1 // btree version threadskv1 sched_yield version
2 // with reworked bt_deletekey code
3 // and phase-fair reader writer lock
6 // author: karl malbrain, malbrain@cal.berkeley.edu
9 This work, including the source code, documentation
10 and related data, is placed into the public domain.
12 The orginal author is Karl Malbrain.
14 THIS SOFTWARE IS PROVIDED AS-IS WITHOUT WARRANTY
15 OF ANY KIND, NOT EVEN THE IMPLIED WARRANTY OF
16 MERCHANTABILITY. THE AUTHOR OF THIS SOFTWARE,
17 ASSUMES _NO_ RESPONSIBILITY FOR ANY CONSEQUENCE
18 RESULTING FROM THE USE, MODIFICATION, OR
19 REDISTRIBUTION OF THIS SOFTWARE.
22 // Please see the project home page for documentation
23 // code.google.com/p/high-concurrency-btree
25 #define _FILE_OFFSET_BITS 64
26 #define _LARGEFILE64_SOURCE
42 #define WIN32_LEAN_AND_MEAN
56 typedef unsigned long long uid;
59 typedef unsigned long long off64_t;
60 typedef unsigned short ushort;
61 typedef unsigned int uint;
64 #define BT_latchtable 128 // number of latch manager slots
66 #define BT_ro 0x6f72 // ro
67 #define BT_rw 0x7772 // rw
69 #define BT_maxbits 24 // maximum page size in bits
70 #define BT_minbits 9 // minimum page size in bits
71 #define BT_minpage (1 << BT_minbits) // minimum page size
72 #define BT_maxpage (1 << BT_maxbits) // maximum page size
75 There are five lock types for each node in three independent sets:
76 1. (set 1) AccessIntent: Sharable. Going to Read the node. Incompatible with NodeDelete.
77 2. (set 1) NodeDelete: Exclusive. About to release the node. Incompatible with AccessIntent.
78 3. (set 2) ReadLock: Sharable. Read the node. Incompatible with WriteLock.
79 4. (set 2) WriteLock: Exclusive. Modify the node. Incompatible with ReadLock and other WriteLocks.
80 5. (set 3) ParentModification: Exclusive. Change the node's parent keys. Incompatible with another ParentModification.
91 // definition for phase-fair reader/writer lock implementation
105 // definition for spin latch implementation
107 // exclusive is set for write access
108 // share is count of read accessors
109 // grant write lock when share == 0
111 volatile typedef struct {
122 // hash table entries
125 BtSpinLatch latch[1];
126 volatile ushort slot; // Latch table entry at head of chain
129 // latch manager table structure
132 RWLock readwr[1]; // read/write page lock
133 RWLock access[1]; // Access Intent/Page delete
134 RWLock parent[1]; // Posting of fence key in parent
135 BtSpinLatch busy[1]; // slot is being moved between chains
136 volatile ushort next; // next entry in hash table chain
137 volatile ushort prev; // prev entry in hash table chain
138 volatile ushort pin; // number of outstanding locks
139 volatile ushort hash; // hash slot entry is under
140 volatile uid page_no; // latch set page number
143 // Define the length of the page and key pointers
147 // Page key slot definition.
149 // If BT_maxbits is 15 or less, you can save 2 bytes
150 // for each key stored by making the two uints
153 // Keys are marked dead, but remain on the page until
154 // it cleanup is called. The fence key (highest key) for
155 // the page is always present, even after cleanup.
158 uint off:BT_maxbits; // page offset for key start
159 uint dead:1; // set for deleted key
162 // The key structure occupies space at the upper end of
163 // each page. It's a length byte followed by the key
168 unsigned char key[1];
171 // the value structure also occupies space at the upper
176 unsigned char value[1];
179 // The first part of an index page.
180 // It is immediately followed
181 // by the BtSlot array of keys.
183 typedef struct BtPage_ {
184 uint cnt; // count of keys in page
185 uint act; // count of active keys
186 uint min; // next key offset
187 unsigned char bits:7; // page size in bits
188 unsigned char free:1; // page is on free chain
189 unsigned char lvl:6; // level of page
190 unsigned char kill:1; // page is being deleted
191 unsigned char dirty:1; // page has deleted keys
192 unsigned char right[BtId]; // page number to right
195 // The memory mapping pool table buffer manager entry
198 uid basepage; // mapped base page number
199 char *map; // mapped memory pointer
200 ushort slot; // slot index in this array
201 ushort pin; // mapped page pin counter
202 void *hashprev; // previous pool entry for the same hash idx
203 void *hashnext; // next pool entry for the same hash idx
205 HANDLE hmap; // Windows memory mapping handle
209 #define CLOCK_bit 0x8000 // bit in pool->pin
211 // The loadpage interface object
214 uid page_no; // current page number
215 BtPage page; // current page pointer
216 BtPool *pool; // current page pool
217 BtLatchSet *latch; // current page latch set
220 // structure for latch manager on ALLOC_page
223 struct BtPage_ alloc[1]; // next page_no in right ptr
224 unsigned char chain[BtId]; // head of free page_nos chain
225 BtSpinLatch lock[1]; // allocation area lite latch
226 ushort latchdeployed; // highest number of latch entries deployed
227 ushort nlatchpage; // number of latch pages at BT_latch
228 ushort latchtotal; // number of page latch entries
229 ushort latchhash; // number of latch hash table slots
230 ushort latchvictim; // next latch entry to examine
231 BtHashEntry table[0]; // the hash table
234 // The object structure for Btree access
237 uint page_size; // page size
238 uint page_bits; // page size in bits
239 uint seg_bits; // seg size in pages in bits
240 uint mode; // read-write mode
246 ushort poolcnt; // highest page pool node in use
247 ushort poolmax; // highest page pool node allocated
248 ushort poolmask; // total number of pages in mmap segment - 1
249 ushort hashsize; // size of Hash Table for pool entries
250 volatile uint evicted; // last evicted hash table slot
251 ushort *hash; // pool index for hash entries
252 BtSpinLatch *latch; // latches for hash table slots
253 BtLatchMgr *latchmgr; // mapped latch page from allocation page
254 BtLatchSet *latchsets; // mapped latch set from latch pages
255 BtPool *pool; // memory pool page segments
257 HANDLE halloc; // allocation and latch table handle
262 BtMgr *mgr; // buffer manager for thread
263 BtPage cursor; // cached frame for start/next (never mapped)
264 BtPage frame; // spare frame for the page split (never mapped)
265 uid cursor_page; // current cursor page number
266 unsigned char *mem; // frame, cursor, page memory buffer
267 int found; // last delete or insert was found
268 int err; // last error
282 extern void bt_close (BtDb *bt);
283 extern BtDb *bt_open (BtMgr *mgr);
284 extern BTERR bt_insertkey (BtDb *bt, unsigned char *key, uint len, uint lvl, void *value, uint vallen);
285 extern BTERR bt_deletekey (BtDb *bt, unsigned char *key, uint len, uint lvl);
286 extern int bt_findkey (BtDb *bt, unsigned char *key, uint keylen, unsigned char *value, uint vallen);
287 extern uint bt_startkey (BtDb *bt, unsigned char *key, uint len);
288 extern uint bt_nextkey (BtDb *bt, uint slot);
291 extern BtMgr *bt_mgr (char *name, uint mode, uint bits, uint poolsize, uint segsize, uint hashsize);
292 void bt_mgrclose (BtMgr *mgr);
294 // Helper functions to return slot values
295 // from the cursor page.
297 extern BtKey bt_key (BtDb *bt, uint slot);
298 extern BtVal bt_val (BtDb *bt, uint slot);
300 // BTree page number constants
301 #define ALLOC_page 0 // allocation & latch manager hash table
302 #define ROOT_page 1 // root of the btree
303 #define LEAF_page 2 // first page of leaves
304 #define LATCH_page 3 // pages for latch manager
306 // Number of levels to create in a new BTree
310 // The page is allocated from low and hi ends.
311 // The key slots are allocated from the bottom,
312 // while the text and value of the key
313 // are allocated from the top. When the two
314 // areas meet, the page is split into two.
316 // A key consists of a length byte, two bytes of
317 // index number (0 - 65534), and up to 253 bytes
318 // of key value. Duplicate keys are discarded.
319 // Associated with each key is a value byte string
320 // containing any value desired.
322 // The b-tree root is always located at page 1.
323 // The first leaf page of level zero is always
324 // located on page 2.
326 // The b-tree pages are linked with next
327 // pointers to facilitate enumerators,
328 // and provide for concurrency.
330 // When to root page fills, it is split in two and
331 // the tree height is raised by a new root at page
332 // one with two keys.
334 // Deleted keys are marked with a dead bit until
335 // page cleanup. The fence key for a leaf node is
338 // Groups of pages called segments from the btree are optionally
339 // cached with a memory mapped pool. A hash table is used to keep
340 // track of the cached segments. This behaviour is controlled
341 // by the cache block size parameter to bt_open.
343 // To achieve maximum concurrency one page is locked at a time
344 // as the tree is traversed to find leaf key in question. The right
345 // page numbers are used in cases where the page is being split,
348 // Page 0 is dedicated to lock for new page extensions,
349 // and chains empty pages together for reuse. It also
350 // contains the latch manager hash table.
