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 2011-2013 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.
36 #include <sys/types.h>
38 #include <sys/param.h>
44 #ifdef HAVE_SYS_FILE_H
61 #if !(defined(BYTE_ORDER) || defined(__BYTE_ORDER))
62 #include <netinet/in.h>
63 #include <resolv.h> /* defines BYTE_ORDER on HPUX and Solaris */
66 #if defined(__APPLE__) || defined (BSD)
67 # define MDB_USE_POSIX_SEM 1
68 # define MDB_FDATASYNC fsync
69 #elif defined(ANDROID)
70 # define MDB_FDATASYNC fsync
75 #ifdef MDB_USE_POSIX_SEM
76 #include <semaphore.h>
81 #include <valgrind/memcheck.h>
82 #define VGMEMP_CREATE(h,r,z) VALGRIND_CREATE_MEMPOOL(h,r,z)
83 #define VGMEMP_ALLOC(h,a,s) VALGRIND_MEMPOOL_ALLOC(h,a,s)
84 #define VGMEMP_FREE(h,a) VALGRIND_MEMPOOL_FREE(h,a)
85 #define VGMEMP_DESTROY(h) VALGRIND_DESTROY_MEMPOOL(h)
86 #define VGMEMP_DEFINED(a,s) VALGRIND_MAKE_MEM_DEFINED(a,s)
88 #define VGMEMP_CREATE(h,r,z)
89 #define VGMEMP_ALLOC(h,a,s)
90 #define VGMEMP_FREE(h,a)
91 #define VGMEMP_DESTROY(h)
92 #define VGMEMP_DEFINED(a,s)
96 # if (defined(_LITTLE_ENDIAN) || defined(_BIG_ENDIAN)) && !(defined(_LITTLE_ENDIAN) && defined(_BIG_ENDIAN))
97 /* Solaris just defines one or the other */
98 # define LITTLE_ENDIAN 1234
99 # define BIG_ENDIAN 4321
100 # ifdef _LITTLE_ENDIAN
101 # define BYTE_ORDER LITTLE_ENDIAN
103 # define BYTE_ORDER BIG_ENDIAN
106 # define BYTE_ORDER __BYTE_ORDER
110 #ifndef LITTLE_ENDIAN
111 #define LITTLE_ENDIAN __LITTLE_ENDIAN
114 #define BIG_ENDIAN __BIG_ENDIAN
117 #if defined(__i386) || defined(__x86_64) || defined(_M_IX86)
118 #define MISALIGNED_OK 1
124 #if (BYTE_ORDER == LITTLE_ENDIAN) == (BYTE_ORDER == BIG_ENDIAN)
125 # error "Unknown or unsupported endianness (BYTE_ORDER)"
126 #elif (-6 & 5) || CHAR_BIT != 8 || UINT_MAX < 0xffffffff || ULONG_MAX % 0xFFFF
127 # error "Two's complement, reasonably sized integer types, please"
130 /** @defgroup internal MDB Internals
133 /** @defgroup compat Windows Compatibility Macros
134 * A bunch of macros to minimize the amount of platform-specific ifdefs
135 * needed throughout the rest of the code. When the features this library
136 * needs are similar enough to POSIX to be hidden in a one-or-two line
137 * replacement, this macro approach is used.
141 #define pthread_t DWORD
142 #define pthread_mutex_t HANDLE
143 #define pthread_key_t DWORD
144 #define pthread_self() GetCurrentThreadId()
145 #define pthread_key_create(x,y) \
146 ((*(x) = TlsAlloc()) == TLS_OUT_OF_INDEXES ? ErrCode() : 0)
147 #define pthread_key_delete(x) TlsFree(x)
148 #define pthread_getspecific(x) TlsGetValue(x)
149 #define pthread_setspecific(x,y) (TlsSetValue(x,y) ? 0 : ErrCode())
150 #define pthread_mutex_unlock(x) ReleaseMutex(x)
151 #define pthread_mutex_lock(x) WaitForSingleObject(x, INFINITE)
152 #define LOCK_MUTEX_R(env) pthread_mutex_lock((env)->me_rmutex)
153 #define UNLOCK_MUTEX_R(env) pthread_mutex_unlock((env)->me_rmutex)
154 #define LOCK_MUTEX_W(env) pthread_mutex_lock((env)->me_wmutex)
155 #define UNLOCK_MUTEX_W(env) pthread_mutex_unlock((env)->me_wmutex)
156 #define getpid() GetCurrentProcessId()
157 #define MDB_FDATASYNC(fd) (!FlushFileBuffers(fd))
158 #define MDB_MSYNC(addr,len,flags) (!FlushViewOfFile(addr,len))
159 #define ErrCode() GetLastError()
160 #define GET_PAGESIZE(x) {SYSTEM_INFO si; GetSystemInfo(&si); (x) = si.dwPageSize;}
161 #define close(fd) CloseHandle(fd)
162 #define munmap(ptr,len) UnmapViewOfFile(ptr)
165 #ifdef MDB_USE_POSIX_SEM
167 #define LOCK_MUTEX_R(env) mdb_sem_wait((env)->me_rmutex)
168 #define UNLOCK_MUTEX_R(env) sem_post((env)->me_rmutex)
169 #define LOCK_MUTEX_W(env) mdb_sem_wait((env)->me_wmutex)
170 #define UNLOCK_MUTEX_W(env) sem_post((env)->me_wmutex)
173 mdb_sem_wait(sem_t *sem)
176 while ((rc = sem_wait(sem)) && (rc = errno) == EINTR) ;
181 /** Lock the reader mutex.
183 #define LOCK_MUTEX_R(env) pthread_mutex_lock(&(env)->me_txns->mti_mutex)
184 /** Unlock the reader mutex.
186 #define UNLOCK_MUTEX_R(env) pthread_mutex_unlock(&(env)->me_txns->mti_mutex)
188 /** Lock the writer mutex.
189 * Only a single write transaction is allowed at a time. Other writers
190 * will block waiting for this mutex.
192 #define LOCK_MUTEX_W(env) pthread_mutex_lock(&(env)->me_txns->mti_wmutex)
193 /** Unlock the writer mutex.
195 #define UNLOCK_MUTEX_W(env) pthread_mutex_unlock(&(env)->me_txns->mti_wmutex)
196 #endif /* MDB_USE_POSIX_SEM */
198 /** Get the error code for the last failed system function.
200 #define ErrCode() errno
202 /** A value for an invalid file handle.
203 * Mainly used to initialize file variables and signify that they are
206 #define INVALID_HANDLE_VALUE (-1)
208 /** Get the size of a memory page for the system.
209 * This is the basic size that the platform's memory manager uses, and is
210 * fundamental to the use of memory-mapped files.
212 #define GET_PAGESIZE(x) ((x) = sysconf(_SC_PAGE_SIZE))
215 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
218 #define MNAME_LEN (sizeof(pthread_mutex_t))
224 /** A flag for opening a file and requesting synchronous data writes.
225 * This is only used when writing a meta page. It's not strictly needed;
226 * we could just do a normal write and then immediately perform a flush.
227 * But if this flag is available it saves us an extra system call.
229 * @note If O_DSYNC is undefined but exists in /usr/include,
230 * preferably set some compiler flag to get the definition.
231 * Otherwise compile with the less efficient -DMDB_DSYNC=O_SYNC.
234 # define MDB_DSYNC O_DSYNC
238 /** Function for flushing the data of a file. Define this to fsync
239 * if fdatasync() is not supported.
241 #ifndef MDB_FDATASYNC
242 # define MDB_FDATASYNC fdatasync
246 # define MDB_MSYNC(addr,len,flags) msync(addr,len,flags)
257 /** A page number in the database.
258 * Note that 64 bit page numbers are overkill, since pages themselves
259 * already represent 12-13 bits of addressable memory, and the OS will
260 * always limit applications to a maximum of 63 bits of address space.
262 * @note In the #MDB_node structure, we only store 48 bits of this value,
263 * which thus limits us to only 60 bits of addressable data.
265 typedef MDB_ID pgno_t;
267 /** A transaction ID.
268 * See struct MDB_txn.mt_txnid for details.
270 typedef MDB_ID txnid_t;
272 /** @defgroup debug Debug Macros
276 /** Enable debug output.
277 * Set this to 1 for copious tracing. Set to 2 to add dumps of all IDLs
278 * read from and written to the database (used for free space management).
283 #if !(__STDC_VERSION__ >= 199901L || defined(__GNUC__))
286 # define DPRINTF (void) /* Vararg macros may be unsupported */
288 static int mdb_debug;
289 static txnid_t mdb_debug_start;
291 /** Print a debug message with printf formatting. */
292 # define DPRINTF(fmt, ...) /**< Requires 2 or more args */ \
293 ((void) ((mdb_debug) && \
294 fprintf(stderr, "%s:%d " fmt "\n", __func__, __LINE__, __VA_ARGS__)))
296 # define DPRINTF(fmt, ...) ((void) 0)
297 # define MDB_DEBUG_SKIP
299 /** Print a debug string.
300 * The string is printed literally, with no format processing.
302 #define DPUTS(arg) DPRINTF("%s", arg)
305 /** A default memory page size.
306 * The actual size is platform-dependent, but we use this for
307 * boot-strapping. We probably should not be using this any more.
308 * The #GET_PAGESIZE() macro is used to get the actual size.
310 * Note that we don't currently support Huge pages. On Linux,
311 * regular data files cannot use Huge pages, and in general
312 * Huge pages aren't actually pageable. We rely on the OS
313 * demand-pager to read our data and page it out when memory
314 * pressure from other processes is high. So until OSs have
315 * actual paging support for Huge pages, they're not viable.
317 #define MDB_PAGESIZE 4096
319 /** The minimum number of keys required in a database page.
320 * Setting this to a larger value will place a smaller bound on the
321 * maximum size of a data item. Data items larger than this size will
322 * be pushed into overflow pages instead of being stored directly in
323 * the B-tree node. This value used to default to 4. With a page size
324 * of 4096 bytes that meant that any item larger than 1024 bytes would
325 * go into an overflow page. That also meant that on average 2-3KB of
326 * each overflow page was wasted space. The value cannot be lower than
327 * 2 because then there would no longer be a tree structure. With this
328 * value, items larger than 2KB will go into overflow pages, and on
329 * average only 1KB will be wasted.
331 #define MDB_MINKEYS 2
333 /** A stamp that identifies a file as an MDB file.
334 * There's nothing special about this value other than that it is easily
335 * recognizable, and it will reflect any byte order mismatches.
337 #define MDB_MAGIC 0xBEEFC0DE
339 /** The version number for a database's file format. */
340 #define MDB_VERSION 1
342 /** @brief The maximum size of a key in the database.
344 * We require that keys all fit onto a regular page. This limit
345 * could be raised a bit further if needed; to something just
346 * under #MDB_PAGESIZE / #MDB_MINKEYS.
348 * Note that data items in an #MDB_DUPSORT database are actually keys
349 * of a subDB, so they're also limited to this size.
351 #ifndef MDB_MAXKEYSIZE
352 #define MDB_MAXKEYSIZE 511
355 /** @brief The maximum size of a data item.
357 * We only store a 32 bit value for node sizes.
359 #define MAXDATASIZE 0xffffffffUL
364 * This is used for printing a hex dump of a key's contents.
366 #define DKBUF char kbuf[(MDB_MAXKEYSIZE*2+1)]
367 /** Display a key in hex.
369 * Invoke a function to display a key in hex.
371 #define DKEY(x) mdb_dkey(x, kbuf)
373 #define DKBUF typedef int dummy_kbuf /* so we can put ';' after */
377 /** An invalid page number.
378 * Mainly used to denote an empty tree.
380 #define P_INVALID (~(pgno_t)0)
382 /** Test if the flags \b f are set in a flag word \b w. */
383 #define F_ISSET(w, f) (((w) & (f)) == (f))
385 /** Used for offsets within a single page.
386 * Since memory pages are typically 4 or 8KB in size, 12-13 bits,
389 typedef uint16_t indx_t;
391 /** Default size of memory map.
392 * This is certainly too small for any actual applications. Apps should always set
393 * the size explicitly using #mdb_env_set_mapsize().
395 #define DEFAULT_MAPSIZE 1048576
397 /** @defgroup readers Reader Lock Table
398 * Readers don't acquire any locks for their data access. Instead, they
399 * simply record their transaction ID in the reader table. The reader
400 * mutex is needed just to find an empty slot in the reader table. The
401 * slot's address is saved in thread-specific data so that subsequent read
402 * transactions started by the same thread need no further locking to proceed.
404 * If #MDB_NOTLS is set, the slot address is not saved in thread-specific data.
406 * No reader table is used if the database is on a read-only filesystem.
408 * Since the database uses multi-version concurrency control, readers don't
409 * actually need any locking. This table is used to keep track of which
410 * readers are using data from which old transactions, so that we'll know
411 * when a particular old transaction is no longer in use. Old transactions
412 * that have discarded any data pages can then have those pages reclaimed
413 * for use by a later write transaction.
415 * The lock table is constructed such that reader slots are aligned with the
416 * processor's cache line size. Any slot is only ever used by one thread.
417 * This alignment guarantees that there will be no contention or cache
418 * thrashing as threads update their own slot info, and also eliminates
419 * any need for locking when accessing a slot.
421 * A writer thread will scan every slot in the table to determine the oldest
422 * outstanding reader transaction. Any freed pages older than this will be
423 * reclaimed by the writer. The writer doesn't use any locks when scanning
424 * this table. This means that there's no guarantee that the writer will
425 * see the most up-to-date reader info, but that's not required for correct
426 * operation - all we need is to know the upper bound on the oldest reader,
427 * we don't care at all about the newest reader. So the only consequence of
428 * reading stale information here is that old pages might hang around a
429 * while longer before being reclaimed. That's actually good anyway, because
430 * the longer we delay reclaiming old pages, the more likely it is that a
431 * string of contiguous pages can be found after coalescing old pages from
432 * many old transactions together.
435 /** Number of slots in the reader table.
436 * This value was chosen somewhat arbitrarily. 126 readers plus a
437 * couple mutexes fit exactly into 8KB on my development machine.
438 * Applications should set the table size using #mdb_env_set_maxreaders().
440 #define DEFAULT_READERS 126
442 /** The size of a CPU cache line in bytes. We want our lock structures
443 * aligned to this size to avoid false cache line sharing in the
445 * This value works for most CPUs. For Itanium this should be 128.
451 /** The information we store in a single slot of the reader table.
452 * In addition to a transaction ID, we also record the process and
453 * thread ID that owns a slot, so that we can detect stale information,
454 * e.g. threads or processes that went away without cleaning up.
455 * @note We currently don't check for stale records. We simply re-init
456 * the table when we know that we're the only process opening the
459 typedef struct MDB_rxbody {
460 /** Current Transaction ID when this transaction began, or (txnid_t)-1.
461 * Multiple readers that start at the same time will probably have the
462 * same ID here. Again, it's not important to exclude them from
463 * anything; all we need to know is which version of the DB they
464 * started from so we can avoid overwriting any data used in that
465 * particular version.
468 /** The process ID of the process owning this reader txn. */
470 /** The thread ID of the thread owning this txn. */
474 /** The actual reader record, with cacheline padding. */
475 typedef struct MDB_reader {
478 /** shorthand for mrb_txnid */
479 #define mr_txnid mru.mrx.mrb_txnid
480 #define mr_pid mru.mrx.mrb_pid
481 #define mr_tid mru.mrx.mrb_tid
482 /** cache line alignment */
483 char pad[(sizeof(MDB_rxbody)+CACHELINE-1) & ~(CACHELINE-1)];
487 /** The header for the reader table.
488 * The table resides in a memory-mapped file. (This is a different file
489 * than is used for the main database.)
491 * For POSIX the actual mutexes reside in the shared memory of this
492 * mapped file. On Windows, mutexes are named objects allocated by the
493 * kernel; we store the mutex names in this mapped file so that other
494 * processes can grab them. This same approach is also used on
495 * MacOSX/Darwin (using named semaphores) since MacOSX doesn't support
496 * process-shared POSIX mutexes. For these cases where a named object
497 * is used, the object name is derived from a 64 bit FNV hash of the
498 * environment pathname. As such, naming collisions are extremely
499 * unlikely. If a collision occurs, the results are unpredictable.
501 typedef struct MDB_txbody {
502 /** Stamp identifying this as an MDB file. It must be set
505 /** Version number of this lock file. Must be set to #MDB_VERSION. */
506 uint32_t mtb_version;
507 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
508 char mtb_rmname[MNAME_LEN];
510 /** Mutex protecting access to this table.
511 * This is the reader lock that #LOCK_MUTEX_R acquires.
513 pthread_mutex_t mtb_mutex;
515 /** The ID of the last transaction committed to the database.
516 * This is recorded here only for convenience; the value can always
517 * be determined by reading the main database meta pages.
520 /** The number of slots that have been used in the reader table.
521 * This always records the maximum count, it is not decremented
522 * when readers release their slots.
524 unsigned mtb_numreaders;
527 /** The actual reader table definition. */
528 typedef struct MDB_txninfo {
531 #define mti_magic mt1.mtb.mtb_magic
532 #define mti_version mt1.mtb.mtb_version
533 #define mti_mutex mt1.mtb.mtb_mutex
534 #define mti_rmname mt1.mtb.mtb_rmname
535 #define mti_txnid mt1.mtb.mtb_txnid
536 #define mti_numreaders mt1.mtb.mtb_numreaders
537 char pad[(sizeof(MDB_txbody)+CACHELINE-1) & ~(CACHELINE-1)];
540 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
541 char mt2_wmname[MNAME_LEN];
542 #define mti_wmname mt2.mt2_wmname
544 pthread_mutex_t mt2_wmutex;
545 #define mti_wmutex mt2.mt2_wmutex
547 char pad[(MNAME_LEN+CACHELINE-1) & ~(CACHELINE-1)];
549 MDB_reader mti_readers[1];
553 /** Common header for all page types.
554 * Overflow records occupy a number of contiguous pages with no
555 * headers on any page after the first.
557 typedef struct MDB_page {
558 #define mp_pgno mp_p.p_pgno
559 #define mp_next mp_p.p_next
561 pgno_t p_pgno; /**< page number */
562 void * p_next; /**< for in-memory list of freed structs */
565 /** @defgroup mdb_page Page Flags
567 * Flags for the page headers.
570 #define P_BRANCH 0x01 /**< branch page */
571 #define P_LEAF 0x02 /**< leaf page */
572 #define P_OVERFLOW 0x04 /**< overflow page */
573 #define P_META 0x08 /**< meta page */
574 #define P_DIRTY 0x10 /**< dirty page */
575 #define P_LEAF2 0x20 /**< for #MDB_DUPFIXED records */
576 #define P_SUBP 0x40 /**< for #MDB_DUPSORT sub-pages */
578 uint16_t mp_flags; /**< @ref mdb_page */
579 #define mp_lower mp_pb.pb.pb_lower
580 #define mp_upper mp_pb.pb.pb_upper
581 #define mp_pages mp_pb.pb_pages
584 indx_t pb_lower; /**< lower bound of free space */
585 indx_t pb_upper; /**< upper bound of free space */
587 uint32_t pb_pages; /**< number of overflow pages */
589 indx_t mp_ptrs[1]; /**< dynamic size */
592 /** Size of the page header, excluding dynamic data at the end */
593 #define PAGEHDRSZ ((unsigned) offsetof(MDB_page, mp_ptrs))
595 /** Address of first usable data byte in a page, after the header */
596 #define METADATA(p) ((void *)((char *)(p) + PAGEHDRSZ))
598 /** Number of nodes on a page */
599 #define NUMKEYS(p) (((p)->mp_lower - PAGEHDRSZ) >> 1)
601 /** The amount of space remaining in the page */
602 #define SIZELEFT(p) (indx_t)((p)->mp_upper - (p)->mp_lower)
604 /** The percentage of space used in the page, in tenths of a percent. */
605 #define PAGEFILL(env, p) (1000L * ((env)->me_psize - PAGEHDRSZ - SIZELEFT(p)) / \
606 ((env)->me_psize - PAGEHDRSZ))
607 /** The minimum page fill factor, in tenths of a percent.
608 * Pages emptier than this are candidates for merging.
610 #define FILL_THRESHOLD 250
612 /** Test if a page is a leaf page */
613 #define IS_LEAF(p) F_ISSET((p)->mp_flags, P_LEAF)
614 /** Test if a page is a LEAF2 page */
615 #define IS_LEAF2(p) F_ISSET((p)->mp_flags, P_LEAF2)
616 /** Test if a page is a branch page */
617 #define IS_BRANCH(p) F_ISSET((p)->mp_flags, P_BRANCH)
618 /** Test if a page is an overflow page */
619 #define IS_OVERFLOW(p) F_ISSET((p)->mp_flags, P_OVERFLOW)
620 /** Test if a page is a sub page */
621 #define IS_SUBP(p) F_ISSET((p)->mp_flags, P_SUBP)
623 /** The number of overflow pages needed to store the given size. */
624 #define OVPAGES(size, psize) ((PAGEHDRSZ-1 + (size)) / (psize) + 1)
626 /** Header for a single key/data pair within a page.
627 * We guarantee 2-byte alignment for nodes.
629 typedef struct MDB_node {
630 /** lo and hi are used for data size on leaf nodes and for
631 * child pgno on branch nodes. On 64 bit platforms, flags
632 * is also used for pgno. (Branch nodes have no flags).
633 * They are in host byte order in case that lets some
634 * accesses be optimized into a 32-bit word access.
636 #define mn_lo mn_offset[BYTE_ORDER!=LITTLE_ENDIAN]
637 #define mn_hi mn_offset[BYTE_ORDER==LITTLE_ENDIAN] /**< part of dsize or pgno */
638 unsigned short mn_offset[2]; /**< storage for #mn_lo and #mn_hi */
639 /** @defgroup mdb_node Node Flags
641 * Flags for node headers.
644 #define F_BIGDATA 0x01 /**< data put on overflow page */
645 #define F_SUBDATA 0x02 /**< data is a sub-database */
646 #define F_DUPDATA 0x04 /**< data has duplicates */
648 /** valid flags for #mdb_node_add() */
649 #define NODE_ADD_FLAGS (F_DUPDATA|F_SUBDATA|MDB_RESERVE|MDB_APPEND)
652 unsigned short mn_flags; /**< @ref mdb_node */
653 unsigned short mn_ksize; /**< key size */
654 char mn_data[1]; /**< key and data are appended here */
657 /** Size of the node header, excluding dynamic data at the end */
658 #define NODESIZE offsetof(MDB_node, mn_data)
660 /** Bit position of top word in page number, for shifting mn_flags */
661 #define PGNO_TOPWORD ((pgno_t)-1 > 0xffffffffu ? 32 : 0)
663 /** Size of a node in a branch page with a given key.
664 * This is just the node header plus the key, there is no data.
666 #define INDXSIZE(k) (NODESIZE + ((k) == NULL ? 0 : (k)->mv_size))
668 /** Size of a node in a leaf page with a given key and data.
669 * This is node header plus key plus data size.
671 #define LEAFSIZE(k, d) (NODESIZE + (k)->mv_size + (d)->mv_size)
673 /** Address of node \b i in page \b p */
674 #define NODEPTR(p, i) ((MDB_node *)((char *)(p) + (p)->mp_ptrs[i]))
676 /** Address of the key for the node */
677 #define NODEKEY(node) (void *)((node)->mn_data)
679 /** Address of the data for a node */
680 #define NODEDATA(node) (void *)((char *)(node)->mn_data + (node)->mn_ksize)
682 /** Get the page number pointed to by a branch node */
683 #define NODEPGNO(node) \
684 ((node)->mn_lo | ((pgno_t) (node)->mn_hi << 16) | \
685 (PGNO_TOPWORD ? ((pgno_t) (node)->mn_flags << PGNO_TOPWORD) : 0))
686 /** Set the page number in a branch node */
687 #define SETPGNO(node,pgno) do { \
688 (node)->mn_lo = (pgno) & 0xffff; (node)->mn_hi = (pgno) >> 16; \
689 if (PGNO_TOPWORD) (node)->mn_flags = (pgno) >> PGNO_TOPWORD; } while(0)
691 /** Get the size of the data in a leaf node */
692 #define NODEDSZ(node) ((node)->mn_lo | ((unsigned)(node)->mn_hi << 16))
693 /** Set the size of the data for a leaf node */
694 #define SETDSZ(node,size) do { \
695 (node)->mn_lo = (size) & 0xffff; (node)->mn_hi = (size) >> 16;} while(0)
696 /** The size of a key in a node */
697 #define NODEKSZ(node) ((node)->mn_ksize)
699 /** Copy a page number from src to dst */
701 #define COPY_PGNO(dst,src) dst = src
703 #if SIZE_MAX > 4294967295UL
704 #define COPY_PGNO(dst,src) do { \
705 unsigned short *s, *d; \
706 s = (unsigned short *)&(src); \
707 d = (unsigned short *)&(dst); \
714 #define COPY_PGNO(dst,src) do { \
715 unsigned short *s, *d; \
716 s = (unsigned short *)&(src); \
717 d = (unsigned short *)&(dst); \
723 /** The address of a key in a LEAF2 page.
724 * LEAF2 pages are used for #MDB_DUPFIXED sorted-duplicate sub-DBs.
725 * There are no node headers, keys are stored contiguously.
727 #define LEAF2KEY(p, i, ks) ((char *)(p) + PAGEHDRSZ + ((i)*(ks)))
729 /** Set the \b node's key into \b key, if requested. */
730 #define MDB_GET_KEY(node, key) { if ((key) != NULL) { \
731 (key)->mv_size = NODEKSZ(node); (key)->mv_data = NODEKEY(node); } }
733 /** Information about a single database in the environment. */
734 typedef struct MDB_db {
735 uint32_t md_pad; /**< also ksize for LEAF2 pages */
736 uint16_t md_flags; /**< @ref mdb_dbi_open */
737 uint16_t md_depth; /**< depth of this tree */
738 pgno_t md_branch_pages; /**< number of internal pages */
739 pgno_t md_leaf_pages; /**< number of leaf pages */
740 pgno_t md_overflow_pages; /**< number of overflow pages */
741 size_t md_entries; /**< number of data items */
742 pgno_t md_root; /**< the root page of this tree */
745 /** mdb_dbi_open flags */
746 #define MDB_VALID 0x8000 /**< DB handle is valid, for me_dbflags */
747 #define PERSISTENT_FLAGS (0xffff & ~(MDB_VALID))
748 #define VALID_FLAGS (MDB_REVERSEKEY|MDB_DUPSORT|MDB_INTEGERKEY|MDB_DUPFIXED|\
749 MDB_INTEGERDUP|MDB_REVERSEDUP|MDB_CREATE)
751 /** Handle for the DB used to track free pages. */
753 /** Handle for the default DB. */
756 /** Meta page content. */
757 typedef struct MDB_meta {
758 /** Stamp identifying this as an MDB file. It must be set
761 /** Version number of this lock file. Must be set to #MDB_VERSION. */
763 void *mm_address; /**< address for fixed mapping */
764 size_t mm_mapsize; /**< size of mmap region */
765 MDB_db mm_dbs[2]; /**< first is free space, 2nd is main db */
766 /** The size of pages used in this DB */
767 #define mm_psize mm_dbs[0].md_pad
768 /** Any persistent environment flags. @ref mdb_env */
769 #define mm_flags mm_dbs[0].md_flags
770 pgno_t mm_last_pg; /**< last used page in file */
771 txnid_t mm_txnid; /**< txnid that committed this page */
774 /** Buffer for a stack-allocated dirty page.
775 * The members define size and alignment, and silence type
776 * aliasing warnings. They are not used directly; that could
777 * mean incorrectly using several union members in parallel.
779 typedef union MDB_pagebuf {
780 char mb_raw[MDB_PAGESIZE];
783 char mm_pad[PAGEHDRSZ];
788 /** Auxiliary DB info.
789 * The information here is mostly static/read-only. There is
790 * only a single copy of this record in the environment.
792 typedef struct MDB_dbx {
793 MDB_val md_name; /**< name of the database */
794 MDB_cmp_func *md_cmp; /**< function for comparing keys */
795 MDB_cmp_func *md_dcmp; /**< function for comparing data items */
796 MDB_rel_func *md_rel; /**< user relocate function */
797 void *md_relctx; /**< user-provided context for md_rel */
800 /** A database transaction.
801 * Every operation requires a transaction handle.
804 MDB_txn *mt_parent; /**< parent of a nested txn */
805 MDB_txn *mt_child; /**< nested txn under this txn */
806 pgno_t mt_next_pgno; /**< next unallocated page */
807 /** The ID of this transaction. IDs are integers incrementing from 1.
808 * Only committed write transactions increment the ID. If a transaction
809 * aborts, the ID may be re-used by the next writer.
812 MDB_env *mt_env; /**< the DB environment */
813 /** The list of pages that became unused during this transaction.
