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 %lu on mdbenv %p, root page %lu", txn->mt_txnid, env, txn->mt_root);
514 mdb_txn_abort(MDB_txn *txn)
523 DPRINTF("abort transaction %lu on mdbenv %p, root page %lu", txn->mt_txnid, 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 %lu on mdbenv %p, root page %lu",
590 txn->mt_txnid, env, txn->mt_root);
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 if (meta[toggle]->mm_txnid > env->me_meta.mm_txnid) {
847 bcopy(meta[toggle], &env->me_meta, sizeof(env->me_meta));
848 env->me_metatoggle = toggle;
851 DPRINTF("Using meta page %d", toggle);
857 mdbenv_create(MDB_env **env)
861 e = calloc(1, sizeof(*e));
862 if (!e) return ENOMEM;
864 e->me_head.mh_magic = MDB_MAGIC;
865 e->me_head.mh_version = MDB_VERSION;
866 e->me_head.mh_mapsize = DEFAULT_MAPSIZE;
867 e->me_maxreaders = DEFAULT_READERS;
874 mdbenv_set_mapsize(MDB_env *env, size_t size)
878 env->me_mapsize = env->me_head.mh_mapsize = size;
883 mdbenv_set_maxreaders(MDB_env *env, int readers)
885 env->me_maxreaders = readers;
890 mdbenv_get_maxreaders(MDB_env *env, int *readers)
892 if (!env || !readers)
894 *readers = env->me_maxreaders;
899 mdbenv_open2(MDB_env *env, unsigned int flags)
903 env->me_flags = flags;
904 env->me_meta.mm_root = P_INVALID;
906 if ((i = mdbenv_read_header(env)) != 0) {
909 DPRINTF("new mdbenv");
913 if (!env->me_mapsize)
914 env->me_mapsize = env->me_head.mh_mapsize;
917 if (env->me_head.mh_address && (flags & MDB_FIXEDMAP))
919 env->me_map = mmap(env->me_head.mh_address, env->me_mapsize, PROT_READ, i,
921 if (env->me_map == MAP_FAILED)
925 env->me_head.mh_mapsize = env->me_mapsize;
926 if (flags & MDB_FIXEDMAP)
927 env->me_head.mh_address = env->me_map;
928 i = mdbenv_write_header(env);
929 if (i != MDB_SUCCESS) {
930 munmap(env->me_map, env->me_mapsize);
933 i = mdbenv_init_meta(env);
934 if (i != MDB_SUCCESS) {
935 munmap(env->me_map, env->me_mapsize);
940 if ((i = mdbenv_read_meta(env)) != 0)
943 DPRINTF("opened database version %u, pagesize %u",
944 env->me_head.mh_version, env->me_head.mh_psize);
945 DPRINTF("depth: %u", env->me_meta.mm_stat.ms_depth);
946 DPRINTF("entries: %lu", env->me_meta.mm_stat.ms_entries);
947 DPRINTF("branch pages: %lu", env->me_meta.mm_stat.ms_branch_pages);
948 DPRINTF("leaf pages: %lu", env->me_meta.mm_stat.ms_leaf_pages);
949 DPRINTF("overflow pages: %lu", env->me_meta.mm_stat.ms_overflow_pages);
950 DPRINTF("root: %lu", env->me_meta.mm_root);
956 mdbenv_reader_dest(void *ptr)
958 MDB_reader *reader = ptr;
960 reader->mr_txnid = 0;
966 mdbenv_open(MDB_env *env, const char *path, unsigned int flags, mode_t mode)
968 int oflags, rc, shmid;
972 if (F_ISSET(flags, MDB_RDONLY))
975 oflags = O_RDWR | O_CREAT;
977 if ((env->me_fd = open(path, oflags, mode)) == -1)
980 env->me_shmkey = ftok(path, 'm');
981 size = (env->me_maxreaders-1) * sizeof(MDB_reader) + sizeof(MDB_txninfo);
982 shmid = shmget(env->me_shmkey, size, IPC_CREAT|IPC_EXCL|mode);
984 if (errno == EEXIST) {
985 shmid = shmget(env->me_shmkey, size, IPC_CREAT|mode);
988 env->me_txns = shmat(shmid, NULL, 0);
989 if (env->me_txns->mt_magic != MDB_MAGIC ||
990 env->me_txns->mt_version != MDB_VERSION) {
991 DPRINTF("invalid lock region %d", shmid);
1000 pthread_mutexattr_t mattr;
1002 env->me_txns = shmat(shmid, NULL, 0);
1003 pthread_mutexattr_init(&mattr);
1004 pthread_mutexattr_setpshared(&mattr, PTHREAD_PROCESS_SHARED);
1005 pthread_mutex_init(&env->me_txns->mt_mutex, &mattr);
1006 pthread_mutex_init(&env->me_txns->mt_wmutex, &mattr);
1007 env->me_txns->mt_version = MDB_VERSION;
1008 env->me_txns->mt_magic = MDB_MAGIC;
1011 if ((rc = mdbenv_open2(env, flags)) != MDB_SUCCESS) {
1015 env->me_path = strdup(path);
1016 DPRINTF("opened dbenv %p", env);
1019 pthread_key_create(&env->me_txkey, mdbenv_reader_dest);
1025 mdbenv_close(MDB_env *env)
1033 munmap(env->me_map, env->me_mapsize);
1037 shmdt(env->me_txns);
1042 /* Search for key within a leaf page, using binary search.
1043 * Returns the smallest entry larger or equal to the key.
1044 * If exactp is non-null, stores whether the found entry was an exact match
1045 * in *exactp (1 or 0).
1046 * If kip is non-null, stores the index of the found entry in *kip.
1047 * If no entry larger or equal to the key is found, returns NULL.
