32 #if (DEBUG +0) && defined(__GNUC__)
33 # define DPRINTF(fmt, ...) \
34 fprintf(stderr, "%s:%d: " fmt "\n", __func__, __LINE__, ##__VA_ARGS__)
36 # define DPRINTF(...) ((void) 0)
41 #define MDB_MAGIC 0xBEEFC0DE
43 #define MAXKEYSIZE 255
45 #define P_INVALID (~0UL)
47 #define F_ISSET(w, f) (((w) & (f)) == (f))
50 typedef uint16_t indx_t;
52 #define DEFAULT_READERS 126
53 #define DEFAULT_MAPSIZE 1048576
55 /* Lock descriptor stuff */
60 typedef struct MDB_rxbody {
65 #define CACHELINE 64 /* most CPUs. Itanium uses 128 */
68 typedef struct MDB_reader {
70 /* cache line alignment */
71 char pad[CACHELINE-sizeof(MDB_rxbody)];
76 uint32_t mt_version; \
77 pthread_mutex_t mt_mutex; \
79 uint32_t mt_numreaders
80 typedef struct MDB_txbody {
84 typedef struct MDB_txninfo {
86 char pad[CACHELINE-sizeof(MDB_txbody)];
87 pthread_mutex_t mt_wmutex;
88 char pad2[CACHELINE-sizeof(pthread_mutex_t)];
89 MDB_reader mt_readers[1];
92 /* Common header for all page types. Overflow pages
93 * occupy a number of contiguous pages with no
94 * headers on any page after the first.
96 typedef struct MDB_page { /* represents a page of storage */
97 pgno_t mp_pgno; /* page number */
98 #define P_BRANCH 0x01 /* branch page */
99 #define P_LEAF 0x02 /* leaf page */
100 #define P_OVERFLOW 0x04 /* overflow page */
101 #define P_META 0x08 /* meta page */
102 #define P_DIRTY 0x10 /* dirty page */
104 #define mp_lower mp_pb.pb.pb_lower
105 #define mp_upper mp_pb.pb.pb_upper
106 #define mp_pages mp_pb.pb_pages
109 indx_t pb_lower; /* lower bound of free space */
110 indx_t pb_upper; /* upper bound of free space */
112 uint32_t pb_pages; /* number of overflow pages */
114 indx_t mp_ptrs[1]; /* dynamic size */
117 #define PAGEHDRSZ ((unsigned) offsetof(MDB_page, mp_ptrs))
119 #define NUMKEYS(p) (((p)->mp_lower - PAGEHDRSZ) >> 1)
120 #define SIZELEFT(p) (indx_t)((p)->mp_upper - (p)->mp_lower)
121 #define PAGEFILL(env, p) (1000L * ((env)->me_meta.mm_stat.ms_psize - PAGEHDRSZ - SIZELEFT(p)) / \
122 ((env)->me_meta.mm_stat.ms_psize - PAGEHDRSZ))
123 #define IS_LEAF(p) F_ISSET((p)->mp_flags, P_LEAF)
124 #define IS_BRANCH(p) F_ISSET((p)->mp_flags, P_BRANCH)
125 #define IS_OVERFLOW(p) F_ISSET((p)->mp_flags, P_OVERFLOW)
127 typedef struct MDB_meta { /* meta (footer) page content */
130 void *mm_address; /* address for fixed mapping */
131 size_t mm_mapsize; /* size of mmap region */
132 pgno_t mm_last_pg; /* last used page in file */
133 ulong mm_txnid; /* txnid that committed this page */
135 pgno_t mm_root; /* page number of root page */
138 typedef struct MDB_dhead { /* a dirty page */
139 SIMPLEQ_ENTRY(MDB_dpage) md_next; /* queue of dirty pages */
141 unsigned md_pi; /* parent index */
145 typedef struct MDB_dpage {
150 SIMPLEQ_HEAD(dirty_queue, MDB_dpage);
152 typedef struct MDB_pageparent {
158 static MDB_dpage *mdb_newpage(MDB_txn *txn, MDB_page *parent, unsigned int parent_idx, int num);
159 static int mdb_touch(MDB_txn *txn, MDB_pageparent *mp);
161 typedef struct MDB_ppage { /* ordered list of pages */
162 SLIST_ENTRY(MDB_ppage) mp_entry;
164 unsigned int mp_ki; /* cursor index on page */
166 SLIST_HEAD(page_stack, MDB_ppage);
168 #define CURSOR_EMPTY(c) SLIST_EMPTY(&(c)->mc_stack)
169 #define CURSOR_TOP(c) SLIST_FIRST(&(c)->mc_stack)
170 #define CURSOR_POP(c) SLIST_REMOVE_HEAD(&(c)->mc_stack, mp_entry)
171 #define CURSOR_PUSH(c,p) SLIST_INSERT_HEAD(&(c)->mc_stack, p, mp_entry)
176 struct page_stack mc_stack; /* stack of parent pages */
177 short mc_initialized; /* 1 if initialized */
178 short mc_eof; /* 1 if end is reached */
181 #define METAHASHLEN offsetof(MDB_meta, mm_hash)
182 #define METADATA(p) ((void *)((char *)p + PAGEHDRSZ))
184 typedef struct MDB_node {
185 #define mn_pgno mn_p.np_pgno
186 #define mn_dsize mn_p.np_dsize
188 pgno_t np_pgno; /* child page number */
189 uint32_t np_dsize; /* leaf data size */
191 unsigned int mn_flags:4;
192 unsigned int mn_ksize:12; /* key size */
193 #define F_BIGDATA 0x01 /* data put on overflow page */
198 pgno_t mt_root; /* current / new root page */
199 pgno_t mt_next_pgno; /* next unallocated page */
200 pgno_t mt_first_pgno;
204 struct dirty_queue *dirty_queue; /* modified pages */
207 #define MDB_TXN_RDONLY 0x01 /* read-only transaction */
208 #define MDB_TXN_ERROR 0x02 /* an error has occurred */
209 unsigned int mt_flags;
212 /* Must be same as MDB_db, minus md_root/md_stat */
213 typedef struct MDB_db0 {
214 unsigned int md_flags;
215 MDB_cmp_func *md_cmp; /* user compare function */
216 MDB_rel_func *md_rel; /* user relocate function */
217 MDB_db *md_parent; /* parent tree */
222 unsigned int md_flags;
223 MDB_cmp_func *md_cmp; /* user compare function */
224 MDB_rel_func *md_rel; /* user relocate function */
225 MDB_db *md_parent; /* parent tree */
228 pgno_t md_root; /* page number of root page */
239 MDB_txninfo *me_txns;
240 MDB_db0 me_db; /* first DB, overlaps with meta */
242 MDB_txn *me_txn; /* current write transaction */
244 off_t me_size; /* current file size */
245 pthread_key_t me_txkey; /* thread-key for readers */
248 #define NODESIZE offsetof(MDB_node, mn_data)
250 #define INDXSIZE(k) (NODESIZE + ((k) == NULL ? 0 : (k)->mv_size))
251 #define LEAFSIZE(k, d) (NODESIZE + (k)->mv_size + (d)->mv_size)
252 #define NODEPTR(p, i) ((MDB_node *)((char *)(p) + (p)->mp_ptrs[i]))
253 #define NODEKEY(node) (void *)((node)->mn_data)
254 #define NODEDATA(node) (void *)((char *)(node)->mn_data + (node)->mn_ksize)
255 #define NODEPGNO(node) ((node)->mn_pgno)
256 #define NODEDSZ(node) ((node)->mn_dsize)
258 #define MDB_COMMIT_PAGES 64 /* max number of pages to write in one commit */
259 #define MDB_MAXCACHE_DEF 1024 /* max number of pages to keep in cache */
261 static int mdb_search_page_root(MDB_db *db,
263 MDB_cursor *cursor, int modify,
264 MDB_pageparent *mpp);
265 static int mdb_search_page(MDB_db *db,
266 MDB_txn *txn, MDB_val *key,
267 MDB_cursor *cursor, int modify,
268 MDB_pageparent *mpp);
270 static int mdbenv_read_header(MDB_env *env);
271 static int mdb_check_meta_page(MDB_page *p);
272 static int mdbenv_read_meta(MDB_env *env);
273 static int mdbenv_write_meta(MDB_txn *txn);
274 static MDB_page *mdbenv_get_page(MDB_env *env, pgno_t pgno);
276 static MDB_node *mdb_search_node(MDB_db *db, MDB_page *mp,
277 MDB_val *key, int *exactp, unsigned int *kip);
278 static int mdb_add_node(MDB_db *bt, MDB_page *mp,
