35 #if (DEBUG +0) && defined(__GNUC__)
36 # define DPRINTF(fmt, ...) \
37 fprintf(stderr, "%s:%d: " fmt "\n", __func__, __LINE__, ##__VA_ARGS__)
39 # define DPRINTF(...) ((void) 0)
44 #define MDB_MAGIC 0xBEEFC0DE
46 #define MAXKEYSIZE 255
48 #define P_INVALID (~0UL)
50 #define F_ISSET(w, f) (((w) & (f)) == (f))
52 typedef uint16_t indx_t;
54 #define DEFAULT_READERS 126
55 #define DEFAULT_MAPSIZE 1048576
57 /* Lock descriptor stuff */
62 typedef struct MDB_rxbody {
67 #define CACHELINE 64 /* most CPUs. Itanium uses 128 */
70 typedef struct MDB_reader {
72 /* cache line alignment */
73 char pad[CACHELINE-sizeof(MDB_rxbody)];
78 uint32_t mt_version; \
79 pthread_mutex_t mt_mutex; \
81 uint32_t mt_numreaders
82 typedef struct MDB_txbody {
86 typedef struct MDB_txninfo {
88 char pad[CACHELINE-sizeof(MDB_txbody)];
89 pthread_mutex_t mt_wmutex;
90 char pad2[CACHELINE-sizeof(pthread_mutex_t)];
91 MDB_reader mt_readers[1];
94 /* Common header for all page types. Overflow pages
95 * occupy a number of contiguous pages with no
96 * headers on any page after the first.
98 typedef struct MDB_page { /* represents a page of storage */
99 pgno_t mp_pgno; /* page number */
100 #define P_BRANCH 0x01 /* branch page */
101 #define P_LEAF 0x02 /* leaf page */
102 #define P_OVERFLOW 0x04 /* overflow page */
103 #define P_META 0x08 /* meta page */
104 #define P_DIRTY 0x10 /* dirty page */
106 #define mp_lower mp_pb.pb.pb_lower
107 #define mp_upper mp_pb.pb.pb_upper
108 #define mp_pages mp_pb.pb_pages
111 indx_t pb_lower; /* lower bound of free space */
112 indx_t pb_upper; /* upper bound of free space */
114 uint32_t pb_pages; /* number of overflow pages */
116 indx_t mp_ptrs[1]; /* dynamic size */
119 #define PAGEHDRSZ ((unsigned) offsetof(MDB_page, mp_ptrs))
121 #define NUMKEYS(p) (((p)->mp_lower - PAGEHDRSZ) >> 1)
122 #define SIZELEFT(p) (indx_t)((p)->mp_upper - (p)->mp_lower)
123 #define PAGEFILL(env, p) (1000L * ((env)->me_meta.mm_stat.ms_psize - PAGEHDRSZ - SIZELEFT(p)) / \
124 ((env)->me_meta.mm_stat.ms_psize - PAGEHDRSZ))
125 #define IS_LEAF(p) F_ISSET((p)->mp_flags, P_LEAF)
126 #define IS_BRANCH(p) F_ISSET((p)->mp_flags, P_BRANCH)
127 #define IS_OVERFLOW(p) F_ISSET((p)->mp_flags, P_OVERFLOW)
129 #define OVPAGES(size, psize) (PAGEHDRSZ + size + psize - 1) / psize;
131 typedef struct MDB_meta { /* meta (footer) page content */
134 void *mm_address; /* address for fixed mapping */
135 size_t mm_mapsize; /* size of mmap region */
136 pgno_t mm_last_pg; /* last used page in file */
137 ulong mm_txnid; /* txnid that committed this page */
139 pgno_t mm_root; /* page number of root page */
142 typedef struct MDB_dhead { /* a dirty page */
143 SIMPLEQ_ENTRY(MDB_dpage) md_next; /* queue of dirty pages */
145 unsigned md_pi; /* parent index */
149 typedef struct MDB_dpage {
154 SIMPLEQ_HEAD(dirty_queue, MDB_dpage);
156 typedef struct MDB_oldpages {
157 struct MDB_oldpages *mo_next;
159 pgno_t mo_pages[1]; /* dynamic */
162 typedef struct MDB_pageparent {
168 static MDB_dpage *mdb_newpage(MDB_txn *txn, MDB_page *parent, unsigned int parent_idx, int num);
169 static int mdb_touch(MDB_txn *txn, MDB_pageparent *mp);
171 typedef struct MDB_ppage { /* ordered list of pages */
172 SLIST_ENTRY(MDB_ppage) mp_entry;
174 unsigned int mp_ki; /* cursor index on page */
176 SLIST_HEAD(page_stack, MDB_ppage);
178 #define CURSOR_EMPTY(c) SLIST_EMPTY(&(c)->mc_stack)
179 #define CURSOR_TOP(c) SLIST_FIRST(&(c)->mc_stack)
180 #define CURSOR_POP(c) SLIST_REMOVE_HEAD(&(c)->mc_stack, mp_entry)
181 #define CURSOR_PUSH(c,p) SLIST_INSERT_HEAD(&(c)->mc_stack, p, mp_entry)
186 struct page_stack mc_stack; /* stack of parent pages */
187 short mc_initialized; /* 1 if initialized */
188 short mc_eof; /* 1 if end is reached */
191 #define METAHASHLEN offsetof(MDB_meta, mm_hash)
192 #define METADATA(p) ((void *)((char *)p + PAGEHDRSZ))
194 typedef struct MDB_node {
195 #define mn_pgno mn_p.np_pgno
196 #define mn_dsize mn_p.np_dsize
198 pgno_t np_pgno; /* child page number */
199 uint32_t np_dsize; /* leaf data size */
201 unsigned int mn_flags:4;
202 unsigned int mn_ksize:12; /* key size */
203 #define F_BIGDATA 0x01 /* data put on overflow page */
208 pgno_t mt_root; /* current / new root page */
209 pgno_t mt_next_pgno; /* next unallocated page */
213 pgno_t *mt_free_pgs; /* this is an IDL */
215 struct dirty_queue *dirty_queue; /* modified pages */
218 #define MDB_TXN_RDONLY 0x01 /* read-only transaction */
219 #define MDB_TXN_ERROR 0x02 /* an error has occurred */
220 #define MDB_TXN_METOGGLE 0x04 /* used meta page 1 */
221 unsigned int mt_flags;
225 unsigned int md_flags;
226 MDB_cmp_func *md_cmp; /* user compare function */
227 MDB_rel_func *md_rel; /* user relocate function */
228 MDB_db *md_parent; /* parent tree */
231 pgno_t md_root; /* page number of root page */
238 unsigned int me_maxreaders;
241 MDB_txninfo *me_txns;
242 MDB_db me_db; /* first DB */
244 MDB_txn *me_txn; /* current write transaction */
246 off_t me_size; /* current file size */
247 pthread_key_t me_txkey; /* thread-key for readers */
248 MDB_oldpages *me_pghead;
249 MDB_oldpages *me_pgtail;
252 #define NODESIZE offsetof(MDB_node, mn_data)
254 #define INDXSIZE(k) (NODESIZE + ((k) == NULL ? 0 : (k)->mv_size))
255 #define LEAFSIZE(k, d) (NODESIZE + (k)->mv_size + (d)->mv_size)
256 #define NODEPTR(p, i) ((MDB_node *)((char *)(p) + (p)->mp_ptrs[i]))
257 #define NODEKEY(node) (void *)((node)->mn_data)
258 #define NODEDATA(node) (void *)((char *)(node)->mn_data + (node)->mn_ksize)
259 #define NODEPGNO(node) ((node)->mn_pgno)
260 #define NODEDSZ(node) ((node)->mn_dsize)
262 #define MDB_COMMIT_PAGES 64 /* max number of pages to write in one commit */
263 #define MDB_MAXCACHE_DEF 1024 /* max number of pages to keep in cache */
265 static int mdb_search_page_root(MDB_db *db,
267 MDB_cursor *cursor, int modify,
268 MDB_pageparent *mpp);
269 static int mdb_search_page(MDB_db *db,
270 MDB_txn *txn, MDB_val *key,
271 MDB_cursor *cursor, int modify,
272 MDB_pageparent *mpp);
274 static int mdbenv_read_header(MDB_env *env);
275 static int mdb_check_meta_page(MDB_page *p);
276 static int mdbenv_read_meta(MDB_env *env, int *which);
277 static int mdbenv_write_meta(MDB_txn *txn);
278 static MDB_page *mdbenv_get_page(MDB_env *env, pgno_t pgno);
280 static MDB_node *mdb_search_node(MDB_db *db, MDB_page *mp,
281 MDB_val *key, int *exactp, unsigned int *kip);
282 static int mdb_add_node(MDB_db *bt, MDB_page *mp,
283 indx_t indx, MDB_val *key, MDB_val *data,
284 pgno_t pgno, uint8_t flags);
285 static void mdb_del_node(MDB_db *bt, MDB_page *mp,
287 static int mdb_read_data(MDB_db *bt, MDB_page *mp,
288 MDB_node *leaf, MDB_val *data);
290 static int mdb_rebalance(MDB_db *bt, MDB_pageparent *mp);
291 static int mdb_update_key(MDB_db *bt, MDB_page *mp,
292 indx_t indx, MDB_val *key);
293 static int mdb_move_node(MDB_db *bt,
294 MDB_pageparent *src, indx_t srcindx,
295 MDB_pageparent *dst, indx_t dstindx);
296 static int mdb_merge(MDB_db *bt, MDB_pageparent *src,
297 MDB_pageparent *dst);
298 static int mdb_split(MDB_db *bt, MDB_page **mpp,
299 unsigned int *newindxp, MDB_val *newkey,
