1 // btree version threads2h pthread rw lock/SRW version
2 // with fixed bt_deletekey code
5 // author: karl malbrain, malbrain@cal.berkeley.edu
8 This work, including the source code, documentation
9 and related data, is placed into the public domain.
11 The orginal author is Karl Malbrain.
13 THIS SOFTWARE IS PROVIDED AS-IS WITHOUT WARRANTY
14 OF ANY KIND, NOT EVEN THE IMPLIED WARRANTY OF
15 MERCHANTABILITY. THE AUTHOR OF THIS SOFTWARE,
16 ASSUMES _NO_ RESPONSIBILITY FOR ANY CONSEQUENCE
17 RESULTING FROM THE USE, MODIFICATION, OR
18 REDISTRIBUTION OF THIS SOFTWARE.
21 // Please see the project home page for documentation
22 // code.google.com/p/high-concurrency-btree
24 #define _FILE_OFFSET_BITS 64
25 #define _LARGEFILE64_SOURCE
41 #define WIN32_LEAN_AND_MEAN
55 typedef unsigned long long uid;
58 typedef unsigned long long off64_t;
59 typedef unsigned short ushort;
60 typedef unsigned int uint;
63 #define BT_latchtable 128 // number of latch manager slots
65 #define BT_ro 0x6f72 // ro
66 #define BT_rw 0x7772 // rw
68 #define BT_maxbits 24 // maximum page size in bits
69 #define BT_minbits 9 // minimum page size in bits
70 #define BT_minpage (1 << BT_minbits) // minimum page size
71 #define BT_maxpage (1 << BT_maxbits) // maximum page size
74 There are five lock types for each node in three independent sets:
75 1. (set 1) AccessIntent: Sharable. Going to Read the node. Incompatible with NodeDelete.
76 2. (set 1) NodeDelete: Exclusive. About to release the node. Incompatible with AccessIntent.
77 3. (set 2) ReadLock: Sharable. Read the node. Incompatible with WriteLock.
78 4. (set 2) WriteLock: Exclusive. Modify the node. Incompatible with ReadLock and other WriteLocks.
79 5. (set 3) ParentModification: Exclusive. Change the node's parent keys. Incompatible with another ParentModification.
90 // mode & definition for latch implementation
92 // exclusive is set for write access
93 // share is count of read accessors
94 // grant write lock when share == 0
97 volatile unsigned char mutex;
98 volatile unsigned char exclusive:1;
99 volatile unsigned char pending:1;
100 volatile ushort share;
103 // hash table entries
106 BtSpinLatch latch[1];
107 volatile ushort slot; // Latch table entry at head of chain
110 // latch manager table structure
114 pthread_rwlock_t lock[1];
121 BtLatch readwr[1]; // read/write page lock
122 BtLatch access[1]; // Access Intent/Page delete
123 BtLatch parent[1]; // Posting of fence key in parent
124 BtSpinLatch busy[1]; // slot is being moved between chains
125 volatile ushort next; // next entry in hash table chain
126 volatile ushort prev; // prev entry in hash table chain
127 volatile ushort pin; // number of outstanding locks
128 volatile ushort hash; // hash slot entry is under
129 volatile uid page_no; // latch set page number
132 // Define the length of the page and key pointers
136 // Page key slot definition.
138 // If BT_maxbits is 15 or less, you can save 4 bytes
139 // for each key stored by making the first two uints
140 // into ushorts. You can also save 4 bytes by removing
141 // the tod field from the key.
143 // Keys are marked dead, but remain on the page until
144 // it cleanup is called. The fence key (highest key) for
145 // the page is always present, even after cleanup.
148 uint off:BT_maxbits; // page offset for key start
149 uint dead:1; // set for deleted key
150 uint tod; // time-stamp for key
151 unsigned char id[BtId]; // id associated with key
154 // The key structure occupies space at the upper end of
155 // each page. It's a length byte followed by the value
160 unsigned char key[1];
163 // The first part of an index page.
164 // It is immediately followed
165 // by the BtSlot array of keys.
167 typedef struct BtPage_ {
168 uint cnt; // count of keys in page
169 uint act; // count of active keys
170 uint min; // next key offset
171 unsigned char bits:7; // page size in bits
172 unsigned char free:1; // page is on free list
173 unsigned char lvl:6; // level of page
174 unsigned char kill:1; // page is being killed
175 unsigned char dirty:1; // page has deleted keys
176 unsigned char right[BtId]; // page number to right
179 // The memory mapping pool table buffer manager entry
182 unsigned long long int lru; // number of times accessed
183 uid basepage; // mapped base page number
184 char *map; // mapped memory pointer
185 ushort slot; // slot index in this array
186 ushort pin; // mapped page pin counter
187 void *hashprev; // previous pool entry for the same hash idx
188 void *hashnext; // next pool entry for the same hash idx
190 HANDLE hmap; // Windows memory mapping handle
194 // The loadpage interface object
197 uid page_no; // current page number
198 BtPage page; // current page pointer
199 BtPool *pool; // current page pool
200 BtLatchSet *latch; // current page latch set
203 // structure for latch manager on ALLOC_page
206 struct BtPage_ alloc[2]; // next & free page_nos in right ptr
207 BtSpinLatch lock[1]; // allocation area lite latch
208 ushort latchdeployed; // highest number of latch entries deployed
209 ushort nlatchpage; // number of latch pages at BT_latch
210 ushort latchtotal; // number of page latch entries
211 ushort latchhash; // number of latch hash table slots
212 ushort latchvictim; // next latch entry to examine
213 BtHashEntry table[0]; // the hash table
216 // The object structure for Btree access
219 uint page_size; // page size
220 uint page_bits; // page size in bits
221 uint seg_bits; // seg size in pages in bits
222 uint mode; // read-write mode
228 ushort poolcnt; // highest page pool node in use
229 ushort poolmax; // highest page pool node allocated
230 ushort poolmask; // total number of pages in mmap segment - 1
231 ushort hashsize; // size of Hash Table for pool entries
232 volatile uint evicted; // last evicted hash table slot
233 ushort *hash; // pool index for hash entries
234 BtSpinLatch *latch; // latches for hash table slots
235 BtLatchMgr *latchmgr; // mapped latch page from allocation page
236 BtLatchSet *latchsets; // mapped latch set from latch pages
237 BtPool *pool; // memory pool page segments
239 HANDLE halloc; // allocation and latch table handle
244 BtMgr *mgr; // buffer manager for thread
245 BtPage cursor; // cached frame for start/next (never mapped)
246 BtPage frame; // spare frame for the page split (never mapped)
247 BtPage zero; // page frame for zeroes at end of file
248 uid cursor_page; // current cursor page number
249 unsigned char *mem; // frame, cursor, page memory buffer
250 int found; // last delete or insert was found
251 int err; // last error
265 extern void bt_close (BtDb *bt);
266 extern BtDb *bt_open (BtMgr *mgr);
267 extern BTERR bt_insertkey (BtDb *bt, unsigned char *key, uint len, uint lvl, uid id, uint tod);
268 extern BTERR bt_deletekey (BtDb *bt, unsigned char *key, uint len, uint lvl);
269 extern uid bt_findkey (BtDb *bt, unsigned char *key, uint len);
270 extern uint bt_startkey (BtDb *bt, unsigned char *key, uint len);
271 extern uint bt_nextkey (BtDb *bt, uint slot);
274 extern BtMgr *bt_mgr (char *name, uint mode, uint bits, uint poolsize, uint segsize, uint hashsize);
275 void bt_mgrclose (BtMgr *mgr);
277 // Helper functions to return slot values
279 extern BtKey bt_key (BtDb *bt, uint slot);
280 extern uid bt_uid (BtDb *bt, uint slot);
281 extern uint bt_tod (BtDb *bt, uint slot);
283 // BTree page number constants
284 #define ALLOC_page 0 // allocation & lock manager hash table
285 #define ROOT_page 1 // root of the btree
286 #define LEAF_page 2 // first page of leaves
287 #define LATCH_page 3 // pages for lock manager
289 // Number of levels to create in a new BTree
293 // The page is allocated from low and hi ends.
294 // The key offsets and row-id's are allocated
295 // from the bottom, while the text of the key
296 // is allocated from the top. When the two
297 // areas meet, the page is split into two.
299 // A key consists of a length byte, two bytes of
300 // index number (0 - 65534), and up to 253 bytes
301 // of key value. Duplicate keys are discarded.
302 // Associated with each key is a 48 bit row-id.
304 // The b-tree root is always located at page 1.
305 // The first leaf page of level zero is always
306 // located on page 2.
308 // The b-tree pages are linked with next
309 // pointers to facilitate enumerators,
310 // and provide for concurrency.
312 // When to root page fills, it is split in two and
313 // the tree height is raised by a new root at page
314 // one with two keys.