817 MDB_ID2L dirty_list; /**< for write txns: modified pages */
818 MDB_reader *reader; /**< this thread's reader table slot or NULL */
820 /** Array of records for each DB known in the environment. */
822 /** Array of MDB_db records for each known DB */
824 /** @defgroup mt_dbflag Transaction DB Flags
828 #define DB_DIRTY 0x01 /**< DB was written in this txn */
829 #define DB_STALE 0x02 /**< DB record is older than txnID */
830 #define DB_NEW 0x04 /**< DB handle opened in this txn */
831 #define DB_VALID 0x08 /**< DB handle is valid, see also #MDB_VALID */
833 /** In write txns, array of cursors for each DB */
834 MDB_cursor **mt_cursors;
835 /** Array of flags for each DB */
836 unsigned char *mt_dbflags;
837 /** Number of DB records in use. This number only ever increments;
838 * we don't decrement it when individual DB handles are closed.
842 /** @defgroup mdb_txn Transaction Flags
846 #define MDB_TXN_RDONLY 0x01 /**< read-only transaction */
847 #define MDB_TXN_ERROR 0x02 /**< an error has occurred */
848 #define MDB_TXN_DIRTY 0x04 /**< must write, even if dirty list is empty */
850 unsigned int mt_flags; /**< @ref mdb_txn */
851 /** dirty_list maxsize - # of allocated pages allowed, including in parent txns */
852 unsigned int mt_dirty_room;
853 /** Tracks which of the two meta pages was used at the start
854 * of this transaction.
856 unsigned int mt_toggle;
859 /** Enough space for 2^32 nodes with minimum of 2 keys per node. I.e., plenty.
860 * At 4 keys per node, enough for 2^64 nodes, so there's probably no need to
861 * raise this on a 64 bit machine.
863 #define CURSOR_STACK 32
867 /** Cursors are used for all DB operations */
869 /** Next cursor on this DB in this txn */
871 /** Original cursor if this is a shadow */
873 /** Context used for databases with #MDB_DUPSORT, otherwise NULL */
874 struct MDB_xcursor *mc_xcursor;
875 /** The transaction that owns this cursor */
877 /** The database handle this cursor operates on */
879 /** The database record for this cursor */
881 /** The database auxiliary record for this cursor */
883 /** The @ref mt_dbflag for this database */
884 unsigned char *mc_dbflag;
885 unsigned short mc_snum; /**< number of pushed pages */
886 unsigned short mc_top; /**< index of top page, normally mc_snum-1 */
887 /** @defgroup mdb_cursor Cursor Flags
889 * Cursor state flags.
892 #define C_INITIALIZED 0x01 /**< cursor has been initialized and is valid */
893 #define C_EOF 0x02 /**< No more data */
894 #define C_SUB 0x04 /**< Cursor is a sub-cursor */
895 #define C_SHADOW 0x08 /**< Cursor is a dup from a parent txn */
896 #define C_ALLOCD 0x10 /**< Cursor was malloc'd */
897 #define C_SPLITTING 0x20 /**< Cursor is in page_split */
898 #define C_UNTRACK 0x40 /**< Un-track cursor when closing */
900 unsigned int mc_flags; /**< @ref mdb_cursor */
901 MDB_page *mc_pg[CURSOR_STACK]; /**< stack of pushed pages */
902 indx_t mc_ki[CURSOR_STACK]; /**< stack of page indices */
905 /** Context for sorted-dup records.
906 * We could have gone to a fully recursive design, with arbitrarily
907 * deep nesting of sub-databases. But for now we only handle these
908 * levels - main DB, optional sub-DB, sorted-duplicate DB.
910 typedef struct MDB_xcursor {
911 /** A sub-cursor for traversing the Dup DB */
912 MDB_cursor mx_cursor;
913 /** The database record for this Dup DB */
915 /** The auxiliary DB record for this Dup DB */
917 /** The @ref mt_dbflag for this Dup DB */
918 unsigned char mx_dbflag;
921 /** State of FreeDB old pages, stored in the MDB_env */
922 typedef struct MDB_pgstate {
923 txnid_t mf_pglast; /**< ID of last old page record we used */
924 pgno_t *mf_pghead; /**< old pages reclaimed from freelist */
925 pgno_t *mf_pgfree; /**< memory to free when dropping me_pghead */
928 /** The database environment. */
930 HANDLE me_fd; /**< The main data file */
931 HANDLE me_lfd; /**< The lock file */
932 HANDLE me_mfd; /**< just for writing the meta pages */
933 /** Failed to update the meta page. Probably an I/O error. */
934 #define MDB_FATAL_ERROR 0x80000000U
935 /** Some fields are initialized. */
936 #define MDB_ENV_ACTIVE 0x20000000U
937 /** me_txkey is set */
938 #define MDB_ENV_TXKEY 0x10000000U
939 uint32_t me_flags; /**< @ref mdb_env */
940 unsigned int me_psize; /**< size of a page, from #GET_PAGESIZE */
941 unsigned int me_maxreaders; /**< size of the reader table */
942 unsigned int me_numreaders; /**< max numreaders set by this env */
943 MDB_dbi me_numdbs; /**< number of DBs opened */
944 MDB_dbi me_maxdbs; /**< size of the DB table */
945 pid_t me_pid; /**< process ID of this env */
946 char *me_path; /**< path to the DB files */
947 char *me_map; /**< the memory map of the data file */
948 MDB_txninfo *me_txns; /**< the memory map of the lock file or NULL */
949 MDB_meta *me_metas[2]; /**< pointers to the two meta pages */
950 MDB_txn *me_txn; /**< current write transaction */
951 size_t me_mapsize; /**< size of the data memory map */
952 off_t me_size; /**< current file size */
953 pgno_t me_maxpg; /**< me_mapsize / me_psize */
954 MDB_dbx *me_dbxs; /**< array of static DB info */
955 uint16_t *me_dbflags; /**< array of flags from MDB_db.md_flags */
956 pthread_key_t me_txkey; /**< thread-key for readers */
957 MDB_pgstate me_pgstate; /**< state of old pages from freeDB */
958 # define me_pglast me_pgstate.mf_pglast
959 # define me_pghead me_pgstate.mf_pghead
960 # define me_pgfree me_pgstate.mf_pgfree
961 MDB_page *me_dpages; /**< list of malloc'd blocks for re-use */
962 /** IDL of pages that became unused in a write txn */
964 /** ID2L of pages written during a write txn. Length MDB_IDL_UM_SIZE. */
965 MDB_ID2L me_dirty_list;
966 /** Max number of freelist items that can fit in a single overflow page */
967 unsigned int me_maxfree_1pg;
968 /** Max size of a node on a page */
969 unsigned int me_nodemax;
971 HANDLE me_rmutex; /* Windows mutexes don't reside in shared mem */
973 #elif defined(MDB_USE_POSIX_SEM)
974 sem_t *me_rmutex; /* Shared mutexes are not supported */
979 /** Nested transaction */
980 typedef struct MDB_ntxn {
981 MDB_txn mnt_txn; /* the transaction */
982 MDB_pgstate mnt_pgstate; /* parent transaction's saved freestate */
985 /** max number of pages to commit in one writev() call */
986 #define MDB_COMMIT_PAGES 64
987 #if defined(IOV_MAX) && IOV_MAX < MDB_COMMIT_PAGES
988 #undef MDB_COMMIT_PAGES
989 #define MDB_COMMIT_PAGES IOV_MAX
992 static int mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp);
993 static int mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp);
994 static int mdb_page_touch(MDB_cursor *mc);
996 static int mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **mp, int *lvl);
997 static int mdb_page_search_root(MDB_cursor *mc,
998 MDB_val *key, int modify);
999 #define MDB_PS_MODIFY 1
1000 #define MDB_PS_ROOTONLY 2
1001 static int mdb_page_search(MDB_cursor *mc,
1002 MDB_val *key, int flags);
1003 static int mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst);
1005 #define MDB_SPLIT_REPLACE MDB_APPENDDUP /**< newkey is not new */
1006 static int mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata,
1007 pgno_t newpgno, unsigned int nflags);
1009 static int mdb_env_read_header(MDB_env *env, MDB_meta *meta);
1010 static int mdb_env_pick_meta(const MDB_env *env);
1011 static int mdb_env_write_meta(MDB_txn *txn);
1012 #if !(defined(_WIN32) || defined(MDB_USE_POSIX_SEM)) /* Drop unused excl arg */
1013 # define mdb_env_close0(env, excl) mdb_env_close1(env)
1015 static void mdb_env_close0(MDB_env *env, int excl);
1017 static MDB_node *mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp);
1018 static int mdb_node_add(MDB_cursor *mc, indx_t indx,
1019 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags);
1020 static void mdb_node_del(MDB_page *mp, indx_t indx, int ksize);
1021 static void mdb_node_shrink(MDB_page *mp, indx_t indx);
1022 static int mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst);
1023 static int mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data);
1024 static size_t mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data);
1025 static size_t mdb_branch_size(MDB_env *env, MDB_val *key);
1027 static int mdb_rebalance(MDB_cursor *mc);
1028 static int mdb_update_key(MDB_cursor *mc, MDB_val *key);
1030 static void mdb_cursor_pop(MDB_cursor *mc);
1031 static int mdb_cursor_push(MDB_cursor *mc, MDB_page *mp);
1033 static int mdb_cursor_del0(MDB_cursor *mc, MDB_node *leaf);
1034 static int mdb_cursor_sibling(MDB_cursor *mc, int move_right);
1035 static int mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1036 static int mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1037 static int mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op,
1039 static int mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1040 static int mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1042 static void mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx);
1043 static void mdb_xcursor_init0(MDB_cursor *mc);
1044 static void mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node);
1046 static int mdb_drop0(MDB_cursor *mc, int subs);
1047 static void mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi);
1050 static MDB_cmp_func mdb_cmp_memn, mdb_cmp_memnr, mdb_cmp_int, mdb_cmp_cint, mdb_cmp_long;
1054 static SECURITY_DESCRIPTOR mdb_null_sd;
1055 static SECURITY_ATTRIBUTES mdb_all_sa;
1056 static int mdb_sec_inited;
1059 /** Return the library version info. */
1061 mdb_version(int *major, int *minor, int *patch)
1063 if (major) *major = MDB_VERSION_MAJOR;
1064 if (minor) *minor = MDB_VERSION_MINOR;
1065 if (patch) *patch = MDB_VERSION_PATCH;
1066 return MDB_VERSION_STRING;
1069 /** Table of descriptions for MDB @ref errors */
1070 static char *const mdb_errstr[] = {
1071 "MDB_KEYEXIST: Key/data pair already exists",
1072 "MDB_NOTFOUND: No matching key/data pair found",
1073 "MDB_PAGE_NOTFOUND: Requested page not found",
1074 "MDB_CORRUPTED: Located page was wrong type",
1075 "MDB_PANIC: Update of meta page failed",
1076 "MDB_VERSION_MISMATCH: Database environment version mismatch",
1077 "MDB_INVALID: File is not an MDB file",
1078 "MDB_MAP_FULL: Environment mapsize limit reached",
1079 "MDB_DBS_FULL: Environment maxdbs limit reached",
1080 "MDB_READERS_FULL: Environment maxreaders limit reached",
1081 "MDB_TLS_FULL: Thread-local storage keys full - too many environments open",
1082 "MDB_TXN_FULL: Transaction has too many dirty pages - transaction too big",
1083 "MDB_CURSOR_FULL: Internal error - cursor stack limit reached",
1084 "MDB_PAGE_FULL: Internal error - page has no more space",
1085 "MDB_MAP_RESIZED: Database contents grew beyond environment mapsize",
1086 "MDB_INCOMPATIBLE: Database flags changed or would change",
1087 "MDB_BAD_RSLOT: Invalid reuse of reader locktable slot",
1091 mdb_strerror(int err)
1095 return ("Successful return: 0");
1097 if (err >= MDB_KEYEXIST && err <= MDB_LAST_ERRCODE) {
1098 i = err - MDB_KEYEXIST;
1099 return mdb_errstr[i];
1102 return strerror(err);
1106 /** Display a key in hexadecimal and return the address of the result.
1107 * @param[in] key the key to display
1108 * @param[in] buf the buffer to write into. Should always be #DKBUF.
1109 * @return The key in hexadecimal form.
1112 mdb_dkey(MDB_val *key, char *buf)
1115 unsigned char *c = key->mv_data;
1121 if (key->mv_size > MDB_MAXKEYSIZE)
1122 return "MDB_MAXKEYSIZE";
1123 /* may want to make this a dynamic check: if the key is mostly
1124 * printable characters, print it as-is instead of converting to hex.
1128 for (i=0; i<key->mv_size; i++)
1129 ptr += sprintf(ptr, "%02x", *c++);
1131 sprintf(buf, "%.*s", key->mv_size, key->mv_data);
1136 /** Display all the keys in the page. */
1138 mdb_page_list(MDB_page *mp)
1141 unsigned int i, nkeys, nsize;
1145 nkeys = NUMKEYS(mp);
1146 fprintf(stderr, "Page %zu numkeys %d\n", mp->mp_pgno, nkeys);
1147 for (i=0; i<nkeys; i++) {
1148 node = NODEPTR(mp, i);
1149 key.mv_size = node->mn_ksize;
1150 key.mv_data = node->mn_data;
1151 nsize = NODESIZE + NODEKSZ(node) + sizeof(indx_t);
1152 if (IS_BRANCH(mp)) {
1153 fprintf(stderr, "key %d: page %zu, %s\n", i, NODEPGNO(node),
1156 if (F_ISSET(node->mn_flags, F_BIGDATA))
1157 nsize += sizeof(pgno_t);
1159 nsize += NODEDSZ(node);
1160 fprintf(stderr, "key %d: nsize %d, %s\n", i, nsize, DKEY(&key));
1166 mdb_cursor_chk(MDB_cursor *mc)
1172 if (!mc->mc_snum && !(mc->mc_flags & C_INITIALIZED)) return;
1173 for (i=0; i<mc->mc_top; i++) {
1175 node = NODEPTR(mp, mc->mc_ki[i]);
1176 if (NODEPGNO(node) != mc->mc_pg[i+1]->mp_pgno)
1179 if (mc->mc_ki[i] >= NUMKEYS(mc->mc_pg[i]))
1185 /** Count all the pages in each DB and in the freelist
1186 * and make sure it matches the actual number of pages
1189 static void mdb_audit(MDB_txn *txn)
1193 MDB_ID freecount, count;
1198 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
1199 while ((rc = mdb_cursor_get(&mc, &key, &data, MDB_NEXT)) == 0)
1200 freecount += *(MDB_ID *)data.mv_data;
1203 for (i = 0; i<txn->mt_numdbs; i++) {
1205 mdb_cursor_init(&mc, txn, i, &mx);
1206 if (txn->mt_dbs[i].md_root == P_INVALID)
1208 count += txn->mt_dbs[i].md_branch_pages +
1209 txn->mt_dbs[i].md_leaf_pages +
1210 txn->mt_dbs[i].md_overflow_pages;
1211 if (txn->mt_dbs[i].md_flags & MDB_DUPSORT) {
1212 mdb_page_search(&mc, NULL, 0);
1216 mp = mc.mc_pg[mc.mc_top];
1217 for (j=0; j<NUMKEYS(mp); j++) {
1218 MDB_node *leaf = NODEPTR(mp, j);
1219 if (leaf->mn_flags & F_SUBDATA) {
1221 memcpy(&db, NODEDATA(leaf), sizeof(db));
1222 count += db.md_branch_pages + db.md_leaf_pages +
1223 db.md_overflow_pages;
1227 while (mdb_cursor_sibling(&mc, 1) == 0);
1230 if (freecount + count + 2 /* metapages */ != txn->mt_next_pgno) {
1231 fprintf(stderr, "audit: %lu freecount: %lu count: %lu total: %lu next_pgno: %lu\n",
1232 txn->mt_txnid, freecount, count+2, freecount+count+2, txn->mt_next_pgno);
1238 mdb_cmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1240 return txn->mt_dbxs[dbi].md_cmp(a, b);
1244 mdb_dcmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1246 return txn->mt_dbxs[dbi].md_dcmp(a, b);
1249 /** Allocate a page.
1250 * Re-use old malloc'd pages first for singletons, otherwise just malloc.
1253 mdb_page_malloc(MDB_cursor *mc, unsigned num)
1255 MDB_env *env = mc->mc_txn->mt_env;
1256 MDB_page *ret = env->me_dpages;
1257 size_t sz = env->me_psize;
1260 VGMEMP_ALLOC(env, ret, sz);
1261 VGMEMP_DEFINED(ret, sizeof(ret->mp_next));
1262 env->me_dpages = ret->mp_next;
1268 if ((ret = malloc(sz)) != NULL) {
1269 VGMEMP_ALLOC(env, ret, sz);
1274 /** Free a single page.
1275 * Saves single pages to a list, for future reuse.
1276 * (This is not used for multi-page overflow pages.)
1279 mdb_page_free(MDB_env *env, MDB_page *mp)
1281 mp->mp_next = env->me_dpages;
1282 VGMEMP_FREE(env, mp);
1283 env->me_dpages = mp;
1286 /* Return all dirty pages to dpage list */
1288 mdb_dlist_free(MDB_txn *txn)
1290 MDB_env *env = txn->mt_env;
1291 MDB_ID2L dl = txn->mt_u.dirty_list;
1292 unsigned i, n = dl[0].mid;
1294 for (i = 1; i <= n; i++) {
1295 MDB_page *dp = dl[i].mptr;
1296 if (!IS_OVERFLOW(dp) || dp->mp_pages == 1) {
1297 mdb_page_free(env, dp);
1299 /* large pages just get freed directly */
1300 VGMEMP_FREE(env, dp);
1307 /** Find oldest txnid still referenced. Expects txn->mt_txnid > 0. */
1309 mdb_find_oldest(MDB_txn *txn)
1312 txnid_t mr, oldest = txn->mt_txnid - 1;
1313 MDB_reader *r = txn->mt_env->me_txns->mti_readers;
1314 for (i = txn->mt_env->me_txns->mti_numreaders; --i >= 0; ) {
1324 /** Allocate pages for writing.
1325 * If there are free pages available from older transactions, they
1326 * will be re-used first. Otherwise a new page will be allocated.
1327 * @param[in] mc cursor A cursor handle identifying the transaction and
1328 * database for which we are allocating.
1329 * @param[in] num the number of pages to allocate.
1330 * @param[out] mp Address of the allocated page(s). Requests for multiple pages
1331 * will always be satisfied by a single contiguous chunk of memory.
1332 * @return 0 on success, non-zero on failure.
1335 mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp)
1337 MDB_txn *txn = mc->mc_txn;
1339 pgno_t pgno = P_INVALID;
1341 txnid_t oldest = 0, last;
1346 /* If our dirty list is already full, we can't do anything */
1347 if (txn->mt_dirty_room == 0)
1348 return MDB_TXN_FULL;
1350 /* The free list won't have any content at all until txn 2 has
1351 * committed. The pages freed by txn 2 will be unreferenced
1352 * after txn 3 commits, and so will be safe to re-use in txn 4.
1354 if (txn->mt_txnid > 3) {
1355 if (!txn->mt_env->me_pghead &&
1356 txn->mt_dbs[FREE_DBI].md_root != P_INVALID) {
1357 /* See if there's anything in the free DB */
1363 mdb_cursor_init(&m2, txn, FREE_DBI, NULL);
1364 if (!txn->mt_env->me_pglast) {
1365 mdb_page_search(&m2, NULL, 0);
1366 leaf = NODEPTR(m2.mc_pg[m2.mc_top], 0);
1367 kptr = (txnid_t *)NODEKEY(leaf);
1372 last = txn->mt_env->me_pglast + 1;
1374 key.mv_data = &last;
1375 key.mv_size = sizeof(last);
1376 rc = mdb_cursor_set(&m2, &key, &data, MDB_SET_RANGE, NULL);
1379 last = *(txnid_t *)key.mv_data;
1383 oldest = mdb_find_oldest(txn);
1385 if (oldest > last) {
1386 /* It's usable, grab it.
1390 if (!txn->mt_env->me_pglast) {
1391 mdb_node_read(txn, leaf, &data);
1393 idl = (MDB_ID *) data.mv_data;
1394 /* We might have a zero-length IDL due to freelist growth
1395 * during a prior commit
1398 txn->mt_env->me_pglast = last;
1401 mop = malloc(MDB_IDL_SIZEOF(idl));
1404 txn->mt_env->me_pglast = last;
1405 txn->mt_env->me_pghead = txn->mt_env->me_pgfree = mop;
1406 memcpy(mop, idl, MDB_IDL_SIZEOF(idl));
1411 DPRINTF("IDL read txn %zu root %zu num %zu",
1412 last, txn->mt_dbs[FREE_DBI].md_root, idl[0]);
1413 for (i=0; i<idl[0]; i++) {
1414 DPRINTF("IDL %zu", idl[i+1]);
1421 if (txn->mt_env->me_pghead) {
1422 pgno_t *mop = txn->mt_env->me_pghead;
1425 int retry = 1, readit = 0, n2 = num-1;
1426 unsigned int i, j, k;
1428 /* If current list is too short, must fetch more and coalesce */
1429 if (mop[0] < (unsigned)num)
1432 mdb_cursor_init(&m2, txn, FREE_DBI, NULL);
1434 #ifdef MDB_PARANOID /* Seems like we can ignore this now */
1435 /* If on freelist, don't try to read more. If what we have
1436 * right now isn't enough just use new pages.
1437 * TODO: get all of this working. Many circular dependencies...
1439 if (mc->mc_dbi == FREE_DBI) {
1448 last = txn->mt_env->me_pglast + 1;
1450 /* We haven't hit the readers list yet? */
1452 oldest = mdb_find_oldest(txn);
1455 /* There's nothing we can use on the freelist */
1456 if (oldest - last < 1)
1459 key.mv_data = &last;
1460 key.mv_size = sizeof(last);
1461 rc = mdb_cursor_set(&m2,&key,&data,MDB_SET_RANGE,NULL);
1463 if (rc == MDB_NOTFOUND)
1467 last = *(txnid_t*)key.mv_data;
1470 idl = (MDB_ID *) data.mv_data;
1471 mop2 = malloc(MDB_IDL_SIZEOF(idl) + MDB_IDL_SIZEOF(mop));
1474 /* merge in sorted order */
1475 i = idl[0]; j = mop[0]; mop2[0] = k = i+j;
1477 while (i>0 || j>0) {
1478 if (i && idl[i] < mop[j])
1479 mop2[k--] = idl[i--];
1481 mop2[k--] = mop[j--];
1483 txn->mt_env->me_pglast = last;
1484 free(txn->mt_env->me_pgfree);
1485 txn->mt_env->me_pghead = txn->mt_env->me_pgfree = mop2;
1487 /* Keep trying to read until we have enough */
1488 if (mop[0] < (unsigned)num) {
1493 /* current list has enough pages, but are they contiguous? */
1494 for (i=mop[0]; i>=(unsigned)num; i--) {
1495 if (mop[i-n2] == mop[i] + n2) {
1498 /* move any stragglers down */
1499 for (j=i+num; j<=mop[0]; j++)
1506 /* Stop if we succeeded, or no retries */
1507 if (!retry || pgno != P_INVALID)
1513 /* peel pages off tail, so we only have to truncate the list */
1514 pgno = MDB_IDL_LAST(mop);
1517 if (MDB_IDL_IS_ZERO(mop)) {
1518 free(txn->mt_env->me_pgfree);
1519 txn->mt_env->me_pghead = txn->mt_env->me_pgfree = NULL;
1524 if (pgno == P_INVALID) {
1525 /* DB size is maxed out */
1526 if (txn->mt_next_pgno + num >= txn->mt_env->me_maxpg) {
1527 DPUTS("DB size maxed out");
1528 return MDB_MAP_FULL;
1531 if (txn->mt_env->me_flags & MDB_WRITEMAP) {
1532 if (pgno == P_INVALID) {
1533 pgno = txn->mt_next_pgno;
1534 txn->mt_next_pgno += num;
1536 np = (MDB_page *)(txn->mt_env->me_map + txn->mt_env->me_psize * pgno);
1539 if (!(np = mdb_page_malloc(mc, num)))
1541 if (pgno == P_INVALID) {
1542 np->mp_pgno = txn->mt_next_pgno;
1543 txn->mt_next_pgno += num;
1548 mid.mid = np->mp_pgno;
1550 if (txn->mt_env->me_flags & MDB_WRITEMAP) {
1551 mdb_mid2l_append(txn->mt_u.dirty_list, &mid);
1553 mdb_mid2l_insert(txn->mt_u.dirty_list, &mid);
1555 txn->mt_dirty_room--;
1561 /** Copy a page: avoid copying unused portions of the page.
1562 * @param[in] dst page to copy into
1563 * @param[in] src page to copy from
1564 * @param[in] psize size of a page
1567 mdb_page_copy(MDB_page *dst, MDB_page *src, unsigned int psize)
1569 dst->mp_flags = src->mp_flags | P_DIRTY;
1570 dst->mp_pages = src->mp_pages;
1572 if (IS_LEAF2(src)) {
1573 memcpy(dst->mp_ptrs, src->mp_ptrs, psize - PAGEHDRSZ - SIZELEFT(src));
1575 unsigned int i, nkeys = NUMKEYS(src);
1576 for (i=0; i<nkeys; i++)
1577 dst->mp_ptrs[i] = src->mp_ptrs[i];
1578 memcpy((char *)dst+src->mp_upper, (char *)src+src->mp_upper,
1579 psize - src->mp_upper);
1583 /** Touch a page: make it dirty and re-insert into tree with updated pgno.
1584 * @param[in] mc cursor pointing to the page to be touched
1585 * @return 0 on success, non-zero on failure.
1588 mdb_page_touch(MDB_cursor *mc)
1590 MDB_page *mp = mc->mc_pg[mc->mc_top];
1594 if (!F_ISSET(mp->mp_flags, P_DIRTY)) {
1596 if ((rc = mdb_page_alloc(mc, 1, &np)))
1598 DPRINTF("touched db %u page %zu -> %zu", mc->mc_dbi, mp->mp_pgno, np->mp_pgno);
1599 assert(mp->mp_pgno != np->mp_pgno);
1600 mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
1602 /* If page isn't full, just copy the used portion */
1603 mdb_page_copy(np, mp, mc->mc_txn->mt_env->me_psize);
1606 memcpy(np, mp, mc->mc_txn->mt_env->me_psize);
1608 np->mp_flags |= P_DIRTY;
1613 /* Adjust other cursors pointing to mp */
1614 if (mc->mc_flags & C_SUB) {
1615 MDB_cursor *m2, *m3;
1616 MDB_dbi dbi = mc->mc_dbi-1;
1618 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
1619 if (m2 == mc) continue;
1620 m3 = &m2->mc_xcursor->mx_cursor;
1621 if (m3->mc_snum < mc->mc_snum) continue;
1622 if (m3->mc_pg[mc->mc_top] == mc->mc_pg[mc->mc_top]) {
1623 m3->mc_pg[mc->mc_top] = mp;
1629 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
1630 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
1631 if (m2->mc_pg[mc->mc_top] == mc->mc_pg[mc->mc_top]) {
1632 m2->mc_pg[mc->mc_top] = mp;
1633 if (mc->mc_db->md_flags & MDB_DUPSORT)
1634 m2->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
1638 mc->mc_pg[mc->mc_top] = mp;
1639 /** If this page has a parent, update the parent to point to
1643 SETPGNO(NODEPTR(mc->mc_pg[mc->mc_top-1], mc->mc_ki[mc->mc_top-1]), mp->mp_pgno);
1645 mc->mc_db->md_root = mp->mp_pgno;
1646 } else if (mc->mc_txn->mt_parent && !(mp->mp_flags & P_SUBP)) {
1648 MDB_ID2 mid, *dl = mc->mc_txn->mt_u.dirty_list;
1649 /* If txn has a parent, make sure the page is in our
1653 unsigned x = mdb_mid2l_search(dl, mp->mp_pgno);
1654 if (x <= dl[0].mid && dl[x].mid == mp->mp_pgno) {
1657 mc->mc_pg[mc->mc_top] = np;
1661 assert(dl[0].mid < MDB_IDL_UM_MAX);
1663 np = mdb_page_malloc(mc, 1);
1666 memcpy(np, mp, mc->mc_txn->mt_env->me_psize);
1667 mid.mid = np->mp_pgno;
1669 mdb_mid2l_insert(dl, &mid);
1677 mdb_env_sync(MDB_env *env, int force)
1680 if (force || !F_ISSET(env->me_flags, MDB_NOSYNC)) {
1681 if (env->me_flags & MDB_WRITEMAP) {
1682 int flags = ((env->me_flags & MDB_MAPASYNC) && !force)
1683 ? MS_ASYNC : MS_SYNC;
1684 if (MDB_MSYNC(env->me_map, env->me_mapsize, flags))
1687 else if (flags == MS_SYNC && MDB_FDATASYNC(env->me_fd))
1691 if (MDB_FDATASYNC(env->me_fd))
1698 /** Make shadow copies of all of parent txn's cursors */
1700 mdb_cursor_shadow(MDB_txn *src, MDB_txn *dst)
1702 MDB_cursor *mc, *m2;
1703 unsigned int i, j, size;
1705 for (i=0;i<src->mt_numdbs; i++) {
1706 if (src->mt_cursors[i]) {
1707 size = sizeof(MDB_cursor);
1708 if (src->mt_cursors[i]->mc_xcursor)
1709 size += sizeof(MDB_xcursor);
1710 for (m2 = src->mt_cursors[i]; m2; m2=m2->mc_next) {
1717 mc->mc_db = &dst->mt_dbs[i];
1718 mc->mc_dbx = m2->mc_dbx;
1719 mc->mc_dbflag = &dst->mt_dbflags[i];
1720 mc->mc_snum = m2->mc_snum;
1721 mc->mc_top = m2->mc_top;
1722 mc->mc_flags = m2->mc_flags | (C_SHADOW|C_ALLOCD);
1723 for (j=0; j<mc->mc_snum; j++) {
1724 mc->mc_pg[j] = m2->mc_pg[j];
1725 mc->mc_ki[j] = m2->mc_ki[j];
1727 if (m2->mc_xcursor) {
1728 MDB_xcursor *mx, *mx2;
1729 mx = (MDB_xcursor *)(mc+1);
1730 mc->mc_xcursor = mx;
1731 mx2 = m2->mc_xcursor;
1732 mx->mx_db = mx2->mx_db;
1733 mx->mx_dbx = mx2->mx_dbx;
1734 mx->mx_dbflag = mx2->mx_dbflag;
1735 mx->mx_cursor.mc_txn = dst;
1736 mx->mx_cursor.mc_dbi = mx2->mx_cursor.mc_dbi;
1737 mx->mx_cursor.mc_db = &mx->mx_db;
1738 mx->mx_cursor.mc_dbx = &mx->mx_dbx;
1739 mx->mx_cursor.mc_dbflag = &mx->mx_dbflag;
1740 mx->mx_cursor.mc_snum = mx2->mx_cursor.mc_snum;
1741 mx->mx_cursor.mc_top = mx2->mx_cursor.mc_top;
1742 mx->mx_cursor.mc_flags = mx2->mx_cursor.mc_flags | C_SHADOW;
1743 for (j=0; j<mx2->mx_cursor.mc_snum; j++) {
1744 mx->mx_cursor.mc_pg[j] = mx2->mx_cursor.mc_pg[j];
1745 mx->mx_cursor.mc_ki[j] = mx2->mx_cursor.mc_ki[j];
1748 mc->mc_xcursor = NULL;
1750 mc->mc_next = dst->mt_cursors[i];
1751 dst->mt_cursors[i] = mc;
1758 /** Close this write txn's cursors, after optionally merging its shadow
1759 * cursors back into parent's.
