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 ((unsigned) 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) (1000L * ((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)
336 int diff, len_diff = -1;
339 len_diff = (n1 > n2);
342 diff = memcmp(s1, s2, n1);
343 return diff ? diff : len_diff;
347 memnrcmp(const void *s1, size_t n1, const void *s2, size_t n2)
349 const unsigned char *p1, *p2, *p1_lim;
356 p1 = (const unsigned char *)s1 + n1 - 1;
357 p2 = (const unsigned char *)s2 + n2 - 1;
359 for (p1_lim = (n1 <= n2 ? s1 : s2); *p1 == *p2; p1--, p2--) {
361 return (p1 != s1) ? (p1 != p2) : (p2 != s2) ? -1 : 0;
367 mdb_cmp(MDB_db *db, const MDB_val *a, const MDB_val *b)
369 return db->md_cmp(a, b);
373 _mdb_cmp(MDB_db *db, const MDB_val *key1, const MDB_val *key2)
375 if (F_ISSET(db->md_flags, MDB_REVERSEKEY))
376 return memnrcmp(key1->mv_data, key1->mv_size, key2->mv_data, key2->mv_size);
378 return memncmp((char *)key1->mv_data, key1->mv_size, key2->mv_data, key2->mv_size);
381 /* Allocate new page(s) for writing */
383 mdb_newpage(MDB_txn *txn, MDB_page *parent, int parent_idx, int num)
387 if ((dp = malloc(txn->mt_env->me_head.mh_psize * num + sizeof(MDB_dhead))) == NULL)
390 dp->h.md_parent = parent;
391 dp->h.md_pi = parent_idx;
392 SIMPLEQ_INSERT_TAIL(txn->mt_u.dirty_queue, dp, h.md_next);
393 dp->p.mp_pgno = txn->mt_next_pgno;
394 txn->mt_next_pgno += num;
399 /* Touch a page: make it dirty and re-insert into tree with updated pgno.
402 mdb_touch(MDB_txn *txn, MDB_pageparent *pp)
405 MDB_page *mp = pp->mp_page;
410 if (!F_ISSET(mp->mp_flags, P_DIRTY)) {
412 DPRINTF("touching page %lu -> %lu", mp->mp_pgno, txn->mt_next_pgno);
413 if ((dp = mdb_newpage(txn, pp->mp_parent, pp->mp_pi, 1)) == NULL)
415 pgno = dp->p.mp_pgno;
416 bcopy(mp, &dp->p, txn->mt_env->me_head.mh_psize);
419 mp->mp_flags |= P_DIRTY;
421 /* Update the page number to new touched page. */
422 if (pp->mp_parent != NULL)
423 NODEPGNO(NODEPTR(pp->mp_parent, pp->mp_pi)) = mp->mp_pgno;
430 mdbenv_sync(MDB_env *env)
433 if (!F_ISSET(env->me_flags, MDB_NOSYNC)) {
434 if (fsync(env->me_fd))
441 mdb_txn_begin(MDB_env *env, int rdonly, MDB_txn **ret)
446 if ((txn = calloc(1, sizeof(*txn))) == NULL) {
447 DPRINTF("calloc: %s", strerror(errno));
452 txn->mt_flags |= MDB_TXN_RDONLY;
454 txn->mt_u.dirty_queue = calloc(1, sizeof(*txn->mt_u.dirty_queue));
455 if (txn->mt_u.dirty_queue == NULL) {
459 SIMPLEQ_INIT(txn->mt_u.dirty_queue);
461 pthread_mutex_lock(&env->me_txns->mt_wmutex);
462 env->me_txns->mt_txnid++;
464 txn->mt_txnid = env->me_txns->mt_txnid;
466 MDB_reader *r = pthread_getspecific(env->me_txkey);
469 pthread_mutex_lock(&env->me_txns->mt_mutex);
470 for (i=0; i<env->me_maxreaders; i++) {
471 if (env->me_txns->mt_readers[i].mr_pid == 0) {
472 env->me_txns->mt_readers[i].mr_pid = getpid();
473 env->me_txns->mt_readers[i].mr_tid = pthread_self();
474 pthread_setspecific(env->me_txkey, &env->me_txns->mt_readers[i]);
475 if (i >= env->me_txns->mt_numreaders)
476 env->me_txns->mt_numreaders = i+1;
480 pthread_mutex_unlock(&env->me_txns->mt_mutex);
481 if (i == env->me_maxreaders) {
485 r->mr_txnid = txn->mt_txnid;
486 txn->mt_u.reader = r;
493 if ((rc = mdbenv_read_meta(env)) != MDB_SUCCESS) {
498 txn->mt_next_pgno = env->me_meta.mm_last_pg+1;
499 txn->mt_first_pgno = txn->mt_next_pgno;
500 txn->mt_root = env->me_meta.mm_root;
501 DPRINTF("begin transaction %lu on mdbenv %p, root page %lu", txn->mt_txnid, env, txn->mt_root);
508 mdb_txn_abort(MDB_txn *txn)
517 DPRINTF("abort transaction %lu on mdbenv %p, root page %lu", txn->mt_txnid, env, txn->mt_root);
519 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
520 txn->mt_u.reader->mr_txnid = 0;
522 /* Discard all dirty pages.
524 while (!SIMPLEQ_EMPTY(txn->mt_u.dirty_queue)) {
525 dp = SIMPLEQ_FIRST(txn->mt_u.dirty_queue);
526 SIMPLEQ_REMOVE_HEAD(txn->mt_u.dirty_queue, h.md_next);
531 DPRINTF("releasing write lock on txn %p", txn);
533 if (flock(txn->bt->fd, LOCK_UN) != 0) {
534 DPRINTF("failed to unlock fd %d: %s",
535 txn->bt->fd, strerror(errno));
538 free(txn->mt_u.dirty_queue);
540 pthread_mutex_unlock(&env->me_txns->mt_wmutex);
547 mdb_txn_commit(MDB_txn *txn)
555 struct iovec iov[MDB_COMMIT_PAGES];
558 assert(txn->mt_env != NULL);
562 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
563 DPRINTF("attempt to commit read-only transaction");
568 if (txn != env->me_txn) {
569 DPRINTF("attempt to commit unknown transaction");
574 if (F_ISSET(txn->mt_flags, MDB_TXN_ERROR)) {
575 DPRINTF("error flag is set, can't commit");
580 if (SIMPLEQ_EMPTY(txn->mt_u.dirty_queue))
583 DPRINTF("committing transaction %lu on mdbenv %p, root page %lu",
584 txn->mt_txnid, env, txn->mt_root);
586 /* Commit up to MDB_COMMIT_PAGES dirty pages to disk until done.
593 SIMPLEQ_FOREACH(dp, txn->mt_u.dirty_queue, h.md_next) {
594 if (dp->p.mp_pgno != next) {
595 lseek(env->me_fd, dp->p.mp_pgno * env->me_head.mh_psize, SEEK_SET);
596 next = dp->p.mp_pgno;
600 DPRINTF("committing page %lu", dp->p.mp_pgno);
601 iov[n].iov_len = env->me_head.mh_psize * dp->h.md_num;
602 iov[n].iov_base = &dp->p;
603 size += iov[n].iov_len;
604 next = dp->p.mp_pgno + dp->h.md_num;
605 /* clear dirty flag */
606 dp->p.mp_flags &= ~P_DIRTY;
607 if (++n >= MDB_COMMIT_PAGES) {
616 DPRINTF("committing %u dirty pages", n);
617 rc = writev(env->me_fd, iov, n);
621 DPRINTF("short write, filesystem full?");
623 DPRINTF("writev: %s", strerror(errno));
628 /* Drop the dirty pages.
