31 # define DPRINTF(...) do { fprintf(stderr, "%s:%d: ", __func__, __LINE__); \
32 fprintf(stderr, __VA_ARGS__); \
33 fprintf(stderr, "\n"); } while(0)
40 #define MDB_MAGIC 0xBEEFC0DE
42 #define MAXKEYSIZE 255
44 #define P_INVALID (~0L)
46 #define F_ISSET(w, f) (((w) & (f)) == (f))
49 typedef uint16_t indx_t;
51 #define DEFAULT_READERS 126
52 #define DEFAULT_MAPSIZE 1048576
54 /* Lock descriptor stuff */
59 typedef struct MDB_rxbody {
64 #define CACHELINE 64 /* most CPUs. Itanium uses 128 */
67 typedef struct MDB_reader {
69 /* cache line alignment */
70 char pad[CACHELINE-sizeof(MDB_rxbody)];
75 uint32_t mt_version; \
76 pthread_mutex_t mt_mutex; \
78 uint32_t mt_numreaders
79 typedef struct MDB_txbody {
83 typedef struct MDB_txninfo {
85 char pad[CACHELINE-sizeof(MDB_txbody)];
86 pthread_mutex_t mt_wmutex;
87 char pad2[CACHELINE-sizeof(pthread_mutex_t)];
88 MDB_reader mt_readers[1];
91 /* Common header for all page types. Overflow pages
92 * occupy a number of contiguous pages with no
93 * headers on any page after the first.
95 typedef struct MDB_page { /* represents a page of storage */
96 pgno_t mp_pgno; /* page number */
97 #define P_BRANCH 0x01 /* branch page */
98 #define P_LEAF 0x02 /* leaf page */
99 #define P_OVERFLOW 0x04 /* overflow page */
100 #define P_META 0x08 /* meta page */
101 #define P_HEAD 0x10 /* header page */
102 #define P_DIRTY 0x20 /* dirty page */
104 #define mp_lower mp_pb.pb.pb_lower
105 #define mp_upper mp_pb.pb.pb_upper
106 #define mp_pages mp_pb.pb_pages
109 indx_t pb_lower; /* lower bound of free space */
110 indx_t pb_upper; /* upper bound of free space */
112 uint32_t pb_pages; /* number of overflow pages */
114 indx_t mp_ptrs[1]; /* dynamic size */
117 #define PAGEHDRSZ offsetof(MDB_page, mp_ptrs)
119 #define NUMKEYS(p) (((p)->mp_lower - PAGEHDRSZ) >> 1)
120 #define SIZELEFT(p) (indx_t)((p)->mp_upper - (p)->mp_lower)
121 #define PAGEFILL(env, p) (1000 * ((env)->me_head.mh_psize - PAGEHDRSZ - SIZELEFT(p)) / \
122 ((env)->me_head.mh_psize - PAGEHDRSZ))
123 #define IS_LEAF(p) F_ISSET((p)->mp_flags, P_LEAF)
124 #define IS_BRANCH(p) F_ISSET((p)->mp_flags, P_BRANCH)
125 #define IS_OVERFLOW(p) F_ISSET((p)->mp_flags, P_OVERFLOW)
127 typedef struct MDB_head { /* header page content */
131 uint32_t mh_psize; /* page size */
132 void *mh_address; /* address for fixed mapping */
133 size_t mh_mapsize; /* size of mmap region */
136 typedef struct MDB_meta { /* meta (footer) page content */
138 pgno_t mm_root; /* page number of root page */
139 pgno_t mm_last_pg; /* last used page in file */
140 ulong mm_txnid; /* txnid that committed this page */
143 typedef struct MDB_dhead { /* a dirty page */
144 SIMPLEQ_ENTRY(MDB_dpage) md_next; /* queue of dirty pages */
146 int md_pi; /* parent index */
150 typedef struct MDB_dpage {
155 SIMPLEQ_HEAD(dirty_queue, MDB_dpage);
157 typedef struct MDB_pageparent {
163 static MDB_dpage *mdb_newpage(MDB_txn *txn, MDB_page *parent, int parent_idx, int num);
164 static int mdb_touch(MDB_txn *txn, MDB_pageparent *mp);
166 typedef struct MDB_ppage { /* ordered list of pages */
167 SLIST_ENTRY(MDB_ppage) mp_entry;
169 unsigned int mp_ki; /* cursor index on page */
171 SLIST_HEAD(page_stack, MDB_ppage);
173 #define CURSOR_EMPTY(c) SLIST_EMPTY(&(c)->mc_stack)
174 #define CURSOR_TOP(c) SLIST_FIRST(&(c)->mc_stack)
175 #define CURSOR_POP(c) SLIST_REMOVE_HEAD(&(c)->mc_stack, mp_entry)
176 #define CURSOR_PUSH(c,p) SLIST_INSERT_HEAD(&(c)->mc_stack, p, mp_entry)
181 struct page_stack mc_stack; /* stack of parent pages */
182 short mc_initialized; /* 1 if initialized */
183 short mc_eof; /* 1 if end is reached */
186 #define METAHASHLEN offsetof(MDB_meta, mm_hash)
187 #define METADATA(p) ((void *)((char *)p + PAGEHDRSZ))
189 typedef struct MDB_node {
190 #define mn_pgno mn_p.np_pgno
191 #define mn_dsize mn_p.np_dsize
193 pgno_t np_pgno; /* child page number */
194 uint32_t np_dsize; /* leaf data size */
196 uint16_t mn_ksize; /* key size */
197 #define F_BIGDATA 0x01 /* data put on overflow page */
203 pgno_t mt_root; /* current / new root page */
204 pgno_t mt_next_pgno; /* next unallocated page */
205 pgno_t mt_first_pgno;
209 struct dirty_queue *dirty_queue; /* modified pages */
212 #define MDB_TXN_RDONLY 0x01 /* read-only transaction */
213 #define MDB_TXN_ERROR 0x02 /* an error has occurred */
214 unsigned int mt_flags;
217 /* Must be same as MDB_db, minus md_root/md_stat */
218 typedef struct MDB_db0 {
219 unsigned int md_flags;
220 MDB_cmp_func *md_cmp; /* user compare function */
221 MDB_rel_func *md_rel; /* user relocate function */
222 MDB_db *md_parent; /* parent tree */
227 unsigned int md_flags;
228 MDB_cmp_func *md_cmp; /* user compare function */
229 MDB_rel_func *md_rel; /* user relocate function */
230 MDB_db *md_parent; /* parent tree */
233 pgno_t md_root; /* page number of root page */
244 MDB_txninfo *me_txns;
246 MDB_db0 me_db; /* first DB, overlaps with meta */
248 MDB_txn *me_txn; /* current write transaction */
250 off_t me_size; /* current file size */
251 pthread_key_t me_txkey; /* thread-key for readers */
254 #define NODESIZE offsetof(MDB_node, mn_data)
256 #define INDXSIZE(k) (NODESIZE + ((k) == NULL ? 0 : (k)->mv_size))
257 #define LEAFSIZE(k, d) (NODESIZE + (k)->mv_size + (d)->mv_size)
258 #define NODEPTR(p, i) ((MDB_node *)((char *)(p) + (p)->mp_ptrs[i]))
259 #define NODEKEY(node) (void *)((node)->mn_data)
260 #define NODEDATA(node) (void *)((char *)(node)->mn_data + (node)->mn_ksize)
261 #define NODEPGNO(node) ((node)->mn_pgno)
262 #define NODEDSZ(node) ((node)->mn_dsize)
264 #define MDB_COMMIT_PAGES 64 /* max number of pages to write in one commit */
265 #define MDB_MAXCACHE_DEF 1024 /* max number of pages to keep in cache */
267 static int mdb_search_page_root(MDB_db *db,
269 MDB_cursor *cursor, int modify,
270 MDB_pageparent *mpp);
271 static int mdb_search_page(MDB_db *db,
272 MDB_txn *txn, MDB_val *key,
273 MDB_cursor *cursor, int modify,
274 MDB_pageparent *mpp);
276 static int mdbenv_write_header(MDB_env *env);
277 static int mdbenv_read_header(MDB_env *env);
278 static int mdb_check_meta_page(MDB_page *p);
279 static int mdbenv_read_meta(MDB_env *env);
280 static int mdbenv_write_meta(MDB_txn *txn);
281 static MDB_page *mdbenv_get_page(MDB_env *env, pgno_t pgno);
283 static MDB_node *mdb_search_node(MDB_db *db, MDB_page *mp,
