1 // btree version threads2j linux futex concurrency version
2 // with reworked bt_deletekey
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
29 #include <linux/futex.h>
44 #define WIN32_LEAN_AND_MEAN
58 typedef unsigned long long uid;
61 typedef unsigned long long off64_t;
62 typedef unsigned short ushort;
63 typedef unsigned int uint;
66 #define BT_ro 0x6f72 // ro
67 #define BT_rw 0x7772 // rw
69 #define BT_latchtable 128 // number of latch manager slots
71 #define BT_maxbits 24 // maximum page size in bits
72 #define BT_minbits 9 // minimum page size in bits
73 #define BT_minpage (1 << BT_minbits) // minimum page size
74 #define BT_maxpage (1 << BT_maxbits) // maximum page size
77 There are five lock types for each node in three independent sets:
78 1. (set 1) AccessIntent: Sharable. Going to Read the node. Incompatible with NodeDelete.
79 2. (set 1) NodeDelete: Exclusive. About to release the node. Incompatible with AccessIntent.
80 3. (set 2) ReadLock: Sharable. Read the node. Incompatible with WriteLock.
81 4. (set 2) WriteLock: Exclusive. Modify the node. Incompatible with ReadLock and other WriteLocks.
82 5. (set 3) ParentModification: Exclusive. Change the node's parent keys. Incompatible with another ParentModification.
94 // mode & definition for latch implementation
97 QueRd = 1, // reader queue
98 QueWr = 2 // writer queue
101 // share is count of read accessors
102 // grant write lock when share == 0
104 volatile typedef struct {
105 unsigned char mutex[1]; // 1 = busy
106 unsigned char write:1; // 1 = exclusive
107 unsigned char readwait:1; // readers are waiting
108 unsigned char writewait:1; // writers are waiting
109 unsigned char filler:5;
110 ushort share; // count of readers holding locks
111 ushort rcnt; // count of waiting readers
112 ushort wcnt; // count of waiting writers
115 // Define the length of the page and key pointers
119 // Page key slot definition.
121 // If BT_maxbits is 15 or less, you can save 4 bytes
122 // for each key stored by making the first two uints
123 // into ushorts. You can also save 4 bytes by removing
124 // the tod field from the key.
126 // Keys are marked dead, but remain on the page until
127 // it cleanup is called. The fence key (highest key) for
128 // the page is always present, even after cleanup.
131 uint off:BT_maxbits; // page offset for key start
132 uint dead:1; // set for deleted key
133 uint tod; // time-stamp for key
134 unsigned char id[BtId]; // id associated with key
137 // The key structure occupies space at the upper end of
138 // each page. It's a length byte followed by the key
143 unsigned char key[1];
146 // The first part of an index page.
147 // It is immediately followed
148 // by the BtSlot array of keys.
150 typedef struct BtPage_ {
151 uint cnt; // count of keys in page
152 uint act; // count of active keys
153 uint min; // next key offset
154 unsigned char bits:7; // page size in bits
155 unsigned char free:1; // page is on free list
156 unsigned char lvl:6; // level of page
157 unsigned char kill:1; // page is being deleted
158 unsigned char dirty:1; // page has deleted keys
159 unsigned char right[BtId]; // page number to right
162 // hash table entries
166 volatile ushort slot; // Latch table entry at head of chain
169 // latch manager table structure
172 BtLatch readwr[1]; // read/write page lock
173 BtLatch access[1]; // Access Intent/Page delete
174 BtLatch parent[1]; // adoption of foster children
175 BtLatch busy[1]; // slot is being moved between chains
176 volatile ushort next; // next entry in hash table chain
177 volatile ushort prev; // prev entry in hash table chain
178 volatile ushort pin; // number of outstanding locks
179 volatile ushort hash; // hash slot entry is under
180 volatile uid page_no; // latch set page number
183 // The memory mapping pool table buffer manager entry
186 unsigned long long int lru; // number of times accessed
187 uid basepage; // mapped base page number
188 char *map; // mapped memory pointer
189 ushort slot; // slot index in this array
190 ushort pin; // mapped page pin counter
191 void *hashprev; // previous pool entry for the same hash idx
192 void *hashnext; // next pool entry for the same hash idx
194 HANDLE hmap; // Windows memory mapping handle
198 // The loadpage interface object
201 uid page_no; // current page number
202 BtPage page; // current page pointer
203 BtPool *pool; // current page pool
204 BtLatchSet *latch; // current page latch set
207 // structure for latch manager on ALLOC_page
210 struct BtPage_ alloc[2]; // next & free page_nos in right ptr
211 BtLatch lock[1]; // allocation area lite latch
212 ushort latchdeployed; // highest number of latch entries deployed
213 ushort nlatchpage; // number of latch pages at BT_latch
214 ushort latchtotal; // number of page latch entries
215 ushort latchhash; // number of latch hash table slots
216 ushort latchvictim; // next latch entry to examine
217 BtHashEntry table[0]; // the hash table
220 // The object structure for Btree access
223 uint page_size; // page size
224 uint page_bits; // page size in bits
225 uint seg_bits; // seg size in pages in bits
226 uint mode; // read-write mode
232 ushort poolcnt; // highest page pool node in use
233 ushort poolmax; // highest page pool node allocated
234 ushort poolmask; // total number of pages in mmap segment - 1
235 ushort evicted; // last evicted hash table slot
236 ushort hashsize; // size of Hash Table for pool entries
237 ushort *hash; // pool index for hash entries
238 BtLatch *latch; // latches for pool hash slots
239 BtLatchMgr *latchmgr; // mapped latch page from allocation page
240 BtLatchSet *latchsets; // mapped latch set from latch pages
241 BtPool *pool; // memory pool page segments
243 HANDLE halloc; // allocation and latch table handle
248 BtMgr *mgr; // buffer manager for thread
249 BtPage cursor; // cached frame for start/next (never mapped)
250 BtPage frame; // spare frame for the page split (never mapped)
251 BtPage zero; // page of zeroes to extend the file (never mapped)
252 uid cursor_page; // current cursor page number
253 unsigned char *mem; // frame, cursor, page memory buffer
254 int found; // last delete or insert was found
255 int err; // last error
269 extern void bt_close (BtDb *bt);
270 extern BtDb *bt_open (BtMgr *mgr);
271 extern BTERR bt_insertkey (BtDb *bt, unsigned char *key, uint len, uint lvl, uid id, uint tod);
272 extern BTERR bt_deletekey (BtDb *bt, unsigned char *key, uint len, uint lvl);
273 extern uid bt_findkey (BtDb *bt, unsigned char *key, uint len);
274 extern uint bt_startkey (BtDb *bt, unsigned char *key, uint len);
275 extern uint bt_nextkey (BtDb *bt, uint slot);
278 extern BtMgr *bt_mgr (char *name, uint mode, uint bits, uint poolsize, uint segsize, uint hashsize);
279 void bt_mgrclose (BtMgr *mgr);
281 // Helper functions to return slot values
282 extern BtKey bt_key (BtDb *bt, uint slot);
283 extern uid bt_uid (BtDb *bt, uint slot);
284 extern uint bt_tod (BtDb *bt, uint slot);
286 // BTree page number constants
287 #define ALLOC_page 0 // allocation & lock manager hash table
288 #define ROOT_page 1 // root of the btree
289 #define LEAF_page 2 // first page of leaves
290 #define LATCH_page 3 // pages for lock manager
292 // Number of levels to create in a new BTree
296 // The page is allocated from low and hi ends.
297 // The key offsets and row-id's are allocated
298 // from the bottom, while the text of the key
299 // is allocated from the top. When the two
300 // areas meet, the page is split into two.
302 // A key consists of a length byte, two bytes of
303 // index number (0 - 65534), and up to 253 bytes
304 // of key value. Duplicate keys are discarded.
305 // Associated with each key is a 48 bit row-id,
306 // or any other value desired.
308 // The b-tree root is always located at page 1.
309 // The first leaf page of level zero is always
310 // located on page 2.
312 // The b-tree pages are linked with next
313 // pointers to facilitate enumerators,
314 // and provide for concurrency.
316 // When to root page fills, it is split in two and
317 // the tree height is raised by a new root at page
318 // one with two keys.
320 // Deleted keys are marked with a dead bit until
321 // page cleanup. The fence key for a node is
324 // Groups of pages called segments from the btree are optionally
325 // cached with a memory mapped pool. A hash table is used to keep
326 // track of the cached segments. This behaviour is controlled
327 // by the cache block size parameter to bt_open.
329 // To achieve maximum concurrency one page is locked at a time
330 // as the tree is traversed to find leaf key in question. The right
331 // page numbers are used in cases where the page is being split,
334 // Page 0 is dedicated to lock for new page extensions,
335 // and chains empty pages together for reuse.
337 // The ParentModification lock on a node is obtained to serialize posting
338 // or changing the fence key for a node.