300 MDB_val *newdata, pgno_t newpgno);
301 static MDB_dpage *mdbenv_new_page(MDB_env *env, uint32_t flags, int num);
303 static void cursor_pop_page(MDB_cursor *cursor);
304 static MDB_ppage *cursor_push_page(MDB_cursor *cursor,
307 static int mdb_set_key(MDB_db *bt, MDB_page *mp,
308 MDB_node *node, MDB_val *key);
309 static int mdb_sibling(MDB_cursor *cursor, int move_right);
310 static int mdb_cursor_next(MDB_cursor *cursor,
311 MDB_val *key, MDB_val *data);
312 static int mdb_cursor_set(MDB_cursor *cursor,
313 MDB_val *key, MDB_val *data, int *exactp);
314 static int mdb_cursor_first(MDB_cursor *cursor,
315 MDB_val *key, MDB_val *data);
317 static size_t mdb_leaf_size(MDB_db *bt, MDB_val *key,
319 static size_t mdb_branch_size(MDB_db *bt, MDB_val *key);
321 static int memncmp(const void *s1, size_t n1,
322 const void *s2, size_t n2);
323 static int memnrcmp(const void *s1, size_t n1,
324 const void *s2, size_t n2);
327 memncmp(const void *s1, size_t n1, const void *s2, size_t n2)
329 int diff, len_diff = -1;
332 len_diff = (n1 > n2);
335 diff = memcmp(s1, s2, n1);
336 return diff ? diff : len_diff;
340 memnrcmp(const void *s1, size_t n1, const void *s2, size_t n2)
342 const unsigned char *p1, *p2, *p1_lim;
349 p1 = (const unsigned char *)s1 + n1 - 1;
350 p2 = (const unsigned char *)s2 + n2 - 1;
352 for (p1_lim = (n1 <= n2 ? s1 : s2); *p1 == *p2; p1--, p2--) {
354 return (p1 != s1) ? (p1 != p2) : (p2 != s2) ? -1 : 0;
360 mdb_cmp(MDB_db *db, const MDB_val *a, const MDB_val *b)
362 return db->md_cmp(a, b);
366 _mdb_cmp(MDB_db *db, const MDB_val *key1, const MDB_val *key2)
368 if (F_ISSET(db->md_flags, MDB_REVERSEKEY))
369 return memnrcmp(key1->mv_data, key1->mv_size, key2->mv_data, key2->mv_size);
371 return memncmp((char *)key1->mv_data, key1->mv_size, key2->mv_data, key2->mv_size);
374 /* Allocate new page(s) for writing */
376 mdb_newpage(MDB_txn *txn, MDB_page *parent, unsigned int parent_idx, int num)
379 pgno_t pgno = P_INVALID;
381 if (txn->mt_env->me_pghead) {
382 ulong oldest = txn->mt_txnid - 2;
384 for (i=0; i<txn->mt_env->me_txns->mt_numreaders; i++) {
385 if (txn->mt_env->me_txns->mt_readers[i].mr_txnid < oldest)
386 oldest = txn->mt_env->me_txns->mt_readers[i].mr_txnid;
388 if (oldest > txn->mt_env->me_pghead->mo_txnid) {
389 MDB_oldpages *mop = txn->mt_env->me_pghead;
390 txn->mt_oldest = oldest;
395 /* peel pages off tail, so we only have to truncate the list */
396 pgno = MDB_IDL_LAST(mop->mo_pages);
397 if (MDB_IDL_IS_RANGE(mop->mo_pages)) {
399 if (mop->mo_pages[2] > mop->mo_pages[1])
400 mop->mo_pages[0] = 0;
404 if (MDB_IDL_IS_ZERO(mop->mo_pages)) {
405 txn->mt_env->me_pghead = mop->mo_next;
406 if (!txn->mt_env->me_pghead)
407 txn->mt_env->me_pgtail = NULL;
414 if ((dp = malloc(txn->mt_env->me_meta.mm_stat.ms_psize * num + sizeof(MDB_dhead))) == NULL)
417 dp->h.md_parent = parent;
418 dp->h.md_pi = parent_idx;
419 SIMPLEQ_INSERT_TAIL(txn->mt_u.dirty_queue, dp, h.md_next);
420 if (pgno == P_INVALID) {
421 dp->p.mp_pgno = txn->mt_next_pgno;
422 txn->mt_next_pgno += num;
424 dp->p.mp_pgno = pgno;
430 /* Touch a page: make it dirty and re-insert into tree with updated pgno.
433 mdb_touch(MDB_txn *txn, MDB_pageparent *pp)
435 MDB_page *mp = pp->mp_page;
440 if (!F_ISSET(mp->mp_flags, P_DIRTY)) {
442 if ((dp = mdb_newpage(txn, pp->mp_parent, pp->mp_pi, 1)) == NULL)
444 DPRINTF("touched page %lu -> %lu", mp->mp_pgno, dp->p.mp_pgno);
445 mdb_idl_insert(txn->mt_free_pgs, mp->mp_pgno);
446 pgno = dp->p.mp_pgno;
447 bcopy(mp, &dp->p, txn->mt_env->me_meta.mm_stat.ms_psize);
450 mp->mp_flags |= P_DIRTY;
452 /* Update the page number to new touched page. */
453 if (pp->mp_parent != NULL)
454 NODEPGNO(NODEPTR(pp->mp_parent, pp->mp_pi)) = mp->mp_pgno;
461 mdbenv_sync(MDB_env *env)
464 if (!F_ISSET(env->me_flags, MDB_NOSYNC)) {
465 if (fsync(env->me_fd))
472 mdb_txn_begin(MDB_env *env, int rdonly, MDB_txn **ret)
477 if ((txn = calloc(1, sizeof(*txn))) == NULL) {
478 DPRINTF("calloc: %s", strerror(errno));
483 txn->mt_flags |= MDB_TXN_RDONLY;
485 txn->mt_u.dirty_queue = calloc(1, sizeof(*txn->mt_u.dirty_queue));
486 if (txn->mt_u.dirty_queue == NULL) {
490 SIMPLEQ_INIT(txn->mt_u.dirty_queue);
492 pthread_mutex_lock(&env->me_txns->mt_wmutex);
493 env->me_txns->mt_txnid++;
494 txn->mt_free_pgs = malloc(MDB_IDL_UM_SIZEOF);
495 if (txn->mt_free_pgs == NULL) {
496 free(txn->mt_u.dirty_queue);
500 txn->mt_free_pgs[0] = 0;
502 txn->mt_txnid = env->me_txns->mt_txnid;
504 MDB_reader *r = pthread_getspecific(env->me_txkey);
507 pthread_mutex_lock(&env->me_txns->mt_mutex);
508 for (i=0; i<env->me_maxreaders; i++) {
509 if (env->me_txns->mt_readers[i].mr_pid == 0) {
510 env->me_txns->mt_readers[i].mr_pid = getpid();
511 env->me_txns->mt_readers[i].mr_tid = pthread_self();
512 pthread_setspecific(env->me_txkey, &env->me_txns->mt_readers[i]);
513 if (i >= env->me_txns->mt_numreaders)
514 env->me_txns->mt_numreaders = i+1;
518 pthread_mutex_unlock(&env->me_txns->mt_mutex);
519 if (i == env->me_maxreaders) {
523 r->mr_txnid = txn->mt_txnid;
524 txn->mt_u.reader = r;
531 if ((rc = mdbenv_read_meta(env, &toggle)) != MDB_SUCCESS) {
537 txn->mt_flags |= MDB_TXN_METOGGLE;
539 txn->mt_next_pgno = env->me_meta.mm_last_pg+1;
540 txn->mt_root = env->me_meta.mm_root;
541 DPRINTF("begin transaction %lu on mdbenv %p, root page %lu",
542 txn->mt_txnid, (void *) env, txn->mt_root);
549 mdb_txn_abort(MDB_txn *txn)
558 DPRINTF("abort transaction %lu on mdbenv %p, root page %lu",
559 txn->mt_txnid, (void *) env, txn->mt_root);
561 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
562 txn->mt_u.reader->mr_txnid = 0;
564 /* Discard all dirty pages. Return any re-used pages
567 MDB_IDL_ZERO(txn->mt_free_pgs);
568 while (!SIMPLEQ_EMPTY(txn->mt_u.dirty_queue)) {
569 dp = SIMPLEQ_FIRST(txn->mt_u.dirty_queue);
570 SIMPLEQ_REMOVE_HEAD(txn->mt_u.dirty_queue, h.md_next);
571 if (dp->p.mp_pgno <= env->me_meta.mm_last_pg)
572 mdb_idl_insert(txn->mt_free_pgs, dp->p.mp_pgno);
575 /* put back to head of free list */
576 if (!MDB_IDL_IS_ZERO(txn->mt_free_pgs)) {
579 mop = malloc(sizeof(MDB_oldpages) + MDB_IDL_SIZEOF(txn->mt_free_pgs) - sizeof(pgno_t));
580 mop->mo_next = env->me_pghead;
581 mop->mo_txnid = txn->mt_oldest - 1;
582 if (!env->me_pghead) {
583 env->me_pgtail = mop;
585 env->me_pghead = mop;
586 bcopy(txn->mt_free_pgs, mop->mo_pages, MDB_IDL_SIZEOF(txn->mt_free_pgs));
589 free(txn->mt_free_pgs);
590 free(txn->mt_u.dirty_queue);
592 env->me_txns->mt_txnid--;
593 pthread_mutex_unlock(&env->me_txns->mt_wmutex);
600 mdb_txn_commit(MDB_txn *txn)
608 struct iovec iov[MDB_COMMIT_PAGES];
611 assert(txn->mt_env != NULL);
615 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
616 DPRINTF("attempt to commit read-only transaction");
621 if (txn != env->me_txn) {
622 DPRINTF("attempt to commit unknown transaction");
627 if (F_ISSET(txn->mt_flags, MDB_TXN_ERROR)) {
628 DPRINTF("error flag is set, can't commit");
633 if (SIMPLEQ_EMPTY(txn->mt_u.dirty_queue))
636 DPRINTF("committing transaction %lu on mdbenv %p, root page %lu",
637 txn->mt_txnid, (void *) env, txn->mt_root);
639 /* Commit up to MDB_COMMIT_PAGES dirty pages to disk until done.
