35 #if (DEBUG +0) && defined(__GNUC__)
36 # define DPRINTF(fmt, ...) \
37 fprintf(stderr, "%s:%d: " fmt "\n", __func__, __LINE__, ##__VA_ARGS__)
39 # define DPRINTF(...) ((void) 0)
44 #define MDB_MAGIC 0xBEEFC0DE
46 #define MAXKEYSIZE 255
48 #define P_INVALID (~0UL)
50 #define F_ISSET(w, f) (((w) & (f)) == (f))
52 typedef uint16_t indx_t;
54 #define DEFAULT_READERS 126
55 #define DEFAULT_MAPSIZE 1048576
57 /* Lock descriptor stuff */
62 typedef struct MDB_rxbody {
67 #define CACHELINE 64 /* most CPUs. Itanium uses 128 */
70 typedef struct MDB_reader {
72 /* cache line alignment */
73 char pad[CACHELINE-sizeof(MDB_rxbody)];
78 uint32_t mt_version; \
79 pthread_mutex_t mt_mutex; \
81 uint32_t mt_numreaders
82 typedef struct MDB_txbody {
86 typedef struct MDB_txninfo {
88 char pad[CACHELINE-sizeof(MDB_txbody)];
89 pthread_mutex_t mt_wmutex;
90 char pad2[CACHELINE-sizeof(pthread_mutex_t)];
91 MDB_reader mt_readers[1];
94 /* Common header for all page types. Overflow pages
95 * occupy a number of contiguous pages with no
96 * headers on any page after the first.
98 typedef struct MDB_page { /* represents a page of storage */
99 pgno_t mp_pgno; /* page number */
100 #define P_BRANCH 0x01 /* branch page */
101 #define P_LEAF 0x02 /* leaf page */
102 #define P_OVERFLOW 0x04 /* overflow page */
103 #define P_META 0x08 /* meta page */
104 #define P_DIRTY 0x10 /* dirty page */
106 #define mp_lower mp_pb.pb.pb_lower
107 #define mp_upper mp_pb.pb.pb_upper
108 #define mp_pages mp_pb.pb_pages
111 indx_t pb_lower; /* lower bound of free space */
112 indx_t pb_upper; /* upper bound of free space */
114 uint32_t pb_pages; /* number of overflow pages */
116 indx_t mp_ptrs[1]; /* dynamic size */
119 #define PAGEHDRSZ ((unsigned) offsetof(MDB_page, mp_ptrs))
121 #define NUMKEYS(p) (((p)->mp_lower - PAGEHDRSZ) >> 1)
122 #define SIZELEFT(p) (indx_t)((p)->mp_upper - (p)->mp_lower)
123 #define PAGEFILL(env, p) (1000L * ((env)->me_meta.mm_psize - PAGEHDRSZ - SIZELEFT(p)) / \
124 ((env)->me_meta.mm_psize - PAGEHDRSZ))
125 #define IS_LEAF(p) F_ISSET((p)->mp_flags, P_LEAF)
126 #define IS_BRANCH(p) F_ISSET((p)->mp_flags, P_BRANCH)
127 #define IS_OVERFLOW(p) F_ISSET((p)->mp_flags, P_OVERFLOW)
129 #define OVPAGES(size, psize) (PAGEHDRSZ + size + psize - 1) / psize;
131 typedef struct MDB_meta { /* meta (footer) page content */
134 void *mm_address; /* address for fixed mapping */
135 size_t mm_mapsize; /* size of mmap region */
139 ulong mm_branch_pages;
141 ulong mm_overflow_pages;
144 pgno_t mm_last_pg; /* last used page in file */
145 ulong mm_txnid; /* txnid that committed this page */
148 typedef struct MDB_dhead { /* a dirty page */
149 SIMPLEQ_ENTRY(MDB_dpage) md_next; /* queue of dirty pages */
151 unsigned md_pi; /* parent index */
155 typedef struct MDB_dpage {
160 SIMPLEQ_HEAD(dirty_queue, MDB_dpage);
162 typedef struct MDB_oldpages {
163 struct MDB_oldpages *mo_next;
165 pgno_t mo_pages[1]; /* dynamic */
168 typedef struct MDB_pageparent {
174 static MDB_dpage *mdb_alloc_page(MDB_txn *txn, MDB_page *parent, unsigned int parent_idx, int num);
175 static int mdb_touch(MDB_txn *txn, MDB_pageparent *mp);
177 typedef struct MDB_ppage { /* ordered list of pages */
178 SLIST_ENTRY(MDB_ppage) mp_entry;
180 unsigned int mp_ki; /* cursor index on page */
182 SLIST_HEAD(page_stack, MDB_ppage);
184 #define CURSOR_EMPTY(c) SLIST_EMPTY(&(c)->mc_stack)
185 #define CURSOR_TOP(c) SLIST_FIRST(&(c)->mc_stack)
186 #define CURSOR_POP(c) SLIST_REMOVE_HEAD(&(c)->mc_stack, mp_entry)
187 #define CURSOR_PUSH(c,p) SLIST_INSERT_HEAD(&(c)->mc_stack, p, mp_entry)
192 struct page_stack mc_stack; /* stack of parent pages */
193 short mc_initialized; /* 1 if initialized */
194 short mc_eof; /* 1 if end is reached */
197 #define METAHASHLEN offsetof(MDB_meta, mm_hash)
198 #define METADATA(p) ((void *)((char *)p + PAGEHDRSZ))
200 typedef struct MDB_node {
201 #define mn_pgno mn_p.np_pgno
202 #define mn_dsize mn_p.np_dsize
204 pgno_t np_pgno; /* child page number */
205 uint32_t np_dsize; /* leaf data size */
207 unsigned int mn_flags:4;
208 unsigned int mn_ksize:12; /* key size */
209 #define F_BIGDATA 0x01 /* data put on overflow page */
214 pgno_t mt_root; /* current / new root page */
215 pgno_t mt_next_pgno; /* next unallocated page */
219 pgno_t *mt_free_pgs; /* this is an IDL */
221 struct dirty_queue *dirty_queue; /* modified pages */
224 #define MDB_TXN_RDONLY 0x01 /* read-only transaction */
225 #define MDB_TXN_ERROR 0x02 /* an error has occurred */
226 #define MDB_TXN_METOGGLE 0x04 /* used meta page 1 */
227 unsigned int mt_flags;
233 MDB_cmp_func *md_cmp; /* user compare function */
234 MDB_rel_func *md_rel; /* user relocate function */
235 MDB_db *md_parent; /* parent tree */
237 ulong md_branch_pages;
239 ulong md_overflow_pages;
250 unsigned int me_maxreaders;
253 MDB_txninfo *me_txns;
254 MDB_db me_db; /* first DB */
256 MDB_txn *me_txn; /* current write transaction */
258 off_t me_size; /* current file size */
259 pthread_key_t me_txkey; /* thread-key for readers */
260 MDB_oldpages *me_pghead;
261 MDB_oldpages *me_pgtail;
264 #define NODESIZE offsetof(MDB_node, mn_data)
266 #define INDXSIZE(k) (NODESIZE + ((k) == NULL ? 0 : (k)->mv_size))
267 #define LEAFSIZE(k, d) (NODESIZE + (k)->mv_size + (d)->mv_size)
268 #define NODEPTR(p, i) ((MDB_node *)((char *)(p) + (p)->mp_ptrs[i]))
269 #define NODEKEY(node) (void *)((node)->mn_data)
270 #define NODEDATA(node) (void *)((char *)(node)->mn_data + (node)->mn_ksize)
271 #define NODEPGNO(node) ((node)->mn_pgno)
272 #define NODEDSZ(node) ((node)->mn_dsize)
274 #define MDB_COMMIT_PAGES 64 /* max number of pages to write in one commit */
275 #define MDB_MAXCACHE_DEF 1024 /* max number of pages to keep in cache */
277 static int mdb_search_page_root(MDB_db *db,
279 MDB_cursor *cursor, int modify,
280 MDB_pageparent *mpp);
281 static int mdb_search_page(MDB_db *db,
282 MDB_txn *txn, MDB_val *key,
283 MDB_cursor *cursor, int modify,
284 MDB_pageparent *mpp);
286 static int mdbenv_read_header(MDB_env *env);
287 static int mdb_check_meta_page(MDB_page *p);
288 static int mdbenv_read_meta(MDB_env *env, int *which);
289 static int mdbenv_write_meta(MDB_txn *txn);
290 static MDB_page *mdbenv_get_page(MDB_env *env, pgno_t pgno);
292 static MDB_node *mdb_search_node(MDB_db *db, MDB_page *mp,
293 MDB_val *key, int *exactp, unsigned int *kip);
294 static int mdb_add_node(MDB_db *bt, MDB_page *mp,
295 indx_t indx, MDB_val *key, MDB_val *data,
296 pgno_t pgno, uint8_t flags);
297 static void mdb_del_node(MDB_db *bt, MDB_page *mp,
299 static int mdb_read_data(MDB_db *bt, MDB_page *mp,
300 MDB_node *leaf, MDB_val *data);
302 static int mdb_rebalance(MDB_db *bt, MDB_pageparent *mp);
303 static int mdb_update_key(MDB_db *bt, MDB_page *mp,
304 indx_t indx, MDB_val *key);
305 static int mdb_move_node(MDB_db *bt,
306 MDB_pageparent *src, indx_t srcindx,
307 MDB_pageparent *dst, indx_t dstindx);
308 static int mdb_merge(MDB_db *bt, MDB_pageparent *src,
309 MDB_pageparent *dst);
310 static int mdb_split(MDB_db *bt, MDB_page **mpp,
311 unsigned int *newindxp, MDB_val *newkey,
312 MDB_val *newdata, pgno_t newpgno);
