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 */
319 pgno_t me_free_pgs[MDB_IDL_UM_SIZE];
322 #define NODESIZE offsetof(MDB_node, mn_data)
324 #define INDXSIZE(k) (NODESIZE + ((k) == NULL ? 0 : (k)->mv_size))
325 #define LEAFSIZE(k, d) (NODESIZE + (k)->mv_size + (d)->mv_size)
326 #define NODEPTR(p, i) ((MDB_node *)((char *)(p) + (p)->mp_ptrs[i]))
327 #define NODEKEY(node) (void *)((node)->mn_data)
328 #define NODEDATA(node) (void *)((char *)(node)->mn_data + (node)->mn_ksize)
329 #define NODEPGNO(node) ((node)->mn_pgno)
330 #define NODEDSZ(node) ((node)->mn_dsize)
332 #define MDB_COMMIT_PAGES 64 /* max number of pages to write in one commit */
333 #define MDB_MAXCACHE_DEF 1024 /* max number of pages to keep in cache */
335 static int mdb_search_page_root(MDB_txn *txn,
336 MDB_dbi dbi, MDB_val *key,
337 MDB_cursor *cursor, int modify,
338 MDB_pageparent *mpp);
339 static int mdb_search_page(MDB_txn *txn,
340 MDB_dbi dbi, MDB_val *key,
341 MDB_cursor *cursor, int modify,
342 MDB_pageparent *mpp);
344 static int mdbenv_read_header(MDB_env *env, MDB_meta *meta);
345 static int mdbenv_read_meta(MDB_env *env, int *which);
346 static int mdbenv_write_meta(MDB_txn *txn);
347 static MDB_page *mdb_get_page(MDB_txn *txn, pgno_t pgno);
349 static MDB_node *mdb_search_node(MDB_txn *txn, MDB_dbi dbi, MDB_page *mp,
350 MDB_val *key, int *exactp, unsigned int *kip);
351 static int mdb_add_node(MDB_txn *txn, MDB_dbi dbi, MDB_page *mp,
352 indx_t indx, MDB_val *key, MDB_val *data,
353 pgno_t pgno, uint8_t flags);
354 static void mdb_del_node(MDB_page *mp, indx_t indx);
355 static int mdb_del0(MDB_txn *txn, MDB_dbi dbi, unsigned int ki,
356 MDB_pageparent *mpp, MDB_node *leaf);
357 static int mdb_read_data(MDB_txn *txn, MDB_node *leaf, MDB_val *data);
359 static int mdb_rebalance(MDB_txn *txn, MDB_dbi dbi, MDB_pageparent *mp);
360 static int mdb_update_key(MDB_page *mp, indx_t indx, MDB_val *key);
361 static int mdb_move_node(MDB_txn *txn, MDB_dbi dbi,
362 MDB_pageparent *src, indx_t srcindx,
363 MDB_pageparent *dst, indx_t dstindx);
364 static int mdb_merge(MDB_txn *txn, MDB_dbi dbi, MDB_pageparent *src,
365 MDB_pageparent *dst);
366 static int mdb_split(MDB_txn *txn, MDB_dbi dbi, MDB_page **mpp,
367 unsigned int *newindxp, MDB_val *newkey,
368 MDB_val *newdata, pgno_t newpgno);
369 static MDB_dpage *mdb_new_page(MDB_txn *txn, MDB_dbi dbi, uint32_t flags, int num);
371 static void cursor_pop_page(MDB_cursor *cursor);
372 static MDB_ppage *cursor_push_page(MDB_cursor *cursor,
375 static int mdb_set_key(MDB_node *node, MDB_val *key);
376 static int mdb_sibling(MDB_cursor *cursor, int move_right);
377 static int mdb_cursor_next(MDB_cursor *cursor,
378 MDB_val *key, MDB_val *data);
379 static int mdb_cursor_prev(MDB_cursor *cursor,
380 MDB_val *key, MDB_val *data);
381 static int mdb_cursor_set(MDB_cursor *cursor,
382 MDB_val *key, MDB_val *data, int *exactp);
383 static int mdb_cursor_first(MDB_cursor *cursor,
384 MDB_val *key, MDB_val *data);
385 static int mdb_cursor_last(MDB_cursor *cursor,
386 MDB_val *key, MDB_val *data);
388 static size_t mdb_leaf_size(MDB_env *env, MDB_val *key,
390 static size_t mdb_branch_size(MDB_env *env, MDB_val *key);
392 static int memncmp(const void *s1, size_t n1,
393 const void *s2, size_t n2);
394 static int memnrcmp(const void *s1, size_t n1,
395 const void *s2, size_t n2);
398 memncmp(const void *s1, size_t n1, const void *s2, size_t n2)
400 int diff, len_diff = -1;
403 len_diff = (n1 > n2);
406 diff = memcmp(s1, s2, n1);
407 return diff ? diff : len_diff;
411 memnrcmp(const void *s1, size_t n1, const void *s2, size_t n2)
413 const unsigned char *p1, *p2, *p1_lim;
420 p1 = (const unsigned char *)s1 + n1 - 1;
421 p2 = (const unsigned char *)s2 + n2 - 1;
423 for (p1_lim = (n1 <= n2 ? s1 : s2); *p1 == *p2; p1--, p2--) {
425 return (p1 != s1) ? (p1 != p2) : (p2 != s2) ? -1 : 0;
431 mdb_cmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
433 return txn->mt_dbxs[dbi].md_cmp(a, b);
437 _mdb_cmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *key1, const MDB_val *key2)
439 if (F_ISSET(txn->mt_dbs[dbi].md_flags, MDB_REVERSEKEY))
440 return memnrcmp(key1->mv_data, key1->mv_size, key2->mv_data, key2->mv_size);
442 return memncmp((char *)key1->mv_data, key1->mv_size, key2->mv_data, key2->mv_size);
445 /* Allocate new page(s) for writing */
447 mdb_alloc_page(MDB_txn *txn, MDB_page *parent, unsigned int parent_idx, int num)
450 pgno_t pgno = P_INVALID;
451 ULONG oldest = txn->mt_txnid - 2;
453 if (!txn->mt_env->me_pghead && txn->mt_dbs[FREE_DBI].md_root != P_INVALID) {
454 /* See if there's anything in the free DB */
459 mpp.mp_parent = NULL;
461 mdb_search_page(txn, FREE_DBI, NULL, NULL, 0, &mpp);
462 leaf = NODEPTR(mpp.mp_page, 0);
463 kptr = (ULONG *)NODEKEY(leaf);
465 /* It's potentially usable, unless there are still
466 * older readers outstanding. Grab it.
468 if (oldest > *kptr) {
473 mdb_read_data(txn, leaf, &data);
474 idl = (ULONG *)data.mv_data;
475 mop = malloc(sizeof(MDB_oldpages) + MDB_IDL_SIZEOF(idl) - sizeof(pgno_t));
476 mop->mo_next = txn->mt_env->me_pghead;
477 mop->mo_txnid = *kptr;
478 txn->mt_env->me_pghead = mop;
479 memcpy(mop->mo_pages, idl, MDB_IDL_SIZEOF(idl));
484 DPRINTF("IDL read txn %lu root %lu num %lu",
485 mop->mo_txnid, txn->mt_dbs[FREE_DBI].md_root, idl[0]);
486 for (i=0; i<idl[0]; i++) {
487 DPRINTF("IDL %lu", idl[i+1]);
491 /* drop this IDL from the DB */
492 mpp.mp_parent = NULL;
494 mdb_search_page(txn, FREE_DBI, NULL, NULL, 1, &mpp);
495 leaf = NODEPTR(mpp.mp_page, 0);
496 mdb_del0(txn, FREE_DBI, 0, &mpp, leaf);
499 if (txn->mt_env->me_pghead) {
501 for (i=0; i<txn->mt_env->me_txns->mt_numreaders; i++) {
502 ULONG mr = txn->mt_env->me_txns->mt_readers[i].mr_txnid;
505 oldest = txn->mt_env->me_txns->mt_readers[i].mr_txnid;
507 if (oldest > txn->mt_env->me_pghead->mo_txnid) {
508 MDB_oldpages *mop = txn->mt_env->me_pghead;
509 txn->mt_oldest = oldest;
511 /* FIXME: For now, always use fresh pages. We
512 * really ought to search the free list for a
517 /* peel pages off tail, so we only have to truncate the list */
518 pgno = MDB_IDL_LAST(mop->mo_pages);
519 if (MDB_IDL_IS_RANGE(mop->mo_pages)) {
521 if (mop->mo_pages[2] > mop->mo_pages[1])
522 mop->mo_pages[0] = 0;
526 if (MDB_IDL_IS_ZERO(mop->mo_pages)) {
527 txn->mt_env->me_pghead = mop->mo_next;
534 if ((dp = malloc(txn->mt_env->me_psize * num + sizeof(MDB_dhead))) == NULL)
537 dp->h.md_parent = parent;
538 dp->h.md_pi = parent_idx;
539 STAILQ_INSERT_TAIL(txn->mt_u.dirty_queue, dp, h.md_next);
540 if (pgno == P_INVALID) {
541 dp->p.mp_pgno = txn->mt_next_pgno;
542 txn->mt_next_pgno += num;
544 dp->p.mp_pgno = pgno;
550 /* Touch a page: make it dirty and re-insert into tree with updated pgno.