316 // Deleted keys are marked with a dead bit until
317 // page cleanup The fence key for a node is
318 // present in a special array.
320 // Groups of pages called segments from the btree are optionally
321 // cached with a memory mapped pool. A hash table is used to keep
322 // track of the cached segments. This behaviour is controlled
323 // by the cache block size parameter to bt_open.
325 // To achieve maximum concurrency one page is locked at a time
326 // as the tree is traversed to find leaf key in question. The right
327 // page numbers are used in cases where the page is being split,
330 // Page 0 is dedicated to lock for new page extensions,
331 // and chains empty pages together for reuse.
333 // The ParentModification lock on a node is obtained to serialize posting
334 // or changing the fence key for a node.
336 // Empty pages are chained together through the ALLOC page and reused.
338 // Access macros to address slot and key values from the page.
339 // Page slots use 1 based indexing.
341 #define slotptr(page, slot) (((BtSlot *)(page+1)) + (slot-1))
342 #define keyptr(page, slot) ((BtKey)((unsigned char*)(page) + slotptr(page, slot)->off))
344 void bt_putid(unsigned char *dest, uid id)
349 dest[i] = (unsigned char)id, id >>= 8;
352 uid bt_getid(unsigned char *src)
357 for( i = 0; i < BtId; i++ )
358 id <<= 8, id |= *src++;
365 // wait until write lock mode is clear
366 // and add 1 to the share count
368 void bt_spinreadlock(BtSpinLatch *latch)
373 // obtain latch mutex
375 if( __sync_lock_test_and_set(&latch->mutex, 1) )
378 if( _InterlockedExchange8(&latch->mutex, 1) )
381 // see if exclusive request is granted or pending
383 if( prev = !(latch->exclusive | latch->pending) )
387 __sync_lock_release (&latch->mutex);
389 _InterlockedExchange8(&latch->mutex, 0);
396 } while( sched_yield(), 1 );
398 } while( SwitchToThread(), 1 );
402 // wait for other read and write latches to relinquish
404 void bt_spinwritelock(BtSpinLatch *latch)
410 if( __sync_lock_test_and_set(&latch->mutex, 1) )
413 if( _InterlockedExchange8(&latch->mutex, 1) )
416 if( prev = !(latch->share | latch->exclusive) )
417 latch->exclusive = 1, latch->pending = 0;
421 __sync_lock_release (&latch->mutex);
423 _InterlockedExchange8(&latch->mutex, 0);
428 } while( sched_yield(), 1 );
430 } while( SwitchToThread(), 1 );
434 // try to obtain write lock
436 // return 1 if obtained,
439 int bt_spinwritetry(BtSpinLatch *latch)
444 if( __sync_lock_test_and_set(&latch->mutex, 1) )
447 if( _InterlockedExchange8(&latch->mutex, 1) )
450 // take write access if all bits are clear
452 if( prev = !(latch->exclusive | latch->share) )
453 latch->exclusive = 1;
456 __sync_lock_release (&latch->mutex);
458 _InterlockedExchange8(&latch->mutex, 0);
465 void bt_spinreleasewrite(BtSpinLatch *latch)
467 // obtain latch mutex
469 while( __sync_lock_test_and_set(&latch->mutex, 1) )
472 while( _InterlockedExchange8(&latch->mutex, 1) )
475 latch->exclusive = 0;
477 __sync_lock_release (&latch->mutex);
479 _InterlockedExchange8(&latch->mutex, 0);
483 // decrement reader count
485 void bt_spinreleaseread(BtSpinLatch *latch)
488 while( __sync_lock_test_and_set(&latch->mutex, 1) )
491 while( _InterlockedExchange8(&latch->mutex, 1) )
496 __sync_lock_release (&latch->mutex);
498 _InterlockedExchange8(&latch->mutex, 0);
502 void bt_readlock(BtLatch *latch)
505 pthread_rwlock_rdlock (latch->lock);
507 AcquireSRWLockShared (latch->srw);
511 // wait for other read and write latches to relinquish
513 void bt_writelock(BtLatch *latch)
516 pthread_rwlock_wrlock (latch->lock);
518 AcquireSRWLockExclusive (latch->srw);
522 // try to obtain write lock
524 // return 1 if obtained,
525 // 0 if already write or read locked
527 int bt_writetry(BtLatch *latch)
532 result = !pthread_rwlock_trywrlock (latch->lock);
534 result = TryAcquireSRWLockExclusive (latch->srw);
541 void bt_releasewrite(BtLatch *latch)
544 pthread_rwlock_unlock (latch->lock);
546 ReleaseSRWLockExclusive (latch->srw);
550 // decrement reader count
552 void bt_releaseread(BtLatch *latch)
555 pthread_rwlock_unlock (latch->lock);
557 ReleaseSRWLockShared (latch->srw);
561 void bt_initlockset (BtLatchSet *set)
564 pthread_rwlockattr_t rwattr[1];
566 pthread_rwlockattr_init (rwattr);
567 pthread_rwlockattr_setpshared (rwattr, PTHREAD_PROCESS_SHARED);
569 pthread_rwlock_init (set->readwr->lock, rwattr);
570 pthread_rwlock_init (set->access->lock, rwattr);
571 pthread_rwlock_init (set->parent->lock, rwattr);
572 pthread_rwlockattr_destroy (rwattr);
574 InitializeSRWLock (set->readwr->srw);
575 InitializeSRWLock (set->access->srw);
576 InitializeSRWLock (set->parent->srw);
580 // link latch table entry into latch hash table
582 void bt_latchlink (BtDb *bt, ushort hashidx, ushort victim, uid page_no)
584 BtLatchSet *set = bt->mgr->latchsets + victim;
586 if( set->next = bt->mgr->latchmgr->table[hashidx].slot )
587 bt->mgr->latchsets[set->next].prev = victim;
589 bt->mgr->latchmgr->table[hashidx].slot = victim;
590 set->page_no = page_no;
597 void bt_unpinlatch (BtLatchSet *set)
600 __sync_fetch_and_add(&set->pin, -1);
602 _InterlockedDecrement16 (&set->pin);
606 // find existing latchset or inspire new one
607 // return with latchset pinned
609 BtLatchSet *bt_pinlatch (BtDb *bt, uid page_no)
611 ushort hashidx = page_no % bt->mgr->latchmgr->latchhash;
612 ushort slot, avail = 0, victim, idx;
615 // obtain read lock on hash table entry
617 bt_spinreadlock(bt->mgr->latchmgr->table[hashidx].latch);
619 if( slot = bt->mgr->latchmgr->table[hashidx].slot ) do
621 set = bt->mgr->latchsets + slot;
622 if( page_no == set->page_no )
624 } while( slot = set->next );
628 __sync_fetch_and_add(&set->pin, 1);
630 _InterlockedIncrement16 (&set->pin);
634 bt_spinreleaseread (bt->mgr->latchmgr->table[hashidx].latch);
639 // try again, this time with write lock
641 bt_spinwritelock(bt->mgr->latchmgr->table[hashidx].latch);
643 if( slot = bt->mgr->latchmgr->table[hashidx].slot ) do
645 set = bt->mgr->latchsets + slot;
646 if( page_no == set->page_no )
648 if( !set->pin && !avail )
650 } while( slot = set->next );
652 // found our entry, or take over an unpinned one
654 if( slot || (slot = avail) ) {
655 set = bt->mgr->latchsets + slot;
657 __sync_fetch_and_add(&set->pin, 1);
659 _InterlockedIncrement16 (&set->pin);
661 set->page_no = page_no;
662 bt_spinreleasewrite(bt->mgr->latchmgr->table[hashidx].latch);
666 // see if there are any unused entries
668 victim = __sync_fetch_and_add (&bt->mgr->latchmgr->latchdeployed, 1) + 1;
670 victim = _InterlockedIncrement16 (&bt->mgr->latchmgr->latchdeployed);
673 if( victim < bt->mgr->latchmgr->latchtotal ) {
674 set = bt->mgr->latchsets + victim;
676 __sync_fetch_and_add(&set->pin, 1);
678 _InterlockedIncrement16 (&set->pin);
680 bt_initlockset (set);
681 bt_latchlink (bt, hashidx, victim, page_no);
682 bt_spinreleasewrite (bt->mgr->latchmgr->table[hashidx].latch);
687 victim = __sync_fetch_and_add (&bt->mgr->latchmgr->latchdeployed, -1);
689 victim = _InterlockedDecrement16 (&bt->mgr->latchmgr->latchdeployed);
691 // find and reuse previous lock entry
