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 (~0L)
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_HEAD 0x10 /* header page */
103 #define P_DIRTY 0x20 /* dirty page */
105 #define mp_lower mp_pb.pb.pb_lower
106 #define mp_upper mp_pb.pb.pb_upper
107 #define mp_pages mp_pb.pb_pages
110 indx_t pb_lower; /* lower bound of free space */
111 indx_t pb_upper; /* upper bound of free space */
113 uint32_t pb_pages; /* number of overflow pages */
115 indx_t mp_ptrs[1]; /* dynamic size */
118 #define PAGEHDRSZ offsetof(MDB_page, mp_ptrs)
120 #define NUMKEYS(p) (((p)->mp_lower - PAGEHDRSZ) >> 1)
121 #define SIZELEFT(p) (indx_t)((p)->mp_upper - (p)->mp_lower)
122 #define PAGEFILL(env, p) (1000 * ((env)->me_head.mh_psize - PAGEHDRSZ - SIZELEFT(p)) / \
123 ((env)->me_head.mh_psize - PAGEHDRSZ))
124 #define IS_LEAF(p) F_ISSET((p)->mp_flags, P_LEAF)
125 #define IS_BRANCH(p) F_ISSET((p)->mp_flags, P_BRANCH)
126 #define IS_OVERFLOW(p) F_ISSET((p)->mp_flags, P_OVERFLOW)
128 typedef struct MDB_head { /* header page content */
132 uint32_t mh_psize; /* page size */
133 void *mh_address; /* address for fixed mapping */
134 size_t mh_mapsize; /* size of mmap region */
137 typedef struct MDB_meta { /* meta (footer) page content */
139 pgno_t mm_root; /* page number of root page */
140 pgno_t mm_last_pg; /* last used page in file */
141 ulong mm_txnid; /* txnid that committed this page */
144 typedef struct MDB_dhead { /* a dirty page */
145 SIMPLEQ_ENTRY(MDB_dpage) md_next; /* queue of dirty pages */
147 int md_pi; /* parent index */
151 typedef struct MDB_dpage {
156 SIMPLEQ_HEAD(dirty_queue, MDB_dpage);
158 typedef struct MDB_pageparent {
164 static MDB_dpage *mdb_newpage(MDB_txn *txn, MDB_page *parent, int parent_idx, int num);
165 static int mdb_touch(MDB_txn *txn, MDB_pageparent *mp);
167 typedef struct MDB_ppage { /* ordered list of pages */
168 SLIST_ENTRY(MDB_ppage) mp_entry;
170 unsigned int mp_ki; /* cursor index on page */
172 SLIST_HEAD(page_stack, MDB_ppage);
174 #define CURSOR_EMPTY(c) SLIST_EMPTY(&(c)->mc_stack)
175 #define CURSOR_TOP(c) SLIST_FIRST(&(c)->mc_stack)
176 #define CURSOR_POP(c) SLIST_REMOVE_HEAD(&(c)->mc_stack, mp_entry)
177 #define CURSOR_PUSH(c,p) SLIST_INSERT_HEAD(&(c)->mc_stack, p, mp_entry)
182 struct page_stack mc_stack; /* stack of parent pages */
183 short mc_initialized; /* 1 if initialized */
184 short mc_eof; /* 1 if end is reached */
187 #define METAHASHLEN offsetof(MDB_meta, mm_hash)
188 #define METADATA(p) ((void *)((char *)p + PAGEHDRSZ))
190 typedef struct MDB_node {
191 #define mn_pgno mn_p.np_pgno
192 #define mn_dsize mn_p.np_dsize
194 pgno_t np_pgno; /* child page number */
195 uint32_t np_dsize; /* leaf data size */
197 unsigned int mn_flags:4;
198 unsigned int mn_ksize:12; /* key size */
199 #define F_BIGDATA 0x01 /* data put on overflow page */
204 pgno_t mt_root; /* current / new root page */
205 pgno_t mt_next_pgno; /* next unallocated page */
206 pgno_t mt_first_pgno;
210 struct dirty_queue *dirty_queue; /* modified pages */
213 #define MDB_TXN_RDONLY 0x01 /* read-only transaction */
214 #define MDB_TXN_ERROR 0x02 /* an error has occurred */
215 unsigned int mt_flags;
218 /* Must be same as MDB_db, minus md_root/md_stat */
219 typedef struct MDB_db0 {
220 unsigned int md_flags;
221 MDB_cmp_func *md_cmp; /* user compare function */
222 MDB_rel_func *md_rel; /* user relocate function */
223 MDB_db *md_parent; /* parent tree */
228 unsigned int md_flags;
229 MDB_cmp_func *md_cmp; /* user compare function */
230 MDB_rel_func *md_rel; /* user relocate function */
231 MDB_db *md_parent; /* parent tree */
234 pgno_t md_root; /* page number of root page */
245 MDB_txninfo *me_txns;
247 MDB_db0 me_db; /* first DB, overlaps with meta */
249 MDB_txn *me_txn; /* current write transaction */
251 off_t me_size; /* current file size */
252 pthread_key_t me_txkey; /* thread-key for readers */
255 #define NODESIZE offsetof(MDB_node, mn_data)
257 #define INDXSIZE(k) (NODESIZE + ((k) == NULL ? 0 : (k)->mv_size))
258 #define LEAFSIZE(k, d) (NODESIZE + (k)->mv_size + (d)->mv_size)
259 #define NODEPTR(p, i) ((MDB_node *)((char *)(p) + (p)->mp_ptrs[i]))
260 #define NODEKEY(node) (void *)((node)->mn_data)
261 #define NODEDATA(node) (void *)((char *)(node)->mn_data + (node)->mn_ksize)
262 #define NODEPGNO(node) ((node)->mn_pgno)
263 #define NODEDSZ(node) ((node)->mn_dsize)
265 #define MDB_COMMIT_PAGES 64 /* max number of pages to write in one commit */
266 #define MDB_MAXCACHE_DEF 1024 /* max number of pages to keep in cache */
268 static int mdb_search_page_root(MDB_db *db,
270 MDB_cursor *cursor, int modify,
271 MDB_pageparent *mpp);
272 static int mdb_search_page(MDB_db *db,
273 MDB_txn *txn, MDB_val *key,
274 MDB_cursor *cursor, int modify,
275 MDB_pageparent *mpp);
277 static int mdbenv_write_header(MDB_env *env);
278 static int mdbenv_read_header(MDB_env *env);
279 static int mdb_check_meta_page(MDB_page *p);
280 static int mdbenv_read_meta(MDB_env *env);
281 static int mdbenv_write_meta(MDB_txn *txn);
282 static MDB_page *mdbenv_get_page(MDB_env *env, pgno_t pgno);
284 static MDB_node *mdb_search_node(MDB_db *db, MDB_page *mp,
285 MDB_val *key, int *exactp, unsigned int *kip);
286 static int mdb_add_node(MDB_db *bt, MDB_page *mp,
287 indx_t indx, MDB_val *key, MDB_val *data,
288 pgno_t pgno, uint8_t flags);
289 static void mdb_del_node(MDB_db *bt, MDB_page *mp,
291 static int mdb_read_data(MDB_db *bt, MDB_page *mp,
292 MDB_node *leaf, MDB_val *data);
294 static int mdb_rebalance(MDB_db *bt, MDB_pageparent *mp);
295 static int mdb_update_key(MDB_db *bt, MDB_page *mp,
296 indx_t indx, MDB_val *key);
297 static int mdb_move_node(MDB_db *bt,
298 MDB_pageparent *src, indx_t srcindx,
299 MDB_pageparent *dst, indx_t dstindx);
300 static int mdb_merge(MDB_db *bt, MDB_pageparent *src,
301 MDB_pageparent *dst);
302 static int mdb_split(MDB_db *bt, MDB_page **mpp,
303 unsigned int *newindxp, MDB_val *newkey,
304 MDB_val *newdata, pgno_t newpgno);
305 static MDB_dpage *mdbenv_new_page(MDB_env *env, uint32_t flags, int num);
307 static void cursor_pop_page(MDB_cursor *cursor);