553 mdb_touch(MDB_txn *txn, MDB_pageparent *pp)
555 MDB_page *mp = pp->mp_page;
560 if (!F_ISSET(mp->mp_flags, P_DIRTY)) {
562 if ((dp = mdb_alloc_page(txn, pp->mp_parent, pp->mp_pi, 1)) == NULL)
564 DPRINTF("touched page %lu -> %lu", mp->mp_pgno, dp->p.mp_pgno);
565 mdb_idl_insert(txn->mt_free_pgs, mp->mp_pgno);
566 pgno = dp->p.mp_pgno;
567 memcpy(&dp->p, mp, txn->mt_env->me_psize);
570 mp->mp_flags |= P_DIRTY;
572 /* Update the page number to new touched page. */
573 if (pp->mp_parent != NULL)
574 NODEPGNO(NODEPTR(pp->mp_parent, pp->mp_pi)) = mp->mp_pgno;
581 mdbenv_sync(MDB_env *env)
584 if (!F_ISSET(env->me_flags, MDB_NOSYNC)) {
585 if (fsync(env->me_fd))
592 mdb_txn_begin(MDB_env *env, int rdonly, MDB_txn **ret)
597 if ((txn = calloc(1, sizeof(*txn))) == NULL) {
598 DPRINTF("calloc: %s", strerror(errno));
603 txn->mt_flags |= MDB_TXN_RDONLY;
605 txn->mt_u.dirty_queue = calloc(1, sizeof(*txn->mt_u.dirty_queue));
606 if (txn->mt_u.dirty_queue == NULL) {
610 STAILQ_INIT(txn->mt_u.dirty_queue);
612 pthread_mutex_lock(&env->me_txns->mt_wmutex);
613 env->me_txns->mt_txnid++;
614 txn->mt_free_pgs = env->me_free_pgs;
615 txn->mt_free_pgs[0] = 0;
618 txn->mt_txnid = env->me_txns->mt_txnid;
620 MDB_reader *r = pthread_getspecific(env->me_txkey);
623 pthread_mutex_lock(&env->me_txns->mt_mutex);
624 for (i=0; i<env->me_maxreaders; i++) {
625 if (env->me_txns->mt_readers[i].mr_pid == 0) {
626 env->me_txns->mt_readers[i].mr_pid = getpid();
627 env->me_txns->mt_readers[i].mr_tid = pthread_self();
628 r = &env->me_txns->mt_readers[i];
629 pthread_setspecific(env->me_txkey, r);
630 if (i >= env->me_txns->mt_numreaders)
631 env->me_txns->mt_numreaders = i+1;
635 pthread_mutex_unlock(&env->me_txns->mt_mutex);
636 if (i == env->me_maxreaders) {
640 r->mr_txnid = txn->mt_txnid;
641 txn->mt_u.reader = r;
648 if ((rc = mdbenv_read_meta(env, &toggle)) != MDB_SUCCESS) {
653 /* Copy the DB arrays */
654 txn->mt_numdbs = env->me_numdbs;
655 txn->mt_dbxs = env->me_dbxs; /* mostly static anyway */
656 txn->mt_dbs = malloc(env->me_maxdbs * sizeof(MDB_db));
657 memcpy(txn->mt_dbs, env->me_meta->mm_dbs, 2 * sizeof(MDB_db));
658 if (txn->mt_numdbs > 2)
659 memcpy(txn->mt_dbs+2, env->me_dbs[env->me_db_toggle]+2,
660 (txn->mt_numdbs - 2) * sizeof(MDB_db));
664 txn->mt_flags |= MDB_TXN_METOGGLE;
665 txn->mt_next_pgno = env->me_meta->mm_last_pg+1;
668 DPRINTF("begin transaction %lu on mdbenv %p, root page %lu",
669 txn->mt_txnid, (void *) env, txn->mt_dbs[MAIN_DBI].md_root);
676 mdb_txn_abort(MDB_txn *txn)
685 DPRINTF("abort transaction %lu on mdbenv %p, root page %lu",
686 txn->mt_txnid, (void *) env, txn->mt_dbs[MAIN_DBI].md_root);
690 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
691 txn->mt_u.reader->mr_txnid = 0;
696 /* Discard all dirty pages. */
697 while (!STAILQ_EMPTY(txn->mt_u.dirty_queue)) {
698 dp = STAILQ_FIRST(txn->mt_u.dirty_queue);
699 STAILQ_REMOVE_HEAD(txn->mt_u.dirty_queue, h.md_next);
702 free(txn->mt_u.dirty_queue);
704 while ((mop = txn->mt_env->me_pghead)) {
705 txn->mt_env->me_pghead = mop->mo_next;
710 env->me_txns->mt_txnid--;
711 for (i=2; i<env->me_numdbs; i++)
712 env->me_dbxs[i].md_dirty = 0;
713 pthread_mutex_unlock(&env->me_txns->mt_wmutex);
720 mdb_txn_commit(MDB_txn *txn)
729 struct iovec iov[MDB_COMMIT_PAGES];
732 assert(txn->mt_env != NULL);
736 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
737 DPRINTF("attempt to commit read-only transaction");
742 if (txn != env->me_txn) {
743 DPRINTF("attempt to commit unknown transaction");
748 if (F_ISSET(txn->mt_flags, MDB_TXN_ERROR)) {
749 DPRINTF("error flag is set, can't commit");
754 if (STAILQ_EMPTY(txn->mt_u.dirty_queue))
757 DPRINTF("committing transaction %lu on mdbenv %p, root page %lu",
758 txn->mt_txnid, (void *) env, txn->mt_dbs[MAIN_DBI].md_root);
760 /* should only be one record now */
761 if (env->me_pghead) {
765 mop = env->me_pghead;
766 key.mv_size = sizeof(pgno_t);
767 key.mv_data = (char *)&mop->mo_txnid;
768 data.mv_size = MDB_IDL_SIZEOF(mop->mo_pages);
769 data.mv_data = mop->mo_pages;
770 mdb_put(txn, FREE_DBI, &key, &data, 0);
771 free(env->me_pghead);
772 env->me_pghead = NULL;
774 /* save to free list */
775 if (!MDB_IDL_IS_ZERO(txn->mt_free_pgs)) {
779 /* make sure last page of freeDB is touched and on freelist */
780 key.mv_size = MAXKEYSIZE+1;
782 mpp.mp_parent = NULL;
784 mdb_search_page(txn, FREE_DBI, &key, NULL, 1, &mpp);
789 ULONG *idl = txn->mt_free_pgs;
790 DPRINTF("IDL write txn %lu root %lu num %lu",
791 txn->mt_txnid, txn->mt_dbs[FREE_DBI].md_root, idl[0]);
792 for (i=0; i<idl[0]; i++) {
793 DPRINTF("IDL %lu", idl[i+1]);
797 /* write to last page of freeDB */
798 key.mv_size = sizeof(pgno_t);
799 key.mv_data = (char *)&txn->mt_txnid;
800 data.mv_size = MDB_IDL_SIZEOF(txn->mt_free_pgs);
801 data.mv_data = txn->mt_free_pgs;
802 mdb_put(txn, FREE_DBI, &key, &data, 0);
805 /* Update DB root pointers. Their pages have already been
806 * touched so this is all in-place and cannot fail.
810 data.mv_size = sizeof(MDB_db);
812 for (i = 2; i < txn->mt_numdbs; i++) {
813 if (txn->mt_dbxs[i].md_dirty) {
814 data.mv_data = &txn->mt_dbs[i];
815 mdb_put(txn, i, &txn->mt_dbxs[i].md_name, &data, 0);
820 /* Commit up to MDB_COMMIT_PAGES dirty pages to disk until done.
827 STAILQ_FOREACH(dp, txn->mt_u.dirty_queue, h.md_next) {
828 if (dp->p.mp_pgno != next) {
830 DPRINTF("committing %u dirty pages", n);
831 rc = writev(env->me_fd, iov, n);
835 DPRINTF("short write, filesystem full?");
837 DPRINTF("writev: %s", strerror(errno));
844 lseek(env->me_fd, dp->p.mp_pgno * env->me_psize, SEEK_SET);
845 next = dp->p.mp_pgno;
847 DPRINTF("committing page %lu", dp->p.mp_pgno);
848 iov[n].iov_len = env->me_psize * dp->h.md_num;
849 iov[n].iov_base = &dp->p;
850 size += iov[n].iov_len;
851 next = dp->p.mp_pgno + dp->h.md_num;
852 /* clear dirty flag */
853 dp->p.mp_flags &= ~P_DIRTY;
854 if (++n >= MDB_COMMIT_PAGES) {
863 DPRINTF("committing %u dirty pages", n);
864 rc = writev(env->me_fd, iov, n);
868 DPRINTF("short write, filesystem full?");
870 DPRINTF("writev: %s", strerror(errno));
877 /* Drop the dirty pages.
879 while (!STAILQ_EMPTY(txn->mt_u.dirty_queue)) {
880 dp = STAILQ_FIRST(txn->mt_u.dirty_queue);
881 STAILQ_REMOVE_HEAD(txn->mt_u.dirty_queue, h.md_next);
885 if ((n = mdbenv_sync(env)) != 0 ||
886 (n = mdbenv_write_meta(txn)) != MDB_SUCCESS ||
887 (n = mdbenv_sync(env)) != 0) {
893 /* update the DB tables */
895 int toggle = !env->me_db_toggle;
897 for (i = 2; i < env->me_numdbs; i++) {
898 if (txn->mt_dbxs[i].md_dirty) {
899 env->me_dbs[toggle][i] = txn->mt_dbs[i];
900 txn->mt_dbxs[i].md_dirty = 0;
903 for (i = env->me_numdbs; i < txn->mt_numdbs; i++) {
904 txn->mt_dbxs[i].md_dirty = 0;
905 env->me_dbxs[i] = txn->mt_dbxs[i];
906 env->me_dbs[toggle][i] = txn->mt_dbs[i];
908 env->me_db_toggle = toggle;
909 env->me_numdbs = txn->mt_numdbs;
914 pthread_mutex_unlock(&env->me_txns->mt_wmutex);
915 free(txn->mt_u.dirty_queue);
926 mdbenv_read_header(MDB_env *env, MDB_meta *meta)
935 /* We don't know the page size yet, so use a minimum value.