695 victim = __sync_fetch_and_add(&bt->mgr->latchmgr->latchvictim, 1);
697 victim = _InterlockedIncrement16 (&bt->mgr->latchmgr->latchvictim) - 1;
699 // we don't use slot zero
701 if( victim %= bt->mgr->latchmgr->latchtotal )
702 set = bt->mgr->latchsets + victim;
706 // take control of our slot
707 // from other threads
709 if( set->pin || !bt_spinwritetry (set->busy) )
714 // try to get write lock on hash chain
715 // skip entry if not obtained
716 // or has outstanding locks
718 if( !bt_spinwritetry (bt->mgr->latchmgr->table[idx].latch) ) {
719 bt_spinreleasewrite (set->busy);
724 bt_spinreleasewrite (set->busy);
725 bt_spinreleasewrite (bt->mgr->latchmgr->table[idx].latch);
729 // unlink our available victim from its hash chain
732 bt->mgr->latchsets[set->prev].next = set->next;
734 bt->mgr->latchmgr->table[idx].slot = set->next;
737 bt->mgr->latchsets[set->next].prev = set->prev;
739 bt_spinreleasewrite (bt->mgr->latchmgr->table[idx].latch);
741 __sync_fetch_and_add(&set->pin, 1);
743 _InterlockedIncrement16 (&set->pin);
745 bt_latchlink (bt, hashidx, victim, page_no);
746 bt_spinreleasewrite (bt->mgr->latchmgr->table[hashidx].latch);
747 bt_spinreleasewrite (set->busy);
752 void bt_mgrclose (BtMgr *mgr)
757 // release mapped pages
758 // note that slot zero is never used
760 for( slot = 1; slot < mgr->poolmax; slot++ ) {
761 pool = mgr->pool + slot;
764 munmap (pool->map, (mgr->poolmask+1) << mgr->page_bits);
767 FlushViewOfFile(pool->map, 0);
768 UnmapViewOfFile(pool->map);
769 CloseHandle(pool->hmap);
775 munmap (mgr->latchsets, mgr->latchmgr->nlatchpage * mgr->page_size);
776 munmap (mgr->latchmgr, mgr->page_size);
778 FlushViewOfFile(mgr->latchmgr, 0);
779 UnmapViewOfFile(mgr->latchmgr);
780 CloseHandle(mgr->halloc);
789 FlushFileBuffers(mgr->idx);
790 CloseHandle(mgr->idx);
791 GlobalFree (mgr->pool);
792 GlobalFree (mgr->hash);
793 GlobalFree (mgr->latch);
798 // close and release memory
800 void bt_close (BtDb *bt)
807 VirtualFree (bt->mem, 0, MEM_RELEASE);
812 // open/create new btree buffer manager
814 // call with file_name, BT_openmode, bits in page size (e.g. 16),
815 // size of mapped page pool (e.g. 8192)
817 BtMgr *bt_mgr (char *name, uint mode, uint bits, uint poolmax, uint segsize, uint hashsize)
819 uint lvl, attr, cacheblk, last, slot, idx;
820 uint nlatchpage, latchhash;
821 BtLatchMgr *latchmgr;
829 SYSTEM_INFO sysinfo[1];
832 // determine sanity of page size and buffer pool
834 if( bits > BT_maxbits )
836 else if( bits < BT_minbits )
840 return NULL; // must have buffer pool
843 mgr = calloc (1, sizeof(BtMgr));
845 mgr->idx = open ((char*)name, O_RDWR | O_CREAT, 0666);
848 return free(mgr), NULL;
850 cacheblk = 4096; // minimum mmap segment size for unix
853 mgr = GlobalAlloc (GMEM_FIXED|GMEM_ZEROINIT, sizeof(BtMgr));
854 attr = FILE_ATTRIBUTE_NORMAL;
855 mgr->idx = CreateFile(name, GENERIC_READ| GENERIC_WRITE, FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, OPEN_ALWAYS, attr, NULL);
857 if( mgr->idx == INVALID_HANDLE_VALUE )
858 return GlobalFree(mgr), NULL;
860 // normalize cacheblk to multiple of sysinfo->dwAllocationGranularity
861 GetSystemInfo(sysinfo);
862 cacheblk = sysinfo->dwAllocationGranularity;
866 latchmgr = malloc (BT_maxpage);
869 // read minimum page size to get root info
871 if( size = lseek (mgr->idx, 0L, 2) ) {
872 if( pread(mgr->idx, latchmgr, BT_minpage, 0) == BT_minpage )
873 bits = latchmgr->alloc->bits;
875 return free(mgr), free(latchmgr), NULL;
876 } else if( mode == BT_ro )
877 return free(latchmgr), bt_mgrclose (mgr), NULL;
879 latchmgr = VirtualAlloc(NULL, BT_maxpage, MEM_COMMIT, PAGE_READWRITE);
880 size = GetFileSize(mgr->idx, amt);
883 if( !ReadFile(mgr->idx, (char *)latchmgr, BT_minpage, amt, NULL) )
884 return bt_mgrclose (mgr), NULL;
885 bits = latchmgr->alloc->bits;
886 } else if( mode == BT_ro )
887 return bt_mgrclose (mgr), NULL;
890 mgr->page_size = 1 << bits;
891 mgr->page_bits = bits;
893 mgr->poolmax = poolmax;
896 if( cacheblk < mgr->page_size )
897 cacheblk = mgr->page_size;
899 // mask for partial memmaps
901 mgr->poolmask = (cacheblk >> bits) - 1;
903 // see if requested size of pages per memmap is greater
905 if( (1 << segsize) > mgr->poolmask )
906 mgr->poolmask = (1 << segsize) - 1;
910 while( (1 << mgr->seg_bits) <= mgr->poolmask )
913 mgr->hashsize = hashsize;
916 mgr->pool = calloc (poolmax, sizeof(BtPool));
917 mgr->hash = calloc (hashsize, sizeof(ushort));
918 mgr->latch = calloc (hashsize, sizeof(BtSpinLatch));
920 mgr->pool = GlobalAlloc (GMEM_FIXED|GMEM_ZEROINIT, poolmax * sizeof(BtPool));
921 mgr->hash = GlobalAlloc (GMEM_FIXED|GMEM_ZEROINIT, hashsize * sizeof(ushort));
922 mgr->latch = GlobalAlloc (GMEM_FIXED|GMEM_ZEROINIT, hashsize * sizeof(BtSpinLatch));
928 // initialize an empty b-tree with latch page, root page, page of leaves
929 // and page(s) of latches
931 memset (latchmgr, 0, 1 << bits);
932 nlatchpage = BT_latchtable / (mgr->page_size / sizeof(BtLatchSet)) + 1;
933 bt_putid(latchmgr->alloc->right, MIN_lvl+1+nlatchpage);
934 latchmgr->alloc->bits = mgr->page_bits;
936 latchmgr->nlatchpage = nlatchpage;
937 latchmgr->latchtotal = nlatchpage * (mgr->page_size / sizeof(BtLatchSet));
939 // initialize latch manager
941 latchhash = (mgr->page_size - sizeof(BtLatchMgr)) / sizeof(BtHashEntry);
943 // size of hash table = total number of latchsets
945 if( latchhash > latchmgr->latchtotal )
946 latchhash = latchmgr->latchtotal;
948 latchmgr->latchhash = latchhash;
951 if( write (mgr->idx, latchmgr, mgr->page_size) < mgr->page_size )
952 return bt_mgrclose (mgr), NULL;
954 if( !WriteFile (mgr->idx, (char *)latchmgr, mgr->page_size, amt, NULL) )
955 return bt_mgrclose (mgr), NULL;
957 if( *amt < mgr->page_size )
958 return bt_mgrclose (mgr), NULL;
961 memset (latchmgr, 0, 1 << bits);
962 latchmgr->alloc->bits = mgr->page_bits;
964 for( lvl=MIN_lvl; lvl--; ) {
965 slotptr(latchmgr->alloc, 1)->off = mgr->page_size - 3;
966 bt_putid(slotptr(latchmgr->alloc, 1)->id, lvl ? MIN_lvl - lvl + 1 : 0); // next(lower) page number
967 key = keyptr(latchmgr->alloc, 1);
968 key->len = 2; // create stopper key
971 latchmgr->alloc->min = mgr->page_size - 3;
972 latchmgr->alloc->lvl = lvl;
973 latchmgr->alloc->cnt = 1;
974 latchmgr->alloc->act = 1;
976 if( write (mgr->idx, latchmgr, mgr->page_size) < mgr->page_size )
977 return bt_mgrclose (mgr), NULL;
979 if( !WriteFile (mgr->idx, (char *)latchmgr, mgr->page_size, amt, NULL) )
980 return bt_mgrclose (mgr), NULL;
982 if( *amt < mgr->page_size )
983 return bt_mgrclose (mgr), NULL;
987 // clear out latch manager locks
988 // and rest of pages to round out segment
990 memset(latchmgr, 0, mgr->page_size);
993 while( last <= ((MIN_lvl + 1 + nlatchpage) | mgr->poolmask) ) {
995 pwrite(mgr->idx, latchmgr, mgr->page_size, last << mgr->page_bits);
997 SetFilePointer (mgr->idx, last << mgr->page_bits, NULL, FILE_BEGIN);
998 if( !WriteFile (mgr->idx, (char *)latchmgr, mgr->page_size, amt, NULL) )
999 return bt_mgrclose (mgr), NULL;
1000 if( *amt < mgr->page_size )
1001 return bt_mgrclose (mgr), NULL;
1008 flag = PROT_READ | PROT_WRITE;