646 SIMPLEQ_FOREACH(dp, txn->mt_u.dirty_queue, h.md_next) {
647 if (dp->p.mp_pgno != next) {
649 DPRINTF("committing %u dirty pages", n);
650 rc = writev(env->me_fd, iov, n);
654 DPRINTF("short write, filesystem full?");
656 DPRINTF("writev: %s", strerror(errno));
663 lseek(env->me_fd, dp->p.mp_pgno * env->me_meta.mm_stat.ms_psize, SEEK_SET);
664 next = dp->p.mp_pgno;
666 DPRINTF("committing page %lu", dp->p.mp_pgno);
667 iov[n].iov_len = env->me_meta.mm_stat.ms_psize * dp->h.md_num;
668 iov[n].iov_base = &dp->p;
669 size += iov[n].iov_len;
670 next = dp->p.mp_pgno + dp->h.md_num;
671 /* clear dirty flag */
672 dp->p.mp_flags &= ~P_DIRTY;
673 if (++n >= MDB_COMMIT_PAGES) {
682 DPRINTF("committing %u dirty pages", n);
683 rc = writev(env->me_fd, iov, n);
687 DPRINTF("short write, filesystem full?");
689 DPRINTF("writev: %s", strerror(errno));
696 /* Drop the dirty pages.
698 while (!SIMPLEQ_EMPTY(txn->mt_u.dirty_queue)) {
699 dp = SIMPLEQ_FIRST(txn->mt_u.dirty_queue);
700 SIMPLEQ_REMOVE_HEAD(txn->mt_u.dirty_queue, h.md_next);
704 if ((n = mdbenv_sync(env)) != 0 ||
705 (n = mdbenv_write_meta(txn)) != MDB_SUCCESS ||
706 (n = mdbenv_sync(env)) != 0) {
712 /* add to tail of free list */
713 if (!MDB_IDL_IS_ZERO(txn->mt_free_pgs)) {
716 mop = malloc(sizeof(MDB_oldpages) + MDB_IDL_SIZEOF(txn->mt_free_pgs) - sizeof(pgno_t));
718 if (env->me_pghead) {
719 env->me_pgtail->mo_next = mop;
721 env->me_pghead = mop;
723 env->me_pgtail = mop;
724 bcopy(txn->mt_free_pgs, mop->mo_pages, MDB_IDL_SIZEOF(txn->mt_free_pgs));
725 mop->mo_txnid = txn->mt_txnid;
728 pthread_mutex_unlock(&env->me_txns->mt_wmutex);
729 free(txn->mt_free_pgs);
730 free(txn->mt_u.dirty_queue);
741 mdbenv_read_header(MDB_env *env)
750 /* We don't know the page size yet, so use a minimum value.
753 if ((rc = pread(env->me_fd, page, PAGESIZE, 0)) == 0) {
755 } else if (rc != PAGESIZE) {
758 DPRINTF("read: %s", strerror(errno));
762 p = (MDB_page *)page;
764 if (!F_ISSET(p->mp_flags, P_META)) {
765 DPRINTF("page %lu not a meta page", p->mp_pgno);
770 if (m->mm_magic != MDB_MAGIC) {
771 DPRINTF("meta has invalid magic");
775 if (m->mm_version != MDB_VERSION) {
776 DPRINTF("database is version %u, expected version %u",
777 m->mm_version, MDB_VERSION);
781 bcopy(m, &env->me_meta, sizeof(*m));
786 mdbenv_init_meta(MDB_env *env)
793 DPRINTF("writing new meta page");
794 psize = sysconf(_SC_PAGE_SIZE);
796 env->me_meta.mm_magic = MDB_MAGIC;
797 env->me_meta.mm_version = MDB_VERSION;
798 env->me_meta.mm_stat.ms_psize = psize;
799 env->me_meta.mm_stat.ms_flags = env->me_flags & 0xffff;
800 env->me_meta.mm_root = P_INVALID;
801 env->me_meta.mm_last_pg = 1;
803 p = calloc(2, psize);
805 p->mp_flags = P_META;
808 bcopy(&env->me_meta, meta, sizeof(*meta));
810 q = (MDB_page *)((char *)p + psize);
813 q->mp_flags = P_META;
816 bcopy(&env->me_meta, meta, sizeof(*meta));
818 rc = write(env->me_fd, p, psize * 2);
820 return (rc == (int)psize * 2) ? MDB_SUCCESS : errno;
824 mdbenv_write_meta(MDB_txn *txn)
832 assert(txn->mt_env != NULL);
834 DPRINTF("writing meta page for root page %lu", txn->mt_root);
838 bcopy(&env->me_meta, &meta, sizeof(meta));
839 meta.mm_root = txn->mt_root;
840 meta.mm_last_pg = txn->mt_next_pgno - 1;
841 meta.mm_txnid = txn->mt_txnid;
844 if (!F_ISSET(txn->mt_flags, MDB_TXN_METOGGLE))
845 off = env->me_meta.mm_stat.ms_psize;
848 lseek(env->me_fd, off, SEEK_SET);
849 rc = write(env->me_fd, &meta, sizeof(meta));
850 if (rc != sizeof(meta)) {
851 DPRINTF("write failed, disk error?");
858 /* Returns true if page p is a valid meta page, false otherwise.
861 mdb_check_meta_page(MDB_page *p)
863 if (!F_ISSET(p->mp_flags, P_META)) {
864 DPRINTF("page %lu not a meta page", p->mp_pgno);
872 mdbenv_read_meta(MDB_env *env, int *which)
880 if ((mp0 = mdbenv_get_page(env, 0)) == NULL ||
881 (mp1 = mdbenv_get_page(env, 1)) == NULL)
884 rc = mdb_check_meta_page(mp0);
887 rc = mdb_check_meta_page(mp1);
890 meta[0] = METADATA(mp0);
891 meta[1] = METADATA(mp1);
893 if (meta[0]->mm_txnid < meta[1]->mm_txnid)
896 if (meta[toggle]->mm_txnid > env->me_meta.mm_txnid) {
897 bcopy(meta[toggle], &env->me_meta, sizeof(env->me_meta));
902 DPRINTF("Using meta page %d", toggle);
908 mdbenv_create(MDB_env **env)
912 e = calloc(1, sizeof(*e));
913 if (!e) return ENOMEM;
915 e->me_meta.mm_mapsize = DEFAULT_MAPSIZE;
916 e->me_maxreaders = DEFAULT_READERS;
925 mdbenv_set_mapsize(MDB_env *env, size_t size)
929 env->me_mapsize = env->me_meta.mm_mapsize = size;
934 mdbenv_set_maxreaders(MDB_env *env, int readers)
936 env->me_maxreaders = readers;
941 mdbenv_get_maxreaders(MDB_env *env, int *readers)
943 if (!env || !readers)
945 *readers = env->me_maxreaders;
950 mdbenv_open2(MDB_env *env, unsigned int flags)
954 env->me_flags = flags;
956 if ((i = mdbenv_read_header(env)) != 0) {
959 DPRINTF("new mdbenv");
963 if (!env->me_mapsize)
964 env->me_mapsize = env->me_meta.mm_mapsize;
967 if (env->me_meta.mm_address && (flags & MDB_FIXEDMAP))
969 env->me_map = mmap(env->me_meta.mm_address, env->me_mapsize, PROT_READ, i,
971 if (env->me_map == MAP_FAILED)
975 env->me_meta.mm_mapsize = env->me_mapsize;
976 if (flags & MDB_FIXEDMAP)
977 env->me_meta.mm_address = env->me_map;
978 i = mdbenv_init_meta(env);
979 if (i != MDB_SUCCESS) {
980 munmap(env->me_map, env->me_mapsize);
985 if ((i = mdbenv_read_meta(env, NULL)) != 0)
988 DPRINTF("opened database version %u, pagesize %u",
989 env->me_meta.mm_version, env->me_meta.mm_stat.ms_psize);
990 DPRINTF("depth: %u", env->me_meta.mm_stat.ms_depth);
991 DPRINTF("entries: %lu", env->me_meta.mm_stat.ms_entries);
992 DPRINTF("branch pages: %lu", env->me_meta.mm_stat.ms_branch_pages);
993 DPRINTF("leaf pages: %lu", env->me_meta.mm_stat.ms_leaf_pages);
994 DPRINTF("overflow pages: %lu", env->me_meta.mm_stat.ms_overflow_pages);
995 DPRINTF("root: %lu", env->me_meta.mm_root);
1001 mdbenv_reader_dest(void *ptr)
1003 MDB_reader *reader = ptr;
1005 reader->mr_txnid = 0;
1011 mdbenv_share_locks(MDB_env *env)
1013 struct flock lock_info;
1015 env->me_txns->mt_txnid = env->me_meta.mm_txnid;
1017 memset((void *)&lock_info, 0, sizeof(lock_info));
1018 lock_info.l_type = F_RDLCK;
1019 lock_info.l_whence = SEEK_SET;
1020 lock_info.l_start = 0;
1021 lock_info.l_len = 1;
1022 fcntl(env->me_lfd, F_SETLK, &lock_info);
1026 mdbenv_setup_locks(MDB_env *env, char *lpath, int mode, int *excl)
1030 struct flock lock_info;
1034 if ((env->me_lfd = open(lpath, O_RDWR|O_CREAT, mode)) == -1) {
1038 /* Try to get exclusive lock. If we succeed, then
1039 * nobody is using the lock region and we should initialize it.
