23 #include <openssl/sha.h>
30 # define DPRINTF(...) do { fprintf(stderr, "%s:%d: ", __func__, __LINE__); \
31 fprintf(stderr, __VA_ARGS__); \
32 fprintf(stderr, "\n"); } while(0)
39 #define MDB_MAGIC 0xBEEFC0DE
41 #define MAXKEYSIZE 255
43 #define P_INVALID 0xFFFFFFFF
45 #define F_ISSET(w, f) (((w) & (f)) == (f))
48 typedef uint16_t indx_t;
50 /* Common header for all page types. Overflow pages
51 * occupy a number of contiguous pages with no
52 * headers on any page after the first.
54 typedef struct MDB_page { /* represents a page of storage */
55 pgno_t mp_pgno; /* page number */
56 #define P_BRANCH 0x01 /* branch page */
57 #define P_LEAF 0x02 /* leaf page */
58 #define P_OVERFLOW 0x04 /* overflow page */
59 #define P_META 0x08 /* meta page */
60 #define P_HEAD 0x10 /* header page */
61 #define P_DIRTY 0x20 /* dirty page */
63 #define mp_lower mp_pb.pb.pb_lower
64 #define mp_upper mp_pb.pb.pb_upper
65 #define mp_pages mp_pb.pb_pages
68 indx_t pb_lower; /* lower bound of free space */
69 indx_t pb_upper; /* upper bound of free space */
71 uint32_t pb_pages; /* number of overflow pages */
73 indx_t mp_ptrs[1]; /* dynamic size */
76 #define PAGEHDRSZ offsetof(MDB_page, mp_ptrs)
78 #define NUMKEYS(p) (((p)->mp_lower - PAGEHDRSZ) >> 1)
79 #define SIZELEFT(p) (indx_t)((p)->mp_upper - (p)->mp_lower)
80 #define PAGEFILL(env, p) (1000 * ((env)->me_head.mh_psize - PAGEHDRSZ - SIZELEFT(p)) / \
81 ((env)->me_head.mh_psize - PAGEHDRSZ))
82 #define IS_LEAF(p) F_ISSET((p)->mp_flags, P_LEAF)
83 #define IS_BRANCH(p) F_ISSET((p)->mp_flags, P_BRANCH)
84 #define IS_OVERFLOW(p) F_ISSET((p)->mp_flags, P_OVERFLOW)
86 typedef struct MDB_head { /* header page content */
90 uint32_t mh_psize; /* page size */
91 void *mh_address; /* address for fixed mapping */
92 size_t mh_mapsize; /* size of mmap region */
95 typedef struct MDB_rootstat {
96 pgno_t mr_root; /* page number of root page */
100 typedef struct MDB_meta { /* meta (footer) page content */
101 pgno_t mm_root; /* page number of root page */
103 pgno_t mm_prev_meta; /* previous meta page number */
104 #define MDB_TOMBSTONE 0x01 /* file is replaced */
106 #define mm_revisions mm_stat.ms_revisions
107 #define mm_created_at mm_stat.ms_created_at
108 uint8_t mm_hash[SHA_DIGEST_LENGTH];
111 typedef struct MDB_dhead { /* a dirty page */
112 SIMPLEQ_ENTRY(MDB_dpage) md_next; /* queue of dirty pages */
114 int md_pi; /* parent index */
118 typedef struct MDB_dpage {
123 SIMPLEQ_HEAD(dirty_queue, MDB_dpage);
125 typedef struct MDB_pageparent {
131 static MDB_dpage *mdb_newpage(MDB_txn *txn, MDB_page *parent, int parent_idx, int num);
132 static int mdb_touch(MDB_txn *txn, MDB_pageparent *mp);
134 typedef struct MDB_ppage { /* ordered list of pages */
135 SLIST_ENTRY(MDB_ppage) mp_entry;
137 unsigned int mp_ki; /* cursor index on page */
139 SLIST_HEAD(page_stack, MDB_ppage);
141 #define CURSOR_EMPTY(c) SLIST_EMPTY(&(c)->mc_stack)
142 #define CURSOR_TOP(c) SLIST_FIRST(&(c)->mc_stack)
143 #define CURSOR_POP(c) SLIST_REMOVE_HEAD(&(c)->mc_stack, mp_entry)
144 #define CURSOR_PUSH(c,p) SLIST_INSERT_HEAD(&(c)->mc_stack, p, mp_entry)
149 struct page_stack mc_stack; /* stack of parent pages */
150 short mc_initialized; /* 1 if initialized */
151 short mc_eof; /* 1 if end is reached */
154 #define METAHASHLEN offsetof(MDB_meta, mm_hash)
155 #define METADATA(p) ((void *)((char *)p + PAGEHDRSZ))
157 typedef struct MDB_node {
158 #define mn_pgno mn_p.np_pgno
159 #define mn_dsize mn_p.np_dsize
161 pgno_t np_pgno; /* child page number */
162 uint32_t np_dsize; /* leaf data size */
164 uint16_t mn_ksize; /* key size */
165 #define F_BIGDATA 0x01 /* data put on overflow page */
171 pgno_t mt_root; /* current / new root page */
172 pgno_t mt_next_pgno; /* next unallocated page */
173 pgno_t mt_first_pgno;
175 struct dirty_queue *mt_dirty_queue; /* modified pages */
176 #define MDB_TXN_RDONLY 0x01 /* read-only transaction */
177 #define MDB_TXN_ERROR 0x02 /* an error has occurred */
178 unsigned int mt_flags;
181 /* Must be same as MDB_db, minus md_root/md_stat */
182 typedef struct MDB_db0 {
183 unsigned int md_flags;
184 MDB_cmp_func *md_cmp; /* user compare function */
185 MDB_rel_func *md_rel; /* user relocate function */
186 MDB_db *md_parent; /* parent tree */
191 unsigned int md_flags;
192 MDB_cmp_func *md_cmp; /* user compare function */
193 MDB_rel_func *md_rel; /* user relocate function */
194 MDB_db *md_parent; /* parent tree */
196 pgno_t md_root; /* page number of root page */
202 #define MDB_FIXPADDING 0x01 /* internal */
207 MDB_db0 me_db; /* first DB, overlaps with meta */
209 MDB_txn *me_txn; /* current write transaction */
211 off_t me_size; /* current file size */
214 #define NODESIZE offsetof(MDB_node, mn_data)
216 #define INDXSIZE(k) (NODESIZE + ((k) == NULL ? 0 : (k)->mv_size))
217 #define LEAFSIZE(k, d) (NODESIZE + (k)->mv_size + (d)->mv_size)
218 #define NODEPTR(p, i) ((MDB_node *)((char *)(p) + (p)->mp_ptrs[i]))
219 #define NODEKEY(node) (void *)((node)->mn_data)
220 #define NODEDATA(node) (void *)((char *)(node)->mn_data + (node)->mn_ksize)
221 #define NODEPGNO(node) ((node)->mn_pgno)
222 #define NODEDSZ(node) ((node)->mn_dsize)
224 #define MDB_COMMIT_PAGES 64 /* max number of pages to write in one commit */
225 #define MDB_MAXCACHE_DEF 1024 /* max number of pages to keep in cache */
227 static int mdb_search_page_root(MDB_db *db,
229 MDB_cursor *cursor, int modify,
230 MDB_pageparent *mpp);
231 static int mdb_search_page(MDB_db *db,
232 MDB_txn *txn, MDB_val *key,
233 MDB_cursor *cursor, int modify,
234 MDB_pageparent *mpp);
236 static int mdbenv_write_header(MDB_env *env);
237 static int mdbenv_read_header(MDB_env *env);
238 static int mdb_check_meta_page(MDB_page *p);
239 static int mdbenv_read_meta(MDB_env *env, pgno_t *p_next);
240 static int mdbenv_write_meta(MDB_env *env, pgno_t root,
242 static MDB_page *mdbenv_get_page(MDB_env *env, pgno_t pgno);
244 static MDB_node *mdb_search_node(MDB_db *db, MDB_page *mp,
245 MDB_val *key, int *exactp, unsigned int *kip);
246 static int mdb_add_node(MDB_db *bt, MDB_page *mp,
247 indx_t indx, MDB_val *key, MDB_val *data,
248 pgno_t pgno, uint8_t flags);
249 static void mdb_del_node(MDB_db *bt, MDB_page *mp,
251 static int mdb_read_data(MDB_db *bt, MDB_page *mp,
252 MDB_node *leaf, MDB_val *data);