313 static MDB_dpage *mdb_new_page(MDB_db *db, uint32_t flags, int num);
315 static void cursor_pop_page(MDB_cursor *cursor);
316 static MDB_ppage *cursor_push_page(MDB_cursor *cursor,
319 static int mdb_set_key(MDB_db *bt, MDB_page *mp,
320 MDB_node *node, MDB_val *key);
321 static int mdb_sibling(MDB_cursor *cursor, int move_right);
322 static int mdb_cursor_next(MDB_cursor *cursor,
323 MDB_val *key, MDB_val *data);
324 static int mdb_cursor_set(MDB_cursor *cursor,
325 MDB_val *key, MDB_val *data, int *exactp);
326 static int mdb_cursor_first(MDB_cursor *cursor,
327 MDB_val *key, MDB_val *data);
329 static size_t mdb_leaf_size(MDB_db *bt, MDB_val *key,
331 static size_t mdb_branch_size(MDB_db *bt, MDB_val *key);
333 static int memncmp(const void *s1, size_t n1,
334 const void *s2, size_t n2);
335 static int memnrcmp(const void *s1, size_t n1,
336 const void *s2, size_t n2);
339 memncmp(const void *s1, size_t n1, const void *s2, size_t n2)
341 int diff, len_diff = -1;
344 len_diff = (n1 > n2);
347 diff = memcmp(s1, s2, n1);
348 return diff ? diff : len_diff;
352 memnrcmp(const void *s1, size_t n1, const void *s2, size_t n2)
354 const unsigned char *p1, *p2, *p1_lim;
361 p1 = (const unsigned char *)s1 + n1 - 1;
362 p2 = (const unsigned char *)s2 + n2 - 1;
364 for (p1_lim = (n1 <= n2 ? s1 : s2); *p1 == *p2; p1--, p2--) {
366 return (p1 != s1) ? (p1 != p2) : (p2 != s2) ? -1 : 0;
372 mdb_cmp(MDB_db *db, const MDB_val *a, const MDB_val *b)
374 return db->md_cmp(a, b);
378 _mdb_cmp(MDB_db *db, const MDB_val *key1, const MDB_val *key2)
380 if (F_ISSET(db->md_flags, MDB_REVERSEKEY))
381 return memnrcmp(key1->mv_data, key1->mv_size, key2->mv_data, key2->mv_size);
383 return memncmp((char *)key1->mv_data, key1->mv_size, key2->mv_data, key2->mv_size);
386 /* Allocate new page(s) for writing */
388 mdb_alloc_page(MDB_txn *txn, MDB_page *parent, unsigned int parent_idx, int num)
391 pgno_t pgno = P_INVALID;
393 if (txn->mt_env->me_pghead) {
394 ulong oldest = txn->mt_txnid - 2;
396 for (i=0; i<txn->mt_env->me_txns->mt_numreaders; i++) {
397 if (txn->mt_env->me_txns->mt_readers[i].mr_txnid < oldest)
398 oldest = txn->mt_env->me_txns->mt_readers[i].mr_txnid;
400 if (oldest > txn->mt_env->me_pghead->mo_txnid) {
401 MDB_oldpages *mop = txn->mt_env->me_pghead;
402 txn->mt_oldest = oldest;
404 /* FIXME: For now, always use fresh pages. We
405 * really ought to search the free list for a
410 /* peel pages off tail, so we only have to truncate the list */
411 pgno = MDB_IDL_LAST(mop->mo_pages);
412 if (MDB_IDL_IS_RANGE(mop->mo_pages)) {
414 if (mop->mo_pages[2] > mop->mo_pages[1])
415 mop->mo_pages[0] = 0;
419 if (MDB_IDL_IS_ZERO(mop->mo_pages)) {
420 txn->mt_env->me_pghead = mop->mo_next;
421 if (!txn->mt_env->me_pghead)
422 txn->mt_env->me_pgtail = NULL;
429 if ((dp = malloc(txn->mt_env->me_meta.mm_psize * num + sizeof(MDB_dhead))) == NULL)
432 dp->h.md_parent = parent;
433 dp->h.md_pi = parent_idx;
434 SIMPLEQ_INSERT_TAIL(txn->mt_u.dirty_queue, dp, h.md_next);
435 if (pgno == P_INVALID) {
436 dp->p.mp_pgno = txn->mt_next_pgno;
437 txn->mt_next_pgno += num;
439 dp->p.mp_pgno = pgno;
445 /* Touch a page: make it dirty and re-insert into tree with updated pgno.
448 mdb_touch(MDB_txn *txn, MDB_pageparent *pp)
450 MDB_page *mp = pp->mp_page;
455 if (!F_ISSET(mp->mp_flags, P_DIRTY)) {
457 if ((dp = mdb_alloc_page(txn, pp->mp_parent, pp->mp_pi, 1)) == NULL)
459 DPRINTF("touched page %lu -> %lu", mp->mp_pgno, dp->p.mp_pgno);
460 mdb_idl_insert(txn->mt_free_pgs, mp->mp_pgno);
461 pgno = dp->p.mp_pgno;
462 bcopy(mp, &dp->p, txn->mt_env->me_meta.mm_psize);
465 mp->mp_flags |= P_DIRTY;
467 /* Update the page number to new touched page. */
468 if (pp->mp_parent != NULL)
469 NODEPGNO(NODEPTR(pp->mp_parent, pp->mp_pi)) = mp->mp_pgno;
476 mdbenv_sync(MDB_env *env)
479 if (!F_ISSET(env->me_flags, MDB_NOSYNC)) {
480 if (fsync(env->me_fd))
487 mdb_txn_begin(MDB_env *env, int rdonly, MDB_txn **ret)
492 if ((txn = calloc(1, sizeof(*txn))) == NULL) {
493 DPRINTF("calloc: %s", strerror(errno));
498 txn->mt_flags |= MDB_TXN_RDONLY;
500 txn->mt_u.dirty_queue = calloc(1, sizeof(*txn->mt_u.dirty_queue));
501 if (txn->mt_u.dirty_queue == NULL) {
505 SIMPLEQ_INIT(txn->mt_u.dirty_queue);
507 pthread_mutex_lock(&env->me_txns->mt_wmutex);
508 env->me_txns->mt_txnid++;
509 txn->mt_free_pgs = malloc(MDB_IDL_UM_SIZEOF);
510 if (txn->mt_free_pgs == NULL) {
511 free(txn->mt_u.dirty_queue);
515 txn->mt_free_pgs[0] = 0;
517 txn->mt_txnid = env->me_txns->mt_txnid;
519 MDB_reader *r = pthread_getspecific(env->me_txkey);
522 pthread_mutex_lock(&env->me_txns->mt_mutex);
523 for (i=0; i<env->me_maxreaders; i++) {
524 if (env->me_txns->mt_readers[i].mr_pid == 0) {
525 env->me_txns->mt_readers[i].mr_pid = getpid();
526 env->me_txns->mt_readers[i].mr_tid = pthread_self();
527 pthread_setspecific(env->me_txkey, &env->me_txns->mt_readers[i]);
528 if (i >= env->me_txns->mt_numreaders)
529 env->me_txns->mt_numreaders = i+1;
533 pthread_mutex_unlock(&env->me_txns->mt_mutex);
534 if (i == env->me_maxreaders) {
538 r->mr_txnid = txn->mt_txnid;
539 txn->mt_u.reader = r;
546 if ((rc = mdbenv_read_meta(env, &toggle)) != MDB_SUCCESS) {
552 txn->mt_flags |= MDB_TXN_METOGGLE;
554 txn->mt_next_pgno = env->me_meta.mm_last_pg+1;
555 txn->mt_root = env->me_meta.mm_root;
556 DPRINTF("begin transaction %lu on mdbenv %p, root page %lu",
557 txn->mt_txnid, (void *) env, txn->mt_root);
564 mdb_txn_abort(MDB_txn *txn)
573 DPRINTF("abort transaction %lu on mdbenv %p, root page %lu",
574 txn->mt_txnid, (void *) env, txn->mt_root);
576 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
577 txn->mt_u.reader->mr_txnid = 0;
579 /* Discard all dirty pages. Return any re-used pages
582 MDB_IDL_ZERO(txn->mt_free_pgs);
583 while (!SIMPLEQ_EMPTY(txn->mt_u.dirty_queue)) {
584 dp = SIMPLEQ_FIRST(txn->mt_u.dirty_queue);
585 SIMPLEQ_REMOVE_HEAD(txn->mt_u.dirty_queue, h.md_next);
586 if (dp->p.mp_pgno <= env->me_meta.mm_last_pg)
587 mdb_idl_insert(txn->mt_free_pgs, dp->p.mp_pgno);
590 /* put back to head of free list */
591 if (!MDB_IDL_IS_ZERO(txn->mt_free_pgs)) {
594 mop = malloc(sizeof(MDB_oldpages) + MDB_IDL_SIZEOF(txn->mt_free_pgs) - sizeof(pgno_t));
595 mop->mo_next = env->me_pghead;
596 mop->mo_txnid = txn->mt_oldest - 1;
597 if (!env->me_pghead) {
598 env->me_pgtail = mop;
600 env->me_pghead = mop;
601 bcopy(txn->mt_free_pgs, mop->mo_pages, MDB_IDL_SIZEOF(txn->mt_free_pgs));
604 free(txn->mt_free_pgs);
605 free(txn->mt_u.dirty_queue);
607 env->me_txns->mt_txnid--;
608 pthread_mutex_unlock(&env->me_txns->mt_wmutex);
615 mdb_txn_commit(MDB_txn *txn)
623 struct iovec iov[MDB_COMMIT_PAGES];
626 assert(txn->mt_env != NULL);
630 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
631 DPRINTF("attempt to commit read-only transaction");
636 if (txn != env->me_txn) {
637 DPRINTF("attempt to commit unknown transaction");
642 if (F_ISSET(txn->mt_flags, MDB_TXN_ERROR)) {
643 DPRINTF("error flag is set, can't commit");
648 if (SIMPLEQ_EMPTY(txn->mt_u.dirty_queue))
651 DPRINTF("committing transaction %lu on mdbenv %p, root page %lu",
652 txn->mt_txnid, (void *) env, txn->mt_root);
654 /* Commit up to MDB_COMMIT_PAGES dirty pages to disk until done.