1760 * @param[in] txn the transaction handle.
1761 * @param[in] merge 0 to not merge cursors, C_SHADOW to merge.
1762 * @return 0 on success, non-zero on failure.
1765 mdb_cursors_close(MDB_txn *txn, unsigned merge)
1767 MDB_cursor **cursors = txn->mt_cursors, *mc, *next;
1770 for (i = txn->mt_numdbs; --i >= 0; ) {
1771 for (mc = cursors[i]; mc; mc = next) {
1773 if (mc->mc_flags & merge) {
1774 MDB_cursor *m2 = mc->mc_orig;
1775 m2->mc_snum = mc->mc_snum;
1776 m2->mc_top = mc->mc_top;
1777 for (j = mc->mc_snum; --j >= 0; ) {
1778 m2->mc_pg[j] = mc->mc_pg[j];
1779 m2->mc_ki[j] = mc->mc_ki[j];
1782 if (mc->mc_flags & C_ALLOCD)
1790 mdb_txn_reset0(MDB_txn *txn);
1792 /** Common code for #mdb_txn_begin() and #mdb_txn_renew().
1793 * @param[in] txn the transaction handle to initialize
1794 * @return 0 on success, non-zero on failure.
1797 mdb_txn_renew0(MDB_txn *txn)
1799 MDB_env *env = txn->mt_env;
1802 int rc, new_notls = 0;
1805 txn->mt_numdbs = env->me_numdbs;
1806 txn->mt_dbxs = env->me_dbxs; /* mostly static anyway */
1808 if (txn->mt_flags & MDB_TXN_RDONLY) {
1809 if (!env->me_txns) {
1810 i = mdb_env_pick_meta(env);
1811 txn->mt_txnid = env->me_metas[i]->mm_txnid;
1812 txn->mt_u.reader = NULL;
1814 MDB_reader *r = (env->me_flags & MDB_NOTLS) ? txn->mt_u.reader :
1815 pthread_getspecific(env->me_txkey);
1817 if (r->mr_pid != env->me_pid || r->mr_txnid != (txnid_t)-1)
1818 return MDB_BAD_RSLOT;
1820 pid_t pid = env->me_pid;
1821 pthread_t tid = pthread_self();
1824 for (i=0; i<env->me_txns->mti_numreaders; i++)
1825 if (env->me_txns->mti_readers[i].mr_pid == 0)
1827 if (i == env->me_maxreaders) {
1828 UNLOCK_MUTEX_R(env);
1829 return MDB_READERS_FULL;
1831 env->me_txns->mti_readers[i].mr_pid = pid;
1832 env->me_txns->mti_readers[i].mr_tid = tid;
1833 if (i >= env->me_txns->mti_numreaders)
1834 env->me_txns->mti_numreaders = i+1;
1835 /* Save numreaders for un-mutexed mdb_env_close() */
1836 env->me_numreaders = env->me_txns->mti_numreaders;
1837 UNLOCK_MUTEX_R(env);
1838 r = &env->me_txns->mti_readers[i];
1839 new_notls = (env->me_flags & MDB_NOTLS);
1840 if (!new_notls && (rc=pthread_setspecific(env->me_txkey, r))) {
1845 txn->mt_txnid = r->mr_txnid = env->me_txns->mti_txnid;
1846 txn->mt_u.reader = r;
1848 txn->mt_toggle = txn->mt_txnid & 1;
1849 txn->mt_next_pgno = env->me_metas[txn->mt_toggle]->mm_last_pg+1;
1853 txn->mt_txnid = env->me_txns->mti_txnid;
1854 txn->mt_toggle = txn->mt_txnid & 1;
1855 txn->mt_next_pgno = env->me_metas[txn->mt_toggle]->mm_last_pg+1;
1858 if (txn->mt_txnid == mdb_debug_start)
1861 txn->mt_dirty_room = MDB_IDL_UM_MAX;
1862 txn->mt_u.dirty_list = env->me_dirty_list;
1863 txn->mt_u.dirty_list[0].mid = 0;
1864 txn->mt_free_pgs = env->me_free_pgs;
1865 txn->mt_free_pgs[0] = 0;
1869 /* Copy the DB info and flags */
1870 memcpy(txn->mt_dbs, env->me_metas[txn->mt_toggle]->mm_dbs, 2 * sizeof(MDB_db));
1871 for (i=2; i<txn->mt_numdbs; i++) {
1872 x = env->me_dbflags[i];
1873 txn->mt_dbs[i].md_flags = x & PERSISTENT_FLAGS;
1874 txn->mt_dbflags[i] = (x & MDB_VALID) ? DB_VALID|DB_STALE : 0;
1876 txn->mt_dbflags[0] = txn->mt_dbflags[1] = DB_VALID;
1878 if (env->me_maxpg < txn->mt_next_pgno) {
1879 mdb_txn_reset0(txn);
1881 txn->mt_u.reader->mr_pid = 0;
1882 txn->mt_u.reader = NULL;
1884 return MDB_MAP_RESIZED;
1891 mdb_txn_renew(MDB_txn *txn)
1895 if (!txn || txn->mt_dbxs) /* A reset txn has mt_dbxs==NULL */
1898 if (txn->mt_env->me_flags & MDB_FATAL_ERROR) {
1899 DPUTS("environment had fatal error, must shutdown!");
1903 rc = mdb_txn_renew0(txn);
1904 if (rc == MDB_SUCCESS) {
1905 DPRINTF("renew txn %zu%c %p on mdbenv %p, root page %zu",
1906 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
1907 (void *)txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root);
1913 mdb_txn_begin(MDB_env *env, MDB_txn *parent, unsigned int flags, MDB_txn **ret)
1917 int rc, size, tsize = sizeof(MDB_txn);
1919 if (env->me_flags & MDB_FATAL_ERROR) {
1920 DPUTS("environment had fatal error, must shutdown!");
1923 if ((env->me_flags & MDB_RDONLY) && !(flags & MDB_RDONLY))
1926 /* Nested transactions: Max 1 child, write txns only, no writemap */
1927 if (parent->mt_child ||
1928 (flags & MDB_RDONLY) || (parent->mt_flags & MDB_TXN_RDONLY) ||
1929 (env->me_flags & MDB_WRITEMAP))
1933 tsize = sizeof(MDB_ntxn);
1935 size = tsize + env->me_maxdbs * (sizeof(MDB_db)+1);
1936 if (!(flags & MDB_RDONLY))
1937 size += env->me_maxdbs * sizeof(MDB_cursor *);
1939 if ((txn = calloc(1, size)) == NULL) {
1940 DPRINTF("calloc: %s", strerror(ErrCode()));
1943 txn->mt_dbs = (MDB_db *) ((char *)txn + tsize);
1944 if (flags & MDB_RDONLY) {
1945 txn->mt_flags |= MDB_TXN_RDONLY;
1946 txn->mt_dbflags = (unsigned char *)(txn->mt_dbs + env->me_maxdbs);
1948 txn->mt_cursors = (MDB_cursor **)(txn->mt_dbs + env->me_maxdbs);
1949 txn->mt_dbflags = (unsigned char *)(txn->mt_cursors + env->me_maxdbs);
1955 txn->mt_u.dirty_list = malloc(sizeof(MDB_ID2)*MDB_IDL_UM_SIZE);
1956 if (!txn->mt_u.dirty_list ||
1957 !(txn->mt_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)))
1959 free(txn->mt_u.dirty_list);
1963 txn->mt_txnid = parent->mt_txnid;
1964 txn->mt_toggle = parent->mt_toggle;
1965 txn->mt_dirty_room = parent->mt_dirty_room;
1966 txn->mt_u.dirty_list[0].mid = 0;
1967 txn->mt_free_pgs[0] = 0;
1968 txn->mt_next_pgno = parent->mt_next_pgno;
1969 parent->mt_child = txn;
1970 txn->mt_parent = parent;
1971 txn->mt_numdbs = parent->mt_numdbs;
1972 txn->mt_dbxs = parent->mt_dbxs;
1973 memcpy(txn->mt_dbs, parent->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
1974 /* Copy parent's mt_dbflags, but clear DB_NEW */
1975 for (i=0; i<txn->mt_numdbs; i++)
1976 txn->mt_dbflags[i] = parent->mt_dbflags[i] & ~DB_NEW;
1978 ntxn = (MDB_ntxn *)txn;
1979 ntxn->mnt_pgstate = env->me_pgstate; /* save parent me_pghead & co */
1980 if (env->me_pghead) {
1981 size = MDB_IDL_SIZEOF(env->me_pghead);
1982 env->me_pghead = malloc(size);
1984 memcpy(env->me_pghead, ntxn->mnt_pgstate.mf_pghead, size);
1988 env->me_pgfree = env->me_pghead;
1990 rc = mdb_cursor_shadow(parent, txn);
1992 mdb_txn_reset0(txn);
1994 rc = mdb_txn_renew0(txn);
2000 DPRINTF("begin txn %zu%c %p on mdbenv %p, root page %zu",
2001 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2002 (void *) txn, (void *) env, txn->mt_dbs[MAIN_DBI].md_root);
2008 /** Export or close DBI handles opened in this txn. */
2010 mdb_dbis_update(MDB_txn *txn, int keep)
2013 MDB_dbi n = txn->mt_numdbs;
2014 MDB_env *env = txn->mt_env;
2015 unsigned char *tdbflags = txn->mt_dbflags;
2017 for (i = n; --i >= 2;) {
2018 if (tdbflags[i] & DB_NEW) {
2020 env->me_dbflags[i] = txn->mt_dbs[i].md_flags | MDB_VALID;
2022 char *ptr = env->me_dbxs[i].md_name.mv_data;
2023 env->me_dbxs[i].md_name.mv_data = NULL;
2024 env->me_dbxs[i].md_name.mv_size = 0;
2025 env->me_dbflags[i] = 0;
2030 if (keep && env->me_numdbs < n)
2034 /** Common code for #mdb_txn_reset() and #mdb_txn_abort().
2035 * May be called twice for readonly txns: First reset it, then abort.
2036 * @param[in] txn the transaction handle to reset
2039 mdb_txn_reset0(MDB_txn *txn)
2041 MDB_env *env = txn->mt_env;
2043 /* Close any DBI handles opened in this txn */
2044 mdb_dbis_update(txn, 0);
2046 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
2047 if (txn->mt_u.reader) {
2048 txn->mt_u.reader->mr_txnid = (txnid_t)-1;
2049 if (!(env->me_flags & MDB_NOTLS))
2050 txn->mt_u.reader = NULL; /* txn does not own reader */
2052 txn->mt_numdbs = 0; /* close nothing if called again */
2053 txn->mt_dbxs = NULL; /* mark txn as reset */
2055 mdb_cursors_close(txn, 0);
2057 if (!(env->me_flags & MDB_WRITEMAP)) {
2058 mdb_dlist_free(txn);
2060 free(env->me_pgfree);
2062 if (txn->mt_parent) {
2063 txn->mt_parent->mt_child = NULL;
2064 env->me_pgstate = ((MDB_ntxn *)txn)->mnt_pgstate;
2065 mdb_midl_free(txn->mt_free_pgs);
2066 free(txn->mt_u.dirty_list);
2069 if (mdb_midl_shrink(&txn->mt_free_pgs))
2070 env->me_free_pgs = txn->mt_free_pgs;
2073 txn->mt_env->me_pghead = txn->mt_env->me_pgfree = NULL;
2074 txn->mt_env->me_pglast = 0;
2077 /* The writer mutex was locked in mdb_txn_begin. */
2078 UNLOCK_MUTEX_W(env);
2083 mdb_txn_reset(MDB_txn *txn)
2088 DPRINTF("reset txn %zu%c %p on mdbenv %p, root page %zu",
2089 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2090 (void *) txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root);
2092 /* This call is only valid for read-only txns */
2093 if (!(txn->mt_flags & MDB_TXN_RDONLY))
2096 mdb_txn_reset0(txn);
2100 mdb_txn_abort(MDB_txn *txn)
2105 DPRINTF("abort txn %zu%c %p on mdbenv %p, root page %zu",
2106 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2107 (void *)txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root);
2110 mdb_txn_abort(txn->mt_child);
2112 mdb_txn_reset0(txn);
2113 /* Free reader slot tied to this txn (if MDB_NOTLS && writable FS) */
2114 if ((txn->mt_flags & MDB_TXN_RDONLY) && txn->mt_u.reader)
2115 txn->mt_u.reader->mr_pid = 0;
2121 mdb_txn_commit(MDB_txn *txn)
2129 pgno_t next, freecnt;
2130 txnid_t oldpg_txnid, id;
2133 assert(txn != NULL);
2134 assert(txn->mt_env != NULL);
2136 if (txn->mt_child) {
2137 rc = mdb_txn_commit(txn->mt_child);
2138 txn->mt_child = NULL;
2147 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
2148 mdb_dbis_update(txn, 1);
2149 txn->mt_numdbs = 2; /* so txn_abort() doesn't close any new handles */
2154 if (F_ISSET(txn->mt_flags, MDB_TXN_ERROR)) {
2155 DPUTS("error flag is set, can't commit");
2157 txn->mt_parent->mt_flags |= MDB_TXN_ERROR;
2162 if (txn->mt_parent) {
2163 MDB_txn *parent = txn->mt_parent;
2167 /* Append our free list to parent's */
2168 if (mdb_midl_append_list(&parent->mt_free_pgs, txn->mt_free_pgs)) {
2172 mdb_midl_free(txn->mt_free_pgs);
2174 parent->mt_next_pgno = txn->mt_next_pgno;
2175 parent->mt_flags = txn->mt_flags;
2177 /* Merge our cursors into parent's and close them */
2178 mdb_cursors_close(txn, C_SHADOW);
2180 /* Update parent's DB table. */
2181 memcpy(parent->mt_dbs, txn->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
2182 txn->mt_parent->mt_numdbs = txn->mt_numdbs;
2183 txn->mt_parent->mt_dbflags[0] = txn->mt_dbflags[0];
2184 txn->mt_parent->mt_dbflags[1] = txn->mt_dbflags[1];
2185 for (i=2; i<txn->mt_numdbs; i++) {
2186 /* preserve parent's DB_NEW status */
2187 x = txn->mt_parent->mt_dbflags[i] & DB_NEW;
2188 txn->mt_parent->mt_dbflags[i] = txn->mt_dbflags[i] | x;
2191 dst = txn->mt_parent->mt_u.dirty_list;
2192 src = txn->mt_u.dirty_list;
2193 /* Find len = length of merging our dirty list with parent's */
2195 dst[0].mid = 0; /* simplify loops */
2196 if (parent->mt_parent) {
2197 len = x + src[0].mid;
2198 y = mdb_mid2l_search(src, dst[x].mid + 1) - 1;
2199 for (i = x; y && i; y--) {
2200 pgno_t yp = src[y].mid;
2201 while (yp < dst[i].mid)
2203 if (yp == dst[i].mid) {
2208 } else { /* Simplify the above for single-ancestor case */
2209 len = MDB_IDL_UM_MAX - txn->mt_dirty_room;
2211 /* Merge our dirty list with parent's */
2213 for (i = len; y; dst[i--] = src[y--]) {
2214 pgno_t yp = src[y].mid;
2215 while (yp < dst[x].mid)
2216 dst[i--] = dst[x--];
2217 if (yp == dst[x].mid)
2218 free(dst[x--].mptr);
2222 free(txn->mt_u.dirty_list);
2223 parent->mt_dirty_room = txn->mt_dirty_room;
2225 txn->mt_parent->mt_child = NULL;
2226 free(((MDB_ntxn *)txn)->mnt_pgstate.mf_pgfree);
2231 if (txn != env->me_txn) {
2232 DPUTS("attempt to commit unknown transaction");
2237 mdb_cursors_close(txn, 0);
2239 if (!txn->mt_u.dirty_list[0].mid && !(txn->mt_flags & MDB_TXN_DIRTY))
2242 DPRINTF("committing txn %zu %p on mdbenv %p, root page %zu",
2243 txn->mt_txnid, (void *)txn, (void *)env, txn->mt_dbs[MAIN_DBI].md_root);
2245 /* Update DB root pointers */
2246 if (txn->mt_numdbs > 2) {
2249 data.mv_size = sizeof(MDB_db);
2251 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
2252 for (i = 2; i < txn->mt_numdbs; i++) {
2253 if (txn->mt_dbflags[i] & DB_DIRTY) {
2254 data.mv_data = &txn->mt_dbs[i];
2255 rc = mdb_cursor_put(&mc, &txn->mt_dbxs[i].md_name, &data, 0);
2262 /* Save the freelist as of this transaction to the freeDB. This
2263 * can change the freelist, so keep trying until it stabilizes.
2265 * env->me_pglast and the length of txn->mt_free_pgs cannot decrease,
2266 * except the code below can decrease env->me_pglast to split pghead.
2267 * Page numbers cannot disappear from txn->mt_free_pgs. New pages
2268 * can only appear in env->me_pghead when env->me_pglast increases.
2269 * Until then, the me_pghead pointer won't move but can become NULL.
2272 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
2273 oldpg_txnid = id = 0;
2276 /* should only be one record now */
2277 if (env->me_pghead || env->me_pglast) {
2278 /* make sure first page of freeDB is touched and on freelist */
2279 rc = mdb_page_search(&mc, NULL, MDB_PS_MODIFY);
2280 if (rc && rc != MDB_NOTFOUND) {
2287 /* Delete IDLs we used from the free list */
2288 if (env->me_pglast) {
2293 rc = mdb_cursor_first(&mc, &key, NULL);
2296 oldpg_txnid = *(txnid_t *)key.mv_data;
2298 assert(oldpg_txnid <= env->me_pglast);
2300 rc = mdb_cursor_del(&mc, 0);
2303 } while (oldpg_txnid < env->me_pglast);
2306 /* Save IDL of pages freed by this txn, to freeDB */
2308 if (freecnt != txn->mt_free_pgs[0]) {
2311 /* make sure last page of freeDB is touched and on freelist */
2312 key.mv_size = MDB_MAXKEYSIZE+1;
2314 rc = mdb_page_search(&mc, &key, MDB_PS_MODIFY);
2315 if (rc && rc != MDB_NOTFOUND)
2321 MDB_IDL idl = txn->mt_free_pgs;
2322 mdb_midl_sort(txn->mt_free_pgs);
2323 DPRINTF("IDL write txn %zu root %zu num %zu",
2324 txn->mt_txnid, txn->mt_dbs[FREE_DBI].md_root, idl[0]);
2325 for (i=1; i<=idl[0]; i++) {
2326 DPRINTF("IDL %zu", idl[i]);
2330 /* write to last page of freeDB */
2331 key.mv_size = sizeof(pgno_t);
2332 key.mv_data = &txn->mt_txnid;
2333 /* The free list can still grow during this call,
2334 * despite the pre-emptive touches above. So retry
2335 * until the reserved space remains big enough.
2338 assert(freecnt < txn->mt_free_pgs[0]);
2339 freecnt = txn->mt_free_pgs[0];
2340 data.mv_size = MDB_IDL_SIZEOF(txn->mt_free_pgs);
2341 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
2344 } while (freecnt != txn->mt_free_pgs[0]);
2345 mdb_midl_sort(txn->mt_free_pgs);
2346 memcpy(data.mv_data, txn->mt_free_pgs, data.mv_size);
2347 if (oldpg_txnid < env->me_pglast || (!env->me_pghead && id))
2348 goto free_pgfirst; /* used up freeDB[oldpg_txnid] */
2351 /* Put back page numbers we took from freeDB but did not use */
2352 if (env->me_pghead) {
2357 mop = env->me_pghead;
2358 id = env->me_pglast;
2359 key.mv_size = sizeof(id);
2361 /* These steps may grow the freelist again
2362 * due to freed overflow pages...
2367 if (orig > env->me_maxfree_1pg && id > 4)
2368 orig = env->me_maxfree_1pg; /* Do not use more than 1 page */
2369 data.mv_size = (orig + 1) * sizeof(pgno_t);
2370 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
2373 assert(!env->me_pghead || env->me_pglast);
2374 /* mop could have been used again here */
2375 if (id != env->me_pglast || env->me_pghead == NULL)
2376 goto again; /* was completely used up */
2377 assert(mop == env->me_pghead);
2378 } while (mop[0] < orig && --i);
2379 memcpy(data.mv_data, mop, data.mv_size);
2382 *(pgno_t *)data.mv_data = orig;
2383 mop[orig] = mop[0] - orig;
2384 env->me_pghead = mop += orig;
2385 /* Save more oldpages at the previous txnid. */
2386 assert(env->me_pglast == id && id == oldpg_txnid);
2387 env->me_pglast = --oldpg_txnid;
2391 /* Check for growth of freelist again */
2392 if (freecnt != txn->mt_free_pgs[0])
2395 free(env->me_pgfree);
2396 env->me_pghead = env->me_pgfree = NULL;
2398 if (!MDB_IDL_IS_ZERO(txn->mt_free_pgs)) {
2399 if (mdb_midl_shrink(&txn->mt_free_pgs))
2400 env->me_free_pgs = txn->mt_free_pgs;
2407 if (env->me_flags & MDB_WRITEMAP) {
2408 for (i=1; i<=txn->mt_u.dirty_list[0].mid; i++) {
2409 dp = txn->mt_u.dirty_list[i].mptr;
2410 /* clear dirty flag */
2411 dp->mp_flags &= ~P_DIRTY;
2413 txn->mt_u.dirty_list[0].mid = 0;
2417 /* Commit up to MDB_COMMIT_PAGES dirty pages to disk until done.
2423 /* Windows actually supports scatter/gather I/O, but only on
2424 * unbuffered file handles. Since we're relying on the OS page
2425 * cache for all our data, that's self-defeating. So we just
2426 * write pages one at a time. We use the ov structure to set
2427 * the write offset, to at least save the overhead of a Seek
2431 memset(&ov, 0, sizeof(ov));
2432 for (; i<=txn->mt_u.dirty_list[0].mid; i++) {
2434 dp = txn->mt_u.dirty_list[i].mptr;
2435 DPRINTF("committing page %zu", dp->mp_pgno);
2436 size = dp->mp_pgno * env->me_psize;
2437 ov.Offset = size & 0xffffffff;
2438 ov.OffsetHigh = size >> 16;
2439 ov.OffsetHigh >>= 16;
2440 /* clear dirty flag */
2441 dp->mp_flags &= ~P_DIRTY;
2442 wsize = env->me_psize;
2443 if (IS_OVERFLOW(dp)) wsize *= dp->mp_pages;
2444 rc = WriteFile(env->me_fd, dp, wsize, NULL, &ov);
2447 DPRINTF("WriteFile: %d", n);
2454 struct iovec iov[MDB_COMMIT_PAGES];
2458 for (; i<=txn->mt_u.dirty_list[0].mid; i++) {
2459 dp = txn->mt_u.dirty_list[i].mptr;
2460 if (dp->mp_pgno != next) {
2462 rc = writev(env->me_fd, iov, n);
2466 DPUTS("short write, filesystem full?");
2468 DPRINTF("writev: %s", strerror(n));
2475 lseek(env->me_fd, dp->mp_pgno * env->me_psize, SEEK_SET);
2478 DPRINTF("committing page %zu", dp->mp_pgno);
2479 iov[n].iov_len = env->me_psize;
2480 if (IS_OVERFLOW(dp)) iov[n].iov_len *= dp->mp_pages;
2481 iov[n].iov_base = (char *)dp;
2482 size += iov[n].iov_len;
2483 next = dp->mp_pgno + (IS_OVERFLOW(dp) ? dp->mp_pages : 1);
2484 /* clear dirty flag */
2485 dp->mp_flags &= ~P_DIRTY;
2486 if (++n >= MDB_COMMIT_PAGES) {
2496 rc = writev(env->me_fd, iov, n);
2500 DPUTS("short write, filesystem full?");
2502 DPRINTF("writev: %s", strerror(n));
2509 mdb_dlist_free(txn);
2512 if ((n = mdb_env_sync(env, 0)) != 0 ||
2513 (n = mdb_env_write_meta(txn)) != MDB_SUCCESS) {
2521 mdb_dbis_update(txn, 1);
2523 UNLOCK_MUTEX_W(env);
2529 /** Read the environment parameters of a DB environment before
2530 * mapping it into memory.
2531 * @param[in] env the environment handle
2532 * @param[out] meta address of where to store the meta information
2533 * @return 0 on success, non-zero on failure.
2536 mdb_env_read_header(MDB_env *env, MDB_meta *meta)
2543 /* We don't know the page size yet, so use a minimum value.
2544 * Read both meta pages so we can use the latest one.
2547 for (i=0; i<2; i++) {
2549 if (!ReadFile(env->me_fd, &pbuf, MDB_PAGESIZE, (DWORD *)&rc, NULL) || rc == 0)
2551 if ((rc = read(env->me_fd, &pbuf, MDB_PAGESIZE)) == 0)
2556 else if (rc != MDB_PAGESIZE) {
2560 DPRINTF("read: %s", strerror(err));
2564 p = (MDB_page *)&pbuf;
2566 if (!F_ISSET(p->mp_flags, P_META)) {
2567 DPRINTF("page %zu not a meta page", p->mp_pgno);
2572 if (m->mm_magic != MDB_MAGIC) {
2573 DPUTS("meta has invalid magic");
2577 if (m->mm_version != MDB_VERSION) {
2578 DPRINTF("database is version %u, expected version %u",
2579 m->mm_version, MDB_VERSION);
2580 return MDB_VERSION_MISMATCH;
2584 if (m->mm_txnid > meta->mm_txnid)
2585 memcpy(meta, m, sizeof(*m));
2587 memcpy(meta, m, sizeof(*m));
2589 if (SetFilePointer(env->me_fd, meta->mm_psize, NULL, FILE_BEGIN) != meta->mm_psize)
2591 if (lseek(env->me_fd, meta->mm_psize, SEEK_SET) != meta->mm_psize)
2599 /** Write the environment parameters of a freshly created DB environment.
