1 /* mdb.c - memory-mapped database library */
3 * Copyright 2011 Howard Chu, Symas Corp.
6 * Redistribution and use in source and binary forms, with or without
7 * modification, are permitted only as authorized by the OpenLDAP
10 * A copy of this license is available in the file LICENSE in the
11 * top-level directory of the distribution or, alternatively, at
12 * <http://www.OpenLDAP.org/license.html>.
14 * This code is derived from btree.c written by Martin Hedenfalk.
16 * Copyright (c) 2009, 2010 Martin Hedenfalk <martin@bzero.se>
18 * Permission to use, copy, modify, and distribute this software for any
19 * purpose with or without fee is hereby granted, provided that the above
20 * copyright notice and this permission notice appear in all copies.
22 * THE SOFTWARE IS PROVIDED "AS IS" AND THE AUTHOR DISCLAIMS ALL WARRANTIES
23 * WITH REGARD TO THIS SOFTWARE INCLUDING ALL IMPLIED WARRANTIES OF
24 * MERCHANTABILITY AND FITNESS. IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR
25 * ANY SPECIAL, DIRECT, INDIRECT, OR CONSEQUENTIAL DAMAGES OR ANY DAMAGES
26 * WHATSOEVER RESULTING FROM LOSS OF USE, DATA OR PROFITS, WHETHER IN AN
27 * ACTION OF CONTRACT, NEGLIGENCE OR OTHER TORTIOUS ACTION, ARISING OUT OF
28 * OR IN CONNECTION WITH THE USE OR PERFORMANCE OF THIS SOFTWARE.
30 #include <sys/types.h>
32 #include <sys/queue.h>
33 #include <sys/param.h>
36 #ifdef HAVE_SYS_FILE_H
57 #define ULONG unsigned long
66 #if DEBUG && defined(__GNUC__)
67 # define DPRINTF(fmt, ...) \
68 fprintf(stderr, "%s:%d: " fmt "\n", __func__, __LINE__, ##__VA_ARGS__)
70 # define DPRINTF(...) ((void) 0)
75 #define MDB_MAGIC 0xBEEFC0DE
77 #define MAXKEYSIZE 255
79 #define P_INVALID (~0UL)
81 #define F_ISSET(w, f) (((w) & (f)) == (f))
83 typedef uint16_t indx_t;
85 #define DEFAULT_READERS 126
86 #define DEFAULT_MAPSIZE 1048576
88 /* Lock descriptor stuff */
93 typedef struct MDB_rxbody {
98 #define CACHELINE 64 /* most CPUs. Itanium uses 128 */
101 typedef struct MDB_reader {
103 /* cache line alignment */
104 char pad[CACHELINE-sizeof(MDB_rxbody)];
109 uint32_t mt_version; \
110 pthread_mutex_t mt_mutex; \
112 uint32_t mt_numreaders
113 typedef struct MDB_txbody {
117 typedef struct MDB_txninfo {
119 char pad[CACHELINE-sizeof(MDB_txbody)];
120 pthread_mutex_t mt_wmutex;
121 char pad2[CACHELINE-sizeof(pthread_mutex_t)];
122 MDB_reader mt_readers[1];
125 /* Common header for all page types. Overflow pages
126 * occupy a number of contiguous pages with no
127 * headers on any page after the first.
129 typedef struct MDB_page { /* represents a page of storage */
130 #define mp_pgno mp_p.p_pgno
132 pgno_t p_pgno; /* page number */
135 #define P_BRANCH 0x01 /* branch page */
136 #define P_LEAF 0x02 /* leaf page */
137 #define P_OVERFLOW 0x04 /* overflow page */
138 #define P_META 0x08 /* meta page */
139 #define P_DIRTY 0x10 /* dirty page */
141 #define mp_lower mp_pb.pb.pb_lower
142 #define mp_upper mp_pb.pb.pb_upper
143 #define mp_pages mp_pb.pb_pages
146 indx_t pb_lower; /* lower bound of free space */
147 indx_t pb_upper; /* upper bound of free space */
149 uint32_t pb_pages; /* number of overflow pages */
151 indx_t mp_ptrs[1]; /* dynamic size */
154 #define PAGEHDRSZ ((unsigned) offsetof(MDB_page, mp_ptrs))
156 #define NUMKEYS(p) (((p)->mp_lower - PAGEHDRSZ) >> 1)
157 #define SIZELEFT(p) (indx_t)((p)->mp_upper - (p)->mp_lower)
158 #define PAGEFILL(env, p) (1000L * ((env)->me_psize - PAGEHDRSZ - SIZELEFT(p)) / \
159 ((env)->me_psize - PAGEHDRSZ))
160 #define IS_LEAF(p) F_ISSET((p)->mp_flags, P_LEAF)
161 #define IS_BRANCH(p) F_ISSET((p)->mp_flags, P_BRANCH)
162 #define IS_OVERFLOW(p) F_ISSET((p)->mp_flags, P_OVERFLOW)
164 #define OVPAGES(size, psize) (PAGEHDRSZ + size + psize - 1) / psize;
166 typedef struct MDB_db {
170 ULONG md_branch_pages;
172 ULONG md_overflow_pages;
180 typedef struct MDB_meta { /* meta (footer) page content */
183 void *mm_address; /* address for fixed mapping */
184 size_t mm_mapsize; /* size of mmap region */
185 MDB_db mm_dbs[2]; /* first is free space, 2nd is main db */
186 #define mm_psize mm_dbs[0].md_pad
187 #define mm_flags mm_dbs[0].md_flags
188 pgno_t mm_last_pg; /* last used page in file */
189 ULONG mm_txnid; /* txnid that committed this page */
192 typedef struct MDB_dhead { /* a dirty page */
193 STAILQ_ENTRY(MDB_dpage) md_next; /* queue of dirty pages */
195 unsigned md_pi; /* parent index */
199 typedef struct MDB_dpage {
204 STAILQ_HEAD(dirty_queue, MDB_dpage); /* FIXME: use a sorted data structure */
206 typedef struct MDB_oldpages {
207 struct MDB_oldpages *mo_next;
209 pgno_t mo_pages[1]; /* dynamic */
212 typedef struct MDB_pageparent {
218 static MDB_dpage *mdb_alloc_page(MDB_txn *txn, MDB_page *parent, unsigned int parent_idx, int num);
219 static int mdb_touch(MDB_txn *txn, MDB_pageparent *mp);
221 typedef struct MDB_ppage { /* ordered list of pages */
222 SLIST_ENTRY(MDB_ppage) mp_entry;
224 unsigned int mp_ki; /* cursor index on page */
226 SLIST_HEAD(page_stack, MDB_ppage);
228 #define CURSOR_EMPTY(c) SLIST_EMPTY(&(c)->mc_stack)
229 #define CURSOR_TOP(c) SLIST_FIRST(&(c)->mc_stack)
230 #define CURSOR_POP(c) SLIST_REMOVE_HEAD(&(c)->mc_stack, mp_entry)
231 #define CURSOR_PUSH(c,p) SLIST_INSERT_HEAD(&(c)->mc_stack, p, mp_entry)
237 struct page_stack mc_stack; /* stack of parent pages */
239 short mc_initialized; /* 1 if initialized */
240 short mc_eof; /* 1 if end is reached */
241 struct MDB_xcursor *mc_xcursor;
244 #define METAHASHLEN offsetof(MDB_meta, mm_hash)
245 #define METADATA(p) ((void *)((char *)p + PAGEHDRSZ))
247 typedef struct MDB_node {
248 #define mn_pgno mn_p.np_pgno
249 #define mn_dsize mn_p.np_dsize
251 pgno_t np_pgno; /* child page number */
252 uint32_t np_dsize; /* leaf data size */
254 unsigned int mn_flags:4;
255 unsigned int mn_ksize:12; /* key size */
256 #define F_BIGDATA 0x01 /* data put on overflow page */
261 typedef struct MDB_dbx {
263 MDB_cmp_func *md_cmp; /* user compare function */
264 MDB_cmp_func *md_dcmp; /* user dupsort function */
265 MDB_rel_func *md_rel; /* user relocate function */
267 unsigned int md_dirty;
271 pgno_t mt_next_pgno; /* next unallocated page */
275 pgno_t *mt_free_pgs; /* this is an IDL */
277 struct dirty_queue *dirty_queue; /* modified pages */
280 MDB_dbx *mt_dbxs; /* array */
282 unsigned int mt_numdbs;
284 #define MDB_TXN_RDONLY 0x01 /* read-only transaction */
285 #define MDB_TXN_ERROR 0x02 /* an error has occurred */
286 #define MDB_TXN_METOGGLE 0x04 /* used meta page 1 */
287 unsigned int mt_flags;
290 /* Context for sorted-dup records */
291 typedef struct MDB_xcursor {
292 MDB_cursor mx_cursor;
302 unsigned int me_maxreaders;
303 unsigned int me_numdbs;
304 unsigned int me_maxdbs;
307 MDB_txninfo *me_txns;
308 MDB_meta *me_metas[2];
310 MDB_txn *me_txn; /* current write transaction */
312 off_t me_size; /* current file size */
313 unsigned int me_psize;
315 MDB_dbx *me_dbxs; /* array */
317 MDB_oldpages *me_pghead;
318 pthread_key_t me_txkey; /* thread-key for readers */
321 #define NODESIZE offsetof(MDB_node, mn_data)
323 #define INDXSIZE(k) (NODESIZE + ((k) == NULL ? 0 : (k)->mv_size))
324 #define LEAFSIZE(k, d) (NODESIZE + (k)->mv_size + (d)->mv_size)
325 #define NODEPTR(p, i) ((MDB_node *)((char *)(p) + (p)->mp_ptrs[i]))
326 #define NODEKEY(node) (void *)((node)->mn_data)
327 #define NODEDATA(node) (void *)((char *)(node)->mn_data + (node)->mn_ksize)
328 #define NODEPGNO(node) ((node)->mn_pgno)
329 #define NODEDSZ(node) ((node)->mn_dsize)
331 #define MDB_COMMIT_PAGES 64 /* max number of pages to write in one commit */
332 #define MDB_MAXCACHE_DEF 1024 /* max number of pages to keep in cache */
334 static int mdb_search_page_root(MDB_txn *txn,
335 MDB_dbi dbi, MDB_val *key,
336 MDB_cursor *cursor, int modify,
337 MDB_pageparent *mpp);
338 static int mdb_search_page(MDB_txn *txn,
339 MDB_dbi dbi, MDB_val *key,
340 MDB_cursor *cursor, int modify,
341 MDB_pageparent *mpp);
343 static int mdbenv_read_header(MDB_env *env, MDB_meta *meta);
344 static int mdbenv_read_meta(MDB_env *env, int *which);
345 static int mdbenv_write_meta(MDB_txn *txn);
346 static MDB_page *mdb_get_page(MDB_txn *txn, pgno_t pgno);
348 static MDB_node *mdb_search_node(MDB_txn *txn, MDB_dbi dbi, MDB_page *mp,
349 MDB_val *key, int *exactp, unsigned int *kip);
350 static int mdb_add_node(MDB_txn *txn, MDB_dbi dbi, MDB_page *mp,
351 indx_t indx, MDB_val *key, MDB_val *data,
352 pgno_t pgno, uint8_t flags);
353 static void mdb_del_node(MDB_page *mp, indx_t indx);
354 static int mdb_del0(MDB_txn *txn, MDB_dbi dbi, unsigned int ki,
355 MDB_pageparent *mpp, MDB_node *leaf);
356 static int mdb_read_data(MDB_txn *txn, MDB_node *leaf, MDB_val *data);
