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
252 #define mp_next mp_p.p_align
254 pgno_t p_pgno; /* page number */
255 void * p_align; /* for IL32P64 */
257 #define P_BRANCH 0x01 /* branch page */
258 #define P_LEAF 0x02 /* leaf page */
259 #define P_OVERFLOW 0x04 /* overflow page */
260 #define P_META 0x08 /* meta page */
261 #define P_DIRTY 0x10 /* dirty page */
262 #define P_LEAF2 0x20 /* DB with small, fixed size keys and no data */
264 #define mp_lower mp_pb.pb.pb_lower
265 #define mp_upper mp_pb.pb.pb_upper
266 #define mp_pages mp_pb.pb_pages
269 indx_t pb_lower; /* lower bound of free space */
270 indx_t pb_upper; /* upper bound of free space */
272 uint32_t pb_pages; /* number of overflow pages */
274 indx_t mp_ptrs[1]; /* dynamic size */
277 #define PAGEHDRSZ ((unsigned) offsetof(MDB_page, mp_ptrs))
279 #define NUMKEYS(p) (((p)->mp_lower - PAGEHDRSZ) >> 1)
280 #define SIZELEFT(p) (indx_t)((p)->mp_upper - (p)->mp_lower)
281 #define PAGEFILL(env, p) (1000L * ((env)->me_psize - PAGEHDRSZ - SIZELEFT(p)) / \
282 ((env)->me_psize - PAGEHDRSZ))
283 #define IS_LEAF(p) F_ISSET((p)->mp_flags, P_LEAF)
284 #define IS_LEAF2(p) F_ISSET((p)->mp_flags, P_LEAF2)
285 #define IS_BRANCH(p) F_ISSET((p)->mp_flags, P_BRANCH)
286 #define IS_OVERFLOW(p) F_ISSET((p)->mp_flags, P_OVERFLOW)
288 #define OVPAGES(size, psize) ((PAGEHDRSZ-1 + (size)) / (psize) + 1)
290 typedef struct MDB_db {
291 uint32_t md_pad; /* also ksize for LEAF2 pages */
294 ULONG md_branch_pages;
296 ULONG md_overflow_pages;
304 typedef struct MDB_meta { /* meta (footer) page content */
307 void *mm_address; /* address for fixed mapping */
308 size_t mm_mapsize; /* size of mmap region */
309 MDB_db mm_dbs[2]; /* first is free space, 2nd is main db */
310 #define mm_psize mm_dbs[0].md_pad
311 #define mm_flags mm_dbs[0].md_flags
312 pgno_t mm_last_pg; /* last used page in file */
313 ULONG mm_txnid; /* txnid that committed this page */
316 typedef struct MDB_oldpages {
317 struct MDB_oldpages *mo_next;
319 pgno_t mo_pages[1]; /* dynamic */
322 static MDB_page *mdb_alloc_page(MDB_cursor *mc, int num);
323 static int mdb_touch(MDB_cursor *mc);
325 /* Enough space for 2^32 nodes with minimum of 2 keys per node. I.e., plenty.
326 * At 4 keys per node, enough for 2^64 nodes, so there's probably no need to
327 * raise this on a 64 bit machine.
329 #define CURSOR_STACK 32
334 struct MDB_xcursor *mc_xcursor;
337 unsigned short mc_snum; /* number of pushed pages */
338 unsigned short mc_top; /* index of top page, mc_snum-1 */
339 unsigned int mc_flags;
340 #define C_INITIALIZED 0x01
342 #define C_XDIRTY 0x04
343 MDB_page *mc_pg[CURSOR_STACK]; /* stack of pushed pages */
344 indx_t mc_ki[CURSOR_STACK]; /* stack of page indices */
347 #define METADATA(p) ((void *)((char *)(p) + PAGEHDRSZ))
349 /* We guarantee 2-byte alignment for nodes */
350 typedef struct MDB_node {
351 /* lo and hi are used for data size on leaf nodes and for
352 * child pgno on branch nodes. On 64 bit platforms, flags
353 * is also used for pgno. (branch nodes ignore flags)
355 unsigned short mn_lo;
356 unsigned short mn_hi;
357 unsigned short mn_flags;
358 #define F_BIGDATA 0x01 /* data put on overflow page */
359 #define F_SUBDATA 0x02 /* data is a sub-database */
360 #define F_DUPDATA 0x04 /* data has duplicates */
361 unsigned short mn_ksize; /* key size */
365 typedef struct MDB_dbx {
367 MDB_cmp_func *md_cmp; /* user compare function */
368 MDB_cmp_func *md_dcmp; /* user dupsort function */
369 MDB_rel_func *md_rel; /* user relocate function */
371 unsigned int md_dirty;
375 pgno_t mt_next_pgno; /* next unallocated page */
378 pgno_t *mt_free_pgs; /* this is an IDL */
380 ID2L dirty_list; /* modified pages */
383 MDB_dbx *mt_dbxs; /* array */
385 unsigned int mt_numdbs;
387 #define MDB_TXN_RDONLY 0x01 /* read-only transaction */
388 #define MDB_TXN_ERROR 0x02 /* an error has occurred */
389 unsigned int mt_flags;
390 unsigned int mt_toggle;
393 /* Context for sorted-dup records */
394 typedef struct MDB_xcursor {
395 MDB_cursor mx_cursor;
404 HANDLE me_mfd; /* just for writing the meta pages */
405 #define MDB_FATAL_ERROR 0x80000000U
407 uint32_t me_extrapad; /* unused for now */
408 unsigned int me_maxreaders;
409 unsigned int me_numdbs;
410 unsigned int me_maxdbs;
413 MDB_txninfo *me_txns;
414 MDB_meta *me_metas[2];
415 MDB_txn *me_txn; /* current write transaction */
417 off_t me_size; /* current file size */
418 pgno_t me_maxpg; /* me_mapsize / me_psize */
419 unsigned int me_psize;
420 unsigned int me_db_toggle;
421 MDB_dbx *me_dbxs; /* array */
423 MDB_oldpages *me_pghead;
424 pthread_key_t me_txkey; /* thread-key for readers */
426 pgno_t me_free_pgs[MDB_IDL_UM_SIZE];
427 ID2 me_dirty_list[MDB_IDL_UM_SIZE];
428 LAZY_RWLOCK_DEF(me_dblock);
430 HANDLE me_rmutex; /* Windows mutexes don't reside in shared mem */
435 #define NODESIZE offsetof(MDB_node, mn_data)
437 #define INDXSIZE(k) (NODESIZE + ((k) == NULL ? 0 : (k)->mv_size))
438 #define LEAFSIZE(k, d) (NODESIZE + (k)->mv_size + (d)->mv_size)
439 #define NODEPTR(p, i) ((MDB_node *)((char *)(p) + (p)->mp_ptrs[i]))
440 #define NODEKEY(node) (void *)((node)->mn_data)
441 #define NODEDATA(node) (void *)((char *)(node)->mn_data + (node)->mn_ksize)
442 #if LONG_MAX == 0x7fffffff
443 #define NODEPGNO(node) ((node)->mn_lo | ((node)->mn_hi << 16))
444 #define SETPGNO(node,pgno) do { \
445 (node)->mn_lo = (pgno) & 0xffff; (node)->mn_hi = (pgno) >> 16;} while(0)
447 #define NODEPGNO(node) ((node)->mn_lo | ((node)->mn_hi << 16) | ((unsigned long)(node)->mn_flags << 32))
448 #define SETPGNO(node,pgno) do { \
449 (node)->mn_lo = (pgno) & 0xffff; (node)->mn_hi = (pgno) >> 16; \
450 (node)->mn_flags = (pgno) >> 32; } while(0)
452 #define NODEDSZ(node) ((node)->mn_lo | ((unsigned)(node)->mn_hi << 16))
453 #define SETDSZ(node,size) do { \
454 (node)->mn_lo = (size) & 0xffff; (node)->mn_hi = (size) >> 16;} while(0)
455 #define NODEKSZ(node) ((node)->mn_ksize)
456 #define LEAF2KEY(p, i, ks) ((char *)(p) + PAGEHDRSZ + ((i)*(ks)))
458 #define MDB_SET_KEY(node, key) if (key!=NULL) {(key)->mv_size = NODEKSZ(node); (key)->mv_data = NODEKEY(node);}
460 #define MDB_COMMIT_PAGES 64 /* max number of pages to write in one commit */
462 static int mdb_search_page_root(MDB_cursor *mc,
463 MDB_val *key, int modify);
464 static int mdb_search_page(MDB_cursor *mc,
465 MDB_val *key, int modify);
467 static int mdb_env_read_header(MDB_env *env, MDB_meta *meta);
468 static int mdb_env_read_meta(MDB_env *env, int *which);
469 static int mdb_env_write_meta(MDB_txn *txn);
470 static int mdb_get_page(MDB_txn *txn, pgno_t pgno, MDB_page **mp);
472 static MDB_node *mdb_search_node(MDB_cursor *mc, MDB_val *key, int *exactp);
473 static int mdb_add_node(MDB_cursor *mc, indx_t indx,
474 MDB_val *key, MDB_val *data, pgno_t pgno, uint8_t flags);
475 static void mdb_del_node(MDB_page *mp, indx_t indx, int ksize);
476 static int mdb_del0(MDB_cursor *mc, MDB_node *leaf);
477 static int mdb_read_data(MDB_txn *txn, MDB_node *leaf, MDB_val *data);
479 static int mdb_rebalance(MDB_cursor *mc);
480 static int mdb_update_key(MDB_page *mp, indx_t indx, MDB_val *key);
481 static int mdb_move_node(MDB_cursor *csrcrc, MDB_cursor *cdstst);
482 static int mdb_merge(MDB_cursor *csrcrc, MDB_cursor *cdstst);
483 static int mdb_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata,
485 static MDB_page *mdb_new_page(MDB_cursor *mc, uint32_t flags, int num);
487 static void cursor_pop_page(MDB_cursor *mc);
488 static int cursor_push_page(MDB_cursor *mc, MDB_page *mp);
490 static int mdb_sibling(MDB_cursor *mc, int move_right);
491 static int mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
492 static int mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
493 static int mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op,
495 static int mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data);
496 static int mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data);
498 static void mdb_xcursor_init0(MDB_cursor *mc);
499 static void mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node);
500 static void mdb_xcursor_init2(MDB_cursor *mc);
501 static void mdb_xcursor_fini(MDB_cursor *mc);
503 static size_t mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data);
504 static size_t mdb_branch_size(MDB_env *env, MDB_val *key);
506 static void mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi);
508 static MDB_cmp_func memncmp, memnrcmp, intcmp, cintcmp;
511 static SECURITY_DESCRIPTOR mdb_null_sd;
512 static SECURITY_ATTRIBUTES mdb_all_sa;
513 static int mdb_sec_inited;
517 mdb_version(int *maj, int *min, int *pat)
519 if (maj) *maj = MDB_VERSION_MAJOR;
520 if (min) *min = MDB_VERSION_MINOR;
521 if (pat) *pat = MDB_VERSION_PATCH;
522 return MDB_VERSION_STRING;
525 static char *const mdb_errstr[] = {
526 "MDB_KEYEXIST: Key/data pair already exists",
527 "MDB_NOTFOUND: No matching key/data pair found",
528 "MDB_PAGE_NOTFOUND: Requested page not found",
529 "MDB_CORRUPTED: Located page was wrong type",
530 "MDB_PANIC: Update of meta page failed",
531 "MDB_VERSION_MISMATCH: Database environment version mismatch"
535 mdb_strerror(int err)
538 return ("Successful return: 0");
540 if (err >= MDB_KEYEXIST && err <= MDB_VERSION_MISMATCH)
541 return mdb_errstr[err - MDB_KEYEXIST];
543 return strerror(err);
548 mdb_dkey(MDB_val *key, char *buf)
551 unsigned char *c = key->mv_data;
553 if (key->mv_size > MAXKEYSIZE)
556 for (i=0; i<key->mv_size; i++)
557 ptr += sprintf(ptr, "%02x", *c++);
559 sprintf(buf, "%.*s", key->mv_size, key->mv_data);
566 mdb_cmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
568 return txn->mt_dbxs[dbi].md_cmp(a, b);
572 mdb_dcmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
574 if (txn->mt_dbxs[dbi].md_dcmp)
575 return txn->mt_dbxs[dbi].md_dcmp(a, b);
577 return EINVAL; /* too bad you can't distinguish this from a valid result */
580 /* Allocate new page(s) for writing */
582 mdb_alloc_page(MDB_cursor *mc, int num)
584 MDB_txn *txn = mc->mc_txn;
586 pgno_t pgno = P_INVALID;
589 if (txn->mt_txnid > 2) {
591 if (!txn->mt_env->me_pghead && mc->mc_dbi != FREE_DBI &&
592 txn->mt_dbs[FREE_DBI].md_root != P_INVALID) {
593 /* See if there's anything in the free DB */
599 m2.mc_dbi = FREE_DBI;
602 mdb_search_page(&m2, NULL, 0);
603 leaf = NODEPTR(m2.mc_pg[m2.mc_top], 0);
604 kptr = (ULONG *)NODEKEY(leaf);
608 oldest = txn->mt_txnid - 1;
609 for (i=0; i<txn->mt_env->me_txns->mti_numreaders; i++) {
610 ULONG mr = txn->mt_env->me_txns->mti_readers[i].mr_txnid;
611 if (mr && mr < oldest)
616 if (oldest > *kptr) {
617 /* It's usable, grab it.
623 mdb_read_data(txn, leaf, &data);
624 idl = (ULONG *)data.mv_data;
625 mop = malloc(sizeof(MDB_oldpages) + MDB_IDL_SIZEOF(idl) - sizeof(pgno_t));
626 mop->mo_next = txn->mt_env->me_pghead;
627 mop->mo_txnid = *kptr;
628 txn->mt_env->me_pghead = mop;
629 memcpy(mop->mo_pages, idl, MDB_IDL_SIZEOF(idl));
634 DPRINTF("IDL read txn %lu root %lu num %lu",
635 mop->mo_txnid, txn->mt_dbs[FREE_DBI].md_root, idl[0]);
636 for (i=0; i<idl[0]; i++) {
637 DPRINTF("IDL %lu", idl[i+1]);
641 /* drop this IDL from the DB */
642 m2.mc_ki[m2.mc_top] = 0;
643 m2.mc_flags = C_INITIALIZED;
644 mdb_cursor_del(&m2, 0);
647 if (txn->mt_env->me_pghead) {
648 MDB_oldpages *mop = txn->mt_env->me_pghead;
650 /* FIXME: For now, always use fresh pages. We
651 * really ought to search the free list for a
656 /* peel pages off tail, so we only have to truncate the list */
657 pgno = MDB_IDL_LAST(mop->mo_pages);
658 if (MDB_IDL_IS_RANGE(mop->mo_pages)) {
660 if (mop->mo_pages[2] > mop->mo_pages[1])
661 mop->mo_pages[0] = 0;
665 if (MDB_IDL_IS_ZERO(mop->mo_pages)) {
666 txn->mt_env->me_pghead = mop->mo_next;
673 if (pgno == P_INVALID) {
674 /* DB size is maxed out */
675 if (txn->mt_next_pgno + num >= txn->mt_env->me_maxpg)
678 if (txn->mt_env->me_dpages && num == 1) {
679 np = txn->mt_env->me_dpages;
680 txn->mt_env->me_dpages = np->mp_next;
682 if ((np = malloc(txn->mt_env->me_psize * num )) == NULL)
685 if (pgno == P_INVALID) {
686 np->mp_pgno = txn->mt_next_pgno;
687 txn->mt_next_pgno += num;
691 mid.mid = np->mp_pgno;
693 mdb_mid2l_insert(txn->mt_u.dirty_list, &mid);
698 /* Touch a page: make it dirty and re-insert into tree with updated pgno.