630 while (!SIMPLEQ_EMPTY(txn->mt_u.dirty_queue)) {
631 dp = SIMPLEQ_FIRST(txn->mt_u.dirty_queue);
632 SIMPLEQ_REMOVE_HEAD(txn->mt_u.dirty_queue, h.md_next);
639 if ((n = mdbenv_sync(env)) != 0 ||
640 (n = mdbenv_write_meta(txn)) != MDB_SUCCESS ||
641 (n = mdbenv_sync(env)) != 0) {
646 pthread_mutex_unlock(&env->me_txns->mt_wmutex);
647 free(txn->mt_u.dirty_queue);
658 mdbenv_write_header(MDB_env *env)
666 DPRINTF("writing header page");
669 psize = sysconf(_SC_PAGE_SIZE);
671 if ((p = calloc(1, psize)) == NULL)
673 p->mp_flags = P_HEAD;
675 env->me_head.mh_psize = psize;
676 env->me_head.mh_flags = env->me_flags & 0xffff;
679 bcopy(&env->me_head, h, sizeof(*h));
681 rc = write(env->me_fd, p, env->me_head.mh_psize);
683 if (rc != (ssize_t)env->me_head.mh_psize) {
686 DPRINTF("short write, filesystem full?");
694 mdbenv_read_header(MDB_env *env)
703 /* We don't know the page size yet, so use a minimum value.
706 if ((rc = pread(env->me_fd, page, PAGESIZE, 0)) == 0) {
708 } else if (rc != PAGESIZE) {
711 DPRINTF("read: %s", strerror(errno));
715 p = (MDB_page *)page;
717 if (!F_ISSET(p->mp_flags, P_HEAD)) {
718 DPRINTF("page %lu not a header page", p->mp_pgno);
723 if (h->mh_magic != MDB_MAGIC) {
724 DPRINTF("header has invalid magic");
728 if (h->mh_version != MDB_VERSION) {
729 DPRINTF("database is version %u, expected version %u",
730 h->mh_version, MDB_VERSION);
734 bcopy(h, &env->me_head, sizeof(*h));
739 mdbenv_init_meta(MDB_env *env)
745 p = calloc(2, env->me_head.mh_psize);
747 p->mp_flags = P_META;
750 meta->mm_root = P_INVALID;
751 meta->mm_last_pg = 2;
753 q = (MDB_page *)((char *)p + env->me_head.mh_psize);
756 q->mp_flags = P_META;
759 meta->mm_root = P_INVALID;
760 meta->mm_last_pg = 2;
762 rc = write(env->me_fd, p, env->me_head.mh_psize * 2);
764 return (rc == env->me_head.mh_psize * 2) ? MDB_SUCCESS : errno;
768 mdbenv_write_meta(MDB_txn *txn)
776 assert(txn->mt_env != NULL);
778 DPRINTF("writing meta page for root page %lu", txn->mt_root);
782 bcopy(&env->me_meta, &meta, sizeof(meta));
783 meta.mm_root = txn->mt_root;
784 meta.mm_last_pg = txn->mt_next_pgno - 1;
785 meta.mm_txnid = txn->mt_txnid;
787 off = env->me_head.mh_psize;
788 if (!env->me_metatoggle)
792 lseek(env->me_fd, off, SEEK_SET);
793 rc = write(env->me_fd, &meta, sizeof(meta));
794 if (rc != sizeof(meta)) {
795 DPRINTF("write failed, disk error?");
802 /* Returns true if page p is a valid meta page, false otherwise.
805 mdb_check_meta_page(MDB_page *p)
807 if (!F_ISSET(p->mp_flags, P_META)) {
808 DPRINTF("page %lu not a meta page", p->mp_pgno);
816 mdbenv_read_meta(MDB_env *env)
824 if ((mp0 = mdbenv_get_page(env, 1)) == NULL ||
825 (mp1 = mdbenv_get_page(env, 2)) == NULL)
828 rc = mdb_check_meta_page(mp0);
831 rc = mdb_check_meta_page(mp1);
834 meta[0] = METADATA(mp0);
835 meta[1] = METADATA(mp1);
837 if (meta[0]->mm_txnid < meta[1]->mm_txnid)
840 if (meta[toggle]->mm_txnid > env->me_meta.mm_txnid) {
841 bcopy(meta[toggle], &env->me_meta, sizeof(env->me_meta));
842 env->me_metatoggle = toggle;
845 DPRINTF("Using meta page %d", toggle);
851 mdbenv_create(MDB_env **env)
855 e = calloc(1, sizeof(*e));
856 if (!e) return ENOMEM;
858 e->me_head.mh_magic = MDB_MAGIC;
859 e->me_head.mh_version = MDB_VERSION;
860 e->me_head.mh_mapsize = DEFAULT_MAPSIZE;
861 e->me_maxreaders = DEFAULT_READERS;
868 mdbenv_set_mapsize(MDB_env *env, size_t size)
872 env->me_mapsize = env->me_head.mh_mapsize = size;
877 mdbenv_set_maxreaders(MDB_env *env, int readers)
879 env->me_maxreaders = readers;
884 mdbenv_get_maxreaders(MDB_env *env, int *readers)
886 if (!env || !readers)
888 *readers = env->me_maxreaders;
893 mdbenv_open2(MDB_env *env, unsigned int flags)
897 env->me_flags = flags;
898 env->me_meta.mm_root = P_INVALID;
899 env->me_meta.mm_last_pg = 2;
901 if ((i = mdbenv_read_header(env)) != 0) {
904 DPRINTF("new mdbenv");
908 if (!env->me_mapsize)
909 env->me_mapsize = env->me_head.mh_mapsize;
912 if (env->me_head.mh_address && (flags & MDB_FIXEDMAP))
914 env->me_map = mmap(env->me_head.mh_address, env->me_mapsize, PROT_READ, i,
916 if (env->me_map == MAP_FAILED)
920 env->me_head.mh_mapsize = env->me_mapsize;
921 if (flags & MDB_FIXEDMAP)
922 env->me_head.mh_address = env->me_map;
923 i = mdbenv_write_header(env);
924 if (i != MDB_SUCCESS) {
925 munmap(env->me_map, env->me_mapsize);
928 i = mdbenv_init_meta(env);
929 if (i != MDB_SUCCESS) {
930 munmap(env->me_map, env->me_mapsize);
935 if ((i = mdbenv_read_meta(env)) != 0)
938 DPRINTF("opened database version %u, pagesize %u",
939 env->me_head.mh_version, env->me_head.mh_psize);
940 DPRINTF("depth: %u", env->me_meta.mm_stat.ms_depth);
941 DPRINTF("entries: %lu", env->me_meta.mm_stat.ms_entries);
942 DPRINTF("branch pages: %lu", env->me_meta.mm_stat.ms_branch_pages);
943 DPRINTF("leaf pages: %lu", env->me_meta.mm_stat.ms_leaf_pages);
944 DPRINTF("overflow pages: %lu", env->me_meta.mm_stat.ms_overflow_pages);
945 DPRINTF("root: %lu", env->me_meta.mm_root);
951 mdbenv_reader_dest(void *ptr)