661 SIMPLEQ_FOREACH(dp, txn->mt_u.dirty_queue, h.md_next) {
662 if (dp->p.mp_pgno != next) {
664 DPRINTF("committing %u dirty pages", n);
665 rc = writev(env->me_fd, iov, n);
669 DPRINTF("short write, filesystem full?");
671 DPRINTF("writev: %s", strerror(errno));
678 lseek(env->me_fd, dp->p.mp_pgno * env->me_meta.mm_psize, SEEK_SET);
679 next = dp->p.mp_pgno;
681 DPRINTF("committing page %lu", dp->p.mp_pgno);
682 iov[n].iov_len = env->me_meta.mm_psize * dp->h.md_num;
683 iov[n].iov_base = &dp->p;
684 size += iov[n].iov_len;
685 next = dp->p.mp_pgno + dp->h.md_num;
686 /* clear dirty flag */
687 dp->p.mp_flags &= ~P_DIRTY;
688 if (++n >= MDB_COMMIT_PAGES) {
697 DPRINTF("committing %u dirty pages", n);
698 rc = writev(env->me_fd, iov, n);
702 DPRINTF("short write, filesystem full?");
704 DPRINTF("writev: %s", strerror(errno));
711 /* Drop the dirty pages.
713 while (!SIMPLEQ_EMPTY(txn->mt_u.dirty_queue)) {
714 dp = SIMPLEQ_FIRST(txn->mt_u.dirty_queue);
715 SIMPLEQ_REMOVE_HEAD(txn->mt_u.dirty_queue, h.md_next);
719 if ((n = mdbenv_sync(env)) != 0 ||
720 (n = mdbenv_write_meta(txn)) != MDB_SUCCESS ||
721 (n = mdbenv_sync(env)) != 0) {
727 /* add to tail of free list */
728 if (!MDB_IDL_IS_ZERO(txn->mt_free_pgs)) {
731 mop = malloc(sizeof(MDB_oldpages) + MDB_IDL_SIZEOF(txn->mt_free_pgs) - sizeof(pgno_t));
733 if (env->me_pghead) {
734 env->me_pgtail->mo_next = mop;
736 env->me_pghead = mop;
738 env->me_pgtail = mop;
739 bcopy(txn->mt_free_pgs, mop->mo_pages, MDB_IDL_SIZEOF(txn->mt_free_pgs));
740 mop->mo_txnid = txn->mt_txnid;
743 pthread_mutex_unlock(&env->me_txns->mt_wmutex);
744 free(txn->mt_free_pgs);
745 free(txn->mt_u.dirty_queue);
756 mdbenv_read_header(MDB_env *env)
765 /* We don't know the page size yet, so use a minimum value.
768 if ((rc = pread(env->me_fd, page, PAGESIZE, 0)) == 0) {
770 } else if (rc != PAGESIZE) {
773 DPRINTF("read: %s", strerror(errno));
777 p = (MDB_page *)page;
779 if (!F_ISSET(p->mp_flags, P_META)) {
780 DPRINTF("page %lu not a meta page", p->mp_pgno);
785 if (m->mm_magic != MDB_MAGIC) {
786 DPRINTF("meta has invalid magic");
790 if (m->mm_version != MDB_VERSION) {
791 DPRINTF("database is version %u, expected version %u",
792 m->mm_version, MDB_VERSION);
796 bcopy(m, &env->me_meta, sizeof(*m));
801 mdbenv_init_meta(MDB_env *env)
808 DPRINTF("writing new meta page");
809 psize = sysconf(_SC_PAGE_SIZE);
811 env->me_meta.mm_magic = MDB_MAGIC;
812 env->me_meta.mm_version = MDB_VERSION;
813 env->me_meta.mm_psize = psize;
814 env->me_meta.mm_flags = env->me_flags & 0xffff;
815 env->me_meta.mm_root = P_INVALID;
816 env->me_meta.mm_last_pg = 1;
818 p = calloc(2, psize);
820 p->mp_flags = P_META;
823 bcopy(&env->me_meta, meta, sizeof(*meta));
825 q = (MDB_page *)((char *)p + psize);
828 q->mp_flags = P_META;
831 bcopy(&env->me_meta, meta, sizeof(*meta));
833 rc = write(env->me_fd, p, psize * 2);
835 return (rc == (int)psize * 2) ? MDB_SUCCESS : errno;
839 mdbenv_write_meta(MDB_txn *txn)
848 assert(txn->mt_env != NULL);
850 DPRINTF("writing meta page for root page %lu", txn->mt_root);
855 off = offsetof(MDB_meta, mm_depth);
856 len = sizeof(MDB_meta) - off;
859 meta.mm_depth = env->me_db.md_depth;
860 meta.mm_branch_pages = env->me_db.md_branch_pages;
861 meta.mm_leaf_pages = env->me_db.md_leaf_pages;
862 meta.mm_overflow_pages = env->me_db.md_overflow_pages;
863 meta.mm_entries = env->me_db.md_entries;
864 meta.mm_root = txn->mt_root;
865 meta.mm_last_pg = txn->mt_next_pgno - 1;
866 meta.mm_txnid = txn->mt_txnid;
868 if (!F_ISSET(txn->mt_flags, MDB_TXN_METOGGLE))
869 off += env->me_meta.mm_psize;
872 lseek(env->me_fd, off, SEEK_SET);
873 rc = write(env->me_fd, ptr, len);
875 DPRINTF("write failed, disk error?");
882 /* Returns true if page p is a valid meta page, false otherwise.
885 mdb_check_meta_page(MDB_page *p)
887 if (!F_ISSET(p->mp_flags, P_META)) {
888 DPRINTF("page %lu not a meta page", p->mp_pgno);
896 mdbenv_read_meta(MDB_env *env, int *which)
904 if ((mp0 = mdbenv_get_page(env, 0)) == NULL ||
905 (mp1 = mdbenv_get_page(env, 1)) == NULL)
908 rc = mdb_check_meta_page(mp0);
911 rc = mdb_check_meta_page(mp1);
914 meta[0] = METADATA(mp0);
915 meta[1] = METADATA(mp1);
917 if (meta[0]->mm_txnid < meta[1]->mm_txnid)
920 if (meta[toggle]->mm_txnid > env->me_meta.mm_txnid) {
921 bcopy(meta[toggle], &env->me_meta, sizeof(env->me_meta));
926 DPRINTF("Using meta page %d", toggle);
932 mdbenv_create(MDB_env **env)
936 e = calloc(1, sizeof(*e));
937 if (!e) return ENOMEM;
939 e->me_meta.mm_mapsize = DEFAULT_MAPSIZE;
940 e->me_maxreaders = DEFAULT_READERS;
949 mdbenv_set_mapsize(MDB_env *env, size_t size)
953 env->me_mapsize = env->me_meta.mm_mapsize = size;
958 mdbenv_set_maxreaders(MDB_env *env, int readers)
960 env->me_maxreaders = readers;
965 mdbenv_get_maxreaders(MDB_env *env, int *readers)
967 if (!env || !readers)
969 *readers = env->me_maxreaders;
974 mdbenv_open2(MDB_env *env, unsigned int flags)
978 env->me_flags = flags;
980 if ((i = mdbenv_read_header(env)) != 0) {
983 DPRINTF("new mdbenv");
987 if (!env->me_mapsize)
988 env->me_mapsize = env->me_meta.mm_mapsize;
991 if (env->me_meta.mm_address && (flags & MDB_FIXEDMAP))
993 env->me_map = mmap(env->me_meta.mm_address, env->me_mapsize, PROT_READ, i,
995 if (env->me_map == MAP_FAILED)
999 env->me_meta.mm_mapsize = env->me_mapsize;
1000 if (flags & MDB_FIXEDMAP)
1001 env->me_meta.mm_address = env->me_map;
1002 i = mdbenv_init_meta(env);
1003 if (i != MDB_SUCCESS) {
1004 munmap(env->me_map, env->me_mapsize);
1009 if ((i = mdbenv_read_meta(env, NULL)) != 0)
1012 DPRINTF("opened database version %u, pagesize %u",
1013 env->me_meta.mm_version, env->me_meta.mm_psize);
1014 DPRINTF("depth: %u", env->me_meta.mm_depth);
1015 DPRINTF("entries: %lu", env->me_meta.mm_entries);
1016 DPRINTF("branch pages: %lu", env->me_meta.mm_branch_pages);
1017 DPRINTF("leaf pages: %lu", env->me_meta.mm_leaf_pages);
1018 DPRINTF("overflow pages: %lu", env->me_meta.mm_overflow_pages);
1019 DPRINTF("root: %lu", env->me_meta.mm_root);
1025 mdbenv_reader_dest(void *ptr)
1027 MDB_reader *reader = ptr;
1029 reader->mr_txnid = 0;
1035 mdbenv_share_locks(MDB_env *env)
1037 struct flock lock_info;
1039 env->me_txns->mt_txnid = env->me_meta.mm_txnid;
1041 memset((void *)&lock_info, 0, sizeof(lock_info));
1042 lock_info.l_type = F_RDLCK;
1043 lock_info.l_whence = SEEK_SET;
1044 lock_info.l_start = 0;
1045 lock_info.l_len = 1;
1046 fcntl(env->me_lfd, F_SETLK, &lock_info);
1050 mdbenv_setup_locks(MDB_env *env, char *lpath, int mode, int *excl)
1054 struct flock lock_info;
1058 if ((env->me_lfd = open(lpath, O_RDWR|O_CREAT, mode)) == -1) {
1062 /* Try to get exclusive lock. If we succeed, then
1063 * nobody is using the lock region and we should initialize it.