284 MDB_val *key, int *exactp, unsigned int *kip);
285 static int mdb_add_node(MDB_db *bt, MDB_page *mp,
286 indx_t indx, MDB_val *key, MDB_val *data,
287 pgno_t pgno, uint8_t flags);
288 static void mdb_del_node(MDB_db *bt, MDB_page *mp,
290 static int mdb_read_data(MDB_db *bt, MDB_page *mp,
291 MDB_node *leaf, MDB_val *data);
293 static int mdb_rebalance(MDB_db *bt, MDB_pageparent *mp);
294 static int mdb_update_key(MDB_db *bt, MDB_page *mp,
295 indx_t indx, MDB_val *key);
296 static int mdb_move_node(MDB_db *bt,
297 MDB_pageparent *src, indx_t srcindx,
298 MDB_pageparent *dst, indx_t dstindx);
299 static int mdb_merge(MDB_db *bt, MDB_pageparent *src,
300 MDB_pageparent *dst);
301 static int mdb_split(MDB_db *bt, MDB_page **mpp,
302 unsigned int *newindxp, MDB_val *newkey,
303 MDB_val *newdata, pgno_t newpgno);
304 static MDB_dpage *mdbenv_new_page(MDB_env *env, uint32_t flags, int num);
306 static void cursor_pop_page(MDB_cursor *cursor);
307 static MDB_ppage *cursor_push_page(MDB_cursor *cursor,
310 static int mdb_set_key(MDB_db *bt, MDB_page *mp,
311 MDB_node *node, MDB_val *key);
312 static int mdb_sibling(MDB_cursor *cursor, int move_right);
313 static int mdb_cursor_next(MDB_cursor *cursor,
314 MDB_val *key, MDB_val *data);
315 static int mdb_cursor_set(MDB_cursor *cursor,
316 MDB_val *key, MDB_val *data, int *exactp);
317 static int mdb_cursor_first(MDB_cursor *cursor,
318 MDB_val *key, MDB_val *data);
320 static size_t mdb_leaf_size(MDB_db *bt, MDB_val *key,
322 static size_t mdb_branch_size(MDB_db *bt, MDB_val *key);
324 static pgno_t mdbenv_compact_tree(MDB_env *env, pgno_t pgno,
327 static int memncmp(const void *s1, size_t n1,
328 const void *s2, size_t n2);
329 static int memnrcmp(const void *s1, size_t n1,
330 const void *s2, size_t n2);
333 memncmp(const void *s1, size_t n1, const void *s2, size_t n2)
336 if (memcmp(s1, s2, n1) == 0)
340 if (memcmp(s1, s2, n2) == 0)
343 return memcmp(s1, s2, n1);
347 memnrcmp(const void *s1, size_t n1, const void *s2, size_t n2)
349 const unsigned char *p1;
350 const unsigned char *p2;
353 return n2 == 0 ? 0 : -1;
356 return n1 == 0 ? 0 : 1;
358 p1 = (const unsigned char *)s1 + n1 - 1;
359 p2 = (const unsigned char *)s2 + n2 - 1;
363 return (p2 == s2) ? 0 : -1;
365 return (p1 == p2) ? 0 : 1;
373 mdb_cmp(MDB_db *db, const MDB_val *a, const MDB_val *b)
375 return db->md_cmp(a, b);
379 _mdb_cmp(MDB_db *db, const MDB_val *key1, const MDB_val *key2)
381 if (F_ISSET(db->md_flags, MDB_REVERSEKEY))
382 return memnrcmp(key1->mv_data, key1->mv_size, key2->mv_data, key2->mv_size);
384 return memncmp((char *)key1->mv_data, key1->mv_size, key2->mv_data, key2->mv_size);
387 /* Allocate new page(s) for writing */
389 mdb_newpage(MDB_txn *txn, MDB_page *parent, int parent_idx, int num)
393 if ((dp = malloc(txn->mt_env->me_head.mh_psize * num + sizeof(MDB_dhead))) == NULL)
396 dp->h.md_parent = parent;
397 dp->h.md_pi = parent_idx;
398 SIMPLEQ_INSERT_TAIL(txn->mt_u.dirty_queue, dp, h.md_next);
399 dp->p.mp_pgno = txn->mt_next_pgno;
400 txn->mt_next_pgno += num;
405 /* Touch a page: make it dirty and re-insert into tree with updated pgno.
408 mdb_touch(MDB_txn *txn, MDB_pageparent *pp)
411 MDB_page *mp = pp->mp_page;
416 if (!F_ISSET(mp->mp_flags, P_DIRTY)) {
418 DPRINTF("touching page %lu -> %lu", mp->mp_pgno, txn->mt_next_pgno);
419 if ((dp = mdb_newpage(txn, pp->mp_parent, pp->mp_pi, 1)) == NULL)
421 pgno = dp->p.mp_pgno;
422 bcopy(mp, &dp->p, txn->mt_env->me_head.mh_psize);
425 mp->mp_flags |= P_DIRTY;
427 /* Update the page number to new touched page. */
428 if (pp->mp_parent != NULL)
429 NODEPGNO(NODEPTR(pp->mp_parent, pp->mp_pi)) = mp->mp_pgno;
436 mdbenv_sync(MDB_env *env)
439 if (!F_ISSET(env->me_flags, MDB_NOSYNC)) {
440 if (fsync(env->me_fd))
447 mdb_txn_begin(MDB_env *env, int rdonly, MDB_txn **ret)
452 if ((txn = calloc(1, sizeof(*txn))) == NULL) {
453 DPRINTF("calloc: %s", strerror(errno));
458 txn->mt_flags |= MDB_TXN_RDONLY;
460 txn->mt_u.dirty_queue = calloc(1, sizeof(*txn->mt_u.dirty_queue));
461 if (txn->mt_u.dirty_queue == NULL) {
465 SIMPLEQ_INIT(txn->mt_u.dirty_queue);
467 pthread_mutex_lock(&env->me_txns->mt_wmutex);
468 env->me_txns->mt_txnid++;
470 txn->mt_txnid = env->me_txns->mt_txnid;
472 MDB_reader *r = pthread_getspecific(env->me_txkey);
475 pthread_mutex_lock(&env->me_txns->mt_mutex);
476 for (i=0; i<env->me_maxreaders; i++) {
477 if (env->me_txns->mt_readers[i].mr_pid == 0) {
478 env->me_txns->mt_readers[i].mr_pid = getpid();
479 env->me_txns->mt_readers[i].mr_tid = pthread_self();
480 pthread_setspecific(env->me_txkey, &env->me_txns->mt_readers[i]);
481 if (i >= env->me_txns->mt_numreaders)
482 env->me_txns->mt_numreaders = i+1;
486 pthread_mutex_unlock(&env->me_txns->mt_mutex);
487 if (i == env->me_maxreaders) {
491 r->mr_txnid = txn->mt_txnid;
492 txn->mt_u.reader = r;
499 if ((rc = mdbenv_read_meta(env)) != MDB_SUCCESS) {
504 txn->mt_next_pgno = env->me_meta.mm_last_pg+1;
505 txn->mt_first_pgno = txn->mt_next_pgno;
506 txn->mt_root = env->me_meta.mm_root;
507 DPRINTF("begin transaction on mdbenv %p, root page %lu", env, txn->mt_root);
514 mdb_txn_abort(MDB_txn *txn)
523 DPRINTF("abort transaction on mdbenv %p, root page %lu", env, txn->mt_root);
525 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
526 txn->mt_u.reader->mr_txnid = 0;
528 /* Discard all dirty pages.
530 while (!SIMPLEQ_EMPTY(txn->mt_u.dirty_queue)) {
531 dp = SIMPLEQ_FIRST(txn->mt_u.dirty_queue);
532 SIMPLEQ_REMOVE_HEAD(txn->mt_u.dirty_queue, h.md_next);
537 DPRINTF("releasing write lock on txn %p", txn);
539 if (flock(txn->bt->fd, LOCK_UN) != 0) {
540 DPRINTF("failed to unlock fd %d: %s",
541 txn->bt->fd, strerror(errno));
544 free(txn->mt_u.dirty_queue);
546 pthread_mutex_unlock(&env->me_txns->mt_wmutex);
553 mdb_txn_commit(MDB_txn *txn)
561 struct iovec iov[MDB_COMMIT_PAGES];
564 assert(txn->mt_env != NULL);
568 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
569 DPRINTF("attempt to commit read-only transaction");
574 if (txn != env->me_txn) {
575 DPRINTF("attempt to commit unknown transaction");
580 if (F_ISSET(txn->mt_flags, MDB_TXN_ERROR)) {
581 DPRINTF("error flag is set, can't commit");
586 if (SIMPLEQ_EMPTY(txn->mt_u.dirty_queue))
589 DPRINTF("committing transaction on mdbenv %p, root page %lu",
592 /* Commit up to MDB_COMMIT_PAGES dirty pages to disk until done.
