE)_beginthreadex(NULL, 65536, index_file, args + idx, 0, NULL); +#endif + } + + // wait for termination + +#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); + + cnt = 0; + len = key[0] = 0; + bt = bt_open (mgr); + + 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) ) + cnt++; + + fprintf(stderr, " Total keys read %d\n", cnt); + + bt_close (bt); + bt_mgrclose (mgr); +} + +#endif //STANDALONE diff --git a/fosterbtreeb.c b/fosterbtreeb.c new file mode 100644 index 0000000..2211cb0 --- /dev/null +++ b/fosterbtreeb.c @@ -0,0 +1,2192 @@ +// foster btree version b +// 22 DEC 2013 + +// author: karl malbrain, malbrain@cal.berkeley.edu + +/* +This work, including the source code, documentation +and related data, is placed into the public domain. + +The orginal author is Karl Malbrain. + +THIS SOFTWARE IS PROVIDED AS-IS WITHOUT WARRANTY +OF ANY KIND, NOT EVEN THE IMPLIED WARRANTY OF +MERCHANTABILITY. THE AUTHOR OF THIS SOFTWARE, +ASSUMES _NO_ RESPONSIBILITY FOR ANY CONSEQUENCE +RESULTING FROM THE USE, MODIFICATION, OR +REDISTRIBUTION OF THIS SOFTWARE. +*/ + +// Please see the project home page for documentation +// code.google.com/p/high-concurrency-btree + +#define _FILE_OFFSET_BITS 64 +#define _LARGEFILE64_SOURCE + +#ifdef linux +#define _GNU_SOURCE +#endif + +#ifdef unix +#include +#include +#include +#include +#include +#include +#include +#include +#else +#define WIN32_LEAN_AND_MEAN +#include +#include +#include +#include +#include +#include +#include +#endif + +#include +#include + +typedef unsigned long long uid; + +#ifndef unix +typedef unsigned long long off64_t; +typedef unsigned short ushort; +typedef unsigned int uint; +#endif + +#define BT_ro 0x6f72 // ro +#define BT_rw 0x7772 // rw + +#define BT_maxbits 24 // maximum page size in bits +#define BT_minbits 9 // minimum page size in bits +#define BT_minpage (1 << BT_minbits) // minimum page size +#define BT_maxpage (1 << BT_maxbits) // maximum page size + +/* +There are five lock types for each node in three independent sets: +1. (set 1) AccessIntent: Sharable. Going to Read the node. Incompatible with NodeDelete. +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. +*/ + +typedef enum{ + BtLockAccess, + BtLockDelete, + BtLockRead, + BtLockWrite, + BtLockParent +}BtLock; + +// Define the length of the page and key pointers + +#define BtId 6 + +// Page key slot definition. + +// If BT_maxbits is 15 or less, you can save 4 bytes +// for each key stored by making the first two uints +// into ushorts. You can also save 4 bytes by removing +// the tod field from the key. + +// Keys are marked dead, but remain on the page until +// it cleanup is called. The fence key (highest key) for +// the page is always present, even after cleanup. + +typedef struct { + uint off:BT_maxbits; // page offset for key start + uint dead:1; // set for deleted key + uint tod; // time-stamp for key + unsigned char id[BtId]; // id associated with key +} BtSlot; + +// The key structure occupies space at the upper end of +// each page. It's a length byte followed by the value +// bytes. + +typedef struct { + unsigned char len; + unsigned char key[1]; +} *BtKey; + +// The first part of an index page. +// It is immediately followed +// by the BtSlot array of keys. + +typedef struct Page { + uint cnt; // count of keys in page + uint act; // count of active keys + uint min; // next key offset + uint foster; // count of foster children + unsigned char bits; // page size in bits + unsigned char lvl:6; // level of page + unsigned char kill:1; // page is being deleted + unsigned char dirty:1; // page needs to be cleaned + unsigned char right[BtId]; // page number to right +} *BtPage; + +// mode & definition for latch table implementation + +enum { + Write = 1, + Pending = 2, + Share = 4 +} LockMode; + +// latch table lock structure + +// exclusive is set for write access +// share is count of read accessors +// grant write lock when share == 0 + +typedef struct { + volatile uint exclusive:1; + volatile uint request:1; + volatile uint share:30; +} BtLatch; + +typedef struct { + BtLatch readwr[1]; // read/write page lock + BtLatch access[1]; // Access Intent/Page delete + BtLatch parent[1]; // adoption of foster children +} BtLatchSet; + +// 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 pin; // mapped page pin counter + uint slot; // slot index in this array + void *hashprev; // previous pool entry for the same hash idx + void *hashnext; // next pool entry for the same hash idx +#ifndef unix + HANDLE hmap; +#endif +// array of page latch sets, one for each page in map segment + BtLatchSet pagelatch[0]; +} BtPool; + +// The object structure for Btree access + +typedef struct { + uint page_size; // page size + uint page_bits; // page size in bits + 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 +#else + HANDLE idx; +#endif + uint poolcnt; // highest page pool node in use + 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 + volatile uint evicted; // last evicted hash table slot + ushort *hash; // hash table of pool entries + BtLatch *latch; // latches for hash table slots + char *nodes; // memory pool page segments +} 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 ( page 0 ) + 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 + uid page_no; // current page number + uid cursor_page; // current cursor page number + unsigned char *mem; // frame, cursor, page memory buffer + int err; // last error +} BtDb; + +typedef enum { + BTERR_ok = 0, + BTERR_struct, + BTERR_ovflw, + BTERR_lock, + BTERR_map, + BTERR_wrt, + BTERR_hash +} BTERR; + +// 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); +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); + +// manager functions +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 + +extern BtKey bt_key (BtDb *bt, uint slot); +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 + +// Number of levels to create in a new BTree + +#define MIN_lvl 2 + +// The page is allocated from low and hi ends. +// The key offsets and row-id's are allocated +// from the bottom, while the text of the key +// is allocated from the top. When the two +// areas meet, the page is split into two. + +// 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. + +// The b-tree root is always located at page 1. +// The first leaf page of level zero is always +// located on page 2. + +// 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. + +// 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. + +// Groups of pages called segments from the btree are +// cached with memory mapping. 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. + +// Page 0 is dedicated to lock for new page extensions, +// and chains empty pages together for reuse. + +// Empty pages are chained together through the ALLOC page and reused. + +// Access macros to address slot and key values from the page + +#define slotptr(page, slot) (((BtSlot *)(page+1)) + (slot-1)) +#define keyptr(page, slot) ((BtKey)((unsigned char*)(page) + slotptr(page, slot)->off)) + +void bt_putid(unsigned char *dest, uid id) +{ +int i = BtId; + + while( i-- ) + dest[i] = (unsigned char)id, id >>= 8; +} + +uid bt_getid(unsigned char *src) +{ +uid id = 0; +int i; + + for( i = 0; i < BtId; i++ ) + id <<= 8, id |= *src++; + + return id; +} + +void bt_mgrclose (BtMgr *mgr) +{ +BtPool *pool; +uint slot; + + // release mapped pages + // note that slot zero is never used + + for( slot = 1; slot < mgr->poolmax; slot++ ) { + pool = (BtPool *)(mgr->nodes + slot * (sizeof(BtPool) + (mgr->poolmask + 1) * sizeof(BtLatchSet))); + if( pool->slot ) +#ifdef unix + munmap (pool->map, (mgr->poolmask+1) << mgr->page_bits); +#else + { + FlushViewOfFile(pool->map, 0); + UnmapViewOfFile(pool->map); + CloseHandle(pool->hmap); + } +#endif + } + +#ifdef unix + close (mgr->idx); + free (mgr->nodes); + free (mgr->hash); + free (mgr->latch); + free (mgr->pooladvise); + free (mgr); +#else + FlushFileBuffers(mgr->idx); + CloseHandle(mgr->idx); + GlobalFree (mgr->nodes); + GlobalFree (mgr->hash); + GlobalFree (mgr->latch); + GlobalFree (mgr); +#endif +} + +// close and release memory + +void bt_close (BtDb *bt) +{ +#ifdef unix + if ( bt->mem ) + free (bt->mem); +#else + if ( bt->mem) + VirtualFree (bt->mem, 0, MEM_RELEASE); +#endif + free (bt); +} + +// open/create new btree buffer manager + +// call with file_name, BT_openmode, bits in page size (e.g. 16), +// size of mapped page pool (e.g. 8192) + +BtMgr *bt_mgr (char *name, uint mode, uint bits, uint poolmax, uint segsize, uint hashsize) +{ +uint lvl, attr, cacheblk, last; +BtPage alloc; +int lockmode; +off64_t size; +uint amt[1]; +BtMgr* mgr; +BtKey key; + +#ifndef unix +SYSTEM_INFO sysinfo[1]; +#endif + + // determine sanity of page size and buffer pool + + if( bits > BT_maxbits ) + bits = BT_maxbits; + else if( bits < BT_minbits ) + bits = BT_minbits; + + if( !poolmax ) + return NULL; // must have buffer pool + +#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; + + case BT_ro: + default: + mgr->idx = open ((char*)name, O_RDONLY); + lockmode = 0; + break; + } + if( mgr->idx == -1 ) + return free(mgr), NULL; + + cacheblk = 4096; // minimum mmap segment size for unix + +#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; + + 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; + + // normalize cacheblk to multiple of sysinfo->dwAllocationGranularity + GetSystemInfo(sysinfo); + cacheblk = sysinfo->dwAllocationGranularity; +#endif + +#ifdef unix + alloc = 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; + else + return free(mgr), free(alloc), NULL; + } else if( mode == BT_ro ) + return bt_mgrclose (mgr), NULL; +#else + alloc = 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) ) + return bt_mgrclose (mgr), NULL; + bits = alloc->bits; + } else if( mode == BT_ro ) + return bt_mgrclose (mgr), NULL; +#endif + + mgr->page_size = 1 << bits; + mgr->page_bits = bits; + + mgr->poolmax = poolmax; + mgr->mode = mode; + + if( cacheblk < mgr->page_size ) + cacheblk = mgr->page_size; + + // mask for partial memmaps + + mgr->poolmask = (cacheblk >> bits) - 1; + + // see if requested size of pages per memmap is greater + + if( (1 << segsize) > mgr->poolmask ) + mgr->poolmask = (1 << segsize) - 1; + + mgr->seg_bits = 0; + + while( (1 << mgr->seg_bits) <= mgr->poolmask ) + mgr->seg_bits++; + + mgr->hashsize = hashsize; + +#ifdef unix + mgr->nodes = calloc (poolmax, (sizeof(BtPool) + (mgr->poolmask + 1) * sizeof(BtLatchSet))); + mgr->hash = calloc (hashsize, sizeof(ushort)); + mgr->latch = calloc (hashsize, sizeof(BtLatch)); + mgr->pooladvise = calloc (poolmax, (mgr->poolmask + 1) / 8); +#else + mgr->nodes = GlobalAlloc (GMEM_FIXED|GMEM_ZEROINIT, poolmax * (sizeof(BtPool) + (mgr->poolmask + 1) * sizeof(BtLatchSet))); + mgr->hash = GlobalAlloc (GMEM_FIXED|GMEM_ZEROINIT, hashsize * sizeof(ushort)); + mgr->latch = GlobalAlloc (GMEM_FIXED|GMEM_ZEROINIT, hashsize * sizeof(BtLatch)); +#endif + + if( size || *amt ) + goto mgrxit; + + // initializes an empty b-tree with root page and page of leaves + + memset (alloc, 0, 1 << bits); + bt_putid(alloc->right, MIN_lvl+1); + alloc->bits = mgr->page_bits; + +#ifdef unix + if( write (mgr->idx, alloc, mgr->page_size) < mgr->page_size ) + return bt_mgrclose (mgr), NULL; +#else + if( !WriteFile (mgr->idx, (char *)alloc, 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; + + 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 + key->key[0] = 0xff; + key->key[1] = 0xff; + alloc->min = mgr->page_size - 3; + alloc->lvl = lvl; + alloc->cnt = 1; + alloc->act = 1; +#ifdef unix + if( write (mgr->idx, alloc, mgr->page_size) < mgr->page_size ) + return bt_mgrclose (mgr), NULL; +#else + if( !WriteFile (mgr->idx, (char *)alloc, mgr->page_size, amt, NULL) ) + return bt_mgrclose (mgr), NULL; + + if( *amt < mgr->page_size ) + return bt_mgrclose (mgr), NULL; +#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; + + while( last < MIN_lvl + 1 ) + last += mgr->poolmask + 1; + +#ifdef unix + pwrite(mgr->idx, alloc, 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) ) + return bt_mgrclose (mgr), NULL; + if( *amt < mgr->page_size ) + return bt_mgrclose (mgr), NULL; +#endif + } + +mgrxit: +#ifdef unix + free (alloc); +#else + VirtualFree (alloc, 0, MEM_RELEASE); +#endif + return mgr; +} + +// open BTree access method +// based on buffer manager + +BtDb *bt_open (BtMgr *mgr) +{ +BtDb *bt = malloc (sizeof(*bt)); + + memset (bt, 0, sizeof(*bt)); + bt->mgr = mgr; +#ifdef unix + bt->mem = malloc (3 *mgr->page_size); +#else + bt->mem = VirtualAlloc(NULL, 3 * mgr->page_size, MEM_COMMIT, PAGE_READWRITE); +#endif + bt->frame = (BtPage)bt->mem; + bt->zero = (BtPage)(bt->mem + 1 * mgr->page_size); + bt->cursor = (BtPage)(bt->mem + 2 * mgr->page_size); + return bt; +} + +// compare two keys, returning > 0, = 0, or < 0 +// as the comparison value + +int keycmp (BtKey key1, unsigned char *key2, uint len2) +{ +uint len1 = key1->len; +int ans; + + if( ans = memcmp (key1->key, key2, len1 > len2 ? len2 : len1) ) + return ans; + + if( len1 > len2 ) + return 1; + if( len1 < len2 ) + return -1; + + return 0; +} + +// Latch Manager + +// wait until write lock mode is clear +// and add 1 to the share count + +void bt_readlock(BtLatch *latch) +{ + do { + // see if exclusive request is pending, or granted + + if( !(volatile int)latch->request && !(volatile int)latch->exclusive ) { + // add one to counter, check write bit +#ifdef unix + if( ~__sync_fetch_and_add((volatile int *)latch, Share) & Write ) + return; +#else + if( ~_InterlockedExchangeAdd((volatile int *)latch, Share) & Write ) + return; +#endif + // didn't get latch, reduce counter by one + +#ifdef unix + __sync_fetch_and_add((volatile int *)latch, -Share); +#else + _InterlockedExchangeAdd ((volatile int *)latch, -Share); +#endif + } + + // and yield +#ifdef unix + sched_yield(); +#else + SwitchToThread(); +#endif + } while( 1 ); +} + +// wait for other read and write latches to relinquish + +void bt_writelock(BtLatch *latch) +{ +int prev; + + do { + // set exclusive access pending + +#ifdef unix + __sync_fetch_and_or((int *)latch, Pending); +#else + _InterlockedOr((int *)latch, Pending); +#endif + + // see if we can get write access + // with no readers +#ifdef unix + prev = __sync_fetch_and_or((volatile int *)latch, Write); +#else + prev = _InterlockedOr((volatile int *)latch, Write); +#endif + + // did we get exclusive access? + // if so, clear write pending + + if( !(prev & ~Pending) ) { +#ifdef unix + __sync_fetch_and_and((volatile int *)latch, ~Pending); +#else + _InterlockedAnd((volatile int *)latch, ~Pending); +#endif + return; + } + + // reset our Write mode if it was clear before + + if( !(prev & Write) ) { +#ifdef unix + __sync_fetch_and_and((volatile int *)latch, ~Write); +#else + _InterlockedAnd((volatile int *)latch, ~Write); +#endif + } + + // otherwise yield + +#ifdef unix + sched_yield(); +#else + SwitchToThread(); +#endif + } while( 1 ); +} + +// try to obtain write lock + +// return 1 if obtained, +// 0 if already write locked + +int bt_writetry(BtLatch *latch) +{ +int prev; + + // see if we can get write access + // with no readers +#ifdef unix + prev = __sync_fetch_and_or((volatile int *)latch, Write); +#else + prev = _InterlockedOr((volatile int *)latch, Write); +#endif + + // did we get exclusive access? + // if so, return OK + + if( !(prev & ~Pending) ) + return 1; + + // reset our Write mode if it was clear before + + if( !(prev & Write) ) { +#ifdef unix + __sync_fetch_and_and((volatile int *)latch, ~Write); +#else + _InterlockedAnd((volatile int *)latch, ~Write); +#endif + } + return 0; +} + +// clear write mode + +void bt_releasewrite(BtLatch *latch) +{ +#ifdef unix + __sync_fetch_and_and((int *)latch, ~Write); +#else + _InterlockedAnd ((int *)latch, ~Write); +#endif +} + +// decrement reader count + +void bt_releaseread(BtLatch *latch) +{ +#ifdef unix + __sync_fetch_and_add((int *)latch, -Share); +#else + _InterlockedExchangeAdd((int *)latch, -Share); +#endif +} + +// Buffer Pool mgr + +// find segment in pool +// must be called with hashslot idx locked +// return NULL if not there +// otherwise return node + +BtPool *bt_findpool(BtDb *bt, uid page_no, uint idx) +{ +BtPool *pool; +uint slot; + + // compute start of hash chain in pool + + if( slot = bt->mgr->hash[idx] ) + pool = (BtPool *)(bt->mgr->nodes + slot * (sizeof(BtPool) + (bt->mgr->poolmask + 1) * sizeof(BtLatchSet))); + else + return NULL; + + page_no &= ~bt->mgr->poolmask; + + while( pool->basepage != page_no ) + if( pool = pool->hashnext ) + continue; + else + return NULL; + + return pool; +} + +// add segment to hash table + +void bt_linkhash(BtDb *bt, BtPool *pool, uid page_no, int idx) +{ +BtPool *node; +uint slot; + + pool->hashprev = pool->hashnext = NULL; + pool->basepage = page_no & ~bt->mgr->poolmask; + pool->lru = 1; + + if( slot = bt->mgr->hash[idx] ) { + node = (BtPool *)(bt->mgr->nodes + slot * (sizeof(BtPool) + (bt->mgr->poolmask + 1) * sizeof(BtLatchSet))); + pool->hashnext = node; + node->hashprev = pool; + } + + 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 = (BtPool *)(bt->mgr->nodes + hashslot * (sizeof(BtPool) + (bt->mgr->poolmask + 1) * sizeof(BtLatchSet))); + + // 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) +{ +off64_t off = (page_no & ~bt->mgr->poolmask) << bt->mgr->page_bits; +off64_t limit = off + ((bt->mgr->poolmask+1) << bt->mgr->page_bits); +int flag; + +#ifdef unix + flag = PROT_READ | ( bt->mgr->mode == BT_ro ? 0 : PROT_WRITE ); + 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 + 1) / 8, 0, (bt->mgr->poolmask + 1)/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); + if( !pool->hmap ) + return bt->err = BTERR_map; + + flag = ( bt->mgr->mode == BT_ro ? FILE_MAP_READ : FILE_MAP_WRITE ); + pool->map = MapViewOfFile(pool->hmap, flag, (DWORD)(off >> 32), (DWORD)off, (bt->mgr->poolmask+1) << bt->mgr->page_bits); + if( !pool->map ) + return bt->err = BTERR_map; +#endif + return bt->err = 0; +} + +// find or place requested page in segment-pool +// return pool table entry, incrementing pin + +BtPool *bt_pinpage(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_readlock (&bt->mgr->latch[idx]); + + // look up in hash table + + if( pool = bt_findpool(bt, page_no, idx) ) { +#ifdef unix + __sync_fetch_and_add(&pool->pin, 1); +#else + _InterlockedIncrement (&pool->pin); +#endif + bt_releaseread (&bt->mgr->latch[idx]); + pool->lru++; + return pool; + } + + // upgrade to write lock + + bt_releaseread (&bt->mgr->latch[idx]); + bt_writelock (&bt->mgr->latch[idx]); + + // try to find page in pool with write lock + + if( pool = bt_findpool(bt, page_no, idx) ) { +#ifdef unix + __sync_fetch_and_add(&pool->pin, 1); +#else + _InterlockedIncrement (&pool->pin); +#endif + bt_releasewrite (&bt->mgr->latch[idx]); + pool->lru++; + return pool; + } + + // allocate a new pool node + // and add to hash table + +#ifdef unix + slot = __sync_fetch_and_add(&bt->mgr->poolcnt, 1); +#else + slot = _InterlockedIncrement (&bt->mgr->poolcnt) - 1; +#endif + + if( ++slot < bt->mgr->poolmax ) { + pool = (BtPool *)(bt->mgr->nodes + slot * (sizeof(BtPool) + (bt->mgr->poolmask + 1) * sizeof(BtLatchSet))); + pool->slot = slot; + + 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]); + return pool; + } + + // pool table is full + // find best pool entry to evict + +#ifdef unix + __sync_fetch_and_add(&bt->mgr->poolcnt, -1); +#else + _InterlockedDecrement (&bt->mgr->poolcnt); +#endif + + while( 1 ) { +#ifdef unix + victim = __sync_fetch_and_add(&bt->mgr->evicted, 1); +#else + victim = _InterlockedIncrement (&bt->mgr->evicted) - 1; +#endif + victim %= bt->mgr->hashsize; + + // try to get write lock + // skip entry if not obtained + + if( !bt_writetry (&bt->mgr->latch[victim]) ) + continue; + + // if cache entry is empty + // or no slots are unpinned + // skip this entry + + if( !(pool = bt_findlru(bt, bt->mgr->hash[victim])) ) { + bt_releasewrite (&bt->mgr->latch[victim]); + continue; + } + + // unlink victim pool node from hash table + + if( node = pool->hashprev ) + node->hashnext = pool->hashnext; + else if( node = pool->hashnext ) + bt->mgr->hash[victim] = node->slot; + else + bt->mgr->hash[victim] = 0; + + if( node = pool->hashnext ) + node->hashprev = pool->hashprev; + + bt_releasewrite (&bt->mgr->latch[victim]); + + // remove old file mapping +#ifdef unix + munmap (pool->map, (bt->mgr->poolmask+1) << bt->mgr->page_bits); +#else + FlushViewOfFile(pool->map, 0); + UnmapViewOfFile(pool->map); + CloseHandle(pool->hmap); +#endif + pool->map = NULL; + + // create new pool mapping + // and link into hash table + + 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]); + 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) +{ +BtLatchSet *set; +BtPool *pool; +uint subpage; +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; + + set = pool->pagelatch + subpage; + 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 + 1)/8)[idx] >> bit) & 1 ) { + madvise (page, bt->mgr->page_size, MADV_WILLNEED); + (bt->mgr->pooladvise + pool->slot * (bt->mgr->poolmask + 1)/8)[idx] |= 1 << bit; + } + } +#endif + + switch( mode ) { + case BtLockRead: + bt_readlock (set->readwr); + break; + case BtLockWrite: + bt_writelock (set->readwr); + break; + case BtLockAccess: + bt_readlock (set->access); + break; + case BtLockDelete: + bt_writelock (set->access); + break; + case BtLockParent: + bt_writelock (set->parent); + break; + default: + return bt->err = BTERR_lock; + } + + if( pageptr ) + *pageptr = page; + + return bt->err = 0; +} + +// remove write, read, or parent lock on requested page_no. + +BTERR bt_unlockpage(BtDb *bt, uid page_no, BtLock mode) +{ +uint subpage, idx; +BtLatchSet *set; +BtPool *pool; + + // 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]); + + 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]); + set = pool->pagelatch + subpage; + + switch( mode ) { + case BtLockRead: + bt_releaseread (set->readwr); + break; + case BtLockWrite: + bt_releasewrite (set->readwr); + break; + case BtLockAccess: + bt_releaseread (set->access); + break; + case BtLockDelete: + bt_releasewrite (set->access); + break; + case BtLockParent: + bt_releasewrite (set->parent); + 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 + + if( bt_unlockpage(bt, page_no, BtLockDelete) ) + return bt->err; + + return 0; +} + +// allocate a new page and write page into it + +uid bt_newpage(BtDb *bt, BtPage page) +{ +uid new_page; +BtPage pmap; +int reuse; + + // lock page zero + + if ( bt_lockpage(bt, ALLOC_page, BtLockWrite, &bt->alloc) ) + return 0; + + // 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) ) + return 0; + reuse = 1; + } else { + new_page = bt_getid(bt->alloc->right); + bt_putid(bt->alloc->right, new_page+1); + 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; + } +#else + // bring new page into pool and copy page. + // this will extend the file into the new pages. + + if( bt_lockpage(bt, new_page, BtLockWrite, &pmap) ) + return 0; + + memcpy(pmap, page, bt->mgr->page_size); + + if( bt_unlockpage (bt, new_page, BtLockWrite) ) + return 0; +#endif + // unlock page zero + + if ( bt_unlockpage(bt, ALLOC_page, BtLockWrite) ) + return 0; + + return new_page; +} + +// find slot in page for given key at a given level + +int bt_findslot (BtDb *bt, unsigned char *key, uint len) +{ +uint diff, higher = bt->page->cnt, low = 1, slot; + + // low is the lowest candidate, higher is already + // tested as .ge. the given key, loop ends when they meet + + while( diff = higher - low ) { + slot = low + ( diff >> 1 ); + if( keycmp (keyptr(bt->page, slot), key, len) < 0 ) + low = slot + 1; + else + higher = slot; + } + + return higher; +} + +// 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) +{ +uid page_no = ROOT_page, prevpage = 0; +uint drill = 0xff, slot; +uint mode, prevmode; + + // start at root of btree and drill down + + do { + // determine lock mode of drill level + mode = (lock == BtLockWrite) && (drill == lvl) ? BtLockWrite : BtLockRead; + + bt->page_no = page_no; + + // obtain access lock using lock chaining with Access mode + + if( page_no > ROOT_page ) + if( bt_lockpage(bt, page_no, BtLockAccess, NULL) ) + return 0; + + if( prevpage ) + if( bt_unlockpage(bt, prevpage, prevmode) ) + return 0; + + // obtain read lock using lock chaining + // and pin page contents + + if( bt_lockpage(bt, page_no, mode, &bt->page) ) + return 0; + + if( page_no > ROOT_page ) + if( bt_unlockpage(bt, page_no, BtLockAccess) ) + return 0; + + // re-read and re-lock root after determining actual level of root + + if( bt->page_no == ROOT_page ) + if( bt->page->lvl != drill) { + drill = bt->page->lvl; + + if( lock == BtLockWrite && drill == lvl ) + if( bt_unlockpage(bt, page_no, mode) ) + return 0; + else + continue; + } + + // if page is being deleted, + // move back to preceeding page + + if( bt->page->kill ) { + page_no = bt_getid (bt->page->right); + continue; + } + + // find key on page at this level + // and descend to requested level + + slot = bt_findslot (bt, key, len); + + // is this slot a foster child? + + if( slot <= bt->page->cnt - bt->page->foster ) + if( drill == lvl ) + return slot; + else + drill--; + + while( slotptr(bt->page, slot)->dead ) + if( slot++ < bt->page->cnt ) + continue; + else + return bt->err = BTERR_struct, 0; + + // continue down / right using overlapping locks + // to protect pages being killed or split. + + prevmode = mode; + prevpage = bt->page_no; + page_no = bt_getid(slotptr(bt->page, slot)->id); + } while( page_no ); + + // return error on end of chain + + bt->err = BTERR_struct; + return 0; // return error +} + +// find and delete key on page by marking delete flag bit +// when page becomes empty, delete it from the btree + +BTERR bt_deletekey (BtDb *bt, unsigned char *key, uint len, uint lvl) +{ +unsigned char leftkey[256], rightkey[256]; +uid page_no, right; +uint slot, tod; +BtKey ptr; + + 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 + + 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--; + } + + // return if page is not empty, or it has no right sibling + + right = bt_getid(bt->page->right); + page_no = bt->page_no; + + if( !right || bt->page->act ) + return bt_unlockpage(bt, page_no, BtLockWrite); + + // obtain Parent lock over write lock + + if( bt_lockpage(bt, page_no, BtLockParent, NULL) ) + return bt->err; + + // cache copy of key to delete + + ptr = keyptr(bt->page, bt->page->cnt); + memcpy(leftkey, ptr, ptr->len + 1); + + // lock and map right page + + if ( bt_lockpage(bt, right, BtLockWrite, &bt->temp) ) + return bt->err; + + // pull contents of next page into current empty page + memcpy (bt->page, bt->temp, bt->mgr->page_size); + + // cache copy of key to update + ptr = keyptr(bt->temp, bt->temp->cnt); + memcpy(rightkey, ptr, ptr->len + 1); + + // Mark right page as deleted and point it to left page + // until we can post updates at higher level. + + bt_putid(bt->temp->right, page_no); + bt->temp->kill = 1; + bt->temp->cnt = 0; + + if( bt_unlockpage(bt, right, BtLockWrite) ) + return bt->err; + if( bt_unlockpage(bt, page_no, BtLockWrite) ) + return bt->err; + + // delete old lower key to consolidated node + + if( bt_deletekey (bt, leftkey + 1, *leftkey, lvl + 1) ) + return bt->err; + + // redirect higher key directly to consolidated node + + if( slot = bt_loadpage (bt, rightkey+1, *rightkey, lvl+1, BtLockWrite) ) + ptr = keyptr(bt->page, slot); + else + return bt->err; + + // since key already exists, update id + + if( keycmp (ptr, rightkey+1, *rightkey) ) + return bt->err = BTERR_struct; + + slotptr(bt->page, slot)->dead = 0; + bt_putid(slotptr(bt->page,slot)->id, page_no); + bt_unlockpage(bt, bt->page_no, BtLockWrite); + + // obtain write lock and + // add right block to free chain + + if( bt_freepage (bt, right) ) + return bt->err; + + // remove ParentModify lock + + if( bt_unlockpage(bt, page_no, BtLockParent) ) + return bt->err; + + return 0; +} + +// find key in leaf level and return row-id + +uid bt_findkey (BtDb *bt, unsigned char *key, uint len) +{ +uint slot; +BtKey ptr; +uid id; + + if( slot = bt_loadpage (bt, key, len, 0, BtLockRead) ) + ptr = keyptr(bt->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; + + return id; +} + +// check page for space available, +// clean if necessary and return +// 0 - page needs splitting +// 1 - go ahead + +uint bt_cleanpage(BtDb *bt, uint amt) +{ +uint nxt = bt->mgr->page_size; +BtPage page = bt->page; +uint cnt = 0, idx = 0; +uint max = page->cnt; +BtKey key; + + if( page->min >= (max+1) * sizeof(BtSlot) + sizeof(*page) + amt + 1 ) + return 1; + + // skip cleanup if nothing to reclaim + + if( !page->dirty ) + return 0; + + memcpy (bt->frame, page, bt->mgr->page_size); + + // skip page info and set rest of page to zero + + memset (page+1, 0, bt->mgr->page_size - sizeof(*page)); + page->dirty = 0; + page->act = 0; + + // try cleaning up page first + + while( cnt++ < max ) { + // always leave fence key and foster children in list + if( cnt < max - page->foster && slotptr(bt->frame,cnt)->dead ) + continue; + + // copy key + 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; + + // see if page has enough space now, or does it need splitting? + + if( page->min >= (idx+1) * sizeof(BtSlot) + sizeof(*page) + amt + 1 ) + return 1; + + return 0; +} + +// add key to page +// return with page unlocked + +BTERR bt_addkeytopage (BtDb *bt, uint slot, unsigned char *key, uint len, uid id, uint tod) +{ +BtPage page = bt->page; +uint idx; + + // calculate next available slot and copy key into page + + page->min -= len + 1; + ((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; + + return bt_unlockpage(bt, bt->page_no, BtLockWrite); +} + +// split the root and raise the height of the btree + +BTERR bt_splitroot(BtDb *bt, uid right) +{ +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) + + root->act--; + root->cnt--; + root->foster--; + + // Save left fence key. + + key = keyptr(root, root->cnt); + memcpy (fencekey, key, key->len + 1); + + // copy the lower keys into a new left page + + 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 + + memset (root+1, 0, bt->mgr->page_size - sizeof(*root)); + + // insert left fence key on empty newroot page + + nxt -= *fencekey + 1; + memcpy ((unsigned char *)root + nxt, fencekey, *fencekey + 1); + bt_putid(slotptr(root, 1)->id, new_page); + slotptr(root, 1)->off = nxt; + + // insert stopper 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); + slotptr(root, 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++; + + // release root (bt->page) + + return bt_unlockpage(bt, bt->page_no, BtLockWrite); +} + +// split already locked full node +// 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; +BtPage page = bt->page; +uint tod = time(NULL); +uint lvl = page->lvl; +uid new_page, right; +BtKey key; + + // initialize frame buffer + + memset (bt->frame, 0, bt->mgr->page_size); + max = page->cnt - page->foster; + tod = (uint)time(NULL); + cnt = max / 2; + idx = 0; + + // split higher half of keys to bt->frame + // leaving foster children in the left node. + + while( cnt++ < max ) { + key = keyptr(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); + slotptr(bt->frame, idx)->tod = slotptr(page, cnt)->tod; + slotptr(bt->frame, idx)->off = nxt; + bt->frame->act++; + } + + // transfer right link node + + if( page_no > ROOT_page ) { + right = bt_getid (page->right); + bt_putid(bt->frame->right, right); + } + + 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 frame to it. + + if( !(new_page = bt_newpage(bt, bt->frame)) ) + return bt->err; + + // remember fence key for new page to add + // as foster child + + key = keyptr(bt->frame, idx); + memcpy (fencekey, key, key->len + 1); + + // update lower keys and foster children 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->act = 0; + cnt = 0; + idx = 0; + + // assemble page of smaller keys + // to remain in the old page + + 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++; + } + + // insert new foster child at beginning 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 if any + + 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); + slotptr(page, idx)->tod = slotptr(bt->frame, cnt)->tod; + slotptr(page, idx)->off = nxt; + page->act++; + } + + 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); + + // release wr lock on page + + if( bt_unlockpage (bt, page_no, BtLockWrite) ) + return bt->err; + + // obtain ParentModification lock for current page + // to fix fence key and highest foster child on page + + if( bt_lockpage (bt, page_no, BtLockParent, NULL) ) + return bt->err; + + // get our highest foster child key to find in parent node + + if( bt_lockpage (bt, page_no, BtLockRead, &page) ) + return bt->err; + + key = keyptr(page, page->cnt); + memcpy (fencekey, key, key->len+1); + + if( bt_unlockpage (bt, page_no, BtLockRead) ) + return bt->err; + +try_again: + + do { + slot = bt_loadpage (bt, fencekey + 1, *fencekey, lvl + 1, BtLockWrite); + + if( !slot ) + return bt->err; + + // check if parent page has enough space for any possible key + + if( bt_cleanpage (bt, 256) ) + break; + + if( bt_splitpage (bt) ) + return bt->err; + } while( 1 ); + + // see if we are still a foster child from another node + + if( bt_getid (slotptr(bt->page, slot)->id) != page_no ) { + bt_unlockpage (bt, bt->page_no, BtLockWrite); +#ifdef unix + sched_yield(); +#else + SwitchToThread(); +#endif + goto try_again; + } + + // wait until readers from parent get their locks + + if( bt_lockpage (bt, page_no, BtLockDelete, NULL) ) + return bt->err; + + if( bt_lockpage (bt, page_no, BtLockWrite, &page) ) + return bt->err; + + // switch parent fence key to foster child + + if( slotptr(page, page->cnt)->dead ) + slotptr(bt->page, slot)->dead = 1; + else + bt_putid (slotptr(bt->page, slot)->id, bt_getid(slotptr(page, page->cnt)->id)); + + // remove highest foster child from our page + // add our new fence key to parent + + page->cnt--; + page->act--; + page->foster--; + page->dirty = 1; + key = keyptr(page, page->cnt); + + if( bt_addkeytopage (bt, slot, key->key, key->len, page_no, tod) ) + return bt->err; + + if( bt_unlockpage (bt, page_no, BtLockDelete) ) + return bt->err; + + if( bt_unlockpage (bt, page_no, BtLockWrite) ) + return bt->err; + + return bt_unlockpage (bt, page_no, BtLockParent); +} + +// Insert new key into the btree at leaf level. + +BTERR bt_insertkey (BtDb *bt, unsigned char *key, uint len, uid id, uint tod) +{ +uint slot, idx; +BtPage page; +BtKey ptr; + + while( 1 ) { + if( slot = bt_loadpage (bt, key, len, 0, 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; + + 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); + } + + // check if page has enough space + + if( bt_cleanpage (bt, len) ) + break; + + if( bt_splitpage (bt) ) + return bt->err; + } + + return bt_addkeytopage (bt, slot, key, len, id, tod); +} + +// cache page of keys into cursor and return starting slot for given key + +uint bt_startkey (BtDb *bt, unsigned char *key, uint len) +{ +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; + + return slot; +} + +// return next slot for cursor page +// or slide cursor right into next page + +uint bt_nextkey (BtDb *bt, uint slot) +{ +BtPage page; +uid right; + + do { + right = bt_getid(bt->cursor->right); + while( slot++ < bt->cursor->cnt - bt->cursor->foster ) + if( slotptr(bt->cursor,slot)->dead ) + continue; + else if( right || (slot < bt->cursor->cnt - bt->cursor->foster) ) + return slot; + else + break; + + if( !right ) + break; + + bt->cursor_page = right; + + if( bt_lockpage(bt, right, BtLockRead, &page) ) + return 0; + + memcpy (bt->cursor, page, bt->mgr->page_size); + + if ( bt_unlockpage(bt, right, BtLockRead) ) + return 0; + + slot = 0; + } while( 1 ); + + return bt->err = 0; +} + +BtKey bt_key(BtDb *bt, uint slot) +{ + return keyptr(bt->cursor, slot); +} + +uid bt_uid(BtDb *bt, uint slot) +{ + return bt_getid(slotptr(bt->cursor,slot)->id); +} + +uint bt_tod(BtDb *bt, uint slot) +{ + return slotptr(bt->cursor,slot)->tod; +} + + +#ifdef STANDALONE + +typedef struct { + char type, idx; + char *infile; + BtMgr *mgr; + int num; +} ThreadArg; + +// standalone program to index file of keys +// then list them onto std-out + +#ifdef unix +void *index_file (void *arg) +#else +uint __stdcall index_file (void *arg) +#endif +{ +int line = 0, found = 0, cnt = 0; +uid next, page_no = LEAF_page; // start on first page of leaves +unsigned char key[256]; +ThreadArg *args = arg; +int ch, len = 0, slot; +time_t tod[1]; +BtPage page; +BtKey ptr; +BtDb *bt; +FILE *in; + + bt = bt_open (args->mgr); + time (tod); + + switch(args->type | 0x20) + { + case 'w': + fprintf(stderr, "started indexing for %s\n", args->infile); + if( in = fopen (args->infile, "rb") ) + while( ch = getc(in), ch != EOF ) + if( ch == '\n' ) + { + 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; + + if( bt_insertkey (bt, key, len, line, *tod) ) + fprintf(stderr, "Error %d Line: %d\n", bt->err, line), exit(0); + len = 0; + } + else if( len < 255 ) + key[len++] = ch; + fprintf(stderr, "finished %s for %d keys\n", args->infile, line); + break; + + case 'd': + fprintf(stderr, "started deleting keys for %s\n", args->infile); + if( in = fopen (args->infile, "rb") ) + while( ch = getc(in), ch != EOF ) + if( ch == '\n' ) + { + 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; + + if( bt_deletekey (bt, key, len, 0) ) + fprintf(stderr, "Error %d Line: %d\n", bt->err, line), exit(0); + len = 0; + } + else if( len < 255 ) + key[len++] = ch; + fprintf(stderr, "finished %s for keys, %d \n", args->infile, line); + break; + + case 'f': + fprintf(stderr, "started finding keys for %s\n", args->infile); + if( in = fopen (args->infile, "rb") ) + while( ch = getc(in), ch != EOF ) + if( ch == '\n' ) + { + 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; + + if( bt_findkey (bt, key, len) ) + found++; + else if( bt->err ) + fprintf(stderr, "Error %d Syserr %d Line: %d\n", bt->err, errno, line), exit(0); + len = 0; + } + else if( len < 255 ) + key[len++] = ch; + fprintf(stderr, "finished %s for %d keys, found %d\n", args->infile, line, found); + 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); + } + + break; + + case 'c': + fprintf(stderr, "started reading\n"); + + 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 ); + + cnt--; // remove stopper key + fprintf(stderr, " Total keys read %d\n", cnt); + break; + } + + bt_close (bt); +#ifdef unix + return NULL; +#else + return 0; +#endif +} + +typedef struct timeval timer; + +int main (int argc, char **argv) +{ +int idx, cnt, len, slot, err; +int segsize, bits = 16; +#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; +int num = 0; +char key[1]; +BtMgr *mgr; +BtKey ptr; +BtDb *bt; + + if( argc < 3 ) { + fprintf (stderr, "Usage: %s idx_file Read/Write/Scan/Delete/Find [page_bits mapped_segments seg_bits line_numbers src_file1 src_file2 ... ]\n", argv[0]); + 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 = 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 + + if( argc > 3 ) + bits = atoi(argv[3]); + + if( argc > 4 ) + poolsize = atoi(argv[4]); + + if( !poolsize ) + fprintf (stderr, "Warning: no mapped_pool\n"); + + if( poolsize > 65535 ) + fprintf (stderr, "Warning: mapped_pool > 65535 segments\n"); + + if( argc > 5 ) + segsize = atoi(argv[5]); + else + segsize = 4; // 16 pages per mmap segment + + if( argc > 6 ) + num = atoi(argv[6]); + + cnt = argc - 7; +#ifdef unix + threads = malloc (cnt * sizeof(pthread_t)); +#else + threads = GlobalAlloc (GMEM_FIXED|GMEM_ZEROINIT, cnt * sizeof(HANDLE)); +#endif + args = malloc (cnt * sizeof(ThreadArg)); + + mgr = bt_mgr ((argv[1]), BT_rw, bits, poolsize, segsize, poolsize / 8); + + if( !mgr ) { + fprintf(stderr, "Index Open Error %s\n", argv[1]); + exit (1); + } + + // fire off threads + + for( idx = 0; idx < cnt; idx++ ) { + args[idx].infile = argv[idx + 7]; + args[idx].type = argv[2][0]; + args[idx].mgr = mgr; + args[idx].num = num; + args[idx].idx = idx; +#ifdef unix + if( err = pthread_create (threads + idx, NULL, index_file, args + idx) ) + fprintf(stderr, "Error creating thread %d\n", err); +#else + threads[idx] = (HANDLE)_beginthreadex(NULL, 65536, index_file, args + idx, 0, NULL); +#endif + } + + // wait for termination + +#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); + bt_mgrclose (mgr); +} + +#endif //STANDALONE diff --git a/fosterbtreec.c b/fosterbtreec.c new file mode 100644 index 0000000..3b07013 --- /dev/null +++ b/fosterbtreec.c @@ -0,0 +1,2166 @@ +// foster btree version d +// 26 DEC 2013 + +// author: karl malbrain, malbrain@cal.berkeley.edu + +/* +This work, including the source code, documentation +and related data, is placed into the public domain. + +The orginal author is Karl Malbrain. + +THIS SOFTWARE IS PROVIDED AS-IS WITHOUT WARRANTY +OF ANY KIND, NOT EVEN THE IMPLIED WARRANTY OF +MERCHANTABILITY. THE AUTHOR OF THIS SOFTWARE, +ASSUMES _NO_ RESPONSIBILITY FOR ANY CONSEQUENCE +RESULTING FROM THE USE, MODIFICATION, OR +REDISTRIBUTION OF THIS SOFTWARE. +*/ + +// Please see the project home page for documentation +// code.google.com/p/high-concurrency-btree + +#define _FILE_OFFSET_BITS 64 +#define _LARGEFILE64_SOURCE + +#ifdef linux +#define _GNU_SOURCE +#endif + +#ifdef unix +#include +#include +#include +#include +#include +#include +#include +#include +#else +#define WIN32_LEAN_AND_MEAN +#include +#include +#include +#include +#include +#include +#include +#endif + +#include +#include + +typedef unsigned long long uid; + +#ifndef unix +typedef unsigned long long off64_t; +typedef unsigned short ushort; +typedef unsigned int uint; +#endif + +#define BT_ro 0x6f72 // ro +#define BT_rw 0x7772 // rw + +#define BT_maxbits 24 // maximum page size in bits +#define BT_minbits 9 // minimum page size in bits +#define BT_minpage (1 << BT_minbits) // minimum page size +#define BT_maxpage (1 << BT_maxbits) // maximum page size + +/* +There are five lock types for each node in three independent sets: +1. (set 1) AccessIntent: Sharable. Going to Read the node. Incompatible with NodeDelete. +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. +*/ + +typedef enum{ + BtLockAccess, + BtLockDelete, + BtLockRead, + BtLockWrite, + BtLockParent +}BtLock; + +// Define the length of the page and key pointers + +#define BtId 6 + +// Page key slot definition. + +// If BT_maxbits is 15 or less, you can save 4 bytes +// for each key stored by making the first two uints +// into ushorts. You can also save 4 bytes by removing +// the tod field from the key. + +// Keys are marked dead, but remain on the page until +// it cleanup is called. The fence key (highest key) for +// the page is always present, even after cleanup. + +typedef struct { + uint off:BT_maxbits; // page offset for key start + uint dead:1; // set for deleted key + uint tod; // time-stamp for key + unsigned char id[BtId]; // id associated with key +} BtSlot; + +// The key structure occupies space at the upper end of +// each page. It's a length byte followed by the value +// bytes. + +typedef struct { + unsigned char len; + unsigned char key[1]; +} *BtKey; + +// The first part of an index page. +// It is immediately followed +// by the BtSlot array of keys. + +typedef struct Page { + uint cnt; // count of keys in page + uint act; // count of active keys + uint min; // next key offset + uint foster; // count of foster children + unsigned char bits; // page size in bits + unsigned char lvl:6; // level of page + unsigned char kill:1; // page is being deleted + unsigned char dirty:1; // page needs to be cleaned + unsigned char right[BtId]; // page number to right +} *BtPage; + +// latch table lock structure + +// exclusive is set for write access +// share is count of read accessors +// pending is count of waiting writers +// grant write lock when share == 0 + +typedef struct { +#ifdef unix + volatile uint exclusive:1; + volatile uint pending:15; + volatile uint share:16; + pthread_mutex_t mut[1]; + pthread_cond_t cond[1]; +#else + SRWLOCK srw[1]; +#endif +} BtLatch; + +typedef struct { + BtLatch readwr[1]; // read/write page lock + BtLatch access[1]; // Access Intent/Page delete + BtLatch parent[1]; // adoption of foster children +} BtLatchSet; + +// 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 pin; // mapped page pin counter + uint slot; // slot index in this array + void *hashprev; // previous pool entry for the same hash idx + void *hashnext; // next pool entry for the same hash idx +#ifndef unix + HANDLE hmap; +#endif +// array of page latch sets, one for each page in map segment + BtLatchSet pagelatch[0]; +} BtPool; + +// The object structure for Btree access + +typedef struct { + uint page_size; // page size + uint page_bits; // page size in bits + 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 +#else + HANDLE idx; +#endif + uint poolcnt; // highest page pool node in use + 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 + volatile uint evicted; // last evicted hash table slot + ushort *hash; // hash table of pool entries + BtLatch *latch; // latches for hash table slots + char *nodes; // memory pool page segments +} 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 ( page 0 ) + 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 + uid page_no; // current page number + uid cursor_page; // current cursor page number + unsigned char *mem; // frame, cursor, page memory buffer + int err; // last error +} BtDb; + +typedef enum { + BTERR_ok = 0, + BTERR_struct, + BTERR_ovflw, + BTERR_lock, + BTERR_map, + BTERR_wrt, + BTERR_hash +} BTERR; + +// 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); +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); + +// manager functions +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 + +extern BtKey bt_key (BtDb *bt, uint slot); +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 + +// Number of levels to create in a new BTree + +#define MIN_lvl 2 + +// The page is allocated from low and hi ends. +// The key offsets and row-id's are allocated +// from the bottom, while the text of the key +// is allocated from the top. When the two +// areas meet, the page is split into two. + +// 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. + +// The b-tree root is always located at page 1. +// The first leaf page of level zero is always +// located on page 2. + +// 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. + +// 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. + +// Groups of pages called segments from the btree are +// cached with memory mapping. 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. + +// Page 0 is dedicated to lock for new page extensions, +// and chains empty pages together for reuse. + +// Empty pages are chained together through the ALLOC page and reused. + +// Access macros to address slot and key values from the page + +#define slotptr(page, slot) (((BtSlot *)(page+1)) + (slot-1)) +#define keyptr(page, slot) ((BtKey)((unsigned char*)(page) + slotptr(page, slot)->off)) + +void bt_putid(unsigned char *dest, uid id) +{ +int i = BtId; + + while( i-- ) + dest[i] = (unsigned char)id, id >>= 8; +} + +uid bt_getid(unsigned char *src) +{ +uid id = 0; +int i; + + for( i = 0; i < BtId; i++ ) + id <<= 8, id |= *src++; + + return id; +} + +void bt_mgrclose (BtMgr *mgr) +{ +BtPool *pool; +uint slot; + + // release mapped pages + // note that slot zero is never used + + for( slot = 1; slot < mgr->poolmax; slot++ ) { + pool = (BtPool *)(mgr->nodes + slot * (sizeof(BtPool) + (mgr->poolmask + 1) * sizeof(BtLatchSet))); + if( pool->slot ) +#ifdef unix + munmap (pool->map, (mgr->poolmask+1) << mgr->page_bits); +#else + { + FlushViewOfFile(pool->map, 0); + UnmapViewOfFile(pool->map); + CloseHandle(pool->hmap); + } +#endif + } + +#ifdef unix + close (mgr->idx); + free (mgr->nodes); + free (mgr->hash); + free (mgr->latch); + free (mgr->pooladvise); + free (mgr); +#else + FlushFileBuffers(mgr->idx); + CloseHandle(mgr->idx); + GlobalFree (mgr->nodes); + GlobalFree (mgr->hash); + GlobalFree (mgr->latch); + GlobalFree (mgr); +#endif +} + +// close and release memory + +void bt_close (BtDb *bt) +{ +#ifdef unix + if ( bt->mem ) + free (bt->mem); +#else + if ( bt->mem) + VirtualFree (bt->mem, 0, MEM_RELEASE); +#endif + free (bt); +} + +// open/create new btree buffer manager + +// call with file_name, BT_openmode, bits in page size (e.g. 16), +// size of mapped page pool (e.g. 8192) + +BtMgr *bt_mgr (char *name, uint mode, uint bits, uint poolmax, uint segsize, uint hashsize) +{ +uint lvl, attr, cacheblk, last, slot, idx; +BtPage alloc; +int lockmode; +off64_t size; +uint amt[1]; +BtMgr* mgr; +BtKey key; + +#ifndef unix +SYSTEM_INFO sysinfo[1]; +#endif + + // determine sanity of page size and buffer pool + + if( bits > BT_maxbits ) + bits = BT_maxbits; + else if( bits < BT_minbits ) + bits = BT_minbits; + + if( !poolmax ) + return NULL; // must have buffer pool + +#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; + + case BT_ro: + default: + mgr->idx = open ((char*)name, O_RDONLY); + lockmode = 0; + break; + } + if( mgr->idx == -1 ) + return free(mgr), NULL; + + cacheblk = 4096; // minimum mmap segment size for unix + +#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; + + 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; + + // normalize cacheblk to multiple of sysinfo->dwAllocationGranularity + GetSystemInfo(sysinfo); + cacheblk = sysinfo->dwAllocationGranularity; +#endif + +#ifdef unix + alloc = 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; + else + return free(mgr), free(alloc), NULL; + } else if( mode == BT_ro ) + return bt_mgrclose (mgr), NULL; +#else + alloc = 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) ) + return bt_mgrclose (mgr), NULL; + bits = alloc->bits; + } else if( mode == BT_ro ) + return bt_mgrclose (mgr), NULL; +#endif + + mgr->page_size = 1 << bits; + mgr->page_bits = bits; + + mgr->poolmax = poolmax; + mgr->mode = mode; + + if( cacheblk < mgr->page_size ) + cacheblk = mgr->page_size; + + // mask for partial memmaps + + mgr->poolmask = (cacheblk >> bits) - 1; + + // see if requested size of pages per memmap is greater + + if( (1 << segsize) > mgr->poolmask ) + mgr->poolmask = (1 << segsize) - 1; + + mgr->seg_bits = 0; + + while( (1 << mgr->seg_bits) <= mgr->poolmask ) + mgr->seg_bits++; + + mgr->hashsize = hashsize; + +#ifdef unix + mgr->nodes = calloc (poolmax, (sizeof(BtPool) + (mgr->poolmask + 1) * sizeof(BtLatchSet))); + mgr->hash = calloc (hashsize, sizeof(ushort)); + mgr->latch = calloc (hashsize, sizeof(BtLatch)); + mgr->pooladvise = calloc (poolmax, (mgr->poolmask + 1) / 8); +#else + mgr->nodes = GlobalAlloc (GMEM_FIXED|GMEM_ZEROINIT, poolmax * (sizeof(BtPool) + (mgr->poolmask + 1) * sizeof(BtLatchSet))); + mgr->hash = GlobalAlloc (GMEM_FIXED|GMEM_ZEROINIT, hashsize * sizeof(ushort)); + mgr->latch = GlobalAlloc (GMEM_FIXED|GMEM_ZEROINIT, hashsize * sizeof(BtLatch)); +#endif + + // initialize buffer pool page latches + + for( slot = 1; slot < poolmax; slot++ ) { + BtLatchSet *latchset = (BtLatchSet *)(mgr->nodes + slot * (sizeof(BtPool) + (mgr->poolmask + 1) * sizeof(BtLatchSet))); + for( idx = 0; idx < mgr->poolmask + 1; idx++ ) { +#ifdef unix + pthread_mutex_init (latchset[idx].readwr->mut, NULL); + pthread_cond_init (latchset[idx].readwr->cond, NULL); + pthread_mutex_init (latchset[idx].access->mut, NULL); + pthread_cond_init (latchset[idx].access->cond, NULL); + pthread_mutex_init (latchset[idx].parent->mut, NULL); + pthread_cond_init (latchset[idx].parent->cond, NULL); +#else + InitializeSRWLock (latchset[idx].readwr->srw); + InitializeSRWLock (latchset[idx].access->srw); + InitializeSRWLock (latchset[idx].parent->srw); +#endif + } + } + + // initialize buffer pool mgr latches + + for( slot = 0; slot < hashsize; slot++ ) { +#ifdef unix + pthread_mutex_init (mgr->latch[slot].mut, NULL); + pthread_cond_init (mgr->latch[slot].cond, NULL); +#else + InitializeSRWLock (mgr->latch[slot].srw); +#endif + } + + if( size || *amt ) + goto mgrxit; + + // initializes an empty b-tree with root page and page of leaves + + memset (alloc, 0, 1 << bits); + bt_putid(alloc->right, MIN_lvl+1); + alloc->bits = mgr->page_bits; + +#ifdef unix + if( write (mgr->idx, alloc, mgr->page_size) < mgr->page_size ) + return bt_mgrclose (mgr), NULL; +#else + if( !WriteFile (mgr->idx, (char *)alloc, 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; + + 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 + key->key[0] = 0xff; + key->key[1] = 0xff; + alloc->min = mgr->page_size - 3; + alloc->lvl = lvl; + alloc->cnt = 1; + alloc->act = 1; +#ifdef unix + if( write (mgr->idx, alloc, mgr->page_size) < mgr->page_size ) + return bt_mgrclose (mgr), NULL; +#else + if( !WriteFile (mgr->idx, (char *)alloc, mgr->page_size, amt, NULL) ) + return bt_mgrclose (mgr), NULL; + + if( *amt < mgr->page_size ) + return bt_mgrclose (mgr), NULL; +#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; + + while( last < MIN_lvl + 1 ) + last += mgr->poolmask + 1; + +#ifdef unix + pwrite(mgr->idx, alloc, 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) ) + return bt_mgrclose (mgr), NULL; + if( *amt < mgr->page_size ) + return bt_mgrclose (mgr), NULL; +#endif + } + +mgrxit: +#ifdef unix + free (alloc); +#else + VirtualFree (alloc, 0, MEM_RELEASE); +#endif + return mgr; +} + +// open BTree access method +// based on buffer manager + +BtDb *bt_open (BtMgr *mgr) +{ +BtDb *bt = malloc (sizeof(*bt)); + + memset (bt, 0, sizeof(*bt)); + bt->mgr = mgr; +#ifdef unix + bt->mem = malloc (3 *mgr->page_size); +#else + bt->mem = VirtualAlloc(NULL, 3 * mgr->page_size, MEM_COMMIT, PAGE_READWRITE); +#endif + bt->frame = (BtPage)bt->mem; + bt->zero = (BtPage)(bt->mem + 1 * mgr->page_size); + bt->cursor = (BtPage)(bt->mem + 2 * mgr->page_size); + return bt; +} + +// compare two keys, returning > 0, = 0, or < 0 +// as the comparison value + +int keycmp (BtKey key1, unsigned char *key2, uint len2) +{ +uint len1 = key1->len; +int ans; + + if( ans = memcmp (key1->key, key2, len1 > len2 ? len2 : len1) ) + return ans; + + if( len1 > len2 ) + return 1; + if( len1 < len2 ) + return -1; + + return 0; +} + +// Latch Manager + +// wait if exclusive request is pending, or granted +// and add 1 to the share count + +void bt_readlock(BtLatch *latch) +{ +#ifdef unix + pthread_mutex_lock (latch->mut); + + while( latch->pending || latch->exclusive ) + pthread_cond_wait (latch->cond, latch->mut); + + // add one to readers counter + + latch->share++; + pthread_mutex_unlock (latch->mut); +#else + AcquireSRWLockShared (latch->srw); +#endif +} + +// wait for other read and write latches to relinquish + +void bt_writelock(BtLatch *latch) +{ +#ifdef unix + pthread_mutex_lock (latch->mut); + latch->pending++; + + while( latch->share || latch->exclusive ) + pthread_cond_wait (latch->cond, latch->mut); + + latch->exclusive = 1; + latch->pending--; + pthread_mutex_unlock (latch->mut); +#else + AcquireSRWLockExclusive (latch->srw); +#endif +} + +// try to obtain write lock + +// return 1 if obtained, +// 0 if already write or read locked + +int bt_writetry(BtLatch *latch) +{ +int result = 0; + +#ifdef unix + pthread_mutex_lock (latch->mut); + + if( !latch->share && !latch->exclusive ) + result = latch->exclusive = 1; + + pthread_mutex_unlock (latch->mut); +#else + result = TryAcquireSRWLockExclusive (latch->srw); +#endif + return result; +} + +// clear write mode + +void bt_releasewrite(BtLatch *latch) +{ +#ifdef unix + pthread_mutex_lock (latch->mut); + latch->exclusive = 0; + pthread_cond_broadcast (latch->cond); + pthread_mutex_unlock (latch->mut); +#else + ReleaseSRWLockExclusive (latch->srw); +#endif +} + +// decrement reader count + +void bt_releaseread(BtLatch *latch) +{ +#ifdef unix + pthread_mutex_lock (latch->mut); + + if( !--latch->share && latch->pending ) + pthread_cond_broadcast (latch->cond); + + pthread_mutex_unlock (latch->mut); +#else + ReleaseSRWLockShared (latch->srw); +#endif +} + +// Buffer Pool mgr + +// find segment in pool +// must be called with hashslot idx locked +// return NULL if not there +// otherwise return node + +BtPool *bt_findpool(BtDb *bt, uid page_no, uint idx) +{ +BtPool *pool; +uint slot; + + // compute start of hash chain in pool + + if( slot = bt->mgr->hash[idx] ) + pool = (BtPool *)(bt->mgr->nodes + slot * (sizeof(BtPool) + (bt->mgr->poolmask + 1) * sizeof(BtLatchSet))); + else + return NULL; + + page_no &= ~bt->mgr->poolmask; + + while( pool->basepage != page_no ) + if( pool = pool->hashnext ) + continue; + else + return NULL; + + return pool; +} + +// add segment to hash table + +void bt_linkhash(BtDb *bt, BtPool *pool, uid page_no, int idx) +{ +BtPool *node; +uint slot; + + pool->hashprev = pool->hashnext = NULL; + pool->basepage = page_no & ~bt->mgr->poolmask; + pool->lru = 1; + + if( slot = bt->mgr->hash[idx] ) { + node = (BtPool *)(bt->mgr->nodes + slot * (sizeof(BtPool) + (bt->mgr->poolmask + 1) * sizeof(BtLatchSet))); + pool->hashnext = node; + node->hashprev = pool; + } + + 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 = (BtPool *)(bt->mgr->nodes + hashslot * (sizeof(BtPool) + (bt->mgr->poolmask + 1) * sizeof(BtLatchSet))); + + // 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) +{ +off64_t off = (page_no & ~bt->mgr->poolmask) << bt->mgr->page_bits; +off64_t limit = off + ((bt->mgr->poolmask+1) << bt->mgr->page_bits); +int flag; + +#ifdef unix + flag = PROT_READ | ( bt->mgr->mode == BT_ro ? 0 : PROT_WRITE ); + 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 + 1) / 8, 0, (bt->mgr->poolmask + 1)/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); + if( !pool->hmap ) + return bt->err = BTERR_map; + + flag = ( bt->mgr->mode == BT_ro ? FILE_MAP_READ : FILE_MAP_WRITE ); + pool->map = MapViewOfFile(pool->hmap, flag, (DWORD)(off >> 32), (DWORD)off, (bt->mgr->poolmask+1) << bt->mgr->page_bits); + if( !pool->map ) + return bt->err = BTERR_map; +#endif + return bt->err = 0; +} + +// find or place requested page in segment-pool +// return pool table entry, incrementing pin + +BtPool *bt_pinpage(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_readlock (&bt->mgr->latch[idx]); + + // look up in hash table + + if( pool = bt_findpool(bt, page_no, idx) ) { +#ifdef unix + __sync_fetch_and_add(&pool->pin, 1); +#else + _InterlockedIncrement (&pool->pin); +#endif + bt_releaseread (&bt->mgr->latch[idx]); + pool->lru++; + return pool; + } + + // upgrade to write lock + + bt_releaseread (&bt->mgr->latch[idx]); + bt_writelock (&bt->mgr->latch[idx]); + + // try to find page in pool with write lock + + if( pool = bt_findpool(bt, page_no, idx) ) { +#ifdef unix + __sync_fetch_and_add(&pool->pin, 1); +#else + _InterlockedIncrement (&pool->pin); +#endif + bt_releasewrite (&bt->mgr->latch[idx]); + pool->lru++; + return pool; + } + + // allocate a new pool node + // and add to hash table + +#ifdef unix + slot = __sync_fetch_and_add(&bt->mgr->poolcnt, 1); +#else + slot = _InterlockedIncrement (&bt->mgr->poolcnt) - 1; +#endif + + if( ++slot < bt->mgr->poolmax ) { + pool = (BtPool *)(bt->mgr->nodes + slot * (sizeof(BtPool) + (bt->mgr->poolmask + 1) * sizeof(BtLatchSet))); + pool->slot = slot; + + 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]); + return pool; + } + + // pool table is full + // find best pool entry to evict + +#ifdef unix + __sync_fetch_and_add(&bt->mgr->poolcnt, -1); +#else + _InterlockedDecrement (&bt->mgr->poolcnt); +#endif + + while( 1 ) { +#ifdef unix + victim = __sync_fetch_and_add(&bt->mgr->evicted, 1); +#else + victim = _InterlockedIncrement (&bt->mgr->evicted) - 1; +#endif + victim %= bt->mgr->hashsize; + + // try to get write lock + // skip entry if not obtained + + if( !bt_writetry (&bt->mgr->latch[victim]) ) + continue; + + // if cache entry is empty + // or no slots are unpinned + // skip this entry + + if( !(pool = bt_findlru(bt, bt->mgr->hash[victim])) ) { + bt_releasewrite (&bt->mgr->latch[victim]); + continue; + } + + // unlink victim pool node from hash table + + if( node = pool->hashprev ) + node->hashnext = pool->hashnext; + else if( node = pool->hashnext ) + bt->mgr->hash[victim] = node->slot; + else + bt->mgr->hash[victim] = 0; + + if( node = pool->hashnext ) + node->hashprev = pool->hashprev; + + bt_releasewrite (&bt->mgr->latch[victim]); + + // remove old file mapping +#ifdef unix + munmap (pool->map, (bt->mgr->poolmask+1) << bt->mgr->page_bits); +#else + FlushViewOfFile(pool->map, 0); + UnmapViewOfFile(pool->map); + CloseHandle(pool->hmap); +#endif + pool->map = NULL; + + // create new pool mapping + // and link into hash table + + 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]); + 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) +{ +BtLatchSet *set; +BtPool *pool; +uint subpage; +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; + + set = pool->pagelatch + subpage; + 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 + 1)/8)[idx] >> bit) & 1 ) { + madvise (page, bt->mgr->page_size, MADV_WILLNEED); + (bt->mgr->pooladvise + pool->slot * (bt->mgr->poolmask + 1)/8)[idx] |= 1 << bit; + } + } +#endif + + switch( mode ) { + case BtLockRead: + bt_readlock (set->readwr); + break; + case BtLockWrite: + bt_writelock (set->readwr); + break; + case BtLockAccess: + bt_readlock (set->access); + break; + case BtLockDelete: + bt_writelock (set->access); + break; + case BtLockParent: + bt_writelock (set->parent); + break; + default: + return bt->err = BTERR_lock; + } + + if( pageptr ) + *pageptr = page; + + return bt->err = 0; +} + +// remove write, read, or parent lock on requested page_no. + +BTERR bt_unlockpage(BtDb *bt, uid page_no, BtLock mode) +{ +uint subpage, idx; +BtLatchSet *set; +BtPool *pool; + + // 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]); + + 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]); + set = pool->pagelatch + subpage; + + switch( mode ) { + case BtLockRead: + bt_releaseread (set->readwr); + break; + case BtLockWrite: + bt_releasewrite (set->readwr); + break; + case BtLockAccess: + bt_releaseread (set->access); + break; + case BtLockDelete: + bt_releasewrite (set->access); + break; + case BtLockParent: + bt_releasewrite (set->parent); + 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 + + if( bt_unlockpage(bt, page_no, BtLockDelete) ) + return bt->err; + + return 0; +} + +// allocate a new page and write page into it + +uid bt_newpage(BtDb *bt, BtPage page) +{ +uid new_page; +BtPage pmap; +int reuse; + + // lock page zero + + if ( bt_lockpage(bt, ALLOC_page, BtLockWrite, &bt->alloc) ) + return 0; + + // 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) ) + return 0; + reuse = 1; + } else { + new_page = bt_getid(bt->alloc->right); + bt_putid(bt->alloc->right, new_page+1); + 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; + } +#else + // bring new page into pool and copy page. + // this will extend the file into the new pages. + + if( bt_lockpage(bt, new_page, BtLockWrite, &pmap) ) + return 0; + + memcpy(pmap, page, bt->mgr->page_size); + + if( bt_unlockpage (bt, new_page, BtLockWrite) ) + return 0; +#endif + // unlock page zero + + if ( bt_unlockpage(bt, ALLOC_page, BtLockWrite) ) + return 0; + + return new_page; +} + +// find slot in page for given key at a given level + +int bt_findslot (BtDb *bt, unsigned char *key, uint len) +{ +uint diff, higher = bt->page->cnt, low = 1, slot; + + // low is the lowest candidate, higher is already + // tested as .ge. the given key, loop ends when they meet + + while( diff = higher - low ) { + slot = low + ( diff >> 1 ); + if( keycmp (keyptr(bt->page, slot), key, len) < 0 ) + low = slot + 1; + else + higher = slot; + } + + return higher; +} + +// 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) +{ +uid page_no = ROOT_page, prevpage = 0; +uint drill = 0xff, slot; +uint mode, prevmode; + + // start at root of btree and drill down + + do { + // determine lock mode of drill level + mode = (lock == BtLockWrite) && (drill == lvl) ? BtLockWrite : BtLockRead; + + bt->page_no = page_no; + + // obtain access lock using lock chaining with Access mode + + if( page_no > ROOT_page ) + if( bt_lockpage(bt, page_no, BtLockAccess, NULL) ) + return 0; + + if( prevpage ) + if( bt_unlockpage(bt, prevpage, prevmode) ) + return 0; + + // obtain read lock using lock chaining + // and pin page contents + + if( bt_lockpage(bt, page_no, mode, &bt->page) ) + return 0; + + if( page_no > ROOT_page ) + if( bt_unlockpage(bt, page_no, BtLockAccess) ) + return 0; + + // re-read and re-lock root after determining actual level of root + + if( bt->page_no == ROOT_page ) + if( bt->page->lvl != drill) { + drill = bt->page->lvl; + + if( lock == BtLockWrite && drill == lvl ) + if( bt_unlockpage(bt, page_no, mode) ) + return 0; + else + continue; + } + + // if page is being deleted, + // move back to preceeding page + + if( bt->page->kill ) { + page_no = bt_getid (bt->page->right); + continue; + } + + // find key on page at this level + // and descend to requested level + + slot = bt_findslot (bt, key, len); + + // is this slot a foster child? + + if( slot <= bt->page->cnt - bt->page->foster ) + if( drill == lvl ) + return slot; + else + drill--; + + while( slotptr(bt->page, slot)->dead ) + if( slot++ < bt->page->cnt ) + continue; + else + return bt->err = BTERR_struct, 0; + + // continue down / right using overlapping locks + // to protect pages being killed or split. + + prevmode = mode; + prevpage = bt->page_no; + page_no = bt_getid(slotptr(bt->page, slot)->id); + } while( page_no ); + + // return error on end of chain + + bt->err = BTERR_struct; + return 0; // return error +} + +// find and delete key on page by marking delete flag bit +// when page becomes empty, delete it from the btree + +BTERR bt_deletekey (BtDb *bt, unsigned char *key, uint len, uint lvl) +{ +unsigned char leftkey[256], rightkey[256]; +uid page_no, right; +uint slot, tod; +BtKey ptr; + + 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 + + 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--; + } + + // return if page is not empty, or it has no right sibling + + right = bt_getid(bt->page->right); + page_no = bt->page_no; + + if( !right || bt->page->act ) + return bt_unlockpage(bt, page_no, BtLockWrite); + + // obtain Parent lock over write lock + + if( bt_lockpage(bt, page_no, BtLockParent, NULL) ) + return bt->err; + + // cache copy of key to delete + + ptr = keyptr(bt->page, bt->page->cnt); + memcpy(leftkey, ptr, ptr->len + 1); + + // lock and map right page + + if ( bt_lockpage(bt, right, BtLockWrite, &bt->temp) ) + return bt->err; + + // pull contents of next page into current empty page + memcpy (bt->page, bt->temp, bt->mgr->page_size); + + // cache copy of key to update + ptr = keyptr(bt->temp, bt->temp->cnt); + memcpy(rightkey, ptr, ptr->len + 1); + + // Mark right page as deleted and point it to left page + // until we can post updates at higher level. + + bt_putid(bt->temp->right, page_no); + bt->temp->kill = 1; + bt->temp->cnt = 0; + + if( bt_unlockpage(bt, right, BtLockWrite) ) + return bt->err; + if( bt_unlockpage(bt, page_no, BtLockWrite) ) + return bt->err; + + // delete old lower key to consolidated node + + if( bt_deletekey (bt, leftkey + 1, *leftkey, lvl + 1) ) + return bt->err; + + // redirect higher key directly to consolidated node + + if( slot = bt_loadpage (bt, rightkey+1, *rightkey, lvl+1, BtLockWrite) ) + ptr = keyptr(bt->page, slot); + else + return bt->err; + + // since key already exists, update id + + if( keycmp (ptr, rightkey+1, *rightkey) ) + return bt->err = BTERR_struct; + + slotptr(bt->page, slot)->dead = 0; + bt_putid(slotptr(bt->page,slot)->id, page_no); + bt_unlockpage(bt, bt->page_no, BtLockWrite); + + // obtain write lock and + // add right block to free chain + + if( bt_freepage (bt, right) ) + return bt->err; + + // remove ParentModify lock + + if( bt_unlockpage(bt, page_no, BtLockParent) ) + return bt->err; + + return 0; +} + +// find key in leaf level and return row-id + +uid bt_findkey (BtDb *bt, unsigned char *key, uint len) +{ +uint slot; +BtKey ptr; +uid id; + + if( slot = bt_loadpage (bt, key, len, 0, BtLockRead) ) + ptr = keyptr(bt->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; + + return id; +} + +// check page for space available, +// clean if necessary and return +// 0 - page needs splitting +// 1 - go ahead + +uint bt_cleanpage(BtDb *bt, uint amt) +{ +uint nxt = bt->mgr->page_size; +BtPage page = bt->page; +uint cnt = 0, idx = 0; +uint max = page->cnt; +BtKey key; + + if( page->min >= (max+1) * sizeof(BtSlot) + sizeof(*page) + amt + 1 ) + return 1; + + // skip cleanup if nothing to reclaim + + if( !page->dirty ) + return 0; + + memcpy (bt->frame, page, bt->mgr->page_size); + + // skip page info and set rest of page to zero + + memset (page+1, 0, bt->mgr->page_size - sizeof(*page)); + page->dirty = 0; + page->act = 0; + + // try cleaning up page first + + while( cnt++ < max ) { + // always leave fence key and foster children in list + if( cnt < max - page->foster && slotptr(bt->frame,cnt)->dead ) + continue; + + // copy key + 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; + + // see if page has enough space now, or does it need splitting? + + if( page->min >= (idx+1) * sizeof(BtSlot) + sizeof(*page) + amt + 1 ) + return 1; + + return 0; +} + +// add key to page +// return with page unlocked + +BTERR bt_addkeytopage (BtDb *bt, uint slot, unsigned char *key, uint len, uid id, uint tod) +{ +BtPage page = bt->page; +uint idx; + + // calculate next available slot and copy key into page + + page->min -= len + 1; + ((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; + + return bt_unlockpage(bt, bt->page_no, BtLockWrite); +} + +// split the root and raise the height of the btree + +BTERR bt_splitroot(BtDb *bt, uid right) +{ +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) + + root->act--; + root->cnt--; + root->foster--; + + // Save left fence key. + + key = keyptr(root, root->cnt); + memcpy (fencekey, key, key->len + 1); + + // copy the lower keys into a new left page + + 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 + + memset (root+1, 0, bt->mgr->page_size - sizeof(*root)); + + // insert left fence key on empty newroot page + + nxt -= *fencekey + 1; + memcpy ((unsigned char *)root + nxt, fencekey, *fencekey + 1); + bt_putid(slotptr(root, 1)->id, new_page); + slotptr(root, 1)->off = nxt; + + // insert stopper 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); + slotptr(root, 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++; + + // release root (bt->page) + + return bt_unlockpage(bt, bt->page_no, BtLockWrite); +} + +// split already locked full node +// 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; +BtPage page = bt->page; +uint tod = time(NULL); +uint lvl = page->lvl; +uid new_page, right; +BtKey key; + + // initialize frame buffer + + memset (bt->frame, 0, bt->mgr->page_size); + max = page->cnt - page->foster; + tod = (uint)time(NULL); + cnt = max / 2; + idx = 0; + + // split higher half of keys to bt->frame + // leaving foster children in the left node. + + while( cnt++ < max ) { + key = keyptr(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); + slotptr(bt->frame, idx)->tod = slotptr(page, cnt)->tod; + slotptr(bt->frame, idx)->off = nxt; + bt->frame->act++; + } + + // transfer right link node + + if( page_no > ROOT_page ) { + right = bt_getid (page->right); + bt_putid(bt->frame->right, right); + } + + 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 frame to it. + + if( !