1065 memset((void *)&lock_info, 0, sizeof(lock_info));
1066 lock_info.l_type = F_WRLCK;
1067 lock_info.l_whence = SEEK_SET;
1068 lock_info.l_start = 0;
1069 lock_info.l_len = 1;
1070 rc = fcntl(env->me_lfd, F_SETLK, &lock_info);
1074 lock_info.l_type = F_RDLCK;
1075 rc = fcntl(env->me_lfd, F_SETLK, &lock_info);
1081 size = lseek(env->me_lfd, 0, SEEK_END);
1082 rsize = (env->me_maxreaders-1) * sizeof(MDB_reader) + sizeof(MDB_txninfo);
1083 if (size < rsize && *excl) {
1084 if (ftruncate(env->me_lfd, rsize) != 0) {
1090 size = rsize - sizeof(MDB_txninfo);
1091 env->me_maxreaders = size/sizeof(MDB_reader) + 1;
1093 env->me_txns = mmap(0, rsize, PROT_READ|PROT_WRITE, MAP_SHARED,
1095 if (env->me_txns == MAP_FAILED) {
1100 pthread_mutexattr_t mattr;
1102 pthread_mutexattr_init(&mattr);
1103 pthread_mutexattr_setpshared(&mattr, PTHREAD_PROCESS_SHARED);
1104 pthread_mutex_init(&env->me_txns->mt_mutex, &mattr);
1105 pthread_mutex_init(&env->me_txns->mt_wmutex, &mattr);
1106 env->me_txns->mt_version = MDB_VERSION;
1107 env->me_txns->mt_magic = MDB_MAGIC;
1108 env->me_txns->mt_txnid = 0;
1109 env->me_txns->mt_numreaders = 0;
1112 if (env->me_txns->mt_magic != MDB_MAGIC) {
1113 DPRINTF("lock region has invalid magic");
1116 if (env->me_txns->mt_version != MDB_VERSION) {
1117 DPRINTF("lock region is version %u, expected version %u",
1118 env->me_txns->mt_version, MDB_VERSION);
1121 if (errno != EACCES && errno != EAGAIN) {
1135 mdbenv_open(MDB_env *env, const char *path, unsigned int flags, mode_t mode)
1137 int oflags, rc, len, excl;
1138 char *lpath, *dpath;
1141 lpath = malloc(len + sizeof("/lock.mdb") + len + sizeof("/data.db"));
1144 dpath = lpath + len + sizeof("/lock.mdb");
1145 sprintf(lpath, "%s/lock.mdb", path);
1146 sprintf(dpath, "%s/data.mdb", path);
1148 rc = mdbenv_setup_locks(env, lpath, mode, &excl);
1152 if (F_ISSET(flags, MDB_RDONLY))
1155 oflags = O_RDWR | O_CREAT;
1157 if ((env->me_fd = open(dpath, oflags, mode)) == -1)
1160 if ((rc = mdbenv_open2(env, flags)) != MDB_SUCCESS) {
1164 env->me_path = strdup(path);
1165 DPRINTF("opened dbenv %p", (void *) env);
1166 pthread_key_create(&env->me_txkey, mdbenv_reader_dest);
1168 mdbenv_share_locks(env);
1178 mdbenv_close(MDB_env *env)
1186 munmap(env->me_map, env->me_mapsize);
1190 size_t size = (env->me_maxreaders-1) * sizeof(MDB_reader) + sizeof(MDB_txninfo);
1191 munmap(env->me_txns, size);
1197 /* Search for key within a leaf page, using binary search.
1198 * Returns the smallest entry larger or equal to the key.
1199 * If exactp is non-null, stores whether the found entry was an exact match
1200 * in *exactp (1 or 0).
1201 * If kip is non-null, stores the index of the found entry in *kip.
1202 * If no entry larger or equal to the key is found, returns NULL.
1205 mdb_search_node(MDB_db *bt, MDB_page *mp, MDB_val *key,
1206 int *exactp, unsigned int *kip)
1214 DPRINTF("searching %u keys in %s page %lu",
1216 IS_LEAF(mp) ? "leaf" : "branch",
1219 assert(NUMKEYS(mp) > 0);
1221 bzero(&nodekey, sizeof(nodekey));
1223 low = IS_LEAF(mp) ? 0 : 1;
1224 high = NUMKEYS(mp) - 1;
1225 while (low <= high) {
1226 i = (low + high) >> 1;
1227 node = NODEPTR(mp, i);
1229 nodekey.mv_size = node->mn_ksize;
1230 nodekey.mv_data = NODEKEY(node);
1233 rc = bt->md_cmp(key, &nodekey);
1235 rc = _mdb_cmp(bt, key, &nodekey);
1238 DPRINTF("found leaf index %u [%.*s], rc = %i",
1239 i, (int)nodekey.mv_size, (char *)nodekey.mv_data, rc);
1241 DPRINTF("found branch index %u [%.*s -> %lu], rc = %i",
1242 i, (int)node->mn_ksize, (char *)NODEKEY(node),
1253 if (rc > 0) { /* Found entry is less than the key. */
1254 i++; /* Skip to get the smallest entry larger than key. */
1255 if (i >= NUMKEYS(mp))
1256 /* There is no entry larger or equal to the key. */
1260 *exactp = (rc == 0);
1261 if (kip) /* Store the key index if requested. */
1264 return NODEPTR(mp, i);
1268 cursor_pop_page(MDB_cursor *cursor)
1272 top = CURSOR_TOP(cursor);
1275 DPRINTF("popped page %lu off cursor %p", top->mp_page->mp_pgno, (void *) cursor);
1281 cursor_push_page(MDB_cursor *cursor, MDB_page *mp)
1285 DPRINTF("pushing page %lu on cursor %p", mp->mp_pgno, (void *) cursor);
1287 if ((ppage = calloc(1, sizeof(*ppage))) == NULL)
1289 ppage->mp_page = mp;
1290 CURSOR_PUSH(cursor, ppage);
1295 mdbenv_get_page(MDB_env *env, pgno_t pgno)
1298 MDB_txn *txn = env->me_txn;
1301 if (txn && !SIMPLEQ_EMPTY(txn->mt_u.dirty_queue)) {
1303 SIMPLEQ_FOREACH(dp, txn->mt_u.dirty_queue, h.md_next) {
1304 if (dp->p.mp_pgno == pgno) {
1312 p = (MDB_page *)(env->me_map + env->me_meta.mm_psize * pgno);
1318 mdb_search_page_root(MDB_db *bt, MDB_val *key,
1319 MDB_cursor *cursor, int modify, MDB_pageparent *mpp)
1321 MDB_page *mp = mpp->mp_page;
1324 if (cursor && cursor_push_page(cursor, mp) == NULL)
1327 while (IS_BRANCH(mp)) {
1331 DPRINTF("branch page %lu has %u keys", mp->mp_pgno, NUMKEYS(mp));
1332 assert(NUMKEYS(mp) > 1);
1333 DPRINTF("found index 0 to page %lu", NODEPGNO(NODEPTR(mp, 0)));
1335 if (key == NULL) /* Initialize cursor to first page. */
1339 node = mdb_search_node(bt, mp, key, &exact, &i);
1341 i = NUMKEYS(mp) - 1;
1349 DPRINTF("following index %u for key %.*s",
1350 i, (int)key->mv_size, (char *)key->mv_data);
1351 assert(i < NUMKEYS(mp));
1352 node = NODEPTR(mp, i);
1355 CURSOR_TOP(cursor)->mp_ki = i;
1357 mpp->mp_parent = mp;
1358 if ((mp = mdbenv_get_page(bt->md_env, NODEPGNO(node))) == NULL)
1363 if (cursor && cursor_push_page(cursor, mp) == NULL)
1367 MDB_dhead *dh = ((MDB_dhead *)mp)-1;
1368 if ((rc = mdb_touch(bt->md_env->me_txn, mpp)) != 0)
1370 dh = ((MDB_dhead *)mpp->mp_page)-1;
1371 dh->md_parent = mpp->mp_parent;
1372 dh->md_pi = mpp->mp_pi;
1379 DPRINTF("internal error, index points to a %02X page!?",
1384 DPRINTF("found leaf page %lu for key %.*s", mp->mp_pgno,
1385 key ? (int)key->mv_size : 0, key ? (char *)key->mv_data : NULL);
1390 /* Search for the page a given key should be in.