1050 mdb_search_node(MDB_db *bt, MDB_page *mp, MDB_val *key,
1051 int *exactp, unsigned int *kip)
1059 DPRINTF("searching %lu keys in %s page %lu",
1061 IS_LEAF(mp) ? "leaf" : "branch",
1064 assert(NUMKEYS(mp) > 0);
1066 bzero(&nodekey, sizeof(nodekey));
1068 low = IS_LEAF(mp) ? 0 : 1;
1069 high = NUMKEYS(mp) - 1;
1070 while (low <= high) {
1071 i = (low + high) >> 1;
1072 node = NODEPTR(mp, i);
1074 nodekey.mv_size = node->mn_ksize;
1075 nodekey.mv_data = NODEKEY(node);
1078 rc = bt->md_cmp(key, &nodekey);
1080 rc = _mdb_cmp(bt, key, &nodekey);
1083 DPRINTF("found leaf index %u [%.*s], rc = %i",
1084 i, (int)nodekey.mv_size, (char *)nodekey.mv_data, rc);
1086 DPRINTF("found branch index %u [%.*s -> %lu], rc = %i",
1087 i, (int)node->mn_ksize, (char *)NODEKEY(node),
1098 if (rc > 0) { /* Found entry is less than the key. */
1099 i++; /* Skip to get the smallest entry larger than key. */
1100 if (i >= NUMKEYS(mp))
1101 /* There is no entry larger or equal to the key. */
1105 *exactp = (rc == 0);
1106 if (kip) /* Store the key index if requested. */
1109 return NODEPTR(mp, i);
1113 cursor_pop_page(MDB_cursor *cursor)
1117 top = CURSOR_TOP(cursor);
1120 DPRINTF("popped page %lu off cursor %p", top->mp_page->mp_pgno, cursor);
1126 cursor_push_page(MDB_cursor *cursor, MDB_page *mp)
1130 DPRINTF("pushing page %lu on cursor %p", mp->mp_pgno, cursor);
1132 if ((ppage = calloc(1, sizeof(*ppage))) == NULL)
1134 ppage->mp_page = mp;
1135 CURSOR_PUSH(cursor, ppage);
1140 mdbenv_get_page(MDB_env *env, pgno_t pgno)
1143 MDB_txn *txn = env->me_txn;
1145 if (txn && pgno >= txn->mt_first_pgno &&
1146 !SIMPLEQ_EMPTY(txn->mt_u.dirty_queue)) {
1148 SIMPLEQ_FOREACH(dp, txn->mt_u.dirty_queue, h.md_next) {
1149 if (dp->p.mp_pgno == pgno) {
1155 p = (MDB_page *)(env->me_map + env->me_head.mh_psize * pgno);
1161 mdb_search_page_root(MDB_db *bt, MDB_val *key,
1162 MDB_cursor *cursor, int modify, MDB_pageparent *mpp)
1164 MDB_page *mp = mpp->mp_page;
1167 if (cursor && cursor_push_page(cursor, mp) == NULL)
1170 while (IS_BRANCH(mp)) {
1174 DPRINTF("branch page %lu has %lu keys", mp->mp_pgno, NUMKEYS(mp));
1175 assert(NUMKEYS(mp) > 1);
1176 DPRINTF("found index 0 to page %lu", NODEPGNO(NODEPTR(mp, 0)));
1178 if (key == NULL) /* Initialize cursor to first page. */
1182 node = mdb_search_node(bt, mp, key, &exact, &i);
1184 i = NUMKEYS(mp) - 1;
1192 DPRINTF("following index %u for key %.*s",
1193 i, (int)key->mv_size, (char *)key->mv_data);
1194 assert(i >= 0 && i < NUMKEYS(mp));
1195 node = NODEPTR(mp, i);
1198 CURSOR_TOP(cursor)->mp_ki = i;
1200 mpp->mp_parent = mp;
1201 if ((mp = mdbenv_get_page(bt->md_env, NODEPGNO(node))) == NULL)
1206 if (cursor && cursor_push_page(cursor, mp) == NULL)
1210 MDB_dhead *dh = ((MDB_dhead *)mp)-1;
1211 if (rc = mdb_touch(bt->md_env->me_txn, mpp))
1213 dh = ((MDB_dhead *)mpp->mp_page)-1;
1214 dh->md_parent = mpp->mp_parent;
1215 dh->md_pi = mpp->mp_pi;
1222 DPRINTF("internal error, index points to a %02X page!?",
1227 DPRINTF("found leaf page %lu for key %.*s", mp->mp_pgno,
1228 key ? (int)key->mv_size : 0, key ? (char *)key->mv_data : NULL);
1233 /* Search for the page a given key should be in.
1234 * Stores a pointer to the found page in *mpp.
1235 * If key is NULL, search for the lowest page (used by mdb_cursor_first).
1236 * If cursor is non-null, pushes parent pages on the cursor stack.
1237 * If modify is true, visited pages are updated with new page numbers.
1240 mdb_search_page(MDB_db *db, MDB_txn *txn, MDB_val *key,
1241 MDB_cursor *cursor, int modify, MDB_pageparent *mpp)
1246 /* Can't modify pages outside a transaction. */
1247 if (txn == NULL && modify) {
1251 /* Choose which root page to start with. If a transaction is given
1252 * use the root page from the transaction, otherwise read the last
1253 * committed root page.
