27 #define ULONG unsigned long
36 #if (DEBUG +0) && defined(__GNUC__)
37 # define DPRINTF(fmt, ...) \
38 fprintf(stderr, "%s:%d: " fmt "\n", __func__, __LINE__, ##__VA_ARGS__)
40 # define DPRINTF(...) ((void) 0)
45 #define MDB_MAGIC 0xBEEFC0DE
47 #define MAXKEYSIZE 255
49 #define P_INVALID (~0UL)
51 #define F_ISSET(w, f) (((w) & (f)) == (f))
53 typedef uint16_t indx_t;
55 #define DEFAULT_READERS 126
56 #define DEFAULT_MAPSIZE 1048576
58 /* Lock descriptor stuff */
63 typedef struct MDB_rxbody {
68 #define CACHELINE 64 /* most CPUs. Itanium uses 128 */
71 typedef struct MDB_reader {
73 /* cache line alignment */
74 char pad[CACHELINE-sizeof(MDB_rxbody)];
79 uint32_t mt_version; \
80 pthread_mutex_t mt_mutex; \
82 uint32_t mt_numreaders
83 typedef struct MDB_txbody {
87 typedef struct MDB_txninfo {
89 char pad[CACHELINE-sizeof(MDB_txbody)];
90 pthread_mutex_t mt_wmutex;
91 char pad2[CACHELINE-sizeof(pthread_mutex_t)];
92 MDB_reader mt_readers[1];
95 /* Common header for all page types. Overflow pages
96 * occupy a number of contiguous pages with no
97 * headers on any page after the first.
99 typedef struct MDB_page { /* represents a page of storage */
100 pgno_t mp_pgno; /* page number */
101 #define P_BRANCH 0x01 /* branch page */
102 #define P_LEAF 0x02 /* leaf page */
103 #define P_OVERFLOW 0x04 /* overflow page */
104 #define P_META 0x08 /* meta page */
105 #define P_DIRTY 0x10 /* dirty page */
107 #define mp_lower mp_pb.pb.pb_lower
108 #define mp_upper mp_pb.pb.pb_upper
109 #define mp_pages mp_pb.pb_pages
112 indx_t pb_lower; /* lower bound of free space */
113 indx_t pb_upper; /* upper bound of free space */
115 uint32_t pb_pages; /* number of overflow pages */
117 indx_t mp_ptrs[1]; /* dynamic size */
120 #define PAGEHDRSZ ((unsigned) offsetof(MDB_page, mp_ptrs))
122 #define NUMKEYS(p) (((p)->mp_lower - PAGEHDRSZ) >> 1)
123 #define SIZELEFT(p) (indx_t)((p)->mp_upper - (p)->mp_lower)
124 #define PAGEFILL(env, p) (1000L * ((env)->me_meta.mm_psize - PAGEHDRSZ - SIZELEFT(p)) / \
125 ((env)->me_meta.mm_psize - PAGEHDRSZ))
126 #define IS_LEAF(p) F_ISSET((p)->mp_flags, P_LEAF)
127 #define IS_BRANCH(p) F_ISSET((p)->mp_flags, P_BRANCH)
128 #define IS_OVERFLOW(p) F_ISSET((p)->mp_flags, P_OVERFLOW)
130 #define OVPAGES(size, psize) (PAGEHDRSZ + size + psize - 1) / psize;
132 typedef struct MDB_meta { /* meta (footer) page content */
135 void *mm_address; /* address for fixed mapping */
136 size_t mm_mapsize; /* size of mmap region */
137 pgno_t mm_last_pg; /* last used page in file */
138 ULONG mm_txnid; /* txnid that committed this page */
142 ULONG mm_branch_pages;
144 ULONG mm_overflow_pages;
149 typedef struct MDB_dhead { /* a dirty page */
150 SIMPLEQ_ENTRY(MDB_dpage) md_next; /* queue of dirty pages */
152 unsigned md_pi; /* parent index */
156 typedef struct MDB_dpage {
161 SIMPLEQ_HEAD(dirty_queue, MDB_dpage);
163 typedef struct MDB_oldpages {
164 struct MDB_oldpages *mo_next;
166 pgno_t mo_pages[1]; /* dynamic */
169 typedef struct MDB_pageparent {
175 static MDB_dpage *mdb_alloc_page(MDB_txn *txn, MDB_page *parent, unsigned int parent_idx, int num);
176 static int mdb_touch(MDB_txn *txn, MDB_pageparent *mp);
178 typedef struct MDB_ppage { /* ordered list of pages */
179 SLIST_ENTRY(MDB_ppage) mp_entry;
181 unsigned int mp_ki; /* cursor index on page */
183 SLIST_HEAD(page_stack, MDB_ppage);
185 #define CURSOR_EMPTY(c) SLIST_EMPTY(&(c)->mc_stack)
186 #define CURSOR_TOP(c) SLIST_FIRST(&(c)->mc_stack)
187 #define CURSOR_POP(c) SLIST_REMOVE_HEAD(&(c)->mc_stack, mp_entry)
188 #define CURSOR_PUSH(c,p) SLIST_INSERT_HEAD(&(c)->mc_stack, p, mp_entry)
192 struct page_stack mc_stack; /* stack of parent pages */
194 short mc_initialized; /* 1 if initialized */
195 short mc_eof; /* 1 if end is reached */
198 #define METAHASHLEN offsetof(MDB_meta, mm_hash)
199 #define METADATA(p) ((void *)((char *)p + PAGEHDRSZ))
201 typedef struct MDB_node {
202 #define mn_pgno mn_p.np_pgno
203 #define mn_dsize mn_p.np_dsize
205 pgno_t np_pgno; /* child page number */
206 uint32_t np_dsize; /* leaf data size */
208 unsigned int mn_flags:4;
209 unsigned int mn_ksize:12; /* key size */
210 #define F_BIGDATA 0x01 /* data put on overflow page */
214 typedef struct MDB_dbx {
216 MDB_cmp_func *md_cmp; /* user compare function */
217 MDB_rel_func *md_rel; /* user relocate function */
218 MDB_db *md_parent; /* parent tree */
221 typedef struct MDB_db {
225 ULONG md_branch_pages;
227 ULONG md_overflow_pages;
233 pgno_t mt_root; /* current / new root page */
234 pgno_t mt_next_pgno; /* next unallocated page */
238 pgno_t *mt_free_pgs; /* this is an IDL */
240 struct dirty_queue *dirty_queue; /* modified pages */
243 MDB_dbx *mt_dbxs; /* array */
244 MDB_db **mt_dbs; /* array of ptrs */
245 unsigned int mt_numdbs;
247 #define MDB_TXN_RDONLY 0x01 /* read-only transaction */
248 #define MDB_TXN_ERROR 0x02 /* an error has occurred */
249 #define MDB_TXN_METOGGLE 0x04 /* used meta page 1 */
250 unsigned int mt_flags;
257 unsigned int me_maxreaders;
260 MDB_txninfo *me_txns;
262 MDB_txn *me_txn; /* current write transaction */
264 off_t me_size; /* current file size */
265 pthread_key_t me_txkey; /* thread-key for readers */
266 MDB_oldpages *me_pghead;
267 MDB_oldpages *me_pgtail;
268 MDB_dbx *me_dbxs; /* array */
269 MDB_db **me_dbs; /* array of ptrs */
270 unsigned int me_numdbs;
273 #define NODESIZE offsetof(MDB_node, mn_data)
275 #define INDXSIZE(k) (NODESIZE + ((k) == NULL ? 0 : (k)->mv_size))
276 #define LEAFSIZE(k, d) (NODESIZE + (k)->mv_size + (d)->mv_size)
277 #define NODEPTR(p, i) ((MDB_node *)((char *)(p) + (p)->mp_ptrs[i]))
278 #define NODEKEY(node) (void *)((node)->mn_data)
279 #define NODEDATA(node) (void *)((char *)(node)->mn_data + (node)->mn_ksize)
280 #define NODEPGNO(node) ((node)->mn_pgno)
281 #define NODEDSZ(node) ((node)->mn_dsize)
283 #define MDB_COMMIT_PAGES 64 /* max number of pages to write in one commit */
284 #define MDB_MAXCACHE_DEF 1024 /* max number of pages to keep in cache */
286 static int mdb_search_page_root(MDB_db *db,
288 MDB_cursor *cursor, int modify,
289 MDB_pageparent *mpp);
290 static int mdb_search_page(MDB_db *db,
291 MDB_txn *txn, MDB_val *key,
292 MDB_cursor *cursor, int modify,
293 MDB_pageparent *mpp);
295 static int mdbenv_read_header(MDB_env *env);
296 static int mdb_check_meta_page(MDB_page *p);
297 static int mdbenv_read_meta(MDB_env *env, int *which);
298 static int mdbenv_write_meta(MDB_txn *txn);
299 static MDB_page *mdbenv_get_page(MDB_env *env, pgno_t pgno);
301 static MDB_node *mdb_search_node(MDB_db *db, MDB_page *mp,
302 MDB_val *key, int *exactp, unsigned int *kip);
303 static int mdb_add_node(MDB_db *bt, MDB_page *mp,
304 indx_t indx, MDB_val *key, MDB_val *data,
305 pgno_t pgno, uint8_t flags);
306 static void mdb_del_node(MDB_db *bt, MDB_page *mp,
308 static int mdb_read_data(MDB_db *bt, MDB_page *mp,
309 MDB_node *leaf, MDB_val *data);
311 static int mdb_rebalance(MDB_db *bt, MDB_pageparent *mp);
312 static int mdb_update_key(MDB_db *bt, MDB_page *mp,
313 indx_t indx, MDB_val *key);
314 static int mdb_move_node(MDB_db *bt,
315 MDB_pageparent *src, indx_t srcindx,
316 MDB_pageparent *dst, indx_t dstindx);
317 static int mdb_merge(MDB_db *bt, MDB_pageparent *src,
318 MDB_pageparent *dst);
319 static int mdb_split(MDB_db *bt, MDB_page **mpp,