599 SIMPLEQ_FOREACH(dp, txn->mt_u.dirty_queue, h.md_next) {
600 if (dp->p.mp_pgno != next) {
601 lseek(env->me_fd, dp->p.mp_pgno * env->me_head.mh_psize, SEEK_SET);
602 next = dp->p.mp_pgno;
606 DPRINTF("committing page %lu", dp->p.mp_pgno);
607 iov[n].iov_len = env->me_head.mh_psize * dp->h.md_num;
608 iov[n].iov_base = &dp->p;
609 size += iov[n].iov_len;
610 next = dp->p.mp_pgno + dp->h.md_num;
611 /* clear dirty flag */
612 dp->p.mp_flags &= ~P_DIRTY;
613 if (++n >= MDB_COMMIT_PAGES) {
622 DPRINTF("committing %u dirty pages", n);
623 rc = writev(env->me_fd, iov, n);
627 DPRINTF("short write, filesystem full?");
629 DPRINTF("writev: %s", strerror(errno));
634 /* Drop the dirty pages.
636 while (!SIMPLEQ_EMPTY(txn->mt_u.dirty_queue)) {
637 dp = SIMPLEQ_FIRST(txn->mt_u.dirty_queue);
638 SIMPLEQ_REMOVE_HEAD(txn->mt_u.dirty_queue, h.md_next);
645 if ((n = mdbenv_sync(env)) != 0 ||
646 (n = mdbenv_write_meta(txn)) != MDB_SUCCESS ||
647 (n = mdbenv_sync(env)) != 0) {
652 pthread_mutex_unlock(&env->me_txns->mt_wmutex);
653 free(txn->mt_u.dirty_queue);
664 mdbenv_write_header(MDB_env *env)
672 DPRINTF("writing header page");
675 psize = sysconf(_SC_PAGE_SIZE);
677 if ((p = calloc(1, psize)) == NULL)
679 p->mp_flags = P_HEAD;
681 env->me_head.mh_psize = psize;
682 env->me_head.mh_flags = env->me_flags & 0xffff;
685 bcopy(&env->me_head, h, sizeof(*h));
687 rc = write(env->me_fd, p, env->me_head.mh_psize);
689 if (rc != (ssize_t)env->me_head.mh_psize) {
692 DPRINTF("short write, filesystem full?");
700 mdbenv_read_header(MDB_env *env)
709 /* We don't know the page size yet, so use a minimum value.
712 if ((rc = pread(env->me_fd, page, PAGESIZE, 0)) == 0) {
714 } else if (rc != PAGESIZE) {
717 DPRINTF("read: %s", strerror(errno));
721 p = (MDB_page *)page;
723 if (!F_ISSET(p->mp_flags, P_HEAD)) {
724 DPRINTF("page %lu not a header page", p->mp_pgno);
729 if (h->mh_magic != MDB_MAGIC) {
730 DPRINTF("header has invalid magic");
734 if (h->mh_version != MDB_VERSION) {
735 DPRINTF("database is version %u, expected version %u",
736 h->mh_version, MDB_VERSION);
740 bcopy(h, &env->me_head, sizeof(*h));
745 mdbenv_init_meta(MDB_env *env)
751 p = calloc(2, env->me_head.mh_psize);
753 p->mp_flags = P_META;
756 meta->mm_root = P_INVALID;
757 meta->mm_last_pg = 2;
759 q = (MDB_page *)((char *)p + env->me_head.mh_psize);
762 q->mp_flags = P_META;
765 meta->mm_root = P_INVALID;
766 meta->mm_last_pg = 2;
768 rc = write(env->me_fd, p, env->me_head.mh_psize * 2);
770 return (rc == env->me_head.mh_psize * 2) ? MDB_SUCCESS : errno;
774 mdbenv_write_meta(MDB_txn *txn)
782 assert(txn->mt_env != NULL);
784 DPRINTF("writing meta page for root page %lu", txn->mt_root);
788 bcopy(&env->me_meta, &meta, sizeof(meta));
789 meta.mm_root = txn->mt_root;
790 meta.mm_last_pg = txn->mt_next_pgno - 1;
791 meta.mm_txnid = txn->mt_txnid;
793 off = env->me_head.mh_psize;
794 if (!env->me_metatoggle)
798 lseek(env->me_fd, off, SEEK_SET);
799 rc = write(env->me_fd, &meta, sizeof(meta));
800 if (rc != sizeof(meta)) {
801 DPRINTF("write failed, disk error?");
808 /* Returns true if page p is a valid meta page, false otherwise.
811 mdb_check_meta_page(MDB_page *p)
813 if (!F_ISSET(p->mp_flags, P_META)) {
814 DPRINTF("page %lu not a meta page", p->mp_pgno);
822 mdbenv_read_meta(MDB_env *env)
830 if ((mp0 = mdbenv_get_page(env, 1)) == NULL ||
831 (mp1 = mdbenv_get_page(env, 2)) == NULL)
834 rc = mdb_check_meta_page(mp0);
837 rc = mdb_check_meta_page(mp1);
840 meta[0] = METADATA(mp0);
841 meta[1] = METADATA(mp1);
843 if (meta[0]->mm_txnid < meta[1]->mm_txnid)
846 bcopy(meta[toggle], &env->me_meta, sizeof(env->me_meta));
847 env->me_metatoggle = toggle;
849 DPRINTF("Using meta page %d", toggle);
855 mdbenv_create(MDB_env **env)
859 e = calloc(1, sizeof(*e));
860 if (!e) return ENOMEM;
862 e->me_head.mh_magic = MDB_MAGIC;
863 e->me_head.mh_version = MDB_VERSION;
864 e->me_head.mh_mapsize = DEFAULT_MAPSIZE;
865 e->me_maxreaders = DEFAULT_READERS;
872 mdbenv_set_mapsize(MDB_env *env, size_t size)
876 env->me_mapsize = env->me_head.mh_mapsize = size;
881 mdbenv_set_maxreaders(MDB_env *env, int readers)
883 env->me_maxreaders = readers;
888 mdbenv_get_maxreaders(MDB_env *env, int *readers)
890 if (!env || !readers)
892 *readers = env->me_maxreaders;
897 mdbenv_open2(MDB_env *env, unsigned int flags)
901 env->me_flags = flags;
902 env->me_meta.mm_root = P_INVALID;
904 if ((i = mdbenv_read_header(env)) != 0) {
907 DPRINTF("new mdbenv");
911 if (!env->me_mapsize)
912 env->me_mapsize = env->me_head.mh_mapsize;
915 if (env->me_head.mh_address && (flags & MDB_FIXEDMAP))
917 env->me_map = mmap(env->me_head.mh_address, env->me_mapsize, PROT_READ, i,
919 if (env->me_map == MAP_FAILED)
923 env->me_head.mh_mapsize = env->me_mapsize;
924 if (flags & MDB_FIXEDMAP)
925 env->me_head.mh_address = env->me_map;
926 i = mdbenv_write_header(env);
927 if (i != MDB_SUCCESS) {
928 munmap(env->me_map, env->me_mapsize);
931 i = mdbenv_init_meta(env);
932 if (i != MDB_SUCCESS) {
933 munmap(env->me_map, env->me_mapsize);
938 if ((i = mdbenv_read_meta(env)) != 0)
941 DPRINTF("opened database version %u, pagesize %u",
942 env->me_head.mh_version, env->me_head.mh_psize);
943 DPRINTF("depth: %u", env->me_meta.mm_stat.ms_depth);
944 DPRINTF("entries: %lu", env->me_meta.mm_stat.ms_entries);
945 DPRINTF("branch pages: %lu", env->me_meta.mm_stat.ms_branch_pages);
946 DPRINTF("leaf pages: %lu", env->me_meta.mm_stat.ms_leaf_pages);
947 DPRINTF("overflow pages: %lu", env->me_meta.mm_stat.ms_overflow_pages);
948 DPRINTF("root: %lu", env->me_meta.mm_root);
954 mdbenv_reader_dest(void *ptr)
956 MDB_reader *reader = ptr;
958 reader->mr_txnid = 0;
964 mdbenv_open(MDB_env *env, const char *path, unsigned int flags, mode_t mode)
966 int oflags, rc, shmid;
970 if (F_ISSET(flags, MDB_RDONLY))
973 oflags = O_RDWR | O_CREAT;
975 if ((env->me_fd = open(path, oflags, mode)) == -1)
978 env->me_shmkey = ftok(path, 'm');
979 size = (env->me_maxreaders-1) * sizeof(MDB_reader) + sizeof(MDB_txninfo);
980 shmid = shmget(env->me_shmkey, size, IPC_CREAT|IPC_EXCL|mode);
982 if (errno == EEXIST) {
983 shmid = shmget(env->me_shmkey, size, IPC_CREAT|mode);
986 env->me_txns = shmat(shmid, NULL, 0);
987 if (env->me_txns->mt_magic != MDB_MAGIC ||
988 env->me_txns->mt_version != MDB_VERSION) {
989 DPRINTF("invalid lock region %d", shmid);
998 pthread_mutexattr_t mattr;
1000 env->me_txns = shmat(shmid, NULL, 0);
1001 pthread_mutexattr_init(&mattr);
1002 pthread_mutexattr_setpshared(&mattr, PTHREAD_PROCESS_SHARED);
1003 pthread_mutex_init(&env->me_txns->mt_mutex, &mattr);
1004 pthread_mutex_init(&env->me_txns->mt_wmutex, &mattr);
1005 env->me_txns->mt_version = MDB_VERSION;
1006 env->me_txns->mt_magic = MDB_MAGIC;
1009 if ((rc = mdbenv_open2(env, flags)) != MDB_SUCCESS) {
1013 env->me_path = strdup(path);
1014 DPRINTF("opened dbenv %p", env);
1017 pthread_key_create(&env->me_txkey, mdbenv_reader_dest);
1023 mdbenv_close(MDB_env *env)
1031 munmap(env->me_map, env->me_mapsize);
1035 shmdt(env->me_txns);
1040 /* Search for key within a leaf page, using binary search.