352 // The ParentModification lock on a node is obtained to serialize posting
353 // or changing the fence key for a node.
355 // Empty pages are chained together through the ALLOC page and reused.
357 // Access macros to address slot and key values from the page
358 // Page slots use 1 based indexing.
360 #define slotptr(page, slot) (((BtSlot *)(page+1)) + (slot-1))
361 #define keyptr(page, slot) ((BtKey)((unsigned char*)(page) + slotptr(page, slot)->off))
362 #define valptr(page, slot) ((BtVal)(keyptr(page,slot)->key + keyptr(page,slot)->len))
364 void bt_putid(unsigned char *dest, uid id)
369 dest[i] = (unsigned char)id, id >>= 8;
372 uid bt_getid(unsigned char *src)
377 for( i = 0; i < BtId; i++ )
378 id <<= 8, id |= *src++;
383 // Phase-Fair reader/writer lock implementation
385 void WriteLock (RWLock *lock)
390 tix = __sync_fetch_and_add (lock->ticket, 1);
392 tix = _InterlockedExchangeAdd16 (lock->ticket, 1);
394 // wait for our ticket to come up
396 while( tix != lock->serving[0] )
403 w = PRES | (tix & PHID);
405 r = __sync_fetch_and_add (lock->rin, w);
407 r = _InterlockedExchangeAdd16 (lock->rin, w);
409 while( r != *lock->rout )
417 void WriteRelease (RWLock *lock)
420 __sync_fetch_and_and (lock->rin, ~MASK);
422 _InterlockedAnd16 (lock->rin, ~MASK);
427 void ReadLock (RWLock *lock)
431 w = __sync_fetch_and_add (lock->rin, RINC) & MASK;
433 w = _InterlockedExchangeAdd16 (lock->rin, RINC) & MASK;
436 while( w == (*lock->rin & MASK) )
444 void ReadRelease (RWLock *lock)
447 __sync_fetch_and_add (lock->rout, RINC);
449 _InterlockedExchangeAdd16 (lock->rout, RINC);
453 // Spin Latch Manager
455 // wait until write lock mode is clear
456 // and add 1 to the share count
458 void bt_spinreadlock(BtSpinLatch *latch)
464 prev = __sync_fetch_and_add ((ushort *)latch, SHARE);
466 prev = _InterlockedExchangeAdd16((ushort *)latch, SHARE);
468 // see if exclusive request is granted or pending
473 prev = __sync_fetch_and_add ((ushort *)latch, -SHARE);
475 prev = _InterlockedExchangeAdd16((ushort *)latch, -SHARE);
478 } while( sched_yield(), 1 );
480 } while( SwitchToThread(), 1 );
484 // wait for other read and write latches to relinquish
486 void bt_spinwritelock(BtSpinLatch *latch)
492 prev = __sync_fetch_and_or((ushort *)latch, PEND | XCL);
494 prev = _InterlockedOr16((ushort *)latch, PEND | XCL);
497 if( !(prev & ~BOTH) )
501 __sync_fetch_and_and ((ushort *)latch, ~XCL);
503 _InterlockedAnd16((ushort *)latch, ~XCL);
506 } while( sched_yield(), 1 );
508 } while( SwitchToThread(), 1 );
512 // try to obtain write lock
514 // return 1 if obtained,
517 int bt_spinwritetry(BtSpinLatch *latch)
522 prev = __sync_fetch_and_or((ushort *)latch, XCL);
524 prev = _InterlockedOr16((ushort *)latch, XCL);
526 // take write access if all bits are clear
529 if( !(prev & ~BOTH) )
533 __sync_fetch_and_and ((ushort *)latch, ~XCL);
535 _InterlockedAnd16((ushort *)latch, ~XCL);
542 void bt_spinreleasewrite(BtSpinLatch *latch)
545 __sync_fetch_and_and((ushort *)latch, ~BOTH);
547 _InterlockedAnd16((ushort *)latch, ~BOTH);
551 // decrement reader count
553 void bt_spinreleaseread(BtSpinLatch *latch)
556 __sync_fetch_and_add((ushort *)latch, -SHARE);
558 _InterlockedExchangeAdd16((ushort *)latch, -SHARE);
562 // link latch table entry into latch hash table
564 void bt_latchlink (BtDb *bt, ushort hashidx, ushort victim, uid page_no)
566 BtLatchSet *set = bt->mgr->latchsets + victim;
568 if( set->next = bt->mgr->latchmgr->table[hashidx].slot )
569 bt->mgr->latchsets[set->next].prev = victim;
571 bt->mgr->latchmgr->table[hashidx].slot = victim;
572 set->page_no = page_no;
579 void bt_unpinlatch (BtLatchSet *set)
582 __sync_fetch_and_add(&set->pin, -1);
584 _InterlockedDecrement16 (&set->pin);
588 // find existing latchset or inspire new one
589 // return with latchset pinned
591 BtLatchSet *bt_pinlatch (BtDb *bt, uid page_no)
593 ushort hashidx = page_no % bt->mgr->latchmgr->latchhash;
594 ushort slot, avail = 0, victim, idx;
597 // obtain read lock on hash table entry
599 bt_spinreadlock(bt->mgr->latchmgr->table[hashidx].latch);
601 if( slot = bt->mgr->latchmgr->table[hashidx].slot ) do
603 set = bt->mgr->latchsets + slot;
604 if( page_no == set->page_no )
606 } while( slot = set->next );
610 __sync_fetch_and_add(&set->pin, 1);
612 _InterlockedIncrement16 (&set->pin);
616 bt_spinreleaseread (bt->mgr->latchmgr->table[hashidx].latch);
621 // try again, this time with write lock
623 bt_spinwritelock(bt->mgr->latchmgr->table[hashidx].latch);
625 if( slot = bt->mgr->latchmgr->table[hashidx].slot ) do
627 set = bt->mgr->latchsets + slot;
628 if( page_no == set->page_no )
630 if( !set->pin && !avail )
632 } while( slot = set->next );
634 // found our entry, or take over an unpinned one
636 if( slot || (slot = avail) ) {
637 set = bt->mgr->latchsets + slot;
639 __sync_fetch_and_add(&set->pin, 1);
641 _InterlockedIncrement16 (&set->pin);
643 set->page_no = page_no;
644 bt_spinreleasewrite(bt->mgr->latchmgr->table[hashidx].latch);
648 // see if there are any unused entries
650 victim = __sync_fetch_and_add (&bt->mgr->latchmgr->latchdeployed, 1) + 1;
652 victim = _InterlockedIncrement16 (&bt->mgr->latchmgr->latchdeployed);
655 if( victim < bt->mgr->latchmgr->latchtotal ) {
656 set = bt->mgr->latchsets + victim;
658 __sync_fetch_and_add(&set->pin, 1);
660 _InterlockedIncrement16 (&set->pin);
662 bt_latchlink (bt, hashidx, victim, page_no);
663 bt_spinreleasewrite (bt->mgr->latchmgr->table[hashidx].latch);
668 victim = __sync_fetch_and_add (&bt->mgr->latchmgr->latchdeployed, -1);
670 victim = _InterlockedDecrement16 (&bt->mgr->latchmgr->latchdeployed);
672 // find and reuse previous lock entry
676 victim = __sync_fetch_and_add(&bt->mgr->latchmgr->latchvictim, 1);
678 victim = _InterlockedIncrement16 (&bt->mgr->latchmgr->latchvictim) - 1;
680 // we don't use slot zero
682 if( victim %= bt->mgr->latchmgr->latchtotal )
683 set = bt->mgr->latchsets + victim;
687 // take control of our slot
688 // from other threads
690 if( set->pin || !bt_spinwritetry (set->busy) )
695 // try to get write lock on hash chain
696 // skip entry if not obtained
697 // or has outstanding locks
699 if( !bt_spinwritetry (bt->mgr->latchmgr->table[idx].latch) ) {
700 bt_spinreleasewrite (set->busy);
705 bt_spinreleasewrite (set->busy);
706 bt_spinreleasewrite (bt->mgr->latchmgr->table[idx].latch);
710 // unlink our available victim from its hash chain
713 bt->mgr->latchsets[set->prev].next = set->next;
715 bt->mgr->latchmgr->table[idx].slot = set->next;
718 bt->mgr->latchsets[set->next].prev = set->prev;
720 bt_spinreleasewrite (bt->mgr->latchmgr->table[idx].latch);
722 __sync_fetch_and_add(&set->pin, 1);
724 _InterlockedIncrement16 (&set->pin);
726 bt_latchlink (bt, hashidx, victim, page_no);
727 bt_spinreleasewrite (bt->mgr->latchmgr->table[hashidx].latch);
728 bt_spinreleasewrite (set->busy);
733 void bt_mgrclose (BtMgr *mgr)
738 // release mapped pages
739 // note that slot zero is never used
741 for( slot = 1; slot < mgr->poolmax; slot++ ) {