320 unsigned int *newindxp, MDB_val *newkey,
321 MDB_val *newdata, pgno_t newpgno);
322 static MDB_dpage *mdb_new_page(MDB_db *db, uint32_t flags, int num);
324 static void cursor_pop_page(MDB_cursor *cursor);
325 static MDB_ppage *cursor_push_page(MDB_cursor *cursor,
328 static int mdb_set_key(MDB_db *bt, MDB_page *mp,
329 MDB_node *node, MDB_val *key);
330 static int mdb_sibling(MDB_cursor *cursor, int move_right);
331 static int mdb_cursor_next(MDB_cursor *cursor,
332 MDB_val *key, MDB_val *data);
333 static int mdb_cursor_set(MDB_cursor *cursor,
334 MDB_val *key, MDB_val *data, int *exactp);
335 static int mdb_cursor_first(MDB_cursor *cursor,
336 MDB_val *key, MDB_val *data);
338 static size_t mdb_leaf_size(MDB_db *bt, MDB_val *key,
340 static size_t mdb_branch_size(MDB_db *bt, MDB_val *key);
342 static int memncmp(const void *s1, size_t n1,
343 const void *s2, size_t n2);
344 static int memnrcmp(const void *s1, size_t n1,
345 const void *s2, size_t n2);
348 memncmp(const void *s1, size_t n1, const void *s2, size_t n2)
350 int diff, len_diff = -1;
353 len_diff = (n1 > n2);
356 diff = memcmp(s1, s2, n1);
357 return diff ? diff : len_diff;
361 memnrcmp(const void *s1, size_t n1, const void *s2, size_t n2)
363 const unsigned char *p1, *p2, *p1_lim;
370 p1 = (const unsigned char *)s1 + n1 - 1;
371 p2 = (const unsigned char *)s2 + n2 - 1;
373 for (p1_lim = (n1 <= n2 ? s1 : s2); *p1 == *p2; p1--, p2--) {
375 return (p1 != s1) ? (p1 != p2) : (p2 != s2) ? -1 : 0;
381 mdb_cmp(MDB_db *db, const MDB_val *a, const MDB_val *b)
383 return db->md_cmp(a, b);
387 _mdb_cmp(MDB_db *db, const MDB_val *key1, const MDB_val *key2)
389 if (F_ISSET(db->md_flags, MDB_REVERSEKEY))
390 return memnrcmp(key1->mv_data, key1->mv_size, key2->mv_data, key2->mv_size);
392 return memncmp((char *)key1->mv_data, key1->mv_size, key2->mv_data, key2->mv_size);
395 /* Allocate new page(s) for writing */
397 mdb_alloc_page(MDB_txn *txn, MDB_page *parent, unsigned int parent_idx, int num)
400 pgno_t pgno = P_INVALID;
402 if (txn->mt_env->me_pghead) {
403 ULONG oldest = txn->mt_txnid - 2;
405 for (i=0; i<txn->mt_env->me_txns->mt_numreaders; i++) {
406 if (txn->mt_env->me_txns->mt_readers[i].mr_txnid < oldest)
407 oldest = txn->mt_env->me_txns->mt_readers[i].mr_txnid;
409 if (oldest > txn->mt_env->me_pghead->mo_txnid) {
410 MDB_oldpages *mop = txn->mt_env->me_pghead;
411 txn->mt_oldest = oldest;
413 /* FIXME: For now, always use fresh pages. We
414 * really ought to search the free list for a
419 /* peel pages off tail, so we only have to truncate the list */
420 pgno = MDB_IDL_LAST(mop->mo_pages);
421 if (MDB_IDL_IS_RANGE(mop->mo_pages)) {
423 if (mop->mo_pages[2] > mop->mo_pages[1])
424 mop->mo_pages[0] = 0;
428 if (MDB_IDL_IS_ZERO(mop->mo_pages)) {
429 txn->mt_env->me_pghead = mop->mo_next;
430 if (!txn->mt_env->me_pghead)
431 txn->mt_env->me_pgtail = NULL;
438 if ((dp = malloc(txn->mt_env->me_meta.mm_psize * num + sizeof(MDB_dhead))) == NULL)
441 dp->h.md_parent = parent;
442 dp->h.md_pi = parent_idx;
443 SIMPLEQ_INSERT_TAIL(txn->mt_u.dirty_queue, dp, h.md_next);
444 if (pgno == P_INVALID) {
445 dp->p.mp_pgno = txn->mt_next_pgno;
446 txn->mt_next_pgno += num;
448 dp->p.mp_pgno = pgno;
454 /* Touch a page: make it dirty and re-insert into tree with updated pgno.
457 mdb_touch(MDB_txn *txn, MDB_pageparent *pp)
459 MDB_page *mp = pp->mp_page;
464 if (!F_ISSET(mp->mp_flags, P_DIRTY)) {
466 if ((dp = mdb_alloc_page(txn, pp->mp_parent, pp->mp_pi, 1)) == NULL)
468 DPRINTF("touched page %lu -> %lu", mp->mp_pgno, dp->p.mp_pgno);
469 mdb_idl_insert(txn->mt_free_pgs, mp->mp_pgno);
470 pgno = dp->p.mp_pgno;
471 bcopy(mp, &dp->p, txn->mt_env->me_meta.mm_psize);
474 mp->mp_flags |= P_DIRTY;
476 /* Update the page number to new touched page. */
477 if (pp->mp_parent != NULL)
478 NODEPGNO(NODEPTR(pp->mp_parent, pp->mp_pi)) = mp->mp_pgno;
485 mdbenv_sync(MDB_env *env)
488 if (!F_ISSET(env->me_flags, MDB_NOSYNC)) {
489 if (fsync(env->me_fd))
496 mdb_txn_begin(MDB_env *env, int rdonly, MDB_txn **ret)
501 if ((txn = calloc(1, sizeof(*txn))) == NULL) {
502 DPRINTF("calloc: %s", strerror(errno));
507 txn->mt_flags |= MDB_TXN_RDONLY;
509 txn->mt_u.dirty_queue = calloc(1, sizeof(*txn->mt_u.dirty_queue));
510 if (txn->mt_u.dirty_queue == NULL) {
514 SIMPLEQ_INIT(txn->mt_u.dirty_queue);
516 pthread_mutex_lock(&env->me_txns->mt_wmutex);
517 env->me_txns->mt_txnid++;
518 txn->mt_free_pgs = malloc(MDB_IDL_UM_SIZEOF);
519 if (txn->mt_free_pgs == NULL) {
520 free(txn->mt_u.dirty_queue);
524 txn->mt_free_pgs[0] = 0;
526 txn->mt_txnid = env->me_txns->mt_txnid;
528 MDB_reader *r = pthread_getspecific(env->me_txkey);
531 pthread_mutex_lock(&env->me_txns->mt_mutex);
532 for (i=0; i<env->me_maxreaders; i++) {
533 if (env->me_txns->mt_readers[i].mr_pid == 0) {
534 env->me_txns->mt_readers[i].mr_pid = getpid();
535 env->me_txns->mt_readers[i].mr_tid = pthread_self();
536 pthread_setspecific(env->me_txkey, &env->me_txns->mt_readers[i]);
537 if (i >= env->me_txns->mt_numreaders)
538 env->me_txns->mt_numreaders = i+1;
542 pthread_mutex_unlock(&env->me_txns->mt_mutex);
543 if (i == env->me_maxreaders) {
547 r->mr_txnid = txn->mt_txnid;
548 txn->mt_u.reader = r;
555 if ((rc = mdbenv_read_meta(env, &toggle)) != MDB_SUCCESS) {
561 txn->mt_flags |= MDB_TXN_METOGGLE;
563 txn->mt_next_pgno = env->me_meta.mm_last_pg+1;
564 txn->mt_root = env->me_meta.mm_root;
565 DPRINTF("begin transaction %lu on mdbenv %p, root page %lu",
566 txn->mt_txnid, (void *) env, txn->mt_root);
573 mdb_txn_abort(MDB_txn *txn)
582 DPRINTF("abort transaction %lu on mdbenv %p, root page %lu",
583 txn->mt_txnid, (void *) env, txn->mt_root);
585 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
586 txn->mt_u.reader->mr_txnid = 0;
588 /* Discard all dirty pages. Return any re-used pages
591 MDB_IDL_ZERO(txn->mt_free_pgs);
592 while (!SIMPLEQ_EMPTY(txn->mt_u.dirty_queue)) {
593 dp = SIMPLEQ_FIRST(txn->mt_u.dirty_queue);
594 SIMPLEQ_REMOVE_HEAD(txn->mt_u.dirty_queue, h.md_next);
595 if (dp->p.mp_pgno <= env->me_meta.mm_last_pg)
596 mdb_idl_insert(txn->mt_free_pgs, dp->p.mp_pgno);
599 /* put back to head of free list */
600 if (!MDB_IDL_IS_ZERO(txn->mt_free_pgs)) {
603 mop = malloc(sizeof(MDB_oldpages) + MDB_IDL_SIZEOF(txn->mt_free_pgs) - sizeof(pgno_t));
604 mop->mo_next = env->me_pghead;
605 mop->mo_txnid = txn->mt_oldest - 1;
606 if (!env->me_pghead) {
607 env->me_pgtail = mop;
609 env->me_pghead = mop;
610 bcopy(txn->mt_free_pgs, mop->mo_pages, MDB_IDL_SIZEOF(txn->mt_free_pgs));
613 free(txn->mt_free_pgs);
614 free(txn->mt_u.dirty_queue);
616 env->me_txns->mt_txnid--;
617 pthread_mutex_unlock(&env->me_txns->mt_wmutex);
624 mdb_txn_commit(MDB_txn *txn)
632 struct iovec iov[MDB_COMMIT_PAGES];
635 assert(txn->mt_env != NULL);
639 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
640 DPRINTF("attempt to commit read-only transaction");
645 if (txn != env->me_txn) {
646 DPRINTF("attempt to commit unknown transaction");
651 if (F_ISSET(txn->mt_flags, MDB_TXN_ERROR)) {
652 DPRINTF("error flag is set, can't commit");
657 if (SIMPLEQ_EMPTY(txn->mt_u.dirty_queue))
660 DPRINTF("committing transaction %lu on mdbenv %p, root page %lu",
661 txn->mt_txnid, (void *) env, txn->mt_root);
663 /* Commit up to MDB_COMMIT_PAGES dirty pages to disk until done.