358 static int mdb_rebalance(MDB_txn *txn, MDB_dbi dbi, MDB_pageparent *mp);
359 static int mdb_update_key(MDB_page *mp, indx_t indx, MDB_val *key);
360 static int mdb_move_node(MDB_txn *txn, MDB_dbi dbi,
361 MDB_pageparent *src, indx_t srcindx,
362 MDB_pageparent *dst, indx_t dstindx);
363 static int mdb_merge(MDB_txn *txn, MDB_dbi dbi, MDB_pageparent *src,
364 MDB_pageparent *dst);
365 static int mdb_split(MDB_txn *txn, MDB_dbi dbi, MDB_page **mpp,
366 unsigned int *newindxp, MDB_val *newkey,
367 MDB_val *newdata, pgno_t newpgno);
368 static MDB_dpage *mdb_new_page(MDB_txn *txn, MDB_dbi dbi, uint32_t flags, int num);
370 static void cursor_pop_page(MDB_cursor *cursor);
371 static MDB_ppage *cursor_push_page(MDB_cursor *cursor,
374 static int mdb_set_key(MDB_node *node, MDB_val *key);
375 static int mdb_sibling(MDB_cursor *cursor, int move_right);
376 static int mdb_cursor_next(MDB_cursor *cursor,
377 MDB_val *key, MDB_val *data);
378 static int mdb_cursor_prev(MDB_cursor *cursor,
379 MDB_val *key, MDB_val *data);
380 static int mdb_cursor_set(MDB_cursor *cursor,
381 MDB_val *key, MDB_val *data, int *exactp);
382 static int mdb_cursor_first(MDB_cursor *cursor,
383 MDB_val *key, MDB_val *data);
384 static int mdb_cursor_last(MDB_cursor *cursor,
385 MDB_val *key, MDB_val *data);
387 static size_t mdb_leaf_size(MDB_env *env, MDB_val *key,
389 static size_t mdb_branch_size(MDB_env *env, MDB_val *key);
391 static int memncmp(const void *s1, size_t n1,
392 const void *s2, size_t n2);
393 static int memnrcmp(const void *s1, size_t n1,
394 const void *s2, size_t n2);
397 memncmp(const void *s1, size_t n1, const void *s2, size_t n2)
399 int diff, len_diff = -1;
402 len_diff = (n1 > n2);
405 diff = memcmp(s1, s2, n1);
406 return diff ? diff : len_diff;
410 memnrcmp(const void *s1, size_t n1, const void *s2, size_t n2)
412 const unsigned char *p1, *p2, *p1_lim;
419 p1 = (const unsigned char *)s1 + n1 - 1;
420 p2 = (const unsigned char *)s2 + n2 - 1;
422 for (p1_lim = (n1 <= n2 ? s1 : s2); *p1 == *p2; p1--, p2--) {
424 return (p1 != s1) ? (p1 != p2) : (p2 != s2) ? -1 : 0;
430 mdb_cmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
432 return txn->mt_dbxs[dbi].md_cmp(a, b);
436 _mdb_cmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *key1, const MDB_val *key2)
438 if (F_ISSET(txn->mt_dbs[dbi].md_flags, MDB_REVERSEKEY))
439 return memnrcmp(key1->mv_data, key1->mv_size, key2->mv_data, key2->mv_size);
441 return memncmp((char *)key1->mv_data, key1->mv_size, key2->mv_data, key2->mv_size);
444 /* Allocate new page(s) for writing */
446 mdb_alloc_page(MDB_txn *txn, MDB_page *parent, unsigned int parent_idx, int num)
449 pgno_t pgno = P_INVALID;
450 ULONG oldest = txn->mt_txnid - 2;
452 if (!txn->mt_env->me_pghead && txn->mt_dbs[FREE_DBI].md_root != P_INVALID) {
453 /* See if there's anything in the free DB */
458 mpp.mp_parent = NULL;
460 mdb_search_page(txn, FREE_DBI, NULL, NULL, 0, &mpp);
461 leaf = NODEPTR(mpp.mp_page, 0);
462 kptr = (ULONG *)NODEKEY(leaf);
464 /* It's potentially usable, unless there are still
465 * older readers outstanding. Grab it.
467 if (oldest > *kptr) {
472 mdb_read_data(txn, leaf, &data);
473 idl = (ULONG *)data.mv_data;
474 mop = malloc(sizeof(MDB_oldpages) + MDB_IDL_SIZEOF(idl) - sizeof(pgno_t));
475 mop->mo_next = txn->mt_env->me_pghead;
476 mop->mo_txnid = *kptr;
477 txn->mt_env->me_pghead = mop;
478 memcpy(mop->mo_pages, idl, MDB_IDL_SIZEOF(idl));
483 DPRINTF("IDL read txn %lu root %lu num %lu",
484 mop->mo_txnid, txn->mt_dbs[FREE_DBI].md_root, idl[0]);
485 for (i=0; i<idl[0]; i++) {
486 DPRINTF("IDL %lu", idl[i+1]);
490 /* drop this IDL from the DB */
491 mpp.mp_parent = NULL;
493 mdb_search_page(txn, FREE_DBI, NULL, NULL, 1, &mpp);
494 leaf = NODEPTR(mpp.mp_page, 0);
495 mdb_del0(txn, FREE_DBI, 0, &mpp, leaf);
498 if (txn->mt_env->me_pghead) {
500 for (i=0; i<txn->mt_env->me_txns->mt_numreaders; i++) {
501 ULONG mr = txn->mt_env->me_txns->mt_readers[i].mr_txnid;
504 oldest = txn->mt_env->me_txns->mt_readers[i].mr_txnid;
506 if (oldest > txn->mt_env->me_pghead->mo_txnid) {
507 MDB_oldpages *mop = txn->mt_env->me_pghead;
508 txn->mt_oldest = oldest;
510 /* FIXME: For now, always use fresh pages. We
511 * really ought to search the free list for a
516 /* peel pages off tail, so we only have to truncate the list */
517 pgno = MDB_IDL_LAST(mop->mo_pages);
518 if (MDB_IDL_IS_RANGE(mop->mo_pages)) {
520 if (mop->mo_pages[2] > mop->mo_pages[1])
521 mop->mo_pages[0] = 0;
525 if (MDB_IDL_IS_ZERO(mop->mo_pages)) {
526 txn->mt_env->me_pghead = mop->mo_next;
533 if ((dp = malloc(txn->mt_env->me_psize * num + sizeof(MDB_dhead))) == NULL)
536 dp->h.md_parent = parent;
537 dp->h.md_pi = parent_idx;
538 STAILQ_INSERT_TAIL(txn->mt_u.dirty_queue, dp, h.md_next);
539 if (pgno == P_INVALID) {
540 dp->p.mp_pgno = txn->mt_next_pgno;
541 txn->mt_next_pgno += num;
543 dp->p.mp_pgno = pgno;
549 /* Touch a page: make it dirty and re-insert into tree with updated pgno.
552 mdb_touch(MDB_txn *txn, MDB_pageparent *pp)
554 MDB_page *mp = pp->mp_page;
559 if (!F_ISSET(mp->mp_flags, P_DIRTY)) {
561 if ((dp = mdb_alloc_page(txn, pp->mp_parent, pp->mp_pi, 1)) == NULL)
563 DPRINTF("touched page %lu -> %lu", mp->mp_pgno, dp->p.mp_pgno);
564 mdb_idl_insert(txn->mt_free_pgs, mp->mp_pgno);
565 pgno = dp->p.mp_pgno;
566 memcpy(&dp->p, mp, txn->mt_env->me_psize);
569 mp->mp_flags |= P_DIRTY;
571 /* Update the page number to new touched page. */
572 if (pp->mp_parent != NULL)
573 NODEPGNO(NODEPTR(pp->mp_parent, pp->mp_pi)) = mp->mp_pgno;
580 mdbenv_sync(MDB_env *env)
583 if (!F_ISSET(env->me_flags, MDB_NOSYNC)) {
584 if (fsync(env->me_fd))
591 mdb_txn_begin(MDB_env *env, int rdonly, MDB_txn **ret)
596 if ((txn = calloc(1, sizeof(*txn))) == NULL) {
597 DPRINTF("calloc: %s", strerror(errno));
602 txn->mt_flags |= MDB_TXN_RDONLY;
604 txn->mt_u.dirty_queue = calloc(1, sizeof(*txn->mt_u.dirty_queue));
605 if (txn->mt_u.dirty_queue == NULL) {
609 STAILQ_INIT(txn->mt_u.dirty_queue);
611 pthread_mutex_lock(&env->me_txns->mt_wmutex);
612 env->me_txns->mt_txnid++;
613 txn->mt_free_pgs = malloc(MDB_IDL_UM_SIZEOF);
614 if (txn->mt_free_pgs == NULL) {
615 free(txn->mt_u.dirty_queue);
619 txn->mt_free_pgs[0] = 0;
622 txn->mt_txnid = env->me_txns->mt_txnid;
624 MDB_reader *r = pthread_getspecific(env->me_txkey);
627 pthread_mutex_lock(&env->me_txns->mt_mutex);
628 for (i=0; i<env->me_maxreaders; i++) {
629 if (env->me_txns->mt_readers[i].mr_pid == 0) {
630 env->me_txns->mt_readers[i].mr_pid = getpid();
631 env->me_txns->mt_readers[i].mr_tid = pthread_self();
632 r = &env->me_txns->mt_readers[i];
633 pthread_setspecific(env->me_txkey, r);
634 if (i >= env->me_txns->mt_numreaders)
635 env->me_txns->mt_numreaders = i+1;
639 pthread_mutex_unlock(&env->me_txns->mt_mutex);
640 if (i == env->me_maxreaders) {
644 r->mr_txnid = txn->mt_txnid;
645 txn->mt_u.reader = r;
652 if ((rc = mdbenv_read_meta(env, &toggle)) != MDB_SUCCESS) {
657 /* Copy the DB arrays */
658 txn->mt_numdbs = env->me_numdbs;
659 txn->mt_dbxs = env->me_dbxs; /* mostly static anyway */
660 txn->mt_dbs = malloc(env->me_maxdbs * sizeof(MDB_db));
661 memcpy(txn->mt_dbs, env->me_meta->mm_dbs, 2 * sizeof(MDB_db));
662 if (txn->mt_numdbs > 2)
663 memcpy(txn->mt_dbs+2, env->me_dbs[env->me_db_toggle]+2,
664 (txn->mt_numdbs - 2) * sizeof(MDB_db));
668 txn->mt_flags |= MDB_TXN_METOGGLE;
669 txn->mt_next_pgno = env->me_meta->mm_last_pg+1;
672 DPRINTF("begin transaction %lu on mdbenv %p, root page %lu",
673 txn->mt_txnid, (void *) env, txn->mt_dbs[MAIN_DBI].md_root);
680 mdb_txn_abort(MDB_txn *txn)
689 DPRINTF("abort transaction %lu on mdbenv %p, root page %lu",
690 txn->mt_txnid, (void *) env, txn->mt_dbs[MAIN_DBI].md_root);
694 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
695 txn->mt_u.reader->mr_txnid = 0;
700 /* Discard all dirty pages. */
701 while (!STAILQ_EMPTY(txn->mt_u.dirty_queue)) {
702 dp = STAILQ_FIRST(txn->mt_u.dirty_queue);
703 STAILQ_REMOVE_HEAD(txn->mt_u.dirty_queue, h.md_next);
706 free(txn->mt_free_pgs);
707 free(txn->mt_u.dirty_queue);
709 while ((mop = txn->mt_env->me_pghead)) {
710 txn->mt_env->me_pghead = mop->mo_next;
715 env->me_txns->mt_txnid--;
716 for (i=2; i<env->me_numdbs; i++)
717 env->me_dbxs[i].md_dirty = 0;
718 pthread_mutex_unlock(&env->me_txns->mt_wmutex);
725 mdb_txn_commit(MDB_txn *txn)
734 struct iovec iov[MDB_COMMIT_PAGES];
737 assert(txn->mt_env != NULL);
741 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
742 DPRINTF("attempt to commit read-only transaction");
747 if (txn != env->me_txn) {
748 DPRINTF("attempt to commit unknown transaction");