701 mdb_touch(MDB_cursor *mc)
703 MDB_page *mp = mc->mc_pg[mc->mc_top];
706 if (!F_ISSET(mp->mp_flags, P_DIRTY)) {
708 if ((np = mdb_alloc_page(mc, 1)) == NULL)
710 DPRINTF("touched db %u page %lu -> %lu", mc->mc_dbi, mp->mp_pgno, np->mp_pgno);
711 assert(mp->mp_pgno != np->mp_pgno);
712 mdb_midl_append(mc->mc_txn->mt_free_pgs, mp->mp_pgno);
714 memcpy(np, mp, mc->mc_txn->mt_env->me_psize);
717 mp->mp_flags |= P_DIRTY;
719 mc->mc_pg[mc->mc_top] = mp;
720 /* Update the page number to new touched page. */
722 SETPGNO(NODEPTR(mc->mc_pg[mc->mc_top-1], mc->mc_ki[mc->mc_top-1]), mp->mp_pgno);
728 mdb_env_sync(MDB_env *env, int force)
731 if (force || !F_ISSET(env->me_flags, MDB_NOSYNC)) {
732 if (fdatasync(env->me_fd))
739 mdb_txn_reset0(MDB_txn *txn);
742 mdb_txn_renew0(MDB_txn *txn)
744 MDB_env *env = txn->mt_env;
746 if (txn->mt_flags & MDB_TXN_RDONLY) {
747 MDB_reader *r = pthread_getspecific(env->me_txkey);
750 pid_t pid = getpid();
751 pthread_t tid = pthread_self();
754 for (i=0; i<env->me_txns->mti_numreaders; i++)
755 if (env->me_txns->mti_readers[i].mr_pid == 0)
757 if (i == env->me_maxreaders) {
761 env->me_txns->mti_readers[i].mr_pid = pid;
762 env->me_txns->mti_readers[i].mr_tid = tid;
763 if (i >= env->me_txns->mti_numreaders)
764 env->me_txns->mti_numreaders = i+1;
766 r = &env->me_txns->mti_readers[i];
767 pthread_setspecific(env->me_txkey, r);
769 txn->mt_txnid = env->me_txns->mti_txnid;
770 txn->mt_toggle = env->me_txns->mti_me_toggle;
771 r->mr_txnid = txn->mt_txnid;
772 txn->mt_u.reader = r;
776 txn->mt_txnid = env->me_txns->mti_txnid+1;
777 txn->mt_toggle = env->me_txns->mti_me_toggle;
778 txn->mt_u.dirty_list = env->me_dirty_list;
779 txn->mt_u.dirty_list[0].mid = 0;
780 txn->mt_free_pgs = env->me_free_pgs;
781 txn->mt_free_pgs[0] = 0;
782 txn->mt_next_pgno = env->me_metas[txn->mt_toggle]->mm_last_pg+1;
786 /* Copy the DB arrays */
787 LAZY_RWLOCK_RDLOCK(&env->me_dblock);
788 txn->mt_numdbs = env->me_numdbs;
789 txn->mt_dbxs = env->me_dbxs; /* mostly static anyway */
790 memcpy(txn->mt_dbs, env->me_metas[txn->mt_toggle]->mm_dbs, 2 * sizeof(MDB_db));
791 if (txn->mt_numdbs > 2)
792 memcpy(txn->mt_dbs+2, env->me_dbs[env->me_db_toggle]+2,
793 (txn->mt_numdbs - 2) * sizeof(MDB_db));
794 LAZY_RWLOCK_UNLOCK(&env->me_dblock);
800 mdb_txn_renew(MDB_txn *txn)
807 if (txn->mt_env->me_flags & MDB_FATAL_ERROR) {
808 DPUTS("environment had fatal error, must shutdown!");
812 rc = mdb_txn_renew0(txn);
813 if (rc == MDB_SUCCESS) {
814 DPRINTF("renew txn %lu%c %p on mdbenv %p, root page %lu",
815 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w', txn,
816 (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root);
822 mdb_txn_begin(MDB_env *env, unsigned int flags, MDB_txn **ret)
827 if (env->me_flags & MDB_FATAL_ERROR) {
828 DPUTS("environment had fatal error, must shutdown!");
831 if ((txn = calloc(1, sizeof(MDB_txn) + env->me_maxdbs * sizeof(MDB_db))) == NULL) {
832 DPRINTF("calloc: %s", strerror(ErrCode()));
835 txn->mt_dbs = (MDB_db *)(txn+1);
836 if (flags & MDB_RDONLY) {
837 txn->mt_flags |= MDB_TXN_RDONLY;
841 rc = mdb_txn_renew0(txn);
846 DPRINTF("begin txn %lu%c %p on mdbenv %p, root page %lu",
847 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w', txn,
848 (void *) env, txn->mt_dbs[MAIN_DBI].md_root);
855 mdb_txn_reset0(MDB_txn *txn)
857 MDB_env *env = txn->mt_env;
859 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
860 txn->mt_u.reader->mr_txnid = 0;
866 /* return all dirty pages to dpage list */
867 for (i=1; i<=txn->mt_u.dirty_list[0].mid; i++) {
868 dp = txn->mt_u.dirty_list[i].mptr;
869 if (!IS_OVERFLOW(dp) || dp->mp_pages == 1) {
870 dp->mp_next = txn->mt_env->me_dpages;
871 txn->mt_env->me_dpages = dp;
873 /* large pages just get freed directly */
878 while ((mop = txn->mt_env->me_pghead)) {
879 txn->mt_env->me_pghead = mop->mo_next;
884 for (i=2; i<env->me_numdbs; i++)
885 env->me_dbxs[i].md_dirty = 0;
891 mdb_txn_reset(MDB_txn *txn)
896 DPRINTF("reset txn %lu%c %p on mdbenv %p, root page %lu",
897 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w', txn,
898 (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root);
904 mdb_txn_abort(MDB_txn *txn)
909 DPRINTF("abort txn %lu%c %p on mdbenv %p, root page %lu",
910 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w', txn,
911 (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root);
918 mdb_txn_commit(MDB_txn *txn)
930 assert(txn->mt_env != NULL);
934 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
939 if (txn != env->me_txn) {
940 DPUTS("attempt to commit unknown transaction");
945 if (F_ISSET(txn->mt_flags, MDB_TXN_ERROR)) {
946 DPUTS("error flag is set, can't commit");
951 if (!txn->mt_u.dirty_list[0].mid)
954 DPRINTF("committing txn %lu %p on mdbenv %p, root page %lu",
955 txn->mt_txnid, txn, (void *)env, txn->mt_dbs[MAIN_DBI].md_root);
958 mc.mc_dbi = FREE_DBI;
961 /* should only be one record now */
962 if (env->me_pghead) {
963 /* make sure first page of freeDB is touched and on freelist */
964 mdb_search_page(&mc, NULL, 1);
966 /* save to free list */
967 if (!MDB_IDL_IS_ZERO(txn->mt_free_pgs)) {
971 /* make sure last page of freeDB is touched and on freelist */
972 key.mv_size = MAXKEYSIZE+1;
974 mdb_search_page(&mc, &key, 1);
976 mdb_midl_sort(txn->mt_free_pgs);
980 ULONG *idl = txn->mt_free_pgs;
981 DPRINTF("IDL write txn %lu root %lu num %lu",
982 txn->mt_txnid, txn->mt_dbs[FREE_DBI].md_root, idl[0]);
983 for (i=0; i<idl[0]; i++) {
984 DPRINTF("IDL %lu", idl[i+1]);
988 /* write to last page of freeDB */
989 key.mv_size = sizeof(pgno_t);
990 key.mv_data = (char *)&txn->mt_txnid;
991 data.mv_data = txn->mt_free_pgs;
992 /* The free list can still grow during this call,
993 * despite the pre-emptive touches above. So check
994 * and make sure the entire thing got written.
997 i = txn->mt_free_pgs[0];
998 data.mv_size = MDB_IDL_SIZEOF(txn->mt_free_pgs);
999 rc = mdb_cursor_put(&mc, &key, &data, 0);
1004 } while (i != txn->mt_free_pgs[0]);
1006 /* should only be one record now */
1007 if (env->me_pghead) {
1011 mop = env->me_pghead;
1012 key.mv_size = sizeof(pgno_t);
1013 key.mv_data = (char *)&mop->mo_txnid;
1014 data.mv_size = MDB_IDL_SIZEOF(mop->mo_pages);
1015 data.mv_data = mop->mo_pages;
1016 mdb_cursor_put(&mc, &key, &data, 0);
1017 free(env->me_pghead);
1018 env->me_pghead = NULL;
1021 /* Update DB root pointers. Their pages have already been
1022 * touched so this is all in-place and cannot fail.
1026 data.mv_size = sizeof(MDB_db);
1028 mc.mc_dbi = MAIN_DBI;
1030 for (i = 2; i < txn->mt_numdbs; i++) {
1031 if (txn->mt_dbxs[i].md_dirty) {
1032 data.mv_data = &txn->mt_dbs[i];
1033 mdb_cursor_put(&mc, &txn->mt_dbxs[i].md_name, &data, 0);
1038 /* Commit up to MDB_COMMIT_PAGES dirty pages to disk until done.
1044 /* Windows actually supports scatter/gather I/O, but only on
1045 * unbuffered file handles. Since we're relying on the OS page
1046 * cache for all our data, that's self-defeating. So we just
1047 * write pages one at a time. We use the ov structure to set
1048 * the write offset, to at least save the overhead of a Seek
1052 memset(&ov, 0, sizeof(ov));
1053 for (; i<=txn->mt_u.dirty_list[0].mid; i++) {
1055 dp = txn->mt_u.dirty_list[i].mptr;
1056 DPRINTF("committing page %lu", dp->mp_pgno);
1057 size = dp->mp_pgno * env->me_psize;
1058 ov.Offset = size & 0xffffffff;
1059 ov.OffsetHigh = size >> 16;
1060 ov.OffsetHigh >>= 16;
1061 /* clear dirty flag */
1062 dp->mp_flags &= ~P_DIRTY;
1063 wsize = env->me_psize;
1064 if (IS_OVERFLOW(dp)) wsize *= dp->mp_pages;
1065 rc = WriteFile(env->me_fd, dp, wsize, NULL, &ov);
1068 DPRINTF("WriteFile: %d", n);
1075 struct iovec iov[MDB_COMMIT_PAGES];
1079 for (; i<=txn->mt_u.dirty_list[0].mid; i++) {
1080 dp = txn->mt_u.dirty_list[i].mptr;
1081 if (dp->mp_pgno != next) {
1083 DPRINTF("committing %u dirty pages", n);
1084 rc = writev(env->me_fd, iov, n);
1088 DPUTS("short write, filesystem full?");
1090 DPRINTF("writev: %s", strerror(n));
1097 lseek(env->me_fd, dp->mp_pgno * env->me_psize, SEEK_SET);
1100 DPRINTF("committing page %lu", dp->mp_pgno);
1101 iov[n].iov_len = env->me_psize;
1102 if (IS_OVERFLOW(dp)) iov[n].iov_len *= dp->mp_pages;
1103 iov[n].iov_base = dp;
1104 size += iov[n].iov_len;
1105 next = dp->mp_pgno + (IS_OVERFLOW(dp) ? dp->mp_pages : 1);
1106 /* clear dirty flag */
1107 dp->mp_flags &= ~P_DIRTY;
1108 if (++n >= MDB_COMMIT_PAGES) {
1118 DPRINTF("committing %u dirty pages", n);
1119 rc = writev(env->me_fd, iov, n);
1123 DPUTS("short write, filesystem full?");
1125 DPRINTF("writev: %s", strerror(n));
1132 /* Drop the dirty pages.
1134 for (i=1; i<=txn->mt_u.dirty_list[0].mid; i++) {
1135 dp = txn->mt_u.dirty_list[i].mptr;
1136 if (!IS_OVERFLOW(dp) || dp->mp_pages == 1) {
1137 dp->mp_next = txn->mt_env->me_dpages;
1138 txn->mt_env->me_dpages = dp;
1142 txn->mt_u.dirty_list[i].mid = 0;
1144 txn->mt_u.dirty_list[0].mid = 0;
1146 if ((n = mdb_env_sync(env, 0)) != 0 ||
1147 (n = mdb_env_write_meta(txn)) != MDB_SUCCESS) {
1154 /* update the DB tables */
1156 int toggle = !env->me_db_toggle;
1159 ip = &env->me_dbs[toggle][2];
1160 jp = &txn->mt_dbs[2];
1161 LAZY_RWLOCK_WRLOCK(&env->me_dblock);
1162 for (i = 2; i < txn->mt_numdbs; i++) {
1163 if (ip->md_root != jp->md_root)
1168 for (i = 2; i < txn->mt_numdbs; i++) {
1169 if (txn->mt_dbxs[i].md_dirty)
1170 txn->mt_dbxs[i].md_dirty = 0;
1172 env->me_db_toggle = toggle;
1173 env->me_numdbs = txn->mt_numdbs;
1174 LAZY_RWLOCK_UNLOCK(&env->me_dblock);
1177 UNLOCK_MUTEX_W(env);
1184 mdb_env_read_header(MDB_env *env, MDB_meta *meta)
1186 char page[PAGESIZE];
1191 /* We don't know the page size yet, so use a minimum value.
1195 if (!ReadFile(env->me_fd, page, PAGESIZE, (DWORD *)&rc, NULL) || rc == 0)
1197 if ((rc = read(env->me_fd, page, PAGESIZE)) == 0)
1202 else if (rc != PAGESIZE) {
1206 DPRINTF("read: %s", strerror(err));
1210 p = (MDB_page *)page;
1212 if (!F_ISSET(p->mp_flags, P_META)) {
1213 DPRINTF("page %lu not a meta page", p->mp_pgno);
1218 if (m->mm_magic != MDB_MAGIC) {
1219 DPUTS("meta has invalid magic");
1223 if (m->mm_version != MDB_VERSION) {
1224 DPRINTF("database is version %u, expected version %u",
1225 m->mm_version, MDB_VERSION);
1226 return MDB_VERSION_MISMATCH;
1229 memcpy(meta, m, sizeof(*m));
1234 mdb_env_init_meta(MDB_env *env, MDB_meta *meta)
1241 DPUTS("writing new meta page");
1245 meta->mm_magic = MDB_MAGIC;
1246 meta->mm_version = MDB_VERSION;
1247 meta->mm_psize = psize;
1248 meta->mm_last_pg = 1;
1249 meta->mm_flags = env->me_flags & 0xffff;
1250 meta->mm_flags |= MDB_INTEGERKEY;
1251 meta->mm_dbs[0].md_root = P_INVALID;
1252 meta->mm_dbs[1].md_root = P_INVALID;
1254 p = calloc(2, psize);
1256 p->mp_flags = P_META;
1259 memcpy(m, meta, sizeof(*meta));
1261 q = (MDB_page *)((char *)p + psize);
1264 q->mp_flags = P_META;
1267 memcpy(m, meta, sizeof(*meta));
1272 rc = WriteFile(env->me_fd, p, psize * 2, &len, NULL);
1273 rc = (len == psize * 2) ? MDB_SUCCESS : ErrCode();
1276 rc = write(env->me_fd, p, psize * 2);
1277 rc = (rc == (int)psize * 2) ? MDB_SUCCESS : ErrCode();
1284 mdb_env_write_meta(MDB_txn *txn)
1287 MDB_meta meta, metab;
1289 int rc, len, toggle;
1295 assert(txn != NULL);
1296 assert(txn->mt_env != NULL);
1298 toggle = !txn->mt_toggle;
1299 DPRINTF("writing meta page %d for root page %lu",
1300 toggle, txn->mt_dbs[MAIN_DBI].md_root);
1304 metab.mm_txnid = env->me_metas[toggle]->mm_txnid;
1305 metab.mm_last_pg = env->me_metas[toggle]->mm_last_pg;
1307 ptr = (char *)&meta;
1308 off = offsetof(MDB_meta, mm_dbs[0].md_depth);
1309 len = sizeof(MDB_meta) - off;
1312 meta.mm_dbs[0] = txn->mt_dbs[0];
1313 meta.mm_dbs[1] = txn->mt_dbs[1];
1314 meta.mm_last_pg = txn->mt_next_pgno - 1;
1315 meta.mm_txnid = txn->mt_txnid;
1318 off += env->me_psize;
1321 /* Write to the SYNC fd */
1324 memset(&ov, 0, sizeof(ov));
1326 WriteFile(env->me_mfd, ptr, len, (DWORD *)&rc, &ov);
1329 rc = pwrite(env->me_mfd, ptr, len, off);
1334 DPUTS("write failed, disk error?");
1335 /* On a failure, the pagecache still contains the new data.