(new_page = bt_newpage(bt, bt->frame)) ) + return bt->err; + + // remember fence key for new page to add + // as foster child + + key = keyptr(bt->frame, idx); + memcpy (fencekey, key, key->len + 1); + + // update lower keys and foster children 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->act = 0; + cnt = 0; + idx = 0; + + // assemble page of smaller keys + // to remain in the old page + + 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++; + } + + // insert new foster child at beginning 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 if any + + 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); + slotptr(page, idx)->tod = slotptr(bt->frame, cnt)->tod; + slotptr(page, idx)->off = nxt; + page->act++; + } + + 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); + + // release wr lock on page + + if( bt_unlockpage (bt, page_no, BtLockWrite) ) + return bt->err; + + // obtain ParentModification lock for current page + // to fix fence key and highest foster child on page + + if( bt_lockpage (bt, page_no, BtLockParent, NULL) ) + return bt->err; + + // get our highest foster child key to find in parent node + + if( bt_lockpage (bt, page_no, BtLockRead, &page) ) + return bt->err; + + key = keyptr(page, page->cnt); + memcpy (fencekey, key, key->len+1); + + if( bt_unlockpage (bt, page_no, BtLockRead) ) + return bt->err; + +try_again: + + do { + slot = bt_loadpage (bt, fencekey + 1, *fencekey, lvl + 1, BtLockWrite); + + if( !slot ) + return bt->err; + + // check if parent page has enough space for any possible key + + if( bt_cleanpage (bt, 256) ) + break; + + if( bt_splitpage (bt) ) + return bt->err; + } while( 1 ); + + // see if we are still a foster child from another node + + if( bt_getid (slotptr(bt->page, slot)->id) != page_no ) { + bt_unlockpage (bt, bt->page_no, BtLockWrite); +#ifdef unix + sched_yield(); +#else + SwitchToThread(); +#endif + goto try_again; + } + + // wait until readers from parent get their locks + + if( bt_lockpage (bt, page_no, BtLockDelete, NULL) ) + return bt->err; + + if( bt_lockpage (bt, page_no, BtLockWrite, &page) ) + return bt->err; + + // switch parent fence key to foster child + + if( slotptr(page, page->cnt)->dead ) + slotptr(bt->page, slot)->dead = 1; + else + bt_putid (slotptr(bt->page, slot)->id, bt_getid(slotptr(page, page->cnt)->id)); + + // remove highest foster child from our page + // add our new fence key to parent + + page->cnt--; + page->act--; + page->foster--; + page->dirty = 1; + key = keyptr(page, page->cnt); + + if( bt_addkeytopage (bt, slot, key->key, key->len, page_no, tod) ) + return bt->err; + + if( bt_unlockpage (bt, page_no, BtLockDelete) ) + return bt->err; + + if( bt_unlockpage (bt, page_no, BtLockWrite) ) + return bt->err; + + return bt_unlockpage (bt, page_no, BtLockParent); +} + +// Insert new key into the btree at leaf level. + +BTERR bt_insertkey (BtDb *bt, unsigned char *key, uint len, uid id, uint tod) +{ +uint slot, idx; +BtPage page; +BtKey ptr; + + while( 1 ) { + if( slot = bt_loadpage (bt, key, len, 0, 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; + + 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); + } + + // check if page has enough space + + if( bt_cleanpage (bt, len) ) + break; + + if( bt_splitpage (bt) ) + return bt->err; + } + + return bt_addkeytopage (bt, slot, key, len, id, tod); +} + +// cache page of keys into cursor and return starting slot for given key + +uint bt_startkey (BtDb *bt, unsigned char *key, uint len) +{ +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; + + return slot; +} + +// return next slot for cursor page +// or slide cursor right into next page + +uint bt_nextkey (BtDb *bt, uint slot) +{ +BtPage page; +uid right; + + do { + right = bt_getid(bt->cursor->right); + while( slot++ < bt->cursor->cnt - bt->cursor->foster ) + if( slotptr(bt->cursor,slot)->dead ) + continue; + else if( right || (slot < bt->cursor->cnt - bt->cursor->foster) ) + return slot; + else + break; + + if( !right ) + break; + + bt->cursor_page = right; + + if( bt_lockpage(bt, right, BtLockRead, &page) ) + return 0; + + memcpy (bt->cursor, page, bt->mgr->page_size); + + if ( bt_unlockpage(bt, right, BtLockRead) ) + return 0; + + slot = 0; + } while( 1 ); + + return bt->err = 0; +} + +BtKey bt_key(BtDb *bt, uint slot) +{ + return keyptr(bt->cursor, slot); +} + +uid bt_uid(BtDb *bt, uint slot) +{ + return bt_getid(slotptr(bt->cursor,slot)->id); +} + +uint bt_tod(BtDb *bt, uint slot) +{ + return slotptr(bt->cursor,slot)->tod; +} + + +#ifdef STANDALONE + +typedef struct { + char type, idx; + char *infile; + BtMgr *mgr; + int num; +} ThreadArg; + +// standalone program to index file of keys +// then list them onto std-out + +#ifdef unix +void *index_file (void *arg) +#else +uint __stdcall index_file (void *arg) +#endif +{ +int line = 0, found = 0, cnt = 0; +uid next, page_no = LEAF_page; // start on first page of leaves +unsigned char key[256]; +ThreadArg *args = arg; +int ch, len = 0, slot; +time_t tod[1]; +BtPage page; +BtKey ptr; +BtDb *bt; +FILE *in; + + bt = bt_open (args->mgr); + time (tod); + + switch(args->type | 0x20) + { + case 'w': + fprintf(stderr, "started indexing for %s\n", args->infile); + if( in = fopen (args->infile, "rb") ) + while( ch = getc(in), ch != EOF ) + if( ch == '\n' ) + { + 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; + + if( bt_insertkey (bt, key, len, line, *tod) ) + fprintf(stderr, "Error %d Line: %d\n", bt->err, line), exit(0); + len = 0; + } + else if( len < 255 ) + key[len++] = ch; + fprintf(stderr, "finished %s for %d keys\n", args->infile, line); + break; + + case 'd': + fprintf(stderr, "started deleting keys for %s\n", args->infile); + if( in = fopen (args->infile, "rb") ) + while( ch = getc(in), ch != EOF ) + if( ch == '\n' ) + { + 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; + + if( bt_deletekey (bt, key, len, 0) ) + fprintf(stderr, "Error %d Line: %d\n", bt->err, line), exit(0); + len = 0; + } + else if( len < 255 ) + key[len++] = ch; + fprintf(stderr, "finished %s for keys, %d \n", args->infile, line); + break; + + case 'f': + fprintf(stderr, "started finding keys for %s\n", args->infile); + if( in = fopen (args->infile, "rb") ) + while( ch = getc(in), ch != EOF ) + if( ch == '\n' ) + { + 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; + + if( bt_findkey (bt, key, len) ) + found++; + else if( bt->err ) + fprintf(stderr, "Error %d Syserr %d Line: %d\n", bt->err, errno, line), exit(0); + len = 0; + } + else if( len < 255 ) + key[len++] = ch; + fprintf(stderr, "finished %s for %d keys, found %d\n", args->infile, line, found); + 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); + } + + break; + + case 'c': + fprintf(stderr, "started reading\n"); + + 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 ); + + cnt--; // remove stopper key + fprintf(stderr, " Total keys read %d\n", cnt); + break; + } + + bt_close (bt); +#ifdef unix + return NULL; +#else + return 0; +#endif +} + +typedef struct timeval timer; + +int main (int argc, char **argv) +{ +int idx, cnt, len, slot, err; +int segsize, bits = 16; +#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; +int num = 0; +char key[1]; +BtMgr *mgr; +BtKey ptr; +BtDb *bt; + + if( argc < 3 ) { + fprintf (stderr, "Usage: %s idx_file Read/Write/Scan/Delete/Find [page_bits mapped_segments seg_bits line_numbers src_file1 src_file2 ... ]\n", argv[0]); + 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 = 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 + + if( argc > 3 ) + bits = atoi(argv[3]); + + if( argc > 4 ) + poolsize = atoi(argv[4]); + + if( !poolsize ) + fprintf (stderr, "Warning: no mapped_pool\n"); + + if( poolsize > 65535 ) + fprintf (stderr, "Warning: mapped_pool > 65535 segments\n"); + + if( argc > 5 ) + segsize = atoi(argv[5]); + else + segsize = 4; // 16 pages per mmap segment + + if( argc > 6 ) + num = atoi(argv[6]); + + cnt = argc - 7; +#ifdef unix + threads = malloc (cnt * sizeof(pthread_t)); +#else + threads = GlobalAlloc (GMEM_FIXED|GMEM_ZEROINIT, cnt * sizeof(HANDLE)); +#endif + args = malloc (cnt * sizeof(ThreadArg)); + + mgr = bt_mgr ((argv[1]), BT_rw, bits, poolsize, segsize, poolsize / 8); + + if( !mgr ) { + fprintf(stderr, "Index Open Error %s\n", argv[1]); + exit (1); + } + + // fire off threads + + for( idx = 0; idx < cnt; idx++ ) { + args[idx].infile = argv[idx + 7]; + args[idx].type = argv[2][0]; + args[idx].mgr = mgr; + args[idx].num = num; + args[idx].idx = idx; +#ifdef unix + if( err = pthread_create (threads + idx, NULL, index_file, args + idx) ) + fprintf(stderr, "Error creating thread %d\n", err); +#else + threads[idx] = (HANDLE)_beginthreadex(NULL, 65536, index_file, args + idx, 0, NULL); +#endif + } + + // wait for termination + +#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); + bt_mgrclose (mgr); +} + +#endif //STANDALONE diff --git a/fosterbtreee.c b/fosterbtreee.c index 8d18fed..86f836d 100644 --- a/fosterbtreee.c +++ b/fosterbtreee.c @@ -1,5 +1,5 @@ -// foster btree version e -// 30 DEC 2013 +// foster btree version e2 +// 18 JAN 2014 // author: karl malbrain, malbrain@cal.berkeley.edu @@ -97,7 +97,7 @@ typedef enum{ // the tod field from the key. // Keys are marked dead, but remain on the page until -// it cleanup is called. The fence key (highest key) for +// cleanup is called. The fence key (highest key) for // the page is always present, even after cleanup. typedef struct { @@ -121,13 +121,12 @@ typedef struct { // by the BtSlot array of keys. typedef struct Page { - uint cnt; // count of keys in page - uint act; // count of active keys - uint min; // next key offset - uint foster; // count of foster children + 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 unsigned char bits; // page size in bits - unsigned char lvl:6; // level of page - unsigned char kill:1; // page is being deleted + unsigned char lvl:7; // level of page unsigned char dirty:1; // page needs to be cleaned unsigned char right[BtId]; // page number to right } *BtPage; @@ -159,8 +158,7 @@ typedef struct { volatile ushort slot; // Latch table entry at head of chain } BtHashEntry; -// latch table lock structure -// implements a fair read-write lock +// latch manager table structure typedef struct { #ifdef unix @@ -225,14 +223,14 @@ typedef struct { #endif ushort poolcnt; // highest page pool node in use ushort poolmax; // highest page pool node allocated - ushort poolmask; // total size of pages in mmap segment - 1 + 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 BtPool *pool; // memory pool page segments BtSpinLatch *latch; // latches for pool hash slots BtLatchMgr *latchmgr; // mapped latch page from allocation page - BtLatchSet *latchset; // first mapped latch set from latch pages + BtLatchSet *latchsets; // mapped latch set from latch pages #ifndef unix HANDLE halloc; // allocation and latch table handle #endif @@ -240,14 +238,16 @@ typedef struct { typedef struct { BtMgr *mgr; // buffer manager for thread - BtPage temp; // temporary frame buffer (memory mapped/file IO) 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 uid page_no; // current page number uid cursor_page; // current cursor page number + BtLatchSet *set; // current page latch set + BtPool *pool; // current page pool unsigned char *mem; // frame, cursor, page memory buffer + int found; // last delete was found int err; // last error } BtDb; @@ -265,8 +265,8 @@ 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); -extern BTERR bt_deletekey (BtDb *bt, unsigned char *key, uint len, uint lvl); +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 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); @@ -397,30 +397,30 @@ ushort prev; void bt_spinwritelock(BtSpinLatch *latch) { -ushort prev; - do { -#ifdef unix +#ifdef unix while( __sync_fetch_and_or((ushort *)latch, Mutex | Pending) & Mutex ) sched_yield(); #else while( _InterlockedOr16((ushort *)latch, Mutex | Pending) & Mutex ) SwitchToThread(); #endif - if( prev = !(latch->share | latch->exclusive) ) + 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; + } #ifdef unix - __sync_fetch_and_and ((ushort *)latch, ~(Mutex | Pending)); + __sync_fetch_and_and ((ushort *)latch, ~Mutex); #else - _InterlockedAnd16((ushort *)latch, ~(Mutex | Pending)); + _InterlockedAnd16((ushort *)latch, ~Mutex); #endif - if( prev ) - return; #ifdef unix sched_yield(); #else @@ -484,11 +484,17 @@ void bt_spinreleaseread(BtSpinLatch *latch) #endif } -void bt_initlockset (BtLatchSet *set) +void bt_initlockset (BtLatchSet *set, int reuse) { #ifdef unix pthread_rwlockattr_t rwattr[1]; + if( reuse ) { + pthread_rwlock_destroy (set->readwr->lock); + pthread_rwlock_destroy (set->access->lock); + pthread_rwlock_destroy (set->parent->lock); + } + pthread_rwlockattr_init (rwattr); pthread_rwlockattr_setkind_np (rwattr, PTHREAD_RWLOCK_PREFER_WRITER_NONRECURSIVE_NP); pthread_rwlockattr_setpshared (rwattr, PTHREAD_PROCESS_SHARED); @@ -508,10 +514,10 @@ pthread_rwlockattr_t rwattr[1]; void bt_latchlink (BtDb *bt, ushort hashidx, ushort victim, uid page_no) { -BtLatchSet *set = bt->mgr->latchset + victim; +BtLatchSet *set = bt->mgr->latchsets + victim; if( set->next = bt->mgr->latchmgr->table[hashidx].slot ) - bt->mgr->latchset[set->next].prev = victim; + bt->mgr->latchsets[set->next].prev = victim; bt->mgr->latchmgr->table[hashidx].slot = victim; set->page_no = page_no; @@ -519,10 +525,19 @@ BtLatchSet *set = bt->mgr->latchset + victim; set->prev = 0; } +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_bindlatch (BtDb *bt, uid page_no, int incr) +BtLatchSet *bt_pinlatch (BtDb *bt, uid page_no) { ushort hashidx = page_no % bt->mgr->latchmgr->latchhash; ushort slot, avail = 0, victim, idx; @@ -534,12 +549,12 @@ BtLatchSet *set; if( slot = bt->mgr->latchmgr->table[hashidx].slot ) do { - set = bt->mgr->latchset + slot; + set = bt->mgr->latchsets + slot; if( page_no == set->page_no ) break; } while( slot = set->next ); - if( slot && incr ) { + if( slot ) { #ifdef unix __sync_fetch_and_add(&set->pin, 1); #else @@ -558,7 +573,7 @@ BtLatchSet *set; if( slot = bt->mgr->latchmgr->table[hashidx].slot ) do { - set = bt->mgr->latchset + slot; + set = bt->mgr->latchsets + slot; if( page_no == set->page_no ) break; if( !set->pin && !avail ) @@ -568,12 +583,11 @@ BtLatchSet *set; // found our entry, or take over an unpinned one if( slot || (slot = avail) ) { - set = bt->mgr->latchset + slot; - if( incr ) + set = bt->mgr->latchsets + slot; #ifdef unix - __sync_fetch_and_add(&set->pin, 1); + __sync_fetch_and_add(&set->pin, 1); #else - _InterlockedIncrement16 (&set->pin); + _InterlockedIncrement16 (&set->pin); #endif set->page_no = page_no; bt_spinreleasewrite(bt->mgr->latchmgr->table[hashidx].latch); @@ -588,14 +602,13 @@ BtLatchSet *set; #endif if( victim < bt->mgr->latchmgr->latchtotal ) { - set = bt->mgr->latchset + victim; - if( incr ) + set = bt->mgr->latchsets + victim; #ifdef unix - __sync_fetch_and_add(&set->pin, 1); + __sync_fetch_and_add(&set->pin, 1); #else - _InterlockedIncrement16 (&set->pin); + _InterlockedIncrement16 (&set->pin); #endif - bt_initlockset (set); + bt_initlockset (set, 0); bt_latchlink (bt, hashidx, victim, page_no); bt_spinreleasewrite (bt->mgr->latchmgr->table[hashidx].latch); return set; @@ -617,7 +630,7 @@ BtLatchSet *set; // we don't use slot zero if( victim %= bt->mgr->latchmgr->latchtotal ) - set = bt->mgr->latchset + victim; + set = bt->mgr->latchsets + victim; else continue; @@ -647,22 +660,20 @@ BtLatchSet *set; // unlink our available victim from its hash chain if( set->prev ) - bt->mgr->latchset[set->prev].next = set->next; + bt->mgr->latchsets[set->prev].next = set->next; else bt->mgr->latchmgr->table[idx].slot = set->next; if( set->next ) - bt->mgr->latchset[set->next].prev = set->prev; + bt->mgr->latchsets[set->next].prev = set->prev; bt_spinreleasewrite (bt->mgr->latchmgr->table[idx].latch); - - if( incr ) #ifdef unix - __sync_fetch_and_add(&set->pin, 1); + __sync_fetch_and_add(&set->pin, 1); #else - _InterlockedIncrement16 (&set->pin); + _InterlockedIncrement16 (&set->pin); #endif - + bt_initlockset (set, 1); bt_latchlink (bt, hashidx, victim, page_no); bt_spinreleasewrite (bt->mgr->latchmgr->table[hashidx].latch); bt_spinreleasewrite (set->busy); @@ -692,6 +703,14 @@ 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); @@ -733,7 +752,6 @@ BtMgr *bt_mgr (char *name, uint mode, uint bits, uint poolmax, uint segsize, uin uint lvl, attr, cacheblk, last, slot, idx; uint nlatchpage, latchhash; BtLatchMgr *latchmgr; -int lockmode; off64_t size; uint amt[1]; BtMgr* mgr; @@ -757,19 +775,8 @@ 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; @@ -778,19 +785,8 @@ 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; @@ -811,7 +807,7 @@ SYSTEM_INFO sysinfo[1]; else return free(mgr), free(latchmgr), NULL; } else if( mode == BT_ro ) - return 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); @@ -887,7 +883,7 @@ SYSTEM_INFO sysinfo[1]; #ifdef unix if( write (mgr->idx, latchmgr, mgr->page_size) < mgr->page_size ) - return bt_mgrclose (mgr), NULL; + return free(latchmgr), bt_mgrclose (mgr), NULL; #else if( !WriteFile (mgr->idx, (char *)latchmgr, mgr->page_size, amt, NULL) ) return bt_mgrclose (mgr), NULL; @@ -943,25 +939,25 @@ SYSTEM_INFO sysinfo[1]; mgrlatch: #ifdef unix - flag = PROT_READ | ( mgr->mode == BT_ro ? 0 : PROT_WRITE ); + 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->latchset = (BtLatchSet *)mmap (0, mgr->latchmgr->nlatchpage * mgr->page_size, flag, MAP_SHARED, mgr->idx, LATCH_page * mgr->page_size); - if( mgr->latchset == MAP_FAILED ) + 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 = ( mgr->mode == BT_ro ? PAGE_READONLY : PAGE_READWRITE ); + 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 = ( mgr->mode == BT_ro ? FILE_MAP_READ : FILE_MAP_WRITE ); + 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->latchset = (void *)((char *)mgr->latchmgr + LATCH_page * mgr->page_size); + mgr->latchsets = (void *)((char *)mgr->latchmgr + LATCH_page * mgr->page_size); #endif #ifdef unix @@ -989,6 +985,8 @@ BtDb *bt = malloc (sizeof(*bt)); bt->frame = (BtPage)bt->mem; 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); return bt; } @@ -1177,10 +1175,43 @@ int flag; return bt->err = 0; } +// calculate page within pool + +BtPage bt_page (BtDb *bt, BtPool *pool, uid page_no) +{ +uint subpage = (uint)(page_no & bt->mgr->poolmask); // page within mapping +BtPage page; + + 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 + return page; +} + +// release pool pin + +void bt_unpinpool (BtPool *pool) +{ +#ifdef unix + __sync_fetch_and_add(&pool->pin, -1); +#else + _InterlockedDecrement16 (&pool->pin); +#endif +} + // find or place requested page in segment-pool // return pool table entry, incrementing pin -BtPool *bt_pinpage(BtDb *bt, uid page_no) +BtPool *bt_pinpool(BtDb *bt, uid page_no) { BtPool *pool, *node, *next; uint slot, idx, victim; @@ -1322,41 +1353,10 @@ BtLatchSet *set; } // place write, read, or parent lock on requested page_no. -// pin to buffer pool and return page pointer +// pin to buffer pool and return latchset pointer -BTERR bt_lockpage(BtDb *bt, uid page_no, BtLock mode, BtPage *pageptr) +void bt_lockpage(BtLock mode, BtLatchSet *set) { -BtLatchSet *set; -BtPool *pool; -uint subpage; -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; - - if( !(set = bt_bindlatch (bt, page_no, 1)) ) - return bt->err; - - page = (BtPage)(pool->map + (subpage << bt->mgr->page_bits)); - -#ifdef unix - { - uint idx = subpage / 8; - uint bit = subpage % 8; - - if( mode == BtLockRead || mode == BtLockWrite ) - 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 (set->readwr); @@ -1373,39 +1373,13 @@ BtPage page; case BtLockParent: bt_writelock (set->parent); break; - default: - return bt->err = BTERR_lock; } - - if( pageptr ) - *pageptr = page; - - return bt->err = 0; } // remove write, read, or parent lock on requested page_no. -BTERR bt_unlockpage(BtDb *bt, uid page_no, BtLock mode) +void bt_unlockpage(BtLock mode, BtLatchSet *set) { -BtLatchSet *set; -BtPool *pool; -uint idx; - - // since page is pinned - // it should still be in the buffer pool - // and is in no danger of being a victim for reuse - - if( !(set = bt_bindlatch (bt, page_no, 0)) ) - return bt->err = BTERR_latch; - - idx = (uint)(page_no >> bt->mgr->seg_bits) % bt->mgr->hashsize; - bt_spinreadlock (&bt->mgr->latch[idx]); - - if( !(pool = bt_findpool(bt, page_no, idx)) ) - return bt->err = BTERR_hash; - - bt_spinreleaseread (&bt->mgr->latch[idx]); - switch( mode ) { case BtLockRead: bt_releaseread (set->readwr); @@ -1422,65 +1396,15 @@ uint idx; case BtLockParent: bt_releasewrite (set->parent); break; - default: - return bt->err = BTERR_lock; } - -#ifdef unix - __sync_fetch_and_add(&pool->pin, -1); - __sync_fetch_and_add (&set->pin, -1); -#else - _InterlockedDecrement16 (&pool->pin); - _InterlockedDecrement16 (&set->pin); -#endif - return bt->err = 0; -} - -// deallocate a deleted page -// place on free chain out of allocator page -// fence key must already be removed from parent - -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 - - bt_spinwritelock(bt->mgr->latchmgr->lock); - - // store free chain in allocation page second right - bt_putid(bt->temp->right, bt_getid(bt->mgr->latchmgr->alloc[1].right)); - bt_putid(bt->mgr->latchmgr->alloc[1].right, page_no); - - // unlock page zero - - bt_spinreleasewrite(bt->mgr->latchmgr->lock); - - // remove write lock on deleted node - - if( bt_unlockpage(bt, page_no, BtLockWrite) ) - return bt->err; - - // remove delete lock on deleted node - - if( bt_unlockpage(bt, page_no, BtLockDelete) ) - return bt->err; - - return 0; } // allocate a new page and write page into it uid bt_newpage(BtDb *bt, BtPage page) { +BtLatchSet *set; +BtPool *pool; uid new_page; BtPage pmap; int reuse; @@ -1493,11 +1417,12 @@ int reuse; // else allocate empty page if( new_page = bt_getid(bt->mgr->latchmgr->alloc[1].right) ) { - if( bt_lockpage (bt, new_page, BtLockWrite, &bt->temp) ) - return 0; - bt_putid(bt->mgr->latchmgr->alloc[1].right, bt_getid(bt->temp->right)); - if( bt_unlockpage (bt, new_page, BtLockWrite) ) + if( pool = bt_pinpool (bt, new_page) ) + pmap = bt_page (bt, pool, new_page); + else return 0; + bt_putid(bt->mgr->latchmgr->alloc[1].right, bt_getid(pmap->right)); + bt_unpinpool (pool); reuse = 1; } else { new_page = bt_getid(bt->mgr->latchmgr->alloc->right); @@ -1505,33 +1430,39 @@ 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( bt_lockpage(bt, new_page, BtLockWrite, &pmap) ) + if( pool = bt_pinpool (bt, new_page) ) + pmap = bt_page (bt, pool, new_page); + else return 0; memcpy(pmap, page, bt->mgr->page_size); - - if( bt_unlockpage (bt, new_page, BtLockWrite) ) - return 0; + bt_unpinpool (pool); #endif - // unlock allocation latch and return new page no - - bt_spinreleasewrite(bt->mgr->latchmgr->lock); return new_page; } @@ -1558,11 +1489,13 @@ uint diff, higher = bt->page->cnt, low = 1, slot; // 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, unsigned char *key, uint len, uint lvl, BtLock lock) { uid page_no = ROOT_page, prevpage = 0; +BtLatchSet *set, *prevset; uint drill = 0xff, slot; uint mode, prevmode; +BtPool *prevpool; // start at root of btree and drill down @@ -1570,84 +1503,94 @@ uint mode, prevmode; // 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; + // pin page contents + + if( bt->pool = bt_pinpool (bt, page_no) ) + bt->page = bt_page (bt, bt->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, bt->set); - // now unlock our (possibly foster) parent + // now unlock and unpin our (possibly foster) parent - if( prevpage ) - if( bt_unlockpage(bt, prevpage, prevmode) ) - return 0; - else - prevpage = 0; + if( prevpage ) { + bt_unlockpage(prevmode, prevset); + bt_unpinlatch (prevset); + 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, bt->set); if( page_no > ROOT_page ) - if( bt_unlockpage(bt, page_no, BtLockAccess) ) - return 0; + bt_unlockpage(BtLockAccess, bt->set); // re-read and re-lock root after determining actual level of root - if( bt->page_no == ROOT_page ) + if( page_no == ROOT_page ) if( bt->page->lvl != drill) { drill = bt->page->lvl; - if( lock == BtLockWrite && drill == lvl ) - if( bt_unlockpage(bt, page_no, mode) ) - return 0; - else - continue; + 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; - // if page is being deleted, - // move back to preceeding page - - if( bt->page->kill ) { - page_no = bt_getid (bt->page->right); - continue; - } - // find key on page at this level - // and descend to requested level + // and either descend to requested level + // or return key slot slot = bt_findslot (bt, key, len); - // is this slot a foster child? + // 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( 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 - goto slideright; + return bt->err = BTERR_struct, 0; // last key shouldn't be deleted + + // is this slot < foster child area + // if so, drill to next level if( slot <= bt->page->cnt - bt->page->foster ) drill--; - // continue down / right using overlapping locks - // to protect pages being killed or split. + // continue right onto foster child + // or down to next level. page_no = bt_getid(slotptr(bt->page, slot)->id); - continue; - -slideright: - page_no = bt_getid(bt->page->right); } while( page_no ); @@ -1658,108 +1601,182 @@ slideright: } // find and delete key on page by marking delete flag bit -// when page becomes empty, delete it from the btree +// when leaf page becomes empty, delete it from the btree -BTERR bt_deletekey (BtDb *bt, unsigned char *key, uint len, uint lvl) +BTERR bt_deletekey (BtDb *bt, unsigned char *key, uint len) { -unsigned char leftkey[256], rightkey[256]; +unsigned char leftkey[256]; +BtLatchSet *rset, *set; +BtPool *pool, *rpool; +BtPage rpage, page; uid page_no, right; uint slot, tod; BtKey ptr; - if( slot = bt_loadpage (bt, key, len, lvl, BtLockWrite) ) + if( slot = bt_loadpage (bt, key, len, 0, 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( !keycmp (ptr, key, len) ) - if( slotptr(bt->page, slot)->dead == 0 ) { + if( bt->found = !keycmp (ptr, key, len) ) + if( bt->found = slotptr(bt->page, slot)->dead == 0 ) { slotptr(bt->page,slot)->dead = 1; if( slot < bt->page->cnt ) 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; - if( !right || bt->page->act ) - return bt_unlockpage(bt, page_no, BtLockWrite); - - // obtain Parent lock over write lock + // return if page is not empty or not found - if( bt_lockpage(bt, page_no, BtLockParent, NULL) ) + if( page->act || !bt->found ) { + bt_unlockpage(BtLockWrite, set); + bt_unpinlatch (set); + bt_unpinpool (pool); return bt->err; + } - // cache copy of key to delete + // cache copy of fence key of empty node - ptr = keyptr(bt->page, bt->page->cnt); + ptr = keyptr(page, page->cnt); memcpy(leftkey, ptr, ptr->len + 1); - // lock and map right page + // release write lock on empty node + // obtain Parent lock - if( bt_lockpage(bt, right, BtLockWrite, &bt->temp) ) - return bt->err; + bt_unlockpage(BtLockWrite, set); + bt_lockpage(BtLockParent, set); - // pull contents of next page into current empty page - memcpy (bt->page, bt->temp, bt->mgr->page_size); + // load and lock parent to see + // if delete of empty node is OK + // ie, not a fence key of parent - // cache copy of key to update - ptr = keyptr(bt->temp, bt->temp->cnt); - memcpy(rightkey, ptr, ptr->len + 1); + while( 1 ) { + if( slot = bt_loadpage (bt, leftkey+1, *leftkey, 1, BtLockWrite) ) + ptr = keyptr(bt->page, slot); + else + return bt->err; - // Mark right page as deleted and point it to left page - // until we can post updates at higher level. + // does parent level contain our fence key yet? + // and is it free of foster children? - bt_putid(bt->temp->right, page_no); - bt->temp->kill = 1; - bt->temp->cnt = 0; + if( !bt->page->foster ) + if( !keycmp (ptr, leftkey+1, *leftkey) ) + break; - if( bt_unlockpage(bt, right, BtLockWrite) ) - return bt->err; - if( bt_unlockpage(bt, page_no, BtLockWrite) ) - return bt->err; + bt_unlockpage(BtLockWrite, bt->set); + bt_unpinlatch (bt->set); + bt_unpinpool (bt->pool); +#ifdef unix + sched_yield(); +#else + SwitchToThread(); +#endif + } - // delete old lower key to consolidated node + // find our left fence key - if( bt_deletekey (bt, leftkey + 1, *leftkey, lvl + 1) ) + 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); + + // 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; + } - // redirect higher key directly to consolidated node + // lock and map our right page + // note that it cannot be our foster child + // since the our node is empty - if( slot = bt_loadpage (bt, rightkey+1, *rightkey, lvl+1, BtLockWrite) ) - ptr = keyptr(bt->page, slot); + right = bt_getid(page->right); + + if( rpool = bt_pinpool (bt, right) ) + rpage = bt_page (bt, rpool, right); else return bt->err; - // since key already exists, update id + 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); - if( keycmp (ptr, rightkey+1, *rightkey) ) - return bt->err = BTERR_struct; + // delete left parent slot for old empty page + // and redirect right parent slot to it + + bt->page->act--; + bt->page->dirty = 1; + slotptr(bt->page, slot)->dead = 1; + + while( slot++ < bt->page->cnt ) + if( !slotptr(bt->page, slot)->dead ) + break; - slotptr(bt->page, slot)->dead = 0; bt_putid(slotptr(bt->page,slot)->id, page_no); - if( bt_unlockpage(bt, bt->page_no, BtLockWrite) ) - return bt->err; + // release parent level lock + // and our empty node lock - // obtain write lock and - // add right block to free chain + bt_unlockpage(BtLockWrite, set); + bt_unlockpage(BtLockWrite, bt->set); + bt_unpinlatch (bt->set); + bt_unpinpool (bt->pool); - if( bt_freepage (bt, right) ) - return bt->err; + // add killed right block to free chain + // lock latch mgr + + bt_spinwritelock(bt->mgr->latchmgr->lock); + + // store free chain in allocation page second right + bt_putid(rpage->right, bt_getid(bt->mgr->latchmgr->alloc[1].right)); + bt_putid(bt->mgr->latchmgr->alloc[1].right, right); + + // unlock latch mgr and right page + + bt_spinreleasewrite(bt->mgr->latchmgr->lock); + + bt_unlockpage(BtLockWrite, rset); + bt_unlockpage(BtLockDelete, rset); + bt_unpinlatch (rset); + bt_unpinpool (rpool); // remove ParentModify lock - if( bt_unlockpage(bt, page_no, BtLockParent) ) - return bt->err; - + 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 @@ -1777,32 +1794,33 @@ uid id; // if key exists, return row-id // otherwise return 0 - if( ptr->len == len && !memcmp (ptr->key, key, len) ) + if( slot <= bt->page->cnt && !keycmp (ptr, key, len) ) id = bt_getid(slotptr(bt->page,slot)->id); else id = 0; - if( bt_unlockpage (bt, bt->page_no, BtLockRead) ) - return 0; - + bt_unlockpage (BtLockRead, bt->set); + bt_unpinlatch (bt->set); + bt_unpinpool (bt->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, 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; BtKey key; if( page->min >= (max+1) * sizeof(BtSlot) + sizeof(*page) + amt + 1 ) - return 1; + return slot; // skip cleanup if nothing to reclaim @@ -1819,12 +1837,20 @@ BtKey key; // 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 + while( cnt++ < max ) { - // always leave fence key and foster children in list - if( cnt < max - page->foster && slotptr(bt->frame,cnt)->dead ) + if( cnt == slot ) + newslot = idx + 1; + else if( cnt < max && slotptr(bt->frame,cnt)->dead ) continue; // copy key + key = keyptr(bt->frame, cnt); nxt -= key->len + 1; memcpy ((unsigned char *)page + nxt, key, key->len + 1); @@ -1843,7 +1869,7 @@ BtKey key; // see if page has enough space now, or does it need splitting? if( page->min >= (idx+1) * sizeof(BtSlot) + sizeof(*page) + amt + 1 ) - return 1; + return newslot; return 0; } @@ -1856,27 +1882,29 @@ void bt_addkeytopage (BtDb *bt, uint slot, unsigned char *key, uint len, uid id, BtPage page = bt->page; uint idx; - // calculate next available slot and copy key into page + // find next available dead slot and copy key onto page + // note that foster children on the page are never dead - page->min -= len + 1; - ((unsigned char *)page)[page->min] = len; - memcpy ((unsigned char *)page + page->min +1, key, len ); + // look for next hole, but stay back from the fence key 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++; + // 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; @@ -1899,9 +1927,11 @@ BtKey key; // 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--; - root->foster--; // Save left fence key. @@ -1942,27 +1972,32 @@ BtKey key; root->act = 2; root->lvl++; - // release root (bt->page) + // release and unpin root (bt->page) - return bt_unlockpage(bt, ROOT_page, BtLockWrite); + bt_unlockpage(BtLockWrite, bt->set); + bt_unpinlatch (bt->set); + bt_unpinpool (bt->pool); + return 0; } // split already locked full node // in current page variables -// return unlocked. +// return unlocked and unpinned. 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; +BtPool *pool = bt->pool; BtPage page = bt->page; uint tod = time(NULL); uint lvl = page->lvl; uid new_page, right; BtKey key; - // initialize frame buffer + // initialize frame buffer for right node memset (bt->frame, 0, bt->mgr->page_size); max = page->cnt - page->foster; @@ -1971,19 +2006,21 @@ BtKey key; idx = 0; // split higher half of keys to bt->frame - // leaving foster children in the left node. + // 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; 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) ) + bt->frame->act++; slotptr(bt->frame, idx)->tod = slotptr(page, cnt)->tod; slotptr(bt->frame, idx)->off = nxt; - bt->frame->act++; } - // transfer right link node + // transfer right link node to new right node if( page_no > ROOT_page ) { right = bt_getid (page->right); @@ -1995,13 +2032,13 @@ BtKey key; bt->frame->cnt = idx; bt->frame->lvl = lvl; - // get new free page and write frame to it. + // get new free page and write right frame to it. if( !(new_page = bt_newpage(bt, bt->frame)) ) return bt->err; - // remember fence key for new page to add - // as foster child + // remember fence key for new right page to add + // as foster child to the left node key = keyptr(bt->frame, idx); memcpy (fencekey, key, key->len + 1); @@ -2011,6 +2048,7 @@ BtKey key; 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; @@ -2023,22 +2061,26 @@ BtKey key; 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; - page->act++; } - // insert new foster child at beginning of the current foster children + // 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 if any + // continue with old foster child keys + // note that none will be dead cnt = bt->frame->cnt - bt->frame->foster; @@ -2066,109 +2108,69 @@ BtKey key; // release wr lock on our page - if( bt_unlockpage (bt, page_no, BtLockWrite) ) - return bt->err; + bt_unlockpage (BtLockWrite, set); // obtain ParentModification lock for current page - // to fix fence key and highest foster child on page + // to fix new fence key and oldest foster child on page - if( bt_lockpage (bt, page_no, BtLockParent, NULL) ) - return bt->err; + bt_lockpage (BtLockParent, set); - // get our highest foster child key to find in parent node + // get our new fence key to insert in parent node - if( bt_lockpage (bt, page_no, BtLockRead, &page) ) - return bt->err; + bt_lockpage (BtLockRead, set); - key = keyptr(page, page->cnt); + key = keyptr(page, page->cnt-1); memcpy (fencekey, key, key->len+1); - if( bt_unlockpage (bt, page_no, BtLockRead) ) - return bt->err; - - // update our parent -try_again: - - do { - slot = bt_loadpage (bt, fencekey + 1, *fencekey, lvl + 1, BtLockWrite); + bt_unlockpage (BtLockRead, set); - if( !slot ) + if( bt_insertkey (bt, fencekey + 1, *fencekey, page_no, tod, lvl + 1) ) return bt->err; - // check if parent page has enough space for any possible key + // lock our page for writing - if( bt_cleanpage (bt, 256) ) - break; + bt_lockpage (BtLockRead, set); - if( bt_splitpage (bt) ) - return bt->err; - } while( 1 ); + // switch old parent key from us to our oldest foster child - // see if we are still a foster child from another node - - if( bt_getid (slotptr(bt->page, slot)->id) != page_no ) { - if( bt_unlockpage (bt, bt->page_no, BtLockWrite) ) - return bt->err; -#ifdef unix - sched_yield(); -#else - SwitchToThread(); -#endif - goto try_again; - } - - // wait until readers from parent get their locks - // on our page - - if( bt_lockpage (bt, page_no, BtLockDelete, NULL) ) - return bt->err; + key = keyptr(page, page->cnt); + memcpy (fencekey, key, key->len+1); - // lock our page for writing + new_page = bt_getid (slotptr(page, page->cnt)->id); + bt_unlockpage (BtLockRead, set); - if( bt_lockpage (bt, page_no, BtLockWrite, &page) ) + if( bt_insertkey (bt, fencekey + 1, *fencekey, new_page, tod, lvl + 1) ) return bt->err; - // switch parent fence key to foster child - - if( slotptr(page, page->cnt)->dead ) - slotptr(bt->page, slot)->dead = 1; - else - bt_putid (slotptr(bt->page, slot)->id, bt_getid(slotptr(page, page->cnt)->id)); - - // remove highest foster child from our page + // 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--; - page->foster--; - page->dirty = 1; - key = keyptr(page, page->cnt); - - // add our new fence key for foster child to our parent - bt_addkeytopage (bt, slot, key->key, key->len, page_no, tod); - - if( bt_unlockpage (bt, bt->page_no, BtLockWrite) ) - return bt->err; + // unlock and unpin - if( bt_unlockpage (bt, page_no, BtLockDelete) ) - return bt->err; - - if( bt_unlockpage (bt, page_no, BtLockWrite) ) - return bt->err; - - return bt_unlockpage (bt, page_no, BtLockParent); + bt_unlockpage (BtLockWrite, set); + bt_unlockpage (BtLockParent, set); + bt_unpinlatch (set); + bt_unpinpool (pool); + return 0; } // Insert new key into the btree at leaf level. -BTERR bt_insertkey (BtDb *bt, unsigned char *key, uint len, uid id, uint tod) +BTERR bt_insertkey (BtDb *bt, unsigned char *key, uint len, uid id, uint tod, uint lvl) { uint slot, idx; BtPage page; BtKey ptr; while( 1 ) { - if( slot = bt_loadpage (bt, key, len, 0, BtLockWrite) ) + if( slot = bt_loadpage (bt, key, len, lvl, BtLockWrite) ) ptr = keyptr(bt->page, slot); else { @@ -2182,15 +2184,20 @@ BtKey ptr; 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); - return bt_unlockpage(bt, bt->page_no, BtLockWrite); + bt_unlockpage(BtLockWrite, bt->set); + bt_unpinlatch (bt->set); + bt_unpinpool (bt->pool); + return bt->err; } // check if page has enough space - if( bt_cleanpage (bt, len) ) + if( slot = bt_cleanpage (bt, len, slot) ) break; if( bt_splitpage (bt) ) @@ -2199,7 +2206,10 @@ BtKey ptr; bt_addkeytopage (bt, slot, key, len, id, tod); - return bt_unlockpage (bt, bt->page_no, BtLockWrite); + 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 @@ -2211,10 +2221,12 @@ 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; + bt_unlockpage(BtLockRead, bt->set); + bt_unpinlatch (bt->set); + bt_unpinpool (bt->pool); return slot; } @@ -2223,6 +2235,8 @@ uint slot; uint bt_nextkey (BtDb *bt, uint slot) { +BtLatchSet *set; +BtPool *pool; BtPage page; uid right; @@ -2240,15 +2254,19 @@ uid right; break; bt->cursor_page = right; - - if( bt_lockpage(bt, right, BtLockRead, &page) ) + if( pool = bt_pinpool (bt, right) ) + page = bt_page (bt, pool, right); + else return 0; - memcpy (bt->cursor, page, bt->mgr->page_size); + set = bt_pinlatch (bt, right); + bt_lockpage(BtLockRead, set); - if ( bt_unlockpage(bt, right, BtLockRead) ) - return 0; + memcpy (bt->cursor, page, bt->mgr->page_size); + bt_unlockpage(BtLockRead, set); + bt_unpinlatch (set); + bt_unpinpool (pool); slot = 0; } while( 1 ); @@ -2294,7 +2312,9 @@ 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; BtKey ptr; BtDb *bt; @@ -2319,7 +2339,7 @@ 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) ) + if( bt_insertkey (bt, key, len, line, *tod, 0) ) fprintf(stderr, "Error %d Line: %d\n", bt->err, line), exit(0); len = 0; } @@ -2341,7 +2361,7 @@ FILE *in; else if( args->num ) sprintf((char *)key+len, "%.9d", line + args->idx * args->num), len += 9; - if( bt_deletekey (bt, key, len, 0) ) + if( bt_deletekey (bt, key, len) ) fprintf(stderr, "Error %d Line: %d\n", bt->err, line), exit(0); len = 0; } @@ -2396,10 +2416,17 @@ FILE *in; fprintf(stderr, "started reading\n"); do { - bt_lockpage (bt, page_no, BtLockRead, &page); + if( pool = bt_pinpool (bt, page_no) ) + page = bt_page (bt, pool, page_no); + else + break; + set = bt_pinlatch (bt, page_no); + bt_lockpage (BtLockRead, set); cnt += page->act; next = bt_getid (page->right); - bt_unlockpage (bt, page_no, BtLockRead); + bt_unlockpage (BtLockRead, set); + bt_unpinlatch (set); + bt_unpinpool (pool); } while( page_no = next ); cnt--; // remove stopper key diff --git a/fosterbtreee1.c b/fosterbtreee1.c new file mode 100644 index 0000000..2b45ce8 --- /dev/null +++ b/fosterbtreee1.c @@ -0,0 +1,2529 @@ +// foster btree version e +// 17 JAN 2014 + +// author: karl malbrain, malbrain@cal.berkeley.edu + +/* +This work, including the source code, documentation +and related data, is placed into the public domain. + +The orginal author is Karl Malbrain. + +THIS SOFTWARE IS PROVIDED AS-IS WITHOUT WARRANTY +OF ANY KIND, NOT EVEN THE IMPLIED WARRANTY OF +MERCHANTABILITY. THE AUTHOR OF THIS SOFTWARE, +ASSUMES _NO_ RESPONSIBILITY FOR ANY CONSEQUENCE +RESULTING FROM THE USE, MODIFICATION, OR +REDISTRIBUTION OF THIS SOFTWARE. +*/ + +// Please see the project home page for documentation +// code.google.com/p/high-concurrency-btree + +#define _FILE_OFFSET_BITS 64 +#define _LARGEFILE64_SOURCE + +#ifdef linux +#define _GNU_SOURCE +#endif + +#ifdef unix +#include +#include +#include +#include +#include +#include +#include +#include +#else +#define WIN32_LEAN_AND_MEAN +#include +#include +#include +#include +#include +#include +#include +#endif + +#include +#include + +typedef unsigned long long uid; + +#ifndef unix +typedef unsigned long long off64_t; +typedef unsigned short ushort; +typedef unsigned int uint; +#endif + +#define BT_ro 0x6f72 // ro +#define BT_rw 0x7772 // rw + +#define BT_latchtable 128 // number of latch manager slots + +#define BT_maxbits 24 // maximum page size in bits +#define BT_minbits 9 // minimum page size in bits +#define BT_minpage (1 << BT_minbits) // minimum page size +#define BT_maxpage (1 << BT_maxbits) // maximum page size + +/* +There are five lock types for each node in three independent sets: +1. (set 1) AccessIntent: Sharable. Going to Read the node. Incompatible with NodeDelete. +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. +*/ + +typedef enum{ + BtLockAccess, + BtLockDelete, + BtLockRead, + BtLockWrite, + BtLockParent, + BtLockPin +}BtLock; + +// Define the length of the page and key pointers + +#define BtId 6 + +// Page key slot definition. + +// If BT_maxbits is 15 or less, you can save 4 bytes +// for each key stored by making the first two uints +// into ushorts. You can also save 4 bytes by removing +// the tod field from the key. + +// Keys are marked dead, but remain on the page until +// it cleanup is called. The fence key (highest key) for +// the page is always present, even after cleanup. + +typedef struct { + uint off:BT_maxbits; // page offset for key start + uint dead:1; // set for deleted key + uint tod; // time-stamp for key + unsigned char id[BtId]; // id associated with key +} BtSlot; + +// The key structure occupies space at the upper end of +// each page. It's a length byte followed by the value +// bytes. + +typedef struct { + unsigned char len; + unsigned char key[1]; +} *BtKey; + +// The first part of an index page. +// It is immediately followed +// by the BtSlot array of keys. + +typedef struct Page { + uint cnt; // count of keys in page + uint act; // count of active keys + uint min; // next key offset + uint foster; // count of foster children + unsigned char bits; // page size in bits + unsigned char lvl:6; // level of page + unsigned char kill:1; // page is being deleted + unsigned char dirty:1; // page needs to be cleaned + 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]; + volatile ushort slot; // Latch table entry at head of chain +} BtHashEntry; + +// latch table lock structure +// implements a fair read-write lock + +typedef struct { +#ifdef unix + pthread_rwlock_t lock[1]; +#else + SRWLOCK srw[1]; +#endif +} BtLatch; + +typedef struct { + BtLatch readwr[1]; // read/write page lock + BtLatch access[1]; // Access Intent/Page delete + BtLatch parent[1]; // adoption of foster children + 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; + +// 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 + ushort pin; // mapped page pin counter + ushort slot; // slot index in this array + void *hashprev; // previous pool entry for the same hash idx + void *hashnext; // next pool entry for the same hash idx +#ifndef unix + HANDLE hmap; // Windows memory mapping handle +#endif +} BtPool; + +// structure for latch manager on ALLOC_page + +typedef struct { + struct Page 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 { + uint page_size; // page size + uint page_bits; // page size in bits + 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 +#else + HANDLE idx; +#endif + ushort poolcnt; // highest page pool node in use + ushort poolmax; // highest page pool node allocated + ushort poolmask; // total size 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 + BtPool *pool; // memory pool page segments + BtSpinLatch *latch; // latches for pool hash slots + BtLatchMgr *latchmgr; // mapped latch page from allocation page + BtLatchSet *latchsets; // mapped latch set from latch pages +#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 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 + uid page_no; // current page number + uid cursor_page; // current cursor page number + BtLatchSet *set; // current page latch set + unsigned char *mem; // frame, cursor, page memory buffer + int err; // last error +} BtDb; + +typedef enum { + BTERR_ok = 0, + BTERR_struct, + BTERR_ovflw, + BTERR_lock, + BTERR_map, + BTERR_wrt, + BTERR_hash, + BTERR_latch +} BTERR; + +// 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); +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); + +// manager functions +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 + +extern BtKey bt_key (BtDb *bt, uint slot); +extern uid bt_uid (BtDb *bt, uint slot); +extern uint bt_tod (BtDb *bt, uint slot); + +// BTree page number constants +#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 + +#define MIN_lvl 2 + +// The page is allocated from low and hi ends. +// The key offsets and row-id's are allocated +// from the bottom, while the text of the key +// is allocated from the top. When the two +// areas meet, the page is split into two. + +// 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. + +// The b-tree root is always located at page 1. +// The first leaf page of level zero is always +// located on page 2. + +// 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. + +// 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. + +// Groups of pages called segments from the btree are +// cached with memory mapping. 