1256 if ((rc = mdbenv_read_meta(db->md_env)) != MDB_SUCCESS)
1258 root = db->md_env->me_meta.mm_root;
1259 } else if (F_ISSET(txn->mt_flags, MDB_TXN_ERROR)) {
1260 DPRINTF("transaction has failed, must abort");
1263 root = txn->mt_root;
1265 if (root == P_INVALID) { /* Tree is empty. */
1266 DPRINTF("tree is empty");
1270 if ((mpp->mp_page = mdbenv_get_page(db->md_env, root)) == NULL)
1273 DPRINTF("root page has flags 0x%X", mpp->mp_page->mp_flags);
1275 if (modify && !F_ISSET(mpp->mp_page->mp_flags, P_DIRTY)) {
1276 mpp->mp_parent = NULL;
1278 if ((rc = mdb_touch(txn, mpp)))
1280 txn->mt_root = mpp->mp_page->mp_pgno;
1283 return mdb_search_page_root(db, key, cursor, modify, mpp);
1287 mdb_read_data(MDB_db *db, MDB_page *mp, MDB_node *leaf,
1290 MDB_page *omp; /* overflow mpage */
1296 bzero(data, sizeof(*data));
1297 max = db->md_env->me_head.mh_psize - PAGEHDRSZ;
1299 if (!F_ISSET(leaf->mn_flags, F_BIGDATA)) {
1300 data->mv_size = leaf->mn_dsize;
1301 data->mv_data = NODEDATA(leaf);
1305 /* Read overflow data.
1307 data->mv_size = leaf->mn_dsize;
1308 bcopy(NODEDATA(leaf), &pgno, sizeof(pgno));
1309 if ((omp = mdbenv_get_page(db->md_env, pgno)) == NULL) {
1310 DPRINTF("read overflow page %lu failed", pgno);
1313 data->mv_data = omp;
1319 mdb_get(MDB_db *db, MDB_txn *txn,
1320 MDB_val *key, MDB_val *data)
1328 DPRINTF("===> get key [%.*s]", (int)key->mv_size, (char *)key->mv_data);
1333 if (txn != NULL && db->md_env != txn->mt_env) {
1337 if (key->mv_size == 0 || key->mv_size > MAXKEYSIZE) {
1341 if ((rc = mdb_search_page(db, txn, key, NULL, 0, &mpp)) != MDB_SUCCESS)
1344 leaf = mdb_search_node(db, mpp.mp_page, key, &exact, NULL);
1346 rc = mdb_read_data(db, mpp.mp_page, leaf, data);
1355 mdb_sibling(MDB_cursor *cursor, int move_right)
1359 MDB_ppage *parent, *top;
1362 top = CURSOR_TOP(cursor);
1363 if ((parent = SLIST_NEXT(top, mp_entry)) == NULL) {
1364 return ENOENT; /* root has no siblings */
1367 DPRINTF("parent page is page %lu, index %u",
1368 parent->mp_page->mp_pgno, parent->mp_ki);
1370 cursor_pop_page(cursor);
1371 if (move_right ? (parent->mp_ki + 1 >= NUMKEYS(parent->mp_page))
1372 : (parent->mp_ki == 0)) {
1373 DPRINTF("no more keys left, moving to %s sibling",
1374 move_right ? "right" : "left");
1375 if ((rc = mdb_sibling(cursor, move_right)) != MDB_SUCCESS)
1377 parent = CURSOR_TOP(cursor);
1383 DPRINTF("just moving to %s index key %u",
1384 move_right ? "right" : "left", parent->mp_ki);
1386 assert(IS_BRANCH(parent->mp_page));
1388 indx = NODEPTR(parent->mp_page, parent->mp_ki);
1389 if ((mp = mdbenv_get_page(cursor->mc_db->md_env, indx->mn_pgno)) == NULL)
1392 mp->parent = parent->mp_page;
1393 mp->parent_index = parent->mp_ki;
1396 cursor_push_page(cursor, mp);
1402 mdb_set_key(MDB_db *bt, MDB_page *mp, MDB_node *node,
1408 key->mv_size = node->mn_ksize;
1409 key->mv_data = NODEKEY(node);
1415 mdb_cursor_next(MDB_cursor *cursor, MDB_val *key, MDB_val *data)
1421 if (cursor->mc_eof) {
1425 assert(cursor->mc_initialized);
1427 top = CURSOR_TOP(cursor);
1430 DPRINTF("cursor_next: top page is %lu in cursor %p", mp->mp_pgno, cursor);
1432 if (top->mp_ki + 1 >= NUMKEYS(mp)) {
1433 DPRINTF("=====> move to next sibling page");
1434 if (mdb_sibling(cursor, 1) != MDB_SUCCESS) {
1438 top = CURSOR_TOP(cursor);
1440 DPRINTF("next page is %lu, key index %u", mp->mp_pgno, top->mp_ki);
1444 DPRINTF("==> cursor points to page %lu with %lu keys, key index %u",
1445 mp->mp_pgno, NUMKEYS(mp), top->mp_ki);
1447 assert(IS_LEAF(mp));
1448 leaf = NODEPTR(mp, top->mp_ki);
1450 if (data && mdb_read_data(cursor->mc_db, mp, leaf, data) != MDB_SUCCESS)
1453 return mdb_set_key(cursor->mc_db, mp, leaf, key);
1457 mdb_cursor_set(MDB_cursor *cursor, MDB_val *key, MDB_val *data,
1467 assert(key->mv_size > 0);
1469 rc = mdb_search_page(cursor->mc_db, cursor->mc_txn, key, cursor, 0, &mpp);
1470 if (rc != MDB_SUCCESS)
1472 assert(IS_LEAF(mpp.mp_page));
1474 top = CURSOR_TOP(cursor);
1475 leaf = mdb_search_node(cursor->mc_db, mpp.mp_page, key, exactp, &top->mp_ki);
1476 if (exactp != NULL && !*exactp) {
1477 /* MDB_CURSOR_EXACT specified and not an exact match. */