308 static MDB_ppage *cursor_push_page(MDB_cursor *cursor,
311 static int mdb_set_key(MDB_db *bt, MDB_page *mp,
312 MDB_node *node, MDB_val *key);
313 static int mdb_sibling(MDB_cursor *cursor, int move_right);
314 static int mdb_cursor_next(MDB_cursor *cursor,
315 MDB_val *key, MDB_val *data);
316 static int mdb_cursor_set(MDB_cursor *cursor,
317 MDB_val *key, MDB_val *data, int *exactp);
318 static int mdb_cursor_first(MDB_cursor *cursor,
319 MDB_val *key, MDB_val *data);
321 static size_t mdb_leaf_size(MDB_db *bt, MDB_val *key,
323 static size_t mdb_branch_size(MDB_db *bt, MDB_val *key);
325 static pgno_t mdbenv_compact_tree(MDB_env *env, pgno_t pgno,
328 static int memncmp(const void *s1, size_t n1,
329 const void *s2, size_t n2);
330 static int memnrcmp(const void *s1, size_t n1,
331 const void *s2, size_t n2);
334 memncmp(const void *s1, size_t n1, const void *s2, size_t n2)
337 if (memcmp(s1, s2, n1) == 0)
341 if (memcmp(s1, s2, n2) == 0)
344 return memcmp(s1, s2, n1);
348 memnrcmp(const void *s1, size_t n1, const void *s2, size_t n2)
350 const unsigned char *p1;
351 const unsigned char *p2;
354 return n2 == 0 ? 0 : -1;
357 return n1 == 0 ? 0 : 1;
359 p1 = (const unsigned char *)s1 + n1 - 1;
360 p2 = (const unsigned char *)s2 + n2 - 1;
364 return (p2 == s2) ? 0 : -1;
366 return (p1 == p2) ? 0 : 1;
374 mdb_cmp(MDB_db *db, const MDB_val *a, const MDB_val *b)
376 return db->md_cmp(a, b);
380 _mdb_cmp(MDB_db *db, const MDB_val *key1, const MDB_val *key2)
382 if (F_ISSET(db->md_flags, MDB_REVERSEKEY))
383 return memnrcmp(key1->mv_data, key1->mv_size, key2->mv_data, key2->mv_size);
385 return memncmp((char *)key1->mv_data, key1->mv_size, key2->mv_data, key2->mv_size);
388 /* Allocate new page(s) for writing */
390 mdb_newpage(MDB_txn *txn, MDB_page *parent, int parent_idx, int num)
394 if ((dp = malloc(txn->mt_env->me_head.mh_psize * num + sizeof(MDB_dhead))) == NULL)
397 dp->h.md_parent = parent;
398 dp->h.md_pi = parent_idx;
399 SIMPLEQ_INSERT_TAIL(txn->mt_u.dirty_queue, dp, h.md_next);
400 dp->p.mp_pgno = txn->mt_next_pgno;
401 txn->mt_next_pgno += num;
406 /* Touch a page: make it dirty and re-insert into tree with updated pgno.
409 mdb_touch(MDB_txn *txn, MDB_pageparent *pp)
412 MDB_page *mp = pp->mp_page;
417 if (!F_ISSET(mp->mp_flags, P_DIRTY)) {
419 DPRINTF("touching page %lu -> %lu", mp->mp_pgno, txn->mt_next_pgno);
420 if ((dp = mdb_newpage(txn, pp->mp_parent, pp->mp_pi, 1)) == NULL)
422 pgno = dp->p.mp_pgno;
423 bcopy(mp, &dp->p, txn->mt_env->me_head.mh_psize);
426 mp->mp_flags |= P_DIRTY;
428 /* Update the page number to new touched page. */
429 if (pp->mp_parent != NULL)
430 NODEPGNO(NODEPTR(pp->mp_parent, pp->mp_pi)) = mp->mp_pgno;
437 mdbenv_sync(MDB_env *env)
440 if (!F_ISSET(env->me_flags, MDB_NOSYNC)) {
441 if (fsync(env->me_fd))
448 mdb_txn_begin(MDB_env *env, int rdonly, MDB_txn **ret)
453 if ((txn = calloc(1, sizeof(*txn))) == NULL) {
454 DPRINTF("calloc: %s", strerror(errno));
459 txn->mt_flags |= MDB_TXN_RDONLY;
461 txn->mt_u.dirty_queue = calloc(1, sizeof(*txn->mt_u.dirty_queue));
462 if (txn->mt_u.dirty_queue == NULL) {
466 SIMPLEQ_INIT(txn->mt_u.dirty_queue);
468 pthread_mutex_lock(&env->me_txns->mt_wmutex);
469 env->me_txns->mt_txnid++;
471 txn->mt_txnid = env->me_txns->mt_txnid;
473 MDB_reader *r = pthread_getspecific(env->me_txkey);
476 pthread_mutex_lock(&env->me_txns->mt_mutex);
477 for (i=0; i<env->me_maxreaders; i++) {
478 if (env->me_txns->mt_readers[i].mr_pid == 0) {
479 env->me_txns->mt_readers[i].mr_pid = getpid();
480 env->me_txns->mt_readers[i].mr_tid = pthread_self();
481 pthread_setspecific(env->me_txkey, &env->me_txns->mt_readers[i]);
482 if (i >= env->me_txns->mt_numreaders)
483 env->me_txns->mt_numreaders = i+1;
487 pthread_mutex_unlock(&env->me_txns->mt_mutex);
488 if (i == env->me_maxreaders) {
492 r->mr_txnid = txn->mt_txnid;
493 txn->mt_u.reader = r;
500 if ((rc = mdbenv_read_meta(env)) != MDB_SUCCESS) {
505 txn->mt_next_pgno = env->me_meta.mm_last_pg+1;
506 txn->mt_first_pgno = txn->mt_next_pgno;
507 txn->mt_root = env->me_meta.mm_root;
508 DPRINTF("begin transaction %lu on mdbenv %p, root page %lu", txn->mt_txnid, env, txn->mt_root);
515 mdb_txn_abort(MDB_txn *txn)
524 DPRINTF("abort transaction %lu on mdbenv %p, root page %lu", txn->mt_txnid, env, txn->mt_root);
526 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
527 txn->mt_u.reader->mr_txnid = 0;
529 /* Discard all dirty pages.
531 while (!SIMPLEQ_EMPTY(txn->mt_u.dirty_queue)) {
532 dp = SIMPLEQ_FIRST(txn->mt_u.dirty_queue);
533 SIMPLEQ_REMOVE_HEAD(txn->mt_u.dirty_queue, h.md_next);
538 DPRINTF("releasing write lock on txn %p", txn);
540 if (flock(txn->bt->fd, LOCK_UN) != 0) {
541 DPRINTF("failed to unlock fd %d: %s",
542 txn->bt->fd, strerror(errno));
545 free(txn->mt_u.dirty_queue);
547 pthread_mutex_unlock(&env->me_txns->mt_wmutex);
554 mdb_txn_commit(MDB_txn *txn)
562 struct iovec iov[MDB_COMMIT_PAGES];
565 assert(txn->mt_env != NULL);
569 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
570 DPRINTF("attempt to commit read-only transaction");
575 if (txn != env->me_txn) {
576 DPRINTF("attempt to commit unknown transaction");
581 if (F_ISSET(txn->mt_flags, MDB_TXN_ERROR)) {
582 DPRINTF("error flag is set, can't commit");
587 if (SIMPLEQ_EMPTY(txn->mt_u.dirty_queue))
590 DPRINTF("committing transaction %lu on mdbenv %p, root page %lu",
591 txn->mt_txnid, env, txn->mt_root);
593 /* Commit up to MDB_COMMIT_PAGES dirty pages to disk until done.
600 SIMPLEQ_FOREACH(dp, txn->mt_u.dirty_queue, h.md_next) {
601 if (dp->p.mp_pgno != next) {
602 lseek(env->me_fd, dp->p.mp_pgno * env->me_head.mh_psize, SEEK_SET);
603 next = dp->p.mp_pgno;
607 DPRINTF("committing page %lu", dp->p.mp_pgno);
608 iov[n].iov_len = env->me_head.mh_psize * dp->h.md_num;
609 iov[n].iov_base = &dp->p;
610 size += iov[n].iov_len;
611 next = dp->p.mp_pgno + dp->h.md_num;
612 /* clear dirty flag */
613 dp->p.mp_flags &= ~P_DIRTY;
614 if (++n >= MDB_COMMIT_PAGES) {
623 DPRINTF("committing %u dirty pages", n);
624 rc = writev(env->me_fd, iov, n);
628 DPRINTF("short write, filesystem full?");
630 DPRINTF("writev: %s", strerror(errno));
635 /* Drop the dirty pages.