742 pool = mgr->pool + slot;
745 munmap (pool->map, (uid)(mgr->poolmask+1) << mgr->page_bits);
748 FlushViewOfFile(pool->map, 0);
749 UnmapViewOfFile(pool->map);
750 CloseHandle(pool->hmap);
756 munmap (mgr->latchsets, mgr->latchmgr->nlatchpage * mgr->page_size);
757 munmap (mgr->latchmgr, 2 * mgr->page_size);
759 FlushViewOfFile(mgr->latchmgr, 0);
760 UnmapViewOfFile(mgr->latchmgr);
761 CloseHandle(mgr->halloc);
767 free ((void *)mgr->latch);
770 FlushFileBuffers(mgr->idx);
771 CloseHandle(mgr->idx);
772 GlobalFree (mgr->pool);
773 GlobalFree (mgr->hash);
774 GlobalFree ((void *)mgr->latch);
779 // close and release memory
781 void bt_close (BtDb *bt)
788 VirtualFree (bt->mem, 0, MEM_RELEASE);
793 // open/create new btree buffer manager
795 // call with file_name, BT_openmode, bits in page size (e.g. 16),
796 // size of mapped page pool (e.g. 8192)
798 BtMgr *bt_mgr (char *name, uint mode, uint bits, uint poolmax, uint segsize, uint hashsize)
800 uint lvl, attr, cacheblk, last, slot, idx;
801 uint nlatchpage, latchhash;
802 unsigned char value[BtId];
803 BtLatchMgr *latchmgr;
812 SYSTEM_INFO sysinfo[1];
815 // determine sanity of page size and buffer pool
817 if( bits > BT_maxbits )
819 else if( bits < BT_minbits )
823 return NULL; // must have buffer pool
826 mgr = calloc (1, sizeof(BtMgr));
828 mgr->idx = open ((char*)name, O_RDWR | O_CREAT, 0666);
831 return free(mgr), NULL;
833 cacheblk = 4096; // minimum mmap segment size for unix
836 mgr = GlobalAlloc (GMEM_FIXED|GMEM_ZEROINIT, sizeof(BtMgr));
837 attr = FILE_ATTRIBUTE_NORMAL;
838 mgr->idx = CreateFile(name, GENERIC_READ| GENERIC_WRITE, FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, OPEN_ALWAYS, attr, NULL);
840 if( mgr->idx == INVALID_HANDLE_VALUE )
841 return GlobalFree(mgr), NULL;
843 // normalize cacheblk to multiple of sysinfo->dwAllocationGranularity
844 GetSystemInfo(sysinfo);
845 cacheblk = sysinfo->dwAllocationGranularity;
849 latchmgr = malloc (BT_maxpage);
852 // read minimum page size to get root info
854 if( size = lseek (mgr->idx, 0L, 2) ) {
855 if( pread(mgr->idx, latchmgr, BT_minpage, 0) == BT_minpage )
856 bits = latchmgr->alloc->bits;
858 return free(mgr), free(latchmgr), NULL;
859 } else if( mode == BT_ro )
860 return free(latchmgr), bt_mgrclose (mgr), NULL;
862 latchmgr = VirtualAlloc(NULL, BT_maxpage, MEM_COMMIT, PAGE_READWRITE);
863 size = GetFileSize(mgr->idx, amt);
866 if( !ReadFile(mgr->idx, (char *)latchmgr, BT_minpage, amt, NULL) )
867 return bt_mgrclose (mgr), NULL;
868 bits = latchmgr->alloc->bits;
869 } else if( mode == BT_ro )
870 return bt_mgrclose (mgr), NULL;
873 mgr->page_size = 1 << bits;
874 mgr->page_bits = bits;
876 mgr->poolmax = poolmax;
879 if( cacheblk < mgr->page_size )
880 cacheblk = mgr->page_size;
882 // mask for partial memmaps
884 mgr->poolmask = (cacheblk >> bits) - 1;
886 // see if requested size of pages per memmap is greater
888 if( (1 << segsize) > mgr->poolmask )
889 mgr->poolmask = (1 << segsize) - 1;
893 while( (1 << mgr->seg_bits) <= mgr->poolmask )
896 mgr->hashsize = hashsize;
899 mgr->pool = calloc (poolmax, sizeof(BtPool));
900 mgr->hash = calloc (hashsize, sizeof(ushort));
901 mgr->latch = calloc (hashsize, sizeof(BtSpinLatch));
903 mgr->pool = GlobalAlloc (GMEM_FIXED|GMEM_ZEROINIT, poolmax * sizeof(BtPool));
904 mgr->hash = GlobalAlloc (GMEM_FIXED|GMEM_ZEROINIT, hashsize * sizeof(ushort));
905 mgr->latch = GlobalAlloc (GMEM_FIXED|GMEM_ZEROINIT, hashsize * sizeof(BtSpinLatch));
911 // initialize an empty b-tree with latch page, root page, page of leaves
912 // and page(s) of latches
914 memset (latchmgr, 0, 1 << bits);
915 nlatchpage = BT_latchtable / (mgr->page_size / sizeof(BtLatchSet)) + 1;
916 bt_putid(latchmgr->alloc->right, MIN_lvl+1+nlatchpage);
917 latchmgr->alloc->bits = mgr->page_bits;
919 latchmgr->nlatchpage = nlatchpage;
920 latchmgr->latchtotal = nlatchpage * (mgr->page_size / sizeof(BtLatchSet));
922 // initialize latch manager
924 latchhash = (mgr->page_size - sizeof(BtLatchMgr)) / sizeof(BtHashEntry);
926 // size of hash table = total number of latchsets
928 if( latchhash > latchmgr->latchtotal )
929 latchhash = latchmgr->latchtotal;
931 latchmgr->latchhash = latchhash;
934 if( write (mgr->idx, latchmgr, mgr->page_size) < mgr->page_size )
935 return bt_mgrclose (mgr), NULL;
937 if( !WriteFile (mgr->idx, (char *)latchmgr, mgr->page_size, amt, NULL) )
938 return bt_mgrclose (mgr), NULL;
940 if( *amt < mgr->page_size )
941 return bt_mgrclose (mgr), NULL;
944 memset (latchmgr, 0, 1 << bits);
945 latchmgr->alloc->bits = mgr->page_bits;
947 for( lvl=MIN_lvl; lvl--; ) {
948 slotptr(latchmgr->alloc, 1)->off = mgr->page_size - 3 - (lvl ? BtId + 1: 1);
949 key = keyptr(latchmgr->alloc, 1);
950 key->len = 2; // create stopper key
954 bt_putid(value, MIN_lvl - lvl + 1);
955 val = valptr(latchmgr->alloc, 1);
956 val->len = lvl ? BtId : 0;
957 memcpy (val->value, value, val->len);
959 latchmgr->alloc->min = slotptr(latchmgr->alloc, 1)->off;
960 latchmgr->alloc->lvl = lvl;
961 latchmgr->alloc->cnt = 1;
962 latchmgr->alloc->act = 1;
964 if( write (mgr->idx, latchmgr, mgr->page_size) < mgr->page_size )
965 return bt_mgrclose (mgr), NULL;
967 if( !WriteFile (mgr->idx, (char *)latchmgr, mgr->page_size, amt, NULL) )
968 return bt_mgrclose (mgr), NULL;
970 if( *amt < mgr->page_size )
971 return bt_mgrclose (mgr), NULL;
975 // clear out latch manager locks
976 // and rest of pages to round out segment
978 memset(latchmgr, 0, mgr->page_size);
981 while( last <= ((MIN_lvl + 1 + nlatchpage) | mgr->poolmask) ) {
983 pwrite(mgr->idx, latchmgr, mgr->page_size, last << mgr->page_bits);
985 SetFilePointer (mgr->idx, last << mgr->page_bits, NULL, FILE_BEGIN);
986 if( !WriteFile (mgr->idx, (char *)latchmgr, mgr->page_size, amt, NULL) )
987 return bt_mgrclose (mgr), NULL;
988 if( *amt < mgr->page_size )
989 return bt_mgrclose (mgr), NULL;
996 // mlock the root page and the latchmgr page
998 flag = PROT_READ | PROT_WRITE;
999 mgr->latchmgr = mmap (0, 2 * mgr->page_size, flag, MAP_SHARED, mgr->idx, ALLOC_page * mgr->page_size);
1000 if( mgr->latchmgr == MAP_FAILED )
1001 return bt_mgrclose (mgr), NULL;
1002 mlock (mgr->latchmgr, 2 * mgr->page_size);
1004 mgr->latchsets = (BtLatchSet *)mmap (0, mgr->latchmgr->nlatchpage * mgr->page_size, flag, MAP_SHARED, mgr->idx, LATCH_page * mgr->page_size);
1005 if( mgr->latchsets == MAP_FAILED )
1006 return bt_mgrclose (mgr), NULL;
1007 mlock (mgr->latchsets, mgr->latchmgr->nlatchpage * mgr->page_size);
1009 flag = PAGE_READWRITE;
1010 mgr->halloc = CreateFileMapping(mgr->idx, NULL, flag, 0, (BT_latchtable / (mgr->page_size / sizeof(BtLatchSet)) + 1 + LATCH_page) * mgr->page_size, NULL);
1012 return bt_mgrclose (mgr), NULL;
1014 flag = FILE_MAP_WRITE;
1015 mgr->latchmgr = MapViewOfFile(mgr->halloc, flag, 0, 0, (BT_latchtable / (mgr->page_size / sizeof(BtLatchSet)) + 1 + LATCH_page) * mgr->page_size);
1016 if( !mgr->latchmgr )
1017 return GetLastError(), bt_mgrclose (mgr), NULL;
1019 mgr->latchsets = (void *)((char *)mgr->latchmgr + LATCH_page * mgr->page_size);