670 SIMPLEQ_FOREACH(dp, txn->mt_u.dirty_queue, h.md_next) {
671 if (dp->p.mp_pgno != next) {
673 DPRINTF("committing %u dirty pages", n);
674 rc = writev(env->me_fd, iov, n);
678 DPRINTF("short write, filesystem full?");
680 DPRINTF("writev: %s", strerror(errno));
687 lseek(env->me_fd, dp->p.mp_pgno * env->me_meta.mm_psize, SEEK_SET);
688 next = dp->p.mp_pgno;
690 DPRINTF("committing page %lu", dp->p.mp_pgno);
691 iov[n].iov_len = env->me_meta.mm_psize * dp->h.md_num;
692 iov[n].iov_base = &dp->p;
693 size += iov[n].iov_len;
694 next = dp->p.mp_pgno + dp->h.md_num;
695 /* clear dirty flag */
696 dp->p.mp_flags &= ~P_DIRTY;
697 if (++n >= MDB_COMMIT_PAGES) {
706 DPRINTF("committing %u dirty pages", n);
707 rc = writev(env->me_fd, iov, n);
711 DPRINTF("short write, filesystem full?");
713 DPRINTF("writev: %s", strerror(errno));
720 /* Drop the dirty pages.
722 while (!SIMPLEQ_EMPTY(txn->mt_u.dirty_queue)) {
723 dp = SIMPLEQ_FIRST(txn->mt_u.dirty_queue);
724 SIMPLEQ_REMOVE_HEAD(txn->mt_u.dirty_queue, h.md_next);
728 if ((n = mdbenv_sync(env)) != 0 ||
729 (n = mdbenv_write_meta(txn)) != MDB_SUCCESS ||
730 (n = mdbenv_sync(env)) != 0) {
736 /* add to tail of free list */
737 if (!MDB_IDL_IS_ZERO(txn->mt_free_pgs)) {
740 mop = malloc(sizeof(MDB_oldpages) + MDB_IDL_SIZEOF(txn->mt_free_pgs) - sizeof(pgno_t));
742 if (env->me_pghead) {
743 env->me_pgtail->mo_next = mop;
745 env->me_pghead = mop;
747 env->me_pgtail = mop;
748 bcopy(txn->mt_free_pgs, mop->mo_pages, MDB_IDL_SIZEOF(txn->mt_free_pgs));
749 mop->mo_txnid = txn->mt_txnid;
752 pthread_mutex_unlock(&env->me_txns->mt_wmutex);
753 free(txn->mt_free_pgs);
754 free(txn->mt_u.dirty_queue);
765 mdbenv_read_header(MDB_env *env)
774 /* We don't know the page size yet, so use a minimum value.
777 if ((rc = pread(env->me_fd, page, PAGESIZE, 0)) == 0) {
779 } else if (rc != PAGESIZE) {
782 DPRINTF("read: %s", strerror(errno));
786 p = (MDB_page *)page;
788 if (!F_ISSET(p->mp_flags, P_META)) {
789 DPRINTF("page %lu not a meta page", p->mp_pgno);
794 if (m->mm_magic != MDB_MAGIC) {
795 DPRINTF("meta has invalid magic");
799 if (m->mm_version != MDB_VERSION) {
800 DPRINTF("database is version %u, expected version %u",
801 m->mm_version, MDB_VERSION);
805 bcopy(m, &env->me_meta, sizeof(*m));
810 mdbenv_init_meta(MDB_env *env)
817 DPRINTF("writing new meta page");
818 psize = sysconf(_SC_PAGE_SIZE);
820 env->me_meta.mm_magic = MDB_MAGIC;
821 env->me_meta.mm_version = MDB_VERSION;
822 env->me_meta.mm_psize = psize;
823 env->me_meta.mm_flags = env->me_flags & 0xffff;
824 env->me_meta.mm_root = P_INVALID;
825 env->me_meta.mm_last_pg = 1;
827 p = calloc(2, psize);
829 p->mp_flags = P_META;
832 bcopy(&env->me_meta, meta, sizeof(*meta));
834 q = (MDB_page *)((char *)p + psize);
837 q->mp_flags = P_META;
840 bcopy(&env->me_meta, meta, sizeof(*meta));
842 rc = write(env->me_fd, p, psize * 2);
844 return (rc == (int)psize * 2) ? MDB_SUCCESS : errno;
848 mdbenv_write_meta(MDB_txn *txn)
857 assert(txn->mt_env != NULL);
859 DPRINTF("writing meta page for root page %lu", txn->mt_root);
864 off = offsetof(MDB_meta, mm_depth);
865 len = sizeof(MDB_meta) - off;
868 meta.mm_depth = env->me_db.md_depth;
869 meta.mm_branch_pages = env->me_db.md_branch_pages;
870 meta.mm_leaf_pages = env->me_db.md_leaf_pages;
871 meta.mm_overflow_pages = env->me_db.md_overflow_pages;
872 meta.mm_entries = env->me_db.md_entries;
873 meta.mm_root = txn->mt_root;
874 meta.mm_last_pg = txn->mt_next_pgno - 1;
875 meta.mm_txnid = txn->mt_txnid;
877 if (!F_ISSET(txn->mt_flags, MDB_TXN_METOGGLE))
878 off += env->me_meta.mm_psize;
881 lseek(env->me_fd, off, SEEK_SET);
882 rc = write(env->me_fd, ptr, len);
884 DPRINTF("write failed, disk error?");
891 /* Returns true if page p is a valid meta page, false otherwise.
894 mdb_check_meta_page(MDB_page *p)
896 if (!F_ISSET(p->mp_flags, P_META)) {
897 DPRINTF("page %lu not a meta page", p->mp_pgno);
905 mdbenv_read_meta(MDB_env *env, int *which)
913 if ((mp0 = mdbenv_get_page(env, 0)) == NULL ||
914 (mp1 = mdbenv_get_page(env, 1)) == NULL)
917 rc = mdb_check_meta_page(mp0);
920 rc = mdb_check_meta_page(mp1);
923 meta[0] = METADATA(mp0);
924 meta[1] = METADATA(mp1);
926 if (meta[0]->mm_txnid < meta[1]->mm_txnid)
929 if (meta[toggle]->mm_txnid > env->me_meta.mm_txnid) {
930 bcopy(meta[toggle], &env->me_meta, sizeof(env->me_meta));
935 DPRINTF("Using meta page %d", toggle);
941 mdbenv_create(MDB_env **env)
945 e = calloc(1, sizeof(*e));
946 if (!e) return ENOMEM;
948 e->me_meta.mm_mapsize = DEFAULT_MAPSIZE;
949 e->me_maxreaders = DEFAULT_READERS;
958 mdbenv_set_mapsize(MDB_env *env, size_t size)
962 env->me_mapsize = env->me_meta.mm_mapsize = size;
967 mdbenv_set_maxreaders(MDB_env *env, int readers)
969 env->me_maxreaders = readers;
974 mdbenv_get_maxreaders(MDB_env *env, int *readers)
976 if (!env || !readers)
978 *readers = env->me_maxreaders;
983 mdbenv_open2(MDB_env *env, unsigned int flags)
987 env->me_flags = flags;
989 if ((i = mdbenv_read_header(env)) != 0) {
992 DPRINTF("new mdbenv");
996 if (!env->me_mapsize)
997 env->me_mapsize = env->me_meta.mm_mapsize;
1000 if (env->me_meta.mm_address && (flags & MDB_FIXEDMAP))
1002 env->me_map = mmap(env->me_meta.mm_address, env->me_mapsize, PROT_READ, i,
1004 if (env->me_map == MAP_FAILED)
1008 env->me_meta.mm_mapsize = env->me_mapsize;
1009 if (flags & MDB_FIXEDMAP)
1010 env->me_meta.mm_address = env->me_map;
1011 i = mdbenv_init_meta(env);
1012 if (i != MDB_SUCCESS) {
1013 munmap(env->me_map, env->me_mapsize);
1018 if ((i = mdbenv_read_meta(env, NULL)) != 0)
1021 DPRINTF("opened database version %u, pagesize %u",
1022 env->me_meta.mm_version, env->me_meta.mm_psize);
1023 DPRINTF("depth: %u", env->me_meta.mm_depth);
1024 DPRINTF("entries: %lu", env->me_meta.mm_entries);
1025 DPRINTF("branch pages: %lu", env->me_meta.mm_branch_pages);
1026 DPRINTF("leaf pages: %lu", env->me_meta.mm_leaf_pages);
1027 DPRINTF("overflow pages: %lu", env->me_meta.mm_overflow_pages);
1028 DPRINTF("root: %lu", env->me_meta.mm_root);
1034 mdbenv_reader_dest(void *ptr)
1036 MDB_reader *reader = ptr;
1038 reader->mr_txnid = 0;
1044 mdbenv_share_locks(MDB_env *env)
1046 struct flock lock_info;
1048 env->me_txns->mt_txnid = env->me_meta.mm_txnid;
1050 memset((void *)&lock_info, 0, sizeof(lock_info));
1051 lock_info.l_type = F_RDLCK;
1052 lock_info.l_whence = SEEK_SET;
1053 lock_info.l_start = 0;
1054 lock_info.l_len = 1;
1055 fcntl(env->me_lfd, F_SETLK, &lock_info);
1059 mdbenv_setup_locks(MDB_env *env, char *lpath, int mode, int *excl)
1063 struct flock lock_info;
1067 if ((env->me_lfd = open(lpath, O_RDWR|O_CREAT, mode)) == -1) {
1071 /* Try to get exclusive lock. If we succeed, then
1072 * nobody is using the lock region and we should initialize it.
