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 env->me_meta.mm_stat.ms_psize = env->me_head.mh_psize;
786 return MDB_SUCCESS; /* new file */
789 next_pgno = size / env->me_head.mh_psize;
790 if (next_pgno == 0) {
791 DPRINTF("corrupt file");
796 meta_pgno = next_pgno - 1;
798 if (size % env->me_head.mh_psize != 0) {
799 DPRINTF("filesize not a multiple of the page size!");
800 env->me_flags |= MDB_FIXPADDING;
807 if (size == env->me_size) {
808 DPRINTF("size unchanged, keeping current meta page");
809 if (F_ISSET(env->me_meta.mm_flags, MDB_TOMBSTONE)) {
810 DPRINTF("file is dead");
819 while (meta_pgno > 0) {
820 if ((mp = mdbenv_get_page(env, meta_pgno)) == NULL)
822 if (!mdb_check_meta_page(mp)) {
824 DPRINTF("flags = 0x%x", meta->mm_flags);
825 if (F_ISSET(meta->mm_flags, MDB_TOMBSTONE)) {
826 DPRINTF("file is dead");
830 /* Make copy of last meta page. */
831 bcopy(meta, &env->me_meta, sizeof(env->me_meta));
835 --meta_pgno; /* scan backwards to first valid meta page */
846 mdbenv_create(MDB_env **env, size_t size)
850 e = calloc(1, sizeof(*e));
851 if (!e) return ENOMEM;
853 e->me_head.mh_magic = MDB_MAGIC;
854 e->me_head.mh_version = MDB_VERSION;
855 e->me_mapsize = e->me_head.mh_mapsize = size;
862 mdbenv_open2(MDB_env *env, unsigned int flags)
866 i = fcntl(env->me_fd, F_GETFL, 0);
867 if (fcntl(env->me_fd, F_SETFL, i | O_APPEND) == -1)
870 env->me_flags = flags;
871 env->me_flags &= ~MDB_FIXPADDING;
872 env->me_meta.mm_root = P_INVALID;
874 if ((i = mdbenv_read_header(env)) != 0) {
877 DPRINTF("new mdbenv");
881 if (!env->me_mapsize)
882 env->me_mapsize = env->me_head.mh_mapsize;
885 if (env->me_head.mh_address && (flags & MDB_FIXEDMAP))
887 env->me_map = mmap(env->me_head.mh_address, env->me_mapsize, PROT_READ, i,
889 if (env->me_map == MAP_FAILED)
893 env->me_head.mh_mapsize = env->me_mapsize;
894 if (flags & MDB_FIXEDMAP)
895 env->me_head.mh_address = env->me_map;
896 i = mdbenv_write_header(env);
897 if (i != MDB_SUCCESS) {
898 munmap(env->me_map, env->me_mapsize);
903 if ((i = mdbenv_read_meta(env, NULL)) != 0)
906 DPRINTF("opened database version %u, pagesize %u",
907 env->me_head.mh_version, env->me_head.mh_psize);
908 DPRINTF("timestamp: %s", ctime(&env->me_meta.mm_created_at));
909 DPRINTF("depth: %u", env->me_meta.mm_stat.ms_depth);
910 DPRINTF("entries: %lu", env->me_meta.mm_stat.ms_entries);
911 DPRINTF("revisions: %lu", env->me_meta.mm_stat.ms_revisions);
912 DPRINTF("branch pages: %lu", env->me_meta.mm_stat.ms_branch_pages);
913 DPRINTF("leaf pages: %lu", env->me_meta.mm_stat.ms_leaf_pages);
914 DPRINTF("overflow pages: %lu", env->me_meta.mm_stat.ms_overflow_pages);
915 DPRINTF("root: %lu", env->me_meta.mm_root);
916 DPRINTF("previous me_meta.me_page: %lu", env->me_meta.mm_prev_meta);
922 mdbenv_open(MDB_env *env, const char *path, unsigned int flags, mode_t mode)
926 if (F_ISSET(flags, MDB_RDONLY))
929 oflags = O_RDWR | O_CREAT | O_APPEND;
931 if ((env->me_fd = open(path, oflags, mode)) == -1)
934 if ((rc = mdbenv_open2(env, flags)) != MDB_SUCCESS) {
938 env->me_path = strdup(path);
939 DPRINTF("opened dbenv %p", env);
946 mdbenv_close(MDB_env *env)
954 munmap(env->me_map, env->me_mapsize);
959 /* Search for key within a leaf page, using binary search.
960 * Returns the smallest entry larger or equal to the key.
961 * If exactp is non-null, stores whether the found entry was an exact match
962 * in *exactp (1 or 0).
963 * If kip is non-null, stores the index of the found entry in *kip.
964 * If no entry larger or equal to the key is found, returns NULL.
