26 #include <openssl/sha.h>
33 # define DPRINTF(...) do { fprintf(stderr, "%s:%d: ", __func__, __LINE__); \
34 fprintf(stderr, __VA_ARGS__); \
35 fprintf(stderr, "\n"); } while(0)
42 #define MDB_MAGIC 0xBEEFC0DE
44 #define MAXKEYSIZE 255
46 #define P_INVALID (~0L)
48 #define F_ISSET(w, f) (((w) & (f)) == (f))
51 typedef uint16_t indx_t;
53 #define DEFAULT_READERS 126
54 #define DEFAULT_MAPSIZE 1048576
56 /* Lock descriptor stuff */
61 typedef struct MDB_rxbody {
66 #define CACHELINE 64 /* most CPUs. Itanium uses 128 */
69 typedef struct MDB_reader {
71 /* cache line alignment */
72 char pad[CACHELINE-sizeof(MDB_rxbody)];
77 uint32_t mt_version; \
78 pthread_mutex_t mt_mutex; \
80 uint32_t mt_numreaders
81 typedef struct MDB_txbody {
85 typedef struct MDB_txninfo {
87 char pad[CACHELINE-sizeof(MDB_txbody)];
88 pthread_mutex_t mt_wmutex;
89 char pad2[CACHELINE-sizeof(pthread_mutex_t)];
90 MDB_reader mt_readers[1];
93 /* Common header for all page types. Overflow pages
94 * occupy a number of contiguous pages with no
95 * headers on any page after the first.
97 typedef struct MDB_page { /* represents a page of storage */
98 pgno_t mp_pgno; /* page number */
99 #define P_BRANCH 0x01 /* branch page */
100 #define P_LEAF 0x02 /* leaf page */
101 #define P_OVERFLOW 0x04 /* overflow page */
102 #define P_META 0x08 /* meta page */
103 #define P_HEAD 0x10 /* header page */
104 #define P_DIRTY 0x20 /* dirty page */
106 #define mp_lower mp_pb.pb.pb_lower
107 #define mp_upper mp_pb.pb.pb_upper
108 #define mp_pages mp_pb.pb_pages
111 indx_t pb_lower; /* lower bound of free space */
112 indx_t pb_upper; /* upper bound of free space */
114 uint32_t pb_pages; /* number of overflow pages */
116 indx_t mp_ptrs[1]; /* dynamic size */
119 #define PAGEHDRSZ offsetof(MDB_page, mp_ptrs)
121 #define NUMKEYS(p) (((p)->mp_lower - PAGEHDRSZ) >> 1)
122 #define SIZELEFT(p) (indx_t)((p)->mp_upper - (p)->mp_lower)
123 #define PAGEFILL(env, p) (1000 * ((env)->me_head.mh_psize - PAGEHDRSZ - SIZELEFT(p)) / \
124 ((env)->me_head.mh_psize - PAGEHDRSZ))
125 #define IS_LEAF(p) F_ISSET((p)->mp_flags, P_LEAF)
126 #define IS_BRANCH(p) F_ISSET((p)->mp_flags, P_BRANCH)
127 #define IS_OVERFLOW(p) F_ISSET((p)->mp_flags, P_OVERFLOW)
129 typedef struct MDB_head { /* header page content */
133 uint32_t mh_psize; /* page size */
134 void *mh_address; /* address for fixed mapping */
135 size_t mh_mapsize; /* size of mmap region */
138 typedef struct MDB_rootstat {
139 pgno_t mr_root; /* page number of root page */
143 typedef struct MDB_meta { /* meta (footer) page content */
144 pgno_t mm_root; /* page number of root page */
146 pgno_t mm_prev_meta; /* previous meta page number */
148 #define MDB_TOMBSTONE 0x01 /* file is replaced */
150 #define mm_revisions mm_stat.ms_revisions
151 #define mm_created_at mm_stat.ms_created_at
152 uint8_t mm_hash[SHA_DIGEST_LENGTH];
155 typedef struct MDB_dhead { /* a dirty page */
156 SIMPLEQ_ENTRY(MDB_dpage) md_next; /* queue of dirty pages */
158 int md_pi; /* parent index */
162 typedef struct MDB_dpage {
167 SIMPLEQ_HEAD(dirty_queue, MDB_dpage);
169 typedef struct MDB_pageparent {
175 static MDB_dpage *mdb_newpage(MDB_txn *txn, MDB_page *parent, int parent_idx, int num);
176 static int mdb_touch(MDB_txn *txn, MDB_pageparent *mp);
178 typedef struct MDB_ppage { /* ordered list of pages */
179 SLIST_ENTRY(MDB_ppage) mp_entry;
181 unsigned int mp_ki; /* cursor index on page */
183 SLIST_HEAD(page_stack, MDB_ppage);
185 #define CURSOR_EMPTY(c) SLIST_EMPTY(&(c)->mc_stack)
186 #define CURSOR_TOP(c) SLIST_FIRST(&(c)->mc_stack)
187 #define CURSOR_POP(c) SLIST_REMOVE_HEAD(&(c)->mc_stack, mp_entry)
188 #define CURSOR_PUSH(c,p) SLIST_INSERT_HEAD(&(c)->mc_stack, p, mp_entry)
193 struct page_stack mc_stack; /* stack of parent pages */
194 short mc_initialized; /* 1 if initialized */
195 short mc_eof; /* 1 if end is reached */
198 #define METAHASHLEN offsetof(MDB_meta, mm_hash)
199 #define METADATA(p) ((void *)((char *)p + PAGEHDRSZ))
201 typedef struct MDB_node {
202 #define mn_pgno mn_p.np_pgno
203 #define mn_dsize mn_p.np_dsize
205 pgno_t np_pgno; /* child page number */
206 uint32_t np_dsize; /* leaf data size */
208 uint16_t mn_ksize; /* key size */
209 #define F_BIGDATA 0x01 /* data put on overflow page */
215 pgno_t mt_root; /* current / new root page */
216 pgno_t mt_next_pgno; /* next unallocated page */
217 pgno_t mt_first_pgno;
221 struct dirty_queue *dirty_queue; /* modified pages */
224 #define MDB_TXN_RDONLY 0x01 /* read-only transaction */
225 #define MDB_TXN_ERROR 0x02 /* an error has occurred */
226 unsigned int mt_flags;
229 /* Must be same as MDB_db, minus md_root/md_stat */
230 typedef struct MDB_db0 {
231 unsigned int md_flags;
232 MDB_cmp_func *md_cmp; /* user compare function */
233 MDB_rel_func *md_rel; /* user relocate function */
234 MDB_db *md_parent; /* parent tree */
239 unsigned int md_flags;
240 MDB_cmp_func *md_cmp; /* user compare function */
241 MDB_rel_func *md_rel; /* user relocate function */
242 MDB_db *md_parent; /* parent tree */
244 pgno_t md_root; /* page number of root page */
251 #define MDB_FIXPADDING 0x01 /* internal */
256 MDB_txninfo *me_txns;
258 MDB_db0 me_db; /* first DB, overlaps with meta */
260 MDB_txn *me_txn; /* current write transaction */
262 off_t me_size; /* current file size */
263 pthread_key_t me_txkey; /* thread-key for readers */
266 #define NODESIZE offsetof(MDB_node, mn_data)
268 #define INDXSIZE(k) (NODESIZE + ((k) == NULL ? 0 : (k)->mv_size))
269 #define LEAFSIZE(k, d) (NODESIZE + (k)->mv_size + (d)->mv_size)
270 #define NODEPTR(p, i) ((MDB_node *)((char *)(p) + (p)->mp_ptrs[i]))
271 #define NODEKEY(node) (void *)((node)->mn_data)
272 #define NODEDATA(node) (void *)((char *)(node)->mn_data + (node)->mn_ksize)
273 #define NODEPGNO(node) ((node)->mn_pgno)
274 #define NODEDSZ(node) ((node)->mn_dsize)
276 #define MDB_COMMIT_PAGES 64 /* max number of pages to write in one commit */
277 #define MDB_MAXCACHE_DEF 1024 /* max number of pages to keep in cache */
279 static int mdb_search_page_root(MDB_db *db,
281 MDB_cursor *cursor, int modify,
282 MDB_pageparent *mpp);
283 static int mdb_search_page(MDB_db *db,
284 MDB_txn *txn, MDB_val *key,
285 MDB_cursor *cursor, int modify,
286 MDB_pageparent *mpp);
288 static int mdbenv_write_header(MDB_env *env);
289 static int mdbenv_read_header(MDB_env *env);
290 static int mdb_check_meta_page(MDB_page *p);
291 static int mdbenv_read_meta(MDB_env *env, pgno_t *p_next);
292 static int mdbenv_write_meta(MDB_env *env, pgno_t root,
294 static MDB_page *mdbenv_get_page(MDB_env *env, pgno_t pgno);
296 static MDB_node *mdb_search_node(MDB_db *db, MDB_page *mp,
297 MDB_val *key, int *exactp, unsigned int *kip);
298 static int mdb_add_node(MDB_db *bt, MDB_page *mp,
299 indx_t indx, MDB_val *key, MDB_val *data,
300 pgno_t pgno, uint8_t flags);
301 static void mdb_del_node(MDB_db *bt, MDB_page *mp,
303 static int mdb_read_data(MDB_db *bt, MDB_page *mp,