340 // Empty pages are chained together through the ALLOC page and reused.
342 // Access macros to address slot and key values from the page
343 // Page slots use 1 based indexing.
345 #define slotptr(page, slot) (((BtSlot *)(page+1)) + (slot-1))
346 #define keyptr(page, slot) ((BtKey)((unsigned char*)(page) + slotptr(page, slot)->off))
348 void bt_putid(unsigned char *dest, uid id)
353 dest[i] = (unsigned char)id, id >>= 8;
356 uid bt_getid(unsigned char *src)
361 for( i = 0; i < BtId; i++ )
362 id <<= 8, id |= *src++;
369 int sys_futex(void *addr1, int op, int val1, struct timespec *timeout, void *addr2, int val3)
371 return syscall(SYS_futex, addr1, op, val1, timeout, addr2, val3);
374 // wait until write lock mode is clear
375 // and add 1 to the share count
377 void bt_spinreadlock(BtLatch *latch, int private)
383 private = FUTEX_PRIVATE_FLAG;
386 // obtain latch mutex
387 while( __sync_lock_test_and_set(latch->mutex, 1) )
391 latch->rcnt--, decr = 0;
393 // wait for writers to clear
394 // increment read waiters and wait
396 if( latch->write || latch->writewait ) {
399 prev = *(uint *)latch & ~1;
400 __sync_lock_release (latch->mutex);
401 sys_futex( (uint *)latch, FUTEX_WAIT_BITSET | private, prev, NULL, NULL, QueRd );
406 // increment reader lock count
407 // and release latch mutex
411 __sync_lock_release (latch->mutex);
416 // wait for other read and write latches to relinquish
418 void bt_spinwritelock(BtLatch *latch, int private)
424 private = FUTEX_PRIVATE_FLAG;
427 // obtain latch mutex
428 while( __sync_lock_test_and_set(latch->mutex, 1) )
432 latch->wcnt--, decr = 0;
434 // wait for write and reader count to clear
436 if( latch->write || latch->share ) {
437 latch->writewait = 1;
439 prev = *(uint *)latch & ~1;
440 __sync_lock_release (latch->mutex);
441 sys_futex( (uint *)latch, FUTEX_WAIT_BITSET | private, prev, NULL, NULL, QueWr );
447 // release latch mutex
450 latch->writewait = 0;
453 __sync_lock_release (latch->mutex);
458 // try to obtain write lock
460 // return 1 if obtained,
463 int bt_spinwritetry(BtLatch *latch)
468 // abandon request if not taken
470 if( __sync_lock_test_and_set(latch->mutex, 1) )
473 // see if write mode is available
475 if( !latch->write && !latch->share )
476 ans = latch->write = 1;
480 // release latch mutex
482 __sync_lock_release (latch->mutex);
488 void bt_spinreleasewrite(BtLatch *latch, int private)
491 private = FUTEX_PRIVATE_FLAG;
493 // obtain latch mutex
495 while( __sync_lock_test_and_set(latch->mutex, 1) )
503 if( sys_futex( (uint *)latch, FUTEX_WAKE_BITSET | private, 1, NULL, NULL, QueWr ) )
507 sys_futex( (uint *)latch, FUTEX_WAKE_BITSET | private, INT_MAX, NULL, NULL, QueRd );
509 // release latch mutex
512 __sync_lock_release (latch->mutex);
515 // decrement reader count
517 void bt_spinreleaseread(BtLatch *latch, int private)
520 private = FUTEX_PRIVATE_FLAG;
522 // obtain latch mutex
524 while( __sync_lock_test_and_set(latch->mutex, 1) )
529 // wake one waiting writer
531 if( !latch->share && latch->wcnt )
532 sys_futex( (uint *)latch, FUTEX_WAKE_BITSET | private, 1, NULL, NULL, QueWr );
534 // release latch mutex
536 __sync_lock_release (latch->mutex);
539 // link latch table entry into latch hash table
541 void bt_latchlink (BtDb *bt, ushort hashidx, ushort victim, uid page_no)
543 BtLatchSet *set = bt->mgr->latchsets + victim;
545 if( set->next = bt->mgr->latchmgr->table[hashidx].slot )
546 bt->mgr->latchsets[set->next].prev = victim;
548 bt->mgr->latchmgr->table[hashidx].slot = victim;
549 set->page_no = page_no;
556 void bt_unpinlatch (BtLatchSet *set)
559 __sync_fetch_and_add(&set->pin, -1);
561 _InterlockedDecrement16 (&set->pin);
565 // find existing latchset or inspire new one
566 // return with latchset pinned
568 BtLatchSet *bt_pinlatch (BtDb *bt, uid page_no)
570 ushort hashidx = page_no % bt->mgr->latchmgr->latchhash;
571 ushort slot, avail = 0, victim, idx;
574 // obtain read lock on hash table entry
576 bt_spinreadlock(bt->mgr->latchmgr->table[hashidx].latch, 0);
578 if( slot = bt->mgr->latchmgr->table[hashidx].slot ) do
580 set = bt->mgr->latchsets + slot;
581 if( page_no == set->page_no )
583 } while( slot = set->next );
587 __sync_fetch_and_add(&set->pin, 1);
589 _InterlockedIncrement16 (&set->pin);
593 bt_spinreleaseread (bt->mgr->latchmgr->table[hashidx].latch, 0);
598 // try again, this time with write lock
600 bt_spinwritelock(bt->mgr->latchmgr->table[hashidx].latch, 0);
602 if( slot = bt->mgr->latchmgr->table[hashidx].slot ) do
604 set = bt->mgr->latchsets + slot;
605 if( page_no == set->page_no )
607 if( !set->pin && !avail )
609 } while( slot = set->next );
611 // found our entry, or take over an unpinned one
613 if( slot || (slot = avail) ) {
614 set = bt->mgr->latchsets + slot;
616 __sync_fetch_and_add(&set->pin, 1);
618 _InterlockedIncrement16 (&set->pin);
620 set->page_no = page_no;
621 bt_spinreleasewrite(bt->mgr->latchmgr->table[hashidx].latch, 0);
625 // see if there are any unused entries
627 victim = __sync_fetch_and_add (&bt->mgr->latchmgr->latchdeployed, 1) + 1;
629 victim = _InterlockedIncrement16 (&bt->mgr->latchmgr->latchdeployed);
632 if( victim < bt->mgr->latchmgr->latchtotal ) {
633 set = bt->mgr->latchsets + victim;
635 __sync_fetch_and_add(&set->pin, 1);
637 _InterlockedIncrement16 (&set->pin);
639 bt_latchlink (bt, hashidx, victim, page_no);
640 bt_spinreleasewrite (bt->mgr->latchmgr->table[hashidx].latch, 0);
645 victim = __sync_fetch_and_add (&bt->mgr->latchmgr->latchdeployed, -1);
647 victim = _InterlockedDecrement16 (&bt->mgr->latchmgr->latchdeployed);
649 // find and reuse previous lock entry
653 victim = __sync_fetch_and_add(&bt->mgr->latchmgr->latchvictim, 1);
655 victim = _InterlockedIncrement16 (&bt->mgr->latchmgr->latchvictim) - 1;
657 // we don't use slot zero
659 if( victim %= bt->mgr->latchmgr->latchtotal )
660 set = bt->mgr->latchsets + victim;
664 // take control of our slot
665 // from other threads
667 if( set->pin || !bt_spinwritetry (set->busy) )
672 // try to get write lock on hash chain
673 // skip entry if not obtained
674 // or has outstanding locks
676 if( !bt_spinwritetry (bt->mgr->latchmgr->table[idx].latch) ) {
677 bt_spinreleasewrite (set->busy, 0);
682 bt_spinreleasewrite (set->busy, 0);
683 bt_spinreleasewrite (bt->mgr->latchmgr->table[idx].latch, 0);
687 // unlink our available victim from its hash chain
690 bt->mgr->latchsets[set->prev].next = set->next;
692 bt->mgr->latchmgr->table[idx].slot = set->next;
695 bt->mgr->latchsets[set->next].prev = set->prev;
697 bt_spinreleasewrite (bt->mgr->latchmgr->table[idx].latch, 0);