938 if ((rc = pread(env->me_fd, page, PAGESIZE, 0)) == 0) {
940 } else if (rc != PAGESIZE) {
943 DPRINTF("read: %s", strerror(errno));
947 p = (MDB_page *)page;
949 if (!F_ISSET(p->mp_flags, P_META)) {
950 DPRINTF("page %lu not a meta page", p->mp_pgno);
955 if (m->mm_magic != MDB_MAGIC) {
956 DPRINTF("meta has invalid magic");
960 if (m->mm_version != MDB_VERSION) {
961 DPRINTF("database is version %u, expected version %u",
962 m->mm_version, MDB_VERSION);
966 memcpy(meta, m, sizeof(*m));
971 mdbenv_init_meta(MDB_env *env, MDB_meta *meta)
978 DPRINTF("writing new meta page");
979 psize = sysconf(_SC_PAGE_SIZE);
981 meta->mm_magic = MDB_MAGIC;
982 meta->mm_version = MDB_VERSION;
983 meta->mm_psize = psize;
984 meta->mm_last_pg = 1;
985 meta->mm_flags = env->me_flags & 0xffff;
986 #if __BYTE_ORDER == __LITTLE_ENDIAN
987 /* freeDB keys are pgno_t's, must compare in int order */
988 meta->mm_flags |= MDB_REVERSEKEY;
990 meta->mm_dbs[0].md_root = P_INVALID;
991 meta->mm_dbs[1].md_root = P_INVALID;
993 p = calloc(2, psize);
995 p->mp_flags = P_META;
998 memcpy(m, meta, sizeof(*meta));
1000 q = (MDB_page *)((char *)p + psize);
1003 q->mp_flags = P_META;
1006 memcpy(m, meta, sizeof(*meta));
1008 rc = write(env->me_fd, p, psize * 2);
1010 return (rc == (int)psize * 2) ? MDB_SUCCESS : errno;
1014 mdbenv_write_meta(MDB_txn *txn)
1022 assert(txn != NULL);
1023 assert(txn->mt_env != NULL);
1025 DPRINTF("writing meta page for root page %lu", txn->mt_dbs[MAIN_DBI].md_root);
1029 ptr = (char *)&meta;
1030 off = offsetof(MDB_meta, mm_dbs[0].md_depth);
1031 len = sizeof(MDB_meta) - off;
1034 meta.mm_dbs[0] = txn->mt_dbs[0];
1035 meta.mm_dbs[1] = txn->mt_dbs[1];
1036 meta.mm_last_pg = txn->mt_next_pgno - 1;
1037 meta.mm_txnid = txn->mt_txnid;
1039 if (!F_ISSET(txn->mt_flags, MDB_TXN_METOGGLE))
1040 off += env->me_psize;
1043 lseek(env->me_fd, off, SEEK_SET);
1044 rc = write(env->me_fd, ptr, len);
1046 DPRINTF("write failed, disk error?");
1054 mdbenv_read_meta(MDB_env *env, int *which)
1058 assert(env != NULL);
1060 if (env->me_metas[0]->mm_txnid < env->me_metas[1]->mm_txnid)
1063 if (env->me_meta != env->me_metas[toggle])
1064 env->me_meta = env->me_metas[toggle];
1068 DPRINTF("Using meta page %d", toggle);
1074 mdbenv_create(MDB_env **env)
1078 e = calloc(1, sizeof(*e));
1079 if (!e) return ENOMEM;
1081 e->me_maxreaders = DEFAULT_READERS;
1090 mdbenv_set_mapsize(MDB_env *env, size_t size)
1094 env->me_mapsize = size;
1099 mdbenv_set_maxdbs(MDB_env *env, int dbs)
1103 env->me_maxdbs = dbs;
1108 mdbenv_set_maxreaders(MDB_env *env, int readers)
1110 env->me_maxreaders = readers;
1115 mdbenv_get_maxreaders(MDB_env *env, int *readers)
1117 if (!env || !readers)
1119 *readers = env->me_maxreaders;
1124 mdbenv_open2(MDB_env *env, unsigned int flags)
1130 env->me_flags = flags;
1132 memset(&meta, 0, sizeof(meta));
1134 if ((i = mdbenv_read_header(env, &meta)) != 0) {
1137 DPRINTF("new mdbenv");
1141 if (!env->me_mapsize) {
1142 env->me_mapsize = newenv ? DEFAULT_MAPSIZE : meta.mm_mapsize;
1146 if (meta.mm_address && (flags & MDB_FIXEDMAP))
1148 env->me_map = mmap(meta.mm_address, env->me_mapsize, PROT_READ, i,
1150 if (env->me_map == MAP_FAILED)
1154 meta.mm_mapsize = env->me_mapsize;
1155 if (flags & MDB_FIXEDMAP)
1156 meta.mm_address = env->me_map;
1157 i = mdbenv_init_meta(env, &meta);
1158 if (i != MDB_SUCCESS) {
1159 munmap(env->me_map, env->me_mapsize);
1163 env->me_psize = meta.mm_psize;
1165 p = (MDB_page *)(MDB_page *)(MDB_page *)(MDB_page *)(MDB_page *)(MDB_page *)(MDB_page *)(MDB_page *)(MDB_page *)env->me_map;
1166 env->me_metas[0] = METADATA(p);
1167 env->me_metas[1] = (MDB_meta *)((char *)env->me_metas[0] + meta.mm_psize);
1169 if ((i = mdbenv_read_meta(env, NULL)) != 0)
1172 DPRINTF("opened database version %u, pagesize %u",
1173 env->me_meta->mm_version, env->me_psize);
1174 DPRINTF("depth: %u", env->me_meta->mm_dbs[MAIN_DBI].md_depth);
1175 DPRINTF("entries: %lu", env->me_meta->mm_dbs[MAIN_DBI].md_entries);
1176 DPRINTF("branch pages: %lu", env->me_meta->mm_dbs[MAIN_DBI].md_branch_pages);
1177 DPRINTF("leaf pages: %lu", env->me_meta->mm_dbs[MAIN_DBI].md_leaf_pages);
1178 DPRINTF("overflow pages: %lu", env->me_meta->mm_dbs[MAIN_DBI].md_overflow_pages);
1179 DPRINTF("root: %lu", env->me_meta->mm_dbs[MAIN_DBI].md_root);
1185 mdbenv_reader_dest(void *ptr)
1187 MDB_reader *reader = ptr;
1189 reader->mr_txnid = 0;
1194 /* downgrade the exclusive lock on the region back to shared */
1196 mdbenv_share_locks(MDB_env *env)
1198 struct flock lock_info;
1200 env->me_txns->mt_txnid = env->me_meta->mm_txnid;
1202 memset((void *)&lock_info, 0, sizeof(lock_info));
1203 lock_info.l_type = F_RDLCK;
1204 lock_info.l_whence = SEEK_SET;
1205 lock_info.l_start = 0;
1206 lock_info.l_len = 1;
1207 fcntl(env->me_lfd, F_SETLK, &lock_info);
1211 mdbenv_setup_locks(MDB_env *env, char *lpath, int mode, int *excl)
1215 struct flock lock_info;
1219 if ((env->me_lfd = open(lpath, O_RDWR|O_CREAT, mode)) == -1) {
1223 /* Try to get exclusive lock. If we succeed, then
1224 * nobody is using the lock region and we should initialize it.
1226 memset((void *)&lock_info, 0, sizeof(lock_info));
1227 lock_info.l_type = F_WRLCK;
1228 lock_info.l_whence = SEEK_SET;
1229 lock_info.l_start = 0;
1230 lock_info.l_len = 1;
1231 rc = fcntl(env->me_lfd, F_SETLK, &lock_info);
1235 lock_info.l_type = F_RDLCK;
1236 rc = fcntl(env->me_lfd, F_SETLK, &lock_info);
1242 size = lseek(env->me_lfd, 0, SEEK_END);
1243 rsize = (env->me_maxreaders-1) * sizeof(MDB_reader) + sizeof(MDB_txninfo);
1244 if (size < rsize && *excl) {
1245 if (ftruncate(env->me_lfd, rsize) != 0) {
1251 size = rsize - sizeof(MDB_txninfo);
1252 env->me_maxreaders = size/sizeof(MDB_reader) + 1;
1254 env->me_txns = mmap(0, rsize, PROT_READ|PROT_WRITE, MAP_SHARED,
1256 if (env->me_txns == MAP_FAILED) {
1261 pthread_mutexattr_t mattr;
1263 pthread_mutexattr_init(&mattr);
1264 pthread_mutexattr_setpshared(&mattr, PTHREAD_PROCESS_SHARED);
1265 pthread_mutex_init(&env->me_txns->mt_mutex, &mattr);
1266 pthread_mutex_init(&env->me_txns->mt_wmutex, &mattr);
1267 env->me_txns->mt_version = MDB_VERSION;
1268 env->me_txns->mt_magic = MDB_MAGIC;
1269 env->me_txns->mt_txnid = 0;
1270 env->me_txns->mt_numreaders = 0;
1273 if (env->me_txns->mt_magic != MDB_MAGIC) {
1274 DPRINTF("lock region has invalid magic");
1277 if (env->me_txns->mt_version != MDB_VERSION) {
1278 DPRINTF("lock region is version %u, expected version %u",
1279 env->me_txns->mt_version, MDB_VERSION);
1282 if (errno != EACCES && errno != EAGAIN) {
1296 mdbenv_open(MDB_env *env, const char *path, unsigned int flags, mode_t mode)
1298 int oflags, rc, len, excl;
1299 char *lpath, *dpath;
1302 lpath = malloc(len + sizeof("/lock.mdb") + len + sizeof("/data.db"));
1305 dpath = lpath + len + sizeof("/lock.mdb");
1306 sprintf(lpath, "%s/lock.mdb", path);
1307 sprintf(dpath, "%s/data.mdb", path);
1309 rc = mdbenv_setup_locks(env, lpath, mode, &excl);
1313 if (F_ISSET(flags, MDB_RDONLY))
1316 oflags = O_RDWR | O_CREAT;
1318 if ((env->me_fd = open(dpath, oflags, mode)) == -1)
1321 if ((rc = mdbenv_open2(env, flags)) != MDB_SUCCESS) {
1325 env->me_path = strdup(path);
1326 DPRINTF("opened dbenv %p", (void *) env);
1327 pthread_key_create(&env->me_txkey, mdbenv_reader_dest);
1329 mdbenv_share_locks(env);
1330 env->me_dbxs = calloc(env->me_maxdbs, sizeof(MDB_dbx));
1331 env->me_dbs[0] = calloc(env->me_maxdbs, sizeof(MDB_db));
1332 env->me_dbs[1] = calloc(env->me_maxdbs, sizeof(MDB_db));
1342 mdbenv_close(MDB_env *env)
1351 munmap(env->me_map, env->me_mapsize);
1355 size_t size = (env->me_maxreaders-1) * sizeof(MDB_reader) + sizeof(MDB_txninfo);
1356 munmap(env->me_txns, size);
1362 /* Search for key within a leaf page, using binary search.