279 indx_t indx, MDB_val *key, MDB_val *data,
280 pgno_t pgno, uint8_t flags);
281 static void mdb_del_node(MDB_db *bt, MDB_page *mp,
283 static int mdb_read_data(MDB_db *bt, MDB_page *mp,
284 MDB_node *leaf, MDB_val *data);
286 static int mdb_rebalance(MDB_db *bt, MDB_pageparent *mp);
287 static int mdb_update_key(MDB_db *bt, MDB_page *mp,
288 indx_t indx, MDB_val *key);
289 static int mdb_move_node(MDB_db *bt,
290 MDB_pageparent *src, indx_t srcindx,
291 MDB_pageparent *dst, indx_t dstindx);
292 static int mdb_merge(MDB_db *bt, MDB_pageparent *src,
293 MDB_pageparent *dst);
294 static int mdb_split(MDB_db *bt, MDB_page **mpp,
295 unsigned int *newindxp, MDB_val *newkey,
296 MDB_val *newdata, pgno_t newpgno);
297 static MDB_dpage *mdbenv_new_page(MDB_env *env, uint32_t flags, int num);
299 static void cursor_pop_page(MDB_cursor *cursor);
300 static MDB_ppage *cursor_push_page(MDB_cursor *cursor,
303 static int mdb_set_key(MDB_db *bt, MDB_page *mp,
304 MDB_node *node, MDB_val *key);
305 static int mdb_sibling(MDB_cursor *cursor, int move_right);
306 static int mdb_cursor_next(MDB_cursor *cursor,
307 MDB_val *key, MDB_val *data);
308 static int mdb_cursor_set(MDB_cursor *cursor,
309 MDB_val *key, MDB_val *data, int *exactp);
310 static int mdb_cursor_first(MDB_cursor *cursor,
311 MDB_val *key, MDB_val *data);
313 static size_t mdb_leaf_size(MDB_db *bt, MDB_val *key,
315 static size_t mdb_branch_size(MDB_db *bt, MDB_val *key);
317 static pgno_t mdbenv_compact_tree(MDB_env *env, pgno_t pgno,
320 static int memncmp(const void *s1, size_t n1,
321 const void *s2, size_t n2);
322 static int memnrcmp(const void *s1, size_t n1,
323 const void *s2, size_t n2);
326 memncmp(const void *s1, size_t n1, const void *s2, size_t n2)
328 int diff, len_diff = -1;
331 len_diff = (n1 > n2);
334 diff = memcmp(s1, s2, n1);
335 return diff ? diff : len_diff;
339 memnrcmp(const void *s1, size_t n1, const void *s2, size_t n2)
341 const unsigned char *p1, *p2, *p1_lim;
348 p1 = (const unsigned char *)s1 + n1 - 1;
349 p2 = (const unsigned char *)s2 + n2 - 1;
351 for (p1_lim = (n1 <= n2 ? s1 : s2); *p1 == *p2; p1--, p2--) {
353 return (p1 != s1) ? (p1 != p2) : (p2 != s2) ? -1 : 0;
359 mdb_cmp(MDB_db *db, const MDB_val *a, const MDB_val *b)
361 return db->md_cmp(a, b);
365 _mdb_cmp(MDB_db *db, const MDB_val *key1, const MDB_val *key2)
367 if (F_ISSET(db->md_flags, MDB_REVERSEKEY))
368 return memnrcmp(key1->mv_data, key1->mv_size, key2->mv_data, key2->mv_size);
370 return memncmp((char *)key1->mv_data, key1->mv_size, key2->mv_data, key2->mv_size);
373 /* Allocate new page(s) for writing */
375 mdb_newpage(MDB_txn *txn, MDB_page *parent, unsigned int parent_idx, int num)
379 if ((dp = malloc(txn->mt_env->me_meta.mm_stat.ms_psize * num + sizeof(MDB_dhead))) == NULL)
382 dp->h.md_parent = parent;
383 dp->h.md_pi = parent_idx;
384 SIMPLEQ_INSERT_TAIL(txn->mt_u.dirty_queue, dp, h.md_next);
385 dp->p.mp_pgno = txn->mt_next_pgno;
386 txn->mt_next_pgno += num;
391 /* Touch a page: make it dirty and re-insert into tree with updated pgno.
394 mdb_touch(MDB_txn *txn, MDB_pageparent *pp)
396 MDB_page *mp = pp->mp_page;
401 if (!F_ISSET(mp->mp_flags, P_DIRTY)) {
403 DPRINTF("touching page %lu -> %lu", mp->mp_pgno, txn->mt_next_pgno);
404 if ((dp = mdb_newpage(txn, pp->mp_parent, pp->mp_pi, 1)) == NULL)
406 pgno = dp->p.mp_pgno;
407 bcopy(mp, &dp->p, txn->mt_env->me_meta.mm_stat.ms_psize);
410 mp->mp_flags |= P_DIRTY;
412 /* Update the page number to new touched page. */
413 if (pp->mp_parent != NULL)
414 NODEPGNO(NODEPTR(pp->mp_parent, pp->mp_pi)) = mp->mp_pgno;
421 mdbenv_sync(MDB_env *env)
424 if (!F_ISSET(env->me_flags, MDB_NOSYNC)) {
425 if (fsync(env->me_fd))
432 mdb_txn_begin(MDB_env *env, int rdonly, MDB_txn **ret)
437 if ((txn = calloc(1, sizeof(*txn))) == NULL) {
438 DPRINTF("calloc: %s", strerror(errno));
443 txn->mt_flags |= MDB_TXN_RDONLY;
445 txn->mt_u.dirty_queue = calloc(1, sizeof(*txn->mt_u.dirty_queue));
446 if (txn->mt_u.dirty_queue == NULL) {
450 SIMPLEQ_INIT(txn->mt_u.dirty_queue);
452 pthread_mutex_lock(&env->me_txns->mt_wmutex);
453 env->me_txns->mt_txnid++;
455 txn->mt_txnid = env->me_txns->mt_txnid;
457 MDB_reader *r = pthread_getspecific(env->me_txkey);
460 pthread_mutex_lock(&env->me_txns->mt_mutex);
461 for (i=0; i<env->me_maxreaders; i++) {
462 if (env->me_txns->mt_readers[i].mr_pid == 0) {
463 env->me_txns->mt_readers[i].mr_pid = getpid();
464 env->me_txns->mt_readers[i].mr_tid = pthread_self();
465 pthread_setspecific(env->me_txkey, &env->me_txns->mt_readers[i]);
466 if (i >= env->me_txns->mt_numreaders)
467 env->me_txns->mt_numreaders = i+1;
471 pthread_mutex_unlock(&env->me_txns->mt_mutex);
472 if (i == env->me_maxreaders) {
476 r->mr_txnid = txn->mt_txnid;
477 txn->mt_u.reader = r;
484 if ((rc = mdbenv_read_meta(env)) != MDB_SUCCESS) {
489 txn->mt_next_pgno = env->me_meta.mm_last_pg+1;
490 txn->mt_first_pgno = txn->mt_next_pgno;
491 txn->mt_root = env->me_meta.mm_root;
492 DPRINTF("begin transaction %lu on mdbenv %p, root page %lu",
493 txn->mt_txnid, (void *) env, txn->mt_root);
500 mdb_txn_abort(MDB_txn *txn)
509 DPRINTF("abort transaction %lu on mdbenv %p, root page %lu",
510 txn->mt_txnid, (void *) env, txn->mt_root);
512 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
513 txn->mt_u.reader->mr_txnid = 0;
515 /* Discard all dirty pages.
517 while (!SIMPLEQ_EMPTY(txn->mt_u.dirty_queue)) {
518 dp = SIMPLEQ_FIRST(txn->mt_u.dirty_queue);
519 SIMPLEQ_REMOVE_HEAD(txn->mt_u.dirty_queue, h.md_next);
524 DPRINTF("releasing write lock on txn %p", txn);
526 if (flock(txn->bt->fd, LOCK_UN) != 0) {
527 DPRINTF("failed to unlock fd %d: %s",
528 txn->bt->fd, strerror(errno));
531 free(txn->mt_u.dirty_queue);
533 pthread_mutex_unlock(&env->me_txns->mt_wmutex);
540 mdb_txn_commit(MDB_txn *txn)
548 struct iovec iov[MDB_COMMIT_PAGES];
551 assert(txn->mt_env != NULL);
555 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
556 DPRINTF("attempt to commit read-only transaction");
561 if (txn != env->me_txn) {
562 DPRINTF("attempt to commit unknown transaction");
567 if (F_ISSET(txn->mt_flags, MDB_TXN_ERROR)) {
568 DPRINTF("error flag is set, can't commit");
573 if (SIMPLEQ_EMPTY(txn->mt_u.dirty_queue))
576 DPRINTF("committing transaction %lu on mdbenv %p, root page %lu",
577 txn->mt_txnid, (void *) env, txn->mt_root);
579 /* Commit up to MDB_COMMIT_PAGES dirty pages to disk until done.