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. + +// Page 0 is dedicated to lock for new page extensions, +// and chains empty pages together for reuse. + +// Empty pages are chained together through the ALLOC page and reused. + +// Access macros to address slot and key values from the page + +#define slotptr(page, slot) (((BtSlot *)(page+1)) + (slot-1)) +#define keyptr(page, slot) ((BtKey)((unsigned char*)(page) + slotptr(page, slot)->off)) + +void bt_putid(unsigned char *dest, uid id) +{ +int i = BtId; + + while( i-- ) + dest[i] = (unsigned char)id, id >>= 8; +} + +uid bt_getid(unsigned char *src) +{ +uid id = 0; +int i; + + for( i = 0; i < BtId; i++ ) + id <<= 8, id |= *src++; + + return id; +} + +// wait until write lock mode is clear +// and add 1 to the share count + +void bt_spinreadlock(BtSpinLatch *latch) +{ +ushort prev; + + do { +#ifdef unix + while( __sync_fetch_and_or((ushort *)latch, Mutex) & Mutex ) + sched_yield(); +#else + while( _InterlockedOr16((ushort *)latch, Mutex) & Mutex ) + SwitchToThread(); +#endif + + // see if exclusive request is granted or pending + + if( prev = !(latch->exclusive | latch->pending) ) +#ifdef unix + __sync_fetch_and_add((ushort *)latch, Share); +#else + _InterlockedExchangeAdd16 ((ushort *)latch, Share); +#endif + +#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 +} + +// wait for other read and write latches to relinquish + +void bt_spinwritelock(BtSpinLatch *latch) +{ +ushort prev; + + do { +#ifdef unix + while( __sync_fetch_and_or((ushort *)latch, Mutex | Pending) & Mutex ) + sched_yield(); +#else + while( _InterlockedOr16((ushort *)latch, Mutex | Pending) & Mutex ) + SwitchToThread(); +#endif + if( prev = !(latch->share | latch->exclusive) ) +#ifdef unix + __sync_fetch_and_or((ushort *)latch, Write); +#else + _InterlockedOr16((ushort *)latch, Write); +#endif + +#ifdef unix + __sync_fetch_and_and ((ushort *)latch, ~(Mutex | Pending)); +#else + _InterlockedAnd16((ushort *)latch, ~(Mutex | Pending)); +#endif + if( prev ) + return; +#ifdef unix + sched_yield(); +#else + SwitchToThread(); +#endif + } while( 1 ); +} + +// try to obtain write lock + +// return 1 if obtained, +// 0 otherwise + +int bt_spinwritetry(BtSpinLatch *latch) +{ +ushort prev; + +#ifdef unix + if( prev = __sync_fetch_and_or((ushort *)latch, Mutex), prev & Mutex ) + return 0; +#else + if( prev = _InterlockedOr16((ushort *)latch, Mutex), prev & 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 + +#ifdef unix + __sync_fetch_and_and ((ushort *)latch, ~Mutex); +#else + _InterlockedAnd16((ushort *)latch, ~Mutex); +#endif + return !prev; +} + +// clear write mode + +void bt_spinreleasewrite(BtSpinLatch *latch) +{ +#ifdef unix + __sync_fetch_and_and ((ushort *)latch, ~Write); +#else + _InterlockedAnd16((ushort *)latch, ~Write); +#endif +} + +// decrement reader count + +void bt_spinreleaseread(BtSpinLatch *latch) +{ +#ifdef unix + __sync_fetch_and_add((ushort *)latch, -Share); +#else + _InterlockedExchangeAdd16 ((ushort *)latch, -Share); +#endif +} + +void bt_initlockset (BtLatchSet *set) +{ +#ifdef unix +pthread_rwlockattr_t rwattr[1]; + + pthread_rwlockattr_init (rwattr); + pthread_rwlockattr_setkind_np (rwattr, PTHREAD_RWLOCK_PREFER_WRITER_NONRECURSIVE_NP); + pthread_rwlockattr_setpshared (rwattr, PTHREAD_PROCESS_SHARED); + + pthread_rwlock_init (set->readwr->lock, rwattr); + pthread_rwlock_init (set->access->lock, rwattr); + pthread_rwlock_init (set->parent->lock, rwattr); + pthread_rwlockattr_destroy (rwattr); +#else + InitializeSRWLock (set->readwr->srw); + InitializeSRWLock (set->access->srw); + InitializeSRWLock (set->parent->srw); +#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; +} + +// find existing latchset or inspire new one +// return with latchset pinned + +BtLatchSet *bt_bindlatch (BtDb *bt, uid page_no, int incr) +{ +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 && incr ) { +#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; + if( incr ) +#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; + if( incr ) +#ifdef unix + __sync_fetch_and_add(&set->pin, 1); +#else + _InterlockedIncrement16 (&set->pin); +#endif + bt_initlockset (set); + 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); + + if( incr ) +#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; +uint slot; + + // release mapped pages + // note that slot zero is never used + + for( slot = 1; slot < mgr->poolmax; slot++ ) { + pool = mgr->pool + slot; + if( pool->slot ) +#ifdef unix + munmap (pool->map, (mgr->poolmask+1) << mgr->page_bits); +#else + { + FlushViewOfFile(pool->map, 0); + UnmapViewOfFile(pool->map); + CloseHandle(pool->hmap); + } +#endif + } + +#ifdef unix + close (mgr->idx); + free (mgr->pool); + free (mgr->hash); + free (mgr->latch); + free (mgr->pooladvise); + free (mgr); +#else + FlushFileBuffers(mgr->idx); + CloseHandle(mgr->idx); + GlobalFree (mgr->pool); + GlobalFree (mgr->hash); + GlobalFree (mgr->latch); + GlobalFree (mgr); +#endif +} + +// close and release memory + +void bt_close (BtDb *bt) +{ +#ifdef unix + if ( bt->mem ) + free (bt->mem); +#else + if ( bt->mem) + VirtualFree (bt->mem, 0, MEM_RELEASE); +#endif + free (bt); +} + +// open/create new btree buffer manager + +// call with file_name, BT_openmode, bits in page size (e.g. 16), +// size of mapped page pool (e.g. 8192) + +BtMgr *bt_mgr (char *name, uint mode, uint bits, uint poolmax, uint segsize, uint hashsize) +{ +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]; +#endif + + // determine sanity of page size and buffer pool + + if( bits > BT_maxbits ) + bits = BT_maxbits; + else if( bits < BT_minbits ) + bits = BT_minbits; + + if( !poolmax ) + return NULL; // must have buffer pool + +#ifdef unix + mgr = calloc (1, sizeof(BtMgr)); + + mgr->idx = open ((char*)name, O_RDWR | O_CREAT, 0666); + + if( mgr->idx == -1 ) + return free(mgr), NULL; + + cacheblk = 4096; // minimum mmap segment size for unix + +#else + mgr = GlobalAlloc (GMEM_FIXED|GMEM_ZEROINIT, sizeof(BtMgr)); + attr = FILE_ATTRIBUTE_NORMAL; + mgr->idx = CreateFile(name, GENERIC_READ| GENERIC_WRITE, FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, OPEN_ALWAYS, attr, NULL); + + if( mgr->idx == INVALID_HANDLE_VALUE ) + return GlobalFree(mgr), NULL; + + // normalize cacheblk to multiple of sysinfo->dwAllocationGranularity + GetSystemInfo(sysinfo); + cacheblk = sysinfo->dwAllocationGranularity; +#endif + +#ifdef unix + 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, latchmgr, BT_minpage, 0) == BT_minpage ) + bits = latchmgr->alloc->bits; + else + return free(mgr), free(latchmgr), NULL; + } else if( mode == BT_ro ) + return bt_mgrclose (mgr), NULL; +#else + latchmgr = VirtualAlloc(NULL, BT_maxpage, MEM_COMMIT, PAGE_READWRITE); + size = GetFileSize(mgr->idx, amt); + + if( size || *amt ) { + if( !ReadFile(mgr->idx, (char *)latchmgr, BT_minpage, amt, NULL) ) + return bt_mgrclose (mgr), NULL; + bits = latchmgr->alloc->bits; + } else if( mode == BT_ro ) + return bt_mgrclose (mgr), NULL; +#endif + + mgr->page_size = 1 << bits; + mgr->page_bits = bits; + + mgr->poolmax = poolmax; + mgr->mode = mode; + + if( cacheblk < mgr->page_size ) + cacheblk = mgr->page_size; + + // mask for partial memmaps + + mgr->poolmask = (cacheblk >> bits) - 1; + + // see if requested size of pages per memmap is greater + + if( (1 << segsize) > mgr->poolmask ) + mgr->poolmask = (1 << segsize) - 1; + + mgr->seg_bits = 0; + + while( (1 << mgr->seg_bits) <= mgr->poolmask ) + mgr->seg_bits++; + + mgr->hashsize = hashsize; + +#ifdef unix + mgr->pool = calloc (poolmax, sizeof(BtPool)); + mgr->hash = calloc (hashsize, sizeof(ushort)); + mgr->latch = calloc (hashsize, sizeof(BtSpinLatch)); + 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)); +#endif + + if( size || *amt ) + 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)); + + // initialize latch manager + + 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, latchmgr, mgr->page_size) < mgr->page_size ) + return bt_mgrclose (mgr), NULL; +#else + 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 (latchmgr, 0, 1 << bits); + latchmgr->alloc->bits = mgr->page_bits; + + for( lvl=MIN_lvl; lvl--; ) { + 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; + 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, latchmgr, mgr->page_size) < mgr->page_size ) + return bt_mgrclose (mgr), NULL; +#else + 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 + } + + // clear out latch manager locks + // and rest of pages to round out segment + + memset(latchmgr, 0, mgr->page_size); + last = MIN_lvl + 1; + + while( last <= ((MIN_lvl + 1 + nlatchpage) | mgr->poolmask) ) { +#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 + last++; + } + +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 (latchmgr); +#else + VirtualFree (latchmgr, 0, MEM_RELEASE); +#endif + return mgr; +} + +// open BTree access method +// based on buffer manager + +BtDb *bt_open (BtMgr *mgr) +{ +BtDb *bt = malloc (sizeof(*bt)); + + memset (bt, 0, sizeof(*bt)); + bt->mgr = mgr; +#ifdef unix + bt->mem = malloc (3 *mgr->page_size); +#else + bt->mem = VirtualAlloc(NULL, 3 * mgr->page_size, MEM_COMMIT, PAGE_READWRITE); +#endif + bt->frame = (BtPage)bt->mem; + bt->zero = (BtPage)(bt->mem + 1 * mgr->page_size); + bt->cursor = (BtPage)(bt->mem + 2 * mgr->page_size); + return bt; +} + +// compare two keys, returning > 0, = 0, or < 0 +// as the comparison value + +int keycmp (BtKey key1, unsigned char *key2, uint len2) +{ +uint len1 = key1->len; +int ans; + + if( ans = memcmp (key1->key, key2, len1 > len2 ? len2 : len1) ) + return ans; + + if( len1 > len2 ) + return 1; + if( len1 < len2 ) + return -1; + + return 0; +} + +// Latch Manager + +void bt_readlock(BtLatch *latch) +{ +#ifdef unix + pthread_rwlock_rdlock (latch->lock); +#else + AcquireSRWLockShared (latch->srw); +#endif +} + +// wait for other read and write latches to relinquish + +void bt_writelock(BtLatch *latch) +{ +#ifdef unix + pthread_rwlock_wrlock (latch->lock); +#else + AcquireSRWLockExclusive (latch->srw); +#endif +} + +// try to obtain write lock + +// return 1 if obtained, +// 0 if already write or read locked + +int bt_writetry(BtLatch *latch) +{ +int result = 0; + +#ifdef unix + result = !pthread_rwlock_trywrlock (latch->lock); +#else + result = TryAcquireSRWLockExclusive (latch->srw); +#endif + return result; +} + +// clear write mode + +void bt_releasewrite(BtLatch *latch) +{ +#ifdef unix + pthread_rwlock_unlock (latch->lock); +#else + ReleaseSRWLockExclusive (latch->srw); +#endif +} + +// decrement reader count + +void bt_releaseread(BtLatch *latch) +{ +#ifdef unix + pthread_rwlock_unlock (latch->lock); +#else + ReleaseSRWLockShared (latch->srw); +#endif +} + +// Buffer Pool mgr + +// find segment in pool +// must be called with hashslot idx locked +// return NULL if not there +// otherwise return node + +BtPool *bt_findpool(BtDb *bt, uid page_no, uint idx) +{ +BtPool *pool; +uint slot; + + // compute start of hash chain in pool + + if( slot = bt->mgr->hash[idx] ) + pool = bt->mgr->pool + slot; + else + return NULL; + + page_no &= ~bt->mgr->poolmask; + + while( pool->basepage != page_no ) + if( pool = pool->hashnext ) + continue; + else + return NULL; + + return pool; +} + +// add segment to hash table + +void bt_linkhash(BtDb *bt, BtPool *pool, uid page_no, int idx) +{ +BtPool *node; +uint slot; + + pool->hashprev = pool->hashnext = NULL; + pool->basepage = page_no & ~bt->mgr->poolmask; + pool->lru = 1; + + if( slot = bt->mgr->hash[idx] ) { + node = bt->mgr->pool + slot; + pool->hashnext = node; + node->hashprev = pool; + } + + 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) +{ +off64_t off = (page_no & ~bt->mgr->poolmask) << bt->mgr->page_bits; +off64_t limit = off + ((bt->mgr->poolmask+1) << bt->mgr->page_bits); +int flag; + +#ifdef unix + flag = PROT_READ | ( bt->mgr->mode == BT_ro ? 0 : PROT_WRITE ); + 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 + flag = ( bt->mgr->mode == BT_ro ? PAGE_READONLY : PAGE_READWRITE ); + pool->hmap = CreateFileMapping(bt->mgr->idx, NULL, flag, (DWORD)(limit >> 32), (DWORD)limit, NULL); + if( !pool->hmap ) + return bt->err = BTERR_map; + + flag = ( bt->mgr->mode == BT_ro ? FILE_MAP_READ : FILE_MAP_WRITE ); + pool->map = MapViewOfFile(pool->hmap, flag, (DWORD)(off >> 32), (DWORD)off, (bt->mgr->poolmask+1) << bt->mgr->page_bits); + if( !pool->map ) + return bt->err = BTERR_map; +#endif + return bt->err = 0; +} + +// find or place requested page in segment-pool +// return pool table entry, incrementing pin + +BtPool *bt_pinpage(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]); + + // look up in hash table + + if( pool = bt_findpool(bt, page_no, idx) ) { +#ifdef unix + __sync_fetch_and_add(&pool->pin, 1); +#else + _InterlockedIncrement16 (&pool->pin); +#endif + bt_spinreleaseread (&bt->mgr->latch[idx]); + pool->lru++; + return pool; + } + + // upgrade to write lock + + bt_spinreleaseread (&bt->mgr->latch[idx]); + bt_spinwritelock (&bt->mgr->latch[idx]); + + // try to find page in pool with write lock + + if( pool = bt_findpool(bt, page_no, idx) ) { +#ifdef unix + __sync_fetch_and_add(&pool->pin, 1); +#else + _InterlockedIncrement16 (&pool->pin); +#endif + bt_spinreleasewrite (&bt->mgr->latch[idx]); + pool->lru++; + return pool; + } + + // allocate a new pool node + // and add to hash table + +#ifdef unix + slot = __sync_fetch_and_add(&bt->mgr->poolcnt, 1); +#else + slot = _InterlockedIncrement16 (&bt->mgr->poolcnt) - 1; +#endif + + if( ++slot < bt->mgr->poolmax ) { + pool = bt->mgr->pool + slot; + pool->slot = slot; + + 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 + _InterlockedIncrement16 (&pool->pin); +#endif + bt_spinreleasewrite (&bt->mgr->latch[idx]); + return pool; + } + + // pool table is full + // find best pool entry to evict + +#ifdef unix + __sync_fetch_and_add(&bt->mgr->poolcnt, -1); +#else + _InterlockedDecrement16 (&bt->mgr->poolcnt); +#endif + + while( 1 ) { +#ifdef unix + victim = __sync_fetch_and_add(&bt->mgr->evicted, 1); +#else + victim = _InterlockedIncrement16 (&bt->mgr->evicted) - 1; +#endif + victim %= bt->mgr->hashsize; + + // try to get write lock + // skip entry if not obtained + + if( !bt_spinwritetry (&bt->mgr->latch[victim]) ) + continue; + + // if cache entry is empty + // or no slots are unpinned + // skip this entry + + if( !(pool = bt_findlru(bt, bt->mgr->hash[victim])) ) { + bt_spinreleasewrite (&bt->mgr->latch[victim]); + continue; + } + + // unlink victim pool node from hash table + + if( node = pool->hashprev ) + node->hashnext = pool->hashnext; + else if( node = pool->hashnext ) + bt->mgr->hash[victim] = node->slot; + else + bt->mgr->hash[victim] = 0; + + if( node = pool->hashnext ) + node->hashprev = pool->hashprev; + + bt_spinreleasewrite (&bt->mgr->latch[victim]); + + // remove old file mapping +#ifdef unix + munmap (pool->map, (bt->mgr->poolmask+1) << bt->mgr->page_bits); +#else + FlushViewOfFile(pool->map, 0); + UnmapViewOfFile(pool->map); + CloseHandle(pool->hmap); +#endif + pool->map = NULL; + + // create new pool mapping + // and link into hash table + + 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 + _InterlockedIncrement16 (&pool->pin); +#endif + bt_spinreleasewrite (&bt->mgr->latch[idx]); + return pool; + } +} + +// place write, read, or parent lock on requested page_no. +// pin to buffer pool and return latchset pointer + +BtLatchSet *bt_lockpage(BtDb *bt, uid page_no, BtLock mode, BtPage *pageptr, BtLatchSet *set) +{ +BtPool *pool; +uint subpage; +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 NULL; + + if( set ) +#ifdef unix + __sync_fetch_and_add(&set->pin, 1); +#else + _InterlockedIncrement16 (&set->pin); +#endif + else if( !(set = bt_bindlatch (bt, page_no, 1)) ) + return NULL; + + page = (BtPage)(pool->map + (subpage << bt->mgr->page_bits)); + +#ifdef unix + { + uint idx = subpage / 8; + uint bit = subpage % 8; + + if( mode == BtLockRead || mode == BtLockWrite ) + 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 (set->readwr); + break; + case BtLockWrite: + bt_writelock (set->readwr); + break; + case BtLockAccess: + bt_readlock (set->access); + break; + case BtLockDelete: + bt_writelock (set->access); + break; + case BtLockParent: + bt_writelock (set->parent); + break; + case BtLockPin: + break; + default: + return bt->err = BTERR_lock, NULL; + } + + if( pageptr ) + *pageptr = page; + + return set; +} + +// remove write, read, or parent lock on requested page_no. + +BTERR bt_unlockpage(BtDb *bt, uid page_no, BtLock mode, BtLatchSet *set) +{ +BtPool *pool; +uint idx; + + // 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_spinreadlock (&bt->mgr->latch[idx]); + + if( !(pool = bt_findpool(bt, page_no, idx)) ) + return bt->err = BTERR_hash; + + bt_spinreleaseread (&bt->mgr->latch[idx]); + + switch( mode ) { + case BtLockRead: + bt_releaseread (set->readwr); + break; + case BtLockWrite: + bt_releasewrite (set->readwr); + break; + case BtLockAccess: + bt_releaseread (set->access); + break; + case BtLockDelete: + bt_releasewrite (set->access); + break; + case BtLockParent: + bt_releasewrite (set->parent); + break; + case BtLockPin: + break; + default: + return bt->err = BTERR_lock; + } + +#ifdef unix + __sync_fetch_and_add(&pool->pin, -1); + __sync_fetch_and_add (&set->pin, -1); +#else + _InterlockedDecrement16 (&pool->pin); + _InterlockedDecrement16 (&set->pin); +#endif + return bt->err = 0; +} + +// deallocate a deleted page +// place on free chain out of allocator page +// fence key must already be removed from parent + +BTERR bt_freepage(BtDb *bt, uid page_no, BtLatchSet *set) +{ + // obtain delete lock on deleted page + + if( !bt_lockpage(bt, page_no, BtLockDelete, NULL, set) ) + return bt->err; + + // obtain write lock on deleted page + + if( !bt_lockpage(bt, page_no, BtLockWrite, &bt->temp, set) ) + return bt->err; + + // lock allocation page + + bt_spinwritelock(bt->mgr->latchmgr->lock); + + // store free chain in allocation page second right + bt_putid(bt->temp->right, bt_getid(bt->mgr->latchmgr->alloc[1].right)); + bt_putid(bt->mgr->latchmgr->alloc[1].right, page_no); + + // unlock page zero + + bt_spinreleasewrite(bt->mgr->latchmgr->lock); + + // remove write lock on deleted node + + if( bt_unlockpage(bt, page_no, BtLockWrite, set) ) + return bt->err; + + // remove delete lock on deleted node + + if( bt_unlockpage(bt, page_no, BtLockDelete, set) ) + return bt->err; + + return 0; +} + +// allocate a new page and write page into it + +uid bt_newpage(BtDb *bt, BtPage page) +{ +BtLatchSet *set; +uid new_page; +BtPage pmap; +int reuse; + + // lock allocation page + + bt_spinwritelock(bt->mgr->latchmgr->lock); + + // use empty chain first + // else allocate empty page + + if( new_page = bt_getid(bt->mgr->latchmgr->alloc[1].right) ) { + if( !(set = bt_lockpage (bt, new_page, BtLockWrite, &bt->temp, NULL)) ) + return 0; + bt_putid(bt->mgr->latchmgr->alloc[1].right, bt_getid(bt->temp->right)); + if( bt_unlockpage (bt, new_page, BtLockWrite, set) ) + return 0; + reuse = 1; + } else { + new_page = bt_getid(bt->mgr->latchmgr->alloc->right); + bt_putid(bt->mgr->latchmgr->alloc->right, new_page+1); + 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; + } +#else + // bring new page into pool and copy page. + // this will extend the file into the new pages. + + if( !(set = bt_lockpage(bt, new_page, BtLockWrite, &pmap, NULL)) ) + return 0; + + memcpy(pmap, page, bt->mgr->page_size); + + if( bt_unlockpage (bt, new_page, BtLockWrite, set) ) + return 0; +#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) +{ +uint diff, higher = bt->page->cnt, low = 1, slot; + + // low is the lowest candidate, higher is already + // tested as .ge. the given key, loop ends when they meet + + while( diff = higher - low ) { + slot = low + ( diff >> 1 ); + if( keycmp (keyptr(bt->page, slot), key, len) < 0 ) + low = slot + 1; + else + higher = slot; + } + + return higher; +} + +// 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) +{ +uid page_no = ROOT_page, prevpage = 0; +BtLatchSet *set, *prevset; +uint drill = 0xff, slot; +uint mode, prevmode; + + bt->set = NULL; + + // start at root of btree and drill down + + do { + // determine lock mode of drill level + mode = (lock == BtLockWrite) && (drill == lvl) ? BtLockWrite : BtLockRead; + + bt->page_no = page_no; + + // obtain access lock using lock chaining with Access mode + + if( page_no > ROOT_page ) + if( !(bt->set = bt_lockpage(bt, page_no, BtLockAccess, NULL, NULL)) ) + return 0; + + // now unlock our (possibly foster) parent + + if( prevpage ) + if( bt_unlockpage(bt, prevpage, prevmode, prevset) ) + return 0; + else + prevpage = 0; + + // obtain read lock using lock chaining + // and pin page contents + + if( !(bt->set = bt_lockpage(bt, page_no, mode, &bt->page, bt->set)) ) + return 0; + + if( page_no > ROOT_page ) + if( bt_unlockpage(bt, page_no, BtLockAccess, bt->set) ) + return 0; + + // re-read and re-lock root after determining actual level of root + + if( bt->page_no == ROOT_page ) + if( bt->page->lvl != drill) { + drill = bt->page->lvl; + + if( lock == BtLockWrite && drill == lvl ) + if( bt_unlockpage(bt, page_no, mode, bt->set) ) + return 0; + else + continue; + } + + prevpage = bt->page_no; + prevset = bt->set; + prevmode = mode; + + // if page is being deleted, + // move back to preceeding page + + if( bt->page->kill ) { + page_no = bt_getid (bt->page->right); + continue; + } + + // find key on page at this level + // and descend to requested level + + slot = bt_findslot (bt, key, len); + + // is this slot a foster child? + + if( slot <= bt->page->cnt - bt->page->foster ) + if( drill == lvl ) + return slot; + + while( slotptr(bt->page, slot)->dead ) + if( slot++ < bt->page->cnt ) + continue; + else + goto slideright; + + if( slot <= bt->page->cnt - bt->page->foster ) + drill--; + + // continue down / right using overlapping locks + // to protect pages being killed or split. + + page_no = bt_getid(slotptr(bt->page, slot)->id); + continue; + +slideright: + page_no = bt_getid(bt->page->right); + + } while( page_no ); + + // return error on end of chain + + bt->err = BTERR_struct; + return 0; // return error +} + +// find and delete key on page by marking delete flag bit +// when page becomes empty, delete it from the btree + +BTERR bt_deletekey (BtDb *bt, unsigned char *key, uint len, uint lvl) +{ +unsigned char leftkey[256], rightkey[256]; +BtLatchSet *rset, *set; +uid page_no, right; +uint slot, tod; +BtKey ptr; + + 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 + + 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--; + } + + // return if page is not empty, or it has no right sibling + + right = bt_getid(bt->page->right); + page_no = bt->page_no; + set = bt->set; + + if( !right || bt->page->act ) + return bt_unlockpage(bt, page_no, BtLockWrite, set); + + // obtain Parent lock over write lock + + if( !bt_lockpage(bt, page_no, BtLockParent, NULL, set) ) + return bt->err; + + // cache copy of key to delete + + ptr = keyptr(bt->page, bt->page->cnt); + memcpy(leftkey, ptr, ptr->len + 1); + + // lock and map right page + + if( !(rset = bt_lockpage(bt, right, BtLockWrite, &bt->temp, NULL)) ) + return bt->err; + + // pull contents of next page into current empty page + memcpy (bt->page, bt->temp, bt->mgr->page_size); + + // cache copy of key to update + ptr = keyptr(bt->temp, bt->temp->cnt); + memcpy(rightkey, ptr, ptr->len + 1); + + // Mark right page as deleted and point it to left page + // until we can post updates at higher level. + + bt_putid(bt->temp->right, page_no); + bt->temp->kill = 1; + bt->temp->cnt = 0; + + if( bt_unlockpage(bt, right, BtLockWrite, rset) ) + return bt->err; + if( bt_unlockpage(bt, page_no, BtLockWrite, set) ) + return bt->err; + + // delete old lower key to consolidated node + + if( bt_deletekey (bt, leftkey + 1, *leftkey, lvl + 1) ) + return bt->err; + + // redirect higher key directly to consolidated node + + if( slot = bt_loadpage (bt, rightkey+1, *rightkey, lvl+1, BtLockWrite) ) + ptr = keyptr(bt->page, slot); + else + return bt->err; + + // since key already exists, update id + + if( keycmp (ptr, rightkey+1, *rightkey) ) + return bt->err = BTERR_struct; + + slotptr(bt->page, slot)->dead = 0; + bt_putid(slotptr(bt->page,slot)->id, page_no); + + if( bt_unlockpage(bt, bt->page_no, BtLockWrite, bt->set) ) + return bt->err; + + // obtain write lock and + // add right block to free chain + + if( bt_freepage (bt, right, rset) ) + return bt->err; + + // remove ParentModify lock + + if( bt_unlockpage(bt, page_no, BtLockParent, set) ) + return bt->err; + + return 0; +} + +// find key in leaf level and return row-id + +uid bt_findkey (BtDb *bt, unsigned char *key, uint len) +{ +uint slot; +BtKey ptr; +uid id; + + if( slot = bt_loadpage (bt, key, len, 0, BtLockRead) ) + ptr = keyptr(bt->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, bt->set) ) + return 0; + + return id; +} + +// check page for space available, +// clean if necessary and return +// 0 - page needs splitting +// 1 - go ahead + +uint bt_cleanpage(BtDb *bt, uint amt) +{ +uint nxt = bt->mgr->page_size; +BtPage page = bt->page; +uint cnt = 0, idx = 0; +uint max = page->cnt; +BtKey key; + + if( page->min >= (max+1) * sizeof(BtSlot) + sizeof(*page) + amt + 1 ) + return 1; + + // skip cleanup if nothing to reclaim + + if( !page->dirty ) + return 0; + + memcpy (bt->frame, page, bt->mgr->page_size); + + // skip page info and set rest of page to zero + + memset (page+1, 0, bt->mgr->page_size - sizeof(*page)); + page->dirty = 0; + page->act = 0; + + // try cleaning up page first + + while( cnt++ < max ) { + // always leave fence key and foster children in list + if( cnt < max - page->foster && slotptr(bt->frame,cnt)->dead ) + continue; + + // copy key + 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; + + // see if page has enough space now, or does it need splitting? + + if( page->min >= (idx+1) * sizeof(BtSlot) + sizeof(*page) + amt + 1 ) + return 1; + + 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; + + // calculate next available slot and copy key into page + + page->min -= len + 1; + ((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; +} + +// 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) +{ +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) + + root->act--; + root->cnt--; + root->foster--; + + // Save left fence key. + + key = keyptr(root, root->cnt); + memcpy (fencekey, key, key->len + 1); + + // copy the lower keys into a new left page + + 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 + + memset (root+1, 0, bt->mgr->page_size - sizeof(*root)); + + // insert left fence key on empty newroot page + + nxt -= *fencekey + 1; + memcpy ((unsigned char *)root + nxt, fencekey, *fencekey + 1); + bt_putid(slotptr(root, 1)->id, new_page); + slotptr(root, 1)->off = nxt; + + // insert stopper 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); + slotptr(root, 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++; + + // release root (bt->page) + + return bt_unlockpage(bt, ROOT_page, BtLockWrite, bt->set); +} + +// split already locked full node +// in current page variables +// 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; +BtPage page = bt->page; +uint tod = time(NULL); +uint lvl = page->lvl; +uid new_page, right; +BtKey key; + + // initialize frame buffer + + memset (bt->frame, 0, bt->mgr->page_size); + max = page->cnt - page->foster; + tod = (uint)time(NULL); + cnt = max / 2; + idx = 0; + + // split higher half of keys to bt->frame + // leaving foster children in the left node. + + while( cnt++ < max ) { + key = keyptr(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); + slotptr(bt->frame, idx)->tod = slotptr(page, cnt)->tod; + slotptr(bt->frame, idx)->off = nxt; + bt->frame->act++; + } + + // transfer right link node + + if( page_no > ROOT_page ) { + right = bt_getid (page->right); + bt_putid(bt->frame->right, right); + } + + 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 frame to it. + + if( !(new_page = bt_newpage(bt, bt->frame)) ) + return bt->err; + + // remember fence key for new page to add + // as foster child + + key = keyptr(bt->frame, idx); + memcpy (fencekey, key, key->len + 1); + + // update lower keys and foster children 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->act = 0; + cnt = 0; + idx = 0; + + // assemble page of smaller keys + // to remain in the old page + + 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++; + } + + // insert new foster child at beginning 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 if any + + 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); + slotptr(page, idx)->tod = slotptr(bt->frame, cnt)->tod; + slotptr(page, idx)->off = nxt; + page->act++; + } + + 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); + + // keep our latch set + // release wr lock on our page + + if( !bt_lockpage (bt, page_no, BtLockPin, NULL, set) ) + return bt->err; + + if( bt_unlockpage (bt, page_no, BtLockWrite, set) ) + return bt->err; + + // obtain ParentModification lock for current page + // to fix fence key and highest foster child on page + + if( !bt_lockpage (bt, page_no, BtLockParent, NULL, set) ) + return bt->err; + + // get our highest foster child key to find in parent node + + if( !bt_lockpage (bt, page_no, BtLockRead, &page, set) ) + return bt->err; + + key = keyptr(page, page->cnt); + memcpy (fencekey, key, key->len+1); + + if( bt_unlockpage (bt, page_no, BtLockRead, set) ) + return bt->err; + + // update our parent +try_again: + + do { + slot = bt_loadpage (bt, fencekey + 1, *fencekey, lvl + 1, BtLockWrite); + + if( !slot ) + return bt->err; + + // check if parent page has enough space for any possible key + + if( bt_cleanpage (bt, 256) ) + break; + + if( bt_splitpage (bt) ) + return bt->err; + } while( 1 ); + + // see if we are still a foster child from another node + + if( bt_getid (slotptr(bt->page, slot)->id) != page_no ) { + if( bt_unlockpage (bt, bt->page_no, BtLockWrite, bt->set) ) + return bt->err; +#ifdef unix + sched_yield(); +#else + SwitchToThread(); +#endif + goto try_again; + } + + // wait until readers from parent get their locks + // on our page + + if( !bt_lockpage (bt, page_no, BtLockDelete, NULL, set) ) + return bt->err; + + // lock our page for writing + + if( !bt_lockpage (bt, page_no, BtLockWrite, &page, set) ) + return bt->err; + + // switch parent fence key to foster child + + if( slotptr(page, page->cnt)->dead ) + slotptr(bt->page, slot)->dead = 1; + else + bt_putid (slotptr(bt->page, slot)->id, bt_getid(slotptr(page, page->cnt)->id)); + + // remove highest foster child from our page + + page->cnt--; + page->act--; + page->foster--; + page->dirty = 1; + key = keyptr(page, page->cnt); + + // add our new fence key for foster child to our parent + + bt_addkeytopage (bt, slot, key->key, key->len, page_no, tod); + + if( bt_unlockpage (bt, bt->page_no, BtLockWrite, bt->set) ) + return bt->err; + + if( bt_unlockpage (bt, page_no, BtLockDelete, set) ) + return bt->err; + + if( bt_unlockpage (bt, page_no, BtLockWrite, set) ) + return bt->err; + + if( bt_unlockpage (bt, page_no, BtLockParent, set) ) + return bt->err; + + // release extra latch pin + + return bt_unlockpage (bt, page_no, BtLockPin, set); +} + +// Insert new key into the btree at leaf level. + +BTERR bt_insertkey (BtDb *bt, unsigned char *key, uint len, uid id, uint tod) +{ +uint slot, idx; +BtPage page; +BtKey ptr; + + while( 1 ) { + if( slot = bt_loadpage (bt, key, len, 0, 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; + + 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, bt->set); + } + + // check if page has enough space + + if( bt_cleanpage (bt, len) ) + break; + + if( bt_splitpage (bt) ) + return bt->err; + } + + bt_addkeytopage (bt, slot, key, len, id, tod); + + return bt_unlockpage (bt, bt->page_no, BtLockWrite, bt->set); +} + +// cache page of keys into cursor and return starting slot for given key + +uint bt_startkey (BtDb *bt, unsigned char *key, uint len) +{ +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, bt->set) ) + return 0; + + return slot; +} + +// return next slot for cursor page +// or slide cursor right into next page + +uint bt_nextkey (BtDb *bt, uint slot) +{ +BtLatchSet *rset; +BtPage page; +uid right; + + do { + right = bt_getid(bt->cursor->right); + while( slot++ < bt->cursor->cnt - bt->cursor->foster ) + if( slotptr(bt->cursor,slot)->dead ) + continue; + else if( right || (slot < bt->cursor->cnt - bt->cursor->foster) ) + return slot; + else + break; + + if( !right ) + break; + + bt->cursor_page = right; + + if( !