699 __sync_fetch_and_add(&set->pin, 1);
701 _InterlockedIncrement16 (&set->pin);
703 bt_latchlink (bt, hashidx, victim, page_no);
704 bt_spinreleasewrite (bt->mgr->latchmgr->table[hashidx].latch, 0);
705 bt_spinreleasewrite (set->busy, 0);
710 void bt_mgrclose (BtMgr *mgr)
715 // release mapped pages
716 // note that slot zero is never used
718 for( slot = 1; slot < mgr->poolmax; slot++ ) {
719 pool = mgr->pool + slot;
722 munmap (pool->map, (mgr->poolmask+1) << mgr->page_bits);
725 FlushViewOfFile(pool->map, 0);
726 UnmapViewOfFile(pool->map);
727 CloseHandle(pool->hmap);
733 munmap (mgr->latchsets, mgr->latchmgr->nlatchpage * mgr->page_size);
734 munmap (mgr->latchmgr, mgr->page_size);
736 FlushViewOfFile(mgr->latchmgr, 0);
737 UnmapViewOfFile(mgr->latchmgr);
738 CloseHandle(mgr->halloc);
744 free ((void *)mgr->latch);
747 FlushFileBuffers(mgr->idx);
748 CloseHandle(mgr->idx);
749 GlobalFree (mgr->pool);
750 GlobalFree (mgr->hash);
751 GlobalFree ((void *)mgr->latch);
756 // close and release memory
758 void bt_close (BtDb *bt)
765 VirtualFree (bt->mem, 0, MEM_RELEASE);
770 // open/create new btree buffer manager
772 // call with file_name, BT_openmode, bits in page size (e.g. 16),
773 // size of mapped page pool (e.g. 8192)
775 BtMgr *bt_mgr (char *name, uint mode, uint bits, uint poolmax, uint segsize, uint hashsize)
777 uint lvl, attr, cacheblk, last, slot, idx;
778 uint nlatchpage, latchhash;
779 BtLatchMgr *latchmgr;
786 SYSTEM_INFO sysinfo[1];
789 // determine sanity of page size and buffer pool
791 if( bits > BT_maxbits )
793 else if( bits < BT_minbits )
797 return NULL; // must have buffer pool
800 mgr = calloc (1, sizeof(BtMgr));
801 mgr->idx = open ((char*)name, O_RDWR | O_CREAT, 0666);
804 return free(mgr), NULL;
806 cacheblk = 4096; // minimum mmap segment size for unix
809 mgr = GlobalAlloc (GMEM_FIXED|GMEM_ZEROINIT, sizeof(BtMgr));
810 attr = FILE_ATTRIBUTE_NORMAL;
811 mgr->idx = CreateFile(name, GENERIC_READ| GENERIC_WRITE, FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, OPEN_ALWAYS, attr, NULL);
813 if( mgr->idx == INVALID_HANDLE_VALUE )
814 return GlobalFree(mgr), NULL;
816 // normalize cacheblk to multiple of sysinfo->dwAllocationGranularity
817 GetSystemInfo(sysinfo);
818 cacheblk = sysinfo->dwAllocationGranularity;
822 latchmgr = malloc (BT_maxpage);
825 // read minimum page size to get root info
827 if( size = lseek (mgr->idx, 0L, 2) ) {
828 if( pread(mgr->idx, latchmgr, BT_minpage, 0) == BT_minpage )
829 bits = latchmgr->alloc->bits;
831 return free(mgr), free(latchmgr), NULL;
832 } else if( mode == BT_ro )
833 return free(latchmgr), free (mgr), NULL;
835 latchmgr = VirtualAlloc(NULL, BT_maxpage, MEM_COMMIT, PAGE_READWRITE);
836 size = GetFileSize(mgr->idx, amt);
839 if( !ReadFile(mgr->idx, (char *)latchmgr, BT_minpage, amt, NULL) )
840 return bt_mgrclose (mgr), NULL;
841 bits = latchmgr->alloc->bits;
842 } else if( mode == BT_ro )
843 return bt_mgrclose (mgr), NULL;
846 mgr->page_size = 1 << bits;
847 mgr->page_bits = bits;
849 mgr->poolmax = poolmax;
852 if( cacheblk < mgr->page_size )
853 cacheblk = mgr->page_size;
855 // mask for partial memmaps
857 mgr->poolmask = (cacheblk >> bits) - 1;
859 // see if requested size of pages per memmap is greater
861 if( (1 << segsize) > mgr->poolmask )
862 mgr->poolmask = (1 << segsize) - 1;
866 while( (1 << mgr->seg_bits) <= mgr->poolmask )
869 mgr->hashsize = hashsize;
872 mgr->pool = calloc (poolmax, sizeof(BtPool));
873 mgr->hash = calloc (hashsize, sizeof(ushort));
874 mgr->latch = calloc (hashsize, sizeof(BtLatch));
876 mgr->pool = GlobalAlloc (GMEM_FIXED|GMEM_ZEROINIT, poolmax * sizeof(BtPool));
877 mgr->hash = GlobalAlloc (GMEM_FIXED|GMEM_ZEROINIT, hashsize * sizeof(ushort));
878 mgr->latch = GlobalAlloc (GMEM_FIXED|GMEM_ZEROINIT, hashsize * sizeof(BtLatch));
884 // initialize an empty b-tree with latch page, root page, page of leaves
885 // and page(s) of latches
887 memset (latchmgr, 0, 1 << bits);
888 nlatchpage = BT_latchtable / (mgr->page_size / sizeof(BtLatchSet)) + 1;
889 bt_putid(latchmgr->alloc->right, MIN_lvl+1+nlatchpage);
890 latchmgr->alloc->bits = mgr->page_bits;
892 latchmgr->nlatchpage = nlatchpage;
893 latchmgr->latchtotal = nlatchpage * (mgr->page_size / sizeof(BtLatchSet));
895 // initialize latch manager
897 latchhash = (mgr->page_size - sizeof(BtLatchMgr)) / sizeof(BtHashEntry);
899 // size of hash table = total number of latchsets
901 if( latchhash > latchmgr->latchtotal )
902 latchhash = latchmgr->latchtotal;
904 latchmgr->latchhash = latchhash;
907 if( write (mgr->idx, latchmgr, mgr->page_size) < mgr->page_size )
908 return bt_mgrclose (mgr), NULL;
910 if( !WriteFile (mgr->idx, (char *)latchmgr, mgr->page_size, amt, NULL) )
911 return bt_mgrclose (mgr), NULL;
913 if( *amt < mgr->page_size )
914 return bt_mgrclose (mgr), NULL;
917 memset (latchmgr, 0, 1 << bits);
918 latchmgr->alloc->bits = mgr->page_bits;
920 for( lvl=MIN_lvl; lvl--; ) {
921 slotptr(latchmgr->alloc, 1)->off = mgr->page_size - 3;
922 bt_putid(slotptr(latchmgr->alloc, 1)->id, lvl ? MIN_lvl - lvl + 1 : 0); // next(lower) page number
923 key = keyptr(latchmgr->alloc, 1);
924 key->len = 2; // create stopper key
927 latchmgr->alloc->min = mgr->page_size - 3;
928 latchmgr->alloc->lvl = lvl;
929 latchmgr->alloc->cnt = 1;
930 latchmgr->alloc->act = 1;
932 if( write (mgr->idx, latchmgr, mgr->page_size) < mgr->page_size )
933 return bt_mgrclose (mgr), NULL;
935 if( !WriteFile (mgr->idx, (char *)latchmgr, mgr->page_size, amt, NULL) )
936 return bt_mgrclose (mgr), NULL;
938 if( *amt < mgr->page_size )
939 return bt_mgrclose (mgr), NULL;
943 // clear out latch manager locks
944 // and rest of pages to round out segment
946 memset(latchmgr, 0, mgr->page_size);
949 while( last <= ((MIN_lvl + 1 + nlatchpage) | mgr->poolmask) ) {
951 pwrite(mgr->idx, latchmgr, mgr->page_size, last << mgr->page_bits);
953 SetFilePointer (mgr->idx, last << mgr->page_bits, NULL, FILE_BEGIN);
954 if( !WriteFile (mgr->idx, (char *)latchmgr, mgr->page_size, amt, NULL) )
955 return bt_mgrclose (mgr), NULL;
956 if( *amt < mgr->page_size )
957 return bt_mgrclose (mgr), NULL;
964 flag = PROT_READ | PROT_WRITE;
965 mgr->latchmgr = mmap (0, mgr->page_size, flag, MAP_SHARED, mgr->idx, ALLOC_page * mgr->page_size);
966 if( mgr->latchmgr == MAP_FAILED )
967 return bt_mgrclose (mgr), NULL;
968 mgr->latchsets = (BtLatchSet *)mmap (0, mgr->latchmgr->nlatchpage * mgr->page_size, flag, MAP_SHARED, mgr->idx, LATCH_page * mgr->page_size);
969 if( mgr->latchsets == MAP_FAILED )
970 return bt_mgrclose (mgr), NULL;
972 flag = PAGE_READWRITE;