1025 VirtualFree (latchmgr, 0, MEM_RELEASE);
1030 // open BTree access method
1031 // based on buffer manager
1033 BtDb *bt_open (BtMgr *mgr)
1035 BtDb *bt = malloc (sizeof(*bt));
1037 memset (bt, 0, sizeof(*bt));
1040 bt->mem = malloc (2 *mgr->page_size);
1042 bt->mem = VirtualAlloc(NULL, 2 * mgr->page_size, MEM_COMMIT, PAGE_READWRITE);
1044 bt->frame = (BtPage)bt->mem;
1045 bt->cursor = (BtPage)(bt->mem + 1 * mgr->page_size);
1049 // compare two keys, returning > 0, = 0, or < 0
1050 // as the comparison value
1052 int keycmp (BtKey key1, unsigned char *key2, uint len2)
1054 uint len1 = key1->len;
1057 if( ans = memcmp (key1->key, key2, len1 > len2 ? len2 : len1) )
1070 // find segment in pool
1071 // must be called with hashslot idx locked
1072 // return NULL if not there
1073 // otherwise return node
1075 BtPool *bt_findpool(BtDb *bt, uid page_no, uint idx)
1080 // compute start of hash chain in pool
1082 if( slot = bt->mgr->hash[idx] )
1083 pool = bt->mgr->pool + slot;
1087 page_no &= ~bt->mgr->poolmask;
1089 while( pool->basepage != page_no )
1090 if( pool = pool->hashnext )
1098 // add segment to hash table
1100 void bt_linkhash(BtDb *bt, BtPool *pool, uid page_no, int idx)
1105 pool->hashprev = pool->hashnext = NULL;
1106 pool->basepage = page_no & ~bt->mgr->poolmask;
1107 pool->pin = CLOCK_bit + 1;
1109 if( slot = bt->mgr->hash[idx] ) {
1110 node = bt->mgr->pool + slot;
1111 pool->hashnext = node;
1112 node->hashprev = pool;
1115 bt->mgr->hash[idx] = pool->slot;
1118 // map new buffer pool segment to virtual memory
1120 BTERR bt_mapsegment(BtDb *bt, BtPool *pool, uid page_no)
1122 off64_t off = (page_no & ~bt->mgr->poolmask) << bt->mgr->page_bits;
1123 off64_t limit = off + ((bt->mgr->poolmask+1) << bt->mgr->page_bits);
1127 flag = PROT_READ | ( bt->mgr->mode == BT_ro ? 0 : PROT_WRITE );
1128 pool->map = mmap (0, (uid)(bt->mgr->poolmask+1) << bt->mgr->page_bits, flag, MAP_SHARED, bt->mgr->idx, off);
1129 if( pool->map == MAP_FAILED )
1130 return bt->err = BTERR_map;
1133 flag = ( bt->mgr->mode == BT_ro ? PAGE_READONLY : PAGE_READWRITE );
1134 pool->hmap = CreateFileMapping(bt->mgr->idx, NULL, flag, (DWORD)(limit >> 32), (DWORD)limit, NULL);
1136 return bt->err = BTERR_map;
1138 flag = ( bt->mgr->mode == BT_ro ? FILE_MAP_READ : FILE_MAP_WRITE );
1139 pool->map = MapViewOfFile(pool->hmap, flag, (DWORD)(off >> 32), (DWORD)off, (uid)(bt->mgr->poolmask+1) << bt->mgr->page_bits);
1141 return bt->err = BTERR_map;
1146 // calculate page within pool
1148 BtPage bt_page (BtDb *bt, BtPool *pool, uid page_no)
1150 uint subpage = (uint)(page_no & bt->mgr->poolmask); // page within mapping
1153 page = (BtPage)(pool->map + (subpage << bt->mgr->page_bits));
1154 // madvise (page, bt->mgr->page_size, MADV_WILLNEED);
1160 void bt_unpinpool (BtPool *pool)
1163 __sync_fetch_and_add(&pool->pin, -1);
1165 _InterlockedDecrement16 (&pool->pin);
1169 // find or place requested page in segment-pool
1170 // return pool table entry, incrementing pin
1172 BtPool *bt_pinpool(BtDb *bt, uid page_no)
1174 uint slot, hashidx, idx, victim;
1175 BtPool *pool, *node, *next;
1177 // lock hash table chain
1179 hashidx = (uint)(page_no >> bt->mgr->seg_bits) % bt->mgr->hashsize;
1180 bt_spinwritelock (&bt->mgr->latch[hashidx]);
1182 // look up in hash table
1184 if( pool = bt_findpool(bt, page_no, hashidx) ) {
1186 __sync_fetch_and_or(&pool->pin, CLOCK_bit);
1187 __sync_fetch_and_add(&pool->pin, 1);
1189 _InterlockedOr16 (&pool->pin, CLOCK_bit);
1190 _InterlockedIncrement16 (&pool->pin);
1192 bt_spinreleasewrite (&bt->mgr->latch[hashidx]);
1196 // allocate a new pool node
1197 // and add to hash table
1200 slot = __sync_fetch_and_add(&bt->mgr->poolcnt, 1);
1202 slot = _InterlockedIncrement16 (&bt->mgr->poolcnt) - 1;
1205 if( ++slot < bt->mgr->poolmax ) {
1206 pool = bt->mgr->pool + slot;
1209 if( bt_mapsegment(bt, pool, page_no) )
1212 bt_linkhash(bt, pool, page_no, hashidx);
1213 bt_spinreleasewrite (&bt->mgr->latch[hashidx]);
1217 // pool table is full
1218 // find best pool entry to evict
1221 __sync_fetch_and_add(&bt->mgr->poolcnt, -1);
1223 _InterlockedDecrement16 (&bt->mgr->poolcnt);
1228 victim = __sync_fetch_and_add(&bt->mgr->evicted, 1);
1230 victim = _InterlockedIncrement (&bt->mgr->evicted) - 1;
1232 victim %= bt->mgr->poolmax;
1233 pool = bt->mgr->pool + victim;
1234 idx = (uint)(pool->basepage >> bt->mgr->seg_bits) % bt->mgr->hashsize;
1239 // try to get write lock
1240 // skip entry if not obtained
1242 if( !bt_spinwritetry (&bt->mgr->latch[idx]) )
1245 // skip this entry if
1247 // or clock bit is set
1251 __sync_fetch_and_and(&pool->pin, ~CLOCK_bit);
1253 _InterlockedAnd16 (&pool->pin, ~CLOCK_bit);
1255 bt_spinreleasewrite (&bt->mgr->latch[idx]);
1259 // unlink victim pool node from hash table
1261 if( node = pool->hashprev )
1262 node->hashnext = pool->hashnext;
1263 else if( node = pool->hashnext )
1264 bt->mgr->hash[idx] = node->slot;
1266 bt->mgr->hash[idx] = 0;
1268 if( node = pool->hashnext )
1269 node->hashprev = pool->hashprev;
1271 bt_spinreleasewrite (&bt->mgr->latch[idx]);
1273 // remove old file mapping
1275 munmap (pool->map, (uid)(bt->mgr->poolmask+1) << bt->mgr->page_bits);
1277 // FlushViewOfFile(pool->map, 0);
1278 UnmapViewOfFile(pool->map);
1279 CloseHandle(pool->hmap);
1283 // create new pool mapping
1284 // and link into hash table
1286 if( bt_mapsegment(bt, pool, page_no) )
1289 bt_linkhash(bt, pool, page_no, hashidx);
1290 bt_spinreleasewrite (&bt->mgr->latch[hashidx]);
1295 // place write, read, or parent lock on requested page_no.
1297 void bt_lockpage(BtLock mode, BtLatchSet *set)
1301 ReadLock (set->readwr);
1304 WriteLock (set->readwr);
1307 ReadLock (set->access);
1310 WriteLock (set->access);
1313 WriteLock (set->parent);
1318 // remove write, read, or parent lock on requested page
1320 void bt_unlockpage(BtLock mode, BtLatchSet *set)
1324 ReadRelease (set->readwr);
1327 WriteRelease (set->readwr);
1330 ReadRelease (set->access);
1333 WriteRelease (set->access);
1336 WriteRelease (set->parent);
1341 // allocate a new page and write page into it
1343 uid bt_newpage(BtDb *bt, BtPage page)
1349 // lock allocation page
1351 bt_spinwritelock(bt->mgr->latchmgr->lock);
1353 // use empty chain first
1354 // else allocate empty page
1356 if( new_page = bt_getid(bt->mgr->latchmgr->chain) ) {
1357 if( set->pool = bt_pinpool (bt, new_page) )
1358 set->page = bt_page (bt, set->pool, new_page);
1362 bt_putid(bt->mgr->latchmgr->chain, bt_getid(set->page->right));
1363 bt_unpinpool (set->pool);
1365 new_page = bt_getid(bt->mgr->latchmgr->alloc->right);
1366 bt_putid(bt->mgr->latchmgr->alloc->right, new_page+1);
1368 // if writing first page of pool block, set file length thru last page
1370 if( (new_page & bt->mgr->poolmask) == 0 )
1371 ftruncate (bt->mgr->idx, (new_page + bt->mgr->poolmask + 1) << bt->mgr->page_bits);
1375 // unlock allocation latch
1377 bt_spinreleasewrite(bt->mgr->latchmgr->lock);
1380 // bring new page into pool and copy page.