1074 memset((void *)&lock_info, 0, sizeof(lock_info));
1075 lock_info.l_type = F_WRLCK;
1076 lock_info.l_whence = SEEK_SET;
1077 lock_info.l_start = 0;
1078 lock_info.l_len = 1;
1079 rc = fcntl(env->me_lfd, F_SETLK, &lock_info);
1083 lock_info.l_type = F_RDLCK;
1084 rc = fcntl(env->me_lfd, F_SETLK, &lock_info);
1090 size = lseek(env->me_lfd, 0, SEEK_END);
1091 rsize = (env->me_maxreaders-1) * sizeof(MDB_reader) + sizeof(MDB_txninfo);
1092 if (size < rsize && *excl) {
1093 if (ftruncate(env->me_lfd, rsize) != 0) {
1099 size = rsize - sizeof(MDB_txninfo);
1100 env->me_maxreaders = size/sizeof(MDB_reader) + 1;
1102 env->me_txns = mmap(0, rsize, PROT_READ|PROT_WRITE, MAP_SHARED,
1104 if (env->me_txns == MAP_FAILED) {
1109 pthread_mutexattr_t mattr;
1111 pthread_mutexattr_init(&mattr);
1112 pthread_mutexattr_setpshared(&mattr, PTHREAD_PROCESS_SHARED);
1113 pthread_mutex_init(&env->me_txns->mt_mutex, &mattr);
1114 pthread_mutex_init(&env->me_txns->mt_wmutex, &mattr);
1115 env->me_txns->mt_version = MDB_VERSION;
1116 env->me_txns->mt_magic = MDB_MAGIC;
1117 env->me_txns->mt_txnid = 0;
1118 env->me_txns->mt_numreaders = 0;
1121 if (env->me_txns->mt_magic != MDB_MAGIC) {
1122 DPRINTF("lock region has invalid magic");
1125 if (env->me_txns->mt_version != MDB_VERSION) {
1126 DPRINTF("lock region is version %u, expected version %u",
1127 env->me_txns->mt_version, MDB_VERSION);
1130 if (errno != EACCES && errno != EAGAIN) {
1144 mdbenv_open(MDB_env *env, const char *path, unsigned int flags, mode_t mode)
1146 int oflags, rc, len, excl;
1147 char *lpath, *dpath;
1150 lpath = malloc(len + sizeof("/lock.mdb") + len + sizeof("/data.db"));
1153 dpath = lpath + len + sizeof("/lock.mdb");
1154 sprintf(lpath, "%s/lock.mdb", path);
1155 sprintf(dpath, "%s/data.mdb", path);
1157 rc = mdbenv_setup_locks(env, lpath, mode, &excl);
1161 if (F_ISSET(flags, MDB_RDONLY))
1164 oflags = O_RDWR | O_CREAT;
1166 if ((env->me_fd = open(dpath, oflags, mode)) == -1)
1169 if ((rc = mdbenv_open2(env, flags)) != MDB_SUCCESS) {
1173 env->me_path = strdup(path);
1174 DPRINTF("opened dbenv %p", (void *) env);
1175 pthread_key_create(&env->me_txkey, mdbenv_reader_dest);
1177 mdbenv_share_locks(env);
1187 mdbenv_close(MDB_env *env)
1195 munmap(env->me_map, env->me_mapsize);
1199 size_t size = (env->me_maxreaders-1) * sizeof(MDB_reader) + sizeof(MDB_txninfo);
1200 munmap(env->me_txns, size);
1206 /* Search for key within a leaf page, using binary search.
1207 * Returns the smallest entry larger or equal to the key.
1208 * If exactp is non-null, stores whether the found entry was an exact match
1209 * in *exactp (1 or 0).
1210 * If kip is non-null, stores the index of the found entry in *kip.
1211 * If no entry larger or equal to the key is found, returns NULL.
1214 mdb_search_node(MDB_db *bt, MDB_page *mp, MDB_val *key,
1215 int *exactp, unsigned int *kip)
1223 DPRINTF("searching %u keys in %s page %lu",
1225 IS_LEAF(mp) ? "leaf" : "branch",
1228 assert(NUMKEYS(mp) > 0);
1230 bzero(&nodekey, sizeof(nodekey));
1232 low = IS_LEAF(mp) ? 0 : 1;
1233 high = NUMKEYS(mp) - 1;
1234 while (low <= high) {
1235 i = (low + high) >> 1;
1236 node = NODEPTR(mp, i);
1238 nodekey.mv_size = node->mn_ksize;
1239 nodekey.mv_data = NODEKEY(node);
1242 rc = bt->md_cmp(key, &nodekey);
1244 rc = _mdb_cmp(bt, key, &nodekey);
1247 DPRINTF("found leaf index %u [%.*s], rc = %i",
1248 i, (int)nodekey.mv_size, (char *)nodekey.mv_data, rc);
1250 DPRINTF("found branch index %u [%.*s -> %lu], rc = %i",
1251 i, (int)node->mn_ksize, (char *)NODEKEY(node),
1262 if (rc > 0) { /* Found entry is less than the key. */
1263 i++; /* Skip to get the smallest entry larger than key. */
1264 if (i >= NUMKEYS(mp))
1265 /* There is no entry larger or equal to the key. */
1269 *exactp = (rc == 0);
1270 if (kip) /* Store the key index if requested. */
1273 return NODEPTR(mp, i);
1277 cursor_pop_page(MDB_cursor *cursor)
1281 top = CURSOR_TOP(cursor);
1284 DPRINTF("popped page %lu off cursor %p", top->mp_page->mp_pgno, (void *) cursor);
1290 cursor_push_page(MDB_cursor *cursor, MDB_page *mp)
1294 DPRINTF("pushing page %lu on cursor %p", mp->mp_pgno, (void *) cursor);
1296 if ((ppage = calloc(1, sizeof(*ppage))) == NULL)
1298 ppage->mp_page = mp;
1299 CURSOR_PUSH(cursor, ppage);
1304 mdbenv_get_page(MDB_env *env, pgno_t pgno)
1307 MDB_txn *txn = env->me_txn;
1310 if (txn && !SIMPLEQ_EMPTY(txn->mt_u.dirty_queue)) {
1312 SIMPLEQ_FOREACH(dp, txn->mt_u.dirty_queue, h.md_next) {
1313 if (dp->p.mp_pgno == pgno) {
1321 p = (MDB_page *)(env->me_map + env->me_meta.mm_psize * pgno);
1327 mdb_search_page_root(MDB_db *bt, MDB_val *key,
1328 MDB_cursor *cursor, int modify, MDB_pageparent *mpp)
1330 MDB_page *mp = mpp->mp_page;
1333 if (cursor && cursor_push_page(cursor, mp) == NULL)
1336 while (IS_BRANCH(mp)) {
1340 DPRINTF("branch page %lu has %u keys", mp->mp_pgno, NUMKEYS(mp));
1341 assert(NUMKEYS(mp) > 1);
1342 DPRINTF("found index 0 to page %lu", NODEPGNO(NODEPTR(mp, 0)));
1344 if (key == NULL) /* Initialize cursor to first page. */
1348 node = mdb_search_node(bt, mp, key, &exact, &i);
1350 i = NUMKEYS(mp) - 1;
1358 DPRINTF("following index %u for key %.*s",
1359 i, (int)key->mv_size, (char *)key->mv_data);
1360 assert(i < NUMKEYS(mp));
1361 node = NODEPTR(mp, i);
1364 CURSOR_TOP(cursor)->mp_ki = i;
1366 mpp->mp_parent = mp;
1367 if ((mp = mdbenv_get_page(bt->md_env, NODEPGNO(node))) == NULL)
1372 if (cursor && cursor_push_page(cursor, mp) == NULL)
1376 MDB_dhead *dh = ((MDB_dhead *)mp)-1;
1377 if ((rc = mdb_touch(bt->md_env->me_txn, mpp)) != 0)
1379 dh = ((MDB_dhead *)mpp->mp_page)-1;
1380 dh->md_parent = mpp->mp_parent;
1381 dh->md_pi = mpp->mp_pi;
1388 DPRINTF("internal error, index points to a %02X page!?",
1393 DPRINTF("found leaf page %lu for key %.*s", mp->mp_pgno,
1394 key ? (int)key->mv_size : 0, key ? (char *)key->mv_data : NULL);
1399 /* Search for the page a given key should be in.