2600 * @param[in] env the environment handle
2601 * @param[out] meta address of where to store the meta information
2602 * @return 0 on success, non-zero on failure.
2605 mdb_env_init_meta(MDB_env *env, MDB_meta *meta)
2612 DPUTS("writing new meta page");
2614 GET_PAGESIZE(psize);
2616 meta->mm_magic = MDB_MAGIC;
2617 meta->mm_version = MDB_VERSION;
2618 meta->mm_mapsize = env->me_mapsize;
2619 meta->mm_psize = psize;
2620 meta->mm_last_pg = 1;
2621 meta->mm_flags = env->me_flags & 0xffff;
2622 meta->mm_flags |= MDB_INTEGERKEY;
2623 meta->mm_dbs[0].md_root = P_INVALID;
2624 meta->mm_dbs[1].md_root = P_INVALID;
2626 p = calloc(2, psize);
2628 p->mp_flags = P_META;
2631 memcpy(m, meta, sizeof(*meta));
2633 q = (MDB_page *)((char *)p + psize);
2636 q->mp_flags = P_META;
2639 memcpy(m, meta, sizeof(*meta));
2644 SetFilePointer(env->me_fd, 0, NULL, FILE_BEGIN);
2645 rc = WriteFile(env->me_fd, p, psize * 2, &len, NULL);
2646 rc = (len == psize * 2) ? MDB_SUCCESS : ErrCode();
2649 lseek(env->me_fd, 0, SEEK_SET);
2650 rc = write(env->me_fd, p, psize * 2);
2651 rc = (rc == (int)psize * 2) ? MDB_SUCCESS : ErrCode();
2657 /** Update the environment info to commit a transaction.
2658 * @param[in] txn the transaction that's being committed
2659 * @return 0 on success, non-zero on failure.
2662 mdb_env_write_meta(MDB_txn *txn)
2665 MDB_meta meta, metab, *mp;
2667 int rc, len, toggle;
2674 assert(txn != NULL);
2675 assert(txn->mt_env != NULL);
2677 toggle = !txn->mt_toggle;
2678 DPRINTF("writing meta page %d for root page %zu",
2679 toggle, txn->mt_dbs[MAIN_DBI].md_root);
2682 mp = env->me_metas[toggle];
2684 if (env->me_flags & MDB_WRITEMAP) {
2685 /* Persist any increases of mapsize config */
2686 if (env->me_mapsize > mp->mm_mapsize)
2687 mp->mm_mapsize = env->me_mapsize;
2688 mp->mm_dbs[0] = txn->mt_dbs[0];
2689 mp->mm_dbs[1] = txn->mt_dbs[1];
2690 mp->mm_last_pg = txn->mt_next_pgno - 1;
2691 mp->mm_txnid = txn->mt_txnid;
2692 if (!(env->me_flags & (MDB_NOMETASYNC|MDB_NOSYNC))) {
2693 rc = (env->me_flags & MDB_MAPASYNC) ? MS_ASYNC : MS_SYNC;
2696 ptr += env->me_psize;
2697 if (MDB_MSYNC(ptr, env->me_psize, rc)) {
2704 metab.mm_txnid = env->me_metas[toggle]->mm_txnid;
2705 metab.mm_last_pg = env->me_metas[toggle]->mm_last_pg;
2707 ptr = (char *)&meta;
2708 if (env->me_mapsize > mp->mm_mapsize) {
2709 /* Persist any increases of mapsize config */
2710 meta.mm_mapsize = env->me_mapsize;
2711 off = offsetof(MDB_meta, mm_mapsize);
2713 off = offsetof(MDB_meta, mm_dbs[0].md_depth);
2715 len = sizeof(MDB_meta) - off;
2718 meta.mm_dbs[0] = txn->mt_dbs[0];
2719 meta.mm_dbs[1] = txn->mt_dbs[1];
2720 meta.mm_last_pg = txn->mt_next_pgno - 1;
2721 meta.mm_txnid = txn->mt_txnid;
2724 off += env->me_psize;
2727 /* Write to the SYNC fd */
2728 mfd = env->me_flags & (MDB_NOSYNC|MDB_NOMETASYNC) ?
2729 env->me_fd : env->me_mfd;
2732 memset(&ov, 0, sizeof(ov));
2734 WriteFile(mfd, ptr, len, (DWORD *)&rc, &ov);
2737 rc = pwrite(mfd, ptr, len, off);
2742 DPUTS("write failed, disk error?");
2743 /* On a failure, the pagecache still contains the new data.
2744 * Write some old data back, to prevent it from being used.
2745 * Use the non-SYNC fd; we know it will fail anyway.
2747 meta.mm_last_pg = metab.mm_last_pg;
2748 meta.mm_txnid = metab.mm_txnid;
2750 WriteFile(env->me_fd, ptr, len, NULL, &ov);
2752 r2 = pwrite(env->me_fd, ptr, len, off);
2755 env->me_flags |= MDB_FATAL_ERROR;
2759 /* Memory ordering issues are irrelevant; since the entire writer
2760 * is wrapped by wmutex, all of these changes will become visible
2761 * after the wmutex is unlocked. Since the DB is multi-version,
2762 * readers will get consistent data regardless of how fresh or
2763 * how stale their view of these values is.
2765 txn->mt_env->me_txns->mti_txnid = txn->mt_txnid;
2770 /** Check both meta pages to see which one is newer.
2771 * @param[in] env the environment handle
2772 * @return meta toggle (0 or 1).
2775 mdb_env_pick_meta(const MDB_env *env)
2777 return (env->me_metas[0]->mm_txnid < env->me_metas[1]->mm_txnid);
2781 mdb_env_create(MDB_env **env)
2785 e = calloc(1, sizeof(MDB_env));
2789 e->me_maxreaders = DEFAULT_READERS;
2790 e->me_maxdbs = e->me_numdbs = 2;
2791 e->me_fd = INVALID_HANDLE_VALUE;
2792 e->me_lfd = INVALID_HANDLE_VALUE;
2793 e->me_mfd = INVALID_HANDLE_VALUE;
2794 #ifdef MDB_USE_POSIX_SEM
2795 e->me_rmutex = SEM_FAILED;
2796 e->me_wmutex = SEM_FAILED;
2798 e->me_pid = getpid();
2799 VGMEMP_CREATE(e,0,0);
2805 mdb_env_set_mapsize(MDB_env *env, size_t size)
2809 env->me_mapsize = size;
2811 env->me_maxpg = env->me_mapsize / env->me_psize;
2816 mdb_env_set_maxdbs(MDB_env *env, MDB_dbi dbs)
2820 env->me_maxdbs = dbs + 2; /* Named databases + main and free DB */
2825 mdb_env_set_maxreaders(MDB_env *env, unsigned int readers)
2827 if (env->me_map || readers < 1)
2829 env->me_maxreaders = readers;
2834 mdb_env_get_maxreaders(MDB_env *env, unsigned int *readers)
2836 if (!env || !readers)
2838 *readers = env->me_maxreaders;
2842 /** Further setup required for opening an MDB environment
2845 mdb_env_open2(MDB_env *env)
2847 unsigned int flags = env->me_flags;
2848 int i, newenv = 0, prot;
2852 memset(&meta, 0, sizeof(meta));
2854 if ((i = mdb_env_read_header(env, &meta)) != 0) {
2857 DPUTS("new mdbenv");
2861 /* Was a mapsize configured? */
2862 if (!env->me_mapsize) {
2863 /* If this is a new environment, take the default,
2864 * else use the size recorded in the existing env.
2866 env->me_mapsize = newenv ? DEFAULT_MAPSIZE : meta.mm_mapsize;
2867 } else if (env->me_mapsize < meta.mm_mapsize) {
2868 /* If the configured size is smaller, make sure it's
2869 * still big enough. Silently round up to minimum if not.
2871 size_t minsize = (meta.mm_last_pg + 1) * meta.mm_psize;
2872 if (env->me_mapsize < minsize)
2873 env->me_mapsize = minsize;
2879 LONG sizelo, sizehi;
2880 sizelo = env->me_mapsize & 0xffffffff;
2881 sizehi = env->me_mapsize >> 16; /* pointless on WIN32, only needed on W64 */
2883 /* Windows won't create mappings for zero length files.
2884 * Just allocate the maxsize right now.
2887 SetFilePointer(env->me_fd, sizelo, sizehi ? &sizehi : NULL, 0);
2888 if (!SetEndOfFile(env->me_fd))
2890 SetFilePointer(env->me_fd, 0, NULL, 0);
2892 mh = CreateFileMapping(env->me_fd, NULL, flags & MDB_WRITEMAP ?
2893 PAGE_READWRITE : PAGE_READONLY,
2894 sizehi, sizelo, NULL);
2897 env->me_map = MapViewOfFileEx(mh, flags & MDB_WRITEMAP ?
2898 FILE_MAP_WRITE : FILE_MAP_READ,
2899 0, 0, env->me_mapsize, meta.mm_address);
2907 if (flags & MDB_WRITEMAP) {
2909 if (ftruncate(env->me_fd, env->me_mapsize) < 0)
2912 env->me_map = mmap(meta.mm_address, env->me_mapsize, prot, i,
2914 if (env->me_map == MAP_FAILED) {
2918 /* Turn off readahead. It's harmful when the DB is larger than RAM. */
2920 madvise(env->me_map, env->me_mapsize, MADV_RANDOM);
2922 #ifdef POSIX_MADV_RANDOM
2923 posix_madvise(env->me_map, env->me_mapsize, POSIX_MADV_RANDOM);
2924 #endif /* POSIX_MADV_RANDOM */
2925 #endif /* MADV_RANDOM */
2929 if (flags & MDB_FIXEDMAP)
2930 meta.mm_address = env->me_map;
2931 i = mdb_env_init_meta(env, &meta);
2932 if (i != MDB_SUCCESS) {
2935 } else if (meta.mm_address && env->me_map != meta.mm_address) {
2936 /* Can happen because the address argument to mmap() is just a
2937 * hint. mmap() can pick another, e.g. if the range is in use.
2938 * The MAP_FIXED flag would prevent that, but then mmap could
2939 * instead unmap existing pages to make room for the new map.
2941 return EBUSY; /* TODO: Make a new MDB_* error code? */
2943 env->me_psize = meta.mm_psize;
2944 env->me_maxfree_1pg = (env->me_psize - PAGEHDRSZ) / sizeof(pgno_t) - 1;
2945 env->me_nodemax = (env->me_psize - PAGEHDRSZ) / MDB_MINKEYS;
2947 env->me_maxpg = env->me_mapsize / env->me_psize;
2949 p = (MDB_page *)env->me_map;
2950 env->me_metas[0] = METADATA(p);
2951 env->me_metas[1] = (MDB_meta *)((char *)env->me_metas[0] + meta.mm_psize);
2955 int toggle = mdb_env_pick_meta(env);
2956 MDB_db *db = &env->me_metas[toggle]->mm_dbs[MAIN_DBI];
2958 DPRINTF("opened database version %u, pagesize %u",
2959 env->me_metas[0]->mm_version, env->me_psize);
2960 DPRINTF("using meta page %d", toggle);
2961 DPRINTF("depth: %u", db->md_depth);
2962 DPRINTF("entries: %zu", db->md_entries);
2963 DPRINTF("branch pages: %zu", db->md_branch_pages);
2964 DPRINTF("leaf pages: %zu", db->md_leaf_pages);
2965 DPRINTF("overflow pages: %zu", db->md_overflow_pages);
2966 DPRINTF("root: %zu", db->md_root);
2974 /** Release a reader thread's slot in the reader lock table.
2975 * This function is called automatically when a thread exits.
2976 * @param[in] ptr This points to the slot in the reader lock table.
2979 mdb_env_reader_dest(void *ptr)
2981 MDB_reader *reader = ptr;
2987 /** Junk for arranging thread-specific callbacks on Windows. This is
2988 * necessarily platform and compiler-specific. Windows supports up
2989 * to 1088 keys. Let's assume nobody opens more than 64 environments
2990 * in a single process, for now. They can override this if needed.
2992 #ifndef MAX_TLS_KEYS
2993 #define MAX_TLS_KEYS 64
2995 static pthread_key_t mdb_tls_keys[MAX_TLS_KEYS];
2996 static int mdb_tls_nkeys;
2998 static void NTAPI mdb_tls_callback(PVOID module, DWORD reason, PVOID ptr)
3002 case DLL_PROCESS_ATTACH: break;
3003 case DLL_THREAD_ATTACH: break;
3004 case DLL_THREAD_DETACH:
3005 for (i=0; i<mdb_tls_nkeys; i++) {
3006 MDB_reader *r = pthread_getspecific(mdb_tls_keys[i]);
3007 mdb_env_reader_dest(r);
3010 case DLL_PROCESS_DETACH: break;
3015 const PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
3017 PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
3021 /* Force some symbol references.
3022 * _tls_used forces the linker to create the TLS directory if not already done
3023 * mdb_tls_cbp prevents whole-program-optimizer from dropping the symbol.
3025 #pragma comment(linker, "/INCLUDE:_tls_used")
3026 #pragma comment(linker, "/INCLUDE:mdb_tls_cbp")
3027 #pragma const_seg(".CRT$XLB")
3028 extern const PIMAGE_TLS_CALLBACK mdb_tls_callback;
3029 const PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
3032 #pragma comment(linker, "/INCLUDE:__tls_used")
3033 #pragma comment(linker, "/INCLUDE:_mdb_tls_cbp")
3034 #pragma data_seg(".CRT$XLB")
3035 PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
3037 #endif /* WIN 32/64 */
3038 #endif /* !__GNUC__ */
3041 /** Downgrade the exclusive lock on the region back to shared */
3043 mdb_env_share_locks(MDB_env *env, int *excl)
3045 int rc = 0, toggle = mdb_env_pick_meta(env);
3047 env->me_txns->mti_txnid = env->me_metas[toggle]->mm_txnid;
3052 /* First acquire a shared lock. The Unlock will
3053 * then release the existing exclusive lock.
3055 memset(&ov, 0, sizeof(ov));
3056 if (!LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
3059 UnlockFile(env->me_lfd, 0, 0, 1, 0);
3065 struct flock lock_info;
3066 /* The shared lock replaces the existing lock */
3067 memset((void *)&lock_info, 0, sizeof(lock_info));
3068 lock_info.l_type = F_RDLCK;
3069 lock_info.l_whence = SEEK_SET;
3070 lock_info.l_start = 0;
3071 lock_info.l_len = 1;
3072 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
3073 (rc = ErrCode()) == EINTR) ;
3074 *excl = rc ? -1 : 0; /* error may mean we lost the lock */
3081 /** Try to get exlusive lock, otherwise shared.
3082 * Maintain *excl = -1: no/unknown lock, 0: shared, 1: exclusive.
3085 mdb_env_excl_lock(MDB_env *env, int *excl)
3089 if (LockFile(env->me_lfd, 0, 0, 1, 0)) {
3093 memset(&ov, 0, sizeof(ov));
3094 if (LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
3101 struct flock lock_info;
3102 memset((void *)&lock_info, 0, sizeof(lock_info));
3103 lock_info.l_type = F_WRLCK;
3104 lock_info.l_whence = SEEK_SET;
3105 lock_info.l_start = 0;
3106 lock_info.l_len = 1;
3107 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
3108 (rc = ErrCode()) == EINTR) ;
3112 # ifdef MDB_USE_POSIX_SEM
3113 if (*excl < 0) /* always true when !MDB_USE_POSIX_SEM */
3116 lock_info.l_type = F_RDLCK;
3117 while ((rc = fcntl(env->me_lfd, F_SETLKW, &lock_info)) &&
3118 (rc = ErrCode()) == EINTR) ;
3126 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
3128 * hash_64 - 64 bit Fowler/Noll/Vo-0 FNV-1a hash code
3130 * @(#) $Revision: 5.1 $
3131 * @(#) $Id: hash_64a.c,v 5.1 2009/06/30 09:01:38 chongo Exp $
3132 * @(#) $Source: /usr/local/src/cmd/fnv/RCS/hash_64a.c,v $
3134 * http://www.isthe.com/chongo/tech/comp/fnv/index.html
3138 * Please do not copyright this code. This code is in the public domain.
3140 * LANDON CURT NOLL DISCLAIMS ALL WARRANTIES WITH REGARD TO THIS SOFTWARE,
3141 * INCLUDING ALL IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS. IN NO
3142 * EVENT SHALL LANDON CURT NOLL BE LIABLE FOR ANY SPECIAL, INDIRECT OR
3143 * CONSEQUENTIAL DAMAGES OR ANY DAMAGES WHATSOEVER RESULTING FROM LOSS OF
3144 * USE, DATA OR PROFITS, WHETHER IN AN ACTION OF CONTRACT, NEGLIGENCE OR
3145 * OTHER TORTIOUS ACTION, ARISING OUT OF OR IN CONNECTION WITH THE USE OR
3146 * PERFORMANCE OF THIS SOFTWARE.
3149 * chongo <Landon Curt Noll> /\oo/\
3150 * http://www.isthe.com/chongo/
3152 * Share and Enjoy! :-)
3155 typedef unsigned long long mdb_hash_t;
3156 #define MDB_HASH_INIT ((mdb_hash_t)0xcbf29ce484222325ULL)
3158 /** perform a 64 bit Fowler/Noll/Vo FNV-1a hash on a buffer
3159 * @param[in] str string to hash
3160 * @param[in] hval initial value for hash
3161 * @return 64 bit hash
3163 * NOTE: To use the recommended 64 bit FNV-1a hash, use MDB_HASH_INIT as the
3164 * hval arg on the first call.
3167 mdb_hash_val(MDB_val *val, mdb_hash_t hval)
3169 unsigned char *s = (unsigned char *)val->mv_data; /* unsigned string */
3170 unsigned char *end = s + val->mv_size;
3172 * FNV-1a hash each octet of the string
3175 /* xor the bottom with the current octet */
3176 hval ^= (mdb_hash_t)*s++;
3178 /* multiply by the 64 bit FNV magic prime mod 2^64 */
3179 hval += (hval << 1) + (hval << 4) + (hval << 5) +
3180 (hval << 7) + (hval << 8) + (hval << 40);
3182 /* return our new hash value */
3186 /** Hash the string and output the hash in hex.
3187 * @param[in] str string to hash
3188 * @param[out] hexbuf an array of 17 chars to hold the hash
3191 mdb_hash_hex(MDB_val *val, char *hexbuf)
3194 mdb_hash_t h = mdb_hash_val(val, MDB_HASH_INIT);
3195 for (i=0; i<8; i++) {
3196 hexbuf += sprintf(hexbuf, "%02x", (unsigned int)h & 0xff);
3202 /** Open and/or initialize the lock region for the environment.
3203 * @param[in] env The MDB environment.
3204 * @param[in] lpath The pathname of the file used for the lock region.
3205 * @param[in] mode The Unix permissions for the file, if we create it.
3206 * @param[out] excl Resulting file lock type: -1 none, 0 shared, 1 exclusive
3207 * @param[in,out] excl In -1, out lock type: -1 none, 0 shared, 1 exclusive
3208 * @return 0 on success, non-zero on failure.
3211 mdb_env_setup_locks(MDB_env *env, char *lpath, int mode, int *excl)
3214 # define MDB_ERRCODE_ROFS ERROR_WRITE_PROTECT
3216 # define MDB_ERRCODE_ROFS EROFS
3217 #ifdef O_CLOEXEC /* Linux: Open file and set FD_CLOEXEC atomically */
3218 # define MDB_CLOEXEC O_CLOEXEC
3221 # define MDB_CLOEXEC 0
3228 env->me_lfd = CreateFile(lpath, GENERIC_READ|GENERIC_WRITE,
3229 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, OPEN_ALWAYS,
3230 FILE_ATTRIBUTE_NORMAL, NULL);
3232 env->me_lfd = open(lpath, O_RDWR|O_CREAT|MDB_CLOEXEC, mode);
3234 if (env->me_lfd == INVALID_HANDLE_VALUE) {
3236 if (rc == MDB_ERRCODE_ROFS && (env->me_flags & MDB_RDONLY)) {
3241 #if ! ((MDB_CLOEXEC) || defined(_WIN32))
3242 /* Lose record locks when exec*() */
3243 if ((fdflags = fcntl(env->me_lfd, F_GETFD) | FD_CLOEXEC) >= 0)
3244 fcntl(env->me_lfd, F_SETFD, fdflags);
3247 if (!(env->me_flags & MDB_NOTLS)) {
3248 rc = pthread_key_create(&env->me_txkey, mdb_env_reader_dest);
3251 env->me_flags |= MDB_ENV_TXKEY;
3253 /* Windows TLS callbacks need help finding their TLS info. */
3254 if (mdb_tls_nkeys >= MAX_TLS_KEYS) {
3258 mdb_tls_keys[mdb_tls_nkeys++] = env->me_txkey;
3262 /* Try to get exclusive lock. If we succeed, then
3263 * nobody is using the lock region and we should initialize it.
3265 if ((rc = mdb_env_excl_lock(env, excl))) goto fail;
3268 size = GetFileSize(env->me_lfd, NULL);
3270 size = lseek(env->me_lfd, 0, SEEK_END);
3272 rsize = (env->me_maxreaders-1) * sizeof(MDB_reader) + sizeof(MDB_txninfo);
3273 if (size < rsize && *excl > 0) {
3275 SetFilePointer(env->me_lfd, rsize, NULL, 0);
3276 if (!SetEndOfFile(env->me_lfd)) goto fail_errno;
3278 if (ftruncate(env->me_lfd, rsize) != 0) goto fail_errno;
3282 size = rsize - sizeof(MDB_txninfo);
3283 env->me_maxreaders = size/sizeof(MDB_reader) + 1;
3288 mh = CreateFileMapping(env->me_lfd, NULL, PAGE_READWRITE,
3290 if (!mh) goto fail_errno;
3291 env->me_txns = MapViewOfFileEx(mh, FILE_MAP_WRITE, 0, 0, rsize, NULL);
3293 if (!env->me_txns) goto fail_errno;
3295 void *m = mmap(NULL, rsize, PROT_READ|PROT_WRITE, MAP_SHARED,
3297 if (m == MAP_FAILED) goto fail_errno;
3303 BY_HANDLE_FILE_INFORMATION stbuf;
3312 if (!mdb_sec_inited) {
3313 InitializeSecurityDescriptor(&mdb_null_sd,
3314 SECURITY_DESCRIPTOR_REVISION);
3315 SetSecurityDescriptorDacl(&mdb_null_sd, TRUE, 0, FALSE);
3316 mdb_all_sa.nLength = sizeof(SECURITY_ATTRIBUTES);
3317 mdb_all_sa.bInheritHandle = FALSE;
3318 mdb_all_sa.lpSecurityDescriptor = &mdb_null_sd;
3321 if (!GetFileInformationByHandle(env->me_lfd, &stbuf)) goto fail_errno;
3322 idbuf.volume = stbuf.dwVolumeSerialNumber;
3323 idbuf.nhigh = stbuf.nFileIndexHigh;
3324 idbuf.nlow = stbuf.nFileIndexLow;
3325 val.mv_data = &idbuf;
3326 val.mv_size = sizeof(idbuf);
3327 mdb_hash_hex(&val, hexbuf);
3328 sprintf(env->me_txns->mti_rmname, "Global\\MDBr%s", hexbuf);
3329 sprintf(env->me_txns->mti_wmname, "Global\\MDBw%s", hexbuf);
3330 env->me_rmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_rmname);
3331 if (!env->me_rmutex) goto fail_errno;
3332 env->me_wmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_wmname);
3333 if (!env->me_wmutex) goto fail_errno;
3334 #elif defined(MDB_USE_POSIX_SEM)
3343 if (fstat(env->me_lfd, &stbuf)) goto fail_errno;
3344 idbuf.dev = stbuf.st_dev;
3345 idbuf.ino = stbuf.st_ino;
3346 val.mv_data = &idbuf;
3347 val.mv_size = sizeof(idbuf);
3348 mdb_hash_hex(&val, hexbuf);
3349 sprintf(env->me_txns->mti_rmname, "/MDBr%s", hexbuf);
3350 sprintf(env->me_txns->mti_wmname, "/MDBw%s", hexbuf);
3351 /* Clean up after a previous run, if needed: Try to
3352 * remove both semaphores before doing anything else.
3354 sem_unlink(env->me_txns->mti_rmname);
3355 sem_unlink(env->me_txns->mti_wmname);
3356 env->me_rmutex = sem_open(env->me_txns->mti_rmname,
3357 O_CREAT|O_EXCL, mode, 1);
3358 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
3359 env->me_wmutex = sem_open(env->me_txns->mti_wmname,
3360 O_CREAT|O_EXCL, mode, 1);
3361 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
3362 #else /* MDB_USE_POSIX_SEM */
3363 pthread_mutexattr_t mattr;
3365 if ((rc = pthread_mutexattr_init(&mattr))
3366 || (rc = pthread_mutexattr_setpshared(&mattr, PTHREAD_PROCESS_SHARED))
3367 || (rc = pthread_mutex_init(&env->me_txns->mti_mutex, &mattr))
3368 || (rc = pthread_mutex_init(&env->me_txns->mti_wmutex, &mattr)))
3370 pthread_mutexattr_destroy(&mattr);
3371 #endif /* _WIN32 || MDB_USE_POSIX_SEM */
3373 env->me_txns->mti_version = MDB_VERSION;
3374 env->me_txns->mti_magic = MDB_MAGIC;
3375 env->me_txns->mti_txnid = 0;
3376 env->me_txns->mti_numreaders = 0;
3379 if (env->me_txns->mti_magic != MDB_MAGIC) {
3380 DPUTS("lock region has invalid magic");
3384 if (env->me_txns->mti_version != MDB_VERSION) {
3385 DPRINTF("lock region is version %u, expected version %u",
3386 env->me_txns->mti_version, MDB_VERSION);
3387 rc = MDB_VERSION_MISMATCH;
3391 if (rc != EACCES && rc != EAGAIN) {
3395 env->me_rmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_rmname);
3396 if (!env->me_rmutex) goto fail_errno;
3397 env->me_wmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_wmname);
3398 if (!env->me_wmutex) goto fail_errno;
3399 #elif defined(MDB_USE_POSIX_SEM)
3400 env->me_rmutex = sem_open(env->me_txns->mti_rmname, 0);
3401 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
3402 env->me_wmutex = sem_open(env->me_txns->mti_wmname, 0);
3403 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
3414 /** The name of the lock file in the DB environment */
3415 #define LOCKNAME "/lock.mdb"
3416 /** The name of the data file in the DB environment */
3417 #define DATANAME "/data.mdb"
3418 /** The suffix of the lock file when no subdir is used */
3419 #define LOCKSUFF "-lock"
3420 /** Only a subset of the @ref mdb_env flags can be changed
3421 * at runtime. Changing other flags requires closing the
3422 * environment and re-opening it with the new flags.
3424 #define CHANGEABLE (MDB_NOSYNC|MDB_NOMETASYNC|MDB_MAPASYNC)
3425 #define CHANGELESS (MDB_FIXEDMAP|MDB_NOSUBDIR|MDB_RDONLY|MDB_WRITEMAP|MDB_NOTLS)
3428 mdb_env_open(MDB_env *env, const char *path, unsigned int flags, mdb_mode_t mode)
3430 int oflags, rc, len, excl = -1;
3431 char *lpath, *dpath;
3433 if (env->me_fd!=INVALID_HANDLE_VALUE || (flags & ~(CHANGEABLE|CHANGELESS)))
3437 if (flags & MDB_NOSUBDIR) {
3438 rc = len + sizeof(LOCKSUFF) + len + 1;
3440 rc = len + sizeof(LOCKNAME) + len + sizeof(DATANAME);
3445 if (flags & MDB_NOSUBDIR) {
3446 dpath = lpath + len + sizeof(LOCKSUFF);
3447 sprintf(lpath, "%s" LOCKSUFF, path);
3448 strcpy(dpath, path);
3450 dpath = lpath + len + sizeof(LOCKNAME);
3451 sprintf(lpath, "%s" LOCKNAME, path);
3452 sprintf(dpath, "%s" DATANAME, path);
3456 flags |= env->me_flags;
3457 if (flags & MDB_RDONLY) {
3458 /* silently ignore WRITEMAP when we're only getting read access */
3459 flags &= ~MDB_WRITEMAP;
3461 if (!((env->me_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)) &&
3462 (env->me_dirty_list = calloc(MDB_IDL_UM_SIZE, sizeof(MDB_ID2)))))
3465 env->me_flags = flags |= MDB_ENV_ACTIVE;
3469 env->me_path = strdup(path);
3470 env->me_dbxs = calloc(env->me_maxdbs, sizeof(MDB_dbx));
3471 env->me_dbflags = calloc(env->me_maxdbs, sizeof(uint16_t));
3472 if (!(env->me_dbxs && env->me_path && env->me_dbflags)) {
3477 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
3482 if (F_ISSET(flags, MDB_RDONLY)) {
3483 oflags = GENERIC_READ;
3484 len = OPEN_EXISTING;
3486 oflags = GENERIC_READ|GENERIC_WRITE;
3489 mode = FILE_ATTRIBUTE_NORMAL;
3490 env->me_fd = CreateFile(dpath, oflags, FILE_SHARE_READ|FILE_SHARE_WRITE,
3491 NULL, len, mode, NULL);
3493 if (F_ISSET(flags, MDB_RDONLY))
3496 oflags = O_RDWR | O_CREAT;
3498 env->me_fd = open(dpath, oflags, mode);
3500 if (env->me_fd == INVALID_HANDLE_VALUE) {
3505 if ((rc = mdb_env_open2(env)) == MDB_SUCCESS) {
3506 if (flags & (MDB_RDONLY|MDB_WRITEMAP)) {
3507 env->me_mfd = env->me_fd;
3509 /* Synchronous fd for meta writes. Needed even with
3510 * MDB_NOSYNC/MDB_NOMETASYNC, in case these get reset.