753 if (F_ISSET(txn->mt_flags, MDB_TXN_ERROR)) {
754 DPRINTF("error flag is set, can't commit");
759 if (STAILQ_EMPTY(txn->mt_u.dirty_queue))
762 DPRINTF("committing transaction %lu on mdbenv %p, root page %lu",
763 txn->mt_txnid, (void *) env, txn->mt_dbs[MAIN_DBI].md_root);
765 /* should only be one record now */
766 if (env->me_pghead) {
770 mop = env->me_pghead;
771 key.mv_size = sizeof(pgno_t);
772 key.mv_data = (char *)&mop->mo_txnid;
773 data.mv_size = MDB_IDL_SIZEOF(mop->mo_pages);
774 data.mv_data = mop->mo_pages;
775 mdb_put(txn, FREE_DBI, &key, &data, 0);
776 free(env->me_pghead);
777 env->me_pghead = NULL;
779 /* save to free list */
780 if (!MDB_IDL_IS_ZERO(txn->mt_free_pgs)) {
784 /* make sure last page of freeDB is touched and on freelist */
785 key.mv_size = MAXKEYSIZE+1;
787 mpp.mp_parent = NULL;
789 mdb_search_page(txn, FREE_DBI, &key, NULL, 1, &mpp);
794 ULONG *idl = txn->mt_free_pgs;
795 DPRINTF("IDL write txn %lu root %lu num %lu",
796 txn->mt_txnid, txn->mt_dbs[FREE_DBI].md_root, idl[0]);
797 for (i=0; i<idl[0]; i++) {
798 DPRINTF("IDL %lu", idl[i+1]);
802 /* write to last page of freeDB */
803 key.mv_size = sizeof(pgno_t);
804 key.mv_data = (char *)&txn->mt_txnid;
805 data.mv_size = MDB_IDL_SIZEOF(txn->mt_free_pgs);
806 data.mv_data = txn->mt_free_pgs;
807 mdb_put(txn, FREE_DBI, &key, &data, 0);
810 /* Update DB root pointers. Their pages have already been
811 * touched so this is all in-place and cannot fail.
815 data.mv_size = sizeof(MDB_db);
817 for (i = 2; i < txn->mt_numdbs; i++) {
818 if (txn->mt_dbxs[i].md_dirty) {
819 data.mv_data = &txn->mt_dbs[i];
820 mdb_put(txn, i, &txn->mt_dbxs[i].md_name, &data, 0);
825 /* Commit up to MDB_COMMIT_PAGES dirty pages to disk until done.
832 STAILQ_FOREACH(dp, txn->mt_u.dirty_queue, h.md_next) {
833 if (dp->p.mp_pgno != next) {
835 DPRINTF("committing %u dirty pages", n);
836 rc = writev(env->me_fd, iov, n);
840 DPRINTF("short write, filesystem full?");
842 DPRINTF("writev: %s", strerror(errno));
849 lseek(env->me_fd, dp->p.mp_pgno * env->me_psize, SEEK_SET);
850 next = dp->p.mp_pgno;
852 DPRINTF("committing page %lu", dp->p.mp_pgno);
853 iov[n].iov_len = env->me_psize * dp->h.md_num;
854 iov[n].iov_base = &dp->p;
855 size += iov[n].iov_len;
856 next = dp->p.mp_pgno + dp->h.md_num;
857 /* clear dirty flag */
858 dp->p.mp_flags &= ~P_DIRTY;
859 if (++n >= MDB_COMMIT_PAGES) {
868 DPRINTF("committing %u dirty pages", n);
869 rc = writev(env->me_fd, iov, n);
873 DPRINTF("short write, filesystem full?");
875 DPRINTF("writev: %s", strerror(errno));
882 /* Drop the dirty pages.
884 while (!STAILQ_EMPTY(txn->mt_u.dirty_queue)) {
885 dp = STAILQ_FIRST(txn->mt_u.dirty_queue);
886 STAILQ_REMOVE_HEAD(txn->mt_u.dirty_queue, h.md_next);
890 if ((n = mdbenv_sync(env)) != 0 ||
891 (n = mdbenv_write_meta(txn)) != MDB_SUCCESS ||
892 (n = mdbenv_sync(env)) != 0) {
898 /* update the DB tables */
900 int toggle = !env->me_db_toggle;
902 for (i = 2; i < env->me_numdbs; i++) {
903 if (txn->mt_dbxs[i].md_dirty) {
904 env->me_dbs[toggle][i] = txn->mt_dbs[i];
905 txn->mt_dbxs[i].md_dirty = 0;
908 for (i = env->me_numdbs; i < txn->mt_numdbs; i++) {
909 txn->mt_dbxs[i].md_dirty = 0;
910 env->me_dbxs[i] = txn->mt_dbxs[i];
911 env->me_dbs[toggle][i] = txn->mt_dbs[i];
913 env->me_db_toggle = toggle;
914 env->me_numdbs = txn->mt_numdbs;
919 pthread_mutex_unlock(&env->me_txns->mt_wmutex);
920 free(txn->mt_free_pgs);
921 free(txn->mt_u.dirty_queue);
932 mdbenv_read_header(MDB_env *env, MDB_meta *meta)
941 /* We don't know the page size yet, so use a minimum value.
944 if ((rc = pread(env->me_fd, page, PAGESIZE, 0)) == 0) {
946 } else if (rc != PAGESIZE) {
949 DPRINTF("read: %s", strerror(errno));
953 p = (MDB_page *)page;
955 if (!F_ISSET(p->mp_flags, P_META)) {
956 DPRINTF("page %lu not a meta page", p->mp_pgno);
961 if (m->mm_magic != MDB_MAGIC) {
962 DPRINTF("meta has invalid magic");
966 if (m->mm_version != MDB_VERSION) {
967 DPRINTF("database is version %u, expected version %u",
968 m->mm_version, MDB_VERSION);
972 memcpy(meta, m, sizeof(*m));
977 mdbenv_init_meta(MDB_env *env, MDB_meta *meta)
984 DPRINTF("writing new meta page");
985 psize = sysconf(_SC_PAGE_SIZE);
987 meta->mm_magic = MDB_MAGIC;
988 meta->mm_version = MDB_VERSION;
989 meta->mm_psize = psize;
990 meta->mm_last_pg = 1;
991 meta->mm_flags = env->me_flags & 0xffff;
992 #if __BYTE_ORDER == __LITTLE_ENDIAN
993 /* freeDB keys are pgno_t's, must compare in int order */
994 meta->mm_flags |= MDB_REVERSEKEY;
996 meta->mm_dbs[0].md_root = P_INVALID;
997 meta->mm_dbs[1].md_root = P_INVALID;
999 p = calloc(2, psize);
1001 p->mp_flags = P_META;
1004 memcpy(m, meta, sizeof(*meta));
1006 q = (MDB_page *)((char *)p + psize);
1009 q->mp_flags = P_META;
1012 memcpy(m, meta, sizeof(*meta));
1014 rc = write(env->me_fd, p, psize * 2);
1016 return (rc == (int)psize * 2) ? MDB_SUCCESS : errno;
1020 mdbenv_write_meta(MDB_txn *txn)
1028 assert(txn != NULL);
1029 assert(txn->mt_env != NULL);
1031 DPRINTF("writing meta page for root page %lu", txn->mt_dbs[MAIN_DBI].md_root);
1035 ptr = (char *)&meta;
1036 off = offsetof(MDB_meta, mm_dbs[0].md_depth);
1037 len = sizeof(MDB_meta) - off;
1040 meta.mm_dbs[0] = txn->mt_dbs[0];
1041 meta.mm_dbs[1] = txn->mt_dbs[1];
1042 meta.mm_last_pg = txn->mt_next_pgno - 1;
1043 meta.mm_txnid = txn->mt_txnid;
1045 if (!F_ISSET(txn->mt_flags, MDB_TXN_METOGGLE))
1046 off += env->me_psize;
1049 lseek(env->me_fd, off, SEEK_SET);
1050 rc = write(env->me_fd, ptr, len);
1052 DPRINTF("write failed, disk error?");
1060 mdbenv_read_meta(MDB_env *env, int *which)
1064 assert(env != NULL);
1066 if (env->me_metas[0]->mm_txnid < env->me_metas[1]->mm_txnid)
1069 if (env->me_meta != env->me_metas[toggle])
1070 env->me_meta = env->me_metas[toggle];
1074 DPRINTF("Using meta page %d", toggle);
1080 mdbenv_create(MDB_env **env)
1084 e = calloc(1, sizeof(*e));
1085 if (!e) return ENOMEM;
1087 e->me_maxreaders = DEFAULT_READERS;
1096 mdbenv_set_mapsize(MDB_env *env, size_t size)
1100 env->me_mapsize = size;
1105 mdbenv_set_maxdbs(MDB_env *env, int dbs)
1109 env->me_maxdbs = dbs;
1114 mdbenv_set_maxreaders(MDB_env *env, int readers)
1116 env->me_maxreaders = readers;
1121 mdbenv_get_maxreaders(MDB_env *env, int *readers)
1123 if (!env || !readers)
1125 *readers = env->me_maxreaders;
1130 mdbenv_open2(MDB_env *env, unsigned int flags)
1136 env->me_flags = flags;
1138 memset(&meta, 0, sizeof(meta));
1140 if ((i = mdbenv_read_header(env, &meta)) != 0) {
1143 DPRINTF("new mdbenv");
1147 if (!env->me_mapsize) {
1148 env->me_mapsize = newenv ? DEFAULT_MAPSIZE : meta.mm_mapsize;
1152 if (meta.mm_address && (flags & MDB_FIXEDMAP))
1154 env->me_map = mmap(meta.mm_address, env->me_mapsize, PROT_READ, i,
1156 if (env->me_map == MAP_FAILED)
1160 meta.mm_mapsize = env->me_mapsize;
1161 if (flags & MDB_FIXEDMAP)
1162 meta.mm_address = env->me_map;
1163 i = mdbenv_init_meta(env, &meta);
1164 if (i != MDB_SUCCESS) {
1165 munmap(env->me_map, env->me_mapsize);
1169 env->me_psize = meta.mm_psize;
1171 p = (MDB_page *)(MDB_page *)(MDB_page *)(MDB_page *)(MDB_page *)(MDB_page *)(MDB_page *)(MDB_page *)(MDB_page *)env->me_map;
1172 env->me_metas[0] = METADATA(p);
1173 env->me_metas[1] = (MDB_meta *)((char *)env->me_metas[0] + meta.mm_psize);
1175 if ((i = mdbenv_read_meta(env, NULL)) != 0)
1178 DPRINTF("opened database version %u, pagesize %u",
1179 env->me_meta->mm_version, env->me_psize);
1180 DPRINTF("depth: %u", env->me_meta->mm_dbs[MAIN_DBI].md_depth);
1181 DPRINTF("entries: %lu", env->me_meta->mm_dbs[MAIN_DBI].md_entries);
1182 DPRINTF("branch pages: %lu", env->me_meta->mm_dbs[MAIN_DBI].md_branch_pages);
1183 DPRINTF("leaf pages: %lu", env->me_meta->mm_dbs[MAIN_DBI].md_leaf_pages);
1184 DPRINTF("overflow pages: %lu", env->me_meta->mm_dbs[MAIN_DBI].md_overflow_pages);
1185 DPRINTF("root: %lu", env->me_meta->mm_dbs[MAIN_DBI].md_root);
1191 mdbenv_reader_dest(void *ptr)
1193 MDB_reader *reader = ptr;
1195 reader->mr_txnid = 0;
1200 /* downgrade the exclusive lock on the region back to shared */
1202 mdbenv_share_locks(MDB_env *env)
1204 struct flock lock_info;
1206 env->me_txns->mt_txnid = env->me_meta->mm_txnid;
1208 memset((void *)&lock_info, 0, sizeof(lock_info));
1209 lock_info.l_type = F_RDLCK;
1210 lock_info.l_whence = SEEK_SET;
1211 lock_info.l_start = 0;
1212 lock_info.l_len = 1;
1213 fcntl(env->me_lfd, F_SETLK, &lock_info);
1217 mdbenv_setup_locks(MDB_env *env, char *lpath, int mode, int *excl)
1221 struct flock lock_info;
1225 if ((env->me_lfd = open(lpath, O_RDWR|O_CREAT, mode)) == -1) {
1229 /* Try to get exclusive lock. If we succeed, then
1230 * nobody is using the lock region and we should initialize it.