1041 memset((void *)&lock_info, 0, sizeof(lock_info));
1042 lock_info.l_type = F_WRLCK;
1043 lock_info.l_whence = SEEK_SET;
1044 lock_info.l_start = 0;
1045 lock_info.l_len = 1;
1046 rc = fcntl(env->me_lfd, F_SETLK, &lock_info);
1050 lock_info.l_type = F_RDLCK;
1051 rc = fcntl(env->me_lfd, F_SETLK, &lock_info);
1057 size = lseek(env->me_lfd, 0, SEEK_END);
1058 rsize = (env->me_maxreaders-1) * sizeof(MDB_reader) + sizeof(MDB_txninfo);
1059 if (size < rsize && *excl) {
1060 if (ftruncate(env->me_lfd, rsize) != 0) {
1066 size = rsize - sizeof(MDB_txninfo);
1067 env->me_maxreaders = size/sizeof(MDB_reader) + 1;
1069 env->me_txns = mmap(0, rsize, PROT_READ|PROT_WRITE, MAP_SHARED,
1071 if (env->me_txns == MAP_FAILED) {
1076 pthread_mutexattr_t mattr;
1078 pthread_mutexattr_init(&mattr);
1079 pthread_mutexattr_setpshared(&mattr, PTHREAD_PROCESS_SHARED);
1080 pthread_mutex_init(&env->me_txns->mt_mutex, &mattr);
1081 pthread_mutex_init(&env->me_txns->mt_wmutex, &mattr);
1082 env->me_txns->mt_version = MDB_VERSION;
1083 env->me_txns->mt_magic = MDB_MAGIC;
1084 env->me_txns->mt_txnid = 0;
1085 env->me_txns->mt_numreaders = 0;
1088 if (env->me_txns->mt_magic != MDB_MAGIC) {
1089 DPRINTF("lock region has invalid magic");
1092 if (env->me_txns->mt_version != MDB_VERSION) {
1093 DPRINTF("lock region is version %u, expected version %u",
1094 env->me_txns->mt_version, MDB_VERSION);
1097 if (errno != EACCES && errno != EAGAIN) {
1111 mdbenv_open(MDB_env *env, const char *path, unsigned int flags, mode_t mode)
1113 int oflags, rc, len, excl;
1114 char *lpath, *dpath;
1117 lpath = malloc(len + sizeof("/lock.mdb") + len + sizeof("/data.db"));
1120 dpath = lpath + len + sizeof("/lock.mdb");
1121 sprintf(lpath, "%s/lock.mdb", path);
1122 sprintf(dpath, "%s/data.mdb", path);
1124 rc = mdbenv_setup_locks(env, lpath, mode, &excl);
1128 if (F_ISSET(flags, MDB_RDONLY))
1131 oflags = O_RDWR | O_CREAT;
1133 if ((env->me_fd = open(dpath, oflags, mode)) == -1)
1136 if ((rc = mdbenv_open2(env, flags)) != MDB_SUCCESS) {
1140 env->me_path = strdup(path);
1141 DPRINTF("opened dbenv %p", (void *) env);
1142 pthread_key_create(&env->me_txkey, mdbenv_reader_dest);
1144 mdbenv_share_locks(env);
1154 mdbenv_close(MDB_env *env)
1162 munmap(env->me_map, env->me_mapsize);
1166 size_t size = (env->me_maxreaders-1) * sizeof(MDB_reader) + sizeof(MDB_txninfo);
1167 munmap(env->me_txns, size);
1173 /* Search for key within a leaf page, using binary search.
1174 * Returns the smallest entry larger or equal to the key.
1175 * If exactp is non-null, stores whether the found entry was an exact match
1176 * in *exactp (1 or 0).
1177 * If kip is non-null, stores the index of the found entry in *kip.
1178 * If no entry larger or equal to the key is found, returns NULL.
1181 mdb_search_node(MDB_db *bt, MDB_page *mp, MDB_val *key,
1182 int *exactp, unsigned int *kip)
1190 DPRINTF("searching %u keys in %s page %lu",
1192 IS_LEAF(mp) ? "leaf" : "branch",
1195 assert(NUMKEYS(mp) > 0);
1197 bzero(&nodekey, sizeof(nodekey));
1199 low = IS_LEAF(mp) ? 0 : 1;
1200 high = NUMKEYS(mp) - 1;
1201 while (low <= high) {
1202 i = (low + high) >> 1;
1203 node = NODEPTR(mp, i);
1205 nodekey.mv_size = node->mn_ksize;
1206 nodekey.mv_data = NODEKEY(node);
1209 rc = bt->md_cmp(key, &nodekey);
1211 rc = _mdb_cmp(bt, key, &nodekey);
1214 DPRINTF("found leaf index %u [%.*s], rc = %i",
1215 i, (int)nodekey.mv_size, (char *)nodekey.mv_data, rc);
1217 DPRINTF("found branch index %u [%.*s -> %lu], rc = %i",
1218 i, (int)node->mn_ksize, (char *)NODEKEY(node),
1229 if (rc > 0) { /* Found entry is less than the key. */
1230 i++; /* Skip to get the smallest entry larger than key. */
1231 if (i >= NUMKEYS(mp))
1232 /* There is no entry larger or equal to the key. */
1236 *exactp = (rc == 0);
1237 if (kip) /* Store the key index if requested. */
1240 return NODEPTR(mp, i);
1244 cursor_pop_page(MDB_cursor *cursor)
1248 top = CURSOR_TOP(cursor);
1251 DPRINTF("popped page %lu off cursor %p", top->mp_page->mp_pgno, (void *) cursor);
1257 cursor_push_page(MDB_cursor *cursor, MDB_page *mp)
1261 DPRINTF("pushing page %lu on cursor %p", mp->mp_pgno, (void *) cursor);
1263 if ((ppage = calloc(1, sizeof(*ppage))) == NULL)
1265 ppage->mp_page = mp;
1266 CURSOR_PUSH(cursor, ppage);
1271 mdbenv_get_page(MDB_env *env, pgno_t pgno)
1274 MDB_txn *txn = env->me_txn;
1277 if (txn && !SIMPLEQ_EMPTY(txn->mt_u.dirty_queue)) {
1279 SIMPLEQ_FOREACH(dp, txn->mt_u.dirty_queue, h.md_next) {
1280 if (dp->p.mp_pgno == pgno) {
1288 p = (MDB_page *)(env->me_map + env->me_meta.mm_stat.ms_psize * pgno);
1294 mdb_search_page_root(MDB_db *bt, MDB_val *key,
1295 MDB_cursor *cursor, int modify, MDB_pageparent *mpp)
1297 MDB_page *mp = mpp->mp_page;
1300 if (cursor && cursor_push_page(cursor, mp) == NULL)
1303 while (IS_BRANCH(mp)) {
1307 DPRINTF("branch page %lu has %u keys", mp->mp_pgno, NUMKEYS(mp));
1308 assert(NUMKEYS(mp) > 1);
1309 DPRINTF("found index 0 to page %lu", NODEPGNO(NODEPTR(mp, 0)));
1311 if (key == NULL) /* Initialize cursor to first page. */
1315 node = mdb_search_node(bt, mp, key, &exact, &i);
1317 i = NUMKEYS(mp) - 1;
1325 DPRINTF("following index %u for key %.*s",
1326 i, (int)key->mv_size, (char *)key->mv_data);
1327 assert(i < NUMKEYS(mp));
1328 node = NODEPTR(mp, i);
1331 CURSOR_TOP(cursor)->mp_ki = i;
1333 mpp->mp_parent = mp;
1334 if ((mp = mdbenv_get_page(bt->md_env, NODEPGNO(node))) == NULL)
1339 if (cursor && cursor_push_page(cursor, mp) == NULL)
1343 MDB_dhead *dh = ((MDB_dhead *)mp)-1;
1344 if ((rc = mdb_touch(bt->md_env->me_txn, mpp)) != 0)
1346 dh = ((MDB_dhead *)mpp->mp_page)-1;
1347 dh->md_parent = mpp->mp_parent;
1348 dh->md_pi = mpp->mp_pi;
1355 DPRINTF("internal error, index points to a %02X page!?",
1360 DPRINTF("found leaf page %lu for key %.*s", mp->mp_pgno,
1361 key ? (int)key->mv_size : 0, key ? (char *)key->mv_data : NULL);
1366 /* Search for the page a given key should be in.