1381 // this will extend the file into the new pages on WIN32.
1383 if( set->pool = bt_pinpool (bt, new_page) )
1384 set->page = bt_page (bt, set->pool, new_page);
1388 memcpy(set->page, page, bt->mgr->page_size);
1389 bt_unpinpool (set->pool);
1392 // unlock allocation latch
1394 bt_spinreleasewrite(bt->mgr->latchmgr->lock);
1399 // find slot in page for given key at a given level
1401 int bt_findslot (BtPageSet *set, unsigned char *key, uint len)
1403 uint diff, higher = set->page->cnt, low = 1, slot;
1406 // make stopper key an infinite fence value
1408 if( bt_getid (set->page->right) )
1413 // low is the lowest candidate.
1414 // loop ends when they meet
1416 // higher is already
1417 // tested as .ge. the passed key.
1419 while( diff = higher - low ) {
1420 slot = low + ( diff >> 1 );
1421 if( keycmp (keyptr(set->page, slot), key, len) < 0 )
1424 higher = slot, good++;
1427 // return zero if key is on right link page
1429 return good ? higher : 0;
1432 // find and load page at given level for given key
1433 // leave page rd or wr locked as requested
1435 int bt_loadpage (BtDb *bt, BtPageSet *set, unsigned char *key, uint len, uint lvl, BtLock lock)
1437 uid page_no = ROOT_page, prevpage = 0;
1438 uint drill = 0xff, slot;
1439 BtLatchSet *prevlatch;
1440 uint mode, prevmode;
1443 // start at root of btree and drill down
1446 // determine lock mode of drill level
1447 mode = (drill == lvl) ? lock : BtLockRead;
1449 set->latch = bt_pinlatch (bt, page_no);
1450 set->page_no = page_no;
1452 // pin page contents
1454 if( set->pool = bt_pinpool (bt, page_no) )
1455 set->page = bt_page (bt, set->pool, page_no);
1459 // obtain access lock using lock chaining with Access mode
1461 if( page_no > ROOT_page )
1462 bt_lockpage(BtLockAccess, set->latch);
1464 // release & unpin parent page
1467 bt_unlockpage(prevmode, prevlatch);
1468 bt_unpinlatch (prevlatch);
1469 bt_unpinpool (prevpool);
1473 // obtain read lock using lock chaining
1475 bt_lockpage(mode, set->latch);
1477 if( set->page->free )
1478 return bt->err = BTERR_struct, 0;
1480 if( page_no > ROOT_page )
1481 bt_unlockpage(BtLockAccess, set->latch);
1483 // re-read and re-lock root after determining actual level of root
1485 if( set->page->lvl != drill) {
1486 if( set->page_no != ROOT_page )
1487 return bt->err = BTERR_struct, 0;
1489 drill = set->page->lvl;
1491 if( lock != BtLockRead && drill == lvl ) {
1492 bt_unlockpage(mode, set->latch);
1493 bt_unpinlatch (set->latch);
1494 bt_unpinpool (set->pool);
1499 prevpage = set->page_no;
1500 prevlatch = set->latch;
1501 prevpool = set->pool;
1504 // find key on page at this level
1505 // and descend to requested level
1507 if( !set->page->kill )
1508 if( slot = bt_findslot (set, key, len) ) {
1512 while( slotptr(set->page, slot)->dead )
1513 if( slot++ < set->page->cnt )
1518 page_no = bt_getid(valptr(set->page, slot)->value);
1523 // or slide right into next page
1526 page_no = bt_getid(set->page->right);
1530 // return error on end of right chain
1532 bt->err = BTERR_struct;
1533 return 0; // return error
1536 // return page to free list
1537 // page must be delete & write locked
1539 void bt_freepage (BtDb *bt, BtPageSet *set)
1541 // lock allocation page
1543 bt_spinwritelock (bt->mgr->latchmgr->lock);
1546 memcpy(set->page->right, bt->mgr->latchmgr->chain, BtId);
1547 bt_putid(bt->mgr->latchmgr->chain, set->page_no);
1548 set->page->free = 1;
1550 // unlock released page
1552 bt_unlockpage (BtLockDelete, set->latch);
1553 bt_unlockpage (BtLockWrite, set->latch);
1554 bt_unpinlatch (set->latch);
1555 bt_unpinpool (set->pool);
1557 // unlock allocation page
1559 bt_spinreleasewrite (bt->mgr->latchmgr->lock);
1562 // a fence key was deleted from a page
1563 // push new fence value upwards
1565 BTERR bt_fixfence (BtDb *bt, BtPageSet *set, uint lvl)
1567 unsigned char leftkey[256], rightkey[256];
1568 unsigned char value[BtId];
1572 // remove the old fence value
1574 ptr = keyptr(set->page, set->page->cnt);
1575 memcpy (rightkey, ptr, ptr->len + 1);
1577 memset (slotptr(set->page, set->page->cnt--), 0, sizeof(BtSlot));
1578 set->page->dirty = 1;
1580 ptr = keyptr(set->page, set->page->cnt);
1581 memcpy (leftkey, ptr, ptr->len + 1);
1582 page_no = set->page_no;
1584 bt_lockpage (BtLockParent, set->latch);
1585 bt_unlockpage (BtLockWrite, set->latch);
1587 // insert new (now smaller) fence key
1589 bt_putid (value, page_no);
1591 if( bt_insertkey (bt, leftkey+1, *leftkey, lvl+1, value, BtId) )
1594 // now delete old fence key
1596 if( bt_deletekey (bt, rightkey+1, *rightkey, lvl+1) )
1599 bt_unlockpage (BtLockParent, set->latch);
1600 bt_unpinlatch(set->latch);
1601 bt_unpinpool (set->pool);
1605 // root has a single child
1606 // collapse a level from the tree
1608 BTERR bt_collapseroot (BtDb *bt, BtPageSet *root)
1613 // find the child entry and promote as new root contents
1616 for( idx = 0; idx++ < root->page->cnt; )
1617 if( !slotptr(root->page, idx)->dead )
1620 child->page_no = bt_getid (valptr(root->page, idx)->value);
1622 child->latch = bt_pinlatch (bt, child->page_no);
1623 bt_lockpage (BtLockDelete, child->latch);
1624 bt_lockpage (BtLockWrite, child->latch);
1626 if( child->pool = bt_pinpool (bt, child->page_no) )
1627 child->page = bt_page (bt, child->pool, child->page_no);
1631 memcpy (root->page, child->page, bt->mgr->page_size);
1632 bt_freepage (bt, child);
1634 } while( root->page->lvl > 1 && root->page->act == 1 );
1636 bt_unlockpage (BtLockWrite, root->latch);
1637 bt_unpinlatch (root->latch);
1638 bt_unpinpool (root->pool);
1642 // find and delete key on page by marking delete flag bit
1643 // if page becomes empty, delete it from the btree
1645 BTERR bt_deletekey (BtDb *bt, unsigned char *key, uint len, uint lvl)
1647 unsigned char lowerfence[256], higherfence[256];
1648 uint slot, idx, dirty = 0, fence, found;
1649 BtPageSet set[1], right[1];
1650 unsigned char value[BtId];
1653 if( slot = bt_loadpage (bt, set, key, len, lvl, BtLockWrite) )
1654 ptr = keyptr(set->page, slot);
1658 // are we deleting a fence slot?
1660 fence = slot == set->page->cnt;
1662 // if key is found delete it, otherwise ignore request
1664 if( found = !keycmp (ptr, key, len) )
1665 if( found = slotptr(set->page, slot)->dead == 0 ) {
1666 dirty = slotptr(set->page, slot)->dead = 1;
1667 set->page->dirty = 1;
1670 // collapse empty slots
1672 while( idx = set->page->cnt - 1 )
1673 if( slotptr(set->page, idx)->dead ) {
1674 *slotptr(set->page, idx) = *slotptr(set->page, idx + 1);
1675 memset (slotptr(set->page, set->page->cnt--), 0, sizeof(BtSlot));
1680 // did we delete a fence key in an upper level?
1682 if( dirty && lvl && set->page->act && fence )
1683 if( bt_fixfence (bt, set, lvl) )
1686 return bt->found = found, 0;
1688 // is this a collapsed root?