1232 memset((void *)&lock_info, 0, sizeof(lock_info));
1233 lock_info.l_type = F_WRLCK;
1234 lock_info.l_whence = SEEK_SET;
1235 lock_info.l_start = 0;
1236 lock_info.l_len = 1;
1237 rc = fcntl(env->me_lfd, F_SETLK, &lock_info);
1241 lock_info.l_type = F_RDLCK;
1242 rc = fcntl(env->me_lfd, F_SETLK, &lock_info);
1248 size = lseek(env->me_lfd, 0, SEEK_END);
1249 rsize = (env->me_maxreaders-1) * sizeof(MDB_reader) + sizeof(MDB_txninfo);
1250 if (size < rsize && *excl) {
1251 if (ftruncate(env->me_lfd, rsize) != 0) {
1257 size = rsize - sizeof(MDB_txninfo);
1258 env->me_maxreaders = size/sizeof(MDB_reader) + 1;
1260 env->me_txns = mmap(0, rsize, PROT_READ|PROT_WRITE, MAP_SHARED,
1262 if (env->me_txns == MAP_FAILED) {
1267 pthread_mutexattr_t mattr;
1269 pthread_mutexattr_init(&mattr);
1270 pthread_mutexattr_setpshared(&mattr, PTHREAD_PROCESS_SHARED);
1271 pthread_mutex_init(&env->me_txns->mt_mutex, &mattr);
1272 pthread_mutex_init(&env->me_txns->mt_wmutex, &mattr);
1273 env->me_txns->mt_version = MDB_VERSION;
1274 env->me_txns->mt_magic = MDB_MAGIC;
1275 env->me_txns->mt_txnid = 0;
1276 env->me_txns->mt_numreaders = 0;
1279 if (env->me_txns->mt_magic != MDB_MAGIC) {
1280 DPRINTF("lock region has invalid magic");
1283 if (env->me_txns->mt_version != MDB_VERSION) {
1284 DPRINTF("lock region is version %u, expected version %u",
1285 env->me_txns->mt_version, MDB_VERSION);
1288 if (errno != EACCES && errno != EAGAIN) {
1302 mdbenv_open(MDB_env *env, const char *path, unsigned int flags, mode_t mode)
1304 int oflags, rc, len, excl;
1305 char *lpath, *dpath;
1308 lpath = malloc(len + sizeof("/lock.mdb") + len + sizeof("/data.db"));
1311 dpath = lpath + len + sizeof("/lock.mdb");
1312 sprintf(lpath, "%s/lock.mdb", path);
1313 sprintf(dpath, "%s/data.mdb", path);
1315 rc = mdbenv_setup_locks(env, lpath, mode, &excl);
1319 if (F_ISSET(flags, MDB_RDONLY))
1322 oflags = O_RDWR | O_CREAT;
1324 if ((env->me_fd = open(dpath, oflags, mode)) == -1)
1327 if ((rc = mdbenv_open2(env, flags)) != MDB_SUCCESS) {
1331 env->me_path = strdup(path);
1332 DPRINTF("opened dbenv %p", (void *) env);
1333 pthread_key_create(&env->me_txkey, mdbenv_reader_dest);
1335 mdbenv_share_locks(env);
1336 env->me_dbxs = calloc(env->me_maxdbs, sizeof(MDB_dbx));
1337 env->me_dbs[0] = calloc(env->me_maxdbs, sizeof(MDB_db));
1338 env->me_dbs[1] = calloc(env->me_maxdbs, sizeof(MDB_db));
1348 mdbenv_close(MDB_env *env)
1357 munmap(env->me_map, env->me_mapsize);
1361 size_t size = (env->me_maxreaders-1) * sizeof(MDB_reader) + sizeof(MDB_txninfo);
1362 munmap(env->me_txns, size);
1368 /* Search for key within a leaf page, using binary search.
1369 * Returns the smallest entry larger or equal to the key.
1370 * If exactp is non-null, stores whether the found entry was an exact match
1371 * in *exactp (1 or 0).
1372 * If kip is non-null, stores the index of the found entry in *kip.
1373 * If no entry larger or equal to the key is found, returns NULL.
1376 mdb_search_node(MDB_txn *txn, MDB_dbi dbi, MDB_page *mp, MDB_val *key,
1377 int *exactp, unsigned int *kip)
1385 DPRINTF("searching %u keys in %s page %lu",
1387 IS_LEAF(mp) ? "leaf" : "branch",
1390 assert(NUMKEYS(mp) > 0);
1392 memset(&nodekey, 0, sizeof(nodekey));
1394 low = IS_LEAF(mp) ? 0 : 1;
1395 high = NUMKEYS(mp) - 1;
1396 while (low <= high) {
1397 i = (low + high) >> 1;
1398 node = NODEPTR(mp, i);
1400 nodekey.mv_size = node->mn_ksize;
1401 nodekey.mv_data = NODEKEY(node);
1403 if (txn->mt_dbxs[dbi].md_cmp)
1404 rc = txn->mt_dbxs[dbi].md_cmp(key, &nodekey);
1406 rc = _mdb_cmp(txn, dbi, key, &nodekey);
1409 DPRINTF("found leaf index %u [%.*s], rc = %i",
1410 i, (int)nodekey.mv_size, (char *)nodekey.mv_data, rc);
1412 DPRINTF("found branch index %u [%.*s -> %lu], rc = %i",
1413 i, (int)node->mn_ksize, (char *)NODEKEY(node),
1424 if (rc > 0) { /* Found entry is less than the key. */
1425 i++; /* Skip to get the smallest entry larger than key. */
1426 if (i >= NUMKEYS(mp))
1427 /* There is no entry larger or equal to the key. */
1431 *exactp = (rc == 0);
1432 if (kip) /* Store the key index if requested. */
1435 return NODEPTR(mp, i);
1439 cursor_pop_page(MDB_cursor *cursor)
1443 top = CURSOR_TOP(cursor);
1446 DPRINTF("popped page %lu off cursor %p", top->mp_page->mp_pgno, (void *) cursor);
1452 cursor_push_page(MDB_cursor *cursor, MDB_page *mp)
1456 DPRINTF("pushing page %lu on cursor %p", mp->mp_pgno, (void *) cursor);
1458 if ((ppage = calloc(1, sizeof(*ppage))) == NULL)
1460 ppage->mp_page = mp;
1461 CURSOR_PUSH(cursor, ppage);
1466 mdb_get_page(MDB_txn *txn, pgno_t pgno)
1471 if (!F_ISSET(txn->mt_flags, MDB_TXN_RDONLY) && !STAILQ_EMPTY(txn->mt_u.dirty_queue)) {
1473 STAILQ_FOREACH(dp, txn->mt_u.dirty_queue, h.md_next) {
1474 if (dp->p.mp_pgno == pgno) {
1482 p = (MDB_page *)(txn->mt_env->me_map + txn->mt_env->me_psize * pgno);
1488 mdb_search_page_root(MDB_txn *txn, MDB_dbi dbi, MDB_val *key,
1489 MDB_cursor *cursor, int modify, MDB_pageparent *mpp)
1491 MDB_page *mp = mpp->mp_page;
1494 if (cursor && cursor_push_page(cursor, mp) == NULL)
1497 while (IS_BRANCH(mp)) {
1501 DPRINTF("branch page %lu has %u keys", mp->mp_pgno, NUMKEYS(mp));
1502 assert(NUMKEYS(mp) > 1);
1503 DPRINTF("found index 0 to page %lu", NODEPGNO(NODEPTR(mp, 0)));
1505 if (key == NULL) /* Initialize cursor to first page. */
1507 else if (key->mv_size > MAXKEYSIZE && key->mv_data == NULL) {
1508 /* cursor to last page */
1512 node = mdb_search_node(txn, dbi, mp, key, &exact, &i);
1514 i = NUMKEYS(mp) - 1;
1522 DPRINTF("following index %u for key %.*s",
1523 i, (int)key->mv_size, (char *)key->mv_data);
1524 assert(i < NUMKEYS(mp));
1525 node = NODEPTR(mp, i);
1528 CURSOR_TOP(cursor)->mp_ki = i;
1530 mpp->mp_parent = mp;
1531 if ((mp = mdb_get_page(txn, NODEPGNO(node))) == NULL)
1536 if (cursor && cursor_push_page(cursor, mp) == NULL)
1540 MDB_dhead *dh = ((MDB_dhead *)mp)-1;
1541 if ((rc = mdb_touch(txn, mpp)) != 0)
1543 dh = ((MDB_dhead *)mpp->mp_page)-1;
1544 dh->md_parent = mpp->mp_parent;
1545 dh->md_pi = mpp->mp_pi;
1552 DPRINTF("internal error, index points to a %02X page!?",
1557 DPRINTF("found leaf page %lu for key %.*s", mp->mp_pgno,
1558 key ? (int)key->mv_size : 0, key ? (char *)key->mv_data : NULL);
1563 /* Search for the page a given key should be in.