1367 * Stores a pointer to the found page in *mpp.
1368 * If key is NULL, search for the lowest page (used by mdb_cursor_first).
1369 * If cursor is non-null, pushes parent pages on the cursor stack.
1370 * If modify is true, visited pages are updated with new page numbers.
1373 mdb_search_page(MDB_db *db, MDB_txn *txn, MDB_val *key,
1374 MDB_cursor *cursor, int modify, MDB_pageparent *mpp)
1379 /* Can't modify pages outside a transaction. */
1380 if (txn == NULL && modify) {
1384 /* Choose which root page to start with. If a transaction is given
1385 * use the root page from the transaction, otherwise read the last
1386 * committed root page.
1389 if ((rc = mdbenv_read_meta(db->md_env, NULL)) != MDB_SUCCESS)
1391 root = db->md_env->me_meta.mm_root;
1392 } else if (F_ISSET(txn->mt_flags, MDB_TXN_ERROR)) {
1393 DPRINTF("transaction has failed, must abort");
1396 root = txn->mt_root;
1398 if (root == P_INVALID) { /* Tree is empty. */
1399 DPRINTF("tree is empty");
1403 if ((mpp->mp_page = mdbenv_get_page(db->md_env, root)) == NULL)
1406 DPRINTF("root page has flags 0x%X", mpp->mp_page->mp_flags);
1408 if (modify && !F_ISSET(mpp->mp_page->mp_flags, P_DIRTY)) {
1409 mpp->mp_parent = NULL;
1411 if ((rc = mdb_touch(txn, mpp)))
1413 txn->mt_root = mpp->mp_page->mp_pgno;
1416 return mdb_search_page_root(db, key, cursor, modify, mpp);
1420 mdb_read_data(MDB_db *db, MDB_page *mp, MDB_node *leaf,
1423 MDB_page *omp; /* overflow mpage */
1427 bzero(data, sizeof(*data));
1428 max = db->md_env->me_meta.mm_stat.ms_psize - PAGEHDRSZ;
1430 if (!F_ISSET(leaf->mn_flags, F_BIGDATA)) {
1431 data->mv_size = leaf->mn_dsize;
1432 data->mv_data = NODEDATA(leaf);
1436 /* Read overflow data.
1438 data->mv_size = leaf->mn_dsize;
1439 bcopy(NODEDATA(leaf), &pgno, sizeof(pgno));
1440 if ((omp = mdbenv_get_page(db->md_env, pgno)) == NULL) {
1441 DPRINTF("read overflow page %lu failed", pgno);
1444 data->mv_data = omp;
1450 mdb_get(MDB_db *db, MDB_txn *txn,
1451 MDB_val *key, MDB_val *data)
1459 DPRINTF("===> get key [%.*s]", (int)key->mv_size, (char *)key->mv_data);
1464 if (txn != NULL && db->md_env != txn->mt_env) {
1468 if (key->mv_size == 0 || key->mv_size > MAXKEYSIZE) {
1472 if ((rc = mdb_search_page(db, txn, key, NULL, 0, &mpp)) != MDB_SUCCESS)
1475 leaf = mdb_search_node(db, mpp.mp_page, key, &exact, NULL);
1477 rc = mdb_read_data(db, mpp.mp_page, leaf, data);
1486 mdb_sibling(MDB_cursor *cursor, int move_right)
1490 MDB_ppage *parent, *top;
1493 top = CURSOR_TOP(cursor);
1494 if ((parent = SLIST_NEXT(top, mp_entry)) == NULL) {
1495 return ENOENT; /* root has no siblings */
1498 DPRINTF("parent page is page %lu, index %u",
1499 parent->mp_page->mp_pgno, parent->mp_ki);
1501 cursor_pop_page(cursor);
1502 if (move_right ? (parent->mp_ki + 1 >= NUMKEYS(parent->mp_page))
1503 : (parent->mp_ki == 0)) {
1504 DPRINTF("no more keys left, moving to %s sibling",
1505 move_right ? "right" : "left");
1506 if ((rc = mdb_sibling(cursor, move_right)) != MDB_SUCCESS)
1508 parent = CURSOR_TOP(cursor);
1514 DPRINTF("just moving to %s index key %u",
1515 move_right ? "right" : "left", parent->mp_ki);
1517 assert(IS_BRANCH(parent->mp_page));
1519 indx = NODEPTR(parent->mp_page, parent->mp_ki);
1520 if ((mp = mdbenv_get_page(cursor->mc_db->md_env, indx->mn_pgno)) == NULL)
1523 mp->parent = parent->mp_page;
1524 mp->parent_index = parent->mp_ki;
1527 cursor_push_page(cursor, mp);
1533 mdb_set_key(MDB_db *bt, MDB_page *mp, MDB_node *node,
1539 key->mv_size = node->mn_ksize;
1540 key->mv_data = NODEKEY(node);
1546 mdb_cursor_next(MDB_cursor *cursor, MDB_val *key, MDB_val *data)
1552 if (cursor->mc_eof) {
1556 assert(cursor->mc_initialized);
1558 top = CURSOR_TOP(cursor);
1561 DPRINTF("cursor_next: top page is %lu in cursor %p", mp->mp_pgno, (void *) cursor);
1563 if (top->mp_ki + 1 >= NUMKEYS(mp)) {
1564 DPRINTF("=====> move to next sibling page");
1565 if (mdb_sibling(cursor, 1) != MDB_SUCCESS) {
1569 top = CURSOR_TOP(cursor);
1571 DPRINTF("next page is %lu, key index %u", mp->mp_pgno, top->mp_ki);
1575 DPRINTF("==> cursor points to page %lu with %u keys, key index %u",
1576 mp->mp_pgno, NUMKEYS(mp), top->mp_ki);
1578 assert(IS_LEAF(mp));
1579 leaf = NODEPTR(mp, top->mp_ki);
1581 if (data && mdb_read_data(cursor->mc_db, mp, leaf, data) != MDB_SUCCESS)
1584 return mdb_set_key(cursor->mc_db, mp, leaf, key);
1588 mdb_cursor_set(MDB_cursor *cursor, MDB_val *key, MDB_val *data,
1598 assert(key->mv_size > 0);
1600 rc = mdb_search_page(cursor->mc_db, cursor->mc_txn, key, cursor, 0, &mpp);
1601 if (rc != MDB_SUCCESS)
1603 assert(IS_LEAF(mpp.mp_page));
1605 top = CURSOR_TOP(cursor);
1606 leaf = mdb_search_node(cursor->mc_db, mpp.mp_page, key, exactp, &top->mp_ki);
1607 if (exactp != NULL && !*exactp) {
1608 /* MDB_CURSOR_EXACT specified and not an exact match. */
1613 DPRINTF("===> inexact leaf not found, goto sibling");
1614 if ((rc = mdb_sibling(cursor, 1)) != MDB_SUCCESS)
1615 return rc; /* no entries matched */
1616 top = CURSOR_TOP(cursor);
1618 mpp.mp_page = top->mp_page;
1619 assert(IS_LEAF(mpp.mp_page));
1620 leaf = NODEPTR(mpp.mp_page, 0);
1623 cursor->mc_initialized = 1;
1626 if (data && (rc = mdb_read_data(cursor->mc_db, mpp.mp_page, leaf, data)) != MDB_SUCCESS)
1629 rc = mdb_set_key(cursor->mc_db, mpp.mp_page, leaf, key);
1630 if (rc == MDB_SUCCESS) {
1631 DPRINTF("==> cursor placed on key %.*s",
1632 (int)key->mv_size, (char *)key->mv_data);
1640 mdb_cursor_first(MDB_cursor *cursor, MDB_val *key, MDB_val *data)
1646 rc = mdb_search_page(cursor->mc_db, cursor->mc_txn, NULL, cursor, 0, &mpp);
1647 if (rc != MDB_SUCCESS)
1649 assert(IS_LEAF(mpp.mp_page));
1651 leaf = NODEPTR(mpp.mp_page, 0);
1652 cursor->mc_initialized = 1;
1655 if (data && (rc = mdb_read_data(cursor->mc_db, mpp.mp_page, leaf, data)) != MDB_SUCCESS)
1658 return mdb_set_key(cursor->mc_db, mpp.mp_page, leaf, key);
1662 mdb_cursor_get(MDB_cursor *cursor, MDB_val *key, MDB_val *data,
1672 case MDB_CURSOR_EXACT:
1673 while (CURSOR_TOP(cursor) != NULL)
1674 cursor_pop_page(cursor);
1675 if (key == NULL || key->mv_size == 0 || key->mv_size > MAXKEYSIZE) {
1677 } else if (op == MDB_CURSOR_EXACT)
1678 rc = mdb_cursor_set(cursor, key, data, &exact);
1680 rc = mdb_cursor_set(cursor, key, data, NULL);
1683 if (!cursor->mc_initialized)
1684 rc = mdb_cursor_first(cursor, key, data);