1041 * Returns the smallest entry larger or equal to the key.
1042 * If exactp is non-null, stores whether the found entry was an exact match
1043 * in *exactp (1 or 0).
1044 * If kip is non-null, stores the index of the found entry in *kip.
1045 * If no entry larger or equal to the key is found, returns NULL.
1048 mdb_search_node(MDB_db *bt, MDB_page *mp, MDB_val *key,
1049 int *exactp, unsigned int *kip)
1057 DPRINTF("searching %lu keys in %s page %lu",
1059 IS_LEAF(mp) ? "leaf" : "branch",
1062 assert(NUMKEYS(mp) > 0);
1064 bzero(&nodekey, sizeof(nodekey));
1066 low = IS_LEAF(mp) ? 0 : 1;
1067 high = NUMKEYS(mp) - 1;
1068 while (low <= high) {
1069 i = (low + high) >> 1;
1070 node = NODEPTR(mp, i);
1072 nodekey.mv_size = node->mn_ksize;
1073 nodekey.mv_data = NODEKEY(node);
1076 rc = bt->md_cmp(key, &nodekey);
1078 rc = _mdb_cmp(bt, key, &nodekey);
1081 DPRINTF("found leaf index %u [%.*s], rc = %i",
1082 i, (int)nodekey.mv_size, (char *)nodekey.mv_data, rc);
1084 DPRINTF("found branch index %u [%.*s -> %lu], rc = %i",
1085 i, (int)node->mn_ksize, (char *)NODEKEY(node),
1096 if (rc > 0) { /* Found entry is less than the key. */
1097 i++; /* Skip to get the smallest entry larger than key. */
1098 if (i >= NUMKEYS(mp))
1099 /* There is no entry larger or equal to the key. */
1103 *exactp = (rc == 0);
1104 if (kip) /* Store the key index if requested. */
1107 return NODEPTR(mp, i);
1111 cursor_pop_page(MDB_cursor *cursor)
1115 top = CURSOR_TOP(cursor);
1118 DPRINTF("popped page %lu off cursor %p", top->mp_page->mp_pgno, cursor);
1124 cursor_push_page(MDB_cursor *cursor, MDB_page *mp)
1128 DPRINTF("pushing page %lu on cursor %p", mp->mp_pgno, cursor);
1130 if ((ppage = calloc(1, sizeof(*ppage))) == NULL)
1132 ppage->mp_page = mp;
1133 CURSOR_PUSH(cursor, ppage);
1138 mdbenv_get_page(MDB_env *env, pgno_t pgno)
1141 MDB_txn *txn = env->me_txn;
1143 if (txn && pgno >= txn->mt_first_pgno &&
1144 !SIMPLEQ_EMPTY(txn->mt_u.dirty_queue)) {
1146 SIMPLEQ_FOREACH(dp, txn->mt_u.dirty_queue, h.md_next) {
1147 if (dp->p.mp_pgno == pgno) {
1153 p = (MDB_page *)(env->me_map + env->me_head.mh_psize * pgno);
1159 mdb_search_page_root(MDB_db *bt, MDB_val *key,
1160 MDB_cursor *cursor, int modify, MDB_pageparent *mpp)
1162 MDB_page *mp = mpp->mp_page;
1165 if (cursor && cursor_push_page(cursor, mp) == NULL)
1168 while (IS_BRANCH(mp)) {
1172 DPRINTF("branch page %lu has %lu keys", mp->mp_pgno, NUMKEYS(mp));
1173 assert(NUMKEYS(mp) > 1);
1174 DPRINTF("found index 0 to page %lu", NODEPGNO(NODEPTR(mp, 0)));
1176 if (key == NULL) /* Initialize cursor to first page. */
1180 node = mdb_search_node(bt, mp, key, &exact, &i);
1182 i = NUMKEYS(mp) - 1;
1190 DPRINTF("following index %u for key %.*s",
1191 i, (int)key->mv_size, (char *)key->mv_data);
1192 assert(i >= 0 && i < NUMKEYS(mp));
1193 node = NODEPTR(mp, i);
1196 CURSOR_TOP(cursor)->mp_ki = i;
1198 mpp->mp_parent = mp;
1199 if ((mp = mdbenv_get_page(bt->md_env, NODEPGNO(node))) == NULL)
1204 if (cursor && cursor_push_page(cursor, mp) == NULL)
1208 MDB_dhead *dh = ((MDB_dhead *)mp)-1;
1209 if (rc = mdb_touch(bt->md_env->me_txn, mpp))
1211 dh = ((MDB_dhead *)mpp->mp_page)-1;
1212 dh->md_parent = mpp->mp_parent;
1213 dh->md_pi = mpp->mp_pi;
1220 DPRINTF("internal error, index points to a %02X page!?",
1225 DPRINTF("found leaf page %lu for key %.*s", mp->mp_pgno,
1226 key ? (int)key->mv_size : 0, key ? (char *)key->mv_data : NULL);
1231 /* Search for the page a given key should be in.
1232 * Stores a pointer to the found page in *mpp.
1233 * If key is NULL, search for the lowest page (used by mdb_cursor_first).
1234 * If cursor is non-null, pushes parent pages on the cursor stack.
1235 * If modify is true, visited pages are updated with new page numbers.
1238 mdb_search_page(MDB_db *db, MDB_txn *txn, MDB_val *key,
1239 MDB_cursor *cursor, int modify, MDB_pageparent *mpp)
1244 /* Can't modify pages outside a transaction. */
1245 if (txn == NULL && modify) {
1249 /* Choose which root page to start with. If a transaction is given
1250 * use the root page from the transaction, otherwise read the last
1251 * committed root page.
1254 if ((rc = mdbenv_read_meta(db->md_env)) != MDB_SUCCESS)
1256 root = db->md_env->me_meta.mm_root;
1257 } else if (F_ISSET(txn->mt_flags, MDB_TXN_ERROR)) {
1258 DPRINTF("transaction has failed, must abort");
1261 root = txn->mt_root;
1263 if (root == P_INVALID) { /* Tree is empty. */
1264 DPRINTF("tree is empty");
1268 if ((mpp->mp_page = mdbenv_get_page(db->md_env, root)) == NULL)
1271 DPRINTF("root page has flags 0x%X", mpp->mp_page->mp_flags);
1273 if (modify && !F_ISSET(mpp->mp_page->mp_flags, P_DIRTY)) {
1274 mpp->mp_parent = NULL;
1276 if ((rc = mdb_touch(txn, mpp)))
1278 txn->mt_root = mpp->mp_page->mp_pgno;
1281 return mdb_search_page_root(db, key, cursor, modify, mpp);
1285 mdb_read_data(MDB_db *db, MDB_page *mp, MDB_node *leaf,
1288 MDB_page *omp; /* overflow mpage */
1294 bzero(data, sizeof(*data));
1295 max = db->md_env->me_head.mh_psize - PAGEHDRSZ;
1297 if (!F_ISSET(leaf->mn_flags, F_BIGDATA)) {
1298 data->mv_size = leaf->mn_dsize;
1299 data->mv_data = NODEDATA(leaf);
1303 /* Read overflow data.