1363 * Returns the smallest entry larger or equal to the key.
1364 * If exactp is non-null, stores whether the found entry was an exact match
1365 * in *exactp (1 or 0).
1366 * If kip is non-null, stores the index of the found entry in *kip.
1367 * If no entry larger or equal to the key is found, returns NULL.
1370 mdb_search_node(MDB_txn *txn, MDB_dbi dbi, MDB_page *mp, MDB_val *key,
1371 int *exactp, unsigned int *kip)
1379 DPRINTF("searching %u keys in %s page %lu",
1381 IS_LEAF(mp) ? "leaf" : "branch",
1384 assert(NUMKEYS(mp) > 0);
1386 memset(&nodekey, 0, sizeof(nodekey));
1388 low = IS_LEAF(mp) ? 0 : 1;
1389 high = NUMKEYS(mp) - 1;
1390 while (low <= high) {
1391 i = (low + high) >> 1;
1392 node = NODEPTR(mp, i);
1394 nodekey.mv_size = node->mn_ksize;
1395 nodekey.mv_data = NODEKEY(node);
1397 if (txn->mt_dbxs[dbi].md_cmp)
1398 rc = txn->mt_dbxs[dbi].md_cmp(key, &nodekey);
1400 rc = _mdb_cmp(txn, dbi, key, &nodekey);
1403 DPRINTF("found leaf index %u [%.*s], rc = %i",
1404 i, (int)nodekey.mv_size, (char *)nodekey.mv_data, rc);
1406 DPRINTF("found branch index %u [%.*s -> %lu], rc = %i",
1407 i, (int)node->mn_ksize, (char *)NODEKEY(node),
1418 if (rc > 0) { /* Found entry is less than the key. */
1419 i++; /* Skip to get the smallest entry larger than key. */
1420 if (i >= NUMKEYS(mp))
1421 /* There is no entry larger or equal to the key. */
1425 *exactp = (rc == 0);
1426 if (kip) /* Store the key index if requested. */
1429 return NODEPTR(mp, i);
1433 cursor_pop_page(MDB_cursor *cursor)
1437 top = CURSOR_TOP(cursor);
1440 DPRINTF("popped page %lu off cursor %p", top->mp_page->mp_pgno, (void *) cursor);
1446 cursor_push_page(MDB_cursor *cursor, MDB_page *mp)
1450 DPRINTF("pushing page %lu on cursor %p", mp->mp_pgno, (void *) cursor);
1452 if ((ppage = calloc(1, sizeof(*ppage))) == NULL)
1454 ppage->mp_page = mp;
1455 CURSOR_PUSH(cursor, ppage);
1460 mdb_get_page(MDB_txn *txn, pgno_t pgno)
1465 if (!F_ISSET(txn->mt_flags, MDB_TXN_RDONLY) && !STAILQ_EMPTY(txn->mt_u.dirty_queue)) {
1467 STAILQ_FOREACH(dp, txn->mt_u.dirty_queue, h.md_next) {
1468 if (dp->p.mp_pgno == pgno) {
1476 p = (MDB_page *)(txn->mt_env->me_map + txn->mt_env->me_psize * pgno);
1482 mdb_search_page_root(MDB_txn *txn, MDB_dbi dbi, MDB_val *key,
1483 MDB_cursor *cursor, int modify, MDB_pageparent *mpp)
1485 MDB_page *mp = mpp->mp_page;
1488 if (cursor && cursor_push_page(cursor, mp) == NULL)
1491 while (IS_BRANCH(mp)) {
1495 DPRINTF("branch page %lu has %u keys", mp->mp_pgno, NUMKEYS(mp));
1496 assert(NUMKEYS(mp) > 1);
1497 DPRINTF("found index 0 to page %lu", NODEPGNO(NODEPTR(mp, 0)));
1499 if (key == NULL) /* Initialize cursor to first page. */
1501 else if (key->mv_size > MAXKEYSIZE && key->mv_data == NULL) {
1502 /* cursor to last page */
1506 node = mdb_search_node(txn, dbi, mp, key, &exact, &i);
1508 i = NUMKEYS(mp) - 1;
1516 DPRINTF("following index %u for key %.*s",
1517 i, (int)key->mv_size, (char *)key->mv_data);
1518 assert(i < NUMKEYS(mp));
1519 node = NODEPTR(mp, i);
1522 CURSOR_TOP(cursor)->mp_ki = i;
1524 mpp->mp_parent = mp;
1525 if ((mp = mdb_get_page(txn, NODEPGNO(node))) == NULL)
1530 if (cursor && cursor_push_page(cursor, mp) == NULL)
1534 MDB_dhead *dh = ((MDB_dhead *)mp)-1;
1535 if ((rc = mdb_touch(txn, mpp)) != 0)
1537 dh = ((MDB_dhead *)mpp->mp_page)-1;
1538 dh->md_parent = mpp->mp_parent;
1539 dh->md_pi = mpp->mp_pi;
1546 DPRINTF("internal error, index points to a %02X page!?",
1551 DPRINTF("found leaf page %lu for key %.*s", mp->mp_pgno,
1552 key ? (int)key->mv_size : 0, key ? (char *)key->mv_data : NULL);
1557 /* Search for the page a given key should be in.
1558 * Stores a pointer to the found page in *mpp.
1559 * If key is NULL, search for the lowest page (used by mdb_cursor_first).
1560 * If cursor is non-null, pushes parent pages on the cursor stack.
1561 * If modify is true, visited pages are updated with new page numbers.
1564 mdb_search_page(MDB_txn *txn, MDB_dbi dbi, MDB_val *key,
1565 MDB_cursor *cursor, int modify, MDB_pageparent *mpp)
1570 /* Choose which root page to start with. If a transaction is given
1571 * use the root page from the transaction, otherwise read the last
1572 * committed root page.
1574 if (F_ISSET(txn->mt_flags, MDB_TXN_ERROR)) {
1575 DPRINTF("transaction has failed, must abort");
1578 root = txn->mt_dbs[dbi].md_root;
1580 if (root == P_INVALID) { /* Tree is empty. */
1581 DPRINTF("tree is empty");
1585 if ((mpp->mp_page = mdb_get_page(txn, root)) == NULL)
1588 DPRINTF("root page has flags 0x%X", mpp->mp_page->mp_flags);
1591 /* For sub-databases, update main root first */
1592 if (dbi > MAIN_DBI && !txn->mt_dbxs[dbi].md_dirty) {
1594 rc = mdb_search_page(txn, 0, &txn->mt_dbxs[dbi].md_name,
1598 txn->mt_dbxs[dbi].md_dirty = 1;
1600 if (!F_ISSET(mpp->mp_page->mp_flags, P_DIRTY)) {
1601 mpp->mp_parent = NULL;
1603 if ((rc = mdb_touch(txn, mpp)))
1605 txn->mt_dbs[dbi].md_root = mpp->mp_page->mp_pgno;
1609 return mdb_search_page_root(txn, dbi, key, cursor, modify, mpp);
1613 mdb_read_data(MDB_txn *txn, MDB_node *leaf, MDB_val *data)
1615 MDB_page *omp; /* overflow mpage */
1618 if (!F_ISSET(leaf->mn_flags, F_BIGDATA)) {
1619 data->mv_size = leaf->mn_dsize;
1620 data->mv_data = NODEDATA(leaf);
1624 /* Read overflow data.