1336 * Write some old data back, to prevent it from being used.
1337 * Use the non-SYNC fd; we know it will fail anyway.
1339 meta.mm_last_pg = metab.mm_last_pg;
1340 meta.mm_txnid = metab.mm_txnid;
1342 WriteFile(env->me_fd, ptr, len, NULL, &ov);
1344 r2 = pwrite(env->me_fd, ptr, len, off);
1346 env->me_flags |= MDB_FATAL_ERROR;
1349 /* Memory ordering issues are irrelevant; since the entire writer
1350 * is wrapped by wmutex, all of these changes will become visible
1351 * after the wmutex is unlocked. Since the DB is multi-version,
1352 * readers will get consistent data regardless of how fresh or
1353 * how stale their view of these values is.
1355 LAZY_MUTEX_LOCK(&env->me_txns->mti_mutex);
1356 txn->mt_env->me_txns->mti_me_toggle = toggle;
1357 txn->mt_env->me_txns->mti_txnid = txn->mt_txnid;
1358 LAZY_MUTEX_UNLOCK(&env->me_txns->mti_mutex);
1364 mdb_env_read_meta(MDB_env *env, int *which)
1368 assert(env != NULL);
1370 if (env->me_metas[0]->mm_txnid < env->me_metas[1]->mm_txnid)
1373 DPRINTF("Using meta page %d", toggle);
1380 mdb_env_create(MDB_env **env)
1384 e = calloc(1, sizeof(MDB_env));
1385 if (!e) return ENOMEM;
1387 e->me_maxreaders = DEFAULT_READERS;
1389 e->me_fd = INVALID_HANDLE_VALUE;
1390 e->me_lfd = INVALID_HANDLE_VALUE;
1391 e->me_mfd = INVALID_HANDLE_VALUE;
1397 mdb_env_set_mapsize(MDB_env *env, size_t size)
1401 env->me_mapsize = size;
1406 mdb_env_set_maxdbs(MDB_env *env, int dbs)
1410 env->me_maxdbs = dbs;
1415 mdb_env_set_maxreaders(MDB_env *env, int readers)
1419 env->me_maxreaders = readers;
1424 mdb_env_get_maxreaders(MDB_env *env, int *readers)
1426 if (!env || !readers)
1428 *readers = env->me_maxreaders;
1433 mdb_env_open2(MDB_env *env, unsigned int flags)
1435 int i, newenv = 0, toggle;
1439 env->me_flags = flags;
1441 memset(&meta, 0, sizeof(meta));
1443 if ((i = mdb_env_read_header(env, &meta)) != 0) {
1446 DPUTS("new mdbenv");
1450 if (!env->me_mapsize) {
1451 env->me_mapsize = newenv ? DEFAULT_MAPSIZE : meta.mm_mapsize;
1457 LONG sizelo, sizehi;
1458 sizelo = env->me_mapsize & 0xffffffff;
1459 sizehi = env->me_mapsize >> 16; /* pointless on WIN32, only needed on W64 */
1461 /* Windows won't create mappings for zero length files.
1462 * Just allocate the maxsize right now.
1465 SetFilePointer(env->me_fd, sizelo, sizehi ? &sizehi : NULL, 0);
1466 if (!SetEndOfFile(env->me_fd))
1468 SetFilePointer(env->me_fd, 0, NULL, 0);
1470 mh = CreateFileMapping(env->me_fd, NULL, PAGE_READONLY,
1471 sizehi, sizelo, NULL);
1474 env->me_map = MapViewOfFileEx(mh, FILE_MAP_READ, 0, 0, env->me_mapsize,
1482 if (meta.mm_address && (flags & MDB_FIXEDMAP))
1484 env->me_map = mmap(meta.mm_address, env->me_mapsize, PROT_READ, i,
1486 if (env->me_map == MAP_FAILED)
1491 meta.mm_mapsize = env->me_mapsize;
1492 if (flags & MDB_FIXEDMAP)
1493 meta.mm_address = env->me_map;
1494 i = mdb_env_init_meta(env, &meta);
1495 if (i != MDB_SUCCESS) {
1496 munmap(env->me_map, env->me_mapsize);
1500 env->me_psize = meta.mm_psize;
1502 env->me_maxpg = env->me_mapsize / env->me_psize;
1504 p = (MDB_page *)env->me_map;
1505 env->me_metas[0] = METADATA(p);
1506 env->me_metas[1] = (MDB_meta *)((char *)env->me_metas[0] + meta.mm_psize);
1508 if ((i = mdb_env_read_meta(env, &toggle)) != 0)
1511 DPRINTF("opened database version %u, pagesize %u",
1512 env->me_metas[toggle]->mm_version, env->me_psize);
1513 DPRINTF("depth: %u", env->me_metas[toggle]->mm_dbs[MAIN_DBI].md_depth);
1514 DPRINTF("entries: %lu", env->me_metas[toggle]->mm_dbs[MAIN_DBI].md_entries);
1515 DPRINTF("branch pages: %lu", env->me_metas[toggle]->mm_dbs[MAIN_DBI].md_branch_pages);
1516 DPRINTF("leaf pages: %lu", env->me_metas[toggle]->mm_dbs[MAIN_DBI].md_leaf_pages);
1517 DPRINTF("overflow pages: %lu", env->me_metas[toggle]->mm_dbs[MAIN_DBI].md_overflow_pages);
1518 DPRINTF("root: %lu", env->me_metas[toggle]->mm_dbs[MAIN_DBI].md_root);
1524 /* Windows doesn't support destructor callbacks for thread-specific storage */
1526 mdb_env_reader_dest(void *ptr)
1528 MDB_reader *reader = ptr;
1530 reader->mr_txnid = 0;
1536 /* downgrade the exclusive lock on the region back to shared */
1538 mdb_env_share_locks(MDB_env *env)
1542 if (env->me_metas[0]->mm_txnid < env->me_metas[1]->mm_txnid)
1544 env->me_txns->mti_me_toggle = toggle;
1545 env->me_txns->mti_txnid = env->me_metas[toggle]->mm_txnid;
1550 /* First acquire a shared lock. The Unlock will
1551 * then release the existing exclusive lock.
1553 memset(&ov, 0, sizeof(ov));
1554 LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov);
1555 UnlockFile(env->me_lfd, 0, 0, 1, 0);
1559 struct flock lock_info;
1560 /* The shared lock replaces the existing lock */
1561 memset((void *)&lock_info, 0, sizeof(lock_info));
1562 lock_info.l_type = F_RDLCK;
1563 lock_info.l_whence = SEEK_SET;
1564 lock_info.l_start = 0;
1565 lock_info.l_len = 1;
1566 fcntl(env->me_lfd, F_SETLK, &lock_info);
1572 mdb_env_setup_locks(MDB_env *env, char *lpath, int mode, int *excl)
1580 if ((env->me_lfd = CreateFile(lpath, GENERIC_READ|GENERIC_WRITE,
1581 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, OPEN_ALWAYS,
1582 FILE_ATTRIBUTE_NORMAL, NULL)) == INVALID_HANDLE_VALUE) {
1586 /* Try to get exclusive lock. If we succeed, then
1587 * nobody is using the lock region and we should initialize it.
1590 if (LockFile(env->me_lfd, 0, 0, 1, 0)) {
1594 memset(&ov, 0, sizeof(ov));
1595 if (!LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
1601 size = GetFileSize(env->me_lfd, NULL);
1603 if ((env->me_lfd = open(lpath, O_RDWR|O_CREAT, mode)) == -1) {
1607 /* Try to get exclusive lock. If we succeed, then
1608 * nobody is using the lock region and we should initialize it.
1611 struct flock lock_info;
1612 memset((void *)&lock_info, 0, sizeof(lock_info));
1613 lock_info.l_type = F_WRLCK;
1614 lock_info.l_whence = SEEK_SET;
1615 lock_info.l_start = 0;
1616 lock_info.l_len = 1;
1617 rc = fcntl(env->me_lfd, F_SETLK, &lock_info);
1621 lock_info.l_type = F_RDLCK;
1622 rc = fcntl(env->me_lfd, F_SETLKW, &lock_info);
1629 size = lseek(env->me_lfd, 0, SEEK_END);
1631 rsize = (env->me_maxreaders-1) * sizeof(MDB_reader) + sizeof(MDB_txninfo);
1632 if (size < rsize && *excl) {
1634 SetFilePointer(env->me_lfd, rsize, NULL, 0);
1635 if (!SetEndOfFile(env->me_lfd)) {
1640 if (ftruncate(env->me_lfd, rsize) != 0) {
1647 size = rsize - sizeof(MDB_txninfo);
1648 env->me_maxreaders = size/sizeof(MDB_reader) + 1;
1653 mh = CreateFileMapping(env->me_lfd, NULL, PAGE_READWRITE,
1659 env->me_txns = MapViewOfFileEx(mh, FILE_MAP_WRITE, 0, 0, rsize, NULL);
1661 if (!env->me_txns) {
1667 env->me_txns = mmap(0, rsize, PROT_READ|PROT_WRITE, MAP_SHARED,
1669 if (env->me_txns == MAP_FAILED) {
1677 if (!mdb_sec_inited) {
1678 InitializeSecurityDescriptor(&mdb_null_sd,
1679 SECURITY_DESCRIPTOR_REVISION);
1680 SetSecurityDescriptorDacl(&mdb_null_sd, TRUE, 0, FALSE);
1681 mdb_all_sa.nLength = sizeof(SECURITY_ATTRIBUTES);
1682 mdb_all_sa.bInheritHandle = FALSE;
1683 mdb_all_sa.lpSecurityDescriptor = &mdb_null_sd;
1686 /* FIXME: only using up to 20 characters of the env path here,
1687 * probably not enough to assure uniqueness...
1689 sprintf(env->me_txns->mti_rmname, "Global\\MDBr%.20s", lpath);
1690 ptr = env->me_txns->mti_rmname + sizeof("Global\\MDBr");
1691 while ((ptr = strchr(ptr, '\\')))
1693 env->me_rmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_rmname);
1694 if (!env->me_rmutex) {
1698 sprintf(env->me_txns->mti_rmname, "Global\\MDBw%.20s", lpath);
1699 ptr = env->me_txns->mti_rmname + sizeof("Global\\MDBw");
1700 while ((ptr = strchr(ptr, '\\')))
1702 env->me_wmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_rmname);
1703 if (!env->me_wmutex) {
1708 pthread_mutexattr_t mattr;
1710 pthread_mutexattr_init(&mattr);
1711 rc = pthread_mutexattr_setpshared(&mattr, PTHREAD_PROCESS_SHARED);
1715 pthread_mutex_init(&env->me_txns->mti_mutex, &mattr);
1716 pthread_mutex_init(&env->me_txns->mti_wmutex, &mattr);
1718 env->me_txns->mti_version = MDB_VERSION;
1719 env->me_txns->mti_magic = MDB_MAGIC;
1720 env->me_txns->mti_txnid = 0;
1721 env->me_txns->mti_numreaders = 0;
1722 env->me_txns->mti_me_toggle = 0;
1725 if (env->me_txns->mti_magic != MDB_MAGIC) {
1726 DPUTS("lock region has invalid magic");
1730 if (env->me_txns->mti_version != MDB_VERSION) {
1731 DPRINTF("lock region is version %u, expected version %u",
1732 env->me_txns->mti_version, MDB_VERSION);
1733 rc = MDB_VERSION_MISMATCH;
1737 if (rc != EACCES && rc != EAGAIN) {
1741 env->me_rmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_rmname);
1742 if (!env->me_rmutex) {
1746 env->me_wmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_wmname);
1747 if (!env->me_wmutex) {
1757 env->me_lfd = INVALID_HANDLE_VALUE;
1762 #define LOCKNAME "/lock.mdb"
1763 #define DATANAME "/data.mdb"
1765 mdb_env_open(MDB_env *env, const char *path, unsigned int flags, mode_t mode)
1767 int oflags, rc, len, excl;
1768 char *lpath, *dpath;
1771 lpath = malloc(len + sizeof(LOCKNAME) + len + sizeof(DATANAME));
1774 dpath = lpath + len + sizeof(LOCKNAME);
1775 sprintf(lpath, "%s" LOCKNAME, path);
1776 sprintf(dpath, "%s" DATANAME, path);
1778 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
1783 if (F_ISSET(flags, MDB_RDONLY)) {
1784 oflags = GENERIC_READ;
1785 len = OPEN_EXISTING;
1787 oflags = GENERIC_READ|GENERIC_WRITE;
1790 mode = FILE_ATTRIBUTE_NORMAL;
1791 if ((env->me_fd = CreateFile(dpath, oflags, FILE_SHARE_READ|FILE_SHARE_WRITE,
1792 NULL, len, mode, NULL)) == INVALID_HANDLE_VALUE) {
1797 if (F_ISSET(flags, MDB_RDONLY))
1800 oflags = O_RDWR | O_CREAT;
1802 if ((env->me_fd = open(dpath, oflags, mode)) == -1) {
1808 if ((rc = mdb_env_open2(env, flags)) == MDB_SUCCESS) {
1809 /* synchronous fd for meta writes */
1811 if (!(flags & (MDB_RDONLY|MDB_NOSYNC)))
1812 mode |= FILE_FLAG_WRITE_THROUGH;
1813 if ((env->me_mfd = CreateFile(dpath, oflags, FILE_SHARE_READ|FILE_SHARE_WRITE,
1814 NULL, len, mode, NULL)) == INVALID_HANDLE_VALUE) {
1819 if (!(flags & (MDB_RDONLY|MDB_NOSYNC)))
1820 oflags |= MDB_DSYNC;
1821 if ((env->me_mfd = open(dpath, oflags, mode)) == -1) {
1826 env->me_path = strdup(path);
1827 DPRINTF("opened dbenv %p", (void *) env);
1828 pthread_key_create(&env->me_txkey, mdb_env_reader_dest);
1829 LAZY_RWLOCK_INIT(&env->me_dblock, NULL);
1831 mdb_env_share_locks(env);
1832 env->me_dbxs = calloc(env->me_maxdbs, sizeof(MDB_dbx));
1833 env->me_dbs[0] = calloc(env->me_maxdbs, sizeof(MDB_db));
1834 env->me_dbs[1] = calloc(env->me_maxdbs, sizeof(MDB_db));
1840 if (env->me_fd != INVALID_HANDLE_VALUE) {
1842 env->me_fd = INVALID_HANDLE_VALUE;
1844 if (env->me_lfd != INVALID_HANDLE_VALUE) {
1846 env->me_lfd = INVALID_HANDLE_VALUE;
1854 mdb_env_close(MDB_env *env)
1861 while (env->me_dpages) {
1862 dp = env->me_dpages;
1863 env->me_dpages = dp->mp_next;
1867 free(env->me_dbs[1]);
1868 free(env->me_dbs[0]);
1872 LAZY_RWLOCK_DESTROY(&env->me_dblock);
1873 pthread_key_delete(env->me_txkey);
1876 munmap(env->me_map, env->me_mapsize);
1881 pid_t pid = getpid();
1883 for (i=0; i<env->me_txns->mti_numreaders; i++)
1884 if (env->me_txns->mti_readers[i].mr_pid == pid)
1885 env->me_txns->mti_readers[i].mr_pid = 0;
1886 munmap(env->me_txns, (env->me_maxreaders-1)*sizeof(MDB_reader)+sizeof(MDB_txninfo));
1892 /* only for aligned ints */
1894 intcmp(const MDB_val *a, const MDB_val *b)
1896 if (a->mv_size == sizeof(long))
1898 unsigned long *la, *lb;
1903 unsigned int *ia, *ib;
1910 /* ints must always be the same size */
1912 cintcmp(const MDB_val *a, const MDB_val *b)
1914 #if __BYTE_ORDER == __LITTLE_ENDIAN
1915 unsigned char *u, *c;
1918 u = a->mv_data + a->mv_size;
1919 c = b->mv_data + a->mv_size;
1922 } while(!x && u > (unsigned char *)a->mv_data);
1925 return memcmp(a->mv_data, b->mv_data, a->mv_size);
1930 memncmp(const MDB_val *a, const MDB_val *b)
1936 len_diff = a->mv_size - b->mv_size;
1939 diff = memcmp(a->mv_data, b->mv_data, len);
1940 return diff ? diff : len_diff;
1944 memnrcmp(const MDB_val *a, const MDB_val *b)
1946 const unsigned char *p1, *p2, *p1_lim;
1949 if (b->mv_size == 0)
1950 return a->mv_size != 0;
1951 if (a->mv_size == 0)
1954 p1 = (const unsigned char *)a->mv_data + a->mv_size - 1;
1955 p2 = (const unsigned char *)b->mv_data + b->mv_size - 1;
1957 len_diff = a->mv_size - b->mv_size;
1959 p1_lim = p1 - a->mv_size;
1961 p1_lim = p1 - b->mv_size;
1963 while (p1 > p1_lim) {
1973 /* Search for key within a leaf page, using binary search.