(bt->set = bt_lockpage(bt, right, BtLockRead, &page, NULL)) ) + return 0; + + memcpy (bt->cursor, page, bt->mgr->page_size); + + if ( bt_unlockpage(bt, right, BtLockRead, bt->set) ) + return 0; + + slot = 0; + } while( 1 ); + + return bt->err = 0; +} + +BtKey bt_key(BtDb *bt, uint slot) +{ + return keyptr(bt->cursor, slot); +} + +uid bt_uid(BtDb *bt, uint slot) +{ + return bt_getid(slotptr(bt->cursor,slot)->id); +} + +uint bt_tod(BtDb *bt, uint slot) +{ + return slotptr(bt->cursor,slot)->tod; +} + + +#ifdef STANDALONE + +typedef struct { + char type, idx; + char *infile; + BtMgr *mgr; + int num; +} ThreadArg; + +// standalone program to index file of keys +// then list them onto std-out + +#ifdef unix +void *index_file (void *arg) +#else +uint __stdcall index_file (void *arg) +#endif +{ +int line = 0, found = 0, cnt = 0; +uid next, page_no = LEAF_page; // start on first page of leaves +unsigned char key[256]; +ThreadArg *args = arg; +int ch, len = 0, slot; +time_t tod[1]; +BtPage page; +BtKey ptr; +BtDb *bt; +FILE *in; + + bt = bt_open (args->mgr); + time (tod); + + switch(args->type | 0x20) + { + case 'w': + fprintf(stderr, "started indexing for %s\n", args->infile); + if( in = fopen (args->infile, "rb") ) + while( ch = getc(in), ch != EOF ) + if( ch == '\n' ) + { + 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; + + if( bt_insertkey (bt, key, len, line, *tod) ) + fprintf(stderr, "Error %d Line: %d\n", bt->err, line), exit(0); + len = 0; + } + else if( len < 255 ) + key[len++] = ch; + fprintf(stderr, "finished %s for %d keys\n", args->infile, line); + break; + + case 'd': + fprintf(stderr, "started deleting keys for %s\n", args->infile); + if( in = fopen (args->infile, "rb") ) + while( ch = getc(in), ch != EOF ) + if( ch == '\n' ) + { + 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; + + if( bt_deletekey (bt, key, len, 0) ) + fprintf(stderr, "Error %d Line: %d\n", bt->err, line), exit(0); + len = 0; + } + else if( len < 255 ) + key[len++] = ch; + fprintf(stderr, "finished %s for keys, %d \n", args->infile, line); + break; + + case 'f': + fprintf(stderr, "started finding keys for %s\n", args->infile); + if( in = fopen (args->infile, "rb") ) + while( ch = getc(in), ch != EOF ) + if( ch == '\n' ) + { + 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; + + if( bt_findkey (bt, key, len) ) + found++; + else if( bt->err ) + fprintf(stderr, "Error %d Syserr %d Line: %d\n", bt->err, errno, line), exit(0); + len = 0; + } + else if( len < 255 ) + key[len++] = ch; + fprintf(stderr, "finished %s for %d keys, found %d\n", args->infile, line, found); + 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); + } + + break; + + case 'c': + fprintf(stderr, "started reading\n"); + + do { + bt->set = bt_lockpage (bt, page_no, BtLockRead, &page, NULL); + cnt += page->act; + next = bt_getid (page->right); + bt_unlockpage (bt, page_no, BtLockRead, bt->set); + } while( page_no = next ); + + cnt--; // remove stopper key + fprintf(stderr, " Total keys read %d\n", cnt); + break; + } + + bt_close (bt); +#ifdef unix + return NULL; +#else + return 0; +#endif +} + +typedef struct timeval timer; + +int main (int argc, char **argv) +{ +int idx, cnt, len, slot, err; +int segsize, bits = 16; +#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; +int num = 0; +char key[1]; +BtMgr *mgr; +BtKey ptr; +BtDb *bt; + + if( argc < 3 ) { + fprintf (stderr, "Usage: %s idx_file Read/Write/Scan/Delete/Find [page_bits mapped_segments seg_bits line_numbers src_file1 src_file2 ... ]\n", argv[0]); + 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 = 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 + + if( argc > 3 ) + bits = atoi(argv[3]); + + if( argc > 4 ) + poolsize = atoi(argv[4]); + + if( !poolsize ) + fprintf (stderr, "Warning: no mapped_pool\n"); + + if( poolsize > 65535 ) + fprintf (stderr, "Warning: mapped_pool > 65535 segments\n"); + + if( argc > 5 ) + segsize = atoi(argv[5]); + else + segsize = 4; // 16 pages per mmap segment + + if( argc > 6 ) + num = atoi(argv[6]); + + cnt = argc - 7; +#ifdef unix + threads = malloc (cnt * sizeof(pthread_t)); +#else + threads = GlobalAlloc (GMEM_FIXED|GMEM_ZEROINIT, cnt * sizeof(HANDLE)); +#endif + args = malloc (cnt * sizeof(ThreadArg)); + + mgr = bt_mgr ((argv[1]), BT_rw, bits, poolsize, segsize, poolsize / 8); + + if( !mgr ) { + fprintf(stderr, "Index Open Error %s\n", argv[1]); + exit (1); + } + + // fire off threads + + for( idx = 0; idx < cnt; idx++ ) { + args[idx].infile = argv[idx + 7]; + args[idx].type = argv[2][0]; + args[idx].mgr = mgr; + args[idx].num = num; + args[idx].idx = idx; +#ifdef unix + if( err = pthread_create (threads + idx, NULL, index_file, args + idx) ) + fprintf(stderr, "Error creating thread %d\n", err); +#else + threads[idx] = (HANDLE)_beginthreadex(NULL, 65536, index_file, args + idx, 0, NULL); +#endif + } + + // wait for termination + +#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); + bt_mgrclose (mgr); +} + +#endif //STANDALONE diff --git a/fosterbtreee2.c b/fosterbtreee2.c new file mode 100644 index 0000000..86f836d --- /dev/null +++ b/fosterbtreee2.c @@ -0,0 +1,2554 @@ +// foster btree version e2 +// 18 JAN 2014 + +// author: karl malbrain, malbrain@cal.berkeley.edu + +/* +This work, including the source code, documentation +and related data, is placed into the public domain. + +The orginal author is Karl Malbrain. + +THIS SOFTWARE IS PROVIDED AS-IS WITHOUT WARRANTY +OF ANY KIND, NOT EVEN THE IMPLIED WARRANTY OF +MERCHANTABILITY. THE AUTHOR OF THIS SOFTWARE, +ASSUMES _NO_ RESPONSIBILITY FOR ANY CONSEQUENCE +RESULTING FROM THE USE, MODIFICATION, OR +REDISTRIBUTION OF THIS SOFTWARE. +*/ + +// Please see the project home page for documentation +// code.google.com/p/high-concurrency-btree + +#define _FILE_OFFSET_BITS 64 +#define _LARGEFILE64_SOURCE + +#ifdef linux +#define _GNU_SOURCE +#endif + +#ifdef unix +#include +#include +#include +#include +#include +#include +#include +#include +#else +#define WIN32_LEAN_AND_MEAN +#include +#include +#include +#include +#include +#include +#include +#endif + +#include +#include + +typedef unsigned long long uid; + +#ifndef unix +typedef unsigned long long off64_t; +typedef unsigned short ushort; +typedef unsigned int uint; +#endif + +#define BT_ro 0x6f72 // ro +#define BT_rw 0x7772 // rw + +#define BT_latchtable 128 // number of latch manager slots + +#define BT_maxbits 24 // maximum page size in bits +#define BT_minbits 9 // minimum page size in bits +#define BT_minpage (1 << BT_minbits) // minimum page size +#define BT_maxpage (1 << BT_maxbits) // maximum page size + +/* +There are five lock types for each node in three independent sets: +1. (set 1) AccessIntent: Sharable. Going to Read the node. Incompatible with NodeDelete. +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. +*/ + +typedef enum{ + BtLockAccess, + BtLockDelete, + BtLockRead, + BtLockWrite, + BtLockParent +}BtLock; + +// Define the length of the page and key pointers + +#define BtId 6 + +// Page key slot definition. + +// If BT_maxbits is 15 or less, you can save 4 bytes +// for each key stored by making the first two uints +// into ushorts. You can also save 4 bytes by removing +// the tod field from the key. + +// Keys are marked dead, but remain on the page until +// cleanup is called. The fence key (highest key) for +// the page is always present, even after cleanup. + +typedef struct { + uint off:BT_maxbits; // page offset for key start + uint dead:1; // set for deleted key + uint tod; // time-stamp for key + unsigned char id[BtId]; // id associated with key +} BtSlot; + +// The key structure occupies space at the upper end of +// each page. It's a length byte followed by the value +// bytes. + +typedef struct { + unsigned char len; + unsigned char key[1]; +} *BtKey; + +// The first part of an index page. +// It is immediately followed +// 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 + 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 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]; + volatile ushort slot; // Latch table entry at head of chain +} BtHashEntry; + +// latch manager table structure + +typedef struct { +#ifdef unix + pthread_rwlock_t lock[1]; +#else + SRWLOCK srw[1]; +#endif +} BtLatch; + +typedef struct { + BtLatch readwr[1]; // read/write page lock + BtLatch access[1]; // Access Intent/Page delete + BtLatch parent[1]; // adoption of foster children + 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; + +// 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 + ushort pin; // mapped page pin counter + ushort slot; // slot index in this array + void *hashprev; // previous pool entry for the same hash idx + void *hashnext; // next pool entry for the same hash idx +#ifndef unix + HANDLE hmap; // Windows memory mapping handle +#endif +} BtPool; + +// structure for latch manager on ALLOC_page + +typedef struct { + struct Page 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 { + uint page_size; // page size + uint page_bits; // page size in bits + 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 +#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 + BtPool *pool; // memory pool page segments + BtSpinLatch *latch; // latches for pool hash slots + BtLatchMgr *latchmgr; // mapped latch page from allocation page + BtLatchSet *latchsets; // mapped latch set from latch pages +#ifndef unix + HANDLE halloc; // allocation and latch table handle +#endif +} BtMgr; + +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 + uid page_no; // current page number + uid cursor_page; // current cursor page number + BtLatchSet *set; // current page latch set + BtPool *pool; // current page pool + unsigned char *mem; // frame, cursor, page memory buffer + int found; // last delete was found + int err; // last error +} BtDb; + +typedef enum { + BTERR_ok = 0, + BTERR_struct, + BTERR_ovflw, + BTERR_lock, + BTERR_map, + BTERR_wrt, + BTERR_hash, + BTERR_latch +} BTERR; + +// 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 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); + +// manager functions +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 + +extern BtKey bt_key (BtDb *bt, uint slot); +extern uid bt_uid (BtDb *bt, uint slot); +extern uint bt_tod (BtDb *bt, uint slot); + +// BTree page number constants +#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 + +#define MIN_lvl 2 + +// The page is allocated from low and hi ends. +// The key offsets and row-id's are allocated +// from the bottom, while the text of the key +// is allocated from the top. When the two +// areas meet, the page is split into two. + +// 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. + +// The b-tree root is always located at page 1. +// The first leaf page of level zero is always +// located on page 2. + +// 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. + +// 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. + +// Groups of pages called segments from the btree are +// cached with memory mapping. 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. + +// Page 0 is dedicated to lock for new page extensions, +// and chains empty pages together for reuse. + +// Empty pages are chained together through the ALLOC page and reused. + +// Access macros to address slot and key values from the page + +#define slotptr(page, slot) (((BtSlot *)(page+1)) + (slot-1)) +#define keyptr(page, slot) ((BtKey)((unsigned char*)(page) + slotptr(page, slot)->off)) + +void bt_putid(unsigned char *dest, uid id) +{ +int i = BtId; + + while( i-- ) + dest[i] = (unsigned char)id, id >>= 8; +} + +uid bt_getid(unsigned char *src) +{ +uid id = 0; +int i; + + for( i = 0; i < BtId; i++ ) + id <<= 8, id |= *src++; + + return id; +} + +// wait until write lock mode is clear +// and add 1 to the share count + +void bt_spinreadlock(BtSpinLatch *latch) +{ +ushort prev; + + do { +#ifdef unix + while( __sync_fetch_and_or((ushort *)latch, Mutex) & Mutex ) + sched_yield(); +#else + while( _InterlockedOr16((ushort *)latch, Mutex) & Mutex ) + SwitchToThread(); +#endif + + // see if exclusive request is granted or pending + + if( prev = !(latch->exclusive | latch->pending) ) +#ifdef unix + __sync_fetch_and_add((ushort *)latch, Share); +#else + _InterlockedExchangeAdd16 ((ushort *)latch, Share); +#endif + +#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 +} + +// wait for other read and write latches to relinquish + +void bt_spinwritelock(BtSpinLatch *latch) +{ + do { +#ifdef unix + while( __sync_fetch_and_or((ushort *)latch, Mutex | Pending) & 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; + } + +#ifdef unix + __sync_fetch_and_and ((ushort *)latch, ~Mutex); +#else + _InterlockedAnd16((ushort *)latch, ~Mutex); +#endif +#ifdef unix + sched_yield(); +#else + SwitchToThread(); +#endif + } while( 1 ); +} + +// try to obtain write lock + +// return 1 if obtained, +// 0 otherwise + +int bt_spinwritetry(BtSpinLatch *latch) +{ +ushort prev; + +#ifdef unix + if( prev = __sync_fetch_and_or((ushort *)latch, Mutex), prev & Mutex ) + return 0; +#else + if( prev = _InterlockedOr16((ushort *)latch, Mutex), prev & 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 + +#ifdef unix + __sync_fetch_and_and ((ushort *)latch, ~Mutex); +#else + _InterlockedAnd16((ushort *)latch, ~Mutex); +#endif + return !prev; +} + +// clear write mode + +void bt_spinreleasewrite(BtSpinLatch *latch) +{ +#ifdef unix + __sync_fetch_and_and ((ushort *)latch, ~Write); +#else + _InterlockedAnd16((ushort *)latch, ~Write); +#endif +} + +// decrement reader count + +void bt_spinreleaseread(BtSpinLatch *latch) +{ +#ifdef unix + __sync_fetch_and_add((ushort *)latch, -Share); +#else + _InterlockedExchangeAdd16 ((ushort *)latch, -Share); +#endif +} + +void bt_initlockset (BtLatchSet *set, int reuse) +{ +#ifdef unix +pthread_rwlockattr_t rwattr[1]; + + if( reuse ) { + pthread_rwlock_destroy (set->readwr->lock); + pthread_rwlock_destroy (set->access->lock); + pthread_rwlock_destroy (set->parent->lock); + } + + pthread_rwlockattr_init (rwattr); + pthread_rwlockattr_setkind_np (rwattr, PTHREAD_RWLOCK_PREFER_WRITER_NONRECURSIVE_NP); + pthread_rwlockattr_setpshared (rwattr, PTHREAD_PROCESS_SHARED); + + pthread_rwlock_init (set->readwr->lock, rwattr); + pthread_rwlock_init (set->access->lock, rwattr); + pthread_rwlock_init (set->parent->lock, rwattr); + pthread_rwlockattr_destroy (rwattr); +#else + InitializeSRWLock (set->readwr->srw); + InitializeSRWLock (set->access->srw); + InitializeSRWLock (set->parent->srw); +#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; +} + +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_initlockset (set, 0); + 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_initlockset (set, 1); + 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; +uint slot; + + // release mapped pages + // note that slot zero is never used + + for( slot = 1; slot < mgr->poolmax; slot++ ) { + pool = mgr->pool + slot; + if( pool->slot ) +#ifdef unix + munmap (pool->map, (mgr->poolmask+1) << mgr->page_bits); +#else + { + FlushViewOfFile(pool->map, 0); + UnmapViewOfFile(pool->map); + CloseHandle(pool->hmap); + } +#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 (mgr); +#else + FlushFileBuffers(mgr->idx); + CloseHandle(mgr->idx); + GlobalFree (mgr->pool); + GlobalFree (mgr->hash); + GlobalFree (mgr->latch); + GlobalFree (mgr); +#endif +} + +// close and release memory + +void bt_close (BtDb *bt) +{ +#ifdef unix + if ( bt->mem ) + free (bt->mem); +#else + if ( bt->mem) + VirtualFree (bt->mem, 0, MEM_RELEASE); +#endif + free (bt); +} + +// open/create new btree buffer manager + +// call with file_name, BT_openmode, bits in page size (e.g. 16), +// size of mapped page pool (e.g. 8192) + +BtMgr *bt_mgr (char *name, uint mode, uint bits, uint poolmax, uint segsize, uint hashsize) +{ +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]; +#endif + + // determine sanity of page size and buffer pool + + if( bits > BT_maxbits ) + bits = BT_maxbits; + else if( bits < BT_minbits ) + bits = BT_minbits; + + if( !poolmax ) + return NULL; // must have buffer pool + +#ifdef unix + mgr = calloc (1, sizeof(BtMgr)); + + mgr->idx = open ((char*)name, O_RDWR | O_CREAT, 0666); + + if( mgr->idx == -1 ) + return free(mgr), NULL; + + cacheblk = 4096; // minimum mmap segment size for unix + +#else + mgr = GlobalAlloc (GMEM_FIXED|GMEM_ZEROINIT, sizeof(BtMgr)); + attr = FILE_ATTRIBUTE_NORMAL; + mgr->idx = CreateFile(name, GENERIC_READ| GENERIC_WRITE, FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, OPEN_ALWAYS, attr, NULL); + + if( mgr->idx == INVALID_HANDLE_VALUE ) + return GlobalFree(mgr), NULL; + + // normalize cacheblk to multiple of sysinfo->dwAllocationGranularity + GetSystemInfo(sysinfo); + cacheblk = sysinfo->dwAllocationGranularity; +#endif + +#ifdef unix + 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, latchmgr, BT_minpage, 0) == BT_minpage ) + bits = latchmgr->alloc->bits; + else + return free(mgr), free(latchmgr), NULL; + } else if( mode == BT_ro ) + return free(latchmgr), free (mgr), NULL; +#else + latchmgr = VirtualAlloc(NULL, BT_maxpage, MEM_COMMIT, PAGE_READWRITE); + size = GetFileSize(mgr->idx, amt); + + if( size || *amt ) { + if( !ReadFile(mgr->idx, (char *)latchmgr, BT_minpage, amt, NULL) ) + return bt_mgrclose (mgr), NULL; + bits = latchmgr->alloc->bits; + } else if( mode == BT_ro ) + return bt_mgrclose (mgr), NULL; +#endif + + mgr->page_size = 1 << bits; + mgr->page_bits = bits; + + mgr->poolmax = poolmax; + mgr->mode = mode; + + if( cacheblk < mgr->page_size ) + cacheblk = mgr->page_size; + + // mask for partial memmaps + + mgr->poolmask = (cacheblk >> bits) - 1; + + // see if requested size of pages per memmap is greater + + if( (1 << segsize) > mgr->poolmask ) + mgr->poolmask = (1 << segsize) - 1; + + mgr->seg_bits = 0; + + while( (1 << mgr->seg_bits) <= mgr->poolmask ) + mgr->seg_bits++; + + mgr->hashsize = hashsize; + +#ifdef unix + mgr->pool = calloc (poolmax, sizeof(BtPool)); + mgr->hash = calloc (hashsize, sizeof(ushort)); + mgr->latch = calloc (hashsize, sizeof(BtSpinLatch)); + 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)); +#endif + + if( size || *amt ) + 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)); + + // initialize latch manager + + 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, latchmgr, mgr->page_size) < mgr->page_size ) + return free(latchmgr), bt_mgrclose (mgr), NULL; +#else + 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 (latchmgr, 0, 1 << bits); + latchmgr->alloc->bits = mgr->page_bits; + + for( lvl=MIN_lvl; lvl--; ) { + 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; + 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, latchmgr, mgr->page_size) < mgr->page_size ) + return bt_mgrclose (mgr), NULL; +#else + 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 + } + + // clear out latch manager locks + // and rest of pages to round out segment + + memset(latchmgr, 0, mgr->page_size); + last = MIN_lvl + 1; + + while( last <= ((MIN_lvl + 1 + nlatchpage) | mgr->poolmask) ) { +#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 + last++; + } + +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 (latchmgr); +#else + VirtualFree (latchmgr, 0, MEM_RELEASE); +#endif + return mgr; +} + +// open BTree access method +// based on buffer manager + +BtDb *bt_open (BtMgr *mgr) +{ +BtDb *bt = malloc (sizeof(*bt)); + + memset (bt, 0, sizeof(*bt)); + bt->mgr = mgr; +#ifdef unix + bt->mem = malloc (3 *mgr->page_size); +#else + bt->mem = VirtualAlloc(NULL, 3 * mgr->page_size, MEM_COMMIT, PAGE_READWRITE); +#endif + bt->frame = (BtPage)bt->mem; + 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); + return bt; +} + +// compare two keys, returning > 0, = 0, or < 0 +// as the comparison value + +int keycmp (BtKey key1, unsigned char *key2, uint len2) +{ +uint len1 = key1->len; +int ans; + + if( ans = memcmp (key1->key, key2, len1 > len2 ? len2 : len1) ) + return ans; + + if( len1 > len2 ) + return 1; + if( len1 < len2 ) + return -1; + + return 0; +} + +// Latch Manager + +void bt_readlock(BtLatch *latch) +{ +#ifdef unix + pthread_rwlock_rdlock (latch->lock); +#else + AcquireSRWLockShared (latch->srw); +#endif +} + +// wait for other read and write latches to relinquish + +void bt_writelock(BtLatch *latch) +{ +#ifdef unix + pthread_rwlock_wrlock (latch->lock); +#else + AcquireSRWLockExclusive (latch->srw); +#endif +} + +// try to obtain write lock + +// return 1 if obtained, +// 0 if already write or read locked + +int bt_writetry(BtLatch *latch) +{ +int result = 0; + +#ifdef unix + result = !pthread_rwlock_trywrlock (latch->lock); +#else + result = TryAcquireSRWLockExclusive (latch->srw); +#endif + return result; +} + +// clear write mode + +void bt_releasewrite(BtLatch *latch) +{ +#ifdef unix + pthread_rwlock_unlock (latch->lock); +#else + ReleaseSRWLockExclusive (latch->srw); +#endif +} + +// decrement reader count + +void bt_releaseread(BtLatch *latch) +{ +#ifdef unix + pthread_rwlock_unlock (latch->lock); +#else + ReleaseSRWLockShared (latch->srw); +#endif +} + +// Buffer Pool mgr + +// find segment in pool +// must be called with hashslot idx locked +// return NULL if not there +// otherwise return node + +BtPool *bt_findpool(BtDb *bt, uid page_no, uint idx) +{ +BtPool *pool; +uint slot; + + // compute start of hash chain in pool + + if( slot = bt->mgr->hash[idx] ) + pool = bt->mgr->pool + slot; + else + return NULL; + + page_no &= ~bt->mgr->poolmask; + + while( pool->basepage != page_no ) + if( pool = pool->hashnext ) + continue; + else + return NULL; + + return pool; +} + +// add segment to hash table + +void bt_linkhash(BtDb *bt, BtPool *pool, uid page_no, int idx) +{ +BtPool *node; +uint slot; + + pool->hashprev = pool->hashnext = NULL; + pool->basepage = page_no & ~bt->mgr->poolmask; + pool->lru = 1; + + if( slot = bt->mgr->hash[idx] ) { + node = bt->mgr->pool + slot; + pool->hashnext = node; + node->hashprev = pool; + } + + 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) +{ +off64_t off = (page_no & ~bt->mgr->poolmask) << bt->mgr->page_bits; +off64_t limit = off + ((bt->mgr->poolmask+1) << bt->mgr->page_bits); +int flag; + +#ifdef unix + flag = PROT_READ | ( bt->mgr->mode == BT_ro ? 0 : PROT_WRITE ); + 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 + flag = ( bt->mgr->mode == BT_ro ? PAGE_READONLY : PAGE_READWRITE ); + pool->hmap = CreateFileMapping(bt->mgr->idx, NULL, flag, (DWORD)(limit >> 32), (DWORD)limit, NULL); + if( !pool->hmap ) + return bt->err = BTERR_map; + + flag = ( bt->mgr->mode == BT_ro ? FILE_MAP_READ : FILE_MAP_WRITE ); + pool->map = MapViewOfFile(pool->hmap, flag, (DWORD)(off >> 32), (DWORD)off, (bt->mgr->poolmask+1) << bt->mgr->page_bits); + if( !pool->map ) + return bt->err = BTERR_map; +#endif + return bt->err = 0; +} + +// calculate page within pool + +BtPage bt_page (BtDb *bt, BtPool *pool, uid page_no) +{ +uint subpage = (uint)(page_no & bt->mgr->poolmask); // page within mapping +BtPage page; + + 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 + return page; +} + +// release pool pin + +void bt_unpinpool (BtPool *pool) +{ +#ifdef unix + __sync_fetch_and_add(&pool->pin, -1); +#else + _InterlockedDecrement16 (&pool->pin); +#endif +} + +// find or place requested page in segment-pool +// return pool table entry, incrementing pin + +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]); + + // look up in hash table + + if( pool = bt_findpool(bt, page_no, idx) ) { +#ifdef unix + __sync_fetch_and_add(&pool->pin, 1); +#else + _InterlockedIncrement16 (&pool->pin); +#endif + bt_spinreleaseread (&bt->mgr->latch[idx]); + pool->lru++; + return pool; + } + + // upgrade to write lock + + bt_spinreleaseread (&bt->mgr->latch[idx]); + bt_spinwritelock (&bt->mgr->latch[idx]); + + // try to find page in pool with write lock + + if( pool = bt_findpool(bt, page_no, idx) ) { +#ifdef unix + __sync_fetch_and_add(&pool->pin, 1); +#else + _InterlockedIncrement16 (&pool->pin); +#endif + bt_spinreleasewrite (&bt->mgr->latch[idx]); + pool->lru++; + return pool; + } + + // allocate a new pool node + // and add to hash table + +#ifdef unix + slot = __sync_fetch_and_add(&bt->mgr->poolcnt, 1); +#else + slot = _InterlockedIncrement16 (&bt->mgr->poolcnt) - 1; +#endif + + if( ++slot < bt->mgr->poolmax ) { + pool = bt->mgr->pool + slot; + pool->slot = slot; + + 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 + _InterlockedIncrement16 (&pool->pin); +#endif + bt_spinreleasewrite (&bt->mgr->latch[idx]); + return pool; + } + + // pool table is full + // find best pool entry to evict + +#ifdef unix + __sync_fetch_and_add(&bt->mgr->poolcnt, -1); +#else + _InterlockedDecrement16 (&bt->mgr->poolcnt); +#endif + + while( 1 ) { +#ifdef unix + victim = __sync_fetch_and_add(&bt->mgr->evicted, 1); +#else + victim = _InterlockedIncrement16 (&bt->mgr->evicted) - 1; +#endif + victim %= bt->mgr->hashsize; + + // try to get write lock + // skip entry if not obtained + + if( !bt_spinwritetry (&bt->mgr->latch[victim]) ) + continue; + + // if cache entry is empty + // or no slots are unpinned + // skip this entry + + if( !(pool = bt_findlru(bt, bt->mgr->hash[victim])) ) { + bt_spinreleasewrite (&bt->mgr->latch[victim]); + continue; + } + + // unlink victim pool node from hash table + + if( node = pool->hashprev ) + node->hashnext = pool->hashnext; + else if( node = pool->hashnext ) + bt->mgr->hash[victim] = node->slot; + else + bt->mgr->hash[victim] = 0; + + if( node = pool->hashnext ) + node->hashprev = pool->hashprev; + + bt_spinreleasewrite (&bt->mgr->latch[victim]); + + // remove old file mapping +#ifdef unix + munmap (pool->map, (bt->mgr->poolmask+1) << bt->mgr->page_bits); +#else + FlushViewOfFile(pool->map, 0); + UnmapViewOfFile(pool->map); + CloseHandle(pool->hmap); +#endif + pool->map = NULL; + + // create new pool mapping + // and link into hash table + + 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 + _InterlockedIncrement16 (&pool->pin); +#endif + bt_spinreleasewrite (&bt->mgr->latch[idx]); + return pool; + } +} + +// place write, read, or parent lock on requested page_no. +// pin to buffer pool and return latchset pointer + +void bt_lockpage(BtLock mode, BtLatchSet *set) +{ + switch( mode ) { + case BtLockRead: + bt_readlock (set->readwr); + break; + case BtLockWrite: + bt_writelock (set->readwr); + break; + case BtLockAccess: + bt_readlock (set->access); + break; + case BtLockDelete: + bt_writelock (set->access); + break; + case BtLockParent: + bt_writelock (set->parent); + break; + } +} + +// remove write, read, or parent lock on requested page_no. + +void bt_unlockpage(BtLock mode, BtLatchSet *set) +{ + switch( mode ) { + case BtLockRead: + bt_releaseread (set->readwr); + break; + case BtLockWrite: + bt_releasewrite (set->readwr); + break; + case BtLockAccess: + bt_releaseread (set->access); + break; + case BtLockDelete: + bt_releasewrite (set->access); + break; + case BtLockParent: + bt_releasewrite (set->parent); + break; + } +} + +// allocate a new page and write page into it + +uid bt_newpage(BtDb *bt, BtPage page) +{ +BtLatchSet *set; +BtPool *pool; +uid new_page; +BtPage pmap; +int reuse; + + // lock allocation page + + bt_spinwritelock(bt->mgr->latchmgr->lock); + + // use empty chain first + // else allocate empty page + + if( new_page = bt_getid(bt->mgr->latchmgr->alloc[1].right) ) { + if( pool = bt_pinpool (bt, new_page) ) + pmap = bt_page (bt, pool, new_page); + else + return 0; + bt_putid(bt->mgr->latchmgr->alloc[1].right, bt_getid(pmap->right)); + bt_unpinpool (pool); + reuse = 1; + } else { + 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, 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; + } + + // 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); + else + return 0; + + memcpy(pmap, page, bt->mgr->page_size); + bt_unpinpool (pool); +#endif + return new_page; +} + +// find slot in page for given key at a given level + +int bt_findslot (BtDb *bt, unsigned char *key, uint len) +{ +uint diff, higher = bt->page->cnt, low = 1, slot; + + // low is the lowest candidate, higher is already + // tested as .ge. the given key, loop ends when they meet + + while( diff = higher - low ) { + slot = low + ( diff >> 1 ); + if( keycmp (keyptr(bt->page, slot), key, len) < 0 ) + low = slot + 1; + else + higher = slot; + } + + return higher; +} + +// 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) +{ +uid page_no = ROOT_page, prevpage = 0; +BtLatchSet *set, *prevset; +uint drill = 0xff, slot; +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; + + // obtain latch set for this page + + bt->set = bt_pinlatch (bt, page_no); + bt->page_no = page_no; + + // pin page contents + + if( bt->pool = bt_pinpool (bt, page_no) ) + bt->page = bt_page (bt, bt->pool, page_no); + else + return 0; + + // obtain access lock using lock chaining with Access mode + + if( page_no > ROOT_page ) + bt_lockpage(BtLockAccess, bt->set); + + // now unlock and unpin our (possibly foster) parent + + if( prevpage ) { + bt_unlockpage(prevmode, prevset); + bt_unpinlatch (prevset); + bt_unpinpool (prevpool); + prevpage = 0; + } + + // obtain read lock using lock chaining + + bt_lockpage(mode, bt->set); + + if( page_no > ROOT_page ) + bt_unlockpage(BtLockAccess, bt->set); + + // re-read and re-lock root after determining actual level of root + + if( page_no == ROOT_page ) + if( bt->page->lvl != drill) { + drill = bt->page->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); + + // 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( 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 + + // is this slot < foster child area + // if so, drill to next level + + if( slot <= bt->page->cnt - bt->page->foster ) + drill--; + + // continue right onto foster child + // or down to next level. + + page_no = bt_getid(slotptr(bt->page, slot)->id); + + } while( page_no ); + + // return error on end of 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 + +BTERR bt_deletekey (BtDb *bt, unsigned char *key, uint len) +{ +unsigned char leftkey[256]; +BtLatchSet *rset, *set; +BtPool *pool, *rpool; +BtPage rpage, page; +uid page_no, right; +uint slot, tod; +BtKey ptr; + + if( slot = bt_loadpage (bt, key, len, 0, 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; + if( slot < bt->page->cnt ) + bt->page->dirty = 1; + bt->page->act--; + } + + 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 ) { + 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); + + // 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); + + // 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; + } + + // 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); + + if( rpool = bt_pinpool (bt, right) ) + rpage = bt_page (bt, rpool, right); + else + return bt->err; + + 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); + + // delete left parent slot for old empty page + // and redirect right parent slot to it + + bt->page->act--; + bt->page->dirty = 1; + slotptr(bt->page, slot)->dead = 1; + + while( slot++ < bt->page->cnt ) + if( !slotptr(bt->page, slot)->dead ) + break; + + bt_putid(slotptr(bt->page,slot)->id, page_no); + + // release parent level lock + // and our empty node lock + + bt_unlockpage(BtLockWrite, set); + bt_unlockpage(BtLockWrite, bt->set); + bt_unpinlatch (bt->set); + bt_unpinpool (bt->pool); + + // add killed right block to free chain + // lock latch mgr + + bt_spinwritelock(bt->mgr->latchmgr->lock); + + // store free chain in allocation page second right + bt_putid(rpage->right, bt_getid(bt->mgr->latchmgr->alloc[1].right)); + bt_putid(bt->mgr->latchmgr->alloc[1].right, right); + + // unlock latch mgr and right page + + bt_spinreleasewrite(bt->mgr->latchmgr->lock); + + bt_unlockpage(BtLockWrite, rset); + bt_unlockpage(BtLockDelete, rset); + bt_unpinlatch (rset); + bt_unpinpool (rpool); + + // 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 + +uid bt_findkey (BtDb *bt, unsigned char *key, uint len) +{ +uint slot; +BtKey ptr; +uid id; + + if( slot = bt_loadpage (bt, key, len, 0, BtLockRead) ) + ptr = keyptr(bt->page, slot); + else + return 0; + + // if key exists, return row-id + // otherwise return 0 + + if( slot <= bt->page->cnt && !keycmp (ptr, key, len) ) + id = bt_getid(slotptr(bt->page,slot)->id); + else + id = 0; + + bt_unlockpage (BtLockRead, bt->set); + bt_unpinlatch (bt->set); + bt_unpinpool (bt->pool); + return id; +} + +// check page for space available, +// clean if necessary and return +// 0 - page needs splitting +// >0 new slot value + +uint bt_cleanpage(BtDb *bt, 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; +BtKey key; + + if( page->min >= (max+1) * sizeof(BtSlot) + sizeof(*page) + amt + 1 ) + return slot; + + // skip cleanup if nothing to reclaim + + if( !page->dirty ) + return 0; + + memcpy (bt->frame, page, bt->mgr->page_size); + + // skip page info and set rest of page to zero + + memset (page+1, 0, bt->mgr->page_size - sizeof(*page)); + 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 + + while( cnt++ < max ) { + if( cnt == slot ) + newslot = idx + 1; + else if( cnt < max && slotptr(bt->frame,cnt)->dead ) + continue; + + // copy key + + 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; + + // 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) +{ +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 + + 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 + + memset (root+1, 0, bt->mgr->page_size - sizeof(*root)); + + // insert left fence key on empty newroot page + + nxt -= *fencekey + 1; + memcpy ((unsigned char *)root + nxt, fencekey, *fencekey + 1); + bt_putid(slotptr(root, 1)->id, new_page); + slotptr(root, 1)->off = nxt; + + // insert stopper 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); + slotptr(root, 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++; + + // release and unpin root (bt->page) + + bt_unlockpage(BtLockWrite, bt->set); + bt_unpinlatch (bt->set); + bt_unpinpool (bt->pool); + return 0; +} + +// split already locked full node +// in current page variables +// return unlocked and unpinned. + +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; +BtPool *pool = bt->pool; +BtPage page = bt->page; +uint tod = time(NULL); +uint lvl = page->lvl; +uid new_page, right; +BtKey key; + + // initialize frame buffer for right node + + memset (bt->frame, 0, bt->mgr->page_size); + max = page->cnt - page->foster; + tod = (uint)time(NULL); + 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; + 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) ) + bt->frame->act++; + slotptr(bt->frame, idx)->tod = slotptr(page, cnt)->tod; + slotptr(bt->frame, idx)->off = nxt; + } + + // transfer right link node to new right node + + if( page_no > ROOT_page ) { + right = bt_getid (page->right); + bt_putid(bt->frame->right, right); + } + + 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. + + if( !(new_page = bt_newpage(bt, bt->frame)) ) + return bt->err; + + // remember fence key for new right page to add + // as foster child to the left node + + key = keyptr(bt->frame, idx); + memcpy (fencekey, key, key->len + 1); + + // update lower keys and foster children 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 + + 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); + slotptr(page, idx)->tod = slotptr(bt->frame, cnt)->tod; + slotptr(page, idx)->off = nxt; + page->act++; + } + + 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); + + // release wr lock on our page + + bt_unlockpage (BtLockWrite, set); + + // obtain ParentModification lock for current page + // to fix new fence key and oldest foster child on page + + bt_lockpage (BtLockParent, set); + + // 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_unlockpage (BtLockRead, set); + + if( bt_insertkey (bt, fencekey + 1, *fencekey, page_no, tod, lvl + 1) ) + return bt->err; + + // lock our page for writing + + bt_lockpage (BtLockRead, set); + + // switch old parent key from us to our oldest foster child + + 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) ) + 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--; + + // unlock and unpin + + bt_unlockpage (BtLockWrite, set); + bt_unlockpage (BtLockParent, set); + bt_unpinlatch (set); + bt_unpinpool (pool); + return 0; +} + +// Insert new key into the btree at leaf level. + +BTERR bt_insertkey (BtDb *bt, unsigned char *key, uint len, uid id, uint tod, uint lvl) +{ +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; + + 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; + } + + // check if page has enough space + + if( slot = bt_cleanpage (bt, len, slot) ) + break; + + if( bt_splitpage (bt) ) + return bt->err; + } + + bt_addkeytopage (bt, slot, key, len, id, tod); + + 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 + +uint bt_startkey (BtDb *bt, unsigned char *key, uint len) +{ +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); + return slot; +} + +// return next slot for cursor page +// or slide cursor right into next page + +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 ) + if( slotptr(bt->cursor,slot)->dead ) + continue; + else if( right || (slot < bt->cursor->cnt - bt->cursor->foster) ) + return slot; + else + break; + + if( !