1626 data->mv_size = leaf->mn_dsize;
1627 memcpy(&pgno, NODEDATA(leaf), sizeof(pgno));
1628 if ((omp = mdb_get_page(txn, pgno)) == NULL) {
1629 DPRINTF("read overflow page %lu failed", pgno);
1632 data->mv_data = omp;
1638 mdb_get(MDB_txn *txn, MDB_dbi dbi,
1639 MDB_val *key, MDB_val *data)
1647 DPRINTF("===> get key [%.*s]", (int)key->mv_size, (char *)key->mv_data);
1649 if (key->mv_size == 0 || key->mv_size > MAXKEYSIZE) {
1653 if ((rc = mdb_search_page(txn, dbi, key, NULL, 0, &mpp)) != MDB_SUCCESS)
1656 leaf = mdb_search_node(txn, dbi, mpp.mp_page, key, &exact, NULL);
1658 rc = mdb_read_data(txn, leaf, data);
1667 mdb_sibling(MDB_cursor *cursor, int move_right)
1671 MDB_ppage *parent, *top;
1674 top = CURSOR_TOP(cursor);
1675 if ((parent = SLIST_NEXT(top, mp_entry)) == NULL) {
1676 return ENOENT; /* root has no siblings */
1679 DPRINTF("parent page is page %lu, index %u",
1680 parent->mp_page->mp_pgno, parent->mp_ki);
1682 cursor_pop_page(cursor);
1683 if (move_right ? (parent->mp_ki + 1 >= NUMKEYS(parent->mp_page))
1684 : (parent->mp_ki == 0)) {
1685 DPRINTF("no more keys left, moving to %s sibling",
1686 move_right ? "right" : "left");
1687 if ((rc = mdb_sibling(cursor, move_right)) != MDB_SUCCESS)
1689 parent = CURSOR_TOP(cursor);
1695 DPRINTF("just moving to %s index key %u",
1696 move_right ? "right" : "left", parent->mp_ki);
1698 assert(IS_BRANCH(parent->mp_page));
1700 indx = NODEPTR(parent->mp_page, parent->mp_ki);
1701 if ((mp = mdb_get_page(cursor->mc_txn, indx->mn_pgno)) == NULL)
1704 mp->parent = parent->mp_page;
1705 mp->parent_index = parent->mp_ki;
1708 cursor_push_page(cursor, mp);
1714 mdb_set_key(MDB_node *node, MDB_val *key)
1719 key->mv_size = node->mn_ksize;
1720 key->mv_data = NODEKEY(node);
1726 mdb_cursor_next(MDB_cursor *cursor, MDB_val *key, MDB_val *data)
1732 if (cursor->mc_eof) {
1736 assert(cursor->mc_initialized);
1738 top = CURSOR_TOP(cursor);
1741 DPRINTF("cursor_next: top page is %lu in cursor %p", mp->mp_pgno, (void *) cursor);
1743 if (top->mp_ki + 1 >= NUMKEYS(mp)) {
1744 DPRINTF("=====> move to next sibling page");
1745 if (mdb_sibling(cursor, 1) != MDB_SUCCESS) {
1749 top = CURSOR_TOP(cursor);
1751 DPRINTF("next page is %lu, key index %u", mp->mp_pgno, top->mp_ki);
1755 DPRINTF("==> cursor points to page %lu with %u keys, key index %u",
1756 mp->mp_pgno, NUMKEYS(mp), top->mp_ki);
1758 assert(IS_LEAF(mp));
1759 leaf = NODEPTR(mp, top->mp_ki);
1761 if (data && mdb_read_data(cursor->mc_txn, leaf, data) != MDB_SUCCESS)
1764 return mdb_set_key(leaf, key);
1768 mdb_cursor_prev(MDB_cursor *cursor, MDB_val *key, MDB_val *data)
1774 assert(cursor->mc_initialized);
1776 top = CURSOR_TOP(cursor);
1779 DPRINTF("cursor_prev: top page is %lu in cursor %p", mp->mp_pgno, (void *) cursor);
1781 if (top->mp_ki == 0) {
1782 DPRINTF("=====> move to prev sibling page");
1783 if (mdb_sibling(cursor, 0) != MDB_SUCCESS) {
1786 top = CURSOR_TOP(cursor);
1788 top->mp_ki = NUMKEYS(mp) - 1;
1789 DPRINTF("prev page is %lu, key index %u", mp->mp_pgno, top->mp_ki);
1795 DPRINTF("==> cursor points to page %lu with %u keys, key index %u",
1796 mp->mp_pgno, NUMKEYS(mp), top->mp_ki);
1798 assert(IS_LEAF(mp));
1799 leaf = NODEPTR(mp, top->mp_ki);
1801 if (data && mdb_read_data(cursor->mc_txn, leaf, data) != MDB_SUCCESS)
1804 return mdb_set_key(leaf, key);
1808 mdb_cursor_set(MDB_cursor *cursor, MDB_val *key, MDB_val *data,
1818 assert(key->mv_size > 0);
1820 rc = mdb_search_page(cursor->mc_txn, cursor->mc_dbi, key, cursor, 0, &mpp);
1821 if (rc != MDB_SUCCESS)
1823 assert(IS_LEAF(mpp.mp_page));
1825 top = CURSOR_TOP(cursor);
1826 leaf = mdb_search_node(cursor->mc_txn, cursor->mc_dbi, mpp.mp_page, key, exactp, &top->mp_ki);
1827 if (exactp != NULL && !*exactp) {
1828 /* MDB_SET specified and not an exact match. */
1833 DPRINTF("===> inexact leaf not found, goto sibling");
1834 if ((rc = mdb_sibling(cursor, 1)) != MDB_SUCCESS)
1835 return rc; /* no entries matched */
1836 top = CURSOR_TOP(cursor);
1838 mpp.mp_page = top->mp_page;
1839 assert(IS_LEAF(mpp.mp_page));
1840 leaf = NODEPTR(mpp.mp_page, 0);
1843 cursor->mc_initialized = 1;
1846 if (data && (rc = mdb_read_data(cursor->mc_txn, leaf, data)) != MDB_SUCCESS)
1849 rc = mdb_set_key(leaf, key);
1850 if (rc == MDB_SUCCESS) {
1851 DPRINTF("==> cursor placed on key %.*s",
1852 (int)key->mv_size, (char *)key->mv_data);
1860 mdb_cursor_first(MDB_cursor *cursor, MDB_val *key, MDB_val *data)
1866 rc = mdb_search_page(cursor->mc_txn, cursor->mc_dbi, NULL, cursor, 0, &mpp);
1867 if (rc != MDB_SUCCESS)
1869 assert(IS_LEAF(mpp.mp_page));
1871 leaf = NODEPTR(mpp.mp_page, 0);
1872 cursor->mc_initialized = 1;
1875 if (data && (rc = mdb_read_data(cursor->mc_txn, leaf, data)) != MDB_SUCCESS)
1878 return mdb_set_key(leaf, key);
1882 mdb_cursor_last(MDB_cursor *cursor, MDB_val *key, MDB_val *data)
1890 lkey.mv_size = MAXKEYSIZE+1;
1891 lkey.mv_data = NULL;
1893 rc = mdb_search_page(cursor->mc_txn, cursor->mc_dbi, &lkey, cursor, 0, &mpp);
1894 if (rc != MDB_SUCCESS)
1896 assert(IS_LEAF(mpp.mp_page));
1898 leaf = NODEPTR(mpp.mp_page, NUMKEYS(mpp.mp_page)-1);
1899 cursor->mc_initialized = 1;
1902 top = CURSOR_TOP(cursor);
1903 top->mp_ki = NUMKEYS(top->mp_page) - 1;
1905 if (data && (rc = mdb_read_data(cursor->mc_txn, leaf, data)) != MDB_SUCCESS)
1908 return mdb_set_key(leaf, key);
1912 mdb_cursor_get(MDB_cursor *cursor, MDB_val *key, MDB_val *data,
1923 while (CURSOR_TOP(cursor) != NULL)
1924 cursor_pop_page(cursor);
1925 if (key == NULL || key->mv_size == 0 || key->mv_size > MAXKEYSIZE) {
1927 } else if (op == MDB_SET)
1928 rc = mdb_cursor_set(cursor, key, data, &exact);
1930 rc = mdb_cursor_set(cursor, key, data, NULL);
1933 if (!cursor->mc_initialized)
1934 rc = mdb_cursor_first(cursor, key, data);
1936 rc = mdb_cursor_next(cursor, key, data);
1939 if (!cursor->mc_initialized || cursor->mc_eof) {
1940 while (CURSOR_TOP(cursor) != NULL)
1941 cursor_pop_page(cursor);
1942 rc = mdb_cursor_last(cursor, key, data);
1944 rc = mdb_cursor_prev(cursor, key, data);
1947 while (CURSOR_TOP(cursor) != NULL)
1948 cursor_pop_page(cursor);
1949 rc = mdb_cursor_first(cursor, key, data);
1952 while (CURSOR_TOP(cursor) != NULL)
1953 cursor_pop_page(cursor);
1954 rc = mdb_cursor_last(cursor, key, data);
1957 DPRINTF("unhandled/unimplemented cursor operation %u", op);
1965 /* Allocate a page and initialize it
1968 mdb_new_page(MDB_txn *txn, MDB_dbi dbi, uint32_t flags, int num)
1972 if ((dp = mdb_alloc_page(txn, NULL, 0, num)) == NULL)
1974 DPRINTF("allocated new mpage %lu, page size %u",
1975 dp->p.mp_pgno, txn->mt_env->me_psize);
1976 dp->p.mp_flags = flags | P_DIRTY;
1977 dp->p.mp_lower = PAGEHDRSZ;
1978 dp->p.mp_upper = txn->mt_env->me_psize;
1980 if (IS_BRANCH(&dp->p))
1981 txn->mt_dbs[dbi].md_branch_pages++;
1982 else if (IS_LEAF(&dp->p))
1983 txn->mt_dbs[dbi].md_leaf_pages++;
1984 else if (IS_OVERFLOW(&dp->p)) {
1985 txn->mt_dbs[dbi].md_overflow_pages += num;
1986 dp->p.mp_pages = num;
1993 mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data)
1997 sz = LEAFSIZE(key, data);
1998 if (data->mv_size >= env->me_psize / MDB_MINKEYS) {
1999 /* put on overflow page */
2000 sz -= data->mv_size - sizeof(pgno_t);
2003 return sz + sizeof(indx_t);
2007 mdb_branch_size(MDB_env *env, MDB_val *key)