1391 * Stores a pointer to the found page in *mpp.
1392 * If key is NULL, search for the lowest page (used by mdb_cursor_first).
1393 * If cursor is non-null, pushes parent pages on the cursor stack.
1394 * If modify is true, visited pages are updated with new page numbers.
1397 mdb_search_page(MDB_db *db, MDB_txn *txn, MDB_val *key,
1398 MDB_cursor *cursor, int modify, MDB_pageparent *mpp)
1403 /* Can't modify pages outside a transaction. */
1404 if (txn == NULL && modify) {
1408 /* Choose which root page to start with. If a transaction is given
1409 * use the root page from the transaction, otherwise read the last
1410 * committed root page.
1413 if ((rc = mdbenv_read_meta(db->md_env, NULL)) != MDB_SUCCESS)
1415 root = db->md_env->me_meta.mm_root;
1416 } else if (F_ISSET(txn->mt_flags, MDB_TXN_ERROR)) {
1417 DPRINTF("transaction has failed, must abort");
1420 root = txn->mt_root;
1422 if (root == P_INVALID) { /* Tree is empty. */
1423 DPRINTF("tree is empty");
1427 if ((mpp->mp_page = mdbenv_get_page(db->md_env, root)) == NULL)
1430 DPRINTF("root page has flags 0x%X", mpp->mp_page->mp_flags);
1432 if (modify && !F_ISSET(mpp->mp_page->mp_flags, P_DIRTY)) {
1433 mpp->mp_parent = NULL;
1435 if ((rc = mdb_touch(txn, mpp)))
1437 txn->mt_root = mpp->mp_page->mp_pgno;
1440 return mdb_search_page_root(db, key, cursor, modify, mpp);
1444 mdb_read_data(MDB_db *db, MDB_page *mp, MDB_node *leaf,
1447 MDB_page *omp; /* overflow mpage */
1451 bzero(data, sizeof(*data));
1452 max = db->md_env->me_meta.mm_psize - PAGEHDRSZ;
1454 if (!F_ISSET(leaf->mn_flags, F_BIGDATA)) {
1455 data->mv_size = leaf->mn_dsize;
1456 data->mv_data = NODEDATA(leaf);
1460 /* Read overflow data.
1462 data->mv_size = leaf->mn_dsize;
1463 bcopy(NODEDATA(leaf), &pgno, sizeof(pgno));
1464 if ((omp = mdbenv_get_page(db->md_env, pgno)) == NULL) {
1465 DPRINTF("read overflow page %lu failed", pgno);
1468 data->mv_data = omp;
1474 mdb_get(MDB_db *db, MDB_txn *txn,
1475 MDB_val *key, MDB_val *data)
1483 DPRINTF("===> get key [%.*s]", (int)key->mv_size, (char *)key->mv_data);
1488 if (txn != NULL && db->md_env != txn->mt_env) {
1492 if (key->mv_size == 0 || key->mv_size > MAXKEYSIZE) {
1496 if ((rc = mdb_search_page(db, txn, key, NULL, 0, &mpp)) != MDB_SUCCESS)
1499 leaf = mdb_search_node(db, mpp.mp_page, key, &exact, NULL);
1501 rc = mdb_read_data(db, mpp.mp_page, leaf, data);
1510 mdb_sibling(MDB_cursor *cursor, int move_right)
1514 MDB_ppage *parent, *top;
1517 top = CURSOR_TOP(cursor);
1518 if ((parent = SLIST_NEXT(top, mp_entry)) == NULL) {
1519 return ENOENT; /* root has no siblings */
1522 DPRINTF("parent page is page %lu, index %u",
1523 parent->mp_page->mp_pgno, parent->mp_ki);
1525 cursor_pop_page(cursor);
1526 if (move_right ? (parent->mp_ki + 1 >= NUMKEYS(parent->mp_page))
1527 : (parent->mp_ki == 0)) {
1528 DPRINTF("no more keys left, moving to %s sibling",
1529 move_right ? "right" : "left");
1530 if ((rc = mdb_sibling(cursor, move_right)) != MDB_SUCCESS)
1532 parent = CURSOR_TOP(cursor);
1538 DPRINTF("just moving to %s index key %u",
1539 move_right ? "right" : "left", parent->mp_ki);
1541 assert(IS_BRANCH(parent->mp_page));
1543 indx = NODEPTR(parent->mp_page, parent->mp_ki);
1544 if ((mp = mdbenv_get_page(cursor->mc_db->md_env, indx->mn_pgno)) == NULL)
1547 mp->parent = parent->mp_page;
1548 mp->parent_index = parent->mp_ki;
1551 cursor_push_page(cursor, mp);
1557 mdb_set_key(MDB_db *bt, MDB_page *mp, MDB_node *node,
1563 key->mv_size = node->mn_ksize;
1564 key->mv_data = NODEKEY(node);
1570 mdb_cursor_next(MDB_cursor *cursor, MDB_val *key, MDB_val *data)
1576 if (cursor->mc_eof) {
1580 assert(cursor->mc_initialized);
1582 top = CURSOR_TOP(cursor);
1585 DPRINTF("cursor_next: top page is %lu in cursor %p", mp->mp_pgno, (void *) cursor);
1587 if (top->mp_ki + 1 >= NUMKEYS(mp)) {
1588 DPRINTF("=====> move to next sibling page");
1589 if (mdb_sibling(cursor, 1) != MDB_SUCCESS) {
1593 top = CURSOR_TOP(cursor);
1595 DPRINTF("next page is %lu, key index %u", mp->mp_pgno, top->mp_ki);
1599 DPRINTF("==> cursor points to page %lu with %u keys, key index %u",
1600 mp->mp_pgno, NUMKEYS(mp), top->mp_ki);
1602 assert(IS_LEAF(mp));
1603 leaf = NODEPTR(mp, top->mp_ki);
1605 if (data && mdb_read_data(cursor->mc_db, mp, leaf, data) != MDB_SUCCESS)
1608 return mdb_set_key(cursor->mc_db, mp, leaf, key);
1612 mdb_cursor_set(MDB_cursor *cursor, MDB_val *key, MDB_val *data,
1622 assert(key->mv_size > 0);
1624 rc = mdb_search_page(cursor->mc_db, cursor->mc_txn, key, cursor, 0, &mpp);
1625 if (rc != MDB_SUCCESS)
1627 assert(IS_LEAF(mpp.mp_page));
1629 top = CURSOR_TOP(cursor);
1630 leaf = mdb_search_node(cursor->mc_db, mpp.mp_page, key, exactp, &top->mp_ki);
1631 if (exactp != NULL && !*exactp) {
1632 /* MDB_CURSOR_EXACT specified and not an exact match. */
1637 DPRINTF("===> inexact leaf not found, goto sibling");
1638 if ((rc = mdb_sibling(cursor, 1)) != MDB_SUCCESS)
1639 return rc; /* no entries matched */
1640 top = CURSOR_TOP(cursor);
1642 mpp.mp_page = top->mp_page;
1643 assert(IS_LEAF(mpp.mp_page));
1644 leaf = NODEPTR(mpp.mp_page, 0);
1647 cursor->mc_initialized = 1;
1650 if (data && (rc = mdb_read_data(cursor->mc_db, mpp.mp_page, leaf, data)) != MDB_SUCCESS)
1653 rc = mdb_set_key(cursor->mc_db, mpp.mp_page, leaf, key);
1654 if (rc == MDB_SUCCESS) {
1655 DPRINTF("==> cursor placed on key %.*s",
1656 (int)key->mv_size, (char *)key->mv_data);
1664 mdb_cursor_first(MDB_cursor *cursor, MDB_val *key, MDB_val *data)
1670 rc = mdb_search_page(cursor->mc_db, cursor->mc_txn, NULL, cursor, 0, &mpp);
1671 if (rc != MDB_SUCCESS)
1673 assert(IS_LEAF(mpp.mp_page));
1675 leaf = NODEPTR(mpp.mp_page, 0);
1676 cursor->mc_initialized = 1;
1679 if (data && (rc = mdb_read_data(cursor->mc_db, mpp.mp_page, leaf, data)) != MDB_SUCCESS)
1682 return mdb_set_key(cursor->mc_db, mpp.mp_page, leaf, key);
1686 mdb_cursor_get(MDB_cursor *cursor, MDB_val *key, MDB_val *data,
1696 case MDB_CURSOR_EXACT:
1697 while (CURSOR_TOP(cursor) != NULL)
1698 cursor_pop_page(cursor);
1699 if (key == NULL || key->mv_size == 0 || key->mv_size > MAXKEYSIZE) {
1701 } else if (op == MDB_CURSOR_EXACT)
1702 rc = mdb_cursor_set(cursor, key, data, &exact);
1704 rc = mdb_cursor_set(cursor, key, data, NULL);
1707 if (!cursor->mc_initialized)
1708 rc = mdb_cursor_first(cursor, key, data);
1710 rc = mdb_cursor_next(cursor, key, data);
1713 while (CURSOR_TOP(cursor) != NULL)
1714 cursor_pop_page(cursor);