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); + + memcpy (bt->cursor, page, bt->mgr->page_size); + + bt_unlockpage(BtLockRead, set); + bt_unpinlatch (set); + bt_unpinpool (pool); + slot = 0; + } while( 1 ); + + return bt->err = 0; +} + +BtKey bt_key(BtDb *bt, uint slot) +{ + return keyptr(bt->cursor, slot); +} + +uid bt_uid(BtDb *bt, uint slot) +{ + return bt_getid(slotptr(bt->cursor,slot)->id); +} + +uint bt_tod(BtDb *bt, uint slot) +{ + return slotptr(bt->cursor,slot)->tod; +} + + +#ifdef STANDALONE + +typedef struct { + char type, idx; + char *infile; + BtMgr *mgr; + int num; +} ThreadArg; + +// standalone program to index file of keys +// then list them onto std-out + +#ifdef unix +void *index_file (void *arg) +#else +uint __stdcall index_file (void *arg) +#endif +{ +int line = 0, found = 0, cnt = 0; +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; +BtKey ptr; +BtDb *bt; +FILE *in; + + bt = bt_open (args->mgr); + time (tod); + + switch(args->type | 0x20) + { + case 'w': + fprintf(stderr, "started indexing for %s\n", args->infile); + if( in = fopen (args->infile, "rb") ) + while( ch = getc(in), ch != EOF ) + if( ch == '\n' ) + { + 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; + + if( bt_insertkey (bt, key, len, line, *tod, 0) ) + fprintf(stderr, "Error %d Line: %d\n", bt->err, line), exit(0); + len = 0; + } + else if( len < 255 ) + key[len++] = ch; + fprintf(stderr, "finished %s for %d keys\n", args->infile, line); + break; + + case 'd': + fprintf(stderr, "started deleting keys for %s\n", args->infile); + if( in = fopen (args->infile, "rb") ) + while( ch = getc(in), ch != EOF ) + if( ch == '\n' ) + { + 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; + + if( bt_deletekey (bt, key, len) ) + fprintf(stderr, "Error %d Line: %d\n", bt->err, line), exit(0); + len = 0; + } + else if( len < 255 ) + key[len++] = ch; + fprintf(stderr, "finished %s for keys, %d \n", args->infile, line); + break; + + case 'f': + fprintf(stderr, "started finding keys for %s\n", args->infile); + if( in = fopen (args->infile, "rb") ) + while( ch = getc(in), ch != EOF ) + if( ch == '\n' ) + { + 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; + + if( bt_findkey (bt, key, len) ) + found++; + else if( bt->err ) + fprintf(stderr, "Error %d Syserr %d Line: %d\n", bt->err, errno, line), exit(0); + len = 0; + } + else if( len < 255 ) + key[len++] = ch; + fprintf(stderr, "finished %s for %d keys, found %d\n", args->infile, line, found); + 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); + } + + break; + + case 'c': + fprintf(stderr, "started reading\n"); + + do { + if( pool = bt_pinpool (bt, page_no) ) + page = bt_page (bt, pool, page_no); + else + break; + set = bt_pinlatch (bt, page_no); + bt_lockpage (BtLockRead, set); + cnt += page->act; + next = bt_getid (page->right); + bt_unlockpage (BtLockRead, set); + bt_unpinlatch (set); + bt_unpinpool (pool); + } while( page_no = next ); + + cnt--; // remove stopper key + fprintf(stderr, " Total keys read %d\n", cnt); + break; + } + + bt_close (bt); +#ifdef unix + return NULL; +#else + return 0; +#endif +} + +typedef struct timeval timer; + +int main (int argc, char **argv) +{ +int idx, cnt, len, slot, err; +int segsize, bits = 16; +#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; +int num = 0; +char key[1]; +BtMgr *mgr; +BtKey ptr; +BtDb *bt; + + if( argc < 3 ) { + fprintf (stderr, "Usage: %s idx_file Read/Write/Scan/Delete/Find [page_bits mapped_segments seg_bits line_numbers src_file1 src_file2 ... ]\n", argv[0]); + 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 = 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 + + if( argc > 3 ) + bits = atoi(argv[3]); + + if( argc > 4 ) + poolsize = atoi(argv[4]); + + if( !poolsize ) + fprintf (stderr, "Warning: no mapped_pool\n"); + + if( poolsize > 65535 ) + fprintf (stderr, "Warning: mapped_pool > 65535 segments\n"); + + if( argc > 5 ) + segsize = atoi(argv[5]); + else + segsize = 4; // 16 pages per mmap segment + + if( argc > 6 ) + num = atoi(argv[6]); + + cnt = argc - 7; +#ifdef unix + threads = malloc (cnt * sizeof(pthread_t)); +#else + threads = GlobalAlloc (GMEM_FIXED|GMEM_ZEROINIT, cnt * sizeof(HANDLE)); +#endif + args = malloc (cnt * sizeof(ThreadArg)); + + mgr = bt_mgr ((argv[1]), BT_rw, bits, poolsize, segsize, poolsize / 8); + + if( !mgr ) { + fprintf(stderr, "Index Open Error %s\n", argv[1]); + exit (1); + } + + // fire off threads + + for( idx = 0; idx < cnt; idx++ ) { + args[idx].infile = argv[idx + 7]; + args[idx].type = argv[2][0]; + args[idx].mgr = mgr; + args[idx].num = num; + args[idx].idx = idx; +#ifdef unix + if( err = pthread_create (threads + idx, NULL, index_file, args + idx) ) + fprintf(stderr, "Error creating thread %d\n", err); +#else + threads[idx] = (HANDLE)_beginthreadex(NULL, 65536, index_file, args + idx, 0, NULL); +#endif + } + + // wait for termination + +#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); + bt_mgrclose (mgr); +} + +#endif //STANDALONE diff --git a/fosterbtreef.c b/fosterbtreef.c index 2b3eecb..e4592d2 100644 --- a/fosterbtreef.c +++ b/fosterbtreef.c @@ -1,5 +1,5 @@ -// foster btree version f -// 30 DEC 2013 +// foster btree version f2 +// 18 JAN 2014 // author: karl malbrain, malbrain@cal.berkeley.edu @@ -121,18 +121,17 @@ typedef struct { // by the BtSlot array of keys. typedef struct Page { - uint cnt; // count of keys in page - uint act; // count of active keys - uint min; // next key offset - uint foster; // count of foster children + 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 unsigned char bits; // page size in bits - unsigned char lvl:6; // level of page - unsigned char kill:1; // page is being deleted + unsigned char lvl:7; // level of page unsigned char dirty:1; // page needs to be cleaned unsigned char right[BtId]; // page number to right } *BtPage; -// mode & definition for spin latch implementation +// mode & definition for hash latch implementation enum { Mutex = 1, @@ -159,10 +158,12 @@ typedef struct { 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 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 volatile ushort next; // next entry in hash table chain volatile ushort prev; // prev entry in hash table chain @@ -214,29 +215,31 @@ typedef struct { #endif ushort poolcnt; // highest page pool node in use ushort poolmax; // highest page pool node allocated - ushort poolmask; // total size of pages in mmap segment - 1 + 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 BtPool *pool; // memory pool page segments BtSpinLatch *latch; // latches for pool hash slots BtLatchMgr *latchmgr; // mapped latch page from allocation page - BtLatchSet *latchset; // first mapped latch set from latch pages + BtLatchSet *latchsets; // mapped latch set from latch pages #ifndef unix - HANDLE halloc; // allocation and latch table + 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 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 uid page_no; // current page number uid cursor_page; // current cursor page number + BtLatchSet *set; // current page latch set + BtPool *pool; // current page pool unsigned char *mem; // frame, cursor, page memory buffer + int found; // last delete was found int err; // last error } BtDb; @@ -254,8 +257,8 @@ 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); -extern BTERR bt_deletekey (BtDb *bt, unsigned char *key, uint len, uint lvl); +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 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); @@ -373,7 +376,6 @@ ushort prev; #else _InterlockedAnd16((ushort *)latch, ~Mutex); #endif - if( prev ) return; #ifdef unix @@ -387,8 +389,6 @@ ushort prev; void bt_spinwritelock(BtSpinLatch *latch) { -ushort prev; - do { #ifdef unix while( __sync_fetch_and_or((ushort *)latch, Mutex | Pending) & Mutex ) @@ -397,23 +397,22 @@ ushort prev; while( _InterlockedOr16((ushort *)latch, Mutex | Pending) & Mutex ) SwitchToThread(); #endif - if( prev = !(latch->share | latch->exclusive) ) + 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; + } #ifdef unix - __sync_fetch_and_and ((ushort *)latch, ~(Mutex | Pending)); -#else - _InterlockedAnd16((ushort *)latch, ~(Mutex | Pending)); -#endif - if( prev ) - return; -#ifdef unix + __sync_fetch_and_and ((ushort *)latch, ~Mutex); sched_yield(); #else + _InterlockedAnd16((ushort *)latch, ~Mutex); SwitchToThread(); #endif } while( 1 ); @@ -478,10 +477,10 @@ void bt_spinreleaseread(BtSpinLatch *latch) void bt_latchlink (BtDb *bt, ushort hashidx, ushort victim, uid page_no) { -BtLatchSet *set = bt->mgr->latchset + victim; +BtLatchSet *set = bt->mgr->latchsets + victim; if( set->next = bt->mgr->latchmgr->table[hashidx].slot ) - bt->mgr->latchset[set->next].prev = victim; + bt->mgr->latchsets[set->next].prev = victim; bt->mgr->latchmgr->table[hashidx].slot = victim; set->page_no = page_no; @@ -489,10 +488,21 @@ BtLatchSet *set = bt->mgr->latchset + victim; 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_bindlatch (BtDb *bt, uid page_no, int incr) +BtLatchSet *bt_pinlatch (BtDb *bt, uid page_no) { ushort hashidx = page_no % bt->mgr->latchmgr->latchhash; ushort slot, avail = 0, victim, idx; @@ -504,12 +514,12 @@ BtLatchSet *set; if( slot = bt->mgr->latchmgr->table[hashidx].slot ) do { - set = bt->mgr->latchset + slot; + set = bt->mgr->latchsets + slot; if( page_no == set->page_no ) break; } while( slot = set->next ); - if( slot && incr ) { + if( slot ) { #ifdef unix __sync_fetch_and_add(&set->pin, 1); #else @@ -528,7 +538,7 @@ BtLatchSet *set; if( slot = bt->mgr->latchmgr->table[hashidx].slot ) do { - set = bt->mgr->latchset + slot; + set = bt->mgr->latchsets + slot; if( page_no == set->page_no ) break; if( !set->pin && !avail ) @@ -538,14 +548,12 @@ BtLatchSet *set; // found our entry, or take over an unpinned one if( slot || (slot = avail) ) { - set = bt->mgr->latchset + slot; - if( incr ) { + set = bt->mgr->latchsets + slot; #ifdef unix - __sync_fetch_and_add(&set->pin, 1); + __sync_fetch_and_add(&set->pin, 1); #else - _InterlockedIncrement16 (&set->pin); + _InterlockedIncrement16 (&set->pin); #endif - } set->page_no = page_no; bt_spinreleasewrite(bt->mgr->latchmgr->table[hashidx].latch); return set; @@ -559,7 +567,7 @@ BtLatchSet *set; #endif if( victim < bt->mgr->latchmgr->latchtotal ) { - set = bt->mgr->latchset + victim; + set = bt->mgr->latchsets + victim; #ifdef unix __sync_fetch_and_add(&set->pin, 1); #else @@ -586,7 +594,7 @@ BtLatchSet *set; // we don't use slot zero if( victim %= bt->mgr->latchmgr->latchtotal ) - set = bt->mgr->latchset + victim; + set = bt->mgr->latchsets + victim; else continue; @@ -616,18 +624,19 @@ BtLatchSet *set; // unlink our available victim from its hash chain if( set->prev ) - bt->mgr->latchset[set->prev].next = set->next; + bt->mgr->latchsets[set->prev].next = set->next; else bt->mgr->latchmgr->table[idx].slot = set->next; if( set->next ) - bt->mgr->latchset[set->next].prev = set->prev; + bt->mgr->latchsets[set->next].prev = set->prev; bt_spinreleasewrite (bt->mgr->latchmgr->table[idx].latch); - - if( incr ) - set->pin++; - +#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); @@ -657,6 +666,14 @@ 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); @@ -698,7 +715,6 @@ BtMgr *bt_mgr (char *name, uint mode, uint bits, uint poolmax, uint segsize, uin uint lvl, attr, cacheblk, last, slot, idx; uint nlatchpage, latchhash; BtLatchMgr *latchmgr; -int lockmode; off64_t size; uint amt[1]; BtMgr* mgr; @@ -722,19 +738,8 @@ 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; @@ -743,19 +748,8 @@ 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; @@ -776,7 +770,7 @@ SYSTEM_INFO sysinfo[1]; else return free(mgr), free(latchmgr), NULL; } else if( mode == BT_ro ) - return bt_mgrclose (mgr), NULL; + return free(latchmgr), bt_mgrclose (mgr), NULL; #else latchmgr = VirtualAlloc(NULL, BT_maxpage, MEM_COMMIT, PAGE_READWRITE); size = GetFileSize(mgr->idx, amt); @@ -908,25 +902,25 @@ SYSTEM_INFO sysinfo[1]; mgrlatch: #ifdef unix - flag = PROT_READ | ( mgr->mode == BT_ro ? 0 : PROT_WRITE ); + 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->latchset = (BtLatchSet *)mmap (0, mgr->latchmgr->nlatchpage * mgr->page_size, flag, MAP_SHARED, mgr->idx, LATCH_page * mgr->page_size); - if( mgr->latchset == MAP_FAILED ) + 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 = ( mgr->mode == BT_ro ? PAGE_READONLY : PAGE_READWRITE ); + 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 = ( mgr->mode == BT_ro ? FILE_MAP_READ : FILE_MAP_WRITE ); + 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->latchset = (void *)((char *)mgr->latchmgr + LATCH_page * mgr->page_size); + mgr->latchsets = (void *)((char *)mgr->latchmgr + LATCH_page * mgr->page_size); #endif #ifdef unix @@ -954,6 +948,8 @@ BtDb *bt = malloc (sizeof(*bt)); bt->frame = (BtPage)bt->mem; 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); return bt; } @@ -1081,10 +1077,43 @@ int flag; return bt->err = 0; } +// calculate page within pool + +BtPage bt_page (BtDb *bt, BtPool *pool, uid page_no) +{ +uint subpage = (uint)(page_no & bt->mgr->poolmask); // page within mapping +BtPage page; + + 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 + return page; +} + +// release pool pin + +void bt_unpinpool (BtPool *pool) +{ +#ifdef unix + __sync_fetch_and_add(&pool->pin, -1); +#else + _InterlockedDecrement16 (&pool->pin); +#endif +} + // find or place requested page in segment-pool // return pool table entry, incrementing pin -BtPool *bt_pinpage(BtDb *bt, uid page_no) +BtPool *bt_pinpool(BtDb *bt, uid page_no) { BtPool *pool, *node, *next; uint slot, idx, victim; @@ -1226,41 +1255,10 @@ BtLatchSet *set; } // place write, read, or parent lock on requested page_no. -// pin to buffer pool and return page pointer +// pin to buffer pool and return latchset pointer -BTERR bt_lockpage(BtDb *bt, uid page_no, BtLock mode, BtPage *pageptr) +void bt_lockpage(BtLock mode, BtLatchSet *set) { -BtLatchSet *set; -BtPool *pool; -uint subpage; -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; - - if( !(set = bt_bindlatch (bt, page_no, 1)) ) - return bt->err; - - page = (BtPage)(pool->map + (subpage << bt->mgr->page_bits)); - -#ifdef unix - { - uint idx = subpage / 8; - uint bit = subpage % 8; - - if( mode == BtLockRead || mode == BtLockWrite ) - 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_spinreadlock (set->readwr); @@ -1277,39 +1275,13 @@ BtPage page; case BtLockParent: bt_spinwritelock (set->parent); break; - default: - return bt->err = BTERR_lock; } - - if( pageptr ) - *pageptr = page; - - return bt->err = 0; } // remove write, read, or parent lock on requested page_no. -BTERR bt_unlockpage(BtDb *bt, uid page_no, BtLock mode) +void bt_unlockpage(BtLock mode, BtLatchSet *set) { -BtLatchSet *set; -BtPool *pool; -uint idx; - - // since page is pinned - // it should still be in the buffer pool - // and is in no danger of being a victim for reuse - - if( !(set = bt_bindlatch (bt, page_no, 0)) ) - return bt->err = BTERR_latch; - - idx = (uint)(page_no >> bt->mgr->seg_bits) % bt->mgr->hashsize; - bt_spinreadlock (&bt->mgr->latch[idx]); - - if( !(pool = bt_findpool(bt, page_no, idx)) ) - return bt->err = BTERR_hash; - - bt_spinreleaseread (&bt->mgr->latch[idx]); - switch( mode ) { case BtLockRead: bt_spinreleaseread (set->readwr); @@ -1326,65 +1298,15 @@ uint idx; case BtLockParent: bt_spinreleasewrite (set->parent); break; - default: - return bt->err = BTERR_lock; } - -#ifdef unix - __sync_fetch_and_add(&pool->pin, -1); - __sync_fetch_and_add (&set->pin, -1); -#else - _InterlockedDecrement16 (&pool->pin); - _InterlockedDecrement16 (&set->pin); -#endif - return bt->err = 0; -} - -// deallocate a deleted page -// place on free chain out of allocator page -// fence key must already be removed from parent - -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 - - bt_spinwritelock(bt->mgr->latchmgr->lock); - - // store free chain in allocation page second right - bt_putid(bt->temp->right, bt_getid(bt->mgr->latchmgr->alloc[1].right)); - bt_putid(bt->mgr->latchmgr->alloc[1].right, page_no); - - // unlock page zero - - bt_spinreleasewrite(bt->mgr->latchmgr->lock); - - // remove write lock on deleted node - - if( bt_unlockpage(bt, page_no, BtLockWrite) ) - return bt->err; - - // remove delete lock on deleted node - - if( bt_unlockpage(bt, page_no, BtLockDelete) ) - return bt->err; - - return 0; } // allocate a new page and write page into it uid bt_newpage(BtDb *bt, BtPage page) { +BtLatchSet *set; +BtPool *pool; uid new_page; BtPage pmap; int reuse; @@ -1397,11 +1319,12 @@ int reuse; // else allocate empty page if( new_page = bt_getid(bt->mgr->latchmgr->alloc[1].right) ) { - if( bt_lockpage (bt, new_page, BtLockWrite, &bt->temp) ) - return 0; - bt_putid(bt->mgr->latchmgr->alloc[1].right, bt_getid(bt->temp->right)); - if( bt_unlockpage (bt, new_page, BtLockWrite) ) + if( pool = bt_pinpool (bt, new_page) ) + pmap = bt_page (bt, pool, new_page); + else return 0; + bt_putid(bt->mgr->latchmgr->alloc[1].right, bt_getid(pmap->right)); + bt_unpinpool (pool); reuse = 1; } else { new_page = bt_getid(bt->mgr->latchmgr->alloc->right); @@ -1409,33 +1332,39 @@ 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( bt_lockpage(bt, new_page, BtLockWrite, &pmap) ) + if( pool = bt_pinpool (bt, new_page) ) + pmap = bt_page (bt, pool, new_page); + else return 0; memcpy(pmap, page, bt->mgr->page_size); - - if( bt_unlockpage (bt, new_page, BtLockWrite) ) - return 0; + bt_unpinpool (pool); #endif - // unlock allocation latch and return new page no - - bt_spinreleasewrite(bt->mgr->latchmgr->lock); return new_page; } @@ -1462,11 +1391,13 @@ uint diff, higher = bt->page->cnt, low = 1, slot; // 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, unsigned char *key, uint len, uint lvl, BtLock lock) { uid page_no = ROOT_page, prevpage = 0; +BtLatchSet *set, *prevset; uint drill = 0xff, slot; uint mode, prevmode; +BtPool *prevpool; // start at root of btree and drill down @@ -1474,84 +1405,94 @@ uint mode, prevmode; // 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; + // pin page contents + + if( bt->pool = bt_pinpool (bt, page_no) ) + bt->page = bt_page (bt, bt->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, bt->set); - // now unlock our (possibly foster) parent + // now unlock and unpin our (possibly foster) parent - if( prevpage ) - if( bt_unlockpage(bt, prevpage, prevmode) ) - return 0; - else - prevpage = 0; + if( prevpage ) { + bt_unlockpage(prevmode, prevset); + bt_unpinlatch (prevset); + 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, bt->set); if( page_no > ROOT_page ) - if( bt_unlockpage(bt, page_no, BtLockAccess) ) - return 0; + bt_unlockpage(BtLockAccess, bt->set); // re-read and re-lock root after determining actual level of root - if( bt->page_no == ROOT_page ) + if( page_no == ROOT_page ) if( bt->page->lvl != drill) { drill = bt->page->lvl; - if( lock == BtLockWrite && drill == lvl ) - if( bt_unlockpage(bt, page_no, mode) ) - return 0; - else - continue; + 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; - // if page is being deleted, - // move back to preceeding page - - if( bt->page->kill ) { - page_no = bt_getid (bt->page->right); - continue; - } - // find key on page at this level - // and descend to requested level + // and either descend to requested level + // or return key slot slot = bt_findslot (bt, key, len); - // is this slot a foster child? + // 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( 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 - goto slideright; + return bt->err = BTERR_struct, 0; // last key shouldn't be deleted + + // is this slot < foster child area + // if so, drill to next level if( slot <= bt->page->cnt - bt->page->foster ) drill--; - // continue down / right using overlapping locks - // to protect pages being killed or split. + // continue right onto foster child + // or down to next level. page_no = bt_getid(slotptr(bt->page, slot)->id); - continue; - -slideright: - page_no = bt_getid(bt->page->right); } while( page_no ); @@ -1562,108 +1503,182 @@ slideright: } // find and delete key on page by marking delete flag bit -// when page becomes empty, delete it from the btree +// when leaf page becomes empty, delete it from the btree -BTERR bt_deletekey (BtDb *bt, unsigned char *key, uint len, uint lvl) +BTERR bt_deletekey (BtDb *bt, unsigned char *key, uint len) { -unsigned char leftkey[256], rightkey[256]; +unsigned char leftkey[256]; +BtLatchSet *rset, *set; +BtPool *pool, *rpool; +BtPage rpage, page; uid page_no, right; uint slot, tod; BtKey ptr; - if( slot = bt_loadpage (bt, key, len, lvl, BtLockWrite) ) + if( slot = bt_loadpage (bt, key, len, 0, 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( !keycmp (ptr, key, len) ) - if( slotptr(bt->page, slot)->dead == 0 ) { + if( bt->found = !keycmp (ptr, key, len) ) + if( bt->found = slotptr(bt->page, slot)->dead == 0 ) { slotptr(bt->page,slot)->dead = 1; if( slot < bt->page->cnt ) 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; - if( !right || bt->page->act ) - return bt_unlockpage(bt, page_no, BtLockWrite); + // return if page is not empty or not found - // obtain Parent lock over write lock - - if( bt_lockpage(bt, page_no, BtLockParent, NULL) ) + if( page->act || !bt->found ) { + bt_unlockpage(BtLockWrite, set); + bt_unpinlatch (set); + bt_unpinpool (pool); return bt->err; + } - // cache copy of key to delete + // cache copy of fence key of empty node - ptr = keyptr(bt->page, bt->page->cnt); + ptr = keyptr(page, page->cnt); memcpy(leftkey, ptr, ptr->len + 1); - // lock and map right page + // release write lock on empty node + // obtain Parent lock - if( bt_lockpage(bt, right, BtLockWrite, &bt->temp) ) - return bt->err; + bt_unlockpage(BtLockWrite, set); + bt_lockpage(BtLockParent, set); - // pull contents of next page into current empty page - memcpy (bt->page, bt->temp, bt->mgr->page_size); + // load and lock parent to see + // if delete of empty node is OK + // ie, not a fence key of parent - // cache copy of key to update - ptr = keyptr(bt->temp, bt->temp->cnt); - memcpy(rightkey, ptr, ptr->len + 1); + while( 1 ) { + if( slot = bt_loadpage (bt, leftkey+1, *leftkey, 1, BtLockWrite) ) + ptr = keyptr(bt->page, slot); + else + return bt->err; - // Mark right page as deleted and point it to left page - // until we can post updates at higher level. + // does parent level contain our fence key yet? + // and is it free of foster children? - bt_putid(bt->temp->right, page_no); - bt->temp->kill = 1; - bt->temp->cnt = 0; + if( !bt->page->foster ) + if( !keycmp (ptr, leftkey+1, *leftkey) ) + break; - if( bt_unlockpage(bt, right, BtLockWrite) ) - return bt->err; - if( bt_unlockpage(bt, page_no, BtLockWrite) ) - return bt->err; + bt_unlockpage(BtLockWrite, bt->set); + bt_unpinlatch (bt->set); + bt_unpinpool (bt->pool); +#ifdef unix + sched_yield(); +#else + SwitchToThread(); +#endif + } - // delete old lower key to consolidated node + // find our left fence key - if( bt_deletekey (bt, leftkey + 1, *leftkey, lvl + 1) ) + 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); + + // 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; + } - // redirect higher key directly to consolidated node + // lock and map our right page + // note that it cannot be our foster child + // since the our node is empty - if( slot = bt_loadpage (bt, rightkey+1, *rightkey, lvl+1, BtLockWrite) ) - ptr = keyptr(bt->page, slot); + right = bt_getid(page->right); + + if( rpool = bt_pinpool (bt, right) ) + rpage = bt_page (bt, rpool, right); else return bt->err; - // since key already exists, update id + rset = bt_pinlatch (bt, right); + bt_lockpage(BtLockWrite, rset); + bt_lockpage(BtLockDelete, rset); + + // pull contents of right page into empty page - if( keycmp (ptr, rightkey+1, *rightkey) ) - return bt->err = BTERR_struct; + memcpy (page, rpage, bt->mgr->page_size); + + // delete left parent slot for old empty page + // and redirect right parent slot to it + + bt->page->act--; + bt->page->dirty = 1; + slotptr(bt->page, slot)->dead = 1; + + while( slot++ < bt->page->cnt ) + if( !slotptr(bt->page, slot)->dead ) + break; - slotptr(bt->page, slot)->dead = 0; bt_putid(slotptr(bt->page,slot)->id, page_no); - if( bt_unlockpage(bt, bt->page_no, BtLockWrite) ) - return bt->err; + // release parent level lock + // and our empty node lock - // obtain write lock and - // add right block to free chain + bt_unlockpage(BtLockWrite, set); + bt_unlockpage(BtLockWrite, bt->set); + bt_unpinlatch (bt->set); + bt_unpinpool (bt->pool); - if( bt_freepage (bt, right) ) - return bt->err; + // add killed right block to free chain + // lock latch mgr + + bt_spinwritelock(bt->mgr->latchmgr->lock); + + // store free chain in allocation page second right + bt_putid(rpage->right, bt_getid(bt->mgr->latchmgr->alloc[1].right)); + bt_putid(bt->mgr->latchmgr->alloc[1].right, right); + + // unlock latch mgr and right page + + bt_spinreleasewrite(bt->mgr->latchmgr->lock); + + bt_unlockpage(BtLockWrite, rset); + bt_unlockpage(BtLockDelete, rset); + bt_unpinlatch (rset); + bt_unpinpool (rpool); // remove ParentModify lock - if( bt_unlockpage(bt, page_no, BtLockParent) ) - return bt->err; - + 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 @@ -1681,32 +1696,33 @@ uid id; // if key exists, return row-id // otherwise return 0 - if( ptr->len == len && !memcmp (ptr->key, key, len) ) + if( slot <= bt->page->cnt && !keycmp (ptr, key, len) ) id = bt_getid(slotptr(bt->page,slot)->id); else id = 0; - if( bt_unlockpage (bt, bt->page_no, BtLockRead) ) - return 0; - + bt_unlockpage (BtLockRead, bt->set); + bt_unpinlatch (bt->set); + bt_unpinpool (bt->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, 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; BtKey key; if( page->min >= (max+1) * sizeof(BtSlot) + sizeof(*page) + amt + 1 ) - return 1; + return slot; // skip cleanup if nothing to reclaim @@ -1723,12 +1739,20 @@ BtKey key; // 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 + while( cnt++ < max ) { - // always leave fence key and foster children in list - if( cnt < max - page->foster && slotptr(bt->frame,cnt)->dead ) + if( cnt == slot ) + newslot = idx + 1; + else if( cnt < max && slotptr(bt->frame,cnt)->dead ) continue; // copy key + key = keyptr(bt->frame, cnt); nxt -= key->len + 1; memcpy ((unsigned char *)page + nxt, key, key->len + 1); @@ -1747,7 +1771,7 @@ BtKey key; // see if page has enough space now, or does it need splitting? if( page->min >= (idx+1) * sizeof(BtSlot) + sizeof(*page) + amt + 1 ) - return 1; + return newslot; return 0; } @@ -1760,27 +1784,29 @@ void bt_addkeytopage (BtDb *bt, uint slot, unsigned char *key, uint len, uid id, BtPage page = bt->page; uint idx; - // calculate next available slot and copy key into page + // find next available dead slot and copy key onto page + // note that foster children on the page are never dead - page->min -= len + 1; - ((unsigned char *)page)[page->min] = len; - memcpy ((unsigned char *)page + page->min +1, key, len ); + // look for next hole, but stay back from the fence key 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++; + // 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; @@ -1803,9 +1829,11 @@ BtKey key; // 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--; - root->foster--; // Save left fence key. @@ -1846,27 +1874,32 @@ BtKey key; root->act = 2; root->lvl++; - // release root (bt->page) + // release and unpin root (bt->page) - return bt_unlockpage(bt, ROOT_page, BtLockWrite); + bt_unlockpage(BtLockWrite, bt->set); + bt_unpinlatch (bt->set); + bt_unpinpool (bt->pool); + return 0; } // split already locked full node // in current page variables -// return unlocked. +// return unlocked and unpinned. 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; +BtPool *pool = bt->pool; BtPage page = bt->page; uint tod = time(NULL); uint lvl = page->lvl; uid new_page, right; BtKey key; - // initialize frame buffer + // initialize frame buffer for right node memset (bt->frame, 0, bt->mgr->page_size); max = page->cnt - page->foster; @@ -1875,19 +1908,21 @@ BtKey key; idx = 0; // split higher half of keys to bt->frame - // leaving foster children in the left node. + // 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; 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) ) + bt->frame->act++; slotptr(bt->frame, idx)->tod = slotptr(page, cnt)->tod; slotptr(bt->frame, idx)->off = nxt; - bt->frame->act++; } - // transfer right link node + // transfer right link node to new right node if( page_no > ROOT_page ) { right = bt_getid (page->right); @@ -1899,13 +1934,13 @@ BtKey key; bt->frame->cnt = idx; bt->frame->lvl = lvl; - // get new free page and write frame to it. + // get new free page and write right frame to it. if( !(new_page = bt_newpage(bt, bt->frame)) ) return bt->err; - // remember fence key for new page to add - // as foster child + // remember fence key for new right page to add + // as foster child to the left node key = keyptr(bt->frame, idx); memcpy (fencekey, key, key->len + 1); @@ -1915,6 +1950,7 @@ BtKey key; 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; @@ -1927,22 +1963,26 @@ BtKey key; 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; - page->act++; } - // insert new foster child at beginning of the current foster children + // 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 if any + // continue with old foster child keys + // note that none will be dead cnt = bt->frame->cnt - bt->frame->foster; @@ -1970,109 +2010,69 @@ BtKey key; // release wr lock on our page - if( bt_unlockpage (bt, page_no, BtLockWrite) ) - return bt->err; + bt_unlockpage (BtLockWrite, set); // obtain ParentModification lock for current page - // to fix fence key and highest foster child on page + // to fix new fence key and oldest foster child on page - if( bt_lockpage (bt, page_no, BtLockParent, NULL) ) - return bt->err; + bt_lockpage (BtLockParent, set); - // get our highest foster child key to find in parent node + // get our new fence key to insert in parent node - if( bt_lockpage (bt, page_no, BtLockRead, &page) ) - return bt->err; + bt_lockpage (BtLockRead, set); - key = keyptr(page, page->cnt); + key = keyptr(page, page->cnt-1); memcpy (fencekey, key, key->len+1); - if( bt_unlockpage (bt, page_no, BtLockRead) ) - return bt->err; - - // update our parent -try_again: - - do { - slot = bt_loadpage (bt, fencekey + 1, *fencekey, lvl + 1, BtLockWrite); - - if( !slot ) - return bt->err; - - // check if parent page has enough space for any possible key + bt_unlockpage (BtLockRead, set); - if( bt_cleanpage (bt, 256) ) - break; - - if( bt_splitpage (bt) ) + if( bt_insertkey (bt, fencekey + 1, *fencekey, page_no, tod, lvl + 1) ) return bt->err; - } while( 1 ); - // see if we are still a foster child from another node + // lock our page for writing - if( bt_getid (slotptr(bt->page, slot)->id) != page_no ) { - if( bt_unlockpage (bt, bt->page_no, BtLockWrite) ) - return bt->err; -#ifdef unix - sched_yield(); -#else - SwitchToThread(); -#endif - goto try_again; - } + bt_lockpage (BtLockRead, set); - // wait until readers from parent get their locks - // on our page + // switch old parent key from us to our oldest foster child - if( bt_lockpage (bt, page_no, BtLockDelete, NULL) ) - return bt->err; + key = keyptr(page, page->cnt); + memcpy (fencekey, key, key->len+1); - // lock our page for writing + new_page = bt_getid (slotptr(page, page->cnt)->id); + bt_unlockpage (BtLockRead, set); - if( bt_lockpage (bt, page_no, BtLockWrite, &page) ) + if( bt_insertkey (bt, fencekey + 1, *fencekey, new_page, tod, lvl + 1) ) return bt->err; - // switch parent fence key to foster child - - if( slotptr(page, page->cnt)->dead ) - slotptr(bt->page, slot)->dead = 1; - else - bt_putid (slotptr(bt->page, slot)->id, bt_getid(slotptr(page, page->cnt)->id)); - - // remove highest foster child from our page + // 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--; - page->foster--; - page->dirty = 1; - key = keyptr(page, page->cnt); - - // add our new fence key for foster child to our parent - - bt_addkeytopage (bt, slot, key->key, key->len, page_no, tod); - - if( bt_unlockpage (bt, bt->page_no, BtLockWrite) ) - return bt->err; - if( bt_unlockpage (bt, page_no, BtLockDelete) ) - return bt->err; + // unlock and unpin - if( bt_unlockpage (bt, page_no, BtLockWrite) ) - return bt->err; - - return bt_unlockpage (bt, page_no, BtLockParent); + bt_unlockpage (BtLockWrite, set); + bt_unlockpage (BtLockParent, set); + bt_unpinlatch (set); + bt_unpinpool (pool); + return 0; } // Insert new key into the btree at leaf level. -BTERR bt_insertkey (BtDb *bt, unsigned char *key, uint len, uid id, uint tod) +BTERR bt_insertkey (BtDb *bt, unsigned char *key, uint len, uid id, uint tod, uint lvl) { uint slot, idx; BtPage page; BtKey ptr; while( 1 ) { - if( slot = bt_loadpage (bt, key, len, 0, BtLockWrite) ) + if( slot = bt_loadpage (bt, key, len, lvl, BtLockWrite) ) ptr = keyptr(bt->page, slot); else { @@ -2086,15 +2086,20 @@ BtKey ptr; 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); - return bt_unlockpage(bt, bt->page_no, BtLockWrite); + bt_unlockpage(BtLockWrite, bt->set); + bt_unpinlatch (bt->set); + bt_unpinpool (bt->pool); + return bt->err; } // check if page has enough space - if( bt_cleanpage (bt, len) ) + if( slot = bt_cleanpage (bt, len, slot) ) break; if( bt_splitpage (bt) ) @@ -2103,7 +2108,10 @@ BtKey ptr; bt_addkeytopage (bt, slot, key, len, id, tod); - return bt_unlockpage (bt, bt->page_no, BtLockWrite); + 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 @@ -2115,10 +2123,12 @@ 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; + bt_unlockpage(BtLockRead, bt->set); + bt_unpinlatch (bt->set); + bt_unpinpool (bt->pool); return slot; } @@ -2127,6 +2137,8 @@ uint slot; uint bt_nextkey (BtDb *bt, uint slot) { +BtLatchSet *set; +BtPool *pool; BtPage page; uid right; @@ -2144,15 +2156,19 @@ uid right; break; bt->cursor_page = right; - - if( bt_lockpage(bt, right, BtLockRead, &page) ) + if( pool = bt_pinpool (bt, right) ) + page = bt_page (bt, pool, right); + else return 0; - memcpy (bt->cursor, page, bt->mgr->page_size); + set = bt_pinlatch (bt, right); + bt_lockpage(BtLockRead, set); - if ( bt_unlockpage(bt, right, BtLockRead) ) - return 0; + memcpy (bt->cursor, page, bt->mgr->page_size); + bt_unlockpage(BtLockRead, set); + bt_unpinlatch (set); + bt_unpinpool (pool); slot = 0; } while( 1 ); @@ -2177,6 +2193,54 @@ uint bt_tod(BtDb *bt, uint slot) #ifdef STANDALONE +void bt_latchaudit (BtDb *bt) +{ +ushort idx, hashidx; +BtLatchSet *set; +BtPool *pool; +BtPage page; +uid page_no; + +#ifdef unix + for( idx = 1; idx < bt->mgr->latchmgr->latchdeployed; idx++ ) { + set = bt->mgr->latchsets + idx; + if( *(ushort *)set->readwr || *(ushort *)set->access || *(ushort *)set->parent ) { + fprintf(stderr, "latchset %d locked for page %6x\n", idx, set->page_no); + *(ushort *)set->readwr = 0; + *(ushort *)set->access = 0; + *(ushort *)set->parent = 0; + } + if( set->pin ) { + fprintf(stderr, "latchset %d pinned\n", idx); + set->pin = 0; + } + } + + for( hashidx = 0; hashidx < bt->mgr->latchmgr->latchhash; hashidx++ ) { + if( *(ushort *)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 ) + fprintf(stderr, "latchset %d locked\n", idx); + if( set->hash != hashidx ) + fprintf(stderr, "latchset %d wrong hashidx\n", idx); + if( set->pin ) + fprintf(stderr, "latchset %d pinned\n", idx); + } while( idx = set->next ); + } + page_no = bt_getid(bt->mgr->latchmgr->alloc[1].right); + + while( page_no ) { + fprintf(stderr, "free: %.6x\n", (uint)page_no); + pool = bt_pinpool (bt, page_no); + page = bt_page (bt, pool, page_no); + page_no = bt_getid(page->right); + bt_unpinpool (pool); + } +#endif +} + typedef struct { char type, idx; char *infile; @@ -2198,7 +2262,9 @@ 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; BtKey ptr; BtDb *bt; @@ -2209,6 +2275,11 @@ 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") ) @@ -2223,7 +2294,7 @@ 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) ) + if( bt_insertkey (bt, key, len, line, *tod, 0) ) fprintf(stderr, "Error %d Line: %d\n", bt->err, line), exit(0); len = 0; } @@ -2245,7 +2316,7 @@ FILE *in; else if( args->num ) sprintf((char *)key+len, "%.9d", line + args->idx * args->num), len += 9; - if( bt_deletekey (bt, key, len, 0) ) + if( bt_deletekey (bt, key, len) ) fprintf(stderr, "Error %d Line: %d\n", bt->err, line), exit(0); len = 0; } @@ -2300,10 +2371,17 @@ FILE *in; fprintf(stderr, "started reading\n"); do { - bt_lockpage (bt, page_no, BtLockRead, &page); + if( pool = bt_pinpool (bt, page_no) ) + page = bt_page (bt, pool, page_no); + else + break; + set = bt_pinlatch (bt, page_no); + bt_lockpage (BtLockRead, set); cnt += page->act; next = bt_getid (page->right); - bt_unlockpage (bt, page_no, BtLockRead); + bt_unlockpage (BtLockRead, set); + bt_unpinlatch (set); + bt_unpinpool (pool); } while( page_no = next ); cnt--; // remove stopper key diff --git a/fosterbtreef1.c b/fosterbtreef1.c new file mode 100644 index 0000000..f2fea9d --- /dev/null +++ b/fosterbtreef1.c @@ -0,0 +1,2447 @@ +// foster btree version f +// 17 JAN 2014 + +// author: karl malbrain, malbrain@cal.berkeley.edu + +/* +This work, including the source code, documentation +and related data, is placed into the public domain. + +The orginal author is Karl Malbrain. + +THIS SOFTWARE IS PROVIDED AS-IS WITHOUT WARRANTY +OF ANY KIND, NOT EVEN THE IMPLIED WARRANTY OF +MERCHANTABILITY. THE AUTHOR OF THIS SOFTWARE, +ASSUMES _NO_ RESPONSIBILITY FOR ANY CONSEQUENCE +RESULTING FROM THE USE, MODIFICATION, OR +REDISTRIBUTION OF THIS SOFTWARE. +*/ + +// Please see the project home page for documentation +// code.google.com/p/high-concurrency-btree + +#define _FILE_OFFSET_BITS 64 +#define _LARGEFILE64_SOURCE + +#ifdef linux +#define _GNU_SOURCE +#endif + +#ifdef unix +#include +#include +#include +#include +#include +#include +#include +#include +#else +#define WIN32_LEAN_AND_MEAN +#include +#include +#include +#include +#include +#include +#include +#endif + +#include +#include + +typedef unsigned long long uid; + +#ifndef unix +typedef unsigned long long off64_t; +typedef unsigned short ushort; +typedef unsigned int uint; +#endif + +#define BT_ro 0x6f72 // ro +#define BT_rw 0x7772 // rw + +#define BT_latchtable 128 // number of latch manager slots + +#define BT_maxbits 24 // maximum page size in bits +#define BT_minbits 9 // minimum page size in bits +#define BT_minpage (1 << BT_minbits) // minimum page size +#define BT_maxpage (1 << BT_maxbits) // maximum page size + +/* +There are five lock types for each node in three independent sets: +1. (set 1) AccessIntent: Sharable. Going to Read the node. Incompatible with NodeDelete. +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. +*/ + +typedef enum{ + BtLockAccess, + BtLockDelete, + BtLockRead, + BtLockWrite, + BtLockParent, + BtLockPin +}BtLock; + +// Define the length of the page and key pointers + +#define BtId 6 + +// Page key slot definition. + +// If BT_maxbits is 15 or less, you can save 4 bytes +// for each key stored by making the first two uints +// into ushorts. You can also save 4 bytes by removing +// the tod field from the key. + +// Keys are marked dead, but remain on the page until +// it cleanup is called. The fence key (highest key) for +// the page is always present, even after cleanup. + +typedef struct { + uint off:BT_maxbits; // page offset for key start + uint dead:1; // set for deleted key + uint tod; // time-stamp for key + unsigned char id[BtId]; // id associated with key +} BtSlot; + +// The key structure occupies space at the upper end of +// each page. It's a length byte followed by the value +// bytes. + +typedef struct { + unsigned char len; + unsigned char key[1]; +} *BtKey; + +// The first part of an index page. +// It is immediately followed +// by the BtSlot array of keys. + +typedef struct Page { + uint cnt; // count of keys in page + uint act; // count of active keys + uint min; // next key offset + uint foster; // count of foster children + unsigned char bits; // page size in bits + unsigned char lvl:6; // level of page + unsigned char kill:1; // page is being deleted + unsigned char dirty:1; // page needs to be cleaned + 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]; + volatile ushort slot; // Latch table entry at head of chain +} BtHashEntry; + +// latch table lock structure +// implements a fair read-write lock + +typedef struct { +#ifdef unix + pthread_rwlock_t lock[1]; +#else + SRWLOCK srw[1]; +#endif +} BtLatch; + +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 + 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; + +// 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 + ushort pin; // mapped page pin counter + ushort slot; // slot index in this array + void *hashprev; // previous pool entry for the same hash idx + void *hashnext; // next pool entry for the same hash idx +#ifndef unix + HANDLE hmap; // Windows memory mapping handle +#endif +} BtPool; + +// structure for latch manager on ALLOC_page + +typedef struct { + struct Page 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 { + uint page_size; // page size + uint page_bits; // page size in bits + 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 +#else + HANDLE idx; +#endif + ushort poolcnt; // highest page pool node in use + ushort poolmax; // highest page pool node allocated + ushort poolmask; // total size 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 + BtPool *pool; // memory pool page segments + BtSpinLatch *latch; // latches for pool hash slots + BtLatchMgr *latchmgr; // mapped latch page from allocation page + BtLatchSet *latchsets; // mapped latch set from latch pages +#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 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 + uid page_no; // current page number + uid cursor_page; // current cursor page number + BtLatchSet *set; // current page latch set + unsigned char *mem; // frame, cursor, page memory buffer + int err; // last error +} BtDb; + +typedef enum { + BTERR_ok = 0, + BTERR_struct, + BTERR_ovflw, + BTERR_lock, + BTERR_map, + BTERR_wrt, + BTERR_hash, + BTERR_latch +} BTERR; + +// 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); +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); + +// manager functions +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 + +extern BtKey bt_key (BtDb *bt, uint slot); +extern uid bt_uid (BtDb *bt, uint slot); +extern uint bt_tod (BtDb *bt, uint slot); + +// BTree page number constants +#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 + +#define MIN_lvl 2 + +// The page is allocated from low and hi ends. +// The key offsets and row-id's are allocated +// from the bottom, while the text of the key +// is allocated from the top. When the two +// areas meet, the page is split into two. + +// 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. + +// The b-tree root is always located at page 1. +// The first leaf page of level zero is always +// located on page 2. + +// 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. + +// 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. + +// Groups of pages called segments from the btree are +// cached with memory mapping. 