1305 data->mv_size = leaf->mn_dsize;
1306 bcopy(NODEDATA(leaf), &pgno, sizeof(pgno));
1307 if ((omp = mdbenv_get_page(db->md_env, pgno)) == NULL) {
1308 DPRINTF("read overflow page %lu failed", pgno);
1311 data->mv_data = omp;
1317 mdb_get(MDB_db *db, MDB_txn *txn,
1318 MDB_val *key, MDB_val *data)
1326 DPRINTF("===> get key [%.*s]", (int)key->mv_size, (char *)key->mv_data);
1331 if (txn != NULL && db->md_env != txn->mt_env) {
1335 if (key->mv_size == 0 || key->mv_size > MAXKEYSIZE) {
1339 if ((rc = mdb_search_page(db, txn, key, NULL, 0, &mpp)) != MDB_SUCCESS)
1342 leaf = mdb_search_node(db, mpp.mp_page, key, &exact, NULL);
1344 rc = mdb_read_data(db, mpp.mp_page, leaf, data);
1353 mdb_sibling(MDB_cursor *cursor, int move_right)
1357 MDB_ppage *parent, *top;
1360 top = CURSOR_TOP(cursor);
1361 if ((parent = SLIST_NEXT(top, mp_entry)) == NULL) {
1362 return ENOENT; /* root has no siblings */
1365 DPRINTF("parent page is page %lu, index %u",
1366 parent->mp_page->mp_pgno, parent->mp_ki);
1368 cursor_pop_page(cursor);
1369 if (move_right ? (parent->mp_ki + 1 >= NUMKEYS(parent->mp_page))
1370 : (parent->mp_ki == 0)) {
1371 DPRINTF("no more keys left, moving to %s sibling",
1372 move_right ? "right" : "left");
1373 if ((rc = mdb_sibling(cursor, move_right)) != MDB_SUCCESS)
1375 parent = CURSOR_TOP(cursor);
1381 DPRINTF("just moving to %s index key %u",
1382 move_right ? "right" : "left", parent->mp_ki);
1384 assert(IS_BRANCH(parent->mp_page));
1386 indx = NODEPTR(parent->mp_page, parent->mp_ki);
1387 if ((mp = mdbenv_get_page(cursor->mc_db->md_env, indx->mn_pgno)) == NULL)
1390 mp->parent = parent->mp_page;
1391 mp->parent_index = parent->mp_ki;
1394 cursor_push_page(cursor, mp);
1400 mdb_set_key(MDB_db *bt, MDB_page *mp, MDB_node *node,
1406 key->mv_size = node->mn_ksize;
1407 key->mv_data = NODEKEY(node);
1413 mdb_cursor_next(MDB_cursor *cursor, MDB_val *key, MDB_val *data)
1419 if (cursor->mc_eof) {
1423 assert(cursor->mc_initialized);
1425 top = CURSOR_TOP(cursor);
1428 DPRINTF("cursor_next: top page is %lu in cursor %p", mp->mp_pgno, cursor);
1430 if (top->mp_ki + 1 >= NUMKEYS(mp)) {
1431 DPRINTF("=====> move to next sibling page");
1432 if (mdb_sibling(cursor, 1) != MDB_SUCCESS) {
1436 top = CURSOR_TOP(cursor);
1438 DPRINTF("next page is %lu, key index %u", mp->mp_pgno, top->mp_ki);
1442 DPRINTF("==> cursor points to page %lu with %lu keys, key index %u",
1443 mp->mp_pgno, NUMKEYS(mp), top->mp_ki);
1445 assert(IS_LEAF(mp));
1446 leaf = NODEPTR(mp, top->mp_ki);
1448 if (data && mdb_read_data(cursor->mc_db, mp, leaf, data) != MDB_SUCCESS)
1451 return mdb_set_key(cursor->mc_db, mp, leaf, key);
1455 mdb_cursor_set(MDB_cursor *cursor, MDB_val *key, MDB_val *data,
1465 assert(key->mv_size > 0);
1467 rc = mdb_search_page(cursor->mc_db, cursor->mc_txn, key, cursor, 0, &mpp);
1468 if (rc != MDB_SUCCESS)
1470 assert(IS_LEAF(mpp.mp_page));
1472 top = CURSOR_TOP(cursor);
1473 leaf = mdb_search_node(cursor->mc_db, mpp.mp_page, key, exactp, &top->mp_ki);
1474 if (exactp != NULL && !*exactp) {
1475 /* MDB_CURSOR_EXACT specified and not an exact match. */
1480 DPRINTF("===> inexact leaf not found, goto sibling");
1481 if ((rc = mdb_sibling(cursor, 1)) != MDB_SUCCESS)
1482 return rc; /* no entries matched */
1483 top = CURSOR_TOP(cursor);
1485 mpp.mp_page = top->mp_page;
1486 assert(IS_LEAF(mpp.mp_page));
1487 leaf = NODEPTR(mpp.mp_page, 0);
1490 cursor->mc_initialized = 1;
1493 if (data && (rc = mdb_read_data(cursor->mc_db, mpp.mp_page, leaf, data)) != MDB_SUCCESS)
1496 rc = mdb_set_key(cursor->mc_db, mpp.mp_page, leaf, key);
1497 if (rc == MDB_SUCCESS) {
1498 DPRINTF("==> cursor placed on key %.*s",
1499 (int)key->mv_size, (char *)key->mv_data);
1507 mdb_cursor_first(MDB_cursor *cursor, MDB_val *key, MDB_val *data)
1513 rc = mdb_search_page(cursor->mc_db, cursor->mc_txn, NULL, cursor, 0, &mpp);
1514 if (rc != MDB_SUCCESS)
1516 assert(IS_LEAF(mpp.mp_page));
1518 leaf = NODEPTR(mpp.mp_page, 0);
1519 cursor->mc_initialized = 1;
1522 if (data && (rc = mdb_read_data(cursor->mc_db, mpp.mp_page, leaf, data)) != MDB_SUCCESS)
1525 return mdb_set_key(cursor->mc_db, mpp.mp_page, leaf, key);
1529 mdb_cursor_get(MDB_cursor *cursor, MDB_val *key, MDB_val *data,
1539 case MDB_CURSOR_EXACT:
1540 while (CURSOR_TOP(cursor) != NULL)
1541 cursor_pop_page(cursor);
1542 if (key == NULL || key->mv_size == 0 || key->mv_size > MAXKEYSIZE) {
1544 } else if (op == MDB_CURSOR_EXACT)
1545 rc = mdb_cursor_set(cursor, key, data, &exact);
1547 rc = mdb_cursor_set(cursor, key, data, NULL);
1550 if (!cursor->mc_initialized)
1551 rc = mdb_cursor_first(cursor, key, data);
1553 rc = mdb_cursor_next(cursor, key, data);
1556 while (CURSOR_TOP(cursor) != NULL)
1557 cursor_pop_page(cursor);
1558 rc = mdb_cursor_first(cursor, key, data);
1561 DPRINTF("unhandled/unimplemented cursor operation %u", op);
1569 /* Allocate a page and initialize it
1572 mdbenv_new_page(MDB_env *env, uint32_t flags, int num)
1576 assert(env != NULL);
1577 assert(env->me_txn != NULL);
1579 DPRINTF("allocating new mpage %lu, page size %u",
1580 env->me_txn->mt_next_pgno, env->me_head.mh_psize);
1581 if ((dp = mdb_newpage(env->me_txn, NULL, 0, num)) == NULL)
1583 dp->p.mp_flags = flags | P_DIRTY;
1584 dp->p.mp_lower = PAGEHDRSZ;
1585 dp->p.mp_upper = env->me_head.mh_psize;
1587 if (IS_BRANCH(&dp->p))
1588 env->me_meta.mm_stat.ms_branch_pages++;
1589 else if (IS_LEAF(&dp->p))
1590 env->me_meta.mm_stat.ms_leaf_pages++;
1591 else if (IS_OVERFLOW(&dp->p)) {
1592 env->me_meta.mm_stat.ms_overflow_pages += num;
1593 dp->p.mp_pages = num;
1600 mdb_leaf_size(MDB_db *db, MDB_val *key, MDB_val *data)
1604 sz = LEAFSIZE(key, data);
1605 if (data->mv_size >= db->md_env->me_head.mh_psize / MDB_MINKEYS) {