953 MDB_reader *reader = ptr;
955 reader->mr_txnid = 0;
961 mdbenv_open(MDB_env *env, const char *path, unsigned int flags, mode_t mode)
963 int oflags, rc, shmid;
967 if (F_ISSET(flags, MDB_RDONLY))
970 oflags = O_RDWR | O_CREAT;
972 if ((env->me_fd = open(path, oflags, mode)) == -1)
975 env->me_shmkey = ftok(path, 'm');
976 size = (env->me_maxreaders-1) * sizeof(MDB_reader) + sizeof(MDB_txninfo);
977 shmid = shmget(env->me_shmkey, size, IPC_CREAT|IPC_EXCL|mode);
979 if (errno == EEXIST) {
980 shmid = shmget(env->me_shmkey, size, IPC_CREAT|mode);
983 env->me_txns = shmat(shmid, NULL, 0);
984 if (env->me_txns->mt_magic != MDB_MAGIC ||
985 env->me_txns->mt_version != MDB_VERSION) {
986 DPRINTF("invalid lock region %d", shmid);
995 pthread_mutexattr_t mattr;
997 env->me_txns = shmat(shmid, NULL, 0);
998 pthread_mutexattr_init(&mattr);
999 pthread_mutexattr_setpshared(&mattr, PTHREAD_PROCESS_SHARED);
1000 pthread_mutex_init(&env->me_txns->mt_mutex, &mattr);
1001 pthread_mutex_init(&env->me_txns->mt_wmutex, &mattr);
1002 env->me_txns->mt_version = MDB_VERSION;
1003 env->me_txns->mt_magic = MDB_MAGIC;
1006 if ((rc = mdbenv_open2(env, flags)) != MDB_SUCCESS) {
1010 env->me_path = strdup(path);
1011 DPRINTF("opened dbenv %p", env);
1014 pthread_key_create(&env->me_txkey, mdbenv_reader_dest);
1020 mdbenv_close(MDB_env *env)
1028 munmap(env->me_map, env->me_mapsize);
1032 shmdt(env->me_txns);
1037 /* Search for key within a leaf page, using binary search.
1038 * Returns the smallest entry larger or equal to the key.
1039 * If exactp is non-null, stores whether the found entry was an exact match
1040 * in *exactp (1 or 0).
1041 * If kip is non-null, stores the index of the found entry in *kip.
1042 * If no entry larger or equal to the key is found, returns NULL.
1045 mdb_search_node(MDB_db *bt, MDB_page *mp, MDB_val *key,
1046 int *exactp, unsigned int *kip)
1054 DPRINTF("searching %u keys in %s page %lu",
1056 IS_LEAF(mp) ? "leaf" : "branch",
1059 assert(NUMKEYS(mp) > 0);
1061 bzero(&nodekey, sizeof(nodekey));
1063 low = IS_LEAF(mp) ? 0 : 1;
1064 high = NUMKEYS(mp) - 1;
1065 while (low <= high) {
1066 i = (low + high) >> 1;
1067 node = NODEPTR(mp, i);
1069 nodekey.mv_size = node->mn_ksize;
1070 nodekey.mv_data = NODEKEY(node);
1073 rc = bt->md_cmp(key, &nodekey);
1075 rc = _mdb_cmp(bt, key, &nodekey);
1078 DPRINTF("found leaf index %u [%.*s], rc = %i",
1079 i, (int)nodekey.mv_size, (char *)nodekey.mv_data, rc);
1081 DPRINTF("found branch index %u [%.*s -> %lu], rc = %i",
1082 i, (int)node->mn_ksize, (char *)NODEKEY(node),
1093 if (rc > 0) { /* Found entry is less than the key. */
1094 i++; /* Skip to get the smallest entry larger than key. */
1095 if (i >= NUMKEYS(mp))
1096 /* There is no entry larger or equal to the key. */
1100 *exactp = (rc == 0);
1101 if (kip) /* Store the key index if requested. */
1104 return NODEPTR(mp, i);
1108 cursor_pop_page(MDB_cursor *cursor)
1112 top = CURSOR_TOP(cursor);
1115 DPRINTF("popped page %lu off cursor %p", top->mp_page->mp_pgno, cursor);
1121 cursor_push_page(MDB_cursor *cursor, MDB_page *mp)
1125 DPRINTF("pushing page %lu on cursor %p", mp->mp_pgno, cursor);
1127 if ((ppage = calloc(1, sizeof(*ppage))) == NULL)
1129 ppage->mp_page = mp;
1130 CURSOR_PUSH(cursor, ppage);
1135 mdbenv_get_page(MDB_env *env, pgno_t pgno)
1138 MDB_txn *txn = env->me_txn;
1140 if (txn && pgno >= txn->mt_first_pgno &&
1141 !SIMPLEQ_EMPTY(txn->mt_u.dirty_queue)) {
1143 SIMPLEQ_FOREACH(dp, txn->mt_u.dirty_queue, h.md_next) {
1144 if (dp->p.mp_pgno == pgno) {
1150 p = (MDB_page *)(env->me_map + env->me_head.mh_psize * pgno);
1156 mdb_search_page_root(MDB_db *bt, MDB_val *key,
1157 MDB_cursor *cursor, int modify, MDB_pageparent *mpp)
1159 MDB_page *mp = mpp->mp_page;
1162 if (cursor && cursor_push_page(cursor, mp) == NULL)
1165 while (IS_BRANCH(mp)) {
1169 DPRINTF("branch page %lu has %u keys", mp->mp_pgno, NUMKEYS(mp));
1170 assert(NUMKEYS(mp) > 1);
1171 DPRINTF("found index 0 to page %lu", NODEPGNO(NODEPTR(mp, 0)));
1173 if (key == NULL) /* Initialize cursor to first page. */
1177 node = mdb_search_node(bt, mp, key, &exact, &i);
1179 i = NUMKEYS(mp) - 1;
1187 DPRINTF("following index %u for key %.*s",
1188 i, (int)key->mv_size, (char *)key->mv_data);
1189 assert(i >= 0 && i < NUMKEYS(mp));
1190 node = NODEPTR(mp, i);
1193 CURSOR_TOP(cursor)->mp_ki = i;
1195 mpp->mp_parent = mp;
1196 if ((mp = mdbenv_get_page(bt->md_env, NODEPGNO(node))) == NULL)
1201 if (cursor && cursor_push_page(cursor, mp) == NULL)
1205 MDB_dhead *dh = ((MDB_dhead *)mp)-1;
1206 if (rc = mdb_touch(bt->md_env->me_txn, mpp))
1208 dh = ((MDB_dhead *)mpp->mp_page)-1;
1209 dh->md_parent = mpp->mp_parent;
1210 dh->md_pi = mpp->mp_pi;
1217 DPRINTF("internal error, index points to a %02X page!?",
1222 DPRINTF("found leaf page %lu for key %.*s", mp->mp_pgno,
1223 key ? (int)key->mv_size : 0, key ? (char *)key->mv_data : NULL);
1228 /* Search for the page a given key should be in.
1229 * Stores a pointer to the found page in *mpp.