1974 * Returns the smallest entry larger or equal to the key.
1975 * If exactp is non-null, stores whether the found entry was an exact match
1976 * in *exactp (1 or 0).
1977 * If kip is non-null, stores the index of the found entry in *kip.
1978 * If no entry larger or equal to the key is found, returns NULL.
1981 mdb_search_node(MDB_cursor *mc, MDB_val *key, int *exactp)
1983 unsigned int i = 0, nkeys;
1986 MDB_page *mp = mc->mc_pg[mc->mc_top];
1987 MDB_node *node = NULL;
1992 nkeys = NUMKEYS(mp);
1994 DPRINTF("searching %u keys in %s page %lu",
1995 nkeys, IS_LEAF(mp) ? "leaf" : "branch",
2000 low = IS_LEAF(mp) ? 0 : 1;
2002 cmp = mc->mc_txn->mt_dbxs[mc->mc_dbi].md_cmp;
2004 nodekey.mv_size = mc->mc_txn->mt_dbs[mc->mc_dbi].md_pad;
2005 node = NODEPTR(mp, 0); /* fake */
2007 while (low <= high) {
2008 i = (low + high) >> 1;
2011 nodekey.mv_data = LEAF2KEY(mp, i, nodekey.mv_size);
2013 node = NODEPTR(mp, i);
2015 nodekey.mv_size = node->mn_ksize;
2016 nodekey.mv_data = NODEKEY(node);
2019 rc = cmp(key, &nodekey);
2023 DPRINTF("found leaf index %u [%s], rc = %i",
2024 i, DKEY(&nodekey), rc);
2026 DPRINTF("found branch index %u [%s -> %lu], rc = %i",
2027 i, DKEY(&nodekey), NODEPGNO(node), rc);
2038 if (rc > 0) { /* Found entry is less than the key. */
2039 i++; /* Skip to get the smallest entry larger than key. */
2041 node = NODEPTR(mp, i);
2044 *exactp = (rc == 0);
2045 /* store the key index */
2046 mc->mc_ki[mc->mc_top] = i;
2048 /* There is no entry larger or equal to the key. */
2051 /* nodeptr is fake for LEAF2 */
2056 cursor_pop_page(MDB_cursor *mc)
2061 top = mc->mc_pg[mc->mc_top];
2066 DPRINTF("popped page %lu off db %u cursor %p", top->mp_pgno,
2067 mc->mc_dbi, (void *) mc);
2072 cursor_push_page(MDB_cursor *mc, MDB_page *mp)
2074 DPRINTF("pushing page %lu on db %u cursor %p", mp->mp_pgno,
2075 mc->mc_dbi, (void *) mc);
2077 if (mc->mc_snum >= CURSOR_STACK)
2080 mc->mc_top = mc->mc_snum++;
2081 mc->mc_pg[mc->mc_top] = mp;
2082 mc->mc_ki[mc->mc_top] = 0;
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) {
2094 x = mdb_mid2l_search(txn->mt_u.dirty_list, pgno);
2095 if (x <= txn->mt_u.dirty_list[0].mid && txn->mt_u.dirty_list[x].mid == pgno) {
2096 p = txn->mt_u.dirty_list[x].mptr;
2100 if (pgno <= txn->mt_env->me_metas[txn->mt_toggle]->mm_last_pg)
2101 p = (MDB_page *)(txn->mt_env->me_map + txn->mt_env->me_psize * pgno);
2105 DPRINTF("page %lu not found", pgno);
2108 return (p != NULL) ? MDB_SUCCESS : MDB_PAGE_NOTFOUND;
2112 mdb_search_page_root(MDB_cursor *mc, MDB_val *key, int modify)
2114 MDB_page *mp = mc->mc_pg[mc->mc_top];
2119 while (IS_BRANCH(mp)) {
2122 DPRINTF("branch page %lu has %u keys", mp->mp_pgno, NUMKEYS(mp));
2123 assert(NUMKEYS(mp) > 1);
2124 DPRINTF("found index 0 to page %lu", NODEPGNO(NODEPTR(mp, 0)));
2126 if (key == NULL) /* Initialize cursor to first page. */
2127 mc->mc_ki[mc->mc_top] = 0;
2128 else if (key->mv_size > MAXKEYSIZE && key->mv_data == NULL) {
2129 /* cursor to last page */
2130 mc->mc_ki[mc->mc_top] = NUMKEYS(mp)-1;
2133 node = mdb_search_node(mc, key, &exact);
2135 mc->mc_ki[mc->mc_top] = NUMKEYS(mp) - 1;
2137 assert(mc->mc_ki[mc->mc_top] > 0);
2138 mc->mc_ki[mc->mc_top]--;
2143 DPRINTF("following index %u for key [%s]",
2144 mc->mc_ki[mc->mc_top], DKEY(key));
2145 assert(mc->mc_ki[mc->mc_top] < NUMKEYS(mp));
2146 node = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
2148 if ((rc = mdb_get_page(mc->mc_txn, NODEPGNO(node), &mp)))
2151 if ((rc = cursor_push_page(mc, mp)))
2155 if ((rc = mdb_touch(mc)) != 0)
2157 mp = mc->mc_pg[mc->mc_top];
2162 DPRINTF("internal error, index points to a %02X page!?",
2164 return MDB_CORRUPTED;
2167 DPRINTF("found leaf page %lu for key [%s]", mp->mp_pgno,
2168 key ? DKEY(key) : NULL);
2173 /* Search for the page a given key should be in.
2174 * Pushes parent pages on the cursor stack.
2175 * If key is NULL, search for the lowest page (used by mdb_cursor_first).
2176 * If modify is true, visited pages are updated with new page numbers.
2179 mdb_search_page(MDB_cursor *mc, MDB_val *key, int modify)
2184 /* Make sure the txn is still viable, then find the root from
2185 * the txn's db table.
2187 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_ERROR)) {
2188 DPUTS("transaction has failed, must abort");
2191 root = mc->mc_txn->mt_dbs[mc->mc_dbi].md_root;
2193 if (root == P_INVALID) { /* Tree is empty. */
2194 DPUTS("tree is empty");
2195 return MDB_NOTFOUND;
2198 if ((rc = mdb_get_page(mc->mc_txn, root, &mc->mc_pg[0])))
2204 DPRINTF("db %u root page %lu has flags 0x%X",
2205 mc->mc_dbi, root, mc->mc_pg[0]->mp_flags);
2208 /* For sub-databases, update main root first */
2209 if (mc->mc_dbi > MAIN_DBI && !mc->mc_txn->mt_dbxs[mc->mc_dbi].md_dirty) {
2211 mc2.mc_txn = mc->mc_txn;
2212 mc2.mc_dbi = MAIN_DBI;
2213 rc = mdb_search_page(&mc2, &mc->mc_txn->mt_dbxs[mc->mc_dbi].md_name, 1);
2216 mc->mc_txn->mt_dbxs[mc->mc_dbi].md_dirty = 1;
2218 if (!F_ISSET(mc->mc_pg[0]->mp_flags, P_DIRTY)) {
2219 if ((rc = mdb_touch(mc)))
2221 mc->mc_txn->mt_dbs[mc->mc_dbi].md_root = mc->mc_pg[0]->mp_pgno;
2225 return mdb_search_page_root(mc, key, modify);
2229 mdb_read_data(MDB_txn *txn, MDB_node *leaf, MDB_val *data)
2231 MDB_page *omp; /* overflow mpage */
2235 if (!F_ISSET(leaf->mn_flags, F_BIGDATA)) {
2236 data->mv_size = NODEDSZ(leaf);
2237 data->mv_data = NODEDATA(leaf);
2241 /* Read overflow data.
2243 data->mv_size = NODEDSZ(leaf);
2244 memcpy(&pgno, NODEDATA(leaf), sizeof(pgno));
2245 if ((rc = mdb_get_page(txn, pgno, &omp))) {
2246 DPRINTF("read overflow page %lu failed", pgno);
2249 data->mv_data = METADATA(omp);
2255 mdb_get(MDB_txn *txn, MDB_dbi dbi,
2256 MDB_val *key, MDB_val *data)
2265 DPRINTF("===> get db %u key [%s]", dbi, DKEY(key));
2267 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
2270 if (key->mv_size == 0 || key->mv_size > MAXKEYSIZE) {
2277 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
2278 mc.mc_xcursor = &mx;
2279 mdb_xcursor_init0(&mc);
2281 mc.mc_xcursor = NULL;
2283 return mdb_cursor_set(&mc, key, data, MDB_SET, &exact);
2287 mdb_sibling(MDB_cursor *mc, int move_right)
2294 if (mc->mc_snum < 2) {
2295 return MDB_NOTFOUND; /* root has no siblings */
2297 ptop = mc->mc_top-1;
2299 DPRINTF("parent page is page %lu, index %u",
2300 mc->mc_pg[ptop]->mp_pgno, mc->mc_ki[ptop]);
2302 cursor_pop_page(mc);
2303 if (move_right ? (mc->mc_ki[ptop] + 1u >= NUMKEYS(mc->mc_pg[ptop]))
2304 : (mc->mc_ki[ptop] == 0)) {
2305 DPRINTF("no more keys left, moving to %s sibling",
2306 move_right ? "right" : "left");
2307 if ((rc = mdb_sibling(mc, move_right)) != MDB_SUCCESS)
2314 DPRINTF("just moving to %s index key %u",
2315 move_right ? "right" : "left", mc->mc_ki[ptop]);
2317 assert(IS_BRANCH(mc->mc_pg[ptop]));
2319 indx = NODEPTR(mc->mc_pg[ptop], mc->mc_ki[ptop]);
2320 if ((rc = mdb_get_page(mc->mc_txn, NODEPGNO(indx), &mp)))
2323 cursor_push_page(mc, mp);
2329 mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
2335 if (mc->mc_flags & C_EOF) {
2336 return MDB_NOTFOUND;
2339 assert(mc->mc_flags & C_INITIALIZED);
2341 mp = mc->mc_pg[mc->mc_top];
2343 if (mc->mc_txn->mt_dbs[mc->mc_dbi].md_flags & MDB_DUPSORT) {
2344 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
2345 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
2346 if (op == MDB_NEXT || op == MDB_NEXT_DUP) {
2347 rc = mdb_cursor_next(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_NEXT);
2348 if (op != MDB_NEXT || rc == MDB_SUCCESS)
2352 mc->mc_xcursor->mx_cursor.mc_flags = 0;
2353 if (op == MDB_NEXT_DUP)
2354 return MDB_NOTFOUND;
2358 DPRINTF("cursor_next: top page is %lu in cursor %p", mp->mp_pgno, (void *) mc);
2360 if (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mp)) {
2361 DPUTS("=====> move to next sibling page");
2362 if (mdb_sibling(mc, 1) != MDB_SUCCESS) {
2363 mc->mc_flags |= C_EOF;
2364 return MDB_NOTFOUND;
2366 mp = mc->mc_pg[mc->mc_top];
2367 DPRINTF("next page is %lu, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]);
2369 mc->mc_ki[mc->mc_top]++;
2371 DPRINTF("==> cursor points to page %lu with %u keys, key index %u",
2372 mp->mp_pgno, NUMKEYS(mp), mc->mc_ki[mc->mc_top]);
2375 key->mv_size = mc->mc_txn->mt_dbs[mc->mc_dbi].md_pad;
2376 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
2380 assert(IS_LEAF(mp));
2381 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
2383 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
2384 mdb_xcursor_init1(mc, leaf);
2387 if ((rc = mdb_read_data(mc->mc_txn, leaf, data) != MDB_SUCCESS))
2390 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
2391 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
2392 if (rc != MDB_SUCCESS)
2397 MDB_SET_KEY(leaf, key);
2402 mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
2408 assert(mc->mc_flags & C_INITIALIZED);
2410 mp = mc->mc_pg[mc->mc_top];
2412 if (mc->mc_txn->mt_dbs[mc->mc_dbi].md_flags & MDB_DUPSORT) {
2413 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
2414 if (op == MDB_PREV || op == MDB_PREV_DUP) {
2415 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
2416 rc = mdb_cursor_prev(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_PREV);
2417 if (op != MDB_PREV || rc == MDB_SUCCESS)
2420 mc->mc_xcursor->mx_cursor.mc_flags = 0;
2421 if (op == MDB_PREV_DUP)
2422 return MDB_NOTFOUND;
2427 DPRINTF("cursor_prev: top page is %lu in cursor %p", mp->mp_pgno, (void *) mc);
2429 if (mc->mc_ki[mc->mc_top] == 0) {
2430 DPUTS("=====> move to prev sibling page");
2431 if (mdb_sibling(mc, 0) != MDB_SUCCESS) {
2432 mc->mc_flags &= ~C_INITIALIZED;
2433 return MDB_NOTFOUND;
2435 mp = mc->mc_pg[mc->mc_top];
2436 mc->mc_ki[mc->mc_top] = NUMKEYS(mp) - 1;
2437 DPRINTF("prev page is %lu, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]);
2439 mc->mc_ki[mc->mc_top]--;
2441 mc->mc_flags &= ~C_EOF;
2443 DPRINTF("==> cursor points to page %lu with %u keys, key index %u",
2444 mp->mp_pgno, NUMKEYS(mp), mc->mc_ki[mc->mc_top]);
2447 key->mv_size = mc->mc_txn->mt_dbs[mc->mc_dbi].md_pad;
2448 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
2452 assert(IS_LEAF(mp));
2453 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
2455 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
2456 mdb_xcursor_init1(mc, leaf);
2459 if ((rc = mdb_read_data(mc->mc_txn, leaf, data) != MDB_SUCCESS))
2462 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
2463 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
2464 if (rc != MDB_SUCCESS)
2469 MDB_SET_KEY(leaf, key);
2474 mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data,
2475 MDB_cursor_op op, int *exactp)
2483 assert(key->mv_size > 0);
2485 /* See if we're already on the right page */
2486 if (mc->mc_flags & C_INITIALIZED) {
2489 if (mc->mc_pg[mc->mc_top]->mp_flags & P_LEAF2) {
2490 nodekey.mv_size = mc->mc_txn->mt_dbs[mc->mc_dbi].md_pad;
2491 nodekey.mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], 0, nodekey.mv_size);
2493 leaf = NODEPTR(mc->mc_pg[mc->mc_top], 0);
2494 MDB_SET_KEY(leaf, &nodekey);
2496 rc = mc->mc_txn->mt_dbxs[mc->mc_dbi].md_cmp(key, &nodekey);
2498 /* Probably happens rarely, but first node on the page
2499 * was the one we wanted.