586 SIMPLEQ_FOREACH(dp, txn->mt_u.dirty_queue, h.md_next) {
587 if (dp->p.mp_pgno != next) {
588 lseek(env->me_fd, dp->p.mp_pgno * env->me_meta.mm_stat.ms_psize, SEEK_SET);
589 next = dp->p.mp_pgno;
593 DPRINTF("committing page %lu", dp->p.mp_pgno);
594 iov[n].iov_len = env->me_meta.mm_stat.ms_psize * dp->h.md_num;
595 iov[n].iov_base = &dp->p;
596 size += iov[n].iov_len;
597 next = dp->p.mp_pgno + dp->h.md_num;
598 /* clear dirty flag */
599 dp->p.mp_flags &= ~P_DIRTY;
600 if (++n >= MDB_COMMIT_PAGES) {
609 DPRINTF("committing %u dirty pages", n);
610 rc = writev(env->me_fd, iov, n);
614 DPRINTF("short write, filesystem full?");
616 DPRINTF("writev: %s", strerror(errno));
621 /* Drop the dirty pages.
623 while (!SIMPLEQ_EMPTY(txn->mt_u.dirty_queue)) {
624 dp = SIMPLEQ_FIRST(txn->mt_u.dirty_queue);
625 SIMPLEQ_REMOVE_HEAD(txn->mt_u.dirty_queue, h.md_next);
632 if ((n = mdbenv_sync(env)) != 0 ||
633 (n = mdbenv_write_meta(txn)) != MDB_SUCCESS ||
634 (n = mdbenv_sync(env)) != 0) {
639 pthread_mutex_unlock(&env->me_txns->mt_wmutex);
640 free(txn->mt_u.dirty_queue);
651 mdbenv_read_header(MDB_env *env)
660 /* We don't know the page size yet, so use a minimum value.
663 if ((rc = pread(env->me_fd, page, PAGESIZE, 0)) == 0) {
665 } else if (rc != PAGESIZE) {
668 DPRINTF("read: %s", strerror(errno));
672 p = (MDB_page *)page;
674 if (!F_ISSET(p->mp_flags, P_META)) {
675 DPRINTF("page %lu not a meta page", p->mp_pgno);
680 if (m->mm_magic != MDB_MAGIC) {
681 DPRINTF("meta has invalid magic");
685 if (m->mm_version != MDB_VERSION) {
686 DPRINTF("database is version %u, expected version %u",
687 m->mm_version, MDB_VERSION);
691 bcopy(m, &env->me_meta, sizeof(*m));
696 mdbenv_init_meta(MDB_env *env)
703 DPRINTF("writing new meta page");
704 psize = sysconf(_SC_PAGE_SIZE);
706 env->me_meta.mm_magic = MDB_MAGIC;
707 env->me_meta.mm_version = MDB_VERSION;
708 env->me_meta.mm_stat.ms_psize = psize;
709 env->me_meta.mm_stat.ms_flags = env->me_flags & 0xffff;
710 env->me_meta.mm_root = P_INVALID;
711 env->me_meta.mm_last_pg = 1;
713 p = calloc(2, psize);
715 p->mp_flags = P_META;
718 bcopy(&env->me_meta, meta, sizeof(*meta));
720 q = (MDB_page *)((char *)p + psize);
723 q->mp_flags = P_META;
726 bcopy(&env->me_meta, meta, sizeof(*meta));
728 rc = write(env->me_fd, p, psize * 2);
730 return (rc == psize * 2) ? MDB_SUCCESS : errno;
734 mdbenv_write_meta(MDB_txn *txn)
742 assert(txn->mt_env != NULL);
744 DPRINTF("writing meta page for root page %lu", txn->mt_root);
748 bcopy(&env->me_meta, &meta, sizeof(meta));
749 meta.mm_root = txn->mt_root;
750 meta.mm_last_pg = txn->mt_next_pgno - 1;
751 meta.mm_txnid = txn->mt_txnid;
754 if (env->me_metatoggle)
755 off = env->me_meta.mm_stat.ms_psize;
758 lseek(env->me_fd, off, SEEK_SET);
759 rc = write(env->me_fd, &meta, sizeof(meta));
760 if (rc != sizeof(meta)) {
761 DPRINTF("write failed, disk error?");
768 /* Returns true if page p is a valid meta page, false otherwise.
771 mdb_check_meta_page(MDB_page *p)
773 if (!F_ISSET(p->mp_flags, P_META)) {
774 DPRINTF("page %lu not a meta page", p->mp_pgno);
782 mdbenv_read_meta(MDB_env *env)
790 if ((mp0 = mdbenv_get_page(env, 0)) == NULL ||
791 (mp1 = mdbenv_get_page(env, 1)) == NULL)
794 rc = mdb_check_meta_page(mp0);
797 rc = mdb_check_meta_page(mp1);
800 meta[0] = METADATA(mp0);
801 meta[1] = METADATA(mp1);
803 if (meta[0]->mm_txnid < meta[1]->mm_txnid)
806 if (meta[toggle]->mm_txnid > env->me_meta.mm_txnid) {
807 bcopy(meta[toggle], &env->me_meta, sizeof(env->me_meta));
808 env->me_metatoggle = toggle;
811 DPRINTF("Using meta page %d", toggle);
817 mdbenv_create(MDB_env **env)
821 e = calloc(1, sizeof(*e));
822 if (!e) return ENOMEM;
824 e->me_meta.mm_mapsize = DEFAULT_MAPSIZE;
825 e->me_maxreaders = DEFAULT_READERS;
832 mdbenv_set_mapsize(MDB_env *env, size_t size)
836 env->me_mapsize = env->me_meta.mm_mapsize = size;
841 mdbenv_set_maxreaders(MDB_env *env, int readers)
843 env->me_maxreaders = readers;
848 mdbenv_get_maxreaders(MDB_env *env, int *readers)
850 if (!env || !readers)
852 *readers = env->me_maxreaders;
857 mdbenv_open2(MDB_env *env, unsigned int flags)
861 env->me_flags = flags;
863 if ((i = mdbenv_read_header(env)) != 0) {
866 DPRINTF("new mdbenv");
870 if (!env->me_mapsize)
871 env->me_mapsize = env->me_meta.mm_mapsize;
874 if (env->me_meta.mm_address && (flags & MDB_FIXEDMAP))
876 env->me_map = mmap(env->me_meta.mm_address, env->me_mapsize, PROT_READ, i,
878 if (env->me_map == MAP_FAILED)
882 env->me_meta.mm_mapsize = env->me_mapsize;
883 if (flags & MDB_FIXEDMAP)
884 env->me_meta.mm_address = env->me_map;
885 i = mdbenv_init_meta(env);
886 if (i != MDB_SUCCESS) {
887 munmap(env->me_map, env->me_mapsize);
892 if ((i = mdbenv_read_meta(env)) != 0)
895 DPRINTF("opened database version %u, pagesize %u",
896 env->me_meta.mm_version, env->me_meta.mm_stat.ms_psize);
897 DPRINTF("depth: %u", env->me_meta.mm_stat.ms_depth);
898 DPRINTF("entries: %lu", env->me_meta.mm_stat.ms_entries);
899 DPRINTF("branch pages: %lu", env->me_meta.mm_stat.ms_branch_pages);
900 DPRINTF("leaf pages: %lu", env->me_meta.mm_stat.ms_leaf_pages);
901 DPRINTF("overflow pages: %lu", env->me_meta.mm_stat.ms_overflow_pages);
902 DPRINTF("root: %lu", env->me_meta.mm_root);
908 mdbenv_reader_dest(void *ptr)
910 MDB_reader *reader = ptr;
912 reader->mr_txnid = 0;
918 mdbenv_open(MDB_env *env, const char *path, unsigned int flags, mode_t mode)
920 int oflags, rc, shmid;
924 if (F_ISSET(flags, MDB_RDONLY))
927 oflags = O_RDWR | O_CREAT;
929 if ((env->me_fd = open(path, oflags, mode)) == -1)
932 env->me_shmkey = ftok(path, 'm');
933 size = (env->me_maxreaders-1) * sizeof(MDB_reader) + sizeof(MDB_txninfo);
934 shmid = shmget(env->me_shmkey, size, IPC_CREAT|IPC_EXCL|mode);
936 if (errno == EEXIST) {
937 shmid = shmget(env->me_shmkey, size, IPC_CREAT|mode);
940 env->me_txns = shmat(shmid, NULL, 0);
941 if (env->me_txns->mt_magic != MDB_MAGIC ||
942 env->me_txns->mt_version != MDB_VERSION) {
943 DPRINTF("invalid lock region %d", shmid);
952 pthread_mutexattr_t mattr;
954 env->me_txns = shmat(shmid, NULL, 0);
955 pthread_mutexattr_init(&mattr);
956 pthread_mutexattr_setpshared(&mattr, PTHREAD_PROCESS_SHARED);
957 pthread_mutex_init(&env->me_txns->mt_mutex, &mattr);
958 pthread_mutex_init(&env->me_txns->mt_wmutex, &mattr);
959 env->me_txns->mt_version = MDB_VERSION;
960 env->me_txns->mt_magic = MDB_MAGIC;
963 if ((rc = mdbenv_open2(env, flags)) != MDB_SUCCESS) {
967 env->me_path = strdup(path);
968 DPRINTF("opened dbenv %p", (void *) env);
971 pthread_key_create(&env->me_txkey, mdbenv_reader_dest);
977 mdbenv_close(MDB_env *env)
985 munmap(env->me_map, env->me_mapsize);
994 /* Search for key within a leaf page, using binary search.