637 while (!SIMPLEQ_EMPTY(txn->mt_u.dirty_queue)) {
638 dp = SIMPLEQ_FIRST(txn->mt_u.dirty_queue);
639 SIMPLEQ_REMOVE_HEAD(txn->mt_u.dirty_queue, h.md_next);
646 if ((n = mdbenv_sync(env)) != 0 ||
647 (n = mdbenv_write_meta(txn)) != MDB_SUCCESS ||
648 (n = mdbenv_sync(env)) != 0) {
653 pthread_mutex_unlock(&env->me_txns->mt_wmutex);
654 free(txn->mt_u.dirty_queue);
665 mdbenv_write_header(MDB_env *env)
673 DPRINTF("writing header page");
676 psize = sysconf(_SC_PAGE_SIZE);
678 if ((p = calloc(1, psize)) == NULL)
680 p->mp_flags = P_HEAD;
682 env->me_head.mh_psize = psize;
683 env->me_head.mh_flags = env->me_flags & 0xffff;
686 bcopy(&env->me_head, h, sizeof(*h));
688 rc = write(env->me_fd, p, env->me_head.mh_psize);
690 if (rc != (ssize_t)env->me_head.mh_psize) {
693 DPRINTF("short write, filesystem full?");
701 mdbenv_read_header(MDB_env *env)
710 /* We don't know the page size yet, so use a minimum value.
713 if ((rc = pread(env->me_fd, page, PAGESIZE, 0)) == 0) {
715 } else if (rc != PAGESIZE) {
718 DPRINTF("read: %s", strerror(errno));
722 p = (MDB_page *)page;
724 if (!F_ISSET(p->mp_flags, P_HEAD)) {
725 DPRINTF("page %lu not a header page", p->mp_pgno);
730 if (h->mh_magic != MDB_MAGIC) {
731 DPRINTF("header has invalid magic");
735 if (h->mh_version != MDB_VERSION) {
736 DPRINTF("database is version %u, expected version %u",
737 h->mh_version, MDB_VERSION);
741 bcopy(h, &env->me_head, sizeof(*h));
746 mdbenv_init_meta(MDB_env *env)
752 p = calloc(2, env->me_head.mh_psize);
754 p->mp_flags = P_META;
757 meta->mm_root = P_INVALID;
758 meta->mm_last_pg = 2;
760 q = (MDB_page *)((char *)p + env->me_head.mh_psize);
763 q->mp_flags = P_META;
766 meta->mm_root = P_INVALID;
767 meta->mm_last_pg = 2;
769 rc = write(env->me_fd, p, env->me_head.mh_psize * 2);
771 return (rc == env->me_head.mh_psize * 2) ? MDB_SUCCESS : errno;
775 mdbenv_write_meta(MDB_txn *txn)
783 assert(txn->mt_env != NULL);
785 DPRINTF("writing meta page for root page %lu", txn->mt_root);
789 bcopy(&env->me_meta, &meta, sizeof(meta));
790 meta.mm_root = txn->mt_root;
791 meta.mm_last_pg = txn->mt_next_pgno - 1;
792 meta.mm_txnid = txn->mt_txnid;
794 off = env->me_head.mh_psize;
795 if (!env->me_metatoggle)
799 lseek(env->me_fd, off, SEEK_SET);
800 rc = write(env->me_fd, &meta, sizeof(meta));
801 if (rc != sizeof(meta)) {
802 DPRINTF("write failed, disk error?");
809 /* Returns true if page p is a valid meta page, false otherwise.
812 mdb_check_meta_page(MDB_page *p)
814 if (!F_ISSET(p->mp_flags, P_META)) {
815 DPRINTF("page %lu not a meta page", p->mp_pgno);
823 mdbenv_read_meta(MDB_env *env)
831 if ((mp0 = mdbenv_get_page(env, 1)) == NULL ||
832 (mp1 = mdbenv_get_page(env, 2)) == NULL)
835 rc = mdb_check_meta_page(mp0);
838 rc = mdb_check_meta_page(mp1);
841 meta[0] = METADATA(mp0);
842 meta[1] = METADATA(mp1);
844 if (meta[0]->mm_txnid < meta[1]->mm_txnid)
847 if (meta[toggle]->mm_txnid > env->me_meta.mm_txnid) {
848 bcopy(meta[toggle], &env->me_meta, sizeof(env->me_meta));
849 env->me_metatoggle = toggle;
852 DPRINTF("Using meta page %d", toggle);
858 mdbenv_create(MDB_env **env)
862 e = calloc(1, sizeof(*e));
863 if (!e) return ENOMEM;
865 e->me_head.mh_magic = MDB_MAGIC;
866 e->me_head.mh_version = MDB_VERSION;
867 e->me_head.mh_mapsize = DEFAULT_MAPSIZE;
868 e->me_maxreaders = DEFAULT_READERS;
875 mdbenv_set_mapsize(MDB_env *env, size_t size)
879 env->me_mapsize = env->me_head.mh_mapsize = size;
884 mdbenv_set_maxreaders(MDB_env *env, int readers)
886 env->me_maxreaders = readers;
891 mdbenv_get_maxreaders(MDB_env *env, int *readers)
893 if (!env || !readers)
895 *readers = env->me_maxreaders;
900 mdbenv_open2(MDB_env *env, unsigned int flags)
904 env->me_flags = flags;
905 env->me_meta.mm_root = P_INVALID;
906 env->me_meta.mm_last_pg = 2;
908 if ((i = mdbenv_read_header(env)) != 0) {
911 DPRINTF("new mdbenv");
915 if (!env->me_mapsize)
916 env->me_mapsize = env->me_head.mh_mapsize;
919 if (env->me_head.mh_address && (flags & MDB_FIXEDMAP))
921 env->me_map = mmap(env->me_head.mh_address, env->me_mapsize, PROT_READ, i,
923 if (env->me_map == MAP_FAILED)
927 env->me_head.mh_mapsize = env->me_mapsize;
928 if (flags & MDB_FIXEDMAP)
929 env->me_head.mh_address = env->me_map;
930 i = mdbenv_write_header(env);
931 if (i != MDB_SUCCESS) {
932 munmap(env->me_map, env->me_mapsize);
935 i = mdbenv_init_meta(env);
936 if (i != MDB_SUCCESS) {
937 munmap(env->me_map, env->me_mapsize);
942 if ((i = mdbenv_read_meta(env)) != 0)
945 DPRINTF("opened database version %u, pagesize %u",
946 env->me_head.mh_version, env->me_head.mh_psize);
947 DPRINTF("depth: %u", env->me_meta.mm_stat.ms_depth);
948 DPRINTF("entries: %lu", env->me_meta.mm_stat.ms_entries);
949 DPRINTF("branch pages: %lu", env->me_meta.mm_stat.ms_branch_pages);
950 DPRINTF("leaf pages: %lu", env->me_meta.mm_stat.ms_leaf_pages);
951 DPRINTF("overflow pages: %lu", env->me_meta.mm_stat.ms_overflow_pages);
952 DPRINTF("root: %lu", env->me_meta.mm_root);
958 mdbenv_reader_dest(void *ptr)
960 MDB_reader *reader = ptr;
962 reader->mr_txnid = 0;
968 mdbenv_open(MDB_env *env, const char *path, unsigned int flags, mode_t mode)
970 int oflags, rc, shmid;
974 if (F_ISSET(flags, MDB_RDONLY))
977 oflags = O_RDWR | O_CREAT;
979 if ((env->me_fd = open(path, oflags, mode)) == -1)
982 env->me_shmkey = ftok(path, 'm');
983 size = (env->me_maxreaders-1) * sizeof(MDB_reader) + sizeof(MDB_txninfo);
984 shmid = shmget(env->me_shmkey, size, IPC_CREAT|IPC_EXCL|mode);
986 if (errno == EEXIST) {
987 shmid = shmget(env->me_shmkey, size, IPC_CREAT|mode);
990 env->me_txns = shmat(shmid, NULL, 0);
991 if (env->me_txns->mt_magic != MDB_MAGIC ||
992 env->me_txns->mt_version != MDB_VERSION) {
993 DPRINTF("invalid lock region %d", shmid);
1002 pthread_mutexattr_t mattr;
1004 env->me_txns = shmat(shmid, NULL, 0);
1005 pthread_mutexattr_init(&mattr);
1006 pthread_mutexattr_setpshared(&mattr, PTHREAD_PROCESS_SHARED);
1007 pthread_mutex_init(&env->me_txns->mt_mutex, &mattr);
1008 pthread_mutex_init(&env->me_txns->mt_wmutex, &mattr);
1009 env->me_txns->mt_version = MDB_VERSION;
1010 env->me_txns->mt_magic = MDB_MAGIC;
1013 if ((rc = mdbenv_open2(env, flags)) != MDB_SUCCESS) {
1017 env->me_path = strdup(path);
1018 DPRINTF("opened dbenv %p", env);
1021 pthread_key_create(&env->me_txkey, mdbenv_reader_dest);
1027 mdbenv_close(MDB_env *env)
1035 munmap(env->me_map, env->me_mapsize);
1039 shmdt(env->me_txns);
1044 /* Search for key within a leaf page, using binary search.