967 mdb_search_node(MDB_db *bt, MDB_page *mp, MDB_val *key,
968 int *exactp, unsigned int *kip)
976 DPRINTF("searching %lu keys in %s page %lu",
978 IS_LEAF(mp) ? "leaf" : "branch",
981 assert(NUMKEYS(mp) > 0);
983 bzero(&nodekey, sizeof(nodekey));
985 low = IS_LEAF(mp) ? 0 : 1;
986 high = NUMKEYS(mp) - 1;
987 while (low <= high) {
988 i = (low + high) >> 1;
989 node = NODEPTR(mp, i);
991 nodekey.mv_size = node->mn_ksize;
992 nodekey.mv_data = NODEKEY(node);
995 rc = bt->md_cmp(key, &nodekey);
997 rc = _mdb_cmp(bt, key, &nodekey);
1000 DPRINTF("found leaf index %u [%.*s], rc = %i",
1001 i, (int)nodekey.mv_size, (char *)nodekey.mv_data, rc);
1003 DPRINTF("found branch index %u [%.*s -> %lu], rc = %i",
1004 i, (int)node->mn_ksize, (char *)NODEKEY(node),
1015 if (rc > 0) { /* Found entry is less than the key. */
1016 i++; /* Skip to get the smallest entry larger than key. */
1017 if (i >= NUMKEYS(mp))
1018 /* There is no entry larger or equal to the key. */
1022 *exactp = (rc == 0);
1023 if (kip) /* Store the key index if requested. */
1026 return NODEPTR(mp, i);
1030 cursor_pop_page(MDB_cursor *cursor)
1034 top = CURSOR_TOP(cursor);
1037 DPRINTF("popped page %lu off cursor %p", top->mp_page->mp_pgno, cursor);
1043 cursor_push_page(MDB_cursor *cursor, MDB_page *mp)
1047 DPRINTF("pushing page %lu on cursor %p", mp->mp_pgno, cursor);
1049 if ((ppage = calloc(1, sizeof(*ppage))) == NULL)
1051 ppage->mp_page = mp;
1052 CURSOR_PUSH(cursor, ppage);
1057 mdbenv_get_page(MDB_env *env, pgno_t pgno)
1060 MDB_txn *txn = env->me_txn;
1062 if (txn && pgno >= txn->mt_first_pgno &&
1063 !SIMPLEQ_EMPTY(txn->mt_dirty_queue)) {
1065 SIMPLEQ_FOREACH(dp, txn->mt_dirty_queue, h.md_next) {
1066 if (dp->p.mp_pgno == pgno) {
1072 p = (MDB_page *)(env->me_map + env->me_head.mh_psize * pgno);
1078 mdb_search_page_root(MDB_db *bt, MDB_val *key,
1079 MDB_cursor *cursor, int modify, MDB_pageparent *mpp)
1081 MDB_page *mp = mpp->mp_page;
1084 if (cursor && cursor_push_page(cursor, mp) == NULL)
1087 while (IS_BRANCH(mp)) {
1091 DPRINTF("branch page %lu has %lu keys", mp->mp_pgno, NUMKEYS(mp));
1092 assert(NUMKEYS(mp) > 1);
1093 DPRINTF("found index 0 to page %lu", NODEPGNO(NODEPTR(mp, 0)));
1095 if (key == NULL) /* Initialize cursor to first page. */
1099 node = mdb_search_node(bt, mp, key, &exact, &i);
1101 i = NUMKEYS(mp) - 1;
1109 DPRINTF("following index %u for key %.*s",
1110 i, (int)key->mv_size, (char *)key->mv_data);
1111 assert(i >= 0 && i < NUMKEYS(mp));
1112 node = NODEPTR(mp, i);
1115 CURSOR_TOP(cursor)->mp_ki = i;
1117 mpp->mp_parent = mp;
1118 if ((mp = mdbenv_get_page(bt->md_env, NODEPGNO(node))) == NULL)
1123 if (cursor && cursor_push_page(cursor, mp) == NULL)
1126 if (modify && (rc = mdb_touch(bt->md_env->me_txn, mpp)))
1132 DPRINTF("internal error, index points to a %02X page!?",
1137 DPRINTF("found leaf page %lu for key %.*s", mp->mp_pgno,
1138 key ? (int)key->mv_size : 0, key ? (char *)key->mv_data : NULL);
1143 /* Search for the page a given key should be in.
1144 * Stores a pointer to the found page in *mpp.
1145 * If key is NULL, search for the lowest page (used by mdb_cursor_first).
1146 * If cursor is non-null, pushes parent pages on the cursor stack.
1147 * If modify is true, visited pages are updated with new page numbers.
1150 mdb_search_page(MDB_db *db, MDB_txn *txn, MDB_val *key,
1151 MDB_cursor *cursor, int modify, MDB_pageparent *mpp)
1156 /* Can't modify pages outside a transaction. */
1157 if (txn == NULL && modify) {
1161 /* Choose which root page to start with. If a transaction is given
1162 * use the root page from the transaction, otherwise read the last
1163 * committed root page.
1166 if ((rc = mdbenv_read_meta(db->md_env, NULL)) != MDB_SUCCESS)
1168 root = db->md_env->me_meta.mm_root;
1169 } else if (F_ISSET(txn->mt_flags, MDB_TXN_ERROR)) {
1170 DPRINTF("transaction has failed, must abort");
1173 root = txn->mt_root;
1175 if (root == P_INVALID) { /* Tree is empty. */
1176 DPRINTF("tree is empty");
1180 if ((mpp->mp_page = mdbenv_get_page(db->md_env, root)) == NULL)
1183 DPRINTF("root page has flags 0x%X", mpp->mp_page->mp_flags);
1185 if (modify && !F_ISSET(mpp->mp_page->mp_flags, P_DIRTY)) {
1186 mpp->mp_parent = NULL;
1188 if ((rc = mdb_touch(txn, mpp)))
1190 txn->mt_root = mpp->mp_page->mp_pgno;
1193 return mdb_search_page_root(db, key, cursor, modify, mpp);
1197 mdb_read_data(MDB_db *db, MDB_page *mp, MDB_node *leaf,
1200 MDB_page *omp; /* overflow mpage */
1206 bzero(data, sizeof(*data));
1207 max = db->md_env->me_head.mh_psize - PAGEHDRSZ;
1209 if (!F_ISSET(leaf->mn_flags, F_BIGDATA)) {
1210 data->mv_size = leaf->mn_dsize;
1211 data->mv_data = NODEDATA(leaf);
1215 /* Read overflow data.
