2 * @brief memory-mapped database library
4 * A Btree-based database management library modeled loosely on the
5 * BerkeleyDB API, but much simplified.
8 * Copyright 2011 Howard Chu, Symas Corp.
11 * Redistribution and use in source and binary forms, with or without
12 * modification, are permitted only as authorized by the OpenLDAP
15 * A copy of this license is available in the file LICENSE in the
16 * top-level directory of the distribution or, alternatively, at
17 * <http://www.OpenLDAP.org/license.html>.
19 * This code is derived from btree.c written by Martin Hedenfalk.
21 * Copyright (c) 2009, 2010 Martin Hedenfalk <martin@bzero.se>
23 * Permission to use, copy, modify, and distribute this software for any
24 * purpose with or without fee is hereby granted, provided that the above
25 * copyright notice and this permission notice appear in all copies.
27 * THE SOFTWARE IS PROVIDED "AS IS" AND THE AUTHOR DISCLAIMS ALL WARRANTIES
28 * WITH REGARD TO THIS SOFTWARE INCLUDING ALL IMPLIED WARRANTIES OF
29 * MERCHANTABILITY AND FITNESS. IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR
30 * ANY SPECIAL, DIRECT, INDIRECT, OR CONSEQUENTIAL DAMAGES OR ANY DAMAGES
31 * WHATSOEVER RESULTING FROM LOSS OF USE, DATA OR PROFITS, WHETHER IN AN
32 * ACTION OF CONTRACT, NEGLIGENCE OR OTHER TORTIOUS ACTION, ARISING OUT OF
33 * OR IN CONNECTION WITH THE USE OR PERFORMANCE OF THIS SOFTWARE.
35 #include <sys/types.h>
37 #include <sys/param.h>
43 #ifdef HAVE_SYS_FILE_H
66 /** @defgroup internal MDB Internals
69 /** @defgroup compat Windows Compatibility Macros
73 #define pthread_t DWORD
74 #define pthread_mutex_t HANDLE
75 #define pthread_key_t DWORD
76 #define pthread_self() GetCurrentThreadId()
77 #define pthread_key_create(x,y) *(x) = TlsAlloc()
78 #define pthread_key_delete(x) TlsFree(x)
79 #define pthread_getspecific(x) TlsGetValue(x)
80 #define pthread_setspecific(x,y) TlsSetValue(x,y)
81 #define pthread_mutex_unlock(x) ReleaseMutex(x)
82 #define pthread_mutex_lock(x) WaitForSingleObject(x, INFINITE)
83 #define LOCK_MUTEX_R(env) pthread_mutex_lock(env->me_rmutex)
84 #define UNLOCK_MUTEX_R(env) pthread_mutex_unlock(env->me_rmutex)
85 #define LOCK_MUTEX_W(env) pthread_mutex_lock(env->me_wmutex)
86 #define UNLOCK_MUTEX_W(env) pthread_mutex_unlock(env->me_wmutex)
87 #define getpid() GetCurrentProcessId()
88 #define fdatasync(fd) !FlushFileBuffers(fd)
89 #define ErrCode() GetLastError()
90 #define GetPageSize(x) {SYSTEM_INFO si; GetSystemInfo(&si); (x) = si.dwPageSize;}
91 #define close(fd) CloseHandle(fd)
92 #define munmap(ptr,len) UnmapViewOfFile(ptr)
94 #define LOCK_MUTEX_R(env) pthread_mutex_lock(&env->me_txns->mti_mutex)
95 #define UNLOCK_MUTEX_R(env) pthread_mutex_unlock(&env->me_txns->mti_mutex)
96 #define LOCK_MUTEX_W(env) pthread_mutex_lock(&env->me_txns->mti_wmutex)
97 #define UNLOCK_MUTEX_W(env) pthread_mutex_unlock(&env->me_txns->mti_wmutex)
98 #define ErrCode() errno
100 #define INVALID_HANDLE_VALUE -1
101 #define GetPageSize(x) (x) = sysconf(_SC_PAGE_SIZE)
107 /* Note: If O_DSYNC is undefined but exists in /usr/include,
108 * preferably set some compiler flag to get the definition.
109 * Otherwise compile with the less efficient -DMDB_DSYNC=O_SYNC.
112 # define MDB_DSYNC O_DSYNC
116 typedef ULONG pgno_t;
122 #if !(__STDC_VERSION__ >= 199901L || defined(__GNUC__))
123 # define DPRINTF (void) /* Vararg macros may be unsupported */
125 # define DPRINTF(fmt, ...) /* Requires 2 or more args */ \
126 fprintf(stderr, "%s:%d:(%p) " fmt "\n", __func__, __LINE__, pthread_self(), __VA_ARGS__)
128 # define DPRINTF(fmt, ...) ((void) 0)
130 #define DPUTS(arg) DPRINTF("%s", arg)
132 #define PAGESIZE 4096
133 #define MDB_MINKEYS 2
134 #define MDB_MAGIC 0xBEEFC0DE
135 #define MDB_VERSION 1
136 #define MAXKEYSIZE 511
138 #define KBUF (MAXKEYSIZE*2+1)
139 #define DKBUF char kbuf[KBUF]
140 #define DKEY(x) mdb_dkey(x, kbuf)
146 /* The DB view is always consistent because all writes are wrapped in
147 * the wmutex. Finer-grained locks aren't necessary.
153 #define LAZY_MUTEX_LOCK(x)
154 #define LAZY_MUTEX_UNLOCK(x)
155 #define LAZY_RWLOCK_UNLOCK(x)
156 #define LAZY_RWLOCK_WRLOCK(x)
157 #define LAZY_RWLOCK_RDLOCK(x)
158 #define LAZY_RWLOCK_DEF(x)
159 #define LAZY_RWLOCK_INIT(x,y)
160 #define LAZY_RWLOCK_DESTROY(x)
162 #define LAZY_MUTEX_LOCK(x) pthread_mutex_lock(x)
163 #define LAZY_MUTEX_UNLOCK(x) pthread_mutex_unlock(x)
164 #define LAZY_RWLOCK_UNLOCK(x) pthread_rwlock_unlock(x)
165 #define LAZY_RWLOCK_WRLOCK(x) pthread_rwlock_wrlock(x)
166 #define LAZY_RWLOCK_RDLOCK(x) pthread_rwlock_rdlock(x)
167 #define LAZY_RWLOCK_DEF(x) pthread_rwlock_t x
168 #define LAZY_RWLOCK_INIT(x,y) pthread_rwlock_init(x,y)
169 #define LAZY_RWLOCK_DESTROY(x) pthread_rwlock_destroy(x)
172 #define P_INVALID (~0UL)
174 #define F_ISSET(w, f) (((w) & (f)) == (f))
176 typedef uint16_t indx_t;
178 #define DEFAULT_READERS 126
179 #define DEFAULT_MAPSIZE 1048576
181 /* Lock descriptor stuff */
183 #define CACHELINE 64 /* most CPUs. Itanium uses 128 */
186 typedef struct MDB_rxbody {
192 typedef struct MDB_reader {
195 #define mr_txnid mru.mrx.mrb_txnid
196 #define mr_pid mru.mrx.mrb_pid
197 #define mr_tid mru.mrx.mrb_tid
198 /* cache line alignment */
199 char pad[(sizeof(MDB_rxbody)+CACHELINE-1) & ~(CACHELINE-1)];
203 typedef struct MDB_txbody {
205 uint32_t mtb_version;
206 /* For POSIX the actual mutexes reside in shared memory.
207 * On Windows, mutexes are allocated by the kernel; we store
208 * the name in shared memory so that other processes can
214 pthread_mutex_t mtb_mutex;
217 uint32_t mtb_numreaders;
218 uint32_t mtb_me_toggle;
221 typedef struct MDB_txninfo {
224 #define mti_magic mt1.mtb.mtb_magic
225 #define mti_version mt1.mtb.mtb_version
226 #define mti_mutex mt1.mtb.mtb_mutex
227 #define mti_rmname mt1.mtb.mtb_rmname
228 #define mti_txnid mt1.mtb.mtb_txnid
229 #define mti_numreaders mt1.mtb.mtb_numreaders
230 #define mti_me_toggle mt1.mtb.mtb_me_toggle
231 char pad[(sizeof(MDB_txbody)+CACHELINE-1) & ~(CACHELINE-1)];
236 #define mti_wmname mt2.mt2_wmname
238 pthread_mutex_t mt2_wmutex;
239 #define mti_wmutex mt2.mt2_wmutex
241 char pad[(sizeof(pthread_mutex_t)+CACHELINE-1) & ~(CACHELINE-1)];
243 MDB_reader mti_readers[1];
246 /* Common header for all page types. Overflow pages
247 * occupy a number of contiguous pages with no
248 * headers on any page after the first.
250 typedef struct MDB_page { /* represents a page of storage */
251 #define mp_pgno mp_p.p_pgno
253 pgno_t p_pgno; /* page number */
254 void * p_align; /* for IL32P64 */
256 #define P_BRANCH 0x01 /* branch page */
257 #define P_LEAF 0x02 /* leaf page */
258 #define P_OVERFLOW 0x04 /* overflow page */
259 #define P_META 0x08 /* meta page */
260 #define P_DIRTY 0x10 /* dirty page */
261 #define P_LEAF2 0x20 /* DB with small, fixed size keys and no data */
263 #define mp_lower mp_pb.pb.pb_lower
264 #define mp_upper mp_pb.pb.pb_upper
265 #define mp_pages mp_pb.pb_pages
268 indx_t pb_lower; /* lower bound of free space */
269 indx_t pb_upper; /* upper bound of free space */
271 uint32_t pb_pages; /* number of overflow pages */
273 indx_t mp_ptrs[1]; /* dynamic size */
276 #define PAGEHDRSZ ((unsigned) offsetof(MDB_page, mp_ptrs))
278 #define NUMKEYS(p) (((p)->mp_lower - PAGEHDRSZ) >> 1)
279 #define SIZELEFT(p) (indx_t)((p)->mp_upper - (p)->mp_lower)
280 #define PAGEFILL(env, p) (1000L * ((env)->me_psize - PAGEHDRSZ - SIZELEFT(p)) / \
281 ((env)->me_psize - PAGEHDRSZ))
282 #define IS_LEAF(p) F_ISSET((p)->mp_flags, P_LEAF)
283 #define IS_LEAF2(p) F_ISSET((p)->mp_flags, P_LEAF2)
284 #define IS_BRANCH(p) F_ISSET((p)->mp_flags, P_BRANCH)
285 #define IS_OVERFLOW(p) F_ISSET((p)->mp_flags, P_OVERFLOW)
287 #define OVPAGES(size, psize) ((PAGEHDRSZ-1 + (size)) / (psize) + 1)
289 typedef struct MDB_db {
290 uint32_t md_pad; /* also ksize for LEAF2 pages */
293 ULONG md_branch_pages;
295 ULONG md_overflow_pages;
303 typedef struct MDB_meta { /* meta (footer) page content */
306 void *mm_address; /* address for fixed mapping */
307 size_t mm_mapsize; /* size of mmap region */
308 MDB_db mm_dbs[2]; /* first is free space, 2nd is main db */
309 #define mm_psize mm_dbs[0].md_pad
310 #define mm_flags mm_dbs[0].md_flags
311 pgno_t mm_last_pg; /* last used page in file */
312 ULONG mm_txnid; /* txnid that committed this page */
315 typedef struct MDB_dhead { /* a dirty page */
317 unsigned md_pi; /* parent index */
321 typedef struct MDB_dpage {
326 typedef struct MDB_oldpages {
327 struct MDB_oldpages *mo_next;
329 pgno_t mo_pages[1]; /* dynamic */
332 typedef struct MDB_pageparent {
338 static MDB_dpage *mdb_alloc_page(MDB_txn *txn, MDB_dbi dbi, MDB_page *parent, unsigned int parent_idx, int num);
339 static int mdb_touch(MDB_txn *txn, MDB_dbi dbi, MDB_pageparent *mp);
341 typedef struct MDB_ppage { /* ordered list of pages */
343 unsigned int mp_ki; /* cursor index on page */
346 #define CURSOR_TOP(c) (&(c)->mc_stack[(c)->mc_snum-1])
347 #define CURSOR_PARENT(c) (&(c)->mc_stack[(c)->mc_snum-2])
348 #define CURSOR_STACK 32
353 struct MDB_xcursor *mc_xcursor;
356 unsigned int mc_snum; /* number of pushed pages */
357 unsigned int mc_flags;
358 #define C_INITIALIZED 0x01
360 #define C_XDIRTY 0x04
361 MDB_ppage mc_stack[CURSOR_STACK]; /* stack of parent pages */
364 #define METADATA(p) ((void *)((char *)(p) + PAGEHDRSZ))
366 typedef struct MDB_node {
367 #define mn_pgno mn_p.np_pgno
368 #define mn_dsize mn_p.np_dsize
370 pgno_t np_pgno; /* child page number */
371 uint32_t np_dsize; /* leaf data size */
373 unsigned int mn_flags:4;
374 unsigned int mn_ksize:12; /* key size */
375 #define F_BIGDATA 0x01 /* data put on overflow page */
376 #define F_SUBDATA 0x02 /* data is a sub-database */
377 #define F_DUPDATA 0x04 /* data has duplicates */
381 typedef struct MDB_dbx {
383 MDB_cmp_func *md_cmp; /* user compare function */
384 MDB_cmp_func *md_dcmp; /* user dupsort function */
385 MDB_rel_func *md_rel; /* user relocate function */
387 unsigned int md_dirty;
391 pgno_t mt_next_pgno; /* next unallocated page */
394 pgno_t *mt_free_pgs; /* this is an IDL */
396 ID2L dirty_list; /* modified pages */
399 MDB_dbx *mt_dbxs; /* array */
401 unsigned int mt_numdbs;
403 #define MDB_TXN_RDONLY 0x01 /* read-only transaction */
404 #define MDB_TXN_ERROR 0x02 /* an error has occurred */
405 unsigned int mt_flags;
406 unsigned int mt_toggle;
409 /* Context for sorted-dup records */
410 typedef struct MDB_xcursor {
411 MDB_cursor mx_cursor;
420 HANDLE me_mfd; /* just for writing the meta pages */
421 #define MDB_FATAL_ERROR 0x80000000U
423 uint32_t me_extrapad; /* unused for now */
424 unsigned int me_maxreaders;
425 unsigned int me_numdbs;
426 unsigned int me_maxdbs;
429 MDB_txninfo *me_txns;
430 MDB_meta *me_metas[2];
431 MDB_txn *me_txn; /* current write transaction */
433 off_t me_size; /* current file size */
434 pgno_t me_maxpg; /* me_mapsize / me_psize */
435 unsigned int me_psize;
436 unsigned int me_db_toggle;
437 MDB_dbx *me_dbxs; /* array */
439 MDB_oldpages *me_pghead;
440 pthread_key_t me_txkey; /* thread-key for readers */
441 MDB_dpage *me_dpages;
442 pgno_t me_free_pgs[MDB_IDL_UM_SIZE];
443 ID2 me_dirty_list[MDB_IDL_DB_SIZE];
444 LAZY_RWLOCK_DEF(me_dblock);
446 HANDLE me_rmutex; /* Windows mutexes don't reside in shared mem */
451 #define NODESIZE offsetof(MDB_node, mn_data)
453 #define INDXSIZE(k) (NODESIZE + ((k) == NULL ? 0 : (k)->mv_size))
454 #define LEAFSIZE(k, d) (NODESIZE + (k)->mv_size + (d)->mv_size)
455 #define NODEPTR(p, i) ((MDB_node *)((char *)(p) + (p)->mp_ptrs[i]))
456 #define NODEKEY(node) (void *)((node)->mn_data)
457 #define NODEDATA(node) (void *)((char *)(node)->mn_data + (node)->mn_ksize)
458 #define NODEPGNO(node) ((node)->mn_pgno)
459 #define NODEDSZ(node) ((node)->mn_dsize)
460 #define NODEKSZ(node) ((node)->mn_ksize)
461 #define LEAF2KEY(p, i, ks) ((char *)(p) + PAGEHDRSZ + ((i)*(ks)))
463 #define MDB_SET_KEY(node, key) if (key!=NULL) {(key)->mv_size = NODEKSZ(node); (key)->mv_data = NODEKEY(node);}
465 #define MDB_COMMIT_PAGES 64 /* max number of pages to write in one commit */
467 static int mdb_search_page_root(MDB_txn *txn,
468 MDB_dbi dbi, MDB_val *key,
469 MDB_cursor *cursor, int modify,
470 MDB_pageparent *mpp);
471 static int mdb_search_page(MDB_txn *txn,
472 MDB_dbi dbi, MDB_val *key,
473 MDB_cursor *cursor, int modify,
474 MDB_pageparent *mpp);
476 static int mdb_env_read_header(MDB_env *env, MDB_meta *meta);
477 static int mdb_env_read_meta(MDB_env *env, int *which);
478 static int mdb_env_write_meta(MDB_txn *txn);
479 static int mdb_get_page(MDB_txn *txn, pgno_t pgno, MDB_page **mp);
481 static MDB_node *mdb_search_node(MDB_txn *txn, MDB_dbi dbi, MDB_page *mp,
482 MDB_val *key, int *exactp, unsigned int *kip);
483 static int mdb_add_node(MDB_txn *txn, MDB_dbi dbi, MDB_page *mp,
484 indx_t indx, MDB_val *key, MDB_val *data,
485 pgno_t pgno, uint8_t flags);
486 static void mdb_del_node(MDB_page *mp, indx_t indx, int ksize);
487 static int mdb_del0(MDB_cursor *mc, unsigned int ki,