254 static int mdb_rebalance(MDB_db *bt, MDB_pageparent *mp);
255 static int mdb_update_key(MDB_db *bt, MDB_page *mp,
256 indx_t indx, MDB_val *key);
257 static int mdb_move_node(MDB_db *bt,
258 MDB_pageparent *src, indx_t srcindx,
259 MDB_pageparent *dst, indx_t dstindx);
260 static int mdb_merge(MDB_db *bt, MDB_pageparent *src,
261 MDB_pageparent *dst);
262 static int mdb_split(MDB_db *bt, MDB_page **mpp,
263 unsigned int *newindxp, MDB_val *newkey,
264 MDB_val *newdata, pgno_t newpgno);
265 static MDB_dpage *mdbenv_new_page(MDB_env *env, uint32_t flags, int num);
267 static void cursor_pop_page(MDB_cursor *cursor);
268 static MDB_ppage *cursor_push_page(MDB_cursor *cursor,
271 static int mdb_set_key(MDB_db *bt, MDB_page *mp,
272 MDB_node *node, MDB_val *key);
273 static int mdb_sibling(MDB_cursor *cursor, int move_right);
274 static int mdb_cursor_next(MDB_cursor *cursor,
275 MDB_val *key, MDB_val *data);
276 static int mdb_cursor_set(MDB_cursor *cursor,
277 MDB_val *key, MDB_val *data, int *exactp);
278 static int mdb_cursor_first(MDB_cursor *cursor,
279 MDB_val *key, MDB_val *data);
281 static size_t mdb_leaf_size(MDB_db *bt, MDB_val *key,
283 static size_t mdb_branch_size(MDB_db *bt, MDB_val *key);
285 static pgno_t mdbenv_compact_tree(MDB_env *env, pgno_t pgno,
288 static int memncmp(const void *s1, size_t n1,
289 const void *s2, size_t n2);
290 static int memnrcmp(const void *s1, size_t n1,
291 const void *s2, size_t n2);
294 memncmp(const void *s1, size_t n1, const void *s2, size_t n2)
297 if (memcmp(s1, s2, n1) == 0)
301 if (memcmp(s1, s2, n2) == 0)
304 return memcmp(s1, s2, n1);
308 memnrcmp(const void *s1, size_t n1, const void *s2, size_t n2)
310 const unsigned char *p1;
311 const unsigned char *p2;
314 return n2 == 0 ? 0 : -1;
317 return n1 == 0 ? 0 : 1;
319 p1 = (const unsigned char *)s1 + n1 - 1;
320 p2 = (const unsigned char *)s2 + n2 - 1;
324 return (p2 == s2) ? 0 : -1;
326 return (p1 == p2) ? 0 : 1;
334 mdb_cmp(MDB_db *db, const MDB_val *a, const MDB_val *b)
336 return db->md_cmp(a, b);
340 _mdb_cmp(MDB_db *db, const MDB_val *key1, const MDB_val *key2)
342 if (F_ISSET(db->md_flags, MDB_REVERSEKEY))
343 return memnrcmp(key1->mv_data, key1->mv_size, key2->mv_data, key2->mv_size);
345 return memncmp((char *)key1->mv_data, key1->mv_size, key2->mv_data, key2->mv_size);
348 /* Allocate new page(s) for writing */
350 mdb_newpage(MDB_txn *txn, MDB_page *parent, int parent_idx, int num)
354 if ((dp = malloc(txn->mt_env->me_head.mh_psize * num + sizeof(MDB_dhead))) == NULL)
357 dp->h.md_parent = parent;
358 dp->h.md_pi = parent_idx;
359 SIMPLEQ_INSERT_TAIL(txn->mt_dirty_queue, dp, h.md_next);
360 dp->p.mp_pgno = txn->mt_next_pgno;
361 txn->mt_next_pgno += num;
366 /* Touch a page: make it dirty and re-insert into tree with updated pgno.
369 mdb_touch(MDB_txn *txn, MDB_pageparent *pp)
372 MDB_page *mp = pp->mp_page;
377 if (!F_ISSET(mp->mp_flags, P_DIRTY)) {
379 DPRINTF("touching page %lu -> %lu", mp->mp_pgno, txn->mt_next_pgno);
380 if ((dp = mdb_newpage(txn, pp->mp_parent, pp->mp_pi, 1)) == NULL)
382 pgno = dp->p.mp_pgno;
383 bcopy(mp, &dp->p, txn->mt_env->me_head.mh_psize);
386 mp->mp_flags |= P_DIRTY;
388 /* Update the page number to new touched page. */
389 if (pp->mp_parent != NULL)
390 NODEPGNO(NODEPTR(pp->mp_parent, pp->mp_pi)) = mp->mp_pgno;
397 mdbenv_sync(MDB_env *env)
400 if (!F_ISSET(env->me_flags, MDB_NOSYNC)) {
401 if (fsync(env->me_fd))
408 mdb_txn_begin(MDB_env *env, int rdonly, MDB_txn **ret)
413 if (!rdonly && env->me_txn != NULL) {
414 DPRINTF("write transaction already begun");
418 if ((txn = calloc(1, sizeof(*txn))) == NULL) {
419 DPRINTF("calloc: %s", strerror(errno));
424 txn->mt_flags |= MDB_TXN_RDONLY;
426 txn->mt_dirty_queue = calloc(1, sizeof(*txn->mt_dirty_queue));
427 if (txn->mt_dirty_queue == NULL) {
431 SIMPLEQ_INIT(txn->mt_dirty_queue);
434 DPRINTF("taking write lock on txn %p", txn);
435 if (flock(bt->fd, LOCK_EX | LOCK_NB) != 0) {
436 DPRINTF("flock: %s", strerror(errno));
438 free(txn->dirty_queue);
448 if ((rc = mdbenv_read_meta(env, &txn->mt_next_pgno)) != MDB_SUCCESS) {
453 txn->mt_first_pgno = txn->mt_next_pgno;
454 txn->mt_root = env->me_meta.mm_root;
455 DPRINTF("begin transaction on mdbenv %p, root page %lu", env, txn->mt_root);
462 mdb_txn_abort(MDB_txn *txn)
471 DPRINTF("abort transaction on mdbenv %p, root page %lu", env, txn->mt_root);
473 if (!F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
474 /* Discard all dirty pages.
476 while (!SIMPLEQ_EMPTY(txn->mt_dirty_queue)) {
477 dp = SIMPLEQ_FIRST(txn->mt_dirty_queue);
478 SIMPLEQ_REMOVE_HEAD(txn->mt_dirty_queue, h.md_next);
483 DPRINTF("releasing write lock on txn %p", txn);
485 if (flock(txn->bt->fd, LOCK_UN) != 0) {
486 DPRINTF("failed to unlock fd %d: %s",
487 txn->bt->fd, strerror(errno));
490 free(txn->mt_dirty_queue);
493 if (txn == env->me_txn)
499 mdb_txn_commit(MDB_txn *txn)
506 struct iovec iov[MDB_COMMIT_PAGES];
509 assert(txn->mt_env != NULL);
513 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
514 DPRINTF("attempt to commit read-only transaction");
519 if (txn != env->me_txn) {
520 DPRINTF("attempt to commit unknown transaction");
525 if (F_ISSET(txn->mt_flags, MDB_TXN_ERROR)) {
526 DPRINTF("error flag is set, can't commit");
531 if (SIMPLEQ_EMPTY(txn->mt_dirty_queue))
534 if (F_ISSET(env->me_flags, MDB_FIXPADDING)) {
535 size = lseek(env->me_fd, 0, SEEK_END);
536 size += env->me_head.mh_psize - (size % env->me_head.mh_psize);
537 DPRINTF("extending to multiple of page size: %lu", size);
538 if (ftruncate(env->me_fd, size) != 0) {
540 DPRINTF("ftruncate: %s", strerror(errno));
544 env->me_flags ^= MDB_FIXPADDING;
547 DPRINTF("committing transaction on mdbenv %p, root page %lu",
550 /* Commit up to MDB_COMMIT_PAGES dirty pages to disk until done.
555 SIMPLEQ_FOREACH(dp, txn->mt_dirty_queue, h.md_next) {
556 DPRINTF("committing page %lu", dp->p.mp_pgno);
557 iov[n].iov_len = env->me_head.mh_psize * dp->h.md_num;
558 iov[n].iov_base = &dp->p;
559 /* clear dirty flag */
560 dp->p.mp_flags &= ~P_DIRTY;
561 if (++n >= MDB_COMMIT_PAGES) {
570 DPRINTF("committing %u dirty pages", n);
571 rc = writev(env->me_fd, iov, n);
572 if (rc != (ssize_t)env->me_head.mh_psize*n) {
575 DPRINTF("short write, filesystem full?");
577 DPRINTF("writev: %s", strerror(errno));
582 /* Drop the dirty pages.