1564 * Stores a pointer to the found page in *mpp.
1565 * If key is NULL, search for the lowest page (used by mdb_cursor_first).
1566 * If cursor is non-null, pushes parent pages on the cursor stack.
1567 * If modify is true, visited pages are updated with new page numbers.
1570 mdb_search_page(MDB_txn *txn, MDB_dbi dbi, MDB_val *key,
1571 MDB_cursor *cursor, int modify, MDB_pageparent *mpp)
1576 /* Choose which root page to start with. If a transaction is given
1577 * use the root page from the transaction, otherwise read the last
1578 * committed root page.
1580 if (F_ISSET(txn->mt_flags, MDB_TXN_ERROR)) {
1581 DPRINTF("transaction has failed, must abort");
1584 root = txn->mt_dbs[dbi].md_root;
1586 if (root == P_INVALID) { /* Tree is empty. */
1587 DPRINTF("tree is empty");
1591 if ((mpp->mp_page = mdb_get_page(txn, root)) == NULL)
1594 DPRINTF("root page has flags 0x%X", mpp->mp_page->mp_flags);
1597 /* For sub-databases, update main root first */
1598 if (dbi > MAIN_DBI && !txn->mt_dbxs[dbi].md_dirty) {
1600 rc = mdb_search_page(txn, 0, &txn->mt_dbxs[dbi].md_name,
1604 txn->mt_dbxs[dbi].md_dirty = 1;
1606 if (!F_ISSET(mpp->mp_page->mp_flags, P_DIRTY)) {
1607 mpp->mp_parent = NULL;
1609 if ((rc = mdb_touch(txn, mpp)))
1611 txn->mt_dbs[dbi].md_root = mpp->mp_page->mp_pgno;
1615 return mdb_search_page_root(txn, dbi, key, cursor, modify, mpp);
1619 mdb_read_data(MDB_txn *txn, MDB_node *leaf, MDB_val *data)
1621 MDB_page *omp; /* overflow mpage */
1624 if (!F_ISSET(leaf->mn_flags, F_BIGDATA)) {
1625 data->mv_size = leaf->mn_dsize;
1626 data->mv_data = NODEDATA(leaf);
1630 /* Read overflow data.
1632 data->mv_size = leaf->mn_dsize;
1633 memcpy(&pgno, NODEDATA(leaf), sizeof(pgno));
1634 if ((omp = mdb_get_page(txn, pgno)) == NULL) {
1635 DPRINTF("read overflow page %lu failed", pgno);
1638 data->mv_data = omp;
1644 mdb_get(MDB_txn *txn, MDB_dbi dbi,
1645 MDB_val *key, MDB_val *data)
1653 DPRINTF("===> get key [%.*s]", (int)key->mv_size, (char *)key->mv_data);
1655 if (key->mv_size == 0 || key->mv_size > MAXKEYSIZE) {
1659 if ((rc = mdb_search_page(txn, dbi, key, NULL, 0, &mpp)) != MDB_SUCCESS)
1662 leaf = mdb_search_node(txn, dbi, mpp.mp_page, key, &exact, NULL);
1664 rc = mdb_read_data(txn, leaf, data);
1673 mdb_sibling(MDB_cursor *cursor, int move_right)
1677 MDB_ppage *parent, *top;
1680 top = CURSOR_TOP(cursor);
1681 if ((parent = SLIST_NEXT(top, mp_entry)) == NULL) {
1682 return ENOENT; /* root has no siblings */
1685 DPRINTF("parent page is page %lu, index %u",
1686 parent->mp_page->mp_pgno, parent->mp_ki);
1688 cursor_pop_page(cursor);
1689 if (move_right ? (parent->mp_ki + 1 >= NUMKEYS(parent->mp_page))
1690 : (parent->mp_ki == 0)) {
1691 DPRINTF("no more keys left, moving to %s sibling",
1692 move_right ? "right" : "left");
1693 if ((rc = mdb_sibling(cursor, move_right)) != MDB_SUCCESS)
1695 parent = CURSOR_TOP(cursor);
1701 DPRINTF("just moving to %s index key %u",
1702 move_right ? "right" : "left", parent->mp_ki);
1704 assert(IS_BRANCH(parent->mp_page));
1706 indx = NODEPTR(parent->mp_page, parent->mp_ki);
1707 if ((mp = mdb_get_page(cursor->mc_txn, indx->mn_pgno)) == NULL)
1710 mp->parent = parent->mp_page;
1711 mp->parent_index = parent->mp_ki;
1714 cursor_push_page(cursor, mp);
1720 mdb_set_key(MDB_node *node, MDB_val *key)
1725 key->mv_size = node->mn_ksize;
1726 key->mv_data = NODEKEY(node);
1732 mdb_cursor_next(MDB_cursor *cursor, MDB_val *key, MDB_val *data)
1738 if (cursor->mc_eof) {
1742 assert(cursor->mc_initialized);
1744 top = CURSOR_TOP(cursor);
1747 DPRINTF("cursor_next: top page is %lu in cursor %p", mp->mp_pgno, (void *) cursor);
1749 if (top->mp_ki + 1 >= NUMKEYS(mp)) {
1750 DPRINTF("=====> move to next sibling page");
1751 if (mdb_sibling(cursor, 1) != MDB_SUCCESS) {
1755 top = CURSOR_TOP(cursor);
1757 DPRINTF("next page is %lu, key index %u", mp->mp_pgno, top->mp_ki);
1761 DPRINTF("==> cursor points to page %lu with %u keys, key index %u",
1762 mp->mp_pgno, NUMKEYS(mp), top->mp_ki);
1764 assert(IS_LEAF(mp));
1765 leaf = NODEPTR(mp, top->mp_ki);
1767 if (data && mdb_read_data(cursor->mc_txn, leaf, data) != MDB_SUCCESS)
1770 return mdb_set_key(leaf, key);
1774 mdb_cursor_prev(MDB_cursor *cursor, MDB_val *key, MDB_val *data)
1780 assert(cursor->mc_initialized);
1782 top = CURSOR_TOP(cursor);
1785 DPRINTF("cursor_prev: top page is %lu in cursor %p", mp->mp_pgno, (void *) cursor);
1787 if (top->mp_ki == 0) {
1788 DPRINTF("=====> move to prev sibling page");
1789 if (mdb_sibling(cursor, 0) != MDB_SUCCESS) {
1792 top = CURSOR_TOP(cursor);
1794 top->mp_ki = NUMKEYS(mp) - 1;
1795 DPRINTF("prev page is %lu, key index %u", mp->mp_pgno, top->mp_ki);
1801 DPRINTF("==> cursor points to page %lu with %u keys, key index %u",
1802 mp->mp_pgno, NUMKEYS(mp), top->mp_ki);
1804 assert(IS_LEAF(mp));
1805 leaf = NODEPTR(mp, top->mp_ki);
1807 if (data && mdb_read_data(cursor->mc_txn, leaf, data) != MDB_SUCCESS)
1810 return mdb_set_key(leaf, key);
1814 mdb_cursor_set(MDB_cursor *cursor, MDB_val *key, MDB_val *data,
1824 assert(key->mv_size > 0);
1826 rc = mdb_search_page(cursor->mc_txn, cursor->mc_dbi, key, cursor, 0, &mpp);
1827 if (rc != MDB_SUCCESS)
1829 assert(IS_LEAF(mpp.mp_page));
1831 top = CURSOR_TOP(cursor);
1832 leaf = mdb_search_node(cursor->mc_txn, cursor->mc_dbi, mpp.mp_page, key, exactp, &top->mp_ki);
1833 if (exactp != NULL && !*exactp) {
1834 /* MDB_SET specified and not an exact match. */
1839 DPRINTF("===> inexact leaf not found, goto sibling");
1840 if ((rc = mdb_sibling(cursor, 1)) != MDB_SUCCESS)
1841 return rc; /* no entries matched */
1842 top = CURSOR_TOP(cursor);
1844 mpp.mp_page = top->mp_page;
1845 assert(IS_LEAF(mpp.mp_page));
1846 leaf = NODEPTR(mpp.mp_page, 0);
1849 cursor->mc_initialized = 1;
1852 if (data && (rc = mdb_read_data(cursor->mc_txn, leaf, data)) != MDB_SUCCESS)
1855 rc = mdb_set_key(leaf, key);
1856 if (rc == MDB_SUCCESS) {
1857 DPRINTF("==> cursor placed on key %.*s",
1858 (int)key->mv_size, (char *)key->mv_data);
1866 mdb_cursor_first(MDB_cursor *cursor, MDB_val *key, MDB_val *data)
1872 rc = mdb_search_page(cursor->mc_txn, cursor->mc_dbi, NULL, cursor, 0, &mpp);
1873 if (rc != MDB_SUCCESS)
1875 assert(IS_LEAF(mpp.mp_page));
1877 leaf = NODEPTR(mpp.mp_page, 0);
1878 cursor->mc_initialized = 1;
1881 if (data && (rc = mdb_read_data(cursor->mc_txn, leaf, data)) != MDB_SUCCESS)
1884 return mdb_set_key(leaf, key);
1888 mdb_cursor_last(MDB_cursor *cursor, MDB_val *key, MDB_val *data)
1896 lkey.mv_size = MAXKEYSIZE+1;
1897 lkey.mv_data = NULL;
1899 rc = mdb_search_page(cursor->mc_txn, cursor->mc_dbi, &lkey, cursor, 0, &mpp);
1900 if (rc != MDB_SUCCESS)
1902 assert(IS_LEAF(mpp.mp_page));
1904 leaf = NODEPTR(mpp.mp_page, NUMKEYS(mpp.mp_page)-1);
1905 cursor->mc_initialized = 1;
1908 top = CURSOR_TOP(cursor);
1909 top->mp_ki = NUMKEYS(top->mp_page) - 1;
1911 if (data && (rc = mdb_read_data(cursor->mc_txn, leaf, data)) != MDB_SUCCESS)
1914 return mdb_set_key(leaf, key);
1918 mdb_cursor_get(MDB_cursor *cursor, MDB_val *key, MDB_val *data,
1929 while (CURSOR_TOP(cursor) != NULL)
1930 cursor_pop_page(cursor);
1931 if (key == NULL || key->mv_size == 0 || key->mv_size > MAXKEYSIZE) {
1933 } else if (op == MDB_SET)
1934 rc = mdb_cursor_set(cursor, key, data, &exact);