1482 DPRINTF("===> inexact leaf not found, goto sibling");
1483 if ((rc = mdb_sibling(cursor, 1)) != MDB_SUCCESS)
1484 return rc; /* no entries matched */
1485 top = CURSOR_TOP(cursor);
1487 mpp.mp_page = top->mp_page;
1488 assert(IS_LEAF(mpp.mp_page));
1489 leaf = NODEPTR(mpp.mp_page, 0);
1492 cursor->mc_initialized = 1;
1495 if (data && (rc = mdb_read_data(cursor->mc_db, mpp.mp_page, leaf, data)) != MDB_SUCCESS)
1498 rc = mdb_set_key(cursor->mc_db, mpp.mp_page, leaf, key);
1499 if (rc == MDB_SUCCESS) {
1500 DPRINTF("==> cursor placed on key %.*s",
1501 (int)key->mv_size, (char *)key->mv_data);
1509 mdb_cursor_first(MDB_cursor *cursor, MDB_val *key, MDB_val *data)
1515 rc = mdb_search_page(cursor->mc_db, cursor->mc_txn, NULL, cursor, 0, &mpp);
1516 if (rc != MDB_SUCCESS)
1518 assert(IS_LEAF(mpp.mp_page));
1520 leaf = NODEPTR(mpp.mp_page, 0);
1521 cursor->mc_initialized = 1;
1524 if (data && (rc = mdb_read_data(cursor->mc_db, mpp.mp_page, leaf, data)) != MDB_SUCCESS)
1527 return mdb_set_key(cursor->mc_db, mpp.mp_page, leaf, key);
1531 mdb_cursor_get(MDB_cursor *cursor, MDB_val *key, MDB_val *data,
1541 case MDB_CURSOR_EXACT:
1542 while (CURSOR_TOP(cursor) != NULL)
1543 cursor_pop_page(cursor);
1544 if (key == NULL || key->mv_size == 0 || key->mv_size > MAXKEYSIZE) {
1546 } else if (op == MDB_CURSOR_EXACT)
1547 rc = mdb_cursor_set(cursor, key, data, &exact);
1549 rc = mdb_cursor_set(cursor, key, data, NULL);
1552 if (!cursor->mc_initialized)
1553 rc = mdb_cursor_first(cursor, key, data);
1555 rc = mdb_cursor_next(cursor, key, data);
1558 while (CURSOR_TOP(cursor) != NULL)
1559 cursor_pop_page(cursor);
1560 rc = mdb_cursor_first(cursor, key, data);
1563 DPRINTF("unhandled/unimplemented cursor operation %u", op);
1571 /* Allocate a page and initialize it
1574 mdbenv_new_page(MDB_env *env, uint32_t flags, int num)
1578 assert(env != NULL);
1579 assert(env->me_txn != NULL);
1581 DPRINTF("allocating new mpage %lu, page size %u",
1582 env->me_txn->mt_next_pgno, env->me_head.mh_psize);
1583 if ((dp = mdb_newpage(env->me_txn, NULL, 0, num)) == NULL)
1585 dp->p.mp_flags = flags | P_DIRTY;
1586 dp->p.mp_lower = PAGEHDRSZ;
1587 dp->p.mp_upper = env->me_head.mh_psize;
1589 if (IS_BRANCH(&dp->p))
1590 env->me_meta.mm_stat.ms_branch_pages++;
1591 else if (IS_LEAF(&dp->p))
1592 env->me_meta.mm_stat.ms_leaf_pages++;
1593 else if (IS_OVERFLOW(&dp->p)) {
1594 env->me_meta.mm_stat.ms_overflow_pages += num;
1595 dp->p.mp_pages = num;
1602 mdb_leaf_size(MDB_db *db, MDB_val *key, MDB_val *data)
1606 sz = LEAFSIZE(key, data);
1607 if (data->mv_size >= db->md_env->me_head.mh_psize / MDB_MINKEYS) {
1608 /* put on overflow page */
1609 sz -= data->mv_size - sizeof(pgno_t);
1612 return sz + sizeof(indx_t);
1616 mdb_branch_size(MDB_db *db, MDB_val *key)
1621 if (sz >= db->md_env->me_head.mh_psize / MDB_MINKEYS) {
1622 /* put on overflow page */
1623 /* not implemented */
1624 /* sz -= key->size - sizeof(pgno_t); */
1627 return sz + sizeof(indx_t);
1631 mdb_add_node(MDB_db *db, MDB_page *mp, indx_t indx,
1632 MDB_val *key, MDB_val *data, pgno_t pgno, uint8_t flags)
1635 size_t node_size = NODESIZE;
1638 MDB_dpage *ofp = NULL; /* overflow page */
1640 assert(mp->mp_upper >= mp->mp_lower);
1642 DPRINTF("add node [%.*s] to %s page %lu at index %i, key size %zu",
1643 key ? (int)key->mv_size : 0, key ? (char *)key->mv_data : NULL,
1644 IS_LEAF(mp) ? "leaf" : "branch",
1645 mp->mp_pgno, indx, key ? key->mv_size : 0);
1648 node_size += key->mv_size;
1652 node_size += data->mv_size;
1653 if (F_ISSET(flags, F_BIGDATA)) {
1654 /* Data already on overflow page. */
1655 node_size -= data->mv_size - sizeof(pgno_t);
1656 } else if (data->mv_size >= db->md_env->me_head.mh_psize / MDB_MINKEYS) {
1657 int ovpages = PAGEHDRSZ + data->mv_size + db->md_env->me_head.mh_psize - 1;
1658 ovpages /= db->md_env->me_head.mh_psize;
1659 /* Put data on overflow page. */
1660 DPRINTF("data size is %zu, put on overflow page",
1662 node_size -= data->mv_size - sizeof(pgno_t);
1663 if ((ofp = mdbenv_new_page(db->md_env, P_OVERFLOW, ovpages)) == NULL)
1665 DPRINTF("allocated overflow page %lu", ofp->p.mp_pgno);
1670 if (node_size + sizeof(indx_t) > SIZELEFT(mp)) {
