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/param.h>
35 #ifdef HAVE_SYS_FILE_H
53 #define ULONG unsigned long
58 /* Note: If O_DSYNC is undefined but exists in /usr/include,
59 * preferably set some compiler flag to get the definition.
60 * Otherwise compile with the less efficient -DMDB_DSYNC=O_SYNC.
63 # define MDB_DSYNC O_DSYNC
70 #if !(__STDC_VERSION__ >= 199901L || defined(__GNUC__))
71 # define DPRINTF (void) /* Vararg macros may be unsupported */
73 # define DPRINTF(fmt, ...) /* Requires 2 or more args */ \
74 fprintf(stderr, "%s:%d: " fmt "\n", __func__, __LINE__, __VA_ARGS__)
76 # define DPRINTF(fmt, ...) ((void) 0)
78 #define DPUTS(arg) DPRINTF("%s", arg)
82 #define MDB_MAGIC 0xBEEFC0DE
84 #define MAXKEYSIZE 511
86 #define KBUF (MAXKEYSIZE*2+1)
87 #define DKBUF char kbuf[KBUF]
88 #define DKEY(x) mdb_dkey(x, kbuf)
94 #define P_INVALID (~0UL)
96 #define F_ISSET(w, f) (((w) & (f)) == (f))
98 typedef uint16_t indx_t;
100 #define DEFAULT_READERS 126
101 #define DEFAULT_MAPSIZE 1048576
103 /* Lock descriptor stuff */
105 #define CACHELINE 64 /* most CPUs. Itanium uses 128 */
108 typedef struct MDB_rxbody {
114 typedef struct MDB_reader {
117 #define mr_txnid mru.mrx.mrb_txnid
118 #define mr_pid mru.mrx.mrb_pid
119 #define mr_tid mru.mrx.mrb_tid
120 /* cache line alignment */
121 char pad[(sizeof(MDB_rxbody)+CACHELINE-1) & ~(CACHELINE-1)];
125 typedef struct MDB_txbody {
127 uint32_t mtb_version;
128 pthread_mutex_t mtb_mutex;
130 uint32_t mtb_numreaders;
131 uint32_t mtb_me_toggle;
134 typedef struct MDB_txninfo {
137 #define mti_magic mt1.mtb.mtb_magic
138 #define mti_version mt1.mtb.mtb_version
139 #define mti_mutex mt1.mtb.mtb_mutex
140 #define mti_txnid mt1.mtb.mtb_txnid
141 #define mti_numreaders mt1.mtb.mtb_numreaders
142 #define mti_me_toggle mt1.mtb.mtb_me_toggle
143 char pad[(sizeof(MDB_txbody)+CACHELINE-1) & ~(CACHELINE-1)];
146 pthread_mutex_t mt2_wmutex;
147 #define mti_wmutex mt2.mt2_wmutex
148 char pad[(sizeof(pthread_mutex_t)+CACHELINE-1) & ~(CACHELINE-1)];
150 MDB_reader mti_readers[1];
153 /* Common header for all page types. Overflow pages
154 * occupy a number of contiguous pages with no
155 * headers on any page after the first.
157 typedef struct MDB_page { /* represents a page of storage */
158 #define mp_pgno mp_p.p_pgno
160 pgno_t p_pgno; /* page number */
161 void * p_align; /* for IL32P64 */
163 #define P_BRANCH 0x01 /* branch page */
164 #define P_LEAF 0x02 /* leaf page */
165 #define P_OVERFLOW 0x04 /* overflow page */
166 #define P_META 0x08 /* meta page */
167 #define P_DIRTY 0x10 /* dirty page */
168 #define P_LEAF2 0x20 /* DB with small, fixed size keys and no data */
170 #define mp_lower mp_pb.pb.pb_lower
171 #define mp_upper mp_pb.pb.pb_upper
172 #define mp_pages mp_pb.pb_pages
175 indx_t pb_lower; /* lower bound of free space */
176 indx_t pb_upper; /* upper bound of free space */
178 uint32_t pb_pages; /* number of overflow pages */
180 indx_t mp_ptrs[1]; /* dynamic size */
183 #define PAGEHDRSZ ((unsigned) offsetof(MDB_page, mp_ptrs))
185 #define NUMKEYS(p) (((p)->mp_lower - PAGEHDRSZ) >> 1)
186 #define SIZELEFT(p) (indx_t)((p)->mp_upper - (p)->mp_lower)
187 #define PAGEFILL(env, p) (1000L * ((env)->me_psize - PAGEHDRSZ - SIZELEFT(p)) / \
188 ((env)->me_psize - PAGEHDRSZ))
189 #define IS_LEAF(p) F_ISSET((p)->mp_flags, P_LEAF)
190 #define IS_LEAF2(p) F_ISSET((p)->mp_flags, P_LEAF2)
191 #define IS_BRANCH(p) F_ISSET((p)->mp_flags, P_BRANCH)
192 #define IS_OVERFLOW(p) F_ISSET((p)->mp_flags, P_OVERFLOW)
194 #define OVPAGES(size, psize) ((PAGEHDRSZ-1 + (size)) / (psize) + 1)
196 typedef struct MDB_db {
197 uint32_t md_pad; /* also ksize for LEAF2 pages */
200 ULONG md_branch_pages;
202 ULONG md_overflow_pages;
210 typedef struct MDB_meta { /* meta (footer) page content */
213 void *mm_address; /* address for fixed mapping */
214 size_t mm_mapsize; /* size of mmap region */
215 MDB_db mm_dbs[2]; /* first is free space, 2nd is main db */
216 #define mm_psize mm_dbs[0].md_pad
217 #define mm_flags mm_dbs[0].md_flags
218 pgno_t mm_last_pg; /* last used page in file */
219 ULONG mm_txnid; /* txnid that committed this page */
222 typedef struct MDB_dhead { /* a dirty page */
224 unsigned md_pi; /* parent index */
228 typedef struct MDB_dpage {
233 typedef struct MDB_oldpages {
234 struct MDB_oldpages *mo_next;
236 pgno_t mo_pages[1]; /* dynamic */
239 typedef struct MDB_pageparent {
245 static MDB_dpage *mdb_alloc_page(MDB_txn *txn, MDB_page *parent, unsigned int parent_idx, int num);
246 static int mdb_touch(MDB_txn *txn, MDB_pageparent *mp);
248 typedef struct MDB_ppage { /* ordered list of pages */
250 unsigned int mp_ki; /* cursor index on page */
253 #define CURSOR_TOP(c) (&(c)->mc_stack[(c)->mc_snum-1])
254 #define CURSOR_PARENT(c) (&(c)->mc_stack[(c)->mc_snum-2])
260 MDB_ppage mc_stack[32]; /* stack of parent pages */
261 unsigned int mc_snum; /* number of pushed pages */
263 short mc_initialized; /* 1 if initialized */
264 short mc_eof; /* 1 if end is reached */
265 struct MDB_xcursor *mc_xcursor;
268 #define METADATA(p) ((void *)((char *)(p) + PAGEHDRSZ))
270 typedef struct MDB_node {
271 #define mn_pgno mn_p.np_pgno
272 #define mn_dsize mn_p.np_dsize
274 pgno_t np_pgno; /* child page number */
275 uint32_t np_dsize; /* leaf data size */
277 unsigned int mn_flags:4;
278 unsigned int mn_ksize:12; /* key size */
279 #define F_BIGDATA 0x01 /* data put on overflow page */
280 #define F_SUBDATA 0x02 /* data is a sub-database */
281 #define F_DUPDATA 0x04 /* data has duplicates */
285 typedef struct MDB_dbx {
287 MDB_cmp_func *md_cmp; /* user compare function */
288 MDB_cmp_func *md_dcmp; /* user dupsort function */
289 MDB_rel_func *md_rel; /* user relocate function */
291 unsigned int md_dirty;
295 pgno_t mt_next_pgno; /* next unallocated page */
299 pgno_t *mt_free_pgs; /* this is an IDL */
301 MIDL2 *dirty_list; /* modified pages */
304 MDB_dbx *mt_dbxs; /* array */
306 unsigned int mt_numdbs;
308 #define MDB_TXN_RDONLY 0x01 /* read-only transaction */
309 #define MDB_TXN_ERROR 0x02 /* an error has occurred */
310 #define MDB_TXN_METOGGLE 0x04 /* used meta page 1 */
311 unsigned int mt_flags;
314 /* Context for sorted-dup records */
315 typedef struct MDB_xcursor {
316 MDB_cursor mx_cursor;
325 int me_mfd; /* just for writing the meta pages */
326 #define MDB_FATAL_ERROR 0x80000000U
328 uint32_t me_extrapad; /* unused for now */
329 unsigned int me_maxreaders;
330 unsigned int me_numdbs;
331 unsigned int me_maxdbs;
334 MDB_txninfo *me_txns;
335 MDB_meta *me_metas[2];
337 MDB_txn *me_txn; /* current write transaction */
339 off_t me_size; /* current file size */
340 pgno_t me_maxpg; /* me_mapsize / me_psize */
341 unsigned int me_psize;
342 unsigned int me_db_toggle;
343 MDB_dbx *me_dbxs; /* array */
345 MDB_oldpages *me_pghead;
346 pthread_key_t me_txkey; /* thread-key for readers */
347 MDB_dpage *me_dpages;
348 pgno_t me_free_pgs[MDB_IDL_UM_SIZE];
349 MIDL2 me_dirty_list[MDB_IDL_DB_SIZE];
352 #define NODESIZE offsetof(MDB_node, mn_data)
354 #define INDXSIZE(k) (NODESIZE + ((k) == NULL ? 0 : (k)->mv_size))
355 #define LEAFSIZE(k, d) (NODESIZE + (k)->mv_size + (d)->mv_size)
356 #define NODEPTR(p, i) ((MDB_node *)((char *)(p) + (p)->mp_ptrs[i]))
357 #define NODEKEY(node) (void *)((node)->mn_data)
358 #define NODEDATA(node) (void *)((char *)(node)->mn_data + (node)->mn_ksize)
359 #define NODEPGNO(node) ((node)->mn_pgno)
360 #define NODEDSZ(node) ((node)->mn_dsize)
361 #define NODEKSZ(node) ((node)->mn_ksize)
362 #define LEAF2KEY(p, i, ks) ((char *)(p) + PAGEHDRSZ + ((i)*(ks)))
364 #define MDB_SET_KEY(node, key) if (key!=NULL) {(key)->mv_size = NODEKSZ(node); (key)->mv_data = NODEKEY(node);}
366 #define MDB_COMMIT_PAGES 64 /* max number of pages to write in one commit */
368 static int mdb_search_page_root(MDB_txn *txn,
369 MDB_dbi dbi, MDB_val *key,
370 MDB_cursor *cursor, int modify,
371 MDB_pageparent *mpp);
372 static int mdb_search_page(MDB_txn *txn,
373 MDB_dbi dbi, MDB_val *key,
374 MDB_cursor *cursor, int modify,
375 MDB_pageparent *mpp);
377 static int mdb_env_read_header(MDB_env *env, MDB_meta *meta);
378 static int mdb_env_read_meta(MDB_env *env, int *which);
379 static int mdb_env_write_meta(MDB_txn *txn);
380 static MDB_page *mdb_get_page(MDB_txn *txn, pgno_t pgno);
382 static MDB_node *mdb_search_node(MDB_txn *txn, MDB_dbi dbi, MDB_page *mp,
383 MDB_val *key, int *exactp, unsigned int *kip);
384 static int mdb_add_node(MDB_txn *txn, MDB_dbi dbi, MDB_page *mp,
385 indx_t indx, MDB_val *key, MDB_val *data,
386 pgno_t pgno, uint8_t flags);
387 static void mdb_del_node(MDB_page *mp, indx_t indx, int ksize);
388 static int mdb_del0(MDB_txn *txn, MDB_dbi dbi, unsigned int ki,
389 MDB_pageparent *mpp, MDB_node *leaf);
390 static int mdb_put0(MDB_txn *txn, MDB_dbi dbi,
391 MDB_val *key, MDB_val *data, unsigned int flags);
392 static int mdb_read_data(MDB_txn *txn, MDB_node *leaf, MDB_val *data);
394 static int mdb_rebalance(MDB_txn *txn, MDB_dbi dbi, MDB_pageparent *mp);
395 static int mdb_update_key(MDB_page *mp, indx_t indx, MDB_val *key);
396 static int mdb_move_node(MDB_txn *txn, MDB_dbi dbi,
397 MDB_pageparent *src, indx_t srcindx,
398 MDB_pageparent *dst, indx_t dstindx);
399 static int mdb_merge(MDB_txn *txn, MDB_dbi dbi, MDB_pageparent *src,
400 MDB_pageparent *dst);
401 static int mdb_split(MDB_txn *txn, MDB_dbi dbi, MDB_page **mpp,
402 unsigned int *newindxp, MDB_val *newkey,
403 MDB_val *newdata, pgno_t newpgno);
404 static MDB_dpage *mdb_new_page(MDB_txn *txn, MDB_dbi dbi, uint32_t flags, int num);
406 static void cursor_pop_page(MDB_cursor *cursor);
407 static MDB_ppage *cursor_push_page(MDB_cursor *cursor,
410 static int mdb_sibling(MDB_cursor *cursor, int move_right);
411 static int mdb_cursor_next(MDB_cursor *cursor,
412 MDB_val *key, MDB_val *data, MDB_cursor_op op);
413 static int mdb_cursor_prev(MDB_cursor *cursor,
414 MDB_val *key, MDB_val *data, MDB_cursor_op op);
415 static int mdb_cursor_set(MDB_cursor *cursor,
416 MDB_val *key, MDB_val *data, MDB_cursor_op op, int *exactp);
417 static int mdb_cursor_first(MDB_cursor *cursor,
418 MDB_val *key, MDB_val *data);
419 static int mdb_cursor_last(MDB_cursor *cursor,
420 MDB_val *key, MDB_val *data);
422 static void mdb_xcursor_init0(MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx);
423 static void mdb_xcursor_init1(MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx, MDB_node *node);
424 static void mdb_xcursor_fini(MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx);
426 static size_t mdb_leaf_size(MDB_env *env, MDB_val *key,
428 static size_t mdb_branch_size(MDB_env *env, MDB_val *key);
430 static int memncmp(const void *s1, size_t n1,
431 const void *s2, size_t n2);
432 static int memnrcmp(const void *s1, size_t n1,
433 const void *s2, size_t n2);
436 memncmp(const void *s1, size_t n1, const void *s2, size_t n2)
438 int diff, len_diff = -1;
441 len_diff = (n1 > n2);
444 diff = memcmp(s1, s2, n1);
445 return diff ? diff : len_diff;
449 memnrcmp(const void *s1, size_t n1, const void *s2, size_t n2)
451 const unsigned char *p1, *p2, *p1_lim;
458 p1 = (const unsigned char *)s1 + n1 - 1;
459 p2 = (const unsigned char *)s2 + n2 - 1;
461 for (p1_lim = (n1 <= n2 ? s1 : s2); *p1 == *p2; p1--, p2--) {
463 return (p1 != s1) ? (p1 != p2) : (p2 != s2) ? -1 : 0;
469 mdb_version(int *maj, int *min, int *pat)
471 *maj = MDB_VERSION_MAJOR;
472 *min = MDB_VERSION_MINOR;
473 *pat = MDB_VERSION_PATCH;
474 return MDB_VERSION_STRING;
477 static char *const errstr[] = {
478 "MDB_KEYEXIST: Key/data pair already exists",
479 "MDB_NOTFOUND: No matching key/data pair found",
480 "MDB_PAGE_NOTFOUND: Requested page not found",
481 "MDB_CORRUPTED: Located page was wrong type",
482 "MDB_PANIC: Update of meta page failed",
483 "MDB_VERSION_MISMATCH: Database environment version mismatch"
487 mdb_strerror(int err)
490 return ("Successful return: 0");
492 if (err >= MDB_KEYEXIST && err <= MDB_VERSION_MISMATCH)
493 return errstr[err - MDB_KEYEXIST];
495 return strerror(err);
499 mdb_dkey(MDB_val *key, char *buf)
502 unsigned char *c = key->mv_data;
504 if (key->mv_size > MAXKEYSIZE)
506 for (i=0; i<key->mv_size; i++)
507 ptr += sprintf(ptr, "%02x", *c++);
512 mdb_cmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
514 if (txn->mt_dbxs[dbi].md_cmp)
515 return txn->mt_dbxs[dbi].md_cmp(a, b);
517 if (txn->mt_dbs[dbi].md_flags & (MDB_REVERSEKEY
518 #if __BYTE_ORDER == __LITTLE_ENDIAN
522 return memnrcmp(a->mv_data, a->mv_size, b->mv_data, b->mv_size);
524 return memncmp((char *)a->mv_data, a->mv_size, b->mv_data, b->mv_size);
528 mdb_dcmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
530 if (txn->mt_dbxs[dbi].md_dcmp)
531 return txn->mt_dbxs[dbi].md_dcmp(a, b);
533 if (txn->mt_dbs[dbi].md_flags & (0
534 #if __BYTE_ORDER == __LITTLE_ENDIAN
538 return memnrcmp(a->mv_data, a->mv_size, b->mv_data, b->mv_size);
540 return memncmp((char *)a->mv_data, a->mv_size, b->mv_data, b->mv_size);
543 /* Allocate new page(s) for writing */
545 mdb_alloc_page(MDB_txn *txn, MDB_page *parent, unsigned int parent_idx, int num)
548 pgno_t pgno = P_INVALID;
552 if (txn->mt_txnid > 2) {
554 oldest = txn->mt_txnid - 2;
555 if (!txn->mt_env->me_pghead && txn->mt_dbs[FREE_DBI].md_root != P_INVALID) {
556 /* See if there's anything in the free DB */
561 mpp.mp_parent = NULL;
563 mdb_search_page(txn, FREE_DBI, NULL, NULL, 0, &mpp);
564 leaf = NODEPTR(mpp.mp_page, 0);
565 kptr = (ULONG *)NODEKEY(leaf);
567 /* It's potentially usable, unless there are still
568 * older readers outstanding. Grab it.