1230 * If key is NULL, search for the lowest page (used by mdb_cursor_first).
1231 * If cursor is non-null, pushes parent pages on the cursor stack.
1232 * If modify is true, visited pages are updated with new page numbers.
1235 mdb_search_page(MDB_db *db, MDB_txn *txn, MDB_val *key,
1236 MDB_cursor *cursor, int modify, MDB_pageparent *mpp)
1241 /* Can't modify pages outside a transaction. */
1242 if (txn == NULL && modify) {
1246 /* Choose which root page to start with. If a transaction is given
1247 * use the root page from the transaction, otherwise read the last
1248 * committed root page.
1251 if ((rc = mdbenv_read_meta(db->md_env)) != MDB_SUCCESS)
1253 root = db->md_env->me_meta.mm_root;
1254 } else if (F_ISSET(txn->mt_flags, MDB_TXN_ERROR)) {
1255 DPRINTF("transaction has failed, must abort");
1258 root = txn->mt_root;
1260 if (root == P_INVALID) { /* Tree is empty. */
1261 DPRINTF("tree is empty");
1265 if ((mpp->mp_page = mdbenv_get_page(db->md_env, root)) == NULL)
1268 DPRINTF("root page has flags 0x%X", mpp->mp_page->mp_flags);
1270 if (modify && !F_ISSET(mpp->mp_page->mp_flags, P_DIRTY)) {
1271 mpp->mp_parent = NULL;
1273 if ((rc = mdb_touch(txn, mpp)))
1275 txn->mt_root = mpp->mp_page->mp_pgno;
1278 return mdb_search_page_root(db, key, cursor, modify, mpp);
1282 mdb_read_data(MDB_db *db, MDB_page *mp, MDB_node *leaf,
1285 MDB_page *omp; /* overflow mpage */
1291 bzero(data, sizeof(*data));
1292 max = db->md_env->me_head.mh_psize - PAGEHDRSZ;
1294 if (!F_ISSET(leaf->mn_flags, F_BIGDATA)) {
1295 data->mv_size = leaf->mn_dsize;
1296 data->mv_data = NODEDATA(leaf);
1300 /* Read overflow data.
1302 data->mv_size = leaf->mn_dsize;
1303 bcopy(NODEDATA(leaf), &pgno, sizeof(pgno));
1304 if ((omp = mdbenv_get_page(db->md_env, pgno)) == NULL) {
1305 DPRINTF("read overflow page %lu failed", pgno);
1308 data->mv_data = omp;
1314 mdb_get(MDB_db *db, MDB_txn *txn,
1315 MDB_val *key, MDB_val *data)
1323 DPRINTF("===> get key [%.*s]", (int)key->mv_size, (char *)key->mv_data);
1328 if (txn != NULL && db->md_env != txn->mt_env) {
1332 if (key->mv_size == 0 || key->mv_size > MAXKEYSIZE) {
1336 if ((rc = mdb_search_page(db, txn, key, NULL, 0, &mpp)) != MDB_SUCCESS)
1339 leaf = mdb_search_node(db, mpp.mp_page, key, &exact, NULL);
1341 rc = mdb_read_data(db, mpp.mp_page, leaf, data);
1350 mdb_sibling(MDB_cursor *cursor, int move_right)
1354 MDB_ppage *parent, *top;
1357 top = CURSOR_TOP(cursor);
1358 if ((parent = SLIST_NEXT(top, mp_entry)) == NULL) {
1359 return ENOENT; /* root has no siblings */
1362 DPRINTF("parent page is page %lu, index %u",
1363 parent->mp_page->mp_pgno, parent->mp_ki);
1365 cursor_pop_page(cursor);
1366 if (move_right ? (parent->mp_ki + 1 >= NUMKEYS(parent->mp_page))
1367 : (parent->mp_ki == 0)) {
1368 DPRINTF("no more keys left, moving to %s sibling",
1369 move_right ? "right" : "left");
1370 if ((rc = mdb_sibling(cursor, move_right)) != MDB_SUCCESS)
1372 parent = CURSOR_TOP(cursor);
1378 DPRINTF("just moving to %s index key %u",
1379 move_right ? "right" : "left", parent->mp_ki);
1381 assert(IS_BRANCH(parent->mp_page));
1383 indx = NODEPTR(parent->mp_page, parent->mp_ki);
1384 if ((mp = mdbenv_get_page(cursor->mc_db->md_env, indx->mn_pgno)) == NULL)
1387 mp->parent = parent->mp_page;
1388 mp->parent_index = parent->mp_ki;
1391 cursor_push_page(cursor, mp);
1397 mdb_set_key(MDB_db *bt, MDB_page *mp, MDB_node *node,
1403 key->mv_size = node->mn_ksize;
1404 key->mv_data = NODEKEY(node);
1410 mdb_cursor_next(MDB_cursor *cursor, MDB_val *key, MDB_val *data)
1416 if (cursor->mc_eof) {
1420 assert(cursor->mc_initialized);
1422 top = CURSOR_TOP(cursor);
1425 DPRINTF("cursor_next: top page is %lu in cursor %p", mp->mp_pgno, cursor);
1427 if (top->mp_ki + 1 >= NUMKEYS(mp)) {
1428 DPRINTF("=====> move to next sibling page");
1429 if (mdb_sibling(cursor, 1) != MDB_SUCCESS) {
1433 top = CURSOR_TOP(cursor);
1435 DPRINTF("next page is %lu, key index %u", mp->mp_pgno, top->mp_ki);
1439 DPRINTF("==> cursor points to page %lu with %u keys, key index %u",
1440 mp->mp_pgno, NUMKEYS(mp), top->mp_ki);
1442 assert(IS_LEAF(mp));
1443 leaf = NODEPTR(mp, top->mp_ki);
1445 if (data && mdb_read_data(cursor->mc_db, mp, leaf, data) != MDB_SUCCESS)
1448 return mdb_set_key(cursor->mc_db, mp, leaf, key);
1452 mdb_cursor_set(MDB_cursor *cursor, MDB_val *key, MDB_val *data,
1462 assert(key->mv_size > 0);
1464 rc = mdb_search_page(cursor->mc_db, cursor->mc_txn, key, cursor, 0, &mpp);
1465 if (rc != MDB_SUCCESS)
1467 assert(IS_LEAF(mpp.mp_page));
1469 top = CURSOR_TOP(cursor);
1470 leaf = mdb_search_node(cursor->mc_db, mpp.mp_page, key, exactp, &top->mp_ki);
1471 if (exactp != NULL && !*exactp) {
1472 /* MDB_CURSOR_EXACT specified and not an exact match. */
1477 DPRINTF("===> inexact leaf not found, goto sibling");
1478 if ((rc = mdb_sibling(cursor, 1)) != MDB_SUCCESS)
1479 return rc; /* no entries matched */
1480 top = CURSOR_TOP(cursor);
1482 mpp.mp_page = top->mp_page;
1483 assert(IS_LEAF(mpp.mp_page));