973 mgr->halloc = CreateFileMapping(mgr->idx, NULL, flag, 0, (BT_latchtable / (mgr->page_size / sizeof(BtLatchSet)) + 1 + LATCH_page) * mgr->page_size, NULL);
975 return bt_mgrclose (mgr), NULL;
977 flag = FILE_MAP_WRITE;
978 mgr->latchmgr = MapViewOfFile(mgr->halloc, flag, 0, 0, (BT_latchtable / (mgr->page_size / sizeof(BtLatchSet)) + 1 + LATCH_page) * mgr->page_size);
980 return GetLastError(), bt_mgrclose (mgr), NULL;
982 mgr->latchsets = (void *)((char *)mgr->latchmgr + LATCH_page * mgr->page_size);
988 VirtualFree (latchmgr, 0, MEM_RELEASE);
993 // open BTree access method
994 // based on buffer manager
996 BtDb *bt_open (BtMgr *mgr)
998 BtDb *bt = malloc (sizeof(*bt));
1000 memset (bt, 0, sizeof(*bt));
1003 bt->mem = malloc (3 *mgr->page_size);
1005 bt->mem = VirtualAlloc(NULL, 3 * mgr->page_size, MEM_COMMIT, PAGE_READWRITE);
1007 bt->frame = (BtPage)bt->mem;
1008 bt->zero = (BtPage)(bt->mem + 1 * mgr->page_size);
1009 bt->cursor = (BtPage)(bt->mem + 2 * mgr->page_size);
1011 memset (bt->zero, 0, mgr->page_size);
1015 // compare two keys, returning > 0, = 0, or < 0
1016 // as the comparison value
1018 int keycmp (BtKey key1, unsigned char *key2, uint len2)
1020 uint len1 = key1->len;
1023 if( ans = memcmp (key1->key, key2, len1 > len2 ? len2 : len1) )
1036 // find segment in pool
1037 // must be called with hashslot idx locked
1038 // return NULL if not there
1039 // otherwise return node
1041 BtPool *bt_findpool(BtDb *bt, uid page_no, uint idx)
1046 // compute start of hash chain in pool
1048 if( slot = bt->mgr->hash[idx] )
1049 pool = bt->mgr->pool + slot;
1053 page_no &= ~bt->mgr->poolmask;
1055 while( pool->basepage != page_no )
1056 if( pool = pool->hashnext )
1064 // add segment to hash table
1066 void bt_linkhash(BtDb *bt, BtPool *pool, uid page_no, int idx)
1071 pool->hashprev = pool->hashnext = NULL;
1072 pool->basepage = page_no & ~bt->mgr->poolmask;
1075 if( slot = bt->mgr->hash[idx] ) {
1076 node = bt->mgr->pool + slot;
1077 pool->hashnext = node;
1078 node->hashprev = pool;
1081 bt->mgr->hash[idx] = pool->slot;
1084 // find best segment to evict from buffer pool
1086 BtPool *bt_findlru (BtDb *bt, uint hashslot)
1088 unsigned long long int target = ~0LL;
1089 BtPool *pool = NULL, *node;
1094 node = bt->mgr->pool + hashslot;
1096 // scan pool entries under hash table slot
1101 if( node->lru > target )
1105 } while( node = node->hashnext );
1110 // map new buffer pool segment to virtual memory
1112 BTERR bt_mapsegment(BtDb *bt, BtPool *pool, uid page_no)
1114 off64_t off = (page_no & ~bt->mgr->poolmask) << bt->mgr->page_bits;
1115 off64_t limit = off + ((bt->mgr->poolmask+1) << bt->mgr->page_bits);
1119 flag = PROT_READ | ( bt->mgr->mode == BT_ro ? 0 : PROT_WRITE );
1120 pool->map = mmap (0, (bt->mgr->poolmask+1) << bt->mgr->page_bits, flag, MAP_SHARED, bt->mgr->idx, off);
1121 if( pool->map == MAP_FAILED )
1122 return bt->err = BTERR_map;
1125 flag = ( bt->mgr->mode == BT_ro ? PAGE_READONLY : PAGE_READWRITE );
1126 pool->hmap = CreateFileMapping(bt->mgr->idx, NULL, flag, (DWORD)(limit >> 32), (DWORD)limit, NULL);
1128 return bt->err = BTERR_map;
1130 flag = ( bt->mgr->mode == BT_ro ? FILE_MAP_READ : FILE_MAP_WRITE );
1131 pool->map = MapViewOfFile(pool->hmap, flag, (DWORD)(off >> 32), (DWORD)off, (bt->mgr->poolmask+1) << bt->mgr->page_bits);
1133 return bt->err = BTERR_map;
1138 // calculate page within pool
1140 BtPage bt_page (BtDb *bt, BtPool *pool, uid page_no)
1142 uint subpage = (uint)(page_no & bt->mgr->poolmask); // page within mapping
1145 page = (BtPage)(pool->map + (subpage << bt->mgr->page_bits));
1151 void bt_unpinpool (BtPool *pool)
1154 __sync_fetch_and_add(&pool->pin, -1);
1156 _InterlockedDecrement16 (&pool->pin);
1160 // find or place requested page in segment-pool
1161 // return pool table entry, incrementing pin
1163 BtPool *bt_pinpool(BtDb *bt, uid page_no)
1165 BtPool *pool, *node, *next;
1166 uint slot, idx, victim;
1168 // lock hash table chain
1170 idx = (uint)(page_no >> bt->mgr->seg_bits) % bt->mgr->hashsize;
1171 bt_spinreadlock (&bt->mgr->latch[idx], 1);
1173 // look up in hash table
1175 if( pool = bt_findpool(bt, page_no, idx) ) {
1177 __sync_fetch_and_add(&pool->pin, 1);
1179 _InterlockedIncrement16 (&pool->pin);
1181 bt_spinreleaseread (&bt->mgr->latch[idx], 1);
1186 // upgrade to write lock
1188 bt_spinreleaseread (&bt->mgr->latch[idx], 1);
1189 bt_spinwritelock (&bt->mgr->latch[idx], 1);
1191 // try to find page in pool with write lock
1193 if( pool = bt_findpool(bt, page_no, idx) ) {
1195 __sync_fetch_and_add(&pool->pin, 1);
1197 _InterlockedIncrement16 (&pool->pin);
1199 bt_spinreleasewrite (&bt->mgr->latch[idx], 1);
1204 // allocate a new pool node
1205 // and add to hash table
1208 slot = __sync_fetch_and_add(&bt->mgr->poolcnt, 1);
1210 slot = _InterlockedIncrement16 (&bt->mgr->poolcnt) - 1;
1213 if( ++slot < bt->mgr->poolmax ) {
1214 pool = bt->mgr->pool + slot;
1217 if( bt_mapsegment(bt, pool, page_no) )
1220 bt_linkhash(bt, pool, page_no, idx);
1222 __sync_fetch_and_add(&pool->pin, 1);
1224 _InterlockedIncrement16 (&pool->pin);
1226 bt_spinreleasewrite (&bt->mgr->latch[idx], 1);
1230 // pool table is full
1231 // find best pool entry to evict
1234 __sync_fetch_and_add(&bt->mgr->poolcnt, -1);
1236 _InterlockedDecrement16 (&bt->mgr->poolcnt);
1241 victim = __sync_fetch_and_add(&bt->mgr->evicted, 1);
1243 victim = _InterlockedIncrement16 (&bt->mgr->evicted) - 1;
1245 victim %= bt->mgr->hashsize;
1247 // try to get write lock
1248 // skip entry if not obtained
1250 if( !bt_spinwritetry (&bt->mgr->latch[victim]) )
1253 // if pool entry is empty
1254 // or any pages are pinned
1257 if( !(pool = bt_findlru(bt, bt->mgr->hash[victim])) ) {
1258 bt_spinreleasewrite (&bt->mgr->latch[victim], 1);
1262 // unlink victim pool node from hash table
1264 if( node = pool->hashprev )
1265 node->hashnext = pool->hashnext;
1266 else if( node = pool->hashnext )
1267 bt->mgr->hash[victim] = node->slot;
1269 bt->mgr->hash[victim] = 0;
1271 if( node = pool->hashnext )
1272 node->hashprev = pool->hashprev;
1274 bt_spinreleasewrite (&bt->mgr->latch[victim], 1);
1276 // remove old file mapping
1278 munmap (pool->map, (bt->mgr->poolmask+1) << bt->mgr->page_bits);
1280 FlushViewOfFile(pool->map, 0);
1281 UnmapViewOfFile(pool->map);
1282 CloseHandle(pool->hmap);
1286 // create new pool mapping
1287 // and link into hash table
1289 if( bt_mapsegment(bt, pool, page_no) )
1292 bt_linkhash(bt, pool, page_no, idx);
1294 __sync_fetch_and_add(&pool->pin, 1);
1296 _InterlockedIncrement16 (&pool->pin);
1298 bt_spinreleasewrite (&bt->mgr->latch[idx], 1);
1303 // place write, read, or parent lock on requested page_no.