1671 DPRINTF("not enough room in page %lu, got %lu ptrs",
1672 mp->mp_pgno, NUMKEYS(mp));
1673 DPRINTF("upper - lower = %u - %u = %u", mp->mp_upper, mp->mp_lower,
1674 mp->mp_upper - mp->mp_lower);
1675 DPRINTF("node size = %zu", node_size);
1679 /* Move higher pointers up one slot. */
1680 for (i = NUMKEYS(mp); i > indx; i--)
1681 mp->mp_ptrs[i] = mp->mp_ptrs[i - 1];
1683 /* Adjust free space offsets. */
1684 ofs = mp->mp_upper - node_size;
1685 assert(ofs >= mp->mp_lower + sizeof(indx_t));
1686 mp->mp_ptrs[indx] = ofs;
1688 mp->mp_lower += sizeof(indx_t);
1690 /* Write the node data. */
1691 node = NODEPTR(mp, indx);
1692 node->mn_ksize = (key == NULL) ? 0 : key->mv_size;
1693 node->mn_flags = flags;
1695 node->mn_dsize = data->mv_size;
1697 node->mn_pgno = pgno;
1700 bcopy(key->mv_data, NODEKEY(node), key->mv_size);
1705 if (F_ISSET(flags, F_BIGDATA))
1706 bcopy(data->mv_data, node->mn_data + key->mv_size,
1709 bcopy(data->mv_data, node->mn_data + key->mv_size,
1712 bcopy(&ofp->p.mp_pgno, node->mn_data + key->mv_size,
1714 bcopy(data->mv_data, METADATA(&ofp->p), data->mv_size);
1722 mdb_del_node(MDB_db *db, MDB_page *mp, indx_t indx)
1725 indx_t i, j, numkeys, ptr;
1729 DPRINTF("delete node %u on %s page %lu", indx,
1730 IS_LEAF(mp) ? "leaf" : "branch", mp->mp_pgno);
1731 assert(indx < NUMKEYS(mp));
1733 node = NODEPTR(mp, indx);
1734 sz = NODESIZE + node->mn_ksize;
1736 if (F_ISSET(node->mn_flags, F_BIGDATA))
1737 sz += sizeof(pgno_t);
1739 sz += NODEDSZ(node);
1742 ptr = mp->mp_ptrs[indx];
1743 numkeys = NUMKEYS(mp);
1744 for (i = j = 0; i < numkeys; i++) {
1746 mp->mp_ptrs[j] = mp->mp_ptrs[i];
1747 if (mp->mp_ptrs[i] < ptr)
1748 mp->mp_ptrs[j] += sz;
1753 base = (char *)mp + mp->mp_upper;
1754 bcopy(base, base + sz, ptr - mp->mp_upper);
1756 mp->mp_lower -= sizeof(indx_t);
1761 mdb_cursor_open(MDB_db *db, MDB_txn *txn, MDB_cursor **ret)
1765 if (db == NULL || ret == NULL)
1768 if (txn != NULL && db->md_env != txn->mt_env) {
1772 if ((cursor = calloc(1, sizeof(*cursor))) != NULL) {
1773 SLIST_INIT(&cursor->mc_stack);
1775 cursor->mc_txn = txn;
1784 mdb_cursor_close(MDB_cursor *cursor)
1786 if (cursor != NULL) {
1787 while (!CURSOR_EMPTY(cursor))
1788 cursor_pop_page(cursor);
1790 /* btree_close(cursor->bt); */
1796 mdb_update_key(MDB_db *db, MDB_page *mp, indx_t indx,
1799 indx_t ptr, i, numkeys;
1805 node = NODEPTR(mp, indx);
1806 ptr = mp->mp_ptrs[indx];
1807 DPRINTF("update key %u (ofs %u) [%.*s] to [%.*s] on page %lu",
1809 (int)node->mn_ksize, (char *)NODEKEY(node),
1810 (int)key->mv_size, (char *)key->mv_data,
1813 delta = key->mv_size - node->mn_ksize;
1815 if (delta > 0 && SIZELEFT(mp) < delta) {
1816 DPRINTF("OUCH! Not enough room, delta = %d", delta);
1820 numkeys = NUMKEYS(mp);
1821 for (i = 0; i < numkeys; i++) {
1822 if (mp->mp_ptrs[i] <= ptr)
1823 mp->mp_ptrs[i] -= delta;
1826 base = (char *)mp + mp->mp_upper;
1827 len = ptr - mp->mp_upper + NODESIZE;
1828 bcopy(base, base - delta, len);
1829 mp->mp_upper -= delta;
1831 node = NODEPTR(mp, indx);
1832 node->mn_ksize = key->mv_size;
1835 bcopy(key->mv_data, NODEKEY(node), key->mv_size);
1840 /* Move a node from src to dst.
1843 mdb_move_node(MDB_db *bt, MDB_pageparent *src, indx_t srcindx,
1844 MDB_pageparent *dst, indx_t dstindx)
1850 srcnode = NODEPTR(src->mp_page, srcindx);
1851 DPRINTF("moving %s node %u [%.*s] on page %lu to node %u on page %lu",
1852 IS_LEAF(src->mp_page) ? "leaf" : "branch",
1854 (int)srcnode->mn_ksize, (char *)NODEKEY(srcnode),
1855 src->mp_page->mp_pgno,
1856 dstindx, dst->mp_page->mp_pgno);
1858 /* Mark src and dst as dirty. */
1859 if ((rc = mdb_touch(bt->md_env->me_txn, src)) ||
1860 (rc = mdb_touch(bt->md_env->me_txn, dst)))
1863 /* Add the node to the destination page.
1865 key.mv_size = srcnode->mn_ksize;
1866 key.mv_data = NODEKEY(srcnode);
1867 data.mv_size = NODEDSZ(srcnode);
1868 data.mv_data = NODEDATA(srcnode);
1869 rc = mdb_add_node(bt, dst->mp_page, dstindx, &key, &data, NODEPGNO(srcnode),
1871 if (rc != MDB_SUCCESS)
1874 /* Delete the node from the source page.
1876 mdb_del_node(bt, src->mp_page, srcindx);
1878 /* Update the parent separators.