570 if (oldest > *kptr) {
575 mdb_read_data(txn, leaf, &data);
576 idl = (ULONG *)data.mv_data;
577 mop = malloc(sizeof(MDB_oldpages) + MDB_IDL_SIZEOF(idl) - sizeof(pgno_t));
578 mop->mo_next = txn->mt_env->me_pghead;
579 mop->mo_txnid = *kptr;
580 txn->mt_env->me_pghead = mop;
581 memcpy(mop->mo_pages, idl, MDB_IDL_SIZEOF(idl));
586 DPRINTF("IDL read txn %lu root %lu num %lu",
587 mop->mo_txnid, txn->mt_dbs[FREE_DBI].md_root, idl[0]);
588 for (i=0; i<idl[0]; i++) {
589 DPRINTF("IDL %lu", idl[i+1]);
593 /* drop this IDL from the DB */
594 mpp.mp_parent = NULL;
596 mdb_search_page(txn, FREE_DBI, NULL, NULL, 1, &mpp);
597 leaf = NODEPTR(mpp.mp_page, 0);
598 mdb_del0(txn, FREE_DBI, 0, &mpp, leaf);
601 if (txn->mt_env->me_pghead) {
603 for (i=0; i<txn->mt_env->me_txns->mti_numreaders; i++) {
604 ULONG mr = txn->mt_env->me_txns->mti_readers[i].mr_txnid;
607 oldest = txn->mt_env->me_txns->mti_readers[i].mr_txnid;
609 if (oldest > txn->mt_env->me_pghead->mo_txnid) {
610 MDB_oldpages *mop = txn->mt_env->me_pghead;
611 txn->mt_oldest = oldest;
613 /* FIXME: For now, always use fresh pages. We
614 * really ought to search the free list for a
619 /* peel pages off tail, so we only have to truncate the list */
620 pgno = MDB_IDL_LAST(mop->mo_pages);
621 if (MDB_IDL_IS_RANGE(mop->mo_pages)) {
623 if (mop->mo_pages[2] > mop->mo_pages[1])
624 mop->mo_pages[0] = 0;
628 if (MDB_IDL_IS_ZERO(mop->mo_pages)) {
629 txn->mt_env->me_pghead = mop->mo_next;
637 if (pgno == P_INVALID) {
638 /* DB size is maxed out */
639 if (txn->mt_next_pgno + num >= txn->mt_env->me_maxpg)
642 if (txn->mt_env->me_dpages && num == 1) {
643 dp = txn->mt_env->me_dpages;
644 txn->mt_env->me_dpages = (MDB_dpage *)dp->h.md_parent;
646 if ((dp = malloc(txn->mt_env->me_psize * num + sizeof(MDB_dhead))) == NULL)
650 dp->h.md_parent = parent;
651 dp->h.md_pi = parent_idx;
652 if (pgno == P_INVALID) {
653 dp->p.mp_pgno = txn->mt_next_pgno;
654 txn->mt_next_pgno += num;
656 dp->p.mp_pgno = pgno;
658 mid.mid = dp->p.mp_pgno;
660 mdb_midl2_insert(txn->mt_u.dirty_list, &mid);
665 /* Touch a page: make it dirty and re-insert into tree with updated pgno.
668 mdb_touch(MDB_txn *txn, MDB_pageparent *pp)
670 MDB_page *mp = pp->mp_page;
675 if (!F_ISSET(mp->mp_flags, P_DIRTY)) {
677 if ((dp = mdb_alloc_page(txn, pp->mp_parent, pp->mp_pi, 1)) == NULL)
679 DPRINTF("touched page %lu -> %lu", mp->mp_pgno, dp->p.mp_pgno);
680 mdb_midl_insert(txn->mt_free_pgs, mp->mp_pgno);
681 pgno = dp->p.mp_pgno;
682 memcpy(&dp->p, mp, txn->mt_env->me_psize);
685 mp->mp_flags |= P_DIRTY;
687 /* Update the page number to new touched page. */
688 if (pp->mp_parent != NULL)
689 NODEPGNO(NODEPTR(pp->mp_parent, pp->mp_pi)) = mp->mp_pgno;
696 mdb_env_sync(MDB_env *env, int force)
699 if (force || !F_ISSET(env->me_flags, MDB_NOSYNC)) {
700 if (fdatasync(env->me_fd))
707 mdb_txn_begin(MDB_env *env, int rdonly, MDB_txn **ret)
712 if (env->me_flags & MDB_FATAL_ERROR) {
713 DPUTS("mdb_txn_begin: environment had fatal error, must shutdown!");
716 if ((txn = calloc(1, sizeof(MDB_txn))) == NULL) {
717 DPRINTF("calloc: %s", strerror(errno));
722 txn->mt_flags |= MDB_TXN_RDONLY;
724 txn->mt_u.dirty_list = env->me_dirty_list;
725 txn->mt_u.dirty_list[0].mid = 0;
726 txn->mt_free_pgs = env->me_free_pgs;
727 txn->mt_free_pgs[0] = 0;
729 pthread_mutex_lock(&env->me_txns->mti_wmutex);
730 env->me_txns->mti_txnid++;
733 txn->mt_txnid = env->me_txns->mti_txnid;
735 MDB_reader *r = pthread_getspecific(env->me_txkey);
738 pthread_mutex_lock(&env->me_txns->mti_mutex);
739 for (i=0; i<env->me_txns->mti_numreaders; i++)
740 if (env->me_txns->mti_readers[i].mr_pid == 0)
742 if (i == env->me_maxreaders) {
743 pthread_mutex_unlock(&env->me_txns->mti_mutex);
746 env->me_txns->mti_readers[i].mr_pid = getpid();
747 env->me_txns->mti_readers[i].mr_tid = pthread_self();
748 r = &env->me_txns->mti_readers[i];
749 pthread_setspecific(env->me_txkey, r);
750 if (i >= env->me_txns->mti_numreaders)
751 env->me_txns->mti_numreaders = i+1;
752 pthread_mutex_unlock(&env->me_txns->mti_mutex);
754 r->mr_txnid = txn->mt_txnid;
755 txn->mt_u.reader = r;
762 toggle = env->me_txns->mti_me_toggle;
763 if ((rc = mdb_env_read_meta(env, &toggle)) != MDB_SUCCESS) {
768 /* Copy the DB arrays */
769 txn->mt_numdbs = env->me_numdbs;
770 txn->mt_dbxs = env->me_dbxs; /* mostly static anyway */
771 txn->mt_dbs = malloc(env->me_maxdbs * sizeof(MDB_db));
772 memcpy(txn->mt_dbs, env->me_meta->mm_dbs, 2 * sizeof(MDB_db));
773 if (txn->mt_numdbs > 2)
774 memcpy(txn->mt_dbs+2, env->me_dbs[env->me_db_toggle]+2,
775 (txn->mt_numdbs - 2) * sizeof(MDB_db));
779 txn->mt_flags |= MDB_TXN_METOGGLE;
780 txn->mt_next_pgno = env->me_meta->mm_last_pg+1;
783 DPRINTF("begin transaction %lu on mdbenv %p, root page %lu",
784 txn->mt_txnid, (void *) env, txn->mt_dbs[MAIN_DBI].md_root);
791 mdb_txn_abort(MDB_txn *txn)
799 DPRINTF("abort transaction %lu on mdbenv %p, root page %lu",
800 txn->mt_txnid, (void *) env, txn->mt_dbs[MAIN_DBI].md_root);
804 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
805 txn->mt_u.reader->mr_txnid = 0;
811 /* return all dirty pages to dpage list */
812 for (i=1; i<=txn->mt_u.dirty_list[0].mid; i++) {
813 dp = txn->mt_u.dirty_list[i].mptr;
814 if (dp->h.md_num == 1) {
815 dp->h.md_parent = (MDB_page *)txn->mt_env->me_dpages;
816 txn->mt_env->me_dpages = dp;
818 /* large pages just get freed directly */
823 while ((mop = txn->mt_env->me_pghead)) {
824 txn->mt_env->me_pghead = mop->mo_next;
829 env->me_txns->mti_txnid--;
830 for (i=2; i<env->me_numdbs; i++)
831 env->me_dbxs[i].md_dirty = 0;
832 pthread_mutex_unlock(&env->me_txns->mti_wmutex);
839 mdb_txn_commit(MDB_txn *txn)
848 struct iovec iov[MDB_COMMIT_PAGES];
851 assert(txn->mt_env != NULL);
855 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
860 if (txn != env->me_txn) {
861 DPUTS("attempt to commit unknown transaction");
866 if (F_ISSET(txn->mt_flags, MDB_TXN_ERROR)) {
867 DPUTS("error flag is set, can't commit");
872 if (!txn->mt_u.dirty_list[0].mid)
875 DPRINTF("committing transaction %lu on mdbenv %p, root page %lu",
876 txn->mt_txnid, (void *) env, txn->mt_dbs[MAIN_DBI].md_root);
878 /* should only be one record now */
879 if (env->me_pghead) {
883 mop = env->me_pghead;
884 key.mv_size = sizeof(pgno_t);
885 key.mv_data = (char *)&mop->mo_txnid;
886 data.mv_size = MDB_IDL_SIZEOF(mop->mo_pages);
887 data.mv_data = mop->mo_pages;
888 mdb_put0(txn, FREE_DBI, &key, &data, 0);
889 free(env->me_pghead);
890 env->me_pghead = NULL;
892 /* save to free list */
893 if (!MDB_IDL_IS_ZERO(txn->mt_free_pgs)) {
897 /* make sure last page of freeDB is touched and on freelist */
898 key.mv_size = MAXKEYSIZE+1;
900 mpp.mp_parent = NULL;
902 mdb_search_page(txn, FREE_DBI, &key, NULL, 1, &mpp);
907 ULONG *idl = txn->mt_free_pgs;
908 DPRINTF("IDL write txn %lu root %lu num %lu",
909 txn->mt_txnid, txn->mt_dbs[FREE_DBI].md_root, idl[0]);
910 for (i=0; i<idl[0]; i++) {
911 DPRINTF("IDL %lu", idl[i+1]);
915 /* write to last page of freeDB */
916 key.mv_size = sizeof(pgno_t);
917 key.mv_data = (char *)&txn->mt_txnid;
918 data.mv_size = MDB_IDL_SIZEOF(txn->mt_free_pgs);
919 data.mv_data = txn->mt_free_pgs;
920 mdb_put0(txn, FREE_DBI, &key, &data, 0);
923 /* Update DB root pointers. Their pages have already been
924 * touched so this is all in-place and cannot fail.