1009 mgr->latchmgr = mmap (0, mgr->page_size, flag, MAP_SHARED, mgr->idx, ALLOC_page * mgr->page_size);
1010 if( mgr->latchmgr == MAP_FAILED )
1011 return bt_mgrclose (mgr), NULL;
1012 mgr->latchsets = (BtLatchSet *)mmap (0, mgr->latchmgr->nlatchpage * mgr->page_size, flag, MAP_SHARED, mgr->idx, LATCH_page * mgr->page_size);
1013 if( mgr->latchsets == MAP_FAILED )
1014 return bt_mgrclose (mgr), NULL;
1016 flag = PAGE_READWRITE;
1017 mgr->halloc = CreateFileMapping(mgr->idx, NULL, flag, 0, (BT_latchtable / (mgr->page_size / sizeof(BtLatchSet)) + 1 + LATCH_page) * mgr->page_size, NULL);
1019 return bt_mgrclose (mgr), NULL;
1021 flag = FILE_MAP_WRITE;
1022 mgr->latchmgr = MapViewOfFile(mgr->halloc, flag, 0, 0, (BT_latchtable / (mgr->page_size / sizeof(BtLatchSet)) + 1 + LATCH_page) * mgr->page_size);
1023 if( !mgr->latchmgr )
1024 return GetLastError(), bt_mgrclose (mgr), NULL;
1026 mgr->latchsets = (void *)((char *)mgr->latchmgr + LATCH_page * mgr->page_size);
1032 VirtualFree (latchmgr, 0, MEM_RELEASE);
1037 // open BTree access method
1038 // based on buffer manager
1040 BtDb *bt_open (BtMgr *mgr)
1042 BtDb *bt = malloc (sizeof(*bt));
1044 memset (bt, 0, sizeof(*bt));
1047 bt->mem = malloc (3 *mgr->page_size);
1049 bt->mem = VirtualAlloc(NULL, 3 * mgr->page_size, MEM_COMMIT, PAGE_READWRITE);
1051 bt->frame = (BtPage)bt->mem;
1052 bt->zero = (BtPage)(bt->mem + 1 * mgr->page_size);
1053 bt->cursor = (BtPage)(bt->mem + 2 * mgr->page_size);
1055 memset (bt->zero, 0, mgr->page_size);
1059 // compare two keys, returning > 0, = 0, or < 0
1060 // as the comparison value
1062 int keycmp (BtKey key1, unsigned char *key2, uint len2)
1064 uint len1 = key1->len;
1067 if( ans = memcmp (key1->key, key2, len1 > len2 ? len2 : len1) )
1080 // find segment in pool
1081 // must be called with hashslot idx locked
1082 // return NULL if not there
1083 // otherwise return node
1085 BtPool *bt_findpool(BtDb *bt, uid page_no, uint idx)
1090 // compute start of hash chain in pool
1092 if( slot = bt->mgr->hash[idx] )
1093 pool = bt->mgr->pool + slot;
1097 page_no &= ~bt->mgr->poolmask;
1099 while( pool->basepage != page_no )
1100 if( pool = pool->hashnext )
1108 // add segment to hash table
1110 void bt_linkhash(BtDb *bt, BtPool *pool, uid page_no, int idx)
1115 pool->hashprev = pool->hashnext = NULL;
1116 pool->basepage = page_no & ~bt->mgr->poolmask;
1119 if( slot = bt->mgr->hash[idx] ) {
1120 node = bt->mgr->pool + slot;
1121 pool->hashnext = node;
1122 node->hashprev = pool;
1125 bt->mgr->hash[idx] = pool->slot;
1128 // find best segment to evict from buffer pool
1130 BtPool *bt_findlru (BtDb *bt, uint hashslot)
1132 unsigned long long int target = ~0LL;
1133 BtPool *pool = NULL, *node;
1138 node = bt->mgr->pool + hashslot;
1140 // scan pool entries under hash table slot
1145 if( node->lru > target )
1149 } while( node = node->hashnext );
1154 // map new buffer pool segment to virtual memory
1156 BTERR bt_mapsegment(BtDb *bt, BtPool *pool, uid page_no)
1158 off64_t off = (page_no & ~bt->mgr->poolmask) << bt->mgr->page_bits;
1159 off64_t limit = off + ((bt->mgr->poolmask+1) << bt->mgr->page_bits);
1163 flag = PROT_READ | ( bt->mgr->mode == BT_ro ? 0 : PROT_WRITE );
1164 pool->map = mmap (0, (bt->mgr->poolmask+1) << bt->mgr->page_bits, flag, MAP_SHARED, bt->mgr->idx, off);
1165 if( pool->map == MAP_FAILED )
1166 return bt->err = BTERR_map;
1168 flag = ( bt->mgr->mode == BT_ro ? PAGE_READONLY : PAGE_READWRITE );
1169 pool->hmap = CreateFileMapping(bt->mgr->idx, NULL, flag, (DWORD)(limit >> 32), (DWORD)limit, NULL);
1171 return bt->err = BTERR_map;
1173 flag = ( bt->mgr->mode == BT_ro ? FILE_MAP_READ : FILE_MAP_WRITE );
1174 pool->map = MapViewOfFile(pool->hmap, flag, (DWORD)(off >> 32), (DWORD)off, (bt->mgr->poolmask+1) << bt->mgr->page_bits);
1176 return bt->err = BTERR_map;
1181 // calculate page within pool
1183 BtPage bt_page (BtDb *bt, BtPool *pool, uid page_no)
1185 uint subpage = (uint)(page_no & bt->mgr->poolmask); // page within mapping
1188 page = (BtPage)(pool->map + (subpage << bt->mgr->page_bits));
1194 void bt_unpinpool (BtPool *pool)
1197 __sync_fetch_and_add(&pool->pin, -1);
1199 _InterlockedDecrement16 (&pool->pin);
1203 // find or place requested page in segment-pool
1204 // return pool table entry, incrementing pin
1206 BtPool *bt_pinpool(BtDb *bt, uid page_no)
1208 BtPool *pool, *node, *next;
1209 uint slot, idx, victim;
1211 // lock hash table chain
1213 idx = (uint)(page_no >> bt->mgr->seg_bits) % bt->mgr->hashsize;
1214 bt_spinreadlock (&bt->mgr->latch[idx]);
1216 // look up in hash table
1218 if( pool = bt_findpool(bt, page_no, idx) ) {
1220 __sync_fetch_and_add(&pool->pin, 1);
1222 _InterlockedIncrement16 (&pool->pin);
1224 bt_spinreleaseread (&bt->mgr->latch[idx]);
1229 // upgrade to write lock
1231 bt_spinreleaseread (&bt->mgr->latch[idx]);
1232 bt_spinwritelock (&bt->mgr->latch[idx]);
1234 // try to find page in pool with write lock
1236 if( pool = bt_findpool(bt, page_no, idx) ) {
1238 __sync_fetch_and_add(&pool->pin, 1);
1240 _InterlockedIncrement16 (&pool->pin);
1242 bt_spinreleasewrite (&bt->mgr->latch[idx]);
1247 // allocate a new pool node
1248 // and add to hash table
1251 slot = __sync_fetch_and_add(&bt->mgr->poolcnt, 1);
1253 slot = _InterlockedIncrement16 (&bt->mgr->poolcnt) - 1;
1256 if( ++slot < bt->mgr->poolmax ) {
1257 pool = bt->mgr->pool + slot;
1260 if( bt_mapsegment(bt, pool, page_no) )
1263 bt_linkhash(bt, pool, page_no, idx);
1265 __sync_fetch_and_add(&pool->pin, 1);
1267 _InterlockedIncrement16 (&pool->pin);
1269 bt_spinreleasewrite (&bt->mgr->latch[idx]);
1273 // pool table is full
1274 // find best pool entry to evict
1277 __sync_fetch_and_add(&bt->mgr->poolcnt, -1);
1279 _InterlockedDecrement16 (&bt->mgr->poolcnt);
1284 victim = __sync_fetch_and_add(&bt->mgr->evicted, 1);
1286 victim = _InterlockedIncrement (&bt->mgr->evicted) - 1;
1288 victim %= bt->mgr->hashsize;
1290 // try to get write lock
1291 // skip entry if not obtained
1293 if( !bt_spinwritetry (&bt->mgr->latch[victim]) )
1296 // if pool entry is empty
1297 // or any pages are pinned
1300 if( !(pool = bt_findlru(bt, bt->mgr->hash[victim])) ) {
1301 bt_spinreleasewrite (&bt->mgr->latch[victim]);
1305 // unlink victim pool node from hash table
1307 if( node = pool->hashprev )
1308 node->hashnext = pool->hashnext;
1309 else if( node = pool->hashnext )
1310 bt->mgr->hash[victim] = node->slot;
1312 bt->mgr->hash[victim] = 0;
1314 if( node = pool->hashnext )
1315 node->hashprev = pool->hashprev;
1317 bt_spinreleasewrite (&bt->mgr->latch[victim]);
1319 // remove old file mapping
1321 munmap (pool->map, (bt->mgr->poolmask+1) << bt->mgr->page_bits);
1323 FlushViewOfFile(pool->map, 0);
1324 UnmapViewOfFile(pool->map);
1325 CloseHandle(pool->hmap);
1329 // create new pool mapping
1330 // and link into hash table
1332 if( bt_mapsegment(bt, pool, page_no) )
1335 bt_linkhash(bt, pool, page_no, idx);
1337 __sync_fetch_and_add(&pool->pin, 1);
1339 _InterlockedIncrement16 (&pool->pin);
1341 bt_spinreleasewrite (&bt->mgr->latch[idx]);
1346 // place write, read, or parent lock on requested page_no.