1880 if (srcindx == 0 && src->mp_pi != 0) {
1881 DPRINTF("update separator for source page %lu to [%.*s]",
1882 src->mp_page->mp_pgno, (int)key.mv_size, (char *)key.mv_data);
1883 if ((rc = mdb_update_key(bt, src->mp_parent, src->mp_pi,
1884 &key)) != MDB_SUCCESS)
1888 if (srcindx == 0 && IS_BRANCH(src->mp_page)) {
1890 nullkey.mv_size = 0;
1891 assert(mdb_update_key(bt, src->mp_page, 0, &nullkey) == MDB_SUCCESS);
1894 if (dstindx == 0 && dst->mp_pi != 0) {
1895 DPRINTF("update separator for destination page %lu to [%.*s]",
1896 dst->mp_page->mp_pgno, (int)key.mv_size, (char *)key.mv_data);
1897 if ((rc = mdb_update_key(bt, dst->mp_parent, dst->mp_pi,
1898 &key)) != MDB_SUCCESS)
1902 if (dstindx == 0 && IS_BRANCH(dst->mp_page)) {
1904 nullkey.mv_size = 0;
1905 assert(mdb_update_key(bt, dst->mp_page, 0, &nullkey) == MDB_SUCCESS);
1912 mdb_merge(MDB_db *bt, MDB_pageparent *src, MDB_pageparent *dst)
1921 DPRINTF("merging page %lu and %lu", src->mp_page->mp_pgno, dst->mp_page->mp_pgno);
1923 assert(src->mp_parent); /* can't merge root page */
1924 assert(dst->mp_parent);
1925 assert(bt->md_env->me_txn != NULL);
1927 /* Mark src and dst as dirty. */
1928 if ((rc = mdb_touch(bt->md_env->me_txn, src)) ||
1929 (rc = mdb_touch(bt->md_env->me_txn, dst)))
1932 /* Move all nodes from src to dst.
1934 for (i = 0; i < NUMKEYS(src->mp_page); i++) {
1935 srcnode = NODEPTR(src->mp_page, i);
1937 key.mv_size = srcnode->mn_ksize;
1938 key.mv_data = NODEKEY(srcnode);
1939 data.mv_size = NODEDSZ(srcnode);
1940 data.mv_data = NODEDATA(srcnode);
1941 rc = mdb_add_node(bt, dst->mp_page, NUMKEYS(dst->mp_page), &key,
1942 &data, NODEPGNO(srcnode), srcnode->mn_flags);
1943 if (rc != MDB_SUCCESS)
1947 DPRINTF("dst page %lu now has %lu keys (%.1f%% filled)",
1948 dst->mp_page->mp_pgno, NUMKEYS(dst->mp_page), (float)PAGEFILL(bt->md_env, dst->mp_page) / 10);
1950 /* Unlink the src page from parent.
1952 mdb_del_node(bt, src->mp_parent, src->mp_pi);
1953 if (src->mp_pi == 0) {
1955 if ((rc = mdb_update_key(bt, src->mp_parent, 0, &key)) != MDB_SUCCESS)
1959 if (IS_LEAF(src->mp_page))
1960 bt->md_env->me_meta.mm_stat.ms_leaf_pages--;
1962 bt->md_env->me_meta.mm_stat.ms_branch_pages--;
1964 mpp.mp_page = src->mp_parent;
1965 dh = (MDB_dhead *)src->mp_parent;
1967 mpp.mp_parent = dh->md_parent;
1968 mpp.mp_pi = dh->md_pi;
1970 return mdb_rebalance(bt, &mpp);
1973 #define FILL_THRESHOLD 250
1976 mdb_rebalance(MDB_db *db, MDB_pageparent *mpp)
1981 indx_t si = 0, di = 0;
1984 assert(db->md_env->me_txn != NULL);
1985 assert(mpp != NULL);
1987 DPRINTF("rebalancing %s page %lu (has %lu keys, %.1f%% full)",
1988 IS_LEAF(mpp->mp_page) ? "leaf" : "branch",
1989 mpp->mp_page->mp_pgno, NUMKEYS(mpp->mp_page), (float)PAGEFILL(db->md_env, mpp->mp_page) / 10);
1991 if (PAGEFILL(db->md_env, mpp->mp_page) >= FILL_THRESHOLD) {
1992 DPRINTF("no need to rebalance page %lu, above fill threshold",
1993 mpp->mp_page->mp_pgno);
1997 if (mpp->mp_parent == NULL) {
1998 if (NUMKEYS(mpp->mp_page) == 0) {
1999 DPRINTF("tree is completely empty");
2000 db->md_env->me_txn->mt_root = P_INVALID;
2001 db->md_env->me_meta.mm_stat.ms_depth--;
2002 db->md_env->me_meta.mm_stat.ms_leaf_pages--;
2003 } else if (IS_BRANCH(mpp->mp_page) && NUMKEYS(mpp->mp_page) == 1) {
2004 DPRINTF("collapsing root page!");
2005 db->md_env->me_txn->mt_root = NODEPGNO(NODEPTR(mpp->mp_page, 0));
2006 if ((root = mdbenv_get_page(db->md_env, db->md_env->me_txn->mt_root)) == NULL)
2008 db->md_env->me_meta.mm_stat.ms_depth--;
2009 db->md_env->me_meta.mm_stat.ms_branch_pages--;
2011 DPRINTF("root page doesn't need rebalancing");
2015 /* The parent (branch page) must have at least 2 pointers,
2016 * otherwise the tree is invalid.
2018 assert(NUMKEYS(mpp->mp_parent) > 1);
2020 /* Leaf page fill factor is below the threshold.
2021 * Try to move keys from left or right neighbor, or
2022 * merge with a neighbor page.
2027 if (mpp->mp_pi == 0) {
2028 /* We're the leftmost leaf in our parent.
2030 DPRINTF("reading right neighbor");
2031 node = NODEPTR(mpp->mp_parent, mpp->mp_pi + 1);
2032 if ((npp.mp_page = mdbenv_get_page(db->md_env, NODEPGNO(node))) == NULL)
2034 npp.mp_pi = mpp->mp_pi + 1;
2036 di = NUMKEYS(mpp->mp_page);
2038 /* There is at least one neighbor to the left.