584 while (!SIMPLEQ_EMPTY(txn->mt_dirty_queue)) {
585 dp = SIMPLEQ_FIRST(txn->mt_dirty_queue);
586 SIMPLEQ_REMOVE_HEAD(txn->mt_dirty_queue, h.md_next);
593 if ((n = mdbenv_sync(env)) != 0 ||
594 (n = mdbenv_write_meta(env, txn->mt_root, 0)) != MDB_SUCCESS ||
595 (n = mdbenv_sync(env)) != 0) {
609 mdbenv_write_header(MDB_env *env)
617 DPRINTF("writing header page");
620 psize = sysconf(_SC_PAGE_SIZE);
622 if ((p = calloc(1, psize)) == NULL)
624 p->mp_flags = P_HEAD;
626 env->me_head.mh_psize = psize;
627 env->me_head.mh_flags = env->me_flags & 0xffff;
630 bcopy(&env->me_head, h, sizeof(*h));
632 rc = write(env->me_fd, p, env->me_head.mh_psize);
634 if (rc != (ssize_t)env->me_head.mh_psize) {
637 DPRINTF("short write, filesystem full?");
645 mdbenv_read_header(MDB_env *env)
654 /* We don't know the page size yet, so use a minimum value.
657 if ((rc = pread(env->me_fd, page, PAGESIZE, 0)) == 0) {
659 } else if (rc != PAGESIZE) {
662 DPRINTF("read: %s", strerror(errno));
666 p = (MDB_page *)page;
668 if (!F_ISSET(p->mp_flags, P_HEAD)) {
669 DPRINTF("page %lu not a header page", p->mp_pgno);
674 if (h->mh_magic != MDB_MAGIC) {
675 DPRINTF("header has invalid magic");
679 if (h->mh_version != MDB_VERSION) {
680 DPRINTF("database is version %u, expected version %u",
681 h->mh_version, MDB_VERSION);
685 bcopy(h, &env->me_head, sizeof(*h));
690 mdbenv_write_meta(MDB_env *env, pgno_t root, unsigned int flags)
696 DPRINTF("writing meta page for root page %lu", root);
699 assert(env->me_txn != NULL);
701 if ((dp = mdbenv_new_page(env, P_META, 1)) == NULL)
704 env->me_meta.mm_prev_meta = env->me_meta.mm_root;
705 env->me_meta.mm_root = root;
706 env->me_meta.mm_flags = flags;
707 env->me_meta.mm_created_at = time(0);
708 env->me_meta.mm_revisions++;
709 SHA1((unsigned char *)&env->me_meta, METAHASHLEN, env->me_meta.mm_hash);
711 /* Copy the meta data changes to the new meta page. */
712 meta = METADATA(&dp->p);
713 bcopy(&env->me_meta, meta, sizeof(*meta));
715 rc = write(env->me_fd, &dp->p, env->me_head.mh_psize);
716 SIMPLEQ_REMOVE_HEAD(env->me_txn->mt_dirty_queue, h.md_next);
718 if (rc != (ssize_t)env->me_head.mh_psize) {
721 DPRINTF("short write, filesystem full?");
725 if ((env->me_size = lseek(env->me_fd, 0, SEEK_END)) == -1) {
726 DPRINTF("failed to update file size: %s", strerror(errno));
733 /* Returns true if page p is a valid meta page, false otherwise.
736 mdb_check_meta_page(MDB_page *p)
739 unsigned char hash[SHA_DIGEST_LENGTH];
742 if (!F_ISSET(p->mp_flags, P_META)) {
743 DPRINTF("page %lu not a meta page", p->mp_pgno);
747 if (m->mm_root >= p->mp_pgno && m->mm_root != P_INVALID) {
748 DPRINTF("page %lu points to an invalid root page", p->mp_pgno);
752 SHA1((unsigned char *)m, METAHASHLEN, hash);
753 if (bcmp(hash, m->mm_hash, SHA_DIGEST_LENGTH) != 0) {
754 DPRINTF("page %lu has an invalid digest", p->mp_pgno);
762 mdbenv_read_meta(MDB_env *env, pgno_t *p_next)
766 pgno_t meta_pgno, next_pgno;
771 if ((size = lseek(env->me_fd, 0, SEEK_END)) == -1)
774 DPRINTF("mdbenv_read_meta: size = %lu", size);
776 if (size < env->me_size) {
777 DPRINTF("file has shrunk!");
782 if (size == env->me_head.mh_psize) { /* there is only the header */
785 return MDB_SUCCESS; /* new file */
788 next_pgno = size / env->me_head.mh_psize;
789 if (next_pgno == 0) {
790 DPRINTF("corrupt file");
795 meta_pgno = next_pgno - 1;
797 if (size % env->me_head.mh_psize != 0) {
798 DPRINTF("filesize not a multiple of the page size!");
799 env->me_flags |= MDB_FIXPADDING;
806 if (size == env->me_size) {
807 DPRINTF("size unchanged, keeping current meta page");
808 if (F_ISSET(env->me_meta.mm_flags, MDB_TOMBSTONE)) {
809 DPRINTF("file is dead");
818 while (meta_pgno > 0) {
819 if ((mp = mdbenv_get_page(env, meta_pgno)) == NULL)
821 if (!mdb_check_meta_page(mp)) {
823 DPRINTF("flags = 0x%x", meta->mm_flags);
824 if (F_ISSET(meta->mm_flags, MDB_TOMBSTONE)) {
825 DPRINTF("file is dead");
829 /* Make copy of last meta page. */
830 bcopy(meta, &env->me_meta, sizeof(env->me_meta));
834 --meta_pgno; /* scan backwards to first valid meta page */
845 mdbenv_create(MDB_env **env, size_t size)
849 if (!size) return EINVAL;
851 e = calloc(1, sizeof(*e));
852 if (!e) return ENOMEM;
854 e->me_head.mh_magic = MDB_MAGIC;
855 e->me_head.mh_version = MDB_VERSION;
856 e->me_mapsize = e->me_head.mh_mapsize = size;
863 mdbenv_open2(MDB_env *env, unsigned int flags)
867 i = fcntl(env->me_fd, F_GETFL, 0);
868 if (fcntl(env->me_fd, F_SETFL, i | O_APPEND) == -1)
871 env->me_flags = flags;
872 env->me_flags &= ~MDB_FIXPADDING;
873 env->me_meta.mm_root = P_INVALID;
875 if ((i = mdbenv_read_header(env)) != 0) {
878 DPRINTF("new mdbenv");
883 if (env->me_head.mh_address && (flags & MDB_FIXEDMAP))
885 env->me_map = mmap(env->me_head.mh_address, env->me_mapsize, PROT_READ, i,
887 if (env->me_map == MAP_FAILED)
891 env->me_head.mh_mapsize = env->me_mapsize;
892 if (flags & MDB_FIXEDMAP)
893 env->me_head.mh_address = env->me_map;
894 i = mdbenv_write_header(env);
895 if (i != MDB_SUCCESS) {
896 munmap(env->me_map, env->me_mapsize);
901 if ((i = mdbenv_read_meta(env, NULL)) != 0)
904 DPRINTF("opened database version %u, pagesize %u",
905 env->me_head.mh_version, env->me_head.mh_psize);
906 DPRINTF("timestamp: %s", ctime(&env->me_meta.mm_created_at));
907 DPRINTF("depth: %u", env->me_meta.mm_stat.ms_depth);
908 DPRINTF("entries: %lu", env->me_meta.mm_stat.ms_entries);
909 DPRINTF("revisions: %lu", env->me_meta.mm_stat.ms_revisions);
910 DPRINTF("branch pages: %lu", env->me_meta.mm_stat.ms_branch_pages);
911 DPRINTF("leaf pages: %lu", env->me_meta.mm_stat.ms_leaf_pages);
912 DPRINTF("overflow pages: %lu", env->me_meta.mm_stat.ms_overflow_pages);
913 DPRINTF("root: %lu", env->me_meta.mm_root);
914 DPRINTF("previous me_meta.me_page: %lu", env->me_meta.mm_prev_meta);
920 mdbenv_open(MDB_env *env, const char *path, unsigned int flags, mode_t mode)
924 if (F_ISSET(flags, MDB_RDONLY))
927 oflags = O_RDWR | O_CREAT | O_APPEND;
929 if ((env->me_fd = open(path, oflags, mode)) == -1)
932 if ((rc = mdbenv_open2(env, flags)) != MDB_SUCCESS) {
936 env->me_path = strdup(path);
937 DPRINTF("opened dbenv %p", env);
944 mdbenv_close(MDB_env *env)
952 munmap(env->me_map, env->me_mapsize);
957 /* Search for key within a leaf page, using binary search.
958 * Returns the smallest entry larger or equal to the key.
959 * If exactp is non-null, stores whether the found entry was an exact match
960 * in *exactp (1 or 0).
961 * If kip is non-null, stores the index of the found entry in *kip.
962 * If no entry larger or equal to the key is found, returns NULL.