3513 env->me_mfd = CreateFile(dpath, oflags,
3514 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, len,
3515 mode | FILE_FLAG_WRITE_THROUGH, NULL);
3517 env->me_mfd = open(dpath, oflags | MDB_DSYNC, mode);
3519 if (env->me_mfd == INVALID_HANDLE_VALUE) {
3524 DPRINTF("opened dbenv %p", (void *) env);
3526 rc = mdb_env_share_locks(env, &excl);
3532 mdb_env_close0(env, excl);
3538 /** Destroy resources from mdb_env_open(), clear our readers & DBIs */
3540 mdb_env_close0(MDB_env *env, int excl)
3544 if (!(env->me_flags & MDB_ENV_ACTIVE))
3547 /* Doing this here since me_dbxs may not exist during mdb_env_close */
3548 for (i = env->me_maxdbs; --i > MAIN_DBI; )
3549 free(env->me_dbxs[i].md_name.mv_data);
3551 free(env->me_dbflags);
3554 free(env->me_dirty_list);
3555 if (env->me_free_pgs)
3556 mdb_midl_free(env->me_free_pgs);
3558 if (env->me_flags & MDB_ENV_TXKEY) {
3559 pthread_key_delete(env->me_txkey);
3561 /* Delete our key from the global list */
3562 for (i=0; i<mdb_tls_nkeys; i++)
3563 if (mdb_tls_keys[i] == env->me_txkey) {
3564 mdb_tls_keys[i] = mdb_tls_keys[mdb_tls_nkeys-1];
3572 munmap(env->me_map, env->me_mapsize);
3574 if (env->me_mfd != env->me_fd && env->me_mfd != INVALID_HANDLE_VALUE)
3576 if (env->me_fd != INVALID_HANDLE_VALUE)
3579 pid_t pid = env->me_pid;
3580 /* Clearing readers is done in this function because
3581 * me_txkey with its destructor must be disabled first.
3583 for (i = env->me_numreaders; --i >= 0; )
3584 if (env->me_txns->mti_readers[i].mr_pid == pid)
3585 env->me_txns->mti_readers[i].mr_pid = 0;
3587 if (env->me_rmutex) {
3588 CloseHandle(env->me_rmutex);
3589 if (env->me_wmutex) CloseHandle(env->me_wmutex);
3591 /* Windows automatically destroys the mutexes when
3592 * the last handle closes.
3594 #elif defined(MDB_USE_POSIX_SEM)
3595 if (env->me_rmutex != SEM_FAILED) {
3596 sem_close(env->me_rmutex);
3597 if (env->me_wmutex != SEM_FAILED)
3598 sem_close(env->me_wmutex);
3599 /* If we have the filelock: If we are the
3600 * only remaining user, clean up semaphores.
3603 mdb_env_excl_lock(env, &excl);
3605 sem_unlink(env->me_txns->mti_rmname);
3606 sem_unlink(env->me_txns->mti_wmname);
3610 munmap((void *)env->me_txns, (env->me_maxreaders-1)*sizeof(MDB_reader)+sizeof(MDB_txninfo));
3612 if (env->me_lfd != INVALID_HANDLE_VALUE) {
3615 /* Unlock the lockfile. Windows would have unlocked it
3616 * after closing anyway, but not necessarily at once.
3618 UnlockFile(env->me_lfd, 0, 0, 1, 0);
3624 env->me_flags &= ~(MDB_ENV_ACTIVE|MDB_ENV_TXKEY);
3628 mdb_env_copyfd(MDB_env *env, int fd)
3630 MDB_txn *txn = NULL;
3635 /* Do the lock/unlock of the reader mutex before starting the
3636 * write txn. Otherwise other read txns could block writers.
3638 rc = mdb_txn_begin(env, NULL, MDB_RDONLY, &txn);
3643 /* We must start the actual read txn after blocking writers */
3644 mdb_txn_reset0(txn);
3646 /* Temporarily block writers until we snapshot the meta pages */
3649 rc = mdb_txn_renew0(txn);
3651 UNLOCK_MUTEX_W(env);
3656 wsize = env->me_psize * 2;
3660 rc = WriteFile(fd, env->me_map, wsize, &len, NULL);
3661 rc = (len == wsize) ? MDB_SUCCESS : ErrCode();
3664 rc = write(fd, env->me_map, wsize);
3665 rc = (rc == (int)wsize) ? MDB_SUCCESS : ErrCode();
3668 UNLOCK_MUTEX_W(env);
3673 ptr = env->me_map + wsize;
3674 wsize = txn->mt_next_pgno * env->me_psize - wsize;
3675 #define MAX_WRITE 2147483648U
3679 if (wsize > MAX_WRITE)
3683 rc = WriteFile(fd, ptr, w2, &len, NULL);
3684 rc = (len == w2) ? MDB_SUCCESS : ErrCode();
3693 if (wsize > MAX_WRITE)
3697 wres = write(fd, ptr, w2);
3698 rc = (wres > 0) ? MDB_SUCCESS : ErrCode();
3711 mdb_env_copy(MDB_env *env, const char *path)
3715 HANDLE newfd = INVALID_HANDLE_VALUE;
3717 if (env->me_flags & MDB_NOSUBDIR) {
3718 lpath = (char *)path;
3721 len += sizeof(DATANAME);
3722 lpath = malloc(len);
3725 sprintf(lpath, "%s" DATANAME, path);
3728 /* The destination path must exist, but the destination file must not.
3729 * We don't want the OS to cache the writes, since the source data is
3730 * already in the OS cache.
3733 newfd = CreateFile(lpath, GENERIC_WRITE, 0, NULL, CREATE_NEW,
3734 FILE_FLAG_NO_BUFFERING|FILE_FLAG_WRITE_THROUGH, NULL);
3736 newfd = open(lpath, O_WRONLY|O_CREAT|O_EXCL
3742 if (!(env->me_flags & MDB_NOSUBDIR))
3744 if (newfd == INVALID_HANDLE_VALUE) {
3749 #ifdef F_NOCACHE /* __APPLE__ */
3750 rc = fcntl(newfd, F_NOCACHE, 1);
3757 rc = mdb_env_copyfd(env, newfd);
3760 if (newfd != INVALID_HANDLE_VALUE)
3767 mdb_env_close(MDB_env *env)
3774 VGMEMP_DESTROY(env);
3775 while ((dp = env->me_dpages) != NULL) {
3776 VGMEMP_DEFINED(&dp->mp_next, sizeof(dp->mp_next));
3777 env->me_dpages = dp->mp_next;
3781 mdb_env_close0(env, 0);
3785 /** Compare two items pointing at aligned size_t's */
3787 mdb_cmp_long(const MDB_val *a, const MDB_val *b)
3789 return (*(size_t *)a->mv_data < *(size_t *)b->mv_data) ? -1 :
3790 *(size_t *)a->mv_data > *(size_t *)b->mv_data;
3793 /** Compare two items pointing at aligned int's */
3795 mdb_cmp_int(const MDB_val *a, const MDB_val *b)
3797 return (*(unsigned int *)a->mv_data < *(unsigned int *)b->mv_data) ? -1 :
3798 *(unsigned int *)a->mv_data > *(unsigned int *)b->mv_data;
3801 /** Compare two items pointing at ints of unknown alignment.
3802 * Nodes and keys are guaranteed to be 2-byte aligned.
3805 mdb_cmp_cint(const MDB_val *a, const MDB_val *b)
3807 #if BYTE_ORDER == LITTLE_ENDIAN
3808 unsigned short *u, *c;
3811 u = (unsigned short *) ((char *) a->mv_data + a->mv_size);
3812 c = (unsigned short *) ((char *) b->mv_data + a->mv_size);
3815 } while(!x && u > (unsigned short *)a->mv_data);
3818 return memcmp(a->mv_data, b->mv_data, a->mv_size);
3822 /** Compare two items lexically */
3824 mdb_cmp_memn(const MDB_val *a, const MDB_val *b)
3831 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
3837 diff = memcmp(a->mv_data, b->mv_data, len);
3838 return diff ? diff : len_diff<0 ? -1 : len_diff;
3841 /** Compare two items in reverse byte order */
3843 mdb_cmp_memnr(const MDB_val *a, const MDB_val *b)
3845 const unsigned char *p1, *p2, *p1_lim;
3849 p1_lim = (const unsigned char *)a->mv_data;
3850 p1 = (const unsigned char *)a->mv_data + a->mv_size;
3851 p2 = (const unsigned char *)b->mv_data + b->mv_size;
3853 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
3859 while (p1 > p1_lim) {
3860 diff = *--p1 - *--p2;
3864 return len_diff<0 ? -1 : len_diff;
3867 /** Search for key within a page, using binary search.
3868 * Returns the smallest entry larger or equal to the key.
3869 * If exactp is non-null, stores whether the found entry was an exact match
3870 * in *exactp (1 or 0).
3871 * Updates the cursor index with the index of the found entry.
3872 * If no entry larger or equal to the key is found, returns NULL.
3875 mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp)
3877 unsigned int i = 0, nkeys;
3880 MDB_page *mp = mc->mc_pg[mc->mc_top];
3881 MDB_node *node = NULL;
3886 nkeys = NUMKEYS(mp);
3891 COPY_PGNO(pgno, mp->mp_pgno);
3892 DPRINTF("searching %u keys in %s %spage %zu",
3893 nkeys, IS_LEAF(mp) ? "leaf" : "branch", IS_SUBP(mp) ? "sub-" : "",
3900 low = IS_LEAF(mp) ? 0 : 1;
3902 cmp = mc->mc_dbx->md_cmp;
3904 /* Branch pages have no data, so if using integer keys,
3905 * alignment is guaranteed. Use faster mdb_cmp_int.
3907 if (cmp == mdb_cmp_cint && IS_BRANCH(mp)) {
3908 if (NODEPTR(mp, 1)->mn_ksize == sizeof(size_t))
3915 nodekey.mv_size = mc->mc_db->md_pad;
3916 node = NODEPTR(mp, 0); /* fake */
3917 while (low <= high) {
3918 i = (low + high) >> 1;
3919 nodekey.mv_data = LEAF2KEY(mp, i, nodekey.mv_size);
3920 rc = cmp(key, &nodekey);
3921 DPRINTF("found leaf index %u [%s], rc = %i",
3922 i, DKEY(&nodekey), rc);
3931 while (low <= high) {
3932 i = (low + high) >> 1;
3934 node = NODEPTR(mp, i);
3935 nodekey.mv_size = NODEKSZ(node);
3936 nodekey.mv_data = NODEKEY(node);
3938 rc = cmp(key, &nodekey);
3941 DPRINTF("found leaf index %u [%s], rc = %i",
3942 i, DKEY(&nodekey), rc);
3944 DPRINTF("found branch index %u [%s -> %zu], rc = %i",
3945 i, DKEY(&nodekey), NODEPGNO(node), rc);
3956 if (rc > 0) { /* Found entry is less than the key. */
3957 i++; /* Skip to get the smallest entry larger than key. */
3959 node = NODEPTR(mp, i);
3962 *exactp = (rc == 0);
3963 /* store the key index */
3964 mc->mc_ki[mc->mc_top] = i;
3966 /* There is no entry larger or equal to the key. */
3969 /* nodeptr is fake for LEAF2 */
3975 mdb_cursor_adjust(MDB_cursor *mc, func)
3979 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
3980 if (m2->mc_pg[m2->mc_top] == mc->mc_pg[mc->mc_top]) {
3987 /** Pop a page off the top of the cursor's stack. */
3989 mdb_cursor_pop(MDB_cursor *mc)
3992 #ifndef MDB_DEBUG_SKIP
3993 MDB_page *top = mc->mc_pg[mc->mc_top];
3999 DPRINTF("popped page %zu off db %u cursor %p", top->mp_pgno,
4000 mc->mc_dbi, (void *) mc);
4004 /** Push a page onto the top of the cursor's stack. */
4006 mdb_cursor_push(MDB_cursor *mc, MDB_page *mp)
4008 DPRINTF("pushing page %zu on db %u cursor %p", mp->mp_pgno,
4009 mc->mc_dbi, (void *) mc);
4011 if (mc->mc_snum >= CURSOR_STACK) {
4012 assert(mc->mc_snum < CURSOR_STACK);
4013 return MDB_CURSOR_FULL;
4016 mc->mc_top = mc->mc_snum++;
4017 mc->mc_pg[mc->mc_top] = mp;
4018 mc->mc_ki[mc->mc_top] = 0;
4023 /** Find the address of the page corresponding to a given page number.
4024 * @param[in] txn the transaction for this access.
4025 * @param[in] pgno the page number for the page to retrieve.
4026 * @param[out] ret address of a pointer where the page's address will be stored.
4027 * @param[out] lvl dirty_list inheritance level of found page. 1=current txn, 0=mapped page.
4028 * @return 0 on success, non-zero on failure.
4031 mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **ret, int *lvl)
4036 if (!((txn->mt_flags & MDB_TXN_RDONLY) |
4037 (txn->mt_env->me_flags & MDB_WRITEMAP)))
4042 MDB_ID2L dl = tx2->mt_u.dirty_list;
4044 unsigned x = mdb_mid2l_search(dl, pgno);
4045 if (x <= dl[0].mid && dl[x].mid == pgno) {
4051 } while ((tx2 = tx2->mt_parent) != NULL);
4054 if (pgno < txn->mt_next_pgno) {
4056 p = (MDB_page *)(txn->mt_env->me_map + txn->mt_env->me_psize * pgno);
4058 DPRINTF("page %zu not found", pgno);
4060 return MDB_PAGE_NOTFOUND;
4070 /** Search for the page a given key should be in.
4071 * Pushes parent pages on the cursor stack. This function continues a
4072 * search on a cursor that has already been initialized. (Usually by
4073 * #mdb_page_search() but also by #mdb_node_move().)
4074 * @param[in,out] mc the cursor for this operation.
4075 * @param[in] key the key to search for. If NULL, search for the lowest
4076 * page. (This is used by #mdb_cursor_first().)
4077 * @param[in] modify If true, visited pages are updated with new page numbers.
4078 * @return 0 on success, non-zero on failure.
4081 mdb_page_search_root(MDB_cursor *mc, MDB_val *key, int modify)
4083 MDB_page *mp = mc->mc_pg[mc->mc_top];
4088 while (IS_BRANCH(mp)) {
4092 DPRINTF("branch page %zu has %u keys", mp->mp_pgno, NUMKEYS(mp));
4093 assert(NUMKEYS(mp) > 1);
4094 DPRINTF("found index 0 to page %zu", NODEPGNO(NODEPTR(mp, 0)));
4096 if (key == NULL) /* Initialize cursor to first page. */
4098 else if (key->mv_size > MDB_MAXKEYSIZE && key->mv_data == NULL) {
4099 /* cursor to last page */
4103 node = mdb_node_search(mc, key, &exact);
4105 i = NUMKEYS(mp) - 1;
4107 i = mc->mc_ki[mc->mc_top];
4116 DPRINTF("following index %u for key [%s]",
4118 assert(i < NUMKEYS(mp));
4119 node = NODEPTR(mp, i);
4121 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp, NULL)) != 0)
4124 mc->mc_ki[mc->mc_top] = i;
4125 if ((rc = mdb_cursor_push(mc, mp)))
4129 if ((rc = mdb_page_touch(mc)) != 0)
4131 mp = mc->mc_pg[mc->mc_top];
4136 DPRINTF("internal error, index points to a %02X page!?",
4138 return MDB_CORRUPTED;
4141 DPRINTF("found leaf page %zu for key [%s]", mp->mp_pgno,
4142 key ? DKEY(key) : NULL);
4147 /** Search for the lowest key under the current branch page.
4148 * This just bypasses a NUMKEYS check in the current page
4149 * before calling mdb_page_search_root(), because the callers
4150 * are all in situations where the current page is known to
4154 mdb_page_search_lowest(MDB_cursor *mc)
4156 MDB_page *mp = mc->mc_pg[mc->mc_top];
4157 MDB_node *node = NODEPTR(mp, 0);
4160 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp, NULL)) != 0)
4163 mc->mc_ki[mc->mc_top] = 0;
4164 if ((rc = mdb_cursor_push(mc, mp)))
4166 return mdb_page_search_root(mc, NULL, 0);
4169 /** Search for the page a given key should be in.
4170 * Pushes parent pages on the cursor stack. This function just sets up
4171 * the search; it finds the root page for \b mc's database and sets this
4172 * as the root of the cursor's stack. Then #mdb_page_search_root() is
4173 * called to complete the search.
4174 * @param[in,out] mc the cursor for this operation.
4175 * @param[in] key the key to search for. If NULL, search for the lowest
4176 * page. (This is used by #mdb_cursor_first().)
4177 * @param[in] flags If MDB_PS_MODIFY set, visited pages are updated with new page numbers.
4178 * If MDB_PS_ROOTONLY set, just fetch root node, no further lookups.
4179 * @return 0 on success, non-zero on failure.
4182 mdb_page_search(MDB_cursor *mc, MDB_val *key, int flags)
4187 /* Make sure the txn is still viable, then find the root from
4188 * the txn's db table.
4190 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_ERROR)) {
4191 DPUTS("transaction has failed, must abort");
4194 /* Make sure we're using an up-to-date root */
4195 if (mc->mc_dbi > MAIN_DBI) {
4196 if ((*mc->mc_dbflag & DB_STALE) ||
4197 ((flags & MDB_PS_MODIFY) && !(*mc->mc_dbflag & DB_DIRTY))) {
4199 unsigned char dbflag = 0;
4200 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, NULL);
4201 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, flags & MDB_PS_MODIFY);
4204 if (*mc->mc_dbflag & DB_STALE) {
4208 MDB_node *leaf = mdb_node_search(&mc2,
4209 &mc->mc_dbx->md_name, &exact);
4211 return MDB_NOTFOUND;
4212 mdb_node_read(mc->mc_txn, leaf, &data);
4213 memcpy(&flags, ((char *) data.mv_data + offsetof(MDB_db, md_flags)),
4215 /* The txn may not know this DBI, or another process may
4216 * have dropped and recreated the DB with other flags.
4218 if ((mc->mc_db->md_flags & PERSISTENT_FLAGS) != flags)
4219 return MDB_INCOMPATIBLE;
4220 memcpy(mc->mc_db, data.mv_data, sizeof(MDB_db));
4222 if (flags & MDB_PS_MODIFY)
4224 *mc->mc_dbflag &= ~DB_STALE;
4225 *mc->mc_dbflag |= dbflag;
4228 root = mc->mc_db->md_root;
4230 if (root == P_INVALID) { /* Tree is empty. */
4231 DPUTS("tree is empty");
4232 return MDB_NOTFOUND;
4237 if (!mc->mc_pg[0] || mc->mc_pg[0]->mp_pgno != root)
4238 if ((rc = mdb_page_get(mc->mc_txn, root, &mc->mc_pg[0], NULL)) != 0)
4244 DPRINTF("db %u root page %zu has flags 0x%X",
4245 mc->mc_dbi, root, mc->mc_pg[0]->mp_flags);
4247 if (flags & MDB_PS_MODIFY) {
4248 if ((rc = mdb_page_touch(mc)))
4252 if (flags & MDB_PS_ROOTONLY)
4255 return mdb_page_search_root(mc, key, flags);
4258 /** Return the data associated with a given node.
4259 * @param[in] txn The transaction for this operation.
4260 * @param[in] leaf The node being read.
4261 * @param[out] data Updated to point to the node's data.
4262 * @return 0 on success, non-zero on failure.
4265 mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data)
4267 MDB_page *omp; /* overflow page */
4271 if (!F_ISSET(leaf->mn_flags, F_BIGDATA)) {
4272 data->mv_size = NODEDSZ(leaf);
4273 data->mv_data = NODEDATA(leaf);
4277 /* Read overflow data.
4279 data->mv_size = NODEDSZ(leaf);
4280 memcpy(&pgno, NODEDATA(leaf), sizeof(pgno));
4281 if ((rc = mdb_page_get(txn, pgno, &omp, NULL)) != 0) {
4282 DPRINTF("read overflow page %zu failed", pgno);
4285 data->mv_data = METADATA(omp);
4291 mdb_get(MDB_txn *txn, MDB_dbi dbi,
4292 MDB_val *key, MDB_val *data)
4301 DPRINTF("===> get db %u key [%s]", dbi, DKEY(key));
4303 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
4306 if (key->mv_size == 0 || key->mv_size > MDB_MAXKEYSIZE) {
4310 mdb_cursor_init(&mc, txn, dbi, &mx);
4311 return mdb_cursor_set(&mc, key, data, MDB_SET, &exact);
4314 /** Find a sibling for a page.
4315 * Replaces the page at the top of the cursor's stack with the
4316 * specified sibling, if one exists.
4317 * @param[in] mc The cursor for this operation.
4318 * @param[in] move_right Non-zero if the right sibling is requested,
4319 * otherwise the left sibling.
4320 * @return 0 on success, non-zero on failure.
4323 mdb_cursor_sibling(MDB_cursor *mc, int move_right)
4329 if (mc->mc_snum < 2) {
4330 return MDB_NOTFOUND; /* root has no siblings */
4334 DPRINTF("parent page is page %zu, index %u",
4335 mc->mc_pg[mc->mc_top]->mp_pgno, mc->mc_ki[mc->mc_top]);
4337 if (move_right ? (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mc->mc_pg[mc->mc_top]))
4338 : (mc->mc_ki[mc->mc_top] == 0)) {
4339 DPRINTF("no more keys left, moving to %s sibling",
4340 move_right ? "right" : "left");
4341 if ((rc = mdb_cursor_sibling(mc, move_right)) != MDB_SUCCESS) {
4342 /* undo cursor_pop before returning */
4349 mc->mc_ki[mc->mc_top]++;
4351 mc->mc_ki[mc->mc_top]--;
4352 DPRINTF("just moving to %s index key %u",
4353 move_right ? "right" : "left", mc->mc_ki[mc->mc_top]);
4355 assert(IS_BRANCH(mc->mc_pg[mc->mc_top]));
4357 indx = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
4358 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(indx), &mp, NULL) != 0))
4361 mdb_cursor_push(mc, mp);
4363 mc->mc_ki[mc->mc_top] = NUMKEYS(mp)-1;
4368 /** Move the cursor to the next data item. */
4370 mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
4376 if (mc->mc_flags & C_EOF) {
4377 return MDB_NOTFOUND;
4380 assert(mc->mc_flags & C_INITIALIZED);
4382 mp = mc->mc_pg[mc->mc_top];
4384 if (mc->mc_db->md_flags & MDB_DUPSORT) {
4385 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4386 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4387 if (op == MDB_NEXT || op == MDB_NEXT_DUP) {
4388 rc = mdb_cursor_next(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_NEXT);
4389 if (op != MDB_NEXT || rc == MDB_SUCCESS)
4393 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
4394 if (op == MDB_NEXT_DUP)
4395 return MDB_NOTFOUND;
4399 DPRINTF("cursor_next: top page is %zu in cursor %p", mp->mp_pgno, (void *) mc);
4401 if (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mp)) {
4402 DPUTS("=====> move to next sibling page");
4403 if (mdb_cursor_sibling(mc, 1) != MDB_SUCCESS) {
4404 mc->mc_flags |= C_EOF;
4405 mc->mc_flags &= ~C_INITIALIZED;
4406 return MDB_NOTFOUND;
4408 mp = mc->mc_pg[mc->mc_top];
4409 DPRINTF("next page is %zu, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]);
4411 mc->mc_ki[mc->mc_top]++;
4413 DPRINTF("==> cursor points to page %zu with %u keys, key index %u",
4414 mp->mp_pgno, NUMKEYS(mp), mc->mc_ki[mc->mc_top]);
4417 key->mv_size = mc->mc_db->md_pad;
4418 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
4422 assert(IS_LEAF(mp));
4423 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4425 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4426 mdb_xcursor_init1(mc, leaf);
4429 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
4432 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4433 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
4434 if (rc != MDB_SUCCESS)
4439 MDB_GET_KEY(leaf, key);
4443 /** Move the cursor to the previous data item. */
4445 mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
4451 assert(mc->mc_flags & C_INITIALIZED);
4453 mp = mc->mc_pg[mc->mc_top];
4455 if (mc->mc_db->md_flags & MDB_DUPSORT) {
4456 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4457 if (op == MDB_PREV || op == MDB_PREV_DUP) {
4458 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4459 rc = mdb_cursor_prev(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_PREV);
4460 if (op != MDB_PREV || rc == MDB_SUCCESS)
4463 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
4464 if (op == MDB_PREV_DUP)
4465 return MDB_NOTFOUND;
4470 DPRINTF("cursor_prev: top page is %zu in cursor %p", mp->mp_pgno, (void *) mc);
4472 if (mc->mc_ki[mc->mc_top] == 0) {
4473 DPUTS("=====> move to prev sibling page");
4474 if (mdb_cursor_sibling(mc, 0) != MDB_SUCCESS) {
4475 mc->mc_flags &= ~C_INITIALIZED;
4476 return MDB_NOTFOUND;
4478 mp = mc->mc_pg[mc->mc_top];
4479 mc->mc_ki[mc->mc_top] = NUMKEYS(mp) - 1;
4480 DPRINTF("prev page is %zu, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]);
4482 mc->mc_ki[mc->mc_top]--;
4484 mc->mc_flags &= ~C_EOF;
4486 DPRINTF("==> cursor points to page %zu with %u keys, key index %u",
4487 mp->mp_pgno, NUMKEYS(mp), mc->mc_ki[mc->mc_top]);
4490 key->mv_size = mc->mc_db->md_pad;
4491 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
4495 assert(IS_LEAF(mp));
4496 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4498 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4499 mdb_xcursor_init1(mc, leaf);
4502 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
4505 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4506 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
4507 if (rc != MDB_SUCCESS)
4512 MDB_GET_KEY(leaf, key);
4516 /** Set the cursor on a specific data item. */
4518 mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data,
4519 MDB_cursor_op op, int *exactp)
4523 MDB_node *leaf = NULL;
4528 assert(key->mv_size > 0);
4530 /* See if we're already on the right page */
4531 if (mc->mc_flags & C_INITIALIZED) {
4534 mp = mc->mc_pg[mc->mc_top];
4536 mc->mc_ki[mc->mc_top] = 0;
4537 return MDB_NOTFOUND;
4539 if (mp->mp_flags & P_LEAF2) {
4540 nodekey.mv_size = mc->mc_db->md_pad;
4541 nodekey.mv_data = LEAF2KEY(mp, 0, nodekey.mv_size);
4543 leaf = NODEPTR(mp, 0);
4544 MDB_GET_KEY(leaf, &nodekey);
4546 rc = mc->mc_dbx->md_cmp(key, &nodekey);
4548 /* Probably happens rarely, but first node on the page
4549 * was the one we wanted.
4551 mc->mc_ki[mc->mc_top] = 0;
4558 unsigned int nkeys = NUMKEYS(mp);
4560 if (mp->mp_flags & P_LEAF2) {
4561 nodekey.mv_data = LEAF2KEY(mp,
4562 nkeys-1, nodekey.mv_size);
4564 leaf = NODEPTR(mp, nkeys-1);
4565 MDB_GET_KEY(leaf, &nodekey);
4567 rc = mc->mc_dbx->md_cmp(key, &nodekey);
4569 /* last node was the one we wanted */
4570 mc->mc_ki[mc->mc_top] = nkeys-1;
4576 if (mc->mc_ki[mc->mc_top] < NUMKEYS(mp)) {
4577 /* This is definitely the right page, skip search_page */
4578 if (mp->mp_flags & P_LEAF2) {
4579 nodekey.mv_data = LEAF2KEY(mp,
4580 mc->mc_ki[mc->mc_top], nodekey.mv_size);
4582 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4583 MDB_GET_KEY(leaf, &nodekey);
4585 rc = mc->mc_dbx->md_cmp(key, &nodekey);
4587 /* current node was the one we wanted */
4597 /* If any parents have right-sibs, search.