1686 rc = mdb_cursor_next(cursor, key, data);
1689 while (CURSOR_TOP(cursor) != NULL)
1690 cursor_pop_page(cursor);
1691 rc = mdb_cursor_first(cursor, key, data);
1694 DPRINTF("unhandled/unimplemented cursor operation %u", op);
1702 /* Allocate a page and initialize it
1705 mdbenv_new_page(MDB_env *env, uint32_t flags, int num)
1709 assert(env != NULL);
1710 assert(env->me_txn != NULL);
1712 if ((dp = mdb_newpage(env->me_txn, NULL, 0, num)) == NULL)
1714 DPRINTF("allocated new mpage %lu, page size %u",
1715 dp->p.mp_pgno, env->me_meta.mm_stat.ms_psize);
1716 dp->p.mp_flags = flags | P_DIRTY;
1717 dp->p.mp_lower = PAGEHDRSZ;
1718 dp->p.mp_upper = env->me_meta.mm_stat.ms_psize;
1720 if (IS_BRANCH(&dp->p))
1721 env->me_meta.mm_stat.ms_branch_pages++;
1722 else if (IS_LEAF(&dp->p))
1723 env->me_meta.mm_stat.ms_leaf_pages++;
1724 else if (IS_OVERFLOW(&dp->p)) {
1725 env->me_meta.mm_stat.ms_overflow_pages += num;
1726 dp->p.mp_pages = num;
1733 mdb_leaf_size(MDB_db *db, MDB_val *key, MDB_val *data)
1737 sz = LEAFSIZE(key, data);
1738 if (data->mv_size >= db->md_env->me_meta.mm_stat.ms_psize / MDB_MINKEYS) {
1739 /* put on overflow page */
1740 sz -= data->mv_size - sizeof(pgno_t);
1743 return sz + sizeof(indx_t);
1747 mdb_branch_size(MDB_db *db, MDB_val *key)
1752 if (sz >= db->md_env->me_meta.mm_stat.ms_psize / MDB_MINKEYS) {
1753 /* put on overflow page */
1754 /* not implemented */
1755 /* sz -= key->size - sizeof(pgno_t); */
1758 return sz + sizeof(indx_t);
1762 mdb_add_node(MDB_db *db, MDB_page *mp, indx_t indx,
1763 MDB_val *key, MDB_val *data, pgno_t pgno, uint8_t flags)
1766 size_t node_size = NODESIZE;
1769 MDB_dpage *ofp = NULL; /* overflow page */
1771 assert(mp->mp_upper >= mp->mp_lower);
1773 DPRINTF("add node [%.*s] to %s page %lu at index %i, key size %zu",
1774 key ? (int)key->mv_size : 0, key ? (char *)key->mv_data : NULL,
1775 IS_LEAF(mp) ? "leaf" : "branch",
1776 mp->mp_pgno, indx, key ? key->mv_size : 0);
1779 node_size += key->mv_size;
1783 if (F_ISSET(flags, F_BIGDATA)) {
1784 /* Data already on overflow page. */
1785 node_size += sizeof(pgno_t);
1786 } else if (data->mv_size >= db->md_env->me_meta.mm_stat.ms_psize / MDB_MINKEYS) {
1787 int ovpages = OVPAGES(data->mv_size, db->md_env->me_meta.mm_stat.ms_psize);
1788 /* Put data on overflow page. */
1789 DPRINTF("data size is %zu, put on overflow page",
1791 node_size += sizeof(pgno_t);
1792 if ((ofp = mdbenv_new_page(db->md_env, P_OVERFLOW, ovpages)) == NULL)
1794 DPRINTF("allocated overflow page %lu", ofp->p.mp_pgno);
1797 node_size += data->mv_size;
1801 if (node_size + sizeof(indx_t) > SIZELEFT(mp)) {
1802 DPRINTF("not enough room in page %lu, got %u ptrs",
1803 mp->mp_pgno, NUMKEYS(mp));
1804 DPRINTF("upper - lower = %u - %u = %u", mp->mp_upper, mp->mp_lower,
1805 mp->mp_upper - mp->mp_lower);
1806 DPRINTF("node size = %zu", node_size);
1810 /* Move higher pointers up one slot. */
1811 for (i = NUMKEYS(mp); i > indx; i--)
1812 mp->mp_ptrs[i] = mp->mp_ptrs[i - 1];
1814 /* Adjust free space offsets. */
1815 ofs = mp->mp_upper - node_size;
1816 assert(ofs >= mp->mp_lower + sizeof(indx_t));
1817 mp->mp_ptrs[indx] = ofs;
1819 mp->mp_lower += sizeof(indx_t);
1821 /* Write the node data. */
1822 node = NODEPTR(mp, indx);
1823 node->mn_ksize = (key == NULL) ? 0 : key->mv_size;
1824 node->mn_flags = flags;
1826 node->mn_dsize = data->mv_size;
1828 node->mn_pgno = pgno;
1831 bcopy(key->mv_data, NODEKEY(node), key->mv_size);
1836 if (F_ISSET(flags, F_BIGDATA))
1837 bcopy(data->mv_data, node->mn_data + key->mv_size,
1840 bcopy(data->mv_data, node->mn_data + key->mv_size,
1843 bcopy(&ofp->p.mp_pgno, node->mn_data + key->mv_size,
1845 bcopy(data->mv_data, METADATA(&ofp->p), data->mv_size);
1853 mdb_del_node(MDB_db *db, MDB_page *mp, indx_t indx)
1856 indx_t i, j, numkeys, ptr;
1860 DPRINTF("delete node %u on %s page %lu", indx,
1861 IS_LEAF(mp) ? "leaf" : "branch", mp->mp_pgno);
1862 assert(indx < NUMKEYS(mp));
1864 node = NODEPTR(mp, indx);
1865 sz = NODESIZE + node->mn_ksize;
1867 if (F_ISSET(node->mn_flags, F_BIGDATA))
1868 sz += sizeof(pgno_t);
1870 sz += NODEDSZ(node);
1873 ptr = mp->mp_ptrs[indx];
1874 numkeys = NUMKEYS(mp);
1875 for (i = j = 0; i < numkeys; i++) {
1877 mp->mp_ptrs[j] = mp->mp_ptrs[i];
1878 if (mp->mp_ptrs[i] < ptr)
1879 mp->mp_ptrs[j] += sz;
1884 base = (char *)mp + mp->mp_upper;
1885 bcopy(base, base + sz, ptr - mp->mp_upper);
1887 mp->mp_lower -= sizeof(indx_t);
1892 mdb_cursor_open(MDB_db *db, MDB_txn *txn, MDB_cursor **ret)
1896 if (db == NULL || ret == NULL)
1899 if (txn != NULL && db->md_env != txn->mt_env) {
1903 if ((cursor = calloc(1, sizeof(*cursor))) != NULL) {
1904 SLIST_INIT(&cursor->mc_stack);
1906 cursor->mc_txn = txn;
1915 mdb_cursor_close(MDB_cursor *cursor)
1917 if (cursor != NULL) {
1918 while (!CURSOR_EMPTY(cursor))
1919 cursor_pop_page(cursor);
1921 /* btree_close(cursor->bt); */
1927 mdb_update_key(MDB_db *db, MDB_page *mp, indx_t indx,
1930 indx_t ptr, i, numkeys;
1936 node = NODEPTR(mp, indx);
1937 ptr = mp->mp_ptrs[indx];
1938 DPRINTF("update key %u (ofs %u) [%.*s] to [%.*s] on page %lu",
1940 (int)node->mn_ksize, (char *)NODEKEY(node),
1941 (int)key->mv_size, (char *)key->mv_data,
1944 delta = key->mv_size - node->mn_ksize;
1946 if (delta > 0 && SIZELEFT(mp) < delta) {
1947 DPRINTF("OUCH! Not enough room, delta = %d", delta);
1951 numkeys = NUMKEYS(mp);
1952 for (i = 0; i < numkeys; i++) {
1953 if (mp->mp_ptrs[i] <= ptr)
1954 mp->mp_ptrs[i] -= delta;
1957 base = (char *)mp + mp->mp_upper;
1958 len = ptr - mp->mp_upper + NODESIZE;
1959 bcopy(base, base - delta, len);
1960 mp->mp_upper -= delta;
1962 node = NODEPTR(mp, indx);
1963 node->mn_ksize = key->mv_size;
1966 bcopy(key->mv_data, NODEKEY(node), key->mv_size);
1971 /* Move a node from src to dst.
1974 mdb_move_node(MDB_db *bt, MDB_pageparent *src, indx_t srcindx,
1975 MDB_pageparent *dst, indx_t dstindx)
1981 srcnode = NODEPTR(src->mp_page, srcindx);
1982 DPRINTF("moving %s node %u [%.*s] on page %lu to node %u on page %lu",
1983 IS_LEAF(src->mp_page) ? "leaf" : "branch",
1985 (int)srcnode->mn_ksize, (char *)NODEKEY(srcnode),
1986 src->mp_page->mp_pgno,
1987 dstindx, dst->mp_page->mp_pgno);
1989 /* Mark src and dst as dirty. */
1990 if ((rc = mdb_touch(bt->md_env->me_txn, src)) ||
1991 (rc = mdb_touch(bt->md_env->me_txn, dst)))
1994 /* Add the node to the destination page.