965 mdb_search_node(MDB_db *bt, MDB_page *mp, MDB_val *key,
966 int *exactp, unsigned int *kip)
974 DPRINTF("searching %lu keys in %s page %lu",
976 IS_LEAF(mp) ? "leaf" : "branch",
979 assert(NUMKEYS(mp) > 0);
981 bzero(&nodekey, sizeof(nodekey));
983 low = IS_LEAF(mp) ? 0 : 1;
984 high = NUMKEYS(mp) - 1;
985 while (low <= high) {
986 i = (low + high) >> 1;
987 node = NODEPTR(mp, i);
989 nodekey.mv_size = node->mn_ksize;
990 nodekey.mv_data = NODEKEY(node);
993 rc = bt->md_cmp(key, &nodekey);
995 rc = _mdb_cmp(bt, key, &nodekey);
998 DPRINTF("found leaf index %u [%.*s], rc = %i",
999 i, (int)nodekey.mv_size, (char *)nodekey.mv_data, rc);
1001 DPRINTF("found branch index %u [%.*s -> %lu], rc = %i",
1002 i, (int)node->mn_ksize, (char *)NODEKEY(node),
1013 if (rc > 0) { /* Found entry is less than the key. */
1014 i++; /* Skip to get the smallest entry larger than key. */
1015 if (i >= NUMKEYS(mp))
1016 /* There is no entry larger or equal to the key. */
1020 *exactp = (rc == 0);
1021 if (kip) /* Store the key index if requested. */
1024 return NODEPTR(mp, i);
1028 cursor_pop_page(MDB_cursor *cursor)
1032 top = CURSOR_TOP(cursor);
1035 DPRINTF("popped page %lu off cursor %p", top->mp_page->mp_pgno, cursor);
1041 cursor_push_page(MDB_cursor *cursor, MDB_page *mp)
1045 DPRINTF("pushing page %lu on cursor %p", mp->mp_pgno, cursor);
1047 if ((ppage = calloc(1, sizeof(*ppage))) == NULL)
1049 ppage->mp_page = mp;
1050 CURSOR_PUSH(cursor, ppage);
1055 mdbenv_get_page(MDB_env *env, pgno_t pgno)
1058 MDB_txn *txn = env->me_txn;
1060 if (txn && pgno >= txn->mt_first_pgno &&
1061 !SIMPLEQ_EMPTY(txn->mt_dirty_queue)) {
1063 SIMPLEQ_FOREACH(dp, txn->mt_dirty_queue, h.md_next) {
1064 if (dp->p.mp_pgno == pgno) {
1070 p = (MDB_page *)(env->me_map + env->me_head.mh_psize * pgno);
1076 mdb_search_page_root(MDB_db *bt, MDB_val *key,
1077 MDB_cursor *cursor, int modify, MDB_pageparent *mpp)
1079 MDB_page *mp = mpp->mp_page;
1082 if (cursor && cursor_push_page(cursor, mp) == NULL)
1085 while (IS_BRANCH(mp)) {
1089 DPRINTF("branch page %lu has %lu keys", mp->mp_pgno, NUMKEYS(mp));
1090 assert(NUMKEYS(mp) > 1);
1091 DPRINTF("found index 0 to page %lu", NODEPGNO(NODEPTR(mp, 0)));
1093 if (key == NULL) /* Initialize cursor to first page. */
1097 node = mdb_search_node(bt, mp, key, &exact, &i);
1099 i = NUMKEYS(mp) - 1;
1107 DPRINTF("following index %u for key %.*s",
1108 i, (int)key->mv_size, (char *)key->mv_data);
1109 assert(i >= 0 && i < NUMKEYS(mp));
1110 node = NODEPTR(mp, i);
1113 CURSOR_TOP(cursor)->mp_ki = i;
1115 mpp->mp_parent = mp;
1116 if ((mp = mdbenv_get_page(bt->md_env, NODEPGNO(node))) == NULL)
1121 if (cursor && cursor_push_page(cursor, mp) == NULL)
1124 if (modify && (rc = mdb_touch(bt->md_env->me_txn, mpp)))
1130 DPRINTF("internal error, index points to a %02X page!?",
1135 DPRINTF("found leaf page %lu for key %.*s", mp->mp_pgno,
1136 key ? (int)key->mv_size : 0, key ? (char *)key->mv_data : NULL);
1141 /* Search for the page a given key should be in.
1142 * Stores a pointer to the found page in *mpp.
1143 * If key is NULL, search for the lowest page (used by mdb_cursor_first).
1144 * If cursor is non-null, pushes parent pages on the cursor stack.
1145 * If modify is true, visited pages are updated with new page numbers.
1148 mdb_search_page(MDB_db *db, MDB_txn *txn, MDB_val *key,
1149 MDB_cursor *cursor, int modify, MDB_pageparent *mpp)
1154 /* Can't modify pages outside a transaction. */
1155 if (txn == NULL && modify) {
1159 /* Choose which root page to start with. If a transaction is given
1160 * use the root page from the transaction, otherwise read the last
1161 * committed root page.
1164 if ((rc = mdbenv_read_meta(db->md_env, NULL)) != MDB_SUCCESS)
1166 root = db->md_env->me_meta.mm_root;
1167 } else if (F_ISSET(txn->mt_flags, MDB_TXN_ERROR)) {
1168 DPRINTF("transaction has failed, must abort");
1171 root = txn->mt_root;
1173 if (root == P_INVALID) { /* Tree is empty. */
1174 DPRINTF("tree is empty");
1178 if ((mpp->mp_page = mdbenv_get_page(db->md_env, root)) == NULL)
1181 DPRINTF("root page has flags 0x%X", mpp->mp_page->mp_flags);
1183 if (modify && !F_ISSET(mpp->mp_page->mp_flags, P_DIRTY)) {
1184 mpp->mp_parent = NULL;
1186 if ((rc = mdb_touch(txn, mpp)))
1188 txn->mt_root = mpp->mp_page->mp_pgno;
1191 return mdb_search_page_root(db, key, cursor, modify, mpp);
1195 mdb_read_data(MDB_db *db, MDB_page *mp, MDB_node *leaf,
1198 MDB_page *omp; /* overflow mpage */
1204 bzero(data, sizeof(*data));
1205 max = db->md_env->me_head.mh_psize - PAGEHDRSZ;
1207 if (!F_ISSET(leaf->mn_flags, F_BIGDATA)) {
1208 data->mv_size = leaf->mn_dsize;
1209 data->mv_data = NODEDATA(leaf);
1213 /* Read overflow data.
1215 data->mv_size = leaf->mn_dsize;
1216 bcopy(NODEDATA(leaf), &pgno, sizeof(pgno));
1217 if ((omp = mdbenv_get_page(db->md_env, pgno)) == NULL) {
1218 DPRINTF("read overflow page %lu failed", pgno);
1221 data->mv_data = omp;
1227 mdb_get(MDB_db *db, MDB_txn *txn,
1228 MDB_val *key, MDB_val *data)
1236 DPRINTF("===> get key [%.*s]", (int)key->mv_size, (char *)key->mv_data);
1241 if (txn != NULL && db->md_env != txn->mt_env) {
1245 if (key->mv_size == 0 || key->mv_size > MAXKEYSIZE) {
1249 if ((rc = mdb_search_page(db, txn, key, NULL, 0, &mpp)) != MDB_SUCCESS)
1252 leaf = mdb_search_node(db, mpp.mp_page, key, &exact, NULL);
1254 rc = mdb_read_data(db, mpp.mp_page, leaf, data);
1263 mdb_sibling(MDB_cursor *cursor, int move_right)
1267 MDB_ppage *parent, *top;
1270 top = CURSOR_TOP(cursor);
1271 if ((parent = SLIST_NEXT(top, mp_entry)) == NULL) {
1272 return ENOENT; /* root has no siblings */
1275 DPRINTF("parent page is page %lu, index %u",
1276 parent->mp_page->mp_pgno, parent->mp_ki);
1278 cursor_pop_page(cursor);
1279 if (move_right ? (parent->mp_ki + 1 >= NUMKEYS(parent->mp_page))
1280 : (parent->mp_ki == 0)) {
1281 DPRINTF("no more keys left, moving to %s sibling",
1282 move_right ? "right" : "left");
1283 if ((rc = mdb_sibling(cursor, move_right)) != MDB_SUCCESS)