928 data.mv_size = sizeof(MDB_db);
930 for (i = 2; i < txn->mt_numdbs; i++) {
931 if (txn->mt_dbxs[i].md_dirty) {
932 data.mv_data = &txn->mt_dbs[i];
933 mdb_put0(txn, MAIN_DBI, &txn->mt_dbxs[i].md_name, &data, 0);
938 /* Commit up to MDB_COMMIT_PAGES dirty pages to disk until done.
946 for (; i<=txn->mt_u.dirty_list[0].mid; i++) {
947 dp = txn->mt_u.dirty_list[i].mptr;
948 if (dp->p.mp_pgno != next) {
950 DPRINTF("committing %u dirty pages", n);
951 rc = writev(env->me_fd, iov, n);
955 DPUTS("short write, filesystem full?");
957 DPRINTF("writev: %s", strerror(errno));
964 lseek(env->me_fd, dp->p.mp_pgno * env->me_psize, SEEK_SET);
965 next = dp->p.mp_pgno;
967 DPRINTF("committing page %lu", dp->p.mp_pgno);
968 iov[n].iov_len = env->me_psize * dp->h.md_num;
969 iov[n].iov_base = &dp->p;
970 size += iov[n].iov_len;
971 next = dp->p.mp_pgno + dp->h.md_num;
972 /* clear dirty flag */
973 dp->p.mp_flags &= ~P_DIRTY;
974 if (++n >= MDB_COMMIT_PAGES) {
983 DPRINTF("committing %u dirty pages", n);
984 rc = writev(env->me_fd, iov, n);
988 DPUTS("short write, filesystem full?");
990 DPRINTF("writev: %s", strerror(errno));
997 /* Drop the dirty pages.
999 for (i=1; i<=txn->mt_u.dirty_list[0].mid; i++) {
1000 dp = txn->mt_u.dirty_list[i].mptr;
1001 if (dp->h.md_num == 1) {
1002 dp->h.md_parent = (MDB_page *)txn->mt_env->me_dpages;
1003 txn->mt_env->me_dpages = dp;
1007 txn->mt_u.dirty_list[i].mid = 0;
1010 if ((n = mdb_env_sync(env, 0)) != 0 ||
1011 (n = mdb_env_write_meta(txn)) != MDB_SUCCESS) {
1018 /* update the DB tables */
1020 int toggle = !env->me_db_toggle;
1022 for (i = 2; i < env->me_numdbs; i++) {
1023 if (txn->mt_dbxs[i].md_dirty) {
1024 env->me_dbs[toggle][i] = txn->mt_dbs[i];
1025 txn->mt_dbxs[i].md_dirty = 0;
1028 for (i = env->me_numdbs; i < txn->mt_numdbs; i++) {
1029 txn->mt_dbxs[i].md_dirty = 0;
1030 env->me_dbxs[i] = txn->mt_dbxs[i];
1031 env->me_dbs[toggle][i] = txn->mt_dbs[i];
1033 env->me_db_toggle = toggle;
1034 env->me_numdbs = txn->mt_numdbs;
1039 pthread_mutex_unlock(&env->me_txns->mti_wmutex);
1046 mdb_env_read_header(MDB_env *env, MDB_meta *meta)
1048 char page[PAGESIZE];
1053 assert(env != NULL);
1055 /* We don't know the page size yet, so use a minimum value.
1058 if ((rc = pread(env->me_fd, page, PAGESIZE, 0)) == 0) {
1060 } else if (rc != PAGESIZE) {
1063 DPRINTF("read: %s", strerror(errno));
1067 p = (MDB_page *)page;
1069 if (!F_ISSET(p->mp_flags, P_META)) {
1070 DPRINTF("page %lu not a meta page", p->mp_pgno);
1075 if (m->mm_magic != MDB_MAGIC) {
1076 DPUTS("meta has invalid magic");
1080 if (m->mm_version != MDB_VERSION) {
1081 DPRINTF("database is version %u, expected version %u",
1082 m->mm_version, MDB_VERSION);
1083 return MDB_VERSION_MISMATCH;
1086 memcpy(meta, m, sizeof(*m));
1091 mdb_env_init_meta(MDB_env *env, MDB_meta *meta)
1098 DPUTS("writing new meta page");
1099 psize = sysconf(_SC_PAGE_SIZE);
1101 meta->mm_magic = MDB_MAGIC;
1102 meta->mm_version = MDB_VERSION;
1103 meta->mm_psize = psize;
1104 meta->mm_last_pg = 1;
1105 meta->mm_flags = env->me_flags & 0xffff;
1106 meta->mm_flags |= MDB_INTEGERKEY;
1107 meta->mm_dbs[0].md_root = P_INVALID;
1108 meta->mm_dbs[1].md_root = P_INVALID;
1110 p = calloc(2, psize);
1112 p->mp_flags = P_META;
1115 memcpy(m, meta, sizeof(*meta));
1117 q = (MDB_page *)((char *)p + psize);
1120 q->mp_flags = P_META;
1123 memcpy(m, meta, sizeof(*meta));
1125 rc = write(env->me_fd, p, psize * 2);
1127 return (rc == (int)psize * 2) ? MDB_SUCCESS : errno;
1131 mdb_env_write_meta(MDB_txn *txn)
1134 MDB_meta meta, metab;
1136 int rc, len, toggle;
1139 assert(txn != NULL);
1140 assert(txn->mt_env != NULL);
1142 toggle = !F_ISSET(txn->mt_flags, MDB_TXN_METOGGLE);
1143 DPRINTF("writing meta page %d for root page %lu",
1144 toggle, txn->mt_dbs[MAIN_DBI].md_root);
1148 metab.mm_txnid = env->me_metas[toggle]->mm_txnid;
1149 metab.mm_last_pg = env->me_metas[toggle]->mm_last_pg;
1151 ptr = (char *)&meta;
1152 off = offsetof(MDB_meta, mm_dbs[0].md_depth);
1153 len = sizeof(MDB_meta) - off;
1156 meta.mm_dbs[0] = txn->mt_dbs[0];
1157 meta.mm_dbs[1] = txn->mt_dbs[1];
1158 meta.mm_last_pg = txn->mt_next_pgno - 1;
1159 meta.mm_txnid = txn->mt_txnid;
1162 off += env->me_psize;
1165 /* Write to the SYNC fd */
1166 rc = pwrite(env->me_mfd, ptr, len, off);
1170 DPUTS("write failed, disk error?");
1171 /* On a failure, the pagecache still contains the new data.
1172 * Write some old data back, to prevent it from being used.
1173 * Use the non-SYNC fd; we know it will fail anyway.
1175 meta.mm_last_pg = metab.mm_last_pg;
1176 meta.mm_txnid = metab.mm_txnid;
1177 r2 = pwrite(env->me_fd, ptr, len, off);
1178 env->me_flags |= MDB_FATAL_ERROR;
1181 txn->mt_env->me_txns->mti_me_toggle = toggle;
1187 mdb_env_read_meta(MDB_env *env, int *which)
1191 assert(env != NULL);
1195 else if (env->me_metas[0]->mm_txnid < env->me_metas[1]->mm_txnid)
1198 if (env->me_meta != env->me_metas[toggle])
1199 env->me_meta = env->me_metas[toggle];
1201 DPRINTF("Using meta page %d", toggle);
1207 mdb_env_create(MDB_env **env)
1211 e = calloc(1, sizeof(MDB_env));
1212 if (!e) return ENOMEM;
1214 e->me_maxreaders = DEFAULT_READERS;
1224 mdb_env_set_mapsize(MDB_env *env, size_t size)
1228 env->me_mapsize = size;
1233 mdb_env_set_maxdbs(MDB_env *env, int dbs)
1235 env->me_maxdbs = dbs;
1240 mdb_env_set_maxreaders(MDB_env *env, int readers)
1242 env->me_maxreaders = readers;
1247 mdb_env_get_maxreaders(MDB_env *env, int *readers)
1249 if (!env || !readers)
1251 *readers = env->me_maxreaders;
1256 mdb_env_open2(MDB_env *env, unsigned int flags)
1262 env->me_flags = flags;
1264 memset(&meta, 0, sizeof(meta));
1266 if ((i = mdb_env_read_header(env, &meta)) != 0) {
1269 DPUTS("new mdbenv");
1273 if (!env->me_mapsize) {
1274 env->me_mapsize = newenv ? DEFAULT_MAPSIZE : meta.mm_mapsize;
1278 if (meta.mm_address && (flags & MDB_FIXEDMAP))
1280 env->me_map = mmap(meta.mm_address, env->me_mapsize, PROT_READ, i,
1282 if (env->me_map == MAP_FAILED)
1286 meta.mm_mapsize = env->me_mapsize;
1287 if (flags & MDB_FIXEDMAP)
1288 meta.mm_address = env->me_map;
1289 i = mdb_env_init_meta(env, &meta);
1290 if (i != MDB_SUCCESS) {
1291 munmap(env->me_map, env->me_mapsize);
1295 env->me_psize = meta.mm_psize;
1297 env->me_maxpg = env->me_mapsize / env->me_psize;
1299 p = (MDB_page *)env->me_map;
1300 env->me_metas[0] = METADATA(p);
1301 env->me_metas[1] = (MDB_meta *)((char *)env->me_metas[0] + meta.mm_psize);
1303 if ((i = mdb_env_read_meta(env, NULL)) != 0)
1306 DPRINTF("opened database version %u, pagesize %u",
1307 env->me_meta->mm_version, env->me_psize);
1308 DPRINTF("depth: %u", env->me_meta->mm_dbs[MAIN_DBI].md_depth);
1309 DPRINTF("entries: %lu", env->me_meta->mm_dbs[MAIN_DBI].md_entries);
1310 DPRINTF("branch pages: %lu", env->me_meta->mm_dbs[MAIN_DBI].md_branch_pages);
1311 DPRINTF("leaf pages: %lu", env->me_meta->mm_dbs[MAIN_DBI].md_leaf_pages);
1312 DPRINTF("overflow pages: %lu", env->me_meta->mm_dbs[MAIN_DBI].md_overflow_pages);
1313 DPRINTF("root: %lu", env->me_meta->mm_dbs[MAIN_DBI].md_root);
1319 mdb_env_reader_dest(void *ptr)
1321 MDB_reader *reader = ptr;
1323 reader->mr_txnid = 0;
1328 /* downgrade the exclusive lock on the region back to shared */
1330 mdb_env_share_locks(MDB_env *env)
1332 struct flock lock_info;
1334 env->me_txns->mti_txnid = env->me_meta->mm_txnid;
1335 if (env->me_metas[0]->mm_txnid < env->me_metas[1]->mm_txnid)
1336 env->me_txns->mti_me_toggle = 1;
1338 memset((void *)&lock_info, 0, sizeof(lock_info));
1339 lock_info.l_type = F_RDLCK;
1340 lock_info.l_whence = SEEK_SET;
1341 lock_info.l_start = 0;
1342 lock_info.l_len = 1;
1343 fcntl(env->me_lfd, F_SETLK, &lock_info);
1347 mdb_env_setup_locks(MDB_env *env, char *lpath, int mode, int *excl)
1351 struct flock lock_info;
1355 if ((env->me_lfd = open(lpath, O_RDWR|O_CREAT, mode)) == -1) {
1359 /* Try to get exclusive lock. If we succeed, then
1360 * nobody is using the lock region and we should initialize it.
1362 memset((void *)&lock_info, 0, sizeof(lock_info));
1363 lock_info.l_type = F_WRLCK;
1364 lock_info.l_whence = SEEK_SET;
1365 lock_info.l_start = 0;
1366 lock_info.l_len = 1;
1367 rc = fcntl(env->me_lfd, F_SETLK, &lock_info);
1371 lock_info.l_type = F_RDLCK;
1372 rc = fcntl(env->me_lfd, F_SETLK, &lock_info);
1378 size = lseek(env->me_lfd, 0, SEEK_END);
1379 rsize = (env->me_maxreaders-1) * sizeof(MDB_reader) + sizeof(MDB_txninfo);
1380 if (size < rsize && *excl) {
1381 if (ftruncate(env->me_lfd, rsize) != 0) {
1387 size = rsize - sizeof(MDB_txninfo);
1388 env->me_maxreaders = size/sizeof(MDB_reader) + 1;
1390 env->me_txns = mmap(0, rsize, PROT_READ|PROT_WRITE, MAP_SHARED,
1392 if (env->me_txns == MAP_FAILED) {
1397 pthread_mutexattr_t mattr;
1399 pthread_mutexattr_init(&mattr);
1400 rc = pthread_mutexattr_setpshared(&mattr, PTHREAD_PROCESS_SHARED);
1404 pthread_mutex_init(&env->me_txns->mti_mutex, &mattr);
1405 pthread_mutex_init(&env->me_txns->mti_wmutex, &mattr);
1406 env->me_txns->mti_version = MDB_VERSION;
1407 env->me_txns->mti_magic = MDB_MAGIC;
1408 env->me_txns->mti_txnid = 0;
1409 env->me_txns->mti_numreaders = 0;
1410 env->me_txns->mti_me_toggle = 0;
1413 if (env->me_txns->mti_magic != MDB_MAGIC) {
1414 DPUTS("lock region has invalid magic");
1418 if (env->me_txns->mti_version != MDB_VERSION) {
1419 DPRINTF("lock region is version %u, expected version %u",
1420 env->me_txns->mti_version, MDB_VERSION);
1421 rc = MDB_VERSION_MISMATCH;
1424 if (errno != EACCES && errno != EAGAIN) {
1438 #define LOCKNAME "/lock.mdb"
1439 #define DATANAME "/data.mdb"
1441 mdb_env_open(MDB_env *env, const char *path, unsigned int flags, mode_t mode)
1443 int oflags, rc, len, excl;
1444 char *lpath, *dpath;
1447 lpath = malloc(len + sizeof(LOCKNAME) + len + sizeof(DATANAME));
1450 dpath = lpath + len + sizeof(LOCKNAME);
1451 sprintf(lpath, "%s" LOCKNAME, path);
1452 sprintf(dpath, "%s" DATANAME, path);
1454 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
1458 if (F_ISSET(flags, MDB_RDONLY))
1461 oflags = O_RDWR | O_CREAT;
1463 if ((env->me_fd = open(dpath, oflags, mode)) == -1) {
1468 if ((rc = mdb_env_open2(env, flags)) == MDB_SUCCESS) {
1469 /* synchronous fd for meta writes */
1470 if (!(flags & (MDB_RDONLY|MDB_NOSYNC)))
1471 oflags |= MDB_DSYNC;
1472 if ((env->me_mfd = open(dpath, oflags, mode)) == -1) {
1477 env->me_path = strdup(path);
1478 DPRINTF("opened dbenv %p", (void *) env);
1479 pthread_key_create(&env->me_txkey, mdb_env_reader_dest);
1481 mdb_env_share_locks(env);
1482 env->me_dbxs = calloc(env->me_maxdbs, sizeof(MDB_dbx));
1483 env->me_dbs[0] = calloc(env->me_maxdbs, sizeof(MDB_db));
1484 env->me_dbs[1] = calloc(env->me_maxdbs, sizeof(MDB_db));
1490 if (env->me_fd >= 0) {
1494 if (env->me_lfd >= 0) {
1504 mdb_env_close(MDB_env *env)
1511 while (env->me_dpages) {
1512 dp = env->me_dpages;
1513 env->me_dpages = (MDB_dpage *)dp->h.md_parent;
1517 free(env->me_dbs[1]);
1518 free(env->me_dbs[0]);
1522 pthread_key_delete(env->me_txkey);
1525 munmap(env->me_map, env->me_mapsize);
1530 pid_t pid = getpid();
1531 size_t size = (env->me_maxreaders-1) * sizeof(MDB_reader) + sizeof(MDB_txninfo);
1533 for (i=0; i<env->me_txns->mti_numreaders; i++)
1534 if (env->me_txns->mti_readers[i].mr_pid == pid)
1535 env->me_txns->mti_readers[i].mr_pid = 0;
1536 munmap(env->me_txns, size);
1542 /* Search for key within a leaf page, using binary search.