1400 * Stores a pointer to the found page in *mpp.
1401 * If key is NULL, search for the lowest page (used by mdb_cursor_first).
1402 * If cursor is non-null, pushes parent pages on the cursor stack.
1403 * If modify is true, visited pages are updated with new page numbers.
1406 mdb_search_page(MDB_db *db, MDB_txn *txn, MDB_val *key,
1407 MDB_cursor *cursor, int modify, MDB_pageparent *mpp)
1412 /* Can't modify pages outside a transaction. */
1413 if (txn == NULL && modify) {
1417 /* Choose which root page to start with. If a transaction is given
1418 * use the root page from the transaction, otherwise read the last
1419 * committed root page.
1422 if ((rc = mdbenv_read_meta(db->md_env, NULL)) != MDB_SUCCESS)
1424 root = db->md_env->me_meta.mm_root;
1425 } else if (F_ISSET(txn->mt_flags, MDB_TXN_ERROR)) {
1426 DPRINTF("transaction has failed, must abort");
1429 root = txn->mt_root;
1431 if (root == P_INVALID) { /* Tree is empty. */
1432 DPRINTF("tree is empty");
1436 if ((mpp->mp_page = mdbenv_get_page(db->md_env, root)) == NULL)
1439 DPRINTF("root page has flags 0x%X", mpp->mp_page->mp_flags);
1441 if (modify && !F_ISSET(mpp->mp_page->mp_flags, P_DIRTY)) {
1442 mpp->mp_parent = NULL;
1444 if ((rc = mdb_touch(txn, mpp)))
1446 txn->mt_root = mpp->mp_page->mp_pgno;
1449 return mdb_search_page_root(db, key, cursor, modify, mpp);
1453 mdb_read_data(MDB_db *db, MDB_page *mp, MDB_node *leaf,
1456 MDB_page *omp; /* overflow mpage */
1460 bzero(data, sizeof(*data));
1461 max = db->md_env->me_meta.mm_psize - PAGEHDRSZ;
1463 if (!F_ISSET(leaf->mn_flags, F_BIGDATA)) {
1464 data->mv_size = leaf->mn_dsize;
1465 data->mv_data = NODEDATA(leaf);
1469 /* Read overflow data.
1471 data->mv_size = leaf->mn_dsize;
1472 bcopy(NODEDATA(leaf), &pgno, sizeof(pgno));
1473 if ((omp = mdbenv_get_page(db->md_env, pgno)) == NULL) {
1474 DPRINTF("read overflow page %lu failed", pgno);
1477 data->mv_data = omp;
1483 mdb_get(MDB_db *db, MDB_txn *txn,
1484 MDB_val *key, MDB_val *data)
1492 DPRINTF("===> get key [%.*s]", (int)key->mv_size, (char *)key->mv_data);
1497 if (txn != NULL && db->md_env != txn->mt_env) {
1501 if (key->mv_size == 0 || key->mv_size > MAXKEYSIZE) {
1505 if ((rc = mdb_search_page(db, txn, key, NULL, 0, &mpp)) != MDB_SUCCESS)
1508 leaf = mdb_search_node(db, mpp.mp_page, key, &exact, NULL);
1510 rc = mdb_read_data(db, mpp.mp_page, leaf, data);
1519 mdb_sibling(MDB_cursor *cursor, int move_right)
1523 MDB_ppage *parent, *top;
1526 top = CURSOR_TOP(cursor);
1527 if ((parent = SLIST_NEXT(top, mp_entry)) == NULL) {
1528 return ENOENT; /* root has no siblings */
1531 DPRINTF("parent page is page %lu, index %u",
1532 parent->mp_page->mp_pgno, parent->mp_ki);
1534 cursor_pop_page(cursor);
1535 if (move_right ? (parent->mp_ki + 1 >= NUMKEYS(parent->mp_page))
1536 : (parent->mp_ki == 0)) {
1537 DPRINTF("no more keys left, moving to %s sibling",
1538 move_right ? "right" : "left");
1539 if ((rc = mdb_sibling(cursor, move_right)) != MDB_SUCCESS)
1541 parent = CURSOR_TOP(cursor);
1547 DPRINTF("just moving to %s index key %u",
1548 move_right ? "right" : "left", parent->mp_ki);
1550 assert(IS_BRANCH(parent->mp_page));
1552 indx = NODEPTR(parent->mp_page, parent->mp_ki);
1553 if ((mp = mdbenv_get_page(cursor->mc_db->md_env, indx->mn_pgno)) == NULL)
1556 mp->parent = parent->mp_page;
1557 mp->parent_index = parent->mp_ki;
1560 cursor_push_page(cursor, mp);
1566 mdb_set_key(MDB_db *bt, MDB_page *mp, MDB_node *node,
1572 key->mv_size = node->mn_ksize;
1573 key->mv_data = NODEKEY(node);
1579 mdb_cursor_next(MDB_cursor *cursor, MDB_val *key, MDB_val *data)
1585 if (cursor->mc_eof) {
1589 assert(cursor->mc_initialized);
1591 top = CURSOR_TOP(cursor);
1594 DPRINTF("cursor_next: top page is %lu in cursor %p", mp->mp_pgno, (void *) cursor);
1596 if (top->mp_ki + 1 >= NUMKEYS(mp)) {
1597 DPRINTF("=====> move to next sibling page");
1598 if (mdb_sibling(cursor, 1) != MDB_SUCCESS) {
1602 top = CURSOR_TOP(cursor);
1604 DPRINTF("next page is %lu, key index %u", mp->mp_pgno, top->mp_ki);
1608 DPRINTF("==> cursor points to page %lu with %u keys, key index %u",
1609 mp->mp_pgno, NUMKEYS(mp), top->mp_ki);
1611 assert(IS_LEAF(mp));
1612 leaf = NODEPTR(mp, top->mp_ki);
1614 if (data && mdb_read_data(cursor->mc_db, mp, leaf, data) != MDB_SUCCESS)
1617 return mdb_set_key(cursor->mc_db, mp, leaf, key);
1621 mdb_cursor_set(MDB_cursor *cursor, MDB_val *key, MDB_val *data,
1631 assert(key->mv_size > 0);
1633 rc = mdb_search_page(cursor->mc_db, cursor->mc_txn, key, cursor, 0, &mpp);
1634 if (rc != MDB_SUCCESS)
1636 assert(IS_LEAF(mpp.mp_page));
1638 top = CURSOR_TOP(cursor);
1639 leaf = mdb_search_node(cursor->mc_db, mpp.mp_page, key, exactp, &top->mp_ki);
1640 if (exactp != NULL && !*exactp) {
1641 /* MDB_CURSOR_EXACT specified and not an exact match. */
1646 DPRINTF("===> inexact leaf not found, goto sibling");
1647 if ((rc = mdb_sibling(cursor, 1)) != MDB_SUCCESS)
1648 return rc; /* no entries matched */
1649 top = CURSOR_TOP(cursor);
1651 mpp.mp_page = top->mp_page;
1652 assert(IS_LEAF(mpp.mp_page));
1653 leaf = NODEPTR(mpp.mp_page, 0);
1656 cursor->mc_initialized = 1;
1659 if (data && (rc = mdb_read_data(cursor->mc_db, mpp.mp_page, leaf, data)) != MDB_SUCCESS)
1662 rc = mdb_set_key(cursor->mc_db, mpp.mp_page, leaf, key);
1663 if (rc == MDB_SUCCESS) {
1664 DPRINTF("==> cursor placed on key %.*s",
1665 (int)key->mv_size, (char *)key->mv_data);
1673 mdb_cursor_first(MDB_cursor *cursor, MDB_val *key, MDB_val *data)
1679 rc = mdb_search_page(cursor->mc_db, cursor->mc_txn, NULL, cursor, 0, &mpp);
1680 if (rc != MDB_SUCCESS)
1682 assert(IS_LEAF(mpp.mp_page));
1684 leaf = NODEPTR(mpp.mp_page, 0);
1685 cursor->mc_initialized = 1;
1688 if (data && (rc = mdb_read_data(cursor->mc_db, mpp.mp_page, leaf, data)) != MDB_SUCCESS)
1691 return mdb_set_key(cursor->mc_db, mpp.mp_page, leaf, key);
1695 mdb_cursor_get(MDB_cursor *cursor, MDB_val *key, MDB_val *data,
1705 case MDB_CURSOR_EXACT:
1706 while (CURSOR_TOP(cursor) != NULL)
1707 cursor_pop_page(cursor);
1708 if (key == NULL || key->mv_size == 0 || key->mv_size > MAXKEYSIZE) {
1710 } else if (op == MDB_CURSOR_EXACT)
1711 rc = mdb_cursor_set(cursor, key, data, &exact);
1713 rc = mdb_cursor_set(cursor, key, data, NULL);
1716 if (!cursor->mc_initialized)
1717 rc = mdb_cursor_first(cursor, key, data);