1045 * Returns the smallest entry larger or equal to the key.
1046 * If exactp is non-null, stores whether the found entry was an exact match
1047 * in *exactp (1 or 0).
1048 * If kip is non-null, stores the index of the found entry in *kip.
1049 * If no entry larger or equal to the key is found, returns NULL.
1052 mdb_search_node(MDB_db *bt, MDB_page *mp, MDB_val *key,
1053 int *exactp, unsigned int *kip)
1061 DPRINTF("searching %lu keys in %s page %lu",
1063 IS_LEAF(mp) ? "leaf" : "branch",
1066 assert(NUMKEYS(mp) > 0);
1068 bzero(&nodekey, sizeof(nodekey));
1070 low = IS_LEAF(mp) ? 0 : 1;
1071 high = NUMKEYS(mp) - 1;
1072 while (low <= high) {
1073 i = (low + high) >> 1;
1074 node = NODEPTR(mp, i);
1076 nodekey.mv_size = node->mn_ksize;
1077 nodekey.mv_data = NODEKEY(node);
1080 rc = bt->md_cmp(key, &nodekey);
1082 rc = _mdb_cmp(bt, key, &nodekey);
1085 DPRINTF("found leaf index %u [%.*s], rc = %i",
1086 i, (int)nodekey.mv_size, (char *)nodekey.mv_data, rc);
1088 DPRINTF("found branch index %u [%.*s -> %lu], rc = %i",
1089 i, (int)node->mn_ksize, (char *)NODEKEY(node),
1100 if (rc > 0) { /* Found entry is less than the key. */
1101 i++; /* Skip to get the smallest entry larger than key. */
1102 if (i >= NUMKEYS(mp))
1103 /* There is no entry larger or equal to the key. */
1107 *exactp = (rc == 0);
1108 if (kip) /* Store the key index if requested. */
1111 return NODEPTR(mp, i);
1115 cursor_pop_page(MDB_cursor *cursor)
1119 top = CURSOR_TOP(cursor);
1122 DPRINTF("popped page %lu off cursor %p", top->mp_page->mp_pgno, cursor);
1128 cursor_push_page(MDB_cursor *cursor, MDB_page *mp)
1132 DPRINTF("pushing page %lu on cursor %p", mp->mp_pgno, cursor);
1134 if ((ppage = calloc(1, sizeof(*ppage))) == NULL)
1136 ppage->mp_page = mp;
1137 CURSOR_PUSH(cursor, ppage);
1142 mdbenv_get_page(MDB_env *env, pgno_t pgno)
1145 MDB_txn *txn = env->me_txn;
1147 if (txn && pgno >= txn->mt_first_pgno &&
1148 !SIMPLEQ_EMPTY(txn->mt_u.dirty_queue)) {
1150 SIMPLEQ_FOREACH(dp, txn->mt_u.dirty_queue, h.md_next) {
1151 if (dp->p.mp_pgno == pgno) {
1157 p = (MDB_page *)(env->me_map + env->me_head.mh_psize * pgno);
1163 mdb_search_page_root(MDB_db *bt, MDB_val *key,
1164 MDB_cursor *cursor, int modify, MDB_pageparent *mpp)
1166 MDB_page *mp = mpp->mp_page;
1169 if (cursor && cursor_push_page(cursor, mp) == NULL)
1172 while (IS_BRANCH(mp)) {
1176 DPRINTF("branch page %lu has %lu keys", mp->mp_pgno, NUMKEYS(mp));
1177 assert(NUMKEYS(mp) > 1);
1178 DPRINTF("found index 0 to page %lu", NODEPGNO(NODEPTR(mp, 0)));
1180 if (key == NULL) /* Initialize cursor to first page. */
1184 node = mdb_search_node(bt, mp, key, &exact, &i);
1186 i = NUMKEYS(mp) - 1;
1194 DPRINTF("following index %u for key %.*s",
1195 i, (int)key->mv_size, (char *)key->mv_data);
1196 assert(i >= 0 && i < NUMKEYS(mp));
1197 node = NODEPTR(mp, i);
1200 CURSOR_TOP(cursor)->mp_ki = i;
1202 mpp->mp_parent = mp;
1203 if ((mp = mdbenv_get_page(bt->md_env, NODEPGNO(node))) == NULL)
1208 if (cursor && cursor_push_page(cursor, mp) == NULL)
1212 MDB_dhead *dh = ((MDB_dhead *)mp)-1;
1213 if (rc = mdb_touch(bt->md_env->me_txn, mpp))
1215 dh = ((MDB_dhead *)mpp->mp_page)-1;
1216 dh->md_parent = mpp->mp_parent;
1217 dh->md_pi = mpp->mp_pi;
1224 DPRINTF("internal error, index points to a %02X page!?",
1229 DPRINTF("found leaf page %lu for key %.*s", mp->mp_pgno,
1230 key ? (int)key->mv_size : 0, key ? (char *)key->mv_data : NULL);
1235 /* Search for the page a given key should be in.
1236 * Stores a pointer to the found page in *mpp.
1237 * If key is NULL, search for the lowest page (used by mdb_cursor_first).
1238 * If cursor is non-null, pushes parent pages on the cursor stack.
1239 * If modify is true, visited pages are updated with new page numbers.
1242 mdb_search_page(MDB_db *db, MDB_txn *txn, MDB_val *key,
1243 MDB_cursor *cursor, int modify, MDB_pageparent *mpp)
1248 /* Can't modify pages outside a transaction. */
1249 if (txn == NULL && modify) {
1253 /* Choose which root page to start with. If a transaction is given
1254 * use the root page from the transaction, otherwise read the last
1255 * committed root page.
1258 if ((rc = mdbenv_read_meta(db->md_env)) != MDB_SUCCESS)
1260 root = db->md_env->me_meta.mm_root;
1261 } else if (F_ISSET(txn->mt_flags, MDB_TXN_ERROR)) {
1262 DPRINTF("transaction has failed, must abort");
1265 root = txn->mt_root;
1267 if (root == P_INVALID) { /* Tree is empty. */
1268 DPRINTF("tree is empty");
1272 if ((mpp->mp_page = mdbenv_get_page(db->md_env, root)) == NULL)
1275 DPRINTF("root page has flags 0x%X", mpp->mp_page->mp_flags);
1277 if (modify && !F_ISSET(mpp->mp_page->mp_flags, P_DIRTY)) {
1278 mpp->mp_parent = NULL;
1280 if ((rc = mdb_touch(txn, mpp)))
1282 txn->mt_root = mpp->mp_page->mp_pgno;
1285 return mdb_search_page_root(db, key, cursor, modify, mpp);
1289 mdb_read_data(MDB_db *db, MDB_page *mp, MDB_node *leaf,
1292 MDB_page *omp; /* overflow mpage */
1298 bzero(data, sizeof(*data));
1299 max = db->md_env->me_head.mh_psize - PAGEHDRSZ;
1301 if (!F_ISSET(leaf->mn_flags, F_BIGDATA)) {
1302 data->mv_size = leaf->mn_dsize;
1303 data->mv_data = NODEDATA(leaf);
1307 /* Read overflow data.