1217 data->mv_size = leaf->mn_dsize;
1218 bcopy(NODEDATA(leaf), &pgno, sizeof(pgno));
1219 if ((omp = mdbenv_get_page(db->md_env, pgno)) == NULL) {
1220 DPRINTF("read overflow page %lu failed", pgno);
1223 data->mv_data = omp;
1229 mdb_get(MDB_db *db, MDB_txn *txn,
1230 MDB_val *key, MDB_val *data)
1238 DPRINTF("===> get key [%.*s]", (int)key->mv_size, (char *)key->mv_data);
1243 if (txn != NULL && db->md_env != txn->mt_env) {
1247 if (key->mv_size == 0 || key->mv_size > MAXKEYSIZE) {
1251 if ((rc = mdb_search_page(db, txn, key, NULL, 0, &mpp)) != MDB_SUCCESS)
1254 leaf = mdb_search_node(db, mpp.mp_page, key, &exact, NULL);
1256 rc = mdb_read_data(db, mpp.mp_page, leaf, data);
1265 mdb_sibling(MDB_cursor *cursor, int move_right)
1269 MDB_ppage *parent, *top;
1272 top = CURSOR_TOP(cursor);
1273 if ((parent = SLIST_NEXT(top, mp_entry)) == NULL) {
1274 return ENOENT; /* root has no siblings */
1277 DPRINTF("parent page is page %lu, index %u",
1278 parent->mp_page->mp_pgno, parent->mp_ki);
1280 cursor_pop_page(cursor);
1281 if (move_right ? (parent->mp_ki + 1 >= NUMKEYS(parent->mp_page))
1282 : (parent->mp_ki == 0)) {
1283 DPRINTF("no more keys left, moving to %s sibling",
1284 move_right ? "right" : "left");
1285 if ((rc = mdb_sibling(cursor, move_right)) != MDB_SUCCESS)
1287 parent = CURSOR_TOP(cursor);
1293 DPRINTF("just moving to %s index key %u",
1294 move_right ? "right" : "left", parent->mp_ki);
1296 assert(IS_BRANCH(parent->mp_page));
1299 indx = NODEPTR(parent->mp_page, parent->mp_ki);
1300 if ((mp = mdbenv_get_page(cursor->bt, indx->n_pgno)) == NULL)
1302 mp->parent = parent->mp_page;
1303 mp->parent_index = parent->mp_ki;
1306 cursor_push_page(cursor, mp);
1312 mdb_set_key(MDB_db *bt, MDB_page *mp, MDB_node *node,
1318 key->mv_size = node->mn_ksize;
1319 key->mv_data = NODEKEY(node);
1325 mdb_cursor_next(MDB_cursor *cursor, MDB_val *key, MDB_val *data)
1331 if (cursor->mc_eof) {
1335 assert(cursor->mc_initialized);
1337 top = CURSOR_TOP(cursor);
1340 DPRINTF("cursor_next: top page is %lu in cursor %p", mp->mp_pgno, cursor);
1342 if (top->mp_ki + 1 >= NUMKEYS(mp)) {
1343 DPRINTF("=====> move to next sibling page");
1344 if (mdb_sibling(cursor, 1) != MDB_SUCCESS) {
1348 top = CURSOR_TOP(cursor);
1350 DPRINTF("next page is %lu, key index %u", mp->mp_pgno, top->mp_ki);
1354 DPRINTF("==> cursor points to page %lu with %lu keys, key index %u",
1355 mp->mp_pgno, NUMKEYS(mp), top->mp_ki);
1357 assert(IS_LEAF(mp));
1358 leaf = NODEPTR(mp, top->mp_ki);
1360 if (data && mdb_read_data(cursor->mc_db, mp, leaf, data) != MDB_SUCCESS)
1363 return mdb_set_key(cursor->mc_db, mp, leaf, key);
1367 mdb_cursor_set(MDB_cursor *cursor, MDB_val *key, MDB_val *data,
1377 assert(key->mv_size > 0);
1379 rc = mdb_search_page(cursor->mc_db, cursor->mc_txn, key, cursor, 0, &mpp);
1380 if (rc != MDB_SUCCESS)
1382 assert(IS_LEAF(mpp.mp_page));
1384 top = CURSOR_TOP(cursor);
1385 leaf = mdb_search_node(cursor->mc_db, mpp.mp_page, key, exactp, &top->mp_ki);
1386 if (exactp != NULL && !*exactp) {
1387 /* MDB_CURSOR_EXACT specified and not an exact match. */
1392 DPRINTF("===> inexact leaf not found, goto sibling");
1393 if ((rc = mdb_sibling(cursor, 1)) != MDB_SUCCESS)
1394 return rc; /* no entries matched */
1395 top = CURSOR_TOP(cursor);
1397 mpp.mp_page = top->mp_page;
1398 assert(IS_LEAF(mpp.mp_page));
1399 leaf = NODEPTR(mpp.mp_page, 0);
1402 cursor->mc_initialized = 1;
1405 if (data && (rc = mdb_read_data(cursor->mc_db, mpp.mp_page, leaf, data)) != MDB_SUCCESS)
1408 rc = mdb_set_key(cursor->mc_db, mpp.mp_page, leaf, key);
1409 if (rc == MDB_SUCCESS) {
1410 DPRINTF("==> cursor placed on key %.*s",
1411 (int)key->mv_size, (char *)key->mv_data);
1419 mdb_cursor_first(MDB_cursor *cursor, MDB_val *key, MDB_val *data)
1425 rc = mdb_search_page(cursor->mc_db, cursor->mc_txn, NULL, cursor, 0, &mpp);
1426 if (rc != MDB_SUCCESS)
1428 assert(IS_LEAF(mpp.mp_page));
1430 leaf = NODEPTR(mpp.mp_page, 0);
1431 cursor->mc_initialized = 1;
1434 if (data && (rc = mdb_read_data(cursor->mc_db, mpp.mp_page, leaf, data)) != MDB_SUCCESS)
1437 return mdb_set_key(cursor->mc_db, mpp.mp_page, leaf, key);
1441 mdb_cursor_get(MDB_cursor *cursor, MDB_val *key, MDB_val *data,
1451 case MDB_CURSOR_EXACT:
1452 while (CURSOR_TOP(cursor) != NULL)
1453 cursor_pop_page(cursor);
1454 if (key == NULL || key->mv_size == 0 || key->mv_size > MAXKEYSIZE) {
1456 } else if (op == MDB_CURSOR_EXACT)
1457 rc = mdb_cursor_set(cursor, key, data, &exact);
1459 rc = mdb_cursor_set(cursor, key, data, NULL);
1462 if (!cursor->mc_initialized)
1463 rc = mdb_cursor_first(cursor, key, data);
1465 rc = mdb_cursor_next(cursor, key, data);
1468 while (CURSOR_TOP(cursor) != NULL)
1469 cursor_pop_page(cursor);
1470 rc = mdb_cursor_first(cursor, key, data);
1473 DPRINTF("unhandled/unimplemented cursor operation %u", op);
1481 /* Allocate a page and initialize it
1484 mdbenv_new_page(MDB_env *env, uint32_t flags, int num)
1488 assert(env != NULL);
1489 assert(env->me_txn != NULL);
1491 DPRINTF("allocating new mpage %lu, page size %u",
1492 env->me_txn->mt_next_pgno, env->me_head.mh_psize);
1493 if ((dp = mdb_newpage(env->me_txn, NULL, 0, num)) == NULL)
1495 dp->p.mp_flags = flags;