488 MDB_pageparent *mpp, MDB_node *leaf);
490 static int mdb_put0(MDB_txn *txn, MDB_dbi dbi,
491 MDB_val *key, MDB_val *data, unsigned int flags);
493 static int mdb_read_data(MDB_txn *txn, MDB_node *leaf, MDB_val *data);
495 static int mdb_rebalance(MDB_txn *txn, MDB_dbi dbi, MDB_pageparent *mp);
496 static int mdb_update_key(MDB_page *mp, indx_t indx, MDB_val *key);
497 static int mdb_move_node(MDB_txn *txn, MDB_dbi dbi,
498 MDB_pageparent *src, indx_t srcindx,
499 MDB_pageparent *dst, indx_t dstindx);
500 static int mdb_merge(MDB_txn *txn, MDB_dbi dbi, MDB_pageparent *src,
501 MDB_pageparent *dst);
502 static int mdb_split(MDB_txn *txn, MDB_dbi dbi, MDB_page **mpp,
503 unsigned int *newindxp, MDB_val *newkey,
504 MDB_val *newdata, pgno_t newpgno);
505 static MDB_dpage *mdb_new_page(MDB_txn *txn, MDB_dbi dbi, uint32_t flags, int num);
507 static void cursor_pop_page(MDB_cursor *cursor);
508 static MDB_ppage *cursor_push_page(MDB_cursor *cursor,
511 static int mdb_sibling(MDB_cursor *cursor, int move_right);
512 static void mdb_cursor_reset(MDB_cursor *mc);
513 static int mdb_cursor_next(MDB_cursor *cursor,
514 MDB_val *key, MDB_val *data, MDB_cursor_op op);
515 static int mdb_cursor_prev(MDB_cursor *cursor,
516 MDB_val *key, MDB_val *data, MDB_cursor_op op);
517 static int mdb_cursor_set(MDB_cursor *cursor,
518 MDB_val *key, MDB_val *data, MDB_cursor_op op, int *exactp);
519 static int mdb_cursor_first(MDB_cursor *cursor,
520 MDB_val *key, MDB_val *data);
521 static int mdb_cursor_last(MDB_cursor *cursor,
522 MDB_val *key, MDB_val *data);
524 static void mdb_xcursor_init0(MDB_cursor *mc);
525 static void mdb_xcursor_init1(MDB_cursor *mc, MDB_page *mp, MDB_node *node);
526 static void mdb_xcursor_init2(MDB_cursor *mc);
527 static void mdb_xcursor_fini(MDB_cursor *mc);
529 static size_t mdb_leaf_size(MDB_env *env, MDB_val *key,
531 static size_t mdb_branch_size(MDB_env *env, MDB_val *key);
533 static int memncmp(const void *s1, size_t n1,
534 const void *s2, size_t n2);
535 static int memnrcmp(const void *s1, size_t n1,
536 const void *s2, size_t n2);
539 static SECURITY_DESCRIPTOR mdb_null_sd;
540 static SECURITY_ATTRIBUTES mdb_all_sa;
541 static int mdb_sec_inited;
545 memncmp(const void *s1, size_t n1, const void *s2, size_t n2)
547 int diff, len_diff = -1;
550 len_diff = (n1 > n2);
553 diff = memcmp(s1, s2, n1);
554 return diff ? diff : len_diff;
558 memnrcmp(const void *s1, size_t n1, const void *s2, size_t n2)
560 const unsigned char *p1, *p2, *p1_lim;
567 p1 = (const unsigned char *)s1 + n1 - 1;
568 p2 = (const unsigned char *)s2 + n2 - 1;
570 for (p1_lim = (n1 <= n2 ? s1 : s2); *p1 == *p2; p1--, p2--) {
572 return (p1 != s1) ? (p1 != p2) : (p2 != s2) ? -1 : 0;
578 mdb_version(int *maj, int *min, int *pat)
580 if (maj) *maj = MDB_VERSION_MAJOR;
581 if (min) *min = MDB_VERSION_MINOR;
582 if (pat) *pat = MDB_VERSION_PATCH;
583 return MDB_VERSION_STRING;
586 static char *const mdb_errstr[] = {
587 "MDB_KEYEXIST: Key/data pair already exists",
588 "MDB_NOTFOUND: No matching key/data pair found",
589 "MDB_PAGE_NOTFOUND: Requested page not found",
590 "MDB_CORRUPTED: Located page was wrong type",
591 "MDB_PANIC: Update of meta page failed",
592 "MDB_VERSION_MISMATCH: Database environment version mismatch"
596 mdb_strerror(int err)
599 return ("Successful return: 0");
601 if (err >= MDB_KEYEXIST && err <= MDB_VERSION_MISMATCH)
602 return mdb_errstr[err - MDB_KEYEXIST];
604 return strerror(err);
609 mdb_dkey(MDB_val *key, char *buf)
612 unsigned char *c = key->mv_data;
614 if (key->mv_size > MAXKEYSIZE)
617 for (i=0; i<key->mv_size; i++)
618 ptr += sprintf(ptr, "%02x", *c++);
620 sprintf(buf, "%.*s", key->mv_size, key->mv_data);
627 mdb_cmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
629 if (txn->mt_dbxs[dbi].md_cmp)
630 return txn->mt_dbxs[dbi].md_cmp(a, b);
632 if (txn->mt_dbs[dbi].md_flags & (MDB_REVERSEKEY
633 #if __BYTE_ORDER == __LITTLE_ENDIAN
637 return memnrcmp(a->mv_data, a->mv_size, b->mv_data, b->mv_size);
639 return memncmp((char *)a->mv_data, a->mv_size, b->mv_data, b->mv_size);
643 mdb_dcmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
645 if (txn->mt_dbxs[dbi].md_dcmp)
646 return txn->mt_dbxs[dbi].md_dcmp(a, b);
648 if (txn->mt_dbs[dbi].md_flags & (0
649 #if __BYTE_ORDER == __LITTLE_ENDIAN
653 return memnrcmp(a->mv_data, a->mv_size, b->mv_data, b->mv_size);
655 return memncmp((char *)a->mv_data, a->mv_size, b->mv_data, b->mv_size);
658 /* Allocate new page(s) for writing */
660 mdb_alloc_page(MDB_txn *txn, MDB_dbi dbi, MDB_page *parent, unsigned int parent_idx, int num)
663 pgno_t pgno = P_INVALID;
666 if (txn->mt_txnid > 2) {
668 if (!txn->mt_env->me_pghead && dbi != FREE_DBI &&
669 txn->mt_dbs[FREE_DBI].md_root != P_INVALID) {
670 /* See if there's anything in the free DB */
676 mpp.mp_parent = NULL;
679 mc.mc_dbi = FREE_DBI;
682 mdb_search_page(txn, FREE_DBI, NULL, &mc, 0, &mpp);
683 leaf = NODEPTR(mpp.mp_page, 0);
684 kptr = (ULONG *)NODEKEY(leaf);
688 oldest = txn->mt_txnid - 1;
689 for (i=0; i<txn->mt_env->me_txns->mti_numreaders; i++) {
690 ULONG mr = txn->mt_env->me_txns->mti_readers[i].mr_txnid;
691 if (mr && mr < oldest)
696 if (oldest > *kptr) {
697 /* It's usable, grab it.
704 mdb_read_data(txn, leaf, &data);
705 idl = (ULONG *)data.mv_data;
706 mop = malloc(sizeof(MDB_oldpages) + MDB_IDL_SIZEOF(idl) - sizeof(pgno_t));
707 mop->mo_next = txn->mt_env->me_pghead;
708 mop->mo_txnid = *kptr;
709 txn->mt_env->me_pghead = mop;
710 memcpy(mop->mo_pages, idl, MDB_IDL_SIZEOF(idl));
715 DPRINTF("IDL read txn %lu root %lu num %lu",
716 mop->mo_txnid, txn->mt_dbs[FREE_DBI].md_root, idl[0]);
717 for (i=0; i<idl[0]; i++) {
718 DPRINTF("IDL %lu", idl[i+1]);
722 /* drop this IDL from the DB */
723 top = CURSOR_TOP(&mc);
725 mc.mc_flags = C_INITIALIZED;
726 mdb_cursor_del(&mc, 0);
729 if (txn->mt_env->me_pghead) {
730 MDB_oldpages *mop = txn->mt_env->me_pghead;
732 /* FIXME: For now, always use fresh pages. We
733 * really ought to search the free list for a
738 /* peel pages off tail, so we only have to truncate the list */
739 pgno = MDB_IDL_LAST(mop->mo_pages);
740 if (MDB_IDL_IS_RANGE(mop->mo_pages)) {
742 if (mop->mo_pages[2] > mop->mo_pages[1])
743 mop->mo_pages[0] = 0;
747 if (MDB_IDL_IS_ZERO(mop->mo_pages)) {
748 txn->mt_env->me_pghead = mop->mo_next;
755 if (pgno == P_INVALID) {
756 /* DB size is maxed out */
757 if (txn->mt_next_pgno + num >= txn->mt_env->me_maxpg)
760 if (txn->mt_env->me_dpages && num == 1) {
761 dp = txn->mt_env->me_dpages;
762 txn->mt_env->me_dpages = (MDB_dpage *)dp->h.md_parent;
764 if ((dp = malloc(txn->mt_env->me_psize * num + sizeof(MDB_dhead))) == NULL)
768 dp->h.md_parent = parent;
769 dp->h.md_pi = parent_idx;
770 if (pgno == P_INVALID) {
771 dp->p.mp_pgno = txn->mt_next_pgno;
772 txn->mt_next_pgno += num;
774 dp->p.mp_pgno = pgno;
776 mid.mid = dp->p.mp_pgno;
778 mdb_mid2l_insert(txn->mt_u.dirty_list, &mid);
783 /* Touch a page: make it dirty and re-insert into tree with updated pgno.
786 mdb_touch(MDB_txn *txn, MDB_dbi dbi, MDB_pageparent *pp)
788 MDB_page *mp = pp->mp_page;
793 if (!F_ISSET(mp->mp_flags, P_DIRTY)) {
795 if ((dp = mdb_alloc_page(txn, dbi, pp->mp_parent, pp->mp_pi, 1)) == NULL)
797 DPRINTF("touched db %u page %lu -> %lu", dbi, mp->mp_pgno, dp->p.mp_pgno);
798 assert(mp->mp_pgno != dp->p.mp_pgno);
799 mdb_midl_insert(txn->mt_free_pgs, mp->mp_pgno);
800 pgno = dp->p.mp_pgno;
801 memcpy(&dp->p, mp, txn->mt_env->me_psize);
804 mp->mp_flags |= P_DIRTY;
806 /* Update the page number to new touched page. */
807 if (pp->mp_parent != NULL)
808 NODEPGNO(NODEPTR(pp->mp_parent, pp->mp_pi)) = mp->mp_pgno;
815 mdb_env_sync(MDB_env *env, int force)
818 if (force || !F_ISSET(env->me_flags, MDB_NOSYNC)) {
819 if (fdatasync(env->me_fd))
826 mdb_txn_reset0(MDB_txn *txn);
829 mdb_txn_renew0(MDB_txn *txn)
831 MDB_env *env = txn->mt_env;
833 if (txn->mt_flags & MDB_TXN_RDONLY) {
834 MDB_reader *r = pthread_getspecific(env->me_txkey);
837 pid_t pid = getpid();
838 pthread_t tid = pthread_self();
841 for (i=0; i<env->me_txns->mti_numreaders; i++)
842 if (env->me_txns->mti_readers[i].mr_pid == 0)
844 if (i == env->me_maxreaders) {
848 env->me_txns->mti_readers[i].mr_pid = pid;
849 env->me_txns->mti_readers[i].mr_tid = tid;
850 if (i >= env->me_txns->mti_numreaders)
851 env->me_txns->mti_numreaders = i+1;
853 r = &env->me_txns->mti_readers[i];
854 pthread_setspecific(env->me_txkey, r);
856 txn->mt_txnid = env->me_txns->mti_txnid;
857 txn->mt_toggle = env->me_txns->mti_me_toggle;
858 r->mr_txnid = txn->mt_txnid;
859 txn->mt_u.reader = r;
863 txn->mt_txnid = env->me_txns->mti_txnid+1;
864 txn->mt_toggle = env->me_txns->mti_me_toggle;
865 txn->mt_u.dirty_list = env->me_dirty_list;
866 txn->mt_u.dirty_list[0].mid = 0;
867 txn->mt_free_pgs = env->me_free_pgs;
868 txn->mt_free_pgs[0] = 0;
869 txn->mt_next_pgno = env->me_metas[txn->mt_toggle]->mm_last_pg+1;
873 /* Copy the DB arrays */
874 LAZY_RWLOCK_RDLOCK(&env->me_dblock);
875 txn->mt_numdbs = env->me_numdbs;
876 txn->mt_dbxs = env->me_dbxs; /* mostly static anyway */
877 memcpy(txn->mt_dbs, env->me_metas[txn->mt_toggle]->mm_dbs, 2 * sizeof(MDB_db));
878 if (txn->mt_numdbs > 2)
879 memcpy(txn->mt_dbs+2, env->me_dbs[env->me_db_toggle]+2,
880 (txn->mt_numdbs - 2) * sizeof(MDB_db));
881 LAZY_RWLOCK_UNLOCK(&env->me_dblock);
887 mdb_txn_renew(MDB_txn *txn)
894 if (txn->mt_env->me_flags & MDB_FATAL_ERROR) {
895 DPUTS("environment had fatal error, must shutdown!");
899 rc = mdb_txn_renew0(txn);
900 if (rc == MDB_SUCCESS) {
901 DPRINTF("renew txn %lu%c %p on mdbenv %p, root page %lu",
902 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w', txn,
903 (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root);
909 mdb_txn_begin(MDB_env *env, unsigned int flags, MDB_txn **ret)
914 if (env->me_flags & MDB_FATAL_ERROR) {
915 DPUTS("environment had fatal error, must shutdown!");
918 if ((txn = calloc(1, sizeof(MDB_txn) + env->me_maxdbs * sizeof(MDB_db))) == NULL) {
919 DPRINTF("calloc: %s", strerror(ErrCode()));
922 txn->mt_dbs = (MDB_db *)(txn+1);
923 if (flags & MDB_RDONLY) {
924 txn->mt_flags |= MDB_TXN_RDONLY;
928 rc = mdb_txn_renew0(txn);
933 DPRINTF("begin txn %lu%c %p on mdbenv %p, root page %lu",
934 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w', txn,
935 (void *) env, txn->mt_dbs[MAIN_DBI].md_root);
942 mdb_txn_reset0(MDB_txn *txn)
944 MDB_env *env = txn->mt_env;
946 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
947 txn->mt_u.reader->mr_txnid = 0;
953 /* return all dirty pages to dpage list */
954 for (i=1; i<=txn->mt_u.dirty_list[0].mid; i++) {
955 dp = txn->mt_u.dirty_list[i].mptr;
956 if (dp->h.md_num == 1) {
957 dp->h.md_parent = (MDB_page *)txn->mt_env->me_dpages;
958 txn->mt_env->me_dpages = dp;
960 /* large pages just get freed directly */
965 while ((mop = txn->mt_env->me_pghead)) {
966 txn->mt_env->me_pghead = mop->mo_next;
971 for (i=2; i<env->me_numdbs; i++)
972 env->me_dbxs[i].md_dirty = 0;
978 mdb_txn_reset(MDB_txn *txn)
983 DPRINTF("reset txn %lu%c %p on mdbenv %p, root page %lu",
984 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w', txn,
985 (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root);
991 mdb_txn_abort(MDB_txn *txn)
996 DPRINTF("abort txn %lu%c %p on mdbenv %p, root page %lu",
997 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w', txn,
998 (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root);
1000 mdb_txn_reset0(txn);
1005 mdb_txn_commit(MDB_txn *txn)
1016 assert(txn != NULL);
1017 assert(txn->mt_env != NULL);
1021 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
1026 if (txn != env->me_txn) {
1027 DPUTS("attempt to commit unknown transaction");
1032 if (F_ISSET(txn->mt_flags, MDB_TXN_ERROR)) {
1033 DPUTS("error flag is set, can't commit");
1038 if (!txn->mt_u.dirty_list[0].mid)
1041 DPRINTF("committing txn %lu %p on mdbenv %p, root page %lu",
1042 txn->mt_txnid, txn, (void *)env, txn->mt_dbs[MAIN_DBI].md_root);
1045 mc.mc_dbi = FREE_DBI;
1048 /* should only be one record now */
1049 if (env->me_pghead) {
1052 /* make sure first page of freeDB is touched and on freelist */
1053 mpp.mp_parent = NULL;
1056 mdb_search_page(txn, FREE_DBI, NULL, &mc, 1, &mpp);
1058 /* save to free list */
1059 if (!MDB_IDL_IS_ZERO(txn->mt_free_pgs)) {
1064 /* make sure last page of freeDB is touched and on freelist */
1065 key.mv_size = MAXKEYSIZE+1;
1067 mpp.mp_parent = NULL;
1070 mdb_search_page(txn, FREE_DBI, &key, &mc, 1, &mpp);
1075 ULONG *idl = txn->mt_free_pgs;
1076 DPRINTF("IDL write txn %lu root %lu num %lu",
1077 txn->mt_txnid, txn->mt_dbs[FREE_DBI].md_root, idl[0]);
1078 for (i=0; i<idl[0]; i++) {
1079 DPRINTF("IDL %lu", idl[i+1]);
1083 /* write to last page of freeDB */
1084 key.mv_size = sizeof(pgno_t);
1085 key.mv_data = (char *)&txn->mt_txnid;
1086 data.mv_data = txn->mt_free_pgs;
1087 /* The free list can still grow during this call,
1088 * despite the pre-emptive touches above. So check
1089 * and make sure the entire thing got written.