2040 DPRINTF("reading left neighbor");
2041 node = NODEPTR(mpp->mp_parent, mpp->mp_pi - 1);
2042 if ((npp.mp_page = mdbenv_get_page(db->md_env, NODEPGNO(node))) == NULL)
2044 npp.mp_pi = mpp->mp_pi - 1;
2045 si = NUMKEYS(npp.mp_page) - 1;
2048 npp.mp_parent = mpp->mp_parent;
2050 DPRINTF("found neighbor page %lu (%lu keys, %.1f%% full)",
2051 npp.mp_page->mp_pgno, NUMKEYS(npp.mp_page), (float)PAGEFILL(db->md_env, npp.mp_page) / 10);
2053 /* If the neighbor page is above threshold and has at least two
2054 * keys, move one key from it.
2056 * Otherwise we should try to merge them.
2058 if (PAGEFILL(db->md_env, npp.mp_page) >= FILL_THRESHOLD && NUMKEYS(npp.mp_page) >= 2)
2059 return mdb_move_node(db, &npp, si, mpp, di);
2060 else { /* FIXME: if (has_enough_room()) */
2061 if (mpp->mp_pi == 0)
2062 return mdb_merge(db, &npp, mpp);
2064 return mdb_merge(db, mpp, &npp);
2069 mdb_del(MDB_db *bt, MDB_txn *txn,
2070 MDB_val *key, MDB_val *data)
2077 DPRINTF("========> delete key %.*s", (int)key->mv_size, (char *)key->mv_data);
2079 assert(key != NULL);
2081 if (bt == NULL || txn == NULL)
2084 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
2088 if (bt->md_env->me_txn != txn) {
2092 if (key->mv_size == 0 || key->mv_size > MAXKEYSIZE) {
2096 if ((rc = mdb_search_page(bt, txn, key, NULL, 1, &mpp)) != MDB_SUCCESS)
2099 leaf = mdb_search_node(bt, mpp.mp_page, key, &exact, &ki);
2100 if (leaf == NULL || !exact) {
2104 if (data && (rc = mdb_read_data(bt, NULL, leaf, data)) != MDB_SUCCESS)
2107 mdb_del_node(bt, mpp.mp_page, ki);
2108 bt->md_env->me_meta.mm_stat.ms_entries--;
2109 rc = mdb_rebalance(bt, &mpp);
2110 if (rc != MDB_SUCCESS)
2111 txn->mt_flags |= MDB_TXN_ERROR;
2116 /* Split page <*mpp>, and insert <key,(data|newpgno)> in either left or
2117 * right sibling, at index <*newindxp> (as if unsplit). Updates *mpp and
2118 * *newindxp with the actual values after split, ie if *mpp and *newindxp
2119 * refer to a node in the new right sibling page.
2122 mdb_split(MDB_db *bt, MDB_page **mpp, unsigned int *newindxp,
2123 MDB_val *newkey, MDB_val *newdata, pgno_t newpgno)
2126 int rc = MDB_SUCCESS, ins_new = 0;
2129 unsigned int i, j, split_indx;
2131 MDB_page *pright, *p;
2132 MDB_val sepkey, rkey, rdata;
2134 MDB_dpage *mdp, *rdp, *pdp;
2138 assert(bt->md_env != NULL);
2140 dh = ((MDB_dhead *)*mpp) - 1;
2141 mdp = (MDB_dpage *)dh;
2142 newindx = *newindxp;
2144 DPRINTF("-----> splitting %s page %lu and adding [%.*s] at index %i",
2145 IS_LEAF(&mdp->p) ? "leaf" : "branch", mdp->p.mp_pgno,
2146 (int)newkey->mv_size, (char *)newkey->mv_data, *newindxp);
2148 if (mdp->h.md_parent == NULL) {
2149 if ((pdp = mdbenv_new_page(bt->md_env, P_BRANCH, 1)) == NULL)
2152 mdp->h.md_parent = &pdp->p;
2153 bt->md_env->me_txn->mt_root = pdp->p.mp_pgno;
2154 DPRINTF("root split! new root = %lu", pdp->p.mp_pgno);
2155 bt->md_env->me_meta.mm_stat.ms_depth++;
2157 /* Add left (implicit) pointer. */
2158 if (mdb_add_node(bt, &pdp->p, 0, NULL, NULL,
2159 mdp->p.mp_pgno, 0) != MDB_SUCCESS)
2162 DPRINTF("parent branch page is %lu", mdp->h.md_parent->mp_pgno);
2165 /* Create a right sibling. */
2166 if ((rdp = mdbenv_new_page(bt->md_env, mdp->p.mp_flags, 1)) == NULL)
2168 rdp->h.md_parent = mdp->h.md_parent;
2169 rdp->h.md_pi = mdp->h.md_pi + 1;
2170 DPRINTF("new right sibling: page %lu", rdp->p.mp_pgno);
2172 /* Move half of the keys to the right sibling. */
2173 if ((copy = malloc(bt->md_env->me_head.mh_psize)) == NULL)
2175 bcopy(&mdp->p, copy, bt->md_env->me_head.mh_psize);
2176 bzero(&mdp->p.mp_ptrs, bt->md_env->me_head.mh_psize - PAGEHDRSZ);
2177 mdp->p.mp_lower = PAGEHDRSZ;
2178 mdp->p.mp_upper = bt->md_env->me_head.mh_psize;
2180 split_indx = NUMKEYS(copy) / 2 + 1;
2182 /* First find the separating key between the split pages.
2184 bzero(&sepkey, sizeof(sepkey));
2185 if (newindx == split_indx) {
2186 sepkey.mv_size = newkey->mv_size;
2187 sepkey.mv_data = newkey->mv_data;
2189 node = NODEPTR(copy, split_indx);
2190 sepkey.mv_size = node->mn_ksize;
2191 sepkey.mv_data = NODEKEY(node);
2194 DPRINTF("separator is [%.*s]", (int)sepkey.mv_size, (char *)sepkey.mv_data);
2196 /* Copy separator key to the parent.