1690 if( lvl > 1 && set->page_no == ROOT_page && set->page->act == 1 )
1691 if( bt_collapseroot (bt, set) )
1694 return bt->found = found, 0;
1696 // return if page is not empty
1698 if( set->page->act ) {
1699 bt_unlockpage(BtLockWrite, set->latch);
1700 bt_unpinlatch (set->latch);
1701 bt_unpinpool (set->pool);
1702 return bt->found = found, 0;
1705 // cache copy of fence key
1706 // to post in parent
1708 ptr = keyptr(set->page, set->page->cnt);
1709 memcpy (lowerfence, ptr, ptr->len + 1);
1711 // obtain lock on right page
1713 right->page_no = bt_getid(set->page->right);
1714 right->latch = bt_pinlatch (bt, right->page_no);
1715 bt_lockpage (BtLockWrite, right->latch);
1717 // pin page contents
1719 if( right->pool = bt_pinpool (bt, right->page_no) )
1720 right->page = bt_page (bt, right->pool, right->page_no);
1724 if( right->page->kill )
1725 return bt->err = BTERR_struct;
1727 // pull contents of right peer into our empty page
1729 memcpy (set->page, right->page, bt->mgr->page_size);
1731 // cache copy of key to update
1733 ptr = keyptr(right->page, right->page->cnt);
1734 memcpy (higherfence, ptr, ptr->len + 1);
1736 // mark right page deleted and point it to left page
1737 // until we can post parent updates
1739 bt_putid (right->page->right, set->page_no);
1740 right->page->kill = 1;
1742 bt_lockpage (BtLockParent, right->latch);
1743 bt_unlockpage (BtLockWrite, right->latch);
1745 bt_lockpage (BtLockParent, set->latch);
1746 bt_unlockpage (BtLockWrite, set->latch);
1748 // redirect higher key directly to our new node contents
1750 bt_putid (value, set->page_no);
1752 if( bt_insertkey (bt, higherfence+1, *higherfence, lvl+1, value, BtId) )
1755 // delete old lower key to our node
1757 if( bt_deletekey (bt, lowerfence+1, *lowerfence, lvl+1) )
1760 // obtain delete and write locks to right node
1762 bt_unlockpage (BtLockParent, right->latch);
1763 bt_lockpage (BtLockDelete, right->latch);
1764 bt_lockpage (BtLockWrite, right->latch);
1765 bt_freepage (bt, right);
1767 bt_unlockpage (BtLockParent, set->latch);
1768 bt_unpinlatch (set->latch);
1769 bt_unpinpool (set->pool);
1774 // find key in leaf level and return number of value bytes
1775 // or (-1) if not found
1777 int bt_findkey (BtDb *bt, unsigned char *key, uint keylen, unsigned char *value, uint valmax)
1785 if( slot = bt_loadpage (bt, set, key, keylen, 0, BtLockRead) )
1786 ptr = keyptr(set->page, slot);
1790 // if key exists, return >= 0 value bytes copied
1791 // otherwise return (-1)
1793 if( !slotptr(set->page, slot)->dead && !keycmp (ptr, key, keylen) ) {
1794 val = valptr (set->page,slot);
1795 if( valmax > val->len )
1797 memcpy (value, val->value, valmax);
1802 bt_unlockpage (BtLockRead, set->latch);
1803 bt_unpinlatch (set->latch);
1804 bt_unpinpool (set->pool);
1808 // check page for space available,
1809 // clean if necessary and return
1810 // 0 - page needs splitting
1811 // >0 new slot value
1813 uint bt_cleanpage(BtDb *bt, BtPage page, uint keylen, uint slot, uint vallen)
1815 uint nxt = bt->mgr->page_size;
1816 uint cnt = 0, idx = 0;
1817 uint max = page->cnt;
1822 if( page->min >= (max+1) * sizeof(BtSlot) + sizeof(*page) + keylen + 1 + vallen + 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);
1853 // copy the value across
1855 val = valptr(bt->frame, cnt);
1856 nxt -= val->len + 1;
1857 ((unsigned char *)page)[nxt] = val->len;
1858 memcpy ((unsigned char *)page + nxt + 1, val->value, val->len);
1862 slotptr(page, ++idx)->off = nxt;
1864 if( !(slotptr(page, idx)->dead = slotptr(bt->frame, cnt)->dead) )
1871 // see if page has enough space now, or does it need splitting?
1873 if( page->min >= (idx+1) * sizeof(BtSlot) + sizeof(*page) + keylen + 1 + vallen + 1 )
1879 // split the root and raise the height of the btree
1881 BTERR bt_splitroot(BtDb *bt, BtPageSet *root, unsigned char *leftkey, uid page_no2)
1883 uint nxt = bt->mgr->page_size;
1884 unsigned char value[BtId];
1887 // Obtain an empty page to use, and copy the current
1888 // root contents into it, e.g. lower keys
1890 if( !(left = bt_newpage(bt, root->page)) )
1893 // preserve the page info at the bottom
1894 // of higher keys and set rest to zero
1896 memset(root->page+1, 0, bt->mgr->page_size - sizeof(*root->page));
1898 // insert lower keys page fence key on newroot page as first key
1901 bt_putid (value, left);
1902 ((unsigned char *)root->page)[nxt] = BtId;
1903 memcpy ((unsigned char *)root->page + nxt + 1, value, BtId);
1905 nxt -= *leftkey + 1;
1906 memcpy ((unsigned char *)root->page + nxt, leftkey, *leftkey + 1);
1907 slotptr(root->page, 1)->off = nxt;
1909 // insert stopper key on newroot page
1910 // and increase the root height
1912 nxt -= 3 + BtId + 1;
1913 ((unsigned char *)root->page)[nxt] = 2;
1914 ((unsigned char *)root->page)[nxt+1] = 0xff;
1915 ((unsigned char *)root->page)[nxt+2] = 0xff;
1917 bt_putid (value, page_no2);
1918 ((unsigned char *)root->page)[nxt+3] = BtId;
1919 memcpy ((unsigned char *)root->page + nxt + 4, value, BtId);
1920 slotptr(root->page, 2)->off = nxt;
1922 bt_putid(root->page->right, 0);
1923 root->page->min = nxt; // reset lowest used offset and key count
1924 root->page->cnt = 2;
1925 root->page->act = 2;
1928 // release and unpin root
1930 bt_unlockpage(BtLockWrite, root->latch);
1931 bt_unpinlatch (root->latch);
1932 bt_unpinpool (root->pool);
1936 // split already locked full node
1939 BTERR bt_splitpage (BtDb *bt, BtPageSet *set)
1941 uint cnt = 0, idx = 0, max, nxt = bt->mgr->page_size;
1942 unsigned char fencekey[256], rightkey[256];
1943 unsigned char value[BtId];
1944 uint lvl = set->page->lvl;
1950 // split higher half of keys to bt->frame
1952 memset (bt->frame, 0, bt->mgr->page_size);
1953 max = set->page->cnt;
1957 while( cnt++ < max ) {
1958 val = valptr(set->page, cnt);
1959 nxt -= val->len + 1;
1960 ((unsigned char *)bt->frame)[nxt] = val->len;
1961 memcpy ((unsigned char *)bt->frame + nxt + 1, val->value, val->len);
1963 key = keyptr(set->page, cnt);
1964 nxt -= key->len + 1;
1965 memcpy ((unsigned char *)bt->frame + nxt, key, key->len + 1);
1967 slotptr(bt->frame, ++idx)->off = nxt;
1969 if( !(slotptr(bt->frame, idx)->dead = slotptr(set->page, cnt)->dead) )
1973 // remember existing fence key for new page to the right
1975 memcpy (rightkey, key, key->len + 1);
1977 bt->frame->bits = bt->mgr->page_bits;
1978 bt->frame->min = nxt;
1979 bt->frame->cnt = idx;
1980 bt->frame->lvl = lvl;
1984 if( set->page_no > ROOT_page )
1985 memcpy (bt->frame->right, set->page->right, BtId);
1987 // get new free page and write higher keys to it.
1989 if( !(right->page_no = bt_newpage(bt, bt->frame)) )
1992 // update lower keys to continue in old page
1994 memcpy (bt->frame, set->page, bt->mgr->page_size);
1995 memset (set->page+1, 0, bt->mgr->page_size - sizeof(*set->page));
1996 nxt = bt->mgr->page_size;
1997 set->page->dirty = 0;
2002 // assemble page of smaller keys
2004 while( cnt++ < max / 2 ) {
2005 val = valptr(bt->frame, cnt);
2006 nxt -= val->len + 1;
2007 ((unsigned char *)set->page)[nxt] = val->len;
2008 memcpy ((unsigned char *)set->page + nxt + 1, val->value, val->len);
2010 key = keyptr(bt->frame, cnt);
2011 nxt -= key->len + 1;
2012 memcpy ((unsigned char *)set->page + nxt, key, key->len + 1);
2013 slotptr(set->page, ++idx)->off = nxt;
2017 // remember fence key for smaller page
2019 memcpy(fencekey, key, key->len + 1);
2021 bt_putid(set->page->right, right->page_no);
2022 set->page->min = nxt;
2023 set->page->cnt = idx;
2025 // if current page is the root page, split it
2027 if( set->page_no == ROOT_page )
2028 return bt_splitroot (bt, set, fencekey, right->page_no);
2030 // insert new fences in their parent pages
2032 right->latch = bt_pinlatch (bt, right->page_no);
2033 bt_lockpage (BtLockParent, right->latch);
2035 bt_lockpage (BtLockParent, set->latch);
2036 bt_unlockpage (BtLockWrite, set->latch);
2038 // insert new fence for reformulated left block of smaller keys
2040 bt_putid (value, set->page_no);
2042 if( bt_insertkey (bt, fencekey+1, *fencekey, lvl+1, value, BtId) )
2045 // switch fence for right block of larger keys to new right page
2047 bt_putid (value, right->page_no);
2049 if( bt_insertkey (bt, rightkey+1, *rightkey, lvl+1, value, BtId) )
2052 bt_unlockpage (BtLockParent, set->latch);
2053 bt_unpinlatch (set->latch);
2054 bt_unpinpool (set->pool);
2056 bt_unlockpage (BtLockParent, right->latch);
2057 bt_unpinlatch (right->latch);
2060 // Insert new key into the btree at given level.