1092 i = txn->mt_free_pgs[0];
1093 data.mv_size = MDB_IDL_SIZEOF(txn->mt_free_pgs);
1094 rc = mdb_cursor_put(&mc, &key, &data, 0);
1099 } while (i != txn->mt_free_pgs[0]);
1101 /* should only be one record now */
1102 if (env->me_pghead) {
1106 mop = env->me_pghead;
1107 key.mv_size = sizeof(pgno_t);
1108 key.mv_data = (char *)&mop->mo_txnid;
1109 data.mv_size = MDB_IDL_SIZEOF(mop->mo_pages);
1110 data.mv_data = mop->mo_pages;
1111 mdb_cursor_put(&mc, &key, &data, 0);
1112 free(env->me_pghead);
1113 env->me_pghead = NULL;
1116 /* Update DB root pointers. Their pages have already been
1117 * touched so this is all in-place and cannot fail.
1121 data.mv_size = sizeof(MDB_db);
1123 mc.mc_dbi = MAIN_DBI;
1125 for (i = 2; i < txn->mt_numdbs; i++) {
1126 if (txn->mt_dbxs[i].md_dirty) {
1127 data.mv_data = &txn->mt_dbs[i];
1128 mdb_cursor_put(&mc, &txn->mt_dbxs[i].md_name, &data, 0);
1133 /* Commit up to MDB_COMMIT_PAGES dirty pages to disk until done.
1139 /* Windows actually supports scatter/gather I/O, but only on
1140 * unbuffered file handles. Since we're relying on the OS page
1141 * cache for all our data, that's self-defeating. So we just
1142 * write pages one at a time. We use the ov structure to set
1143 * the write offset, to at least save the overhead of a Seek
1147 memset(&ov, 0, sizeof(ov));
1148 for (; i<=txn->mt_u.dirty_list[0].mid; i++) {
1149 dp = txn->mt_u.dirty_list[i].mptr;
1150 DPRINTF("committing page %lu", dp->p.mp_pgno);
1151 size = dp->p.mp_pgno * env->me_psize;
1152 ov.Offset = size & 0xffffffff;
1153 ov.OffsetHigh = size >> 16;
1154 ov.OffsetHigh >>= 16;
1155 /* clear dirty flag */
1156 dp->p.mp_flags &= ~P_DIRTY;
1157 rc = WriteFile(env->me_fd, &dp->p, env->me_psize * dp->h.md_num, NULL, &ov);
1160 DPRINTF("WriteFile: %d", n);
1167 struct iovec iov[MDB_COMMIT_PAGES];
1171 for (; i<=txn->mt_u.dirty_list[0].mid; i++) {
1172 dp = txn->mt_u.dirty_list[i].mptr;
1173 if (dp->p.mp_pgno != next) {
1175 DPRINTF("committing %u dirty pages", n);
1176 rc = writev(env->me_fd, iov, n);
1180 DPUTS("short write, filesystem full?");
1182 DPRINTF("writev: %s", strerror(n));
1189 lseek(env->me_fd, dp->p.mp_pgno * env->me_psize, SEEK_SET);
1190 next = dp->p.mp_pgno;
1192 DPRINTF("committing page %lu", dp->p.mp_pgno);
1193 iov[n].iov_len = env->me_psize * dp->h.md_num;
1194 iov[n].iov_base = &dp->p;
1195 size += iov[n].iov_len;
1196 next = dp->p.mp_pgno + dp->h.md_num;
1197 /* clear dirty flag */
1198 dp->p.mp_flags &= ~P_DIRTY;
1199 if (++n >= MDB_COMMIT_PAGES) {
1209 DPRINTF("committing %u dirty pages", n);
1210 rc = writev(env->me_fd, iov, n);
1214 DPUTS("short write, filesystem full?");
1216 DPRINTF("writev: %s", strerror(n));
1223 /* Drop the dirty pages.
1225 for (i=1; i<=txn->mt_u.dirty_list[0].mid; i++) {
1226 dp = txn->mt_u.dirty_list[i].mptr;
1227 if (dp->h.md_num == 1) {
1228 dp->h.md_parent = (MDB_page *)txn->mt_env->me_dpages;
1229 txn->mt_env->me_dpages = dp;
1233 txn->mt_u.dirty_list[i].mid = 0;
1235 txn->mt_u.dirty_list[0].mid = 0;
1237 if ((n = mdb_env_sync(env, 0)) != 0 ||
1238 (n = mdb_env_write_meta(txn)) != MDB_SUCCESS) {
1245 /* update the DB tables */
1247 int toggle = !env->me_db_toggle;
1250 ip = &env->me_dbs[toggle][2];
1251 jp = &txn->mt_dbs[2];
1252 LAZY_RWLOCK_WRLOCK(&env->me_dblock);
1253 for (i = 2; i < txn->mt_numdbs; i++) {
1254 if (ip->md_root != jp->md_root)
1259 for (i = 2; i < txn->mt_numdbs; i++) {
1260 if (txn->mt_dbxs[i].md_dirty)
1261 txn->mt_dbxs[i].md_dirty = 0;
1263 env->me_db_toggle = toggle;
1264 env->me_numdbs = txn->mt_numdbs;
1265 LAZY_RWLOCK_UNLOCK(&env->me_dblock);
1268 UNLOCK_MUTEX_W(env);
1275 mdb_env_read_header(MDB_env *env, MDB_meta *meta)
1277 char page[PAGESIZE];
1282 /* We don't know the page size yet, so use a minimum value.
1286 if (!ReadFile(env->me_fd, page, PAGESIZE, (DWORD *)&rc, NULL) || rc == 0)
1288 if ((rc = read(env->me_fd, page, PAGESIZE)) == 0)
1293 else if (rc != PAGESIZE) {
1297 DPRINTF("read: %s", strerror(err));
1301 p = (MDB_page *)page;
1303 if (!F_ISSET(p->mp_flags, P_META)) {
1304 DPRINTF("page %lu not a meta page", p->mp_pgno);
1309 if (m->mm_magic != MDB_MAGIC) {
1310 DPUTS("meta has invalid magic");
1314 if (m->mm_version != MDB_VERSION) {
1315 DPRINTF("database is version %u, expected version %u",
1316 m->mm_version, MDB_VERSION);
1317 return MDB_VERSION_MISMATCH;
1320 memcpy(meta, m, sizeof(*m));
1325 mdb_env_init_meta(MDB_env *env, MDB_meta *meta)
1332 DPUTS("writing new meta page");
1336 meta->mm_magic = MDB_MAGIC;
1337 meta->mm_version = MDB_VERSION;
1338 meta->mm_psize = psize;
1339 meta->mm_last_pg = 1;
1340 meta->mm_flags = env->me_flags & 0xffff;
1341 meta->mm_flags |= MDB_INTEGERKEY;
1342 meta->mm_dbs[0].md_root = P_INVALID;
1343 meta->mm_dbs[1].md_root = P_INVALID;
1345 p = calloc(2, psize);
1347 p->mp_flags = P_META;
1350 memcpy(m, meta, sizeof(*meta));
1352 q = (MDB_page *)((char *)p + psize);
1355 q->mp_flags = P_META;
1358 memcpy(m, meta, sizeof(*meta));
1363 rc = WriteFile(env->me_fd, p, psize * 2, &len, NULL);
1364 rc = (len == psize * 2) ? MDB_SUCCESS : ErrCode();
1367 rc = write(env->me_fd, p, psize * 2);
1368 rc = (rc == (int)psize * 2) ? MDB_SUCCESS : ErrCode();
1375 mdb_env_write_meta(MDB_txn *txn)
1378 MDB_meta meta, metab;
1380 int rc, len, toggle;
1386 assert(txn != NULL);
1387 assert(txn->mt_env != NULL);
1389 toggle = !txn->mt_toggle;
1390 DPRINTF("writing meta page %d for root page %lu",
1391 toggle, txn->mt_dbs[MAIN_DBI].md_root);
1395 metab.mm_txnid = env->me_metas[toggle]->mm_txnid;
1396 metab.mm_last_pg = env->me_metas[toggle]->mm_last_pg;
1398 ptr = (char *)&meta;
1399 off = offsetof(MDB_meta, mm_dbs[0].md_depth);
1400 len = sizeof(MDB_meta) - off;
1403 meta.mm_dbs[0] = txn->mt_dbs[0];
1404 meta.mm_dbs[1] = txn->mt_dbs[1];
1405 meta.mm_last_pg = txn->mt_next_pgno - 1;
1406 meta.mm_txnid = txn->mt_txnid;
1409 off += env->me_psize;
1412 /* Write to the SYNC fd */
1415 memset(&ov, 0, sizeof(ov));
1417 WriteFile(env->me_mfd, ptr, len, (DWORD *)&rc, &ov);
1420 rc = pwrite(env->me_mfd, ptr, len, off);
1425 DPUTS("write failed, disk error?");
1426 /* On a failure, the pagecache still contains the new data.
1427 * Write some old data back, to prevent it from being used.
1428 * Use the non-SYNC fd; we know it will fail anyway.
1430 meta.mm_last_pg = metab.mm_last_pg;
1431 meta.mm_txnid = metab.mm_txnid;
1433 WriteFile(env->me_fd, ptr, len, NULL, &ov);
1435 r2 = pwrite(env->me_fd, ptr, len, off);
1437 env->me_flags |= MDB_FATAL_ERROR;
1440 /* Memory ordering issues are irrelevant; since the entire writer
1441 * is wrapped by wmutex, all of these changes will become visible
1442 * after the wmutex is unlocked. Since the DB is multi-version,
1443 * readers will get consistent data regardless of how fresh or
1444 * how stale their view of these values is.
1446 LAZY_MUTEX_LOCK(&env->me_txns->mti_mutex);
1447 txn->mt_env->me_txns->mti_me_toggle = toggle;
1448 txn->mt_env->me_txns->mti_txnid = txn->mt_txnid;
1449 LAZY_MUTEX_UNLOCK(&env->me_txns->mti_mutex);
1455 mdb_env_read_meta(MDB_env *env, int *which)
1459 assert(env != NULL);
1461 if (env->me_metas[0]->mm_txnid < env->me_metas[1]->mm_txnid)
1464 DPRINTF("Using meta page %d", toggle);
1471 mdb_env_create(MDB_env **env)
1475 e = calloc(1, sizeof(MDB_env));
1476 if (!e) return ENOMEM;
1478 e->me_maxreaders = DEFAULT_READERS;
1480 e->me_fd = INVALID_HANDLE_VALUE;
1481 e->me_lfd = INVALID_HANDLE_VALUE;
1482 e->me_mfd = INVALID_HANDLE_VALUE;
1488 mdb_env_set_mapsize(MDB_env *env, size_t size)
1492 env->me_mapsize = size;
1497 mdb_env_set_maxdbs(MDB_env *env, int dbs)
1501 env->me_maxdbs = dbs;
1506 mdb_env_set_maxreaders(MDB_env *env, int readers)
1510 env->me_maxreaders = readers;
1515 mdb_env_get_maxreaders(MDB_env *env, int *readers)
1517 if (!env || !readers)
1519 *readers = env->me_maxreaders;
1524 mdb_env_open2(MDB_env *env, unsigned int flags)
1526 int i, newenv = 0, toggle;
1530 env->me_flags = flags;
1532 memset(&meta, 0, sizeof(meta));
1534 if ((i = mdb_env_read_header(env, &meta)) != 0) {
1537 DPUTS("new mdbenv");
1541 if (!env->me_mapsize) {
1542 env->me_mapsize = newenv ? DEFAULT_MAPSIZE : meta.mm_mapsize;
1548 LONG sizelo, sizehi;
1549 sizelo = env->me_mapsize & 0xffffffff;
1550 sizehi = env->me_mapsize >> 16; /* pointless on WIN32, only needed on W64 */
1552 /* Windows won't create mappings for zero length files.
1553 * Just allocate the maxsize right now.
1556 SetFilePointer(env->me_fd, sizelo, sizehi ? &sizehi : NULL, 0);
1557 if (!SetEndOfFile(env->me_fd))
1559 SetFilePointer(env->me_fd, 0, NULL, 0);
1561 mh = CreateFileMapping(env->me_fd, NULL, PAGE_READONLY,
1562 sizehi, sizelo, NULL);
1565 env->me_map = MapViewOfFileEx(mh, FILE_MAP_READ, 0, 0, env->me_mapsize,
1573 if (meta.mm_address && (flags & MDB_FIXEDMAP))
1575 env->me_map = mmap(meta.mm_address, env->me_mapsize, PROT_READ, i,
1577 if (env->me_map == MAP_FAILED)
1582 meta.mm_mapsize = env->me_mapsize;
1583 if (flags & MDB_FIXEDMAP)
1584 meta.mm_address = env->me_map;
1585 i = mdb_env_init_meta(env, &meta);
1586 if (i != MDB_SUCCESS) {
1587 munmap(env->me_map, env->me_mapsize);
1591 env->me_psize = meta.mm_psize;
1593 env->me_maxpg = env->me_mapsize / env->me_psize;
1595 p = (MDB_page *)env->me_map;
1596 env->me_metas[0] = METADATA(p);
1597 env->me_metas[1] = (MDB_meta *)((char *)env->me_metas[0] + meta.mm_psize);
1599 if ((i = mdb_env_read_meta(env, &toggle)) != 0)
1602 DPRINTF("opened database version %u, pagesize %u",
1603 env->me_metas[toggle]->mm_version, env->me_psize);
1604 DPRINTF("depth: %u", env->me_metas[toggle]->mm_dbs[MAIN_DBI].md_depth);
1605 DPRINTF("entries: %lu", env->me_metas[toggle]->mm_dbs[MAIN_DBI].md_entries);
1606 DPRINTF("branch pages: %lu", env->me_metas[toggle]->mm_dbs[MAIN_DBI].md_branch_pages);
1607 DPRINTF("leaf pages: %lu", env->me_metas[toggle]->mm_dbs[MAIN_DBI].md_leaf_pages);
1608 DPRINTF("overflow pages: %lu", env->me_metas[toggle]->mm_dbs[MAIN_DBI].md_overflow_pages);
1609 DPRINTF("root: %lu", env->me_metas[toggle]->mm_dbs[MAIN_DBI].md_root);
1615 /* Windows doesn't support destructor callbacks for thread-specific storage */
1617 mdb_env_reader_dest(void *ptr)
1619 MDB_reader *reader = ptr;
1621 reader->mr_txnid = 0;
1627 /* downgrade the exclusive lock on the region back to shared */
1629 mdb_env_share_locks(MDB_env *env)
1633 if (env->me_metas[0]->mm_txnid < env->me_metas[1]->mm_txnid)
1635 env->me_txns->mti_me_toggle = toggle;
1636 env->me_txns->mti_txnid = env->me_metas[toggle]->mm_txnid;
1641 /* First acquire a shared lock. The Unlock will
1642 * then release the existing exclusive lock.
1644 memset(&ov, 0, sizeof(ov));
1645 LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov);
1646 UnlockFile(env->me_lfd, 0, 0, 1, 0);
1650 struct flock lock_info;
1651 /* The shared lock replaces the existing lock */
1652 memset((void *)&lock_info, 0, sizeof(lock_info));
1653 lock_info.l_type = F_RDLCK;
1654 lock_info.l_whence = SEEK_SET;
1655 lock_info.l_start = 0;
1656 lock_info.l_len = 1;
1657 fcntl(env->me_lfd, F_SETLK, &lock_info);
1663 mdb_env_setup_locks(MDB_env *env, char *lpath, int mode, int *excl)
1671 if ((env->me_lfd = CreateFile(lpath, GENERIC_READ|GENERIC_WRITE,
1672 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, OPEN_ALWAYS,
1673 FILE_ATTRIBUTE_NORMAL, NULL)) == INVALID_HANDLE_VALUE) {
1677 /* Try to get exclusive lock. If we succeed, then
1678 * nobody is using the lock region and we should initialize it.
1681 if (LockFile(env->me_lfd, 0, 0, 1, 0)) {
1685 memset(&ov, 0, sizeof(ov));
1686 if (!LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
1692 size = GetFileSize(env->me_lfd, NULL);
1694 if ((env->me_lfd = open(lpath, O_RDWR|O_CREAT, mode)) == -1) {
1698 /* Try to get exclusive lock. If we succeed, then
1699 * nobody is using the lock region and we should initialize it.
1702 struct flock lock_info;
1703 memset((void *)&lock_info, 0, sizeof(lock_info));
1704 lock_info.l_type = F_WRLCK;
1705 lock_info.l_whence = SEEK_SET;
1706 lock_info.l_start = 0;
1707 lock_info.l_len = 1;
1708 rc = fcntl(env->me_lfd, F_SETLK, &lock_info);
1712 lock_info.l_type = F_RDLCK;
1713 rc = fcntl(env->me_lfd, F_SETLKW, &lock_info);
1720 size = lseek(env->me_lfd, 0, SEEK_END);
1722 rsize = (env->me_maxreaders-1) * sizeof(MDB_reader) + sizeof(MDB_txninfo);
1723 if (size < rsize && *excl) {
1725 SetFilePointer(env->me_lfd, rsize, NULL, 0);
1726 if (!SetEndOfFile(env->me_lfd)) {
1731 if (ftruncate(env->me_lfd, rsize) != 0) {
1738 size = rsize - sizeof(MDB_txninfo);
1739 env->me_maxreaders = size/sizeof(MDB_reader) + 1;
1744 mh = CreateFileMapping(env->me_lfd, NULL, PAGE_READWRITE,
1750 env->me_txns = MapViewOfFileEx(mh, FILE_MAP_WRITE, 0, 0, rsize, NULL);
1752 if (!env->me_txns) {
1758 env->me_txns = mmap(0, rsize, PROT_READ|PROT_WRITE, MAP_SHARED,
1760 if (env->me_txns == MAP_FAILED) {
1768 if (!mdb_sec_inited) {
1769 InitializeSecurityDescriptor(&mdb_null_sd,
1770 SECURITY_DESCRIPTOR_REVISION);
1771 SetSecurityDescriptorDacl(&mdb_null_sd, TRUE, 0, FALSE);
1772 mdb_all_sa.nLength = sizeof(SECURITY_ATTRIBUTES);
1773 mdb_all_sa.bInheritHandle = FALSE;
1774 mdb_all_sa.lpSecurityDescriptor = &mdb_null_sd;
1777 /* FIXME: only using up to 20 characters of the env path here,
1778 * probably not enough to assure uniqueness...