995 * Returns the smallest entry larger or equal to the key.
996 * If exactp is non-null, stores whether the found entry was an exact match
997 * in *exactp (1 or 0).
998 * If kip is non-null, stores the index of the found entry in *kip.
999 * If no entry larger or equal to the key is found, returns NULL.
1002 mdb_search_node(MDB_db *bt, MDB_page *mp, MDB_val *key,
1003 int *exactp, unsigned int *kip)
1011 DPRINTF("searching %u keys in %s page %lu",
1013 IS_LEAF(mp) ? "leaf" : "branch",
1016 assert(NUMKEYS(mp) > 0);
1018 bzero(&nodekey, sizeof(nodekey));
1020 low = IS_LEAF(mp) ? 0 : 1;
1021 high = NUMKEYS(mp) - 1;
1022 while (low <= high) {
1023 i = (low + high) >> 1;
1024 node = NODEPTR(mp, i);
1026 nodekey.mv_size = node->mn_ksize;
1027 nodekey.mv_data = NODEKEY(node);
1030 rc = bt->md_cmp(key, &nodekey);
1032 rc = _mdb_cmp(bt, key, &nodekey);
1035 DPRINTF("found leaf index %u [%.*s], rc = %i",
1036 i, (int)nodekey.mv_size, (char *)nodekey.mv_data, rc);
1038 DPRINTF("found branch index %u [%.*s -> %lu], rc = %i",
1039 i, (int)node->mn_ksize, (char *)NODEKEY(node),
1050 if (rc > 0) { /* Found entry is less than the key. */
1051 i++; /* Skip to get the smallest entry larger than key. */
1052 if (i >= NUMKEYS(mp))
1053 /* There is no entry larger or equal to the key. */
1057 *exactp = (rc == 0);
1058 if (kip) /* Store the key index if requested. */
1061 return NODEPTR(mp, i);
1065 cursor_pop_page(MDB_cursor *cursor)
1069 top = CURSOR_TOP(cursor);
1072 DPRINTF("popped page %lu off cursor %p", top->mp_page->mp_pgno, (void *) cursor);
1078 cursor_push_page(MDB_cursor *cursor, MDB_page *mp)
1082 DPRINTF("pushing page %lu on cursor %p", mp->mp_pgno, (void *) cursor);
1084 if ((ppage = calloc(1, sizeof(*ppage))) == NULL)
1086 ppage->mp_page = mp;
1087 CURSOR_PUSH(cursor, ppage);
1092 mdbenv_get_page(MDB_env *env, pgno_t pgno)
1095 MDB_txn *txn = env->me_txn;
1097 if (txn && pgno >= txn->mt_first_pgno &&
1098 !SIMPLEQ_EMPTY(txn->mt_u.dirty_queue)) {
1100 SIMPLEQ_FOREACH(dp, txn->mt_u.dirty_queue, h.md_next) {
1101 if (dp->p.mp_pgno == pgno) {
1107 p = (MDB_page *)(env->me_map + env->me_meta.mm_stat.ms_psize * pgno);
1113 mdb_search_page_root(MDB_db *bt, MDB_val *key,
1114 MDB_cursor *cursor, int modify, MDB_pageparent *mpp)
1116 MDB_page *mp = mpp->mp_page;
1119 if (cursor && cursor_push_page(cursor, mp) == NULL)
1122 while (IS_BRANCH(mp)) {
1126 DPRINTF("branch page %lu has %u keys", mp->mp_pgno, NUMKEYS(mp));
1127 assert(NUMKEYS(mp) > 1);
1128 DPRINTF("found index 0 to page %lu", NODEPGNO(NODEPTR(mp, 0)));
1130 if (key == NULL) /* Initialize cursor to first page. */
1134 node = mdb_search_node(bt, mp, key, &exact, &i);
1136 i = NUMKEYS(mp) - 1;
1144 DPRINTF("following index %u for key %.*s",
1145 i, (int)key->mv_size, (char *)key->mv_data);
1146 assert(i < NUMKEYS(mp));
1147 node = NODEPTR(mp, i);
1150 CURSOR_TOP(cursor)->mp_ki = i;
1152 mpp->mp_parent = mp;
1153 if ((mp = mdbenv_get_page(bt->md_env, NODEPGNO(node))) == NULL)
1158 if (cursor && cursor_push_page(cursor, mp) == NULL)
1162 MDB_dhead *dh = ((MDB_dhead *)mp)-1;
1163 if ((rc = mdb_touch(bt->md_env->me_txn, mpp)) != 0)
1165 dh = ((MDB_dhead *)mpp->mp_page)-1;
1166 dh->md_parent = mpp->mp_parent;
1167 dh->md_pi = mpp->mp_pi;
1174 DPRINTF("internal error, index points to a %02X page!?",
1179 DPRINTF("found leaf page %lu for key %.*s", mp->mp_pgno,
1180 key ? (int)key->mv_size : 0, key ? (char *)key->mv_data : NULL);
1185 /* Search for the page a given key should be in.
1186 * Stores a pointer to the found page in *mpp.
1187 * If key is NULL, search for the lowest page (used by mdb_cursor_first).
1188 * If cursor is non-null, pushes parent pages on the cursor stack.
1189 * If modify is true, visited pages are updated with new page numbers.
1192 mdb_search_page(MDB_db *db, MDB_txn *txn, MDB_val *key,
1193 MDB_cursor *cursor, int modify, MDB_pageparent *mpp)
1198 /* Can't modify pages outside a transaction. */
1199 if (txn == NULL && modify) {
1203 /* Choose which root page to start with. If a transaction is given
1204 * use the root page from the transaction, otherwise read the last
1205 * committed root page.
1208 if ((rc = mdbenv_read_meta(db->md_env)) != MDB_SUCCESS)
1210 root = db->md_env->me_meta.mm_root;
1211 } else if (F_ISSET(txn->mt_flags, MDB_TXN_ERROR)) {
1212 DPRINTF("transaction has failed, must abort");
1215 root = txn->mt_root;
1217 if (root == P_INVALID) { /* Tree is empty. */
1218 DPRINTF("tree is empty");
1222 if ((mpp->mp_page = mdbenv_get_page(db->md_env, root)) == NULL)
1225 DPRINTF("root page has flags 0x%X", mpp->mp_page->mp_flags);
1227 if (modify && !F_ISSET(mpp->mp_page->mp_flags, P_DIRTY)) {
1228 mpp->mp_parent = NULL;
1230 if ((rc = mdb_touch(txn, mpp)))
1232 txn->mt_root = mpp->mp_page->mp_pgno;
1235 return mdb_search_page_root(db, key, cursor, modify, mpp);
1239 mdb_read_data(MDB_db *db, MDB_page *mp, MDB_node *leaf,
1242 MDB_page *omp; /* overflow mpage */
1246 bzero(data, sizeof(*data));
1247 max = db->md_env->me_meta.mm_stat.ms_psize - PAGEHDRSZ;
1249 if (!F_ISSET(leaf->mn_flags, F_BIGDATA)) {
1250 data->mv_size = leaf->mn_dsize;
1251 data->mv_data = NODEDATA(leaf);
1255 /* Read overflow data.