1719 rc = mdb_cursor_next(cursor, key, data);
1722 while (CURSOR_TOP(cursor) != NULL)
1723 cursor_pop_page(cursor);
1724 rc = mdb_cursor_first(cursor, key, data);
1727 DPRINTF("unhandled/unimplemented cursor operation %u", op);
1735 /* Allocate a page and initialize it
1738 mdb_new_page(MDB_db *db, uint32_t flags, int num)
1743 assert(db->md_env != NULL);
1744 assert(db->md_env->me_txn != NULL);
1746 if ((dp = mdb_alloc_page(db->md_env->me_txn, NULL, 0, num)) == NULL)
1748 DPRINTF("allocated new mpage %lu, page size %u",
1749 dp->p.mp_pgno, db->md_env->me_meta.mm_psize);
1750 dp->p.mp_flags = flags | P_DIRTY;
1751 dp->p.mp_lower = PAGEHDRSZ;
1752 dp->p.mp_upper = db->md_env->me_meta.mm_psize;
1754 if (IS_BRANCH(&dp->p))
1755 db->md_branch_pages++;
1756 else if (IS_LEAF(&dp->p))
1757 db->md_leaf_pages++;
1758 else if (IS_OVERFLOW(&dp->p)) {
1759 db->md_overflow_pages += num;
1760 dp->p.mp_pages = num;
1767 mdb_leaf_size(MDB_db *db, MDB_val *key, MDB_val *data)
1771 sz = LEAFSIZE(key, data);
1772 if (data->mv_size >= db->md_env->me_meta.mm_psize / MDB_MINKEYS) {
1773 /* put on overflow page */
1774 sz -= data->mv_size - sizeof(pgno_t);
1777 return sz + sizeof(indx_t);
1781 mdb_branch_size(MDB_db *db, MDB_val *key)
1786 if (sz >= db->md_env->me_meta.mm_psize / MDB_MINKEYS) {
1787 /* put on overflow page */
1788 /* not implemented */
1789 /* sz -= key->size - sizeof(pgno_t); */
1792 return sz + sizeof(indx_t);
1796 mdb_add_node(MDB_db *db, MDB_page *mp, indx_t indx,
1797 MDB_val *key, MDB_val *data, pgno_t pgno, uint8_t flags)
1800 size_t node_size = NODESIZE;
1803 MDB_dpage *ofp = NULL; /* overflow page */
1805 assert(mp->mp_upper >= mp->mp_lower);
1807 DPRINTF("add node [%.*s] to %s page %lu at index %i, key size %zu",
1808 key ? (int)key->mv_size : 0, key ? (char *)key->mv_data : NULL,
1809 IS_LEAF(mp) ? "leaf" : "branch",
1810 mp->mp_pgno, indx, key ? key->mv_size : 0);
1813 node_size += key->mv_size;
1817 if (F_ISSET(flags, F_BIGDATA)) {
1818 /* Data already on overflow page. */
1819 node_size += sizeof(pgno_t);
1820 } else if (data->mv_size >= db->md_env->me_meta.mm_psize / MDB_MINKEYS) {
1821 int ovpages = OVPAGES(data->mv_size, db->md_env->me_meta.mm_psize);
1822 /* Put data on overflow page. */
1823 DPRINTF("data size is %zu, put on overflow page",
1825 node_size += sizeof(pgno_t);
1826 if ((ofp = mdb_new_page(db, P_OVERFLOW, ovpages)) == NULL)
1828 DPRINTF("allocated overflow page %lu", ofp->p.mp_pgno);
1831 node_size += data->mv_size;
1835 if (node_size + sizeof(indx_t) > SIZELEFT(mp)) {
1836 DPRINTF("not enough room in page %lu, got %u ptrs",
1837 mp->mp_pgno, NUMKEYS(mp));
1838 DPRINTF("upper - lower = %u - %u = %u", mp->mp_upper, mp->mp_lower,
1839 mp->mp_upper - mp->mp_lower);
1840 DPRINTF("node size = %zu", node_size);
1844 /* Move higher pointers up one slot. */
1845 for (i = NUMKEYS(mp); i > indx; i--)
1846 mp->mp_ptrs[i] = mp->mp_ptrs[i - 1];
1848 /* Adjust free space offsets. */
1849 ofs = mp->mp_upper - node_size;
1850 assert(ofs >= mp->mp_lower + sizeof(indx_t));
1851 mp->mp_ptrs[indx] = ofs;
1853 mp->mp_lower += sizeof(indx_t);
1855 /* Write the node data. */
1856 node = NODEPTR(mp, indx);
1857 node->mn_ksize = (key == NULL) ? 0 : key->mv_size;
1858 node->mn_flags = flags;
1860 node->mn_dsize = data->mv_size;
1862 node->mn_pgno = pgno;
1865 bcopy(key->mv_data, NODEKEY(node), key->mv_size);
1870 if (F_ISSET(flags, F_BIGDATA))
1871 bcopy(data->mv_data, node->mn_data + key->mv_size,
1874 bcopy(data->mv_data, node->mn_data + key->mv_size,
1877 bcopy(&ofp->p.mp_pgno, node->mn_data + key->mv_size,
1879 bcopy(data->mv_data, METADATA(&ofp->p), data->mv_size);
1887 mdb_del_node(MDB_db *db, MDB_page *mp, indx_t indx)
1890 indx_t i, j, numkeys, ptr;
1894 DPRINTF("delete node %u on %s page %lu", indx,
1895 IS_LEAF(mp) ? "leaf" : "branch", mp->mp_pgno);
1896 assert(indx < NUMKEYS(mp));
1898 node = NODEPTR(mp, indx);
1899 sz = NODESIZE + node->mn_ksize;
1901 if (F_ISSET(node->mn_flags, F_BIGDATA))
1902 sz += sizeof(pgno_t);
1904 sz += NODEDSZ(node);
1907 ptr = mp->mp_ptrs[indx];
1908 numkeys = NUMKEYS(mp);
1909 for (i = j = 0; i < numkeys; i++) {
1911 mp->mp_ptrs[j] = mp->mp_ptrs[i];
1912 if (mp->mp_ptrs[i] < ptr)
1913 mp->mp_ptrs[j] += sz;
1918 base = (char *)mp + mp->mp_upper;
1919 bcopy(base, base + sz, ptr - mp->mp_upper);
1921 mp->mp_lower -= sizeof(indx_t);
1926 mdb_cursor_open(MDB_db *db, MDB_txn *txn, MDB_cursor **ret)
1930 if (db == NULL || ret == NULL)
1933 if (txn != NULL && db->md_env != txn->mt_env) {
1937 if ((cursor = calloc(1, sizeof(*cursor))) != NULL) {
1938 SLIST_INIT(&cursor->mc_stack);
1940 cursor->mc_txn = txn;
1949 mdb_cursor_close(MDB_cursor *cursor)
1951 if (cursor != NULL) {
1952 while (!CURSOR_EMPTY(cursor))
1953 cursor_pop_page(cursor);
1955 /* btree_close(cursor->bt); */
1961 mdb_update_key(MDB_db *db, MDB_page *mp, indx_t indx,
1964 indx_t ptr, i, numkeys;
1970 node = NODEPTR(mp, indx);
1971 ptr = mp->mp_ptrs[indx];
1972 DPRINTF("update key %u (ofs %u) [%.*s] to [%.*s] on page %lu",
1974 (int)node->mn_ksize, (char *)NODEKEY(node),
1975 (int)key->mv_size, (char *)key->mv_data,
1978 delta = key->mv_size - node->mn_ksize;
1980 if (delta > 0 && SIZELEFT(mp) < delta) {
1981 DPRINTF("OUCH! Not enough room, delta = %d", delta);
1985 numkeys = NUMKEYS(mp);
1986 for (i = 0; i < numkeys; i++) {
1987 if (mp->mp_ptrs[i] <= ptr)
1988 mp->mp_ptrs[i] -= delta;
1991 base = (char *)mp + mp->mp_upper;
1992 len = ptr - mp->mp_upper + NODESIZE;
1993 bcopy(base, base - delta, len);
1994 mp->mp_upper -= delta;
1996 node = NODEPTR(mp, indx);
1997 node->mn_ksize = key->mv_size;
2000 bcopy(key->mv_data, NODEKEY(node), key->mv_size);
2005 /* Move a node from src to dst.
2008 mdb_move_node(MDB_db *bt, MDB_pageparent *src, indx_t srcindx,
2009 MDB_pageparent *dst, indx_t dstindx)
2015 srcnode = NODEPTR(src->mp_page, srcindx);
2016 DPRINTF("moving %s node %u [%.*s] on page %lu to node %u on page %lu",
2017 IS_LEAF(src->mp_page) ? "leaf" : "branch",
2019 (int)srcnode->mn_ksize, (char *)NODEKEY(srcnode),
2020 src->mp_page->mp_pgno,
2021 dstindx, dst->mp_page->mp_pgno);
2023 /* Mark src and dst as dirty. */
2024 if ((rc = mdb_touch(bt->md_env->me_txn, src)) ||
2025 (rc = mdb_touch(bt->md_env->me_txn, dst)))
2028 /* Add the node to the destination page.