1543 * Returns the smallest entry larger or equal to the key.
1544 * If exactp is non-null, stores whether the found entry was an exact match
1545 * in *exactp (1 or 0).
1546 * If kip is non-null, stores the index of the found entry in *kip.
1547 * If no entry larger or equal to the key is found, returns NULL.
1550 mdb_search_node(MDB_txn *txn, MDB_dbi dbi, MDB_page *mp, MDB_val *key,
1551 int *exactp, unsigned int *kip)
1556 MDB_node *node = NULL;
1560 DPRINTF("searching %u keys in %s page %lu",
1562 IS_LEAF(mp) ? "leaf" : "branch",
1565 assert(NUMKEYS(mp) > 0);
1567 memset(&nodekey, 0, sizeof(nodekey));
1569 low = IS_LEAF(mp) ? 0 : 1;
1570 high = NUMKEYS(mp) - 1;
1571 while (low <= high) {
1572 i = (low + high) >> 1;
1575 nodekey.mv_size = txn->mt_dbs[dbi].md_pad;
1576 nodekey.mv_data = LEAF2KEY(mp, i, nodekey.mv_size);
1578 node = NODEPTR(mp, i);
1580 nodekey.mv_size = node->mn_ksize;
1581 nodekey.mv_data = NODEKEY(node);
1584 rc = mdb_cmp(txn, dbi, key, &nodekey);
1587 DPRINTF("found leaf index %u [%s], rc = %i",
1588 i, DKEY(&nodekey), rc);
1590 DPRINTF("found branch index %u [%s -> %lu], rc = %i",
1591 i, DKEY(&nodekey), NODEPGNO(node), rc);
1601 if (rc > 0) { /* Found entry is less than the key. */
1602 i++; /* Skip to get the smallest entry larger than key. */
1603 if (i >= NUMKEYS(mp))
1604 /* There is no entry larger or equal to the key. */
1608 *exactp = (rc == 0);
1609 if (kip) /* Store the key index if requested. */
1612 /* nodeptr is fake for LEAF2 */
1613 return IS_LEAF2(mp) ? NODEPTR(mp, 0) : NODEPTR(mp, i);
1617 cursor_pop_page(MDB_cursor *cursor)
1621 if (cursor->mc_snum) {
1622 top = CURSOR_TOP(cursor);
1625 DPRINTF("popped page %lu off db %u cursor %p", top->mp_page->mp_pgno,
1626 cursor->mc_dbi, (void *) cursor);
1631 cursor_push_page(MDB_cursor *cursor, MDB_page *mp)
1635 DPRINTF("pushing page %lu on db %u cursor %p", mp->mp_pgno,
1636 cursor->mc_dbi, (void *) cursor);
1638 ppage = &cursor->mc_stack[cursor->mc_snum++];
1639 ppage->mp_page = mp;
1645 mdb_get_page(MDB_txn *txn, pgno_t pgno)
1650 if (!F_ISSET(txn->mt_flags, MDB_TXN_RDONLY) && txn->mt_u.dirty_list[0].mid) {
1655 x = mdb_midl2_search(txn->mt_u.dirty_list, &id);
1656 if (x <= txn->mt_u.dirty_list[0].mid && txn->mt_u.dirty_list[x].mid == pgno) {
1657 dp = txn->mt_u.dirty_list[x].mptr;
1663 if (pgno > txn->mt_env->me_meta->mm_last_pg)
1665 p = (MDB_page *)(txn->mt_env->me_map + txn->mt_env->me_psize * pgno);
1671 mdb_search_page_root(MDB_txn *txn, MDB_dbi dbi, MDB_val *key,
1672 MDB_cursor *cursor, int modify, MDB_pageparent *mpp)
1674 MDB_page *mp = mpp->mp_page;
1678 if (cursor && cursor_push_page(cursor, mp) == NULL)
1681 while (IS_BRANCH(mp)) {
1685 DPRINTF("branch page %lu has %u keys", mp->mp_pgno, NUMKEYS(mp));
1686 assert(NUMKEYS(mp) > 1);
1687 DPRINTF("found index 0 to page %lu", NODEPGNO(NODEPTR(mp, 0)));
1689 if (key == NULL) /* Initialize cursor to first page. */
1691 else if (key->mv_size > MAXKEYSIZE && key->mv_data == NULL) {
1692 /* cursor to last page */
1696 node = mdb_search_node(txn, dbi, mp, key, &exact, &i);
1698 i = NUMKEYS(mp) - 1;
1706 DPRINTF("following index %u for key [%s]",
1708 assert(i < NUMKEYS(mp));
1709 node = NODEPTR(mp, i);
1712 CURSOR_TOP(cursor)->mp_ki = i;
1714 mpp->mp_parent = mp;
1715 if ((mp = mdb_get_page(txn, NODEPGNO(node))) == NULL)
1716 return MDB_PAGE_NOTFOUND;
1720 if (cursor && cursor_push_page(cursor, mp) == NULL)
1724 MDB_dhead *dh = ((MDB_dhead *)mp)-1;
1725 if ((rc = mdb_touch(txn, mpp)) != 0)
1727 dh = ((MDB_dhead *)mpp->mp_page)-1;
1728 dh->md_parent = mpp->mp_parent;
1729 dh->md_pi = mpp->mp_pi;
1736 DPRINTF("internal error, index points to a %02X page!?",
1738 return MDB_CORRUPTED;
1741 DPRINTF("found leaf page %lu for key [%s]", mp->mp_pgno,
1742 key ? DKEY(key) : NULL);
1747 /* Search for the page a given key should be in.
1748 * Stores a pointer to the found page in *mpp.
1749 * If key is NULL, search for the lowest page (used by mdb_cursor_first).
1750 * If cursor is non-null, pushes parent pages on the cursor stack.
1751 * If modify is true, visited pages are updated with new page numbers.
1754 mdb_search_page(MDB_txn *txn, MDB_dbi dbi, MDB_val *key,
1755 MDB_cursor *cursor, int modify, MDB_pageparent *mpp)
1760 /* Choose which root page to start with. If a transaction is given
1761 * use the root page from the transaction, otherwise read the last
1762 * committed root page.
1764 if (F_ISSET(txn->mt_flags, MDB_TXN_ERROR)) {
1765 DPUTS("transaction has failed, must abort");
1768 root = txn->mt_dbs[dbi].md_root;
1770 if (root == P_INVALID) { /* Tree is empty. */
1771 DPUTS("tree is empty");
1772 return MDB_NOTFOUND;
1775 if ((mpp->mp_page = mdb_get_page(txn, root)) == NULL)
1776 return MDB_PAGE_NOTFOUND;
1778 DPRINTF("root page has flags 0x%X", mpp->mp_page->mp_flags);
1781 /* For sub-databases, update main root first */
1782 if (dbi > MAIN_DBI && !txn->mt_dbxs[dbi].md_dirty) {
1784 rc = mdb_search_page(txn, MAIN_DBI, &txn->mt_dbxs[dbi].md_name,
1788 txn->mt_dbxs[dbi].md_dirty = 1;
1790 if (!F_ISSET(mpp->mp_page->mp_flags, P_DIRTY)) {
1791 mpp->mp_parent = NULL;
1793 if ((rc = mdb_touch(txn, mpp)))
1795 txn->mt_dbs[dbi].md_root = mpp->mp_page->mp_pgno;
1799 return mdb_search_page_root(txn, dbi, key, cursor, modify, mpp);
1803 mdb_read_data(MDB_txn *txn, MDB_node *leaf, MDB_val *data)
1805 MDB_page *omp; /* overflow mpage */
1808 if (!F_ISSET(leaf->mn_flags, F_BIGDATA)) {
1809 data->mv_size = leaf->mn_dsize;
1810 data->mv_data = NODEDATA(leaf);
1814 /* Read overflow data.
1816 data->mv_size = leaf->mn_dsize;
1817 memcpy(&pgno, NODEDATA(leaf), sizeof(pgno));
1818 if ((omp = mdb_get_page(txn, pgno)) == NULL) {
1819 DPRINTF("read overflow page %lu failed", pgno);
1820 return MDB_PAGE_NOTFOUND;
1822 data->mv_data = METADATA(omp);
1828 mdb_get(MDB_txn *txn, MDB_dbi dbi,
1829 MDB_val *key, MDB_val *data)
1838 DPRINTF("===> get db %u key [%s]", dbi, DKEY(key));
1840 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
1843 if (key->mv_size == 0 || key->mv_size > MAXKEYSIZE) {
1847 if ((rc = mdb_search_page(txn, dbi, key, NULL, 0, &mpp)) != MDB_SUCCESS)
1850 leaf = mdb_search_node(txn, dbi, mpp.mp_page, key, &exact, NULL);
1851 if (leaf && exact) {
1852 /* Return first duplicate data item */
1853 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
1856 mdb_xcursor_init0(txn, dbi, &mx);
1857 mdb_xcursor_init1(txn, dbi, &mx, leaf);
1858 rc = mdb_search_page(&mx.mx_txn, mx.mx_cursor.mc_dbi, NULL, NULL, 0, &mpp);
1859 if (rc != MDB_SUCCESS)
1861 if (IS_LEAF2(mpp.mp_page)) {
1862 data->mv_size = txn->mt_dbs[dbi].md_pad;
1863 data->mv_data = LEAF2KEY(mpp.mp_page, 0, data->mv_size);
1865 leaf = NODEPTR(mpp.mp_page, 0);
1866 data->mv_size = NODEKSZ(leaf);
1867 data->mv_data = NODEKEY(leaf);
1870 rc = mdb_read_data(txn, leaf, data);
1880 mdb_sibling(MDB_cursor *cursor, int move_right)
1887 if (cursor->mc_snum < 2) {
1888 return MDB_NOTFOUND; /* root has no siblings */
1890 parent = CURSOR_PARENT(cursor);
1892 DPRINTF("parent page is page %lu, index %u",
1893 parent->mp_page->mp_pgno, parent->mp_ki);
1895 cursor_pop_page(cursor);
1896 if (move_right ? (parent->mp_ki + 1 >= NUMKEYS(parent->mp_page))
1897 : (parent->mp_ki == 0)) {
1898 DPRINTF("no more keys left, moving to %s sibling",
1899 move_right ? "right" : "left");
1900 if ((rc = mdb_sibling(cursor, move_right)) != MDB_SUCCESS)
1902 parent = CURSOR_TOP(cursor);
1908 DPRINTF("just moving to %s index key %u",
1909 move_right ? "right" : "left", parent->mp_ki);
1911 assert(IS_BRANCH(parent->mp_page));
1913 indx = NODEPTR(parent->mp_page, parent->mp_ki);
1914 if ((mp = mdb_get_page(cursor->mc_txn, NODEPGNO(indx))) == NULL)
1915 return MDB_PAGE_NOTFOUND;
1917 mp->parent = parent->mp_page;
1918 mp->parent_index = parent->mp_ki;
1921 cursor_push_page(cursor, mp);
1927 mdb_cursor_next(MDB_cursor *cursor, MDB_val *key, MDB_val *data, MDB_cursor_op op)
1934 if (cursor->mc_eof) {
1935 return MDB_NOTFOUND;
1938 assert(cursor->mc_initialized);
1940 top = CURSOR_TOP(cursor);
1943 if (cursor->mc_txn->mt_dbs[cursor->mc_dbi].md_flags & MDB_DUPSORT) {
1944 leaf = NODEPTR(mp, top->mp_ki);
1945 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
1946 if (op == MDB_NEXT || op == MDB_NEXT_DUP) {
1947 rc = mdb_cursor_next(&cursor->mc_xcursor->mx_cursor, data, NULL, MDB_NEXT);
1948 if (op != MDB_NEXT || rc == MDB_SUCCESS)
1952 cursor->mc_xcursor->mx_cursor.mc_initialized = 0;
1953 if (op == MDB_NEXT_DUP)
1954 return MDB_NOTFOUND;
1958 DPRINTF("cursor_next: top page is %lu in cursor %p", mp->mp_pgno, (void *) cursor);
1960 if (top->mp_ki + 1 >= NUMKEYS(mp)) {
1961 DPUTS("=====> move to next sibling page");
1962 if (mdb_sibling(cursor, 1) != MDB_SUCCESS) {
1964 return MDB_NOTFOUND;
1966 top = CURSOR_TOP(cursor);
1968 DPRINTF("next page is %lu, key index %u", mp->mp_pgno, top->mp_ki);
1972 DPRINTF("==> cursor points to page %lu with %u keys, key index %u",
1973 mp->mp_pgno, NUMKEYS(mp), top->mp_ki);
1976 key->mv_size = cursor->mc_txn->mt_dbs[cursor->mc_dbi].md_pad;
1977 key->mv_data = LEAF2KEY(mp, top->mp_ki, key->mv_size);
1981 assert(IS_LEAF(mp));
1982 leaf = NODEPTR(mp, top->mp_ki);
1984 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
1985 mdb_xcursor_init1(cursor->mc_txn, cursor->mc_dbi, cursor->mc_xcursor, leaf);
1988 if ((rc = mdb_read_data(cursor->mc_txn, leaf, data) != MDB_SUCCESS))
1991 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