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. + +// Page 0 is dedicated to lock for new page extensions, +// and chains empty pages together for reuse. + +// Empty pages are chained together through the ALLOC page and reused. + +// Access macros to address slot and key values from the page + +#define slotptr(page, slot) (((BtSlot *)(page+1)) + (slot-1)) +#define keyptr(page, slot) ((BtKey)((unsigned char*)(page) + slotptr(page, slot)->off)) + +void bt_putid(unsigned char *dest, uid id) +{ +int i = BtId; + + while( i-- ) + dest[i] = (unsigned char)id, id >>= 8; +} + +uid bt_getid(unsigned char *src) +{ +uid id = 0; +int i; + + for( i = 0; i < BtId; i++ ) + id <<= 8, id |= *src++; + + return id; +} + +// wait until write lock mode is clear +// and add 1 to the share count + +void bt_spinreadlock(BtSpinLatch *latch) +{ +ushort prev; + + do { +#ifdef unix + while( __sync_fetch_and_or((ushort *)latch, Mutex) & Mutex ) + sched_yield(); +#else + while( _InterlockedOr16((ushort *)latch, Mutex) & Mutex ) + SwitchToThread(); +#endif + + // see if exclusive request is granted or pending + + if( prev = !(latch->exclusive | latch->pending) ) +#ifdef unix + __sync_fetch_and_add((ushort *)latch, Share); +#else + _InterlockedExchangeAdd16 ((ushort *)latch, Share); +#endif + +#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 +} + +// wait for other read and write latches to relinquish + +void bt_spinwritelock(BtSpinLatch *latch) +{ +ushort prev; + + do { +#ifdef unix + while( __sync_fetch_and_or((ushort *)latch, Mutex | Pending) & Mutex ) + sched_yield(); +#else + while( _InterlockedOr16((ushort *)latch, Mutex | Pending) & Mutex ) + SwitchToThread(); +#endif + if( prev = !(latch->share | latch->exclusive) ) +#ifdef unix + __sync_fetch_and_or((ushort *)latch, Write); +#else + _InterlockedOr16((ushort *)latch, Write); +#endif + +#ifdef unix + __sync_fetch_and_and ((ushort *)latch, ~(Mutex | Pending)); +#else + _InterlockedAnd16((ushort *)latch, ~(Mutex | Pending)); +#endif + if( prev ) + return; +#ifdef unix + sched_yield(); +#else + SwitchToThread(); +#endif + } while( 1 ); +} + +// try to obtain write lock + +// return 1 if obtained, +// 0 otherwise + +int bt_spinwritetry(BtSpinLatch *latch) +{ +ushort prev; + +#ifdef unix + if( prev = __sync_fetch_and_or((ushort *)latch, Mutex), prev & Mutex ) + return 0; +#else + if( prev = _InterlockedOr16((ushort *)latch, Mutex), prev & 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 + +#ifdef unix + __sync_fetch_and_and ((ushort *)latch, ~Mutex); +#else + _InterlockedAnd16((ushort *)latch, ~Mutex); +#endif + return !prev; +} + +// clear write mode + +void bt_spinreleasewrite(BtSpinLatch *latch) +{ +#ifdef unix + __sync_fetch_and_and ((ushort *)latch, ~Write); +#else + _InterlockedAnd16((ushort *)latch, ~Write); +#endif +} + +// decrement reader count + +void bt_spinreleaseread(BtSpinLatch *latch) +{ +#ifdef unix + __sync_fetch_and_add((ushort *)latch, -Share); +#else + _InterlockedExchangeAdd16 ((ushort *)latch, -Share); +#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; +} + +// find existing latchset or inspire new one +// return with latchset pinned + +BtLatchSet *bt_bindlatch (BtDb *bt, uid page_no, int incr) +{ +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 && incr ) { +#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; + if( incr ) +#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; + if( incr ) +#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); + + if( incr ) +#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; +uint slot; + + // release mapped pages + // note that slot zero is never used + + for( slot = 1; slot < mgr->poolmax; slot++ ) { + pool = mgr->pool + slot; + if( pool->slot ) +#ifdef unix + munmap (pool->map, (mgr->poolmask+1) << mgr->page_bits); +#else + { + FlushViewOfFile(pool->map, 0); + UnmapViewOfFile(pool->map); + CloseHandle(pool->hmap); + } +#endif + } + +#ifdef unix + close (mgr->idx); + free (mgr->pool); + free (mgr->hash); + free (mgr->latch); + free (mgr->pooladvise); + free (mgr); +#else + FlushFileBuffers(mgr->idx); + CloseHandle(mgr->idx); + GlobalFree (mgr->pool); + GlobalFree (mgr->hash); + GlobalFree (mgr->latch); + GlobalFree (mgr); +#endif +} + +// close and release memory + +void bt_close (BtDb *bt) +{ +#ifdef unix + if ( bt->mem ) + free (bt->mem); +#else + if ( bt->mem) + VirtualFree (bt->mem, 0, MEM_RELEASE); +#endif + free (bt); +} + +// open/create new btree buffer manager + +// call with file_name, BT_openmode, bits in page size (e.g. 16), +// size of mapped page pool (e.g. 8192) + +BtMgr *bt_mgr (char *name, uint mode, uint bits, uint poolmax, uint segsize, uint hashsize) +{ +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]; +#endif + + // determine sanity of page size and buffer pool + + if( bits > BT_maxbits ) + bits = BT_maxbits; + else if( bits < BT_minbits ) + bits = BT_minbits; + + if( !poolmax ) + return NULL; // must have buffer pool + +#ifdef unix + mgr = calloc (1, sizeof(BtMgr)); + + mgr->idx = open ((char*)name, O_RDWR | O_CREAT, 0666); + + if( mgr->idx == -1 ) + return free(mgr), NULL; + + cacheblk = 4096; // minimum mmap segment size for unix + +#else + mgr = GlobalAlloc (GMEM_FIXED|GMEM_ZEROINIT, sizeof(BtMgr)); + attr = FILE_ATTRIBUTE_NORMAL; + mgr->idx = CreateFile(name, GENERIC_READ| GENERIC_WRITE, FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, OPEN_ALWAYS, attr, NULL); + + if( mgr->idx == INVALID_HANDLE_VALUE ) + return GlobalFree(mgr), NULL; + + // normalize cacheblk to multiple of sysinfo->dwAllocationGranularity + GetSystemInfo(sysinfo); + cacheblk = sysinfo->dwAllocationGranularity; +#endif + +#ifdef unix + 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, latchmgr, BT_minpage, 0) == BT_minpage ) + bits = latchmgr->alloc->bits; + else + return free(mgr), free(latchmgr), NULL; + } else if( mode == BT_ro ) + return bt_mgrclose (mgr), NULL; +#else + latchmgr = VirtualAlloc(NULL, BT_maxpage, MEM_COMMIT, PAGE_READWRITE); + size = GetFileSize(mgr->idx, amt); + + if( size || *amt ) { + if( !ReadFile(mgr->idx, (char *)latchmgr, BT_minpage, amt, NULL) ) + return bt_mgrclose (mgr), NULL; + bits = latchmgr->alloc->bits; + } else if( mode == BT_ro ) + return bt_mgrclose (mgr), NULL; +#endif + + mgr->page_size = 1 << bits; + mgr->page_bits = bits; + + mgr->poolmax = poolmax; + mgr->mode = mode; + + if( cacheblk < mgr->page_size ) + cacheblk = mgr->page_size; + + // mask for partial memmaps + + mgr->poolmask = (cacheblk >> bits) - 1; + + // see if requested size of pages per memmap is greater + + if( (1 << segsize) > mgr->poolmask ) + mgr->poolmask = (1 << segsize) - 1; + + mgr->seg_bits = 0; + + while( (1 << mgr->seg_bits) <= mgr->poolmask ) + mgr->seg_bits++; + + mgr->hashsize = hashsize; + +#ifdef unix + mgr->pool = calloc (poolmax, sizeof(BtPool)); + mgr->hash = calloc (hashsize, sizeof(ushort)); + mgr->latch = calloc (hashsize, sizeof(BtSpinLatch)); + 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)); +#endif + + if( size || *amt ) + 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)); + + // initialize latch manager + + 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, latchmgr, mgr->page_size) < mgr->page_size ) + return bt_mgrclose (mgr), NULL; +#else + 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 (latchmgr, 0, 1 << bits); + latchmgr->alloc->bits = mgr->page_bits; + + for( lvl=MIN_lvl; lvl--; ) { + 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; + 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, latchmgr, mgr->page_size) < mgr->page_size ) + return bt_mgrclose (mgr), NULL; +#else + 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 + } + + // clear out latch manager locks + // and rest of pages to round out segment + + memset(latchmgr, 0, mgr->page_size); + last = MIN_lvl + 1; + + while( last <= ((MIN_lvl + 1 + nlatchpage) | mgr->poolmask) ) { +#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 + last++; + } + +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 (latchmgr); +#else + VirtualFree (latchmgr, 0, MEM_RELEASE); +#endif + return mgr; +} + +// open BTree access method +// based on buffer manager + +BtDb *bt_open (BtMgr *mgr) +{ +BtDb *bt = malloc (sizeof(*bt)); + + memset (bt, 0, sizeof(*bt)); + bt->mgr = mgr; +#ifdef unix + bt->mem = malloc (3 *mgr->page_size); +#else + bt->mem = VirtualAlloc(NULL, 3 * mgr->page_size, MEM_COMMIT, PAGE_READWRITE); +#endif + bt->frame = (BtPage)bt->mem; + bt->zero = (BtPage)(bt->mem + 1 * mgr->page_size); + bt->cursor = (BtPage)(bt->mem + 2 * mgr->page_size); + return bt; +} + +// compare two keys, returning > 0, = 0, or < 0 +// as the comparison value + +int keycmp (BtKey key1, unsigned char *key2, uint len2) +{ +uint len1 = key1->len; +int ans; + + if( ans = memcmp (key1->key, key2, len1 > len2 ? len2 : len1) ) + return ans; + + if( len1 > len2 ) + return 1; + if( len1 < len2 ) + return -1; + + return 0; +} + +// Buffer Pool mgr + +// find segment in pool +// must be called with hashslot idx locked +// return NULL if not there +// otherwise return node + +BtPool *bt_findpool(BtDb *bt, uid page_no, uint idx) +{ +BtPool *pool; +uint slot; + + // compute start of hash chain in pool + + if( slot = bt->mgr->hash[idx] ) + pool = bt->mgr->pool + slot; + else + return NULL; + + page_no &= ~bt->mgr->poolmask; + + while( pool->basepage != page_no ) + if( pool = pool->hashnext ) + continue; + else + return NULL; + + return pool; +} + +// add segment to hash table + +void bt_linkhash(BtDb *bt, BtPool *pool, uid page_no, int idx) +{ +BtPool *node; +uint slot; + + pool->hashprev = pool->hashnext = NULL; + pool->basepage = page_no & ~bt->mgr->poolmask; + pool->lru = 1; + + if( slot = bt->mgr->hash[idx] ) { + node = bt->mgr->pool + slot; + pool->hashnext = node; + node->hashprev = pool; + } + + 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) +{ +off64_t off = (page_no & ~bt->mgr->poolmask) << bt->mgr->page_bits; +off64_t limit = off + ((bt->mgr->poolmask+1) << bt->mgr->page_bits); +int flag; + +#ifdef unix + flag = PROT_READ | ( bt->mgr->mode == BT_ro ? 0 : PROT_WRITE ); + 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 + flag = ( bt->mgr->mode == BT_ro ? PAGE_READONLY : PAGE_READWRITE ); + pool->hmap = CreateFileMapping(bt->mgr->idx, NULL, flag, (DWORD)(limit >> 32), (DWORD)limit, NULL); + if( !pool->hmap ) + return bt->err = BTERR_map; + + flag = ( bt->mgr->mode == BT_ro ? FILE_MAP_READ : FILE_MAP_WRITE ); + pool->map = MapViewOfFile(pool->hmap, flag, (DWORD)(off >> 32), (DWORD)off, (bt->mgr->poolmask+1) << bt->mgr->page_bits); + if( !pool->map ) + return bt->err = BTERR_map; +#endif + return bt->err = 0; +} + +// find or place requested page in segment-pool +// return pool table entry, incrementing pin + +BtPool *bt_pinpage(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]); + + // look up in hash table + + if( pool = bt_findpool(bt, page_no, idx) ) { +#ifdef unix + __sync_fetch_and_add(&pool->pin, 1); +#else + _InterlockedIncrement16 (&pool->pin); +#endif + bt_spinreleaseread (&bt->mgr->latch[idx]); + pool->lru++; + return pool; + } + + // upgrade to write lock + + bt_spinreleaseread (&bt->mgr->latch[idx]); + bt_spinwritelock (&bt->mgr->latch[idx]); + + // try to find page in pool with write lock + + if( pool = bt_findpool(bt, page_no, idx) ) { +#ifdef unix + __sync_fetch_and_add(&pool->pin, 1); +#else + _InterlockedIncrement16 (&pool->pin); +#endif + bt_spinreleasewrite (&bt->mgr->latch[idx]); + pool->lru++; + return pool; + } + + // allocate a new pool node + // and add to hash table + +#ifdef unix + slot = __sync_fetch_and_add(&bt->mgr->poolcnt, 1); +#else + slot = _InterlockedIncrement16 (&bt->mgr->poolcnt) - 1; +#endif + + if( ++slot < bt->mgr->poolmax ) { + pool = bt->mgr->pool + slot; + pool->slot = slot; + + 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 + _InterlockedIncrement16 (&pool->pin); +#endif + bt_spinreleasewrite (&bt->mgr->latch[idx]); + return pool; + } + + // pool table is full + // find best pool entry to evict + +#ifdef unix + __sync_fetch_and_add(&bt->mgr->poolcnt, -1); +#else + _InterlockedDecrement16 (&bt->mgr->poolcnt); +#endif + + while( 1 ) { +#ifdef unix + victim = __sync_fetch_and_add(&bt->mgr->evicted, 1); +#else + victim = _InterlockedIncrement16 (&bt->mgr->evicted) - 1; +#endif + victim %= bt->mgr->hashsize; + + // try to get write lock + // skip entry if not obtained + + if( !bt_spinwritetry (&bt->mgr->latch[victim]) ) + continue; + + // if cache entry is empty + // or no slots are unpinned + // skip this entry + + if( !(pool = bt_findlru(bt, bt->mgr->hash[victim])) ) { + bt_spinreleasewrite (&bt->mgr->latch[victim]); + continue; + } + + // unlink victim pool node from hash table + + if( node = pool->hashprev ) + node->hashnext = pool->hashnext; + else if( node = pool->hashnext ) + bt->mgr->hash[victim] = node->slot; + else + bt->mgr->hash[victim] = 0; + + if( node = pool->hashnext ) + node->hashprev = pool->hashprev; + + bt_spinreleasewrite (&bt->mgr->latch[victim]); + + // remove old file mapping +#ifdef unix + munmap (pool->map, (bt->mgr->poolmask+1) << bt->mgr->page_bits); +#else + FlushViewOfFile(pool->map, 0); + UnmapViewOfFile(pool->map); + CloseHandle(pool->hmap); +#endif + pool->map = NULL; + + // create new pool mapping + // and link into hash table + + 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 + _InterlockedIncrement16 (&pool->pin); +#endif + bt_spinreleasewrite (&bt->mgr->latch[idx]); + return pool; + } +} + +// place write, read, or parent lock on requested page_no. +// pin to buffer pool and return latchset pointer + +BtLatchSet *bt_lockpage(BtDb *bt, uid page_no, BtLock mode, BtPage *pageptr, BtLatchSet *set) +{ +BtPool *pool; +uint subpage; +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 NULL; + + if( set ) +#ifdef unix + __sync_fetch_and_add(&set->pin, 1); +#else + _InterlockedIncrement16 (&set->pin); +#endif + else if( !(set = bt_bindlatch (bt, page_no, 1)) ) + return NULL; + + page = (BtPage)(pool->map + (subpage << bt->mgr->page_bits)); + +#ifdef unix + { + uint idx = subpage / 8; + uint bit = subpage % 8; + + if( mode == BtLockRead || mode == BtLockWrite ) + 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_spinreadlock (set->readwr); + break; + case BtLockWrite: + bt_spinwritelock (set->readwr); + break; + case BtLockAccess: + bt_spinreadlock (set->access); + break; + case BtLockDelete: + bt_spinwritelock (set->access); + break; + case BtLockParent: + bt_spinwritelock (set->parent); + break; + case BtLockPin: + break; + default: + return bt->err = BTERR_lock, NULL; + } + + if( pageptr ) + *pageptr = page; + + return set; +} + +// remove write, read, or parent lock on requested page_no. + +BTERR bt_unlockpage(BtDb *bt, uid page_no, BtLock mode, BtLatchSet *set) +{ +BtPool *pool; +uint idx; + + // 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_spinreadlock (&bt->mgr->latch[idx]); + + if( !(pool = bt_findpool(bt, page_no, idx)) ) + return bt->err = BTERR_hash; + + bt_spinreleaseread (&bt->mgr->latch[idx]); + + 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; + case BtLockPin: + break; + default: + return bt->err = BTERR_lock; + } + +#ifdef unix + __sync_fetch_and_add(&pool->pin, -1); + __sync_fetch_and_add (&set->pin, -1); +#else + _InterlockedDecrement16 (&pool->pin); + _InterlockedDecrement16 (&set->pin); +#endif + return bt->err = 0; +} + +// deallocate a deleted page +// place on free chain out of allocator page +// fence key must already be removed from parent + +BTERR bt_freepage(BtDb *bt, uid page_no, BtLatchSet *set) +{ + // obtain delete lock on deleted page + + if( !bt_lockpage(bt, page_no, BtLockDelete, NULL, set) ) + return bt->err; + + // obtain write lock on deleted page + + if( !bt_lockpage(bt, page_no, BtLockWrite, &bt->temp, set) ) + return bt->err; + + // lock allocation page + + bt_spinwritelock(bt->mgr->latchmgr->lock); + + // store free chain in allocation page second right + bt_putid(bt->temp->right, bt_getid(bt->mgr->latchmgr->alloc[1].right)); + bt_putid(bt->mgr->latchmgr->alloc[1].right, page_no); + + // unlock page zero + + bt_spinreleasewrite(bt->mgr->latchmgr->lock); + + // remove write lock on deleted node + + if( bt_unlockpage(bt, page_no, BtLockWrite, set) ) + return bt->err; + + // remove delete lock on deleted node + + if( bt_unlockpage(bt, page_no, BtLockDelete, set) ) + return bt->err; + + return 0; +} + +// allocate a new page and write page into it + +uid bt_newpage(BtDb *bt, BtPage page) +{ +BtLatchSet *set; +uid new_page; +BtPage pmap; +int reuse; + + // lock allocation page + + bt_spinwritelock(bt->mgr->latchmgr->lock); + + // use empty chain first + // else allocate empty page + + if( new_page = bt_getid(bt->mgr->latchmgr->alloc[1].right) ) { + if( !(set = bt_lockpage (bt, new_page, BtLockWrite, &bt->temp, NULL)) ) + return 0; + bt_putid(bt->mgr->latchmgr->alloc[1].right, bt_getid(bt->temp->right)); + if( bt_unlockpage (bt, new_page, BtLockWrite, set) ) + return 0; + reuse = 1; + } else { + new_page = bt_getid(bt->mgr->latchmgr->alloc->right); + bt_putid(bt->mgr->latchmgr->alloc->right, new_page+1); + 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; + } +#else + // bring new page into pool and copy page. + // this will extend the file into the new pages. + + if( !(set = bt_lockpage(bt, new_page, BtLockWrite, &pmap, NULL)) ) + return 0; + + memcpy(pmap, page, bt->mgr->page_size); + + if( bt_unlockpage (bt, new_page, BtLockWrite, set) ) + return 0; +#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) +{ +uint diff, higher = bt->page->cnt, low = 1, slot; + + // low is the lowest candidate, higher is already + // tested as .ge. the given key, loop ends when they meet + + while( diff = higher - low ) { + slot = low + ( diff >> 1 ); + if( keycmp (keyptr(bt->page, slot), key, len) < 0 ) + low = slot + 1; + else + higher = slot; + } + + return higher; +} + +// 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) +{ +uid page_no = ROOT_page, prevpage = 0; +BtLatchSet *set, *prevset; +uint drill = 0xff, slot; +uint mode, prevmode; + + bt->set = NULL; + + // start at root of btree and drill down + + do { + // determine lock mode of drill level + mode = (lock == BtLockWrite) && (drill == lvl) ? BtLockWrite : BtLockRead; + + bt->page_no = page_no; + + // obtain access lock using lock chaining with Access mode + + if( page_no > ROOT_page ) + if( !(bt->set = bt_lockpage(bt, page_no, BtLockAccess, NULL, NULL)) ) + return 0; + + // now unlock our (possibly foster) parent + + if( prevpage ) + if( bt_unlockpage(bt, prevpage, prevmode, prevset) ) + return 0; + else + prevpage = 0; + + // obtain read lock using lock chaining + // and pin page contents + + if( !(bt->set = bt_lockpage(bt, page_no, mode, &bt->page, bt->set)) ) + return 0; + + if( page_no > ROOT_page ) + if( bt_unlockpage(bt, page_no, BtLockAccess, bt->set) ) + return 0; + + // re-read and re-lock root after determining actual level of root + + if( bt->page_no == ROOT_page ) + if( bt->page->lvl != drill) { + drill = bt->page->lvl; + + if( lock == BtLockWrite && drill == lvl ) + if( bt_unlockpage(bt, page_no, mode, bt->set) ) + return 0; + else + continue; + } + + prevpage = bt->page_no; + prevset = bt->set; + prevmode = mode; + + // if page is being deleted, + // move back to preceeding page + + if( bt->page->kill ) { + page_no = bt_getid (bt->page->right); + continue; + } + + // find key on page at this level + // and descend to requested level + + slot = bt_findslot (bt, key, len); + + // is this slot a foster child? + + if( slot <= bt->page->cnt - bt->page->foster ) + if( drill == lvl ) + return slot; + + while( slotptr(bt->page, slot)->dead ) + if( slot++ < bt->page->cnt ) + continue; + else + goto slideright; + + if( slot <= bt->page->cnt - bt->page->foster ) + drill--; + + // continue down / right using overlapping locks + // to protect pages being killed or split. + + page_no = bt_getid(slotptr(bt->page, slot)->id); + continue; + +slideright: + page_no = bt_getid(bt->page->right); + + } while( page_no ); + + // return error on end of chain + + bt->err = BTERR_struct; + return 0; // return error +} + +// find and delete key on page by marking delete flag bit +// when page becomes empty, delete it from the btree + +BTERR bt_deletekey (BtDb *bt, unsigned char *key, uint len, uint lvl) +{ +unsigned char leftkey[256], rightkey[256]; +BtLatchSet *rset, *set; +uid page_no, right; +uint slot, tod; +BtKey ptr; + + 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 + + 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--; + } + + // return if page is not empty, or it has no right sibling + + right = bt_getid(bt->page->right); + page_no = bt->page_no; + set = bt->set; + + if( !right || bt->page->act ) + return bt_unlockpage(bt, page_no, BtLockWrite, set); + + // obtain Parent lock over write lock + + if( !bt_lockpage(bt, page_no, BtLockParent, NULL, set) ) + return bt->err; + + // cache copy of key to delete + + ptr = keyptr(bt->page, bt->page->cnt); + memcpy(leftkey, ptr, ptr->len + 1); + + // lock and map right page + + if( !(rset = bt_lockpage(bt, right, BtLockWrite, &bt->temp, NULL)) ) + return bt->err; + + // pull contents of next page into current empty page + memcpy (bt->page, bt->temp, bt->mgr->page_size); + + // cache copy of key to update + ptr = keyptr(bt->temp, bt->temp->cnt); + memcpy(rightkey, ptr, ptr->len + 1); + + // Mark right page as deleted and point it to left page + // until we can post updates at higher level. + + bt_putid(bt->temp->right, page_no); + bt->temp->kill = 1; + bt->temp->cnt = 0; + + if( bt_unlockpage(bt, right, BtLockWrite, rset) ) + return bt->err; + if( bt_unlockpage(bt, page_no, BtLockWrite, set) ) + return bt->err; + + // delete old lower key to consolidated node + + if( bt_deletekey (bt, leftkey + 1, *leftkey, lvl + 1) ) + return bt->err; + + // redirect higher key directly to consolidated node + + if( slot = bt_loadpage (bt, rightkey+1, *rightkey, lvl+1, BtLockWrite) ) + ptr = keyptr(bt->page, slot); + else + return bt->err; + + // since key already exists, update id + + if( keycmp (ptr, rightkey+1, *rightkey) ) + return bt->err = BTERR_struct; + + slotptr(bt->page, slot)->dead = 0; + bt_putid(slotptr(bt->page,slot)->id, page_no); + + if( bt_unlockpage(bt, bt->page_no, BtLockWrite, bt->set) ) + return bt->err; + + // obtain write lock and + // add right block to free chain + + if( bt_freepage (bt, right, rset) ) + return bt->err; + + // remove ParentModify lock + + if( bt_unlockpage(bt, page_no, BtLockParent, set) ) + return bt->err; + + return 0; +} + +// find key in leaf level and return row-id + +uid bt_findkey (BtDb *bt, unsigned char *key, uint len) +{ +uint slot; +BtKey ptr; +uid id; + + if( slot = bt_loadpage (bt, key, len, 0, BtLockRead) ) + ptr = keyptr(bt->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, bt->set) ) + return 0; + + return id; +} + +// check page for space available, +// clean if necessary and return +// 0 - page needs splitting +// 1 - go ahead + +uint bt_cleanpage(BtDb *bt, uint amt) +{ +uint nxt = bt->mgr->page_size; +BtPage page = bt->page; +uint cnt = 0, idx = 0; +uint max = page->cnt; +BtKey key; + + if( page->min >= (max+1) * sizeof(BtSlot) + sizeof(*page) + amt + 1 ) + return 1; + + // skip cleanup if nothing to reclaim + + if( !page->dirty ) + return 0; + + memcpy (bt->frame, page, bt->mgr->page_size); + + // skip page info and set rest of page to zero + + memset (page+1, 0, bt->mgr->page_size - sizeof(*page)); + page->dirty = 0; + page->act = 0; + + // try cleaning up page first + + while( cnt++ < max ) { + // always leave fence key and foster children in list + if( cnt < max - page->foster && slotptr(bt->frame,cnt)->dead ) + continue; + + // copy key + 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; + + // see if page has enough space now, or does it need splitting? + + if( page->min >= (idx+1) * sizeof(BtSlot) + sizeof(*page) + amt + 1 ) + return 1; + + 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; + + // calculate next available slot and copy key into page + + page->min -= len + 1; + ((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; +} + +// 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) +{ +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) + + root->act--; + root->cnt--; + root->foster--; + + // Save left fence key. + + key = keyptr(root, root->cnt); + memcpy (fencekey, key, key->len + 1); + + // copy the lower keys into a new left page + + 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 + + memset (root+1, 0, bt->mgr->page_size - sizeof(*root)); + + // insert left fence key on empty newroot page + + nxt -= *fencekey + 1; + memcpy ((unsigned char *)root + nxt, fencekey, *fencekey + 1); + bt_putid(slotptr(root, 1)->id, new_page); + slotptr(root, 1)->off = nxt; + + // insert stopper 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); + slotptr(root, 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++; + + // release root (bt->page) + + return bt_unlockpage(bt, ROOT_page, BtLockWrite, bt->set); +} + +// split already locked full node +// in current page variables +// 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; +BtPage page = bt->page; +uint tod = time(NULL); +uint lvl = page->lvl; +uid new_page, right; +BtKey key; + + // initialize frame buffer + + memset (bt->frame, 0, bt->mgr->page_size); + max = page->cnt - page->foster; + tod = (uint)time(NULL); + cnt = max / 2; + idx = 0; + + // split higher half of keys to bt->frame + // leaving foster children in the left node. + + while( cnt++ < max ) { + key = keyptr(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); + slotptr(bt->frame, idx)->tod = slotptr(page, cnt)->tod; + slotptr(bt->frame, idx)->off = nxt; + bt->frame->act++; + } + + // transfer right link node + + if( page_no > ROOT_page ) { + right = bt_getid (page->right); + bt_putid(bt->frame->right, right); + } + + 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 frame to it. + + if( !(new_page = bt_newpage(bt, bt->frame)) ) + return bt->err; + + // remember fence key for new page to add + // as foster child + + key = keyptr(bt->frame, idx); + memcpy (fencekey, key, key->len + 1); + + // update lower keys and foster children 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->act = 0; + cnt = 0; + idx = 0; + + // assemble page of smaller keys + // to remain in the old page + + 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++; + } + + // insert new foster child at beginning 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 if any + + 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); + slotptr(page, idx)->tod = slotptr(bt->frame, cnt)->tod; + slotptr(page, idx)->off = nxt; + page->act++; + } + + 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); + + // keep our latch set + // release wr lock on our page + + if( !bt_lockpage (bt, page_no, BtLockPin, NULL, set) ) + return bt->err; + + if( bt_unlockpage (bt, page_no, BtLockWrite, set) ) + return bt->err; + + // obtain ParentModification lock for current page + // to fix fence key and highest foster child on page + + if( !bt_lockpage (bt, page_no, BtLockParent, NULL, set) ) + return bt->err; + + // get our highest foster child key to find in parent node + + if( !bt_lockpage (bt, page_no, BtLockRead, &page, set) ) + return bt->err; + + key = keyptr(page, page->cnt); + memcpy (fencekey, key, key->len+1); + + if( bt_unlockpage (bt, page_no, BtLockRead, set) ) + return bt->err; + + // update our parent +try_again: + + do { + slot = bt_loadpage (bt, fencekey + 1, *fencekey, lvl + 1, BtLockWrite); + + if( !slot ) + return bt->err; + + // check if parent page has enough space for any possible key + + if( bt_cleanpage (bt, 256) ) + break; + + if( bt_splitpage (bt) ) + return bt->err; + } while( 1 ); + + // see if we are still a foster child from another node + + if( bt_getid (slotptr(bt->page, slot)->id) != page_no ) { + if( bt_unlockpage (bt, bt->page_no, BtLockWrite, bt->set) ) + return bt->err; +#ifdef unix + sched_yield(); +#else + SwitchToThread(); +#endif + goto try_again; + } + + // wait until readers from parent get their locks + // on our page + + if( !bt_lockpage (bt, page_no, BtLockDelete, NULL, set) ) + return bt->err; + + // lock our page for writing + + if( !bt_lockpage (bt, page_no, BtLockWrite, &page, set) ) + return bt->err; + + // switch parent fence key to foster child + + if( slotptr(page, page->cnt)->dead ) + slotptr(bt->page, slot)->dead = 1; + else + bt_putid (slotptr(bt->page, slot)->id, bt_getid(slotptr(page, page->cnt)->id)); + + // remove highest foster child from our page + + page->cnt--; + page->act--; + page->foster--; + page->dirty = 1; + key = keyptr(page, page->cnt); + + // add our new fence key for foster child to our parent + + bt_addkeytopage (bt, slot, key->key, key->len, page_no, tod); + + if( bt_unlockpage (bt, bt->page_no, BtLockWrite, bt->set) ) + return bt->err; + + if( bt_unlockpage (bt, page_no, BtLockDelete, set) ) + return bt->err; + + if( bt_unlockpage (bt, page_no, BtLockWrite, set) ) + return bt->err; + + if( bt_unlockpage (bt, page_no, BtLockParent, set) ) + return bt->err; + + // release extra latch pin + + return bt_unlockpage (bt, page_no, BtLockPin, set); +} + +// Insert new key into the btree at leaf level. + +BTERR bt_insertkey (BtDb *bt, unsigned char *key, uint len, uid id, uint tod) +{ +uint slot, idx; +BtPage page; +BtKey ptr; + + while( 1 ) { + if( slot = bt_loadpage (bt, key, len, 0, 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; + + 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, bt->set); + } + + // check if page has enough space + + if( bt_cleanpage (bt, len) ) + break; + + if( bt_splitpage (bt) ) + return bt->err; + } + + bt_addkeytopage (bt, slot, key, len, id, tod); + + return bt_unlockpage (bt, bt->page_no, BtLockWrite, bt->set); +} + +// cache page of keys into cursor and return starting slot for given key + +uint bt_startkey (BtDb *bt, unsigned char *key, uint len) +{ +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, bt->set) ) + return 0; + + return slot; +} + +// return next slot for cursor page +// or slide cursor right into next page + +uint bt_nextkey (BtDb *bt, uint slot) +{ +BtLatchSet *rset; +BtPage page; +uid right; + + do { + right = bt_getid(bt->cursor->right); + while( slot++ < bt->cursor->cnt - bt->cursor->foster ) + if( slotptr(bt->cursor,slot)->dead ) + continue; + else if( right || (slot < bt->cursor->cnt - bt->cursor->foster) ) + return slot; + else + break; + + if( !right ) + break; + + bt->cursor_page = right; + + if( !