1305 void bt_lockpage(BtLock mode, BtLatchSet *set)
1309 bt_spinreadlock (set->readwr, 0);
1312 bt_spinwritelock (set->readwr, 0);
1315 bt_spinreadlock (set->access, 0);
1318 bt_spinwritelock (set->access, 0);
1321 bt_spinwritelock (set->parent, 0);
1323 case BtLockParentWrt:
1324 bt_spinwritelock (set->parent, 0);
1325 bt_spinwritelock (set->readwr, 0);
1330 // remove write, read, or parent lock on requested page
1332 void bt_unlockpage(BtLock mode, BtLatchSet *set)
1336 bt_spinreleaseread (set->readwr, 0);
1339 bt_spinreleasewrite (set->readwr, 0);
1342 bt_spinreleaseread (set->access, 0);
1345 bt_spinreleasewrite (set->access, 0);
1348 bt_spinreleasewrite (set->parent, 0);
1350 case BtLockParentWrt:
1351 bt_spinreleasewrite (set->parent, 0);
1352 bt_spinreleasewrite (set->readwr, 0);
1357 // allocate a new page and write page into it
1359 uid bt_newpage(BtDb *bt, BtPage page)
1365 // lock allocation page
1367 bt_spinwritelock(bt->mgr->latchmgr->lock, 0);
1369 // use empty chain first
1370 // else allocate empty page
1372 if( new_page = bt_getid(bt->mgr->latchmgr->alloc[1].right) ) {
1373 if( set->pool = bt_pinpool (bt, new_page) )
1374 set->page = bt_page (bt, set->pool, new_page);
1378 bt_putid(bt->mgr->latchmgr->alloc[1].right, bt_getid(set->page->right));
1379 bt_unpinpool (set->pool);
1382 new_page = bt_getid(bt->mgr->latchmgr->alloc->right);
1383 bt_putid(bt->mgr->latchmgr->alloc->right, new_page+1);
1387 if ( pwrite(bt->mgr->idx, page, bt->mgr->page_size, new_page << bt->mgr->page_bits) < bt->mgr->page_size )
1388 return bt->err = BTERR_wrt, 0;
1390 // if writing first page of pool block, zero last page in the block
1392 if ( !reuse && bt->mgr->poolmask > 0 && (new_page & bt->mgr->poolmask) == 0 )
1394 // use zero buffer to write zeros
1395 if ( pwrite(bt->mgr->idx,bt->zero, bt->mgr->page_size, (new_page | bt->mgr->poolmask) << bt->mgr->page_bits) < bt->mgr->page_size )
1396 return bt->err = BTERR_wrt, 0;
1399 // bring new page into pool and copy page.
1400 // this will extend the file into the new pages.
1402 if( set->pool = bt_pinpool (bt, new_page) )
1403 set->page = bt_page (bt, set->pool, new_page);
1407 memcpy(set->page, page, bt->mgr->page_size);
1408 bt_unpinpool (set->pool);
1410 // unlock allocation latch and return new page no
1412 bt_spinreleasewrite(bt->mgr->latchmgr->lock, 0);
1416 // find slot in page for given key at a given level
1418 int bt_findslot (BtPageSet *set, unsigned char *key, uint len)
1420 uint diff, higher = set->page->cnt, low = 1, slot;
1423 // make stopper key an infinite fence value
1425 if( bt_getid (set->page->right) )
1430 // low is the lowest candidate.
1431 // loop ends when they meet
1433 // higher is already
1434 // tested as .ge. the passed key.
1436 while( diff = higher - low ) {
1437 slot = low + ( diff >> 1 );
1438 if( keycmp (keyptr(set->page, slot), key, len) < 0 )
1441 higher = slot, good++;
1444 // return zero if key is on right link page
1446 return good ? higher : 0;
1449 // find and load page at given level for given key
1450 // leave page rd or wr locked as requested
1452 int bt_loadpage (BtDb *bt, BtPageSet *set, unsigned char *key, uint len, uint lvl, BtLock lock)
1454 uid page_no = ROOT_page, prevpage = 0;
1455 uint drill = 0xff, slot;
1456 BtLatchSet *prevlatch;
1457 uint mode, prevmode;
1460 // start at root of btree and drill down
1463 // determine lock mode of drill level
1464 mode = (drill == lvl) ? lock : BtLockRead;
1466 set->latch = bt_pinlatch (bt, page_no);
1467 set->page_no = page_no;
1469 // pin page contents
1471 if( set->pool = bt_pinpool (bt, page_no) )
1472 set->page = bt_page (bt, set->pool, page_no);
1476 // obtain access lock using lock chaining with Access mode
1478 if( page_no > ROOT_page )
1479 bt_lockpage(BtLockAccess, set->latch);
1481 // release & unpin parent page
1484 bt_unlockpage(prevmode, prevlatch);
1485 bt_unpinlatch (prevlatch);
1486 bt_unpinpool (prevpool);
1490 // obtain read lock using lock chaining
1492 bt_lockpage(mode, set->latch);
1494 if( set->page->free )
1495 return bt->err = BTERR_struct, 0;
1497 if( page_no > ROOT_page )
1498 bt_unlockpage(BtLockAccess, set->latch);
1500 // re-read and re-lock root after determining actual level of root
1502 if( set->page->lvl != drill) {
1503 if ( set->page_no != ROOT_page )
1504 return bt->err = BTERR_struct, 0;
1506 drill = set->page->lvl;
1508 if( lock != BtLockRead && drill == lvl ) {
1509 bt_unlockpage(mode, set->latch);
1510 bt_unpinlatch (set->latch);
1511 bt_unpinpool (set->pool);
1516 prevpage = set->page_no;
1517 prevlatch = set->latch;
1518 prevpool = set->pool;
1521 // find key on page at this level
1522 // and descend to requested level
1524 if( !set->page->kill )
1525 if( slot = bt_findslot (set, key, len) ) {
1529 while( slotptr(set->page, slot)->dead )
1530 if( slot++ < set->page->cnt )
1535 page_no = bt_getid(slotptr(set->page, slot)->id);
1540 // or slide right into next page
1543 page_no = bt_getid(set->page->right);
1547 // return error on end of right chain
1549 bt->err = BTERR_struct;
1550 return 0; // return error
1553 // return page to free list
1554 // page must be delete & write locked
1556 void bt_freepage (BtDb *bt, BtPageSet *set)
1558 // lock allocation page
1560 bt_spinwritelock (bt->mgr->latchmgr->lock, 0);
1562 // store chain in second right
1563 bt_putid(set->page->right, bt_getid(bt->mgr->latchmgr->alloc[1].right));
1564 bt_putid(bt->mgr->latchmgr->alloc[1].right, set->page_no);
1565 set->page->free = 1;
1567 // unlock released page
1569 bt_unlockpage (BtLockDelete, set->latch);
1570 bt_unlockpage (BtLockWrite, set->latch);
1571 bt_unpinlatch (set->latch);
1572 bt_unpinpool (set->pool);
1574 // unlock allocation page
1576 bt_spinreleasewrite (bt->mgr->latchmgr->lock, 0);
1579 // a fence key was deleted from a page
1580 // push new fence value upwards
1582 BTERR bt_fixfence (BtDb *bt, BtPageSet *set, uint lvl)
1584 unsigned char leftkey[256], rightkey[256];
1589 // remove the old fence value
1591 ptr = keyptr(set->page, set->page->cnt);
1592 memcpy (rightkey, ptr, ptr->len + 1);
1594 memset (slotptr(set->page, set->page->cnt--), 0, sizeof(BtSlot));
1595 set->page->dirty = 1;
1597 ptr = keyptr(set->page, set->page->cnt);
1598 memcpy (leftkey, ptr, ptr->len + 1);
1599 page_no = set->page_no;
1601 bt_unlockpage (BtLockWrite, set->latch);
1603 // insert new (now smaller) fence key
1605 if( bt_insertkey (bt, leftkey+1, *leftkey, lvl+1, page_no, time(NULL)) )
1608 // now delete old fence key
1610 if( bt_deletekey (bt, rightkey+1, *rightkey, lvl+1) )
1613 bt_unlockpage (BtLockParent, set->latch);
1614 bt_unpinlatch(set->latch);
1615 bt_unpinpool (set->pool);
1619 // root has a single child
1620 // collapse a level from the tree
1622 BTERR bt_collapseroot (BtDb *bt, BtPageSet *root)
1627 // find the child entry and promote as new root contents
1630 for( idx = 0; idx++ < root->page->cnt; )
1631 if( !slotptr(root->page, idx)->dead )
1634 child->page_no = bt_getid (slotptr(root->page, idx)->id);
1636 child->latch = bt_pinlatch (bt, child->page_no);
1637 bt_lockpage (BtLockDelete, child->latch);
1638 bt_lockpage (BtLockWrite, child->latch);
1640 if( child->pool = bt_pinpool (bt, child->page_no) )
1641 child->page = bt_page (bt, child->pool, child->page_no);
1645 memcpy (root->page, child->page, bt->mgr->page_size);
1646 bt_freepage (bt, child);
1648 } while( root->page->lvl > 1 && root->page->act == 1 );
1650 bt_unlockpage (BtLockParentWrt, root->latch);
1651 bt_unpinlatch (root->latch);
1652 bt_unpinpool (root->pool);
1656 // find and delete key on page by marking delete flag bit
1657 // if page becomes empty, delete it from the btree
1659 BTERR bt_deletekey (BtDb *bt, unsigned char *key, uint len, uint lvl)
1661 unsigned char lowerfence[256], higherfence[256];
1662 uint slot, idx, dirty = 0, fence, found;
1663 BtPageSet set[1], right[1];
1666 if( slot = bt_loadpage (bt, set, key, len, lvl, BtLockParentWrt) )
1667 ptr = keyptr(set->page, slot);
1671 // are we deleting a fence slot?