1496 dp->p.mp_lower = PAGEHDRSZ;
1497 dp->p.mp_upper = env->me_head.mh_psize;
1499 if (IS_BRANCH(&dp->p))
1500 env->me_meta.mm_stat.ms_branch_pages++;
1501 else if (IS_LEAF(&dp->p))
1502 env->me_meta.mm_stat.ms_leaf_pages++;
1503 else if (IS_OVERFLOW(&dp->p)) {
1504 env->me_meta.mm_stat.ms_overflow_pages += num;
1505 dp->p.mp_pages = num;
1512 mdb_leaf_size(MDB_db *db, MDB_val *key, MDB_val *data)
1516 sz = LEAFSIZE(key, data);
1517 if (data->mv_size >= db->md_env->me_head.mh_psize / MDB_MINKEYS) {
1518 /* put on overflow page */
1519 sz -= data->mv_size - sizeof(pgno_t);
1522 return sz + sizeof(indx_t);
1526 mdb_branch_size(MDB_db *db, MDB_val *key)
1531 if (sz >= db->md_env->me_head.mh_psize / MDB_MINKEYS) {
1532 /* put on overflow page */
1533 /* not implemented */
1534 /* sz -= key->size - sizeof(pgno_t); */
1537 return sz + sizeof(indx_t);
1541 mdb_add_node(MDB_db *db, MDB_page *mp, indx_t indx,
1542 MDB_val *key, MDB_val *data, pgno_t pgno, uint8_t flags)
1545 size_t node_size = NODESIZE;
1548 MDB_dpage *ofp = NULL; /* overflow page */
1550 assert(mp->mp_upper >= mp->mp_lower);
1552 DPRINTF("add node [%.*s] to %s page %lu at index %i, key size %zu",
1553 key ? (int)key->mv_size : 0, key ? (char *)key->mv_data : NULL,
1554 IS_LEAF(mp) ? "leaf" : "branch",
1555 mp->mp_pgno, indx, key ? key->mv_size : 0);
1558 node_size += key->mv_size;
1562 node_size += data->mv_size;
1563 if (F_ISSET(flags, F_BIGDATA)) {
1564 /* Data already on overflow page. */
1565 node_size -= data->mv_size - sizeof(pgno_t);
1566 } else if (data->mv_size >= db->md_env->me_head.mh_psize / MDB_MINKEYS) {
1567 int ovpages = PAGEHDRSZ + data->mv_size + db->md_env->me_head.mh_psize - 1;
1568 ovpages /= db->md_env->me_head.mh_psize;
1569 /* Put data on overflow page. */
1570 DPRINTF("data size is %zu, put on overflow page",
1572 node_size -= data->mv_size - sizeof(pgno_t);
1573 if ((ofp = mdbenv_new_page(db->md_env, P_OVERFLOW, ovpages)) == NULL)
1575 DPRINTF("allocated overflow page %lu", ofp->p.mp_pgno);
1580 if (node_size + sizeof(indx_t) > SIZELEFT(mp)) {
1581 DPRINTF("not enough room in page %lu, got %lu ptrs",
1582 mp->mp_pgno, NUMKEYS(mp));
1583 DPRINTF("upper - lower = %u - %u = %u", mp->mp_upper, mp->mp_lower,
1584 mp->mp_upper - mp->mp_lower);
1585 DPRINTF("node size = %zu", node_size);
1589 /* Move higher pointers up one slot. */
1590 for (i = NUMKEYS(mp); i > indx; i--)
1591 mp->mp_ptrs[i] = mp->mp_ptrs[i - 1];
1593 /* Adjust free space offsets. */
1594 ofs = mp->mp_upper - node_size;
1595 assert(ofs >= mp->mp_lower + sizeof(indx_t));
1596 mp->mp_ptrs[indx] = ofs;
1598 mp->mp_lower += sizeof(indx_t);
1600 /* Write the node data. */
1601 node = NODEPTR(mp, indx);
1602 node->mn_ksize = (key == NULL) ? 0 : key->mv_size;
1603 node->mn_flags = flags;
1605 node->mn_dsize = data->mv_size;
1607 node->mn_pgno = pgno;
1610 bcopy(key->mv_data, NODEKEY(node), key->mv_size);
1615 if (F_ISSET(flags, F_BIGDATA))
1616 bcopy(data->mv_data, node->mn_data + key->mv_size,
1619 bcopy(data->mv_data, node->mn_data + key->mv_size,
1622 bcopy(&ofp->p.mp_pgno, node->mn_data + key->mv_size,
1624 bcopy(data->mv_data, METADATA(&ofp->p), data->mv_size);
1632 mdb_del_node(MDB_db *db, MDB_page *mp, indx_t indx)
1635 indx_t i, j, numkeys, ptr;
1639 DPRINTF("delete node %u on %s page %lu", indx,
1640 IS_LEAF(mp) ? "leaf" : "branch", mp->mp_pgno);
1641 assert(indx < NUMKEYS(mp));
1643 node = NODEPTR(mp, indx);
1644 sz = NODESIZE + node->mn_ksize;
1646 if (F_ISSET(node->mn_flags, F_BIGDATA))
1647 sz += sizeof(pgno_t);
1649 sz += NODEDSZ(node);
1652 ptr = mp->mp_ptrs[indx];
1653 numkeys = NUMKEYS(mp);
1654 for (i = j = 0; i < numkeys; i++) {
1656 mp->mp_ptrs[j] = mp->mp_ptrs[i];
1657 if (mp->mp_ptrs[i] < ptr)
1658 mp->mp_ptrs[j] += sz;
1663 base = (char *)mp + mp->mp_upper;
1664 bcopy(base, base + sz, ptr - mp->mp_upper);
1666 mp->mp_lower -= sizeof(indx_t);
1671 mdb_cursor_open(MDB_db *db, MDB_txn *txn, MDB_cursor **ret)
1675 if (db == NULL || ret == NULL)
1678 if (txn != NULL && db->md_env != txn->mt_env) {
1682 if ((cursor = calloc(1, sizeof(*cursor))) != NULL) {
1683 SLIST_INIT(&cursor->mc_stack);
1685 cursor->mc_txn = txn;
1694 mdb_cursor_close(MDB_cursor *cursor)
1696 if (cursor != NULL) {
1697 while (!CURSOR_EMPTY(cursor))
1698 cursor_pop_page(cursor);
1700 /* btree_close(cursor->bt); */
1706 mdb_update_key(MDB_db *db, MDB_page *mp, indx_t indx,
1709 indx_t ptr, i, numkeys;
1715 node = NODEPTR(mp, indx);
1716 ptr = mp->mp_ptrs[indx];
1717 DPRINTF("update key %u (ofs %u) [%.*s] to [%.*s] on page %lu",
1719 (int)node->mn_ksize, (char *)NODEKEY(node),
1720 (int)key->mv_size, (char *)key->mv_data,
1723 delta = key->mv_size - node->mn_ksize;
1725 if (delta > 0 && SIZELEFT(mp) < delta) {
1726 DPRINTF("OUCH! Not enough room, delta = %d", delta);
1730 numkeys = NUMKEYS(mp);
1731 for (i = 0; i < numkeys; i++) {
1732 if (mp->mp_ptrs[i] <= ptr)
1733 mp->mp_ptrs[i] -= delta;
1736 base = (char *)mp + mp->mp_upper;
1737 len = ptr - mp->mp_upper + NODESIZE;
1738 bcopy(base, base - delta, len);
1739 mp->mp_upper -= delta;
1741 node = NODEPTR(mp, indx);
1742 node->mn_ksize = key->mv_size;
1745 bcopy(key->mv_data, NODEKEY(node), key->mv_size);
1750 /* Move a node from src to dst.