1348 void bt_lockpage(BtLock mode, BtLatchSet *set)
1352 bt_readlock (set->readwr);
1355 bt_writelock (set->readwr);
1358 bt_readlock (set->access);
1361 bt_writelock (set->access);
1364 bt_writelock (set->parent);
1369 // remove write, read, or parent lock on requested page
1371 void bt_unlockpage(BtLock mode, BtLatchSet *set)
1375 bt_releaseread (set->readwr);
1378 bt_releasewrite (set->readwr);
1381 bt_releaseread (set->access);
1384 bt_releasewrite (set->access);
1387 bt_releasewrite (set->parent);
1392 // allocate a new page and write page into it
1394 uid bt_newpage(BtDb *bt, BtPage page)
1400 // lock allocation page
1402 bt_spinwritelock(bt->mgr->latchmgr->lock);
1404 // use empty chain first
1405 // else allocate empty page
1407 if( new_page = bt_getid(bt->mgr->latchmgr->alloc[1].right) ) {
1408 if( set->pool = bt_pinpool (bt, new_page) )
1409 set->page = bt_page (bt, set->pool, new_page);
1413 bt_putid(bt->mgr->latchmgr->alloc[1].right, bt_getid(set->page->right));
1414 bt_unpinpool (set->pool);
1417 new_page = bt_getid(bt->mgr->latchmgr->alloc->right);
1418 bt_putid(bt->mgr->latchmgr->alloc->right, new_page+1);
1422 if ( pwrite(bt->mgr->idx, page, bt->mgr->page_size, new_page << bt->mgr->page_bits) < bt->mgr->page_size )
1423 return bt->err = BTERR_wrt, 0;
1425 // if writing first page of pool block, zero last page in the block
1427 if ( !reuse && bt->mgr->poolmask > 0 && (new_page & bt->mgr->poolmask) == 0 )
1429 // use zero buffer to write zeros
1430 if ( pwrite(bt->mgr->idx,bt->zero, bt->mgr->page_size, (new_page | bt->mgr->poolmask) << bt->mgr->page_bits) < bt->mgr->page_size )
1431 return bt->err = BTERR_wrt, 0;
1434 // bring new page into pool and copy page.
1435 // this will extend the file into the new pages.
1437 if( set->pool = bt_pinpool (bt, new_page) )
1438 set->page = bt_page (bt, set->pool, new_page);
1442 memcpy(set->page, page, bt->mgr->page_size);
1443 bt_unpinpool (set->pool);
1445 // unlock allocation latch and return new page no
1447 bt_spinreleasewrite(bt->mgr->latchmgr->lock);
1451 // find slot in page for given key at a given level
1453 int bt_findslot (BtPageSet *set, unsigned char *key, uint len)
1455 uint diff, higher = set->page->cnt, low = 1, slot;
1458 // make stopper key an infinite fence value
1460 if( bt_getid (set->page->right) )
1465 // low is the lowest candidate.
1466 // loop ends when they meet
1468 // higher is already
1469 // tested as .ge. the passed key.
1471 while( diff = higher - low ) {
1472 slot = low + ( diff >> 1 );
1473 if( keycmp (keyptr(set->page, slot), key, len) < 0 )
1476 higher = slot, good++;
1479 // return zero if key is on right link page
1481 return good ? higher : 0;
1484 // find and load page at given level for given key
1485 // leave page rd or wr locked as requested
1487 int bt_loadpage (BtDb *bt, BtPageSet *set, unsigned char *key, uint len, uint lvl, BtLock lock)
1489 uid page_no = ROOT_page, prevpage = 0;
1490 uint drill = 0xff, slot;
1491 BtLatchSet *prevlatch;
1492 uint mode, prevmode;
1495 // start at root of btree and drill down
1498 // determine lock mode of drill level
1499 mode = (drill == lvl) ? lock : BtLockRead;
1501 set->latch = bt_pinlatch (bt, page_no);
1502 set->page_no = page_no;
1504 // pin page contents
1506 if( set->pool = bt_pinpool (bt, page_no) )
1507 set->page = bt_page (bt, set->pool, page_no);
1511 // obtain access lock using lock chaining with Access mode
1513 if( page_no > ROOT_page )
1514 bt_lockpage(BtLockAccess, set->latch);
1516 // release & unpin parent page
1519 bt_unlockpage(prevmode, prevlatch);
1520 bt_unpinlatch (prevlatch);
1521 bt_unpinpool (prevpool);
1525 // obtain read lock using lock chaining
1527 bt_lockpage(mode, set->latch);
1529 if( set->page->free )
1530 return bt->err = BTERR_struct, 0;
1532 if( page_no > ROOT_page )
1533 bt_unlockpage(BtLockAccess, set->latch);
1535 // re-read and re-lock root after determining actual level of root
1537 if( set->page->lvl != drill) {
1538 if ( set->page_no != ROOT_page )
1539 return bt->err = BTERR_struct, 0;
1541 drill = set->page->lvl;
1543 if( lock != BtLockRead && drill == lvl ) {
1544 bt_unlockpage(mode, set->latch);
1545 bt_unpinlatch (set->latch);
1546 bt_unpinpool (set->pool);
1551 prevpage = set->page_no;
1552 prevlatch = set->latch;
1553 prevpool = set->pool;
1556 // find key on page at this level
1557 // and descend to requested level
1559 if( !set->page->kill )
1560 if( slot = bt_findslot (set, key, len) ) {
1564 while( slotptr(set->page, slot)->dead )
1565 if( slot++ < set->page->cnt )
1570 page_no = bt_getid(slotptr(set->page, slot)->id);
1575 // or slide right into next page
1578 page_no = bt_getid(set->page->right);
1582 // return error on end of right chain
1584 bt->err = BTERR_struct;
1585 return 0; // return error
1588 // return page to free list
1589 // page must be delete & write locked
1591 void bt_freepage (BtDb *bt, BtPageSet *set)
1593 // lock allocation page
1595 bt_spinwritelock (bt->mgr->latchmgr->lock);
1597 // store chain in second right
1598 bt_putid(set->page->right, bt_getid(bt->mgr->latchmgr->alloc[1].right));
1599 bt_putid(bt->mgr->latchmgr->alloc[1].right, set->page_no);
1600 set->page->free = 1;
1602 // unlock released page
1604 bt_unlockpage (BtLockDelete, set->latch);
1605 bt_unlockpage (BtLockWrite, set->latch);
1606 bt_unpinlatch (set->latch);
1607 bt_unpinpool (set->pool);
1609 // unlock allocation page
1611 bt_spinreleasewrite (bt->mgr->latchmgr->lock);
1614 // a fence key was deleted from a page
1615 // push new fence value upwards
1617 BTERR bt_fixfence (BtDb *bt, BtPageSet *set, uint lvl)
1619 unsigned char leftkey[256], rightkey[256];
1623 // remove the old fence value
1625 ptr = keyptr(set->page, set->page->cnt);
1626 memcpy (rightkey, ptr, ptr->len + 1);
1628 memset (slotptr(set->page, set->page->cnt--), 0, sizeof(BtSlot));
1629 set->page->dirty = 1;
1631 ptr = keyptr(set->page, set->page->cnt);
1632 memcpy (leftkey, ptr, ptr->len + 1);
1633 page_no = set->page_no;
1635 bt_lockpage (BtLockParent, set->latch);
1636 bt_unlockpage (BtLockWrite, set->latch);
1638 // insert new (now smaller) fence key
1640 if( bt_insertkey (bt, leftkey+1, *leftkey, lvl+1, page_no, time(NULL)) )
1643 // now delete old fence key
1645 if( bt_deletekey (bt, rightkey+1, *rightkey, lvl+1) )
1648 bt_unlockpage (BtLockParent, set->latch);
1649 bt_unpinlatch(set->latch);
1650 bt_unpinpool (set->pool);
1654 // root has a single child
1655 // collapse a level from the tree
1657 BTERR bt_collapseroot (BtDb *bt, BtPageSet *root)
1662 // find the child entry and promote as new root contents
1665 for( idx = 0; idx++ < root->page->cnt; )
1666 if( !slotptr(root->page, idx)->dead )
1669 child->page_no = bt_getid (slotptr(root->page, idx)->id);
1671 child->latch = bt_pinlatch (bt, child->page_no);
1672 bt_lockpage (BtLockDelete, child->latch);
1673 bt_lockpage (BtLockWrite, child->latch);
1675 if( child->pool = bt_pinpool (bt, child->page_no) )
1676 child->page = bt_page (bt, child->pool, child->page_no);
1680 memcpy (root->page, child->page, bt->mgr->page_size);
1681 bt_freepage (bt, child);
1683 } while( root->page->lvl > 1 && root->page->act == 1 );
1685 bt_unlockpage (BtLockWrite, root->latch);
1686 bt_unpinlatch (root->latch);
1687 bt_unpinpool (root->pool);
1691 // find and delete key on page by marking delete flag bit
1692 // if page becomes empty, delete it from the btree
1694 BTERR bt_deletekey (BtDb *bt, unsigned char *key, uint len, uint lvl)
1696 unsigned char lowerfence[256], higherfence[256];
1697 uint slot, idx, dirty = 0, fence, found;
1698 BtPageSet set[1], right[1];
1701 if( slot = bt_loadpage (bt, set, key, len, lvl, BtLockWrite) )
1702 ptr = keyptr(set->page, slot);
1706 // are we deleting a fence slot?