1484 leaf = NODEPTR(mpp.mp_page, 0);
1487 cursor->mc_initialized = 1;
1490 if (data && (rc = mdb_read_data(cursor->mc_db, mpp.mp_page, leaf, data)) != MDB_SUCCESS)
1493 rc = mdb_set_key(cursor->mc_db, mpp.mp_page, leaf, key);
1494 if (rc == MDB_SUCCESS) {
1495 DPRINTF("==> cursor placed on key %.*s",
1496 (int)key->mv_size, (char *)key->mv_data);
1504 mdb_cursor_first(MDB_cursor *cursor, MDB_val *key, MDB_val *data)
1510 rc = mdb_search_page(cursor->mc_db, cursor->mc_txn, NULL, cursor, 0, &mpp);
1511 if (rc != MDB_SUCCESS)
1513 assert(IS_LEAF(mpp.mp_page));
1515 leaf = NODEPTR(mpp.mp_page, 0);
1516 cursor->mc_initialized = 1;
1519 if (data && (rc = mdb_read_data(cursor->mc_db, mpp.mp_page, leaf, data)) != MDB_SUCCESS)
1522 return mdb_set_key(cursor->mc_db, mpp.mp_page, leaf, key);
1526 mdb_cursor_get(MDB_cursor *cursor, MDB_val *key, MDB_val *data,
1536 case MDB_CURSOR_EXACT:
1537 while (CURSOR_TOP(cursor) != NULL)
1538 cursor_pop_page(cursor);
1539 if (key == NULL || key->mv_size == 0 || key->mv_size > MAXKEYSIZE) {
1541 } else if (op == MDB_CURSOR_EXACT)
1542 rc = mdb_cursor_set(cursor, key, data, &exact);
1544 rc = mdb_cursor_set(cursor, key, data, NULL);
1547 if (!cursor->mc_initialized)
1548 rc = mdb_cursor_first(cursor, key, data);
1550 rc = mdb_cursor_next(cursor, key, data);
1553 while (CURSOR_TOP(cursor) != NULL)
1554 cursor_pop_page(cursor);
1555 rc = mdb_cursor_first(cursor, key, data);
1558 DPRINTF("unhandled/unimplemented cursor operation %u", op);
1566 /* Allocate a page and initialize it
1569 mdbenv_new_page(MDB_env *env, uint32_t flags, int num)
1573 assert(env != NULL);
1574 assert(env->me_txn != NULL);
1576 DPRINTF("allocating new mpage %lu, page size %u",
1577 env->me_txn->mt_next_pgno, env->me_head.mh_psize);
1578 if ((dp = mdb_newpage(env->me_txn, NULL, 0, num)) == NULL)
1580 dp->p.mp_flags = flags | P_DIRTY;
1581 dp->p.mp_lower = PAGEHDRSZ;
1582 dp->p.mp_upper = env->me_head.mh_psize;
1584 if (IS_BRANCH(&dp->p))
1585 env->me_meta.mm_stat.ms_branch_pages++;
1586 else if (IS_LEAF(&dp->p))
1587 env->me_meta.mm_stat.ms_leaf_pages++;
1588 else if (IS_OVERFLOW(&dp->p)) {
1589 env->me_meta.mm_stat.ms_overflow_pages += num;
1590 dp->p.mp_pages = num;
1597 mdb_leaf_size(MDB_db *db, MDB_val *key, MDB_val *data)
1601 sz = LEAFSIZE(key, data);
1602 if (data->mv_size >= db->md_env->me_head.mh_psize / MDB_MINKEYS) {
1603 /* put on overflow page */
1604 sz -= data->mv_size - sizeof(pgno_t);
1607 return sz + sizeof(indx_t);
1611 mdb_branch_size(MDB_db *db, MDB_val *key)
1616 if (sz >= db->md_env->me_head.mh_psize / MDB_MINKEYS) {
1617 /* put on overflow page */
1618 /* not implemented */
1619 /* sz -= key->size - sizeof(pgno_t); */
1622 return sz + sizeof(indx_t);
1626 mdb_add_node(MDB_db *db, MDB_page *mp, indx_t indx,
1627 MDB_val *key, MDB_val *data, pgno_t pgno, uint8_t flags)
1630 size_t node_size = NODESIZE;
1633 MDB_dpage *ofp = NULL; /* overflow page */
1635 assert(mp->mp_upper >= mp->mp_lower);
1637 DPRINTF("add node [%.*s] to %s page %lu at index %i, key size %zu",
1638 key ? (int)key->mv_size : 0, key ? (char *)key->mv_data : NULL,
1639 IS_LEAF(mp) ? "leaf" : "branch",
1640 mp->mp_pgno, indx, key ? key->mv_size : 0);
1643 node_size += key->mv_size;
1647 if (F_ISSET(flags, F_BIGDATA)) {
1648 /* Data already on overflow page. */
1649 node_size += sizeof(pgno_t);
1650 } else if (data->mv_size >= db->md_env->me_head.mh_psize / MDB_MINKEYS) {
1651 int ovpages = PAGEHDRSZ + data->mv_size + db->md_env->me_head.mh_psize - 1;
1652 ovpages /= db->md_env->me_head.mh_psize;
1653 /* Put data on overflow page. */
1654 DPRINTF("data size is %zu, put on overflow page",
1656 node_size += sizeof(pgno_t);
1657 if ((ofp = mdbenv_new_page(db->md_env, P_OVERFLOW, ovpages)) == NULL)
1659 DPRINTF("allocated overflow page %lu", ofp->p.mp_pgno);
1662 node_size += data->mv_size;
1666 if (node_size + sizeof(indx_t) > SIZELEFT(mp)) {
1667 DPRINTF("not enough room in page %lu, got %u ptrs",
1668 mp->mp_pgno, NUMKEYS(mp));
1669 DPRINTF("upper - lower = %u - %u = %u", mp->mp_upper, mp->mp_lower,
1670 mp->mp_upper - mp->mp_lower);
1671 DPRINTF("node size = %zu", node_size);
1675 /* Move higher pointers up one slot. */
1676 for (i = NUMKEYS(mp); i > indx; i--)
1677 mp->mp_ptrs[i] = mp->mp_ptrs[i - 1];
1679 /* Adjust free space offsets. */
1680 ofs = mp->mp_upper - node_size;
1681 assert(ofs >= mp->mp_lower + sizeof(indx_t));
1682 mp->mp_ptrs[indx] = ofs;
1684 mp->mp_lower += sizeof(indx_t);
1686 /* Write the node data. */
1687 node = NODEPTR(mp, indx);
1688 node->mn_ksize = (key == NULL) ? 0 : key->mv_size;
1689 node->mn_flags = flags;
1691 node->mn_dsize = data->mv_size;
1693 node->mn_pgno = pgno;
1696 bcopy(key->mv_data, NODEKEY(node), key->mv_size);
1701 if (F_ISSET(flags, F_BIGDATA))
1702 bcopy(data->mv_data, node->mn_data + key->mv_size,
1705 bcopy(data->mv_data, node->mn_data + key->mv_size,
1708 bcopy(&ofp->p.mp_pgno, node->mn_data + key->mv_size,