1309 data->mv_size = leaf->mn_dsize;
1310 bcopy(NODEDATA(leaf), &pgno, sizeof(pgno));
1311 if ((omp = mdbenv_get_page(db->md_env, pgno)) == NULL) {
1312 DPRINTF("read overflow page %lu failed", pgno);
1315 data->mv_data = omp;
1321 mdb_get(MDB_db *db, MDB_txn *txn,
1322 MDB_val *key, MDB_val *data)
1330 DPRINTF("===> get key [%.*s]", (int)key->mv_size, (char *)key->mv_data);
1335 if (txn != NULL && db->md_env != txn->mt_env) {
1339 if (key->mv_size == 0 || key->mv_size > MAXKEYSIZE) {
1343 if ((rc = mdb_search_page(db, txn, key, NULL, 0, &mpp)) != MDB_SUCCESS)
1346 leaf = mdb_search_node(db, mpp.mp_page, key, &exact, NULL);
1348 rc = mdb_read_data(db, mpp.mp_page, leaf, data);
1357 mdb_sibling(MDB_cursor *cursor, int move_right)
1361 MDB_ppage *parent, *top;
1364 top = CURSOR_TOP(cursor);
1365 if ((parent = SLIST_NEXT(top, mp_entry)) == NULL) {
1366 return ENOENT; /* root has no siblings */
1369 DPRINTF("parent page is page %lu, index %u",
1370 parent->mp_page->mp_pgno, parent->mp_ki);
1372 cursor_pop_page(cursor);
1373 if (move_right ? (parent->mp_ki + 1 >= NUMKEYS(parent->mp_page))
1374 : (parent->mp_ki == 0)) {
1375 DPRINTF("no more keys left, moving to %s sibling",
1376 move_right ? "right" : "left");
1377 if ((rc = mdb_sibling(cursor, move_right)) != MDB_SUCCESS)
1379 parent = CURSOR_TOP(cursor);
1385 DPRINTF("just moving to %s index key %u",
1386 move_right ? "right" : "left", parent->mp_ki);
1388 assert(IS_BRANCH(parent->mp_page));
1390 indx = NODEPTR(parent->mp_page, parent->mp_ki);
1391 if ((mp = mdbenv_get_page(cursor->mc_db->md_env, indx->mn_pgno)) == NULL)
1394 mp->parent = parent->mp_page;
1395 mp->parent_index = parent->mp_ki;
1398 cursor_push_page(cursor, mp);
1404 mdb_set_key(MDB_db *bt, MDB_page *mp, MDB_node *node,
1410 key->mv_size = node->mn_ksize;
1411 key->mv_data = NODEKEY(node);
1417 mdb_cursor_next(MDB_cursor *cursor, MDB_val *key, MDB_val *data)
1423 if (cursor->mc_eof) {
1427 assert(cursor->mc_initialized);
1429 top = CURSOR_TOP(cursor);
1432 DPRINTF("cursor_next: top page is %lu in cursor %p", mp->mp_pgno, cursor);
1434 if (top->mp_ki + 1 >= NUMKEYS(mp)) {
1435 DPRINTF("=====> move to next sibling page");
1436 if (mdb_sibling(cursor, 1) != MDB_SUCCESS) {
1440 top = CURSOR_TOP(cursor);
1442 DPRINTF("next page is %lu, key index %u", mp->mp_pgno, top->mp_ki);
1446 DPRINTF("==> cursor points to page %lu with %lu keys, key index %u",
1447 mp->mp_pgno, NUMKEYS(mp), top->mp_ki);
1449 assert(IS_LEAF(mp));
1450 leaf = NODEPTR(mp, top->mp_ki);
1452 if (data && mdb_read_data(cursor->mc_db, mp, leaf, data) != MDB_SUCCESS)
1455 return mdb_set_key(cursor->mc_db, mp, leaf, key);
1459 mdb_cursor_set(MDB_cursor *cursor, MDB_val *key, MDB_val *data,
1469 assert(key->mv_size > 0);
1471 rc = mdb_search_page(cursor->mc_db, cursor->mc_txn, key, cursor, 0, &mpp);
1472 if (rc != MDB_SUCCESS)
1474 assert(IS_LEAF(mpp.mp_page));
1476 top = CURSOR_TOP(cursor);
1477 leaf = mdb_search_node(cursor->mc_db, mpp.mp_page, key, exactp, &top->mp_ki);
1478 if (exactp != NULL && !*exactp) {
1479 /* MDB_CURSOR_EXACT specified and not an exact match. */
1484 DPRINTF("===> inexact leaf not found, goto sibling");
1485 if ((rc = mdb_sibling(cursor, 1)) != MDB_SUCCESS)
1486 return rc; /* no entries matched */
1487 top = CURSOR_TOP(cursor);
1489 mpp.mp_page = top->mp_page;
1490 assert(IS_LEAF(mpp.mp_page));
1491 leaf = NODEPTR(mpp.mp_page, 0);
1494 cursor->mc_initialized = 1;
1497 if (data && (rc = mdb_read_data(cursor->mc_db, mpp.mp_page, leaf, data)) != MDB_SUCCESS)
1500 rc = mdb_set_key(cursor->mc_db, mpp.mp_page, leaf, key);
1501 if (rc == MDB_SUCCESS) {
1502 DPRINTF("==> cursor placed on key %.*s",
1503 (int)key->mv_size, (char *)key->mv_data);
1511 mdb_cursor_first(MDB_cursor *cursor, MDB_val *key, MDB_val *data)
1517 rc = mdb_search_page(cursor->mc_db, cursor->mc_txn, NULL, cursor, 0, &mpp);
1518 if (rc != MDB_SUCCESS)
1520 assert(IS_LEAF(mpp.mp_page));
1522 leaf = NODEPTR(mpp.mp_page, 0);
1523 cursor->mc_initialized = 1;
1526 if (data && (rc = mdb_read_data(cursor->mc_db, mpp.mp_page, leaf, data)) != MDB_SUCCESS)
1529 return mdb_set_key(cursor->mc_db, mpp.mp_page, leaf, key);
1533 mdb_cursor_get(MDB_cursor *cursor, MDB_val *key, MDB_val *data,
1543 case MDB_CURSOR_EXACT:
1544 while (CURSOR_TOP(cursor) != NULL)
1545 cursor_pop_page(cursor);
1546 if (key == NULL || key->mv_size == 0 || key->mv_size > MAXKEYSIZE) {
1548 } else if (op == MDB_CURSOR_EXACT)
1549 rc = mdb_cursor_set(cursor, key, data, &exact);
1551 rc = mdb_cursor_set(cursor, key, data, NULL);
1554 if (!cursor->mc_initialized)
1555 rc = mdb_cursor_first(cursor, key, data);
1557 rc = mdb_cursor_next(cursor, key, data);
1560 while (CURSOR_TOP(cursor) != NULL)
1561 cursor_pop_page(cursor);
1562 rc = mdb_cursor_first(cursor, key, data);
1565 DPRINTF("unhandled/unimplemented cursor operation %u", op);
1573 /* Allocate a page and initialize it
1576 mdbenv_new_page(MDB_env *env, uint32_t flags, int num)
1580 assert(env != NULL);
1581 assert(env->me_txn != NULL);
1583 DPRINTF("allocating new mpage %lu, page size %u",
1584 env->me_txn->mt_next_pgno, env->me_head.mh_psize);
1585 if ((dp = mdb_newpage(env->me_txn, NULL, 0, num)) == NULL)
1587 dp->p.mp_flags = flags | P_DIRTY;
1588 dp->p.mp_lower = PAGEHDRSZ;
1589 dp->p.mp_upper = env->me_head.mh_psize;
1591 if (IS_BRANCH(&dp->p))
1592 env->me_meta.mm_stat.ms_branch_pages++;
1593 else if (IS_LEAF(&dp->p))
1594 env->me_meta.mm_stat.ms_leaf_pages++;
1595 else if (IS_OVERFLOW(&dp->p)) {
1596 env->me_meta.mm_stat.ms_overflow_pages += num;
1597 dp->p.mp_pages = num;
1604 mdb_leaf_size(MDB_db *db, MDB_val *key, MDB_val *data)
1608 sz = LEAFSIZE(key, data);
1609 if (data->mv_size >= db->md_env->me_head.mh_psize / MDB_MINKEYS) {