2030 key.mv_size = srcnode->mn_ksize;
2031 key.mv_data = NODEKEY(srcnode);
2032 data.mv_size = NODEDSZ(srcnode);
2033 data.mv_data = NODEDATA(srcnode);
2034 rc = mdb_add_node(bt, dst->mp_page, dstindx, &key, &data, NODEPGNO(srcnode),
2036 if (rc != MDB_SUCCESS)
2039 /* Delete the node from the source page.
2041 mdb_del_node(bt, src->mp_page, srcindx);
2043 /* Update the parent separators.
2045 if (srcindx == 0 && src->mp_pi != 0) {
2046 DPRINTF("update separator for source page %lu to [%.*s]",
2047 src->mp_page->mp_pgno, (int)key.mv_size, (char *)key.mv_data);
2048 if ((rc = mdb_update_key(bt, src->mp_parent, src->mp_pi,
2049 &key)) != MDB_SUCCESS)
2053 if (srcindx == 0 && IS_BRANCH(src->mp_page)) {
2055 nullkey.mv_size = 0;
2056 assert(mdb_update_key(bt, src->mp_page, 0, &nullkey) == MDB_SUCCESS);
2059 if (dstindx == 0 && dst->mp_pi != 0) {
2060 DPRINTF("update separator for destination page %lu to [%.*s]",
2061 dst->mp_page->mp_pgno, (int)key.mv_size, (char *)key.mv_data);
2062 if ((rc = mdb_update_key(bt, dst->mp_parent, dst->mp_pi,
2063 &key)) != MDB_SUCCESS)
2067 if (dstindx == 0 && IS_BRANCH(dst->mp_page)) {
2069 nullkey.mv_size = 0;
2070 assert(mdb_update_key(bt, dst->mp_page, 0, &nullkey) == MDB_SUCCESS);
2077 mdb_merge(MDB_db *bt, MDB_pageparent *src, MDB_pageparent *dst)
2086 DPRINTF("merging page %lu and %lu", src->mp_page->mp_pgno, dst->mp_page->mp_pgno);
2088 assert(src->mp_parent); /* can't merge root page */
2089 assert(dst->mp_parent);
2090 assert(bt->md_env->me_txn != NULL);
2092 /* Mark src and dst as dirty. */
2093 if ((rc = mdb_touch(bt->md_env->me_txn, src)) ||
2094 (rc = mdb_touch(bt->md_env->me_txn, dst)))
2097 /* Move all nodes from src to dst.
2099 for (i = 0; i < NUMKEYS(src->mp_page); i++) {
2100 srcnode = NODEPTR(src->mp_page, i);
2102 key.mv_size = srcnode->mn_ksize;
2103 key.mv_data = NODEKEY(srcnode);
2104 data.mv_size = NODEDSZ(srcnode);
2105 data.mv_data = NODEDATA(srcnode);
2106 rc = mdb_add_node(bt, dst->mp_page, NUMKEYS(dst->mp_page), &key,
2107 &data, NODEPGNO(srcnode), srcnode->mn_flags);
2108 if (rc != MDB_SUCCESS)
2112 DPRINTF("dst page %lu now has %u keys (%.1f%% filled)",
2113 dst->mp_page->mp_pgno, NUMKEYS(dst->mp_page), (float)PAGEFILL(bt->md_env, dst->mp_page) / 10);
2115 /* Unlink the src page from parent.
2117 mdb_del_node(bt, src->mp_parent, src->mp_pi);
2118 if (src->mp_pi == 0) {
2120 if ((rc = mdb_update_key(bt, src->mp_parent, 0, &key)) != MDB_SUCCESS)
2124 if (IS_LEAF(src->mp_page))
2125 bt->md_leaf_pages--;
2127 bt->md_branch_pages--;
2129 mpp.mp_page = src->mp_parent;
2130 dh = (MDB_dhead *)src->mp_parent;
2132 mpp.mp_parent = dh->md_parent;
2133 mpp.mp_pi = dh->md_pi;
2135 return mdb_rebalance(bt, &mpp);
2138 #define FILL_THRESHOLD 250
2141 mdb_rebalance(MDB_db *db, MDB_pageparent *mpp)
2146 indx_t si = 0, di = 0;
2149 assert(db->md_env->me_txn != NULL);
2150 assert(mpp != NULL);
2152 DPRINTF("rebalancing %s page %lu (has %u keys, %.1f%% full)",
2153 IS_LEAF(mpp->mp_page) ? "leaf" : "branch",
2154 mpp->mp_page->mp_pgno, NUMKEYS(mpp->mp_page), (float)PAGEFILL(db->md_env, mpp->mp_page) / 10);
2156 if (PAGEFILL(db->md_env, mpp->mp_page) >= FILL_THRESHOLD) {
2157 DPRINTF("no need to rebalance page %lu, above fill threshold",
2158 mpp->mp_page->mp_pgno);
2162 if (mpp->mp_parent == NULL) {
2163 if (NUMKEYS(mpp->mp_page) == 0) {
2164 DPRINTF("tree is completely empty");
2165 db->md_root = P_INVALID;
2167 db->md_leaf_pages--;
2168 } else if (IS_BRANCH(mpp->mp_page) && NUMKEYS(mpp->mp_page) == 1) {
2169 DPRINTF("collapsing root page!");
2170 db->md_root = NODEPGNO(NODEPTR(mpp->mp_page, 0));
2171 if ((root = mdbenv_get_page(db->md_env, db->md_root)) == NULL)
2174 db->md_branch_pages--;
2176 DPRINTF("root page doesn't need rebalancing");
2180 /* The parent (branch page) must have at least 2 pointers,
2181 * otherwise the tree is invalid.
2183 assert(NUMKEYS(mpp->mp_parent) > 1);
2185 /* Leaf page fill factor is below the threshold.
2186 * Try to move keys from left or right neighbor, or
2187 * merge with a neighbor page.
2192 if (mpp->mp_pi == 0) {
2193 /* We're the leftmost leaf in our parent.
2195 DPRINTF("reading right neighbor");
2196 node = NODEPTR(mpp->mp_parent, mpp->mp_pi + 1);
2197 if ((npp.mp_page = mdbenv_get_page(db->md_env, NODEPGNO(node))) == NULL)
2199 npp.mp_pi = mpp->mp_pi + 1;
2201 di = NUMKEYS(mpp->mp_page);
2203 /* There is at least one neighbor to the left.
2205 DPRINTF("reading left neighbor");
2206 node = NODEPTR(mpp->mp_parent, mpp->mp_pi - 1);
2207 if ((npp.mp_page = mdbenv_get_page(db->md_env, NODEPGNO(node))) == NULL)
2209 npp.mp_pi = mpp->mp_pi - 1;
2210 si = NUMKEYS(npp.mp_page) - 1;
2213 npp.mp_parent = mpp->mp_parent;
2215 DPRINTF("found neighbor page %lu (%u keys, %.1f%% full)",
2216 npp.mp_page->mp_pgno, NUMKEYS(npp.mp_page), (float)PAGEFILL(db->md_env, npp.mp_page) / 10);
2218 /* If the neighbor page is above threshold and has at least two
2219 * keys, move one key from it.
2221 * Otherwise we should try to merge them.
2223 if (PAGEFILL(db->md_env, npp.mp_page) >= FILL_THRESHOLD && NUMKEYS(npp.mp_page) >= 2)
2224 return mdb_move_node(db, &npp, si, mpp, di);
2225 else { /* FIXME: if (has_enough_room()) */
2226 if (mpp->mp_pi == 0)
2227 return mdb_merge(db, &npp, mpp);
2229 return mdb_merge(db, mpp, &npp);
2234 mdb_del(MDB_db *bt, MDB_txn *txn,
2235 MDB_val *key, MDB_val *data)
2242 DPRINTF("========> delete key %.*s", (int)key->mv_size, (char *)key->mv_data);
2244 assert(key != NULL);
2246 if (bt == NULL || txn == NULL)
2249 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
2253 if (bt->md_env->me_txn != txn) {
2257 if (key->mv_size == 0 || key->mv_size > MAXKEYSIZE) {
2261 if ((rc = mdb_search_page(bt, txn, key, NULL, 1, &mpp)) != MDB_SUCCESS)
2264 leaf = mdb_search_node(bt, mpp.mp_page, key, &exact, &ki);
2265 if (leaf == NULL || !exact) {
2269 if (data && (rc = mdb_read_data(bt, NULL, leaf, data)) != MDB_SUCCESS)
2272 mdb_del_node(bt, mpp.mp_page, ki);
2273 /* add overflow pages to free list */
2274 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) {
2278 bcopy(NODEDATA(leaf), &pg, sizeof(pg));
2279 ovpages = OVPAGES(NODEDSZ(leaf), bt->md_env->me_meta.mm_psize);
2280 for (i=0; i<ovpages; i++) {
2281 mdb_idl_insert(txn->mt_free_pgs, pg);
2286 rc = mdb_rebalance(bt, &mpp);
2287 if (rc != MDB_SUCCESS)
2288 txn->mt_flags |= MDB_TXN_ERROR;
2293 /* Split page <*mpp>, and insert <key,(data|newpgno)> in either left or
2294 * right sibling, at index <*newindxp> (as if unsplit). Updates *mpp and
2295 * *newindxp with the actual values after split, ie if *mpp and *newindxp
2296 * refer to a node in the new right sibling page.