1606 /* put on overflow page */
1607 sz -= data->mv_size - sizeof(pgno_t);
1610 return sz + sizeof(indx_t);
1614 mdb_branch_size(MDB_db *db, MDB_val *key)
1619 if (sz >= db->md_env->me_head.mh_psize / MDB_MINKEYS) {
1620 /* put on overflow page */
1621 /* not implemented */
1622 /* sz -= key->size - sizeof(pgno_t); */
1625 return sz + sizeof(indx_t);
1629 mdb_add_node(MDB_db *db, MDB_page *mp, indx_t indx,
1630 MDB_val *key, MDB_val *data, pgno_t pgno, uint8_t flags)
1633 size_t node_size = NODESIZE;
1636 MDB_dpage *ofp = NULL; /* overflow page */
1638 assert(mp->mp_upper >= mp->mp_lower);
1640 DPRINTF("add node [%.*s] to %s page %lu at index %i, key size %zu",
1641 key ? (int)key->mv_size : 0, key ? (char *)key->mv_data : NULL,
1642 IS_LEAF(mp) ? "leaf" : "branch",
1643 mp->mp_pgno, indx, key ? key->mv_size : 0);
1646 node_size += key->mv_size;
1650 node_size += data->mv_size;
1651 if (F_ISSET(flags, F_BIGDATA)) {
1652 /* Data already on overflow page. */
1653 node_size -= data->mv_size - sizeof(pgno_t);
1654 } else if (data->mv_size >= db->md_env->me_head.mh_psize / MDB_MINKEYS) {
1655 int ovpages = PAGEHDRSZ + data->mv_size + db->md_env->me_head.mh_psize - 1;
1656 ovpages /= db->md_env->me_head.mh_psize;
1657 /* Put data on overflow page. */
1658 DPRINTF("data size is %zu, put on overflow page",
1660 node_size -= data->mv_size - sizeof(pgno_t);
1661 if ((ofp = mdbenv_new_page(db->md_env, P_OVERFLOW, ovpages)) == NULL)
1663 DPRINTF("allocated overflow page %lu", ofp->p.mp_pgno);
1668 if (node_size + sizeof(indx_t) > SIZELEFT(mp)) {
1669 DPRINTF("not enough room in page %lu, got %lu ptrs",
1670 mp->mp_pgno, NUMKEYS(mp));
1671 DPRINTF("upper - lower = %u - %u = %u", mp->mp_upper, mp->mp_lower,
1672 mp->mp_upper - mp->mp_lower);
1673 DPRINTF("node size = %zu", node_size);
1677 /* Move higher pointers up one slot. */
1678 for (i = NUMKEYS(mp); i > indx; i--)
1679 mp->mp_ptrs[i] = mp->mp_ptrs[i - 1];
1681 /* Adjust free space offsets. */
1682 ofs = mp->mp_upper - node_size;
1683 assert(ofs >= mp->mp_lower + sizeof(indx_t));
1684 mp->mp_ptrs[indx] = ofs;
1686 mp->mp_lower += sizeof(indx_t);
1688 /* Write the node data. */
1689 node = NODEPTR(mp, indx);
1690 node->mn_ksize = (key == NULL) ? 0 : key->mv_size;
1691 node->mn_flags = flags;
1693 node->mn_dsize = data->mv_size;
1695 node->mn_pgno = pgno;
1698 bcopy(key->mv_data, NODEKEY(node), key->mv_size);
1703 if (F_ISSET(flags, F_BIGDATA))
1704 bcopy(data->mv_data, node->mn_data + key->mv_size,
1707 bcopy(data->mv_data, node->mn_data + key->mv_size,
1710 bcopy(&ofp->p.mp_pgno, node->mn_data + key->mv_size,
1712 bcopy(data->mv_data, METADATA(&ofp->p), data->mv_size);
1720 mdb_del_node(MDB_db *db, MDB_page *mp, indx_t indx)
1723 indx_t i, j, numkeys, ptr;
1727 DPRINTF("delete node %u on %s page %lu", indx,
1728 IS_LEAF(mp) ? "leaf" : "branch", mp->mp_pgno);
1729 assert(indx < NUMKEYS(mp));
1731 node = NODEPTR(mp, indx);
1732 sz = NODESIZE + node->mn_ksize;
1734 if (F_ISSET(node->mn_flags, F_BIGDATA))
1735 sz += sizeof(pgno_t);
1737 sz += NODEDSZ(node);
1740 ptr = mp->mp_ptrs[indx];
1741 numkeys = NUMKEYS(mp);
1742 for (i = j = 0; i < numkeys; i++) {
1744 mp->mp_ptrs[j] = mp->mp_ptrs[i];
1745 if (mp->mp_ptrs[i] < ptr)
1746 mp->mp_ptrs[j] += sz;
1751 base = (char *)mp + mp->mp_upper;
1752 bcopy(base, base + sz, ptr - mp->mp_upper);
1754 mp->mp_lower -= sizeof(indx_t);
1759 mdb_cursor_open(MDB_db *db, MDB_txn *txn, MDB_cursor **ret)
1763 if (db == NULL || ret == NULL)
1766 if (txn != NULL && db->md_env != txn->mt_env) {
1770 if ((cursor = calloc(1, sizeof(*cursor))) != NULL) {
1771 SLIST_INIT(&cursor->mc_stack);
1773 cursor->mc_txn = txn;
1782 mdb_cursor_close(MDB_cursor *cursor)
1784 if (cursor != NULL) {
1785 while (!CURSOR_EMPTY(cursor))
1786 cursor_pop_page(cursor);
1788 /* btree_close(cursor->bt); */
1794 mdb_update_key(MDB_db *db, MDB_page *mp, indx_t indx,
1797 indx_t ptr, i, numkeys;
1803 node = NODEPTR(mp, indx);
1804 ptr = mp->mp_ptrs[indx];
1805 DPRINTF("update key %u (ofs %u) [%.*s] to [%.*s] on page %lu",
1807 (int)node->mn_ksize, (char *)NODEKEY(node),
1808 (int)key->mv_size, (char *)key->mv_data,
1811 delta = key->mv_size - node->mn_ksize;
1813 if (delta > 0 && SIZELEFT(mp) < delta) {
1814 DPRINTF("OUCH! Not enough room, delta = %d", delta);
1818 numkeys = NUMKEYS(mp);
1819 for (i = 0; i < numkeys; i++) {
1820 if (mp->mp_ptrs[i] <= ptr)
1821 mp->mp_ptrs[i] -= delta;
1824 base = (char *)mp + mp->mp_upper;
1825 len = ptr - mp->mp_upper + NODESIZE;
1826 bcopy(base, base - delta, len);
1827 mp->mp_upper -= delta;
1829 node = NODEPTR(mp, indx);
1830 node->mn_ksize = key->mv_size;
1833 bcopy(key->mv_data, NODEKEY(node), key->mv_size);
1838 /* Move a node from src to dst.
1841 mdb_move_node(MDB_db *bt, MDB_pageparent *src, indx_t srcindx,
1842 MDB_pageparent *dst, indx_t dstindx)
1848 srcnode = NODEPTR(src->mp_page, srcindx);
1849 DPRINTF("moving %s node %u [%.*s] on page %lu to node %u on page %lu",
1850 IS_LEAF(src->mp_page) ? "leaf" : "branch",
1852 (int)srcnode->mn_ksize, (char *)NODEKEY(srcnode),
1853 src->mp_page->mp_pgno,
1854 dstindx, dst->mp_page->mp_pgno);
1856 /* Mark src and dst as dirty. */
1857 if ((rc = mdb_touch(bt->md_env->me_txn, src)) ||
1858 (rc = mdb_touch(bt->md_env->me_txn, dst)))
1861 /* Add the node to the destination page.
1863 key.mv_size = srcnode->mn_ksize;
1864 key.mv_data = NODEKEY(srcnode);
1865 data.mv_size = NODEDSZ(srcnode);
1866 data.mv_data = NODEDATA(srcnode);
1867 rc = mdb_add_node(bt, dst->mp_page, dstindx, &key, &data, NODEPGNO(srcnode),
1869 if (rc != MDB_SUCCESS)
1872 /* Delete the node from the source page.
1874 mdb_del_node(bt, src->mp_page, srcindx);
1876 /* Update the parent separators.