1780 sprintf(env->me_txns->mti_rmname, "Global\\MDBr%.20s", lpath);
1781 ptr = env->me_txns->mti_rmname + sizeof("Global\\MDBr");
1782 while ((ptr = strchr(ptr, '\\')))
1784 env->me_rmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_rmname);
1785 if (!env->me_rmutex) {
1789 sprintf(env->me_txns->mti_rmname, "Global\\MDBw%.20s", lpath);
1790 ptr = env->me_txns->mti_rmname + sizeof("Global\\MDBw");
1791 while ((ptr = strchr(ptr, '\\')))
1793 env->me_wmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_rmname);
1794 if (!env->me_wmutex) {
1799 pthread_mutexattr_t mattr;
1801 pthread_mutexattr_init(&mattr);
1802 rc = pthread_mutexattr_setpshared(&mattr, PTHREAD_PROCESS_SHARED);
1806 pthread_mutex_init(&env->me_txns->mti_mutex, &mattr);
1807 pthread_mutex_init(&env->me_txns->mti_wmutex, &mattr);
1809 env->me_txns->mti_version = MDB_VERSION;
1810 env->me_txns->mti_magic = MDB_MAGIC;
1811 env->me_txns->mti_txnid = 0;
1812 env->me_txns->mti_numreaders = 0;
1813 env->me_txns->mti_me_toggle = 0;
1816 if (env->me_txns->mti_magic != MDB_MAGIC) {
1817 DPUTS("lock region has invalid magic");
1821 if (env->me_txns->mti_version != MDB_VERSION) {
1822 DPRINTF("lock region is version %u, expected version %u",
1823 env->me_txns->mti_version, MDB_VERSION);
1824 rc = MDB_VERSION_MISMATCH;
1828 if (rc != EACCES && rc != EAGAIN) {
1832 env->me_rmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_rmname);
1833 if (!env->me_rmutex) {
1837 env->me_wmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_wmname);
1838 if (!env->me_wmutex) {
1848 env->me_lfd = INVALID_HANDLE_VALUE;
1853 #define LOCKNAME "/lock.mdb"
1854 #define DATANAME "/data.mdb"
1856 mdb_env_open(MDB_env *env, const char *path, unsigned int flags, mode_t mode)
1858 int oflags, rc, len, excl;
1859 char *lpath, *dpath;
1862 lpath = malloc(len + sizeof(LOCKNAME) + len + sizeof(DATANAME));
1865 dpath = lpath + len + sizeof(LOCKNAME);
1866 sprintf(lpath, "%s" LOCKNAME, path);
1867 sprintf(dpath, "%s" DATANAME, path);
1869 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
1874 if (F_ISSET(flags, MDB_RDONLY)) {
1875 oflags = GENERIC_READ;
1876 len = OPEN_EXISTING;
1878 oflags = GENERIC_READ|GENERIC_WRITE;
1881 mode = FILE_ATTRIBUTE_NORMAL;
1882 if ((env->me_fd = CreateFile(dpath, oflags, FILE_SHARE_READ|FILE_SHARE_WRITE,
1883 NULL, len, mode, NULL)) == INVALID_HANDLE_VALUE) {
1888 if (F_ISSET(flags, MDB_RDONLY))
1891 oflags = O_RDWR | O_CREAT;
1893 if ((env->me_fd = open(dpath, oflags, mode)) == -1) {
1899 if ((rc = mdb_env_open2(env, flags)) == MDB_SUCCESS) {
1900 /* synchronous fd for meta writes */
1902 if (!(flags & (MDB_RDONLY|MDB_NOSYNC)))
1903 mode |= FILE_FLAG_WRITE_THROUGH;
1904 if ((env->me_mfd = CreateFile(dpath, oflags, FILE_SHARE_READ|FILE_SHARE_WRITE,
1905 NULL, len, mode, NULL)) == INVALID_HANDLE_VALUE) {
1910 if (!(flags & (MDB_RDONLY|MDB_NOSYNC)))
1911 oflags |= MDB_DSYNC;
1912 if ((env->me_mfd = open(dpath, oflags, mode)) == -1) {
1917 env->me_path = strdup(path);
1918 DPRINTF("opened dbenv %p", (void *) env);
1919 pthread_key_create(&env->me_txkey, mdb_env_reader_dest);
1920 LAZY_RWLOCK_INIT(&env->me_dblock, NULL);
1922 mdb_env_share_locks(env);
1923 env->me_dbxs = calloc(env->me_maxdbs, sizeof(MDB_dbx));
1924 env->me_dbs[0] = calloc(env->me_maxdbs, sizeof(MDB_db));
1925 env->me_dbs[1] = calloc(env->me_maxdbs, sizeof(MDB_db));
1931 if (env->me_fd != INVALID_HANDLE_VALUE) {
1933 env->me_fd = INVALID_HANDLE_VALUE;
1935 if (env->me_lfd != INVALID_HANDLE_VALUE) {
1937 env->me_lfd = INVALID_HANDLE_VALUE;
1945 mdb_env_close(MDB_env *env)
1952 while (env->me_dpages) {
1953 dp = env->me_dpages;
1954 env->me_dpages = (MDB_dpage *)dp->h.md_parent;
1958 free(env->me_dbs[1]);
1959 free(env->me_dbs[0]);
1963 LAZY_RWLOCK_DESTROY(&env->me_dblock);
1964 pthread_key_delete(env->me_txkey);
1967 munmap(env->me_map, env->me_mapsize);
1972 pid_t pid = getpid();
1974 for (i=0; i<env->me_txns->mti_numreaders; i++)
1975 if (env->me_txns->mti_readers[i].mr_pid == pid)
1976 env->me_txns->mti_readers[i].mr_pid = 0;
1977 munmap(env->me_txns, (env->me_maxreaders-1)*sizeof(MDB_reader)+sizeof(MDB_txninfo));
1983 /* Search for key within a leaf page, using binary search.
1984 * Returns the smallest entry larger or equal to the key.
1985 * If exactp is non-null, stores whether the found entry was an exact match
1986 * in *exactp (1 or 0).
1987 * If kip is non-null, stores the index of the found entry in *kip.
1988 * If no entry larger or equal to the key is found, returns NULL.
1991 mdb_search_node(MDB_txn *txn, MDB_dbi dbi, MDB_page *mp, MDB_val *key,
1992 int *exactp, unsigned int *kip)
1997 MDB_node *node = NULL;
2001 DPRINTF("searching %u keys in %s page %lu",
2003 IS_LEAF(mp) ? "leaf" : "branch",
2006 assert(NUMKEYS(mp) > 0);
2008 memset(&nodekey, 0, sizeof(nodekey));
2010 low = IS_LEAF(mp) ? 0 : 1;
2011 high = NUMKEYS(mp) - 1;
2012 while (low <= high) {
2013 i = (low + high) >> 1;
2016 nodekey.mv_size = txn->mt_dbs[dbi].md_pad;
2017 nodekey.mv_data = LEAF2KEY(mp, i, nodekey.mv_size);
2019 node = NODEPTR(mp, i);
2021 nodekey.mv_size = node->mn_ksize;
2022 nodekey.mv_data = NODEKEY(node);
2025 rc = mdb_cmp(txn, dbi, key, &nodekey);
2028 DPRINTF("found leaf index %u [%s], rc = %i",
2029 i, DKEY(&nodekey), rc);
2031 DPRINTF("found branch index %u [%s -> %lu], rc = %i",
2032 i, DKEY(&nodekey), NODEPGNO(node), rc);
2042 if (rc > 0) { /* Found entry is less than the key. */
2043 i++; /* Skip to get the smallest entry larger than key. */
2046 *exactp = (rc == 0);
2047 if (kip) /* Store the key index if requested. */
2049 if (i >= NUMKEYS(mp))
2050 /* There is no entry larger or equal to the key. */
2053 /* nodeptr is fake for LEAF2 */
2054 return IS_LEAF2(mp) ? NODEPTR(mp, 0) : NODEPTR(mp, i);
2058 cursor_pop_page(MDB_cursor *cursor)
2062 if (cursor->mc_snum) {
2063 top = CURSOR_TOP(cursor);
2066 DPRINTF("popped page %lu off db %u cursor %p", top->mp_page->mp_pgno,
2067 cursor->mc_dbi, (void *) cursor);
2072 cursor_push_page(MDB_cursor *cursor, MDB_page *mp)
2076 DPRINTF("pushing page %lu on db %u cursor %p", mp->mp_pgno,
2077 cursor->mc_dbi, (void *) cursor);
2079 assert(cursor->mc_snum < CURSOR_STACK);
2081 ppage = &cursor->mc_stack[cursor->mc_snum++];
2082 ppage->mp_page = mp;
2088 mdb_get_page(MDB_txn *txn, pgno_t pgno, MDB_page **ret)
2092 if (!F_ISSET(txn->mt_flags, MDB_TXN_RDONLY) && txn->mt_u.dirty_list[0].mid) {
2095 x = mdb_mid2l_search(txn->mt_u.dirty_list, pgno);
2096 if (x <= txn->mt_u.dirty_list[0].mid && txn->mt_u.dirty_list[x].mid == pgno) {
2097 dp = txn->mt_u.dirty_list[x].mptr;
2102 if (pgno <= txn->mt_env->me_metas[txn->mt_toggle]->mm_last_pg)
2103 p = (MDB_page *)(txn->mt_env->me_map + txn->mt_env->me_psize * pgno);
2107 DPRINTF("page %lu not found", pgno);
2110 return (p != NULL) ? MDB_SUCCESS : MDB_PAGE_NOTFOUND;
2114 mdb_search_page_root(MDB_txn *txn, MDB_dbi dbi, MDB_val *key,
2115 MDB_cursor *cursor, int modify, MDB_pageparent *mpp)
2117 MDB_page *mp = mpp->mp_page;
2121 if (cursor && cursor_push_page(cursor, mp) == NULL)
2124 while (IS_BRANCH(mp)) {
2128 DPRINTF("branch page %lu has %u keys", mp->mp_pgno, NUMKEYS(mp));
2129 assert(NUMKEYS(mp) > 1);
2130 DPRINTF("found index 0 to page %lu", NODEPGNO(NODEPTR(mp, 0)));
2132 if (key == NULL) /* Initialize cursor to first page. */
2134 else if (key->mv_size > MAXKEYSIZE && key->mv_data == NULL) {
2135 /* cursor to last page */
2139 node = mdb_search_node(txn, dbi, mp, key, &exact, &i);
2141 i = NUMKEYS(mp) - 1;
2149 DPRINTF("following index %u for key [%s]",
2151 assert(i < NUMKEYS(mp));
2152 node = NODEPTR(mp, i);
2155 CURSOR_TOP(cursor)->mp_ki = i;
2157 mpp->mp_parent = mp;
2158 if ((rc = mdb_get_page(txn, NODEPGNO(node), &mp)))
2163 if (cursor && cursor_push_page(cursor, mp) == NULL)
2167 MDB_dhead *dh = ((MDB_dhead *)mp)-1;
2168 if ((rc = mdb_touch(txn, dbi, mpp)) != 0)
2170 dh = ((MDB_dhead *)mpp->mp_page)-1;
2171 dh->md_parent = mpp->mp_parent;
2172 dh->md_pi = mpp->mp_pi;
2179 DPRINTF("internal error, index points to a %02X page!?",
2181 return MDB_CORRUPTED;
2184 DPRINTF("found leaf page %lu for key [%s]", mp->mp_pgno,
2185 key ? DKEY(key) : NULL);
2190 /* Search for the page a given key should be in.
2191 * Stores a pointer to the found page in *mpp.
2192 * If key is NULL, search for the lowest page (used by mdb_cursor_first).
2193 * If cursor is non-null, pushes parent pages on the cursor stack.
2194 * If modify is true, visited pages are updated with new page numbers.
2197 mdb_search_page(MDB_txn *txn, MDB_dbi dbi, MDB_val *key,
2198 MDB_cursor *cursor, int modify, MDB_pageparent *mpp)
2203 /* Choose which root page to start with. If a transaction is given
2204 * use the root page from the transaction, otherwise read the last
2205 * committed root page.
2207 if (F_ISSET(txn->mt_flags, MDB_TXN_ERROR)) {
2208 DPUTS("transaction has failed, must abort");
2211 root = txn->mt_dbs[dbi].md_root;
2213 if (root == P_INVALID) { /* Tree is empty. */
2214 DPUTS("tree is empty");
2215 return MDB_NOTFOUND;
2218 if ((rc = mdb_get_page(txn, root, &mpp->mp_page)))
2221 DPRINTF("db %u root page %lu has flags 0x%X",
2222 dbi, root, mpp->mp_page->mp_flags);
2225 /* For sub-databases, update main root first */
2226 if (dbi > MAIN_DBI && !txn->mt_dbxs[dbi].md_dirty) {
2228 rc = mdb_search_page(txn, MAIN_DBI, &txn->mt_dbxs[dbi].md_name,
2232 txn->mt_dbxs[dbi].md_dirty = 1;
2234 if (!F_ISSET(mpp->mp_page->mp_flags, P_DIRTY)) {
2235 mpp->mp_parent = NULL;
2237 if ((rc = mdb_touch(txn, dbi, mpp)))
2239 txn->mt_dbs[dbi].md_root = mpp->mp_page->mp_pgno;
2243 return mdb_search_page_root(txn, dbi, key, cursor, modify, mpp);
2247 mdb_read_data(MDB_txn *txn, MDB_node *leaf, MDB_val *data)
2249 MDB_page *omp; /* overflow mpage */
2253 if (!F_ISSET(leaf->mn_flags, F_BIGDATA)) {
2254 data->mv_size = leaf->mn_dsize;
2255 data->mv_data = NODEDATA(leaf);
2259 /* Read overflow data.