1708 fence = slot == set->page->cnt;
1710 // if key is found delete it, otherwise ignore request
1712 if( found = !keycmp (ptr, key, len) )
1713 if( found = slotptr(set->page, slot)->dead == 0 ) {
1714 dirty = slotptr(set->page, slot)->dead = 1;
1715 set->page->dirty = 1;
1718 // collapse empty slots
1720 while( idx = set->page->cnt - 1 )
1721 if( slotptr(set->page, idx)->dead ) {
1722 *slotptr(set->page, idx) = *slotptr(set->page, idx + 1);
1723 memset (slotptr(set->page, set->page->cnt--), 0, sizeof(BtSlot));
1728 // did we delete a fence key in an upper level?
1730 if( dirty && lvl && set->page->act && fence )
1731 if( bt_fixfence (bt, set, lvl) )
1734 return bt->found = found, 0;
1736 // is this a collapsed root?
1738 if( lvl > 1 && set->page_no == ROOT_page && set->page->act == 1 )
1739 if( bt_collapseroot (bt, set) )
1742 return bt->found = found, 0;
1744 // return if page is not empty
1746 if( set->page->act ) {
1747 bt_unlockpage(BtLockWrite, set->latch);
1748 bt_unpinlatch (set->latch);
1749 bt_unpinpool (set->pool);
1750 return bt->found = found, 0;
1753 // cache copy of fence key
1754 // to post in parent
1756 ptr = keyptr(set->page, set->page->cnt);
1757 memcpy (lowerfence, ptr, ptr->len + 1);
1759 // obtain lock on right page
1761 right->page_no = bt_getid(set->page->right);
1762 right->latch = bt_pinlatch (bt, right->page_no);
1763 bt_lockpage (BtLockWrite, right->latch);
1765 // pin page contents
1767 if( right->pool = bt_pinpool (bt, right->page_no) )
1768 right->page = bt_page (bt, right->pool, right->page_no);
1772 if( right->page->kill )
1773 return bt->err = BTERR_struct;
1775 // pull contents of right peer into our empty page
1777 memcpy (set->page, right->page, bt->mgr->page_size);
1779 // cache copy of key to update
1781 ptr = keyptr(right->page, right->page->cnt);
1782 memcpy (higherfence, ptr, ptr->len + 1);
1784 // mark right page deleted and point it to left page
1785 // until we can post parent updates
1787 bt_putid (right->page->right, set->page_no);
1788 right->page->kill = 1;
1790 bt_lockpage (BtLockParent, right->latch);
1791 bt_unlockpage (BtLockWrite, right->latch);
1793 bt_lockpage (BtLockParent, set->latch);
1794 bt_unlockpage (BtLockWrite, set->latch);
1796 // redirect higher key directly to our new node contents
1798 if( bt_insertkey (bt, higherfence+1, *higherfence, lvl+1, set->page_no, time(NULL)) )
1801 // delete old lower key to our node
1803 if( bt_deletekey (bt, lowerfence+1, *lowerfence, lvl+1) )
1806 // obtain delete and write locks to right node
1808 bt_unlockpage (BtLockParent, right->latch);
1809 bt_lockpage (BtLockDelete, right->latch);
1810 bt_lockpage (BtLockWrite, right->latch);
1811 bt_freepage (bt, right);
1813 bt_unlockpage (BtLockParent, set->latch);
1814 bt_unpinlatch (set->latch);
1815 bt_unpinpool (set->pool);
1820 // find key in leaf level and return row-id
1822 uid bt_findkey (BtDb *bt, unsigned char *key, uint len)
1829 if( slot = bt_loadpage (bt, set, key, len, 0, BtLockRead) )
1830 ptr = keyptr(set->page, slot);
1834 // if key exists, return row-id
1835 // otherwise return 0
1837 if( slot <= set->page->cnt )
1838 if( !keycmp (ptr, key, len) )
1839 id = bt_getid(slotptr(set->page,slot)->id);
1841 bt_unlockpage (BtLockRead, set->latch);
1842 bt_unpinlatch (set->latch);
1843 bt_unpinpool (set->pool);
1847 // check page for space available,
1848 // clean if necessary and return
1849 // 0 - page needs splitting
1850 // >0 new slot value
1852 uint bt_cleanpage(BtDb *bt, BtPage page, uint amt, uint slot)
1854 uint nxt = bt->mgr->page_size;
1855 uint cnt = 0, idx = 0;
1856 uint max = page->cnt;
1860 if( page->min >= (max+1) * sizeof(BtSlot) + sizeof(*page) + amt + 1 )
1863 // skip cleanup if nothing to reclaim
1868 memcpy (bt->frame, page, bt->mgr->page_size);
1870 // skip page info and set rest of page to zero
1872 memset (page+1, 0, bt->mgr->page_size - sizeof(*page));
1876 // try cleaning up page first
1877 // by removing deleted keys
1879 while( cnt++ < max ) {
1882 if( cnt < max && slotptr(bt->frame,cnt)->dead )
1885 // copy the key across
1887 key = keyptr(bt->frame, cnt);
1888 nxt -= key->len + 1;
1889 memcpy ((unsigned char *)page + nxt, key, key->len + 1);
1893 memcpy(slotptr(page, ++idx)->id, slotptr(bt->frame, cnt)->id, BtId);
1894 if( !(slotptr(page, idx)->dead = slotptr(bt->frame, cnt)->dead) )
1896 slotptr(page, idx)->tod = slotptr(bt->frame, cnt)->tod;
1897 slotptr(page, idx)->off = nxt;
1903 // see if page has enough space now, or does it need splitting?