1936 rc = mdb_cursor_set(cursor, key, data, NULL);
1939 if (!cursor->mc_initialized)
1940 rc = mdb_cursor_first(cursor, key, data);
1942 rc = mdb_cursor_next(cursor, key, data);
1945 if (!cursor->mc_initialized || cursor->mc_eof) {
1946 while (CURSOR_TOP(cursor) != NULL)
1947 cursor_pop_page(cursor);
1948 rc = mdb_cursor_last(cursor, key, data);
1950 rc = mdb_cursor_prev(cursor, key, data);
1953 while (CURSOR_TOP(cursor) != NULL)
1954 cursor_pop_page(cursor);
1955 rc = mdb_cursor_first(cursor, key, data);
1958 while (CURSOR_TOP(cursor) != NULL)
1959 cursor_pop_page(cursor);
1960 rc = mdb_cursor_last(cursor, key, data);
1963 DPRINTF("unhandled/unimplemented cursor operation %u", op);
1971 /* Allocate a page and initialize it
1974 mdb_new_page(MDB_txn *txn, MDB_dbi dbi, uint32_t flags, int num)
1978 if ((dp = mdb_alloc_page(txn, NULL, 0, num)) == NULL)
1980 DPRINTF("allocated new mpage %lu, page size %u",
1981 dp->p.mp_pgno, txn->mt_env->me_psize);
1982 dp->p.mp_flags = flags | P_DIRTY;
1983 dp->p.mp_lower = PAGEHDRSZ;
1984 dp->p.mp_upper = txn->mt_env->me_psize;
1986 if (IS_BRANCH(&dp->p))
1987 txn->mt_dbs[dbi].md_branch_pages++;
1988 else if (IS_LEAF(&dp->p))
1989 txn->mt_dbs[dbi].md_leaf_pages++;
1990 else if (IS_OVERFLOW(&dp->p)) {
1991 txn->mt_dbs[dbi].md_overflow_pages += num;
1992 dp->p.mp_pages = num;
1999 mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data)
2003 sz = LEAFSIZE(key, data);
2004 if (data->mv_size >= env->me_psize / MDB_MINKEYS) {
2005 /* put on overflow page */
2006 sz -= data->mv_size - sizeof(pgno_t);
2009 return sz + sizeof(indx_t);
2013 mdb_branch_size(MDB_env *env, MDB_val *key)
2018 if (sz >= env->me_psize / MDB_MINKEYS) {
2019 /* put on overflow page */
2020 /* not implemented */
2021 /* sz -= key->size - sizeof(pgno_t); */
2024 return sz + sizeof(indx_t);
2028 mdb_add_node(MDB_txn *txn, MDB_dbi dbi, MDB_page *mp, indx_t indx,
2029 MDB_val *key, MDB_val *data, pgno_t pgno, uint8_t flags)
2032 size_t node_size = NODESIZE;
2035 MDB_dpage *ofp = NULL; /* overflow page */
2037 assert(mp->mp_upper >= mp->mp_lower);
2039 DPRINTF("add node [%.*s] to %s page %lu at index %i, key size %zu",
2040 key ? (int)key->mv_size : 0, key ? (char *)key->mv_data : NULL,
2041 IS_LEAF(mp) ? "leaf" : "branch",
2042 mp->mp_pgno, indx, key ? key->mv_size : 0);
2045 node_size += key->mv_size;
2049 if (F_ISSET(flags, F_BIGDATA)) {
2050 /* Data already on overflow page. */
2051 node_size += sizeof(pgno_t);
2052 } else if (data->mv_size >= txn->mt_env->me_psize / MDB_MINKEYS) {
2053 int ovpages = OVPAGES(data->mv_size, txn->mt_env->me_psize);
2054 /* Put data on overflow page. */
2055 DPRINTF("data size is %zu, put on overflow page",
2057 node_size += sizeof(pgno_t);
2058 if ((ofp = mdb_new_page(txn, dbi, P_OVERFLOW, ovpages)) == NULL)
2060 DPRINTF("allocated overflow page %lu", ofp->p.mp_pgno);
2063 node_size += data->mv_size;
2067 if (node_size + sizeof(indx_t) > SIZELEFT(mp)) {
2068 DPRINTF("not enough room in page %lu, got %u ptrs",
2069 mp->mp_pgno, NUMKEYS(mp));
2070 DPRINTF("upper - lower = %u - %u = %u", mp->mp_upper, mp->mp_lower,
2071 mp->mp_upper - mp->mp_lower);
2072 DPRINTF("node size = %zu", node_size);
2076 /* Move higher pointers up one slot. */
2077 for (i = NUMKEYS(mp); i > indx; i--)
2078 mp->mp_ptrs[i] = mp->mp_ptrs[i - 1];
2080 /* Adjust free space offsets. */
2081 ofs = mp->mp_upper - node_size;
2082 assert(ofs >= mp->mp_lower + sizeof(indx_t));
2083 mp->mp_ptrs[indx] = ofs;
2085 mp->mp_lower += sizeof(indx_t);
2087 /* Write the node data. */
2088 node = NODEPTR(mp, indx);
2089 node->mn_ksize = (key == NULL) ? 0 : key->mv_size;
2090 node->mn_flags = flags;
2092 node->mn_dsize = data->mv_size;
2094 node->mn_pgno = pgno;
2097 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
2102 if (F_ISSET(flags, F_BIGDATA))
2103 memcpy(node->mn_data + key->mv_size, data->mv_data,
2106 memcpy(node->mn_data + key->mv_size, data->mv_data,
2109 memcpy(node->mn_data + key->mv_size, &ofp->p.mp_pgno,
2111 memcpy(METADATA(&ofp->p), data->mv_data, data->mv_size);
2119 mdb_del_node(MDB_page *mp, indx_t indx)
2122 indx_t i, j, numkeys, ptr;
2126 DPRINTF("delete node %u on %s page %lu", indx,
2127 IS_LEAF(mp) ? "leaf" : "branch", mp->mp_pgno);
2128 assert(indx < NUMKEYS(mp));
2130 node = NODEPTR(mp, indx);
2131 sz = NODESIZE + node->mn_ksize;
2133 if (F_ISSET(node->mn_flags, F_BIGDATA))
2134 sz += sizeof(pgno_t);
2136 sz += NODEDSZ(node);
2139 ptr = mp->mp_ptrs[indx];
2140 numkeys = NUMKEYS(mp);
2141 for (i = j = 0; i < numkeys; i++) {
2143 mp->mp_ptrs[j] = mp->mp_ptrs[i];
2144 if (mp->mp_ptrs[i] < ptr)
2145 mp->mp_ptrs[j] += sz;
2150 base = (char *)mp + mp->mp_upper;
2151 memmove(base + sz, base, ptr - mp->mp_upper);
2153 mp->mp_lower -= sizeof(indx_t);
2158 mdb_cursor_open(MDB_txn *txn, MDB_dbi dbi, MDB_cursor **ret)
2161 size_t size = sizeof(MDB_cursor);
2163 if (txn == NULL || ret == NULL || !dbi || dbi >= txn->mt_numdbs)
2166 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT)
2167 size += sizeof(MDB_xcursor);
2169 if ((cursor = calloc(1, size)) != NULL) {
2170 SLIST_INIT(&cursor->mc_stack);
2171 cursor->mc_dbi = dbi;
2172 cursor->mc_txn = txn;
2173 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
2174 MDB_xcursor *mx = (MDB_xcursor *)(cursor + 1);
2175 cursor->mc_xcursor = mx;
2176 mx->mx_cursor.mc_txn = &mx->mx_txn;
2178 mx->mx_txn.mt_dbxs = mx->mx_dbxs;
2179 mx->mx_txn.mt_dbs = mx->mx_dbs;
2180 mx->mx_dbxs[0] = txn->mt_dbxs[0];
2181 mx->mx_dbxs[1] = txn->mt_dbxs[1];
2183 mx->mx_dbxs[2] = txn->mt_dbxs[dbi];
2184 mx->mx_txn.mt_numdbs = 4;
2186 mx->mx_txn.mt_numdbs = 3;
2199 mdb_cursor_close(MDB_cursor *cursor)
2201 if (cursor != NULL) {
2202 while (!CURSOR_EMPTY(cursor))
2203 cursor_pop_page(cursor);
2205 /* btree_close(cursor->bt); */
2211 mdb_update_key(MDB_page *mp, indx_t indx, MDB_val *key)
2213 indx_t ptr, i, numkeys;
2219 node = NODEPTR(mp, indx);
2220 ptr = mp->mp_ptrs[indx];
2221 DPRINTF("update key %u (ofs %u) [%.*s] to [%.*s] on page %lu",
2223 (int)node->mn_ksize, (char *)NODEKEY(node),
2224 (int)key->mv_size, (char *)key->mv_data,
2227 delta = key->mv_size - node->mn_ksize;
2229 if (delta > 0 && SIZELEFT(mp) < delta) {
2230 DPRINTF("OUCH! Not enough room, delta = %d", delta);
2234 numkeys = NUMKEYS(mp);
2235 for (i = 0; i < numkeys; i++) {
2236 if (mp->mp_ptrs[i] <= ptr)
2237 mp->mp_ptrs[i] -= delta;
2240 base = (char *)mp + mp->mp_upper;
2241 len = ptr - mp->mp_upper + NODESIZE;
2242 memmove(base - delta, base, len);
2243 mp->mp_upper -= delta;
2245 node = NODEPTR(mp, indx);
2246 node->mn_ksize = key->mv_size;
2249 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
2254 /* Move a node from src to dst.
2257 mdb_move_node(MDB_txn *txn, MDB_dbi dbi, MDB_pageparent *src, indx_t srcindx,
2258 MDB_pageparent *dst, indx_t dstindx)
2264 srcnode = NODEPTR(src->mp_page, srcindx);
2265 DPRINTF("moving %s node %u [%.*s] on page %lu to node %u on page %lu",
2266 IS_LEAF(src->mp_page) ? "leaf" : "branch",
2268 (int)srcnode->mn_ksize, (char *)NODEKEY(srcnode),
2269 src->mp_page->mp_pgno,
2270 dstindx, dst->mp_page->mp_pgno);
2272 /* Mark src and dst as dirty. */
2273 if ((rc = mdb_touch(txn, src)) ||
2274 (rc = mdb_touch(txn, dst)))
2277 /* Add the node to the destination page.