304 MDB_node *leaf, MDB_val *data);
306 static int mdb_rebalance(MDB_db *bt, MDB_pageparent *mp);
307 static int mdb_update_key(MDB_db *bt, MDB_page *mp,
308 indx_t indx, MDB_val *key);
309 static int mdb_move_node(MDB_db *bt,
310 MDB_pageparent *src, indx_t srcindx,
311 MDB_pageparent *dst, indx_t dstindx);
312 static int mdb_merge(MDB_db *bt, MDB_pageparent *src,
313 MDB_pageparent *dst);
314 static int mdb_split(MDB_db *bt, MDB_page **mpp,
315 unsigned int *newindxp, MDB_val *newkey,
316 MDB_val *newdata, pgno_t newpgno);
317 static MDB_dpage *mdbenv_new_page(MDB_env *env, uint32_t flags, int num);
319 static void cursor_pop_page(MDB_cursor *cursor);
320 static MDB_ppage *cursor_push_page(MDB_cursor *cursor,
323 static int mdb_set_key(MDB_db *bt, MDB_page *mp,
324 MDB_node *node, MDB_val *key);
325 static int mdb_sibling(MDB_cursor *cursor, int move_right);
326 static int mdb_cursor_next(MDB_cursor *cursor,
327 MDB_val *key, MDB_val *data);
328 static int mdb_cursor_set(MDB_cursor *cursor,
329 MDB_val *key, MDB_val *data, int *exactp);
330 static int mdb_cursor_first(MDB_cursor *cursor,
331 MDB_val *key, MDB_val *data);
333 static size_t mdb_leaf_size(MDB_db *bt, MDB_val *key,
335 static size_t mdb_branch_size(MDB_db *bt, MDB_val *key);
337 static pgno_t mdbenv_compact_tree(MDB_env *env, pgno_t pgno,
340 static int memncmp(const void *s1, size_t n1,
341 const void *s2, size_t n2);
342 static int memnrcmp(const void *s1, size_t n1,
343 const void *s2, size_t n2);
346 memncmp(const void *s1, size_t n1, const void *s2, size_t n2)
349 if (memcmp(s1, s2, n1) == 0)
353 if (memcmp(s1, s2, n2) == 0)
356 return memcmp(s1, s2, n1);
360 memnrcmp(const void *s1, size_t n1, const void *s2, size_t n2)
362 const unsigned char *p1;
363 const unsigned char *p2;
366 return n2 == 0 ? 0 : -1;
369 return n1 == 0 ? 0 : 1;
371 p1 = (const unsigned char *)s1 + n1 - 1;
372 p2 = (const unsigned char *)s2 + n2 - 1;
376 return (p2 == s2) ? 0 : -1;
378 return (p1 == p2) ? 0 : 1;
386 mdb_cmp(MDB_db *db, const MDB_val *a, const MDB_val *b)
388 return db->md_cmp(a, b);
392 _mdb_cmp(MDB_db *db, const MDB_val *key1, const MDB_val *key2)
394 if (F_ISSET(db->md_flags, MDB_REVERSEKEY))
395 return memnrcmp(key1->mv_data, key1->mv_size, key2->mv_data, key2->mv_size);
397 return memncmp((char *)key1->mv_data, key1->mv_size, key2->mv_data, key2->mv_size);
400 /* Allocate new page(s) for writing */
402 mdb_newpage(MDB_txn *txn, MDB_page *parent, int parent_idx, int num)
406 if ((dp = malloc(txn->mt_env->me_head.mh_psize * num + sizeof(MDB_dhead))) == NULL)
409 dp->h.md_parent = parent;
410 dp->h.md_pi = parent_idx;
411 SIMPLEQ_INSERT_TAIL(txn->mt_u.dirty_queue, dp, h.md_next);
412 dp->p.mp_pgno = txn->mt_next_pgno;
413 txn->mt_next_pgno += num;
418 /* Touch a page: make it dirty and re-insert into tree with updated pgno.
421 mdb_touch(MDB_txn *txn, MDB_pageparent *pp)
424 MDB_page *mp = pp->mp_page;
429 if (!F_ISSET(mp->mp_flags, P_DIRTY)) {
431 DPRINTF("touching page %lu -> %lu", mp->mp_pgno, txn->mt_next_pgno);
432 if ((dp = mdb_newpage(txn, pp->mp_parent, pp->mp_pi, 1)) == NULL)
434 pgno = dp->p.mp_pgno;
435 bcopy(mp, &dp->p, txn->mt_env->me_head.mh_psize);
438 mp->mp_flags |= P_DIRTY;
440 /* Update the page number to new touched page. */
441 if (pp->mp_parent != NULL)
442 NODEPGNO(NODEPTR(pp->mp_parent, pp->mp_pi)) = mp->mp_pgno;
449 mdbenv_sync(MDB_env *env)
452 if (!F_ISSET(env->me_flags, MDB_NOSYNC)) {
453 if (fsync(env->me_fd))
460 mdb_txn_begin(MDB_env *env, int rdonly, MDB_txn **ret)
465 if ((txn = calloc(1, sizeof(*txn))) == NULL) {
466 DPRINTF("calloc: %s", strerror(errno));
471 txn->mt_flags |= MDB_TXN_RDONLY;
473 txn->mt_u.dirty_queue = calloc(1, sizeof(*txn->mt_u.dirty_queue));
474 if (txn->mt_u.dirty_queue == NULL) {
478 SIMPLEQ_INIT(txn->mt_u.dirty_queue);
480 pthread_mutex_lock(&env->me_txns->mt_wmutex);
481 env->me_txns->mt_txnid++;
483 txn->mt_txnid = env->me_txns->mt_txnid;
485 MDB_reader *r = pthread_getspecific(env->me_txkey);
488 pthread_mutex_lock(&env->me_txns->mt_mutex);
489 for (i=0; i<env->me_maxreaders; i++) {
490 if (env->me_txns->mt_readers[i].mr_pid == 0) {
491 env->me_txns->mt_readers[i].mr_pid = getpid();
492 env->me_txns->mt_readers[i].mr_tid = pthread_self();
493 pthread_setspecific(env->me_txkey, &env->me_txns->mt_readers[i]);
494 if (i >= env->me_txns->mt_numreaders)
495 env->me_txns->mt_numreaders = i+1;
499 pthread_mutex_unlock(&env->me_txns->mt_mutex);
500 if (i == env->me_maxreaders) {
504 r->mr_txnid = txn->mt_txnid;
505 txn->mt_u.reader = r;
512 if ((rc = mdbenv_read_meta(env, &txn->mt_next_pgno)) != MDB_SUCCESS) {
517 txn->mt_first_pgno = txn->mt_next_pgno;
518 txn->mt_root = env->me_meta.mm_root;
519 DPRINTF("begin transaction on mdbenv %p, root page %lu", env, txn->mt_root);
526 mdb_txn_abort(MDB_txn *txn)
535 DPRINTF("abort transaction on mdbenv %p, root page %lu", env, txn->mt_root);
537 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
538 txn->mt_u.reader->mr_txnid = 0;
540 /* Discard all dirty pages.
542 while (!SIMPLEQ_EMPTY(txn->mt_u.dirty_queue)) {
543 dp = SIMPLEQ_FIRST(txn->mt_u.dirty_queue);
544 SIMPLEQ_REMOVE_HEAD(txn->mt_u.dirty_queue, h.md_next);
549 DPRINTF("releasing write lock on txn %p", txn);
551 if (flock(txn->bt->fd, LOCK_UN) != 0) {
552 DPRINTF("failed to unlock fd %d: %s",
553 txn->bt->fd, strerror(errno));
556 free(txn->mt_u.dirty_queue);
558 pthread_mutex_unlock(&env->me_txns->mt_wmutex);
565 mdb_txn_commit(MDB_txn *txn)
572 struct iovec iov[MDB_COMMIT_PAGES];
575 assert(txn->mt_env != NULL);
579 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
580 DPRINTF("attempt to commit read-only transaction");
585 if (txn != env->me_txn) {
586 DPRINTF("attempt to commit unknown transaction");
591 if (F_ISSET(txn->mt_flags, MDB_TXN_ERROR)) {
592 DPRINTF("error flag is set, can't commit");
597 if (SIMPLEQ_EMPTY(txn->mt_u.dirty_queue))
600 if (F_ISSET(env->me_flags, MDB_FIXPADDING)) {
601 size = lseek(env->me_fd, 0, SEEK_END);
602 size += env->me_head.mh_psize - (size % env->me_head.mh_psize);
603 DPRINTF("extending to multiple of page size: %lu", size);
604 if (ftruncate(env->me_fd, size) != 0) {
606 DPRINTF("ftruncate: %s", strerror(errno));
610 env->me_flags ^= MDB_FIXPADDING;
613 DPRINTF("committing transaction on mdbenv %p, root page %lu",
616 /* Commit up to MDB_COMMIT_PAGES dirty pages to disk until done.
621 SIMPLEQ_FOREACH(dp, txn->mt_u.dirty_queue, h.md_next) {
622 DPRINTF("committing page %lu", dp->p.mp_pgno);
623 iov[n].iov_len = env->me_head.mh_psize * dp->h.md_num;
624 iov[n].iov_base = &dp->p;
625 /* clear dirty flag */
626 dp->p.mp_flags &= ~P_DIRTY;
627 if (++n >= MDB_COMMIT_PAGES) {
636 DPRINTF("committing %u dirty pages", n);
637 rc = writev(env->me_fd, iov, n);
638 if (rc != (ssize_t)env->me_head.mh_psize*n) {
641 DPRINTF("short write, filesystem full?");
643 DPRINTF("writev: %s", strerror(errno));
648 /* Drop the dirty pages.