1673 fence = slot == set->page->cnt;
1675 // if key is found delete it, otherwise ignore request
1677 if( found = !keycmp (ptr, key, len) )
1678 if( found = slotptr(set->page, slot)->dead == 0 ) {
1679 dirty = slotptr(set->page, slot)->dead = 1;
1680 set->page->dirty = 1;
1683 // collapse empty slots
1685 while( idx = set->page->cnt - 1 )
1686 if( slotptr(set->page, idx)->dead ) {
1687 *slotptr(set->page, idx) = *slotptr(set->page, idx + 1);
1688 memset (slotptr(set->page, set->page->cnt--), 0, sizeof(BtSlot));
1693 // did we delete a fence key in an upper level?
1695 if( dirty && lvl && set->page->act && fence )
1696 if( bt_fixfence (bt, set, lvl) )
1699 return bt->found = found, 0;
1701 // is this a collapsed root?
1703 if( lvl > 1 && set->page_no == ROOT_page && set->page->act == 1 )
1704 if( bt_collapseroot (bt, set) )
1707 return bt->found = found, 0;
1709 // return if page is not empty
1711 if( set->page->act ) {
1712 bt_unlockpage(BtLockParentWrt, set->latch);
1713 bt_unpinlatch (set->latch);
1714 bt_unpinpool (set->pool);
1715 return bt->found = found, 0;
1718 // cache copy of fence key
1719 // to post in parent
1721 ptr = keyptr(set->page, set->page->cnt);
1722 memcpy (lowerfence, ptr, ptr->len + 1);
1724 // obtain lock on right page
1726 right->page_no = bt_getid(set->page->right);
1727 right->latch = bt_pinlatch (bt, right->page_no);
1728 bt_lockpage (BtLockParentWrt, right->latch);
1730 // pin page contents
1732 if( right->pool = bt_pinpool (bt, right->page_no) )
1733 right->page = bt_page (bt, right->pool, right->page_no);
1737 if( right->page->kill )
1738 return bt->err = BTERR_struct;
1740 // pull contents of right peer into our empty page
1742 memcpy (set->page, right->page, bt->mgr->page_size);
1744 // cache copy of key to update
1746 ptr = keyptr(right->page, right->page->cnt);
1747 memcpy (higherfence, ptr, ptr->len + 1);
1749 // mark right page deleted and point it to left page
1750 // until we can post parent updates
1752 bt_putid (right->page->right, set->page_no);
1753 right->page->kill = 1;
1755 bt_unlockpage (BtLockWrite, right->latch);
1756 bt_unlockpage (BtLockWrite, set->latch);
1758 // redirect higher key directly to our new node contents
1760 if( bt_insertkey (bt, higherfence+1, *higherfence, lvl+1, set->page_no, time(NULL)) )
1763 // delete old lower key to our node
1765 if( bt_deletekey (bt, lowerfence+1, *lowerfence, lvl+1) )
1768 // obtain delete and write locks to right node
1770 bt_unlockpage (BtLockParent, right->latch);
1771 bt_lockpage (BtLockDelete, right->latch);
1772 bt_lockpage (BtLockWrite, right->latch);
1773 bt_freepage (bt, right);
1775 bt_unlockpage (BtLockParent, set->latch);
1776 bt_unpinlatch (set->latch);
1777 bt_unpinpool (set->pool);
1782 // find key in leaf level and return row-id
1784 uid bt_findkey (BtDb *bt, unsigned char *key, uint len)
1791 if( slot = bt_loadpage (bt, set, key, len, 0, BtLockRead) )
1792 ptr = keyptr(set->page, slot);
1796 // if key exists, return row-id
1797 // otherwise return 0
1799 if( slot <= set->page->cnt )
1800 if( !keycmp (ptr, key, len) )
1801 id = bt_getid(slotptr(set->page,slot)->id);
1803 bt_unlockpage (BtLockRead, set->latch);
1804 bt_unpinlatch (set->latch);
1805 bt_unpinpool (set->pool);
1809 // check page for space available,
1810 // clean if necessary and return
1811 // 0 - page needs splitting
1812 // >0 new slot value
1814 uint bt_cleanpage(BtDb *bt, BtPage page, uint amt, uint slot)
1816 uint nxt = bt->mgr->page_size;
1817 uint cnt = 0, idx = 0;
1818 uint max = page->cnt;
1822 if( page->min >= (max+1) * sizeof(BtSlot) + sizeof(*page) + amt + 1 )
1825 // skip cleanup if nothing to reclaim
1830 memcpy (bt->frame, page, bt->mgr->page_size);
1832 // skip page info and set rest of page to zero
1834 memset (page+1, 0, bt->mgr->page_size - sizeof(*page));
1838 // try cleaning up page first
1839 // by removing deleted keys
1841 while( cnt++ < max ) {
1844 if( cnt < max && slotptr(bt->frame,cnt)->dead )
1847 // copy the key across
1849 key = keyptr(bt->frame, cnt);
1850 nxt -= key->len + 1;
1851 memcpy ((unsigned char *)page + nxt, key, key->len + 1);
1855 memcpy(slotptr(page, ++idx)->id, slotptr(bt->frame, cnt)->id, BtId);
1856 if( !(slotptr(page, idx)->dead = slotptr(bt->frame, cnt)->dead) )
1858 slotptr(page, idx)->tod = slotptr(bt->frame, cnt)->tod;
1859 slotptr(page, idx)->off = nxt;
1865 // see if page has enough space now, or does it need splitting?
1867 if( page->min >= (idx+1) * sizeof(BtSlot) + sizeof(*page) + amt + 1 )
1873 // split the root and raise the height of the btree
1875 BTERR bt_splitroot(BtDb *bt, BtPageSet *root, unsigned char *leftkey, uid page_no2)
1877 uint nxt = bt->mgr->page_size;
1880 // Obtain an empty page to use, and copy the current
1881 // root contents into it, e.g. lower keys
1883 if( !(left = bt_newpage(bt, root->page)) )
1886 // preserve the page info at the bottom
1887 // of higher keys and set rest to zero
1889 memset(root->page+1, 0, bt->mgr->page_size - sizeof(*root->page));
1891 // insert lower keys page fence key on newroot page as first key
1893 nxt -= *leftkey + 1;
1894 memcpy ((unsigned char *)root->page + nxt, leftkey, *leftkey + 1);
1895 bt_putid(slotptr(root->page, 1)->id, left);
1896 slotptr(root->page, 1)->off = nxt;
1898 // insert stopper key on newroot page
1899 // and increase the root height
1902 ((unsigned char *)root->page)[nxt] = 2;
1903 ((unsigned char *)root->page)[nxt+1] = 0xff;
1904 ((unsigned char *)root->page)[nxt+2] = 0xff;
1905 bt_putid(slotptr(root->page, 2)->id, page_no2);
1906 slotptr(root->page, 2)->off = nxt;
1908 bt_putid(root->page->right, 0);
1909 root->page->min = nxt; // reset lowest used offset and key count
1910 root->page->cnt = 2;
1911 root->page->act = 2;
1914 // release and unpin root
1916 bt_unlockpage(BtLockWrite, root->latch);
1917 bt_unpinlatch (root->latch);
1918 bt_unpinpool (root->pool);
1922 // split already locked full node
1925 BTERR bt_splitpage (BtDb *bt, BtPageSet *set)
1927 uint cnt = 0, idx = 0, max, nxt = bt->mgr->page_size;
1928 unsigned char fencekey[256], rightkey[256];
1929 uint lvl = set->page->lvl;
1934 // split higher half of keys to bt->frame
1936 memset (bt->frame, 0, bt->mgr->page_size);
1937 max = set->page->cnt;
1941 while( cnt++ < max ) {
1942 key = keyptr(set->page, cnt);
1943 nxt -= key->len + 1;
1944 memcpy ((unsigned char *)bt->frame + nxt, key, key->len + 1);
1946 memcpy(slotptr(bt->frame,++idx)->id, slotptr(set->page,cnt)->id, BtId);
1947 if( !(slotptr(bt->frame, idx)->dead = slotptr(set->page, cnt)->dead) )
1949 slotptr(bt->frame, idx)->tod = slotptr(set->page, cnt)->tod;
1950 slotptr(bt->frame, idx)->off = nxt;
1953 // remember existing fence key for new page to the right
1955 memcpy (rightkey, key, key->len + 1);
1957 bt->frame->bits = bt->mgr->page_bits;
1958 bt->frame->min = nxt;
1959 bt->frame->cnt = idx;
1960 bt->frame->lvl = lvl;
1964 if( set->page_no > ROOT_page )
1965 memcpy (bt->frame->right, set->page->right, BtId);
1967 // get new free page and write higher keys to it.