1715 rc = mdb_cursor_first(cursor, key, data);
1718 DPRINTF("unhandled/unimplemented cursor operation %u", op);
1726 /* Allocate a page and initialize it
1729 mdb_new_page(MDB_db *db, uint32_t flags, int num)
1734 assert(db->md_env != NULL);
1735 assert(db->md_env->me_txn != NULL);
1737 if ((dp = mdb_alloc_page(db->md_env->me_txn, NULL, 0, num)) == NULL)
1739 DPRINTF("allocated new mpage %lu, page size %u",
1740 dp->p.mp_pgno, db->md_env->me_meta.mm_psize);
1741 dp->p.mp_flags = flags | P_DIRTY;
1742 dp->p.mp_lower = PAGEHDRSZ;
1743 dp->p.mp_upper = db->md_env->me_meta.mm_psize;
1745 if (IS_BRANCH(&dp->p))
1746 db->md_branch_pages++;
1747 else if (IS_LEAF(&dp->p))
1748 db->md_leaf_pages++;
1749 else if (IS_OVERFLOW(&dp->p)) {
1750 db->md_overflow_pages += num;
1751 dp->p.mp_pages = num;
1758 mdb_leaf_size(MDB_db *db, MDB_val *key, MDB_val *data)
1762 sz = LEAFSIZE(key, data);
1763 if (data->mv_size >= db->md_env->me_meta.mm_psize / MDB_MINKEYS) {
1764 /* put on overflow page */
1765 sz -= data->mv_size - sizeof(pgno_t);
1768 return sz + sizeof(indx_t);
1772 mdb_branch_size(MDB_db *db, MDB_val *key)
1777 if (sz >= db->md_env->me_meta.mm_psize / MDB_MINKEYS) {
1778 /* put on overflow page */
1779 /* not implemented */
1780 /* sz -= key->size - sizeof(pgno_t); */
1783 return sz + sizeof(indx_t);
1787 mdb_add_node(MDB_db *db, MDB_page *mp, indx_t indx,
1788 MDB_val *key, MDB_val *data, pgno_t pgno, uint8_t flags)
1791 size_t node_size = NODESIZE;
1794 MDB_dpage *ofp = NULL; /* overflow page */
1796 assert(mp->mp_upper >= mp->mp_lower);
1798 DPRINTF("add node [%.*s] to %s page %lu at index %i, key size %zu",
1799 key ? (int)key->mv_size : 0, key ? (char *)key->mv_data : NULL,
1800 IS_LEAF(mp) ? "leaf" : "branch",
1801 mp->mp_pgno, indx, key ? key->mv_size : 0);
1804 node_size += key->mv_size;
1808 if (F_ISSET(flags, F_BIGDATA)) {
1809 /* Data already on overflow page. */
1810 node_size += sizeof(pgno_t);
1811 } else if (data->mv_size >= db->md_env->me_meta.mm_psize / MDB_MINKEYS) {
1812 int ovpages = OVPAGES(data->mv_size, db->md_env->me_meta.mm_psize);
1813 /* Put data on overflow page. */
1814 DPRINTF("data size is %zu, put on overflow page",
1816 node_size += sizeof(pgno_t);
1817 if ((ofp = mdb_new_page(db, P_OVERFLOW, ovpages)) == NULL)
1819 DPRINTF("allocated overflow page %lu", ofp->p.mp_pgno);
1822 node_size += data->mv_size;
1826 if (node_size + sizeof(indx_t) > SIZELEFT(mp)) {
1827 DPRINTF("not enough room in page %lu, got %u ptrs",
1828 mp->mp_pgno, NUMKEYS(mp));
1829 DPRINTF("upper - lower = %u - %u = %u", mp->mp_upper, mp->mp_lower,
1830 mp->mp_upper - mp->mp_lower);
1831 DPRINTF("node size = %zu", node_size);
1835 /* Move higher pointers up one slot. */
1836 for (i = NUMKEYS(mp); i > indx; i--)
1837 mp->mp_ptrs[i] = mp->mp_ptrs[i - 1];
1839 /* Adjust free space offsets. */
1840 ofs = mp->mp_upper - node_size;
1841 assert(ofs >= mp->mp_lower + sizeof(indx_t));
1842 mp->mp_ptrs[indx] = ofs;
1844 mp->mp_lower += sizeof(indx_t);
1846 /* Write the node data. */
1847 node = NODEPTR(mp, indx);
1848 node->mn_ksize = (key == NULL) ? 0 : key->mv_size;
1849 node->mn_flags = flags;
1851 node->mn_dsize = data->mv_size;
1853 node->mn_pgno = pgno;
1856 bcopy(key->mv_data, NODEKEY(node), key->mv_size);
1861 if (F_ISSET(flags, F_BIGDATA))
1862 bcopy(data->mv_data, node->mn_data + key->mv_size,
1865 bcopy(data->mv_data, node->mn_data + key->mv_size,
1868 bcopy(&ofp->p.mp_pgno, node->mn_data + key->mv_size,
1870 bcopy(data->mv_data, METADATA(&ofp->p), data->mv_size);
1878 mdb_del_node(MDB_db *db, MDB_page *mp, indx_t indx)
1881 indx_t i, j, numkeys, ptr;
1885 DPRINTF("delete node %u on %s page %lu", indx,
1886 IS_LEAF(mp) ? "leaf" : "branch", mp->mp_pgno);
1887 assert(indx < NUMKEYS(mp));
1889 node = NODEPTR(mp, indx);
1890 sz = NODESIZE + node->mn_ksize;
1892 if (F_ISSET(node->mn_flags, F_BIGDATA))
1893 sz += sizeof(pgno_t);
1895 sz += NODEDSZ(node);
1898 ptr = mp->mp_ptrs[indx];
1899 numkeys = NUMKEYS(mp);
1900 for (i = j = 0; i < numkeys; i++) {
1902 mp->mp_ptrs[j] = mp->mp_ptrs[i];
1903 if (mp->mp_ptrs[i] < ptr)
1904 mp->mp_ptrs[j] += sz;
1909 base = (char *)mp + mp->mp_upper;
1910 bcopy(base, base + sz, ptr - mp->mp_upper);
1912 mp->mp_lower -= sizeof(indx_t);
1917 mdb_cursor_open(MDB_db *db, MDB_txn *txn, MDB_cursor **ret)
1921 if (db == NULL || ret == NULL)
1924 if (txn != NULL && db->md_env != txn->mt_env) {
1928 if ((cursor = calloc(1, sizeof(*cursor))) != NULL) {
1929 SLIST_INIT(&cursor->mc_stack);
1931 cursor->mc_txn = txn;
1940 mdb_cursor_close(MDB_cursor *cursor)
1942 if (cursor != NULL) {
1943 while (!CURSOR_EMPTY(cursor))
1944 cursor_pop_page(cursor);
1946 /* btree_close(cursor->bt); */
1952 mdb_update_key(MDB_db *db, MDB_page *mp, indx_t indx,
1955 indx_t ptr, i, numkeys;
1961 node = NODEPTR(mp, indx);
1962 ptr = mp->mp_ptrs[indx];
1963 DPRINTF("update key %u (ofs %u) [%.*s] to [%.*s] on page %lu",
1965 (int)node->mn_ksize, (char *)NODEKEY(node),
1966 (int)key->mv_size, (char *)key->mv_data,
1969 delta = key->mv_size - node->mn_ksize;
1971 if (delta > 0 && SIZELEFT(mp) < delta) {
1972 DPRINTF("OUCH! Not enough room, delta = %d", delta);
1976 numkeys = NUMKEYS(mp);
1977 for (i = 0; i < numkeys; i++) {
1978 if (mp->mp_ptrs[i] <= ptr)
1979 mp->mp_ptrs[i] -= delta;
1982 base = (char *)mp + mp->mp_upper;
1983 len = ptr - mp->mp_upper + NODESIZE;
1984 bcopy(base, base - delta, len);
1985 mp->mp_upper -= delta;
1987 node = NODEPTR(mp, indx);
1988 node->mn_ksize = key->mv_size;
1991 bcopy(key->mv_data, NODEKEY(node), key->mv_size);
1996 /* Move a node from src to dst.
1999 mdb_move_node(MDB_db *bt, MDB_pageparent *src, indx_t srcindx,
2000 MDB_pageparent *dst, indx_t dstindx)
2006 srcnode = NODEPTR(src->mp_page, srcindx);
2007 DPRINTF("moving %s node %u [%.*s] on page %lu to node %u on page %lu",
2008 IS_LEAF(src->mp_page) ? "leaf" : "branch",
2010 (int)srcnode->mn_ksize, (char *)NODEKEY(srcnode),
2011 src->mp_page->mp_pgno,
2012 dstindx, dst->mp_page->mp_pgno);
2014 /* Mark src and dst as dirty. */
2015 if ((rc = mdb_touch(bt->md_env->me_txn, src)) ||
2016 (rc = mdb_touch(bt->md_env->me_txn, dst)))
2019 /* Add the node to the destination page.
2021 key.mv_size = srcnode->mn_ksize;
2022 key.mv_data = NODEKEY(srcnode);
2023 data.mv_size = NODEDSZ(srcnode);
2024 data.mv_data = NODEDATA(srcnode);
2025 rc = mdb_add_node(bt, dst->mp_page, dstindx, &key, &data, NODEPGNO(srcnode),
2027 if (rc != MDB_SUCCESS)
2030 /* Delete the node from the source page.