650 while (!SIMPLEQ_EMPTY(txn->mt_u.dirty_queue)) {
651 dp = SIMPLEQ_FIRST(txn->mt_u.dirty_queue);
652 SIMPLEQ_REMOVE_HEAD(txn->mt_u.dirty_queue, h.md_next);
659 if ((n = mdbenv_sync(env)) != 0 ||
660 (n = mdbenv_write_meta(env, txn->mt_root, 0)) != MDB_SUCCESS ||
661 (n = mdbenv_sync(env)) != 0) {
666 pthread_mutex_unlock(&env->me_txns->mt_wmutex);
676 mdbenv_write_header(MDB_env *env)
684 DPRINTF("writing header page");
687 psize = sysconf(_SC_PAGE_SIZE);
689 if ((p = calloc(1, psize)) == NULL)
691 p->mp_flags = P_HEAD;
693 env->me_head.mh_psize = psize;
694 env->me_head.mh_flags = env->me_flags & 0xffff;
697 bcopy(&env->me_head, h, sizeof(*h));
699 rc = write(env->me_fd, p, env->me_head.mh_psize);
701 if (rc != (ssize_t)env->me_head.mh_psize) {
704 DPRINTF("short write, filesystem full?");
712 mdbenv_read_header(MDB_env *env)
721 /* We don't know the page size yet, so use a minimum value.
724 if ((rc = pread(env->me_fd, page, PAGESIZE, 0)) == 0) {
726 } else if (rc != PAGESIZE) {
729 DPRINTF("read: %s", strerror(errno));
733 p = (MDB_page *)page;
735 if (!F_ISSET(p->mp_flags, P_HEAD)) {
736 DPRINTF("page %lu not a header page", p->mp_pgno);
741 if (h->mh_magic != MDB_MAGIC) {
742 DPRINTF("header has invalid magic");
746 if (h->mh_version != MDB_VERSION) {
747 DPRINTF("database is version %u, expected version %u",
748 h->mh_version, MDB_VERSION);
752 bcopy(h, &env->me_head, sizeof(*h));
757 mdbenv_write_meta(MDB_env *env, pgno_t root, unsigned int flags)
763 DPRINTF("writing meta page for root page %lu", root);
766 assert(env->me_txn != NULL);
768 if ((dp = mdbenv_new_page(env, P_META, 1)) == NULL)
771 env->me_meta.mm_prev_meta = env->me_meta.mm_root;
772 env->me_meta.mm_root = root;
773 env->me_meta.mm_flags = flags;
774 env->me_meta.mm_created_at = time(0);
775 env->me_meta.mm_revisions++;
776 SHA1((unsigned char *)&env->me_meta, METAHASHLEN, env->me_meta.mm_hash);
778 /* Copy the meta data changes to the new meta page. */
779 meta = METADATA(&dp->p);
780 bcopy(&env->me_meta, meta, sizeof(*meta));
782 rc = write(env->me_fd, &dp->p, env->me_head.mh_psize);
783 SIMPLEQ_REMOVE_HEAD(env->me_txn->mt_u.dirty_queue, h.md_next);
785 if (rc != (ssize_t)env->me_head.mh_psize) {
788 DPRINTF("short write, filesystem full?");
792 if ((env->me_size = lseek(env->me_fd, 0, SEEK_END)) == -1) {
793 DPRINTF("failed to update file size: %s", strerror(errno));
800 /* Returns true if page p is a valid meta page, false otherwise.
803 mdb_check_meta_page(MDB_page *p)
806 unsigned char hash[SHA_DIGEST_LENGTH];
809 if (!F_ISSET(p->mp_flags, P_META)) {
810 DPRINTF("page %lu not a meta page", p->mp_pgno);
814 if (m->mm_root >= p->mp_pgno && m->mm_root != P_INVALID) {
815 DPRINTF("page %lu points to an invalid root page", p->mp_pgno);
819 SHA1((unsigned char *)m, METAHASHLEN, hash);
820 if (bcmp(hash, m->mm_hash, SHA_DIGEST_LENGTH) != 0) {
821 DPRINTF("page %lu has an invalid digest", p->mp_pgno);
829 mdbenv_read_meta(MDB_env *env, pgno_t *p_next)
833 pgno_t meta_pgno, next_pgno;
838 if ((size = lseek(env->me_fd, 0, SEEK_END)) == -1)
841 DPRINTF("mdbenv_read_meta: size = %lu", size);
843 if (size < env->me_size) {
844 DPRINTF("file has shrunk!");
849 if (size == env->me_head.mh_psize) { /* there is only the header */
852 env->me_meta.mm_stat.ms_psize = env->me_head.mh_psize;
853 return MDB_SUCCESS; /* new file */
856 next_pgno = size / env->me_head.mh_psize;
857 if (next_pgno == 0) {
858 DPRINTF("corrupt file");
863 meta_pgno = next_pgno - 1;
865 if (size % env->me_head.mh_psize != 0) {
866 DPRINTF("filesize not a multiple of the page size!");
867 env->me_flags |= MDB_FIXPADDING;
874 if (size == env->me_size) {
875 DPRINTF("size unchanged, keeping current meta page");
876 if (F_ISSET(env->me_meta.mm_flags, MDB_TOMBSTONE)) {
877 DPRINTF("file is dead");
886 while (meta_pgno > 0) {
887 if ((mp = mdbenv_get_page(env, meta_pgno)) == NULL)
889 if (!mdb_check_meta_page(mp)) {
891 DPRINTF("flags = 0x%x", meta->mm_flags);
892 if (F_ISSET(meta->mm_flags, MDB_TOMBSTONE)) {
893 DPRINTF("file is dead");
897 /* Make copy of last meta page. */
898 bcopy(meta, &env->me_meta, sizeof(env->me_meta));
902 --meta_pgno; /* scan backwards to first valid meta page */
913 mdbenv_create(MDB_env **env)
917 e = calloc(1, sizeof(*e));
918 if (!e) return ENOMEM;
920 e->me_head.mh_magic = MDB_MAGIC;
921 e->me_head.mh_version = MDB_VERSION;
922 e->me_head.mh_mapsize = DEFAULT_MAPSIZE;
923 e->me_maxreaders = DEFAULT_READERS;
930 mdbenv_set_mapsize(MDB_env *env, size_t size)
934 env->me_mapsize = env->me_head.mh_mapsize = size;
939 mdbenv_set_maxreaders(MDB_env *env, int readers)
941 env->me_maxreaders = readers;
946 mdbenv_get_maxreaders(MDB_env *env, int *readers)
948 if (!env || !readers)
950 *readers = env->me_maxreaders;
955 mdbenv_open2(MDB_env *env, unsigned int flags)
959 i = fcntl(env->me_fd, F_GETFL, 0);
960 if (fcntl(env->me_fd, F_SETFL, i | O_APPEND) == -1)
963 env->me_flags = flags;
964 env->me_flags &= ~MDB_FIXPADDING;
965 env->me_meta.mm_root = P_INVALID;
967 if ((i = mdbenv_read_header(env)) != 0) {
970 DPRINTF("new mdbenv");
974 if (!env->me_mapsize)
975 env->me_mapsize = env->me_head.mh_mapsize;
978 if (env->me_head.mh_address && (flags & MDB_FIXEDMAP))
980 env->me_map = mmap(env->me_head.mh_address, env->me_mapsize, PROT_READ, i,
982 if (env->me_map == MAP_FAILED)
986 env->me_head.mh_mapsize = env->me_mapsize;
987 if (flags & MDB_FIXEDMAP)
988 env->me_head.mh_address = env->me_map;
989 i = mdbenv_write_header(env);
990 if (i != MDB_SUCCESS) {
991 munmap(env->me_map, env->me_mapsize);
996 if ((i = mdbenv_read_meta(env, NULL)) != 0)
999 DPRINTF("opened database version %u, pagesize %u",
1000 env->me_head.mh_version, env->me_head.mh_psize);
1001 DPRINTF("timestamp: %s", ctime(&env->me_meta.mm_created_at));
1002 DPRINTF("depth: %u", env->me_meta.mm_stat.ms_depth);
1003 DPRINTF("entries: %lu", env->me_meta.mm_stat.ms_entries);
1004 DPRINTF("revisions: %lu", env->me_meta.mm_stat.ms_revisions);
1005 DPRINTF("branch pages: %lu", env->me_meta.mm_stat.ms_branch_pages);
1006 DPRINTF("leaf pages: %lu", env->me_meta.mm_stat.ms_leaf_pages);
1007 DPRINTF("overflow pages: %lu", env->me_meta.mm_stat.ms_overflow_pages);
1008 DPRINTF("root: %lu", env->me_meta.mm_root);
1009 DPRINTF("previous me_meta.me_page: %lu", env->me_meta.mm_prev_meta);
1015 mdbenv_reader_dest(void *ptr)
1017 MDB_reader *reader = ptr;
1019 reader->mr_txnid = 0;
1025 mdbenv_open(MDB_env *env, const char *path, unsigned int flags, mode_t mode)
1027 int oflags, rc, shmid;
1031 if (F_ISSET(flags, MDB_RDONLY))
1034 oflags = O_RDWR | O_CREAT | O_APPEND;
1036 if ((env->me_fd = open(path, oflags, mode)) == -1)
1039 env->me_shmkey = ftok(path, 'm');
1040 size = (env->me_maxreaders-1) * sizeof(MDB_reader) + sizeof(MDB_txninfo);
1041 shmid = shmget(env->me_shmkey, size, IPC_CREAT|IPC_EXCL|mode);
1043 if (errno == EEXIST) {
1044 shmid = shmget(env->me_shmkey, size, IPC_CREAT|mode);
1047 env->me_txns = shmat(shmid, NULL, 0);
1052 pthread_mutexattr_t mattr;
1054 env->me_txns = shmat(shmid, NULL, 0);
1055 pthread_mutexattr_init(&mattr);
1056 pthread_mutexattr_setpshared(&mattr, PTHREAD_PROCESS_SHARED);
1057 pthread_mutex_init(&env->me_txns->mt_mutex, &mattr);
1058 pthread_mutex_init(&env->me_txns->mt_wmutex, &mattr);
1061 if ((rc = mdbenv_open2(env, flags)) != MDB_SUCCESS) {
1065 env->me_path = strdup(path);
1066 DPRINTF("opened dbenv %p", env);
1069 pthread_key_create(&env->me_txkey, mdbenv_reader_dest);
1075 mdbenv_close(MDB_env *env)
1083 munmap(env->me_map, env->me_mapsize);
1087 shmdt(env->me_txns);
1092 /* Search for key within a leaf page, using binary search.