1610 /* put on overflow page */
1611 sz -= data->mv_size - sizeof(pgno_t);
1614 return sz + sizeof(indx_t);
1618 mdb_branch_size(MDB_db *db, MDB_val *key)
1623 if (sz >= db->md_env->me_head.mh_psize / MDB_MINKEYS) {
1624 /* put on overflow page */
1625 /* not implemented */
1626 /* sz -= key->size - sizeof(pgno_t); */
1629 return sz + sizeof(indx_t);
1633 mdb_add_node(MDB_db *db, MDB_page *mp, indx_t indx,
1634 MDB_val *key, MDB_val *data, pgno_t pgno, uint8_t flags)
1637 size_t node_size = NODESIZE;
1640 MDB_dpage *ofp = NULL; /* overflow page */
1642 assert(mp->mp_upper >= mp->mp_lower);
1644 DPRINTF("add node [%.*s] to %s page %lu at index %i, key size %zu",
1645 key ? (int)key->mv_size : 0, key ? (char *)key->mv_data : NULL,
1646 IS_LEAF(mp) ? "leaf" : "branch",
1647 mp->mp_pgno, indx, key ? key->mv_size : 0);
1650 node_size += key->mv_size;
1654 if (F_ISSET(flags, F_BIGDATA)) {
1655 /* Data already on overflow page. */
1656 node_size += sizeof(pgno_t);
1657 } else if (data->mv_size >= db->md_env->me_head.mh_psize / MDB_MINKEYS) {
1658 int ovpages = PAGEHDRSZ + data->mv_size + db->md_env->me_head.mh_psize - 1;
1659 ovpages /= db->md_env->me_head.mh_psize;
1660 /* Put data on overflow page. */
1661 DPRINTF("data size is %zu, put on overflow page",
1663 node_size += sizeof(pgno_t);
1664 if ((ofp = mdbenv_new_page(db->md_env, P_OVERFLOW, ovpages)) == NULL)
1666 DPRINTF("allocated overflow page %lu", ofp->p.mp_pgno);
1669 node_size += data->mv_size;
1673 if (node_size + sizeof(indx_t) > SIZELEFT(mp)) {
1674 DPRINTF("not enough room in page %lu, got %lu ptrs",
1675 mp->mp_pgno, NUMKEYS(mp));
1676 DPRINTF("upper - lower = %u - %u = %u", mp->mp_upper, mp->mp_lower,
1677 mp->mp_upper - mp->mp_lower);
1678 DPRINTF("node size = %zu", node_size);
1682 /* Move higher pointers up one slot. */
1683 for (i = NUMKEYS(mp); i > indx; i--)
1684 mp->mp_ptrs[i] = mp->mp_ptrs[i - 1];
1686 /* Adjust free space offsets. */
1687 ofs = mp->mp_upper - node_size;
1688 assert(ofs >= mp->mp_lower + sizeof(indx_t));
1689 mp->mp_ptrs[indx] = ofs;
1691 mp->mp_lower += sizeof(indx_t);
1693 /* Write the node data. */
1694 node = NODEPTR(mp, indx);
1695 node->mn_ksize = (key == NULL) ? 0 : key->mv_size;
1696 node->mn_flags = flags;
1698 node->mn_dsize = data->mv_size;
1700 node->mn_pgno = pgno;
1703 bcopy(key->mv_data, NODEKEY(node), key->mv_size);
1708 if (F_ISSET(flags, F_BIGDATA))
1709 bcopy(data->mv_data, node->mn_data + key->mv_size,
1712 bcopy(data->mv_data, node->mn_data + key->mv_size,
1715 bcopy(&ofp->p.mp_pgno, node->mn_data + key->mv_size,
1717 bcopy(data->mv_data, METADATA(&ofp->p), data->mv_size);
1725 mdb_del_node(MDB_db *db, MDB_page *mp, indx_t indx)
1728 indx_t i, j, numkeys, ptr;
1732 DPRINTF("delete node %u on %s page %lu", indx,
1733 IS_LEAF(mp) ? "leaf" : "branch", mp->mp_pgno);
1734 assert(indx < NUMKEYS(mp));
1736 node = NODEPTR(mp, indx);
1737 sz = NODESIZE + node->mn_ksize;
1739 if (F_ISSET(node->mn_flags, F_BIGDATA))
1740 sz += sizeof(pgno_t);
1742 sz += NODEDSZ(node);
1745 ptr = mp->mp_ptrs[indx];
1746 numkeys = NUMKEYS(mp);
1747 for (i = j = 0; i < numkeys; i++) {
1749 mp->mp_ptrs[j] = mp->mp_ptrs[i];
1750 if (mp->mp_ptrs[i] < ptr)
1751 mp->mp_ptrs[j] += sz;
1756 base = (char *)mp + mp->mp_upper;
1757 bcopy(base, base + sz, ptr - mp->mp_upper);
1759 mp->mp_lower -= sizeof(indx_t);
1764 mdb_cursor_open(MDB_db *db, MDB_txn *txn, MDB_cursor **ret)
1768 if (db == NULL || ret == NULL)
1771 if (txn != NULL && db->md_env != txn->mt_env) {
1775 if ((cursor = calloc(1, sizeof(*cursor))) != NULL) {
1776 SLIST_INIT(&cursor->mc_stack);
1778 cursor->mc_txn = txn;
1787 mdb_cursor_close(MDB_cursor *cursor)
1789 if (cursor != NULL) {
1790 while (!CURSOR_EMPTY(cursor))
1791 cursor_pop_page(cursor);
1793 /* btree_close(cursor->bt); */
1799 mdb_update_key(MDB_db *db, MDB_page *mp, indx_t indx,
1802 indx_t ptr, i, numkeys;
1808 node = NODEPTR(mp, indx);
1809 ptr = mp->mp_ptrs[indx];
1810 DPRINTF("update key %u (ofs %u) [%.*s] to [%.*s] on page %lu",
1812 (int)node->mn_ksize, (char *)NODEKEY(node),
1813 (int)key->mv_size, (char *)key->mv_data,
1816 delta = key->mv_size - node->mn_ksize;
1818 if (delta > 0 && SIZELEFT(mp) < delta) {
1819 DPRINTF("OUCH! Not enough room, delta = %d", delta);
1823 numkeys = NUMKEYS(mp);
1824 for (i = 0; i < numkeys; i++) {
1825 if (mp->mp_ptrs[i] <= ptr)
1826 mp->mp_ptrs[i] -= delta;
1829 base = (char *)mp + mp->mp_upper;
1830 len = ptr - mp->mp_upper + NODESIZE;
1831 bcopy(base, base - delta, len);
1832 mp->mp_upper -= delta;
1834 node = NODEPTR(mp, indx);
1835 node->mn_ksize = key->mv_size;
1838 bcopy(key->mv_data, NODEKEY(node), key->mv_size);
1843 /* Move a node from src to dst.
1846 mdb_move_node(MDB_db *bt, MDB_pageparent *src, indx_t srcindx,
1847 MDB_pageparent *dst, indx_t dstindx)
1853 srcnode = NODEPTR(src->mp_page, srcindx);
1854 DPRINTF("moving %s node %u [%.*s] on page %lu to node %u on page %lu",
1855 IS_LEAF(src->mp_page) ? "leaf" : "branch",
1857 (int)srcnode->mn_ksize, (char *)NODEKEY(srcnode),
1858 src->mp_page->mp_pgno,
1859 dstindx, dst->mp_page->mp_pgno);
1861 /* Mark src and dst as dirty. */
1862 if ((rc = mdb_touch(bt->md_env->me_txn, src)) ||
1863 (rc = mdb_touch(bt->md_env->me_txn, dst)))
1866 /* Add the node to the destination page.
1868 key.mv_size = srcnode->mn_ksize;
1869 key.mv_data = NODEKEY(srcnode);
1870 data.mv_size = NODEDSZ(srcnode);
1871 data.mv_data = NODEDATA(srcnode);
1872 rc = mdb_add_node(bt, dst->mp_page, dstindx, &key, &data, NODEPGNO(srcnode),
1874 if (rc != MDB_SUCCESS)
1877 /* Delete the node from the source page.
1879 mdb_del_node(bt, src->mp_page, srcindx);
1881 /* Update the parent separators.