2032 mdb_del_node(bt, src->mp_page, srcindx);
2034 /* Update the parent separators.
2036 if (srcindx == 0 && src->mp_pi != 0) {
2037 DPRINTF("update separator for source page %lu to [%.*s]",
2038 src->mp_page->mp_pgno, (int)key.mv_size, (char *)key.mv_data);
2039 if ((rc = mdb_update_key(bt, src->mp_parent, src->mp_pi,
2040 &key)) != MDB_SUCCESS)
2044 if (srcindx == 0 && IS_BRANCH(src->mp_page)) {
2046 nullkey.mv_size = 0;
2047 assert(mdb_update_key(bt, src->mp_page, 0, &nullkey) == MDB_SUCCESS);
2050 if (dstindx == 0 && dst->mp_pi != 0) {
2051 DPRINTF("update separator for destination page %lu to [%.*s]",
2052 dst->mp_page->mp_pgno, (int)key.mv_size, (char *)key.mv_data);
2053 if ((rc = mdb_update_key(bt, dst->mp_parent, dst->mp_pi,
2054 &key)) != MDB_SUCCESS)
2058 if (dstindx == 0 && IS_BRANCH(dst->mp_page)) {
2060 nullkey.mv_size = 0;
2061 assert(mdb_update_key(bt, dst->mp_page, 0, &nullkey) == MDB_SUCCESS);
2068 mdb_merge(MDB_db *bt, MDB_pageparent *src, MDB_pageparent *dst)
2077 DPRINTF("merging page %lu and %lu", src->mp_page->mp_pgno, dst->mp_page->mp_pgno);
2079 assert(src->mp_parent); /* can't merge root page */
2080 assert(dst->mp_parent);
2081 assert(bt->md_env->me_txn != NULL);
2083 /* Mark src and dst as dirty. */
2084 if ((rc = mdb_touch(bt->md_env->me_txn, src)) ||
2085 (rc = mdb_touch(bt->md_env->me_txn, dst)))
2088 /* Move all nodes from src to dst.
2090 for (i = 0; i < NUMKEYS(src->mp_page); i++) {
2091 srcnode = NODEPTR(src->mp_page, i);
2093 key.mv_size = srcnode->mn_ksize;
2094 key.mv_data = NODEKEY(srcnode);
2095 data.mv_size = NODEDSZ(srcnode);
2096 data.mv_data = NODEDATA(srcnode);
2097 rc = mdb_add_node(bt, dst->mp_page, NUMKEYS(dst->mp_page), &key,
2098 &data, NODEPGNO(srcnode), srcnode->mn_flags);
2099 if (rc != MDB_SUCCESS)
2103 DPRINTF("dst page %lu now has %u keys (%.1f%% filled)",
2104 dst->mp_page->mp_pgno, NUMKEYS(dst->mp_page), (float)PAGEFILL(bt->md_env, dst->mp_page) / 10);
2106 /* Unlink the src page from parent.
2108 mdb_del_node(bt, src->mp_parent, src->mp_pi);
2109 if (src->mp_pi == 0) {
2111 if ((rc = mdb_update_key(bt, src->mp_parent, 0, &key)) != MDB_SUCCESS)
2115 if (IS_LEAF(src->mp_page))
2116 bt->md_leaf_pages--;
2118 bt->md_branch_pages--;
2120 mpp.mp_page = src->mp_parent;
2121 dh = (MDB_dhead *)src->mp_parent;
2123 mpp.mp_parent = dh->md_parent;
2124 mpp.mp_pi = dh->md_pi;
2126 return mdb_rebalance(bt, &mpp);
2129 #define FILL_THRESHOLD 250
2132 mdb_rebalance(MDB_db *db, MDB_pageparent *mpp)
2137 indx_t si = 0, di = 0;
2140 assert(db->md_env->me_txn != NULL);
2141 assert(mpp != NULL);
2143 DPRINTF("rebalancing %s page %lu (has %u keys, %.1f%% full)",
2144 IS_LEAF(mpp->mp_page) ? "leaf" : "branch",
2145 mpp->mp_page->mp_pgno, NUMKEYS(mpp->mp_page), (float)PAGEFILL(db->md_env, mpp->mp_page) / 10);
2147 if (PAGEFILL(db->md_env, mpp->mp_page) >= FILL_THRESHOLD) {
2148 DPRINTF("no need to rebalance page %lu, above fill threshold",
2149 mpp->mp_page->mp_pgno);
2153 if (mpp->mp_parent == NULL) {
2154 if (NUMKEYS(mpp->mp_page) == 0) {
2155 DPRINTF("tree is completely empty");
2156 db->md_root = P_INVALID;
2158 db->md_leaf_pages--;
2159 } else if (IS_BRANCH(mpp->mp_page) && NUMKEYS(mpp->mp_page) == 1) {
2160 DPRINTF("collapsing root page!");
2161 db->md_root = NODEPGNO(NODEPTR(mpp->mp_page, 0));
2162 if ((root = mdbenv_get_page(db->md_env, db->md_root)) == NULL)
2165 db->md_branch_pages--;
2167 DPRINTF("root page doesn't need rebalancing");
2171 /* The parent (branch page) must have at least 2 pointers,
2172 * otherwise the tree is invalid.
2174 assert(NUMKEYS(mpp->mp_parent) > 1);
2176 /* Leaf page fill factor is below the threshold.
2177 * Try to move keys from left or right neighbor, or
2178 * merge with a neighbor page.
2183 if (mpp->mp_pi == 0) {
2184 /* We're the leftmost leaf in our parent.
2186 DPRINTF("reading right neighbor");
2187 node = NODEPTR(mpp->mp_parent, mpp->mp_pi + 1);
2188 if ((npp.mp_page = mdbenv_get_page(db->md_env, NODEPGNO(node))) == NULL)
2190 npp.mp_pi = mpp->mp_pi + 1;
2192 di = NUMKEYS(mpp->mp_page);
2194 /* There is at least one neighbor to the left.
2196 DPRINTF("reading left neighbor");
2197 node = NODEPTR(mpp->mp_parent, mpp->mp_pi - 1);
2198 if ((npp.mp_page = mdbenv_get_page(db->md_env, NODEPGNO(node))) == NULL)
2200 npp.mp_pi = mpp->mp_pi - 1;
2201 si = NUMKEYS(npp.mp_page) - 1;
2204 npp.mp_parent = mpp->mp_parent;
2206 DPRINTF("found neighbor page %lu (%u keys, %.1f%% full)",
2207 npp.mp_page->mp_pgno, NUMKEYS(npp.mp_page), (float)PAGEFILL(db->md_env, npp.mp_page) / 10);
2209 /* If the neighbor page is above threshold and has at least two
2210 * keys, move one key from it.
2212 * Otherwise we should try to merge them.
2214 if (PAGEFILL(db->md_env, npp.mp_page) >= FILL_THRESHOLD && NUMKEYS(npp.mp_page) >= 2)
2215 return mdb_move_node(db, &npp, si, mpp, di);
2216 else { /* FIXME: if (has_enough_room()) */
2217 if (mpp->mp_pi == 0)
2218 return mdb_merge(db, &npp, mpp);
2220 return mdb_merge(db, mpp, &npp);
2225 mdb_del(MDB_db *bt, MDB_txn *txn,
2226 MDB_val *key, MDB_val *data)
2233 DPRINTF("========> delete key %.*s", (int)key->mv_size, (char *)key->mv_data);
2235 assert(key != NULL);
2237 if (bt == NULL || txn == NULL)
2240 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
2244 if (bt->md_env->me_txn != txn) {
2248 if (key->mv_size == 0 || key->mv_size > MAXKEYSIZE) {
2252 if ((rc = mdb_search_page(bt, txn, key, NULL, 1, &mpp)) != MDB_SUCCESS)
2255 leaf = mdb_search_node(bt, mpp.mp_page, key, &exact, &ki);
2256 if (leaf == NULL || !exact) {
2260 if (data && (rc = mdb_read_data(bt, NULL, leaf, data)) != MDB_SUCCESS)
2263 mdb_del_node(bt, mpp.mp_page, ki);
2264 /* add overflow pages to free list */
2265 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) {
2269 bcopy(NODEDATA(leaf), &pg, sizeof(pg));
2270 ovpages = OVPAGES(NODEDSZ(leaf), bt->md_env->me_meta.mm_psize);
2271 for (i=0; i<ovpages; i++) {
2272 mdb_idl_insert(txn->mt_free_pgs, pg);
2277 rc = mdb_rebalance(bt, &mpp);
2278 if (rc != MDB_SUCCESS)
2279 txn->mt_flags |= MDB_TXN_ERROR;
2284 /* Split page <*mpp>, and insert <key,(data|newpgno)> in either left or
2285 * right sibling, at index <*newindxp> (as if unsplit). Updates *mpp and
2286 * *newindxp with the actual values after split, ie if *mpp and *newindxp
2287 * refer to a node in the new right sibling page.