1093 * Returns the smallest entry larger or equal to the key.
1094 * If exactp is non-null, stores whether the found entry was an exact match
1095 * in *exactp (1 or 0).
1096 * If kip is non-null, stores the index of the found entry in *kip.
1097 * If no entry larger or equal to the key is found, returns NULL.
1100 mdb_search_node(MDB_db *bt, MDB_page *mp, MDB_val *key,
1101 int *exactp, unsigned int *kip)
1109 DPRINTF("searching %lu keys in %s page %lu",
1111 IS_LEAF(mp) ? "leaf" : "branch",
1114 assert(NUMKEYS(mp) > 0);
1116 bzero(&nodekey, sizeof(nodekey));
1118 low = IS_LEAF(mp) ? 0 : 1;
1119 high = NUMKEYS(mp) - 1;
1120 while (low <= high) {
1121 i = (low + high) >> 1;
1122 node = NODEPTR(mp, i);
1124 nodekey.mv_size = node->mn_ksize;
1125 nodekey.mv_data = NODEKEY(node);
1128 rc = bt->md_cmp(key, &nodekey);
1130 rc = _mdb_cmp(bt, key, &nodekey);
1133 DPRINTF("found leaf index %u [%.*s], rc = %i",
1134 i, (int)nodekey.mv_size, (char *)nodekey.mv_data, rc);
1136 DPRINTF("found branch index %u [%.*s -> %lu], rc = %i",
1137 i, (int)node->mn_ksize, (char *)NODEKEY(node),
1148 if (rc > 0) { /* Found entry is less than the key. */
1149 i++; /* Skip to get the smallest entry larger than key. */
1150 if (i >= NUMKEYS(mp))
1151 /* There is no entry larger or equal to the key. */
1155 *exactp = (rc == 0);
1156 if (kip) /* Store the key index if requested. */
1159 return NODEPTR(mp, i);
1163 cursor_pop_page(MDB_cursor *cursor)
1167 top = CURSOR_TOP(cursor);
1170 DPRINTF("popped page %lu off cursor %p", top->mp_page->mp_pgno, cursor);
1176 cursor_push_page(MDB_cursor *cursor, MDB_page *mp)
1180 DPRINTF("pushing page %lu on cursor %p", mp->mp_pgno, cursor);
1182 if ((ppage = calloc(1, sizeof(*ppage))) == NULL)
1184 ppage->mp_page = mp;
1185 CURSOR_PUSH(cursor, ppage);
1190 mdbenv_get_page(MDB_env *env, pgno_t pgno)
1193 MDB_txn *txn = env->me_txn;
1195 if (txn && pgno >= txn->mt_first_pgno &&
1196 !SIMPLEQ_EMPTY(txn->mt_u.dirty_queue)) {
1198 SIMPLEQ_FOREACH(dp, txn->mt_u.dirty_queue, h.md_next) {
1199 if (dp->p.mp_pgno == pgno) {
1205 p = (MDB_page *)(env->me_map + env->me_head.mh_psize * pgno);
1211 mdb_search_page_root(MDB_db *bt, MDB_val *key,
1212 MDB_cursor *cursor, int modify, MDB_pageparent *mpp)
1214 MDB_page *mp = mpp->mp_page;
1217 if (cursor && cursor_push_page(cursor, mp) == NULL)
1220 while (IS_BRANCH(mp)) {
1224 DPRINTF("branch page %lu has %lu keys", mp->mp_pgno, NUMKEYS(mp));
1225 assert(NUMKEYS(mp) > 1);
1226 DPRINTF("found index 0 to page %lu", NODEPGNO(NODEPTR(mp, 0)));
1228 if (key == NULL) /* Initialize cursor to first page. */
1232 node = mdb_search_node(bt, mp, key, &exact, &i);
1234 i = NUMKEYS(mp) - 1;
1242 DPRINTF("following index %u for key %.*s",
1243 i, (int)key->mv_size, (char *)key->mv_data);
1244 assert(i >= 0 && i < NUMKEYS(mp));
1245 node = NODEPTR(mp, i);
1248 CURSOR_TOP(cursor)->mp_ki = i;
1250 mpp->mp_parent = mp;
1251 if ((mp = mdbenv_get_page(bt->md_env, NODEPGNO(node))) == NULL)
1256 if (cursor && cursor_push_page(cursor, mp) == NULL)
1260 MDB_dhead *dh = ((MDB_dhead *)mp)-1;
1261 if (rc = mdb_touch(bt->md_env->me_txn, mpp))
1263 dh = ((MDB_dhead *)mpp->mp_page)-1;
1264 dh->md_parent = mpp->mp_parent;
1265 dh->md_pi = mpp->mp_pi;
1272 DPRINTF("internal error, index points to a %02X page!?",
1277 DPRINTF("found leaf page %lu for key %.*s", mp->mp_pgno,
1278 key ? (int)key->mv_size : 0, key ? (char *)key->mv_data : NULL);
1283 /* Search for the page a given key should be in.
1284 * Stores a pointer to the found page in *mpp.
1285 * If key is NULL, search for the lowest page (used by mdb_cursor_first).
1286 * If cursor is non-null, pushes parent pages on the cursor stack.
1287 * If modify is true, visited pages are updated with new page numbers.
1290 mdb_search_page(MDB_db *db, MDB_txn *txn, MDB_val *key,
1291 MDB_cursor *cursor, int modify, MDB_pageparent *mpp)
1296 /* Can't modify pages outside a transaction. */
1297 if (txn == NULL && modify) {
1301 /* Choose which root page to start with. If a transaction is given
1302 * use the root page from the transaction, otherwise read the last
1303 * committed root page.
1306 if ((rc = mdbenv_read_meta(db->md_env, NULL)) != MDB_SUCCESS)
1308 root = db->md_env->me_meta.mm_root;
1309 } else if (F_ISSET(txn->mt_flags, MDB_TXN_ERROR)) {
1310 DPRINTF("transaction has failed, must abort");
1313 root = txn->mt_root;
1315 if (root == P_INVALID) { /* Tree is empty. */
1316 DPRINTF("tree is empty");
1320 if ((mpp->mp_page = mdbenv_get_page(db->md_env, root)) == NULL)
1323 DPRINTF("root page has flags 0x%X", mpp->mp_page->mp_flags);
1325 if (modify && !F_ISSET(mpp->mp_page->mp_flags, P_DIRTY)) {
1326 mpp->mp_parent = NULL;
1328 if ((rc = mdb_touch(txn, mpp)))
1330 txn->mt_root = mpp->mp_page->mp_pgno;
1333 return mdb_search_page_root(db, key, cursor, modify, mpp);
1337 mdb_read_data(MDB_db *db, MDB_page *mp, MDB_node *leaf,
1340 MDB_page *omp; /* overflow mpage */
1346 bzero(data, sizeof(*data));
1347 max = db->md_env->me_head.mh_psize - PAGEHDRSZ;
1349 if (!F_ISSET(leaf->mn_flags, F_BIGDATA)) {
1350 data->mv_size = leaf->mn_dsize;
1351 data->mv_data = NODEDATA(leaf);
1355 /* Read overflow data.