2261 data->mv_size = leaf->mn_dsize;
2262 memcpy(&pgno, NODEDATA(leaf), sizeof(pgno));
2263 if ((rc = mdb_get_page(txn, pgno, &omp))) {
2264 DPRINTF("read overflow page %lu failed", pgno);
2267 data->mv_data = METADATA(omp);
2273 mdb_get(MDB_txn *txn, MDB_dbi dbi,
2274 MDB_val *key, MDB_val *data)
2283 DPRINTF("===> get db %u key [%s]", dbi, DKEY(key));
2285 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
2288 if (key->mv_size == 0 || key->mv_size > MAXKEYSIZE) {
2295 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
2296 mc.mc_xcursor = &mx;
2297 mdb_xcursor_init0(&mc);
2299 mc.mc_xcursor = NULL;
2301 return mdb_cursor_set(&mc, key, data, MDB_SET, &exact);
2305 mdb_sibling(MDB_cursor *cursor, int move_right)
2312 if (cursor->mc_snum < 2) {
2313 return MDB_NOTFOUND; /* root has no siblings */
2315 parent = CURSOR_PARENT(cursor);
2317 DPRINTF("parent page is page %lu, index %u",
2318 parent->mp_page->mp_pgno, parent->mp_ki);
2320 cursor_pop_page(cursor);
2321 if (move_right ? (parent->mp_ki + 1 >= NUMKEYS(parent->mp_page))
2322 : (parent->mp_ki == 0)) {
2323 DPRINTF("no more keys left, moving to %s sibling",
2324 move_right ? "right" : "left");
2325 if ((rc = mdb_sibling(cursor, move_right)) != MDB_SUCCESS)
2327 parent = CURSOR_TOP(cursor);
2333 DPRINTF("just moving to %s index key %u",
2334 move_right ? "right" : "left", parent->mp_ki);
2336 assert(IS_BRANCH(parent->mp_page));
2338 indx = NODEPTR(parent->mp_page, parent->mp_ki);
2339 if ((rc = mdb_get_page(cursor->mc_txn, NODEPGNO(indx), &mp)))
2342 mp->parent = parent->mp_page;
2343 mp->parent_index = parent->mp_ki;
2346 cursor_push_page(cursor, mp);
2352 mdb_cursor_next(MDB_cursor *cursor, MDB_val *key, MDB_val *data, MDB_cursor_op op)
2359 if (cursor->mc_flags & C_EOF) {
2360 return MDB_NOTFOUND;
2363 assert(cursor->mc_flags & C_INITIALIZED);
2365 top = CURSOR_TOP(cursor);
2368 if (cursor->mc_txn->mt_dbs[cursor->mc_dbi].md_flags & MDB_DUPSORT) {
2369 leaf = NODEPTR(mp, top->mp_ki);
2370 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
2371 if (op == MDB_NEXT || op == MDB_NEXT_DUP) {
2372 rc = mdb_cursor_next(&cursor->mc_xcursor->mx_cursor, data, NULL, MDB_NEXT);
2373 if (op != MDB_NEXT || rc == MDB_SUCCESS)
2377 cursor->mc_xcursor->mx_cursor.mc_flags = 0;
2378 if (op == MDB_NEXT_DUP)
2379 return MDB_NOTFOUND;
2383 if (cursor->mc_flags & C_XDIRTY)
2384 mdb_cursor_reset(cursor);
2386 DPRINTF("cursor_next: top page is %lu in cursor %p", mp->mp_pgno, (void *) cursor);
2388 if (top->mp_ki + 1 >= NUMKEYS(mp)) {
2389 DPUTS("=====> move to next sibling page");
2390 if (mdb_sibling(cursor, 1) != MDB_SUCCESS) {
2391 cursor->mc_flags |= C_EOF;
2392 return MDB_NOTFOUND;
2394 top = CURSOR_TOP(cursor);
2396 DPRINTF("next page is %lu, key index %u", mp->mp_pgno, top->mp_ki);
2400 DPRINTF("==> cursor points to page %lu with %u keys, key index %u",
2401 mp->mp_pgno, NUMKEYS(mp), top->mp_ki);
2404 key->mv_size = cursor->mc_txn->mt_dbs[cursor->mc_dbi].md_pad;
2405 key->mv_data = LEAF2KEY(mp, top->mp_ki, key->mv_size);
2409 assert(IS_LEAF(mp));
2410 leaf = NODEPTR(mp, top->mp_ki);
2412 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
2413 mdb_xcursor_init1(cursor, mp, leaf);
2416 if ((rc = mdb_read_data(cursor->mc_txn, leaf, data) != MDB_SUCCESS))
2419 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
2420 rc = mdb_cursor_first(&cursor->mc_xcursor->mx_cursor, data, NULL);
2421 if (rc != MDB_SUCCESS)
2426 MDB_SET_KEY(leaf, key);
2431 mdb_cursor_prev(MDB_cursor *cursor, MDB_val *key, MDB_val *data, MDB_cursor_op op)
2438 assert(cursor->mc_flags & C_INITIALIZED);
2440 top = CURSOR_TOP(cursor);
2443 if (cursor->mc_txn->mt_dbs[cursor->mc_dbi].md_flags & MDB_DUPSORT) {
2444 leaf = NODEPTR(mp, top->mp_ki);
2445 if (op == MDB_PREV || op == MDB_PREV_DUP) {
2446 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
2447 rc = mdb_cursor_prev(&cursor->mc_xcursor->mx_cursor, data, NULL, MDB_PREV);
2448 if (op != MDB_PREV || rc == MDB_SUCCESS)
2451 cursor->mc_xcursor->mx_cursor.mc_flags = 0;
2452 if (op == MDB_PREV_DUP)
2453 return MDB_NOTFOUND;
2458 if (cursor->mc_flags & C_XDIRTY)
2459 mdb_cursor_reset(cursor);
2461 DPRINTF("cursor_prev: top page is %lu in cursor %p", mp->mp_pgno, (void *) cursor);
2463 if (top->mp_ki == 0) {
2464 DPUTS("=====> move to prev sibling page");
2465 if (mdb_sibling(cursor, 0) != MDB_SUCCESS) {
2466 cursor->mc_flags &= ~C_INITIALIZED;
2467 return MDB_NOTFOUND;
2469 top = CURSOR_TOP(cursor);
2471 top->mp_ki = NUMKEYS(mp) - 1;
2472 DPRINTF("prev page is %lu, key index %u", mp->mp_pgno, top->mp_ki);
2476 cursor->mc_flags &= ~C_EOF;
2478 DPRINTF("==> cursor points to page %lu with %u keys, key index %u",
2479 mp->mp_pgno, NUMKEYS(mp), top->mp_ki);
2482 key->mv_size = cursor->mc_txn->mt_dbs[cursor->mc_dbi].md_pad;
2483 key->mv_data = LEAF2KEY(mp, top->mp_ki, key->mv_size);
2487 assert(IS_LEAF(mp));
2488 leaf = NODEPTR(mp, top->mp_ki);
2490 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
2491 mdb_xcursor_init1(cursor, mp, leaf);
2494 if ((rc = mdb_read_data(cursor->mc_txn, leaf, data) != MDB_SUCCESS))
2497 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
2498 rc = mdb_cursor_last(&cursor->mc_xcursor->mx_cursor, data, NULL);
2499 if (rc != MDB_SUCCESS)
2504 MDB_SET_KEY(leaf, key);
2509 mdb_cursor_set(MDB_cursor *cursor, MDB_val *key, MDB_val *data,
2510 MDB_cursor_op op, int *exactp)
2520 assert(key->mv_size > 0);
2522 /* See if we're already on the right page */
2523 if (cursor->mc_flags & C_INITIALIZED) {
2525 top = CURSOR_TOP(cursor);
2526 /* Don't try this for LEAF2 pages. Maybe support that later. */
2527 if ((top->mp_page->mp_flags & (P_LEAF|P_LEAF2)) == P_LEAF) {
2528 leaf = NODEPTR(top->mp_page, 0);
2529 MDB_SET_KEY(leaf, &nodekey);
2530 rc = mdb_cmp(cursor->mc_txn, cursor->mc_dbi, key, &nodekey);
2532 leaf = NODEPTR(top->mp_page, NUMKEYS(top->mp_page)-1);
2533 MDB_SET_KEY(leaf, &nodekey);
2534 rc = mdb_cmp(cursor->mc_txn, cursor->mc_dbi, key, &nodekey);
2536 /* we're already on the right page */
2537 mpp.mp_page = top->mp_page;
2544 if (cursor->mc_flags & C_XDIRTY)
2545 mdb_cursor_reset(cursor);
2547 cursor->mc_snum = 0;
2549 rc = mdb_search_page(cursor->mc_txn, cursor->mc_dbi, key, cursor, 0, &mpp);
2550 if (rc != MDB_SUCCESS)
2553 assert(IS_LEAF(mpp.mp_page));
2555 top = CURSOR_TOP(cursor);
2557 leaf = mdb_search_node(cursor->mc_txn, cursor->mc_dbi, mpp.mp_page, key, exactp, &top->mp_ki);
2558 if (exactp != NULL && !*exactp) {
2559 /* MDB_SET specified and not an exact match. */
2560 return MDB_NOTFOUND;
2564 DPUTS("===> inexact leaf not found, goto sibling");
2565 if ((rc = mdb_sibling(cursor, 1)) != MDB_SUCCESS)
2566 return rc; /* no entries matched */
2567 top = CURSOR_TOP(cursor);
2569 mpp.mp_page = top->mp_page;
2570 assert(IS_LEAF(mpp.mp_page));
2571 leaf = NODEPTR(mpp.mp_page, 0);
2574 cursor->mc_flags |= C_INITIALIZED;
2575 cursor->mc_flags &= ~C_EOF;
2577 if (IS_LEAF2(mpp.mp_page)) {
2578 key->mv_size = cursor->mc_txn->mt_dbs[cursor->mc_dbi].md_pad;
2579 key->mv_data = LEAF2KEY(mpp.mp_page, top->mp_ki, key->mv_size);
2583 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
2584 mdb_xcursor_init1(cursor, mpp.mp_page, leaf);
2587 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
2588 if (op == MDB_SET || op == MDB_SET_RANGE) {
2589 rc = mdb_cursor_first(&cursor->mc_xcursor->mx_cursor, data, NULL);
2592 if (op == MDB_GET_BOTH) {
2598 rc = mdb_cursor_set(&cursor->mc_xcursor->mx_cursor, data, NULL, MDB_SET_RANGE, ex2p);
2599 if (rc != MDB_SUCCESS)
2602 } else if (op == MDB_GET_BOTH || op == MDB_GET_BOTH_RANGE) {
2604 if ((rc = mdb_read_data(cursor->mc_txn, leaf, &d2)) != MDB_SUCCESS)
2606 rc = mdb_dcmp(cursor->mc_txn, cursor->mc_dbi, data, &d2);
2608 if (op == MDB_GET_BOTH || rc > 0)
2609 return MDB_NOTFOUND;
2613 if ((rc = mdb_read_data(cursor->mc_txn, leaf, data)) != MDB_SUCCESS)
2618 /* The key already matches in all other cases */
2619 if (op == MDB_SET_RANGE)
2620 MDB_SET_KEY(leaf, key);
2621 DPRINTF("==> cursor placed on key [%s]", DKEY(key));
2627 mdb_cursor_first(MDB_cursor *cursor, MDB_val *key, MDB_val *data)
2633 if (cursor->mc_flags & C_XDIRTY)
2634 mdb_cursor_reset(cursor);
2636 cursor->mc_snum = 0;
2638 rc = mdb_search_page(cursor->mc_txn, cursor->mc_dbi, NULL, cursor, 0, &mpp);
2639 if (rc != MDB_SUCCESS)
2641 assert(IS_LEAF(mpp.mp_page));
2643 leaf = NODEPTR(mpp.mp_page, 0);
2644 cursor->mc_flags |= C_INITIALIZED;
2645 cursor->mc_flags &= ~C_EOF;
2647 if (IS_LEAF2(mpp.mp_page)) {
2648 key->mv_size = cursor->mc_txn->mt_dbs[cursor->mc_dbi].md_pad;
2649 key->mv_data = LEAF2KEY(mpp.mp_page, 0, key->mv_size);
2654 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
2655 mdb_xcursor_init1(cursor, mpp.mp_page, leaf);
2656 rc = mdb_cursor_first(&cursor->mc_xcursor->mx_cursor, data, NULL);
2660 if (cursor->mc_xcursor)
2661 cursor->mc_xcursor->mx_cursor.mc_flags = 0;
2662 if ((rc = mdb_read_data(cursor->mc_txn, leaf, data)) != MDB_SUCCESS)
2666 MDB_SET_KEY(leaf, key);
2671 mdb_cursor_last(MDB_cursor *cursor, MDB_val *key, MDB_val *data)
2679 if (cursor->mc_flags & C_XDIRTY)
2680 mdb_cursor_reset(cursor);
2682 cursor->mc_snum = 0;
2684 lkey.mv_size = MAXKEYSIZE+1;
2685 lkey.mv_data = NULL;
2687 rc = mdb_search_page(cursor->mc_txn, cursor->mc_dbi, &lkey, cursor, 0, &mpp);
2688 if (rc != MDB_SUCCESS)
2690 assert(IS_LEAF(mpp.mp_page));
2692 leaf = NODEPTR(mpp.mp_page, NUMKEYS(mpp.mp_page)-1);
2693 cursor->mc_flags |= C_INITIALIZED;
2694 cursor->mc_flags &= ~C_EOF;
2696 top = CURSOR_TOP(cursor);
2697 top->mp_ki = NUMKEYS(top->mp_page) - 1;
2699 if (IS_LEAF2(mpp.mp_page)) {
2700 key->mv_size = cursor->mc_txn->mt_dbs[cursor->mc_dbi].md_pad;
2701 key->mv_data = LEAF2KEY(mpp.mp_page, top->mp_ki, key->mv_size);
2706 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
2707 mdb_xcursor_init1(cursor, mpp.mp_page, leaf);
2708 rc = mdb_cursor_last(&cursor->mc_xcursor->mx_cursor, data, NULL);
2712 if ((rc = mdb_read_data(cursor->mc_txn, leaf, data)) != MDB_SUCCESS)
2717 MDB_SET_KEY(leaf, key);
2722 mdb_cursor_get(MDB_cursor *cursor, MDB_val *key, MDB_val *data,
2732 case MDB_GET_BOTH_RANGE:
2733 if (data == NULL || cursor->mc_xcursor == NULL) {
2740 if (key == NULL || key->mv_size == 0 || key->mv_size > MAXKEYSIZE) {
2742 } else if (op == MDB_SET_RANGE)
2743 rc = mdb_cursor_set(cursor, key, data, op, NULL);
2745 rc = mdb_cursor_set(cursor, key, data, op, &exact);
2747 case MDB_GET_MULTIPLE:
2749 !(cursor->mc_txn->mt_dbs[cursor->mc_dbi].md_flags & MDB_DUPFIXED) ||
2750 !(cursor->mc_flags & C_INITIALIZED)) {
2755 if (!(cursor->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) ||
2756 (cursor->mc_xcursor->mx_cursor.mc_flags & C_EOF))
2759 case MDB_NEXT_MULTIPLE:
2761 !(cursor->mc_txn->mt_dbs[cursor->mc_dbi].md_flags & MDB_DUPFIXED)) {
2765 if (!(cursor->mc_flags & C_INITIALIZED))
2766 rc = mdb_cursor_first(cursor, key, data);
2768 rc = mdb_cursor_next(cursor, key, data, MDB_NEXT_DUP);
2769 if (rc == MDB_SUCCESS) {
2770 if (cursor->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) {
2773 top = CURSOR_TOP(&cursor->mc_xcursor->mx_cursor);
2774 data->mv_size = NUMKEYS(top->mp_page) *
2775 cursor->mc_xcursor->mx_txn.mt_dbs[cursor->mc_xcursor->mx_cursor.mc_dbi].md_pad;
2776 data->mv_data = METADATA(top->mp_page);
2777 top->mp_ki = NUMKEYS(top->mp_page)-1;
2785 case MDB_NEXT_NODUP:
2786 if (!(cursor->mc_flags & C_INITIALIZED))
2787 rc = mdb_cursor_first(cursor, key, data);
2789 rc = mdb_cursor_next(cursor, key, data, op);
2793 case MDB_PREV_NODUP:
2794 if (!(cursor->mc_flags & C_INITIALIZED) || (cursor->mc_flags & C_EOF))
2795 rc = mdb_cursor_last(cursor, key, data);
2797 rc = mdb_cursor_prev(cursor, key, data, op);
2800 rc = mdb_cursor_first(cursor, key, data);
2804 !(cursor->mc_txn->mt_dbs[cursor->mc_dbi].md_flags & MDB_DUPSORT) ||
2805 !(cursor->mc_flags & C_INITIALIZED) ||
2806 !(cursor->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
2810 rc = mdb_cursor_first(&cursor->mc_xcursor->mx_cursor, data, NULL);
2813 rc = mdb_cursor_last(cursor, key, data);
2817 !(cursor->mc_txn->mt_dbs[cursor->mc_dbi].md_flags & MDB_DUPSORT) ||
2818 !(cursor->mc_flags & C_INITIALIZED) ||
2819 !(cursor->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
2823 rc = mdb_cursor_last(&cursor->mc_xcursor->mx_cursor, data, NULL);
2826 DPRINTF("unhandled/unimplemented cursor operation %u", op);
2835 mdb_cursor_touch(MDB_cursor *mc)
2842 mpp.mp_parent = NULL;
2844 if (mc->mc_dbi > MAIN_DBI && !mc->mc_txn->mt_dbxs[mc->mc_dbi].md_dirty) {
2845 rc = mdb_search_page(mc->mc_txn, MAIN_DBI, &mc->mc_txn->mt_dbxs[mc->mc_dbi].md_name,
2848 mc->mc_txn->mt_dbxs[mc->mc_dbi].md_dirty = 1;
2850 for(i=0; i<mc->mc_snum; i++) {
2851 mpp.mp_page = mc->mc_stack[i].mp_page;
2852 if (!F_ISSET(mpp.mp_page->mp_flags, P_DIRTY)) {
2853 rc = mdb_touch(mc->mc_txn, mc->mc_dbi, &mpp);
2855 mc->mc_stack[i].mp_page = mpp.mp_page;
2857 mc->mc_txn->mt_dbs[mc->mc_dbi].md_root = mpp.mp_page->mp_pgno;
2860 dh = ((MDB_dhead *)mpp.mp_page)-1;
2861 dh->md_parent = mpp.mp_parent;
2862 dh->md_pi = mpp.mp_pi;
2863 mpp.mp_parent = mpp.mp_page;
2864 mpp.mp_pi = mc->mc_stack[i].mp_ki;
2870 mdb_cursor_put(MDB_cursor *mc, MDB_val *key, MDB_val *data,
2875 MDB_val xdata, *rdata, dkey;
2877 char dbuf[PAGESIZE];
2883 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_RDONLY))
2886 DPRINTF("==> put db %u key [%s], size %zu, data size %zu",
2887 mc->mc_dbi, DKEY(key), key->mv_size, data->mv_size);
2891 if (flags == MDB_CURRENT) {
2892 if (!(mc->mc_flags & C_INITIALIZED))
2895 } else if (mc->mc_txn->mt_dbs[mc->mc_dbi].md_root == P_INVALID) {
2897 /* new database, write a root leaf page */
2898 DPUTS("allocating new root leaf page");
2899 if ((dp = mdb_new_page(mc->mc_txn, mc->mc_dbi, P_LEAF, 1)) == NULL) {
2903 cursor_push_page(mc, &dp->p);
2904 mc->mc_txn->mt_dbs[mc->mc_dbi].md_root = dp->p.mp_pgno;
2905 mc->mc_txn->mt_dbs[mc->mc_dbi].md_depth++;
2906 mc->mc_txn->mt_dbxs[mc->mc_dbi].md_dirty = 1;
2907 if ((mc->mc_txn->mt_dbs[mc->mc_dbi].md_flags & (MDB_DUPSORT|MDB_DUPFIXED))
2909 dp->p.mp_flags |= P_LEAF2;
2913 rc = mdb_cursor_set(mc, key, NULL, MDB_SET, &exact);
2914 if (flags == MDB_NOOVERWRITE && rc == 0) {
2915 DPRINTF("duplicate key [%s]", DKEY(key));
2916 return MDB_KEYEXIST;
2918 if (rc && rc != MDB_NOTFOUND)
2922 /* Cursor is positioned, now make sure all pages are writable */
2923 rc2 = mdb_cursor_touch(mc);
2924 if (rc2) return rc2;
2926 top = CURSOR_TOP(mc);
2928 /* The key already exists */
2929 if (rc == MDB_SUCCESS) {
2930 /* there's only a key anyway, so this is a no-op */
2931 if (IS_LEAF2(top->mp_page))
2934 leaf = NODEPTR(top->mp_page, top->mp_ki);
2937 if (F_ISSET(mc->mc_txn->mt_dbs[mc->mc_dbi].md_flags, MDB_DUPSORT)) {
2938 /* Was a single item before, must convert now */
2939 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
2940 dkey.mv_size = NODEDSZ(leaf);
2941 dkey.mv_data = dbuf;
2942 memcpy(dbuf, NODEDATA(leaf), dkey.mv_size);
2943 /* data matches, ignore it */
2944 if (!mdb_dcmp(mc->mc_txn, mc->mc_dbi, data, &dkey))
2945 return (flags == MDB_NODUPDATA) ? MDB_KEYEXIST : MDB_SUCCESS;
2946 memset(&dummy, 0, sizeof(dummy));
2947 if (mc->mc_txn->mt_dbs[mc->mc_dbi].md_flags & MDB_DUPFIXED) {
2948 dummy.md_pad = data->mv_size;
2949 dummy.md_flags = MDB_DUPFIXED;
2950 if (mc->mc_txn->mt_dbs[mc->mc_dbi].md_flags & MDB_INTEGERDUP)
2951 dummy.md_flags |= MDB_INTEGERKEY;
2953 dummy.md_root = P_INVALID;
2954 if (dkey.mv_size == sizeof(MDB_db)) {
2955 memcpy(NODEDATA(leaf), &dummy, sizeof(dummy));
2958 mdb_del_node(top->mp_page, top->mp_ki, 0);
2961 xdata.mv_size = sizeof(MDB_db);
2962 xdata.mv_data = &dummy;
2967 /* same size, just replace it */
2968 if (!F_ISSET(leaf->mn_flags, F_BIGDATA) &&
2969 NODEDSZ(leaf) == data->mv_size) {
2970 memcpy(NODEDATA(leaf), data->mv_data, data->mv_size);
2973 mdb_del_node(top->mp_page, top->mp_ki, 0);
2975 DPRINTF("appending key at index %i", top->mp_ki);
2981 nsize = IS_LEAF2(top->mp_page) ? key->mv_size : mdb_leaf_size(mc->mc_txn->mt_env, key, rdata);
2982 if (SIZELEFT(top->mp_page) < nsize) {
2983 rc = mdb_split(mc->mc_txn, mc->mc_dbi, &top->mp_page, &top->mp_ki, key, rdata, P_INVALID);
2985 /* There is room already in this leaf page. */
2986 rc = mdb_add_node(mc->mc_txn, mc->mc_dbi, top->mp_page, top->mp_ki, key, rdata, 0, 0);
2989 if (rc != MDB_SUCCESS)
2990 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
2992 /* Remember if we just added a subdatabase */
2993 if (flags & F_SUBDATA) {
2994 leaf = NODEPTR(top->mp_page, top->mp_ki);
2995 leaf->mn_flags |= F_SUBDATA;
2998 /* Now store the actual data in the child DB. Note that we're
2999 * storing the user data in the keys field, so there are strict
3000 * size limits on dupdata. The actual data fields of the child
3001 * DB are all zero size.