1257 data->mv_size = leaf->mn_dsize;
1258 bcopy(NODEDATA(leaf), &pgno, sizeof(pgno));
1259 if ((omp = mdbenv_get_page(db->md_env, pgno)) == NULL) {
1260 DPRINTF("read overflow page %lu failed", pgno);
1263 data->mv_data = omp;
1269 mdb_get(MDB_db *db, MDB_txn *txn,
1270 MDB_val *key, MDB_val *data)
1278 DPRINTF("===> get key [%.*s]", (int)key->mv_size, (char *)key->mv_data);
1283 if (txn != NULL && db->md_env != txn->mt_env) {
1287 if (key->mv_size == 0 || key->mv_size > MAXKEYSIZE) {
1291 if ((rc = mdb_search_page(db, txn, key, NULL, 0, &mpp)) != MDB_SUCCESS)
1294 leaf = mdb_search_node(db, mpp.mp_page, key, &exact, NULL);
1296 rc = mdb_read_data(db, mpp.mp_page, leaf, data);
1305 mdb_sibling(MDB_cursor *cursor, int move_right)
1309 MDB_ppage *parent, *top;
1312 top = CURSOR_TOP(cursor);
1313 if ((parent = SLIST_NEXT(top, mp_entry)) == NULL) {
1314 return ENOENT; /* root has no siblings */
1317 DPRINTF("parent page is page %lu, index %u",
1318 parent->mp_page->mp_pgno, parent->mp_ki);
1320 cursor_pop_page(cursor);
1321 if (move_right ? (parent->mp_ki + 1 >= NUMKEYS(parent->mp_page))
1322 : (parent->mp_ki == 0)) {
1323 DPRINTF("no more keys left, moving to %s sibling",
1324 move_right ? "right" : "left");
1325 if ((rc = mdb_sibling(cursor, move_right)) != MDB_SUCCESS)
1327 parent = CURSOR_TOP(cursor);
1333 DPRINTF("just moving to %s index key %u",
1334 move_right ? "right" : "left", parent->mp_ki);
1336 assert(IS_BRANCH(parent->mp_page));
1338 indx = NODEPTR(parent->mp_page, parent->mp_ki);
1339 if ((mp = mdbenv_get_page(cursor->mc_db->md_env, indx->mn_pgno)) == NULL)
1342 mp->parent = parent->mp_page;
1343 mp->parent_index = parent->mp_ki;
1346 cursor_push_page(cursor, mp);
1352 mdb_set_key(MDB_db *bt, MDB_page *mp, MDB_node *node,
1358 key->mv_size = node->mn_ksize;
1359 key->mv_data = NODEKEY(node);
1365 mdb_cursor_next(MDB_cursor *cursor, MDB_val *key, MDB_val *data)
1371 if (cursor->mc_eof) {
1375 assert(cursor->mc_initialized);
1377 top = CURSOR_TOP(cursor);
1380 DPRINTF("cursor_next: top page is %lu in cursor %p", mp->mp_pgno, (void *) cursor);
1382 if (top->mp_ki + 1 >= NUMKEYS(mp)) {
1383 DPRINTF("=====> move to next sibling page");
1384 if (mdb_sibling(cursor, 1) != MDB_SUCCESS) {
1388 top = CURSOR_TOP(cursor);
1390 DPRINTF("next page is %lu, key index %u", mp->mp_pgno, top->mp_ki);
1394 DPRINTF("==> cursor points to page %lu with %u keys, key index %u",
1395 mp->mp_pgno, NUMKEYS(mp), top->mp_ki);
1397 assert(IS_LEAF(mp));
1398 leaf = NODEPTR(mp, top->mp_ki);
1400 if (data && mdb_read_data(cursor->mc_db, mp, leaf, data) != MDB_SUCCESS)
1403 return mdb_set_key(cursor->mc_db, mp, leaf, key);
1407 mdb_cursor_set(MDB_cursor *cursor, MDB_val *key, MDB_val *data,
1417 assert(key->mv_size > 0);
1419 rc = mdb_search_page(cursor->mc_db, cursor->mc_txn, key, cursor, 0, &mpp);
1420 if (rc != MDB_SUCCESS)
1422 assert(IS_LEAF(mpp.mp_page));
1424 top = CURSOR_TOP(cursor);
1425 leaf = mdb_search_node(cursor->mc_db, mpp.mp_page, key, exactp, &top->mp_ki);
1426 if (exactp != NULL && !*exactp) {
1427 /* MDB_CURSOR_EXACT specified and not an exact match. */
1432 DPRINTF("===> inexact leaf not found, goto sibling");
1433 if ((rc = mdb_sibling(cursor, 1)) != MDB_SUCCESS)
1434 return rc; /* no entries matched */
1435 top = CURSOR_TOP(cursor);
1437 mpp.mp_page = top->mp_page;
1438 assert(IS_LEAF(mpp.mp_page));
1439 leaf = NODEPTR(mpp.mp_page, 0);
1442 cursor->mc_initialized = 1;
1445 if (data && (rc = mdb_read_data(cursor->mc_db, mpp.mp_page, leaf, data)) != MDB_SUCCESS)
1448 rc = mdb_set_key(cursor->mc_db, mpp.mp_page, leaf, key);
1449 if (rc == MDB_SUCCESS) {
1450 DPRINTF("==> cursor placed on key %.*s",
1451 (int)key->mv_size, (char *)key->mv_data);
1459 mdb_cursor_first(MDB_cursor *cursor, MDB_val *key, MDB_val *data)
1465 rc = mdb_search_page(cursor->mc_db, cursor->mc_txn, NULL, cursor, 0, &mpp);
1466 if (rc != MDB_SUCCESS)
1468 assert(IS_LEAF(mpp.mp_page));
1470 leaf = NODEPTR(mpp.mp_page, 0);
1471 cursor->mc_initialized = 1;
1474 if (data && (rc = mdb_read_data(cursor->mc_db, mpp.mp_page, leaf, data)) != MDB_SUCCESS)
1477 return mdb_set_key(cursor->mc_db, mpp.mp_page, leaf, key);
1481 mdb_cursor_get(MDB_cursor *cursor, MDB_val *key, MDB_val *data,
1491 case MDB_CURSOR_EXACT:
1492 while (CURSOR_TOP(cursor) != NULL)
1493 cursor_pop_page(cursor);
1494 if (key == NULL || key->mv_size == 0 || key->mv_size > MAXKEYSIZE) {
1496 } else if (op == MDB_CURSOR_EXACT)
1497 rc = mdb_cursor_set(cursor, key, data, &exact);
1499 rc = mdb_cursor_set(cursor, key, data, NULL);
1502 if (!cursor->mc_initialized)
1503 rc = mdb_cursor_first(cursor, key, data);
1505 rc = mdb_cursor_next(cursor, key, data);
1508 while (CURSOR_TOP(cursor) != NULL)
1509 cursor_pop_page(cursor);
1510 rc = mdb_cursor_first(cursor, key, data);
1513 DPRINTF("unhandled/unimplemented cursor operation %u", op);
1521 /* Allocate a page and initialize it
1524 mdbenv_new_page(MDB_env *env, uint32_t flags, int num)
1528 assert(env != NULL);
1529 assert(env->me_txn != NULL);
1531 DPRINTF("allocating new mpage %lu, page size %u",
1532 env->me_txn->mt_next_pgno, env->me_meta.mm_stat.ms_psize);
1533 if ((dp = mdb_newpage(env->me_txn, NULL, 0, num)) == NULL)
1535 dp->p.mp_flags = flags | P_DIRTY;
1536 dp->p.mp_lower = PAGEHDRSZ;
1537 dp->p.mp_upper = env->me_meta.mm_stat.ms_psize;
1539 if (IS_BRANCH(&dp->p))
1540 env->me_meta.mm_stat.ms_branch_pages++;
1541 else if (IS_LEAF(&dp->p))
1542 env->me_meta.mm_stat.ms_leaf_pages++;
1543 else if (IS_OVERFLOW(&dp->p)) {
1544 env->me_meta.mm_stat.ms_overflow_pages += num;
1545 dp->p.mp_pages = num;
1552 mdb_leaf_size(MDB_db *db, MDB_val *key, MDB_val *data)
1556 sz = LEAFSIZE(key, data);
1557 if (data->mv_size >= db->md_env->me_meta.mm_stat.ms_psize / MDB_MINKEYS) {