(bt->set = bt_lockpage(bt, right, BtLockRead, &page, NULL)) ) + return 0; + + memcpy (bt->cursor, page, bt->mgr->page_size); + + if ( bt_unlockpage(bt, right, BtLockRead, bt->set) ) + return 0; + + slot = 0; + } while( 1 ); + + return bt->err = 0; +} + +BtKey bt_key(BtDb *bt, uint slot) +{ + return keyptr(bt->cursor, slot); +} + +uid bt_uid(BtDb *bt, uint slot) +{ + return bt_getid(slotptr(bt->cursor,slot)->id); +} + +uint bt_tod(BtDb *bt, uint slot) +{ + return slotptr(bt->cursor,slot)->tod; +} + + +#ifdef STANDALONE + +typedef struct { + char type, idx; + char *infile; + BtMgr *mgr; + int num; +} ThreadArg; + +// standalone program to index file of keys +// then list them onto std-out + +#ifdef unix +void *index_file (void *arg) +#else +uint __stdcall index_file (void *arg) +#endif +{ +int line = 0, found = 0, cnt = 0; +uid next, page_no = LEAF_page; // start on first page of leaves +unsigned char key[256]; +ThreadArg *args = arg; +int ch, len = 0, slot; +time_t tod[1]; +BtPage page; +BtKey ptr; +BtDb *bt; +FILE *in; + + bt = bt_open (args->mgr); + time (tod); + + switch(args->type | 0x20) + { + case 'w': + fprintf(stderr, "started indexing for %s\n", args->infile); + if( in = fopen (args->infile, "rb") ) + while( ch = getc(in), ch != EOF ) + if( ch == '\n' ) + { + 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; + + if( bt_insertkey (bt, key, len, line, *tod) ) + fprintf(stderr, "Error %d Line: %d\n", bt->err, line), exit(0); + len = 0; + } + else if( len < 255 ) + key[len++] = ch; + fprintf(stderr, "finished %s for %d keys\n", args->infile, line); + break; + + case 'd': + fprintf(stderr, "started deleting keys for %s\n", args->infile); + if( in = fopen (args->infile, "rb") ) + while( ch = getc(in), ch != EOF ) + if( ch == '\n' ) + { + 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; + + if( bt_deletekey (bt, key, len, 0) ) + fprintf(stderr, "Error %d Line: %d\n", bt->err, line), exit(0); + len = 0; + } + else if( len < 255 ) + key[len++] = ch; + fprintf(stderr, "finished %s for keys, %d \n", args->infile, line); + break; + + case 'f': + fprintf(stderr, "started finding keys for %s\n", args->infile); + if( in = fopen (args->infile, "rb") ) + while( ch = getc(in), ch != EOF ) + if( ch == '\n' ) + { + 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; + + if( bt_findkey (bt, key, len) ) + found++; + else if( bt->err ) + fprintf(stderr, "Error %d Syserr %d Line: %d\n", bt->err, errno, line), exit(0); + len = 0; + } + else if( len < 255 ) + key[len++] = ch; + fprintf(stderr, "finished %s for %d keys, found %d\n", args->infile, line, found); + 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); + } + + break; + + case 'c': + fprintf(stderr, "started reading\n"); + + do { + bt->set = bt_lockpage (bt, page_no, BtLockRead, &page, NULL); + cnt += page->act; + next = bt_getid (page->right); + bt_unlockpage (bt, page_no, BtLockRead, bt->set); + } while( page_no = next ); + + cnt--; // remove stopper key + fprintf(stderr, " Total keys read %d\n", cnt); + break; + } + + bt_close (bt); +#ifdef unix + return NULL; +#else + return 0; +#endif +} + +typedef struct timeval timer; + +int main (int argc, char **argv) +{ +int idx, cnt, len, slot, err; +int segsize, bits = 16; +#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; +int num = 0; +char key[1]; +BtMgr *mgr; +BtKey ptr; +BtDb *bt; + + if( argc < 3 ) { + fprintf (stderr, "Usage: %s idx_file Read/Write/Scan/Delete/Find [page_bits mapped_segments seg_bits line_numbers src_file1 src_file2 ... ]\n", argv[0]); + 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 = 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 + + if( argc > 3 ) + bits = atoi(argv[3]); + + if( argc > 4 ) + poolsize = atoi(argv[4]); + + if( !poolsize ) + fprintf (stderr, "Warning: no mapped_pool\n"); + + if( poolsize > 65535 ) + fprintf (stderr, "Warning: mapped_pool > 65535 segments\n"); + + if( argc > 5 ) + segsize = atoi(argv[5]); + else + segsize = 4; // 16 pages per mmap segment + + if( argc > 6 ) + num = atoi(argv[6]); + + cnt = argc - 7; +#ifdef unix + threads = malloc (cnt * sizeof(pthread_t)); +#else + threads = GlobalAlloc (GMEM_FIXED|GMEM_ZEROINIT, cnt * sizeof(HANDLE)); +#endif + args = malloc (cnt * sizeof(ThreadArg)); + + mgr = bt_mgr ((argv[1]), BT_rw, bits, poolsize, segsize, poolsize / 8); + + if( !mgr ) { + fprintf(stderr, "Index Open Error %s\n", argv[1]); + exit (1); + } + + // fire off threads + + for( idx = 0; idx < cnt; idx++ ) { + args[idx].infile = argv[idx + 7]; + args[idx].type = argv[2][0]; + args[idx].mgr = mgr; + args[idx].num = num; + args[idx].idx = idx; +#ifdef unix + if( err = pthread_create (threads + idx, NULL, index_file, args + idx) ) + fprintf(stderr, "Error creating thread %d\n", err); +#else + threads[idx] = (HANDLE)_beginthreadex(NULL, 65536, index_file, args + idx, 0, NULL); +#endif + } + + // wait for termination + +#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); + bt_mgrclose (mgr); +} + +#endif //STANDALONE diff --git a/fosterbtreef2.c b/fosterbtreef2.c new file mode 100644 index 0000000..e4592d2 --- /dev/null +++ b/fosterbtreef2.c @@ -0,0 +1,2509 @@ +// foster btree version f2 +// 18 JAN 2014 + +// author: karl malbrain, malbrain@cal.berkeley.edu + +/* +This work, including the source code, documentation +and related data, is placed into the public domain. + +The orginal author is Karl Malbrain. + +THIS SOFTWARE IS PROVIDED AS-IS WITHOUT WARRANTY +OF ANY KIND, NOT EVEN THE IMPLIED WARRANTY OF +MERCHANTABILITY. THE AUTHOR OF THIS SOFTWARE, +ASSUMES _NO_ RESPONSIBILITY FOR ANY CONSEQUENCE +RESULTING FROM THE USE, MODIFICATION, OR +REDISTRIBUTION OF THIS SOFTWARE. +*/ + +// Please see the project home page for documentation +// code.google.com/p/high-concurrency-btree + +#define _FILE_OFFSET_BITS 64 +#define _LARGEFILE64_SOURCE + +#ifdef linux +#define _GNU_SOURCE +#endif + +#ifdef unix +#include +#include +#include +#include +#include +#include +#include +#include +#else +#define WIN32_LEAN_AND_MEAN +#include +#include +#include +#include +#include +#include +#include +#endif + +#include +#include + +typedef unsigned long long uid; + +#ifndef unix +typedef unsigned long long off64_t; +typedef unsigned short ushort; +typedef unsigned int uint; +#endif + +#define BT_ro 0x6f72 // ro +#define BT_rw 0x7772 // rw + +#define BT_latchtable 128 // number of latch manager slots + +#define BT_maxbits 24 // maximum page size in bits +#define BT_minbits 9 // minimum page size in bits +#define BT_minpage (1 << BT_minbits) // minimum page size +#define BT_maxpage (1 << BT_maxbits) // maximum page size + +/* +There are five lock types for each node in three independent sets: +1. (set 1) AccessIntent: Sharable. Going to Read the node. Incompatible with NodeDelete. +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. +*/ + +typedef enum{ + BtLockAccess, + BtLockDelete, + BtLockRead, + BtLockWrite, + BtLockParent +}BtLock; + +// Define the length of the page and key pointers + +#define BtId 6 + +// Page key slot definition. + +// If BT_maxbits is 15 or less, you can save 4 bytes +// for each key stored by making the first two uints +// into ushorts. You can also save 4 bytes by removing +// the tod field from the key. + +// Keys are marked dead, but remain on the page until +// it cleanup is called. The fence key (highest key) for +// the page is always present, even after cleanup. + +typedef struct { + uint off:BT_maxbits; // page offset for key start + uint dead:1; // set for deleted key + uint tod; // time-stamp for key + unsigned char id[BtId]; // id associated with key +} BtSlot; + +// The key structure occupies space at the upper end of +// each page. It's a length byte followed by the value +// bytes. + +typedef struct { + unsigned char len; + unsigned char key[1]; +} *BtKey; + +// The first part of an index page. +// It is immediately followed +// 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 + 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 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]; + 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 + 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; + +// 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 + ushort pin; // mapped page pin counter + ushort slot; // slot index in this array + void *hashprev; // previous pool entry for the same hash idx + void *hashnext; // next pool entry for the same hash idx +#ifndef unix + HANDLE hmap; // Windows memory mapping handle +#endif +} BtPool; + +// structure for latch manager on ALLOC_page + +typedef struct { + struct Page 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 { + uint page_size; // page size + uint page_bits; // page size in bits + 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 +#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 + BtPool *pool; // memory pool page segments + BtSpinLatch *latch; // latches for pool hash slots + BtLatchMgr *latchmgr; // mapped latch page from allocation page + BtLatchSet *latchsets; // mapped latch set from latch pages +#ifndef unix + HANDLE halloc; // allocation and latch table handle +#endif +} BtMgr; + +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 + uid page_no; // current page number + uid cursor_page; // current cursor page number + BtLatchSet *set; // current page latch set + BtPool *pool; // current page pool + unsigned char *mem; // frame, cursor, page memory buffer + int found; // last delete was found + int err; // last error +} BtDb; + +typedef enum { + BTERR_ok = 0, + BTERR_struct, + BTERR_ovflw, + BTERR_lock, + BTERR_map, + BTERR_wrt, + BTERR_hash, + BTERR_latch +} BTERR; + +// 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 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); + +// manager functions +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 + +extern BtKey bt_key (BtDb *bt, uint slot); +extern uid bt_uid (BtDb *bt, uint slot); +extern uint bt_tod (BtDb *bt, uint slot); + +// BTree page number constants +#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 + +#define MIN_lvl 2 + +// The page is allocated from low and hi ends. +// The key offsets and row-id's are allocated +// from the bottom, while the text of the key +// is allocated from the top. When the two +// areas meet, the page is split into two. + +// 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. + +// The b-tree root is always located at page 1. +// The first leaf page of level zero is always +// located on page 2. + +// 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. + +// 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. + +// Groups of pages called segments from the btree are +// cached with memory mapping. 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. + +// Page 0 is dedicated to lock for new page extensions, +// and chains empty pages together for reuse. + +// Empty pages are chained together through the ALLOC page and reused. + +// Access macros to address slot and key values from the page + +#define slotptr(page, slot) (((BtSlot *)(page+1)) + (slot-1)) +#define keyptr(page, slot) ((BtKey)((unsigned char*)(page) + slotptr(page, slot)->off)) + +void bt_putid(unsigned char *dest, uid id) +{ +int i = BtId; + + while( i-- ) + dest[i] = (unsigned char)id, id >>= 8; +} + +uid bt_getid(unsigned char *src) +{ +uid id = 0; +int i; + + for( i = 0; i < BtId; i++ ) + id <<= 8, id |= *src++; + + return id; +} + +// wait until write lock mode is clear +// and add 1 to the share count + +void bt_spinreadlock(BtSpinLatch *latch) +{ +ushort prev; + + do { +#ifdef unix + while( __sync_fetch_and_or((ushort *)latch, Mutex) & Mutex ) + sched_yield(); +#else + while( _InterlockedOr16((ushort *)latch, Mutex) & Mutex ) + SwitchToThread(); +#endif + + // see if exclusive request is granted or pending + + if( prev = !(latch->exclusive | latch->pending) ) +#ifdef unix + __sync_fetch_and_add((ushort *)latch, Share); +#else + _InterlockedExchangeAdd16 ((ushort *)latch, Share); +#endif + +#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 +} + +// wait for other read and write latches to relinquish + +void bt_spinwritelock(BtSpinLatch *latch) +{ + do { +#ifdef unix + while( __sync_fetch_and_or((ushort *)latch, Mutex | Pending) & 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; + } + +#ifdef unix + __sync_fetch_and_and ((ushort *)latch, ~Mutex); + sched_yield(); +#else + _InterlockedAnd16((ushort *)latch, ~Mutex); + SwitchToThread(); +#endif + } while( 1 ); +} + +// try to obtain write lock + +// return 1 if obtained, +// 0 otherwise + +int bt_spinwritetry(BtSpinLatch *latch) +{ +ushort prev; + +#ifdef unix + if( prev = __sync_fetch_and_or((ushort *)latch, Mutex), prev & Mutex ) + return 0; +#else + if( prev = _InterlockedOr16((ushort *)latch, Mutex), prev & 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 + +#ifdef unix + __sync_fetch_and_and ((ushort *)latch, ~Mutex); +#else + _InterlockedAnd16((ushort *)latch, ~Mutex); +#endif + return !prev; +} + +// clear write mode + +void bt_spinreleasewrite(BtSpinLatch *latch) +{ +#ifdef unix + __sync_fetch_and_and ((ushort *)latch, ~Write); +#else + _InterlockedAnd16((ushort *)latch, ~Write); +#endif +} + +// decrement reader count + +void bt_spinreleaseread(BtSpinLatch *latch) +{ +#ifdef unix + __sync_fetch_and_add((ushort *)latch, -Share); +#else + _InterlockedExchangeAdd16 ((ushort *)latch, -Share); +#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; +uint slot; + + // release mapped pages + // note that slot zero is never used + + for( slot = 1; slot < mgr->poolmax; slot++ ) { + pool = mgr->pool + slot; + if( pool->slot ) +#ifdef unix + munmap (pool->map, (mgr->poolmask+1) << mgr->page_bits); +#else + { + FlushViewOfFile(pool->map, 0); + UnmapViewOfFile(pool->map); + CloseHandle(pool->hmap); + } +#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 (mgr); +#else + FlushFileBuffers(mgr->idx); + CloseHandle(mgr->idx); + GlobalFree (mgr->pool); + GlobalFree (mgr->hash); + GlobalFree (mgr->latch); + GlobalFree (mgr); +#endif +} + +// close and release memory + +void bt_close (BtDb *bt) +{ +#ifdef unix + if ( bt->mem ) + free (bt->mem); +#else + if ( bt->mem) + VirtualFree (bt->mem, 0, MEM_RELEASE); +#endif + free (bt); +} + +// open/create new btree buffer manager + +// call with file_name, BT_openmode, bits in page size (e.g. 16), +// size of mapped page pool (e.g. 8192) + +BtMgr *bt_mgr (char *name, uint mode, uint bits, uint poolmax, uint segsize, uint hashsize) +{ +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]; +#endif + + // determine sanity of page size and buffer pool + + if( bits > BT_maxbits ) + bits = BT_maxbits; + else if( bits < BT_minbits ) + bits = BT_minbits; + + if( !poolmax ) + return NULL; // must have buffer pool + +#ifdef unix + mgr = calloc (1, sizeof(BtMgr)); + + mgr->idx = open ((char*)name, O_RDWR | O_CREAT, 0666); + + if( mgr->idx == -1 ) + return free(mgr), NULL; + + cacheblk = 4096; // minimum mmap segment size for unix + +#else + mgr = GlobalAlloc (GMEM_FIXED|GMEM_ZEROINIT, sizeof(BtMgr)); + attr = FILE_ATTRIBUTE_NORMAL; + mgr->idx = CreateFile(name, GENERIC_READ| GENERIC_WRITE, FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, OPEN_ALWAYS, attr, NULL); + + if( mgr->idx == INVALID_HANDLE_VALUE ) + return GlobalFree(mgr), NULL; + + // normalize cacheblk to multiple of sysinfo->dwAllocationGranularity + GetSystemInfo(sysinfo); + cacheblk = sysinfo->dwAllocationGranularity; +#endif + +#ifdef unix + 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, latchmgr, BT_minpage, 0) == BT_minpage ) + bits = latchmgr->alloc->bits; + else + return free(mgr), free(latchmgr), NULL; + } else if( mode == BT_ro ) + return free(latchmgr), bt_mgrclose (mgr), NULL; +#else + latchmgr = VirtualAlloc(NULL, BT_maxpage, MEM_COMMIT, PAGE_READWRITE); + size = GetFileSize(mgr->idx, amt); + + if( size || *amt ) { + if( !ReadFile(mgr->idx, (char *)latchmgr, BT_minpage, amt, NULL) ) + return bt_mgrclose (mgr), NULL; + bits = latchmgr->alloc->bits; + } else if( mode == BT_ro ) + return bt_mgrclose (mgr), NULL; +#endif + + mgr->page_size = 1 << bits; + mgr->page_bits = bits; + + mgr->poolmax = poolmax; + mgr->mode = mode; + + if( cacheblk < mgr->page_size ) + cacheblk = mgr->page_size; + + // mask for partial memmaps + + mgr->poolmask = (cacheblk >> bits) - 1; + + // see if requested size of pages per memmap is greater + + if( (1 << segsize) > mgr->poolmask ) + mgr->poolmask = (1 << segsize) - 1; + + mgr->seg_bits = 0; + + while( (1 << mgr->seg_bits) <= mgr->poolmask ) + mgr->seg_bits++; + + mgr->hashsize = hashsize; + +#ifdef unix + mgr->pool = calloc (poolmax, sizeof(BtPool)); + mgr->hash = calloc (hashsize, sizeof(ushort)); + mgr->latch = calloc (hashsize, sizeof(BtSpinLatch)); + 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)); +#endif + + if( size || *amt ) + 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)); + + // initialize latch manager + + 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, latchmgr, mgr->page_size) < mgr->page_size ) + return bt_mgrclose (mgr), NULL; +#else + 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 (latchmgr, 0, 1 << bits); + latchmgr->alloc->bits = mgr->page_bits; + + for( lvl=MIN_lvl; lvl--; ) { + 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; + 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, latchmgr, mgr->page_size) < mgr->page_size ) + return bt_mgrclose (mgr), NULL; +#else + 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 + } + + // clear out latch manager locks + // and rest of pages to round out segment + + memset(latchmgr, 0, mgr->page_size); + last = MIN_lvl + 1; + + while( last <= ((MIN_lvl + 1 + nlatchpage) | mgr->poolmask) ) { +#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 + last++; + } + +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 (latchmgr); +#else + VirtualFree (latchmgr, 0, MEM_RELEASE); +#endif + return mgr; +} + +// open BTree access method +// based on buffer manager + +BtDb *bt_open (BtMgr *mgr) +{ +BtDb *bt = malloc (sizeof(*bt)); + + memset (bt, 0, sizeof(*bt)); + bt->mgr = mgr; +#ifdef unix + bt->mem = malloc (3 *mgr->page_size); +#else + bt->mem = VirtualAlloc(NULL, 3 * mgr->page_size, MEM_COMMIT, PAGE_READWRITE); +#endif + bt->frame = (BtPage)bt->mem; + 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); + return bt; +} + +// compare two keys, returning > 0, = 0, or < 0 +// as the comparison value + +int keycmp (BtKey key1, unsigned char *key2, uint len2) +{ +uint len1 = key1->len; +int ans; + + if( ans = memcmp (key1->key, key2, len1 > len2 ? len2 : len1) ) + return ans; + + if( len1 > len2 ) + return 1; + if( len1 < len2 ) + return -1; + + return 0; +} + +// Buffer Pool mgr + +// find segment in pool +// must be called with hashslot idx locked +// return NULL if not there +// otherwise return node + +BtPool *bt_findpool(BtDb *bt, uid page_no, uint idx) +{ +BtPool *pool; +uint slot; + + // compute start of hash chain in pool + + if( slot = bt->mgr->hash[idx] ) + pool = bt->mgr->pool + slot; + else + return NULL; + + page_no &= ~bt->mgr->poolmask; + + while( pool->basepage != page_no ) + if( pool = pool->hashnext ) + continue; + else + return NULL; + + return pool; +} + +// add segment to hash table + +void bt_linkhash(BtDb *bt, BtPool *pool, uid page_no, int idx) +{ +BtPool *node; +uint slot; + + pool->hashprev = pool->hashnext = NULL; + pool->basepage = page_no & ~bt->mgr->poolmask; + pool->lru = 1; + + if( slot = bt->mgr->hash[idx] ) { + node = bt->mgr->pool + slot; + pool->hashnext = node; + node->hashprev = pool; + } + + 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) +{ +off64_t off = (page_no & ~bt->mgr->poolmask) << bt->mgr->page_bits; +off64_t limit = off + ((bt->mgr->poolmask+1) << bt->mgr->page_bits); +int flag; + +#ifdef unix + flag = PROT_READ | ( bt->mgr->mode == BT_ro ? 0 : PROT_WRITE ); + 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 + flag = ( bt->mgr->mode == BT_ro ? PAGE_READONLY : PAGE_READWRITE ); + pool->hmap = CreateFileMapping(bt->mgr->idx, NULL, flag, (DWORD)(limit >> 32), (DWORD)limit, NULL); + if( !pool->hmap ) + return bt->err = BTERR_map; + + flag = ( bt->mgr->mode == BT_ro ? FILE_MAP_READ : FILE_MAP_WRITE ); + pool->map = MapViewOfFile(pool->hmap, flag, (DWORD)(off >> 32), (DWORD)off, (bt->mgr->poolmask+1) << bt->mgr->page_bits); + if( !pool->map ) + return bt->err = BTERR_map; +#endif + return bt->err = 0; +} + +// calculate page within pool + +BtPage bt_page (BtDb *bt, BtPool *pool, uid page_no) +{ +uint subpage = (uint)(page_no & bt->mgr->poolmask); // page within mapping +BtPage page; + + 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 + return page; +} + +// release pool pin + +void bt_unpinpool (BtPool *pool) +{ +#ifdef unix + __sync_fetch_and_add(&pool->pin, -1); +#else + _InterlockedDecrement16 (&pool->pin); +#endif +} + +// find or place requested page in segment-pool +// return pool table entry, incrementing pin + +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]); + + // look up in hash table + + if( pool = bt_findpool(bt, page_no, idx) ) { +#ifdef unix + __sync_fetch_and_add(&pool->pin, 1); +#else + _InterlockedIncrement16 (&pool->pin); +#endif + bt_spinreleaseread (&bt->mgr->latch[idx]); + pool->lru++; + return pool; + } + + // upgrade to write lock + + bt_spinreleaseread (&bt->mgr->latch[idx]); + bt_spinwritelock (&bt->mgr->latch[idx]); + + // try to find page in pool with write lock + + if( pool = bt_findpool(bt, page_no, idx) ) { +#ifdef unix + __sync_fetch_and_add(&pool->pin, 1); +#else + _InterlockedIncrement16 (&pool->pin); +#endif + bt_spinreleasewrite (&bt->mgr->latch[idx]); + pool->lru++; + return pool; + } + + // allocate a new pool node + // and add to hash table + +#ifdef unix + slot = __sync_fetch_and_add(&bt->mgr->poolcnt, 1); +#else + slot = _InterlockedIncrement16 (&bt->mgr->poolcnt) - 1; +#endif + + if( ++slot < bt->mgr->poolmax ) { + pool = bt->mgr->pool + slot; + pool->slot = slot; + + 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 + _InterlockedIncrement16 (&pool->pin); +#endif + bt_spinreleasewrite (&bt->mgr->latch[idx]); + return pool; + } + + // pool table is full + // find best pool entry to evict + +#ifdef unix + __sync_fetch_and_add(&bt->mgr->poolcnt, -1); +#else + _InterlockedDecrement16 (&bt->mgr->poolcnt); +#endif + + while( 1 ) { +#ifdef unix + victim = __sync_fetch_and_add(&bt->mgr->evicted, 1); +#else + victim = _InterlockedIncrement16 (&bt->mgr->evicted) - 1; +#endif + victim %= bt->mgr->hashsize; + + // try to get write lock + // skip entry if not obtained + + if( !bt_spinwritetry (&bt->mgr->latch[victim]) ) + continue; + + // if cache entry is empty + // or no slots are unpinned + // skip this entry + + if( !(pool = bt_findlru(bt, bt->mgr->hash[victim])) ) { + bt_spinreleasewrite (&bt->mgr->latch[victim]); + continue; + } + + // unlink victim pool node from hash table + + if( node = pool->hashprev ) + node->hashnext = pool->hashnext; + else if( node = pool->hashnext ) + bt->mgr->hash[victim] = node->slot; + else + bt->mgr->hash[victim] = 0; + + if( node = pool->hashnext ) + node->hashprev = pool->hashprev; + + bt_spinreleasewrite (&bt->mgr->latch[victim]); + + // remove old file mapping +#ifdef unix + munmap (pool->map, (bt->mgr->poolmask+1) << bt->mgr->page_bits); +#else + FlushViewOfFile(pool->map, 0); + UnmapViewOfFile(pool->map); + CloseHandle(pool->hmap); +#endif + pool->map = NULL; + + // create new pool mapping + // and link into hash table + + 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 + _InterlockedIncrement16 (&pool->pin); +#endif + bt_spinreleasewrite (&bt->mgr->latch[idx]); + return pool; + } +} + +// place write, read, or parent lock on requested page_no. +// pin to buffer pool and return latchset pointer + +void bt_lockpage(BtLock mode, BtLatchSet *set) +{ + switch( mode ) { + case BtLockRead: + bt_spinreadlock (set->readwr); + break; + case BtLockWrite: + bt_spinwritelock (set->readwr); + break; + case BtLockAccess: + bt_spinreadlock (set->access); + break; + case BtLockDelete: + bt_spinwritelock (set->access); + break; + case BtLockParent: + bt_spinwritelock (set->parent); + break; + } +} + +// remove write, read, or parent lock on requested page_no. + +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) +{ +BtLatchSet *set; +BtPool *pool; +uid new_page; +BtPage pmap; +int reuse; + + // lock allocation page + + bt_spinwritelock(bt->mgr->latchmgr->lock); + + // use empty chain first + // else allocate empty page + + if( new_page = bt_getid(bt->mgr->latchmgr->alloc[1].right) ) { + if( pool = bt_pinpool (bt, new_page) ) + pmap = bt_page (bt, pool, new_page); + else + return 0; + bt_putid(bt->mgr->latchmgr->alloc[1].right, bt_getid(pmap->right)); + bt_unpinpool (pool); + reuse = 1; + } else { + 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, 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; + } + + // 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); + else + return 0; + + memcpy(pmap, page, bt->mgr->page_size); + bt_unpinpool (pool); +#endif + return new_page; +} + +// find slot in page for given key at a given level + +int bt_findslot (BtDb *bt, unsigned char *key, uint len) +{ +uint diff, higher = bt->page->cnt, low = 1, slot; + + // low is the lowest candidate, higher is already + // tested as .ge. the given key, loop ends when they meet + + while( diff = higher - low ) { + slot = low + ( diff >> 1 ); + if( keycmp (keyptr(bt->page, slot), key, len) < 0 ) + low = slot + 1; + else + higher = slot; + } + + return higher; +} + +// 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) +{ +uid page_no = ROOT_page, prevpage = 0; +BtLatchSet *set, *prevset; +uint drill = 0xff, slot; +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; + + // obtain latch set for this page + + bt->set = bt_pinlatch (bt, page_no); + bt->page_no = page_no; + + // pin page contents + + if( bt->pool = bt_pinpool (bt, page_no) ) + bt->page = bt_page (bt, bt->pool, page_no); + else + return 0; + + // obtain access lock using lock chaining with Access mode + + if( page_no > ROOT_page ) + bt_lockpage(BtLockAccess, bt->set); + + // now unlock and unpin our (possibly foster) parent + + if( prevpage ) { + bt_unlockpage(prevmode, prevset); + bt_unpinlatch (prevset); + bt_unpinpool (prevpool); + prevpage = 0; + } + + // obtain read lock using lock chaining + + bt_lockpage(mode, bt->set); + + if( page_no > ROOT_page ) + bt_unlockpage(BtLockAccess, bt->set); + + // re-read and re-lock root after determining actual level of root + + if( page_no == ROOT_page ) + if( bt->page->lvl != drill) { + drill = bt->page->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); + + // 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( 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 + + // is this slot < foster child area + // if so, drill to next level + + if( slot <= bt->page->cnt - bt->page->foster ) + drill--; + + // continue right onto foster child + // or down to next level. + + page_no = bt_getid(slotptr(bt->page, slot)->id); + + } while( page_no ); + + // return error on end of 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 + +BTERR bt_deletekey (BtDb *bt, unsigned char *key, uint len) +{ +unsigned char leftkey[256]; +BtLatchSet *rset, *set; +BtPool *pool, *rpool; +BtPage rpage, page; +uid page_no, right; +uint slot, tod; +BtKey ptr; + + if( slot = bt_loadpage (bt, key, len, 0, 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; + if( slot < bt->page->cnt ) + bt->page->dirty = 1; + bt->page->act--; + } + + 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 ) { + 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); + + // 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); + + // 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; + } + + // 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); + + if( rpool = bt_pinpool (bt, right) ) + rpage = bt_page (bt, rpool, right); + else + return bt->err; + + 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); + + // delete left parent slot for old empty page + // and redirect right parent slot to it + + bt->page->act--; + bt->page->dirty = 1; + slotptr(bt->page, slot)->dead = 1; + + while( slot++ < bt->page->cnt ) + if( !slotptr(bt->page, slot)->dead ) + break; + + bt_putid(slotptr(bt->page,slot)->id, page_no); + + // release parent level lock + // and our empty node lock + + bt_unlockpage(BtLockWrite, set); + bt_unlockpage(BtLockWrite, bt->set); + bt_unpinlatch (bt->set); + bt_unpinpool (bt->pool); + + // add killed right block to free chain + // lock latch mgr + + bt_spinwritelock(bt->mgr->latchmgr->lock); + + // store free chain in allocation page second right + bt_putid(rpage->right, bt_getid(bt->mgr->latchmgr->alloc[1].right)); + bt_putid(bt->mgr->latchmgr->alloc[1].right, right); + + // unlock latch mgr and right page + + bt_spinreleasewrite(bt->mgr->latchmgr->lock); + + bt_unlockpage(BtLockWrite, rset); + bt_unlockpage(BtLockDelete, rset); + bt_unpinlatch (rset); + bt_unpinpool (rpool); + + // 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 + +uid bt_findkey (BtDb *bt, unsigned char *key, uint len) +{ +uint slot; +BtKey ptr; +uid id; + + if( slot = bt_loadpage (bt, key, len, 0, BtLockRead) ) + ptr = keyptr(bt->page, slot); + else + return 0; + + // if key exists, return row-id + // otherwise return 0 + + if( slot <= bt->page->cnt && !keycmp (ptr, key, len) ) + id = bt_getid(slotptr(bt->page,slot)->id); + else + id = 0; + + bt_unlockpage (BtLockRead, bt->set); + bt_unpinlatch (bt->set); + bt_unpinpool (bt->pool); + return id; +} + +// check page for space available, +// clean if necessary and return +// 0 - page needs splitting +// >0 new slot value + +uint bt_cleanpage(BtDb *bt, 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; +BtKey key; + + if( page->min >= (max+1) * sizeof(BtSlot) + sizeof(*page) + amt + 1 ) + return slot; + + // skip cleanup if nothing to reclaim + + if( !page->dirty ) + return 0; + + memcpy (bt->frame, page, bt->mgr->page_size); + + // skip page info and set rest of page to zero + + memset (page+1, 0, bt->mgr->page_size - sizeof(*page)); + 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 + + while( cnt++ < max ) { + if( cnt == slot ) + newslot = idx + 1; + else if( cnt < max && slotptr(bt->frame,cnt)->dead ) + continue; + + // copy key + + 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; + + // 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) +{ +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 + + 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 + + memset (root+1, 0, bt->mgr->page_size - sizeof(*root)); + + // insert left fence key on empty newroot page + + nxt -= *fencekey + 1; + memcpy ((unsigned char *)root + nxt, fencekey, *fencekey + 1); + bt_putid(slotptr(root, 1)->id, new_page); + slotptr(root, 1)->off = nxt; + + // insert stopper 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); + slotptr(root, 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++; + + // release and unpin root (bt->page) + + bt_unlockpage(BtLockWrite, bt->set); + bt_unpinlatch (bt->set); + bt_unpinpool (bt->pool); + return 0; +} + +// split already locked full node +// in current page variables +// return unlocked and unpinned. + +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; +BtPool *pool = bt->pool; +BtPage page = bt->page; +uint tod = time(NULL); +uint lvl = page->lvl; +uid new_page, right; +BtKey key; + + // initialize frame buffer for right node + + memset (bt->frame, 0, bt->mgr->page_size); + max = page->cnt - page->foster; + tod = (uint)time(NULL); + 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; + 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) ) + bt->frame->act++; + slotptr(bt->frame, idx)->tod = slotptr(page, cnt)->tod; + slotptr(bt->frame, idx)->off = nxt; + } + + // transfer right link node to new right node + + if( page_no > ROOT_page ) { + right = bt_getid (page->right); + bt_putid(bt->frame->right, right); + } + + 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. + + if( !(new_page = bt_newpage(bt, bt->frame)) ) + return bt->err; + + // remember fence key for new right page to add + // as foster child to the left node + + key = keyptr(bt->frame, idx); + memcpy (fencekey, key, key->len + 1); + + // update lower keys and foster children 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 + + 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); + slotptr(page, idx)->tod = slotptr(bt->frame, cnt)->tod; + slotptr(page, idx)->off = nxt; + page->act++; + } + + 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); + + // release wr lock on our page + + bt_unlockpage (BtLockWrite, set); + + // obtain ParentModification lock for current page + // to fix new fence key and oldest foster child on page + + bt_lockpage (BtLockParent, set); + + // 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_unlockpage (BtLockRead, set); + + if( bt_insertkey (bt, fencekey + 1, *fencekey, page_no, tod, lvl + 1) ) + return bt->err; + + // lock our page for writing + + bt_lockpage (BtLockRead, set); + + // switch old parent key from us to our oldest foster child + + 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) ) + 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--; + + // unlock and unpin + + bt_unlockpage (BtLockWrite, set); + bt_unlockpage (BtLockParent, set); + bt_unpinlatch (set); + bt_unpinpool (pool); + return 0; +} + +// Insert new key into the btree at leaf level. + +BTERR bt_insertkey (BtDb *bt, unsigned char *key, uint len, uid id, uint tod, uint lvl) +{ +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; + + 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; + } + + // check if page has enough space + + if( slot = bt_cleanpage (bt, len, slot) ) + break; + + if( bt_splitpage (bt) ) + return bt->err; + } + + bt_addkeytopage (bt, slot, key, len, id, tod); + + 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 + +uint bt_startkey (BtDb *bt, unsigned char *key, uint len) +{ +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); + return slot; +} + +// return next slot for cursor page +// or slide cursor right into next page + +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 ) + if( slotptr(bt->cursor,slot)->dead ) + continue; + else if( right || (slot < bt->cursor->cnt - bt->cursor->foster) ) + return slot; + else + break; + + if( !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); + + memcpy (bt->cursor, page, bt->mgr->page_size); + + bt_unlockpage(BtLockRead, set); + bt_unpinlatch (set); + bt_unpinpool (pool); + slot = 0; + } while( 1 ); + + return bt->err = 0; +} + +BtKey bt_key(BtDb *bt, uint slot) +{ + return keyptr(bt->cursor, slot); +} + +uid bt_uid(BtDb *bt, uint slot) +{ + return bt_getid(slotptr(bt->cursor,slot)->id); +} + +uint bt_tod(BtDb *bt, uint slot) +{ + return slotptr(bt->cursor,slot)->tod; +} + + +#ifdef STANDALONE + +void bt_latchaudit (BtDb *bt) +{ +ushort idx, hashidx; +BtLatchSet *set; +BtPool *pool; +BtPage page; +uid page_no; + +#ifdef unix + for( idx = 1; idx < bt->mgr->latchmgr->latchdeployed; idx++ ) { + set = bt->mgr->latchsets + idx; + if( *(ushort *)set->readwr || *(ushort *)set->access || *(ushort *)set->parent ) { + fprintf(stderr, "latchset %d locked for page %6x\n", idx, set->page_no); + *(ushort *)set->readwr = 0; + *(ushort *)set->access = 0; + *(ushort *)set->parent = 0; + } + if( set->pin ) { + fprintf(stderr, "latchset %d pinned\n", idx); + set->pin = 0; + } + } + + for( hashidx = 0; hashidx < bt->mgr->latchmgr->latchhash; hashidx++ ) { + if( *(ushort *)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 ) + fprintf(stderr, "latchset %d locked\n", idx); + if( set->hash != hashidx ) + fprintf(stderr, "latchset %d wrong hashidx\n", idx); + if( set->pin ) + fprintf(stderr, "latchset %d pinned\n", idx); + } while( idx = set->next ); + } + page_no = bt_getid(bt->mgr->latchmgr->alloc[1].right); + + while( page_no ) { + fprintf(stderr, "free: %.6x\n", (uint)page_no); + pool = bt_pinpool (bt, page_no); + page = bt_page (bt, pool, page_no); + page_no = bt_getid(page->right); + bt_unpinpool (pool); + } +#endif +} + +typedef struct { + char type, idx; + char *infile; + BtMgr *mgr; + int num; +} ThreadArg; + +// standalone program to index file of keys +// then list them onto std-out + +#ifdef unix +void *index_file (void *arg) +#else +uint __stdcall index_file (void *arg) +#endif +{ +int line = 0, found = 0, cnt = 0; +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; +BtKey ptr; +BtDb *bt; +FILE *in; + + bt = bt_open (args->mgr); + time (tod); + + 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") ) + while( ch = getc(in), ch != EOF ) + if( ch == '\n' ) + { + 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; + + if( bt_insertkey (bt, key, len, line, *tod, 0) ) + fprintf(stderr, "Error %d Line: %d\n", bt->err, line), exit(0); + len = 0; + } + else if( len < 255 ) + key[len++] = ch; + fprintf(stderr, "finished %s for %d keys\n", args->infile, line); + break; + + case 'd': + fprintf(stderr, "started deleting keys for %s\n", args->infile); + if( in = fopen (args->infile, "rb") ) + while( ch = getc(in), ch != EOF ) + if( ch == '\n' ) + { + 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; + + if( bt_deletekey (bt, key, len) ) + fprintf(stderr, "Error %d Line: %d\n", bt->err, line), exit(0); + len = 0; + } + else if( len < 255 ) + key[len++] = ch; + fprintf(stderr, "finished %s for keys, %d \n", args->infile, line); + break; + + case 'f': + fprintf(stderr, "started finding keys for %s\n", args->infile); + if( in = fopen (args->infile, "rb") ) + while( ch = getc(in), ch != EOF ) + if( ch == '\n' ) + { + 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; + + if( bt_findkey (bt, key, len) ) + found++; + else if( bt->err ) + fprintf(stderr, "Error %d Syserr %d Line: %d\n", bt->err, errno, line), exit(0); + len = 0; + } + else if( len < 255 ) + key[len++] = ch; + fprintf(stderr, "finished %s for %d keys, found %d\n", args->infile, line, found); + 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); + } + + break; + + case 'c': + fprintf(stderr, "started reading\n"); + + do { + if( pool = bt_pinpool (bt, page_no) ) + page = bt_page (bt, pool, page_no); + else + break; + set = bt_pinlatch (bt, page_no); + bt_lockpage (BtLockRead, set); + cnt += page->act; + next = bt_getid (page->right); + bt_unlockpage (BtLockRead, set); + bt_unpinlatch (set); + bt_unpinpool (pool); + } while( page_no = next ); + + cnt--; // remove stopper key + fprintf(stderr, " Total keys read %d\n", cnt); + break; + } + + bt_close (bt); +#ifdef unix + return NULL; +#else + return 0; +#endif +} + +typedef struct timeval timer; + +int main (int argc, char **argv) +{ +int idx, cnt, len, slot, err; +int segsize, bits = 16; +#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; +int num = 0; +char key[1]; +BtMgr *mgr; +BtKey ptr; +BtDb *bt; + + if( argc < 3 ) { + fprintf (stderr, "Usage: %s idx_file Read/Write/Scan/Delete/Find [page_bits mapped_segments seg_bits line_numbers src_file1 src_file2 ... ]\n", argv[0]); + 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 = 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 + + if( argc > 3 ) + bits = atoi(argv[3]); + + if( argc > 4 ) + poolsize = atoi(argv[4]); + + if( !poolsize ) + fprintf (stderr, "Warning: no mapped_pool\n"); + + if( poolsize > 65535 ) + fprintf (stderr, "Warning: mapped_pool > 65535 segments\n"); + + if( argc > 5 ) + segsize = atoi(argv[5]); + else + segsize = 4; // 16 pages per mmap segment + + if( argc > 6 ) + num = atoi(argv[6]); + + cnt = argc - 7; +#ifdef unix + threads = malloc (cnt * sizeof(pthread_t)); +#else + threads = GlobalAlloc (GMEM_FIXED|GMEM_ZEROINIT, cnt * sizeof(HANDLE)); +#endif + args = malloc (cnt * sizeof(ThreadArg)); + + mgr = bt_mgr ((argv[1]), BT_rw, bits, poolsize, segsize, poolsize / 8); + + if( !mgr ) { + fprintf(stderr, "Index Open Error %s\n", argv[1]); + exit (1); + } + + // fire off threads + + for( idx = 0; idx < cnt; idx++ ) { + args[idx].infile = argv[idx + 7]; + args[idx].type = argv[2][0]; + args[idx].mgr = mgr; + args[idx].num = num; + args[idx].idx = idx; +#ifdef unix + if( err = pthread_create (threads + idx, NULL, index_file, args + idx) ) + fprintf(stderr, "Error creating thread %d\n", err); +#else + threads[idx] = (HANDLE)_beginthreadex(NULL, 65536, index_file, args + idx, 0, NULL); +#endif + } + + // wait for termination + +#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); + bt_mgrclose (mgr); +} + +#endif //STANDALONE -- 2.40.0