1996 key.mv_size = srcnode->mn_ksize;
1997 key.mv_data = NODEKEY(srcnode);
1998 data.mv_size = NODEDSZ(srcnode);
1999 data.mv_data = NODEDATA(srcnode);
2000 rc = mdb_add_node(bt, dst->mp_page, dstindx, &key, &data, NODEPGNO(srcnode),
2002 if (rc != MDB_SUCCESS)
2005 /* Delete the node from the source page.
2007 mdb_del_node(bt, src->mp_page, srcindx);
2009 /* Update the parent separators.
2011 if (srcindx == 0 && src->mp_pi != 0) {
2012 DPRINTF("update separator for source page %lu to [%.*s]",
2013 src->mp_page->mp_pgno, (int)key.mv_size, (char *)key.mv_data);
2014 if ((rc = mdb_update_key(bt, src->mp_parent, src->mp_pi,
2015 &key)) != MDB_SUCCESS)
2019 if (srcindx == 0 && IS_BRANCH(src->mp_page)) {
2021 nullkey.mv_size = 0;
2022 assert(mdb_update_key(bt, src->mp_page, 0, &nullkey) == MDB_SUCCESS);
2025 if (dstindx == 0 && dst->mp_pi != 0) {
2026 DPRINTF("update separator for destination page %lu to [%.*s]",
2027 dst->mp_page->mp_pgno, (int)key.mv_size, (char *)key.mv_data);
2028 if ((rc = mdb_update_key(bt, dst->mp_parent, dst->mp_pi,
2029 &key)) != MDB_SUCCESS)
2033 if (dstindx == 0 && IS_BRANCH(dst->mp_page)) {
2035 nullkey.mv_size = 0;
2036 assert(mdb_update_key(bt, dst->mp_page, 0, &nullkey) == MDB_SUCCESS);
2043 mdb_merge(MDB_db *bt, MDB_pageparent *src, MDB_pageparent *dst)
2052 DPRINTF("merging page %lu and %lu", src->mp_page->mp_pgno, dst->mp_page->mp_pgno);
2054 assert(src->mp_parent); /* can't merge root page */
2055 assert(dst->mp_parent);
2056 assert(bt->md_env->me_txn != NULL);
2058 /* Mark src and dst as dirty. */
2059 if ((rc = mdb_touch(bt->md_env->me_txn, src)) ||
2060 (rc = mdb_touch(bt->md_env->me_txn, dst)))
2063 /* Move all nodes from src to dst.
2065 for (i = 0; i < NUMKEYS(src->mp_page); i++) {
2066 srcnode = NODEPTR(src->mp_page, i);
2068 key.mv_size = srcnode->mn_ksize;
2069 key.mv_data = NODEKEY(srcnode);
2070 data.mv_size = NODEDSZ(srcnode);
2071 data.mv_data = NODEDATA(srcnode);
2072 rc = mdb_add_node(bt, dst->mp_page, NUMKEYS(dst->mp_page), &key,
2073 &data, NODEPGNO(srcnode), srcnode->mn_flags);
2074 if (rc != MDB_SUCCESS)
2078 DPRINTF("dst page %lu now has %u keys (%.1f%% filled)",
2079 dst->mp_page->mp_pgno, NUMKEYS(dst->mp_page), (float)PAGEFILL(bt->md_env, dst->mp_page) / 10);
2081 /* Unlink the src page from parent.
2083 mdb_del_node(bt, src->mp_parent, src->mp_pi);
2084 if (src->mp_pi == 0) {
2086 if ((rc = mdb_update_key(bt, src->mp_parent, 0, &key)) != MDB_SUCCESS)
2090 if (IS_LEAF(src->mp_page))
2091 bt->md_env->me_meta.mm_stat.ms_leaf_pages--;
2093 bt->md_env->me_meta.mm_stat.ms_branch_pages--;
2095 mpp.mp_page = src->mp_parent;
2096 dh = (MDB_dhead *)src->mp_parent;
2098 mpp.mp_parent = dh->md_parent;
2099 mpp.mp_pi = dh->md_pi;
2101 return mdb_rebalance(bt, &mpp);
2104 #define FILL_THRESHOLD 250
2107 mdb_rebalance(MDB_db *db, MDB_pageparent *mpp)
2112 indx_t si = 0, di = 0;
2115 assert(db->md_env->me_txn != NULL);
2116 assert(mpp != NULL);
2118 DPRINTF("rebalancing %s page %lu (has %u keys, %.1f%% full)",
2119 IS_LEAF(mpp->mp_page) ? "leaf" : "branch",
2120 mpp->mp_page->mp_pgno, NUMKEYS(mpp->mp_page), (float)PAGEFILL(db->md_env, mpp->mp_page) / 10);
2122 if (PAGEFILL(db->md_env, mpp->mp_page) >= FILL_THRESHOLD) {
2123 DPRINTF("no need to rebalance page %lu, above fill threshold",
2124 mpp->mp_page->mp_pgno);
2128 if (mpp->mp_parent == NULL) {
2129 if (NUMKEYS(mpp->mp_page) == 0) {
2130 DPRINTF("tree is completely empty");
2131 db->md_env->me_txn->mt_root = P_INVALID;
2132 db->md_env->me_meta.mm_stat.ms_depth--;
2133 db->md_env->me_meta.mm_stat.ms_leaf_pages--;
2134 } else if (IS_BRANCH(mpp->mp_page) && NUMKEYS(mpp->mp_page) == 1) {
2135 DPRINTF("collapsing root page!");
2136 db->md_env->me_txn->mt_root = NODEPGNO(NODEPTR(mpp->mp_page, 0));
2137 if ((root = mdbenv_get_page(db->md_env, db->md_env->me_txn->mt_root)) == NULL)
2139 db->md_env->me_meta.mm_stat.ms_depth--;
2140 db->md_env->me_meta.mm_stat.ms_branch_pages--;
2142 DPRINTF("root page doesn't need rebalancing");
2146 /* The parent (branch page) must have at least 2 pointers,
2147 * otherwise the tree is invalid.
2149 assert(NUMKEYS(mpp->mp_parent) > 1);
2151 /* Leaf page fill factor is below the threshold.
2152 * Try to move keys from left or right neighbor, or
2153 * merge with a neighbor page.
2158 if (mpp->mp_pi == 0) {
2159 /* We're the leftmost leaf in our parent.
2161 DPRINTF("reading right neighbor");
2162 node = NODEPTR(mpp->mp_parent, mpp->mp_pi + 1);
2163 if ((npp.mp_page = mdbenv_get_page(db->md_env, NODEPGNO(node))) == NULL)
2165 npp.mp_pi = mpp->mp_pi + 1;
2167 di = NUMKEYS(mpp->mp_page);
2169 /* There is at least one neighbor to the left.
2171 DPRINTF("reading left neighbor");
2172 node = NODEPTR(mpp->mp_parent, mpp->mp_pi - 1);
2173 if ((npp.mp_page = mdbenv_get_page(db->md_env, NODEPGNO(node))) == NULL)
2175 npp.mp_pi = mpp->mp_pi - 1;
2176 si = NUMKEYS(npp.mp_page) - 1;
2179 npp.mp_parent = mpp->mp_parent;
2181 DPRINTF("found neighbor page %lu (%u keys, %.1f%% full)",
2182 npp.mp_page->mp_pgno, NUMKEYS(npp.mp_page), (float)PAGEFILL(db->md_env, npp.mp_page) / 10);
2184 /* If the neighbor page is above threshold and has at least two
2185 * keys, move one key from it.
2187 * Otherwise we should try to merge them.
2189 if (PAGEFILL(db->md_env, npp.mp_page) >= FILL_THRESHOLD && NUMKEYS(npp.mp_page) >= 2)
2190 return mdb_move_node(db, &npp, si, mpp, di);
2191 else { /* FIXME: if (has_enough_room()) */
2192 if (mpp->mp_pi == 0)
2193 return mdb_merge(db, &npp, mpp);
2195 return mdb_merge(db, mpp, &npp);
2200 mdb_del(MDB_db *bt, MDB_txn *txn,
2201 MDB_val *key, MDB_val *data)
2208 DPRINTF("========> delete key %.*s", (int)key->mv_size, (char *)key->mv_data);
2210 assert(key != NULL);
2212 if (bt == NULL || txn == NULL)
2215 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
2219 if (bt->md_env->me_txn != txn) {
2223 if (key->mv_size == 0 || key->mv_size > MAXKEYSIZE) {
2227 if ((rc = mdb_search_page(bt, txn, key, NULL, 1, &mpp)) != MDB_SUCCESS)
2230 leaf = mdb_search_node(bt, mpp.mp_page, key, &exact, &ki);
2231 if (leaf == NULL || !exact) {
2235 if (data && (rc = mdb_read_data(bt, NULL, leaf, data)) != MDB_SUCCESS)
2238 mdb_del_node(bt, mpp.mp_page, ki);
2239 /* add overflow pages to free list */
2240 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) {
2244 bcopy(NODEDATA(leaf), &pg, sizeof(pg));
2245 ovpages = OVPAGES(NODEDSZ(leaf), bt->md_env->me_meta.mm_stat.ms_psize);
2246 for (i=0; i<ovpages; i++) {
2247 mdb_idl_insert(txn->mt_free_pgs, pg);
2251 bt->md_env->me_meta.mm_stat.ms_entries--;
2252 rc = mdb_rebalance(bt, &mpp);
2253 if (rc != MDB_SUCCESS)
2254 txn->mt_flags |= MDB_TXN_ERROR;
2259 /* Split page <*mpp>, and insert <key,(data|newpgno)> in either left or
2260 * right sibling, at index <*newindxp> (as if unsplit). Updates *mpp and
2261 * *newindxp with the actual values after split, ie if *mpp and *newindxp
2262 * refer to a node in the new right sibling page.