1558 /* put on overflow page */
1559 sz -= data->mv_size - sizeof(pgno_t);
1562 return sz + sizeof(indx_t);
1566 mdb_branch_size(MDB_db *db, MDB_val *key)
1571 if (sz >= db->md_env->me_meta.mm_stat.ms_psize / MDB_MINKEYS) {
1572 /* put on overflow page */
1573 /* not implemented */
1574 /* sz -= key->size - sizeof(pgno_t); */
1577 return sz + sizeof(indx_t);
1581 mdb_add_node(MDB_db *db, MDB_page *mp, indx_t indx,
1582 MDB_val *key, MDB_val *data, pgno_t pgno, uint8_t flags)
1585 size_t node_size = NODESIZE;
1588 MDB_dpage *ofp = NULL; /* overflow page */
1590 assert(mp->mp_upper >= mp->mp_lower);
1592 DPRINTF("add node [%.*s] to %s page %lu at index %i, key size %zu",
1593 key ? (int)key->mv_size : 0, key ? (char *)key->mv_data : NULL,
1594 IS_LEAF(mp) ? "leaf" : "branch",
1595 mp->mp_pgno, indx, key ? key->mv_size : 0);
1598 node_size += key->mv_size;
1602 if (F_ISSET(flags, F_BIGDATA)) {
1603 /* Data already on overflow page. */
1604 node_size += sizeof(pgno_t);
1605 } else if (data->mv_size >= db->md_env->me_meta.mm_stat.ms_psize / MDB_MINKEYS) {
1606 int ovpages = PAGEHDRSZ + data->mv_size + db->md_env->me_meta.mm_stat.ms_psize - 1;
1607 ovpages /= db->md_env->me_meta.mm_stat.ms_psize;
1608 /* Put data on overflow page. */
1609 DPRINTF("data size is %zu, put on overflow page",
1611 node_size += sizeof(pgno_t);
1612 if ((ofp = mdbenv_new_page(db->md_env, P_OVERFLOW, ovpages)) == NULL)
1614 DPRINTF("allocated overflow page %lu", ofp->p.mp_pgno);
1617 node_size += data->mv_size;
1621 if (node_size + sizeof(indx_t) > SIZELEFT(mp)) {
1622 DPRINTF("not enough room in page %lu, got %u ptrs",
1623 mp->mp_pgno, NUMKEYS(mp));
1624 DPRINTF("upper - lower = %u - %u = %u", mp->mp_upper, mp->mp_lower,
1625 mp->mp_upper - mp->mp_lower);
1626 DPRINTF("node size = %zu", node_size);
1630 /* Move higher pointers up one slot. */
1631 for (i = NUMKEYS(mp); i > indx; i--)
1632 mp->mp_ptrs[i] = mp->mp_ptrs[i - 1];
1634 /* Adjust free space offsets. */
1635 ofs = mp->mp_upper - node_size;
1636 assert(ofs >= mp->mp_lower + sizeof(indx_t));
1637 mp->mp_ptrs[indx] = ofs;
1639 mp->mp_lower += sizeof(indx_t);
1641 /* Write the node data. */
1642 node = NODEPTR(mp, indx);
1643 node->mn_ksize = (key == NULL) ? 0 : key->mv_size;
1644 node->mn_flags = flags;
1646 node->mn_dsize = data->mv_size;
1648 node->mn_pgno = pgno;
1651 bcopy(key->mv_data, NODEKEY(node), key->mv_size);
1656 if (F_ISSET(flags, F_BIGDATA))
1657 bcopy(data->mv_data, node->mn_data + key->mv_size,
1660 bcopy(data->mv_data, node->mn_data + key->mv_size,
1663 bcopy(&ofp->p.mp_pgno, node->mn_data + key->mv_size,
1665 bcopy(data->mv_data, METADATA(&ofp->p), data->mv_size);
1673 mdb_del_node(MDB_db *db, MDB_page *mp, indx_t indx)
1676 indx_t i, j, numkeys, ptr;
1680 DPRINTF("delete node %u on %s page %lu", indx,
1681 IS_LEAF(mp) ? "leaf" : "branch", mp->mp_pgno);
1682 assert(indx < NUMKEYS(mp));
1684 node = NODEPTR(mp, indx);
1685 sz = NODESIZE + node->mn_ksize;
1687 if (F_ISSET(node->mn_flags, F_BIGDATA))
1688 sz += sizeof(pgno_t);
1690 sz += NODEDSZ(node);
1693 ptr = mp->mp_ptrs[indx];
1694 numkeys = NUMKEYS(mp);
1695 for (i = j = 0; i < numkeys; i++) {
1697 mp->mp_ptrs[j] = mp->mp_ptrs[i];
1698 if (mp->mp_ptrs[i] < ptr)
1699 mp->mp_ptrs[j] += sz;
1704 base = (char *)mp + mp->mp_upper;
1705 bcopy(base, base + sz, ptr - mp->mp_upper);
1707 mp->mp_lower -= sizeof(indx_t);
1712 mdb_cursor_open(MDB_db *db, MDB_txn *txn, MDB_cursor **ret)
1716 if (db == NULL || ret == NULL)
1719 if (txn != NULL && db->md_env != txn->mt_env) {
1723 if ((cursor = calloc(1, sizeof(*cursor))) != NULL) {
1724 SLIST_INIT(&cursor->mc_stack);
1726 cursor->mc_txn = txn;
1735 mdb_cursor_close(MDB_cursor *cursor)
1737 if (cursor != NULL) {
1738 while (!CURSOR_EMPTY(cursor))
1739 cursor_pop_page(cursor);
1741 /* btree_close(cursor->bt); */
1747 mdb_update_key(MDB_db *db, MDB_page *mp, indx_t indx,
1750 indx_t ptr, i, numkeys;
1756 node = NODEPTR(mp, indx);
1757 ptr = mp->mp_ptrs[indx];
1758 DPRINTF("update key %u (ofs %u) [%.*s] to [%.*s] on page %lu",
1760 (int)node->mn_ksize, (char *)NODEKEY(node),
1761 (int)key->mv_size, (char *)key->mv_data,
1764 delta = key->mv_size - node->mn_ksize;
1766 if (delta > 0 && SIZELEFT(mp) < delta) {
1767 DPRINTF("OUCH! Not enough room, delta = %d", delta);
1771 numkeys = NUMKEYS(mp);
1772 for (i = 0; i < numkeys; i++) {
1773 if (mp->mp_ptrs[i] <= ptr)
1774 mp->mp_ptrs[i] -= delta;
1777 base = (char *)mp + mp->mp_upper;
1778 len = ptr - mp->mp_upper + NODESIZE;
1779 bcopy(base, base - delta, len);
1780 mp->mp_upper -= delta;
1782 node = NODEPTR(mp, indx);
1783 node->mn_ksize = key->mv_size;
1786 bcopy(key->mv_data, NODEKEY(node), key->mv_size);
1791 /* Move a node from src to dst.
1794 mdb_move_node(MDB_db *bt, MDB_pageparent *src, indx_t srcindx,
1795 MDB_pageparent *dst, indx_t dstindx)
1801 srcnode = NODEPTR(src->mp_page, srcindx);
1802 DPRINTF("moving %s node %u [%.*s] on page %lu to node %u on page %lu",
1803 IS_LEAF(src->mp_page) ? "leaf" : "branch",
1805 (int)srcnode->mn_ksize, (char *)NODEKEY(srcnode),
1806 src->mp_page->mp_pgno,
1807 dstindx, dst->mp_page->mp_pgno);
1809 /* Mark src and dst as dirty. */
1810 if ((rc = mdb_touch(bt->md_env->me_txn, src)) ||
1811 (rc = mdb_touch(bt->md_env->me_txn, dst)))
1814 /* Add the node to the destination page.
1816 key.mv_size = srcnode->mn_ksize;
1817 key.mv_data = NODEKEY(srcnode);
1818 data.mv_size = NODEDSZ(srcnode);
1819 data.mv_data = NODEDATA(srcnode);
1820 rc = mdb_add_node(bt, dst->mp_page, dstindx, &key, &data, NODEPGNO(srcnode),
1822 if (rc != MDB_SUCCESS)
1825 /* Delete the node from the source page.
1827 mdb_del_node(bt, src->mp_page, srcindx);
1829 /* Update the parent separators.