1710 bcopy(data->mv_data, METADATA(&ofp->p), data->mv_size);
1718 mdb_del_node(MDB_db *db, MDB_page *mp, indx_t indx)
1721 indx_t i, j, numkeys, ptr;
1725 DPRINTF("delete node %u on %s page %lu", indx,
1726 IS_LEAF(mp) ? "leaf" : "branch", mp->mp_pgno);
1727 assert(indx < NUMKEYS(mp));
1729 node = NODEPTR(mp, indx);
1730 sz = NODESIZE + node->mn_ksize;
1732 if (F_ISSET(node->mn_flags, F_BIGDATA))
1733 sz += sizeof(pgno_t);
1735 sz += NODEDSZ(node);
1738 ptr = mp->mp_ptrs[indx];
1739 numkeys = NUMKEYS(mp);
1740 for (i = j = 0; i < numkeys; i++) {
1742 mp->mp_ptrs[j] = mp->mp_ptrs[i];
1743 if (mp->mp_ptrs[i] < ptr)
1744 mp->mp_ptrs[j] += sz;
1749 base = (char *)mp + mp->mp_upper;
1750 bcopy(base, base + sz, ptr - mp->mp_upper);
1752 mp->mp_lower -= sizeof(indx_t);
1757 mdb_cursor_open(MDB_db *db, MDB_txn *txn, MDB_cursor **ret)
1761 if (db == NULL || ret == NULL)
1764 if (txn != NULL && db->md_env != txn->mt_env) {
1768 if ((cursor = calloc(1, sizeof(*cursor))) != NULL) {
1769 SLIST_INIT(&cursor->mc_stack);
1771 cursor->mc_txn = txn;
1780 mdb_cursor_close(MDB_cursor *cursor)
1782 if (cursor != NULL) {
1783 while (!CURSOR_EMPTY(cursor))
1784 cursor_pop_page(cursor);
1786 /* btree_close(cursor->bt); */
1792 mdb_update_key(MDB_db *db, MDB_page *mp, indx_t indx,
1795 indx_t ptr, i, numkeys;
1801 node = NODEPTR(mp, indx);
1802 ptr = mp->mp_ptrs[indx];
1803 DPRINTF("update key %u (ofs %u) [%.*s] to [%.*s] on page %lu",
1805 (int)node->mn_ksize, (char *)NODEKEY(node),
1806 (int)key->mv_size, (char *)key->mv_data,
1809 delta = key->mv_size - node->mn_ksize;
1811 if (delta > 0 && SIZELEFT(mp) < delta) {
1812 DPRINTF("OUCH! Not enough room, delta = %d", delta);
1816 numkeys = NUMKEYS(mp);
1817 for (i = 0; i < numkeys; i++) {
1818 if (mp->mp_ptrs[i] <= ptr)
1819 mp->mp_ptrs[i] -= delta;
1822 base = (char *)mp + mp->mp_upper;
1823 len = ptr - mp->mp_upper + NODESIZE;
1824 bcopy(base, base - delta, len);
1825 mp->mp_upper -= delta;
1827 node = NODEPTR(mp, indx);
1828 node->mn_ksize = key->mv_size;
1831 bcopy(key->mv_data, NODEKEY(node), key->mv_size);
1836 /* Move a node from src to dst.
1839 mdb_move_node(MDB_db *bt, MDB_pageparent *src, indx_t srcindx,
1840 MDB_pageparent *dst, indx_t dstindx)
1846 srcnode = NODEPTR(src->mp_page, srcindx);
1847 DPRINTF("moving %s node %u [%.*s] on page %lu to node %u on page %lu",
1848 IS_LEAF(src->mp_page) ? "leaf" : "branch",
1850 (int)srcnode->mn_ksize, (char *)NODEKEY(srcnode),
1851 src->mp_page->mp_pgno,
1852 dstindx, dst->mp_page->mp_pgno);
1854 /* Mark src and dst as dirty. */
1855 if ((rc = mdb_touch(bt->md_env->me_txn, src)) ||
1856 (rc = mdb_touch(bt->md_env->me_txn, dst)))
1859 /* Add the node to the destination page.
1861 key.mv_size = srcnode->mn_ksize;
1862 key.mv_data = NODEKEY(srcnode);
1863 data.mv_size = NODEDSZ(srcnode);
1864 data.mv_data = NODEDATA(srcnode);
1865 rc = mdb_add_node(bt, dst->mp_page, dstindx, &key, &data, NODEPGNO(srcnode),
1867 if (rc != MDB_SUCCESS)
1870 /* Delete the node from the source page.
1872 mdb_del_node(bt, src->mp_page, srcindx);
1874 /* Update the parent separators.
1876 if (srcindx == 0 && src->mp_pi != 0) {
1877 DPRINTF("update separator for source page %lu to [%.*s]",
1878 src->mp_page->mp_pgno, (int)key.mv_size, (char *)key.mv_data);
1879 if ((rc = mdb_update_key(bt, src->mp_parent, src->mp_pi,
1880 &key)) != MDB_SUCCESS)
1884 if (srcindx == 0 && IS_BRANCH(src->mp_page)) {
1886 nullkey.mv_size = 0;
1887 assert(mdb_update_key(bt, src->mp_page, 0, &nullkey) == MDB_SUCCESS);
1890 if (dstindx == 0 && dst->mp_pi != 0) {
1891 DPRINTF("update separator for destination page %lu to [%.*s]",
1892 dst->mp_page->mp_pgno, (int)key.mv_size, (char *)key.mv_data);
1893 if ((rc = mdb_update_key(bt, dst->mp_parent, dst->mp_pi,
1894 &key)) != MDB_SUCCESS)
1898 if (dstindx == 0 && IS_BRANCH(dst->mp_page)) {
1900 nullkey.mv_size = 0;
1901 assert(mdb_update_key(bt, dst->mp_page, 0, &nullkey) == MDB_SUCCESS);
1908 mdb_merge(MDB_db *bt, MDB_pageparent *src, MDB_pageparent *dst)
1917 DPRINTF("merging page %lu and %lu", src->mp_page->mp_pgno, dst->mp_page->mp_pgno);
1919 assert(src->mp_parent); /* can't merge root page */
1920 assert(dst->mp_parent);
1921 assert(bt->md_env->me_txn != NULL);
1923 /* Mark src and dst as dirty. */
1924 if ((rc = mdb_touch(bt->md_env->me_txn, src)) ||
1925 (rc = mdb_touch(bt->md_env->me_txn, dst)))
1928 /* Move all nodes from src to dst.
1930 for (i = 0; i < NUMKEYS(src->mp_page); i++) {
1931 srcnode = NODEPTR(src->mp_page, i);
1933 key.mv_size = srcnode->mn_ksize;
1934 key.mv_data = NODEKEY(srcnode);
1935 data.mv_size = NODEDSZ(srcnode);
1936 data.mv_data = NODEDATA(srcnode);
1937 rc = mdb_add_node(bt, dst->mp_page, NUMKEYS(dst->mp_page), &key,
1938 &data, NODEPGNO(srcnode), srcnode->mn_flags);
1939 if (rc != MDB_SUCCESS)
1943 DPRINTF("dst page %lu now has %u keys (%.1f%% filled)",
1944 dst->mp_page->mp_pgno, NUMKEYS(dst->mp_page), (float)PAGEFILL(bt->md_env, dst->mp_page) / 10);
1946 /* Unlink the src page from parent.