1357 data->mv_size = leaf->mn_dsize;
1358 bcopy(NODEDATA(leaf), &pgno, sizeof(pgno));
1359 if ((omp = mdbenv_get_page(db->md_env, pgno)) == NULL) {
1360 DPRINTF("read overflow page %lu failed", pgno);
1363 data->mv_data = omp;
1369 mdb_get(MDB_db *db, MDB_txn *txn,
1370 MDB_val *key, MDB_val *data)
1378 DPRINTF("===> get key [%.*s]", (int)key->mv_size, (char *)key->mv_data);
1383 if (txn != NULL && db->md_env != txn->mt_env) {
1387 if (key->mv_size == 0 || key->mv_size > MAXKEYSIZE) {
1391 if ((rc = mdb_search_page(db, txn, key, NULL, 0, &mpp)) != MDB_SUCCESS)
1394 leaf = mdb_search_node(db, mpp.mp_page, key, &exact, NULL);
1396 rc = mdb_read_data(db, mpp.mp_page, leaf, data);
1405 mdb_sibling(MDB_cursor *cursor, int move_right)
1409 MDB_ppage *parent, *top;
1412 top = CURSOR_TOP(cursor);
1413 if ((parent = SLIST_NEXT(top, mp_entry)) == NULL) {
1414 return ENOENT; /* root has no siblings */
1417 DPRINTF("parent page is page %lu, index %u",
1418 parent->mp_page->mp_pgno, parent->mp_ki);
1420 cursor_pop_page(cursor);
1421 if (move_right ? (parent->mp_ki + 1 >= NUMKEYS(parent->mp_page))
1422 : (parent->mp_ki == 0)) {
1423 DPRINTF("no more keys left, moving to %s sibling",
1424 move_right ? "right" : "left");
1425 if ((rc = mdb_sibling(cursor, move_right)) != MDB_SUCCESS)
1427 parent = CURSOR_TOP(cursor);
1433 DPRINTF("just moving to %s index key %u",
1434 move_right ? "right" : "left", parent->mp_ki);
1436 assert(IS_BRANCH(parent->mp_page));
1438 indx = NODEPTR(parent->mp_page, parent->mp_ki);
1439 if ((mp = mdbenv_get_page(cursor->mc_db->md_env, indx->mn_pgno)) == NULL)
1442 mp->parent = parent->mp_page;
1443 mp->parent_index = parent->mp_ki;
1446 cursor_push_page(cursor, mp);
1452 mdb_set_key(MDB_db *bt, MDB_page *mp, MDB_node *node,
1458 key->mv_size = node->mn_ksize;
1459 key->mv_data = NODEKEY(node);
1465 mdb_cursor_next(MDB_cursor *cursor, MDB_val *key, MDB_val *data)
1471 if (cursor->mc_eof) {
1475 assert(cursor->mc_initialized);
1477 top = CURSOR_TOP(cursor);
1480 DPRINTF("cursor_next: top page is %lu in cursor %p", mp->mp_pgno, cursor);
1482 if (top->mp_ki + 1 >= NUMKEYS(mp)) {
1483 DPRINTF("=====> move to next sibling page");
1484 if (mdb_sibling(cursor, 1) != MDB_SUCCESS) {
1488 top = CURSOR_TOP(cursor);
1490 DPRINTF("next page is %lu, key index %u", mp->mp_pgno, top->mp_ki);
1494 DPRINTF("==> cursor points to page %lu with %lu keys, key index %u",
1495 mp->mp_pgno, NUMKEYS(mp), top->mp_ki);
1497 assert(IS_LEAF(mp));
1498 leaf = NODEPTR(mp, top->mp_ki);
1500 if (data && mdb_read_data(cursor->mc_db, mp, leaf, data) != MDB_SUCCESS)
1503 return mdb_set_key(cursor->mc_db, mp, leaf, key);
1507 mdb_cursor_set(MDB_cursor *cursor, MDB_val *key, MDB_val *data,
1517 assert(key->mv_size > 0);
1519 rc = mdb_search_page(cursor->mc_db, cursor->mc_txn, key, cursor, 0, &mpp);
1520 if (rc != MDB_SUCCESS)
1522 assert(IS_LEAF(mpp.mp_page));
1524 top = CURSOR_TOP(cursor);
1525 leaf = mdb_search_node(cursor->mc_db, mpp.mp_page, key, exactp, &top->mp_ki);
1526 if (exactp != NULL && !*exactp) {
1527 /* MDB_CURSOR_EXACT specified and not an exact match. */
1532 DPRINTF("===> inexact leaf not found, goto sibling");
1533 if ((rc = mdb_sibling(cursor, 1)) != MDB_SUCCESS)
1534 return rc; /* no entries matched */
1535 top = CURSOR_TOP(cursor);
1537 mpp.mp_page = top->mp_page;
1538 assert(IS_LEAF(mpp.mp_page));
1539 leaf = NODEPTR(mpp.mp_page, 0);
1542 cursor->mc_initialized = 1;
1545 if (data && (rc = mdb_read_data(cursor->mc_db, mpp.mp_page, leaf, data)) != MDB_SUCCESS)
1548 rc = mdb_set_key(cursor->mc_db, mpp.mp_page, leaf, key);
1549 if (rc == MDB_SUCCESS) {
1550 DPRINTF("==> cursor placed on key %.*s",
1551 (int)key->mv_size, (char *)key->mv_data);
1559 mdb_cursor_first(MDB_cursor *cursor, MDB_val *key, MDB_val *data)
1565 rc = mdb_search_page(cursor->mc_db, cursor->mc_txn, NULL, cursor, 0, &mpp);
1566 if (rc != MDB_SUCCESS)
1568 assert(IS_LEAF(mpp.mp_page));
1570 leaf = NODEPTR(mpp.mp_page, 0);
1571 cursor->mc_initialized = 1;
1574 if (data && (rc = mdb_read_data(cursor->mc_db, mpp.mp_page, leaf, data)) != MDB_SUCCESS)
1577 return mdb_set_key(cursor->mc_db, mpp.mp_page, leaf, key);
1581 mdb_cursor_get(MDB_cursor *cursor, MDB_val *key, MDB_val *data,
1591 case MDB_CURSOR_EXACT:
1592 while (CURSOR_TOP(cursor) != NULL)
1593 cursor_pop_page(cursor);
1594 if (key == NULL || key->mv_size == 0 || key->mv_size > MAXKEYSIZE) {
1596 } else if (op == MDB_CURSOR_EXACT)
1597 rc = mdb_cursor_set(cursor, key, data, &exact);
1599 rc = mdb_cursor_set(cursor, key, data, NULL);
1602 if (!cursor->mc_initialized)
1603 rc = mdb_cursor_first(cursor, key, data);
1605 rc = mdb_cursor_next(cursor, key, data);
1608 while (CURSOR_TOP(cursor) != NULL)
1609 cursor_pop_page(cursor);
1610 rc = mdb_cursor_first(cursor, key, data);
1613 DPRINTF("unhandled/unimplemented cursor operation %u", op);
1621 /* Allocate a page and initialize it
1624 mdbenv_new_page(MDB_env *env, uint32_t flags, int num)
1628 assert(env != NULL);
1629 assert(env->me_txn != NULL);
1631 DPRINTF("allocating new mpage %lu, page size %u",
1632 env->me_txn->mt_next_pgno, env->me_head.mh_psize);
1633 if ((dp = mdb_newpage(env->me_txn, NULL, 0, num)) == NULL)
1635 dp->p.mp_flags = flags | P_DIRTY;
1636 dp->p.mp_lower = PAGEHDRSZ;
1637 dp->p.mp_upper = env->me_head.mh_psize;
1639 if (IS_BRANCH(&dp->p))
1640 env->me_meta.mm_stat.ms_branch_pages++;
1641 else if (IS_LEAF(&dp->p))
1642 env->me_meta.mm_stat.ms_leaf_pages++;
1643 else if (IS_OVERFLOW(&dp->p)) {
1644 env->me_meta.mm_stat.ms_overflow_pages += num;
1645 dp->p.mp_pages = num;
1652 mdb_leaf_size(MDB_db *db, MDB_val *key, MDB_val *data)
1656 sz = LEAFSIZE(key, data);
1657 if (data->mv_size >= db->md_env->me_head.mh_psize / MDB_MINKEYS) {
1658 /* put on overflow page */
1659 sz -= data->mv_size - sizeof(pgno_t);
1662 return sz + sizeof(indx_t);
1666 mdb_branch_size(MDB_db *db, MDB_val *key)
1671 if (sz >= db->md_env->me_head.mh_psize / MDB_MINKEYS) {
1672 /* put on overflow page */
1673 /* not implemented */
1674 /* sz -= key->size - sizeof(pgno_t); */
1677 return sz + sizeof(indx_t);
1681 mdb_add_node(MDB_db *db, MDB_page *mp, indx_t indx,
1682 MDB_val *key, MDB_val *data, pgno_t pgno, uint8_t flags)
1685 size_t node_size = NODESIZE;
1688 MDB_dpage *ofp = NULL; /* overflow page */
1690 assert(mp->mp_upper >= mp->mp_lower);
1692 DPRINTF("add node [%.*s] to %s page %lu at index %i, key size %zu",
1693 key ? (int)key->mv_size : 0, key ? (char *)key->mv_data : NULL,
1694 IS_LEAF(mp) ? "leaf" : "branch",
1695 mp->mp_pgno, indx, key ? key->mv_size : 0);
1698 node_size += key->mv_size;
1702 node_size += data->mv_size;
1703 if (F_ISSET(flags, F_BIGDATA)) {