1878 if (srcindx == 0 && src->mp_pi != 0) {
1879 DPRINTF("update separator for source page %lu to [%.*s]",
1880 src->mp_page->mp_pgno, (int)key.mv_size, (char *)key.mv_data);
1881 if ((rc = mdb_update_key(bt, src->mp_parent, src->mp_pi,
1882 &key)) != MDB_SUCCESS)
1886 if (srcindx == 0 && IS_BRANCH(src->mp_page)) {
1888 nullkey.mv_size = 0;
1889 assert(mdb_update_key(bt, src->mp_page, 0, &nullkey) == MDB_SUCCESS);
1892 if (dstindx == 0 && dst->mp_pi != 0) {
1893 DPRINTF("update separator for destination page %lu to [%.*s]",
1894 dst->mp_page->mp_pgno, (int)key.mv_size, (char *)key.mv_data);
1895 if ((rc = mdb_update_key(bt, dst->mp_parent, dst->mp_pi,
1896 &key)) != MDB_SUCCESS)
1900 if (dstindx == 0 && IS_BRANCH(dst->mp_page)) {
1902 nullkey.mv_size = 0;
1903 assert(mdb_update_key(bt, dst->mp_page, 0, &nullkey) == MDB_SUCCESS);
1910 mdb_merge(MDB_db *bt, MDB_pageparent *src, MDB_pageparent *dst)
1919 DPRINTF("merging page %lu and %lu", src->mp_page->mp_pgno, dst->mp_page->mp_pgno);
1921 assert(src->mp_parent); /* can't merge root page */
1922 assert(dst->mp_parent);
1923 assert(bt->md_env->me_txn != NULL);
1925 /* Mark src and dst as dirty. */
1926 if ((rc = mdb_touch(bt->md_env->me_txn, src)) ||
1927 (rc = mdb_touch(bt->md_env->me_txn, dst)))
1930 /* Move all nodes from src to dst.
1932 for (i = 0; i < NUMKEYS(src->mp_page); i++) {
1933 srcnode = NODEPTR(src->mp_page, i);
1935 key.mv_size = srcnode->mn_ksize;
1936 key.mv_data = NODEKEY(srcnode);
1937 data.mv_size = NODEDSZ(srcnode);
1938 data.mv_data = NODEDATA(srcnode);
1939 rc = mdb_add_node(bt, dst->mp_page, NUMKEYS(dst->mp_page), &key,
1940 &data, NODEPGNO(srcnode), srcnode->mn_flags);
1941 if (rc != MDB_SUCCESS)
1945 DPRINTF("dst page %lu now has %lu keys (%.1f%% filled)",
1946 dst->mp_page->mp_pgno, NUMKEYS(dst->mp_page), (float)PAGEFILL(bt->md_env, dst->mp_page) / 10);
1948 /* Unlink the src page from parent.
1950 mdb_del_node(bt, src->mp_parent, src->mp_pi);
1951 if (src->mp_pi == 0) {
1953 if ((rc = mdb_update_key(bt, src->mp_parent, 0, &key)) != MDB_SUCCESS)
1957 if (IS_LEAF(src->mp_page))
1958 bt->md_env->me_meta.mm_stat.ms_leaf_pages--;
1960 bt->md_env->me_meta.mm_stat.ms_branch_pages--;
1962 mpp.mp_page = src->mp_parent;
1963 dh = (MDB_dhead *)src->mp_parent;
1965 mpp.mp_parent = dh->md_parent;
1966 mpp.mp_pi = dh->md_pi;
1968 return mdb_rebalance(bt, &mpp);
1971 #define FILL_THRESHOLD 250
1974 mdb_rebalance(MDB_db *db, MDB_pageparent *mpp)
1979 indx_t si = 0, di = 0;
1982 assert(db->md_env->me_txn != NULL);
1983 assert(mpp != NULL);
1985 DPRINTF("rebalancing %s page %lu (has %lu keys, %.1f%% full)",
1986 IS_LEAF(mpp->mp_page) ? "leaf" : "branch",
1987 mpp->mp_page->mp_pgno, NUMKEYS(mpp->mp_page), (float)PAGEFILL(db->md_env, mpp->mp_page) / 10);
1989 if (PAGEFILL(db->md_env, mpp->mp_page) >= FILL_THRESHOLD) {
1990 DPRINTF("no need to rebalance page %lu, above fill threshold",
1991 mpp->mp_page->mp_pgno);
1995 if (mpp->mp_parent == NULL) {
1996 if (NUMKEYS(mpp->mp_page) == 0) {
1997 DPRINTF("tree is completely empty");
1998 db->md_env->me_txn->mt_root = P_INVALID;
1999 db->md_env->me_meta.mm_stat.ms_depth--;
2000 db->md_env->me_meta.mm_stat.ms_leaf_pages--;
2001 } else if (IS_BRANCH(mpp->mp_page) && NUMKEYS(mpp->mp_page) == 1) {
2002 DPRINTF("collapsing root page!");
2003 db->md_env->me_txn->mt_root = NODEPGNO(NODEPTR(mpp->mp_page, 0));
2004 if ((root = mdbenv_get_page(db->md_env, db->md_env->me_txn->mt_root)) == NULL)
2006 db->md_env->me_meta.mm_stat.ms_depth--;
2007 db->md_env->me_meta.mm_stat.ms_branch_pages--;
2009 DPRINTF("root page doesn't need rebalancing");
2013 /* The parent (branch page) must have at least 2 pointers,
2014 * otherwise the tree is invalid.
2016 assert(NUMKEYS(mpp->mp_parent) > 1);
2018 /* Leaf page fill factor is below the threshold.
2019 * Try to move keys from left or right neighbor, or
2020 * merge with a neighbor page.
2025 if (mpp->mp_pi == 0) {
2026 /* We're the leftmost leaf in our parent.
2028 DPRINTF("reading right neighbor");
2029 node = NODEPTR(mpp->mp_parent, mpp->mp_pi + 1);
2030 if ((npp.mp_page = mdbenv_get_page(db->md_env, NODEPGNO(node))) == NULL)
2032 npp.mp_pi = mpp->mp_pi + 1;
2034 di = NUMKEYS(mpp->mp_page);
2036 /* There is at least one neighbor to the left.
2038 DPRINTF("reading left neighbor");
2039 node = NODEPTR(mpp->mp_parent, mpp->mp_pi - 1);
2040 if ((npp.mp_page = mdbenv_get_page(db->md_env, NODEPGNO(node))) == NULL)
2042 npp.mp_pi = mpp->mp_pi - 1;
2043 si = NUMKEYS(npp.mp_page) - 1;
2046 npp.mp_parent = mpp->mp_parent;
2048 DPRINTF("found neighbor page %lu (%lu keys, %.1f%% full)",
2049 npp.mp_page->mp_pgno, NUMKEYS(npp.mp_page), (float)PAGEFILL(db->md_env, npp.mp_page) / 10);
2051 /* If the neighbor page is above threshold and has at least two
2052 * keys, move one key from it.
2054 * Otherwise we should try to merge them.
2056 if (PAGEFILL(db->md_env, npp.mp_page) >= FILL_THRESHOLD && NUMKEYS(npp.mp_page) >= 2)
2057 return mdb_move_node(db, &npp, si, mpp, di);
2058 else { /* FIXME: if (has_enough_room()) */
2059 if (mpp->mp_pi == 0)
2060 return mdb_merge(db, &npp, mpp);
2062 return mdb_merge(db, mpp, &npp);
2067 mdb_del(MDB_db *bt, MDB_txn *txn,
2068 MDB_val *key, MDB_val *data)
2075 DPRINTF("========> delete key %.*s", (int)key->mv_size, (char *)key->mv_data);
2077 assert(key != NULL);
2079 if (bt == NULL || txn == NULL)
2082 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
2086 if (bt->md_env->me_txn != txn) {
2090 if (key->mv_size == 0 || key->mv_size > MAXKEYSIZE) {
2094 if ((rc = mdb_search_page(bt, txn, key, NULL, 1, &mpp)) != MDB_SUCCESS)
2097 leaf = mdb_search_node(bt, mpp.mp_page, key, &exact, &ki);
2098 if (leaf == NULL || !exact) {
2102 if (data && (rc = mdb_read_data(bt, NULL, leaf, data)) != MDB_SUCCESS)
2105 mdb_del_node(bt, mpp.mp_page, ki);
2106 bt->md_env->me_meta.mm_stat.ms_entries--;
2107 rc = mdb_rebalance(bt, &mpp);
2108 if (rc != MDB_SUCCESS)
2109 txn->mt_flags |= MDB_TXN_ERROR;
2114 /* Split page <*mpp>, and insert <key,(data|newpgno)> in either left or
2115 * right sibling, at index <*newindxp> (as if unsplit). Updates *mpp and
2116 * *newindxp with the actual values after split, ie if *mpp and *newindxp
2117 * refer to a node in the new right sibling page.