2062 BTERR bt_insertkey (BtDb *bt, unsigned char *key, uint keylen, uint lvl, void *value, uint vallen)
2071 if( slot = bt_loadpage (bt, set, key, keylen, lvl, BtLockWrite) )
2072 ptr = keyptr(set->page, slot);
2076 bt->err = BTERR_ovflw;
2080 // if key already exists, update value and return
2082 if( reuse = !keycmp (ptr, key, keylen) )
2083 if( val = valptr(set->page, slot), val->len >= vallen ) {
2084 if( slotptr(set->page, slot)->dead )
2086 slotptr(set->page, slot)->dead = 0;
2088 memcpy (val->value, value, vallen);
2089 bt_unlockpage(BtLockWrite, set->latch);
2090 bt_unpinlatch (set->latch);
2091 bt_unpinpool (set->pool);
2094 if( !slotptr(set->page, slot)->dead )
2096 slotptr(set->page, slot)->dead = 1;
2097 set->page->dirty = 1;
2100 // check if page has enough space
2102 if( slot = bt_cleanpage (bt, set->page, keylen, slot, vallen) )
2105 if( bt_splitpage (bt, set) )
2109 // calculate next available slot and copy key into page
2111 set->page->min -= vallen + 1; // reset lowest used offset
2112 ((unsigned char *)set->page)[set->page->min] = vallen;
2113 memcpy ((unsigned char *)set->page + set->page->min +1, value, vallen );
2115 set->page->min -= keylen + 1; // reset lowest used offset
2116 ((unsigned char *)set->page)[set->page->min] = keylen;
2117 memcpy ((unsigned char *)set->page + set->page->min +1, key, keylen );
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( !reuse && idx == set->page->cnt )
2126 idx++, set->page->cnt++;
2131 *slotptr(set->page, idx) = *slotptr(set->page, idx -1), idx--;
2133 slotptr(set->page, slot)->off = set->page->min;
2134 slotptr(set->page, slot)->dead = 0;
2136 bt_unlockpage (BtLockWrite, set->latch);
2137 bt_unpinlatch (set->latch);
2138 bt_unpinpool (set->pool);
2142 // cache page of keys into cursor and return starting slot for given key
2144 uint bt_startkey (BtDb *bt, unsigned char *key, uint len)
2149 // cache page for retrieval
2151 if( slot = bt_loadpage (bt, set, key, len, 0, BtLockRead) )
2152 memcpy (bt->cursor, set->page, bt->mgr->page_size);
2156 bt->cursor_page = set->page_no;
2158 bt_unlockpage(BtLockRead, set->latch);
2159 bt_unpinlatch (set->latch);
2160 bt_unpinpool (set->pool);
2164 // return next slot for cursor page
2165 // or slide cursor right into next page
2167 uint bt_nextkey (BtDb *bt, uint slot)
2173 right = bt_getid(bt->cursor->right);
2175 while( slot++ < bt->cursor->cnt )
2176 if( slotptr(bt->cursor,slot)->dead )
2178 else if( right || (slot < bt->cursor->cnt) ) // skip infinite stopper
2186 bt->cursor_page = right;
2188 if( set->pool = bt_pinpool (bt, right) )
2189 set->page = bt_page (bt, set->pool, right);
2193 set->latch = bt_pinlatch (bt, right);
2194 bt_lockpage(BtLockRead, set->latch);
2196 memcpy (bt->cursor, set->page, bt->mgr->page_size);
2198 bt_unlockpage(BtLockRead, set->latch);
2199 bt_unpinlatch (set->latch);
2200 bt_unpinpool (set->pool);
2208 BtKey bt_key(BtDb *bt, uint slot)
2210 return keyptr(bt->cursor, slot);
2213 BtVal bt_val(BtDb *bt, uint slot)
2215 return valptr(bt->cursor,slot);
2221 double getCpuTime(int type)
2224 FILETIME xittime[1];
2225 FILETIME systime[1];
2226 FILETIME usrtime[1];
2227 SYSTEMTIME timeconv[1];
2230 memset (timeconv, 0, sizeof(SYSTEMTIME));
2234 GetSystemTimeAsFileTime (xittime);
2235 FileTimeToSystemTime (xittime, timeconv);
2236 ans = (double)timeconv->wDayOfWeek * 3600 * 24;
2239 GetProcessTimes (GetCurrentProcess(), crtime, xittime, systime, usrtime);
2240 FileTimeToSystemTime (usrtime, timeconv);
2243 GetProcessTimes (GetCurrentProcess(), crtime, xittime, systime, usrtime);
2244 FileTimeToSystemTime (systime, timeconv);
2248 ans += (double)timeconv->wHour * 3600;
2249 ans += (double)timeconv->wMinute * 60;
2250 ans += (double)timeconv->wSecond;
2251 ans += (double)timeconv->wMilliseconds / 1000;
2256 #include <sys/resource.h>
2258 double getCpuTime(int type)
2260 struct rusage used[1];
2261 struct timeval tv[1];
2265 gettimeofday(tv, NULL);
2266 return (double)tv->tv_sec + (double)tv->tv_usec / 1000000;
2269 getrusage(RUSAGE_SELF, used);
2270 return (double)used->ru_utime.tv_sec + (double)used->ru_utime.tv_usec / 1000000;
2273 getrusage(RUSAGE_SELF, used);
2274 return (double)used->ru_stime.tv_sec + (double)used->ru_stime.tv_usec / 1000000;
2281 uint bt_latchaudit (BtDb *bt)
2283 ushort idx, hashidx;
2290 posix_fadvise( bt->mgr->idx, 0, 0, POSIX_FADV_SEQUENTIAL);
2292 if( *(ushort *)(bt->mgr->latchmgr->lock) )
2293 fprintf(stderr, "Alloc page locked\n");
2294 *(ushort *)(bt->mgr->latchmgr->lock) = 0;
2296 for( idx = 1; idx <= bt->mgr->latchmgr->latchdeployed; idx++ ) {
2297 latch = bt->mgr->latchsets + idx;
2298 if( *latch->readwr->rin & MASK )
2299 fprintf(stderr, "latchset %d rwlocked for page %.8x\n", idx, latch->page_no);
2300 memset ((ushort *)latch->readwr, 0, sizeof(RWLock));
2302 if( *latch->access->rin & MASK )
2303 fprintf(stderr, "latchset %d accesslocked for page %.8x\n", idx, latch->page_no);
2304 memset ((ushort *)latch->access, 0, sizeof(RWLock));
2306 if( *latch->parent->rin & MASK )
2307 fprintf(stderr, "latchset %d parentlocked for page %.8x\n", idx, latch->page_no);
2308 memset ((ushort *)latch->parent, 0, sizeof(RWLock));
2311 fprintf(stderr, "latchset %d pinned for page %.8x\n", idx, latch->page_no);
2316 for( hashidx = 0; hashidx < bt->mgr->latchmgr->latchhash; hashidx++ ) {
2317 if( *(ushort *)(bt->mgr->latchmgr->table[hashidx].latch) )
2318 fprintf(stderr, "hash entry %d locked\n", hashidx);
2320 *(ushort *)(bt->mgr->latchmgr->table[hashidx].latch) = 0;
2322 if( idx = bt->mgr->latchmgr->table[hashidx].slot ) do {
2323 latch = bt->mgr->latchsets + idx;
2324 if( *(ushort *)latch->busy )
2325 fprintf(stderr, "latchset %d busylocked for page %.8x\n", idx, latch->page_no);
2326 *(ushort *)latch->busy = 0;
2328 fprintf(stderr, "latchset %d pinned for page %.8x\n", idx, latch->page_no);
2329 } while( idx = latch->next );
2332 next = bt->mgr->latchmgr->nlatchpage + LATCH_page;
2333 page_no = LEAF_page;
2335 while( page_no < bt_getid(bt->mgr->latchmgr->alloc->right) ) {
2336 off64_t off = page_no << bt->mgr->page_bits;
2338 pread (bt->mgr->idx, bt->frame, bt->mgr->page_size, off);
2342 SetFilePointer (bt->mgr->idx, (long)off, (long*)(&off)+1, FILE_BEGIN);
2344 if( !ReadFile(bt->mgr->idx, bt->frame, bt->mgr->page_size, amt, NULL))
2345 fprintf(stderr, "page %.8x unable to read\n", page_no);
2347 if( *amt < bt->mgr->page_size )
2348 fprintf(stderr, "page %.8x unable to read\n", page_no);
2350 if( !bt->frame->free ) {
2351 for( idx = 0; idx++ < bt->frame->cnt - 1; ) {
2352 ptr = keyptr(bt->frame, idx+1);
2353 if( keycmp (keyptr(bt->frame, idx), ptr->key, ptr->len) >= 0 )
2354 fprintf(stderr, "page %.8x idx %.2x out of order\n", page_no, idx);