1704 /* Data already on overflow page. */
1705 node_size -= data->mv_size - sizeof(pgno_t);
1706 } else if (data->mv_size >= db->md_env->me_head.mh_psize / MDB_MINKEYS) {
1707 int ovpages = PAGEHDRSZ + data->mv_size + db->md_env->me_head.mh_psize - 1;
1708 ovpages /= db->md_env->me_head.mh_psize;
1709 /* Put data on overflow page. */
1710 DPRINTF("data size is %zu, put on overflow page",
1712 node_size -= data->mv_size - sizeof(pgno_t);
1713 if ((ofp = mdbenv_new_page(db->md_env, P_OVERFLOW, ovpages)) == NULL)
1715 DPRINTF("allocated overflow page %lu", ofp->p.mp_pgno);
1720 if (node_size + sizeof(indx_t) > SIZELEFT(mp)) {
1721 DPRINTF("not enough room in page %lu, got %lu ptrs",
1722 mp->mp_pgno, NUMKEYS(mp));
1723 DPRINTF("upper - lower = %u - %u = %u", mp->mp_upper, mp->mp_lower,
1724 mp->mp_upper - mp->mp_lower);
1725 DPRINTF("node size = %zu", node_size);
1729 /* Move higher pointers up one slot. */
1730 for (i = NUMKEYS(mp); i > indx; i--)
1731 mp->mp_ptrs[i] = mp->mp_ptrs[i - 1];
1733 /* Adjust free space offsets. */
1734 ofs = mp->mp_upper - node_size;
1735 assert(ofs >= mp->mp_lower + sizeof(indx_t));
1736 mp->mp_ptrs[indx] = ofs;
1738 mp->mp_lower += sizeof(indx_t);
1740 /* Write the node data. */
1741 node = NODEPTR(mp, indx);
1742 node->mn_ksize = (key == NULL) ? 0 : key->mv_size;
1743 node->mn_flags = flags;
1745 node->mn_dsize = data->mv_size;
1747 node->mn_pgno = pgno;
1750 bcopy(key->mv_data, NODEKEY(node), key->mv_size);
1755 if (F_ISSET(flags, F_BIGDATA))
1756 bcopy(data->mv_data, node->mn_data + key->mv_size,
1759 bcopy(data->mv_data, node->mn_data + key->mv_size,
1762 bcopy(&ofp->p.mp_pgno, node->mn_data + key->mv_size,
1764 bcopy(data->mv_data, METADATA(&ofp->p), data->mv_size);
1772 mdb_del_node(MDB_db *db, MDB_page *mp, indx_t indx)
1775 indx_t i, j, numkeys, ptr;
1779 DPRINTF("delete node %u on %s page %lu", indx,
1780 IS_LEAF(mp) ? "leaf" : "branch", mp->mp_pgno);
1781 assert(indx < NUMKEYS(mp));
1783 node = NODEPTR(mp, indx);
1784 sz = NODESIZE + node->mn_ksize;
1786 if (F_ISSET(node->mn_flags, F_BIGDATA))
1787 sz += sizeof(pgno_t);
1789 sz += NODEDSZ(node);
1792 ptr = mp->mp_ptrs[indx];
1793 numkeys = NUMKEYS(mp);
1794 for (i = j = 0; i < numkeys; i++) {
1796 mp->mp_ptrs[j] = mp->mp_ptrs[i];
1797 if (mp->mp_ptrs[i] < ptr)
1798 mp->mp_ptrs[j] += sz;
1803 base = (char *)mp + mp->mp_upper;
1804 bcopy(base, base + sz, ptr - mp->mp_upper);
1806 mp->mp_lower -= sizeof(indx_t);
1811 mdb_cursor_open(MDB_db *db, MDB_txn *txn, MDB_cursor **ret)
1815 if (db == NULL || ret == NULL)
1818 if (txn != NULL && db->md_env != txn->mt_env) {
1822 if ((cursor = calloc(1, sizeof(*cursor))) != NULL) {
1823 SLIST_INIT(&cursor->mc_stack);
1825 cursor->mc_txn = txn;
1834 mdb_cursor_close(MDB_cursor *cursor)
1836 if (cursor != NULL) {
1837 while (!CURSOR_EMPTY(cursor))
1838 cursor_pop_page(cursor);
1840 /* btree_close(cursor->bt); */
1846 mdb_update_key(MDB_db *db, MDB_page *mp, indx_t indx,
1849 indx_t ptr, i, numkeys;
1855 node = NODEPTR(mp, indx);
1856 ptr = mp->mp_ptrs[indx];
1857 DPRINTF("update key %u (ofs %u) [%.*s] to [%.*s] on page %lu",
1859 (int)node->mn_ksize, (char *)NODEKEY(node),
1860 (int)key->mv_size, (char *)key->mv_data,
1863 delta = key->mv_size - node->mn_ksize;
1865 if (delta > 0 && SIZELEFT(mp) < delta) {
1866 DPRINTF("OUCH! Not enough room, delta = %d", delta);
1870 numkeys = NUMKEYS(mp);
1871 for (i = 0; i < numkeys; i++) {
1872 if (mp->mp_ptrs[i] <= ptr)
1873 mp->mp_ptrs[i] -= delta;
1876 base = (char *)mp + mp->mp_upper;
1877 len = ptr - mp->mp_upper + NODESIZE;
1878 bcopy(base, base - delta, len);
1879 mp->mp_upper -= delta;
1881 node = NODEPTR(mp, indx);
1882 node->mn_ksize = key->mv_size;
1885 bcopy(key->mv_data, NODEKEY(node), key->mv_size);
1890 /* Move a node from src to dst.
1893 mdb_move_node(MDB_db *bt, MDB_pageparent *src, indx_t srcindx,
1894 MDB_pageparent *dst, indx_t dstindx)
1900 srcnode = NODEPTR(src->mp_page, srcindx);
1901 DPRINTF("moving %s node %u [%.*s] on page %lu to node %u on page %lu",
1902 IS_LEAF(src->mp_page) ? "leaf" : "branch",
1904 (int)srcnode->mn_ksize, (char *)NODEKEY(srcnode),
1905 src->mp_page->mp_pgno,
1906 dstindx, dst->mp_page->mp_pgno);
1908 /* Mark src and dst as dirty. */
1909 if ((rc = mdb_touch(bt->md_env->me_txn, src)) ||
1910 (rc = mdb_touch(bt->md_env->me_txn, dst)))
1913 /* Add the node to the destination page.
1915 key.mv_size = srcnode->mn_ksize;
1916 key.mv_data = NODEKEY(srcnode);
1917 data.mv_size = NODEDSZ(srcnode);
1918 data.mv_data = NODEDATA(srcnode);
1919 rc = mdb_add_node(bt, dst->mp_page, dstindx, &key, &data, NODEPGNO(srcnode),
1921 if (rc != MDB_SUCCESS)
1924 /* Delete the node from the source page.
1926 mdb_del_node(bt, src->mp_page, srcindx);
1928 /* Update the parent separators.
1930 if (srcindx == 0 && src->mp_pi != 0) {
1931 DPRINTF("update separator for source page %lu to [%.*s]",
1932 src->mp_page->mp_pgno, (int)key.mv_size, (char *)key.mv_data);
1933 if ((rc = mdb_update_key(bt, src->mp_parent, src->mp_pi,
1934 &key)) != MDB_SUCCESS)
1938 if (srcindx == 0 && IS_BRANCH(src->mp_page)) {
1940 nullkey.mv_size = 0;
1941 assert(mdb_update_key(bt, src->mp_page, 0, &nullkey) == MDB_SUCCESS);
1944 if (dstindx == 0 && dst->mp_pi != 0) {
1945 DPRINTF("update separator for destination page %lu to [%.*s]",
1946 dst->mp_page->mp_pgno, (int)key.mv_size, (char *)key.mv_data);
1947 if ((rc = mdb_update_key(bt, dst->mp_parent, dst->mp_pi,
1948 &key)) != MDB_SUCCESS)
1952 if (dstindx == 0 && IS_BRANCH(dst->mp_page)) {
1954 nullkey.mv_size = 0;
1955 assert(mdb_update_key(bt, dst->mp_page, 0, &nullkey) == MDB_SUCCESS);
1962 mdb_merge(MDB_db *bt, MDB_pageparent *src, MDB_pageparent *dst)
1971 DPRINTF("merging page %lu and %lu", src->mp_page->mp_pgno, dst->mp_page->mp_pgno);
1973 assert(src->mp_parent); /* can't merge root page */
1974 assert(dst->mp_parent);
1975 assert(bt->md_env->me_txn != NULL);
1977 /* Mark src and dst as dirty. */
1978 if ((rc = mdb_touch(bt->md_env->me_txn, src)) ||
1979 (rc = mdb_touch(bt->md_env->me_txn, dst)))
1982 /* Move all nodes from src to dst.
1984 for (i = 0; i < NUMKEYS(src->mp_page); i++) {
1985 srcnode = NODEPTR(src->mp_page, i);
1987 key.mv_size = srcnode->mn_ksize;
1988 key.mv_data = NODEKEY(srcnode);
1989 data.mv_size = NODEDSZ(srcnode);
1990 data.mv_data = NODEDATA(srcnode);
1991 rc = mdb_add_node(bt, dst->mp_page, NUMKEYS(dst->mp_page), &key,
1992 &data, NODEPGNO(srcnode), srcnode->mn_flags);
1993 if (rc != MDB_SUCCESS)
1997 DPRINTF("dst page %lu now has %lu keys (%.1f%% filled)",
1998 dst->mp_page->mp_pgno, NUMKEYS(dst->mp_page), (float)PAGEFILL(bt->md_env, dst->mp_page) / 10);
2000 /* Unlink the src page from parent.