2198 if (SIZELEFT(rdp->h.md_parent) < mdb_branch_size(bt, &sepkey)) {
2199 rc = mdb_split(bt, &rdp->h.md_parent, &rdp->h.md_pi,
2200 &sepkey, NULL, rdp->p.mp_pgno);
2202 /* Right page might now have changed parent.
2203 * Check if left page also changed parent.
2205 if (rdp->h.md_parent != mdp->h.md_parent &&
2206 mdp->h.md_pi >= NUMKEYS(mdp->h.md_parent)) {
2207 mdp->h.md_parent = rdp->h.md_parent;
2208 mdp->h.md_pi = rdp->h.md_pi - 1;
2211 rc = mdb_add_node(bt, rdp->h.md_parent, rdp->h.md_pi,
2212 &sepkey, NULL, rdp->p.mp_pgno, 0);
2214 if (rc != MDB_SUCCESS) {
2219 for (i = j = 0; i <= NUMKEYS(copy); j++) {
2220 if (i < split_indx) {
2221 /* Re-insert in left sibling. */
2224 /* Insert in right sibling. */
2225 if (i == split_indx)
2226 /* Reset insert index for right sibling. */
2227 j = (i == newindx && ins_new);
2231 if (i == newindx && !ins_new) {
2232 /* Insert the original entry that caused the split. */
2233 rkey.mv_data = newkey->mv_data;
2234 rkey.mv_size = newkey->mv_size;
2235 if (IS_LEAF(&mdp->p)) {
2236 rdata.mv_data = newdata->mv_data;
2237 rdata.mv_size = newdata->mv_size;
2244 /* Update page and index for the new key. */
2247 } else if (i == NUMKEYS(copy)) {
2250 node = NODEPTR(copy, i);
2251 rkey.mv_data = NODEKEY(node);
2252 rkey.mv_size = node->mn_ksize;
2253 if (IS_LEAF(&mdp->p)) {
2254 rdata.mv_data = NODEDATA(node);
2255 rdata.mv_size = node->mn_dsize;
2257 pgno = node->mn_pgno;
2258 flags = node->mn_flags;
2263 if (!IS_LEAF(&mdp->p) && j == 0) {
2264 /* First branch index doesn't need key data. */
2268 rc = mdb_add_node(bt, &pdp->p, j, &rkey, &rdata, pgno,flags);
2276 mdb_put(MDB_db *bt, MDB_txn *txn,
2277 MDB_val *key, MDB_val *data, unsigned int flags)
2279 int rc = MDB_SUCCESS, exact;
2284 assert(key != NULL);
2285 assert(data != NULL);
2287 if (bt == NULL || txn == NULL)
2290 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
2294 if (bt->md_env->me_txn != txn) {
2298 if (key->mv_size == 0 || key->mv_size > MAXKEYSIZE) {
2302 DPRINTF("==> put key %.*s, size %zu, data size %zu",
2303 (int)key->mv_size, (char *)key->mv_data, key->mv_size, data->mv_size);
2305 rc = mdb_search_page(bt, txn, key, NULL, 1, &mpp);
2306 if (rc == MDB_SUCCESS) {
2307 leaf = mdb_search_node(bt, mpp.mp_page, key, &exact, &ki);
2308 if (leaf && exact) {
2309 if (F_ISSET(flags, MDB_NOOVERWRITE)) {
2310 DPRINTF("duplicate key %.*s",
2311 (int)key->mv_size, (char *)key->mv_data);
2314 mdb_del_node(bt, mpp.mp_page, ki);
2316 if (leaf == NULL) { /* append if not found */
2317 ki = NUMKEYS(mpp.mp_page);
2318 DPRINTF("appending key at index %i", ki);
2320 } else if (rc == ENOENT) {
2322 /* new file, just write a root leaf page */
2323 DPRINTF("allocating new root leaf page");
2324 if ((dp = mdbenv_new_page(bt->md_env, P_LEAF, 1)) == NULL) {
2327 mpp.mp_page = &dp->p;
2328 txn->mt_root = mpp.mp_page->mp_pgno;
2329 bt->md_env->me_meta.mm_stat.ms_depth++;
2335 assert(IS_LEAF(mpp.mp_page));
2336 DPRINTF("there are %lu keys, should insert new key at index %i",
2337 NUMKEYS(mpp.mp_page), ki);
2339 if (SIZELEFT(mpp.mp_page) < mdb_leaf_size(bt, key, data)) {
2340 rc = mdb_split(bt, &mpp.mp_page, &ki, key, data, P_INVALID);
2342 /* There is room already in this leaf page. */
2343 rc = mdb_add_node(bt, mpp.mp_page, ki, key, data, 0, 0);
2346 if (rc != MDB_SUCCESS)
2347 txn->mt_flags |= MDB_TXN_ERROR;
2349 bt->md_env->me_meta.mm_stat.ms_entries++;
2356 mdbenv_get_flags(MDB_env *env, unsigned int *arg)
2361 *arg = env->me_flags;
2366 mdbenv_get_path(MDB_env *env, const char **arg)
2371 *arg = env->me_path;
2376 mdbenv_stat(MDB_env *env, MDB_stat **arg)
2378 if (env == NULL || arg == NULL)
2381 *arg = &env->me_meta.mm_stat;
2386 int mdb_open(MDB_env *env, MDB_txn *txn, const char *name, unsigned int flags, MDB_db **db)
2389 *db = (MDB_db *)&env->me_db;
2395 int mdb_stat(MDB_db *db, MDB_stat **arg)
2397 if (db == NULL || arg == NULL)
2400 *arg = &db->md_stat;
2405 void mdb_close(MDB_db *db)