1831 if (srcindx == 0 && src->mp_pi != 0) {
1832 DPRINTF("update separator for source page %lu to [%.*s]",
1833 src->mp_page->mp_pgno, (int)key.mv_size, (char *)key.mv_data);
1834 if ((rc = mdb_update_key(bt, src->mp_parent, src->mp_pi,
1835 &key)) != MDB_SUCCESS)
1839 if (srcindx == 0 && IS_BRANCH(src->mp_page)) {
1841 nullkey.mv_size = 0;
1842 assert(mdb_update_key(bt, src->mp_page, 0, &nullkey) == MDB_SUCCESS);
1845 if (dstindx == 0 && dst->mp_pi != 0) {
1846 DPRINTF("update separator for destination page %lu to [%.*s]",
1847 dst->mp_page->mp_pgno, (int)key.mv_size, (char *)key.mv_data);
1848 if ((rc = mdb_update_key(bt, dst->mp_parent, dst->mp_pi,
1849 &key)) != MDB_SUCCESS)
1853 if (dstindx == 0 && IS_BRANCH(dst->mp_page)) {
1855 nullkey.mv_size = 0;
1856 assert(mdb_update_key(bt, dst->mp_page, 0, &nullkey) == MDB_SUCCESS);
1863 mdb_merge(MDB_db *bt, MDB_pageparent *src, MDB_pageparent *dst)
1872 DPRINTF("merging page %lu and %lu", src->mp_page->mp_pgno, dst->mp_page->mp_pgno);
1874 assert(src->mp_parent); /* can't merge root page */
1875 assert(dst->mp_parent);
1876 assert(bt->md_env->me_txn != NULL);
1878 /* Mark src and dst as dirty. */
1879 if ((rc = mdb_touch(bt->md_env->me_txn, src)) ||
1880 (rc = mdb_touch(bt->md_env->me_txn, dst)))
1883 /* Move all nodes from src to dst.
1885 for (i = 0; i < NUMKEYS(src->mp_page); i++) {
1886 srcnode = NODEPTR(src->mp_page, i);
1888 key.mv_size = srcnode->mn_ksize;
1889 key.mv_data = NODEKEY(srcnode);
1890 data.mv_size = NODEDSZ(srcnode);
1891 data.mv_data = NODEDATA(srcnode);
1892 rc = mdb_add_node(bt, dst->mp_page, NUMKEYS(dst->mp_page), &key,
1893 &data, NODEPGNO(srcnode), srcnode->mn_flags);
1894 if (rc != MDB_SUCCESS)
1898 DPRINTF("dst page %lu now has %u keys (%.1f%% filled)",
1899 dst->mp_page->mp_pgno, NUMKEYS(dst->mp_page), (float)PAGEFILL(bt->md_env, dst->mp_page) / 10);
1901 /* Unlink the src page from parent.
1903 mdb_del_node(bt, src->mp_parent, src->mp_pi);
1904 if (src->mp_pi == 0) {
1906 if ((rc = mdb_update_key(bt, src->mp_parent, 0, &key)) != MDB_SUCCESS)
1910 if (IS_LEAF(src->mp_page))
1911 bt->md_env->me_meta.mm_stat.ms_leaf_pages--;
1913 bt->md_env->me_meta.mm_stat.ms_branch_pages--;
1915 mpp.mp_page = src->mp_parent;
1916 dh = (MDB_dhead *)src->mp_parent;
1918 mpp.mp_parent = dh->md_parent;
1919 mpp.mp_pi = dh->md_pi;
1921 return mdb_rebalance(bt, &mpp);
1924 #define FILL_THRESHOLD 250
1927 mdb_rebalance(MDB_db *db, MDB_pageparent *mpp)
1932 indx_t si = 0, di = 0;
1935 assert(db->md_env->me_txn != NULL);
1936 assert(mpp != NULL);
1938 DPRINTF("rebalancing %s page %lu (has %u keys, %.1f%% full)",
1939 IS_LEAF(mpp->mp_page) ? "leaf" : "branch",
1940 mpp->mp_page->mp_pgno, NUMKEYS(mpp->mp_page), (float)PAGEFILL(db->md_env, mpp->mp_page) / 10);
1942 if (PAGEFILL(db->md_env, mpp->mp_page) >= FILL_THRESHOLD) {
1943 DPRINTF("no need to rebalance page %lu, above fill threshold",
1944 mpp->mp_page->mp_pgno);
1948 if (mpp->mp_parent == NULL) {
1949 if (NUMKEYS(mpp->mp_page) == 0) {
1950 DPRINTF("tree is completely empty");
1951 db->md_env->me_txn->mt_root = P_INVALID;
1952 db->md_env->me_meta.mm_stat.ms_depth--;
1953 db->md_env->me_meta.mm_stat.ms_leaf_pages--;
1954 } else if (IS_BRANCH(mpp->mp_page) && NUMKEYS(mpp->mp_page) == 1) {
1955 DPRINTF("collapsing root page!");
1956 db->md_env->me_txn->mt_root = NODEPGNO(NODEPTR(mpp->mp_page, 0));
1957 if ((root = mdbenv_get_page(db->md_env, db->md_env->me_txn->mt_root)) == NULL)
1959 db->md_env->me_meta.mm_stat.ms_depth--;
1960 db->md_env->me_meta.mm_stat.ms_branch_pages--;
1962 DPRINTF("root page doesn't need rebalancing");
1966 /* The parent (branch page) must have at least 2 pointers,
1967 * otherwise the tree is invalid.
1969 assert(NUMKEYS(mpp->mp_parent) > 1);
1971 /* Leaf page fill factor is below the threshold.
1972 * Try to move keys from left or right neighbor, or
1973 * merge with a neighbor page.
1978 if (mpp->mp_pi == 0) {
1979 /* We're the leftmost leaf in our parent.
1981 DPRINTF("reading right neighbor");
1982 node = NODEPTR(mpp->mp_parent, mpp->mp_pi + 1);
1983 if ((npp.mp_page = mdbenv_get_page(db->md_env, NODEPGNO(node))) == NULL)
1985 npp.mp_pi = mpp->mp_pi + 1;
1987 di = NUMKEYS(mpp->mp_page);
1989 /* There is at least one neighbor to the left.
1991 DPRINTF("reading left neighbor");
1992 node = NODEPTR(mpp->mp_parent, mpp->mp_pi - 1);
1993 if ((npp.mp_page = mdbenv_get_page(db->md_env, NODEPGNO(node))) == NULL)
1995 npp.mp_pi = mpp->mp_pi - 1;
1996 si = NUMKEYS(npp.mp_page) - 1;
1999 npp.mp_parent = mpp->mp_parent;
2001 DPRINTF("found neighbor page %lu (%u keys, %.1f%% full)",
2002 npp.mp_page->mp_pgno, NUMKEYS(npp.mp_page), (float)PAGEFILL(db->md_env, npp.mp_page) / 10);
2004 /* If the neighbor page is above threshold and has at least two
2005 * keys, move one key from it.
2007 * Otherwise we should try to merge them.
2009 if (PAGEFILL(db->md_env, npp.mp_page) >= FILL_THRESHOLD && NUMKEYS(npp.mp_page) >= 2)
2010 return mdb_move_node(db, &npp, si, mpp, di);
2011 else { /* FIXME: if (has_enough_room()) */
2012 if (mpp->mp_pi == 0)
2013 return mdb_merge(db, &npp, mpp);
2015 return mdb_merge(db, mpp, &npp);
2020 mdb_del(MDB_db *bt, MDB_txn *txn,
2021 MDB_val *key, MDB_val *data)
2028 DPRINTF("========> delete key %.*s", (int)key->mv_size, (char *)key->mv_data);
2030 assert(key != NULL);
2032 if (bt == NULL || txn == NULL)
2035 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
2039 if (bt->md_env->me_txn != txn) {
2043 if (key->mv_size == 0 || key->mv_size > MAXKEYSIZE) {
2047 if ((rc = mdb_search_page(bt, txn, key, NULL, 1, &mpp)) != MDB_SUCCESS)
2050 leaf = mdb_search_node(bt, mpp.mp_page, key, &exact, &ki);
2051 if (leaf == NULL || !exact) {
2055 if (data && (rc = mdb_read_data(bt, NULL, leaf, data)) != MDB_SUCCESS)
2058 mdb_del_node(bt, mpp.mp_page, ki);
2059 bt->md_env->me_meta.mm_stat.ms_entries--;
2060 rc = mdb_rebalance(bt, &mpp);
2061 if (rc != MDB_SUCCESS)
2062 txn->mt_flags |= MDB_TXN_ERROR;
2067 /* Split page <*mpp>, and insert <key,(data|newpgno)> in either left or
2068 * right sibling, at index <*newindxp> (as if unsplit). Updates *mpp and
2069 * *newindxp with the actual values after split, ie if *mpp and *newindxp
2070 * refer to a node in the new right sibling page.