1992 rc = mdb_cursor_first(&cursor->mc_xcursor->mx_cursor, data, NULL);
1993 if (rc != MDB_SUCCESS)
1998 MDB_SET_KEY(leaf, key);
2003 mdb_cursor_prev(MDB_cursor *cursor, MDB_val *key, MDB_val *data, MDB_cursor_op op)
2010 assert(cursor->mc_initialized);
2012 top = CURSOR_TOP(cursor);
2015 if (cursor->mc_txn->mt_dbs[cursor->mc_dbi].md_flags & MDB_DUPSORT) {
2016 leaf = NODEPTR(mp, top->mp_ki);
2017 if (op == MDB_PREV || op == MDB_PREV_DUP) {
2018 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
2019 rc = mdb_cursor_prev(&cursor->mc_xcursor->mx_cursor, data, NULL, MDB_PREV);
2020 if (op != MDB_PREV || rc == MDB_SUCCESS)
2023 cursor->mc_xcursor->mx_cursor.mc_initialized = 0;
2024 if (op == MDB_PREV_DUP)
2025 return MDB_NOTFOUND;
2030 DPRINTF("cursor_prev: top page is %lu in cursor %p", mp->mp_pgno, (void *) cursor);
2032 if (top->mp_ki == 0) {
2033 DPUTS("=====> move to prev sibling page");
2034 if (mdb_sibling(cursor, 0) != MDB_SUCCESS) {
2035 cursor->mc_initialized = 0;
2036 return MDB_NOTFOUND;
2038 top = CURSOR_TOP(cursor);
2040 top->mp_ki = NUMKEYS(mp) - 1;
2041 DPRINTF("prev page is %lu, key index %u", mp->mp_pgno, top->mp_ki);
2047 DPRINTF("==> cursor points to page %lu with %u keys, key index %u",
2048 mp->mp_pgno, NUMKEYS(mp), top->mp_ki);
2051 key->mv_size = cursor->mc_txn->mt_dbs[cursor->mc_dbi].md_pad;
2052 key->mv_data = LEAF2KEY(mp, top->mp_ki, key->mv_size);
2056 assert(IS_LEAF(mp));
2057 leaf = NODEPTR(mp, top->mp_ki);
2059 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
2060 mdb_xcursor_init1(cursor->mc_txn, cursor->mc_dbi, cursor->mc_xcursor, leaf);
2063 if ((rc = mdb_read_data(cursor->mc_txn, leaf, data) != MDB_SUCCESS))
2066 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
2067 rc = mdb_cursor_last(&cursor->mc_xcursor->mx_cursor, data, NULL);
2068 if (rc != MDB_SUCCESS)
2073 MDB_SET_KEY(leaf, key);
2078 mdb_cursor_set(MDB_cursor *cursor, MDB_val *key, MDB_val *data,
2079 MDB_cursor_op op, int *exactp)
2089 assert(key->mv_size > 0);
2091 /* See if we're already on the right page */
2092 if (cursor->mc_initialized) {
2094 top = CURSOR_TOP(cursor);
2095 /* Don't try this for LEAF2 pages. Maybe support that later. */
2096 if ((top->mp_page->mp_flags & (P_LEAF|P_LEAF2)) == P_LEAF) {
2097 leaf = NODEPTR(top->mp_page, 0);
2098 MDB_SET_KEY(leaf, &nodekey);
2099 rc = mdb_cmp(cursor->mc_txn, cursor->mc_dbi, key, &nodekey);
2101 leaf = NODEPTR(top->mp_page, NUMKEYS(top->mp_page)-1);
2102 MDB_SET_KEY(leaf, &nodekey);
2103 rc = mdb_cmp(cursor->mc_txn, cursor->mc_dbi, key, &nodekey);
2105 /* we're already on the right page */
2106 mpp.mp_page = top->mp_page;
2113 cursor->mc_snum = 0;
2115 rc = mdb_search_page(cursor->mc_txn, cursor->mc_dbi, key, cursor, 0, &mpp);
2116 if (rc != MDB_SUCCESS)
2119 assert(IS_LEAF(mpp.mp_page));
2121 top = CURSOR_TOP(cursor);
2123 leaf = mdb_search_node(cursor->mc_txn, cursor->mc_dbi, mpp.mp_page, key, exactp, &top->mp_ki);
2124 if (exactp != NULL && !*exactp) {
2125 /* MDB_SET specified and not an exact match. */
2126 return MDB_NOTFOUND;
2130 DPUTS("===> inexact leaf not found, goto sibling");
2131 if ((rc = mdb_sibling(cursor, 1)) != MDB_SUCCESS)
2132 return rc; /* no entries matched */
2133 top = CURSOR_TOP(cursor);
2135 mpp.mp_page = top->mp_page;
2136 assert(IS_LEAF(mpp.mp_page));
2137 leaf = NODEPTR(mpp.mp_page, 0);
2140 cursor->mc_initialized = 1;
2143 if (IS_LEAF2(mpp.mp_page)) {
2144 key->mv_size = cursor->mc_txn->mt_dbs[cursor->mc_dbi].md_pad;
2145 key->mv_data = LEAF2KEY(mpp.mp_page, top->mp_ki, key->mv_size);
2149 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
2150 mdb_xcursor_init1(cursor->mc_txn, cursor->mc_dbi, cursor->mc_xcursor, leaf);
2153 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
2154 if (op == MDB_SET || op == MDB_SET_RANGE) {
2155 rc = mdb_cursor_first(&cursor->mc_xcursor->mx_cursor, data, NULL);
2158 if (op == MDB_GET_BOTH) {
2163 rc = mdb_cursor_set(&cursor->mc_xcursor->mx_cursor, data, NULL, MDB_SET_RANGE, ex2p);
2164 if (rc != MDB_SUCCESS)
2167 } else if (op == MDB_GET_BOTH || op == MDB_GET_BOTH_RANGE) {
2169 if ((rc = mdb_read_data(cursor->mc_txn, leaf, &d2)) != MDB_SUCCESS)
2171 rc = mdb_dcmp(cursor->mc_txn, cursor->mc_dbi, data, &d2);
2173 if (op == MDB_GET_BOTH || rc > 0)
2174 return MDB_NOTFOUND;
2178 if ((rc = mdb_read_data(cursor->mc_txn, leaf, data)) != MDB_SUCCESS)
2183 /* The key already matches in all other cases */
2184 if (op == MDB_SET_RANGE)
2185 MDB_SET_KEY(leaf, key);
2186 DPRINTF("==> cursor placed on key [%s]", DKEY(key));
2192 mdb_cursor_first(MDB_cursor *cursor, MDB_val *key, MDB_val *data)
2198 cursor->mc_snum = 0;
2200 rc = mdb_search_page(cursor->mc_txn, cursor->mc_dbi, NULL, cursor, 0, &mpp);
2201 if (rc != MDB_SUCCESS)
2203 assert(IS_LEAF(mpp.mp_page));
2205 leaf = NODEPTR(mpp.mp_page, 0);
2206 cursor->mc_initialized = 1;
2209 if (IS_LEAF2(mpp.mp_page)) {
2210 key->mv_size = cursor->mc_txn->mt_dbs[cursor->mc_dbi].md_pad;
2211 key->mv_data = LEAF2KEY(mpp.mp_page, 0, key->mv_size);
2216 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
2217 mdb_xcursor_init1(cursor->mc_txn, cursor->mc_dbi, cursor->mc_xcursor, leaf);
2218 rc = mdb_cursor_first(&cursor->mc_xcursor->mx_cursor, data, NULL);
2222 if (cursor->mc_xcursor)
2223 cursor->mc_xcursor->mx_cursor.mc_initialized = 0;
2224 if ((rc = mdb_read_data(cursor->mc_txn, leaf, data)) != MDB_SUCCESS)
2228 MDB_SET_KEY(leaf, key);
2233 mdb_cursor_last(MDB_cursor *cursor, MDB_val *key, MDB_val *data)
2241 cursor->mc_snum = 0;
2243 lkey.mv_size = MAXKEYSIZE+1;
2244 lkey.mv_data = NULL;
2246 rc = mdb_search_page(cursor->mc_txn, cursor->mc_dbi, &lkey, cursor, 0, &mpp);
2247 if (rc != MDB_SUCCESS)
2249 assert(IS_LEAF(mpp.mp_page));
2251 leaf = NODEPTR(mpp.mp_page, NUMKEYS(mpp.mp_page)-1);
2252 cursor->mc_initialized = 1;
2255 top = CURSOR_TOP(cursor);
2256 top->mp_ki = NUMKEYS(top->mp_page) - 1;
2258 if (IS_LEAF2(mpp.mp_page)) {
2259 key->mv_size = cursor->mc_txn->mt_dbs[cursor->mc_dbi].md_pad;
2260 key->mv_data = LEAF2KEY(mpp.mp_page, top->mp_ki, key->mv_size);
2265 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
2266 mdb_xcursor_init1(cursor->mc_txn, cursor->mc_dbi, cursor->mc_xcursor, leaf);
2267 rc = mdb_cursor_last(&cursor->mc_xcursor->mx_cursor, data, NULL);
2271 if ((rc = mdb_read_data(cursor->mc_txn, leaf, data)) != MDB_SUCCESS)
2276 MDB_SET_KEY(leaf, key);
2281 mdb_cursor_get(MDB_cursor *cursor, MDB_val *key, MDB_val *data,
2291 case MDB_GET_BOTH_RANGE:
2292 if (data == NULL || cursor->mc_xcursor == NULL) {
2299 if (key == NULL || key->mv_size == 0 || key->mv_size > MAXKEYSIZE) {
2301 } else if (op == MDB_SET_RANGE)
2302 rc = mdb_cursor_set(cursor, key, data, op, NULL);
2304 rc = mdb_cursor_set(cursor, key, data, op, &exact);
2306 case MDB_GET_MULTIPLE:
2308 !(cursor->mc_txn->mt_dbs[cursor->mc_dbi].md_flags & MDB_DUPFIXED) ||
2309 !cursor->mc_initialized) {
2314 if (!cursor->mc_xcursor->mx_cursor.mc_initialized || cursor->mc_xcursor->mx_cursor.mc_eof)
2317 case MDB_NEXT_MULTIPLE:
2319 !(cursor->mc_txn->mt_dbs[cursor->mc_dbi].md_flags & MDB_DUPFIXED)) {
2323 if (!cursor->mc_initialized)
2324 rc = mdb_cursor_first(cursor, key, data);
2326 rc = mdb_cursor_next(cursor, key, data, MDB_NEXT_DUP);
2327 if (rc == MDB_SUCCESS) {
2328 if (cursor->mc_xcursor->mx_cursor.mc_initialized) {
2331 top = CURSOR_TOP(&cursor->mc_xcursor->mx_cursor);
2332 data->mv_size = NUMKEYS(top->mp_page) *
2333 cursor->mc_xcursor->mx_txn.mt_dbs[cursor->mc_xcursor->mx_cursor.mc_dbi].md_pad;
2334 data->mv_data = METADATA(top->mp_page);
2335 top->mp_ki = NUMKEYS(top->mp_page)-1;
2343 case MDB_NEXT_NODUP:
2344 if (!cursor->mc_initialized)
2345 rc = mdb_cursor_first(cursor, key, data);
2347 rc = mdb_cursor_next(cursor, key, data, op);
2351 case MDB_PREV_NODUP:
2352 if (!cursor->mc_initialized || cursor->mc_eof)
2353 rc = mdb_cursor_last(cursor, key, data);
2355 rc = mdb_cursor_prev(cursor, key, data, op);
2358 rc = mdb_cursor_first(cursor, key, data);
2361 rc = mdb_cursor_last(cursor, key, data);
2364 DPRINTF("unhandled/unimplemented cursor operation %u", op);
2372 /* Allocate a page and initialize it
2375 mdb_new_page(MDB_txn *txn, MDB_dbi dbi, uint32_t flags, int num)
2379 if ((dp = mdb_alloc_page(txn, NULL, 0, num)) == NULL)
2381 DPRINTF("allocated new mpage %lu, page size %u",
2382 dp->p.mp_pgno, txn->mt_env->me_psize);
2383 dp->p.mp_flags = flags | P_DIRTY;
2384 dp->p.mp_lower = PAGEHDRSZ;
2385 dp->p.mp_upper = txn->mt_env->me_psize;
2387 if (IS_BRANCH(&dp->p))
2388 txn->mt_dbs[dbi].md_branch_pages++;
2389 else if (IS_LEAF(&dp->p))
2390 txn->mt_dbs[dbi].md_leaf_pages++;
2391 else if (IS_OVERFLOW(&dp->p)) {
2392 txn->mt_dbs[dbi].md_overflow_pages += num;
2393 dp->p.mp_pages = num;
2400 mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data)
2404 sz = LEAFSIZE(key, data);
2405 if (data->mv_size >= env->me_psize / MDB_MINKEYS) {
2406 /* put on overflow page */
2407 sz -= data->mv_size - sizeof(pgno_t);
2410 return sz + sizeof(indx_t);
2414 mdb_branch_size(MDB_env *env, MDB_val *key)
2419 if (sz >= env->me_psize / MDB_MINKEYS) {
2420 /* put on overflow page */
2421 /* not implemented */
2422 /* sz -= key->size - sizeof(pgno_t); */
2425 return sz + sizeof(indx_t);
2429 mdb_add_node(MDB_txn *txn, MDB_dbi dbi, MDB_page *mp, indx_t indx,
2430 MDB_val *key, MDB_val *data, pgno_t pgno, uint8_t flags)
2433 size_t node_size = NODESIZE;
2436 MDB_dpage *ofp = NULL; /* overflow page */