1969 if( !(right->page_no = bt_newpage(bt, bt->frame)) )
1972 // update lower keys to continue in old page
1974 memcpy (bt->frame, set->page, bt->mgr->page_size);
1975 memset (set->page+1, 0, bt->mgr->page_size - sizeof(*set->page));
1976 nxt = bt->mgr->page_size;
1977 set->page->dirty = 0;
1982 // assemble page of smaller keys
1984 while( cnt++ < max / 2 ) {
1985 key = keyptr(bt->frame, cnt);
1986 nxt -= key->len + 1;
1987 memcpy ((unsigned char *)set->page + nxt, key, key->len + 1);
1988 memcpy(slotptr(set->page,++idx)->id, slotptr(bt->frame,cnt)->id, BtId);
1989 slotptr(set->page, idx)->tod = slotptr(bt->frame, cnt)->tod;
1990 slotptr(set->page, idx)->off = nxt;
1994 // remember fence key for smaller page
1996 memcpy(fencekey, key, key->len + 1);
1998 bt_putid(set->page->right, right->page_no);
1999 set->page->min = nxt;
2000 set->page->cnt = idx;
2002 // if current page is the root page, split it
2004 if( set->page_no == ROOT_page )
2005 return bt_splitroot (bt, set, fencekey, right->page_no);
2007 // insert new fences in their parent pages
2009 right->latch = bt_pinlatch (bt, right->page_no);
2010 bt_lockpage (BtLockParent, right->latch);
2012 bt_lockpage (BtLockParent, set->latch);
2013 bt_unlockpage (BtLockWrite, set->latch);
2015 // insert new fence for reformulated left block of smaller keys
2017 if( bt_insertkey (bt, fencekey+1, *fencekey, lvl+1, set->page_no, time(NULL)) )
2020 // switch fence for right block of larger keys to new right page
2022 if( bt_insertkey (bt, rightkey+1, *rightkey, lvl+1, right->page_no, time(NULL)) )
2025 bt_unlockpage (BtLockParent, set->latch);
2026 bt_unpinlatch (set->latch);
2027 bt_unpinpool (set->pool);
2029 bt_unlockpage (BtLockParent, right->latch);
2030 bt_unpinlatch (right->latch);
2033 // Insert new key into the btree at given level.
2035 BTERR bt_insertkey (BtDb *bt, unsigned char *key, uint len, uint lvl, uid id, uint tod)
2042 if( slot = bt_loadpage (bt, set, key, len, lvl, BtLockWrite) )
2043 ptr = keyptr(set->page, slot);
2047 bt->err = BTERR_ovflw;
2051 // if key already exists, update id and return
2053 if( !keycmp (ptr, key, len) ) {
2054 if( slotptr(set->page, slot)->dead )
2056 slotptr(set->page, slot)->dead = 0;
2057 slotptr(set->page, slot)->tod = tod;
2058 bt_putid(slotptr(set->page,slot)->id, id);
2059 bt_unlockpage(BtLockWrite, set->latch);
2060 bt_unpinlatch (set->latch);
2061 bt_unpinpool (set->pool);
2065 // check if page has enough space
2067 if( slot = bt_cleanpage (bt, set->page, len, slot) )
2070 if( bt_splitpage (bt, set) )
2074 // calculate next available slot and copy key into page
2076 set->page->min -= len + 1; // reset lowest used offset
2077 ((unsigned char *)set->page)[set->page->min] = len;
2078 memcpy ((unsigned char *)set->page + set->page->min +1, key, len );
2080 for( idx = slot; idx < set->page->cnt; idx++ )
2081 if( slotptr(set->page, idx)->dead )
2084 // now insert key into array before slot
2086 if( idx == set->page->cnt )
2087 idx++, set->page->cnt++;
2092 *slotptr(set->page, idx) = *slotptr(set->page, idx -1), idx--;
2094 bt_putid(slotptr(set->page,slot)->id, id);
2095 slotptr(set->page, slot)->off = set->page->min;
2096 slotptr(set->page, slot)->tod = tod;
2097 slotptr(set->page, slot)->dead = 0;
2099 bt_unlockpage (BtLockWrite, set->latch);
2100 bt_unpinlatch (set->latch);
2101 bt_unpinpool (set->pool);
2105 // cache page of keys into cursor and return starting slot for given key
2107 uint bt_startkey (BtDb *bt, unsigned char *key, uint len)
2112 // cache page for retrieval
2114 if( slot = bt_loadpage (bt, set, key, len, 0, BtLockRead) )
2115 memcpy (bt->cursor, set->page, bt->mgr->page_size);
2119 bt->cursor_page = set->page_no;
2121 bt_unlockpage(BtLockRead, set->latch);
2122 bt_unpinlatch (set->latch);
2123 bt_unpinpool (set->pool);
2127 // return next slot for cursor page
2128 // or slide cursor right into next page
2130 uint bt_nextkey (BtDb *bt, uint slot)
2136 right = bt_getid(bt->cursor->right);
2138 while( slot++ < bt->cursor->cnt )
2139 if( slotptr(bt->cursor,slot)->dead )
2141 else if( right || (slot < bt->cursor->cnt) ) // skip infinite stopper
2149 bt->cursor_page = right;
2151 if( set->pool = bt_pinpool (bt, right) )
2152 set->page = bt_page (bt, set->pool, right);
2156 set->latch = bt_pinlatch (bt, right);
2157 bt_lockpage(BtLockRead, set->latch);
2159 memcpy (bt->cursor, set->page, bt->mgr->page_size);
2161 bt_unlockpage(BtLockRead, set->latch);
2162 bt_unpinlatch (set->latch);
2163 bt_unpinpool (set->pool);
2171 BtKey bt_key(BtDb *bt, uint slot)
2173 return keyptr(bt->cursor, slot);
2176 uid bt_uid(BtDb *bt, uint slot)
2178 return bt_getid(slotptr(bt->cursor,slot)->id);
2181 uint bt_tod(BtDb *bt, uint slot)
2183 return slotptr(bt->cursor,slot)->tod;
2189 double getCpuTime(int type)
2192 FILETIME xittime[1];
2193 FILETIME systime[1];
2194 FILETIME usrtime[1];
2195 SYSTEMTIME timeconv[1];
2198 GetProcessTimes (GetCurrentProcess(), crtime, xittime, systime, usrtime);
2199 memset (timeconv, 0, sizeof(SYSTEMTIME));
2203 FileTimeToSystemTime (usrtime, timeconv);
2206 FileTimeToSystemTime (systime, timeconv);
2210 ans = (double)timeconv->wHour * 3600;
2211 ans += (double)timeconv->wMinute * 60;
2212 ans += (double)timeconv->wSecond;
2213 ans += (double)timeconv->wMilliseconds / 1000;
2217 #include <sys/time.h>
2218 #include <sys/resource.h>
2220 double getCpuTime(int type)
2222 struct rusage used[1];
2224 getrusage(RUSAGE_SELF, used);
2227 return (double)used->ru_utime.tv_sec + (double)used->ru_utime.tv_usec / 1000000;
2230 return (double)used->ru_stime.tv_sec + (double)used->ru_stime.tv_usec / 1000000;
2237 void bt_latchaudit (BtDb *bt)
2239 ushort idx, hashidx;
2245 for( idx = 1; idx < bt->mgr->latchmgr->latchdeployed; idx++ ) {
2246 set->latch = bt->mgr->latchsets + idx;
2247 if( set->latch->pin ) {
2248 fprintf(stderr, "latchset %d pinned for page %.6x\n", idx, set->latch->page_no);
2249 set->latch->pin = 0;
2253 for( hashidx = 0; hashidx < bt->mgr->latchmgr->latchhash; hashidx++ ) {
2254 if( idx = bt->mgr->latchmgr->table[hashidx].slot ) do {
2255 set->latch = bt->mgr->latchsets + idx;
2256 if( set->latch->hash != hashidx )
2257 fprintf(stderr, "latchset %d wrong hashidx\n", idx);
2258 if( set->latch->pin )
2259 fprintf(stderr, "latchset %d pinned for page %.8x\n", idx, set->latch->page_no);
2260 } while( idx = set->latch->next );
2263 next = bt->mgr->latchmgr->nlatchpage + LATCH_page;
2264 page_no = LEAF_page;
2266 while( page_no < bt_getid(bt->mgr->latchmgr->alloc->right) ) {
2267 pread (bt->mgr->idx, bt->frame, bt->mgr->page_size, page_no << bt->mgr->page_bits);