2073 mdb_split(MDB_db *bt, MDB_page **mpp, unsigned int *newindxp,
2074 MDB_val *newkey, MDB_val *newdata, pgno_t newpgno)
2077 int rc = MDB_SUCCESS, ins_new = 0;
2080 unsigned int i, j, split_indx;
2082 MDB_val sepkey, rkey, rdata;
2084 MDB_dpage *mdp, *rdp, *pdp;
2088 assert(bt->md_env != NULL);
2090 dh = ((MDB_dhead *)*mpp) - 1;
2091 mdp = (MDB_dpage *)dh;
2092 newindx = *newindxp;
2094 DPRINTF("-----> splitting %s page %lu and adding [%.*s] at index %i",
2095 IS_LEAF(&mdp->p) ? "leaf" : "branch", mdp->p.mp_pgno,
2096 (int)newkey->mv_size, (char *)newkey->mv_data, *newindxp);
2098 if (mdp->h.md_parent == NULL) {
2099 if ((pdp = mdbenv_new_page(bt->md_env, P_BRANCH, 1)) == NULL)
2102 mdp->h.md_parent = &pdp->p;
2103 bt->md_env->me_txn->mt_root = pdp->p.mp_pgno;
2104 DPRINTF("root split! new root = %lu", pdp->p.mp_pgno);
2105 bt->md_env->me_meta.mm_stat.ms_depth++;
2107 /* Add left (implicit) pointer. */
2108 if (mdb_add_node(bt, &pdp->p, 0, NULL, NULL,
2109 mdp->p.mp_pgno, 0) != MDB_SUCCESS)
2112 DPRINTF("parent branch page is %lu", mdp->h.md_parent->mp_pgno);
2115 /* Create a right sibling. */
2116 if ((rdp = mdbenv_new_page(bt->md_env, mdp->p.mp_flags, 1)) == NULL)
2118 rdp->h.md_parent = mdp->h.md_parent;
2119 rdp->h.md_pi = mdp->h.md_pi + 1;
2120 DPRINTF("new right sibling: page %lu", rdp->p.mp_pgno);
2122 /* Move half of the keys to the right sibling. */
2123 if ((copy = malloc(bt->md_env->me_meta.mm_stat.ms_psize)) == NULL)
2125 bcopy(&mdp->p, copy, bt->md_env->me_meta.mm_stat.ms_psize);
2126 bzero(&mdp->p.mp_ptrs, bt->md_env->me_meta.mm_stat.ms_psize - PAGEHDRSZ);
2127 mdp->p.mp_lower = PAGEHDRSZ;
2128 mdp->p.mp_upper = bt->md_env->me_meta.mm_stat.ms_psize;
2130 split_indx = NUMKEYS(copy) / 2 + 1;
2132 /* First find the separating key between the split pages.
2134 bzero(&sepkey, sizeof(sepkey));
2135 if (newindx == split_indx) {
2136 sepkey.mv_size = newkey->mv_size;
2137 sepkey.mv_data = newkey->mv_data;
2139 node = NODEPTR(copy, split_indx);
2140 sepkey.mv_size = node->mn_ksize;
2141 sepkey.mv_data = NODEKEY(node);
2144 DPRINTF("separator is [%.*s]", (int)sepkey.mv_size, (char *)sepkey.mv_data);
2146 /* Copy separator key to the parent.
2148 if (SIZELEFT(rdp->h.md_parent) < mdb_branch_size(bt, &sepkey)) {
2149 rc = mdb_split(bt, &rdp->h.md_parent, &rdp->h.md_pi,
2150 &sepkey, NULL, rdp->p.mp_pgno);
2152 /* Right page might now have changed parent.
2153 * Check if left page also changed parent.
2155 if (rdp->h.md_parent != mdp->h.md_parent &&
2156 mdp->h.md_pi >= NUMKEYS(mdp->h.md_parent)) {
2157 mdp->h.md_parent = rdp->h.md_parent;
2158 mdp->h.md_pi = rdp->h.md_pi - 1;
2161 rc = mdb_add_node(bt, rdp->h.md_parent, rdp->h.md_pi,
2162 &sepkey, NULL, rdp->p.mp_pgno, 0);
2164 if (rc != MDB_SUCCESS) {
2169 for (i = j = 0; i <= NUMKEYS(copy); j++) {
2170 if (i < split_indx) {
2171 /* Re-insert in left sibling. */
2174 /* Insert in right sibling. */
2175 if (i == split_indx)
2176 /* Reset insert index for right sibling. */
2177 j = (i == newindx && ins_new);
2181 if (i == newindx && !ins_new) {
2182 /* Insert the original entry that caused the split. */
2183 rkey.mv_data = newkey->mv_data;
2184 rkey.mv_size = newkey->mv_size;
2185 if (IS_LEAF(&mdp->p)) {
2186 rdata.mv_data = newdata->mv_data;
2187 rdata.mv_size = newdata->mv_size;
2194 /* Update page and index for the new key. */
2197 } else if (i == NUMKEYS(copy)) {
2200 node = NODEPTR(copy, i);
2201 rkey.mv_data = NODEKEY(node);
2202 rkey.mv_size = node->mn_ksize;
2203 if (IS_LEAF(&mdp->p)) {
2204 rdata.mv_data = NODEDATA(node);
2205 rdata.mv_size = node->mn_dsize;
2207 pgno = node->mn_pgno;
2208 flags = node->mn_flags;
2213 if (!IS_LEAF(&mdp->p) && j == 0) {
2214 /* First branch index doesn't need key data. */
2218 rc = mdb_add_node(bt, &pdp->p, j, &rkey, &rdata, pgno,flags);
2226 mdb_put(MDB_db *bt, MDB_txn *txn,
2227 MDB_val *key, MDB_val *data, unsigned int flags)
2229 int rc = MDB_SUCCESS, exact;
2234 assert(key != NULL);
2235 assert(data != 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 DPRINTF("==> put key %.*s, size %zu, data size %zu",
2253 (int)key->mv_size, (char *)key->mv_data, key->mv_size, data->mv_size);
2255 rc = mdb_search_page(bt, txn, key, NULL, 1, &mpp);
2256 if (rc == MDB_SUCCESS) {
2257 leaf = mdb_search_node(bt, mpp.mp_page, key, &exact, &ki);
2258 if (leaf && exact) {
2259 if (F_ISSET(flags, MDB_NOOVERWRITE)) {
2260 DPRINTF("duplicate key %.*s",
2261 (int)key->mv_size, (char *)key->mv_data);
2264 mdb_del_node(bt, mpp.mp_page, ki);
2266 if (leaf == NULL) { /* append if not found */
2267 ki = NUMKEYS(mpp.mp_page);
2268 DPRINTF("appending key at index %i", ki);
2270 } else if (rc == ENOENT) {
2272 /* new file, just write a root leaf page */
2273 DPRINTF("allocating new root leaf page");
2274 if ((dp = mdbenv_new_page(bt->md_env, P_LEAF, 1)) == NULL) {
2277 mpp.mp_page = &dp->p;
2278 txn->mt_root = mpp.mp_page->mp_pgno;
2279 bt->md_env->me_meta.mm_stat.ms_depth++;
2285 assert(IS_LEAF(mpp.mp_page));
2286 DPRINTF("there are %u keys, should insert new key at index %i",
2287 NUMKEYS(mpp.mp_page), ki);
2289 if (SIZELEFT(mpp.mp_page) < mdb_leaf_size(bt, key, data)) {
2290 rc = mdb_split(bt, &mpp.mp_page, &ki, key, data, P_INVALID);
2292 /* There is room already in this leaf page. */
2293 rc = mdb_add_node(bt, mpp.mp_page, ki, key, data, 0, 0);
2296 if (rc != MDB_SUCCESS)
2297 txn->mt_flags |= MDB_TXN_ERROR;
2299 bt->md_env->me_meta.mm_stat.ms_entries++;
2306 mdbenv_get_flags(MDB_env *env, unsigned int *arg)
2311 *arg = env->me_flags;
2316 mdbenv_get_path(MDB_env *env, const char **arg)
2321 *arg = env->me_path;
2326 mdbenv_stat(MDB_env *env, MDB_stat **arg)
2328 if (env == NULL || arg == NULL)
2331 *arg = &env->me_meta.mm_stat;
2336 int mdb_open(MDB_env *env, MDB_txn *txn, const char *name, unsigned int flags, MDB_db **db)
2339 *db = (MDB_db *)&env->me_db;
2345 int mdb_stat(MDB_db *db, MDB_stat **arg)
2347 if (db == NULL || arg == NULL)
2350 *arg = &db->md_stat;
2355 void mdb_close(MDB_db *db)