2439 assert(mp->mp_upper >= mp->mp_lower);
2441 DPRINTF("add node [%s] to %s page %lu at index %i, key size %zu",
2442 key ? DKEY(key) : NULL,
2443 IS_LEAF(mp) ? "leaf" : "branch",
2444 mp->mp_pgno, indx, key ? key->mv_size : 0);
2447 /* Move higher keys up one slot. */
2448 int ksize = txn->mt_dbs[dbi].md_pad, dif;
2449 char *ptr = LEAF2KEY(mp, indx, ksize);
2450 dif = NUMKEYS(mp) - indx;
2452 memmove(ptr+ksize, ptr, dif*ksize);
2453 /* insert new key */
2454 memcpy(ptr, key->mv_data, ksize);
2456 /* Just using these for counting */
2457 mp->mp_lower += sizeof(indx_t);
2458 mp->mp_upper -= ksize - sizeof(indx_t);
2463 node_size += key->mv_size;
2467 if (F_ISSET(flags, F_BIGDATA)) {
2468 /* Data already on overflow page. */
2469 node_size += sizeof(pgno_t);
2470 } else if (data->mv_size >= txn->mt_env->me_psize / MDB_MINKEYS) {
2471 int ovpages = OVPAGES(data->mv_size, txn->mt_env->me_psize);
2472 /* Put data on overflow page. */
2473 DPRINTF("data size is %zu, put on overflow page",
2475 node_size += sizeof(pgno_t);
2476 if ((ofp = mdb_new_page(txn, dbi, P_OVERFLOW, ovpages)) == NULL)
2478 DPRINTF("allocated overflow page %lu", ofp->p.mp_pgno);
2481 node_size += data->mv_size;
2485 if (node_size + sizeof(indx_t) > SIZELEFT(mp)) {
2486 DPRINTF("not enough room in page %lu, got %u ptrs",
2487 mp->mp_pgno, NUMKEYS(mp));
2488 DPRINTF("upper - lower = %u - %u = %u", mp->mp_upper, mp->mp_lower,
2489 mp->mp_upper - mp->mp_lower);
2490 DPRINTF("node size = %zu", node_size);
2494 /* Move higher pointers up one slot. */
2495 for (i = NUMKEYS(mp); i > indx; i--)
2496 mp->mp_ptrs[i] = mp->mp_ptrs[i - 1];
2498 /* Adjust free space offsets. */
2499 ofs = mp->mp_upper - node_size;
2500 assert(ofs >= mp->mp_lower + sizeof(indx_t));
2501 mp->mp_ptrs[indx] = ofs;
2503 mp->mp_lower += sizeof(indx_t);
2505 /* Write the node data. */
2506 node = NODEPTR(mp, indx);
2507 node->mn_ksize = (key == NULL) ? 0 : key->mv_size;
2508 node->mn_flags = flags;
2510 node->mn_dsize = data->mv_size;
2512 NODEPGNO(node) = pgno;
2515 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
2520 if (F_ISSET(flags, F_BIGDATA))
2521 memcpy(node->mn_data + key->mv_size, data->mv_data,
2524 memcpy(node->mn_data + key->mv_size, data->mv_data,
2527 memcpy(node->mn_data + key->mv_size, &ofp->p.mp_pgno,
2529 memcpy(METADATA(&ofp->p), data->mv_data, data->mv_size);
2537 mdb_del_node(MDB_page *mp, indx_t indx, int ksize)
2540 indx_t i, j, numkeys, ptr;
2544 DPRINTF("delete node %u on %s page %lu", indx,
2545 IS_LEAF(mp) ? "leaf" : "branch", mp->mp_pgno);
2546 assert(indx < NUMKEYS(mp));
2549 int x = NUMKEYS(mp) - 1 - indx;
2550 base = LEAF2KEY(mp, indx, ksize);
2552 memmove(base, base + ksize, x * ksize);
2553 mp->mp_lower -= sizeof(indx_t);
2554 mp->mp_upper += ksize - sizeof(indx_t);
2558 node = NODEPTR(mp, indx);
2559 sz = NODESIZE + node->mn_ksize;
2561 if (F_ISSET(node->mn_flags, F_BIGDATA))
2562 sz += sizeof(pgno_t);
2564 sz += NODEDSZ(node);
2567 ptr = mp->mp_ptrs[indx];
2568 numkeys = NUMKEYS(mp);
2569 for (i = j = 0; i < numkeys; i++) {
2571 mp->mp_ptrs[j] = mp->mp_ptrs[i];
2572 if (mp->mp_ptrs[i] < ptr)
2573 mp->mp_ptrs[j] += sz;
2578 base = (char *)mp + mp->mp_upper;
2579 memmove(base + sz, base, ptr - mp->mp_upper);
2581 mp->mp_lower -= sizeof(indx_t);
2586 mdb_xcursor_init0(MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx)
2591 mx->mx_txn.mt_dbxs = mx->mx_dbxs;
2592 mx->mx_txn.mt_dbs = mx->mx_dbs;
2593 mx->mx_dbxs[0] = txn->mt_dbxs[0];
2594 mx->mx_dbxs[1] = txn->mt_dbxs[1];
2596 mx->mx_dbxs[2] = txn->mt_dbxs[dbi];
2601 mx->mx_dbxs[dbn+1].md_parent = dbn;
2602 mx->mx_dbxs[dbn+1].md_cmp = mx->mx_dbxs[dbn].md_dcmp;
2603 mx->mx_dbxs[dbn+1].md_rel = mx->mx_dbxs[dbn].md_rel;
2604 mx->mx_dbxs[dbn+1].md_dirty = 0;
2605 mx->mx_txn.mt_numdbs = dbn+2;
2607 mx->mx_cursor.mc_snum = 0;
2608 mx->mx_cursor.mc_txn = &mx->mx_txn;
2609 mx->mx_cursor.mc_dbi = dbn+1;
2613 mdb_xcursor_init1(MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx, MDB_node *node)
2615 MDB_db *db = NODEDATA(node);
2617 mx->mx_dbs[0] = txn->mt_dbs[0];
2618 mx->mx_dbs[1] = txn->mt_dbs[1];
2620 mx->mx_dbs[2] = txn->mt_dbs[dbi];
2625 mx->mx_dbs[dbn] = *db;
2626 mx->mx_dbxs[dbn].md_name.mv_data = NODEKEY(node);
2627 mx->mx_dbxs[dbn].md_name.mv_size = node->mn_ksize;
2628 mx->mx_txn.mt_next_pgno = txn->mt_next_pgno;
2629 mx->mx_txn.mt_oldest = txn->mt_oldest;
2630 mx->mx_txn.mt_u = txn->mt_u;
2631 mx->mx_cursor.mc_initialized = 0;
2632 mx->mx_cursor.mc_eof = 0;
2636 mdb_xcursor_fini(MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx)
2638 txn->mt_next_pgno = mx->mx_txn.mt_next_pgno;
2639 txn->mt_oldest = mx->mx_txn.mt_oldest;
2640 txn->mt_u = mx->mx_txn.mt_u;
2641 txn->mt_dbs[0] = mx->mx_dbs[0];
2642 txn->mt_dbs[1] = mx->mx_dbs[1];
2643 txn->mt_dbxs[0].md_dirty = mx->mx_dbxs[0].md_dirty;
2644 txn->mt_dbxs[1].md_dirty = mx->mx_dbxs[1].md_dirty;
2646 txn->mt_dbs[dbi] = mx->mx_dbs[2];
2647 txn->mt_dbxs[dbi].md_dirty = mx->mx_dbxs[2].md_dirty;
2652 mdb_cursor_open(MDB_txn *txn, MDB_dbi dbi, MDB_cursor **ret)
2655 size_t size = sizeof(MDB_cursor);
2657 if (txn == NULL || ret == NULL || !dbi || dbi >= txn->mt_numdbs)
2660 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT)
2661 size += sizeof(MDB_xcursor);
2663 if ((cursor = calloc(1, size)) != NULL) {
2664 cursor->mc_dbi = dbi;
2665 cursor->mc_txn = txn;
2666 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
2667 MDB_xcursor *mx = (MDB_xcursor *)(cursor + 1);
2668 cursor->mc_xcursor = mx;
2669 mdb_xcursor_init0(txn, dbi, mx);
2680 /* Return the count of duplicate data items for the current key */
2682 mdb_cursor_count(MDB_cursor *mc, unsigned long *countp)
2687 if (mc == NULL || countp == NULL)
2690 if (!(mc->mc_txn->mt_dbs[mc->mc_dbi].md_flags & MDB_DUPSORT))
2693 top = CURSOR_TOP(mc);
2694 leaf = NODEPTR(top->mp_page, top->mp_ki);
2695 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
2698 if (!mc->mc_xcursor->mx_cursor.mc_initialized)
2701 *countp = mc->mc_xcursor->mx_txn.mt_dbs[mc->mc_xcursor->mx_cursor.mc_dbi].md_entries;
2707 mdb_cursor_close(MDB_cursor *cursor)
2709 if (cursor != NULL) {
2715 mdb_update_key(MDB_page *mp, indx_t indx, MDB_val *key)
2717 indx_t ptr, i, numkeys;
2724 node = NODEPTR(mp, indx);
2725 ptr = mp->mp_ptrs[indx];
2726 DPRINTF("update key %u (ofs %u) [%.*s] to [%s] on page %lu",
2728 (int)node->mn_ksize, (char *)NODEKEY(node),
2732 delta = key->mv_size - node->mn_ksize;
2734 if (delta > 0 && SIZELEFT(mp) < delta) {
2735 DPRINTF("OUCH! Not enough room, delta = %d", delta);
2739 numkeys = NUMKEYS(mp);
2740 for (i = 0; i < numkeys; i++) {
2741 if (mp->mp_ptrs[i] <= ptr)
2742 mp->mp_ptrs[i] -= delta;
2745 base = (char *)mp + mp->mp_upper;
2746 len = ptr - mp->mp_upper + NODESIZE;
2747 memmove(base - delta, base, len);
2748 mp->mp_upper -= delta;
2750 node = NODEPTR(mp, indx);
2751 node->mn_ksize = key->mv_size;
2754 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
2759 /* Move a node from src to dst.
2762 mdb_move_node(MDB_txn *txn, MDB_dbi dbi, MDB_pageparent *src, indx_t srcindx,
2763 MDB_pageparent *dst, indx_t dstindx)
2770 /* Mark src and dst as dirty. */
2771 if ((rc = mdb_touch(txn, src)) ||
2772 (rc = mdb_touch(txn, dst)))
2775 if (IS_LEAF2(src->mp_page)) {
2776 srcnode = NODEPTR(src->mp_page, 0); /* fake */
2777 key.mv_size = txn->mt_dbs[dbi].md_pad;
2778 key.mv_data = LEAF2KEY(src->mp_page, srcindx, key.mv_size);
2780 data.mv_data = NULL;
2782 srcnode = NODEPTR(src->mp_page, srcindx);
2783 key.mv_size = NODEKSZ(srcnode);
2784 key.mv_data = NODEKEY(srcnode);
2785 data.mv_size = NODEDSZ(srcnode);
2786 data.mv_data = NODEDATA(srcnode);
2788 DPRINTF("moving %s node %u [%s] on page %lu to node %u on page %lu",
2789 IS_LEAF(src->mp_page) ? "leaf" : "branch",
2792 src->mp_page->mp_pgno,
2793 dstindx, dst->mp_page->mp_pgno);
2795 /* Add the node to the destination page.
2797 rc = mdb_add_node(txn, dbi, dst->mp_page, dstindx, &key, &data, NODEPGNO(srcnode),
2799 if (rc != MDB_SUCCESS)
2802 /* Delete the node from the source page.
2804 mdb_del_node(src->mp_page, srcindx, key.mv_size);
2806 /* The key value just changed due to del_node, find it again.
2808 if (!IS_LEAF2(src->mp_page)) {
2809 srcnode = NODEPTR(src->mp_page, srcindx);
2810 key.mv_data = NODEKEY(srcnode);
2813 /* Update the parent separators.
2816 if (src->mp_pi != 0) {
2817 DPRINTF("update separator for source page %lu to [%s]",
2818 src->mp_page->mp_pgno, DKEY(&key));
2819 if ((rc = mdb_update_key(src->mp_parent, src->mp_pi,
2820 &key)) != MDB_SUCCESS)
2823 if (IS_BRANCH(src->mp_page)) {
2825 nullkey.mv_size = 0;
2826 assert(mdb_update_key(src->mp_page, 0, &nullkey) == MDB_SUCCESS);
2831 if (dst->mp_pi != 0) {
2832 DPRINTF("update separator for destination page %lu to [%s]",
2833 dst->mp_page->mp_pgno, DKEY(&key));
2834 if ((rc = mdb_update_key(dst->mp_parent, dst->mp_pi,
2835 &key)) != MDB_SUCCESS)
2838 if (IS_BRANCH(dst->mp_page)) {
2840 nullkey.mv_size = 0;
2841 assert(mdb_update_key(dst->mp_page, 0, &nullkey) == MDB_SUCCESS);
2849 mdb_merge(MDB_txn *txn, MDB_dbi dbi, MDB_pageparent *src, MDB_pageparent *dst)
2858 DPRINTF("merging page %lu and %lu", src->mp_page->mp_pgno, dst->mp_page->mp_pgno);
2860 assert(txn != NULL);
2861 assert(src->mp_parent); /* can't merge root page */
2862 assert(dst->mp_parent);
2864 /* Mark src and dst as dirty. */
2865 if ((rc = mdb_touch(txn, src)) ||
2866 (rc = mdb_touch(txn, dst)))
2869 /* Move all nodes from src to dst.