2501 mc->mc_ki[mc->mc_top] = 0;
2506 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
2511 if (NUMKEYS(mc->mc_pg[mc->mc_top]) > 1) {
2512 if (mc->mc_pg[mc->mc_top]->mp_flags & P_LEAF2) {
2513 nodekey.mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top],
2514 NUMKEYS(mc->mc_pg[mc->mc_top])-1, nodekey.mv_size);
2516 leaf = NODEPTR(mc->mc_pg[mc->mc_top], NUMKEYS(mc->mc_pg[mc->mc_top])-1);
2517 MDB_SET_KEY(leaf, &nodekey);
2519 rc = mc->mc_txn->mt_dbxs[mc->mc_dbi].md_cmp(key, &nodekey);
2521 /* last node was the one we wanted */
2522 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top])-1;
2526 /* This is definitely the right page, skip search_page */
2531 /* If any parents have right-sibs, search.
2532 * Otherwise, there's nothing further.
2534 for (i=0; i<mc->mc_top; i++)
2536 NUMKEYS(mc->mc_pg[i])-1)
2538 if (i == mc->mc_top) {
2539 /* There are no other pages */
2540 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]);
2541 return MDB_NOTFOUND;
2546 rc = mdb_search_page(mc, key, 0);
2547 if (rc != MDB_SUCCESS)
2550 assert(IS_LEAF(mc->mc_pg[mc->mc_top]));
2553 leaf = mdb_search_node(mc, key, exactp);
2554 if (exactp != NULL && !*exactp) {
2555 /* MDB_SET specified and not an exact match. */
2556 return MDB_NOTFOUND;
2560 DPUTS("===> inexact leaf not found, goto sibling");
2561 if ((rc = mdb_sibling(mc, 1)) != MDB_SUCCESS)
2562 return rc; /* no entries matched */
2563 mc->mc_ki[mc->mc_top] = 0;
2564 assert(IS_LEAF(mc->mc_pg[mc->mc_top]));
2565 leaf = NODEPTR(mc->mc_pg[mc->mc_top], 0);
2569 mc->mc_flags |= C_INITIALIZED;
2570 mc->mc_flags &= ~C_EOF;
2572 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
2573 key->mv_size = mc->mc_txn->mt_dbs[mc->mc_dbi].md_pad;
2574 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], key->mv_size);
2578 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
2579 mdb_xcursor_init1(mc, leaf);
2582 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
2583 if (op == MDB_SET || op == MDB_SET_RANGE) {
2584 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
2587 if (op == MDB_GET_BOTH) {
2593 rc = mdb_cursor_set(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_SET_RANGE, ex2p);
2594 if (rc != MDB_SUCCESS)
2597 } else if (op == MDB_GET_BOTH || op == MDB_GET_BOTH_RANGE) {
2599 if ((rc = mdb_read_data(mc->mc_txn, leaf, &d2)) != MDB_SUCCESS)
2601 rc = mc->mc_txn->mt_dbxs[mc->mc_dbi].md_dcmp(data, &d2);
2603 if (op == MDB_GET_BOTH || rc > 0)
2604 return MDB_NOTFOUND;
2608 if ((rc = mdb_read_data(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
2613 /* The key already matches in all other cases */
2614 if (op == MDB_SET_RANGE)
2615 MDB_SET_KEY(leaf, key);
2616 DPRINTF("==> cursor placed on key [%s]", DKEY(key));
2622 mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data)
2627 rc = mdb_search_page(mc, NULL, 0);
2628 if (rc != MDB_SUCCESS)
2630 assert(IS_LEAF(mc->mc_pg[mc->mc_top]));
2632 leaf = NODEPTR(mc->mc_pg[mc->mc_top], 0);
2633 mc->mc_flags |= C_INITIALIZED;
2634 mc->mc_flags &= ~C_EOF;
2636 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
2637 key->mv_size = mc->mc_txn->mt_dbs[mc->mc_dbi].md_pad;
2638 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], 0, key->mv_size);
2643 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
2644 mdb_xcursor_init1(mc, leaf);
2645 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
2650 mc->mc_xcursor->mx_cursor.mc_flags = 0;
2651 if ((rc = mdb_read_data(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
2655 MDB_SET_KEY(leaf, key);
2660 mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data)
2666 lkey.mv_size = MAXKEYSIZE+1;
2667 lkey.mv_data = NULL;
2669 rc = mdb_search_page(mc, &lkey, 0);
2670 if (rc != MDB_SUCCESS)
2672 assert(IS_LEAF(mc->mc_pg[mc->mc_top]));
2674 leaf = NODEPTR(mc->mc_pg[mc->mc_top], NUMKEYS(mc->mc_pg[mc->mc_top])-1);
2675 mc->mc_flags |= C_INITIALIZED;
2676 mc->mc_flags &= ~C_EOF;
2678 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]) - 1;
2680 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
2681 key->mv_size = mc->mc_txn->mt_dbs[mc->mc_dbi].md_pad;
2682 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], key->mv_size);
2687 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
2688 mdb_xcursor_init1(mc, leaf);
2689 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
2693 if ((rc = mdb_read_data(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
2698 MDB_SET_KEY(leaf, key);
2703 mdb_cursor_get(MDB_cursor *mc, MDB_val *key, MDB_val *data,
2713 case MDB_GET_BOTH_RANGE:
2714 if (data == NULL || mc->mc_xcursor == NULL) {
2721 if (key == NULL || key->mv_size == 0 || key->mv_size > MAXKEYSIZE) {
2723 } else if (op == MDB_SET_RANGE)
2724 rc = mdb_cursor_set(mc, key, data, op, NULL);
2726 rc = mdb_cursor_set(mc, key, data, op, &exact);
2728 case MDB_GET_MULTIPLE:
2730 !(mc->mc_txn->mt_dbs[mc->mc_dbi].md_flags & MDB_DUPFIXED) ||
2731 !(mc->mc_flags & C_INITIALIZED)) {
2736 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) ||
2737 (mc->mc_xcursor->mx_cursor.mc_flags & C_EOF))
2740 case MDB_NEXT_MULTIPLE:
2742 !(mc->mc_txn->mt_dbs[mc->mc_dbi].md_flags & MDB_DUPFIXED)) {
2746 if (!(mc->mc_flags & C_INITIALIZED))
2747 rc = mdb_cursor_first(mc, key, data);
2749 rc = mdb_cursor_next(mc, key, data, MDB_NEXT_DUP);
2750 if (rc == MDB_SUCCESS) {
2751 if (mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) {
2754 mx = &mc->mc_xcursor->mx_cursor;
2755 data->mv_size = NUMKEYS(mx->mc_pg[mx->mc_top]) *
2756 mx->mc_txn->mt_dbs[mx->mc_dbi].md_pad;
2757 data->mv_data = METADATA(mx->mc_pg[mx->mc_top]);
2758 mx->mc_ki[mx->mc_top] = NUMKEYS(mx->mc_pg[mx->mc_top])-1;
2766 case MDB_NEXT_NODUP:
2767 if (!(mc->mc_flags & C_INITIALIZED))
2768 rc = mdb_cursor_first(mc, key, data);
2770 rc = mdb_cursor_next(mc, key, data, op);
2774 case MDB_PREV_NODUP:
2775 if (!(mc->mc_flags & C_INITIALIZED) || (mc->mc_flags & C_EOF))
2776 rc = mdb_cursor_last(mc, key, data);
2778 rc = mdb_cursor_prev(mc, key, data, op);
2781 rc = mdb_cursor_first(mc, key, data);
2785 !(mc->mc_txn->mt_dbs[mc->mc_dbi].md_flags & MDB_DUPSORT) ||
2786 !(mc->mc_flags & C_INITIALIZED) ||
2787 !(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
2791 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
2794 rc = mdb_cursor_last(mc, key, data);
2798 !(mc->mc_txn->mt_dbs[mc->mc_dbi].md_flags & MDB_DUPSORT) ||
2799 !(mc->mc_flags & C_INITIALIZED) ||
2800 !(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
2804 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
2807 DPRINTF("unhandled/unimplemented cursor operation %u", op);
2816 mdb_cursor_touch(MDB_cursor *mc)
2820 if (mc->mc_dbi > MAIN_DBI && !mc->mc_txn->mt_dbxs[mc->mc_dbi].md_dirty) {
2822 mc2.mc_txn = mc->mc_txn;
2823 mc2.mc_dbi = MAIN_DBI;
2824 rc = mdb_search_page(&mc2, &mc->mc_txn->mt_dbxs[mc->mc_dbi].md_name, 1);
2826 mc->mc_txn->mt_dbxs[mc->mc_dbi].md_dirty = 1;
2828 for (mc->mc_top = 0; mc->mc_top < mc->mc_snum; mc->mc_top++) {
2829 if (!F_ISSET(mc->mc_pg[mc->mc_top]->mp_flags, P_DIRTY)) {
2833 mc->mc_txn->mt_dbs[mc->mc_dbi].md_root =
2834 mc->mc_pg[mc->mc_top]->mp_pgno;
2838 mc->mc_top = mc->mc_snum-1;
2843 mdb_cursor_put(MDB_cursor *mc, MDB_val *key, MDB_val *data,
2847 MDB_val xdata, *rdata, dkey;
2849 char dbuf[PAGESIZE];
2855 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_RDONLY))
2858 DPRINTF("==> put db %u key [%s], size %zu, data size %zu",
2859 mc->mc_dbi, DKEY(key), key->mv_size, data->mv_size);
2863 if (flags == MDB_CURRENT) {
2864 if (!(mc->mc_flags & C_INITIALIZED))
2867 } else if (mc->mc_txn->mt_dbs[mc->mc_dbi].md_root == P_INVALID) {
2869 /* new database, write a root leaf page */
2870 DPUTS("allocating new root leaf page");
2871 if ((np = mdb_new_page(mc, P_LEAF, 1)) == NULL) {
2875 cursor_push_page(mc, np);
2876 mc->mc_txn->mt_dbs[mc->mc_dbi].md_root = np->mp_pgno;
2877 mc->mc_txn->mt_dbs[mc->mc_dbi].md_depth++;
2878 mc->mc_txn->mt_dbxs[mc->mc_dbi].md_dirty = 1;
2879 if ((mc->mc_txn->mt_dbs[mc->mc_dbi].md_flags & (MDB_DUPSORT|MDB_DUPFIXED))
2881 np->mp_flags |= P_LEAF2;
2882 mc->mc_flags |= C_INITIALIZED;
2887 rc = mdb_cursor_set(mc, key, NULL, MDB_SET, &exact);
2888 if (flags == MDB_NOOVERWRITE && rc == 0) {
2889 DPRINTF("duplicate key [%s]", DKEY(key));
2890 return MDB_KEYEXIST;
2892 if (rc && rc != MDB_NOTFOUND)
2896 /* Cursor is positioned, now make sure all pages are writable */
2897 rc2 = mdb_cursor_touch(mc);
2898 if (rc2) return rc2;
2901 /* The key already exists */
2902 if (rc == MDB_SUCCESS) {
2903 /* there's only a key anyway, so this is a no-op */
2904 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
2905 unsigned int ksize = mc->mc_txn->mt_dbs[mc->mc_dbi].md_pad;
2906 if (key->mv_size != ksize)
2908 if (flags == MDB_CURRENT) {
2909 char *ptr = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], ksize);
2910 memcpy(ptr, key->mv_data, ksize);
2915 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
2918 if (F_ISSET(mc->mc_txn->mt_dbs[mc->mc_dbi].md_flags, MDB_DUPSORT)) {
2919 /* Was a single item before, must convert now */
2920 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
2921 dkey.mv_size = NODEDSZ(leaf);
2922 dkey.mv_data = dbuf;
2923 memcpy(dbuf, NODEDATA(leaf), dkey.mv_size);
2924 /* data matches, ignore it */
2925 if (!mdb_dcmp(mc->mc_txn, mc->mc_dbi, data, &dkey))
2926 return (flags == MDB_NODUPDATA) ? MDB_KEYEXIST : MDB_SUCCESS;
2927 memset(&dummy, 0, sizeof(dummy));
2928 if (mc->mc_txn->mt_dbs[mc->mc_dbi].md_flags & MDB_DUPFIXED) {
2929 dummy.md_pad = data->mv_size;
2930 dummy.md_flags = MDB_DUPFIXED;
2931 if (mc->mc_txn->mt_dbs[mc->mc_dbi].md_flags & MDB_INTEGERDUP)
2932 dummy.md_flags |= MDB_INTEGERKEY;
2934 dummy.md_root = P_INVALID;
2935 if (dkey.mv_size == sizeof(MDB_db)) {
2936 memcpy(NODEDATA(leaf), &dummy, sizeof(dummy));
2939 mdb_del_node(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
2942 xdata.mv_size = sizeof(MDB_db);
2943 xdata.mv_data = &dummy;
2948 /* same size, just replace it */
2949 if (!F_ISSET(leaf->mn_flags, F_BIGDATA) &&
2950 NODEDSZ(leaf) == data->mv_size) {
2951 memcpy(NODEDATA(leaf), data->mv_data, data->mv_size);
2954 mdb_del_node(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
2956 DPRINTF("inserting key at index %i", mc->mc_ki[mc->mc_top]);
2962 nsize = IS_LEAF2(mc->mc_pg[mc->mc_top]) ? key->mv_size : mdb_leaf_size(mc->mc_txn->mt_env, key, rdata);
2963 if (SIZELEFT(mc->mc_pg[mc->mc_top]) < nsize) {
2964 rc = mdb_split(mc, key, rdata, P_INVALID);
2966 /* There is room already in this leaf page. */
2967 rc = mdb_add_node(mc, mc->mc_ki[mc->mc_top], key, rdata, 0, 0);
2970 if (rc != MDB_SUCCESS)
2971 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
2973 /* Remember if we just added a subdatabase */
2974 if (flags & F_SUBDATA) {
2975 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
2976 leaf->mn_flags |= F_SUBDATA;
2979 /* Now store the actual data in the child DB. Note that we're
2980 * storing the user data in the keys field, so there are strict
2981 * size limits on dupdata. The actual data fields of the child
2982 * DB are all zero size.