2299 mdb_split(MDB_db *bt, MDB_page **mpp, unsigned int *newindxp,
2300 MDB_val *newkey, MDB_val *newdata, pgno_t newpgno)
2303 int rc = MDB_SUCCESS, ins_new = 0;
2306 unsigned int i, j, split_indx;
2308 MDB_val sepkey, rkey, rdata;
2310 MDB_dpage *mdp, *rdp, *pdp;
2314 assert(bt->md_env != NULL);
2316 dh = ((MDB_dhead *)*mpp) - 1;
2317 mdp = (MDB_dpage *)dh;
2318 newindx = *newindxp;
2320 DPRINTF("-----> splitting %s page %lu and adding [%.*s] at index %i",
2321 IS_LEAF(&mdp->p) ? "leaf" : "branch", mdp->p.mp_pgno,
2322 (int)newkey->mv_size, (char *)newkey->mv_data, *newindxp);
2324 if (mdp->h.md_parent == NULL) {
2325 if ((pdp = mdb_new_page(bt, P_BRANCH, 1)) == NULL)
2328 mdp->h.md_parent = &pdp->p;
2329 bt->md_env->me_txn->mt_root = pdp->p.mp_pgno;
2330 DPRINTF("root split! new root = %lu", pdp->p.mp_pgno);
2333 /* Add left (implicit) pointer. */
2334 if (mdb_add_node(bt, &pdp->p, 0, NULL, NULL,
2335 mdp->p.mp_pgno, 0) != MDB_SUCCESS)
2338 DPRINTF("parent branch page is %lu", mdp->h.md_parent->mp_pgno);
2341 /* Create a right sibling. */
2342 if ((rdp = mdb_new_page(bt, mdp->p.mp_flags, 1)) == NULL)
2344 rdp->h.md_parent = mdp->h.md_parent;
2345 rdp->h.md_pi = mdp->h.md_pi + 1;
2346 DPRINTF("new right sibling: page %lu", rdp->p.mp_pgno);
2348 /* Move half of the keys to the right sibling. */
2349 if ((copy = malloc(bt->md_env->me_meta.mm_psize)) == NULL)
2351 bcopy(&mdp->p, copy, bt->md_env->me_meta.mm_psize);
2352 bzero(&mdp->p.mp_ptrs, bt->md_env->me_meta.mm_psize - PAGEHDRSZ);
2353 mdp->p.mp_lower = PAGEHDRSZ;
2354 mdp->p.mp_upper = bt->md_env->me_meta.mm_psize;
2356 split_indx = NUMKEYS(copy) / 2 + 1;
2358 /* First find the separating key between the split pages.
2360 bzero(&sepkey, sizeof(sepkey));
2361 if (newindx == split_indx) {
2362 sepkey.mv_size = newkey->mv_size;
2363 sepkey.mv_data = newkey->mv_data;
2365 node = NODEPTR(copy, split_indx);
2366 sepkey.mv_size = node->mn_ksize;
2367 sepkey.mv_data = NODEKEY(node);
2370 DPRINTF("separator is [%.*s]", (int)sepkey.mv_size, (char *)sepkey.mv_data);
2372 /* Copy separator key to the parent.
2374 if (SIZELEFT(rdp->h.md_parent) < mdb_branch_size(bt, &sepkey)) {
2375 rc = mdb_split(bt, &rdp->h.md_parent, &rdp->h.md_pi,
2376 &sepkey, NULL, rdp->p.mp_pgno);
2378 /* Right page might now have changed parent.
2379 * Check if left page also changed parent.
2381 if (rdp->h.md_parent != mdp->h.md_parent &&
2382 mdp->h.md_pi >= NUMKEYS(mdp->h.md_parent)) {
2383 mdp->h.md_parent = rdp->h.md_parent;
2384 mdp->h.md_pi = rdp->h.md_pi - 1;
2387 rc = mdb_add_node(bt, rdp->h.md_parent, rdp->h.md_pi,
2388 &sepkey, NULL, rdp->p.mp_pgno, 0);
2390 if (rc != MDB_SUCCESS) {
2395 for (i = j = 0; i <= NUMKEYS(copy); j++) {
2396 if (i < split_indx) {
2397 /* Re-insert in left sibling. */
2400 /* Insert in right sibling. */
2401 if (i == split_indx)
2402 /* Reset insert index for right sibling. */
2403 j = (i == newindx && ins_new);
2407 if (i == newindx && !ins_new) {
2408 /* Insert the original entry that caused the split. */
2409 rkey.mv_data = newkey->mv_data;
2410 rkey.mv_size = newkey->mv_size;
2411 if (IS_LEAF(&mdp->p)) {
2412 rdata.mv_data = newdata->mv_data;
2413 rdata.mv_size = newdata->mv_size;
2420 /* Update page and index for the new key. */
2423 } else if (i == NUMKEYS(copy)) {
2426 node = NODEPTR(copy, i);
2427 rkey.mv_data = NODEKEY(node);
2428 rkey.mv_size = node->mn_ksize;
2429 if (IS_LEAF(&mdp->p)) {
2430 rdata.mv_data = NODEDATA(node);
2431 rdata.mv_size = node->mn_dsize;
2433 pgno = node->mn_pgno;
2434 flags = node->mn_flags;
2439 if (!IS_LEAF(&mdp->p) && j == 0) {
2440 /* First branch index doesn't need key data. */
2444 rc = mdb_add_node(bt, &pdp->p, j, &rkey, &rdata, pgno,flags);
2452 mdb_put(MDB_db *bt, MDB_txn *txn,
2453 MDB_val *key, MDB_val *data, unsigned int flags)
2455 int rc = MDB_SUCCESS, exact;
2460 assert(key != NULL);
2461 assert(data != NULL);
2463 if (bt == NULL || txn == NULL)
2466 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
2470 if (bt->md_env->me_txn != txn) {
2474 if (key->mv_size == 0 || key->mv_size > MAXKEYSIZE) {
2478 DPRINTF("==> put key %.*s, size %zu, data size %zu",
2479 (int)key->mv_size, (char *)key->mv_data, key->mv_size, data->mv_size);
2481 rc = mdb_search_page(bt, txn, key, NULL, 1, &mpp);
2482 if (rc == MDB_SUCCESS) {
2483 leaf = mdb_search_node(bt, mpp.mp_page, key, &exact, &ki);
2484 if (leaf && exact) {
2485 if (F_ISSET(flags, MDB_NOOVERWRITE)) {
2486 DPRINTF("duplicate key %.*s",
2487 (int)key->mv_size, (char *)key->mv_data);
2490 mdb_del_node(bt, mpp.mp_page, ki);
2492 if (leaf == NULL) { /* append if not found */
2493 ki = NUMKEYS(mpp.mp_page);
2494 DPRINTF("appending key at index %i", ki);
2496 } else if (rc == ENOENT) {
2498 /* new file, just write a root leaf page */
2499 DPRINTF("allocating new root leaf page");
2500 if ((dp = mdb_new_page(bt, P_LEAF, 1)) == NULL) {
2503 mpp.mp_page = &dp->p;
2504 txn->mt_root = mpp.mp_page->mp_pgno;
2511 assert(IS_LEAF(mpp.mp_page));
2512 DPRINTF("there are %u keys, should insert new key at index %i",
2513 NUMKEYS(mpp.mp_page), ki);
2515 if (SIZELEFT(mpp.mp_page) < mdb_leaf_size(bt, key, data)) {
2516 rc = mdb_split(bt, &mpp.mp_page, &ki, key, data, P_INVALID);
2518 /* There is room already in this leaf page. */
2519 rc = mdb_add_node(bt, mpp.mp_page, ki, key, data, 0, 0);
2522 if (rc != MDB_SUCCESS)
2523 txn->mt_flags |= MDB_TXN_ERROR;
2532 mdbenv_get_flags(MDB_env *env, unsigned int *arg)
2537 *arg = env->me_flags;
2542 mdbenv_get_path(MDB_env *env, const char **arg)
2547 *arg = env->me_path;
2552 mdbenv_stat(MDB_env *env, MDB_stat *arg)
2554 if (env == NULL || arg == NULL)
2557 arg->ms_psize = env->me_meta.mm_psize;
2558 arg->ms_depth = env->me_meta.mm_depth;
2559 arg->ms_branch_pages = env->me_meta.mm_branch_pages;
2560 arg->ms_leaf_pages = env->me_meta.mm_leaf_pages;
2561 arg->ms_overflow_pages = env->me_meta.mm_overflow_pages;
2562 arg->ms_entries = env->me_meta.mm_entries;
2567 int mdb_open(MDB_txn *txn, const char *name, unsigned int flags, MDB_dbi *dbi)
2576 int mdb_stat(MDB_db *db, MDB_stat *arg)
2578 if (db == NULL || arg == NULL)
2581 arg->ms_psize = db->md_env->me_meta.mm_psize;
2582 if (!db->md_parent) {
2583 arg->ms_depth = db->md_env->me_meta.mm_depth;
2584 arg->ms_branch_pages = db->md_env->me_meta.mm_branch_pages;
2585 arg->ms_leaf_pages = db->md_env->me_meta.mm_leaf_pages;
2586 arg->ms_overflow_pages = db->md_env->me_meta.mm_overflow_pages;
2587 arg->ms_entries = db->md_env->me_meta.mm_entries;
2589 arg->ms_depth = db->md_depth;
2590 arg->ms_branch_pages = db->md_branch_pages;
2591 arg->ms_leaf_pages = db->md_leaf_pages;
2592 arg->ms_overflow_pages = db->md_overflow_pages;
2593 arg->ms_entries = db->md_entries;
2599 void mdb_close(MDB_db *db)