2871 if (IS_LEAF2(src->mp_page)) {
2872 key.mv_size = txn->mt_dbs[dbi].md_pad;
2873 key.mv_data = METADATA(src->mp_page);
2874 for (i = 0; i < NUMKEYS(src->mp_page); i++) {
2875 rc = mdb_add_node(txn, dbi, dst->mp_page, NUMKEYS(dst->mp_page), &key,
2877 if (rc != MDB_SUCCESS)
2879 key.mv_data = (char *)key.mv_data + key.mv_size;
2882 for (i = 0; i < NUMKEYS(src->mp_page); i++) {
2883 srcnode = NODEPTR(src->mp_page, i);
2885 key.mv_size = srcnode->mn_ksize;
2886 key.mv_data = NODEKEY(srcnode);
2887 data.mv_size = NODEDSZ(srcnode);
2888 data.mv_data = NODEDATA(srcnode);
2889 rc = mdb_add_node(txn, dbi, dst->mp_page, NUMKEYS(dst->mp_page), &key,
2890 &data, NODEPGNO(srcnode), srcnode->mn_flags);
2891 if (rc != MDB_SUCCESS)
2896 DPRINTF("dst page %lu now has %u keys (%.1f%% filled)",
2897 dst->mp_page->mp_pgno, NUMKEYS(dst->mp_page), (float)PAGEFILL(txn->mt_env, dst->mp_page) / 10);
2899 /* Unlink the src page from parent.
2901 mdb_del_node(src->mp_parent, src->mp_pi, 0);
2902 if (src->mp_pi == 0) {
2904 if ((rc = mdb_update_key(src->mp_parent, 0, &key)) != MDB_SUCCESS)
2908 if (IS_LEAF(src->mp_page))
2909 txn->mt_dbs[dbi].md_leaf_pages--;
2911 txn->mt_dbs[dbi].md_branch_pages--;
2913 mpp.mp_page = src->mp_parent;
2914 dh = (MDB_dhead *)src->mp_parent;
2916 mpp.mp_parent = dh->md_parent;
2917 mpp.mp_pi = dh->md_pi;
2919 return mdb_rebalance(txn, dbi, &mpp);
2922 #define FILL_THRESHOLD 250
2925 mdb_rebalance(MDB_txn *txn, MDB_dbi dbi, MDB_pageparent *mpp)
2930 indx_t si = 0, di = 0;
2932 assert(txn != NULL);
2933 assert(mpp != NULL);
2935 DPRINTF("rebalancing %s page %lu (has %u keys, %.1f%% full)",
2936 IS_LEAF(mpp->mp_page) ? "leaf" : "branch",
2937 mpp->mp_page->mp_pgno, NUMKEYS(mpp->mp_page), (float)PAGEFILL(txn->mt_env, mpp->mp_page) / 10);
2939 if (PAGEFILL(txn->mt_env, mpp->mp_page) >= FILL_THRESHOLD) {
2940 DPRINTF("no need to rebalance page %lu, above fill threshold",
2941 mpp->mp_page->mp_pgno);
2945 if (mpp->mp_parent == NULL) {
2946 if (NUMKEYS(mpp->mp_page) == 0) {
2947 DPUTS("tree is completely empty");
2948 txn->mt_dbs[dbi].md_root = P_INVALID;
2949 txn->mt_dbs[dbi].md_depth = 0;
2950 txn->mt_dbs[dbi].md_leaf_pages = 0;
2951 } else if (IS_BRANCH(mpp->mp_page) && NUMKEYS(mpp->mp_page) == 1) {
2952 DPUTS("collapsing root page!");
2953 txn->mt_dbs[dbi].md_root = NODEPGNO(NODEPTR(mpp->mp_page, 0));
2954 if ((root = mdb_get_page(txn, txn->mt_dbs[dbi].md_root)) == NULL)
2955 return MDB_PAGE_NOTFOUND;
2956 txn->mt_dbs[dbi].md_depth--;
2957 txn->mt_dbs[dbi].md_branch_pages--;
2959 DPUTS("root page doesn't need rebalancing");
2963 /* The parent (branch page) must have at least 2 pointers,
2964 * otherwise the tree is invalid.
2966 assert(NUMKEYS(mpp->mp_parent) > 1);
2968 /* Leaf page fill factor is below the threshold.
2969 * Try to move keys from left or right neighbor, or
2970 * merge with a neighbor page.
2975 if (mpp->mp_pi == 0) {
2976 /* We're the leftmost leaf in our parent.
2978 DPUTS("reading right neighbor");
2979 node = NODEPTR(mpp->mp_parent, mpp->mp_pi + 1);
2980 if ((npp.mp_page = mdb_get_page(txn, NODEPGNO(node))) == NULL)
2981 return MDB_PAGE_NOTFOUND;
2982 npp.mp_pi = mpp->mp_pi + 1;
2984 di = NUMKEYS(mpp->mp_page);
2986 /* There is at least one neighbor to the left.
2988 DPUTS("reading left neighbor");
2989 node = NODEPTR(mpp->mp_parent, mpp->mp_pi - 1);
2990 if ((npp.mp_page = mdb_get_page(txn, NODEPGNO(node))) == NULL)
2991 return MDB_PAGE_NOTFOUND;
2992 npp.mp_pi = mpp->mp_pi - 1;
2993 si = NUMKEYS(npp.mp_page) - 1;
2996 npp.mp_parent = mpp->mp_parent;
2998 DPRINTF("found neighbor page %lu (%u keys, %.1f%% full)",
2999 npp.mp_page->mp_pgno, NUMKEYS(npp.mp_page), (float)PAGEFILL(txn->mt_env, npp.mp_page) / 10);
3001 /* If the neighbor page is above threshold and has at least two
3002 * keys, move one key from it.
3004 * Otherwise we should try to merge them.
3006 if (PAGEFILL(txn->mt_env, npp.mp_page) >= FILL_THRESHOLD && NUMKEYS(npp.mp_page) >= 2)
3007 return mdb_move_node(txn, dbi, &npp, si, mpp, di);
3008 else { /* FIXME: if (has_enough_room()) */
3009 if (mpp->mp_pi == 0)
3010 return mdb_merge(txn, dbi, &npp, mpp);
3012 return mdb_merge(txn, dbi, mpp, &npp);
3017 mdb_del0(MDB_txn *txn, MDB_dbi dbi, unsigned int ki, MDB_pageparent *mpp, MDB_node *leaf)
3021 /* add overflow pages to free list */
3022 if (!IS_LEAF2(mpp->mp_page) && F_ISSET(leaf->mn_flags, F_BIGDATA)) {
3026 memcpy(&pg, NODEDATA(leaf), sizeof(pg));
3027 ovpages = OVPAGES(NODEDSZ(leaf), txn->mt_env->me_psize);
3028 for (i=0; i<ovpages; i++) {
3029 DPRINTF("freed ov page %lu", pg);
3030 mdb_midl_insert(txn->mt_free_pgs, pg);
3034 mdb_del_node(mpp->mp_page, ki, txn->mt_dbs[dbi].md_pad);
3035 txn->mt_dbs[dbi].md_entries--;
3036 rc = mdb_rebalance(txn, dbi, mpp);
3037 if (rc != MDB_SUCCESS)
3038 txn->mt_flags |= MDB_TXN_ERROR;
3044 mdb_del(MDB_txn *txn, MDB_dbi dbi,
3045 MDB_val *key, MDB_val *data,
3054 assert(key != NULL);
3056 DPRINTF("====> delete db %u key [%s]", dbi, DKEY(key));
3058 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
3061 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
3065 if (key->mv_size == 0 || key->mv_size > MAXKEYSIZE) {
3069 mpp.mp_parent = NULL;
3071 if ((rc = mdb_search_page(txn, dbi, key, NULL, 1, &mpp)) != MDB_SUCCESS)
3074 leaf = mdb_search_node(txn, dbi, mpp.mp_page, key, &exact, &ki);
3075 if (leaf == NULL || !exact) {
3076 return MDB_NOTFOUND;
3079 if (!IS_LEAF2(mpp.mp_page) && F_ISSET(leaf->mn_flags, F_DUPDATA)) {
3083 mdb_xcursor_init0(txn, dbi, &mx);
3084 mdb_xcursor_init1(txn, dbi, &mx, leaf);
3085 if (flags == MDB_DEL_DUP) {
3086 rc = mdb_del(&mx.mx_txn, mx.mx_cursor.mc_dbi, data, NULL, 0);
3087 mdb_xcursor_fini(txn, dbi, &mx);
3088 if (rc != MDB_SUCCESS)
3090 /* If sub-DB still has entries, we're done */
3091 if (mx.mx_txn.mt_dbs[mx.mx_cursor.mc_dbi].md_root != P_INVALID) {
3092 memcpy(NODEDATA(leaf), &mx.mx_txn.mt_dbs[mx.mx_cursor.mc_dbi],
3094 txn->mt_dbs[dbi].md_entries--;
3097 /* otherwise fall thru and delete the sub-DB */
3099 /* add all the child DB's pages to the free list */
3100 rc = mdb_search_page(&mx.mx_txn, mx.mx_cursor.mc_dbi,
3101 NULL, &mx.mx_cursor, 0, &mp2);
3102 if (rc == MDB_SUCCESS) {
3103 MDB_ppage *top, *parent;
3107 cursor_pop_page(&mx.mx_cursor);
3108 if (mx.mx_cursor.mc_snum) {
3109 top = CURSOR_TOP(&mx.mx_cursor);
3110 while (mx.mx_cursor.mc_snum > 1) {
3111 parent = CURSOR_PARENT(&mx.mx_cursor);
3112 for (i=0; i<NUMKEYS(top->mp_page); i++) {
3113 ni = NODEPTR(top->mp_page, i);
3114 mdb_midl_insert(txn->mt_free_pgs, NODEPGNO(ni));
3117 if (parent->mp_ki >= NUMKEYS(parent->mp_page)) {
3118 cursor_pop_page(&mx.mx_cursor);
3121 ni = NODEPTR(parent->mp_page, parent->mp_ki);
3122 top->mp_page = mdb_get_page(&mx.mx_txn, NODEPGNO(ni));
3126 mdb_midl_insert(txn->mt_free_pgs, mx.mx_txn.mt_dbs[mx.mx_cursor.mc_dbi].md_root);
3131 if (data && (rc = mdb_read_data(txn, leaf, data)) != MDB_SUCCESS)
3134 return mdb_del0(txn, dbi, ki, &mpp, leaf);
3137 /* Split page <*mpp>, and insert <key,(data|newpgno)> in either left or
3138 * right sibling, at index <*newindxp> (as if unsplit). Updates *mpp and
3139 * *newindxp with the actual values after split, ie if *mpp and *newindxp
3140 * refer to a node in the new right sibling page.
3143 mdb_split(MDB_txn *txn, MDB_dbi dbi, MDB_page **mpp, unsigned int *newindxp,
3144 MDB_val *newkey, MDB_val *newdata, pgno_t newpgno)
3147 int rc = MDB_SUCCESS, ins_new = 0;
3150 unsigned int i, j, split_indx;
3152 MDB_val sepkey, rkey, rdata;
3154 MDB_dpage *mdp, *rdp, *pdp;
3158 assert(txn != NULL);
3160 dh = ((MDB_dhead *)*mpp) - 1;
3161 mdp = (MDB_dpage *)dh;
3162 newindx = *newindxp;
3164 DPRINTF("-----> splitting %s page %lu and adding [%s] at index %i",
3165 IS_LEAF(&mdp->p) ? "leaf" : "branch", mdp->p.mp_pgno,
3166 DKEY(newkey), *newindxp);
3168 if (mdp->h.md_parent == NULL) {
3169 if ((pdp = mdb_new_page(txn, dbi, P_BRANCH, 1)) == NULL)
3172 mdp->h.md_parent = &pdp->p;
3173 txn->mt_dbs[dbi].md_root = pdp->p.mp_pgno;
3174 DPRINTF("root split! new root = %lu", pdp->p.mp_pgno);
3175 txn->mt_dbs[dbi].md_depth++;
3177 /* Add left (implicit) pointer. */
3178 if ((rc = mdb_add_node(txn, dbi, &pdp->p, 0, NULL, NULL,
3179 mdp->p.mp_pgno, 0)) != MDB_SUCCESS)
3182 DPRINTF("parent branch page is %lu", mdp->h.md_parent->mp_pgno);
3185 /* Create a right sibling. */
3186 if ((rdp = mdb_new_page(txn, dbi, mdp->p.mp_flags, 1)) == NULL)
3188 rdp->h.md_parent = mdp->h.md_parent;
3189 rdp->h.md_pi = mdp->h.md_pi + 1;
3190 DPRINTF("new right sibling: page %lu", rdp->p.mp_pgno);
3192 split_indx = NUMKEYS(&mdp->p) / 2 + 1;
3194 if (IS_LEAF2(&rdp->p)) {
3197 unsigned int nkeys = NUMKEYS(&mdp->p), lsize, rsize, ksize;
3198 /* Move half of the keys to the right sibling */
3200 x = *newindxp - split_indx;
3201 ksize = txn->mt_dbs[dbi].md_pad;
3202 split = LEAF2KEY(&mdp->p, split_indx, ksize);
3203 rsize = (nkeys - split_indx) * ksize;
3204 lsize = (nkeys - split_indx) * sizeof(indx_t);
3205 mdp->p.mp_lower -= lsize;
3206 rdp->p.mp_lower += lsize;
3207 mdp->p.mp_upper += rsize - lsize;
3208 rdp->p.mp_upper -= rsize - lsize;
3209 sepkey.mv_size = ksize;
3210 if (newindx == split_indx) {
3211 sepkey.mv_data = newkey->mv_data;
3213 sepkey.mv_data = split;
3216 ins = LEAF2KEY(&mdp->p, *newindxp, ksize);
3217 memcpy(&rdp->p.mp_ptrs, split, rsize);
3218 sepkey.mv_data = &rdp->p.mp_ptrs;
3219 memmove(ins+ksize, ins, (split_indx - *newindxp) * ksize);
3220 memcpy(ins, newkey->mv_data, ksize);
3221 mdp->p.mp_lower += sizeof(indx_t);
3222 mdp->p.mp_upper -= ksize - sizeof(indx_t);
3225 memcpy(&rdp->p.mp_ptrs, split, x * ksize);
3226 ins = LEAF2KEY(&rdp->p, x, ksize);
3227 memcpy(ins, newkey->mv_data, ksize);
3228 memcpy(ins+ksize, split + x * ksize, rsize - x * ksize);
3229 rdp->p.mp_lower += sizeof(indx_t);
3230 rdp->p.mp_upper -= ksize - sizeof(indx_t);
3237 /* Move half of the keys to the right sibling. */
3238 if ((copy = malloc(txn->mt_env->me_psize)) == NULL)
3240 memcpy(copy, &mdp->p, txn->mt_env->me_psize);
3241 memset(&mdp->p.mp_ptrs, 0, txn->mt_env->me_psize - PAGEHDRSZ);
3242 mdp->p.mp_lower = PAGEHDRSZ;
3243 mdp->p.mp_upper = txn->mt_env->me_psize;
3245 /* First find the separating key between the split pages.