1883 if (srcindx == 0 && src->mp_pi != 0) {
1884 DPRINTF("update separator for source page %lu to [%.*s]",
1885 src->mp_page->mp_pgno, (int)key.mv_size, (char *)key.mv_data);
1886 if ((rc = mdb_update_key(bt, src->mp_parent, src->mp_pi,
1887 &key)) != MDB_SUCCESS)
1891 if (srcindx == 0 && IS_BRANCH(src->mp_page)) {
1893 nullkey.mv_size = 0;
1894 assert(mdb_update_key(bt, src->mp_page, 0, &nullkey) == MDB_SUCCESS);
1897 if (dstindx == 0 && dst->mp_pi != 0) {
1898 DPRINTF("update separator for destination page %lu to [%.*s]",
1899 dst->mp_page->mp_pgno, (int)key.mv_size, (char *)key.mv_data);
1900 if ((rc = mdb_update_key(bt, dst->mp_parent, dst->mp_pi,
1901 &key)) != MDB_SUCCESS)
1905 if (dstindx == 0 && IS_BRANCH(dst->mp_page)) {
1907 nullkey.mv_size = 0;
1908 assert(mdb_update_key(bt, dst->mp_page, 0, &nullkey) == MDB_SUCCESS);
1915 mdb_merge(MDB_db *bt, MDB_pageparent *src, MDB_pageparent *dst)
1924 DPRINTF("merging page %lu and %lu", src->mp_page->mp_pgno, dst->mp_page->mp_pgno);
1926 assert(src->mp_parent); /* can't merge root page */
1927 assert(dst->mp_parent);
1928 assert(bt->md_env->me_txn != NULL);
1930 /* Mark src and dst as dirty. */
1931 if ((rc = mdb_touch(bt->md_env->me_txn, src)) ||
1932 (rc = mdb_touch(bt->md_env->me_txn, dst)))
1935 /* Move all nodes from src to dst.
1937 for (i = 0; i < NUMKEYS(src->mp_page); i++) {
1938 srcnode = NODEPTR(src->mp_page, i);
1940 key.mv_size = srcnode->mn_ksize;
1941 key.mv_data = NODEKEY(srcnode);
1942 data.mv_size = NODEDSZ(srcnode);
1943 data.mv_data = NODEDATA(srcnode);
1944 rc = mdb_add_node(bt, dst->mp_page, NUMKEYS(dst->mp_page), &key,
1945 &data, NODEPGNO(srcnode), srcnode->mn_flags);
1946 if (rc != MDB_SUCCESS)
1950 DPRINTF("dst page %lu now has %lu keys (%.1f%% filled)",
1951 dst->mp_page->mp_pgno, NUMKEYS(dst->mp_page), (float)PAGEFILL(bt->md_env, dst->mp_page) / 10);
1953 /* Unlink the src page from parent.
1955 mdb_del_node(bt, src->mp_parent, src->mp_pi);
1956 if (src->mp_pi == 0) {
1958 if ((rc = mdb_update_key(bt, src->mp_parent, 0, &key)) != MDB_SUCCESS)
1962 if (IS_LEAF(src->mp_page))
1963 bt->md_env->me_meta.mm_stat.ms_leaf_pages--;
1965 bt->md_env->me_meta.mm_stat.ms_branch_pages--;
1967 mpp.mp_page = src->mp_parent;
1968 dh = (MDB_dhead *)src->mp_parent;
1970 mpp.mp_parent = dh->md_parent;
1971 mpp.mp_pi = dh->md_pi;
1973 return mdb_rebalance(bt, &mpp);
1976 #define FILL_THRESHOLD 250
1979 mdb_rebalance(MDB_db *db, MDB_pageparent *mpp)
1984 indx_t si = 0, di = 0;
1987 assert(db->md_env->me_txn != NULL);
1988 assert(mpp != NULL);
1990 DPRINTF("rebalancing %s page %lu (has %lu keys, %.1f%% full)",
1991 IS_LEAF(mpp->mp_page) ? "leaf" : "branch",
1992 mpp->mp_page->mp_pgno, NUMKEYS(mpp->mp_page), (float)PAGEFILL(db->md_env, mpp->mp_page) / 10);
1994 if (PAGEFILL(db->md_env, mpp->mp_page) >= FILL_THRESHOLD) {
1995 DPRINTF("no need to rebalance page %lu, above fill threshold",
1996 mpp->mp_page->mp_pgno);
2000 if (mpp->mp_parent == NULL) {
2001 if (NUMKEYS(mpp->mp_page) == 0) {
2002 DPRINTF("tree is completely empty");
2003 db->md_env->me_txn->mt_root = P_INVALID;
2004 db->md_env->me_meta.mm_stat.ms_depth--;
2005 db->md_env->me_meta.mm_stat.ms_leaf_pages--;
2006 } else if (IS_BRANCH(mpp->mp_page) && NUMKEYS(mpp->mp_page) == 1) {
2007 DPRINTF("collapsing root page!");
2008 db->md_env->me_txn->mt_root = NODEPGNO(NODEPTR(mpp->mp_page, 0));
2009 if ((root = mdbenv_get_page(db->md_env, db->md_env->me_txn->mt_root)) == NULL)
2011 db->md_env->me_meta.mm_stat.ms_depth--;
2012 db->md_env->me_meta.mm_stat.ms_branch_pages--;
2014 DPRINTF("root page doesn't need rebalancing");
2018 /* The parent (branch page) must have at least 2 pointers,
2019 * otherwise the tree is invalid.
2021 assert(NUMKEYS(mpp->mp_parent) > 1);
2023 /* Leaf page fill factor is below the threshold.
2024 * Try to move keys from left or right neighbor, or
2025 * merge with a neighbor page.
2030 if (mpp->mp_pi == 0) {
2031 /* We're the leftmost leaf in our parent.
2033 DPRINTF("reading right neighbor");
2034 node = NODEPTR(mpp->mp_parent, mpp->mp_pi + 1);
2035 if ((npp.mp_page = mdbenv_get_page(db->md_env, NODEPGNO(node))) == NULL)
2037 npp.mp_pi = mpp->mp_pi + 1;
2039 di = NUMKEYS(mpp->mp_page);
2041 /* There is at least one neighbor to the left.
2043 DPRINTF("reading left neighbor");
2044 node = NODEPTR(mpp->mp_parent, mpp->mp_pi - 1);
2045 if ((npp.mp_page = mdbenv_get_page(db->md_env, NODEPGNO(node))) == NULL)
2047 npp.mp_pi = mpp->mp_pi - 1;
2048 si = NUMKEYS(npp.mp_page) - 1;
2051 npp.mp_parent = mpp->mp_parent;
2053 DPRINTF("found neighbor page %lu (%lu keys, %.1f%% full)",
2054 npp.mp_page->mp_pgno, NUMKEYS(npp.mp_page), (float)PAGEFILL(db->md_env, npp.mp_page) / 10);
2056 /* If the neighbor page is above threshold and has at least two
2057 * keys, move one key from it.
2059 * Otherwise we should try to merge them.
2061 if (PAGEFILL(db->md_env, npp.mp_page) >= FILL_THRESHOLD && NUMKEYS(npp.mp_page) >= 2)
2062 return mdb_move_node(db, &npp, si, mpp, di);
2063 else { /* FIXME: if (has_enough_room()) */
2064 if (mpp->mp_pi == 0)
2065 return mdb_merge(db, &npp, mpp);
2067 return mdb_merge(db, mpp, &npp);
2072 mdb_del(MDB_db *bt, MDB_txn *txn,
2073 MDB_val *key, MDB_val *data)
2080 DPRINTF("========> delete key %.*s", (int)key->mv_size, (char *)key->mv_data);
2082 assert(key != NULL);
2084 if (bt == NULL || txn == NULL)
2087 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
2091 if (bt->md_env->me_txn != txn) {
2095 if (key->mv_size == 0 || key->mv_size > MAXKEYSIZE) {
2099 if ((rc = mdb_search_page(bt, txn, key, NULL, 1, &mpp)) != MDB_SUCCESS)
2102 leaf = mdb_search_node(bt, mpp.mp_page, key, &exact, &ki);
2103 if (leaf == NULL || !exact) {
2107 if (data && (rc = mdb_read_data(bt, NULL, leaf, data)) != MDB_SUCCESS)
2110 mdb_del_node(bt, mpp.mp_page, ki);
2111 bt->md_env->me_meta.mm_stat.ms_entries--;
2112 rc = mdb_rebalance(bt, &mpp);
2113 if (rc != MDB_SUCCESS)
2114 txn->mt_flags |= MDB_TXN_ERROR;
2119 /* Split page <*mpp>, and insert <key,(data|newpgno)> in either left or
2120 * right sibling, at index <*newindxp> (as if unsplit). Updates *mpp and
2121 * *newindxp with the actual values after split, ie if *mpp and *newindxp
2122 * refer to a node in the new right sibling page.