4598 * Otherwise, there's nothing further.
4600 for (i=0; i<mc->mc_top; i++)
4602 NUMKEYS(mc->mc_pg[i])-1)
4604 if (i == mc->mc_top) {
4605 /* There are no other pages */
4606 mc->mc_ki[mc->mc_top] = nkeys;
4607 return MDB_NOTFOUND;
4611 /* There are no other pages */
4612 mc->mc_ki[mc->mc_top] = 0;
4613 return MDB_NOTFOUND;
4617 rc = mdb_page_search(mc, key, 0);
4618 if (rc != MDB_SUCCESS)
4621 mp = mc->mc_pg[mc->mc_top];
4622 assert(IS_LEAF(mp));
4625 leaf = mdb_node_search(mc, key, exactp);
4626 if (exactp != NULL && !*exactp) {
4627 /* MDB_SET specified and not an exact match. */
4628 return MDB_NOTFOUND;
4632 DPUTS("===> inexact leaf not found, goto sibling");
4633 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS)
4634 return rc; /* no entries matched */
4635 mp = mc->mc_pg[mc->mc_top];
4636 assert(IS_LEAF(mp));
4637 leaf = NODEPTR(mp, 0);
4641 mc->mc_flags |= C_INITIALIZED;
4642 mc->mc_flags &= ~C_EOF;
4645 key->mv_size = mc->mc_db->md_pad;
4646 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
4650 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4651 mdb_xcursor_init1(mc, leaf);
4654 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4655 if (op == MDB_SET || op == MDB_SET_KEY || op == MDB_SET_RANGE) {
4656 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
4659 if (op == MDB_GET_BOTH) {
4665 rc = mdb_cursor_set(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_SET_RANGE, ex2p);
4666 if (rc != MDB_SUCCESS)
4669 } else if (op == MDB_GET_BOTH || op == MDB_GET_BOTH_RANGE) {
4671 if ((rc = mdb_node_read(mc->mc_txn, leaf, &d2)) != MDB_SUCCESS)
4673 rc = mc->mc_dbx->md_dcmp(data, &d2);
4675 if (op == MDB_GET_BOTH || rc > 0)
4676 return MDB_NOTFOUND;
4681 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
4682 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
4687 /* The key already matches in all other cases */
4688 if (op == MDB_SET_RANGE || op == MDB_SET_KEY)
4689 MDB_GET_KEY(leaf, key);
4690 DPRINTF("==> cursor placed on key [%s]", DKEY(key));
4695 /** Move the cursor to the first item in the database. */
4697 mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data)
4702 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
4703 rc = mdb_page_search(mc, NULL, 0);
4704 if (rc != MDB_SUCCESS)
4707 assert(IS_LEAF(mc->mc_pg[mc->mc_top]));
4709 leaf = NODEPTR(mc->mc_pg[mc->mc_top], 0);
4710 mc->mc_flags |= C_INITIALIZED;
4711 mc->mc_flags &= ~C_EOF;
4713 mc->mc_ki[mc->mc_top] = 0;
4715 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
4716 key->mv_size = mc->mc_db->md_pad;
4717 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], 0, key->mv_size);
4722 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4723 mdb_xcursor_init1(mc, leaf);
4724 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
4729 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
4730 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
4734 MDB_GET_KEY(leaf, key);
4738 /** Move the cursor to the last item in the database. */
4740 mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data)
4745 if (!(mc->mc_flags & C_EOF)) {
4747 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
4750 lkey.mv_size = MDB_MAXKEYSIZE+1;
4751 lkey.mv_data = NULL;
4752 rc = mdb_page_search(mc, &lkey, 0);
4753 if (rc != MDB_SUCCESS)
4756 assert(IS_LEAF(mc->mc_pg[mc->mc_top]));
4759 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]) - 1;
4760 mc->mc_flags |= C_INITIALIZED|C_EOF;
4761 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
4763 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
4764 key->mv_size = mc->mc_db->md_pad;
4765 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], key->mv_size);
4770 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4771 mdb_xcursor_init1(mc, leaf);
4772 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
4777 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
4778 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
4783 MDB_GET_KEY(leaf, key);
4788 mdb_cursor_get(MDB_cursor *mc, MDB_val *key, MDB_val *data,
4797 case MDB_GET_CURRENT:
4798 if (!(mc->mc_flags & C_INITIALIZED)) {
4801 MDB_page *mp = mc->mc_pg[mc->mc_top];
4803 mc->mc_ki[mc->mc_top] = 0;
4809 key->mv_size = mc->mc_db->md_pad;
4810 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
4812 MDB_node *leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4813 MDB_GET_KEY(leaf, key);
4815 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4816 rc = mdb_cursor_get(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_GET_CURRENT);
4818 rc = mdb_node_read(mc->mc_txn, leaf, data);
4825 case MDB_GET_BOTH_RANGE:
4826 if (data == NULL || mc->mc_xcursor == NULL) {
4834 if (key == NULL || key->mv_size == 0 || key->mv_size > MDB_MAXKEYSIZE) {
4836 } else if (op == MDB_SET_RANGE)
4837 rc = mdb_cursor_set(mc, key, data, op, NULL);
4839 rc = mdb_cursor_set(mc, key, data, op, &exact);
4841 case MDB_GET_MULTIPLE:
4843 !(mc->mc_db->md_flags & MDB_DUPFIXED) ||
4844 !(mc->mc_flags & C_INITIALIZED)) {
4849 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) ||
4850 (mc->mc_xcursor->mx_cursor.mc_flags & C_EOF))
4853 case MDB_NEXT_MULTIPLE:
4855 !(mc->mc_db->md_flags & MDB_DUPFIXED)) {
4859 if (!(mc->mc_flags & C_INITIALIZED))
4860 rc = mdb_cursor_first(mc, key, data);
4862 rc = mdb_cursor_next(mc, key, data, MDB_NEXT_DUP);
4863 if (rc == MDB_SUCCESS) {
4864 if (mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) {
4867 mx = &mc->mc_xcursor->mx_cursor;
4868 data->mv_size = NUMKEYS(mx->mc_pg[mx->mc_top]) *
4870 data->mv_data = METADATA(mx->mc_pg[mx->mc_top]);
4871 mx->mc_ki[mx->mc_top] = NUMKEYS(mx->mc_pg[mx->mc_top])-1;
4879 case MDB_NEXT_NODUP:
4880 if (!(mc->mc_flags & C_INITIALIZED))
4881 rc = mdb_cursor_first(mc, key, data);
4883 rc = mdb_cursor_next(mc, key, data, op);
4887 case MDB_PREV_NODUP:
4888 if (!(mc->mc_flags & C_INITIALIZED)) {
4889 rc = mdb_cursor_last(mc, key, data);
4890 mc->mc_flags |= C_INITIALIZED;
4891 mc->mc_ki[mc->mc_top]++;
4893 rc = mdb_cursor_prev(mc, key, data, op);
4896 rc = mdb_cursor_first(mc, key, data);
4900 !(mc->mc_db->md_flags & MDB_DUPSORT) ||
4901 !(mc->mc_flags & C_INITIALIZED) ||
4902 !(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
4906 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
4909 rc = mdb_cursor_last(mc, key, data);
4913 !(mc->mc_db->md_flags & MDB_DUPSORT) ||
4914 !(mc->mc_flags & C_INITIALIZED) ||
4915 !(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
4919 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
4922 DPRINTF("unhandled/unimplemented cursor operation %u", op);
4930 /** Touch all the pages in the cursor stack.
4931 * Makes sure all the pages are writable, before attempting a write operation.
4932 * @param[in] mc The cursor to operate on.
4935 mdb_cursor_touch(MDB_cursor *mc)
4939 if (mc->mc_dbi > MAIN_DBI && !(*mc->mc_dbflag & DB_DIRTY)) {
4942 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, &mcx);
4943 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, MDB_PS_MODIFY);
4946 *mc->mc_dbflag |= DB_DIRTY;
4948 for (mc->mc_top = 0; mc->mc_top < mc->mc_snum; mc->mc_top++) {
4949 rc = mdb_page_touch(mc);
4953 mc->mc_top = mc->mc_snum-1;
4958 mdb_cursor_put(MDB_cursor *mc, MDB_val *key, MDB_val *data,
4961 MDB_node *leaf = NULL;
4962 MDB_val xdata, *rdata, dkey;
4965 int do_sub = 0, insert = 0;
4966 unsigned int mcount = 0;
4970 char dbuf[MDB_MAXKEYSIZE+1];
4971 unsigned int nflags;
4974 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_RDONLY))
4977 if (flags != MDB_CURRENT && (key->mv_size == 0 || key->mv_size > MDB_MAXKEYSIZE))
4980 if (F_ISSET(mc->mc_db->md_flags, MDB_DUPSORT) && data->mv_size > MDB_MAXKEYSIZE)
4983 #if SIZE_MAX > MAXDATASIZE
4984 if (data->mv_size > MAXDATASIZE)
4988 DPRINTF("==> put db %u key [%s], size %zu, data size %zu",
4989 mc->mc_dbi, DKEY(key), key ? key->mv_size:0, data->mv_size);
4993 if (flags == MDB_CURRENT) {
4994 if (!(mc->mc_flags & C_INITIALIZED))
4997 } else if (mc->mc_db->md_root == P_INVALID) {
4999 /* new database, write a root leaf page */
5000 DPUTS("allocating new root leaf page");
5001 if ((rc = mdb_page_new(mc, P_LEAF, 1, &np))) {
5005 mdb_cursor_push(mc, np);
5006 mc->mc_db->md_root = np->mp_pgno;
5007 mc->mc_db->md_depth++;
5008 *mc->mc_dbflag |= DB_DIRTY;
5009 if ((mc->mc_db->md_flags & (MDB_DUPSORT|MDB_DUPFIXED))
5011 np->mp_flags |= P_LEAF2;
5012 mc->mc_flags |= C_INITIALIZED;
5018 if (flags & MDB_APPEND) {
5020 rc = mdb_cursor_last(mc, &k2, &d2);
5022 rc = mc->mc_dbx->md_cmp(key, &k2);
5025 mc->mc_ki[mc->mc_top]++;
5027 /* new key is <= last key */
5032 rc = mdb_cursor_set(mc, key, &d2, MDB_SET, &exact);
5034 if ((flags & MDB_NOOVERWRITE) && rc == 0) {
5035 DPRINTF("duplicate key [%s]", DKEY(key));
5037 return MDB_KEYEXIST;
5039 if (rc && rc != MDB_NOTFOUND)
5043 /* Cursor is positioned, now make sure all pages are writable */
5044 rc2 = mdb_cursor_touch(mc);
5049 /* The key already exists */
5050 if (rc == MDB_SUCCESS) {
5051 /* there's only a key anyway, so this is a no-op */
5052 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
5053 unsigned int ksize = mc->mc_db->md_pad;
5054 if (key->mv_size != ksize)
5056 if (flags == MDB_CURRENT) {
5057 char *ptr = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], ksize);
5058 memcpy(ptr, key->mv_data, ksize);
5063 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5066 if (F_ISSET(mc->mc_db->md_flags, MDB_DUPSORT)) {
5067 /* Was a single item before, must convert now */
5069 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5070 /* Just overwrite the current item */
5071 if (flags == MDB_CURRENT)
5074 dkey.mv_size = NODEDSZ(leaf);
5075 dkey.mv_data = NODEDATA(leaf);
5076 #if UINT_MAX < SIZE_MAX
5077 if (mc->mc_dbx->md_dcmp == mdb_cmp_int && dkey.mv_size == sizeof(size_t))
5078 #ifdef MISALIGNED_OK
5079 mc->mc_dbx->md_dcmp = mdb_cmp_long;
5081 mc->mc_dbx->md_dcmp = mdb_cmp_cint;
5084 /* if data matches, ignore it */
5085 if (!mc->mc_dbx->md_dcmp(data, &dkey))
5086 return (flags == MDB_NODUPDATA) ? MDB_KEYEXIST : MDB_SUCCESS;
5088 /* create a fake page for the dup items */
5089 memcpy(dbuf, dkey.mv_data, dkey.mv_size);
5090 dkey.mv_data = dbuf;
5091 fp = (MDB_page *)&pbuf;
5092 fp->mp_pgno = mc->mc_pg[mc->mc_top]->mp_pgno;
5093 fp->mp_flags = P_LEAF|P_DIRTY|P_SUBP;
5094 fp->mp_lower = PAGEHDRSZ;
5095 fp->mp_upper = PAGEHDRSZ + dkey.mv_size + data->mv_size;
5096 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
5097 fp->mp_flags |= P_LEAF2;
5098 fp->mp_pad = data->mv_size;
5099 fp->mp_upper += 2 * data->mv_size; /* leave space for 2 more */
5101 fp->mp_upper += 2 * sizeof(indx_t) + 2 * NODESIZE +
5102 (dkey.mv_size & 1) + (data->mv_size & 1);
5104 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
5107 xdata.mv_size = fp->mp_upper;
5112 if (!F_ISSET(leaf->mn_flags, F_SUBDATA)) {
5113 /* See if we need to convert from fake page to subDB */
5115 unsigned int offset;
5118 fp = NODEDATA(leaf);
5119 if (flags == MDB_CURRENT) {
5121 fp->mp_flags |= P_DIRTY;
5122 COPY_PGNO(fp->mp_pgno, mc->mc_pg[mc->mc_top]->mp_pgno);
5123 mc->mc_xcursor->mx_cursor.mc_pg[0] = fp;
5127 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
5128 offset = fp->mp_pad;
5129 if (SIZELEFT(fp) >= offset)
5131 offset *= 4; /* space for 4 more */
5133 offset = NODESIZE + sizeof(indx_t) + data->mv_size;
5135 offset += offset & 1;
5136 if (NODESIZE + sizeof(indx_t) + NODEKSZ(leaf) + NODEDSZ(leaf) +
5137 offset >= mc->mc_txn->mt_env->me_nodemax) {
5138 /* yes, convert it */
5140 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
5141 dummy.md_pad = fp->mp_pad;
5142 dummy.md_flags = MDB_DUPFIXED;
5143 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
5144 dummy.md_flags |= MDB_INTEGERKEY;
5147 dummy.md_branch_pages = 0;
5148 dummy.md_leaf_pages = 1;
5149 dummy.md_overflow_pages = 0;
5150 dummy.md_entries = NUMKEYS(fp);
5152 xdata.mv_size = sizeof(MDB_db);
5153 xdata.mv_data = &dummy;
5154 if ((rc = mdb_page_alloc(mc, 1, &mp)))
5156 offset = mc->mc_txn->mt_env->me_psize - NODEDSZ(leaf);
5157 flags |= F_DUPDATA|F_SUBDATA;
5158 dummy.md_root = mp->mp_pgno;
5160 /* no, just grow it */
5162 xdata.mv_size = NODEDSZ(leaf) + offset;
5163 xdata.mv_data = &pbuf;
5164 mp = (MDB_page *)&pbuf;
5165 mp->mp_pgno = mc->mc_pg[mc->mc_top]->mp_pgno;
5168 mp->mp_flags = fp->mp_flags | P_DIRTY;
5169 mp->mp_pad = fp->mp_pad;
5170 mp->mp_lower = fp->mp_lower;
5171 mp->mp_upper = fp->mp_upper + offset;
5173 memcpy(METADATA(mp), METADATA(fp), NUMKEYS(fp) * fp->mp_pad);
5175 nsize = NODEDSZ(leaf) - fp->mp_upper;
5176 memcpy((char *)mp + mp->mp_upper, (char *)fp + fp->mp_upper, nsize);
5177 for (i=0; i<NUMKEYS(fp); i++)
5178 mp->mp_ptrs[i] = fp->mp_ptrs[i] + offset;
5180 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
5184 /* data is on sub-DB, just store it */
5185 flags |= F_DUPDATA|F_SUBDATA;
5189 /* overflow page overwrites need special handling */
5190 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) {
5193 unsigned psize = mc->mc_txn->mt_env->me_psize;
5194 int level, ovpages, dpages = OVPAGES(data->mv_size, psize);
5196 memcpy(&pg, NODEDATA(leaf), sizeof(pg));
5197 if ((rc2 = mdb_page_get(mc->mc_txn, pg, &omp, &level)) != 0)
5199 ovpages = omp->mp_pages;
5201 /* Is the ov page writable and large enough? */
5202 if ((omp->mp_flags & P_DIRTY) && ovpages >= dpages) {
5203 /* yes, overwrite it. Note in this case we don't
5204 * bother to try shrinking the page if the new data
5205 * is smaller than the overflow threshold.
5208 /* It is writable only in a parent txn */
5209 size_t sz = (size_t) psize * ovpages, off;
5210 MDB_page *np = mdb_page_malloc(mc, ovpages);
5216 mdb_mid2l_insert(mc->mc_txn->mt_u.dirty_list, &id2);
5217 if (!(flags & MDB_RESERVE)) {
5218 /* Copy end of page, adjusting alignment so
5219 * compiler may copy words instead of bytes.
5221 off = (PAGEHDRSZ + data->mv_size) & -sizeof(size_t);
5222 memcpy((size_t *)((char *)np + off),
5223 (size_t *)((char *)omp + off), sz - off);
5226 memcpy(np, omp, sz); /* Copy beginning of page */
5229 SETDSZ(leaf, data->mv_size);
5230 if (F_ISSET(flags, MDB_RESERVE))
5231 data->mv_data = METADATA(omp);
5233 memcpy(METADATA(omp), data->mv_data, data->mv_size);
5236 /* no, free ovpages */
5238 mc->mc_db->md_overflow_pages -= ovpages;
5239 for (i=0; i<ovpages; i++) {
5240 DPRINTF("freed ov page %zu", pg);
5241 mdb_midl_append(&mc->mc_txn->mt_free_pgs, pg);
5245 } else if (NODEDSZ(leaf) == data->mv_size) {
5246 /* same size, just replace it. Note that we could
5247 * also reuse this node if the new data is smaller,
5248 * but instead we opt to shrink the node in that case.
5250 if (F_ISSET(flags, MDB_RESERVE))
5251 data->mv_data = NODEDATA(leaf);
5252 else if (data->mv_size)
5253 memcpy(NODEDATA(leaf), data->mv_data, data->mv_size);
5255 memcpy(NODEKEY(leaf), key->mv_data, key->mv_size);
5258 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
5259 mc->mc_db->md_entries--;
5261 DPRINTF("inserting key at index %i", mc->mc_ki[mc->mc_top]);
5268 nflags = flags & NODE_ADD_FLAGS;
5269 nsize = IS_LEAF2(mc->mc_pg[mc->mc_top]) ? key->mv_size : mdb_leaf_size(mc->mc_txn->mt_env, key, rdata);
5270 if (SIZELEFT(mc->mc_pg[mc->mc_top]) < nsize) {
5271 if (( flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA )
5272 nflags &= ~MDB_APPEND;
5274 nflags |= MDB_SPLIT_REPLACE;
5275 rc = mdb_page_split(mc, key, rdata, P_INVALID, nflags);
5277 /* There is room already in this leaf page. */
5278 rc = mdb_node_add(mc, mc->mc_ki[mc->mc_top], key, rdata, 0, nflags);
5279 if (rc == 0 && !do_sub && insert) {
5280 /* Adjust other cursors pointing to mp */
5281 MDB_cursor *m2, *m3;
5282 MDB_dbi dbi = mc->mc_dbi;
5283 unsigned i = mc->mc_top;
5284 MDB_page *mp = mc->mc_pg[i];
5286 if (mc->mc_flags & C_SUB)
5289 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
5290 if (mc->mc_flags & C_SUB)
5291 m3 = &m2->mc_xcursor->mx_cursor;
5294 if (m3 == mc || m3->mc_snum < mc->mc_snum) continue;
5295 if (m3->mc_pg[i] == mp && m3->mc_ki[i] >= mc->mc_ki[i]) {
5302 if (rc != MDB_SUCCESS)
5303 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
5305 /* Now store the actual data in the child DB. Note that we're
5306 * storing the user data in the keys field, so there are strict
5307 * size limits on dupdata. The actual data fields of the child
5308 * DB are all zero size.
5315 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5316 if (flags & MDB_CURRENT) {
5317 xflags = MDB_CURRENT;
5319 mdb_xcursor_init1(mc, leaf);
5320 xflags = (flags & MDB_NODUPDATA) ? MDB_NOOVERWRITE : 0;
5322 /* converted, write the original data first */
5324 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, &dkey, &xdata, xflags);
5328 /* Adjust other cursors pointing to mp */
5330 unsigned i = mc->mc_top;
5331 MDB_page *mp = mc->mc_pg[i];
5333 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
5334 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
5335 if (m2->mc_pg[i] == mp && m2->mc_ki[i] == mc->mc_ki[i]) {
5336 mdb_xcursor_init1(m2, leaf);
5340 /* we've done our job */
5343 if (flags & MDB_APPENDDUP)
5344 xflags |= MDB_APPEND;
5345 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, data, &xdata, xflags);
5346 if (flags & F_SUBDATA) {
5347 void *db = NODEDATA(leaf);
5348 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
5351 /* sub-writes might have failed so check rc again.
5352 * Don't increment count if we just replaced an existing item.
5354 if (!rc && !(flags & MDB_CURRENT))
5355 mc->mc_db->md_entries++;
5356 if (flags & MDB_MULTIPLE) {
5358 if (mcount < data[1].mv_size) {
5359 data[0].mv_data = (char *)data[0].mv_data + data[0].mv_size;
5360 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5366 /* If we succeeded and the key didn't exist before, make sure
5367 * the cursor is marked valid.
5370 mc->mc_flags |= C_INITIALIZED;
5375 mdb_cursor_del(MDB_cursor *mc, unsigned int flags)
5380 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_RDONLY))
5383 if (!(mc->mc_flags & C_INITIALIZED))
5386 rc = mdb_cursor_touch(mc);
5390 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5392 if (!IS_LEAF2(mc->mc_pg[mc->mc_top]) && F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5393 if (flags != MDB_NODUPDATA) {
5394 if (!F_ISSET(leaf->mn_flags, F_SUBDATA)) {
5395 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
5397 rc = mdb_cursor_del(&mc->mc_xcursor->mx_cursor, 0);
5398 /* If sub-DB still has entries, we're done */
5399 if (mc->mc_xcursor->mx_db.md_entries) {
5400 if (leaf->mn_flags & F_SUBDATA) {
5401 /* update subDB info */
5402 void *db = NODEDATA(leaf);
5403 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
5405 /* shrink fake page */
5406 mdb_node_shrink(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5408 mc->mc_db->md_entries--;
5411 /* otherwise fall thru and delete the sub-DB */
5414 if (leaf->mn_flags & F_SUBDATA) {
5415 /* add all the child DB's pages to the free list */
5416 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
5417 if (rc == MDB_SUCCESS) {
5418 mc->mc_db->md_entries -=
5419 mc->mc_xcursor->mx_db.md_entries;
5424 return mdb_cursor_del0(mc, leaf);
5427 /** Allocate and initialize new pages for a database.
5428 * @param[in] mc a cursor on the database being added to.
5429 * @param[in] flags flags defining what type of page is being allocated.
5430 * @param[in] num the number of pages to allocate. This is usually 1,
5431 * unless allocating overflow pages for a large record.
5432 * @param[out] mp Address of a page, or NULL on failure.
5433 * @return 0 on success, non-zero on failure.
5436 mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp)
5441 if ((rc = mdb_page_alloc(mc, num, &np)))
5443 DPRINTF("allocated new mpage %zu, page size %u",
5444 np->mp_pgno, mc->mc_txn->mt_env->me_psize);
5445 np->mp_flags = flags | P_DIRTY;
5446 np->mp_lower = PAGEHDRSZ;
5447 np->mp_upper = mc->mc_txn->mt_env->me_psize;
5450 mc->mc_db->md_branch_pages++;
5451 else if (IS_LEAF(np))
5452 mc->mc_db->md_leaf_pages++;
5453 else if (IS_OVERFLOW(np)) {
5454 mc->mc_db->md_overflow_pages += num;
5462 /** Calculate the size of a leaf node.
5463 * The size depends on the environment's page size; if a data item
5464 * is too large it will be put onto an overflow page and the node
5465 * size will only include the key and not the data. Sizes are always
5466 * rounded up to an even number of bytes, to guarantee 2-byte alignment
5467 * of the #MDB_node headers.
5468 * @param[in] env The environment handle.
5469 * @param[in] key The key for the node.
5470 * @param[in] data The data for the node.
5471 * @return The number of bytes needed to store the node.
5474 mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data)
5478 sz = LEAFSIZE(key, data);
5479 if (sz >= env->me_nodemax) {
5480 /* put on overflow page */
5481 sz -= data->mv_size - sizeof(pgno_t);
5485 return sz + sizeof(indx_t);
5488 /** Calculate the size of a branch node.
5489 * The size should depend on the environment's page size but since
5490 * we currently don't support spilling large keys onto overflow
5491 * pages, it's simply the size of the #MDB_node header plus the
5492 * size of the key. Sizes are always rounded up to an even number
5493 * of bytes, to guarantee 2-byte alignment of the #MDB_node headers.
5494 * @param[in] env The environment handle.
5495 * @param[in] key The key for the node.
5496 * @return The number of bytes needed to store the node.
5499 mdb_branch_size(MDB_env *env, MDB_val *key)
5504 if (sz >= env->me_nodemax) {
5505 /* put on overflow page */
5506 /* not implemented */
5507 /* sz -= key->size - sizeof(pgno_t); */
5510 return sz + sizeof(indx_t);
5513 /** Add a node to the page pointed to by the cursor.
5514 * @param[in] mc The cursor for this operation.
5515 * @param[in] indx The index on the page where the new node should be added.
5516 * @param[in] key The key for the new node.
5517 * @param[in] data The data for the new node, if any.
5518 * @param[in] pgno The page number, if adding a branch node.
5519 * @param[in] flags Flags for the node.
5520 * @return 0 on success, non-zero on failure. Possible errors are:
5522 * <li>ENOMEM - failed to allocate overflow pages for the node.
5523 * <li>MDB_PAGE_FULL - there is insufficient room in the page. This error
5524 * should never happen since all callers already calculate the
5525 * page's free space before calling this function.
5529 mdb_node_add(MDB_cursor *mc, indx_t indx,
5530 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags)
5533 size_t node_size = NODESIZE;
5536 MDB_page *mp = mc->mc_pg[mc->mc_top];
5537 MDB_page *ofp = NULL; /* overflow page */
5540 assert(mp->mp_upper >= mp->mp_lower);
5542 DPRINTF("add to %s %spage %zu index %i, data size %zu key size %zu [%s]",
5543 IS_LEAF(mp) ? "leaf" : "branch",
5544 IS_SUBP(mp) ? "sub-" : "",
5545 mp->mp_pgno, indx, data ? data->mv_size : 0,
5546 key ? key->mv_size : 0, key ? DKEY(key) : NULL);
5549 /* Move higher keys up one slot. */
5550 int ksize = mc->mc_db->md_pad, dif;
5551 char *ptr = LEAF2KEY(mp, indx, ksize);
5552 dif = NUMKEYS(mp) - indx;
5554 memmove(ptr+ksize, ptr, dif*ksize);
5555 /* insert new key */
5556 memcpy(ptr, key->mv_data, ksize);
5558 /* Just using these for counting */
5559 mp->mp_lower += sizeof(indx_t);
5560 mp->mp_upper -= ksize - sizeof(indx_t);
5565 node_size += key->mv_size;
5569 if (F_ISSET(flags, F_BIGDATA)) {
5570 /* Data already on overflow page. */
5571 node_size += sizeof(pgno_t);
5572 } else if (node_size + data->mv_size >= mc->mc_txn->mt_env->me_nodemax) {
5573 int ovpages = OVPAGES(data->mv_size, mc->mc_txn->mt_env->me_psize);
5575 /* Put data on overflow page. */
5576 DPRINTF("data size is %zu, node would be %zu, put data on overflow page",
5577 data->mv_size, node_size+data->mv_size);
5578 node_size += sizeof(pgno_t);
5579 if ((rc = mdb_page_new(mc, P_OVERFLOW, ovpages, &ofp)))
5581 DPRINTF("allocated overflow page %zu", ofp->mp_pgno);
5584 node_size += data->mv_size;
5587 node_size += node_size & 1;
5589 if (node_size + sizeof(indx_t) > SIZELEFT(mp)) {
5590 DPRINTF("not enough room in page %zu, got %u ptrs",
5591 mp->mp_pgno, NUMKEYS(mp));
5592 DPRINTF("upper - lower = %u - %u = %u", mp->mp_upper, mp->mp_lower,
5593 mp->mp_upper - mp->mp_lower);
5594 DPRINTF("node size = %zu", node_size);
5595 return MDB_PAGE_FULL;
5598 /* Move higher pointers up one slot. */
5599 for (i = NUMKEYS(mp); i > indx; i--)
5600 mp->mp_ptrs[i] = mp->mp_ptrs[i - 1];
5602 /* Adjust free space offsets. */
5603 ofs = mp->mp_upper - node_size;
5604 assert(ofs >= mp->mp_lower + sizeof(indx_t));
5605 mp->mp_ptrs[indx] = ofs;
5607 mp->mp_lower += sizeof(indx_t);
5609 /* Write the node data. */
5610 node = NODEPTR(mp, indx);
5611 node->mn_ksize = (key == NULL) ? 0 : key->mv_size;
5612 node->mn_flags = flags;
5614 SETDSZ(node,data->mv_size);
5619 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
5624 if (F_ISSET(flags, F_BIGDATA))
5625 memcpy(node->mn_data + key->mv_size, data->mv_data,
5627 else if (F_ISSET(flags, MDB_RESERVE))
5628 data->mv_data = node->mn_data + key->mv_size;
5630 memcpy(node->mn_data + key->mv_size, data->mv_data,
5633 memcpy(node->mn_data + key->mv_size, &ofp->mp_pgno,
5635 if (F_ISSET(flags, MDB_RESERVE))
5636 data->mv_data = METADATA(ofp);
5638 memcpy(METADATA(ofp), data->mv_data, data->mv_size);
5645 /** Delete the specified node from a page.