1753 mdb_move_node(MDB_db *bt, MDB_pageparent *src, indx_t srcindx,
1754 MDB_pageparent *dst, indx_t dstindx)
1760 srcnode = NODEPTR(src->mp_page, srcindx);
1761 DPRINTF("moving %s node %u [%.*s] on page %lu to node %u on page %lu",
1762 IS_LEAF(src->mp_page) ? "leaf" : "branch",
1764 (int)srcnode->mn_ksize, (char *)NODEKEY(srcnode),
1765 src->mp_page->mp_pgno,
1766 dstindx, dst->mp_page->mp_pgno);
1768 /* Mark src and dst as dirty. */
1769 if ((rc = mdb_touch(bt->md_env->me_txn, src)) ||
1770 (rc = mdb_touch(bt->md_env->me_txn, dst)))
1773 /* Add the node to the destination page.
1775 key.mv_size = srcnode->mn_ksize;
1776 key.mv_data = NODEKEY(srcnode);
1777 data.mv_size = NODEDSZ(srcnode);
1778 data.mv_data = NODEDATA(srcnode);
1779 rc = mdb_add_node(bt, dst->mp_page, dstindx, &key, &data, NODEPGNO(srcnode),
1781 if (rc != MDB_SUCCESS)
1784 /* Delete the node from the source page.
1786 mdb_del_node(bt, src->mp_page, srcindx);
1788 /* Update the parent separators.
1790 if (srcindx == 0 && src->mp_pi != 0) {
1791 DPRINTF("update separator for source page %lu to [%.*s]",
1792 src->mp_page->mp_pgno, (int)key.mv_size, (char *)key.mv_data);
1793 if ((rc = mdb_update_key(bt, src->mp_parent, src->mp_pi,
1794 &key)) != MDB_SUCCESS)
1798 if (srcindx == 0 && IS_BRANCH(src->mp_page)) {
1800 nullkey.mv_size = 0;
1801 assert(mdb_update_key(bt, src->mp_page, 0, &nullkey) == MDB_SUCCESS);
1804 if (dstindx == 0 && dst->mp_pi != 0) {
1805 DPRINTF("update separator for destination page %lu to [%.*s]",
1806 dst->mp_page->mp_pgno, (int)key.mv_size, (char *)key.mv_data);
1807 if ((rc = mdb_update_key(bt, dst->mp_parent, dst->mp_pi,
1808 &key)) != MDB_SUCCESS)
1812 if (dstindx == 0 && IS_BRANCH(dst->mp_page)) {
1814 nullkey.mv_size = 0;
1815 assert(mdb_update_key(bt, dst->mp_page, 0, &nullkey) == MDB_SUCCESS);
1822 mdb_merge(MDB_db *bt, MDB_pageparent *src, MDB_pageparent *dst)
1831 DPRINTF("merging page %lu and %lu", src->mp_page->mp_pgno, dst->mp_page->mp_pgno);
1833 assert(src->mp_parent); /* can't merge root page */
1834 assert(dst->mp_parent);
1835 assert(bt->md_env->me_txn != NULL);
1837 /* Mark src and dst as dirty. */
1838 if ((rc = mdb_touch(bt->md_env->me_txn, src)) ||
1839 (rc = mdb_touch(bt->md_env->me_txn, dst)))
1842 /* Move all nodes from src to dst.
1844 for (i = 0; i < NUMKEYS(src->mp_page); i++) {
1845 srcnode = NODEPTR(src->mp_page, i);
1847 key.mv_size = srcnode->mn_ksize;
1848 key.mv_data = NODEKEY(srcnode);
1849 data.mv_size = NODEDSZ(srcnode);
1850 data.mv_data = NODEDATA(srcnode);
1851 rc = mdb_add_node(bt, dst->mp_page, NUMKEYS(dst->mp_page), &key,
1852 &data, NODEPGNO(srcnode), srcnode->mn_flags);
1853 if (rc != MDB_SUCCESS)
1857 DPRINTF("dst page %lu now has %lu keys (%.1f%% filled)",
1858 dst->mp_page->mp_pgno, NUMKEYS(dst->mp_page), (float)PAGEFILL(bt->md_env, dst->mp_page) / 10);
1860 /* Unlink the src page from parent.