2125 mdb_split(MDB_db *bt, MDB_page **mpp, unsigned int *newindxp,
2126 MDB_val *newkey, MDB_val *newdata, pgno_t newpgno)
2129 int rc = MDB_SUCCESS, ins_new = 0;
2132 unsigned int i, j, split_indx;
2134 MDB_page *pright, *p;
2135 MDB_val sepkey, rkey, rdata;
2137 MDB_dpage *mdp, *rdp, *pdp;
2141 assert(bt->md_env != NULL);
2143 dh = ((MDB_dhead *)*mpp) - 1;
2144 mdp = (MDB_dpage *)dh;
2145 newindx = *newindxp;
2147 DPRINTF("-----> splitting %s page %lu and adding [%.*s] at index %i",
2148 IS_LEAF(&mdp->p) ? "leaf" : "branch", mdp->p.mp_pgno,
2149 (int)newkey->mv_size, (char *)newkey->mv_data, *newindxp);
2151 if (mdp->h.md_parent == NULL) {
2152 if ((pdp = mdbenv_new_page(bt->md_env, P_BRANCH, 1)) == NULL)
2155 mdp->h.md_parent = &pdp->p;
2156 bt->md_env->me_txn->mt_root = pdp->p.mp_pgno;
2157 DPRINTF("root split! new root = %lu", pdp->p.mp_pgno);
2158 bt->md_env->me_meta.mm_stat.ms_depth++;
2160 /* Add left (implicit) pointer. */
2161 if (mdb_add_node(bt, &pdp->p, 0, NULL, NULL,
2162 mdp->p.mp_pgno, 0) != MDB_SUCCESS)
2165 DPRINTF("parent branch page is %lu", mdp->h.md_parent->mp_pgno);
2168 /* Create a right sibling. */
2169 if ((rdp = mdbenv_new_page(bt->md_env, mdp->p.mp_flags, 1)) == NULL)
2171 rdp->h.md_parent = mdp->h.md_parent;
2172 rdp->h.md_pi = mdp->h.md_pi + 1;
2173 DPRINTF("new right sibling: page %lu", rdp->p.mp_pgno);
2175 /* Move half of the keys to the right sibling. */
2176 if ((copy = malloc(bt->md_env->me_head.mh_psize)) == NULL)
2178 bcopy(&mdp->p, copy, bt->md_env->me_head.mh_psize);
2179 bzero(&mdp->p.mp_ptrs, bt->md_env->me_head.mh_psize - PAGEHDRSZ);
2180 mdp->p.mp_lower = PAGEHDRSZ;
2181 mdp->p.mp_upper = bt->md_env->me_head.mh_psize;
2183 split_indx = NUMKEYS(copy) / 2 + 1;
2185 /* First find the separating key between the split pages.
2187 bzero(&sepkey, sizeof(sepkey));
2188 if (newindx == split_indx) {
2189 sepkey.mv_size = newkey->mv_size;
2190 sepkey.mv_data = newkey->mv_data;
2192 node = NODEPTR(copy, split_indx);
2193 sepkey.mv_size = node->mn_ksize;
2194 sepkey.mv_data = NODEKEY(node);
2197 DPRINTF("separator is [%.*s]", (int)sepkey.mv_size, (char *)sepkey.mv_data);
2199 /* Copy separator key to the parent.
2201 if (SIZELEFT(rdp->h.md_parent) < mdb_branch_size(bt, &sepkey)) {
2202 rc = mdb_split(bt, &rdp->h.md_parent, &rdp->h.md_pi,
2203 &sepkey, NULL, rdp->p.mp_pgno);
2205 /* Right page might now have changed parent.
2206 * Check if left page also changed parent.
2208 if (rdp->h.md_parent != mdp->h.md_parent &&
2209 mdp->h.md_pi >= NUMKEYS(mdp->h.md_parent)) {
2210 mdp->h.md_parent = rdp->h.md_parent;
2211 mdp->h.md_pi = rdp->h.md_pi - 1;
2214 rc = mdb_add_node(bt, rdp->h.md_parent, rdp->h.md_pi,
2215 &sepkey, NULL, rdp->p.mp_pgno, 0);
2217 if (rc != MDB_SUCCESS) {
2222 for (i = j = 0; i <= NUMKEYS(copy); j++) {
2223 if (i < split_indx) {
2224 /* Re-insert in left sibling. */
2227 /* Insert in right sibling. */
2228 if (i == split_indx)
2229 /* Reset insert index for right sibling. */
2230 j = (i == newindx && ins_new);
2234 if (i == newindx && !ins_new) {
2235 /* Insert the original entry that caused the split. */
2236 rkey.mv_data = newkey->mv_data;
2237 rkey.mv_size = newkey->mv_size;
2238 if (IS_LEAF(&mdp->p)) {
2239 rdata.mv_data = newdata->mv_data;
2240 rdata.mv_size = newdata->mv_size;
2247 /* Update page and index for the new key. */
2250 } else if (i == NUMKEYS(copy)) {
2253 node = NODEPTR(copy, i);
2254 rkey.mv_data = NODEKEY(node);
2255 rkey.mv_size = node->mn_ksize;
2256 if (IS_LEAF(&mdp->p)) {
2257 rdata.mv_data = NODEDATA(node);
2258 rdata.mv_size = node->mn_dsize;
2260 pgno = node->mn_pgno;
2261 flags = node->mn_flags;
2266 if (!IS_LEAF(&mdp->p) && j == 0) {
2267 /* First branch index doesn't need key data. */
2271 rc = mdb_add_node(bt, &pdp->p, j, &rkey, &rdata, pgno,flags);
2279 mdb_put(MDB_db *bt, MDB_txn *txn,
2280 MDB_val *key, MDB_val *data, unsigned int flags)
2282 int rc = MDB_SUCCESS, exact;
2287 assert(key != NULL);
2288 assert(data != NULL);
2290 if (bt == NULL || txn == NULL)
2293 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
2297 if (bt->md_env->me_txn != txn) {
2301 if (key->mv_size == 0 || key->mv_size > MAXKEYSIZE) {
2305 DPRINTF("==> put key %.*s, size %zu, data size %zu",
2306 (int)key->mv_size, (char *)key->mv_data, key->mv_size, data->mv_size);
2308 rc = mdb_search_page(bt, txn, key, NULL, 1, &mpp);
2309 if (rc == MDB_SUCCESS) {
2310 leaf = mdb_search_node(bt, mpp.mp_page, key, &exact, &ki);
2311 if (leaf && exact) {
2312 if (F_ISSET(flags, MDB_NOOVERWRITE)) {
2313 DPRINTF("duplicate key %.*s",
2314 (int)key->mv_size, (char *)key->mv_data);
2317 mdb_del_node(bt, mpp.mp_page, ki);
2319 if (leaf == NULL) { /* append if not found */
2320 ki = NUMKEYS(mpp.mp_page);
2321 DPRINTF("appending key at index %i", ki);
2323 } else if (rc == ENOENT) {
2325 /* new file, just write a root leaf page */
2326 DPRINTF("allocating new root leaf page");
2327 if ((dp = mdbenv_new_page(bt->md_env, P_LEAF, 1)) == NULL) {
2330 mpp.mp_page = &dp->p;
2331 txn->mt_root = mpp.mp_page->mp_pgno;
2332 bt->md_env->me_meta.mm_stat.ms_depth++;
2338 assert(IS_LEAF(mpp.mp_page));
2339 DPRINTF("there are %lu keys, should insert new key at index %i",
2340 NUMKEYS(mpp.mp_page), ki);
2342 if (SIZELEFT(mpp.mp_page) < mdb_leaf_size(bt, key, data)) {
2343 rc = mdb_split(bt, &mpp.mp_page, &ki, key, data, P_INVALID);
2345 /* There is room already in this leaf page. */
2346 rc = mdb_add_node(bt, mpp.mp_page, ki, key, data, 0, 0);
2349 if (rc != MDB_SUCCESS)
2350 txn->mt_flags |= MDB_TXN_ERROR;
2352 bt->md_env->me_meta.mm_stat.ms_entries++;
2359 mdbenv_get_flags(MDB_env *env, unsigned int *arg)
2364 *arg = env->me_flags;
2369 mdbenv_get_path(MDB_env *env, const char **arg)
2374 *arg = env->me_path;
2379 mdbenv_stat(MDB_env *env, MDB_stat **arg)
2381 if (env == NULL || arg == NULL)
2384 *arg = &env->me_meta.mm_stat;
2389 int mdb_open(MDB_env *env, MDB_txn *txn, const char *name, unsigned int flags, MDB_db **db)
2392 *db = (MDB_db *)&env->me_db;
2398 int mdb_stat(MDB_db *db, MDB_stat **arg)
2400 if (db == NULL || arg == NULL)
2403 *arg = &db->md_stat;
2408 void mdb_close(MDB_db *db)