1948 mdb_del_node(bt, src->mp_parent, src->mp_pi);
1949 if (src->mp_pi == 0) {
1951 if ((rc = mdb_update_key(bt, src->mp_parent, 0, &key)) != MDB_SUCCESS)
1955 if (IS_LEAF(src->mp_page))
1956 bt->md_env->me_meta.mm_stat.ms_leaf_pages--;
1958 bt->md_env->me_meta.mm_stat.ms_branch_pages--;
1960 mpp.mp_page = src->mp_parent;
1961 dh = (MDB_dhead *)src->mp_parent;
1963 mpp.mp_parent = dh->md_parent;
1964 mpp.mp_pi = dh->md_pi;
1966 return mdb_rebalance(bt, &mpp);
1969 #define FILL_THRESHOLD 250
1972 mdb_rebalance(MDB_db *db, MDB_pageparent *mpp)
1977 indx_t si = 0, di = 0;
1980 assert(db->md_env->me_txn != NULL);
1981 assert(mpp != NULL);
1983 DPRINTF("rebalancing %s page %lu (has %u keys, %.1f%% full)",
1984 IS_LEAF(mpp->mp_page) ? "leaf" : "branch",
1985 mpp->mp_page->mp_pgno, NUMKEYS(mpp->mp_page), (float)PAGEFILL(db->md_env, mpp->mp_page) / 10);
1987 if (PAGEFILL(db->md_env, mpp->mp_page) >= FILL_THRESHOLD) {
1988 DPRINTF("no need to rebalance page %lu, above fill threshold",
1989 mpp->mp_page->mp_pgno);
1993 if (mpp->mp_parent == NULL) {
1994 if (NUMKEYS(mpp->mp_page) == 0) {
1995 DPRINTF("tree is completely empty");
1996 db->md_env->me_txn->mt_root = P_INVALID;
1997 db->md_env->me_meta.mm_stat.ms_depth--;
1998 db->md_env->me_meta.mm_stat.ms_leaf_pages--;
1999 } else if (IS_BRANCH(mpp->mp_page) && NUMKEYS(mpp->mp_page) == 1) {
2000 DPRINTF("collapsing root page!");
2001 db->md_env->me_txn->mt_root = NODEPGNO(NODEPTR(mpp->mp_page, 0));
2002 if ((root = mdbenv_get_page(db->md_env, db->md_env->me_txn->mt_root)) == NULL)
2004 db->md_env->me_meta.mm_stat.ms_depth--;
2005 db->md_env->me_meta.mm_stat.ms_branch_pages--;
2007 DPRINTF("root page doesn't need rebalancing");
2011 /* The parent (branch page) must have at least 2 pointers,
2012 * otherwise the tree is invalid.
2014 assert(NUMKEYS(mpp->mp_parent) > 1);
2016 /* Leaf page fill factor is below the threshold.
2017 * Try to move keys from left or right neighbor, or
2018 * merge with a neighbor page.
2023 if (mpp->mp_pi == 0) {
2024 /* We're the leftmost leaf in our parent.
2026 DPRINTF("reading right neighbor");
2027 node = NODEPTR(mpp->mp_parent, mpp->mp_pi + 1);
2028 if ((npp.mp_page = mdbenv_get_page(db->md_env, NODEPGNO(node))) == NULL)
2030 npp.mp_pi = mpp->mp_pi + 1;
2032 di = NUMKEYS(mpp->mp_page);
2034 /* There is at least one neighbor to the left.
2036 DPRINTF("reading left neighbor");
2037 node = NODEPTR(mpp->mp_parent, mpp->mp_pi - 1);
2038 if ((npp.mp_page = mdbenv_get_page(db->md_env, NODEPGNO(node))) == NULL)
2040 npp.mp_pi = mpp->mp_pi - 1;
2041 si = NUMKEYS(npp.mp_page) - 1;
2044 npp.mp_parent = mpp->mp_parent;
2046 DPRINTF("found neighbor page %lu (%u keys, %.1f%% full)",
2047 npp.mp_page->mp_pgno, NUMKEYS(npp.mp_page), (float)PAGEFILL(db->md_env, npp.mp_page) / 10);
2049 /* If the neighbor page is above threshold and has at least two
2050 * keys, move one key from it.
2052 * Otherwise we should try to merge them.
2054 if (PAGEFILL(db->md_env, npp.mp_page) >= FILL_THRESHOLD && NUMKEYS(npp.mp_page) >= 2)
2055 return mdb_move_node(db, &npp, si, mpp, di);
2056 else { /* FIXME: if (has_enough_room()) */
2057 if (mpp->mp_pi == 0)
2058 return mdb_merge(db, &npp, mpp);
2060 return mdb_merge(db, mpp, &npp);
2065 mdb_del(MDB_db *bt, MDB_txn *txn,
2066 MDB_val *key, MDB_val *data)
2073 DPRINTF("========> delete key %.*s", (int)key->mv_size, (char *)key->mv_data);
2075 assert(key != NULL);
2077 if (bt == NULL || txn == NULL)
2080 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
2084 if (bt->md_env->me_txn != txn) {
2088 if (key->mv_size == 0 || key->mv_size > MAXKEYSIZE) {
2092 if ((rc = mdb_search_page(bt, txn, key, NULL, 1, &mpp)) != MDB_SUCCESS)
2095 leaf = mdb_search_node(bt, mpp.mp_page, key, &exact, &ki);
2096 if (leaf == NULL || !exact) {
2100 if (data && (rc = mdb_read_data(bt, NULL, leaf, data)) != MDB_SUCCESS)
2103 mdb_del_node(bt, mpp.mp_page, ki);
2104 bt->md_env->me_meta.mm_stat.ms_entries--;
2105 rc = mdb_rebalance(bt, &mpp);
2106 if (rc != MDB_SUCCESS)
2107 txn->mt_flags |= MDB_TXN_ERROR;
2112 /* Split page <*mpp>, and insert <key,(data|newpgno)> in either left or
2113 * right sibling, at index <*newindxp> (as if unsplit). Updates *mpp and
2114 * *newindxp with the actual values after split, ie if *mpp and *newindxp
2115 * refer to a node in the new right sibling page.
2118 mdb_split(MDB_db *bt, MDB_page **mpp, unsigned int *newindxp,
2119 MDB_val *newkey, MDB_val *newdata, pgno_t newpgno)
2122 int rc = MDB_SUCCESS, ins_new = 0;
2125 unsigned int i, j, split_indx;
2127 MDB_page *pright, *p;
2128 MDB_val sepkey, rkey, rdata;
2130 MDB_dpage *mdp, *rdp, *pdp;
2134 assert(bt->md_env != NULL);
2136 dh = ((MDB_dhead *)*mpp) - 1;
2137 mdp = (MDB_dpage *)dh;
2138 newindx = *newindxp;
2140 DPRINTF("-----> splitting %s page %lu and adding [%.*s] at index %i",
2141 IS_LEAF(&mdp->p) ? "leaf" : "branch", mdp->p.mp_pgno,
2142 (int)newkey->mv_size, (char *)newkey->mv_data, *newindxp);
2144 if (mdp->h.md_parent == NULL) {
2145 if ((pdp = mdbenv_new_page(bt->md_env, P_BRANCH, 1)) == NULL)
2148 mdp->h.md_parent = &pdp->p;
2149 bt->md_env->me_txn->mt_root = pdp->p.mp_pgno;
2150 DPRINTF("root split! new root = %lu", pdp->p.mp_pgno);
2151 bt->md_env->me_meta.mm_stat.ms_depth++;
2153 /* Add left (implicit) pointer. */
2154 if (mdb_add_node(bt, &pdp->p, 0, NULL, NULL,
2155 mdp->p.mp_pgno, 0) != MDB_SUCCESS)
2158 DPRINTF("parent branch page is %lu", mdp->h.md_parent->mp_pgno);
2161 /* Create a right sibling. */
2162 if ((rdp = mdbenv_new_page(bt->md_env, mdp->p.mp_flags, 1)) == NULL)
2164 rdp->h.md_parent = mdp->h.md_parent;
2165 rdp->h.md_pi = mdp->h.md_pi + 1;
2166 DPRINTF("new right sibling: page %lu", rdp->p.mp_pgno);
2168 /* Move half of the keys to the right sibling. */
2169 if ((copy = malloc(bt->md_env->me_head.mh_psize)) == NULL)
2171 bcopy(&mdp->p, copy, bt->md_env->me_head.mh_psize);
2172 bzero(&mdp->p.mp_ptrs, bt->md_env->me_head.mh_psize - PAGEHDRSZ);
2173 mdp->p.mp_lower = PAGEHDRSZ;
2174 mdp->p.mp_upper = bt->md_env->me_head.mh_psize;
2176 split_indx = NUMKEYS(copy) / 2 + 1;
2178 /* First find the separating key between the split pages.