2985 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
2987 if (flags == MDB_CURRENT)
2988 mdb_xcursor_init2(mc);
2990 mdb_xcursor_init1(mc, leaf);
2993 if (flags == MDB_NODUPDATA)
2994 flags = MDB_NOOVERWRITE;
2995 /* converted, write the original data first */
2997 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, &dkey, &xdata, flags);
2999 leaf->mn_flags |= F_DUPDATA;
3001 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, data, &xdata, flags);
3002 mdb_xcursor_fini(mc);
3003 memcpy(NODEDATA(leaf),
3004 &mc->mc_xcursor->mx_txn.mt_dbs[mc->mc_xcursor->mx_cursor.mc_dbi],
3007 mc->mc_txn->mt_dbs[mc->mc_dbi].md_entries++;
3014 mdb_cursor_del(MDB_cursor *mc, unsigned int flags)
3019 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_RDONLY))
3022 if (!mc->mc_flags & C_INITIALIZED)
3025 rc = mdb_cursor_touch(mc);
3028 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
3030 if (!IS_LEAF2(mc->mc_pg[mc->mc_top]) && F_ISSET(leaf->mn_flags, F_DUPDATA)) {
3031 if (flags != MDB_NODUPDATA) {
3032 mdb_xcursor_init2(mc);
3033 rc = mdb_cursor_del(&mc->mc_xcursor->mx_cursor, 0);
3034 mdb_xcursor_fini(mc);
3035 /* If sub-DB still has entries, we're done */
3036 if (mc->mc_xcursor->mx_txn.mt_dbs[mc->mc_xcursor->mx_cursor.mc_dbi].md_root
3038 memcpy(NODEDATA(leaf),
3039 &mc->mc_xcursor->mx_txn.mt_dbs[mc->mc_xcursor->mx_cursor.mc_dbi],
3041 mc->mc_txn->mt_dbs[mc->mc_dbi].md_entries--;
3044 /* otherwise fall thru and delete the sub-DB */
3047 /* add all the child DB's pages to the free list */
3048 rc = mdb_search_page(&mc->mc_xcursor->mx_cursor, NULL, 0);
3049 if (rc == MDB_SUCCESS) {
3054 mx = &mc->mc_xcursor->mx_cursor;
3055 mc->mc_txn->mt_dbs[mc->mc_dbi].md_entries -=
3056 mx->mc_txn->mt_dbs[mx->mc_dbi].md_entries;
3058 cursor_pop_page(mx);
3060 while (mx->mc_snum > 1) {
3061 for (i=0; i<NUMKEYS(mx->mc_pg[mx->mc_top]); i++) {
3063 ni = NODEPTR(mx->mc_pg[mx->mc_top], i);
3066 mdb_midl_append(mc->mc_txn->mt_free_pgs, pg);
3068 rc = mdb_sibling(mx, 1);
3073 mdb_midl_append(mc->mc_txn->mt_free_pgs,
3074 mx->mc_txn->mt_dbs[mx->mc_dbi].md_root);
3078 return mdb_del0(mc, leaf);
3081 /* Allocate a page and initialize it
3084 mdb_new_page(MDB_cursor *mc, uint32_t flags, int num)
3088 if ((np = mdb_alloc_page(mc, num)) == NULL)
3090 DPRINTF("allocated new mpage %lu, page size %u",
3091 np->mp_pgno, mc->mc_txn->mt_env->me_psize);
3092 np->mp_flags = flags | P_DIRTY;
3093 np->mp_lower = PAGEHDRSZ;
3094 np->mp_upper = mc->mc_txn->mt_env->me_psize;
3097 mc->mc_txn->mt_dbs[mc->mc_dbi].md_branch_pages++;
3098 else if (IS_LEAF(np))
3099 mc->mc_txn->mt_dbs[mc->mc_dbi].md_leaf_pages++;
3100 else if (IS_OVERFLOW(np)) {
3101 mc->mc_txn->mt_dbs[mc->mc_dbi].md_overflow_pages += num;
3109 mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data)
3113 sz = LEAFSIZE(key, data);
3114 if (data->mv_size >= env->me_psize / MDB_MINKEYS) {
3115 /* put on overflow page */
3116 sz -= data->mv_size - sizeof(pgno_t);
3120 return sz + sizeof(indx_t);
3124 mdb_branch_size(MDB_env *env, MDB_val *key)
3129 if (sz >= env->me_psize / MDB_MINKEYS) {
3130 /* put on overflow page */
3131 /* not implemented */
3132 /* sz -= key->size - sizeof(pgno_t); */
3135 return sz + sizeof(indx_t);
3139 mdb_add_node(MDB_cursor *mc, indx_t indx,
3140 MDB_val *key, MDB_val *data, pgno_t pgno, uint8_t flags)
3143 size_t node_size = NODESIZE;
3146 MDB_page *mp = mc->mc_pg[mc->mc_top];
3147 MDB_page *ofp = NULL; /* overflow page */
3150 assert(mp->mp_upper >= mp->mp_lower);
3152 DPRINTF("add to %s page %lu index %i, data size %zu key size %zu [%s]",
3153 IS_LEAF(mp) ? "leaf" : "branch",
3154 mp->mp_pgno, indx, data ? data->mv_size : 0,
3155 key ? key->mv_size : 0, key ? DKEY(key) : NULL);
3158 /* Move higher keys up one slot. */
3159 int ksize = mc->mc_txn->mt_dbs[mc->mc_dbi].md_pad, dif;
3160 char *ptr = LEAF2KEY(mp, indx, ksize);
3161 dif = NUMKEYS(mp) - indx;
3163 memmove(ptr+ksize, ptr, dif*ksize);
3164 /* insert new key */
3165 memcpy(ptr, key->mv_data, ksize);
3167 /* Just using these for counting */
3168 mp->mp_lower += sizeof(indx_t);
3169 mp->mp_upper -= ksize - sizeof(indx_t);
3174 node_size += key->mv_size;
3178 if (F_ISSET(flags, F_BIGDATA)) {
3179 /* Data already on overflow page. */
3180 node_size += sizeof(pgno_t);
3181 } else if (data->mv_size >= mc->mc_txn->mt_env->me_psize / MDB_MINKEYS) {
3182 int ovpages = OVPAGES(data->mv_size, mc->mc_txn->mt_env->me_psize);
3183 /* Put data on overflow page. */
3184 DPRINTF("data size is %zu, put on overflow page",
3186 node_size += sizeof(pgno_t);
3187 if ((ofp = mdb_new_page(mc, P_OVERFLOW, ovpages)) == NULL)
3189 DPRINTF("allocated overflow page %lu", ofp->mp_pgno);
3192 node_size += data->mv_size;
3195 node_size += node_size & 1;
3197 if (node_size + sizeof(indx_t) > SIZELEFT(mp)) {
3198 DPRINTF("not enough room in page %lu, got %u ptrs",
3199 mp->mp_pgno, NUMKEYS(mp));
3200 DPRINTF("upper - lower = %u - %u = %u", mp->mp_upper, mp->mp_lower,
3201 mp->mp_upper - mp->mp_lower);
3202 DPRINTF("node size = %zu", node_size);
3206 /* Move higher pointers up one slot. */
3207 for (i = NUMKEYS(mp); i > indx; i--)
3208 mp->mp_ptrs[i] = mp->mp_ptrs[i - 1];
3210 /* Adjust free space offsets. */
3211 ofs = mp->mp_upper - node_size;
3212 assert(ofs >= mp->mp_lower + sizeof(indx_t));
3213 mp->mp_ptrs[indx] = ofs;
3215 mp->mp_lower += sizeof(indx_t);
3217 /* Write the node data. */
3218 node = NODEPTR(mp, indx);
3219 node->mn_ksize = (key == NULL) ? 0 : key->mv_size;
3220 node->mn_flags = flags;
3222 SETDSZ(node,data->mv_size);
3227 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
3232 if (F_ISSET(flags, F_BIGDATA))
3233 memcpy(node->mn_data + key->mv_size, data->mv_data,
3236 memcpy(node->mn_data + key->mv_size, data->mv_data,
3239 memcpy(node->mn_data + key->mv_size, &ofp->mp_pgno,
3241 memcpy(METADATA(ofp), data->mv_data, data->mv_size);
3249 mdb_del_node(MDB_page *mp, indx_t indx, int ksize)
3252 indx_t i, j, numkeys, ptr;
3256 DPRINTF("delete node %u on %s page %lu", indx,
3257 IS_LEAF(mp) ? "leaf" : "branch", mp->mp_pgno);
3258 assert(indx < NUMKEYS(mp));
3261 int x = NUMKEYS(mp) - 1 - indx;
3262 base = LEAF2KEY(mp, indx, ksize);
3264 memmove(base, base + ksize, x * ksize);
3265 mp->mp_lower -= sizeof(indx_t);
3266 mp->mp_upper += ksize - sizeof(indx_t);
3270 node = NODEPTR(mp, indx);
3271 sz = NODESIZE + node->mn_ksize;
3273 if (F_ISSET(node->mn_flags, F_BIGDATA))
3274 sz += sizeof(pgno_t);
3276 sz += NODEDSZ(node);
3280 ptr = mp->mp_ptrs[indx];
3281 numkeys = NUMKEYS(mp);
3282 for (i = j = 0; i < numkeys; i++) {
3284 mp->mp_ptrs[j] = mp->mp_ptrs[i];
3285 if (mp->mp_ptrs[i] < ptr)
3286 mp->mp_ptrs[j] += sz;
3291 base = (char *)mp + mp->mp_upper;
3292 memmove(base + sz, base, ptr - mp->mp_upper);
3294 mp->mp_lower -= sizeof(indx_t);
3299 mdb_xcursor_init0(MDB_cursor *mc)
3301 MDB_xcursor *mx = mc->mc_xcursor;
3304 mx->mx_txn = *mc->mc_txn;
3305 mx->mx_txn.mt_dbxs = mx->mx_dbxs;
3306 mx->mx_txn.mt_dbs = mx->mx_dbs;
3307 mx->mx_dbxs[0] = mc->mc_txn->mt_dbxs[0];
3308 mx->mx_dbxs[1] = mc->mc_txn->mt_dbxs[1];
3309 if (mc->mc_dbi > 1) {
3310 mx->mx_dbxs[2] = mc->mc_txn->mt_dbxs[mc->mc_dbi];
3315 mx->mx_dbxs[dbn+1].md_parent = dbn;
3316 mx->mx_dbxs[dbn+1].md_cmp = mx->mx_dbxs[dbn].md_dcmp;
3317 mx->mx_dbxs[dbn+1].md_rel = mx->mx_dbxs[dbn].md_rel;
3318 mx->mx_dbxs[dbn+1].md_dirty = 0;
3319 mx->mx_txn.mt_numdbs = dbn+2;
3320 mx->mx_txn.mt_u = mc->mc_txn->mt_u;
3322 mx->mx_cursor.mc_xcursor = NULL;
3323 mx->mx_cursor.mc_txn = &mx->mx_txn;
3324 mx->mx_cursor.mc_dbi = dbn+1;
3328 mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node)
3330 MDB_db *db = NODEDATA(node);
3331 MDB_xcursor *mx = mc->mc_xcursor;
3333 mx->mx_dbs[0] = mc->mc_txn->mt_dbs[0];
3334 mx->mx_dbs[1] = mc->mc_txn->mt_dbs[1];
3335 if (mc->mc_dbi > 1) {
3336 mx->mx_dbs[2] = mc->mc_txn->mt_dbs[mc->mc_dbi];
3337 mx->mx_dbxs[2].md_dirty = mc->mc_txn->mt_dbxs[mc->mc_dbi].md_dirty;
3342 DPRINTF("Sub-db %u for db %u root page %lu", dbn, mc->mc_dbi, db->md_root);
3343 mx->mx_dbs[dbn] = *db;
3344 if (F_ISSET(mc->mc_pg[mc->mc_top]->mp_flags, P_DIRTY))
3345 mx->mx_dbxs[dbn].md_dirty = 1;
3346 mx->mx_dbxs[dbn].md_name.mv_data = NODEKEY(node);
3347 mx->mx_dbxs[dbn].md_name.mv_size = node->mn_ksize;
3348 mx->mx_txn.mt_next_pgno = mc->mc_txn->mt_next_pgno;
3349 mx->mx_cursor.mc_snum = 0;
3350 mx->mx_cursor.mc_flags = 0;
3354 mdb_xcursor_init2(MDB_cursor *mc)
3356 MDB_xcursor *mx = mc->mc_xcursor;
3358 mx->mx_dbs[0] = mc->mc_txn->mt_dbs[0];
3359 mx->mx_dbs[1] = mc->mc_txn->mt_dbs[1];
3360 if (mc->mc_dbi > 1) {
3361 mx->mx_dbs[2] = mc->mc_txn->mt_dbs[mc->mc_dbi];
3362 mx->mx_dbxs[2].md_dirty = mc->mc_txn->mt_dbxs[mc->mc_dbi].md_dirty;
3367 DPRINTF("Sub-db %u for db %u root page %lu", dbn, mc->mc_dbi,
3368 mx->mx_dbs[dbn].md_root);
3369 mx->mx_txn.mt_next_pgno = mc->mc_txn->mt_next_pgno;
3373 mdb_xcursor_fini(MDB_cursor *mc)
3375 MDB_xcursor *mx = mc->mc_xcursor;
3376 mc->mc_txn->mt_next_pgno = mx->mx_txn.mt_next_pgno;
3377 mc->mc_txn->mt_dbs[0] = mx->mx_dbs[0];
3378 mc->mc_txn->mt_dbs[1] = mx->mx_dbs[1];
3379 if (mc->mc_dbi > 1) {
3380 mc->mc_txn->mt_dbs[mc->mc_dbi] = mx->mx_dbs[2];
3381 mc->mc_txn->mt_dbxs[mc->mc_dbi].md_dirty = mx->mx_dbxs[2].md_dirty;
3386 mdb_cursor_open(MDB_txn *txn, MDB_dbi dbi, MDB_cursor **ret)
3389 size_t size = sizeof(MDB_cursor);
3391 if (txn == NULL || ret == NULL || !dbi || dbi >= txn->mt_numdbs)
3394 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT)
3395 size += sizeof(MDB_xcursor);
3397 if ((mc = calloc(1, size)) != NULL) {
3400 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
3401 MDB_xcursor *mx = (MDB_xcursor *)(mc + 1);
3402 mc->mc_xcursor = mx;
3403 mdb_xcursor_init0(mc);
3414 /* Return the count of duplicate data items for the current key */
3416 mdb_cursor_count(MDB_cursor *mc, unsigned long *countp)
3420 if (mc == NULL || countp == NULL)
3423 if (!(mc->mc_txn->mt_dbs[mc->mc_dbi].md_flags & MDB_DUPSORT))
3426 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
3427 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
3430 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED))
3433 *countp = mc->mc_xcursor->mx_txn.mt_dbs[mc->mc_xcursor->mx_cursor.mc_dbi].md_entries;
3439 mdb_cursor_close(MDB_cursor *mc)
3447 mdb_update_key(MDB_page *mp, indx_t indx, MDB_val *key)
3449 indx_t ptr, i, numkeys;
3456 node = NODEPTR(mp, indx);
3457 ptr = mp->mp_ptrs[indx];
3458 DPRINTF("update key %u (ofs %u) [%.*s] to [%s] on page %lu",
3460 (int)node->mn_ksize, (char *)NODEKEY(node),
3464 delta = key->mv_size - node->mn_ksize;
3466 if (delta > 0 && SIZELEFT(mp) < delta) {
3467 DPRINTF("OUCH! Not enough room, delta = %d", delta);
3471 numkeys = NUMKEYS(mp);
3472 for (i = 0; i < numkeys; i++) {
3473 if (mp->mp_ptrs[i] <= ptr)
3474 mp->mp_ptrs[i] -= delta;
3477 base = (char *)mp + mp->mp_upper;
3478 len = ptr - mp->mp_upper + NODESIZE;
3479 memmove(base - delta, base, len);
3480 mp->mp_upper -= delta;
3482 node = NODEPTR(mp, indx);
3483 node->mn_ksize = key->mv_size;
3486 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
3491 /* Move a node from csrc to cdst.