2356 if( !bt->frame->lvl )
2357 cnt += bt->frame->act;
2359 if( page_no > LEAF_page )
2374 // standalone program to index file of keys
2375 // then list them onto std-out
2378 void *index_file (void *arg)
2380 uint __stdcall index_file (void *arg)
2383 int line = 0, found = 0, cnt = 0;
2384 uid next, page_no = LEAF_page; // start on first page of leaves
2385 unsigned char key[256];
2386 ThreadArg *args = arg;
2387 int ch, len = 0, slot;
2393 bt = bt_open (args->mgr);
2395 switch(args->type[0] | 0x20)
2398 fprintf(stderr, "started latch mgr audit\n");
2399 cnt = bt_latchaudit (bt);
2400 fprintf(stderr, "finished latch mgr audit, found %d keys\n", cnt);
2404 fprintf(stderr, "started pennysort for %s\n", args->infile);
2405 if( in = fopen (args->infile, "rb") )
2406 while( ch = getc(in), ch != EOF )
2411 if( bt_insertkey (bt, key, 10, 0, key + 10, len - 10) )
2412 fprintf(stderr, "Error %d Line: %d\n", bt->err, line), exit(0);
2415 else if( len < 255 )
2417 fprintf(stderr, "finished %s for %d keys\n", args->infile, line);
2421 fprintf(stderr, "started indexing for %s\n", args->infile);
2422 if( in = fopen (args->infile, "rb") )
2423 while( ch = getc(in), ch != EOF )
2428 if( args->num == 1 )
2429 sprintf((char *)key+len, "%.9d", 1000000000 - line), len += 9;
2431 else if( args->num )
2432 sprintf((char *)key+len, "%.9d", line + args->idx * args->num), len += 9;
2434 if( bt_insertkey (bt, key, len, 0, NULL, 0) )
2435 fprintf(stderr, "Error %d Line: %d\n", bt->err, line), exit(0);
2438 else if( len < 255 )
2440 fprintf(stderr, "finished %s for %d keys\n", args->infile, line);
2444 fprintf(stderr, "started deleting keys for %s\n", args->infile);
2445 if( in = fopen (args->infile, "rb") )
2446 while( ch = getc(in), ch != EOF )
2450 if( args->num == 1 )
2451 sprintf((char *)key+len, "%.9d", 1000000000 - line), len += 9;
2453 else if( args->num )
2454 sprintf((char *)key+len, "%.9d", line + args->idx * args->num), len += 9;
2456 if( (args->type[1] | 0x20) == 'p' )
2458 if( bt_deletekey (bt, key, len, 0) )
2459 fprintf(stderr, "Error %d Line: %d\n", bt->err, line), exit(0);
2462 else if( len < 255 )
2464 fprintf(stderr, "finished %s for keys, %d \n", args->infile, line);
2468 fprintf(stderr, "started finding keys for %s\n", args->infile);
2469 if( in = fopen (args->infile, "rb") )
2470 while( ch = getc(in), ch != EOF )
2474 if( args->num == 1 )
2475 sprintf((char *)key+len, "%.9d", 1000000000 - line), len += 9;
2477 else if( args->num )
2478 sprintf((char *)key+len, "%.9d", line + args->idx * args->num), len += 9;
2480 if( (args->type[1] | 0x20) == 'p' )
2482 if( bt_findkey (bt, key, len, NULL, 0) == 0 )
2485 fprintf(stderr, "Error %d Syserr %d Line: %d\n", bt->err, errno, line), exit(0);
2488 else if( len < 255 )
2490 fprintf(stderr, "finished %s for %d keys, found %d\n", args->infile, line, found);
2494 fprintf(stderr, "started scanning\n");
2496 if( set->pool = bt_pinpool (bt, page_no) )
2497 set->page = bt_page (bt, set->pool, page_no);
2500 set->latch = bt_pinlatch (bt, page_no);
2501 bt_lockpage (BtLockRead, set->latch);
2502 next = bt_getid (set->page->right);
2503 cnt += set->page->act;
2505 for( slot = 0; slot++ < set->page->cnt; )
2506 if( next || slot < set->page->cnt )
2507 if( !slotptr(set->page, slot)->dead ) {
2508 ptr = keyptr(set->page, slot);
2509 fwrite (ptr->key, ptr->len, 1, stdout);
2510 fputc ('\n', stdout);
2513 bt_unlockpage (BtLockRead, set->latch);
2514 bt_unpinlatch (set->latch);
2515 bt_unpinpool (set->pool);
2516 } while( page_no = next );
2518 cnt--; // remove stopper key
2519 fprintf(stderr, " Total keys read %d\n", cnt);
2524 posix_fadvise( bt->mgr->idx, 0, 0, POSIX_FADV_SEQUENTIAL);
2526 fprintf(stderr, "started counting\n");
2527 next = bt->mgr->latchmgr->nlatchpage + LATCH_page;
2528 page_no = LEAF_page;
2530 while( page_no < bt_getid(bt->mgr->latchmgr->alloc->right) ) {
2531 uid off = page_no << bt->mgr->page_bits;
2533 pread (bt->mgr->idx, bt->frame, bt->mgr->page_size, off);
2537 SetFilePointer (bt->mgr->idx, (long)off, (long*)(&off)+1, FILE_BEGIN);
2539 if( !ReadFile(bt->mgr->idx, bt->frame, bt->mgr->page_size, amt, NULL))
2540 return bt->err = BTERR_map;
2542 if( *amt < bt->mgr->page_size )
2543 return bt->err = BTERR_map;
2545 if( !bt->frame->free && !bt->frame->lvl )
2546 cnt += bt->frame->act;
2547 if( page_no > LEAF_page )
2552 cnt--; // remove stopper key
2553 fprintf(stderr, " Total keys read %d\n", cnt);
2565 typedef struct timeval timer;
2567 int main (int argc, char **argv)
2569 int idx, cnt, len, slot, err;
2570 int segsize, bits = 16;
2587 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]);
2588 fprintf (stderr, " where page_bits is the page size in bits\n");
2589 fprintf (stderr, " mapped_segments is the number of mmap segments in buffer pool\n");
2590 fprintf (stderr, " seg_bits is the size of individual segments in buffer pool in pages in bits\n");
2591 fprintf (stderr, " line_numbers = 1 to append line numbers to keys\n");
2592 fprintf (stderr, " src_file1 thru src_filen are files of keys separated by newline\n");
2596 start = getCpuTime(0);
2599 bits = atoi(argv[3]);
2602 poolsize = atoi(argv[4]);
2605 fprintf (stderr, "Warning: no mapped_pool\n");
2607 if( poolsize > 65535 )
2608 fprintf (stderr, "Warning: mapped_pool > 65535 segments\n");
2611 segsize = atoi(argv[5]);
2613 segsize = 4; // 16 pages per mmap segment
2616 num = atoi(argv[6]);
2620 threads = malloc (cnt * sizeof(pthread_t));
2622 threads = GlobalAlloc (GMEM_FIXED|GMEM_ZEROINIT, cnt * sizeof(HANDLE));
2624 args = malloc (cnt * sizeof(ThreadArg));
2626 mgr = bt_mgr ((argv[1]), BT_rw, bits, poolsize, segsize, poolsize / 8);
2629 fprintf(stderr, "Index Open Error %s\n", argv[1]);
2635 for( idx = 0; idx < cnt; idx++ ) {
2636 args[idx].infile = argv[idx + 7];
2637 args[idx].type = argv[2];
2638 args[idx].mgr = mgr;
2639 args[idx].num = num;
2640 args[idx].idx = idx;
2642 if( err = pthread_create (threads + idx, NULL, index_file, args + idx) )
2643 fprintf(stderr, "Error creating thread %d\n", err);
2645 threads[idx] = (HANDLE)_beginthreadex(NULL, 65536, index_file, args + idx, 0, NULL);
2649 // wait for termination
2652 for( idx = 0; idx < cnt; idx++ )
2653 pthread_join (threads[idx], NULL);
2655 WaitForMultipleObjects (cnt, threads, TRUE, INFINITE);
2657 for( idx = 0; idx < cnt; idx++ )
2658 CloseHandle(threads[idx]);
2661 elapsed = getCpuTime(0) - start;
2662 fprintf(stderr, " real %dm%.3fs\n", (int)(elapsed/60), elapsed - (int)(elapsed/60)*60);
2663 elapsed = getCpuTime(1);
2664 fprintf(stderr, " user %dm%.3fs\n", (int)(elapsed/60), elapsed - (int)(elapsed/60)*60);
2665 elapsed = getCpuTime(2);
2666 fprintf(stderr, " sys %dm%.3fs\n", (int)(elapsed/60), elapsed - (int)(elapsed/60)*60);