2265 mdb_split(MDB_db *bt, MDB_page **mpp, unsigned int *newindxp,
2266 MDB_val *newkey, MDB_val *newdata, pgno_t newpgno)
2269 int rc = MDB_SUCCESS, ins_new = 0;
2272 unsigned int i, j, split_indx;
2274 MDB_val sepkey, rkey, rdata;
2276 MDB_dpage *mdp, *rdp, *pdp;
2280 assert(bt->md_env != NULL);
2282 dh = ((MDB_dhead *)*mpp) - 1;
2283 mdp = (MDB_dpage *)dh;
2284 newindx = *newindxp;
2286 DPRINTF("-----> splitting %s page %lu and adding [%.*s] at index %i",
2287 IS_LEAF(&mdp->p) ? "leaf" : "branch", mdp->p.mp_pgno,
2288 (int)newkey->mv_size, (char *)newkey->mv_data, *newindxp);
2290 if (mdp->h.md_parent == NULL) {
2291 if ((pdp = mdbenv_new_page(bt->md_env, P_BRANCH, 1)) == NULL)
2294 mdp->h.md_parent = &pdp->p;
2295 bt->md_env->me_txn->mt_root = pdp->p.mp_pgno;
2296 DPRINTF("root split! new root = %lu", pdp->p.mp_pgno);
2297 bt->md_env->me_meta.mm_stat.ms_depth++;
2299 /* Add left (implicit) pointer. */
2300 if (mdb_add_node(bt, &pdp->p, 0, NULL, NULL,
2301 mdp->p.mp_pgno, 0) != MDB_SUCCESS)
2304 DPRINTF("parent branch page is %lu", mdp->h.md_parent->mp_pgno);
2307 /* Create a right sibling. */
2308 if ((rdp = mdbenv_new_page(bt->md_env, mdp->p.mp_flags, 1)) == NULL)
2310 rdp->h.md_parent = mdp->h.md_parent;
2311 rdp->h.md_pi = mdp->h.md_pi + 1;
2312 DPRINTF("new right sibling: page %lu", rdp->p.mp_pgno);
2314 /* Move half of the keys to the right sibling. */
2315 if ((copy = malloc(bt->md_env->me_meta.mm_stat.ms_psize)) == NULL)
2317 bcopy(&mdp->p, copy, bt->md_env->me_meta.mm_stat.ms_psize);
2318 bzero(&mdp->p.mp_ptrs, bt->md_env->me_meta.mm_stat.ms_psize - PAGEHDRSZ);
2319 mdp->p.mp_lower = PAGEHDRSZ;
2320 mdp->p.mp_upper = bt->md_env->me_meta.mm_stat.ms_psize;
2322 split_indx = NUMKEYS(copy) / 2 + 1;
2324 /* First find the separating key between the split pages.
2326 bzero(&sepkey, sizeof(sepkey));
2327 if (newindx == split_indx) {
2328 sepkey.mv_size = newkey->mv_size;
2329 sepkey.mv_data = newkey->mv_data;
2331 node = NODEPTR(copy, split_indx);
2332 sepkey.mv_size = node->mn_ksize;
2333 sepkey.mv_data = NODEKEY(node);
2336 DPRINTF("separator is [%.*s]", (int)sepkey.mv_size, (char *)sepkey.mv_data);
2338 /* Copy separator key to the parent.
2340 if (SIZELEFT(rdp->h.md_parent) < mdb_branch_size(bt, &sepkey)) {
2341 rc = mdb_split(bt, &rdp->h.md_parent, &rdp->h.md_pi,
2342 &sepkey, NULL, rdp->p.mp_pgno);
2344 /* Right page might now have changed parent.
2345 * Check if left page also changed parent.
2347 if (rdp->h.md_parent != mdp->h.md_parent &&
2348 mdp->h.md_pi >= NUMKEYS(mdp->h.md_parent)) {
2349 mdp->h.md_parent = rdp->h.md_parent;
2350 mdp->h.md_pi = rdp->h.md_pi - 1;
2353 rc = mdb_add_node(bt, rdp->h.md_parent, rdp->h.md_pi,
2354 &sepkey, NULL, rdp->p.mp_pgno, 0);
2356 if (rc != MDB_SUCCESS) {
2361 for (i = j = 0; i <= NUMKEYS(copy); j++) {
2362 if (i < split_indx) {
2363 /* Re-insert in left sibling. */
2366 /* Insert in right sibling. */
2367 if (i == split_indx)
2368 /* Reset insert index for right sibling. */
2369 j = (i == newindx && ins_new);
2373 if (i == newindx && !ins_new) {
2374 /* Insert the original entry that caused the split. */
2375 rkey.mv_data = newkey->mv_data;
2376 rkey.mv_size = newkey->mv_size;
2377 if (IS_LEAF(&mdp->p)) {
2378 rdata.mv_data = newdata->mv_data;
2379 rdata.mv_size = newdata->mv_size;
2386 /* Update page and index for the new key. */
2389 } else if (i == NUMKEYS(copy)) {
2392 node = NODEPTR(copy, i);
2393 rkey.mv_data = NODEKEY(node);
2394 rkey.mv_size = node->mn_ksize;
2395 if (IS_LEAF(&mdp->p)) {
2396 rdata.mv_data = NODEDATA(node);
2397 rdata.mv_size = node->mn_dsize;
2399 pgno = node->mn_pgno;
2400 flags = node->mn_flags;
2405 if (!IS_LEAF(&mdp->p) && j == 0) {
2406 /* First branch index doesn't need key data. */
2410 rc = mdb_add_node(bt, &pdp->p, j, &rkey, &rdata, pgno,flags);
2418 mdb_put(MDB_db *bt, MDB_txn *txn,
2419 MDB_val *key, MDB_val *data, unsigned int flags)
2421 int rc = MDB_SUCCESS, exact;
2426 assert(key != NULL);
2427 assert(data != NULL);
2429 if (bt == NULL || txn == NULL)
2432 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
2436 if (bt->md_env->me_txn != txn) {
2440 if (key->mv_size == 0 || key->mv_size > MAXKEYSIZE) {
2444 DPRINTF("==> put key %.*s, size %zu, data size %zu",
2445 (int)key->mv_size, (char *)key->mv_data, key->mv_size, data->mv_size);
2447 rc = mdb_search_page(bt, txn, key, NULL, 1, &mpp);
2448 if (rc == MDB_SUCCESS) {
2449 leaf = mdb_search_node(bt, mpp.mp_page, key, &exact, &ki);
2450 if (leaf && exact) {
2451 if (F_ISSET(flags, MDB_NOOVERWRITE)) {
2452 DPRINTF("duplicate key %.*s",
2453 (int)key->mv_size, (char *)key->mv_data);
2456 mdb_del_node(bt, mpp.mp_page, ki);
2458 if (leaf == NULL) { /* append if not found */
2459 ki = NUMKEYS(mpp.mp_page);
2460 DPRINTF("appending key at index %i", ki);
2462 } else if (rc == ENOENT) {
2464 /* new file, just write a root leaf page */
2465 DPRINTF("allocating new root leaf page");
2466 if ((dp = mdbenv_new_page(bt->md_env, P_LEAF, 1)) == NULL) {
2469 mpp.mp_page = &dp->p;
2470 txn->mt_root = mpp.mp_page->mp_pgno;
2471 bt->md_env->me_meta.mm_stat.ms_depth++;
2477 assert(IS_LEAF(mpp.mp_page));
2478 DPRINTF("there are %u keys, should insert new key at index %i",
2479 NUMKEYS(mpp.mp_page), ki);
2481 if (SIZELEFT(mpp.mp_page) < mdb_leaf_size(bt, key, data)) {
2482 rc = mdb_split(bt, &mpp.mp_page, &ki, key, data, P_INVALID);
2484 /* There is room already in this leaf page. */
2485 rc = mdb_add_node(bt, mpp.mp_page, ki, key, data, 0, 0);
2488 if (rc != MDB_SUCCESS)
2489 txn->mt_flags |= MDB_TXN_ERROR;
2491 bt->md_env->me_meta.mm_stat.ms_entries++;
2498 mdbenv_get_flags(MDB_env *env, unsigned int *arg)
2503 *arg = env->me_flags;
2508 mdbenv_get_path(MDB_env *env, const char **arg)
2513 *arg = env->me_path;
2518 mdbenv_stat(MDB_env *env, MDB_stat **arg)
2520 if (env == NULL || arg == NULL)
2523 *arg = &env->me_meta.mm_stat;
2528 int mdb_open(MDB_env *env, MDB_txn *txn, const char *name, unsigned int flags, MDB_db **db)
2531 *db = (MDB_db *)&env->me_db;
2537 int mdb_stat(MDB_db *db, MDB_stat **arg)
2539 if (db == NULL || arg == NULL)
2543 bcopy(&db->md_env->me_meta.mm_stat, &db->md_stat, sizeof(db->md_stat));
2545 *arg = &db->md_stat;
2550 void mdb_close(MDB_db *db)