5646 * @param[in] mp The page to operate on.
5647 * @param[in] indx The index of the node to delete.
5648 * @param[in] ksize The size of a node. Only used if the page is
5649 * part of a #MDB_DUPFIXED database.
5652 mdb_node_del(MDB_page *mp, indx_t indx, int ksize)
5655 indx_t i, j, numkeys, ptr;
5662 COPY_PGNO(pgno, mp->mp_pgno);
5663 DPRINTF("delete node %u on %s page %zu", indx,
5664 IS_LEAF(mp) ? "leaf" : "branch", pgno);
5667 assert(indx < NUMKEYS(mp));
5670 int x = NUMKEYS(mp) - 1 - indx;
5671 base = LEAF2KEY(mp, indx, ksize);
5673 memmove(base, base + ksize, x * ksize);
5674 mp->mp_lower -= sizeof(indx_t);
5675 mp->mp_upper += ksize - sizeof(indx_t);
5679 node = NODEPTR(mp, indx);
5680 sz = NODESIZE + node->mn_ksize;
5682 if (F_ISSET(node->mn_flags, F_BIGDATA))
5683 sz += sizeof(pgno_t);
5685 sz += NODEDSZ(node);
5689 ptr = mp->mp_ptrs[indx];
5690 numkeys = NUMKEYS(mp);
5691 for (i = j = 0; i < numkeys; i++) {
5693 mp->mp_ptrs[j] = mp->mp_ptrs[i];
5694 if (mp->mp_ptrs[i] < ptr)
5695 mp->mp_ptrs[j] += sz;
5700 base = (char *)mp + mp->mp_upper;
5701 memmove(base + sz, base, ptr - mp->mp_upper);
5703 mp->mp_lower -= sizeof(indx_t);
5707 /** Compact the main page after deleting a node on a subpage.
5708 * @param[in] mp The main page to operate on.
5709 * @param[in] indx The index of the subpage on the main page.
5712 mdb_node_shrink(MDB_page *mp, indx_t indx)
5719 indx_t i, numkeys, ptr;
5721 node = NODEPTR(mp, indx);
5722 sp = (MDB_page *)NODEDATA(node);
5723 osize = NODEDSZ(node);
5725 delta = sp->mp_upper - sp->mp_lower;
5726 SETDSZ(node, osize - delta);
5727 xp = (MDB_page *)((char *)sp + delta);
5729 /* shift subpage upward */
5731 nsize = NUMKEYS(sp) * sp->mp_pad;
5732 memmove(METADATA(xp), METADATA(sp), nsize);
5735 nsize = osize - sp->mp_upper;
5736 numkeys = NUMKEYS(sp);
5737 for (i=numkeys-1; i>=0; i--)
5738 xp->mp_ptrs[i] = sp->mp_ptrs[i] - delta;
5740 xp->mp_upper = sp->mp_lower;
5741 xp->mp_lower = sp->mp_lower;
5742 xp->mp_flags = sp->mp_flags;
5743 xp->mp_pad = sp->mp_pad;
5744 COPY_PGNO(xp->mp_pgno, mp->mp_pgno);
5746 /* shift lower nodes upward */
5747 ptr = mp->mp_ptrs[indx];
5748 numkeys = NUMKEYS(mp);
5749 for (i = 0; i < numkeys; i++) {
5750 if (mp->mp_ptrs[i] <= ptr)
5751 mp->mp_ptrs[i] += delta;
5754 base = (char *)mp + mp->mp_upper;
5755 memmove(base + delta, base, ptr - mp->mp_upper + NODESIZE + NODEKSZ(node));
5756 mp->mp_upper += delta;
5759 /** Initial setup of a sorted-dups cursor.
5760 * Sorted duplicates are implemented as a sub-database for the given key.
5761 * The duplicate data items are actually keys of the sub-database.
5762 * Operations on the duplicate data items are performed using a sub-cursor
5763 * initialized when the sub-database is first accessed. This function does
5764 * the preliminary setup of the sub-cursor, filling in the fields that
5765 * depend only on the parent DB.
5766 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
5769 mdb_xcursor_init0(MDB_cursor *mc)
5771 MDB_xcursor *mx = mc->mc_xcursor;
5773 mx->mx_cursor.mc_xcursor = NULL;
5774 mx->mx_cursor.mc_txn = mc->mc_txn;
5775 mx->mx_cursor.mc_db = &mx->mx_db;
5776 mx->mx_cursor.mc_dbx = &mx->mx_dbx;
5777 mx->mx_cursor.mc_dbi = mc->mc_dbi+1;
5778 mx->mx_cursor.mc_dbflag = &mx->mx_dbflag;
5779 mx->mx_cursor.mc_snum = 0;
5780 mx->mx_cursor.mc_top = 0;
5781 mx->mx_cursor.mc_flags = C_SUB;
5782 mx->mx_dbx.md_cmp = mc->mc_dbx->md_dcmp;
5783 mx->mx_dbx.md_dcmp = NULL;
5784 mx->mx_dbx.md_rel = mc->mc_dbx->md_rel;
5787 /** Final setup of a sorted-dups cursor.
5788 * Sets up the fields that depend on the data from the main cursor.
5789 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
5790 * @param[in] node The data containing the #MDB_db record for the
5791 * sorted-dup database.
5794 mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node)
5796 MDB_xcursor *mx = mc->mc_xcursor;
5798 if (node->mn_flags & F_SUBDATA) {
5799 memcpy(&mx->mx_db, NODEDATA(node), sizeof(MDB_db));
5800 mx->mx_cursor.mc_pg[0] = 0;
5801 mx->mx_cursor.mc_snum = 0;
5802 mx->mx_cursor.mc_flags = C_SUB;
5804 MDB_page *fp = NODEDATA(node);
5805 mx->mx_db.md_pad = mc->mc_pg[mc->mc_top]->mp_pad;
5806 mx->mx_db.md_flags = 0;
5807 mx->mx_db.md_depth = 1;
5808 mx->mx_db.md_branch_pages = 0;
5809 mx->mx_db.md_leaf_pages = 1;
5810 mx->mx_db.md_overflow_pages = 0;
5811 mx->mx_db.md_entries = NUMKEYS(fp);
5812 COPY_PGNO(mx->mx_db.md_root, fp->mp_pgno);
5813 mx->mx_cursor.mc_snum = 1;
5814 mx->mx_cursor.mc_flags = C_INITIALIZED|C_SUB;
5815 mx->mx_cursor.mc_top = 0;
5816 mx->mx_cursor.mc_pg[0] = fp;
5817 mx->mx_cursor.mc_ki[0] = 0;
5818 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
5819 mx->mx_db.md_flags = MDB_DUPFIXED;
5820 mx->mx_db.md_pad = fp->mp_pad;
5821 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
5822 mx->mx_db.md_flags |= MDB_INTEGERKEY;
5825 DPRINTF("Sub-db %u for db %u root page %zu", mx->mx_cursor.mc_dbi, mc->mc_dbi,
5827 mx->mx_dbflag = DB_VALID | (F_ISSET(mc->mc_pg[mc->mc_top]->mp_flags, P_DIRTY) ?
5829 mx->mx_dbx.md_name.mv_data = NODEKEY(node);
5830 mx->mx_dbx.md_name.mv_size = node->mn_ksize;
5831 #if UINT_MAX < SIZE_MAX
5832 if (mx->mx_dbx.md_cmp == mdb_cmp_int && mx->mx_db.md_pad == sizeof(size_t))
5833 #ifdef MISALIGNED_OK
5834 mx->mx_dbx.md_cmp = mdb_cmp_long;
5836 mx->mx_dbx.md_cmp = mdb_cmp_cint;
5841 /** Initialize a cursor for a given transaction and database. */
5843 mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx)
5848 mc->mc_db = &txn->mt_dbs[dbi];
5849 mc->mc_dbx = &txn->mt_dbxs[dbi];
5850 mc->mc_dbflag = &txn->mt_dbflags[dbi];
5855 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
5857 mc->mc_xcursor = mx;
5858 mdb_xcursor_init0(mc);
5860 mc->mc_xcursor = NULL;
5862 if (*mc->mc_dbflag & DB_STALE) {
5863 mdb_page_search(mc, NULL, MDB_PS_ROOTONLY);
5868 mdb_cursor_open(MDB_txn *txn, MDB_dbi dbi, MDB_cursor **ret)
5871 size_t size = sizeof(MDB_cursor);
5873 if (txn == NULL || ret == NULL || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
5876 /* Allow read access to the freelist */
5877 if (!dbi && !F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
5880 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT)
5881 size += sizeof(MDB_xcursor);
5883 if ((mc = malloc(size)) != NULL) {
5884 mdb_cursor_init(mc, txn, dbi, (MDB_xcursor *)(mc + 1));
5885 if (txn->mt_cursors) {
5886 mc->mc_next = txn->mt_cursors[dbi];
5887 txn->mt_cursors[dbi] = mc;
5888 mc->mc_flags |= C_UNTRACK;
5890 mc->mc_flags |= C_ALLOCD;
5901 mdb_cursor_renew(MDB_txn *txn, MDB_cursor *mc)
5905 if (txn == NULL || mc == NULL || mc->mc_dbi >= txn->mt_numdbs)
5908 if ((mc->mc_flags & C_UNTRACK) || txn->mt_cursors)
5911 flags = mc->mc_flags;
5913 mdb_cursor_init(mc, txn, mc->mc_dbi, mc->mc_xcursor);
5915 mc->mc_flags |= (flags & C_ALLOCD);
5919 /* Return the count of duplicate data items for the current key */
5921 mdb_cursor_count(MDB_cursor *mc, size_t *countp)
5925 if (mc == NULL || countp == NULL)
5928 if (!(mc->mc_db->md_flags & MDB_DUPSORT))
5931 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5932 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5935 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED))
5938 *countp = mc->mc_xcursor->mx_db.md_entries;
5944 mdb_cursor_close(MDB_cursor *mc)
5947 /* remove from txn, if tracked */
5948 if ((mc->mc_flags & C_UNTRACK) && mc->mc_txn->mt_cursors) {
5949 MDB_cursor **prev = &mc->mc_txn->mt_cursors[mc->mc_dbi];
5950 while (*prev && *prev != mc) prev = &(*prev)->mc_next;
5952 *prev = mc->mc_next;
5954 if (mc->mc_flags & C_ALLOCD)
5960 mdb_cursor_txn(MDB_cursor *mc)
5962 if (!mc) return NULL;
5967 mdb_cursor_dbi(MDB_cursor *mc)
5973 /** Replace the key for a node with a new key.
5974 * @param[in] mc Cursor pointing to the node to operate on.
5975 * @param[in] key The new key to use.
5976 * @return 0 on success, non-zero on failure.
5979 mdb_update_key(MDB_cursor *mc, MDB_val *key)
5986 indx_t ptr, i, numkeys, indx;
5989 indx = mc->mc_ki[mc->mc_top];
5990 mp = mc->mc_pg[mc->mc_top];
5991 node = NODEPTR(mp, indx);
5992 ptr = mp->mp_ptrs[indx];
5996 char kbuf2[(MDB_MAXKEYSIZE*2+1)];
5997 k2.mv_data = NODEKEY(node);
5998 k2.mv_size = node->mn_ksize;
5999 DPRINTF("update key %u (ofs %u) [%s] to [%s] on page %zu",
6001 mdb_dkey(&k2, kbuf2),
6007 delta0 = delta = key->mv_size - node->mn_ksize;
6009 /* Must be 2-byte aligned. If new key is
6010 * shorter by 1, the shift will be skipped.
6012 delta += (delta & 1);
6014 if (delta > 0 && SIZELEFT(mp) < delta) {
6016 /* not enough space left, do a delete and split */
6017 DPRINTF("Not enough room, delta = %d, splitting...", delta);
6018 pgno = NODEPGNO(node);
6019 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
6020 return mdb_page_split(mc, key, NULL, pgno, MDB_SPLIT_REPLACE);
6023 numkeys = NUMKEYS(mp);
6024 for (i = 0; i < numkeys; i++) {
6025 if (mp->mp_ptrs[i] <= ptr)
6026 mp->mp_ptrs[i] -= delta;
6029 base = (char *)mp + mp->mp_upper;
6030 len = ptr - mp->mp_upper + NODESIZE;
6031 memmove(base - delta, base, len);
6032 mp->mp_upper -= delta;
6034 node = NODEPTR(mp, indx);
6037 /* But even if no shift was needed, update ksize */
6039 node->mn_ksize = key->mv_size;
6042 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
6048 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst);
6050 /** Move a node from csrc to cdst.
6053 mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst)
6060 unsigned short flags;
6064 /* Mark src and dst as dirty. */
6065 if ((rc = mdb_page_touch(csrc)) ||
6066 (rc = mdb_page_touch(cdst)))
6069 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6070 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0); /* fake */
6071 key.mv_size = csrc->mc_db->md_pad;
6072 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
6074 data.mv_data = NULL;
6078 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top]);
6079 assert(!((long)srcnode&1));
6080 srcpg = NODEPGNO(srcnode);
6081 flags = srcnode->mn_flags;
6082 if (csrc->mc_ki[csrc->mc_top] == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
6083 unsigned int snum = csrc->mc_snum;
6085 /* must find the lowest key below src */
6086 mdb_page_search_lowest(csrc);
6087 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6088 key.mv_size = csrc->mc_db->md_pad;
6089 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
6091 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
6092 key.mv_size = NODEKSZ(s2);
6093 key.mv_data = NODEKEY(s2);
6095 csrc->mc_snum = snum--;
6096 csrc->mc_top = snum;
6098 key.mv_size = NODEKSZ(srcnode);
6099 key.mv_data = NODEKEY(srcnode);
6101 data.mv_size = NODEDSZ(srcnode);
6102 data.mv_data = NODEDATA(srcnode);
6104 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top]) && cdst->mc_ki[cdst->mc_top] == 0) {
6105 unsigned int snum = cdst->mc_snum;
6108 /* must find the lowest key below dst */
6109 mdb_page_search_lowest(cdst);
6110 if (IS_LEAF2(cdst->mc_pg[cdst->mc_top])) {
6111 bkey.mv_size = cdst->mc_db->md_pad;
6112 bkey.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, bkey.mv_size);
6114 s2 = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
6115 bkey.mv_size = NODEKSZ(s2);
6116 bkey.mv_data = NODEKEY(s2);
6118 cdst->mc_snum = snum--;
6119 cdst->mc_top = snum;
6120 mdb_cursor_copy(cdst, &mn);
6122 rc = mdb_update_key(&mn, &bkey);
6127 DPRINTF("moving %s node %u [%s] on page %zu to node %u on page %zu",
6128 IS_LEAF(csrc->mc_pg[csrc->mc_top]) ? "leaf" : "branch",
6129 csrc->mc_ki[csrc->mc_top],
6131 csrc->mc_pg[csrc->mc_top]->mp_pgno,
6132 cdst->mc_ki[cdst->mc_top], cdst->mc_pg[cdst->mc_top]->mp_pgno);
6134 /* Add the node to the destination page.
6136 rc = mdb_node_add(cdst, cdst->mc_ki[cdst->mc_top], &key, &data, srcpg, flags);
6137 if (rc != MDB_SUCCESS)
6140 /* Delete the node from the source page.
6142 mdb_node_del(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
6145 /* Adjust other cursors pointing to mp */
6146 MDB_cursor *m2, *m3;
6147 MDB_dbi dbi = csrc->mc_dbi;
6148 MDB_page *mp = csrc->mc_pg[csrc->mc_top];
6150 if (csrc->mc_flags & C_SUB)
6153 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6154 if (m2 == csrc) continue;
6155 if (csrc->mc_flags & C_SUB)
6156 m3 = &m2->mc_xcursor->mx_cursor;
6159 if (m3->mc_pg[csrc->mc_top] == mp && m3->mc_ki[csrc->mc_top] ==
6160 csrc->mc_ki[csrc->mc_top]) {
6161 m3->mc_pg[csrc->mc_top] = cdst->mc_pg[cdst->mc_top];
6162 m3->mc_ki[csrc->mc_top] = cdst->mc_ki[cdst->mc_top];
6167 /* Update the parent separators.
6169 if (csrc->mc_ki[csrc->mc_top] == 0) {
6170 if (csrc->mc_ki[csrc->mc_top-1] != 0) {
6171 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6172 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
6174 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
6175 key.mv_size = NODEKSZ(srcnode);
6176 key.mv_data = NODEKEY(srcnode);
6178 DPRINTF("update separator for source page %zu to [%s]",
6179 csrc->mc_pg[csrc->mc_top]->mp_pgno, DKEY(&key));
6180 mdb_cursor_copy(csrc, &mn);
6183 if ((rc = mdb_update_key(&mn, &key)) != MDB_SUCCESS)
6186 if (IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
6188 indx_t ix = csrc->mc_ki[csrc->mc_top];
6189 nullkey.mv_size = 0;
6190 csrc->mc_ki[csrc->mc_top] = 0;
6191 rc = mdb_update_key(csrc, &nullkey);
6192 csrc->mc_ki[csrc->mc_top] = ix;
6193 assert(rc == MDB_SUCCESS);
6197 if (cdst->mc_ki[cdst->mc_top] == 0) {
6198 if (cdst->mc_ki[cdst->mc_top-1] != 0) {
6199 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6200 key.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, key.mv_size);
6202 srcnode = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
6203 key.mv_size = NODEKSZ(srcnode);
6204 key.mv_data = NODEKEY(srcnode);
6206 DPRINTF("update separator for destination page %zu to [%s]",
6207 cdst->mc_pg[cdst->mc_top]->mp_pgno, DKEY(&key));
6208 mdb_cursor_copy(cdst, &mn);
6211 if ((rc = mdb_update_key(&mn, &key)) != MDB_SUCCESS)
6214 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top])) {
6216 indx_t ix = cdst->mc_ki[cdst->mc_top];
6217 nullkey.mv_size = 0;
6218 cdst->mc_ki[cdst->mc_top] = 0;
6219 rc = mdb_update_key(cdst, &nullkey);
6220 cdst->mc_ki[cdst->mc_top] = ix;
6221 assert(rc == MDB_SUCCESS);
6228 /** Merge one page into another.
6229 * The nodes from the page pointed to by \b csrc will
6230 * be copied to the page pointed to by \b cdst and then
6231 * the \b csrc page will be freed.
6232 * @param[in] csrc Cursor pointing to the source page.
6233 * @param[in] cdst Cursor pointing to the destination page.
6236 mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst)
6244 DPRINTF("merging page %zu into %zu", csrc->mc_pg[csrc->mc_top]->mp_pgno,
6245 cdst->mc_pg[cdst->mc_top]->mp_pgno);
6247 assert(csrc->mc_snum > 1); /* can't merge root page */
6248 assert(cdst->mc_snum > 1);
6250 /* Mark dst as dirty. */
6251 if ((rc = mdb_page_touch(cdst)))
6254 /* Move all nodes from src to dst.
6256 j = nkeys = NUMKEYS(cdst->mc_pg[cdst->mc_top]);
6257 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6258 key.mv_size = csrc->mc_db->md_pad;
6259 key.mv_data = METADATA(csrc->mc_pg[csrc->mc_top]);
6260 for (i = 0; i < NUMKEYS(csrc->mc_pg[csrc->mc_top]); i++, j++) {
6261 rc = mdb_node_add(cdst, j, &key, NULL, 0, 0);
6262 if (rc != MDB_SUCCESS)
6264 key.mv_data = (char *)key.mv_data + key.mv_size;
6267 for (i = 0; i < NUMKEYS(csrc->mc_pg[csrc->mc_top]); i++, j++) {
6268 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], i);
6269 if (i == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
6270 unsigned int snum = csrc->mc_snum;
6272 /* must find the lowest key below src */
6273 mdb_page_search_lowest(csrc);
6274 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6275 key.mv_size = csrc->mc_db->md_pad;
6276 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
6278 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
6279 key.mv_size = NODEKSZ(s2);
6280 key.mv_data = NODEKEY(s2);
6282 csrc->mc_snum = snum--;
6283 csrc->mc_top = snum;
6285 key.mv_size = srcnode->mn_ksize;
6286 key.mv_data = NODEKEY(srcnode);
6289 data.mv_size = NODEDSZ(srcnode);
6290 data.mv_data = NODEDATA(srcnode);
6291 rc = mdb_node_add(cdst, j, &key, &data, NODEPGNO(srcnode), srcnode->mn_flags);
6292 if (rc != MDB_SUCCESS)
6297 DPRINTF("dst page %zu now has %u keys (%.1f%% filled)",
6298 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);
6300 /* Unlink the src page from parent and add to free list.
6302 mdb_node_del(csrc->mc_pg[csrc->mc_top-1], csrc->mc_ki[csrc->mc_top-1], 0);
6303 if (csrc->mc_ki[csrc->mc_top-1] == 0) {
6306 rc = mdb_update_key(csrc, &key);
6312 mdb_midl_append(&csrc->mc_txn->mt_free_pgs, csrc->mc_pg[csrc->mc_top]->mp_pgno);
6313 if (IS_LEAF(csrc->mc_pg[csrc->mc_top]))
6314 csrc->mc_db->md_leaf_pages--;
6316 csrc->mc_db->md_branch_pages--;
6318 /* Adjust other cursors pointing to mp */
6319 MDB_cursor *m2, *m3;
6320 MDB_dbi dbi = csrc->mc_dbi;
6321 MDB_page *mp = cdst->mc_pg[cdst->mc_top];
6323 if (csrc->mc_flags & C_SUB)
6326 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6327 if (csrc->mc_flags & C_SUB)
6328 m3 = &m2->mc_xcursor->mx_cursor;
6331 if (m3 == csrc) continue;
6332 if (m3->mc_snum < csrc->mc_snum) continue;
6333 if (m3->mc_pg[csrc->mc_top] == csrc->mc_pg[csrc->mc_top]) {
6334 m3->mc_pg[csrc->mc_top] = mp;
6335 m3->mc_ki[csrc->mc_top] += nkeys;
6339 mdb_cursor_pop(csrc);
6341 return mdb_rebalance(csrc);
6344 /** Copy the contents of a cursor.
6345 * @param[in] csrc The cursor to copy from.
6346 * @param[out] cdst The cursor to copy to.
6349 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst)
6353 cdst->mc_txn = csrc->mc_txn;
6354 cdst->mc_dbi = csrc->mc_dbi;
6355 cdst->mc_db = csrc->mc_db;
6356 cdst->mc_dbx = csrc->mc_dbx;
6357 cdst->mc_snum = csrc->mc_snum;
6358 cdst->mc_top = csrc->mc_top;
6359 cdst->mc_flags = csrc->mc_flags;
6361 for (i=0; i<csrc->mc_snum; i++) {
6362 cdst->mc_pg[i] = csrc->mc_pg[i];
6363 cdst->mc_ki[i] = csrc->mc_ki[i];
6367 /** Rebalance the tree after a delete operation.
6368 * @param[in] mc Cursor pointing to the page where rebalancing
6370 * @return 0 on success, non-zero on failure.
6373 mdb_rebalance(MDB_cursor *mc)
6377 unsigned int ptop, minkeys;
6380 minkeys = 1 + (IS_BRANCH(mc->mc_pg[mc->mc_top]));
6384 COPY_PGNO(pgno, mc->mc_pg[mc->mc_top]->mp_pgno);
6385 DPRINTF("rebalancing %s page %zu (has %u keys, %.1f%% full)",
6386 IS_LEAF(mc->mc_pg[mc->mc_top]) ? "leaf" : "branch",
6387 pgno, NUMKEYS(mc->mc_pg[mc->mc_top]), (float)PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) / 10);
6391 if (PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) >= FILL_THRESHOLD &&
6392 NUMKEYS(mc->mc_pg[mc->mc_top]) >= minkeys) {
6395 COPY_PGNO(pgno, mc->mc_pg[mc->mc_top]->mp_pgno);
6396 DPRINTF("no need to rebalance page %zu, above fill threshold",
6402 if (mc->mc_snum < 2) {
6403 MDB_page *mp = mc->mc_pg[0];
6405 DPUTS("Can't rebalance a subpage, ignoring");
6408 if (NUMKEYS(mp) == 0) {
6409 DPUTS("tree is completely empty");
6410 mc->mc_db->md_root = P_INVALID;
6411 mc->mc_db->md_depth = 0;
6412 mc->mc_db->md_leaf_pages = 0;
6413 mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
6417 /* Adjust other cursors pointing to mp */
6418 MDB_cursor *m2, *m3;
6419 MDB_dbi dbi = mc->mc_dbi;
6421 if (mc->mc_flags & C_SUB)
6424 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6425 if (m2 == mc) continue;
6426 if (mc->mc_flags & C_SUB)
6427 m3 = &m2->mc_xcursor->mx_cursor;
6430 if (m3->mc_snum < mc->mc_snum) continue;
6431 if (m3->mc_pg[0] == mp) {
6437 } else if (IS_BRANCH(mp) && NUMKEYS(mp) == 1) {
6438 DPUTS("collapsing root page!");
6439 mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
6440 mc->mc_db->md_root = NODEPGNO(NODEPTR(mp, 0));
6441 rc = mdb_page_get(mc->mc_txn,mc->mc_db->md_root,&mc->mc_pg[0],NULL);
6444 mc->mc_db->md_depth--;
6445 mc->mc_db->md_branch_pages--;
6447 /* Adjust other cursors pointing to mp */
6448 MDB_cursor *m2, *m3;
6449 MDB_dbi dbi = mc->mc_dbi;
6451 if (mc->mc_flags & C_SUB)
6454 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6455 if (m2 == mc) continue;
6456 if (mc->mc_flags & C_SUB)
6457 m3 = &m2->mc_xcursor->mx_cursor;
6460 if (m3->mc_snum < mc->mc_snum) continue;
6461 if (m3->mc_pg[0] == mp) {
6462 m3->mc_pg[0] = mc->mc_pg[0];
6469 DPUTS("root page doesn't need rebalancing");
6473 /* The parent (branch page) must have at least 2 pointers,
6474 * otherwise the tree is invalid.
6476 ptop = mc->mc_top-1;
6477 assert(NUMKEYS(mc->mc_pg[ptop]) > 1);
6479 /* Leaf page fill factor is below the threshold.
6480 * Try to move keys from left or right neighbor, or
6481 * merge with a neighbor page.
6486 mdb_cursor_copy(mc, &mn);
6487 mn.mc_xcursor = NULL;
6489 if (mc->mc_ki[ptop] == 0) {
6490 /* We're the leftmost leaf in our parent.
6492 DPUTS("reading right neighbor");
6494 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
6495 rc = mdb_page_get(mc->mc_txn,NODEPGNO(node),&mn.mc_pg[mn.mc_top],NULL);
6498 mn.mc_ki[mn.mc_top] = 0;
6499 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]);
6501 /* There is at least one neighbor to the left.
6503 DPUTS("reading left neighbor");
6505 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
6506 rc = mdb_page_get(mc->mc_txn,NODEPGNO(node),&mn.mc_pg[mn.mc_top],NULL);
6509 mn.mc_ki[mn.mc_top] = NUMKEYS(mn.mc_pg[mn.mc_top]) - 1;
6510 mc->mc_ki[mc->mc_top] = 0;
6513 DPRINTF("found neighbor page %zu (%u keys, %.1f%% full)",
6514 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);
6516 /* If the neighbor page is above threshold and has enough keys,
6517 * move one key from it. Otherwise we should try to merge them.