1862 mdb_del_node(bt, src->mp_parent, src->mp_pi);
1863 if (src->mp_pi == 0) {
1865 if ((rc = mdb_update_key(bt, src->mp_parent, 0, &key)) != MDB_SUCCESS)
1869 if (IS_LEAF(src->mp_page))
1870 bt->md_env->me_meta.mm_stat.ms_leaf_pages--;
1872 bt->md_env->me_meta.mm_stat.ms_branch_pages--;
1874 mpp.mp_page = src->mp_parent;
1875 dh = (MDB_dhead *)src->mp_parent;
1877 mpp.mp_parent = dh->md_parent;
1878 mpp.mp_pi = dh->md_pi;
1880 return mdb_rebalance(bt, &mpp);
1883 #define FILL_THRESHOLD 250
1886 mdb_rebalance(MDB_db *db, MDB_pageparent *mpp)
1891 indx_t si = 0, di = 0;
1894 assert(db->md_env->me_txn != NULL);
1895 assert(mpp != NULL);
1897 DPRINTF("rebalancing %s page %lu (has %lu keys, %.1f%% full)",
1898 IS_LEAF(mpp->mp_page) ? "leaf" : "branch",
1899 mpp->mp_page->mp_pgno, NUMKEYS(mpp->mp_page), (float)PAGEFILL(db->md_env, mpp->mp_page) / 10);
1901 if (PAGEFILL(db->md_env, mpp->mp_page) >= FILL_THRESHOLD) {
1902 DPRINTF("no need to rebalance page %lu, above fill threshold",
1903 mpp->mp_page->mp_pgno);
1907 if (mpp->mp_parent == NULL) {
1908 if (NUMKEYS(mpp->mp_page) == 0) {
1909 DPRINTF("tree is completely empty");
1910 db->md_env->me_txn->mt_root = P_INVALID;
1911 db->md_env->me_meta.mm_stat.ms_depth--;
1912 db->md_env->me_meta.mm_stat.ms_leaf_pages--;
1913 } else if (IS_BRANCH(mpp->mp_page) && NUMKEYS(mpp->mp_page) == 1) {
1914 DPRINTF("collapsing root page!");
1915 db->md_env->me_txn->mt_root = NODEPGNO(NODEPTR(mpp->mp_page, 0));
1916 if ((root = mdbenv_get_page(db->md_env, db->md_env->me_txn->mt_root)) == NULL)
1918 db->md_env->me_meta.mm_stat.ms_depth--;
1919 db->md_env->me_meta.mm_stat.ms_branch_pages--;
1921 DPRINTF("root page doesn't need rebalancing");
1925 /* The parent (branch page) must have at least 2 pointers,
1926 * otherwise the tree is invalid.
1928 assert(NUMKEYS(mpp->mp_parent) > 1);
1930 /* Leaf page fill factor is below the threshold.
1931 * Try to move keys from left or right neighbor, or
1932 * merge with a neighbor page.
1937 if (mpp->mp_pi == 0) {
1938 /* We're the leftmost leaf in our parent.
1940 DPRINTF("reading right neighbor");
1941 node = NODEPTR(mpp->mp_parent, mpp->mp_pi + 1);
1942 if ((npp.mp_page = mdbenv_get_page(db->md_env, NODEPGNO(node))) == NULL)
1944 npp.mp_pi = mpp->mp_pi + 1;
1946 di = NUMKEYS(mpp->mp_page);
1948 /* There is at least one neighbor to the left.
1950 DPRINTF("reading left neighbor");
1951 node = NODEPTR(mpp->mp_parent, mpp->mp_pi - 1);
1952 if ((npp.mp_page = mdbenv_get_page(db->md_env, NODEPGNO(node))) == NULL)
1954 npp.mp_pi = mpp->mp_pi - 1;
1955 si = NUMKEYS(npp.mp_page) - 1;
1958 npp.mp_parent = mpp->mp_parent;
1960 DPRINTF("found neighbor page %lu (%lu keys, %.1f%% full)",
1961 npp.mp_page->mp_pgno, NUMKEYS(npp.mp_page), (float)PAGEFILL(db->md_env, npp.mp_page) / 10);
1963 /* If the neighbor page is above threshold and has at least two
1964 * keys, move one key from it.
1966 * Otherwise we should try to merge them.
1968 if (PAGEFILL(db->md_env, npp.mp_page) >= FILL_THRESHOLD && NUMKEYS(npp.mp_page) >= 2)
1969 return mdb_move_node(db, &npp, si, mpp, di);
1970 else { /* FIXME: if (has_enough_room()) */
1971 if (mpp->mp_pi == 0)
1972 return mdb_merge(db, &npp, mpp);
1974 return mdb_merge(db, mpp, &npp);
1979 mdb_del(MDB_db *bt, MDB_txn *txn,
1980 MDB_val *key, MDB_val *data)
1987 DPRINTF("========> delete key %.*s", (int)key->mv_size, (char *)key->mv_data);
1989 assert(key != NULL);
1991 if (bt == NULL || txn == NULL)
1994 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
1998 if (bt->md_env->me_txn != txn) {
2002 if (key->mv_size == 0 || key->mv_size > MAXKEYSIZE) {
2006 if ((rc = mdb_search_page(bt, txn, key, NULL, 1, &mpp)) != MDB_SUCCESS)
2009 leaf = mdb_search_node(bt, mpp.mp_page, key, &exact, &ki);
2010 if (leaf == NULL || !exact) {
2014 if (data && (rc = mdb_read_data(bt, NULL, leaf, data)) != MDB_SUCCESS)
2017 mdb_del_node(bt, mpp.mp_page, ki);
2018 bt->md_env->me_meta.mm_stat.ms_entries--;
2019 rc = mdb_rebalance(bt, &mpp);
2020 if (rc != MDB_SUCCESS)
2021 txn->mt_flags |= MDB_TXN_ERROR;
2026 /* Split page <*mpp>, and insert <key,(data|newpgno)> in either left or
2027 * right sibling, at index <*newindxp> (as if unsplit). Updates *mpp and
2028 * *newindxp with the actual values after split, ie if *mpp and *newindxp
2029 * refer to a node in the new right sibling page.