3494 mdb_move_node(MDB_cursor *csrc, MDB_cursor *cdst)
3501 /* Mark src and dst as dirty. */
3502 if ((rc = mdb_touch(csrc)) ||
3503 (rc = mdb_touch(cdst)))
3506 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
3507 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0); /* fake */
3508 key.mv_size = csrc->mc_txn->mt_dbs[csrc->mc_dbi].md_pad;
3509 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
3511 data.mv_data = NULL;
3513 if (csrc->mc_ki[csrc->mc_top] == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
3514 unsigned int snum = csrc->mc_snum;
3515 /* must find the lowest key below src */
3516 mdb_search_page_root(csrc, NULL, 0);
3517 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
3518 csrc->mc_snum = snum--;
3519 csrc->mc_top = snum;
3521 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top]);
3523 key.mv_size = NODEKSZ(srcnode);
3524 key.mv_data = NODEKEY(srcnode);
3525 data.mv_size = NODEDSZ(srcnode);
3526 data.mv_data = NODEDATA(srcnode);
3528 DPRINTF("moving %s node %u [%s] on page %lu to node %u on page %lu",
3529 IS_LEAF(csrc->mc_pg[csrc->mc_top]) ? "leaf" : "branch",
3530 csrc->mc_ki[csrc->mc_top],
3532 csrc->mc_pg[csrc->mc_top]->mp_pgno,
3533 cdst->mc_ki[cdst->mc_top], cdst->mc_pg[cdst->mc_top]->mp_pgno);
3535 /* Add the node to the destination page.
3537 rc = mdb_add_node(cdst, cdst->mc_ki[cdst->mc_top], &key, &data, NODEPGNO(srcnode),
3539 if (rc != MDB_SUCCESS)
3542 /* Delete the node from the source page.
3544 mdb_del_node(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
3546 /* Update the parent separators.
3548 if (csrc->mc_ki[csrc->mc_top] == 0) {
3549 if (csrc->mc_ki[csrc->mc_top-1] != 0) {
3550 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
3551 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
3553 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top]);
3554 key.mv_size = NODEKSZ(srcnode);
3555 key.mv_data = NODEKEY(srcnode);
3557 DPRINTF("update separator for source page %lu to [%s]",
3558 csrc->mc_pg[csrc->mc_top]->mp_pgno, DKEY(&key));
3559 if ((rc = mdb_update_key(csrc->mc_pg[csrc->mc_top-1], csrc->mc_ki[csrc->mc_top-1],
3560 &key)) != MDB_SUCCESS)
3563 if (IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
3565 nullkey.mv_size = 0;
3566 assert(mdb_update_key(csrc->mc_pg[csrc->mc_top], 0, &nullkey) == MDB_SUCCESS);
3570 if (cdst->mc_ki[cdst->mc_top] == 0) {
3571 if (cdst->mc_ki[cdst->mc_top-1] != 0) {
3572 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
3573 key.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, key.mv_size);
3575 srcnode = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
3576 key.mv_size = NODEKSZ(srcnode);
3577 key.mv_data = NODEKEY(srcnode);
3579 DPRINTF("update separator for destination page %lu to [%s]",
3580 cdst->mc_pg[cdst->mc_top]->mp_pgno, DKEY(&key));
3581 if ((rc = mdb_update_key(cdst->mc_pg[cdst->mc_top-1], cdst->mc_ki[cdst->mc_top-1],
3582 &key)) != MDB_SUCCESS)
3585 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top])) {
3587 nullkey.mv_size = 0;
3588 assert(mdb_update_key(cdst->mc_pg[cdst->mc_top], 0, &nullkey) == MDB_SUCCESS);
3596 mdb_merge(MDB_cursor *csrc, MDB_cursor *cdst)
3603 DPRINTF("merging page %lu into %lu", csrc->mc_pg[csrc->mc_top]->mp_pgno, cdst->mc_pg[cdst->mc_top]->mp_pgno);
3605 assert(csrc->mc_snum > 1); /* can't merge root page */
3606 assert(cdst->mc_snum > 1);
3608 /* Mark dst as dirty. */
3609 if ((rc = mdb_touch(cdst)))
3612 /* Move all nodes from src to dst.
3614 j = NUMKEYS(cdst->mc_pg[cdst->mc_top]);
3615 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
3616 key.mv_size = csrc->mc_txn->mt_dbs[csrc->mc_dbi].md_pad;
3617 key.mv_data = METADATA(csrc->mc_pg[csrc->mc_top]);
3618 for (i = 0; i < NUMKEYS(csrc->mc_pg[csrc->mc_top]); i++, j++) {
3619 rc = mdb_add_node(cdst, j, &key, NULL, 0, 0);
3620 if (rc != MDB_SUCCESS)
3622 key.mv_data = (char *)key.mv_data + key.mv_size;
3625 for (i = 0; i < NUMKEYS(csrc->mc_pg[csrc->mc_top]); i++, j++) {
3626 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], i);
3628 key.mv_size = srcnode->mn_ksize;
3629 key.mv_data = NODEKEY(srcnode);
3630 data.mv_size = NODEDSZ(srcnode);
3631 data.mv_data = NODEDATA(srcnode);
3632 rc = mdb_add_node(cdst, j, &key, &data, NODEPGNO(srcnode), srcnode->mn_flags);
3633 if (rc != MDB_SUCCESS)
3638 DPRINTF("dst page %lu now has %u keys (%.1f%% filled)",
3639 cdst->mc_pg[cdst->mc_top]->mp_pgno, NUMKEYS(cdst->mc_pg[cdst->mc_top]), (float)PAGEFILL(cdst->mc_txn->mt_env, cdst->mc_pg[cdst->mc_top]) / 10);
3641 /* Unlink the src page from parent and add to free list.
3643 mdb_del_node(csrc->mc_pg[csrc->mc_top-1], csrc->mc_ki[csrc->mc_top-1], 0);
3644 if (csrc->mc_ki[csrc->mc_top-1] == 0) {
3646 if ((rc = mdb_update_key(csrc->mc_pg[csrc->mc_top-1], 0, &key)) != MDB_SUCCESS)
3650 mdb_midl_append(csrc->mc_txn->mt_free_pgs, csrc->mc_pg[csrc->mc_top]->mp_pgno);
3651 if (IS_LEAF(csrc->mc_pg[csrc->mc_top]))
3652 csrc->mc_txn->mt_dbs[csrc->mc_dbi].md_leaf_pages--;
3654 csrc->mc_txn->mt_dbs[csrc->mc_dbi].md_branch_pages--;
3655 cursor_pop_page(csrc);
3657 return mdb_rebalance(csrc);
3661 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst)
3665 cdst->mc_txn = csrc->mc_txn;
3666 cdst->mc_dbi = csrc->mc_dbi;
3667 cdst->mc_snum = csrc->mc_snum;
3668 cdst->mc_top = csrc->mc_top;
3669 cdst->mc_flags = csrc->mc_flags;
3671 for (i=0; i<csrc->mc_snum; i++) {
3672 cdst->mc_pg[i] = csrc->mc_pg[i];
3673 cdst->mc_ki[i] = csrc->mc_ki[i];
3677 #define FILL_THRESHOLD 250
3680 mdb_rebalance(MDB_cursor *mc)
3688 DPRINTF("rebalancing %s page %lu (has %u keys, %.1f%% full)",
3689 IS_LEAF(mc->mc_pg[mc->mc_top]) ? "leaf" : "branch",
3690 mc->mc_pg[mc->mc_top]->mp_pgno, NUMKEYS(mc->mc_pg[mc->mc_top]), (float)PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) / 10);
3692 if (PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) >= FILL_THRESHOLD) {
3693 DPRINTF("no need to rebalance page %lu, above fill threshold",
3694 mc->mc_pg[mc->mc_top]->mp_pgno);
3698 if (mc->mc_snum < 2) {
3699 if (NUMKEYS(mc->mc_pg[mc->mc_top]) == 0) {
3700 DPUTS("tree is completely empty");
3701 mc->mc_txn->mt_dbs[mc->mc_dbi].md_root = P_INVALID;
3702 mc->mc_txn->mt_dbs[mc->mc_dbi].md_depth = 0;
3703 mc->mc_txn->mt_dbs[mc->mc_dbi].md_leaf_pages = 0;
3704 mdb_midl_append(mc->mc_txn->mt_free_pgs, mc->mc_pg[mc->mc_top]->mp_pgno);
3705 } else if (IS_BRANCH(mc->mc_pg[mc->mc_top]) && NUMKEYS(mc->mc_pg[mc->mc_top]) == 1) {
3706 DPUTS("collapsing root page!");
3707 mdb_midl_append(mc->mc_txn->mt_free_pgs, mc->mc_pg[mc->mc_top]->mp_pgno);
3708 mc->mc_txn->mt_dbs[mc->mc_dbi].md_root = NODEPGNO(NODEPTR(mc->mc_pg[mc->mc_top], 0));
3709 if ((rc = mdb_get_page(mc->mc_txn, mc->mc_txn->mt_dbs[mc->mc_dbi].md_root, &root)))
3711 mc->mc_txn->mt_dbs[mc->mc_dbi].md_depth--;
3712 mc->mc_txn->mt_dbs[mc->mc_dbi].md_branch_pages--;
3714 DPUTS("root page doesn't need rebalancing");
3718 /* The parent (branch page) must have at least 2 pointers,
3719 * otherwise the tree is invalid.
3721 ptop = mc->mc_top-1;
3722 assert(NUMKEYS(mc->mc_pg[ptop]) > 1);
3724 /* Leaf page fill factor is below the threshold.
3725 * Try to move keys from left or right neighbor, or
3726 * merge with a neighbor page.
3731 mdb_cursor_copy(mc, &mn);
3732 mn.mc_xcursor = NULL;
3734 if (mc->mc_ki[ptop] == 0) {
3735 /* We're the leftmost leaf in our parent.
3737 DPUTS("reading right neighbor");
3739 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
3740 if ((rc = mdb_get_page(mc->mc_txn, NODEPGNO(node), &mn.mc_pg[mn.mc_top])))
3742 mn.mc_ki[mn.mc_top] = 0;
3743 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]);
3745 /* There is at least one neighbor to the left.
3747 DPUTS("reading left neighbor");
3749 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
3750 if ((rc = mdb_get_page(mc->mc_txn, NODEPGNO(node), &mn.mc_pg[mn.mc_top])))
3752 mn.mc_ki[mn.mc_top] = NUMKEYS(mn.mc_pg[mn.mc_top]) - 1;
3753 mc->mc_ki[mc->mc_top] = 0;
3756 DPRINTF("found neighbor page %lu (%u keys, %.1f%% full)",
3757 mn.mc_pg[mn.mc_top]->mp_pgno, NUMKEYS(mn.mc_pg[mn.mc_top]), (float)PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) / 10);
3759 /* If the neighbor page is above threshold and has at least two
3760 * keys, move one key from it.
3762 * Otherwise we should try to merge them.
3764 if (PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) >= FILL_THRESHOLD && NUMKEYS(mn.mc_pg[mn.mc_top]) >= 2)
3765 return mdb_move_node(&mn, mc);
3766 else { /* FIXME: if (has_enough_room()) */
3767 if (mc->mc_ki[ptop] == 0)
3768 return mdb_merge(&mn, mc);
3770 return mdb_merge(mc, &mn);
3775 mdb_del0(MDB_cursor *mc, MDB_node *leaf)
3779 /* add overflow pages to free list */
3780 if (!IS_LEAF2(mc->mc_pg[mc->mc_top]) && F_ISSET(leaf->mn_flags, F_BIGDATA)) {
3784 memcpy(&pg, NODEDATA(leaf), sizeof(pg));
3785 ovpages = OVPAGES(NODEDSZ(leaf), mc->mc_txn->mt_env->me_psize);
3786 for (i=0; i<ovpages; i++) {
3787 DPRINTF("freed ov page %lu", pg);
3788 mdb_midl_append(mc->mc_txn->mt_free_pgs, pg);
3792 mdb_del_node(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], mc->mc_txn->mt_dbs[mc->mc_dbi].md_pad);
3793 mc->mc_txn->mt_dbs[mc->mc_dbi].md_entries--;
3794 rc = mdb_rebalance(mc);
3795 if (rc != MDB_SUCCESS)
3796 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
3802 mdb_del(MDB_txn *txn, MDB_dbi dbi,
3803 MDB_val *key, MDB_val *data)
3808 MDB_val rdata, *xdata;
3812 assert(key != NULL);
3814 DPRINTF("====> delete db %u key [%s]", dbi, DKEY(key));
3816 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
3819 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
3823 if (key->mv_size == 0 || key->mv_size > MAXKEYSIZE) {
3830 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
3831 mc.mc_xcursor = &mx;
3832 mdb_xcursor_init0(&mc);
3834 mc.mc_xcursor = NULL;
3846 rc = mdb_cursor_set(&mc, key, xdata, op, &exact);
3848 rc = mdb_cursor_del(&mc, data ? 0 : MDB_NODUPDATA);
3852 /* Split page <mc->top>, and insert <key,(data|newpgno)> in either left or
3853 * right sibling, at index <mc->ki> (as if unsplit). Updates mc->top and
3854 * mc->ki with the actual values after split, ie if mc->top and mc->ki
3855 * refer to a node in the new right sibling page.