6518 * (A branch page must never have less than 2 keys.)
6520 minkeys = 1 + (IS_BRANCH(mn.mc_pg[mn.mc_top]));
6521 if (PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) >= FILL_THRESHOLD && NUMKEYS(mn.mc_pg[mn.mc_top]) > minkeys)
6522 return mdb_node_move(&mn, mc);
6524 if (mc->mc_ki[ptop] == 0)
6525 rc = mdb_page_merge(&mn, mc);
6527 rc = mdb_page_merge(mc, &mn);
6528 mc->mc_flags &= ~C_INITIALIZED;
6533 /** Complete a delete operation started by #mdb_cursor_del(). */
6535 mdb_cursor_del0(MDB_cursor *mc, MDB_node *leaf)
6539 /* add overflow pages to free list */
6540 if (!IS_LEAF2(mc->mc_pg[mc->mc_top]) && F_ISSET(leaf->mn_flags, F_BIGDATA)) {
6545 memcpy(&pg, NODEDATA(leaf), sizeof(pg));
6546 if ((rc = mdb_page_get(mc->mc_txn, pg, &omp, NULL)) != 0)
6548 assert(IS_OVERFLOW(omp));
6549 ovpages = omp->mp_pages;
6550 mc->mc_db->md_overflow_pages -= ovpages;
6551 for (i=0; i<ovpages; i++) {
6552 DPRINTF("freed ov page %zu", pg);
6553 mdb_midl_append(&mc->mc_txn->mt_free_pgs, pg);
6557 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], mc->mc_db->md_pad);
6558 mc->mc_db->md_entries--;
6559 rc = mdb_rebalance(mc);
6560 if (rc != MDB_SUCCESS)
6561 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
6562 /* if mc points past last node in page, invalidate */
6563 else if (mc->mc_ki[mc->mc_top] >= NUMKEYS(mc->mc_pg[mc->mc_top]))
6564 mc->mc_flags &= ~C_INITIALIZED;
6570 mdb_del(MDB_txn *txn, MDB_dbi dbi,
6571 MDB_val *key, MDB_val *data)
6576 MDB_val rdata, *xdata;
6580 assert(key != NULL);
6582 DPRINTF("====> delete db %u key [%s]", dbi, DKEY(key));
6584 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
6587 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
6591 if (key->mv_size == 0 || key->mv_size > MDB_MAXKEYSIZE) {
6595 mdb_cursor_init(&mc, txn, dbi, &mx);
6606 rc = mdb_cursor_set(&mc, key, xdata, op, &exact);
6608 /* let mdb_page_split know about this cursor if needed:
6609 * delete will trigger a rebalance; if it needs to move
6610 * a node from one page to another, it will have to
6611 * update the parent's separator key(s). If the new sepkey
6612 * is larger than the current one, the parent page may
6613 * run out of space, triggering a split. We need this
6614 * cursor to be consistent until the end of the rebalance.
6616 mc.mc_next = txn->mt_cursors[dbi];
6617 txn->mt_cursors[dbi] = &mc;
6618 rc = mdb_cursor_del(&mc, data ? 0 : MDB_NODUPDATA);
6619 txn->mt_cursors[dbi] = mc.mc_next;
6624 /** Split a page and insert a new node.
6625 * @param[in,out] mc Cursor pointing to the page and desired insertion index.
6626 * The cursor will be updated to point to the actual page and index where
6627 * the node got inserted after the split.
6628 * @param[in] newkey The key for the newly inserted node.
6629 * @param[in] newdata The data for the newly inserted node.
6630 * @param[in] newpgno The page number, if the new node is a branch node.
6631 * @param[in] nflags The #NODE_ADD_FLAGS for the new node.
6632 * @return 0 on success, non-zero on failure.
6635 mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata, pgno_t newpgno,
6636 unsigned int nflags)
6639 int rc = MDB_SUCCESS, ins_new = 0, new_root = 0, newpos = 1, did_split = 0;
6642 unsigned int i, j, split_indx, nkeys, pmax;
6644 MDB_val sepkey, rkey, xdata, *rdata = &xdata;
6646 MDB_page *mp, *rp, *pp;
6651 mp = mc->mc_pg[mc->mc_top];
6652 newindx = mc->mc_ki[mc->mc_top];
6654 DPRINTF("-----> splitting %s page %zu and adding [%s] at index %i",
6655 IS_LEAF(mp) ? "leaf" : "branch", mp->mp_pgno,
6656 DKEY(newkey), mc->mc_ki[mc->mc_top]);
6658 /* Create a right sibling. */
6659 if ((rc = mdb_page_new(mc, mp->mp_flags, 1, &rp)))
6661 DPRINTF("new right sibling: page %zu", rp->mp_pgno);
6663 if (mc->mc_snum < 2) {
6664 if ((rc = mdb_page_new(mc, P_BRANCH, 1, &pp)))
6666 /* shift current top to make room for new parent */
6667 mc->mc_pg[1] = mc->mc_pg[0];
6668 mc->mc_ki[1] = mc->mc_ki[0];
6671 mc->mc_db->md_root = pp->mp_pgno;
6672 DPRINTF("root split! new root = %zu", pp->mp_pgno);
6673 mc->mc_db->md_depth++;
6676 /* Add left (implicit) pointer. */
6677 if ((rc = mdb_node_add(mc, 0, NULL, NULL, mp->mp_pgno, 0)) != MDB_SUCCESS) {
6678 /* undo the pre-push */
6679 mc->mc_pg[0] = mc->mc_pg[1];
6680 mc->mc_ki[0] = mc->mc_ki[1];
6681 mc->mc_db->md_root = mp->mp_pgno;
6682 mc->mc_db->md_depth--;
6689 ptop = mc->mc_top-1;
6690 DPRINTF("parent branch page is %zu", mc->mc_pg[ptop]->mp_pgno);
6693 mc->mc_flags |= C_SPLITTING;
6694 mdb_cursor_copy(mc, &mn);
6695 mn.mc_pg[mn.mc_top] = rp;
6696 mn.mc_ki[ptop] = mc->mc_ki[ptop]+1;
6698 if (nflags & MDB_APPEND) {
6699 mn.mc_ki[mn.mc_top] = 0;
6701 split_indx = newindx;
6706 nkeys = NUMKEYS(mp);
6707 split_indx = nkeys / 2;
6708 if (newindx < split_indx)
6714 unsigned int lsize, rsize, ksize;
6715 /* Move half of the keys to the right sibling */
6717 x = mc->mc_ki[mc->mc_top] - split_indx;
6718 ksize = mc->mc_db->md_pad;
6719 split = LEAF2KEY(mp, split_indx, ksize);
6720 rsize = (nkeys - split_indx) * ksize;
6721 lsize = (nkeys - split_indx) * sizeof(indx_t);
6722 mp->mp_lower -= lsize;
6723 rp->mp_lower += lsize;
6724 mp->mp_upper += rsize - lsize;
6725 rp->mp_upper -= rsize - lsize;
6726 sepkey.mv_size = ksize;
6727 if (newindx == split_indx) {
6728 sepkey.mv_data = newkey->mv_data;
6730 sepkey.mv_data = split;
6733 ins = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], ksize);
6734 memcpy(rp->mp_ptrs, split, rsize);
6735 sepkey.mv_data = rp->mp_ptrs;
6736 memmove(ins+ksize, ins, (split_indx - mc->mc_ki[mc->mc_top]) * ksize);
6737 memcpy(ins, newkey->mv_data, ksize);
6738 mp->mp_lower += sizeof(indx_t);
6739 mp->mp_upper -= ksize - sizeof(indx_t);
6742 memcpy(rp->mp_ptrs, split, x * ksize);
6743 ins = LEAF2KEY(rp, x, ksize);
6744 memcpy(ins, newkey->mv_data, ksize);
6745 memcpy(ins+ksize, split + x * ksize, rsize - x * ksize);
6746 rp->mp_lower += sizeof(indx_t);
6747 rp->mp_upper -= ksize - sizeof(indx_t);
6748 mc->mc_ki[mc->mc_top] = x;
6749 mc->mc_pg[mc->mc_top] = rp;
6754 /* For leaf pages, check the split point based on what
6755 * fits where, since otherwise mdb_node_add can fail.
6757 * This check is only needed when the data items are
6758 * relatively large, such that being off by one will
6759 * make the difference between success or failure.
6761 * It's also relevant if a page happens to be laid out
6762 * such that one half of its nodes are all "small" and
6763 * the other half of its nodes are "large." If the new
6764 * item is also "large" and falls on the half with
6765 * "large" nodes, it also may not fit.
6768 unsigned int psize, nsize;
6769 /* Maximum free space in an empty page */
6770 pmax = mc->mc_txn->mt_env->me_psize - PAGEHDRSZ;
6771 nsize = mdb_leaf_size(mc->mc_txn->mt_env, newkey, newdata);
6772 if ((nkeys < 20) || (nsize > pmax/16)) {
6773 if (newindx <= split_indx) {
6776 for (i=0; i<split_indx; i++) {
6777 node = NODEPTR(mp, i);
6778 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
6779 if (F_ISSET(node->mn_flags, F_BIGDATA))
6780 psize += sizeof(pgno_t);
6782 psize += NODEDSZ(node);
6786 split_indx = newindx;
6797 for (i=nkeys-1; i>=split_indx; i--) {
6798 node = NODEPTR(mp, i);
6799 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
6800 if (F_ISSET(node->mn_flags, F_BIGDATA))
6801 psize += sizeof(pgno_t);
6803 psize += NODEDSZ(node);
6807 split_indx = newindx;
6818 /* First find the separating key between the split pages.
6819 * The case where newindx == split_indx is ambiguous; the
6820 * new item could go to the new page or stay on the original
6821 * page. If newpos == 1 it goes to the new page.
6823 if (newindx == split_indx && newpos) {
6824 sepkey.mv_size = newkey->mv_size;
6825 sepkey.mv_data = newkey->mv_data;
6827 node = NODEPTR(mp, split_indx);
6828 sepkey.mv_size = node->mn_ksize;
6829 sepkey.mv_data = NODEKEY(node);
6833 DPRINTF("separator is [%s]", DKEY(&sepkey));
6835 /* Copy separator key to the parent.
6837 if (SIZELEFT(mn.mc_pg[ptop]) < mdb_branch_size(mc->mc_txn->mt_env, &sepkey)) {
6841 rc = mdb_page_split(&mn, &sepkey, NULL, rp->mp_pgno, 0);
6844 if (mn.mc_snum == mc->mc_snum) {
6845 mc->mc_pg[mc->mc_snum] = mc->mc_pg[mc->mc_top];
6846 mc->mc_ki[mc->mc_snum] = mc->mc_ki[mc->mc_top];
6847 mc->mc_pg[mc->mc_top] = mc->mc_pg[ptop];
6848 mc->mc_ki[mc->mc_top] = mc->mc_ki[ptop];
6853 /* Right page might now have changed parent.
6854 * Check if left page also changed parent.
6856 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
6857 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
6858 for (i=0; i<ptop; i++) {
6859 mc->mc_pg[i] = mn.mc_pg[i];
6860 mc->mc_ki[i] = mn.mc_ki[i];
6862 mc->mc_pg[ptop] = mn.mc_pg[ptop];
6863 mc->mc_ki[ptop] = mn.mc_ki[ptop] - 1;
6867 rc = mdb_node_add(&mn, mn.mc_ki[ptop], &sepkey, NULL, rp->mp_pgno, 0);
6870 mc->mc_flags ^= C_SPLITTING;
6871 if (rc != MDB_SUCCESS) {
6874 if (nflags & MDB_APPEND) {
6875 mc->mc_pg[mc->mc_top] = rp;
6876 mc->mc_ki[mc->mc_top] = 0;
6877 rc = mdb_node_add(mc, 0, newkey, newdata, newpgno, nflags);
6880 for (i=0; i<mc->mc_top; i++)
6881 mc->mc_ki[i] = mn.mc_ki[i];
6888 /* Move half of the keys to the right sibling. */
6890 /* grab a page to hold a temporary copy */
6891 copy = mdb_page_malloc(mc, 1);
6895 copy->mp_pgno = mp->mp_pgno;
6896 copy->mp_flags = mp->mp_flags;
6897 copy->mp_lower = PAGEHDRSZ;
6898 copy->mp_upper = mc->mc_txn->mt_env->me_psize;
6899 mc->mc_pg[mc->mc_top] = copy;
6900 for (i = j = 0; i <= nkeys; j++) {
6901 if (i == split_indx) {
6902 /* Insert in right sibling. */
6903 /* Reset insert index for right sibling. */
6904 if (i != newindx || (newpos ^ ins_new)) {
6906 mc->mc_pg[mc->mc_top] = rp;
6910 if (i == newindx && !ins_new) {
6911 /* Insert the original entry that caused the split. */
6912 rkey.mv_data = newkey->mv_data;
6913 rkey.mv_size = newkey->mv_size;
6922 /* Update index for the new key. */
6923 mc->mc_ki[mc->mc_top] = j;
6924 } else if (i == nkeys) {
6927 node = NODEPTR(mp, i);
6928 rkey.mv_data = NODEKEY(node);
6929 rkey.mv_size = node->mn_ksize;
6931 xdata.mv_data = NODEDATA(node);
6932 xdata.mv_size = NODEDSZ(node);
6935 pgno = NODEPGNO(node);
6936 flags = node->mn_flags;
6941 if (!IS_LEAF(mp) && j == 0) {
6942 /* First branch index doesn't need key data. */
6946 rc = mdb_node_add(mc, j, &rkey, rdata, pgno, flags);
6950 nkeys = NUMKEYS(copy);
6951 for (i=0; i<nkeys; i++)
6952 mp->mp_ptrs[i] = copy->mp_ptrs[i];
6953 mp->mp_lower = copy->mp_lower;
6954 mp->mp_upper = copy->mp_upper;
6955 memcpy(NODEPTR(mp, nkeys-1), NODEPTR(copy, nkeys-1),
6956 mc->mc_txn->mt_env->me_psize - copy->mp_upper);
6958 /* reset back to original page */
6959 if (newindx < split_indx || (!newpos && newindx == split_indx)) {
6960 mc->mc_pg[mc->mc_top] = mp;
6961 if (nflags & MDB_RESERVE) {
6962 node = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6963 if (!(node->mn_flags & F_BIGDATA))
6964 newdata->mv_data = NODEDATA(node);
6968 /* Make sure mc_ki is still valid.
6970 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
6971 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
6972 for (i=0; i<ptop; i++) {
6973 mc->mc_pg[i] = mn.mc_pg[i];
6974 mc->mc_ki[i] = mn.mc_ki[i];
6976 mc->mc_pg[ptop] = mn.mc_pg[ptop];
6977 mc->mc_ki[ptop] = mn.mc_ki[ptop] - 1;
6981 /* return tmp page to freelist */
6982 mdb_page_free(mc->mc_txn->mt_env, copy);
6985 /* Adjust other cursors pointing to mp */
6986 MDB_cursor *m2, *m3;
6987 MDB_dbi dbi = mc->mc_dbi;
6988 int fixup = NUMKEYS(mp);
6990 if (mc->mc_flags & C_SUB)
6993 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6994 if (m2 == mc) continue;
6995 if (mc->mc_flags & C_SUB)
6996 m3 = &m2->mc_xcursor->mx_cursor;
6999 if (!(m3->mc_flags & C_INITIALIZED))
7001 if (m3->mc_flags & C_SPLITTING)
7006 for (k=m3->mc_top; k>=0; k--) {
7007 m3->mc_ki[k+1] = m3->mc_ki[k];
7008 m3->mc_pg[k+1] = m3->mc_pg[k];
7010 if (m3->mc_ki[0] >= split_indx) {
7015 m3->mc_pg[0] = mc->mc_pg[0];
7019 if (m3->mc_pg[mc->mc_top] == mp) {
7020 if (m3->mc_ki[mc->mc_top] >= newindx && !(nflags & MDB_SPLIT_REPLACE))
7021 m3->mc_ki[mc->mc_top]++;
7022 if (m3->mc_ki[mc->mc_top] >= fixup) {
7023 m3->mc_pg[mc->mc_top] = rp;
7024 m3->mc_ki[mc->mc_top] -= fixup;
7025 m3->mc_ki[ptop] = mn.mc_ki[ptop];
7027 } else if (!did_split && m3->mc_pg[ptop] == mc->mc_pg[ptop] &&
7028 m3->mc_ki[ptop] >= mc->mc_ki[ptop]) {
7037 mdb_put(MDB_txn *txn, MDB_dbi dbi,
7038 MDB_val *key, MDB_val *data, unsigned int flags)
7043 assert(key != NULL);
7044 assert(data != NULL);
7046 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
7049 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
7053 if (key->mv_size == 0 || key->mv_size > MDB_MAXKEYSIZE) {
7057 if ((flags & (MDB_NOOVERWRITE|MDB_NODUPDATA|MDB_RESERVE|MDB_APPEND|MDB_APPENDDUP)) != flags)
7060 mdb_cursor_init(&mc, txn, dbi, &mx);
7061 return mdb_cursor_put(&mc, key, data, flags);
7065 mdb_env_set_flags(MDB_env *env, unsigned int flag, int onoff)
7067 if ((flag & CHANGEABLE) != flag)
7070 env->me_flags |= flag;
7072 env->me_flags &= ~flag;
7077 mdb_env_get_flags(MDB_env *env, unsigned int *arg)
7082 *arg = env->me_flags;
7087 mdb_env_get_path(MDB_env *env, const char **arg)
7092 *arg = env->me_path;
7096 /** Common code for #mdb_stat() and #mdb_env_stat().
7097 * @param[in] env the environment to operate in.
7098 * @param[in] db the #MDB_db record containing the stats to return.
7099 * @param[out] arg the address of an #MDB_stat structure to receive the stats.
7100 * @return 0, this function always succeeds.
7103 mdb_stat0(MDB_env *env, MDB_db *db, MDB_stat *arg)
7105 arg->ms_psize = env->me_psize;
7106 arg->ms_depth = db->md_depth;
7107 arg->ms_branch_pages = db->md_branch_pages;
7108 arg->ms_leaf_pages = db->md_leaf_pages;
7109 arg->ms_overflow_pages = db->md_overflow_pages;
7110 arg->ms_entries = db->md_entries;
7115 mdb_env_stat(MDB_env *env, MDB_stat *arg)
7119 if (env == NULL || arg == NULL)
7122 toggle = mdb_env_pick_meta(env);
7124 return mdb_stat0(env, &env->me_metas[toggle]->mm_dbs[MAIN_DBI], arg);
7128 mdb_env_info(MDB_env *env, MDB_envinfo *arg)
7132 if (env == NULL || arg == NULL)
7135 toggle = mdb_env_pick_meta(env);
7136 arg->me_mapaddr = (env->me_flags & MDB_FIXEDMAP) ? env->me_map : 0;
7137 arg->me_mapsize = env->me_mapsize;
7138 arg->me_maxreaders = env->me_maxreaders;
7139 arg->me_numreaders = env->me_numreaders;
7140 arg->me_last_pgno = env->me_metas[toggle]->mm_last_pg;
7141 arg->me_last_txnid = env->me_metas[toggle]->mm_txnid;
7145 /** Set the default comparison functions for a database.
7146 * Called immediately after a database is opened to set the defaults.
7147 * The user can then override them with #mdb_set_compare() or
7148 * #mdb_set_dupsort().
7149 * @param[in] txn A transaction handle returned by #mdb_txn_begin()
7150 * @param[in] dbi A database handle returned by #mdb_dbi_open()
7153 mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi)
7155 uint16_t f = txn->mt_dbs[dbi].md_flags;
7157 txn->mt_dbxs[dbi].md_cmp =
7158 (f & MDB_REVERSEKEY) ? mdb_cmp_memnr :
7159 (f & MDB_INTEGERKEY) ? mdb_cmp_cint : mdb_cmp_memn;
7161 txn->mt_dbxs[dbi].md_dcmp =
7162 !(f & MDB_DUPSORT) ? 0 :
7163 ((f & MDB_INTEGERDUP)
7164 ? ((f & MDB_DUPFIXED) ? mdb_cmp_int : mdb_cmp_cint)
7165 : ((f & MDB_REVERSEDUP) ? mdb_cmp_memnr : mdb_cmp_memn));
7168 int mdb_dbi_open(MDB_txn *txn, const char *name, unsigned int flags, MDB_dbi *dbi)
7173 int rc, dbflag, exact;
7174 unsigned int unused = 0;
7177 if (txn->mt_dbxs[FREE_DBI].md_cmp == NULL) {
7178 mdb_default_cmp(txn, FREE_DBI);
7181 if ((flags & VALID_FLAGS) != flags)
7187 if (flags & PERSISTENT_FLAGS) {
7188 uint16_t f2 = flags & PERSISTENT_FLAGS;
7189 /* make sure flag changes get committed */
7190 if ((txn->mt_dbs[MAIN_DBI].md_flags | f2) != txn->mt_dbs[MAIN_DBI].md_flags) {
7191 txn->mt_dbs[MAIN_DBI].md_flags |= f2;
7192 txn->mt_flags |= MDB_TXN_DIRTY;
7195 mdb_default_cmp(txn, MAIN_DBI);
7199 if (txn->mt_dbxs[MAIN_DBI].md_cmp == NULL) {
7200 mdb_default_cmp(txn, MAIN_DBI);
7203 /* Is the DB already open? */
7205 for (i=2; i<txn->mt_numdbs; i++) {
7206 if (!txn->mt_dbxs[i].md_name.mv_size) {
7207 /* Remember this free slot */
7208 if (!unused) unused = i;
7211 if (len == txn->mt_dbxs[i].md_name.mv_size &&
7212 !strncmp(name, txn->mt_dbxs[i].md_name.mv_data, len)) {
7218 /* If no free slot and max hit, fail */
7219 if (!unused && txn->mt_numdbs >= txn->mt_env->me_maxdbs)
7220 return MDB_DBS_FULL;
7222 /* Cannot mix named databases with some mainDB flags */
7223 if (txn->mt_dbs[MAIN_DBI].md_flags & (MDB_DUPSORT|MDB_INTEGERKEY))
7224 return (flags & MDB_CREATE) ? MDB_INCOMPATIBLE : MDB_NOTFOUND;
7226 /* Find the DB info */
7227 dbflag = DB_NEW|DB_VALID;
7230 key.mv_data = (void *)name;
7231 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
7232 rc = mdb_cursor_set(&mc, &key, &data, MDB_SET, &exact);
7233 if (rc == MDB_SUCCESS) {
7234 /* make sure this is actually a DB */
7235 MDB_node *node = NODEPTR(mc.mc_pg[mc.mc_top], mc.mc_ki[mc.mc_top]);
7236 if (!(node->mn_flags & F_SUBDATA))
7238 } else if (rc == MDB_NOTFOUND && (flags & MDB_CREATE)) {
7239 /* Create if requested */
7241 data.mv_size = sizeof(MDB_db);
7242 data.mv_data = &dummy;
7243 memset(&dummy, 0, sizeof(dummy));
7244 dummy.md_root = P_INVALID;
7245 dummy.md_flags = flags & PERSISTENT_FLAGS;
7246 rc = mdb_cursor_put(&mc, &key, &data, F_SUBDATA);
7250 /* OK, got info, add to table */
7251 if (rc == MDB_SUCCESS) {
7252 unsigned int slot = unused ? unused : txn->mt_numdbs;
7253 txn->mt_dbxs[slot].md_name.mv_data = strdup(name);
7254 txn->mt_dbxs[slot].md_name.mv_size = len;
7255 txn->mt_dbxs[slot].md_rel = NULL;
7256 txn->mt_dbflags[slot] = dbflag;
7257 memcpy(&txn->mt_dbs[slot], data.mv_data, sizeof(MDB_db));
7259 txn->mt_env->me_dbflags[slot] = txn->mt_dbs[slot].md_flags;
7260 mdb_default_cmp(txn, slot);
7269 int mdb_stat(MDB_txn *txn, MDB_dbi dbi, MDB_stat *arg)
7271 if (txn == NULL || arg == NULL || dbi >= txn->mt_numdbs)
7274 return mdb_stat0(txn->mt_env, &txn->mt_dbs[dbi], arg);
7277 void mdb_dbi_close(MDB_env *env, MDB_dbi dbi)
7280 if (dbi <= MAIN_DBI || dbi >= env->me_maxdbs)
7282 ptr = env->me_dbxs[dbi].md_name.mv_data;
7283 env->me_dbxs[dbi].md_name.mv_data = NULL;
7284 env->me_dbxs[dbi].md_name.mv_size = 0;
7285 env->me_dbflags[dbi] = 0;
7289 /** Add all the DB's pages to the free list.
7290 * @param[in] mc Cursor on the DB to free.
7291 * @param[in] subs non-Zero to check for sub-DBs in this DB.
7292 * @return 0 on success, non-zero on failure.
7295 mdb_drop0(MDB_cursor *mc, int subs)
7299 rc = mdb_page_search(mc, NULL, 0);
7300 if (rc == MDB_SUCCESS) {
7305 /* LEAF2 pages have no nodes, cannot have sub-DBs */
7306 if (IS_LEAF2(mc->mc_pg[mc->mc_top]))
7309 mdb_cursor_copy(mc, &mx);
7310 while (mc->mc_snum > 0) {
7311 if (IS_LEAF(mc->mc_pg[mc->mc_top])) {
7312 for (i=0; i<NUMKEYS(mc->mc_pg[mc->mc_top]); i++) {
7313 ni = NODEPTR(mc->mc_pg[mc->mc_top], i);
7314 if (ni->mn_flags & F_BIGDATA) {
7318 memcpy(&pg, NODEDATA(ni), sizeof(pg));
7319 rc = mdb_page_get(mc->mc_txn, pg, &omp, NULL);
7322 assert(IS_OVERFLOW(omp));
7323 ovpages = omp->mp_pages;
7324 for (j=0; j<ovpages; j++) {
7325 mdb_midl_append(&mc->mc_txn->mt_free_pgs, pg);
7328 } else if (subs && (ni->mn_flags & F_SUBDATA)) {
7329 mdb_xcursor_init1(mc, ni);
7330 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
7336 for (i=0; i<NUMKEYS(mc->mc_pg[mc->mc_top]); i++) {
7338 ni = NODEPTR(mc->mc_pg[mc->mc_top], i);
7341 mdb_midl_append(&mc->mc_txn->mt_free_pgs, pg);
7346 mc->mc_ki[mc->mc_top] = i;
7347 rc = mdb_cursor_sibling(mc, 1);
7349 /* no more siblings, go back to beginning
7350 * of previous level.
7354 for (i=1; i<mc->mc_snum; i++) {
7356 mc->mc_pg[i] = mx.mc_pg[i];
7361 mdb_midl_append(&mc->mc_txn->mt_free_pgs,
7362 mc->mc_db->md_root);
7367 int mdb_drop(MDB_txn *txn, MDB_dbi dbi, int del)
7372 if (!txn || !dbi || dbi >= txn->mt_numdbs || (unsigned)del > 1 || !(txn->mt_dbflags[dbi] & DB_VALID))
7375 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
7378 rc = mdb_cursor_open(txn, dbi, &mc);
7382 rc = mdb_drop0(mc, mc->mc_db->md_flags & MDB_DUPSORT);
7386 /* Can't delete the main DB */
7387 if (del && dbi > MAIN_DBI) {
7388 rc = mdb_del(txn, MAIN_DBI, &mc->mc_dbx->md_name, NULL);
7390 txn->mt_dbflags[dbi] = DB_STALE;
7391 mdb_dbi_close(txn->mt_env, dbi);
7394 /* reset the DB record, mark it dirty */
7395 txn->mt_dbflags[dbi] |= DB_DIRTY;
7396 txn->mt_dbs[dbi].md_depth = 0;
7397 txn->mt_dbs[dbi].md_branch_pages = 0;
7398 txn->mt_dbs[dbi].md_leaf_pages = 0;
7399 txn->mt_dbs[dbi].md_overflow_pages = 0;
7400 txn->mt_dbs[dbi].md_entries = 0;
7401 txn->mt_dbs[dbi].md_root = P_INVALID;
7403 txn->mt_flags |= MDB_TXN_DIRTY;
7406 mdb_cursor_close(mc);
7410 int mdb_set_compare(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
7412 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
7415 txn->mt_dbxs[dbi].md_cmp = cmp;
7419 int mdb_set_dupsort(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
7421 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
7424 txn->mt_dbxs[dbi].md_dcmp = cmp;
7428 int mdb_set_relfunc(MDB_txn *txn, MDB_dbi dbi, MDB_rel_func *rel)
7430 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
7433 txn->mt_dbxs[dbi].md_rel = rel;
7437 int mdb_set_relctx(MDB_txn *txn, MDB_dbi dbi, void *ctx)
7439 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
7442 txn->mt_dbxs[dbi].md_relctx = ctx;