1285 parent = CURSOR_TOP(cursor);
1291 DPRINTF("just moving to %s index key %u",
1292 move_right ? "right" : "left", parent->mp_ki);
1294 assert(IS_BRANCH(parent->mp_page));
1297 indx = NODEPTR(parent->mp_page, parent->mp_ki);
1298 if ((mp = mdbenv_get_page(cursor->bt, indx->n_pgno)) == NULL)
1300 mp->parent = parent->mp_page;
1301 mp->parent_index = parent->mp_ki;
1304 cursor_push_page(cursor, mp);
1310 mdb_set_key(MDB_db *bt, MDB_page *mp, MDB_node *node,
1316 key->mv_size = node->mn_ksize;
1317 key->mv_data = NODEKEY(node);
1323 mdb_cursor_next(MDB_cursor *cursor, MDB_val *key, MDB_val *data)
1329 if (cursor->mc_eof) {
1333 assert(cursor->mc_initialized);
1335 top = CURSOR_TOP(cursor);
1338 DPRINTF("cursor_next: top page is %lu in cursor %p", mp->mp_pgno, cursor);
1340 if (top->mp_ki + 1 >= NUMKEYS(mp)) {
1341 DPRINTF("=====> move to next sibling page");
1342 if (mdb_sibling(cursor, 1) != MDB_SUCCESS) {
1346 top = CURSOR_TOP(cursor);
1348 DPRINTF("next page is %lu, key index %u", mp->mp_pgno, top->mp_ki);
1352 DPRINTF("==> cursor points to page %lu with %lu keys, key index %u",
1353 mp->mp_pgno, NUMKEYS(mp), top->mp_ki);
1355 assert(IS_LEAF(mp));
1356 leaf = NODEPTR(mp, top->mp_ki);
1358 if (data && mdb_read_data(cursor->mc_db, mp, leaf, data) != MDB_SUCCESS)
1361 return mdb_set_key(cursor->mc_db, mp, leaf, key);
1365 mdb_cursor_set(MDB_cursor *cursor, MDB_val *key, MDB_val *data,
1375 assert(key->mv_size > 0);
1377 rc = mdb_search_page(cursor->mc_db, cursor->mc_txn, key, cursor, 0, &mpp);
1378 if (rc != MDB_SUCCESS)
1380 assert(IS_LEAF(mpp.mp_page));
1382 top = CURSOR_TOP(cursor);
1383 leaf = mdb_search_node(cursor->mc_db, mpp.mp_page, key, exactp, &top->mp_ki);
1384 if (exactp != NULL && !*exactp) {
1385 /* MDB_CURSOR_EXACT specified and not an exact match. */
1390 DPRINTF("===> inexact leaf not found, goto sibling");
1391 if ((rc = mdb_sibling(cursor, 1)) != MDB_SUCCESS)
1392 return rc; /* no entries matched */
1393 top = CURSOR_TOP(cursor);
1395 mpp.mp_page = top->mp_page;
1396 assert(IS_LEAF(mpp.mp_page));
1397 leaf = NODEPTR(mpp.mp_page, 0);
1400 cursor->mc_initialized = 1;
1403 if (data && (rc = mdb_read_data(cursor->mc_db, mpp.mp_page, leaf, data)) != MDB_SUCCESS)
1406 rc = mdb_set_key(cursor->mc_db, mpp.mp_page, leaf, key);
1407 if (rc == MDB_SUCCESS) {
1408 DPRINTF("==> cursor placed on key %.*s",
1409 (int)key->mv_size, (char *)key->mv_data);
1417 mdb_cursor_first(MDB_cursor *cursor, MDB_val *key, MDB_val *data)
1423 rc = mdb_search_page(cursor->mc_db, cursor->mc_txn, NULL, cursor, 0, &mpp);
1424 if (rc != MDB_SUCCESS)
1426 assert(IS_LEAF(mpp.mp_page));
1428 leaf = NODEPTR(mpp.mp_page, 0);
1429 cursor->mc_initialized = 1;
1432 if (data && (rc = mdb_read_data(cursor->mc_db, mpp.mp_page, leaf, data)) != MDB_SUCCESS)
1435 return mdb_set_key(cursor->mc_db, mpp.mp_page, leaf, key);
1439 mdb_cursor_get(MDB_cursor *cursor, MDB_val *key, MDB_val *data,
1449 case MDB_CURSOR_EXACT:
1450 while (CURSOR_TOP(cursor) != NULL)
1451 cursor_pop_page(cursor);
1452 if (key == NULL || key->mv_size == 0 || key->mv_size > MAXKEYSIZE) {
1454 } else if (op == MDB_CURSOR_EXACT)
1455 rc = mdb_cursor_set(cursor, key, data, &exact);
1457 rc = mdb_cursor_set(cursor, key, data, NULL);
1460 if (!cursor->mc_initialized)
1461 rc = mdb_cursor_first(cursor, key, data);
1463 rc = mdb_cursor_next(cursor, key, data);
1466 while (CURSOR_TOP(cursor) != NULL)
1467 cursor_pop_page(cursor);
1468 rc = mdb_cursor_first(cursor, key, data);
1471 DPRINTF("unhandled/unimplemented cursor operation %u", op);
1479 /* Allocate a page and initialize it
1482 mdbenv_new_page(MDB_env *env, uint32_t flags, int num)
1486 assert(env != NULL);
1487 assert(env->me_txn != NULL);
1489 DPRINTF("allocating new mpage %lu, page size %u",
1490 env->me_txn->mt_next_pgno, env->me_head.mh_psize);
1491 if ((dp = mdb_newpage(env->me_txn, NULL, 0, num)) == NULL)
1493 dp->p.mp_flags = flags;
1494 dp->p.mp_lower = PAGEHDRSZ;
1495 dp->p.mp_upper = env->me_head.mh_psize;
1497 if (IS_BRANCH(&dp->p))
1498 env->me_meta.mm_stat.ms_branch_pages++;
1499 else if (IS_LEAF(&dp->p))
1500 env->me_meta.mm_stat.ms_leaf_pages++;
1501 else if (IS_OVERFLOW(&dp->p)) {
1502 env->me_meta.mm_stat.ms_overflow_pages += num;
1503 dp->p.mp_pages = num;
1510 mdb_leaf_size(MDB_db *db, MDB_val *key, MDB_val *data)
1514 sz = LEAFSIZE(key, data);
1515 if (data->mv_size >= db->md_env->me_head.mh_psize / MDB_MINKEYS) {
1516 /* put on overflow page */
1517 sz -= data->mv_size - sizeof(pgno_t);
1520 return sz + sizeof(indx_t);
1524 mdb_branch_size(MDB_db *db, MDB_val *key)
1529 if (sz >= db->md_env->me_head.mh_psize / MDB_MINKEYS) {
1530 /* put on overflow page */
1531 /* not implemented */
1532 /* sz -= key->size - sizeof(pgno_t); */
1535 return sz + sizeof(indx_t);
1539 mdb_add_node(MDB_db *db, MDB_page *mp, indx_t indx,
1540 MDB_val *key, MDB_val *data, pgno_t pgno, uint8_t flags)
1543 size_t node_size = NODESIZE;
1546 MDB_dpage *ofp = NULL; /* overflow page */
1548 assert(mp->mp_upper >= mp->mp_lower);
1550 DPRINTF("add node [%.*s] to %s page %lu at index %i, key size %zu",
1551 key ? (int)key->mv_size : 0, key ? (char *)key->mv_data : NULL,
1552 IS_LEAF(mp) ? "leaf" : "branch",
1553 mp->mp_pgno, indx, key ? key->mv_size : 0);
1556 node_size += key->mv_size;
1560 node_size += data->mv_size;
1561 if (F_ISSET(flags, F_BIGDATA)) {
1562 /* Data already on overflow page. */
1563 node_size -= data->mv_size - sizeof(pgno_t);
1564 } else if (data->mv_size >= db->md_env->me_head.mh_psize / MDB_MINKEYS) {
1565 int ovpages = PAGEHDRSZ + data->mv_size + db->md_env->me_head.mh_psize - 1;
1566 ovpages /= db->md_env->me_head.mh_psize;
1567 /* Put data on overflow page. */
1568 DPRINTF("data size is %zu, put on overflow page",
1570 node_size -= data->mv_size - sizeof(pgno_t);
1571 if ((ofp = mdbenv_new_page(db->md_env, P_OVERFLOW, ovpages)) == NULL)
1573 DPRINTF("allocated overflow page %lu", ofp->p.mp_pgno);
1578 if (node_size + sizeof(indx_t) > SIZELEFT(mp)) {
1579 DPRINTF("not enough room in page %lu, got %lu ptrs",