3247 memset(&sepkey, 0, sizeof(sepkey));
3248 if (newindx == split_indx) {
3249 sepkey.mv_size = newkey->mv_size;
3250 sepkey.mv_data = newkey->mv_data;
3252 node = NODEPTR(copy, split_indx);
3253 sepkey.mv_size = node->mn_ksize;
3254 sepkey.mv_data = NODEKEY(node);
3258 DPRINTF("separator is [%s]", DKEY(&sepkey));
3260 /* Copy separator key to the parent.
3262 if (SIZELEFT(rdp->h.md_parent) < mdb_branch_size(txn->mt_env, &sepkey)) {
3263 rc = mdb_split(txn, dbi, &rdp->h.md_parent, &rdp->h.md_pi,
3264 &sepkey, NULL, rdp->p.mp_pgno);
3266 /* Right page might now have changed parent.
3267 * Check if left page also changed parent.
3269 if (rdp->h.md_parent != mdp->h.md_parent &&
3270 mdp->h.md_pi >= NUMKEYS(mdp->h.md_parent)) {
3271 mdp->h.md_parent = rdp->h.md_parent;
3272 mdp->h.md_pi = rdp->h.md_pi - 1;
3275 rc = mdb_add_node(txn, dbi, rdp->h.md_parent, rdp->h.md_pi,
3276 &sepkey, NULL, rdp->p.mp_pgno, 0);
3278 if (IS_LEAF2(&rdp->p)) {
3281 if (rc != MDB_SUCCESS) {
3286 for (i = j = 0; i <= NUMKEYS(copy); j++) {
3287 if (i < split_indx) {
3288 /* Re-insert in left sibling. */
3291 /* Insert in right sibling. */
3292 if (i == split_indx)
3293 /* Reset insert index for right sibling. */
3294 j = (i == newindx && ins_new);
3298 if (i == newindx && !ins_new) {
3299 /* Insert the original entry that caused the split. */
3300 rkey.mv_data = newkey->mv_data;
3301 rkey.mv_size = newkey->mv_size;
3302 if (IS_LEAF(&mdp->p)) {
3303 rdata.mv_data = newdata->mv_data;
3304 rdata.mv_size = newdata->mv_size;
3311 /* Update page and index for the new key. */
3314 } else if (i == NUMKEYS(copy)) {
3317 node = NODEPTR(copy, i);
3318 rkey.mv_data = NODEKEY(node);
3319 rkey.mv_size = node->mn_ksize;
3320 if (IS_LEAF(&mdp->p)) {
3321 rdata.mv_data = NODEDATA(node);
3322 rdata.mv_size = node->mn_dsize;
3324 pgno = NODEPGNO(node);
3325 flags = node->mn_flags;
3330 if (!IS_LEAF(&mdp->p) && j == 0) {
3331 /* First branch index doesn't need key data. */
3335 rc = mdb_add_node(txn, dbi, &pdp->p, j, &rkey, &rdata, pgno,flags);
3343 mdb_put0(MDB_txn *txn, MDB_dbi dbi,
3344 MDB_val *key, MDB_val *data, unsigned int flags)
3346 int rc = MDB_SUCCESS, exact;
3350 MDB_val xdata, *rdata, dkey;
3352 char dbuf[PAGESIZE];
3357 DPRINTF("==> put db %u key [%s], size %zu, data size %zu",
3358 dbi, DKEY(key), key->mv_size, data->mv_size);
3361 mpp.mp_parent = NULL;
3363 rc = mdb_search_page(txn, dbi, key, NULL, 1, &mpp);
3364 if (rc == MDB_SUCCESS) {
3365 leaf = mdb_search_node(txn, dbi, mpp.mp_page, key, &exact, &ki);
3366 if (leaf && exact) {
3367 if (flags == MDB_NOOVERWRITE) {
3368 DPRINTF("duplicate key [%s]", DKEY(key));
3369 return MDB_KEYEXIST;
3371 /* there's only a key anyway, so this is a no-op */
3372 if (IS_LEAF2(mpp.mp_page))
3375 if (F_ISSET(txn->mt_dbs[dbi].md_flags, MDB_DUPSORT)) {
3376 /* Was a single item before, must convert now */
3377 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
3378 dkey.mv_size = NODEDSZ(leaf);
3379 dkey.mv_data = dbuf;
3380 memcpy(dbuf, NODEDATA(leaf), dkey.mv_size);
3381 /* data matches, ignore it */
3382 if (!mdb_dcmp(txn, dbi, data, &dkey))
3383 return (flags == MDB_NODUPDATA) ? MDB_KEYEXIST : MDB_SUCCESS;
3384 memset(&dummy, 0, sizeof(dummy));
3385 if (txn->mt_dbs[dbi].md_flags & MDB_DUPFIXED) {
3386 dummy.md_pad = data->mv_size;
3387 dummy.md_flags = MDB_DUPFIXED;
3388 if (txn->mt_dbs[dbi].md_flags & MDB_INTEGERDUP)
3389 dummy.md_flags |= MDB_INTEGERKEY;
3391 dummy.md_root = P_INVALID;
3392 if (dkey.mv_size == sizeof(MDB_db)) {
3393 memcpy(NODEDATA(leaf), &dummy, sizeof(dummy));
3396 mdb_del_node(mpp.mp_page, ki, 0);
3399 xdata.mv_size = sizeof(MDB_db);
3400 xdata.mv_data = &dummy;
3405 /* same size, just replace it */
3406 if (NODEDSZ(leaf) == data->mv_size) {
3407 memcpy(NODEDATA(leaf), data->mv_data, data->mv_size);
3410 mdb_del_node(mpp.mp_page, ki, 0);
3412 if (leaf == NULL) { /* append if not found */
3413 ki = NUMKEYS(mpp.mp_page);
3414 DPRINTF("appending key at index %i", ki);
3416 } else if (rc == MDB_NOTFOUND) {
3418 /* new file, just write a root leaf page */
3419 DPUTS("allocating new root leaf page");
3420 if ((dp = mdb_new_page(txn, dbi, P_LEAF, 1)) == NULL) {
3423 mpp.mp_page = &dp->p;
3424 txn->mt_dbs[dbi].md_root = mpp.mp_page->mp_pgno;
3425 txn->mt_dbs[dbi].md_depth++;
3426 txn->mt_dbxs[dbi].md_dirty = 1;
3427 if ((txn->mt_dbs[dbi].md_flags & (MDB_DUPSORT|MDB_DUPFIXED)) == MDB_DUPFIXED)
3428 mpp.mp_page->mp_flags |= P_LEAF2;
3434 assert(IS_LEAF(mpp.mp_page));
3435 DPRINTF("there are %u keys, should insert new key at index %i",
3436 NUMKEYS(mpp.mp_page), ki);
3441 nsize = IS_LEAF2(mpp.mp_page) ? key->mv_size : mdb_leaf_size(txn->mt_env, key, rdata);
3442 if (SIZELEFT(mpp.mp_page) < nsize) {
3443 rc = mdb_split(txn, dbi, &mpp.mp_page, &ki, key, rdata, P_INVALID);
3445 /* There is room already in this leaf page. */
3446 rc = mdb_add_node(txn, dbi, mpp.mp_page, ki, key, rdata, 0, 0);
3449 if (rc != MDB_SUCCESS)
3450 txn->mt_flags |= MDB_TXN_ERROR;
3452 /* Remember if we just added a subdatabase */
3453 if (flags & F_SUBDATA) {
3454 leaf = NODEPTR(mpp.mp_page, ki);
3455 leaf->mn_flags |= F_SUBDATA;
3458 /* Now store the actual data in the child DB. Note that we're
3459 * storing the user data in the keys field, so there are strict
3460 * size limits on dupdata. The actual data fields of the child
3461 * DB are all zero size.
3466 leaf = NODEPTR(mpp.mp_page, ki);
3468 mdb_xcursor_init0(txn, dbi, &mx);
3469 mdb_xcursor_init1(txn, dbi, &mx, leaf);
3472 if (flags == MDB_NODUPDATA)
3473 flags = MDB_NOOVERWRITE;
3474 /* converted, write the original data first */
3476 rc = mdb_put0(&mx.mx_txn, mx.mx_cursor.mc_dbi, &dkey, &xdata, flags);
3478 leaf->mn_flags |= F_DUPDATA;
3480 rc = mdb_put0(&mx.mx_txn, mx.mx_cursor.mc_dbi, data, &xdata, flags);
3481 mdb_xcursor_fini(txn, dbi, &mx);
3482 memcpy(NODEDATA(leaf), &mx.mx_txn.mt_dbs[mx.mx_cursor.mc_dbi],
3485 txn->mt_dbs[dbi].md_entries++;
3493 mdb_put(MDB_txn *txn, MDB_dbi dbi,
3494 MDB_val *key, MDB_val *data, unsigned int flags)
3496 assert(key != NULL);
3497 assert(data != NULL);
3499 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
3502 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
3506 if (key->mv_size == 0 || key->mv_size > MAXKEYSIZE) {
3510 if ((flags & (MDB_NOOVERWRITE|MDB_NODUPDATA)) != flags)
3513 return mdb_put0(txn, dbi, key, data, flags);
3517 mdb_env_set_flags(MDB_env *env, unsigned int flag, int onoff)
3519 #define CHANGEABLE (MDB_NOSYNC)
3520 if ((flag & CHANGEABLE) != flag)
3523 env->me_flags |= flag;
3525 env->me_flags &= ~flag;
3530 mdb_env_get_flags(MDB_env *env, unsigned int *arg)
3535 *arg = env->me_flags;
3540 mdb_env_get_path(MDB_env *env, const char **arg)
3545 *arg = env->me_path;
3550 mdb_stat0(MDB_env *env, MDB_db *db, MDB_stat *arg)
3552 arg->ms_psize = env->me_psize;
3553 arg->ms_depth = db->md_depth;
3554 arg->ms_branch_pages = db->md_branch_pages;
3555 arg->ms_leaf_pages = db->md_leaf_pages;
3556 arg->ms_overflow_pages = db->md_overflow_pages;
3557 arg->ms_entries = db->md_entries;
3562 mdb_env_stat(MDB_env *env, MDB_stat *arg)
3564 if (env == NULL || arg == NULL)
3567 return mdb_stat0(env, &env->me_meta->mm_dbs[MAIN_DBI], arg);
3570 int mdb_open(MDB_txn *txn, const char *name, unsigned int flags, MDB_dbi *dbi)
3580 if (flags & (MDB_DUPSORT|MDB_REVERSEKEY|MDB_INTEGERKEY))
3581 txn->mt_dbs[MAIN_DBI].md_flags |= (flags & (MDB_DUPSORT|MDB_REVERSEKEY|MDB_INTEGERKEY));
3585 /* Is the DB already open? */
3587 for (i=2; i<txn->mt_numdbs; i++) {
3588 if (len == txn->mt_dbxs[i].md_name.mv_size &&
3589 !strncmp(name, txn->mt_dbxs[i].md_name.mv_data, len)) {
3595 if (txn->mt_numdbs >= txn->mt_env->me_maxdbs - 1)
3598 /* Find the DB info */
3600 key.mv_data = (void *)name;
3601 rc = mdb_get(txn, MAIN_DBI, &key, &data);
3603 /* Create if requested */
3604 if (rc == MDB_NOTFOUND && (flags & MDB_CREATE)) {
3606 data.mv_size = sizeof(MDB_db);
3607 data.mv_data = &dummy;
3608 memset(&dummy, 0, sizeof(dummy));
3609 dummy.md_root = P_INVALID;
3610 dummy.md_flags = flags & 0xffff;
3611 rc = mdb_put0(txn, MAIN_DBI, &key, &data, F_SUBDATA);
3615 /* OK, got info, add to table */
3616 if (rc == MDB_SUCCESS) {
3617 txn->mt_dbxs[txn->mt_numdbs].md_name.mv_data = strdup(name);
3618 txn->mt_dbxs[txn->mt_numdbs].md_name.mv_size = len;
3619 txn->mt_dbxs[txn->mt_numdbs].md_cmp = NULL;
3620 txn->mt_dbxs[txn->mt_numdbs].md_dcmp = NULL;
3621 txn->mt_dbxs[txn->mt_numdbs].md_rel = NULL;
3622 txn->mt_dbxs[txn->mt_numdbs].md_parent = MAIN_DBI;
3623 txn->mt_dbxs[txn->mt_numdbs].md_dirty = dirty;
3624 memcpy(&txn->mt_dbs[txn->mt_numdbs], data.mv_data, sizeof(MDB_db));
3625 *dbi = txn->mt_numdbs;
3626 txn->mt_env->me_dbs[0][txn->mt_numdbs] = txn->mt_dbs[txn->mt_numdbs];
3627 txn->mt_env->me_dbs[1][txn->mt_numdbs] = txn->mt_dbs[txn->mt_numdbs];
3634 int mdb_stat(MDB_txn *txn, MDB_dbi dbi, MDB_stat *arg)
3636 if (txn == NULL || arg == NULL || dbi >= txn->mt_numdbs)
3639 return mdb_stat0(txn->mt_env, &txn->mt_dbs[dbi], arg);
3642 void mdb_close(MDB_txn *txn, MDB_dbi dbi)
3645 if (dbi <= MAIN_DBI || dbi >= txn->mt_numdbs)
3647 ptr = txn->mt_dbxs[dbi].md_name.mv_data;
3648 txn->mt_dbxs[dbi].md_name.mv_data = NULL;
3649 txn->mt_dbxs[dbi].md_name.mv_size = 0;
3653 int mdb_set_compare(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
3655 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
3658 txn->mt_dbxs[dbi].md_cmp = cmp;
3662 int mdb_set_dupsort(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
3664 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
3667 txn->mt_dbxs[dbi].md_dcmp = cmp;
3671 int mdb_set_relfunc(MDB_txn *txn, MDB_dbi dbi, MDB_rel_func *rel)
3673 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
3676 txn->mt_dbxs[dbi].md_rel = rel;