2012 if (sz >= env->me_psize / MDB_MINKEYS) {
2013 /* put on overflow page */
2014 /* not implemented */
2015 /* sz -= key->size - sizeof(pgno_t); */
2018 return sz + sizeof(indx_t);
2022 mdb_add_node(MDB_txn *txn, MDB_dbi dbi, MDB_page *mp, indx_t indx,
2023 MDB_val *key, MDB_val *data, pgno_t pgno, uint8_t flags)
2026 size_t node_size = NODESIZE;
2029 MDB_dpage *ofp = NULL; /* overflow page */
2031 assert(mp->mp_upper >= mp->mp_lower);
2033 DPRINTF("add node [%.*s] to %s page %lu at index %i, key size %zu",
2034 key ? (int)key->mv_size : 0, key ? (char *)key->mv_data : NULL,
2035 IS_LEAF(mp) ? "leaf" : "branch",
2036 mp->mp_pgno, indx, key ? key->mv_size : 0);
2039 node_size += key->mv_size;
2043 if (F_ISSET(flags, F_BIGDATA)) {
2044 /* Data already on overflow page. */
2045 node_size += sizeof(pgno_t);
2046 } else if (data->mv_size >= txn->mt_env->me_psize / MDB_MINKEYS) {
2047 int ovpages = OVPAGES(data->mv_size, txn->mt_env->me_psize);
2048 /* Put data on overflow page. */
2049 DPRINTF("data size is %zu, put on overflow page",
2051 node_size += sizeof(pgno_t);
2052 if ((ofp = mdb_new_page(txn, dbi, P_OVERFLOW, ovpages)) == NULL)
2054 DPRINTF("allocated overflow page %lu", ofp->p.mp_pgno);
2057 node_size += data->mv_size;
2061 if (node_size + sizeof(indx_t) > SIZELEFT(mp)) {
2062 DPRINTF("not enough room in page %lu, got %u ptrs",
2063 mp->mp_pgno, NUMKEYS(mp));
2064 DPRINTF("upper - lower = %u - %u = %u", mp->mp_upper, mp->mp_lower,
2065 mp->mp_upper - mp->mp_lower);
2066 DPRINTF("node size = %zu", node_size);
2070 /* Move higher pointers up one slot. */
2071 for (i = NUMKEYS(mp); i > indx; i--)
2072 mp->mp_ptrs[i] = mp->mp_ptrs[i - 1];
2074 /* Adjust free space offsets. */
2075 ofs = mp->mp_upper - node_size;
2076 assert(ofs >= mp->mp_lower + sizeof(indx_t));
2077 mp->mp_ptrs[indx] = ofs;
2079 mp->mp_lower += sizeof(indx_t);
2081 /* Write the node data. */
2082 node = NODEPTR(mp, indx);
2083 node->mn_ksize = (key == NULL) ? 0 : key->mv_size;
2084 node->mn_flags = flags;
2086 node->mn_dsize = data->mv_size;
2088 node->mn_pgno = pgno;
2091 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
2096 if (F_ISSET(flags, F_BIGDATA))
2097 memcpy(node->mn_data + key->mv_size, data->mv_data,
2100 memcpy(node->mn_data + key->mv_size, data->mv_data,
2103 memcpy(node->mn_data + key->mv_size, &ofp->p.mp_pgno,
2105 memcpy(METADATA(&ofp->p), data->mv_data, data->mv_size);
2113 mdb_del_node(MDB_page *mp, indx_t indx)
2116 indx_t i, j, numkeys, ptr;
2120 DPRINTF("delete node %u on %s page %lu", indx,
2121 IS_LEAF(mp) ? "leaf" : "branch", mp->mp_pgno);
2122 assert(indx < NUMKEYS(mp));
2124 node = NODEPTR(mp, indx);
2125 sz = NODESIZE + node->mn_ksize;
2127 if (F_ISSET(node->mn_flags, F_BIGDATA))
2128 sz += sizeof(pgno_t);
2130 sz += NODEDSZ(node);
2133 ptr = mp->mp_ptrs[indx];
2134 numkeys = NUMKEYS(mp);
2135 for (i = j = 0; i < numkeys; i++) {
2137 mp->mp_ptrs[j] = mp->mp_ptrs[i];
2138 if (mp->mp_ptrs[i] < ptr)
2139 mp->mp_ptrs[j] += sz;
2144 base = (char *)mp + mp->mp_upper;
2145 memmove(base + sz, base, ptr - mp->mp_upper);
2147 mp->mp_lower -= sizeof(indx_t);
2152 mdb_cursor_open(MDB_txn *txn, MDB_dbi dbi, MDB_cursor **ret)
2155 size_t size = sizeof(MDB_cursor);
2157 if (txn == NULL || ret == NULL || !dbi || dbi >= txn->mt_numdbs)
2160 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT)
2161 size += sizeof(MDB_xcursor);
2163 if ((cursor = calloc(1, size)) != NULL) {
2164 SLIST_INIT(&cursor->mc_stack);
2165 cursor->mc_dbi = dbi;
2166 cursor->mc_txn = txn;
2167 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
2168 MDB_xcursor *mx = (MDB_xcursor *)(cursor + 1);
2169 cursor->mc_xcursor = mx;
2170 mx->mx_cursor.mc_txn = &mx->mx_txn;
2172 mx->mx_txn.mt_dbxs = mx->mx_dbxs;
2173 mx->mx_txn.mt_dbs = mx->mx_dbs;
2174 mx->mx_dbxs[0] = txn->mt_dbxs[0];
2175 mx->mx_dbxs[1] = txn->mt_dbxs[1];
2177 mx->mx_dbxs[2] = txn->mt_dbxs[dbi];
2178 mx->mx_txn.mt_numdbs = 4;
2180 mx->mx_txn.mt_numdbs = 3;
2193 mdb_cursor_close(MDB_cursor *cursor)
2195 if (cursor != NULL) {
2196 while (!CURSOR_EMPTY(cursor))
2197 cursor_pop_page(cursor);
2199 /* btree_close(cursor->bt); */
2205 mdb_update_key(MDB_page *mp, indx_t indx, MDB_val *key)
2207 indx_t ptr, i, numkeys;
2213 node = NODEPTR(mp, indx);
2214 ptr = mp->mp_ptrs[indx];
2215 DPRINTF("update key %u (ofs %u) [%.*s] to [%.*s] on page %lu",
2217 (int)node->mn_ksize, (char *)NODEKEY(node),
2218 (int)key->mv_size, (char *)key->mv_data,
2221 delta = key->mv_size - node->mn_ksize;
2223 if (delta > 0 && SIZELEFT(mp) < delta) {
2224 DPRINTF("OUCH! Not enough room, delta = %d", delta);
2228 numkeys = NUMKEYS(mp);
2229 for (i = 0; i < numkeys; i++) {
2230 if (mp->mp_ptrs[i] <= ptr)
2231 mp->mp_ptrs[i] -= delta;
2234 base = (char *)mp + mp->mp_upper;
2235 len = ptr - mp->mp_upper + NODESIZE;
2236 memmove(base - delta, base, len);
2237 mp->mp_upper -= delta;
2239 node = NODEPTR(mp, indx);
2240 node->mn_ksize = key->mv_size;
2243 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
2248 /* Move a node from src to dst.
2251 mdb_move_node(MDB_txn *txn, MDB_dbi dbi, MDB_pageparent *src, indx_t srcindx,
2252 MDB_pageparent *dst, indx_t dstindx)
2258 srcnode = NODEPTR(src->mp_page, srcindx);
2259 DPRINTF("moving %s node %u [%.*s] on page %lu to node %u on page %lu",
2260 IS_LEAF(src->mp_page) ? "leaf" : "branch",
2262 (int)srcnode->mn_ksize, (char *)NODEKEY(srcnode),
2263 src->mp_page->mp_pgno,
2264 dstindx, dst->mp_page->mp_pgno);
2266 /* Mark src and dst as dirty. */
2267 if ((rc = mdb_touch(txn, src)) ||
2268 (rc = mdb_touch(txn, dst)))
2271 /* Add the node to the destination page.
2273 key.mv_size = srcnode->mn_ksize;
2274 key.mv_data = NODEKEY(srcnode);
2275 data.mv_size = NODEDSZ(srcnode);
2276 data.mv_data = NODEDATA(srcnode);
2277 rc = mdb_add_node(txn, dbi, dst->mp_page, dstindx, &key, &data, NODEPGNO(srcnode),
2279 if (rc != MDB_SUCCESS)
2282 /* Delete the node from the source page.
2284 mdb_del_node(src->mp_page, srcindx);
2286 /* Update the parent separators.
2288 if (srcindx == 0 && src->mp_pi != 0) {
2289 DPRINTF("update separator for source page %lu to [%.*s]",
2290 src->mp_page->mp_pgno, (int)key.mv_size, (char *)key.mv_data);
2291 if ((rc = mdb_update_key(src->mp_parent, src->mp_pi,
2292 &key)) != MDB_SUCCESS)
2296 if (srcindx == 0 && IS_BRANCH(src->mp_page)) {
2298 nullkey.mv_size = 0;
2299 assert(mdb_update_key(src->mp_page, 0, &nullkey) == MDB_SUCCESS);
2302 if (dstindx == 0 && dst->mp_pi != 0) {
2303 DPRINTF("update separator for destination page %lu to [%.*s]",
2304 dst->mp_page->mp_pgno, (int)key.mv_size, (char *)key.mv_data);
2305 if ((rc = mdb_update_key(dst->mp_parent, dst->mp_pi,
2306 &key)) != MDB_SUCCESS)
2310 if (dstindx == 0 && IS_BRANCH(dst->mp_page)) {
2312 nullkey.mv_size = 0;
2313 assert(mdb_update_key(dst->mp_page, 0, &nullkey) == MDB_SUCCESS);
2320 mdb_merge(MDB_txn *txn, MDB_dbi dbi, MDB_pageparent *src, MDB_pageparent *dst)
2329 DPRINTF("merging page %lu and %lu", src->mp_page->mp_pgno, dst->mp_page->mp_pgno);
2331 assert(txn != NULL);
2332 assert(src->mp_parent); /* can't merge root page */
2333 assert(dst->mp_parent);
2335 /* Mark src and dst as dirty. */
2336 if ((rc = mdb_touch(txn, src)) ||
2337 (rc = mdb_touch(txn, dst)))
2340 /* Move all nodes from src to dst.