3004 leaf = NODEPTR(top->mp_page, top->mp_ki);
3006 mdb_xcursor_init1(mc, top->mp_page, leaf);
3009 if (flags == MDB_NODUPDATA)
3010 flags = MDB_NOOVERWRITE;
3011 /* converted, write the original data first */
3013 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, &dkey, &xdata, flags);
3015 leaf->mn_flags |= F_DUPDATA;
3017 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, data, &xdata, flags);
3018 mc->mc_flags |= C_XDIRTY;
3020 mc->mc_txn->mt_dbs[mc->mc_dbi].md_entries++;
3027 mdb_cursor_del(MDB_cursor *mc, unsigned int flags)
3034 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_RDONLY))
3037 if (!mc->mc_flags & C_INITIALIZED)
3040 rc = mdb_cursor_touch(mc);
3043 top = CURSOR_TOP(mc);
3044 leaf = NODEPTR(top->mp_page, top->mp_ki);
3045 mpp.mp_page = top->mp_page;
3046 if (mc->mc_snum > 1) {
3047 MDB_ppage *parent = CURSOR_PARENT(mc);
3048 mpp.mp_parent = parent->mp_page;
3049 mpp.mp_pi = parent->mp_ki;
3051 mpp.mp_parent = NULL;
3055 if (!IS_LEAF2(top->mp_page) && F_ISSET(leaf->mn_flags, F_DUPDATA)) {
3058 if (flags != MDB_NODUPDATA) {
3059 mdb_xcursor_init2(mc);
3060 rc = mdb_cursor_del(&mc->mc_xcursor->mx_cursor, 0);
3061 mc->mc_flags |= C_XDIRTY;
3062 /* If sub-DB still has entries, we're done */
3063 if (mc->mc_xcursor->mx_txn.mt_dbs[mc->mc_xcursor->mx_cursor.mc_dbi].md_root
3065 mc->mc_txn->mt_dbs[mc->mc_dbi].md_entries--;
3068 /* otherwise fall thru and delete the sub-DB */
3071 /* add all the child DB's pages to the free list */
3072 rc = mdb_search_page(&mc->mc_xcursor->mx_txn, mc->mc_xcursor->mx_cursor.mc_dbi,
3073 NULL, &mc->mc_xcursor->mx_cursor, 0, &mp2);
3074 if (rc == MDB_SUCCESS) {
3075 MDB_ppage *top, *parent;
3079 mc->mc_txn->mt_dbs[mc->mc_dbi].md_entries -=
3080 mc->mc_xcursor->mx_txn.mt_dbs[mc->mc_xcursor->mx_cursor.mc_dbi].md_entries;
3082 cursor_pop_page(&mc->mc_xcursor->mx_cursor);
3083 if (mc->mc_xcursor->mx_cursor.mc_snum) {
3084 top = CURSOR_TOP(&mc->mc_xcursor->mx_cursor);
3085 while (mc->mc_xcursor->mx_cursor.mc_snum > 1) {
3086 parent = CURSOR_PARENT(&mc->mc_xcursor->mx_cursor);
3087 for (i=0; i<NUMKEYS(top->mp_page); i++) {
3088 ni = NODEPTR(top->mp_page, i);
3089 mdb_midl_insert(mc->mc_txn->mt_free_pgs, NODEPGNO(ni));
3092 if (parent->mp_ki >= NUMKEYS(parent->mp_page)) {
3093 cursor_pop_page(&mc->mc_xcursor->mx_cursor);
3096 ni = NODEPTR(parent->mp_page, parent->mp_ki);
3097 rc = mdb_get_page(&mc->mc_xcursor->mx_txn, NODEPGNO(ni), &top->mp_page);
3101 mdb_midl_insert(mc->mc_txn->mt_free_pgs,
3102 mc->mc_xcursor->mx_txn.mt_dbs[mc->mc_xcursor->mx_cursor.mc_dbi].md_root);
3106 return mdb_del0(mc, top->mp_ki, &mpp, leaf);
3109 /* Allocate a page and initialize it
3112 mdb_new_page(MDB_txn *txn, MDB_dbi dbi, uint32_t flags, int num)
3116 if ((dp = mdb_alloc_page(txn, dbi, NULL, 0, num)) == NULL)
3118 DPRINTF("allocated new mpage %lu, page size %u",
3119 dp->p.mp_pgno, txn->mt_env->me_psize);
3120 dp->p.mp_flags = flags | P_DIRTY;
3121 dp->p.mp_lower = PAGEHDRSZ;
3122 dp->p.mp_upper = txn->mt_env->me_psize;
3124 if (IS_BRANCH(&dp->p))
3125 txn->mt_dbs[dbi].md_branch_pages++;
3126 else if (IS_LEAF(&dp->p))
3127 txn->mt_dbs[dbi].md_leaf_pages++;
3128 else if (IS_OVERFLOW(&dp->p)) {
3129 txn->mt_dbs[dbi].md_overflow_pages += num;
3130 dp->p.mp_pages = num;
3137 mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data)
3141 sz = LEAFSIZE(key, data);
3142 if (data->mv_size >= env->me_psize / MDB_MINKEYS) {
3143 /* put on overflow page */
3144 sz -= data->mv_size - sizeof(pgno_t);
3147 return sz + sizeof(indx_t);
3151 mdb_branch_size(MDB_env *env, MDB_val *key)
3156 if (sz >= env->me_psize / MDB_MINKEYS) {
3157 /* put on overflow page */
3158 /* not implemented */
3159 /* sz -= key->size - sizeof(pgno_t); */
3162 return sz + sizeof(indx_t);
3166 mdb_add_node(MDB_txn *txn, MDB_dbi dbi, MDB_page *mp, indx_t indx,
3167 MDB_val *key, MDB_val *data, pgno_t pgno, uint8_t flags)
3170 size_t node_size = NODESIZE;
3173 MDB_dpage *ofp = NULL; /* overflow page */
3176 assert(mp->mp_upper >= mp->mp_lower);
3178 DPRINTF("add to %s page %lu index %i, data size %zu key size %zu [%s]",
3179 IS_LEAF(mp) ? "leaf" : "branch",
3180 mp->mp_pgno, indx, data ? data->mv_size : 0,
3181 key ? key->mv_size : 0, key ? DKEY(key) : NULL);
3184 /* Move higher keys up one slot. */
3185 int ksize = txn->mt_dbs[dbi].md_pad, dif;
3186 char *ptr = LEAF2KEY(mp, indx, ksize);
3187 dif = NUMKEYS(mp) - indx;
3189 memmove(ptr+ksize, ptr, dif*ksize);
3190 /* insert new key */
3191 memcpy(ptr, key->mv_data, ksize);
3193 /* Just using these for counting */
3194 mp->mp_lower += sizeof(indx_t);
3195 mp->mp_upper -= ksize - sizeof(indx_t);
3200 node_size += key->mv_size;
3204 if (F_ISSET(flags, F_BIGDATA)) {
3205 /* Data already on overflow page. */
3206 node_size += sizeof(pgno_t);
3207 } else if (data->mv_size >= txn->mt_env->me_psize / MDB_MINKEYS) {
3208 int ovpages = OVPAGES(data->mv_size, txn->mt_env->me_psize);
3209 /* Put data on overflow page. */
3210 DPRINTF("data size is %zu, put on overflow page",
3212 node_size += sizeof(pgno_t);
3213 if ((ofp = mdb_new_page(txn, dbi, P_OVERFLOW, ovpages)) == NULL)
3215 DPRINTF("allocated overflow page %lu", ofp->p.mp_pgno);
3218 node_size += data->mv_size;
3222 if (node_size + sizeof(indx_t) > SIZELEFT(mp)) {
3223 DPRINTF("not enough room in page %lu, got %u ptrs",
3224 mp->mp_pgno, NUMKEYS(mp));
3225 DPRINTF("upper - lower = %u - %u = %u", mp->mp_upper, mp->mp_lower,
3226 mp->mp_upper - mp->mp_lower);
3227 DPRINTF("node size = %zu", node_size);
3231 /* Move higher pointers up one slot. */
3232 for (i = NUMKEYS(mp); i > indx; i--)
3233 mp->mp_ptrs[i] = mp->mp_ptrs[i - 1];
3235 /* Adjust free space offsets. */
3236 ofs = mp->mp_upper - node_size;
3237 assert(ofs >= mp->mp_lower + sizeof(indx_t));
3238 mp->mp_ptrs[indx] = ofs;
3240 mp->mp_lower += sizeof(indx_t);
3242 /* Write the node data. */
3243 node = NODEPTR(mp, indx);
3244 node->mn_ksize = (key == NULL) ? 0 : key->mv_size;
3245 node->mn_flags = flags;
3247 node->mn_dsize = data->mv_size;
3249 NODEPGNO(node) = pgno;
3252 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
3257 if (F_ISSET(flags, F_BIGDATA))
3258 memcpy(node->mn_data + key->mv_size, data->mv_data,
3261 memcpy(node->mn_data + key->mv_size, data->mv_data,
3264 memcpy(node->mn_data + key->mv_size, &ofp->p.mp_pgno,
3266 memcpy(METADATA(&ofp->p), data->mv_data, data->mv_size);
3274 mdb_del_node(MDB_page *mp, indx_t indx, int ksize)
3277 indx_t i, j, numkeys, ptr;
3281 DPRINTF("delete node %u on %s page %lu", indx,
3282 IS_LEAF(mp) ? "leaf" : "branch", mp->mp_pgno);
3283 assert(indx < NUMKEYS(mp));
3286 int x = NUMKEYS(mp) - 1 - indx;
3287 base = LEAF2KEY(mp, indx, ksize);
3289 memmove(base, base + ksize, x * ksize);
3290 mp->mp_lower -= sizeof(indx_t);
3291 mp->mp_upper += ksize - sizeof(indx_t);
3295 node = NODEPTR(mp, indx);
3296 sz = NODESIZE + node->mn_ksize;
3298 if (F_ISSET(node->mn_flags, F_BIGDATA))
3299 sz += sizeof(pgno_t);
3301 sz += NODEDSZ(node);
3304 ptr = mp->mp_ptrs[indx];
3305 numkeys = NUMKEYS(mp);
3306 for (i = j = 0; i < numkeys; i++) {
3308 mp->mp_ptrs[j] = mp->mp_ptrs[i];
3309 if (mp->mp_ptrs[i] < ptr)
3310 mp->mp_ptrs[j] += sz;
3315 base = (char *)mp + mp->mp_upper;
3316 memmove(base + sz, base, ptr - mp->mp_upper);
3318 mp->mp_lower -= sizeof(indx_t);
3323 mdb_xcursor_init0(MDB_cursor *mc)
3325 MDB_xcursor *mx = mc->mc_xcursor;
3328 mx->mx_txn = *mc->mc_txn;
3329 mx->mx_txn.mt_dbxs = mx->mx_dbxs;
3330 mx->mx_txn.mt_dbs = mx->mx_dbs;
3331 mx->mx_dbxs[0] = mc->mc_txn->mt_dbxs[0];
3332 mx->mx_dbxs[1] = mc->mc_txn->mt_dbxs[1];
3333 if (mc->mc_dbi > 1) {
3334 mx->mx_dbxs[2] = mc->mc_txn->mt_dbxs[mc->mc_dbi];
3339 mx->mx_dbxs[dbn+1].md_parent = dbn;
3340 mx->mx_dbxs[dbn+1].md_cmp = mx->mx_dbxs[dbn].md_dcmp;
3341 mx->mx_dbxs[dbn+1].md_rel = mx->mx_dbxs[dbn].md_rel;
3342 mx->mx_dbxs[dbn+1].md_dirty = 0;
3343 mx->mx_txn.mt_numdbs = dbn+2;
3345 mx->mx_cursor.mc_xcursor = NULL;
3346 mx->mx_cursor.mc_txn = &mx->mx_txn;
3347 mx->mx_cursor.mc_snum = 0;
3348 mx->mx_cursor.mc_dbi = dbn+1;
3352 mdb_xcursor_init1(MDB_cursor *mc, MDB_page *mp, MDB_node *node)
3354 MDB_db *db = NODEDATA(node);
3355 MDB_xcursor *mx = mc->mc_xcursor;
3357 mx->mx_dbs[0] = mc->mc_txn->mt_dbs[0];
3358 mx->mx_dbs[1] = mc->mc_txn->mt_dbs[1];
3359 if (mc->mc_dbi > 1) {
3360 mx->mx_dbs[2] = mc->mc_txn->mt_dbs[mc->mc_dbi];
3365 DPRINTF("Sub-db %u for db %u root page %lu", dbn, mc->mc_dbi, db->md_root);
3366 mx->mx_dbs[dbn] = *db;
3367 if (F_ISSET(mp->mp_flags, P_DIRTY))
3368 mx->mx_dbxs[dbn].md_dirty = 1;
3369 mx->mx_dbxs[dbn].md_name.mv_data = NODEKEY(node);
3370 mx->mx_dbxs[dbn].md_name.mv_size = node->mn_ksize;
3371 mx->mx_txn.mt_next_pgno = mc->mc_txn->mt_next_pgno;
3372 mx->mx_txn.mt_u = mc->mc_txn->mt_u;
3373 mx->mx_cursor.mc_flags = 0;
3377 mdb_xcursor_init2(MDB_cursor *mc)
3379 MDB_xcursor *mx = mc->mc_xcursor;
3380 mx->mx_dbs[0] = mc->mc_txn->mt_dbs[0];
3381 mx->mx_dbs[1] = mc->mc_txn->mt_dbs[1];
3382 if (mc->mc_dbi > 1) {
3383 mx->mx_dbs[2] = mc->mc_txn->mt_dbs[mc->mc_dbi];
3385 mx->mx_txn.mt_next_pgno = mc->mc_txn->mt_next_pgno;
3389 mdb_xcursor_fini(MDB_cursor *mc)
3391 MDB_xcursor *mx = mc->mc_xcursor;
3392 mc->mc_txn->mt_next_pgno = mx->mx_txn.mt_next_pgno;
3393 mc->mc_txn->mt_dbs[0] = mx->mx_dbs[0];
3394 mc->mc_txn->mt_dbs[1] = mx->mx_dbs[1];
3395 mc->mc_txn->mt_dbxs[0].md_dirty = mx->mx_dbxs[0].md_dirty;
3396 mc->mc_txn->mt_dbxs[1].md_dirty = mx->mx_dbxs[1].md_dirty;
3397 if (mc->mc_dbi > 1) {
3398 mc->mc_txn->mt_dbs[mc->mc_dbi] = mx->mx_dbs[2];
3399 mc->mc_txn->mt_dbxs[mc->mc_dbi].md_dirty = mx->mx_dbxs[2].md_dirty;
3404 mdb_cursor_open(MDB_txn *txn, MDB_dbi dbi, MDB_cursor **ret)
3407 size_t size = sizeof(MDB_cursor);
3409 if (txn == NULL || ret == NULL || !dbi || dbi >= txn->mt_numdbs)
3412 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT)
3413 size += sizeof(MDB_xcursor);
3415 if ((cursor = calloc(1, size)) != NULL) {
3416 cursor->mc_dbi = dbi;
3417 cursor->mc_txn = txn;
3418 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
3419 MDB_xcursor *mx = (MDB_xcursor *)(cursor + 1);
3420 cursor->mc_xcursor = mx;
3421 mdb_xcursor_init0(cursor);
3432 /* Return the count of duplicate data items for the current key */
3434 mdb_cursor_count(MDB_cursor *mc, unsigned long *countp)
3439 if (mc == NULL || countp == NULL)
3442 if (!(mc->mc_txn->mt_dbs[mc->mc_dbi].md_flags & MDB_DUPSORT))
3445 top = CURSOR_TOP(mc);
3446 leaf = NODEPTR(top->mp_page, top->mp_ki);
3447 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
3450 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED))
3453 *countp = mc->mc_xcursor->mx_txn.mt_dbs[mc->mc_xcursor->mx_cursor.mc_dbi].md_entries;
3459 mdb_cursor_reset(MDB_cursor *mc)
3462 if (mc->mc_flags & C_XDIRTY) {
3464 mdb_xcursor_fini(mc);
3466 /* If sub-DB still has entries, update root record */
3467 if (mc->mc_xcursor->mx_txn.mt_dbs[mc->mc_xcursor->mx_cursor.mc_dbi].md_root
3471 top = CURSOR_TOP(mc);
3472 leaf = NODEPTR(top->mp_page, top->mp_ki);
3473 memcpy(NODEDATA(leaf),
3474 &mc->mc_xcursor->mx_txn.mt_dbs[mc->mc_xcursor->mx_cursor.mc_dbi],
3482 mdb_cursor_close(MDB_cursor *cursor)
3484 if (cursor != NULL) {
3485 mdb_cursor_reset(cursor);
3491 mdb_update_key(MDB_page *mp, indx_t indx, MDB_val *key)
3493 indx_t ptr, i, numkeys;
3500 node = NODEPTR(mp, indx);
3501 ptr = mp->mp_ptrs[indx];
3502 DPRINTF("update key %u (ofs %u) [%.*s] to [%s] on page %lu",
3504 (int)node->mn_ksize, (char *)NODEKEY(node),
3508 delta = key->mv_size - node->mn_ksize;
3510 if (delta > 0 && SIZELEFT(mp) < delta) {
3511 DPRINTF("OUCH! Not enough room, delta = %d", delta);
3515 numkeys = NUMKEYS(mp);
3516 for (i = 0; i < numkeys; i++) {
3517 if (mp->mp_ptrs[i] <= ptr)
3518 mp->mp_ptrs[i] -= delta;
3521 base = (char *)mp + mp->mp_upper;
3522 len = ptr - mp->mp_upper + NODESIZE;
3523 memmove(base - delta, base, len);
3524 mp->mp_upper -= delta;
3526 node = NODEPTR(mp, indx);
3527 node->mn_ksize = key->mv_size;
3530 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
3535 /* Move a node from src to dst.