2002 mdb_del_node(bt, src->mp_parent, src->mp_pi);
2003 if (src->mp_pi == 0) {
2005 if ((rc = mdb_update_key(bt, src->mp_parent, 0, &key)) != MDB_SUCCESS)
2009 if (IS_LEAF(src->mp_page))
2010 bt->md_env->me_meta.mm_stat.ms_leaf_pages--;
2012 bt->md_env->me_meta.mm_stat.ms_branch_pages--;
2014 mpp.mp_page = src->mp_parent;
2015 dh = (MDB_dhead *)src->mp_parent;
2017 mpp.mp_parent = dh->md_parent;
2018 mpp.mp_pi = dh->md_pi;
2020 return mdb_rebalance(bt, &mpp);
2023 #define FILL_THRESHOLD 250
2026 mdb_rebalance(MDB_db *db, MDB_pageparent *mpp)
2031 indx_t si = 0, di = 0;
2034 assert(db->md_env->me_txn != NULL);
2035 assert(mpp != NULL);
2037 DPRINTF("rebalancing %s page %lu (has %lu keys, %.1f%% full)",
2038 IS_LEAF(mpp->mp_page) ? "leaf" : "branch",
2039 mpp->mp_page->mp_pgno, NUMKEYS(mpp->mp_page), (float)PAGEFILL(db->md_env, mpp->mp_page) / 10);
2041 if (PAGEFILL(db->md_env, mpp->mp_page) >= FILL_THRESHOLD) {
2042 DPRINTF("no need to rebalance page %lu, above fill threshold",
2043 mpp->mp_page->mp_pgno);
2047 if (mpp->mp_parent == NULL) {
2048 if (NUMKEYS(mpp->mp_page) == 0) {
2049 DPRINTF("tree is completely empty");
2050 db->md_env->me_txn->mt_root = P_INVALID;
2051 db->md_env->me_meta.mm_stat.ms_depth--;
2052 db->md_env->me_meta.mm_stat.ms_leaf_pages--;
2053 } else if (IS_BRANCH(mpp->mp_page) && NUMKEYS(mpp->mp_page) == 1) {
2054 DPRINTF("collapsing root page!");
2055 db->md_env->me_txn->mt_root = NODEPGNO(NODEPTR(mpp->mp_page, 0));
2056 if ((root = mdbenv_get_page(db->md_env, db->md_env->me_txn->mt_root)) == NULL)
2058 db->md_env->me_meta.mm_stat.ms_depth--;
2059 db->md_env->me_meta.mm_stat.ms_branch_pages--;
2061 DPRINTF("root page doesn't need rebalancing");
2065 /* The parent (branch page) must have at least 2 pointers,
2066 * otherwise the tree is invalid.
2068 assert(NUMKEYS(mpp->mp_parent) > 1);
2070 /* Leaf page fill factor is below the threshold.
2071 * Try to move keys from left or right neighbor, or
2072 * merge with a neighbor page.
2077 if (mpp->mp_pi == 0) {
2078 /* We're the leftmost leaf in our parent.
2080 DPRINTF("reading right neighbor");
2081 node = NODEPTR(mpp->mp_parent, mpp->mp_pi + 1);
2082 if ((npp.mp_page = mdbenv_get_page(db->md_env, NODEPGNO(node))) == NULL)
2084 npp.mp_pi = mpp->mp_pi + 1;
2086 di = NUMKEYS(mpp->mp_page);
2088 /* There is at least one neighbor to the left.
2090 DPRINTF("reading left neighbor");
2091 node = NODEPTR(mpp->mp_parent, mpp->mp_pi - 1);
2092 if ((npp.mp_page = mdbenv_get_page(db->md_env, NODEPGNO(node))) == NULL)
2094 npp.mp_pi = mpp->mp_pi - 1;
2095 si = NUMKEYS(npp.mp_page) - 1;
2098 npp.mp_parent = mpp->mp_parent;
2100 DPRINTF("found neighbor page %lu (%lu keys, %.1f%% full)",
2101 npp.mp_page->mp_pgno, NUMKEYS(npp.mp_page), (float)PAGEFILL(db->md_env, npp.mp_page) / 10);
2103 /* If the neighbor page is above threshold and has at least two
2104 * keys, move one key from it.
2106 * Otherwise we should try to merge them.
2108 if (PAGEFILL(db->md_env, npp.mp_page) >= FILL_THRESHOLD && NUMKEYS(npp.mp_page) >= 2)
2109 return mdb_move_node(db, &npp, si, mpp, di);
2110 else { /* FIXME: if (has_enough_room()) */
2111 if (mpp->mp_pi == 0)
2112 return mdb_merge(db, &npp, mpp);
2114 return mdb_merge(db, mpp, &npp);
2119 mdb_del(MDB_db *bt, MDB_txn *txn,
2120 MDB_val *key, MDB_val *data)
2127 DPRINTF("========> delete key %.*s", (int)key->mv_size, (char *)key->mv_data);
2129 assert(key != NULL);
2131 if (bt == NULL || txn == NULL)
2134 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
2138 if (bt->md_env->me_txn != txn) {
2142 if (key->mv_size == 0 || key->mv_size > MAXKEYSIZE) {
2146 if ((rc = mdb_search_page(bt, txn, key, NULL, 1, &mpp)) != MDB_SUCCESS)
2149 leaf = mdb_search_node(bt, mpp.mp_page, key, &exact, &ki);
2150 if (leaf == NULL || !exact) {
2154 if (data && (rc = mdb_read_data(bt, NULL, leaf, data)) != MDB_SUCCESS)
2157 mdb_del_node(bt, mpp.mp_page, ki);
2158 bt->md_env->me_meta.mm_stat.ms_entries--;
2159 rc = mdb_rebalance(bt, &mpp);
2160 if (rc != MDB_SUCCESS)
2161 txn->mt_flags |= MDB_TXN_ERROR;
2166 /* Split page <*mpp>, and insert <key,(data|newpgno)> in either left or
2167 * right sibling, at index <*newindxp> (as if unsplit). Updates *mpp and
2168 * *newindxp with the actual values after split, ie if *mpp and *newindxp
2169 * refer to a node in the new right sibling page.
2172 mdb_split(MDB_db *bt, MDB_page **mpp, unsigned int *newindxp,
2173 MDB_val *newkey, MDB_val *newdata, pgno_t newpgno)
2176 int rc = MDB_SUCCESS, ins_new = 0;
2179 unsigned int i, j, split_indx;
2181 MDB_page *pright, *p;
2182 MDB_val sepkey, rkey, rdata;
2184 MDB_dpage *mdp, *rdp, *pdp;
2188 assert(bt->md_env != NULL);
2190 dh = ((MDB_dhead *)*mpp) - 1;
2191 mdp = (MDB_dpage *)dh;
2192 newindx = *newindxp;
2194 DPRINTF("-----> splitting %s page %lu and adding [%.*s] at index %i",
2195 IS_LEAF(&mdp->p) ? "leaf" : "branch", mdp->p.mp_pgno,
2196 (int)newkey->mv_size, (char *)newkey->mv_data, *newindxp);
2198 if (mdp->h.md_parent == NULL) {
2199 if ((pdp = mdbenv_new_page(bt->md_env, P_BRANCH, 1)) == NULL)
2202 mdp->h.md_parent = &pdp->p;
2203 bt->md_env->me_txn->mt_root = pdp->p.mp_pgno;
2204 DPRINTF("root split! new root = %lu", pdp->p.mp_pgno);
2205 bt->md_env->me_meta.mm_stat.ms_depth++;
2207 /* Add left (implicit) pointer. */
2208 if (mdb_add_node(bt, &pdp->p, 0, NULL, NULL,
2209 mdp->p.mp_pgno, 0) != MDB_SUCCESS)
2212 DPRINTF("parent branch page is %lu", mdp->h.md_parent->mp_pgno);
2215 /* Create a right sibling. */
2216 if ((rdp = mdbenv_new_page(bt->md_env, mdp->p.mp_flags, 1)) == NULL)
2218 rdp->h.md_parent = mdp->h.md_parent;
2219 rdp->h.md_pi = mdp->h.md_pi + 1;
2220 DPRINTF("new right sibling: page %lu", rdp->p.mp_pgno);
2222 /* Move half of the keys to the right sibling. */
2223 if ((copy = malloc(bt->md_env->me_head.mh_psize)) == NULL)
2225 bcopy(&mdp->p, copy, bt->md_env->me_head.mh_psize);
2226 bzero(&mdp->p.mp_ptrs, bt->md_env->me_head.mh_psize - PAGEHDRSZ);
2227 mdp->p.mp_lower = PAGEHDRSZ;
2228 mdp->p.mp_upper = bt->md_env->me_head.mh_psize;
2230 split_indx = NUMKEYS(copy) / 2 + 1;
2232 /* First find the separating key between the split pages.
2234 bzero(&sepkey, sizeof(sepkey));
2235 if (newindx == split_indx) {
2236 sepkey.mv_size = newkey->mv_size;
2237 sepkey.mv_data = newkey->mv_data;
2239 node = NODEPTR(copy, split_indx);
2240 sepkey.mv_size = node->mn_ksize;
2241 sepkey.mv_data = NODEKEY(node);
2244 DPRINTF("separator is [%.*s]", (int)sepkey.mv_size, (char *)sepkey.mv_data);
2246 /* Copy separator key to the parent.