2032 mdb_split(MDB_db *bt, MDB_page **mpp, unsigned int *newindxp,
2033 MDB_val *newkey, MDB_val *newdata, pgno_t newpgno)
2036 int rc = MDB_SUCCESS, ins_new = 0;
2039 unsigned int i, j, split_indx;
2041 MDB_page *pright, *p;
2042 MDB_val sepkey, rkey, rdata;
2044 MDB_dpage *mdp, *rdp, *pdp;
2048 assert(bt->md_env != NULL);
2050 dh = ((MDB_dhead *)*mpp) - 1;
2051 mdp = (MDB_dpage *)dh;
2052 newindx = *newindxp;
2054 DPRINTF("-----> splitting %s page %lu and adding [%.*s] at index %i",
2055 IS_LEAF(&mdp->p) ? "leaf" : "branch", mdp->p.mp_pgno,
2056 (int)newkey->mv_size, (char *)newkey->mv_data, *newindxp);
2058 if (mdp->h.md_parent == NULL) {
2059 if ((pdp = mdbenv_new_page(bt->md_env, P_BRANCH, 1)) == NULL)
2062 mdp->h.md_parent = &pdp->p;
2063 bt->md_env->me_txn->mt_root = pdp->p.mp_pgno;
2064 DPRINTF("root split! new root = %lu", pdp->p.mp_pgno);
2065 bt->md_env->me_meta.mm_stat.ms_depth++;
2067 /* Add left (implicit) pointer. */
2068 if (mdb_add_node(bt, &pdp->p, 0, NULL, NULL,
2069 mdp->p.mp_pgno, 0) != MDB_SUCCESS)
2072 DPRINTF("parent branch page is %lu", mdp->h.md_parent->mp_pgno);
2075 /* Create a right sibling. */
2076 if ((rdp = mdbenv_new_page(bt->md_env, mdp->p.mp_flags, 1)) == NULL)
2078 rdp->h.md_parent = mdp->h.md_parent;
2079 rdp->h.md_pi = mdp->h.md_pi + 1;
2080 DPRINTF("new right sibling: page %lu", rdp->p.mp_pgno);
2082 /* Move half of the keys to the right sibling. */
2083 if ((copy = malloc(bt->md_env->me_head.mh_psize)) == NULL)
2085 bcopy(&mdp->p, copy, bt->md_env->me_head.mh_psize);
2086 bzero(&mdp->p.mp_ptrs, bt->md_env->me_head.mh_psize - PAGEHDRSZ);
2087 mdp->p.mp_lower = PAGEHDRSZ;
2088 mdp->p.mp_upper = bt->md_env->me_head.mh_psize;
2090 split_indx = NUMKEYS(copy) / 2 + 1;
2092 /* First find the separating key between the split pages.
2094 bzero(&sepkey, sizeof(sepkey));
2095 if (newindx == split_indx) {
2096 sepkey.mv_size = newkey->mv_size;
2097 sepkey.mv_data = newkey->mv_data;
2099 node = NODEPTR(copy, split_indx);
2100 sepkey.mv_size = node->mn_ksize;
2101 sepkey.mv_data = NODEKEY(node);
2104 DPRINTF("separator is [%.*s]", (int)sepkey.mv_size, (char *)sepkey.mv_data);
2106 /* Copy separator key to the parent.
2108 if (SIZELEFT(rdp->h.md_parent) < mdb_branch_size(bt, &sepkey)) {
2109 rc = mdb_split(bt, &rdp->h.md_parent, &rdp->h.md_pi,
2110 &sepkey, NULL, rdp->p.mp_pgno);
2112 /* Right page might now have changed parent.
2113 * Check if left page also changed parent.
2115 if (rdp->h.md_parent != mdp->h.md_parent &&
2116 mdp->h.md_pi >= NUMKEYS(mdp->h.md_parent)) {
2117 mdp->h.md_parent = rdp->h.md_parent;
2118 mdp->h.md_pi = rdp->h.md_pi - 1;
2121 rc = mdb_add_node(bt, rdp->h.md_parent, rdp->h.md_pi,
2122 &sepkey, NULL, rdp->p.mp_pgno, 0);
2124 if (rc != MDB_SUCCESS) {
2129 for (i = j = 0; i <= NUMKEYS(copy); j++) {
2130 if (i < split_indx) {
2131 /* Re-insert in left sibling. */
2134 /* Insert in right sibling. */
2135 if (i == split_indx)
2136 /* Reset insert index for right sibling. */
2137 j = (i == newindx && ins_new);
2141 if (i == newindx && !ins_new) {
2142 /* Insert the original entry that caused the split. */
2143 rkey.mv_data = newkey->mv_data;
2144 rkey.mv_size = newkey->mv_size;
2145 if (IS_LEAF(&mdp->p)) {
2146 rdata.mv_data = newdata->mv_data;
2147 rdata.mv_size = newdata->mv_size;
2154 /* Update page and index for the new key. */
2157 } else if (i == NUMKEYS(copy)) {
2160 node = NODEPTR(copy, i);
2161 rkey.mv_data = NODEKEY(node);
2162 rkey.mv_size = node->mn_ksize;
2163 if (IS_LEAF(&mdp->p)) {
2164 rdata.mv_data = NODEDATA(node);
2165 rdata.mv_size = node->mn_dsize;
2167 pgno = node->mn_pgno;
2168 flags = node->mn_flags;
2173 if (!IS_LEAF(&mdp->p) && j == 0) {
2174 /* First branch index doesn't need key data. */
2178 rc = mdb_add_node(bt, &pdp->p, j, &rkey, &rdata, pgno,flags);
2186 mdb_put(MDB_db *bt, MDB_txn *txn,
2187 MDB_val *key, MDB_val *data, unsigned int flags)
2189 int rc = MDB_SUCCESS, exact;
2194 assert(key != NULL);
2195 assert(data != NULL);
2197 if (bt == NULL || txn == NULL)
2200 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
2204 if (bt->md_env->me_txn != txn) {
2208 if (key->mv_size == 0 || key->mv_size > MAXKEYSIZE) {
2212 DPRINTF("==> put key %.*s, size %zu, data size %zu",
2213 (int)key->mv_size, (char *)key->mv_data, key->mv_size, data->mv_size);
2215 rc = mdb_search_page(bt, txn, key, NULL, 1, &mpp);
2216 if (rc == MDB_SUCCESS) {
2217 leaf = mdb_search_node(bt, mpp.mp_page, key, &exact, &ki);
2218 if (leaf && exact) {
2219 if (F_ISSET(flags, MDB_NOOVERWRITE)) {
2220 DPRINTF("duplicate key %.*s",
2221 (int)key->mv_size, (char *)key->mv_data);
2224 mdb_del_node(bt, mpp.mp_page, ki);
2226 if (leaf == NULL) { /* append if not found */
2227 ki = NUMKEYS(mpp.mp_page);
2228 DPRINTF("appending key at index %i", ki);
2230 } else if (rc == ENOENT) {
2232 /* new file, just write a root leaf page */
2233 DPRINTF("allocating new root leaf page");
2234 if ((dp = mdbenv_new_page(bt->md_env, P_LEAF, 1)) == NULL) {
2237 mpp.mp_page = &dp->p;
2238 txn->mt_root = mpp.mp_page->mp_pgno;
2239 bt->md_env->me_meta.mm_stat.ms_depth++;
2245 assert(IS_LEAF(mpp.mp_page));
2246 DPRINTF("there are %lu keys, should insert new key at index %i",
2247 NUMKEYS(mpp.mp_page), ki);
2249 if (SIZELEFT(mpp.mp_page) < mdb_leaf_size(bt, key, data)) {
2250 rc = mdb_split(bt, &mpp.mp_page, &ki, key, data, P_INVALID);
2252 /* There is room already in this leaf page. */
2253 rc = mdb_add_node(bt, mpp.mp_page, ki, key, data, 0, 0);
2256 if (rc != MDB_SUCCESS)
2257 txn->mt_flags |= MDB_TXN_ERROR;
2259 bt->md_env->me_meta.mm_stat.ms_entries++;
2266 mdbenv_compact_tree(MDB_env *env, pgno_t pgno, MDB_env *envc)
2270 pgno_t *pnext, next;
2275 /* Get the page and make a copy of it.