2268 if( !bt->frame->free )
2269 for( idx = 0; idx++ < bt->frame->cnt - 1; ) {
2270 ptr = keyptr(bt->frame, idx+1);
2271 if( keycmp (keyptr(bt->frame, idx), ptr->key, ptr->len) >= 0 )
2272 fprintf(stderr, "page %.8x idx %.2x out of order\n", page_no, idx);
2275 if( page_no > LEAF_page )
2289 // standalone program to index file of keys
2290 // then list them onto std-out
2293 void *index_file (void *arg)
2295 uint __stdcall index_file (void *arg)
2298 int line = 0, found = 0, cnt = 0;
2299 uid next, page_no = LEAF_page; // start on first page of leaves
2300 unsigned char key[256];
2301 ThreadArg *args = arg;
2302 int ch, len = 0, slot;
2309 bt = bt_open (args->mgr);
2312 switch(args->type | 0x20)
2315 fprintf(stderr, "started latch mgr audit\n");
2317 fprintf(stderr, "finished latch mgr audit\n");
2321 fprintf(stderr, "started indexing for %s\n", args->infile);
2322 if( in = fopen (args->infile, "rb") )
2323 while( ch = getc(in), ch != EOF )
2328 if( args->num == 1 )
2329 sprintf((char *)key+len, "%.9d", 1000000000 - line), len += 9;
2331 else if( args->num )
2332 sprintf((char *)key+len, "%.9d", line + args->idx * args->num), len += 9;
2334 if( bt_insertkey (bt, key, len, 0, line, *tod) )
2335 fprintf(stderr, "Error %d Line: %d\n", bt->err, line), exit(0);
2338 else if( len < 255 )
2340 fprintf(stderr, "finished %s for %d keys\n", args->infile, line);
2344 fprintf(stderr, "started deleting keys for %s\n", args->infile);
2345 if( in = fopen (args->infile, "rb") )
2346 while( ch = getc(in), ch != EOF )
2350 if( args->num == 1 )
2351 sprintf((char *)key+len, "%.9d", 1000000000 - line), len += 9;
2353 else if( args->num )
2354 sprintf((char *)key+len, "%.9d", line + args->idx * args->num), len += 9;
2356 if( bt_deletekey (bt, key, len, 0) )
2357 fprintf(stderr, "Error %d Line: %d\n", bt->err, line), exit(0);
2360 else if( len < 255 )
2362 fprintf(stderr, "finished %s for keys, %d \n", args->infile, line);
2366 fprintf(stderr, "started finding keys for %s\n", args->infile);
2367 if( in = fopen (args->infile, "rb") )
2368 while( ch = getc(in), ch != EOF )
2372 if( args->num == 1 )
2373 sprintf((char *)key+len, "%.9d", 1000000000 - line), len += 9;
2375 else if( args->num )
2376 sprintf((char *)key+len, "%.9d", line + args->idx * args->num), len += 9;
2378 if( bt_findkey (bt, key, len) )
2381 fprintf(stderr, "Error %d Syserr %d Line: %d\n", bt->err, errno, line), exit(0);
2383 fprintf(stderr, "Unable to find key %.*s line %d\n", len, key, line);
2386 else if( len < 255 )
2388 fprintf(stderr, "finished %s for %d keys, found %d\n", args->infile, line, found);
2392 fprintf(stderr, "started scanning\n");
2394 if( set->pool = bt_pinpool (bt, page_no) )
2395 set->page = bt_page (bt, set->pool, page_no);
2398 set->latch = bt_pinlatch (bt, page_no);
2399 bt_lockpage (BtLockRead, set->latch);
2400 next = bt_getid (set->page->right);
2401 cnt += set->page->act;
2403 for( slot = 0; slot++ < set->page->cnt; )
2404 if( next || slot < set->page->cnt )
2405 if( !slotptr(set->page, slot)->dead ) {
2406 ptr = keyptr(set->page, slot);
2407 fwrite (ptr->key, ptr->len, 1, stdout);
2408 fputc ('\n', stdout);
2411 bt_unlockpage (BtLockRead, set->latch);
2412 bt_unpinlatch (set->latch);
2413 bt_unpinpool (set->pool);
2414 } while( page_no = next );
2416 cnt--; // remove stopper key
2417 fprintf(stderr, " Total keys read %d\n", cnt);
2421 fprintf(stderr, "started counting\n");
2422 next = bt->mgr->latchmgr->nlatchpage + LATCH_page;
2423 page_no = LEAF_page;
2425 while( page_no < bt_getid(bt->mgr->latchmgr->alloc->right) ) {
2426 uid off = page_no << bt->mgr->page_bits;
2428 pread (bt->mgr->idx, bt->frame, bt->mgr->page_size, off);
2432 SetFilePointer (bt->mgr->idx, (long)off, (long*)(&off)+1, FILE_BEGIN);
2434 if( !ReadFile(bt->mgr->idx, bt->frame, bt->mgr->page_size, amt, NULL))
2435 return bt->err = BTERR_map;
2437 if( *amt < bt->mgr->page_size )
2438 return bt->err = BTERR_map;
2440 if( !bt->frame->free && !bt->frame->lvl )
2441 cnt += bt->frame->act;
2442 if( page_no > LEAF_page )
2447 cnt--; // remove stopper key
2448 fprintf(stderr, " Total keys read %d\n", cnt);
2460 typedef struct timeval timer;
2462 int main (int argc, char **argv)
2464 int idx, cnt, len, slot, err;
2465 int segsize, bits = 16;
2470 time_t start[1], stop[1];
2484 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]);
2485 fprintf (stderr, " where page_bits is the page size in bits\n");
2486 fprintf (stderr, " mapped_segments is the number of mmap segments in buffer pool\n");
2487 fprintf (stderr, " seg_bits is the size of individual segments in buffer pool in pages in bits\n");
2488 fprintf (stderr, " line_numbers = 1 to append line numbers to keys\n");
2489 fprintf (stderr, " src_file1 thru src_filen are files of keys separated by newline\n");
2494 gettimeofday(&start, NULL);
2500 bits = atoi(argv[3]);
2503 poolsize = atoi(argv[4]);
2506 fprintf (stderr, "Warning: no mapped_pool\n");
2508 if( poolsize > 65535 )
2509 fprintf (stderr, "Warning: mapped_pool > 65535 segments\n");
2512 segsize = atoi(argv[5]);
2514 segsize = 4; // 16 pages per mmap segment
2517 num = atoi(argv[6]);
2521 threads = malloc (cnt * sizeof(pthread_t));
2523 threads = GlobalAlloc (GMEM_FIXED|GMEM_ZEROINIT, cnt * sizeof(HANDLE));
2525 args = malloc (cnt * sizeof(ThreadArg));
2527 mgr = bt_mgr ((argv[1]), BT_rw, bits, poolsize, segsize, poolsize / 8);
2530 fprintf(stderr, "Index Open Error %s\n", argv[1]);
2536 for( idx = 0; idx < cnt; idx++ ) {
2537 args[idx].infile = argv[idx + 7];
2538 args[idx].type = argv[2][0];
2539 args[idx].mgr = mgr;
2540 args[idx].num = num;
2541 args[idx].idx = idx;
2543 if( err = pthread_create (threads + idx, NULL, index_file, args + idx) )
2544 fprintf(stderr, "Error creating thread %d\n", err);
2546 threads[idx] = (HANDLE)_beginthreadex(NULL, 65536, index_file, args + idx, 0, NULL);
2550 // wait for termination
2553 for( idx = 0; idx < cnt; idx++ )
2554 pthread_join (threads[idx], NULL);
2555 gettimeofday(&stop, NULL);
2556 real_time = 1000.0 * ( stop.tv_sec - start.tv_sec ) + 0.001 * (stop.tv_usec - start.tv_usec );
2558 WaitForMultipleObjects (cnt, threads, TRUE, INFINITE);
2560 for( idx = 0; idx < cnt; idx++ )
2561 CloseHandle(threads[idx]);
2564 real_time = 1000 * (*stop - *start);
2566 elapsed = real_time / 1000;
2567 fprintf(stderr, " real %dm%.3fs\n", (int)(elapsed/60), elapsed - (int)(elapsed/60)*60);
2568 elapsed = getCpuTime(1);
2569 fprintf(stderr, " user %dm%.3fs\n", (int)(elapsed/60), elapsed - (int)(elapsed/60)*60);
2570 elapsed = getCpuTime(2);
2571 fprintf(stderr, " sys %dm%.3fs\n", (int)(elapsed/60), elapsed - (int)(elapsed/60)*60);