3538 mdb_move_node(MDB_txn *txn, MDB_dbi dbi, MDB_pageparent *src, indx_t srcindx,
3539 MDB_pageparent *dst, indx_t dstindx)
3546 /* Mark src and dst as dirty. */
3547 if ((rc = mdb_touch(txn, dbi, src)) ||
3548 (rc = mdb_touch(txn, dbi, dst)))
3551 if (IS_LEAF2(src->mp_page)) {
3552 srcnode = NODEPTR(src->mp_page, 0); /* fake */
3553 key.mv_size = txn->mt_dbs[dbi].md_pad;
3554 key.mv_data = LEAF2KEY(src->mp_page, srcindx, key.mv_size);
3556 data.mv_data = NULL;
3558 srcnode = NODEPTR(src->mp_page, srcindx);
3559 key.mv_size = NODEKSZ(srcnode);
3560 key.mv_data = NODEKEY(srcnode);
3561 data.mv_size = NODEDSZ(srcnode);
3562 data.mv_data = NODEDATA(srcnode);
3564 DPRINTF("moving %s node %u [%s] on page %lu to node %u on page %lu",
3565 IS_LEAF(src->mp_page) ? "leaf" : "branch",
3568 src->mp_page->mp_pgno,
3569 dstindx, dst->mp_page->mp_pgno);
3571 /* Add the node to the destination page.
3573 rc = mdb_add_node(txn, dbi, dst->mp_page, dstindx, &key, &data, NODEPGNO(srcnode),
3575 if (rc != MDB_SUCCESS)
3578 /* Delete the node from the source page.
3580 mdb_del_node(src->mp_page, srcindx, key.mv_size);
3582 /* The key value just changed due to del_node, find it again.
3584 if (!IS_LEAF2(src->mp_page)) {
3585 srcnode = NODEPTR(src->mp_page, srcindx);
3586 key.mv_data = NODEKEY(srcnode);
3589 /* Update the parent separators.
3592 if (src->mp_pi != 0) {
3593 DPRINTF("update separator for source page %lu to [%s]",
3594 src->mp_page->mp_pgno, DKEY(&key));
3595 if ((rc = mdb_update_key(src->mp_parent, src->mp_pi,
3596 &key)) != MDB_SUCCESS)
3599 if (IS_BRANCH(src->mp_page)) {
3601 nullkey.mv_size = 0;
3602 assert(mdb_update_key(src->mp_page, 0, &nullkey) == MDB_SUCCESS);
3607 if (dst->mp_pi != 0) {
3608 DPRINTF("update separator for destination page %lu to [%s]",
3609 dst->mp_page->mp_pgno, DKEY(&key));
3610 if ((rc = mdb_update_key(dst->mp_parent, dst->mp_pi,
3611 &key)) != MDB_SUCCESS)
3614 if (IS_BRANCH(dst->mp_page)) {
3616 nullkey.mv_size = 0;
3617 assert(mdb_update_key(dst->mp_page, 0, &nullkey) == MDB_SUCCESS);
3625 mdb_merge(MDB_txn *txn, MDB_dbi dbi, MDB_pageparent *src, MDB_pageparent *dst)
3634 DPRINTF("merging page %lu and %lu", src->mp_page->mp_pgno, dst->mp_page->mp_pgno);
3636 assert(txn != NULL);
3637 assert(src->mp_parent); /* can't merge root page */
3638 assert(dst->mp_parent);
3640 /* Mark src and dst as dirty. */
3641 if ((rc = mdb_touch(txn, dbi, src)) ||
3642 (rc = mdb_touch(txn, dbi, dst)))
3645 /* Move all nodes from src to dst.
3647 if (IS_LEAF2(src->mp_page)) {
3648 key.mv_size = txn->mt_dbs[dbi].md_pad;
3649 key.mv_data = METADATA(src->mp_page);
3650 for (i = 0; i < NUMKEYS(src->mp_page); i++) {
3651 rc = mdb_add_node(txn, dbi, dst->mp_page, NUMKEYS(dst->mp_page), &key,
3653 if (rc != MDB_SUCCESS)
3655 key.mv_data = (char *)key.mv_data + key.mv_size;
3658 for (i = 0; i < NUMKEYS(src->mp_page); i++) {
3659 srcnode = NODEPTR(src->mp_page, i);
3661 key.mv_size = srcnode->mn_ksize;
3662 key.mv_data = NODEKEY(srcnode);
3663 data.mv_size = NODEDSZ(srcnode);
3664 data.mv_data = NODEDATA(srcnode);
3665 rc = mdb_add_node(txn, dbi, dst->mp_page, NUMKEYS(dst->mp_page), &key,
3666 &data, NODEPGNO(srcnode), srcnode->mn_flags);
3667 if (rc != MDB_SUCCESS)
3672 DPRINTF("dst page %lu now has %u keys (%.1f%% filled)",
3673 dst->mp_page->mp_pgno, NUMKEYS(dst->mp_page), (float)PAGEFILL(txn->mt_env, dst->mp_page) / 10);
3675 /* Unlink the src page from parent.
3677 mdb_del_node(src->mp_parent, src->mp_pi, 0);
3678 if (src->mp_pi == 0) {
3680 if ((rc = mdb_update_key(src->mp_parent, 0, &key)) != MDB_SUCCESS)
3684 if (IS_LEAF(src->mp_page))
3685 txn->mt_dbs[dbi].md_leaf_pages--;
3687 txn->mt_dbs[dbi].md_branch_pages--;
3689 mpp.mp_page = src->mp_parent;
3690 dh = (MDB_dhead *)src->mp_parent;
3692 mpp.mp_parent = dh->md_parent;
3693 mpp.mp_pi = dh->md_pi;
3695 return mdb_rebalance(txn, dbi, &mpp);
3698 #define FILL_THRESHOLD 250
3701 mdb_rebalance(MDB_txn *txn, MDB_dbi dbi, MDB_pageparent *mpp)
3706 indx_t si = 0, di = 0;
3709 assert(txn != NULL);
3710 assert(mpp != NULL);
3712 DPRINTF("rebalancing %s page %lu (has %u keys, %.1f%% full)",
3713 IS_LEAF(mpp->mp_page) ? "leaf" : "branch",
3714 mpp->mp_page->mp_pgno, NUMKEYS(mpp->mp_page), (float)PAGEFILL(txn->mt_env, mpp->mp_page) / 10);
3716 if (PAGEFILL(txn->mt_env, mpp->mp_page) >= FILL_THRESHOLD) {
3717 DPRINTF("no need to rebalance page %lu, above fill threshold",
3718 mpp->mp_page->mp_pgno);
3722 if (mpp->mp_parent == NULL) {
3723 if (NUMKEYS(mpp->mp_page) == 0) {
3724 DPUTS("tree is completely empty");
3725 txn->mt_dbs[dbi].md_root = P_INVALID;
3726 txn->mt_dbs[dbi].md_depth = 0;
3727 txn->mt_dbs[dbi].md_leaf_pages = 0;
3728 } else if (IS_BRANCH(mpp->mp_page) && NUMKEYS(mpp->mp_page) == 1) {
3729 DPUTS("collapsing root page!");
3730 txn->mt_dbs[dbi].md_root = NODEPGNO(NODEPTR(mpp->mp_page, 0));
3731 if ((rc = mdb_get_page(txn, txn->mt_dbs[dbi].md_root, &root)))
3733 txn->mt_dbs[dbi].md_depth--;
3734 txn->mt_dbs[dbi].md_branch_pages--;
3736 DPUTS("root page doesn't need rebalancing");
3740 /* The parent (branch page) must have at least 2 pointers,
3741 * otherwise the tree is invalid.
3743 assert(NUMKEYS(mpp->mp_parent) > 1);
3745 /* Leaf page fill factor is below the threshold.
3746 * Try to move keys from left or right neighbor, or
3747 * merge with a neighbor page.
3752 if (mpp->mp_pi == 0) {
3753 /* We're the leftmost leaf in our parent.
3755 DPUTS("reading right neighbor");
3756 node = NODEPTR(mpp->mp_parent, mpp->mp_pi + 1);
3757 if ((rc = mdb_get_page(txn, NODEPGNO(node), &npp.mp_page)))
3759 npp.mp_pi = mpp->mp_pi + 1;
3761 di = NUMKEYS(mpp->mp_page);
3763 /* There is at least one neighbor to the left.
3765 DPUTS("reading left neighbor");
3766 node = NODEPTR(mpp->mp_parent, mpp->mp_pi - 1);
3767 if ((rc = mdb_get_page(txn, NODEPGNO(node), &npp.mp_page)))
3769 npp.mp_pi = mpp->mp_pi - 1;
3770 si = NUMKEYS(npp.mp_page) - 1;
3773 npp.mp_parent = mpp->mp_parent;
3775 DPRINTF("found neighbor page %lu (%u keys, %.1f%% full)",
3776 npp.mp_page->mp_pgno, NUMKEYS(npp.mp_page), (float)PAGEFILL(txn->mt_env, npp.mp_page) / 10);
3778 /* If the neighbor page is above threshold and has at least two
3779 * keys, move one key from it.
3781 * Otherwise we should try to merge them.
3783 if (PAGEFILL(txn->mt_env, npp.mp_page) >= FILL_THRESHOLD && NUMKEYS(npp.mp_page) >= 2)
3784 return mdb_move_node(txn, dbi, &npp, si, mpp, di);
3785 else { /* FIXME: if (has_enough_room()) */
3786 if (mpp->mp_pi == 0)
3787 return mdb_merge(txn, dbi, &npp, mpp);
3789 return mdb_merge(txn, dbi, mpp, &npp);
3794 mdb_del0(MDB_cursor *mc, unsigned int ki, MDB_pageparent *mpp, MDB_node *leaf)
3798 /* add overflow pages to free list */
3799 if (!IS_LEAF2(mpp->mp_page) && F_ISSET(leaf->mn_flags, F_BIGDATA)) {
3803 memcpy(&pg, NODEDATA(leaf), sizeof(pg));
3804 ovpages = OVPAGES(NODEDSZ(leaf), mc->mc_txn->mt_env->me_psize);
3805 for (i=0; i<ovpages; i++) {
3806 DPRINTF("freed ov page %lu", pg);
3807 mdb_midl_insert(mc->mc_txn->mt_free_pgs, pg);
3811 mdb_del_node(mpp->mp_page, ki, mc->mc_txn->mt_dbs[mc->mc_dbi].md_pad);
3812 mc->mc_txn->mt_dbs[mc->mc_dbi].md_entries--;
3813 rc = mdb_rebalance(mc->mc_txn, mc->mc_dbi, mpp);
3814 if (rc != MDB_SUCCESS)
3815 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
3821 mdb_del(MDB_txn *txn, MDB_dbi dbi,
3822 MDB_val *key, MDB_val *data)
3827 MDB_val rdata, *xdata;
3831 assert(key != NULL);
3833 DPRINTF("====> delete db %u key [%s]", dbi, DKEY(key));
3835 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
3838 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
3842 if (key->mv_size == 0 || key->mv_size > MAXKEYSIZE) {
3849 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
3850 mc.mc_xcursor = &mx;
3851 mdb_xcursor_init0(&mc);
3853 mc.mc_xcursor = NULL;
3865 rc = mdb_cursor_set(&mc, key, xdata, op, &exact);
3867 rc = mdb_cursor_del(&mc, data ? 0 : MDB_NODUPDATA);
3868 if (mc.mc_flags & C_XDIRTY) {
3869 mdb_cursor_reset(&mc);
3874 /* Split page <*mpp>, and insert <key,(data|newpgno)> in either left or
3875 * right sibling, at index <*newindxp> (as if unsplit). Updates *mpp and
3876 * *newindxp with the actual values after split, ie if *mpp and *newindxp
3877 * refer to a node in the new right sibling page.
3880 mdb_split(MDB_txn *txn, MDB_dbi dbi, MDB_page **mpp, unsigned int *newindxp,
3881 MDB_val *newkey, MDB_val *newdata, pgno_t newpgno)
3884 int rc = MDB_SUCCESS, ins_new = 0;
3887 unsigned int i, j, split_indx, nkeys, pmax;
3889 MDB_val sepkey, rkey, rdata;
3890 MDB_page *copy, *cptr;
3891 MDB_dpage *mdp, *rdp, *pdp;
3895 assert(txn != NULL);
3897 dh = ((MDB_dhead *)*mpp) - 1;
3898 mdp = (MDB_dpage *)dh;
3899 newindx = *newindxp;
3901 DPRINTF("-----> splitting %s page %lu and adding [%s] at index %i",
3902 IS_LEAF(&mdp->p) ? "leaf" : "branch", mdp->p.mp_pgno,
3903 DKEY(newkey), *newindxp);
3905 if (mdp->h.md_parent == NULL) {
3906 if ((pdp = mdb_new_page(txn, dbi, P_BRANCH, 1)) == NULL)
3909 mdp->h.md_parent = &pdp->p;
3910 txn->mt_dbs[dbi].md_root = pdp->p.mp_pgno;
3911 DPRINTF("root split! new root = %lu", pdp->p.mp_pgno);
3912 txn->mt_dbs[dbi].md_depth++;
3914 /* Add left (implicit) pointer. */
3915 if ((rc = mdb_add_node(txn, dbi, &pdp->p, 0, NULL, NULL,
3916 mdp->p.mp_pgno, 0)) != MDB_SUCCESS)
3919 DPRINTF("parent branch page is %lu", mdp->h.md_parent->mp_pgno);
3922 /* Create a right sibling. */
3923 if ((rdp = mdb_new_page(txn, dbi, mdp->p.mp_flags, 1)) == NULL)
3925 rdp->h.md_parent = mdp->h.md_parent;
3926 rdp->h.md_pi = mdp->h.md_pi + 1;
3927 DPRINTF("new right sibling: page %lu", rdp->p.mp_pgno);
3929 nkeys = NUMKEYS(&mdp->p);
3930 split_indx = nkeys / 2 + 1;
3932 if (IS_LEAF2(&rdp->p)) {
3935 unsigned int lsize, rsize, ksize;
3936 /* Move half of the keys to the right sibling */
3938 x = *newindxp - split_indx;
3939 ksize = txn->mt_dbs[dbi].md_pad;
3940 split = LEAF2KEY(&mdp->p, split_indx, ksize);
3941 rsize = (nkeys - split_indx) * ksize;
3942 lsize = (nkeys - split_indx) * sizeof(indx_t);
3943 mdp->p.mp_lower -= lsize;
3944 rdp->p.mp_lower += lsize;
3945 mdp->p.mp_upper += rsize - lsize;
3946 rdp->p.mp_upper -= rsize - lsize;
3947 sepkey.mv_size = ksize;
3948 if (newindx == split_indx) {
3949 sepkey.mv_data = newkey->mv_data;
3951 sepkey.mv_data = split;
3954 ins = LEAF2KEY(&mdp->p, *newindxp, ksize);
3955 memcpy(&rdp->p.mp_ptrs, split, rsize);
3956 sepkey.mv_data = &rdp->p.mp_ptrs;
3957 memmove(ins+ksize, ins, (split_indx - *newindxp) * ksize);
3958 memcpy(ins, newkey->mv_data, ksize);
3959 mdp->p.mp_lower += sizeof(indx_t);
3960 mdp->p.mp_upper -= ksize - sizeof(indx_t);
3963 memcpy(&rdp->p.mp_ptrs, split, x * ksize);
3964 ins = LEAF2KEY(&rdp->p, x, ksize);
3965 memcpy(ins, newkey->mv_data, ksize);
3966 memcpy(ins+ksize, split + x * ksize, rsize - x * ksize);
3967 rdp->p.mp_lower += sizeof(indx_t);
3968 rdp->p.mp_upper -= ksize - sizeof(indx_t);
3975 /* For leaf pages, check the split point based on what
3976 * fits where, since otherwise add_node can fail.