2180 bzero(&sepkey, sizeof(sepkey));
2181 if (newindx == split_indx) {
2182 sepkey.mv_size = newkey->mv_size;
2183 sepkey.mv_data = newkey->mv_data;
2185 node = NODEPTR(copy, split_indx);
2186 sepkey.mv_size = node->mn_ksize;
2187 sepkey.mv_data = NODEKEY(node);
2190 DPRINTF("separator is [%.*s]", (int)sepkey.mv_size, (char *)sepkey.mv_data);
2192 /* Copy separator key to the parent.
2194 if (SIZELEFT(rdp->h.md_parent) < mdb_branch_size(bt, &sepkey)) {
2195 rc = mdb_split(bt, &rdp->h.md_parent, &rdp->h.md_pi,
2196 &sepkey, NULL, rdp->p.mp_pgno);
2198 /* Right page might now have changed parent.
2199 * Check if left page also changed parent.
2201 if (rdp->h.md_parent != mdp->h.md_parent &&
2202 mdp->h.md_pi >= NUMKEYS(mdp->h.md_parent)) {
2203 mdp->h.md_parent = rdp->h.md_parent;
2204 mdp->h.md_pi = rdp->h.md_pi - 1;
2207 rc = mdb_add_node(bt, rdp->h.md_parent, rdp->h.md_pi,
2208 &sepkey, NULL, rdp->p.mp_pgno, 0);
2210 if (rc != MDB_SUCCESS) {
2215 for (i = j = 0; i <= NUMKEYS(copy); j++) {
2216 if (i < split_indx) {
2217 /* Re-insert in left sibling. */
2220 /* Insert in right sibling. */
2221 if (i == split_indx)
2222 /* Reset insert index for right sibling. */
2223 j = (i == newindx && ins_new);
2227 if (i == newindx && !ins_new) {
2228 /* Insert the original entry that caused the split. */
2229 rkey.mv_data = newkey->mv_data;
2230 rkey.mv_size = newkey->mv_size;
2231 if (IS_LEAF(&mdp->p)) {
2232 rdata.mv_data = newdata->mv_data;
2233 rdata.mv_size = newdata->mv_size;
2240 /* Update page and index for the new key. */
2243 } else if (i == NUMKEYS(copy)) {
2246 node = NODEPTR(copy, i);
2247 rkey.mv_data = NODEKEY(node);
2248 rkey.mv_size = node->mn_ksize;
2249 if (IS_LEAF(&mdp->p)) {
2250 rdata.mv_data = NODEDATA(node);
2251 rdata.mv_size = node->mn_dsize;
2253 pgno = node->mn_pgno;
2254 flags = node->mn_flags;
2259 if (!IS_LEAF(&mdp->p) && j == 0) {
2260 /* First branch index doesn't need key data. */
2264 rc = mdb_add_node(bt, &pdp->p, j, &rkey, &rdata, pgno,flags);
2272 mdb_put(MDB_db *bt, MDB_txn *txn,
2273 MDB_val *key, MDB_val *data, unsigned int flags)
2275 int rc = MDB_SUCCESS, exact;
2280 assert(key != NULL);
2281 assert(data != NULL);
2283 if (bt == NULL || txn == NULL)
2286 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
2290 if (bt->md_env->me_txn != txn) {
2294 if (key->mv_size == 0 || key->mv_size > MAXKEYSIZE) {
2298 DPRINTF("==> put key %.*s, size %zu, data size %zu",
2299 (int)key->mv_size, (char *)key->mv_data, key->mv_size, data->mv_size);
2301 rc = mdb_search_page(bt, txn, key, NULL, 1, &mpp);
2302 if (rc == MDB_SUCCESS) {
2303 leaf = mdb_search_node(bt, mpp.mp_page, key, &exact, &ki);
2304 if (leaf && exact) {
2305 if (F_ISSET(flags, MDB_NOOVERWRITE)) {
2306 DPRINTF("duplicate key %.*s",
2307 (int)key->mv_size, (char *)key->mv_data);
2310 mdb_del_node(bt, mpp.mp_page, ki);
2312 if (leaf == NULL) { /* append if not found */
2313 ki = NUMKEYS(mpp.mp_page);
2314 DPRINTF("appending key at index %i", ki);
2316 } else if (rc == ENOENT) {
2318 /* new file, just write a root leaf page */
2319 DPRINTF("allocating new root leaf page");
2320 if ((dp = mdbenv_new_page(bt->md_env, P_LEAF, 1)) == NULL) {
2323 mpp.mp_page = &dp->p;
2324 txn->mt_root = mpp.mp_page->mp_pgno;
2325 bt->md_env->me_meta.mm_stat.ms_depth++;
2331 assert(IS_LEAF(mpp.mp_page));
2332 DPRINTF("there are %u keys, should insert new key at index %i",
2333 NUMKEYS(mpp.mp_page), ki);
2335 if (SIZELEFT(mpp.mp_page) < mdb_leaf_size(bt, key, data)) {
2336 rc = mdb_split(bt, &mpp.mp_page, &ki, key, data, P_INVALID);
2338 /* There is room already in this leaf page. */
2339 rc = mdb_add_node(bt, mpp.mp_page, ki, key, data, 0, 0);
2342 if (rc != MDB_SUCCESS)
2343 txn->mt_flags |= MDB_TXN_ERROR;
2345 bt->md_env->me_meta.mm_stat.ms_entries++;
2352 mdbenv_get_flags(MDB_env *env, unsigned int *arg)
2357 *arg = env->me_flags;
2362 mdbenv_get_path(MDB_env *env, const char **arg)
2367 *arg = env->me_path;
2372 mdbenv_stat(MDB_env *env, MDB_stat **arg)
2374 if (env == NULL || arg == NULL)
2377 *arg = &env->me_meta.mm_stat;
2382 int mdb_open(MDB_env *env, MDB_txn *txn, const char *name, unsigned int flags, MDB_db **db)
2385 *db = (MDB_db *)&env->me_db;
2391 int mdb_stat(MDB_db *db, MDB_stat **arg)
2393 if (db == NULL || arg == NULL)
2396 *arg = &db->md_stat;
2401 void mdb_close(MDB_db *db)