2290 mdb_split(MDB_db *bt, MDB_page **mpp, unsigned int *newindxp,
2291 MDB_val *newkey, MDB_val *newdata, pgno_t newpgno)
2294 int rc = MDB_SUCCESS, ins_new = 0;
2297 unsigned int i, j, split_indx;
2299 MDB_val sepkey, rkey, rdata;
2301 MDB_dpage *mdp, *rdp, *pdp;
2305 assert(bt->md_env != NULL);
2307 dh = ((MDB_dhead *)*mpp) - 1;
2308 mdp = (MDB_dpage *)dh;
2309 newindx = *newindxp;
2311 DPRINTF("-----> splitting %s page %lu and adding [%.*s] at index %i",
2312 IS_LEAF(&mdp->p) ? "leaf" : "branch", mdp->p.mp_pgno,
2313 (int)newkey->mv_size, (char *)newkey->mv_data, *newindxp);
2315 if (mdp->h.md_parent == NULL) {
2316 if ((pdp = mdb_new_page(bt, P_BRANCH, 1)) == NULL)
2319 mdp->h.md_parent = &pdp->p;
2320 bt->md_env->me_txn->mt_root = pdp->p.mp_pgno;
2321 DPRINTF("root split! new root = %lu", pdp->p.mp_pgno);
2324 /* Add left (implicit) pointer. */
2325 if (mdb_add_node(bt, &pdp->p, 0, NULL, NULL,
2326 mdp->p.mp_pgno, 0) != MDB_SUCCESS)
2329 DPRINTF("parent branch page is %lu", mdp->h.md_parent->mp_pgno);
2332 /* Create a right sibling. */
2333 if ((rdp = mdb_new_page(bt, mdp->p.mp_flags, 1)) == NULL)
2335 rdp->h.md_parent = mdp->h.md_parent;
2336 rdp->h.md_pi = mdp->h.md_pi + 1;
2337 DPRINTF("new right sibling: page %lu", rdp->p.mp_pgno);
2339 /* Move half of the keys to the right sibling. */
2340 if ((copy = malloc(bt->md_env->me_meta.mm_psize)) == NULL)
2342 bcopy(&mdp->p, copy, bt->md_env->me_meta.mm_psize);
2343 bzero(&mdp->p.mp_ptrs, bt->md_env->me_meta.mm_psize - PAGEHDRSZ);
2344 mdp->p.mp_lower = PAGEHDRSZ;
2345 mdp->p.mp_upper = bt->md_env->me_meta.mm_psize;
2347 split_indx = NUMKEYS(copy) / 2 + 1;
2349 /* First find the separating key between the split pages.
2351 bzero(&sepkey, sizeof(sepkey));
2352 if (newindx == split_indx) {
2353 sepkey.mv_size = newkey->mv_size;
2354 sepkey.mv_data = newkey->mv_data;
2356 node = NODEPTR(copy, split_indx);
2357 sepkey.mv_size = node->mn_ksize;
2358 sepkey.mv_data = NODEKEY(node);
2361 DPRINTF("separator is [%.*s]", (int)sepkey.mv_size, (char *)sepkey.mv_data);
2363 /* Copy separator key to the parent.
2365 if (SIZELEFT(rdp->h.md_parent) < mdb_branch_size(bt, &sepkey)) {
2366 rc = mdb_split(bt, &rdp->h.md_parent, &rdp->h.md_pi,
2367 &sepkey, NULL, rdp->p.mp_pgno);
2369 /* Right page might now have changed parent.
2370 * Check if left page also changed parent.
2372 if (rdp->h.md_parent != mdp->h.md_parent &&
2373 mdp->h.md_pi >= NUMKEYS(mdp->h.md_parent)) {
2374 mdp->h.md_parent = rdp->h.md_parent;
2375 mdp->h.md_pi = rdp->h.md_pi - 1;
2378 rc = mdb_add_node(bt, rdp->h.md_parent, rdp->h.md_pi,
2379 &sepkey, NULL, rdp->p.mp_pgno, 0);
2381 if (rc != MDB_SUCCESS) {
2386 for (i = j = 0; i <= NUMKEYS(copy); j++) {
2387 if (i < split_indx) {
2388 /* Re-insert in left sibling. */
2391 /* Insert in right sibling. */
2392 if (i == split_indx)
2393 /* Reset insert index for right sibling. */
2394 j = (i == newindx && ins_new);
2398 if (i == newindx && !ins_new) {
2399 /* Insert the original entry that caused the split. */
2400 rkey.mv_data = newkey->mv_data;
2401 rkey.mv_size = newkey->mv_size;
2402 if (IS_LEAF(&mdp->p)) {
2403 rdata.mv_data = newdata->mv_data;
2404 rdata.mv_size = newdata->mv_size;
2411 /* Update page and index for the new key. */
2414 } else if (i == NUMKEYS(copy)) {
2417 node = NODEPTR(copy, i);
2418 rkey.mv_data = NODEKEY(node);
2419 rkey.mv_size = node->mn_ksize;
2420 if (IS_LEAF(&mdp->p)) {
2421 rdata.mv_data = NODEDATA(node);
2422 rdata.mv_size = node->mn_dsize;
2424 pgno = node->mn_pgno;
2425 flags = node->mn_flags;
2430 if (!IS_LEAF(&mdp->p) && j == 0) {
2431 /* First branch index doesn't need key data. */
2435 rc = mdb_add_node(bt, &pdp->p, j, &rkey, &rdata, pgno,flags);
2443 mdb_put(MDB_db *bt, MDB_txn *txn,
2444 MDB_val *key, MDB_val *data, unsigned int flags)
2446 int rc = MDB_SUCCESS, exact;
2451 assert(key != NULL);
2452 assert(data != NULL);
2454 if (bt == NULL || txn == NULL)
2457 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
2461 if (bt->md_env->me_txn != txn) {
2465 if (key->mv_size == 0 || key->mv_size > MAXKEYSIZE) {
2469 DPRINTF("==> put key %.*s, size %zu, data size %zu",
2470 (int)key->mv_size, (char *)key->mv_data, key->mv_size, data->mv_size);
2472 rc = mdb_search_page(bt, txn, key, NULL, 1, &mpp);
2473 if (rc == MDB_SUCCESS) {
2474 leaf = mdb_search_node(bt, mpp.mp_page, key, &exact, &ki);
2475 if (leaf && exact) {
2476 if (F_ISSET(flags, MDB_NOOVERWRITE)) {
2477 DPRINTF("duplicate key %.*s",
2478 (int)key->mv_size, (char *)key->mv_data);
2481 mdb_del_node(bt, mpp.mp_page, ki);
2483 if (leaf == NULL) { /* append if not found */
2484 ki = NUMKEYS(mpp.mp_page);
2485 DPRINTF("appending key at index %i", ki);
2487 } else if (rc == ENOENT) {
2489 /* new file, just write a root leaf page */
2490 DPRINTF("allocating new root leaf page");
2491 if ((dp = mdb_new_page(bt, P_LEAF, 1)) == NULL) {
2494 mpp.mp_page = &dp->p;
2495 txn->mt_root = mpp.mp_page->mp_pgno;
2502 assert(IS_LEAF(mpp.mp_page));
2503 DPRINTF("there are %u keys, should insert new key at index %i",
2504 NUMKEYS(mpp.mp_page), ki);
2506 if (SIZELEFT(mpp.mp_page) < mdb_leaf_size(bt, key, data)) {
2507 rc = mdb_split(bt, &mpp.mp_page, &ki, key, data, P_INVALID);
2509 /* There is room already in this leaf page. */
2510 rc = mdb_add_node(bt, mpp.mp_page, ki, key, data, 0, 0);
2513 if (rc != MDB_SUCCESS)
2514 txn->mt_flags |= MDB_TXN_ERROR;
2523 mdbenv_get_flags(MDB_env *env, unsigned int *arg)
2528 *arg = env->me_flags;
2533 mdbenv_get_path(MDB_env *env, const char **arg)
2538 *arg = env->me_path;
2543 mdbenv_stat(MDB_env *env, MDB_stat *arg)
2545 if (env == NULL || arg == NULL)
2548 arg->ms_psize = env->me_meta.mm_psize;
2549 arg->ms_depth = env->me_meta.mm_depth;
2550 arg->ms_branch_pages = env->me_meta.mm_branch_pages;
2551 arg->ms_leaf_pages = env->me_meta.mm_leaf_pages;
2552 arg->ms_overflow_pages = env->me_meta.mm_overflow_pages;
2553 arg->ms_entries = env->me_meta.mm_entries;
2558 int mdb_open(MDB_env *env, MDB_txn *txn, const char *name, unsigned int flags, MDB_db **db)
2561 *db = (MDB_db *)&env->me_db;
2567 int mdb_stat(MDB_db *db, MDB_stat *arg)
2569 if (db == NULL || arg == NULL)
2572 arg->ms_psize = db->md_env->me_meta.mm_psize;
2573 if (!db->md_parent) {
2574 arg->ms_depth = db->md_env->me_meta.mm_depth;
2575 arg->ms_branch_pages = db->md_env->me_meta.mm_branch_pages;
2576 arg->ms_leaf_pages = db->md_env->me_meta.mm_leaf_pages;
2577 arg->ms_overflow_pages = db->md_env->me_meta.mm_overflow_pages;
2578 arg->ms_entries = db->md_env->me_meta.mm_entries;
2580 arg->ms_depth = db->md_depth;
2581 arg->ms_branch_pages = db->md_branch_pages;
2582 arg->ms_leaf_pages = db->md_leaf_pages;
2583 arg->ms_overflow_pages = db->md_overflow_pages;
2584 arg->ms_entries = db->md_entries;
2590 void mdb_close(MDB_db *db)