2277 if ((mp = mdbenv_get_page(env, pgno)) == NULL)
2279 if ((p = malloc(env->me_head.mh_psize)) == NULL)
2281 bcopy(mp, p, env->me_head.mh_psize);
2283 /* Go through all nodes in the (copied) page and update the
2286 if (F_ISSET(p->mp_flags, P_BRANCH)) {
2287 for (i = 0; i < NUMKEYS(p); i++) {
2288 node = NODEPTR(p, i);
2289 node->mn_pgno = mdbenv_compact_tree(env, node->mn_pgno, envc);
2290 if (node->mn_pgno == P_INVALID) {
2295 } else if (F_ISSET(p->mp_flags, P_LEAF)) {
2296 for (i = 0; i < NUMKEYS(p); i++) {
2297 node = NODEPTR(p, i);
2298 if (F_ISSET(node->mn_flags, F_BIGDATA)) {
2299 bcopy(NODEDATA(node), &next, sizeof(next));
2300 next = mdbenv_compact_tree(env, next, envc);
2301 if (next == P_INVALID) {
2305 bcopy(&next, NODEDATA(node), sizeof(next));
2308 } else if (F_ISSET(p->mp_flags, P_OVERFLOW)) {
2309 ; /* handled below */
2313 pgno = p->mp_pgno = envc->me_txn->mt_next_pgno++;
2314 rc = write(envc->me_fd, p, env->me_head.mh_psize);
2316 if (rc != (ssize_t)env->me_head.mh_psize)
2319 if (F_ISSET(mp->mp_flags, P_OVERFLOW)) {
2320 if (mp->mp_pages > 1) {
2321 size_t len = (mp->mp_pages-1) * env->me_head.mh_psize;
2322 envc->me_txn->mt_next_pgno += mp->mp_pages-1;
2323 rc = write(envc->me_fd, mp+1, len);
2332 mdbenv_compact(MDB_env *env)
2334 char *compact_path = NULL;
2336 MDB_txn *txn, *txnc = NULL;
2340 assert(env != NULL);
2342 if (env->me_path == NULL) {
2346 DPRINTF("compacting mdbenv %p with path %s", env, env->me_path);
2348 rc = mdb_txn_begin(env, 0, &txn);
2351 asprintf(&compact_path, "%s.compact.XXXXXX", env->me_path);
2352 fd = mkstemp(compact_path);
2360 rc = mdbenv_create(&envc, env->me_mapsize);
2361 if (rc) goto failed;
2364 rc = mdbenv_open2(envc, env->me_flags);
2365 if (rc) goto failed;
2367 bcopy(&env->me_meta, &envc->me_meta, sizeof(env->me_meta));
2368 envc->me_meta.mm_stat.ms_revisions = 0;
2370 rc = mdb_txn_begin(envc, 0, &txnc);
2371 if (rc) goto failed;
2373 if (env->me_meta.mm_root != P_INVALID) {
2374 root = mdbenv_compact_tree(env, env->me_meta.mm_root, envc);
2375 if (root == P_INVALID)
2377 if ((rc = mdbenv_write_meta(envc, root, 0)) != MDB_SUCCESS)
2383 DPRINTF("renaming %s to %s", compact_path, env->me_path);
2384 if (rename(compact_path, env->me_path) != 0) {
2389 /* Write a "tombstone" meta page so other processes can pick up
2390 * the change and re-open the file.
2392 if (mdbenv_write_meta(env, P_INVALID, MDB_TOMBSTONE) != MDB_SUCCESS)
2396 mdb_txn_abort(txnc);
2403 mdb_txn_abort(txnc);
2404 unlink(compact_path);
2411 mdbenv_get_flags(MDB_env *env, unsigned int *arg)
2416 *arg = (env->me_flags & ~MDB_FIXPADDING);
2421 mdbenv_get_path(MDB_env *env, const char **arg)
2426 *arg = env->me_path;
2431 mdbenv_stat(MDB_env *env, MDB_stat **arg)
2433 if (env == NULL || arg == NULL)
2436 *arg = &env->me_meta.mm_stat;
2441 int mdb_open(MDB_env *env, MDB_txn *txn, const char *name, unsigned int flags, MDB_db **db)
2444 *db = (MDB_db *)&env->me_db;
2450 int mdb_stat(MDB_db *db, MDB_stat **arg)
2452 if (db == NULL || arg == NULL)
2455 *arg = &db->md_stat;
2460 void mdb_close(MDB_db *db)