2279 key.mv_size = srcnode->mn_ksize;
2280 key.mv_data = NODEKEY(srcnode);
2281 data.mv_size = NODEDSZ(srcnode);
2282 data.mv_data = NODEDATA(srcnode);
2283 rc = mdb_add_node(txn, dbi, dst->mp_page, dstindx, &key, &data, NODEPGNO(srcnode),
2285 if (rc != MDB_SUCCESS)
2288 /* Delete the node from the source page.
2290 mdb_del_node(src->mp_page, srcindx);
2292 /* Update the parent separators.
2294 if (srcindx == 0 && src->mp_pi != 0) {
2295 DPRINTF("update separator for source page %lu to [%.*s]",
2296 src->mp_page->mp_pgno, (int)key.mv_size, (char *)key.mv_data);
2297 if ((rc = mdb_update_key(src->mp_parent, src->mp_pi,
2298 &key)) != MDB_SUCCESS)
2302 if (srcindx == 0 && IS_BRANCH(src->mp_page)) {
2304 nullkey.mv_size = 0;
2305 assert(mdb_update_key(src->mp_page, 0, &nullkey) == MDB_SUCCESS);
2308 if (dstindx == 0 && dst->mp_pi != 0) {
2309 DPRINTF("update separator for destination page %lu to [%.*s]",
2310 dst->mp_page->mp_pgno, (int)key.mv_size, (char *)key.mv_data);
2311 if ((rc = mdb_update_key(dst->mp_parent, dst->mp_pi,
2312 &key)) != MDB_SUCCESS)
2316 if (dstindx == 0 && IS_BRANCH(dst->mp_page)) {
2318 nullkey.mv_size = 0;
2319 assert(mdb_update_key(dst->mp_page, 0, &nullkey) == MDB_SUCCESS);
2326 mdb_merge(MDB_txn *txn, MDB_dbi dbi, MDB_pageparent *src, MDB_pageparent *dst)
2335 DPRINTF("merging page %lu and %lu", src->mp_page->mp_pgno, dst->mp_page->mp_pgno);
2337 assert(txn != NULL);
2338 assert(src->mp_parent); /* can't merge root page */
2339 assert(dst->mp_parent);
2341 /* Mark src and dst as dirty. */
2342 if ((rc = mdb_touch(txn, src)) ||
2343 (rc = mdb_touch(txn, dst)))
2346 /* Move all nodes from src to dst.
2348 for (i = 0; i < NUMKEYS(src->mp_page); i++) {
2349 srcnode = NODEPTR(src->mp_page, i);
2351 key.mv_size = srcnode->mn_ksize;
2352 key.mv_data = NODEKEY(srcnode);
2353 data.mv_size = NODEDSZ(srcnode);
2354 data.mv_data = NODEDATA(srcnode);
2355 rc = mdb_add_node(txn, dbi, dst->mp_page, NUMKEYS(dst->mp_page), &key,
2356 &data, NODEPGNO(srcnode), srcnode->mn_flags);
2357 if (rc != MDB_SUCCESS)
2361 DPRINTF("dst page %lu now has %u keys (%.1f%% filled)",
2362 dst->mp_page->mp_pgno, NUMKEYS(dst->mp_page), (float)PAGEFILL(txn->mt_env, dst->mp_page) / 10);
2364 /* Unlink the src page from parent.
2366 mdb_del_node(src->mp_parent, src->mp_pi);
2367 if (src->mp_pi == 0) {
2369 if ((rc = mdb_update_key(src->mp_parent, 0, &key)) != MDB_SUCCESS)
2373 if (IS_LEAF(src->mp_page))
2374 txn->mt_dbs[dbi].md_leaf_pages--;
2376 txn->mt_dbs[dbi].md_branch_pages--;
2378 mpp.mp_page = src->mp_parent;
2379 dh = (MDB_dhead *)src->mp_parent;
2381 mpp.mp_parent = dh->md_parent;
2382 mpp.mp_pi = dh->md_pi;
2384 return mdb_rebalance(txn, dbi, &mpp);
2387 #define FILL_THRESHOLD 250
2390 mdb_rebalance(MDB_txn *txn, MDB_dbi dbi, MDB_pageparent *mpp)
2395 indx_t si = 0, di = 0;
2397 assert(txn != NULL);
2398 assert(mpp != NULL);
2400 DPRINTF("rebalancing %s page %lu (has %u keys, %.1f%% full)",
2401 IS_LEAF(mpp->mp_page) ? "leaf" : "branch",
2402 mpp->mp_page->mp_pgno, NUMKEYS(mpp->mp_page), (float)PAGEFILL(txn->mt_env, mpp->mp_page) / 10);
2404 if (PAGEFILL(txn->mt_env, mpp->mp_page) >= FILL_THRESHOLD) {
2405 DPRINTF("no need to rebalance page %lu, above fill threshold",
2406 mpp->mp_page->mp_pgno);
2410 if (mpp->mp_parent == NULL) {
2411 if (NUMKEYS(mpp->mp_page) == 0) {
2412 DPRINTF("tree is completely empty");
2413 txn->mt_dbs[dbi].md_root = P_INVALID;
2414 txn->mt_dbs[dbi].md_depth--;
2415 txn->mt_dbs[dbi].md_leaf_pages--;
2416 } else if (IS_BRANCH(mpp->mp_page) && NUMKEYS(mpp->mp_page) == 1) {
2417 DPRINTF("collapsing root page!");
2418 txn->mt_dbs[dbi].md_root = NODEPGNO(NODEPTR(mpp->mp_page, 0));
2419 if ((root = mdb_get_page(txn, txn->mt_dbs[dbi].md_root)) == NULL)
2421 txn->mt_dbs[dbi].md_depth--;
2422 txn->mt_dbs[dbi].md_branch_pages--;
2424 DPRINTF("root page doesn't need rebalancing");
2428 /* The parent (branch page) must have at least 2 pointers,
2429 * otherwise the tree is invalid.
2431 assert(NUMKEYS(mpp->mp_parent) > 1);
2433 /* Leaf page fill factor is below the threshold.
2434 * Try to move keys from left or right neighbor, or
2435 * merge with a neighbor page.
2440 if (mpp->mp_pi == 0) {
2441 /* We're the leftmost leaf in our parent.
2443 DPRINTF("reading right neighbor");
2444 node = NODEPTR(mpp->mp_parent, mpp->mp_pi + 1);
2445 if ((npp.mp_page = mdb_get_page(txn, NODEPGNO(node))) == NULL)
2447 npp.mp_pi = mpp->mp_pi + 1;
2449 di = NUMKEYS(mpp->mp_page);
2451 /* There is at least one neighbor to the left.
2453 DPRINTF("reading left neighbor");
2454 node = NODEPTR(mpp->mp_parent, mpp->mp_pi - 1);
2455 if ((npp.mp_page = mdb_get_page(txn, NODEPGNO(node))) == NULL)
2457 npp.mp_pi = mpp->mp_pi - 1;
2458 si = NUMKEYS(npp.mp_page) - 1;
2461 npp.mp_parent = mpp->mp_parent;
2463 DPRINTF("found neighbor page %lu (%u keys, %.1f%% full)",
2464 npp.mp_page->mp_pgno, NUMKEYS(npp.mp_page), (float)PAGEFILL(txn->mt_env, npp.mp_page) / 10);
2466 /* If the neighbor page is above threshold and has at least two
2467 * keys, move one key from it.
2469 * Otherwise we should try to merge them.
2471 if (PAGEFILL(txn->mt_env, npp.mp_page) >= FILL_THRESHOLD && NUMKEYS(npp.mp_page) >= 2)
2472 return mdb_move_node(txn, dbi, &npp, si, mpp, di);
2473 else { /* FIXME: if (has_enough_room()) */
2474 if (mpp->mp_pi == 0)
2475 return mdb_merge(txn, dbi, &npp, mpp);
2477 return mdb_merge(txn, dbi, mpp, &npp);
2482 mdb_del0(MDB_txn *txn, MDB_dbi dbi, unsigned int ki, MDB_pageparent *mpp, MDB_node *leaf)
2486 /* add overflow pages to free list */
2487 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) {
2491 memcpy(&pg, NODEDATA(leaf), sizeof(pg));
2492 ovpages = OVPAGES(NODEDSZ(leaf), txn->mt_env->me_psize);
2493 for (i=0; i<ovpages; i++) {
2494 DPRINTF("freed ov page %lu", pg);
2495 mdb_idl_insert(txn->mt_free_pgs, pg);
2499 mdb_del_node(mpp->mp_page, ki);
2500 txn->mt_dbs[dbi].md_entries--;
2501 rc = mdb_rebalance(txn, dbi, mpp);
2502 if (rc != MDB_SUCCESS)
2503 txn->mt_flags |= MDB_TXN_ERROR;
2509 mdb_del(MDB_txn *txn, MDB_dbi dbi,
2510 MDB_val *key, MDB_val *data)
2517 DPRINTF("========> delete key %.*s", (int)key->mv_size, (char *)key->mv_data);
2519 assert(key != NULL);
2521 if (txn == NULL || dbi >= txn->mt_numdbs)
2524 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
2528 if (key->mv_size == 0 || key->mv_size > MAXKEYSIZE) {
2532 mpp.mp_parent = NULL;
2534 if ((rc = mdb_search_page(txn, dbi, key, NULL, 1, &mpp)) != MDB_SUCCESS)
2537 leaf = mdb_search_node(txn, dbi, mpp.mp_page, key, &exact, &ki);
2538 if (leaf == NULL || !exact) {
2542 if (data && (rc = mdb_read_data(txn, leaf, data)) != MDB_SUCCESS)
2545 return mdb_del0(txn, dbi, ki, &mpp, leaf);
2548 /* Split page <*mpp>, and insert <key,(data|newpgno)> in either left or
2549 * right sibling, at index <*newindxp> (as if unsplit). Updates *mpp and
2550 * *newindxp with the actual values after split, ie if *mpp and *newindxp
2551 * refer to a node in the new right sibling page.