1580 mp->mp_pgno, NUMKEYS(mp));
1581 DPRINTF("upper - lower = %u - %u = %u", mp->mp_upper, mp->mp_lower,
1582 mp->mp_upper - mp->mp_lower);
1583 DPRINTF("node size = %zu", node_size);
1587 /* Move higher pointers up one slot. */
1588 for (i = NUMKEYS(mp); i > indx; i--)
1589 mp->mp_ptrs[i] = mp->mp_ptrs[i - 1];
1591 /* Adjust free space offsets. */
1592 ofs = mp->mp_upper - node_size;
1593 assert(ofs >= mp->mp_lower + sizeof(indx_t));
1594 mp->mp_ptrs[indx] = ofs;
1596 mp->mp_lower += sizeof(indx_t);
1598 /* Write the node data. */
1599 node = NODEPTR(mp, indx);
1600 node->mn_ksize = (key == NULL) ? 0 : key->mv_size;
1601 node->mn_flags = flags;
1603 node->mn_dsize = data->mv_size;
1605 node->mn_pgno = pgno;
1608 bcopy(key->mv_data, NODEKEY(node), key->mv_size);
1613 if (F_ISSET(flags, F_BIGDATA))
1614 bcopy(data->mv_data, node->mn_data + key->mv_size,
1617 bcopy(data->mv_data, node->mn_data + key->mv_size,
1620 bcopy(&ofp->p.mp_pgno, node->mn_data + key->mv_size,
1622 bcopy(data->mv_data, METADATA(&ofp->p), data->mv_size);
1630 mdb_del_node(MDB_db *db, MDB_page *mp, indx_t indx)
1633 indx_t i, j, numkeys, ptr;
1637 DPRINTF("delete node %u on %s page %lu", indx,
1638 IS_LEAF(mp) ? "leaf" : "branch", mp->mp_pgno);
1639 assert(indx < NUMKEYS(mp));
1641 node = NODEPTR(mp, indx);
1642 sz = NODESIZE + node->mn_ksize;
1644 if (F_ISSET(node->mn_flags, F_BIGDATA))
1645 sz += sizeof(pgno_t);
1647 sz += NODEDSZ(node);
1650 ptr = mp->mp_ptrs[indx];
1651 numkeys = NUMKEYS(mp);
1652 for (i = j = 0; i < numkeys; i++) {
1654 mp->mp_ptrs[j] = mp->mp_ptrs[i];
1655 if (mp->mp_ptrs[i] < ptr)
1656 mp->mp_ptrs[j] += sz;
1661 base = (char *)mp + mp->mp_upper;
1662 bcopy(base, base + sz, ptr - mp->mp_upper);
1664 mp->mp_lower -= sizeof(indx_t);
1669 mdb_cursor_open(MDB_db *db, MDB_txn *txn, MDB_cursor **ret)
1673 if (db == NULL || ret == NULL)
1676 if (txn != NULL && db->md_env != txn->mt_env) {
1680 if ((cursor = calloc(1, sizeof(*cursor))) != NULL) {
1681 SLIST_INIT(&cursor->mc_stack);
1683 cursor->mc_txn = txn;
1692 mdb_cursor_close(MDB_cursor *cursor)
1694 if (cursor != NULL) {
1695 while (!CURSOR_EMPTY(cursor))
1696 cursor_pop_page(cursor);
1698 /* btree_close(cursor->bt); */
1704 mdb_update_key(MDB_db *db, MDB_page *mp, indx_t indx,
1707 indx_t ptr, i, numkeys;
1713 node = NODEPTR(mp, indx);
1714 ptr = mp->mp_ptrs[indx];
1715 DPRINTF("update key %u (ofs %u) [%.*s] to [%.*s] on page %lu",
1717 (int)node->mn_ksize, (char *)NODEKEY(node),
1718 (int)key->mv_size, (char *)key->mv_data,
1721 delta = key->mv_size - node->mn_ksize;
1723 if (delta > 0 && SIZELEFT(mp) < delta) {
1724 DPRINTF("OUCH! Not enough room, delta = %d", delta);
1728 numkeys = NUMKEYS(mp);
1729 for (i = 0; i < numkeys; i++) {
1730 if (mp->mp_ptrs[i] <= ptr)
1731 mp->mp_ptrs[i] -= delta;
1734 base = (char *)mp + mp->mp_upper;
1735 len = ptr - mp->mp_upper + NODESIZE;
1736 bcopy(base, base - delta, len);
1737 mp->mp_upper -= delta;
1739 node = NODEPTR(mp, indx);
1740 node->mn_ksize = key->mv_size;
1743 bcopy(key->mv_data, NODEKEY(node), key->mv_size);
1748 /* Move a node from src to dst.
1751 mdb_move_node(MDB_db *bt, MDB_pageparent *src, indx_t srcindx,
1752 MDB_pageparent *dst, indx_t dstindx)
1758 srcnode = NODEPTR(src->mp_page, srcindx);
1759 DPRINTF("moving %s node %u [%.*s] on page %lu to node %u on page %lu",
1760 IS_LEAF(src->mp_page) ? "leaf" : "branch",
1762 (int)srcnode->mn_ksize, (char *)NODEKEY(srcnode),
1763 src->mp_page->mp_pgno,
1764 dstindx, dst->mp_page->mp_pgno);
1766 /* Mark src and dst as dirty. */
1767 if ((rc = mdb_touch(bt->md_env->me_txn, src)) ||
1768 (rc = mdb_touch(bt->md_env->me_txn, dst)))
1771 /* Add the node to the destination page.
1773 key.mv_size = srcnode->mn_ksize;
1774 key.mv_data = NODEKEY(srcnode);
1775 data.mv_size = NODEDSZ(srcnode);
1776 data.mv_data = NODEDATA(srcnode);
1777 rc = mdb_add_node(bt, dst->mp_page, dstindx, &key, &data, NODEPGNO(srcnode),
1779 if (rc != MDB_SUCCESS)
1782 /* Delete the node from the source page.
1784 mdb_del_node(bt, src->mp_page, srcindx);
1786 /* Update the parent separators.
1788 if (srcindx == 0 && src->mp_pi != 0) {
1789 DPRINTF("update separator for source page %lu to [%.*s]",
1790 src->mp_page->mp_pgno, (int)key.mv_size, (char *)key.mv_data);
1791 if ((rc = mdb_update_key(bt, src->mp_parent, src->mp_pi,
1792 &key)) != MDB_SUCCESS)
1796 if (srcindx == 0 && IS_BRANCH(src->mp_page)) {
1798 nullkey.mv_size = 0;
1799 assert(mdb_update_key(bt, src->mp_page, 0, &nullkey) == MDB_SUCCESS);
1802 if (dstindx == 0 && dst->mp_pi != 0) {
1803 DPRINTF("update separator for destination page %lu to [%.*s]",
1804 dst->mp_page->mp_pgno, (int)key.mv_size, (char *)key.mv_data);
1805 if ((rc = mdb_update_key(bt, dst->mp_parent, dst->mp_pi,
1806 &key)) != MDB_SUCCESS)
1810 if (dstindx == 0 && IS_BRANCH(dst->mp_page)) {
1812 nullkey.mv_size = 0;
1813 assert(mdb_update_key(bt, dst->mp_page, 0, &nullkey) == MDB_SUCCESS);
1820 mdb_merge(MDB_db *bt, MDB_pageparent *src, MDB_pageparent *dst)
1829 DPRINTF("merging page %lu and %lu", src->mp_page->mp_pgno, dst->mp_page->mp_pgno);
1831 assert(src->mp_parent); /* can't merge root page */
1832 assert(dst->mp_parent);
1833 assert(bt->md_env->me_txn != NULL);
1835 /* Mark src and dst as dirty. */
1836 if ((rc = mdb_touch(bt->md_env->me_txn, src)) ||
1837 (rc = mdb_touch(bt->md_env->me_txn, dst)))
1840 /* Move all nodes from src to dst.
1842 for (i = 0; i < NUMKEYS(src->mp_page); i++) {
1843 srcnode = NODEPTR(src->mp_page, i);
1845 key.mv_size = srcnode->mn_ksize;
1846 key.mv_data = NODEKEY(srcnode);
1847 data.mv_size = NODEDSZ(srcnode);
1848 data.mv_data = NODEDATA(srcnode);
1849 rc = mdb_add_node(bt, dst->mp_page, NUMKEYS(dst->mp_page), &key,
1850 &data, NODEPGNO(srcnode), srcnode->mn_flags);
1851 if (rc != MDB_SUCCESS)
1855 DPRINTF("dst page %lu now has %lu keys (%.1f%% filled)",
1856 dst->mp_page->mp_pgno, NUMKEYS(dst->mp_page), (float)PAGEFILL(bt->md_env, dst->mp_page) / 10);
1858 /* Unlink the src page from parent.