2342 for (i = 0; i < NUMKEYS(src->mp_page); i++) {
2343 srcnode = NODEPTR(src->mp_page, i);
2345 key.mv_size = srcnode->mn_ksize;
2346 key.mv_data = NODEKEY(srcnode);
2347 data.mv_size = NODEDSZ(srcnode);
2348 data.mv_data = NODEDATA(srcnode);
2349 rc = mdb_add_node(txn, dbi, dst->mp_page, NUMKEYS(dst->mp_page), &key,
2350 &data, NODEPGNO(srcnode), srcnode->mn_flags);
2351 if (rc != MDB_SUCCESS)
2355 DPRINTF("dst page %lu now has %u keys (%.1f%% filled)",
2356 dst->mp_page->mp_pgno, NUMKEYS(dst->mp_page), (float)PAGEFILL(txn->mt_env, dst->mp_page) / 10);
2358 /* Unlink the src page from parent.
2360 mdb_del_node(src->mp_parent, src->mp_pi);
2361 if (src->mp_pi == 0) {
2363 if ((rc = mdb_update_key(src->mp_parent, 0, &key)) != MDB_SUCCESS)
2367 if (IS_LEAF(src->mp_page))
2368 txn->mt_dbs[dbi].md_leaf_pages--;
2370 txn->mt_dbs[dbi].md_branch_pages--;
2372 mpp.mp_page = src->mp_parent;
2373 dh = (MDB_dhead *)src->mp_parent;
2375 mpp.mp_parent = dh->md_parent;
2376 mpp.mp_pi = dh->md_pi;
2378 return mdb_rebalance(txn, dbi, &mpp);
2381 #define FILL_THRESHOLD 250
2384 mdb_rebalance(MDB_txn *txn, MDB_dbi dbi, MDB_pageparent *mpp)
2389 indx_t si = 0, di = 0;
2391 assert(txn != NULL);
2392 assert(mpp != NULL);
2394 DPRINTF("rebalancing %s page %lu (has %u keys, %.1f%% full)",
2395 IS_LEAF(mpp->mp_page) ? "leaf" : "branch",
2396 mpp->mp_page->mp_pgno, NUMKEYS(mpp->mp_page), (float)PAGEFILL(txn->mt_env, mpp->mp_page) / 10);
2398 if (PAGEFILL(txn->mt_env, mpp->mp_page) >= FILL_THRESHOLD) {
2399 DPRINTF("no need to rebalance page %lu, above fill threshold",
2400 mpp->mp_page->mp_pgno);
2404 if (mpp->mp_parent == NULL) {
2405 if (NUMKEYS(mpp->mp_page) == 0) {
2406 DPRINTF("tree is completely empty");
2407 txn->mt_dbs[dbi].md_root = P_INVALID;
2408 txn->mt_dbs[dbi].md_depth--;
2409 txn->mt_dbs[dbi].md_leaf_pages--;
2410 } else if (IS_BRANCH(mpp->mp_page) && NUMKEYS(mpp->mp_page) == 1) {
2411 DPRINTF("collapsing root page!");
2412 txn->mt_dbs[dbi].md_root = NODEPGNO(NODEPTR(mpp->mp_page, 0));
2413 if ((root = mdb_get_page(txn, txn->mt_dbs[dbi].md_root)) == NULL)
2415 txn->mt_dbs[dbi].md_depth--;
2416 txn->mt_dbs[dbi].md_branch_pages--;
2418 DPRINTF("root page doesn't need rebalancing");
2422 /* The parent (branch page) must have at least 2 pointers,
2423 * otherwise the tree is invalid.
2425 assert(NUMKEYS(mpp->mp_parent) > 1);
2427 /* Leaf page fill factor is below the threshold.
2428 * Try to move keys from left or right neighbor, or
2429 * merge with a neighbor page.
2434 if (mpp->mp_pi == 0) {
2435 /* We're the leftmost leaf in our parent.
2437 DPRINTF("reading right neighbor");
2438 node = NODEPTR(mpp->mp_parent, mpp->mp_pi + 1);
2439 if ((npp.mp_page = mdb_get_page(txn, NODEPGNO(node))) == NULL)
2441 npp.mp_pi = mpp->mp_pi + 1;
2443 di = NUMKEYS(mpp->mp_page);
2445 /* There is at least one neighbor to the left.
2447 DPRINTF("reading left neighbor");
2448 node = NODEPTR(mpp->mp_parent, mpp->mp_pi - 1);
2449 if ((npp.mp_page = mdb_get_page(txn, NODEPGNO(node))) == NULL)
2451 npp.mp_pi = mpp->mp_pi - 1;
2452 si = NUMKEYS(npp.mp_page) - 1;
2455 npp.mp_parent = mpp->mp_parent;
2457 DPRINTF("found neighbor page %lu (%u keys, %.1f%% full)",
2458 npp.mp_page->mp_pgno, NUMKEYS(npp.mp_page), (float)PAGEFILL(txn->mt_env, npp.mp_page) / 10);
2460 /* If the neighbor page is above threshold and has at least two
2461 * keys, move one key from it.
2463 * Otherwise we should try to merge them.
2465 if (PAGEFILL(txn->mt_env, npp.mp_page) >= FILL_THRESHOLD && NUMKEYS(npp.mp_page) >= 2)
2466 return mdb_move_node(txn, dbi, &npp, si, mpp, di);
2467 else { /* FIXME: if (has_enough_room()) */
2468 if (mpp->mp_pi == 0)
2469 return mdb_merge(txn, dbi, &npp, mpp);
2471 return mdb_merge(txn, dbi, mpp, &npp);
2476 mdb_del0(MDB_txn *txn, MDB_dbi dbi, unsigned int ki, MDB_pageparent *mpp, MDB_node *leaf)
2480 /* add overflow pages to free list */
2481 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) {
2485 memcpy(&pg, NODEDATA(leaf), sizeof(pg));
2486 ovpages = OVPAGES(NODEDSZ(leaf), txn->mt_env->me_psize);
2487 for (i=0; i<ovpages; i++) {
2488 DPRINTF("freed ov page %lu", pg);
2489 mdb_idl_insert(txn->mt_free_pgs, pg);
2493 mdb_del_node(mpp->mp_page, ki);
2494 txn->mt_dbs[dbi].md_entries--;
2495 rc = mdb_rebalance(txn, dbi, mpp);
2496 if (rc != MDB_SUCCESS)
2497 txn->mt_flags |= MDB_TXN_ERROR;
2503 mdb_del(MDB_txn *txn, MDB_dbi dbi,
2504 MDB_val *key, MDB_val *data)
2511 DPRINTF("========> delete key %.*s", (int)key->mv_size, (char *)key->mv_data);
2513 assert(key != NULL);
2515 if (txn == NULL || dbi >= txn->mt_numdbs)
2518 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
2522 if (key->mv_size == 0 || key->mv_size > MAXKEYSIZE) {
2526 mpp.mp_parent = NULL;
2528 if ((rc = mdb_search_page(txn, dbi, key, NULL, 1, &mpp)) != MDB_SUCCESS)
2531 leaf = mdb_search_node(txn, dbi, mpp.mp_page, key, &exact, &ki);
2532 if (leaf == NULL || !exact) {
2536 if (data && (rc = mdb_read_data(txn, leaf, data)) != MDB_SUCCESS)
2539 return mdb_del0(txn, dbi, ki, &mpp, leaf);
2542 /* Split page <*mpp>, and insert <key,(data|newpgno)> in either left or
2543 * right sibling, at index <*newindxp> (as if unsplit). Updates *mpp and
2544 * *newindxp with the actual values after split, ie if *mpp and *newindxp
2545 * refer to a node in the new right sibling page.
2548 mdb_split(MDB_txn *txn, MDB_dbi dbi, MDB_page **mpp, unsigned int *newindxp,
2549 MDB_val *newkey, MDB_val *newdata, pgno_t newpgno)
2552 int rc = MDB_SUCCESS, ins_new = 0;
2555 unsigned int i, j, split_indx;
2557 MDB_val sepkey, rkey, rdata;
2559 MDB_dpage *mdp, *rdp, *pdp;
2562 assert(txn != NULL);
2564 dh = ((MDB_dhead *)*mpp) - 1;
2565 mdp = (MDB_dpage *)dh;
2566 newindx = *newindxp;
2568 DPRINTF("-----> splitting %s page %lu and adding [%.*s] at index %i",
2569 IS_LEAF(&mdp->p) ? "leaf" : "branch", mdp->p.mp_pgno,
2570 (int)newkey->mv_size, (char *)newkey->mv_data, *newindxp);
2572 if (mdp->h.md_parent == NULL) {
2573 if ((pdp = mdb_new_page(txn, dbi, P_BRANCH, 1)) == NULL)
2576 mdp->h.md_parent = &pdp->p;
2577 txn->mt_dbs[dbi].md_root = pdp->p.mp_pgno;
2578 DPRINTF("root split! new root = %lu", pdp->p.mp_pgno);
2579 txn->mt_dbs[dbi].md_depth++;
2581 /* Add left (implicit) pointer. */
2582 if (mdb_add_node(txn, dbi, &pdp->p, 0, NULL, NULL,
2583 mdp->p.mp_pgno, 0) != MDB_SUCCESS)
2586 DPRINTF("parent branch page is %lu", mdp->h.md_parent->mp_pgno);
2589 /* Create a right sibling. */
2590 if ((rdp = mdb_new_page(txn, dbi, mdp->p.mp_flags, 1)) == NULL)
2592 rdp->h.md_parent = mdp->h.md_parent;
2593 rdp->h.md_pi = mdp->h.md_pi + 1;
2594 DPRINTF("new right sibling: page %lu", rdp->p.mp_pgno);
2596 /* Move half of the keys to the right sibling. */
2597 if ((copy = malloc(txn->mt_env->me_psize)) == NULL)
2599 memcpy(copy, &mdp->p, txn->mt_env->me_psize);
2600 memset(&mdp->p.mp_ptrs, 0, txn->mt_env->me_psize - PAGEHDRSZ);
2601 mdp->p.mp_lower = PAGEHDRSZ;
2602 mdp->p.mp_upper = txn->mt_env->me_psize;
2604 split_indx = NUMKEYS(copy) / 2 + 1;
2606 /* First find the separating key between the split pages.