2120 mdb_split(MDB_db *bt, MDB_page **mpp, unsigned int *newindxp,
2121 MDB_val *newkey, MDB_val *newdata, pgno_t newpgno)
2124 int rc = MDB_SUCCESS, ins_new = 0;
2127 unsigned int i, j, split_indx;
2129 MDB_page *pright, *p;
2130 MDB_val sepkey, rkey, rdata;
2132 MDB_dpage *mdp, *rdp, *pdp;
2136 assert(bt->md_env != NULL);
2138 dh = ((MDB_dhead *)*mpp) - 1;
2139 mdp = (MDB_dpage *)dh;
2140 newindx = *newindxp;
2142 DPRINTF("-----> splitting %s page %lu and adding [%.*s] at index %i",
2143 IS_LEAF(&mdp->p) ? "leaf" : "branch", mdp->p.mp_pgno,
2144 (int)newkey->mv_size, (char *)newkey->mv_data, *newindxp);
2146 if (mdp->h.md_parent == NULL) {
2147 if ((pdp = mdbenv_new_page(bt->md_env, P_BRANCH, 1)) == NULL)
2150 mdp->h.md_parent = &pdp->p;
2151 bt->md_env->me_txn->mt_root = pdp->p.mp_pgno;
2152 DPRINTF("root split! new root = %lu", pdp->p.mp_pgno);
2153 bt->md_env->me_meta.mm_stat.ms_depth++;
2155 /* Add left (implicit) pointer. */
2156 if (mdb_add_node(bt, &pdp->p, 0, NULL, NULL,
2157 mdp->p.mp_pgno, 0) != MDB_SUCCESS)
2160 DPRINTF("parent branch page is %lu", mdp->h.md_parent->mp_pgno);
2163 /* Create a right sibling. */
2164 if ((rdp = mdbenv_new_page(bt->md_env, mdp->p.mp_flags, 1)) == NULL)
2166 rdp->h.md_parent = mdp->h.md_parent;
2167 rdp->h.md_pi = mdp->h.md_pi + 1;
2168 DPRINTF("new right sibling: page %lu", rdp->p.mp_pgno);
2170 /* Move half of the keys to the right sibling. */
2171 if ((copy = malloc(bt->md_env->me_head.mh_psize)) == NULL)
2173 bcopy(&mdp->p, copy, bt->md_env->me_head.mh_psize);
2174 bzero(&mdp->p.mp_ptrs, bt->md_env->me_head.mh_psize - PAGEHDRSZ);
2175 mdp->p.mp_lower = PAGEHDRSZ;
2176 mdp->p.mp_upper = bt->md_env->me_head.mh_psize;
2178 split_indx = NUMKEYS(copy) / 2 + 1;
2180 /* First find the separating key between the split pages.
2182 bzero(&sepkey, sizeof(sepkey));
2183 if (newindx == split_indx) {
2184 sepkey.mv_size = newkey->mv_size;
2185 sepkey.mv_data = newkey->mv_data;
2187 node = NODEPTR(copy, split_indx);
2188 sepkey.mv_size = node->mn_ksize;
2189 sepkey.mv_data = NODEKEY(node);
2192 DPRINTF("separator is [%.*s]", (int)sepkey.mv_size, (char *)sepkey.mv_data);
2194 /* Copy separator key to the parent.
2196 if (SIZELEFT(rdp->h.md_parent) < mdb_branch_size(bt, &sepkey)) {
2197 rc = mdb_split(bt, &rdp->h.md_parent, &rdp->h.md_pi,
2198 &sepkey, NULL, rdp->p.mp_pgno);
2200 /* Right page might now have changed parent.
2201 * Check if left page also changed parent.
2203 if (rdp->h.md_parent != mdp->h.md_parent &&
2204 mdp->h.md_pi >= NUMKEYS(mdp->h.md_parent)) {
2205 mdp->h.md_parent = rdp->h.md_parent;
2206 mdp->h.md_pi = rdp->h.md_pi - 1;
2209 rc = mdb_add_node(bt, rdp->h.md_parent, rdp->h.md_pi,
2210 &sepkey, NULL, rdp->p.mp_pgno, 0);
2212 if (rc != MDB_SUCCESS) {
2217 for (i = j = 0; i <= NUMKEYS(copy); j++) {
2218 if (i < split_indx) {
2219 /* Re-insert in left sibling. */
2222 /* Insert in right sibling. */
2223 if (i == split_indx)
2224 /* Reset insert index for right sibling. */
2225 j = (i == newindx && ins_new);
2229 if (i == newindx && !ins_new) {
2230 /* Insert the original entry that caused the split. */
2231 rkey.mv_data = newkey->mv_data;
2232 rkey.mv_size = newkey->mv_size;
2233 if (IS_LEAF(&mdp->p)) {
2234 rdata.mv_data = newdata->mv_data;
2235 rdata.mv_size = newdata->mv_size;
2242 /* Update page and index for the new key. */
2245 } else if (i == NUMKEYS(copy)) {
2248 node = NODEPTR(copy, i);
2249 rkey.mv_data = NODEKEY(node);
2250 rkey.mv_size = node->mn_ksize;
2251 if (IS_LEAF(&mdp->p)) {
2252 rdata.mv_data = NODEDATA(node);
2253 rdata.mv_size = node->mn_dsize;
2255 pgno = node->mn_pgno;
2256 flags = node->mn_flags;
2261 if (!IS_LEAF(&mdp->p) && j == 0) {
2262 /* First branch index doesn't need key data. */
2266 rc = mdb_add_node(bt, &pdp->p, j, &rkey, &rdata, pgno,flags);
2274 mdb_put(MDB_db *bt, MDB_txn *txn,
2275 MDB_val *key, MDB_val *data, unsigned int flags)
2277 int rc = MDB_SUCCESS, exact;
2282 assert(key != NULL);
2283 assert(data != NULL);
2285 if (bt == NULL || txn == NULL)
2288 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
2292 if (bt->md_env->me_txn != txn) {
2296 if (key->mv_size == 0 || key->mv_size > MAXKEYSIZE) {
2300 DPRINTF("==> put key %.*s, size %zu, data size %zu",
2301 (int)key->mv_size, (char *)key->mv_data, key->mv_size, data->mv_size);
2303 rc = mdb_search_page(bt, txn, key, NULL, 1, &mpp);
2304 if (rc == MDB_SUCCESS) {
2305 leaf = mdb_search_node(bt, mpp.mp_page, key, &exact, &ki);
2306 if (leaf && exact) {
2307 if (F_ISSET(flags, MDB_NOOVERWRITE)) {
2308 DPRINTF("duplicate key %.*s",
2309 (int)key->mv_size, (char *)key->mv_data);
2312 mdb_del_node(bt, mpp.mp_page, ki);
2314 if (leaf == NULL) { /* append if not found */
2315 ki = NUMKEYS(mpp.mp_page);
2316 DPRINTF("appending key at index %i", ki);
2318 } else if (rc == ENOENT) {
2320 /* new file, just write a root leaf page */
2321 DPRINTF("allocating new root leaf page");
2322 if ((dp = mdbenv_new_page(bt->md_env, P_LEAF, 1)) == NULL) {
2325 mpp.mp_page = &dp->p;
2326 txn->mt_root = mpp.mp_page->mp_pgno;
2327 bt->md_env->me_meta.mm_stat.ms_depth++;
2333 assert(IS_LEAF(mpp.mp_page));
2334 DPRINTF("there are %lu keys, should insert new key at index %i",
2335 NUMKEYS(mpp.mp_page), ki);
2337 if (SIZELEFT(mpp.mp_page) < mdb_leaf_size(bt, key, data)) {
2338 rc = mdb_split(bt, &mpp.mp_page, &ki, key, data, P_INVALID);
2340 /* There is room already in this leaf page. */
2341 rc = mdb_add_node(bt, mpp.mp_page, ki, key, data, 0, 0);
2344 if (rc != MDB_SUCCESS)
2345 txn->mt_flags |= MDB_TXN_ERROR;
2347 bt->md_env->me_meta.mm_stat.ms_entries++;
2354 mdbenv_get_flags(MDB_env *env, unsigned int *arg)
2359 *arg = env->me_flags;
2364 mdbenv_get_path(MDB_env *env, const char **arg)
2369 *arg = env->me_path;
2374 mdbenv_stat(MDB_env *env, MDB_stat **arg)
2376 if (env == NULL || arg == NULL)
2379 *arg = &env->me_meta.mm_stat;
2384 int mdb_open(MDB_env *env, MDB_txn *txn, const char *name, unsigned int flags, MDB_db **db)
2387 *db = (MDB_db *)&env->me_db;
2393 int mdb_stat(MDB_db *db, MDB_stat **arg)
2395 if (db == NULL || arg == NULL)
2398 *arg = &db->md_stat;
2403 void mdb_close(MDB_db *db)