2248 if (SIZELEFT(rdp->h.md_parent) < mdb_branch_size(bt, &sepkey)) {
2249 rc = mdb_split(bt, &rdp->h.md_parent, &rdp->h.md_pi,
2250 &sepkey, NULL, rdp->p.mp_pgno);
2252 /* Right page might now have changed parent.
2253 * Check if left page also changed parent.
2255 if (rdp->h.md_parent != mdp->h.md_parent &&
2256 mdp->h.md_pi >= NUMKEYS(mdp->h.md_parent)) {
2257 mdp->h.md_parent = rdp->h.md_parent;
2258 mdp->h.md_pi = rdp->h.md_pi - 1;
2261 rc = mdb_add_node(bt, rdp->h.md_parent, rdp->h.md_pi,
2262 &sepkey, NULL, rdp->p.mp_pgno, 0);
2264 if (rc != MDB_SUCCESS) {
2269 for (i = j = 0; i <= NUMKEYS(copy); j++) {
2270 if (i < split_indx) {
2271 /* Re-insert in left sibling. */
2274 /* Insert in right sibling. */
2275 if (i == split_indx)
2276 /* Reset insert index for right sibling. */
2277 j = (i == newindx && ins_new);
2281 if (i == newindx && !ins_new) {
2282 /* Insert the original entry that caused the split. */
2283 rkey.mv_data = newkey->mv_data;
2284 rkey.mv_size = newkey->mv_size;
2285 if (IS_LEAF(&mdp->p)) {
2286 rdata.mv_data = newdata->mv_data;
2287 rdata.mv_size = newdata->mv_size;
2294 /* Update page and index for the new key. */
2297 } else if (i == NUMKEYS(copy)) {
2300 node = NODEPTR(copy, i);
2301 rkey.mv_data = NODEKEY(node);
2302 rkey.mv_size = node->mn_ksize;
2303 if (IS_LEAF(&mdp->p)) {
2304 rdata.mv_data = NODEDATA(node);
2305 rdata.mv_size = node->mn_dsize;
2307 pgno = node->mn_pgno;
2308 flags = node->mn_flags;
2313 if (!IS_LEAF(&mdp->p) && j == 0) {
2314 /* First branch index doesn't need key data. */
2318 rc = mdb_add_node(bt, &pdp->p, j, &rkey, &rdata, pgno,flags);
2326 mdb_put(MDB_db *bt, MDB_txn *txn,
2327 MDB_val *key, MDB_val *data, unsigned int flags)
2329 int rc = MDB_SUCCESS, exact;
2334 assert(key != NULL);
2335 assert(data != NULL);
2337 if (bt == NULL || txn == NULL)
2340 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
2344 if (bt->md_env->me_txn != txn) {
2348 if (key->mv_size == 0 || key->mv_size > MAXKEYSIZE) {
2352 DPRINTF("==> put key %.*s, size %zu, data size %zu",
2353 (int)key->mv_size, (char *)key->mv_data, key->mv_size, data->mv_size);
2355 rc = mdb_search_page(bt, txn, key, NULL, 1, &mpp);
2356 if (rc == MDB_SUCCESS) {
2357 leaf = mdb_search_node(bt, mpp.mp_page, key, &exact, &ki);
2358 if (leaf && exact) {
2359 if (F_ISSET(flags, MDB_NOOVERWRITE)) {
2360 DPRINTF("duplicate key %.*s",
2361 (int)key->mv_size, (char *)key->mv_data);
2364 mdb_del_node(bt, mpp.mp_page, ki);
2366 if (leaf == NULL) { /* append if not found */
2367 ki = NUMKEYS(mpp.mp_page);
2368 DPRINTF("appending key at index %i", ki);
2370 } else if (rc == ENOENT) {
2372 /* new file, just write a root leaf page */
2373 DPRINTF("allocating new root leaf page");
2374 if ((dp = mdbenv_new_page(bt->md_env, P_LEAF, 1)) == NULL) {
2377 mpp.mp_page = &dp->p;
2378 txn->mt_root = mpp.mp_page->mp_pgno;
2379 bt->md_env->me_meta.mm_stat.ms_depth++;
2385 assert(IS_LEAF(mpp.mp_page));
2386 DPRINTF("there are %lu keys, should insert new key at index %i",
2387 NUMKEYS(mpp.mp_page), ki);
2389 if (SIZELEFT(mpp.mp_page) < mdb_leaf_size(bt, key, data)) {
2390 rc = mdb_split(bt, &mpp.mp_page, &ki, key, data, P_INVALID);
2392 /* There is room already in this leaf page. */
2393 rc = mdb_add_node(bt, mpp.mp_page, ki, key, data, 0, 0);
2396 if (rc != MDB_SUCCESS)
2397 txn->mt_flags |= MDB_TXN_ERROR;
2399 bt->md_env->me_meta.mm_stat.ms_entries++;
2406 mdbenv_compact_tree(MDB_env *env, pgno_t pgno, MDB_env *envc)
2410 pgno_t *pnext, next;
2415 /* Get the page and make a copy of it.
2417 if ((mp = mdbenv_get_page(env, pgno)) == NULL)
2419 if ((p = malloc(env->me_head.mh_psize)) == NULL)
2421 bcopy(mp, p, env->me_head.mh_psize);
2423 /* Go through all nodes in the (copied) page and update the
2426 if (F_ISSET(p->mp_flags, P_BRANCH)) {
2427 for (i = 0; i < NUMKEYS(p); i++) {
2428 node = NODEPTR(p, i);
2429 node->mn_pgno = mdbenv_compact_tree(env, node->mn_pgno, envc);
2430 if (node->mn_pgno == P_INVALID) {
2435 } else if (F_ISSET(p->mp_flags, P_LEAF)) {
2436 for (i = 0; i < NUMKEYS(p); i++) {
2437 node = NODEPTR(p, i);
2438 if (F_ISSET(node->mn_flags, F_BIGDATA)) {
2439 bcopy(NODEDATA(node), &next, sizeof(next));
2440 next = mdbenv_compact_tree(env, next, envc);
2441 if (next == P_INVALID) {
2445 bcopy(&next, NODEDATA(node), sizeof(next));
2448 } else if (F_ISSET(p->mp_flags, P_OVERFLOW)) {
2449 ; /* handled below */
2453 pgno = p->mp_pgno = envc->me_txn->mt_next_pgno++;
2454 rc = write(envc->me_fd, p, env->me_head.mh_psize);
2456 if (rc != (ssize_t)env->me_head.mh_psize)
2459 if (F_ISSET(mp->mp_flags, P_OVERFLOW)) {
2460 if (mp->mp_pages > 1) {
2461 size_t len = (mp->mp_pages-1) * env->me_head.mh_psize;
2462 envc->me_txn->mt_next_pgno += mp->mp_pages-1;
2463 rc = write(envc->me_fd, mp+1, len);
2472 mdbenv_compact(MDB_env *env)
2474 char *compact_path = NULL;
2476 MDB_txn *txn, *txnc = NULL;
2480 assert(env != NULL);
2482 if (env->me_path == NULL) {
2486 DPRINTF("compacting mdbenv %p with path %s", env, env->me_path);
2488 rc = mdb_txn_begin(env, 0, &txn);
2491 asprintf(&compact_path, "%s.compact.XXXXXX", env->me_path);
2492 fd = mkstemp(compact_path);
2500 rc = mdbenv_create(&envc);
2501 if (rc) goto failed;
2502 rc = mdbenv_set_mapsize(envc, env->me_mapsize);
2505 rc = mdbenv_open2(envc, env->me_flags);
2506 if (rc) goto failed;
2508 bcopy(&env->me_meta, &envc->me_meta, sizeof(env->me_meta));
2509 envc->me_meta.mm_stat.ms_revisions = 0;
2511 rc = mdb_txn_begin(envc, 0, &txnc);
2512 if (rc) goto failed;
2514 if (env->me_meta.mm_root != P_INVALID) {
2515 root = mdbenv_compact_tree(env, env->me_meta.mm_root, envc);
2516 if (root == P_INVALID)
2518 if ((rc = mdbenv_write_meta(envc, root, 0)) != MDB_SUCCESS)
2524 DPRINTF("renaming %s to %s", compact_path, env->me_path);
2525 if (rename(compact_path, env->me_path) != 0) {
2530 /* Write a "tombstone" meta page so other processes can pick up
2531 * the change and re-open the file.
2533 if (mdbenv_write_meta(env, P_INVALID, MDB_TOMBSTONE) != MDB_SUCCESS)
2537 mdb_txn_abort(txnc);
2544 mdb_txn_abort(txnc);
2545 unlink(compact_path);
2552 mdbenv_get_flags(MDB_env *env, unsigned int *arg)
2557 *arg = (env->me_flags & ~MDB_FIXPADDING);
2562 mdbenv_get_path(MDB_env *env, const char **arg)
2567 *arg = env->me_path;
2572 mdbenv_stat(MDB_env *env, MDB_stat **arg)
2574 if (env == NULL || arg == NULL)
2577 *arg = &env->me_meta.mm_stat;
2582 int mdb_open(MDB_env *env, MDB_txn *txn, const char *name, unsigned int flags, MDB_db **db)
2585 *db = (MDB_db *)&env->me_db;
2591 int mdb_stat(MDB_db *db, MDB_stat **arg)
2593 if (db == NULL || arg == NULL)
2596 *arg = &db->md_stat;
2601 void mdb_close(MDB_db *db)