3978 if (IS_LEAF(&mdp->p)) {
3979 unsigned int psize, nsize;
3980 /* Maximum free space in an empty page */
3981 pmax = txn->mt_env->me_psize - PAGEHDRSZ;
3982 nsize = mdb_leaf_size(txn->mt_env, newkey, newdata);
3983 if (newindx < split_indx) {
3985 for (i=0; i<split_indx; i++) {
3986 node = NODEPTR(&mdp->p, i);
3987 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
3988 if (F_ISSET(node->mn_flags, F_BIGDATA))
3989 psize += sizeof(pgno_t);
3991 psize += NODEDSZ(node);
3999 for (i=nkeys-1; i>=split_indx; i--) {
4000 node = NODEPTR(&mdp->p, i);
4001 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
4002 if (F_ISSET(node->mn_flags, F_BIGDATA))
4003 psize += sizeof(pgno_t);
4005 psize += NODEDSZ(node);
4014 /* First find the separating key between the split pages.
4016 if (newindx == split_indx) {
4017 sepkey.mv_size = newkey->mv_size;
4018 sepkey.mv_data = newkey->mv_data;
4020 node = NODEPTR(&mdp->p, split_indx);
4021 sepkey.mv_size = node->mn_ksize;
4022 sepkey.mv_data = NODEKEY(node);
4026 DPRINTF("separator is [%s]", DKEY(&sepkey));
4028 /* Copy separator key to the parent.
4030 if (SIZELEFT(rdp->h.md_parent) < mdb_branch_size(txn->mt_env, &sepkey)) {
4031 rc = mdb_split(txn, dbi, &rdp->h.md_parent, &rdp->h.md_pi,
4032 &sepkey, NULL, rdp->p.mp_pgno);
4034 /* Right page might now have changed parent.
4035 * Check if left page also changed parent.
4037 if (rdp->h.md_parent != mdp->h.md_parent &&
4038 mdp->h.md_pi >= NUMKEYS(mdp->h.md_parent)) {
4039 mdp->h.md_parent = rdp->h.md_parent;
4040 mdp->h.md_pi = rdp->h.md_pi - 1;
4043 rc = mdb_add_node(txn, dbi, rdp->h.md_parent, rdp->h.md_pi,
4044 &sepkey, NULL, rdp->p.mp_pgno, 0);
4046 if (IS_LEAF2(&rdp->p)) {
4049 if (rc != MDB_SUCCESS) {
4053 /* Move half of the keys to the right sibling. */
4054 if ((copy = malloc(txn->mt_env->me_psize)) == NULL)
4057 copy->mp_pgno = mdp->p.mp_pgno;
4058 copy->mp_flags = mdp->p.mp_flags;
4059 copy->mp_lower = PAGEHDRSZ;
4060 copy->mp_upper = txn->mt_env->me_psize;
4062 for (i = j = 0; i <= nkeys; j++) {
4063 if (i == split_indx) {
4064 /* Insert in right sibling. */
4065 /* Reset insert index for right sibling. */
4066 j = (i == newindx && ins_new);
4070 if (i == newindx && !ins_new) {
4071 /* Insert the original entry that caused the split. */
4072 rkey.mv_data = newkey->mv_data;
4073 rkey.mv_size = newkey->mv_size;
4074 if (IS_LEAF(&mdp->p)) {
4075 rdata.mv_data = newdata->mv_data;
4076 rdata.mv_size = newdata->mv_size;
4083 /* Update page and index for the new key. */
4085 if (cptr == &rdp->p)
4087 } else if (i == nkeys) {
4090 node = NODEPTR(&mdp->p, i);
4091 rkey.mv_data = NODEKEY(node);
4092 rkey.mv_size = node->mn_ksize;
4093 if (IS_LEAF(&mdp->p)) {
4094 rdata.mv_data = NODEDATA(node);
4095 rdata.mv_size = node->mn_dsize;
4097 pgno = NODEPGNO(node);
4098 flags = node->mn_flags;
4103 if (!IS_LEAF(&mdp->p) && j == 0) {
4104 /* First branch index doesn't need key data. */
4108 rc = mdb_add_node(txn, dbi, cptr, j, &rkey, &rdata, pgno, flags);
4110 nkeys = NUMKEYS(copy);
4111 for (i=0; i<nkeys; i++)
4112 mdp->p.mp_ptrs[i] = copy->mp_ptrs[i];
4113 mdp->p.mp_lower = copy->mp_lower;
4114 mdp->p.mp_upper = copy->mp_upper;
4115 memcpy(NODEPTR(&mdp->p, nkeys-1), NODEPTR(copy, nkeys-1),
4116 txn->mt_env->me_psize - copy->mp_upper);
4124 mdb_put0(MDB_txn *txn, MDB_dbi dbi,
4125 MDB_val *key, MDB_val *data, unsigned int flags)
4127 int rc = MDB_SUCCESS, exact;
4131 MDB_val xdata, *rdata, dkey;
4133 char dbuf[PAGESIZE];
4138 DPRINTF("==> put db %u key [%s], size %zu, data size %zu",
4139 dbi, DKEY(key), key->mv_size, data->mv_size);
4142 mpp.mp_parent = NULL;
4144 rc = mdb_search_page(txn, dbi, key, NULL, 1, &mpp);
4145 if (rc == MDB_SUCCESS) {
4146 leaf = mdb_search_node(txn, dbi, mpp.mp_page, key, &exact, &ki);
4147 if (leaf && exact) {
4148 if (flags == MDB_NOOVERWRITE) {
4149 DPRINTF("duplicate key [%s]", DKEY(key));
4150 return MDB_KEYEXIST;
4152 /* there's only a key anyway, so this is a no-op */
4153 if (IS_LEAF2(mpp.mp_page))
4156 if (F_ISSET(txn->mt_dbs[dbi].md_flags, MDB_DUPSORT)) {
4157 /* Was a single item before, must convert now */
4158 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4159 dkey.mv_size = NODEDSZ(leaf);
4160 dkey.mv_data = dbuf;
4161 memcpy(dbuf, NODEDATA(leaf), dkey.mv_size);
4162 /* data matches, ignore it */
4163 if (!mdb_dcmp(txn, dbi, data, &dkey))
4164 return (flags == MDB_NODUPDATA) ? MDB_KEYEXIST : MDB_SUCCESS;
4165 memset(&dummy, 0, sizeof(dummy));
4166 if (txn->mt_dbs[dbi].md_flags & MDB_DUPFIXED) {
4167 dummy.md_pad = data->mv_size;
4168 dummy.md_flags = MDB_DUPFIXED;
4169 if (txn->mt_dbs[dbi].md_flags & MDB_INTEGERDUP)
4170 dummy.md_flags |= MDB_INTEGERKEY;
4172 dummy.md_root = P_INVALID;
4173 if (dkey.mv_size == sizeof(MDB_db)) {
4174 memcpy(NODEDATA(leaf), &dummy, sizeof(dummy));
4177 mdb_del_node(mpp.mp_page, ki, 0);
4180 xdata.mv_size = sizeof(MDB_db);
4181 xdata.mv_data = &dummy;
4186 /* same size, just replace it */
4187 if (!F_ISSET(leaf->mn_flags, F_BIGDATA) &&
4188 NODEDSZ(leaf) == data->mv_size) {
4189 memcpy(NODEDATA(leaf), data->mv_data, data->mv_size);
4192 mdb_del_node(mpp.mp_page, ki, 0);
4194 if (leaf == NULL) { /* append if not found */
4195 ki = NUMKEYS(mpp.mp_page);
4196 DPRINTF("appending key at index %i", ki);
4198 } else if (rc == MDB_NOTFOUND) {
4200 /* new file, just write a root leaf page */
4201 DPUTS("allocating new root leaf page");
4202 if ((dp = mdb_new_page(txn, dbi, P_LEAF, 1)) == NULL) {
4205 mpp.mp_page = &dp->p;
4206 txn->mt_dbs[dbi].md_root = mpp.mp_page->mp_pgno;
4207 txn->mt_dbs[dbi].md_depth++;
4208 txn->mt_dbxs[dbi].md_dirty = 1;
4209 if ((txn->mt_dbs[dbi].md_flags & (MDB_DUPSORT|MDB_DUPFIXED)) == MDB_DUPFIXED)
4210 mpp.mp_page->mp_flags |= P_LEAF2;
4216 assert(IS_LEAF(mpp.mp_page));
4217 DPRINTF("there are %u keys, should insert new key at index %i",
4218 NUMKEYS(mpp.mp_page), ki);
4223 nsize = IS_LEAF2(mpp.mp_page) ? key->mv_size : mdb_leaf_size(txn->mt_env, key, rdata);
4224 if (SIZELEFT(mpp.mp_page) < nsize) {
4225 rc = mdb_split(txn, dbi, &mpp.mp_page, &ki, key, rdata, P_INVALID);
4227 /* There is room already in this leaf page. */
4228 rc = mdb_add_node(txn, dbi, mpp.mp_page, ki, key, rdata, 0, 0);
4231 if (rc != MDB_SUCCESS)
4232 txn->mt_flags |= MDB_TXN_ERROR;
4234 /* Remember if we just added a subdatabase */
4235 if (flags & F_SUBDATA) {
4236 leaf = NODEPTR(mpp.mp_page, ki);
4237 leaf->mn_flags |= F_SUBDATA;
4240 /* Now store the actual data in the child DB. Note that we're
4241 * storing the user data in the keys field, so there are strict
4242 * size limits on dupdata. The actual data fields of the child
4243 * DB are all zero size.
4249 leaf = NODEPTR(mpp.mp_page, ki);
4254 mc.mc_xcursor = &mx;
4255 mdb_xcursor_init0(&mc);
4256 mdb_xcursor_init1(txn, dbi, &mx, mpp.mp_page, leaf);
4259 if (flags == MDB_NODUPDATA)
4260 flags = MDB_NOOVERWRITE;
4261 /* converted, write the original data first */
4263 rc = mdb_put0(&mx.mx_txn, mx.mx_cursor.mc_dbi, &dkey, &xdata, flags);
4265 leaf->mn_flags |= F_DUPDATA;
4267 rc = mdb_put0(&mx.mx_txn, mx.mx_cursor.mc_dbi, data, &xdata, flags);
4268 mdb_xcursor_fini(&mc);
4269 memcpy(NODEDATA(leaf), &mx.mx_txn.mt_dbs[mx.mx_cursor.mc_dbi],
4272 txn->mt_dbs[dbi].md_entries++;
4281 mdb_put(MDB_txn *txn, MDB_dbi dbi,
4282 MDB_val *key, MDB_val *data, unsigned int flags)
4288 assert(key != NULL);
4289 assert(data != NULL);
4291 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
4294 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
4298 if (key->mv_size == 0 || key->mv_size > MAXKEYSIZE) {
4302 if ((flags & (MDB_NOOVERWRITE|MDB_NODUPDATA)) != flags)
4309 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
4310 mc.mc_xcursor = &mx;
4311 mdb_xcursor_init0(&mc);
4313 mc.mc_xcursor = NULL;
4315 rc = mdb_cursor_put(&mc, key, data, flags);
4316 if (mc.mc_flags & C_XDIRTY) {
4317 mdb_cursor_reset(&mc);
4323 mdb_env_set_flags(MDB_env *env, unsigned int flag, int onoff)
4325 #define CHANGEABLE (MDB_NOSYNC)
4326 if ((flag & CHANGEABLE) != flag)
4329 env->me_flags |= flag;
4331 env->me_flags &= ~flag;
4336 mdb_env_get_flags(MDB_env *env, unsigned int *arg)
4341 *arg = env->me_flags;
4346 mdb_env_get_path(MDB_env *env, const char **arg)
4351 *arg = env->me_path;
4356 mdb_stat0(MDB_env *env, MDB_db *db, MDB_stat *arg)
4358 arg->ms_psize = env->me_psize;
4359 arg->ms_depth = db->md_depth;
4360 arg->ms_branch_pages = db->md_branch_pages;
4361 arg->ms_leaf_pages = db->md_leaf_pages;
4362 arg->ms_overflow_pages = db->md_overflow_pages;
4363 arg->ms_entries = db->md_entries;
4368 mdb_env_stat(MDB_env *env, MDB_stat *arg)
4372 if (env == NULL || arg == NULL)
4375 mdb_env_read_meta(env, &toggle);
4377 return mdb_stat0(env, &env->me_metas[toggle]->mm_dbs[MAIN_DBI], arg);
4380 int mdb_open(MDB_txn *txn, const char *name, unsigned int flags, MDB_dbi *dbi)
4390 if (flags & (MDB_DUPSORT|MDB_REVERSEKEY|MDB_INTEGERKEY))
4391 txn->mt_dbs[MAIN_DBI].md_flags |= (flags & (MDB_DUPSORT|MDB_REVERSEKEY|MDB_INTEGERKEY));
4395 /* Is the DB already open? */
4397 for (i=2; i<txn->mt_numdbs; i++) {
4398 if (len == txn->mt_dbxs[i].md_name.mv_size &&
4399 !strncmp(name, txn->mt_dbxs[i].md_name.mv_data, len)) {
4405 if (txn->mt_numdbs >= txn->mt_env->me_maxdbs - 1)
4408 /* Find the DB info */
4410 key.mv_data = (void *)name;
4411 rc = mdb_get(txn, MAIN_DBI, &key, &data);
4413 /* Create if requested */
4414 if (rc == MDB_NOTFOUND && (flags & MDB_CREATE)) {
4417 data.mv_size = sizeof(MDB_db);
4418 data.mv_data = &dummy;
4419 memset(&dummy, 0, sizeof(dummy));
4420 dummy.md_root = P_INVALID;
4421 dummy.md_flags = flags & 0xffff;
4423 mc.mc_dbi = MAIN_DBI;
4425 rc = mdb_cursor_put(&mc, &key, &data, F_SUBDATA);
4429 /* OK, got info, add to table */
4430 if (rc == MDB_SUCCESS) {
4431 txn->mt_dbxs[txn->mt_numdbs].md_name.mv_data = strdup(name);
4432 txn->mt_dbxs[txn->mt_numdbs].md_name.mv_size = len;
4433 txn->mt_dbxs[txn->mt_numdbs].md_cmp = NULL;
4434 txn->mt_dbxs[txn->mt_numdbs].md_dcmp = NULL;
4435 txn->mt_dbxs[txn->mt_numdbs].md_rel = NULL;
4436 txn->mt_dbxs[txn->mt_numdbs].md_parent = MAIN_DBI;
4437 txn->mt_dbxs[txn->mt_numdbs].md_dirty = dirty;
4438 memcpy(&txn->mt_dbs[txn->mt_numdbs], data.mv_data, sizeof(MDB_db));
4439 *dbi = txn->mt_numdbs;
4440 txn->mt_env->me_dbs[0][txn->mt_numdbs] = txn->mt_dbs[txn->mt_numdbs];
4441 txn->mt_env->me_dbs[1][txn->mt_numdbs] = txn->mt_dbs[txn->mt_numdbs];
4448 int mdb_stat(MDB_txn *txn, MDB_dbi dbi, MDB_stat *arg)
4450 if (txn == NULL || arg == NULL || dbi >= txn->mt_numdbs)
4453 return mdb_stat0(txn->mt_env, &txn->mt_dbs[dbi], arg);
4456 void mdb_close(MDB_txn *txn, MDB_dbi dbi)
4459 if (dbi <= MAIN_DBI || dbi >= txn->mt_numdbs)
4461 ptr = txn->mt_dbxs[dbi].md_name.mv_data;
4462 txn->mt_dbxs[dbi].md_name.mv_data = NULL;
4463 txn->mt_dbxs[dbi].md_name.mv_size = 0;
4467 int mdb_set_compare(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
4469 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
4472 txn->mt_dbxs[dbi].md_cmp = cmp;
4476 int mdb_set_dupsort(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
4478 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
4481 txn->mt_dbxs[dbi].md_dcmp = cmp;
4485 int mdb_set_relfunc(MDB_txn *txn, MDB_dbi dbi, MDB_rel_func *rel)
4487 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
4490 txn->mt_dbxs[dbi].md_rel = rel;