3858 mdb_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata, pgno_t newpgno)
3861 int rc = MDB_SUCCESS, ins_new = 0;
3864 unsigned int i, j, split_indx, nkeys, pmax;
3866 MDB_val sepkey, rkey, rdata;
3868 MDB_page *mp, *rp, *pp;
3873 mp = mc->mc_pg[mc->mc_top];
3874 newindx = mc->mc_ki[mc->mc_top];
3876 DPRINTF("-----> splitting %s page %lu and adding [%s] at index %i",
3877 IS_LEAF(mp) ? "leaf" : "branch", mp->mp_pgno,
3878 DKEY(newkey), mc->mc_ki[mc->mc_top]);
3880 if (mc->mc_snum < 2) {
3881 if ((pp = mdb_new_page(mc, P_BRANCH, 1)) == NULL)
3883 /* shift current top to make room for new parent */
3884 mc->mc_pg[1] = mc->mc_pg[0];
3885 mc->mc_ki[1] = mc->mc_ki[0];
3888 mc->mc_txn->mt_dbs[mc->mc_dbi].md_root = pp->mp_pgno;
3889 DPRINTF("root split! new root = %lu", pp->mp_pgno);
3890 mc->mc_txn->mt_dbs[mc->mc_dbi].md_depth++;
3892 /* Add left (implicit) pointer. */
3893 if ((rc = mdb_add_node(mc, 0, NULL, NULL, mp->mp_pgno, 0)) != MDB_SUCCESS) {
3894 /* undo the pre-push */
3895 mc->mc_pg[0] = mc->mc_pg[1];
3896 mc->mc_ki[0] = mc->mc_ki[1];
3897 mc->mc_txn->mt_dbs[mc->mc_dbi].md_root = mp->mp_pgno;
3898 mc->mc_txn->mt_dbs[mc->mc_dbi].md_depth--;
3905 ptop = mc->mc_top-1;
3906 DPRINTF("parent branch page is %lu", mc->mc_pg[ptop]->mp_pgno);
3909 /* Create a right sibling. */
3910 if ((rp = mdb_new_page(mc, mp->mp_flags, 1)) == NULL)
3912 mdb_cursor_copy(mc, &mn);
3913 mn.mc_pg[mn.mc_top] = rp;
3914 mn.mc_ki[ptop] = mc->mc_ki[ptop]+1;
3915 DPRINTF("new right sibling: page %lu", rp->mp_pgno);
3917 nkeys = NUMKEYS(mp);
3918 split_indx = nkeys / 2 + 1;
3923 unsigned int lsize, rsize, ksize;
3924 /* Move half of the keys to the right sibling */
3926 x = mc->mc_ki[mc->mc_top] - split_indx;
3927 ksize = mc->mc_txn->mt_dbs[mc->mc_dbi].md_pad;
3928 split = LEAF2KEY(mp, split_indx, ksize);
3929 rsize = (nkeys - split_indx) * ksize;
3930 lsize = (nkeys - split_indx) * sizeof(indx_t);
3931 mp->mp_lower -= lsize;
3932 rp->mp_lower += lsize;
3933 mp->mp_upper += rsize - lsize;
3934 rp->mp_upper -= rsize - lsize;
3935 sepkey.mv_size = ksize;
3936 if (newindx == split_indx) {
3937 sepkey.mv_data = newkey->mv_data;
3939 sepkey.mv_data = split;
3942 ins = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], ksize);
3943 memcpy(rp->mp_ptrs, split, rsize);
3944 sepkey.mv_data = rp->mp_ptrs;
3945 memmove(ins+ksize, ins, (split_indx - mc->mc_ki[mc->mc_top]) * ksize);
3946 memcpy(ins, newkey->mv_data, ksize);
3947 mp->mp_lower += sizeof(indx_t);
3948 mp->mp_upper -= ksize - sizeof(indx_t);
3951 memcpy(rp->mp_ptrs, split, x * ksize);
3952 ins = LEAF2KEY(rp, x, ksize);
3953 memcpy(ins, newkey->mv_data, ksize);
3954 memcpy(ins+ksize, split + x * ksize, rsize - x * ksize);
3955 rp->mp_lower += sizeof(indx_t);
3956 rp->mp_upper -= ksize - sizeof(indx_t);
3957 mc->mc_ki[mc->mc_top] = x;
3958 mc->mc_pg[mc->mc_top] = rp;
3963 /* For leaf pages, check the split point based on what
3964 * fits where, since otherwise add_node can fail.
3967 unsigned int psize, nsize;
3968 /* Maximum free space in an empty page */
3969 pmax = mc->mc_txn->mt_env->me_psize - PAGEHDRSZ;
3970 nsize = mdb_leaf_size(mc->mc_txn->mt_env, newkey, newdata);
3971 if (newindx < split_indx) {
3973 for (i=0; i<split_indx; i++) {
3974 node = NODEPTR(mp, i);
3975 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
3976 if (F_ISSET(node->mn_flags, F_BIGDATA))
3977 psize += sizeof(pgno_t);
3979 psize += NODEDSZ(node);
3988 for (i=nkeys-1; i>=split_indx; i--) {
3989 node = NODEPTR(mp, i);
3990 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
3991 if (F_ISSET(node->mn_flags, F_BIGDATA))
3992 psize += sizeof(pgno_t);
3994 psize += NODEDSZ(node);
4004 /* First find the separating key between the split pages.
4006 if (newindx == split_indx) {
4007 sepkey.mv_size = newkey->mv_size;
4008 sepkey.mv_data = newkey->mv_data;
4010 node = NODEPTR(mp, split_indx);
4011 sepkey.mv_size = node->mn_ksize;
4012 sepkey.mv_data = NODEKEY(node);
4016 DPRINTF("separator is [%s]", DKEY(&sepkey));
4018 /* Copy separator key to the parent.
4020 if (SIZELEFT(mn.mc_pg[ptop]) < mdb_branch_size(mc->mc_txn->mt_env, &sepkey)) {
4023 rc = mdb_split(&mn, &sepkey, NULL, rp->mp_pgno);
4025 /* Right page might now have changed parent.
4026 * Check if left page also changed parent.
4028 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
4029 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
4030 mc->mc_pg[ptop] = mn.mc_pg[ptop];
4031 mc->mc_ki[ptop] = mn.mc_ki[ptop] - 1;
4035 rc = mdb_add_node(&mn, mn.mc_ki[ptop], &sepkey, NULL, rp->mp_pgno, 0);
4041 if (rc != MDB_SUCCESS) {
4045 /* Move half of the keys to the right sibling. */
4047 /* grab a page to hold a temporary copy */
4048 if (mc->mc_txn->mt_env->me_dpages) {
4049 copy = mc->mc_txn->mt_env->me_dpages;
4050 mc->mc_txn->mt_env->me_dpages = copy->mp_next;
4052 if ((copy = malloc(mc->mc_txn->mt_env->me_psize)) == NULL)
4056 copy->mp_pgno = mp->mp_pgno;
4057 copy->mp_flags = mp->mp_flags;
4058 copy->mp_lower = PAGEHDRSZ;
4059 copy->mp_upper = mc->mc_txn->mt_env->me_psize;
4060 mc->mc_pg[mc->mc_top] = copy;
4061 for (i = j = 0; i <= nkeys; j++) {
4062 if (i == split_indx) {
4063 /* Insert in right sibling. */
4064 /* Reset insert index for right sibling. */
4065 j = (i == newindx && ins_new);
4066 mc->mc_pg[mc->mc_top] = rp;
4069 if (i == newindx && !ins_new) {
4070 /* Insert the original entry that caused the split. */
4071 rkey.mv_data = newkey->mv_data;
4072 rkey.mv_size = newkey->mv_size;
4074 rdata.mv_data = newdata->mv_data;
4075 rdata.mv_size = newdata->mv_size;
4082 /* Update page and index for the new key. */
4083 mc->mc_ki[mc->mc_top] = j;
4084 } else if (i == nkeys) {
4087 node = NODEPTR(mp, i);
4088 rkey.mv_data = NODEKEY(node);
4089 rkey.mv_size = node->mn_ksize;
4091 rdata.mv_data = NODEDATA(node);
4092 rdata.mv_size = NODEDSZ(node);
4094 pgno = NODEPGNO(node);
4095 flags = node->mn_flags;
4100 if (!IS_LEAF(mp) && j == 0) {
4101 /* First branch index doesn't need key data. */
4105 rc = mdb_add_node(mc, j, &rkey, &rdata, pgno, flags);
4108 /* reset back to original page */
4109 if (newindx < split_indx)
4110 mc->mc_pg[mc->mc_top] = mp;
4112 nkeys = NUMKEYS(copy);
4113 for (i=0; i<nkeys; i++)
4114 mp->mp_ptrs[i] = copy->mp_ptrs[i];
4115 mp->mp_lower = copy->mp_lower;
4116 mp->mp_upper = copy->mp_upper;
4117 memcpy(NODEPTR(mp, nkeys-1), NODEPTR(copy, nkeys-1),
4118 mc->mc_txn->mt_env->me_psize - copy->mp_upper);
4120 /* return tmp page to freelist */
4121 copy->mp_next = mc->mc_txn->mt_env->me_dpages;
4122 mc->mc_txn->mt_env->me_dpages = copy;
4127 mdb_put(MDB_txn *txn, MDB_dbi dbi,
4128 MDB_val *key, MDB_val *data, unsigned int flags)
4133 assert(key != NULL);
4134 assert(data != NULL);
4136 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
4139 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
4143 if (key->mv_size == 0 || key->mv_size > MAXKEYSIZE) {
4147 if ((flags & (MDB_NOOVERWRITE|MDB_NODUPDATA)) != flags)
4154 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
4155 mc.mc_xcursor = &mx;
4156 mdb_xcursor_init0(&mc);
4158 mc.mc_xcursor = NULL;
4160 return mdb_cursor_put(&mc, key, data, flags);
4164 mdb_env_set_flags(MDB_env *env, unsigned int flag, int onoff)
4166 #define CHANGEABLE (MDB_NOSYNC)
4167 if ((flag & CHANGEABLE) != flag)
4170 env->me_flags |= flag;
4172 env->me_flags &= ~flag;
4177 mdb_env_get_flags(MDB_env *env, unsigned int *arg)
4182 *arg = env->me_flags;
4187 mdb_env_get_path(MDB_env *env, const char **arg)
4192 *arg = env->me_path;
4197 mdb_stat0(MDB_env *env, MDB_db *db, MDB_stat *arg)
4199 arg->ms_psize = env->me_psize;
4200 arg->ms_depth = db->md_depth;
4201 arg->ms_branch_pages = db->md_branch_pages;
4202 arg->ms_leaf_pages = db->md_leaf_pages;
4203 arg->ms_overflow_pages = db->md_overflow_pages;
4204 arg->ms_entries = db->md_entries;
4209 mdb_env_stat(MDB_env *env, MDB_stat *arg)
4213 if (env == NULL || arg == NULL)
4216 mdb_env_read_meta(env, &toggle);
4218 return mdb_stat0(env, &env->me_metas[toggle]->mm_dbs[MAIN_DBI], arg);
4222 mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi)
4224 if (txn->mt_dbs[dbi].md_flags & MDB_REVERSEKEY)
4225 txn->mt_dbxs[dbi].md_cmp = memnrcmp;
4226 else if (txn->mt_dbs[dbi].md_flags & MDB_INTEGERKEY)
4227 txn->mt_dbxs[dbi].md_cmp = cintcmp;
4229 txn->mt_dbxs[dbi].md_cmp = memncmp;
4231 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
4232 if (txn->mt_dbs[dbi].md_flags & MDB_INTEGERDUP) {
4233 if (txn->mt_dbs[dbi].md_flags & MDB_DUPFIXED)
4234 txn->mt_dbxs[dbi].md_dcmp = intcmp;
4236 txn->mt_dbxs[dbi].md_dcmp = cintcmp;
4237 } else if (txn->mt_dbs[dbi].md_flags & MDB_REVERSEDUP) {
4238 txn->mt_dbxs[dbi].md_dcmp = memnrcmp;
4240 txn->mt_dbxs[dbi].md_dcmp = memncmp;
4243 txn->mt_dbxs[dbi].md_dcmp = NULL;
4247 int mdb_open(MDB_txn *txn, const char *name, unsigned int flags, MDB_dbi *dbi)
4254 if (txn->mt_dbxs[FREE_DBI].md_cmp == NULL) {
4255 mdb_default_cmp(txn, FREE_DBI);
4261 if (flags & (MDB_DUPSORT|MDB_REVERSEKEY|MDB_INTEGERKEY))
4262 txn->mt_dbs[MAIN_DBI].md_flags |= (flags & (MDB_DUPSORT|MDB_REVERSEKEY|MDB_INTEGERKEY));
4263 mdb_default_cmp(txn, MAIN_DBI);
4267 if (txn->mt_dbxs[MAIN_DBI].md_cmp == NULL) {
4268 mdb_default_cmp(txn, MAIN_DBI);
4271 /* Is the DB already open? */
4273 for (i=2; i<txn->mt_numdbs; i++) {
4274 if (len == txn->mt_dbxs[i].md_name.mv_size &&
4275 !strncmp(name, txn->mt_dbxs[i].md_name.mv_data, len)) {
4281 if (txn->mt_numdbs >= txn->mt_env->me_maxdbs - 1)
4284 /* Find the DB info */
4286 key.mv_data = (void *)name;
4287 rc = mdb_get(txn, MAIN_DBI, &key, &data);
4289 /* Create if requested */
4290 if (rc == MDB_NOTFOUND && (flags & MDB_CREATE)) {
4293 data.mv_size = sizeof(MDB_db);
4294 data.mv_data = &dummy;
4295 memset(&dummy, 0, sizeof(dummy));
4296 dummy.md_root = P_INVALID;
4297 dummy.md_flags = flags & 0xffff;
4299 mc.mc_dbi = MAIN_DBI;
4301 rc = mdb_cursor_put(&mc, &key, &data, F_SUBDATA);
4305 /* OK, got info, add to table */
4306 if (rc == MDB_SUCCESS) {
4307 txn->mt_dbxs[txn->mt_numdbs].md_name.mv_data = strdup(name);
4308 txn->mt_dbxs[txn->mt_numdbs].md_name.mv_size = len;
4309 txn->mt_dbxs[txn->mt_numdbs].md_rel = NULL;
4310 txn->mt_dbxs[txn->mt_numdbs].md_parent = MAIN_DBI;
4311 txn->mt_dbxs[txn->mt_numdbs].md_dirty = dirty;
4312 memcpy(&txn->mt_dbs[txn->mt_numdbs], data.mv_data, sizeof(MDB_db));
4313 *dbi = txn->mt_numdbs;
4314 txn->mt_env->me_dbs[0][txn->mt_numdbs] = txn->mt_dbs[txn->mt_numdbs];
4315 txn->mt_env->me_dbs[1][txn->mt_numdbs] = txn->mt_dbs[txn->mt_numdbs];
4316 mdb_default_cmp(txn, txn->mt_numdbs);
4323 int mdb_stat(MDB_txn *txn, MDB_dbi dbi, MDB_stat *arg)
4325 if (txn == NULL || arg == NULL || dbi >= txn->mt_numdbs)
4328 return mdb_stat0(txn->mt_env, &txn->mt_dbs[dbi], arg);
4331 void mdb_close(MDB_txn *txn, MDB_dbi dbi)
4334 if (dbi <= MAIN_DBI || dbi >= txn->mt_numdbs)
4336 ptr = txn->mt_dbxs[dbi].md_name.mv_data;
4337 txn->mt_dbxs[dbi].md_name.mv_data = NULL;
4338 txn->mt_dbxs[dbi].md_name.mv_size = 0;
4342 int mdb_set_compare(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
4344 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
4347 txn->mt_dbxs[dbi].md_cmp = cmp;
4351 int mdb_set_dupsort(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
4353 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
4356 txn->mt_dbxs[dbi].md_dcmp = cmp;
4360 int mdb_set_relfunc(MDB_txn *txn, MDB_dbi dbi, MDB_rel_func *rel)
4362 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
4365 txn->mt_dbxs[dbi].md_rel = rel;