1860 mdb_del_node(bt, src->mp_parent, src->mp_pi);
1861 if (src->mp_pi == 0) {
1863 if ((rc = mdb_update_key(bt, src->mp_parent, 0, &key)) != MDB_SUCCESS)
1867 if (IS_LEAF(src->mp_page))
1868 bt->md_env->me_meta.mm_stat.ms_leaf_pages--;
1870 bt->md_env->me_meta.mm_stat.ms_branch_pages--;
1872 mpp.mp_page = src->mp_parent;
1873 dh = (MDB_dhead *)src->mp_parent;
1875 mpp.mp_parent = dh->md_parent;
1876 mpp.mp_pi = dh->md_pi;
1878 return mdb_rebalance(bt, &mpp);
1881 #define FILL_THRESHOLD 250
1884 mdb_rebalance(MDB_db *db, MDB_pageparent *mpp)
1889 indx_t si = 0, di = 0;
1892 assert(db->md_env->me_txn != NULL);
1893 assert(mpp != NULL);
1895 DPRINTF("rebalancing %s page %lu (has %lu keys, %.1f%% full)",
1896 IS_LEAF(mpp->mp_page) ? "leaf" : "branch",
1897 mpp->mp_page->mp_pgno, NUMKEYS(mpp->mp_page), (float)PAGEFILL(db->md_env, mpp->mp_page) / 10);
1899 if (PAGEFILL(db->md_env, mpp->mp_page) >= FILL_THRESHOLD) {
1900 DPRINTF("no need to rebalance page %lu, above fill threshold",
1901 mpp->mp_page->mp_pgno);
1905 if (mpp->mp_parent == NULL) {
1906 if (NUMKEYS(mpp->mp_page) == 0) {
1907 DPRINTF("tree is completely empty");
1908 db->md_env->me_txn->mt_root = P_INVALID;
1909 db->md_env->me_meta.mm_stat.ms_depth--;
1910 db->md_env->me_meta.mm_stat.ms_leaf_pages--;
1911 } else if (IS_BRANCH(mpp->mp_page) && NUMKEYS(mpp->mp_page) == 1) {
1912 DPRINTF("collapsing root page!");
1913 db->md_env->me_txn->mt_root = NODEPGNO(NODEPTR(mpp->mp_page, 0));
1914 if ((root = mdbenv_get_page(db->md_env, db->md_env->me_txn->mt_root)) == NULL)
1916 db->md_env->me_meta.mm_stat.ms_depth--;
1917 db->md_env->me_meta.mm_stat.ms_branch_pages--;
1919 DPRINTF("root page doesn't need rebalancing");
1923 /* The parent (branch page) must have at least 2 pointers,
1924 * otherwise the tree is invalid.
1926 assert(NUMKEYS(mpp->mp_parent) > 1);
1928 /* Leaf page fill factor is below the threshold.
1929 * Try to move keys from left or right neighbor, or
1930 * merge with a neighbor page.
1935 if (mpp->mp_pi == 0) {
1936 /* We're the leftmost leaf in our parent.
1938 DPRINTF("reading right neighbor");
1939 node = NODEPTR(mpp->mp_parent, mpp->mp_pi + 1);
1940 if ((npp.mp_page = mdbenv_get_page(db->md_env, NODEPGNO(node))) == NULL)
1942 npp.mp_pi = mpp->mp_pi + 1;
1944 di = NUMKEYS(mpp->mp_page);
1946 /* There is at least one neighbor to the left.
1948 DPRINTF("reading left neighbor");
1949 node = NODEPTR(mpp->mp_parent, mpp->mp_pi - 1);
1950 if ((npp.mp_page = mdbenv_get_page(db->md_env, NODEPGNO(node))) == NULL)
1952 npp.mp_pi = mpp->mp_pi - 1;
1953 si = NUMKEYS(npp.mp_page) - 1;
1956 npp.mp_parent = mpp->mp_parent;
1958 DPRINTF("found neighbor page %lu (%lu keys, %.1f%% full)",
1959 npp.mp_page->mp_pgno, NUMKEYS(npp.mp_page), (float)PAGEFILL(db->md_env, npp.mp_page) / 10);
1961 /* If the neighbor page is above threshold and has at least two
1962 * keys, move one key from it.
1964 * Otherwise we should try to merge them.
1966 if (PAGEFILL(db->md_env, npp.mp_page) >= FILL_THRESHOLD && NUMKEYS(npp.mp_page) >= 2)
1967 return mdb_move_node(db, &npp, si, mpp, di);
1968 else { /* FIXME: if (has_enough_room()) */
1969 if (mpp->mp_pi == 0)
1970 return mdb_merge(db, &npp, mpp);
1972 return mdb_merge(db, mpp, &npp);
1977 mdb_del(MDB_db *bt, MDB_txn *txn,
1978 MDB_val *key, MDB_val *data)
1985 DPRINTF("========> delete key %.*s", (int)key->mv_size, (char *)key->mv_data);
1987 assert(key != NULL);
1989 if (bt == NULL || txn == NULL)
1992 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
1996 if (bt->md_env->me_txn != txn) {
2000 if (key->mv_size == 0 || key->mv_size > MAXKEYSIZE) {
2004 if ((rc = mdb_search_page(bt, txn, key, NULL, 1, &mpp)) != MDB_SUCCESS)
2007 leaf = mdb_search_node(bt, mpp.mp_page, key, &exact, &ki);
2008 if (leaf == NULL || !exact) {
2012 if (data && (rc = mdb_read_data(bt, NULL, leaf, data)) != MDB_SUCCESS)
2015 mdb_del_node(bt, mpp.mp_page, ki);
2016 bt->md_env->me_meta.mm_stat.ms_entries--;
2017 rc = mdb_rebalance(bt, &mpp);
2018 if (rc != MDB_SUCCESS)
2019 txn->mt_flags |= MDB_TXN_ERROR;
2024 /* Split page <*mpp>, and insert <key,(data|newpgno)> in either left or
2025 * right sibling, at index <*newindxp> (as if unsplit). Updates *mpp and
2026 * *newindxp with the actual values after split, ie if *mpp and *newindxp
2027 * refer to a node in the new right sibling page.
2030 mdb_split(MDB_db *bt, MDB_page **mpp, unsigned int *newindxp,
2031 MDB_val *newkey, MDB_val *newdata, pgno_t newpgno)
2034 int rc = MDB_SUCCESS, ins_new = 0;
2037 unsigned int i, j, split_indx;
2039 MDB_page *pright, *p;
2040 MDB_val sepkey, rkey, rdata;
2042 MDB_dpage *mdp, *rdp, *pdp;
2046 assert(bt->md_env != NULL);
2048 dh = ((MDB_dhead *)*mpp) - 1;
2049 mdp = (MDB_dpage *)dh;
2050 newindx = *newindxp;
2052 DPRINTF("-----> splitting %s page %lu and adding [%.*s] at index %i",
2053 IS_LEAF(&mdp->p) ? "leaf" : "branch", mdp->p.mp_pgno,
2054 (int)newkey->mv_size, (char *)newkey->mv_data, *newindxp);
2056 if (mdp->h.md_parent == NULL) {
2057 if ((pdp = mdbenv_new_page(bt->md_env, P_BRANCH, 1)) == NULL)
2060 mdp->h.md_parent = &pdp->p;
2061 bt->md_env->me_txn->mt_root = pdp->p.mp_pgno;
2062 DPRINTF("root split! new root = %lu", pdp->p.mp_pgno);
2063 bt->md_env->me_meta.mm_stat.ms_depth++;
2065 /* Add left (implicit) pointer. */
2066 if (mdb_add_node(bt, &pdp->p, 0, NULL, NULL,
2067 mdp->p.mp_pgno, 0) != MDB_SUCCESS)
2070 DPRINTF("parent branch page is %lu", mdp->h.md_parent->mp_pgno);
2073 /* Create a right sibling. */
2074 if ((rdp = mdbenv_new_page(bt->md_env, mdp->p.mp_flags, 1)) == NULL)
2076 rdp->h.md_parent = mdp->h.md_parent;
2077 rdp->h.md_pi = mdp->h.md_pi + 1;
2078 DPRINTF("new right sibling: page %lu", rdp->p.mp_pgno);
2080 /* Move half of the keys to the right sibling. */
2081 if ((copy = malloc(bt->md_env->me_head.mh_psize)) == NULL)
2083 bcopy(&mdp->p, copy, bt->md_env->me_head.mh_psize);
2084 bzero(&mdp->p.mp_ptrs, bt->md_env->me_head.mh_psize - PAGEHDRSZ);
2085 mdp->p.mp_lower = PAGEHDRSZ;
2086 mdp->p.mp_upper = bt->md_env->me_head.mh_psize;
2088 split_indx = NUMKEYS(copy) / 2 + 1;
2090 /* First find the separating key between the split pages.
2092 bzero(&sepkey, sizeof(sepkey));
2093 if (newindx == split_indx) {
2094 sepkey.mv_size = newkey->mv_size;
2095 sepkey.mv_data = newkey->mv_data;
2097 node = NODEPTR(copy, split_indx);
2098 sepkey.mv_size = node->mn_ksize;
2099 sepkey.mv_data = NODEKEY(node);
2102 DPRINTF("separator is [%.*s]", (int)sepkey.mv_size, (char *)sepkey.mv_data);
2104 /* Copy separator key to the parent.
2106 if (SIZELEFT(rdp->h.md_parent) < mdb_branch_size(bt, &sepkey)) {
2107 rc = mdb_split(bt, &rdp->h.md_parent, &rdp->h.md_pi,
2108 &sepkey, NULL, rdp->p.mp_pgno);
2110 /* Right page might now have changed parent.
2111 * Check if left page also changed parent.