1905 if( page->min >= (idx+1) * sizeof(BtSlot) + sizeof(*page) + amt + 1 )
1911 // split the root and raise the height of the btree
1913 BTERR bt_splitroot(BtDb *bt, BtPageSet *root, unsigned char *leftkey, uid page_no2)
1915 uint nxt = bt->mgr->page_size;
1918 // Obtain an empty page to use, and copy the current
1919 // root contents into it, e.g. lower keys
1921 if( !(left = bt_newpage(bt, root->page)) )
1924 // preserve the page info at the bottom
1925 // of higher keys and set rest to zero
1927 memset(root->page+1, 0, bt->mgr->page_size - sizeof(*root->page));
1929 // insert lower keys page fence key on newroot page as first key
1931 nxt -= *leftkey + 1;
1932 memcpy ((unsigned char *)root->page + nxt, leftkey, *leftkey + 1);
1933 bt_putid(slotptr(root->page, 1)->id, left);
1934 slotptr(root->page, 1)->off = nxt;
1936 // insert stopper key on newroot page
1937 // and increase the root height
1940 ((unsigned char *)root->page)[nxt] = 2;
1941 ((unsigned char *)root->page)[nxt+1] = 0xff;
1942 ((unsigned char *)root->page)[nxt+2] = 0xff;
1943 bt_putid(slotptr(root->page, 2)->id, page_no2);
1944 slotptr(root->page, 2)->off = nxt;
1946 bt_putid(root->page->right, 0);
1947 root->page->min = nxt; // reset lowest used offset and key count
1948 root->page->cnt = 2;
1949 root->page->act = 2;
1952 // release and unpin root
1954 bt_unlockpage(BtLockWrite, root->latch);
1955 bt_unpinlatch (root->latch);
1956 bt_unpinpool (root->pool);
1960 // split already locked full node
1963 BTERR bt_splitpage (BtDb *bt, BtPageSet *set)
1965 uint cnt = 0, idx = 0, max, nxt = bt->mgr->page_size;
1966 unsigned char fencekey[256], rightkey[256];
1967 uint lvl = set->page->lvl;
1972 // split higher half of keys to bt->frame
1974 memset (bt->frame, 0, bt->mgr->page_size);
1975 max = set->page->cnt;
1979 while( cnt++ < max ) {
1980 key = keyptr(set->page, cnt);
1981 nxt -= key->len + 1;
1982 memcpy ((unsigned char *)bt->frame + nxt, key, key->len + 1);
1984 memcpy(slotptr(bt->frame,++idx)->id, slotptr(set->page,cnt)->id, BtId);
1985 if( !(slotptr(bt->frame, idx)->dead = slotptr(set->page, cnt)->dead) )
1987 slotptr(bt->frame, idx)->tod = slotptr(set->page, cnt)->tod;
1988 slotptr(bt->frame, idx)->off = nxt;
1991 // remember existing fence key for new page to the right
1993 memcpy (rightkey, key, key->len + 1);
1995 bt->frame->bits = bt->mgr->page_bits;
1996 bt->frame->min = nxt;
1997 bt->frame->cnt = idx;
1998 bt->frame->lvl = lvl;
2002 if( set->page_no > ROOT_page )
2003 memcpy (bt->frame->right, set->page->right, BtId);
2005 // get new free page and write higher keys to it.
2007 if( !(right->page_no = bt_newpage(bt, bt->frame)) )
2010 // update lower keys to continue in old page
2012 memcpy (bt->frame, set->page, bt->mgr->page_size);
2013 memset (set->page+1, 0, bt->mgr->page_size - sizeof(*set->page));
2014 nxt = bt->mgr->page_size;
2015 set->page->dirty = 0;
2020 // assemble page of smaller keys
2022 while( cnt++ < max / 2 ) {
2023 key = keyptr(bt->frame, cnt);
2024 nxt -= key->len + 1;
2025 memcpy ((unsigned char *)set->page + nxt, key, key->len + 1);
2026 memcpy(slotptr(set->page,++idx)->id, slotptr(bt->frame,cnt)->id, BtId);
2027 slotptr(set->page, idx)->tod = slotptr(bt->frame, cnt)->tod;
2028 slotptr(set->page, idx)->off = nxt;
2032 // remember fence key for smaller page
2034 memcpy(fencekey, key, key->len + 1);
2036 bt_putid(set->page->right, right->page_no);
2037 set->page->min = nxt;
2038 set->page->cnt = idx;
2040 // if current page is the root page, split it
2042 if( set->page_no == ROOT_page )
2043 return bt_splitroot (bt, set, fencekey, right->page_no);
2045 // insert new fences in their parent pages
2047 right->latch = bt_pinlatch (bt, right->page_no);
2048 bt_lockpage (BtLockParent, right->latch);
2050 bt_lockpage (BtLockParent, set->latch);
2051 bt_unlockpage (BtLockWrite, set->latch);
2053 // insert new fence for reformulated left block of smaller keys
2055 if( bt_insertkey (bt, fencekey+1, *fencekey, lvl+1, set->page_no, time(NULL)) )
2058 // switch fence for right block of larger keys to new right page
2060 if( bt_insertkey (bt, rightkey+1, *rightkey, lvl+1, right->page_no, time(NULL)) )
2063 bt_unlockpage (BtLockParent, set->latch);
2064 bt_unpinlatch (set->latch);
2065 bt_unpinpool (set->pool);
2067 bt_unlockpage (BtLockParent, right->latch);
2068 bt_unpinlatch (right->latch);
2072 // Insert new key into the btree at given level.
2074 BTERR bt_insertkey (BtDb *bt, unsigned char *key, uint len, uint lvl, uid id, uint tod)
2081 if( slot = bt_loadpage (bt, set, key, len, lvl, BtLockWrite) )
2082 ptr = keyptr(set->page, slot);
2086 bt->err = BTERR_ovflw;
2090 // if key already exists, update id and return
2092 if( !keycmp (ptr, key, len) ) {
2093 if( slotptr(set->page, slot)->dead )
2095 slotptr(set->page, slot)->dead = 0;
2096 slotptr(set->page, slot)->tod = tod;
2097 bt_putid(slotptr(set->page,slot)->id, id);
2098 bt_unlockpage(BtLockWrite, set->latch);
2099 bt_unpinlatch (set->latch);
2100 bt_unpinpool (set->pool);
2104 // check if page has enough space
2106 if( slot = bt_cleanpage (bt, set->page, len, slot) )
2109 if( bt_splitpage (bt, set) )
2113 // calculate next available slot and copy key into page
2115 set->page->min -= len + 1; // reset lowest used offset
2116 ((unsigned char *)set->page)[set->page->min] = len;
2117 memcpy ((unsigned char *)set->page + set->page->min +1, key, len );
2119 for( idx = slot; idx < set->page->cnt; idx++ )
2120 if( slotptr(set->page, idx)->dead )
2123 // now insert key into array before slot
2125 if( idx == set->page->cnt )
2126 idx++, set->page->cnt++;
2131 *slotptr(set->page, idx) = *slotptr(set->page, idx -1), idx--;
2133 bt_putid(slotptr(set->page,slot)->id, id);
2134 slotptr(set->page, slot)->off = set->page->min;
2135 slotptr(set->page, slot)->tod = tod;
2136 slotptr(set->page, slot)->dead = 0;
2138 bt_unlockpage (BtLockWrite, set->latch);
2139 bt_unpinlatch (set->latch);
2140 bt_unpinpool (set->pool);
2144 // cache page of keys into cursor and return starting slot for given key
2146 uint bt_startkey (BtDb *bt, unsigned char *key, uint len)
2151 // cache page for retrieval
2153 if( slot = bt_loadpage (bt, set, key, len, 0, BtLockRead) )
2154 memcpy (bt->cursor, set->page, bt->mgr->page_size);
2158 bt->cursor_page = set->page_no;
2160 bt_unlockpage(BtLockRead, set->latch);
2161 bt_unpinlatch (set->latch);
2162 bt_unpinpool (set->pool);
2166 // return next slot for cursor page
2167 // or slide cursor right into next page
2169 uint bt_nextkey (BtDb *bt, uint slot)
2175 right = bt_getid(bt->cursor->right);
2177 while( slot++ < bt->cursor->cnt )
2178 if( slotptr(bt->cursor,slot)->dead )
2180 else if( right || (slot < bt->cursor->cnt) ) // skip infinite stopper
2188 bt->cursor_page = right;
2190 if( set->pool = bt_pinpool (bt, right) )
2191 set->page = bt_page (bt, set->pool, right);
2195 set->latch = bt_pinlatch (bt, right);
2196 bt_lockpage(BtLockRead, set->latch);
2198 memcpy (bt->cursor, set->page, bt->mgr->page_size);
2200 bt_unlockpage(BtLockRead, set->latch);
2201 bt_unpinlatch (set->latch);
2202 bt_unpinpool (set->pool);
2210 BtKey bt_key(BtDb *bt, uint slot)
2212 return keyptr(bt->cursor, slot);
2215 uid bt_uid(BtDb *bt, uint slot)
2217 return bt_getid(slotptr(bt->cursor,slot)->id);
2220 uint bt_tod(BtDb *bt, uint slot)
2222 return slotptr(bt->cursor,slot)->tod;