2608 memset(&sepkey, 0, sizeof(sepkey));
2609 if (newindx == split_indx) {
2610 sepkey.mv_size = newkey->mv_size;
2611 sepkey.mv_data = newkey->mv_data;
2613 node = NODEPTR(copy, split_indx);
2614 sepkey.mv_size = node->mn_ksize;
2615 sepkey.mv_data = NODEKEY(node);
2618 DPRINTF("separator is [%.*s]", (int)sepkey.mv_size, (char *)sepkey.mv_data);
2620 /* Copy separator key to the parent.
2622 if (SIZELEFT(rdp->h.md_parent) < mdb_branch_size(txn->mt_env, &sepkey)) {
2623 rc = mdb_split(txn, dbi, &rdp->h.md_parent, &rdp->h.md_pi,
2624 &sepkey, NULL, rdp->p.mp_pgno);
2626 /* Right page might now have changed parent.
2627 * Check if left page also changed parent.
2629 if (rdp->h.md_parent != mdp->h.md_parent &&
2630 mdp->h.md_pi >= NUMKEYS(mdp->h.md_parent)) {
2631 mdp->h.md_parent = rdp->h.md_parent;
2632 mdp->h.md_pi = rdp->h.md_pi - 1;
2635 rc = mdb_add_node(txn, dbi, rdp->h.md_parent, rdp->h.md_pi,
2636 &sepkey, NULL, rdp->p.mp_pgno, 0);
2638 if (rc != MDB_SUCCESS) {
2643 for (i = j = 0; i <= NUMKEYS(copy); j++) {
2644 if (i < split_indx) {
2645 /* Re-insert in left sibling. */
2648 /* Insert in right sibling. */
2649 if (i == split_indx)
2650 /* Reset insert index for right sibling. */
2651 j = (i == newindx && ins_new);
2655 if (i == newindx && !ins_new) {
2656 /* Insert the original entry that caused the split. */
2657 rkey.mv_data = newkey->mv_data;
2658 rkey.mv_size = newkey->mv_size;
2659 if (IS_LEAF(&mdp->p)) {
2660 rdata.mv_data = newdata->mv_data;
2661 rdata.mv_size = newdata->mv_size;
2668 /* Update page and index for the new key. */
2671 } else if (i == NUMKEYS(copy)) {
2674 node = NODEPTR(copy, i);
2675 rkey.mv_data = NODEKEY(node);
2676 rkey.mv_size = node->mn_ksize;
2677 if (IS_LEAF(&mdp->p)) {
2678 rdata.mv_data = NODEDATA(node);
2679 rdata.mv_size = node->mn_dsize;
2681 pgno = node->mn_pgno;
2682 flags = node->mn_flags;
2687 if (!IS_LEAF(&mdp->p) && j == 0) {
2688 /* First branch index doesn't need key data. */
2692 rc = mdb_add_node(txn, dbi, &pdp->p, j, &rkey, &rdata, pgno,flags);
2700 mdb_put(MDB_txn *txn, MDB_dbi dbi,
2701 MDB_val *key, MDB_val *data, unsigned int flags)
2703 int rc = MDB_SUCCESS, exact;
2708 assert(key != NULL);
2709 assert(data != NULL);
2714 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
2718 if (txn->mt_env->me_txn != txn) {
2722 if (key->mv_size == 0 || key->mv_size > MAXKEYSIZE) {
2726 DPRINTF("==> put key %.*s, size %zu, data size %zu",
2727 (int)key->mv_size, (char *)key->mv_data, key->mv_size, data->mv_size);
2729 mpp.mp_parent = NULL;
2731 rc = mdb_search_page(txn, dbi, key, NULL, 1, &mpp);
2732 if (rc == MDB_SUCCESS) {
2733 leaf = mdb_search_node(txn, dbi, mpp.mp_page, key, &exact, &ki);
2734 if (leaf && exact) {
2735 if (F_ISSET(flags, MDB_NOOVERWRITE)) {
2736 DPRINTF("duplicate key %.*s",
2737 (int)key->mv_size, (char *)key->mv_data);
2740 /* same size, just replace it */
2741 if (NODEDSZ(leaf) == data->mv_size) {
2742 memcpy(NODEDATA(leaf), data->mv_data, data->mv_size);
2745 mdb_del_node(mpp.mp_page, ki);
2747 if (leaf == NULL) { /* append if not found */
2748 ki = NUMKEYS(mpp.mp_page);
2749 DPRINTF("appending key at index %i", ki);
2751 } else if (rc == ENOENT) {
2753 /* new file, just write a root leaf page */
2754 DPRINTF("allocating new root leaf page");
2755 if ((dp = mdb_new_page(txn, dbi, P_LEAF, 1)) == NULL) {
2758 mpp.mp_page = &dp->p;
2759 txn->mt_dbs[dbi].md_root = mpp.mp_page->mp_pgno;
2760 txn->mt_dbs[dbi].md_depth++;
2766 assert(IS_LEAF(mpp.mp_page));
2767 DPRINTF("there are %u keys, should insert new key at index %i",
2768 NUMKEYS(mpp.mp_page), ki);
2770 if (SIZELEFT(mpp.mp_page) < mdb_leaf_size(txn->mt_env, key, data)) {
2771 rc = mdb_split(txn, dbi, &mpp.mp_page, &ki, key, data, P_INVALID);
2773 /* There is room already in this leaf page. */
2774 rc = mdb_add_node(txn, dbi, mpp.mp_page, ki, key, data, 0, 0);
2777 if (rc != MDB_SUCCESS)
2778 txn->mt_flags |= MDB_TXN_ERROR;
2780 txn->mt_dbs[dbi].md_entries++;
2787 mdbenv_get_flags(MDB_env *env, unsigned int *arg)
2792 *arg = env->me_flags;
2797 mdbenv_get_path(MDB_env *env, const char **arg)
2802 *arg = env->me_path;
2807 mdbenv_stat(MDB_env *env, MDB_stat *arg)
2809 if (env == NULL || arg == NULL)
2812 arg->ms_psize = env->me_psize;
2813 arg->ms_depth = env->me_meta->mm_dbs[MAIN_DBI].md_depth;
2814 arg->ms_branch_pages = env->me_meta->mm_dbs[MAIN_DBI].md_branch_pages;
2815 arg->ms_leaf_pages = env->me_meta->mm_dbs[MAIN_DBI].md_leaf_pages;
2816 arg->ms_overflow_pages = env->me_meta->mm_dbs[MAIN_DBI].md_overflow_pages;
2817 arg->ms_entries = env->me_meta->mm_dbs[MAIN_DBI].md_entries;
2822 int mdb_open(MDB_txn *txn, const char *name, unsigned int flags, MDB_dbi *dbi)
2835 /* Is the DB already open? */
2837 for (i=2; i<txn->mt_numdbs; i++) {
2838 if (len == txn->mt_dbxs[i].md_name.mv_size &&
2839 !strncmp(name, txn->mt_dbxs[i].md_name.mv_data, len)) {
2845 if (txn->mt_numdbs >= txn->mt_env->me_maxdbs - 1)
2848 /* Find the DB info */
2850 key.mv_data = (void *)name;
2851 rc = mdb_get(txn, MAIN_DBI, &key, &data);
2853 /* Create if requested */
2854 if (rc == ENOENT && (flags & MDB_CREATE)) {
2856 data.mv_size = sizeof(MDB_db);
2857 data.mv_data = &dummy;
2858 memset(&dummy, 0, sizeof(dummy));
2859 dummy.md_root = P_INVALID;
2860 dummy.md_flags = flags & 0xffff;
2861 rc = mdb_put(txn, 0, &key, &data, 0);
2864 /* OK, got info, add to table */
2865 if (rc == MDB_SUCCESS) {
2866 txn->mt_dbxs[txn->mt_numdbs].md_name.mv_data = strdup(name);
2867 txn->mt_dbxs[txn->mt_numdbs].md_name.mv_size = len;
2868 txn->mt_dbxs[txn->mt_numdbs].md_cmp = NULL;
2869 txn->mt_dbxs[txn->mt_numdbs].md_dcmp = NULL;
2870 txn->mt_dbxs[txn->mt_numdbs].md_rel = NULL;
2871 txn->mt_dbxs[txn->mt_numdbs].md_parent = MAIN_DBI;
2872 txn->mt_dbxs[txn->mt_numdbs].md_dirty = 0;
2873 memcpy(&txn->mt_dbs[txn->mt_numdbs], data.mv_data, sizeof(MDB_db));
2874 *dbi = txn->mt_numdbs;
2881 int mdb_stat(MDB_txn *txn, MDB_dbi dbi, MDB_stat *arg)
2883 if (txn == NULL || arg == NULL)
2886 arg->ms_psize = txn->mt_env->me_psize;
2887 arg->ms_depth = txn->mt_dbs[dbi].md_depth;
2888 arg->ms_branch_pages = txn->mt_dbs[dbi].md_branch_pages;
2889 arg->ms_leaf_pages = txn->mt_dbs[dbi].md_leaf_pages;
2890 arg->ms_overflow_pages = txn->mt_dbs[dbi].md_overflow_pages;
2891 arg->ms_entries = txn->mt_dbs[dbi].md_entries;
2896 void mdb_close(MDB_txn *txn, MDB_dbi dbi)
2899 if (dbi <= MAIN_DBI || dbi >= txn->mt_numdbs)
2901 ptr = txn->mt_dbxs[dbi].md_name.mv_data;
2902 txn->mt_dbxs[dbi].md_name.mv_data = NULL;
2903 txn->mt_dbxs[dbi].md_name.mv_size = 0;