2554 mdb_split(MDB_txn *txn, MDB_dbi dbi, MDB_page **mpp, unsigned int *newindxp,
2555 MDB_val *newkey, MDB_val *newdata, pgno_t newpgno)
2558 int rc = MDB_SUCCESS, ins_new = 0;
2561 unsigned int i, j, split_indx;
2563 MDB_val sepkey, rkey, rdata;
2565 MDB_dpage *mdp, *rdp, *pdp;
2568 assert(txn != NULL);
2570 dh = ((MDB_dhead *)*mpp) - 1;
2571 mdp = (MDB_dpage *)dh;
2572 newindx = *newindxp;
2574 DPRINTF("-----> splitting %s page %lu and adding [%.*s] at index %i",
2575 IS_LEAF(&mdp->p) ? "leaf" : "branch", mdp->p.mp_pgno,
2576 (int)newkey->mv_size, (char *)newkey->mv_data, *newindxp);
2578 if (mdp->h.md_parent == NULL) {
2579 if ((pdp = mdb_new_page(txn, dbi, P_BRANCH, 1)) == NULL)
2582 mdp->h.md_parent = &pdp->p;
2583 txn->mt_dbs[dbi].md_root = pdp->p.mp_pgno;
2584 DPRINTF("root split! new root = %lu", pdp->p.mp_pgno);
2585 txn->mt_dbs[dbi].md_depth++;
2587 /* Add left (implicit) pointer. */
2588 if (mdb_add_node(txn, dbi, &pdp->p, 0, NULL, NULL,
2589 mdp->p.mp_pgno, 0) != MDB_SUCCESS)
2592 DPRINTF("parent branch page is %lu", mdp->h.md_parent->mp_pgno);
2595 /* Create a right sibling. */
2596 if ((rdp = mdb_new_page(txn, dbi, mdp->p.mp_flags, 1)) == NULL)
2598 rdp->h.md_parent = mdp->h.md_parent;
2599 rdp->h.md_pi = mdp->h.md_pi + 1;
2600 DPRINTF("new right sibling: page %lu", rdp->p.mp_pgno);
2602 /* Move half of the keys to the right sibling. */
2603 if ((copy = malloc(txn->mt_env->me_psize)) == NULL)
2605 memcpy(copy, &mdp->p, txn->mt_env->me_psize);
2606 memset(&mdp->p.mp_ptrs, 0, txn->mt_env->me_psize - PAGEHDRSZ);
2607 mdp->p.mp_lower = PAGEHDRSZ;
2608 mdp->p.mp_upper = txn->mt_env->me_psize;
2610 split_indx = NUMKEYS(copy) / 2 + 1;
2612 /* First find the separating key between the split pages.
2614 memset(&sepkey, 0, sizeof(sepkey));
2615 if (newindx == split_indx) {
2616 sepkey.mv_size = newkey->mv_size;
2617 sepkey.mv_data = newkey->mv_data;
2619 node = NODEPTR(copy, split_indx);
2620 sepkey.mv_size = node->mn_ksize;
2621 sepkey.mv_data = NODEKEY(node);
2624 DPRINTF("separator is [%.*s]", (int)sepkey.mv_size, (char *)sepkey.mv_data);
2626 /* Copy separator key to the parent.
2628 if (SIZELEFT(rdp->h.md_parent) < mdb_branch_size(txn->mt_env, &sepkey)) {
2629 rc = mdb_split(txn, dbi, &rdp->h.md_parent, &rdp->h.md_pi,
2630 &sepkey, NULL, rdp->p.mp_pgno);
2632 /* Right page might now have changed parent.
2633 * Check if left page also changed parent.
2635 if (rdp->h.md_parent != mdp->h.md_parent &&
2636 mdp->h.md_pi >= NUMKEYS(mdp->h.md_parent)) {
2637 mdp->h.md_parent = rdp->h.md_parent;
2638 mdp->h.md_pi = rdp->h.md_pi - 1;
2641 rc = mdb_add_node(txn, dbi, rdp->h.md_parent, rdp->h.md_pi,
2642 &sepkey, NULL, rdp->p.mp_pgno, 0);
2644 if (rc != MDB_SUCCESS) {
2649 for (i = j = 0; i <= NUMKEYS(copy); j++) {
2650 if (i < split_indx) {
2651 /* Re-insert in left sibling. */
2654 /* Insert in right sibling. */
2655 if (i == split_indx)
2656 /* Reset insert index for right sibling. */
2657 j = (i == newindx && ins_new);
2661 if (i == newindx && !ins_new) {
2662 /* Insert the original entry that caused the split. */
2663 rkey.mv_data = newkey->mv_data;
2664 rkey.mv_size = newkey->mv_size;
2665 if (IS_LEAF(&mdp->p)) {
2666 rdata.mv_data = newdata->mv_data;
2667 rdata.mv_size = newdata->mv_size;
2674 /* Update page and index for the new key. */
2677 } else if (i == NUMKEYS(copy)) {
2680 node = NODEPTR(copy, i);
2681 rkey.mv_data = NODEKEY(node);
2682 rkey.mv_size = node->mn_ksize;
2683 if (IS_LEAF(&mdp->p)) {
2684 rdata.mv_data = NODEDATA(node);
2685 rdata.mv_size = node->mn_dsize;
2687 pgno = node->mn_pgno;
2688 flags = node->mn_flags;
2693 if (!IS_LEAF(&mdp->p) && j == 0) {
2694 /* First branch index doesn't need key data. */
2698 rc = mdb_add_node(txn, dbi, &pdp->p, j, &rkey, &rdata, pgno,flags);
2706 mdb_put(MDB_txn *txn, MDB_dbi dbi,
2707 MDB_val *key, MDB_val *data, unsigned int flags)
2709 int rc = MDB_SUCCESS, exact;
2714 assert(key != NULL);
2715 assert(data != NULL);
2720 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
2724 if (txn->mt_env->me_txn != txn) {
2728 if (key->mv_size == 0 || key->mv_size > MAXKEYSIZE) {
2732 DPRINTF("==> put key %.*s, size %zu, data size %zu",
2733 (int)key->mv_size, (char *)key->mv_data, key->mv_size, data->mv_size);
2735 mpp.mp_parent = NULL;
2737 rc = mdb_search_page(txn, dbi, key, NULL, 1, &mpp);
2738 if (rc == MDB_SUCCESS) {
2739 leaf = mdb_search_node(txn, dbi, mpp.mp_page, key, &exact, &ki);
2740 if (leaf && exact) {
2741 if (F_ISSET(flags, MDB_NOOVERWRITE)) {
2742 DPRINTF("duplicate key %.*s",
2743 (int)key->mv_size, (char *)key->mv_data);
2746 /* same size, just replace it */
2747 if (NODEDSZ(leaf) == data->mv_size) {
2748 memcpy(NODEDATA(leaf), data->mv_data, data->mv_size);
2751 mdb_del_node(mpp.mp_page, ki);
2753 if (leaf == NULL) { /* append if not found */
2754 ki = NUMKEYS(mpp.mp_page);
2755 DPRINTF("appending key at index %i", ki);
2757 } else if (rc == ENOENT) {
2759 /* new file, just write a root leaf page */
2760 DPRINTF("allocating new root leaf page");
2761 if ((dp = mdb_new_page(txn, dbi, P_LEAF, 1)) == NULL) {
2764 mpp.mp_page = &dp->p;
2765 txn->mt_dbs[dbi].md_root = mpp.mp_page->mp_pgno;
2766 txn->mt_dbs[dbi].md_depth++;
2772 assert(IS_LEAF(mpp.mp_page));
2773 DPRINTF("there are %u keys, should insert new key at index %i",
2774 NUMKEYS(mpp.mp_page), ki);
2776 if (SIZELEFT(mpp.mp_page) < mdb_leaf_size(txn->mt_env, key, data)) {
2777 rc = mdb_split(txn, dbi, &mpp.mp_page, &ki, key, data, P_INVALID);
2779 /* There is room already in this leaf page. */
2780 rc = mdb_add_node(txn, dbi, mpp.mp_page, ki, key, data, 0, 0);
2783 if (rc != MDB_SUCCESS)
2784 txn->mt_flags |= MDB_TXN_ERROR;
2786 txn->mt_dbs[dbi].md_entries++;
2793 mdbenv_get_flags(MDB_env *env, unsigned int *arg)
2798 *arg = env->me_flags;
2803 mdbenv_get_path(MDB_env *env, const char **arg)
2808 *arg = env->me_path;
2813 mdbenv_stat(MDB_env *env, MDB_stat *arg)
2815 if (env == NULL || arg == NULL)
2818 arg->ms_psize = env->me_psize;
2819 arg->ms_depth = env->me_meta->mm_dbs[MAIN_DBI].md_depth;
2820 arg->ms_branch_pages = env->me_meta->mm_dbs[MAIN_DBI].md_branch_pages;
2821 arg->ms_leaf_pages = env->me_meta->mm_dbs[MAIN_DBI].md_leaf_pages;
2822 arg->ms_overflow_pages = env->me_meta->mm_dbs[MAIN_DBI].md_overflow_pages;
2823 arg->ms_entries = env->me_meta->mm_dbs[MAIN_DBI].md_entries;
2828 int mdb_open(MDB_txn *txn, const char *name, unsigned int flags, MDB_dbi *dbi)
2841 /* Is the DB already open? */
2843 for (i=2; i<txn->mt_numdbs; i++) {
2844 if (len == txn->mt_dbxs[i].md_name.mv_size &&
2845 !strncmp(name, txn->mt_dbxs[i].md_name.mv_data, len)) {
2851 if (txn->mt_numdbs >= txn->mt_env->me_maxdbs - 1)
2854 /* Find the DB info */
2856 key.mv_data = (void *)name;
2857 rc = mdb_get(txn, MAIN_DBI, &key, &data);
2859 /* Create if requested */
2860 if (rc == ENOENT && (flags & MDB_CREATE)) {
2862 data.mv_size = sizeof(MDB_db);
2863 data.mv_data = &dummy;
2864 memset(&dummy, 0, sizeof(dummy));
2865 dummy.md_root = P_INVALID;
2866 dummy.md_flags = flags & 0xffff;
2867 rc = mdb_put(txn, 0, &key, &data, 0);
2870 /* OK, got info, add to table */
2871 if (rc == MDB_SUCCESS) {
2872 txn->mt_dbxs[txn->mt_numdbs].md_name.mv_data = strdup(name);
2873 txn->mt_dbxs[txn->mt_numdbs].md_name.mv_size = len;
2874 txn->mt_dbxs[txn->mt_numdbs].md_cmp = NULL;
2875 txn->mt_dbxs[txn->mt_numdbs].md_dcmp = NULL;
2876 txn->mt_dbxs[txn->mt_numdbs].md_rel = NULL;
2877 txn->mt_dbxs[txn->mt_numdbs].md_parent = MAIN_DBI;
2878 txn->mt_dbxs[txn->mt_numdbs].md_dirty = 0;
2879 memcpy(&txn->mt_dbs[txn->mt_numdbs], data.mv_data, sizeof(MDB_db));
2880 *dbi = txn->mt_numdbs;
2887 int mdb_stat(MDB_txn *txn, MDB_dbi dbi, MDB_stat *arg)
2889 if (txn == NULL || arg == NULL)
2892 arg->ms_psize = txn->mt_env->me_psize;
2893 arg->ms_depth = txn->mt_dbs[dbi].md_depth;
2894 arg->ms_branch_pages = txn->mt_dbs[dbi].md_branch_pages;
2895 arg->ms_leaf_pages = txn->mt_dbs[dbi].md_leaf_pages;
2896 arg->ms_overflow_pages = txn->mt_dbs[dbi].md_overflow_pages;
2897 arg->ms_entries = txn->mt_dbs[dbi].md_entries;
2902 void mdb_close(MDB_txn *txn, MDB_dbi dbi)
2905 if (dbi <= MAIN_DBI || dbi >= txn->mt_numdbs)
2907 ptr = txn->mt_dbxs[dbi].md_name.mv_data;
2908 txn->mt_dbxs[dbi].md_name.mv_data = NULL;
2909 txn->mt_dbxs[dbi].md_name.mv_size = 0;