2113 if (rdp->h.md_parent != mdp->h.md_parent &&
2114 mdp->h.md_pi >= NUMKEYS(mdp->h.md_parent)) {
2115 mdp->h.md_parent = rdp->h.md_parent;
2116 mdp->h.md_pi = rdp->h.md_pi - 1;
2119 rc = mdb_add_node(bt, rdp->h.md_parent, rdp->h.md_pi,
2120 &sepkey, NULL, rdp->p.mp_pgno, 0);
2122 if (rc != MDB_SUCCESS) {
2127 for (i = j = 0; i <= NUMKEYS(copy); j++) {
2128 if (i < split_indx) {
2129 /* Re-insert in left sibling. */
2132 /* Insert in right sibling. */
2133 if (i == split_indx)
2134 /* Reset insert index for right sibling. */
2135 j = (i == newindx && ins_new);
2139 if (i == newindx && !ins_new) {
2140 /* Insert the original entry that caused the split. */
2141 rkey.mv_data = newkey->mv_data;
2142 rkey.mv_size = newkey->mv_size;
2143 if (IS_LEAF(&mdp->p)) {
2144 rdata.mv_data = newdata->mv_data;
2145 rdata.mv_size = newdata->mv_size;
2152 /* Update page and index for the new key. */
2155 } else if (i == NUMKEYS(copy)) {
2158 node = NODEPTR(copy, i);
2159 rkey.mv_data = NODEKEY(node);
2160 rkey.mv_size = node->mn_ksize;
2161 if (IS_LEAF(&mdp->p)) {
2162 rdata.mv_data = NODEDATA(node);
2163 rdata.mv_size = node->mn_dsize;
2165 pgno = node->mn_pgno;
2166 flags = node->mn_flags;
2171 if (!IS_LEAF(&mdp->p) && j == 0) {
2172 /* First branch index doesn't need key data. */
2176 rc = mdb_add_node(bt, &pdp->p, j, &rkey, &rdata, pgno,flags);
2184 mdb_put(MDB_db *bt, MDB_txn *txn,
2185 MDB_val *key, MDB_val *data, unsigned int flags)
2187 int rc = MDB_SUCCESS, exact;
2192 assert(key != NULL);
2193 assert(data != NULL);
2195 if (bt == NULL || txn == NULL)
2198 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
2202 if (bt->md_env->me_txn != txn) {
2206 if (key->mv_size == 0 || key->mv_size > MAXKEYSIZE) {
2210 DPRINTF("==> put key %.*s, size %zu, data size %zu",
2211 (int)key->mv_size, (char *)key->mv_data, key->mv_size, data->mv_size);
2213 rc = mdb_search_page(bt, txn, key, NULL, 1, &mpp);
2214 if (rc == MDB_SUCCESS) {
2215 leaf = mdb_search_node(bt, mpp.mp_page, key, &exact, &ki);
2216 if (leaf && exact) {
2217 if (F_ISSET(flags, MDB_NOOVERWRITE)) {
2218 DPRINTF("duplicate key %.*s",
2219 (int)key->mv_size, (char *)key->mv_data);
2222 mdb_del_node(bt, mpp.mp_page, ki);
2224 if (leaf == NULL) { /* append if not found */
2225 ki = NUMKEYS(mpp.mp_page);
2226 DPRINTF("appending key at index %i", ki);
2228 } else if (rc == ENOENT) {
2230 /* new file, just write a root leaf page */
2231 DPRINTF("allocating new root leaf page");
2232 if ((dp = mdbenv_new_page(bt->md_env, P_LEAF, 1)) == NULL) {
2235 mpp.mp_page = &dp->p;
2236 txn->mt_root = mpp.mp_page->mp_pgno;
2237 bt->md_env->me_meta.mm_stat.ms_depth++;
2243 assert(IS_LEAF(mpp.mp_page));
2244 DPRINTF("there are %lu keys, should insert new key at index %i",
2245 NUMKEYS(mpp.mp_page), ki);
2247 if (SIZELEFT(mpp.mp_page) < mdb_leaf_size(bt, key, data)) {
2248 rc = mdb_split(bt, &mpp.mp_page, &ki, key, data, P_INVALID);
2250 /* There is room already in this leaf page. */
2251 rc = mdb_add_node(bt, mpp.mp_page, ki, key, data, 0, 0);
2254 if (rc != MDB_SUCCESS)
2255 txn->mt_flags |= MDB_TXN_ERROR;
2257 bt->md_env->me_meta.mm_stat.ms_entries++;
2264 mdbenv_compact_tree(MDB_env *env, pgno_t pgno, MDB_env *envc)
2268 pgno_t *pnext, next;
2273 /* Get the page and make a copy of it.
2275 if ((mp = mdbenv_get_page(env, pgno)) == NULL)
2277 if ((p = malloc(env->me_head.mh_psize)) == NULL)
2279 bcopy(mp, p, env->me_head.mh_psize);
2281 /* Go through all nodes in the (copied) page and update the
2284 if (F_ISSET(p->mp_flags, P_BRANCH)) {
2285 for (i = 0; i < NUMKEYS(p); i++) {
2286 node = NODEPTR(p, i);
2287 node->mn_pgno = mdbenv_compact_tree(env, node->mn_pgno, envc);
2288 if (node->mn_pgno == P_INVALID) {
2293 } else if (F_ISSET(p->mp_flags, P_LEAF)) {
2294 for (i = 0; i < NUMKEYS(p); i++) {
2295 node = NODEPTR(p, i);
2296 if (F_ISSET(node->mn_flags, F_BIGDATA)) {
2297 bcopy(NODEDATA(node), &next, sizeof(next));
2298 next = mdbenv_compact_tree(env, next, envc);
2299 if (next == P_INVALID) {
2303 bcopy(&next, NODEDATA(node), sizeof(next));
2306 } else if (F_ISSET(p->mp_flags, P_OVERFLOW)) {
2307 ; /* handled below */
2311 pgno = p->mp_pgno = envc->me_txn->mt_next_pgno++;
2312 rc = write(envc->me_fd, p, env->me_head.mh_psize);
2314 if (rc != (ssize_t)env->me_head.mh_psize)
2317 if (F_ISSET(mp->mp_flags, P_OVERFLOW)) {
2318 if (mp->mp_pages > 1) {
2319 size_t len = (mp->mp_pages-1) * env->me_head.mh_psize;
2320 envc->me_txn->mt_next_pgno += mp->mp_pages-1;
2321 rc = write(envc->me_fd, mp+1, len);
2330 mdbenv_compact(MDB_env *env)
2332 char *compact_path = NULL;
2334 MDB_txn *txn, *txnc = NULL;
2338 assert(env != NULL);
2340 if (env->me_path == NULL) {
2344 DPRINTF("compacting mdbenv %p with path %s", env, env->me_path);
2346 rc = mdb_txn_begin(env, 0, &txn);
2349 asprintf(&compact_path, "%s.compact.XXXXXX", env->me_path);
2350 fd = mkstemp(compact_path);
2358 rc = mdbenv_create(&envc, env->me_mapsize);
2359 if (rc) goto failed;
2362 rc = mdbenv_open2(envc, env->me_flags);
2363 if (rc) goto failed;
2365 bcopy(&env->me_meta, &envc->me_meta, sizeof(env->me_meta));
2366 envc->me_meta.mm_stat.ms_revisions = 0;
2368 rc = mdb_txn_begin(envc, 0, &txnc);
2369 if (rc) goto failed;
2371 if (env->me_meta.mm_root != P_INVALID) {
2372 root = mdbenv_compact_tree(env, env->me_meta.mm_root, envc);
2373 if (root == P_INVALID)
2375 if ((rc = mdbenv_write_meta(envc, root, 0)) != MDB_SUCCESS)
2381 DPRINTF("renaming %s to %s", compact_path, env->me_path);
2382 if (rename(compact_path, env->me_path) != 0) {
2387 /* Write a "tombstone" meta page so other processes can pick up
2388 * the change and re-open the file.
2390 if (mdbenv_write_meta(env, P_INVALID, MDB_TOMBSTONE) != MDB_SUCCESS)
2394 mdb_txn_abort(txnc);
2401 mdb_txn_abort(txnc);
2402 unlink(compact_path);
2409 mdbenv_get_flags(MDB_env *env, unsigned int *arg)
2414 *arg = (env->me_flags & ~MDB_FIXPADDING);
2419 mdbenv_get_path(MDB_env *env, const char **arg)
2424 *arg = env->me_path;
2429 mdbenv_stat(MDB_env *env, MDB_stat **arg)
2431 if (env == NULL || arg == NULL)
2434 *arg = &env->me_meta.mm_stat;
2439 int mdb_open(MDB_env *env, MDB_txn *txn, const char *name, unsigned int flags, MDB_db **db)
2442 *db = (MDB_db *)&env->me_db;
2448 int mdb_stat(MDB_db *db, MDB_stat **arg)
2450 if (db == NULL || arg == NULL)
2453 *arg = &db->md_stat;
2458 void mdb_close(MDB_db *db)