2228 void bt_latchaudit (BtDb *bt)
2230 ushort idx, hashidx;
2236 for( idx = 1; idx < bt->mgr->latchmgr->latchdeployed; idx++ ) {
2237 set->latch = bt->mgr->latchsets + idx;
2238 if( set->latch->pin ) {
2239 fprintf(stderr, "latchset %d pinned for page %.6x\n", idx, set->latch->page_no);
2240 set->latch->pin = 0;
2244 for( hashidx = 0; hashidx < bt->mgr->latchmgr->latchhash; hashidx++ ) {
2245 if( idx = bt->mgr->latchmgr->table[hashidx].slot ) do {
2246 set->latch = bt->mgr->latchsets + idx;
2247 if( set->latch->hash != hashidx )
2248 fprintf(stderr, "latchset %d wrong hashidx\n", idx);
2249 if( set->latch->pin )
2250 fprintf(stderr, "latchset %d pinned for page %.8x\n", idx, set->latch->page_no);
2251 } while( idx = set->latch->next );
2254 next = bt->mgr->latchmgr->nlatchpage + LATCH_page;
2255 page_no = LEAF_page;
2257 while( page_no < bt_getid(bt->mgr->latchmgr->alloc->right) ) {
2258 pread (bt->mgr->idx, bt->frame, bt->mgr->page_size, page_no << bt->mgr->page_bits);
2259 if( !bt->frame->free )
2260 for( idx = 0; idx++ < bt->frame->cnt - 1; ) {
2261 ptr = keyptr(bt->frame, idx+1);
2262 if( keycmp (keyptr(bt->frame, idx), ptr->key, ptr->len) >= 0 )
2263 fprintf(stderr, "page %.8x idx %.2x out of order\n", page_no, idx);
2266 if( page_no > LEAF_page )
2280 // standalone program to index file of keys
2281 // then list them onto std-out
2284 void *index_file (void *arg)
2286 uint __stdcall index_file (void *arg)
2289 int line = 0, found = 0, cnt = 0;
2290 uid next, page_no = LEAF_page; // start on first page of leaves
2291 unsigned char key[256];
2292 ThreadArg *args = arg;
2293 int ch, len = 0, slot;
2300 bt = bt_open (args->mgr);
2303 switch(args->type | 0x20)
2306 fprintf(stderr, "started latch mgr audit\n");
2308 fprintf(stderr, "finished latch mgr audit\n");
2312 fprintf(stderr, "started indexing for %s\n", args->infile);
2313 if( in = fopen (args->infile, "rb") )
2314 while( ch = getc(in), ch != EOF )
2319 if( args->num == 1 )
2320 sprintf((char *)key+len, "%.9d", 1000000000 - line), len += 9;
2322 else if( args->num )
2323 sprintf((char *)key+len, "%.9d", line + args->idx * args->num), len += 9;
2325 if( bt_insertkey (bt, key, len, 0, line, *tod) )
2326 fprintf(stderr, "Error %d Line: %d\n", bt->err, line), exit(0);
2329 else if( len < 255 )
2331 fprintf(stderr, "finished %s for %d keys\n", args->infile, line);
2335 fprintf(stderr, "started deleting keys for %s\n", args->infile);
2336 if( in = fopen (args->infile, "rb") )
2337 while( ch = getc(in), ch != EOF )
2341 if( args->num == 1 )
2342 sprintf((char *)key+len, "%.9d", 1000000000 - line), len += 9;
2344 else if( args->num )
2345 sprintf((char *)key+len, "%.9d", line + args->idx * args->num), len += 9;
2347 if( bt_deletekey (bt, key, len, 0) )
2348 fprintf(stderr, "Error %d Line: %d\n", bt->err, line), exit(0);
2351 else if( len < 255 )
2353 fprintf(stderr, "finished %s for keys, %d \n", args->infile, line);
2357 fprintf(stderr, "started finding keys for %s\n", args->infile);
2358 if( in = fopen (args->infile, "rb") )
2359 while( ch = getc(in), ch != EOF )
2363 if( args->num == 1 )
2364 sprintf((char *)key+len, "%.9d", 1000000000 - line), len += 9;
2366 else if( args->num )
2367 sprintf((char *)key+len, "%.9d", line + args->idx * args->num), len += 9;
2369 if( bt_findkey (bt, key, len) )
2372 fprintf(stderr, "Error %d Syserr %d Line: %d\n", bt->err, errno, line), exit(0);
2374 fprintf(stderr, "Unable to find key %.*s line %d\n", len, key, line);
2377 else if( len < 255 )
2379 fprintf(stderr, "finished %s for %d keys, found %d\n", args->infile, line, found);
2383 fprintf(stderr, "started scanning\n");
2385 if( set->pool = bt_pinpool (bt, page_no) )
2386 set->page = bt_page (bt, set->pool, page_no);
2389 set->latch = bt_pinlatch (bt, page_no);
2390 bt_lockpage (BtLockRead, set->latch);
2391 next = bt_getid (set->page->right);
2392 cnt += set->page->act;
2394 for( slot = 0; slot++ < set->page->cnt; )
2395 if( next || slot < set->page->cnt )
2396 if( !slotptr(set->page, slot)->dead ) {
2397 ptr = keyptr(set->page, slot);
2398 fwrite (ptr->key, ptr->len, 1, stdout);
2399 fputc ('\n', stdout);
2402 bt_unlockpage (BtLockRead, set->latch);
2403 bt_unpinlatch (set->latch);
2404 bt_unpinpool (set->pool);
2405 } while( page_no = next );
2407 cnt--; // remove stopper key
2408 fprintf(stderr, " Total keys read %d\n", cnt);
2412 fprintf(stderr, "started counting\n");
2413 next = bt->mgr->latchmgr->nlatchpage + LATCH_page;
2414 page_no = LEAF_page;
2416 while( page_no < bt_getid(bt->mgr->latchmgr->alloc->right) ) {
2417 uid off = page_no << bt->mgr->page_bits;
2419 pread (bt->mgr->idx, bt->frame, bt->mgr->page_size, off);
2423 SetFilePointer (bt->mgr->idx, (long)off, (long*)(&off)+1, FILE_BEGIN);
2425 if( !ReadFile(bt->mgr->idx, bt->frame, bt->mgr->page_size, amt, NULL))
2426 return bt->err = BTERR_map;
2428 if( *amt < bt->mgr->page_size )
2429 return bt->err = BTERR_map;
2431 if( !bt->frame->free && !bt->frame->lvl )
2432 cnt += bt->frame->act;
2433 if( page_no > LEAF_page )
2438 cnt--; // remove stopper key
2439 fprintf(stderr, " Total keys read %d\n", cnt);
2451 typedef struct timeval timer;
2453 int main (int argc, char **argv)
2455 int idx, cnt, len, slot, err;
2456 int segsize, bits = 16;
2461 time_t start[1], stop[1];
2474 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]);
2475 fprintf (stderr, " where page_bits is the page size in bits\n");
2476 fprintf (stderr, " mapped_segments is the number of mmap segments in buffer pool\n");
2477 fprintf (stderr, " seg_bits is the size of individual segments in buffer pool in pages in bits\n");
2478 fprintf (stderr, " line_numbers = 1 to append line numbers to keys\n");
2479 fprintf (stderr, " src_file1 thru src_filen are files of keys separated by newline\n");
2484 gettimeofday(&start, NULL);
2490 bits = atoi(argv[3]);
2493 poolsize = atoi(argv[4]);
2496 fprintf (stderr, "Warning: no mapped_pool\n");
2498 if( poolsize > 65535 )
2499 fprintf (stderr, "Warning: mapped_pool > 65535 segments\n");
2502 segsize = atoi(argv[5]);
2504 segsize = 4; // 16 pages per mmap segment
2507 num = atoi(argv[6]);
2511 threads = malloc (cnt * sizeof(pthread_t));
2513 threads = GlobalAlloc (GMEM_FIXED|GMEM_ZEROINIT, cnt * sizeof(HANDLE));
2515 args = malloc (cnt * sizeof(ThreadArg));
2517 mgr = bt_mgr ((argv[1]), BT_rw, bits, poolsize, segsize, poolsize / 8);
2520 fprintf(stderr, "Index Open Error %s\n", argv[1]);
2526 for( idx = 0; idx < cnt; idx++ ) {
2527 args[idx].infile = argv[idx + 7];
2528 args[idx].type = argv[2][0];
2529 args[idx].mgr = mgr;
2530 args[idx].num = num;
2531 args[idx].idx = idx;
2533 if( err = pthread_create (threads + idx, NULL, index_file, args + idx) )
2534 fprintf(stderr, "Error creating thread %d\n", err);
2536 threads[idx] = (HANDLE)_beginthreadex(NULL, 65536, index_file, args + idx, 0, NULL);
2540 // wait for termination
2543 for( idx = 0; idx < cnt; idx++ )
2544 pthread_join (threads[idx], NULL);
2545 gettimeofday(&stop, NULL);
2546 real_time = 1000.0 * ( stop.tv_sec - start.tv_sec ) + 0.001 * (stop.tv_usec - start.tv_usec );
2548 WaitForMultipleObjects (cnt, threads, TRUE, INFINITE);
2550 for( idx = 0; idx < cnt; idx++ )
2551 CloseHandle(threads[idx]);
2554 real_time = 1000 * (*stop - *start);
2556 fprintf(stderr, " Time to complete: %.2f seconds\n", real_time/1000);